From 68b0fb0c620c7451ef1a6380c43c39de6614db51 Mon Sep 17 00:00:00 2001
From: Dave Barach <dave@barachs.net>
Date: Tue, 28 Feb 2017 15:15:56 -0500
Subject: VPP-598: tcp stack initial commit

Change-Id: I49e5ce0aae6e4ff634024387ceaf7dbc432a0351
Signed-off-by: Dave Barach <dave@barachs.net>
Signed-off-by: Florin Coras <fcoras@cisco.com>
---
 src/svm/svm_fifo.c | 568 +++++++++++++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 568 insertions(+)
 create mode 100644 src/svm/svm_fifo.c

(limited to 'src/svm/svm_fifo.c')

diff --git a/src/svm/svm_fifo.c b/src/svm/svm_fifo.c
new file mode 100644
index 00000000..11f90193
--- /dev/null
+++ b/src/svm/svm_fifo.c
@@ -0,0 +1,568 @@
+/*
+ * Copyright (c) 2016 Cisco and/or its affiliates.
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at:
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include "svm_fifo.h"
+
+/** create an svm fifo, in the current heap. Fails vs blow up the process */
+svm_fifo_t *
+svm_fifo_create (u32 data_size_in_bytes)
+{
+  svm_fifo_t *f;
+  pthread_mutexattr_t attr;
+  pthread_condattr_t cattr;
+
+  f = clib_mem_alloc_aligned_or_null (sizeof (*f) + data_size_in_bytes,
+				      CLIB_CACHE_LINE_BYTES);
+  if (f == 0)
+    return 0;
+
+  memset (f, 0, sizeof (*f) + data_size_in_bytes);
+  f->nitems = data_size_in_bytes;
+  f->ooos_list_head = OOO_SEGMENT_INVALID_INDEX;
+
+  memset (&attr, 0, sizeof (attr));
+  memset (&cattr, 0, sizeof (cattr));
+
+  if (pthread_mutexattr_init (&attr))
+    clib_unix_warning ("mutexattr_init");
+  if (pthread_mutexattr_setpshared (&attr, PTHREAD_PROCESS_SHARED))
+    clib_unix_warning ("pthread_mutexattr_setpshared");
+  if (pthread_mutex_init (&f->mutex, &attr))
+    clib_unix_warning ("mutex_init");
+  if (pthread_mutexattr_destroy (&attr))
+    clib_unix_warning ("mutexattr_destroy");
+  if (pthread_condattr_init (&cattr))
+    clib_unix_warning ("condattr_init");
+  if (pthread_condattr_setpshared (&cattr, PTHREAD_PROCESS_SHARED))
+    clib_unix_warning ("condattr_setpshared");
+  if (pthread_cond_init (&f->condvar, &cattr))
+    clib_unix_warning ("cond_init1");
+  if (pthread_condattr_destroy (&cattr))
+    clib_unix_warning ("cond_init2");
+
+  return (f);
+}
+
+always_inline ooo_segment_t *
+ooo_segment_new (svm_fifo_t * f, u32 start, u32 length)
+{
+  ooo_segment_t *s;
+
+  pool_get (f->ooo_segments, s);
+
+  s->fifo_position = start;
+  s->length = length;
+
+  s->prev = s->next = OOO_SEGMENT_INVALID_INDEX;
+
+  return s;
+}
+
+always_inline void
+ooo_segment_del (svm_fifo_t * f, u32 index)
+{
+  ooo_segment_t *cur, *prev = 0, *next = 0;
+  cur = pool_elt_at_index (f->ooo_segments, index);
+
+  if (cur->next != OOO_SEGMENT_INVALID_INDEX)
+    {
+      next = pool_elt_at_index (f->ooo_segments, cur->next);
+      next->prev = cur->prev;
+    }
+
+  if (cur->prev != OOO_SEGMENT_INVALID_INDEX)
+    {
+      prev = pool_elt_at_index (f->ooo_segments, cur->prev);
+      prev->next = cur->next;
+    }
+  else
+    {
+      f->ooos_list_head = cur->next;
+    }
+
+  pool_put (f->ooo_segments, cur);
+}
+
+/**
+ * Add segment to fifo's out-of-order segment list. Takes care of merging
+ * adjacent segments and removing overlapping ones.
+ */
+static void
+ooo_segment_add (svm_fifo_t * f, u32 offset, u32 length)
+{
+  ooo_segment_t *s, *new_s, *prev, *next, *it;
+  u32 new_index, position, end_offset, s_sof, s_eof, s_index;
+
+  position = (f->tail + offset) % f->nitems;
+  end_offset = offset + length;
+
+  if (f->ooos_list_head == OOO_SEGMENT_INVALID_INDEX)
+    {
+      s = ooo_segment_new (f, position, length);
+      f->ooos_list_head = s - f->ooo_segments;
+      f->ooos_newest = f->ooos_list_head;
+      return;
+    }
+
+  /* Find first segment that starts after new segment */
+  s = pool_elt_at_index (f->ooo_segments, f->ooos_list_head);
+  while (s->next != OOO_SEGMENT_INVALID_INDEX
+	 && ooo_segment_offset (f, s) <= offset)
+    s = pool_elt_at_index (f->ooo_segments, s->next);
+
+  s_index = s - f->ooo_segments;
+  s_sof = ooo_segment_offset (f, s);
+  s_eof = ooo_segment_end_offset (f, s);
+
+  /* No overlap, add before current segment */
+  if (end_offset < s_sof)
+    {
+      new_s = ooo_segment_new (f, position, length);
+      new_index = new_s - f->ooo_segments;
+
+      /* Pool might've moved, get segment again */
+      s = pool_elt_at_index (f->ooo_segments, s_index);
+
+      if (s->prev != OOO_SEGMENT_INVALID_INDEX)
+	{
+	  new_s->prev = s->prev;
+
+	  prev = pool_elt_at_index (f->ooo_segments, new_s->prev);
+	  prev->next = new_index;
+	}
+      else
+	{
+	  /* New head */
+	  f->ooos_list_head = new_index;
+	}
+
+      new_s->next = s - f->ooo_segments;
+      s->prev = new_index;
+      f->ooos_newest = new_index;
+      return;
+    }
+  /* No overlap, add after current segment */
+  else if (s_eof < offset)
+    {
+      new_s = ooo_segment_new (f, position, length);
+      new_index = new_s - f->ooo_segments;
+
+      /* Pool might've moved, get segment again */
+      s = pool_elt_at_index (f->ooo_segments, s_index);
+
+      if (s->next != OOO_SEGMENT_INVALID_INDEX)
+	{
+	  new_s->next = s->next;
+
+	  next = pool_elt_at_index (f->ooo_segments, new_s->next);
+	  next->prev = new_index;
+	}
+
+      new_s->prev = s - f->ooo_segments;
+      s->next = new_index;
+      f->ooos_newest = new_index;
+
+      return;
+    }
+
+  /*
+   * Merge needed
+   */
+
+  /* Merge at head */
+  if (offset <= s_sof)
+    {
+      /* If we have a previous, check if we overlap */
+      if (s->prev != OOO_SEGMENT_INVALID_INDEX)
+	{
+	  prev = pool_elt_at_index (f->ooo_segments, s->prev);
+
+	  /* New segment merges prev and current. Remove previous and
+	   * update position of current. */
+	  if (ooo_segment_end_offset (f, prev) >= offset)
+	    {
+	      s->fifo_position = prev->fifo_position;
+	      s->length = s_eof - ooo_segment_offset (f, prev);
+	      ooo_segment_del (f, s->prev);
+	    }
+	}
+      else
+	{
+	  s->fifo_position = position;
+	  s->length = s_eof - ooo_segment_offset (f, s);
+	}
+
+      /* The new segment's tail may cover multiple smaller ones */
+      if (s_eof < end_offset)
+	{
+	  /* Remove segments completely covered */
+	  it = (s->next != OOO_SEGMENT_INVALID_INDEX) ?
+	    pool_elt_at_index (f->ooo_segments, s->next) : 0;
+	  while (it && ooo_segment_end_offset (f, it) < end_offset)
+	    {
+	      next = (it->next != OOO_SEGMENT_INVALID_INDEX) ?
+		pool_elt_at_index (f->ooo_segments, it->next) : 0;
+	      ooo_segment_del (f, it - f->ooo_segments);
+	      it = next;
+	    }
+
+	  /* Update length. Segment's start might have changed. */
+	  s->length = end_offset - ooo_segment_offset (f, s);
+
+	  /* If partial overlap with last, merge */
+	  if (it && ooo_segment_offset (f, it) < end_offset)
+	    {
+	      s->length +=
+		it->length - (ooo_segment_offset (f, it) - end_offset);
+	      ooo_segment_del (f, it - f->ooo_segments);
+	    }
+	}
+    }
+  /* Last but overlapping previous */
+  else if (s_eof <= end_offset)
+    {
+      s->length = end_offset - ooo_segment_offset (f, s);
+    }
+  /* New segment completely covered by current one */
+  else
+    {
+      /* Do Nothing */
+    }
+
+  /* Most recently updated segment */
+  f->ooos_newest = s - f->ooo_segments;
+}
+
+/**
+ * Removes segments that can now be enqueued because the fifo's tail has
+ * advanced. Returns the number of bytes added to tail.
+ */
+static int
+ooo_segment_try_collect (svm_fifo_t * f, u32 n_bytes_enqueued)
+{
+  ooo_segment_t *s;
+  u32 index, bytes = 0, diff;
+
+  s = pool_elt_at_index (f->ooo_segments, f->ooos_list_head);
+
+  /* If last tail update overlaps one/multiple ooo segments, remove them */
+  diff = (f->nitems + f->tail - s->fifo_position) % f->nitems;
+  while (0 < diff && diff < n_bytes_enqueued)
+    {
+      /* Segment end is beyond the tail. Advance tail and be done */
+      if (diff < s->length)
+	{
+	  f->tail += s->length - diff;
+	  f->tail %= f->nitems;
+	  break;
+	}
+      /* If we have next go on */
+      else if (s->next != OOO_SEGMENT_INVALID_INDEX)
+	{
+	  index = s - f->ooo_segments;
+	  s = pool_elt_at_index (f->ooo_segments, s->next);
+	  diff = (f->nitems + f->tail - s->fifo_position) % f->nitems;
+	  ooo_segment_del (f, index);
+	}
+      /* End of search */
+      else
+	{
+	  break;
+	}
+    }
+
+  /* If tail is adjacent to an ooo segment, 'consume' it */
+  if (diff == 0)
+    {
+      bytes = ((f->nitems - f->cursize) >= s->length) ? s->length :
+	f->nitems - f->cursize;
+
+      f->tail += bytes;
+      f->tail %= f->nitems;
+
+      ooo_segment_del (f, s - f->ooo_segments);
+    }
+
+  return bytes;
+}
+
+static int
+svm_fifo_enqueue_internal (svm_fifo_t * f,
+			   int pid, u32 max_bytes, u8 * copy_from_here)
+{
+  u32 total_copy_bytes, first_copy_bytes, second_copy_bytes;
+  u32 cursize, nitems;
+
+  if (PREDICT_FALSE (f->cursize == f->nitems))
+    return -2;			/* fifo stuffed */
+
+  /* read cursize, which can only decrease while we're working */
+  cursize = f->cursize;
+  nitems = f->nitems;
+
+  /* Number of bytes we're going to copy */
+  total_copy_bytes = (nitems - cursize) < max_bytes ?
+    (nitems - cursize) : max_bytes;
+
+  if (PREDICT_TRUE (copy_from_here != 0))
+    {
+      /* Number of bytes in first copy segment */
+      first_copy_bytes = ((nitems - f->tail) < total_copy_bytes)
+	? (nitems - f->tail) : total_copy_bytes;
+
+      clib_memcpy (&f->data[f->tail], copy_from_here, first_copy_bytes);
+      f->tail += first_copy_bytes;
+      f->tail = (f->tail == nitems) ? 0 : f->tail;
+
+      /* Number of bytes in second copy segment, if any */
+      second_copy_bytes = total_copy_bytes - first_copy_bytes;
+      if (second_copy_bytes)
+	{
+	  clib_memcpy (&f->data[f->tail], copy_from_here + first_copy_bytes,
+		       second_copy_bytes);
+	  f->tail += second_copy_bytes;
+	  f->tail = (f->tail == nitems) ? 0 : f->tail;
+	}
+    }
+  else
+    {
+      /* Account for a zero-copy enqueue done elsewhere */
+      ASSERT (max_bytes <= (nitems - cursize));
+      f->tail += max_bytes;
+      f->tail = f->tail % nitems;
+      total_copy_bytes = max_bytes;
+    }
+
+  /* Any out-of-order segments to collect? */
+  if (PREDICT_FALSE (f->ooos_list_head != OOO_SEGMENT_INVALID_INDEX))
+    total_copy_bytes += ooo_segment_try_collect (f, total_copy_bytes);
+
+  /* Atomically increase the queue length */
+  __sync_fetch_and_add (&f->cursize, total_copy_bytes);
+
+  return (total_copy_bytes);
+}
+
+int
+svm_fifo_enqueue_nowait (svm_fifo_t * f,
+			 int pid, u32 max_bytes, u8 * copy_from_here)
+{
+  return svm_fifo_enqueue_internal (f, pid, max_bytes, copy_from_here);
+}
+
+/** Enqueue a future segment.
+ * Two choices: either copies the entire segment, or copies nothing
+ * Returns 0 of the entire segment was copied
+ * Returns -1 if none of the segment was copied due to lack of space
+ */
+
+static int
+svm_fifo_enqueue_with_offset_internal2 (svm_fifo_t * f,
+					int pid,
+					u32 offset,
+					u32 required_bytes,
+					u8 * copy_from_here)
+{
+  u32 total_copy_bytes, first_copy_bytes, second_copy_bytes;
+  u32 cursize, nitems;
+  u32 tail_plus_offset;
+
+  ASSERT (offset > 0);
+
+  /* read cursize, which can only decrease while we're working */
+  cursize = f->cursize;
+  nitems = f->nitems;
+
+  /* Will this request fit? */
+  if ((required_bytes + offset) > (nitems - cursize))
+    return -1;
+
+  ooo_segment_add (f, offset, required_bytes);
+
+  /* Number of bytes we're going to copy */
+  total_copy_bytes = required_bytes;
+  tail_plus_offset = (f->tail + offset) % nitems;
+
+  /* Number of bytes in first copy segment */
+  first_copy_bytes = ((nitems - tail_plus_offset) < total_copy_bytes)
+    ? (nitems - tail_plus_offset) : total_copy_bytes;
+
+  clib_memcpy (&f->data[tail_plus_offset], copy_from_here, first_copy_bytes);
+
+  /* Number of bytes in second copy segment, if any */
+  second_copy_bytes = total_copy_bytes - first_copy_bytes;
+  if (second_copy_bytes)
+    {
+      tail_plus_offset += first_copy_bytes;
+      tail_plus_offset %= nitems;
+
+      ASSERT (tail_plus_offset == 0);
+
+      clib_memcpy (&f->data[tail_plus_offset],
+		   copy_from_here + first_copy_bytes, second_copy_bytes);
+    }
+
+  return (0);
+}
+
+
+int
+svm_fifo_enqueue_with_offset (svm_fifo_t * f,
+			      int pid,
+			      u32 offset,
+			      u32 required_bytes, u8 * copy_from_here)
+{
+  return svm_fifo_enqueue_with_offset_internal2
+    (f, pid, offset, required_bytes, copy_from_here);
+}
+
+
+static int
+svm_fifo_dequeue_internal2 (svm_fifo_t * f,
+			    int pid, u32 max_bytes, u8 * copy_here)
+{
+  u32 total_copy_bytes, first_copy_bytes, second_copy_bytes;
+  u32 cursize, nitems;
+
+  if (PREDICT_FALSE (f->cursize == 0))
+    return -2;			/* nothing in the fifo */
+
+  /* read cursize, which can only increase while we're working */
+  cursize = f->cursize;
+  nitems = f->nitems;
+
+  /* Number of bytes we're going to copy */
+  total_copy_bytes = (cursize < max_bytes) ? cursize : max_bytes;
+
+  if (PREDICT_TRUE (copy_here != 0))
+    {
+      /* Number of bytes in first copy segment */
+      first_copy_bytes = ((nitems - f->head) < total_copy_bytes)
+	? (nitems - f->head) : total_copy_bytes;
+      clib_memcpy (copy_here, &f->data[f->head], first_copy_bytes);
+      f->head += first_copy_bytes;
+      f->head = (f->head == nitems) ? 0 : f->head;
+
+      /* Number of bytes in second copy segment, if any */
+      second_copy_bytes = total_copy_bytes - first_copy_bytes;
+      if (second_copy_bytes)
+	{
+	  clib_memcpy (copy_here + first_copy_bytes,
+		       &f->data[f->head], second_copy_bytes);
+	  f->head += second_copy_bytes;
+	  f->head = (f->head == nitems) ? 0 : f->head;
+	}
+    }
+  else
+    {
+      /* Account for a zero-copy dequeue done elsewhere */
+      ASSERT (max_bytes <= cursize);
+      f->head += max_bytes;
+      f->head = f->head % nitems;
+      cursize -= max_bytes;
+      total_copy_bytes = max_bytes;
+    }
+
+  __sync_fetch_and_sub (&f->cursize, total_copy_bytes);
+
+  return (total_copy_bytes);
+}
+
+int
+svm_fifo_dequeue_nowait (svm_fifo_t * f,
+			 int pid, u32 max_bytes, u8 * copy_here)
+{
+  return svm_fifo_dequeue_internal2 (f, pid, max_bytes, copy_here);
+}
+
+int
+svm_fifo_peek (svm_fifo_t * f, int pid, u32 offset, u32 max_bytes,
+	       u8 * copy_here)
+{
+  u32 total_copy_bytes, first_copy_bytes, second_copy_bytes;
+  u32 cursize, nitems;
+
+  if (PREDICT_FALSE (f->cursize == 0))
+    return -2;			/* nothing in the fifo */
+
+  /* read cursize, which can only increase while we're working */
+  cursize = f->cursize;
+  nitems = f->nitems;
+
+  /* Number of bytes we're going to copy */
+  total_copy_bytes = (cursize < max_bytes) ? cursize : max_bytes;
+
+  if (PREDICT_TRUE (copy_here != 0))
+    {
+      /* Number of bytes in first copy segment */
+      first_copy_bytes =
+	((nitems - f->head) < total_copy_bytes) ?
+	(nitems - f->head) : total_copy_bytes;
+      clib_memcpy (copy_here, &f->data[f->head], first_copy_bytes);
+
+      /* Number of bytes in second copy segment, if any */
+      second_copy_bytes = total_copy_bytes - first_copy_bytes;
+      if (second_copy_bytes)
+	{
+	  clib_memcpy (copy_here + first_copy_bytes, &f->data[0],
+		       second_copy_bytes);
+	}
+    }
+  return total_copy_bytes;
+}
+
+int
+svm_fifo_dequeue_drop (svm_fifo_t * f, int pid, u32 max_bytes)
+{
+  u32 total_drop_bytes, first_drop_bytes, second_drop_bytes;
+  u32 cursize, nitems;
+
+  if (PREDICT_FALSE (f->cursize == 0))
+    return -2;			/* nothing in the fifo */
+
+  /* read cursize, which can only increase while we're working */
+  cursize = f->cursize;
+  nitems = f->nitems;
+
+  /* Number of bytes we're going to drop */
+  total_drop_bytes = (cursize < max_bytes) ? cursize : max_bytes;
+
+  /* Number of bytes in first copy segment */
+  first_drop_bytes =
+    ((nitems - f->head) < total_drop_bytes) ?
+    (nitems - f->head) : total_drop_bytes;
+  f->head += first_drop_bytes;
+  f->head = (f->head == nitems) ? 0 : f->head;
+
+  /* Number of bytes in second drop segment, if any */
+  second_drop_bytes = total_drop_bytes - first_drop_bytes;
+  if (second_drop_bytes)
+    {
+      f->head += second_drop_bytes;
+      f->head = (f->head == nitems) ? 0 : f->head;
+    }
+
+  __sync_fetch_and_sub (&f->cursize, total_drop_bytes);
+
+  return total_drop_bytes;
+}
+
+/*
+ * fd.io coding-style-patch-verification: ON
+ *
+ * Local Variables:
+ * eval: (c-set-style "gnu")
+ * End:
+ */
-- 
cgit 1.2.3-korg


From e04c29942af6a130591059679531c9ffa3d7237a Mon Sep 17 00:00:00 2001
From: Florin Coras <fcoras@cisco.com>
Date: Wed, 1 Mar 2017 08:17:34 -0800
Subject: Cleanup URI code and TCP bugfixing

- Add CLI/API to enable session layer, by default it's disabled
- Improve rcv wnd computation
- Improvements to tx path
- URI code cleanup
- Builtin test tcp server
- Improve src port allocation

Change-Id: I2ace498e76a0771d4c31a8075cc14fe33d7dfa38
Signed-off-by: Florin Coras <fcoras@cisco.com>
---
 src/scripts/vnet/uri/dummy_app.py        |  65 +++
 src/scripts/vnet/uri/tcp_server          |   1 +
 src/svm/svm_fifo.c                       |   6 +-
 src/uri.am                               |  10 +-
 src/uri/uri_tcp_test.c                   | 792 +++++++++++++------------
 src/uri/uri_udp_test.c                   | 442 +++++++++++++-
 src/uri/uri_udp_test2.c                  | 954 -------------------------------
 src/uri/uritest.c                        | 484 ----------------
 src/vnet.am                              |   1 +
 src/vnet/api_errno.h                     |   3 +-
 src/vnet/session/application.c           |  27 +-
 src/vnet/session/application.h           |   1 +
 src/vnet/session/application_interface.c |   6 +-
 src/vnet/session/node.c                  |  57 +-
 src/vnet/session/session.api             |  22 +
 src/vnet/session/session.c               |  86 ++-
 src/vnet/session/session.h               |  23 +-
 src/vnet/session/session_api.c           |  59 +-
 src/vnet/session/session_cli.c           |  63 +-
 src/vnet/tcp/builtin_server.c            | 135 +++++
 src/vnet/tcp/tcp.c                       |  48 +-
 src/vnet/tcp/tcp.h                       |   4 +-
 src/vnet/tcp/tcp_input.c                 |  56 +-
 src/vnet/tcp/tcp_output.c                |  90 ++-
 24 files changed, 1460 insertions(+), 1975 deletions(-)
 create mode 100644 src/scripts/vnet/uri/dummy_app.py
 delete mode 100644 src/uri/uri_udp_test2.c
 delete mode 100644 src/uri/uritest.c
 create mode 100644 src/vnet/tcp/builtin_server.c

(limited to 'src/svm/svm_fifo.c')

diff --git a/src/scripts/vnet/uri/dummy_app.py b/src/scripts/vnet/uri/dummy_app.py
new file mode 100644
index 00000000..b80fbb28
--- /dev/null
+++ b/src/scripts/vnet/uri/dummy_app.py
@@ -0,0 +1,65 @@
+#!/usr/bin/env python
+
+import socket
+import sys
+import bitstring
+
+# action can be reflect or drop 
+action = "drop"
+
+def handle_connection (connection, client_address):
+    print("Received connection from {}".format(repr(client_address)))
+    try:
+        while True:
+            data = connection.recv(4096)
+            if not data:
+                break;
+            if (action != "drop"):
+                connection.sendall(data)
+    finally:
+        connection.close()
+        
+def run_server(ip, port):
+    print("Starting server {}:{}".format(repr(ip), repr(port)))
+    sock = socket.socket(socket.AF_INET, socket.SOCK_STREAM)
+    server_address = (ip, int(port))
+    sock.bind(server_address)
+    sock.listen(1)
+    
+    while True:
+        connection, client_address = sock.accept()
+        handle_connection (connection, client_address)
+
+def prepare_data():
+    buf = []
+    for i in range (0, pow(2, 16)):
+        buf.append(i & 0xff)
+    return bytearray(buf)
+
+def run_client(ip, port):
+    print("Starting client {}:{}".format(repr(ip), repr(port)))
+    sock = socket.socket(socket.AF_INET, socket.SOCK_STREAM)
+    server_address = ("6.0.1.1", 1234)
+    sock.connect(server_address)
+    
+    data = prepare_data()
+    try:
+        sock.sendall(data)
+    finally:
+        sock.close()
+    
+def run(mode, ip, port):
+    if (mode == "server"):
+        run_server (ip, port)
+    elif (mode == "client"):
+        run_client (ip, port)
+    else:
+        raise Exception("Unknown mode. Only client and server supported")
+
+if __name__ == "__main__":
+    if (len(sys.argv)) < 4:
+        raise Exception("Usage: ./dummy_app <mode> <ip> <port> [<action>]")
+    if (len(sys.argv) == 5):
+        action = sys.argv[4]
+
+    run (sys.argv[1], sys.argv[2], sys.argv[3])
diff --git a/src/scripts/vnet/uri/tcp_server b/src/scripts/vnet/uri/tcp_server
index 7f5a86de..c29afc6f 100644
--- a/src/scripts/vnet/uri/tcp_server
+++ b/src/scripts/vnet/uri/tcp_server
@@ -2,3 +2,4 @@ create host-interface name vpp1
 set int state host-vpp1 up
 set int ip address host-vpp1 6.0.1.1/24
 trace add af-packet-input 10
+session enable
diff --git a/src/svm/svm_fifo.c b/src/svm/svm_fifo.c
index 11f90193..e3f534b1 100644
--- a/src/svm/svm_fifo.c
+++ b/src/svm/svm_fifo.c
@@ -508,9 +508,9 @@ svm_fifo_peek (svm_fifo_t * f, int pid, u32 offset, u32 max_bytes,
     {
       /* Number of bytes in first copy segment */
       first_copy_bytes =
-	((nitems - f->head) < total_copy_bytes) ?
-	(nitems - f->head) : total_copy_bytes;
-      clib_memcpy (copy_here, &f->data[f->head], first_copy_bytes);
+	((nitems - f->head + offset) < total_copy_bytes) ?
+	(nitems - f->head + offset) : total_copy_bytes;
+      clib_memcpy (copy_here, &f->data[f->head + offset], first_copy_bytes);
 
       /* Number of bytes in second copy segment, if any */
       second_copy_bytes = total_copy_bytes - first_copy_bytes;
diff --git a/src/uri.am b/src/uri.am
index 8cdd77c6..09b5b15b 100644
--- a/src/uri.am
+++ b/src/uri.am
@@ -11,12 +11,12 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-noinst_PROGRAMS += uri_udp_test2 uri_tcp_test
+noinst_PROGRAMS += uri_udp_test uri_tcp_test
 
-uri_udp_test2_SOURCES = uri/uri_udp_test2.c			
-uri_udp_test2_LDADD = libvlibmemoryclient.la libvlibapi.la libsvm.la \
-	libvppinfra.la -lpthread -lm -lrt 
+uri_udp_test_SOURCES = uri/uri_udp_test.c
+uri_udp_test_LDADD = libvlibmemoryclient.la libvlibapi.la libsvm.la \
+	libvppinfra.la -lpthread -lm -lrt
 
 uri_tcp_test_SOURCES = uri/uri_tcp_test.c
 uri_tcp_test_LDADD = libvlibmemoryclient.la libvlibapi.la libsvm.la \
-	libvppinfra.la -lpthread -lm -lrt 
+	libvppinfra.la -lpthread -lm -lrt
diff --git a/src/uri/uri_tcp_test.c b/src/uri/uri_tcp_test.c
index ed5a37d8..6c9cf1db 100644
--- a/src/uri/uri_tcp_test.c
+++ b/src/uri/uri_tcp_test.c
@@ -20,16 +20,15 @@
 #include <svm/svm_fifo_segment.h>
 #include <vlibmemory/api.h>
 #include <vpp/api/vpe_msg_enum.h>
+#include <vnet/session/application_interface.h>
 
-#include "../vnet/session/application_interface.h"
-
-#define vl_typedefs             /* define message structures */
+#define vl_typedefs		/* define message structures */
 #include <vpp/api/vpe_all_api_h.h>
 #undef vl_typedefs
 
 /* declare message handlers for each api */
 
-#define vl_endianfun            /* define message structures */
+#define vl_endianfun		/* define message structures */
 #include <vpp/api/vpe_all_api_h.h>
 #undef vl_endianfun
 
@@ -45,8 +44,8 @@ vlib_main_t **vlib_mains;
 
 typedef struct
 {
-  svm_fifo_t * server_rx_fifo;
-  svm_fifo_t * server_tx_fifo;
+  svm_fifo_t *server_rx_fifo;
+  svm_fifo_t *server_tx_fifo;
 
   u32 vpp_session_index;
   u32 vpp_session_thread;
@@ -69,19 +68,19 @@ typedef struct
   u32 my_client_index;
 
   /* The URI we're playing with */
-  u8 * uri;
+  u8 *uri;
 
   /* Session pool */
-  session_t * sessions;
+  session_t *sessions;
 
   /* Hash table for disconnect processing */
-  uword * session_index_by_vpp_handles;
+  uword *session_index_by_vpp_handles;
 
   /* intermediate rx buffer */
-  u8 * rx_buf;
+  u8 *rx_buf;
 
   /* URI for slave's connect */
-  u8 * connect_uri;
+  u8 *connect_uri;
 
   u32 connected_session_index;
 
@@ -91,10 +90,10 @@ typedef struct
   int drop_packets;
 
   /* Our event queue */
-  unix_shared_memory_queue_t * our_event_queue;
+  unix_shared_memory_queue_t *our_event_queue;
 
   /* $$$ single thread only for the moment */
-  unix_shared_memory_queue_t * vpp_event_queue;
+  unix_shared_memory_queue_t *vpp_event_queue;
 
   pid_t my_pid;
 
@@ -111,12 +110,15 @@ typedef struct
   u32 configured_segment_size;
 
   /* VNET_API_ERROR_FOO -> "Foo" hash table */
-  uword * error_string_by_error_number;
-
-  /* convenience */
-  svm_fifo_segment_main_t * segment_main;
+  uword *error_string_by_error_number;
 
   u8 *connect_test_data;
+  pthread_t client_rx_thread_handle;
+  u32 client_bytes_received;
+  u8 test_return_packets;
+
+  /* convenience */
+  svm_fifo_segment_main_t *segment_main;
 } uri_tcp_test_main_t;
 
 uri_tcp_test_main_t uri_tcp_test_main;
@@ -141,7 +143,7 @@ wait_for_state_change (uri_tcp_test_main_t * utm, connection_state_t state)
   while (clib_time_now (&utm->clib_time) < timeout)
     {
       if (utm->state == state)
-        return 0;
+	return 0;
       if (utm->state == STATE_FAILED)
 	return -1;
     }
@@ -209,7 +211,7 @@ connect_to_vpp (char *name)
 }
 
 static void
-vl_api_map_another_segment_t_handler (vl_api_map_another_segment_t *mp)
+vl_api_map_another_segment_t_handler (vl_api_map_another_segment_t * mp)
 {
   svm_fifo_segment_create_args_t _a, *a = &_a;
   int rv;
@@ -221,24 +223,24 @@ vl_api_map_another_segment_t_handler (vl_api_map_another_segment_t *mp)
   if (rv)
     {
       clib_warning ("svm_fifo_segment_attach ('%s') failed",
-                    mp->segment_name);
+		    mp->segment_name);
       return;
     }
   clib_warning ("Mapped new segment '%s' size %d", mp->segment_name,
-                mp->segment_size);
+		mp->segment_size);
 }
 
 static void
 vl_api_disconnect_session_t_handler (vl_api_disconnect_session_t * mp)
 {
   uri_tcp_test_main_t *utm = &uri_tcp_test_main;
-  session_t * session;
-  vl_api_disconnect_session_reply_t * rmp;
-  uword * p;
+  session_t *session;
+  vl_api_disconnect_session_reply_t *rmp;
+  uword *p;
   int rv = 0;
   u64 key;
 
-  key = (((u64)mp->session_thread_index) << 32) | (u64)mp->session_index;
+  key = (((u64) mp->session_thread_index) << 32) | (u64) mp->session_index;
 
   p = hash_get (utm->session_index_by_vpp_handles, key);
 
@@ -254,6 +256,8 @@ vl_api_disconnect_session_t_handler (vl_api_disconnect_session_t * mp)
       rv = -11;
     }
 
+  utm->time_to_stop = 1;
+
   rmp = vl_msg_api_alloc (sizeof (*rmp));
   memset (rmp, 0, sizeof (*rmp));
 
@@ -261,32 +265,32 @@ vl_api_disconnect_session_t_handler (vl_api_disconnect_session_t * mp)
   rmp->retval = rv;
   rmp->session_index = mp->session_index;
   rmp->session_thread_index = mp->session_thread_index;
-  vl_msg_api_send_shmem (utm->vl_input_queue, (u8 *)&rmp);
+  vl_msg_api_send_shmem (utm->vl_input_queue, (u8 *) & rmp);
 }
 
 static void
 vl_api_reset_session_t_handler (vl_api_reset_session_t * mp)
 {
   uri_tcp_test_main_t *utm = &uri_tcp_test_main;
-  session_t * session;
-  vl_api_reset_session_reply_t * rmp;
-  uword * p;
+  session_t *session;
+  vl_api_reset_session_reply_t *rmp;
+  uword *p;
   int rv = 0;
   u64 key;
 
-  key = (((u64)mp->session_thread_index) << 32) | (u64)mp->session_index;
+  key = (((u64) mp->session_thread_index) << 32) | (u64) mp->session_index;
 
-  p = hash_get(utm->session_index_by_vpp_handles, key);
+  p = hash_get (utm->session_index_by_vpp_handles, key);
 
   if (p)
     {
-      session = pool_elt_at_index(utm->sessions, p[0]);
-      hash_unset(utm->session_index_by_vpp_handles, key);
-      pool_put(utm->sessions, session);
+      session = pool_elt_at_index (utm->sessions, p[0]);
+      hash_unset (utm->session_index_by_vpp_handles, key);
+      pool_put (utm->sessions, session);
     }
   else
     {
-      clib_warning("couldn't find session key %llx", key);
+      clib_warning ("couldn't find session key %llx", key);
       rv = -11;
     }
 
@@ -296,301 +300,95 @@ vl_api_reset_session_t_handler (vl_api_reset_session_t * mp)
   rmp->retval = rv;
   rmp->session_index = mp->session_index;
   rmp->session_thread_index = mp->session_thread_index;
-  vl_msg_api_send_shmem (utm->vl_input_queue, (u8 *)&rmp);
+  vl_msg_api_send_shmem (utm->vl_input_queue, (u8 *) & rmp);
 }
 
 void
-handle_fifo_event_connect_rx (uri_tcp_test_main_t *utm, session_fifo_event_t * e)
+client_handle_fifo_event_rx (uri_tcp_test_main_t * utm,
+			     session_fifo_event_t * e)
 {
-  svm_fifo_t * rx_fifo;
-  int n_read, bytes;
+  svm_fifo_t *rx_fifo;
+  int n_read, bytes, i;
 
   rx_fifo = e->fifo;
 
   bytes = e->enqueue_length;
   do
     {
-      n_read = svm_fifo_dequeue_nowait (rx_fifo, 0, vec_len(utm->rx_buf),
-                                         utm->rx_buf);
+      n_read = svm_fifo_dequeue_nowait (rx_fifo, 0, vec_len (utm->rx_buf),
+					utm->rx_buf);
       if (n_read > 0)
-        bytes -= n_read;
+	{
+	  bytes -= n_read;
+	  for (i = 0; i < n_read; i++)
+	    {
+	      if (utm->rx_buf[i] != ((utm->client_bytes_received + i) & 0xff))
+		{
+		  clib_warning ("error at byte %lld, 0x%x not 0x%x",
+				utm->client_bytes_received + i,
+				utm->rx_buf[i],
+				((utm->client_bytes_received + i) & 0xff));
+		}
+	    }
+	  utm->client_bytes_received += n_read;
+	}
+
     }
   while (n_read < 0 || bytes > 0);
-
-  //      bytes_to_read = svm_fifo_max_dequeue (rx_fifo);
-  //
-  //      bytes_to_read = vec_len(utm->rx_buf) > bytes_to_read ?
-  //        bytes_to_read : vec_len(utm->rx_buf);
-  //
-  //      buffer_offset = 0;
-  //      while (bytes_to_read > 0)
-  //        {
-  //          rv = svm_fifo_dequeue_nowait2 (rx_fifo, mypid,
-  //                                         bytes_to_read,
-  //                                         utm->rx_buf + buffer_offset);
-  //          if (rv > 0)
-  //            {
-  //              bytes_to_read -= rv;
-  //              buffer_offset += rv;
-  //              bytes_received += rv;
-  //            }
-  //        }
-
-
-  //  while (bytes_received < bytes_sent)
-  //    {
-  //      rv = svm_fifo_dequeue_nowait2 (rx_fifo, mypid,
-  //                                     vec_len (utm->rx_buf),
-  //                                     utm->rx_buf);
-  //      if (rv > 0)
-  //        {
-  //#if CLIB_DEBUG > 0
-  //          int j;
-  //          for (j = 0; j < rv; j++)
-  //            {
-  //              if (utm->rx_buf[j] != ((bytes_received + j) & 0xff))
-  //                {
-  //                  clib_warning ("error at byte %lld, 0x%x not 0x%x",
-  //                                bytes_received + j,
-  //                                utm->rx_buf[j],
-  //                                ((bytes_received + j )&0xff));
-  //                }
-  //            }
-  //#endif
-  //          bytes_received += (u64) rv;
-  //        }
-  //    }
 }
 
 void
-handle_connect_event_queue (uri_tcp_test_main_t * utm)
+client_handle_event_queue (uri_tcp_test_main_t * utm)
 {
   session_fifo_event_t _e, *e = &_e;;
 
-  unix_shared_memory_queue_sub (utm->our_event_queue, (u8 *) e, 0 /* nowait */);
+  unix_shared_memory_queue_sub (utm->our_event_queue, (u8 *) e,
+				0 /* nowait */ );
   switch (e->event_type)
     {
     case FIFO_EVENT_SERVER_RX:
-      handle_fifo_event_connect_rx (utm, e);
+      client_handle_fifo_event_rx (utm, e);
       break;
 
     case FIFO_EVENT_SERVER_EXIT:
       return;
 
     default:
-      clib_warning("unknown event type %d", e->event_type);
+      clib_warning ("unknown event type %d", e->event_type);
       break;
     }
 }
 
-void
-uri_tcp_connect_send (uri_tcp_test_main_t *utm)
-{
-  u8 *test_data = utm->connect_test_data;
-  u64 bytes_sent = 0;
-  int rv;
-  int mypid = getpid();
-  session_t * session;
-  svm_fifo_t *tx_fifo;
-  int buffer_offset, bytes_to_send = 0;
-  session_fifo_event_t evt;
-  static int serial_number = 0;
-  int i;
-  u32 max_chunk = 64 << 10, write;
-
-  session = pool_elt_at_index (utm->sessions, utm->connected_session_index);
-  tx_fifo = session->server_tx_fifo;
-
-  vec_validate (utm->rx_buf, vec_len (test_data) - 1);
-
-  for (i = 0; i < 10; i++)
-    {
-      bytes_to_send = vec_len (test_data);
-      buffer_offset = 0;
-      while (bytes_to_send > 0)
-        {
-          write = bytes_to_send > max_chunk ? max_chunk : bytes_to_send;
-          rv = svm_fifo_enqueue_nowait (tx_fifo, mypid, write,
-                                         test_data + buffer_offset);
-
-          if (rv > 0)
-            {
-              bytes_to_send -= rv;
-              buffer_offset += rv;
-              bytes_sent += rv;
-
-              /* Fabricate TX event, send to vpp */
-              evt.fifo = tx_fifo;
-              evt.event_type = FIFO_EVENT_SERVER_TX;
-              /* $$$$ for event logging */
-              evt.enqueue_length = rv;
-              evt.event_id = serial_number++;
-
-              unix_shared_memory_queue_add (utm->vpp_event_queue, (u8 *) &evt,
-                                            0 /* do wait for mutex */);
-            }
-        }
-    }
-}
-
-static void
-uri_tcp_client_test (uri_tcp_test_main_t * utm)
-{
-  vl_api_connect_uri_t * cmp;
-  vl_api_disconnect_session_t *dmp;
-  session_t *connected_session;
-  int i;
-
-  cmp = vl_msg_api_alloc (sizeof (*cmp));
-  memset (cmp, 0, sizeof (*cmp));
-
-  cmp->_vl_msg_id = ntohs (VL_API_CONNECT_URI);
-  cmp->client_index = utm->my_client_index;
-  cmp->context = ntohl(0xfeedface);
-  memcpy (cmp->uri, utm->connect_uri, vec_len (utm->connect_uri));
-  vl_msg_api_send_shmem (utm->vl_input_queue, (u8 *)&cmp);
-
-  if (wait_for_state_change (utm, STATE_READY))
-    {
-      return;
-    }
-
-  /* Init test data */
-  vec_validate (utm->connect_test_data, 64 * 1024 - 1);
-  for (i = 0; i < vec_len (utm->connect_test_data); i++)
-    utm->connect_test_data[i] = i & 0xff;
-
-  /* Start reader thread */
-  /* handle_connect_event_queue (utm); */
-
-  /* Start send */
-  uri_tcp_connect_send (utm);
-
-  /* Disconnect */
-  connected_session = pool_elt_at_index(utm->sessions,
-					utm->connected_session_index);
-  dmp = vl_msg_api_alloc (sizeof (*dmp));
-  memset (dmp, 0, sizeof (*dmp));
-  dmp->_vl_msg_id = ntohs (VL_API_DISCONNECT_SESSION);
-  dmp->client_index = utm->my_client_index;
-  dmp->session_index = connected_session->vpp_session_index;
-  dmp->session_thread_index = connected_session->vpp_session_thread;
-  vl_msg_api_send_shmem (utm->vl_input_queue, (u8 *)&dmp);
-}
-
-void
-handle_fifo_event_server_rx (uri_tcp_test_main_t *utm, session_fifo_event_t * e)
-{
-  svm_fifo_t * rx_fifo, * tx_fifo;
-  int n_read;
-
-  session_fifo_event_t evt;
-  unix_shared_memory_queue_t *q;
-  int rv, bytes;
-
-  rx_fifo = e->fifo;
-  tx_fifo = utm->sessions[rx_fifo->client_session_index].server_tx_fifo;
-
-  bytes = e->enqueue_length;
-  do
-    {
-      n_read = svm_fifo_dequeue_nowait (rx_fifo, 0, vec_len(utm->rx_buf),
-                                         utm->rx_buf);
-
-      /* Reflect if a non-drop session */
-      if (!utm->drop_packets && n_read > 0)
-        {
-          do
-            {
-              rv = svm_fifo_enqueue_nowait (tx_fifo, 0, n_read, utm->rx_buf);
-            }
-          while (rv == -2);
-
-          /* Fabricate TX event, send to vpp */
-          evt.fifo = tx_fifo;
-          evt.event_type = FIFO_EVENT_SERVER_TX;
-          /* $$$$ for event logging */
-          evt.enqueue_length = n_read;
-          evt.event_id = e->event_id;
-          q = utm->vpp_event_queue;
-          unix_shared_memory_queue_add (q, (u8 *) &evt, 0 /* do wait for mutex */);
-        }
-
-      if (n_read > 0)
-        bytes -= n_read;
-    }
-  while (n_read < 0 || bytes > 0);
-}
-
-void
-handle_event_queue (uri_tcp_test_main_t * utm)
+static void *
+client_rx_thread_fn (void *arg)
 {
-  session_fifo_event_t _e, *e = &_e;;
+  session_fifo_event_t _e, *e = &_e;
+  uri_tcp_test_main_t *utm = &uri_tcp_test_main;
 
+  utm->client_bytes_received = 0;
   while (1)
     {
-      unix_shared_memory_queue_sub (utm->our_event_queue, (u8 *)e,
-                                    0 /* nowait */);
+      unix_shared_memory_queue_sub (utm->our_event_queue, (u8 *) e,
+				    0 /* nowait */ );
       switch (e->event_type)
-        {
-        case FIFO_EVENT_SERVER_RX:
-          handle_fifo_event_server_rx (utm, e);
-          break;
-
-        case FIFO_EVENT_SERVER_EXIT:
-          return;
-
-        default:
-          clib_warning ("unknown event type %d", e->event_type);
-          break;
-        }
-      if (PREDICT_FALSE(utm->time_to_stop == 1))
-        break;
-      if (PREDICT_FALSE(utm->time_to_print_stats == 1))
-        {
-          utm->time_to_print_stats = 0;
-          fformat(stdout, "%d connections\n", pool_elts (utm->sessions));
-        }
+	{
+	case FIFO_EVENT_SERVER_RX:
+	  client_handle_fifo_event_rx (utm, e);
+	  break;
+
+	case FIFO_EVENT_SERVER_EXIT:
+	  return 0;
+	default:
+	  clib_warning ("unknown event type %d", e->event_type);
+	  break;
+	}
+
+      if (PREDICT_FALSE (utm->time_to_stop == 1))
+	break;
     }
+  pthread_exit (0);
 }
 
-static void
-vl_api_bind_uri_reply_t_handler (vl_api_bind_uri_reply_t * mp)
-{
-  uri_tcp_test_main_t *utm = &uri_tcp_test_main;
-  svm_fifo_segment_create_args_t _a, *a = &_a;
-  int rv;
-
-  if (mp->retval)
-    {
-      clib_warning("bind failed: %d", mp->retval);
-      return;
-    }
-
-  if (mp->segment_name_length == 0)
-    {
-      clib_warning("segment_name_length zero");
-      return;
-    }
-
-  a->segment_name = (char *) mp->segment_name;
-  a->segment_size = mp->segment_size;
-
-  ASSERT(mp->server_event_queue_address);
-
-  /* Attach to the segment vpp created */
-  rv = svm_fifo_segment_attach (a);
-  if (rv)
-    {
-      clib_warning("svm_fifo_segment_attach ('%s') failed", mp->segment_name);
-      return;
-    }
-
-  utm->our_event_queue =
-      (unix_shared_memory_queue_t *) mp->server_event_queue_address;
-
-  utm->state = STATE_READY;
-}
 
 static void
 vl_api_connect_uri_reply_t_handler (vl_api_connect_uri_reply_t * mp)
@@ -601,6 +399,7 @@ vl_api_connect_uri_reply_t_handler (vl_api_connect_uri_reply_t * mp)
   u32 session_index;
   svm_fifo_t *rx_fifo, *tx_fifo;
   int rv;
+  u64 key;
 
   if (mp->retval)
     {
@@ -608,6 +407,7 @@ vl_api_connect_uri_reply_t_handler (vl_api_connect_uri_reply_t * mp)
       utm->state = STATE_FAILED;
       return;
     }
+
   /*
    * Attatch to segment
    */
@@ -622,14 +422,14 @@ vl_api_connect_uri_reply_t_handler (vl_api_connect_uri_reply_t * mp)
   a->segment_name = (char *) mp->segment_name;
   a->segment_size = mp->segment_size;
 
-  ASSERT(mp->client_event_queue_address);
+  ASSERT (mp->client_event_queue_address);
 
   /* Attach to the segment vpp created */
   rv = svm_fifo_segment_attach (a);
   if (rv)
     {
       clib_warning ("svm_fifo_segment_attach ('%s') failed",
-                    mp->segment_name);
+		    mp->segment_name);
       return;
     }
 
@@ -650,9 +450,9 @@ vl_api_connect_uri_reply_t_handler (vl_api_connect_uri_reply_t * mp)
   pool_get (utm->sessions, session);
   session_index = session - utm->sessions;
 
-  rx_fifo = (svm_fifo_t *)mp->server_rx_fifo;
+  rx_fifo = (svm_fifo_t *) mp->server_rx_fifo;
   rx_fifo->client_session_index = session_index;
-  tx_fifo = (svm_fifo_t *)mp->server_tx_fifo;
+  tx_fifo = (svm_fifo_t *) mp->server_tx_fifo;
   tx_fifo->client_session_index = session_index;
 
   session->server_rx_fifo = rx_fifo;
@@ -662,54 +462,193 @@ vl_api_connect_uri_reply_t_handler (vl_api_connect_uri_reply_t * mp)
 
   /* Save handle */
   utm->connected_session_index = session_index;
-
   utm->state = STATE_READY;
+
+  /* Add it to lookup table */
+  key = (((u64) mp->session_thread_index) << 32) | (u64) mp->session_index;
+  hash_set (utm->session_index_by_vpp_handles, key, session_index);
+
+  /* Start RX thread */
+  rv = pthread_create (&utm->client_rx_thread_handle,
+		       NULL /*attr */ , client_rx_thread_fn, 0);
+  if (rv)
+    {
+      clib_warning ("pthread_create returned %d", rv);
+      rv = VNET_API_ERROR_SYSCALL_ERROR_1;
+    }
 }
 
 void
-uri_tcp_bind (uri_tcp_test_main_t *utm)
+client_send_data (uri_tcp_test_main_t * utm)
 {
-  vl_api_bind_uri_t * bmp;
-  u32 fifo_size = 3 << 20;
-  bmp = vl_msg_api_alloc (sizeof (*bmp));
-  memset (bmp, 0, sizeof (*bmp));
+  u8 *test_data = utm->connect_test_data;
+  u64 bytes_sent = 0;
+  int rv;
+  int mypid = getpid ();
+  session_t *session;
+  svm_fifo_t *tx_fifo;
+  int buffer_offset, bytes_to_send = 0;
+  session_fifo_event_t evt;
+  static int serial_number = 0;
+  int i;
+  u32 max_chunk = 64 << 10, write;
 
-  bmp->_vl_msg_id = ntohs (VL_API_BIND_URI);
-  bmp->client_index = utm->my_client_index;
-  bmp->context = ntohl(0xfeedface);
-  bmp->initial_segment_size = 256<<20;    /* size of initial segment */
-  bmp->options[SESSION_OPTIONS_FLAGS] =
-    SESSION_OPTIONS_FLAGS_USE_FIFO | SESSION_OPTIONS_FLAGS_ADD_SEGMENT;
-  bmp->options[SESSION_OPTIONS_RX_FIFO_SIZE] = fifo_size;
-  bmp->options[SESSION_OPTIONS_TX_FIFO_SIZE] = fifo_size;
-  bmp->options[SESSION_OPTIONS_ADD_SEGMENT_SIZE] = 128<<20;
-  memcpy (bmp->uri, utm->uri, vec_len (utm->uri));
-  vl_msg_api_send_shmem (utm->vl_input_queue, (u8 *)&bmp);
+  session = pool_elt_at_index (utm->sessions, utm->connected_session_index);
+  tx_fifo = session->server_tx_fifo;
+
+  vec_validate (utm->rx_buf, vec_len (test_data) - 1);
+
+  for (i = 0; i < 1; i++)
+    {
+      bytes_to_send = vec_len (test_data);
+      buffer_offset = 0;
+      while (bytes_to_send > 0)
+	{
+	  write = bytes_to_send > max_chunk ? max_chunk : bytes_to_send;
+	  rv = svm_fifo_enqueue_nowait (tx_fifo, mypid, write,
+					test_data + buffer_offset);
+
+	  if (rv > 0)
+	    {
+	      bytes_to_send -= rv;
+	      buffer_offset += rv;
+	      bytes_sent += rv;
+
+	      /* Fabricate TX event, send to vpp */
+	      evt.fifo = tx_fifo;
+	      evt.event_type = FIFO_EVENT_SERVER_TX;
+	      /* $$$$ for event logging */
+	      evt.enqueue_length = rv;
+	      evt.event_id = serial_number++;
+
+	      unix_shared_memory_queue_add (utm->vpp_event_queue,
+					    (u8 *) & evt,
+					    0 /* do wait for mutex */ );
+	    }
+	}
+    }
+
+  if (utm->test_return_packets)
+    {
+      f64 timeout = clib_time_now (&utm->clib_time) + 2;
+
+      /* Wait for the outstanding packets */
+      while (utm->client_bytes_received < vec_len (test_data))
+	{
+	  if (clib_time_now (&utm->clib_time) > timeout)
+	    {
+	      clib_warning ("timed out waiting for the missing packets");
+	      break;
+	    }
+	}
+
+      utm->time_to_stop = 1;
+    }
+}
+
+void
+client_connect (uri_tcp_test_main_t * utm)
+{
+  vl_api_connect_uri_t *cmp;
+  cmp = vl_msg_api_alloc (sizeof (*cmp));
+  memset (cmp, 0, sizeof (*cmp));
+
+  cmp->_vl_msg_id = ntohs (VL_API_CONNECT_URI);
+  cmp->client_index = utm->my_client_index;
+  cmp->context = ntohl (0xfeedface);
+  memcpy (cmp->uri, utm->connect_uri, vec_len (utm->connect_uri));
+  vl_msg_api_send_shmem (utm->vl_input_queue, (u8 *) & cmp);
+}
+
+void
+client_disconnect (uri_tcp_test_main_t * utm)
+{
+  session_t *connected_session;
+  vl_api_disconnect_session_t *dmp;
+  connected_session = pool_elt_at_index (utm->sessions,
+					 utm->connected_session_index);
+  dmp = vl_msg_api_alloc (sizeof (*dmp));
+  memset (dmp, 0, sizeof (*dmp));
+  dmp->_vl_msg_id = ntohs (VL_API_DISCONNECT_SESSION);
+  dmp->client_index = utm->my_client_index;
+  dmp->session_index = connected_session->vpp_session_index;
+  dmp->session_thread_index = connected_session->vpp_session_thread;
+  vl_msg_api_send_shmem (utm->vl_input_queue, (u8 *) & dmp);
+}
+
+static void
+client_test (uri_tcp_test_main_t * utm)
+{
+  int i;
+
+  client_connect (utm);
+
+  if (wait_for_state_change (utm, STATE_READY))
+    {
+      return;
+    }
+
+  /* Init test data */
+  vec_validate (utm->connect_test_data, 64 * 1024 - 1);
+  for (i = 0; i < vec_len (utm->connect_test_data); i++)
+    utm->connect_test_data[i] = i & 0xff;
+
+  /* Start send */
+  client_send_data (utm);
+
+  /* Disconnect */
+  client_disconnect (utm);
 }
 
 static void
-vl_api_unbind_uri_reply_t_handler (vl_api_unbind_uri_reply_t *mp)
+vl_api_bind_uri_reply_t_handler (vl_api_bind_uri_reply_t * mp)
 {
   uri_tcp_test_main_t *utm = &uri_tcp_test_main;
+  svm_fifo_segment_create_args_t _a, *a = &_a;
+  int rv;
 
-  if (mp->retval != 0)
-    clib_warning ("returned %d", ntohl(mp->retval));
+  if (mp->retval)
+    {
+      clib_warning ("bind failed: %d", mp->retval);
+      utm->state = STATE_FAILED;
+      return;
+    }
 
-  utm->state = STATE_START;
+  if (mp->segment_name_length == 0)
+    {
+      clib_warning ("segment_name_length zero");
+      return;
+    }
+
+  a->segment_name = (char *) mp->segment_name;
+  a->segment_size = mp->segment_size;
+
+  ASSERT (mp->server_event_queue_address);
+
+  /* Attach to the segment vpp created */
+  rv = svm_fifo_segment_attach (a);
+  if (rv)
+    {
+      clib_warning ("svm_fifo_segment_attach ('%s') failed",
+		    mp->segment_name);
+      return;
+    }
+
+  utm->our_event_queue =
+    (unix_shared_memory_queue_t *) mp->server_event_queue_address;
+
+  utm->state = STATE_READY;
 }
 
-void
-uri_tcp_unbind (uri_tcp_test_main_t *utm)
+static void
+vl_api_unbind_uri_reply_t_handler (vl_api_unbind_uri_reply_t * mp)
 {
-  vl_api_unbind_uri_t * ump;
+  uri_tcp_test_main_t *utm = &uri_tcp_test_main;
 
-  ump = vl_msg_api_alloc (sizeof (*ump));
-  memset (ump, 0, sizeof (*ump));
+  if (mp->retval != 0)
+    clib_warning ("returned %d", ntohl (mp->retval));
 
-  ump->_vl_msg_id = ntohs (VL_API_UNBIND_URI);
-  ump->client_index = utm->my_client_index;
-  memcpy (ump->uri, utm->uri, vec_len (utm->uri));
-  vl_msg_api_send_shmem (utm->vl_input_queue, (u8 *)&ump);
+  utm->state = STATE_START;
 }
 
 static void
@@ -717,14 +656,14 @@ vl_api_accept_session_t_handler (vl_api_accept_session_t * mp)
 {
   uri_tcp_test_main_t *utm = &uri_tcp_test_main;
   vl_api_accept_session_reply_t *rmp;
-  svm_fifo_t * rx_fifo, * tx_fifo;
-  session_t * session;
+  svm_fifo_t *rx_fifo, *tx_fifo;
+  session_t *session;
   static f64 start_time;
   u64 key;
   u32 session_index;
 
   if (start_time == 0.0)
-      start_time = clib_time_now (&utm->clib_time);
+    start_time = clib_time_now (&utm->clib_time);
 
   utm->vpp_event_queue = (unix_shared_memory_queue_t *)
     mp->vpp_event_queue_address;
@@ -733,45 +672,159 @@ vl_api_accept_session_t_handler (vl_api_accept_session_t * mp)
   pool_get (utm->sessions, session);
   session_index = session - utm->sessions;
 
-  rx_fifo = (svm_fifo_t *)mp->server_rx_fifo;
+  rx_fifo = (svm_fifo_t *) mp->server_rx_fifo;
   rx_fifo->client_session_index = session_index;
-  tx_fifo = (svm_fifo_t *)mp->server_tx_fifo;
+  tx_fifo = (svm_fifo_t *) mp->server_tx_fifo;
   tx_fifo->client_session_index = session_index;
 
   session->server_rx_fifo = rx_fifo;
   session->server_tx_fifo = tx_fifo;
 
   /* Add it to lookup table */
-  key = (((u64)mp->session_thread_index) << 32) | (u64)mp->session_index;
+  key = (((u64) mp->session_thread_index) << 32) | (u64) mp->session_index;
   hash_set (utm->session_index_by_vpp_handles, key, session_index);
 
   utm->state = STATE_READY;
 
   /* Stats printing */
-  if (pool_elts (utm->sessions) && (pool_elts(utm->sessions) % 20000) == 0)
+  if (pool_elts (utm->sessions) && (pool_elts (utm->sessions) % 20000) == 0)
     {
       f64 now = clib_time_now (&utm->clib_time);
       fformat (stdout, "%d active sessions in %.2f seconds, %.2f/sec...\n",
-               pool_elts(utm->sessions), now - start_time,
-               (f64)pool_elts(utm->sessions) / (now - start_time));
+	       pool_elts (utm->sessions), now - start_time,
+	       (f64) pool_elts (utm->sessions) / (now - start_time));
     }
 
-  /* Send accept reply to vpp */
+  /*
+   * Send accept reply to vpp
+   */
   rmp = vl_msg_api_alloc (sizeof (*rmp));
   memset (rmp, 0, sizeof (*rmp));
   rmp->_vl_msg_id = ntohs (VL_API_ACCEPT_SESSION_REPLY);
   rmp->session_type = mp->session_type;
   rmp->session_index = mp->session_index;
   rmp->session_thread_index = mp->session_thread_index;
-  vl_msg_api_send_shmem (utm->vl_input_queue, (u8 *)&rmp);
+  vl_msg_api_send_shmem (utm->vl_input_queue, (u8 *) & rmp);
 }
 
 void
-uri_tcp_server_test (uri_tcp_test_main_t * utm)
+server_handle_fifo_event_rx (uri_tcp_test_main_t * utm,
+			     session_fifo_event_t * e)
 {
+  svm_fifo_t *rx_fifo, *tx_fifo;
+  int n_read;
+
+  session_fifo_event_t evt;
+  unix_shared_memory_queue_t *q;
+  int rv, bytes;
+
+  rx_fifo = e->fifo;
+  tx_fifo = utm->sessions[rx_fifo->client_session_index].server_tx_fifo;
+
+  bytes = e->enqueue_length;
+  do
+    {
+      n_read = svm_fifo_dequeue_nowait (rx_fifo, 0, vec_len (utm->rx_buf),
+					utm->rx_buf);
+
+      /* Reflect if a non-drop session */
+      if (!utm->drop_packets && n_read > 0)
+	{
+	  do
+	    {
+	      rv = svm_fifo_enqueue_nowait (tx_fifo, 0, n_read, utm->rx_buf);
+	    }
+	  while (rv == -2);
+
+	  /* Fabricate TX event, send to vpp */
+	  evt.fifo = tx_fifo;
+	  evt.event_type = FIFO_EVENT_SERVER_TX;
+	  /* $$$$ for event logging */
+	  evt.enqueue_length = n_read;
+	  evt.event_id = e->event_id;
+	  q = utm->vpp_event_queue;
+	  unix_shared_memory_queue_add (q, (u8 *) & evt,
+					0 /* do wait for mutex */ );
+	}
+
+      if (n_read > 0)
+	bytes -= n_read;
+    }
+  while (n_read < 0 || bytes > 0);
+}
+
+void
+server_handle_event_queue (uri_tcp_test_main_t * utm)
+{
+  session_fifo_event_t _e, *e = &_e;;
 
+  while (1)
+    {
+      unix_shared_memory_queue_sub (utm->our_event_queue, (u8 *) e,
+				    0 /* nowait */ );
+      switch (e->event_type)
+	{
+	case FIFO_EVENT_SERVER_RX:
+	  server_handle_fifo_event_rx (utm, e);
+	  break;
+
+	case FIFO_EVENT_SERVER_EXIT:
+	  return;
+
+	default:
+	  clib_warning ("unknown event type %d", e->event_type);
+	  break;
+	}
+      if (PREDICT_FALSE (utm->time_to_stop == 1))
+	break;
+      if (PREDICT_FALSE (utm->time_to_print_stats == 1))
+	{
+	  utm->time_to_print_stats = 0;
+	  fformat (stdout, "%d connections\n", pool_elts (utm->sessions));
+	}
+    }
+}
+
+void
+server_bind (uri_tcp_test_main_t * utm)
+{
+  vl_api_bind_uri_t *bmp;
+  u32 fifo_size = 3 << 20;
+  bmp = vl_msg_api_alloc (sizeof (*bmp));
+  memset (bmp, 0, sizeof (*bmp));
+
+  bmp->_vl_msg_id = ntohs (VL_API_BIND_URI);
+  bmp->client_index = utm->my_client_index;
+  bmp->context = ntohl (0xfeedface);
+  bmp->initial_segment_size = 256 << 20;	/* size of initial segment */
+  bmp->options[SESSION_OPTIONS_FLAGS] =
+    SESSION_OPTIONS_FLAGS_USE_FIFO | SESSION_OPTIONS_FLAGS_ADD_SEGMENT;
+  bmp->options[SESSION_OPTIONS_RX_FIFO_SIZE] = fifo_size;
+  bmp->options[SESSION_OPTIONS_TX_FIFO_SIZE] = fifo_size;
+  bmp->options[SESSION_OPTIONS_ADD_SEGMENT_SIZE] = 128 << 20;
+  memcpy (bmp->uri, utm->uri, vec_len (utm->uri));
+  vl_msg_api_send_shmem (utm->vl_input_queue, (u8 *) & bmp);
+}
+
+void
+server_unbind (uri_tcp_test_main_t * utm)
+{
+  vl_api_unbind_uri_t *ump;
+
+  ump = vl_msg_api_alloc (sizeof (*ump));
+  memset (ump, 0, sizeof (*ump));
+
+  ump->_vl_msg_id = ntohs (VL_API_UNBIND_URI);
+  ump->client_index = utm->my_client_index;
+  memcpy (ump->uri, utm->uri, vec_len (utm->uri));
+  vl_msg_api_send_shmem (utm->vl_input_queue, (u8 *) & ump);
+}
+
+void
+server_test (uri_tcp_test_main_t * utm)
+{
   /* Bind to uri */
-  uri_tcp_bind (utm);
+  server_bind (utm);
 
   if (wait_for_state_change (utm, STATE_READY))
     {
@@ -780,10 +833,10 @@ uri_tcp_server_test (uri_tcp_test_main_t * utm)
     }
 
   /* Enter handle event loop */
-  handle_event_queue (utm);
+  server_handle_event_queue (utm);
 
   /* Cleanup */
-  uri_tcp_unbind (utm);
+  server_unbind (utm);
 
   if (wait_for_state_change (utm, STATE_START))
     {
@@ -824,12 +877,12 @@ main (int argc, char **argv)
   unformat_input_t _argv, *a = &_argv;
   u8 *chroot_prefix;
   u8 *heap;
-  u8 * bind_name = (u8 *) "tcp://0.0.0.0/1234";
+  u8 *bind_name = (u8 *) "tcp://0.0.0.0/1234";
   u32 tmp;
   mheap_t *h;
-  session_t * session;
+  session_t *session;
   int i;
-  int i_am_master = 1, drop_packets = 0;
+  int i_am_master = 1, drop_packets = 0, test_return_packets = 0;
 
   clib_mem_init (0, 256 << 20);
 
@@ -841,53 +894,54 @@ main (int argc, char **argv)
 
   vec_validate (utm->rx_buf, 65536);
 
-  utm->session_index_by_vpp_handles =
-    hash_create (0, sizeof(uword));
+  utm->session_index_by_vpp_handles = hash_create (0, sizeof (uword));
 
-  utm->my_pid = getpid();
-  utm->configured_segment_size = 1<<20;
+  utm->my_pid = getpid ();
+  utm->configured_segment_size = 1 << 20;
 
   clib_time_init (&utm->clib_time);
   init_error_string_table (utm);
-  svm_fifo_segment_init(0x200000000ULL, 20);
+  svm_fifo_segment_init (0x200000000ULL, 20);
   unformat_init_command_line (a, argv);
 
   while (unformat_check_input (a) != UNFORMAT_END_OF_INPUT)
     {
       if (unformat (a, "chroot prefix %s", &chroot_prefix))
-        {
-          vl_set_memory_root_path ((char *) chroot_prefix);
-        }
+	{
+	  vl_set_memory_root_path ((char *) chroot_prefix);
+	}
       else if (unformat (a, "uri %s", &bind_name))
-        ;
+	;
       else if (unformat (a, "segment-size %dM", &tmp))
-        utm->configured_segment_size = tmp<<20;
+	utm->configured_segment_size = tmp << 20;
       else if (unformat (a, "segment-size %dG", &tmp))
-        utm->configured_segment_size = tmp<<30;
+	utm->configured_segment_size = tmp << 30;
       else if (unformat (a, "master"))
-        i_am_master = 1;
+	i_am_master = 1;
       else if (unformat (a, "slave"))
-        i_am_master = 0;
+	i_am_master = 0;
       else if (unformat (a, "drop"))
-        drop_packets = 1;
+	drop_packets = 1;
+      else if (unformat (a, "test"))
+	test_return_packets = 1;
       else
-        {
-          fformat (stderr, "%s: usage [master|slave]\n");
-          exit (1);
-        }
+	{
+	  fformat (stderr, "%s: usage [master|slave]\n");
+	  exit (1);
+	}
     }
 
   utm->uri = format (0, "%s%c", bind_name, 0);
   utm->i_am_master = i_am_master;
   utm->segment_main = &svm_fifo_segment_main;
   utm->drop_packets = drop_packets;
-
+  utm->test_return_packets = test_return_packets;
   utm->connect_uri = format (0, "tcp://6.0.1.2/1234%c", 0);
 
-  setup_signal_handlers();
+  setup_signal_handlers ();
   uri_api_hookup (utm);
 
-  if (connect_to_vpp (i_am_master? "uri_tcp_server":"uri_tcp_client") < 0)
+  if (connect_to_vpp (i_am_master ? "uri_tcp_server" : "uri_tcp_client") < 0)
     {
       svm_region_exit ();
       fformat (stderr, "Couldn't connect to vpe, exiting...\n");
@@ -896,7 +950,7 @@ main (int argc, char **argv)
 
   if (i_am_master == 0)
     {
-      uri_tcp_client_test (utm);
+      client_test (utm);
       exit (0);
     }
 
@@ -909,8 +963,16 @@ main (int argc, char **argv)
   for (i = 0; i < 200000; i++)
     pool_put_index (utm->sessions, i);
 
-  uri_tcp_server_test (utm);
+  server_test (utm);
 
   vl_client_disconnect_from_vlib ();
   exit (0);
 }
+
+/*
+ * fd.io coding-style-patch-verification: ON
+ *
+ * Local Variables:
+ * eval: (c-set-style "gnu")
+ * End:
+ */
diff --git a/src/uri/uri_udp_test.c b/src/uri/uri_udp_test.c
index 6f5284c9..54625d64 100644
--- a/src/uri/uri_udp_test.c
+++ b/src/uri/uri_udp_test.c
@@ -26,25 +26,25 @@
 #include <vlib/unix/unix.h>
 #include <vlibapi/api.h>
 #include <vlibmemory/api.h>
-#include <vpp-api/vpe_msg_enum.h>
-#include <svm_fifo_segment.h>
-
-#include <vnet/uri/uri.h>
+#include <vpp/api/vpe_msg_enum.h>
+#include <svm/svm_fifo_segment.h>
+#include <pthread.h>
+#include <vnet/session/application_interface.h>
 
 #define vl_typedefs		/* define message structures */
-#include <vpp-api/vpe_all_api_h.h>
+#include <vpp/api/vpe_all_api_h.h>
 #undef vl_typedefs
 
 /* declare message handlers for each api */
 
 #define vl_endianfun		/* define message structures */
-#include <vpp-api/vpe_all_api_h.h>
+#include <vpp/api/vpe_all_api_h.h>
 #undef vl_endianfun
 
 /* instantiate all the print functions we know about */
 #define vl_print(handle, ...)
 #define vl_printfun
-#include <vpp-api/vpe_all_api_h.h>
+#include <vpp/api/vpe_all_api_h.h>
 #undef vl_printfun
 
 /* Satisfy external references when not linking with -lvlib */
@@ -87,12 +87,28 @@ typedef struct
   /* intermediate rx buffer */
   u8 *rx_buf;
 
+  /* URI for connect */
+  u8 *connect_uri;
+
+  int i_am_master;
+
   /* Our event queue */
   unix_shared_memory_queue_t *our_event_queue;
 
   /* $$$ single thread only for the moment */
   unix_shared_memory_queue_t *vpp_event_queue;
 
+  /* $$$$ hack: cut-through session index */
+  volatile u32 cut_through_session_index;
+
+  /* unique segment name counter */
+  u32 unique_segment_index;
+
+  pid_t my_pid;
+
+  /* pthread handle */
+  pthread_t cut_through_thread_handle;
+
   /* For deadman timers */
   clib_time_t clib_time;
 
@@ -102,14 +118,20 @@ typedef struct
   volatile int time_to_stop;
   volatile int time_to_print_stats;
 
+  u32 configured_segment_size;
+
   /* VNET_API_ERROR_FOO -> "Foo" hash table */
   uword *error_string_by_error_number;
+
+  /* convenience */
+  svm_fifo_segment_main_t *segment_main;
+
 } uri_udp_test_main_t;
 
 #if CLIB_DEBUG > 0
-#define NITER 1000
+#define NITER 10000
 #else
-#define NITER 1000000
+#define NITER 4000000
 #endif
 
 uri_udp_test_main_t uri_udp_test_main;
@@ -159,7 +181,13 @@ format_api_error (u8 * s, va_list * args)
 int
 wait_for_state_change (uri_udp_test_main_t * utm, connection_state_t state)
 {
-  f64 timeout = clib_time_now (&utm->clib_time) + 5.0;
+#if CLIB_DEBUG > 0
+#define TIMEOUT 600.0
+#else
+#define TIMEOUT 600.0
+#endif
+
+  f64 timeout = clib_time_now (&utm->clib_time) + TIMEOUT;
 
   while (clib_time_now (&utm->clib_time) < timeout)
     {
@@ -169,6 +197,183 @@ wait_for_state_change (uri_udp_test_main_t * utm, connection_state_t state)
   return -1;
 }
 
+u64 server_bytes_received, server_bytes_sent;
+
+static void *
+cut_through_thread_fn (void *arg)
+{
+  session_t *s;
+  svm_fifo_t *rx_fifo;
+  svm_fifo_t *tx_fifo;
+  u8 *my_copy_buffer = 0;
+  uri_udp_test_main_t *utm = &uri_udp_test_main;
+  i32 actual_transfer;
+  int rv;
+  u32 buffer_offset;
+
+  while (utm->cut_through_session_index == ~0)
+    ;
+
+  s = pool_elt_at_index (utm->sessions, utm->cut_through_session_index);
+
+  rx_fifo = s->server_rx_fifo;
+  tx_fifo = s->server_tx_fifo;
+
+  vec_validate (my_copy_buffer, 64 * 1024 - 1);
+
+  while (true)
+    {
+      /* We read from the tx fifo and write to the rx fifo */
+      do
+	{
+	  actual_transfer = svm_fifo_dequeue_nowait (tx_fifo, 0,
+						     vec_len (my_copy_buffer),
+						     my_copy_buffer);
+	}
+      while (actual_transfer <= 0);
+
+      server_bytes_received += actual_transfer;
+
+      buffer_offset = 0;
+      while (actual_transfer > 0)
+	{
+	  rv = svm_fifo_enqueue_nowait (rx_fifo, 0, actual_transfer,
+					my_copy_buffer + buffer_offset);
+	  if (rv > 0)
+	    {
+	      actual_transfer -= rv;
+	      buffer_offset += rv;
+	      server_bytes_sent += rv;
+	    }
+
+	}
+      if (PREDICT_FALSE (utm->time_to_stop))
+	break;
+    }
+
+  pthread_exit (0);
+}
+
+static void
+uri_udp_slave_test (uri_udp_test_main_t * utm)
+{
+  vl_api_connect_uri_t *cmp;
+  int i;
+  u8 *test_data = 0;
+  u64 bytes_received = 0, bytes_sent = 0;
+  i32 bytes_to_read;
+  int rv;
+  int mypid = getpid ();
+  f64 before, after, delta, bytes_per_second;
+  session_t *session;
+  svm_fifo_t *rx_fifo, *tx_fifo;
+  int buffer_offset, bytes_to_send = 0;
+
+  vec_validate (test_data, 64 * 1024 - 1);
+  for (i = 0; i < vec_len (test_data); i++)
+    test_data[i] = i & 0xff;
+
+  cmp = vl_msg_api_alloc (sizeof (*cmp));
+  memset (cmp, 0, sizeof (*cmp));
+
+  cmp->_vl_msg_id = ntohs (VL_API_CONNECT_URI);
+  cmp->client_index = utm->my_client_index;
+  cmp->context = ntohl (0xfeedface);
+  memcpy (cmp->uri, utm->connect_uri, vec_len (utm->connect_uri));
+  vl_msg_api_send_shmem (utm->vl_input_queue, (u8 *) & cmp);
+
+  if (wait_for_state_change (utm, STATE_READY))
+    {
+      clib_warning ("timeout waiting for STATE_READY");
+      return;
+    }
+
+  session = pool_elt_at_index (utm->sessions, utm->cut_through_session_index);
+  rx_fifo = session->server_rx_fifo;
+  tx_fifo = session->server_tx_fifo;
+
+  before = clib_time_now (&utm->clib_time);
+
+  vec_validate (utm->rx_buf, vec_len (test_data) - 1);
+
+  for (i = 0; i < NITER; i++)
+    {
+      bytes_to_send = vec_len (test_data);
+      buffer_offset = 0;
+      while (bytes_to_send > 0)
+	{
+	  rv = svm_fifo_enqueue_nowait (tx_fifo, mypid,
+					bytes_to_send,
+					test_data + buffer_offset);
+
+	  if (rv > 0)
+	    {
+	      bytes_to_send -= rv;
+	      buffer_offset += rv;
+	      bytes_sent += rv;
+	    }
+	}
+
+      bytes_to_read = svm_fifo_max_dequeue (rx_fifo);
+
+      bytes_to_read = vec_len (utm->rx_buf) > bytes_to_read ?
+	bytes_to_read : vec_len (utm->rx_buf);
+
+      buffer_offset = 0;
+      while (bytes_to_read > 0)
+	{
+	  rv = svm_fifo_dequeue_nowait (rx_fifo, mypid,
+					bytes_to_read,
+					utm->rx_buf + buffer_offset);
+	  if (rv > 0)
+	    {
+	      bytes_to_read -= rv;
+	      buffer_offset += rv;
+	      bytes_received += rv;
+	    }
+	}
+    }
+  while (bytes_received < bytes_sent)
+    {
+      rv = svm_fifo_dequeue_nowait (rx_fifo, mypid,
+				    vec_len (utm->rx_buf), utm->rx_buf);
+      if (rv > 0)
+	{
+#if CLIB_DEBUG > 0
+	  int j;
+	  for (j = 0; j < rv; j++)
+	    {
+	      if (utm->rx_buf[j] != ((bytes_received + j) & 0xff))
+		{
+		  clib_warning ("error at byte %lld, 0x%x not 0x%x",
+				bytes_received + j,
+				utm->rx_buf[j],
+				((bytes_received + j) & 0xff));
+		}
+	    }
+#endif
+	  bytes_received += (u64) rv;
+	}
+    }
+
+  after = clib_time_now (&utm->clib_time);
+  delta = after - before;
+  bytes_per_second = 0.0;
+
+  if (delta > 0.0)
+    bytes_per_second = (f64) bytes_received / delta;
+
+  fformat (stdout,
+	   "Done: %lld recv bytes in %.2f seconds, %.2f bytes/sec...\n\n",
+	   bytes_received, delta, bytes_per_second);
+  fformat (stdout,
+	   "Done: %lld sent bytes in %.2f seconds, %.2f bytes/sec...\n\n",
+	   bytes_sent, delta, bytes_per_second);
+  fformat (stdout,
+	   "client -> server -> client round trip: %.2f Gbit/sec \n\n",
+	   (bytes_per_second * 8.0) / 1e9);
+}
+
 static void
 vl_api_bind_uri_reply_t_handler (vl_api_bind_uri_reply_t * mp)
 {
@@ -183,12 +388,16 @@ vl_api_bind_uri_reply_t_handler (vl_api_bind_uri_reply_t * mp)
     }
 
   a->segment_name = (char *) mp->segment_name;
+  a->segment_size = mp->segment_size;
+
+  ASSERT (mp->server_event_queue_address);
 
   /* Attach to the segment vpp created */
   rv = svm_fifo_segment_attach (a);
   if (rv)
     {
-      clib_warning ("sm_fifo_segment_create ('%s') failed", mp->segment_name);
+      clib_warning ("svm_fifo_segment_attach ('%s') failed",
+		    mp->segment_name);
       return;
     }
 
@@ -198,6 +407,101 @@ vl_api_bind_uri_reply_t_handler (vl_api_bind_uri_reply_t * mp)
   utm->state = STATE_READY;
 }
 
+static void
+vl_api_map_another_segment_t_handler (vl_api_map_another_segment_t * mp)
+{
+  svm_fifo_segment_create_args_t _a, *a = &_a;
+  int rv;
+
+  a->segment_name = (char *) mp->segment_name;
+  a->segment_size = mp->segment_size;
+  /* Attach to the segment vpp created */
+  rv = svm_fifo_segment_attach (a);
+  if (rv)
+    {
+      clib_warning ("svm_fifo_segment_attach ('%s') failed",
+		    mp->segment_name);
+      return;
+    }
+  clib_warning ("Mapped new segment '%s' size %d", mp->segment_name,
+		mp->segment_size);
+}
+
+static void
+vl_api_connect_uri_t_handler (vl_api_connect_uri_t * mp)
+{
+  u32 segment_index;
+  uri_udp_test_main_t *utm = &uri_udp_test_main;
+  svm_fifo_segment_main_t *sm = &svm_fifo_segment_main;
+  svm_fifo_segment_create_args_t _a, *a = &_a;
+  svm_fifo_segment_private_t *seg;
+  unix_shared_memory_queue_t *client_q;
+  vl_api_connect_uri_reply_t *rmp;
+  session_t *session;
+  int rv = 0;
+
+  /* Create the segment */
+  a->segment_name = (char *) format (0, "%d:segment%d%c", utm->my_pid,
+				     utm->unique_segment_index++, 0);
+  a->segment_size = utm->configured_segment_size;
+
+  rv = svm_fifo_segment_create (a);
+  if (rv)
+    {
+      clib_warning ("sm_fifo_segment_create ('%s') failed", a->segment_name);
+      rv = VNET_API_ERROR_URI_FIFO_CREATE_FAILED;
+      goto send_reply;
+    }
+
+  vec_add2 (utm->seg, seg, 1);
+
+  segment_index = vec_len (sm->segments) - 1;
+
+  memcpy (seg, sm->segments + segment_index, sizeof (utm->seg[0]));
+
+  pool_get (utm->sessions, session);
+
+  /*
+   * By construction the master's idea of the rx fifo ends up in
+   * fsh->fifos[0], and the master's idea of the tx fifo ends up in
+   * fsh->fifos[1].
+   */
+  session->server_rx_fifo = svm_fifo_segment_alloc_fifo (utm->seg,
+							 128 * 1024);
+  ASSERT (session->server_rx_fifo);
+
+  session->server_tx_fifo = svm_fifo_segment_alloc_fifo (utm->seg,
+							 128 * 1024);
+  ASSERT (session->server_tx_fifo);
+
+  session->server_rx_fifo->server_session_index = session - utm->sessions;
+  session->server_tx_fifo->server_session_index = session - utm->sessions;
+  utm->cut_through_session_index = session - utm->sessions;
+
+  rv = pthread_create (&utm->cut_through_thread_handle,
+		       NULL /*attr */ , cut_through_thread_fn, 0);
+  if (rv)
+    {
+      clib_warning ("pthread_create returned %d", rv);
+      rv = VNET_API_ERROR_SYSCALL_ERROR_1;
+    }
+
+send_reply:
+  rmp = vl_msg_api_alloc (sizeof (*rmp));
+  memset (rmp, 0, sizeof (*rmp));
+
+  rmp->_vl_msg_id = ntohs (VL_API_CONNECT_URI_REPLY);
+  rmp->context = mp->context;
+  rmp->retval = ntohl (rv);
+  rmp->segment_name_length = vec_len (a->segment_name);
+  memcpy (rmp->segment_name, a->segment_name, vec_len (a->segment_name));
+
+  vec_free (a->segment_name);
+
+  client_q = (unix_shared_memory_queue_t *) mp->client_queue_address;
+  vl_msg_api_send_shmem (client_q, (u8 *) & rmp);
+}
+
 static void
 vl_api_unbind_uri_reply_t_handler (vl_api_unbind_uri_reply_t * mp)
 {
@@ -293,18 +597,79 @@ vl_api_disconnect_session_t_handler (vl_api_disconnect_session_t * mp)
   vl_msg_api_send_shmem (utm->vl_input_queue, (u8 *) & rmp);
 }
 
+static void
+vl_api_connect_uri_reply_t_handler (vl_api_connect_uri_reply_t * mp)
+{
+  svm_fifo_segment_main_t *sm = &svm_fifo_segment_main;
+  uri_udp_test_main_t *utm = &uri_udp_test_main;
+  svm_fifo_segment_create_args_t _a, *a = &_a;
+  ssvm_shared_header_t *sh;
+  svm_fifo_segment_private_t *seg;
+  svm_fifo_segment_header_t *fsh;
+  session_t *session;
+  u32 segment_index;
+  int rv;
+
+  ASSERT (utm->i_am_master == 0);
+
+  if (mp->segment_name_length == 0)
+    {
+      clib_warning ("segment_name_length zero");
+      return;
+    }
+
+  memset (a, 0, sizeof (*a));
+
+  a->segment_name = (char *) mp->segment_name;
+
+  sleep (1);
+
+  rv = svm_fifo_segment_attach (a);
+  if (rv)
+    {
+      clib_warning ("sm_fifo_segment_create ('%v') failed", mp->segment_name);
+      return;
+    }
+
+  segment_index = vec_len (sm->segments) - 1;
+
+  vec_add2 (utm->seg, seg, 1);
+
+  memcpy (seg, sm->segments + segment_index, sizeof (*seg));
+  sh = seg->ssvm.sh;
+  fsh = (svm_fifo_segment_header_t *) sh->opaque[0];
+
+  while (vec_len (fsh->fifos) < 2)
+    sleep (1);
+
+  pool_get (utm->sessions, session);
+  utm->cut_through_session_index = session - utm->sessions;
+
+  session->server_rx_fifo = (svm_fifo_t *) fsh->fifos[0];
+  ASSERT (session->server_rx_fifo);
+  session->server_tx_fifo = (svm_fifo_t *) fsh->fifos[1];
+  ASSERT (session->server_tx_fifo);
+
+  /* security: could unlink /dev/shm/<mp->segment_name> here, maybe */
+
+  utm->state = STATE_READY;
+}
+
 #define foreach_uri_msg                         \
 _(BIND_URI_REPLY, bind_uri_reply)               \
+_(CONNECT_URI, connect_uri)                     \
+_(CONNECT_URI_REPLY, connect_uri_reply)         \
 _(UNBIND_URI_REPLY, unbind_uri_reply)           \
 _(ACCEPT_SESSION, accept_session)		\
-_(DISCONNECT_SESSION, disconnect_session)
+_(DISCONNECT_SESSION, disconnect_session)	\
+_(MAP_ANOTHER_SEGMENT, map_another_segment)
 
 void
 uri_api_hookup (uri_udp_test_main_t * utm)
 {
 #define _(N,n)                                                  \
     vl_msg_api_set_handlers(VL_API_##N, #n,                     \
-                           vl_api_##n##_t_handler,	        \
+                           vl_api_##n##_t_handler,              \
                            vl_noop_handler,                     \
                            vl_api_##n##_t_endian,               \
                            vl_api_##n##_t_print,                \
@@ -349,7 +714,7 @@ init_error_string_table (uri_udp_test_main_t * utm)
 }
 
 void
-handle_fifo_event_server_rx (uri_udp_test_main_t * utm,
+server_handle_fifo_event_rx (uri_udp_test_main_t * utm,
 			     session_fifo_event_t * e)
 {
   svm_fifo_t *rx_fifo, *tx_fifo;
@@ -385,7 +750,7 @@ handle_fifo_event_server_rx (uri_udp_test_main_t * utm,
 }
 
 void
-handle_event_queue (uri_udp_test_main_t * utm)
+server_handle_event_queue (uri_udp_test_main_t * utm)
 {
   session_fifo_event_t _e, *e = &_e;;
 
@@ -396,7 +761,7 @@ handle_event_queue (uri_udp_test_main_t * utm)
       switch (e->event_type)
 	{
 	case FIFO_EVENT_SERVER_RX:
-	  handle_fifo_event_server_rx (utm, e);
+	  server_handle_fifo_event_rx (utm, e);
 	  break;
 
 	case FIFO_EVENT_SERVER_EXIT:
@@ -428,7 +793,12 @@ uri_udp_test (uri_udp_test_main_t * utm)
   bmp->_vl_msg_id = ntohs (VL_API_BIND_URI);
   bmp->client_index = utm->my_client_index;
   bmp->context = ntohl (0xfeedface);
-  bmp->segment_size = 2 << 30;
+  bmp->initial_segment_size = 256 << 20;	/* size of initial segment */
+  bmp->options[SESSION_OPTIONS_FLAGS] =
+    SESSION_OPTIONS_FLAGS_USE_FIFO | SESSION_OPTIONS_FLAGS_ADD_SEGMENT;
+  bmp->options[SESSION_OPTIONS_RX_FIFO_SIZE] = 16 << 10;
+  bmp->options[SESSION_OPTIONS_TX_FIFO_SIZE] = 16 << 10;
+  bmp->options[SESSION_OPTIONS_ADD_SEGMENT_SIZE] = 128 << 20;
   memcpy (bmp->uri, utm->uri, vec_len (utm->uri));
   vl_msg_api_send_shmem (utm->vl_input_queue, (u8 *) & bmp);
 
@@ -438,7 +808,7 @@ uri_udp_test (uri_udp_test_main_t * utm)
       return;
     }
 
-  handle_event_queue (utm);
+  server_handle_event_queue (utm);
 
   ump = vl_msg_api_alloc (sizeof (*ump));
   memset (ump, 0, sizeof (*ump));
@@ -464,10 +834,12 @@ main (int argc, char **argv)
   unformat_input_t _argv, *a = &_argv;
   u8 *chroot_prefix;
   u8 *heap;
-  u8 *bind_name = (u8 *) "udp4:1234";
+  u8 *bind_name = (u8 *) "udp://0.0.0.0/1234";
+  u32 tmp;
   mheap_t *h;
   session_t *session;
   int i;
+  int i_am_master = 1;
 
   clib_mem_init (0, 256 << 20);
 
@@ -481,6 +853,9 @@ main (int argc, char **argv)
 
   utm->session_index_by_vpp_handles = hash_create (0, sizeof (uword));
 
+  utm->my_pid = getpid ();
+  utm->configured_segment_size = 1 << 20;
+
   clib_time_init (&utm->clib_time);
   init_error_string_table (utm);
   svm_fifo_segment_init (0x200000000ULL, 20);
@@ -494,6 +869,14 @@ main (int argc, char **argv)
 	}
       else if (unformat (a, "uri %s", &bind_name))
 	;
+      else if (unformat (a, "segment-size %dM", &tmp))
+	utm->configured_segment_size = tmp << 20;
+      else if (unformat (a, "segment-size %dG", &tmp))
+	utm->configured_segment_size = tmp << 30;
+      else if (unformat (a, "master"))
+	i_am_master = 1;
+      else if (unformat (a, "slave"))
+	i_am_master = 0;
       else
 	{
 	  fformat (stderr, "%s: usage [master|slave]\n");
@@ -501,19 +884,30 @@ main (int argc, char **argv)
 	}
     }
 
+  utm->cut_through_session_index = ~0;
   utm->uri = format (0, "%s%c", bind_name, 0);
+  utm->i_am_master = i_am_master;
+  utm->segment_main = &svm_fifo_segment_main;
+
+  utm->connect_uri = format (0, "udp://10.0.0.1/1234%c", 0);
 
   setup_signal_handlers ();
 
   uri_api_hookup (utm);
 
-  if (connect_to_vpp ("uri_udp_test") < 0)
+  if (connect_to_vpp (i_am_master ? "uri_udp_master" : "uri_udp_slave") < 0)
     {
       svm_region_exit ();
       fformat (stderr, "Couldn't connect to vpe, exiting...\n");
       exit (1);
     }
 
+  if (i_am_master == 0)
+    {
+      uri_udp_slave_test (utm);
+      exit (0);
+    }
+
   /* $$$$ hack preallocation */
   for (i = 0; i < 200000; i++)
     {
@@ -531,7 +925,7 @@ main (int argc, char **argv)
 
 #undef vl_api_version
 #define vl_api_version(n,v) static u32 vpe_api_version = v;
-#include <vpp-api/vpe.api.h>
+#include <vpp/api/vpe.api.h>
 #undef vl_api_version
 
 void
@@ -544,6 +938,12 @@ vl_client_add_api_signatures (vl_api_memclnt_create_t * mp)
   mp->api_versions[0] = clib_host_to_net_u32 (vpe_api_version);
 }
 
+u32
+vl (void *p)
+{
+  return vec_len (p);
+}
+
 /*
  * fd.io coding-style-patch-verification: ON
  *
diff --git a/src/uri/uri_udp_test2.c b/src/uri/uri_udp_test2.c
deleted file mode 100644
index ddfffaa6..00000000
--- a/src/uri/uri_udp_test2.c
+++ /dev/null
@@ -1,954 +0,0 @@
-/*
- * Copyright (c) 2016 Cisco and/or its affiliates.
- * Licensed under the Apache License, Version 2.0 (the "License");
- * you may not use this file except in compliance with the License.
- * You may obtain a copy of the License at:
- *
- *     http://www.apache.org/licenses/LICENSE-2.0
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS,
- * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
- * See the License for the specific language governing permissions and
- * limitations under the License.
- */
-
-#include <stdio.h>
-#include <setjmp.h>
-#include <signal.h>
-#include <vppinfra/clib.h>
-#include <vppinfra/format.h>
-#include <vppinfra/error.h>
-#include <vppinfra/time.h>
-#include <vppinfra/macros.h>
-#include <vnet/vnet.h>
-#include <vlib/vlib.h>
-#include <vlib/unix/unix.h>
-#include <vlibapi/api.h>
-#include <vlibmemory/api.h>
-#include <vpp/api/vpe_msg_enum.h>
-#include <svm/svm_fifo_segment.h>
-#include <pthread.h>
-
-#include "../vnet/session/application_interface.h"
-
-#define vl_typedefs		/* define message structures */
-#include <vpp/api/vpe_all_api_h.h>
-#undef vl_typedefs
-
-/* declare message handlers for each api */
-
-#define vl_endianfun		/* define message structures */
-#include <vpp/api/vpe_all_api_h.h>
-#undef vl_endianfun
-
-/* instantiate all the print functions we know about */
-#define vl_print(handle, ...)
-#define vl_printfun
-#include <vpp/api/vpe_all_api_h.h>
-#undef vl_printfun
-
-/* Satisfy external references when not linking with -lvlib */
-vlib_main_t vlib_global_main;
-vlib_main_t **vlib_mains;
-
-typedef enum
-{
-  STATE_START,
-  STATE_READY,
-  STATE_DISCONNECTING,
-} connection_state_t;
-
-typedef struct
-{
-  svm_fifo_t *server_rx_fifo;
-  svm_fifo_t *server_tx_fifo;
-} session_t;
-
-typedef struct
-{
-  /* vpe input queue */
-  unix_shared_memory_queue_t *vl_input_queue;
-
-  /* API client handle */
-  u32 my_client_index;
-
-  /* The URI we're playing with */
-  u8 *uri;
-
-  /* Session pool */
-  session_t *sessions;
-
-  /* Hash table for disconnect processing */
-  uword *session_index_by_vpp_handles;
-
-  /* fifo segment */
-  svm_fifo_segment_private_t *seg;
-
-  /* intermediate rx buffer */
-  u8 *rx_buf;
-
-  /* URI for connect */
-  u8 *connect_uri;
-
-  int i_am_master;
-
-  /* Our event queue */
-  unix_shared_memory_queue_t *our_event_queue;
-
-  /* $$$ single thread only for the moment */
-  unix_shared_memory_queue_t *vpp_event_queue;
-
-  /* $$$$ hack: cut-through session index */
-  volatile u32 cut_through_session_index;
-
-  /* unique segment name counter */
-  u32 unique_segment_index;
-
-  pid_t my_pid;
-
-  /* pthread handle */
-  pthread_t cut_through_thread_handle;
-
-  /* For deadman timers */
-  clib_time_t clib_time;
-
-  /* State of the connection, shared between msg RX thread and main thread */
-  volatile connection_state_t state;
-
-  volatile int time_to_stop;
-  volatile int time_to_print_stats;
-
-  u32 configured_segment_size;
-
-  /* VNET_API_ERROR_FOO -> "Foo" hash table */
-  uword *error_string_by_error_number;
-
-  /* convenience */
-  svm_fifo_segment_main_t *segment_main;
-
-} uri_udp_test_main_t;
-
-#if CLIB_DEBUG > 0
-#define NITER 10000
-#else
-#define NITER 4000000
-#endif
-
-uri_udp_test_main_t uri_udp_test_main;
-
-static void
-stop_signal (int signum)
-{
-  uri_udp_test_main_t *um = &uri_udp_test_main;
-
-  um->time_to_stop = 1;
-}
-
-static void
-stats_signal (int signum)
-{
-  uri_udp_test_main_t *um = &uri_udp_test_main;
-
-  um->time_to_print_stats = 1;
-}
-
-static clib_error_t *
-setup_signal_handlers (void)
-{
-  signal (SIGINT, stats_signal);
-  signal (SIGQUIT, stop_signal);
-  signal (SIGTERM, stop_signal);
-
-  return 0;
-}
-
-u8 *
-format_api_error (u8 * s, va_list * args)
-{
-  uri_udp_test_main_t *utm = va_arg (*args, uri_udp_test_main_t *);
-  i32 error = va_arg (*args, u32);
-  uword *p;
-
-  p = hash_get (utm->error_string_by_error_number, -error);
-
-  if (p)
-    s = format (s, "%s", p[0]);
-  else
-    s = format (s, "%d", error);
-  return s;
-}
-
-int
-wait_for_state_change (uri_udp_test_main_t * utm, connection_state_t state)
-{
-#if CLIB_DEBUG > 0
-#define TIMEOUT 600.0
-#else
-#define TIMEOUT 600.0
-#endif
-
-  f64 timeout = clib_time_now (&utm->clib_time) + TIMEOUT;
-
-  while (clib_time_now (&utm->clib_time) < timeout)
-    {
-      if (utm->state == state)
-	return 0;
-    }
-  return -1;
-}
-
-u64 server_bytes_received, server_bytes_sent;
-
-static void *
-cut_through_thread_fn (void *arg)
-{
-  session_t *s;
-  svm_fifo_t *rx_fifo;
-  svm_fifo_t *tx_fifo;
-  u8 *my_copy_buffer = 0;
-  uri_udp_test_main_t *utm = &uri_udp_test_main;
-  i32 actual_transfer;
-  int rv;
-  u32 buffer_offset;
-
-  while (utm->cut_through_session_index == ~0)
-    ;
-
-  s = pool_elt_at_index (utm->sessions, utm->cut_through_session_index);
-
-  rx_fifo = s->server_rx_fifo;
-  tx_fifo = s->server_tx_fifo;
-
-  vec_validate (my_copy_buffer, 64 * 1024 - 1);
-
-  while (true)
-    {
-      /* We read from the tx fifo and write to the rx fifo */
-      do
-	{
-	  actual_transfer = svm_fifo_dequeue_nowait (tx_fifo, 0,
-						     vec_len (my_copy_buffer),
-						     my_copy_buffer);
-	}
-      while (actual_transfer <= 0);
-
-      server_bytes_received += actual_transfer;
-
-      buffer_offset = 0;
-      while (actual_transfer > 0)
-	{
-	  rv = svm_fifo_enqueue_nowait (rx_fifo, 0, actual_transfer,
-					my_copy_buffer + buffer_offset);
-	  if (rv > 0)
-	    {
-	      actual_transfer -= rv;
-	      buffer_offset += rv;
-	      server_bytes_sent += rv;
-	    }
-
-	}
-      if (PREDICT_FALSE (utm->time_to_stop))
-	break;
-    }
-
-  pthread_exit (0);
-}
-
-static void
-uri_udp_slave_test (uri_udp_test_main_t * utm)
-{
-  vl_api_connect_uri_t *cmp;
-  int i;
-  u8 *test_data = 0;
-  u64 bytes_received = 0, bytes_sent = 0;
-  i32 bytes_to_read;
-  int rv;
-  int mypid = getpid ();
-  f64 before, after, delta, bytes_per_second;
-  session_t *session;
-  svm_fifo_t *rx_fifo, *tx_fifo;
-  int buffer_offset, bytes_to_send = 0;
-
-  vec_validate (test_data, 64 * 1024 - 1);
-  for (i = 0; i < vec_len (test_data); i++)
-    test_data[i] = i & 0xff;
-
-  cmp = vl_msg_api_alloc (sizeof (*cmp));
-  memset (cmp, 0, sizeof (*cmp));
-
-  cmp->_vl_msg_id = ntohs (VL_API_CONNECT_URI);
-  cmp->client_index = utm->my_client_index;
-  cmp->context = ntohl (0xfeedface);
-  memcpy (cmp->uri, utm->connect_uri, vec_len (utm->connect_uri));
-  vl_msg_api_send_shmem (utm->vl_input_queue, (u8 *) & cmp);
-
-  if (wait_for_state_change (utm, STATE_READY))
-    {
-      clib_warning ("timeout waiting for STATE_READY");
-      return;
-    }
-
-  session = pool_elt_at_index (utm->sessions, utm->cut_through_session_index);
-  rx_fifo = session->server_rx_fifo;
-  tx_fifo = session->server_tx_fifo;
-
-  before = clib_time_now (&utm->clib_time);
-
-  vec_validate (utm->rx_buf, vec_len (test_data) - 1);
-
-  for (i = 0; i < NITER; i++)
-    {
-      bytes_to_send = vec_len (test_data);
-      buffer_offset = 0;
-      while (bytes_to_send > 0)
-	{
-	  rv = svm_fifo_enqueue_nowait (tx_fifo, mypid,
-					bytes_to_send,
-					test_data + buffer_offset);
-
-	  if (rv > 0)
-	    {
-	      bytes_to_send -= rv;
-	      buffer_offset += rv;
-	      bytes_sent += rv;
-	    }
-	}
-
-      bytes_to_read = svm_fifo_max_dequeue (rx_fifo);
-
-      bytes_to_read = vec_len (utm->rx_buf) > bytes_to_read ?
-	bytes_to_read : vec_len (utm->rx_buf);
-
-      buffer_offset = 0;
-      while (bytes_to_read > 0)
-	{
-	  rv = svm_fifo_dequeue_nowait (rx_fifo, mypid,
-					bytes_to_read,
-					utm->rx_buf + buffer_offset);
-	  if (rv > 0)
-	    {
-	      bytes_to_read -= rv;
-	      buffer_offset += rv;
-	      bytes_received += rv;
-	    }
-	}
-    }
-  while (bytes_received < bytes_sent)
-    {
-      rv = svm_fifo_dequeue_nowait (rx_fifo, mypid,
-				    vec_len (utm->rx_buf), utm->rx_buf);
-      if (rv > 0)
-	{
-#if CLIB_DEBUG > 0
-	  int j;
-	  for (j = 0; j < rv; j++)
-	    {
-	      if (utm->rx_buf[j] != ((bytes_received + j) & 0xff))
-		{
-		  clib_warning ("error at byte %lld, 0x%x not 0x%x",
-				bytes_received + j,
-				utm->rx_buf[j],
-				((bytes_received + j) & 0xff));
-		}
-	    }
-#endif
-	  bytes_received += (u64) rv;
-	}
-    }
-
-  after = clib_time_now (&utm->clib_time);
-  delta = after - before;
-  bytes_per_second = 0.0;
-
-  if (delta > 0.0)
-    bytes_per_second = (f64) bytes_received / delta;
-
-  fformat (stdout,
-	   "Done: %lld recv bytes in %.2f seconds, %.2f bytes/sec...\n\n",
-	   bytes_received, delta, bytes_per_second);
-  fformat (stdout,
-	   "Done: %lld sent bytes in %.2f seconds, %.2f bytes/sec...\n\n",
-	   bytes_sent, delta, bytes_per_second);
-  fformat (stdout,
-	   "client -> server -> client round trip: %.2f Gbit/sec \n\n",
-	   (bytes_per_second * 8.0) / 1e9);
-}
-
-static void
-vl_api_bind_uri_reply_t_handler (vl_api_bind_uri_reply_t * mp)
-{
-  uri_udp_test_main_t *utm = &uri_udp_test_main;
-  svm_fifo_segment_create_args_t _a, *a = &_a;
-  int rv;
-
-  if (mp->segment_name_length == 0)
-    {
-      clib_warning ("segment_name_length zero");
-      return;
-    }
-
-  a->segment_name = (char *) mp->segment_name;
-  a->segment_size = mp->segment_size;
-
-  ASSERT (mp->server_event_queue_address);
-
-  /* Attach to the segment vpp created */
-  rv = svm_fifo_segment_attach (a);
-  if (rv)
-    {
-      clib_warning ("svm_fifo_segment_attach ('%s') failed",
-		    mp->segment_name);
-      return;
-    }
-
-  utm->our_event_queue = (unix_shared_memory_queue_t *)
-    mp->server_event_queue_address;
-
-  utm->state = STATE_READY;
-}
-
-static void
-vl_api_map_another_segment_t_handler (vl_api_map_another_segment_t * mp)
-{
-  svm_fifo_segment_create_args_t _a, *a = &_a;
-  int rv;
-
-  a->segment_name = (char *) mp->segment_name;
-  a->segment_size = mp->segment_size;
-  /* Attach to the segment vpp created */
-  rv = svm_fifo_segment_attach (a);
-  if (rv)
-    {
-      clib_warning ("svm_fifo_segment_attach ('%s') failed",
-		    mp->segment_name);
-      return;
-    }
-  clib_warning ("Mapped new segment '%s' size %d", mp->segment_name,
-		mp->segment_size);
-}
-
-static void
-vl_api_connect_uri_t_handler (vl_api_connect_uri_t * mp)
-{
-  u32 segment_index;
-  uri_udp_test_main_t *utm = &uri_udp_test_main;
-  svm_fifo_segment_main_t *sm = &svm_fifo_segment_main;
-  svm_fifo_segment_create_args_t _a, *a = &_a;
-  svm_fifo_segment_private_t *seg;
-  unix_shared_memory_queue_t *client_q;
-  vl_api_connect_uri_reply_t *rmp;
-  session_t *session;
-  int rv = 0;
-
-  /* Create the segment */
-  a->segment_name = (char *) format (0, "%d:segment%d%c", utm->my_pid,
-				     utm->unique_segment_index++, 0);
-  a->segment_size = utm->configured_segment_size;
-
-  rv = svm_fifo_segment_create (a);
-  if (rv)
-    {
-      clib_warning ("sm_fifo_segment_create ('%s') failed", a->segment_name);
-      rv = VNET_API_ERROR_URI_FIFO_CREATE_FAILED;
-      goto send_reply;
-    }
-
-  vec_add2 (utm->seg, seg, 1);
-
-  segment_index = vec_len (sm->segments) - 1;
-
-  memcpy (seg, sm->segments + segment_index, sizeof (utm->seg[0]));
-
-  pool_get (utm->sessions, session);
-
-  /*
-   * By construction the master's idea of the rx fifo ends up in
-   * fsh->fifos[0], and the master's idea of the tx fifo ends up in
-   * fsh->fifos[1].
-   */
-  session->server_rx_fifo = svm_fifo_segment_alloc_fifo (utm->seg,
-							 128 * 1024);
-  ASSERT (session->server_rx_fifo);
-
-  session->server_tx_fifo = svm_fifo_segment_alloc_fifo (utm->seg,
-							 128 * 1024);
-  ASSERT (session->server_tx_fifo);
-
-  session->server_rx_fifo->server_session_index = session - utm->sessions;
-  session->server_tx_fifo->server_session_index = session - utm->sessions;
-  utm->cut_through_session_index = session - utm->sessions;
-
-  rv = pthread_create (&utm->cut_through_thread_handle,
-		       NULL /*attr */ , cut_through_thread_fn, 0);
-  if (rv)
-    {
-      clib_warning ("pthread_create returned %d", rv);
-      rv = VNET_API_ERROR_SYSCALL_ERROR_1;
-    }
-
-send_reply:
-  rmp = vl_msg_api_alloc (sizeof (*rmp));
-  memset (rmp, 0, sizeof (*rmp));
-
-  rmp->_vl_msg_id = ntohs (VL_API_CONNECT_URI_REPLY);
-  rmp->context = mp->context;
-  rmp->retval = ntohl (rv);
-  rmp->segment_name_length = vec_len (a->segment_name);
-  memcpy (rmp->segment_name, a->segment_name, vec_len (a->segment_name));
-
-  vec_free (a->segment_name);
-
-  client_q = (unix_shared_memory_queue_t *) mp->client_queue_address;
-  vl_msg_api_send_shmem (client_q, (u8 *) & rmp);
-}
-
-static void
-vl_api_unbind_uri_reply_t_handler (vl_api_unbind_uri_reply_t * mp)
-{
-  uri_udp_test_main_t *utm = &uri_udp_test_main;
-
-  if (mp->retval != 0)
-    clib_warning ("returned %d", ntohl (mp->retval));
-
-  utm->state = STATE_START;
-}
-
-static void
-vl_api_accept_session_t_handler (vl_api_accept_session_t * mp)
-{
-  uri_udp_test_main_t *utm = &uri_udp_test_main;
-  vl_api_accept_session_reply_t *rmp;
-  svm_fifo_t *rx_fifo, *tx_fifo;
-  session_t *session;
-  static f64 start_time;
-  u64 key;
-
-  if (start_time == 0.0)
-    start_time = clib_time_now (&utm->clib_time);
-
-  utm->vpp_event_queue = (unix_shared_memory_queue_t *)
-    mp->vpp_event_queue_address;
-
-  pool_get (utm->sessions, session);
-
-  rx_fifo = (svm_fifo_t *) mp->server_rx_fifo;
-  rx_fifo->client_session_index = session - utm->sessions;
-  tx_fifo = (svm_fifo_t *) mp->server_tx_fifo;
-  tx_fifo->client_session_index = session - utm->sessions;
-
-  session->server_rx_fifo = rx_fifo;
-  session->server_tx_fifo = tx_fifo;
-
-  key = (((u64) mp->session_thread_index) << 32) | (u64) mp->session_index;
-
-  hash_set (utm->session_index_by_vpp_handles, key, session - utm->sessions);
-
-  utm->state = STATE_READY;
-
-  if (pool_elts (utm->sessions) && (pool_elts (utm->sessions) % 20000) == 0)
-    {
-      f64 now = clib_time_now (&utm->clib_time);
-      fformat (stdout, "%d active sessions in %.2f seconds, %.2f/sec...\n",
-	       pool_elts (utm->sessions), now - start_time,
-	       (f64) pool_elts (utm->sessions) / (now - start_time));
-    }
-
-  rmp = vl_msg_api_alloc (sizeof (*rmp));
-  memset (rmp, 0, sizeof (*rmp));
-  rmp->_vl_msg_id = ntohs (VL_API_ACCEPT_SESSION_REPLY);
-  rmp->session_type = mp->session_type;
-  rmp->session_index = mp->session_index;
-  rmp->session_thread_index = mp->session_thread_index;
-  vl_msg_api_send_shmem (utm->vl_input_queue, (u8 *) & rmp);
-}
-
-static void
-vl_api_disconnect_session_t_handler (vl_api_disconnect_session_t * mp)
-{
-  uri_udp_test_main_t *utm = &uri_udp_test_main;
-  session_t *session;
-  vl_api_disconnect_session_reply_t *rmp;
-  uword *p;
-  int rv = 0;
-  u64 key;
-
-  key = (((u64) mp->session_thread_index) << 32) | (u64) mp->session_index;
-
-  p = hash_get (utm->session_index_by_vpp_handles, key);
-
-  if (p)
-    {
-      session = pool_elt_at_index (utm->sessions, p[0]);
-      hash_unset (utm->session_index_by_vpp_handles, key);
-      pool_put (utm->sessions, session);
-    }
-  else
-    {
-      clib_warning ("couldn't find session key %llx", key);
-      rv = -11;
-    }
-
-  rmp = vl_msg_api_alloc (sizeof (*rmp));
-  memset (rmp, 0, sizeof (*rmp));
-  rmp->_vl_msg_id = ntohs (VL_API_DISCONNECT_SESSION_REPLY);
-  rmp->retval = rv;
-  rmp->session_index = mp->session_index;
-  rmp->session_thread_index = mp->session_thread_index;
-  vl_msg_api_send_shmem (utm->vl_input_queue, (u8 *) & rmp);
-}
-
-static void
-vl_api_connect_uri_reply_t_handler (vl_api_connect_uri_reply_t * mp)
-{
-  svm_fifo_segment_main_t *sm = &svm_fifo_segment_main;
-  uri_udp_test_main_t *utm = &uri_udp_test_main;
-  svm_fifo_segment_create_args_t _a, *a = &_a;
-  ssvm_shared_header_t *sh;
-  svm_fifo_segment_private_t *seg;
-  svm_fifo_segment_header_t *fsh;
-  session_t *session;
-  u32 segment_index;
-  int rv;
-
-  ASSERT (utm->i_am_master == 0);
-
-  if (mp->segment_name_length == 0)
-    {
-      clib_warning ("segment_name_length zero");
-      return;
-    }
-
-  memset (a, 0, sizeof (*a));
-
-  a->segment_name = (char *) mp->segment_name;
-
-  sleep (1);
-
-  rv = svm_fifo_segment_attach (a);
-  if (rv)
-    {
-      clib_warning ("sm_fifo_segment_create ('%v') failed", mp->segment_name);
-      return;
-    }
-
-  segment_index = vec_len (sm->segments) - 1;
-
-  vec_add2 (utm->seg, seg, 1);
-
-  memcpy (seg, sm->segments + segment_index, sizeof (*seg));
-  sh = seg->ssvm.sh;
-  fsh = (svm_fifo_segment_header_t *) sh->opaque[0];
-
-  while (vec_len (fsh->fifos) < 2)
-    sleep (1);
-
-  pool_get (utm->sessions, session);
-  utm->cut_through_session_index = session - utm->sessions;
-
-  session->server_rx_fifo = (svm_fifo_t *) fsh->fifos[0];
-  ASSERT (session->server_rx_fifo);
-  session->server_tx_fifo = (svm_fifo_t *) fsh->fifos[1];
-  ASSERT (session->server_tx_fifo);
-
-  /* security: could unlink /dev/shm/<mp->segment_name> here, maybe */
-
-  utm->state = STATE_READY;
-}
-
-#define foreach_uri_msg                         \
-_(BIND_URI_REPLY, bind_uri_reply)               \
-_(CONNECT_URI, connect_uri)                     \
-_(CONNECT_URI_REPLY, connect_uri_reply)         \
-_(UNBIND_URI_REPLY, unbind_uri_reply)           \
-_(ACCEPT_SESSION, accept_session)		\
-_(DISCONNECT_SESSION, disconnect_session)	\
-_(MAP_ANOTHER_SEGMENT, map_another_segment)
-
-void
-uri_api_hookup (uri_udp_test_main_t * utm)
-{
-#define _(N,n)                                                  \
-    vl_msg_api_set_handlers(VL_API_##N, #n,                     \
-                           vl_api_##n##_t_handler,              \
-                           vl_noop_handler,                     \
-                           vl_api_##n##_t_endian,               \
-                           vl_api_##n##_t_print,                \
-                           sizeof(vl_api_##n##_t), 1);
-  foreach_uri_msg;
-#undef _
-
-}
-
-
-int
-connect_to_vpp (char *name)
-{
-  uri_udp_test_main_t *utm = &uri_udp_test_main;
-  api_main_t *am = &api_main;
-
-  if (vl_client_connect_to_vlib ("/vpe-api", name, 32) < 0)
-    return -1;
-
-  utm->vl_input_queue = am->shmem_hdr->vl_input_queue;
-  utm->my_client_index = am->my_client_index;
-
-  return 0;
-}
-
-void
-vlib_cli_output (struct vlib_main_t *vm, char *fmt, ...)
-{
-  clib_warning ("BUG");
-}
-
-static void
-init_error_string_table (uri_udp_test_main_t * utm)
-{
-  utm->error_string_by_error_number = hash_create (0, sizeof (uword));
-
-#define _(n,v,s) hash_set (utm->error_string_by_error_number, -v, s);
-  foreach_vnet_api_error;
-#undef _
-
-  hash_set (utm->error_string_by_error_number, 99, "Misc");
-}
-
-void
-handle_fifo_event_server_rx (uri_udp_test_main_t * utm,
-			     session_fifo_event_t * e)
-{
-  svm_fifo_t *rx_fifo, *tx_fifo;
-  int nbytes;
-
-  session_fifo_event_t evt;
-  unix_shared_memory_queue_t *q;
-  int rv;
-
-  rx_fifo = e->fifo;
-  tx_fifo = utm->sessions[rx_fifo->client_session_index].server_tx_fifo;
-
-  do
-    {
-      nbytes = svm_fifo_dequeue_nowait (rx_fifo, 0,
-					vec_len (utm->rx_buf), utm->rx_buf);
-    }
-  while (nbytes <= 0);
-  do
-    {
-      rv = svm_fifo_enqueue_nowait (tx_fifo, 0, nbytes, utm->rx_buf);
-    }
-  while (rv == -2);
-
-  /* Fabricate TX event, send to vpp */
-  evt.fifo = tx_fifo;
-  evt.event_type = FIFO_EVENT_SERVER_TX;
-  /* $$$$ for event logging */
-  evt.enqueue_length = nbytes;
-  evt.event_id = e->event_id;
-  q = utm->vpp_event_queue;
-  unix_shared_memory_queue_add (q, (u8 *) & evt, 0 /* do wait for mutex */ );
-}
-
-void
-handle_event_queue (uri_udp_test_main_t * utm)
-{
-  session_fifo_event_t _e, *e = &_e;;
-
-  while (1)
-    {
-      unix_shared_memory_queue_sub (utm->our_event_queue, (u8 *) e,
-				    0 /* nowait */ );
-      switch (e->event_type)
-	{
-	case FIFO_EVENT_SERVER_RX:
-	  handle_fifo_event_server_rx (utm, e);
-	  break;
-
-	case FIFO_EVENT_SERVER_EXIT:
-	  return;
-
-	default:
-	  clib_warning ("unknown event type %d", e->event_type);
-	  break;
-	}
-      if (PREDICT_FALSE (utm->time_to_stop == 1))
-	break;
-      if (PREDICT_FALSE (utm->time_to_print_stats == 1))
-	{
-	  utm->time_to_print_stats = 0;
-	  fformat (stdout, "%d connections\n", pool_elts (utm->sessions));
-	}
-    }
-}
-
-void
-uri_udp_test (uri_udp_test_main_t * utm)
-{
-  vl_api_bind_uri_t *bmp;
-  vl_api_unbind_uri_t *ump;
-
-  bmp = vl_msg_api_alloc (sizeof (*bmp));
-  memset (bmp, 0, sizeof (*bmp));
-
-  bmp->_vl_msg_id = ntohs (VL_API_BIND_URI);
-  bmp->client_index = utm->my_client_index;
-  bmp->context = ntohl (0xfeedface);
-  bmp->initial_segment_size = 256 << 20;	/* size of initial segment */
-  bmp->options[SESSION_OPTIONS_FLAGS] =
-    SESSION_OPTIONS_FLAGS_USE_FIFO | SESSION_OPTIONS_FLAGS_ADD_SEGMENT;
-  bmp->options[SESSION_OPTIONS_RX_FIFO_SIZE] = 16 << 10;
-  bmp->options[SESSION_OPTIONS_TX_FIFO_SIZE] = 16 << 10;
-  bmp->options[SESSION_OPTIONS_ADD_SEGMENT_SIZE] = 128 << 20;
-  memcpy (bmp->uri, utm->uri, vec_len (utm->uri));
-  vl_msg_api_send_shmem (utm->vl_input_queue, (u8 *) & bmp);
-
-  if (wait_for_state_change (utm, STATE_READY))
-    {
-      clib_warning ("timeout waiting for STATE_READY");
-      return;
-    }
-
-  handle_event_queue (utm);
-
-  ump = vl_msg_api_alloc (sizeof (*ump));
-  memset (ump, 0, sizeof (*ump));
-
-  ump->_vl_msg_id = ntohs (VL_API_UNBIND_URI);
-  ump->client_index = utm->my_client_index;
-  memcpy (ump->uri, utm->uri, vec_len (utm->uri));
-  vl_msg_api_send_shmem (utm->vl_input_queue, (u8 *) & ump);
-
-  if (wait_for_state_change (utm, STATE_START))
-    {
-      clib_warning ("timeout waiting for STATE_START");
-      return;
-    }
-
-  fformat (stdout, "Test complete...\n");
-}
-
-int
-main (int argc, char **argv)
-{
-  uri_udp_test_main_t *utm = &uri_udp_test_main;
-  unformat_input_t _argv, *a = &_argv;
-  u8 *chroot_prefix;
-  u8 *heap;
-  u8 *bind_name = (u8 *) "udp://0.0.0.0/1234";
-  u32 tmp;
-  mheap_t *h;
-  session_t *session;
-  int i;
-  int i_am_master = 1;
-
-  clib_mem_init (0, 256 << 20);
-
-  heap = clib_mem_get_per_cpu_heap ();
-  h = mheap_header (heap);
-
-  /* make the main heap thread-safe */
-  h->flags |= MHEAP_FLAG_THREAD_SAFE;
-
-  vec_validate (utm->rx_buf, 8192);
-
-  utm->session_index_by_vpp_handles = hash_create (0, sizeof (uword));
-
-  utm->my_pid = getpid ();
-  utm->configured_segment_size = 1 << 20;
-
-  clib_time_init (&utm->clib_time);
-  init_error_string_table (utm);
-  svm_fifo_segment_init (0x200000000ULL, 20);
-  unformat_init_command_line (a, argv);
-
-  while (unformat_check_input (a) != UNFORMAT_END_OF_INPUT)
-    {
-      if (unformat (a, "chroot prefix %s", &chroot_prefix))
-	{
-	  vl_set_memory_root_path ((char *) chroot_prefix);
-	}
-      else if (unformat (a, "uri %s", &bind_name))
-	;
-      else if (unformat (a, "segment-size %dM", &tmp))
-	utm->configured_segment_size = tmp << 20;
-      else if (unformat (a, "segment-size %dG", &tmp))
-	utm->configured_segment_size = tmp << 30;
-      else if (unformat (a, "master"))
-	i_am_master = 1;
-      else if (unformat (a, "slave"))
-	i_am_master = 0;
-      else
-	{
-	  fformat (stderr, "%s: usage [master|slave]\n");
-	  exit (1);
-	}
-    }
-
-  utm->cut_through_session_index = ~0;
-  utm->uri = format (0, "%s%c", bind_name, 0);
-  utm->i_am_master = i_am_master;
-  utm->segment_main = &svm_fifo_segment_main;
-
-  utm->connect_uri = format (0, "udp://10.0.0.1/1234%c", 0);
-
-  setup_signal_handlers ();
-
-  uri_api_hookup (utm);
-
-  if (connect_to_vpp (i_am_master ? "uri_udp_master" : "uri_udp_slave") < 0)
-    {
-      svm_region_exit ();
-      fformat (stderr, "Couldn't connect to vpe, exiting...\n");
-      exit (1);
-    }
-
-  if (i_am_master == 0)
-    {
-      uri_udp_slave_test (utm);
-      exit (0);
-    }
-
-  /* $$$$ hack preallocation */
-  for (i = 0; i < 200000; i++)
-    {
-      pool_get (utm->sessions, session);
-      memset (session, 0, sizeof (*session));
-    }
-  for (i = 0; i < 200000; i++)
-    pool_put_index (utm->sessions, i);
-
-  uri_udp_test (utm);
-
-  vl_client_disconnect_from_vlib ();
-  exit (0);
-}
-
-#undef vl_api_version
-#define vl_api_version(n,v) static u32 vpe_api_version = v;
-#include <vpp/api/vpe.api.h>
-#undef vl_api_version
-
-void
-vl_client_add_api_signatures (vl_api_memclnt_create_t * mp)
-{
-  /*
-   * Send the main API signature in slot 0. This bit of code must
-   * match the checks in ../vpe/api/api.c: vl_msg_api_version_check().
-   */
-  mp->api_versions[0] = clib_host_to_net_u32 (vpe_api_version);
-}
-
-u32
-vl (void *p)
-{
-  return vec_len (p);
-}
-
-/*
- * fd.io coding-style-patch-verification: ON
- *
- * Local Variables:
- * eval: (c-set-style "gnu")
- * End:
- */
diff --git a/src/uri/uritest.c b/src/uri/uritest.c
deleted file mode 100644
index edcdb3ad..00000000
--- a/src/uri/uritest.c
+++ /dev/null
@@ -1,484 +0,0 @@
-/*
- * Copyright (c) 2016 Cisco and/or its affiliates.
- * Licensed under the Apache License, Version 2.0 (the "License");
- * you may not use this file except in compliance with the License.
- * You may obtain a copy of the License at:
- *
- *     http://www.apache.org/licenses/LICENSE-2.0
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS,
- * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
- * See the License for the specific language governing permissions and
- * limitations under the License.
- */
-
-#include <stdio.h>
-#include <setjmp.h>
-#include <vppinfra/clib.h>
-#include <vppinfra/format.h>
-#include <vppinfra/error.h>
-#include <vppinfra/time.h>
-#include <vppinfra/macros.h>
-#include <vnet/vnet.h>
-#include <vlib/vlib.h>
-#include <vlib/unix/unix.h>
-#include <vlibapi/api.h>
-#include <vlibmemory/api.h>
-#include <vpp-api/vpe_msg_enum.h>
-#include <svm_fifo_segment.h>
-
-#define vl_typedefs		/* define message structures */
-#include <vpp-api/vpe_all_api_h.h>
-#undef vl_typedefs
-
-/* declare message handlers for each api */
-
-#define vl_endianfun		/* define message structures */
-#include <vpp-api/vpe_all_api_h.h>
-#undef vl_endianfun
-
-/* instantiate all the print functions we know about */
-#define vl_print(handle, ...)
-#define vl_printfun
-#include <vpp-api/vpe_all_api_h.h>
-#undef vl_printfun
-
-typedef enum
-{
-  STATE_START,
-  STATE_READY,
-  STATE_DISCONNECTING,
-} connection_state_t;
-
-typedef struct
-{
-  /* vpe input queue */
-  unix_shared_memory_queue_t *vl_input_queue;
-
-  /* API client handle */
-  u32 my_client_index;
-
-  /* role */
-  int i_am_master;
-
-  /* The URI we're playing with */
-  u8 *uri;
-
-  /* fifo segment */
-  svm_fifo_segment_private_t *seg;
-
-  svm_fifo_t *rx_fifo;
-  svm_fifo_t *tx_fifo;
-
-  /* For deadman timers */
-  clib_time_t clib_time;
-
-  /* State of the connection, shared between msg RX thread and main thread */
-  volatile connection_state_t state;
-
-  /* VNET_API_ERROR_FOO -> "Foo" hash table */
-  uword *error_string_by_error_number;
-} uritest_main_t;
-
-#if CLIB_DEBUG > 0
-#define NITER 1000
-#else
-#define NITER 1000000
-#endif
-
-uritest_main_t uritest_main;
-
-u8 *
-format_api_error (u8 * s, va_list * args)
-{
-  uritest_main_t *utm = va_arg (*args, uritest_main_t *);
-  i32 error = va_arg (*args, u32);
-  uword *p;
-
-  p = hash_get (utm->error_string_by_error_number, -error);
-
-  if (p)
-    s = format (s, "%s", p[0]);
-  else
-    s = format (s, "%d", error);
-  return s;
-}
-
-int
-wait_for_state_change (uritest_main_t * utm, connection_state_t state)
-{
-  f64 timeout = clib_time_now (&utm->clib_time) + 1.0;
-
-  while (clib_time_now (&utm->clib_time) < timeout)
-    {
-      if (utm->state == state)
-	return 0;
-    }
-  return -1;
-}
-
-static void
-vl_api_bind_uri_reply_t_handler (vl_api_bind_uri_reply_t * mp)
-{
-  uritest_main_t *utm = &uritest_main;
-  svm_fifo_segment_create_args_t _a, *a = &_a;
-  int rv;
-
-  ASSERT (utm->i_am_master);
-
-  if (mp->segment_name_length == 0)
-    {
-      clib_warning ("segment_name_length zero");
-      return;
-    }
-
-  a->segment_name = (char *) mp->segment_name;
-  a->segment_size = mp->segment_size;
-
-  /* Create the segment */
-  rv = svm_fifo_segment_create (a);
-  if (rv)
-    {
-      clib_warning ("sm_fifo_segment_create ('%s') failed", mp->segment_name);
-      return;
-    }
-
-  vec_validate (utm->seg, 0);
-
-  memcpy (utm->seg, a->rv, sizeof (*utm->seg));
-
-  /*
-   * By construction the master's idea of the rx fifo ends up in
-   * fsh->fifos[0], and the master's idea of the tx fifo ends up in
-   * fsh->fifos[1].
-   */
-  utm->rx_fifo = svm_fifo_segment_alloc_fifo (utm->seg, 10240);
-  ASSERT (utm->rx_fifo);
-
-  utm->tx_fifo = svm_fifo_segment_alloc_fifo (utm->seg, 10240);
-  ASSERT (utm->tx_fifo);
-
-  utm->state = STATE_READY;
-}
-
-static void
-vl_api_connect_uri_reply_t_handler (vl_api_connect_uri_reply_t * mp)
-{
-  uritest_main_t *utm = &uritest_main;
-  svm_fifo_segment_create_args_t _a, *a = &_a;
-  ssvm_shared_header_t *sh;
-  svm_fifo_segment_header_t *fsh;
-  int rv;
-
-  ASSERT (utm->i_am_master == 0);
-
-  if (mp->segment_name_length == 0)
-    {
-      clib_warning ("segment_name_length zero");
-      return;
-    }
-
-  memset (a, 0, sizeof (*a));
-
-  a->segment_name = (char *) mp->segment_name;
-
-  rv = svm_fifo_segment_attach (a);
-  if (rv)
-    {
-      clib_warning ("sm_fifo_segment_create ('%s') failed", mp->segment_name);
-      return;
-    }
-
-  vec_validate (utm->seg, 0);
-
-  memcpy (utm->seg, a->rv, sizeof (*utm->seg));
-  sh = utm->seg->ssvm.sh;
-  fsh = (svm_fifo_segment_header_t *) sh->opaque[0];
-
-  while (vec_len (fsh->fifos) < 2)
-    sleep (1);
-
-  utm->rx_fifo = (svm_fifo_t *) fsh->fifos[1];
-  ASSERT (utm->rx_fifo);
-  utm->tx_fifo = (svm_fifo_t *) fsh->fifos[0];
-  ASSERT (utm->tx_fifo);
-
-  /* security: could unlink /dev/shm/<mp->segment_name> here, maybe */
-
-  utm->state = STATE_READY;
-}
-
-static void
-vl_api_unbind_uri_reply_t_handler (vl_api_unbind_uri_reply_t * mp)
-{
-  uritest_main_t *utm = &uritest_main;
-
-  if (mp->retval != 0)
-    clib_warning ("returned %d", ntohl (mp->retval));
-
-  utm->state = STATE_START;
-}
-
-#define foreach_uri_msg                         \
-_(BIND_URI_REPLY, bind_uri_reply)               \
-_(CONNECT_URI_REPLY, connect_uri_reply)         \
-_(UNBIND_URI_REPLY, unbind_uri_reply)
-
-void
-uri_api_hookup (uritest_main_t * utm)
-{
-#define _(N,n)                                                  \
-    vl_msg_api_set_handlers(VL_API_##N, #n,                     \
-                           vl_api_##n##_t_handler,	        \
-                           vl_noop_handler,                     \
-                           vl_api_##n##_t_endian,               \
-                           vl_api_##n##_t_print,                \
-                           sizeof(vl_api_##n##_t), 1);
-  foreach_uri_msg;
-#undef _
-
-}
-
-
-int
-connect_to_vpp (char *name)
-{
-  uritest_main_t *utm = &uritest_main;
-  api_main_t *am = &api_main;
-
-  if (vl_client_connect_to_vlib ("/vpe-api", name, 32) < 0)
-    return -1;
-
-  utm->vl_input_queue = am->shmem_hdr->vl_input_queue;
-  utm->my_client_index = am->my_client_index;
-
-  return 0;
-}
-
-void
-vlib_cli_output (struct vlib_main_t *vm, char *fmt, ...)
-{
-  clib_warning ("BUG");
-}
-
-static void
-init_error_string_table (uritest_main_t * utm)
-{
-  utm->error_string_by_error_number = hash_create (0, sizeof (uword));
-
-#define _(n,v,s) hash_set (utm->error_string_by_error_number, -v, s);
-  foreach_vnet_api_error;
-#undef _
-
-  hash_set (utm->error_string_by_error_number, 99, "Misc");
-}
-
-void
-uritest_master (uritest_main_t * utm)
-{
-  vl_api_bind_uri_t *bmp;
-  vl_api_unbind_uri_t *ump;
-  int i;
-  u8 *test_data = 0;
-  u8 *reply = 0;
-  u32 reply_len;
-  int mypid = getpid ();
-
-  for (i = 0; i < 2048; i++)
-    vec_add1 (test_data, 'a' + (i % 32));
-
-  bmp = vl_msg_api_alloc (sizeof (*bmp));
-  memset (bmp, 0, sizeof (*bmp));
-
-  bmp->_vl_msg_id = ntohs (VL_API_BIND_URI);
-  bmp->client_index = utm->my_client_index;
-  bmp->context = ntohl (0xfeedface);
-  bmp->segment_size = 256 << 10;
-  memcpy (bmp->uri, utm->uri, vec_len (utm->uri));
-  vl_msg_api_send_shmem (utm->vl_input_queue, (u8 *) & bmp);
-
-  if (wait_for_state_change (utm, STATE_READY))
-    {
-      clib_warning ("timeout waiting for STATE_READY");
-      return;
-    }
-
-  for (i = 0; i < NITER; i++)
-    svm_fifo_enqueue (utm->tx_fifo, mypid, vec_len (test_data), test_data);
-
-  vec_validate (reply, 0);
-
-  reply_len = svm_fifo_dequeue (utm->rx_fifo, mypid, vec_len (reply), reply);
-
-  if (reply_len != 1)
-    clib_warning ("reply length %d", reply_len);
-
-  if (reply[0] == 1)
-    fformat (stdout, "Test OK...");
-
-  ump = vl_msg_api_alloc (sizeof (*ump));
-  memset (ump, 0, sizeof (*ump));
-
-  ump->_vl_msg_id = ntohs (VL_API_UNBIND_URI);
-  ump->client_index = utm->my_client_index;
-  memcpy (ump->uri, utm->uri, vec_len (utm->uri));
-  vl_msg_api_send_shmem (utm->vl_input_queue, (u8 *) & ump);
-
-  if (wait_for_state_change (utm, STATE_START))
-    {
-      clib_warning ("timeout waiting for STATE_READY");
-      return;
-    }
-
-  fformat (stdout, "Master done...\n");
-}
-
-void
-uritest_slave (uritest_main_t * utm)
-{
-  vl_api_connect_uri_t *cmp;
-  int i, j;
-  u8 *test_data = 0;
-  u8 *reply = 0;
-  u32 bytes_received = 0;
-  u32 actual_bytes;
-  int mypid = getpid ();
-  u8 ok;
-  f64 before, after, delta, bytes_per_second;
-
-  vec_validate (test_data, 4095);
-
-  cmp = vl_msg_api_alloc (sizeof (*cmp));
-  memset (cmp, 0, sizeof (*cmp));
-
-  cmp->_vl_msg_id = ntohs (VL_API_CONNECT_URI);
-  cmp->client_index = utm->my_client_index;
-  cmp->context = ntohl (0xfeedface);
-  memcpy (cmp->uri, utm->uri, vec_len (utm->uri));
-  vl_msg_api_send_shmem (utm->vl_input_queue, (u8 *) & cmp);
-
-  if (wait_for_state_change (utm, STATE_READY))
-    {
-      clib_warning ("timeout waiting for STATE_READY");
-      return;
-    }
-
-  ok = 1;
-  before = clib_time_now (&utm->clib_time);
-  for (i = 0; i < NITER; i++)
-    {
-      actual_bytes = svm_fifo_dequeue (utm->rx_fifo, mypid,
-				       vec_len (test_data), test_data);
-      j = 0;
-      while (j < actual_bytes)
-	{
-	  if (test_data[j] != ('a' + (bytes_received % 32)))
-	    ok = 0;
-	  bytes_received++;
-	  j++;
-	}
-      if (bytes_received == NITER * 2048)
-	break;
-    }
-
-  vec_add1 (reply, ok);
-
-  svm_fifo_enqueue (utm->tx_fifo, mypid, vec_len (reply), reply);
-  after = clib_time_now (&utm->clib_time);
-  delta = after - before;
-  bytes_per_second = 0.0;
-
-  if (delta > 0.0)
-    bytes_per_second = (f64) bytes_received / delta;
-
-  fformat (stdout,
-	   "Slave done, %d bytes in %.2f seconds, %.2f bytes/sec...\n",
-	   bytes_received, delta, bytes_per_second);
-}
-
-int
-main (int argc, char **argv)
-{
-  uritest_main_t *utm = &uritest_main;
-  unformat_input_t _argv, *a = &_argv;
-  u8 *chroot_prefix;
-  u8 *heap;
-  char *bind_name = "fifo:uritest";
-  mheap_t *h;
-  int i_am_master = 0;
-
-  clib_mem_init (0, 128 << 20);
-
-  heap = clib_mem_get_per_cpu_heap ();
-  h = mheap_header (heap);
-
-  /* make the main heap thread-safe */
-  h->flags |= MHEAP_FLAG_THREAD_SAFE;
-
-  clib_time_init (&utm->clib_time);
-  init_error_string_table (utm);
-  svm_fifo_segment_init (0x200000000ULL, 20);
-  unformat_init_command_line (a, argv);
-
-  utm->uri = format (0, "%s%c", bind_name, 0);
-
-  while (unformat_check_input (a) != UNFORMAT_END_OF_INPUT)
-    {
-      if (unformat (a, "master"))
-	i_am_master = 1;
-      else if (unformat (a, "slave"))
-	i_am_master = 0;
-      else if (unformat (a, "chroot prefix %s", &chroot_prefix))
-	{
-	  vl_set_memory_root_path ((char *) chroot_prefix);
-	}
-      else
-	{
-	  fformat (stderr, "%s: usage [master|slave]\n");
-	  exit (1);
-	}
-    }
-
-  uri_api_hookup (utm);
-
-  if (connect_to_vpp (i_am_master ? "uritest_master" : "uritest_slave") < 0)
-    {
-      svm_region_exit ();
-      fformat (stderr, "Couldn't connect to vpe, exiting...\n");
-      exit (1);
-    }
-
-  utm->i_am_master = i_am_master;
-
-  if (i_am_master)
-    uritest_master (utm);
-  else
-    uritest_slave (utm);
-
-  vl_client_disconnect_from_vlib ();
-  exit (0);
-}
-
-#undef vl_api_version
-#define vl_api_version(n,v) static u32 vpe_api_version = v;
-#include <vpp-api/vpe.api.h>
-#undef vl_api_version
-
-void
-vl_client_add_api_signatures (vl_api_memclnt_create_t * mp)
-{
-  /*
-   * Send the main API signature in slot 0. This bit of code must
-   * match the checks in ../vpe/api/api.c: vl_msg_api_version_check().
-   */
-  mp->api_versions[0] = clib_host_to_net_u32 (vpe_api_version);
-}
-
-/*
- * fd.io coding-style-patch-verification: ON
- *
- * Local Variables:
- * eval: (c-set-style "gnu")
- * End:
- */
diff --git a/src/vnet.am b/src/vnet.am
index 7125a122..4e30ee92 100644
--- a/src/vnet.am
+++ b/src/vnet.am
@@ -461,6 +461,7 @@ libvnet_la_SOURCES +=				\
  vnet/tcp/tcp_output.c				\
  vnet/tcp/tcp_input.c				\
  vnet/tcp/tcp_newreno.c				\
+ vnet/tcp/builtin_server.c			\
  vnet/tcp/tcp.c
 
 nobase_include_HEADERS +=			\
diff --git a/src/vnet/api_errno.h b/src/vnet/api_errno.h
index 5e65ac7b..74d39bdb 100644
--- a/src/vnet/api_errno.h
+++ b/src/vnet/api_errno.h
@@ -103,7 +103,8 @@ _(LISP_RLOC_LOCAL, -110, "RLOC address is local")                       \
 _(BFD_EAGAIN, -111, "BFD object cannot be manipulated at this time")	\
 _(INVALID_GPE_MODE, -112, "Invalid GPE mode")                           \
 _(LISP_GPE_ENTRIES_PRESENT, -113, "LISP GPE entries are present")       \
-_(ADDRESS_FOUND_FOR_INTERFACE, -114, "Address found for interface")
+_(ADDRESS_FOUND_FOR_INTERFACE, -114, "Address found for interface")	\
+_(SESSION_CONNECT_FAIL, -115, "Session failed to connect")
 
 typedef enum
 {
diff --git a/src/vnet/session/application.c b/src/vnet/session/application.c
index a561e7d1..a542eebe 100644
--- a/src/vnet/session/application.c
+++ b/src/vnet/session/application.c
@@ -154,6 +154,15 @@ application_get (u32 index)
   return pool_elt_at_index (app_pool, index);
 }
 
+application_t *
+application_get_if_valid (u32 index)
+{
+  if (pool_is_free_index (app_pool, index))
+    return 0;
+
+  return pool_elt_at_index (app_pool, index);
+}
+
 u32
 application_get_index (application_t * app)
 {
@@ -209,7 +218,7 @@ format_application_server (u8 * s, va_list * args)
 
   regp = vl_api_client_index_to_registration (srv->api_client_index);
   if (!regp)
-    server_name = format (0, "%s%c", regp->name, 0);
+    server_name = format (0, "builtin-%d%c", srv->index, 0);
   else
     server_name = regp->name;
 
@@ -269,11 +278,17 @@ static clib_error_t *
 show_app_command_fn (vlib_main_t * vm, unformat_input_t * input,
 		     vlib_cli_command_t * cmd)
 {
+  session_manager_main_t *smm = &session_manager_main;
   application_t *app;
   int do_server = 0;
   int do_client = 0;
   int verbose = 0;
 
+  if (!smm->is_enabled)
+    {
+      clib_error_return (0, "session layer is not enabled");
+    }
+
   while (unformat_check_input (input) != UNFORMAT_END_OF_INPUT)
     {
       if (unformat (input, "server"))
@@ -323,16 +338,20 @@ show_app_command_fn (vlib_main_t * vm, unformat_input_t * input,
           /* *INDENT-ON* */
 	}
       else
-	vlib_cli_output (vm, "No active server bindings");
+	vlib_cli_output (vm, "No active client bindings");
     }
 
   return 0;
 }
 
+/* *INDENT-OFF* */
 VLIB_CLI_COMMAND (show_app_command, static) =
 {
-.path = "show app",.short_help =
-    "show app [server|client] [verbose]",.function = show_app_command_fn,};
+  .path = "show app",
+  .short_help = "show app [server|client] [verbose]",
+  .function = show_app_command_fn,
+};
+/* *INDENT-ON* */
 
 /*
  * fd.io coding-style-patch-verification: ON
diff --git a/src/vnet/session/application.h b/src/vnet/session/application.h
index 027d6967..480828f7 100644
--- a/src/vnet/session/application.h
+++ b/src/vnet/session/application.h
@@ -100,6 +100,7 @@ application_t *application_new (application_type_t type, session_type_t sst,
 				session_cb_vft_t * cb_fns);
 void application_del (application_t * app);
 application_t *application_get (u32 index);
+application_t *application_get_if_valid (u32 index);
 application_t *application_lookup (u32 api_client_index);
 u32 application_get_index (application_t * app);
 
diff --git a/src/vnet/session/application_interface.c b/src/vnet/session/application_interface.c
index 0ea77fd8..6ddfb70f 100644
--- a/src/vnet/session/application_interface.c
+++ b/src/vnet/session/application_interface.c
@@ -51,7 +51,7 @@ ip_is_local (ip46_address_t * ip46_address, u8 is_ip4)
       prefix.fp_proto = FIB_PROTOCOL_IP6;
     }
 
-  clib_memcpy (&prefix.fp_addr, ip46_address, sizeof (ip46_address));
+  clib_memcpy (&prefix.fp_addr, ip46_address, sizeof (ip46_address_t));
   fei = fib_table_lookup (0, &prefix);
   flags = fib_entry_get_flags (fei);
 
@@ -186,9 +186,7 @@ vnet_connect_i (u32 api_client_index, u32 api_context, session_type_t sst,
   /*
    * Not connecting to a local server. Create regular session
    */
-  stream_session_open (sst, ip46, port, app->index);
-
-  return 0;
+  return stream_session_open (sst, ip46, port, app->index);
 }
 
 /**
diff --git a/src/vnet/session/node.c b/src/vnet/session/node.c
index e467f4e9..399077de 100644
--- a/src/vnet/session/node.c
+++ b/src/vnet/session/node.c
@@ -104,9 +104,13 @@ session_fifo_rx_i (vlib_main_t * vm, vlib_node_runtime_t * node,
   snd_space0 = transport_vft->send_space (tc0);
   snd_mss0 = transport_vft->send_mss (tc0);
 
+  /* Can't make any progress */
   if (snd_space0 == 0 || svm_fifo_max_dequeue (s0->server_tx_fifo) == 0
       || snd_mss0 == 0)
-    return 0;
+    {
+      vec_add1 (smm->evts_partially_read[thread_index], *e0);
+      return 0;
+    }
 
   ASSERT (e0->enqueue_length > 0);
 
@@ -143,7 +147,12 @@ session_fifo_rx_i (vlib_main_t * vm, vlib_node_runtime_t * node,
 	  if (PREDICT_FALSE (n_bufs < 0.9 * VLIB_FRAME_SIZE))
 	    {
 	      /* Keep track of how much we've dequeued and exit */
-	      e0->enqueue_length -= max_len_to_snd0 - left_to_snd0;
+	      if (left_to_snd0 != max_len_to_snd0)
+		{
+		  e0->enqueue_length -= max_len_to_snd0 - left_to_snd0;
+		  vec_add1 (smm->evts_partially_read[thread_index], *e0);
+		}
+
 	      return -1;
 	    }
 
@@ -185,12 +194,13 @@ session_fifo_rx_i (vlib_main_t * vm, vlib_node_runtime_t * node,
 	      t0->server_thread_index = s0->thread_index;
 	    }
 
+	  /* *INDENT-OFF* */
 	  if (1)
 	    {
-	      ELOG_TYPE_DECLARE (e) =
-	      {
-	      .format = "evt-dequeue: id %d length %d",.format_args =
-		  "i4i4",};
+	      ELOG_TYPE_DECLARE (e) = {
+		  .format = "evt-dequeue: id %d length %d",
+		  .format_args = "i4i4",
+	      };
 	      struct
 	      {
 		u32 data[2];
@@ -199,6 +209,7 @@ session_fifo_rx_i (vlib_main_t * vm, vlib_node_runtime_t * node,
 	      ed->data[0] = e0->event_id;
 	      ed->data[1] = e0->enqueue_length;
 	    }
+	  /* *INDENT-ON* */
 
 	  len_to_deq0 = (left_to_snd0 < snd_mss0) ? left_to_snd0 : snd_mss0;
 
@@ -289,7 +300,7 @@ session_queue_node_fn (vlib_main_t * vm, vlib_node_runtime_t * node,
 {
   session_manager_main_t *smm = vnet_get_session_manager_main ();
   session_fifo_event_t *my_fifo_events, *e;
-  u32 n_to_dequeue;
+  u32 n_to_dequeue, n_events;
   unix_shared_memory_queue_t *q;
   int n_tx_packets = 0;
   u32 my_thread_index = vm->cpu_index;
@@ -309,14 +320,16 @@ session_queue_node_fn (vlib_main_t * vm, vlib_node_runtime_t * node,
 
   /* min number of events we can dequeue without blocking */
   n_to_dequeue = q->cursize;
-  if (n_to_dequeue == 0)
-    return 0;
-
   my_fifo_events = smm->fifo_events[my_thread_index];
 
-  /* If we didn't manage to process previous events try going
+  if (n_to_dequeue == 0 && vec_len (my_fifo_events) == 0)
+    return 0;
+
+  /*
+   * If we didn't manage to process previous events try going
    * over them again without dequeuing new ones.
-   * XXX: Block senders to sessions that can't keep up */
+   */
+  /* XXX: Block senders to sessions that can't keep up */
   if (vec_len (my_fifo_events) >= 100)
     goto skip_dequeue;
 
@@ -338,8 +351,8 @@ session_queue_node_fn (vlib_main_t * vm, vlib_node_runtime_t * node,
   smm->fifo_events[my_thread_index] = my_fifo_events;
 
 skip_dequeue:
-
-  for (i = 0; i < n_to_dequeue; i++)
+  n_events = vec_len (my_fifo_events);
+  for (i = 0; i < n_events; i++)
     {
       svm_fifo_t *f0;		/* $$$ prefetch 1 ahead maybe */
       stream_session_t *s0;
@@ -354,8 +367,13 @@ skip_dequeue:
       /* $$$ add multiple event queues, per vpp worker thread */
       ASSERT (server_thread_index0 == my_thread_index);
 
-      s0 = pool_elt_at_index (smm->sessions[my_thread_index],
-			      server_session_index0);
+      s0 = stream_session_get_if_valid (server_session_index0,
+					my_thread_index);
+      if (!s0)
+	{
+	  clib_warning ("It's dead Jim!");
+	  continue;
+	}
 
       ASSERT (s0->thread_index == my_thread_index);
 
@@ -380,11 +398,11 @@ skip_dequeue:
 done:
 
   /* Couldn't process all events. Probably out of buffers */
-  if (PREDICT_FALSE (i < n_to_dequeue))
+  if (PREDICT_FALSE (i < n_events))
     {
       session_fifo_event_t *partially_read =
 	smm->evts_partially_read[my_thread_index];
-      vec_add (partially_read, &my_fifo_events[i], n_to_dequeue - i);
+      vec_add (partially_read, &my_fifo_events[i], n_events - i);
       vec_free (my_fifo_events);
       smm->fifo_events[my_thread_index] = partially_read;
       smm->evts_partially_read[my_thread_index] = 0;
@@ -413,8 +431,7 @@ VLIB_REGISTER_NODE (session_queue_node) =
   .n_errors = ARRAY_LEN (session_queue_error_strings),
   .error_strings = session_queue_error_strings,
   .n_next_nodes = SESSION_QUEUE_N_NEXT,
-  /* .state = VLIB_NODE_STATE_DISABLED, enable on-demand? */
-  /* edit / add dispositions here */
+  .state = VLIB_NODE_STATE_DISABLED,
   .next_nodes =
   {
       [SESSION_QUEUE_NEXT_DROP] = "error-drop",
diff --git a/src/vnet/session/session.api b/src/vnet/session/session.api
index a7b28c1d..582765b5 100644
--- a/src/vnet/session/session.api
+++ b/src/vnet/session/session.api
@@ -422,6 +422,28 @@ define reset_sock_reply {
   i32 retval;
   u64 handle;
 };
+
+/** \brief enable/disable session layer
+    @param client_index - opaque cookie to identify the sender
+                          client to vpp direction only
+    @param context - sender context, to match reply w/ request
+    @param is_enable - disable session layer if 0, enable otherwise
+*/
+define session_enable_disable {
+  u32 client_index;
+  u32 context;
+  u8 is_enable;
+};
+
+/** \brief Reply for session enable/disable
+    @param context - returned sender context, to match reply w/ request
+    @param retval - return code
+*/
+define session_enable_disable_reply {
+  u32 context;
+  i32 retval;
+};
+
 /*
  * Local Variables:
  * eval: (c-set-style "gnu")
diff --git a/src/vnet/session/session.c b/src/vnet/session/session.c
index 539da613..422527e0 100644
--- a/src/vnet/session/session.c
+++ b/src/vnet/session/session.c
@@ -311,11 +311,11 @@ stream_session_half_open_lookup (session_manager_main_t * smm,
 }
 
 transport_connection_t *
-stream_session_lookup_transport4 (session_manager_main_t * smm,
-				  ip4_address_t * lcl, ip4_address_t * rmt,
+stream_session_lookup_transport4 (ip4_address_t * lcl, ip4_address_t * rmt,
 				  u16 lcl_port, u16 rmt_port, u8 proto,
 				  u32 my_thread_index)
 {
+  session_manager_main_t *smm = &session_manager_main;
   session_kv4_t kv4;
   stream_session_t *s;
   int rv;
@@ -345,11 +345,11 @@ stream_session_lookup_transport4 (session_manager_main_t * smm,
 }
 
 transport_connection_t *
-stream_session_lookup_transport6 (session_manager_main_t * smm,
-				  ip6_address_t * lcl, ip6_address_t * rmt,
+stream_session_lookup_transport6 (ip6_address_t * lcl, ip6_address_t * rmt,
 				  u16 lcl_port, u16 rmt_port, u8 proto,
 				  u32 my_thread_index)
 {
+  session_manager_main_t *smm = &session_manager_main;
   stream_session_t *s;
   session_kv6_t kv6;
   int rv;
@@ -554,7 +554,7 @@ session_manager_allocate_session_fifos (session_manager_main_t * smm,
 					u8 * added_a_segment)
 {
   svm_fifo_segment_private_t *fifo_segment;
-  u32 fifo_size, default_fifo_size = 8192 /* TODO config */ ;
+  u32 fifo_size, default_fifo_size = 128 << 10;	/* TODO config */
   int i;
 
   *added_a_segment = 0;
@@ -948,7 +948,7 @@ void
 connects_session_manager_init (session_manager_main_t * smm, u8 session_type)
 {
   session_manager_t *sm;
-  u32 connect_fifo_size = 8 << 10;	/* Config? */
+  u32 connect_fifo_size = 256 << 10;	/* Config? */
   u32 default_segment_size = 1 << 20;
 
   pool_get (smm->session_managers, sm);
@@ -1055,10 +1055,15 @@ stream_session_delete (stream_session_t * s)
   svm_fifo_segment_free_fifo (fifo_segment, s->server_rx_fifo);
   svm_fifo_segment_free_fifo (fifo_segment, s->server_tx_fifo);
 
-  /* Cleanup app if client */
-  app = application_get (s->app_index);
+  app = application_get_if_valid (s->app_index);
+
+  /* No app. A possibility: after disconnect application called unbind */
+  if (!app)
+    return;
+
   if (app->mode == APP_CLIENT)
     {
+      /* Cleanup app if client */
       application_del (app);
     }
   else if (app->mode == APP_SERVER)
@@ -1068,6 +1073,7 @@ stream_session_delete (stream_session_t * s)
       svm_fifo_t **fifos;
       u32 fifo_index;
 
+      /* For server, see if any segments can be removed */
       sm = session_manager_get (app->session_manager_index);
 
       /* Delete fifo */
@@ -1096,10 +1102,10 @@ stream_session_delete_notify (transport_connection_t * tc)
 {
   stream_session_t *s;
 
+  /* App might've been removed already */
   s = stream_session_get_if_valid (tc->s_index, tc->thread_index);
   if (!s)
     {
-      clib_warning ("Surprised!");
       return;
     }
   stream_session_delete (s);
@@ -1151,16 +1157,24 @@ stream_session_accept (transport_connection_t * tc, u32 listener_index,
   return 0;
 }
 
-void
+int
 stream_session_open (u8 sst, ip46_address_t * addr, u16 port_host_byte_order,
 		     u32 app_index)
 {
   transport_connection_t *tc;
   u32 tci;
   u64 value;
+  int rv;
 
   /* Ask transport to open connection */
-  tci = tp_vfts[sst].open (addr, port_host_byte_order);
+  rv = tp_vfts[sst].open (addr, port_host_byte_order);
+  if (rv < 0)
+    {
+      clib_warning ("Transport failed to open connection.");
+      return VNET_API_ERROR_SESSION_CONNECT_FAIL;
+    }
+
+  tci = rv;
 
   /* Get transport connection */
   tc = tp_vfts[sst].get_half_open (tci);
@@ -1170,6 +1184,8 @@ stream_session_open (u8 sst, ip46_address_t * addr, u16 port_host_byte_order,
 
   /* Add to the half-open lookup table */
   stream_session_half_open_table_add (sst, tc, value);
+
+  return 0;
 }
 
 /**
@@ -1216,16 +1232,13 @@ session_get_transport_vft (u8 type)
 }
 
 static clib_error_t *
-session_manager_main_init (vlib_main_t * vm)
+session_manager_main_enable (vlib_main_t * vm)
 {
-  u32 num_threads;
-  vlib_thread_main_t *vtm = vlib_get_thread_main ();
   session_manager_main_t *smm = &session_manager_main;
+  vlib_thread_main_t *vtm = vlib_get_thread_main ();
+  u32 num_threads;
   int i;
 
-  smm->vlib_main = vm;
-  smm->vnet_main = vnet_get_main ();
-
   num_threads = 1 /* main thread */  + vtm->n_threads;
 
   if (num_threads < 1)
@@ -1272,11 +1285,48 @@ session_manager_main_init (vlib_main_t * vm)
   for (i = 0; i < SESSION_N_TYPES; i++)
     smm->connect_manager_index[i] = INVALID_INDEX;
 
+  smm->is_enabled = 1;
+
   return 0;
 }
 
-VLIB_INIT_FUNCTION (session_manager_main_init);
+clib_error_t *
+vnet_session_enable_disable (vlib_main_t * vm, u8 is_en)
+{
+  if (is_en)
+    {
+      if (session_manager_main.is_enabled)
+	return 0;
+
+      vlib_node_set_state (vm, session_queue_node.index,
+			   VLIB_NODE_STATE_POLLING);
+
+      return session_manager_main_enable (vm);
+    }
+  else
+    {
+      session_manager_main.is_enabled = 0;
+      vlib_node_set_state (vm, session_queue_node.index,
+			   VLIB_NODE_STATE_DISABLED);
+    }
+
+  return 0;
+}
+
+
+clib_error_t *
+session_manager_main_init (vlib_main_t * vm)
+{
+  session_manager_main_t *smm = &session_manager_main;
+
+  smm->vlib_main = vm;
+  smm->vnet_main = vnet_get_main ();
+  smm->is_enabled = 0;
+
+  return 0;
+}
 
+VLIB_INIT_FUNCTION (session_manager_main_init)
 /*
  * fd.io coding-style-patch-verification: ON
  *
diff --git a/src/vnet/session/session.h b/src/vnet/session/session.h
index cf14cca9..46e5ce2c 100644
--- a/src/vnet/session/session.h
+++ b/src/vnet/session/session.h
@@ -213,12 +213,15 @@ struct _session_manager_main
   /** Per transport rx function that can either dequeue or peek */
   session_fifo_rx_fn *session_rx_fns[SESSION_N_TYPES];
 
+  u8 is_enabled;
+
   /* Convenience */
   vlib_main_t *vlib_main;
   vnet_main_t *vnet_main;
 };
 
 extern session_manager_main_t session_manager_main;
+extern vlib_node_registration_t session_queue_node;
 
 /*
  * Session manager function
@@ -276,14 +279,12 @@ stream_session_t *stream_session_lookup6 (ip6_address_t * lcl,
 					  ip6_address_t * rmt, u16 lcl_port,
 					  u16 rmt_port, u8, u32 thread_index);
 transport_connection_t
-  * stream_session_lookup_transport4 (session_manager_main_t * smm,
-				      ip4_address_t * lcl,
+  * stream_session_lookup_transport4 (ip4_address_t * lcl,
 				      ip4_address_t * rmt, u16 lcl_port,
 				      u16 rmt_port, u8 proto,
 				      u32 thread_index);
 transport_connection_t
-  * stream_session_lookup_transport6 (session_manager_main_t * smm,
-				      ip6_address_t * lcl,
+  * stream_session_lookup_transport6 (ip6_address_t * lcl,
 				      ip6_address_t * rmt, u16 lcl_port,
 				      u16 rmt_port, u8 proto,
 				      u32 thread_index);
@@ -338,6 +339,14 @@ stream_session_max_enqueue (transport_connection_t * tc)
   return svm_fifo_max_enqueue (s->server_rx_fifo);
 }
 
+always_inline u32
+stream_session_fifo_size (transport_connection_t * tc)
+{
+  stream_session_t *s = stream_session_get (tc->s_index, tc->thread_index);
+  return s->server_rx_fifo->nitems;
+}
+
+
 int
 stream_session_enqueue_data (transport_connection_t * tc, u8 * data, u16 len,
 			     u8 queue_event);
@@ -356,8 +365,8 @@ void stream_session_reset_notify (transport_connection_t * tc);
 int
 stream_session_accept (transport_connection_t * tc, u32 listener_index,
 		       u8 sst, u8 notify);
-void stream_session_open (u8 sst, ip46_address_t * addr,
-			  u16 port_host_byte_order, u32 api_client_index);
+int stream_session_open (u8 sst, ip46_address_t * addr,
+			 u16 port_host_byte_order, u32 api_client_index);
 void stream_session_disconnect (stream_session_t * s);
 void stream_session_cleanup (stream_session_t * s);
 int
@@ -369,6 +378,8 @@ u8 *format_stream_session (u8 * s, va_list * args);
 void session_register_transport (u8 type, const transport_proto_vft_t * vft);
 transport_proto_vft_t *session_get_transport_vft (u8 type);
 
+clib_error_t *vnet_session_enable_disable (vlib_main_t * vm, u8 is_en);
+
 #endif /* __included_session_h__ */
 
 /*
diff --git a/src/vnet/session/session_api.c b/src/vnet/session/session_api.c
index 9d068684..8852fc6e 100644
--- a/src/vnet/session/session_api.c
+++ b/src/vnet/session/session_api.c
@@ -52,6 +52,8 @@ _(DISCONNECT_SOCK, disconnect_sock)                               	\
 _(DISCONNECT_SOCK_REPLY, disconnect_sock_reply)                        	\
 _(ACCEPT_SOCK_REPLY, accept_sock_reply)                           	\
 _(RESET_SOCK_REPLY, reset_sock_reply)                   		\
+_(SESSION_ENABLE_DISABLE, session_enable_disable)                   	\
+
 
 static int
 send_add_segment_callback (u32 api_client_index, const u8 * segment_name,
@@ -146,7 +148,6 @@ send_session_connected_uri_callback (u32 api_client_index,
   mp = vl_msg_api_alloc (sizeof (*mp));
   mp->_vl_msg_id = clib_host_to_net_u16 (VL_API_CONNECT_URI_REPLY);
   mp->context = app->api_context;
-  mp->retval = is_fail;
   if (!is_fail)
     {
       vpp_queue = session_manager_get_vpp_event_queue (s->thread_index);
@@ -157,6 +158,7 @@ send_session_connected_uri_callback (u32 api_client_index,
       mp->session_type = s->session_type;
       mp->vpp_event_queue_address = (u64) vpp_queue;
       mp->client_event_queue_address = (u64) app->event_queue;
+      mp->retval = 0;
 
       session_manager_get_segment_info (s->server_segment_index, &seg_name,
 					&mp->segment_size);
@@ -164,12 +166,22 @@ send_session_connected_uri_callback (u32 api_client_index,
       if (mp->segment_name_length)
 	clib_memcpy (mp->segment_name, seg_name, mp->segment_name_length);
     }
+  else
+    {
+      mp->retval = VNET_API_ERROR_SESSION_CONNECT_FAIL;
+    }
 
   vl_msg_api_send_shmem (q, (u8 *) & mp);
 
   /* Remove client if connect failed */
   if (is_fail)
-    application_del (app);
+    {
+      application_del (app);
+    }
+  else
+    {
+      s->session_state = SESSION_STATE_READY;
+    }
 
   return 0;
 }
@@ -431,6 +443,17 @@ api_session_not_valid (u32 session_index, u32 thread_index)
   return 0;
 }
 
+static void
+vl_api_session_enable_disable_t_handler (vl_api_session_enable_disable_t * mp)
+{
+  vl_api_session_enable_disable_reply_t *rmp;
+  vlib_main_t *vm = vlib_get_main ();
+  int rv = 0;
+
+  vnet_session_enable_disable (vm, mp->is_enable);
+  REPLY_MACRO (VL_API_SESSION_ENABLE_DISABLE_REPLY);
+}
+
 static void
 vl_api_bind_uri_t_handler (vl_api_bind_uri_t * mp)
 {
@@ -476,7 +499,6 @@ vl_api_bind_uri_t_handler (vl_api_bind_uri_t * mp)
       }
   }));
   /* *INDENT-ON* */
-
 }
 
 static void
@@ -493,7 +515,9 @@ vl_api_unbind_uri_t_handler (vl_api_unbind_uri_t * mp)
 static void
 vl_api_connect_uri_t_handler (vl_api_connect_uri_t * mp)
 {
+  vl_api_connect_uri_reply_t *rmp;
   vnet_connect_args_t _a, *a = &_a;
+  int rv;
 
   a->uri = (char *) mp->uri;
   a->api_client_index = mp->client_index;
@@ -501,7 +525,19 @@ vl_api_connect_uri_t_handler (vl_api_connect_uri_t * mp)
   a->options = mp->options;
   a->session_cb_vft = &uri_session_cb_vft;
   a->mp = mp;
-  vnet_connect_uri (a);
+
+  rv = vnet_connect_uri (a);
+
+  if (rv == 0 || rv == VNET_CONNECT_REDIRECTED)
+    return;
+
+  /* Got some error, relay it */
+
+  /* *INDENT-OFF* */
+  REPLY_MACRO2 (VL_API_CONNECT_URI_REPLY, ({
+    rmp->retval = rv;
+  }));
+  /* *INDENT-ON* */
 }
 
 static void
@@ -662,7 +698,9 @@ vl_api_unbind_sock_t_handler (vl_api_unbind_sock_t * mp)
 static void
 vl_api_connect_sock_t_handler (vl_api_connect_sock_t * mp)
 {
+  vl_api_connect_sock_reply_t *rmp;
   vnet_connect_args_t _a, *a = &_a;
+  int rv;
 
   clib_memcpy (&a->tep.ip, mp->ip,
 	       (mp->is_ip4 ? sizeof (ip4_address_t) :
@@ -675,7 +713,18 @@ vl_api_connect_sock_t_handler (vl_api_connect_sock_t * mp)
   a->api_context = mp->context;
   a->mp = mp;
 
-  vnet_connect (a);
+  rv = vnet_connect (a);
+
+  if (rv == 0 || rv == VNET_CONNECT_REDIRECTED)
+    return;
+
+  /* Got some error, relay it */
+
+  /* *INDENT-OFF* */
+  REPLY_MACRO2 (VL_API_CONNECT_URI_REPLY, ({
+    rmp->retval = rv;
+  }));
+  /* *INDENT-ON* */
 }
 
 static void
diff --git a/src/vnet/session/session_cli.c b/src/vnet/session/session_cli.c
index b2943a1c..b029ee65 100644
--- a/src/vnet/session/session_cli.c
+++ b/src/vnet/session/session_cli.c
@@ -60,7 +60,7 @@ format_stream_session (u8 * s, va_list * args)
     }
   else
     {
-      clib_warning ("Session in unknown state!");
+      clib_warning ("Session in state: %d!", ss->session_state);
     }
 
   vec_free (str);
@@ -78,6 +78,11 @@ show_session_command_fn (vlib_main_t * vm, unformat_input_t * input,
   stream_session_t *s;
   u8 *str = 0;
 
+  if (!smm->is_enabled)
+    {
+      clib_error_return (0, "session layer is not enabled");
+    }
+
   while (unformat_check_input (input) != UNFORMAT_END_OF_INPUT)
     {
       if (unformat (input, "verbose"))
@@ -126,11 +131,14 @@ show_session_command_fn (vlib_main_t * vm, unformat_input_t * input,
   return 0;
 }
 
-VLIB_CLI_COMMAND (show_uri_command, static) =
+/* *INDENT-OFF* */
+VLIB_CLI_COMMAND (show_session_command, static) =
 {
-.path = "show session",.short_help = "show session [verbose]",.function =
-    show_session_command_fn,};
-
+  .path = "show session",
+  .short_help = "show session [verbose]",
+  .function = show_session_command_fn,
+};
+/* *INDENT-ON* */
 
 static clib_error_t *
 clear_session_command_fn (vlib_main_t * vm, unformat_input_t * input,
@@ -142,6 +150,11 @@ clear_session_command_fn (vlib_main_t * vm, unformat_input_t * input,
   stream_session_t *pool, *session;
   application_t *server;
 
+  if (!smm->is_enabled)
+    {
+      clib_error_return (0, "session layer is not enabled");
+    }
+
   while (unformat_check_input (input) != UNFORMAT_END_OF_INPUT)
     {
       if (unformat (input, "thread %d", &thread_index))
@@ -174,11 +187,43 @@ clear_session_command_fn (vlib_main_t * vm, unformat_input_t * input,
   return 0;
 }
 
-VLIB_CLI_COMMAND (clear_uri_session_command, static) =
+/* *INDENT-OFF* */
+VLIB_CLI_COMMAND (clear_session_command, static) =
+{
+  .path = "clear session",
+  .short_help = "clear session thread <thread> session <index>",
+  .function = clear_session_command_fn,
+};
+/* *INDENT-ON* */
+
+static clib_error_t *
+session_enable_disable_fn (vlib_main_t * vm, unformat_input_t * input,
+			   vlib_cli_command_t * cmd)
+{
+  u8 is_en = 1;
+
+  while (unformat_check_input (input) != UNFORMAT_END_OF_INPUT)
+    {
+      if (unformat (input, "enable"))
+	is_en = 1;
+      else if (unformat (input, "disable"))
+	is_en = 0;
+      else
+	return clib_error_return (0, "unknown input `%U'",
+				  format_unformat_error, input);
+    }
+
+  return vnet_session_enable_disable (vm, is_en);
+}
+
+/* *INDENT-OFF* */
+VLIB_CLI_COMMAND (session_enable_disable_command, static) =
 {
-.path = "clear session",.short_help =
-    "clear session thread <thread> session <index>",.function =
-    clear_session_command_fn,};
+  .path = "session",
+  .short_help = "session [enable|disable]",
+  .function = session_enable_disable_fn,
+};
+/* *INDENT-ON* */
 
 /*
  * fd.io coding-style-patch-verification: ON
diff --git a/src/vnet/tcp/builtin_server.c b/src/vnet/tcp/builtin_server.c
new file mode 100644
index 00000000..be65642a
--- /dev/null
+++ b/src/vnet/tcp/builtin_server.c
@@ -0,0 +1,135 @@
+/*
+* Copyright (c) 2015-2017 Cisco and/or its affiliates.
+* Licensed under the Apache License, Version 2.0 (the "License");
+* you may not use this file except in compliance with the License.
+* You may obtain a copy of the License at:
+*
+*     http://www.apache.org/licenses/LICENSE-2.0
+*
+* Unless required by applicable law or agreed to in writing, software
+* distributed under the License is distributed on an "AS IS" BASIS,
+* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+* See the License for the specific language governing permissions and
+* limitations under the License.
+*/
+
+#include <vnet/vnet.h>
+#include <vlibmemory/api.h>
+#include <vnet/session/application.h>
+#include <vnet/session/application_interface.h>
+
+int
+builtin_session_accept_callback (stream_session_t * s)
+{
+  clib_warning ("called...");
+  s->session_state = SESSION_STATE_READY;
+  return 0;
+}
+
+void
+builtin_session_disconnect_callback (stream_session_t * s)
+{
+  clib_warning ("called...");
+}
+
+int
+builtin_session_connected_callback (u32 client_index,
+				    stream_session_t * s, u8 is_fail)
+{
+  clib_warning ("called...");
+  return -1;
+}
+
+int
+builtin_add_segment_callback (u32 client_index,
+			      const u8 * seg_name, u32 seg_size)
+{
+  clib_warning ("called...");
+  return -1;
+}
+
+int
+builtin_redirect_connect_callback (u32 client_index, void *mp)
+{
+  clib_warning ("called...");
+  return -1;
+}
+
+int
+builtin_server_rx_callback (stream_session_t * s)
+{
+  clib_warning ("called...");
+  return 0;
+}
+
+static session_cb_vft_t builtin_session_cb_vft = {
+  .session_accept_callback = builtin_session_accept_callback,
+  .session_disconnect_callback = builtin_session_disconnect_callback,
+  .session_connected_callback = builtin_session_connected_callback,
+  .add_segment_callback = builtin_add_segment_callback,
+  .redirect_connect_callback = builtin_redirect_connect_callback,
+  .builtin_server_rx_callback = builtin_server_rx_callback
+};
+
+static int
+server_create (vlib_main_t * vm)
+{
+  vnet_bind_args_t _a, *a = &_a;
+  u64 options[SESSION_OPTIONS_N_OPTIONS];
+  char segment_name[128];
+
+  memset (a, 0, sizeof (*a));
+  memset (options, 0, sizeof (options));
+
+  a->uri = "tcp://0.0.0.0/80";
+  a->api_client_index = ~0;
+  a->session_cb_vft = &builtin_session_cb_vft;
+  a->options = options;
+  a->options[SESSION_OPTIONS_SEGMENT_SIZE] = 256 << 10;
+  a->options[SESSION_OPTIONS_RX_FIFO_SIZE] = 64 << 10;
+  a->options[SESSION_OPTIONS_TX_FIFO_SIZE] = 64 << 10;
+  a->segment_name = segment_name;
+  a->segment_name_length = ARRAY_LEN (segment_name);
+
+  return vnet_bind_uri (a);
+}
+
+static clib_error_t *
+server_create_command_fn (vlib_main_t * vm,
+			  unformat_input_t * input, vlib_cli_command_t * cmd)
+{
+  int rv;
+#if 0
+  while (unformat_check_input (input) != UNFORMAT_END_OF_INPUT)
+    {
+      if (unformat (input, "whatever %d", &whatever))
+	;
+      else
+	return clib_error_return (0, "unknown input `%U'",
+				  format_unformat_error, input);
+    }
+#endif
+
+  rv = server_create (vm);
+  switch (rv)
+    {
+    case 0:
+      break;
+    default:
+      return clib_error_return (0, "server_create returned %d", rv);
+    }
+  return 0;
+}
+
+VLIB_CLI_COMMAND (server_create_command, static) =
+{
+.path = "test server",.short_help = "test server",.function =
+    server_create_command_fn,};
+
+/*
+* fd.io coding-style-patch-verification: ON
+*
+* Local Variables:
+* eval: (c-set-style "gnu")
+* End:
+*/
diff --git a/src/vnet/tcp/tcp.c b/src/vnet/tcp/tcp.c
index 0f9b7097..e5feaeb1 100644
--- a/src/vnet/tcp/tcp.c
+++ b/src/vnet/tcp/tcp.c
@@ -217,6 +217,7 @@ ip_interface_get_first_ip (u32 sw_if_index, u8 is_ip4)
   return 0;
 }
 
+#define PORT_MASK ((1 << 16)- 1)
 /**
  * Allocate local port and add if successful add entry to local endpoint
  * table to mark the pair as used.
@@ -224,7 +225,6 @@ ip_interface_get_first_ip (u32 sw_if_index, u8 is_ip4)
 u16
 tcp_allocate_local_port (tcp_main_t * tm, ip46_address_t * ip)
 {
-  u8 unique = 0;
   transport_endpoint_t *tep;
   u32 time_now, tei;
   u16 min = 1024, max = 65535, tries;	/* XXX configurable ? */
@@ -235,37 +235,34 @@ tcp_allocate_local_port (tcp_main_t * tm, ip46_address_t * ip)
   /* Start at random point or max */
   pool_get (tm->local_endpoints, tep);
   clib_memcpy (&tep->ip, ip, sizeof (*ip));
-  tep->port = random_u32 (&time_now) << 16;
-  tep->port = tep->port < min ? max : tep->port;
 
   /* Search for first free slot */
-  while (tries)
+  for (; tries >= 0; tries--)
     {
+      u16 port = 0;
+
+      /* Find a port in the specified range */
+      while (1)
+	{
+	  port = random_u32 (&time_now) & PORT_MASK;
+	  if (PREDICT_TRUE (port >= min && port < max))
+	    break;
+	}
+
+      tep->port = port;
+
+      /* Look it up */
       tei = transport_endpoint_lookup (&tm->local_endpoints_table, &tep->ip,
 				       tep->port);
+      /* If not found, we're done */
       if (tei == TRANSPORT_ENDPOINT_INVALID_INDEX)
 	{
-	  unique = 1;
-	  break;
+	  transport_endpoint_table_add (&tm->local_endpoints_table, tep,
+					tep - tm->local_endpoints);
+	  return tep->port;
 	}
-
-      tep->port--;
-
-      if (tep->port < min)
-	tep->port = max;
-
-      tries--;
     }
-
-  if (unique)
-    {
-      transport_endpoint_table_add (&tm->local_endpoints_table, tep,
-				    tep - tm->local_endpoints);
-
-      return tep->port;
-    }
-
-  /* Failed */
+  /* No free ports */
   pool_put (tm->local_endpoints, tep);
   return -1;
 }
@@ -360,7 +357,10 @@ tcp_connection_open (ip46_address_t * rmt_addr, u16 rmt_port, u8 is_ip4)
   /* Allocate source port */
   lcl_port = tcp_allocate_local_port (tm, &lcl_addr);
   if (lcl_port < 1)
-    return -1;
+    {
+      clib_warning ("Failed to allocate src port");
+      return -1;
+    }
 
   /*
    * Create connection and send SYN
diff --git a/src/vnet/tcp/tcp.h b/src/vnet/tcp/tcp.h
index 22f00a63..3560509d 100644
--- a/src/vnet/tcp/tcp.h
+++ b/src/vnet/tcp/tcp.h
@@ -30,7 +30,8 @@
 #define TCP_MAX_OPTION_SPACE 40
 
 #define TCP_DUPACK_THRESHOLD 3
-#define TCP_DEFAULT_RX_FIFO_SIZE 64 << 10
+#define TCP_MAX_RX_FIFO_SIZE 2 << 20
+#define TCP_IW_N_SEGMENTS 10
 
 /** TCP FSM state definitions as per RFC793. */
 #define foreach_tcp_fsm_state   \
@@ -590,7 +591,6 @@ vlib_buffer_push_tcp_net_order (vlib_buffer_t * b, u16 sp, u16 dp, u32 seq,
 /**
  * Push TCP header to buffer
  *
- * @param vm - vlib_main
  * @param b - buffer to write the header to
  * @param sp_net - source port net order
  * @param dp_net - destination port net order
diff --git a/src/vnet/tcp/tcp_input.c b/src/vnet/tcp/tcp_input.c
index daa0683b..0a907d0a 100644
--- a/src/vnet/tcp/tcp_input.c
+++ b/src/vnet/tcp/tcp_input.c
@@ -711,7 +711,7 @@ tcp_rcv_ack (tcp_connection_t * tc, vlib_buffer_t * b,
   if (tcp_opts_sack_permitted (&tc->opt))
     tcp_rcv_sacks (tc, vnet_buffer (b)->tcp.ack_number);
 
-  new_snd_wnd = clib_net_to_host_u32 (th->window) << tc->snd_wscale;
+  new_snd_wnd = clib_net_to_host_u16 (th->window) << tc->snd_wscale;
 
   if (tcp_ack_is_dupack (tc, b, new_snd_wnd))
     {
@@ -1320,7 +1320,6 @@ tcp46_syn_sent_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
 
 	  /* Parse options */
 	  tcp_options_parse (tcp0, &new_tc0->opt);
-	  tcp_connection_init_vars (new_tc0);
 
 	  if (tcp_opts_tstamp (&new_tc0->opt))
 	    {
@@ -1331,11 +1330,13 @@ tcp46_syn_sent_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
 	  if (tcp_opts_wscale (&new_tc0->opt))
 	    new_tc0->snd_wscale = new_tc0->opt.wscale;
 
-	  new_tc0->snd_wnd = clib_net_to_host_u32 (tcp0->window)
-	    << new_tc0->snd_wscale;
+	  /* No scaling */
+	  new_tc0->snd_wnd = clib_net_to_host_u16 (tcp0->window);
 	  new_tc0->snd_wl1 = seq0;
 	  new_tc0->snd_wl2 = ack0;
 
+	  tcp_connection_init_vars (new_tc0);
+
 	  /* SYN-ACK: See if we can switch to ESTABLISHED state */
 	  if (tcp_ack (tcp0))
 	    {
@@ -1345,6 +1346,9 @@ tcp46_syn_sent_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
 	      new_tc0->snd_una = ack0;
 	      new_tc0->state = TCP_STATE_ESTABLISHED;
 
+	      /* Make sure las is initialized for the wnd computation */
+	      new_tc0->rcv_las = new_tc0->rcv_nxt;
+
 	      /* Notify app that we have connection */
 	      stream_session_connect_notify (&new_tc0->connection, sst, 0);
 
@@ -1575,7 +1579,7 @@ tcp46_rcv_process_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
 
 	      /* Initialize session variables */
 	      tc0->snd_una = vnet_buffer (b0)->tcp.ack_number;
-	      tc0->snd_wnd = clib_net_to_host_u32 (tcp0->window)
+	      tc0->snd_wnd = clib_net_to_host_u16 (tcp0->window)
 		<< tc0->opt.wscale;
 	      tc0->snd_wl1 = vnet_buffer (b0)->tcp.seq_number;
 	      tc0->snd_wl2 = vnet_buffer (b0)->tcp.ack_number;
@@ -1899,7 +1903,6 @@ tcp46_listen_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
 	    }
 
 	  tcp_options_parse (th0, &child0->opt);
-	  tcp_connection_init_vars (child0);
 
 	  child0->irs = vnet_buffer (b0)->tcp.seq_number;
 	  child0->rcv_nxt = vnet_buffer (b0)->tcp.seq_number + 1;
@@ -1913,6 +1916,16 @@ tcp46_listen_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
 	      child0->tsval_recent_age = tcp_time_now ();
 	    }
 
+	  if (tcp_opts_wscale (&child0->opt))
+	    child0->snd_wscale = child0->opt.wscale;
+
+	  /* No scaling */
+	  child0->snd_wnd = clib_net_to_host_u16 (th0->window);
+	  child0->snd_wl1 = vnet_buffer (b0)->tcp.seq_number;
+	  child0->snd_wl2 = vnet_buffer (b0)->tcp.ack_number;
+
+	  tcp_connection_init_vars (child0);
+
 	  /* Reuse buffer to make syn-ack and send */
 	  tcp_make_synack (child0, b0);
 	  next0 = tcp_next_output (is_ip4);
@@ -1923,7 +1936,7 @@ tcp46_listen_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
 
 	    }
 
-	  b0->error = error0 ? node->errors[error0] : 0;
+	  b0->error = node->errors[error0];
 
 	  vlib_validate_buffer_enqueue_x1 (vm, node, next_index, to_next,
 					   n_left_to_next, bi0, next0);
@@ -2069,7 +2082,6 @@ tcp46_input_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
   u32 n_left_from, next_index, *from, *to_next;
   u32 my_thread_index = vm->cpu_index;
   tcp_main_t *tm = vnet_get_tcp_main ();
-  session_manager_main_t *ssm = vnet_get_session_manager_main ();
 
   from = vlib_frame_vector_args (from_frame);
   n_left_from = from_frame->n_vectors;
@@ -2109,26 +2121,26 @@ tcp46_input_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
 
 	      /* lookup session */
 	      tc0 =
-		(tcp_connection_t *) stream_session_lookup_transport4 (ssm,
-								       &ip40->dst_address,
-								       &ip40->src_address,
-								       tcp0->dst_port,
-								       tcp0->src_port,
-								       SESSION_TYPE_IP4_TCP,
-								       my_thread_index);
+		(tcp_connection_t *)
+		stream_session_lookup_transport4 (&ip40->dst_address,
+						  &ip40->src_address,
+						  tcp0->dst_port,
+						  tcp0->src_port,
+						  SESSION_TYPE_IP4_TCP,
+						  my_thread_index);
 	    }
 	  else
 	    {
 	      ip60 = vlib_buffer_get_current (b0);
 	      tcp0 = ip6_next_header (ip60);
 	      tc0 =
-		(tcp_connection_t *) stream_session_lookup_transport6 (ssm,
-								       &ip60->src_address,
-								       &ip60->dst_address,
-								       tcp0->src_port,
-								       tcp0->dst_port,
-								       SESSION_TYPE_IP6_TCP,
-								       my_thread_index);
+		(tcp_connection_t *)
+		stream_session_lookup_transport6 (&ip60->src_address,
+						  &ip60->dst_address,
+						  tcp0->src_port,
+						  tcp0->dst_port,
+						  SESSION_TYPE_IP6_TCP,
+						  my_thread_index);
 	    }
 
 	  /* Session exists */
diff --git a/src/vnet/tcp/tcp_output.c b/src/vnet/tcp/tcp_output.c
index dbcf1f74..7e431cd0 100644
--- a/src/vnet/tcp/tcp_output.c
+++ b/src/vnet/tcp/tcp_output.c
@@ -90,6 +90,15 @@ tcp_window_compute_scale (u32 available_space)
   return wnd_scale;
 }
 
+/**
+ * TCP's IW as recommended by RFC6928
+ */
+always_inline u32
+tcp_initial_wnd_unscaled (tcp_connection_t * tc)
+{
+  return TCP_IW_N_SEGMENTS * dummy_mtu;
+}
+
 /**
  * Compute initial window and scale factor. As per RFC1323, window field in
  * SYN and SYN-ACK segments is never scaled.
@@ -97,18 +106,15 @@ tcp_window_compute_scale (u32 available_space)
 u32
 tcp_initial_window_to_advertise (tcp_connection_t * tc)
 {
-  u32 available_space;
+  u32 max_fifo;
 
   /* Initial wnd for SYN. Fifos are not allocated yet.
-   * Use some predefined value */
-  if (tc->state != TCP_STATE_SYN_RCVD)
-    {
-      return TCP_DEFAULT_RX_FIFO_SIZE;
-    }
+   * Use some predefined value. For SYN-ACK we still want the
+   * scale to be computed in the same way */
+  max_fifo = TCP_MAX_RX_FIFO_SIZE;
 
-  available_space = stream_session_max_enqueue (&tc->connection);
-  tc->rcv_wscale = tcp_window_compute_scale (available_space);
-  tc->rcv_wnd = clib_min (available_space, TCP_WND_MAX << tc->rcv_wscale);
+  tc->rcv_wscale = tcp_window_compute_scale (max_fifo);
+  tc->rcv_wnd = tcp_initial_wnd_unscaled (tc);
 
   return clib_min (tc->rcv_wnd, TCP_WND_MAX);
 }
@@ -119,23 +125,43 @@ tcp_initial_window_to_advertise (tcp_connection_t * tc)
 u32
 tcp_window_to_advertise (tcp_connection_t * tc, tcp_state_t state)
 {
-  u32 available_space, wnd, scaled_space;
+  u32 available_space, max_fifo, observed_wnd;
 
-  if (state != TCP_STATE_ESTABLISHED)
+  if (state < TCP_STATE_ESTABLISHED)
     return tcp_initial_window_to_advertise (tc);
 
+  /*
+   * Figure out how much space we have available
+   */
   available_space = stream_session_max_enqueue (&tc->connection);
-  scaled_space = available_space >> tc->rcv_wscale;
+  max_fifo = stream_session_fifo_size (&tc->connection);
+
+  ASSERT (tc->opt.mss < max_fifo);
+
+  if (available_space < tc->opt.mss && available_space < max_fifo / 8)
+    available_space = 0;
 
-  /* Need to update scale */
-  if (PREDICT_FALSE ((scaled_space == 0 && available_space != 0))
-      || (scaled_space >= TCP_WND_MAX))
-    tc->rcv_wscale = tcp_window_compute_scale (available_space);
+  /*
+   * Use the above and what we know about what we've previously advertised
+   * to compute the new window
+   */
+  observed_wnd = tc->rcv_wnd - (tc->rcv_nxt - tc->rcv_las);
 
-  wnd = clib_min (available_space, TCP_WND_MAX << tc->rcv_wscale);
-  tc->rcv_wnd = wnd;
+  /* Bad. Thou shalt not shrink */
+  if (available_space < observed_wnd)
+    {
+      if (available_space == 0)
+	clib_warning ("Didn't shrink rcv window despite not having space");
+    }
+
+  tc->rcv_wnd = clib_min (available_space, TCP_WND_MAX << tc->rcv_wscale);
+
+  if (tc->rcv_wnd == 0)
+    {
+      tc->flags |= TCP_CONN_SENT_RCV_WND0;
+    }
 
-  return wnd >> tc->rcv_wscale;
+  return tc->rcv_wnd >> tc->rcv_wscale;
 }
 
 /**
@@ -225,7 +251,7 @@ tcp_options_write (u8 * data, tcp_options_t * opts)
 }
 
 always_inline int
-tcp_make_syn_options (tcp_options_t * opts, u32 initial_wnd)
+tcp_make_syn_options (tcp_options_t * opts, u8 wnd_scale)
 {
   u8 len = 0;
 
@@ -234,7 +260,7 @@ tcp_make_syn_options (tcp_options_t * opts, u32 initial_wnd)
   len += TCP_OPTION_LEN_MSS;
 
   opts->flags |= TCP_OPTS_FLAG_WSCALE;
-  opts->wscale = tcp_window_compute_scale (initial_wnd);
+  opts->wscale = wnd_scale;
   len += TCP_OPTION_LEN_WINDOW_SCALE;
 
   opts->flags |= TCP_OPTS_FLAG_TSTAMP;
@@ -327,8 +353,7 @@ tcp_make_options (tcp_connection_t * tc, tcp_options_t * opts,
     case TCP_STATE_SYN_RCVD:
       return tcp_make_synack_options (tc, opts);
     case TCP_STATE_SYN_SENT:
-      return tcp_make_syn_options (opts,
-				   tcp_initial_window_to_advertise (tc));
+      return tcp_make_syn_options (opts, tc->rcv_wscale);
     default:
       clib_warning ("Not handled!");
       return 0;
@@ -732,7 +757,7 @@ tcp_send_syn (tcp_connection_t * tc)
 
   /* Make and write options */
   memset (&snd_opts, 0, sizeof (snd_opts));
-  tcp_opts_len = tcp_make_syn_options (&snd_opts, initial_wnd);
+  tcp_opts_len = tcp_make_syn_options (&snd_opts, tc->rcv_wscale);
   tcp_hdr_opts_len = tcp_opts_len + sizeof (tcp_header_t);
 
   th = vlib_buffer_push_tcp (b, tc->c_lcl_port, tc->c_rmt_port, tc->iss,
@@ -900,7 +925,7 @@ tcp_prepare_retransmit_segment (tcp_connection_t * tc, vlib_buffer_t * b,
 
   tcp_reuse_buffer (vm, b);
 
-  ASSERT (tc->state == TCP_STATE_ESTABLISHED);
+  ASSERT (tc->state >= TCP_STATE_ESTABLISHED);
   ASSERT (max_bytes != 0);
 
   if (tcp_opts_sack_permitted (&tc->opt))
@@ -929,7 +954,6 @@ tcp_prepare_retransmit_segment (tcp_connection_t * tc, vlib_buffer_t * b,
 				       max_bytes);
   ASSERT (n_bytes != 0);
 
-  tc->snd_nxt += n_bytes;
   tcp_push_hdr_i (tc, b, tc->state);
 
   return n_bytes;
@@ -967,7 +991,7 @@ tcp_timer_retransmit_handler_i (u32 index, u8 is_syn)
   tcp_get_free_buffer_index (tm, &bi);
   b = vlib_get_buffer (vm, bi);
 
-  if (tc->state == TCP_STATE_ESTABLISHED)
+  if (tc->state >= TCP_STATE_ESTABLISHED)
     {
       tcp_fastrecovery_off (tc);
 
@@ -977,6 +1001,12 @@ tcp_timer_retransmit_handler_i (u32 index, u8 is_syn)
       /* Figure out what and how many bytes we can send */
       snd_space = tcp_available_snd_space (tc);
       max_bytes = clib_min (tc->snd_mss, snd_space);
+
+      if (max_bytes == 0)
+	{
+	  clib_warning ("no wnd to retransmit");
+	  return;
+	}
       tcp_prepare_retransmit_segment (tc, b, max_bytes);
 
       tc->rtx_bytes += max_bytes;
@@ -996,7 +1026,11 @@ tcp_timer_retransmit_handler_i (u32 index, u8 is_syn)
 	tc->rto = clib_min (tc->rto << 1, TCP_RTO_MAX);
 
       vlib_buffer_make_headroom (b, MAX_HDRS_LEN);
+
       tcp_push_hdr_i (tc, b, tc->state);
+
+      /* Account for the SYN */
+      tc->snd_nxt += 1;
     }
 
   if (!is_syn)
@@ -1163,8 +1197,8 @@ tcp46_output_inline (vlib_main_t * vm,
 	  if (PREDICT_FALSE
 	      (vnet_buffer (b0)->tcp.flags & TCP_BUF_FLAG_DUPACK))
 	    {
+	      ASSERT (tc0->snt_dupacks > 0);
 	      tc0->snt_dupacks--;
-	      ASSERT (tc0->snt_dupacks >= 0);
 	      if (!tcp_session_has_ooo_data (tc0))
 		{
 		  error0 = TCP_ERROR_FILTERED_DUPACKS;
-- 
cgit 1.2.3-korg


From 6792ec059696a358b6c98d8d86e9740b34c01e24 Mon Sep 17 00:00:00 2001
From: Florin Coras <fcoras@cisco.com>
Date: Mon, 13 Mar 2017 03:49:51 -0700
Subject: TCP/session improvements

- Added svm fifo flag for tracking fifo dequeue events (replaces event
  length). Updated all code to switch to the new scheme.
- More session debugging
- Fix peek index wrap
- Add a trivial socket test client
- Fast retransmit/cc fixes
- tx and rx SACK fixes and unit testing
- SRTT computation fix
- remove dupack/ack burst filters
- improve ack rx
- improved segment rx
- builtin client test code

Change-Id: Ic4eb2d5ca446eb2260ccd3ccbcdaa73c64e7f4e1
Signed-off-by: Florin Coras <fcoras@cisco.com>
Signed-off-by: Dave Barach <dbarach@cisco.com>
---
 src/svm/svm_fifo.c               |  35 +--
 src/svm/svm_fifo.h               |  28 ++-
 src/svm/svm_fifo_segment.h       |   4 +-
 src/uri.am                       |   5 +-
 src/uri/uri_socket_test.c        | 126 ++++++++++
 src/uri/uri_tcp_test.c           | 161 +++++++++----
 src/uri/uri_udp_test.c           |  13 +-
 src/vnet.am                      |   2 +
 src/vnet/session/application.h   |   3 +-
 src/vnet/session/node.c          | 127 ++++++----
 src/vnet/session/session.c       |  63 +++--
 src/vnet/session/session.h       |  19 +-
 src/vnet/session/session_cli.c   |   2 +-
 src/vnet/session/session_debug.h |  38 ++-
 src/vnet/session/transport.h     |   2 +-
 src/vnet/tcp/builtin_client.c    | 411 +++++++++++++++++++++++++++++++
 src/vnet/tcp/builtin_client.h    | 131 ++++++++++
 src/vnet/tcp/builtin_server.c    |  91 +++++--
 src/vnet/tcp/tcp.c               |  37 ++-
 src/vnet/tcp/tcp.h               | 111 +++++++--
 src/vnet/tcp/tcp_debug.h         | 252 ++++++++++++++++---
 src/vnet/tcp/tcp_error.def       |   7 +-
 src/vnet/tcp/tcp_input.c         | 507 +++++++++++++++++++++++++--------------
 src/vnet/tcp/tcp_output.c        | 295 ++++++++++++++++-------
 src/vnet/tcp/tcp_packet.h        |   2 +-
 src/vnet/tcp/tcp_test.c          | 216 +++++++++++++++++
 src/vnet/udp/builtin_server.c    |  29 ++-
 src/vnet/udp/udp_input.c         |  47 ++--
 28 files changed, 2201 insertions(+), 563 deletions(-)
 create mode 100644 src/uri/uri_socket_test.c
 create mode 100644 src/vnet/tcp/builtin_client.c
 create mode 100644 src/vnet/tcp/builtin_client.h
 create mode 100644 src/vnet/tcp/tcp_test.c

(limited to 'src/svm/svm_fifo.c')

diff --git a/src/svm/svm_fifo.c b/src/svm/svm_fifo.c
index e3f534b1..07b0d2df 100644
--- a/src/svm/svm_fifo.c
+++ b/src/svm/svm_fifo.c
@@ -13,7 +13,7 @@
  * limitations under the License.
  */
 
-#include "svm_fifo.h"
+#include <svm/svm_fifo.h>
 
 /** create an svm fifo, in the current heap. Fails vs blow up the process */
 svm_fifo_t *
@@ -362,18 +362,19 @@ svm_fifo_enqueue_nowait (svm_fifo_t * f,
   return svm_fifo_enqueue_internal (f, pid, max_bytes, copy_from_here);
 }
 
-/** Enqueue a future segment.
+/**
+ * Enqueue a future segment.
+ *
  * Two choices: either copies the entire segment, or copies nothing
  * Returns 0 of the entire segment was copied
  * Returns -1 if none of the segment was copied due to lack of space
  */
-
 static int
-svm_fifo_enqueue_with_offset_internal2 (svm_fifo_t * f,
-					int pid,
-					u32 offset,
-					u32 required_bytes,
-					u8 * copy_from_here)
+svm_fifo_enqueue_with_offset_internal (svm_fifo_t * f,
+				       int pid,
+				       u32 offset,
+				       u32 required_bytes,
+				       u8 * copy_from_here)
 {
   u32 total_copy_bytes, first_copy_bytes, second_copy_bytes;
   u32 cursize, nitems;
@@ -424,14 +425,14 @@ svm_fifo_enqueue_with_offset (svm_fifo_t * f,
 			      u32 offset,
 			      u32 required_bytes, u8 * copy_from_here)
 {
-  return svm_fifo_enqueue_with_offset_internal2
+  return svm_fifo_enqueue_with_offset_internal
     (f, pid, offset, required_bytes, copy_from_here);
 }
 
 
 static int
-svm_fifo_dequeue_internal2 (svm_fifo_t * f,
-			    int pid, u32 max_bytes, u8 * copy_here)
+svm_fifo_dequeue_internal (svm_fifo_t * f,
+			   int pid, u32 max_bytes, u8 * copy_here)
 {
   u32 total_copy_bytes, first_copy_bytes, second_copy_bytes;
   u32 cursize, nitems;
@@ -484,7 +485,7 @@ int
 svm_fifo_dequeue_nowait (svm_fifo_t * f,
 			 int pid, u32 max_bytes, u8 * copy_here)
 {
-  return svm_fifo_dequeue_internal2 (f, pid, max_bytes, copy_here);
+  return svm_fifo_dequeue_internal (f, pid, max_bytes, copy_here);
 }
 
 int
@@ -492,7 +493,7 @@ svm_fifo_peek (svm_fifo_t * f, int pid, u32 offset, u32 max_bytes,
 	       u8 * copy_here)
 {
   u32 total_copy_bytes, first_copy_bytes, second_copy_bytes;
-  u32 cursize, nitems;
+  u32 cursize, nitems, real_head;
 
   if (PREDICT_FALSE (f->cursize == 0))
     return -2;			/* nothing in the fifo */
@@ -500,6 +501,8 @@ svm_fifo_peek (svm_fifo_t * f, int pid, u32 offset, u32 max_bytes,
   /* read cursize, which can only increase while we're working */
   cursize = f->cursize;
   nitems = f->nitems;
+  real_head = f->head + offset;
+  real_head = real_head >= nitems ? real_head - nitems : real_head;
 
   /* Number of bytes we're going to copy */
   total_copy_bytes = (cursize < max_bytes) ? cursize : max_bytes;
@@ -508,9 +511,9 @@ svm_fifo_peek (svm_fifo_t * f, int pid, u32 offset, u32 max_bytes,
     {
       /* Number of bytes in first copy segment */
       first_copy_bytes =
-	((nitems - f->head + offset) < total_copy_bytes) ?
-	(nitems - f->head + offset) : total_copy_bytes;
-      clib_memcpy (copy_here, &f->data[f->head + offset], first_copy_bytes);
+	((nitems - real_head) < total_copy_bytes) ?
+	(nitems - real_head) : total_copy_bytes;
+      clib_memcpy (copy_here, &f->data[real_head], first_copy_bytes);
 
       /* Number of bytes in second copy segment, if any */
       second_copy_bytes = total_copy_bytes - first_copy_bytes;
diff --git a/src/svm/svm_fifo.h b/src/svm/svm_fifo.h
index 70624b74..39556173 100644
--- a/src/svm/svm_fifo.h
+++ b/src/svm/svm_fifo.h
@@ -46,9 +46,11 @@ typedef struct
 {
   pthread_mutex_t mutex;	/* 8 bytes */
   pthread_cond_t condvar;	/* 8 bytes */
-  u32 owner_pid;
   svm_lock_tag_t tag;
-  volatile u32 cursize;
+
+  volatile u32 cursize;		/**< current fifo size */
+  volatile u8 has_event;	/**< non-zero if deq event exists */
+  u32 owner_pid;
   u32 nitems;
 
   /* Backpointers */
@@ -112,6 +114,28 @@ svm_fifo_has_ooo_data (svm_fifo_t * f)
   return f->ooos_list_head != OOO_SEGMENT_INVALID_INDEX;
 }
 
+/**
+ * Sets fifo event flag.
+ *
+ * @return 1 if flag was not set.
+ */
+always_inline u8
+svm_fifo_set_event (svm_fifo_t * f)
+{
+  /* Probably doesn't need to be atomic. Still, better avoid surprises */
+  return __sync_lock_test_and_set (&f->has_event, 1) == 0;
+}
+
+/**
+ * Unsets fifo event flag.
+ */
+always_inline void
+svm_fifo_unset_event (svm_fifo_t * f)
+{
+  /* Probably doesn't need to be atomic. Still, better avoid surprises */
+  __sync_lock_test_and_set (&f->has_event, 0);
+}
+
 svm_fifo_t *svm_fifo_create (u32 data_size_in_bytes);
 
 int svm_fifo_enqueue_nowait (svm_fifo_t * f, int pid, u32 max_bytes,
diff --git a/src/svm/svm_fifo_segment.h b/src/svm/svm_fifo_segment.h
index 793fa7c8..ecb5653a 100644
--- a/src/svm/svm_fifo_segment.h
+++ b/src/svm/svm_fifo_segment.h
@@ -15,8 +15,8 @@
 #ifndef __included_ssvm_fifo_segment_h__
 #define __included_ssvm_fifo_segment_h__
 
-#include "svm_fifo.h"
-#include "ssvm.h"
+#include <svm/svm_fifo.h>
+#include <svm/ssvm.h>
 
 typedef struct
 {
diff --git a/src/uri.am b/src/uri.am
index 09b5b15b..ad4d65d8 100644
--- a/src/uri.am
+++ b/src/uri.am
@@ -11,7 +11,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-noinst_PROGRAMS += uri_udp_test uri_tcp_test
+noinst_PROGRAMS += uri_udp_test uri_tcp_test uri_socket_test
 
 uri_udp_test_SOURCES = uri/uri_udp_test.c
 uri_udp_test_LDADD = libvlibmemoryclient.la libvlibapi.la libsvm.la \
@@ -20,3 +20,6 @@ uri_udp_test_LDADD = libvlibmemoryclient.la libvlibapi.la libsvm.la \
 uri_tcp_test_SOURCES = uri/uri_tcp_test.c
 uri_tcp_test_LDADD = libvlibmemoryclient.la libvlibapi.la libsvm.la \
 	libvppinfra.la -lpthread -lm -lrt
+
+uri_socket_test_SOURCES = uri/uri_socket_test.c
+uri_socket_test_LDADD = libvppinfra.la -lpthread -lm -lrt
diff --git a/src/uri/uri_socket_test.c b/src/uri/uri_socket_test.c
new file mode 100644
index 00000000..9f049bda
--- /dev/null
+++ b/src/uri/uri_socket_test.c
@@ -0,0 +1,126 @@
+/*
+ * Copyright (c) 2017 Cisco and/or its affiliates.
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at:
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include <stdio.h>
+#include <sys/types.h>
+#include <sys/socket.h>
+#include <netinet/in.h>
+#include <netdb.h>
+#include <vppinfra/format.h>
+
+int
+main (int argc, char *argv[])
+{
+  int sockfd, portno, n;
+  struct sockaddr_in serv_addr;
+  struct hostent *server;
+  u8 *rx_buffer = 0, *tx_buffer = 0;
+  u32 offset;
+  int iter, i;
+  if (0 && argc < 3)
+    {
+      fformat (stderr, "usage %s hostname port\n", argv[0]);
+      exit (0);
+    }
+
+  portno = 1234;		// atoi(argv[2]);
+  sockfd = socket (AF_INET, SOCK_STREAM, 0);
+  if (sockfd < 0)
+    {
+      clib_unix_error ("socket");
+      exit (1);
+    }
+  server = gethostbyname ("6.0.1.1" /* argv[1] */ );
+  if (server == NULL)
+    {
+      clib_unix_warning ("gethostbyname");
+      exit (1);
+    }
+  bzero ((char *) &serv_addr, sizeof (serv_addr));
+  serv_addr.sin_family = AF_INET;
+  bcopy ((char *) server->h_addr,
+	 (char *) &serv_addr.sin_addr.s_addr, server->h_length);
+  serv_addr.sin_port = htons (portno);
+  if (connect (sockfd, (const void *) &serv_addr, sizeof (serv_addr)) < 0)
+    {
+      clib_unix_warning ("connect");
+      exit (1);
+    }
+
+  vec_validate (rx_buffer, 1400);
+  vec_validate (tx_buffer, 1400);
+
+  for (i = 0; i < vec_len (tx_buffer); i++)
+    tx_buffer[i] = (i + 1) % 0xff;
+
+  /*
+   * Send one packet to warm up the RX pipeline
+   */
+  n = send (sockfd, tx_buffer, vec_len (tx_buffer), 0 /* flags */ );
+  if (n != vec_len (tx_buffer))
+    {
+      clib_unix_warning ("write");
+      exit (0);
+    }
+
+  for (iter = 0; iter < 100000; iter++)
+    {
+      if (iter < 99999)
+	{
+	  n = send (sockfd, tx_buffer, vec_len (tx_buffer), 0 /* flags */ );
+	  if (n != vec_len (tx_buffer))
+	    {
+	      clib_unix_warning ("write");
+	      exit (0);
+	    }
+	}
+      offset = 0;
+
+      do
+	{
+	  n = recv (sockfd, rx_buffer + offset,
+		    vec_len (rx_buffer) - offset, 0 /* flags */ );
+	  if (n < 0)
+	    {
+	      clib_unix_warning ("read");
+	      exit (0);
+	    }
+	  offset += n;
+	}
+      while (offset < vec_len (rx_buffer));
+
+      for (i = 0; i < vec_len (rx_buffer); i++)
+	{
+	  if (rx_buffer[i] != tx_buffer[i])
+	    {
+	      clib_warning ("[%d] read 0x%x not 0x%x",
+			    rx_buffer[i], tx_buffer[i]);
+	      exit (1);
+	    }
+	}
+
+    }
+  close (sockfd);
+  return 0;
+}
+
+
+/*
+ * fd.io coding-style-patch-verification: ON
+ *
+ * Local Variables:
+ * eval: (c-set-style "gnu")
+ * End:
+ */
diff --git a/src/uri/uri_tcp_test.c b/src/uri/uri_tcp_test.c
index 406a5f4e..e2834817 100644
--- a/src/uri/uri_tcp_test.c
+++ b/src/uri/uri_tcp_test.c
@@ -116,6 +116,7 @@ typedef struct
   pthread_t client_rx_thread_handle;
   u32 client_bytes_received;
   u8 test_return_packets;
+  u32 bytes_to_send;
 
   /* convenience */
   svm_fifo_segment_main_t *segment_main;
@@ -313,11 +314,16 @@ client_handle_fifo_event_rx (uri_tcp_test_main_t * utm,
 
   rx_fifo = e->fifo;
 
-  bytes = e->enqueue_length;
+  bytes = svm_fifo_max_dequeue (rx_fifo);
+  /* Allow enqueuing of new event */
+  svm_fifo_unset_event (rx_fifo);
+
+  /* Read the bytes */
   do
     {
-      n_read = svm_fifo_dequeue_nowait (rx_fifo, 0, vec_len (utm->rx_buf),
-					utm->rx_buf);
+      n_read = svm_fifo_dequeue_nowait (rx_fifo, 0,
+					clib_min (vec_len (utm->rx_buf),
+						  bytes), utm->rx_buf);
       if (n_read > 0)
 	{
 	  bytes -= n_read;
@@ -333,9 +339,17 @@ client_handle_fifo_event_rx (uri_tcp_test_main_t * utm,
 	    }
 	  utm->client_bytes_received += n_read;
 	}
+      else
+	{
+	  if (n_read == -2)
+	    {
+	      clib_warning ("weird!");
+	      break;
+	    }
+	}
 
     }
-  while (n_read < 0 || bytes > 0);
+  while (bytes > 0);
 }
 
 void
@@ -479,47 +493,41 @@ vl_api_connect_uri_reply_t_handler (vl_api_connect_uri_reply_t * mp)
     }
 }
 
-void
-client_send_data (uri_tcp_test_main_t * utm)
+static void
+send_test_chunk (uri_tcp_test_main_t * utm, svm_fifo_t * tx_fifo, int mypid,
+		 u32 bytes)
 {
   u8 *test_data = utm->connect_test_data;
   u64 bytes_sent = 0;
-  int rv;
-  int mypid = getpid ();
-  session_t *session;
-  svm_fifo_t *tx_fifo;
-  int buffer_offset, bytes_to_send = 0;
+  int test_buf_offset = 0;
+  u32 bytes_to_snd;
+  u32 queue_max_chunk = 64 << 10, actual_write;
   session_fifo_event_t evt;
   static int serial_number = 0;
-  int i;
-  u32 max_chunk = 64 << 10, write;
-
-  session = pool_elt_at_index (utm->sessions, utm->connected_session_index);
-  tx_fifo = session->server_tx_fifo;
+  int rv;
 
-  vec_validate (utm->rx_buf, vec_len (test_data) - 1);
+  bytes_to_snd = (bytes == 0) ? vec_len (test_data) : bytes;
+  if (bytes_to_snd > vec_len (test_data))
+    bytes_to_snd = vec_len (test_data);
 
-  for (i = 0; i < 1; i++)
+  while (bytes_to_snd > 0)
     {
-      bytes_to_send = vec_len (test_data);
-      buffer_offset = 0;
-      while (bytes_to_send > 0)
+      actual_write =
+	bytes_to_snd > queue_max_chunk ? queue_max_chunk : bytes_to_snd;
+      rv = svm_fifo_enqueue_nowait (tx_fifo, mypid, actual_write,
+				    test_data + test_buf_offset);
+
+      if (rv > 0)
 	{
-	  write = bytes_to_send > max_chunk ? max_chunk : bytes_to_send;
-	  rv = svm_fifo_enqueue_nowait (tx_fifo, mypid, write,
-					test_data + buffer_offset);
+	  bytes_to_snd -= rv;
+	  test_buf_offset += rv;
+	  bytes_sent += rv;
 
-	  if (rv > 0)
+	  if (svm_fifo_set_event (tx_fifo))
 	    {
-	      bytes_to_send -= rv;
-	      buffer_offset += rv;
-	      bytes_sent += rv;
-
 	      /* Fabricate TX event, send to vpp */
 	      evt.fifo = tx_fifo;
 	      evt.event_type = FIFO_EVENT_SERVER_TX;
-	      /* $$$$ for event logging */
-	      evt.enqueue_length = rv;
 	      evt.event_id = serial_number++;
 
 	      unix_shared_memory_queue_add (utm->vpp_event_queue,
@@ -528,13 +536,40 @@ client_send_data (uri_tcp_test_main_t * utm)
 	    }
 	}
     }
+}
+
+void
+client_send_data (uri_tcp_test_main_t * utm)
+{
+  u8 *test_data = utm->connect_test_data;
+  int mypid = getpid ();
+  session_t *session;
+  svm_fifo_t *tx_fifo;
+  u32 n_iterations, leftover;
+  int i;
+
+  session = pool_elt_at_index (utm->sessions, utm->connected_session_index);
+  tx_fifo = session->server_tx_fifo;
+
+  vec_validate (utm->rx_buf, vec_len (test_data) - 1);
+  n_iterations = utm->bytes_to_send / vec_len (test_data);
+
+  for (i = 0; i < n_iterations; i++)
+    {
+      send_test_chunk (utm, tx_fifo, mypid, 0);
+    }
+
+  leftover = utm->bytes_to_send % vec_len (test_data);
+  if (leftover)
+    send_test_chunk (utm, tx_fifo, mypid, leftover);
 
   if (utm->test_return_packets)
     {
       f64 timeout = clib_time_now (&utm->clib_time) + 2;
 
       /* Wait for the outstanding packets */
-      while (utm->client_bytes_received < vec_len (test_data))
+      while (utm->client_bytes_received <
+	     vec_len (test_data) * n_iterations + leftover)
 	{
 	  if (clib_time_now (&utm->clib_time) > timeout)
 	    {
@@ -542,9 +577,8 @@ client_send_data (uri_tcp_test_main_t * utm)
 	      break;
 	    }
 	}
-
-      utm->time_to_stop = 1;
     }
+  utm->time_to_stop = 1;
 }
 
 void
@@ -599,6 +633,11 @@ client_test (uri_tcp_test_main_t * utm)
 
   /* Disconnect */
   client_disconnect (utm);
+
+  if (wait_for_state_change (utm, STATE_START))
+    {
+      return;
+    }
 }
 
 static void
@@ -714,7 +753,6 @@ server_handle_fifo_event_rx (uri_tcp_test_main_t * utm,
 {
   svm_fifo_t *rx_fifo, *tx_fifo;
   int n_read;
-
   session_fifo_event_t evt;
   unix_shared_memory_queue_t *q;
   int rv, bytes;
@@ -722,34 +760,46 @@ server_handle_fifo_event_rx (uri_tcp_test_main_t * utm,
   rx_fifo = e->fifo;
   tx_fifo = utm->sessions[rx_fifo->client_session_index].server_tx_fifo;
 
-  bytes = e->enqueue_length;
+  bytes = svm_fifo_max_dequeue (rx_fifo);
+  /* Allow enqueuing of a new event */
+  svm_fifo_unset_event (rx_fifo);
+
+  if (bytes == 0)
+    return;
+
+  /* Read the bytes */
   do
     {
       n_read = svm_fifo_dequeue_nowait (rx_fifo, 0, vec_len (utm->rx_buf),
 					utm->rx_buf);
+      if (n_read > 0)
+	bytes -= n_read;
+
+      if (utm->drop_packets)
+	continue;
 
       /* Reflect if a non-drop session */
-      if (!utm->drop_packets && n_read > 0)
+      if (n_read > 0)
 	{
 	  do
 	    {
 	      rv = svm_fifo_enqueue_nowait (tx_fifo, 0, n_read, utm->rx_buf);
 	    }
-	  while (rv == -2 && !utm->time_to_stop);
-
-	  /* Fabricate TX event, send to vpp */
-	  evt.fifo = tx_fifo;
-	  evt.event_type = FIFO_EVENT_SERVER_TX;
-	  /* $$$$ for event logging */
-	  evt.enqueue_length = n_read;
-	  evt.event_id = e->event_id;
-	  q = utm->vpp_event_queue;
-	  unix_shared_memory_queue_add (q, (u8 *) & evt,
-					0 /* do wait for mutex */ );
-	}
+	  while (rv <= 0 && !utm->time_to_stop);
 
-      if (n_read > 0)
-	bytes -= n_read;
+	  /* If event wasn't set, add one */
+	  if (svm_fifo_set_event (tx_fifo))
+	    {
+	      /* Fabricate TX event, send to vpp */
+	      evt.fifo = tx_fifo;
+	      evt.event_type = FIFO_EVENT_SERVER_TX;
+	      evt.event_id = e->event_id;
+
+	      q = utm->vpp_event_queue;
+	      unix_shared_memory_queue_add (q, (u8 *) & evt,
+					    0 /* do wait for mutex */ );
+	    }
+	}
     }
   while ((n_read < 0 || bytes > 0) && !utm->time_to_stop);
 }
@@ -852,7 +902,10 @@ static void
 vl_api_disconnect_session_reply_t_handler (vl_api_disconnect_session_reply_t *
 					   mp)
 {
+  uri_tcp_test_main_t *utm = &uri_tcp_test_main;
+
   clib_warning ("retval %d", ntohl (mp->retval));
+  utm->state = STATE_START;
 }
 
 #define foreach_uri_msg                                 \
@@ -888,6 +941,7 @@ main (int argc, char **argv)
   u8 *heap, *uri = 0;
   u8 *bind_uri = (u8 *) "tcp://0.0.0.0/1234";
   u8 *connect_uri = (u8 *) "tcp://6.0.1.2/1234";
+  u32 bytes_to_send = 64 << 10, mbytes;
   u32 tmp;
   mheap_t *h;
   session_t *session;
@@ -934,6 +988,10 @@ main (int argc, char **argv)
 	drop_packets = 1;
       else if (unformat (a, "test"))
 	test_return_packets = 1;
+      else if (unformat (a, "mbytes %d", &mbytes))
+	{
+	  bytes_to_send = mbytes << 20;
+	}
       else
 	{
 	  fformat (stderr, "%s: usage [master|slave]\n");
@@ -956,6 +1014,7 @@ main (int argc, char **argv)
   utm->segment_main = &svm_fifo_segment_main;
   utm->drop_packets = drop_packets;
   utm->test_return_packets = test_return_packets;
+  utm->bytes_to_send = bytes_to_send;
 
   setup_signal_handlers ();
   uri_api_hookup (utm);
diff --git a/src/uri/uri_udp_test.c b/src/uri/uri_udp_test.c
index 54625d64..e6c239c1 100644
--- a/src/uri/uri_udp_test.c
+++ b/src/uri/uri_udp_test.c
@@ -742,17 +742,20 @@ server_handle_fifo_event_rx (uri_udp_test_main_t * utm,
   /* Fabricate TX event, send to vpp */
   evt.fifo = tx_fifo;
   evt.event_type = FIFO_EVENT_SERVER_TX;
-  /* $$$$ for event logging */
-  evt.enqueue_length = nbytes;
   evt.event_id = e->event_id;
-  q = utm->vpp_event_queue;
-  unix_shared_memory_queue_add (q, (u8 *) & evt, 0 /* do wait for mutex */ );
+
+  if (svm_fifo_set_event (tx_fifo))
+    {
+      q = utm->vpp_event_queue;
+      unix_shared_memory_queue_add (q, (u8 *) & evt,
+				    0 /* do wait for mutex */ );
+    }
 }
 
 void
 server_handle_event_queue (uri_udp_test_main_t * utm)
 {
-  session_fifo_event_t _e, *e = &_e;;
+  session_fifo_event_t _e, *e = &_e;
 
   while (1)
     {
diff --git a/src/vnet.am b/src/vnet.am
index 3e73de8f..9c55e336 100644
--- a/src/vnet.am
+++ b/src/vnet.am
@@ -462,7 +462,9 @@ libvnet_la_SOURCES +=				\
  vnet/tcp/tcp_output.c				\
  vnet/tcp/tcp_input.c				\
  vnet/tcp/tcp_newreno.c				\
+ vnet/tcp/builtin_client.c			\
  vnet/tcp/builtin_server.c			\
+ vnet/tcp/tcp_test.c				\
  vnet/tcp/tcp.c
 
 nobase_include_HEADERS +=			\
diff --git a/src/vnet/session/application.h b/src/vnet/session/application.h
index a60a8b8b..480828f7 100644
--- a/src/vnet/session/application.h
+++ b/src/vnet/session/application.h
@@ -45,8 +45,7 @@ typedef struct _stream_session_cb_vft
   void (*session_reset_callback) (stream_session_t * s);
 
   /* Direct RX callback, for built-in servers */
-  int (*builtin_server_rx_callback) (stream_session_t * session,
-				     session_fifo_event_t * ep);
+  int (*builtin_server_rx_callback) (stream_session_t * session);
 
   /* Redirect connection to local server */
   int (*redirect_connect_callback) (u32 api_client_index, void *mp);
diff --git a/src/vnet/session/node.c b/src/vnet/session/node.c
index 822afebd..8681105c 100644
--- a/src/vnet/session/node.c
+++ b/src/vnet/session/node.c
@@ -13,21 +13,14 @@
  * limitations under the License.
  */
 
+#include <math.h>
 #include <vlib/vlib.h>
 #include <vnet/vnet.h>
-#include <vnet/pg/pg.h>
-#include <vnet/ip/ip.h>
-
 #include <vnet/tcp/tcp.h>
-
-#include <vppinfra/hash.h>
-#include <vppinfra/error.h>
 #include <vppinfra/elog.h>
-#include <vlibmemory/unix_shared_memory_queue.h>
-
-#include <vnet/udp/udp_packet.h>
-#include <math.h>
+#include <vnet/session/application.h>
 #include <vnet/session/session_debug.h>
+#include <vlibmemory/unix_shared_memory_queue.h>
 
 vlib_node_registration_t session_queue_node;
 
@@ -52,8 +45,8 @@ format_session_queue_trace (u8 * s, va_list * args)
 
 vlib_node_registration_t session_queue_node;
 
-#define foreach_session_queue_error                 \
-_(TX, "Packets transmitted")                    \
+#define foreach_session_queue_error		\
+_(TX, "Packets transmitted")                  	\
 _(TIMER, "Timer events")
 
 typedef enum
@@ -91,10 +84,10 @@ session_tx_fifo_read_and_snd_i (vlib_main_t * vm, vlib_node_runtime_t * node,
   transport_proto_vft_t *transport_vft;
   u32 next_index, next0, *to_next, n_left_to_next, bi0;
   vlib_buffer_t *b0;
-  u32 rx_offset;
+  u32 rx_offset = 0, max_dequeue0;
   u16 snd_mss0;
   u8 *data0;
-  int i;
+  int i, n_bytes_read;
 
   next_index = next0 = session_type_to_next[s0->session_type];
 
@@ -106,24 +99,33 @@ session_tx_fifo_read_and_snd_i (vlib_main_t * vm, vlib_node_runtime_t * node,
   snd_mss0 = transport_vft->send_mss (tc0);
 
   /* Can't make any progress */
-  if (snd_space0 == 0 || svm_fifo_max_dequeue (s0->server_tx_fifo) == 0
-      || snd_mss0 == 0)
+  if (snd_space0 == 0 || snd_mss0 == 0)
     {
       vec_add1 (smm->evts_partially_read[thread_index], *e0);
       return 0;
     }
 
-  ASSERT (e0->enqueue_length > 0);
-
-  /* Ensure we're not writing more than transport window allows */
-  max_len_to_snd0 = clib_min (e0->enqueue_length, snd_space0);
-
   if (peek_data)
     {
       /* Offset in rx fifo from where to peek data  */
       rx_offset = transport_vft->tx_fifo_offset (tc0);
     }
 
+  /* Check how much we can pull. If buffering, subtract the offset */
+  max_dequeue0 = svm_fifo_max_dequeue (s0->server_tx_fifo) - rx_offset;
+
+  /* Allow enqueuing of a new event */
+  svm_fifo_unset_event (s0->server_tx_fifo);
+
+  /* Nothing to read return */
+  if (max_dequeue0 == 0)
+    {
+      return 0;
+    }
+
+  /* Ensure we're not writing more than transport window allows */
+  max_len_to_snd0 = clib_min (max_dequeue0, snd_space0);
+
   /* TODO check if transport is willing to send len_to_snd0
    * bytes (Nagle) */
 
@@ -147,13 +149,10 @@ session_tx_fifo_read_and_snd_i (vlib_main_t * vm, vlib_node_runtime_t * node,
 	   * XXX 0.9 because when debugging we might not get a full frame */
 	  if (PREDICT_FALSE (n_bufs < 0.9 * VLIB_FRAME_SIZE))
 	    {
-	      /* Keep track of how much we've dequeued and exit */
-	      if (left_to_snd0 != max_len_to_snd0)
+	      if (svm_fifo_set_event (s0->server_tx_fifo))
 		{
-		  e0->enqueue_length -= max_len_to_snd0 - left_to_snd0;
 		  vec_add1 (smm->evts_partially_read[thread_index], *e0);
 		}
-
 	      return -1;
 	    }
 
@@ -198,9 +197,9 @@ session_tx_fifo_read_and_snd_i (vlib_main_t * vm, vlib_node_runtime_t * node,
 	  len_to_deq0 = (left_to_snd0 < snd_mss0) ? left_to_snd0 : snd_mss0;
 
 	  /* *INDENT-OFF* */
-	  SESSION_EVT_DBG(s0, SESSION_EVT_DEQ, ({
+	  SESSION_EVT_DBG(SESSION_EVT_DEQ, s0, ({
 	      ed->data[0] = e0->event_id;
-	      ed->data[1] = e0->enqueue_length;
+	      ed->data[1] = max_dequeue0;
 	      ed->data[2] = len_to_deq0;
 	      ed->data[3] = left_to_snd0;
 	  }));
@@ -214,29 +213,30 @@ session_tx_fifo_read_and_snd_i (vlib_main_t * vm, vlib_node_runtime_t * node,
 	   *      2) buffer chains */
 	  if (peek_data)
 	    {
-	      int n_bytes_read;
 	      n_bytes_read = svm_fifo_peek (s0->server_tx_fifo, s0->pid,
 					    rx_offset, len_to_deq0, data0);
-	      if (n_bytes_read < 0)
+	      if (n_bytes_read <= 0)
 		goto dequeue_fail;
 
 	      /* Keep track of progress locally, transport is also supposed to
-	       * increment it independently when pushing header */
+	       * increment it independently when pushing the header */
 	      rx_offset += n_bytes_read;
 	    }
 	  else
 	    {
-	      if (svm_fifo_dequeue_nowait (s0->server_tx_fifo, s0->pid,
-					   len_to_deq0, data0) < 0)
+	      n_bytes_read = svm_fifo_dequeue_nowait (s0->server_tx_fifo,
+						      s0->pid, len_to_deq0,
+						      data0);
+	      if (n_bytes_read <= 0)
 		goto dequeue_fail;
 	    }
 
-	  b0->current_length = len_to_deq0;
+	  b0->current_length = n_bytes_read;
 
 	  /* Ask transport to push header */
 	  transport_vft->push_header (tc0, b0);
 
-	  left_to_snd0 -= len_to_deq0;
+	  left_to_snd0 -= n_bytes_read;
 	  *n_tx_packets = *n_tx_packets + 1;
 
 	  vlib_validate_buffer_enqueue_x1 (vm, node, next_index,
@@ -246,25 +246,31 @@ session_tx_fifo_read_and_snd_i (vlib_main_t * vm, vlib_node_runtime_t * node,
       vlib_put_next_frame (vm, node, next_index, n_left_to_next);
     }
 
-  /* If we couldn't dequeue all bytes store progress */
-  if (max_len_to_snd0 < e0->enqueue_length)
+  /* If we couldn't dequeue all bytes mark as partially read */
+  if (max_len_to_snd0 < max_dequeue0)
     {
-      e0->enqueue_length -= max_len_to_snd0;
-      vec_add1 (smm->evts_partially_read[thread_index], *e0);
+      /* If we don't already have new event */
+      if (svm_fifo_set_event (s0->server_tx_fifo))
+	{
+	  vec_add1 (smm->evts_partially_read[thread_index], *e0);
+	}
     }
   return 0;
 
 dequeue_fail:
-  /* Can't read from fifo. Store event rx progress, save as partially read,
-   * return buff to free list and return  */
-  e0->enqueue_length -= max_len_to_snd0 - left_to_snd0;
-  vec_add1 (smm->evts_partially_read[thread_index], *e0);
+  /*
+   * Can't read from fifo. If we don't already have an event, save as partially
+   * read, return buff to free list and return
+   */
+  clib_warning ("dequeue fail");
 
-  to_next -= 1;
-  n_left_to_next += 1;
+  if (svm_fifo_set_event (s0->server_tx_fifo))
+    {
+      vec_add1 (smm->evts_partially_read[thread_index], *e0);
+    }
+  vlib_put_next_frame (vm, node, next_index, n_left_to_next + 1);
   _vec_len (smm->tx_buffers[thread_index]) += 1;
 
-  clib_warning ("dequeue fail");
   return 0;
 }
 
@@ -298,6 +304,7 @@ session_queue_node_fn (vlib_main_t * vm, vlib_node_runtime_t * node,
   session_fifo_event_t *my_fifo_events, *e;
   u32 n_to_dequeue, n_events;
   unix_shared_memory_queue_t *q;
+  application_t *app;
   int n_tx_packets = 0;
   u32 my_thread_index = vm->cpu_index;
   int i, rv;
@@ -321,13 +328,18 @@ session_queue_node_fn (vlib_main_t * vm, vlib_node_runtime_t * node,
   if (n_to_dequeue == 0 && vec_len (my_fifo_events) == 0)
     return 0;
 
+  SESSION_EVT_DBG (SESSION_EVT_DEQ_NODE, 0);
+
   /*
    * If we didn't manage to process previous events try going
    * over them again without dequeuing new ones.
    */
   /* XXX: Block senders to sessions that can't keep up */
   if (vec_len (my_fifo_events) >= 100)
-    goto skip_dequeue;
+    {
+      clib_warning ("too many fifo events unsolved");
+      goto skip_dequeue;
+    }
 
   /* See you in the next life, don't be late */
   if (pthread_mutex_trylock (&q->mutex))
@@ -352,19 +364,17 @@ skip_dequeue:
     {
       svm_fifo_t *f0;		/* $$$ prefetch 1 ahead maybe */
       stream_session_t *s0;
-      u32 server_session_index0, server_thread_index0;
+      u32 session_index0;
       session_fifo_event_t *e0;
 
       e0 = &my_fifo_events[i];
       f0 = e0->fifo;
-      server_session_index0 = f0->server_session_index;
-      server_thread_index0 = f0->server_thread_index;
+      session_index0 = f0->server_session_index;
 
       /* $$$ add multiple event queues, per vpp worker thread */
-      ASSERT (server_thread_index0 == my_thread_index);
+      ASSERT (f0->server_thread_index == my_thread_index);
 
-      s0 = stream_session_get_if_valid (server_session_index0,
-					my_thread_index);
+      s0 = stream_session_get_if_valid (session_index0, my_thread_index);
 
       if (CLIB_DEBUG && !s0)
 	{
@@ -385,11 +395,20 @@ skip_dequeue:
 	  rv = (smm->session_tx_fns[s0->session_type]) (vm, node, smm, e0, s0,
 							my_thread_index,
 							&n_tx_packets);
+	  /* Out of buffers */
 	  if (rv < 0)
 	    goto done;
 
 	  break;
-
+	case FIFO_EVENT_SERVER_EXIT:
+	  stream_session_disconnect (s0);
+	  break;
+	case FIFO_EVENT_BUILTIN_RX:
+	  svm_fifo_unset_event (s0->server_rx_fifo);
+	  /* Get session's server */
+	  app = application_get (s0->app_index);
+	  app->cb_fns.builtin_server_rx_callback (s0);
+	  break;
 	default:
 	  clib_warning ("unhandled event type %d", e0->event_type);
 	}
@@ -418,6 +437,8 @@ done:
   vlib_node_increment_counter (vm, session_queue_node.index,
 			       SESSION_QUEUE_ERROR_TX, n_tx_packets);
 
+  SESSION_EVT_DBG (SESSION_EVT_DEQ_NODE, 1);
+
   return n_tx_packets;
 }
 
diff --git a/src/vnet/session/session.c b/src/vnet/session/session.c
index 06e2a09a..f10918aa 100644
--- a/src/vnet/session/session.c
+++ b/src/vnet/session/session.c
@@ -804,30 +804,36 @@ stream_session_enqueue_notify (stream_session_t * s, u8 block)
   /* Get session's server */
   app = application_get (s->app_index);
 
-  /* Fabricate event */
-  evt.fifo = s->server_rx_fifo;
-  evt.event_type = FIFO_EVENT_SERVER_RX;
-  evt.event_id = serial_number++;
-  evt.enqueue_length = svm_fifo_max_dequeue (s->server_rx_fifo);
-
   /* Built-in server? Hand event to the callback... */
   if (app->cb_fns.builtin_server_rx_callback)
-    return app->cb_fns.builtin_server_rx_callback (s, &evt);
-
-  /* Add event to server's event queue */
-  q = app->event_queue;
+    return app->cb_fns.builtin_server_rx_callback (s);
 
-  /* Based on request block (or not) for lack of space */
-  if (block || PREDICT_TRUE (q->cursize < q->maxsize))
-    unix_shared_memory_queue_add (app->event_queue, (u8 *) & evt,
-				  0 /* do wait for mutex */ );
-  else
-    return -1;
+  /* If no event, send one */
+  if (svm_fifo_set_event (s->server_rx_fifo))
+    {
+      /* Fabricate event */
+      evt.fifo = s->server_rx_fifo;
+      evt.event_type = FIFO_EVENT_SERVER_RX;
+      evt.event_id = serial_number++;
+
+      /* Add event to server's event queue */
+      q = app->event_queue;
+
+      /* Based on request block (or not) for lack of space */
+      if (block || PREDICT_TRUE (q->cursize < q->maxsize))
+	unix_shared_memory_queue_add (app->event_queue, (u8 *) & evt,
+				      0 /* do wait for mutex */ );
+      else
+	{
+	  clib_warning ("fifo full");
+	  return -1;
+	}
+    }
 
   /* *INDENT-OFF* */
-  SESSION_EVT_DBG(s, SESSION_EVT_ENQ, ({
+  SESSION_EVT_DBG(SESSION_EVT_ENQ, s, ({
       ed->data[0] = evt.event_id;
-      ed->data[1] = evt.enqueue_length;
+      ed->data[1] = svm_fifo_max_dequeue (s->server_rx_fifo);
   }));
   /* *INDENT-ON* */
 
@@ -1192,8 +1198,29 @@ stream_session_open (u8 sst, ip46_address_t * addr, u16 port_host_byte_order,
 void
 stream_session_disconnect (stream_session_t * s)
 {
+//  session_fifo_event_t evt;
+
   s->session_state = SESSION_STATE_CLOSED;
+  /* RPC to vpp evt queue in the right thread */
+
   tp_vfts[s->session_type].close (s->connection_index, s->thread_index);
+
+//  {
+//  /* Fabricate event */
+//  evt.fifo = s->server_rx_fifo;
+//  evt.event_type = FIFO_EVENT_SERVER_RX;
+//  evt.event_id = serial_number++;
+//
+//  /* Based on request block (or not) for lack of space */
+//  if (PREDICT_TRUE(q->cursize < q->maxsize))
+//    unix_shared_memory_queue_add (app->event_queue, (u8 *) &evt,
+//                                0 /* do wait for mutex */);
+//  else
+//    {
+//      clib_warning("fifo full");
+//      return -1;
+//    }
+//  }
 }
 
 /**
diff --git a/src/vnet/session/session.h b/src/vnet/session/session.h
index 96c00d87..a39bc06f 100644
--- a/src/vnet/session/session.h
+++ b/src/vnet/session/session.h
@@ -33,6 +33,7 @@ typedef enum
   FIFO_EVENT_SERVER_TX,
   FIFO_EVENT_TIMEOUT,
   FIFO_EVENT_SERVER_EXIT,
+  FIFO_EVENT_BUILTIN_RX
 } fifo_event_type_t;
 
 #define foreach_session_input_error                                         \
@@ -91,14 +92,13 @@ typedef enum
   SESSION_STATE_N_STATES,
 } stream_session_state_t;
 
-typedef CLIB_PACKED (struct
-		     {
-		     svm_fifo_t * fifo;
-		     u8 event_type;
-		     /* $$$$ for event logging */
-		     u16 event_id;
-		     u32 enqueue_length;
-		     }) session_fifo_event_t;
+/* *INDENT-OFF* */
+typedef CLIB_PACKED (struct {
+  svm_fifo_t * fifo;
+  u8 event_type;
+  u16 event_id;
+}) session_fifo_event_t;
+/* *INDENT-ON* */
 
 typedef struct _stream_session_t
 {
@@ -333,7 +333,7 @@ stream_session_get_index (stream_session_t * s)
 }
 
 always_inline u32
-stream_session_max_enqueue (transport_connection_t * tc)
+stream_session_max_rx_enqueue (transport_connection_t * tc)
 {
   stream_session_t *s = stream_session_get (tc->s_index, tc->thread_index);
   return svm_fifo_max_enqueue (s->server_rx_fifo);
@@ -346,7 +346,6 @@ stream_session_fifo_size (transport_connection_t * tc)
   return s->server_rx_fifo->nitems;
 }
 
-
 int
 stream_session_enqueue_data (transport_connection_t * tc, u8 * data, u16 len,
 			     u8 queue_event);
diff --git a/src/vnet/session/session_cli.c b/src/vnet/session/session_cli.c
index b029ee65..38762afc 100644
--- a/src/vnet/session/session_cli.c
+++ b/src/vnet/session/session_cli.c
@@ -107,7 +107,7 @@ show_session_command_fn (vlib_main_t * vm, unformat_input_t * input,
 	    {
 	      if (once_per_pool)
 		{
-		  str = format (str, "%-40s%-20s%-20s%-15s",
+		  str = format (str, "%-50s%-20s%-20s%-15s",
 				"Connection", "Rx fifo", "Tx fifo",
 				"Session Index");
 		  vlib_cli_output (vm, "%v", str);
diff --git a/src/vnet/session/session_debug.h b/src/vnet/session/session_debug.h
index 858f12e0..80a97cd5 100644
--- a/src/vnet/session/session_debug.h
+++ b/src/vnet/session/session_debug.h
@@ -21,7 +21,8 @@
 
 #define foreach_session_dbg_evt		\
   _(ENQ, "enqueue")			\
-  _(DEQ, "dequeue")
+  _(DEQ, "dequeue")			\
+  _(DEQ_NODE, "dequeue")
 
 typedef enum _session_evt_dbg
 {
@@ -30,7 +31,10 @@ typedef enum _session_evt_dbg
 #undef _
 } session_evt_dbg_e;
 
-#if TRANSPORT_DEBUG
+#define SESSION_DBG (0)
+#define SESSION_DEQ_NODE_EVTS (0)
+
+#if TRANSPORT_DEBUG && SESSION_DBG
 
 #define DEC_SESSION_ETD(_s, _e, _size)					\
   struct								\
@@ -44,6 +48,12 @@ typedef enum _session_evt_dbg
   ed = ELOG_TRACK_DATA (&vlib_global_main.elog_main,			\
 			_e, _tc->elog_track)
 
+#define DEC_SESSION_ED(_e, _size)					\
+  struct								\
+  {									\
+    u32 data[_size];							\
+  } * ed;								\
+  ed = ELOG_DATA (&vlib_global_main.elog_main, _e)
 
 #define SESSION_EVT_DEQ_HANDLER(_s, _body)				\
 {									\
@@ -67,13 +77,33 @@ typedef enum _session_evt_dbg
   do { _body; } while (0);						\
 }
 
+#if SESSION_DEQ_NODE_EVTS
+#define SESSION_EVT_DEQ_NODE_HANDLER(_node_evt)				\
+{									\
+  ELOG_TYPE_DECLARE (_e) =						\
+  {									\
+    .format = "deq-node: %s",						\
+    .format_args = "t4",                                      		\
+    .n_enum_strings = 2,                                        	\
+    .enum_strings = {                                           	\
+      "start",                                             		\
+      "end",                                              		\
+    },									\
+  };									\
+  DEC_SESSION_ED(_e, 1);						\
+  ed->data[0] = _node_evt;						\
+}
+#else
+#define SESSION_EVT_DEQ_NODE_HANDLER(_node_evt)
+#endif
+
 #define CONCAT_HELPER(_a, _b) _a##_b
 #define CC(_a, _b) CONCAT_HELPER(_a, _b)
 
-#define SESSION_EVT_DBG(_s, _evt, _body) CC(_evt, _HANDLER)(_s, _body)
+#define SESSION_EVT_DBG(_evt, _args...) CC(_evt, _HANDLER)(_args)
 
 #else
-#define SESSION_EVT_DBG(_s, _evt, _body)
+#define SESSION_EVT_DBG(_evt, _args...)
 #endif
 
 #endif /* SRC_VNET_SESSION_SESSION_DEBUG_H_ */
diff --git a/src/vnet/session/transport.h b/src/vnet/session/transport.h
index 421121d2..2f912cbc 100644
--- a/src/vnet/session/transport.h
+++ b/src/vnet/session/transport.h
@@ -38,7 +38,7 @@ typedef struct _transport_connection
   u32 thread_index;		/**< Worker-thread index */
 
 #if TRANSPORT_DEBUG
-  elog_track_t elog_track;	/**< Debug purposes */
+  elog_track_t elog_track;	/**< Event logging */
 #endif
 
   /** Macros for 'derived classes' where base is named "connection" */
diff --git a/src/vnet/tcp/builtin_client.c b/src/vnet/tcp/builtin_client.c
new file mode 100644
index 00000000..a6eeb775
--- /dev/null
+++ b/src/vnet/tcp/builtin_client.c
@@ -0,0 +1,411 @@
+/*
+ * builtin_client.c - vpp built-in tcp client/connect code
+ *
+ * Copyright (c) 2017 by Cisco and/or its affiliates.
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at:
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include <vnet/vnet.h>
+#include <vnet/plugin/plugin.h>
+#include <vnet/tcp/builtin_client.h>
+
+#include <vlibapi/api.h>
+#include <vlibmemory/api.h>
+#include <vlibsocket/api.h>
+#include <vpp/app/version.h>
+
+/* define message IDs */
+#include <vpp/api/vpe_msg_enum.h>
+
+/* define message structures */
+#define vl_typedefs
+#include <vpp/api/vpe_all_api_h.h>
+#undef vl_typedefs
+
+/* define generated endian-swappers */
+#define vl_endianfun
+#include <vpp/api/vpe_all_api_h.h>
+#undef vl_endianfun
+
+/* instantiate all the print functions we know about */
+#define vl_print(handle, ...) vlib_cli_output (handle, __VA_ARGS__)
+#define vl_printfun
+#include <vpp/api/vpe_all_api_h.h>
+#undef vl_printfun
+
+static void
+send_test_chunk (tclient_main_t * tm, session_t * s)
+{
+  u8 *test_data = tm->connect_test_data;
+  int test_buf_offset = 0;
+  u32 bytes_this_chunk;
+  session_fifo_event_t evt;
+  static int serial_number = 0;
+  int rv;
+
+  while (s->bytes_to_send > 0)
+    {
+      bytes_this_chunk = vec_len (test_data) < s->bytes_to_send
+	? vec_len (test_data) : s->bytes_to_send;
+
+      rv = svm_fifo_enqueue_nowait (s->server_tx_fifo, 0 /*pid */ ,
+				    bytes_this_chunk,
+				    test_data + test_buf_offset);
+
+      if (rv > 0)
+	{
+	  s->bytes_to_send -= rv;
+	  test_buf_offset += rv;
+
+	  if (svm_fifo_set_event (s->server_tx_fifo))
+	    {
+	      /* Fabricate TX event, send to vpp */
+	      evt.fifo = s->server_tx_fifo;
+	      evt.event_type = FIFO_EVENT_SERVER_TX;
+	      evt.event_id = serial_number++;
+
+	      unix_shared_memory_queue_add (tm->vpp_event_queue, (u8 *) & evt,
+					    0 /* do wait for mutex */ );
+	    }
+	}
+    }
+}
+
+static void
+receive_test_chunk (tclient_main_t * tm, session_t * s)
+{
+  svm_fifo_t *rx_fifo = s->server_rx_fifo;
+  int n_read, bytes, i;
+
+  bytes = svm_fifo_max_dequeue (rx_fifo);
+  /* Allow enqueuing of new event */
+  svm_fifo_unset_event (rx_fifo);
+
+  /* Read the bytes */
+  do
+    {
+      n_read = svm_fifo_dequeue_nowait (rx_fifo, 0, vec_len (tm->rx_buf),
+					tm->rx_buf);
+      if (n_read > 0)
+	{
+	  bytes -= n_read;
+	  for (i = 0; i < n_read; i++)
+	    {
+	      if (tm->rx_buf[i] != ((s->bytes_received + i) & 0xff))
+		{
+		  clib_warning ("read %d error at byte %lld, 0x%x not 0x%x",
+				n_read, s->bytes_received + i,
+				tm->rx_buf[i],
+				((s->bytes_received + i) & 0xff));
+		}
+	    }
+	  s->bytes_to_receive -= n_read;
+	  s->bytes_received += n_read;
+	}
+
+    }
+  while (n_read < 0 || bytes > 0);
+}
+
+static void *
+tclient_thread_fn (void *arg)
+{
+  tclient_main_t *tm = &tclient_main;
+  vl_api_disconnect_session_t *dmp;
+  session_t *sp;
+  struct timespec ts, tsrem;
+  int i;
+  int try_tx, try_rx;
+  u32 *session_indices = 0;
+
+  /* stats thread wants no signals. */
+  {
+    sigset_t s;
+    sigfillset (&s);
+    pthread_sigmask (SIG_SETMASK, &s, 0);
+  }
+
+  while (1)
+    {
+      /* Wait until we're told to get busy */
+      while (tm->run_test == 0
+	     || (tm->ready_connections != tm->expected_connections))
+	{
+	  ts.tv_sec = 0;
+	  ts.tv_nsec = 100000000;
+	  while (nanosleep (&ts, &tsrem) < 0)
+	    ts = tsrem;
+	}
+      tm->run_test = 0;
+
+      clib_warning ("Run %d iterations", tm->n_iterations);
+
+      for (i = 0; i < tm->n_iterations; i++)
+	{
+	  session_t *sp;
+
+	  do
+	    {
+	      try_tx = try_rx = 0;
+
+	      /* *INDENT-OFF* */
+	      pool_foreach (sp, tm->sessions, ({
+                if (sp->bytes_to_send > 0)
+                  {
+                    send_test_chunk (tm, sp);
+                    try_tx = 1;
+                  }
+	      }));
+	      pool_foreach (sp, tm->sessions, ({
+		if (sp->bytes_to_receive > 0)
+                  {
+                    receive_test_chunk (tm, sp);
+                    try_rx = 1;
+                  }
+              }));
+	      /* *INDENT-ON* */
+
+	    }
+	  while (try_tx || try_rx);
+	}
+      clib_warning ("Done %d iterations", tm->n_iterations);
+
+      /* Disconnect sessions... */
+      vec_reset_length (session_indices);
+      pool_foreach (sp, tm->sessions, (
+					{
+					vec_add1 (session_indices,
+						  sp - tm->sessions);
+					}
+		    ));
+
+      for (i = 0; i < vec_len (session_indices); i++)
+	{
+	  sp = pool_elt_at_index (tm->sessions, session_indices[i]);
+	  dmp = vl_msg_api_alloc_as_if_client (sizeof (*dmp));
+	  memset (dmp, 0, sizeof (*dmp));
+	  dmp->_vl_msg_id = ntohs (VL_API_DISCONNECT_SESSION);
+	  dmp->client_index = tm->my_client_index;
+	  dmp->session_index = sp->vpp_session_index;
+	  dmp->session_thread_index = sp->vpp_session_thread;
+	  vl_msg_api_send_shmem (tm->vl_input_queue, (u8 *) & dmp);
+	  pool_put (tm->sessions, sp);
+	}
+    }
+  /* NOTREACHED */
+  return 0;
+}
+
+/* So we don't get "no handler for... " msgs */
+static void
+vl_api_memclnt_create_reply_t_handler (vl_api_memclnt_create_reply_t * mp)
+{
+  tclient_main_t *tm = &tclient_main;
+
+  tm->my_client_index = mp->index;
+}
+
+static void
+vl_api_connect_uri_reply_t_handler (vl_api_connect_uri_reply_t * mp)
+{
+  tclient_main_t *tm = &tclient_main;
+  session_t *session;
+  u32 session_index;
+  u64 key;
+  i32 retval = /* clib_net_to_host_u32 ( */ mp->retval /*) */ ;
+
+  if (retval < 0)
+    {
+      clib_warning ("connection failed: retval %d", retval);
+      return;
+    }
+
+  tm->our_event_queue = (unix_shared_memory_queue_t *)
+    mp->vpp_event_queue_address;
+
+  tm->vpp_event_queue = (unix_shared_memory_queue_t *)
+    mp->vpp_event_queue_address;
+
+  /*
+   * Setup session
+   */
+  pool_get (tm->sessions, session);
+  memset (session, 0, sizeof (*session));
+  session_index = session - tm->sessions;
+  session->bytes_to_receive = session->bytes_to_send = tm->bytes_to_send;
+
+  session->server_rx_fifo = (svm_fifo_t *) mp->server_rx_fifo;
+  session->server_rx_fifo->client_session_index = session_index;
+  session->server_tx_fifo = (svm_fifo_t *) mp->server_tx_fifo;
+  session->server_tx_fifo->client_session_index = session_index;
+
+  session->vpp_session_index = mp->session_index;
+  session->vpp_session_thread = mp->session_thread_index;
+
+  /* Add it to the session lookup table */
+  key = (((u64) mp->session_thread_index) << 32) | (u64) mp->session_index;
+  hash_set (tm->session_index_by_vpp_handles, key, session_index);
+
+  tm->ready_connections++;
+}
+
+static void
+create_api_loopback (tclient_main_t * tm)
+{
+  vl_api_memclnt_create_t _m, *mp = &_m;
+  extern void vl_api_memclnt_create_t_handler (vl_api_memclnt_create_t *);
+  api_main_t *am = &api_main;
+  vl_shmem_hdr_t *shmem_hdr;
+
+  /*
+   * Create a "loopback" API client connection
+   * Don't do things like this unless you know what you're doing...
+   */
+
+  shmem_hdr = am->shmem_hdr;
+  tm->vl_input_queue = shmem_hdr->vl_input_queue;
+  memset (mp, 0, sizeof (*mp));
+  mp->_vl_msg_id = VL_API_MEMCLNT_CREATE;
+  mp->context = 0xFEEDFACE;
+  mp->input_queue = (u64) tm->vl_input_queue;
+  strncpy ((char *) mp->name, "tcp_tester", sizeof (mp->name) - 1);
+
+  vl_api_memclnt_create_t_handler (mp);
+}
+
+#define foreach_tclient_static_api_msg       	\
+_(MEMCLNT_CREATE_REPLY, memclnt_create_reply)   \
+_(CONNECT_URI_REPLY, connect_uri_reply)
+
+static clib_error_t *
+tclient_api_hookup (vlib_main_t * vm)
+{
+  tclient_main_t *tm = &tclient_main;
+  vl_msg_api_msg_config_t _c, *c = &_c;
+  int i;
+
+  /* Init test data */
+  vec_validate (tm->connect_test_data, 64 * 1024 - 1);
+  for (i = 0; i < vec_len (tm->connect_test_data); i++)
+    tm->connect_test_data[i] = i & 0xff;
+
+  tm->session_index_by_vpp_handles = hash_create (0, sizeof (uword));
+  vec_validate (tm->rx_buf, vec_len (tm->connect_test_data) - 1);
+
+  /* Hook up client-side static APIs to our handlers */
+#define _(N,n) do {                                             \
+    c->id = VL_API_##N;                                         \
+    c->name = #n;                                               \
+    c->handler = vl_api_##n##_t_handler;                        \
+    c->cleanup = vl_noop_handler;                               \
+    c->endian = vl_api_##n##_t_endian;                          \
+    c->print = vl_api_##n##_t_print;                            \
+    c->size = sizeof(vl_api_##n##_t);                           \
+    c->traced = 1; /* trace, so these msgs print */             \
+    c->replay = 0; /* don't replay client create/delete msgs */ \
+    c->message_bounce = 0; /* don't bounce this message */	\
+    vl_msg_api_config(c);} while (0);
+
+  foreach_tclient_static_api_msg;
+#undef _
+
+  return 0;
+}
+
+VLIB_API_INIT_FUNCTION (tclient_api_hookup);
+
+static clib_error_t *
+test_tcp_clients_command_fn (vlib_main_t * vm,
+			     unformat_input_t * input,
+			     vlib_cli_command_t * cmd)
+{
+  u8 *connect_uri = (u8 *) "tcp://6.0.1.2/1234";
+  u8 *uri;
+  tclient_main_t *tm = &tclient_main;
+  int i;
+  u32 n_clients = 1;
+
+  tm->bytes_to_send = 8192;
+  tm->n_iterations = 1;
+  vec_free (tm->connect_uri);
+
+  while (unformat_check_input (input) != UNFORMAT_END_OF_INPUT)
+    {
+      if (unformat (input, "nclients %d", &n_clients))
+	;
+      else if (unformat (input, "iterations %d", &tm->n_iterations))
+	;
+      else if (unformat (input, "bytes %d", &tm->bytes_to_send))
+	;
+      else if (unformat (input, "uri %s", &tm->connect_uri))
+	;
+      else
+	return clib_error_return (0, "unknown input `%U'",
+				  format_unformat_error, input);
+    }
+
+  tm->ready_connections = 0;
+  tm->expected_connections = n_clients;
+  uri = connect_uri;
+  if (tm->connect_uri)
+    uri = tm->connect_uri;
+
+  create_api_loopback (tm);
+
+  /* Start a transmit thread */
+  if (tm->client_thread_handle == 0)
+    {
+      int rv = pthread_create (&tm->client_thread_handle,
+			       NULL /*attr */ , tclient_thread_fn, 0);
+      if (rv)
+	{
+	  tm->client_thread_handle = 0;
+	  return clib_error_return (0, "pthread_create returned %d", rv);
+	}
+    }
+
+  /* Fire off connect requests, in something approaching a normal manner */
+  for (i = 0; i < n_clients; i++)
+    {
+      vl_api_connect_uri_t *cmp;
+      cmp = vl_msg_api_alloc_as_if_client (sizeof (*cmp));
+      memset (cmp, 0, sizeof (*cmp));
+
+      cmp->_vl_msg_id = ntohs (VL_API_CONNECT_URI);
+      cmp->client_index = tm->my_client_index;
+      cmp->context = ntohl (0xfeedface);
+      memcpy (cmp->uri, uri, strlen ((char *) uri) + 1);
+      vl_msg_api_send_shmem (tm->vl_input_queue, (u8 *) & cmp);
+    }
+
+  tm->run_test = 1;
+
+  return 0;
+}
+
+/* *INDENT-OFF* */
+VLIB_CLI_COMMAND (test_clients_command, static) =
+{
+  .path = "test tcp clients",
+  .short_help = "test tcp clients",
+  .function = test_tcp_clients_command_fn,
+};
+/* *INDENT-ON* */
+
+/*
+ * fd.io coding-style-patch-verification: ON
+ *
+ * Local Variables:
+ * eval: (c-set-style "gnu")
+ * End:
+ */
diff --git a/src/vnet/tcp/builtin_client.h b/src/vnet/tcp/builtin_client.h
new file mode 100644
index 00000000..64030302
--- /dev/null
+++ b/src/vnet/tcp/builtin_client.h
@@ -0,0 +1,131 @@
+
+/*
+ * tclient.h - skeleton vpp engine plug-in header file
+ *
+ * Copyright (c) <current-year> <your-organization>
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at:
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#ifndef __included_tclient_h__
+#define __included_tclient_h__
+
+#include <vnet/vnet.h>
+#include <vnet/ip/ip.h>
+#include <vnet/ethernet/ethernet.h>
+
+#include <vppinfra/hash.h>
+#include <vppinfra/error.h>
+#include <vlibmemory/unix_shared_memory_queue.h>
+#include <svm/svm_fifo_segment.h>
+#include <vnet/session/session.h>
+#include <vnet/session/application_interface.h>
+
+typedef struct
+{
+  u32 bytes_to_send;
+  u32 bytes_sent;
+  u32 bytes_to_receive;
+  u32 bytes_received;
+
+  svm_fifo_t *server_rx_fifo;
+  svm_fifo_t *server_tx_fifo;
+
+  u32 vpp_session_index;
+  u32 vpp_session_thread;
+} session_t;
+
+typedef struct
+{
+  /* API message ID base */
+  u16 msg_id_base;
+
+  /* vpe input queue */
+  unix_shared_memory_queue_t *vl_input_queue;
+
+  /* API client handle */
+  u32 my_client_index;
+
+  /* The URI we're playing with */
+  u8 *uri;
+
+  /* Session pool */
+  session_t *sessions;
+
+  /* Hash table for disconnect processing */
+  uword *session_index_by_vpp_handles;
+
+  /* intermediate rx buffer */
+  u8 *rx_buf;
+
+  /* URI for slave's connect */
+  u8 *connect_uri;
+
+  u32 connected_session_index;
+
+  int i_am_master;
+
+  /* drop all packets */
+  int drop_packets;
+
+  /* Our event queue */
+  unix_shared_memory_queue_t *our_event_queue;
+
+  /* $$$ single thread only for the moment */
+  unix_shared_memory_queue_t *vpp_event_queue;
+
+  pid_t my_pid;
+
+  /* For deadman timers */
+  clib_time_t clib_time;
+
+  /* Connection counts */
+  u32 expected_connections;
+  volatile u32 ready_connections;
+
+  /* Signal variables */
+  volatile int run_test;
+
+  /* Number of iterations */
+  int n_iterations;
+
+  /* Bytes to send */
+  u32 bytes_to_send;
+
+  u32 configured_segment_size;
+
+  /* VNET_API_ERROR_FOO -> "Foo" hash table */
+  uword *error_string_by_error_number;
+
+  u8 *connect_test_data;
+  pthread_t client_thread_handle;
+  u32 client_bytes_received;
+  u8 test_return_packets;
+
+  /* convenience */
+  vlib_main_t *vlib_main;
+  vnet_main_t *vnet_main;
+  ethernet_main_t *ethernet_main;
+} tclient_main_t;
+
+tclient_main_t tclient_main;
+
+vlib_node_registration_t tclient_node;
+
+#endif /* __included_tclient_h__ */
+
+/*
+ * fd.io coding-style-patch-verification: ON
+ *
+ * Local Variables:
+ * eval: (c-set-style "gnu")
+ * End:
+ */
diff --git a/src/vnet/tcp/builtin_server.c b/src/vnet/tcp/builtin_server.c
index dd6759c5..efd26e91 100644
--- a/src/vnet/tcp/builtin_server.c
+++ b/src/vnet/tcp/builtin_server.c
@@ -22,6 +22,7 @@ typedef struct
 {
   u8 *rx_buf;
   unix_shared_memory_queue_t **vpp_queue;
+  u32 byte_index;
   vlib_main_t *vlib_main;
 } builtin_server_main_t;
 
@@ -37,6 +38,7 @@ builtin_session_accept_callback (stream_session_t * s)
   bsm->vpp_queue[s->thread_index] =
     session_manager_get_vpp_event_queue (s->thread_index);
   s->session_state = SESSION_STATE_READY;
+  bsm->byte_index = 0;
   return 0;
 }
 
@@ -80,57 +82,94 @@ builtin_redirect_connect_callback (u32 client_index, void *mp)
   return -1;
 }
 
+void
+test_bytes (builtin_server_main_t * bsm, int actual_transfer)
+{
+  int i;
+
+  for (i = 0; i < actual_transfer; i++)
+    {
+      if (bsm->rx_buf[i] != ((bsm->byte_index + i) & 0xff))
+	{
+	  clib_warning ("at %d expected %d got %d", bsm->byte_index + i,
+			(bsm->byte_index + i) & 0xff, bsm->rx_buf[i]);
+	}
+    }
+  bsm->byte_index += actual_transfer;
+}
+
 int
-builtin_server_rx_callback (stream_session_t * s, session_fifo_event_t * e)
+builtin_server_rx_callback (stream_session_t * s)
 {
-  int n_written, bytes, total_copy_bytes;
-  int n_read;
-  svm_fifo_t *tx_fifo;
+  u32 n_written, max_dequeue, max_enqueue, max_transfer;
+  int actual_transfer;
+  svm_fifo_t *tx_fifo, *rx_fifo;
   builtin_server_main_t *bsm = &builtin_server_main;
   session_fifo_event_t evt;
   static int serial_number = 0;
 
-  bytes = e->enqueue_length;
-  if (PREDICT_FALSE (bytes <= 0))
+  max_dequeue = svm_fifo_max_dequeue (s->server_rx_fifo);
+  max_enqueue = svm_fifo_max_enqueue (s->server_tx_fifo);
+
+  if (PREDICT_FALSE (max_dequeue == 0))
     {
-      clib_warning ("bizarre rx callback: bytes %d", bytes);
       return 0;
     }
 
   tx_fifo = s->server_tx_fifo;
+  rx_fifo = s->server_rx_fifo;
 
   /* Number of bytes we're going to copy */
-  total_copy_bytes = (bytes < (tx_fifo->nitems - tx_fifo->cursize)) ? bytes :
-    tx_fifo->nitems - tx_fifo->cursize;
+  max_transfer = (max_dequeue < max_enqueue) ? max_dequeue : max_enqueue;
 
-  if (PREDICT_FALSE (total_copy_bytes <= 0))
+  /* No space in tx fifo */
+  if (PREDICT_FALSE (max_transfer == 0))
     {
-      clib_warning ("no space in tx fifo, event had %d bytes", bytes);
+      /* XXX timeout for session that are stuck */
+
+      /* Program self-tap to retry */
+      if (svm_fifo_set_event (rx_fifo))
+	{
+	  evt.fifo = rx_fifo;
+	  evt.event_type = FIFO_EVENT_BUILTIN_RX;
+	  evt.event_id = 0;
+	  unix_shared_memory_queue_add (bsm->vpp_queue[s->thread_index],
+					(u8 *) & evt,
+					0 /* do wait for mutex */ );
+	}
+
       return 0;
     }
 
-  vec_validate (bsm->rx_buf, total_copy_bytes - 1);
-  _vec_len (bsm->rx_buf) = total_copy_bytes;
+  svm_fifo_unset_event (rx_fifo);
+
+  vec_validate (bsm->rx_buf, max_transfer - 1);
+  _vec_len (bsm->rx_buf) = max_transfer;
 
-  n_read = svm_fifo_dequeue_nowait (s->server_rx_fifo, 0, total_copy_bytes,
-				    bsm->rx_buf);
-  ASSERT (n_read == total_copy_bytes);
+  actual_transfer = svm_fifo_dequeue_nowait (rx_fifo, 0, max_transfer,
+					     bsm->rx_buf);
+  ASSERT (actual_transfer == max_transfer);
+
+//  test_bytes (bsm, actual_transfer);
 
   /*
    * Echo back
    */
 
-  n_written = svm_fifo_enqueue_nowait (tx_fifo, 0, n_read, bsm->rx_buf);
-  ASSERT (n_written == total_copy_bytes);
+  n_written =
+    svm_fifo_enqueue_nowait (tx_fifo, 0, actual_transfer, bsm->rx_buf);
+  ASSERT (n_written == max_transfer);
 
-  /* Fabricate TX event, send to vpp */
-  evt.fifo = tx_fifo;
-  evt.event_type = FIFO_EVENT_SERVER_TX;
-  evt.enqueue_length = total_copy_bytes;
-  evt.event_id = serial_number++;
+  if (svm_fifo_set_event (tx_fifo))
+    {
+      /* Fabricate TX event, send to vpp */
+      evt.fifo = tx_fifo;
+      evt.event_type = FIFO_EVENT_SERVER_TX;
+      evt.event_id = serial_number++;
 
-  unix_shared_memory_queue_add (bsm->vpp_queue[s->thread_index], (u8 *) & evt,
-				0 /* do wait for mutex */ );
+      unix_shared_memory_queue_add (bsm->vpp_queue[s->thread_index],
+				    (u8 *) & evt, 0 /* do wait for mutex */ );
+    }
 
   return 0;
 }
@@ -164,7 +203,7 @@ server_create (vlib_main_t * vm)
   a->api_client_index = ~0;
   a->session_cb_vft = &builtin_session_cb_vft;
   a->options = options;
-  a->options[SESSION_OPTIONS_SEGMENT_SIZE] = 256 << 10;
+  a->options[SESSION_OPTIONS_SEGMENT_SIZE] = 128 << 20;
   a->options[SESSION_OPTIONS_RX_FIFO_SIZE] = 64 << 10;
   a->options[SESSION_OPTIONS_TX_FIFO_SIZE] = 64 << 10;
   a->segment_name = segment_name;
diff --git a/src/vnet/tcp/tcp.c b/src/vnet/tcp/tcp.c
index 0d2e6d0e..c3df5bc1 100644
--- a/src/vnet/tcp/tcp.c
+++ b/src/vnet/tcp/tcp.c
@@ -328,7 +328,7 @@ tcp_connection_init_vars (tcp_connection_t * tc)
 {
   tcp_connection_timers_init (tc);
   tcp_set_snd_mss (tc);
-  tc->sack_sb.head = TCP_INVALID_SACK_HOLE_INDEX;
+  scoreboard_init (&tc->sack_sb);
   tcp_cc_init (tc);
 }
 
@@ -558,17 +558,48 @@ tcp_session_send_mss (transport_connection_t * trans_conn)
   return tc->snd_mss;
 }
 
+/**
+ * Compute tx window session is allowed to fill.
+ */
 u32
 tcp_session_send_space (transport_connection_t * trans_conn)
 {
+  u32 snd_space;
   tcp_connection_t *tc = (tcp_connection_t *) trans_conn;
-  return tcp_available_snd_space (tc);
+
+  /* If we haven't gotten dupacks or if we did and have gotten sacked bytes
+   * then we can still send */
+  if (PREDICT_TRUE (tcp_in_fastrecovery (tc) == 0
+		    && (tc->rcv_dupacks == 0
+			|| tc->sack_sb.last_sacked_bytes)))
+    {
+      snd_space = tcp_available_snd_space (tc);
+
+      /* If we can't write at least a segment, don't try at all */
+      if (snd_space < tc->snd_mss)
+	return 0;
+      return snd_space;
+    }
+
+  /* If in fast recovery, send 1 SMSS if wnd allows */
+  if (tcp_in_fastrecovery (tc) && tcp_available_snd_space (tc)
+      && tcp_fastrecovery_sent_1_smss (tc))
+    {
+      tcp_fastrecovery_1_smss_on (tc);
+      return tc->snd_mss;
+    }
+
+  return 0;
 }
 
 u32
 tcp_session_tx_fifo_offset (transport_connection_t * trans_conn)
 {
   tcp_connection_t *tc = (tcp_connection_t *) trans_conn;
+
+  ASSERT (seq_geq (tc->snd_nxt, tc->snd_una));
+
+  /* This still works if fast retransmit is on */
   return (tc->snd_nxt - tc->snd_una);
 }
 
@@ -762,7 +793,7 @@ tcp_main_enable (vlib_main_t * vm)
   vec_validate (tm->timer_wheels, num_threads - 1);
   tcp_initialize_timer_wheels (tm);
 
-  vec_validate (tm->delack_connections, num_threads - 1);
+//  vec_validate (tm->delack_connections, num_threads - 1);
 
   /* Initialize clocks per tick for TCP timestamp. Used to compute
    * monotonically increasing timestamps. */
diff --git a/src/vnet/tcp/tcp.h b/src/vnet/tcp/tcp.h
index 082ab1d8..b4286bc4 100644
--- a/src/vnet/tcp/tcp.h
+++ b/src/vnet/tcp/tcp.h
@@ -30,9 +30,10 @@
 #define TCP_PAWS_IDLE 24 * 24 * 60 * 60 * THZ /**< 24 days */
 #define TCP_MAX_OPTION_SPACE 40
 
-#define TCP_DUPACK_THRESHOLD 3
-#define TCP_MAX_RX_FIFO_SIZE 2 << 20
-#define TCP_IW_N_SEGMENTS 10
+#define TCP_DUPACK_THRESHOLD 	3
+#define TCP_MAX_RX_FIFO_SIZE 	2 << 20
+#define TCP_IW_N_SEGMENTS 	10
+#define TCP_ALWAYS_ACK		0	/**< If on, we always ack */
 
 /** TCP FSM state definitions as per RFC793. */
 #define foreach_tcp_fsm_state   \
@@ -102,13 +103,12 @@ void tcp_update_time (f64 now, u32 thread_index);
 
 /** TCP connection flags */
 #define foreach_tcp_connection_flag             \
-  _(DELACK, "Delay ACK")                        \
   _(SNDACK, "Send ACK")                         \
-  _(BURSTACK, "Burst ACK set")                  \
   _(FINSNT, "FIN sent")				\
   _(SENT_RCV_WND0, "Sent 0 receive window")     \
   _(RECOVERY, "Recovery on")                    \
-  _(FAST_RECOVERY, "Fast Recovery on")
+  _(FAST_RECOVERY, "Fast Recovery on")		\
+  _(FR_1_SMSS, "Sent 1 SMSS")
 
 typedef enum _tcp_connection_flag_bits
 {
@@ -160,8 +160,12 @@ typedef struct _sack_scoreboard_hole
 typedef struct _sack_scoreboard
 {
   sack_scoreboard_hole_t *holes;	/**< Pool of holes */
-  u32 head;				/**< Index to first entry */
+  u32 head;				/**< Index of first entry */
+  u32 tail;				/**< Index of last entry */
   u32 sacked_bytes;			/**< Number of bytes sacked in sb */
+  u32 last_sacked_bytes;		/**< Number of bytes last sacked */
+  u32 snd_una_adv;			/**< Bytes to add to snd_una */
+  u32 max_byte_sacked;			/**< Highest byte acked */
 } sack_scoreboard_t;
 
 typedef enum _tcp_cc_algorithm_type
@@ -214,7 +218,7 @@ typedef struct _tcp_connection
   sack_block_t *snd_sacks;	/**< Vector of SACKs to send. XXX Fixed size? */
   sack_scoreboard_t sack_sb;	/**< SACK "scoreboard" that tracks holes */
 
-  u8 rcv_dupacks;	/**< Number of DUPACKs received */
+  u16 rcv_dupacks;	/**< Number of DUPACKs received */
   u8 snt_dupacks;	/**< Number of DUPACKs sent in a burst */
 
   /* Congestion control */
@@ -224,6 +228,7 @@ typedef struct _tcp_connection
   u32 bytes_acked;	/**< Bytes acknowledged by current segment */
   u32 rtx_bytes;	/**< Retransmitted bytes */
   u32 tsecr_last_ack;	/**< Timestamp echoed to us in last healthy ACK */
+  u32 snd_congestion;	/**< snd_una_max when congestion is detected */
   tcp_cc_algorithm_t *cc_algo;	/**< Congestion control algorithm */
 
   /* RTT and RTO */
@@ -250,8 +255,10 @@ struct _tcp_cc_algorithm
 #define tcp_fastrecovery_off(tc) (tc)->flags &= ~TCP_CONN_FAST_RECOVERY
 #define tcp_in_fastrecovery(tc) ((tc)->flags & TCP_CONN_FAST_RECOVERY)
 #define tcp_in_recovery(tc) ((tc)->flags & (TCP_CONN_FAST_RECOVERY | TCP_CONN_RECOVERY))
-#define tcp_recovery_off(tc) ((tc)->flags &= ~(TCP_CONN_FAST_RECOVERY | TCP_CONN_RECOVERY))
 #define tcp_in_slowstart(tc) (tc->cwnd < tc->ssthresh)
+#define tcp_fastrecovery_sent_1_smss(tc) ((tc)->flags & TCP_CONN_FR_1_SMSS)
+#define tcp_fastrecovery_1_smss_on(tc) ((tc)->flags |= TCP_CONN_FR_1_SMSS)
+#define tcp_fastrecovery_1_smss_off(tc) ((tc)->flags &= ~TCP_CONN_FR_1_SMSS)
 
 typedef enum
 {
@@ -293,8 +300,8 @@ typedef struct _tcp_main
   /* Per worker-thread timer wheel for connections timers */
   tw_timer_wheel_16t_2w_512sl_t *timer_wheels;
 
-  /* Convenience per worker-thread vector of connections to DELACK */
-  u32 **delack_connections;
+//  /* Convenience per worker-thread vector of connections to DELACK */
+//  u32 **delack_connections;
 
   /* Pool of half-open connections on which we've sent a SYN */
   tcp_connection_t *half_open_connections;
@@ -397,8 +404,16 @@ tcp_end_seq (tcp_header_t * th, u32 len)
 always_inline u32
 tcp_flight_size (const tcp_connection_t * tc)
 {
-  return tc->snd_una_max - tc->snd_una - tc->sack_sb.sacked_bytes
-    + tc->rtx_bytes;
+  int flight_size;
+
+  flight_size = (int) ((tc->snd_una_max - tc->snd_una) + tc->rtx_bytes)
+    - (tc->rcv_dupacks * tc->snd_mss) /* - tc->sack_sb.sacked_bytes */ ;
+
+  /* Happens if we don't clear sacked bytes */
+  if (flight_size < 0)
+    return 0;
+
+  return flight_size;
 }
 
 /**
@@ -439,9 +454,13 @@ tcp_available_snd_space (const tcp_connection_t * tc)
   return available_wnd - flight_size;
 }
 
+void tcp_update_rcv_wnd (tcp_connection_t * tc);
+
 void tcp_retransmit_first_unacked (tcp_connection_t * tc);
 
 void tcp_fast_retransmit (tcp_connection_t * tc);
+void tcp_cc_congestion (tcp_connection_t * tc);
+void tcp_cc_recover (tcp_connection_t * tc);
 
 always_inline u32
 tcp_time_now (void)
@@ -453,7 +472,7 @@ u32 tcp_push_header (transport_connection_t * tconn, vlib_buffer_t * b);
 
 u32
 tcp_prepare_retransmit_segment (tcp_connection_t * tc, vlib_buffer_t * b,
-				u32 max_bytes);
+				u32 offset, u32 max_bytes);
 
 void tcp_connection_timers_init (tcp_connection_t * tc);
 void tcp_connection_timers_reset (tcp_connection_t * tc);
@@ -476,14 +495,6 @@ tcp_timer_set (tcp_connection_t * tc, u8 timer_id, u32 interval)
 				   tc->c_c_index, timer_id, interval);
 }
 
-always_inline void
-tcp_retransmit_timer_set (tcp_connection_t * tc)
-{
-  /* XXX Switch to faster TW */
-  tcp_timer_set (tc, TCP_TIMER_RETRANSMIT,
-		 clib_max (tc->rto * TCP_TO_TIMER_TICK, 1));
-}
-
 always_inline void
 tcp_timer_reset (tcp_connection_t * tc, u8 timer_id)
 {
@@ -506,6 +517,27 @@ tcp_timer_update (tcp_connection_t * tc, u8 timer_id, u32 interval)
 				 tc->c_c_index, timer_id, interval);
 }
 
+/* XXX Switch retransmit to faster TW */
+always_inline void
+tcp_retransmit_timer_set (tcp_connection_t * tc)
+{
+  tcp_timer_set (tc, TCP_TIMER_RETRANSMIT,
+		 clib_max (tc->rto * TCP_TO_TIMER_TICK, 1));
+}
+
+always_inline void
+tcp_retransmit_timer_update (tcp_connection_t * tc)
+{
+  tcp_timer_update (tc, TCP_TIMER_RETRANSMIT,
+		    clib_max (tc->rto * TCP_TO_TIMER_TICK, 1));
+}
+
+always_inline void
+tcp_retransmit_timer_reset (tcp_connection_t * tc)
+{
+  tcp_timer_reset (tc, TCP_TIMER_RETRANSMIT);
+}
+
 always_inline u8
 tcp_timer_is_active (tcp_connection_t * tc, tcp_timers_e timer)
 {
@@ -516,6 +548,14 @@ void
 scoreboard_remove_hole (sack_scoreboard_t * sb,
 			sack_scoreboard_hole_t * hole);
 
+always_inline sack_scoreboard_hole_t *
+scoreboard_get_hole (sack_scoreboard_t * sb, u32 index)
+{
+  if (index != TCP_INVALID_SACK_HOLE_INDEX)
+    return pool_elt_at_index (sb->holes, index);
+  return 0;
+}
+
 always_inline sack_scoreboard_hole_t *
 scoreboard_next_hole (sack_scoreboard_t * sb, sack_scoreboard_hole_t * hole)
 {
@@ -532,6 +572,14 @@ scoreboard_first_hole (sack_scoreboard_t * sb)
   return 0;
 }
 
+always_inline sack_scoreboard_hole_t *
+scoreboard_last_hole (sack_scoreboard_t * sb)
+{
+  if (sb->tail != TCP_INVALID_SACK_HOLE_INDEX)
+    return pool_elt_at_index (sb->holes, sb->tail);
+  return 0;
+}
+
 always_inline void
 scoreboard_clear (sack_scoreboard_t * sb)
 {
@@ -540,6 +588,10 @@ scoreboard_clear (sack_scoreboard_t * sb)
     {
       scoreboard_remove_hole (sb, hole);
     }
+  sb->sacked_bytes = 0;
+  sb->last_sacked_bytes = 0;
+  sb->snd_una_adv = 0;
+  sb->max_byte_sacked = 0;
 }
 
 always_inline u32
@@ -548,6 +600,21 @@ scoreboard_hole_bytes (sack_scoreboard_hole_t * hole)
   return hole->end - hole->start;
 }
 
+always_inline u32
+scoreboard_hole_index (sack_scoreboard_t * sb, sack_scoreboard_hole_t * hole)
+{
+  return hole - sb->holes;
+}
+
+always_inline void
+scoreboard_init (sack_scoreboard_t * sb)
+{
+  sb->head = TCP_INVALID_SACK_HOLE_INDEX;
+  sb->tail = TCP_INVALID_SACK_HOLE_INDEX;
+}
+
+void tcp_rcv_sacks (tcp_connection_t * tc, u32 ack);
+
 always_inline void
 tcp_cc_algo_register (tcp_cc_algorithm_type_e type,
 		      const tcp_cc_algorithm_t * vft)
diff --git a/src/vnet/tcp/tcp_debug.h b/src/vnet/tcp/tcp_debug.h
index 069c512d..5a71694e 100644
--- a/src/vnet/tcp/tcp_debug.h
+++ b/src/vnet/tcp/tcp_debug.h
@@ -19,6 +19,8 @@
 #include <vlib/vlib.h>
 
 #define TCP_DEBUG (1)
+#define TCP_DEBUG_CC (1)
+#define TCP_DEBUG_VERBOSE (0)
 
 #define foreach_tcp_dbg_evt		\
   _(INIT, "")				\
@@ -30,14 +32,24 @@
   _(DELETE, "delete")			\
   _(SYN_SENT, "SYN sent")		\
   _(FIN_SENT, "FIN sent")		\
+  _(ACK_SENT, "ACK sent")		\
+  _(DUPACK_SENT, "DUPACK sent")		\
   _(RST_SENT, "RST sent")		\
   _(SYN_RCVD, "SYN rcvd")		\
   _(ACK_RCVD, "ACK rcvd")		\
+  _(DUPACK_RCVD, "DUPACK rcvd")		\
   _(FIN_RCVD, "FIN rcvd")		\
   _(RST_RCVD, "RST rcvd")		\
   _(PKTIZE, "packetize")		\
   _(INPUT, "in")			\
-  _(TIMER_POP, "timer pop")
+  _(SND_WND, "snd_wnd update")		\
+  _(OUTPUT, "output")			\
+  _(TIMER_POP, "timer pop")		\
+  _(CC_RTX, "retransmit")		\
+  _(CC_EVT, "cc event")			\
+  _(CC_PACK, "cc partial ack")		\
+  _(SEG_INVALID, "invalid segment")	\
+  _(ACK_RCV_ERR, "invalid ack")		\
 
 typedef enum _tcp_dbg
 {
@@ -73,10 +85,10 @@ typedef enum _tcp_dbg_evt
   ed = ELOG_TRACK_DATA (&vlib_global_main.elog_main,			\
 			_e, _tc->c_elog_track)
 
-#define TCP_EVT_INIT_HANDLER(_tc, ...)					\
+#define TCP_EVT_INIT_HANDLER(_tc, _fmt, ...)				\
 {									\
   _tc->c_elog_track.name = 						\
-	(char *) format (0, "%d%c", _tc->c_c_index, 0);			\
+	(char *) format (0, _fmt, _tc->c_c_index, 0);			\
   elog_track_register (&vlib_global_main.elog_main, &_tc->c_elog_track);\
 }
 
@@ -87,7 +99,7 @@ typedef enum _tcp_dbg_evt
 
 #define TCP_EVT_OPEN_HANDLER(_tc, ...)					\
 {									\
-  TCP_EVT_INIT_HANDLER(_tc);						\
+  TCP_EVT_INIT_HANDLER(_tc, "s%d%c");					\
   ELOG_TYPE_DECLARE (_e) =						\
   {									\
     .format = "open: index %d",						\
@@ -110,7 +122,7 @@ typedef enum _tcp_dbg_evt
 
 #define TCP_EVT_BIND_HANDLER(_tc, ...)					\
 {									\
-  TCP_EVT_INIT_HANDLER(_tc);						\
+  TCP_EVT_INIT_HANDLER(_tc, "l%d%c");					\
   ELOG_TYPE_DECLARE (_e) =						\
   {									\
     .format = "bind: listener %d",					\
@@ -138,16 +150,44 @@ typedef enum _tcp_dbg_evt
     .format = "delete: %d",						\
     .format_args = "i4",						\
   };									\
-  DECLARE_ETD(_tc, _e, 0);						\
+  DECLARE_ETD(_tc, _e, 1);						\
   ed->data[0] = _tc->c_c_index;						\
   TCP_EVT_DEALLOC_HANDLER(_tc);						\
 }
 
+#define TCP_EVT_ACK_SENT_HANDLER(_tc, ...)				\
+{									\
+  ELOG_TYPE_DECLARE (_e) =						\
+  {									\
+    .format = "ack_prep: acked %u rcv_nxt %u rcv_wnd %u snd_nxt %u",	\
+    .format_args = "i4i4i4i4",						\
+  };									\
+  DECLARE_ETD(_tc, _e, 4);						\
+  ed->data[0] = _tc->rcv_nxt - _tc->rcv_las;				\
+  ed->data[1] = _tc->rcv_nxt - _tc->irs;				\
+  ed->data[2] = _tc->rcv_wnd;						\
+  ed->data[3] = _tc->snd_nxt - _tc->iss;				\
+}
+
+#define TCP_EVT_DUPACK_SENT_HANDLER(_tc, ...)				\
+{									\
+  ELOG_TYPE_DECLARE (_e) =						\
+  {									\
+    .format = "dack_tx: rcv_nxt %u rcv_wnd %u snd_nxt %u av-wnd %u",	\
+    .format_args = "i4i4i4i4",						\
+  };									\
+  DECLARE_ETD(_tc, _e, 4);						\
+  ed->data[0] = _tc->rcv_nxt - _tc->irs;				\
+  ed->data[1] = _tc->rcv_wnd;						\
+  ed->data[2] = _tc->snd_nxt - _tc->iss;				\
+  ed->data[3] = tcp_available_wnd(_tc);					\
+}
+
 #define TCP_EVT_SYN_SENT_HANDLER(_tc, ...)				\
 {									\
   ELOG_TYPE_DECLARE (_e) =						\
   {									\
-    .format = "SYN: iss %d",						\
+    .format = "SYNtx: iss %u",					\
     .format_args = "i4",						\
   };									\
   DECLARE_ETD(_tc, _e, 1);						\
@@ -158,7 +198,7 @@ typedef enum _tcp_dbg_evt
 {									\
   ELOG_TYPE_DECLARE (_e) =						\
   {									\
-    .format = "FIN: snd_nxt %d rcv_nxt %d",				\
+    .format = "FINtx: snd_nxt %d rcv_nxt %d",				\
     .format_args = "i4i4",						\
   };									\
   DECLARE_ETD(_tc, _e, 2);						\
@@ -170,7 +210,7 @@ typedef enum _tcp_dbg_evt
 {									\
   ELOG_TYPE_DECLARE (_e) =						\
   {									\
-    .format = "RST: snd_nxt %d rcv_nxt %d",				\
+    .format = "RSTtx: snd_nxt %d rcv_nxt %d",				\
     .format_args = "i4i4",						\
   };									\
   DECLARE_ETD(_tc, _e, 2);						\
@@ -180,10 +220,10 @@ typedef enum _tcp_dbg_evt
 
 #define TCP_EVT_SYN_RCVD_HANDLER(_tc, ...)				\
 {									\
-  TCP_EVT_INIT_HANDLER(_tc);						\
+  TCP_EVT_INIT_HANDLER(_tc, "s%d%c");					\
   ELOG_TYPE_DECLARE (_e) =						\
   {									\
-    .format = "SYN rcvd: irs %d",					\
+    .format = "SYNrx: irs %u",						\
     .format_args = "i4",						\
   };									\
   DECLARE_ETD(_tc, _e, 1);						\
@@ -194,7 +234,7 @@ typedef enum _tcp_dbg_evt
 {									\
   ELOG_TYPE_DECLARE (_e) =						\
   {									\
-    .format = "FIN rcvd: snd_nxt %d rcv_nxt %d",			\
+    .format = "FINrx: snd_nxt %d rcv_nxt %d",				\
     .format_args = "i4i4",						\
   };									\
   DECLARE_ETD(_tc, _e, 2);						\
@@ -206,7 +246,7 @@ typedef enum _tcp_dbg_evt
 {									\
   ELOG_TYPE_DECLARE (_e) =						\
   {									\
-    .format = "RST rcvd: snd_nxt %d rcv_nxt %d",			\
+    .format = "RSTrx: snd_nxt %d rcv_nxt %d",				\
     .format_args = "i4i4",						\
   };									\
   DECLARE_ETD(_tc, _e, 2);						\
@@ -214,54 +254,68 @@ typedef enum _tcp_dbg_evt
   ed->data[1] = _tc->rcv_nxt - _tc->irs;				\
 }
 
-#define TCP_EVT_ACK_RCVD_HANDLER(_tc, ...)				\
+#define TCP_EVT_ACK_RCVD_HANDLER(_tc, _ack, ...)			\
 {									\
   ELOG_TYPE_DECLARE (_e) =						\
   {									\
-    .format = "ACK: acked %u cwnd %u inflight %u",			\
-    .format_args = "i4i4i4",						\
+    .format = "acked: %u snd_una %u ack %u cwnd %u inflight %u",	\
+    .format_args = "i4i4i4i4i4",					\
   };									\
-  DECLARE_ETD(_tc, _e, 3);						\
+  DECLARE_ETD(_tc, _e, 5);						\
   ed->data[0] = _tc->bytes_acked;					\
-  ed->data[1] = _tc->cwnd;						\
-  ed->data[2] = tcp_flight_size(_tc);					\
+  ed->data[1] = _tc->snd_una - _tc->iss;				\
+  ed->data[2] = _ack - _tc->iss;					\
+  ed->data[3] = _tc->cwnd;						\
+  ed->data[4] = tcp_flight_size(_tc);					\
 }
 
-#define TCP_EVT_PKTIZE_HANDLER(_tc, ...)				\
+#define TCP_EVT_DUPACK_RCVD_HANDLER(_tc, ...)				\
 {									\
   ELOG_TYPE_DECLARE (_e) =						\
   {									\
-    .format = "pktize: snd_una %u snd_nxt %u una_max %u",		\
-    .format_args = "i4i4i4",						\
+    .format = "dack_rx: snd_una %u cwnd %u snd_wnd %u inflight %u",	\
+    .format_args = "i4i4i4i4",						\
   };									\
-  DECLARE_ETD(_tc, _e, 3);						\
+  DECLARE_ETD(_tc, _e, 4);						\
   ed->data[0] = _tc->snd_una - _tc->iss;				\
-  ed->data[1] = _tc->snd_nxt - _tc->iss;				\
-  ed->data[2] = _tc->snd_una_max - _tc->iss;				\
+  ed->data[1] = _tc->cwnd;						\
+  ed->data[2] = _tc->snd_wnd;						\
+  ed->data[3] = tcp_flight_size(_tc);					\
 }
 
-#define TCP_EVT_OUTPUT_HANDLER(_tc, flags, n_bytes,...)			\
+#define TCP_EVT_PKTIZE_HANDLER(_tc, ...)				\
 {									\
   ELOG_TYPE_DECLARE (_e) =						\
   {									\
-    .format = "out: flags %x, bytes %u",				\
-    .format_args = "i4i4",						\
+    .format = "pktize: una %u snd_nxt %u space %u flight %u rcv_wnd %u",\
+    .format_args = "i4i4i4i4i4",					\
   };									\
-  DECLARE_ETD(_tc, _e, 2);						\
-  ed->data[0] = flags;							\
-  ed->data[1] = n_bytes;						\
+  DECLARE_ETD(_tc, _e, 5);						\
+  ed->data[0] = _tc->snd_una - _tc->iss;				\
+  ed->data[1] = _tc->snd_nxt - _tc->iss;				\
+  ed->data[2] = tcp_available_snd_space (_tc);				\
+  ed->data[3] = tcp_flight_size (_tc);					\
+  ed->data[4] = _tc->rcv_wnd;						\
 }
 
-#define TCP_EVT_INPUT_HANDLER(_tc, n_bytes, ...)			\
+#define TCP_EVT_INPUT_HANDLER(_tc, _type, _len, _written, ...)		\
 {									\
   ELOG_TYPE_DECLARE (_e) =						\
   {									\
-    .format = "in: bytes %u rcv_nxt %u",				\
-    .format_args = "i4i4",						\
+    .format = "in: %s len %u written %d rcv_nxt %u free wnd %d",	\
+    .format_args = "t4i4i4i4i4",					\
+    .n_enum_strings = 2,                                        	\
+    .enum_strings = {                                           	\
+      "order",   	                                        	\
+      "ooo",   								\
+    },									\
   };									\
-  DECLARE_ETD(_tc, _e, 2);						\
-  ed->data[0] = n_bytes;						\
-  ed->data[1] = _tc->rcv_nxt - _tc->irs;				\
+  DECLARE_ETD(_tc, _e, 5);						\
+  ed->data[0] = _type;							\
+  ed->data[1] = _len;							\
+  ed->data[2] = _written;						\
+  ed->data[3] = (_tc->rcv_nxt - _tc->irs) + _written;			\
+  ed->data[4] = _tc->rcv_wnd - (_tc->rcv_nxt - _tc->rcv_las);		\
 }
 
 #define TCP_EVT_TIMER_POP_HANDLER(_tc_index, _timer_id, ...)            \
@@ -296,9 +350,131 @@ typedef enum _tcp_dbg_evt
   ed->data[1] = _timer_id;                                      	\
 }
 
+#define TCP_EVT_SEG_INVALID_HANDLER(_tc, _seq, _end, ...)		\
+{									\
+  ELOG_TYPE_DECLARE (_e) =						\
+  {									\
+    .format = "seg-inv: seq %u end %u rcv_las %u rcv_nxt %u wnd %u",	\
+    .format_args = "i4i4i4i4i4",					\
+  };									\
+  DECLARE_ETD(_tc, _e, 5);						\
+  ed->data[0] = _seq - _tc->irs;					\
+  ed->data[1] = _end - _tc->irs;					\
+  ed->data[2] = _tc->rcv_las - _tc->irs;				\
+  ed->data[3] = _tc->rcv_nxt - _tc->irs;				\
+  ed->data[4] = _tc->rcv_wnd;						\
+}
+
+#define TCP_EVT_ACK_RCV_ERR_HANDLER(_tc, _type, _ack, ...)		\
+{									\
+  ELOG_TYPE_DECLARE (_e) =						\
+  {									\
+    .format = "ack-err: %s ack %u snd_una %u snd_nxt %u una_max %u",	\
+    .format_args = "t4i4i4i4i4",					\
+    .n_enum_strings = 3,						\
+    .enum_strings = {                                           	\
+      "invalid",                                                 	\
+      "old",                                                 		\
+      "future",								\
+    }, 									\
+  };									\
+  DECLARE_ETD(_tc, _e, 5);						\
+  ed->data[0] = _type;							\
+  ed->data[1] = _ack - _tc->iss;					\
+  ed->data[2] = _tc->snd_una - _tc->iss;				\
+  ed->data[3] = _tc->snd_nxt - _tc->iss;				\
+  ed->data[4] = _tc->snd_una_max - _tc->iss;				\
+}
+
+/*
+ * Congestion Control
+ */
+
+#if TCP_DEBUG_CC
+#define TCP_EVT_CC_RTX_HANDLER(_tc, offset, n_bytes, ...)		\
+{									\
+  ELOG_TYPE_DECLARE (_e) =						\
+  {									\
+    .format = "rtx: snd_nxt %u offset %u snd %u rtx %u",		\
+    .format_args = "i4i4i4i4",						\
+  };									\
+  DECLARE_ETD(_tc, _e, 4);						\
+  ed->data[0] = _tc->snd_nxt - _tc->iss;				\
+  ed->data[1] = offset;							\
+  ed->data[2] = n_bytes;						\
+  ed->data[3] = _tc->rtx_bytes;						\
+}
+
+#define TCP_EVT_CC_EVT_HANDLER(_tc, _sub_evt, ...)			\
+{									\
+  ELOG_TYPE_DECLARE (_e) =						\
+  {									\
+    .format = "cc: %s wnd %u snd_cong %u rtx_bytes %u",			\
+    .format_args = "t4i4i4i4",						\
+    .n_enum_strings = 5,						\
+    .enum_strings = {                                           	\
+      "fast-rtx",	                                             	\
+      "rtx-timeout",                                                 	\
+      "first-rtx",                                                 	\
+      "recovered",							\
+      "congestion",							\
+    },  								\
+  };									\
+  DECLARE_ETD(_tc, _e, 4);						\
+  ed->data[0] = _sub_evt;						\
+  ed->data[1] = tcp_available_snd_space (_tc);				\
+  ed->data[2] = _tc->snd_congestion - _tc->iss;				\
+  ed->data[3] = _tc->rtx_bytes;						\
+}
+
+#define TCP_EVT_CC_PACK_HANDLER(_tc, ...)				\
+{									\
+  ELOG_TYPE_DECLARE (_e) =						\
+  {									\
+    .format = "pack: snd_una %u snd_una_max %u",			\
+    .format_args = "i4i4",						\
+  };									\
+  DECLARE_ETD(_tc, _e, 2);						\
+  ed->data[0] = _tc->snd_una - _tc->iss;				\
+  ed->data[1] = _tc->snd_una_max - _tc->iss;				\
+}
+
+#else
+#define TCP_EVT_CC_RTX_HANDLER(_tc, offset, n_bytes, ...)
+#define TCP_EVT_CC_EVT_HANDLER(_tc, _sub_evt, _snd_space, ...)
+#define TCP_EVT_CC_PACK_HANDLER(_tc, ...)
+#endif
+
+#if TCP_DBG_VERBOSE
+#define TCP_EVT_SND_WND_HANDLER(_tc, ...)				\
+{									\
+  ELOG_TYPE_DECLARE (_e) =						\
+  {									\
+    .format = "snd_wnd update: %u ",					\
+    .format_args = "i4",						\
+  };									\
+  DECLARE_ETD(_tc, _e, 1);						\
+  ed->data[0] = _tc->snd_wnd;						\
+}
+
+#define TCP_EVT_OUTPUT_HANDLER(_tc, flags, n_bytes,...)			\
+{									\
+  ELOG_TYPE_DECLARE (_e) =						\
+  {									\
+    .format = "out: flags %x, bytes %u",				\
+    .format_args = "i4i4",						\
+  };									\
+  DECLARE_ETD(_tc, _e, 2);						\
+  ed->data[0] = flags;							\
+  ed->data[1] = n_bytes;						\
+}
+#else
+#define TCP_EVT_SND_WND_HANDLER(_tc, ...)
+#define TCP_EVT_OUTPUT_HANDLER(_tc, flags, n_bytes,...)
+#endif
+
 #define CONCAT_HELPER(_a, _b) _a##_b
 #define CC(_a, _b) CONCAT_HELPER(_a, _b)
-
 #define TCP_EVT_DBG(_evt, _args...) CC(_evt, _HANDLER)(_args)
 
 #else
diff --git a/src/vnet/tcp/tcp_error.def b/src/vnet/tcp/tcp_error.def
index 2dbdd9b3..b91a08c0 100644
--- a/src/vnet/tcp/tcp_error.def
+++ b/src/vnet/tcp/tcp_error.def
@@ -12,12 +12,12 @@
  * See the License for the specific language governing permissions and
  * limitations under the License.
  */
- 
 tcp_error (NONE, "no error")
 tcp_error (NO_LISTENER, "no listener for dst port")
 tcp_error (LOOKUP_DROPS, "lookup drops")
 tcp_error (DISPATCH, "Dispatch error")
 tcp_error (ENQUEUED, "Packets pushed into rx fifo") 
+tcp_error (PARTIALLY_ENQUEUED, "Packets partially pushed into rx fifo") 
 tcp_error (PURE_ACK, "Pure acks")
 tcp_error (SYNS_RCVD, "SYNs received")
 tcp_error (SYN_ACKS_RCVD, "SYN-ACKs received")
@@ -26,11 +26,14 @@ tcp_error (FIFO_FULL, "Packets dropped for lack of rx fifo space")
 tcp_error (EVENT_FIFO_FULL, "Events not sent for lack of event fifo space") 
 tcp_error (API_QUEUE_FULL, "Sessions not created for lack of API queue space")
 tcp_error (CREATE_SESSION_FAIL, "Sessions couldn't be allocated")
-tcp_error (SEGMENT_INVALID, "Invalid segment")
+tcp_error (SEGMENT_INVALID, "Invalid segments")
+tcp_error (SEGMENT_OLD, "Old segment")
 tcp_error (ACK_INVALID, "Invalid ACK")
 tcp_error (ACK_DUP, "Duplicate ACK")
 tcp_error (ACK_OLD, "Old ACK")
+tcp_error (ACK_FUTURE, "Future ACK")
 tcp_error (PKTS_SENT, "Packets sent")
 tcp_error (FILTERED_DUPACKS, "Filtered duplicate ACKs")
 tcp_error (RST_SENT, "Resets sent")
 tcp_error (INVALID_CONNECTION, "Invalid connection")
+tcp_error (NO_WND, "No window")
\ No newline at end of file
diff --git a/src/vnet/tcp/tcp_input.c b/src/vnet/tcp/tcp_input.c
index 67af4321..5d11985f 100644
--- a/src/vnet/tcp/tcp_input.c
+++ b/src/vnet/tcp/tcp_input.c
@@ -95,13 +95,21 @@ vlib_node_registration_t tcp6_established_node;
  * or the rcv_nxt at last ack sent instead of rcv_nxt since that's the
  * peer's reference when computing our receive window.
  *
- * This accepts only segments within the window.
+ * This:
+ *  seq_leq (end_seq, tc->rcv_las + tc->rcv_wnd) && seq_geq (seq, tc->rcv_las)
+ * however, is too strict when we have retransmits. Instead we just check that
+ * the seq is not beyond the right edge and that the end of the segment is not
+ * less than the left edge.
+ *
+ * N.B. rcv_nxt and rcv_wnd are both updated in this node if acks are sent, so
+ * use rcv_nxt in the right edge window test instead of rcv_las.
+ *
  */
 always_inline u8
 tcp_segment_in_rcv_wnd (tcp_connection_t * tc, u32 seq, u32 end_seq)
 {
-  return seq_leq (end_seq, tc->rcv_las + tc->rcv_wnd)
-    && seq_geq (seq, tc->rcv_nxt);
+  return (seq_geq (end_seq, tc->rcv_las)
+	  && seq_leq (seq, tc->rcv_nxt + tc->rcv_wnd));
 }
 
 void
@@ -253,6 +261,7 @@ tcp_segment_validate (vlib_main_t * vm, tcp_connection_t * tc0,
 	    {
 	      tcp_make_ack (tc0, b0);
 	      *next0 = tcp_next_output (tc0->c_is_ip4);
+	      TCP_EVT_DBG (TCP_EVT_DUPACK_SENT, tc0);
 	      return -1;
 	    }
 	}
@@ -262,13 +271,25 @@ tcp_segment_validate (vlib_main_t * vm, tcp_connection_t * tc0,
   if (!tcp_segment_in_rcv_wnd (tc0, vnet_buffer (b0)->tcp.seq_number,
 			       vnet_buffer (b0)->tcp.seq_end))
     {
-      if (!tcp_rst (th0))
+      /* If our window is 0 and the packet is in sequence, let it pass
+       * through for ack processing. It should be dropped later.*/
+      if (tc0->rcv_wnd == 0
+	  && tc0->rcv_nxt == vnet_buffer (b0)->tcp.seq_number)
 	{
-	  /* Send dup ack */
-	  tcp_make_ack (tc0, b0);
-	  *next0 = tcp_next_output (tc0->c_is_ip4);
+	  /* Make it look as if there's nothing to dequeue */
+	  vnet_buffer (b0)->tcp.seq_end = vnet_buffer (b0)->tcp.seq_number;
+	}
+      else
+	{
+	  /* If not RST, send dup ack */
+	  if (!tcp_rst (th0))
+	    {
+	      tcp_make_ack (tc0, b0);
+	      *next0 = tcp_next_output (tc0->c_is_ip4);
+	      TCP_EVT_DBG (TCP_EVT_DUPACK_SENT, tc0);
+	    }
+	  return -1;
 	}
-      return -1;
     }
 
   /* 2nd: check the RST bit */
@@ -326,13 +347,13 @@ tcp_estimate_rtt (tcp_connection_t * tc, u32 mrtt)
 
       /* XXX Drop in RTT results in RTTVAR increase and bigger RTO.
        * The increase should be bound */
-      tc->rttvar += (clib_abs (err) - tc->rttvar) >> 2;
+      tc->rttvar += ((int) clib_abs (err) - (int) tc->rttvar) >> 2;
     }
   else
     {
       /* First measurement. */
       tc->srtt = mrtt;
-      tc->rttvar = mrtt << 1;
+      tc->rttvar = mrtt >> 1;
     }
 }
 
@@ -394,7 +415,11 @@ tcp_dequeue_acked (tcp_connection_t * tc, u32 ack)
     }
 }
 
-/** Check if dupack as per RFC5681 Sec. 2 */
+/**
+ * Check if dupack as per RFC5681 Sec. 2
+ *
+ * This works only if called before updating snd_wnd.
+ * */
 always_inline u8
 tcp_ack_is_dupack (tcp_connection_t * tc, vlib_buffer_t * b, u32 new_snd_wnd)
 {
@@ -429,10 +454,10 @@ scoreboard_remove_hole (sack_scoreboard_t * sb, sack_scoreboard_hole_t * hole)
 }
 
 sack_scoreboard_hole_t *
-scoreboard_insert_hole (sack_scoreboard_t * sb, sack_scoreboard_hole_t * prev,
+scoreboard_insert_hole (sack_scoreboard_t * sb, u32 prev_index,
 			u32 start, u32 end)
 {
-  sack_scoreboard_hole_t *hole, *next;
+  sack_scoreboard_hole_t *hole, *next, *prev;
   u32 hole_index;
 
   pool_get (sb->holes, hole);
@@ -442,6 +467,7 @@ scoreboard_insert_hole (sack_scoreboard_t * sb, sack_scoreboard_hole_t * prev,
   hole->end = end;
   hole_index = hole - sb->holes;
 
+  prev = scoreboard_get_hole (sb, prev_index);
   if (prev)
     {
       hole->prev = prev - sb->holes;
@@ -462,28 +488,35 @@ scoreboard_insert_hole (sack_scoreboard_t * sb, sack_scoreboard_hole_t * prev,
   return hole;
 }
 
-static void
+void
 tcp_rcv_sacks (tcp_connection_t * tc, u32 ack)
 {
   sack_scoreboard_t *sb = &tc->sack_sb;
   sack_block_t *blk, tmp;
-  sack_scoreboard_hole_t *hole, *next_hole;
-  u32 blk_index = 0;
+  sack_scoreboard_hole_t *hole, *next_hole, *last_hole, *new_hole;
+  u32 blk_index = 0, old_sacked_bytes, hole_index;
   int i, j;
 
-  if (!tcp_opts_sack (tc) && sb->head == TCP_INVALID_SACK_HOLE_INDEX)
+  sb->last_sacked_bytes = 0;
+  sb->snd_una_adv = 0;
+  old_sacked_bytes = sb->sacked_bytes;
+
+  if (!tcp_opts_sack (&tc->opt) && sb->head == TCP_INVALID_SACK_HOLE_INDEX)
     return;
 
   /* Remove invalid blocks */
-  vec_foreach (blk, tc->opt.sacks)
-  {
-    if (seq_lt (blk->start, blk->end)
-	&& seq_gt (blk->start, tc->snd_una)
-	&& seq_gt (blk->start, ack) && seq_lt (blk->end, tc->snd_nxt))
-      continue;
-
-    vec_del1 (tc->opt.sacks, blk - tc->opt.sacks);
-  }
+  blk = tc->opt.sacks;
+  while (blk < vec_end (tc->opt.sacks))
+    {
+      if (seq_lt (blk->start, blk->end)
+	  && seq_gt (blk->start, tc->snd_una)
+	  && seq_gt (blk->start, ack) && seq_leq (blk->end, tc->snd_nxt))
+	{
+	  blk++;
+	  continue;
+	}
+      vec_del1 (tc->opt.sacks, blk - tc->opt.sacks);
+    }
 
   /* Add block for cumulative ack */
   if (seq_gt (ack, tc->snd_una))
@@ -498,7 +531,7 @@ tcp_rcv_sacks (tcp_connection_t * tc, u32 ack)
 
   /* Make sure blocks are ordered */
   for (i = 0; i < vec_len (tc->opt.sacks); i++)
-    for (j = i; j < vec_len (tc->opt.sacks); j++)
+    for (j = i + 1; j < vec_len (tc->opt.sacks); j++)
       if (seq_lt (tc->opt.sacks[j].start, tc->opt.sacks[i].start))
 	{
 	  tmp = tc->opt.sacks[i];
@@ -506,10 +539,22 @@ tcp_rcv_sacks (tcp_connection_t * tc, u32 ack)
 	  tc->opt.sacks[j] = tmp;
 	}
 
-  /* If no holes, insert the first that covers all outstanding bytes */
   if (sb->head == TCP_INVALID_SACK_HOLE_INDEX)
     {
-      scoreboard_insert_hole (sb, 0, tc->snd_una, tc->snd_una_max);
+      /* If no holes, insert the first that covers all outstanding bytes */
+      last_hole = scoreboard_insert_hole (sb, TCP_INVALID_SACK_HOLE_INDEX,
+					  tc->snd_una, tc->snd_una_max);
+      sb->tail = scoreboard_hole_index (sb, last_hole);
+    }
+  else
+    {
+      /* If we have holes but snd_una_max is beyond the last hole, update
+       * last hole end */
+      tmp = tc->opt.sacks[vec_len (tc->opt.sacks) - 1];
+      last_hole = scoreboard_last_hole (sb);
+      if (seq_gt (tc->snd_una_max, sb->max_byte_sacked)
+	  && seq_gt (tc->snd_una_max, last_hole->end))
+	last_hole->end = tc->snd_una_max;
     }
 
   /* Walk the holes with the SACK blocks */
@@ -526,10 +571,10 @@ tcp_rcv_sacks (tcp_connection_t * tc, u32 ack)
 	      next_hole = scoreboard_next_hole (sb, hole);
 
 	      /* Byte accounting */
-	      if (seq_lt (hole->end, ack))
+	      if (seq_leq (hole->end, ack))
 		{
-		  /* Bytes lost because snd wnd left edge advances */
-		  if (seq_lt (next_hole->start, ack))
+		  /* Bytes lost because snd_wnd left edge advances */
+		  if (next_hole && seq_leq (next_hole->start, ack))
 		    sb->sacked_bytes -= next_hole->start - hole->end;
 		  else
 		    sb->sacked_bytes -= ack - hole->end;
@@ -539,35 +584,78 @@ tcp_rcv_sacks (tcp_connection_t * tc, u32 ack)
 		  sb->sacked_bytes += scoreboard_hole_bytes (hole);
 		}
 
+	      /* snd_una needs to be advanced */
+	      if (seq_geq (ack, hole->end))
+		{
+		  if (next_hole && seq_lt (ack, next_hole->start))
+		    sb->snd_una_adv = next_hole->start - ack;
+		  else
+		    sb->snd_una_adv = sb->max_byte_sacked - ack;
+
+		  /* all these can be delivered */
+		  sb->sacked_bytes -= sb->snd_una_adv;
+		}
+
+	      /* About to remove last hole */
+	      if (hole == last_hole)
+		{
+		  sb->tail = hole->prev;
+		  last_hole = scoreboard_last_hole (sb);
+		  /* keep track of max byte sacked in case the last hole
+		   * is acked */
+		  if (seq_gt (hole->end, sb->max_byte_sacked))
+		    sb->max_byte_sacked = hole->end;
+		}
 	      scoreboard_remove_hole (sb, hole);
 	      hole = next_hole;
 	    }
-	  /* Partial overlap */
+	  /* Partial 'head' overlap */
 	  else
 	    {
-	      sb->sacked_bytes += blk->end - hole->start;
-	      hole->start = blk->end;
+	      if (seq_gt (blk->end, hole->start))
+		{
+		  sb->sacked_bytes += blk->end - hole->start;
+		  hole->start = blk->end;
+		}
 	      blk_index++;
 	    }
 	}
       else
 	{
 	  /* Hole must be split */
-	  if (seq_leq (blk->end, hole->end))
+	  if (seq_lt (blk->end, hole->end))
 	    {
 	      sb->sacked_bytes += blk->end - blk->start;
-	      scoreboard_insert_hole (sb, hole, blk->end, hole->end);
-	      hole->end = blk->start - 1;
+	      hole_index = scoreboard_hole_index (sb, hole);
+	      new_hole = scoreboard_insert_hole (sb, hole_index, blk->end,
+						 hole->end);
+
+	      /* Pool might've moved */
+	      hole = scoreboard_get_hole (sb, hole_index);
+	      hole->end = blk->start;
+
+	      /* New or split of tail */
+	      if ((last_hole->end == new_hole->end)
+		  || seq_lt (last_hole->end, new_hole->start))
+		{
+		  last_hole = new_hole;
+		  sb->tail = scoreboard_hole_index (sb, new_hole);
+		}
+
 	      blk_index++;
+	      hole = scoreboard_next_hole (sb, hole);
 	    }
 	  else
 	    {
-	      sb->sacked_bytes += hole->end - blk->start + 1;
-	      hole->end = blk->start - 1;
+	      sb->sacked_bytes += hole->end - blk->start;
+	      hole->end = blk->start;
 	      hole = scoreboard_next_hole (sb, hole);
 	    }
 	}
     }
+
+  sb->last_sacked_bytes = sb->sacked_bytes + sb->snd_una_adv
+    - old_sacked_bytes;
 }
 
 /** Update snd_wnd
@@ -577,72 +665,94 @@ tcp_rcv_sacks (tcp_connection_t * tc, u32 ack)
 static void
 tcp_update_snd_wnd (tcp_connection_t * tc, u32 seq, u32 ack, u32 snd_wnd)
 {
-  if (tc->snd_wl1 < seq || (tc->snd_wl1 == seq && tc->snd_wl2 <= ack))
+  if (seq_lt (tc->snd_wl1, seq)
+      || (tc->snd_wl1 == seq && seq_leq (tc->snd_wl2, ack)))
     {
       tc->snd_wnd = snd_wnd;
       tc->snd_wl1 = seq;
       tc->snd_wl2 = ack;
+      TCP_EVT_DBG (TCP_EVT_SND_WND, tc);
     }
 }
 
-static void
+void
 tcp_cc_congestion (tcp_connection_t * tc)
 {
+  tc->snd_congestion = tc->snd_nxt;
   tc->cc_algo->congestion (tc);
+  TCP_EVT_DBG (TCP_EVT_CC_EVT, tc, 4);
 }
 
-static void
+void
 tcp_cc_recover (tcp_connection_t * tc)
 {
-  if (tcp_in_fastrecovery (tc))
-    {
-      tc->cc_algo->recovered (tc);
-      tcp_recovery_off (tc);
-    }
-  else if (tcp_in_recovery (tc))
-    {
-      tcp_recovery_off (tc);
-      tc->cwnd = tcp_loss_wnd (tc);
-    }
+  tc->cc_algo->recovered (tc);
+
+  tc->rtx_bytes = 0;
+  tc->rcv_dupacks = 0;
+  tc->snd_nxt = tc->snd_una;
+
+  tc->cc_algo->rcv_ack (tc);
+  tc->tsecr_last_ack = tc->opt.tsecr;
+
+  tcp_fastrecovery_1_smss_off (tc);
+  tcp_fastrecovery_off (tc);
+
+  TCP_EVT_DBG (TCP_EVT_CC_EVT, tc, 3);
 }
 
 static void
-tcp_cc_rcv_ack (tcp_connection_t * tc)
+tcp_cc_rcv_ack (tcp_connection_t * tc, vlib_buffer_t * b)
 {
   u8 partial_ack;
 
   if (tcp_in_recovery (tc))
     {
-      partial_ack = seq_lt (tc->snd_una, tc->snd_una_max);
+      partial_ack = seq_lt (tc->snd_una, tc->snd_congestion);
       if (!partial_ack)
 	{
 	  /* Clear retransmitted bytes. */
-	  tc->rtx_bytes = 0;
 	  tcp_cc_recover (tc);
 	}
       else
 	{
+	  TCP_EVT_DBG (TCP_EVT_CC_PACK, tc);
+
 	  /* Clear retransmitted bytes. XXX should we clear all? */
 	  tc->rtx_bytes = 0;
 	  tc->cc_algo->rcv_cong_ack (tc, TCP_CC_PARTIALACK);
 
-	  /* Retransmit first unacked segment */
-	  tcp_retransmit_first_unacked (tc);
+	  /* In case snd_nxt is still in the past and output tries to
+	   * shove some new bytes */
+	  tc->snd_nxt = tc->snd_una;
+
+	  /* XXX need proper RFC6675 support */
+	  if (tc->sack_sb.last_sacked_bytes)
+	    {
+	      tcp_fast_retransmit (tc);
+	    }
+	  else
+	    {
+	      /* Retransmit first unacked segment */
+	      tcp_retransmit_first_unacked (tc);
+	      /* If window allows, send 1 SMSS of new data */
+	      if (seq_lt (tc->snd_nxt, tc->snd_congestion))
+		tc->snd_nxt = tc->snd_congestion;
+	    }
 	}
     }
   else
     {
       tc->cc_algo->rcv_ack (tc);
+      tc->tsecr_last_ack = tc->opt.tsecr;
+      tc->rcv_dupacks = 0;
     }
-
-  tc->rcv_dupacks = 0;
-  tc->tsecr_last_ack = tc->opt.tsecr;
 }
 
 static void
 tcp_cc_rcv_dupack (tcp_connection_t * tc, u32 ack)
 {
-  ASSERT (tc->snd_una == ack);
+//  ASSERT (seq_geq(tc->snd_una, ack));
 
   tc->rcv_dupacks++;
   if (tc->rcv_dupacks == TCP_DUPACK_THRESHOLD)
@@ -688,20 +798,39 @@ tcp_rcv_ack (tcp_connection_t * tc, vlib_buffer_t * b,
 {
   u32 new_snd_wnd;
 
-  /* If the ACK acks something not yet sent (SEG.ACK > SND.NXT) then send an
-   * ACK, drop the segment, and return  */
+  /* If the ACK acks something not yet sent (SEG.ACK > SND.NXT) */
   if (seq_gt (vnet_buffer (b)->tcp.ack_number, tc->snd_nxt))
     {
-      tcp_make_ack (tc, b);
-      *next = tcp_next_output (tc->c_is_ip4);
-      *error = TCP_ERROR_ACK_INVALID;
-      return -1;
+      /* If we have outstanding data and this is within the window, accept it,
+       * probably retransmit has timed out. Otherwise ACK segment and then
+       * drop it */
+      if (seq_gt (vnet_buffer (b)->tcp.ack_number, tc->snd_una_max))
+	{
+	  tcp_make_ack (tc, b);
+	  *next = tcp_next_output (tc->c_is_ip4);
+	  *error = TCP_ERROR_ACK_INVALID;
+	  TCP_EVT_DBG (TCP_EVT_ACK_RCV_ERR, tc, 0,
+		       vnet_buffer (b)->tcp.ack_number);
+	  return -1;
+	}
+
+      tc->snd_nxt = vnet_buffer (b)->tcp.ack_number;
+      *error = TCP_ERROR_ACK_FUTURE;
+      TCP_EVT_DBG (TCP_EVT_ACK_RCV_ERR, tc, 2,
+		   vnet_buffer (b)->tcp.ack_number);
     }
 
-  /* If old ACK, discard */
+  /* If old ACK, probably it's an old dupack */
   if (seq_lt (vnet_buffer (b)->tcp.ack_number, tc->snd_una))
     {
       *error = TCP_ERROR_ACK_OLD;
+      TCP_EVT_DBG (TCP_EVT_ACK_RCV_ERR, tc, 1,
+		   vnet_buffer (b)->tcp.ack_number);
+      if (tcp_in_fastrecovery (tc) && tc->rcv_dupacks == TCP_DUPACK_THRESHOLD)
+	{
+	  TCP_EVT_DBG (TCP_EVT_DUPACK_RCVD, tc);
+	  tcp_cc_rcv_dupack (tc, vnet_buffer (b)->tcp.ack_number);
+	}
       return -1;
     }
 
@@ -712,32 +841,40 @@ tcp_rcv_ack (tcp_connection_t * tc, vlib_buffer_t * b,
 
   if (tcp_ack_is_dupack (tc, b, new_snd_wnd))
     {
+      TCP_EVT_DBG (TCP_EVT_DUPACK_RCVD, tc, 1);
       tcp_cc_rcv_dupack (tc, vnet_buffer (b)->tcp.ack_number);
       *error = TCP_ERROR_ACK_DUP;
       return -1;
     }
 
-  /* Valid ACK */
+  /*
+   * Valid ACK
+   */
+
   tc->bytes_acked = vnet_buffer (b)->tcp.ack_number - tc->snd_una;
-  tc->snd_una = vnet_buffer (b)->tcp.ack_number;
+  tc->snd_una = vnet_buffer (b)->tcp.ack_number + tc->sack_sb.snd_una_adv;
 
-  /* Dequeue ACKed packet and update RTT */
+  /* Dequeue ACKed data and update RTT */
   tcp_dequeue_acked (tc, vnet_buffer (b)->tcp.ack_number);
-
   tcp_update_snd_wnd (tc, vnet_buffer (b)->tcp.seq_number,
 		      vnet_buffer (b)->tcp.ack_number, new_snd_wnd);
 
-  /* Updates congestion control (slow start/congestion avoidance) */
-  tcp_cc_rcv_ack (tc);
+  /* If some of our sent bytes have been acked, update cc and retransmit
+   * timer. */
+  if (tc->bytes_acked)
+    {
+      TCP_EVT_DBG (TCP_EVT_ACK_RCVD, tc, vnet_buffer (b)->tcp.ack_number);
 
-  TCP_EVT_DBG (TCP_EVT_ACK_RCVD, tc);
+      /* Updates congestion control (slow start/congestion avoidance) */
+      tcp_cc_rcv_ack (tc, b);
 
-  /* If everything has been acked, stop retransmit timer
-   * otherwise update */
-  if (tc->snd_una == tc->snd_una_max)
-    tcp_timer_reset (tc, TCP_TIMER_RETRANSMIT);
-  else
-    tcp_timer_update (tc, TCP_TIMER_RETRANSMIT, tc->rto);
+      /* If everything has been acked, stop retransmit timer
+       * otherwise update */
+      if (tc->snd_una == tc->snd_una_max)
+	tcp_retransmit_timer_reset (tc);
+      else
+	tcp_retransmit_timer_update (tc);
+    }
 
   return 0;
 }
@@ -757,9 +894,7 @@ static void
 tcp_update_sack_list (tcp_connection_t * tc, u32 start, u32 end)
 {
   sack_block_t *new_list = 0, block;
-  u32 n_elts;
   int i;
-  u8 new_head = 0;
 
   /* If the first segment is ooo add it to the list. Last write might've moved
    * rcv_nxt over the first segment. */
@@ -768,7 +903,6 @@ tcp_update_sack_list (tcp_connection_t * tc, u32 start, u32 end)
       block.start = start;
       block.end = end;
       vec_add1 (new_list, block);
-      new_head = 1;
     }
 
   /* Find the blocks still worth keeping. */
@@ -782,20 +916,19 @@ tcp_update_sack_list (tcp_connection_t * tc, u32 start, u32 end)
 	  || seq_leq (tc->snd_sacks[i].start, end))
 	continue;
 
-      /* Save subsequent segments to new SACK list. */
-      n_elts = clib_min (vec_len (tc->snd_sacks) - i,
-			 TCP_MAX_SACK_BLOCKS - new_head);
-      vec_insert_elts (new_list, &tc->snd_sacks[i], n_elts, new_head);
-      break;
+      /* Save to new SACK list. */
+      vec_add1 (new_list, tc->snd_sacks[i]);
     }
 
+  ASSERT (vec_len (new_list) < TCP_MAX_SACK_BLOCKS);
+
   /* Replace old vector with new one */
   vec_free (tc->snd_sacks);
   tc->snd_sacks = new_list;
 }
 
 /** Enqueue data for delivery to application */
-always_inline u32
+always_inline int
 tcp_session_enqueue_data (tcp_connection_t * tc, vlib_buffer_t * b,
 			  u16 data_len)
 {
@@ -812,6 +945,8 @@ tcp_session_enqueue_data (tcp_connection_t * tc, vlib_buffer_t * b,
 					 vlib_buffer_get_current (b),
 					 data_len, 1 /* queue event */ );
 
+  TCP_EVT_DBG (TCP_EVT_INPUT, tc, 0, data_len, written);
+
   /* Update rcv_nxt */
   if (PREDICT_TRUE (written == data_len))
     {
@@ -824,38 +959,61 @@ tcp_session_enqueue_data (tcp_connection_t * tc, vlib_buffer_t * b,
 
       /* Send ACK confirming the update */
       tc->flags |= TCP_CONN_SNDACK;
+    }
+  else if (written > 0)
+    {
+      /* We've written something but FIFO is probably full now */
+      tc->rcv_nxt += written;
 
-      /* Update SACK list if need be */
-      if (tcp_opts_sack_permitted (&tc->opt))
-	{
-	  /* Remove SACK blocks that have been delivered */
-	  tcp_update_sack_list (tc, tc->rcv_nxt, tc->rcv_nxt);
-	}
+      /* Depending on how fast the app is, all remaining buffers in burst will
+       * not be enqueued. Should we inform peer of the damage? XXX */
+      return TCP_ERROR_PARTIALLY_ENQUEUED;
     }
   else
     {
-      ASSERT (0);
       return TCP_ERROR_FIFO_FULL;
     }
 
+  /* Update SACK list if need be */
+  if (tcp_opts_sack_permitted (&tc->opt))
+    {
+      /* Remove SACK blocks that have been delivered */
+      tcp_update_sack_list (tc, tc->rcv_nxt, tc->rcv_nxt);
+    }
+
   return TCP_ERROR_ENQUEUED;
 }
 
 /** Enqueue out-of-order data */
-always_inline u32
+always_inline int
 tcp_session_enqueue_ooo (tcp_connection_t * tc, vlib_buffer_t * b,
 			 u16 data_len)
 {
   stream_session_t *s0;
   u32 offset, seq;
+  int rv;
+
+  /* Pure ACK. Do nothing */
+  if (PREDICT_FALSE (data_len == 0))
+    {
+      return TCP_ERROR_PURE_ACK;
+    }
 
   s0 = stream_session_get (tc->c_s_index, tc->c_thread_index);
   seq = vnet_buffer (b)->tcp.seq_number;
   offset = seq - tc->rcv_nxt;
 
-  if (svm_fifo_enqueue_with_offset (s0->server_rx_fifo, s0->pid, offset,
-				    data_len, vlib_buffer_get_current (b)))
-    return TCP_ERROR_FIFO_FULL;
+  rv = svm_fifo_enqueue_with_offset (s0->server_rx_fifo, s0->pid, offset,
+				     data_len, vlib_buffer_get_current (b));
+
+  /* Nothing written */
+  if (rv)
+    {
+      TCP_EVT_DBG (TCP_EVT_INPUT, tc, 1, data_len, 0);
+      return TCP_ERROR_FIFO_FULL;
+    }
+
+  TCP_EVT_DBG (TCP_EVT_INPUT, tc, 1, data_len, data_len);
 
   /* Update SACK list if in use */
   if (tcp_opts_sack_permitted (&tc->opt))
@@ -875,20 +1033,23 @@ tcp_session_enqueue_ooo (tcp_connection_t * tc, vlib_buffer_t * b,
 }
 
 /**
- * Check if ACK could be delayed. DELACK timer is set only after frame is
- * processed so this can return true for a full bursts of packets.
+ * Check if ACK could be delayed. If ack can be delayed, it should return
+ * true for a full frame. If we're always acking return 0.
  */
 always_inline int
 tcp_can_delack (tcp_connection_t * tc)
 {
-  /* If there's no DELACK timer set and the last window sent wasn't 0 we
-   * can safely delay. */
-  if (!tcp_timer_is_active (tc, TCP_TIMER_DELACK)
-      && (tc->flags & TCP_CONN_SENT_RCV_WND0) == 0
-      && (tc->flags & TCP_CONN_SNDACK) == 0)
-    return 1;
+  /* Send ack if ... */
+  if (TCP_ALWAYS_ACK
+      /* just sent a rcv wnd 0 */
+      || (tc->flags & TCP_CONN_SENT_RCV_WND0) != 0
+      /* constrained to send ack */
+      || (tc->flags & TCP_CONN_SNDACK) != 0
+      /* we're almost out of tx wnd */
+      || tcp_available_snd_space (tc) < 2 * tc->snd_mss)
+    return 0;
 
-  return 0;
+  return 1;
 }
 
 static int
@@ -900,23 +1061,33 @@ tcp_segment_rcv (tcp_main_t * tm, tcp_connection_t * tc, vlib_buffer_t * b,
   /* Handle out-of-order data */
   if (PREDICT_FALSE (vnet_buffer (b)->tcp.seq_number != tc->rcv_nxt))
     {
+      /* Old sequence numbers allowed through because they overlapped
+       * the rx window */
+      if (seq_lt (vnet_buffer (b)->tcp.seq_number, tc->rcv_nxt))
+	{
+	  error = TCP_ERROR_SEGMENT_OLD;
+	  *next0 = TCP_NEXT_DROP;
+	  goto done;
+	}
+
       error = tcp_session_enqueue_ooo (tc, b, n_data_bytes);
 
-      /* Don't send more than 3 dupacks per burst
-       * XXX decide if this is good */
-      if (tc->snt_dupacks < 3)
-	{
-	  /* RFC2581: Send DUPACK for fast retransmit */
-	  tcp_make_ack (tc, b);
-	  *next0 = tcp_next_output (tc->c_is_ip4);
+      /* N.B. Should not filter burst of dupacks. Two issues 1) dupacks open
+       * cwnd on remote peer when congested 2) acks leaving should have the
+       * latest rcv_wnd since the burst may eaten up all of it, so only the
+       * old ones could be filtered.
+       */
 
-	  /* Mark as DUPACK. We may filter these in output if
-	   * the burst fills the holes. */
-	  vnet_buffer (b)->tcp.flags = TCP_BUF_FLAG_DUPACK;
+      /* RFC2581: Send DUPACK for fast retransmit */
+      tcp_make_ack (tc, b);
+      *next0 = tcp_next_output (tc->c_is_ip4);
 
-	  tc->snt_dupacks++;
-	}
+      /* Mark as DUPACK. We may filter these in output if
+       * the burst fills the holes. */
+      if (n_data_bytes)
+	vnet_buffer (b)->tcp.flags = TCP_BUF_FLAG_DUPACK;
 
+      TCP_EVT_DBG (TCP_EVT_DUPACK_SENT, tc);
       goto done;
     }
 
@@ -924,63 +1095,45 @@ tcp_segment_rcv (tcp_main_t * tm, tcp_connection_t * tc, vlib_buffer_t * b,
    * segments can be enqueued after fifo tail offset changes. */
   error = tcp_session_enqueue_data (tc, b, n_data_bytes);
 
-  TCP_EVT_DBG (TCP_EVT_INPUT, tc, n_data_bytes);
+  if (n_data_bytes == 0)
+    {
+      *next0 = TCP_NEXT_DROP;
+      goto done;
+    }
+
+  if (PREDICT_FALSE (error == TCP_ERROR_FIFO_FULL))
+    *next0 = TCP_NEXT_DROP;
 
   /* Check if ACK can be delayed */
-  if (tcp_can_delack (tc))
+  if (!tcp_can_delack (tc))
     {
-      /* Nothing to do for pure ACKs */
+      /* Nothing to do for pure ACKs XXX */
       if (n_data_bytes == 0)
 	goto done;
 
-      /* If connection has not been previously marked for delay ack
-       * add it to the list and flag it */
-      if (!tc->flags & TCP_CONN_DELACK)
-	{
-	  vec_add1 (tm->delack_connections[tc->c_thread_index],
-		    tc->c_c_index);
-	  tc->flags |= TCP_CONN_DELACK;
-	}
+      *next0 = tcp_next_output (tc->c_is_ip4);
+      tcp_make_ack (tc, b);
     }
   else
     {
-      /* Check if a packet has already been enqueued to output for burst.
-       * If yes, then drop this one, otherwise, let it pass through to
-       * output */
-      if ((tc->flags & TCP_CONN_BURSTACK) == 0)
-	{
-	  *next0 = tcp_next_output (tc->c_is_ip4);
-	  tcp_make_ack (tc, b);
-	  error = TCP_ERROR_ENQUEUED;
-
-	  /* TODO: maybe add counter to ensure N acks will be sent/burst */
-	  tc->flags |= TCP_CONN_BURSTACK;
-	}
+      if (!tcp_timer_is_active (tc, TCP_TIMER_DELACK))
+	tcp_timer_set (tc, TCP_TIMER_DELACK, TCP_DELACK_TIME);
     }
 
 done:
   return error;
 }
 
-void
-delack_timers_init (tcp_main_t * tm, u32 thread_index)
+always_inline void
+tcp_established_inc_counter (vlib_main_t * vm, u8 is_ip4, u8 evt, u8 val)
 {
-  tcp_connection_t *tc;
-  u32 i, *conns;
-  tw_timer_wheel_16t_2w_512sl_t *tw;
-
-  tw = &tm->timer_wheels[thread_index];
-  conns = tm->delack_connections[thread_index];
-  for (i = 0; i < vec_len (conns); i++)
-    {
-      tc = pool_elt_at_index (tm->connections[thread_index], conns[i]);
-      ASSERT (0 != tc);
+  if (PREDICT_TRUE (!val))
+    return;
 
-      tc->timers[TCP_TIMER_DELACK]
-	= tw_timer_start_16t_2w_512sl (tw, conns[i],
-				       TCP_TIMER_DELACK, TCP_DELACK_TIME);
-    }
-  vec_reset_length (tm->delack_connections[thread_index]);
+  if (is_ip4)
+    vlib_node_increment_counter (vm, tcp4_established_node.index, evt, val);
+  else
+    vlib_node_increment_counter (vm, tcp6_established_node.index, evt, val);
 }
 
 always_inline uword
@@ -1027,7 +1180,7 @@ tcp46_established_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
 	  if (PREDICT_FALSE (tc0 == 0))
 	    {
 	      error0 = TCP_ERROR_INVALID_CONNECTION;
-	      goto drop;
+	      goto done;
 	    }
 
 	  /* Checksum computed by ipx_local no need to compute again */
@@ -1061,18 +1214,22 @@ tcp46_established_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
 	  if (PREDICT_FALSE (tcp_segment_validate (vm, tc0, b0, th0, &next0)))
 	    {
 	      error0 = TCP_ERROR_SEGMENT_INVALID;
-	      goto drop;
+	      TCP_EVT_DBG (TCP_EVT_SEG_INVALID, tc0,
+			   vnet_buffer (b0)->tcp.seq_number,
+			   vnet_buffer (b0)->tcp.seq_end);
+	      goto done;
 	    }
 
 	  /* 5: check the ACK field  */
 	  if (tcp_rcv_ack (tc0, b0, th0, &next0, &error0))
 	    {
-	      goto drop;
+	      goto done;
 	    }
 
 	  /* 6: check the URG bit TODO */
 
 	  /* 7: process the segment text */
+
 	  vlib_buffer_advance (b0, n_advance_bytes0);
 	  error0 = tcp_segment_rcv (tm, tc0, b0, n_data_bytes0, &next0);
 
@@ -1088,7 +1245,7 @@ tcp46_established_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
 	      tcp_timer_set (tc0, TCP_TIMER_WAITCLOSE, TCP_CLOSEWAIT_TIME);
 	    }
 
-	drop:
+	done:
 	  b0->error = node->errors[error0];
 	  if (PREDICT_FALSE (b0->flags & VLIB_BUFFER_IS_TRACED))
 	    {
@@ -1103,17 +1260,7 @@ tcp46_established_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
     }
 
   errors = session_manager_flush_enqueue_events (my_thread_index);
-  if (errors)
-    {
-      if (is_ip4)
-	vlib_node_increment_counter (vm, tcp4_established_node.index,
-				     TCP_ERROR_EVENT_FIFO_FULL, errors);
-      else
-	vlib_node_increment_counter (vm, tcp6_established_node.index,
-				     TCP_ERROR_EVENT_FIFO_FULL, errors);
-    }
-
-  delack_timers_init (tm, my_thread_index);
+  tcp_established_inc_counter (vm, is_ip4, TCP_ERROR_EVENT_FIFO_FULL, errors);
 
   return from_frame->n_vectors;
 }
@@ -1602,7 +1749,7 @@ tcp46_rcv_process_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
 	      stream_session_accept_notify (&tc0->connection);
 
 	      /* Reset SYN-ACK retransmit timer */
-	      tcp_timer_reset (tc0, TCP_TIMER_RETRANSMIT);
+	      tcp_retransmit_timer_reset (tc0);
 	      break;
 	    case TCP_STATE_ESTABLISHED:
 	      /* We can get packets in established state here because they
@@ -1668,7 +1815,7 @@ tcp46_rcv_process_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
 	      tcp_timer_update (tc0, TCP_TIMER_WAITCLOSE, TCP_CLEANUP_TIME);
 
 	      /* Stop retransmit */
-	      tcp_timer_reset (tc0, TCP_TIMER_RETRANSMIT);
+	      tcp_retransmit_timer_reset (tc0);
 
 	      goto drop;
 
diff --git a/src/vnet/tcp/tcp_output.c b/src/vnet/tcp/tcp_output.c
index 114a5b9e..a671f728 100644
--- a/src/vnet/tcp/tcp_output.c
+++ b/src/vnet/tcp/tcp_output.c
@@ -125,15 +125,33 @@ tcp_initial_window_to_advertise (tcp_connection_t * tc)
 u32
 tcp_window_to_advertise (tcp_connection_t * tc, tcp_state_t state)
 {
-  u32 available_space, max_fifo, observed_wnd;
-
   if (state < TCP_STATE_ESTABLISHED)
     return tcp_initial_window_to_advertise (tc);
 
+  tcp_update_rcv_wnd (tc);
+
+  if (tc->rcv_wnd == 0)
+    {
+      tc->flags |= TCP_CONN_SENT_RCV_WND0;
+    }
+  else
+    {
+      tc->flags &= ~TCP_CONN_SENT_RCV_WND0;
+    }
+
+  return tc->rcv_wnd >> tc->rcv_wscale;
+}
+
+void
+tcp_update_rcv_wnd (tcp_connection_t * tc)
+{
+  i32 observed_wnd;
+  u32 available_space, max_fifo, wnd;
+
   /*
    * Figure out how much space we have available
    */
-  available_space = stream_session_max_enqueue (&tc->connection);
+  available_space = stream_session_max_rx_enqueue (&tc->connection);
   max_fifo = stream_session_fifo_size (&tc->connection);
 
   ASSERT (tc->opt.mss < max_fifo);
@@ -145,23 +163,25 @@ tcp_window_to_advertise (tcp_connection_t * tc, tcp_state_t state)
    * Use the above and what we know about what we've previously advertised
    * to compute the new window
    */
-  observed_wnd = tc->rcv_wnd - (tc->rcv_nxt - tc->rcv_las);
+  observed_wnd = (i32) tc->rcv_wnd - (tc->rcv_nxt - tc->rcv_las);
+  if (observed_wnd < 0)
+    observed_wnd = 0;
 
   /* Bad. Thou shalt not shrink */
   if (available_space < observed_wnd)
     {
-      if (available_space == 0)
-	clib_warning ("Didn't shrink rcv window despite not having space");
+      /* Does happen! */
+      wnd = observed_wnd;
     }
-
-  tc->rcv_wnd = clib_min (available_space, TCP_WND_MAX << tc->rcv_wscale);
-
-  if (tc->rcv_wnd == 0)
+  else
     {
-      tc->flags |= TCP_CONN_SENT_RCV_WND0;
+      wnd = available_space;
     }
 
-  return tc->rcv_wnd >> tc->rcv_wscale;
+  if (wnd && ((wnd << tc->rcv_wscale) >> tc->rcv_wscale != wnd))
+    wnd += 1 << tc->rcv_wscale;
+
+  tc->rcv_wnd = clib_min (wnd, TCP_WND_MAX << tc->rcv_wscale);
 }
 
 /**
@@ -363,7 +383,7 @@ tcp_make_options (tcp_connection_t * tc, tcp_options_t * opts,
 #define tcp_get_free_buffer_index(tm, bidx)                             \
 do {                                                                    \
   u32 *my_tx_buffers, n_free_buffers;                                   \
-  u32 cpu_index = tm->vlib_main->cpu_index;                             \
+  u32 cpu_index = os_get_cpu_number();                             	\
   my_tx_buffers = tm->tx_buffers[cpu_index];                            \
   if (PREDICT_FALSE(vec_len (my_tx_buffers) == 0))                      \
     {                                                                   \
@@ -381,6 +401,14 @@ do {                                                                    \
   _vec_len (my_tx_buffers) -= 1;                                        \
 } while (0)
 
+#define tcp_return_buffer(tm)						\
+do {									\
+  u32 *my_tx_buffers;							\
+  u32 cpu_index = os_get_cpu_number();                             	\
+  my_tx_buffers = tm->tx_buffers[cpu_index];                          	\
+  _vec_len (my_tx_buffers) +=1;						\
+} while (0)
+
 always_inline void
 tcp_reuse_buffer (vlib_main_t * vm, vlib_buffer_t * b)
 {
@@ -421,8 +449,6 @@ tcp_make_ack_i (tcp_connection_t * tc, vlib_buffer_t * b, tcp_state_t state,
 			     tc->rcv_nxt, tcp_hdr_opts_len, flags, wnd);
 
   tcp_options_write ((u8 *) (th + 1), snd_opts);
-
-  /* Mark as ACK */
   vnet_buffer (b)->tcp.connection_index = tc->c_c_index;
 }
 
@@ -432,12 +458,12 @@ tcp_make_ack_i (tcp_connection_t * tc, vlib_buffer_t * b, tcp_state_t state,
 void
 tcp_make_ack (tcp_connection_t * tc, vlib_buffer_t * b)
 {
-  tcp_main_t *tm = vnet_get_tcp_main ();
-  vlib_main_t *vm = tm->vlib_main;
+  vlib_main_t *vm = vlib_get_main ();
 
   tcp_reuse_buffer (vm, b);
   tcp_make_ack_i (tc, b, TCP_STATE_ESTABLISHED, TCP_FLAG_ACK);
   vnet_buffer (b)->tcp.flags = TCP_BUF_FLAG_ACK;
+  TCP_EVT_DBG (TCP_EVT_ACK_SENT, tc);
 }
 
 /**
@@ -446,8 +472,7 @@ tcp_make_ack (tcp_connection_t * tc, vlib_buffer_t * b)
 void
 tcp_make_fin (tcp_connection_t * tc, vlib_buffer_t * b)
 {
-  tcp_main_t *tm = vnet_get_tcp_main ();
-  vlib_main_t *vm = tm->vlib_main;
+  vlib_main_t *vm = vlib_get_main ();
   u8 flags = 0;
 
   tcp_reuse_buffer (vm, b);
@@ -467,8 +492,7 @@ tcp_make_fin (tcp_connection_t * tc, vlib_buffer_t * b)
 void
 tcp_make_synack (tcp_connection_t * tc, vlib_buffer_t * b)
 {
-  tcp_main_t *tm = vnet_get_tcp_main ();
-  vlib_main_t *vm = tm->vlib_main;
+  vlib_main_t *vm = vlib_get_main ();
   tcp_options_t _snd_opts, *snd_opts = &_snd_opts;
   u8 tcp_opts_len, tcp_hdr_opts_len;
   tcp_header_t *th;
@@ -631,7 +655,7 @@ tcp_send_reset (vlib_buffer_t * pkt, u8 is_ip4)
   vlib_buffer_t *b;
   u32 bi;
   tcp_main_t *tm = vnet_get_tcp_main ();
-  vlib_main_t *vm = tm->vlib_main;
+  vlib_main_t *vm = vlib_get_main ();
   u8 tcp_hdr_len, flags = 0;
   tcp_header_t *th, *pkt_th;
   u32 seq, ack;
@@ -736,7 +760,7 @@ tcp_send_syn (tcp_connection_t * tc)
   vlib_buffer_t *b;
   u32 bi;
   tcp_main_t *tm = vnet_get_tcp_main ();
-  vlib_main_t *vm = tm->vlib_main;
+  vlib_main_t *vm = vlib_get_main ();
   u8 tcp_hdr_opts_len, tcp_opts_len;
   tcp_header_t *th;
   u32 time_now;
@@ -795,9 +819,9 @@ tcp_enqueue_to_output (vlib_main_t * vm, vlib_buffer_t * b, u32 bi, u8 is_ip4)
 
   /* Decide where to send the packet */
   next_index = is_ip4 ? tcp4_output_node.index : tcp6_output_node.index;
-  f = vlib_get_frame_to_node (vm, next_index);
 
   /* Enqueue the packet */
+  f = vlib_get_frame_to_node (vm, next_index);
   to_next = vlib_frame_vector_args (f);
   to_next[0] = bi;
   f->n_vectors = 1;
@@ -813,7 +837,7 @@ tcp_send_fin (tcp_connection_t * tc)
   vlib_buffer_t *b;
   u32 bi;
   tcp_main_t *tm = vnet_get_tcp_main ();
-  vlib_main_t *vm = tm->vlib_main;
+  vlib_main_t *vm = vlib_get_main ();
 
   tcp_get_free_buffer_index (tm, &bi);
   b = vlib_get_buffer (vm, bi);
@@ -884,22 +908,21 @@ tcp_push_hdr_i (tcp_connection_t * tc, vlib_buffer_t * b,
   vnet_buffer (b)->tcp.connection_index = tc->c_c_index;
 
   tc->snd_nxt += data_len;
+  /* TODO this is updated in output as well ... */
+  if (tc->snd_nxt > tc->snd_una_max)
+    tc->snd_una_max = tc->snd_nxt;
   TCP_EVT_DBG (TCP_EVT_PKTIZE, tc);
 }
 
-/* Send delayed ACK when timer expires */
 void
-tcp_timer_delack_handler (u32 index)
+tcp_send_ack (tcp_connection_t * tc)
 {
   tcp_main_t *tm = vnet_get_tcp_main ();
-  vlib_main_t *vm = tm->vlib_main;
-  u32 thread_index = os_get_cpu_number ();
-  tcp_connection_t *tc;
+  vlib_main_t *vm = vlib_get_main ();
+
   vlib_buffer_t *b;
   u32 bi;
 
-  tc = tcp_connection_get (index, thread_index);
-
   /* Get buffer */
   tcp_get_free_buffer_index (tm, &bi);
   b = vlib_get_buffer (vm, bi);
@@ -907,12 +930,22 @@ tcp_timer_delack_handler (u32 index)
   /* Fill in the ACK */
   tcp_make_ack (tc, b);
 
-  tc->timers[TCP_TIMER_DELACK] = TCP_TIMER_HANDLE_INVALID;
-  tc->flags &= ~TCP_CONN_DELACK;
-
   tcp_enqueue_to_output (vm, b, bi, tc->c_is_ip4);
 }
 
+/* Send delayed ACK when timer expires */
+void
+tcp_timer_delack_handler (u32 index)
+{
+  u32 thread_index = os_get_cpu_number ();
+  tcp_connection_t *tc;
+
+  tc = tcp_connection_get (index, thread_index);
+  tc->timers[TCP_TIMER_DELACK] = TCP_TIMER_HANDLE_INVALID;
+//  tc->flags &= ~TCP_CONN_DELACK;
+  tcp_send_ack (tc);
+}
+
 /** Build a retransmit segment
  *
  * @return the number of bytes in the segment or 0 if there's nothing to
@@ -920,59 +953,74 @@ tcp_timer_delack_handler (u32 index)
  * */
 u32
 tcp_prepare_retransmit_segment (tcp_connection_t * tc, vlib_buffer_t * b,
-				u32 max_bytes)
+				u32 offset, u32 max_bytes)
 {
-  tcp_main_t *tm = vnet_get_tcp_main ();
-  vlib_main_t *vm = tm->vlib_main;
-  u32 n_bytes, offset = 0;
-  sack_scoreboard_hole_t *hole;
-  u32 hole_size;
+  vlib_main_t *vm = vlib_get_main ();
+  u32 n_bytes = 0;
 
   tcp_reuse_buffer (vm, b);
 
   ASSERT (tc->state >= TCP_STATE_ESTABLISHED);
   ASSERT (max_bytes != 0);
 
-  if (tcp_opts_sack_permitted (&tc->opt))
-    {
-      /* XXX get first hole not retransmitted yet  */
-      hole = scoreboard_first_hole (&tc->sack_sb);
-      if (!hole)
-	return 0;
-
-      offset = hole->start - tc->snd_una;
-      hole_size = hole->end - hole->start;
+  max_bytes = clib_min (tc->snd_mss, max_bytes);
 
-      ASSERT (hole_size);
+  /* Start is beyond snd_congestion */
+  if (seq_geq (tc->snd_una + offset, tc->snd_congestion))
+    goto done;
 
-      if (hole_size < max_bytes)
-	max_bytes = hole_size;
-    }
-  else
+  /* Don't overshoot snd_congestion */
+  if (seq_gt (tc->snd_nxt + max_bytes, tc->snd_congestion))
     {
-      if (seq_geq (tc->snd_nxt, tc->snd_una_max))
-	return 0;
+      max_bytes = tc->snd_congestion - tc->snd_nxt;
+      if (max_bytes == 0)
+	goto done;
     }
 
+  ASSERT (max_bytes <= tc->snd_mss);
+
   n_bytes = stream_session_peek_bytes (&tc->connection,
 				       vlib_buffer_get_current (b), offset,
 				       max_bytes);
   ASSERT (n_bytes != 0);
-
+  b->current_length = n_bytes;
   tcp_push_hdr_i (tc, b, tc->state);
 
+done:
+  TCP_EVT_DBG (TCP_EVT_CC_RTX, tc, offset, n_bytes);
   return n_bytes;
 }
 
+/**
+ * Reset congestion control, switch cwnd to loss window and try again.
+ */
+static void
+tcp_rtx_timeout_cc_recover (tcp_connection_t * tc)
+{
+  /* Cleanly recover cc (also clears up fast retransmit) */
+  if (tcp_in_fastrecovery (tc))
+    {
+      tcp_cc_recover (tc);
+    }
+  else
+    {
+      tc->ssthresh = clib_max (tcp_flight_size (tc) / 2, 2 * tc->snd_mss);
+    }
+
+  /* Start again from the beginning */
+  tc->cwnd = tcp_loss_wnd (tc);
+  tc->snd_congestion = tc->snd_una_max;
+}
+
 static void
 tcp_timer_retransmit_handler_i (u32 index, u8 is_syn)
 {
   tcp_main_t *tm = vnet_get_tcp_main ();
-  vlib_main_t *vm = tm->vlib_main;
+  vlib_main_t *vm = vlib_get_main ();
   u32 thread_index = os_get_cpu_number ();
   tcp_connection_t *tc;
   vlib_buffer_t *b;
-  u32 bi, max_bytes, snd_space;
+  u32 bi, snd_space, n_bytes;
 
   if (is_syn)
     {
@@ -998,26 +1046,43 @@ tcp_timer_retransmit_handler_i (u32 index, u8 is_syn)
 
   if (tc->state >= TCP_STATE_ESTABLISHED)
     {
-      tcp_fastrecovery_off (tc);
+      /* First retransmit timeout */
+      if (tc->rto_boff == 1)
+	tcp_rtx_timeout_cc_recover (tc);
 
       /* Exponential backoff */
       tc->rto = clib_min (tc->rto << 1, TCP_RTO_MAX);
 
       /* Figure out what and how many bytes we can send */
       snd_space = tcp_available_snd_space (tc);
-      max_bytes = clib_min (tc->snd_mss, snd_space);
 
-      if (max_bytes == 0)
+      TCP_EVT_DBG (TCP_EVT_CC_EVT, tc, 1);
+
+      if (snd_space == 0)
 	{
 	  clib_warning ("no wnd to retransmit");
+	  tcp_return_buffer (tm);
+
+	  /* Force one segment */
+	  tcp_retransmit_first_unacked (tc);
+
+	  /* Re-enable retransmit timer. Output may be unwilling
+	   * to do it for us */
+	  tcp_retransmit_timer_set (tc);
+
 	  return;
 	}
-      tcp_prepare_retransmit_segment (tc, b, max_bytes);
+      else
+	{
+	  /* No fancy recovery for now! */
+	  n_bytes = tcp_prepare_retransmit_segment (tc, b, 0, snd_space);
+	  scoreboard_clear (&tc->sack_sb);
 
-      tc->rtx_bytes += max_bytes;
+	  if (n_bytes == 0)
+	    return;
 
-      /* No fancy recovery for now! */
-      scoreboard_clear (&tc->sack_sb);
+	  tc->rtx_bytes += n_bytes;
+	}
     }
   else
     {
@@ -1072,63 +1137,110 @@ tcp_timer_retransmit_syn_handler (u32 index)
 }
 
 /**
- * Retansmit first unacked segment */
+ * Retransmit first unacked segment
+ */
 void
 tcp_retransmit_first_unacked (tcp_connection_t * tc)
 {
   tcp_main_t *tm = vnet_get_tcp_main ();
-  u32 snd_nxt = tc->snd_nxt;
+  vlib_main_t *vm = vlib_get_main ();
   vlib_buffer_t *b;
-  u32 bi;
+  u32 bi, n_bytes;
 
   tc->snd_nxt = tc->snd_una;
 
   /* Get buffer */
   tcp_get_free_buffer_index (tm, &bi);
-  b = vlib_get_buffer (tm->vlib_main, bi);
+  b = vlib_get_buffer (vm, bi);
+
+  TCP_EVT_DBG (TCP_EVT_CC_EVT, tc, 2);
 
-  tcp_prepare_retransmit_segment (tc, b, tc->snd_mss);
-  tcp_enqueue_to_output (tm->vlib_main, b, bi, tc->c_is_ip4);
+  n_bytes = tcp_prepare_retransmit_segment (tc, b, 0, tc->snd_mss);
+  if (n_bytes == 0)
+    return;
 
-  tc->snd_nxt = snd_nxt;
-  tc->rtx_bytes += tc->snd_mss;
+  tcp_enqueue_to_output (vm, b, bi, tc->c_is_ip4);
+  tc->rtx_bytes += n_bytes;
+}
+
+sack_scoreboard_hole_t *
+scoreboard_first_rtx_hole (sack_scoreboard_t * sb)
+{
+  sack_scoreboard_hole_t *hole = 0;
+
+//  hole = scoreboard_first_hole (&tc->sack_sb);
+//  if (hole)
+//    {
+//
+//      offset = hole->start - tc->snd_una;
+//      hole_size = hole->end - hole->start;
+//
+//      ASSERT(hole_size);
+//
+//      if (hole_size < max_bytes)
+//      max_bytes = hole_size;
+//    }
+  return hole;
 }
 
+/**
+ * Do fast retransmit.
+ */
 void
 tcp_fast_retransmit (tcp_connection_t * tc)
 {
   tcp_main_t *tm = vnet_get_tcp_main ();
-  u32 snd_space, max_bytes, n_bytes, bi;
+  vlib_main_t *vm = vlib_get_main ();
+  u32 bi;
+  int snd_space;
+  u32 n_written = 0, offset = 0;
   vlib_buffer_t *b;
+  u8 use_sacks = 0;
 
   ASSERT (tcp_in_fastrecovery (tc));
 
-  clib_warning ("fast retransmit!");
-
   /* Start resending from first un-acked segment */
   tc->snd_nxt = tc->snd_una;
 
   snd_space = tcp_available_snd_space (tc);
+  TCP_EVT_DBG (TCP_EVT_CC_EVT, tc, 0);
+
+  /* If we have SACKs use them */
+  if (tcp_opts_sack_permitted (&tc->opt)
+      && scoreboard_first_hole (&tc->sack_sb))
+    use_sacks = 0;
 
-  while (snd_space)
+  while (snd_space > 0)
     {
       tcp_get_free_buffer_index (tm, &bi);
-      b = vlib_get_buffer (tm->vlib_main, bi);
+      b = vlib_get_buffer (vm, bi);
+
+      if (use_sacks)
+	{
+	  scoreboard_first_rtx_hole (&tc->sack_sb);
+	}
+      else
+	{
+	  offset += n_written;
+	}
 
-      max_bytes = clib_min (tc->snd_mss, snd_space);
-      n_bytes = tcp_prepare_retransmit_segment (tc, b, max_bytes);
+      n_written = tcp_prepare_retransmit_segment (tc, b, offset, snd_space);
 
       /* Nothing left to retransmit */
-      if (n_bytes == 0)
-	return;
-
-      tcp_enqueue_to_output (tm->vlib_main, b, bi, tc->c_is_ip4);
+      if (n_written == 0)
+	{
+	  tcp_return_buffer (tm);
+	  break;
+	}
 
-      snd_space -= n_bytes;
+      tcp_enqueue_to_output (vm, b, bi, tc->c_is_ip4);
+      tc->rtx_bytes += n_written;
+      snd_space -= n_written;
     }
 
-  /* If window allows, send new data */
-  tc->snd_nxt = tc->snd_una_max;
+  /* If window allows, send 1 SMSS of new data */
+  if (seq_lt (tc->snd_nxt, tc->snd_congestion))
+    tc->snd_nxt = tc->snd_congestion;
 }
 
 always_inline u32
@@ -1209,8 +1321,6 @@ tcp46_output_inline (vlib_main_t * vm,
 	  if (PREDICT_FALSE
 	      (vnet_buffer (b0)->tcp.flags & TCP_BUF_FLAG_DUPACK))
 	    {
-	      ASSERT (tc0->snt_dupacks > 0);
-	      tc0->snt_dupacks--;
 	      if (!tcp_session_has_ooo_data (tc0))
 		{
 		  error0 = TCP_ERROR_FILTERED_DUPACKS;
@@ -1223,8 +1333,7 @@ tcp46_output_inline (vlib_main_t * vm,
 	  tc0->rcv_las = tc0->rcv_nxt;
 
 	  /* Stop DELACK timer and fix flags */
-	  tc0->flags &=
-	    ~(TCP_CONN_SNDACK | TCP_CONN_DELACK | TCP_CONN_BURSTACK);
+	  tc0->flags &= ~(TCP_CONN_SNDACK);
 	  if (tcp_timer_is_active (tc0, TCP_TIMER_DELACK))
 	    {
 	      tcp_timer_reset (tc0, TCP_TIMER_DELACK);
diff --git a/src/vnet/tcp/tcp_packet.h b/src/vnet/tcp/tcp_packet.h
index 866c5fd6..4f28cf32 100644
--- a/src/vnet/tcp/tcp_packet.h
+++ b/src/vnet/tcp/tcp_packet.h
@@ -137,7 +137,7 @@ enum
 typedef struct _sack_block
 {
   u32 start;		/**< Start sequence number */
-  u32 end;		/**< End sequence number */
+  u32 end;		/**< End sequence number (first outside) */
 } sack_block_t;
 
 typedef struct
diff --git a/src/vnet/tcp/tcp_test.c b/src/vnet/tcp/tcp_test.c
new file mode 100644
index 00000000..0725bb04
--- /dev/null
+++ b/src/vnet/tcp/tcp_test.c
@@ -0,0 +1,216 @@
+/*
+ * Copyright (c) 2017 Cisco and/or its affiliates.
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at:
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include <vnet/tcp/tcp.h>
+
+#define TCP_TEST_I(_cond, _comment, _args...)			\
+({								\
+  int _evald = (_cond);						\
+  if (!(_evald)) {						\
+    fformat(stderr, "FAIL:%d: " _comment "\n",			\
+	    __LINE__, ##_args);					\
+  } else {							\
+    fformat(stderr, "PASS:%d: " _comment "\n",			\
+	    __LINE__, ##_args);					\
+  }								\
+  _evald;							\
+})
+
+#define TCP_TEST(_cond, _comment, _args...)			\
+{								\
+    if (!TCP_TEST_I(_cond, _comment, ##_args)) {		\
+	return 1;                                               \
+    }								\
+}
+
+static int
+tcp_test_sack ()
+{
+  tcp_connection_t _tc, *tc = &_tc;
+  sack_scoreboard_t *sb = &tc->sack_sb;
+  sack_block_t *sacks = 0, block;
+  sack_scoreboard_hole_t *hole;
+  int i;
+
+  memset (tc, 0, sizeof (*tc));
+
+  tc->snd_una = 0;
+  tc->snd_una_max = 1000;
+  tc->snd_nxt = 1000;
+  tc->opt.flags |= TCP_OPTS_FLAG_SACK;
+  scoreboard_init (&tc->sack_sb);
+
+  for (i = 0; i < 1000 / 100; i++)
+    {
+      block.start = i * 100;
+      block.end = (i + 1) * 100;
+      vec_add1 (sacks, block);
+    }
+
+  /*
+   * Inject even blocks
+   */
+
+  for (i = 0; i < 1000 / 200; i++)
+    {
+      vec_add1 (tc->opt.sacks, sacks[i * 2]);
+    }
+  tc->opt.n_sack_blocks = vec_len (tc->opt.sacks);
+  tcp_rcv_sacks (tc, 0);
+
+  TCP_TEST ((pool_elts (sb->holes) == 5),
+	    "scoreboard has %d elements", pool_elts (sb->holes));
+
+  /* First SACK block should be rejected */
+  hole = scoreboard_first_hole (sb);
+  TCP_TEST ((hole->start == 0 && hole->end == 200),
+	    "first hole start %u end %u", hole->start, hole->end);
+  hole = scoreboard_last_hole (sb);
+  TCP_TEST ((hole->start == 900 && hole->end == 1000),
+	    "last hole start %u end %u", hole->start, hole->end);
+  TCP_TEST ((sb->sacked_bytes == 400), "sacked bytes %d", sb->sacked_bytes);
+  TCP_TEST ((sb->snd_una_adv == 0), "snd_una_adv %u", sb->snd_una_adv);
+  TCP_TEST ((sb->last_sacked_bytes == 400),
+	    "last sacked bytes %d", sb->last_sacked_bytes);
+
+  /*
+   * Inject odd blocks
+   */
+
+  vec_reset_length (tc->opt.sacks);
+  for (i = 0; i < 1000 / 200; i++)
+    {
+      vec_add1 (tc->opt.sacks, sacks[i * 2 + 1]);
+    }
+  tc->opt.n_sack_blocks = vec_len (tc->opt.sacks);
+  tcp_rcv_sacks (tc, 0);
+
+  hole = scoreboard_first_hole (sb);
+  TCP_TEST ((pool_elts (sb->holes) == 1),
+	    "scoreboard has %d holes", pool_elts (sb->holes));
+  TCP_TEST ((hole->start == 0 && hole->end == 100),
+	    "first hole start %u end %u", hole->start, hole->end);
+  TCP_TEST ((sb->sacked_bytes == 900), "sacked bytes %d", sb->sacked_bytes);
+  TCP_TEST ((sb->snd_una_adv == 0), "snd_una_adv %u", sb->snd_una_adv);
+  TCP_TEST ((sb->max_byte_sacked == 1000),
+	    "max sacked byte %u", sb->max_byte_sacked);
+  TCP_TEST ((sb->last_sacked_bytes == 500),
+	    "last sacked bytes %d", sb->last_sacked_bytes);
+
+  /*
+   *  Ack until byte 100, all bytes are now acked + sacked
+   */
+  tcp_rcv_sacks (tc, 100);
+
+  TCP_TEST ((pool_elts (sb->holes) == 0),
+	    "scoreboard has %d elements", pool_elts (sb->holes));
+  TCP_TEST ((sb->snd_una_adv == 900),
+	    "snd_una_adv after ack %u", sb->snd_una_adv);
+  TCP_TEST ((sb->max_byte_sacked == 1000),
+	    "max sacked byte %u", sb->max_byte_sacked);
+  TCP_TEST ((sb->sacked_bytes == 0), "sacked bytes %d", sb->sacked_bytes);
+  TCP_TEST ((sb->last_sacked_bytes == 0),
+	    "last sacked bytes %d", sb->last_sacked_bytes);
+
+  /*
+   * Add new block
+   */
+
+  vec_reset_length (tc->opt.sacks);
+
+  block.start = 1200;
+  block.end = 1300;
+  vec_add1 (tc->opt.sacks, block);
+
+  tc->snd_una_max = 1500;
+  tc->snd_una = 1000;
+  tc->snd_nxt = 1500;
+  tcp_rcv_sacks (tc, 1000);
+
+  TCP_TEST ((sb->snd_una_adv == 0),
+	    "snd_una_adv after ack %u", sb->snd_una_adv);
+  TCP_TEST ((pool_elts (sb->holes) == 2),
+	    "scoreboard has %d holes", pool_elts (sb->holes));
+  hole = scoreboard_first_hole (sb);
+  TCP_TEST ((hole->start == 1000 && hole->end == 1200),
+	    "first hole start %u end %u", hole->start, hole->end);
+  hole = scoreboard_last_hole (sb);
+  TCP_TEST ((hole->start == 1300 && hole->end == 1500),
+	    "last hole start %u end %u", hole->start, hole->end);
+  TCP_TEST ((sb->sacked_bytes == 100), "sacked bytes %d", sb->sacked_bytes);
+
+  /*
+   * Ack first hole
+   */
+
+  vec_reset_length (tc->opt.sacks);
+  tcp_rcv_sacks (tc, 1200);
+
+  TCP_TEST ((sb->snd_una_adv == 100),
+	    "snd_una_adv after ack %u", sb->snd_una_adv);
+  TCP_TEST ((sb->sacked_bytes == 0), "sacked bytes %d", sb->sacked_bytes);
+  TCP_TEST ((pool_elts (sb->holes) == 1),
+	    "scoreboard has %d elements", pool_elts (sb->holes));
+
+  /*
+   * Remove all
+   */
+
+  scoreboard_clear (sb);
+  TCP_TEST ((pool_elts (sb->holes) == 0),
+	    "number of holes %d", pool_elts (sb->holes));
+  return 0;
+}
+
+static clib_error_t *
+tcp_test (vlib_main_t * vm,
+	  unformat_input_t * input, vlib_cli_command_t * cmd_arg)
+{
+  int res = 0;
+
+  while (unformat_check_input (input) != UNFORMAT_END_OF_INPUT)
+    {
+      if (unformat (input, "sack"))
+	{
+	  res = tcp_test_sack ();
+	}
+      else
+	{
+	  return clib_error_return (0, "unknown input `%U'",
+				    format_unformat_error, input);
+	}
+    }
+
+  if (res)
+    {
+      return clib_error_return (0, "TCP unit test failed");
+    }
+  else
+    {
+      return 0;
+    }
+}
+
+VLIB_CLI_COMMAND (tcp_test_command, static) =
+{
+.path = "test tcp",.short_help = "internal tcp unit tests",.function =
+    tcp_test,};
+/*
+ * fd.io coding-style-patch-verification: ON
+ *
+ * Local Variables:
+ * eval: (c-set-style "gnu")
+ * End:
+ */
diff --git a/src/vnet/udp/builtin_server.c b/src/vnet/udp/builtin_server.c
index 46c8e734..57f774c5 100644
--- a/src/vnet/udp/builtin_server.c
+++ b/src/vnet/udp/builtin_server.c
@@ -39,10 +39,10 @@ builtin_session_disconnect_callback (stream_session_t * s)
 }
 
 static int
-builtin_server_rx_callback (stream_session_t * s, session_fifo_event_t * ep)
+builtin_server_rx_callback (stream_session_t * s)
 {
   svm_fifo_t *rx_fifo, *tx_fifo;
-  u32 this_transfer;
+  u32 this_transfer, max_deq, max_enq;
   int actual_transfer;
   u8 *my_copy_buffer;
   session_fifo_event_t evt;
@@ -52,9 +52,9 @@ builtin_server_rx_callback (stream_session_t * s, session_fifo_event_t * ep)
   rx_fifo = s->server_rx_fifo;
   tx_fifo = s->server_tx_fifo;
 
-  this_transfer = svm_fifo_max_enqueue (tx_fifo)
-    < svm_fifo_max_dequeue (rx_fifo) ?
-    svm_fifo_max_enqueue (tx_fifo) : svm_fifo_max_dequeue (rx_fifo);
+  max_deq = svm_fifo_max_dequeue (rx_fifo);
+  max_enq = svm_fifo_max_enqueue (tx_fifo);
+  this_transfer = max_enq < max_deq ? max_enq : max_deq;
 
   vec_validate (my_copy_buffer, this_transfer - 1);
   _vec_len (my_copy_buffer) = this_transfer;
@@ -64,17 +64,20 @@ builtin_server_rx_callback (stream_session_t * s, session_fifo_event_t * ep)
   ASSERT (actual_transfer == this_transfer);
   actual_transfer = svm_fifo_enqueue_nowait (tx_fifo, 0, this_transfer,
 					     my_copy_buffer);
+  ASSERT (actual_transfer == this_transfer);
 
   copy_buffers[s->thread_index] = my_copy_buffer;
 
-  /* Fabricate TX event, send to ourselves */
-  evt.fifo = tx_fifo;
-  evt.event_type = FIFO_EVENT_SERVER_TX;
-  /* $$$$ for event logging */
-  evt.enqueue_length = actual_transfer;
-  evt.event_id = 0;
-  q = session_manager_get_vpp_event_queue (s->thread_index);
-  unix_shared_memory_queue_add (q, (u8 *) & evt, 0 /* do wait for mutex */ );
+  if (svm_fifo_set_event (tx_fifo))
+    {
+      /* Fabricate TX event, send to ourselves */
+      evt.fifo = tx_fifo;
+      evt.event_type = FIFO_EVENT_SERVER_TX;
+      evt.event_id = 0;
+      q = session_manager_get_vpp_event_queue (s->thread_index);
+      unix_shared_memory_queue_add (q, (u8 *) & evt,
+				    0 /* do wait for mutex */ );
+    }
 
   return 0;
 }
diff --git a/src/vnet/udp/udp_input.c b/src/vnet/udp/udp_input.c
index 88278735..4b22109b 100644
--- a/src/vnet/udp/udp_input.c
+++ b/src/vnet/udp/udp_input.c
@@ -244,44 +244,53 @@ udp4_uri_input_node_fn (vlib_main_t * vm,
       /* Get session's server */
       server0 = application_get (s0->app_index);
 
-      /* Fabricate event */
-      evt.fifo = s0->server_rx_fifo;
-      evt.event_type = FIFO_EVENT_SERVER_RX;
-      evt.event_id = serial_number++;
-      evt.enqueue_length = svm_fifo_max_dequeue (s0->server_rx_fifo);
-
       /* Built-in server? Deliver the goods... */
       if (server0->cb_fns.builtin_server_rx_callback)
 	{
-	  server0->cb_fns.builtin_server_rx_callback (s0, &evt);
+	  server0->cb_fns.builtin_server_rx_callback (s0);
 	  continue;
 	}
 
-      /* Add event to server's event queue */
-      q = server0->event_queue;
-
-      /* Don't block for lack of space */
-      if (PREDICT_TRUE (q->cursize < q->maxsize))
-	unix_shared_memory_queue_add (server0->event_queue, (u8 *) & evt,
-				      0 /* do wait for mutex */ );
-      else
+      if (svm_fifo_set_event (s0->server_rx_fifo))
 	{
-	  vlib_node_increment_counter (vm, udp4_uri_input_node.index,
-				       SESSION_ERROR_FIFO_FULL, 1);
+	  /* Fabricate event */
+	  evt.fifo = s0->server_rx_fifo;
+	  evt.event_type = FIFO_EVENT_SERVER_RX;
+	  evt.event_id = serial_number++;
+
+	  /* Add event to server's event queue */
+	  q = server0->event_queue;
+
+	  /* Don't block for lack of space */
+	  if (PREDICT_TRUE (q->cursize < q->maxsize))
+	    {
+	      unix_shared_memory_queue_add (server0->event_queue,
+					    (u8 *) & evt,
+					    0 /* do wait for mutex */ );
+	    }
+	  else
+	    {
+	      vlib_node_increment_counter (vm, udp4_uri_input_node.index,
+					   SESSION_ERROR_FIFO_FULL, 1);
+	    }
 	}
+      /* *INDENT-OFF* */
       if (1)
 	{
 	  ELOG_TYPE_DECLARE (e) =
 	  {
-	  .format = "evt-enqueue: id %d length %d",.format_args = "i4i4",};
+	      .format = "evt-enqueue: id %d length %d",
+	      .format_args = "i4i4",};
 	  struct
 	  {
 	    u32 data[2];
 	  } *ed;
 	  ed = ELOG_DATA (&vlib_global_main.elog_main, e);
 	  ed->data[0] = evt.event_id;
-	  ed->data[1] = evt.enqueue_length;
+	  ed->data[1] = svm_fifo_max_dequeue (s0->server_rx_fifo);
 	}
+      /* *INDENT-ON* */
+
     }
 
   vec_reset_length (session_indices_to_enqueue);
-- 
cgit 1.2.3-korg


From 3e350af5d3e9744a4529a28dd293b2d4601442f7 Mon Sep 17 00:00:00 2001
From: Florin Coras <fcoras@cisco.com>
Date: Thu, 30 Mar 2017 02:54:28 -0700
Subject: TCP cc/window management fixes and debugging

- added persist timer
- update rcv_las whenever sending an ack
- moved fifo size to its own cache line
- improved session and builtin client debugging

Change-Id: Ia649cf942cf0c061a713e8b67f0eb6974a6cd55b
Signed-off-by: Florin Coras <fcoras@cisco.com>
Signed-off-by: Dave Barach <dave@barachs.net>
---
 src/svm/svm_fifo.c               |  37 +++++-----
 src/svm/svm_fifo.h               |   8 ++-
 src/vnet/session/node.c          |  24 ++++---
 src/vnet/session/session.c       |   6 +-
 src/vnet/session/session.h       |  12 +++-
 src/vnet/session/session_debug.h |  32 ++++++++-
 src/vnet/tcp/builtin_client.c    | 146 +++++++++++++++++++++++++++------------
 src/vnet/tcp/builtin_server.c    |  12 +++-
 src/vnet/tcp/tcp.c               |   6 +-
 src/vnet/tcp/tcp.h               |  36 +++++++++-
 src/vnet/tcp/tcp_debug.h         |  66 +++++++++++++-----
 src/vnet/tcp/tcp_input.c         |  68 +++++++++---------
 src/vnet/tcp/tcp_output.c        |  79 +++++++++++++++++----
 13 files changed, 386 insertions(+), 146 deletions(-)

(limited to 'src/svm/svm_fifo.c')

diff --git a/src/svm/svm_fifo.c b/src/svm/svm_fifo.c
index 07b0d2df..cc84feb9 100644
--- a/src/svm/svm_fifo.c
+++ b/src/svm/svm_fifo.c
@@ -254,6 +254,10 @@ ooo_segment_try_collect (svm_fifo_t * f, u32 n_bytes_enqueued)
 {
   ooo_segment_t *s;
   u32 index, bytes = 0, diff;
+  u32 cursize;
+
+  /* read cursize, which can only increase while we're working */
+  cursize = svm_fifo_max_dequeue (f);
 
   s = pool_elt_at_index (f->ooo_segments, f->ooos_list_head);
 
@@ -286,8 +290,8 @@ ooo_segment_try_collect (svm_fifo_t * f, u32 n_bytes_enqueued)
   /* If tail is adjacent to an ooo segment, 'consume' it */
   if (diff == 0)
     {
-      bytes = ((f->nitems - f->cursize) >= s->length) ? s->length :
-	f->nitems - f->cursize;
+      bytes = ((f->nitems - cursize) >= s->length) ? s->length :
+	f->nitems - cursize;
 
       f->tail += bytes;
       f->tail %= f->nitems;
@@ -305,11 +309,12 @@ svm_fifo_enqueue_internal (svm_fifo_t * f,
   u32 total_copy_bytes, first_copy_bytes, second_copy_bytes;
   u32 cursize, nitems;
 
-  if (PREDICT_FALSE (f->cursize == f->nitems))
+  /* read cursize, which can only increase while we're working */
+  cursize = svm_fifo_max_dequeue (f);
+
+  if (PREDICT_FALSE (cursize == f->nitems))
     return -2;			/* fifo stuffed */
 
-  /* read cursize, which can only decrease while we're working */
-  cursize = f->cursize;
   nitems = f->nitems;
 
   /* Number of bytes we're going to copy */
@@ -382,8 +387,8 @@ svm_fifo_enqueue_with_offset_internal (svm_fifo_t * f,
 
   ASSERT (offset > 0);
 
-  /* read cursize, which can only decrease while we're working */
-  cursize = f->cursize;
+  /* read cursize, which can only increase while we're working */
+  cursize = svm_fifo_max_dequeue (f);
   nitems = f->nitems;
 
   /* Will this request fit? */
@@ -437,11 +442,11 @@ svm_fifo_dequeue_internal (svm_fifo_t * f,
   u32 total_copy_bytes, first_copy_bytes, second_copy_bytes;
   u32 cursize, nitems;
 
-  if (PREDICT_FALSE (f->cursize == 0))
+  /* read cursize, which can only increase while we're working */
+  cursize = svm_fifo_max_dequeue (f);
+  if (PREDICT_FALSE (cursize == 0))
     return -2;			/* nothing in the fifo */
 
-  /* read cursize, which can only increase while we're working */
-  cursize = f->cursize;
   nitems = f->nitems;
 
   /* Number of bytes we're going to copy */
@@ -495,11 +500,11 @@ svm_fifo_peek (svm_fifo_t * f, int pid, u32 offset, u32 max_bytes,
   u32 total_copy_bytes, first_copy_bytes, second_copy_bytes;
   u32 cursize, nitems, real_head;
 
-  if (PREDICT_FALSE (f->cursize == 0))
+  /* read cursize, which can only increase while we're working */
+  cursize = svm_fifo_max_dequeue (f);
+  if (PREDICT_FALSE (cursize == 0))
     return -2;			/* nothing in the fifo */
 
-  /* read cursize, which can only increase while we're working */
-  cursize = f->cursize;
   nitems = f->nitems;
   real_head = f->head + offset;
   real_head = real_head >= nitems ? real_head - nitems : real_head;
@@ -532,11 +537,11 @@ svm_fifo_dequeue_drop (svm_fifo_t * f, int pid, u32 max_bytes)
   u32 total_drop_bytes, first_drop_bytes, second_drop_bytes;
   u32 cursize, nitems;
 
-  if (PREDICT_FALSE (f->cursize == 0))
+  /* read cursize, which can only increase while we're working */
+  cursize = svm_fifo_max_dequeue (f);
+  if (PREDICT_FALSE (cursize == 0))
     return -2;			/* nothing in the fifo */
 
-  /* read cursize, which can only increase while we're working */
-  cursize = f->cursize;
   nitems = f->nitems;
 
   /* Number of bytes we're going to drop */
diff --git a/src/svm/svm_fifo.h b/src/svm/svm_fifo.h
index 39556173..80e5b0f2 100644
--- a/src/svm/svm_fifo.h
+++ b/src/svm/svm_fifo.h
@@ -44,14 +44,16 @@ typedef struct
 
 typedef struct
 {
+  volatile u32 cursize;		/**< current fifo size */
+  u32 nitems;
+    CLIB_CACHE_LINE_ALIGN_MARK (end_cursize);
+
   pthread_mutex_t mutex;	/* 8 bytes */
   pthread_cond_t condvar;	/* 8 bytes */
   svm_lock_tag_t tag;
 
-  volatile u32 cursize;		/**< current fifo size */
   volatile u8 has_event;	/**< non-zero if deq event exists */
   u32 owner_pid;
-  u32 nitems;
 
   /* Backpointers */
   u32 server_session_index;
@@ -105,7 +107,7 @@ svm_fifo_max_dequeue (svm_fifo_t * f)
 static inline u32
 svm_fifo_max_enqueue (svm_fifo_t * f)
 {
-  return f->nitems - f->cursize;
+  return f->nitems - svm_fifo_max_dequeue (f);
 }
 
 static inline u8
diff --git a/src/vnet/session/node.c b/src/vnet/session/node.c
index 8681105c..b86e87d9 100644
--- a/src/vnet/session/node.c
+++ b/src/vnet/session/node.c
@@ -119,15 +119,20 @@ session_tx_fifo_read_and_snd_i (vlib_main_t * vm, vlib_node_runtime_t * node,
 
   /* Nothing to read return */
   if (max_dequeue0 == 0)
-    {
-      return 0;
-    }
+    return 0;
 
   /* Ensure we're not writing more than transport window allows */
-  max_len_to_snd0 = clib_min (max_dequeue0, snd_space0);
-
-  /* TODO check if transport is willing to send len_to_snd0
-   * bytes (Nagle) */
+  if (max_dequeue0 < snd_space0)
+    {
+      /* Constrained by tx queue. Try to send only fully formed segments */
+      max_len_to_snd0 = (max_dequeue0 > snd_mss0) ?
+	max_dequeue0 - max_dequeue0 % snd_mss0 : max_dequeue0;
+      /* TODO Nagle ? */
+    }
+  else
+    {
+      max_len_to_snd0 = snd_space0;
+    }
 
   n_frame_bytes = snd_mss0 * VLIB_FRAME_SIZE;
   n_frames_per_evt = ceil ((double) max_len_to_snd0 / n_frame_bytes);
@@ -308,11 +313,14 @@ session_queue_node_fn (vlib_main_t * vm, vlib_node_runtime_t * node,
   int n_tx_packets = 0;
   u32 my_thread_index = vm->cpu_index;
   int i, rv;
+  f64 now = vlib_time_now (vm);
+
+  SESSION_EVT_DBG (SESSION_EVT_POLL_GAP_TRACK, smm, my_thread_index);
 
   /*
    *  Update TCP time
    */
-  tcp_update_time (vlib_time_now (vm), my_thread_index);
+  tcp_update_time (now, my_thread_index);
 
   /*
    * Get vpp queue events
diff --git a/src/vnet/session/session.c b/src/vnet/session/session.c
index f10918aa..8e2b2616 100644
--- a/src/vnet/session/session.c
+++ b/src/vnet/session/session.c
@@ -556,7 +556,7 @@ session_manager_allocate_session_fifos (session_manager_main_t * smm,
 					u8 * added_a_segment)
 {
   svm_fifo_segment_private_t *fifo_segment;
-  u32 fifo_size, default_fifo_size = 128 << 10;	/* TODO config */
+  u32 fifo_size, default_fifo_size = 1 << 16;	/* TODO config */
   int i;
 
   *added_a_segment = 0;
@@ -1293,6 +1293,10 @@ session_manager_main_enable (vlib_main_t * vm)
   vec_validate (smm->current_enqueue_epoch, num_threads - 1);
   vec_validate (smm->vpp_event_queues, num_threads - 1);
 
+#if SESSION_DBG
+  vec_validate (smm->last_event_poll_by_thread, num_threads - 1);
+#endif
+
   /* $$$$ preallocate hack config parameter */
   for (i = 0; i < 200000; i++)
     {
diff --git a/src/vnet/session/session.h b/src/vnet/session/session.h
index a39bc06f..6878b4d2 100644
--- a/src/vnet/session/session.h
+++ b/src/vnet/session/session.h
@@ -20,6 +20,7 @@
 #include <vlibmemory/api.h>
 #include <vppinfra/sparse_vec.h>
 #include <svm/svm_fifo_segment.h>
+#include <vnet/session/session_debug.h>
 
 #define HALF_OPEN_LOOKUP_INVALID_VALUE ((u64)~0)
 #define INVALID_INDEX ((u32)~0)
@@ -36,7 +37,7 @@ typedef enum
   FIFO_EVENT_BUILTIN_RX
 } fifo_event_type_t;
 
-#define foreach_session_input_error                                         \
+#define foreach_session_input_error                                    	\
 _(NO_SESSION, "No session drops")                                       \
 _(NO_LISTENER, "No listener for dst port drops")                        \
 _(ENQUEUED, "Packets pushed into rx fifo")                              \
@@ -218,6 +219,15 @@ struct _session_manager_main
   /* Convenience */
   vlib_main_t *vlib_main;
   vnet_main_t *vnet_main;
+
+#if SESSION_DBG
+  /**
+   * last event poll time by thread
+   * Debug only. Will cause false cache-line sharing as-is
+   */
+  f64 *last_event_poll_by_thread;
+#endif
+
 };
 
 extern session_manager_main_t session_manager_main;
diff --git a/src/vnet/session/session_debug.h b/src/vnet/session/session_debug.h
index 80a97cd5..eb11f1a0 100644
--- a/src/vnet/session/session_debug.h
+++ b/src/vnet/session/session_debug.h
@@ -16,13 +16,13 @@
 #define SRC_VNET_SESSION_SESSION_DEBUG_H_
 
 #include <vnet/session/transport.h>
-#include <vnet/session/session.h>
 #include <vlib/vlib.h>
 
 #define foreach_session_dbg_evt		\
   _(ENQ, "enqueue")			\
   _(DEQ, "dequeue")			\
-  _(DEQ_NODE, "dequeue")
+  _(DEQ_NODE, "dequeue")		\
+  _(POLL_GAP_TRACK, "poll gap track")	\
 
 typedef enum _session_evt_dbg
 {
@@ -33,6 +33,7 @@ typedef enum _session_evt_dbg
 
 #define SESSION_DBG (0)
 #define SESSION_DEQ_NODE_EVTS (0)
+#define SESSION_EVT_POLL_DBG (1)
 
 #if TRANSPORT_DEBUG && SESSION_DBG
 
@@ -97,9 +98,34 @@ typedef enum _session_evt_dbg
 #define SESSION_EVT_DEQ_NODE_HANDLER(_node_evt)
 #endif
 
+#if SESSION_DBG && SESSION_EVT_POLL_DBG
+#define SESSION_EVT_POLL_GAP(_smm, _my_thread_index)			\
+{									\
+  ELOG_TYPE_DECLARE (_e) =						\
+  {									\
+    .format = "nixon-gap: %d MS",					\
+    .format_args = "i4",						\
+  };									\
+  DEC_SESSION_ED(_e, 1);						\
+  ed->data[0] =	(u32) ((now -						\
+    _smm->last_event_poll_by_thread[my_thread_index])*1000.0);		\
+}
+#define SESSION_EVT_POLL_GAP_TRACK_HANDLER(_smm, _my_thread_index)	\
+{									\
+  if (PREDICT_TRUE(							\
+	      smm->last_event_poll_by_thread[my_thread_index] != 0.0))	\
+    if (now > smm->last_event_poll_by_thread[_my_thread_index] + 500e-6)\
+	SESSION_EVT_POLL_GAP(smm, my_thread_index);			\
+  _smm->last_event_poll_by_thread[my_thread_index] = now;		\
+}
+
+#else
+#define SESSION_EVT_POLL_GAP(_smm, _my_thread_index)
+#define SESSION_EVT_POLL_GAP_TRACK_HANDLER(_smm, _my_thread_index)
+#endif
+
 #define CONCAT_HELPER(_a, _b) _a##_b
 #define CC(_a, _b) CONCAT_HELPER(_a, _b)
-
 #define SESSION_EVT_DBG(_evt, _args...) CC(_evt, _HANDLER)(_args)
 
 #else
diff --git a/src/vnet/tcp/builtin_client.c b/src/vnet/tcp/builtin_client.c
index 83cdbc1b..e3705060 100644
--- a/src/vnet/tcp/builtin_client.c
+++ b/src/vnet/tcp/builtin_client.c
@@ -43,6 +43,10 @@
 #include <vpp/api/vpe_all_api_h.h>
 #undef vl_printfun
 
+#define TCP_BUILTIN_CLIENT_DBG (1)
+#define TCP_BUILTIN_CLIENT_VPP_THREAD (0)
+#define TCP_BUILTIN_CLIENT_PTHREAD (!TCP_BUILTIN_CLIENT_VPP_THREAD)
+
 static void
 send_test_chunk (tclient_main_t * tm, session_t * s)
 {
@@ -52,35 +56,50 @@ send_test_chunk (tclient_main_t * tm, session_t * s)
   session_fifo_event_t evt;
   static int serial_number = 0;
   int rv;
+  test_buf_offset = s->bytes_sent % vec_len (test_data);
+  bytes_this_chunk = vec_len (test_data) - test_buf_offset;
 
-  while (s->bytes_to_send > 0)
-    {
+  bytes_this_chunk = bytes_this_chunk < s->bytes_to_send
+    ? bytes_this_chunk : s->bytes_to_send;
 
-      test_buf_offset = s->bytes_sent % vec_len (test_data);
-      bytes_this_chunk = vec_len (test_data) - test_buf_offset;
+  rv = svm_fifo_enqueue_nowait (s->server_tx_fifo, 0 /*pid */ ,
+				bytes_this_chunk,
+				test_data + test_buf_offset);
 
-      bytes_this_chunk = bytes_this_chunk < s->bytes_to_send
-	? bytes_this_chunk : s->bytes_to_send;
+  /* If we managed to enqueue data... */
+  if (rv > 0)
+    {
+      if (TCP_BUILTIN_CLIENT_DBG)
+	{
+          /* *INDENT-OFF* */
+          ELOG_TYPE_DECLARE (e) =
+            {
+              .format = "tx-enq: %d bytes",
+              .format_args = "i4",
+            };
+          /* *INDENT-ON* */
+	  struct
+	  {
+	    u32 data[1];
+	  } *ed;
+	  ed = ELOG_DATA (&vlib_global_main.elog_main, e);
+	  ed->data[0] = rv;
+	}
 
-      rv = svm_fifo_enqueue_nowait (s->server_tx_fifo, 0 /*pid */ ,
-				    bytes_this_chunk,
-				    test_data + test_buf_offset);
+      /* Account for it... */
+      s->bytes_to_send -= rv;
+      s->bytes_sent += rv;
 
-      if (rv > 0)
+      /* Poke the TCP state machine */
+      if (svm_fifo_set_event (s->server_tx_fifo))
 	{
-	  s->bytes_to_send -= rv;
-	  s->bytes_sent += rv;
+	  /* Fabricate TX event, send to vpp */
+	  evt.fifo = s->server_tx_fifo;
+	  evt.event_type = FIFO_EVENT_SERVER_TX;
+	  evt.event_id = serial_number++;
 
-	  if (svm_fifo_set_event (s->server_tx_fifo))
-	    {
-	      /* Fabricate TX event, send to vpp */
-	      evt.fifo = s->server_tx_fifo;
-	      evt.event_type = FIFO_EVENT_SERVER_TX;
-	      evt.event_id = serial_number++;
-
-	      unix_shared_memory_queue_add (tm->vpp_event_queue, (u8 *) & evt,
-					    0 /* do wait for mutex */ );
-	    }
+	  unix_shared_memory_queue_add (tm->vpp_event_queue, (u8 *) & evt,
+					0 /* do wait for mutex */ );
 	}
     }
 }
@@ -89,39 +108,55 @@ static void
 receive_test_chunk (tclient_main_t * tm, session_t * s)
 {
   svm_fifo_t *rx_fifo = s->server_rx_fifo;
-  int n_read, bytes, i;
+  int n_read, test_bytes = 0;
 
-  bytes = svm_fifo_max_dequeue (rx_fifo);
   /* Allow enqueuing of new event */
-  svm_fifo_unset_event (rx_fifo);
+  // svm_fifo_unset_event (rx_fifo);
 
-  /* Read the bytes */
-  do
+  n_read = svm_fifo_dequeue_nowait (rx_fifo, 0, vec_len (tm->rx_buf),
+				    tm->rx_buf);
+  if (n_read > 0)
     {
-      n_read = svm_fifo_dequeue_nowait (rx_fifo, 0, vec_len (tm->rx_buf),
-					tm->rx_buf);
-      if (n_read > 0)
+      if (TCP_BUILTIN_CLIENT_DBG)
 	{
-	  bytes -= n_read;
+          /* *INDENT-OFF* */
+          ELOG_TYPE_DECLARE (e) =
+            {
+              .format = "rx-deq: %d bytes",
+              .format_args = "i4",
+            };
+          /* *INDENT-ON* */
+	  struct
+	  {
+	    u32 data[1];
+	  } *ed;
+	  ed = ELOG_DATA (&vlib_global_main.elog_main, e);
+	  ed->data[0] = n_read;
+	}
+
+      if (test_bytes)
+	{
+	  int i;
 	  for (i = 0; i < n_read; i++)
 	    {
 	      if (tm->rx_buf[i] != ((s->bytes_received + i) & 0xff))
 		{
 		  clib_warning ("read %d error at byte %lld, 0x%x not 0x%x",
-				n_read, s->bytes_received + i,
-				tm->rx_buf[i],
+				n_read, s->bytes_received + i, tm->rx_buf[i],
 				((s->bytes_received + i) & 0xff));
 		}
 	    }
-	  s->bytes_to_receive -= n_read;
-	  s->bytes_received += n_read;
 	}
-
+      s->bytes_to_receive -= n_read;
+      s->bytes_received += n_read;
     }
-  while (n_read < 0 || bytes > 0);
 }
 
+#if TCP_BUILTIN_CLIENT_VPP_THREAD
+static void
+#else
 static void *
+#endif
 tclient_thread_fn (void *arg)
 {
   tclient_main_t *tm = &tclient_main;
@@ -139,6 +174,8 @@ tclient_thread_fn (void *arg)
     pthread_sigmask (SIG_SETMASK, &s, 0);
   }
 
+  clib_per_cpu_mheaps[os_get_cpu_number ()] = clib_per_cpu_mheaps[0];
+
   while (1)
     {
       /* Wait until we're told to get busy */
@@ -186,12 +223,12 @@ tclient_thread_fn (void *arg)
 
       /* Disconnect sessions... */
       vec_reset_length (session_indices);
-      pool_foreach (sp, tm->sessions, (
-					{
-					vec_add1 (session_indices,
-						  sp - tm->sessions);
-					}
-		    ));
+
+      /* *INDENT-OFF* */
+      pool_foreach (sp, tm->sessions, ({
+	vec_add1 (session_indices, sp - tm->sessions);
+      }));
+      /* *INDENT-ON* */
 
       for (i = 0; i < vec_len (session_indices); i++)
 	{
@@ -207,7 +244,9 @@ tclient_thread_fn (void *arg)
 	}
     }
   /* NOTREACHED */
+#if TCP_BUILTIN_CLIENT_PTHREAD
   return 0;
+#endif
 }
 
 /* So we don't get "no handler for... " msgs */
@@ -333,7 +372,7 @@ test_tcp_clients_command_fn (vlib_main_t * vm,
 			     unformat_input_t * input,
 			     vlib_cli_command_t * cmd)
 {
-  u8 *connect_uri = (u8 *) "tcp://6.0.1.2/1234";
+  u8 *connect_uri = (u8 *) "tcp://6.0.1.1/1234";
   u8 *uri;
   tclient_main_t *tm = &tclient_main;
   int i;
@@ -349,7 +388,7 @@ test_tcp_clients_command_fn (vlib_main_t * vm,
 	;
       else if (unformat (input, "iterations %d", &tm->n_iterations))
 	;
-      else if (unformat (input, "bytes %d", &tm->bytes_to_send))
+      else if (unformat (input, "bytes %lld", &tm->bytes_to_send))
 	;
       else if (unformat (input, "uri %s", &tm->connect_uri))
 	;
@@ -366,17 +405,20 @@ test_tcp_clients_command_fn (vlib_main_t * vm,
 
   create_api_loopback (tm);
 
+#if TCP_BUILTIN_CLIENT_PTHREAD
   /* Start a transmit thread */
   if (tm->client_thread_handle == 0)
     {
       int rv = pthread_create (&tm->client_thread_handle,
-			       NULL /*attr */ , tclient_thread_fn, 0);
+			       NULL /*attr */ ,
+			       tclient_thread_fn, 0);
       if (rv)
 	{
 	  tm->client_thread_handle = 0;
 	  return clib_error_return (0, "pthread_create returned %d", rv);
 	}
     }
+#endif
 
   /* Fire off connect requests, in something approaching a normal manner */
   for (i = 0; i < n_clients; i++)
@@ -397,6 +439,18 @@ test_tcp_clients_command_fn (vlib_main_t * vm,
   return 0;
 }
 
+#if TCP_BUILTIN_CLIENT_VPP_THREAD
+/* *INDENT-OFF* */
+VLIB_REGISTER_THREAD (builtin_client_reg, static) = {
+  .name = "tcp-builtin-client",
+  .function = tclient_thread_fn,
+  .fixed_count = 1,
+  .count = 1,
+  .no_data_structure_clone = 1,
+};
+/* *INDENT-ON* */
+#endif
+
 /* *INDENT-OFF* */
 VLIB_CLI_COMMAND (test_clients_command, static) =
 {
diff --git a/src/vnet/tcp/builtin_server.c b/src/vnet/tcp/builtin_server.c
index efd26e91..917d4bd3 100644
--- a/src/vnet/tcp/builtin_server.c
+++ b/src/vnet/tcp/builtin_server.c
@@ -127,6 +127,7 @@ builtin_server_rx_callback (stream_session_t * s)
     {
       /* XXX timeout for session that are stuck */
 
+    rx_event:
       /* Program self-tap to retry */
       if (svm_fifo_set_event (rx_fifo))
 	{
@@ -158,7 +159,9 @@ builtin_server_rx_callback (stream_session_t * s)
 
   n_written =
     svm_fifo_enqueue_nowait (tx_fifo, 0, actual_transfer, bsm->rx_buf);
-  ASSERT (n_written == max_transfer);
+
+  if (n_written != max_transfer)
+    clib_warning ("short trout!");
 
   if (svm_fifo_set_event (tx_fifo))
     {
@@ -171,6 +174,9 @@ builtin_server_rx_callback (stream_session_t * s)
 				    (u8 *) & evt, 0 /* do wait for mutex */ );
     }
 
+  if (PREDICT_FALSE (max_enqueue < max_dequeue))
+    goto rx_event;
+
   return 0;
 }
 
@@ -204,8 +210,8 @@ server_create (vlib_main_t * vm)
   a->session_cb_vft = &builtin_session_cb_vft;
   a->options = options;
   a->options[SESSION_OPTIONS_SEGMENT_SIZE] = 128 << 20;
-  a->options[SESSION_OPTIONS_RX_FIFO_SIZE] = 64 << 10;
-  a->options[SESSION_OPTIONS_TX_FIFO_SIZE] = 64 << 10;
+  a->options[SESSION_OPTIONS_RX_FIFO_SIZE] = 1 << 16;
+  a->options[SESSION_OPTIONS_TX_FIFO_SIZE] = 1 << 16;
   a->segment_name = segment_name;
   a->segment_name_length = ARRAY_LEN (segment_name);
 
diff --git a/src/vnet/tcp/tcp.c b/src/vnet/tcp/tcp.c
index c3df5bc1..b2a371e2 100644
--- a/src/vnet/tcp/tcp.c
+++ b/src/vnet/tcp/tcp.c
@@ -578,7 +578,9 @@ tcp_session_send_space (transport_connection_t * trans_conn)
       /* If we can't write at least a segment, don't try at all */
       if (snd_space < tc->snd_mss)
 	return 0;
-      return snd_space;
+
+      /* round down to mss multiple */
+      return snd_space - (snd_space % tc->snd_mss);
     }
 
   /* If in fast recovery, send 1 SMSS if wnd allows */
@@ -706,7 +708,7 @@ static timer_expiration_handler *timer_expiration_handlers[TCP_N_TIMERS] =
 {
     tcp_timer_retransmit_handler,
     tcp_timer_delack_handler,
-    0,
+    tcp_timer_persist_handler,
     tcp_timer_keep_handler,
     tcp_timer_waitclose_handler,
     tcp_timer_retransmit_syn_handler,
diff --git a/src/vnet/tcp/tcp.h b/src/vnet/tcp/tcp.h
index b4286bc4..2f5da108 100644
--- a/src/vnet/tcp/tcp.h
+++ b/src/vnet/tcp/tcp.h
@@ -81,6 +81,7 @@ typedef void (timer_expiration_handler) (u32 index);
 
 extern timer_expiration_handler tcp_timer_delack_handler;
 extern timer_expiration_handler tcp_timer_retransmit_handler;
+extern timer_expiration_handler tcp_timer_persist_handler;
 extern timer_expiration_handler tcp_timer_retransmit_syn_handler;
 
 #define TCP_TIMER_HANDLE_INVALID ((u32) ~0)
@@ -253,13 +254,25 @@ struct _tcp_cc_algorithm
 
 #define tcp_fastrecovery_on(tc) (tc)->flags |= TCP_CONN_FAST_RECOVERY
 #define tcp_fastrecovery_off(tc) (tc)->flags &= ~TCP_CONN_FAST_RECOVERY
+#define tcp_recovery_on(tc) (tc)->flags |= TCP_CONN_RECOVERY
+#define tcp_recovery_off(tc) (tc)->flags &= ~TCP_CONN_RECOVERY
 #define tcp_in_fastrecovery(tc) ((tc)->flags & TCP_CONN_FAST_RECOVERY)
-#define tcp_in_recovery(tc) ((tc)->flags & (TCP_CONN_FAST_RECOVERY | TCP_CONN_RECOVERY))
+#define tcp_in_recovery(tc) ((tc)->flags & (TCP_CONN_RECOVERY))
 #define tcp_in_slowstart(tc) (tc->cwnd < tc->ssthresh)
 #define tcp_fastrecovery_sent_1_smss(tc) ((tc)->flags & TCP_CONN_FR_1_SMSS)
 #define tcp_fastrecovery_1_smss_on(tc) ((tc)->flags |= TCP_CONN_FR_1_SMSS)
 #define tcp_fastrecovery_1_smss_off(tc) ((tc)->flags &= ~TCP_CONN_FR_1_SMSS)
 
+#define tcp_in_cong_recovery(tc) ((tc)->flags & 		\
+	  (TCP_CONN_FAST_RECOVERY | TCP_CONN_RECOVERY))
+
+always_inline void
+tcp_cong_recovery_off (tcp_connection_t * tc)
+{
+  tc->flags &= ~(TCP_CONN_FAST_RECOVERY | TCP_CONN_RECOVERY);
+  tcp_fastrecovery_1_smss_off (tc);
+}
+
 typedef enum
 {
   TCP_IP4,
@@ -538,6 +551,27 @@ tcp_retransmit_timer_reset (tcp_connection_t * tc)
   tcp_timer_reset (tc, TCP_TIMER_RETRANSMIT);
 }
 
+always_inline void
+tcp_persist_timer_set (tcp_connection_t * tc)
+{
+  /* Reuse RTO. It's backed off in handler */
+  tcp_timer_set (tc, TCP_TIMER_PERSIST,
+		 clib_max (tc->rto * TCP_TO_TIMER_TICK, 1));
+}
+
+always_inline void
+tcp_persist_timer_update (tcp_connection_t * tc)
+{
+  tcp_timer_update (tc, TCP_TIMER_PERSIST,
+		    clib_max (tc->rto * TCP_TO_TIMER_TICK, 1));
+}
+
+always_inline void
+tcp_persist_timer_reset (tcp_connection_t * tc)
+{
+  tcp_timer_reset (tc, TCP_TIMER_PERSIST);
+}
+
 always_inline u8
 tcp_timer_is_active (tcp_connection_t * tc, tcp_timers_e timer)
 {
diff --git a/src/vnet/tcp/tcp_debug.h b/src/vnet/tcp/tcp_debug.h
index 5a71694e..0090e15e 100644
--- a/src/vnet/tcp/tcp_debug.h
+++ b/src/vnet/tcp/tcp_debug.h
@@ -31,6 +31,7 @@
   _(UNBIND, "unbind")			\
   _(DELETE, "delete")			\
   _(SYN_SENT, "SYN sent")		\
+  _(SYN_RTX, "SYN retransmit")		\
   _(FIN_SENT, "FIN sent")		\
   _(ACK_SENT, "ACK sent")		\
   _(DUPACK_SENT, "DUPACK sent")		\
@@ -50,6 +51,7 @@
   _(CC_PACK, "cc partial ack")		\
   _(SEG_INVALID, "invalid segment")	\
   _(ACK_RCV_ERR, "invalid ack")		\
+  _(RCV_WND_SHRUNK, "shrunk rcv_wnd")	\
 
 typedef enum _tcp_dbg
 {
@@ -159,35 +161,48 @@ typedef enum _tcp_dbg_evt
 {									\
   ELOG_TYPE_DECLARE (_e) =						\
   {									\
-    .format = "ack_prep: acked %u rcv_nxt %u rcv_wnd %u snd_nxt %u",	\
-    .format_args = "i4i4i4i4",						\
+    .format = "ack_tx: acked %u rcv_nxt %u rcv_wnd %u snd_nxt %u snd_wnd %u",\
+    .format_args = "i4i4i4i4i4",					\
   };									\
-  DECLARE_ETD(_tc, _e, 4);						\
+  DECLARE_ETD(_tc, _e, 5);						\
   ed->data[0] = _tc->rcv_nxt - _tc->rcv_las;				\
   ed->data[1] = _tc->rcv_nxt - _tc->irs;				\
   ed->data[2] = _tc->rcv_wnd;						\
   ed->data[3] = _tc->snd_nxt - _tc->iss;				\
+  ed->data[4] = _tc->snd_wnd;						\
 }
 
 #define TCP_EVT_DUPACK_SENT_HANDLER(_tc, ...)				\
 {									\
   ELOG_TYPE_DECLARE (_e) =						\
   {									\
-    .format = "dack_tx: rcv_nxt %u rcv_wnd %u snd_nxt %u av-wnd %u",	\
-    .format_args = "i4i4i4i4",						\
+    .format = "dack_tx: rcv_nxt %u rcv_wnd %u snd_nxt %u av_wnd %u snd_wnd %u",\
+    .format_args = "i4i4i4i4i4",					\
   };									\
-  DECLARE_ETD(_tc, _e, 4);						\
+  DECLARE_ETD(_tc, _e, 5);						\
   ed->data[0] = _tc->rcv_nxt - _tc->irs;				\
   ed->data[1] = _tc->rcv_wnd;						\
   ed->data[2] = _tc->snd_nxt - _tc->iss;				\
   ed->data[3] = tcp_available_wnd(_tc);					\
+  ed->data[4] = _tc->snd_wnd;						\
 }
 
 #define TCP_EVT_SYN_SENT_HANDLER(_tc, ...)				\
 {									\
   ELOG_TYPE_DECLARE (_e) =						\
   {									\
-    .format = "SYNtx: iss %u",					\
+    .format = "SYNtx: iss %u",						\
+    .format_args = "i4",						\
+  };									\
+  DECLARE_ETD(_tc, _e, 1);						\
+  ed->data[0] = _tc->iss;						\
+}
+
+#define TCP_EVT_SYN_RTX_HANDLER(_tc, ...)				\
+{									\
+  ELOG_TYPE_DECLARE (_e) =						\
+  {									\
+    .format = "SYNrtx: iss %u",						\
     .format_args = "i4",						\
   };									\
   DECLARE_ETD(_tc, _e, 1);						\
@@ -254,17 +269,17 @@ typedef enum _tcp_dbg_evt
   ed->data[1] = _tc->rcv_nxt - _tc->irs;				\
 }
 
-#define TCP_EVT_ACK_RCVD_HANDLER(_tc, _ack, ...)			\
+#define TCP_EVT_ACK_RCVD_HANDLER(_tc, ...)				\
 {									\
   ELOG_TYPE_DECLARE (_e) =						\
   {									\
-    .format = "acked: %u snd_una %u ack %u cwnd %u inflight %u",	\
+    .format = "acked: %u snd_una %u snd_wnd %u cwnd %u inflight %u",	\
     .format_args = "i4i4i4i4i4",					\
   };									\
   DECLARE_ETD(_tc, _e, 5);						\
   ed->data[0] = _tc->bytes_acked;					\
   ed->data[1] = _tc->snd_una - _tc->iss;				\
-  ed->data[2] = _ack - _tc->iss;					\
+  ed->data[2] = _tc->snd_wnd;						\
   ed->data[3] = _tc->cwnd;						\
   ed->data[4] = tcp_flight_size(_tc);					\
 }
@@ -273,14 +288,15 @@ typedef enum _tcp_dbg_evt
 {									\
   ELOG_TYPE_DECLARE (_e) =						\
   {									\
-    .format = "dack_rx: snd_una %u cwnd %u snd_wnd %u inflight %u",	\
-    .format_args = "i4i4i4i4",						\
+    .format = "dack_rx: snd_una %u cwnd %u snd_wnd %u flight %u rcv_wnd %u",\
+    .format_args = "i4i4i4i4i4",					\
   };									\
-  DECLARE_ETD(_tc, _e, 4);						\
+  DECLARE_ETD(_tc, _e, 5);						\
   ed->data[0] = _tc->snd_una - _tc->iss;				\
   ed->data[1] = _tc->cwnd;						\
   ed->data[2] = _tc->snd_wnd;						\
   ed->data[3] = tcp_flight_size(_tc);					\
+  ed->data[4] = _tc->rcv_wnd;						\
 }
 
 #define TCP_EVT_PKTIZE_HANDLER(_tc, ...)				\
@@ -302,7 +318,7 @@ typedef enum _tcp_dbg_evt
 {									\
   ELOG_TYPE_DECLARE (_e) =						\
   {									\
-    .format = "in: %s len %u written %d rcv_nxt %u free wnd %d",	\
+    .format = "in: %s len %u written %d rcv_nxt %u rcv_wnd(o) %d",	\
     .format_args = "t4i4i4i4i4",					\
     .n_enum_strings = 2,                                        	\
     .enum_strings = {                                           	\
@@ -338,7 +354,7 @@ typedef enum _tcp_dbg_evt
     .enum_strings = {                                           	\
       "retransmit",                                             	\
       "delack",                                                 	\
-      "BUG",                                                    	\
+      "persist",                                                    	\
       "keep",                                                   	\
       "waitclose",                                              	\
       "retransmit syn",                                         	\
@@ -354,7 +370,7 @@ typedef enum _tcp_dbg_evt
 {									\
   ELOG_TYPE_DECLARE (_e) =						\
   {									\
-    .format = "seg-inv: seq %u end %u rcv_las %u rcv_nxt %u wnd %u",	\
+    .format = "seg-inv: seq %u end %u rcv_las %u rcv_nxt %u rcv_wnd %u",\
     .format_args = "i4i4i4i4i4",					\
   };									\
   DECLARE_ETD(_tc, _e, 5);						\
@@ -445,6 +461,24 @@ typedef enum _tcp_dbg_evt
 #define TCP_EVT_CC_PACK_HANDLER(_tc, ...)
 #endif
 
+#define TCP_EVT_RCV_WND_SHRUNK_HANDLER(_tc, _obs, _av, ...)		\
+{									\
+if (_av > 0) 								\
+{									\
+  ELOG_TYPE_DECLARE (_e) =						\
+  {									\
+    .format = "huh?: rcv_wnd %u obsd %u av %u rcv_nxt %u rcv_las %u",	\
+    .format_args = "i4i4i4i4i4",					\
+  };									\
+  DECLARE_ETD(_tc, _e, 5);						\
+  ed->data[0] = _tc->rcv_wnd;						\
+  ed->data[1] = _obs;							\
+  ed->data[2] = _av;							\
+  ed->data[3] = _tc->rcv_nxt - _tc->irs;				\
+  ed->data[4] = _tc->rcv_las - _tc->irs;				\
+}									\
+}
+
 #if TCP_DBG_VERBOSE
 #define TCP_EVT_SND_WND_HANDLER(_tc, ...)				\
 {									\
diff --git a/src/vnet/tcp/tcp_input.c b/src/vnet/tcp/tcp_input.c
index 5d11985f..a8224dc2 100644
--- a/src/vnet/tcp/tcp_input.c
+++ b/src/vnet/tcp/tcp_input.c
@@ -276,8 +276,7 @@ tcp_segment_validate (vlib_main_t * vm, tcp_connection_t * tc0,
       if (tc0->rcv_wnd == 0
 	  && tc0->rcv_nxt == vnet_buffer (b0)->tcp.seq_number)
 	{
-	  /* Make it look as if there's nothing to dequeue */
-	  vnet_buffer (b0)->tcp.seq_end = vnet_buffer (b0)->tcp.seq_number;
+	  /* TODO Should segment be tagged?  */
 	}
       else
 	{
@@ -375,7 +374,6 @@ tcp_update_rtt (tcp_connection_t * tc, u32 ack)
   if (tc->rtt_seq && seq_gt (ack, tc->rtt_seq) && !tc->rto_boff)
     {
       mrtt = tcp_time_now () - tc->rtt_ts;
-      tc->rtt_seq = 0;
     }
 
   /* As per RFC7323 TSecr can be used for RTTM only if the segment advances
@@ -395,6 +393,10 @@ tcp_update_rtt (tcp_connection_t * tc, u32 ack)
 
   tc->rto = clib_min (tc->srtt + (tc->rttvar << 2), TCP_RTO_MAX);
 
+  /* Allow measuring of RTT and make sure boff is 0 */
+  tc->rtt_seq = 0;
+  tc->rto_boff = 0;
+
   return 1;
 }
 
@@ -408,11 +410,7 @@ tcp_dequeue_acked (tcp_connection_t * tc, u32 ack)
   stream_session_dequeue_drop (&tc->connection, tc->bytes_acked);
 
   /* Update rtt and rto */
-  if (tcp_update_rtt (tc, ack))
-    {
-      /* Good ACK received and valid RTT, make sure retransmit backoff is 0 */
-      tc->rto_boff = 0;
-    }
+  tcp_update_rtt (tc, ack);
 }
 
 /**
@@ -672,6 +670,13 @@ tcp_update_snd_wnd (tcp_connection_t * tc, u32 seq, u32 ack, u32 snd_wnd)
       tc->snd_wl1 = seq;
       tc->snd_wl2 = ack;
       TCP_EVT_DBG (TCP_EVT_SND_WND, tc);
+
+      /* Set probe timer if we just got 0 wnd */
+      if (tc->snd_wnd < tc->snd_mss
+	  && !tcp_timer_is_active (tc, TCP_TIMER_PERSIST))
+	tcp_persist_timer_set (tc);
+      else
+	tcp_persist_timer_reset (tc);
     }
 }
 
@@ -686,6 +691,10 @@ tcp_cc_congestion (tcp_connection_t * tc)
 void
 tcp_cc_recover (tcp_connection_t * tc)
 {
+  /* TODO: check if time to recover was small. It might be that RTO popped
+   * too soon.
+   */
+
   tc->cc_algo->recovered (tc);
 
   tc->rtx_bytes = 0;
@@ -695,8 +704,7 @@ tcp_cc_recover (tcp_connection_t * tc)
   tc->cc_algo->rcv_ack (tc);
   tc->tsecr_last_ack = tc->opt.tsecr;
 
-  tcp_fastrecovery_1_smss_off (tc);
-  tcp_fastrecovery_off (tc);
+  tcp_cong_recovery_off (tc);
 
   TCP_EVT_DBG (TCP_EVT_CC_EVT, tc, 3);
 }
@@ -706,7 +714,7 @@ tcp_cc_rcv_ack (tcp_connection_t * tc, vlib_buffer_t * b)
 {
   u8 partial_ack;
 
-  if (tcp_in_recovery (tc))
+  if (tcp_in_cong_recovery (tc))
     {
       partial_ack = seq_lt (tc->snd_una, tc->snd_congestion);
       if (!partial_ack)
@@ -724,10 +732,10 @@ tcp_cc_rcv_ack (tcp_connection_t * tc, vlib_buffer_t * b)
 
 	  /* In case snd_nxt is still in the past and output tries to
 	   * shove some new bytes */
-	  tc->snd_nxt = tc->snd_una;
+	  tc->snd_nxt = tc->snd_una_max;
 
 	  /* XXX need proper RFC6675 support */
-	  if (tc->sack_sb.last_sacked_bytes)
+	  if (tc->sack_sb.last_sacked_bytes && !tcp_in_recovery (tc))
 	    {
 	      tcp_fast_retransmit (tc);
 	    }
@@ -735,9 +743,6 @@ tcp_cc_rcv_ack (tcp_connection_t * tc, vlib_buffer_t * b)
 	    {
 	      /* Retransmit first unacked segment */
 	      tcp_retransmit_first_unacked (tc);
-	      /* If window allows, send 1 SMSS of new data */
-	      if (seq_lt (tc->snd_nxt, tc->snd_congestion))
-		tc->snd_nxt = tc->snd_congestion;
 	    }
 	}
     }
@@ -814,10 +819,11 @@ tcp_rcv_ack (tcp_connection_t * tc, vlib_buffer_t * b,
 	  return -1;
 	}
 
-      tc->snd_nxt = vnet_buffer (b)->tcp.ack_number;
-      *error = TCP_ERROR_ACK_FUTURE;
       TCP_EVT_DBG (TCP_EVT_ACK_RCV_ERR, tc, 2,
 		   vnet_buffer (b)->tcp.ack_number);
+
+      tc->snd_nxt = vnet_buffer (b)->tcp.ack_number;
+      *error = TCP_ERROR_ACK_FUTURE;
     }
 
   /* If old ACK, probably it's an old dupack */
@@ -863,7 +869,7 @@ tcp_rcv_ack (tcp_connection_t * tc, vlib_buffer_t * b,
    * timer. */
   if (tc->bytes_acked)
     {
-      TCP_EVT_DBG (TCP_EVT_ACK_RCVD, tc, vnet_buffer (b)->tcp.ack_number);
+      TCP_EVT_DBG (TCP_EVT_ACK_RCVD, tc);
 
       /* Updates congestion control (slow start/congestion avoidance) */
       tcp_cc_rcv_ack (tc, b);
@@ -966,11 +972,14 @@ tcp_session_enqueue_data (tcp_connection_t * tc, vlib_buffer_t * b,
       tc->rcv_nxt += written;
 
       /* Depending on how fast the app is, all remaining buffers in burst will
-       * not be enqueued. Should we inform peer of the damage? XXX */
+       * not be enqueued. Inform peer */
+      tc->flags |= TCP_CONN_SNDACK;
+
       return TCP_ERROR_PARTIALLY_ENQUEUED;
     }
   else
     {
+      tc->flags |= TCP_CONN_SNDACK;
       return TCP_ERROR_FIFO_FULL;
     }
 
@@ -1101,25 +1110,17 @@ tcp_segment_rcv (tcp_main_t * tm, tcp_connection_t * tc, vlib_buffer_t * b,
       goto done;
     }
 
-  if (PREDICT_FALSE (error == TCP_ERROR_FIFO_FULL))
-    *next0 = TCP_NEXT_DROP;
-
   /* Check if ACK can be delayed */
-  if (!tcp_can_delack (tc))
-    {
-      /* Nothing to do for pure ACKs XXX */
-      if (n_data_bytes == 0)
-	goto done;
-
-      *next0 = tcp_next_output (tc->c_is_ip4);
-      tcp_make_ack (tc, b);
-    }
-  else
+  if (tcp_can_delack (tc))
     {
       if (!tcp_timer_is_active (tc, TCP_TIMER_DELACK))
 	tcp_timer_set (tc, TCP_TIMER_DELACK, TCP_DELACK_TIME);
+      goto done;
     }
 
+  *next0 = tcp_next_output (tc->c_is_ip4);
+  tcp_make_ack (tc, b);
+
 done:
   return error;
 }
@@ -2084,6 +2085,7 @@ tcp46_listen_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
 
 	  child0->irs = vnet_buffer (b0)->tcp.seq_number;
 	  child0->rcv_nxt = vnet_buffer (b0)->tcp.seq_number + 1;
+	  child0->rcv_las = child0->rcv_nxt;
 	  child0->state = TCP_STATE_SYN_RCVD;
 
 	  /* RFC1323: TSval timestamps sent on {SYN} and {SYN,ACK}
diff --git a/src/vnet/tcp/tcp_output.c b/src/vnet/tcp/tcp_output.c
index a671f728..ea157bd7 100644
--- a/src/vnet/tcp/tcp_output.c
+++ b/src/vnet/tcp/tcp_output.c
@@ -155,8 +155,7 @@ tcp_update_rcv_wnd (tcp_connection_t * tc)
   max_fifo = stream_session_fifo_size (&tc->connection);
 
   ASSERT (tc->opt.mss < max_fifo);
-
-  if (available_space < tc->opt.mss && available_space < max_fifo / 8)
+  if (available_space < tc->opt.mss && available_space < max_fifo >> 3)
     available_space = 0;
 
   /*
@@ -170,16 +169,21 @@ tcp_update_rcv_wnd (tcp_connection_t * tc)
   /* Bad. Thou shalt not shrink */
   if (available_space < observed_wnd)
     {
-      /* Does happen! */
       wnd = observed_wnd;
+      TCP_EVT_DBG (TCP_EVT_RCV_WND_SHRUNK, tc, observed_wnd, available_space);
     }
   else
     {
       wnd = available_space;
     }
 
-  if (wnd && ((wnd << tc->rcv_wscale) >> tc->rcv_wscale != wnd))
-    wnd += 1 << tc->rcv_wscale;
+  /* Make sure we have a multiple of rcv_wscale */
+  if (wnd && tc->rcv_wscale)
+    {
+      wnd &= ~(1 << tc->rcv_wscale);
+      if (wnd == 0)
+	wnd = 1 << tc->rcv_wscale;
+    }
 
   tc->rcv_wnd = clib_min (wnd, TCP_WND_MAX << tc->rcv_wscale);
 }
@@ -462,8 +466,9 @@ tcp_make_ack (tcp_connection_t * tc, vlib_buffer_t * b)
 
   tcp_reuse_buffer (vm, b);
   tcp_make_ack_i (tc, b, TCP_STATE_ESTABLISHED, TCP_FLAG_ACK);
-  vnet_buffer (b)->tcp.flags = TCP_BUF_FLAG_ACK;
   TCP_EVT_DBG (TCP_EVT_ACK_SENT, tc);
+  vnet_buffer (b)->tcp.flags = TCP_BUF_FLAG_ACK;
+  tc->rcv_las = tc->rcv_nxt;
 }
 
 /**
@@ -908,6 +913,7 @@ tcp_push_hdr_i (tcp_connection_t * tc, vlib_buffer_t * b,
   vnet_buffer (b)->tcp.connection_index = tc->c_c_index;
 
   tc->snd_nxt += data_len;
+
   /* TODO this is updated in output as well ... */
   if (tc->snd_nxt > tc->snd_una_max)
     tc->snd_una_max = tc->snd_nxt;
@@ -929,7 +935,6 @@ tcp_send_ack (tcp_connection_t * tc)
 
   /* Fill in the ACK */
   tcp_make_ack (tc, b);
-
   tcp_enqueue_to_output (vm, b, bi, tc->c_is_ip4);
 }
 
@@ -942,7 +947,6 @@ tcp_timer_delack_handler (u32 index)
 
   tc = tcp_connection_get (index, thread_index);
   tc->timers[TCP_TIMER_DELACK] = TCP_TIMER_HANDLE_INVALID;
-//  tc->flags &= ~TCP_CONN_DELACK;
   tcp_send_ack (tc);
 }
 
@@ -995,7 +999,7 @@ done:
  * Reset congestion control, switch cwnd to loss window and try again.
  */
 static void
-tcp_rtx_timeout_cc_recover (tcp_connection_t * tc)
+tcp_rtx_timeout_cc (tcp_connection_t * tc)
 {
   /* Cleanly recover cc (also clears up fast retransmit) */
   if (tcp_in_fastrecovery (tc))
@@ -1008,6 +1012,7 @@ tcp_rtx_timeout_cc_recover (tcp_connection_t * tc)
     }
 
   /* Start again from the beginning */
+  tcp_recovery_on (tc);
   tc->cwnd = tcp_loss_wnd (tc);
   tc->snd_congestion = tc->snd_una_max;
 }
@@ -1048,7 +1053,7 @@ tcp_timer_retransmit_handler_i (u32 index, u8 is_syn)
     {
       /* First retransmit timeout */
       if (tc->rto_boff == 1)
-	tcp_rtx_timeout_cc_recover (tc);
+	tcp_rtx_timeout_cc (tc);
 
       /* Exponential backoff */
       tc->rto = clib_min (tc->rto << 1, TCP_RTO_MAX);
@@ -1114,6 +1119,8 @@ tcp_timer_retransmit_handler_i (u32 index, u8 is_syn)
     {
       ASSERT (tc->state == TCP_STATE_SYN_SENT);
 
+      TCP_EVT_DBG (TCP_EVT_SYN_RTX, tc);
+
       /* This goes straight to ipx_lookup */
       tcp_push_ip_hdr (tm, tc, b);
       tcp_enqueue_to_ip_lookup (vm, b, bi, tc->c_is_ip4);
@@ -1136,6 +1143,55 @@ tcp_timer_retransmit_syn_handler (u32 index)
   tcp_timer_retransmit_handler_i (index, 1);
 }
 
+/**
+ * Got 0 snd_wnd from peer, try to do something about it.
+ *
+ */
+void
+tcp_timer_persist_handler (u32 index)
+{
+  tcp_main_t *tm = vnet_get_tcp_main ();
+  vlib_main_t *vm = vlib_get_main ();
+  u32 thread_index = os_get_cpu_number ();
+  tcp_connection_t *tc;
+  vlib_buffer_t *b;
+  u32 bi, n_bytes;
+
+  tc = tcp_connection_get (index, thread_index);
+
+  /* Make sure timer handle is set to invalid */
+  tc->timers[TCP_TIMER_PERSIST] = TCP_TIMER_HANDLE_INVALID;
+
+  /* Problem already solved or worse */
+  if (tc->snd_wnd > tc->snd_mss || tcp_in_recovery (tc))
+    return;
+
+  /* Increment RTO backoff */
+  tc->rto_boff += 1;
+  tc->rto = clib_min (tc->rto << 1, TCP_RTO_MAX);
+
+  /* Try to force the first unsent segment  */
+  tcp_get_free_buffer_index (tm, &bi);
+  b = vlib_get_buffer (vm, bi);
+  n_bytes = stream_session_peek_bytes (&tc->connection,
+				       vlib_buffer_get_current (b),
+				       tc->snd_una_max - tc->snd_una,
+				       tc->snd_mss);
+  /* Nothing to send */
+  if (n_bytes == 0)
+    {
+      tcp_return_buffer (tm);
+      return;
+    }
+
+  b->current_length = n_bytes;
+  tcp_push_hdr_i (tc, b, tc->state);
+  tcp_enqueue_to_output (vm, b, bi, tc->c_is_ip4);
+
+  /* Re-enable persist timer */
+  tcp_persist_timer_set (tc);
+}
+
 /**
  * Retransmit first unacked segment
  */
@@ -1329,9 +1385,6 @@ tcp46_output_inline (vlib_main_t * vm,
 		}
 	    }
 
-	  /* Retransmitted SYNs do reach this but it should be harmless */
-	  tc0->rcv_las = tc0->rcv_nxt;
-
 	  /* Stop DELACK timer and fix flags */
 	  tc0->flags &= ~(TCP_CONN_SNDACK);
 	  if (tcp_timer_is_active (tc0, TCP_TIMER_DELACK))
-- 
cgit 1.2.3-korg


From 6cf30adc2cd3aa818e5d97cf71ea8b2fc2aaefa7 Mon Sep 17 00:00:00 2001
From: Florin Coras <fcoras@cisco.com>
Date: Tue, 4 Apr 2017 23:08:23 -0700
Subject: Session layer refactoring

Major refactoring of the session layer api
- Add attatch api for application binding to the the session layer
- Simplify listen/connect calls
- Update application CLI
- Add transport endpoint to accept callback
- Associate segment manager to application and allow for multiple
  binds/connects per app

Additional:
- svm fifo cleanup
- add fifo free, format fns
- add fifo offset enqueue unit test

Change-Id: Id93a65047de61afc2bf3d58c9b544339c02065af
Signed-off-by: Florin Coras <fcoras@cisco.com>
Signed-off-by: Dave Barach <dave@barachs.net>
---
 src/scripts/vnet/uri/udp                 |   3 +-
 src/svm/svm_fifo.c                       |  66 ++-
 src/svm/svm_fifo.h                       |  32 +-
 src/svm/svm_fifo_segment.h               |  14 +-
 src/uri/uri_tcp_test.c                   | 315 +++++++++-----
 src/uri/uri_udp_test.c                   | 326 +++++++++------
 src/vnet.am                              |   2 +
 src/vnet/api_errno.h                     |   4 +-
 src/vnet/session/application.c           | 458 +++++++++++++++------
 src/vnet/session/application.h           |  77 ++--
 src/vnet/session/application_interface.c | 278 ++++++-------
 src/vnet/session/application_interface.h |  45 +-
 src/vnet/session/segment_manager.c       | 342 ++++++++++++++++
 src/vnet/session/segment_manager.h       | 106 +++++
 src/vnet/session/session.api             | 237 +++++------
 src/vnet/session/session.c               | 564 +++++++------------------
 src/vnet/session/session.h               | 175 +++++---
 src/vnet/session/session_api.c           | 678 ++++++++++++-------------------
 src/vnet/session/transport.h             |  23 +-
 src/vnet/tcp/builtin_client.c            | 161 ++++++--
 src/vnet/tcp/builtin_client.h            |   7 +-
 src/vnet/tcp/builtin_server.c            | 206 +++++++++-
 src/vnet/tcp/tcp.c                       |  20 +-
 src/vnet/tcp/tcp.h                       |   9 +-
 src/vnet/tcp/tcp_input.c                 |   7 +-
 src/vnet/tcp/tcp_test.c                  | 127 +++++-
 src/vnet/udp/builtin_server.c            |  34 +-
 27 files changed, 2601 insertions(+), 1715 deletions(-)
 create mode 100644 src/vnet/session/segment_manager.c
 create mode 100644 src/vnet/session/segment_manager.h

(limited to 'src/svm/svm_fifo.c')

diff --git a/src/scripts/vnet/uri/udp b/src/scripts/vnet/uri/udp
index ca13b83c..c7628f49 100644
--- a/src/scripts/vnet/uri/udp
+++ b/src/scripts/vnet/uri/udp
@@ -1,5 +1,5 @@
 loop create
-set int ip address loop0 10.0.0.1/32
+set int ip address loop0 6.0.0.1/32
 set int state loop0 up
 
 packet-generator new {
@@ -17,3 +17,4 @@ packet-generator new {
     incrementing 100
   }
 }
+session enable
diff --git a/src/svm/svm_fifo.c b/src/svm/svm_fifo.c
index cc84feb9..097bab77 100644
--- a/src/svm/svm_fifo.c
+++ b/src/svm/svm_fifo.c
@@ -20,8 +20,6 @@ svm_fifo_t *
 svm_fifo_create (u32 data_size_in_bytes)
 {
   svm_fifo_t *f;
-  pthread_mutexattr_t attr;
-  pthread_condattr_t cattr;
 
   f = clib_mem_alloc_aligned_or_null (sizeof (*f) + data_size_in_bytes,
 				      CLIB_CACHE_LINE_BYTES);
@@ -32,29 +30,16 @@ svm_fifo_create (u32 data_size_in_bytes)
   f->nitems = data_size_in_bytes;
   f->ooos_list_head = OOO_SEGMENT_INVALID_INDEX;
 
-  memset (&attr, 0, sizeof (attr));
-  memset (&cattr, 0, sizeof (cattr));
-
-  if (pthread_mutexattr_init (&attr))
-    clib_unix_warning ("mutexattr_init");
-  if (pthread_mutexattr_setpshared (&attr, PTHREAD_PROCESS_SHARED))
-    clib_unix_warning ("pthread_mutexattr_setpshared");
-  if (pthread_mutex_init (&f->mutex, &attr))
-    clib_unix_warning ("mutex_init");
-  if (pthread_mutexattr_destroy (&attr))
-    clib_unix_warning ("mutexattr_destroy");
-  if (pthread_condattr_init (&cattr))
-    clib_unix_warning ("condattr_init");
-  if (pthread_condattr_setpshared (&cattr, PTHREAD_PROCESS_SHARED))
-    clib_unix_warning ("condattr_setpshared");
-  if (pthread_cond_init (&f->condvar, &cattr))
-    clib_unix_warning ("cond_init1");
-  if (pthread_condattr_destroy (&cattr))
-    clib_unix_warning ("cond_init2");
-
   return (f);
 }
 
+void
+svm_fifo_free (svm_fifo_t * f)
+{
+  pool_free (f->ooo_segments);
+  clib_mem_free (f);
+}
+
 always_inline ooo_segment_t *
 ooo_segment_new (svm_fifo_t * f, u32 start, u32 length)
 {
@@ -567,6 +552,43 @@ svm_fifo_dequeue_drop (svm_fifo_t * f, int pid, u32 max_bytes)
   return total_drop_bytes;
 }
 
+u8 *
+format_svm_fifo (u8 * s, va_list * args)
+{
+  svm_fifo_t *f = va_arg (*args, svm_fifo_t *);
+  int verbose = va_arg (*args, int);
+
+  s = format (s, "cursize %u nitems %u has_event %d\n",
+	      f->cursize, f->nitems, f->has_event);
+  s = format (s, "head %d tail %d\n", f->head, f->tail);
+
+  if (verbose > 1)
+    s = format
+      (s, "server session %d thread %d client session %d thread %d\n",
+       f->server_session_index, f->server_thread_index,
+       f->client_session_index, f->client_thread_index);
+
+  if (verbose)
+    {
+      ooo_segment_t *seg;
+      u32 seg_index;
+
+      s =
+	format (s, "ooo pool %d active elts\n", pool_elts (f->ooo_segments));
+
+      seg_index = f->ooos_list_head;
+
+      while (seg_index != OOO_SEGMENT_INVALID_INDEX)
+	{
+	  seg = pool_elt_at_index (f->ooo_segments, seg_index);
+	  s = format (s, "  pos %u, len %u next %d\n",
+		      seg->fifo_position, seg->length, seg->next);
+	  seg_index = seg->next;
+	}
+    }
+  return s;
+}
+
 /*
  * fd.io coding-style-patch-verification: ON
  *
diff --git a/src/svm/svm_fifo.h b/src/svm/svm_fifo.h
index 80e5b0f2..9beb63f5 100644
--- a/src/svm/svm_fifo.h
+++ b/src/svm/svm_fifo.h
@@ -48,10 +48,6 @@ typedef struct
   u32 nitems;
     CLIB_CACHE_LINE_ALIGN_MARK (end_cursize);
 
-  pthread_mutex_t mutex;	/* 8 bytes */
-  pthread_cond_t condvar;	/* 8 bytes */
-  svm_lock_tag_t tag;
-
   volatile u8 has_event;	/**< non-zero if deq event exists */
   u32 owner_pid;
 
@@ -60,6 +56,7 @@ typedef struct
   u32 client_session_index;
   u8 server_thread_index;
   u8 client_thread_index;
+  u32 segment_manager;
     CLIB_CACHE_LINE_ALIGN_MARK (end_shared);
   u32 head;
     CLIB_CACHE_LINE_ALIGN_MARK (end_consumer);
@@ -74,30 +71,6 @@ typedef struct
     CLIB_CACHE_LINE_ALIGN_MARK (data);
 } svm_fifo_t;
 
-static inline int
-svm_fifo_lock (svm_fifo_t * f, u32 pid, u32 tag, int nowait)
-{
-  if (PREDICT_TRUE (nowait == 0))
-    pthread_mutex_lock (&f->mutex);
-  else
-    {
-      if (pthread_mutex_trylock (&f->mutex))
-	return -1;
-    }
-  f->owner_pid = pid;
-  f->tag = tag;
-  return 0;
-}
-
-static inline void
-svm_fifo_unlock (svm_fifo_t * f)
-{
-  f->owner_pid = 0;
-  f->tag = 0;
-  CLIB_MEMORY_BARRIER ();
-  pthread_mutex_unlock (&f->mutex);
-}
-
 static inline u32
 svm_fifo_max_dequeue (svm_fifo_t * f)
 {
@@ -139,6 +112,7 @@ svm_fifo_unset_event (svm_fifo_t * f)
 }
 
 svm_fifo_t *svm_fifo_create (u32 data_size_in_bytes);
+void svm_fifo_free (svm_fifo_t * f);
 
 int svm_fifo_enqueue_nowait (svm_fifo_t * f, int pid, u32 max_bytes,
 			     u8 * copy_from_here);
@@ -154,6 +128,8 @@ int svm_fifo_peek (svm_fifo_t * f, int pid, u32 offset, u32 max_bytes,
 		   u8 * copy_here);
 int svm_fifo_dequeue_drop (svm_fifo_t * f, int pid, u32 max_bytes);
 
+format_function_t format_svm_fifo;
+
 always_inline ooo_segment_t *
 svm_fifo_newest_ooo_segment (svm_fifo_t * f)
 {
diff --git a/src/svm/svm_fifo_segment.h b/src/svm/svm_fifo_segment.h
index ecb5653a..9ab47a4c 100644
--- a/src/svm/svm_fifo_segment.h
+++ b/src/svm/svm_fifo_segment.h
@@ -55,6 +55,18 @@ svm_fifo_get_segment (u32 segment_index)
   return vec_elt_at_index (ssm->segments, segment_index);
 }
 
+static inline u8
+svm_fifo_segment_has_fifos (svm_fifo_segment_private_t * fifo_segment)
+{
+  return vec_len ((svm_fifo_t **) fifo_segment->h->fifos) != 0;
+}
+
+static inline svm_fifo_t **
+svm_fifo_segment_get_fifos (svm_fifo_segment_private_t * fifo_segment)
+{
+  return (svm_fifo_t **) fifo_segment->h->fifos;
+}
+
 #define foreach_ssvm_fifo_segment_api_error             \
 _(OUT_OF_SPACE, "Out of space in segment", -200)
 
@@ -73,9 +85,7 @@ svm_fifo_t *svm_fifo_segment_alloc_fifo (svm_fifo_segment_private_t * s,
 					 u32 data_size_in_bytes);
 void svm_fifo_segment_free_fifo (svm_fifo_segment_private_t * s,
 				 svm_fifo_t * f);
-
 void svm_fifo_segment_init (u64 baseva, u32 timeout_in_seconds);
-
 u32 svm_fifo_segment_index (svm_fifo_segment_private_t * s);
 
 #endif /* __included_ssvm_fifo_segment_h__ */
diff --git a/src/uri/uri_tcp_test.c b/src/uri/uri_tcp_test.c
index e2834817..c057e06e 100644
--- a/src/uri/uri_tcp_test.c
+++ b/src/uri/uri_tcp_test.c
@@ -15,8 +15,6 @@
 
 #include <stdio.h>
 #include <signal.h>
-#include <vlib/vlib.h>
-#include <vnet/vnet.h>
 #include <svm/svm_fifo_segment.h>
 #include <vlibmemory/api.h>
 #include <vpp/api/vpe_msg_enum.h>
@@ -47,8 +45,7 @@ typedef struct
   svm_fifo_t *server_rx_fifo;
   svm_fifo_t *server_tx_fifo;
 
-  u32 vpp_session_index;
-  u32 vpp_session_thread;
+  u32 vpp_session_handle;
 } session_t;
 
 typedef enum
@@ -116,7 +113,7 @@ typedef struct
   pthread_t client_rx_thread_handle;
   u32 client_bytes_received;
   u8 test_return_packets;
-  u32 bytes_to_send;
+  u64 bytes_to_send;
 
   /* convenience */
   svm_fifo_segment_main_t *segment_main;
@@ -152,6 +149,88 @@ wait_for_state_change (uri_tcp_test_main_t * utm, connection_state_t state)
   return -1;
 }
 
+void
+application_attach (uri_tcp_test_main_t * utm)
+{
+  vl_api_application_attach_t *bmp;
+  u32 fifo_size = 3 << 20;
+  bmp = vl_msg_api_alloc (sizeof (*bmp));
+  memset (bmp, 0, sizeof (*bmp));
+
+  bmp->_vl_msg_id = ntohs (VL_API_APPLICATION_ATTACH);
+  bmp->client_index = utm->my_client_index;
+  bmp->context = ntohl (0xfeedface);
+  bmp->options[SESSION_OPTIONS_FLAGS] =
+    SESSION_OPTIONS_FLAGS_USE_FIFO | SESSION_OPTIONS_FLAGS_ADD_SEGMENT;
+  bmp->options[SESSION_OPTIONS_RX_FIFO_SIZE] = fifo_size;
+  bmp->options[SESSION_OPTIONS_TX_FIFO_SIZE] = fifo_size;
+  bmp->options[SESSION_OPTIONS_ADD_SEGMENT_SIZE] = 128 << 20;
+  bmp->options[SESSION_OPTIONS_SEGMENT_SIZE] = 256 << 20;
+  vl_msg_api_send_shmem (utm->vl_input_queue, (u8 *) & bmp);
+}
+
+void
+application_detach (uri_tcp_test_main_t * utm)
+{
+  vl_api_application_detach_t *bmp;
+  bmp = vl_msg_api_alloc (sizeof (*bmp));
+  memset (bmp, 0, sizeof (*bmp));
+
+  bmp->_vl_msg_id = ntohs (VL_API_APPLICATION_DETACH);
+  bmp->client_index = utm->my_client_index;
+  bmp->context = ntohl (0xfeedface);
+  vl_msg_api_send_shmem (utm->vl_input_queue, (u8 *) & bmp);
+}
+
+static void
+vl_api_application_attach_reply_t_handler (vl_api_application_attach_reply_t *
+					   mp)
+{
+  uri_tcp_test_main_t *utm = &uri_tcp_test_main;
+  svm_fifo_segment_create_args_t _a, *a = &_a;
+  int rv;
+
+  if (mp->retval)
+    {
+      uword *errp = hash_get (utm->error_string_by_error_number, mp->retval);
+      clib_warning ("attach failed: %s", *errp);
+      utm->state = STATE_FAILED;
+      return;
+    }
+
+  if (mp->segment_name_length == 0)
+    {
+      clib_warning ("segment_name_length zero");
+      return;
+    }
+
+  a->segment_name = (char *) mp->segment_name;
+  a->segment_size = mp->segment_size;
+
+  ASSERT (mp->app_event_queue_address);
+
+  /* Attach to the segment vpp created */
+  rv = svm_fifo_segment_attach (a);
+  if (rv)
+    {
+      clib_warning ("svm_fifo_segment_attach ('%s') failed",
+		    mp->segment_name);
+      return;
+    }
+
+  utm->our_event_queue =
+    (unix_shared_memory_queue_t *) mp->app_event_queue_address;
+
+}
+
+static void
+vl_api_application_detach_reply_t_handler (vl_api_application_detach_reply_t *
+					   mp)
+{
+  if (mp->retval)
+    clib_warning ("detach returned with err: %d", mp->retval);
+}
+
 static void
 init_error_string_table (uri_tcp_test_main_t * utm)
 {
@@ -239,21 +318,18 @@ vl_api_disconnect_session_t_handler (vl_api_disconnect_session_t * mp)
   vl_api_disconnect_session_reply_t *rmp;
   uword *p;
   int rv = 0;
-  u64 key;
-
-  key = (((u64) mp->session_thread_index) << 32) | (u64) mp->session_index;
 
-  p = hash_get (utm->session_index_by_vpp_handles, key);
+  p = hash_get (utm->session_index_by_vpp_handles, mp->handle);
 
   if (p)
     {
       session = pool_elt_at_index (utm->sessions, p[0]);
-      hash_unset (utm->session_index_by_vpp_handles, key);
+      hash_unset (utm->session_index_by_vpp_handles, mp->handle);
       pool_put (utm->sessions, session);
     }
   else
     {
-      clib_warning ("couldn't find session key %llx", key);
+      clib_warning ("couldn't find session key %llx", mp->handle);
       rv = -11;
     }
 
@@ -264,8 +340,7 @@ vl_api_disconnect_session_t_handler (vl_api_disconnect_session_t * mp)
 
   rmp->_vl_msg_id = ntohs (VL_API_DISCONNECT_SESSION_REPLY);
   rmp->retval = rv;
-  rmp->session_index = mp->session_index;
-  rmp->session_thread_index = mp->session_thread_index;
+  rmp->handle = mp->handle;
   vl_msg_api_send_shmem (utm->vl_input_queue, (u8 *) & rmp);
 }
 
@@ -277,22 +352,19 @@ vl_api_reset_session_t_handler (vl_api_reset_session_t * mp)
   vl_api_reset_session_reply_t *rmp;
   uword *p;
   int rv = 0;
-  u64 key;
 
-  key = (((u64) mp->session_thread_index) << 32) | (u64) mp->session_index;
-
-  p = hash_get (utm->session_index_by_vpp_handles, key);
+  p = hash_get (utm->session_index_by_vpp_handles, mp->handle);
 
   if (p)
     {
       session = pool_elt_at_index (utm->sessions, p[0]);
-      hash_unset (utm->session_index_by_vpp_handles, key);
+      hash_unset (utm->session_index_by_vpp_handles, mp->handle);
       pool_put (utm->sessions, session);
       utm->time_to_stop = 1;
     }
   else
     {
-      clib_warning ("couldn't find session key %llx", key);
+      clib_warning ("couldn't find session key %llx", mp->handle);
       rv = -11;
     }
 
@@ -300,8 +372,7 @@ vl_api_reset_session_t_handler (vl_api_reset_session_t * mp)
   memset (rmp, 0, sizeof (*rmp));
   rmp->_vl_msg_id = ntohs (VL_API_RESET_SESSION_REPLY);
   rmp->retval = rv;
-  rmp->session_index = mp->session_index;
-  rmp->session_thread_index = mp->session_thread_index;
+  rmp->handle = mp->handle;
   vl_msg_api_send_shmem (utm->vl_input_queue, (u8 *) & rmp);
 }
 
@@ -343,7 +414,7 @@ client_handle_fifo_event_rx (uri_tcp_test_main_t * utm,
 	{
 	  if (n_read == -2)
 	    {
-	      clib_warning ("weird!");
+//            clib_warning ("weird!");
 	      break;
 	    }
 	}
@@ -409,52 +480,19 @@ static void
 vl_api_connect_uri_reply_t_handler (vl_api_connect_uri_reply_t * mp)
 {
   uri_tcp_test_main_t *utm = &uri_tcp_test_main;
-  svm_fifo_segment_create_args_t _a, *a = &_a;
   session_t *session;
   u32 session_index;
   svm_fifo_t *rx_fifo, *tx_fifo;
   int rv;
-  u64 key;
 
   if (mp->retval)
     {
-      clib_warning ("connection failed with code: %d", mp->retval);
-      utm->state = STATE_FAILED;
-      return;
-    }
-
-  /*
-   * Attatch to segment
-   */
-
-  if (mp->segment_name_length == 0)
-    {
-      clib_warning ("segment_name_length zero");
+      uword *errp = hash_get (utm->error_string_by_error_number, -mp->retval);
+      clib_warning ("connection failed with code: %s", *errp);
       utm->state = STATE_FAILED;
       return;
     }
 
-  a->segment_name = (char *) mp->segment_name;
-  a->segment_size = mp->segment_size;
-
-  ASSERT (mp->client_event_queue_address);
-
-  /* Attach to the segment vpp created */
-  rv = svm_fifo_segment_attach (a);
-  if (rv)
-    {
-      clib_warning ("svm_fifo_segment_attach ('%s') failed",
-		    mp->segment_name);
-      return;
-    }
-
-  /*
-   * Save the queues
-   */
-
-  utm->our_event_queue = (unix_shared_memory_queue_t *)
-    mp->client_event_queue_address;
-
   utm->vpp_event_queue = (unix_shared_memory_queue_t *)
     mp->vpp_event_queue_address;
 
@@ -472,16 +510,14 @@ vl_api_connect_uri_reply_t_handler (vl_api_connect_uri_reply_t * mp)
 
   session->server_rx_fifo = rx_fifo;
   session->server_tx_fifo = tx_fifo;
-  session->vpp_session_index = mp->session_index;
-  session->vpp_session_thread = mp->session_thread_index;
+  session->vpp_session_handle = mp->handle;
 
   /* Save handle */
   utm->connected_session_index = session_index;
   utm->state = STATE_READY;
 
   /* Add it to lookup table */
-  key = (((u64) mp->session_thread_index) << 32) | (u64) mp->session_index;
-  hash_set (utm->session_index_by_vpp_handles, key, session_index);
+  hash_set (utm->session_index_by_vpp_handles, mp->handle, session_index);
 
   /* Start RX thread */
   rv = pthread_create (&utm->client_rx_thread_handle,
@@ -606,8 +642,7 @@ client_disconnect (uri_tcp_test_main_t * utm)
   memset (dmp, 0, sizeof (*dmp));
   dmp->_vl_msg_id = ntohs (VL_API_DISCONNECT_SESSION);
   dmp->client_index = utm->my_client_index;
-  dmp->session_index = connected_session->vpp_session_index;
-  dmp->session_thread_index = connected_session->vpp_session_thread;
+  dmp->handle = connected_session->vpp_session_handle;
   vl_msg_api_send_shmem (utm->vl_input_queue, (u8 *) & dmp);
 }
 
@@ -616,6 +651,7 @@ client_test (uri_tcp_test_main_t * utm)
 {
   int i;
 
+  application_attach (utm);
   client_connect (utm);
 
   if (wait_for_state_change (utm, STATE_READY))
@@ -636,47 +672,26 @@ client_test (uri_tcp_test_main_t * utm)
 
   if (wait_for_state_change (utm, STATE_START))
     {
+      clib_warning ("Disconnect failed");
       return;
     }
+  application_detach (utm);
 }
 
 static void
 vl_api_bind_uri_reply_t_handler (vl_api_bind_uri_reply_t * mp)
 {
   uri_tcp_test_main_t *utm = &uri_tcp_test_main;
-  svm_fifo_segment_create_args_t _a, *a = &_a;
-  int rv;
 
   if (mp->retval)
     {
-      clib_warning ("bind failed: %d", mp->retval);
+      uword *errp = hash_get (utm->error_string_by_error_number,
+			      -clib_net_to_host_u32 (mp->retval));
+      clib_warning ("bind failed: %s", (char *) *errp);
       utm->state = STATE_FAILED;
       return;
     }
 
-  if (mp->segment_name_length == 0)
-    {
-      clib_warning ("segment_name_length zero");
-      return;
-    }
-
-  a->segment_name = (char *) mp->segment_name;
-  a->segment_size = mp->segment_size;
-
-  ASSERT (mp->server_event_queue_address);
-
-  /* Attach to the segment vpp created */
-  rv = svm_fifo_segment_attach (a);
-  if (rv)
-    {
-      clib_warning ("svm_fifo_segment_attach ('%s') failed",
-		    mp->segment_name);
-      return;
-    }
-
-  utm->our_event_queue =
-    (unix_shared_memory_queue_t *) mp->server_event_queue_address;
-
   utm->state = STATE_READY;
 }
 
@@ -691,6 +706,89 @@ vl_api_unbind_uri_reply_t_handler (vl_api_unbind_uri_reply_t * mp)
   utm->state = STATE_START;
 }
 
+u8 *
+format_ip4_address (u8 * s, va_list * args)
+{
+  u8 *a = va_arg (*args, u8 *);
+  return format (s, "%d.%d.%d.%d", a[0], a[1], a[2], a[3]);
+}
+
+u8 *
+format_ip6_address (u8 * s, va_list * args)
+{
+  ip6_address_t *a = va_arg (*args, ip6_address_t *);
+  u32 i, i_max_n_zero, max_n_zeros, i_first_zero, n_zeros, last_double_colon;
+
+  i_max_n_zero = ARRAY_LEN (a->as_u16);
+  max_n_zeros = 0;
+  i_first_zero = i_max_n_zero;
+  n_zeros = 0;
+  for (i = 0; i < ARRAY_LEN (a->as_u16); i++)
+    {
+      u32 is_zero = a->as_u16[i] == 0;
+      if (is_zero && i_first_zero >= ARRAY_LEN (a->as_u16))
+	{
+	  i_first_zero = i;
+	  n_zeros = 0;
+	}
+      n_zeros += is_zero;
+      if ((!is_zero && n_zeros > max_n_zeros)
+	  || (i + 1 >= ARRAY_LEN (a->as_u16) && n_zeros > max_n_zeros))
+	{
+	  i_max_n_zero = i_first_zero;
+	  max_n_zeros = n_zeros;
+	  i_first_zero = ARRAY_LEN (a->as_u16);
+	  n_zeros = 0;
+	}
+    }
+
+  last_double_colon = 0;
+  for (i = 0; i < ARRAY_LEN (a->as_u16); i++)
+    {
+      if (i == i_max_n_zero && max_n_zeros > 1)
+	{
+	  s = format (s, "::");
+	  i += max_n_zeros - 1;
+	  last_double_colon = 1;
+	}
+      else
+	{
+	  s = format (s, "%s%x",
+		      (last_double_colon || i == 0) ? "" : ":",
+		      clib_net_to_host_u16 (a->as_u16[i]));
+	  last_double_colon = 0;
+	}
+    }
+
+  return s;
+}
+
+/* Format an IP46 address. */
+u8 *
+format_ip46_address (u8 * s, va_list * args)
+{
+  ip46_address_t *ip46 = va_arg (*args, ip46_address_t *);
+  ip46_type_t type = va_arg (*args, ip46_type_t);
+  int is_ip4 = 1;
+
+  switch (type)
+    {
+    case IP46_TYPE_ANY:
+      is_ip4 = ip46_address_is_ip4 (ip46);
+      break;
+    case IP46_TYPE_IP4:
+      is_ip4 = 1;
+      break;
+    case IP46_TYPE_IP6:
+      is_ip4 = 0;
+      break;
+    }
+
+  return is_ip4 ?
+    format (s, "%U", format_ip4_address, &ip46->ip4) :
+    format (s, "%U", format_ip6_address, &ip46->ip6);
+}
+
 static void
 vl_api_accept_session_t_handler (vl_api_accept_session_t * mp)
 {
@@ -699,12 +797,15 @@ vl_api_accept_session_t_handler (vl_api_accept_session_t * mp)
   svm_fifo_t *rx_fifo, *tx_fifo;
   session_t *session;
   static f64 start_time;
-  u64 key;
   u32 session_index;
+  u8 *ip_str;
 
   if (start_time == 0.0)
     start_time = clib_time_now (&utm->clib_time);
 
+  ip_str = format (0, "%U", format_ip46_address, &mp->ip, mp->is_ip4);
+  clib_warning ("Accepted session from: %s:%d", ip_str,
+		clib_net_to_host_u16 (mp->port));
   utm->vpp_event_queue = (unix_shared_memory_queue_t *)
     mp->vpp_event_queue_address;
 
@@ -721,8 +822,7 @@ vl_api_accept_session_t_handler (vl_api_accept_session_t * mp)
   session->server_tx_fifo = tx_fifo;
 
   /* Add it to lookup table */
-  key = (((u64) mp->session_thread_index) << 32) | (u64) mp->session_index;
-  hash_set (utm->session_index_by_vpp_handles, key, session_index);
+  hash_set (utm->session_index_by_vpp_handles, mp->handle, session_index);
 
   utm->state = STATE_READY;
 
@@ -741,9 +841,7 @@ vl_api_accept_session_t_handler (vl_api_accept_session_t * mp)
   rmp = vl_msg_api_alloc (sizeof (*rmp));
   memset (rmp, 0, sizeof (*rmp));
   rmp->_vl_msg_id = ntohs (VL_API_ACCEPT_SESSION_REPLY);
-  rmp->session_type = mp->session_type;
-  rmp->session_index = mp->session_index;
-  rmp->session_thread_index = mp->session_thread_index;
+  rmp->handle = mp->handle;
   vl_msg_api_send_shmem (utm->vl_input_queue, (u8 *) & rmp);
 }
 
@@ -837,22 +935,15 @@ server_handle_event_queue (uri_tcp_test_main_t * utm)
 }
 
 void
-server_bind (uri_tcp_test_main_t * utm)
+server_listen (uri_tcp_test_main_t * utm)
 {
   vl_api_bind_uri_t *bmp;
-  u32 fifo_size = 3 << 20;
   bmp = vl_msg_api_alloc (sizeof (*bmp));
   memset (bmp, 0, sizeof (*bmp));
 
   bmp->_vl_msg_id = ntohs (VL_API_BIND_URI);
   bmp->client_index = utm->my_client_index;
   bmp->context = ntohl (0xfeedface);
-  bmp->initial_segment_size = 256 << 20;	/* size of initial segment */
-  bmp->options[SESSION_OPTIONS_FLAGS] =
-    SESSION_OPTIONS_FLAGS_USE_FIFO | SESSION_OPTIONS_FLAGS_ADD_SEGMENT;
-  bmp->options[SESSION_OPTIONS_RX_FIFO_SIZE] = fifo_size;
-  bmp->options[SESSION_OPTIONS_TX_FIFO_SIZE] = fifo_size;
-  bmp->options[SESSION_OPTIONS_ADD_SEGMENT_SIZE] = 128 << 20;
   memcpy (bmp->uri, utm->uri, vec_len (utm->uri));
   vl_msg_api_send_shmem (utm->vl_input_queue, (u8 *) & bmp);
 }
@@ -874,8 +965,10 @@ server_unbind (uri_tcp_test_main_t * utm)
 void
 server_test (uri_tcp_test_main_t * utm)
 {
+  application_attach (utm);
+
   /* Bind to uri */
-  server_bind (utm);
+  server_listen (utm);
 
   if (wait_for_state_change (utm, STATE_READY))
     {
@@ -895,6 +988,8 @@ server_test (uri_tcp_test_main_t * utm)
       return;
     }
 
+  application_detach (utm);
+
   fformat (stdout, "Test complete...\n");
 }
 
@@ -916,7 +1011,9 @@ _(CONNECT_URI_REPLY, connect_uri_reply)                 \
 _(DISCONNECT_SESSION, disconnect_session)               \
 _(DISCONNECT_SESSION_REPLY, disconnect_session_reply)   \
 _(RESET_SESSION, reset_session)                         \
-_(MAP_ANOTHER_SEGMENT, map_another_segment)
+_(APPLICATION_ATTACH_REPLY, application_attach_reply)   \
+_(APPLICATION_DETACH_REPLY, application_detach_reply)	\
+_(MAP_ANOTHER_SEGMENT, map_another_segment)		\
 
 void
 uri_api_hookup (uri_tcp_test_main_t * utm)
@@ -941,7 +1038,7 @@ main (int argc, char **argv)
   u8 *heap, *uri = 0;
   u8 *bind_uri = (u8 *) "tcp://0.0.0.0/1234";
   u8 *connect_uri = (u8 *) "tcp://6.0.1.2/1234";
-  u32 bytes_to_send = 64 << 10, mbytes;
+  u64 bytes_to_send = 64 << 10, mbytes;
   u32 tmp;
   mheap_t *h;
   session_t *session;
@@ -988,10 +1085,14 @@ main (int argc, char **argv)
 	drop_packets = 1;
       else if (unformat (a, "test"))
 	test_return_packets = 1;
-      else if (unformat (a, "mbytes %d", &mbytes))
+      else if (unformat (a, "mbytes %lld", &mbytes))
 	{
 	  bytes_to_send = mbytes << 20;
 	}
+      else if (unformat (a, "gbytes %lld", &mbytes))
+	{
+	  bytes_to_send = mbytes << 30;
+	}
       else
 	{
 	  fformat (stderr, "%s: usage [master|slave]\n");
diff --git a/src/uri/uri_udp_test.c b/src/uri/uri_udp_test.c
index e6c239c1..598052bc 100644
--- a/src/uri/uri_udp_test.c
+++ b/src/uri/uri_udp_test.c
@@ -55,6 +55,7 @@ typedef enum
 {
   STATE_START,
   STATE_READY,
+  STATE_FAILED,
   STATE_DISCONNECTING,
 } connection_state_t;
 
@@ -162,6 +163,86 @@ setup_signal_handlers (void)
   return 0;
 }
 
+void
+application_attach (uri_udp_test_main_t * utm)
+{
+  vl_api_application_attach_t *bmp;
+  u32 fifo_size = 3 << 20;
+  bmp = vl_msg_api_alloc (sizeof (*bmp));
+  memset (bmp, 0, sizeof (*bmp));
+
+  bmp->_vl_msg_id = ntohs (VL_API_APPLICATION_ATTACH);
+  bmp->client_index = utm->my_client_index;
+  bmp->context = ntohl (0xfeedface);
+  bmp->options[SESSION_OPTIONS_FLAGS] =
+    SESSION_OPTIONS_FLAGS_USE_FIFO | SESSION_OPTIONS_FLAGS_ADD_SEGMENT;
+  bmp->options[SESSION_OPTIONS_RX_FIFO_SIZE] = fifo_size;
+  bmp->options[SESSION_OPTIONS_TX_FIFO_SIZE] = fifo_size;
+  bmp->options[SESSION_OPTIONS_ADD_SEGMENT_SIZE] = 128 << 20;
+  bmp->options[SESSION_OPTIONS_SEGMENT_SIZE] = 256 << 20;
+  vl_msg_api_send_shmem (utm->vl_input_queue, (u8 *) & bmp);
+}
+
+void
+application_detach (uri_udp_test_main_t * utm)
+{
+  vl_api_application_detach_t *bmp;
+  bmp = vl_msg_api_alloc (sizeof (*bmp));
+  memset (bmp, 0, sizeof (*bmp));
+
+  bmp->_vl_msg_id = ntohs (VL_API_APPLICATION_DETACH);
+  bmp->client_index = utm->my_client_index;
+  bmp->context = ntohl (0xfeedface);
+  vl_msg_api_send_shmem (utm->vl_input_queue, (u8 *) & bmp);
+}
+
+static void
+vl_api_application_attach_reply_t_handler (vl_api_application_attach_reply_t *
+					   mp)
+{
+  uri_udp_test_main_t *utm = &uri_udp_test_main;
+  svm_fifo_segment_create_args_t _a, *a = &_a;
+  int rv;
+
+  if (mp->retval)
+    {
+      clib_warning ("attach failed: %d", mp->retval);
+      utm->state = STATE_FAILED;
+      return;
+    }
+
+  if (mp->segment_name_length == 0)
+    {
+      clib_warning ("segment_name_length zero");
+      return;
+    }
+
+  a->segment_name = (char *) mp->segment_name;
+  a->segment_size = mp->segment_size;
+
+  ASSERT (mp->app_event_queue_address);
+
+  /* Attach to the segment vpp created */
+  rv = svm_fifo_segment_attach (a);
+  if (rv)
+    {
+      clib_warning ("svm_fifo_segment_attach ('%s') failed",
+		    mp->segment_name);
+      return;
+    }
+
+  utm->our_event_queue =
+    (unix_shared_memory_queue_t *) mp->app_event_queue_address;
+}
+
+static void
+vl_api_application_detach_reply_t_handler (vl_api_application_detach_reply_t *
+					   mp)
+{
+  if (mp->retval)
+    clib_warning ("detach returned with err: %d", mp->retval);
+}
+
 u8 *
 format_api_error (u8 * s, va_list * args)
 {
@@ -255,9 +336,22 @@ cut_through_thread_fn (void *arg)
 }
 
 static void
-uri_udp_slave_test (uri_udp_test_main_t * utm)
+udp_client_connect (uri_udp_test_main_t * utm)
 {
   vl_api_connect_uri_t *cmp;
+  cmp = vl_msg_api_alloc (sizeof (*cmp));
+  memset (cmp, 0, sizeof (*cmp));
+
+  cmp->_vl_msg_id = ntohs (VL_API_CONNECT_URI);
+  cmp->client_index = utm->my_client_index;
+  cmp->context = ntohl (0xfeedface);
+  memcpy (cmp->uri, utm->connect_uri, vec_len (utm->connect_uri));
+  vl_msg_api_send_shmem (utm->vl_input_queue, (u8 *) & cmp);
+}
+
+static void
+client_send (uri_udp_test_main_t * utm, session_t * session)
+{
   int i;
   u8 *test_data = 0;
   u64 bytes_received = 0, bytes_sent = 0;
@@ -265,30 +359,16 @@ uri_udp_slave_test (uri_udp_test_main_t * utm)
   int rv;
   int mypid = getpid ();
   f64 before, after, delta, bytes_per_second;
-  session_t *session;
   svm_fifo_t *rx_fifo, *tx_fifo;
   int buffer_offset, bytes_to_send = 0;
 
+  /*
+   * Prepare test data
+   */
   vec_validate (test_data, 64 * 1024 - 1);
   for (i = 0; i < vec_len (test_data); i++)
     test_data[i] = i & 0xff;
 
-  cmp = vl_msg_api_alloc (sizeof (*cmp));
-  memset (cmp, 0, sizeof (*cmp));
-
-  cmp->_vl_msg_id = ntohs (VL_API_CONNECT_URI);
-  cmp->client_index = utm->my_client_index;
-  cmp->context = ntohl (0xfeedface);
-  memcpy (cmp->uri, utm->connect_uri, vec_len (utm->connect_uri));
-  vl_msg_api_send_shmem (utm->vl_input_queue, (u8 *) & cmp);
-
-  if (wait_for_state_change (utm, STATE_READY))
-    {
-      clib_warning ("timeout waiting for STATE_READY");
-      return;
-    }
-
-  session = pool_elt_at_index (utm->sessions, utm->cut_through_session_index);
   rx_fifo = session->server_rx_fifo;
   tx_fifo = session->server_tx_fifo;
 
@@ -375,35 +455,38 @@ uri_udp_slave_test (uri_udp_test_main_t * utm)
 }
 
 static void
-vl_api_bind_uri_reply_t_handler (vl_api_bind_uri_reply_t * mp)
+uri_udp_client_test (uri_udp_test_main_t * utm)
 {
-  uri_udp_test_main_t *utm = &uri_udp_test_main;
-  svm_fifo_segment_create_args_t _a, *a = &_a;
-  int rv;
+  session_t *session;
 
-  if (mp->segment_name_length == 0)
+  application_attach (utm);
+  udp_client_connect (utm);
+
+  if (wait_for_state_change (utm, STATE_READY))
     {
-      clib_warning ("segment_name_length zero");
+      clib_warning ("timeout waiting for STATE_READY");
       return;
     }
 
-  a->segment_name = (char *) mp->segment_name;
-  a->segment_size = mp->segment_size;
+  /* Only works with cut through sessions */
+  session = pool_elt_at_index (utm->sessions, utm->cut_through_session_index);
 
-  ASSERT (mp->server_event_queue_address);
+  client_send (utm, session);
+  application_detach (utm);
+}
 
-  /* Attach to the segment vpp created */
-  rv = svm_fifo_segment_attach (a);
-  if (rv)
+static void
+vl_api_bind_uri_reply_t_handler (vl_api_bind_uri_reply_t * mp)
+{
+  uri_udp_test_main_t *utm = &uri_udp_test_main;
+
+  if (mp->retval)
     {
-      clib_warning ("svm_fifo_segment_attach ('%s') failed",
-		    mp->segment_name);
+      clib_warning ("bind failed: %d", mp->retval);
+      utm->state = STATE_FAILED;
       return;
     }
 
-  utm->our_event_queue = (unix_shared_memory_queue_t *)
-    mp->server_event_queue_address;
-
   utm->state = STATE_READY;
 }
 
@@ -427,6 +510,9 @@ vl_api_map_another_segment_t_handler (vl_api_map_another_segment_t * mp)
 		mp->segment_size);
 }
 
+/**
+ * Acting as server for redirected connect requests
+ */
 static void
 vl_api_connect_uri_t_handler (vl_api_connect_uri_t * mp)
 {
@@ -456,7 +542,6 @@ vl_api_connect_uri_t_handler (vl_api_connect_uri_t * mp)
   vec_add2 (utm->seg, seg, 1);
 
   segment_index = vec_len (sm->segments) - 1;
-
   memcpy (seg, sm->segments + segment_index, sizeof (utm->seg[0]));
 
   pool_get (utm->sessions, session);
@@ -521,7 +606,6 @@ vl_api_accept_session_t_handler (vl_api_accept_session_t * mp)
   svm_fifo_t *rx_fifo, *tx_fifo;
   session_t *session;
   static f64 start_time;
-  u64 key;
 
   if (start_time == 0.0)
     start_time = clib_time_now (&utm->clib_time);
@@ -539,9 +623,8 @@ vl_api_accept_session_t_handler (vl_api_accept_session_t * mp)
   session->server_rx_fifo = rx_fifo;
   session->server_tx_fifo = tx_fifo;
 
-  key = (((u64) mp->session_thread_index) << 32) | (u64) mp->session_index;
-
-  hash_set (utm->session_index_by_vpp_handles, key, session - utm->sessions);
+  hash_set (utm->session_index_by_vpp_handles, mp->handle,
+	    session - utm->sessions);
 
   utm->state = STATE_READY;
 
@@ -556,9 +639,7 @@ vl_api_accept_session_t_handler (vl_api_accept_session_t * mp)
   rmp = vl_msg_api_alloc (sizeof (*rmp));
   memset (rmp, 0, sizeof (*rmp));
   rmp->_vl_msg_id = ntohs (VL_API_ACCEPT_SESSION_REPLY);
-  rmp->session_type = mp->session_type;
-  rmp->session_index = mp->session_index;
-  rmp->session_thread_index = mp->session_thread_index;
+  rmp->handle = mp->handle;
   vl_msg_api_send_shmem (utm->vl_input_queue, (u8 *) & rmp);
 }
 
@@ -570,21 +651,18 @@ vl_api_disconnect_session_t_handler (vl_api_disconnect_session_t * mp)
   vl_api_disconnect_session_reply_t *rmp;
   uword *p;
   int rv = 0;
-  u64 key;
-
-  key = (((u64) mp->session_thread_index) << 32) | (u64) mp->session_index;
 
-  p = hash_get (utm->session_index_by_vpp_handles, key);
+  p = hash_get (utm->session_index_by_vpp_handles, mp->handle);
 
   if (p)
     {
       session = pool_elt_at_index (utm->sessions, p[0]);
-      hash_unset (utm->session_index_by_vpp_handles, key);
+      hash_unset (utm->session_index_by_vpp_handles, mp->handle);
       pool_put (utm->sessions, session);
     }
   else
     {
-      clib_warning ("couldn't find session key %llx", key);
+      clib_warning ("couldn't find session key %llx", mp->handle);
       rv = -11;
     }
 
@@ -592,77 +670,76 @@ vl_api_disconnect_session_t_handler (vl_api_disconnect_session_t * mp)
   memset (rmp, 0, sizeof (*rmp));
   rmp->_vl_msg_id = ntohs (VL_API_DISCONNECT_SESSION_REPLY);
   rmp->retval = rv;
-  rmp->session_index = mp->session_index;
-  rmp->session_thread_index = mp->session_thread_index;
+  rmp->handle = mp->handle;
   vl_msg_api_send_shmem (utm->vl_input_queue, (u8 *) & rmp);
 }
 
 static void
 vl_api_connect_uri_reply_t_handler (vl_api_connect_uri_reply_t * mp)
 {
-  svm_fifo_segment_main_t *sm = &svm_fifo_segment_main;
   uri_udp_test_main_t *utm = &uri_udp_test_main;
-  svm_fifo_segment_create_args_t _a, *a = &_a;
-  ssvm_shared_header_t *sh;
-  svm_fifo_segment_private_t *seg;
-  svm_fifo_segment_header_t *fsh;
-  session_t *session;
-  u32 segment_index;
-  int rv;
 
   ASSERT (utm->i_am_master == 0);
 
-  if (mp->segment_name_length == 0)
+  /* We've been redirected */
+  if (mp->segment_name_length > 0)
     {
-      clib_warning ("segment_name_length zero");
-      return;
-    }
-
-  memset (a, 0, sizeof (*a));
-
-  a->segment_name = (char *) mp->segment_name;
-
-  sleep (1);
-
-  rv = svm_fifo_segment_attach (a);
-  if (rv)
-    {
-      clib_warning ("sm_fifo_segment_create ('%v') failed", mp->segment_name);
-      return;
-    }
-
-  segment_index = vec_len (sm->segments) - 1;
+      svm_fifo_segment_main_t *sm = &svm_fifo_segment_main;
+      svm_fifo_segment_create_args_t _a, *a = &_a;
+      u32 segment_index;
+      session_t *session;
+      ssvm_shared_header_t *sh;
+      svm_fifo_segment_private_t *seg;
+      svm_fifo_segment_header_t *fsh;
+      int rv;
+
+      memset (a, 0, sizeof (*a));
+      a->segment_name = (char *) mp->segment_name;
+
+      sleep (1);
+
+      rv = svm_fifo_segment_attach (a);
+      if (rv)
+	{
+	  clib_warning ("sm_fifo_segment_create ('%v') failed",
+			mp->segment_name);
+	  return;
+	}
 
-  vec_add2 (utm->seg, seg, 1);
+      segment_index = vec_len (sm->segments) - 1;
+      vec_add2 (utm->seg, seg, 1);
 
-  memcpy (seg, sm->segments + segment_index, sizeof (*seg));
-  sh = seg->ssvm.sh;
-  fsh = (svm_fifo_segment_header_t *) sh->opaque[0];
+      memcpy (seg, sm->segments + segment_index, sizeof (*seg));
+      sh = seg->ssvm.sh;
+      fsh = (svm_fifo_segment_header_t *) sh->opaque[0];
 
-  while (vec_len (fsh->fifos) < 2)
-    sleep (1);
+      while (vec_len (fsh->fifos) < 2)
+	sleep (1);
 
-  pool_get (utm->sessions, session);
-  utm->cut_through_session_index = session - utm->sessions;
+      pool_get (utm->sessions, session);
+      utm->cut_through_session_index = session - utm->sessions;
 
-  session->server_rx_fifo = (svm_fifo_t *) fsh->fifos[0];
-  ASSERT (session->server_rx_fifo);
-  session->server_tx_fifo = (svm_fifo_t *) fsh->fifos[1];
-  ASSERT (session->server_tx_fifo);
+      session->server_rx_fifo = (svm_fifo_t *) fsh->fifos[0];
+      ASSERT (session->server_rx_fifo);
+      session->server_tx_fifo = (svm_fifo_t *) fsh->fifos[1];
+      ASSERT (session->server_tx_fifo);
+    }
 
   /* security: could unlink /dev/shm/<mp->segment_name> here, maybe */
 
   utm->state = STATE_READY;
 }
 
-#define foreach_uri_msg                         \
-_(BIND_URI_REPLY, bind_uri_reply)               \
-_(CONNECT_URI, connect_uri)                     \
-_(CONNECT_URI_REPLY, connect_uri_reply)         \
-_(UNBIND_URI_REPLY, unbind_uri_reply)           \
-_(ACCEPT_SESSION, accept_session)		\
-_(DISCONNECT_SESSION, disconnect_session)	\
-_(MAP_ANOTHER_SEGMENT, map_another_segment)
+#define foreach_uri_msg                         	\
+_(BIND_URI_REPLY, bind_uri_reply)               	\
+_(CONNECT_URI, connect_uri)                     	\
+_(CONNECT_URI_REPLY, connect_uri_reply)         	\
+_(UNBIND_URI_REPLY, unbind_uri_reply)           	\
+_(ACCEPT_SESSION, accept_session)			\
+_(DISCONNECT_SESSION, disconnect_session)		\
+_(MAP_ANOTHER_SEGMENT, map_another_segment)		\
+_(APPLICATION_ATTACH_REPLY, application_attach_reply)	\
+_(APPLICATION_DETACH_REPLY, application_detach_reply)	\
 
 void
 uri_api_hookup (uri_udp_test_main_t * utm)
@@ -679,7 +756,6 @@ uri_api_hookup (uri_udp_test_main_t * utm)
 
 }
 
-
 int
 connect_to_vpp (char *name)
 {
@@ -784,26 +860,43 @@ server_handle_event_queue (uri_udp_test_main_t * utm)
     }
 }
 
-void
-uri_udp_test (uri_udp_test_main_t * utm)
+static void
+server_unbind (uri_udp_test_main_t * utm)
 {
-  vl_api_bind_uri_t *bmp;
   vl_api_unbind_uri_t *ump;
 
+  ump = vl_msg_api_alloc (sizeof (*ump));
+  memset (ump, 0, sizeof (*ump));
+
+  ump->_vl_msg_id = ntohs (VL_API_UNBIND_URI);
+  ump->client_index = utm->my_client_index;
+  memcpy (ump->uri, utm->uri, vec_len (utm->uri));
+  vl_msg_api_send_shmem (utm->vl_input_queue, (u8 *) & ump);
+}
+
+static void
+server_listen (uri_udp_test_main_t * utm)
+{
+  vl_api_bind_uri_t *bmp;
+
   bmp = vl_msg_api_alloc (sizeof (*bmp));
   memset (bmp, 0, sizeof (*bmp));
 
   bmp->_vl_msg_id = ntohs (VL_API_BIND_URI);
   bmp->client_index = utm->my_client_index;
   bmp->context = ntohl (0xfeedface);
-  bmp->initial_segment_size = 256 << 20;	/* size of initial segment */
-  bmp->options[SESSION_OPTIONS_FLAGS] =
-    SESSION_OPTIONS_FLAGS_USE_FIFO | SESSION_OPTIONS_FLAGS_ADD_SEGMENT;
-  bmp->options[SESSION_OPTIONS_RX_FIFO_SIZE] = 16 << 10;
-  bmp->options[SESSION_OPTIONS_TX_FIFO_SIZE] = 16 << 10;
-  bmp->options[SESSION_OPTIONS_ADD_SEGMENT_SIZE] = 128 << 20;
   memcpy (bmp->uri, utm->uri, vec_len (utm->uri));
   vl_msg_api_send_shmem (utm->vl_input_queue, (u8 *) & bmp);
+}
+
+void
+udp_server_test (uri_udp_test_main_t * utm)
+{
+
+  application_attach (utm);
+
+  /* Bind to uri */
+  server_listen (utm);
 
   if (wait_for_state_change (utm, STATE_READY))
     {
@@ -813,13 +906,8 @@ uri_udp_test (uri_udp_test_main_t * utm)
 
   server_handle_event_queue (utm);
 
-  ump = vl_msg_api_alloc (sizeof (*ump));
-  memset (ump, 0, sizeof (*ump));
-
-  ump->_vl_msg_id = ntohs (VL_API_UNBIND_URI);
-  ump->client_index = utm->my_client_index;
-  memcpy (ump->uri, utm->uri, vec_len (utm->uri));
-  vl_msg_api_send_shmem (utm->vl_input_queue, (u8 *) & ump);
+  /* Cleanup */
+  server_unbind (utm);
 
   if (wait_for_state_change (utm, STATE_START))
     {
@@ -827,6 +915,8 @@ uri_udp_test (uri_udp_test_main_t * utm)
       return;
     }
 
+  application_detach (utm);
+
   fformat (stdout, "Test complete...\n");
 }
 
@@ -892,7 +982,7 @@ main (int argc, char **argv)
   utm->i_am_master = i_am_master;
   utm->segment_main = &svm_fifo_segment_main;
 
-  utm->connect_uri = format (0, "udp://10.0.0.1/1234%c", 0);
+  utm->connect_uri = format (0, "udp://6.0.0.1/1234%c", 0);
 
   setup_signal_handlers ();
 
@@ -907,7 +997,7 @@ main (int argc, char **argv)
 
   if (i_am_master == 0)
     {
-      uri_udp_slave_test (utm);
+      uri_udp_client_test (utm);
       exit (0);
     }
 
@@ -920,7 +1010,7 @@ main (int argc, char **argv)
   for (i = 0; i < 200000; i++)
     pool_put_index (utm->sessions, i);
 
-  uri_udp_test (utm);
+  udp_server_test (utm);
 
   vl_client_disconnect_from_vlib ();
   exit (0);
diff --git a/src/vnet.am b/src/vnet.am
index bed4902b..25b84616 100644
--- a/src/vnet.am
+++ b/src/vnet.am
@@ -827,6 +827,7 @@ libvnet_la_SOURCES +=				\
   vnet/session/session_cli.c			\
   vnet/session/hashes.c				\
   vnet/session/application_interface.c		\
+  vnet/session/segment_manager.c		\
   vnet/session/session_api.c
 
 nobase_include_HEADERS +=			\
@@ -835,6 +836,7 @@ nobase_include_HEADERS +=			\
   vnet/session/transport.h			\
   vnet/session/application_interface.h		\
   vnet/session/session_debug.h			\
+  vnet/session/segment_manager.h		\
   vnet/session/session.api.h
 
 API_FILES += vnet/session/session.api
diff --git a/src/vnet/api_errno.h b/src/vnet/api_errno.h
index f3ffd2a6..e939404b 100644
--- a/src/vnet/api_errno.h
+++ b/src/vnet/api_errno.h
@@ -105,7 +105,9 @@ _(INVALID_GPE_MODE, -112, "Invalid GPE mode")                           \
 _(LISP_GPE_ENTRIES_PRESENT, -113, "LISP GPE entries are present")       \
 _(ADDRESS_FOUND_FOR_INTERFACE, -114, "Address found for interface")	\
 _(SESSION_CONNECT_FAIL, -115, "Session failed to connect")              \
-_(ENTRY_ALREADY_EXISTS, -116, "Entry already exists")
+_(ENTRY_ALREADY_EXISTS, -116, "Entry already exists")			\
+_(SVM_SEGMENT_CREATE_FAIL, -117, "svm segment create fail")		\
+_(APPLICATION_NOT_ATTACHED, -118, "application not attached")
 
 typedef enum
 {
diff --git a/src/vnet/session/application.c b/src/vnet/session/application.c
index 513e5fac..5a45537b 100644
--- a/src/vnet/session/application.c
+++ b/src/vnet/session/application.c
@@ -14,18 +14,24 @@
  */
 
 #include <vnet/session/application.h>
+#include <vnet/session/application_interface.h>
 #include <vnet/session/session.h>
 
-/*
+/**
  * Pool from which we allocate all applications
  */
 static application_t *app_pool;
 
-/*
+/**
  * Hash table of apps by api client index
  */
 static uword *app_by_api_client_index;
 
+/**
+ * Default application event queue size
+ */
+static u32 default_app_evt_queue_size = 128;
+
 int
 application_api_queue_is_full (application_t * app)
 {
@@ -67,37 +73,71 @@ application_lookup (u32 api_client_index)
   return 0;
 }
 
+application_t *
+application_new ()
+{
+  application_t *app;
+  pool_get (app_pool, app);
+  memset (app, 0, sizeof (*app));
+  app->index = application_get_index (app);
+  app->connects_seg_manager = ~0;
+  return app;
+}
+
 void
 application_del (application_t * app)
 {
-  session_manager_main_t *smm = vnet_get_session_manager_main ();
   api_main_t *am = &api_main;
   void *oldheap;
-  session_manager_t *sm;
+  segment_manager_t *sm;
+  u64 handle;
+  u32 index, *handles = 0;
+  int i;
+  vnet_unbind_args_t _a, *a = &_a;
+
+  /*
+   * Cleanup segment managers
+   */
+  if (app->connects_seg_manager != (u32) ~ 0)
+    {
+      sm = segment_manager_get (app->connects_seg_manager);
+      segment_manager_del (sm);
+    }
 
-  if (app->mode == APP_SERVER)
+  /* *INDENT-OFF* */
+  hash_foreach (handle, index, app->listeners_table,
+  ({
+    vec_add1 (handles, handle);
+  }));
+  /* *INDENT-ON* */
+
+  /* Actual listener cleanup */
+  for (i = 0; i < vec_len (handles); i++)
     {
-      sm = session_manager_get (app->session_manager_index);
-      session_manager_del (smm, sm);
+      a->app_index = app->api_client_index;
+      a->handle = handles[i];
+      /* seg manager is removed when unbind completes */
+      vnet_unbind (a);
     }
 
-  /* Free the event fifo in the /vpe-api shared-memory segment */
+  /*
+   * Free the event fifo in the /vpe-api shared-memory segment
+   */
   oldheap = svm_push_data_heap (am->vlib_rp);
   if (app->event_queue)
     unix_shared_memory_queue_free (app->event_queue);
   svm_pop_heap (oldheap);
 
   application_table_del (app);
-
   pool_put (app_pool, app);
 }
 
 static void
-application_verify_cb_fns (application_type_t type, session_cb_vft_t * cb_fns)
+application_verify_cb_fns (session_cb_vft_t * cb_fns)
 {
-  if (type == APP_SERVER && cb_fns->session_accept_callback == 0)
+  if (cb_fns->session_accept_callback == 0)
     clib_warning ("No accept callback function provided");
-  if (type == APP_CLIENT && cb_fns->session_connected_callback == 0)
+  if (cb_fns->session_connected_callback == 0)
     clib_warning ("No session connected callback function provided");
   if (cb_fns->session_disconnect_callback == 0)
     clib_warning ("No session disconnect callback function provided");
@@ -105,25 +145,26 @@ application_verify_cb_fns (application_type_t type, session_cb_vft_t * cb_fns)
     clib_warning ("No session reset callback function provided");
 }
 
-application_t *
-application_new (application_type_t type, session_type_t sst,
-		 u32 api_client_index, u32 flags, session_cb_vft_t * cb_fns)
+int
+application_init (application_t * app, u32 api_client_index, u64 * options,
+		  session_cb_vft_t * cb_fns)
 {
-  session_manager_main_t *smm = vnet_get_session_manager_main ();
   api_main_t *am = &api_main;
-  application_t *app;
+  segment_manager_t *sm;
+  segment_manager_properties_t *props;
   void *oldheap;
-  session_manager_t *sm;
+  u32 app_evt_queue_size;
+  int rv;
 
-  pool_get (app_pool, app);
-  memset (app, 0, sizeof (*app));
+  app_evt_queue_size = options[APP_EVT_QUEUE_SIZE] > 0 ?
+    options[APP_EVT_QUEUE_SIZE] : default_app_evt_queue_size;
 
   /* Allocate event fifo in the /vpe-api shared-memory segment */
   oldheap = svm_push_data_heap (am->vlib_rp);
 
   /* Allocate server event queue */
   app->event_queue =
-    unix_shared_memory_queue_init (128 /* nels $$$$ config */ ,
+    unix_shared_memory_queue_init (app_evt_queue_size,
 				   sizeof (session_fifo_event_t),
 				   0 /* consumer pid */ ,
 				   0
@@ -132,36 +173,31 @@ application_new (application_type_t type, session_type_t sst,
 
   svm_pop_heap (oldheap);
 
-  /* If a server, allocate session manager */
-  if (type == APP_SERVER)
-    {
-      pool_get (smm->session_managers, sm);
-      memset (sm, 0, sizeof (*sm));
+  /* Setup segment manager */
+  sm = segment_manager_new ();
+  sm->app_index = app->index;
+  props = &app->sm_properties;
+  props->add_segment_size = options[SESSION_OPTIONS_ADD_SEGMENT_SIZE];
+  props->rx_fifo_size = options[SESSION_OPTIONS_RX_FIFO_SIZE];
+  props->tx_fifo_size = options[SESSION_OPTIONS_TX_FIFO_SIZE];
+  props->add_segment = props->add_segment_size != 0;
 
-      app->session_manager_index = sm - smm->session_managers;
-    }
-  else if (type == APP_CLIENT)
-    {
-      /* Allocate connect session manager if needed */
-      if (smm->connect_manager_index[sst] == INVALID_INDEX)
-	connects_session_manager_init (smm, sst);
-      app->session_manager_index = smm->connect_manager_index[sst];
-    }
+  if ((rv = segment_manager_init (sm, props,
+				  options[SESSION_OPTIONS_SEGMENT_SIZE])))
+    return rv;
 
-  app->mode = type;
-  app->index = application_get_index (app);
-  app->session_type = sst;
+  app->first_segment_manager = segment_manager_index (sm);
   app->api_client_index = api_client_index;
-  app->flags = flags;
+  app->flags = options[SESSION_OPTIONS_FLAGS];
   app->cb_fns = *cb_fns;
 
   /* Check that the obvious things are properly set up */
-  application_verify_cb_fns (type, cb_fns);
+  application_verify_cb_fns (cb_fns);
 
   /* Add app to lookup by api_client_index table */
   application_table_add (app);
 
-  return app;
+  return 0;
 }
 
 application_t *
@@ -185,108 +221,286 @@ application_get_index (application_t * app)
   return app - app_pool;
 }
 
+static segment_manager_t *
+application_alloc_segment_manager (application_t * app)
+{
+  segment_manager_t *sm = 0;
+
+  if (app->first_segment_manager != (u32) ~ 0)
+    {
+      sm = segment_manager_get (app->first_segment_manager);
+      app->first_segment_manager = ~0;
+      return sm;
+    }
+
+  sm = segment_manager_new ();
+  if (segment_manager_init (sm, &app->sm_properties, 0))
+    return 0;
+  return sm;
+}
+
+/**
+ * Start listening local transport endpoint for requested transport.
+ *
+ * Creates a 'dummy' stream session with state LISTENING to be used in session
+ * lookups, prior to establishing connection. Requests transport to build
+ * it's own specific listening connection.
+ */
 int
-application_server_init (application_t * server, u32 segment_size,
-			 u32 add_segment_size, u32 rx_fifo_size,
-			 u32 tx_fifo_size, u8 ** segment_name)
+application_start_listen (application_t * srv, session_type_t session_type,
+			  transport_endpoint_t * tep, u64 * res)
 {
-  session_manager_main_t *smm = vnet_get_session_manager_main ();
-  session_manager_t *sm;
-  int rv;
+  segment_manager_t *sm;
+  stream_session_t *s;
+  u64 handle;
+
+  s = listen_session_new (session_type);
+  s->app_index = srv->index;
+
+  if (stream_session_listen (s, tep))
+    goto err;
+
+  /* Allocate segment manager. All sessions derived out of a listen session
+   * have fifos allocated by the same segment manager. */
+  sm = application_alloc_segment_manager (srv);
+  if (sm == 0)
+    goto err;
+
+  /* Add to app's listener table. Useful to find all child listeners
+   * when app goes down, although, just for unbinding this is not needed */
+  handle = listen_session_get_handle (s);
+  hash_set (srv->listeners_table, handle, segment_manager_index (sm));
 
-  sm = session_manager_get (server->session_manager_index);
+  *res = handle;
+  return 0;
+
+err:
+  listen_session_del (s);
+  return -1;
+}
+
+/**
+ * Stop listening on session associated to handle
+ */
+int
+application_stop_listen (application_t * srv, u64 handle)
+{
+  stream_session_t *listener;
+  uword *indexp;
+  segment_manager_t *sm;
 
-  /* Add first segment */
-  if ((rv = session_manager_add_first_segment (smm, sm, segment_size,
-					       segment_name)))
+  if (srv && hash_get (srv->listeners_table, handle) == 0)
     {
-      return rv;
+      clib_warning ("app doesn't own handle %llu!", handle);
+      return -1;
     }
 
-  /* Setup session manager */
-  sm->add_segment_size = add_segment_size;
-  sm->rx_fifo_size = rx_fifo_size;
-  sm->tx_fifo_size = tx_fifo_size;
-  sm->add_segment = sm->add_segment_size != 0;
+  listener = listen_session_get_from_handle (handle);
+  stream_session_stop_listen (listener);
+
+  indexp = hash_get (srv->listeners_table, handle);
+  ASSERT (indexp);
+
+  sm = segment_manager_get (*indexp);
+  segment_manager_del (sm);
+  hash_unset (srv->listeners_table, handle);
+  listen_session_del (listener);
+
   return 0;
 }
 
+int
+application_open_session (application_t * app, session_type_t sst,
+			  transport_endpoint_t * tep, u32 api_context)
+{
+  segment_manager_t *sm;
+  transport_connection_t *tc = 0;
+  int rv;
+
+  /* Make sure we have a segment manager for connects */
+  if (app->connects_seg_manager == (u32) ~ 0)
+    {
+      sm = application_alloc_segment_manager (app);
+      if (sm == 0)
+	return -1;
+      app->connects_seg_manager = segment_manager_index (sm);
+    }
+
+  if ((rv = stream_session_open (app->index, sst, tep, &tc)))
+    return rv;
+
+  /* Store api_context for when the reply comes. Not the nicest thing
+   * but better allocating a separate half-open pool.  */
+  tc->s_index = api_context;
+
+  return 0;
+}
+
+segment_manager_t *
+application_get_connect_segment_manager (application_t * app)
+{
+  ASSERT (app->connects_seg_manager != (u32) ~ 0);
+  return segment_manager_get (app->connects_seg_manager);
+}
+
+segment_manager_t *
+application_get_listen_segment_manager (application_t * app,
+					stream_session_t * s)
+{
+  uword *smp;
+  smp = hash_get (app->listeners_table, listen_session_get_handle (s));
+  ASSERT (smp != 0);
+  return segment_manager_get (*smp);
+}
+
+static u8 *
+app_get_name_from_reg_index (application_t * app)
+{
+  u8 *app_name;
+
+  vl_api_registration_t *regp;
+  regp = vl_api_client_index_to_registration (app->api_client_index);
+  if (!regp)
+    app_name = format (0, "builtin-%d%c", app->index, 0);
+  else
+    app_name = format (0, "%s%c", regp->name, 0);
+
+  return app_name;
+}
+
 u8 *
-format_application_server (u8 * s, va_list * args)
+format_application_listener (u8 * s, va_list * args)
 {
-  application_t *srv = va_arg (*args, application_t *);
+  application_t *app = va_arg (*args, application_t *);
+  u64 handle = va_arg (*args, u64);
+  u32 index = va_arg (*args, u32);
   int verbose = va_arg (*args, int);
-  vl_api_registration_t *regp;
   stream_session_t *listener;
-  u8 *server_name, *str, *seg_name;
-  u32 segment_size;
+  u8 *app_name, *str;
 
-  if (srv == 0)
+  if (app == 0)
     {
       if (verbose)
-	s = format (s, "%-40s%-20s%-15s%-15s%-10s", "Connection", "Server",
-		    "Segment", "API Client", "Cookie");
+	s = format (s, "%-40s%-20s%-15s%-15s%-10s", "Connection", "App",
+		    "API Client", "ListenerID", "SegManager");
       else
-	s = format (s, "%-40s%-20s", "Connection", "Server");
+	s = format (s, "%-40s%-20s", "Connection", "App");
 
       return s;
     }
 
-  regp = vl_api_client_index_to_registration (srv->api_client_index);
-  if (!regp)
-    server_name = format (0, "builtin-%d%c", srv->index, 0);
-  else
-    server_name = regp->name;
-
-  listener = stream_session_listener_get (srv->session_type,
-					  srv->session_index);
+  app_name = app_get_name_from_reg_index (app);
+  listener = listen_session_get_from_handle (handle);
   str = format (0, "%U", format_stream_session, listener, verbose);
 
-  session_manager_get_segment_info (listener->server_segment_index, &seg_name,
-				    &segment_size);
   if (verbose)
     {
-      s = format (s, "%-40s%-20s%-20s%-10d%-10d", str, server_name,
-		  seg_name, srv->api_client_index, srv->accept_cookie);
+      s = format (s, "%-40s%-20s%-15u%-15u%-10u", str, app_name,
+		  app->api_client_index, handle, index);
     }
   else
-    s = format (s, "%-40s%-20s", str, server_name);
+    s = format (s, "%-40s%-20s", str, app_name);
+
+  vec_free (app_name);
   return s;
 }
 
-u8 *
-format_application_client (u8 * s, va_list * args)
+void
+application_format_connects (application_t * app, int verbose)
 {
-  application_t *client = va_arg (*args, application_t *);
-  int verbose = va_arg (*args, int);
-  stream_session_t *session;
-  u8 *str, *seg_name;
-  u32 segment_size;
+  vlib_main_t *vm = vlib_get_main ();
+  segment_manager_t *sm;
+  u8 *app_name, *s = 0;
+  int i, j;
 
-  if (client == 0)
+  /* Header */
+  if (app == 0)
     {
       if (verbose)
-	s =
-	  format (s, "%-40s%-20s%-10s", "Connection", "Segment",
-		  "API Client");
+	vlib_cli_output (vm, "%-40s%-20s%-15s%-10s", "Connection", "App",
+			 "API Client", "SegManager");
       else
-	s = format (s, "%-40s", "Connection");
+	vlib_cli_output (vm, "%-40s%-20s", "Connection", "App");
+      return;
+    }
 
-      return s;
+  /* make sure */
+  if (app->connects_seg_manager == (u32) ~ 0)
+    return;
+
+  app_name = app_get_name_from_reg_index (app);
+
+  /* Across all fifo segments */
+  sm = segment_manager_get (app->connects_seg_manager);
+  for (j = 0; j < vec_len (sm->segment_indices); j++)
+    {
+      svm_fifo_segment_private_t *fifo_segment;
+      svm_fifo_t **fifos;
+      u8 *str;
+
+      fifo_segment = svm_fifo_get_segment (sm->segment_indices[j]);
+      fifos = svm_fifo_segment_get_fifos (fifo_segment);
+      for (i = 0; i < vec_len (fifos); i++)
+	{
+	  svm_fifo_t *fifo;
+	  u32 session_index, thread_index;
+	  stream_session_t *session;
+
+	  /* There are 2 fifos/session. Avoid printing twice. */
+	  if (i % 2)
+	    continue;
+
+	  fifo = fifos[i];
+	  session_index = fifo->server_session_index;
+	  thread_index = fifo->server_thread_index;
+
+	  session = stream_session_get (session_index, thread_index);
+	  str = format (0, "%U", format_stream_session, session, verbose);
+
+	  if (verbose)
+	    s = format (s, "%-40s%-20s%-15u%-10u", str, app_name,
+			app->api_client_index, app->connects_seg_manager);
+	  else
+	    s = format (s, "%-40s%-20s", str, app_name);
+
+	  vlib_cli_output (vm, "%v", s);
+
+	  vec_reset_length (s);
+	  vec_free (str);
+	}
+      vec_free (s);
     }
 
-  session = stream_session_get (client->session_index, client->thread_index);
-  str = format (0, "%U", format_stream_session, session, verbose);
+  vec_free (app_name);
+}
 
-  session_manager_get_segment_info (session->server_segment_index, &seg_name,
-				    &segment_size);
-  if (verbose)
+u8 *
+format_application (u8 * s, va_list * args)
+{
+  application_t *app = va_arg (*args, application_t *);
+  CLIB_UNUSED (int verbose) = va_arg (*args, int);
+  u8 *app_name;
+
+  if (app == 0)
     {
-      s = format (s, "%-40s%-20s%-10d%", str, seg_name,
-		  client->api_client_index);
+      if (verbose)
+	s = format (s, "%-10s%-20s%-15s%-15s%-15s%-15s", "Index", "Name",
+		    "API Client", "Add seg size", "Rx fifo size",
+		    "Tx fifo size");
+      else
+	s = format (s, "%-10s%-20s%-20s", "Index", "Name", "API Client");
+      return s;
     }
+
+  app_name = app_get_name_from_reg_index (app);
+  if (verbose)
+    s = format (s, "%-10d%-20s%-15d%-15d%-15d%-15d", app->index, app_name,
+		app->api_client_index, app->sm_properties.add_segment_size,
+		app->sm_properties.rx_fifo_size,
+		app->sm_properties.tx_fifo_size);
   else
-    s = format (s, "%-40s", str);
+    s = format (s, "%-10d%-20s%-20d", app->index, app_name,
+		app->api_client_index);
   return s;
 }
 
@@ -294,13 +508,12 @@ static clib_error_t *
 show_app_command_fn (vlib_main_t * vm, unformat_input_t * input,
 		     vlib_cli_command_t * cmd)
 {
-  session_manager_main_t *smm = &session_manager_main;
   application_t *app;
   int do_server = 0;
   int do_client = 0;
   int verbose = 0;
 
-  if (!smm->is_enabled)
+  if (!session_manager_is_enabled ())
     {
       clib_error_return (0, "session layer is not enabled");
     }
@@ -319,17 +532,24 @@ show_app_command_fn (vlib_main_t * vm, unformat_input_t * input,
 
   if (do_server)
     {
+      u64 handle;
+      u32 index;
       if (pool_elts (app_pool))
 	{
-	  vlib_cli_output (vm, "%U", format_application_server,
-			   0 /* header */ ,
+	  vlib_cli_output (vm, "%U", format_application_listener,
+			   0 /* header */ , 0, 0,
 			   verbose);
           /* *INDENT-OFF* */
           pool_foreach (app, app_pool,
           ({
-            if (app->mode == APP_SERVER)
-              vlib_cli_output (vm, "%U", format_application_server, app,
-                               verbose);
+            /* App's listener sessions */
+            if (hash_elts (app->listeners_table) == 0)
+              continue;
+            hash_foreach (handle, index, app->listeners_table,
+	    ({
+              vlib_cli_output (vm, "%U", format_application_listener, app,
+              			       handle, index, verbose);
+            }));
           }));
           /* *INDENT-ON* */
 	}
@@ -341,15 +561,14 @@ show_app_command_fn (vlib_main_t * vm, unformat_input_t * input,
     {
       if (pool_elts (app_pool))
 	{
-	  vlib_cli_output (vm, "%U", format_application_client,
-			   0 /* header */ ,
-			   verbose);
+	  application_format_connects (0, verbose);
+
           /* *INDENT-OFF* */
           pool_foreach (app, app_pool,
           ({
-            if (app->mode == APP_CLIENT)
-              vlib_cli_output (vm, "%U", format_application_client, app,
-                               verbose);
+            if (app->connects_seg_manager == (u32)~0)
+              continue;
+            application_format_connects (app, verbose);
           }));
           /* *INDENT-ON* */
 	}
@@ -357,6 +576,19 @@ show_app_command_fn (vlib_main_t * vm, unformat_input_t * input,
 	vlib_cli_output (vm, "No active client bindings");
     }
 
+  /* Print app related info */
+  if (!do_server && !do_client)
+    {
+      vlib_cli_output (vm, "%U", format_application, 0, verbose);
+      pool_foreach (app, app_pool, (
+				     {
+				     vlib_cli_output (vm, "%U",
+						      format_application, app,
+						      verbose);
+				     }
+		    ));
+    }
+
   return 0;
 }
 
diff --git a/src/vnet/session/application.h b/src/vnet/session/application.h
index 480828f7..6bcee9d3 100644
--- a/src/vnet/session/application.h
+++ b/src/vnet/session/application.h
@@ -18,11 +18,13 @@
 
 #include <vnet/vnet.h>
 #include <vnet/session/session.h>
+#include <vnet/session/segment_manager.h>
 
 typedef enum
 {
   APP_SERVER,
-  APP_CLIENT
+  APP_CLIENT,
+  APP_N_TYPES
 } application_type_t;
 
 typedef struct _stream_session_cb_vft
@@ -35,7 +37,7 @@ typedef struct _stream_session_cb_vft
   int (*session_accept_callback) (stream_session_t * new_session);
 
   /* Connection request callback */
-  int (*session_connected_callback) (u32 api_client_index,
+  int (*session_connected_callback) (u32 app_index, u32 api_context,
 				     stream_session_t * s, u8 code);
 
   /** Notify app that session is closing */
@@ -59,45 +61,52 @@ typedef struct _application
   /** Flags */
   u32 flags;
 
+  /* Stream server mode: accept or connect
+   * TODO REMOVE*/
+  u8 mode;
+
+  /** Index of the listen session or connect session
+   * TODO REMOVE*/
+  u32 session_index;
+
+  /** Session thread index for client connect sessions
+   * TODO REMOVE */
+  u32 thread_index;
+
+  /*
+   * Binary API interface to external app
+   */
+
   /** Binary API connection index, ~0 if internal */
   u32 api_client_index;
 
-  /* */
-  u32 api_context;
-
   /** Application listens for events on this svm queue */
   unix_shared_memory_queue_t *event_queue;
 
-  /** Stream session type */
-  u8 session_type;
-
-  /* Stream server mode: accept or connect */
-  u8 mode;
+  /*
+   * Callbacks: shoulder-taps for the server/client
+   */
 
-  u32 session_manager_index;
+  session_cb_vft_t cb_fns;
 
   /*
-   * Bind/Listen specific
+   * svm segment management
    */
+  u32 connects_seg_manager;
 
-  /** Accept cookie, for multiple session flavors ($$$ maybe) */
-  u32 accept_cookie;
+  /* Lookup tables for listeners. Value is segment manager index */
+  uword *listeners_table;
 
-  /** Index of the listen session or connect session */
-  u32 session_index;
+  u32 first_segment_manager;
 
-  /** Session thread index for client connect sessions */
-  u32 thread_index;
-
-  /*
-   * Callbacks: shoulder-taps for the server/client
-   */
-  session_cb_vft_t cb_fns;
+  /** Segment manager properties. Shared by all segment managers */
+  segment_manager_properties_t sm_properties;
 } application_t;
 
-application_t *application_new (application_type_t type, session_type_t sst,
-				u32 api_client_index, u32 flags,
-				session_cb_vft_t * cb_fns);
+application_t *application_new ();
+int
+application_init (application_t * app, u32 api_client_index, u64 * options,
+		  session_cb_vft_t * cb_fns);
 void application_del (application_t * app);
 application_t *application_get (u32 index);
 application_t *application_get_if_valid (u32 index);
@@ -105,11 +114,21 @@ application_t *application_lookup (u32 api_client_index);
 u32 application_get_index (application_t * app);
 
 int
-application_server_init (application_t * server, u32 segment_size,
-			 u32 add_segment_size, u32 rx_fifo_size,
-			 u32 tx_fifo_size, u8 ** segment_name);
+application_start_listen (application_t * app, session_type_t session_type,
+			  transport_endpoint_t * tep, u64 * handle);
+int application_stop_listen (application_t * srv, u64 handle);
+int
+application_open_session (application_t * app, session_type_t sst,
+			  transport_endpoint_t * tep, u32 api_context);
 int application_api_queue_is_full (application_t * app);
 
+segment_manager_t *application_get_listen_segment_manager (application_t *
+							   app,
+							   stream_session_t *
+							   s);
+segment_manager_t *application_get_connect_segment_manager (application_t *
+							    app);
+
 #endif /* SRC_VNET_SESSION_APPLICATION_H_ */
 
 /*
diff --git a/src/vnet/session/application_interface.c b/src/vnet/session/application_interface.c
index 4b30bd87..96d2c621 100644
--- a/src/vnet/session/application_interface.c
+++ b/src/vnet/session/application_interface.c
@@ -79,81 +79,51 @@ api_parse_session_handle (u64 handle, u32 * session_index, u32 * thread_index)
 }
 
 int
-vnet_bind_i (u32 api_client_index, ip46_address_t * ip46, u16 port_host_order,
-	     session_type_t sst, u64 * options, session_cb_vft_t * cb_fns,
-	     application_t ** app, u32 * len_seg_name, char *seg_name)
+vnet_bind_i (u32 app_index, session_type_t sst,
+	     transport_endpoint_t * tep, u64 * handle)
 {
-  u8 *segment_name = 0;
-  application_t *server = 0;
+  application_t *app;
   stream_session_t *listener;
-  u8 is_ip4;
-
-  listener =
-    stream_session_lookup_listener (ip46,
-				    clib_host_to_net_u16 (port_host_order),
-				    sst);
-
-  if (listener)
-    return VNET_API_ERROR_ADDRESS_IN_USE;
 
-  if (application_lookup (api_client_index))
+  app = application_get_if_valid (app_index);
+  if (!app)
     {
-      clib_warning ("Only one connection supported for now");
-      return VNET_API_ERROR_ADDRESS_IN_USE;
+      clib_warning ("app not attached");
+      return VNET_API_ERROR_APPLICATION_NOT_ATTACHED;
     }
 
-  is_ip4 = SESSION_TYPE_IP4_UDP == sst || SESSION_TYPE_IP4_TCP == sst;
-  if (!ip_is_zero (ip46, is_ip4) && !ip_is_local (ip46, is_ip4))
-    return VNET_API_ERROR_INVALID_VALUE;
-
-  /* Allocate and initialize stream server */
-  server = application_new (APP_SERVER, sst, api_client_index,
-			    options[SESSION_OPTIONS_FLAGS], cb_fns);
+  listener = stream_session_lookup_listener (&tep->ip,
+					     clib_host_to_net_u16 (tep->port),
+					     sst);
+  if (listener)
+    return VNET_API_ERROR_ADDRESS_IN_USE;
 
-  application_server_init (server, options[SESSION_OPTIONS_SEGMENT_SIZE],
-			   options[SESSION_OPTIONS_ADD_SEGMENT_SIZE],
-			   options[SESSION_OPTIONS_RX_FIFO_SIZE],
-			   options[SESSION_OPTIONS_TX_FIFO_SIZE],
-			   &segment_name);
+  if (!ip_is_zero (&tep->ip, tep->is_ip4)
+      && !ip_is_local (&tep->ip, tep->is_ip4))
+    return VNET_API_ERROR_INVALID_VALUE_2;
 
   /* Setup listen path down to transport */
-  stream_session_start_listen (server->index, ip46, port_host_order);
-
-  /*
-   * Return values
-   */
-
-  ASSERT (vec_len (segment_name) <= 128);
-  *len_seg_name = vec_len (segment_name);
-  memcpy (seg_name, segment_name, *len_seg_name);
-  *app = server;
-
-  return 0;
+  return application_start_listen (app, sst, tep, handle);
 }
 
 int
-vnet_unbind_i (u32 api_client_index)
+vnet_unbind_i (u32 app_index, u64 handle)
 {
-  application_t *server;
+  application_t *app = application_get_if_valid (app_index);
 
-  /*
-   * Find the stream_server_t corresponding to the api client
-   */
-  server = application_lookup (api_client_index);
-  if (!server)
-    return VNET_API_ERROR_INVALID_VALUE_2;
+  if (!app)
+    {
+      clib_warning ("app not attached");
+      return VNET_API_ERROR_APPLICATION_NOT_ATTACHED;
+    }
 
   /* Clear the listener */
-  stream_session_stop_listen (server->index);
-  application_del (server);
-
-  return 0;
+  return application_stop_listen (app, handle);
 }
 
 int
-vnet_connect_i (u32 api_client_index, u32 api_context, session_type_t sst,
-		ip46_address_t * ip46, u16 port, u64 * options, void *mp,
-		session_cb_vft_t * cb_fns)
+vnet_connect_i (u32 app_index, u32 api_context, session_type_t sst,
+		transport_endpoint_t * tep, void *mp)
 {
   stream_session_t *listener;
   application_t *server, *app;
@@ -161,8 +131,8 @@ vnet_connect_i (u32 api_client_index, u32 api_context, session_type_t sst,
   /*
    * Figure out if connecting to a local server
    */
-  listener = stream_session_lookup_listener (ip46,
-					     clib_host_to_net_u16 (port),
+  listener = stream_session_lookup_listener (&tep->ip,
+					     clib_host_to_net_u16 (tep->port),
 					     sst);
   if (listener)
     {
@@ -177,16 +147,11 @@ vnet_connect_i (u32 api_client_index, u32 api_context, session_type_t sst,
 	  redirect_connect_callback (server->api_client_index, mp);
     }
 
-  /* Create client app */
-  app = application_new (APP_CLIENT, sst, api_client_index,
-			 options[SESSION_OPTIONS_FLAGS], cb_fns);
-
-  app->api_context = api_context;
-
   /*
    * Not connecting to a local server. Create regular session
    */
-  return stream_session_open (sst, ip46, port, app->index);
+  app = application_get (app_index);
+  return application_open_session (app, sst, tep, api_context);
 }
 
 /**
@@ -209,30 +174,31 @@ vnet_connect_i (u32 api_client_index, u32 api_context, session_type_t sst,
 uword
 unformat_vnet_uri (unformat_input_t * input, va_list * args)
 {
-  ip46_address_t *address = va_arg (*args, ip46_address_t *);
   session_type_t *sst = va_arg (*args, session_type_t *);
-  u16 *port = va_arg (*args, u16 *);
+  transport_endpoint_t *tep = va_arg (*args, transport_endpoint_t *);
 
-  if (unformat (input, "tcp://%U/%d", unformat_ip4_address, &address->ip4,
-		port))
+  if (unformat (input, "tcp://%U/%d", unformat_ip4_address, &tep->ip.ip4,
+		&tep->port))
     {
       *sst = SESSION_TYPE_IP4_TCP;
+      tep->is_ip4 = 1;
       return 1;
     }
-  if (unformat (input, "udp://%U/%d", unformat_ip4_address, &address->ip4,
-		port))
+  if (unformat (input, "udp://%U/%d", unformat_ip4_address, &tep->ip.ip4,
+		&tep->port))
     {
       *sst = SESSION_TYPE_IP4_UDP;
+      tep->is_ip4 = 1;
       return 1;
     }
-  if (unformat (input, "udp://%U/%d", unformat_ip6_address, &address->ip6,
-		port))
+  if (unformat (input, "udp://%U/%d", unformat_ip6_address, &tep->ip.ip6,
+		&tep->port))
     {
       *sst = SESSION_TYPE_IP6_UDP;
       return 1;
     }
-  if (unformat (input, "tcp://%U/%d", unformat_ip6_address, &address->ip6,
-		port))
+  if (unformat (input, "tcp://%U/%d", unformat_ip6_address, &tep->ip.ip6,
+		&tep->port))
     {
       *sst = SESSION_TYPE_IP6_TCP;
       return 1;
@@ -242,8 +208,7 @@ unformat_vnet_uri (unformat_input_t * input, va_list * args)
 }
 
 int
-parse_uri (char *uri, session_type_t * sst, ip46_address_t * addr,
-	   u16 * port_number_host_byte_order)
+parse_uri (char *uri, session_type_t * sst, transport_endpoint_t * tep)
 {
   unformat_input_t _input, *input = &_input;
 
@@ -252,8 +217,7 @@ parse_uri (char *uri, session_type_t * sst, ip46_address_t * addr,
 
   /* Parse uri */
   unformat_init_string (input, uri, strlen (uri));
-  if (!unformat (input, "%U", unformat_vnet_uri, addr, sst,
-		 port_number_host_byte_order))
+  if (!unformat (input, "%U", unformat_vnet_uri, sst, tep))
     {
       unformat_free (input);
       return VNET_API_ERROR_INVALID_VALUE;
@@ -263,26 +227,51 @@ parse_uri (char *uri, session_type_t * sst, ip46_address_t * addr,
   return 0;
 }
 
+/**
+ * Attaches application.
+ *
+ * Allocates a vpp app, i.e., a structure that keeps back pointers
+ * to external app and a segment manager for shared memory fifo based
+ * communication with the external app.
+ */
 int
-vnet_bind_uri (vnet_bind_args_t * a)
+vnet_application_attach (vnet_app_attach_args_t * a)
 {
-  application_t *server = 0;
-  u16 port_host_order;
-  session_type_t sst = SESSION_N_TYPES;
-  ip46_address_t ip46;
+  application_t *app = 0;
+  segment_manager_t *sm;
+  u8 *seg_name;
   int rv;
 
-  memset (&ip46, 0, sizeof (ip46));
-  rv = parse_uri (a->uri, &sst, &ip46, &port_host_order);
-  if (rv)
+  app = application_new ();
+  if ((rv = application_init (app, a->api_client_index, a->options,
+			      a->session_cb_vft)))
     return rv;
 
-  if ((rv = vnet_bind_i (a->api_client_index, &ip46, port_host_order, sst,
-			 a->options, a->session_cb_vft, &server,
-			 &a->segment_name_length, a->segment_name)))
-    return rv;
+  a->app_event_queue_address = (u64) app->event_queue;
+  sm = segment_manager_get (app->first_segment_manager);
+  segment_manager_get_segment_info (sm->segment_indices[0],
+				    &seg_name, &a->segment_size);
 
-  a->server_event_queue_address = (u64) server->event_queue;
+  a->segment_name_length = vec_len (seg_name);
+  a->segment_name = seg_name;
+  ASSERT (vec_len (a->segment_name) <= 128);
+  a->app_index = app->index;
+  return 0;
+}
+
+int
+vnet_application_detach (vnet_app_detach_args_t * a)
+{
+  application_t *app;
+  app = application_get_if_valid (a->app_index);
+
+  if (!app)
+    {
+      clib_warning ("app not attached");
+      return VNET_API_ERROR_APPLICATION_NOT_ATTACHED;
+    }
+
+  application_del (app);
   return 0;
 }
 
@@ -308,125 +297,102 @@ session_type_from_proto_and_ip (session_api_proto_t proto, u8 is_ip4)
 }
 
 int
-vnet_unbind_uri (char *uri, u32 api_client_index)
+vnet_bind_uri (vnet_bind_args_t * a)
 {
-  u16 port_number_host_byte_order;
   session_type_t sst = SESSION_N_TYPES;
-  ip46_address_t ip46_address;
-  stream_session_t *listener;
+  transport_endpoint_t tep;
   int rv;
 
-  rv = parse_uri (uri, &sst, &ip46_address, &port_number_host_byte_order);
+  memset (&tep, 0, sizeof (tep));
+  rv = parse_uri (a->uri, &sst, &tep);
   if (rv)
     return rv;
 
-  listener =
-    stream_session_lookup_listener (&ip46_address,
-				    clib_host_to_net_u16
-				    (port_number_host_byte_order), sst);
+  if ((rv = vnet_bind_i (a->app_index, sst, &tep, &a->handle)))
+    return rv;
+
+  return 0;
+}
+
+int
+vnet_unbind_uri (vnet_unbind_args_t * a)
+{
+  session_type_t sst = SESSION_N_TYPES;
+  stream_session_t *listener;
+  transport_endpoint_t tep;
+  int rv;
+
+  rv = parse_uri (a->uri, &sst, &tep);
+  if (rv)
+    return rv;
 
+  listener = stream_session_lookup_listener (&tep.ip,
+					     clib_host_to_net_u16 (tep.port),
+					     sst);
   if (!listener)
     return VNET_API_ERROR_ADDRESS_NOT_IN_USE;
 
-  /* External client? */
-  if (api_client_index != ~0)
-    {
-      ASSERT (vl_api_client_index_to_registration (api_client_index));
-    }
-
-  return vnet_unbind_i (api_client_index);
+  return vnet_unbind_i (a->app_index, listen_session_get_handle (listener));
 }
 
 int
 vnet_connect_uri (vnet_connect_args_t * a)
 {
-  ip46_address_t ip46_address;
-  u16 port;
+  transport_endpoint_t tep;
   session_type_t sst;
-  application_t *app;
   int rv;
 
-  app = application_lookup (a->api_client_index);
-  if (app)
-    {
-      clib_warning ("Already have a connect from this app");
-      return VNET_API_ERROR_INVALID_VALUE_2;
-    }
-
   /* Parse uri */
-  rv = parse_uri (a->uri, &sst, &ip46_address, &port);
+  memset (&tep, 0, sizeof (tep));
+  rv = parse_uri (a->uri, &sst, &tep);
   if (rv)
     return rv;
 
-  return vnet_connect_i (a->api_client_index, a->api_context, sst,
-			 &ip46_address, port, a->options, a->mp,
-			 a->session_cb_vft);
+  return vnet_connect_i (a->app_index, a->api_context, sst, &tep, a->mp);
 }
 
 int
-vnet_disconnect_session (u32 session_index, u32 thread_index)
+vnet_disconnect_session (vnet_disconnect_args_t * a)
 {
-  stream_session_t *session;
+  u32 index, thread_index;
+  stream_session_t *s;
 
-  session = stream_session_get (session_index, thread_index);
-  stream_session_disconnect (session);
+  stream_session_parse_handle (a->handle, &index, &thread_index);
+  s = stream_session_get_if_valid (index, thread_index);
+
+  if (!s || s->app_index != a->app_index)
+    return VNET_API_ERROR_INVALID_VALUE;
 
+  stream_session_disconnect (s);
   return 0;
 }
 
-
 int
 vnet_bind (vnet_bind_args_t * a)
 {
-  application_t *server = 0;
   session_type_t sst = SESSION_N_TYPES;
   int rv;
 
   sst = session_type_from_proto_and_ip (a->proto, a->tep.is_ip4);
-  if ((rv = vnet_bind_i (a->api_client_index, &a->tep.ip, a->tep.port, sst,
-			 a->options, a->session_cb_vft, &server,
-			 &a->segment_name_length, a->segment_name)))
+  if ((rv = vnet_bind_i (a->app_index, sst, &a->tep, &a->handle)))
     return rv;
 
-  a->server_event_queue_address = (u64) server->event_queue;
-  a->handle = (u64) a->tep.vrf << 32 | (u64) server->session_index;
   return 0;
 }
 
 int
 vnet_unbind (vnet_unbind_args_t * a)
 {
-  application_t *server;
-
-  if (a->api_client_index != ~0)
-    {
-      ASSERT (vl_api_client_index_to_registration (a->api_client_index));
-    }
-
-  /* Make sure this is the right one */
-  server = application_lookup (a->api_client_index);
-  ASSERT (server->session_index == (0xFFFFFFFF & a->handle));
-
-  /* TODO use handle to disambiguate namespaces/vrfs */
-  return vnet_unbind_i (a->api_client_index);
+  return vnet_unbind_i (a->app_index, a->handle);
 }
 
 int
 vnet_connect (vnet_connect_args_t * a)
 {
   session_type_t sst;
-  application_t *app;
-
-  app = application_lookup (a->api_client_index);
-  if (app)
-    {
-      clib_warning ("Already have a connect from this app");
-      return VNET_API_ERROR_INVALID_VALUE_2;
-    }
 
   sst = session_type_from_proto_and_ip (a->proto, a->tep.is_ip4);
-  return vnet_connect_i (a->api_client_index, a->api_context, sst, &a->tep.ip,
-			 a->tep.port, a->options, a->mp, a->session_cb_vft);
+  return vnet_connect_i (a->app_index, a->api_context, sst, &a->tep, a->mp);
 }
 
 int
diff --git a/src/vnet/session/application_interface.h b/src/vnet/session/application_interface.h
index a5f2b9a6..2c497531 100644
--- a/src/vnet/session/application_interface.h
+++ b/src/vnet/session/application_interface.h
@@ -28,6 +28,27 @@ typedef enum _session_api_proto
   SESSION_PROTO_UDP
 } session_api_proto_t;
 
+typedef struct _vnet_app_attach_args_t
+{
+  u32 api_client_index;
+  u64 *options;
+  session_cb_vft_t *session_cb_vft;
+
+  /*
+   * Results
+   */
+  u8 *segment_name;
+  u32 segment_name_length;
+  u32 segment_size;
+  u64 app_event_queue_address;
+  u32 app_index;
+} vnet_app_attach_args_t;
+
+typedef struct _vnet_app_detach_args_t
+{
+  u32 app_index;
+} vnet_app_detach_args_t;
+
 typedef struct _vnet_bind_args_t
 {
   union
@@ -40,9 +61,7 @@ typedef struct _vnet_bind_args_t
     };
   };
 
-  u32 api_client_index;
-  u64 *options;
-  session_cb_vft_t *session_cb_vft;
+  u32 app_index;
 
   /*
    * Results
@@ -60,7 +79,7 @@ typedef struct _vnet_unbind_args_t
     char *uri;
     u64 handle;
   };
-  u32 api_client_index;
+  u32 app_index;
 } vnet_unbind_args_t;
 
 typedef struct _vnet_connect_args
@@ -74,10 +93,8 @@ typedef struct _vnet_connect_args
       session_api_proto_t proto;
     };
   };
-  u32 api_client_index;
+  u32 app_index;
   u32 api_context;
-  u64 *options;
-  session_cb_vft_t *session_cb_vft;
 
   /* Used for redirects */
   void *mp;
@@ -86,12 +103,13 @@ typedef struct _vnet_connect_args
 typedef struct _vnet_disconnect_args_t
 {
   u64 handle;
-  u32 api_client_index;
+  u32 app_index;
 } vnet_disconnect_args_t;
 
-/* Bind / connect options */
+/* Application attach options */
 typedef enum
 {
+  APP_EVT_QUEUE_SIZE,
   SESSION_OPTIONS_FLAGS,
   SESSION_OPTIONS_SEGMENT_SIZE,
   SESSION_OPTIONS_ADD_SEGMENT_SIZE,
@@ -99,7 +117,7 @@ typedef enum
   SESSION_OPTIONS_TX_FIFO_SIZE,
   SESSION_OPTIONS_ACCEPT_COOKIE,
   SESSION_OPTIONS_N_OPTIONS
-} session_options_index_t;
+} app_attach_options_index_t;
 
 /** Server can handle delegated connect requests from local clients */
 #define SESSION_OPTIONS_FLAGS_USE_FIFO	(1<<0)
@@ -109,10 +127,13 @@ typedef enum
 
 #define VNET_CONNECT_REDIRECTED	123
 
+int vnet_application_attach (vnet_app_attach_args_t * a);
+int vnet_application_detach (vnet_app_detach_args_t * a);
+
 int vnet_bind_uri (vnet_bind_args_t *);
-int vnet_unbind_uri (char *uri, u32 api_client_index);
+int vnet_unbind_uri (vnet_unbind_args_t * a);
 int vnet_connect_uri (vnet_connect_args_t * a);
-int vnet_disconnect_session (u32 session_index, u32 thread_index);
+int vnet_disconnect_session (vnet_disconnect_args_t * a);
 
 int vnet_bind (vnet_bind_args_t * a);
 int vnet_connect (vnet_connect_args_t * a);
diff --git a/src/vnet/session/segment_manager.c b/src/vnet/session/segment_manager.c
new file mode 100644
index 00000000..16e5bc56
--- /dev/null
+++ b/src/vnet/session/segment_manager.c
@@ -0,0 +1,342 @@
+/*
+ * Copyright (c) 2017 Cisco and/or its affiliates.
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at:
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include <vnet/session/segment_manager.h>
+#include <vnet/session/session.h>
+#include <vnet/session/application.h>
+
+/**
+ * Counter used to build segment names
+ */
+u32 segment_name_counter = 0;
+
+/**
+ * Pool of segment managers
+ */
+segment_manager_t *segment_managers = 0;
+
+/**
+ * Default fifo and segment size. TODO config.
+ */
+u32 default_fifo_size = 1 << 16;
+u32 default_segment_size = 1 << 20;
+
+void
+segment_manager_get_segment_info (u32 index, u8 ** name, u32 * size)
+{
+  svm_fifo_segment_private_t *s;
+  s = svm_fifo_get_segment (index);
+  *name = s->h->segment_name;
+  *size = s->ssvm.ssvm_size;
+}
+
+always_inline int
+session_manager_add_segment_i (segment_manager_t * sm, u32 segment_size,
+			       u8 * segment_name)
+{
+  svm_fifo_segment_create_args_t _ca, *ca = &_ca;
+  int rv;
+
+  memset (ca, 0, sizeof (*ca));
+
+  ca->segment_name = (char *) segment_name;
+  ca->segment_size = segment_size;
+
+  rv = svm_fifo_segment_create (ca);
+  if (rv)
+    {
+      clib_warning ("svm_fifo_segment_create ('%s', %d) failed",
+		    ca->segment_name, ca->segment_size);
+      vec_free (segment_name);
+      return VNET_API_ERROR_SVM_SEGMENT_CREATE_FAIL;
+    }
+
+  vec_add1 (sm->segment_indices, ca->new_segment_index);
+
+  return 0;
+}
+
+int
+session_manager_add_segment (segment_manager_t * sm)
+{
+  u8 *segment_name;
+  svm_fifo_segment_create_args_t _ca, *ca = &_ca;
+  u32 add_segment_size;
+  int rv;
+
+  memset (ca, 0, sizeof (*ca));
+  segment_name = format (0, "%d-%d%c", getpid (), segment_name_counter++, 0);
+  add_segment_size = sm->properties->add_segment_size ?
+    sm->properties->add_segment_size : default_segment_size;
+
+  rv = session_manager_add_segment_i (sm, add_segment_size, segment_name);
+  vec_free (segment_name);
+  return rv;
+}
+
+int
+session_manager_add_first_segment (segment_manager_t * sm, u32 segment_size)
+{
+  svm_fifo_segment_create_args_t _ca, *ca = &_ca;
+  u8 *segment_name;
+  int rv;
+
+  memset (ca, 0, sizeof (*ca));
+  segment_name = format (0, "%d-%d%c", getpid (), segment_name_counter++, 0);
+  rv = session_manager_add_segment_i (sm, segment_size, segment_name);
+  vec_free (segment_name);
+  return rv;
+}
+
+/**
+ * Initializes segment manager based on options provided.
+ * Returns error if svm segment allocation fails.
+ */
+int
+segment_manager_init (segment_manager_t * sm,
+		      segment_manager_properties_t * properties,
+		      u32 first_seg_size)
+{
+  int rv;
+
+  /* app allocates these */
+  sm->properties = properties;
+
+  if (first_seg_size > 0)
+    {
+      rv = session_manager_add_first_segment (sm, first_seg_size);
+      if (rv)
+	{
+	  clib_warning ("Failed to allocate segment");
+	  return rv;
+	}
+    }
+
+  return 0;
+}
+
+/**
+ * Removes segment manager.
+ *
+ * Since the fifos allocated in the segment keep backpointers to the sessions
+ * prior to removing the segment, we call session disconnect. This
+ * subsequently propages into transport.
+ */
+void
+segment_manager_del (segment_manager_t * sm)
+{
+  u32 *deleted_sessions = 0;
+  u32 *deleted_thread_indices = 0;
+  int i, j;
+
+  /* Across all fifo segments used by the server */
+  for (j = 0; j < vec_len (sm->segment_indices); j++)
+    {
+      svm_fifo_segment_private_t *fifo_segment;
+      svm_fifo_t **fifos;
+      /* Vector of fifos allocated in the segment */
+      fifo_segment = svm_fifo_get_segment (sm->segment_indices[j]);
+      fifos = svm_fifo_segment_get_fifos (fifo_segment);
+
+      /*
+       * Remove any residual sessions from the session lookup table
+       * Don't bother deleting the individual fifos, we're going to
+       * throw away the fifo segment in a minute.
+       */
+      for (i = 0; i < vec_len (fifos); i++)
+	{
+	  svm_fifo_t *fifo;
+	  u32 session_index, thread_index;
+	  stream_session_t *session;
+
+	  fifo = fifos[i];
+	  session_index = fifo->server_session_index;
+	  thread_index = fifo->server_thread_index;
+
+	  session = stream_session_get (session_index, thread_index);
+
+	  /* Add to the deleted_sessions vector (once!) */
+	  if (!session->is_deleted)
+	    {
+	      session->is_deleted = 1;
+	      vec_add1 (deleted_sessions, session_index);
+	      vec_add1 (deleted_thread_indices, thread_index);
+	    }
+	}
+
+      for (i = 0; i < vec_len (deleted_sessions); i++)
+	{
+	  stream_session_t *session;
+	  session = stream_session_get (deleted_sessions[i],
+					deleted_thread_indices[i]);
+
+	  /* Instead of directly removing the session call disconnect */
+	  stream_session_disconnect (session);
+
+	  /*
+	     stream_session_table_del (smm, session);
+	     pool_put(smm->sessions[deleted_thread_indices[i]], session);
+	   */
+	}
+
+      vec_reset_length (deleted_sessions);
+      vec_reset_length (deleted_thread_indices);
+
+      /* Instead of removing the segment, test when removing the session if
+       * the segment can be removed
+       */
+      /* svm_fifo_segment_delete (fifo_segment); */
+    }
+
+  vec_free (deleted_sessions);
+  vec_free (deleted_thread_indices);
+  pool_put (segment_managers, sm);
+}
+
+static int
+segment_manager_notify_app_seg_add (segment_manager_t * sm,
+				    u32 fifo_segment_index)
+{
+  application_t *app = application_get (sm->app_index);
+  u32 seg_size = 0;
+  u8 *seg_name;
+
+  /* Send an API message to the external app, to map new segment */
+  ASSERT (app->cb_fns.add_segment_callback);
+
+  segment_manager_get_segment_info (fifo_segment_index, &seg_name, &seg_size);
+  return app->cb_fns.add_segment_callback (app->api_client_index, seg_name,
+					   seg_size);
+}
+
+int
+segment_manager_alloc_session_fifos (segment_manager_t * sm,
+				     svm_fifo_t ** server_rx_fifo,
+				     svm_fifo_t ** server_tx_fifo,
+				     u32 * fifo_segment_index)
+{
+  svm_fifo_segment_private_t *fifo_segment;
+  u32 fifo_size, sm_index;
+  u8 added_a_segment = 0;
+  int i;
+
+  /* Allocate svm fifos */
+  ASSERT (vec_len (sm->segment_indices));
+
+again:
+  for (i = 0; i < vec_len (sm->segment_indices); i++)
+    {
+      *fifo_segment_index = sm->segment_indices[i];
+      fifo_segment = svm_fifo_get_segment (*fifo_segment_index);
+
+      fifo_size = sm->properties->rx_fifo_size;
+      fifo_size = (fifo_size == 0) ? default_fifo_size : fifo_size;
+      *server_rx_fifo = svm_fifo_segment_alloc_fifo (fifo_segment, fifo_size);
+
+      fifo_size = sm->properties->tx_fifo_size;
+      fifo_size = (fifo_size == 0) ? default_fifo_size : fifo_size;
+      *server_tx_fifo = svm_fifo_segment_alloc_fifo (fifo_segment, fifo_size);
+
+      if (*server_rx_fifo == 0)
+	{
+	  /* This would be very odd, but handle it... */
+	  if (*server_tx_fifo != 0)
+	    {
+	      svm_fifo_segment_free_fifo (fifo_segment, *server_tx_fifo);
+	      *server_tx_fifo = 0;
+	    }
+	  continue;
+	}
+      if (*server_tx_fifo == 0)
+	{
+	  if (*server_rx_fifo != 0)
+	    {
+	      svm_fifo_segment_free_fifo (fifo_segment, *server_rx_fifo);
+	      *server_rx_fifo = 0;
+	    }
+	  continue;
+	}
+      break;
+    }
+
+  /* See if we're supposed to create another segment */
+  if (*server_rx_fifo == 0)
+    {
+      if (sm->properties->add_segment)
+	{
+	  if (added_a_segment)
+	    {
+	      clib_warning ("added a segment, still cant allocate a fifo");
+	      return SESSION_ERROR_NEW_SEG_NO_SPACE;
+	    }
+
+	  if (session_manager_add_segment (sm))
+	    return VNET_API_ERROR_URI_FIFO_CREATE_FAILED;
+
+	  added_a_segment = 1;
+	  goto again;
+	}
+      else
+	{
+	  clib_warning ("No space to allocate fifos!");
+	  return SESSION_ERROR_NO_SPACE;
+	}
+    }
+
+  if (added_a_segment)
+    return segment_manager_notify_app_seg_add (sm, *fifo_segment_index);
+
+  /* Backpointers to segment manager */
+  sm_index = segment_manager_index (sm);
+  (*server_tx_fifo)->segment_manager = sm_index;
+  (*server_rx_fifo)->segment_manager = sm_index;
+
+  return 0;
+}
+
+void
+segment_manager_dealloc_fifos (u32 svm_segment_index, svm_fifo_t * rx_fifo,
+			       svm_fifo_t * tx_fifo)
+{
+  segment_manager_t *sm;
+  svm_fifo_segment_private_t *fifo_segment;
+
+  fifo_segment = svm_fifo_get_segment (svm_segment_index);
+  svm_fifo_segment_free_fifo (fifo_segment, rx_fifo);
+  svm_fifo_segment_free_fifo (fifo_segment, tx_fifo);
+
+  /* If we have segment manager, try doing some cleanup.
+   * It's possible to have no segment manager if the session was removed
+   * as result of a detach */
+  sm = segment_manager_get_if_valid (rx_fifo->segment_manager);
+  if (sm)
+    {
+      /* Remove segment only if it holds no fifos and not the first */
+      if (sm->segment_indices[0] != svm_segment_index
+	  && !svm_fifo_segment_has_fifos (fifo_segment))
+	{
+	  svm_fifo_segment_delete (fifo_segment);
+	  vec_del1 (sm->segment_indices, svm_segment_index);
+	}
+    }
+}
+
+/*
+ * fd.io coding-style-patch-verification: ON
+ *
+ * Local Variables:
+ * eval: (c-set-style "gnu")
+ * End:
+ */
diff --git a/src/vnet/session/segment_manager.h b/src/vnet/session/segment_manager.h
new file mode 100644
index 00000000..778d6040
--- /dev/null
+++ b/src/vnet/session/segment_manager.h
@@ -0,0 +1,106 @@
+/*
+ * Copyright (c) 2017 Cisco and/or its affiliates.
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at:
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#ifndef SRC_VNET_SESSION_SEGMENT_MANAGER_H_
+#define SRC_VNET_SESSION_SEGMENT_MANAGER_H_
+
+#include <vnet/vnet.h>
+#include <svm/svm_fifo_segment.h>
+
+typedef struct _segment_manager_properties
+{
+  /** Session fifo sizes.  */
+  u32 rx_fifo_size;
+  u32 tx_fifo_size;
+
+  /** Configured additional segment size */
+  u32 add_segment_size;
+
+  /** Flag that indicates if additional segments should be created */
+  u8 add_segment;
+
+} segment_manager_properties_t;
+
+typedef struct _segment_manager
+{
+  /** segments mapped by this manager */
+  u32 *segment_indices;
+
+  /** Owner app index */
+  u32 app_index;
+
+  /** Pointer to manager properties. Could be shared among all of
+   * an app's segment managers s*/
+  segment_manager_properties_t *properties;
+} segment_manager_t;
+
+/** Pool of segment managers */
+extern segment_manager_t *segment_managers;
+
+always_inline segment_manager_t *
+segment_manager_new ()
+{
+  segment_manager_t *sm;
+  pool_get (segment_managers, sm);
+  memset (sm, 0, sizeof (*sm));
+  return sm;
+}
+
+always_inline segment_manager_t *
+segment_manager_get (u32 index)
+{
+  return pool_elt_at_index (segment_managers, index);
+}
+
+always_inline segment_manager_t *
+segment_manager_get_if_valid (u32 index)
+{
+  if (pool_is_free_index (segment_managers, index))
+    return 0;
+  return pool_elt_at_index (segment_managers, index);
+}
+
+always_inline u32
+segment_manager_index (segment_manager_t * sm)
+{
+  return sm - segment_managers;
+}
+
+int
+segment_manager_init (segment_manager_t * sm,
+		      segment_manager_properties_t * properties,
+		      u32 seg_size);
+
+void segment_manager_get_segment_info (u32 index, u8 ** name, u32 * size);
+int
+session_manager_add_first_segment (segment_manager_t * sm, u32 segment_size);
+int session_manager_add_segment (segment_manager_t * sm);
+void segment_manager_del (segment_manager_t * sm);
+int
+segment_manager_alloc_session_fifos (segment_manager_t * sm,
+				     svm_fifo_t ** server_rx_fifo,
+				     svm_fifo_t ** server_tx_fifo,
+				     u32 * fifo_segment_index);
+void
+segment_manager_dealloc_fifos (u32 svm_segment_index, svm_fifo_t * rx_fifo,
+			       svm_fifo_t * tx_fifo);
+
+#endif /* SRC_VNET_SESSION_SEGMENT_MANAGER_H_ */
+/*
+ * fd.io coding-style-patch-verification: ON
+ *
+ * Local Variables:
+ * eval: (c-set-style "gnu")
+ * End:
+ */
diff --git a/src/vnet/session/session.api b/src/vnet/session/session.api
index 582765b5..e207e46f 100644
--- a/src/vnet/session/session.api
+++ b/src/vnet/session/session.api
@@ -13,6 +13,68 @@
  * limitations under the License.
  */
  
+/** \brief client->vpp, attach application to session layer
+    @param client_index - opaque cookie to identify the sender
+    @param context - sender context, to match reply w/ request
+    @param initial_segment_size - size of the initial shm segment to be 
+    							  allocated
+    @param options - segment size, fifo sizes, etc.
+*/
+ define application_attach {
+    u32 client_index;
+    u32 context;
+    u32 initial_segment_size;
+    u64 options[16];
+ };
+ 
+ /** \brief Application attach reply
+    @param context - sender context, to match reply w/ request
+    @param retval - return code for the request
+    @param app_event_queue_address - vpp event queue address or 0 if this 
+                                 	 connection shouldn't send events
+    @param segment_size - size of first shm segment
+    @param segment_name_length - length of segment name 
+    @param segment_name - name of segment client needs to attach to
+*/
+define application_attach_reply {
+    u32 context;
+    i32 retval;
+    u64 app_event_queue_address;
+    u32 segment_size;
+    u8 segment_name_length;
+    u8 segment_name[128];
+};
+
+ /** \brief client->vpp, attach application to session layer
+    @param client_index - opaque cookie to identify the sender
+    @param context - sender context, to match reply w/ request
+*/
+ define application_detach {
+    u32 client_index;
+    u32 context;
+ };
+ 
+ /** \brief detach reply
+    @param context - sender context, to match reply w/ request
+    @param retval - return code for the request
+*/
+define application_detach_reply {
+    u32 context;
+    i32 retval;
+};
+ 
+/** \brief vpp->client, please map an additional shared memory segment
+    @param client_index - opaque cookie to identify the sender
+    @param context - sender context, to match reply w/ request
+    @param segment_name - 
+*/
+define map_another_segment {
+    u32 client_index;
+    u32 context;
+    u32 segment_size;
+    u8 segment_name[128];
+};
+
  /** \brief Bind to a given URI
     @param client_index - opaque cookie to identify the sender
     @param context - sender context, to match reply w/ request
@@ -25,9 +87,7 @@ define bind_uri {
   u32 client_index;
   u32 context;
   u32 accept_cookie;
-  u32 initial_segment_size;
   u8 uri[128];
-  u64 options[16];
 };
 
 /** \brief Unbind a given URI
@@ -49,7 +109,10 @@ define unbind_uri {
     @param accept_cookie - sender accept cookie, to identify this bind flavor
     @param uri - a URI, e.g. "tcp4://0.0.0.0/0/80"
                  "tcp6://::/0/80" [ipv6], etc.
-    @param options - socket options, fifo sizes, etc.
+    @param options - socket options, fifo sizes, etc. passed by vpp to the
+    				 server when redirecting connects 
+    @param client_queue_address - binary API client queue address. Used by 
+    							  local server when connect was redirected.
 */
 define connect_uri {
   u32 client_index;
@@ -62,18 +125,10 @@ define connect_uri {
 /** \brief Bind reply
     @param context - sender context, to match reply w/ request
     @param retval - return code for the request
-    @param event_queue_address - vpp event queue address or 0 if this 
-                                 connection shouldn't send events
-    @param segment_name_length - length of segment name 
-    @param segment_name - name of segment client needs to attach to
 */
 define bind_uri_reply {
     u32 context;
     i32 retval;
-    u64 server_event_queue_address;
-    u8 segment_name_length;
-    u32 segment_size;
-    u8 segment_name[128];
 };
 
 /** \brief unbind reply
@@ -88,43 +143,28 @@ define unbind_uri_reply {
 /** \brief vpp->client, connect reply
     @param context - sender context, to match reply w/ request
     @param retval - return code for the request
+    @param handle - session handle
     @param server_rx_fifo - rx (vpp -> vpp-client) fifo address 
     @param server_tx_fifo - tx (vpp-client -> vpp) fifo address 
-    @param session_index - session index;
-    @param session_thread_index - session thread index
-    @param session_type - session thread type
     @param vpp_event_queue_address - vpp's event queue address
-    @param client_event_queue_address - client's event queue address
+    @param segment_size - size of segment to be attached. Only for redirects.
     @param segment_name_length - non-zero if the client needs to attach to 
-                                 the fifo segment
+                                 the fifo segment. This should only happen 
+                                 if session was redirected.
     @param segment_name - set if the client needs to attach to the segment
 */
 define connect_uri_reply {
   u32 context;
   i32 retval;
+  u64 handle;
   u64 server_rx_fifo;
   u64 server_tx_fifo;
-  u32 session_index;
-  u32 session_thread_index;
-  u8 session_type;
-  u64 client_event_queue_address;
   u64 vpp_event_queue_address;
   u32 segment_size;
   u8 segment_name_length;
   u8 segment_name[128];
 };
 
-/** \brief vpp->client, please map an additional shared memory segment
-    @param context - sender context, to match reply w/ request
-    @param segment_name - 
-*/
-define map_another_segment {
-    u32 client_index;
-    u32 context;
-    u32 segment_size;
-    u8 segment_name[128];
-};
-
 /** \brief client->vpp
     @param context - sender context, to match reply w/ request
     @param retval - return code for the request
@@ -136,25 +176,27 @@ define map_another_segment_reply {
 
 /** \brief vpp->client, accept this session
     @param context - sender context, to match reply w/ request
-    @param accept_cookie - tells client which bind flavor just occurred
+    @param listener_handle - tells client which listener this pertains to
+    @param handle - unique session identifier
+    @param session_thread_index - thread index of new session
     @param rx_fifo_address - rx (vpp -> vpp-client) fifo address 
     @param tx_fifo_address - tx (vpp-client -> vpp) fifo address 
-    @param session_index - index of new session
-    @param session_thread_index - thread index of new session
     @param vpp_event_queue_address - vpp's event queue address
-    @param session_type - type of session
-    
+    @param port - remote port
+    @param is_ip4 - 1 if the ip is ip4
+    @param ip - remote ip
 */
 define accept_session {
   u32 client_index;
   u32 context;
-  u32 accept_cookie;
+  u64 listener_handle;
+  u64 handle; 
   u64 server_rx_fifo;
   u64 server_tx_fifo;
-  u32 session_index;
-  u32 session_thread_index;
   u64 vpp_event_queue_address;
-  u8 session_type;
+  u16 port;
+  u8 is_ip4;
+  u8 ip[16];
 };
 
 /** \brief client->vpp, reply to an accept message
@@ -167,23 +209,19 @@ define accept_session {
 define accept_session_reply {
   u32 context;
   i32 retval;
-  u8 session_type;
-  u8 session_thread_index;
-  u32 session_index;
+  u64 handle;
 };
 
 /** \brief bidirectional disconnect API
     @param client_index - opaque cookie to identify the sender
                           client to vpp direction only
     @param context - sender context, to match reply w/ request
-    @param session_index - cookie #1 from accept_session / connect_reply
-    @param session_thread_index - cookie #2
+    @param handle - session handle obtained from accept/connect
 */
 define disconnect_session {
   u32 client_index;
   u32 context;
-  u32 session_index;
-  u32 session_thread_index;
+  u64 handle;
 };
 
 /** \brief bidirectional disconnect reply API
@@ -191,31 +229,25 @@ define disconnect_session {
                           client to vpp direction only
     @param context - sender context, to match reply w/ request
     @param retval - return code for the request
-    @param session_index - session index from accept_session / connect_reply
-    @param session_thread_index - thread index from accept_session /
-                                  connect_reply
+    @param handle - session handle
 */
 define disconnect_session_reply {
   u32 client_index;
   u32 context;
   i32 retval;
-  u32 session_index;
-  u32 session_thread_index;
+  u64 handle;
 };
 
 /** \brief vpp->client reset session API
     @param client_index - opaque cookie to identify the sender
                           client to vpp direction only
     @param context - sender context, to match reply w/ request
-    @param session_index - session index from accept_session / connect_reply
-    @param session_thread_index - thread index from accept_session /
-                                  connect_reply
+    @param handle - session handle obtained via accept/connects
 */
 define reset_session {
   u32 client_index;
   u32 context;
-  u32 session_index;
-  u32 session_thread_index;
+  u64 handle;
 };
 
 /** \brief client->vpp reset session reply
@@ -223,16 +255,13 @@ define reset_session {
                           client to vpp direction only
     @param context - sender context, to match reply w/ request
     @param retval - return code for the request
-    @param session_index - session index from accept_session / connect_reply
-    @param session_thread_index - thread index from accept_session /
-                                  connect_reply
+    @param handle - session handle obtained via accept/connect
 */
 define reset_session_reply {
   u32 client_index;
   u32 context;
   i32 retval;
-  u32 session_index;
-  u32 session_thread_index;
+  u64 handle;
 };
 
 /** \brief Bind to an ip:port pair for a given transport protocol
@@ -277,7 +306,7 @@ define unbind_sock {
     @param proto - protocol 0 - TCP 1 - UDP
     @param client_queue_address - client's API queue address. Non-zero when 
                                   used to perform redirects
-    @param options - socket options, fifo sizes, etc.
+    @param options - socket options, fifo sizes, etc. when doing redirects
 */
 define connect_sock {
   u32 client_index;
@@ -326,7 +355,7 @@ define unbind_sock_reply {
     @param server_rx_fifo - rx (vpp -> vpp-client) fifo address 
     @param server_tx_fifo - tx (vpp-client -> vpp) fifo address 
     @param vpp_event_queue_address - vpp's event queue address
-    @param client_event_queue_address - client's event queue address
+    @param segment_size - size of segment to be attached. Only for redirects.
     @param segment_name_length - non-zero if the client needs to attach to 
                                  the fifo segment
     @param segment_name - set if the client needs to attach to the segment
@@ -337,92 +366,12 @@ define connect_sock_reply {
   u64 handle;
   u64 server_rx_fifo;
   u64 server_tx_fifo;
-  u64 client_event_queue_address;
   u64 vpp_event_queue_address;
   u32 segment_size;
   u8 segment_name_length;
   u8 segment_name[128];
 };
 
-/** \brief bidirectional disconnect API
-    @param client_index - opaque cookie to identify the sender
-                          client to vpp direction only
-    @param context - sender context, to match reply w/ request
-    @param handle - session handle obtained through accept/connect
-*/
-define disconnect_sock {
-  u32 client_index;
-  u32 context;
-  u64 handle;
-};
-
-/** \brief bidirectional disconnect reply API
-    @param client_index - opaque cookie to identify the sender
-                          client to vpp direction only
-    @param client_context - sender context, to match reply w/ request
-    @param handle - session handle obtained through accept/connect
-*/
-define disconnect_sock_reply {
-  u32 client_index;
-  u32 context;
-  i32 retval;
-  u64 handle;
-};
-
-/** \brief vpp->client, accept this session
-    @param context - sender context, to match reply w/ request
-    @param accept_cookie - tells client which bind flavor just occurred
-    @param handle - session handle obtained through accept/connect
-    @param rx_fifo_address - rx (vpp -> vpp-client) fifo address 
-    @param tx_fifo_address - tx (vpp-client -> vpp) fifo address 
-    @param vpp_event_queue_address - vpp's event queue address
-*/
-define accept_sock {
-  u32 client_index;
-  u32 context;
-  u32 accept_cookie;
-  u64 handle;
-  u64 server_rx_fifo;
-  u64 server_tx_fifo;
-  u64 vpp_event_queue_address;
-};
-
-/** \brief client->vpp, reply to an accept message
-    @param context - sender context, to match reply w/ request
-    @param retval - return code for the request
-    @param handle - session handle obtained through accept/connect
-*/
-define accept_sock_reply {
-  u32 context;
-  i32 retval;
-  u64 handle;
-};
-
-/** \brief vpp->client reset session API
-    @param client_index - opaque cookie to identify the sender
-                          client to vpp direction only
-    @param context - sender context, to match reply w/ request
-    @param handle - session handle obtained through accept/connect
-*/
-define reset_sock {
-  u32 client_index;
-  u32 context;
-  u64 handle;
-};
-
-/** \brief client->vpp reset session reply
-    @param client_index - opaque cookie to identify the sender
-                          client to vpp direction only
-    @param context - sender context, to match reply w/ request
-    @param handle - session handle obtained through accept/connect
-*/
-define reset_sock_reply {
-  u32 client_index;
-  u32 context;
-  i32 retval;
-  u64 handle;
-};
-
 /** \brief enable/disable session layer
     @param client_index - opaque cookie to identify the sender
                           client to vpp direction only
diff --git a/src/vnet/session/session.c b/src/vnet/session/session.c
index 8e2b2616..e6cfe7da 100644
--- a/src/vnet/session/session.c
+++ b/src/vnet/session/session.c
@@ -36,15 +36,14 @@ session_manager_main_t session_manager_main;
  * Session lookup key; (src-ip, dst-ip, src-port, dst-port, session-type)
  * Value: (owner thread index << 32 | session_index);
  */
-static void
-stream_session_table_add_for_tc (u8 sst, transport_connection_t * tc,
-				 u64 value)
+void
+stream_session_table_add_for_tc (transport_connection_t * tc, u64 value)
 {
   session_manager_main_t *smm = &session_manager_main;
   session_kv4_t kv4;
   session_kv6_t kv6;
 
-  switch (sst)
+  switch (tc->proto)
     {
     case SESSION_TYPE_IP4_UDP:
     case SESSION_TYPE_IP4_TCP:
@@ -72,12 +71,12 @@ stream_session_table_add (session_manager_main_t * smm, stream_session_t * s,
 
   tc = tp_vfts[s->session_type].get_connection (s->connection_index,
 						s->thread_index);
-  stream_session_table_add_for_tc (s->session_type, tc, value);
+  stream_session_table_add_for_tc (tc, value);
 }
 
 static void
-stream_session_half_open_table_add (u8 sst, transport_connection_t * tc,
-				    u64 value)
+stream_session_half_open_table_add (session_type_t sst,
+				    transport_connection_t * tc, u64 value)
 {
   session_manager_main_t *smm = &session_manager_main;
   session_kv4_t kv4;
@@ -105,14 +104,13 @@ stream_session_half_open_table_add (u8 sst, transport_connection_t * tc,
     }
 }
 
-static int
-stream_session_table_del_for_tc (session_manager_main_t * smm, u8 sst,
-				 transport_connection_t * tc)
+int
+stream_session_table_del_for_tc (transport_connection_t * tc)
 {
+  session_manager_main_t *smm = &session_manager_main;
   session_kv4_t kv4;
   session_kv6_t kv6;
-
-  switch (sst)
+  switch (tc->proto)
     {
     case SESSION_TYPE_IP4_UDP:
     case SESSION_TYPE_IP4_TCP:
@@ -141,7 +139,7 @@ stream_session_table_del (session_manager_main_t * smm, stream_session_t * s)
 
   ts = tp_vfts[s->session_type].get_connection (s->connection_index,
 						s->thread_index);
-  return stream_session_table_del_for_tc (smm, s->session_type, ts);
+  return stream_session_table_del_for_tc (ts);
 }
 
 static void
@@ -383,7 +381,7 @@ stream_session_lookup_transport6 (ip6_address_t * lcl, ip6_address_t * rmt,
  * Allocate vpp event queue (once) per worker thread
  */
 void
-vpp_session_event_queue_allocate (session_manager_main_t * smm,
+session_vpp_event_queue_allocate (session_manager_main_t * smm,
 				  u32 thread_index)
 {
   api_main_t *am = &api_main;
@@ -406,266 +404,24 @@ vpp_session_event_queue_allocate (session_manager_main_t * smm,
     }
 }
 
-void
-session_manager_get_segment_info (u32 index, u8 ** name, u32 * size)
-{
-  svm_fifo_segment_private_t *s;
-  s = svm_fifo_get_segment (index);
-  *name = s->h->segment_name;
-  *size = s->ssvm.ssvm_size;
-}
-
-always_inline int
-session_manager_add_segment_i (session_manager_main_t * smm,
-			       session_manager_t * sm,
-			       u32 segment_size, u8 * segment_name)
-{
-  svm_fifo_segment_create_args_t _ca, *ca = &_ca;
-  int rv;
-
-  memset (ca, 0, sizeof (*ca));
-
-  ca->segment_name = (char *) segment_name;
-  ca->segment_size = segment_size;
-
-  rv = svm_fifo_segment_create (ca);
-  if (rv)
-    {
-      clib_warning ("svm_fifo_segment_create ('%s', %d) failed",
-		    ca->segment_name, ca->segment_size);
-      vec_free (segment_name);
-      return -1;
-    }
-
-  vec_add1 (sm->segment_indices, ca->new_segment_index);
-
-  return 0;
-}
-
-static int
-session_manager_add_segment (session_manager_main_t * smm,
-			     session_manager_t * sm)
-{
-  u8 *segment_name;
-  svm_fifo_segment_create_args_t _ca, *ca = &_ca;
-  u32 add_segment_size;
-  u32 default_segment_size = 128 << 10;
-
-  memset (ca, 0, sizeof (*ca));
-  segment_name = format (0, "%d-%d%c", getpid (),
-			 smm->unique_segment_name_counter++, 0);
-  add_segment_size =
-    sm->add_segment_size ? sm->add_segment_size : default_segment_size;
-
-  return session_manager_add_segment_i (smm, sm, add_segment_size,
-					segment_name);
-}
-
-int
-session_manager_add_first_segment (session_manager_main_t * smm,
-				   session_manager_t * sm, u32 segment_size,
-				   u8 ** segment_name)
-{
-  svm_fifo_segment_create_args_t _ca, *ca = &_ca;
-  memset (ca, 0, sizeof (*ca));
-  *segment_name = format (0, "%d-%d%c", getpid (),
-			  smm->unique_segment_name_counter++, 0);
-  return session_manager_add_segment_i (smm, sm, segment_size, *segment_name);
-}
-
-void
-session_manager_del (session_manager_main_t * smm, session_manager_t * sm)
-{
-  u32 *deleted_sessions = 0;
-  u32 *deleted_thread_indices = 0;
-  int i, j;
-
-  /* Across all fifo segments used by the server */
-  for (j = 0; j < vec_len (sm->segment_indices); j++)
-    {
-      svm_fifo_segment_private_t *fifo_segment;
-      svm_fifo_t **fifos;
-      /* Vector of fifos allocated in the segment */
-      fifo_segment = svm_fifo_get_segment (sm->segment_indices[j]);
-      fifos = (svm_fifo_t **) fifo_segment->h->fifos;
-
-      /*
-       * Remove any residual sessions from the session lookup table
-       * Don't bother deleting the individual fifos, we're going to
-       * throw away the fifo segment in a minute.
-       */
-      for (i = 0; i < vec_len (fifos); i++)
-	{
-	  svm_fifo_t *fifo;
-	  u32 session_index, thread_index;
-	  stream_session_t *session;
-
-	  fifo = fifos[i];
-	  session_index = fifo->server_session_index;
-	  thread_index = fifo->server_thread_index;
-
-	  session = pool_elt_at_index (smm->sessions[thread_index],
-				       session_index);
-
-	  /* Add to the deleted_sessions vector (once!) */
-	  if (!session->is_deleted)
-	    {
-	      session->is_deleted = 1;
-	      vec_add1 (deleted_sessions,
-			session - smm->sessions[thread_index]);
-	      vec_add1 (deleted_thread_indices, thread_index);
-	    }
-	}
-
-      for (i = 0; i < vec_len (deleted_sessions); i++)
-	{
-	  stream_session_t *session;
-
-	  session =
-	    pool_elt_at_index (smm->sessions[deleted_thread_indices[i]],
-			       deleted_sessions[i]);
-
-	  /* Instead of directly removing the session call disconnect */
-	  stream_session_disconnect (session);
-
-	  /*
-	     stream_session_table_del (smm, session);
-	     pool_put(smm->sessions[deleted_thread_indices[i]], session);
-	   */
-	}
-
-      vec_reset_length (deleted_sessions);
-      vec_reset_length (deleted_thread_indices);
-
-      /* Instead of removing the segment, test when removing the session if
-       * the segment can be removed
-       */
-      /* svm_fifo_segment_delete (fifo_segment); */
-    }
-
-  vec_free (deleted_sessions);
-  vec_free (deleted_thread_indices);
-}
-
-int
-session_manager_allocate_session_fifos (session_manager_main_t * smm,
-					session_manager_t * sm,
-					svm_fifo_t ** server_rx_fifo,
-					svm_fifo_t ** server_tx_fifo,
-					u32 * fifo_segment_index,
-					u8 * added_a_segment)
-{
-  svm_fifo_segment_private_t *fifo_segment;
-  u32 fifo_size, default_fifo_size = 1 << 16;	/* TODO config */
-  int i;
-
-  *added_a_segment = 0;
-
-  /* Allocate svm fifos */
-  ASSERT (vec_len (sm->segment_indices));
-
-again:
-  for (i = 0; i < vec_len (sm->segment_indices); i++)
-    {
-      *fifo_segment_index = sm->segment_indices[i];
-      fifo_segment = svm_fifo_get_segment (*fifo_segment_index);
-
-      fifo_size = sm->rx_fifo_size;
-      fifo_size = (fifo_size == 0) ? default_fifo_size : fifo_size;
-      *server_rx_fifo = svm_fifo_segment_alloc_fifo (fifo_segment, fifo_size);
-
-      fifo_size = sm->tx_fifo_size;
-      fifo_size = (fifo_size == 0) ? default_fifo_size : fifo_size;
-      *server_tx_fifo = svm_fifo_segment_alloc_fifo (fifo_segment, fifo_size);
-
-      if (*server_rx_fifo == 0)
-	{
-	  /* This would be very odd, but handle it... */
-	  if (*server_tx_fifo != 0)
-	    {
-	      svm_fifo_segment_free_fifo (fifo_segment, *server_tx_fifo);
-	      *server_tx_fifo = 0;
-	    }
-	  continue;
-	}
-      if (*server_tx_fifo == 0)
-	{
-	  if (*server_rx_fifo != 0)
-	    {
-	      svm_fifo_segment_free_fifo (fifo_segment, *server_rx_fifo);
-	      *server_rx_fifo = 0;
-	    }
-	  continue;
-	}
-      break;
-    }
-
-  /* See if we're supposed to create another segment */
-  if (*server_rx_fifo == 0)
-    {
-      if (sm->add_segment)
-	{
-	  if (*added_a_segment)
-	    {
-	      clib_warning ("added a segment, still cant allocate a fifo");
-	      return SESSION_ERROR_NEW_SEG_NO_SPACE;
-	    }
-
-	  if (session_manager_add_segment (smm, sm))
-	    return VNET_API_ERROR_URI_FIFO_CREATE_FAILED;
-
-	  *added_a_segment = 1;
-	  goto again;
-	}
-      else
-	{
-	  clib_warning ("No space to allocate fifos!");
-	  return SESSION_ERROR_NO_SPACE;
-	}
-    }
-  return 0;
-}
-
 int
-stream_session_create_i (session_manager_main_t * smm, application_t * app,
-			 transport_connection_t * tc,
+stream_session_create_i (segment_manager_t * sm, transport_connection_t * tc,
 			 stream_session_t ** ret_s)
 {
-  int rv;
+  session_manager_main_t *smm = &session_manager_main;
   svm_fifo_t *server_rx_fifo = 0, *server_tx_fifo = 0;
   u32 fifo_segment_index;
-  u32 pool_index, seg_size;
+  u32 pool_index;
   stream_session_t *s;
   u64 value;
   u32 thread_index = tc->thread_index;
-  session_manager_t *sm;
-  u8 segment_added;
-  u8 *seg_name;
-
-  sm = session_manager_get (app->session_manager_index);
-
-  /* Check the API queue */
-  if (app->mode == APP_SERVER && application_api_queue_is_full (app))
-    return SESSION_ERROR_API_QUEUE_FULL;
+  int rv;
 
-  if ((rv = session_manager_allocate_session_fifos (smm, sm, &server_rx_fifo,
-						    &server_tx_fifo,
-						    &fifo_segment_index,
-						    &segment_added)))
+  if ((rv = segment_manager_alloc_session_fifos (sm, &server_rx_fifo,
+						 &server_tx_fifo,
+						 &fifo_segment_index)))
     return rv;
 
-  if (segment_added && app->mode == APP_SERVER)
-    {
-      /* Send an API message to the external server, to map new segment */
-      ASSERT (app->cb_fns.add_segment_callback);
-
-      session_manager_get_segment_info (fifo_segment_index, &seg_name,
-					&seg_size);
-      if (app->cb_fns.add_segment_callback (app->api_client_index, seg_name,
-					    seg_size))
-	return VNET_API_ERROR_URI_FIFO_CREATE_FAILED;
-    }
-
   /* Create the session */
   pool_get (smm->sessions[thread_index], s);
   memset (s, 0, sizeof (*s));
@@ -682,10 +438,9 @@ stream_session_create_i (session_manager_main_t * smm, application_t * app,
   s->server_tx_fifo = server_tx_fifo;
 
   /* Initialize state machine, such as it is... */
-  s->session_type = app->session_type;
+  s->session_type = tc->proto;
   s->session_state = SESSION_STATE_CONNECTING;
-  s->app_index = application_get_index (app);
-  s->server_segment_index = fifo_segment_index;
+  s->svm_segment_index = fifo_segment_index;
   s->thread_index = thread_index;
   s->session_index = pool_index;
 
@@ -697,7 +452,7 @@ stream_session_create_i (session_manager_main_t * smm, application_t * app,
 
   /* Add to the main lookup table */
   value = (((u64) thread_index) << 32) | (u64) s->session_index;
-  stream_session_table_add_for_tc (app->session_type, tc, value);
+  stream_session_table_add_for_tc (tc, value);
 
   *ret_s = s;
 
@@ -881,94 +636,6 @@ session_manager_flush_enqueue_events (u32 thread_index)
   return errors;
 }
 
-/*
- * Start listening on server's ip/port pair for requested transport.
- *
- * Creates a 'dummy' stream session with state LISTENING to be used in session
- * lookups, prior to establishing connection. Requests transport to build
- * it's own specific listening connection.
- */
-int
-stream_session_start_listen (u32 server_index, ip46_address_t * ip, u16 port)
-{
-  session_manager_main_t *smm = &session_manager_main;
-  stream_session_t *s;
-  transport_connection_t *tc;
-  application_t *srv;
-  u32 tci;
-
-  srv = application_get (server_index);
-
-  pool_get (smm->listen_sessions[srv->session_type], s);
-  memset (s, 0, sizeof (*s));
-
-  s->session_type = srv->session_type;
-  s->session_state = SESSION_STATE_LISTENING;
-  s->session_index = s - smm->listen_sessions[srv->session_type];
-  s->app_index = srv->index;
-
-  /* Transport bind/listen  */
-  tci = tp_vfts[srv->session_type].bind (s->session_index, ip, port);
-
-  /* Attach transport to session */
-  s->connection_index = tci;
-  tc = tp_vfts[srv->session_type].get_listener (tci);
-
-  srv->session_index = s->session_index;
-
-  /* Add to the main lookup table */
-  stream_session_table_add_for_tc (s->session_type, tc, s->session_index);
-
-  return 0;
-}
-
-void
-stream_session_stop_listen (u32 server_index)
-{
-  session_manager_main_t *smm = &session_manager_main;
-  stream_session_t *listener;
-  transport_connection_t *tc;
-  application_t *srv;
-
-  srv = application_get (server_index);
-  listener = pool_elt_at_index (smm->listen_sessions[srv->session_type],
-				srv->session_index);
-
-  tc = tp_vfts[srv->session_type].get_listener (listener->connection_index);
-  stream_session_table_del_for_tc (smm, listener->session_type, tc);
-
-  tp_vfts[srv->session_type].unbind (listener->connection_index);
-  pool_put (smm->listen_sessions[srv->session_type], listener);
-}
-
-int
-connect_server_add_segment_cb (application_t * ss, char *segment_name,
-			       u32 segment_size)
-{
-  /* Does exactly nothing, but die */
-  ASSERT (0);
-  return 0;
-}
-
-void
-connects_session_manager_init (session_manager_main_t * smm, u8 session_type)
-{
-  session_manager_t *sm;
-  u32 connect_fifo_size = 256 << 10;	/* Config? */
-  u32 default_segment_size = 1 << 20;
-
-  pool_get (smm->session_managers, sm);
-  memset (sm, 0, sizeof (*sm));
-
-  sm->add_segment_size = default_segment_size;
-  sm->rx_fifo_size = connect_fifo_size;
-  sm->tx_fifo_size = connect_fifo_size;
-  sm->add_segment = 1;
-
-  session_manager_add_segment (smm, sm);
-  smm->connect_manager_index[session_type] = sm - smm->session_managers;
-}
-
 void
 stream_session_connect_notify (transport_connection_t * tc, u8 sst,
 			       u8 is_fail)
@@ -976,34 +643,36 @@ stream_session_connect_notify (transport_connection_t * tc, u8 sst,
   session_manager_main_t *smm = &session_manager_main;
   application_t *app;
   stream_session_t *new_s = 0;
-  u64 value;
+  u64 handle;
+  u32 api_context = 0;
 
-  value = stream_session_half_open_lookup (smm, &tc->lcl_ip, &tc->rmt_ip,
-					   tc->lcl_port, tc->rmt_port,
-					   tc->proto);
-  if (value == HALF_OPEN_LOOKUP_INVALID_VALUE)
+  handle = stream_session_half_open_lookup (smm, &tc->lcl_ip, &tc->rmt_ip,
+					    tc->lcl_port, tc->rmt_port,
+					    tc->proto);
+  if (handle == HALF_OPEN_LOOKUP_INVALID_VALUE)
     {
       clib_warning ("This can't be good!");
       return;
     }
 
-  app = application_get (value >> 32);
+  /* Get the app's index from the handle we stored when opening connection */
+  app = application_get (handle >> 32);
+  api_context = tc->s_index;
 
   if (!is_fail)
     {
-      /* Create new session (server segments are allocated if needed) */
-      if (stream_session_create_i (smm, app, tc, &new_s))
-	return;
+      segment_manager_t *sm;
+      sm = application_get_connect_segment_manager (app);
 
-      app->session_index = stream_session_get_index (new_s);
-      app->thread_index = new_s->thread_index;
+      /* Create new session (svm segments are allocated if needed) */
+      if (stream_session_create_i (sm, tc, &new_s))
+	return;
 
-      /* Allocate vpp event queue for this thread if needed */
-      vpp_session_event_queue_allocate (smm, tc->thread_index);
+      new_s->app_index = app->index;
     }
 
   /* Notify client */
-  app->cb_fns.session_connected_callback (app->api_client_index, new_s,
+  app->cb_fns.session_connected_callback (app->index, api_context, new_s,
 					  is_fail);
 
   /* Cleanup session lookup */
@@ -1046,48 +715,13 @@ void
 stream_session_delete (stream_session_t * s)
 {
   session_manager_main_t *smm = vnet_get_session_manager_main ();
-  svm_fifo_segment_private_t *fifo_segment;
-  application_t *app;
 
   /* Delete from the main lookup table. */
   stream_session_table_del (smm, s);
 
   /* Cleanup fifo segments */
-  fifo_segment = svm_fifo_get_segment (s->server_segment_index);
-  svm_fifo_segment_free_fifo (fifo_segment, s->server_rx_fifo);
-  svm_fifo_segment_free_fifo (fifo_segment, s->server_tx_fifo);
-
-  app = application_get_if_valid (s->app_index);
-
-  /* No app. A possibility: after disconnect application called unbind */
-  if (!app)
-    return;
-
-  if (app->mode == APP_CLIENT)
-    {
-      /* Cleanup app if client */
-      application_del (app);
-    }
-  else if (app->mode == APP_SERVER)
-    {
-      session_manager_t *sm;
-      svm_fifo_segment_private_t *fifo_segment;
-      svm_fifo_t **fifos;
-      u32 fifo_index;
-
-      /* For server, see if any segments can be removed */
-      sm = session_manager_get (app->session_manager_index);
-
-      /* Delete fifo */
-      fifo_segment = svm_fifo_get_segment (s->server_segment_index);
-      fifos = (svm_fifo_t **) fifo_segment->h->fifos;
-
-      fifo_index = svm_fifo_segment_index (fifo_segment);
-
-      /* Remove segment only if it holds no fifos and not the first */
-      if (sm->segment_indices[0] != fifo_index && vec_len (fifos) == 0)
-	svm_fifo_segment_delete (fifo_segment);
-    }
+  segment_manager_dealloc_fifos (s->svm_segment_index, s->server_rx_fifo,
+				 s->server_tx_fifo);
 
   pool_put (smm->sessions[s->thread_index], s);
 }
@@ -1134,21 +768,22 @@ int
 stream_session_accept (transport_connection_t * tc, u32 listener_index,
 		       u8 sst, u8 notify)
 {
-  session_manager_main_t *smm = &session_manager_main;
   application_t *server;
   stream_session_t *s, *listener;
+  segment_manager_t *sm;
 
   int rv;
 
   /* Find the server */
-  listener = pool_elt_at_index (smm->listen_sessions[sst], listener_index);
+  listener = listen_session_get (sst, listener_index);
   server = application_get (listener->app_index);
 
-  if ((rv = stream_session_create_i (smm, server, tc, &s)))
+  sm = application_get_listen_segment_manager (server, listener);
+  if ((rv = stream_session_create_i (sm, tc, &s)))
     return rv;
 
-  /* Allocate vpp event queue for this thread if needed */
-  vpp_session_event_queue_allocate (smm, tc->thread_index);
+  s->app_index = server->index;
+  s->listener_index = listener_index;
 
   /* Shoulder-tap the server */
   if (notify)
@@ -1159,37 +794,111 @@ stream_session_accept (transport_connection_t * tc, u32 listener_index,
   return 0;
 }
 
+/**
+ * Ask transport to open connection to remote transport endpoint.
+ *
+ * Stores handle for matching request with reply since the call can be
+ * asynchronous. For instance, for TCP the 3-way handshake must complete
+ * before reply comes. Session is only created once connection is established.
+ *
+ * @param app_index Index of the application requesting the connect
+ * @param st Session type requested.
+ * @param tep Remote transport endpoint
+ * @param res Resulting transport connection .
+ */
 int
-stream_session_open (u8 sst, ip46_address_t * addr, u16 port_host_byte_order,
-		     u32 app_index)
+stream_session_open (u32 app_index, session_type_t st,
+		     transport_endpoint_t * tep,
+		     transport_connection_t ** res)
 {
   transport_connection_t *tc;
-  u32 tci;
-  u64 value;
   int rv;
+  u64 handle;
 
-  /* Ask transport to open connection */
-  rv = tp_vfts[sst].open (addr, port_host_byte_order);
+  rv = tp_vfts[st].open (&tep->ip, tep->port);
   if (rv < 0)
     {
       clib_warning ("Transport failed to open connection.");
       return VNET_API_ERROR_SESSION_CONNECT_FAIL;
     }
 
-  tci = rv;
+  tc = tp_vfts[st].get_half_open ((u32) rv);
 
-  /* Get transport connection */
-  tc = tp_vfts[sst].get_half_open (tci);
-
-  /* Store api_client_index and transport connection index */
-  value = (((u64) app_index) << 32) | (u64) tc->c_index;
+  /* Save app and tc index. The latter is needed to help establish the
+   * connection while the former is needed when the connect notify comes
+   * and we have to notify the external app */
+  handle = (((u64) app_index) << 32) | (u64) tc->c_index;
 
   /* Add to the half-open lookup table */
-  stream_session_half_open_table_add (sst, tc, value);
+  stream_session_half_open_table_add (st, tc, handle);
+
+  *res = tc;
+
+  return 0;
+}
+
+/**
+ * Ask transport to listen on local transport endpoint.
+ *
+ * @param s Session for which listen will be called. Note that unlike
+ * 	    established sessions, listen sessions are not associated to a
+ * 	    thread.
+ * @param tep Local endpoint to be listened on.
+ */
+int
+stream_session_listen (stream_session_t * s, transport_endpoint_t * tep)
+{
+  transport_connection_t *tc;
+  u32 tci;
+
+  /* Transport bind/listen  */
+  tci = tp_vfts[s->session_type].bind (s->session_index, &tep->ip, tep->port);
+
+  if (tci == (u32) ~ 0)
+    return -1;
+
+  /* Attach transport to session */
+  s->connection_index = tci;
+  tc = tp_vfts[s->session_type].get_listener (tci);
+
+  /* Weird but handle it ... */
+  if (tc == 0)
+    return -1;
+
+  /* Add to the main lookup table */
+  stream_session_table_add_for_tc (tc, s->session_index);
 
   return 0;
 }
 
+/**
+ * Ask transport to stop listening on local transport endpoint.
+ *
+ * @param s Session to stop listening on. It must be in state LISTENING.
+ */
+int
+stream_session_stop_listen (stream_session_t * s)
+{
+  transport_connection_t *tc;
+
+  if (s->session_state != SESSION_STATE_LISTENING)
+    {
+      clib_warning ("not a listening session");
+      return -1;
+    }
+
+  tc = tp_vfts[s->session_type].get_listener (s->connection_index);
+  if (!tc)
+    {
+      clib_warning ("no transport");
+      return VNET_API_ERROR_ADDRESS_NOT_IN_USE;
+    }
+
+  stream_session_table_del_for_tc (tc);
+  tp_vfts[s->session_type].unbind (s->connection_index);
+  return 0;
+}
+
 /**
  * Disconnect session and propagate to transport. This should eventually
  * result in a delete notification that allows us to cleanup session state.
@@ -1297,6 +1006,10 @@ session_manager_main_enable (vlib_main_t * vm)
   vec_validate (smm->last_event_poll_by_thread, num_threads - 1);
 #endif
 
+  /* Allocate vpp event queues */
+  for (i = 0; i < vec_len (smm->vpp_event_queues); i++)
+    session_vpp_event_queue_allocate (smm, i);
+
   /* $$$$ preallocate hack config parameter */
   for (i = 0; i < 200000; i++)
     {
@@ -1322,9 +1035,6 @@ session_manager_main_enable (vlib_main_t * vm)
 			 200000 /* $$$$ config parameter nbuckets */ ,
 			 (64 << 20) /*$$$ config parameter table size */ );
 
-  for (i = 0; i < SESSION_N_TYPES; i++)
-    smm->connect_manager_index[i] = INVALID_INDEX;
-
   smm->is_enabled = 1;
 
   /* Enable TCP transport */
diff --git a/src/vnet/session/session.h b/src/vnet/session/session.h
index 6878b4d2..6e4ea96d 100644
--- a/src/vnet/session/session.h
+++ b/src/vnet/session/session.h
@@ -21,6 +21,7 @@
 #include <vppinfra/sparse_vec.h>
 #include <svm/svm_fifo_segment.h>
 #include <vnet/session/session_debug.h>
+#include <vnet/session/segment_manager.h>
 
 #define HALF_OPEN_LOOKUP_INVALID_VALUE ((u64)~0)
 #define INVALID_INDEX ((u32)~0)
@@ -107,6 +108,9 @@ typedef struct _stream_session_t
   svm_fifo_t *server_rx_fifo;
   svm_fifo_t *server_tx_fifo;
 
+  /** svm segment index where fifos were allocated */
+  u32 svm_segment_index;
+
   /** Type */
   u8 session_type;
 
@@ -133,27 +137,10 @@ typedef struct _stream_session_t
   /** stream server pool index */
   u32 app_index;
 
-  /** svm segment index */
-  u32 server_segment_index;
+  /** Parent listener session if the result of an accept */
+  u32 listener_index;
 } stream_session_t;
 
-typedef struct _session_manager
-{
-  /** segments mapped by this server */
-  u32 *segment_indices;
-
-  /** Session fifo sizes. They are provided for binds and take default
-   * values for connects */
-  u32 rx_fifo_size;
-  u32 tx_fifo_size;
-
-  /** Configured additional segment size */
-  u32 add_segment_size;
-
-  /** Flag that indicates if additional segments should be created */
-  u8 add_segment;
-} session_manager_t;
-
 /* Forward definition */
 typedef struct _session_manager_main session_manager_main_t;
 
@@ -206,11 +193,6 @@ struct _session_manager_main
   /** Unique segment name counter */
   u32 unique_segment_name_counter;
 
-  /* Connection manager used by incoming connects */
-  u32 connect_manager_index[SESSION_N_TYPES];
-
-  session_manager_t *session_managers;
-
   /** Per transport rx function that can either dequeue or peek */
   session_fifo_rx_fn *session_tx_fns[SESSION_N_TYPES];
 
@@ -242,37 +224,6 @@ vnet_get_session_manager_main ()
   return &session_manager_main;
 }
 
-always_inline session_manager_t *
-session_manager_get (u32 index)
-{
-  return pool_elt_at_index (session_manager_main.session_managers, index);
-}
-
-always_inline unix_shared_memory_queue_t *
-session_manager_get_vpp_event_queue (u32 thread_index)
-{
-  return session_manager_main.vpp_event_queues[thread_index];
-}
-
-always_inline session_manager_t *
-connects_session_manager_get (session_manager_main_t * smm,
-			      session_type_t session_type)
-{
-  return pool_elt_at_index (smm->session_managers,
-			    smm->connect_manager_index[session_type]);
-}
-
-void session_manager_get_segment_info (u32 index, u8 ** name, u32 * size);
-int session_manager_flush_enqueue_events (u32 thread_index);
-int
-session_manager_add_first_segment (session_manager_main_t * smm,
-				   session_manager_t * sm, u32 segment_size,
-				   u8 ** segment_name);
-void
-session_manager_del (session_manager_main_t * smm, session_manager_t * sm);
-void
-connects_session_manager_init (session_manager_main_t * smm, u8 session_type);
-
 /*
  * Stream session functions
  */
@@ -300,6 +251,8 @@ transport_connection_t
 				      u32 thread_index);
 stream_session_t *stream_session_lookup_listener (ip46_address_t * lcl,
 						  u16 lcl_port, u8 proto);
+void stream_session_table_add_for_tc (transport_connection_t * tc, u64 value);
+int stream_session_table_del_for_tc (transport_connection_t * tc);
 
 always_inline stream_session_t *
 stream_session_get_tsi (u64 ti_and_si, u32 thread_index)
@@ -310,7 +263,7 @@ stream_session_get_tsi (u64 ti_and_si, u32 thread_index)
 }
 
 always_inline stream_session_t *
-stream_session_get (u64 si, u32 thread_index)
+stream_session_get (u32 si, u32 thread_index)
 {
   return pool_elt_at_index (session_manager_main.sessions[thread_index], si);
 }
@@ -327,6 +280,40 @@ stream_session_get_if_valid (u64 si, u32 thread_index)
   return pool_elt_at_index (session_manager_main.sessions[thread_index], si);
 }
 
+always_inline u64
+stream_session_handle (stream_session_t * s)
+{
+  return ((u64) s->thread_index << 32) | (u64) s->session_index;
+}
+
+always_inline u32
+stream_session_index_from_handle (u64 handle)
+{
+  return handle & 0xFFFFFFFF;
+}
+
+always_inline u32
+stream_session_thread_from_handle (u64 handle)
+{
+  return handle >> 32;
+}
+
+always_inline void
+stream_session_parse_handle (u64 handle, u32 * index, u32 * thread_index)
+{
+  *index = stream_session_index_from_handle (handle);
+  *thread_index = stream_session_thread_from_handle (handle);
+}
+
+always_inline stream_session_t *
+stream_session_get_from_handle (u64 handle)
+{
+  session_manager_main_t *smm = &session_manager_main;
+  return pool_elt_at_index (smm->sessions[stream_session_thread_from_handle
+					  (handle)],
+			    stream_session_index_from_handle (handle));
+}
+
 always_inline stream_session_t *
 stream_session_listener_get (u8 sst, u64 si)
 {
@@ -375,13 +362,14 @@ void stream_session_reset_notify (transport_connection_t * tc);
 int
 stream_session_accept (transport_connection_t * tc, u32 listener_index,
 		       u8 sst, u8 notify);
-int stream_session_open (u8 sst, ip46_address_t * addr,
-			 u16 port_host_byte_order, u32 api_client_index);
+int
+stream_session_open (u32 app_index, session_type_t st,
+		     transport_endpoint_t * tep,
+		     transport_connection_t ** tc);
+int stream_session_listen (stream_session_t * s, transport_endpoint_t * tep);
+int stream_session_stop_listen (stream_session_t * s);
 void stream_session_disconnect (stream_session_t * s);
 void stream_session_cleanup (stream_session_t * s);
-int
-stream_session_start_listen (u32 server_index, ip46_address_t * ip, u16 port);
-void stream_session_stop_listen (u32 server_index);
 
 u8 *format_stream_session (u8 * s, va_list * args);
 
@@ -390,6 +378,71 @@ transport_proto_vft_t *session_get_transport_vft (u8 type);
 
 clib_error_t *vnet_session_enable_disable (vlib_main_t * vm, u8 is_en);
 
+always_inline unix_shared_memory_queue_t *
+session_manager_get_vpp_event_queue (u32 thread_index)
+{
+  return session_manager_main.vpp_event_queues[thread_index];
+}
+
+int session_manager_flush_enqueue_events (u32 thread_index);
+
+always_inline u64
+listen_session_get_handle (stream_session_t * s)
+{
+  ASSERT (s->session_state == SESSION_STATE_LISTENING);
+  return ((u64) s->session_type << 32) | s->session_index;
+}
+
+always_inline stream_session_t *
+listen_session_get_from_handle (u64 handle)
+{
+  session_manager_main_t *smm = &session_manager_main;
+  stream_session_t *s;
+  u32 type, index;
+  type = handle >> 32;
+  index = handle & 0xFFFFFFFF;
+
+  if (pool_is_free_index (smm->listen_sessions[type], index))
+    return 0;
+
+  s = pool_elt_at_index (smm->listen_sessions[type], index);
+  ASSERT (s->session_state == SESSION_STATE_LISTENING);
+  return s;
+}
+
+always_inline stream_session_t *
+listen_session_new (session_type_t type)
+{
+  stream_session_t *s;
+  pool_get (session_manager_main.listen_sessions[type], s);
+  memset (s, 0, sizeof (*s));
+
+  s->session_type = type;
+  s->session_state = SESSION_STATE_LISTENING;
+  s->session_index = s - session_manager_main.listen_sessions[type];
+
+  return s;
+}
+
+always_inline stream_session_t *
+listen_session_get (session_type_t type, u32 index)
+{
+  return pool_elt_at_index (session_manager_main.listen_sessions[type],
+			    index);
+}
+
+always_inline void
+listen_session_del (stream_session_t * s)
+{
+  pool_put (session_manager_main.listen_sessions[s->session_type], s);
+}
+
+always_inline u8
+session_manager_is_enabled ()
+{
+  return session_manager_main.is_enabled == 1;
+}
+
 #endif /* __included_session_h__ */
 
 /*
diff --git a/src/vnet/session/session_api.c b/src/vnet/session/session_api.c
index 9c38428a..a82dfe0b 100644
--- a/src/vnet/session/session_api.c
+++ b/src/vnet/session/session_api.c
@@ -38,6 +38,8 @@
 
 #define foreach_session_api_msg                                         \
 _(MAP_ANOTHER_SEGMENT_REPLY, map_another_segment_reply)                 \
+_(APPLICATION_ATTACH, application_attach)				\
+_(APPLICATION_DETACH, application_detach)				\
 _(BIND_URI, bind_uri)                                                   \
 _(UNBIND_URI, unbind_uri)                                               \
 _(CONNECT_URI, connect_uri)                                             \
@@ -48,13 +50,8 @@ _(RESET_SESSION_REPLY, reset_session_reply)                   		\
 _(BIND_SOCK, bind_sock) 		                                \
 _(UNBIND_SOCK, unbind_sock)                                             \
 _(CONNECT_SOCK, connect_sock)                                          	\
-_(DISCONNECT_SOCK, disconnect_sock)                               	\
-_(DISCONNECT_SOCK_REPLY, disconnect_sock_reply)                        	\
-_(ACCEPT_SOCK_REPLY, accept_sock_reply)                           	\
-_(RESET_SOCK_REPLY, reset_sock_reply)                   		\
 _(SESSION_ENABLE_DISABLE, session_enable_disable)                   	\
 
-
 static int
 send_add_segment_callback (u32 api_client_index, const u8 * segment_name,
 			   u32 segment_size)
@@ -80,11 +77,14 @@ send_add_segment_callback (u32 api_client_index, const u8 * segment_name,
 }
 
 static int
-send_session_accept_uri_callback (stream_session_t * s)
+send_session_accept_callback (stream_session_t * s)
 {
   vl_api_accept_session_t *mp;
   unix_shared_memory_queue_t *q, *vpp_queue;
   application_t *server = application_get (s->app_index);
+  transport_connection_t *tc;
+  transport_proto_vft_t *tp_vft;
+  stream_session_t *listener;
 
   q = vl_api_client_index_to_input_queue (server->api_client_index);
   vpp_queue = session_manager_get_vpp_event_queue (s->thread_index);
@@ -93,24 +93,28 @@ send_session_accept_uri_callback (stream_session_t * s)
     return -1;
 
   mp = vl_msg_api_alloc (sizeof (*mp));
-  mp->_vl_msg_id = clib_host_to_net_u16 (VL_API_ACCEPT_SESSION);
+  memset (mp, 0, sizeof (*mp));
 
-  /* Note: session_type is the first octet in all types of sessions */
+  mp->_vl_msg_id = clib_host_to_net_u16 (VL_API_ACCEPT_SESSION);
 
-  mp->accept_cookie = server->accept_cookie;
+  listener = listen_session_get (s->session_type, s->listener_index);
+  tp_vft = session_get_transport_vft (s->session_type);
+  tc = tp_vft->get_connection (s->connection_index, s->thread_index);
+  mp->listener_handle = listen_session_get_handle (listener);
+  mp->handle = stream_session_handle (s);
   mp->server_rx_fifo = (u64) s->server_rx_fifo;
   mp->server_tx_fifo = (u64) s->server_tx_fifo;
-  mp->session_thread_index = s->thread_index;
-  mp->session_index = s->session_index;
-  mp->session_type = s->session_type;
   mp->vpp_event_queue_address = (u64) vpp_queue;
+  mp->port = tc->rmt_port;
+  mp->is_ip4 = tc->is_ip4;
+  clib_memcpy (&mp->ip, &tc->rmt_ip, sizeof (tc->rmt_ip));
   vl_msg_api_send_shmem (q, (u8 *) & mp);
 
   return 0;
 }
 
 static void
-send_session_disconnect_uri_callback (stream_session_t * s)
+send_session_disconnect_callback (stream_session_t * s)
 {
   vl_api_disconnect_session_t *mp;
   unix_shared_memory_queue_t *q;
@@ -124,14 +128,12 @@ send_session_disconnect_uri_callback (stream_session_t * s)
   mp = vl_msg_api_alloc (sizeof (*mp));
   memset (mp, 0, sizeof (*mp));
   mp->_vl_msg_id = clib_host_to_net_u16 (VL_API_DISCONNECT_SESSION);
-
-  mp->session_thread_index = s->thread_index;
-  mp->session_index = s->session_index;
+  mp->handle = stream_session_handle (s);
   vl_msg_api_send_shmem (q, (u8 *) & mp);
 }
 
 static void
-send_session_reset_uri_callback (stream_session_t * s)
+send_session_reset_callback (stream_session_t * s)
 {
   vl_api_reset_session_t *mp;
   unix_shared_memory_queue_t *q;
@@ -145,22 +147,20 @@ send_session_reset_uri_callback (stream_session_t * s)
   mp = vl_msg_api_alloc (sizeof (*mp));
   memset (mp, 0, sizeof (*mp));
   mp->_vl_msg_id = clib_host_to_net_u16 (VL_API_RESET_SESSION);
-
-  mp->session_thread_index = s->thread_index;
-  mp->session_index = s->session_index;
+  mp->handle = stream_session_handle (s);
   vl_msg_api_send_shmem (q, (u8 *) & mp);
 }
 
 static int
-send_session_connected_uri_callback (u32 api_client_index,
-				     stream_session_t * s, u8 is_fail)
+send_session_connected_callback (u32 app_index, u32 api_context,
+				 stream_session_t * s, u8 is_fail)
 {
   vl_api_connect_uri_reply_t *mp;
   unix_shared_memory_queue_t *q;
-  application_t *app = application_lookup (api_client_index);
-  u8 *seg_name;
+  application_t *app;
   unix_shared_memory_queue_t *vpp_queue;
 
+  app = application_get (app_index);
   q = vl_api_client_index_to_input_queue (app->api_client_index);
 
   if (!q)
@@ -168,24 +168,15 @@ send_session_connected_uri_callback (u32 api_client_index,
 
   mp = vl_msg_api_alloc (sizeof (*mp));
   mp->_vl_msg_id = clib_host_to_net_u16 (VL_API_CONNECT_URI_REPLY);
-  mp->context = app->api_context;
+  mp->context = api_context;
   if (!is_fail)
     {
       vpp_queue = session_manager_get_vpp_event_queue (s->thread_index);
       mp->server_rx_fifo = (u64) s->server_rx_fifo;
       mp->server_tx_fifo = (u64) s->server_tx_fifo;
-      mp->session_thread_index = s->thread_index;
-      mp->session_index = s->session_index;
-      mp->session_type = s->session_type;
+      mp->handle = stream_session_handle (s);
       mp->vpp_event_queue_address = (u64) vpp_queue;
-      mp->client_event_queue_address = (u64) app->event_queue;
       mp->retval = 0;
-
-      session_manager_get_segment_info (s->server_segment_index, &seg_name,
-					&mp->segment_size);
-      mp->segment_name_length = vec_len (seg_name);
-      if (mp->segment_name_length)
-	clib_memcpy (mp->segment_name, seg_name, mp->segment_name_length);
     }
   else
     {
@@ -195,199 +186,14 @@ send_session_connected_uri_callback (u32 api_client_index,
   vl_msg_api_send_shmem (q, (u8 *) & mp);
 
   /* Remove client if connect failed */
-  if (is_fail)
-    {
-      application_del (app);
-    }
-  else
-    {
-      s->session_state = SESSION_STATE_READY;
-    }
-
-  return 0;
-}
-
-/**
- * Redirect a connect_uri message to the indicated server.
- * Only sent if the server has bound the related port with
- * URI_OPTIONS_FLAGS_USE_FIFO
- */
-static int
-redirect_connect_uri_callback (u32 server_api_client_index, void *mp_arg)
-{
-  vl_api_connect_uri_t *mp = mp_arg;
-  unix_shared_memory_queue_t *server_q, *client_q;
-  vlib_main_t *vm = vlib_get_main ();
-  f64 timeout = vlib_time_now (vm) + 0.5;
-  int rv = 0;
-
-  server_q = vl_api_client_index_to_input_queue (server_api_client_index);
-
-  if (!server_q)
-    {
-      rv = VNET_API_ERROR_INVALID_VALUE;
-      goto out;
-    }
-
-  client_q = vl_api_client_index_to_input_queue (mp->client_index);
-  if (!client_q)
-    {
-      rv = VNET_API_ERROR_INVALID_VALUE_2;
-      goto out;
-    }
-
-  /* Tell the server the client's API queue address, so it can reply */
-  mp->client_queue_address = (u64) client_q;
-
-  /*
-   * Bounce message handlers MUST NOT block the data-plane.
-   * Spin waiting for the queue lock, but
-   */
-
-  while (vlib_time_now (vm) < timeout)
-    {
-      rv =
-	unix_shared_memory_queue_add (server_q, (u8 *) & mp, 1 /*nowait */ );
-      switch (rv)
-	{
-	  /* correctly enqueued */
-	case 0:
-	  return VNET_CONNECT_REDIRECTED;
-
-	  /* continue spinning, wait for pthread_mutex_trylock to work */
-	case -1:
-	  continue;
-
-	  /* queue stuffed, drop the msg */
-	case -2:
-	  rv = VNET_API_ERROR_QUEUE_FULL;
-	  goto out;
-	}
-    }
-out:
-  /* Dispose of the message */
-  vl_msg_api_free (mp);
-  return rv;
-}
-
-static u64
-make_session_handle (stream_session_t * s)
-{
-  return (u64) s->session_index << 32 | (u64) s->thread_index;
-}
-
-static int
-send_session_accept_callback (stream_session_t * s)
-{
-  vl_api_accept_sock_t *mp;
-  unix_shared_memory_queue_t *q, *vpp_queue;
-  application_t *server = application_get (s->app_index);
-
-  q = vl_api_client_index_to_input_queue (server->api_client_index);
-  vpp_queue = session_manager_get_vpp_event_queue (s->thread_index);
-
-  if (!q)
-    return -1;
-
-  mp = vl_msg_api_alloc (sizeof (*mp));
-  mp->_vl_msg_id = clib_host_to_net_u16 (VL_API_ACCEPT_SOCK);
-
-  /* Note: session_type is the first octet in all types of sessions */
-
-  mp->accept_cookie = server->accept_cookie;
-  mp->server_rx_fifo = (u64) s->server_rx_fifo;
-  mp->server_tx_fifo = (u64) s->server_tx_fifo;
-  mp->handle = make_session_handle (s);
-  mp->vpp_event_queue_address = (u64) vpp_queue;
-  vl_msg_api_send_shmem (q, (u8 *) & mp);
-
-  return 0;
-}
-
-static int
-send_session_connected_callback (u32 api_client_index, stream_session_t * s,
-				 u8 is_fail)
-{
-  vl_api_connect_sock_reply_t *mp;
-  unix_shared_memory_queue_t *q;
-  application_t *app = application_lookup (api_client_index);
-  u8 *seg_name;
-  unix_shared_memory_queue_t *vpp_queue;
-
-  q = vl_api_client_index_to_input_queue (app->api_client_index);
-
-  if (!q)
-    return -1;
-
-  mp = vl_msg_api_alloc (sizeof (*mp));
-  mp->_vl_msg_id = clib_host_to_net_u16 (VL_API_CONNECT_SOCK_REPLY);
-  mp->context = app->api_context;
-  mp->retval = is_fail;
   if (!is_fail)
     {
-      vpp_queue = session_manager_get_vpp_event_queue (s->thread_index);
-      mp->server_rx_fifo = (u64) s->server_rx_fifo;
-      mp->server_tx_fifo = (u64) s->server_tx_fifo;
-      mp->handle = make_session_handle (s);
-      mp->vpp_event_queue_address = (u64) vpp_queue;
-      mp->client_event_queue_address = (u64) app->event_queue;
-
-      session_manager_get_segment_info (s->server_segment_index, &seg_name,
-					&mp->segment_size);
-      mp->segment_name_length = vec_len (seg_name);
-      if (mp->segment_name_length)
-	clib_memcpy (mp->segment_name, seg_name, mp->segment_name_length);
+      s->session_state = SESSION_STATE_READY;
     }
 
-  vl_msg_api_send_shmem (q, (u8 *) & mp);
-
-  /* Remove client if connect failed */
-  if (is_fail)
-    application_del (app);
-
   return 0;
 }
 
-static void
-send_session_disconnect_callback (stream_session_t * s)
-{
-  vl_api_disconnect_sock_t *mp;
-  unix_shared_memory_queue_t *q;
-  application_t *app = application_get (s->app_index);
-
-  q = vl_api_client_index_to_input_queue (app->api_client_index);
-
-  if (!q)
-    return;
-
-  mp = vl_msg_api_alloc (sizeof (*mp));
-  memset (mp, 0, sizeof (*mp));
-  mp->_vl_msg_id = clib_host_to_net_u16 (VL_API_DISCONNECT_SOCK);
-
-  mp->handle = make_session_handle (s);
-  vl_msg_api_send_shmem (q, (u8 *) & mp);
-}
-
-static void
-send_session_reset_callback (stream_session_t * s)
-{
-  vl_api_reset_sock_t *mp;
-  unix_shared_memory_queue_t *q;
-  application_t *app = application_get (s->app_index);
-
-  q = vl_api_client_index_to_input_queue (app->api_client_index);
-
-  if (!q)
-    return;
-
-  mp = vl_msg_api_alloc (sizeof (*mp));
-  memset (mp, 0, sizeof (*mp));
-  mp->_vl_msg_id = clib_host_to_net_u16 (VL_API_RESET_SOCK);
-
-  mp->handle = make_session_handle (s);
-  vl_msg_api_send_shmem (q, (u8 *) & mp);
-}
-
 /**
  * Redirect a connect_uri message to the indicated server.
  * Only sent if the server has bound the related port with
@@ -396,10 +202,11 @@ send_session_reset_callback (stream_session_t * s)
 static int
 redirect_connect_callback (u32 server_api_client_index, void *mp_arg)
 {
-  vl_api_connect_sock_t *mp = mp_arg;
+  vl_api_connect_uri_t *mp = mp_arg;
   unix_shared_memory_queue_t *server_q, *client_q;
   vlib_main_t *vm = vlib_get_main ();
   f64 timeout = vlib_time_now (vm) + 0.5;
+  application_t *app;
   int rv = 0;
 
   server_q = vl_api_client_index_to_input_queue (server_api_client_index);
@@ -419,6 +226,9 @@ redirect_connect_callback (u32 server_api_client_index, void *mp_arg)
 
   /* Tell the server the client's API queue address, so it can reply */
   mp->client_queue_address = (u64) client_q;
+  app = application_lookup (mp->client_index);
+  mp->options[SESSION_OPTIONS_RX_FIFO_SIZE] = app->sm_properties.rx_fifo_size;
+  mp->options[SESSION_OPTIONS_TX_FIFO_SIZE] = app->sm_properties.tx_fifo_size;
 
   /*
    * Bounce message handlers MUST NOT block the data-plane.
@@ -452,15 +262,6 @@ out:
 }
 
 static session_cb_vft_t uri_session_cb_vft = {
-  .session_accept_callback = send_session_accept_uri_callback,
-  .session_disconnect_callback = send_session_disconnect_uri_callback,
-  .session_connected_callback = send_session_connected_uri_callback,
-  .session_reset_callback = send_session_reset_uri_callback,
-  .add_segment_callback = send_add_segment_callback,
-  .redirect_connect_callback = redirect_connect_uri_callback
-};
-
-static session_cb_vft_t session_cb_vft = {
   .session_accept_callback = send_session_accept_callback,
   .session_disconnect_callback = send_session_disconnect_callback,
   .session_connected_callback = send_session_connected_callback,
@@ -498,60 +299,134 @@ vl_api_session_enable_disable_t_handler (vl_api_session_enable_disable_t * mp)
 }
 
 static void
-vl_api_bind_uri_t_handler (vl_api_bind_uri_t * mp)
+vl_api_application_attach_t_handler (vl_api_application_attach_t * mp)
 {
-  vl_api_bind_uri_reply_t *rmp;
-  vnet_bind_args_t _a, *a = &_a;
-  char segment_name[128];
-  u32 segment_name_length;
+  vl_api_application_attach_reply_t *rmp;
+  vnet_app_attach_args_t _a, *a = &_a;
   int rv;
 
-  _Static_assert (sizeof (u64) * SESSION_OPTIONS_N_OPTIONS <=
-		  sizeof (mp->options),
-		  "Out of options, fix api message definition");
+  if (session_manager_is_enabled () == 0)
+    {
+      rv = VNET_API_ERROR_FEATURE_DISABLED;
+      goto done;
+    }
 
-  segment_name_length = ARRAY_LEN (segment_name);
+  STATIC_ASSERT (sizeof (u64) * SESSION_OPTIONS_N_OPTIONS <=
+		 sizeof (mp->options),
+		 "Out of options, fix api message definition");
 
   memset (a, 0, sizeof (*a));
 
-  a->uri = (char *) mp->uri;
   a->api_client_index = mp->client_index;
   a->options = mp->options;
-  a->segment_name = segment_name;
-  a->segment_name_length = segment_name_length;
   a->session_cb_vft = &uri_session_cb_vft;
 
-  a->options[SESSION_OPTIONS_SEGMENT_SIZE] = mp->initial_segment_size;
-  a->options[SESSION_OPTIONS_ACCEPT_COOKIE] = mp->accept_cookie;
-  rv = vnet_bind_uri (a);
+  rv = vnet_application_attach (a);
 
+done:
   /* *INDENT-OFF* */
-  REPLY_MACRO2 (VL_API_BIND_URI_REPLY, ({
+  REPLY_MACRO2 (VL_API_APPLICATION_ATTACH_REPLY, ({
     rmp->retval = rv;
     if (!rv)
       {
 	rmp->segment_name_length = 0;
 	/* $$$$ policy? */
-	rmp->segment_size = mp->initial_segment_size;
-	if (segment_name_length)
+	rmp->segment_size = a->segment_size;
+	if (a->segment_name_length)
 	  {
-	    memcpy (rmp->segment_name, segment_name, segment_name_length);
-	    rmp->segment_name_length = segment_name_length;
+	    memcpy (rmp->segment_name, a->segment_name,
+		    a->segment_name_length);
+	    rmp->segment_name_length = a->segment_name_length;
 	  }
-	rmp->server_event_queue_address = a->server_event_queue_address;
+	rmp->app_event_queue_address = a->app_event_queue_address;
       }
   }));
   /* *INDENT-ON* */
 }
 
+static void
+vl_api_application_detach_t_handler (vl_api_application_detach_t * mp)
+{
+  vl_api_application_detach_reply_t *rmp;
+  int rv = VNET_API_ERROR_INVALID_VALUE_2;
+  vnet_app_detach_args_t _a, *a = &_a;
+  application_t *app;
+
+  if (session_manager_is_enabled () == 0)
+    {
+      rv = VNET_API_ERROR_FEATURE_DISABLED;
+      goto done;
+    }
+
+  app = application_lookup (mp->client_index);
+  if (app)
+    {
+      a->app_index = app->index;
+      rv = vnet_application_detach (a);
+    }
+
+done:
+  REPLY_MACRO (VL_API_APPLICATION_DETACH_REPLY);
+}
+
+static void
+vl_api_bind_uri_t_handler (vl_api_bind_uri_t * mp)
+{
+  vl_api_bind_uri_reply_t *rmp;
+  vnet_bind_args_t _a, *a = &_a;
+  application_t *app;
+  int rv;
+
+  if (session_manager_is_enabled () == 0)
+    {
+      rv = VNET_API_ERROR_FEATURE_DISABLED;
+      goto done;
+    }
+
+  app = application_lookup (mp->client_index);
+  if (app)
+    {
+      memset (a, 0, sizeof (*a));
+      a->uri = (char *) mp->uri;
+      a->app_index = app->index;
+      rv = vnet_bind_uri (a);
+    }
+  else
+    {
+      rv = VNET_API_ERROR_APPLICATION_NOT_ATTACHED;
+    }
+
+done:
+  REPLY_MACRO (VL_API_BIND_URI_REPLY);
+}
+
 static void
 vl_api_unbind_uri_t_handler (vl_api_unbind_uri_t * mp)
 {
   vl_api_unbind_uri_reply_t *rmp;
+  application_t *app;
+  vnet_unbind_args_t _a, *a = &_a;
   int rv;
 
-  rv = vnet_unbind_uri ((char *) mp->uri, mp->client_index);
+  if (session_manager_is_enabled () == 0)
+    {
+      rv = VNET_API_ERROR_FEATURE_DISABLED;
+      goto done;
+    }
+
+  app = application_lookup (mp->client_index);
+  if (app)
+    {
+      a->uri = (char *) mp->uri;
+      a->app_index = app->index;
+      rv = vnet_unbind_uri (a);
+    }
+  else
+    {
+      rv = VNET_API_ERROR_APPLICATION_NOT_ATTACHED;
+    }
 
+done:
   REPLY_MACRO (VL_API_UNBIND_URI_REPLY);
 }
 
@@ -560,26 +435,37 @@ vl_api_connect_uri_t_handler (vl_api_connect_uri_t * mp)
 {
   vl_api_connect_uri_reply_t *rmp;
   vnet_connect_args_t _a, *a = &_a;
+  application_t *app;
   int rv;
 
-  a->uri = (char *) mp->uri;
-  a->api_client_index = mp->client_index;
-  a->api_context = mp->context;
-  a->options = mp->options;
-  a->session_cb_vft = &uri_session_cb_vft;
-  a->mp = mp;
+  if (session_manager_is_enabled () == 0)
+    {
+      rv = VNET_API_ERROR_FEATURE_DISABLED;
+      goto done;
+    }
 
-  rv = vnet_connect_uri (a);
+  app = application_lookup (mp->client_index);
+  if (app)
+    {
+      a->uri = (char *) mp->uri;
+      a->api_context = mp->context;
+      a->app_index = app->index;
+      a->mp = mp;
+      rv = vnet_connect_uri (a);
+    }
+  else
+    {
+      rv = VNET_API_ERROR_APPLICATION_NOT_ATTACHED;
+    }
 
   if (rv == 0 || rv == VNET_CONNECT_REDIRECTED)
     return;
 
   /* Got some error, relay it */
 
+done:
   /* *INDENT-OFF* */
-  REPLY_MACRO2 (VL_API_CONNECT_URI_REPLY, ({
-    rmp->retval = rv;
-  }));
+  REPLY_MACRO (VL_API_CONNECT_URI_REPLY);
   /* *INDENT-ON* */
 }
 
@@ -587,13 +473,29 @@ static void
 vl_api_disconnect_session_t_handler (vl_api_disconnect_session_t * mp)
 {
   vl_api_disconnect_session_reply_t *rmp;
-  int rv;
+  vnet_disconnect_args_t _a, *a = &_a;
+  application_t *app;
+  int rv = 0;
 
-  rv = api_session_not_valid (mp->session_index, mp->session_thread_index);
-  if (!rv)
-    rv =
-      vnet_disconnect_session (mp->session_index, mp->session_thread_index);
+  if (session_manager_is_enabled () == 0)
+    {
+      rv = VNET_API_ERROR_FEATURE_DISABLED;
+      goto done;
+    }
+
+  app = application_lookup (mp->client_index);
+  if (app)
+    {
+      a->handle = mp->handle;
+      a->app_index = app->index;
+      rv = vnet_disconnect_session (a);
+    }
+  else
+    {
+      rv = VNET_API_ERROR_APPLICATION_NOT_ATTACHED;
+    }
 
+done:
   REPLY_MACRO (VL_API_DISCONNECT_SESSION_REPLY);
 }
 
@@ -601,11 +503,8 @@ static void
 vl_api_disconnect_session_reply_t_handler (vl_api_disconnect_session_reply_t *
 					   mp)
 {
-  if (api_session_not_valid (mp->session_index, mp->session_thread_index))
-    {
-      clib_warning ("Invalid session!");
-      return;
-    }
+  vnet_disconnect_args_t _a, *a = &_a;
+  application_t *app;
 
   /* Client objected to disconnecting the session, log and continue */
   if (mp->retval)
@@ -615,15 +514,29 @@ vl_api_disconnect_session_reply_t_handler (vl_api_disconnect_session_reply_t *
     }
 
   /* Disconnect has been confirmed. Confirm close to transport */
-  vnet_disconnect_session (mp->session_index, mp->session_thread_index);
+  app = application_lookup (mp->client_index);
+  if (app)
+    {
+      a->handle = mp->handle;
+      a->app_index = app->index;
+      vnet_disconnect_session (a);
+    }
 }
 
 static void
 vl_api_reset_session_reply_t_handler (vl_api_reset_session_reply_t * mp)
 {
+  application_t *app;
   stream_session_t *s;
+  u32 index, thread_index;
+
+  app = application_lookup (mp->client_index);
+  if (!app)
+    return;
 
-  if (api_session_not_valid (mp->session_index, mp->session_thread_index))
+  stream_session_parse_handle (mp->handle, &index, &thread_index);
+  s = stream_session_get_if_valid (index, thread_index);
+  if (s == 0 || app->index != s->app_index)
     {
       clib_warning ("Invalid session!");
       return;
@@ -636,8 +549,6 @@ vl_api_reset_session_reply_t_handler (vl_api_reset_session_reply_t * mp)
       return;
     }
 
-  s = stream_session_get (mp->session_index, mp->session_thread_index);
-
   /* This comes as a response to a reset, transport only waiting for
    * confirmation to remove connection state, no need to disconnect */
   stream_session_cleanup (s);
@@ -648,11 +559,13 @@ vl_api_accept_session_reply_t_handler (vl_api_accept_session_reply_t * mp)
 {
   stream_session_t *s;
   int rv;
-
-  if (api_session_not_valid (mp->session_index, mp->session_thread_index))
+  u32 session_index, thread_index;
+  session_index = stream_session_index_from_handle (mp->handle);
+  thread_index = stream_session_thread_from_handle (mp->handle);
+  if (api_session_not_valid (session_index, thread_index))
     return;
 
-  s = stream_session_get (mp->session_index, mp->session_thread_index);
+  s = stream_session_get (session_index, thread_index);
   rv = mp->retval;
 
   if (rv)
@@ -677,49 +590,31 @@ vl_api_bind_sock_t_handler (vl_api_bind_sock_t * mp)
 {
   vl_api_bind_sock_reply_t *rmp;
   vnet_bind_args_t _a, *a = &_a;
-  char segment_name[128];
-  u32 segment_name_length;
-  int rv;
-
-  STATIC_ASSERT (sizeof (u64) * SESSION_OPTIONS_N_OPTIONS <=
-		 sizeof (mp->options),
-		 "Out of options, fix api message definition");
-
-  segment_name_length = ARRAY_LEN (segment_name);
-
-  memset (a, 0, sizeof (*a));
-
-  clib_memcpy (&a->tep.ip, mp->ip,
-	       (mp->is_ip4 ? sizeof (ip4_address_t) :
-		sizeof (ip6_address_t)));
-  a->tep.is_ip4 = mp->is_ip4;
-  a->tep.port = mp->port;
-  a->tep.vrf = mp->vrf;
-
-  a->api_client_index = mp->client_index;
-  a->options = mp->options;
-  a->segment_name = segment_name;
-  a->segment_name_length = segment_name_length;
-  a->session_cb_vft = &session_cb_vft;
+  int rv = VNET_API_ERROR_APPLICATION_NOT_ATTACHED;
+  application_t *app;
 
-  rv = vnet_bind_uri (a);
+  if (session_manager_is_enabled () == 0)
+    {
+      rv = VNET_API_ERROR_FEATURE_DISABLED;
+      goto done;
+    }
 
-  /* *INDENT-OFF* */
-  REPLY_MACRO2 (VL_API_BIND_SOCK_REPLY, ({
-    rmp->retval = rv;
-    if (!rv)
-      {
-	rmp->segment_name_length = 0;
-	rmp->segment_size = mp->options[SESSION_OPTIONS_SEGMENT_SIZE];
-	if (segment_name_length)
-	  {
-	    memcpy(rmp->segment_name, segment_name, segment_name_length);
-	    rmp->segment_name_length = segment_name_length;
-	  }
-	rmp->server_event_queue_address = a->server_event_queue_address;
-      }
-  }));
-  /* *INDENT-ON* */
+  app = application_lookup (mp->client_index);
+  if (app)
+    {
+      memset (a, 0, sizeof (*a));
+      clib_memcpy (&a->tep.ip, mp->ip, (mp->is_ip4 ?
+					sizeof (ip4_address_t) :
+					sizeof (ip6_address_t)));
+      a->tep.is_ip4 = mp->is_ip4;
+      a->tep.port = mp->port;
+      a->tep.vrf = mp->vrf;
+      a->app_index = app->index;
+
+      rv = vnet_bind (a);
+    }
+done:
+  REPLY_MACRO (VL_API_BIND_SOCK_REPLY);
 }
 
 static void
@@ -727,13 +622,24 @@ vl_api_unbind_sock_t_handler (vl_api_unbind_sock_t * mp)
 {
   vl_api_unbind_sock_reply_t *rmp;
   vnet_unbind_args_t _a, *a = &_a;
-  int rv;
+  application_t *app;
+  int rv = VNET_API_ERROR_APPLICATION_NOT_ATTACHED;
 
-  a->api_client_index = mp->client_index;
-  a->handle = mp->handle;
+  if (session_manager_is_enabled () == 0)
+    {
+      rv = VNET_API_ERROR_FEATURE_DISABLED;
+      goto done;
+    }
 
-  rv = vnet_unbind (a);
+  app = application_lookup (mp->client_index);
+  if (app)
+    {
+      a->app_index = mp->client_index;
+      a->handle = mp->handle;
+      rv = vnet_unbind (a);
+    }
 
+done:
   REPLY_MACRO (VL_API_UNBIND_SOCK_REPLY);
 }
 
@@ -742,114 +648,55 @@ vl_api_connect_sock_t_handler (vl_api_connect_sock_t * mp)
 {
   vl_api_connect_sock_reply_t *rmp;
   vnet_connect_args_t _a, *a = &_a;
+  application_t *app;
   int rv;
 
-  clib_memcpy (&a->tep.ip, mp->ip,
-	       (mp->is_ip4 ? sizeof (ip4_address_t) :
-		sizeof (ip6_address_t)));
-  a->tep.is_ip4 = mp->is_ip4;
-  a->tep.port = mp->port;
-  a->tep.vrf = mp->vrf;
-  a->options = mp->options;
-  a->session_cb_vft = &session_cb_vft;
-  a->api_context = mp->context;
-  a->mp = mp;
-
-  rv = vnet_connect (a);
-
-  if (rv == 0 || rv == VNET_CONNECT_REDIRECTED)
-    return;
-
-  /* Got some error, relay it */
-
-  /* *INDENT-OFF* */
-  REPLY_MACRO2 (VL_API_CONNECT_URI_REPLY, ({
-    rmp->retval = rv;
-  }));
-  /* *INDENT-ON* */
-}
-
-static void
-vl_api_disconnect_sock_t_handler (vl_api_disconnect_sock_t * mp)
-{
-  vnet_disconnect_args_t _a, *a = &_a;
-  vl_api_disconnect_sock_reply_t *rmp;
-  int rv;
-
-  a->api_client_index = mp->client_index;
-  a->handle = mp->handle;
-  rv = vnet_disconnect (a);
-
-  REPLY_MACRO (VL_API_DISCONNECT_SOCK_REPLY);
-}
-
-static void
-vl_api_disconnect_sock_reply_t_handler (vl_api_disconnect_sock_reply_t * mp)
-{
-  vnet_disconnect_args_t _a, *a = &_a;
-
-  /* Client objected to disconnecting the session, log and continue */
-  if (mp->retval)
+  if (session_manager_is_enabled () == 0)
     {
-      clib_warning ("client retval %d", mp->retval);
-      return;
+      rv = VNET_API_ERROR_FEATURE_DISABLED;
+      goto done;
     }
 
-  a->api_client_index = mp->client_index;
-  a->handle = mp->handle;
-
-  vnet_disconnect (a);
-}
-
-static void
-vl_api_reset_sock_reply_t_handler (vl_api_reset_sock_reply_t * mp)
-{
-  stream_session_t *s;
-  u32 session_index, thread_index;
-
-  /* Client objected to resetting the session, log and continue */
-  if (mp->retval)
+  app = application_lookup (mp->client_index);
+  if (app)
     {
-      clib_warning ("client retval %d", mp->retval);
-      return;
+      clib_memcpy (&a->tep.ip, mp->ip,
+		   (mp->is_ip4 ? sizeof (ip4_address_t) :
+		    sizeof (ip6_address_t)));
+      a->api_context = mp->context;
+      a->app_index = app->index;
+      a->mp = mp;
+      rv = vnet_connect (a);
     }
-
-  if (api_parse_session_handle (mp->handle, &session_index, &thread_index))
+  else
     {
-      clib_warning ("Invalid handle");
-      return;
+      rv = VNET_API_ERROR_APPLICATION_NOT_ATTACHED;
     }
 
-  s = stream_session_get (session_index, thread_index);
+  if (rv == 0 || rv == VNET_CONNECT_REDIRECTED)
+    return;
 
-  /* This comes as a response to a reset, transport only waiting for
-   * confirmation to remove connection state, no need to disconnect */
-  stream_session_cleanup (s);
+  /* Got some error, relay it */
+
+done:
+  REPLY_MACRO (VL_API_CONNECT_URI_REPLY);
 }
 
-static void
-vl_api_accept_sock_reply_t_handler (vl_api_accept_sock_reply_t * mp)
+static clib_error_t *
+application_reaper_cb (u32 client_index)
 {
-  stream_session_t *s;
-  u32 session_index, thread_index;
-
-  if (api_parse_session_handle (mp->handle, &session_index, &thread_index))
-    {
-      clib_warning ("Invalid handle");
-      return;
-    }
-  s = stream_session_get (session_index, thread_index);
-
-  if (mp->retval)
+  application_t *app = application_lookup (client_index);
+  vnet_app_detach_args_t _a, *a = &_a;
+  if (app)
     {
-      /* Server isn't interested, kill the session */
-      stream_session_disconnect (s);
-      return;
+      a->app_index = app->index;
+      vnet_application_detach (a);
     }
-
-  s->session_state = SESSION_STATE_READY;
+  return 0;
 }
 
+VL_MSG_API_REAPER_FUNCTION (application_reaper_cb);
+
 #define vl_msg_name_crc_list
 #include <vnet/vnet_all_api_h.h>
 #undef vl_msg_name_crc_list
@@ -903,6 +750,7 @@ session_api_hookup (vlib_main_t * vm)
 }
 
 VLIB_API_INIT_FUNCTION (session_api_hookup);
+
 /*
  * fd.io coding-style-patch-verification: ON
  *
diff --git a/src/vnet/session/transport.h b/src/vnet/session/transport.h
index 2f912cbc..7ea7af15 100644
--- a/src/vnet/session/transport.h
+++ b/src/vnet/session/transport.h
@@ -30,7 +30,7 @@ typedef struct _transport_connection
   ip46_address_t lcl_ip;	/**< Local IP */
   u16 lcl_port;			/**< Local port */
   u16 rmt_port;			/**< Remote port */
-  u8 proto;			/**< Transport protocol id */
+  u8 proto;			/**< Transport protocol id (also session type) */
 
   u32 s_index;			/**< Parent session index */
   u32 c_index;			/**< Connection index in transport pool */
@@ -103,7 +103,8 @@ typedef CLIB_PACKED (struct {
     {
       struct
 	{
-	  ip4_address_t src; ip4_address_t dst;
+	  ip4_address_t src;
+	  ip4_address_t dst;
 	  u16 src_port;
 	  u16 dst_port;
 	  /* align by making this 4 octets even though its a 1-bit field
@@ -122,10 +123,14 @@ typedef CLIB_PACKED (struct {
       struct
 	{
 	  /* 48 octets */
-	  ip6_address_t src; ip6_address_t dst;
+	  ip6_address_t src;
+	  ip6_address_t dst;
 	  u16 src_port;
-	  u16 dst_port; u32 proto; u8 unused_for_now[8];
-	}; u64 as_u64[6];
+	  u16 dst_port;
+	  u32 proto;
+	  u8 unused_for_now[8];
+	};
+      u64 as_u64[6];
     };
 }) v6_connection_key_t;
 /* *INDENT-ON* */
@@ -233,10 +238,10 @@ make_v6_ss_kv_from_tc (session_kv6_t * kv, transport_connection_t * t)
 
 typedef struct _transport_endpoint
 {
-  ip46_address_t ip;
-  u16 port;
-  u8 is_ip4;
-  u32 vrf;
+  ip46_address_t ip;	/** ip address */
+  u16 port;		/** port in host order */
+  u8 is_ip4;		/** 1 if ip4 */
+  u32 vrf;		/** fib table the endpoint is associated with */
 } transport_endpoint_t;
 
 typedef clib_bihash_24_8_t transport_endpoint_table_t;
diff --git a/src/vnet/tcp/builtin_client.c b/src/vnet/tcp/builtin_client.c
index 9e8e1561..f8fbf28c 100644
--- a/src/vnet/tcp/builtin_client.c
+++ b/src/vnet/tcp/builtin_client.c
@@ -237,8 +237,7 @@ tclient_thread_fn (void *arg)
 	  memset (dmp, 0, sizeof (*dmp));
 	  dmp->_vl_msg_id = ntohs (VL_API_DISCONNECT_SESSION);
 	  dmp->client_index = tm->my_client_index;
-	  dmp->session_index = sp->vpp_session_index;
-	  dmp->session_thread_index = sp->vpp_session_thread;
+	  dmp->handle = sp->vpp_session_handle;
 	  vl_msg_api_send_shmem (tm->vl_input_queue, (u8 *) & dmp);
 	  pool_put (tm->sessions, sp);
 	}
@@ -253,9 +252,10 @@ tclient_thread_fn (void *arg)
 static void
 vl_api_memclnt_create_reply_t_handler (vl_api_memclnt_create_reply_t * mp)
 {
+  vlib_main_t *vm = vlib_get_main ();
   tclient_main_t *tm = &tclient_main;
-
   tm->my_client_index = mp->index;
+  vlib_process_signal_event (vm, tm->node_index, 1 /* evt */ , 0 /* data */ );
 }
 
 static void
@@ -264,7 +264,6 @@ vl_api_connect_uri_reply_t_handler (vl_api_connect_uri_reply_t * mp)
   tclient_main_t *tm = &tclient_main;
   session_t *session;
   u32 session_index;
-  u64 key;
   i32 retval = /* clib_net_to_host_u32 ( */ mp->retval /*) */ ;
 
   if (retval < 0)
@@ -291,24 +290,24 @@ vl_api_connect_uri_reply_t_handler (vl_api_connect_uri_reply_t * mp)
   session->server_rx_fifo->client_session_index = session_index;
   session->server_tx_fifo = (svm_fifo_t *) mp->server_tx_fifo;
   session->server_tx_fifo->client_session_index = session_index;
-
-  session->vpp_session_index = mp->session_index;
-  session->vpp_session_thread = mp->session_thread_index;
+  session->vpp_session_handle = mp->handle;
 
   /* Add it to the session lookup table */
-  key = (((u64) mp->session_thread_index) << 32) | (u64) mp->session_index;
-  hash_set (tm->session_index_by_vpp_handles, key, session_index);
+  hash_set (tm->session_index_by_vpp_handles, mp->handle, session_index);
 
   tm->ready_connections++;
 }
 
-static void
+static int
 create_api_loopback (tclient_main_t * tm)
 {
+  vlib_main_t *vm = vlib_get_main ();
   vl_api_memclnt_create_t _m, *mp = &_m;
   extern void vl_api_memclnt_create_t_handler (vl_api_memclnt_create_t *);
   api_main_t *am = &api_main;
   vl_shmem_hdr_t *shmem_hdr;
+  uword *event_data = 0, event_type;
+  int resolved = 0;
 
   /*
    * Create a "loopback" API client connection
@@ -324,6 +323,25 @@ create_api_loopback (tclient_main_t * tm)
   strncpy ((char *) mp->name, "tcp_tester", sizeof (mp->name) - 1);
 
   vl_api_memclnt_create_t_handler (mp);
+
+  /* Wait for reply */
+  tm->node_index = vlib_get_current_process (vm)->node_runtime.node_index;
+  vlib_process_wait_for_event_or_clock (vm, 1.0);
+  event_type = vlib_process_get_events (vm, &event_data);
+  switch (event_type)
+    {
+    case 1:
+      resolved = 1;
+      break;
+    case ~0:
+      /* timed out */
+      break;
+    default:
+      clib_warning ("unknown event_type %d", event_type);
+    }
+  if (!resolved)
+    return -1;
+  return 0;
 }
 
 #define foreach_tclient_static_api_msg       	\
@@ -333,17 +351,7 @@ _(CONNECT_URI_REPLY, connect_uri_reply)
 static clib_error_t *
 tclient_api_hookup (vlib_main_t * vm)
 {
-  tclient_main_t *tm = &tclient_main;
   vl_msg_api_msg_config_t _c, *c = &_c;
-  int i;
-
-  /* Init test data */
-  vec_validate (tm->connect_test_data, 64 * 1024 - 1);
-  for (i = 0; i < vec_len (tm->connect_test_data); i++)
-    tm->connect_test_data[i] = i & 0xff;
-
-  tm->session_index_by_vpp_handles = hash_create (0, sizeof (uword));
-  vec_validate (tm->rx_buf, vec_len (tm->connect_test_data) - 1);
 
   /* Hook up client-side static APIs to our handlers */
 #define _(N,n) do {                                             \
@@ -365,18 +373,105 @@ tclient_api_hookup (vlib_main_t * vm)
   return 0;
 }
 
-VLIB_API_INIT_FUNCTION (tclient_api_hookup);
+static int
+tcp_test_clients_init (vlib_main_t * vm)
+{
+  tclient_main_t *tm = &tclient_main;
+  int i;
+
+  tclient_api_hookup (vm);
+  if (create_api_loopback (tm))
+    return -1;
+
+  /* Init test data */
+  vec_validate (tm->connect_test_data, 64 * 1024 - 1);
+  for (i = 0; i < vec_len (tm->connect_test_data); i++)
+    tm->connect_test_data[i] = i & 0xff;
+
+  tm->session_index_by_vpp_handles = hash_create (0, sizeof (uword));
+  vec_validate (tm->rx_buf, vec_len (tm->connect_test_data) - 1);
+
+  tm->is_init = 1;
+
+  return 0;
+}
+
+static void
+builtin_session_reset_callback (stream_session_t * s)
+{
+  return;
+}
+
+static int
+builtin_session_connected_callback (u32 app_index, u32 api_context,
+				    stream_session_t * s, u8 code)
+{
+  return 0;
+}
+
+static int
+builtin_session_create_callback (stream_session_t * s)
+{
+  return 0;
+}
+
+static void
+builtin_session_disconnect_callback (stream_session_t * s)
+{
+  return;
+}
+
+static int
+builtin_server_rx_callback (stream_session_t * s)
+{
+  return 0;
+}
+
+/* *INDENT-OFF* */
+static session_cb_vft_t builtin_clients = {
+    .session_reset_callback = builtin_session_reset_callback,
+    .session_connected_callback = builtin_session_connected_callback,
+    .session_accept_callback = builtin_session_create_callback,
+    .session_disconnect_callback = builtin_session_disconnect_callback,
+    .builtin_server_rx_callback = builtin_server_rx_callback
+};
+/* *INDENT-ON* */
+
+static int
+attach_builtin_test_clients ()
+{
+  vnet_app_attach_args_t _a, *a = &_a;
+  u8 segment_name[128];
+  u32 segment_name_length;
+  u64 options[16];
+
+  segment_name_length = ARRAY_LEN (segment_name);
+
+  memset (a, 0, sizeof (*a));
+  memset (options, 0, sizeof (options));
+
+  a->api_client_index = ~0;
+  a->segment_name = segment_name;
+  a->segment_name_length = segment_name_length;
+  a->session_cb_vft = &builtin_clients;
+
+  options[SESSION_OPTIONS_ACCEPT_COOKIE] = 0x12345678;
+  options[SESSION_OPTIONS_SEGMENT_SIZE] = (2 << 30);	/*$$$$ config / arg */
+  a->options = options;
+
+  return vnet_application_attach (a);
+}
 
 static clib_error_t *
 test_tcp_clients_command_fn (vlib_main_t * vm,
 			     unformat_input_t * input,
 			     vlib_cli_command_t * cmd)
 {
+  tclient_main_t *tm = &tclient_main;
   u8 *connect_uri = (u8 *) "tcp://6.0.1.1/1234";
   u8 *uri;
-  tclient_main_t *tm = &tclient_main;
-  int i;
   u32 n_clients = 1;
+  int i;
 
   tm->bytes_to_send = 8192;
   tm->n_iterations = 1;
@@ -397,14 +492,19 @@ test_tcp_clients_command_fn (vlib_main_t * vm,
 				  format_unformat_error, input);
     }
 
+  if (tm->is_init == 0)
+    {
+      if (tcp_test_clients_init (vm))
+	return clib_error_return (0, "failed init");
+    }
+
   tm->ready_connections = 0;
   tm->expected_connections = n_clients;
+
   uri = connect_uri;
   if (tm->connect_uri)
     uri = tm->connect_uri;
 
-  create_api_loopback (tm);
-
 #if TCP_BUILTIN_CLIENT_PTHREAD
   /* Start a transmit thread */
   if (tm->client_thread_handle == 0)
@@ -420,6 +520,7 @@ test_tcp_clients_command_fn (vlib_main_t * vm,
     }
 #endif
   vnet_session_enable_disable (vm, 1 /* turn on TCP, etc. */ );
+  attach_builtin_test_clients ();
 
   /* Fire off connect requests, in something approaching a normal manner */
   for (i = 0; i < n_clients; i++)
@@ -461,6 +562,16 @@ VLIB_CLI_COMMAND (test_clients_command, static) =
 };
 /* *INDENT-ON* */
 
+clib_error_t *
+tcp_test_clients_main_init (vlib_main_t * vm)
+{
+  tclient_main_t *tm = &tclient_main;
+  tm->is_init = 0;
+  return 0;
+}
+
+VLIB_INIT_FUNCTION (tcp_test_clients_main_init);
+
 /*
  * fd.io coding-style-patch-verification: ON
  *
diff --git a/src/vnet/tcp/builtin_client.h b/src/vnet/tcp/builtin_client.h
index 64030302..2bd87c07 100644
--- a/src/vnet/tcp/builtin_client.h
+++ b/src/vnet/tcp/builtin_client.h
@@ -39,8 +39,7 @@ typedef struct
   svm_fifo_t *server_rx_fifo;
   svm_fifo_t *server_tx_fifo;
 
-  u32 vpp_session_index;
-  u32 vpp_session_thread;
+  u64 vpp_session_handle;
 } session_t;
 
 typedef struct
@@ -110,6 +109,10 @@ typedef struct
   u32 client_bytes_received;
   u8 test_return_packets;
 
+  u8 is_init;
+
+  u32 node_index;
+
   /* convenience */
   vlib_main_t *vlib_main;
   vnet_main_t *vnet_main;
diff --git a/src/vnet/tcp/builtin_server.c b/src/vnet/tcp/builtin_server.c
index 917d4bd3..8308e3d9 100644
--- a/src/vnet/tcp/builtin_server.c
+++ b/src/vnet/tcp/builtin_server.c
@@ -18,17 +18,46 @@
 #include <vnet/session/application.h>
 #include <vnet/session/application_interface.h>
 
+/* define message IDs */
+#include <vpp/api/vpe_msg_enum.h>
+
+/* define message structures */
+#define vl_typedefs
+#include <vpp/api/vpe_all_api_h.h>
+#undef vl_typedefs
+
+/* define generated endian-swappers */
+#define vl_endianfun
+#include <vpp/api/vpe_all_api_h.h>
+#undef vl_endianfun
+
+/* instantiate all the print functions we know about */
+#define vl_print(handle, ...) vlib_cli_output (handle, __VA_ARGS__)
+#define vl_printfun
+#include <vpp/api/vpe_all_api_h.h>
+#undef vl_printfun
+
 typedef struct
 {
   u8 *rx_buf;
   unix_shared_memory_queue_t **vpp_queue;
-  u32 byte_index;
+  u64 byte_index;
+
+  /* Sever's event queue */
+  unix_shared_memory_queue_t *vl_input_queue;
+
+  /* API client handle */
+  u32 my_client_index;
+
+  u32 app_index;
+
+  /* process node index for evnt scheduling */
+  u32 node_index;
   vlib_main_t *vlib_main;
 } builtin_server_main_t;
 
 builtin_server_main_t builtin_server_main;
 
-
 int
 builtin_session_accept_callback (stream_session_t * s)
 {
@@ -45,9 +74,13 @@ builtin_session_accept_callback (stream_session_t * s)
 void
 builtin_session_disconnect_callback (stream_session_t * s)
 {
+  builtin_server_main_t *bsm = &builtin_server_main;
+  vnet_disconnect_args_t _a, *a = &_a;
   clib_warning ("called...");
 
-  vnet_disconnect_session (s->session_index, s->thread_index);
+  a->handle = stream_session_handle (s);
+  a->app_index = bsm->app_index;
+  vnet_disconnect_session (a);
 }
 
 void
@@ -60,7 +93,7 @@ builtin_session_reset_callback (stream_session_t * s)
 
 
 int
-builtin_session_connected_callback (u32 client_index,
+builtin_session_connected_callback (u32 app_index, u32 api_context,
 				    stream_session_t * s, u8 is_fail)
 {
   clib_warning ("called...");
@@ -91,7 +124,7 @@ test_bytes (builtin_server_main_t * bsm, int actual_transfer)
     {
       if (bsm->rx_buf[i] != ((bsm->byte_index + i) & 0xff))
 	{
-	  clib_warning ("at %d expected %d got %d", bsm->byte_index + i,
+	  clib_warning ("at %lld expected %d got %d", bsm->byte_index + i,
 			(bsm->byte_index + i) & 0xff, bsm->rx_buf[i]);
 	}
     }
@@ -190,23 +223,66 @@ static session_cb_vft_t builtin_session_cb_vft = {
   .session_reset_callback = builtin_session_reset_callback
 };
 
+/* Abuse VPP's input queue */
 static int
-server_create (vlib_main_t * vm)
+create_api_loopback (vlib_main_t * vm)
 {
-  vnet_bind_args_t _a, *a = &_a;
-  u64 options[SESSION_OPTIONS_N_OPTIONS];
-  char segment_name[128];
-  u32 num_threads;
-  vlib_thread_main_t *vtm = vlib_get_thread_main ();
+  builtin_server_main_t *bsm = &builtin_server_main;
+  vl_api_memclnt_create_t _m, *mp = &_m;
+  extern void vl_api_memclnt_create_t_handler (vl_api_memclnt_create_t *);
+  api_main_t *am = &api_main;
+  vl_shmem_hdr_t *shmem_hdr;
+  uword *event_data = 0, event_type;
+  int resolved = 0;
 
-  num_threads = 1 /* main thread */  + vtm->n_threads;
-  vec_validate (builtin_server_main.vpp_queue, num_threads - 1);
+  /*
+   * Create a "loopback" API client connection
+   * Don't do things like this unless you know what you're doing...
+   */
+
+  shmem_hdr = am->shmem_hdr;
+  bsm->vl_input_queue = shmem_hdr->vl_input_queue;
+  memset (mp, 0, sizeof (*mp));
+  mp->_vl_msg_id = VL_API_MEMCLNT_CREATE;
+  mp->context = 0xFEEDFACE;
+  mp->input_queue = (u64) bsm->vl_input_queue;
+  strncpy ((char *) mp->name, "tcp_test_server", sizeof (mp->name) - 1);
+
+  vl_api_memclnt_create_t_handler (mp);
+
+  /* Wait for reply */
+  bsm->node_index = vlib_get_current_process (vm)->node_runtime.node_index;
+  vlib_process_wait_for_event_or_clock (vm, 1.0);
+  event_type = vlib_process_get_events (vm, &event_data);
+  switch (event_type)
+    {
+    case 1:
+      resolved = 1;
+      break;
+    case ~0:
+      /* timed out */
+      break;
+    default:
+      clib_warning ("unknown event_type %d", event_type);
+    }
+  if (!resolved)
+    return -1;
+
+  return 0;
+}
+
+static int
+server_attach ()
+{
+  builtin_server_main_t *bsm = &builtin_server_main;
+  u8 segment_name[128];
+  u64 options[SESSION_OPTIONS_N_OPTIONS];
+  vnet_app_attach_args_t _a, *a = &_a;
 
   memset (a, 0, sizeof (*a));
   memset (options, 0, sizeof (options));
 
-  a->uri = "tcp://0.0.0.0/1234";
-  a->api_client_index = ~0;
+  a->api_client_index = bsm->my_client_index;
   a->session_cb_vft = &builtin_session_cb_vft;
   a->options = options;
   a->options[SESSION_OPTIONS_SEGMENT_SIZE] = 128 << 20;
@@ -215,9 +291,94 @@ server_create (vlib_main_t * vm)
   a->segment_name = segment_name;
   a->segment_name_length = ARRAY_LEN (segment_name);
 
+  if (vnet_application_attach (a))
+    {
+      clib_warning ("failed to attach server");
+      return -1;
+    }
+  bsm->app_index = a->app_index;
+  return 0;
+}
+
+static int
+server_listen ()
+{
+  builtin_server_main_t *bsm = &builtin_server_main;
+  vnet_bind_args_t _a, *a = &_a;
+  memset (a, 0, sizeof (*a));
+  a->app_index = bsm->app_index;
+  a->uri = "tcp://0.0.0.0/1234";
   return vnet_bind_uri (a);
 }
 
+static int
+server_create (vlib_main_t * vm)
+{
+  builtin_server_main_t *bsm = &builtin_server_main;
+  u32 num_threads;
+  vlib_thread_main_t *vtm = vlib_get_thread_main ();
+
+  if (bsm->my_client_index == (u32) ~ 0)
+    {
+      if (create_api_loopback (vm))
+	return -1;
+    }
+
+  num_threads = 1 /* main thread */  + vtm->n_threads;
+  vec_validate (builtin_server_main.vpp_queue, num_threads - 1);
+
+  if (server_attach ())
+    {
+      clib_warning ("failed to attach server");
+      return -1;
+    }
+  if (server_listen ())
+    {
+      clib_warning ("failed to start listening");
+      return -1;
+    }
+  return 0;
+}
+
+/* Get our api client index */
+static void
+vl_api_memclnt_create_reply_t_handler (vl_api_memclnt_create_reply_t * mp)
+{
+  vlib_main_t *vm = vlib_get_main ();
+  builtin_server_main_t *bsm = &builtin_server_main;
+  bsm->my_client_index = mp->index;
+  vlib_process_signal_event (vm, bsm->node_index, 1 /* evt */ ,
+			     0 /* data */ );
+}
+
+#define foreach_tcp_builtin_server_api_msg      		\
+_(MEMCLNT_CREATE_REPLY, memclnt_create_reply)   		\
+
+static clib_error_t *
+tcp_builtin_server_api_hookup (vlib_main_t * vm)
+{
+  vl_msg_api_msg_config_t _c, *c = &_c;
+
+  /* Hook up client-side static APIs to our handlers */
+#define _(N,n) do {                                             \
+    c->id = VL_API_##N;                                         \
+    c->name = #n;                                               \
+    c->handler = vl_api_##n##_t_handler;                        \
+    c->cleanup = vl_noop_handler;                               \
+    c->endian = vl_api_##n##_t_endian;                          \
+    c->print = vl_api_##n##_t_print;                            \
+    c->size = sizeof(vl_api_##n##_t);                           \
+    c->traced = 1; /* trace, so these msgs print */             \
+    c->replay = 0; /* don't replay client create/delete msgs */ \
+    c->message_bounce = 0; /* don't bounce this message */	\
+    vl_msg_api_config(c);} while (0);
+
+  foreach_tcp_builtin_server_api_msg;
+#undef _
+
+  return 0;
+}
+
 static clib_error_t *
 server_create_command_fn (vlib_main_t * vm,
 			  unformat_input_t * input, vlib_cli_command_t * cmd)
@@ -234,6 +395,7 @@ server_create_command_fn (vlib_main_t * vm,
     }
 #endif
 
+  tcp_builtin_server_api_hookup (vm);
   vnet_session_enable_disable (vm, 1 /* turn on TCP, etc. */ );
   rv = server_create (vm);
   switch (rv)
@@ -249,12 +411,22 @@ server_create_command_fn (vlib_main_t * vm,
 /* *INDENT-OFF* */
 VLIB_CLI_COMMAND (server_create_command, static) =
 {
-  .path = "test server",
-  .short_help = "test server",
+  .path = "test tcp server",
+  .short_help = "test tcp server",
   .function = server_create_command_fn,
 };
 /* *INDENT-ON* */
 
+clib_error_t *
+builtin_tcp_server_main_init (vlib_main_t * vm)
+{
+  builtin_server_main_t *bsm = &builtin_server_main;
+  bsm->my_client_index = ~0;
+  return 0;
+}
+
+VLIB_INIT_FUNCTION (builtin_tcp_server_main_init);
+
 /*
 * fd.io coding-style-patch-verification: ON
 *
diff --git a/src/vnet/tcp/tcp.c b/src/vnet/tcp/tcp.c
index b6c34828..a0c66b9f 100644
--- a/src/vnet/tcp/tcp.c
+++ b/src/vnet/tcp/tcp.c
@@ -34,14 +34,19 @@ tcp_connection_bind (u32 session_index, ip46_address_t * ip,
   listener->c_lcl_port = clib_host_to_net_u16 (port_host_byte_order);
 
   if (is_ip4)
-    listener->c_lcl_ip4.as_u32 = ip->ip4.as_u32;
+    {
+      listener->c_lcl_ip4.as_u32 = ip->ip4.as_u32;
+      listener->c_is_ip4 = 1;
+      listener->c_proto = SESSION_TYPE_IP4_TCP;
+    }
   else
-    clib_memcpy (&listener->c_lcl_ip6, &ip->ip6, sizeof (ip6_address_t));
+    {
+      clib_memcpy (&listener->c_lcl_ip6, &ip->ip6, sizeof (ip6_address_t));
+      listener->c_proto = SESSION_TYPE_IP6_TCP;
+    }
 
   listener->c_s_index = session_index;
-  listener->c_proto = SESSION_TYPE_IP4_TCP;
   listener->state = TCP_STATE_LISTEN;
-  listener->c_is_ip4 = 1;
 
   tcp_connection_timers_init (listener);
 
@@ -62,7 +67,6 @@ tcp_session_bind_ip6 (u32 session_index, ip46_address_t * ip,
 		      u16 port_host_byte_order)
 {
   return tcp_connection_bind (session_index, ip, port_host_byte_order, 0);
-
 }
 
 static void
@@ -397,6 +401,7 @@ tcp_connection_open (ip46_address_t * rmt_addr, u16 rmt_port, u8 is_ip4)
   tc->c_lcl_port = clib_host_to_net_u16 (lcl_port);
   tc->c_c_index = tc - tm->half_open_connections;
   tc->c_is_ip4 = is_ip4;
+  tc->c_proto = is_ip4 ? SESSION_TYPE_IP4_TCP : SESSION_TYPE_IP6_TCP;
 
   /* The other connection vars will be initialized after SYN ACK */
   tcp_connection_timers_init (tc);
@@ -518,7 +523,10 @@ format_tcp_session (u8 * s, va_list * args)
   tcp_connection_t *tc;
 
   tc = tcp_connection_get (tci, thread_index);
-  return format (s, "%U", format_tcp_connection, tc);
+  if (tc)
+    return format (s, "%U", format_tcp_connection, tc);
+  else
+    return format (s, "empty");
 }
 
 u8 *
diff --git a/src/vnet/tcp/tcp.h b/src/vnet/tcp/tcp.h
index 2f5da108..93f3245d 100644
--- a/src/vnet/tcp/tcp.h
+++ b/src/vnet/tcp/tcp.h
@@ -100,8 +100,6 @@ extern timer_expiration_handler tcp_timer_retransmit_syn_handler;
 #define TCP_RTO_SYN_RETRIES 3	/* SYN retries without doubling RTO */
 #define TCP_RTO_INIT 1 * THZ	/* Initial retransmit timer */
 
-void tcp_update_time (f64 now, u32 thread_index);
-
 /** TCP connection flags */
 #define foreach_tcp_connection_flag             \
   _(SNDACK, "Send ACK")                         \
@@ -481,6 +479,13 @@ tcp_time_now (void)
   return clib_cpu_time_now () * tcp_main.tstamp_ticks_per_clock;
 }
 
+always_inline void
+tcp_update_time (f64 now, u32 thread_index)
+{
+  tw_timer_expire_timers_16t_2w_512sl (&tcp_main.timer_wheels[thread_index],
+				       now);
+}
+
 u32 tcp_push_header (transport_connection_t * tconn, vlib_buffer_t * b);
 
 u32
diff --git a/src/vnet/tcp/tcp_input.c b/src/vnet/tcp/tcp_input.c
index 7e9fa47b..ae1f92d5 100644
--- a/src/vnet/tcp/tcp_input.c
+++ b/src/vnet/tcp/tcp_input.c
@@ -1841,6 +1841,7 @@ tcp46_rcv_process_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
 	    case TCP_STATE_ESTABLISHED:
 	    case TCP_STATE_FIN_WAIT_1:
 	    case TCP_STATE_FIN_WAIT_2:
+	      vlib_buffer_advance (b0, n_advance_bytes0);
 	      error0 = tcp_segment_rcv (tm, tc0, b0, n_data_bytes0, &next0);
 	      break;
 	    case TCP_STATE_CLOSE_WAIT:
@@ -2410,12 +2411,6 @@ VLIB_REGISTER_NODE (tcp6_input_node) =
 /* *INDENT-ON* */
 
 VLIB_NODE_FUNCTION_MULTIARCH (tcp6_input_node, tcp6_input);
-void
-tcp_update_time (f64 now, u32 thread_index)
-{
-  tcp_main_t *tm = vnet_get_tcp_main ();
-  tw_timer_expire_timers_16t_2w_512sl (&tm->timer_wheels[thread_index], now);
-}
 
 static void
 tcp_dispatch_table_init (tcp_main_t * tm)
diff --git a/src/vnet/tcp/tcp_test.c b/src/vnet/tcp/tcp_test.c
index 0725bb04..3dbbdf6f 100644
--- a/src/vnet/tcp/tcp_test.c
+++ b/src/vnet/tcp/tcp_test.c
@@ -12,7 +12,6 @@
  * See the License for the specific language governing permissions and
  * limitations under the License.
  */
-
 #include <vnet/tcp/tcp.h>
 
 #define TCP_TEST_I(_cond, _comment, _args...)			\
@@ -174,6 +173,118 @@ tcp_test_sack ()
   return 0;
 }
 
+static int
+tcp_test_fifo (vlib_main_t * vm, unformat_input_t * input)
+{
+  svm_fifo_t *f;
+  u32 fifo_size = 1 << 20;
+  u32 *test_data = 0;
+  u32 offset;
+  int i, rv;
+  u32 data_word, test_data_len;
+
+  /* $$$ parse args */
+  test_data_len = fifo_size / sizeof (u32);
+  vec_validate (test_data, test_data_len - 1);
+
+  for (i = 0; i < vec_len (test_data); i++)
+    test_data[i] = i;
+
+  f = svm_fifo_create (fifo_size);
+
+  /* Paint fifo data vector with -1's */
+  memset (f->data, 0xFF, test_data_len);
+
+  /* Enqueue an initial (un-dequeued) chunk */
+  rv = svm_fifo_enqueue_nowait (f, 0 /* pid */ ,
+				sizeof (u32), (u8 *) test_data);
+
+  if (rv != sizeof (u32))
+    {
+      clib_warning ("enqueue returned %d", rv);
+      goto out;
+    }
+
+  /*
+   * Create 3 chunks in the future. The offsets are relative
+   * to the current fifo tail
+   */
+  for (i = 0; i < 3; i++)
+    {
+      offset = (2 * i + 1) * sizeof (u32);
+      vlib_cli_output (vm, "add offset %d", offset);
+
+      rv = svm_fifo_enqueue_with_offset
+	(f, 0 /* pid */ , offset, sizeof (u32),
+	 (u8 *) (test_data + ((offset + sizeof (u32)) / sizeof (u32))));
+
+      if (rv)
+	{
+	  clib_warning ("enqueue returned %d", rv);
+	  goto out;
+	}
+    }
+
+  /* Paint missing data backwards */
+  for (i = 3; i > 0; i--)
+    {
+      offset = (2 * i + 0) * sizeof (u32);
+
+      vlib_cli_output (vm, "add offset %d", offset);
+
+      rv = svm_fifo_enqueue_with_offset
+	(f, 0 /* pid */ , offset, sizeof (u32),
+	 (u8 *) (test_data + ((offset + sizeof (u32)) / sizeof (u32))));
+
+      if (rv)
+	{
+	  clib_warning ("enqueue returned %d", rv);
+	  goto out;
+	}
+    }
+
+  vlib_cli_output (vm, "fifo before missing link: %U",
+		   format_svm_fifo, f, 1 /* verbose */ );
+
+  /* Enqueue the missing u32 */
+  rv = svm_fifo_enqueue_nowait (f, 0 /* pid */ ,
+				sizeof (u32), (u8 *) (test_data + 1));
+  if (rv != 7 * sizeof (u32))
+    {
+      clib_warning ("enqueue returned %d", rv);
+      goto out;
+    }
+
+  vlib_cli_output (vm, "fifo after missing link: %U",
+		   format_svm_fifo, f, 1 /* verbose */ );
+
+  /* Collect results */
+  for (i = 0; i < 7; i++)
+    {
+      rv = svm_fifo_dequeue_nowait (f, 0 /* pid */ , sizeof (u32),
+				    (u8 *) & data_word);
+      if (rv != sizeof (u32))
+	{
+	  clib_warning ("dequeue returned %d", rv);
+	  goto out;
+	}
+      if (data_word != test_data[i])
+	{
+	  clib_warning ("recovered data %d not %d", data_word, test_data[i]);
+	  goto out;
+	}
+    }
+
+  clib_warning ("test complete...");
+
+out:
+  svm_fifo_free (f);
+  vec_free (test_data);
+  return 0;
+}
+
+
+
 static clib_error_t *
 tcp_test (vlib_main_t * vm,
 	  unformat_input_t * input, vlib_cli_command_t * cmd_arg)
@@ -186,6 +297,10 @@ tcp_test (vlib_main_t * vm,
 	{
 	  res = tcp_test_sack ();
 	}
+      else if (unformat (input, "fifo"))
+	{
+	  res = tcp_test_fifo (vm, input);
+	}
       else
 	{
 	  return clib_error_return (0, "unknown input `%U'",
@@ -203,10 +318,16 @@ tcp_test (vlib_main_t * vm,
     }
 }
 
+/* *INDENT-OFF* */
 VLIB_CLI_COMMAND (tcp_test_command, static) =
 {
-.path = "test tcp",.short_help = "internal tcp unit tests",.function =
-    tcp_test,};
+  .path = "test tcp",
+  .short_help = "internal tcp unit tests",
+  .function = tcp_test,
+};
+/* *INDENT-ON* */
+
+
 /*
  * fd.io coding-style-patch-verification: ON
  *
diff --git a/src/vnet/udp/builtin_server.c b/src/vnet/udp/builtin_server.c
index 57f774c5..8565f04c 100644
--- a/src/vnet/udp/builtin_server.c
+++ b/src/vnet/udp/builtin_server.c
@@ -91,12 +91,11 @@ static session_cb_vft_t builtin_server = {
 /* *INDENT-ON* */
 
 static int
-bind_builtin_uri_server (u8 * uri)
+attach_builtin_uri_server ()
 {
-  vnet_bind_args_t _a, *a = &_a;
-  char segment_name[128];
+  vnet_app_attach_args_t _a, *a = &_a;
+  u8 segment_name[128];
   u32 segment_name_length;
-  int rv;
   u64 options[16];
 
   segment_name_length = ARRAY_LEN (segment_name);
@@ -104,8 +103,7 @@ bind_builtin_uri_server (u8 * uri)
   memset (a, 0, sizeof (*a));
   memset (options, 0, sizeof (options));
 
-  a->uri = (char *) uri;
-  a->api_client_index = ~0;	/* built-in server */
+  a->api_client_index = ~0;
   a->segment_name = segment_name;
   a->segment_name_length = segment_name_length;
   a->session_cb_vft = &builtin_server;
@@ -114,6 +112,23 @@ bind_builtin_uri_server (u8 * uri)
   options[SESSION_OPTIONS_SEGMENT_SIZE] = (2 << 30);	/*$$$$ config / arg */
   a->options = options;
 
+  return vnet_application_attach (a);
+}
+
+static int
+bind_builtin_uri_server (u8 * uri)
+{
+  vnet_bind_args_t _a, *a = &_a;
+  int rv;
+
+  rv = attach_builtin_uri_server ();
+  if (rv)
+    return rv;
+
+  memset (a, 0, sizeof (*a));
+  a->uri = (char *) uri;
+  a->app_index = ~0;		/* built-in server */
+
   rv = vnet_bind_uri (a);
 
   return rv;
@@ -122,11 +137,12 @@ bind_builtin_uri_server (u8 * uri)
 static int
 unbind_builtin_uri_server (u8 * uri)
 {
-  int rv;
+  vnet_unbind_args_t _a, *a = &_a;
 
-  rv = vnet_unbind_uri ((char *) uri, ~0 /* client_index */ );
+  a->app_index = ~0;
+  a->uri = (char *) uri;
 
-  return rv;
+  return vnet_unbind_uri (a);
 }
 
 static clib_error_t *
-- 
cgit 1.2.3-korg


From 1f75cfd73320476a8f821064391fe368dd4bf75b Mon Sep 17 00:00:00 2001
From: Dave Barach <dave@barachs.net>
Date: Fri, 14 Apr 2017 16:46:44 -0400
Subject: Fix fifo ooo bugs and improve testing

Change-Id: If3c01e318bcb740ca5b240c63f712e2167082a80
Signed-off-by: Dave Barach <dave@barachs.net>
Signed-off-by: Florin Coras <fcoras@cisco.com>
---
 src/svm/svm_fifo.c        | 126 +++++++----
 src/svm/svm_fifo.h        |  21 +-
 src/vnet/tcp/tcp.c        |   2 +-
 src/vnet/tcp/tcp.h        |   1 +
 src/vnet/tcp/tcp_format.c |   2 +-
 src/vnet/tcp/tcp_input.c  |  28 ++-
 src/vnet/tcp/tcp_test.c   | 516 ++++++++++++++++++++++++++++++++++++++++++----
 7 files changed, 602 insertions(+), 94 deletions(-)

(limited to 'src/svm/svm_fifo.c')

diff --git a/src/svm/svm_fifo.c b/src/svm/svm_fifo.c
index 097bab77..bd968aea 100644
--- a/src/svm/svm_fifo.c
+++ b/src/svm/svm_fifo.c
@@ -15,6 +15,36 @@
 
 #include <svm/svm_fifo.h>
 
+#define offset_lt(_a, _b) ((i32)((_a)-(_b)) < 0)
+#define offset_leq(_a, _b) ((i32)((_a)-(_b)) <= 0)
+
+u8 *
+format_ooo_segment (u8 * s, va_list * args)
+{
+  ooo_segment_t *seg = va_arg (*args, ooo_segment_t *);
+
+  s = format (s, "pos %u, len %u, next %d, prev %d",
+	      seg->start, seg->length, seg->next, seg->prev);
+  return s;
+}
+
+u8 *
+format_ooo_list (u8 * s, va_list * args)
+{
+  svm_fifo_t *f = va_arg (*args, svm_fifo_t *);
+  u32 ooo_segment_index = f->ooos_list_head;
+  ooo_segment_t *seg;
+
+  while (ooo_segment_index != OOO_SEGMENT_INVALID_INDEX)
+    {
+      seg = pool_elt_at_index (f->ooo_segments, ooo_segment_index);
+      s = format (s, "\n  %U", format_ooo_segment, seg);
+
+      ooo_segment_index = seg->next;
+    }
+  return s;
+}
+
 /** create an svm fifo, in the current heap. Fails vs blow up the process */
 svm_fifo_t *
 svm_fifo_create (u32 data_size_in_bytes)
@@ -47,7 +77,7 @@ ooo_segment_new (svm_fifo_t * f, u32 start, u32 length)
 
   pool_get (f->ooo_segments, s);
 
-  s->fifo_position = start;
+  s->start = start;
   s->length = length;
 
   s->prev = s->next = OOO_SEGMENT_INVALID_INDEX;
@@ -88,14 +118,13 @@ static void
 ooo_segment_add (svm_fifo_t * f, u32 offset, u32 length)
 {
   ooo_segment_t *s, *new_s, *prev, *next, *it;
-  u32 new_index, position, end_offset, s_sof, s_eof, s_index;
+  u32 new_index, end_offset, s_sof, s_eof, s_index;
 
-  position = (f->tail + offset) % f->nitems;
   end_offset = offset + length;
 
   if (f->ooos_list_head == OOO_SEGMENT_INVALID_INDEX)
     {
-      s = ooo_segment_new (f, position, length);
+      s = ooo_segment_new (f, offset, length);
       f->ooos_list_head = s - f->ooo_segments;
       f->ooos_newest = f->ooos_list_head;
       return;
@@ -104,26 +133,26 @@ ooo_segment_add (svm_fifo_t * f, u32 offset, u32 length)
   /* Find first segment that starts after new segment */
   s = pool_elt_at_index (f->ooo_segments, f->ooos_list_head);
   while (s->next != OOO_SEGMENT_INVALID_INDEX
-	 && ooo_segment_offset (f, s) <= offset)
+	 && offset_leq (ooo_segment_offset (f, s), offset))
     s = pool_elt_at_index (f->ooo_segments, s->next);
 
   s_index = s - f->ooo_segments;
   s_sof = ooo_segment_offset (f, s);
   s_eof = ooo_segment_end_offset (f, s);
+  prev = ooo_segment_get_prev (f, s);
 
   /* No overlap, add before current segment */
-  if (end_offset < s_sof)
+  if (offset_lt (end_offset, s_sof)
+      && (!prev || offset_lt (prev->start + prev->length, offset)))
     {
-      new_s = ooo_segment_new (f, position, length);
+      new_s = ooo_segment_new (f, offset, length);
       new_index = new_s - f->ooo_segments;
 
       /* Pool might've moved, get segment again */
       s = pool_elt_at_index (f->ooo_segments, s_index);
-
       if (s->prev != OOO_SEGMENT_INVALID_INDEX)
 	{
 	  new_s->prev = s->prev;
-
 	  prev = pool_elt_at_index (f->ooo_segments, new_s->prev);
 	  prev->next = new_index;
 	}
@@ -139,9 +168,9 @@ ooo_segment_add (svm_fifo_t * f, u32 offset, u32 length)
       return;
     }
   /* No overlap, add after current segment */
-  else if (s_eof < offset)
+  else if (offset_lt (s_eof, offset))
     {
-      new_s = ooo_segment_new (f, position, length);
+      new_s = ooo_segment_new (f, offset, length);
       new_index = new_s - f->ooo_segments;
 
       /* Pool might've moved, get segment again */
@@ -150,7 +179,6 @@ ooo_segment_add (svm_fifo_t * f, u32 offset, u32 length)
       if (s->next != OOO_SEGMENT_INVALID_INDEX)
 	{
 	  new_s->next = s->next;
-
 	  next = pool_elt_at_index (f->ooo_segments, new_s->next);
 	  next->prev = new_index;
 	}
@@ -167,7 +195,7 @@ ooo_segment_add (svm_fifo_t * f, u32 offset, u32 length)
    */
 
   /* Merge at head */
-  if (offset <= s_sof)
+  if (offset_leq (offset, s_sof))
     {
       /* If we have a previous, check if we overlap */
       if (s->prev != OOO_SEGMENT_INVALID_INDEX)
@@ -176,26 +204,31 @@ ooo_segment_add (svm_fifo_t * f, u32 offset, u32 length)
 
 	  /* New segment merges prev and current. Remove previous and
 	   * update position of current. */
-	  if (ooo_segment_end_offset (f, prev) >= offset)
+	  if (offset_leq (offset, ooo_segment_end_offset (f, prev)))
 	    {
-	      s->fifo_position = prev->fifo_position;
+	      s->start = prev->start;
 	      s->length = s_eof - ooo_segment_offset (f, prev);
 	      ooo_segment_del (f, s->prev);
 	    }
+	  else
+	    {
+	      s->start = offset;
+	      s->length = s_eof - ooo_segment_offset (f, s);
+	    }
 	}
       else
 	{
-	  s->fifo_position = position;
+	  s->start = offset;
 	  s->length = s_eof - ooo_segment_offset (f, s);
 	}
 
       /* The new segment's tail may cover multiple smaller ones */
-      if (s_eof < end_offset)
+      if (offset_lt (s_eof, end_offset))
 	{
 	  /* Remove segments completely covered */
 	  it = (s->next != OOO_SEGMENT_INVALID_INDEX) ?
 	    pool_elt_at_index (f->ooo_segments, s->next) : 0;
-	  while (it && ooo_segment_end_offset (f, it) < end_offset)
+	  while (it && offset_lt (ooo_segment_end_offset (f, it), end_offset))
 	    {
 	      next = (it->next != OOO_SEGMENT_INVALID_INDEX) ?
 		pool_elt_at_index (f->ooo_segments, it->next) : 0;
@@ -207,7 +240,7 @@ ooo_segment_add (svm_fifo_t * f, u32 offset, u32 length)
 	  s->length = end_offset - ooo_segment_offset (f, s);
 
 	  /* If partial overlap with last, merge */
-	  if (it && ooo_segment_offset (f, it) < end_offset)
+	  if (it && offset_lt (ooo_segment_offset (f, it), end_offset))
 	    {
 	      s->length +=
 		it->length - (ooo_segment_offset (f, it) - end_offset);
@@ -216,7 +249,7 @@ ooo_segment_add (svm_fifo_t * f, u32 offset, u32 length)
 	}
     }
   /* Last but overlapping previous */
-  else if (s_eof <= end_offset)
+  else if (offset_leq (s_eof, end_offset))
     {
       s->length = end_offset - ooo_segment_offset (f, s);
     }
@@ -247,7 +280,7 @@ ooo_segment_try_collect (svm_fifo_t * f, u32 n_bytes_enqueued)
   s = pool_elt_at_index (f->ooo_segments, f->ooos_list_head);
 
   /* If last tail update overlaps one/multiple ooo segments, remove them */
-  diff = (f->nitems + f->tail - s->fifo_position) % f->nitems;
+  diff = (f->nitems + ((int) s->start - f->tail)) % f->nitems;
   while (0 < diff && diff < n_bytes_enqueued)
     {
       /* Segment end is beyond the tail. Advance tail and be done */
@@ -262,7 +295,7 @@ ooo_segment_try_collect (svm_fifo_t * f, u32 n_bytes_enqueued)
 	{
 	  index = s - f->ooo_segments;
 	  s = pool_elt_at_index (f->ooo_segments, s->next);
-	  diff = (f->nitems + f->tail - s->fifo_position) % f->nitems;
+	  diff = (f->nitems + ((int) s->start - f->tail)) % f->nitems;
 	  ooo_segment_del (f, index);
 	}
       /* End of search */
@@ -368,9 +401,20 @@ svm_fifo_enqueue_with_offset_internal (svm_fifo_t * f,
 {
   u32 total_copy_bytes, first_copy_bytes, second_copy_bytes;
   u32 cursize, nitems;
-  u32 tail_plus_offset;
+  u32 normalized_offset;
+  int rv;
 
-  ASSERT (offset > 0);
+  /* Safety: don't wrap more than nitems/2 */
+  ASSERT ((f->nitems + offset - f->tail) % f->nitems < f->nitems / 2);
+
+  /* Users would do do well to avoid this */
+  if (PREDICT_FALSE (f->tail == (offset % f->nitems)))
+    {
+      rv = svm_fifo_enqueue_internal (f, pid, required_bytes, copy_from_here);
+      if (rv > 0)
+	return 0;
+      return -1;
+    }
 
   /* read cursize, which can only increase while we're working */
   cursize = svm_fifo_max_dequeue (f);
@@ -384,24 +428,24 @@ svm_fifo_enqueue_with_offset_internal (svm_fifo_t * f,
 
   /* Number of bytes we're going to copy */
   total_copy_bytes = required_bytes;
-  tail_plus_offset = (f->tail + offset) % nitems;
+  normalized_offset = offset % nitems;
 
   /* Number of bytes in first copy segment */
-  first_copy_bytes = ((nitems - tail_plus_offset) < total_copy_bytes)
-    ? (nitems - tail_plus_offset) : total_copy_bytes;
+  first_copy_bytes = ((nitems - normalized_offset) < total_copy_bytes)
+    ? (nitems - normalized_offset) : total_copy_bytes;
 
-  clib_memcpy (&f->data[tail_plus_offset], copy_from_here, first_copy_bytes);
+  clib_memcpy (&f->data[normalized_offset], copy_from_here, first_copy_bytes);
 
   /* Number of bytes in second copy segment, if any */
   second_copy_bytes = total_copy_bytes - first_copy_bytes;
   if (second_copy_bytes)
     {
-      tail_plus_offset += first_copy_bytes;
-      tail_plus_offset %= nitems;
+      normalized_offset += first_copy_bytes;
+      normalized_offset %= nitems;
 
-      ASSERT (tail_plus_offset == 0);
+      ASSERT (normalized_offset == 0);
 
-      clib_memcpy (&f->data[tail_plus_offset],
+      clib_memcpy (&f->data[normalized_offset],
 		   copy_from_here + first_copy_bytes, second_copy_bytes);
     }
 
@@ -573,8 +617,8 @@ format_svm_fifo (u8 * s, va_list * args)
       ooo_segment_t *seg;
       u32 seg_index;
 
-      s =
-	format (s, "ooo pool %d active elts\n", pool_elts (f->ooo_segments));
+      s = format (s, "ooo pool %d active elts\n",
+		  pool_elts (f->ooo_segments));
 
       seg_index = f->ooos_list_head;
 
@@ -582,13 +626,25 @@ format_svm_fifo (u8 * s, va_list * args)
 	{
 	  seg = pool_elt_at_index (f->ooo_segments, seg_index);
 	  s = format (s, "  pos %u, len %u next %d\n",
-		      seg->fifo_position, seg->length, seg->next);
+		      seg->start, seg->length, seg->next);
 	  seg_index = seg->next;
 	}
     }
   return s;
 }
 
+u32
+svm_fifo_number_ooo_segments (svm_fifo_t * f)
+{
+  return pool_elts (f->ooo_segments);
+}
+
+ooo_segment_t *
+svm_fifo_first_ooo_segment (svm_fifo_t * f)
+{
+  return pool_elt_at_index (f->ooo_segments, f->ooos_list_head);
+}
+
 /*
  * fd.io coding-style-patch-verification: ON
  *
diff --git a/src/svm/svm_fifo.h b/src/svm/svm_fifo.h
index 9beb63f5..0fff2577 100644
--- a/src/svm/svm_fifo.h
+++ b/src/svm/svm_fifo.h
@@ -36,10 +36,13 @@ typedef struct
   u32 next;	/**< Next linked-list element pool index */
   u32 prev;	/**< Previous linked-list element pool index */
 
-  u32 fifo_position;	/**< Start of segment, normalized*/
+  u32 start;	/**< Start of segment, normalized*/
   u32 length;		/**< Length of segment */
 } ooo_segment_t;
 
+format_function_t format_ooo_segment;
+format_function_t format_ooo_list;
+
 #define OOO_SEGMENT_INVALID_INDEX ((u32)~0)
 
 typedef struct
@@ -127,6 +130,8 @@ int svm_fifo_dequeue_nowait (svm_fifo_t * f, int pid, u32 max_bytes,
 int svm_fifo_peek (svm_fifo_t * f, int pid, u32 offset, u32 max_bytes,
 		   u8 * copy_here);
 int svm_fifo_dequeue_drop (svm_fifo_t * f, int pid, u32 max_bytes);
+u32 svm_fifo_number_ooo_segments (svm_fifo_t * f);
+ooo_segment_t *svm_fifo_first_ooo_segment (svm_fifo_t * f);
 
 format_function_t format_svm_fifo;
 
@@ -139,13 +144,23 @@ svm_fifo_newest_ooo_segment (svm_fifo_t * f)
 always_inline u32
 ooo_segment_offset (svm_fifo_t * f, ooo_segment_t * s)
 {
-  return ((f->nitems + s->fifo_position - f->tail) % f->nitems);
+//  return ((f->nitems + s->fifo_position - f->tail) % f->nitems);
+  return s->start;
 }
 
 always_inline u32
 ooo_segment_end_offset (svm_fifo_t * f, ooo_segment_t * s)
 {
-  return ((f->nitems + s->fifo_position + s->length - f->tail) % f->nitems);
+//  return ((f->nitems + s->fifo_position + s->length - f->tail) % f->nitems);
+  return s->start + s->length;
+}
+
+always_inline ooo_segment_t *
+ooo_segment_get_prev (svm_fifo_t * f, ooo_segment_t * s)
+{
+  if (s->prev == OOO_SEGMENT_INVALID_INDEX)
+    return 0;
+  return pool_elt_at_index (f->ooo_segments, s->prev);
 }
 
 #endif /* __included_ssvm_fifo_h__ */
diff --git a/src/vnet/tcp/tcp.c b/src/vnet/tcp/tcp.c
index a0c66b9f..12982589 100644
--- a/src/vnet/tcp/tcp.c
+++ b/src/vnet/tcp/tcp.c
@@ -447,7 +447,7 @@ format_tcp_state (u8 * s, va_list * args)
   if (*state < TCP_N_STATES)
     s = format (s, "%s", tcp_fsm_states[*state]);
   else
-    s = format (s, "UNKNOWN");
+    s = format (s, "UNKNOWN (%d (0x%x))", *state, *state);
 
   return s;
 }
diff --git a/src/vnet/tcp/tcp.h b/src/vnet/tcp/tcp.h
index 225b26da..2ac6a9b8 100644
--- a/src/vnet/tcp/tcp.h
+++ b/src/vnet/tcp/tcp.h
@@ -58,6 +58,7 @@ typedef enum _tcp_state
 } tcp_state_t;
 
 format_function_t format_tcp_state;
+format_function_t format_tcp_flags;
 
 /** TCP timers */
 #define foreach_tcp_timer               \
diff --git a/src/vnet/tcp/tcp_format.c b/src/vnet/tcp/tcp_format.c
index 7136741d..994ccfd6 100644
--- a/src/vnet/tcp/tcp_format.c
+++ b/src/vnet/tcp/tcp_format.c
@@ -40,7 +40,7 @@
 #include <vnet/ip/ip.h>
 #include <vnet/tcp/tcp.h>
 
-static u8 *
+u8 *
 format_tcp_flags (u8 * s, va_list * args)
 {
   int flags = va_arg (*args, int);
diff --git a/src/vnet/tcp/tcp_input.c b/src/vnet/tcp/tcp_input.c
index a12ad8c0..97679aaf 100644
--- a/src/vnet/tcp/tcp_input.c
+++ b/src/vnet/tcp/tcp_input.c
@@ -211,8 +211,6 @@ tcp_options_parse (tcp_header_t * th, tcp_options_t * to)
 always_inline int
 tcp_segment_check_paws (tcp_connection_t * tc)
 {
-  /* XXX normally test for timestamp should be lt instead of leq, but for
-   * local testing this is not enough */
   return tcp_opts_tstamp (&tc->opt) && tc->tsval_recent
     && timestamp_lt (tc->opt.tsval, tc->tsval_recent);
 }
@@ -999,7 +997,7 @@ tcp_session_enqueue_ooo (tcp_connection_t * tc, vlib_buffer_t * b,
 			 u16 data_len)
 {
   stream_session_t *s0;
-  u32 offset, seq;
+  u32 offset;
   int rv;
 
   /* Pure ACK. Do nothing */
@@ -1009,8 +1007,9 @@ tcp_session_enqueue_ooo (tcp_connection_t * tc, vlib_buffer_t * b,
     }
 
   s0 = stream_session_get (tc->c_s_index, tc->c_thread_index);
-  seq = vnet_buffer (b)->tcp.seq_number;
-  offset = seq - tc->rcv_nxt;
+  offset = vnet_buffer (b)->tcp.seq_number - tc->irs;
+
+  clib_warning ("ooo: offset %d len %d", offset, data_len);
 
   rv = svm_fifo_enqueue_with_offset (s0->server_rx_fifo, s0->pid, offset,
 				     data_len, vlib_buffer_get_current (b));
@@ -1032,8 +1031,8 @@ tcp_session_enqueue_ooo (tcp_connection_t * tc, vlib_buffer_t * b,
 
       /* Get the newest segment from the fifo */
       newest = svm_fifo_newest_ooo_segment (s0->server_rx_fifo);
-      start = tc->rcv_nxt + ooo_segment_offset (s0->server_rx_fifo, newest);
-      end = tc->rcv_nxt + ooo_segment_end_offset (s0->server_rx_fifo, newest);
+      start = ooo_segment_offset (s0->server_rx_fifo, newest);
+      end = ooo_segment_end_offset (s0->server_rx_fifo, newest);
 
       tcp_update_sack_list (tc, start, end);
     }
@@ -1072,6 +1071,7 @@ tcp_segment_rcv (tcp_main_t * tm, tcp_connection_t * tc, vlib_buffer_t * b,
     {
       /* Old sequence numbers allowed through because they overlapped
        * the rx window */
+
       if (seq_lt (vnet_buffer (b)->tcp.seq_number, tc->rcv_nxt))
 	{
 	  error = TCP_ERROR_SEGMENT_OLD;
@@ -1181,6 +1181,7 @@ tcp46_established_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
   u32 n_left_from, next_index, *from, *to_next;
   u32 my_thread_index = vm->thread_index, errors = 0;
   tcp_main_t *tm = vnet_get_tcp_main ();
+  u8 is_fin = 0;
 
   from = vlib_frame_vector_args (from_frame);
   n_left_from = from_frame->n_vectors;
@@ -1243,9 +1244,11 @@ tcp46_established_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
 	      n_advance_bytes0 += sizeof (ip60[0]);
 	    }
 
+	  is_fin = (th0->flags & TCP_FLAG_FIN) != 0;
+
 	  /* SYNs, FINs and data consume sequence numbers */
 	  vnet_buffer (b0)->tcp.seq_end = vnet_buffer (b0)->tcp.seq_number
-	    + tcp_is_syn (th0) + tcp_is_fin (th0) + n_data_bytes0;
+	    + tcp_is_syn (th0) + is_fin + n_data_bytes0;
 
 	  /* TODO header prediction fast path */
 
@@ -1272,8 +1275,11 @@ tcp46_established_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
 	  vlib_buffer_advance (b0, n_advance_bytes0);
 	  error0 = tcp_segment_rcv (tm, tc0, b0, n_data_bytes0, &next0);
 
+	  /* N.B. buffer is rewritten if segment is ooo. Thus, th0 becomes a
+	   * dangling reference. */
+
 	  /* 8: check the FIN bit */
-	  if (tcp_fin (th0))
+	  if (is_fin)
 	    {
 	      /* Enter CLOSE-WAIT and notify session. Don't send ACK, instead
 	       * wait for session to call close. To avoid lingering
@@ -2365,8 +2371,12 @@ tcp46_input_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
 
 	      if (PREDICT_FALSE (error0 == TCP_ERROR_DISPATCH))
 		{
+		  tcp_state_t state0 = tc0->state;
 		  /* Overload tcp flags to store state */
 		  vnet_buffer (b0)->tcp.flags = tc0->state;
+		  clib_warning ("disp error state %U flags %U",
+				format_tcp_state, &state0,
+				format_tcp_flags, flags0);
 		}
 	    }
 	  else
diff --git a/src/vnet/tcp/tcp_test.c b/src/vnet/tcp/tcp_test.c
index 3dbbdf6f..12579632 100644
--- a/src/vnet/tcp/tcp_test.c
+++ b/src/vnet/tcp/tcp_test.c
@@ -173,17 +173,145 @@ tcp_test_sack ()
   return 0;
 }
 
-static int
-tcp_test_fifo (vlib_main_t * vm, unformat_input_t * input)
+typedef struct
+{
+  u32 offset;
+  u32 len;
+} test_pattern_t;
+
+/* *INDENT-OFF* */
+test_pattern_t test_pattern[] = {
+  {380, 8}, {768, 8}, {1156, 8}, {1544, 8}, {1932, 8}, {2320, 8}, {2708, 8},
+  {2992, 8}, {372, 8}, {760, 8}, {1148, 8}, {1536, 8}, {1924, 8}, {2312, 8},
+  {2700, 8}, {2984, 8}, {364, 8}, {752, 8}, {1140, 8}, {1528, 8}, {1916, 8},
+  {2304, 8}, {2692, 8}, {2976, 8}, {356, 8}, {744, 8}, {1132, 8}, {1520, 8},
+  {1908, 8}, {2296, 8}, {2684, 8}, {2968, 8}, {348, 8}, {736, 8}, {1124, 8},
+  {1512, 8}, {1900, 8}, {2288, 8}, {2676, 8}, {2960, 8}, {340, 8}, {728, 8},
+  {1116, 8}, {1504, 8}, {1892, 8}, {2280, 8}, {2668, 8}, {2952, 8}, {332, 8},
+  {720, 8}, {1108, 8}, {1496, 8}, {1884, 8}, {2272, 8}, {2660, 8}, {2944, 8},
+  {324, 8}, {712, 8}, {1100, 8}, {1488, 8}, {1876, 8}, {2264, 8}, {2652, 8},
+  {2936, 8}, {316, 8}, {704, 8}, {1092, 8}, {1480, 8}, {1868, 8}, {2256, 8},
+  {2644, 8}, {2928, 8}, {308, 8}, {696, 8}, {1084, 8}, {1472, 8}, {1860, 8},
+  {2248, 8}, {2636, 8}, {2920, 8}, {300, 8}, {688, 8}, {1076, 8}, {1464, 8},
+  {1852, 8}, {2240, 8}, {2628, 8}, {2912, 8}, {292, 8}, {680, 8}, {1068, 8},
+  {1456, 8}, {1844, 8}, {2232, 8}, {2620, 8}, {2904, 8}, {284, 8}, {672, 8},
+  {1060, 8}, {1448, 8}, {1836, 8}, {2224, 8}, {2612, 8}, {2896, 8}, {276, 8},
+  {664, 8}, {1052, 8}, {1440, 8}, {1828, 8},  {2216, 8}, {2604, 8}, {2888, 8},
+  {268, 8}, {656, 8}, {1044, 8}, {1432, 8}, {1820, 8}, {2208, 8}, {2596, 8},
+  {2880, 8}, {260, 8}, {648, 8}, {1036, 8}, {1424, 8}, {1812, 8}, {2200, 8},
+  {2588, 8}, {2872, 8}, {252, 8}, {640, 8}, {1028, 8}, {1416, 8}, {1804, 8},
+  {2192, 8}, {2580, 8}, {2864, 8}, {244, 8}, {632, 8}, {1020, 8}, {1408, 8},
+  {1796, 8}, {2184, 8}, {2572, 8}, {2856, 8}, {236, 8}, {624, 8}, {1012, 8},
+  {1400, 8}, {1788, 8}, {2176, 8}, {2564, 8}, {2848, 8}, {228, 8}, {616, 8},
+  {1004, 8}, {1392, 8}, {1780, 8}, {2168, 8}, {2556, 8}, {2840, 8}, {220, 8},
+  {608, 8}, {996, 8}, {1384, 8}, {1772, 8}, {2160, 8}, {2548, 8}, {2832, 8},
+  {212, 8}, {600, 8}, {988, 8}, {1376, 8}, {1764, 8}, {2152, 8}, {2540, 8},
+  {2824, 8}, {204, 8}, {592, 8}, {980, 8}, {1368, 8}, {1756, 8}, {2144, 8},
+  {2532, 8}, {2816, 8}, {196, 8}, {584, 8}, {972, 8}, {1360, 8}, {1748, 8},
+  {2136, 8}, {2524, 8}, {2808, 8}, {188, 8}, {576, 8}, {964, 8}, {1352, 8},
+  {1740, 8}, {2128, 8}, {2516, 8}, {2800, 8}, {180, 8}, {568, 8}, {956, 8},
+  {1344, 8}, {1732, 8}, {2120, 8}, {2508, 8}, {2792, 8}, {172, 8}, {560, 8},
+  {948, 8}, {1336, 8}, {1724, 8}, {2112, 8}, {2500, 8}, {2784, 8}, {164, 8},
+  {552, 8}, {940, 8}, {1328, 8}, {1716, 8}, {2104, 8}, {2492, 8}, {2776, 8},
+  {156, 8}, {544, 8}, {932, 8}, {1320, 8}, {1708, 8}, {2096, 8}, {2484, 8},
+  {2768, 8}, {148, 8}, {536, 8}, {924, 8}, {1312, 8}, {1700, 8}, {2088, 8},
+  {2476, 8}, {2760, 8}, {140, 8}, {528, 8}, {916, 8}, {1304, 8}, {1692, 8},
+  {2080, 8}, {2468, 8}, {2752, 8}, {132, 8}, {520, 8}, {908, 8}, {1296, 8},
+  {1684, 8}, {2072, 8}, {2460, 8}, {2744, 8}, {124, 8}, {512, 8}, {900, 8},
+  {1288, 8}, {1676, 8}, {2064, 8}, {2452, 8}, {2736, 8}, {116, 8}, {504, 8},
+  {892, 8}, {1280, 8}, {1668, 8}, {2056, 8}, {2444, 8}, {2728, 8}, {108, 8},
+  {496, 8}, {884, 8}, {1272, 8}, {1660, 8}, {2048, 8}, {2436, 8}, {2720, 8},
+  {100, 8}, {488, 8}, {876, 8}, {1264, 8}, {1652, 8}, {2040, 8}, {2428, 8},
+  {2716, 4}, {92, 8}, {480, 8}, {868, 8}, {1256, 8}, {1644, 8}, {2032, 8},
+  {2420, 8}, {84, 8}, {472, 8}, {860, 8}, {1248, 8}, {1636, 8}, {2024, 8},
+  {2412, 8}, {76, 8}, {464, 8}, {852, 8}, {1240, 8}, {1628, 8}, {2016, 8},
+  {2404, 8}, {68, 8}, {456, 8}, {844, 8}, {1232, 8}, {1620, 8}, {2008, 8},
+  {2396, 8}, {60, 8}, {448, 8}, {836, 8}, {1224, 8}, {1612, 8}, {2000, 8},
+  {2388, 8}, {52, 8}, {440, 8}, {828, 8}, {1216, 8}, {1604, 8}, {1992, 8},
+  {2380, 8}, {44, 8}, {432, 8}, {820, 8}, {1208, 8}, {1596, 8}, {1984, 8},
+  {2372, 8}, {36, 8}, {424, 8}, {812, 8}, {1200, 8}, {1588, 8}, {1976, 8},
+  {2364, 8}, {28, 8}, {416, 8}, {804, 8}, {1192, 8}, {1580, 8}, {1968, 8},
+  {2356, 8}, {20, 8}, {408, 8}, {796, 8}, {1184, 8}, {1572, 8}, {1960, 8},
+  {2348, 8}, {12, 8}, {400, 8}, {788, 8}, {1176, 8}, {1564, 8}, {1952, 8},
+  {2340, 8}, {4, 8}, {392, 8}, {780, 8}, {1168, 8}, {1556, 8}, {1944, 8},
+  {2332, 8},
+  /* missing from original data set */
+  {388, 4}, {776, 4}, {1164, 4}, {1552, 4}, {1940, 4}, {2328, 4},
+};
+/* *INDENT-ON* */
+
+int
+pattern_cmp (const void *arg1, const void *arg2)
+{
+  test_pattern_t *a1 = (test_pattern_t *) arg1;
+  test_pattern_t *a2 = (test_pattern_t *) arg2;
+
+  if (a1->offset < a2->offset)
+    return -1;
+  else if (a1->offset > a2->offset)
+    return 1;
+  return 0;
+}
+
+static u8
+fifo_validate_pattern (vlib_main_t * vm, test_pattern_t * pattern,
+		       u32 pattern_length)
+{
+  test_pattern_t *tp = pattern;
+  int i;
+
+  /* Go through the pattern and make 100% sure it's sane */
+  for (i = 0; i < pattern_length - 1; i++)
+    {
+      if (tp->offset + tp->len != (tp + 1)->offset)
+	{
+	  vlib_cli_output (vm, "[%d] missing {%d, %d}", i,
+			   (tp->offset + tp->len),
+			   (tp + 1)->offset - (tp->offset + tp->len));
+	  return 0;
+	}
+      tp++;
+    }
+  return 1;
+}
+
+static test_pattern_t *
+fifo_get_validate_pattern (vlib_main_t * vm, test_pattern_t * test_data,
+			   u32 test_data_len)
+{
+  test_pattern_t *validate_pattern = 0;
+
+  /* Validate, and try segments in order... */
+  vec_validate (validate_pattern, test_data_len - 1);
+  memcpy (validate_pattern, test_data,
+	  test_data_len * sizeof (test_pattern_t));
+  qsort ((u8 *) validate_pattern, test_data_len, sizeof (test_pattern_t),
+	 pattern_cmp);
+
+  if (fifo_validate_pattern (vm, validate_pattern, test_data_len) == 0)
+    return 0;
+
+  return validate_pattern;
+}
+
+int
+tcp_test_fifo1 (vlib_main_t * vm, unformat_input_t * input)
 {
   svm_fifo_t *f;
   u32 fifo_size = 1 << 20;
   u32 *test_data = 0;
   u32 offset;
-  int i, rv;
+  int i, rv, verbose = 0;
   u32 data_word, test_data_len;
+  ooo_segment_t *ooo_seg;
+  u8 *data;
+
+  while (unformat_check_input (input) != UNFORMAT_END_OF_INPUT)
+    {
+      if (unformat (input, "verbose"))
+	verbose = 1;
+    }
 
-  /* $$$ parse args */
   test_data_len = fifo_size / sizeof (u32);
   vec_validate (test_data, test_data_len - 1);
 
@@ -198,12 +326,8 @@ tcp_test_fifo (vlib_main_t * vm, unformat_input_t * input)
   /* Enqueue an initial (un-dequeued) chunk */
   rv = svm_fifo_enqueue_nowait (f, 0 /* pid */ ,
 				sizeof (u32), (u8 *) test_data);
-
-  if (rv != sizeof (u32))
-    {
-      clib_warning ("enqueue returned %d", rv);
-      goto out;
-    }
+  TCP_TEST ((rv == sizeof (u32)), "enqueued %d", rv);
+  TCP_TEST ((f->tail == 4), "fifo tail %u", f->tail);
 
   /*
    * Create 3 chunks in the future. The offsets are relative
@@ -212,51 +336,62 @@ tcp_test_fifo (vlib_main_t * vm, unformat_input_t * input)
   for (i = 0; i < 3; i++)
     {
       offset = (2 * i + 1) * sizeof (u32);
-      vlib_cli_output (vm, "add offset %d", offset);
-
-      rv = svm_fifo_enqueue_with_offset
-	(f, 0 /* pid */ , offset, sizeof (u32),
-	 (u8 *) (test_data + ((offset + sizeof (u32)) / sizeof (u32))));
-
+      data = (u8 *) (test_data + (2 * i + 1));
+      rv =
+	svm_fifo_enqueue_with_offset (f, 0 /* pid */ , offset, sizeof (u32),
+				      data);
+      if (verbose)
+	vlib_cli_output (vm, "add [%d] [%d, %d]", 2 * i + 1, offset,
+			 offset + sizeof (u32));
       if (rv)
 	{
 	  clib_warning ("enqueue returned %d", rv);
-	  goto out;
+	  goto err;
 	}
     }
 
-  /* Paint missing data backwards */
-  for (i = 3; i > 0; i--)
+  if (verbose)
+    vlib_cli_output (vm, "fifo after odd segs: %U", format_svm_fifo, f, 1);
+  TCP_TEST ((f->tail == 8), "fifo tail %u", f->tail);
+
+  /* Paint some of missing data backwards */
+  for (i = 3; i > 1; i--)
     {
       offset = (2 * i + 0) * sizeof (u32);
-
-      vlib_cli_output (vm, "add offset %d", offset);
-
-      rv = svm_fifo_enqueue_with_offset
-	(f, 0 /* pid */ , offset, sizeof (u32),
-	 (u8 *) (test_data + ((offset + sizeof (u32)) / sizeof (u32))));
-
+      data = (u8 *) (test_data + (2 * i + 0));
+      rv =
+	svm_fifo_enqueue_with_offset (f, 0 /* pid */ , offset, sizeof (u32),
+				      data);
+      if (verbose)
+	vlib_cli_output (vm, "add [%d] [%d, %d]", 2 * i, offset,
+			 offset + sizeof (u32));
       if (rv)
 	{
 	  clib_warning ("enqueue returned %d", rv);
-	  goto out;
+	  goto err;
 	}
     }
 
-  vlib_cli_output (vm, "fifo before missing link: %U",
-		   format_svm_fifo, f, 1 /* verbose */ );
+  if (verbose)
+    vlib_cli_output (vm, "fifo before missing link: %U", format_svm_fifo, f,
+		     1);
+  TCP_TEST ((svm_fifo_number_ooo_segments (f) == 1),
+	    "number of ooo segments %u", svm_fifo_number_ooo_segments (f));
+  ooo_seg = svm_fifo_first_ooo_segment (f);
+  TCP_TEST ((ooo_seg->start == 12),
+	    "first ooo seg position %u", ooo_seg->start);
+  TCP_TEST ((ooo_seg->length == 16),
+	    "first ooo seg length %u", ooo_seg->length);
 
   /* Enqueue the missing u32 */
-  rv = svm_fifo_enqueue_nowait (f, 0 /* pid */ ,
-				sizeof (u32), (u8 *) (test_data + 1));
-  if (rv != 7 * sizeof (u32))
-    {
-      clib_warning ("enqueue returned %d", rv);
-      goto out;
-    }
-
-  vlib_cli_output (vm, "fifo after missing link: %U",
-		   format_svm_fifo, f, 1 /* verbose */ );
+  rv = svm_fifo_enqueue_nowait (f, 0 /* pid */ , sizeof (u32),
+				(u8 *) (test_data + 2));
+  if (verbose)
+    vlib_cli_output (vm, "fifo after missing link: %U", format_svm_fifo, f,
+		     1);
+  TCP_TEST ((rv == 20), "bytes to be enqueued %u", rv);
+  TCP_TEST ((svm_fifo_number_ooo_segments (f) == 0),
+	    "number of ooo segments %u", svm_fifo_number_ooo_segments (f));
 
   /* Collect results */
   for (i = 0; i < 7; i++)
@@ -265,25 +400,316 @@ tcp_test_fifo (vlib_main_t * vm, unformat_input_t * input)
 				    (u8 *) & data_word);
       if (rv != sizeof (u32))
 	{
-	  clib_warning ("dequeue returned %d", rv);
-	  goto out;
+	  clib_warning ("bytes dequeues %u", rv);
+	  goto err;
 	}
       if (data_word != test_data[i])
 	{
-	  clib_warning ("recovered data %d not %d", data_word, test_data[i]);
-	  goto out;
+	  clib_warning ("recovered [%d] %d not %d", i, data_word,
+			test_data[i]);
+	  goto err;
 	}
     }
 
-  clib_warning ("test complete...");
+  svm_fifo_free (f);
+  vec_free (test_data);
+  return 0;
 
-out:
+err:
   svm_fifo_free (f);
   vec_free (test_data);
+  return -1;
+}
+
+static int
+tcp_test_fifo2 (vlib_main_t * vm)
+{
+  svm_fifo_t *f;
+  u32 fifo_size = 1 << 20;
+  int i, rv, test_data_len;
+  u64 data64;
+  test_pattern_t *tp, *vp, *test_data;
+  ooo_segment_t *ooo_seg;
+
+  test_data = test_pattern;
+  test_data_len = ARRAY_LEN (test_pattern);
+
+  vp = fifo_get_validate_pattern (vm, test_data, test_data_len);
+
+  /* Create a fifo */
+  f = svm_fifo_create (fifo_size);
+
+  /* Paint the fifo data vector with -1's */
+  memset (f->data, 0xFF, 1 << 20);
+
+  /*
+   * Try with sorted data
+   */
+  for (i = 0; i < test_data_len; i++)
+    {
+      tp = vp + i;
+      data64 = tp->offset;
+      rv = svm_fifo_enqueue_with_offset (f, 0, tp->offset, tp->len,
+					 (u8 *) & data64);
+    }
+
+  /* Expected result: one big fat chunk at offset 4 */
+  TCP_TEST ((svm_fifo_number_ooo_segments (f) == 1),
+	    "number of ooo segments %u", svm_fifo_number_ooo_segments (f));
+  ooo_seg = svm_fifo_first_ooo_segment (f);
+  TCP_TEST ((ooo_seg->start == 4),
+	    "first ooo seg position %u", ooo_seg->start);
+  TCP_TEST ((ooo_seg->length == 2996),
+	    "first ooo seg length %u", ooo_seg->length);
+
+  data64 = 0;
+  rv = svm_fifo_enqueue_nowait (f, 0, sizeof (u32), (u8 *) & data64);
+  TCP_TEST ((rv == 3000), "bytes to be enqueued %u", rv);
+
+  svm_fifo_free (f);
+  vec_free (vp);
+
+  /*
+   * Now try it again w/ unsorted data...
+   */
+
+  f = svm_fifo_create (fifo_size);
+
+  /* Paint fifo data vector with -1's */
+  memset (f->data, 0xFF, 1 << 20);
+
+  for (i = 0; i < test_data_len; i++)
+    {
+      tp = &test_data[i];
+      data64 = tp->offset;
+      rv = svm_fifo_enqueue_with_offset (f, 0, tp->offset, tp->len,
+					 (u8 *) & data64);
+      if (rv)
+	{
+	  clib_warning ("enqueue returned %d", rv);
+	}
+    }
+
+  /* Expecting the same result: one big fat chunk at offset 4 */
+  TCP_TEST ((svm_fifo_number_ooo_segments (f) == 1),
+	    "number of ooo segments %u", svm_fifo_number_ooo_segments (f));
+  ooo_seg = svm_fifo_first_ooo_segment (f);
+  TCP_TEST ((ooo_seg->start == 4),
+	    "first ooo seg position %u", ooo_seg->start);
+  TCP_TEST ((ooo_seg->length == 2996),
+	    "first ooo seg length %u", ooo_seg->length);
+
+  data64 = 0;
+  rv = svm_fifo_enqueue_nowait (f, 0, sizeof (u32), (u8 *) & data64);
+
+  TCP_TEST ((rv == 3000), "bytes to be enqueued %u", rv);
+
+  svm_fifo_free (f);
+
   return 0;
 }
 
+static int
+tcp_test_fifo3 (vlib_main_t * vm, unformat_input_t * input)
+{
+  svm_fifo_t *f;
+  u32 fifo_size = 4 << 10;
+  u32 fifo_initial_offset = 0;
+  u32 total_size = 2 << 10;
+  int overlap = 0;
+  int i, rv;
+  u8 *data_pattern = 0;
+  test_pattern_t *tp, *generate = 0;
+  u32 nsegs = 2;
+  u32 seg_size, length_so_far;
+  u32 current_offset, offset_increment, len_this_chunk;
+  u32 seed = 0xdeaddabe;
+  int verbose = 0;
+  int randomize = 1;
+
+  while (unformat_check_input (input) != UNFORMAT_END_OF_INPUT)
+    {
+      if (unformat (input, "fifo-size %d", &fifo_size))
+	;
+      else if (unformat (input, "total-size %d", &total_size))
+	;
+      else if (unformat (input, "verbose"))
+	verbose = 1;
+      else if (unformat (input, "overlap"))
+	overlap = 1;
+      else if (unformat (input, "initial-offset %d", &fifo_initial_offset))
+	;
+      else if (unformat (input, "seed %d", &seed))
+	;
+      else if (unformat (input, "nsegs %d", &nsegs))
+	;
+      else if (unformat (input, "no-randomize"))
+	randomize = 0;
+      else
+	{
+	  clib_error_t *e = clib_error_return
+	    (0, "unknown input `%U'", format_unformat_error, input);
+	  clib_error_report (e);
+	  return -1;
+	}
+    }
 
+  /*
+   * Generate data
+   */
+  vec_validate (data_pattern, total_size - 1);
+  for (i = 0; i < vec_len (data_pattern); i++)
+    data_pattern[i] = i & 0xff;
+
+  seg_size = total_size / nsegs;
+  length_so_far = 0;
+  current_offset = 1;
+  while (length_so_far < total_size)
+    {
+      vec_add2 (generate, tp, 1);
+      len_this_chunk = clib_min (seg_size, total_size - length_so_far);
+      tp->offset = current_offset;
+      tp->len = len_this_chunk;
+
+      if (overlap && (len_this_chunk == seg_size))
+	do
+	  {
+	    offset_increment = len_this_chunk
+	      % (1 + (random_u32 (&seed) % len_this_chunk));
+	  }
+	while (offset_increment == 0);
+      else
+	offset_increment = len_this_chunk;
+
+      current_offset += offset_increment;
+      length_so_far = tp->offset + tp->len;
+    }
+
+  /*
+   * Validate segment list. Only valid for non-overlap cases.
+   */
+  if (overlap == 0)
+    fifo_validate_pattern (vm, generate, vec_len (generate));
+
+  if (verbose)
+    {
+      vlib_cli_output (vm, "raw data pattern:");
+      for (i = 0; i < vec_len (generate); i++)
+	{
+	  vlib_cli_output (vm, "[%d] offset %u len %u", i,
+			   generate[i].offset, generate[i].len);
+	}
+    }
+
+  /* Randomize data pattern */
+  if (randomize)
+    {
+      for (i = 0; i < vec_len (generate) / 2; i++)
+	{
+	  u32 src_index, dst_index;
+	  test_pattern_t _tmp, *tmp = &_tmp;
+
+	  src_index = random_u32 (&seed) % vec_len (generate);
+	  dst_index = random_u32 (&seed) % vec_len (generate);
+
+	  tmp[0] = generate[dst_index];
+	  generate[dst_index] = generate[src_index];
+	  generate[src_index] = tmp[0];
+	}
+    }
+
+  if (verbose)
+    {
+      vlib_cli_output (vm, "randomized data pattern:");
+      for (i = 0; i < vec_len (generate); i++)
+	{
+	  vlib_cli_output (vm, "[%d] offset %u len %u", i,
+			   generate[i].offset, generate[i].len);
+	}
+    }
+
+  /* Create a fifo */
+  f = svm_fifo_create (fifo_size);
+
+  /* Paint the fifo data vector with -1's */
+  memset (f->data, 0xFF, fifo_size);
+
+  /* manually set head and tail pointers to validate modular arithmetic */
+  f->head = fifo_initial_offset % fifo_size;
+  f->tail = fifo_initial_offset % fifo_size;
+
+  for (i = 0; i < vec_len (generate); i++)
+    {
+      tp = generate + i;
+      rv = svm_fifo_enqueue_with_offset (f, 0, tp->offset, tp->len,
+					 (u8 *) data_pattern + tp->offset);
+    }
+
+  /* Expected result: one big fat chunk at offset 1 */
+
+  if (verbose)
+    vlib_cli_output (vm, "fifo before missing link: %U",
+		     format_svm_fifo, f, 1 /* verbose */ );
+
+  rv = svm_fifo_enqueue_nowait (f, 0, 1 /* count */ , data_pattern + 0);
+
+  if (verbose)
+    vlib_cli_output (vm, "in-order enqueue returned %d", rv);
+
+  TCP_TEST ((rv == total_size), "retrieved %u expected %u", rv, total_size);
+  TCP_TEST ((svm_fifo_number_ooo_segments (f) == 0),
+	    "number of ooo segments %u", svm_fifo_number_ooo_segments (f));
+  svm_fifo_free (f);
+  vec_free (data_pattern);
+
+  return 0;
+}
+
+static int
+tcp_test_fifo (vlib_main_t * vm, unformat_input_t * input)
+{
+  int res = 0;
+
+  /* Run all tests */
+  if (unformat_check_input (input) == UNFORMAT_END_OF_INPUT)
+    {
+      res = tcp_test_fifo1 (vm, input);
+      if (res)
+	return res;
+
+      res = tcp_test_fifo2 (vm);
+      if (res)
+	return res;
+
+      /* Run a number of fifo3 configs */
+      unformat_init_cstring (input, "nsegs 3 overlap seed 123");
+      if (tcp_test_fifo3 (vm, input))
+	return -1;
+      unformat_free (input);
+
+      unformat_init_cstring (input, "nsegs 10");
+      if (tcp_test_fifo3 (vm, input))
+	return -1;
+      unformat_free (input);
+    }
+  else
+    {
+      if (unformat (input, "fifo3"))
+	{
+	  res = tcp_test_fifo3 (vm, input);
+	}
+      else if (unformat (input, "fifo2"))
+	{
+	  res = tcp_test_fifo2 (vm);
+	}
+      else if (unformat (input, "fifo1"))
+	{
+	  res = tcp_test_fifo1 (vm, input);
+	}
+    }
+
+  return res;
+}
 
 static clib_error_t *
 tcp_test (vlib_main_t * vm,
-- 
cgit 1.2.3-korg


From b59a7055524bae29a328958c493d0b600776dc27 Mon Sep 17 00:00:00 2001
From: Florin Coras <fcoras@cisco.com>
Date: Tue, 18 Apr 2017 22:07:29 -0700
Subject: Add more svm fifo unit tests

Change-Id: Ifc07b3f90ac155c26c3a216e073b474b499ebd44
Signed-off-by: Florin Coras <fcoras@cisco.com>
---
 src/svm/svm_fifo.c      | 100 ++++++++--------
 src/vnet/tcp/tcp_test.c | 305 ++++++++++++++++++++++++++++++++++++++----------
 2 files changed, 292 insertions(+), 113 deletions(-)

(limited to 'src/svm/svm_fifo.c')

diff --git a/src/svm/svm_fifo.c b/src/svm/svm_fifo.c
index bd968aea..f428d3ec 100644
--- a/src/svm/svm_fifo.c
+++ b/src/svm/svm_fifo.c
@@ -38,13 +38,37 @@ format_ooo_list (u8 * s, va_list * args)
   while (ooo_segment_index != OOO_SEGMENT_INVALID_INDEX)
     {
       seg = pool_elt_at_index (f->ooo_segments, ooo_segment_index);
-      s = format (s, "\n  %U", format_ooo_segment, seg);
-
+      s = format (s, "  %U\n", format_ooo_segment, seg);
       ooo_segment_index = seg->next;
     }
   return s;
 }
 
+u8 *
+format_svm_fifo (u8 * s, va_list * args)
+{
+  svm_fifo_t *f = va_arg (*args, svm_fifo_t *);
+  int verbose = va_arg (*args, int);
+
+  s = format (s, "cursize %u nitems %u has_event %d\n",
+	      f->cursize, f->nitems, f->has_event);
+  s = format (s, "head %d tail %d\n", f->head, f->tail);
+
+  if (verbose > 1)
+    s = format
+      (s, "server session %d thread %d client session %d thread %d\n",
+       f->server_session_index, f->server_thread_index,
+       f->client_session_index, f->client_thread_index);
+
+  if (verbose)
+    {
+      s = format (s, "ooo pool %d active elts\n",
+		  pool_elts (f->ooo_segments));
+      s = format (s, "%U", format_ooo_list, f);
+    }
+  return s;
+}
+
 /** create an svm fifo, in the current heap. Fails vs blow up the process */
 svm_fifo_t *
 svm_fifo_create (u32 data_size_in_bytes)
@@ -274,33 +298,41 @@ ooo_segment_try_collect (svm_fifo_t * f, u32 n_bytes_enqueued)
   u32 index, bytes = 0, diff;
   u32 cursize;
 
-  /* read cursize, which can only increase while we're working */
-  cursize = svm_fifo_max_dequeue (f);
+  /* current size has not yet been updated */
+  cursize = svm_fifo_max_dequeue (f) + n_bytes_enqueued;
 
   s = pool_elt_at_index (f->ooo_segments, f->ooos_list_head);
 
+  diff = (f->nitems + (i32) (f->tail - s->start)) % f->nitems;
+  if (diff > cursize)
+    return 0;
+
   /* If last tail update overlaps one/multiple ooo segments, remove them */
-  diff = (f->nitems + ((int) s->start - f->tail)) % f->nitems;
-  while (0 < diff && diff < n_bytes_enqueued)
+  while (0 < diff && diff < cursize)
     {
-      /* Segment end is beyond the tail. Advance tail and be done */
+      index = s - f->ooo_segments;
+
+      /* Segment end is beyond the tail. Advance tail and remove segment */
       if (diff < s->length)
 	{
 	  f->tail += s->length - diff;
 	  f->tail %= f->nitems;
+	  bytes = s->length - diff;
+	  ooo_segment_del (f, index);
 	  break;
 	}
+
       /* If we have next go on */
-      else if (s->next != OOO_SEGMENT_INVALID_INDEX)
+      if (s->next != OOO_SEGMENT_INVALID_INDEX)
 	{
-	  index = s - f->ooo_segments;
 	  s = pool_elt_at_index (f->ooo_segments, s->next);
-	  diff = (f->nitems + ((int) s->start - f->tail)) % f->nitems;
+	  diff = (f->nitems + (i32) (f->tail - s->start)) % f->nitems;
 	  ooo_segment_del (f, index);
 	}
       /* End of search */
       else
 	{
+	  ooo_segment_del (f, index);
 	  break;
 	}
     }
@@ -404,10 +436,7 @@ svm_fifo_enqueue_with_offset_internal (svm_fifo_t * f,
   u32 normalized_offset;
   int rv;
 
-  /* Safety: don't wrap more than nitems/2 */
-  ASSERT ((f->nitems + offset - f->tail) % f->nitems < f->nitems / 2);
-
-  /* Users would do do well to avoid this */
+  /* Users would do well to avoid this */
   if (PREDICT_FALSE (f->tail == (offset % f->nitems)))
     {
       rv = svm_fifo_enqueue_internal (f, pid, required_bytes, copy_from_here);
@@ -421,7 +450,7 @@ svm_fifo_enqueue_with_offset_internal (svm_fifo_t * f,
   nitems = f->nitems;
 
   /* Will this request fit? */
-  if ((required_bytes + offset) > (nitems - cursize))
+  if ((required_bytes + (offset - f->tail) % nitems) > (nitems - cursize))
     return -1;
 
   ooo_segment_add (f, offset, required_bytes);
@@ -523,7 +552,7 @@ svm_fifo_dequeue_nowait (svm_fifo_t * f,
 }
 
 int
-svm_fifo_peek (svm_fifo_t * f, int pid, u32 offset, u32 max_bytes,
+svm_fifo_peek (svm_fifo_t * f, int pid, u32 relative_offset, u32 max_bytes,
 	       u8 * copy_here)
 {
   u32 total_copy_bytes, first_copy_bytes, second_copy_bytes;
@@ -535,7 +564,7 @@ svm_fifo_peek (svm_fifo_t * f, int pid, u32 offset, u32 max_bytes,
     return -2;			/* nothing in the fifo */
 
   nitems = f->nitems;
-  real_head = f->head + offset;
+  real_head = f->head + relative_offset;
   real_head = real_head >= nitems ? real_head - nitems : real_head;
 
   /* Number of bytes we're going to copy */
@@ -596,43 +625,6 @@ svm_fifo_dequeue_drop (svm_fifo_t * f, int pid, u32 max_bytes)
   return total_drop_bytes;
 }
 
-u8 *
-format_svm_fifo (u8 * s, va_list * args)
-{
-  svm_fifo_t *f = va_arg (*args, svm_fifo_t *);
-  int verbose = va_arg (*args, int);
-
-  s = format (s, "cursize %u nitems %u has_event %d\n",
-	      f->cursize, f->nitems, f->has_event);
-  s = format (s, "head %d tail %d\n", f->head, f->tail);
-
-  if (verbose > 1)
-    s = format
-      (s, "server session %d thread %d client session %d thread %d\n",
-       f->server_session_index, f->server_thread_index,
-       f->client_session_index, f->client_thread_index);
-
-  if (verbose)
-    {
-      ooo_segment_t *seg;
-      u32 seg_index;
-
-      s = format (s, "ooo pool %d active elts\n",
-		  pool_elts (f->ooo_segments));
-
-      seg_index = f->ooos_list_head;
-
-      while (seg_index != OOO_SEGMENT_INVALID_INDEX)
-	{
-	  seg = pool_elt_at_index (f->ooo_segments, seg_index);
-	  s = format (s, "  pos %u, len %u next %d\n",
-		      seg->start, seg->length, seg->next);
-	  seg_index = seg->next;
-	}
-    }
-  return s;
-}
-
 u32
 svm_fifo_number_ooo_segments (svm_fifo_t * f)
 {
diff --git a/src/vnet/tcp/tcp_test.c b/src/vnet/tcp/tcp_test.c
index 12579632..890e50b9 100644
--- a/src/vnet/tcp/tcp_test.c
+++ b/src/vnet/tcp/tcp_test.c
@@ -294,6 +294,34 @@ fifo_get_validate_pattern (vlib_main_t * vm, test_pattern_t * test_data,
   return validate_pattern;
 }
 
+static svm_fifo_t *
+fifo_prepare (u32 fifo_size)
+{
+  svm_fifo_t *f;
+  f = svm_fifo_create (fifo_size);
+
+  /* Paint fifo data vector with -1's */
+  memset (f->data, 0xFF, fifo_size);
+
+  return f;
+}
+
+static int
+compare_data (u8 * data1, u8 * data2, u32 start, u32 len, u32 * index)
+{
+  int i;
+
+  for (i = start; i < len; i++)
+    {
+      if (data1[i] != data2[i])
+	{
+	  *index = i;
+	  return 1;
+	}
+    }
+  return 0;
+}
+
 int
 tcp_test_fifo1 (vlib_main_t * vm, unformat_input_t * input)
 {
@@ -302,9 +330,9 @@ tcp_test_fifo1 (vlib_main_t * vm, unformat_input_t * input)
   u32 *test_data = 0;
   u32 offset;
   int i, rv, verbose = 0;
-  u32 data_word, test_data_len;
+  u32 data_word, test_data_len, j;
   ooo_segment_t *ooo_seg;
-  u8 *data;
+  u8 *data, *s, *data_buf = 0;
 
   while (unformat_check_input (input) != UNFORMAT_END_OF_INPUT)
     {
@@ -318,12 +346,11 @@ tcp_test_fifo1 (vlib_main_t * vm, unformat_input_t * input)
   for (i = 0; i < vec_len (test_data); i++)
     test_data[i] = i;
 
-  f = svm_fifo_create (fifo_size);
-
-  /* Paint fifo data vector with -1's */
-  memset (f->data, 0xFF, test_data_len);
+  f = fifo_prepare (fifo_size);
 
-  /* Enqueue an initial (un-dequeued) chunk */
+  /*
+   * Enqueue an initial (un-dequeued) chunk
+   */
   rv = svm_fifo_enqueue_nowait (f, 0 /* pid */ ,
 				sizeof (u32), (u8 *) test_data);
   TCP_TEST ((rv == sizeof (u32)), "enqueued %d", rv);
@@ -337,9 +364,7 @@ tcp_test_fifo1 (vlib_main_t * vm, unformat_input_t * input)
     {
       offset = (2 * i + 1) * sizeof (u32);
       data = (u8 *) (test_data + (2 * i + 1));
-      rv =
-	svm_fifo_enqueue_with_offset (f, 0 /* pid */ , offset, sizeof (u32),
-				      data);
+      rv = svm_fifo_enqueue_with_offset (f, 0, offset, sizeof (u32), data);
       if (verbose)
 	vlib_cli_output (vm, "add [%d] [%d, %d]", 2 * i + 1, offset,
 			 offset + sizeof (u32));
@@ -352,16 +377,23 @@ tcp_test_fifo1 (vlib_main_t * vm, unformat_input_t * input)
 
   if (verbose)
     vlib_cli_output (vm, "fifo after odd segs: %U", format_svm_fifo, f, 1);
+
   TCP_TEST ((f->tail == 8), "fifo tail %u", f->tail);
 
-  /* Paint some of missing data backwards */
+  /*
+   * Make sure format functions are not buggy
+   */
+  s = format (0, "%U", format_svm_fifo, f, 2);
+  vec_free (s);
+
+  /*
+   * Paint some of missing data backwards
+   */
   for (i = 3; i > 1; i--)
     {
       offset = (2 * i + 0) * sizeof (u32);
       data = (u8 *) (test_data + (2 * i + 0));
-      rv =
-	svm_fifo_enqueue_with_offset (f, 0 /* pid */ , offset, sizeof (u32),
-				      data);
+      rv = svm_fifo_enqueue_with_offset (f, 0, offset, sizeof (u32), data);
       if (verbose)
 	vlib_cli_output (vm, "add [%d] [%d, %d]", 2 * i, offset,
 			 offset + sizeof (u32));
@@ -383,7 +415,9 @@ tcp_test_fifo1 (vlib_main_t * vm, unformat_input_t * input)
   TCP_TEST ((ooo_seg->length == 16),
 	    "first ooo seg length %u", ooo_seg->length);
 
-  /* Enqueue the missing u32 */
+  /*
+   * Enqueue the missing u32
+   */
   rv = svm_fifo_enqueue_nowait (f, 0 /* pid */ , sizeof (u32),
 				(u8 *) (test_data + 2));
   if (verbose)
@@ -393,7 +427,9 @@ tcp_test_fifo1 (vlib_main_t * vm, unformat_input_t * input)
   TCP_TEST ((svm_fifo_number_ooo_segments (f) == 0),
 	    "number of ooo segments %u", svm_fifo_number_ooo_segments (f));
 
-  /* Collect results */
+  /*
+   * Collect results
+   */
   for (i = 0; i < 7; i++)
     {
       rv = svm_fifo_dequeue_nowait (f, 0 /* pid */ , sizeof (u32),
@@ -411,8 +447,77 @@ tcp_test_fifo1 (vlib_main_t * vm, unformat_input_t * input)
 	}
     }
 
+  /*
+   * Test segment overlaps: last ooo segment overlaps all
+   */
+  svm_fifo_free (f);
+  f = fifo_prepare (fifo_size);
+
+  for (i = 0; i < 4; i++)
+    {
+      offset = (2 * i + 1) * sizeof (u32);
+      data = (u8 *) (test_data + (2 * i + 1));
+      rv = svm_fifo_enqueue_with_offset (f, 0, offset, sizeof (u32), data);
+      if (verbose)
+	vlib_cli_output (vm, "add [%d] [%d, %d]", 2 * i + 1, offset,
+			 offset + sizeof (u32));
+      if (rv)
+	{
+	  clib_warning ("enqueue returned %d", rv);
+	  goto err;
+	}
+    }
+
+  rv = svm_fifo_enqueue_with_offset (f, 0, 8, 21, data);
+  TCP_TEST ((rv == 0), "ooo enqueued %u", rv);
+  TCP_TEST ((svm_fifo_number_ooo_segments (f) == 1),
+	    "number of ooo segments %u", svm_fifo_number_ooo_segments (f));
+
+  vec_validate (data_buf, vec_len (data));
+  svm_fifo_peek (f, 0, 0, vec_len (data), data_buf);
+  if (compare_data (data_buf, data, 8, vec_len (data), &j))
+    {
+      TCP_TEST (0, "[%d] peeked %u expected %u", j, data_buf[j], data[j]);
+    }
+  vec_reset_length (data_buf);
+
+  /*
+   * Test segment overlaps: enqueue and overlap ooo segments
+   */
+  svm_fifo_free (f);
+  f = fifo_prepare (fifo_size);
+
+  for (i = 0; i < 4; i++)
+    {
+      offset = (2 * i + 1) * sizeof (u32);
+      data = (u8 *) (test_data + (2 * i + 1));
+      rv = svm_fifo_enqueue_with_offset (f, 0, offset, sizeof (u32), data);
+      if (verbose)
+	vlib_cli_output (vm, "add [%d] [%d, %d]", 2 * i + 1, offset,
+			 offset + sizeof (u32));
+      if (rv)
+	{
+	  clib_warning ("enqueue returned %d", rv);
+	  goto err;
+	}
+    }
+
+  rv = svm_fifo_enqueue_nowait (f, 0, 29, data);
+  TCP_TEST ((rv == 32), "ooo enqueued %u", rv);
+  TCP_TEST ((svm_fifo_number_ooo_segments (f) == 0),
+	    "number of ooo segments %u", svm_fifo_number_ooo_segments (f));
+
+  vec_validate (data_buf, vec_len (data));
+  svm_fifo_peek (f, 0, 0, vec_len (data), data_buf);
+  if (compare_data (data_buf, data, 0, vec_len (data), &j))
+    {
+      TCP_TEST (0, "[%d] peeked %u expected %u", j, data_buf[j], data[j]);
+    }
+
+  vec_free (data_buf);
   svm_fifo_free (f);
   vec_free (test_data);
+
   return 0;
 
 err:
@@ -437,10 +542,7 @@ tcp_test_fifo2 (vlib_main_t * vm)
   vp = fifo_get_validate_pattern (vm, test_data, test_data_len);
 
   /* Create a fifo */
-  f = svm_fifo_create (fifo_size);
-
-  /* Paint the fifo data vector with -1's */
-  memset (f->data, 0xFF, 1 << 20);
+  f = fifo_prepare (fifo_size);
 
   /*
    * Try with sorted data
@@ -473,10 +575,7 @@ tcp_test_fifo2 (vlib_main_t * vm)
    * Now try it again w/ unsorted data...
    */
 
-  f = svm_fifo_create (fifo_size);
-
-  /* Paint fifo data vector with -1's */
-  memset (f->data, 0xFF, 1 << 20);
+  f = fifo_prepare (fifo_size);
 
   for (i = 0; i < test_data_len; i++)
     {
@@ -516,16 +615,13 @@ tcp_test_fifo3 (vlib_main_t * vm, unformat_input_t * input)
   u32 fifo_size = 4 << 10;
   u32 fifo_initial_offset = 0;
   u32 total_size = 2 << 10;
-  int overlap = 0;
-  int i, rv;
-  u8 *data_pattern = 0;
+  int overlap = 0, verbose = 0, randomize = 1, drop = 0, in_seq_all = 0;
+  u8 *data_pattern = 0, *data_buf = 0;
   test_pattern_t *tp, *generate = 0;
-  u32 nsegs = 2;
-  u32 seg_size, length_so_far;
+  u32 nsegs = 2, seg_size, length_so_far;
   u32 current_offset, offset_increment, len_this_chunk;
-  u32 seed = 0xdeaddabe;
-  int verbose = 0;
-  int randomize = 1;
+  u32 seed = 0xdeaddabe, j;
+  int i, rv;
 
   while (unformat_check_input (input) != UNFORMAT_END_OF_INPUT)
     {
@@ -545,6 +641,10 @@ tcp_test_fifo3 (vlib_main_t * vm, unformat_input_t * input)
 	;
       else if (unformat (input, "no-randomize"))
 	randomize = 0;
+      else if (unformat (input, "in-seq-all"))
+	in_seq_all = 1;
+      else if (unformat (input, "drop"))
+	drop = 1;
       else
 	{
 	  clib_error_t *e = clib_error_return
@@ -554,6 +654,18 @@ tcp_test_fifo3 (vlib_main_t * vm, unformat_input_t * input)
 	}
     }
 
+  if (total_size > fifo_size)
+    {
+      clib_warning ("total_size %d greater than fifo size %d", total_size,
+		    fifo_size);
+      return -1;
+    }
+  if (overlap && randomize == 0)
+    {
+      clib_warning ("Can't enqueue in-order with overlap");
+      return -1;
+    }
+
   /*
    * Generate data
    */
@@ -561,9 +673,12 @@ tcp_test_fifo3 (vlib_main_t * vm, unformat_input_t * input)
   for (i = 0; i < vec_len (data_pattern); i++)
     data_pattern[i] = i & 0xff;
 
+  /*
+   * Generate segments
+   */
   seg_size = total_size / nsegs;
   length_so_far = 0;
-  current_offset = 1;
+  current_offset = randomize;
   while (length_so_far < total_size)
     {
       vec_add2 (generate, tp, 1);
@@ -616,51 +731,100 @@ tcp_test_fifo3 (vlib_main_t * vm, unformat_input_t * input)
 	  generate[dst_index] = generate[src_index];
 	  generate[src_index] = tmp[0];
 	}
-    }
-
-  if (verbose)
-    {
-      vlib_cli_output (vm, "randomized data pattern:");
-      for (i = 0; i < vec_len (generate); i++)
+      if (verbose)
 	{
-	  vlib_cli_output (vm, "[%d] offset %u len %u", i,
-			   generate[i].offset, generate[i].len);
+	  vlib_cli_output (vm, "randomized data pattern:");
+	  for (i = 0; i < vec_len (generate); i++)
+	    {
+	      vlib_cli_output (vm, "[%d] offset %u len %u", i,
+			       generate[i].offset, generate[i].len);
+	    }
 	}
     }
 
-  /* Create a fifo */
-  f = svm_fifo_create (fifo_size);
-
-  /* Paint the fifo data vector with -1's */
-  memset (f->data, 0xFF, fifo_size);
+  /*
+   * Create a fifo and add segments
+   */
+  f = fifo_prepare (fifo_size);
 
   /* manually set head and tail pointers to validate modular arithmetic */
-  f->head = fifo_initial_offset % fifo_size;
-  f->tail = fifo_initial_offset % fifo_size;
+  fifo_initial_offset = fifo_initial_offset % fifo_size;
+  f->head = fifo_initial_offset;
+  f->tail = fifo_initial_offset;
 
   for (i = 0; i < vec_len (generate); i++)
     {
       tp = generate + i;
-      rv = svm_fifo_enqueue_with_offset (f, 0, tp->offset, tp->len,
+      rv = svm_fifo_enqueue_with_offset (f, 0, fifo_initial_offset
+					 + tp->offset, tp->len,
 					 (u8 *) data_pattern + tp->offset);
     }
 
-  /* Expected result: one big fat chunk at offset 1 */
+  /*
+   * Expected result: one big fat chunk at offset 1 if randomize == 1
+   */
 
   if (verbose)
     vlib_cli_output (vm, "fifo before missing link: %U",
 		     format_svm_fifo, f, 1 /* verbose */ );
 
-  rv = svm_fifo_enqueue_nowait (f, 0, 1 /* count */ , data_pattern + 0);
+  /*
+   * Add the missing byte if segments were randomized
+   */
+  if (randomize)
+    {
+      u32 bytes_to_enq = 1;
+      if (in_seq_all)
+	bytes_to_enq = total_size;
+      rv = svm_fifo_enqueue_nowait (f, 0, bytes_to_enq, data_pattern + 0);
+
+      if (verbose)
+	vlib_cli_output (vm, "in-order enqueue returned %d", rv);
 
-  if (verbose)
-    vlib_cli_output (vm, "in-order enqueue returned %d", rv);
+      TCP_TEST ((rv == total_size), "enqueued %u expected %u", rv,
+		total_size);
+
+    }
+
+  TCP_TEST ((svm_fifo_has_ooo_data (f) == 0), "number of ooo segments %u",
+	    svm_fifo_number_ooo_segments (f));
+
+  /*
+   * Test if peeked data is the same as original data
+   */
+  vec_validate (data_buf, vec_len (data_pattern));
+  svm_fifo_peek (f, 0, 0, vec_len (data_pattern), data_buf);
+  if (compare_data (data_buf, data_pattern, 0, vec_len (data_pattern), &j))
+    {
+      TCP_TEST (0, "[%d] peeked %u expected %u", j, data_buf[j],
+		data_pattern[j]);
+    }
+  vec_reset_length (data_buf);
+
+  /*
+   * Dequeue or drop all data
+   */
+  if (drop)
+    {
+      svm_fifo_dequeue_drop (f, 0, vec_len (data_pattern));
+    }
+  else
+    {
+      svm_fifo_dequeue_nowait (f, 0, vec_len (data_pattern), data_buf);
+      if (compare_data
+	  (data_buf, data_pattern, 0, vec_len (data_pattern), &j))
+	{
+	  TCP_TEST (0, "[%d] dequeued %u expected %u", j, data_buf[j],
+		    data_pattern[j]);
+	}
+    }
+
+  TCP_TEST ((svm_fifo_max_dequeue (f) == 0), "fifo has %d bytes",
+	    svm_fifo_max_dequeue (f));
 
-  TCP_TEST ((rv == total_size), "retrieved %u expected %u", rv, total_size);
-  TCP_TEST ((svm_fifo_number_ooo_segments (f) == 0),
-	    "number of ooo segments %u", svm_fifo_number_ooo_segments (f));
   svm_fifo_free (f);
   vec_free (data_pattern);
+  vec_free (data_buf);
 
   return 0;
 }
@@ -669,6 +833,7 @@ static int
 tcp_test_fifo (vlib_main_t * vm, unformat_input_t * input)
 {
   int res = 0;
+  char *str;
 
   /* Run all tests */
   if (unformat_check_input (input) == UNFORMAT_END_OF_INPUT)
@@ -681,13 +846,35 @@ tcp_test_fifo (vlib_main_t * vm, unformat_input_t * input)
       if (res)
 	return res;
 
-      /* Run a number of fifo3 configs */
-      unformat_init_cstring (input, "nsegs 3 overlap seed 123");
+      /*
+       * Run a number of fifo3 configs
+       */
+      str = "nsegs 10 overlap seed 123";
+      unformat_init_cstring (input, str);
+      if (tcp_test_fifo3 (vm, input))
+	return -1;
+      unformat_free (input);
+
+      str = "nsegs 10 overlap seed 123 in-seq-all";
+      unformat_init_cstring (input, str);
+      if (tcp_test_fifo3 (vm, input))
+	return -1;
+      unformat_free (input);
+
+      str = "nsegs 10 overlap seed 123 initial-offset 3917";
+      unformat_init_cstring (input, str);
+      if (tcp_test_fifo3 (vm, input))
+	return -1;
+      unformat_free (input);
+
+      str = "nsegs 10 overlap seed 123 initial-offset 3917 drop";
+      unformat_init_cstring (input, str);
       if (tcp_test_fifo3 (vm, input))
 	return -1;
       unformat_free (input);
 
-      unformat_init_cstring (input, "nsegs 10");
+      str = "nsegs 10 seed 123 initial-offset 3917 drop no-randomize";
+      unformat_init_cstring (input, str);
       if (tcp_test_fifo3 (vm, input))
 	return -1;
       unformat_free (input);
-- 
cgit 1.2.3-korg


From a5464817522c7a7dc760af4612f1d6a68ed0afc8 Mon Sep 17 00:00:00 2001
From: Florin Coras <fcoras@cisco.com>
Date: Wed, 19 Apr 2017 13:00:05 -0700
Subject: Session layer improvements

Among others:
- Moved app event queue to shared memory segment
- Use private memory segment for builtin apps
- Remove pid from svm fifo
- Protect session fifo (de)allocation
- Use fifo event for session disconnects
- Have session queue node poll in all wk threads

Change-Id: I89dbf7fdfebef12f5ef2b34ba3ef3c2c07f49ff2
Signed-off-by: Florin Coras <fcoras@cisco.com>
---
 src/svm/svm_fifo.c                       |  30 ++---
 src/svm/svm_fifo.h                       |  31 ++---
 src/svm/svm_fifo_segment.c               |  50 +++++++-
 src/svm/svm_fifo_segment.h               |   5 +
 src/svm/test_svm_fifo1.c                 |  27 ++---
 src/uri/uri_tcp_test.c                   | 189 ++++++++++++++++++++-----------
 src/uri/uri_udp_test.c                   |  40 ++++---
 src/vnet/session/application.c           |  48 +++-----
 src/vnet/session/application.h           |  12 --
 src/vnet/session/application_interface.c |  26 +----
 src/vnet/session/application_interface.h |  38 ++++++-
 src/vnet/session/node.c                  |  63 ++++++-----
 src/vnet/session/segment_manager.c       | 134 ++++++++++++++++++----
 src/vnet/session/segment_manager.h       |  12 ++
 src/vnet/session/session.c               | 138 ++++++++++++----------
 src/vnet/session/session.h               |  19 ++--
 src/vnet/session/session_api.c           |  58 ++++------
 src/vnet/tcp/builtin_client.c            |   9 +-
 src/vnet/tcp/builtin_server.c            |   8 +-
 src/vnet/tcp/tcp.c                       |  13 ++-
 src/vnet/tcp/tcp_input.c                 |   8 +-
 src/vnet/tcp/tcp_output.c                |   6 -
 src/vnet/tcp/tcp_test.c                  |  43 ++++---
 src/vnet/udp/builtin_server.c            |   8 +-
 src/vnet/udp/udp_input.c                 |   5 +-
 25 files changed, 604 insertions(+), 416 deletions(-)

(limited to 'src/svm/svm_fifo.c')

diff --git a/src/svm/svm_fifo.c b/src/svm/svm_fifo.c
index f428d3ec..8f2ed0c9 100644
--- a/src/svm/svm_fifo.c
+++ b/src/svm/svm_fifo.c
@@ -57,7 +57,7 @@ format_svm_fifo (u8 * s, va_list * args)
   if (verbose > 1)
     s = format
       (s, "server session %d thread %d client session %d thread %d\n",
-       f->server_session_index, f->server_thread_index,
+       f->master_session_index, f->master_thread_index,
        f->client_session_index, f->client_thread_index);
 
   if (verbose)
@@ -353,8 +353,7 @@ ooo_segment_try_collect (svm_fifo_t * f, u32 n_bytes_enqueued)
 }
 
 static int
-svm_fifo_enqueue_internal (svm_fifo_t * f,
-			   int pid, u32 max_bytes, u8 * copy_from_here)
+svm_fifo_enqueue_internal (svm_fifo_t * f, u32 max_bytes, u8 * copy_from_here)
 {
   u32 total_copy_bytes, first_copy_bytes, second_copy_bytes;
   u32 cursize, nitems;
@@ -411,10 +410,9 @@ svm_fifo_enqueue_internal (svm_fifo_t * f,
 }
 
 int
-svm_fifo_enqueue_nowait (svm_fifo_t * f,
-			 int pid, u32 max_bytes, u8 * copy_from_here)
+svm_fifo_enqueue_nowait (svm_fifo_t * f, u32 max_bytes, u8 * copy_from_here)
 {
-  return svm_fifo_enqueue_internal (f, pid, max_bytes, copy_from_here);
+  return svm_fifo_enqueue_internal (f, max_bytes, copy_from_here);
 }
 
 /**
@@ -426,7 +424,6 @@ svm_fifo_enqueue_nowait (svm_fifo_t * f,
  */
 static int
 svm_fifo_enqueue_with_offset_internal (svm_fifo_t * f,
-				       int pid,
 				       u32 offset,
 				       u32 required_bytes,
 				       u8 * copy_from_here)
@@ -439,7 +436,7 @@ svm_fifo_enqueue_with_offset_internal (svm_fifo_t * f,
   /* Users would do well to avoid this */
   if (PREDICT_FALSE (f->tail == (offset % f->nitems)))
     {
-      rv = svm_fifo_enqueue_internal (f, pid, required_bytes, copy_from_here);
+      rv = svm_fifo_enqueue_internal (f, required_bytes, copy_from_here);
       if (rv > 0)
 	return 0;
       return -1;
@@ -484,18 +481,16 @@ svm_fifo_enqueue_with_offset_internal (svm_fifo_t * f,
 
 int
 svm_fifo_enqueue_with_offset (svm_fifo_t * f,
-			      int pid,
 			      u32 offset,
 			      u32 required_bytes, u8 * copy_from_here)
 {
-  return svm_fifo_enqueue_with_offset_internal
-    (f, pid, offset, required_bytes, copy_from_here);
+  return svm_fifo_enqueue_with_offset_internal (f, offset, required_bytes,
+						copy_from_here);
 }
 
 
 static int
-svm_fifo_dequeue_internal (svm_fifo_t * f,
-			   int pid, u32 max_bytes, u8 * copy_here)
+svm_fifo_dequeue_internal (svm_fifo_t * f, u32 max_bytes, u8 * copy_here)
 {
   u32 total_copy_bytes, first_copy_bytes, second_copy_bytes;
   u32 cursize, nitems;
@@ -545,14 +540,13 @@ svm_fifo_dequeue_internal (svm_fifo_t * f,
 }
 
 int
-svm_fifo_dequeue_nowait (svm_fifo_t * f,
-			 int pid, u32 max_bytes, u8 * copy_here)
+svm_fifo_dequeue_nowait (svm_fifo_t * f, u32 max_bytes, u8 * copy_here)
 {
-  return svm_fifo_dequeue_internal (f, pid, max_bytes, copy_here);
+  return svm_fifo_dequeue_internal (f, max_bytes, copy_here);
 }
 
 int
-svm_fifo_peek (svm_fifo_t * f, int pid, u32 relative_offset, u32 max_bytes,
+svm_fifo_peek (svm_fifo_t * f, u32 relative_offset, u32 max_bytes,
 	       u8 * copy_here)
 {
   u32 total_copy_bytes, first_copy_bytes, second_copy_bytes;
@@ -590,7 +584,7 @@ svm_fifo_peek (svm_fifo_t * f, int pid, u32 relative_offset, u32 max_bytes,
 }
 
 int
-svm_fifo_dequeue_drop (svm_fifo_t * f, int pid, u32 max_bytes)
+svm_fifo_dequeue_drop (svm_fifo_t * f, u32 max_bytes)
 {
   u32 total_drop_bytes, first_drop_bytes, second_drop_bytes;
   u32 cursize, nitems;
diff --git a/src/svm/svm_fifo.h b/src/svm/svm_fifo.h
index 0fff2577..d67237c6 100644
--- a/src/svm/svm_fifo.h
+++ b/src/svm/svm_fifo.h
@@ -23,13 +23,6 @@
 #include <vppinfra/format.h>
 #include <pthread.h>
 
-typedef enum
-{
-  SVM_FIFO_TAG_NOT_HELD = 0,
-  SVM_FIFO_TAG_DEQUEUE,
-  SVM_FIFO_TAG_ENQUEUE,
-} svm_lock_tag_t;
-
 /** Out-of-order segment */
 typedef struct
 {
@@ -37,7 +30,7 @@ typedef struct
   u32 prev;	/**< Previous linked-list element pool index */
 
   u32 start;	/**< Start of segment, normalized*/
-  u32 length;		/**< Length of segment */
+  u32 length;	/**< Length of segment */
 } ooo_segment_t;
 
 format_function_t format_ooo_segment;
@@ -52,12 +45,11 @@ typedef struct
     CLIB_CACHE_LINE_ALIGN_MARK (end_cursize);
 
   volatile u8 has_event;	/**< non-zero if deq event exists */
-  u32 owner_pid;
 
   /* Backpointers */
-  u32 server_session_index;
+  u32 master_session_index;
   u32 client_session_index;
-  u8 server_thread_index;
+  u8 master_thread_index;
   u8 client_thread_index;
   u32 segment_manager;
     CLIB_CACHE_LINE_ALIGN_MARK (end_shared);
@@ -117,19 +109,14 @@ svm_fifo_unset_event (svm_fifo_t * f)
 svm_fifo_t *svm_fifo_create (u32 data_size_in_bytes);
 void svm_fifo_free (svm_fifo_t * f);
 
-int svm_fifo_enqueue_nowait (svm_fifo_t * f, int pid, u32 max_bytes,
+int svm_fifo_enqueue_nowait (svm_fifo_t * f, u32 max_bytes,
 			     u8 * copy_from_here);
+int svm_fifo_enqueue_with_offset (svm_fifo_t * f, u32 offset,
+				  u32 required_bytes, u8 * copy_from_here);
+int svm_fifo_dequeue_nowait (svm_fifo_t * f, u32 max_bytes, u8 * copy_here);
 
-int svm_fifo_enqueue_with_offset (svm_fifo_t * f, int pid,
-				  u32 offset, u32 required_bytes,
-				  u8 * copy_from_here);
-
-int svm_fifo_dequeue_nowait (svm_fifo_t * f, int pid, u32 max_bytes,
-			     u8 * copy_here);
-
-int svm_fifo_peek (svm_fifo_t * f, int pid, u32 offset, u32 max_bytes,
-		   u8 * copy_here);
-int svm_fifo_dequeue_drop (svm_fifo_t * f, int pid, u32 max_bytes);
+int svm_fifo_peek (svm_fifo_t * f, u32 offset, u32 max_bytes, u8 * copy_here);
+int svm_fifo_dequeue_drop (svm_fifo_t * f, u32 max_bytes);
 u32 svm_fifo_number_ooo_segments (svm_fifo_t * f);
 ooo_segment_t *svm_fifo_first_ooo_segment (svm_fifo_t * f);
 
diff --git a/src/svm/svm_fifo_segment.c b/src/svm/svm_fifo_segment.c
index acabb3bd..281fae27 100644
--- a/src/svm/svm_fifo_segment.c
+++ b/src/svm/svm_fifo_segment.c
@@ -70,6 +70,44 @@ svm_fifo_segment_create (svm_fifo_segment_create_args_t * a)
   return (0);
 }
 
+/** Create an svm fifo segment in process-private memory */
+int
+svm_fifo_segment_create_process_private (svm_fifo_segment_create_args_t * a)
+{
+  svm_fifo_segment_private_t *s;
+  svm_fifo_segment_main_t *sm = &svm_fifo_segment_main;
+  ssvm_shared_header_t *sh;
+  svm_fifo_segment_header_t *fsh;
+
+  /* Allocate a fresh segment */
+  pool_get (sm->segments, s);
+  memset (s, 0, sizeof (*s));
+
+  s->ssvm.ssvm_size = ~0;
+  s->ssvm.i_am_master = 1;
+  s->ssvm.my_pid = getpid ();
+  s->ssvm.name = (u8 *) a->segment_name;
+  s->ssvm.requested_va = ~0;
+
+  /* Allocate a [sic] shared memory header, in process memory... */
+  sh = clib_mem_alloc_aligned (sizeof (*sh), CLIB_CACHE_LINE_BYTES);
+  s->ssvm.sh = sh;
+
+  memset (sh, 0, sizeof (*sh));
+  sh->heap = clib_mem_get_heap ();
+
+  /* Set up svm_fifo_segment shared header */
+  fsh = clib_mem_alloc (sizeof (*fsh));
+  memset (fsh, 0, sizeof (*fsh));
+  sh->opaque[0] = fsh;
+  s->h = fsh;
+  fsh->segment_name = format (0, "%s%c", a->segment_name, 0);
+
+  sh->ready = 1;
+  a->new_segment_index = s - sm->segments;
+  return (0);
+}
+
 /** (slave) attach to an svm fifo segment */
 int
 svm_fifo_segment_attach (svm_fifo_segment_create_args_t * a)
@@ -82,7 +120,6 @@ svm_fifo_segment_attach (svm_fifo_segment_create_args_t * a)
 
   /* Allocate a fresh segment */
   pool_get (sm->segments, s);
-
   memset (s, 0, sizeof (*s));
 
   s->ssvm.ssvm_size = a->segment_size;
@@ -126,19 +163,22 @@ svm_fifo_segment_alloc_fifo (svm_fifo_segment_private_t * s,
 
   sh = s->ssvm.sh;
   fsh = (svm_fifo_segment_header_t *) sh->opaque[0];
+
+  ssvm_lock (sh, 1, 0);
   oldheap = ssvm_push_heap (sh);
 
   /* Note: this can fail, in which case: create another segment */
   f = svm_fifo_create (data_size_in_bytes);
-  if (f == 0)
+  if (PREDICT_FALSE (f == 0))
     {
       ssvm_pop_heap (oldheap);
+      ssvm_unlock (sh);
       return (0);
     }
 
   vec_add1 (fsh->fifos, f);
-
   ssvm_pop_heap (oldheap);
+  ssvm_unlock (sh);
   return (f);
 }
 
@@ -152,8 +192,9 @@ svm_fifo_segment_free_fifo (svm_fifo_segment_private_t * s, svm_fifo_t * f)
 
   sh = s->ssvm.sh;
   fsh = (svm_fifo_segment_header_t *) sh->opaque[0];
-  oldheap = ssvm_push_heap (sh);
 
+  ssvm_lock (sh, 1, 0);
+  oldheap = ssvm_push_heap (sh);
   for (i = 0; i < vec_len (fsh->fifos); i++)
     {
       if (fsh->fifos[i] == f)
@@ -167,6 +208,7 @@ svm_fifo_segment_free_fifo (svm_fifo_segment_private_t * s, svm_fifo_t * f)
 found:
   clib_mem_free (f);
   ssvm_pop_heap (oldheap);
+  ssvm_unlock (sh);
 }
 
 void
diff --git a/src/svm/svm_fifo_segment.h b/src/svm/svm_fifo_segment.h
index 9ab47a4c..4218013a 100644
--- a/src/svm/svm_fifo_segment.h
+++ b/src/svm/svm_fifo_segment.h
@@ -17,6 +17,7 @@
 
 #include <svm/svm_fifo.h>
 #include <svm/ssvm.h>
+#include <vppinfra/lock.h>
 
 typedef struct
 {
@@ -32,6 +33,8 @@ typedef struct
 
 typedef struct
 {
+  volatile u32 lock;
+
   /** pool of segments */
   svm_fifo_segment_private_t *segments;
   /* Where to put the next one */
@@ -78,6 +81,8 @@ typedef enum
 } ssvm_fifo_segment_api_error_enum_t;
 
 int svm_fifo_segment_create (svm_fifo_segment_create_args_t * a);
+int svm_fifo_segment_create_process_private (svm_fifo_segment_create_args_t
+					     * a);
 int svm_fifo_segment_attach (svm_fifo_segment_create_args_t * a);
 void svm_fifo_segment_delete (svm_fifo_segment_private_t * s);
 
diff --git a/src/svm/test_svm_fifo1.c b/src/svm/test_svm_fifo1.c
index 355653df..398dd6d7 100644
--- a/src/svm/test_svm_fifo1.c
+++ b/src/svm/test_svm_fifo1.c
@@ -25,7 +25,6 @@ hello_world (int verbose)
   u8 *test_data;
   u8 *retrieved_data = 0;
   clib_error_t *error = 0;
-  int pid = getpid ();
 
   memset (a, 0, sizeof (*a));
 
@@ -48,18 +47,16 @@ hello_world (int verbose)
   vec_validate (retrieved_data, vec_len (test_data) - 1);
 
   while (svm_fifo_max_enqueue (f) >= vec_len (test_data))
-    svm_fifo_enqueue_nowait (f, pid, vec_len (test_data), test_data);
+    svm_fifo_enqueue_nowait (f, vec_len (test_data), test_data);
 
   while (svm_fifo_max_dequeue (f) >= vec_len (test_data))
-    svm_fifo_dequeue_nowait (f, pid, vec_len (retrieved_data),
-			     retrieved_data);
+    svm_fifo_dequeue_nowait (f, vec_len (retrieved_data), retrieved_data);
 
   while (svm_fifo_max_enqueue (f) >= vec_len (test_data))
-    svm_fifo_enqueue_nowait (f, pid, vec_len (test_data), test_data);
+    svm_fifo_enqueue_nowait (f, vec_len (test_data), test_data);
 
   while (svm_fifo_max_dequeue (f) >= vec_len (test_data))
-    svm_fifo_dequeue_nowait (f, pid, vec_len (retrieved_data),
-			     retrieved_data);
+    svm_fifo_dequeue_nowait (f, vec_len (retrieved_data), retrieved_data);
 
   if (!memcmp (retrieved_data, test_data, vec_len (test_data)))
     error = clib_error_return (0, "data test OK, got '%s'", retrieved_data);
@@ -81,7 +78,6 @@ master (int verbose)
   u8 *test_data;
   u8 *retrieved_data = 0;
   int i;
-  int pid = getpid ();
 
   memset (a, 0, sizeof (*a));
 
@@ -104,7 +100,7 @@ master (int verbose)
   vec_validate (retrieved_data, vec_len (test_data) - 1);
 
   for (i = 0; i < 1000; i++)
-    svm_fifo_enqueue_nowait (f, pid, vec_len (test_data), test_data);
+    svm_fifo_enqueue_nowait (f, vec_len (test_data), test_data);
 
   return clib_error_return (0, "master (enqueue) done");
 }
@@ -176,7 +172,6 @@ offset (int verbose)
   u32 *test_data = 0;
   u32 *recovered_data = 0;
   int i;
-  int pid = getpid ();
 
   memset (a, 0, sizeof (*a));
 
@@ -199,19 +194,19 @@ offset (int verbose)
     vec_add1 (test_data, i);
 
   /* Enqueue the first 1024 u32's */
-  svm_fifo_enqueue_nowait (f, pid, 4096 /* bytes to enqueue */ ,
+  svm_fifo_enqueue_nowait (f, 4096 /* bytes to enqueue */ ,
 			   (u8 *) test_data);
 
   /* Enqueue the third 1024 u32's 2048 ahead of the current tail */
-  svm_fifo_enqueue_with_offset (f, pid, 4096, 4096, (u8 *) & test_data[2048]);
+  svm_fifo_enqueue_with_offset (f, 4096, 4096, (u8 *) & test_data[2048]);
 
   /* Enqueue the second 1024 u32's at the current tail */
-  svm_fifo_enqueue_nowait (f, pid, 4096 /* bytes to enqueue */ ,
+  svm_fifo_enqueue_nowait (f, 4096 /* bytes to enqueue */ ,
 			   (u8 *) & test_data[1024]);
 
   vec_validate (recovered_data, (3 * 1024) - 1);
 
-  svm_fifo_dequeue_nowait (f, pid, 3 * 4096, (u8 *) recovered_data);
+  svm_fifo_dequeue_nowait (f, 3 * 4096, (u8 *) recovered_data);
 
   for (i = 0; i < (3 * 1024); i++)
     {
@@ -237,7 +232,6 @@ slave (int verbose)
   int rv;
   u8 *test_data;
   u8 *retrieved_data = 0;
-  int pid = getpid ();
   int i;
 
   memset (a, 0, sizeof (*a));
@@ -262,8 +256,7 @@ slave (int verbose)
 
   for (i = 0; i < 1000; i++)
     {
-      svm_fifo_dequeue_nowait (f, pid, vec_len (retrieved_data),
-			       retrieved_data);
+      svm_fifo_dequeue_nowait (f, vec_len (retrieved_data), retrieved_data);
       if (memcmp (retrieved_data, test_data, vec_len (retrieved_data)))
 	return clib_error_return (0, "retrieved data incorrect, '%s'",
 				  retrieved_data);
diff --git a/src/uri/uri_tcp_test.c b/src/uri/uri_tcp_test.c
index 2e15d36c..686c93f9 100755
--- a/src/uri/uri_tcp_test.c
+++ b/src/uri/uri_tcp_test.c
@@ -45,12 +45,13 @@ typedef struct
   svm_fifo_t *server_rx_fifo;
   svm_fifo_t *server_tx_fifo;
 
-  u32 vpp_session_handle;
+  u64 vpp_session_handle;
 } session_t;
 
 typedef enum
 {
   STATE_START,
+  STATE_ATTACHED,
   STATE_READY,
   STATE_DISCONNECTING,
   STATE_FAILED
@@ -127,6 +128,34 @@ uri_tcp_test_main_t uri_tcp_test_main;
 #define NITER 4000000
 #endif
 
+static u8 *
+format_api_error (u8 * s, va_list * args)
+{
+  uri_tcp_test_main_t *utm = &uri_tcp_test_main;
+  i32 error = va_arg (*args, u32);
+  uword *p;
+
+  p = hash_get (utm->error_string_by_error_number, -error);
+
+  if (p)
+    s = format (s, "%s", p[0]);
+  else
+    s = format (s, "%d", error);
+  return s;
+}
+
+static void
+init_error_string_table (uri_tcp_test_main_t * utm)
+{
+  utm->error_string_by_error_number = hash_create (0, sizeof (uword));
+
+#define _(n,v,s) hash_set (utm->error_string_by_error_number, -v, s);
+  foreach_vnet_api_error;
+#undef _
+
+  hash_set (utm->error_string_by_error_number, 99, "Misc");
+}
+
 int
 wait_for_state_change (uri_tcp_test_main_t * utm, connection_state_t state)
 {
@@ -150,7 +179,7 @@ wait_for_state_change (uri_tcp_test_main_t * utm, connection_state_t state)
 }
 
 void
-application_attach (uri_tcp_test_main_t * utm)
+application_send_attach (uri_tcp_test_main_t * utm)
 {
   vl_api_application_attach_t *bmp;
   u32 fifo_size = 3 << 20;
@@ -160,8 +189,8 @@ application_attach (uri_tcp_test_main_t * utm)
   bmp->_vl_msg_id = ntohs (VL_API_APPLICATION_ATTACH);
   bmp->client_index = utm->my_client_index;
   bmp->context = ntohl (0xfeedface);
-  bmp->options[SESSION_OPTIONS_FLAGS] =
-    SESSION_OPTIONS_FLAGS_USE_FIFO | SESSION_OPTIONS_FLAGS_ADD_SEGMENT;
+  bmp->options[APP_OPTIONS_FLAGS] =
+    APP_OPTIONS_FLAGS_USE_FIFO | APP_OPTIONS_FLAGS_ADD_SEGMENT;
   bmp->options[SESSION_OPTIONS_RX_FIFO_SIZE] = fifo_size;
   bmp->options[SESSION_OPTIONS_TX_FIFO_SIZE] = fifo_size;
   bmp->options[SESSION_OPTIONS_ADD_SEGMENT_SIZE] = 128 << 20;
@@ -169,6 +198,18 @@ application_attach (uri_tcp_test_main_t * utm)
   vl_msg_api_send_shmem (utm->vl_input_queue, (u8 *) & bmp);
 }
 
+int
+application_attach (uri_tcp_test_main_t * utm)
+{
+  application_send_attach (utm);
+  if (wait_for_state_change (utm, STATE_ATTACHED))
+    {
+      clib_warning ("timeout waiting for STATE_ATTACHED");
+      return -1;
+    }
+  return 0;
+}
+
 void
 application_detach (uri_tcp_test_main_t * utm)
 {
@@ -192,8 +233,8 @@ vl_api_application_attach_reply_t_handler (vl_api_application_attach_reply_t *
 
   if (mp->retval)
     {
-      uword *errp = hash_get (utm->error_string_by_error_number, -mp->retval);
-      clib_warning ("attach failed: %s", *errp);
+      clib_warning ("attach failed: %U", format_api_error,
+		    clib_net_to_host_u32 (mp->retval));
       utm->state = STATE_FAILED;
       return;
     }
@@ -220,7 +261,7 @@ vl_api_application_attach_reply_t_handler (vl_api_application_attach_reply_t *
 
   utm->our_event_queue =
     (unix_shared_memory_queue_t *) mp->app_event_queue_address;
-
+  utm->state = STATE_ATTACHED;
 }
 
 static void
@@ -231,18 +272,6 @@ vl_api_application_detach_reply_t_handler (vl_api_application_detach_reply_t *
     clib_warning ("detach returned with err: %d", mp->retval);
 }
 
-static void
-init_error_string_table (uri_tcp_test_main_t * utm)
-{
-  utm->error_string_by_error_number = hash_create (0, sizeof (uword));
-
-#define _(n,v,s) hash_set (utm->error_string_by_error_number, -v, s);
-  foreach_vnet_api_error;
-#undef _
-
-  hash_set (utm->error_string_by_error_number, 99, "Misc");
-}
-
 static void
 stop_signal (int signum)
 {
@@ -392,7 +421,7 @@ client_handle_fifo_event_rx (uri_tcp_test_main_t * utm,
   /* Read the bytes */
   do
     {
-      n_read = svm_fifo_dequeue_nowait (rx_fifo, 0,
+      n_read = svm_fifo_dequeue_nowait (rx_fifo,
 					clib_min (vec_len (utm->rx_buf),
 						  bytes), utm->rx_buf);
       if (n_read > 0)
@@ -432,11 +461,11 @@ client_handle_event_queue (uri_tcp_test_main_t * utm)
 				0 /* nowait */ );
   switch (e->event_type)
     {
-    case FIFO_EVENT_SERVER_RX:
+    case FIFO_EVENT_APP_RX:
       client_handle_fifo_event_rx (utm, e);
       break;
 
-    case FIFO_EVENT_SERVER_EXIT:
+    case FIFO_EVENT_DISCONNECT:
       return;
 
     default:
@@ -458,11 +487,11 @@ client_rx_thread_fn (void *arg)
 				    0 /* nowait */ );
       switch (e->event_type)
 	{
-	case FIFO_EVENT_SERVER_RX:
+	case FIFO_EVENT_APP_RX:
 	  client_handle_fifo_event_rx (utm, e);
 	  break;
 
-	case FIFO_EVENT_SERVER_EXIT:
+	case FIFO_EVENT_DISCONNECT:
 	  return 0;
 	default:
 	  clib_warning ("unknown event type %d", e->event_type);
@@ -487,9 +516,8 @@ vl_api_connect_uri_reply_t_handler (vl_api_connect_uri_reply_t * mp)
 
   if (mp->retval)
     {
-      uword *errp = hash_get (utm->error_string_by_error_number,
-			      -clib_net_to_host_u32 (mp->retval));
-      clib_warning ("connection failed with code: %s", *errp);
+      clib_warning ("connection failed with code: %U", format_api_error,
+		    clib_net_to_host_u32 (mp->retval));
       utm->state = STATE_FAILED;
       return;
     }
@@ -551,7 +579,7 @@ send_test_chunk (uri_tcp_test_main_t * utm, svm_fifo_t * tx_fifo, int mypid,
     {
       actual_write =
 	bytes_to_snd > queue_max_chunk ? queue_max_chunk : bytes_to_snd;
-      rv = svm_fifo_enqueue_nowait (tx_fifo, mypid, actual_write,
+      rv = svm_fifo_enqueue_nowait (tx_fifo, actual_write,
 				    test_data + test_buf_offset);
 
       if (rv > 0)
@@ -564,7 +592,7 @@ send_test_chunk (uri_tcp_test_main_t * utm, svm_fifo_t * tx_fifo, int mypid,
 	    {
 	      /* Fabricate TX event, send to vpp */
 	      evt.fifo = tx_fifo;
-	      evt.event_type = FIFO_EVENT_SERVER_TX;
+	      evt.event_type = FIFO_EVENT_APP_TX;
 	      evt.event_id = serial_number++;
 
 	      unix_shared_memory_queue_add (utm->vpp_event_queue,
@@ -619,7 +647,7 @@ client_send_data (uri_tcp_test_main_t * utm)
 }
 
 void
-client_connect (uri_tcp_test_main_t * utm)
+client_send_connect (uri_tcp_test_main_t * utm)
 {
   vl_api_connect_uri_t *cmp;
   cmp = vl_msg_api_alloc (sizeof (*cmp));
@@ -632,8 +660,20 @@ client_connect (uri_tcp_test_main_t * utm)
   vl_msg_api_send_shmem (utm->vl_input_queue, (u8 *) & cmp);
 }
 
+int
+client_connect (uri_tcp_test_main_t * utm)
+{
+  client_send_connect (utm);
+  if (wait_for_state_change (utm, STATE_READY))
+    {
+      clib_warning ("Connect failed");
+      return -1;
+    }
+  return 0;
+}
+
 void
-client_disconnect (uri_tcp_test_main_t * utm)
+client_send_disconnect (uri_tcp_test_main_t * utm)
 {
   session_t *connected_session;
   vl_api_disconnect_session_t *dmp;
@@ -647,16 +687,29 @@ client_disconnect (uri_tcp_test_main_t * utm)
   vl_msg_api_send_shmem (utm->vl_input_queue, (u8 *) & dmp);
 }
 
+int
+client_disconnect (uri_tcp_test_main_t * utm)
+{
+  client_send_disconnect (utm);
+  if (wait_for_state_change (utm, STATE_START))
+    {
+      clib_warning ("Disconnect failed");
+      return -1;
+    }
+  return 0;
+}
+
 static void
 client_test (uri_tcp_test_main_t * utm)
 {
   int i;
 
-  application_attach (utm);
-  client_connect (utm);
+  if (application_attach (utm))
+    return;
 
-  if (wait_for_state_change (utm, STATE_READY))
+  if (client_connect (utm))
     {
+      application_detach (utm);
       return;
     }
 
@@ -671,11 +724,6 @@ client_test (uri_tcp_test_main_t * utm)
   /* Disconnect */
   client_disconnect (utm);
 
-  if (wait_for_state_change (utm, STATE_START))
-    {
-      clib_warning ("Disconnect failed");
-      return;
-    }
   application_detach (utm);
 }
 
@@ -686,9 +734,8 @@ vl_api_bind_uri_reply_t_handler (vl_api_bind_uri_reply_t * mp)
 
   if (mp->retval)
     {
-      uword *errp = hash_get (utm->error_string_by_error_number,
-			      -clib_net_to_host_u32 (mp->retval));
-      clib_warning ("bind failed: %s", (char *) *errp);
+      clib_warning ("bind failed: %s", format_api_error,
+		    clib_net_to_host_u32 (mp->retval));
       utm->state = STATE_FAILED;
       return;
     }
@@ -869,7 +916,7 @@ server_handle_fifo_event_rx (uri_tcp_test_main_t * utm,
   /* Read the bytes */
   do
     {
-      n_read = svm_fifo_dequeue_nowait (rx_fifo, 0, vec_len (utm->rx_buf),
+      n_read = svm_fifo_dequeue_nowait (rx_fifo, vec_len (utm->rx_buf),
 					utm->rx_buf);
       if (n_read > 0)
 	bytes -= n_read;
@@ -882,7 +929,7 @@ server_handle_fifo_event_rx (uri_tcp_test_main_t * utm,
 	{
 	  do
 	    {
-	      rv = svm_fifo_enqueue_nowait (tx_fifo, 0, n_read, utm->rx_buf);
+	      rv = svm_fifo_enqueue_nowait (tx_fifo, n_read, utm->rx_buf);
 	    }
 	  while (rv <= 0 && !utm->time_to_stop);
 
@@ -891,7 +938,7 @@ server_handle_fifo_event_rx (uri_tcp_test_main_t * utm,
 	    {
 	      /* Fabricate TX event, send to vpp */
 	      evt.fifo = tx_fifo;
-	      evt.event_type = FIFO_EVENT_SERVER_TX;
+	      evt.event_type = FIFO_EVENT_APP_TX;
 	      evt.event_id = e->event_id;
 
 	      q = utm->vpp_event_queue;
@@ -914,11 +961,11 @@ server_handle_event_queue (uri_tcp_test_main_t * utm)
 				    0 /* nowait */ );
       switch (e->event_type)
 	{
-	case FIFO_EVENT_SERVER_RX:
+	case FIFO_EVENT_APP_RX:
 	  server_handle_fifo_event_rx (utm, e);
 	  break;
 
-	case FIFO_EVENT_SERVER_EXIT:
+	case FIFO_EVENT_DISCONNECT:
 	  return;
 
 	default:
@@ -936,7 +983,7 @@ server_handle_event_queue (uri_tcp_test_main_t * utm)
 }
 
 void
-server_listen (uri_tcp_test_main_t * utm)
+server_send_listen (uri_tcp_test_main_t * utm)
 {
   vl_api_bind_uri_t *bmp;
   bmp = vl_msg_api_alloc (sizeof (*bmp));
@@ -949,8 +996,20 @@ server_listen (uri_tcp_test_main_t * utm)
   vl_msg_api_send_shmem (utm->vl_input_queue, (u8 *) & bmp);
 }
 
+int
+server_listen (uri_tcp_test_main_t * utm)
+{
+  server_send_listen (utm);
+  if (wait_for_state_change (utm, STATE_READY))
+    {
+      clib_warning ("timeout waiting for STATE_READY");
+      return -1;
+    }
+  return 0;
+}
+
 void
-server_unbind (uri_tcp_test_main_t * utm)
+server_send_unbind (uri_tcp_test_main_t * utm)
 {
   vl_api_unbind_uri_t *ump;
 
@@ -963,31 +1022,33 @@ server_unbind (uri_tcp_test_main_t * utm)
   vl_msg_api_send_shmem (utm->vl_input_queue, (u8 *) & ump);
 }
 
+int
+server_unbind (uri_tcp_test_main_t * utm)
+{
+  server_send_unbind (utm);
+  if (wait_for_state_change (utm, STATE_START))
+    {
+      clib_warning ("timeout waiting for STATE_START");
+      return -1;
+    }
+  return 0;
+}
+
 void
 server_test (uri_tcp_test_main_t * utm)
 {
-  application_attach (utm);
+  if (application_attach (utm))
+    return;
 
   /* Bind to uri */
-  server_listen (utm);
-
-  if (wait_for_state_change (utm, STATE_READY))
-    {
-      clib_warning ("timeout waiting for STATE_READY");
-      return;
-    }
+  if (server_listen (utm))
+    return;
 
   /* Enter handle event loop */
   server_handle_event_queue (utm);
 
   /* Cleanup */
-  server_unbind (utm);
-
-  if (wait_for_state_change (utm, STATE_START))
-    {
-      clib_warning ("timeout waiting for STATE_START");
-      return;
-    }
+  server_send_unbind (utm);
 
   application_detach (utm);
 
diff --git a/src/uri/uri_udp_test.c b/src/uri/uri_udp_test.c
index 598052bc..266215c8 100644
--- a/src/uri/uri_udp_test.c
+++ b/src/uri/uri_udp_test.c
@@ -164,7 +164,7 @@ setup_signal_handlers (void)
 }
 
 void
-application_attach (uri_udp_test_main_t * utm)
+application_send_attach (uri_udp_test_main_t * utm)
 {
   vl_api_application_attach_t *bmp;
   u32 fifo_size = 3 << 20;
@@ -174,8 +174,8 @@ application_attach (uri_udp_test_main_t * utm)
   bmp->_vl_msg_id = ntohs (VL_API_APPLICATION_ATTACH);
   bmp->client_index = utm->my_client_index;
   bmp->context = ntohl (0xfeedface);
-  bmp->options[SESSION_OPTIONS_FLAGS] =
-    SESSION_OPTIONS_FLAGS_USE_FIFO | SESSION_OPTIONS_FLAGS_ADD_SEGMENT;
+  bmp->options[APP_OPTIONS_FLAGS] =
+    APP_OPTIONS_FLAGS_USE_FIFO | APP_OPTIONS_FLAGS_ADD_SEGMENT;
   bmp->options[SESSION_OPTIONS_RX_FIFO_SIZE] = fifo_size;
   bmp->options[SESSION_OPTIONS_TX_FIFO_SIZE] = fifo_size;
   bmp->options[SESSION_OPTIONS_ADD_SEGMENT_SIZE] = 128 << 20;
@@ -307,7 +307,7 @@ cut_through_thread_fn (void *arg)
       /* We read from the tx fifo and write to the rx fifo */
       do
 	{
-	  actual_transfer = svm_fifo_dequeue_nowait (tx_fifo, 0,
+	  actual_transfer = svm_fifo_dequeue_nowait (tx_fifo,
 						     vec_len (my_copy_buffer),
 						     my_copy_buffer);
 	}
@@ -318,7 +318,7 @@ cut_through_thread_fn (void *arg)
       buffer_offset = 0;
       while (actual_transfer > 0)
 	{
-	  rv = svm_fifo_enqueue_nowait (rx_fifo, 0, actual_transfer,
+	  rv = svm_fifo_enqueue_nowait (rx_fifo, actual_transfer,
 					my_copy_buffer + buffer_offset);
 	  if (rv > 0)
 	    {
@@ -357,7 +357,6 @@ client_send (uri_udp_test_main_t * utm, session_t * session)
   u64 bytes_received = 0, bytes_sent = 0;
   i32 bytes_to_read;
   int rv;
-  int mypid = getpid ();
   f64 before, after, delta, bytes_per_second;
   svm_fifo_t *rx_fifo, *tx_fifo;
   int buffer_offset, bytes_to_send = 0;
@@ -382,8 +381,7 @@ client_send (uri_udp_test_main_t * utm, session_t * session)
       buffer_offset = 0;
       while (bytes_to_send > 0)
 	{
-	  rv = svm_fifo_enqueue_nowait (tx_fifo, mypid,
-					bytes_to_send,
+	  rv = svm_fifo_enqueue_nowait (tx_fifo, bytes_to_send,
 					test_data + buffer_offset);
 
 	  if (rv > 0)
@@ -402,7 +400,7 @@ client_send (uri_udp_test_main_t * utm, session_t * session)
       buffer_offset = 0;
       while (bytes_to_read > 0)
 	{
-	  rv = svm_fifo_dequeue_nowait (rx_fifo, mypid,
+	  rv = svm_fifo_dequeue_nowait (rx_fifo,
 					bytes_to_read,
 					utm->rx_buf + buffer_offset);
 	  if (rv > 0)
@@ -415,8 +413,8 @@ client_send (uri_udp_test_main_t * utm, session_t * session)
     }
   while (bytes_received < bytes_sent)
     {
-      rv = svm_fifo_dequeue_nowait (rx_fifo, mypid,
-				    vec_len (utm->rx_buf), utm->rx_buf);
+      rv =
+	svm_fifo_dequeue_nowait (rx_fifo, vec_len (utm->rx_buf), utm->rx_buf);
       if (rv > 0)
 	{
 #if CLIB_DEBUG > 0
@@ -459,7 +457,7 @@ uri_udp_client_test (uri_udp_test_main_t * utm)
 {
   session_t *session;
 
-  application_attach (utm);
+  application_send_attach (utm);
   udp_client_connect (utm);
 
   if (wait_for_state_change (utm, STATE_READY))
@@ -559,8 +557,8 @@ vl_api_connect_uri_t_handler (vl_api_connect_uri_t * mp)
 							 128 * 1024);
   ASSERT (session->server_tx_fifo);
 
-  session->server_rx_fifo->server_session_index = session - utm->sessions;
-  session->server_tx_fifo->server_session_index = session - utm->sessions;
+  session->server_rx_fifo->master_session_index = session - utm->sessions;
+  session->server_tx_fifo->master_session_index = session - utm->sessions;
   utm->cut_through_session_index = session - utm->sessions;
 
   rv = pthread_create (&utm->cut_through_thread_handle,
@@ -805,19 +803,19 @@ server_handle_fifo_event_rx (uri_udp_test_main_t * utm,
 
   do
     {
-      nbytes = svm_fifo_dequeue_nowait (rx_fifo, 0,
-					vec_len (utm->rx_buf), utm->rx_buf);
+      nbytes = svm_fifo_dequeue_nowait (rx_fifo, vec_len (utm->rx_buf),
+					utm->rx_buf);
     }
   while (nbytes <= 0);
   do
     {
-      rv = svm_fifo_enqueue_nowait (tx_fifo, 0, nbytes, utm->rx_buf);
+      rv = svm_fifo_enqueue_nowait (tx_fifo, nbytes, utm->rx_buf);
     }
   while (rv == -2);
 
   /* Fabricate TX event, send to vpp */
   evt.fifo = tx_fifo;
-  evt.event_type = FIFO_EVENT_SERVER_TX;
+  evt.event_type = FIFO_EVENT_APP_TX;
   evt.event_id = e->event_id;
 
   if (svm_fifo_set_event (tx_fifo))
@@ -839,11 +837,11 @@ server_handle_event_queue (uri_udp_test_main_t * utm)
 				    0 /* nowait */ );
       switch (e->event_type)
 	{
-	case FIFO_EVENT_SERVER_RX:
+	case FIFO_EVENT_APP_RX:
 	  server_handle_fifo_event_rx (utm, e);
 	  break;
 
-	case FIFO_EVENT_SERVER_EXIT:
+	case FIFO_EVENT_DISCONNECT:
 	  return;
 
 	default:
@@ -893,7 +891,7 @@ void
 udp_server_test (uri_udp_test_main_t * utm)
 {
 
-  application_attach (utm);
+  application_send_attach (utm);
 
   /* Bind to uri */
   server_listen (utm);
diff --git a/src/vnet/session/application.c b/src/vnet/session/application.c
index 5a45537b..ccf9837f 100644
--- a/src/vnet/session/application.c
+++ b/src/vnet/session/application.c
@@ -87,14 +87,17 @@ application_new ()
 void
 application_del (application_t * app)
 {
-  api_main_t *am = &api_main;
-  void *oldheap;
   segment_manager_t *sm;
   u64 handle;
   u32 index, *handles = 0;
   int i;
   vnet_unbind_args_t _a, *a = &_a;
 
+  /*
+   * The app event queue allocated in first segment is cleared with
+   * the segment manager. No need to explicitly free it.
+   */
+
   /*
    * Cleanup segment managers
    */
@@ -120,14 +123,6 @@ application_del (application_t * app)
       vnet_unbind (a);
     }
 
-  /*
-   * Free the event fifo in the /vpe-api shared-memory segment
-   */
-  oldheap = svm_push_data_heap (am->vlib_rp);
-  if (app->event_queue)
-    unix_shared_memory_queue_free (app->event_queue);
-  svm_pop_heap (oldheap);
-
   application_table_del (app);
   pool_put (app_pool, app);
 }
@@ -149,30 +144,14 @@ int
 application_init (application_t * app, u32 api_client_index, u64 * options,
 		  session_cb_vft_t * cb_fns)
 {
-  api_main_t *am = &api_main;
   segment_manager_t *sm;
   segment_manager_properties_t *props;
-  void *oldheap;
-  u32 app_evt_queue_size;
+  u32 app_evt_queue_size, first_seg_size;
   int rv;
 
   app_evt_queue_size = options[APP_EVT_QUEUE_SIZE] > 0 ?
     options[APP_EVT_QUEUE_SIZE] : default_app_evt_queue_size;
 
-  /* Allocate event fifo in the /vpe-api shared-memory segment */
-  oldheap = svm_push_data_heap (am->vlib_rp);
-
-  /* Allocate server event queue */
-  app->event_queue =
-    unix_shared_memory_queue_init (app_evt_queue_size,
-				   sizeof (session_fifo_event_t),
-				   0 /* consumer pid */ ,
-				   0
-				   /* (do not) signal when queue non-empty */
-    );
-
-  svm_pop_heap (oldheap);
-
   /* Setup segment manager */
   sm = segment_manager_new ();
   sm->app_index = app->index;
@@ -181,16 +160,21 @@ application_init (application_t * app, u32 api_client_index, u64 * options,
   props->rx_fifo_size = options[SESSION_OPTIONS_RX_FIFO_SIZE];
   props->tx_fifo_size = options[SESSION_OPTIONS_TX_FIFO_SIZE];
   props->add_segment = props->add_segment_size != 0;
+  props->use_private_segment = options[APP_OPTIONS_FLAGS]
+    & APP_OPTIONS_FLAGS_BUILTIN_APP;
 
-  if ((rv = segment_manager_init (sm, props,
-				  options[SESSION_OPTIONS_SEGMENT_SIZE])))
+  first_seg_size = options[SESSION_OPTIONS_SEGMENT_SIZE];
+  if ((rv = segment_manager_init (sm, props, first_seg_size)))
     return rv;
 
   app->first_segment_manager = segment_manager_index (sm);
   app->api_client_index = api_client_index;
-  app->flags = options[SESSION_OPTIONS_FLAGS];
+  app->flags = options[APP_OPTIONS_FLAGS];
   app->cb_fns = *cb_fns;
 
+  /* Allocate app event queue in the first shared-memory segment */
+  app->event_queue = segment_manager_alloc_queue (sm, app_evt_queue_size);
+
   /* Check that the obvious things are properly set up */
   application_verify_cb_fns (cb_fns);
 
@@ -451,8 +435,8 @@ application_format_connects (application_t * app, int verbose)
 	    continue;
 
 	  fifo = fifos[i];
-	  session_index = fifo->server_session_index;
-	  thread_index = fifo->server_thread_index;
+	  session_index = fifo->master_session_index;
+	  thread_index = fifo->master_thread_index;
 
 	  session = stream_session_get (session_index, thread_index);
 	  str = format (0, "%U", format_stream_session, session, verbose);
diff --git a/src/vnet/session/application.h b/src/vnet/session/application.h
index 6bcee9d3..35caae85 100644
--- a/src/vnet/session/application.h
+++ b/src/vnet/session/application.h
@@ -61,18 +61,6 @@ typedef struct _application
   /** Flags */
   u32 flags;
 
-  /* Stream server mode: accept or connect
-   * TODO REMOVE*/
-  u8 mode;
-
-  /** Index of the listen session or connect session
-   * TODO REMOVE*/
-  u32 session_index;
-
-  /** Session thread index for client connect sessions
-   * TODO REMOVE */
-  u32 thread_index;
-
   /*
    * Binary API interface to external app
    */
diff --git a/src/vnet/session/application_interface.c b/src/vnet/session/application_interface.c
index 96d2c621..ad44baa1 100644
--- a/src/vnet/session/application_interface.c
+++ b/src/vnet/session/application_interface.c
@@ -142,7 +142,7 @@ vnet_connect_i (u32 app_index, u32 api_context, session_type_t sst,
        * Server is willing to have a direct fifo connection created
        * instead of going through the state machine, etc.
        */
-      if (server->flags & SESSION_OPTIONS_FLAGS_USE_FIFO)
+      if (server->flags & APP_OPTIONS_FLAGS_USE_FIFO)
 	return server->cb_fns.
 	  redirect_connect_callback (server->api_client_index, mp);
     }
@@ -363,7 +363,11 @@ vnet_disconnect_session (vnet_disconnect_args_t * a)
   if (!s || s->app_index != a->app_index)
     return VNET_API_ERROR_INVALID_VALUE;
 
-  stream_session_disconnect (s);
+  /* We're peeking into another's thread pool. Make sure */
+  ASSERT (s->session_index == index);
+
+  session_send_session_evt_to_thread (a->handle, FIFO_EVENT_DISCONNECT,
+				      thread_index);
   return 0;
 }
 
@@ -395,24 +399,6 @@ vnet_connect (vnet_connect_args_t * a)
   return vnet_connect_i (a->app_index, a->api_context, sst, &a->tep, a->mp);
 }
 
-int
-vnet_disconnect (vnet_disconnect_args_t * a)
-{
-  stream_session_t *session;
-  u32 session_index, thread_index;
-
-  if (api_parse_session_handle (a->handle, &session_index, &thread_index))
-    {
-      clib_warning ("Invalid handle");
-      return -1;
-    }
-
-  session = stream_session_get (session_index, thread_index);
-  stream_session_disconnect (session);
-
-  return 0;
-}
-
 /*
  * fd.io coding-style-patch-verification: ON
  *
diff --git a/src/vnet/session/application_interface.h b/src/vnet/session/application_interface.h
index 2c497531..7d924c14 100644
--- a/src/vnet/session/application_interface.h
+++ b/src/vnet/session/application_interface.h
@@ -30,10 +30,18 @@ typedef enum _session_api_proto
 
 typedef struct _vnet_app_attach_args_t
 {
+  /** Binary API client index */
   u32 api_client_index;
+
+  /** Application and segment manager options */
   u64 *options;
+
+  /** Session to application callback functions */
   session_cb_vft_t *session_cb_vft;
 
+  /** Flag that indicates if app is builtin */
+  u8 builtin;
+
   /*
    * Results
    */
@@ -110,7 +118,7 @@ typedef struct _vnet_disconnect_args_t
 typedef enum
 {
   APP_EVT_QUEUE_SIZE,
-  SESSION_OPTIONS_FLAGS,
+  APP_OPTIONS_FLAGS,
   SESSION_OPTIONS_SEGMENT_SIZE,
   SESSION_OPTIONS_ADD_SEGMENT_SIZE,
   SESSION_OPTIONS_RX_FIFO_SIZE,
@@ -119,11 +127,30 @@ typedef enum
   SESSION_OPTIONS_N_OPTIONS
 } app_attach_options_index_t;
 
-/** Server can handle delegated connect requests from local clients */
-#define SESSION_OPTIONS_FLAGS_USE_FIFO	(1<<0)
+#define foreach_app_options_flags				\
+  _(USE_FIFO, "Use FIFO with redirects")			\
+  _(ADD_SEGMENT, "Add segment and signal app if needed")	\
+  _(BUILTIN_APP, "Application is builtin")			\
+
+typedef enum _app_options
+{
+#define _(sym, str) APP_OPTIONS_##sym,
+  foreach_app_options_flags
+#undef _
+} app_options_t;
+
+typedef enum _app_options_flags
+{
+#define _(sym, str) APP_OPTIONS_FLAGS_##sym = 1 << APP_OPTIONS_##sym,
+  foreach_app_options_flags
+#undef _
+} app_options_flags_t;
 
-/** Server wants vpp to add segments when out of memory for fifos */
-#define SESSION_OPTIONS_FLAGS_ADD_SEGMENT   (1<<1)
+///** Server can handle delegated connect requests from local clients */
+//#define APP_OPTIONS_FLAGS_USE_FIFO    (1<<0)
+//
+///** Server wants vpp to add segments when out of memory for fifos */
+//#define APP_OPTIONS_FLAGS_ADD_SEGMENT   (1<<1)
 
 #define VNET_CONNECT_REDIRECTED	123
 
@@ -138,7 +165,6 @@ int vnet_disconnect_session (vnet_disconnect_args_t * a);
 int vnet_bind (vnet_bind_args_t * a);
 int vnet_connect (vnet_connect_args_t * a);
 int vnet_unbind (vnet_unbind_args_t * a);
-int vnet_disconnect (vnet_disconnect_args_t * a);
 
 int
 api_parse_session_handle (u64 handle, u32 * session_index,
diff --git a/src/vnet/session/node.c b/src/vnet/session/node.c
index dd211c51..210754fa 100644
--- a/src/vnet/session/node.c
+++ b/src/vnet/session/node.c
@@ -218,8 +218,8 @@ session_tx_fifo_read_and_snd_i (vlib_main_t * vm, vlib_node_runtime_t * node,
 	   *      2) buffer chains */
 	  if (peek_data)
 	    {
-	      n_bytes_read = svm_fifo_peek (s0->server_tx_fifo, s0->pid,
-					    rx_offset, len_to_deq0, data0);
+	      n_bytes_read = svm_fifo_peek (s0->server_tx_fifo, rx_offset,
+					    len_to_deq0, data0);
 	      if (n_bytes_read <= 0)
 		goto dequeue_fail;
 
@@ -230,8 +230,7 @@ session_tx_fifo_read_and_snd_i (vlib_main_t * vm, vlib_node_runtime_t * node,
 	  else
 	    {
 	      n_bytes_read = svm_fifo_dequeue_nowait (s0->server_tx_fifo,
-						      s0->pid, len_to_deq0,
-						      data0);
+						      len_to_deq0, data0);
 	      if (n_bytes_read <= 0)
 		goto dequeue_fail;
 	    }
@@ -301,6 +300,26 @@ session_tx_fifo_dequeue_and_snd (vlib_main_t * vm, vlib_node_runtime_t * node,
 					 n_tx_pkts, 0);
 }
 
+stream_session_t *
+session_event_get_session (session_fifo_event_t * e0, u8 thread_index)
+{
+  svm_fifo_t *f0;
+  stream_session_t *s0;
+  u32 session_index0;
+
+  f0 = e0->fifo;
+  session_index0 = f0->master_session_index;
+
+  /* $$$ add multiple event queues, per vpp worker thread */
+  ASSERT (f0->master_thread_index == thread_index);
+
+  s0 = stream_session_get_if_valid (session_index0, thread_index);
+
+  ASSERT (s0->thread_index == thread_index);
+
+  return s0;
+}
+
 static uword
 session_queue_node_fn (vlib_main_t * vm, vlib_node_runtime_t * node,
 		       vlib_frame_t * frame)
@@ -370,34 +389,24 @@ skip_dequeue:
   n_events = vec_len (my_fifo_events);
   for (i = 0; i < n_events; i++)
     {
-      svm_fifo_t *f0;		/* $$$ prefetch 1 ahead maybe */
-      stream_session_t *s0;
-      u32 session_index0;
+      stream_session_t *s0;	/* $$$ prefetch 1 ahead maybe */
       session_fifo_event_t *e0;
 
       e0 = &my_fifo_events[i];
-      f0 = e0->fifo;
-      session_index0 = f0->server_session_index;
-
-      /* $$$ add multiple event queues, per vpp worker thread */
-      ASSERT (f0->server_thread_index == my_thread_index);
 
-      s0 = stream_session_get_if_valid (session_index0, my_thread_index);
-
-      if (CLIB_DEBUG && !s0)
+      switch (e0->event_type)
 	{
-	  clib_warning ("It's dead, Jim!");
-	  continue;
-	}
-
-      if (PREDICT_FALSE (s0->session_state == SESSION_STATE_CLOSED))
-	continue;
+	case FIFO_EVENT_APP_TX:
+	  s0 = session_event_get_session (e0, my_thread_index);
 
-      ASSERT (s0->thread_index == my_thread_index);
+	  if (CLIB_DEBUG && !s0)
+	    {
+	      clib_warning ("It's dead, Jim!");
+	      continue;
+	    }
 
-      switch (e0->event_type)
-	{
-	case FIFO_EVENT_SERVER_TX:
+	  if (PREDICT_FALSE (s0->session_state == SESSION_STATE_CLOSED))
+	    continue;
 	  /* Spray packets in per session type frames, since they go to
 	   * different nodes */
 	  rv = (smm->session_tx_fns[s0->session_type]) (vm, node, smm, e0, s0,
@@ -408,10 +417,12 @@ skip_dequeue:
 	    goto done;
 
 	  break;
-	case FIFO_EVENT_SERVER_EXIT:
+	case FIFO_EVENT_DISCONNECT:
+	  s0 = stream_session_get_from_handle (e0->session_handle);
 	  stream_session_disconnect (s0);
 	  break;
 	case FIFO_EVENT_BUILTIN_RX:
+	  s0 = session_event_get_session (e0, my_thread_index);
 	  svm_fifo_unset_event (s0->server_rx_fifo);
 	  /* Get session's server */
 	  app = application_get (s0->app_index);
diff --git a/src/vnet/session/segment_manager.c b/src/vnet/session/segment_manager.c
index 16e5bc56..e0532320 100644
--- a/src/vnet/session/segment_manager.c
+++ b/src/vnet/session/segment_manager.c
@@ -27,6 +27,11 @@ u32 segment_name_counter = 0;
  */
 segment_manager_t *segment_managers = 0;
 
+/**
+ * Process private segment index
+ */
+u32 private_segment_index = ~0;
+
 /**
  * Default fifo and segment size. TODO config.
  */
@@ -100,6 +105,26 @@ session_manager_add_first_segment (segment_manager_t * sm, u32 segment_size)
   return rv;
 }
 
+static void
+segment_manager_alloc_process_private_segment ()
+{
+  svm_fifo_segment_create_args_t _a, *a = &_a;
+
+  if (private_segment_index != ~0)
+    return;
+
+  memset (a, 0, sizeof (*a));
+  a->segment_name = "process-private-segment";
+  a->segment_size = ~0;
+  a->new_segment_index = ~0;
+
+  if (svm_fifo_segment_create_process_private (a))
+    clib_warning ("Failed to create process private segment");
+
+  private_segment_index = a->new_segment_index;
+  ASSERT (private_segment_index != ~0);
+}
+
 /**
  * Initializes segment manager based on options provided.
  * Returns error if svm segment allocation fails.
@@ -114,7 +139,9 @@ segment_manager_init (segment_manager_t * sm,
   /* app allocates these */
   sm->properties = properties;
 
-  if (first_seg_size > 0)
+  first_seg_size = first_seg_size > 0 ? first_seg_size : default_segment_size;
+
+  if (sm->properties->use_private_segment == 0)
     {
       rv = session_manager_add_first_segment (sm, first_seg_size);
       if (rv)
@@ -123,7 +150,15 @@ segment_manager_init (segment_manager_t * sm,
 	  return rv;
 	}
     }
+  else
+    {
+      if (private_segment_index == ~0)
+	segment_manager_alloc_process_private_segment ();
+      ASSERT (private_segment_index != ~0);
+      vec_add1 (sm->segment_indices, private_segment_index);
+    }
 
+  clib_spinlock_init (&sm->lockp);
   return 0;
 }
 
@@ -162,8 +197,8 @@ segment_manager_del (segment_manager_t * sm)
 	  stream_session_t *session;
 
 	  fifo = fifos[i];
-	  session_index = fifo->server_session_index;
-	  thread_index = fifo->server_thread_index;
+	  session_index = fifo->master_session_index;
+	  thread_index = fifo->master_thread_index;
 
 	  session = stream_session_get (session_index, thread_index);
 
@@ -183,7 +218,9 @@ segment_manager_del (segment_manager_t * sm)
 					deleted_thread_indices[i]);
 
 	  /* Instead of directly removing the session call disconnect */
-	  stream_session_disconnect (session);
+	  session_send_session_evt_to_thread (stream_session_handle (session),
+					      FIFO_EVENT_DISCONNECT,
+					      deleted_thread_indices[i]);
 
 	  /*
 	     stream_session_table_del (smm, session);
@@ -200,6 +237,7 @@ segment_manager_del (segment_manager_t * sm)
       /* svm_fifo_segment_delete (fifo_segment); */
     }
 
+  clib_spinlock_free (&sm->lockp);
   vec_free (deleted_sessions);
   vec_free (deleted_thread_indices);
   pool_put (segment_managers, sm);
@@ -232,9 +270,13 @@ segment_manager_alloc_session_fifos (segment_manager_t * sm,
   u8 added_a_segment = 0;
   int i;
 
-  /* Allocate svm fifos */
   ASSERT (vec_len (sm->segment_indices));
 
+  /* Make sure we don't have multiple threads trying to allocate segments
+   * at the same time. */
+  clib_spinlock_lock (&sm->lockp);
+
+  /* Allocate svm fifos */
 again:
   for (i = 0; i < vec_len (sm->segment_indices); i++)
     {
@@ -283,7 +325,9 @@ again:
 	    }
 
 	  if (session_manager_add_segment (sm))
-	    return VNET_API_ERROR_URI_FIFO_CREATE_FAILED;
+	    {
+	      return VNET_API_ERROR_URI_FIFO_CREATE_FAILED;
+	    }
 
 	  added_a_segment = 1;
 	  goto again;
@@ -295,14 +339,16 @@ again:
 	}
     }
 
-  if (added_a_segment)
-    return segment_manager_notify_app_seg_add (sm, *fifo_segment_index);
-
   /* Backpointers to segment manager */
   sm_index = segment_manager_index (sm);
   (*server_tx_fifo)->segment_manager = sm_index;
   (*server_rx_fifo)->segment_manager = sm_index;
 
+  clib_spinlock_unlock (&sm->lockp);
+
+  if (added_a_segment)
+    return segment_manager_notify_app_seg_add (sm, *fifo_segment_index);
+
   return 0;
 }
 
@@ -313,26 +359,72 @@ segment_manager_dealloc_fifos (u32 svm_segment_index, svm_fifo_t * rx_fifo,
   segment_manager_t *sm;
   svm_fifo_segment_private_t *fifo_segment;
 
+  sm = segment_manager_get_if_valid (rx_fifo->segment_manager);
+
+  /* It's possible to have no segment manager if the session was removed
+   * as result of a detach */
+  if (!sm)
+    return;
+
   fifo_segment = svm_fifo_get_segment (svm_segment_index);
   svm_fifo_segment_free_fifo (fifo_segment, rx_fifo);
   svm_fifo_segment_free_fifo (fifo_segment, tx_fifo);
 
-  /* If we have segment manager, try doing some cleanup.
-   * It's possible to have no segment manager if the session was removed
-   * as result of a detach */
-  sm = segment_manager_get_if_valid (rx_fifo->segment_manager);
-  if (sm)
+  /* Remove segment only if it holds no fifos and not the first */
+  if (sm->segment_indices[0] != svm_segment_index
+      && !svm_fifo_segment_has_fifos (fifo_segment))
     {
-      /* Remove segment only if it holds no fifos and not the first */
-      if (sm->segment_indices[0] != svm_segment_index
-	  && !svm_fifo_segment_has_fifos (fifo_segment))
-	{
-	  svm_fifo_segment_delete (fifo_segment);
-	  vec_del1 (sm->segment_indices, svm_segment_index);
-	}
+      svm_fifo_segment_delete (fifo_segment);
+      vec_del1 (sm->segment_indices, svm_segment_index);
     }
 }
 
+/**
+ * Allocates shm queue in the first segment
+ */
+unix_shared_memory_queue_t *
+segment_manager_alloc_queue (segment_manager_t * sm, u32 queue_size)
+{
+  ssvm_shared_header_t *sh;
+  svm_fifo_segment_private_t *segment;
+  unix_shared_memory_queue_t *q;
+  void *oldheap;
+
+  ASSERT (sm->segment_indices != 0);
+
+  segment = svm_fifo_get_segment (sm->segment_indices[0]);
+  sh = segment->ssvm.sh;
+
+  oldheap = ssvm_push_heap (sh);
+  q =
+    unix_shared_memory_queue_init (queue_size, sizeof (session_fifo_event_t),
+				   0 /* consumer pid */ , 0
+				   /* signal when queue non-empty */ );
+  ssvm_pop_heap (oldheap);
+  return q;
+}
+
+/**
+ * Frees shm queue allocated in the first segment
+ */
+void
+segment_manager_dealloc_queue (segment_manager_t * sm,
+			       unix_shared_memory_queue_t * q)
+{
+  ssvm_shared_header_t *sh;
+  svm_fifo_segment_private_t *segment;
+  void *oldheap;
+
+  ASSERT (sm->segment_indices != 0);
+
+  segment = svm_fifo_get_segment (sm->segment_indices[0]);
+  sh = segment->ssvm.sh;
+
+  oldheap = ssvm_push_heap (sh);
+  unix_shared_memory_queue_free (q);
+  ssvm_pop_heap (oldheap);
+}
+
 /*
  * fd.io coding-style-patch-verification: ON
  *
diff --git a/src/vnet/session/segment_manager.h b/src/vnet/session/segment_manager.h
index 778d6040..2710bb54 100644
--- a/src/vnet/session/segment_manager.h
+++ b/src/vnet/session/segment_manager.h
@@ -18,6 +18,10 @@
 #include <vnet/vnet.h>
 #include <svm/svm_fifo_segment.h>
 
+#include <vlibmemory/unix_shared_memory_queue.h>
+#include <vlibmemory/api.h>
+#include <vppinfra/lock.h>
+
 typedef struct _segment_manager_properties
 {
   /** Session fifo sizes.  */
@@ -30,10 +34,14 @@ typedef struct _segment_manager_properties
   /** Flag that indicates if additional segments should be created */
   u8 add_segment;
 
+  /** Use private memory segment instead of shared memory */
+  u8 use_private_segment;
 } segment_manager_properties_t;
 
 typedef struct _segment_manager
 {
+  clib_spinlock_t lockp;
+
   /** segments mapped by this manager */
   u32 *segment_indices;
 
@@ -95,6 +103,10 @@ segment_manager_alloc_session_fifos (segment_manager_t * sm,
 void
 segment_manager_dealloc_fifos (u32 svm_segment_index, svm_fifo_t * rx_fifo,
 			       svm_fifo_t * tx_fifo);
+unix_shared_memory_queue_t *segment_manager_alloc_queue (segment_manager_t *
+							 sm, u32 queue_size);
+void segment_manager_dealloc_queue (segment_manager_t * sm,
+				    unix_shared_memory_queue_t * q);
 
 #endif /* SRC_VNET_SESSION_SEGMENT_MANAGER_H_ */
 /*
diff --git a/src/vnet/session/session.c b/src/vnet/session/session.c
index e6cfe7da..d17c93f8 100644
--- a/src/vnet/session/session.c
+++ b/src/vnet/session/session.c
@@ -377,33 +377,6 @@ stream_session_lookup_transport6 (ip6_address_t * lcl, ip6_address_t * rmt,
   return 0;
 }
 
-/**
- * Allocate vpp event queue (once) per worker thread
- */
-void
-session_vpp_event_queue_allocate (session_manager_main_t * smm,
-				  u32 thread_index)
-{
-  api_main_t *am = &api_main;
-  void *oldheap;
-
-  if (smm->vpp_event_queues[thread_index] == 0)
-    {
-      /* Allocate event fifo in the /vpe-api shared-memory segment */
-      oldheap = svm_push_data_heap (am->vlib_rp);
-
-      smm->vpp_event_queues[thread_index] =
-	unix_shared_memory_queue_init (2048 /* nels $$$$ config */ ,
-				       sizeof (session_fifo_event_t),
-				       0 /* consumer pid */ ,
-				       0
-				       /* (do not) send signal when queue non-empty */
-	);
-
-      svm_pop_heap (oldheap);
-    }
-}
-
 int
 stream_session_create_i (segment_manager_t * sm, transport_connection_t * tc,
 			 stream_session_t ** ret_s)
@@ -428,11 +401,11 @@ stream_session_create_i (segment_manager_t * sm, transport_connection_t * tc,
 
   /* Initialize backpointers */
   pool_index = s - smm->sessions[thread_index];
-  server_rx_fifo->server_session_index = pool_index;
-  server_rx_fifo->server_thread_index = thread_index;
+  server_rx_fifo->master_session_index = pool_index;
+  server_rx_fifo->master_thread_index = thread_index;
 
-  server_tx_fifo->server_session_index = pool_index;
-  server_tx_fifo->server_thread_index = thread_index;
+  server_tx_fifo->master_session_index = pool_index;
+  server_tx_fifo->master_thread_index = thread_index;
 
   s->server_rx_fifo = server_rx_fifo;
   s->server_tx_fifo = server_tx_fifo;
@@ -485,7 +458,7 @@ stream_session_enqueue_data (transport_connection_t * tc, u8 * data, u16 len,
   if (PREDICT_FALSE (len > svm_fifo_max_enqueue (s->server_rx_fifo)))
     return -1;
 
-  enqueued = svm_fifo_enqueue_nowait (s->server_rx_fifo, s->pid, len, data);
+  enqueued = svm_fifo_enqueue_nowait (s->server_rx_fifo, len, data);
 
   if (queue_event)
     {
@@ -527,14 +500,14 @@ stream_session_peek_bytes (transport_connection_t * tc, u8 * buffer,
 			   u32 offset, u32 max_bytes)
 {
   stream_session_t *s = stream_session_get (tc->s_index, tc->thread_index);
-  return svm_fifo_peek (s->server_tx_fifo, s->pid, offset, max_bytes, buffer);
+  return svm_fifo_peek (s->server_tx_fifo, offset, max_bytes, buffer);
 }
 
 u32
 stream_session_dequeue_drop (transport_connection_t * tc, u32 max_bytes)
 {
   stream_session_t *s = stream_session_get (tc->s_index, tc->thread_index);
-  return svm_fifo_dequeue_drop (s->server_tx_fifo, s->pid, max_bytes);
+  return svm_fifo_dequeue_drop (s->server_tx_fifo, max_bytes);
 }
 
 /**
@@ -568,7 +541,7 @@ stream_session_enqueue_notify (stream_session_t * s, u8 block)
     {
       /* Fabricate event */
       evt.fifo = s->server_rx_fifo;
-      evt.event_type = FIFO_EVENT_SERVER_RX;
+      evt.event_type = FIFO_EVENT_APP_RX;
       evt.event_id = serial_number++;
 
       /* Add event to server's event queue */
@@ -899,37 +872,45 @@ stream_session_stop_listen (stream_session_t * s)
   return 0;
 }
 
+void
+session_send_session_evt_to_thread (u64 session_handle,
+				    fifo_event_type_t evt_type,
+				    u32 thread_index)
+{
+  static u16 serial_number = 0;
+  session_fifo_event_t evt;
+  unix_shared_memory_queue_t *q;
+
+  /* Fabricate event */
+  evt.session_handle = session_handle;
+  evt.event_type = evt_type;
+  evt.event_id = serial_number++;
+
+  q = session_manager_get_vpp_event_queue (thread_index);
+
+  /* Based on request block (or not) for lack of space */
+  if (PREDICT_TRUE (q->cursize < q->maxsize))
+    unix_shared_memory_queue_add (q, (u8 *) & evt,
+				  0 /* do wait for mutex */ );
+  else
+    {
+      clib_warning ("queue full");
+      return;
+    }
+}
+
 /**
  * Disconnect session and propagate to transport. This should eventually
  * result in a delete notification that allows us to cleanup session state.
  * Called for both active/passive disconnects.
+ *
+ * Should be called from the session's thread.
  */
 void
 stream_session_disconnect (stream_session_t * s)
 {
-//  session_fifo_event_t evt;
-
   s->session_state = SESSION_STATE_CLOSED;
-  /* RPC to vpp evt queue in the right thread */
-
   tp_vfts[s->session_type].close (s->connection_index, s->thread_index);
-
-//  {
-//  /* Fabricate event */
-//  evt.fifo = s->server_rx_fifo;
-//  evt.event_type = FIFO_EVENT_SERVER_RX;
-//  evt.event_id = serial_number++;
-//
-//  /* Based on request block (or not) for lack of space */
-//  if (PREDICT_TRUE(q->cursize < q->maxsize))
-//    unix_shared_memory_queue_add (app->event_queue, (u8 *) &evt,
-//                                0 /* do wait for mutex */);
-//  else
-//    {
-//      clib_warning("fifo full");
-//      return -1;
-//    }
-//  }
 }
 
 /**
@@ -976,6 +957,33 @@ session_get_transport_vft (u8 type)
   return &tp_vfts[type];
 }
 
+/**
+ * Allocate vpp event queue (once) per worker thread
+ */
+void
+session_vpp_event_queue_allocate (session_manager_main_t * smm,
+				  u32 thread_index)
+{
+  api_main_t *am = &api_main;
+  void *oldheap;
+
+  if (smm->vpp_event_queues[thread_index] == 0)
+    {
+      /* Allocate event fifo in the /vpe-api shared-memory segment */
+      oldheap = svm_push_data_heap (am->vlib_rp);
+
+      smm->vpp_event_queues[thread_index] =
+	unix_shared_memory_queue_init (2048 /* nels $$$$ config */ ,
+				       sizeof (session_fifo_event_t),
+				       0 /* consumer pid */ ,
+				       0
+				       /* (do not) send signal when queue non-empty */
+	);
+
+      svm_pop_heap (oldheap);
+    }
+}
+
 static clib_error_t *
 session_manager_main_enable (vlib_main_t * vm)
 {
@@ -1043,6 +1051,18 @@ session_manager_main_enable (vlib_main_t * vm)
   return 0;
 }
 
+void
+session_node_enable_disable (u8 is_en)
+{
+  u8 state = is_en ? VLIB_NODE_STATE_POLLING : VLIB_NODE_STATE_DISABLED;
+  /* *INDENT-OFF* */
+  foreach_vlib_main (({
+    vlib_node_set_state (this_vlib_main, session_queue_node.index,
+                         state);
+  }));
+  /* *INDENT-ON* */
+}
+
 clib_error_t *
 vnet_session_enable_disable (vlib_main_t * vm, u8 is_en)
 {
@@ -1051,16 +1071,14 @@ vnet_session_enable_disable (vlib_main_t * vm, u8 is_en)
       if (session_manager_main.is_enabled)
 	return 0;
 
-      vlib_node_set_state (vm, session_queue_node.index,
-			   VLIB_NODE_STATE_POLLING);
+      session_node_enable_disable (is_en);
 
       return session_manager_main_enable (vm);
     }
   else
     {
       session_manager_main.is_enabled = 0;
-      vlib_node_set_state (vm, session_queue_node.index,
-			   VLIB_NODE_STATE_DISABLED);
+      session_node_enable_disable (is_en);
     }
 
   return 0;
diff --git a/src/vnet/session/session.h b/src/vnet/session/session.h
index 6e4ea96d..8cd72f35 100644
--- a/src/vnet/session/session.h
+++ b/src/vnet/session/session.h
@@ -17,9 +17,6 @@
 
 #include <vnet/session/transport.h>
 #include <vlibmemory/unix_shared_memory_queue.h>
-#include <vlibmemory/api.h>
-#include <vppinfra/sparse_vec.h>
-#include <svm/svm_fifo_segment.h>
 #include <vnet/session/session_debug.h>
 #include <vnet/session/segment_manager.h>
 
@@ -31,10 +28,10 @@
 
 typedef enum
 {
-  FIFO_EVENT_SERVER_RX,
-  FIFO_EVENT_SERVER_TX,
+  FIFO_EVENT_APP_RX,
+  FIFO_EVENT_APP_TX,
   FIFO_EVENT_TIMEOUT,
-  FIFO_EVENT_SERVER_EXIT,
+  FIFO_EVENT_DISCONNECT,
   FIFO_EVENT_BUILTIN_RX
 } fifo_event_type_t;
 
@@ -96,7 +93,11 @@ typedef enum
 
 /* *INDENT-OFF* */
 typedef CLIB_PACKED (struct {
-  svm_fifo_t * fifo;
+  union
+    {
+      svm_fifo_t * fifo;
+      u64 session_handle;
+    };
   u8 event_type;
   u16 event_id;
 }) session_fifo_event_t;
@@ -370,7 +371,9 @@ int stream_session_listen (stream_session_t * s, transport_endpoint_t * tep);
 int stream_session_stop_listen (stream_session_t * s);
 void stream_session_disconnect (stream_session_t * s);
 void stream_session_cleanup (stream_session_t * s);
-
+void session_send_session_evt_to_thread (u64 session_handle,
+					 fifo_event_type_t evt_type,
+					 u32 thread_index);
 u8 *format_stream_session (u8 * s, va_list * args);
 
 void session_register_transport (u8 type, const transport_proto_vft_t * vft);
diff --git a/src/vnet/session/session_api.c b/src/vnet/session/session_api.c
index 8116b673..79d67a2f 100755
--- a/src/vnet/session/session_api.c
+++ b/src/vnet/session/session_api.c
@@ -96,7 +96,7 @@ send_session_accept_callback (stream_session_t * s)
   memset (mp, 0, sizeof (*mp));
 
   mp->_vl_msg_id = clib_host_to_net_u16 (VL_API_ACCEPT_SESSION);
-
+  mp->context = server->index;
   listener = listen_session_get (s->session_type, s->listener_index);
   tp_vft = session_get_transport_vft (s->session_type);
   tc = tp_vft->get_connection (s->connection_index, s->thread_index);
@@ -270,23 +270,6 @@ static session_cb_vft_t uri_session_cb_vft = {
   .redirect_connect_callback = redirect_connect_callback
 };
 
-static int
-api_session_not_valid (u32 session_index, u32 thread_index)
-{
-  session_manager_main_t *smm = vnet_get_session_manager_main ();
-  stream_session_t *pool;
-
-  if (thread_index >= vec_len (smm->sessions))
-    return VNET_API_ERROR_INVALID_VALUE;
-
-  pool = smm->sessions[thread_index];
-
-  if (pool_is_free_index (pool, session_index))
-    return VNET_API_ERROR_INVALID_VALUE_2;
-
-  return 0;
-}
-
 static void
 vl_api_session_enable_disable_t_handler (vl_api_session_enable_disable_t * mp)
 {
@@ -324,9 +307,9 @@ vl_api_application_attach_t_handler (vl_api_application_attach_t * mp)
   rv = vnet_application_attach (a);
 
 done:
+
   /* *INDENT-OFF* */
   REPLY_MACRO2 (VL_API_APPLICATION_ATTACH_REPLY, ({
-    rmp->retval = rv;
     if (!rv)
       {
 	rmp->segment_name_length = 0;
@@ -558,24 +541,33 @@ static void
 vl_api_accept_session_reply_t_handler (vl_api_accept_session_reply_t * mp)
 {
   stream_session_t *s;
-  int rv;
   u32 session_index, thread_index;
-  session_index = stream_session_index_from_handle (mp->handle);
-  thread_index = stream_session_thread_from_handle (mp->handle);
-  if (api_session_not_valid (session_index, thread_index))
-    return;
-
-  s = stream_session_get (session_index, thread_index);
-  rv = mp->retval;
+  vnet_disconnect_args_t _a, *a = &_a;
 
-  if (rv)
+  /* Server isn't interested, kill the session */
+  if (mp->retval)
     {
-      /* Server isn't interested, kill the session */
-      stream_session_disconnect (s);
-      return;
+      a->app_index = mp->context;
+      a->handle = mp->handle;
+      vnet_disconnect_session (a);
+    }
+  else
+    {
+      stream_session_parse_handle (mp->handle, &session_index, &thread_index);
+      s = stream_session_get_if_valid (session_index, thread_index);
+      if (!s)
+	{
+	  clib_warning ("session doesn't exist");
+	  return;
+	}
+      if (s->app_index != mp->context)
+	{
+	  clib_warning ("app doesn't own session");
+	  return;
+	}
+      /* XXX volatile? */
+      s->session_state = SESSION_STATE_READY;
     }
-
-  s->session_state = SESSION_STATE_READY;
 }
 
 static void
diff --git a/src/vnet/tcp/builtin_client.c b/src/vnet/tcp/builtin_client.c
index f8fbf28c..276beb21 100644
--- a/src/vnet/tcp/builtin_client.c
+++ b/src/vnet/tcp/builtin_client.c
@@ -62,8 +62,7 @@ send_test_chunk (tclient_main_t * tm, session_t * s)
   bytes_this_chunk = bytes_this_chunk < s->bytes_to_send
     ? bytes_this_chunk : s->bytes_to_send;
 
-  rv = svm_fifo_enqueue_nowait (s->server_tx_fifo, 0 /*pid */ ,
-				bytes_this_chunk,
+  rv = svm_fifo_enqueue_nowait (s->server_tx_fifo, bytes_this_chunk,
 				test_data + test_buf_offset);
 
   /* If we managed to enqueue data... */
@@ -95,7 +94,7 @@ send_test_chunk (tclient_main_t * tm, session_t * s)
 	{
 	  /* Fabricate TX event, send to vpp */
 	  evt.fifo = s->server_tx_fifo;
-	  evt.event_type = FIFO_EVENT_SERVER_TX;
+	  evt.event_type = FIFO_EVENT_APP_TX;
 	  evt.event_id = serial_number++;
 
 	  unix_shared_memory_queue_add (tm->vpp_event_queue, (u8 *) & evt,
@@ -113,7 +112,7 @@ receive_test_chunk (tclient_main_t * tm, session_t * s)
   /* Allow enqueuing of new event */
   // svm_fifo_unset_event (rx_fifo);
 
-  n_read = svm_fifo_dequeue_nowait (rx_fifo, 0, vec_len (tm->rx_buf),
+  n_read = svm_fifo_dequeue_nowait (rx_fifo, vec_len (tm->rx_buf),
 				    tm->rx_buf);
   if (n_read > 0)
     {
@@ -457,6 +456,8 @@ attach_builtin_test_clients ()
 
   options[SESSION_OPTIONS_ACCEPT_COOKIE] = 0x12345678;
   options[SESSION_OPTIONS_SEGMENT_SIZE] = (2 << 30);	/*$$$$ config / arg */
+  options[APP_OPTIONS_FLAGS] = APP_OPTIONS_FLAGS_BUILTIN_APP;
+
   a->options = options;
 
   return vnet_application_attach (a);
diff --git a/src/vnet/tcp/builtin_server.c b/src/vnet/tcp/builtin_server.c
index 8308e3d9..34682699 100644
--- a/src/vnet/tcp/builtin_server.c
+++ b/src/vnet/tcp/builtin_server.c
@@ -180,7 +180,7 @@ builtin_server_rx_callback (stream_session_t * s)
   vec_validate (bsm->rx_buf, max_transfer - 1);
   _vec_len (bsm->rx_buf) = max_transfer;
 
-  actual_transfer = svm_fifo_dequeue_nowait (rx_fifo, 0, max_transfer,
+  actual_transfer = svm_fifo_dequeue_nowait (rx_fifo, max_transfer,
 					     bsm->rx_buf);
   ASSERT (actual_transfer == max_transfer);
 
@@ -190,8 +190,7 @@ builtin_server_rx_callback (stream_session_t * s)
    * Echo back
    */
 
-  n_written =
-    svm_fifo_enqueue_nowait (tx_fifo, 0, actual_transfer, bsm->rx_buf);
+  n_written = svm_fifo_enqueue_nowait (tx_fifo, actual_transfer, bsm->rx_buf);
 
   if (n_written != max_transfer)
     clib_warning ("short trout!");
@@ -200,7 +199,7 @@ builtin_server_rx_callback (stream_session_t * s)
     {
       /* Fabricate TX event, send to vpp */
       evt.fifo = tx_fifo;
-      evt.event_type = FIFO_EVENT_SERVER_TX;
+      evt.event_type = FIFO_EVENT_APP_TX;
       evt.event_id = serial_number++;
 
       unix_shared_memory_queue_add (bsm->vpp_queue[s->thread_index],
@@ -288,6 +287,7 @@ server_attach ()
   a->options[SESSION_OPTIONS_SEGMENT_SIZE] = 128 << 20;
   a->options[SESSION_OPTIONS_RX_FIFO_SIZE] = 1 << 16;
   a->options[SESSION_OPTIONS_TX_FIFO_SIZE] = 1 << 16;
+  a->options[APP_OPTIONS_FLAGS] = APP_OPTIONS_FLAGS_BUILTIN_APP;
   a->segment_name = segment_name;
   a->segment_name_length = ARRAY_LEN (segment_name);
 
diff --git a/src/vnet/tcp/tcp.c b/src/vnet/tcp/tcp.c
index 12982589..245a35ab 100644
--- a/src/vnet/tcp/tcp.c
+++ b/src/vnet/tcp/tcp.c
@@ -487,7 +487,8 @@ u8 *
 format_tcp_connection (u8 * s, va_list * args)
 {
   tcp_connection_t *tc = va_arg (*args, tcp_connection_t *);
-
+  if (!tc)
+    return s;
   if (tc->c_is_ip4)
     {
       s = format (s, "[#%d][%s] %U:%d->%U:%d", tc->c_thread_index, "T",
@@ -747,12 +748,14 @@ void
 tcp_initialize_timer_wheels (tcp_main_t * tm)
 {
   tw_timer_wheel_16t_2w_512sl_t *tw;
-  vec_foreach (tw, tm->timer_wheels)
-  {
+  /* *INDENT-OFF* */
+  foreach_vlib_main (({
+    tw = &tm->timer_wheels[ii];
     tw_timer_wheel_init_16t_2w_512sl (tw, tcp_expired_timers_dispatch,
 				      100e-3 /* timer period 100ms */ , ~0);
-    tw->last_run_time = vlib_time_now (tm->vlib_main);
-  }
+    tw->last_run_time = vlib_time_now (this_vlib_main);
+  }));
+  /* *INDENT-ON* */
 }
 
 clib_error_t *
diff --git a/src/vnet/tcp/tcp_input.c b/src/vnet/tcp/tcp_input.c
index 97679aaf..3bd53878 100644
--- a/src/vnet/tcp/tcp_input.c
+++ b/src/vnet/tcp/tcp_input.c
@@ -1011,8 +1011,8 @@ tcp_session_enqueue_ooo (tcp_connection_t * tc, vlib_buffer_t * b,
 
   clib_warning ("ooo: offset %d len %d", offset, data_len);
 
-  rv = svm_fifo_enqueue_with_offset (s0->server_rx_fifo, s0->pid, offset,
-				     data_len, vlib_buffer_get_current (b));
+  rv = svm_fifo_enqueue_with_offset (s0->server_rx_fifo, offset, data_len,
+				     vlib_buffer_get_current (b));
 
   /* Nothing written */
   if (rv)
@@ -2392,8 +2392,8 @@ tcp46_input_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
 	    {
 	      t0 = vlib_add_trace (vm, node, b0, sizeof (*t0));
 	      clib_memcpy (&t0->tcp_header, tcp0, sizeof (t0->tcp_header));
-	      clib_memcpy (&t0->tcp_connection, tc0,
-			   sizeof (t0->tcp_connection));
+	      if (tc0)
+		clib_memcpy (&t0->tcp_connection, tc0, sizeof (*tc0));
 	    }
 
 	  vlib_validate_buffer_enqueue_x1 (vm, node, next_index, to_next,
diff --git a/src/vnet/tcp/tcp_output.c b/src/vnet/tcp/tcp_output.c
index a7be8bd5..4e1a7aa5 100644
--- a/src/vnet/tcp/tcp_output.c
+++ b/src/vnet/tcp/tcp_output.c
@@ -1558,7 +1558,6 @@ tcp46_send_reset_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
 	  vlib_buffer_t *b0;
 	  tcp_tx_trace_t *t0;
 	  tcp_header_t *th0;
-	  tcp_connection_t *tc0;
 	  u32 error0 = TCP_ERROR_RST_SENT, next0 = TCP_RESET_NEXT_IP_LOOKUP;
 
 	  bi0 = from[0];
@@ -1592,13 +1591,8 @@ tcp46_send_reset_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
 		th0 = ip4_next_header ((ip4_header_t *) th0);
 	      else
 		th0 = ip6_next_header ((ip6_header_t *) th0);
-	      tc0 =
-		tcp_connection_get (vnet_buffer (b0)->tcp.connection_index,
-				    my_thread_index);
 	      t0 = vlib_add_trace (vm, node, b0, sizeof (*t0));
 	      clib_memcpy (&t0->tcp_header, th0, sizeof (t0->tcp_header));
-	      clib_memcpy (&t0->tcp_connection, tc0,
-			   sizeof (t0->tcp_connection));
 	    }
 
 	  vlib_validate_buffer_enqueue_x1 (vm, node, next_index, to_next,
diff --git a/src/vnet/tcp/tcp_test.c b/src/vnet/tcp/tcp_test.c
index 890e50b9..0146154b 100644
--- a/src/vnet/tcp/tcp_test.c
+++ b/src/vnet/tcp/tcp_test.c
@@ -351,8 +351,7 @@ tcp_test_fifo1 (vlib_main_t * vm, unformat_input_t * input)
   /*
    * Enqueue an initial (un-dequeued) chunk
    */
-  rv = svm_fifo_enqueue_nowait (f, 0 /* pid */ ,
-				sizeof (u32), (u8 *) test_data);
+  rv = svm_fifo_enqueue_nowait (f, sizeof (u32), (u8 *) test_data);
   TCP_TEST ((rv == sizeof (u32)), "enqueued %d", rv);
   TCP_TEST ((f->tail == 4), "fifo tail %u", f->tail);
 
@@ -364,7 +363,7 @@ tcp_test_fifo1 (vlib_main_t * vm, unformat_input_t * input)
     {
       offset = (2 * i + 1) * sizeof (u32);
       data = (u8 *) (test_data + (2 * i + 1));
-      rv = svm_fifo_enqueue_with_offset (f, 0, offset, sizeof (u32), data);
+      rv = svm_fifo_enqueue_with_offset (f, offset, sizeof (u32), data);
       if (verbose)
 	vlib_cli_output (vm, "add [%d] [%d, %d]", 2 * i + 1, offset,
 			 offset + sizeof (u32));
@@ -393,7 +392,7 @@ tcp_test_fifo1 (vlib_main_t * vm, unformat_input_t * input)
     {
       offset = (2 * i + 0) * sizeof (u32);
       data = (u8 *) (test_data + (2 * i + 0));
-      rv = svm_fifo_enqueue_with_offset (f, 0, offset, sizeof (u32), data);
+      rv = svm_fifo_enqueue_with_offset (f, offset, sizeof (u32), data);
       if (verbose)
 	vlib_cli_output (vm, "add [%d] [%d, %d]", 2 * i, offset,
 			 offset + sizeof (u32));
@@ -418,8 +417,7 @@ tcp_test_fifo1 (vlib_main_t * vm, unformat_input_t * input)
   /*
    * Enqueue the missing u32
    */
-  rv = svm_fifo_enqueue_nowait (f, 0 /* pid */ , sizeof (u32),
-				(u8 *) (test_data + 2));
+  rv = svm_fifo_enqueue_nowait (f, sizeof (u32), (u8 *) (test_data + 2));
   if (verbose)
     vlib_cli_output (vm, "fifo after missing link: %U", format_svm_fifo, f,
 		     1);
@@ -432,8 +430,7 @@ tcp_test_fifo1 (vlib_main_t * vm, unformat_input_t * input)
    */
   for (i = 0; i < 7; i++)
     {
-      rv = svm_fifo_dequeue_nowait (f, 0 /* pid */ , sizeof (u32),
-				    (u8 *) & data_word);
+      rv = svm_fifo_dequeue_nowait (f, sizeof (u32), (u8 *) & data_word);
       if (rv != sizeof (u32))
 	{
 	  clib_warning ("bytes dequeues %u", rv);
@@ -457,7 +454,7 @@ tcp_test_fifo1 (vlib_main_t * vm, unformat_input_t * input)
     {
       offset = (2 * i + 1) * sizeof (u32);
       data = (u8 *) (test_data + (2 * i + 1));
-      rv = svm_fifo_enqueue_with_offset (f, 0, offset, sizeof (u32), data);
+      rv = svm_fifo_enqueue_with_offset (f, offset, sizeof (u32), data);
       if (verbose)
 	vlib_cli_output (vm, "add [%d] [%d, %d]", 2 * i + 1, offset,
 			 offset + sizeof (u32));
@@ -468,13 +465,13 @@ tcp_test_fifo1 (vlib_main_t * vm, unformat_input_t * input)
 	}
     }
 
-  rv = svm_fifo_enqueue_with_offset (f, 0, 8, 21, data);
+  rv = svm_fifo_enqueue_with_offset (f, 8, 21, data);
   TCP_TEST ((rv == 0), "ooo enqueued %u", rv);
   TCP_TEST ((svm_fifo_number_ooo_segments (f) == 1),
 	    "number of ooo segments %u", svm_fifo_number_ooo_segments (f));
 
   vec_validate (data_buf, vec_len (data));
-  svm_fifo_peek (f, 0, 0, vec_len (data), data_buf);
+  svm_fifo_peek (f, 0, vec_len (data), data_buf);
   if (compare_data (data_buf, data, 8, vec_len (data), &j))
     {
       TCP_TEST (0, "[%d] peeked %u expected %u", j, data_buf[j], data[j]);
@@ -491,7 +488,7 @@ tcp_test_fifo1 (vlib_main_t * vm, unformat_input_t * input)
     {
       offset = (2 * i + 1) * sizeof (u32);
       data = (u8 *) (test_data + (2 * i + 1));
-      rv = svm_fifo_enqueue_with_offset (f, 0, offset, sizeof (u32), data);
+      rv = svm_fifo_enqueue_with_offset (f, offset, sizeof (u32), data);
       if (verbose)
 	vlib_cli_output (vm, "add [%d] [%d, %d]", 2 * i + 1, offset,
 			 offset + sizeof (u32));
@@ -502,13 +499,13 @@ tcp_test_fifo1 (vlib_main_t * vm, unformat_input_t * input)
 	}
     }
 
-  rv = svm_fifo_enqueue_nowait (f, 0, 29, data);
+  rv = svm_fifo_enqueue_nowait (f, 29, data);
   TCP_TEST ((rv == 32), "ooo enqueued %u", rv);
   TCP_TEST ((svm_fifo_number_ooo_segments (f) == 0),
 	    "number of ooo segments %u", svm_fifo_number_ooo_segments (f));
 
   vec_validate (data_buf, vec_len (data));
-  svm_fifo_peek (f, 0, 0, vec_len (data), data_buf);
+  svm_fifo_peek (f, 0, vec_len (data), data_buf);
   if (compare_data (data_buf, data, 0, vec_len (data), &j))
     {
       TCP_TEST (0, "[%d] peeked %u expected %u", j, data_buf[j], data[j]);
@@ -551,7 +548,7 @@ tcp_test_fifo2 (vlib_main_t * vm)
     {
       tp = vp + i;
       data64 = tp->offset;
-      rv = svm_fifo_enqueue_with_offset (f, 0, tp->offset, tp->len,
+      rv = svm_fifo_enqueue_with_offset (f, tp->offset, tp->len,
 					 (u8 *) & data64);
     }
 
@@ -565,7 +562,7 @@ tcp_test_fifo2 (vlib_main_t * vm)
 	    "first ooo seg length %u", ooo_seg->length);
 
   data64 = 0;
-  rv = svm_fifo_enqueue_nowait (f, 0, sizeof (u32), (u8 *) & data64);
+  rv = svm_fifo_enqueue_nowait (f, sizeof (u32), (u8 *) & data64);
   TCP_TEST ((rv == 3000), "bytes to be enqueued %u", rv);
 
   svm_fifo_free (f);
@@ -581,7 +578,7 @@ tcp_test_fifo2 (vlib_main_t * vm)
     {
       tp = &test_data[i];
       data64 = tp->offset;
-      rv = svm_fifo_enqueue_with_offset (f, 0, tp->offset, tp->len,
+      rv = svm_fifo_enqueue_with_offset (f, tp->offset, tp->len,
 					 (u8 *) & data64);
       if (rv)
 	{
@@ -599,7 +596,7 @@ tcp_test_fifo2 (vlib_main_t * vm)
 	    "first ooo seg length %u", ooo_seg->length);
 
   data64 = 0;
-  rv = svm_fifo_enqueue_nowait (f, 0, sizeof (u32), (u8 *) & data64);
+  rv = svm_fifo_enqueue_nowait (f, sizeof (u32), (u8 *) & data64);
 
   TCP_TEST ((rv == 3000), "bytes to be enqueued %u", rv);
 
@@ -755,7 +752,7 @@ tcp_test_fifo3 (vlib_main_t * vm, unformat_input_t * input)
   for (i = 0; i < vec_len (generate); i++)
     {
       tp = generate + i;
-      rv = svm_fifo_enqueue_with_offset (f, 0, fifo_initial_offset
+      rv = svm_fifo_enqueue_with_offset (f, fifo_initial_offset
 					 + tp->offset, tp->len,
 					 (u8 *) data_pattern + tp->offset);
     }
@@ -776,7 +773,7 @@ tcp_test_fifo3 (vlib_main_t * vm, unformat_input_t * input)
       u32 bytes_to_enq = 1;
       if (in_seq_all)
 	bytes_to_enq = total_size;
-      rv = svm_fifo_enqueue_nowait (f, 0, bytes_to_enq, data_pattern + 0);
+      rv = svm_fifo_enqueue_nowait (f, bytes_to_enq, data_pattern + 0);
 
       if (verbose)
 	vlib_cli_output (vm, "in-order enqueue returned %d", rv);
@@ -793,7 +790,7 @@ tcp_test_fifo3 (vlib_main_t * vm, unformat_input_t * input)
    * Test if peeked data is the same as original data
    */
   vec_validate (data_buf, vec_len (data_pattern));
-  svm_fifo_peek (f, 0, 0, vec_len (data_pattern), data_buf);
+  svm_fifo_peek (f, 0, vec_len (data_pattern), data_buf);
   if (compare_data (data_buf, data_pattern, 0, vec_len (data_pattern), &j))
     {
       TCP_TEST (0, "[%d] peeked %u expected %u", j, data_buf[j],
@@ -806,11 +803,11 @@ tcp_test_fifo3 (vlib_main_t * vm, unformat_input_t * input)
    */
   if (drop)
     {
-      svm_fifo_dequeue_drop (f, 0, vec_len (data_pattern));
+      svm_fifo_dequeue_drop (f, vec_len (data_pattern));
     }
   else
     {
-      svm_fifo_dequeue_nowait (f, 0, vec_len (data_pattern), data_buf);
+      svm_fifo_dequeue_nowait (f, vec_len (data_pattern), data_buf);
       if (compare_data
 	  (data_buf, data_pattern, 0, vec_len (data_pattern), &j))
 	{
diff --git a/src/vnet/udp/builtin_server.c b/src/vnet/udp/builtin_server.c
index 8565f04c..18684d54 100644
--- a/src/vnet/udp/builtin_server.c
+++ b/src/vnet/udp/builtin_server.c
@@ -59,10 +59,10 @@ builtin_server_rx_callback (stream_session_t * s)
   vec_validate (my_copy_buffer, this_transfer - 1);
   _vec_len (my_copy_buffer) = this_transfer;
 
-  actual_transfer = svm_fifo_dequeue_nowait (rx_fifo, 0, this_transfer,
+  actual_transfer = svm_fifo_dequeue_nowait (rx_fifo, this_transfer,
 					     my_copy_buffer);
   ASSERT (actual_transfer == this_transfer);
-  actual_transfer = svm_fifo_enqueue_nowait (tx_fifo, 0, this_transfer,
+  actual_transfer = svm_fifo_enqueue_nowait (tx_fifo, this_transfer,
 					     my_copy_buffer);
   ASSERT (actual_transfer == this_transfer);
 
@@ -72,7 +72,7 @@ builtin_server_rx_callback (stream_session_t * s)
     {
       /* Fabricate TX event, send to ourselves */
       evt.fifo = tx_fifo;
-      evt.event_type = FIFO_EVENT_SERVER_TX;
+      evt.event_type = FIFO_EVENT_APP_TX;
       evt.event_id = 0;
       q = session_manager_get_vpp_event_queue (s->thread_index);
       unix_shared_memory_queue_add (q, (u8 *) & evt,
@@ -110,6 +110,8 @@ attach_builtin_uri_server ()
 
   options[SESSION_OPTIONS_ACCEPT_COOKIE] = 0x12345678;
   options[SESSION_OPTIONS_SEGMENT_SIZE] = (2 << 30);	/*$$$$ config / arg */
+  options[APP_OPTIONS_FLAGS] = APP_OPTIONS_FLAGS_BUILTIN_APP;
+
   a->options = options;
 
   return vnet_application_attach (a);
diff --git a/src/vnet/udp/udp_input.c b/src/vnet/udp/udp_input.c
index 810278e6..e6b4f8fc 100644
--- a/src/vnet/udp/udp_input.c
+++ b/src/vnet/udp/udp_input.c
@@ -145,8 +145,7 @@ udp4_uri_input_node_fn (vlib_main_t * vm,
 		  goto trace0;
 		}
 
-	      svm_fifo_enqueue_nowait (f0, 0 /* pid */ ,
-				       udp_len0 - sizeof (*udp0),
+	      svm_fifo_enqueue_nowait (f0, udp_len0 - sizeof (*udp0),
 				       (u8 *) (udp0 + 1));
 
 	      b0->error = node->errors[SESSION_ERROR_ENQUEUED];
@@ -255,7 +254,7 @@ udp4_uri_input_node_fn (vlib_main_t * vm,
 	{
 	  /* Fabricate event */
 	  evt.fifo = s0->server_rx_fifo;
-	  evt.event_type = FIFO_EVENT_SERVER_RX;
+	  evt.event_type = FIFO_EVENT_APP_RX;
 	  evt.event_id = serial_number++;
 
 	  /* Add event to server's event queue */
-- 
cgit 1.2.3-korg


From c28764fd356632763614ea579f678d8f55eca4c7 Mon Sep 17 00:00:00 2001
From: Florin Coras <fcoras@cisco.com>
Date: Wed, 26 Apr 2017 00:08:42 -0700
Subject: TCP ooo reception fixes

- Improve svm fifo handling of out-of-order segments
- Ensure tsval_recent is updated only if rcv_las falls withing the
  segments's sequence space
- Avoid directly dropping old ACKs
- Improve debugging

Change-Id: I88dbe2394a0ad7eb389a4cc12d013a13733953aa
Signed-off-by: Florin Coras <fcoras@cisco.com>
---
 src/svm/svm_fifo.c         | 144 ++++++++++++++++++++++-----------------------
 src/svm/svm_fifo.h         |   1 +
 src/vnet/session/session.c |  15 +++++
 src/vnet/session/session.h |   7 ++-
 src/vnet/tcp/tcp_debug.h   |  15 +++++
 src/vnet/tcp/tcp_error.def |   3 +-
 src/vnet/tcp/tcp_format.c  |   6 +-
 src/vnet/tcp/tcp_input.c   |  81 +++++++++++++++++--------
 src/vnet/tcp/tcp_output.c  |   4 +-
 src/vnet/tcp/tcp_test.c    | 114 ++++++++++++++++++++++++++++++++---
 10 files changed, 275 insertions(+), 115 deletions(-)

(limited to 'src/svm/svm_fifo.c')

diff --git a/src/svm/svm_fifo.c b/src/svm/svm_fifo.c
index 8f2ed0c9..9b09d0c2 100644
--- a/src/svm/svm_fifo.c
+++ b/src/svm/svm_fifo.c
@@ -17,6 +17,8 @@
 
 #define offset_lt(_a, _b) ((i32)((_a)-(_b)) < 0)
 #define offset_leq(_a, _b) ((i32)((_a)-(_b)) <= 0)
+#define offset_gt(_a, _b) ((i32)((_a)-(_b)) > 0)
+#define offset_geq(_a, _b) ((i32)((_a)-(_b)) >= 0)
 
 u8 *
 format_ooo_segment (u8 * s, va_list * args)
@@ -160,14 +162,23 @@ ooo_segment_add (svm_fifo_t * f, u32 offset, u32 length)
 	 && offset_leq (ooo_segment_offset (f, s), offset))
     s = pool_elt_at_index (f->ooo_segments, s->next);
 
+  /* If we have a previous and we overlap it, use it as starting point */
+  prev = ooo_segment_get_prev (f, s);
+  if (prev && offset_leq (offset, ooo_segment_end_offset (f, prev)))
+    {
+      s = prev;
+      prev = ooo_segment_get_prev (f, s);
+      s_sof = ooo_segment_offset (f, s);
+      s_eof = ooo_segment_end_offset (f, s);
+      goto merge;
+    }
+
   s_index = s - f->ooo_segments;
   s_sof = ooo_segment_offset (f, s);
   s_eof = ooo_segment_end_offset (f, s);
-  prev = ooo_segment_get_prev (f, s);
 
   /* No overlap, add before current segment */
-  if (offset_lt (end_offset, s_sof)
-      && (!prev || offset_lt (prev->start + prev->length, offset)))
+  if (offset_lt (end_offset, s_sof))
     {
       new_s = ooo_segment_new (f, offset, length);
       new_index = new_s - f->ooo_segments;
@@ -192,7 +203,7 @@ ooo_segment_add (svm_fifo_t * f, u32 offset, u32 length)
       return;
     }
   /* No overlap, add after current segment */
-  else if (offset_lt (s_eof, offset))
+  else if (offset_gt (offset, s_eof))
     {
       new_s = ooo_segment_new (f, offset, length);
       new_index = new_s - f->ooo_segments;
@@ -218,62 +229,16 @@ ooo_segment_add (svm_fifo_t * f, u32 offset, u32 length)
    * Merge needed
    */
 
+merge:
+
   /* Merge at head */
-  if (offset_leq (offset, s_sof))
+  if (offset_lt (offset, s_sof))
     {
-      /* If we have a previous, check if we overlap */
-      if (s->prev != OOO_SEGMENT_INVALID_INDEX)
-	{
-	  prev = pool_elt_at_index (f->ooo_segments, s->prev);
-
-	  /* New segment merges prev and current. Remove previous and
-	   * update position of current. */
-	  if (offset_leq (offset, ooo_segment_end_offset (f, prev)))
-	    {
-	      s->start = prev->start;
-	      s->length = s_eof - ooo_segment_offset (f, prev);
-	      ooo_segment_del (f, s->prev);
-	    }
-	  else
-	    {
-	      s->start = offset;
-	      s->length = s_eof - ooo_segment_offset (f, s);
-	    }
-	}
-      else
-	{
-	  s->start = offset;
-	  s->length = s_eof - ooo_segment_offset (f, s);
-	}
-
-      /* The new segment's tail may cover multiple smaller ones */
-      if (offset_lt (s_eof, end_offset))
-	{
-	  /* Remove segments completely covered */
-	  it = (s->next != OOO_SEGMENT_INVALID_INDEX) ?
-	    pool_elt_at_index (f->ooo_segments, s->next) : 0;
-	  while (it && offset_lt (ooo_segment_end_offset (f, it), end_offset))
-	    {
-	      next = (it->next != OOO_SEGMENT_INVALID_INDEX) ?
-		pool_elt_at_index (f->ooo_segments, it->next) : 0;
-	      ooo_segment_del (f, it - f->ooo_segments);
-	      it = next;
-	    }
-
-	  /* Update length. Segment's start might have changed. */
-	  s->length = end_offset - ooo_segment_offset (f, s);
-
-	  /* If partial overlap with last, merge */
-	  if (it && offset_lt (ooo_segment_offset (f, it), end_offset))
-	    {
-	      s->length +=
-		it->length - (ooo_segment_offset (f, it) - end_offset);
-	      ooo_segment_del (f, it - f->ooo_segments);
-	    }
-	}
+      s->start = offset;
+      s->length = s_eof - ooo_segment_offset (f, s);
     }
   /* Last but overlapping previous */
-  else if (offset_leq (s_eof, end_offset))
+  else if (offset_gt (end_offset, s_eof))
     {
       s->length = end_offset - ooo_segment_offset (f, s);
     }
@@ -281,8 +246,33 @@ ooo_segment_add (svm_fifo_t * f, u32 offset, u32 length)
   else
     {
       /* Do Nothing */
+      goto done;
+    }
+
+  /* The new segment's tail may cover multiple smaller ones */
+  if (offset_geq (end_offset, s_eof))
+    {
+      /* Remove the completely overlapped segments */
+      it = (s->next != OOO_SEGMENT_INVALID_INDEX) ?
+	pool_elt_at_index (f->ooo_segments, s->next) : 0;
+      while (it && offset_leq (ooo_segment_end_offset (f, it), end_offset))
+	{
+	  next = (it->next != OOO_SEGMENT_INVALID_INDEX) ?
+	    pool_elt_at_index (f->ooo_segments, it->next) : 0;
+	  ooo_segment_del (f, it - f->ooo_segments);
+	  it = next;
+	}
+
+      /* If partial overlap with last, merge */
+      if (it && offset_leq (ooo_segment_offset (f, it), end_offset))
+	{
+	  s->length = ooo_segment_end_offset (f, it) -
+	    ooo_segment_offset (f, s);
+	  ooo_segment_del (f, it - f->ooo_segments);
+	}
     }
 
+done:
   /* Most recently updated segment */
   f->ooos_newest = s - f->ooo_segments;
 }
@@ -296,14 +286,17 @@ ooo_segment_try_collect (svm_fifo_t * f, u32 n_bytes_enqueued)
 {
   ooo_segment_t *s;
   u32 index, bytes = 0, diff;
-  u32 cursize;
+  u32 cursize, norm_start, nitems;
 
   /* current size has not yet been updated */
   cursize = svm_fifo_max_dequeue (f) + n_bytes_enqueued;
+  nitems = f->nitems;
 
   s = pool_elt_at_index (f->ooo_segments, f->ooos_list_head);
 
-  diff = (f->nitems + (i32) (f->tail - s->start)) % f->nitems;
+  norm_start = s->start % nitems;
+  diff = (f->nitems + (i32) (f->tail - norm_start)) % nitems;
+
   if (diff > cursize)
     return 0;
 
@@ -326,7 +319,8 @@ ooo_segment_try_collect (svm_fifo_t * f, u32 n_bytes_enqueued)
       if (s->next != OOO_SEGMENT_INVALID_INDEX)
 	{
 	  s = pool_elt_at_index (f->ooo_segments, s->next);
-	  diff = (f->nitems + (i32) (f->tail - s->start)) % f->nitems;
+	  norm_start = s->start % nitems;
+	  diff = (f->nitems + (i32) (f->tail - norm_start)) % nitems;
 	  ooo_segment_del (f, index);
 	}
       /* End of search */
@@ -340,11 +334,11 @@ ooo_segment_try_collect (svm_fifo_t * f, u32 n_bytes_enqueued)
   /* If tail is adjacent to an ooo segment, 'consume' it */
   if (diff == 0)
     {
-      bytes = ((f->nitems - cursize) >= s->length) ? s->length :
-	f->nitems - cursize;
+      bytes = ((nitems - cursize) >= s->length) ? s->length :
+	nitems - cursize;
 
       f->tail += bytes;
-      f->tail %= f->nitems;
+      f->tail %= nitems;
 
       ooo_segment_del (f, s - f->ooo_segments);
     }
@@ -430,31 +424,22 @@ svm_fifo_enqueue_with_offset_internal (svm_fifo_t * f,
 {
   u32 total_copy_bytes, first_copy_bytes, second_copy_bytes;
   u32 cursize, nitems;
-  u32 normalized_offset;
-  int rv;
-
-  /* Users would do well to avoid this */
-  if (PREDICT_FALSE (f->tail == (offset % f->nitems)))
-    {
-      rv = svm_fifo_enqueue_internal (f, required_bytes, copy_from_here);
-      if (rv > 0)
-	return 0;
-      return -1;
-    }
+  u32 normalized_offset, offset_from_tail;
 
   /* read cursize, which can only increase while we're working */
   cursize = svm_fifo_max_dequeue (f);
   nitems = f->nitems;
+  normalized_offset = offset % nitems;
 
   /* Will this request fit? */
-  if ((required_bytes + (offset - f->tail) % nitems) > (nitems - cursize))
+  offset_from_tail = (nitems + normalized_offset - f->tail) % nitems;
+  if ((required_bytes + offset_from_tail) > (nitems - cursize))
     return -1;
 
   ooo_segment_add (f, offset, required_bytes);
 
   /* Number of bytes we're going to copy */
   total_copy_bytes = required_bytes;
-  normalized_offset = offset % nitems;
 
   /* Number of bytes in first copy segment */
   first_copy_bytes = ((nitems - normalized_offset) < total_copy_bytes)
@@ -631,6 +616,15 @@ svm_fifo_first_ooo_segment (svm_fifo_t * f)
   return pool_elt_at_index (f->ooo_segments, f->ooos_list_head);
 }
 
+/**
+ * Set fifo pointers to requested offset
+ */
+void
+svm_fifo_init_pointers (svm_fifo_t * f, u32 pointer)
+{
+  f->head = f->tail = pointer % f->nitems;
+}
+
 /*
  * fd.io coding-style-patch-verification: ON
  *
diff --git a/src/svm/svm_fifo.h b/src/svm/svm_fifo.h
index d67237c6..36158dc5 100644
--- a/src/svm/svm_fifo.h
+++ b/src/svm/svm_fifo.h
@@ -119,6 +119,7 @@ int svm_fifo_peek (svm_fifo_t * f, u32 offset, u32 max_bytes, u8 * copy_here);
 int svm_fifo_dequeue_drop (svm_fifo_t * f, u32 max_bytes);
 u32 svm_fifo_number_ooo_segments (svm_fifo_t * f);
 ooo_segment_t *svm_fifo_first_ooo_segment (svm_fifo_t * f);
+void svm_fifo_init_pointers (svm_fifo_t * f, u32 pointer);
 
 format_function_t format_svm_fifo;
 
diff --git a/src/vnet/session/session.c b/src/vnet/session/session.c
index d17c93f8..e92bb440 100644
--- a/src/vnet/session/session.c
+++ b/src/vnet/session/session.c
@@ -609,6 +609,21 @@ session_manager_flush_enqueue_events (u32 thread_index)
   return errors;
 }
 
+/**
+ * Init fifo tail and head pointers
+ *
+ * Useful if transport uses absolute offsets for tracking ooo segments.
+ */
+void
+stream_session_init_fifos_pointers (transport_connection_t * tc,
+				    u32 rx_pointer, u32 tx_pointer)
+{
+  stream_session_t *s;
+  s = stream_session_get (tc->s_index, tc->thread_index);
+  svm_fifo_init_pointers (s->server_rx_fifo, rx_pointer);
+  svm_fifo_init_pointers (s->server_tx_fifo, tx_pointer);
+}
+
 void
 stream_session_connect_notify (transport_connection_t * tc, u8 sst,
 			       u8 is_fail)
diff --git a/src/vnet/session/session.h b/src/vnet/session/session.h
index 8cd72f35..f41a8a96 100644
--- a/src/vnet/session/session.h
+++ b/src/vnet/session/session.h
@@ -352,9 +352,10 @@ stream_session_peek_bytes (transport_connection_t * tc, u8 * buffer,
 			   u32 offset, u32 max_bytes);
 u32 stream_session_dequeue_drop (transport_connection_t * tc, u32 max_bytes);
 
-void
-stream_session_connect_notify (transport_connection_t * tc, u8 sst,
-			       u8 is_fail);
+void stream_session_connect_notify (transport_connection_t * tc, u8 sst,
+				    u8 is_fail);
+void stream_session_init_fifos_pointers (transport_connection_t * tc,
+					 u32 rx_pointer, u32 tx_pointer);
 
 void stream_session_accept_notify (transport_connection_t * tc);
 void stream_session_disconnect_notify (transport_connection_t * tc);
diff --git a/src/vnet/tcp/tcp_debug.h b/src/vnet/tcp/tcp_debug.h
index ecbf7887..b4497a3b 100755
--- a/src/vnet/tcp/tcp_debug.h
+++ b/src/vnet/tcp/tcp_debug.h
@@ -50,6 +50,7 @@
   _(CC_EVT, "cc event")			\
   _(CC_PACK, "cc partial ack")		\
   _(SEG_INVALID, "invalid segment")	\
+  _(PAWS_FAIL, "failed paws check")	\
   _(ACK_RCV_ERR, "invalid ack")		\
   _(RCV_WND_SHRUNK, "shrunk rcv_wnd")	\
 
@@ -382,6 +383,20 @@ typedef enum _tcp_dbg_evt
   ed->data[4] = _tc->rcv_wnd;						\
 }
 
+#define TCP_EVT_PAWS_FAIL_HANDLER(_tc, _seq, _end, ...)			\
+{									\
+  ELOG_TYPE_DECLARE (_e) =						\
+  {									\
+    .format = "paws fail: seq %u end %u tsval %u tsval_recent %u",	\
+    .format_args = "i4i4i4i4",						\
+  };									\
+  DECLARE_ETD(_tc, _e, 4);						\
+  ed->data[0] = _seq - _tc->irs;					\
+  ed->data[1] = _end - _tc->irs;					\
+  ed->data[2] = _tc->opt.tsval;						\
+  ed->data[3] = _tc->tsval_recent;					\
+}
+
 #define TCP_EVT_ACK_RCV_ERR_HANDLER(_tc, _type, _ack, ...)		\
 {									\
   ELOG_TYPE_DECLARE (_e) =						\
diff --git a/src/vnet/tcp/tcp_error.def b/src/vnet/tcp/tcp_error.def
index 0d75d975..a4e46d64 100644
--- a/src/vnet/tcp/tcp_error.def
+++ b/src/vnet/tcp/tcp_error.def
@@ -37,4 +37,5 @@ tcp_error (PKTS_SENT, "Packets sent")
 tcp_error (FILTERED_DUPACKS, "Filtered duplicate ACKs")
 tcp_error (RST_SENT, "Resets sent")
 tcp_error (INVALID_CONNECTION, "Invalid connection")
-tcp_error (NO_WND, "No window")
\ No newline at end of file
+tcp_error (NO_WND, "No window")
+tcp_error (CONNECTION_CLOSED, "Connection closed")
\ No newline at end of file
diff --git a/src/vnet/tcp/tcp_format.c b/src/vnet/tcp/tcp_format.c
index 3148fd40..4de99235 100644
--- a/src/vnet/tcp/tcp_format.c
+++ b/src/vnet/tcp/tcp_format.c
@@ -131,11 +131,13 @@ format_tcp_header (u8 * s, va_list * args)
 u8 *
 format_tcp_sacks (u8 * s, va_list * args)
 {
-  sack_block_t *sacks = va_arg (*args, sack_block_t *);
+  tcp_connection_t *tc = va_arg (*args, tcp_connection_t *);
+  sack_block_t *sacks = tc->snd_sacks;
   sack_block_t *block;
   vec_foreach (block, sacks)
   {
-    s = format (s, " start %u end %u\n", block->start, block->end);
+    s = format (s, " start %u end %u\n", block->start - tc->irs,
+		block->end - tc->irs);
   }
   return s;
 }
diff --git a/src/vnet/tcp/tcp_input.c b/src/vnet/tcp/tcp_input.c
index 3c65a5ea..0030cfe2 100644
--- a/src/vnet/tcp/tcp_input.c
+++ b/src/vnet/tcp/tcp_input.c
@@ -208,6 +208,15 @@ tcp_options_parse (tcp_header_t * th, tcp_options_t * to)
     }
 }
 
+/**
+ * RFC1323: Check against wrapped sequence numbers (PAWS). If we have
+ * timestamp to echo and it's less than tsval_recent, drop segment
+ * but still send an ACK in order to retain TCP's mechanism for detecting
+ * and recovering from half-open connections
+ *
+ * Or at least that's what the theory says. It seems that this might not work
+ * very well with packet reordering and fast retransmit. XXX
+ */
 always_inline int
 tcp_segment_check_paws (tcp_connection_t * tc)
 {
@@ -215,6 +224,27 @@ tcp_segment_check_paws (tcp_connection_t * tc)
     && timestamp_lt (tc->opt.tsval, tc->tsval_recent);
 }
 
+/**
+ * Update tsval recent
+ */
+always_inline void
+tcp_update_timestamp (tcp_connection_t * tc, u32 seq, u32 seq_end)
+{
+  /*
+   * RFC1323: If Last.ACK.sent falls within the range of sequence numbers
+   * of an incoming segment:
+   *    SEG.SEQ <= Last.ACK.sent < SEG.SEQ + SEG.LEN
+   * then the TSval from the segment is copied to TS.Recent;
+   * otherwise, the TSval is ignored.
+   */
+  if (tcp_opts_tstamp (&tc->opt) && tc->tsval_recent
+      && seq_leq (seq, tc->rcv_las) && seq_leq (tc->rcv_las, seq_end))
+    {
+      tc->tsval_recent = tc->opt.tsval;
+      tc->tsval_recent_age = tcp_time_now ();
+    }
+}
+
 /**
  * Validate incoming segment as per RFC793 p. 69 and RFC1323 p. 19
  *
@@ -228,21 +258,16 @@ static int
 tcp_segment_validate (vlib_main_t * vm, tcp_connection_t * tc0,
 		      vlib_buffer_t * b0, tcp_header_t * th0, u32 * next0)
 {
-  u8 paws_failed;
-
   if (PREDICT_FALSE (!tcp_ack (th0) && !tcp_rst (th0) && !tcp_syn (th0)))
     return -1;
 
   tcp_options_parse (th0, &tc0->opt);
 
-  /* RFC1323: Check against wrapped sequence numbers (PAWS). If we have
-   * timestamp to echo and it's less than tsval_recent, drop segment
-   * but still send an ACK in order to retain TCP's mechanism for detecting
-   * and recovering from half-open connections */
-  paws_failed = tcp_segment_check_paws (tc0);
-  if (paws_failed)
+  if (tcp_segment_check_paws (tc0))
     {
       clib_warning ("paws failed");
+      TCP_EVT_DBG (TCP_EVT_PAWS_FAIL, tc0, vnet_buffer (b0)->tcp.seq_number,
+		   vnet_buffer (b0)->tcp.seq_end);
 
       /* If it just so happens that a segment updates tsval_recent for a
        * segment over 24 days old, invalidate tsval_recent. */
@@ -251,6 +276,7 @@ tcp_segment_validate (vlib_main_t * vm, tcp_connection_t * tc0,
 	{
 	  /* Age isn't reset until we get a valid tsval (bsd inspired) */
 	  tc0->tsval_recent = 0;
+	  clib_warning ("paws failed - really old segment. REALLY?");
 	}
       else
 	{
@@ -305,12 +331,9 @@ tcp_segment_validate (vlib_main_t * vm, tcp_connection_t * tc0,
       return -1;
     }
 
-  /* If PAWS passed and segment in window, save timestamp */
-  if (!paws_failed)
-    {
-      tc0->tsval_recent = tc0->opt.tsval;
-      tc0->tsval_recent_age = tcp_time_now ();
-    }
+  /* If segment in window, save timestamp */
+  tcp_update_timestamp (tc0, vnet_buffer (b0)->tcp.seq_number,
+			vnet_buffer (b0)->tcp.seq_end);
 
   return 0;
 }
@@ -835,7 +858,8 @@ tcp_rcv_ack (tcp_connection_t * tc, vlib_buffer_t * b,
 	  TCP_EVT_DBG (TCP_EVT_DUPACK_RCVD, tc);
 	  tcp_cc_rcv_dupack (tc, vnet_buffer (b)->tcp.ack_number);
 	}
-      return -1;
+      /* Don't drop yet */
+      return 0;
     }
 
   if (tcp_opts_sack_permitted (&tc->opt))
@@ -932,10 +956,6 @@ tcp_update_sack_list (tcp_connection_t * tc, u32 start, u32 end)
 	{
 	  vec_add1 (new_list, tc->snd_sacks[i]);
 	}
-      else
-	{
-	  clib_warning ("dropped sack blocks");
-	}
     }
 
   ASSERT (vec_len (new_list) <= TCP_MAX_SACK_BLOCKS);
@@ -1011,7 +1031,6 @@ tcp_session_enqueue_ooo (tcp_connection_t * tc, vlib_buffer_t * b,
 			 u16 data_len)
 {
   stream_session_t *s0;
-  u32 offset;
   int rv;
 
   /* Pure ACK. Do nothing */
@@ -1021,12 +1040,11 @@ tcp_session_enqueue_ooo (tcp_connection_t * tc, vlib_buffer_t * b,
     }
 
   s0 = stream_session_get (tc->c_s_index, tc->c_thread_index);
-  offset = vnet_buffer (b)->tcp.seq_number - tc->irs;
 
-  clib_warning ("ooo: offset %d len %d", offset, data_len);
-
-  rv = svm_fifo_enqueue_with_offset (s0->server_rx_fifo, offset, data_len,
-				     vlib_buffer_get_current (b));
+  /* Enqueue out-of-order data with absolute offset */
+  rv = svm_fifo_enqueue_with_offset (s0->server_rx_fifo,
+				     vnet_buffer (b)->tcp.seq_number,
+				     data_len, vlib_buffer_get_current (b));
 
   /* Nothing written */
   if (rv)
@@ -1542,6 +1560,9 @@ tcp46_syn_sent_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
 	      /* Notify app that we have connection */
 	      stream_session_connect_notify (&new_tc0->connection, sst, 0);
 
+	      stream_session_init_fifos_pointers (&new_tc0->connection,
+						  new_tc0->irs + 1,
+						  new_tc0->iss + 1);
 	      /* Make sure after data segment processing ACK is sent */
 	      new_tc0->flags |= TCP_CONN_SNDACK;
 	    }
@@ -1552,7 +1573,9 @@ tcp46_syn_sent_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
 
 	      /* Notify app that we have connection */
 	      stream_session_connect_notify (&new_tc0->connection, sst, 0);
-
+	      stream_session_init_fifos_pointers (&new_tc0->connection,
+						  new_tc0->irs + 1,
+						  new_tc0->iss + 1);
 	      tcp_make_synack (new_tc0, b0);
 	      next0 = tcp_next_output (is_ip4);
 
@@ -2139,6 +2162,10 @@ tcp46_listen_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
 	  tcp_make_synack (child0, b0);
 	  next0 = tcp_next_output (is_ip4);
 
+	  /* Init fifo pointers after we have iss */
+	  stream_session_init_fifos_pointers (&child0->connection,
+					      child0->irs + 1,
+					      child0->iss + 1);
 	drop:
 	  if (PREDICT_FALSE (b0->flags & VLIB_BUFFER_IS_TRACED))
 	    {
@@ -2474,6 +2501,7 @@ do {                                                       	\
   _(LISTEN, TCP_FLAG_SYN, TCP_INPUT_NEXT_LISTEN, TCP_ERROR_NONE);
   /* ACK for for a SYN-ACK -> tcp-rcv-process. */
   _(SYN_RCVD, TCP_FLAG_ACK, TCP_INPUT_NEXT_RCV_PROCESS, TCP_ERROR_NONE);
+  _(SYN_RCVD, TCP_FLAG_RST, TCP_INPUT_NEXT_RCV_PROCESS, TCP_ERROR_NONE);
   /* SYN-ACK for a SYN */
   _(SYN_SENT, TCP_FLAG_SYN | TCP_FLAG_ACK, TCP_INPUT_NEXT_SYN_SENT,
     TCP_ERROR_NONE);
@@ -2499,6 +2527,7 @@ do {                                                       	\
   _(FIN_WAIT_2, TCP_FLAG_FIN | TCP_FLAG_ACK, TCP_INPUT_NEXT_RCV_PROCESS,
     TCP_ERROR_NONE);
   _(LAST_ACK, TCP_FLAG_ACK, TCP_INPUT_NEXT_RCV_PROCESS, TCP_ERROR_NONE);
+  _(CLOSED, TCP_FLAG_ACK, TCP_INPUT_NEXT_DROP, TCP_ERROR_CONNECTION_CLOSED);
 #undef _
 }
 
diff --git a/src/vnet/tcp/tcp_output.c b/src/vnet/tcp/tcp_output.c
index 4e1a7aa5..a85d30da 100644
--- a/src/vnet/tcp/tcp_output.c
+++ b/src/vnet/tcp/tcp_output.c
@@ -359,7 +359,8 @@ tcp_make_established_options (tcp_connection_t * tc, tcp_options_t * opts)
 	{
 	  opts->flags |= TCP_OPTS_FLAG_SACK;
 	  opts->sacks = tc->snd_sacks;
-	  opts->n_sack_blocks = vec_len (tc->snd_sacks);
+	  opts->n_sack_blocks = clib_min (vec_len (tc->snd_sacks),
+					  TCP_OPTS_MAX_SACK_BLOCKS);
 	  len += 2 + TCP_OPTION_LEN_SACK_BLOCK * opts->n_sack_blocks;
 	}
     }
@@ -917,6 +918,7 @@ tcp_push_hdr_i (tcp_connection_t * tc, vlib_buffer_t * b,
   vnet_buffer (b)->tcp.connection_index = tc->c_c_index;
 
   tc->snd_nxt += data_len;
+  tc->rcv_las = tc->rcv_nxt;
 
   /* TODO this is updated in output as well ... */
   if (tc->snd_nxt > tc->snd_una_max)
diff --git a/src/vnet/tcp/tcp_test.c b/src/vnet/tcp/tcp_test.c
index ed032206..a457ac8f 100644
--- a/src/vnet/tcp/tcp_test.c
+++ b/src/vnet/tcp/tcp_test.c
@@ -231,7 +231,7 @@ tcp_test_sack_tx (vlib_main_t * vm, unformat_input_t * input)
   tcp_update_sack_list (tc, 300, 300);
   if (verbose)
     vlib_cli_output (vm, "overlap first 2 segments:\n%U",
-		     format_tcp_sacks, tc->snd_sacks);
+		     format_tcp_sacks, tc);
   TCP_TEST ((vec_len (tc->snd_sacks) == 3), "sack blocks %d expected %d",
 	    vec_len (tc->snd_sacks), 3);
   TCP_TEST ((tc->snd_sacks[0].start == 900),
@@ -244,7 +244,7 @@ tcp_test_sack_tx (vlib_main_t * vm, unformat_input_t * input)
   tcp_update_sack_list (tc, 1100, 1200);
   if (verbose)
     vlib_cli_output (vm, "add new segment [1100, 1200]\n%U",
-		     format_tcp_sacks, tc->snd_sacks);
+		     format_tcp_sacks, tc);
   TCP_TEST ((vec_len (tc->snd_sacks) == 4), "sack blocks %d expected %d",
 	    vec_len (tc->snd_sacks), 4);
   TCP_TEST ((tc->snd_sacks[0].start == 1100),
@@ -257,7 +257,7 @@ tcp_test_sack_tx (vlib_main_t * vm, unformat_input_t * input)
   tcp_update_sack_list (tc, 800, 900);
   if (verbose)
     vlib_cli_output (vm, "join middle segments [800, 900]\n%U",
-		     format_tcp_sacks, tc->snd_sacks);
+		     format_tcp_sacks, tc);
 
   TCP_TEST ((vec_len (tc->snd_sacks) == 3), "sack blocks %d expected %d",
 	    vec_len (tc->snd_sacks), 3);
@@ -271,8 +271,7 @@ tcp_test_sack_tx (vlib_main_t * vm, unformat_input_t * input)
   tc->rcv_nxt = 1200;
   tcp_update_sack_list (tc, 1200, 1200);
   if (verbose)
-    vlib_cli_output (vm, "advance rcv_nxt to 1200\n%U",
-		     format_tcp_sacks, tc->snd_sacks);
+    vlib_cli_output (vm, "advance rcv_nxt to 1200\n%U", format_tcp_sacks, tc);
   TCP_TEST ((vec_len (tc->snd_sacks) == 0), "sack blocks %d expected %d",
 	    vec_len (tc->snd_sacks), 0);
   return 0;
@@ -502,7 +501,13 @@ tcp_test_fifo1 (vlib_main_t * vm, unformat_input_t * input)
     {
       offset = (2 * i + 1) * sizeof (u32);
       data = (u8 *) (test_data + (2 * i + 1));
-      rv = svm_fifo_enqueue_with_offset (f, offset, sizeof (u32), data);
+      if (i == 0)
+	{
+	  rv = svm_fifo_enqueue_nowait (f, sizeof (u32), data);
+	  rv = rv > 0 ? 0 : rv;
+	}
+      else
+	rv = svm_fifo_enqueue_with_offset (f, offset, sizeof (u32), data);
       if (verbose)
 	vlib_cli_output (vm, "add [%d] [%d, %d]", 2 * i + 1, offset,
 			 offset + sizeof (u32));
@@ -517,6 +522,26 @@ tcp_test_fifo1 (vlib_main_t * vm, unformat_input_t * input)
     vlib_cli_output (vm, "fifo after odd segs: %U", format_svm_fifo, f, 1);
 
   TCP_TEST ((f->tail == 8), "fifo tail %u", f->tail);
+  TCP_TEST ((svm_fifo_number_ooo_segments (f) == 2),
+	    "number of ooo segments %u", svm_fifo_number_ooo_segments (f));
+
+  /*
+   * Try adding a completely overlapped segment
+   */
+  offset = 3 * sizeof (u32);
+  data = (u8 *) (test_data + 3);
+  rv = svm_fifo_enqueue_with_offset (f, offset, sizeof (u32), data);
+  if (rv)
+    {
+      clib_warning ("enqueue returned %d", rv);
+      goto err;
+    }
+
+  if (verbose)
+    vlib_cli_output (vm, "fifo after overlap seg: %U", format_svm_fifo, f, 1);
+
+  TCP_TEST ((svm_fifo_number_ooo_segments (f) == 2),
+	    "number of ooo segments %u", svm_fifo_number_ooo_segments (f));
 
   /*
    * Make sure format functions are not buggy
@@ -887,7 +912,7 @@ tcp_test_fifo3 (vlib_main_t * vm, unformat_input_t * input)
   f->head = fifo_initial_offset;
   f->tail = fifo_initial_offset;
 
-  for (i = 0; i < vec_len (generate); i++)
+  for (i = !randomize; i < vec_len (generate); i++)
     {
       tp = generate + i;
       svm_fifo_enqueue_with_offset (f, fifo_initial_offset + tp->offset,
@@ -895,6 +920,10 @@ tcp_test_fifo3 (vlib_main_t * vm, unformat_input_t * input)
 				    (u8 *) data_pattern + tp->offset);
     }
 
+  /* Add the first segment in order for non random data */
+  if (!randomize)
+    svm_fifo_enqueue_nowait (f, generate[0].len, (u8 *) data_pattern);
+
   /*
    * Expected result: one big fat chunk at offset 1 if randomize == 1
    */
@@ -964,6 +993,73 @@ tcp_test_fifo3 (vlib_main_t * vm, unformat_input_t * input)
   return 0;
 }
 
+static int
+tcp_test_fifo4 (vlib_main_t * vm, unformat_input_t * input)
+{
+  svm_fifo_t *f;
+  u32 fifo_size = 6 << 10;
+  u32 fifo_initial_offset = 1000000000;
+  u32 test_n_bytes = 5000, j;
+  u8 *test_data = 0, *data_buf = 0;
+  int i, rv, verbose = 0;
+
+  while (unformat_check_input (input) != UNFORMAT_END_OF_INPUT)
+    {
+      if (unformat (input, "verbose"))
+	verbose = 1;
+      else
+	{
+	  clib_error_t *e = clib_error_return
+	    (0, "unknown input `%U'", format_unformat_error, input);
+	  clib_error_report (e);
+	  return -1;
+	}
+    }
+
+  /*
+   * Create a fifo and add segments
+   */
+  f = fifo_prepare (fifo_size);
+
+  /* Set head and tail pointers */
+  fifo_initial_offset = fifo_initial_offset % fifo_size;
+  svm_fifo_init_pointers (f, fifo_initial_offset);
+
+  vec_validate (test_data, test_n_bytes - 1);
+  for (i = 0; i < vec_len (test_data); i++)
+    test_data[i] = i;
+
+  for (i = test_n_bytes - 1; i > 0; i--)
+    {
+      rv = svm_fifo_enqueue_with_offset (f, fifo_initial_offset + i,
+					 sizeof (u8), &test_data[i]);
+      if (verbose)
+	vlib_cli_output (vm, "add [%d] [%d, %d]", i, i, i + sizeof (u8));
+      if (rv)
+	{
+	  clib_warning ("enqueue returned %d", rv);
+	  svm_fifo_free (f);
+	  vec_free (test_data);
+	  return -1;
+	}
+    }
+
+  svm_fifo_enqueue_nowait (f, sizeof (u8), &test_data[0]);
+
+  vec_validate (data_buf, vec_len (test_data));
+
+  svm_fifo_dequeue_nowait (f, vec_len (test_data), data_buf);
+  rv = compare_data (data_buf, test_data, 0, vec_len (test_data), &j);
+  if (rv)
+    vlib_cli_output (vm, "[%d] dequeued %u expected %u", j, data_buf[j],
+		     test_data[j]);
+  TCP_TEST ((rv == 0), "dequeued compared to original returned %d", rv);
+
+  svm_fifo_free (f);
+  vec_free (test_data);
+  return 0;
+}
+
 static int
 tcp_test_fifo (vlib_main_t * vm, unformat_input_t * input)
 {
@@ -1028,6 +1124,10 @@ tcp_test_fifo (vlib_main_t * vm, unformat_input_t * input)
 	{
 	  res = tcp_test_fifo1 (vm, input);
 	}
+      else if (unformat (input, "fifo4"))
+	{
+	  res = tcp_test_fifo4 (vm, input);
+	}
     }
 
   return res;
-- 
cgit 1.2.3-korg


From bb292f4d3fbecfc6b1bac695f833b0da78369116 Mon Sep 17 00:00:00 2001
From: Florin Coras <fcoras@cisco.com>
Date: Fri, 19 May 2017 09:49:19 -0700
Subject: Improve session debugging

Also improves builtin client code.

Change-Id: I8bca1aa632028f95c373726efb0abf2ee0eff414
Signed-off-by: Florin Coras <fcoras@cisco.com>
---
 src/svm/svm_fifo.c             |  9 +++--
 src/vnet/session/session_cli.c | 47 +++++++++++++--------
 src/vnet/tcp/builtin_client.c  | 73 ++++++++++++++++-----------------
 src/vnet/tcp/builtin_client.h  |  3 --
 src/vnet/tcp/tcp.c             | 92 +++++++++++++++++++++++++++++++++++-------
 src/vnet/tcp/tcp.h             |  4 +-
 src/vnet/tcp/tcp_input.c       | 25 +++++++-----
 src/vnet/tcp/tcp_output.c      | 12 +++++-
 8 files changed, 176 insertions(+), 89 deletions(-)

(limited to 'src/svm/svm_fifo.c')

diff --git a/src/svm/svm_fifo.c b/src/svm/svm_fifo.c
index 9b09d0c2..f13f6fea 100644
--- a/src/svm/svm_fifo.c
+++ b/src/svm/svm_fifo.c
@@ -54,19 +54,20 @@ format_svm_fifo (u8 * s, va_list * args)
 
   s = format (s, "cursize %u nitems %u has_event %d\n",
 	      f->cursize, f->nitems, f->has_event);
-  s = format (s, "head %d tail %d\n", f->head, f->tail);
+  s = format (s, " head %d tail %d\n", f->head, f->tail);
 
   if (verbose > 1)
     s = format
-      (s, "server session %d thread %d client session %d thread %d\n",
+      (s, " server session %d thread %d client session %d thread %d\n",
        f->master_session_index, f->master_thread_index,
        f->client_session_index, f->client_thread_index);
 
   if (verbose)
     {
-      s = format (s, "ooo pool %d active elts\n",
+      s = format (s, " ooo pool %d active elts\n",
 		  pool_elts (f->ooo_segments));
-      s = format (s, "%U", format_ooo_list, f);
+      if (svm_fifo_has_ooo_data (f))
+	s = format (s, " %U", format_ooo_list, f);
     }
   return s;
 }
diff --git a/src/vnet/session/session_cli.c b/src/vnet/session/session_cli.c
index f5990c81..509eedbb 100755
--- a/src/vnet/session/session_cli.c
+++ b/src/vnet/session/session_cli.c
@@ -30,33 +30,35 @@ format_stream_session (u8 * s, va_list * args)
   int verbose = va_arg (*args, int);
   transport_proto_vft_t *tp_vft;
   u8 *str = 0;
-
   tp_vft = session_get_transport_vft (ss->session_type);
 
-  if (verbose)
-    str = format (0, "%-20llp%-20llp%-15lld", ss->server_rx_fifo,
-		  ss->server_tx_fifo, stream_session_get_index (ss));
+  if (verbose == 1)
+    str = format (0, "%-10u%-10u%-10lld",
+		  svm_fifo_max_dequeue (ss->server_rx_fifo),
+		  svm_fifo_max_enqueue (ss->server_tx_fifo),
+		  stream_session_get_index (ss));
 
   if (ss->session_state == SESSION_STATE_READY)
     {
-      s = format (s, "%-40U%v", tp_vft->format_connection,
-		  ss->connection_index, ss->thread_index, str);
+      s = format (s, "%U", tp_vft->format_connection, ss->connection_index,
+		  ss->thread_index, verbose);
+      if (verbose == 1)
+	s = format (s, "%v", str);
     }
   else if (ss->session_state == SESSION_STATE_LISTENING)
     {
       s = format (s, "%-40U%v", tp_vft->format_listener, ss->connection_index,
 		  str);
     }
-  else if (ss->session_state == SESSION_STATE_READY)
+  else if (ss->session_state == SESSION_STATE_CONNECTING)
     {
-      s =
-	format (s, "%-40U%v", tp_vft->format_half_open, ss->connection_index,
-		str);
+      s = format (s, "%-40U%v", tp_vft->format_half_open,
+		  ss->connection_index, str);
     }
   else if (ss->session_state == SESSION_STATE_CLOSED)
     {
       s = format (s, "[CL] %-40U%v", tp_vft->format_connection,
-		  ss->connection_index, ss->thread_index, str);
+		  ss->connection_index, ss->thread_index, verbose, str);
     }
   else
     {
@@ -85,7 +87,9 @@ show_session_command_fn (vlib_main_t * vm, unformat_input_t * input,
 
   while (unformat_check_input (input) != UNFORMAT_END_OF_INPUT)
     {
-      if (unformat (input, "verbose"))
+      if (unformat (input, "verbose %d", &verbose))
+	;
+      else if (unformat (input, "verbose"))
 	verbose = 1;
       else
 	break;
@@ -105,11 +109,11 @@ show_session_command_fn (vlib_main_t * vm, unformat_input_t * input,
 			   i, pool_elts (pool));
 	  if (verbose)
 	    {
-	      if (once_per_pool)
+	      if (once_per_pool && verbose == 1)
 		{
-		  str = format (str, "%-50s%-20s%-20s%-15s",
-				"Connection", "Rx fifo", "Tx fifo",
-				"Session Index");
+		  str =
+		    format (str, "%-50s%-15s%-10s%-10s%-10s", "Connection",
+			    "State", "Rx-f", "Tx-f", "S-idx");
 		  vlib_cli_output (vm, "%v", str);
 		  vec_reset_length (str);
 		  once_per_pool = 0;
@@ -118,7 +122,16 @@ show_session_command_fn (vlib_main_t * vm, unformat_input_t * input,
               /* *INDENT-OFF* */
               pool_foreach (s, pool,
               ({
-                vlib_cli_output (vm, "%U", format_stream_session, s, verbose);
+        	vec_reset_length (str);
+                str = format (str, "%U", format_stream_session, s, verbose);
+                if (verbose > 1)
+                  {
+                    str = format (str, " Rx fifo: %U", format_svm_fifo,
+				  s->server_rx_fifo, 1);
+                    str = format (str, " Tx fifo: %U", format_svm_fifo,
+				  s->server_tx_fifo, 1);
+                  }
+                vlib_cli_output (vm, "%v", str);
               }));
               /* *INDENT-ON* */
 	    }
diff --git a/src/vnet/tcp/builtin_client.c b/src/vnet/tcp/builtin_client.c
index d13fd446..aaefa7eb 100644
--- a/src/vnet/tcp/builtin_client.c
+++ b/src/vnet/tcp/builtin_client.c
@@ -187,6 +187,9 @@ tclient_thread_fn (void *arg)
 
   clib_per_cpu_mheaps[vlib_get_thread_index ()] = clib_per_cpu_mheaps[0];
 
+  vec_validate (session_indices, 0);
+  vec_reset_length (session_indices);
+
   while (1)
     {
       /* Wait until we're told to get busy */
@@ -201,14 +204,12 @@ tclient_thread_fn (void *arg)
       tm->run_test = 0;
       rx_total = 0;
 
-      clib_warning ("Run %d iterations", tm->n_iterations);
+      clib_warning ("Start test...");
 
       before = clib_time_now (&ttime);
 
-      for (i = 0; i < tm->n_iterations; i++)
+      do
 	{
-	  session_t *sp;
-
 	  do
 	    {
 	      try_tx = try_rx = 0;
@@ -229,24 +230,38 @@ tclient_thread_fn (void *arg)
                     receive_test_chunk (tm, sp);
                     try_rx = 1;
                   }
+                else
+                  {
+                    /* Session is complete */
+                    vec_add1 (session_indices, sp - tm->sessions);
+                  }
               }));
+              /* Terminate any completed sessions */
+              if (PREDICT_FALSE (_vec_len(session_indices) != 0))
+                {
+                  for (i = 0; i < _vec_len (session_indices); i++)
+                    {
+                      sp = pool_elt_at_index (tm->sessions, session_indices[i]);
+                      rx_total += sp->bytes_received;
+                      dmp = vl_msg_api_alloc_as_if_client (sizeof (*dmp));
+                      memset (dmp, 0, sizeof (*dmp));
+                      dmp->_vl_msg_id = ntohs (VL_API_DISCONNECT_SESSION);
+                      dmp->client_index = tm->my_client_index;
+                      dmp->handle = sp->vpp_session_handle;
+                      vl_msg_api_send_shmem (tm->vl_input_queue, (u8 *) & dmp);
+                      pool_put (tm->sessions, sp);
+                    }
+                  _vec_len(session_indices) = 0;
+                }
 	      /* *INDENT-ON* */
 	    }
 	  while (try_tx || try_rx);
-
-          /* *INDENT-OFF* */
-          pool_foreach (sp, tm->sessions,
-          ({
-            rx_total += sp->bytes_received;
-            sp->bytes_received = 0;
-            sp->bytes_to_send = tm->bytes_to_send;
-          }));
-          /* *INDENT-ON* */
 	}
+      while (0);
       after = clib_time_now (&ttime);
 
-      clib_warning ("Done %d iterations, %lld bytes in %.2f secs",
-		    tm->n_iterations, rx_total, (after - before));
+      clib_warning ("Test complete %lld bytes in %.2f secs",
+		    rx_total, (after - before));
       if ((after - before) != 0.0)
 	{
 	  clib_warning ("%.2f bytes/second full-duplex",
@@ -255,28 +270,11 @@ tclient_thread_fn (void *arg)
 			(((f64) rx_total * 8.0) / (after - before)) / 1e9);
 	}
 
-      /* Disconnect sessions... */
-      vec_reset_length (session_indices);
-
-      /* *INDENT-OFF* */
-      pool_foreach (sp, tm->sessions,
-      ({
-	vec_add1 (session_indices, sp - tm->sessions);
-      }));
-      /* *INDENT-ON* */
-
-      for (i = 0; i < vec_len (session_indices); i++)
-	{
-	  sp = pool_elt_at_index (tm->sessions, session_indices[i]);
-	  dmp = vl_msg_api_alloc_as_if_client (sizeof (*dmp));
-	  memset (dmp, 0, sizeof (*dmp));
-	  dmp->_vl_msg_id = ntohs (VL_API_DISCONNECT_SESSION);
-	  dmp->client_index = tm->my_client_index;
-	  dmp->handle = sp->vpp_session_handle;
-	  vl_msg_api_send_shmem (tm->vl_input_queue, (u8 *) & dmp);
-	  pool_put (tm->sessions, sp);
-	}
+      if (pool_elts (tm->sessions))
+	clib_warning ("BUG: %d active sessions remain...",
+		      pool_elts (tm->sessions));
     }
+  while (0);
   /* NOTREACHED */
 #if TCP_BUILTIN_CLIENT_PTHREAD
   return 0;
@@ -511,15 +509,12 @@ test_tcp_clients_command_fn (vlib_main_t * vm,
   u64 tmp;
 
   tm->bytes_to_send = 8192;
-  tm->n_iterations = 1;
   vec_free (tm->connect_uri);
 
   while (unformat_check_input (input) != UNFORMAT_END_OF_INPUT)
     {
       if (unformat (input, "nclients %d", &n_clients))
 	;
-      else if (unformat (input, "iterations %d", &tm->n_iterations))
-	;
       else if (unformat (input, "mbytes %lld", &tmp))
 	tm->bytes_to_send = tmp << 20;
       else if (unformat (input, "gbytes %lld", &tmp))
diff --git a/src/vnet/tcp/builtin_client.h b/src/vnet/tcp/builtin_client.h
index 756b3d18..57d112e6 100644
--- a/src/vnet/tcp/builtin_client.h
+++ b/src/vnet/tcp/builtin_client.h
@@ -93,9 +93,6 @@ typedef struct
   /* Signal variables */
   volatile int run_test;
 
-  /* Number of iterations */
-  int n_iterations;
-
   /* Bytes to send */
   u64 bytes_to_send;
 
diff --git a/src/vnet/tcp/tcp.c b/src/vnet/tcp/tcp.c
index 36d85e46..9b7b2f65 100644
--- a/src/vnet/tcp/tcp.c
+++ b/src/vnet/tcp/tcp.c
@@ -461,13 +461,12 @@ const char *tcp_fsm_states[] = {
 u8 *
 format_tcp_state (u8 * s, va_list * args)
 {
-  tcp_state_t *state = va_arg (*args, tcp_state_t *);
+  u32 state = va_arg (*args, u32);
 
-  if (*state < TCP_N_STATES)
-    s = format (s, "%s", tcp_fsm_states[*state]);
+  if (state < TCP_N_STATES)
+    s = format (s, "%s", tcp_fsm_states[state]);
   else
-    s = format (s, "UNKNOWN (%d (0x%x))", *state, *state);
-
+    s = format (s, "UNKNOWN (%d (0x%x))", state, state);
   return s;
 }
 
@@ -503,7 +502,51 @@ format_tcp_timers (u8 * s, va_list * args)
 }
 
 u8 *
-format_tcp_connection (u8 * s, va_list * args)
+format_tcp_congestion_status (u8 * s, va_list * args)
+{
+  tcp_connection_t *tc = va_arg (*args, tcp_connection_t *);
+  if (tcp_in_recovery (tc))
+    s = format (s, "recovery");
+  else if (tcp_in_fastrecovery (tc))
+    s = format (s, "fastrecovery");
+  else
+    s = format (s, "none");
+  return s;
+}
+
+u8 *
+format_tcp_vars (u8 * s, va_list * args)
+{
+  tcp_connection_t *tc = va_arg (*args, tcp_connection_t *);
+  s = format (s, " snd_una %u snd_nxt %u snd_una_max %u\n",
+	      tc->snd_una - tc->iss, tc->snd_nxt - tc->iss,
+	      tc->snd_una_max - tc->iss);
+  s = format (s, " rcv_nxt %u rcv_las %u\n",
+	      tc->rcv_nxt - tc->irs, tc->rcv_las - tc->irs);
+  s = format (s, " snd_wnd %u rcv_wnd %u snd_wl1 %u snd_wl2 %u\n",
+	      tc->snd_wnd, tc->rcv_wnd, tc->snd_wl1 - tc->irs,
+	      tc->snd_wl2 - tc->iss);
+  s = format (s, " flight size %u send space %u rcv_wnd available %d\n",
+	      tcp_flight_size (tc), tcp_snd_space (tc),
+	      tc->rcv_wnd - (tc->rcv_nxt - tc->rcv_las));
+  s = format (s, " cong %U ", format_tcp_congestion_status, tc);
+  s = format (s, "cwnd %u ssthresh %u rtx_bytes %u bytes_acked %u\n",
+	      tc->cwnd, tc->ssthresh, tc->rtx_bytes, tc->bytes_acked);
+  s = format (s, " prev_ssthresh %u snd_congestion %u\n", tc->prev_ssthresh,
+	      tc->snd_congestion - tc->iss);
+  s = format (s, " rto %u rto_boff %u srtt %u rttvar %u rtt_ts %u ", tc->rto,
+	      tc->rto_boff, tc->srtt, tc->rttvar, tc->rtt_ts);
+  s = format (s, "rtt_seq %u\n", tc->rtt_seq);
+  if (scoreboard_first_hole (&tc->sack_sb))
+    s = format (s, " scoreboard: %U\n", format_tcp_scoreboard, &tc->sack_sb);
+  if (vec_len (tc->snd_sacks))
+    s = format (s, " sacks tx: %U\n", format_tcp_sacks, tc);
+
+  return s;
+}
+
+u8 *
+format_tcp_connection_id (u8 * s, va_list * args)
 {
   tcp_connection_t *tc = va_arg (*args, tcp_connection_t *);
   if (!tc)
@@ -527,11 +570,18 @@ format_tcp_connection (u8 * s, va_list * args)
 }
 
 u8 *
-format_tcp_connection_verbose (u8 * s, va_list * args)
+format_tcp_connection (u8 * s, va_list * args)
 {
   tcp_connection_t *tc = va_arg (*args, tcp_connection_t *);
-  s = format (s, "%U %U %U", format_tcp_connection, tc, format_tcp_state,
-	      &tc->state, format_tcp_timers, tc);
+  u32 verbose = va_arg (*args, u32);
+
+  s = format (s, "%-50U", format_tcp_connection_id, tc);
+  if (verbose)
+    {
+      s = format (s, "%-15U", format_tcp_state, tc->state);
+      if (verbose > 1)
+	s = format (s, " %U\n%U", format_tcp_timers, tc, format_tcp_vars, tc);
+    }
   return s;
 }
 
@@ -540,11 +590,12 @@ format_tcp_session (u8 * s, va_list * args)
 {
   u32 tci = va_arg (*args, u32);
   u32 thread_index = va_arg (*args, u32);
+  u32 verbose = va_arg (*args, u32);
   tcp_connection_t *tc;
 
   tc = tcp_connection_get (tci, thread_index);
   if (tc)
-    return format (s, "%U", format_tcp_connection, tc);
+    return format (s, "%U", format_tcp_connection, tc, verbose);
   else
     return format (s, "empty");
 }
@@ -554,7 +605,7 @@ format_tcp_listener_session (u8 * s, va_list * args)
 {
   u32 tci = va_arg (*args, u32);
   tcp_connection_t *tc = tcp_listener_get (tci);
-  return format (s, "%U", format_tcp_connection, tc);
+  return format (s, "%U", format_tcp_connection_id, tc);
 }
 
 u8 *
@@ -562,7 +613,7 @@ format_tcp_half_open_session (u8 * s, va_list * args)
 {
   u32 tci = va_arg (*args, u32);
   tcp_connection_t *tc = tcp_half_open_connection_get (tci);
-  return format (s, "%U", format_tcp_connection, tc);
+  return format (s, "%U", format_tcp_connection_id, tc);
 }
 
 u8 *
@@ -659,12 +710,18 @@ tcp_round_snd_space (tcp_connection_t * tc, u32 snd_space)
 
 /**
  * Compute tx window session is allowed to fill.
+ *
+ * Takes into account available send space, snd_mss and the congestion
+ * state of the connection. If possible, the value returned is a multiple
+ * of snd_mss.
+ *
+ * @param tc tcp connection
+ * @return number of bytes session is allowed to write
  */
 u32
-tcp_session_send_space (transport_connection_t * trans_conn)
+tcp_snd_space (tcp_connection_t * tc)
 {
   int snd_space;
-  tcp_connection_t *tc = (tcp_connection_t *) trans_conn;
 
   /* If we haven't gotten dupacks or if we did and have gotten sacked bytes
    * then we can still send */
@@ -697,6 +754,13 @@ tcp_session_send_space (transport_connection_t * trans_conn)
   return 0;
 }
 
+u32
+tcp_session_send_space (transport_connection_t * trans_conn)
+{
+  tcp_connection_t *tc = (tcp_connection_t *) trans_conn;
+  return tcp_snd_space (tc);
+}
+
 u32
 tcp_session_tx_fifo_offset (transport_connection_t * trans_conn)
 {
diff --git a/src/vnet/tcp/tcp.h b/src/vnet/tcp/tcp.h
index 8d24a70b..c3ebe22b 100644
--- a/src/vnet/tcp/tcp.h
+++ b/src/vnet/tcp/tcp.h
@@ -387,8 +387,8 @@ void tcp_connection_cleanup (tcp_connection_t * tc);
 void tcp_connection_del (tcp_connection_t * tc);
 void tcp_connection_reset (tcp_connection_t * tc);
 
+u8 *format_tcp_connection_id (u8 * s, va_list * args);
 u8 *format_tcp_connection (u8 * s, va_list * args);
-u8 *format_tcp_connection_verbose (u8 * s, va_list * args);
 u8 *format_tcp_scoreboard (u8 * s, va_list * args);
 
 always_inline tcp_connection_t *
@@ -481,6 +481,8 @@ tcp_available_snd_space (const tcp_connection_t * tc)
   return available_wnd - flight_size;
 }
 
+u32 tcp_rcv_wnd_available (tcp_connection_t * tc);
+u32 tcp_snd_space (tcp_connection_t * tc);
 void tcp_update_rcv_wnd (tcp_connection_t * tc);
 
 void tcp_retransmit_first_unacked (tcp_connection_t * tc);
diff --git a/src/vnet/tcp/tcp_input.c b/src/vnet/tcp/tcp_input.c
index 9d3f4cc3..35bc9094 100644
--- a/src/vnet/tcp/tcp_input.c
+++ b/src/vnet/tcp/tcp_input.c
@@ -722,9 +722,11 @@ tcp_update_snd_wnd (tcp_connection_t * tc, u32 seq, u32 ack, u32 snd_wnd)
       TCP_EVT_DBG (TCP_EVT_SND_WND, tc);
 
       /* Set probe timer if we just got 0 wnd */
-      if (tc->snd_wnd < tc->snd_mss
-	  && !tcp_timer_is_active (tc, TCP_TIMER_PERSIST))
-	tcp_persist_timer_set (tc);
+      if (tc->snd_wnd < tc->snd_mss)
+	{
+	  if (!tcp_timer_is_active (tc, TCP_TIMER_PERSIST))
+	    tcp_persist_timer_set (tc);
+	}
       else
 	tcp_persist_timer_reset (tc);
     }
@@ -763,6 +765,7 @@ static void
 tcp_cc_rcv_ack (tcp_connection_t * tc, vlib_buffer_t * b)
 {
   u8 partial_ack;
+  u32 bytes_advanced;
 
   if (tcp_in_fastrecovery (tc))
     {
@@ -804,10 +807,14 @@ tcp_cc_rcv_ack (tcp_connection_t * tc, vlib_buffer_t * b)
       tc->rcv_dupacks = 0;
       if (tcp_in_recovery (tc))
 	{
-	  tc->rtx_bytes -= clib_min (tc->bytes_acked, tc->rtx_bytes);
+	  bytes_advanced = tc->bytes_acked + tc->sack_sb.snd_una_adv;
+	  tc->rtx_bytes -= clib_min (bytes_advanced, tc->rtx_bytes);
 	  tc->rto = clib_min (tc->srtt + (tc->rttvar << 2), TCP_RTO_MAX);
 	  if (seq_geq (tc->snd_una, tc->snd_congestion))
-	    tcp_recovery_off (tc);
+	    {
+	      tc->rtx_bytes = 0;
+	      tcp_recovery_off (tc);
+	    }
 	}
     }
 }
@@ -1221,7 +1228,7 @@ format_tcp_rx_trace (u8 * s, va_list * args)
   s = format (s, "%U\n%U%U",
 	      format_tcp_header, &t->tcp_header, 128,
 	      format_white_space, indent,
-	      format_tcp_connection_verbose, &t->tcp_connection);
+	      format_tcp_connection, &t->tcp_connection, 1);
 
   return s;
 }
@@ -1236,7 +1243,7 @@ format_tcp_rx_trace_short (u8 * s, va_list * args)
   s = format (s, "%d -> %d (%U)",
 	      clib_net_to_host_u16 (t->tcp_header.src_port),
 	      clib_net_to_host_u16 (t->tcp_header.dst_port), format_tcp_state,
-	      &t->tcp_connection.state);
+	      t->tcp_connection.state);
 
   return s;
 }
@@ -2165,6 +2172,7 @@ tcp46_listen_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
 	  child0->c_rmt_port = th0->src_port;
 	  child0->c_is_ip4 = is_ip4;
 	  child0->c_thread_index = my_thread_index;
+	  child0->state = TCP_STATE_SYN_RCVD;
 
 	  if (is_ip4)
 	    {
@@ -2194,7 +2202,6 @@ tcp46_listen_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
 	  child0->irs = vnet_buffer (b0)->tcp.seq_number;
 	  child0->rcv_nxt = vnet_buffer (b0)->tcp.seq_number + 1;
 	  child0->rcv_las = child0->rcv_nxt;
-	  child0->state = TCP_STATE_SYN_RCVD;
 
 	  /* RFC1323: TSval timestamps sent on {SYN} and {SYN,ACK}
 	   * segments are used to initialize PAWS. */
@@ -2450,7 +2457,7 @@ tcp46_input_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
 
 		  if (error0 == TCP_ERROR_DISPATCH)
 		    clib_warning ("disp error state %U flags %U",
-				  format_tcp_state, &state0, format_tcp_flags,
+				  format_tcp_state, state0, format_tcp_flags,
 				  (int) flags0);
 		}
 	    }
diff --git a/src/vnet/tcp/tcp_output.c b/src/vnet/tcp/tcp_output.c
index a462d8da..49fd6bef 100644
--- a/src/vnet/tcp/tcp_output.c
+++ b/src/vnet/tcp/tcp_output.c
@@ -59,7 +59,7 @@ format_tcp_tx_trace (u8 * s, va_list * args)
   s = format (s, "%U\n%U%U",
 	      format_tcp_header, &t->tcp_header, 128,
 	      format_white_space, indent,
-	      format_tcp_connection_verbose, &t->tcp_connection);
+	      format_tcp_connection, &t->tcp_connection, 1);
 
   return s;
 }
@@ -1024,6 +1024,11 @@ tcp_prepare_retransmit_segment (tcp_connection_t * tc, vlib_buffer_t * b,
   ASSERT (n_bytes != 0);
   b->current_length = n_bytes;
   tcp_push_hdr_i (tc, b, tc->state, 0);
+
+  /* Don't count multiple retransmits of the same segment */
+  if (tc->rto_boff > 1)
+    goto done;
+
   tc->rtx_bytes += n_bytes;
 
 done:
@@ -1103,7 +1108,9 @@ tcp_timer_retransmit_handler_i (u32 index, u8 is_syn)
 
       if (n_bytes == 0)
 	{
-	  clib_warning ("could not retransmit");
+	  clib_warning ("could not retransmit anything");
+	  /* Try again eventually */
+	  tcp_retransmit_timer_set (tc);
 	  return;
 	}
     }
@@ -1203,6 +1210,7 @@ tcp_timer_persist_handler (u32 index)
   /* Nothing to send */
   if (n_bytes == 0)
     {
+      clib_warning ("persist found nothing to send");
       tcp_return_buffer (tm);
       return;
     }
-- 
cgit 1.2.3-korg


From 93992a9048cb6e5dcd22de5091e72de778122627 Mon Sep 17 00:00:00 2001
From: Florin Coras <fcoras@cisco.com>
Date: Wed, 24 May 2017 18:03:56 -0700
Subject: Implement sack based tcp loss recovery (RFC 6675)

- refactor existing congestion control code (RFC 6582/5681). Handling of ack
  feedback now consists of: ack parsing, cc event detection, event handling,
  congestion control update
- extend sack scoreboard to support sack based retransmissions
- basic implementation of Eifel detection algorithm (RFC 3522) for
  detecting spurious retransmissions
- actually initialize the per-thread frame freelist hash tables
- increase worker stack size to 2mb
- fix session queue node out-of-buffer handling
  - ensure that the local buffer cache vec_len matches reality
  - avoid 2x spurious event requeues when short of buffers
  - count out-of-buffer events
- make the builtin server thread-safe
- fix bihash template threading issue: need to paint -1 across uninitialized
  working_copy_length vector elements (via rebase from master)

Change-Id: I646cb9f1add9a67d08f4a87badbcb117980ebfc4
Signed-off-by: Florin Coras <fcoras@cisco.com>
Signed-off-by: Dave Barach <dbarach@cisco.com>
---
 src/svm/svm_fifo.c             |   5 +-
 src/vlib/node.c                |   1 +
 src/vlib/threads.c             |   2 +-
 src/vlib/threads.h             |   2 +-
 src/vnet/session/node.c        |  53 ++--
 src/vnet/session/session.c     |  11 +-
 src/vnet/session/session.h     |   6 +-
 src/vnet/session/session_cli.c |  26 +-
 src/vnet/tcp/builtin_client.c  |  40 ++-
 src/vnet/tcp/builtin_server.c  |  20 +-
 src/vnet/tcp/tcp.c             |  57 ++--
 src/vnet/tcp/tcp.h             | 112 +++++--
 src/vnet/tcp/tcp_debug.h       |  16 +-
 src/vnet/tcp/tcp_input.c       | 671 +++++++++++++++++++++++++++++------------
 src/vnet/tcp/tcp_newreno.c     |  20 +-
 src/vnet/tcp/tcp_output.c      | 287 ++++++++++++------
 src/vnet/tcp/tcp_test.c        |  53 ++--
 17 files changed, 973 insertions(+), 409 deletions(-)

(limited to 'src/svm/svm_fifo.c')

diff --git a/src/svm/svm_fifo.c b/src/svm/svm_fifo.c
index f13f6fea..5c8f244a 100644
--- a/src/svm/svm_fifo.c
+++ b/src/svm/svm_fifo.c
@@ -540,7 +540,7 @@ svm_fifo_peek (svm_fifo_t * f, u32 relative_offset, u32 max_bytes,
 
   /* read cursize, which can only increase while we're working */
   cursize = svm_fifo_max_dequeue (f);
-  if (PREDICT_FALSE (cursize == 0))
+  if (PREDICT_FALSE (cursize < relative_offset))
     return -2;			/* nothing in the fifo */
 
   nitems = f->nitems;
@@ -548,7 +548,8 @@ svm_fifo_peek (svm_fifo_t * f, u32 relative_offset, u32 max_bytes,
   real_head = real_head >= nitems ? real_head - nitems : real_head;
 
   /* Number of bytes we're going to copy */
-  total_copy_bytes = (cursize < max_bytes) ? cursize : max_bytes;
+  total_copy_bytes = (cursize - relative_offset < max_bytes) ?
+    cursize - relative_offset : max_bytes;
 
   if (PREDICT_TRUE (copy_here != 0))
     {
diff --git a/src/vlib/node.c b/src/vlib/node.c
index bbd3a42e..eecad274 100644
--- a/src/vlib/node.c
+++ b/src/vlib/node.c
@@ -502,6 +502,7 @@ vlib_node_main_init (vlib_main_t * vm)
   vlib_node_t *n;
   uword ni;
 
+  nm->frame_size_hash = hash_create (0, sizeof (uword));
   nm->flags |= VLIB_NODE_MAIN_RUNTIME_STARTED;
 
   /* Generate sibling relationships */
diff --git a/src/vlib/threads.c b/src/vlib/threads.c
index b7bc9e26..0c775e2d 100644
--- a/src/vlib/threads.c
+++ b/src/vlib/threads.c
@@ -670,7 +670,7 @@ start_workers (vlib_main_t * vm)
 
 	      /* zap the (per worker) frame freelists, etc */
 	      nm_clone->frame_sizes = 0;
-	      nm_clone->frame_size_hash = 0;
+	      nm_clone->frame_size_hash = hash_create (0, sizeof (uword));
 
 	      /* Packet trace buffers are guaranteed to be empty, nothing to do here */
 
diff --git a/src/vlib/threads.h b/src/vlib/threads.h
index 17d35a24..572ce77f 100644
--- a/src/vlib/threads.h
+++ b/src/vlib/threads.h
@@ -62,7 +62,7 @@ typedef struct vlib_thread_registration_
 #define VLIB_CPU_MASK (VLIB_MAX_CPUS - 1)	/* 0x3f, max */
 #define VLIB_OFFSET_MASK (~VLIB_CPU_MASK)
 
-#define VLIB_LOG2_THREAD_STACK_SIZE (20)
+#define VLIB_LOG2_THREAD_STACK_SIZE (21)
 #define VLIB_THREAD_STACK_SIZE (1<<VLIB_LOG2_THREAD_STACK_SIZE)
 
 typedef enum
diff --git a/src/vnet/session/node.c b/src/vnet/session/node.c
index 3053ccc2..07eeae82 100644
--- a/src/vnet/session/node.c
+++ b/src/vnet/session/node.c
@@ -47,7 +47,8 @@ vlib_node_registration_t session_queue_node;
 
 #define foreach_session_queue_error		\
 _(TX, "Packets transmitted")                  	\
-_(TIMER, "Timer events")
+_(TIMER, "Timer events")			\
+_(NO_BUFFER, "Out of buffers")
 
 typedef enum
 {
@@ -141,6 +142,7 @@ session_tx_fifo_read_and_snd_i (vlib_main_t * vm, vlib_node_runtime_t * node,
   u8 *data0;
   int i, n_bytes_read;
   u32 n_bytes_per_buf, deq_per_buf;
+  u32 buffers_allocated, buffers_allocated_this_call;
 
   next_index = next0 = session_type_to_next[s0->session_type];
 
@@ -167,9 +169,6 @@ session_tx_fifo_read_and_snd_i (vlib_main_t * vm, vlib_node_runtime_t * node,
   /* Check how much we can pull. If buffering, subtract the offset */
   max_dequeue0 = svm_fifo_max_dequeue (s0->server_tx_fifo) - rx_offset;
 
-  /* Allow enqueuing of a new event */
-  svm_fifo_unset_event (s0->server_tx_fifo);
-
   /* Nothing to read return */
   if (max_dequeue0 == 0)
     return 0;
@@ -187,8 +186,8 @@ session_tx_fifo_read_and_snd_i (vlib_main_t * vm, vlib_node_runtime_t * node,
       max_len_to_snd0 = snd_space0;
     }
 
-  n_bytes_per_buf = vlib_buffer_free_list_buffer_size (vm,
-						       VLIB_BUFFER_DEFAULT_FREE_LIST_INDEX);
+  n_bytes_per_buf = vlib_buffer_free_list_buffer_size
+    (vm, VLIB_BUFFER_DEFAULT_FREE_LIST_INDEX);
   n_bytes_per_seg = MAX_HDRS_LEN + snd_mss0;
   n_bufs_per_seg = ceil ((double) n_bytes_per_seg / n_bytes_per_buf);
   n_bufs_per_evt = (ceil ((double) max_len_to_snd0 / n_bytes_per_seg))
@@ -205,24 +204,33 @@ session_tx_fifo_read_and_snd_i (vlib_main_t * vm, vlib_node_runtime_t * node,
       if (PREDICT_FALSE (n_bufs < VLIB_FRAME_SIZE))
 	{
 	  vec_validate (smm->tx_buffers[thread_index],
-			n_bufs + VLIB_FRAME_SIZE - 1);
-	  n_bufs += vlib_buffer_alloc (vm,
-				       &smm->tx_buffers[thread_index][n_bufs],
-				       VLIB_FRAME_SIZE);
-
-	  /* buffer shortage
-	   * XXX 0.9 because when debugging we might not get a full frame */
-	  if (PREDICT_FALSE (n_bufs < 0.9 * VLIB_FRAME_SIZE))
+			n_bufs + 2 * VLIB_FRAME_SIZE - 1);
+
+	  buffers_allocated = 0;
+	  do
 	    {
-	      if (svm_fifo_set_event (s0->server_tx_fifo))
-		{
-		  vec_add1 (smm->pending_event_vector[thread_index], *e0);
-		}
-	      return -1;
+	      buffers_allocated_this_call =
+		vlib_buffer_alloc
+		(vm,
+		 &smm->tx_buffers[thread_index][n_bufs + buffers_allocated],
+		 2 * VLIB_FRAME_SIZE - buffers_allocated);
+	      buffers_allocated += buffers_allocated_this_call;
 	    }
+	  while (buffers_allocated_this_call > 0
+		 && ((buffers_allocated + n_bufs < VLIB_FRAME_SIZE)));
+
+	  n_bufs += buffers_allocated;
 
 	  _vec_len (smm->tx_buffers[thread_index]) = n_bufs;
+
+	  if (PREDICT_FALSE (n_bufs < VLIB_FRAME_SIZE))
+	    {
+	      vec_add1 (smm->pending_event_vector[thread_index], *e0);
+	      return -1;
+	    }
 	}
+      /* Allow enqueuing of a new event */
+      svm_fifo_unset_event (s0->server_tx_fifo);
 
       vlib_get_next_frame (vm, node, next_index, to_next, n_left_to_next);
       while (left_to_snd0 && n_left_to_next >= n_bufs_per_seg)
@@ -232,7 +240,9 @@ session_tx_fifo_read_and_snd_i (vlib_main_t * vm, vlib_node_runtime_t * node,
 	   */
 
 	  /* Get free buffer */
+	  ASSERT (n_bufs >= 1);
 	  bi0 = smm->tx_buffers[thread_index][--n_bufs];
+	  ASSERT (bi0);
 	  _vec_len (smm->tx_buffers[thread_index]) = n_bufs;
 
 	  b0 = vlib_get_buffer (vm, bi0);
@@ -545,9 +555,10 @@ skip_dequeue:
 							my_thread_index,
 							&n_tx_packets);
 	  /* Out of buffers */
-	  if (rv < 0)
+	  if (PREDICT_FALSE (rv < 0))
 	    {
-	      vec_add1 (smm->pending_event_vector[my_thread_index], *e0);
+	      vlib_node_increment_counter (vm, node->node_index,
+					   SESSION_QUEUE_ERROR_NO_BUFFER, 1);
 	      continue;
 	    }
 	  break;
diff --git a/src/vnet/session/session.c b/src/vnet/session/session.c
index 02b0cced..534598d6 100644
--- a/src/vnet/session/session.c
+++ b/src/vnet/session/session.c
@@ -551,7 +551,7 @@ u8
 stream_session_no_space (transport_connection_t * tc, u32 thread_index,
 			 u16 data_len)
 {
-  stream_session_t *s = stream_session_get (tc->c_index, thread_index);
+  stream_session_t *s = stream_session_get (tc->s_index, thread_index);
 
   if (PREDICT_FALSE (s->session_state != SESSION_STATE_READY))
     return 1;
@@ -563,6 +563,15 @@ stream_session_no_space (transport_connection_t * tc, u32 thread_index,
 }
 
 u32
+stream_session_tx_fifo_max_dequeue (transport_connection_t * tc)
+{
+  stream_session_t *s = stream_session_get (tc->s_index, tc->thread_index);
+  if (s->session_state != SESSION_STATE_READY)
+    return 0;
+  return svm_fifo_max_dequeue (s->server_tx_fifo);
+}
+
+int
 stream_session_peek_bytes (transport_connection_t * tc, u8 * buffer,
 			   u32 offset, u32 max_bytes)
 {
diff --git a/src/vnet/session/session.h b/src/vnet/session/session.h
index a8728649..d9c38bd1 100644
--- a/src/vnet/session/session.h
+++ b/src/vnet/session/session.h
@@ -352,16 +352,18 @@ stream_session_max_rx_enqueue (transport_connection_t * tc)
 }
 
 always_inline u32
-stream_session_fifo_size (transport_connection_t * tc)
+stream_session_rx_fifo_size (transport_connection_t * tc)
 {
   stream_session_t *s = stream_session_get (tc->s_index, tc->thread_index);
   return s->server_rx_fifo->nitems;
 }
 
+u32 stream_session_tx_fifo_max_dequeue (transport_connection_t * tc);
+
 int
 stream_session_enqueue_data (transport_connection_t * tc, vlib_buffer_t * b,
 			     u32 offset, u8 queue_event, u8 is_in_order);
-u32
+int
 stream_session_peek_bytes (transport_connection_t * tc, u8 * buffer,
 			   u32 offset, u32 max_bytes);
 u32 stream_session_dequeue_drop (transport_connection_t * tc, u32 max_bytes);
diff --git a/src/vnet/session/session_cli.c b/src/vnet/session/session_cli.c
index 509eedbb..6b8341aa 100755
--- a/src/vnet/session/session_cli.c
+++ b/src/vnet/session/session_cli.c
@@ -15,6 +15,15 @@
 #include <vnet/session/application.h>
 #include <vnet/session/session.h>
 
+u8 *
+format_stream_session_fifos (u8 * s, va_list * args)
+{
+  stream_session_t *ss = va_arg (*args, stream_session_t *);
+  s = format (s, " Rx fifo: %U", format_svm_fifo, ss->server_rx_fifo, 1);
+  s = format (s, " Tx fifo: %U", format_svm_fifo, ss->server_tx_fifo, 1);
+  return s;
+}
+
 /**
  * Format stream session as per the following format
  *
@@ -44,6 +53,8 @@ format_stream_session (u8 * s, va_list * args)
 		  ss->thread_index, verbose);
       if (verbose == 1)
 	s = format (s, "%v", str);
+      if (verbose > 1)
+	s = format (s, "%U", format_stream_session_fifos, ss);
     }
   else if (ss->session_state == SESSION_STATE_LISTENING)
     {
@@ -57,8 +68,12 @@ format_stream_session (u8 * s, va_list * args)
     }
   else if (ss->session_state == SESSION_STATE_CLOSED)
     {
-      s = format (s, "[CL] %-40U%v", tp_vft->format_connection,
-		  ss->connection_index, ss->thread_index, verbose, str);
+      s = format (s, "[CL] %-40U", tp_vft->format_connection,
+		  ss->connection_index, ss->thread_index, verbose);
+      if (verbose == 1)
+	s = format (s, "%v", str);
+      if (verbose > 1)
+	s = format (s, "%U", format_stream_session_fifos, ss);
     }
   else
     {
@@ -124,13 +139,6 @@ show_session_command_fn (vlib_main_t * vm, unformat_input_t * input,
               ({
         	vec_reset_length (str);
                 str = format (str, "%U", format_stream_session, s, verbose);
-                if (verbose > 1)
-                  {
-                    str = format (str, " Rx fifo: %U", format_svm_fifo,
-				  s->server_rx_fifo, 1);
-                    str = format (str, " Tx fifo: %U", format_svm_fifo,
-				  s->server_tx_fifo, 1);
-                  }
                 vlib_cli_output (vm, "%v", str);
               }));
               /* *INDENT-ON* */
diff --git a/src/vnet/tcp/builtin_client.c b/src/vnet/tcp/builtin_client.c
index 768f0c3c..7238cda3 100644
--- a/src/vnet/tcp/builtin_client.c
+++ b/src/vnet/tcp/builtin_client.c
@@ -115,8 +115,17 @@ receive_test_chunk (tclient_main_t * tm, session_t * s)
   /* Allow enqueuing of new event */
   // svm_fifo_unset_event (rx_fifo);
 
-  n_read = svm_fifo_dequeue_nowait (rx_fifo, vec_len (tm->rx_buf),
-				    tm->rx_buf);
+  if (test_bytes)
+    {
+      n_read = svm_fifo_dequeue_nowait (rx_fifo, vec_len (tm->rx_buf),
+					tm->rx_buf);
+    }
+  else
+    {
+      n_read = svm_fifo_max_dequeue (rx_fifo);
+      svm_fifo_dequeue_drop (rx_fifo, n_read);
+    }
+
   if (n_read > 0)
     {
       if (TCP_BUILTIN_CLIENT_DBG)
@@ -165,6 +174,8 @@ builtin_client_node_fn (vlib_main_t * vm, vlib_node_runtime_t * node,
   int i;
   int delete_session;
   u32 *connection_indices;
+  u32 tx_quota = 0;
+  u32 delta, prev_bytes_received_this_session;
 
   connection_indices = tm->connection_index_by_thread[my_thread_index];
 
@@ -177,14 +188,19 @@ builtin_client_node_fn (vlib_main_t * vm, vlib_node_runtime_t * node,
 
       sp = pool_elt_at_index (tm->sessions, connection_indices[i]);
 
-      if (sp->bytes_to_send > 0)
+      if (tx_quota < 60 && sp->bytes_to_send > 0)
 	{
 	  send_test_chunk (tm, sp);
 	  delete_session = 0;
+	  tx_quota++;
 	}
       if (sp->bytes_to_receive > 0)
 	{
+	  prev_bytes_received_this_session = sp->bytes_received;
 	  receive_test_chunk (tm, sp);
+	  delta = sp->bytes_received - prev_bytes_received_this_session;
+	  if (delta > 0)
+	    tx_quota--;
 	  delete_session = 0;
 	}
       if (PREDICT_FALSE (delete_session == 1))
@@ -195,11 +211,19 @@ builtin_client_node_fn (vlib_main_t * vm, vlib_node_runtime_t * node,
 	  dmp->_vl_msg_id = ntohs (VL_API_DISCONNECT_SESSION);
 	  dmp->client_index = tm->my_client_index;
 	  dmp->handle = sp->vpp_session_handle;
-	  vl_msg_api_send_shmem (tm->vl_input_queue, (u8 *) & dmp);
-	  vec_delete (connection_indices, 1, i);
-	  tm->connection_index_by_thread[my_thread_index] =
-	    connection_indices;
-	  __sync_fetch_and_add (&tm->ready_connections, -1);
+//        vl_msg_api_send_shmem (tm->vl_input_queue, (u8 *) & dmp);
+	  if (!unix_shared_memory_queue_add (tm->vl_input_queue, (u8 *) & dmp,
+					     1))
+	    {
+	      vec_delete (connection_indices, 1, i);
+	      tm->connection_index_by_thread[my_thread_index] =
+		connection_indices;
+	      __sync_fetch_and_add (&tm->ready_connections, -1);
+	    }
+	  else
+	    {
+	      vl_msg_api_free (dmp);
+	    }
 
 	  /* Kick the debug CLI process */
 	  if (tm->ready_connections == 0)
diff --git a/src/vnet/tcp/builtin_server.c b/src/vnet/tcp/builtin_server.c
index 4f0e211c..8bd2f360 100644
--- a/src/vnet/tcp/builtin_server.c
+++ b/src/vnet/tcp/builtin_server.c
@@ -39,7 +39,8 @@
 
 typedef struct
 {
-  u8 *rx_buf;
+  /* Per-thread RX buffer */
+  u8 **rx_buf;
   unix_shared_memory_queue_t **vpp_queue;
   u64 byte_index;
 
@@ -117,13 +118,15 @@ void
 test_bytes (builtin_server_main_t * bsm, int actual_transfer)
 {
   int i;
+  u32 my_thread_id = vlib_get_thread_index ();
 
   for (i = 0; i < actual_transfer; i++)
     {
-      if (bsm->rx_buf[i] != ((bsm->byte_index + i) & 0xff))
+      if (bsm->rx_buf[my_thread_id][i] != ((bsm->byte_index + i) & 0xff))
 	{
 	  clib_warning ("at %lld expected %d got %d", bsm->byte_index + i,
-			(bsm->byte_index + i) & 0xff, bsm->rx_buf[i]);
+			(bsm->byte_index + i) & 0xff,
+			bsm->rx_buf[my_thread_id][i]);
 	}
     }
   bsm->byte_index += actual_transfer;
@@ -138,6 +141,7 @@ builtin_server_rx_callback (stream_session_t * s)
   builtin_server_main_t *bsm = &builtin_server_main;
   session_fifo_event_t evt;
   static int serial_number = 0;
+  u32 my_thread_id = vlib_get_thread_index ();
 
   tx_fifo = s->server_tx_fifo;
   rx_fifo = s->server_rx_fifo;
@@ -171,11 +175,12 @@ builtin_server_rx_callback (stream_session_t * s)
       return 0;
     }
 
-  vec_validate (bsm->rx_buf, max_transfer - 1);
-  _vec_len (bsm->rx_buf) = max_transfer;
+  vec_validate (bsm->rx_buf, my_thread_id);
+  vec_validate (bsm->rx_buf[my_thread_id], max_transfer - 1);
+  _vec_len (bsm->rx_buf[my_thread_id]) = max_transfer;
 
   actual_transfer = svm_fifo_dequeue_nowait (rx_fifo, max_transfer,
-					     bsm->rx_buf);
+					     bsm->rx_buf[my_thread_id]);
   ASSERT (actual_transfer == max_transfer);
 
 //  test_bytes (bsm, actual_transfer);
@@ -184,7 +189,8 @@ builtin_server_rx_callback (stream_session_t * s)
    * Echo back
    */
 
-  n_written = svm_fifo_enqueue_nowait (tx_fifo, actual_transfer, bsm->rx_buf);
+  n_written = svm_fifo_enqueue_nowait (tx_fifo, actual_transfer,
+				       bsm->rx_buf[my_thread_id]);
 
   if (n_written != max_transfer)
     clib_warning ("short trout!");
diff --git a/src/vnet/tcp/tcp.c b/src/vnet/tcp/tcp.c
index 9b7b2f65..e0b67a8e 100644
--- a/src/vnet/tcp/tcp.c
+++ b/src/vnet/tcp/tcp.c
@@ -195,8 +195,8 @@ tcp_connection_close (tcp_connection_t * tc)
   TCP_EVT_DBG (TCP_EVT_CLOSE, tc);
 
   /* Send FIN if needed */
-  if (tc->state == TCP_STATE_ESTABLISHED || tc->state == TCP_STATE_SYN_RCVD
-      || tc->state == TCP_STATE_CLOSE_WAIT)
+  if (tc->state == TCP_STATE_ESTABLISHED
+      || tc->state == TCP_STATE_SYN_RCVD || tc->state == TCP_STATE_CLOSE_WAIT)
     tcp_send_fin (tc);
 
   /* Switch state */
@@ -480,7 +480,7 @@ u8 *
 format_tcp_timers (u8 * s, va_list * args)
 {
   tcp_connection_t *tc = va_arg (*args, tcp_connection_t *);
-  int i, last = 0;
+  int i, last = -1;
 
   for (i = 0; i < TCP_N_TIMERS; i++)
     if (tc->timers[i] != TCP_TIMER_HANDLE_INVALID)
@@ -493,7 +493,7 @@ format_tcp_timers (u8 * s, va_list * args)
 	s = format (s, "%s,", tcp_conn_timers[i]);
     }
 
-  if (last > 0)
+  if (last >= 0)
     s = format (s, "%s]", tcp_conn_timers[i]);
   else
     s = format (s, "]");
@@ -526,19 +526,19 @@ format_tcp_vars (u8 * s, va_list * args)
   s = format (s, " snd_wnd %u rcv_wnd %u snd_wl1 %u snd_wl2 %u\n",
 	      tc->snd_wnd, tc->rcv_wnd, tc->snd_wl1 - tc->irs,
 	      tc->snd_wl2 - tc->iss);
-  s = format (s, " flight size %u send space %u rcv_wnd available %d\n",
-	      tcp_flight_size (tc), tcp_snd_space (tc),
-	      tc->rcv_wnd - (tc->rcv_nxt - tc->rcv_las));
+  s = format (s, " flight size %u send space %u rcv_wnd_av %d\n",
+	      tcp_flight_size (tc), tcp_available_snd_space (tc),
+	      tcp_rcv_wnd_available (tc));
   s = format (s, " cong %U ", format_tcp_congestion_status, tc);
   s = format (s, "cwnd %u ssthresh %u rtx_bytes %u bytes_acked %u\n",
-	      tc->cwnd, tc->ssthresh, tc->rtx_bytes, tc->bytes_acked);
-  s = format (s, " prev_ssthresh %u snd_congestion %u\n", tc->prev_ssthresh,
-	      tc->snd_congestion - tc->iss);
+	      tc->cwnd, tc->ssthresh, tc->snd_rxt_bytes, tc->bytes_acked);
+  s = format (s, " prev_ssthresh %u snd_congestion %u dupack %u\n",
+	      tc->prev_ssthresh, tc->snd_congestion - tc->iss,
+	      tc->rcv_dupacks);
   s = format (s, " rto %u rto_boff %u srtt %u rttvar %u rtt_ts %u ", tc->rto,
 	      tc->rto_boff, tc->srtt, tc->rttvar, tc->rtt_ts);
   s = format (s, "rtt_seq %u\n", tc->rtt_seq);
-  if (scoreboard_first_hole (&tc->sack_sb))
-    s = format (s, " scoreboard: %U\n", format_tcp_scoreboard, &tc->sack_sb);
+  s = format (s, " scoreboard: %U\n", format_tcp_scoreboard, &tc->sack_sb);
   if (vec_len (tc->snd_sacks))
     s = format (s, " sacks tx: %U\n", format_tcp_sacks, tc);
 
@@ -595,9 +595,10 @@ format_tcp_session (u8 * s, va_list * args)
 
   tc = tcp_connection_get (tci, thread_index);
   if (tc)
-    return format (s, "%U", format_tcp_connection, tc, verbose);
+    s = format (s, "%U", format_tcp_connection, tc, verbose);
   else
-    return format (s, "empty");
+    s = format (s, "empty");
+  return s;
 }
 
 u8 *
@@ -643,13 +644,17 @@ format_tcp_scoreboard (u8 * s, va_list * args)
 {
   sack_scoreboard_t *sb = va_arg (*args, sack_scoreboard_t *);
   sack_scoreboard_hole_t *hole;
-  s = format (s, "head %u tail %u snd_una_adv %u\n", sb->head, sb->tail,
-	      sb->snd_una_adv);
-  s = format (s, "sacked_bytes %u last_sacked_bytes %u", sb->sacked_bytes,
-	      sb->last_sacked_bytes);
-  s = format (s, " max_byte_sacked %u\n", sb->max_byte_sacked);
-  s = format (s, "holes:\n");
+  s = format (s, "sacked_bytes %u last_sacked_bytes %u lost_bytes %u\n",
+	      sb->sacked_bytes, sb->last_sacked_bytes, sb->lost_bytes);
+  s = format (s, " last_bytes_delivered %u high_sacked %u snd_una_adv %u\n",
+	      sb->last_bytes_delivered, sb->high_sacked, sb->snd_una_adv);
+  s = format (s, " cur_rxt_hole %u high_rxt %u rescue_rxt %u",
+	      sb->cur_rxt_hole, sb->high_rxt, sb->rescue_rxt);
+
   hole = scoreboard_first_hole (sb);
+  if (hole)
+    s = format (s, "\n head %u tail %u holes:\n", sb->head, sb->tail);
+
   while (hole)
     {
       s = format (s, "%U", format_tcp_sack_hole, hole);
@@ -736,7 +741,7 @@ tcp_snd_space (tcp_connection_t * tc)
   if (tcp_in_recovery (tc))
     {
       tc->snd_nxt = tc->snd_una_max;
-      snd_space = tcp_available_wnd (tc) - tc->rtx_bytes
+      snd_space = tcp_available_wnd (tc) - tc->snd_rxt_bytes
 	- (tc->snd_una_max - tc->snd_congestion);
       if (snd_space <= 0 || (tc->snd_una_max - tc->snd_una) >= tc->snd_wnd)
 	return 0;
@@ -744,8 +749,8 @@ tcp_snd_space (tcp_connection_t * tc)
     }
 
   /* If in fast recovery, send 1 SMSS if wnd allows */
-  if (tcp_in_fastrecovery (tc) && tcp_available_snd_space (tc)
-      && tcp_fastrecovery_sent_1_smss (tc))
+  if (tcp_in_fastrecovery (tc)
+      && tcp_available_snd_space (tc) && !tcp_fastrecovery_sent_1_smss (tc))
     {
       tcp_fastrecovery_1_smss_on (tc);
       return tc->snd_mss;
@@ -761,6 +766,12 @@ tcp_session_send_space (transport_connection_t * trans_conn)
   return tcp_snd_space (tc);
 }
 
+i32
+tcp_rcv_wnd_available (tcp_connection_t * tc)
+{
+  return (i32) tc->rcv_wnd - (tc->rcv_nxt - tc->rcv_las);
+}
+
 u32
 tcp_session_tx_fifo_offset (transport_connection_t * trans_conn)
 {
diff --git a/src/vnet/tcp/tcp.h b/src/vnet/tcp/tcp.h
index c3ebe22b..071f1ab1 100644
--- a/src/vnet/tcp/tcp.h
+++ b/src/vnet/tcp/tcp.h
@@ -34,6 +34,7 @@
 #define TCP_MAX_RX_FIFO_SIZE 	2 << 20
 #define TCP_IW_N_SEGMENTS 	10
 #define TCP_ALWAYS_ACK		0	/**< If on, we always ack */
+#define TCP_USE_SACKS		1	/**< Disable only for testing */
 
 /** TCP FSM state definitions as per RFC793. */
 #define foreach_tcp_fsm_state   \
@@ -94,7 +95,7 @@ extern timer_expiration_handler tcp_timer_retransmit_syn_handler;
 #define TCP_DELACK_TIME         1	/* 0.1s */
 #define TCP_ESTABLISH_TIME      750	/* 75s */
 #define TCP_2MSL_TIME           300	/* 30s */
-#define TCP_CLOSEWAIT_TIME	1	/* 0.1s */
+#define TCP_CLOSEWAIT_TIME	20	/* 0.1s */
 #define TCP_CLEANUP_TIME	5	/* 0.5s Time to wait before cleanup */
 #define TCP_TIMER_PERSIST_MIN	2	/* 0.2s */
 
@@ -157,6 +158,7 @@ typedef struct _sack_scoreboard_hole
   u32 prev;		/**< Index for previous entry in linked list */
   u32 start;		/**< Start sequence number */
   u32 end;		/**< End sequence number */
+  u8 is_lost;		/**< Mark hole as lost */
 } sack_scoreboard_hole_t;
 
 typedef struct _sack_scoreboard
@@ -166,8 +168,13 @@ typedef struct _sack_scoreboard
   u32 tail;				/**< Index of last entry */
   u32 sacked_bytes;			/**< Number of bytes sacked in sb */
   u32 last_sacked_bytes;		/**< Number of bytes last sacked */
+  u32 last_bytes_delivered;		/**< Number of sack bytes delivered */
   u32 snd_una_adv;			/**< Bytes to add to snd_una */
-  u32 max_byte_sacked;			/**< Highest byte acked */
+  u32 high_sacked;			/**< Highest byte sacked (fack) */
+  u32 high_rxt;				/**< Highest retransmitted sequence */
+  u32 rescue_rxt;			/**< Rescue sequence number */
+  u32 lost_bytes;			/**< Bytes lost as per RFC6675 */
+  u32 cur_rxt_hole;			/**< Retransmitting from this hole */
 } sack_scoreboard_t;
 
 typedef enum _tcp_cc_algorithm_type
@@ -211,7 +218,7 @@ typedef struct _tcp_connection
   u32 irs;		/**< initial remote sequence */
 
   /* Options */
-  tcp_options_t opt;		/**< TCP connection options parsed */
+  tcp_options_t rcv_opts;	/**< Rx options for connection */
   tcp_options_t snd_opts;	/**< Tx options for connection */
   u8 snd_opts_len;		/**< Tx options len */
   u8 rcv_wscale;	/**< Window scale to advertise to peer */
@@ -229,8 +236,10 @@ typedef struct _tcp_connection
   u32 cwnd;		/**< Congestion window */
   u32 ssthresh;		/**< Slow-start threshold */
   u32 prev_ssthresh;	/**< ssthresh before congestion */
+  u32 prev_cwnd;	/**< ssthresh before congestion */
   u32 bytes_acked;	/**< Bytes acknowledged by current segment */
-  u32 rtx_bytes;	/**< Retransmitted bytes */
+  u32 snd_rxt_bytes;	/**< Retransmitted bytes */
+  u32 snd_rxt_ts;	/**< Timestamp when first packet is retransmitted */
   u32 tsecr_last_ack;	/**< Timestamp echoed to us in last healthy ACK */
   u32 snd_congestion;	/**< snd_una_max when congestion is detected */
   tcp_cc_algorithm_t *cc_algo;	/**< Congestion control algorithm */
@@ -411,6 +420,7 @@ void tcp_send_syn (tcp_connection_t * tc);
 void tcp_send_fin (tcp_connection_t * tc);
 void tcp_init_mss (tcp_connection_t * tc);
 void tcp_update_snd_mss (tcp_connection_t * tc);
+void tcp_update_rto (tcp_connection_t * tc);
 
 always_inline u32
 tcp_end_seq (tcp_header_t * th, u32 len)
@@ -428,17 +438,39 @@ tcp_end_seq (tcp_header_t * th, u32 len)
 #define timestamp_lt(_t1, _t2) ((i32)((_t1)-(_t2)) < 0)
 #define timestamp_leq(_t1, _t2) ((i32)((_t1)-(_t2)) <= 0)
 
+/**
+ * Our estimate of the number of bytes that have left the network
+ */
+always_inline u32
+tcp_bytes_out (const tcp_connection_t * tc)
+{
+  if (tcp_opts_sack_permitted (&tc->rcv_opts))
+    return tc->sack_sb.sacked_bytes + tc->sack_sb.lost_bytes;
+  else
+    return tc->rcv_dupacks * tc->snd_mss;
+}
+
+/**
+ * Our estimate of the number of bytes in flight (pipe size)
+ */
 always_inline u32
 tcp_flight_size (const tcp_connection_t * tc)
 {
   int flight_size;
 
-  flight_size = (int) ((tc->snd_una_max - tc->snd_una) + tc->rtx_bytes)
-    - (tc->rcv_dupacks * tc->snd_mss) /* - tc->sack_sb.sacked_bytes */ ;
+  flight_size = (int) (tc->snd_una_max - tc->snd_una) - tcp_bytes_out (tc)
+    + tc->snd_rxt_bytes;
 
-  /* Happens if we don't clear sacked bytes */
   if (flight_size < 0)
-    return 0;
+    {
+      if (0)
+	clib_warning
+	  ("Negative: %u %u %u dupacks %u sacked bytes %u flags %d",
+	   tc->snd_una_max - tc->snd_una, tcp_bytes_out (tc),
+	   tc->snd_rxt_bytes, tc->rcv_dupacks, tc->sack_sb.sacked_bytes,
+	   tc->rcv_opts.flags);
+      return 0;
+    }
 
   return flight_size;
 }
@@ -481,14 +513,17 @@ tcp_available_snd_space (const tcp_connection_t * tc)
   return available_wnd - flight_size;
 }
 
-u32 tcp_rcv_wnd_available (tcp_connection_t * tc);
+i32 tcp_rcv_wnd_available (tcp_connection_t * tc);
 u32 tcp_snd_space (tcp_connection_t * tc);
 void tcp_update_rcv_wnd (tcp_connection_t * tc);
 
 void tcp_retransmit_first_unacked (tcp_connection_t * tc);
+void tcp_fast_retransmit_no_sack (tcp_connection_t * tc);
+void tcp_fast_retransmit_sack (tcp_connection_t * tc);
 void tcp_fast_retransmit (tcp_connection_t * tc);
-void tcp_cc_congestion (tcp_connection_t * tc);
-void tcp_cc_recover (tcp_connection_t * tc);
+void tcp_cc_init_congestion (tcp_connection_t * tc);
+int tcp_cc_recover (tcp_connection_t * tc);
+void tcp_cc_fastrecovery_exit (tcp_connection_t * tc);
 
 /* Made public for unit testing only */
 void tcp_update_sack_list (tcp_connection_t * tc, u32 start, u32 end);
@@ -563,16 +598,16 @@ tcp_retransmit_timer_set (tcp_connection_t * tc)
 }
 
 always_inline void
-tcp_retransmit_timer_update (tcp_connection_t * tc)
+tcp_retransmit_timer_reset (tcp_connection_t * tc)
 {
-  tcp_timer_update (tc, TCP_TIMER_RETRANSMIT,
-		    clib_max (tc->rto * TCP_TO_TIMER_TICK, 1));
+  tcp_timer_reset (tc, TCP_TIMER_RETRANSMIT);
 }
 
 always_inline void
-tcp_retransmit_timer_reset (tcp_connection_t * tc)
+tcp_retransmit_timer_force_update (tcp_connection_t * tc)
 {
-  tcp_timer_reset (tc, TCP_TIMER_RETRANSMIT);
+  tcp_timer_update (tc, TCP_TIMER_RETRANSMIT,
+		    clib_max (tc->rto * TCP_TO_TIMER_TICK, 1));
 }
 
 always_inline void
@@ -598,15 +633,43 @@ tcp_persist_timer_reset (tcp_connection_t * tc)
   tcp_timer_reset (tc, TCP_TIMER_PERSIST);
 }
 
+always_inline void
+tcp_retransmit_timer_update (tcp_connection_t * tc)
+{
+  if (tc->snd_una == tc->snd_una_max)
+    {
+      tcp_retransmit_timer_reset (tc);
+      if (tc->snd_wnd < tc->snd_mss)
+	tcp_persist_timer_set (tc);
+    }
+  else
+    tcp_timer_update (tc, TCP_TIMER_RETRANSMIT,
+		      clib_max (tc->rto * TCP_TO_TIMER_TICK, 1));
+}
+
 always_inline u8
 tcp_timer_is_active (tcp_connection_t * tc, tcp_timers_e timer)
 {
   return tc->timers[timer] != TCP_TIMER_HANDLE_INVALID;
 }
 
+#define tcp_validate_txf_size(_tc, _a) 					\
+  ASSERT(_tc->state != TCP_STATE_ESTABLISHED 				\
+	 || stream_session_tx_fifo_max_dequeue (&_tc->connection) >= _a)
+
 void
 scoreboard_remove_hole (sack_scoreboard_t * sb,
 			sack_scoreboard_hole_t * hole);
+void scoreboard_update_lost (tcp_connection_t * tc, sack_scoreboard_t * sb);
+sack_scoreboard_hole_t *scoreboard_insert_hole (sack_scoreboard_t * sb,
+						u32 prev_index, u32 start,
+						u32 end);
+sack_scoreboard_hole_t *scoreboard_next_rxt_hole (sack_scoreboard_t * sb,
+						  sack_scoreboard_hole_t *
+						  start, u8 have_sent_1_smss,
+						  u8 * can_rescue,
+						  u8 * snd_limited);
+void scoreboard_init_high_rxt (sack_scoreboard_t * sb);
 
 always_inline sack_scoreboard_hole_t *
 scoreboard_get_hole (sack_scoreboard_t * sb, u32 index)
@@ -624,6 +687,14 @@ scoreboard_next_hole (sack_scoreboard_t * sb, sack_scoreboard_hole_t * hole)
   return 0;
 }
 
+always_inline sack_scoreboard_hole_t *
+scoreboard_prev_hole (sack_scoreboard_t * sb, sack_scoreboard_hole_t * hole)
+{
+  if (hole->prev != TCP_INVALID_SACK_HOLE_INDEX)
+    return pool_elt_at_index (sb->holes, hole->prev);
+  return 0;
+}
+
 always_inline sack_scoreboard_hole_t *
 scoreboard_first_hole (sack_scoreboard_t * sb)
 {
@@ -643,15 +714,19 @@ scoreboard_last_hole (sack_scoreboard_t * sb)
 always_inline void
 scoreboard_clear (sack_scoreboard_t * sb)
 {
-  sack_scoreboard_hole_t *hole = scoreboard_first_hole (sb);
+  sack_scoreboard_hole_t *hole;
   while ((hole = scoreboard_first_hole (sb)))
     {
       scoreboard_remove_hole (sb, hole);
     }
   sb->sacked_bytes = 0;
   sb->last_sacked_bytes = 0;
+  sb->last_bytes_delivered = 0;
   sb->snd_una_adv = 0;
-  sb->max_byte_sacked = 0;
+  sb->high_sacked = 0;
+  sb->high_rxt = 0;
+  sb->lost_bytes = 0;
+  sb->cur_rxt_hole = TCP_INVALID_SACK_HOLE_INDEX;
 }
 
 always_inline u32
@@ -671,6 +746,7 @@ scoreboard_init (sack_scoreboard_t * sb)
 {
   sb->head = TCP_INVALID_SACK_HOLE_INDEX;
   sb->tail = TCP_INVALID_SACK_HOLE_INDEX;
+  sb->cur_rxt_hole = TCP_INVALID_SACK_HOLE_INDEX;
 }
 
 void tcp_rcv_sacks (tcp_connection_t * tc, u32 ack);
diff --git a/src/vnet/tcp/tcp_debug.h b/src/vnet/tcp/tcp_debug.h
index b4497a3b..3a16cf63 100755
--- a/src/vnet/tcp/tcp_debug.h
+++ b/src/vnet/tcp/tcp_debug.h
@@ -393,7 +393,7 @@ typedef enum _tcp_dbg_evt
   DECLARE_ETD(_tc, _e, 4);						\
   ed->data[0] = _seq - _tc->irs;					\
   ed->data[1] = _end - _tc->irs;					\
-  ed->data[2] = _tc->opt.tsval;						\
+  ed->data[2] = _tc->rcv_opts.tsval;					\
   ed->data[3] = _tc->tsval_recent;					\
 }
 
@@ -427,27 +427,27 @@ typedef enum _tcp_dbg_evt
 {									\
   ELOG_TYPE_DECLARE (_e) =						\
   {									\
-    .format = "rtx: snd_nxt %u offset %u snd %u rtx %u",		\
+    .format = "rxt: snd_nxt %u offset %u snd %u rxt %u",		\
     .format_args = "i4i4i4i4",						\
   };									\
   DECLARE_ETD(_tc, _e, 4);						\
   ed->data[0] = _tc->snd_nxt - _tc->iss;				\
   ed->data[1] = offset;							\
   ed->data[2] = n_bytes;						\
-  ed->data[3] = _tc->rtx_bytes;						\
+  ed->data[3] = _tc->snd_rxt_bytes;					\
 }
 
 #define TCP_EVT_CC_EVT_HANDLER(_tc, _sub_evt, ...)			\
 {									\
   ELOG_TYPE_DECLARE (_e) =						\
   {									\
-    .format = "cc: %s wnd %u snd_cong %u rtx_bytes %u",			\
+    .format = "cc: %s wnd %u snd_cong %u rxt_bytes %u",			\
     .format_args = "t4i4i4i4",						\
     .n_enum_strings = 5,						\
     .enum_strings = {                                           	\
-      "fast-rtx",	                                             	\
-      "rtx-timeout",                                                 	\
-      "first-rtx",                                                 	\
+      "fast-rxt",	                                             	\
+      "rxt-timeout",                                                 	\
+      "first-rxt",                                                 	\
       "recovered",							\
       "congestion",							\
     },  								\
@@ -456,7 +456,7 @@ typedef enum _tcp_dbg_evt
   ed->data[0] = _sub_evt;						\
   ed->data[1] = tcp_available_snd_space (_tc);				\
   ed->data[2] = _tc->snd_congestion - _tc->iss;				\
-  ed->data[3] = _tc->rtx_bytes;						\
+  ed->data[3] = _tc->snd_rxt_bytes;					\
 }
 
 #define TCP_EVT_CC_PACK_HANDLER(_tc, ...)				\
diff --git a/src/vnet/tcp/tcp_input.c b/src/vnet/tcp/tcp_input.c
index 35bc9094..ff2229b3 100644
--- a/src/vnet/tcp/tcp_input.c
+++ b/src/vnet/tcp/tcp_input.c
@@ -231,8 +231,8 @@ tcp_options_parse (tcp_header_t * th, tcp_options_t * to)
 always_inline int
 tcp_segment_check_paws (tcp_connection_t * tc)
 {
-  return tcp_opts_tstamp (&tc->opt) && tc->tsval_recent
-    && timestamp_lt (tc->opt.tsval, tc->tsval_recent);
+  return tcp_opts_tstamp (&tc->rcv_opts) && tc->tsval_recent
+    && timestamp_lt (tc->rcv_opts.tsval, tc->tsval_recent);
 }
 
 /**
@@ -248,10 +248,10 @@ tcp_update_timestamp (tcp_connection_t * tc, u32 seq, u32 seq_end)
    * then the TSval from the segment is copied to TS.Recent;
    * otherwise, the TSval is ignored.
    */
-  if (tcp_opts_tstamp (&tc->opt) && tc->tsval_recent
+  if (tcp_opts_tstamp (&tc->rcv_opts) && tc->tsval_recent
       && seq_leq (seq, tc->rcv_las) && seq_leq (tc->rcv_las, seq_end))
     {
-      tc->tsval_recent = tc->opt.tsval;
+      tc->tsval_recent = tc->rcv_opts.tsval;
       tc->tsval_recent_age = tcp_time_now ();
     }
 }
@@ -272,14 +272,21 @@ tcp_segment_validate (vlib_main_t * vm, tcp_connection_t * tc0,
   if (PREDICT_FALSE (!tcp_ack (th0) && !tcp_rst (th0) && !tcp_syn (th0)))
     return -1;
 
-  if (PREDICT_FALSE (tcp_options_parse (th0, &tc0->opt)))
+  if (PREDICT_FALSE (tcp_options_parse (th0, &tc0->rcv_opts)))
     {
       return -1;
     }
 
   if (tcp_segment_check_paws (tc0))
     {
-      clib_warning ("paws failed");
+      if (CLIB_DEBUG > 2)
+	{
+	  clib_warning ("paws failed\n%U", format_tcp_connection, tc0, 2);
+	  clib_warning ("seq %u seq_end %u ack %u",
+			vnet_buffer (b0)->tcp.seq_number - tc0->irs,
+			vnet_buffer (b0)->tcp.seq_end - tc0->irs,
+			vnet_buffer (b0)->tcp.ack_number - tc0->iss);
+	}
       TCP_EVT_DBG (TCP_EVT_PAWS_FAIL, tc0, vnet_buffer (b0)->tcp.seq_number,
 		   vnet_buffer (b0)->tcp.seq_end);
 
@@ -348,7 +355,6 @@ tcp_segment_validate (vlib_main_t * vm, tcp_connection_t * tc0,
   /* If segment in window, save timestamp */
   tcp_update_timestamp (tc0, vnet_buffer (b0)->tcp.seq_number,
 			vnet_buffer (b0)->tcp.seq_end);
-
   return 0;
 }
 
@@ -391,6 +397,12 @@ tcp_estimate_rtt (tcp_connection_t * tc, u32 mrtt)
     }
 }
 
+void
+tcp_update_rto (tcp_connection_t * tc)
+{
+  tc->rto = clib_min (tc->srtt + (tc->rttvar << 2), TCP_RTO_MAX);
+}
+
 /** Update RTT estimate and RTO timer
  *
  * Measure RTT: We have two sources of RTT measurements: TSOPT and ACK
@@ -405,7 +417,7 @@ tcp_update_rtt (tcp_connection_t * tc, u32 ack)
   u32 mrtt = 0;
   u8 rtx_acked;
 
-  /* Determine if only rtx bytes are acked. TODO fast retransmit */
+  /* Determine if only rtx bytes are acked. TODO XXX fast retransmit */
   rtx_acked = tc->rto_boff && (tc->bytes_acked <= tc->snd_mss);
 
   /* Karn's rule, part 1. Don't use retransmitted segments to estimate
@@ -418,9 +430,10 @@ tcp_update_rtt (tcp_connection_t * tc, u32 ack)
    * snd_una, i.e., the left side of the send window:
    * seq_lt (tc->snd_una, ack). Note: last condition could be dropped, we don't
    * try to update rtt for dupacks */
-  else if (tcp_opts_tstamp (&tc->opt) && tc->opt.tsecr && tc->bytes_acked)
+  else if (tcp_opts_tstamp (&tc->rcv_opts) && tc->rcv_opts.tsecr
+	   && tc->bytes_acked)
     {
-      mrtt = tcp_time_now () - tc->opt.tsecr;
+      mrtt = tcp_time_now () - tc->rcv_opts.tsecr;
     }
 
   /* Allow measuring of a new RTT */
@@ -436,7 +449,7 @@ tcp_update_rtt (tcp_connection_t * tc, u32 ack)
     return 0;
 
   tcp_estimate_rtt (tc, mrtt);
-  tc->rto = clib_min (tc->srtt + (tc->rttvar << 2), TCP_RTO_MAX);
+  tcp_update_rto (tc);
 
   return 0;
 }
@@ -447,25 +460,46 @@ tcp_update_rtt (tcp_connection_t * tc, u32 ack)
 static void
 tcp_dequeue_acked (tcp_connection_t * tc, u32 ack)
 {
-  /* Dequeue the newly ACKed bytes */
-  stream_session_dequeue_drop (&tc->connection, tc->bytes_acked);
+  /* Dequeue the newly ACKed add SACKed bytes */
+  stream_session_dequeue_drop (&tc->connection,
+			       tc->bytes_acked + tc->sack_sb.snd_una_adv);
+
+  tcp_validate_txf_size (tc, tc->snd_una_max - tc->snd_una);
 
   /* Update rtt and rto */
   tcp_update_rtt (tc, ack);
+
+  /* If everything has been acked, stop retransmit timer
+   * otherwise update. */
+  tcp_retransmit_timer_update (tc);
 }
 
 /**
- * Check if dupack as per RFC5681 Sec. 2
- *
- * This works only if called before updating snd_wnd.
- * */
-always_inline u8
-tcp_ack_is_dupack (tcp_connection_t * tc, vlib_buffer_t * b, u32 new_snd_wnd)
+ * Check if duplicate ack as per RFC5681 Sec. 2
+ */
+static u8
+tcp_ack_is_dupack (tcp_connection_t * tc, vlib_buffer_t * b, u32 prev_snd_wnd,
+		   u32 prev_snd_una)
 {
-  return ((vnet_buffer (b)->tcp.ack_number == tc->snd_una)
+  return ((vnet_buffer (b)->tcp.ack_number == prev_snd_una)
 	  && seq_gt (tc->snd_una_max, tc->snd_una)
 	  && (vnet_buffer (b)->tcp.seq_end == vnet_buffer (b)->tcp.seq_number)
-	  && (new_snd_wnd == tc->snd_wnd));
+	  && (prev_snd_wnd == tc->snd_wnd));
+}
+
+/**
+ * Checks if ack is a congestion control event.
+ */
+static u8
+tcp_ack_is_cc_event (tcp_connection_t * tc, vlib_buffer_t * b,
+		     u32 prev_snd_wnd, u32 prev_snd_una, u8 * is_dack)
+{
+  /* Check if ack is duplicate. Per RFC 6675, ACKs that SACK new data are
+   * defined to be 'duplicate' */
+  *is_dack = tc->sack_sb.last_sacked_bytes
+    || tcp_ack_is_dupack (tc, b, prev_snd_wnd, prev_snd_una);
+
+  return (*is_dack || tcp_in_cong_recovery (tc));
 }
 
 void
@@ -478,6 +512,10 @@ scoreboard_remove_hole (sack_scoreboard_t * sb, sack_scoreboard_hole_t * hole)
       next = pool_elt_at_index (sb->holes, hole->next);
       next->prev = hole->prev;
     }
+  else
+    {
+      sb->tail = hole->prev;
+    }
 
   if (hole->prev != TCP_INVALID_SACK_HOLE_INDEX)
     {
@@ -489,6 +527,9 @@ scoreboard_remove_hole (sack_scoreboard_t * sb, sack_scoreboard_hole_t * hole)
       sb->head = hole->next;
     }
 
+  if (scoreboard_hole_index (sb, hole) == sb->cur_rxt_hole)
+    sb->cur_rxt_hole = TCP_INVALID_SACK_HOLE_INDEX;
+
   pool_put (sb->holes, hole);
 }
 
@@ -527,26 +568,131 @@ scoreboard_insert_hole (sack_scoreboard_t * sb, u32 prev_index,
   return hole;
 }
 
+void
+scoreboard_update_lost (tcp_connection_t * tc, sack_scoreboard_t * sb)
+{
+  sack_scoreboard_hole_t *hole, *prev;
+  u32 bytes = 0, blks = 0;
+
+  sb->lost_bytes = 0;
+  hole = scoreboard_last_hole (sb);
+  if (!hole)
+    return;
+
+  if (seq_gt (sb->high_sacked, hole->end))
+    {
+      bytes = sb->high_sacked - hole->end;
+      blks = 1;
+    }
+
+  while ((prev = scoreboard_prev_hole (sb, hole))
+	 && (bytes < (TCP_DUPACK_THRESHOLD - 1) * tc->snd_mss
+	     && blks < TCP_DUPACK_THRESHOLD))
+    {
+      bytes += hole->start - prev->end;
+      blks++;
+      hole = prev;
+    }
+
+  hole = prev;
+  while (hole)
+    {
+      sb->lost_bytes += scoreboard_hole_bytes (hole);
+      hole->is_lost = 1;
+      hole = scoreboard_prev_hole (sb, hole);
+    }
+}
+
+/**
+ * Figure out the next hole to retransmit
+ *
+ * Follows logic proposed in RFC6675 Sec. 4, NextSeg()
+ */
+sack_scoreboard_hole_t *
+scoreboard_next_rxt_hole (sack_scoreboard_t * sb,
+			  sack_scoreboard_hole_t * start,
+			  u8 have_sent_1_smss,
+			  u8 * can_rescue, u8 * snd_limited)
+{
+  sack_scoreboard_hole_t *hole = 0;
+
+  hole = start ? start : scoreboard_first_hole (sb);
+  while (hole && seq_leq (hole->end, sb->high_rxt) && hole->is_lost)
+    hole = scoreboard_next_hole (sb, hole);
+
+  /* Nothing, return */
+  if (!hole)
+    {
+      sb->cur_rxt_hole = TCP_INVALID_SACK_HOLE_INDEX;
+      return 0;
+    }
+
+  /* Rule (1): if higher than rxt, less than high_sacked and lost */
+  if (hole->is_lost && seq_lt (hole->start, sb->high_sacked))
+    {
+      sb->cur_rxt_hole = scoreboard_hole_index (sb, hole);
+    }
+  else
+    {
+      /* Rule (2): output takes care of transmitting new data */
+      if (!have_sent_1_smss)
+	{
+	  hole = 0;
+	  sb->cur_rxt_hole = TCP_INVALID_SACK_HOLE_INDEX;
+	}
+      /* Rule (3): if hole not lost */
+      else if (seq_lt (hole->start, sb->high_sacked))
+	{
+	  *snd_limited = 1;
+	  sb->cur_rxt_hole = scoreboard_hole_index (sb, hole);
+	}
+      /* Rule (4): if hole beyond high_sacked */
+      else
+	{
+	  ASSERT (seq_geq (hole->start, sb->high_sacked));
+	  *snd_limited = 1;
+	  *can_rescue = 1;
+	  /* HighRxt MUST NOT be updated */
+	  return 0;
+	}
+    }
+
+  if (hole && seq_lt (sb->high_rxt, hole->start))
+    sb->high_rxt = hole->start;
+
+  return hole;
+}
+
+void
+scoreboard_init_high_rxt (sack_scoreboard_t * sb)
+{
+  sack_scoreboard_hole_t *hole;
+  hole = scoreboard_first_hole (sb);
+  sb->high_rxt = hole->start;
+  sb->cur_rxt_hole = sb->head;
+}
+
 void
 tcp_rcv_sacks (tcp_connection_t * tc, u32 ack)
 {
   sack_scoreboard_t *sb = &tc->sack_sb;
   sack_block_t *blk, tmp;
   sack_scoreboard_hole_t *hole, *next_hole, *last_hole, *new_hole;
-  u32 blk_index = 0, old_sacked_bytes, delivered_bytes, hole_index;
+  u32 blk_index = 0, old_sacked_bytes, hole_index;
   int i, j;
 
   sb->last_sacked_bytes = 0;
   sb->snd_una_adv = 0;
   old_sacked_bytes = sb->sacked_bytes;
-  delivered_bytes = 0;
+  sb->last_bytes_delivered = 0;
 
-  if (!tcp_opts_sack (&tc->opt) && sb->head == TCP_INVALID_SACK_HOLE_INDEX)
+  if (!tcp_opts_sack (&tc->rcv_opts)
+      && sb->head == TCP_INVALID_SACK_HOLE_INDEX)
     return;
 
   /* Remove invalid blocks */
-  blk = tc->opt.sacks;
-  while (blk < vec_end (tc->opt.sacks))
+  blk = tc->rcv_opts.sacks;
+  while (blk < vec_end (tc->rcv_opts.sacks))
     {
       if (seq_lt (blk->start, blk->end)
 	  && seq_gt (blk->start, tc->snd_una)
@@ -555,7 +701,7 @@ tcp_rcv_sacks (tcp_connection_t * tc, u32 ack)
 	  blk++;
 	  continue;
 	}
-      vec_del1 (tc->opt.sacks, blk - tc->opt.sacks);
+      vec_del1 (tc->rcv_opts.sacks, blk - tc->rcv_opts.sacks);
     }
 
   /* Add block for cumulative ack */
@@ -563,20 +709,20 @@ tcp_rcv_sacks (tcp_connection_t * tc, u32 ack)
     {
       tmp.start = tc->snd_una;
       tmp.end = ack;
-      vec_add1 (tc->opt.sacks, tmp);
+      vec_add1 (tc->rcv_opts.sacks, tmp);
     }
 
-  if (vec_len (tc->opt.sacks) == 0)
+  if (vec_len (tc->rcv_opts.sacks) == 0)
     return;
 
   /* Make sure blocks are ordered */
-  for (i = 0; i < vec_len (tc->opt.sacks); i++)
-    for (j = i + 1; j < vec_len (tc->opt.sacks); j++)
-      if (seq_lt (tc->opt.sacks[j].start, tc->opt.sacks[i].start))
+  for (i = 0; i < vec_len (tc->rcv_opts.sacks); i++)
+    for (j = i + 1; j < vec_len (tc->rcv_opts.sacks); j++)
+      if (seq_lt (tc->rcv_opts.sacks[j].start, tc->rcv_opts.sacks[i].start))
 	{
-	  tmp = tc->opt.sacks[i];
-	  tc->opt.sacks[i] = tc->opt.sacks[j];
-	  tc->opt.sacks[j] = tmp;
+	  tmp = tc->rcv_opts.sacks[i];
+	  tc->rcv_opts.sacks[i] = tc->rcv_opts.sacks[j];
+	  tc->rcv_opts.sacks[j] = tmp;
 	}
 
   if (sb->head == TCP_INVALID_SACK_HOLE_INDEX)
@@ -585,25 +731,25 @@ tcp_rcv_sacks (tcp_connection_t * tc, u32 ack)
       last_hole = scoreboard_insert_hole (sb, TCP_INVALID_SACK_HOLE_INDEX,
 					  tc->snd_una, tc->snd_una_max);
       sb->tail = scoreboard_hole_index (sb, last_hole);
-      tmp = tc->opt.sacks[vec_len (tc->opt.sacks) - 1];
-      sb->max_byte_sacked = tmp.end;
+      tmp = tc->rcv_opts.sacks[vec_len (tc->rcv_opts.sacks) - 1];
+      sb->high_sacked = tmp.end;
     }
   else
     {
       /* If we have holes but snd_una_max is beyond the last hole, update
        * last hole end */
-      tmp = tc->opt.sacks[vec_len (tc->opt.sacks) - 1];
+      tmp = tc->rcv_opts.sacks[vec_len (tc->rcv_opts.sacks) - 1];
       last_hole = scoreboard_last_hole (sb);
-      if (seq_gt (tc->snd_una_max, sb->max_byte_sacked)
+      if (seq_gt (tc->snd_una_max, sb->high_sacked)
 	  && seq_gt (tc->snd_una_max, last_hole->end))
 	last_hole->end = tc->snd_una_max;
     }
 
   /* Walk the holes with the SACK blocks */
   hole = pool_elt_at_index (sb->holes, sb->head);
-  while (hole && blk_index < vec_len (tc->opt.sacks))
+  while (hole && blk_index < vec_len (tc->rcv_opts.sacks))
     {
-      blk = &tc->opt.sacks[blk_index];
+      blk = &tc->rcv_opts.sacks[blk_index];
 
       if (seq_leq (blk->start, hole->start))
 	{
@@ -617,9 +763,9 @@ tcp_rcv_sacks (tcp_connection_t * tc, u32 ack)
 		{
 		  /* Bytes lost because snd_wnd left edge advances */
 		  if (next_hole && seq_leq (next_hole->start, ack))
-		    delivered_bytes += next_hole->start - hole->end;
+		    sb->last_bytes_delivered += next_hole->start - hole->end;
 		  else
-		    delivered_bytes += ack - hole->end;
+		    sb->last_bytes_delivered += ack - hole->end;
 		}
 	      else
 		{
@@ -633,8 +779,8 @@ tcp_rcv_sacks (tcp_connection_t * tc, u32 ack)
 		  last_hole = scoreboard_last_hole (sb);
 		  /* keep track of max byte sacked for when the last hole
 		   * is acked */
-		  if (seq_gt (hole->end, sb->max_byte_sacked))
-		    sb->max_byte_sacked = hole->end;
+		  if (seq_gt (hole->end, sb->high_sacked))
+		    sb->high_sacked = hole->end;
 		}
 
 	      /* snd_una needs to be advanced */
@@ -645,12 +791,12 @@ tcp_rcv_sacks (tcp_connection_t * tc, u32 ack)
 		      sb->snd_una_adv = next_hole->start - ack;
 
 		      /* all these can be delivered */
-		      delivered_bytes += sb->snd_una_adv;
+		      sb->last_bytes_delivered += sb->snd_una_adv;
 		    }
 		  else if (!next_hole)
 		    {
-		      sb->snd_una_adv = sb->max_byte_sacked - ack;
-		      delivered_bytes += sb->snd_una_adv;
+		      sb->snd_una_adv = sb->high_sacked - ack;
+		      sb->last_bytes_delivered += sb->snd_una_adv;
 		    }
 		}
 
@@ -691,28 +837,33 @@ tcp_rcv_sacks (tcp_connection_t * tc, u32 ack)
 		}
 
 	      blk_index++;
-	      hole = scoreboard_next_hole (sb, hole);
 	    }
-	  else
+	  else if (seq_leq (blk->start, hole->end))
 	    {
 	      sb->sacked_bytes += hole->end - blk->start;
 	      hole->end = blk->start;
-	      hole = scoreboard_next_hole (sb, hole);
 	    }
+
+	  hole = scoreboard_next_hole (sb, hole);
 	}
     }
 
   sb->last_sacked_bytes = sb->sacked_bytes - old_sacked_bytes;
-  sb->sacked_bytes -= delivered_bytes;
+  sb->sacked_bytes -= sb->last_bytes_delivered;
+  scoreboard_update_lost (tc, sb);
 }
 
-/** Update snd_wnd
+/**
+ * Try to update snd_wnd based on feedback received from peer.
  *
- * If (SND.WL1 < SEG.SEQ or (SND.WL1 = SEG.SEQ and SND.WL2 =< SEG.ACK)), set
- * SND.WND <- SEG.WND, set SND.WL1 <- SEG.SEQ, and set SND.WL2 <- SEG.ACK */
+ * If successful, and new window is 'effectively' 0, activate persist
+ * timer.
+ */
 static void
 tcp_update_snd_wnd (tcp_connection_t * tc, u32 seq, u32 ack, u32 snd_wnd)
 {
+  /* If (SND.WL1 < SEG.SEQ or (SND.WL1 = SEG.SEQ and SND.WL2 =< SEG.ACK)), set
+   * SND.WND <- SEG.WND, set SND.WL1 <- SEG.SEQ, and set SND.WL2 <- SEG.ACK */
   if (seq_lt (tc->snd_wl1, seq)
       || (tc->snd_wl1 == seq && seq_leq (tc->snd_wl2, ack)))
     {
@@ -721,138 +872,269 @@ tcp_update_snd_wnd (tcp_connection_t * tc, u32 seq, u32 ack, u32 snd_wnd)
       tc->snd_wl2 = ack;
       TCP_EVT_DBG (TCP_EVT_SND_WND, tc);
 
-      /* Set probe timer if we just got 0 wnd */
       if (tc->snd_wnd < tc->snd_mss)
 	{
-	  if (!tcp_timer_is_active (tc, TCP_TIMER_PERSIST))
+	  /* Set persist timer if not set and we just got 0 wnd */
+	  if (!tcp_timer_is_active (tc, TCP_TIMER_PERSIST)
+	      && !tcp_timer_is_active (tc, TCP_TIMER_RETRANSMIT))
 	    tcp_persist_timer_set (tc);
 	}
       else
-	tcp_persist_timer_reset (tc);
+	{
+	  tcp_persist_timer_reset (tc);
+	  if (!tcp_in_recovery (tc) && tc->rto_boff > 0)
+	    {
+	      tc->rto_boff = 0;
+	      tcp_update_rto (tc);
+	    }
+	}
     }
 }
 
 void
-tcp_cc_congestion (tcp_connection_t * tc)
+tcp_cc_init_congestion (tcp_connection_t * tc)
 {
-  tc->snd_congestion = tc->snd_nxt;
+  tcp_fastrecovery_on (tc);
+  tc->snd_congestion = tc->snd_una_max;
   tc->cc_algo->congestion (tc);
   TCP_EVT_DBG (TCP_EVT_CC_EVT, tc, 4);
 }
 
-void
-tcp_cc_recover (tcp_connection_t * tc)
+static void
+tcp_cc_recovery_exit (tcp_connection_t * tc)
 {
-  /* TODO: check if time to recover was small. It might be that RTO popped
-   * too soon.
-   */
+  /* Deflate rto */
+  tcp_update_rto (tc);
+  tc->rto_boff = 0;
+  tc->snd_rxt_ts = 0;
+  tcp_recovery_off (tc);
+}
 
+void
+tcp_cc_fastrecovery_exit (tcp_connection_t * tc)
+{
   tc->cc_algo->recovered (tc);
+  tc->snd_rxt_bytes = 0;
+  tc->rcv_dupacks = 0;
+  tcp_fastrecovery_off (tc);
+  tcp_fastrecovery_1_smss_off (tc);
+}
 
-  tc->rtx_bytes = 0;
+static void
+tcp_cc_congestion_undo (tcp_connection_t * tc)
+{
+  tc->cwnd = tc->prev_cwnd;
+  tc->ssthresh = tc->prev_ssthresh;
+  tc->snd_nxt = tc->snd_una_max;
   tc->rcv_dupacks = 0;
-  tc->snd_nxt = tc->snd_una;
+  if (tcp_in_recovery (tc))
+    tcp_cc_recovery_exit (tc);
+  ASSERT (tc->rto_boff == 0);
+  /* TODO extend for fastrecovery */
+}
 
-  tc->cc_algo->rcv_ack (tc);
-  tc->tsecr_last_ack = tc->opt.tsecr;
+static u8
+tcp_cc_is_spurious_retransmit (tcp_connection_t * tc)
+{
+  return (tc->snd_rxt_ts
+	  && tcp_opts_tstamp (&tc->rcv_opts)
+	  && timestamp_lt (tc->rcv_opts.tsecr, tc->snd_rxt_ts));
+}
 
-  tcp_cong_recovery_off (tc);
+int
+tcp_cc_recover (tcp_connection_t * tc)
+{
+  ASSERT (tcp_in_cong_recovery (tc));
+  if (tcp_cc_is_spurious_retransmit (tc))
+    {
+      tcp_cc_congestion_undo (tc);
+      return 1;
+    }
+
+  if (tcp_in_recovery (tc))
+    tcp_cc_recovery_exit (tc);
+  else if (tcp_in_fastrecovery (tc))
+    tcp_cc_fastrecovery_exit (tc);
+
+  ASSERT (tc->rto_boff == 0);
+  ASSERT (!tcp_in_cong_recovery (tc));
 
   TCP_EVT_DBG (TCP_EVT_CC_EVT, tc, 3);
+  return 0;
 }
 
 static void
-tcp_cc_rcv_ack (tcp_connection_t * tc, vlib_buffer_t * b)
+tcp_cc_update (tcp_connection_t * tc, vlib_buffer_t * b)
+{
+  ASSERT (!tcp_in_cong_recovery (tc));
+
+  /* Congestion avoidance */
+  tc->cc_algo->rcv_ack (tc);
+  tc->tsecr_last_ack = tc->rcv_opts.tsecr;
+
+  /* If a cumulative ack, make sure dupacks is 0 */
+  tc->rcv_dupacks = 0;
+
+  /* When dupacks hits the threshold we only enter fast retransmit if
+   * cumulative ack covers more than snd_congestion. Should snd_una
+   * wrap this test may fail under otherwise valid circumstances.
+   * Therefore, proactively update snd_congestion when wrap detected. */
+  if (PREDICT_FALSE
+      (seq_leq (tc->snd_congestion, tc->snd_una - tc->bytes_acked)
+       && seq_gt (tc->snd_congestion, tc->snd_una)))
+    tc->snd_congestion = tc->snd_una - 1;
+}
+
+static u8
+tcp_should_fastrecover_sack (tcp_connection_t * tc)
 {
-  u8 partial_ack;
-  u32 bytes_advanced;
+  return (TCP_DUPACK_THRESHOLD - 1) * tc->snd_mss < tc->sack_sb.sacked_bytes;
+}
 
-  if (tcp_in_fastrecovery (tc))
+static u8
+tcp_should_fastrecover (tcp_connection_t * tc)
+{
+  return (tc->rcv_dupacks == TCP_DUPACK_THRESHOLD
+	  || tcp_should_fastrecover_sack (tc));
+}
+
+static void
+tcp_cc_handle_event (tcp_connection_t * tc, u32 is_dack)
+{
+  /*
+   * Duplicate ACK. Check if we should enter fast recovery, or if already in
+   * it account for the bytes that left the network.
+   */
+  if (is_dack)
     {
-      partial_ack = seq_lt (tc->snd_una, tc->snd_congestion);
-      if (!partial_ack)
+      ASSERT (tc->snd_una != tc->snd_una_max
+	      || tc->sack_sb.last_sacked_bytes);
+      tc->rcv_dupacks++;
+
+      if (tc->rcv_dupacks > TCP_DUPACK_THRESHOLD && !tc->bytes_acked)
 	{
-	  /* Clear retransmitted bytes. */
-	  tcp_cc_recover (tc);
+	  ASSERT (tcp_in_fastrecovery (tc));
+	  /* Pure duplicate ack. If some data got acked, it's handled lower */
+	  tc->cc_algo->rcv_cong_ack (tc, TCP_CC_DUPACK);
+	  return;
 	}
-      else
+      else if (tcp_should_fastrecover (tc))
 	{
-	  TCP_EVT_DBG (TCP_EVT_CC_PACK, tc);
+	  /* Things are already bad */
+	  if (tcp_in_cong_recovery (tc))
+	    {
+	      tc->rcv_dupacks = 0;
+	      goto partial_ack_test;
+	    }
 
-	  /* Clear retransmitted bytes. XXX should we clear all? */
-	  tc->rtx_bytes = 0;
+	  /* If of of the two conditions lower hold, reset dupacks
+	   * 1) Cumulative ack does not cover more than congestion threshold
+	   * 2) RFC6582 heuristic to avoid multiple fast retransmits
+	   */
+	  if (seq_leq (tc->snd_una, tc->snd_congestion)
+	      || tc->rcv_opts.tsecr != tc->tsecr_last_ack)
+	    {
+	      tc->rcv_dupacks = 0;
+	      return;
+	    }
+
+	  tcp_cc_init_congestion (tc);
+	  tc->cc_algo->rcv_cong_ack (tc, TCP_CC_DUPACK);
 
-	  tc->cc_algo->rcv_cong_ack (tc, TCP_CC_PARTIALACK);
+	  /* The first segment MUST be retransmitted */
+	  tcp_retransmit_first_unacked (tc);
 
-	  /* In case snd_nxt is still in the past and output tries to
-	   * shove some new bytes */
-	  tc->snd_nxt = tc->snd_una_max;
+	  /* Post retransmit update cwnd to ssthresh and account for the
+	   * three segments that have left the network and should've been
+	   * buffered at the receiver XXX */
+	  tc->cwnd = tc->ssthresh + tc->rcv_dupacks * tc->snd_mss;
 
-	  /* XXX need proper RFC6675 support */
-	  if (tc->sack_sb.last_sacked_bytes && !tcp_in_recovery (tc))
+	  /* If cwnd allows, send more data */
+	  if (tcp_opts_sack_permitted (&tc->rcv_opts)
+	      && scoreboard_first_hole (&tc->sack_sb))
 	    {
-	      tcp_fast_retransmit (tc);
+	      scoreboard_init_high_rxt (&tc->sack_sb);
+	      tcp_fast_retransmit_sack (tc);
 	    }
 	  else
 	    {
-	      /* Retransmit first unacked segment */
-	      tcp_retransmit_first_unacked (tc);
+	      tcp_fast_retransmit_no_sack (tc);
 	    }
+
+	  return;
 	}
-    }
-  else
-    {
-      tc->cc_algo->rcv_ack (tc);
-      tc->tsecr_last_ack = tc->opt.tsecr;
-      tc->rcv_dupacks = 0;
-      if (tcp_in_recovery (tc))
+      else if (!tc->bytes_acked
+	       || (tc->bytes_acked && !tcp_in_cong_recovery (tc)))
 	{
-	  bytes_advanced = tc->bytes_acked + tc->sack_sb.snd_una_adv;
-	  tc->rtx_bytes -= clib_min (bytes_advanced, tc->rtx_bytes);
-	  tc->rto = clib_min (tc->srtt + (tc->rttvar << 2), TCP_RTO_MAX);
-	  if (seq_geq (tc->snd_una, tc->snd_congestion))
-	    {
-	      tc->rtx_bytes = 0;
-	      tcp_recovery_off (tc);
-	    }
+	  tc->cc_algo->rcv_cong_ack (tc, TCP_CC_DUPACK);
+	  return;
 	}
+      else
+	goto partial_ack;
     }
-}
 
-static void
-tcp_cc_rcv_dupack (tcp_connection_t * tc, u32 ack)
-{
-//  ASSERT (seq_geq(tc->snd_una, ack));
+partial_ack_test:
+
+  if (!tc->bytes_acked)
+    return;
+
+partial_ack:
+  /*
+   * Legitimate ACK. 1) See if we can exit recovery
+   */
+  /* XXX limit this only to first partial ack? */
+  tcp_retransmit_timer_update (tc);
 
-  tc->rcv_dupacks++;
-  if (tc->rcv_dupacks == TCP_DUPACK_THRESHOLD)
+  if (seq_geq (tc->snd_una, tc->snd_congestion))
     {
-      /* RFC6582 NewReno heuristic to avoid multiple fast retransmits */
-      if (tc->opt.tsecr != tc->tsecr_last_ack)
-	{
-	  tc->rcv_dupacks = 0;
-	  return;
-	}
+      /* If spurious return, we've already updated everything */
+      if (tcp_cc_recover (tc))
+	return;
+
+      tc->snd_nxt = tc->snd_una_max;
 
-      tcp_fastrecovery_on (tc);
+      /* Treat as congestion avoidance ack */
+      tc->cc_algo->rcv_ack (tc);
+      tc->tsecr_last_ack = tc->rcv_opts.tsecr;
+      return;
+    }
+
+  /*
+   * Legitimate ACK. 2) If PARTIAL ACK try to retransmit
+   */
+  TCP_EVT_DBG (TCP_EVT_CC_PACK, tc);
+
+  /* RFC6675: If the incoming ACK is a cumulative acknowledgment,
+   * reset dupacks to 0 */
+  tc->rcv_dupacks = 0;
 
-      /* Handle congestion and dupack */
-      tcp_cc_congestion (tc);
-      tc->cc_algo->rcv_cong_ack (tc, TCP_CC_DUPACK);
+  tcp_retransmit_first_unacked (tc);
 
-      tcp_fast_retransmit (tc);
+  /* Post RTO timeout don't try anything fancy */
+  if (tcp_in_recovery (tc))
+    return;
 
-      /* Post retransmit update cwnd to ssthresh and account for the
-       * three segments that have left the network and should've been
-       * buffered at the receiver */
-      tc->cwnd = tc->ssthresh + TCP_DUPACK_THRESHOLD * tc->snd_mss;
+  /* Remove retransmitted bytes that have been delivered */
+  if (tc->sack_sb.last_bytes_delivered
+      && seq_gt (tc->sack_sb.high_rxt, tc->snd_una))
+    {
+      /* If we have sacks and we haven't gotten an ack beyond high_rxt,
+       * remove sacked bytes delivered */
+      tc->snd_rxt_bytes -= tc->sack_sb.last_bytes_delivered;
     }
-  else if (tc->rcv_dupacks > TCP_DUPACK_THRESHOLD)
+  else
     {
-      ASSERT (tcp_in_fastrecovery (tc));
-
-      tc->cc_algo->rcv_cong_ack (tc, TCP_CC_DUPACK);
+      /* Either all retransmitted holes have been acked, or we're
+       * "in the blind" and retransmitting segment by segment */
+      tc->snd_rxt_bytes = 0;
     }
+
+  tc->cc_algo->rcv_cong_ack (tc, TCP_CC_PARTIALACK);
+
+  /*
+   * Since this was a partial ack, try to retransmit some more data
+   */
+  tcp_fast_retransmit (tc);
 }
 
 void
@@ -862,14 +1144,18 @@ tcp_cc_init (tcp_connection_t * tc)
   tc->cc_algo->init (tc);
 }
 
+/**
+ * Process incoming ACK
+ */
 static int
 tcp_rcv_ack (tcp_connection_t * tc, vlib_buffer_t * b,
 	     tcp_header_t * th, u32 * next, u32 * error)
 {
-  u32 new_snd_wnd;
+  u32 prev_snd_wnd, prev_snd_una;
+  u8 is_dack;
 
   /* If the ACK acks something not yet sent (SEG.ACK > SND.NXT) */
-  if (seq_gt (vnet_buffer (b)->tcp.ack_number, tc->snd_nxt))
+  if (PREDICT_FALSE (seq_gt (vnet_buffer (b)->tcp.ack_number, tc->snd_nxt)))
     {
       /* If we have outstanding data and this is within the window, accept it,
        * probably retransmit has timed out. Otherwise ACK segment and then
@@ -892,7 +1178,7 @@ tcp_rcv_ack (tcp_connection_t * tc, vlib_buffer_t * b,
     }
 
   /* If old ACK, probably it's an old dupack */
-  if (seq_lt (vnet_buffer (b)->tcp.ack_number, tc->snd_una))
+  if (PREDICT_FALSE (seq_lt (vnet_buffer (b)->tcp.ack_number, tc->snd_una)))
     {
       *error = TCP_ERROR_ACK_OLD;
       TCP_EVT_DBG (TCP_EVT_ACK_RCV_ERR, tc, 1,
@@ -900,54 +1186,50 @@ tcp_rcv_ack (tcp_connection_t * tc, vlib_buffer_t * b,
       if (tcp_in_fastrecovery (tc) && tc->rcv_dupacks == TCP_DUPACK_THRESHOLD)
 	{
 	  TCP_EVT_DBG (TCP_EVT_DUPACK_RCVD, tc);
-	  tcp_cc_rcv_dupack (tc, vnet_buffer (b)->tcp.ack_number);
+	  tcp_cc_handle_event (tc, 1);
 	}
       /* Don't drop yet */
       return 0;
     }
 
-  if (tcp_opts_sack_permitted (&tc->opt))
-    tcp_rcv_sacks (tc, vnet_buffer (b)->tcp.ack_number);
-
-  new_snd_wnd = clib_net_to_host_u16 (th->window) << tc->snd_wscale;
-
-  if (tcp_ack_is_dupack (tc, b, new_snd_wnd))
-    {
-      TCP_EVT_DBG (TCP_EVT_DUPACK_RCVD, tc, 1);
-      tcp_cc_rcv_dupack (tc, vnet_buffer (b)->tcp.ack_number);
-      *error = TCP_ERROR_ACK_DUP;
-      return -1;
-    }
-
   /*
-   * Valid ACK
+   * Looks okay, process feedback
    */
 
-  tc->bytes_acked = vnet_buffer (b)->tcp.ack_number - tc->snd_una;
-  tc->snd_una = vnet_buffer (b)->tcp.ack_number + tc->sack_sb.snd_una_adv;
+  TCP_EVT_DBG (TCP_EVT_ACK_RCVD, tc);
+
+  if (tcp_opts_sack_permitted (&tc->rcv_opts))
+    tcp_rcv_sacks (tc, vnet_buffer (b)->tcp.ack_number);
 
-  /* Dequeue ACKed data and update RTT */
-  tcp_dequeue_acked (tc, vnet_buffer (b)->tcp.ack_number);
+  prev_snd_wnd = tc->snd_wnd;
+  prev_snd_una = tc->snd_una;
   tcp_update_snd_wnd (tc, vnet_buffer (b)->tcp.seq_number,
-		      vnet_buffer (b)->tcp.ack_number, new_snd_wnd);
+		      vnet_buffer (b)->tcp.ack_number,
+		      clib_net_to_host_u16 (th->window) << tc->snd_wscale);
+  tc->bytes_acked = vnet_buffer (b)->tcp.ack_number - tc->snd_una;
+  tc->snd_una = vnet_buffer (b)->tcp.ack_number + tc->sack_sb.snd_una_adv;
+  tcp_validate_txf_size (tc, tc->bytes_acked);
 
-  /* If some of our sent bytes have been acked, update cc and retransmit
-   * timer. */
   if (tc->bytes_acked)
-    {
-      TCP_EVT_DBG (TCP_EVT_ACK_RCVD, tc);
+    tcp_dequeue_acked (tc, vnet_buffer (b)->tcp.ack_number);
 
-      /* Updates congestion control (slow start/congestion avoidance) */
-      tcp_cc_rcv_ack (tc, b);
+  /*
+   * Check if we have congestion event
+   */
 
-      /* If everything has been acked, stop retransmit timer
-       * otherwise update. */
-      if (tc->snd_una == tc->snd_una_max)
-	tcp_retransmit_timer_reset (tc);
-      else
-	tcp_retransmit_timer_update (tc);
+  if (tcp_ack_is_cc_event (tc, b, prev_snd_wnd, prev_snd_una, &is_dack))
+    {
+      tcp_cc_handle_event (tc, is_dack);
+      *error = TCP_ERROR_ACK_DUP;
+      TCP_EVT_DBG (TCP_EVT_DUPACK_RCVD, tc, 1);
+      return vnet_buffer (b)->tcp.data_len ? 0 : -1;
     }
 
+  /*
+   * Update congestion control (slow start/congestion avoidance)
+   */
+  tcp_cc_update (tc, b);
+
   return 0;
 }
 
@@ -1059,7 +1341,7 @@ tcp_session_enqueue_data (tcp_connection_t * tc, vlib_buffer_t * b,
     }
 
   /* Update SACK list if need be */
-  if (tcp_opts_sack_permitted (&tc->opt))
+  if (tcp_opts_sack_permitted (&tc->rcv_opts))
     {
       /* Remove SACK blocks that have been delivered */
       tcp_update_sack_list (tc, tc->rcv_nxt, tc->rcv_nxt);
@@ -1097,7 +1379,7 @@ tcp_session_enqueue_ooo (tcp_connection_t * tc, vlib_buffer_t * b,
   TCP_EVT_DBG (TCP_EVT_INPUT, tc, 1, data_len, data_len);
 
   /* Update SACK list if in use */
-  if (tcp_opts_sack_permitted (&tc->opt))
+  if (tcp_opts_sack_permitted (&tc->rcv_opts))
     {
       ooo_segment_t *newest;
       u32 start, end;
@@ -1294,7 +1576,6 @@ tcp46_established_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
       u32 n_left_to_next;
 
       vlib_get_next_frame (vm, node, next_index, to_next, n_left_to_next);
-
       while (n_left_from > 0 && n_left_to_next > 0)
 	{
 	  u32 bi0;
@@ -1321,7 +1602,6 @@ tcp46_established_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
 	    }
 
 	  th0 = tcp_buffer_hdr (b0);
-
 	  is_fin = (th0->flags & TCP_FLAG_FIN) != 0;
 
 	  /* SYNs, FINs and data consume sequence numbers */
@@ -1387,7 +1667,6 @@ tcp46_established_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
 
   errors = session_manager_flush_enqueue_events (my_thread_index);
   tcp_established_inc_counter (vm, is_ip4, TCP_ERROR_EVENT_FIFO_FULL, errors);
-
   return from_frame->n_vectors;
 }
 
@@ -1582,17 +1861,17 @@ tcp46_syn_sent_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
 	  new_tc0->irs = seq0;
 
 	  /* Parse options */
-	  if (tcp_options_parse (tcp0, &new_tc0->opt))
+	  if (tcp_options_parse (tcp0, &new_tc0->rcv_opts))
 	    goto drop;
 
-	  if (tcp_opts_tstamp (&new_tc0->opt))
+	  if (tcp_opts_tstamp (&new_tc0->rcv_opts))
 	    {
-	      new_tc0->tsval_recent = new_tc0->opt.tsval;
+	      new_tc0->tsval_recent = new_tc0->rcv_opts.tsval;
 	      new_tc0->tsval_recent_age = tcp_time_now ();
 	    }
 
-	  if (tcp_opts_wscale (&new_tc0->opt))
-	    new_tc0->snd_wscale = new_tc0->opt.wscale;
+	  if (tcp_opts_wscale (&new_tc0->rcv_opts))
+	    new_tc0->snd_wscale = new_tc0->rcv_opts.wscale;
 
 	  /* No scaling */
 	  new_tc0->snd_wnd = clib_net_to_host_u16 (tcp0->window);
@@ -1845,7 +2124,7 @@ tcp46_rcv_process_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
 	      /* Initialize session variables */
 	      tc0->snd_una = vnet_buffer (b0)->tcp.ack_number;
 	      tc0->snd_wnd = clib_net_to_host_u16 (tcp0->window)
-		<< tc0->opt.wscale;
+		<< tc0->rcv_opts.wscale;
 	      tc0->snd_wl1 = vnet_buffer (b0)->tcp.seq_number;
 	      tc0->snd_wl2 = vnet_buffer (b0)->tcp.ack_number;
 
@@ -1903,13 +2182,21 @@ tcp46_rcv_process_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
 
 	      break;
 	    case TCP_STATE_LAST_ACK:
-	      /* The only thing that can arrive in this state is an
+	      /* The only thing that [should] arrive in this state is an
 	       * acknowledgment of our FIN. If our FIN is now acknowledged,
 	       * delete the TCB, enter the CLOSED state, and return. */
 
 	      if (!tcp_rcv_ack_is_acceptable (tc0, b0))
 		goto drop;
 
+	      /* Apparently our FIN was lost */
+	      if (tcp_fin (tcp0))
+		{
+		  /* Don't "make" fin since that increments snd_nxt */
+		  tcp_send_fin (tc0);
+		  goto drop;
+		}
+
 	      tc0->state = TCP_STATE_CLOSED;
 
 	      /* Don't delete the connection/session yet. Instead, wait a
@@ -1929,8 +2216,15 @@ tcp46_rcv_process_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
 	       * retransmission of the remote FIN. Acknowledge it, and restart
 	       * the 2 MSL timeout. */
 
-	      /* TODO */
+	      if (tcp_rcv_ack (tc0, b0, tcp0, &next0, &error0))
+		goto drop;
+
+	      tcp_make_ack (tc0, b0);
+	      tcp_timer_reset (tc0, TCP_TIMER_WAITCLOSE);
+	      tcp_timer_set (tc0, TCP_TIMER_WAITCLOSE, TCP_2MSL_TIME);
+
 	      goto drop;
+
 	      break;
 	    default:
 	      ASSERT (0);
@@ -2194,7 +2488,7 @@ tcp46_listen_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
 	      goto drop;
 	    }
 
-	  if (tcp_options_parse (th0, &child0->opt))
+	  if (tcp_options_parse (th0, &child0->rcv_opts))
 	    {
 	      goto drop;
 	    }
@@ -2205,14 +2499,14 @@ tcp46_listen_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
 
 	  /* RFC1323: TSval timestamps sent on {SYN} and {SYN,ACK}
 	   * segments are used to initialize PAWS. */
-	  if (tcp_opts_tstamp (&child0->opt))
+	  if (tcp_opts_tstamp (&child0->rcv_opts))
 	    {
-	      child0->tsval_recent = child0->opt.tsval;
+	      child0->tsval_recent = child0->rcv_opts.tsval;
 	      child0->tsval_recent_age = tcp_time_now ();
 	    }
 
-	  if (tcp_opts_wscale (&child0->opt))
-	    child0->snd_wscale = child0->opt.wscale;
+	  if (tcp_opts_wscale (&child0->rcv_opts))
+	    child0->snd_wscale = child0->rcv_opts.wscale;
 
 	  /* No scaling */
 	  child0->snd_wnd = clib_net_to_host_u16 (th0->window);
@@ -2477,7 +2771,6 @@ tcp46_input_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
 		vlib_add_trace (vm, node, b0, sizeof (*t0));
 	      tcp_set_rx_trace_data (t0, tc0, tcp0, b0, is_ip4);
 	    }
-
 	  vlib_validate_buffer_enqueue_x1 (vm, node, next_index, to_next,
 					   n_left_to_next, bi0, next0);
 	}
@@ -2600,7 +2893,13 @@ do {                                                       	\
   _(FIN_WAIT_2, TCP_FLAG_FIN | TCP_FLAG_ACK, TCP_INPUT_NEXT_RCV_PROCESS,
     TCP_ERROR_NONE);
   _(LAST_ACK, TCP_FLAG_ACK, TCP_INPUT_NEXT_RCV_PROCESS, TCP_ERROR_NONE);
+  _(LAST_ACK, TCP_FLAG_FIN, TCP_INPUT_NEXT_RCV_PROCESS, TCP_ERROR_NONE);
+  _(LAST_ACK, TCP_FLAG_FIN | TCP_FLAG_ACK, TCP_INPUT_NEXT_RCV_PROCESS,
+    TCP_ERROR_NONE);
   _(LAST_ACK, TCP_FLAG_RST, TCP_INPUT_NEXT_RCV_PROCESS, TCP_ERROR_NONE);
+  _(TIME_WAIT, TCP_FLAG_FIN, TCP_INPUT_NEXT_RCV_PROCESS, TCP_ERROR_NONE);
+  _(TIME_WAIT, TCP_FLAG_FIN | TCP_FLAG_ACK, TCP_INPUT_NEXT_RCV_PROCESS,
+    TCP_ERROR_NONE);
   _(CLOSED, TCP_FLAG_ACK, TCP_INPUT_NEXT_RESET, TCP_ERROR_CONNECTION_CLOSED);
   _(CLOSED, TCP_FLAG_RST, TCP_INPUT_NEXT_DROP, TCP_ERROR_CONNECTION_CLOSED);
 #undef _
diff --git a/src/vnet/tcp/tcp_newreno.c b/src/vnet/tcp/tcp_newreno.c
index 3525f4e5..c66250e4 100644
--- a/src/vnet/tcp/tcp_newreno.c
+++ b/src/vnet/tcp/tcp_newreno.c
@@ -51,9 +51,23 @@ newreno_rcv_cong_ack (tcp_connection_t * tc, tcp_cc_ack_t ack_type)
     }
   else if (ack_type == TCP_CC_PARTIALACK)
     {
-      tc->cwnd -= tc->bytes_acked;
-      if (tc->bytes_acked > tc->snd_mss)
-	tc->bytes_acked += tc->snd_mss;
+      /* RFC 6582 Sec. 3.2 */
+      if (!tcp_opts_sack_permitted (&tc->rcv_opts))
+	{
+	  /* Deflate the congestion window by the amount of new data
+	   * acknowledged by the Cumulative Acknowledgment field.
+	   * If the partial ACK acknowledges at least one SMSS of new data,
+	   * then add back SMSS bytes to the congestion window. This
+	   * artificially inflates the congestion window in order to reflect
+	   * the additional segment that has left the network. This "partial
+	   * window deflation" attempts to ensure that, when fast recovery
+	   * eventually ends, approximately ssthresh amount of data will be
+	   * outstanding in the network.*/
+	  tc->cwnd = (tc->cwnd > tc->bytes_acked) ?
+	    tc->cwnd - tc->bytes_acked : 0;
+	  if (tc->bytes_acked > tc->snd_mss)
+	    tc->cwnd += tc->snd_mss;
+	}
     }
 }
 
diff --git a/src/vnet/tcp/tcp_output.c b/src/vnet/tcp/tcp_output.c
index 49fd6bef..47c94e6d 100644
--- a/src/vnet/tcp/tcp_output.c
+++ b/src/vnet/tcp/tcp_output.c
@@ -136,10 +136,10 @@ tcp_update_rcv_wnd (tcp_connection_t * tc)
    * Figure out how much space we have available
    */
   available_space = stream_session_max_rx_enqueue (&tc->connection);
-  max_fifo = stream_session_fifo_size (&tc->connection);
+  max_fifo = stream_session_rx_fifo_size (&tc->connection);
 
-  ASSERT (tc->opt.mss < max_fifo);
-  if (available_space < tc->opt.mss && available_space < max_fifo >> 3)
+  ASSERT (tc->rcv_opts.mss < max_fifo);
+  if (available_space < tc->rcv_opts.mss && available_space < max_fifo >> 3)
     available_space = 0;
 
   /*
@@ -276,8 +276,11 @@ tcp_make_syn_options (tcp_options_t * opts, u8 wnd_scale)
   opts->tsecr = 0;
   len += TCP_OPTION_LEN_TIMESTAMP;
 
-  opts->flags |= TCP_OPTS_FLAG_SACK_PERMITTED;
-  len += TCP_OPTION_LEN_SACK_PERMITTED;
+  if (TCP_USE_SACKS)
+    {
+      opts->flags |= TCP_OPTS_FLAG_SACK_PERMITTED;
+      len += TCP_OPTION_LEN_SACK_PERMITTED;
+    }
 
   /* Align to needed boundary */
   len += (TCP_OPTS_ALIGN - len % TCP_OPTS_ALIGN) % TCP_OPTS_ALIGN;
@@ -293,14 +296,14 @@ tcp_make_synack_options (tcp_connection_t * tc, tcp_options_t * opts)
   opts->mss = tc->mss;
   len += TCP_OPTION_LEN_MSS;
 
-  if (tcp_opts_wscale (&tc->opt))
+  if (tcp_opts_wscale (&tc->rcv_opts))
     {
       opts->flags |= TCP_OPTS_FLAG_WSCALE;
       opts->wscale = tc->rcv_wscale;
       len += TCP_OPTION_LEN_WINDOW_SCALE;
     }
 
-  if (tcp_opts_tstamp (&tc->opt))
+  if (tcp_opts_tstamp (&tc->rcv_opts))
     {
       opts->flags |= TCP_OPTS_FLAG_TSTAMP;
       opts->tsval = tcp_time_now ();
@@ -308,7 +311,7 @@ tcp_make_synack_options (tcp_connection_t * tc, tcp_options_t * opts)
       len += TCP_OPTION_LEN_TIMESTAMP;
     }
 
-  if (tcp_opts_sack_permitted (&tc->opt))
+  if (tcp_opts_sack_permitted (&tc->rcv_opts))
     {
       opts->flags |= TCP_OPTS_FLAG_SACK_PERMITTED;
       len += TCP_OPTION_LEN_SACK_PERMITTED;
@@ -326,14 +329,14 @@ tcp_make_established_options (tcp_connection_t * tc, tcp_options_t * opts)
 
   opts->flags = 0;
 
-  if (tcp_opts_tstamp (&tc->opt))
+  if (tcp_opts_tstamp (&tc->rcv_opts))
     {
       opts->flags |= TCP_OPTS_FLAG_TSTAMP;
       opts->tsval = tcp_time_now ();
       opts->tsecr = tc->tsval_recent;
       len += TCP_OPTION_LEN_TIMESTAMP;
     }
-  if (tcp_opts_sack_permitted (&tc->opt))
+  if (tcp_opts_sack_permitted (&tc->rcv_opts))
     {
       if (vec_len (tc->snd_sacks))
 	{
@@ -395,7 +398,7 @@ tcp_update_snd_mss (tcp_connection_t * tc)
     tcp_make_options (tc, &tc->snd_opts, TCP_STATE_ESTABLISHED);
 
   /* XXX check if MTU has been updated */
-  tc->snd_mss = clib_min (tc->mss, tc->opt.mss) - tc->snd_opts_len;
+  tc->snd_mss = clib_min (tc->mss, tc->rcv_opts.mss) - tc->snd_opts_len;
   ASSERT (tc->snd_mss > 0);
 }
 
@@ -406,21 +409,21 @@ tcp_init_mss (tcp_connection_t * tc)
   tcp_update_rcv_mss (tc);
 
   /* TODO cache mss and consider PMTU discovery */
-  tc->snd_mss = clib_min (tc->opt.mss, tc->mss);
+  tc->snd_mss = clib_min (tc->rcv_opts.mss, tc->mss);
 
   if (tc->snd_mss < 45)
     {
       clib_warning ("snd mss is 0");
       /* Assume that at least the min default mss works */
       tc->snd_mss = default_min_mss;
-      tc->opt.mss = default_min_mss;
+      tc->rcv_opts.mss = default_min_mss;
     }
 
   /* We should have enough space for 40 bytes of options */
   ASSERT (tc->snd_mss > 45);
 
   /* If we use timestamp option, account for it */
-  if (tcp_opts_tstamp (&tc->opt))
+  if (tcp_opts_tstamp (&tc->rcv_opts))
     tc->snd_mss -= TCP_OPTION_LEN_TIMESTAMP;
 }
 
@@ -879,6 +882,7 @@ tcp_send_fin (tcp_connection_t * tc)
   tcp_make_fin (tc, b);
   tcp_enqueue_to_output (vm, b, bi, tc->c_is_ip4);
   tc->flags |= TCP_CONN_FINSNT;
+  tcp_retransmit_timer_force_update (tc);
   TCP_EVT_DBG (TCP_EVT_FIN_SENT, tc);
 }
 
@@ -919,10 +923,7 @@ tcp_push_hdr_i (tcp_connection_t * tc, vlib_buffer_t * b,
   if (compute_opts)
     tc->snd_opts_len = tcp_make_options (tc, &tc->snd_opts, tc->state);
 
-  /* Write pre-computed options */
   tcp_hdr_opts_len = tc->snd_opts_len + sizeof (tcp_header_t);
-
-  /* Get rcv window to advertise */
   advertise_wnd = tcp_window_to_advertise (tc, next_state);
   flags = tcp_make_state_flags (next_state);
 
@@ -930,26 +931,25 @@ tcp_push_hdr_i (tcp_connection_t * tc, vlib_buffer_t * b,
   th = vlib_buffer_push_tcp (b, tc->c_lcl_port, tc->c_rmt_port, tc->snd_nxt,
 			     tc->rcv_nxt, tcp_hdr_opts_len, flags,
 			     advertise_wnd);
-
   opts_write_len = tcp_options_write ((u8 *) (th + 1), &tc->snd_opts);
 
   ASSERT (opts_write_len == tc->snd_opts_len);
-
-  /* Tag the buffer with the connection index  */
   vnet_buffer (b)->tcp.connection_index = tc->c_c_index;
 
+  /*
+   * Update connection variables
+   */
+
   tc->snd_nxt += data_len;
   tc->rcv_las = tc->rcv_nxt;
 
   /* TODO this is updated in output as well ... */
-  if (tc->snd_nxt > tc->snd_una_max)
-    tc->snd_una_max = tc->snd_nxt;
-
-  if (tc->rtt_ts == 0)
+  if (seq_gt (tc->snd_nxt, tc->snd_una_max))
     {
-      tc->rtt_ts = tcp_time_now ();
-      tc->rtt_seq = tc->snd_nxt;
+      tc->snd_una_max = tc->snd_nxt;
+      tcp_validate_txf_size (tc, tc->snd_una_max - tc->snd_una);
     }
+
   TCP_EVT_DBG (TCP_EVT_PKTIZE, tc);
 }
 
@@ -987,13 +987,14 @@ tcp_timer_delack_handler (u32 index)
  *
  * @return the number of bytes in the segment or 0 if there's nothing to
  *         retransmit
- * */
+ */
 u32
 tcp_prepare_retransmit_segment (tcp_connection_t * tc, vlib_buffer_t * b,
 				u32 offset, u32 max_bytes)
 {
   vlib_main_t *vm = vlib_get_main ();
-  u32 n_bytes = 0;
+  int n_bytes = 0;
+  u32 start;
 
   tcp_reuse_buffer (vm, b);
 
@@ -1001,15 +1002,16 @@ tcp_prepare_retransmit_segment (tcp_connection_t * tc, vlib_buffer_t * b,
   ASSERT (max_bytes != 0);
 
   max_bytes = clib_min (tc->snd_mss, max_bytes);
+  start = tc->snd_una + offset;
 
   /* Start is beyond snd_congestion */
-  if (seq_geq (tc->snd_una + offset, tc->snd_congestion))
+  if (seq_geq (start, tc->snd_congestion))
     goto done;
 
   /* Don't overshoot snd_congestion */
-  if (seq_gt (tc->snd_nxt + max_bytes, tc->snd_congestion))
+  if (seq_gt (start + max_bytes, tc->snd_congestion))
     {
-      max_bytes = tc->snd_congestion - tc->snd_nxt;
+      max_bytes = tc->snd_congestion - start;
       if (max_bytes == 0)
 	goto done;
     }
@@ -1021,15 +1023,12 @@ tcp_prepare_retransmit_segment (tcp_connection_t * tc, vlib_buffer_t * b,
   n_bytes = stream_session_peek_bytes (&tc->connection,
 				       vlib_buffer_get_current (b), offset,
 				       max_bytes);
-  ASSERT (n_bytes != 0);
+  ASSERT (n_bytes > 0);
   b->current_length = n_bytes;
   tcp_push_hdr_i (tc, b, tc->state, 0);
 
-  /* Don't count multiple retransmits of the same segment */
-  if (tc->rto_boff > 1)
-    goto done;
-
-  tc->rtx_bytes += n_bytes;
+  if (tcp_in_fastrecovery (tc))
+    tc->snd_rxt_bytes += n_bytes;
 
 done:
   TCP_EVT_DBG (TCP_EVT_CC_RTX, tc, offset, n_bytes);
@@ -1042,18 +1041,15 @@ done:
 static void
 tcp_rtx_timeout_cc (tcp_connection_t * tc)
 {
+  tc->prev_ssthresh = tc->ssthresh;
+  tc->prev_cwnd = tc->cwnd;
+
   /* Cleanly recover cc (also clears up fast retransmit) */
   if (tcp_in_fastrecovery (tc))
-    {
-      tcp_cc_recover (tc);
-    }
-  else
-    {
-      tc->ssthresh = clib_max (tcp_flight_size (tc) / 2, 2 * tc->snd_mss);
-    }
+    tcp_cc_fastrecovery_exit (tc);
 
   /* Start again from the beginning */
-
+  tc->ssthresh = clib_max (tcp_flight_size (tc) / 2, 2 * tc->snd_mss);
   tc->cwnd = tcp_loss_wnd (tc);
   tc->snd_congestion = tc->snd_una_max;
   tcp_recovery_on (tc);
@@ -1081,18 +1077,31 @@ tcp_timer_retransmit_handler_i (u32 index, u8 is_syn)
   /* Make sure timer handle is set to invalid */
   tc->timers[TCP_TIMER_RETRANSMIT] = TCP_TIMER_HANDLE_INVALID;
 
+  if (!tcp_in_recovery (tc) && tc->rto_boff > 0
+      && tc->state >= TCP_STATE_ESTABLISHED)
+    {
+      tc->rto_boff = 0;
+      tcp_update_rto (tc);
+    }
+
   /* Increment RTO backoff (also equal to number of retries) */
   tc->rto_boff += 1;
 
   /* Go back to first un-acked byte */
   tc->snd_nxt = tc->snd_una;
 
-  /* Get buffer */
   tcp_get_free_buffer_index (tm, &bi);
   b = vlib_get_buffer (vm, bi);
 
   if (tc->state >= TCP_STATE_ESTABLISHED)
     {
+      /* Lost FIN, retransmit and return */
+      if (tc->flags & TCP_CONN_FINSNT)
+	{
+	  tcp_send_fin (tc);
+	  return;
+	}
+
       /* First retransmit timeout */
       if (tc->rto_boff == 1)
 	tcp_rtx_timeout_cc (tc);
@@ -1102,24 +1111,30 @@ tcp_timer_retransmit_handler_i (u32 index, u8 is_syn)
 
       TCP_EVT_DBG (TCP_EVT_CC_EVT, tc, 1);
 
-      /* Send one segment. No fancy recovery for now! */
+      /* Send one segment */
       n_bytes = tcp_prepare_retransmit_segment (tc, b, 0, tc->snd_mss);
+      /* TODO be less aggressive about this */
       scoreboard_clear (&tc->sack_sb);
 
       if (n_bytes == 0)
 	{
 	  clib_warning ("could not retransmit anything");
+	  clib_warning ("%U", format_tcp_connection, tc, 2);
+
 	  /* Try again eventually */
 	  tcp_retransmit_timer_set (tc);
+	  ASSERT (0 || (tc->rto_boff > 1
+			&& tc->snd_una == tc->snd_congestion));
 	  return;
 	}
+
+      /* For first retransmit, record timestamp (Eifel detection RFC3522) */
+      if (tc->rto_boff == 1)
+	tc->snd_rxt_ts = tcp_time_now ();
     }
-  else
+  /* Retransmit for SYN/SYNACK */
+  else if (tc->state == TCP_STATE_SYN_RCVD || tc->state == TCP_STATE_SYN_SENT)
     {
-      /* Retransmit for SYN/SYNACK */
-      ASSERT (tc->state == TCP_STATE_SYN_RCVD
-	      || tc->state == TCP_STATE_SYN_SENT);
-
       /* Try without increasing RTO a number of times. If this fails,
        * start growing RTO exponentially */
       if (tc->rto_boff > TCP_RTO_SYN_RETRIES)
@@ -1132,6 +1147,12 @@ tcp_timer_retransmit_handler_i (u32 index, u8 is_syn)
       /* Account for the SYN */
       tc->snd_nxt += 1;
     }
+  else
+    {
+      ASSERT (tc->state == TCP_STATE_CLOSED);
+      clib_warning ("connection closed ...");
+      return;
+    }
 
   if (!is_syn)
     {
@@ -1180,7 +1201,8 @@ tcp_timer_persist_handler (u32 index)
   u32 thread_index = vlib_get_thread_index ();
   tcp_connection_t *tc;
   vlib_buffer_t *b;
-  u32 bi, n_bytes;
+  u32 bi, old_snd_nxt;
+  int n_bytes = 0;
 
   tc = tcp_connection_get_if_valid (index, thread_index);
 
@@ -1202,13 +1224,15 @@ tcp_timer_persist_handler (u32 index)
   /* Try to force the first unsent segment  */
   tcp_get_free_buffer_index (tm, &bi);
   b = vlib_get_buffer (vm, bi);
+
+  tcp_validate_txf_size (tc, tc->snd_una_max - tc->snd_una);
   tc->snd_opts_len = tcp_make_options (tc, &tc->snd_opts, tc->state);
   n_bytes = stream_session_peek_bytes (&tc->connection,
 				       vlib_buffer_get_current (b),
 				       tc->snd_una_max - tc->snd_una,
 				       tc->snd_mss);
   /* Nothing to send */
-  if (n_bytes == 0)
+  if (n_bytes <= 0)
     {
       clib_warning ("persist found nothing to send");
       tcp_return_buffer (tm);
@@ -1216,7 +1240,13 @@ tcp_timer_persist_handler (u32 index)
     }
 
   b->current_length = n_bytes;
+  ASSERT (tc->snd_nxt == tc->snd_una_max || tc->rto_boff > 1
+	  || tcp_timer_is_active (tc, TCP_TIMER_RETRANSMIT));
+
+  /* Allow updating of snd_una_max but don't update snd_nxt */
+  old_snd_nxt = tc->snd_nxt;
   tcp_push_hdr_i (tc, b, tc->state, 0);
+  tc->snd_nxt = old_snd_nxt;
   tcp_enqueue_to_output (vm, b, bi, tc->c_is_ip4);
 
   /* Re-enable persist timer */
@@ -1232,8 +1262,9 @@ tcp_retransmit_first_unacked (tcp_connection_t * tc)
   tcp_main_t *tm = vnet_get_tcp_main ();
   vlib_main_t *vm = vlib_get_main ();
   vlib_buffer_t *b;
-  u32 bi, n_bytes;
+  u32 bi, n_bytes, old_snd_nxt;
 
+  old_snd_nxt = tc->snd_nxt;
   tc->snd_nxt = tc->snd_una;
 
   /* Get buffer */
@@ -1244,75 +1275,117 @@ tcp_retransmit_first_unacked (tcp_connection_t * tc)
 
   n_bytes = tcp_prepare_retransmit_segment (tc, b, 0, tc->snd_mss);
   if (n_bytes == 0)
-    goto done;
+    {
+      tcp_return_buffer (tm);
+      goto done;
+    }
 
   tcp_enqueue_to_output (vm, b, bi, tc->c_is_ip4);
 
 done:
-  tc->snd_nxt = tc->snd_una_max;
+  tc->snd_nxt = old_snd_nxt;
 }
 
-sack_scoreboard_hole_t *
-scoreboard_first_rtx_hole (sack_scoreboard_t * sb)
+/**
+ * Do fast retransmit with SACKs
+ */
+void
+tcp_fast_retransmit_sack (tcp_connection_t * tc)
 {
-  sack_scoreboard_hole_t *hole = 0;
-
-//  hole = scoreboard_first_hole (&tc->sack_sb);
-//  if (hole)
-//    {
-//
-//      offset = hole->start - tc->snd_una;
-//      hole_size = hole->end - hole->start;
-//
-//      ASSERT(hole_size);
-//
-//      if (hole_size < max_bytes)
-//      max_bytes = hole_size;
-//    }
-  return hole;
+  tcp_main_t *tm = vnet_get_tcp_main ();
+  vlib_main_t *vm = vlib_get_main ();
+  u32 n_written = 0, offset = 0, max_bytes;
+  vlib_buffer_t *b;
+  sack_scoreboard_hole_t *hole;
+  sack_scoreboard_t *sb;
+  u32 bi, old_snd_nxt;
+  int snd_space;
+  u8 snd_limited = 0, can_rescue = 0;
+
+  ASSERT (tcp_in_fastrecovery (tc));
+  TCP_EVT_DBG (TCP_EVT_CC_EVT, tc, 0);
+
+  old_snd_nxt = tc->snd_nxt;
+  sb = &tc->sack_sb;
+  snd_space = tcp_available_snd_space (tc);
+
+  hole = scoreboard_get_hole (sb, sb->cur_rxt_hole);
+  while (hole && snd_space > 0)
+    {
+      tcp_get_free_buffer_index (tm, &bi);
+      b = vlib_get_buffer (vm, bi);
+
+      hole = scoreboard_next_rxt_hole (sb, hole,
+				       tcp_fastrecovery_sent_1_smss (tc),
+				       &can_rescue, &snd_limited);
+      if (!hole)
+	{
+	  if (!can_rescue || !(seq_lt (sb->rescue_rxt, tc->snd_una)
+			       || seq_gt (sb->rescue_rxt,
+					  tc->snd_congestion)))
+	    break;
+
+	  /* If rescue rxt undefined or less than snd_una then one segment of
+	   * up to SMSS octets that MUST include the highest outstanding
+	   * unSACKed sequence number SHOULD be returned, and RescueRxt set to
+	   * RecoveryPoint. HighRxt MUST NOT be updated.
+	   */
+	  max_bytes = clib_min (tc->snd_mss, snd_space);
+	  offset = tc->snd_congestion - tc->snd_una - max_bytes;
+	  sb->rescue_rxt = tc->snd_congestion;
+	  tc->snd_nxt = tc->snd_una + offset;
+	  tcp_prepare_retransmit_segment (tc, b, offset, max_bytes);
+	  tcp_enqueue_to_output (vm, b, bi, tc->c_is_ip4);
+	  break;
+	}
+
+      max_bytes = snd_limited ? tc->snd_mss : hole->end - sb->high_rxt;
+      offset = sb->high_rxt - tc->snd_una;
+      tc->snd_nxt = tc->snd_una + offset;
+      n_written = tcp_prepare_retransmit_segment (tc, b, offset, max_bytes);
+
+      /* Nothing left to retransmit */
+      if (n_written == 0)
+	{
+	  tcp_return_buffer (tm);
+	  break;
+	}
+
+      sb->high_rxt += n_written;
+      tcp_enqueue_to_output (vm, b, bi, tc->c_is_ip4);
+      snd_space -= n_written;
+    }
+
+  /* If window allows, send 1 SMSS of new data */
+  tc->snd_nxt = old_snd_nxt;
 }
 
 /**
- * Do fast retransmit.
+ * Fast retransmit without SACK info
  */
 void
-tcp_fast_retransmit (tcp_connection_t * tc)
+tcp_fast_retransmit_no_sack (tcp_connection_t * tc)
 {
   tcp_main_t *tm = vnet_get_tcp_main ();
   vlib_main_t *vm = vlib_get_main ();
-  u32 bi;
+  u32 n_written = 0, offset = 0, bi, old_snd_nxt;
   int snd_space;
-  u32 n_written = 0, offset = 0;
   vlib_buffer_t *b;
-  u8 use_sacks = 0;
 
   ASSERT (tcp_in_fastrecovery (tc));
+  TCP_EVT_DBG (TCP_EVT_CC_EVT, tc, 0);
 
   /* Start resending from first un-acked segment */
+  old_snd_nxt = tc->snd_nxt;
   tc->snd_nxt = tc->snd_una;
-
   snd_space = tcp_available_snd_space (tc);
-  TCP_EVT_DBG (TCP_EVT_CC_EVT, tc, 0);
-
-  /* If we have SACKs use them */
-  if (tcp_opts_sack_permitted (&tc->opt)
-      && scoreboard_first_hole (&tc->sack_sb))
-    use_sacks = 0;
 
   while (snd_space > 0)
     {
       tcp_get_free_buffer_index (tm, &bi);
       b = vlib_get_buffer (vm, bi);
 
-      if (use_sacks)
-	{
-	  scoreboard_first_rtx_hole (&tc->sack_sb);
-	}
-      else
-	{
-	  offset += n_written;
-	}
-
+      offset += n_written;
       n_written = tcp_prepare_retransmit_segment (tc, b, offset, snd_space);
 
       /* Nothing left to retransmit */
@@ -1326,9 +1399,21 @@ tcp_fast_retransmit (tcp_connection_t * tc)
       snd_space -= n_written;
     }
 
-  /* If window allows, send 1 SMSS of new data */
-  if (seq_lt (tc->snd_nxt, tc->snd_congestion))
-    tc->snd_nxt = tc->snd_congestion;
+  /* Restore snd_nxt. If window allows, send 1 SMSS of new data */
+  tc->snd_nxt = old_snd_nxt;
+}
+
+/**
+ * Do fast retransmit
+ */
+void
+tcp_fast_retransmit (tcp_connection_t * tc)
+{
+  if (tcp_opts_sack_permitted (&tc->rcv_opts)
+      && scoreboard_first_hole (&tc->sack_sb))
+    tcp_fast_retransmit_sack (tc);
+  else
+    tcp_fast_retransmit_no_sack (tc);
 }
 
 always_inline u32
@@ -1544,6 +1629,12 @@ tcp_push_header (transport_connection_t * tconn, vlib_buffer_t * b)
 
   tc = (tcp_connection_t *) tconn;
   tcp_push_hdr_i (tc, b, TCP_STATE_ESTABLISHED, 0);
+
+  if (tc->rtt_ts == 0)
+    {
+      tc->rtt_ts = tcp_time_now ();
+      tc->rtt_seq = tc->snd_nxt;
+    }
   return 0;
 }
 
diff --git a/src/vnet/tcp/tcp_test.c b/src/vnet/tcp/tcp_test.c
index 2af38484..3f8afa40 100644
--- a/src/vnet/tcp/tcp_test.c
+++ b/src/vnet/tcp/tcp_test.c
@@ -54,7 +54,7 @@ tcp_test_sack_rx (vlib_main_t * vm, unformat_input_t * input)
   tc->snd_una = 0;
   tc->snd_una_max = 1000;
   tc->snd_nxt = 1000;
-  tc->opt.flags |= TCP_OPTS_FLAG_SACK;
+  tc->rcv_opts.flags |= TCP_OPTS_FLAG_SACK;
   scoreboard_init (&tc->sack_sb);
 
   for (i = 0; i < 1000 / 100; i++)
@@ -70,9 +70,9 @@ tcp_test_sack_rx (vlib_main_t * vm, unformat_input_t * input)
 
   for (i = 0; i < 1000 / 200; i++)
     {
-      vec_add1 (tc->opt.sacks, sacks[i * 2]);
+      vec_add1 (tc->rcv_opts.sacks, sacks[i * 2]);
     }
-  tc->opt.n_sack_blocks = vec_len (tc->opt.sacks);
+  tc->rcv_opts.n_sack_blocks = vec_len (tc->rcv_opts.sacks);
   tcp_rcv_sacks (tc, 0);
 
   if (verbose)
@@ -93,18 +93,17 @@ tcp_test_sack_rx (vlib_main_t * vm, unformat_input_t * input)
   TCP_TEST ((sb->snd_una_adv == 0), "snd_una_adv %u", sb->snd_una_adv);
   TCP_TEST ((sb->last_sacked_bytes == 400),
 	    "last sacked bytes %d", sb->last_sacked_bytes);
-  TCP_TEST ((sb->max_byte_sacked == 900),
-	    "max byte sacked %u", sb->max_byte_sacked);
+  TCP_TEST ((sb->high_sacked == 900), "max byte sacked %u", sb->high_sacked);
   /*
    * Inject odd blocks
    */
 
-  vec_reset_length (tc->opt.sacks);
+  vec_reset_length (tc->rcv_opts.sacks);
   for (i = 0; i < 1000 / 200; i++)
     {
-      vec_add1 (tc->opt.sacks, sacks[i * 2 + 1]);
+      vec_add1 (tc->rcv_opts.sacks, sacks[i * 2 + 1]);
     }
-  tc->opt.n_sack_blocks = vec_len (tc->opt.sacks);
+  tc->rcv_opts.n_sack_blocks = vec_len (tc->rcv_opts.sacks);
   tcp_rcv_sacks (tc, 0);
 
   if (verbose)
@@ -118,8 +117,7 @@ tcp_test_sack_rx (vlib_main_t * vm, unformat_input_t * input)
 	    "first hole start %u end %u", hole->start, hole->end);
   TCP_TEST ((sb->sacked_bytes == 900), "sacked bytes %d", sb->sacked_bytes);
   TCP_TEST ((sb->snd_una_adv == 0), "snd_una_adv %u", sb->snd_una_adv);
-  TCP_TEST ((sb->max_byte_sacked == 1000),
-	    "max sacked byte %u", sb->max_byte_sacked);
+  TCP_TEST ((sb->high_sacked == 1000), "max sacked byte %u", sb->high_sacked);
   TCP_TEST ((sb->last_sacked_bytes == 500),
 	    "last sacked bytes %d", sb->last_sacked_bytes);
 
@@ -135,8 +133,7 @@ tcp_test_sack_rx (vlib_main_t * vm, unformat_input_t * input)
 	    "scoreboard has %d elements", pool_elts (sb->holes));
   TCP_TEST ((sb->snd_una_adv == 900),
 	    "snd_una_adv after ack %u", sb->snd_una_adv);
-  TCP_TEST ((sb->max_byte_sacked == 1000),
-	    "max sacked byte %u", sb->max_byte_sacked);
+  TCP_TEST ((sb->high_sacked == 1000), "max sacked byte %u", sb->high_sacked);
   TCP_TEST ((sb->sacked_bytes == 0), "sacked bytes %d", sb->sacked_bytes);
   TCP_TEST ((sb->last_sacked_bytes == 0),
 	    "last sacked bytes %d", sb->last_sacked_bytes);
@@ -145,11 +142,11 @@ tcp_test_sack_rx (vlib_main_t * vm, unformat_input_t * input)
    * Add new block
    */
 
-  vec_reset_length (tc->opt.sacks);
+  vec_reset_length (tc->rcv_opts.sacks);
 
   block.start = 1200;
   block.end = 1300;
-  vec_add1 (tc->opt.sacks, block);
+  vec_add1 (tc->rcv_opts.sacks, block);
 
   if (verbose)
     vlib_cli_output (vm, "add [1200, 1300]:\n%U", format_tcp_scoreboard, sb);
@@ -171,8 +168,7 @@ tcp_test_sack_rx (vlib_main_t * vm, unformat_input_t * input)
 	    "first hole start %u end %u", hole->start, hole->end);
   TCP_TEST ((sb->snd_una_adv == 0),
 	    "snd_una_adv after ack %u", sb->snd_una_adv);
-  TCP_TEST ((sb->max_byte_sacked == 1300),
-	    "max sacked byte %u", sb->max_byte_sacked);
+  TCP_TEST ((sb->high_sacked == 1300), "max sacked byte %u", sb->high_sacked);
   hole = scoreboard_last_hole (sb);
   TCP_TEST ((hole->start == 1300 && hole->end == 1500),
 	    "last hole start %u end %u", hole->start, hole->end);
@@ -182,7 +178,7 @@ tcp_test_sack_rx (vlib_main_t * vm, unformat_input_t * input)
    * Ack first hole
    */
 
-  vec_reset_length (tc->opt.sacks);
+  vec_reset_length (tc->rcv_opts.sacks);
   tcp_rcv_sacks (tc, 1200);
 
   if (verbose)
@@ -196,8 +192,16 @@ tcp_test_sack_rx (vlib_main_t * vm, unformat_input_t * input)
 	    "scoreboard has %d elements", pool_elts (sb->holes));
 
   /*
-   * Remove all
+   * Add some more blocks and then remove all
    */
+  vec_reset_length (tc->rcv_opts.sacks);
+  for (i = 0; i < 5; i++)
+    {
+      block.start = i * 100 + 1200;
+      block.end = (i + 1) * 100 + 1200;
+      vec_add1 (tc->rcv_opts.sacks, block);
+    }
+  tcp_rcv_sacks (tc, 1900);
 
   scoreboard_clear (sb);
   if (verbose)
@@ -205,6 +209,9 @@ tcp_test_sack_rx (vlib_main_t * vm, unformat_input_t * input)
 
   TCP_TEST ((pool_elts (sb->holes) == 0),
 	    "number of holes %d", pool_elts (sb->holes));
+  TCP_TEST ((sb->head == TCP_INVALID_SACK_HOLE_INDEX), "head %u", sb->head);
+  TCP_TEST ((sb->tail == TCP_INVALID_SACK_HOLE_INDEX), "tail %u", sb->tail);
+
   /*
    * Re-inject odd blocks and ack them all
    */
@@ -214,9 +221,9 @@ tcp_test_sack_rx (vlib_main_t * vm, unformat_input_t * input)
   tc->snd_nxt = 1000;
   for (i = 0; i < 5; i++)
     {
-      vec_add1 (tc->opt.sacks, sacks[i * 2 + 1]);
+      vec_add1 (tc->rcv_opts.sacks, sacks[i * 2 + 1]);
     }
-  tc->opt.n_sack_blocks = vec_len (tc->opt.sacks);
+  tc->rcv_opts.n_sack_blocks = vec_len (tc->rcv_opts.sacks);
   tcp_rcv_sacks (tc, 0);
   if (verbose)
     vlib_cli_output (vm, "sb added odd blocks and ack [0, 950]:\n%U",
@@ -740,6 +747,10 @@ tcp_test_fifo1 (vlib_main_t * vm, unformat_input_t * input)
       TCP_TEST (0, "[%d] peeked %u expected %u", j, data_buf[j], data[j]);
     }
 
+  /* Try to peek beyond the data */
+  rv = svm_fifo_peek (f, svm_fifo_max_dequeue (f), vec_len (data), data_buf);
+  TCP_TEST ((rv == 0), "peeked %u expected 0", rv);
+
   vec_free (data_buf);
   svm_fifo_free (f);
   vec_free (test_data);
@@ -1239,7 +1250,7 @@ tcp_test_session (vlib_main_t * vm, unformat_input_t * input)
       tc0->c_thread_index = 0;
       tc0->c_lcl_ip4.as_u32 = local.as_u32;
       tc0->c_rmt_ip4.as_u32 = remote.as_u32;
-      tc0->opt.mss = 1450;
+      tc0->rcv_opts.mss = 1450;
       tcp_connection_init_vars (tc0);
 
       TCP_EVT_DBG (TCP_EVT_OPEN, tc0);
-- 
cgit 1.2.3-korg


From f03a59ab008908f98fd7d1b187a8c0fb78b01add Mon Sep 17 00:00:00 2001
From: Florin Coras <fcoras@cisco.com>
Date: Fri, 9 Jun 2017 21:07:32 -0700
Subject: Overall tcp performance improvements (VPP-846)

- limit minimum rto per connection
- cleanup sack scoreboard
- switched svm fifo out-of-order data handling from absolute offsets to
  relative offsets.
- improve cwnd handling when using sacks
- add cc event debug stats
- improved uri tcp test client/server:  bugfixes and added half-duplex mode
- expanded builtin client/server
- updated uri socket client/server code to work in half-duplex
- ensure session node unsets fifo event for empty fifo
- fix session detach

Change-Id: Ia446972340e32a65e0694ee2844355167d0c170d
Signed-off-by: Florin Coras <fcoras@cisco.com>
---
 src/svm/svm_fifo.c                 | 152 +++++++++--------
 src/svm/svm_fifo.h                 |  26 ++-
 src/svm/svm_fifo_segment.c         |   7 +-
 src/uri/uri_socket_server.c        |  25 ++-
 src/uri/uri_socket_test.c          |  93 +++++++----
 src/uri/uri_tcp_test.c             | 120 ++++++++++----
 src/vnet/session/application.c     |   2 +-
 src/vnet/session/node.c            |   7 +-
 src/vnet/session/segment_manager.c |   5 +
 src/vnet/session/session.c         |  25 ++-
 src/vnet/session/session.h         |   4 +-
 src/vnet/session/session_api.c     |   2 +-
 src/vnet/session/transport.h       |   2 +
 src/vnet/tcp/builtin_client.c      | 325 +++++++++++++++++++------------------
 src/vnet/tcp/builtin_client.h      | 107 +++++-------
 src/vnet/tcp/builtin_server.c      | 114 ++++++++++---
 src/vnet/tcp/tcp.c                 |  22 ++-
 src/vnet/tcp/tcp.h                 |  12 +-
 src/vnet/tcp/tcp_debug.h           | 186 +++++++++++++++------
 src/vnet/tcp/tcp_input.c           | 193 ++++++++++++----------
 src/vnet/tcp/tcp_newreno.c         |   4 +-
 src/vnet/tcp/tcp_output.c          |  10 +-
 src/vnet/tcp/tcp_test.c            |  70 ++++++--
 23 files changed, 945 insertions(+), 568 deletions(-)

(limited to 'src/svm/svm_fifo.c')

diff --git a/src/svm/svm_fifo.c b/src/svm/svm_fifo.c
index 5c8f244a..6ca437cf 100644
--- a/src/svm/svm_fifo.c
+++ b/src/svm/svm_fifo.c
@@ -15,10 +15,39 @@
 
 #include <svm/svm_fifo.h>
 
-#define offset_lt(_a, _b) ((i32)((_a)-(_b)) < 0)
-#define offset_leq(_a, _b) ((i32)((_a)-(_b)) <= 0)
-#define offset_gt(_a, _b) ((i32)((_a)-(_b)) > 0)
-#define offset_geq(_a, _b) ((i32)((_a)-(_b)) >= 0)
+static inline u8
+position_lt (svm_fifo_t * f, u32 a, u32 b)
+{
+  return (ooo_segment_distance_to_tail (f, a)
+	  < ooo_segment_distance_to_tail (f, b));
+}
+
+static inline u8
+position_leq (svm_fifo_t * f, u32 a, u32 b)
+{
+  return (ooo_segment_distance_to_tail (f, a)
+	  <= ooo_segment_distance_to_tail (f, b));
+}
+
+static inline u8
+position_gt (svm_fifo_t * f, u32 a, u32 b)
+{
+  return (ooo_segment_distance_to_tail (f, a)
+	  > ooo_segment_distance_to_tail (f, b));
+}
+
+static inline u32
+position_diff (svm_fifo_t * f, u32 posa, u32 posb)
+{
+  return ooo_segment_distance_to_tail (f, posa)
+    - ooo_segment_distance_to_tail (f, posb);
+}
+
+static inline u32
+ooo_segment_end_pos (svm_fifo_t * f, ooo_segment_t * s)
+{
+  return (s->start + s->length) % f->nitems;
+}
 
 u8 *
 format_ooo_segment (u8 * s, va_list * args)
@@ -145,13 +174,17 @@ static void
 ooo_segment_add (svm_fifo_t * f, u32 offset, u32 length)
 {
   ooo_segment_t *s, *new_s, *prev, *next, *it;
-  u32 new_index, end_offset, s_sof, s_eof, s_index;
+  u32 new_index, s_end_pos, s_index;
+  u32 normalized_position, normalized_end_position;
+
+  normalized_position = (f->tail + offset) % f->nitems;
+  normalized_end_position = (f->tail + offset + length) % f->nitems;
 
-  end_offset = offset + length;
+  f->ooos_newest = OOO_SEGMENT_INVALID_INDEX;
 
   if (f->ooos_list_head == OOO_SEGMENT_INVALID_INDEX)
     {
-      s = ooo_segment_new (f, offset, length);
+      s = ooo_segment_new (f, normalized_position, length);
       f->ooos_list_head = s - f->ooo_segments;
       f->ooos_newest = f->ooos_list_head;
       return;
@@ -160,28 +193,26 @@ ooo_segment_add (svm_fifo_t * f, u32 offset, u32 length)
   /* Find first segment that starts after new segment */
   s = pool_elt_at_index (f->ooo_segments, f->ooos_list_head);
   while (s->next != OOO_SEGMENT_INVALID_INDEX
-	 && offset_leq (ooo_segment_offset (f, s), offset))
+	 && position_lt (f, s->start, normalized_position))
     s = pool_elt_at_index (f->ooo_segments, s->next);
 
   /* If we have a previous and we overlap it, use it as starting point */
   prev = ooo_segment_get_prev (f, s);
-  if (prev && offset_leq (offset, ooo_segment_end_offset (f, prev)))
+  if (prev
+      && position_leq (f, normalized_position, ooo_segment_end_pos (f, prev)))
     {
       s = prev;
-      prev = ooo_segment_get_prev (f, s);
-      s_sof = ooo_segment_offset (f, s);
-      s_eof = ooo_segment_end_offset (f, s);
+      s_end_pos = ooo_segment_end_pos (f, s);
       goto merge;
     }
 
   s_index = s - f->ooo_segments;
-  s_sof = ooo_segment_offset (f, s);
-  s_eof = ooo_segment_end_offset (f, s);
+  s_end_pos = ooo_segment_end_pos (f, s);
 
   /* No overlap, add before current segment */
-  if (offset_lt (end_offset, s_sof))
+  if (position_lt (f, normalized_end_position, s->start))
     {
-      new_s = ooo_segment_new (f, offset, length);
+      new_s = ooo_segment_new (f, normalized_position, length);
       new_index = new_s - f->ooo_segments;
 
       /* Pool might've moved, get segment again */
@@ -198,28 +229,23 @@ ooo_segment_add (svm_fifo_t * f, u32 offset, u32 length)
 	  f->ooos_list_head = new_index;
 	}
 
-      new_s->next = s - f->ooo_segments;
+      new_s->next = s_index;
       s->prev = new_index;
       f->ooos_newest = new_index;
       return;
     }
   /* No overlap, add after current segment */
-  else if (offset_gt (offset, s_eof))
+  else if (position_gt (f, normalized_position, s_end_pos))
     {
-      new_s = ooo_segment_new (f, offset, length);
+      new_s = ooo_segment_new (f, normalized_position, length);
       new_index = new_s - f->ooo_segments;
 
       /* Pool might've moved, get segment again */
       s = pool_elt_at_index (f->ooo_segments, s_index);
 
-      if (s->next != OOO_SEGMENT_INVALID_INDEX)
-	{
-	  new_s->next = s->next;
-	  next = pool_elt_at_index (f->ooo_segments, new_s->next);
-	  next->prev = new_index;
-	}
+      ASSERT (s->next == OOO_SEGMENT_INVALID_INDEX);
 
-      new_s->prev = s - f->ooo_segments;
+      new_s->prev = s_index;
       s->next = new_index;
       f->ooos_newest = new_index;
 
@@ -233,30 +259,32 @@ ooo_segment_add (svm_fifo_t * f, u32 offset, u32 length)
 merge:
 
   /* Merge at head */
-  if (offset_lt (offset, s_sof))
+  if (position_lt (f, normalized_position, s->start))
     {
-      s->start = offset;
-      s->length = s_eof - ooo_segment_offset (f, s);
+      s->start = normalized_position;
+      s->length = position_diff (f, s_end_pos, s->start);
     }
-  /* Last but overlapping previous */
-  else if (offset_gt (end_offset, s_eof))
+  /* Overlapping tail */
+  else if (position_gt (f, normalized_end_position, s_end_pos))
     {
-      s->length = end_offset - ooo_segment_offset (f, s);
+      s->length = position_diff (f, normalized_end_position, s->start);
     }
   /* New segment completely covered by current one */
   else
     {
       /* Do Nothing */
+      s = 0;
       goto done;
     }
 
   /* The new segment's tail may cover multiple smaller ones */
-  if (offset_geq (end_offset, s_eof))
+  if (position_gt (f, normalized_end_position, s_end_pos))
     {
       /* Remove the completely overlapped segments */
       it = (s->next != OOO_SEGMENT_INVALID_INDEX) ?
 	pool_elt_at_index (f->ooo_segments, s->next) : 0;
-      while (it && offset_leq (ooo_segment_end_offset (f, it), end_offset))
+      while (it && position_leq (f, ooo_segment_end_pos (f, it),
+				 normalized_end_position))
 	{
 	  next = (it->next != OOO_SEGMENT_INVALID_INDEX) ?
 	    pool_elt_at_index (f->ooo_segments, it->next) : 0;
@@ -265,17 +293,17 @@ merge:
 	}
 
       /* If partial overlap with last, merge */
-      if (it && offset_leq (ooo_segment_offset (f, it), end_offset))
+      if (it && position_leq (f, it->start, normalized_end_position))
 	{
-	  s->length = ooo_segment_end_offset (f, it) -
-	    ooo_segment_offset (f, s);
+	  s->length = ooo_segment_end_pos (f, it) - s->start;
 	  ooo_segment_del (f, it - f->ooo_segments);
 	}
     }
 
 done:
   /* Most recently updated segment */
-  f->ooos_newest = s - f->ooo_segments;
+  if (s)
+    f->ooos_newest = s - f->ooo_segments;
 }
 
 /**
@@ -286,32 +314,28 @@ static int
 ooo_segment_try_collect (svm_fifo_t * f, u32 n_bytes_enqueued)
 {
   ooo_segment_t *s;
-  u32 index, bytes = 0, diff;
-  u32 cursize, norm_start, nitems;
-
-  /* current size has not yet been updated */
-  cursize = svm_fifo_max_dequeue (f) + n_bytes_enqueued;
-  nitems = f->nitems;
+  u32 index, bytes = 0;
+  i32 diff;
 
   s = pool_elt_at_index (f->ooo_segments, f->ooos_list_head);
 
-  norm_start = s->start % nitems;
-  diff = (f->nitems + (i32) (f->tail - norm_start)) % nitems;
+  diff = (f->tail >= s->start) ?
+    f->tail - s->start : f->nitems + f->tail - s->start;
 
-  if (diff > cursize)
+  if (diff > n_bytes_enqueued)
     return 0;
 
   /* If last tail update overlaps one/multiple ooo segments, remove them */
-  while (0 < diff && diff < cursize)
+  while (0 <= diff && diff < n_bytes_enqueued)
     {
       index = s - f->ooo_segments;
 
       /* Segment end is beyond the tail. Advance tail and remove segment */
-      if (diff < s->length)
+      if (s->length > diff)
 	{
-	  f->tail += s->length - diff;
-	  f->tail %= f->nitems;
 	  bytes = s->length - diff;
+	  f->tail += bytes;
+	  f->tail %= f->nitems;
 	  ooo_segment_del (f, index);
 	  break;
 	}
@@ -320,8 +344,8 @@ ooo_segment_try_collect (svm_fifo_t * f, u32 n_bytes_enqueued)
       if (s->next != OOO_SEGMENT_INVALID_INDEX)
 	{
 	  s = pool_elt_at_index (f->ooo_segments, s->next);
-	  norm_start = s->start % nitems;
-	  diff = (f->nitems + (i32) (f->tail - norm_start)) % nitems;
+	  diff = (f->tail >= s->start) ?
+	    f->tail - s->start : f->nitems + f->tail - s->start;
 	  ooo_segment_del (f, index);
 	}
       /* End of search */
@@ -332,18 +356,6 @@ ooo_segment_try_collect (svm_fifo_t * f, u32 n_bytes_enqueued)
 	}
     }
 
-  /* If tail is adjacent to an ooo segment, 'consume' it */
-  if (diff == 0)
-    {
-      bytes = ((nitems - cursize) >= s->length) ? s->length :
-	nitems - cursize;
-
-      f->tail += bytes;
-      f->tail %= nitems;
-
-      ooo_segment_del (f, s - f->ooo_segments);
-    }
-
   return bytes;
 }
 
@@ -355,6 +367,7 @@ svm_fifo_enqueue_internal (svm_fifo_t * f, u32 max_bytes, u8 * copy_from_here)
 
   /* read cursize, which can only increase while we're working */
   cursize = svm_fifo_max_dequeue (f);
+  f->ooos_newest = OOO_SEGMENT_INVALID_INDEX;
 
   if (PREDICT_FALSE (cursize == f->nitems))
     return -2;			/* fifo stuffed */
@@ -424,13 +437,16 @@ svm_fifo_enqueue_with_offset_internal (svm_fifo_t * f,
 				       u8 * copy_from_here)
 {
   u32 total_copy_bytes, first_copy_bytes, second_copy_bytes;
-  u32 cursize, nitems;
-  u32 normalized_offset, offset_from_tail;
+  u32 cursize, nitems, normalized_offset;
+  u32 offset_from_tail;
+
+  f->ooos_newest = OOO_SEGMENT_INVALID_INDEX;
 
   /* read cursize, which can only increase while we're working */
   cursize = svm_fifo_max_dequeue (f);
   nitems = f->nitems;
-  normalized_offset = offset % nitems;
+
+  normalized_offset = (f->tail + offset) % nitems;
 
   /* Will this request fit? */
   offset_from_tail = (nitems + normalized_offset - f->tail) % nitems;
diff --git a/src/svm/svm_fifo.h b/src/svm/svm_fifo.h
index 9cb93ff4..f32ef41d 100644
--- a/src/svm/svm_fifo.h
+++ b/src/svm/svm_fifo.h
@@ -127,21 +127,37 @@ format_function_t format_svm_fifo;
 always_inline ooo_segment_t *
 svm_fifo_newest_ooo_segment (svm_fifo_t * f)
 {
-  return f->ooo_segments + f->ooos_newest;
+  if (f->ooos_newest == OOO_SEGMENT_INVALID_INDEX)
+    return 0;
+  return pool_elt_at_index (f->ooo_segments, f->ooos_newest);
+}
+
+always_inline u32
+ooo_segment_distance_to_tail (svm_fifo_t * f, u32 a)
+{
+  /* Ambiguous. Assumption is that ooo segments don't touch tail */
+  if (a == f->tail && f->tail == f->head)
+    return f->nitems;
+
+  return ((f->nitems + a - f->tail) % f->nitems);
 }
 
 always_inline u32
 ooo_segment_offset (svm_fifo_t * f, ooo_segment_t * s)
 {
-//  return ((f->nitems + s->fifo_position - f->tail) % f->nitems);
-  return s->start;
+  return ooo_segment_distance_to_tail (f, s->start);
 }
 
 always_inline u32
 ooo_segment_end_offset (svm_fifo_t * f, ooo_segment_t * s)
 {
-//  return ((f->nitems + s->fifo_position + s->length - f->tail) % f->nitems);
-  return s->start + s->length;
+  return ooo_segment_distance_to_tail (f, s->start) + s->length;
+}
+
+always_inline u32
+ooo_segment_length (svm_fifo_t * f, ooo_segment_t * s)
+{
+  return s->length;
 }
 
 always_inline ooo_segment_t *
diff --git a/src/svm/svm_fifo_segment.c b/src/svm/svm_fifo_segment.c
index eef2168c..c4ac2352 100644
--- a/src/svm/svm_fifo_segment.c
+++ b/src/svm/svm_fifo_segment.c
@@ -305,14 +305,17 @@ svm_fifo_segment_free_fifo (svm_fifo_segment_private_t * s, svm_fifo_t * f,
       /* Remove from active list */
       if (f->prev)
 	f->prev->next = f->next;
+      else
+	fsh->fifos = f->next;
       if (f->next)
 	f->next->prev = f->prev;
-      /* FALLTHROUGH */
+      /* Fall through: we add only rx fifos to active pool */
     case FIFO_SEGMENT_TX_FREELIST:
       /* Add to free list */
       f->next = fsh->free_fifos[list_index];
+      f->prev = 0;
       fsh->free_fifos[list_index] = f;
-      /* FALLTHROUGH */
+      break;
     case FIFO_SEGMENT_FREELIST_NONE:
       break;
 
diff --git a/src/uri/uri_socket_server.c b/src/uri/uri_socket_server.c
index 2366f420..4f4c5f30 100644
--- a/src/uri/uri_socket_server.c
+++ b/src/uri/uri_socket_server.c
@@ -22,6 +22,7 @@
 #include <vppinfra/format.h>
 #include <signal.h>
 #include <sys/ucontext.h>
+#include <sys/time.h>
 
 volatile int signal_received;
 
@@ -78,7 +79,10 @@ main (int argc, char *argv[])
   struct sockaddr_in serv_addr;
   struct sockaddr_in client;
   struct hostent *server;
-  u8 *rx_buffer = 0;
+  u8 *rx_buffer = 0, no_echo = 0;
+  struct timeval start, end;
+  long rcvd = 0;
+  double deltat;
 
   if (argc > 1 && argc < 3)
     {
@@ -86,8 +90,9 @@ main (int argc, char *argv[])
       exit (0);
     }
 
-  if (argc >= 3)
+  if (argc >= 4)
     {
+      no_echo = atoi (argv[3]);
       portno = atoi (argv[2]);
       server = gethostbyname (argv[1]);
       if (server == NULL)
@@ -137,7 +142,7 @@ main (int argc, char *argv[])
       exit (1);
     }
 
-  vec_validate (rx_buffer, 8999 /* jumbo mtu */ );
+  vec_validate (rx_buffer, 128 << 10);
 
   if (listen (sockfd, 5 /* backlog */ ) < 0)
     {
@@ -160,6 +165,8 @@ main (int argc, char *argv[])
 	}
       fformat (stderr, "Accepted connection from: %s : %d\n",
 	       inet_ntoa (client.sin_addr), client.sin_port);
+      gettimeofday (&start, NULL);
+
       while (1)
 	{
 	  n = recv (accfd, rx_buffer, vec_len (rx_buffer), 0 /* flags */ );
@@ -167,6 +174,14 @@ main (int argc, char *argv[])
 	    {
 	      /* Graceful exit */
 	      close (accfd);
+	      gettimeofday (&end, NULL);
+	      deltat = (end.tv_sec - start.tv_sec);
+	      deltat += (end.tv_usec - start.tv_usec) / 1000000.0;
+	      clib_warning ("Finished in %.6f", deltat);
+	      clib_warning ("%.4f Gbit/second %s",
+			    (((f64) rcvd * 8.0) / deltat / 1e9),
+			    no_echo ? "half" : "full");
+	      rcvd = 0;
 	      break;
 	    }
 	  if (n < 0)
@@ -179,6 +194,10 @@ main (int argc, char *argv[])
 	  if (signal_received)
 	    break;
 
+	  rcvd += n;
+	  if (no_echo)
+	    continue;
+
 	  sent = send (accfd, rx_buffer, n, 0 /* flags */ );
 	  if (n < 0)
 	    {
diff --git a/src/uri/uri_socket_test.c b/src/uri/uri_socket_test.c
index 9f049bda..5f7084d5 100644
--- a/src/uri/uri_socket_test.c
+++ b/src/uri/uri_socket_test.c
@@ -19,6 +19,7 @@
 #include <netinet/in.h>
 #include <netdb.h>
 #include <vppinfra/format.h>
+#include <sys/time.h>
 
 int
 main (int argc, char *argv[])
@@ -26,28 +27,44 @@ main (int argc, char *argv[])
   int sockfd, portno, n;
   struct sockaddr_in serv_addr;
   struct hostent *server;
-  u8 *rx_buffer = 0, *tx_buffer = 0;
+  u8 *rx_buffer = 0, *tx_buffer = 0, no_echo = 0, test_bytes = 0;
   u32 offset;
-  int iter, i;
-  if (0 && argc < 3)
+  long bytes = 1 << 20, to_send;
+  int i;
+  struct timeval start, end;
+  double deltat;
+
+  if (argc >= 3)
     {
-      fformat (stderr, "usage %s hostname port\n", argv[0]);
-      exit (0);
+      bytes = ((long) atoi (argv[4])) << 20;
+      no_echo = atoi (argv[3]);
+      portno = atoi (argv[2]);
+      server = gethostbyname (argv[1]);
+      if (server == NULL)
+	{
+	  clib_unix_warning ("gethostbyname");
+	  exit (1);
+	}
+    }
+  else
+    {
+      portno = 1234;		// atoi(argv[2]);
+      server = gethostbyname ("6.0.1.1" /* argv[1] */ );
+      if (server == NULL)
+	{
+	  clib_unix_warning ("gethostbyname");
+	  exit (1);
+	}
     }
 
-  portno = 1234;		// atoi(argv[2]);
+  to_send = bytes;
   sockfd = socket (AF_INET, SOCK_STREAM, 0);
   if (sockfd < 0)
     {
       clib_unix_error ("socket");
       exit (1);
     }
-  server = gethostbyname ("6.0.1.1" /* argv[1] */ );
-  if (server == NULL)
-    {
-      clib_unix_warning ("gethostbyname");
-      exit (1);
-    }
+
   bzero ((char *) &serv_addr, sizeof (serv_addr));
   serv_addr.sin_family = AF_INET;
   bcopy ((char *) server->h_addr,
@@ -59,8 +76,8 @@ main (int argc, char *argv[])
       exit (1);
     }
 
-  vec_validate (rx_buffer, 1400);
-  vec_validate (tx_buffer, 1400);
+  vec_validate (rx_buffer, 128 << 10);
+  vec_validate (tx_buffer, 128 << 10);
 
   for (i = 0; i < vec_len (tx_buffer); i++)
     tx_buffer[i] = (i + 1) % 0xff;
@@ -75,19 +92,28 @@ main (int argc, char *argv[])
       exit (0);
     }
 
-  for (iter = 0; iter < 100000; iter++)
+  gettimeofday (&start, NULL);
+  while (bytes > 0)
     {
-      if (iter < 99999)
+      /*
+       * TX
+       */
+      n = send (sockfd, tx_buffer, vec_len (tx_buffer), 0 /* flags */ );
+      if (n != vec_len (tx_buffer))
 	{
-	  n = send (sockfd, tx_buffer, vec_len (tx_buffer), 0 /* flags */ );
-	  if (n != vec_len (tx_buffer))
-	    {
-	      clib_unix_warning ("write");
-	      exit (0);
-	    }
+	  clib_unix_warning ("write");
+	  exit (0);
 	}
-      offset = 0;
+      bytes -= n;
 
+      if (no_echo)
+	continue;
+
+      /*
+       * RX
+       */
+
+      offset = 0;
       do
 	{
 	  n = recv (sockfd, rx_buffer + offset,
@@ -101,18 +127,27 @@ main (int argc, char *argv[])
 	}
       while (offset < vec_len (rx_buffer));
 
-      for (i = 0; i < vec_len (rx_buffer); i++)
+      if (test_bytes)
 	{
-	  if (rx_buffer[i] != tx_buffer[i])
+	  for (i = 0; i < vec_len (rx_buffer); i++)
 	    {
-	      clib_warning ("[%d] read 0x%x not 0x%x",
-			    rx_buffer[i], tx_buffer[i]);
-	      exit (1);
+	      if (rx_buffer[i] != tx_buffer[i])
+		{
+		  clib_warning ("[%d] read 0x%x not 0x%x", rx_buffer[i],
+				tx_buffer[i]);
+		  exit (1);
+		}
 	    }
 	}
-
     }
   close (sockfd);
+  gettimeofday (&end, NULL);
+
+  deltat = (end.tv_sec - start.tv_sec);
+  deltat += (end.tv_usec - start.tv_usec) / 1000000.0;	// us to ms
+  clib_warning ("Finished in %.6f", deltat);
+  clib_warning ("%.4f Gbit/second %s", (((f64) to_send * 8.0) / deltat / 1e9),
+		no_echo ? "half" : "full");
   return 0;
 }
 
diff --git a/src/uri/uri_tcp_test.c b/src/uri/uri_tcp_test.c
index e201a359..d1694cf4 100755
--- a/src/uri/uri_tcp_test.c
+++ b/src/uri/uri_tcp_test.c
@@ -46,6 +46,8 @@ typedef struct
   svm_fifo_t *server_tx_fifo;
 
   u64 vpp_session_handle;
+  u64 bytes_received;
+  f64 start;
 } session_t;
 
 typedef enum
@@ -174,7 +176,7 @@ wait_for_state_change (uri_tcp_test_main_t * utm, connection_state_t state)
       if (utm->state == STATE_FAILED)
 	return -1;
       if (utm->time_to_stop == 1)
-	return -1;
+	return 0;
     }
   clib_warning ("timeout waiting for STATE_READY");
   return -1;
@@ -184,7 +186,7 @@ void
 application_send_attach (uri_tcp_test_main_t * utm)
 {
   vl_api_application_attach_t *bmp;
-  u32 fifo_size = 3 << 20;
+  u32 fifo_size = 4 << 20;
   bmp = vl_msg_api_alloc (sizeof (*bmp));
   memset (bmp, 0, sizeof (*bmp));
 
@@ -343,11 +345,23 @@ vl_api_map_another_segment_t_handler (vl_api_map_another_segment_t * mp)
 		mp->segment_size);
 }
 
+static void
+session_print_stats (uri_tcp_test_main_t * utm, session_t * session)
+{
+  f64 deltat;
+  u64 bytes;
+
+  deltat = clib_time_now (&utm->clib_time) - session->start;
+  bytes = utm->i_am_master ? session->bytes_received : utm->bytes_to_send;
+  fformat (stdout, "Finished in %.6f\n", deltat);
+  fformat (stdout, "%.4f Gbit/second\n", (bytes * 8.0) / deltat / 1e9);
+}
+
 static void
 vl_api_disconnect_session_t_handler (vl_api_disconnect_session_t * mp)
 {
   uri_tcp_test_main_t *utm = &uri_tcp_test_main;
-  session_t *session;
+  session_t *session = 0;
   vl_api_disconnect_session_reply_t *rmp;
   uword *p;
   int rv = 0;
@@ -366,7 +380,7 @@ vl_api_disconnect_session_t_handler (vl_api_disconnect_session_t * mp)
       rv = -11;
     }
 
-  utm->time_to_stop = 1;
+//  utm->time_to_stop = 1;
 
   rmp = vl_msg_api_alloc (sizeof (*rmp));
   memset (rmp, 0, sizeof (*rmp));
@@ -375,6 +389,9 @@ vl_api_disconnect_session_t_handler (vl_api_disconnect_session_t * mp)
   rmp->retval = rv;
   rmp->handle = mp->handle;
   vl_msg_api_send_shmem (utm->vl_input_queue, (u8 *) & rmp);
+
+  if (session)
+    session_print_stats (utm, session);
 }
 
 static void
@@ -431,14 +448,19 @@ client_handle_fifo_event_rx (uri_tcp_test_main_t * utm,
       if (n_read > 0)
 	{
 	  bytes -= n_read;
-	  for (i = 0; i < n_read; i++)
+	  if (utm->test_return_packets)
 	    {
-	      if (utm->rx_buf[i] != ((utm->client_bytes_received + i) & 0xff))
+	      for (i = 0; i < n_read; i++)
 		{
-		  clib_warning ("error at byte %lld, 0x%x not 0x%x",
-				utm->client_bytes_received + i,
-				utm->rx_buf[i],
-				((utm->client_bytes_received + i) & 0xff));
+		  if (utm->rx_buf[i]
+		      != ((utm->client_bytes_received + i) & 0xff))
+		    {
+		      clib_warning ("error at byte %lld, 0x%x not 0x%x",
+				    utm->client_bytes_received + i,
+				    utm->rx_buf[i],
+				    ((utm->client_bytes_received +
+				      i) & 0xff));
+		    }
 		}
 	    }
 	  utm->client_bytes_received += n_read;
@@ -545,6 +567,7 @@ vl_api_connect_uri_reply_t_handler (vl_api_connect_uri_reply_t * mp)
   session->server_rx_fifo = rx_fifo;
   session->server_tx_fifo = tx_fifo;
   session->vpp_session_handle = mp->handle;
+  session->start = clib_time_now (&utm->clib_time);
 
   /* Save handle */
   utm->connected_session_index = session_index;
@@ -571,7 +594,7 @@ send_test_chunk (uri_tcp_test_main_t * utm, svm_fifo_t * tx_fifo, int mypid,
   u64 bytes_sent = 0;
   int test_buf_offset = 0;
   u32 bytes_to_snd;
-  u32 queue_max_chunk = 64 << 10, actual_write;
+  u32 queue_max_chunk = 128 << 10, actual_write;
   session_fifo_event_t evt;
   static int serial_number = 0;
   int rv;
@@ -582,8 +605,8 @@ send_test_chunk (uri_tcp_test_main_t * utm, svm_fifo_t * tx_fifo, int mypid,
 
   while (bytes_to_snd > 0)
     {
-      actual_write =
-	bytes_to_snd > queue_max_chunk ? queue_max_chunk : bytes_to_snd;
+      actual_write = (bytes_to_snd > queue_max_chunk) ?
+	queue_max_chunk : bytes_to_snd;
       rv = svm_fifo_enqueue_nowait (tx_fifo, actual_write,
 				    test_data + test_buf_offset);
 
@@ -635,9 +658,9 @@ client_send_data (uri_tcp_test_main_t * utm)
   if (leftover)
     send_test_chunk (utm, tx_fifo, mypid, leftover);
 
-  if (utm->test_return_packets)
+  if (!utm->drop_packets)
     {
-      f64 timeout = clib_time_now (&utm->clib_time) + 2;
+      f64 timeout = clib_time_now (&utm->clib_time) + 10;
 
       /* Wait for the outstanding packets */
       while (utm->client_bytes_received <
@@ -698,6 +721,7 @@ int
 client_disconnect (uri_tcp_test_main_t * utm)
 {
   client_send_disconnect (utm);
+  clib_warning ("Sent disconnect");
   if (wait_for_state_change (utm, STATE_START))
     {
       clib_warning ("Disconnect failed");
@@ -721,7 +745,7 @@ client_test (uri_tcp_test_main_t * utm)
     }
 
   /* Init test data */
-  vec_validate (utm->connect_test_data, 64 * 1024 - 1);
+  vec_validate (utm->connect_test_data, 128 * 1024 - 1);
   for (i = 0; i < vec_len (utm->connect_test_data); i++)
     utm->connect_test_data[i] = i & 0xff;
 
@@ -899,6 +923,9 @@ vl_api_accept_session_t_handler (vl_api_accept_session_t * mp)
   rmp->_vl_msg_id = ntohs (VL_API_ACCEPT_SESSION_REPLY);
   rmp->handle = mp->handle;
   vl_msg_api_send_shmem (utm->vl_input_queue, (u8 *) & rmp);
+
+  session->bytes_received = 0;
+  session->start = clib_time_now (&utm->clib_time);
 }
 
 void
@@ -909,37 +936,50 @@ server_handle_fifo_event_rx (uri_tcp_test_main_t * utm,
   int n_read;
   session_fifo_event_t evt;
   unix_shared_memory_queue_t *q;
-  int rv, bytes;
+  session_t *session;
+  int rv;
+  u32 max_dequeue, offset, max_transfer, rx_buf_len;
 
+  rx_buf_len = vec_len (utm->rx_buf);
   rx_fifo = e->fifo;
-  tx_fifo = utm->sessions[rx_fifo->client_session_index].server_tx_fifo;
+  session = &utm->sessions[rx_fifo->client_session_index];
+  tx_fifo = session->server_tx_fifo;
 
-  bytes = svm_fifo_max_dequeue (rx_fifo);
+  max_dequeue = svm_fifo_max_dequeue (rx_fifo);
   /* Allow enqueuing of a new event */
   svm_fifo_unset_event (rx_fifo);
 
-  if (bytes == 0)
-    return;
+  if (PREDICT_FALSE (max_dequeue == 0))
+    {
+      return;
+    }
 
-  /* Read the bytes */
+  /* Read the max_dequeue */
   do
     {
-      n_read = svm_fifo_dequeue_nowait (rx_fifo, vec_len (utm->rx_buf),
-					utm->rx_buf);
+      max_transfer = clib_min (rx_buf_len, max_dequeue);
+      n_read = svm_fifo_dequeue_nowait (rx_fifo, max_transfer, utm->rx_buf);
       if (n_read > 0)
-	bytes -= n_read;
-
-      if (utm->drop_packets)
-	continue;
+	{
+	  max_dequeue -= n_read;
+	  session->bytes_received += n_read;
+	}
 
       /* Reflect if a non-drop session */
-      if (n_read > 0)
+      if (!utm->drop_packets && n_read > 0)
 	{
+	  offset = 0;
 	  do
 	    {
-	      rv = svm_fifo_enqueue_nowait (tx_fifo, n_read, utm->rx_buf);
+	      rv = svm_fifo_enqueue_nowait (tx_fifo, n_read,
+					    &utm->rx_buf[offset]);
+	      if (rv > 0)
+		{
+		  n_read -= rv;
+		  offset += rv;
+		}
 	    }
-	  while (rv <= 0 && !utm->time_to_stop);
+	  while ((rv <= 0 || n_read > 0) && !utm->time_to_stop);
 
 	  /* If event wasn't set, add one */
 	  if (svm_fifo_set_event (tx_fifo))
@@ -951,11 +991,11 @@ server_handle_fifo_event_rx (uri_tcp_test_main_t * utm,
 
 	      q = utm->vpp_event_queue;
 	      unix_shared_memory_queue_add (q, (u8 *) & evt,
-					    0 /* do wait for mutex */ );
+					    1 /* do wait for mutex */ );
 	    }
 	}
     }
-  while ((n_read < 0 || bytes > 0) && !utm->time_to_stop);
+  while ((n_read < 0 || max_dequeue > 0) && !utm->time_to_stop);
 }
 
 void
@@ -1068,9 +1108,18 @@ vl_api_disconnect_session_reply_t_handler (vl_api_disconnect_session_reply_t *
 					   mp)
 {
   uri_tcp_test_main_t *utm = &uri_tcp_test_main;
+  session_t *session;
+
+  if (mp->retval)
+    {
+      clib_warning ("vpp complained about disconnect: %d",
+		    ntohl (mp->retval));
+    }
 
-  clib_warning ("retval %d", ntohl (mp->retval));
   utm->state = STATE_START;
+  session = pool_elt_at_index (utm->sessions, utm->connected_session_index);
+  if (session)
+    session_print_stats (utm, session);
 }
 
 #define foreach_uri_msg                                 \
@@ -1123,7 +1172,7 @@ main (int argc, char **argv)
   /* make the main heap thread-safe */
   h->flags |= MHEAP_FLAG_THREAD_SAFE;
 
-  vec_validate (utm->rx_buf, 65536);
+  vec_validate (utm->rx_buf, 128 << 10);
 
   utm->session_index_by_vpp_handles = hash_create (0, sizeof (uword));
 
@@ -1186,6 +1235,7 @@ main (int argc, char **argv)
   utm->drop_packets = drop_packets;
   utm->test_return_packets = test_return_packets;
   utm->bytes_to_send = bytes_to_send;
+  utm->time_to_stop = 0;
 
   setup_signal_handlers ();
   uri_api_hookup (utm);
diff --git a/src/vnet/session/application.c b/src/vnet/session/application.c
index c679b1f5..4bdb1027 100644
--- a/src/vnet/session/application.c
+++ b/src/vnet/session/application.c
@@ -117,7 +117,7 @@ application_del (application_t * app)
   /* Actual listener cleanup */
   for (i = 0; i < vec_len (handles); i++)
     {
-      a->app_index = app->api_client_index;
+      a->app_index = app->index;
       a->handle = handles[i];
       /* seg manager is removed when unbind completes */
       vnet_unbind (a);
diff --git a/src/vnet/session/node.c b/src/vnet/session/node.c
index 07eeae82..c0ab1bf0 100644
--- a/src/vnet/session/node.c
+++ b/src/vnet/session/node.c
@@ -171,7 +171,10 @@ session_tx_fifo_read_and_snd_i (vlib_main_t * vm, vlib_node_runtime_t * node,
 
   /* Nothing to read return */
   if (max_dequeue0 == 0)
-    return 0;
+    {
+      svm_fifo_unset_event (s0->server_tx_fifo);
+      return 0;
+    }
 
   /* Ensure we're not writing more than transport window allows */
   if (max_dequeue0 < snd_space0)
@@ -393,7 +396,7 @@ session_event_get_session (session_fifo_event_t * e0, u8 thread_index)
 
   s0 = stream_session_get_if_valid (session_index0, thread_index);
 
-  ASSERT (s0->thread_index == thread_index);
+  ASSERT (s0 == 0 || s0->thread_index == thread_index);
 
   return s0;
 }
diff --git a/src/vnet/session/segment_manager.c b/src/vnet/session/segment_manager.c
index caf8eaa3..bf571963 100644
--- a/src/vnet/session/segment_manager.c
+++ b/src/vnet/session/segment_manager.c
@@ -306,11 +306,13 @@ again:
 	  if (added_a_segment)
 	    {
 	      clib_warning ("added a segment, still cant allocate a fifo");
+	      clib_spinlock_unlock (&sm->lockp);
 	      return SESSION_ERROR_NEW_SEG_NO_SPACE;
 	    }
 
 	  if (session_manager_add_segment (sm))
 	    {
+	      clib_spinlock_unlock (&sm->lockp);
 	      return VNET_API_ERROR_URI_FIFO_CREATE_FAILED;
 	    }
 
@@ -320,6 +322,7 @@ again:
       else
 	{
 	  clib_warning ("No space to allocate fifos!");
+	  clib_spinlock_unlock (&sm->lockp);
 	  return SESSION_ERROR_NO_SPACE;
 	}
     }
@@ -361,8 +364,10 @@ segment_manager_dealloc_fifos (u32 svm_segment_index, svm_fifo_t * rx_fifo,
   if (sm->segment_indices[0] != svm_segment_index
       && !svm_fifo_segment_has_fifos (fifo_segment))
     {
+      clib_spinlock_lock (&sm->lockp);
       svm_fifo_segment_delete (fifo_segment);
       vec_del1 (sm->segment_indices, svm_segment_index);
+      clib_spinlock_unlock (&sm->lockp);
     }
 }
 
diff --git a/src/vnet/session/session.c b/src/vnet/session/session.c
index 534598d6..fe198044 100644
--- a/src/vnet/session/session.c
+++ b/src/vnet/session/session.c
@@ -700,7 +700,7 @@ stream_session_init_fifos_pointers (transport_connection_t * tc,
   svm_fifo_init_pointers (s->server_tx_fifo, tx_pointer);
 }
 
-void
+int
 stream_session_connect_notify (transport_connection_t * tc, u8 sst,
 			       u8 is_fail)
 {
@@ -709,6 +709,7 @@ stream_session_connect_notify (transport_connection_t * tc, u8 sst,
   stream_session_t *new_s = 0;
   u64 handle;
   u32 api_context = 0;
+  int error = 0;
 
   handle = stream_session_half_open_lookup (smm, &tc->lcl_ip, &tc->rmt_ip,
 					    tc->lcl_port, tc->rmt_port,
@@ -716,7 +717,7 @@ stream_session_connect_notify (transport_connection_t * tc, u8 sst,
   if (handle == HALF_OPEN_LOOKUP_INVALID_VALUE)
     {
       clib_warning ("This can't be good!");
-      return;
+      return -1;
     }
 
   /* Get the app's index from the handle we stored when opening connection */
@@ -730,9 +731,12 @@ stream_session_connect_notify (transport_connection_t * tc, u8 sst,
 
       /* Create new session (svm segments are allocated if needed) */
       if (stream_session_create_i (sm, tc, &new_s))
-	return;
-
-      new_s->app_index = app->index;
+	{
+	  is_fail = 1;
+	  error = -1;
+	}
+      else
+	new_s->app_index = app->index;
     }
 
   /* Notify client */
@@ -741,6 +745,8 @@ stream_session_connect_notify (transport_connection_t * tc, u8 sst,
 
   /* Cleanup session lookup */
   stream_session_half_open_table_del (smm, sst, tc);
+
+  return error;
 }
 
 void
@@ -981,8 +987,13 @@ session_send_session_evt_to_thread (u64 session_handle,
 
   /* Based on request block (or not) for lack of space */
   if (PREDICT_TRUE (q->cursize < q->maxsize))
-    unix_shared_memory_queue_add (q, (u8 *) & evt,
-				  0 /* do wait for mutex */ );
+    {
+      if (unix_shared_memory_queue_add (q, (u8 *) & evt,
+					1 /* do wait for mutex */ ))
+	{
+	  clib_warning ("failed to enqueue evt");
+	}
+    }
   else
     {
       clib_warning ("queue full");
diff --git a/src/vnet/session/session.h b/src/vnet/session/session.h
index d9c38bd1..5fa4225c 100644
--- a/src/vnet/session/session.h
+++ b/src/vnet/session/session.h
@@ -368,8 +368,8 @@ stream_session_peek_bytes (transport_connection_t * tc, u8 * buffer,
 			   u32 offset, u32 max_bytes);
 u32 stream_session_dequeue_drop (transport_connection_t * tc, u32 max_bytes);
 
-void stream_session_connect_notify (transport_connection_t * tc, u8 sst,
-				    u8 is_fail);
+int stream_session_connect_notify (transport_connection_t * tc, u8 sst,
+				   u8 is_fail);
 void stream_session_init_fifos_pointers (transport_connection_t * tc,
 					 u32 rx_pointer, u32 tx_pointer);
 
diff --git a/src/vnet/session/session_api.c b/src/vnet/session/session_api.c
index f772cb9f..60f764af 100755
--- a/src/vnet/session/session_api.c
+++ b/src/vnet/session/session_api.c
@@ -419,7 +419,7 @@ done:
   REPLY_MACRO (VL_API_UNBIND_URI_REPLY);
 }
 
-void
+static void
 vl_api_connect_uri_t_handler (vl_api_connect_uri_t * mp)
 {
   vl_api_connect_uri_reply_t *rmp;
diff --git a/src/vnet/session/transport.h b/src/vnet/session/transport.h
index e5f788be..04bd5ca0 100644
--- a/src/vnet/session/transport.h
+++ b/src/vnet/session/transport.h
@@ -39,6 +39,7 @@ typedef struct _transport_connection
 
 #if TRANSPORT_DEBUG
   elog_track_t elog_track;	/**< Event logging */
+  u32 cc_stat_tstamp;		/**< CC stats timestamp */
 #endif
 
   /** Macros for 'derived classes' where base is named "connection" */
@@ -57,6 +58,7 @@ typedef struct _transport_connection
 #define c_is_ip4 connection.is_ip4
 #define c_thread_index connection.thread_index
 #define c_elog_track connection.elog_track
+#define c_cc_stat_tstamp connection.cc_stat_tstamp
 } transport_connection_t;
 
 /*
diff --git a/src/vnet/tcp/builtin_client.c b/src/vnet/tcp/builtin_client.c
index 7238cda3..6f8be082 100644
--- a/src/vnet/tcp/builtin_client.c
+++ b/src/vnet/tcp/builtin_client.c
@@ -43,7 +43,7 @@
 #include <vpp/api/vpe_all_api_h.h>
 #undef vl_printfun
 
-#define TCP_BUILTIN_CLIENT_DBG (1)
+#define TCP_BUILTIN_CLIENT_DBG (0)
 
 static void
 send_test_chunk (tclient_main_t * tm, session_t * s)
@@ -92,7 +92,7 @@ send_test_chunk (tclient_main_t * tm, session_t * s)
 	  ed->data[2] = s->bytes_to_send;
 	}
 
-      /* Poke the TCP state machine */
+      /* Poke the session layer */
       if (svm_fifo_set_event (s->server_tx_fifo))
 	{
 	  /* Fabricate TX event, send to vpp */
@@ -100,8 +100,9 @@ send_test_chunk (tclient_main_t * tm, session_t * s)
 	  evt.event_type = FIFO_EVENT_APP_TX;
 	  evt.event_id = serial_number++;
 
-	  unix_shared_memory_queue_add (tm->vpp_event_queue, (u8 *) & evt,
-					0 /* do wait for mutex */ );
+	  if (unix_shared_memory_queue_add (tm->vpp_event_queue, (u8 *) & evt,
+					    0 /* do wait for mutex */ ))
+	    clib_warning ("could not enqueue event");
 	}
     }
 }
@@ -188,13 +189,13 @@ builtin_client_node_fn (vlib_main_t * vm, vlib_node_runtime_t * node,
 
       sp = pool_elt_at_index (tm->sessions, connection_indices[i]);
 
-      if (tx_quota < 60 && sp->bytes_to_send > 0)
+      if ((tm->no_return || tx_quota < 60) && sp->bytes_to_send > 0)
 	{
 	  send_test_chunk (tm, sp);
 	  delete_session = 0;
 	  tx_quota++;
 	}
-      if (sp->bytes_to_receive > 0)
+      if (!tm->no_return && sp->bytes_to_receive > 0)
 	{
 	  prev_bytes_received_this_session = sp->bytes_received;
 	  receive_test_chunk (tm, sp);
@@ -205,13 +206,14 @@ builtin_client_node_fn (vlib_main_t * vm, vlib_node_runtime_t * node,
 	}
       if (PREDICT_FALSE (delete_session == 1))
 	{
+	  __sync_fetch_and_add (&tm->tx_total, tm->bytes_to_send);
 	  __sync_fetch_and_add (&tm->rx_total, sp->bytes_received);
+
 	  dmp = vl_msg_api_alloc_as_if_client (sizeof (*dmp));
 	  memset (dmp, 0, sizeof (*dmp));
 	  dmp->_vl_msg_id = ntohs (VL_API_DISCONNECT_SESSION);
 	  dmp->client_index = tm->my_client_index;
 	  dmp->handle = sp->vpp_session_handle;
-//        vl_msg_api_send_shmem (tm->vl_input_queue, (u8 *) & dmp);
 	  if (!unix_shared_memory_queue_add (tm->vl_input_queue, (u8 *) & dmp,
 					     1))
 	    {
@@ -247,7 +249,6 @@ VLIB_REGISTER_NODE (builtin_client_node) =
 };
 /* *INDENT-ON* */
 
-
 /* So we don't get "no handler for... " msgs */
 static void
 vl_api_memclnt_create_reply_t_handler (vl_api_memclnt_create_reply_t * mp)
@@ -255,76 +256,10 @@ vl_api_memclnt_create_reply_t_handler (vl_api_memclnt_create_reply_t * mp)
   vlib_main_t *vm = vlib_get_main ();
   tclient_main_t *tm = &tclient_main;
   tm->my_client_index = mp->index;
-  vlib_process_signal_event (vm, tm->node_index, 1 /* evt */ ,
+  vlib_process_signal_event (vm, tm->cli_node_index, 1 /* evt */ ,
 			     0 /* data */ );
 }
 
-static void
-vl_api_connect_uri_reply_t_handler (vl_api_connect_uri_reply_t * mp)
-{
-  tclient_main_t *tm = &tclient_main;
-  session_t *session;
-  u32 session_index;
-  i32 retval = /* clib_net_to_host_u32 ( */ mp->retval /*) */ ;
-  int i;
-
-  if (retval < 0)
-    {
-      clib_warning ("connection failed: retval %d", retval);
-      return;
-    }
-
-  tm->our_event_queue =
-    uword_to_pointer (mp->vpp_event_queue_address,
-		      unix_shared_memory_queue_t *);
-  tm->vpp_event_queue =
-    uword_to_pointer (mp->vpp_event_queue_address,
-		      unix_shared_memory_queue_t *);
-
-  /*
-   * Setup session
-   */
-  pool_get (tm->sessions, session);
-  memset (session, 0, sizeof (*session));
-  session_index = session - tm->sessions;
-  session->bytes_to_receive = session->bytes_to_send = tm->bytes_to_send;
-
-  session->server_rx_fifo =
-    uword_to_pointer (mp->server_rx_fifo, svm_fifo_t *);
-  session->server_rx_fifo->client_session_index = session_index;
-  session->server_tx_fifo =
-    uword_to_pointer (mp->server_tx_fifo, svm_fifo_t *);
-  session->server_tx_fifo->client_session_index = session_index;
-  session->vpp_session_handle = mp->handle;
-
-  /* Add it to the session lookup table */
-  hash_set (tm->session_index_by_vpp_handles, mp->handle, session_index);
-
-  if (tm->ready_connections == tm->expected_connections - 1)
-    {
-      vlib_thread_main_t *thread_main = vlib_get_thread_main ();
-      int thread_index;
-
-      thread_index = 0;
-      for (i = 0; i < pool_elts (tm->sessions); i++)
-	{
-	  vec_add1 (tm->connection_index_by_thread[thread_index], i);
-	  thread_index++;
-	  if (thread_index == thread_main->n_vlib_mains)
-	    thread_index = 0;
-	}
-    }
-  __sync_fetch_and_add (&tm->ready_connections, 1);
-  if (tm->ready_connections == tm->expected_connections)
-    {
-      tm->run_test = 1;
-      tm->test_start_time = vlib_time_now (tm->vlib_main);
-      /* Signal the CLI process that the action is starting... */
-      vlib_process_signal_event (tm->vlib_main, tm->cli_node_index,
-				 1, 0 /* data */ );
-    }
-}
-
 static int
 create_api_loopback (tclient_main_t * tm)
 {
@@ -347,12 +282,11 @@ create_api_loopback (tclient_main_t * tm)
   mp->_vl_msg_id = VL_API_MEMCLNT_CREATE;
   mp->context = 0xFEEDFACE;
   mp->input_queue = pointer_to_uword (tm->vl_input_queue);
-  strncpy ((char *) mp->name, "tcp_tester", sizeof (mp->name) - 1);
+  strncpy ((char *) mp->name, "tcp_clients_tester", sizeof (mp->name) - 1);
 
   vl_api_memclnt_create_t_handler (mp);
 
   /* Wait for reply */
-  tm->node_index = vlib_get_current_process (vm)->node_runtime.node_index;
   vlib_process_wait_for_event_or_clock (vm, 1.0);
   event_type = vlib_process_get_events (vm, &event_data);
   switch (event_type)
@@ -373,7 +307,6 @@ create_api_loopback (tclient_main_t * tm)
 
 #define foreach_tclient_static_api_msg       	\
 _(MEMCLNT_CREATE_REPLY, memclnt_create_reply)   \
-_(CONNECT_URI_REPLY, connect_uri_reply)
 
 static clib_error_t *
 tclient_api_hookup (vlib_main_t * vm)
@@ -411,8 +344,8 @@ tcp_test_clients_init (vlib_main_t * vm)
   if (create_api_loopback (tm))
     return -1;
 
-  /* Init test data */
-  vec_validate (tm->connect_test_data, 64 * 1024 - 1);
+  /* Init test data. Big buffer */
+  vec_validate (tm->connect_test_data, 1024 * 1024 - 1);
   for (i = 0; i < vec_len (tm->connect_test_data); i++)
     tm->connect_test_data[i] = i & 0xff;
 
@@ -430,37 +363,66 @@ static int
 builtin_session_connected_callback (u32 app_index, u32 api_context,
 				    stream_session_t * s, u8 is_fail)
 {
-  vl_api_connect_uri_reply_t _m, *mp = &_m;
-  unix_shared_memory_queue_t *q;
-  application_t *app;
-  unix_shared_memory_queue_t *vpp_queue;
+  tclient_main_t *tm = &tclient_main;
+  session_t *session;
+  u32 session_index;
+  int i;
 
-  app = application_get (app_index);
-  q = vl_api_client_index_to_input_queue (app->api_client_index);
+  if (is_fail)
+    {
+      clib_warning ("connection %d failed!", api_context);
+      vlib_process_signal_event (tm->vlib_main, tm->cli_node_index, -1,
+				 0 /* data */ );
+      return -1;
+    }
 
-  if (!q)
-    return -1;
+  /* Mark vpp session as connected */
+  s->session_state = SESSION_STATE_READY;
 
-  memset (mp, 0, sizeof (*mp));
-  mp->_vl_msg_id = clib_host_to_net_u16 (VL_API_CONNECT_URI_REPLY);
-  mp->context = api_context;
-  if (!is_fail)
+  tm->our_event_queue = session_manager_get_vpp_event_queue (s->thread_index);
+  tm->vpp_event_queue = session_manager_get_vpp_event_queue (s->thread_index);
+
+  /*
+   * Setup session
+   */
+  pool_get (tm->sessions, session);
+  memset (session, 0, sizeof (*session));
+  session_index = session - tm->sessions;
+  session->bytes_to_receive = session->bytes_to_send = tm->bytes_to_send;
+  session->server_rx_fifo = s->server_rx_fifo;
+  session->server_rx_fifo->client_session_index = session_index;
+  session->server_tx_fifo = s->server_tx_fifo;
+  session->server_tx_fifo->client_session_index = session_index;
+  session->vpp_session_handle = stream_session_handle (s);
+
+  /* Add it to the session lookup table */
+  hash_set (tm->session_index_by_vpp_handles, session->vpp_session_handle,
+	    session_index);
+
+  if (tm->ready_connections == tm->expected_connections - 1)
     {
-      vpp_queue = session_manager_get_vpp_event_queue (s->thread_index);
-      mp->server_rx_fifo = pointer_to_uword (s->server_rx_fifo);
-      mp->server_tx_fifo = pointer_to_uword (s->server_tx_fifo);
-      mp->handle = stream_session_handle (s);
-      mp->vpp_event_queue_address = pointer_to_uword (vpp_queue);
-      mp->retval = 0;
-      s->session_state = SESSION_STATE_READY;
+      vlib_thread_main_t *thread_main = vlib_get_thread_main ();
+      int thread_index;
+
+      thread_index = 0;
+      for (i = 0; i < pool_elts (tm->sessions); i++)
+	{
+	  vec_add1 (tm->connection_index_by_thread[thread_index], i);
+	  thread_index++;
+	  if (thread_index == thread_main->n_vlib_mains)
+	    thread_index = 0;
+	}
     }
-  else
+  __sync_fetch_and_add (&tm->ready_connections, 1);
+  if (tm->ready_connections == tm->expected_connections)
     {
-      mp->retval = clib_host_to_net_u32 (VNET_API_ERROR_SESSION_CONNECT_FAIL);
+      tm->run_test = 1;
+      tm->test_start_time = vlib_time_now (tm->vlib_main);
+      /* Signal the CLI process that the action is starting... */
+      vlib_process_signal_event (tm->vlib_main, tm->cli_node_index, 1,
+				 0 /* data */ );
     }
 
-  vl_api_connect_uri_reply_t_handler (mp);
-
   return 0;
 }
 
@@ -489,23 +451,22 @@ builtin_server_rx_callback (stream_session_t * s)
 }
 
 /* *INDENT-OFF* */
-static session_cb_vft_t builtin_clients =
-  {
-    .session_reset_callback = builtin_session_reset_callback,
-    .session_connected_callback = builtin_session_connected_callback,
-    .session_accept_callback = builtin_session_create_callback,
-    .session_disconnect_callback = builtin_session_disconnect_callback,
-    .builtin_server_rx_callback = builtin_server_rx_callback
-  };
+static session_cb_vft_t builtin_clients = {
+  .session_reset_callback = builtin_session_reset_callback,
+  .session_connected_callback = builtin_session_connected_callback,
+  .session_accept_callback = builtin_session_create_callback,
+  .session_disconnect_callback = builtin_session_disconnect_callback,
+  .builtin_server_rx_callback = builtin_server_rx_callback
+};
 /* *INDENT-ON* */
 
 static int
-attach_builtin_test_clients ()
+attach_builtin_test_clients_app (void)
 {
   tclient_main_t *tm = &tclient_main;
   vnet_app_attach_args_t _a, *a = &_a;
   u8 segment_name[128];
-  u32 segment_name_length;
+  u32 segment_name_length, prealloc_fifos;
   u64 options[16];
 
   segment_name_length = ARRAY_LEN (segment_name);
@@ -518,13 +479,68 @@ attach_builtin_test_clients ()
   a->segment_name_length = segment_name_length;
   a->session_cb_vft = &builtin_clients;
 
+  prealloc_fifos = tm->prealloc_fifos ? tm->expected_connections : 1;
+
   options[SESSION_OPTIONS_ACCEPT_COOKIE] = 0x12345678;
-  options[SESSION_OPTIONS_SEGMENT_SIZE] = (2 << 30);	/*$$$$ config / arg */
+  options[SESSION_OPTIONS_SEGMENT_SIZE] = (2ULL << 32);
+  options[SESSION_OPTIONS_RX_FIFO_SIZE] = tm->fifo_size;
+  options[SESSION_OPTIONS_TX_FIFO_SIZE] = tm->fifo_size / 2;
+  options[APP_OPTIONS_PREALLOC_FIFO_PAIRS] = prealloc_fifos;
+
   options[APP_OPTIONS_FLAGS] = APP_OPTIONS_FLAGS_BUILTIN_APP;
 
   a->options = options;
 
-  return vnet_application_attach (a);
+  if (vnet_application_attach (a))
+    return -1;
+
+  tm->app_index = a->app_index;
+  return 0;
+}
+
+static void *
+tclient_thread_fn (void *arg)
+{
+  return 0;
+}
+
+/** Start a transmit thread */
+int
+start_tx_pthread (tclient_main_t * tm)
+{
+  if (tm->client_thread_handle == 0)
+    {
+      int rv = pthread_create (&tm->client_thread_handle,
+			       NULL /*attr */ ,
+			       tclient_thread_fn, 0);
+      if (rv)
+	{
+	  tm->client_thread_handle = 0;
+	  return -1;
+	}
+    }
+  return 0;
+}
+
+void
+clients_connect (vlib_main_t * vm, u8 * uri, u32 n_clients)
+{
+  tclient_main_t *tm = &tclient_main;
+  vnet_connect_args_t _a, *a = &_a;
+  int i;
+  for (i = 0; i < n_clients; i++)
+    {
+      memset (a, 0, sizeof (*a));
+
+      a->uri = (char *) uri;
+      a->api_context = i;
+      a->app_index = tm->app_index;
+      a->mp = 0;
+      vnet_connect_uri (a);
+
+      /* Crude pacing for call setups, 100k/sec  */
+      vlib_process_suspend (vm, 10e-6);
+    }
 }
 
 static clib_error_t *
@@ -534,17 +550,18 @@ test_tcp_clients_command_fn (vlib_main_t * vm,
 {
   tclient_main_t *tm = &tclient_main;
   vlib_thread_main_t *thread_main = vlib_get_thread_main ();
-  uword *event_data = 0;
-  uword event_type;
-  u8 *connect_uri = (u8 *) "tcp://6.0.1.1/1234";
-  u8 *uri;
+  uword *event_data = 0, event_type;
+  u8 *default_connect_uri = (u8 *) "tcp://6.0.1.1/1234", *uri;
+  u64 tmp, total_bytes;
+  f64 cli_timeout = 20.0, delta;
   u32 n_clients = 1;
+  char *transfer_type;
   int i;
-  u64 tmp;
-  f64 cli_timeout = 20.0;
-  f64 delta;
 
   tm->bytes_to_send = 8192;
+  tm->no_return = 0;
+  tm->fifo_size = 64 << 10;
+
   vec_free (tm->connect_uri);
 
   while (unformat_check_input (input) != UNFORMAT_END_OF_INPUT)
@@ -561,11 +578,18 @@ test_tcp_clients_command_fn (vlib_main_t * vm,
 	;
       else if (unformat (input, "cli-timeout %f", &cli_timeout))
 	;
+      else if (unformat (input, "no-return"))
+	tm->no_return = 1;
+      else if (unformat (input, "fifo-size %d", &tm->fifo_size))
+	tm->fifo_size <<= 10;
       else
 	return clib_error_return (0, "unknown input `%U'",
 				  format_unformat_error, input);
     }
 
+  /* Store cli process node index for signalling */
+  tm->cli_node_index = vlib_get_current_process (vm)->node_runtime.node_index;
+
   if (tm->is_init == 0)
     {
       if (tcp_test_clients_init (vm))
@@ -575,28 +599,25 @@ test_tcp_clients_command_fn (vlib_main_t * vm,
   tm->ready_connections = 0;
   tm->expected_connections = n_clients;
   tm->rx_total = 0;
+  tm->tx_total = 0;
 
-  uri = connect_uri;
+  uri = default_connect_uri;
   if (tm->connect_uri)
     uri = tm->connect_uri;
 
 #if TCP_BUILTIN_CLIENT_PTHREAD
-  /* Start a transmit thread */
-  if (tm->client_thread_handle == 0)
+  start_tx_pthread ();
+#endif
+
+  vnet_session_enable_disable (vm, 1 /* turn on TCP, etc. */ );
+
+  if (tm->test_client_attached == 0)
     {
-      int rv = pthread_create (&tm->client_thread_handle,
-			       NULL /*attr */ ,
-			       tclient_thread_fn, 0);
-      if (rv)
+      if (attach_builtin_test_clients_app ())
 	{
-	  tm->client_thread_handle = 0;
-	  return clib_error_return (0, "pthread_create returned %d", rv);
+	  return clib_error_return (0, "app attach failed");
 	}
     }
-#endif
-  vnet_session_enable_disable (vm, 1 /* turn on TCP, etc. */ );
-  if (tm->test_client_attached == 0)
-    attach_builtin_test_clients ();
   tm->test_client_attached = 1;
 
   /* Turn on the builtin client input nodes */
@@ -604,25 +625,8 @@ test_tcp_clients_command_fn (vlib_main_t * vm,
     vlib_node_set_state (vlib_mains[i], builtin_client_node.index,
 			 VLIB_NODE_STATE_POLLING);
 
-  tm->cli_node_index = vlib_get_current_process (vm)->node_runtime.node_index;
-
   /* Fire off connect requests */
-  for (i = 0; i < n_clients; i++)
-    {
-      vl_api_connect_uri_t _cmp, *cmp = &_cmp;
-      void vl_api_connect_uri_t_handler (vl_api_connect_uri_t * cmp);
-
-      memset (cmp, 0, sizeof (*cmp));
-
-      cmp->_vl_msg_id = ntohs (VL_API_CONNECT_URI);
-      cmp->client_index = tm->my_client_index;
-      cmp->context = ntohl (0xfeedface);
-      memcpy (cmp->uri, uri, strlen ((char *) uri) + 1);
-
-      vl_api_connect_uri_t_handler (cmp);
-      /* Crude pacing for call setups, 100k/sec  */
-      vlib_process_suspend (vm, 10e-6);
-    }
+  clients_connect (vm, uri, n_clients);
 
   /* Park until the sessions come up, or ten seconds elapse... */
   vlib_process_wait_for_event_or_clock (vm, 10.0 /* timeout, seconds */ );
@@ -668,14 +672,17 @@ test_tcp_clients_command_fn (vlib_main_t * vm,
 
   if (delta != 0.0)
     {
+      total_bytes = (tm->no_return ? tm->tx_total : tm->rx_total);
+      transfer_type = tm->no_return ? "half-duplex" : "full-duplex";
       vlib_cli_output (vm,
 		       "%lld bytes (%lld mbytes, %lld gbytes) in %.2f seconds",
-		       tm->rx_total, tm->rx_total / (1ULL << 20),
-		       tm->rx_total / (1ULL << 30), delta);
-      vlib_cli_output (vm, "%.2f bytes/second full-duplex",
-		       ((f64) tm->rx_total) / (delta));
-      vlib_cli_output (vm, "%.4f gbit/second full-duplex",
-		       (((f64) tm->rx_total * 8.0) / delta / 1e9));
+		       total_bytes, total_bytes / (1ULL << 20),
+		       total_bytes / (1ULL << 30), delta);
+      vlib_cli_output (vm, "%.2f bytes/second %s",
+		       ((f64) total_bytes) / (delta), transfer_type);
+      vlib_cli_output (vm, "%.4f gbit/second %s",
+		       (((f64) total_bytes * 8.0) / delta / 1e9),
+		       transfer_type);
     }
   else
     vlib_cli_output (vm, "zero delta-t?");
diff --git a/src/vnet/tcp/builtin_client.h b/src/vnet/tcp/builtin_client.h
index d5d79e53..3462e0ee 100644
--- a/src/vnet/tcp/builtin_client.h
+++ b/src/vnet/tcp/builtin_client.h
@@ -44,78 +44,59 @@ typedef struct
 
 typedef struct
 {
-  /* API message ID base */
-  u16 msg_id_base;
-
-  /* vpe input queue */
-  unix_shared_memory_queue_t *vl_input_queue;
-
-  /* API client handle */
-  u32 my_client_index;
-
-  /* The URI we're playing with */
-  u8 *uri;
-
-  /* Session pool */
-  session_t *sessions;
-
-  /* Hash table for disconnect processing */
-  uword *session_index_by_vpp_handles;
-
-  /* intermediate rx buffer */
-  u8 *rx_buf;
-
-  /* URI for slave's connect */
-  u8 *connect_uri;
-
-  u32 connected_session_index;
-
-  int i_am_master;
-
-  /* drop all packets */
-  int drop_packets;
-
-  /* Our event queue */
-  unix_shared_memory_queue_t *our_event_queue;
-
-  /* $$$ single thread only for the moment */
-  unix_shared_memory_queue_t *vpp_event_queue;
-
-  pid_t my_pid;
-
-  f64 test_start_time;
-  f64 test_end_time;
-
-  u32 expected_connections;
+  /*
+   * Application setup parameters
+   */
+  unix_shared_memory_queue_t *vl_input_queue;	/**< vpe input queue */
+  unix_shared_memory_queue_t *our_event_queue;	/**< Our event queue */
+  unix_shared_memory_queue_t *vpp_event_queue;	/**< $$$ single thread */
+
+  u32 cli_node_index;			/**< cli process node index */
+  u32 my_client_index;			/**< loopback API client handle */
+  u32 app_index;			/**< app index after attach */
+
+  /*
+   * Configuration params
+   */
+  u8 *connect_uri;			/**< URI for slave's connect */
+  u64 bytes_to_send;			/**< Bytes to send */
+  u32 configured_segment_size;
+  u32 fifo_size;
+  u32 expected_connections;		/**< Number of clients/connections */
+
+  /*
+   * Test state variables
+   */
+  session_t *sessions;			/**< Sessions pool */
+  u8 *rx_buf;				/**< intermediate rx buffer */
+  uword *session_index_by_vpp_handles;	/**< Hash table for disconnecting */
+  u8 *connect_test_data;		/**< Pre-computed test data */
   u32 **connection_index_by_thread;
+  pthread_t client_thread_handle;
+
   volatile u32 ready_connections;
   volatile u32 finished_connections;
-
   volatile u64 rx_total;
-  u32 cli_node_index;
-
-  /* Signal variable */
-  volatile int run_test;
-
-  /* Bytes to send */
-  u64 bytes_to_send;
-
-  u32 configured_segment_size;
+  volatile u64 tx_total;
+  volatile int run_test;		/**< Signal start of test */
 
-  /* VNET_API_ERROR_FOO -> "Foo" hash table */
-  uword *error_string_by_error_number;
-
-  u8 *connect_test_data;
-  pthread_t client_thread_handle;
-  u64 client_bytes_received;
-  u8 test_return_packets;
+  f64 test_start_time;
+  f64 test_end_time;
 
+  /*
+   * Flags
+   */
   u8 is_init;
   u8 test_client_attached;
+  u8 no_return;
+  u8 test_return_packets;
+  int i_am_master;
+  int drop_packets;		/**< drop all packets */
+  u8 prealloc_fifos;		/**< Request fifo preallocation */
 
-  u32 node_index;
-
-  /* convenience */
+  /*
+   * Convenience
+   */
   vlib_main_t *vlib_main;
   vnet_main_t *vnet_main;
   ethernet_main_t *ethernet_main;
diff --git a/src/vnet/tcp/builtin_server.c b/src/vnet/tcp/builtin_server.c
index 8bd2f360..775bfc26 100644
--- a/src/vnet/tcp/builtin_server.c
+++ b/src/vnet/tcp/builtin_server.c
@@ -39,21 +39,30 @@
 
 typedef struct
 {
-  /* Per-thread RX buffer */
-  u8 **rx_buf;
+  /*
+   * Server app parameters
+   */
   unix_shared_memory_queue_t **vpp_queue;
-  u64 byte_index;
+  unix_shared_memory_queue_t *vl_input_queue;	/**< Sever's event queue */
 
-  /* Sever's event queue */
-  unix_shared_memory_queue_t *vl_input_queue;
+  u32 app_index;		/**< Server app index */
+  u32 my_client_index;		/**< API client handle */
+  u32 node_index;		/**< process node index for evnt scheduling */
 
-  /* API client handle */
-  u32 my_client_index;
+  /*
+   * Config params
+   */
+  u8 no_echo;			/**< Don't echo traffic */
+  u32 fifo_size;		/**< Fifo size */
+  u32 rcv_buffer_size;		/**< Rcv buffer size */
+  u32 prealloc_fifos;		/**< Preallocate fifos */
 
-  u32 app_index;
+  /*
+   * Test state
+   */
+  u8 **rx_buf;			/**< Per-thread RX buffer */
+  u64 byte_index;
 
-  /* process node index for evnt scheduling */
-  u32 node_index;
   vlib_main_t *vlib_main;
 } builtin_server_main_t;
 
@@ -132,6 +141,29 @@ test_bytes (builtin_server_main_t * bsm, int actual_transfer)
   bsm->byte_index += actual_transfer;
 }
 
+/*
+ * If no-echo, just read the data and be done with it
+ */
+int
+builtin_server_rx_callback_no_echo (stream_session_t * s)
+{
+  builtin_server_main_t *bsm = &builtin_server_main;
+  u32 my_thread_id = vlib_get_thread_index ();
+  int actual_transfer;
+  svm_fifo_t *rx_fifo;
+
+  rx_fifo = s->server_rx_fifo;
+
+  do
+    {
+      actual_transfer =
+	svm_fifo_dequeue_nowait (rx_fifo, bsm->rcv_buffer_size,
+				 bsm->rx_buf[my_thread_id]);
+    }
+  while (actual_transfer > 0);
+  return 0;
+}
+
 int
 builtin_server_rx_callback (stream_session_t * s)
 {
@@ -143,8 +175,8 @@ builtin_server_rx_callback (stream_session_t * s)
   static int serial_number = 0;
   u32 my_thread_id = vlib_get_thread_index ();
 
-  tx_fifo = s->server_tx_fifo;
   rx_fifo = s->server_rx_fifo;
+  tx_fifo = s->server_tx_fifo;
 
   max_dequeue = svm_fifo_max_dequeue (s->server_rx_fifo);
   max_enqueue = svm_fifo_max_enqueue (s->server_tx_fifo);
@@ -164,19 +196,22 @@ builtin_server_rx_callback (stream_session_t * s)
       /* Program self-tap to retry */
       if (svm_fifo_set_event (rx_fifo))
 	{
+	  unix_shared_memory_queue_t *q;
 	  evt.fifo = rx_fifo;
 	  evt.event_type = FIFO_EVENT_BUILTIN_RX;
 	  evt.event_id = 0;
-	  unix_shared_memory_queue_add (bsm->vpp_queue[s->thread_index],
-					(u8 *) & evt,
-					0 /* do wait for mutex */ );
+
+	  q = bsm->vpp_queue[s->thread_index];
+	  if (PREDICT_FALSE (q->cursize == q->maxsize))
+	    clib_warning ("out of event queue space");
+	  else
+	    unix_shared_memory_queue_add (q, (u8 *) & evt,
+					  0 /* don't wait for mutex */ );
 	}
 
       return 0;
     }
 
-  vec_validate (bsm->rx_buf, my_thread_id);
-  vec_validate (bsm->rx_buf[my_thread_id], max_transfer - 1);
   _vec_len (bsm->rx_buf[my_thread_id]) = max_transfer;
 
   actual_transfer = svm_fifo_dequeue_nowait (rx_fifo, max_transfer,
@@ -281,14 +316,21 @@ server_attach ()
   memset (a, 0, sizeof (*a));
   memset (options, 0, sizeof (options));
 
+  if (bsm->no_echo)
+    builtin_session_cb_vft.builtin_server_rx_callback =
+      builtin_server_rx_callback_no_echo;
+  else
+    builtin_session_cb_vft.builtin_server_rx_callback =
+      builtin_server_rx_callback;
   a->api_client_index = bsm->my_client_index;
   a->session_cb_vft = &builtin_session_cb_vft;
   a->options = options;
   a->options[SESSION_OPTIONS_SEGMENT_SIZE] = 512 << 20;
-  a->options[SESSION_OPTIONS_RX_FIFO_SIZE] = 64 << 10;
-  a->options[SESSION_OPTIONS_TX_FIFO_SIZE] = 64 << 10;
+  a->options[SESSION_OPTIONS_RX_FIFO_SIZE] = bsm->fifo_size;
+  a->options[SESSION_OPTIONS_TX_FIFO_SIZE] = bsm->fifo_size;
   a->options[APP_OPTIONS_FLAGS] = APP_OPTIONS_FLAGS_BUILTIN_APP;
-  a->options[APP_OPTIONS_PREALLOC_FIFO_PAIRS] = 8192;
+  a->options[APP_OPTIONS_PREALLOC_FIFO_PAIRS] =
+    bsm->prealloc_fifos ? bsm->prealloc_fifos : 1;
   a->segment_name = segment_name;
   a->segment_name_length = ARRAY_LEN (segment_name);
 
@@ -316,17 +358,24 @@ static int
 server_create (vlib_main_t * vm)
 {
   builtin_server_main_t *bsm = &builtin_server_main;
-  u32 num_threads;
   vlib_thread_main_t *vtm = vlib_get_thread_main ();
+  u32 num_threads;
+  int i;
 
   if (bsm->my_client_index == (u32) ~ 0)
     {
       if (create_api_loopback (vm))
-	return -1;
+	{
+	  clib_warning ("failed to create api loopback");
+	  return -1;
+	}
     }
 
   num_threads = 1 /* main thread */  + vtm->n_threads;
   vec_validate (builtin_server_main.vpp_queue, num_threads - 1);
+  vec_validate (bsm->rx_buf, num_threads - 1);
+  for (i = 0; i < num_threads; i++)
+    vec_validate (bsm->rx_buf[i], bsm->rcv_buffer_size);
 
   if (server_attach ())
     {
@@ -381,23 +430,35 @@ tcp_builtin_server_api_hookup (vlib_main_t * vm)
 }
 
 static clib_error_t *
-server_create_command_fn (vlib_main_t * vm,
-			  unformat_input_t * input, vlib_cli_command_t * cmd)
+server_create_command_fn (vlib_main_t * vm, unformat_input_t * input,
+			  vlib_cli_command_t * cmd)
 {
+  builtin_server_main_t *bsm = &builtin_server_main;
   int rv;
-#if 0
+
+  bsm->no_echo = 0;
+  bsm->fifo_size = 64 << 10;
+  bsm->rcv_buffer_size = 128 << 10;
+  bsm->prealloc_fifos = 0;
+
   while (unformat_check_input (input) != UNFORMAT_END_OF_INPUT)
     {
-      if (unformat (input, "whatever %d", &whatever))
+      if (unformat (input, "no-echo"))
+	bsm->no_echo = 1;
+      else if (unformat (input, "fifo-size %d", &bsm->fifo_size))
+	bsm->fifo_size <<= 10;
+      else if (unformat (input, "rcv-buf-size %d", &bsm->rcv_buffer_size))
+	;
+      else if (unformat (input, "prealloc-fifos", &bsm->prealloc_fifos))
 	;
       else
 	return clib_error_return (0, "unknown input `%U'",
 				  format_unformat_error, input);
     }
-#endif
 
   tcp_builtin_server_api_hookup (vm);
   vnet_session_enable_disable (vm, 1 /* turn on TCP, etc. */ );
+
   rv = server_create (vm);
   switch (rv)
     {
@@ -406,6 +467,7 @@ server_create_command_fn (vlib_main_t * vm,
     default:
       return clib_error_return (0, "server_create returned %d", rv);
     }
+
   return 0;
 }
 
diff --git a/src/vnet/tcp/tcp.c b/src/vnet/tcp/tcp.c
index e0b67a8e..5c554bac 100644
--- a/src/vnet/tcp/tcp.c
+++ b/src/vnet/tcp/tcp.c
@@ -726,15 +726,25 @@ tcp_round_snd_space (tcp_connection_t * tc, u32 snd_space)
 u32
 tcp_snd_space (tcp_connection_t * tc)
 {
-  int snd_space;
+  int snd_space, snt_limited;
 
-  /* If we haven't gotten dupacks or if we did and have gotten sacked bytes
-   * then we can still send */
-  if (PREDICT_TRUE (tcp_in_cong_recovery (tc) == 0
-		    && (tc->rcv_dupacks == 0
-			|| tc->sack_sb.last_sacked_bytes)))
+  if (PREDICT_TRUE (tcp_in_cong_recovery (tc) == 0))
     {
       snd_space = tcp_available_snd_space (tc);
+
+      /* If we haven't gotten dupacks or if we did and have gotten sacked
+       * bytes then we can still send as per Limited Transmit (RFC3042) */
+      if (PREDICT_FALSE (tc->rcv_dupacks != 0
+			 && (tcp_opts_sack_permitted (tc)
+			     && tc->sack_sb.last_sacked_bytes == 0)))
+	{
+	  if (tc->rcv_dupacks == 1 && tc->limited_transmit != tc->snd_nxt)
+	    tc->limited_transmit = tc->snd_nxt;
+	  ASSERT (seq_leq (tc->limited_transmit, tc->snd_nxt));
+
+	  snt_limited = tc->snd_nxt - tc->limited_transmit;
+	  snd_space = clib_max (2 * tc->snd_mss - snt_limited, 0);
+	}
       return tcp_round_snd_space (tc, snd_space);
     }
 
diff --git a/src/vnet/tcp/tcp.h b/src/vnet/tcp/tcp.h
index 071f1ab1..e8398718 100644
--- a/src/vnet/tcp/tcp.h
+++ b/src/vnet/tcp/tcp.h
@@ -31,9 +31,9 @@
 #define TCP_MAX_OPTION_SPACE 40
 
 #define TCP_DUPACK_THRESHOLD 	3
-#define TCP_MAX_RX_FIFO_SIZE 	2 << 20
+#define TCP_MAX_RX_FIFO_SIZE 	4 << 20
 #define TCP_IW_N_SEGMENTS 	10
-#define TCP_ALWAYS_ACK		0	/**< If on, we always ack */
+#define TCP_ALWAYS_ACK		1	/**< On/off delayed acks */
 #define TCP_USE_SACKS		1	/**< Disable only for testing */
 
 /** TCP FSM state definitions as per RFC793. */
@@ -100,6 +100,7 @@ extern timer_expiration_handler tcp_timer_retransmit_syn_handler;
 #define TCP_TIMER_PERSIST_MIN	2	/* 0.2s */
 
 #define TCP_RTO_MAX 60 * THZ	/* Min max RTO (60s) as per RFC6298 */
+#define TCP_RTO_MIN 0.2 * THZ	/* Min RTO (200ms) - lower than standard */
 #define TCP_RTT_MAX 30 * THZ	/* 30s (probably too much) */
 #define TCP_RTO_SYN_RETRIES 3	/* SYN retries without doubling RTO */
 #define TCP_RTO_INIT 1 * THZ	/* Initial retransmit timer */
@@ -149,7 +150,7 @@ enum
 #undef _
 };
 
-#define TCP_MAX_SACK_BLOCKS 5	/**< Max number of SACK blocks stored */
+#define TCP_MAX_SACK_BLOCKS 15	/**< Max number of SACK blocks stored */
 #define TCP_INVALID_SACK_HOLE_INDEX ((u32)~0)
 
 typedef struct _sack_scoreboard_hole
@@ -208,6 +209,7 @@ typedef struct _tcp_connection
   u32 snd_wl1;		/**< seq number used for last snd.wnd update */
   u32 snd_wl2;		/**< ack number used for last snd.wnd update */
   u32 snd_nxt;		/**< next seq number to be sent */
+  u16 snd_mss;		/**< Effective send max seg (data) size */
 
   /** Receive sequence variables RFC793 */
   u32 rcv_nxt;		/**< next sequence number expected */
@@ -252,8 +254,8 @@ typedef struct _tcp_connection
   u32 rtt_ts;		/**< Timestamp for tracked ACK */
   u32 rtt_seq;		/**< Sequence number for tracked ACK */
 
-  u16 snd_mss;		/**< Effective send max seg (data) size */
   u16 mss;		/**< Our max seg size that includes options */
+  u32 limited_transmit;	/**< snd_nxt when limited transmit starts */
 } tcp_connection_t;
 
 struct _tcp_cc_algorithm
@@ -433,6 +435,7 @@ tcp_end_seq (tcp_header_t * th, u32 len)
 #define seq_leq(_s1, _s2) ((i32)((_s1)-(_s2)) <= 0)
 #define seq_gt(_s1, _s2) ((i32)((_s1)-(_s2)) > 0)
 #define seq_geq(_s1, _s2) ((i32)((_s1)-(_s2)) >= 0)
+#define seq_max(_s1, _s2) (seq_gt((_s1), (_s2)) ? (_s1) : (_s2))
 
 /* Modulo arithmetic for timestamps */
 #define timestamp_lt(_t1, _t2) ((i32)((_t1)-(_t2)) < 0)
@@ -719,6 +722,7 @@ scoreboard_clear (sack_scoreboard_t * sb)
     {
       scoreboard_remove_hole (sb, hole);
     }
+  ASSERT (sb->head == sb->tail && sb->head == TCP_INVALID_SACK_HOLE_INDEX);
   sb->sacked_bytes = 0;
   sb->last_sacked_bytes = 0;
   sb->last_bytes_delivered = 0;
diff --git a/src/vnet/tcp/tcp_debug.h b/src/vnet/tcp/tcp_debug.h
index 3a16cf63..ae68ad1b 100755
--- a/src/vnet/tcp/tcp_debug.h
+++ b/src/vnet/tcp/tcp_debug.h
@@ -19,8 +19,10 @@
 #include <vlib/vlib.h>
 
 #define TCP_DEBUG (1)
+#define TCP_DEBUG_SM (0)
 #define TCP_DEBUG_CC (1)
-#define TCP_DEBUG_VERBOSE (0)
+#define TCP_DEBUG_CC_STAT (1)
+#define TCP_DEBUG_SM_VERBOSE (0)
 
 #define foreach_tcp_dbg_evt		\
   _(INIT, "")				\
@@ -49,6 +51,8 @@
   _(CC_RTX, "retransmit")		\
   _(CC_EVT, "cc event")			\
   _(CC_PACK, "cc partial ack")		\
+  _(CC_STAT, "cc stats")		\
+  _(CC_RTO_STAT, "cc rto stats")	\
   _(SEG_INVALID, "invalid segment")	\
   _(PAWS_FAIL, "failed paws check")	\
   _(ACK_RCV_ERR, "invalid ack")		\
@@ -72,6 +76,10 @@ typedef enum _tcp_dbg_evt
 
 #define TRANSPORT_DEBUG (1)
 
+/*
+ * Infra and evt track setup
+ */
+
 #define TCP_DBG(_tc, _evt, _args...)					\
 {   		            						\
     u8 *_tmp = 0;							\
@@ -158,6 +166,30 @@ typedef enum _tcp_dbg_evt
   TCP_EVT_DEALLOC_HANDLER(_tc);						\
 }
 
+#define TCP_EVT_SYN_RCVD_HANDLER(_tc, ...)				\
+{									\
+  TCP_EVT_INIT_HANDLER(_tc, "s%d%c");					\
+  ELOG_TYPE_DECLARE (_e) =						\
+  {									\
+    .format = "SYNrx: irs %u",						\
+    .format_args = "i4",						\
+  };									\
+  DECLARE_ETD(_tc, _e, 1);						\
+  ed->data[0] = _tc->irs;						\
+}
+
+#define CONCAT_HELPER(_a, _b) _a##_b
+#define CC(_a, _b) CONCAT_HELPER(_a, _b)
+#define TCP_EVT_DBG(_evt, _args...) CC(_evt, _HANDLER)(_args)
+#else
+#define TCP_EVT_DBG(_evt, _args...)
+#endif
+
+/*
+ * State machine
+ */
+#if TCP_DEBUG_SM
+
 #define TCP_EVT_ACK_SENT_HANDLER(_tc, ...)				\
 {									\
   ELOG_TYPE_DECLARE (_e) =						\
@@ -234,18 +266,6 @@ typedef enum _tcp_dbg_evt
   ed->data[1] = _tc->rcv_nxt - _tc->irs;				\
 }
 
-#define TCP_EVT_SYN_RCVD_HANDLER(_tc, ...)				\
-{									\
-  TCP_EVT_INIT_HANDLER(_tc, "s%d%c");					\
-  ELOG_TYPE_DECLARE (_e) =						\
-  {									\
-    .format = "SYNrx: irs %u",						\
-    .format_args = "i4",						\
-  };									\
-  DECLARE_ETD(_tc, _e, 1);						\
-  ed->data[0] = _tc->irs;						\
-}
-
 #define TCP_EVT_FIN_RCVD_HANDLER(_tc, ...)				\
 {									\
   ELOG_TYPE_DECLARE (_e) =						\
@@ -418,6 +438,74 @@ typedef enum _tcp_dbg_evt
   ed->data[4] = _tc->snd_una_max - _tc->iss;				\
 }
 
+#define TCP_EVT_RCV_WND_SHRUNK_HANDLER(_tc, _obs, _av, ...)		\
+{									\
+if (_av > 0) 								\
+{									\
+  ELOG_TYPE_DECLARE (_e) =						\
+  {									\
+    .format = "huh?: rcv_wnd %u obsd %u av %u rcv_nxt %u rcv_las %u",	\
+    .format_args = "i4i4i4i4i4",					\
+  };									\
+  DECLARE_ETD(_tc, _e, 5);						\
+  ed->data[0] = _tc->rcv_wnd;						\
+  ed->data[1] = _obs;							\
+  ed->data[2] = _av;							\
+  ed->data[3] = _tc->rcv_nxt - _tc->irs;				\
+  ed->data[4] = _tc->rcv_las - _tc->irs;				\
+}									\
+}
+#else
+#define TCP_EVT_ACK_SENT_HANDLER(_tc, ...)
+#define TCP_EVT_DUPACK_SENT_HANDLER(_tc, ...)
+#define TCP_EVT_SYN_SENT_HANDLER(_tc, ...)
+#define TCP_EVT_SYN_RTX_HANDLER(_tc, ...)
+#define TCP_EVT_FIN_SENT_HANDLER(_tc, ...)
+#define TCP_EVT_RST_SENT_HANDLER(_tc, ...)
+#define TCP_EVT_FIN_RCVD_HANDLER(_tc, ...)
+#define TCP_EVT_RST_RCVD_HANDLER(_tc, ...)
+#define TCP_EVT_ACK_RCVD_HANDLER(_tc, ...)
+#define TCP_EVT_DUPACK_RCVD_HANDLER(_tc, ...)
+#define TCP_EVT_PKTIZE_HANDLER(_tc, ...)
+#define TCP_EVT_INPUT_HANDLER(_tc, _type, _len, _written, ...)
+#define TCP_EVT_TIMER_POP_HANDLER(_tc_index, _timer_id, ...)
+#define TCP_EVT_SEG_INVALID_HANDLER(_tc, _seq, _end, ...)
+#define TCP_EVT_PAWS_FAIL_HANDLER(_tc, _seq, _end, ...)
+#define TCP_EVT_ACK_RCV_ERR_HANDLER(_tc, _type, _ack, ...)
+#define TCP_EVT_RCV_WND_SHRUNK_HANDLER(_tc, _obs, _av, ...)
+#endif
+
+/*
+ * State machine verbose
+ */
+#if TCP_DBG_SM_VERBOSE
+#define TCP_EVT_SND_WND_HANDLER(_tc, ...)				\
+{									\
+  ELOG_TYPE_DECLARE (_e) =						\
+  {									\
+    .format = "snd_wnd update: %u ",					\
+    .format_args = "i4",						\
+  };									\
+  DECLARE_ETD(_tc, _e, 1);						\
+  ed->data[0] = _tc->snd_wnd;						\
+}
+
+#define TCP_EVT_OUTPUT_HANDLER(_tc, flags, n_bytes,...)			\
+{									\
+  ELOG_TYPE_DECLARE (_e) =						\
+  {									\
+    .format = "out: flags %x, bytes %u",				\
+    .format_args = "i4i4",						\
+  };									\
+  DECLARE_ETD(_tc, _e, 2);						\
+  ed->data[0] = flags;							\
+  ed->data[1] = n_bytes;						\
+}
+#else
+#define TCP_EVT_SND_WND_HANDLER(_tc, ...)
+#define TCP_EVT_OUTPUT_HANDLER(_tc, flags, n_bytes,...)
+#endif
+
 /*
  * Congestion Control
  */
@@ -471,67 +559,59 @@ typedef enum _tcp_dbg_evt
   ed->data[1] = _tc->snd_una_max - _tc->iss;				\
 }
 
-#else
-#define TCP_EVT_CC_RTX_HANDLER(_tc, offset, n_bytes, ...)
-#define TCP_EVT_CC_EVT_HANDLER(_tc, _sub_evt, _snd_space, ...)
-#define TCP_EVT_CC_PACK_HANDLER(_tc, ...)
-#endif
+/*
+ * Congestion control stats
+ */
+#if TCP_DEBUG_CC_STAT
 
-#define TCP_EVT_RCV_WND_SHRUNK_HANDLER(_tc, _obs, _av, ...)		\
+#define STATS_INTERVAL 1
+
+#define TCP_EVT_CC_RTO_STAT_HANDLER(_tc, ...)				\
 {									\
-if (_av > 0) 								\
+if (_tc->c_cc_stat_tstamp + STATS_INTERVAL < tcp_time_now())		\
 {									\
   ELOG_TYPE_DECLARE (_e) =						\
   {									\
-    .format = "huh?: rcv_wnd %u obsd %u av %u rcv_nxt %u rcv_las %u",	\
-    .format_args = "i4i4i4i4i4",					\
+    .format = "rto_stat: rto %u srtt %u rttvar %u ",			\
+    .format_args = "i4i4i4",						\
   };									\
-  DECLARE_ETD(_tc, _e, 5);						\
-  ed->data[0] = _tc->rcv_wnd;						\
-  ed->data[1] = _obs;							\
-  ed->data[2] = _av;							\
-  ed->data[3] = _tc->rcv_nxt - _tc->irs;				\
-  ed->data[4] = _tc->rcv_las - _tc->irs;				\
+  DECLARE_ETD(_tc, _e, 3);						\
+  ed->data[0] = _tc->rto;						\
+  ed->data[1] = _tc->srtt;						\
+  ed->data[2] = _tc->rttvar;						\
 }									\
 }
 
-#if TCP_DBG_VERBOSE
-#define TCP_EVT_SND_WND_HANDLER(_tc, ...)				\
+#define TCP_EVT_CC_STAT_HANDLER(_tc, ...)				\
 {									\
-  ELOG_TYPE_DECLARE (_e) =						\
-  {									\
-    .format = "snd_wnd update: %u ",					\
-    .format_args = "i4",						\
-  };									\
-  DECLARE_ETD(_tc, _e, 1);						\
-  ed->data[0] = _tc->snd_wnd;						\
-}
-
-#define TCP_EVT_OUTPUT_HANDLER(_tc, flags, n_bytes,...)			\
+if (_tc->c_cc_stat_tstamp + STATS_INTERVAL < tcp_time_now())		\
 {									\
   ELOG_TYPE_DECLARE (_e) =						\
   {									\
-    .format = "out: flags %x, bytes %u",				\
-    .format_args = "i4i4",						\
+    .format = "cc_stat: cwnd %u flight %u space %u ssthresh %u snd_wnd %u",\
+    .format_args = "i4i4i4i4i4",					\
   };									\
-  DECLARE_ETD(_tc, _e, 2);						\
-  ed->data[0] = flags;							\
-  ed->data[1] = n_bytes;						\
+  DECLARE_ETD(_tc, _e, 5);						\
+  ed->data[0] = _tc->cwnd;						\
+  ed->data[1] = tcp_flight_size (_tc);					\
+  ed->data[2] = tcp_snd_space (_tc);					\
+  ed->data[3] = _tc->ssthresh;						\
+  ed->data[4] = _tc->snd_wnd;						\
+  TCP_EVT_CC_RTO_STAT_HANDLER (_tc);					\
+  _tc->c_cc_stat_tstamp = tcp_time_now();				\
+}									\
 }
+
 #else
-#define TCP_EVT_SND_WND_HANDLER(_tc, ...)
-#define TCP_EVT_OUTPUT_HANDLER(_tc, flags, n_bytes,...)
+#define TCP_EVT_CC_STAT_HANDLER(_tc, ...)
 #endif
 
-#define CONCAT_HELPER(_a, _b) _a##_b
-#define CC(_a, _b) CONCAT_HELPER(_a, _b)
-#define TCP_EVT_DBG(_evt, _args...) CC(_evt, _HANDLER)(_args)
-
 #else
-#define TCP_EVT_DBG(_evt, _args...)
+#define TCP_EVT_CC_RTX_HANDLER(_tc, offset, n_bytes, ...)
+#define TCP_EVT_CC_EVT_HANDLER(_tc, _sub_evt, ...)
+#define TCP_EVT_CC_PACK_HANDLER(_tc, ...)
 #endif
 
-
 #endif /* SRC_VNET_TCP_TCP_DEBUG_H_ */
 /*
  * fd.io coding-style-patch-verification: ON
diff --git a/src/vnet/tcp/tcp_input.c b/src/vnet/tcp/tcp_input.c
index ff2229b3..a2e6dad1 100644
--- a/src/vnet/tcp/tcp_input.c
+++ b/src/vnet/tcp/tcp_input.c
@@ -378,16 +378,20 @@ tcp_rcv_ack_is_acceptable (tcp_connection_t * tc0, vlib_buffer_t * tb0)
 static void
 tcp_estimate_rtt (tcp_connection_t * tc, u32 mrtt)
 {
-  int err;
+  int err, diff;
 
   if (tc->srtt != 0)
     {
       err = mrtt - tc->srtt;
-      tc->srtt += err >> 3;
+//      tc->srtt += err >> 3;
 
       /* XXX Drop in RTT results in RTTVAR increase and bigger RTO.
        * The increase should be bound */
-      tc->rttvar += ((int) clib_abs (err) - (int) tc->rttvar) >> 2;
+//      tc->rttvar += ((int) clib_abs (err) - (int) tc->rttvar) >> 2;
+
+      tc->srtt = clib_max ((int) tc->srtt + (err >> 3), 1);
+      diff = (clib_abs (err) - (int) tc->rttvar) >> 2;
+      tc->rttvar = clib_max ((int) tc->rttvar + diff, 1);
     }
   else
     {
@@ -401,6 +405,7 @@ void
 tcp_update_rto (tcp_connection_t * tc)
 {
   tc->rto = clib_min (tc->srtt + (tc->rttvar << 2), TCP_RTO_MAX);
+  tc->rto = clib_max (tc->rto, TCP_RTO_MIN);
 }
 
 /** Update RTT estimate and RTO timer
@@ -417,8 +422,8 @@ tcp_update_rtt (tcp_connection_t * tc, u32 ack)
   u32 mrtt = 0;
   u8 rtx_acked;
 
-  /* Determine if only rtx bytes are acked. TODO XXX fast retransmit */
-  rtx_acked = tc->rto_boff && (tc->bytes_acked <= tc->snd_mss);
+  /* Determine if only rtx bytes are acked. */
+  rtx_acked = tcp_in_cong_recovery (tc) || !tc->bytes_acked;
 
   /* Karn's rule, part 1. Don't use retransmitted segments to estimate
    * RTT because they're ambiguous. */
@@ -428,8 +433,7 @@ tcp_update_rtt (tcp_connection_t * tc, u32 ack)
     }
   /* As per RFC7323 TSecr can be used for RTTM only if the segment advances
    * snd_una, i.e., the left side of the send window:
-   * seq_lt (tc->snd_una, ack). Note: last condition could be dropped, we don't
-   * try to update rtt for dupacks */
+   * seq_lt (tc->snd_una, ack). */
   else if (tcp_opts_tstamp (&tc->rcv_opts) && tc->rcv_opts.tsecr
 	   && tc->bytes_acked)
     {
@@ -550,11 +554,13 @@ scoreboard_insert_hole (sack_scoreboard_t * sb, u32 prev_index,
   prev = scoreboard_get_hole (sb, prev_index);
   if (prev)
     {
-      hole->prev = prev - sb->holes;
+      hole->prev = prev_index;
       hole->next = prev->next;
 
       if ((next = scoreboard_next_hole (sb, hole)))
 	next->prev = hole_index;
+      else
+	sb->tail = hole_index;
 
       prev->next = hole_index;
     }
@@ -569,12 +575,13 @@ scoreboard_insert_hole (sack_scoreboard_t * sb, u32 prev_index,
 }
 
 void
-scoreboard_update_lost (tcp_connection_t * tc, sack_scoreboard_t * sb)
+scoreboard_update_bytes (tcp_connection_t * tc, sack_scoreboard_t * sb)
 {
   sack_scoreboard_hole_t *hole, *prev;
   u32 bytes = 0, blks = 0;
 
   sb->lost_bytes = 0;
+  sb->sacked_bytes = 0;
   hole = scoreboard_last_hole (sb);
   if (!hole)
     return;
@@ -594,13 +601,16 @@ scoreboard_update_lost (tcp_connection_t * tc, sack_scoreboard_t * sb)
       hole = prev;
     }
 
-  hole = prev;
   while (hole)
     {
       sb->lost_bytes += scoreboard_hole_bytes (hole);
       hole->is_lost = 1;
+      prev = hole;
       hole = scoreboard_prev_hole (sb, hole);
+      if (hole)
+	bytes += prev->start - hole->end;
     }
+  sb->sacked_bytes = bytes;
 }
 
 /**
@@ -677,7 +687,7 @@ tcp_rcv_sacks (tcp_connection_t * tc, u32 ack)
 {
   sack_scoreboard_t *sb = &tc->sack_sb;
   sack_block_t *blk, tmp;
-  sack_scoreboard_hole_t *hole, *next_hole, *last_hole, *new_hole;
+  sack_scoreboard_hole_t *hole, *next_hole, *last_hole;
   u32 blk_index = 0, old_sacked_bytes, hole_index;
   int i, j;
 
@@ -743,6 +753,10 @@ tcp_rcv_sacks (tcp_connection_t * tc, u32 ack)
       if (seq_gt (tc->snd_una_max, sb->high_sacked)
 	  && seq_gt (tc->snd_una_max, last_hole->end))
 	last_hole->end = tc->snd_una_max;
+      /* keep track of max byte sacked for when the last hole
+       * is acked */
+      if (seq_gt (tmp.end, sb->high_sacked))
+	sb->high_sacked = tmp.end;
     }
 
   /* Walk the holes with the SACK blocks */
@@ -758,45 +772,20 @@ tcp_rcv_sacks (tcp_connection_t * tc, u32 ack)
 	    {
 	      next_hole = scoreboard_next_hole (sb, hole);
 
-	      /* Byte accounting */
-	      if (seq_leq (hole->end, ack))
-		{
-		  /* Bytes lost because snd_wnd left edge advances */
-		  if (next_hole && seq_leq (next_hole->start, ack))
-		    sb->last_bytes_delivered += next_hole->start - hole->end;
-		  else
-		    sb->last_bytes_delivered += ack - hole->end;
-		}
-	      else
-		{
-		  sb->sacked_bytes += scoreboard_hole_bytes (hole);
-		}
-
-	      /* About to remove last hole */
-	      if (hole == last_hole)
-		{
-		  sb->tail = hole->prev;
-		  last_hole = scoreboard_last_hole (sb);
-		  /* keep track of max byte sacked for when the last hole
-		   * is acked */
-		  if (seq_gt (hole->end, sb->high_sacked))
-		    sb->high_sacked = hole->end;
-		}
-
-	      /* snd_una needs to be advanced */
-	      if (blk->end == ack && seq_geq (ack, hole->end))
+	      /* Byte accounting: snd_una needs to be advanced */
+	      if (blk->end == ack)
 		{
-		  if (next_hole && seq_lt (ack, next_hole->start))
+		  if (next_hole)
 		    {
-		      sb->snd_una_adv = next_hole->start - ack;
-
-		      /* all these can be delivered */
-		      sb->last_bytes_delivered += sb->snd_una_adv;
+		      if (seq_lt (ack, next_hole->start))
+			sb->snd_una_adv = next_hole->start - ack;
+		      sb->last_bytes_delivered +=
+			next_hole->start - hole->end;
 		    }
 		  else if (!next_hole)
 		    {
 		      sb->snd_una_adv = sb->high_sacked - ack;
-		      sb->last_bytes_delivered += sb->snd_una_adv;
+		      sb->last_bytes_delivered += sb->high_sacked - hole->end;
 		    }
 		}
 
@@ -808,7 +797,6 @@ tcp_rcv_sacks (tcp_connection_t * tc, u32 ack)
 	    {
 	      if (seq_gt (blk->end, hole->start))
 		{
-		  sb->sacked_bytes += blk->end - hole->start;
 		  hole->start = blk->end;
 		}
 	      blk_index++;
@@ -819,28 +807,16 @@ tcp_rcv_sacks (tcp_connection_t * tc, u32 ack)
 	  /* Hole must be split */
 	  if (seq_lt (blk->end, hole->end))
 	    {
-	      sb->sacked_bytes += blk->end - blk->start;
 	      hole_index = scoreboard_hole_index (sb, hole);
-	      new_hole = scoreboard_insert_hole (sb, hole_index, blk->end,
-						 hole->end);
+	      scoreboard_insert_hole (sb, hole_index, blk->end, hole->end);
 
 	      /* Pool might've moved */
 	      hole = scoreboard_get_hole (sb, hole_index);
 	      hole->end = blk->start;
-
-	      /* New or split of tail */
-	      if ((last_hole->end == new_hole->end)
-		  || seq_lt (last_hole->end, new_hole->start))
-		{
-		  last_hole = new_hole;
-		  sb->tail = scoreboard_hole_index (sb, new_hole);
-		}
-
 	      blk_index++;
 	    }
-	  else if (seq_leq (blk->start, hole->end))
+	  else if (seq_lt (blk->start, hole->end))
 	    {
-	      sb->sacked_bytes += hole->end - blk->start;
 	      hole->end = blk->start;
 	    }
 
@@ -848,9 +824,13 @@ tcp_rcv_sacks (tcp_connection_t * tc, u32 ack)
 	}
     }
 
-  sb->last_sacked_bytes = sb->sacked_bytes - old_sacked_bytes;
-  sb->sacked_bytes -= sb->last_bytes_delivered;
-  scoreboard_update_lost (tc, sb);
+  scoreboard_update_bytes (tc, sb);
+  sb->last_sacked_bytes = sb->sacked_bytes
+    - (old_sacked_bytes - sb->last_bytes_delivered);
+  ASSERT (sb->sacked_bytes == 0
+	  || sb->sacked_bytes < tc->snd_una_max - seq_max (tc->snd_una, ack));
+  ASSERT (sb->last_sacked_bytes + sb->lost_bytes <= tc->snd_una_max
+	  - seq_max (tc->snd_una, ack));
 }
 
 /**
@@ -998,9 +978,14 @@ tcp_should_fastrecover (tcp_connection_t * tc)
 	  || tcp_should_fastrecover_sack (tc));
 }
 
+/**
+ * One function to rule them all ... and in the darkness bind them
+ */
 static void
 tcp_cc_handle_event (tcp_connection_t * tc, u32 is_dack)
 {
+  u32 rxt_delivered;
+
   /*
    * Duplicate ACK. Check if we should enter fast recovery, or if already in
    * it account for the bytes that left the network.
@@ -1028,10 +1013,15 @@ tcp_cc_handle_event (tcp_connection_t * tc, u32 is_dack)
 	    }
 
 	  /* If of of the two conditions lower hold, reset dupacks
-	   * 1) Cumulative ack does not cover more than congestion threshold
+	   * 1) Cumulative ack does not cover more than congestion threshold,
+	   *    and the following doesn't hold: the congestion window is
+	   *    greater than SMSS bytes and the difference between highest_ack
+	   *    and prev_highest_ack is at most 4*SMSS bytes (XXX)
 	   * 2) RFC6582 heuristic to avoid multiple fast retransmits
 	   */
-	  if (seq_leq (tc->snd_una, tc->snd_congestion)
+	  if ((seq_gt (tc->snd_una, tc->snd_congestion)
+	       || !(tc->cwnd > tc->snd_mss
+		    && tc->bytes_acked <= 4 * tc->snd_mss))
 	      || tc->rcv_opts.tsecr != tc->tsecr_last_ack)
 	    {
 	      tc->rcv_dupacks = 0;
@@ -1089,7 +1079,10 @@ partial_ack:
     {
       /* If spurious return, we've already updated everything */
       if (tcp_cc_recover (tc))
-	return;
+	{
+	  tc->tsecr_last_ack = tc->rcv_opts.tsecr;
+	  return;
+	}
 
       tc->snd_nxt = tc->snd_una_max;
 
@@ -1115,12 +1108,16 @@ partial_ack:
     return;
 
   /* Remove retransmitted bytes that have been delivered */
-  if (tc->sack_sb.last_bytes_delivered
-      && seq_gt (tc->sack_sb.high_rxt, tc->snd_una))
+  ASSERT (tc->bytes_acked + tc->sack_sb.snd_una_adv
+	  >= tc->sack_sb.last_bytes_delivered);
+  rxt_delivered = tc->bytes_acked + tc->sack_sb.snd_una_adv
+    - tc->sack_sb.last_bytes_delivered;
+  if (rxt_delivered && seq_gt (tc->sack_sb.high_rxt, tc->snd_una))
     {
       /* If we have sacks and we haven't gotten an ack beyond high_rxt,
        * remove sacked bytes delivered */
-      tc->snd_rxt_bytes -= tc->sack_sb.last_bytes_delivered;
+      ASSERT (tc->snd_rxt_bytes >= rxt_delivered);
+      tc->snd_rxt_bytes -= rxt_delivered;
     }
   else
     {
@@ -1154,6 +1151,8 @@ tcp_rcv_ack (tcp_connection_t * tc, vlib_buffer_t * b,
   u32 prev_snd_wnd, prev_snd_una;
   u8 is_dack;
 
+  TCP_EVT_DBG (TCP_EVT_CC_STAT, tc);
+
   /* If the ACK acks something not yet sent (SEG.ACK > SND.NXT) */
   if (PREDICT_FALSE (seq_gt (vnet_buffer (b)->tcp.ack_number, tc->snd_nxt)))
     {
@@ -1282,6 +1281,10 @@ tcp_update_sack_list (tcp_connection_t * tc, u32 start, u32 end)
 	{
 	  vec_add1 (new_list, tc->snd_sacks[i]);
 	}
+      else
+	{
+	  clib_warning ("sack discarded");
+	}
     }
 
   ASSERT (vec_len (new_list) <= TCP_MAX_SACK_BLOCKS);
@@ -1358,16 +1361,18 @@ tcp_session_enqueue_ooo (tcp_connection_t * tc, vlib_buffer_t * b,
   stream_session_t *s0;
   int rv;
 
+  ASSERT (seq_gt (vnet_buffer (b)->tcp.seq_number, tc->rcv_nxt));
+
   /* Pure ACK. Do nothing */
   if (PREDICT_FALSE (data_len == 0))
     {
       return TCP_ERROR_PURE_ACK;
     }
 
-  /* Enqueue out-of-order data with absolute offset */
+  /* Enqueue out-of-order data with relative offset */
   rv = stream_session_enqueue_data (&tc->connection, b,
-				    vnet_buffer (b)->tcp.seq_number,
-				    0 /* queue event */ , 0);
+				    vnet_buffer (b)->tcp.seq_number -
+				    tc->rcv_nxt, 0 /* queue event */ , 0);
 
   /* Nothing written */
   if (rv)
@@ -1388,10 +1393,15 @@ tcp_session_enqueue_ooo (tcp_connection_t * tc, vlib_buffer_t * b,
 
       /* Get the newest segment from the fifo */
       newest = svm_fifo_newest_ooo_segment (s0->server_rx_fifo);
-      start = ooo_segment_offset (s0->server_rx_fifo, newest);
-      end = ooo_segment_end_offset (s0->server_rx_fifo, newest);
+      if (newest)
+	{
+	  start =
+	    tc->rcv_nxt + ooo_segment_offset (s0->server_rx_fifo, newest);
+	  end = start + ooo_segment_length (s0->server_rx_fifo, newest);
+	  tcp_update_sack_list (tc, start, end);
 
-      tcp_update_sack_list (tc, start, end);
+	  ASSERT (seq_gt (start, tc->rcv_nxt));
+	}
     }
 
   return TCP_ERROR_ENQUEUED;
@@ -1411,7 +1421,7 @@ tcp_can_delack (tcp_connection_t * tc)
       /* constrained to send ack */
       || (tc->flags & TCP_CONN_SNDACK) != 0
       /* we're almost out of tx wnd */
-      || tcp_available_snd_space (tc) < 2 * tc->snd_mss)
+      || tcp_available_snd_space (tc) < 4 * tc->snd_mss)
     return 0;
 
   return 1;
@@ -1434,7 +1444,7 @@ tcp_segment_rcv (tcp_main_t * tm, tcp_connection_t * tc, vlib_buffer_t * b,
 	  *next0 = TCP_NEXT_DROP;
 
 	  /* Completely in the past (possible retransmit) */
-	  if (seq_lt (vnet_buffer (b)->tcp.seq_end, tc->rcv_nxt))
+	  if (seq_leq (vnet_buffer (b)->tcp.seq_end, tc->rcv_nxt))
 	    goto done;
 
 	  /* Chop off the bytes in the past */
@@ -1873,8 +1883,8 @@ tcp46_syn_sent_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
 	  if (tcp_opts_wscale (&new_tc0->rcv_opts))
 	    new_tc0->snd_wscale = new_tc0->rcv_opts.wscale;
 
-	  /* No scaling */
-	  new_tc0->snd_wnd = clib_net_to_host_u16 (tcp0->window);
+	  new_tc0->snd_wnd = clib_net_to_host_u16 (tcp0->window)
+	    << new_tc0->snd_wscale;
 	  new_tc0->snd_wl1 = seq0;
 	  new_tc0->snd_wl2 = ack0;
 
@@ -1892,8 +1902,15 @@ tcp46_syn_sent_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
 	      /* Make sure las is initialized for the wnd computation */
 	      new_tc0->rcv_las = new_tc0->rcv_nxt;
 
-	      /* Notify app that we have connection */
-	      stream_session_connect_notify (&new_tc0->connection, sst, 0);
+	      /* Notify app that we have connection. If session layer can't
+	       * allocate session send reset */
+	      if (stream_session_connect_notify (&new_tc0->connection, sst,
+						 0))
+		{
+		  tcp_connection_cleanup (new_tc0);
+		  tcp_send_reset (b0, is_ip4);
+		  goto drop;
+		}
 
 	      stream_session_init_fifos_pointers (&new_tc0->connection,
 						  new_tc0->irs + 1,
@@ -1907,7 +1924,14 @@ tcp46_syn_sent_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
 	      new_tc0->state = TCP_STATE_SYN_RCVD;
 
 	      /* Notify app that we have connection */
-	      stream_session_connect_notify (&new_tc0->connection, sst, 0);
+	      if (stream_session_connect_notify
+		  (&new_tc0->connection, sst, 0))
+		{
+		  tcp_connection_cleanup (new_tc0);
+		  tcp_send_reset (b0, is_ip4);
+		  goto drop;
+		}
+
 	      stream_session_init_fifos_pointers (&new_tc0->connection,
 						  new_tc0->irs + 1,
 						  new_tc0->iss + 1);
@@ -2508,8 +2532,8 @@ tcp46_listen_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
 	  if (tcp_opts_wscale (&child0->rcv_opts))
 	    child0->snd_wscale = child0->rcv_opts.wscale;
 
-	  /* No scaling */
-	  child0->snd_wnd = clib_net_to_host_u16 (th0->window);
+	  child0->snd_wnd = clib_net_to_host_u16 (th0->window)
+	    << child0->snd_wscale;
 	  child0->snd_wl1 = vnet_buffer (b0)->tcp.seq_number;
 	  child0->snd_wl2 = vnet_buffer (b0)->tcp.ack_number;
 
@@ -2892,6 +2916,9 @@ do {                                                       	\
   _(FIN_WAIT_2, TCP_FLAG_ACK, TCP_INPUT_NEXT_RCV_PROCESS, TCP_ERROR_NONE);
   _(FIN_WAIT_2, TCP_FLAG_FIN | TCP_FLAG_ACK, TCP_INPUT_NEXT_RCV_PROCESS,
     TCP_ERROR_NONE);
+  _(CLOSE_WAIT, TCP_FLAG_ACK, TCP_INPUT_NEXT_RCV_PROCESS, TCP_ERROR_NONE);
+  _(CLOSE_WAIT, TCP_FLAG_FIN | TCP_FLAG_ACK, TCP_INPUT_NEXT_RCV_PROCESS,
+    TCP_ERROR_NONE);
   _(LAST_ACK, TCP_FLAG_ACK, TCP_INPUT_NEXT_RCV_PROCESS, TCP_ERROR_NONE);
   _(LAST_ACK, TCP_FLAG_FIN, TCP_INPUT_NEXT_RCV_PROCESS, TCP_ERROR_NONE);
   _(LAST_ACK, TCP_FLAG_FIN | TCP_FLAG_ACK, TCP_INPUT_NEXT_RCV_PROCESS,
diff --git a/src/vnet/tcp/tcp_newreno.c b/src/vnet/tcp/tcp_newreno.c
index c66250e4..c825e952 100644
--- a/src/vnet/tcp/tcp_newreno.c
+++ b/src/vnet/tcp/tcp_newreno.c
@@ -18,7 +18,6 @@
 void
 newreno_congestion (tcp_connection_t * tc)
 {
-  tc->prev_ssthresh = tc->ssthresh;
   tc->ssthresh = clib_max (tcp_flight_size (tc) / 2, 2 * tc->snd_mss);
 }
 
@@ -47,7 +46,8 @@ newreno_rcv_cong_ack (tcp_connection_t * tc, tcp_cc_ack_t ack_type)
 {
   if (ack_type == TCP_CC_DUPACK)
     {
-      tc->cwnd += tc->snd_mss;
+      if (!tcp_opts_sack_permitted (tc))
+	tc->cwnd += tc->snd_mss;
     }
   else if (ack_type == TCP_CC_PARTIALACK)
     {
diff --git a/src/vnet/tcp/tcp_output.c b/src/vnet/tcp/tcp_output.c
index 47c94e6d..554a981d 100644
--- a/src/vnet/tcp/tcp_output.c
+++ b/src/vnet/tcp/tcp_output.c
@@ -1052,6 +1052,7 @@ tcp_rtx_timeout_cc (tcp_connection_t * tc)
   tc->ssthresh = clib_max (tcp_flight_size (tc) / 2, 2 * tc->snd_mss);
   tc->cwnd = tcp_loss_wnd (tc);
   tc->snd_congestion = tc->snd_una_max;
+
   tcp_recovery_on (tc);
 }
 
@@ -1213,7 +1214,7 @@ tcp_timer_persist_handler (u32 index)
   tc->timers[TCP_TIMER_PERSIST] = TCP_TIMER_HANDLE_INVALID;
 
   /* Problem already solved or worse */
-  if (tc->state == TCP_STATE_CLOSED
+  if (tc->state == TCP_STATE_CLOSED || tc->state > TCP_STATE_ESTABLISHED
       || tc->snd_wnd > tc->snd_mss || tcp_in_recovery (tc))
     return;
 
@@ -1505,10 +1506,7 @@ tcp46_output_inline (vlib_main_t * vm,
 
 	  /* Stop DELACK timer and fix flags */
 	  tc0->flags &= ~(TCP_CONN_SNDACK);
-	  if (tcp_timer_is_active (tc0, TCP_TIMER_DELACK))
-	    {
-	      tcp_timer_reset (tc0, TCP_TIMER_DELACK);
-	    }
+	  tcp_timer_reset (tc0, TCP_TIMER_DELACK);
 
 	  /* If not retransmitting
 	   * 1) update snd_una_max (SYN, SYNACK, FIN)
@@ -1630,7 +1628,7 @@ tcp_push_header (transport_connection_t * tconn, vlib_buffer_t * b)
   tc = (tcp_connection_t *) tconn;
   tcp_push_hdr_i (tc, b, TCP_STATE_ESTABLISHED, 0);
 
-  if (tc->rtt_ts == 0)
+  if (tc->rtt_ts == 0 && !tcp_in_cong_recovery (tc))
     {
       tc->rtt_ts = tcp_time_now ();
       tc->rtt_seq = tc->snd_nxt;
diff --git a/src/vnet/tcp/tcp_test.c b/src/vnet/tcp/tcp_test.c
index 3f8afa40..a461e3b8 100644
--- a/src/vnet/tcp/tcp_test.c
+++ b/src/vnet/tcp/tcp_test.c
@@ -190,11 +190,18 @@ tcp_test_sack_rx (vlib_main_t * vm, unformat_input_t * input)
   TCP_TEST ((sb->sacked_bytes == 0), "sacked bytes %d", sb->sacked_bytes);
   TCP_TEST ((pool_elts (sb->holes) == 1),
 	    "scoreboard has %d elements", pool_elts (sb->holes));
+  hole = scoreboard_first_hole (sb);
+  TCP_TEST ((hole->prev == TCP_INVALID_SACK_HOLE_INDEX
+	     && hole->next == TCP_INVALID_SACK_HOLE_INDEX), "hole is valid");
+  TCP_TEST ((sb->last_bytes_delivered == 100), "last bytes delivered %d",
+	    sb->last_bytes_delivered);
 
   /*
    * Add some more blocks and then remove all
    */
   vec_reset_length (tc->rcv_opts.sacks);
+  tc->snd_una += sb->snd_una_adv;
+  tc->snd_una_max = 1900;
   for (i = 0; i < 5; i++)
     {
       block.start = i * 100 + 1200;
@@ -242,6 +249,39 @@ tcp_test_sack_rx (vlib_main_t * vm, unformat_input_t * input)
   TCP_TEST ((sb->last_sacked_bytes == 0),
 	    "last sacked bytes %d", sb->last_sacked_bytes);
 
+  /*
+   * Inject one block, ack it and overlap hole
+   */
+
+  tc->snd_una = 0;
+  tc->snd_una_max = 1000;
+  tc->snd_nxt = 1000;
+
+  block.start = 100;
+  block.end = 500;
+  vec_add1 (tc->rcv_opts.sacks, block);
+  tc->rcv_opts.n_sack_blocks = vec_len (tc->rcv_opts.sacks);
+
+  tcp_rcv_sacks (tc, 0);
+
+  if (verbose)
+    vlib_cli_output (vm, "sb added [100, 500]:\n%U",
+		     format_tcp_scoreboard, sb);
+
+  tcp_rcv_sacks (tc, 800);
+
+  if (verbose)
+    vlib_cli_output (vm, "sb ack [0, 800]:\n%U", format_tcp_scoreboard, sb);
+
+  TCP_TEST ((pool_elts (sb->holes) == 1),
+	    "scoreboard has %d elements", pool_elts (sb->holes));
+  TCP_TEST ((sb->snd_una_adv == 0), "snd_una_adv %u", sb->snd_una_adv);
+  TCP_TEST ((sb->sacked_bytes == 0), "sacked bytes %d", sb->sacked_bytes);
+  TCP_TEST ((sb->last_sacked_bytes == 0),
+	    "last sacked bytes %d", sb->last_sacked_bytes);
+  TCP_TEST ((sb->last_bytes_delivered == 400),
+	    "last bytes delivered %d", sb->last_bytes_delivered);
+
   return 0;
 }
 
@@ -571,7 +611,7 @@ tcp_test_fifo1 (vlib_main_t * vm, unformat_input_t * input)
    */
   for (i = 0; i < 3; i++)
     {
-      offset = (2 * i + 1) * sizeof (u32);
+      offset = (2 * i + 1) * sizeof (u32) - f->tail;
       data = (u8 *) (test_data + (2 * i + 1));
       if (i == 0)
 	{
@@ -600,7 +640,7 @@ tcp_test_fifo1 (vlib_main_t * vm, unformat_input_t * input)
   /*
    * Try adding a completely overlapped segment
    */
-  offset = 3 * sizeof (u32);
+  offset = 3 * sizeof (u32) - f->tail;
   data = (u8 *) (test_data + 3);
   rv = svm_fifo_enqueue_with_offset (f, offset, sizeof (u32), data);
   if (rv)
@@ -626,7 +666,7 @@ tcp_test_fifo1 (vlib_main_t * vm, unformat_input_t * input)
    */
   for (i = 3; i > 1; i--)
     {
-      offset = (2 * i + 0) * sizeof (u32);
+      offset = (2 * i + 0) * sizeof (u32) - f->tail;
       data = (u8 *) (test_data + (2 * i + 0));
       rv = svm_fifo_enqueue_with_offset (f, offset, sizeof (u32), data);
       if (verbose)
@@ -688,7 +728,7 @@ tcp_test_fifo1 (vlib_main_t * vm, unformat_input_t * input)
 
   for (i = 0; i < 4; i++)
     {
-      offset = (2 * i + 1) * sizeof (u32);
+      offset = (2 * i + 1) * sizeof (u32) - f->tail;
       data = (u8 *) (test_data + (2 * i + 1));
       rv = svm_fifo_enqueue_with_offset (f, offset, sizeof (u32), data);
       if (verbose)
@@ -701,7 +741,7 @@ tcp_test_fifo1 (vlib_main_t * vm, unformat_input_t * input)
 	}
     }
 
-  rv = svm_fifo_enqueue_with_offset (f, 8, 21, data);
+  rv = svm_fifo_enqueue_with_offset (f, 8 - f->tail, 21, data);
   TCP_TEST ((rv == 0), "ooo enqueued %u", rv);
   TCP_TEST ((svm_fifo_number_ooo_segments (f) == 1),
 	    "number of ooo segments %u", svm_fifo_number_ooo_segments (f));
@@ -722,7 +762,7 @@ tcp_test_fifo1 (vlib_main_t * vm, unformat_input_t * input)
 
   for (i = 0; i < 4; i++)
     {
-      offset = (2 * i + 1) * sizeof (u32);
+      offset = (2 * i + 1) * sizeof (u32) - f->tail;
       data = (u8 *) (test_data + (2 * i + 1));
       rv = svm_fifo_enqueue_with_offset (f, offset, sizeof (u32), data);
       if (verbose)
@@ -735,7 +775,13 @@ tcp_test_fifo1 (vlib_main_t * vm, unformat_input_t * input)
 	}
     }
 
+  if (verbose)
+    vlib_cli_output (vm, "fifo after enqueue: %U", format_svm_fifo, f, 1);
+
   rv = svm_fifo_enqueue_nowait (f, 29, data);
+  if (verbose)
+    vlib_cli_output (vm, "fifo after enqueueing 29: %U", format_svm_fifo, f,
+		     1);
   TCP_TEST ((rv == 32), "ooo enqueued %u", rv);
   TCP_TEST ((svm_fifo_number_ooo_segments (f) == 0),
 	    "number of ooo segments %u", svm_fifo_number_ooo_segments (f));
@@ -788,7 +834,8 @@ tcp_test_fifo2 (vlib_main_t * vm)
     {
       tp = vp + i;
       data64 = tp->offset;
-      svm_fifo_enqueue_with_offset (f, tp->offset, tp->len, (u8 *) & data64);
+      svm_fifo_enqueue_with_offset (f, tp->offset - f->tail, tp->len,
+				    (u8 *) & data64);
     }
 
   /* Expected result: one big fat chunk at offset 4 */
@@ -817,7 +864,7 @@ tcp_test_fifo2 (vlib_main_t * vm)
     {
       tp = &test_data[i];
       data64 = tp->offset;
-      rv = svm_fifo_enqueue_with_offset (f, tp->offset, tp->len,
+      rv = svm_fifo_enqueue_with_offset (f, tp->offset - f->tail, tp->len,
 					 (u8 *) & data64);
       if (rv)
 	{
@@ -991,8 +1038,9 @@ tcp_test_fifo3 (vlib_main_t * vm, unformat_input_t * input)
   for (i = !randomize; i < vec_len (generate); i++)
     {
       tp = generate + i;
-      svm_fifo_enqueue_with_offset (f, fifo_initial_offset + tp->offset,
-				    tp->len,
+      svm_fifo_enqueue_with_offset (f,
+				    fifo_initial_offset + tp->offset -
+				    f->tail, tp->len,
 				    (u8 *) data_pattern + tp->offset);
     }
 
@@ -1107,7 +1155,7 @@ tcp_test_fifo4 (vlib_main_t * vm, unformat_input_t * input)
 
   for (i = test_n_bytes - 1; i > 0; i--)
     {
-      rv = svm_fifo_enqueue_with_offset (f, fifo_initial_offset + i,
+      rv = svm_fifo_enqueue_with_offset (f, fifo_initial_offset + i - f->tail,
 					 sizeof (u8), &test_data[i]);
       if (verbose)
 	vlib_cli_output (vm, "add [%d] [%d, %d]", i, i, i + sizeof (u8));
-- 
cgit 1.2.3-korg


From f6359c8cace5b73a813e5f4e3d1bc28f7752fcdf Mon Sep 17 00:00:00 2001
From: Florin Coras <fcoras@cisco.com>
Date: Mon, 19 Jun 2017 12:26:09 -0400
Subject: Improve svm fifo and tcp tx path performance (VPP-846)

- multiarch on svm fifo
- avoid ip lookup on tx

Change-Id: Iab0d85204a710979417bca1d692cc47877131203
Signed-off-by: Florin Coras <fcoras@cisco.com>
Signed-off-by: Dave Barach <dbarach@cisco.com>
---
 src/svm/svm_fifo.c           |  95 +++++++++++++++++++++++++++++++++++++---
 src/uri/uri_tcp_test.c       |  10 ++---
 src/vnet/session/node.c      |  13 +++---
 src/vnet/session/transport.h |   7 ++-
 src/vnet/tcp/tcp.c           | 100 +++++++++++++++++++++++++++++++++++++++++--
 src/vnet/tcp/tcp.h           |   4 ++
 src/vnet/tcp/tcp_output.c    |  29 +++++++++----
 src/vnet/tcp/tcp_packet.h    |   9 ++--
 8 files changed, 232 insertions(+), 35 deletions(-)

(limited to 'src/svm/svm_fifo.c')

diff --git a/src/svm/svm_fifo.c b/src/svm/svm_fifo.c
index 6ca437cf..aed5d6a7 100644
--- a/src/svm/svm_fifo.c
+++ b/src/svm/svm_fifo.c
@@ -14,6 +14,7 @@
  */
 
 #include <svm/svm_fifo.h>
+#include <vppinfra/cpu.h>
 
 static inline u8
 position_lt (svm_fifo_t * f, u32 a, u32 b)
@@ -417,10 +418,38 @@ svm_fifo_enqueue_internal (svm_fifo_t * f, u32 max_bytes, u8 * copy_from_here)
   return (total_copy_bytes);
 }
 
+#define SVM_ENQUEUE_CLONE_TEMPLATE(arch, fn, tgt)                       \
+  uword                                                                 \
+  __attribute__ ((flatten))                                             \
+  __attribute__ ((target (tgt)))                                        \
+  CLIB_CPU_OPTIMIZED                                                    \
+  fn ## _ ## arch ( svm_fifo_t * f, u32 max_bytes, u8 * copy_from_here) \
+  { return fn (f, max_bytes, copy_from_here);}
+
+static int
+svm_fifo_enqueue_nowait_ma (svm_fifo_t * f, u32 max_bytes,
+			    u8 * copy_from_here)
+{
+  return svm_fifo_enqueue_internal (f, max_bytes, copy_from_here);
+}
+
+foreach_march_variant (SVM_ENQUEUE_CLONE_TEMPLATE,
+		       svm_fifo_enqueue_nowait_ma);
+CLIB_MULTIARCH_SELECT_FN (svm_fifo_enqueue_nowait_ma);
+
 int
 svm_fifo_enqueue_nowait (svm_fifo_t * f, u32 max_bytes, u8 * copy_from_here)
 {
-  return svm_fifo_enqueue_internal (f, max_bytes, copy_from_here);
+#if CLIB_DEBUG > 0
+  return svm_fifo_enqueue_nowait_ma (f, max_bytes, copy_from_here);
+#else
+  static int (*fp) (svm_fifo_t *, u32, u8 *);
+
+  if (PREDICT_FALSE (fp == 0))
+    fp = (void *) svm_fifo_enqueue_nowait_ma_multiarch_select ();
+
+  return (*fp) (f, max_bytes, copy_from_here);
+#endif
 }
 
 /**
@@ -541,15 +570,43 @@ svm_fifo_dequeue_internal (svm_fifo_t * f, u32 max_bytes, u8 * copy_here)
   return (total_copy_bytes);
 }
 
-int
-svm_fifo_dequeue_nowait (svm_fifo_t * f, u32 max_bytes, u8 * copy_here)
+static int
+svm_fifo_dequeue_nowait_ma (svm_fifo_t * f, u32 max_bytes, u8 * copy_here)
 {
   return svm_fifo_dequeue_internal (f, max_bytes, copy_here);
 }
 
+#define SVM_FIFO_DEQUEUE_CLONE_TEMPLATE(arch, fn, tgt)          \
+  uword                                                         \
+  __attribute__ ((flatten))                                     \
+  __attribute__ ((target (tgt)))                                \
+  CLIB_CPU_OPTIMIZED                                            \
+  fn ## _ ## arch ( svm_fifo_t * f, u32 max_bytes,              \
+                    u8 * copy_here)                             \
+  { return fn (f, max_bytes, copy_here);}
+
+foreach_march_variant (SVM_FIFO_DEQUEUE_CLONE_TEMPLATE,
+		       svm_fifo_dequeue_nowait_ma);
+CLIB_MULTIARCH_SELECT_FN (svm_fifo_dequeue_nowait_ma);
+
 int
-svm_fifo_peek (svm_fifo_t * f, u32 relative_offset, u32 max_bytes,
-	       u8 * copy_here)
+svm_fifo_dequeue_nowait (svm_fifo_t * f, u32 max_bytes, u8 * copy_here)
+{
+#if CLIB_DEBUG > 0
+  return svm_fifo_dequeue_nowait_ma (f, max_bytes, copy_here);
+#else
+  static int (*fp) (svm_fifo_t *, u32, u8 *);
+
+  if (PREDICT_FALSE (fp == 0))
+    fp = (void *) svm_fifo_dequeue_nowait_ma_multiarch_select ();
+
+  return (*fp) (f, max_bytes, copy_here);
+#endif
+}
+
+static int
+svm_fifo_peek_ma (svm_fifo_t * f, u32 relative_offset, u32 max_bytes,
+		  u8 * copy_here)
 {
   u32 total_copy_bytes, first_copy_bytes, second_copy_bytes;
   u32 cursize, nitems, real_head;
@@ -586,6 +643,34 @@ svm_fifo_peek (svm_fifo_t * f, u32 relative_offset, u32 max_bytes,
   return total_copy_bytes;
 }
 
+#define SVM_FIFO_PEEK_CLONE_TEMPLATE(arch, fn, tgt)                     \
+  uword                                                                 \
+  __attribute__ ((flatten))                                             \
+  __attribute__ ((target (tgt)))                                        \
+  CLIB_CPU_OPTIMIZED                                                    \
+  fn ## _ ## arch ( svm_fifo_t * f, u32 relative_offset, u32 max_bytes, \
+                    u8 * copy_here)                                     \
+  { return fn (f, relative_offset, max_bytes, copy_here);}
+
+foreach_march_variant (SVM_FIFO_PEEK_CLONE_TEMPLATE, svm_fifo_peek_ma);
+CLIB_MULTIARCH_SELECT_FN (svm_fifo_peek_ma);
+
+int
+svm_fifo_peek (svm_fifo_t * f, u32 relative_offset, u32 max_bytes,
+	       u8 * copy_here)
+{
+#if CLIB_DEBUG > 0
+  return svm_fifo_peek_ma (f, relative_offset, max_bytes, copy_here);
+#else
+  static int (*fp) (svm_fifo_t *, u32, u32, u8 *);
+
+  if (PREDICT_FALSE (fp == 0))
+    fp = (void *) svm_fifo_peek_ma_multiarch_select ();
+
+  return (*fp) (f, relative_offset, max_bytes, copy_here);
+#endif
+}
+
 int
 svm_fifo_dequeue_drop (svm_fifo_t * f, u32 max_bytes)
 {
diff --git a/src/uri/uri_tcp_test.c b/src/uri/uri_tcp_test.c
index d1694cf4..80aab183 100755
--- a/src/uri/uri_tcp_test.c
+++ b/src/uri/uri_tcp_test.c
@@ -398,7 +398,6 @@ static void
 vl_api_reset_session_t_handler (vl_api_reset_session_t * mp)
 {
   uri_tcp_test_main_t *utm = &uri_tcp_test_main;
-  session_t *session;
   vl_api_reset_session_reply_t *rmp;
   uword *p;
   int rv = 0;
@@ -407,9 +406,8 @@ vl_api_reset_session_t_handler (vl_api_reset_session_t * mp)
 
   if (p)
     {
-      session = pool_elt_at_index (utm->sessions, p[0]);
-      hash_unset (utm->session_index_by_vpp_handles, mp->handle);
-      pool_put (utm->sessions, session);
+      clib_warning ("got reset");
+      /* Cleanup later */
       utm->time_to_stop = 1;
     }
   else
@@ -603,7 +601,7 @@ send_test_chunk (uri_tcp_test_main_t * utm, svm_fifo_t * tx_fifo, int mypid,
   if (bytes_to_snd > vec_len (test_data))
     bytes_to_snd = vec_len (test_data);
 
-  while (bytes_to_snd > 0)
+  while (bytes_to_snd > 0 && !utm->time_to_stop)
     {
       actual_write = (bytes_to_snd > queue_max_chunk) ?
 	queue_max_chunk : bytes_to_snd;
@@ -652,6 +650,8 @@ client_send_data (uri_tcp_test_main_t * utm)
   for (i = 0; i < n_iterations; i++)
     {
       send_test_chunk (utm, tx_fifo, mypid, 0);
+      if (utm->time_to_stop)
+	break;
     }
 
   leftover = utm->bytes_to_send % vec_len (test_data);
diff --git a/src/vnet/session/node.c b/src/vnet/session/node.c
index c0ab1bf0..b24f5fd9 100644
--- a/src/vnet/session/node.c
+++ b/src/vnet/session/node.c
@@ -248,6 +248,11 @@ session_tx_fifo_read_and_snd_i (vlib_main_t * vm, vlib_node_runtime_t * node,
 	  ASSERT (bi0);
 	  _vec_len (smm->tx_buffers[thread_index]) = n_bufs;
 
+	  /* usual speculation, or the enqueue_x1 macro will barf */
+	  to_next[0] = bi0;
+	  to_next += 1;
+	  n_left_to_next -= 1;
+
 	  b0 = vlib_get_buffer (vm, bi0);
 	  b0->error = 0;
 	  b0->flags = VLIB_BUFFER_TOTAL_LENGTH_VALID
@@ -255,10 +260,6 @@ session_tx_fifo_read_and_snd_i (vlib_main_t * vm, vlib_node_runtime_t * node,
 	  b0->current_data = 0;
 	  b0->total_length_not_including_first_buffer = 0;
 
-	  /* RX on the local interface. tx in default fib */
-	  vnet_buffer (b0)->sw_if_index[VLIB_RX] = 0;
-	  vnet_buffer (b0)->sw_if_index[VLIB_TX] = (u32) ~ 0;
-
 	  len_to_deq0 = clib_min (left_to_snd0, deq_per_buf);
 
 	  data0 = vlib_buffer_make_headroom (b0, MAX_HDRS_LEN);
@@ -307,10 +308,6 @@ session_tx_fifo_read_and_snd_i (vlib_main_t * vm, vlib_node_runtime_t * node,
 	  }));
 	  /* *INDENT-ON* */
 
-	  /* usual speculation, or the enqueue_x1 macro will barf */
-	  to_next[0] = bi0;
-	  to_next += 1;
-	  n_left_to_next -= 1;
 
 	  VLIB_BUFFER_TRACE_TRAJECTORY_INIT (b0);
 	  if (PREDICT_FALSE (n_trace > 0))
diff --git a/src/vnet/session/transport.h b/src/vnet/session/transport.h
index 04bd5ca0..561a9257 100644
--- a/src/vnet/session/transport.h
+++ b/src/vnet/session/transport.h
@@ -30,13 +30,16 @@ typedef struct _transport_connection
   ip46_address_t lcl_ip;	/**< Local IP */
   u16 lcl_port;			/**< Local port */
   u16 rmt_port;			/**< Remote port */
-  u8 proto;			/**< Transport protocol id (also session type) */
+  u8 proto;			/**< Protocol id (also session type) */
 
   u32 s_index;			/**< Parent session index */
   u32 c_index;			/**< Connection index in transport pool */
   u8 is_ip4;			/**< Flag if IP4 connection */
   u32 thread_index;		/**< Worker-thread index */
 
+  fib_node_index_t rmt_fei;	/**< FIB entry index for rmt */
+  dpo_id_t rmt_dpo;		/**< Forwarding DPO for rmt */
+
 #if TRANSPORT_DEBUG
   elog_track_t elog_track;	/**< Event logging */
   u32 cc_stat_tstamp;		/**< CC stats timestamp */
@@ -59,6 +62,8 @@ typedef struct _transport_connection
 #define c_thread_index connection.thread_index
 #define c_elog_track connection.elog_track
 #define c_cc_stat_tstamp connection.cc_stat_tstamp
+#define c_rmt_fei connection.rmt_fei
+#define c_rmt_dpo connection.rmt_dpo
 } transport_connection_t;
 
 /*
diff --git a/src/vnet/tcp/tcp.c b/src/vnet/tcp/tcp.c
index 5c554bac..4e85eb3f 100644
--- a/src/vnet/tcp/tcp.c
+++ b/src/vnet/tcp/tcp.c
@@ -16,6 +16,7 @@
 #include <vnet/tcp/tcp.h>
 #include <vnet/session/session.h>
 #include <vnet/fib/fib.h>
+#include <vnet/dpo/load_balance.h>
 #include <math.h>
 
 tcp_main_t tcp_main;
@@ -342,6 +343,99 @@ tcp_connection_timers_reset (tcp_connection_t * tc)
     }
 }
 
+typedef struct ip4_tcp_hdr
+{
+  ip4_header_t ip;
+  tcp_header_t tcp;
+} ip4_tcp_hdr_t;
+
+typedef struct ip6_tcp_hdr
+{
+  ip6_header_t ip;
+  tcp_header_t tcp;
+} ip6_tcp_hdr_t;
+
+static void
+tcp_connection_select_lb_bucket (tcp_connection_t * tc, const dpo_id_t * dpo,
+				 dpo_id_t * result)
+{
+  const dpo_id_t *choice;
+  load_balance_t *lb;
+  int hash;
+
+  lb = load_balance_get (dpo->dpoi_index);
+  if (tc->c_is_ip4)
+    {
+      ip4_tcp_hdr_t hdr;
+      memset (&hdr, 0, sizeof (hdr));
+      hdr.ip.protocol = IP_PROTOCOL_TCP;
+      hdr.ip.address_pair.src.as_u32 = tc->c_lcl_ip.ip4.as_u32;
+      hdr.ip.address_pair.dst.as_u32 = tc->c_rmt_ip.ip4.as_u32;
+      hdr.tcp.src_port = tc->c_lcl_port;
+      hdr.tcp.dst_port = tc->c_rmt_port;
+      hash = ip4_compute_flow_hash (&hdr.ip, lb->lb_hash_config);
+    }
+  else
+    {
+      ip6_tcp_hdr_t hdr;
+      memset (&hdr, 0, sizeof (hdr));
+      hdr.ip.protocol = IP_PROTOCOL_TCP;
+      clib_memcpy (&hdr.ip.src_address, &tc->c_lcl_ip.ip6,
+		   sizeof (ip6_address_t));
+      clib_memcpy (&hdr.ip.dst_address, &tc->c_rmt_ip.ip6,
+		   sizeof (ip6_address_t));
+      hdr.tcp.src_port = tc->c_lcl_port;
+      hdr.tcp.dst_port = tc->c_rmt_port;
+      hash = ip6_compute_flow_hash (&hdr.ip, lb->lb_hash_config);
+    }
+  choice = load_balance_get_bucket_i (lb, hash & lb->lb_n_buckets_minus_1);
+  dpo_copy (result, choice);
+}
+
+fib_node_index_t
+tcp_lookup_rmt_in_fib (tcp_connection_t * tc)
+{
+  fib_prefix_t prefix;
+
+  clib_memcpy (&prefix.fp_addr, &tc->c_rmt_ip, sizeof (prefix.fp_addr));
+  prefix.fp_proto = tc->c_is_ip4 ? FIB_PROTOCOL_IP4 : FIB_PROTOCOL_IP6;
+  prefix.fp_len = tc->c_is_ip4 ? 32 : 128;
+  return fib_table_lookup (0, &prefix);
+}
+
+static int
+tcp_connection_stack_on_fib_entry (tcp_connection_t * tc)
+{
+  dpo_id_t choice = DPO_INVALID;
+  u32 output_node_index;
+  fib_entry_t *fe;
+
+  fe = fib_entry_get (tc->c_rmt_fei);
+  if (fe->fe_lb.dpoi_type != DPO_LOAD_BALANCE)
+    return -1;
+
+  tcp_connection_select_lb_bucket (tc, &fe->fe_lb, &choice);
+
+  output_node_index =
+    tc->c_is_ip4 ? tcp4_output_node.index : tcp6_output_node.index;
+  dpo_stack_from_node (output_node_index, &tc->c_rmt_dpo, &choice);
+  return 0;
+}
+
+/** Stack tcp connection on peer's fib entry.
+ *
+ * This ultimately populates the dpo the connection will use to send packets.
+ */
+static void
+tcp_connection_fib_attach (tcp_connection_t * tc)
+{
+  tc->c_rmt_fei = tcp_lookup_rmt_in_fib (tc);
+
+  ASSERT (tc->c_rmt_fei != FIB_NODE_INDEX_INVALID);
+
+  tcp_connection_stack_on_fib_entry (tc);
+}
+
 /** Initialize tcp connection variables
  *
  * Should be called after having received a msg from the peer, i.e., a SYN or
@@ -353,6 +447,7 @@ tcp_connection_init_vars (tcp_connection_t * tc)
   tcp_init_mss (tc);
   scoreboard_init (&tc->sack_sb);
   tcp_cc_init (tc);
+  tcp_connection_fib_attach (tc);
 }
 
 int
@@ -361,7 +456,8 @@ tcp_connection_open (ip46_address_t * rmt_addr, u16 rmt_port, u8 is_ip4)
   tcp_main_t *tm = vnet_get_tcp_main ();
   tcp_connection_t *tc;
   fib_prefix_t prefix;
-  u32 fei, sw_if_index;
+  fib_node_index_t fei;
+  u32 sw_if_index;
   ip46_address_t lcl_addr;
   u16 lcl_port;
 
@@ -985,8 +1081,6 @@ tcp_main_enable (vlib_main_t * vm)
   vec_validate (tm->timer_wheels, num_threads - 1);
   tcp_initialize_timer_wheels (tm);
 
-//  vec_validate (tm->delack_connections, num_threads - 1);
-
   /* Initialize clocks per tick for TCP timestamp. Used to compute
    * monotonically increasing timestamps. */
   tm->tstamp_ticks_per_clock = vm->clib_time.seconds_per_clock
diff --git a/src/vnet/tcp/tcp.h b/src/vnet/tcp/tcp.h
index e8398718..12d804b8 100644
--- a/src/vnet/tcp/tcp.h
+++ b/src/vnet/tcp/tcp.h
@@ -28,6 +28,7 @@
 #define THZ (u32) (1/TCP_TICK)		/**< TCP tick frequency */
 #define TCP_TSTAMP_RESOLUTION TCP_TICK	/**< Time stamp resolution */
 #define TCP_PAWS_IDLE 24 * 24 * 60 * 60 * THZ /**< 24 days */
+#define TCP_FIB_RECHECK_PERIOD	1 * THZ	/**< Recheck every 1s */
 #define TCP_MAX_OPTION_SPACE 40
 
 #define TCP_DUPACK_THRESHOLD 	3
@@ -256,6 +257,7 @@ typedef struct _tcp_connection
 
   u16 mss;		/**< Our max seg size that includes options */
   u32 limited_transmit;	/**< snd_nxt when limited transmit starts */
+  u32 last_fib_check;	/**< Last time we checked fib route for peer */
 } tcp_connection_t;
 
 struct _tcp_cc_algorithm
@@ -528,6 +530,8 @@ void tcp_cc_init_congestion (tcp_connection_t * tc);
 int tcp_cc_recover (tcp_connection_t * tc);
 void tcp_cc_fastrecovery_exit (tcp_connection_t * tc);
 
+fib_node_index_t tcp_lookup_rmt_in_fib (tcp_connection_t * tc);
+
 /* Made public for unit testing only */
 void tcp_update_sack_list (tcp_connection_t * tc, u32 start, u32 end);
 
diff --git a/src/vnet/tcp/tcp_output.c b/src/vnet/tcp/tcp_output.c
index 554a981d..41bebcb3 100644
--- a/src/vnet/tcp/tcp_output.c
+++ b/src/vnet/tcp/tcp_output.c
@@ -22,17 +22,14 @@ vlib_node_registration_t tcp6_output_node;
 typedef enum _tcp_output_nect
 {
   TCP_OUTPUT_NEXT_DROP,
-  TCP_OUTPUT_NEXT_IP_LOOKUP,
   TCP_OUTPUT_N_NEXT
 } tcp_output_next_t;
 
 #define foreach_tcp4_output_next              	\
   _ (DROP, "error-drop")                        \
-  _ (IP_LOOKUP, "ip4-lookup")
 
 #define foreach_tcp6_output_next              	\
   _ (DROP, "error-drop")                        \
-  _ (IP_LOOKUP, "ip6-lookup")
 
 static char *tcp_error_strings[] = {
 #define tcp_error(n,s) s,
@@ -1451,7 +1448,7 @@ tcp46_output_inline (vlib_main_t * vm,
 	  tcp_connection_t *tc0;
 	  tcp_tx_trace_t *t0;
 	  tcp_header_t *th0 = 0;
-	  u32 error0 = TCP_ERROR_PKTS_SENT, next0 = TCP_OUTPUT_NEXT_IP_LOOKUP;
+	  u32 error0 = TCP_ERROR_PKTS_SENT, next0 = TCP_OUTPUT_NEXT_DROP;
 
 	  bi0 = from[0];
 	  to_next[0] = bi0;
@@ -1530,10 +1527,26 @@ tcp46_output_inline (vlib_main_t * vm,
 	      tc0->rto_boff = 0;
 	    }
 
-	  /* set fib index to default and lookup node */
-	  /* XXX network virtualization (vrf/vni) */
-	  vnet_buffer (b0)->sw_if_index[VLIB_RX] = 0;
-	  vnet_buffer (b0)->sw_if_index[VLIB_TX] = (u32) ~ 0;
+	  /* Make sure we haven't lost route to our peer */
+	  if (PREDICT_FALSE (tc0->last_fib_check
+			     < tc0->snd_opts.tsval + TCP_FIB_RECHECK_PERIOD))
+	    {
+	      if (PREDICT_TRUE
+		  (tc0->c_rmt_fei == tcp_lookup_rmt_in_fib (tc0)))
+		{
+		  tc0->last_fib_check = tc0->snd_opts.tsval;
+		}
+	      else
+		{
+		  clib_warning ("lost connection to peer");
+		  tcp_connection_reset (tc0);
+		  goto done;
+		}
+	    }
+
+	  /* Use pre-computed dpo to set next node */
+	  next0 = tc0->c_rmt_dpo.dpoi_next_node;
+	  vnet_buffer (b0)->ip.adj_index[VLIB_TX] = tc0->c_rmt_dpo.dpoi_index;
 
 	  b0->flags |= VNET_BUFFER_LOCALLY_ORIGINATED;
 	done:
diff --git a/src/vnet/tcp/tcp_packet.h b/src/vnet/tcp/tcp_packet.h
index 4f28cf32..a6f62ee1 100644
--- a/src/vnet/tcp/tcp_packet.h
+++ b/src/vnet/tcp/tcp_packet.h
@@ -144,12 +144,11 @@ typedef struct
 {
   u8 flags;		/** Option flags, see above */
 
-  /* Received options */
-  u16 mss;		/**< Maximum segment size advertised by peer */
-  u8 wscale;		/**< Window scale advertised by peer */
-  u32 tsval;		/**< Peer's timestamp value */
+  u16 mss;		/**< Maximum segment size advertised */
+  u8 wscale;		/**< Window scale advertised */
+  u32 tsval;		/**< Timestamp value */
   u32 tsecr;		/**< Echoed/reflected time stamp */
-  sack_block_t *sacks;	/**< SACK blocks received */
+  sack_block_t *sacks;	/**< SACK blocks */
   u8 n_sack_blocks;	/**< Number of SACKs blocks */
 } tcp_options_t;
 
-- 
cgit 1.2.3-korg


From 2c25a62cc1cc4937165de740a3b32d78429c72d6 Mon Sep 17 00:00:00 2001
From: Dave Barach <dbarach@cisco.com>
Date: Mon, 26 Jun 2017 11:35:07 -0400
Subject: Horizontal (nSessions) scaling draft

- Data structure preallocation.
- Input state machine fixes for mid-stream 3-way handshake retries.
- Batch connections in the builtin_client
- Multiple private fifo segment support
- Fix elog simultaneous event type registration
- Fix sacks when segment hole is added after highest sacked
- Add "accepting" session state for sessions pending accept
- Add ssvm non-recursive locking
- Estimate RTT for syn-ack
- Don't init fifo pointers. We're using relative offsets for ooo
  segments
- CLI to dump individual session

Change-Id: Ie0598563fd246537bafba4feed7985478ea1d415
Signed-off-by: Dave Barach <dbarach@cisco.com>
Signed-off-by: Florin Coras <fcoras@cisco.com>
---
 src/svm/ssvm.h                           |  17 +++
 src/svm/svm_fifo.c                       |  56 +++++---
 src/svm/svm_fifo.h                       |  16 ++-
 src/svm/svm_fifo_segment.c               | 114 +++++++++++-----
 src/svm/svm_fifo_segment.h               |   4 +-
 src/svm/test_svm_fifo1.c                 |  10 +-
 src/uri/uri_udp_test.c                   |   2 +-
 src/vnet/session/application.c           |   2 +
 src/vnet/session/application_interface.c |  21 ---
 src/vnet/session/application_interface.h |  12 +-
 src/vnet/session/node.c                  |  23 +---
 src/vnet/session/segment_manager.c       |  26 ++--
 src/vnet/session/segment_manager.h       |   4 +
 src/vnet/session/session.c               |  72 +++++++---
 src/vnet/session/session.h               |  30 ++++-
 src/vnet/session/session_cli.c           |  99 +++++++++++---
 src/vnet/session/transport.h             |   6 +
 src/vnet/tcp/builtin_client.c            | 118 +++++++++++-----
 src/vnet/tcp/builtin_client.h            |   7 +-
 src/vnet/tcp/builtin_server.c            |  66 +++++++--
 src/vnet/tcp/tcp.c                       | 225 ++++++++++++++++++++++++++++---
 src/vnet/tcp/tcp.h                       |  13 ++
 src/vnet/tcp/tcp_debug.h                 |  13 +-
 src/vnet/tcp/tcp_input.c                 |  97 ++++++++-----
 src/vnet/tcp/tcp_newreno.c               |   4 +-
 src/vnet/tcp/tcp_output.c                |  53 +++++---
 src/vnet/tcp/tcp_packet.h                |   1 +
 src/vnet/tcp/tcp_test.c                  |  10 +-
 src/vnet/udp/udp_input.c                 |   2 +-
 29 files changed, 838 insertions(+), 285 deletions(-)

(limited to 'src/svm/svm_fifo.c')

diff --git a/src/svm/ssvm.h b/src/svm/ssvm.h
index bccfc164..8466e155 100644
--- a/src/svm/ssvm.h
+++ b/src/svm/ssvm.h
@@ -101,6 +101,15 @@ ssvm_lock (ssvm_shared_header_t * h, u32 my_pid, u32 tag)
   h->tag = tag;
 }
 
+always_inline void
+ssvm_lock_non_recursive (ssvm_shared_header_t * h, u32 tag)
+{
+  while (__sync_lock_test_and_set (&h->lock, 1))
+    ;
+
+  h->tag = tag;
+}
+
 always_inline void
 ssvm_unlock (ssvm_shared_header_t * h)
 {
@@ -113,6 +122,14 @@ ssvm_unlock (ssvm_shared_header_t * h)
     }
 }
 
+always_inline void
+ssvm_unlock_non_recursive (ssvm_shared_header_t * h)
+{
+  h->tag = 0;
+  CLIB_MEMORY_BARRIER ();
+  h->lock = 0;
+}
+
 static inline void *
 ssvm_push_heap (ssvm_shared_header_t * sh)
 {
diff --git a/src/svm/svm_fifo.c b/src/svm/svm_fifo.c
index aed5d6a7..da60fee5 100644
--- a/src/svm/svm_fifo.c
+++ b/src/svm/svm_fifo.c
@@ -19,29 +19,29 @@
 static inline u8
 position_lt (svm_fifo_t * f, u32 a, u32 b)
 {
-  return (ooo_segment_distance_to_tail (f, a)
-	  < ooo_segment_distance_to_tail (f, b));
+  return (ooo_segment_distance_from_tail (f, a)
+	  < ooo_segment_distance_from_tail (f, b));
 }
 
 static inline u8
 position_leq (svm_fifo_t * f, u32 a, u32 b)
 {
-  return (ooo_segment_distance_to_tail (f, a)
-	  <= ooo_segment_distance_to_tail (f, b));
+  return (ooo_segment_distance_from_tail (f, a)
+	  <= ooo_segment_distance_from_tail (f, b));
 }
 
 static inline u8
 position_gt (svm_fifo_t * f, u32 a, u32 b)
 {
-  return (ooo_segment_distance_to_tail (f, a)
-	  > ooo_segment_distance_to_tail (f, b));
+  return (ooo_segment_distance_from_tail (f, a)
+	  > ooo_segment_distance_from_tail (f, b));
 }
 
 static inline u32
 position_diff (svm_fifo_t * f, u32 posa, u32 posb)
 {
-  return ooo_segment_distance_to_tail (f, posa)
-    - ooo_segment_distance_to_tail (f, posb);
+  return ooo_segment_distance_from_tail (f, posa)
+    - ooo_segment_distance_from_tail (f, posb);
 }
 
 static inline u32
@@ -113,7 +113,7 @@ svm_fifo_create (u32 data_size_in_bytes)
   if (f == 0)
     return 0;
 
-  memset (f, 0, sizeof (*f) + data_size_in_bytes);
+  memset (f, 0, sizeof (*f));
   f->nitems = data_size_in_bytes;
   f->ooos_list_head = OOO_SEGMENT_INVALID_INDEX;
 
@@ -204,7 +204,19 @@ ooo_segment_add (svm_fifo_t * f, u32 offset, u32 length)
     {
       s = prev;
       s_end_pos = ooo_segment_end_pos (f, s);
-      goto merge;
+
+      /* Check head and tail now since segment may be wider at both ends so
+       * merge tests lower won't work */
+      if (position_lt (f, normalized_position, s->start))
+	{
+	  s->start = normalized_position;
+	  s->length = position_diff (f, s_end_pos, s->start);
+	}
+      if (position_gt (f, normalized_end_position, s_end_pos))
+	{
+	  s->length = position_diff (f, normalized_end_position, s->start);
+	}
+      goto check_tail;
     }
 
   s_index = s - f->ooo_segments;
@@ -257,8 +269,6 @@ ooo_segment_add (svm_fifo_t * f, u32 offset, u32 length)
    * Merge needed
    */
 
-merge:
-
   /* Merge at head */
   if (position_lt (f, normalized_position, s->start))
     {
@@ -278,6 +288,7 @@ merge:
       goto done;
     }
 
+check_tail:
   /* The new segment's tail may cover multiple smaller ones */
   if (position_gt (f, normalized_end_position, s_end_pos))
     {
@@ -296,7 +307,8 @@ merge:
       /* If partial overlap with last, merge */
       if (it && position_leq (f, it->start, normalized_end_position))
 	{
-	  s->length = ooo_segment_end_pos (f, it) - s->start;
+	  s->length =
+	    position_diff (f, ooo_segment_end_pos (f, it), s->start);
 	  ooo_segment_del (f, it - f->ooo_segments);
 	}
     }
@@ -319,9 +331,9 @@ ooo_segment_try_collect (svm_fifo_t * f, u32 n_bytes_enqueued)
   i32 diff;
 
   s = pool_elt_at_index (f->ooo_segments, f->ooos_list_head);
+  diff = ooo_segment_distance_to_tail (f, s->start);
 
-  diff = (f->tail >= s->start) ?
-    f->tail - s->start : f->nitems + f->tail - s->start;
+  ASSERT (diff != n_bytes_enqueued);
 
   if (diff > n_bytes_enqueued)
     return 0;
@@ -345,8 +357,7 @@ ooo_segment_try_collect (svm_fifo_t * f, u32 n_bytes_enqueued)
       if (s->next != OOO_SEGMENT_INVALID_INDEX)
 	{
 	  s = pool_elt_at_index (f->ooo_segments, s->next);
-	  diff = (f->tail >= s->start) ?
-	    f->tail - s->start : f->nitems + f->tail - s->start;
+	  diff = ooo_segment_distance_to_tail (f, s->start);
 	  ooo_segment_del (f, index);
 	}
       /* End of search */
@@ -357,6 +368,7 @@ ooo_segment_try_collect (svm_fifo_t * f, u32 n_bytes_enqueued)
 	}
     }
 
+  ASSERT (bytes >= 0 && bytes <= f->nitems);
   return bytes;
 }
 
@@ -401,6 +413,8 @@ svm_fifo_enqueue_internal (svm_fifo_t * f, u32 max_bytes, u8 * copy_from_here)
     }
   else
     {
+      ASSERT (0);
+
       /* Account for a zero-copy enqueue done elsewhere */
       ASSERT (max_bytes <= (nitems - cursize));
       f->tail += max_bytes;
@@ -413,6 +427,7 @@ svm_fifo_enqueue_internal (svm_fifo_t * f, u32 max_bytes, u8 * copy_from_here)
     total_copy_bytes += ooo_segment_try_collect (f, total_copy_bytes);
 
   /* Atomically increase the queue length */
+  ASSERT (cursize + total_copy_bytes <= nitems);
   __sync_fetch_and_add (&f->cursize, total_copy_bytes);
 
   return (total_copy_bytes);
@@ -475,6 +490,8 @@ svm_fifo_enqueue_with_offset_internal (svm_fifo_t * f,
   cursize = svm_fifo_max_dequeue (f);
   nitems = f->nitems;
 
+  ASSERT (required_bytes < nitems);
+
   normalized_offset = (f->tail + offset) % nitems;
 
   /* Will this request fit? */
@@ -557,6 +574,7 @@ svm_fifo_dequeue_internal (svm_fifo_t * f, u32 max_bytes, u8 * copy_here)
     }
   else
     {
+      ASSERT (0);
       /* Account for a zero-copy dequeue done elsewhere */
       ASSERT (max_bytes <= cursize);
       f->head += max_bytes;
@@ -565,6 +583,8 @@ svm_fifo_dequeue_internal (svm_fifo_t * f, u32 max_bytes, u8 * copy_here)
       total_copy_bytes = max_bytes;
     }
 
+  ASSERT (f->head <= nitems);
+  ASSERT (cursize >= total_copy_bytes);
   __sync_fetch_and_sub (&f->cursize, total_copy_bytes);
 
   return (total_copy_bytes);
@@ -702,6 +722,8 @@ svm_fifo_dequeue_drop (svm_fifo_t * f, u32 max_bytes)
       f->head = (f->head == nitems) ? 0 : f->head;
     }
 
+  ASSERT (f->head <= nitems);
+  ASSERT (cursize >= total_drop_bytes);
   __sync_fetch_and_sub (&f->cursize, total_drop_bytes);
 
   return total_drop_bytes;
diff --git a/src/svm/svm_fifo.h b/src/svm/svm_fifo.h
index f32ef41d..fe21de47 100644
--- a/src/svm/svm_fifo.h
+++ b/src/svm/svm_fifo.h
@@ -133,25 +133,31 @@ svm_fifo_newest_ooo_segment (svm_fifo_t * f)
 }
 
 always_inline u32
-ooo_segment_distance_to_tail (svm_fifo_t * f, u32 a)
+ooo_segment_distance_from_tail (svm_fifo_t * f, u32 pos)
 {
   /* Ambiguous. Assumption is that ooo segments don't touch tail */
-  if (a == f->tail && f->tail == f->head)
+  if (PREDICT_FALSE (pos == f->tail && f->tail == f->head))
     return f->nitems;
 
-  return ((f->nitems + a - f->tail) % f->nitems);
+  return (((f->nitems + pos) - f->tail) % f->nitems);
+}
+
+always_inline u32
+ooo_segment_distance_to_tail (svm_fifo_t * f, u32 pos)
+{
+  return (((f->nitems + f->tail) - pos) % f->nitems);
 }
 
 always_inline u32
 ooo_segment_offset (svm_fifo_t * f, ooo_segment_t * s)
 {
-  return ooo_segment_distance_to_tail (f, s->start);
+  return ooo_segment_distance_from_tail (f, s->start);
 }
 
 always_inline u32
 ooo_segment_end_offset (svm_fifo_t * f, ooo_segment_t * s)
 {
-  return ooo_segment_distance_to_tail (f, s->start) + s->length;
+  return ooo_segment_distance_from_tail (f, s->start) + s->length;
 }
 
 always_inline u32
diff --git a/src/svm/svm_fifo_segment.c b/src/svm/svm_fifo_segment.c
index c4ac2352..69d4ecb9 100644
--- a/src/svm/svm_fifo_segment.c
+++ b/src/svm/svm_fifo_segment.c
@@ -35,6 +35,11 @@ preallocate_fifo_pairs (svm_fifo_segment_header_t * fsh,
   rx_fifo_size = (sizeof (*f) + a->rx_fifo_size) * a->preallocated_fifo_pairs;
   tx_fifo_size = (sizeof (*f) + a->tx_fifo_size) * a->preallocated_fifo_pairs;
 
+  if (0)
+    clib_warning ("rx_fifo_size %u (%d mb), tx_fifo_size %u (%d mb)",
+		  rx_fifo_size, rx_fifo_size >> 20,
+		  tx_fifo_size, tx_fifo_size >> 20);
+
   /* Allocate rx fifo space. May fail. */
   rx_fifo_space = clib_mem_alloc_aligned_at_offset
     (rx_fifo_size, CLIB_CACHE_LINE_BYTES, 0 /* align_offset */ ,
@@ -129,7 +134,7 @@ svm_fifo_segment_create (svm_fifo_segment_create_args_t * a)
   ssvm_pop_heap (oldheap);
 
   sh->ready = 1;
-  a->new_segment_index = s - sm->segments;
+  vec_add1 (a->new_segment_indices, s - sm->segments);
   return (0);
 }
 
@@ -141,35 +146,81 @@ svm_fifo_segment_create_process_private (svm_fifo_segment_create_args_t * a)
   svm_fifo_segment_main_t *sm = &svm_fifo_segment_main;
   ssvm_shared_header_t *sh;
   svm_fifo_segment_header_t *fsh;
+  void *oldheap;
+  u8 **heaps = 0;
+  mheap_t *heap_header;
+  int segment_count = 1;
+  int i;
 
-  /* Allocate a fresh segment */
-  pool_get (sm->segments, s);
-  memset (s, 0, sizeof (*s));
-
-  s->ssvm.ssvm_size = ~0;
-  s->ssvm.i_am_master = 1;
-  s->ssvm.my_pid = getpid ();
-  s->ssvm.name = (u8 *) a->segment_name;
-  s->ssvm.requested_va = ~0;
-
-  /* Allocate a [sic] shared memory header, in process memory... */
-  sh = clib_mem_alloc_aligned (sizeof (*sh), CLIB_CACHE_LINE_BYTES);
-  s->ssvm.sh = sh;
+  if (a->private_segment_count && a->private_segment_size)
+    {
+      void *mem;
+      u8 *heap;
+      u32 pagesize = clib_mem_get_page_size ();
+      u32 rnd_size;
 
-  memset (sh, 0, sizeof (*sh));
-  sh->heap = clib_mem_get_heap ();
+      for (i = 0; i < a->private_segment_count; i++)
+	{
+	  rnd_size = (a->private_segment_size + (pagesize - 1)) & ~pagesize;
+
+	  mem = mmap (0, rnd_size, PROT_READ | PROT_WRITE,
+		      MAP_PRIVATE | MAP_ANONYMOUS,
+		      -1 /* fd */ , 0 /* offset */ );
+
+	  if (mem == MAP_FAILED)
+	    {
+	      clib_unix_warning ("mmap");
+	      return -1;
+	    }
+	  heap = mheap_alloc (mem, rnd_size);
+	  heap_header = mheap_header (heap);
+	  heap_header->flags |= MHEAP_FLAG_THREAD_SAFE;
+	  vec_add1 (heaps, heap);
+	}
+      segment_count = a->private_segment_count;
+    }
 
-  /* Set up svm_fifo_segment shared header */
-  fsh = clib_mem_alloc (sizeof (*fsh));
-  memset (fsh, 0, sizeof (*fsh));
-  sh->opaque[0] = fsh;
-  s->h = fsh;
-  fsh->segment_name = format (0, "%s%c", a->segment_name, 0);
+  /* Spread preallocated fifo pairs across segments */
+  a->preallocated_fifo_pairs /= segment_count;
 
-  preallocate_fifo_pairs (fsh, a);
+  /* Allocate segments */
+  for (i = 0; i < segment_count; i++)
+    {
+      pool_get (sm->segments, s);
+      memset (s, 0, sizeof (*s));
+
+      s->ssvm.ssvm_size = ~0;
+      s->ssvm.i_am_master = 1;
+      s->ssvm.my_pid = getpid ();
+      s->ssvm.name = (u8 *) a->segment_name;
+      s->ssvm.requested_va = ~0;
+
+      /* Allocate a [sic] shared memory header, in process memory... */
+      sh = clib_mem_alloc_aligned (sizeof (*sh), CLIB_CACHE_LINE_BYTES);
+      s->ssvm.sh = sh;
+
+      memset (sh, 0, sizeof (*sh));
+      sh->heap = a->private_segment_count ? heaps[i] : clib_mem_get_heap ();
+
+      /* Set up svm_fifo_segment shared header */
+      fsh = clib_mem_alloc (sizeof (*fsh));
+      memset (fsh, 0, sizeof (*fsh));
+      sh->opaque[0] = fsh;
+      s->h = fsh;
+      fsh->segment_name = format (0, "%s%c", a->segment_name, 0);
+
+      if (a->private_segment_count)
+	{
+	  oldheap = clib_mem_get_heap ();
+	  clib_mem_set_heap (sh->heap);
+	  preallocate_fifo_pairs (fsh, a);
+	  clib_mem_set_heap (oldheap);
+	}
 
-  sh->ready = 1;
-  a->new_segment_index = s - sm->segments;
+      sh->ready = 1;
+      vec_add1 (a->new_segment_indices, s - sm->segments);
+    }
+  vec_free (heaps);
   return (0);
 }
 
@@ -205,7 +256,7 @@ svm_fifo_segment_attach (svm_fifo_segment_create_args_t * a)
   fsh = (svm_fifo_segment_header_t *) sh->opaque[0];
   s->h = fsh;
 
-  a->new_segment_index = s - sm->segments;
+  vec_add1 (a->new_segment_indices, s - sm->segments);
   return (0);
 }
 
@@ -230,7 +281,7 @@ svm_fifo_segment_alloc_fifo (svm_fifo_segment_private_t * s,
   sh = s->ssvm.sh;
   fsh = (svm_fifo_segment_header_t *) sh->opaque[0];
 
-  ssvm_lock (sh, 1, 0);
+  ssvm_lock_non_recursive (sh, 1);
   oldheap = ssvm_push_heap (sh);
 
   switch (list_index)
@@ -261,7 +312,7 @@ svm_fifo_segment_alloc_fifo (svm_fifo_segment_private_t * s,
   if (PREDICT_FALSE (f == 0))
     {
       ssvm_pop_heap (oldheap);
-      ssvm_unlock (sh);
+      ssvm_unlock_non_recursive (sh);
       return (0);
     }
 
@@ -281,7 +332,7 @@ found:
     }
 
   ssvm_pop_heap (oldheap);
-  ssvm_unlock (sh);
+  ssvm_unlock_non_recursive (sh);
   return (f);
 }
 
@@ -293,10 +344,11 @@ svm_fifo_segment_free_fifo (svm_fifo_segment_private_t * s, svm_fifo_t * f,
   svm_fifo_segment_header_t *fsh;
   void *oldheap;
 
+
   sh = s->ssvm.sh;
   fsh = (svm_fifo_segment_header_t *) sh->opaque[0];
 
-  ssvm_lock (sh, 1, 0);
+  ssvm_lock_non_recursive (sh, 2);
   oldheap = ssvm_push_heap (sh);
 
   switch (list_index)
@@ -325,7 +377,7 @@ svm_fifo_segment_free_fifo (svm_fifo_segment_private_t * s, svm_fifo_t * f,
     }
 
   ssvm_pop_heap (oldheap);
-  ssvm_unlock (sh);
+  ssvm_unlock_non_recursive (sh);
 }
 
 void
diff --git a/src/svm/svm_fifo_segment.h b/src/svm/svm_fifo_segment.h
index 31e14db5..a7a3f469 100644
--- a/src/svm/svm_fifo_segment.h
+++ b/src/svm/svm_fifo_segment.h
@@ -57,10 +57,12 @@ typedef struct
 {
   char *segment_name;
   u32 segment_size;
-  u32 new_segment_index;
+  u32 *new_segment_indices;
   u32 rx_fifo_size;
   u32 tx_fifo_size;
   u32 preallocated_fifo_pairs;
+  u32 private_segment_count;
+  u32 private_segment_size;
 } svm_fifo_segment_create_args_t;
 
 static inline svm_fifo_segment_private_t *
diff --git a/src/svm/test_svm_fifo1.c b/src/svm/test_svm_fifo1.c
index 63b4a9b7..63d75845 100644
--- a/src/svm/test_svm_fifo1.c
+++ b/src/svm/test_svm_fifo1.c
@@ -39,7 +39,7 @@ hello_world (int verbose)
   if (rv)
     return clib_error_return (0, "svm_fifo_segment_create returned %d", rv);
 
-  sp = svm_fifo_get_segment (a->new_segment_index);
+  sp = svm_fifo_get_segment (a->new_segment_indices[0]);
 
   f = svm_fifo_segment_alloc_fifo (sp, 4096, FIFO_SEGMENT_RX_FREELIST);
 
@@ -92,7 +92,7 @@ master (int verbose)
   if (rv)
     return clib_error_return (0, "svm_fifo_segment_create returned %d", rv);
 
-  sp = svm_fifo_get_segment (a->new_segment_index);
+  sp = svm_fifo_get_segment (a->new_segment_indices[0]);
 
   f = svm_fifo_segment_alloc_fifo (sp, 4096, FIFO_SEGMENT_RX_FREELIST);
 
@@ -128,7 +128,7 @@ mempig (int verbose)
   if (rv)
     return clib_error_return (0, "svm_fifo_segment_create returned %d", rv);
 
-  sp = svm_fifo_get_segment (a->new_segment_index);
+  sp = svm_fifo_get_segment (a->new_segment_indices[0]);
 
   for (i = 0; i < 1000; i++)
     {
@@ -186,7 +186,7 @@ offset (int verbose)
   if (rv)
     return clib_error_return (0, "svm_fifo_segment_create returned %d", rv);
 
-  sp = svm_fifo_get_segment (a->new_segment_index);
+  sp = svm_fifo_get_segment (a->new_segment_indices[0]);
 
   f = svm_fifo_segment_alloc_fifo (sp, 200 << 10, FIFO_SEGMENT_RX_FREELIST);
 
@@ -246,7 +246,7 @@ slave (int verbose)
   if (rv)
     return clib_error_return (0, "svm_fifo_segment_attach returned %d", rv);
 
-  sp = svm_fifo_get_segment (a->new_segment_index);
+  sp = svm_fifo_get_segment (a->new_segment_indices[0]);
   sh = sp->ssvm.sh;
   fsh = (svm_fifo_segment_header_t *) sh->opaque[0];
 
diff --git a/src/uri/uri_udp_test.c b/src/uri/uri_udp_test.c
index 45ad35a4..a8e39eaa 100644
--- a/src/uri/uri_udp_test.c
+++ b/src/uri/uri_udp_test.c
@@ -707,7 +707,7 @@ vl_api_connect_uri_reply_t_handler (vl_api_connect_uri_reply_t * mp)
 	  return;
 	}
 
-      segment_index = a->new_segment_index;
+      segment_index = a->new_segment_indices[0];
       vec_add2 (utm->seg, seg, 1);
       memcpy (seg, sm->segments + segment_index, sizeof (*seg));
       sleep (1);
diff --git a/src/vnet/session/application.c b/src/vnet/session/application.c
index 3cc56f37..8a953719 100644
--- a/src/vnet/session/application.c
+++ b/src/vnet/session/application.c
@@ -174,6 +174,8 @@ application_init (application_t * app, u32 api_client_index, u64 * options,
   props->preallocated_fifo_pairs = options[APP_OPTIONS_PREALLOC_FIFO_PAIRS];
   props->use_private_segment = options[APP_OPTIONS_FLAGS]
     & APP_OPTIONS_FLAGS_BUILTIN_APP;
+  props->private_segment_count = options[APP_OPTIONS_PRIVATE_SEGMENT_COUNT];
+  props->private_segment_size = options[APP_OPTIONS_PRIVATE_SEGMENT_SIZE];
 
   first_seg_size = options[SESSION_OPTIONS_SEGMENT_SIZE];
   if ((rv = segment_manager_init (sm, props, first_seg_size)))
diff --git a/src/vnet/session/application_interface.c b/src/vnet/session/application_interface.c
index 338ae857..566a52d7 100644
--- a/src/vnet/session/application_interface.c
+++ b/src/vnet/session/application_interface.c
@@ -275,27 +275,6 @@ vnet_application_detach (vnet_app_detach_args_t * a)
   return 0;
 }
 
-session_type_t
-session_type_from_proto_and_ip (session_api_proto_t proto, u8 is_ip4)
-{
-  if (proto == SESSION_PROTO_TCP)
-    {
-      if (is_ip4)
-	return SESSION_TYPE_IP4_TCP;
-      else
-	return SESSION_TYPE_IP6_TCP;
-    }
-  else
-    {
-      if (is_ip4)
-	return SESSION_TYPE_IP4_UDP;
-      else
-	return SESSION_TYPE_IP6_UDP;
-    }
-
-  return SESSION_N_TYPES;
-}
-
 int
 vnet_bind_uri (vnet_bind_args_t * a)
 {
diff --git a/src/vnet/session/application_interface.h b/src/vnet/session/application_interface.h
index 4d6f9def..ed9f89b3 100644
--- a/src/vnet/session/application_interface.h
+++ b/src/vnet/session/application_interface.h
@@ -22,12 +22,6 @@
 #include <vnet/session/application.h>
 #include <vnet/session/transport.h>
 
-typedef enum _session_api_proto
-{
-  SESSION_PROTO_TCP,
-  SESSION_PROTO_UDP
-} session_api_proto_t;
-
 typedef struct _vnet_app_attach_args_t
 {
   /** Binary API client index */
@@ -65,7 +59,7 @@ typedef struct _vnet_bind_args_t
     struct
     {
       transport_endpoint_t tep;
-      session_api_proto_t proto;
+      transport_proto_t proto;
     };
   };
 
@@ -98,7 +92,7 @@ typedef struct _vnet_connect_args
     struct
     {
       transport_endpoint_t tep;
-      session_api_proto_t proto;
+      transport_proto_t proto;
     };
   };
   u32 app_index;
@@ -120,6 +114,8 @@ typedef enum
   APP_EVT_QUEUE_SIZE,
   APP_OPTIONS_FLAGS,
   APP_OPTIONS_PREALLOC_FIFO_PAIRS,
+  APP_OPTIONS_PRIVATE_SEGMENT_COUNT,
+  APP_OPTIONS_PRIVATE_SEGMENT_SIZE,
   SESSION_OPTIONS_SEGMENT_SIZE,
   SESSION_OPTIONS_ADD_SEGMENT_SIZE,
   SESSION_OPTIONS_RX_FIFO_SIZE,
diff --git a/src/vnet/session/node.c b/src/vnet/session/node.c
index b24f5fd9..56e62637 100644
--- a/src/vnet/session/node.c
+++ b/src/vnet/session/node.c
@@ -378,24 +378,12 @@ session_tx_fifo_dequeue_and_snd (vlib_main_t * vm, vlib_node_runtime_t * node,
 					 n_tx_pkts, 0);
 }
 
-stream_session_t *
-session_event_get_session (session_fifo_event_t * e0, u8 thread_index)
+always_inline stream_session_t *
+session_event_get_session (session_fifo_event_t * e, u8 thread_index)
 {
-  svm_fifo_t *f0;
-  stream_session_t *s0;
-  u32 session_index0;
-
-  f0 = e0->fifo;
-  session_index0 = f0->master_session_index;
-
-  /* $$$ add multiple event queues, per vpp worker thread */
-  ASSERT (f0->master_thread_index == thread_index);
-
-  s0 = stream_session_get_if_valid (session_index0, thread_index);
-
-  ASSERT (s0 == 0 || s0->thread_index == thread_index);
-
-  return s0;
+  ASSERT (e->fifo->master_thread_index == thread_index);
+  return stream_session_get_if_valid (e->fifo->master_session_index,
+				      thread_index);
 }
 
 void
@@ -569,7 +557,6 @@ skip_dequeue:
 	case FIFO_EVENT_BUILTIN_RX:
 	  s0 = session_event_get_session (e0, my_thread_index);
 	  svm_fifo_unset_event (s0->server_rx_fifo);
-	  /* Get session's server */
 	  app = application_get (s0->app_index);
 	  app->cb_fns.builtin_server_rx_callback (s0);
 	  break;
diff --git a/src/vnet/session/segment_manager.c b/src/vnet/session/segment_manager.c
index dcef6261..262b7faa 100644
--- a/src/vnet/session/segment_manager.c
+++ b/src/vnet/session/segment_manager.c
@@ -30,7 +30,7 @@ segment_manager_t *segment_managers = 0;
 /**
  * Process private segment index
  */
-u32 private_segment_index = ~0;
+u32 *private_segment_indices;
 
 /**
  * Default fifo and segment size. TODO config.
@@ -70,7 +70,8 @@ session_manager_add_segment_i (segment_manager_t * sm, u32 segment_size,
       return VNET_API_ERROR_SVM_SEGMENT_CREATE_FAIL;
     }
 
-  vec_add1 (sm->segment_indices, ca->new_segment_index);
+  vec_append (sm->segment_indices, ca->new_segment_indices);
+  vec_free (ca->new_segment_indices);
 
   return 0;
 }
@@ -111,22 +112,23 @@ static void
 {
   svm_fifo_segment_create_args_t _a, *a = &_a;
 
-  if (private_segment_index != ~0)
+  if (private_segment_indices)
     return;
 
   memset (a, 0, sizeof (*a));
   a->segment_name = "process-private-segment";
   a->segment_size = ~0;
-  a->new_segment_index = ~0;
   a->rx_fifo_size = props->rx_fifo_size;
   a->tx_fifo_size = props->tx_fifo_size;
   a->preallocated_fifo_pairs = props->preallocated_fifo_pairs;
+  a->private_segment_count = props->private_segment_count;
+  a->private_segment_size = props->private_segment_size;
 
   if (svm_fifo_segment_create_process_private (a))
     clib_warning ("Failed to create process private segment");
 
-  private_segment_index = a->new_segment_index;
-  ASSERT (private_segment_index != ~0);
+  private_segment_indices = a->new_segment_indices;
+  ASSERT (vec_len (private_segment_indices));
 }
 
 /**
@@ -156,10 +158,10 @@ segment_manager_init (segment_manager_t * sm,
     }
   else
     {
-      if (private_segment_index == ~0)
+      if (vec_len (private_segment_indices) == 0)
 	segment_manager_alloc_process_private_segment (properties);
-      ASSERT (private_segment_index != ~0);
-      vec_add1 (sm->segment_indices, private_segment_index);
+      ASSERT (vec_len (private_segment_indices));
+      vec_append (sm->segment_indices, private_segment_indices);
     }
 
   clib_spinlock_init (&sm->lockp);
@@ -320,7 +322,7 @@ again:
   /* See if we're supposed to create another segment */
   if (*server_rx_fifo == 0)
     {
-      if (sm->properties->add_segment)
+      if (sm->properties->add_segment && !sm->properties->use_private_segment)
 	{
 	  if (added_a_segment)
 	    {
@@ -379,6 +381,10 @@ segment_manager_dealloc_fifos (u32 svm_segment_index, svm_fifo_t * rx_fifo,
   svm_fifo_segment_free_fifo (fifo_segment, tx_fifo,
 			      FIFO_SEGMENT_TX_FREELIST);
 
+  /* Don't try to delete process-private segments */
+  if (sm->properties->private_segment_count > 0)
+    return;
+
   /* Remove segment only if it holds no fifos and not the first */
   if (sm->segment_indices[0] != svm_segment_index
       && !svm_fifo_segment_has_fifos (fifo_segment))
diff --git a/src/vnet/session/segment_manager.h b/src/vnet/session/segment_manager.h
index df38d2b3..41abeb22 100644
--- a/src/vnet/session/segment_manager.h
+++ b/src/vnet/session/segment_manager.h
@@ -39,6 +39,10 @@ typedef struct _segment_manager_properties
 
   /** Use private memory segment instead of shared memory */
   u8 use_private_segment;
+
+  /** Use one or more private mheaps, instead of the global heap */
+  u32 private_segment_count;
+  u32 private_segment_size;
 } segment_manager_properties_t;
 
 typedef struct _segment_manager
diff --git a/src/vnet/session/session.c b/src/vnet/session/session.c
index fe198044..0a86d563 100644
--- a/src/vnet/session/session.c
+++ b/src/vnet/session/session.c
@@ -198,21 +198,28 @@ stream_session_lookup_listener4 (ip4_address_t * lcl, u16 lcl_port, u8 proto)
  */
 stream_session_t *
 stream_session_lookup4 (ip4_address_t * lcl, ip4_address_t * rmt,
-			u16 lcl_port, u16 rmt_port, u8 proto,
-			u32 my_thread_index)
+			u16 lcl_port, u16 rmt_port, u8 proto)
 {
   session_manager_main_t *smm = &session_manager_main;
   session_kv4_t kv4;
+  stream_session_t *s;
   int rv;
 
   /* Lookup session amongst established ones */
   make_v4_ss_kv (&kv4, lcl, rmt, lcl_port, rmt_port, proto);
   rv = clib_bihash_search_inline_16_8 (&smm->v4_session_hash, &kv4);
   if (rv == 0)
-    return stream_session_get_tsi (kv4.value, my_thread_index);
+    return stream_session_get_from_handle (kv4.value);
 
   /* If nothing is found, check if any listener is available */
-  return stream_session_lookup_listener4 (lcl, lcl_port, proto);
+  if ((s = stream_session_lookup_listener4 (lcl, lcl_port, proto)))
+    return s;
+
+  /* Finally, try half-open connections */
+  rv = clib_bihash_search_inline_16_8 (&smm->v4_half_open_hash, &kv4);
+  if (rv == 0)
+    return stream_session_get_from_handle (kv4.value);
+  return 0;
 }
 
 stream_session_t *
@@ -242,20 +249,27 @@ stream_session_lookup_listener6 (ip6_address_t * lcl, u16 lcl_port, u8 proto)
  * wildcarded local source (listener bound to all interfaces) */
 stream_session_t *
 stream_session_lookup6 (ip6_address_t * lcl, ip6_address_t * rmt,
-			u16 lcl_port, u16 rmt_port, u8 proto,
-			u32 my_thread_index)
+			u16 lcl_port, u16 rmt_port, u8 proto)
 {
   session_manager_main_t *smm = vnet_get_session_manager_main ();
   session_kv6_t kv6;
+  stream_session_t *s;
   int rv;
 
   make_v6_ss_kv (&kv6, lcl, rmt, lcl_port, rmt_port, proto);
   rv = clib_bihash_search_inline_48_8 (&smm->v6_session_hash, &kv6);
   if (rv == 0)
-    return stream_session_get_tsi (kv6.value, my_thread_index);
+    return stream_session_get_from_handle (kv6.value);
 
   /* If nothing is found, check if any listener is available */
-  return stream_session_lookup_listener6 (lcl, lcl_port, proto);
+  if ((s = stream_session_lookup_listener6 (lcl, lcl_port, proto)))
+    return s;
+
+  /* Finally, try half-open connections */
+  rv = clib_bihash_search_inline_48_8 (&smm->v6_half_open_hash, &kv6);
+  if (rv == 0)
+    return stream_session_get_from_handle (kv6.value);
+  return 0;
 }
 
 stream_session_t *
@@ -340,7 +354,6 @@ stream_session_lookup_transport4 (ip4_address_t * lcl, ip4_address_t * rmt,
   rv = clib_bihash_search_inline_16_8 (&smm->v4_half_open_hash, &kv4);
   if (rv == 0)
     return tp_vfts[proto].get_half_open (kv4.value & 0xFFFFFFFF);
-
   return 0;
 }
 
@@ -390,6 +403,8 @@ stream_session_create_i (segment_manager_t * sm, transport_connection_t * tc,
   u32 thread_index = tc->thread_index;
   int rv;
 
+  ASSERT (thread_index == vlib_get_thread_index ());
+
   if ((rv = segment_manager_alloc_session_fifos (sm, &server_rx_fifo,
 						 &server_tx_fifo,
 						 &fifo_segment_index)))
@@ -854,6 +869,7 @@ stream_session_accept (transport_connection_t * tc, u32 listener_index,
 
   s->app_index = server->index;
   s->listener_index = listener_index;
+  s->session_state = SESSION_STATE_ACCEPTING;
 
   /* Shoulder-tap the server */
   if (notify)
@@ -1088,6 +1104,27 @@ session_vpp_event_queue_allocate (session_manager_main_t * smm,
     }
 }
 
+session_type_t
+session_type_from_proto_and_ip (transport_proto_t proto, u8 is_ip4)
+{
+  if (proto == TRANSPORT_PROTO_TCP)
+    {
+      if (is_ip4)
+	return SESSION_TYPE_IP4_TCP;
+      else
+	return SESSION_TYPE_IP6_TCP;
+    }
+  else
+    {
+      if (is_ip4)
+	return SESSION_TYPE_IP4_UDP;
+      else
+	return SESSION_TYPE_IP6_UDP;
+    }
+
+  return SESSION_N_TYPES;
+}
+
 static clib_error_t *
 session_manager_main_enable (vlib_main_t * vm)
 {
@@ -1131,14 +1168,13 @@ session_manager_main_enable (vlib_main_t * vm)
     session_vpp_event_queue_allocate (smm, i);
 
   /* $$$$ preallocate hack config parameter */
-  for (i = 0; i < 200000; i++)
+  for (i = 0; i < smm->preallocated_sessions; i++)
     {
-      stream_session_t *ss;
+      stream_session_t *ss __attribute__ ((unused));
       pool_get_aligned (smm->sessions[0], ss, CLIB_CACHE_LINE_BYTES);
-      memset (ss, 0, sizeof (*ss));
     }
 
-  for (i = 0; i < 200000; i++)
+  for (i = 0; i < smm->preallocated_sessions; i++)
     pool_put_index (smm->sessions[0], i);
 
   clib_bihash_init_16_8 (&smm->v4_session_hash, "v4 session table",
@@ -1208,9 +1244,10 @@ session_manager_main_init (vlib_main_t * vm)
   return 0;
 }
 
-VLIB_INIT_FUNCTION (session_manager_main_init)
-     static clib_error_t *session_config_fn (vlib_main_t * vm,
-					     unformat_input_t * input)
+VLIB_INIT_FUNCTION (session_manager_main_init);
+
+static clib_error_t *
+session_config_fn (vlib_main_t * vm, unformat_input_t * input)
 {
   session_manager_main_t *smm = &session_manager_main;
   u32 nitems;
@@ -1224,6 +1261,9 @@ VLIB_INIT_FUNCTION (session_manager_main_init)
 	  else
 	    clib_warning ("event queue length %d too small, ignored", nitems);
 	}
+      if (unformat (input, "preallocated-sessions %d",
+		    &smm->preallocated_sessions))
+	;
       else
 	return clib_error_return (0, "unknown input `%U'",
 				  format_unformat_error, input);
diff --git a/src/vnet/session/session.h b/src/vnet/session/session.h
index 5fa4225c..b4507d4e 100644
--- a/src/vnet/session/session.h
+++ b/src/vnet/session/session.h
@@ -80,6 +80,10 @@ typedef enum
     SESSION_N_TYPES,
 } session_type_t;
 
+
+session_type_t
+session_type_from_proto_and_ip (transport_proto_t proto, u8 is_ip4);
+
 /*
  * Application session state
  */
@@ -87,6 +91,7 @@ typedef enum
 {
   SESSION_STATE_LISTENING,
   SESSION_STATE_CONNECTING,
+  SESSION_STATE_ACCEPTING,
   SESSION_STATE_READY,
   SESSION_STATE_CLOSED,
   SESSION_STATE_N_STATES,
@@ -211,8 +216,12 @@ struct _session_manager_main
   /** Per transport rx function that can either dequeue or peek */
   session_fifo_rx_fn *session_tx_fns[SESSION_N_TYPES];
 
+  /** Session manager is enabled */
   u8 is_enabled;
 
+  /** Preallocate session config parameter */
+  u32 preallocated_sessions;
+
   /* Convenience */
   vlib_main_t *vlib_main;
   vnet_main_t *vnet_main;
@@ -247,13 +256,12 @@ stream_session_t *stream_session_lookup_listener4 (ip4_address_t * lcl,
 						   u16 lcl_port, u8 proto);
 stream_session_t *stream_session_lookup4 (ip4_address_t * lcl,
 					  ip4_address_t * rmt, u16 lcl_port,
-					  u16 rmt_port, u8 proto,
-					  u32 thread_index);
+					  u16 rmt_port, u8 proto);
 stream_session_t *stream_session_lookup_listener6 (ip6_address_t * lcl,
 						   u16 lcl_port, u8 proto);
 stream_session_t *stream_session_lookup6 (ip6_address_t * lcl,
 					  ip6_address_t * rmt, u16 lcl_port,
-					  u16 rmt_port, u8, u32 thread_index);
+					  u16 rmt_port, u8 proto);
 transport_connection_t
   * stream_session_lookup_transport4 (ip4_address_t * lcl,
 				      ip4_address_t * rmt, u16 lcl_port,
@@ -277,9 +285,24 @@ stream_session_get_tsi (u64 ti_and_si, u32 thread_index)
 			    ti_and_si & 0xFFFFFFFFULL);
 }
 
+always_inline u8
+stream_session_is_valid (u32 si, u8 thread_index)
+{
+  stream_session_t *s;
+  s = pool_elt_at_index (session_manager_main.sessions[thread_index], si);
+  if (s->thread_index != thread_index || s->session_index != si
+      || s->server_rx_fifo->master_session_index != si
+      || s->server_tx_fifo->master_session_index != si
+      || s->server_rx_fifo->master_thread_index != thread_index
+      || s->server_tx_fifo->master_thread_index != thread_index)
+    return 0;
+  return 1;
+}
+
 always_inline stream_session_t *
 stream_session_get (u32 si, u32 thread_index)
 {
+  ASSERT (stream_session_is_valid (si, thread_index));
   return pool_elt_at_index (session_manager_main.sessions[thread_index], si);
 }
 
@@ -292,6 +315,7 @@ stream_session_get_if_valid (u64 si, u32 thread_index)
   if (pool_is_free_index (session_manager_main.sessions[thread_index], si))
     return 0;
 
+  ASSERT (stream_session_is_valid (si, thread_index));
   return pool_elt_at_index (session_manager_main.sessions[thread_index], si);
 }
 
diff --git a/src/vnet/session/session_cli.c b/src/vnet/session/session_cli.c
index 6b8341aa..e06bc586 100755
--- a/src/vnet/session/session_cli.c
+++ b/src/vnet/session/session_cli.c
@@ -47,7 +47,8 @@ format_stream_session (u8 * s, va_list * args)
 		  svm_fifo_max_enqueue (ss->server_tx_fifo),
 		  stream_session_get_index (ss));
 
-  if (ss->session_state == SESSION_STATE_READY)
+  if (ss->session_state == SESSION_STATE_READY
+      || ss->session_state == SESSION_STATE_ACCEPTING)
     {
       s = format (s, "%U", tp_vft->format_connection, ss->connection_index,
 		  ss->thread_index, verbose);
@@ -68,8 +69,9 @@ format_stream_session (u8 * s, va_list * args)
     }
   else if (ss->session_state == SESSION_STATE_CLOSED)
     {
-      s = format (s, "[CL] %-40U", tp_vft->format_connection,
-		  ss->connection_index, ss->thread_index, verbose);
+      s =
+	format (s, "[CL] %U", tp_vft->format_connection, ss->connection_index,
+		ss->thread_index, verbose);
       if (verbose == 1)
 	s = format (s, "%v", str);
       if (verbose > 1)
@@ -93,7 +95,13 @@ show_session_command_fn (vlib_main_t * vm, unformat_input_t * input,
   int verbose = 0, i;
   stream_session_t *pool;
   stream_session_t *s;
-  u8 *str = 0;
+  u8 *str = 0, one_session = 0, proto_set = 0, proto = 0;
+  u8 is_ip4 = 0, s_type = 0;
+  ip4_address_t lcl_ip4, rmt_ip4;
+  u32 lcl_port = 0, rmt_port = 0;
+
+  memset (&lcl_ip4, 0, sizeof (lcl_ip4));
+  memset (&rmt_ip4, 0, sizeof (rmt_ip4));
 
   if (!smm->is_enabled)
     {
@@ -106,10 +114,43 @@ show_session_command_fn (vlib_main_t * vm, unformat_input_t * input,
 	;
       else if (unformat (input, "verbose"))
 	verbose = 1;
+      else if (unformat (input, "tcp"))
+	{
+	  proto_set = 1;
+	  proto = TRANSPORT_PROTO_TCP;
+	}
+      else if (unformat (input, "%U:%d->%U:%d",
+			 unformat_ip4_address, &lcl_ip4, &lcl_port,
+			 unformat_ip4_address, &rmt_ip4, &rmt_port))
+	{
+	  one_session = 1;
+	  is_ip4 = 1;
+	}
+
       else
 	break;
     }
 
+  if (one_session)
+    {
+      if (!proto_set)
+	{
+	  vlib_cli_output (vm, "proto not set");
+	  return clib_error_return (0, "proto not set");
+	}
+
+      s_type = session_type_from_proto_and_ip (proto, is_ip4);
+      s = stream_session_lookup4 (&lcl_ip4, &rmt_ip4,
+				  clib_host_to_net_u16 (lcl_port),
+				  clib_host_to_net_u16 (rmt_port), s_type);
+      if (s)
+	vlib_cli_output (vm, "%U", format_stream_session, s, 2);
+      else
+	vlib_cli_output (vm, "session does not exist");
+
+      return 0;
+    }
+
   for (i = 0; i < vec_len (smm->sessions); i++)
     {
       u32 once_per_pool;
@@ -146,6 +187,7 @@ show_session_command_fn (vlib_main_t * vm, unformat_input_t * input,
 	}
       else
 	vlib_cli_output (vm, "Thread %d: no active sessions", i);
+      vec_reset_length (str);
     }
   vec_free (str);
 
@@ -161,15 +203,22 @@ VLIB_CLI_COMMAND (show_session_command, static) =
 };
 /* *INDENT-ON* */
 
+static int
+clear_session (stream_session_t * s)
+{
+  application_t *server = application_get (s->app_index);
+  server->cb_fns.session_disconnect_callback (s);
+  return 0;
+}
+
 static clib_error_t *
 clear_session_command_fn (vlib_main_t * vm, unformat_input_t * input,
 			  vlib_cli_command_t * cmd)
 {
   session_manager_main_t *smm = &session_manager_main;
-  u32 thread_index = 0;
+  u32 thread_index = 0, clear_all = 0;
   u32 session_index = ~0;
-  stream_session_t *pool, *session;
-  application_t *server;
+  stream_session_t **pool, *session;
 
   if (!smm->is_enabled)
     {
@@ -182,28 +231,36 @@ clear_session_command_fn (vlib_main_t * vm, unformat_input_t * input,
 	;
       else if (unformat (input, "session %d", &session_index))
 	;
+      else if (unformat (input, "all"))
+	clear_all = 1;
       else
 	return clib_error_return (0, "unknown input `%U'",
 				  format_unformat_error, input);
     }
 
-  if (session_index == ~0)
+  if (!clear_all && session_index == ~0)
     return clib_error_return (0, "session <nn> required, but not set.");
 
-  if (thread_index > vec_len (smm->sessions))
-    return clib_error_return (0, "thread %d out of range [0-%d]",
-			      thread_index, vec_len (smm->sessions));
-
-  pool = smm->sessions[thread_index];
-
-  if (pool_is_free_index (pool, session_index))
-    return clib_error_return (0, "session %d not active", session_index);
-
-  session = pool_elt_at_index (pool, session_index);
-  server = application_get (session->app_index);
+  if (session_index != ~0)
+    {
+      session = stream_session_get_if_valid (session_index, thread_index);
+      if (!session)
+	return clib_error_return (0, "no session %d on thread %d",
+				  session_index, thread_index);
+      clear_session (session);
+    }
 
-  /* Disconnect both app and transport */
-  server->cb_fns.session_disconnect_callback (session);
+  if (clear_all)
+    {
+      /* *INDENT-OFF* */
+      vec_foreach (pool, smm->sessions)
+	{
+	  pool_foreach(session, *pool, ({
+	    clear_session (session);
+	  }));
+	};
+      /* *INDENT-ON* */
+    }
 
   return 0;
 }
diff --git a/src/vnet/session/transport.h b/src/vnet/session/transport.h
index 561a9257..9c38bab9 100644
--- a/src/vnet/session/transport.h
+++ b/src/vnet/session/transport.h
@@ -225,6 +225,12 @@ make_v6_ss_kv_from_tc (session_kv6_t * kv, transport_connection_t * t)
 		 t->rmt_port, t->proto);
 }
 
+typedef enum _transport_proto
+{
+  TRANSPORT_PROTO_TCP,
+  TRANSPORT_PROTO_UDP
+} transport_proto_t;
+
 typedef struct _transport_endpoint
 {
   ip46_address_t ip;	/** ip address */
diff --git a/src/vnet/tcp/builtin_client.c b/src/vnet/tcp/builtin_client.c
index 6f8be082..a6c8a235 100644
--- a/src/vnet/tcp/builtin_client.c
+++ b/src/vnet/tcp/builtin_client.c
@@ -170,62 +170,90 @@ builtin_client_node_fn (vlib_main_t * vm, vlib_node_runtime_t * node,
 {
   tclient_main_t *tm = &tclient_main;
   int my_thread_index = vlib_get_thread_index ();
-  vl_api_disconnect_session_t *dmp;
   session_t *sp;
   int i;
   int delete_session;
   u32 *connection_indices;
-  u32 tx_quota = 0;
-  u32 delta, prev_bytes_received_this_session;
+  u32 *connections_this_batch;
+  u32 nconnections_this_batch;
 
   connection_indices = tm->connection_index_by_thread[my_thread_index];
+  connections_this_batch =
+    tm->connections_this_batch_by_thread[my_thread_index];
 
-  if (tm->run_test == 0 || vec_len (connection_indices) == 0)
+  if ((tm->run_test == 0) ||
+      ((vec_len (connection_indices) == 0)
+       && vec_len (connections_this_batch) == 0))
     return 0;
 
-  for (i = 0; i < vec_len (connection_indices); i++)
+  /* Grab another pile of connections */
+  if (PREDICT_FALSE (vec_len (connections_this_batch) == 0))
+    {
+      nconnections_this_batch =
+	clib_min (tm->connections_per_batch, vec_len (connection_indices));
+
+      ASSERT (nconnections_this_batch > 0);
+      vec_validate (connections_this_batch, nconnections_this_batch - 1);
+      clib_memcpy (connections_this_batch,
+		   connection_indices + vec_len (connection_indices)
+		   - nconnections_this_batch,
+		   nconnections_this_batch * sizeof (u32));
+      _vec_len (connection_indices) -= nconnections_this_batch;
+    }
+
+  if (PREDICT_FALSE (tm->prev_conns != tm->connections_per_batch
+		     && tm->prev_conns == vec_len (connections_this_batch)))
+    {
+      tm->repeats++;
+      tm->prev_conns = vec_len (connections_this_batch);
+      if (tm->repeats == 500000)
+	{
+	  clib_warning ("stuck clients");
+	}
+    }
+  else
+    {
+      tm->prev_conns = vec_len (connections_this_batch);
+      tm->repeats = 0;
+    }
+
+  for (i = 0; i < vec_len (connections_this_batch); i++)
     {
       delete_session = 1;
 
-      sp = pool_elt_at_index (tm->sessions, connection_indices[i]);
+      sp = pool_elt_at_index (tm->sessions, connections_this_batch[i]);
 
-      if ((tm->no_return || tx_quota < 60) && sp->bytes_to_send > 0)
+      if (sp->bytes_to_send > 0)
 	{
 	  send_test_chunk (tm, sp);
 	  delete_session = 0;
-	  tx_quota++;
 	}
-      if (!tm->no_return && sp->bytes_to_receive > 0)
+      if (sp->bytes_to_receive > 0)
 	{
-	  prev_bytes_received_this_session = sp->bytes_received;
 	  receive_test_chunk (tm, sp);
-	  delta = sp->bytes_received - prev_bytes_received_this_session;
-	  if (delta > 0)
-	    tx_quota--;
 	  delete_session = 0;
 	}
       if (PREDICT_FALSE (delete_session == 1))
 	{
-	  __sync_fetch_and_add (&tm->tx_total, tm->bytes_to_send);
+	  u32 index, thread_index;
+	  stream_session_t *s;
+
+	  __sync_fetch_and_add (&tm->tx_total, sp->bytes_sent);
 	  __sync_fetch_and_add (&tm->rx_total, sp->bytes_received);
 
-	  dmp = vl_msg_api_alloc_as_if_client (sizeof (*dmp));
-	  memset (dmp, 0, sizeof (*dmp));
-	  dmp->_vl_msg_id = ntohs (VL_API_DISCONNECT_SESSION);
-	  dmp->client_index = tm->my_client_index;
-	  dmp->handle = sp->vpp_session_handle;
-	  if (!unix_shared_memory_queue_add (tm->vl_input_queue, (u8 *) & dmp,
-					     1))
+	  stream_session_parse_handle (sp->vpp_session_handle,
+				       &index, &thread_index);
+	  s = stream_session_get_if_valid (index, thread_index);
+
+	  if (s)
 	    {
-	      vec_delete (connection_indices, 1, i);
-	      tm->connection_index_by_thread[my_thread_index] =
-		connection_indices;
+	      stream_session_disconnect (s);
+	      vec_delete (connections_this_batch, 1, i);
+	      i--;
 	      __sync_fetch_and_add (&tm->ready_connections, -1);
 	    }
 	  else
-	    {
-	      vl_msg_api_free (dmp);
-	    }
+	    clib_warning ("session AWOL?");
 
 	  /* Kick the debug CLI process */
 	  if (tm->ready_connections == 0)
@@ -236,6 +264,10 @@ builtin_client_node_fn (vlib_main_t * vm, vlib_node_runtime_t * node,
 	    }
 	}
     }
+
+  tm->connection_index_by_thread[my_thread_index] = connection_indices;
+  tm->connections_this_batch_by_thread[my_thread_index] =
+    connections_this_batch;
   return 0;
 }
 
@@ -356,6 +388,8 @@ tcp_test_clients_init (vlib_main_t * vm)
   tm->vlib_main = vm;
 
   vec_validate (tm->connection_index_by_thread, thread_main->n_vlib_mains);
+  vec_validate (tm->connections_this_batch_by_thread,
+		thread_main->n_vlib_mains);
   return 0;
 }
 
@@ -388,7 +422,8 @@ builtin_session_connected_callback (u32 app_index, u32 api_context,
   pool_get (tm->sessions, session);
   memset (session, 0, sizeof (*session));
   session_index = session - tm->sessions;
-  session->bytes_to_receive = session->bytes_to_send = tm->bytes_to_send;
+  session->bytes_to_send = tm->bytes_to_send;
+  session->bytes_to_receive = tm->no_return ? 0ULL : tm->bytes_to_send;
   session->server_rx_fifo = s->server_rx_fifo;
   session->server_rx_fifo->client_session_index = session_index;
   session->server_tx_fifo = s->server_tx_fifo;
@@ -485,6 +520,8 @@ attach_builtin_test_clients_app (void)
   options[SESSION_OPTIONS_SEGMENT_SIZE] = (2ULL << 32);
   options[SESSION_OPTIONS_RX_FIFO_SIZE] = tm->fifo_size;
   options[SESSION_OPTIONS_TX_FIFO_SIZE] = tm->fifo_size / 2;
+  options[APP_OPTIONS_PRIVATE_SEGMENT_COUNT] = tm->private_segment_count;
+  options[APP_OPTIONS_PRIVATE_SEGMENT_SIZE] = tm->private_segment_size;
   options[APP_OPTIONS_PREALLOC_FIFO_PAIRS] = prealloc_fifos;
 
   options[APP_OPTIONS_FLAGS] = APP_OPTIONS_FLAGS_BUILTIN_APP;
@@ -561,6 +598,9 @@ test_tcp_clients_command_fn (vlib_main_t * vm,
   tm->bytes_to_send = 8192;
   tm->no_return = 0;
   tm->fifo_size = 64 << 10;
+  tm->connections_per_batch = 1000;
+  tm->private_segment_count = 0;
+  tm->private_segment_size = 0;
 
   vec_free (tm->connect_uri);
 
@@ -582,6 +622,20 @@ test_tcp_clients_command_fn (vlib_main_t * vm,
 	tm->no_return = 1;
       else if (unformat (input, "fifo-size %d", &tm->fifo_size))
 	tm->fifo_size <<= 10;
+      else if (unformat (input, "private-segment-count %d",
+			 &tm->private_segment_count))
+	;
+      else if (unformat (input, "private-segment-size %dm", &tmp))
+	tm->private_segment_size = tmp << 20;
+      else if (unformat (input, "private-segment-size %dg", &tmp))
+	tm->private_segment_size = tmp << 30;
+      else if (unformat (input, "private-segment-size %d", &tmp))
+	tm->private_segment_size = tmp;
+      else if (unformat (input, "preallocate-fifos"))
+	tm->prealloc_fifos = 1;
+      else
+	if (unformat (input, "client-batch %d", &tm->connections_per_batch))
+	;
       else
 	return clib_error_return (0, "unknown input `%U'",
 				  format_unformat_error, input);
@@ -688,9 +742,13 @@ test_tcp_clients_command_fn (vlib_main_t * vm,
     vlib_cli_output (vm, "zero delta-t?");
 
 cleanup:
-  pool_free (tm->sessions);
+  tm->run_test = 0;
   for (i = 0; i < vec_len (tm->connection_index_by_thread); i++)
-    vec_reset_length (tm->connection_index_by_thread[i]);
+    {
+      vec_reset_length (tm->connection_index_by_thread[i]);
+      vec_reset_length (tm->connections_this_batch_by_thread[i]);
+    }
+  pool_free (tm->sessions);
 
   return 0;
 }
diff --git a/src/vnet/tcp/builtin_client.h b/src/vnet/tcp/builtin_client.h
index 3462e0ee..38af231d 100644
--- a/src/vnet/tcp/builtin_client.h
+++ b/src/vnet/tcp/builtin_client.h
@@ -63,6 +63,9 @@ typedef struct
   u32 configured_segment_size;
   u32 fifo_size;
   u32 expected_connections;		/**< Number of clients/connections */
+  u32 connections_per_batch;		/**< Connections to rx/tx at once */
+  u32 private_segment_count;		/**< Number of private fifo segs */
+  u32 private_segment_size;		/**< size of private fifo segs */
 
   /*
    * Test state variables
@@ -72,6 +75,7 @@ typedef struct
   uword *session_index_by_vpp_handles;	/**< Hash table for disconnecting */
   u8 *connect_test_data;		/**< Pre-computed test data */
   u32 **connection_index_by_thread;
+  u32 **connections_this_batch_by_thread; /**< active connection batch */
   pthread_t client_thread_handle;
 
   volatile u32 ready_connections;
@@ -82,7 +86,8 @@ typedef struct
 
   f64 test_start_time;
   f64 test_end_time;
-
+  u32 prev_conns;
+  u32 repeats;
   /*
    * Flags
    */
diff --git a/src/vnet/tcp/builtin_server.c b/src/vnet/tcp/builtin_server.c
index 775bfc26..8e958ac0 100644
--- a/src/vnet/tcp/builtin_server.c
+++ b/src/vnet/tcp/builtin_server.c
@@ -56,12 +56,15 @@ typedef struct
   u32 fifo_size;		/**< Fifo size */
   u32 rcv_buffer_size;		/**< Rcv buffer size */
   u32 prealloc_fifos;		/**< Preallocate fifos */
+  u32 private_segment_count;	/**< Number of private segments  */
+  u32 private_segment_size;	/**< Size of private segments  */
 
   /*
    * Test state
    */
   u8 **rx_buf;			/**< Per-thread RX buffer */
   u64 byte_index;
+  u32 **rx_retries;
 
   vlib_main_t *vlib_main;
 } builtin_server_main_t;
@@ -77,6 +80,8 @@ builtin_session_accept_callback (stream_session_t * s)
     session_manager_get_vpp_event_queue (s->thread_index);
   s->session_state = SESSION_STATE_READY;
   bsm->byte_index = 0;
+  vec_validate (bsm->rx_retries[s->thread_index], s->session_index);
+  bsm->rx_retries[s->thread_index][s->session_index] = 0;
   return 0;
 }
 
@@ -173,11 +178,16 @@ builtin_server_rx_callback (stream_session_t * s)
   builtin_server_main_t *bsm = &builtin_server_main;
   session_fifo_event_t evt;
   static int serial_number = 0;
-  u32 my_thread_id = vlib_get_thread_index ();
+  u32 thread_index = vlib_get_thread_index ();
+
+  ASSERT (s->thread_index == thread_index);
 
   rx_fifo = s->server_rx_fifo;
   tx_fifo = s->server_tx_fifo;
 
+  ASSERT (rx_fifo->master_thread_index == thread_index);
+  ASSERT (tx_fifo->master_thread_index == thread_index);
+
   max_dequeue = svm_fifo_max_dequeue (s->server_rx_fifo);
   max_enqueue = svm_fifo_max_enqueue (s->server_tx_fifo);
 
@@ -201,21 +211,31 @@ builtin_server_rx_callback (stream_session_t * s)
 	  evt.event_type = FIFO_EVENT_BUILTIN_RX;
 	  evt.event_id = 0;
 
-	  q = bsm->vpp_queue[s->thread_index];
+	  q = bsm->vpp_queue[thread_index];
 	  if (PREDICT_FALSE (q->cursize == q->maxsize))
 	    clib_warning ("out of event queue space");
-	  else
-	    unix_shared_memory_queue_add (q, (u8 *) & evt,
-					  0 /* don't wait for mutex */ );
+	  else if (unix_shared_memory_queue_add (q, (u8 *) & evt, 0	/* don't wait for mutex */
+		   ))
+	    clib_warning ("failed to enqueue self-tap");
+
+	  bsm->rx_retries[thread_index][s->session_index]++;
+	  if (bsm->rx_retries[thread_index][s->session_index] == 500000)
+	    {
+	      clib_warning ("session stuck: %U", format_stream_session, s, 2);
+	    }
+	}
+      else
+	{
+	  bsm->rx_retries[thread_index][s->session_index] = 0;
 	}
 
       return 0;
     }
 
-  _vec_len (bsm->rx_buf[my_thread_id]) = max_transfer;
+  _vec_len (bsm->rx_buf[thread_index]) = max_transfer;
 
   actual_transfer = svm_fifo_dequeue_nowait (rx_fifo, max_transfer,
-					     bsm->rx_buf[my_thread_id]);
+					     bsm->rx_buf[thread_index]);
   ASSERT (actual_transfer == max_transfer);
 
 //  test_bytes (bsm, actual_transfer);
@@ -225,7 +245,7 @@ builtin_server_rx_callback (stream_session_t * s)
    */
 
   n_written = svm_fifo_enqueue_nowait (tx_fifo, actual_transfer,
-				       bsm->rx_buf[my_thread_id]);
+				       bsm->rx_buf[thread_index]);
 
   if (n_written != max_transfer)
     clib_warning ("short trout!");
@@ -237,11 +257,13 @@ builtin_server_rx_callback (stream_session_t * s)
       evt.event_type = FIFO_EVENT_APP_TX;
       evt.event_id = serial_number++;
 
-      unix_shared_memory_queue_add (bsm->vpp_queue[s->thread_index],
-				    (u8 *) & evt, 0 /* do wait for mutex */ );
+      if (unix_shared_memory_queue_add (bsm->vpp_queue[s->thread_index],
+					(u8 *) & evt,
+					0 /* do wait for mutex */ ))
+	clib_warning ("failed to enqueue tx evt");
     }
 
-  if (PREDICT_FALSE (max_enqueue < max_dequeue))
+  if (PREDICT_FALSE (n_written < max_dequeue))
     goto rx_event;
 
   return 0;
@@ -328,9 +350,13 @@ server_attach ()
   a->options[SESSION_OPTIONS_SEGMENT_SIZE] = 512 << 20;
   a->options[SESSION_OPTIONS_RX_FIFO_SIZE] = bsm->fifo_size;
   a->options[SESSION_OPTIONS_TX_FIFO_SIZE] = bsm->fifo_size;
-  a->options[APP_OPTIONS_FLAGS] = APP_OPTIONS_FLAGS_BUILTIN_APP;
+  a->options[APP_OPTIONS_PRIVATE_SEGMENT_COUNT] = bsm->private_segment_count;
+  a->options[APP_OPTIONS_PRIVATE_SEGMENT_SIZE] = bsm->private_segment_size;
   a->options[APP_OPTIONS_PREALLOC_FIFO_PAIRS] =
     bsm->prealloc_fifos ? bsm->prealloc_fifos : 1;
+
+  a->options[APP_OPTIONS_FLAGS] = APP_OPTIONS_FLAGS_BUILTIN_APP;
+
   a->segment_name = segment_name;
   a->segment_name_length = ARRAY_LEN (segment_name);
 
@@ -374,6 +400,8 @@ server_create (vlib_main_t * vm)
   num_threads = 1 /* main thread */  + vtm->n_threads;
   vec_validate (builtin_server_main.vpp_queue, num_threads - 1);
   vec_validate (bsm->rx_buf, num_threads - 1);
+  vec_validate (bsm->rx_retries, num_threads - 1);
+
   for (i = 0; i < num_threads; i++)
     vec_validate (bsm->rx_buf[i], bsm->rcv_buffer_size);
 
@@ -435,11 +463,14 @@ server_create_command_fn (vlib_main_t * vm, unformat_input_t * input,
 {
   builtin_server_main_t *bsm = &builtin_server_main;
   int rv;
+  u32 tmp;
 
   bsm->no_echo = 0;
   bsm->fifo_size = 64 << 10;
   bsm->rcv_buffer_size = 128 << 10;
   bsm->prealloc_fifos = 0;
+  bsm->private_segment_count = 0;
+  bsm->private_segment_size = 0;
 
   while (unformat_check_input (input) != UNFORMAT_END_OF_INPUT)
     {
@@ -449,8 +480,17 @@ server_create_command_fn (vlib_main_t * vm, unformat_input_t * input,
 	bsm->fifo_size <<= 10;
       else if (unformat (input, "rcv-buf-size %d", &bsm->rcv_buffer_size))
 	;
-      else if (unformat (input, "prealloc-fifos", &bsm->prealloc_fifos))
+      else if (unformat (input, "prealloc-fifos %d", &bsm->prealloc_fifos))
+	;
+      else if (unformat (input, "private-segment-count %d",
+			 &bsm->private_segment_count))
 	;
+      else if (unformat (input, "private-segment-size %dm", &tmp))
+	bsm->private_segment_size = tmp << 20;
+      else if (unformat (input, "private-segment-size %dg", &tmp))
+	bsm->private_segment_size = tmp << 30;
+      else if (unformat (input, "private-segment-size %d", &tmp))
+	bsm->private_segment_size = tmp;
       else
 	return clib_error_return (0, "unknown input `%U'",
 				  format_unformat_error, input);
diff --git a/src/vnet/tcp/tcp.c b/src/vnet/tcp/tcp.c
index 4e85eb3f..f379e699 100644
--- a/src/vnet/tcp/tcp.c
+++ b/src/vnet/tcp/tcp.c
@@ -74,8 +74,16 @@ static void
 tcp_connection_unbind (u32 listener_index)
 {
   tcp_main_t *tm = vnet_get_tcp_main ();
-  TCP_EVT_DBG (TCP_EVT_UNBIND,
-	       pool_elt_at_index (tm->listener_pool, listener_index));
+  tcp_connection_t *tc;
+
+  tc = pool_elt_at_index (tm->listener_pool, listener_index);
+
+  TCP_EVT_DBG (TCP_EVT_UNBIND, tc);
+
+  /* Poison the entry */
+  if (CLIB_DEBUG > 0)
+    memset (tc, 0xFA, sizeof (*tc));
+
   pool_put_index (tm->listener_pool, listener_index);
 }
 
@@ -124,9 +132,20 @@ tcp_connection_cleanup (tcp_connection_t * tc)
 
   /* Check if half-open */
   if (tc->state == TCP_STATE_SYN_SENT)
-    pool_put (tm->half_open_connections, tc);
+    {
+      /* Poison the entry */
+      if (CLIB_DEBUG > 0)
+	memset (tc, 0xFA, sizeof (*tc));
+      pool_put (tm->half_open_connections, tc);
+    }
   else
-    pool_put (tm->connections[tc->c_thread_index], tc);
+    {
+      int thread_index = tc->c_thread_index;
+      /* Poison the entry */
+      if (CLIB_DEBUG > 0)
+	memset (tc, 0xFA, sizeof (*tc));
+      pool_put (tm->connections[thread_index], tc);
+    }
 }
 
 /**
@@ -168,13 +187,14 @@ tcp_connection_reset (tcp_connection_t * tc)
 
       /* Make sure all timers are cleared */
       tcp_connection_timers_reset (tc);
-
       stream_session_reset_notify (&tc->connection);
+
+      /* Wait for cleanup from session layer but not forever */
+      tcp_timer_set (tc, TCP_TIMER_WAITCLOSE, TCP_CLEANUP_TIME);
       break;
     case TCP_STATE_CLOSED:
       return;
     }
-
 }
 
 /**
@@ -278,6 +298,9 @@ tcp_allocate_local_port (tcp_main_t * tm, ip46_address_t * ip)
   tries = max - min;
   time_now = tcp_time_now ();
 
+  /* Only support active opens from thread 0 */
+  ASSERT (vlib_get_thread_index () == 0);
+
   /* Start at random point or max */
   pool_get (tm->local_endpoints, tep);
   clib_memcpy (&tep->ip, ip, sizeof (*ip));
@@ -343,6 +366,7 @@ tcp_connection_timers_reset (tcp_connection_t * tc)
     }
 }
 
+#if 0
 typedef struct ip4_tcp_hdr
 {
   ip4_header_t ip;
@@ -435,6 +459,7 @@ tcp_connection_fib_attach (tcp_connection_t * tc)
 
   tcp_connection_stack_on_fib_entry (tc);
 }
+#endif /* 0 */
 
 /** Initialize tcp connection variables
  *
@@ -447,7 +472,7 @@ tcp_connection_init_vars (tcp_connection_t * tc)
   tcp_init_mss (tc);
   scoreboard_init (&tc->sack_sb);
   tcp_cc_init (tc);
-  tcp_connection_fib_attach (tc);
+  //  tcp_connection_fib_attach (tc);
 }
 
 int
@@ -485,14 +510,38 @@ tcp_connection_open (ip46_address_t * rmt_addr, u16 rmt_port, u8 is_ip4)
   if (is_ip4)
     {
       ip4_address_t *ip4;
-      ip4 = ip_interface_get_first_ip (sw_if_index, 1);
-      lcl_addr.ip4.as_u32 = ip4->as_u32;
+      int index;
+      if (vec_len (tm->ip4_src_addresses))
+	{
+	  index = tm->last_v4_address_rotor++;
+	  if (tm->last_v4_address_rotor >= vec_len (tm->ip4_src_addresses))
+	    tm->last_v4_address_rotor = 0;
+	  lcl_addr.ip4.as_u32 = tm->ip4_src_addresses[index].as_u32;
+	}
+      else
+	{
+	  ip4 = ip_interface_get_first_ip (sw_if_index, 1);
+	  lcl_addr.ip4.as_u32 = ip4->as_u32;
+	}
     }
   else
     {
       ip6_address_t *ip6;
-      ip6 = ip_interface_get_first_ip (sw_if_index, 0);
-      clib_memcpy (&lcl_addr.ip6, ip6, sizeof (*ip6));
+      int index;
+
+      if (vec_len (tm->ip6_src_addresses))
+	{
+	  index = tm->last_v6_address_rotor++;
+	  if (tm->last_v6_address_rotor >= vec_len (tm->ip6_src_addresses))
+	    tm->last_v6_address_rotor = 0;
+	  clib_memcpy (&lcl_addr.ip6, &tm->ip6_src_addresses[index],
+		       sizeof (*ip6));
+	}
+      else
+	{
+	  ip6 = ip_interface_get_first_ip (sw_if_index, 0);
+	  clib_memcpy (&lcl_addr.ip6, ip6, sizeof (*ip6));
+	}
     }
 
   /* Allocate source port */
@@ -614,7 +663,7 @@ u8 *
 format_tcp_vars (u8 * s, va_list * args)
 {
   tcp_connection_t *tc = va_arg (*args, tcp_connection_t *);
-  s = format (s, " snd_una %u snd_nxt %u snd_una_max %u\n",
+  s = format (s, " snd_una %u snd_nxt %u snd_una_max %u",
 	      tc->snd_una - tc->iss, tc->snd_nxt - tc->iss,
 	      tc->snd_una_max - tc->iss);
   s = format (s, " rcv_nxt %u rcv_las %u\n",
@@ -628,12 +677,17 @@ format_tcp_vars (u8 * s, va_list * args)
   s = format (s, " cong %U ", format_tcp_congestion_status, tc);
   s = format (s, "cwnd %u ssthresh %u rtx_bytes %u bytes_acked %u\n",
 	      tc->cwnd, tc->ssthresh, tc->snd_rxt_bytes, tc->bytes_acked);
-  s = format (s, " prev_ssthresh %u snd_congestion %u dupack %u\n",
+  s = format (s, " prev_ssthresh %u snd_congestion %u dupack %u",
 	      tc->prev_ssthresh, tc->snd_congestion - tc->iss,
 	      tc->rcv_dupacks);
+  s = format (s, " limited_transmit %u\n", tc->limited_transmit - tc->iss);
+  s = format (s, " tsecr %u tsecr_last_ack %u\n", tc->rcv_opts.tsecr,
+	      tc->tsecr_last_ack);
   s = format (s, " rto %u rto_boff %u srtt %u rttvar %u rtt_ts %u ", tc->rto,
 	      tc->rto_boff, tc->srtt, tc->rttvar, tc->rtt_ts);
   s = format (s, "rtt_seq %u\n", tc->rtt_seq);
+  s = format (s, " tsval_recent %u tsval_recent_age %u\n", tc->tsval_recent,
+	      tcp_time_now () - tc->tsval_recent_age);
   s = format (s, " scoreboard: %U\n", format_tcp_scoreboard, &tc->sack_sb);
   if (vec_len (tc->snd_sacks))
     s = format (s, " sacks tx: %U\n", format_tcp_sacks, tc);
@@ -719,11 +773,21 @@ format_tcp_sacks (u8 * s, va_list * args)
   tcp_connection_t *tc = va_arg (*args, tcp_connection_t *);
   sack_block_t *sacks = tc->snd_sacks;
   sack_block_t *block;
-  vec_foreach (block, sacks)
-  {
-    s = format (s, " start %u end %u\n", block->start - tc->irs,
-		block->end - tc->irs);
-  }
+  int i, len = 0;
+
+  len = vec_len (sacks);
+  for (i = 0; i < len - 1; i++)
+    {
+      block = &sacks[i];
+      s = format (s, " start %u end %u\n", block->start - tc->irs,
+		  block->end - tc->irs);
+    }
+  if (len)
+    {
+      block = &sacks[len - 1];
+      s = format (s, " start %u end %u", block->start - tc->irs,
+		  block->end - tc->irs);
+    }
   return s;
 }
 
@@ -796,14 +860,18 @@ tcp_session_send_mss (transport_connection_t * trans_conn)
 always_inline u32
 tcp_round_snd_space (tcp_connection_t * tc, u32 snd_space)
 {
-  if (tc->snd_wnd < tc->snd_mss)
+  if (PREDICT_FALSE (tc->snd_wnd < tc->snd_mss))
     {
       return tc->snd_wnd <= snd_space ? tc->snd_wnd : 0;
     }
 
   /* If we can't write at least a segment, don't try at all */
-  if (snd_space < tc->snd_mss)
-    return 0;
+  if (PREDICT_FALSE (snd_space < tc->snd_mss))
+    {
+      if (snd_space > clib_min (tc->mss, tc->rcv_opts.mss) - TCP_HDR_LEN_MAX)
+	return snd_space;
+      return 0;
+    }
 
   /* round down to mss multiple */
   return snd_space - (snd_space % tc->snd_mss);
@@ -1042,6 +1110,8 @@ tcp_main_enable (vlib_main_t * vm)
   vlib_thread_main_t *vtm = vlib_get_thread_main ();
   clib_error_t *error = 0;
   u32 num_threads;
+  int thread, i;
+  tcp_connection_t *tc __attribute__ ((unused));
 
   if ((error = vlib_call_init_function (vm, ip_main_init)))
     return error;
@@ -1074,6 +1144,27 @@ tcp_main_enable (vlib_main_t * vm)
   num_threads = 1 /* main thread */  + vtm->n_threads;
   vec_validate (tm->connections, num_threads - 1);
 
+  /*
+   * Preallocate connections
+   */
+  for (thread = 0; thread < num_threads; thread++)
+    {
+      for (i = 0; i < tm->preallocated_connections; i++)
+	pool_get (tm->connections[thread], tc);
+
+      for (i = 0; i < tm->preallocated_connections; i++)
+	pool_put_index (tm->connections[thread], i);
+    }
+
+  /*
+   * Preallocate half-open connections
+   */
+  for (i = 0; i < tm->preallocated_half_open_connections; i++)
+    pool_get (tm->half_open_connections, tc);
+
+  for (i = 0; i < tm->preallocated_half_open_connections; i++)
+    pool_put_index (tm->half_open_connections, i);
+
   /* Initialize per worker thread tx buffers (used for control messages) */
   vec_validate (tm->tx_buffers, num_threads - 1);
 
@@ -1116,7 +1207,6 @@ tcp_init (vlib_main_t * vm)
 {
   tcp_main_t *tm = vnet_get_tcp_main ();
 
-  tm->vlib_main = vm;
   tm->vnet_main = vnet_get_main ();
   tm->is_enabled = 0;
 
@@ -1125,6 +1215,97 @@ tcp_init (vlib_main_t * vm)
 
 VLIB_INIT_FUNCTION (tcp_init);
 
+
+static clib_error_t *
+tcp_config_fn (vlib_main_t * vm, unformat_input_t * input)
+{
+  tcp_main_t *tm = vnet_get_tcp_main ();
+
+  while (unformat_check_input (input) != UNFORMAT_END_OF_INPUT)
+    {
+      if (unformat
+	  (input, "preallocated-connections %d",
+	   &tm->preallocated_connections))
+	;
+      else if (unformat (input, "preallocated-half-open-connections %d",
+			 &tm->preallocated_half_open_connections))
+	;
+      else
+	return clib_error_return (0, "unknown input `%U'",
+				  format_unformat_error, input);
+    }
+  return 0;
+}
+
+VLIB_CONFIG_FUNCTION (tcp_config_fn, "tcp");
+
+static clib_error_t *
+tcp_src_address (vlib_main_t * vm,
+		 unformat_input_t * input, vlib_cli_command_t * cmd_arg)
+{
+  tcp_main_t *tm = vnet_get_tcp_main ();
+  ip4_address_t v4start, v4end;
+  ip6_address_t v6start, v6end;
+  int v4set = 0;
+  int v6set = 0;
+
+  while (unformat_check_input (input) != UNFORMAT_END_OF_INPUT)
+    {
+      if (unformat (input, "%U - %U", unformat_ip4_address, &v4start,
+		    unformat_ip4_address, &v4end))
+	v4set = 1;
+      else if (unformat (input, "%U", unformat_ip4_address, &v4start))
+	{
+	  memcpy (&v4end, &v4start, sizeof (v4start));
+	  v4set = 1;
+	}
+      else if (unformat (input, "%U - %U", unformat_ip6_address, &v6start,
+			 unformat_ip4_address, &v6end))
+	v6set = 1;
+      else if (unformat (input, "%U", unformat_ip6_address, &v6start))
+	{
+	  memcpy (&v6end, &v6start, sizeof (v4start));
+	  v6set = 1;
+	}
+      else
+	break;
+    }
+
+  if (!v4set && !v6set)
+    return clib_error_return (0, "at least one v4 or v6 address required");
+
+  if (v4set)
+    {
+      u32 tmp;
+
+      do
+	{
+	  vec_add1 (tm->ip4_src_addresses, v4start);
+	  tmp = clib_net_to_host_u32 (v4start.as_u32);
+	  tmp++;
+	  v4start.as_u32 = clib_host_to_net_u32 (tmp);
+	}
+      while (clib_host_to_net_u32 (v4start.as_u32) <=
+	     clib_host_to_net_u32 (v4end.as_u32));
+    }
+  if (v6set)
+    {
+      clib_warning ("v6 src address list unimplemented...");
+    }
+  return 0;
+}
+
+/* *INDENT-OFF* */
+VLIB_CLI_COMMAND (tcp_src_address_command, static) =
+{
+  .path = "tcp src-address",
+  .short_help = "tcp src-address <ip-addr> [- <ip-addr>] add src address range",
+  .function = tcp_src_address,
+};
+/* *INDENT-ON* */
+
+
+
 /*
  * fd.io coding-style-patch-verification: ON
  *
diff --git a/src/vnet/tcp/tcp.h b/src/vnet/tcp/tcp.h
index 12d804b8..37b10fd4 100644
--- a/src/vnet/tcp/tcp.h
+++ b/src/vnet/tcp/tcp.h
@@ -348,6 +348,16 @@ typedef struct _tcp_main
   /* Flag that indicates if stack is on or off */
   u8 is_enabled;
 
+  /** Number of preallocated connections */
+  u32 preallocated_connections;
+  u32 preallocated_half_open_connections;
+
+  /** Vectors of src addresses. Optional unless one needs > 63K active-opens */
+  ip4_address_t *ip4_src_addresses;
+  u32 last_v4_address_rotor;
+  u32 last_v6_address_rotor;
+  ip6_address_t *ip6_src_addresses;
+
   /* convenience */
   vlib_main_t *vlib_main;
   vnet_main_t *vnet_main;
@@ -569,6 +579,7 @@ tcp_connection_force_ack (tcp_connection_t * tc, vlib_buffer_t * b)
 always_inline void
 tcp_timer_set (tcp_connection_t * tc, u8 timer_id, u32 interval)
 {
+  ASSERT (tc->c_thread_index == vlib_get_thread_index ());
   tc->timers[timer_id]
     = tw_timer_start_16t_2w_512sl (&tcp_main.timer_wheels[tc->c_thread_index],
 				   tc->c_c_index, timer_id, interval);
@@ -577,6 +588,7 @@ tcp_timer_set (tcp_connection_t * tc, u8 timer_id, u32 interval)
 always_inline void
 tcp_timer_reset (tcp_connection_t * tc, u8 timer_id)
 {
+  ASSERT (tc->c_thread_index == vlib_get_thread_index ());
   if (tc->timers[timer_id] == TCP_TIMER_HANDLE_INVALID)
     return;
 
@@ -588,6 +600,7 @@ tcp_timer_reset (tcp_connection_t * tc, u8 timer_id)
 always_inline void
 tcp_timer_update (tcp_connection_t * tc, u8 timer_id, u32 interval)
 {
+  ASSERT (tc->c_thread_index == vlib_get_thread_index ());
   if (tc->timers[timer_id] != TCP_TIMER_HANDLE_INVALID)
     tw_timer_stop_16t_2w_512sl (&tcp_main.timer_wheels[tc->c_thread_index],
 				tc->timers[timer_id]);
diff --git a/src/vnet/tcp/tcp_debug.h b/src/vnet/tcp/tcp_debug.h
index ae68ad1b..be51bca2 100755
--- a/src/vnet/tcp/tcp_debug.h
+++ b/src/vnet/tcp/tcp_debug.h
@@ -383,9 +383,16 @@ typedef enum _tcp_dbg_evt
       "establish",                                              	\
     },                                                          	\
   };                                                            	\
-  DECLARE_ETD(_tc, _e, 2);                                      	\
-  ed->data[0] = _timer_id;                                      	\
-  ed->data[1] = _timer_id;                                      	\
+  if (_tc)								\
+    {									\
+      DECLARE_ETD(_tc, _e, 2);                                      	\
+      ed->data[0] = _timer_id;                                      	\
+      ed->data[1] = _timer_id;                                      	\
+    }									\
+  else									\
+    {									\
+      clib_warning ("pop for unexisting connection %d", _tc_index);	\
+    }									\
 }
 
 #define TCP_EVT_SEG_INVALID_HANDLER(_tc, _seq, _end, ...)		\
diff --git a/src/vnet/tcp/tcp_input.c b/src/vnet/tcp/tcp_input.c
index a2e6dad1..45db0da6 100644
--- a/src/vnet/tcp/tcp_input.c
+++ b/src/vnet/tcp/tcp_input.c
@@ -251,6 +251,7 @@ tcp_update_timestamp (tcp_connection_t * tc, u32 seq, u32 seq_end)
   if (tcp_opts_tstamp (&tc->rcv_opts) && tc->tsval_recent
       && seq_leq (seq, tc->rcv_las) && seq_leq (tc->rcv_las, seq_end))
     {
+      ASSERT (timestamp_leq (tc->tsval_recent, tc->rcv_opts.tsval));
       tc->tsval_recent = tc->rcv_opts.tsval;
       tc->tsval_recent_age = tcp_time_now ();
     }
@@ -383,12 +384,9 @@ tcp_estimate_rtt (tcp_connection_t * tc, u32 mrtt)
   if (tc->srtt != 0)
     {
       err = mrtt - tc->srtt;
-//      tc->srtt += err >> 3;
 
       /* XXX Drop in RTT results in RTTVAR increase and bigger RTO.
        * The increase should be bound */
-//      tc->rttvar += ((int) clib_abs (err) - (int) tc->rttvar) >> 2;
-
       tc->srtt = clib_max ((int) tc->srtt + (err >> 3), 1);
       diff = (clib_abs (err) - (int) tc->rttvar) >> 2;
       tc->rttvar = clib_max ((int) tc->rttvar + diff, 1);
@@ -491,6 +489,14 @@ tcp_ack_is_dupack (tcp_connection_t * tc, vlib_buffer_t * b, u32 prev_snd_wnd,
 	  && (prev_snd_wnd == tc->snd_wnd));
 }
 
+static u8
+tcp_is_lost_fin (tcp_connection_t * tc)
+{
+  if ((tc->flags & TCP_CONN_FINSNT) && tc->snd_una_max - tc->snd_una == 1)
+    return 1;
+  return 0;
+}
+
 /**
  * Checks if ack is a congestion control event.
  */
@@ -503,7 +509,7 @@ tcp_ack_is_cc_event (tcp_connection_t * tc, vlib_buffer_t * b,
   *is_dack = tc->sack_sb.last_sacked_bytes
     || tcp_ack_is_dupack (tc, b, prev_snd_wnd, prev_snd_una);
 
-  return (*is_dack || tcp_in_cong_recovery (tc));
+  return ((*is_dack || tcp_in_cong_recovery (tc)) && !tcp_is_lost_fin (tc));
 }
 
 void
@@ -750,10 +756,20 @@ tcp_rcv_sacks (tcp_connection_t * tc, u32 ack)
        * last hole end */
       tmp = tc->rcv_opts.sacks[vec_len (tc->rcv_opts.sacks) - 1];
       last_hole = scoreboard_last_hole (sb);
-      if (seq_gt (tc->snd_una_max, sb->high_sacked)
-	  && seq_gt (tc->snd_una_max, last_hole->end))
-	last_hole->end = tc->snd_una_max;
-      /* keep track of max byte sacked for when the last hole
+      if (seq_gt (tc->snd_una_max, last_hole->end))
+	{
+	  if (seq_geq (last_hole->start, sb->high_sacked))
+	    {
+	      last_hole->end = tc->snd_una_max;
+	    }
+	  /* New hole after high sacked block */
+	  else if (seq_lt (sb->high_sacked, tc->snd_una_max))
+	    {
+	      scoreboard_insert_hole (sb, sb->tail, sb->high_sacked,
+				      tc->snd_una_max);
+	    }
+	}
+      /* Keep track of max byte sacked for when the last hole
        * is acked */
       if (seq_gt (tmp.end, sb->high_sacked))
 	sb->high_sacked = tmp.end;
@@ -764,7 +780,6 @@ tcp_rcv_sacks (tcp_connection_t * tc, u32 ack)
   while (hole && blk_index < vec_len (tc->rcv_opts.sacks))
     {
       blk = &tc->rcv_opts.sacks[blk_index];
-
       if (seq_leq (blk->start, hole->start))
 	{
 	  /* Block covers hole. Remove hole */
@@ -784,6 +799,7 @@ tcp_rcv_sacks (tcp_connection_t * tc, u32 ack)
 		    }
 		  else if (!next_hole)
 		    {
+		      ASSERT (seq_geq (sb->high_sacked, ack));
 		      sb->snd_una_adv = sb->high_sacked - ack;
 		      sb->last_bytes_delivered += sb->high_sacked - hole->end;
 		    }
@@ -819,7 +835,6 @@ tcp_rcv_sacks (tcp_connection_t * tc, u32 ack)
 	    {
 	      hole->end = blk->start;
 	    }
-
 	  hole = scoreboard_next_hole (sb, hole);
 	}
     }
@@ -827,10 +842,13 @@ tcp_rcv_sacks (tcp_connection_t * tc, u32 ack)
   scoreboard_update_bytes (tc, sb);
   sb->last_sacked_bytes = sb->sacked_bytes
     - (old_sacked_bytes - sb->last_bytes_delivered);
+  ASSERT (sb->last_sacked_bytes <= sb->sacked_bytes);
   ASSERT (sb->sacked_bytes == 0
 	  || sb->sacked_bytes < tc->snd_una_max - seq_max (tc->snd_una, ack));
   ASSERT (sb->last_sacked_bytes + sb->lost_bytes <= tc->snd_una_max
 	  - seq_max (tc->snd_una, ack));
+  ASSERT (sb->head == TCP_INVALID_SACK_HOLE_INDEX || tcp_in_recovery (tc)
+	  || sb->holes[sb->head].start == ack + sb->snd_una_adv);
 }
 
 /**
@@ -916,7 +934,8 @@ tcp_cc_congestion_undo (tcp_connection_t * tc)
 static u8
 tcp_cc_is_spurious_retransmit (tcp_connection_t * tc)
 {
-  return (tc->snd_rxt_ts
+  return (tcp_in_recovery (tc)
+	  && tc->snd_rxt_ts
 	  && tcp_opts_tstamp (&tc->rcv_opts)
 	  && timestamp_lt (tc->rcv_opts.tsecr, tc->snd_rxt_ts));
 }
@@ -994,6 +1013,7 @@ tcp_cc_handle_event (tcp_connection_t * tc, u32 is_dack)
     {
       ASSERT (tc->snd_una != tc->snd_una_max
 	      || tc->sack_sb.last_sacked_bytes);
+
       tc->rcv_dupacks++;
 
       if (tc->rcv_dupacks > TCP_DUPACK_THRESHOLD && !tc->bytes_acked)
@@ -1012,17 +1032,20 @@ tcp_cc_handle_event (tcp_connection_t * tc, u32 is_dack)
 	      goto partial_ack_test;
 	    }
 
-	  /* If of of the two conditions lower hold, reset dupacks
-	   * 1) Cumulative ack does not cover more than congestion threshold,
-	   *    and the following doesn't hold: the congestion window is
-	   *    greater than SMSS bytes and the difference between highest_ack
-	   *    and prev_highest_ack is at most 4*SMSS bytes (XXX)
-	   * 2) RFC6582 heuristic to avoid multiple fast retransmits
+	  /* If of of the two conditions lower hold, reset dupacks because
+	   * we're probably after timeout (RFC6582 heuristics).
+	   * If Cumulative ack does not cover more than congestion threshold,
+	   * and:
+	   * 1) The following doesn't hold: The congestion window is greater
+	   *    than SMSS bytes and the difference between highest_ack
+	   *    and prev_highest_ack is at most 4*SMSS bytes
+	   * 2) Echoed timestamp in the last non-dup ack does not equal the
+	   *    stored timestamp
 	   */
-	  if ((seq_gt (tc->snd_una, tc->snd_congestion)
-	       || !(tc->cwnd > tc->snd_mss
-		    && tc->bytes_acked <= 4 * tc->snd_mss))
-	      || tc->rcv_opts.tsecr != tc->tsecr_last_ack)
+	  if (seq_leq (tc->snd_una, tc->snd_congestion)
+	      && ((!(tc->cwnd > tc->snd_mss
+		     && tc->bytes_acked <= 4 * tc->snd_mss))
+		  || (tc->rcv_opts.tsecr != tc->tsecr_last_ack)))
 	    {
 	      tc->rcv_dupacks = 0;
 	      return;
@@ -1038,6 +1061,7 @@ tcp_cc_handle_event (tcp_connection_t * tc, u32 is_dack)
 	   * three segments that have left the network and should've been
 	   * buffered at the receiver XXX */
 	  tc->cwnd = tc->ssthresh + tc->rcv_dupacks * tc->snd_mss;
+	  ASSERT (tc->cwnd >= tc->snd_mss);
 
 	  /* If cwnd allows, send more data */
 	  if (tcp_opts_sack_permitted (&tc->rcv_opts)
@@ -1112,7 +1136,7 @@ partial_ack:
 	  >= tc->sack_sb.last_bytes_delivered);
   rxt_delivered = tc->bytes_acked + tc->sack_sb.snd_una_adv
     - tc->sack_sb.last_bytes_delivered;
-  if (rxt_delivered && seq_gt (tc->sack_sb.high_rxt, tc->snd_una))
+  if (0 && rxt_delivered && seq_gt (tc->sack_sb.high_rxt, tc->snd_una))
     {
       /* If we have sacks and we haven't gotten an ack beyond high_rxt,
        * remove sacked bytes delivered */
@@ -1301,6 +1325,8 @@ tcp_session_enqueue_data (tcp_connection_t * tc, vlib_buffer_t * b,
 {
   int written;
 
+  ASSERT (seq_geq (vnet_buffer (b)->tcp.seq_number, tc->rcv_nxt));
+
   /* Pure ACK. Update rcv_nxt and be done. */
   if (PREDICT_FALSE (data_len == 0))
     {
@@ -1450,6 +1476,7 @@ tcp_segment_rcv (tcp_main_t * tm, tcp_connection_t * tc, vlib_buffer_t * b,
 	  /* Chop off the bytes in the past */
 	  n_bytes_to_drop = tc->rcv_nxt - vnet_buffer (b)->tcp.seq_number;
 	  n_data_bytes -= n_bytes_to_drop;
+	  vnet_buffer (b)->tcp.seq_number = tc->rcv_nxt;
 	  vlib_buffer_advance (b, n_bytes_to_drop);
 
 	  goto in_order;
@@ -1912,11 +1939,12 @@ tcp46_syn_sent_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
 		  goto drop;
 		}
 
-	      stream_session_init_fifos_pointers (&new_tc0->connection,
-						  new_tc0->irs + 1,
-						  new_tc0->iss + 1);
 	      /* Make sure after data segment processing ACK is sent */
 	      new_tc0->flags |= TCP_CONN_SNDACK;
+
+	      /* Update rtt with the syn-ack sample */
+	      new_tc0->bytes_acked = 1;
+	      tcp_update_rtt (new_tc0, vnet_buffer (b0)->tcp.ack_number);
 	    }
 	  /* SYN: Simultaneous open. Change state to SYN-RCVD and send SYN-ACK */
 	  else
@@ -1932,9 +1960,8 @@ tcp46_syn_sent_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
 		  goto drop;
 		}
 
-	      stream_session_init_fifos_pointers (&new_tc0->connection,
-						  new_tc0->irs + 1,
-						  new_tc0->iss + 1);
+	      tc0->rtt_ts = 0;
+
 	      tcp_make_synack (new_tc0, b0);
 	      next0 = tcp_next_output (is_ip4);
 
@@ -2151,8 +2178,6 @@ tcp46_rcv_process_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
 		<< tc0->rcv_opts.wscale;
 	      tc0->snd_wl1 = vnet_buffer (b0)->tcp.seq_number;
 	      tc0->snd_wl2 = vnet_buffer (b0)->tcp.ack_number;
-
-	      /* Shoulder tap the server */
 	      stream_session_accept_notify (&tc0->connection);
 
 	      /* Reset SYN-ACK retransmit timer */
@@ -2175,6 +2200,7 @@ tcp46_rcv_process_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
 	      /* If FIN is ACKed */
 	      if (tc0->snd_una == tc0->snd_una_max)
 		{
+		  ASSERT (tcp_fin (tcp0));
 		  tc0->state = TCP_STATE_FIN_WAIT_2;
 		  /* Stop all timers, 2MSL will be set lower */
 		  tcp_connection_timers_reset (tc0);
@@ -2545,10 +2571,6 @@ tcp46_listen_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
 	  tcp_make_synack (child0, b0);
 	  next0 = tcp_next_output (is_ip4);
 
-	  /* Init fifo pointers after we have iss */
-	  stream_session_init_fifos_pointers (&child0->connection,
-					      child0->irs + 1,
-					      child0->iss + 1);
 	drop:
 	  if (PREDICT_FALSE (b0->flags & VLIB_BUFFER_IS_TRACED))
 	    {
@@ -2886,9 +2908,12 @@ do {                                                       	\
   _(LISTEN, TCP_FLAG_SYN, TCP_INPUT_NEXT_LISTEN, TCP_ERROR_NONE);
   _(LISTEN, TCP_FLAG_ACK, TCP_INPUT_NEXT_RESET, TCP_ERROR_NONE);
   _(LISTEN, TCP_FLAG_RST, TCP_INPUT_NEXT_DROP, TCP_ERROR_NONE);
+  _(LISTEN, TCP_FLAG_FIN | TCP_FLAG_ACK, TCP_INPUT_NEXT_RESET,
+    TCP_ERROR_NONE);
   /* ACK for for a SYN-ACK -> tcp-rcv-process. */
   _(SYN_RCVD, TCP_FLAG_ACK, TCP_INPUT_NEXT_RCV_PROCESS, TCP_ERROR_NONE);
   _(SYN_RCVD, TCP_FLAG_RST, TCP_INPUT_NEXT_RCV_PROCESS, TCP_ERROR_NONE);
+  _(SYN_RCVD, TCP_FLAG_SYN, TCP_INPUT_NEXT_RCV_PROCESS, TCP_ERROR_NONE);
   /* SYN-ACK for a SYN */
   _(SYN_SENT, TCP_FLAG_SYN | TCP_FLAG_ACK, TCP_INPUT_NEXT_SYN_SENT,
     TCP_ERROR_NONE);
@@ -2905,12 +2930,14 @@ do {                                                       	\
   _(ESTABLISHED, TCP_FLAG_RST, TCP_INPUT_NEXT_ESTABLISHED, TCP_ERROR_NONE);
   _(ESTABLISHED, TCP_FLAG_RST | TCP_FLAG_ACK, TCP_INPUT_NEXT_ESTABLISHED,
     TCP_ERROR_NONE);
+  _(ESTABLISHED, TCP_FLAG_SYN, TCP_INPUT_NEXT_ESTABLISHED, TCP_ERROR_NONE);
   /* ACK or FIN-ACK to our FIN */
   _(FIN_WAIT_1, TCP_FLAG_ACK, TCP_INPUT_NEXT_RCV_PROCESS, TCP_ERROR_NONE);
   _(FIN_WAIT_1, TCP_FLAG_ACK | TCP_FLAG_FIN, TCP_INPUT_NEXT_RCV_PROCESS,
     TCP_ERROR_NONE);
   /* FIN in reply to our FIN from the other side */
   _(FIN_WAIT_1, TCP_FLAG_FIN, TCP_INPUT_NEXT_RCV_PROCESS, TCP_ERROR_NONE);
+  _(FIN_WAIT_1, TCP_FLAG_RST, TCP_INPUT_NEXT_RCV_PROCESS, TCP_ERROR_NONE);
   /* FIN confirming that the peer (app) has closed */
   _(FIN_WAIT_2, TCP_FLAG_FIN, TCP_INPUT_NEXT_RCV_PROCESS, TCP_ERROR_NONE);
   _(FIN_WAIT_2, TCP_FLAG_ACK, TCP_INPUT_NEXT_RCV_PROCESS, TCP_ERROR_NONE);
@@ -2929,6 +2956,8 @@ do {                                                       	\
     TCP_ERROR_NONE);
   _(CLOSED, TCP_FLAG_ACK, TCP_INPUT_NEXT_RESET, TCP_ERROR_CONNECTION_CLOSED);
   _(CLOSED, TCP_FLAG_RST, TCP_INPUT_NEXT_DROP, TCP_ERROR_CONNECTION_CLOSED);
+  _(CLOSED, TCP_FLAG_FIN | TCP_FLAG_ACK, TCP_INPUT_NEXT_RESET,
+    TCP_ERROR_CONNECTION_CLOSED);
 #undef _
 }
 
diff --git a/src/vnet/tcp/tcp_newreno.c b/src/vnet/tcp/tcp_newreno.c
index c825e952..103fea4c 100644
--- a/src/vnet/tcp/tcp_newreno.c
+++ b/src/vnet/tcp/tcp_newreno.c
@@ -63,8 +63,8 @@ newreno_rcv_cong_ack (tcp_connection_t * tc, tcp_cc_ack_t ack_type)
 	   * window deflation" attempts to ensure that, when fast recovery
 	   * eventually ends, approximately ssthresh amount of data will be
 	   * outstanding in the network.*/
-	  tc->cwnd = (tc->cwnd > tc->bytes_acked) ?
-	    tc->cwnd - tc->bytes_acked : 0;
+	  tc->cwnd = (tc->cwnd > tc->bytes_acked + tc->snd_mss) ?
+	    tc->cwnd - tc->bytes_acked : tc->snd_mss;
 	  if (tc->bytes_acked > tc->snd_mss)
 	    tc->cwnd += tc->snd_mss;
 	}
diff --git a/src/vnet/tcp/tcp_output.c b/src/vnet/tcp/tcp_output.c
index 41bebcb3..b418e8ba 100644
--- a/src/vnet/tcp/tcp_output.c
+++ b/src/vnet/tcp/tcp_output.c
@@ -19,17 +19,20 @@
 vlib_node_registration_t tcp4_output_node;
 vlib_node_registration_t tcp6_output_node;
 
-typedef enum _tcp_output_nect
+typedef enum _tcp_output_next
 {
   TCP_OUTPUT_NEXT_DROP,
+  TCP_OUTPUT_NEXT_IP_LOOKUP,
   TCP_OUTPUT_N_NEXT
 } tcp_output_next_t;
 
 #define foreach_tcp4_output_next              	\
   _ (DROP, "error-drop")                        \
+  _ (IP_LOOKUP, "ip4-lookup")
 
 #define foreach_tcp6_output_next              	\
   _ (DROP, "error-drop")                        \
+  _ (IP_LOOKUP, "ip6-lookup")
 
 static char *tcp_error_strings[] = {
 #define tcp_error(n,s) s,
@@ -427,16 +430,16 @@ tcp_init_mss (tcp_connection_t * tc)
 #define tcp_get_free_buffer_index(tm, bidx)                             \
 do {                                                                    \
   u32 *my_tx_buffers, n_free_buffers;                                   \
-  u32 thread_index = vlib_get_thread_index();                             	\
-  my_tx_buffers = tm->tx_buffers[thread_index];                            \
+  u32 thread_index = vlib_get_thread_index();                           \
+  my_tx_buffers = tm->tx_buffers[thread_index];                         \
   if (PREDICT_FALSE(vec_len (my_tx_buffers) == 0))                      \
     {                                                                   \
       n_free_buffers = 32;      /* TODO config or macro */              \
       vec_validate (my_tx_buffers, n_free_buffers - 1);                 \
       _vec_len(my_tx_buffers) = vlib_buffer_alloc_from_free_list (      \
-          tm->vlib_main, my_tx_buffers, n_free_buffers,                 \
+       vlib_get_main(), my_tx_buffers, n_free_buffers,                  \
           VLIB_BUFFER_DEFAULT_FREE_LIST_INDEX);                         \
-      tm->tx_buffers[thread_index] = my_tx_buffers;                        \
+      tm->tx_buffers[thread_index] = my_tx_buffers;                     \
     }                                                                   \
   /* buffer shortage */                                                 \
   if (PREDICT_FALSE (vec_len (my_tx_buffers) == 0))                     \
@@ -445,12 +448,12 @@ do {                                                                    \
   _vec_len (my_tx_buffers) -= 1;                                        \
 } while (0)
 
-#define tcp_return_buffer(tm)						\
-do {									\
-  u32 *my_tx_buffers;							\
-  u32 thread_index = vlib_get_thread_index();                             	\
-  my_tx_buffers = tm->tx_buffers[thread_index];                          	\
-  _vec_len (my_tx_buffers) +=1;						\
+#define tcp_return_buffer(tm)                   \
+do {                                            \
+  u32 *my_tx_buffers;                           \
+  u32 thread_index = vlib_get_thread_index();   \
+  my_tx_buffers = tm->tx_buffers[thread_index]; \
+  _vec_len (my_tx_buffers) +=1;                 \
 } while (0)
 
 always_inline void
@@ -757,23 +760,22 @@ void
 tcp_push_ip_hdr (tcp_main_t * tm, tcp_connection_t * tc, vlib_buffer_t * b)
 {
   tcp_header_t *th = vlib_buffer_get_current (b);
-
+  vlib_main_t *vm = vlib_get_main ();
   if (tc->c_is_ip4)
     {
       ip4_header_t *ih;
-      ih = vlib_buffer_push_ip4 (tm->vlib_main, b, &tc->c_lcl_ip4,
+      ih = vlib_buffer_push_ip4 (vm, b, &tc->c_lcl_ip4,
 				 &tc->c_rmt_ip4, IP_PROTOCOL_TCP);
-      th->checksum = ip4_tcp_udp_compute_checksum (tm->vlib_main, b, ih);
+      th->checksum = ip4_tcp_udp_compute_checksum (vm, b, ih);
     }
   else
     {
       ip6_header_t *ih;
       int bogus = ~0;
 
-      ih = vlib_buffer_push_ip6 (tm->vlib_main, b, &tc->c_lcl_ip6,
+      ih = vlib_buffer_push_ip6 (vm, b, &tc->c_lcl_ip6,
 				 &tc->c_rmt_ip6, IP_PROTOCOL_TCP);
-      th->checksum = ip6_tcp_udp_icmp_compute_checksum (tm->vlib_main, b, ih,
-							&bogus);
+      th->checksum = ip6_tcp_udp_icmp_compute_checksum (vm, b, ih, &bogus);
       ASSERT (!bogus);
     }
 }
@@ -851,6 +853,13 @@ tcp_enqueue_to_output (vlib_main_t * vm, vlib_buffer_t * b, u32 bi, u8 is_ip4)
   /* Decide where to send the packet */
   next_index = is_ip4 ? tcp4_output_node.index : tcp6_output_node.index;
 
+  /* Initialize the trajectory trace, if configured */
+  if (VLIB_BUFFER_TRACE_TRAJECTORY > 0)
+    {
+      b->pre_data[0] = 1;
+      b->pre_data[1] = next_index;
+    }
+
   /* Enqueue the packet */
   f = vlib_get_frame_to_node (vm, next_index);
   to_next = vlib_frame_vector_args (f);
@@ -1144,6 +1153,7 @@ tcp_timer_retransmit_handler_i (u32 index, u8 is_syn)
 
       /* Account for the SYN */
       tc->snd_nxt += 1;
+      tc->rtt_ts = 0;
     }
   else
     {
@@ -1232,7 +1242,7 @@ tcp_timer_persist_handler (u32 index)
   /* Nothing to send */
   if (n_bytes <= 0)
     {
-      clib_warning ("persist found nothing to send");
+      // clib_warning ("persist found nothing to send");
       tcp_return_buffer (tm);
       return;
     }
@@ -1448,7 +1458,7 @@ tcp46_output_inline (vlib_main_t * vm,
 	  tcp_connection_t *tc0;
 	  tcp_tx_trace_t *t0;
 	  tcp_header_t *th0 = 0;
-	  u32 error0 = TCP_ERROR_PKTS_SENT, next0 = TCP_OUTPUT_NEXT_DROP;
+	  u32 error0 = TCP_ERROR_PKTS_SENT, next0 = TCP_OUTPUT_NEXT_IP_LOOKUP;
 
 	  bi0 = from[0];
 	  to_next[0] = bi0;
@@ -1527,6 +1537,7 @@ tcp46_output_inline (vlib_main_t * vm,
 	      tc0->rto_boff = 0;
 	    }
 
+#if 0
 	  /* Make sure we haven't lost route to our peer */
 	  if (PREDICT_FALSE (tc0->last_fib_check
 			     < tc0->snd_opts.tsval + TCP_FIB_RECHECK_PERIOD))
@@ -1547,6 +1558,10 @@ tcp46_output_inline (vlib_main_t * vm,
 	  /* Use pre-computed dpo to set next node */
 	  next0 = tc0->c_rmt_dpo.dpoi_next_node;
 	  vnet_buffer (b0)->ip.adj_index[VLIB_TX] = tc0->c_rmt_dpo.dpoi_index;
+#endif
+
+	  vnet_buffer (b0)->sw_if_index[VLIB_RX] = 0;
+	  vnet_buffer (b0)->sw_if_index[VLIB_TX] = ~0;
 
 	  b0->flags |= VNET_BUFFER_LOCALLY_ORIGINATED;
 	done:
diff --git a/src/vnet/tcp/tcp_packet.h b/src/vnet/tcp/tcp_packet.h
index a6f62ee1..9ccfe655 100644
--- a/src/vnet/tcp/tcp_packet.h
+++ b/src/vnet/tcp/tcp_packet.h
@@ -168,6 +168,7 @@ typedef struct
 #define TCP_OPTION_LEN_TIMESTAMP        10
 #define TCP_OPTION_LEN_SACK_BLOCK        8
 
+#define TCP_HDR_LEN_MAX			60
 #define TCP_WND_MAX                     65535U
 #define TCP_MAX_WND_SCALE               14	/* See RFC 1323 */
 #define TCP_OPTS_ALIGN                  4
diff --git a/src/vnet/tcp/tcp_test.c b/src/vnet/tcp/tcp_test.c
index a461e3b8..510deb4f 100644
--- a/src/vnet/tcp/tcp_test.c
+++ b/src/vnet/tcp/tcp_test.c
@@ -290,7 +290,7 @@ tcp_test_sack_tx (vlib_main_t * vm, unformat_input_t * input)
 {
   tcp_connection_t _tc, *tc = &_tc;
   sack_block_t *sacks;
-  int i, verbose = 0;
+  int i, verbose = 0, expected;
 
   while (unformat_check_input (input) != UNFORMAT_END_OF_INPUT)
     {
@@ -326,8 +326,12 @@ tcp_test_sack_tx (vlib_main_t * vm, unformat_input_t * input)
   sacks = vec_dup (tc->snd_sacks);
 
   tcp_update_sack_list (tc, 1100, 1200);
-  TCP_TEST ((vec_len (tc->snd_sacks) == 5), "sack blocks %d expected %d",
-	    vec_len (tc->snd_sacks), 5);
+  if (verbose)
+    vlib_cli_output (vm, "add new segment [1100, 1200]\n%U",
+		     format_tcp_sacks, tc);
+  expected = 5 < TCP_MAX_SACK_BLOCKS ? 6 : 5;
+  TCP_TEST ((vec_len (tc->snd_sacks) == expected),
+	    "sack blocks %d expected %d", vec_len (tc->snd_sacks), expected);
   TCP_TEST ((tc->snd_sacks[0].start == 1100),
 	    "first sack block start %u expected %u", tc->snd_sacks[0].start,
 	    1100);
diff --git a/src/vnet/udp/udp_input.c b/src/vnet/udp/udp_input.c
index e6b4f8fc..9a8ff076 100644
--- a/src/vnet/udp/udp_input.c
+++ b/src/vnet/udp/udp_input.c
@@ -123,7 +123,7 @@ udp4_uri_input_node_fn (vlib_main_t * vm,
 	  /* lookup session */
 	  s0 = stream_session_lookup4 (&ip0->dst_address, &ip0->src_address,
 				       udp0->dst_port, udp0->src_port,
-				       SESSION_TYPE_IP4_UDP, my_thread_index);
+				       SESSION_TYPE_IP4_UDP);
 
 	  /* no listener */
 	  if (PREDICT_FALSE (s0 == 0))
-- 
cgit 1.2.3-korg


From 3eb5062b40feb3002de09a3caff86232d6e1adea Mon Sep 17 00:00:00 2001
From: Florin Coras <fcoras@cisco.com>
Date: Thu, 13 Jul 2017 01:24:57 -0400
Subject: Fixes and improved tcp/session debugging

- Fix rx sack option parsing
- Add session sack scoreboard tracing and replaying
- Add svm fifo tracing and replaying
- Scoreboard/svm fifo ooo segment reception fixes
- Improved overall debugging

Change-Id: Ieae07eba355e66f5935253232bb00f2dfb7ece00
Signed-off-by: Florin Coras <fcoras@cisco.com>
---
 src/svm/svm_fifo.c             | 150 +++++++++++++++------
 src/svm/svm_fifo.h             |  41 ++++++
 src/vnet/session/session.c     |  74 ++++++++++-
 src/vnet/session/session.h     |  38 ++++--
 src/vnet/session/session_cli.c | 266 ++++++++++++++++++++++++++++++++-----
 src/vnet/tcp/builtin_client.c  |   2 +
 src/vnet/tcp/builtin_server.c  |   7 +-
 src/vnet/tcp/tcp.c             | 208 +++++++++++++++++++++++++++++
 src/vnet/tcp/tcp.h             |  53 +++++++-
 src/vnet/tcp/tcp_input.c       | 114 +++++++++++-----
 src/vnet/tcp/tcp_test.c        | 289 +++++++++++++++++++++++++++++++++++++++++
 11 files changed, 1111 insertions(+), 131 deletions(-)

(limited to 'src/svm/svm_fifo.c')

diff --git a/src/svm/svm_fifo.c b/src/svm/svm_fifo.c
index da60fee5..fc2189c5 100644
--- a/src/svm/svm_fifo.c
+++ b/src/svm/svm_fifo.c
@@ -60,6 +60,90 @@ format_ooo_segment (u8 * s, va_list * args)
   return s;
 }
 
+u8 *
+svm_fifo_dump_trace (u8 * s, svm_fifo_t * f)
+{
+#if SVM_FIFO_TRACE
+  svm_fifo_trace_elem_t *seg = 0;
+  int i = 0;
+
+  if (f->trace)
+    {
+      vec_foreach (seg, f->trace)
+      {
+	s = format (s, "{%u, %u, %u}, ", seg->offset, seg->len, seg->action);
+	i++;
+	if (i % 5 == 0)
+	  s = format (s, "\n");
+      }
+      s = format (s, "\n");
+    }
+  return s;
+#else
+  return 0;
+#endif
+}
+
+u8 *
+svm_fifo_replay (u8 * s, svm_fifo_t * f, u8 no_read, u8 verbose)
+{
+  int i, trace_len;
+  u8 *data = 0;
+  svm_fifo_trace_elem_t *trace;
+  u32 offset;
+  svm_fifo_t *dummy_fifo;
+
+  if (!f)
+    return s;
+
+#if SVM_FIFO_TRACE
+  trace = f->trace;
+  trace_len = vec_len (trace);
+#else
+  trace = 0;
+  trace_len = 0;
+#endif
+
+  dummy_fifo = svm_fifo_create (f->nitems);
+  memset (f->data, 0xFF, f->nitems);
+
+  vec_validate (data, f->nitems);
+  for (i = 0; i < vec_len (data); i++)
+    data[i] = i;
+
+  for (i = 0; i < trace_len; i++)
+    {
+      offset = trace[i].offset;
+      if (trace[i].action == 1)
+	{
+	  if (verbose)
+	    s = format (s, "adding [%u, %u]:", trace[i].offset,
+			(trace[i].offset +
+			 trace[i].len) % dummy_fifo->nitems);
+	  svm_fifo_enqueue_with_offset (dummy_fifo, trace[i].offset,
+					trace[i].len, &data[offset]);
+	}
+      else if (trace[i].action == 2)
+	{
+	  if (verbose)
+	    s = format (s, "adding [%u, %u]:", 0, trace[i].len);
+	  svm_fifo_enqueue_nowait (dummy_fifo, trace[i].len, &data[offset]);
+	}
+      else if (!no_read)
+	{
+	  if (verbose)
+	    s = format (s, "read: %u", trace[i].len);
+	  svm_fifo_dequeue_drop (dummy_fifo, trace[i].len);
+	}
+      if (verbose)
+	s = format (s, "%U", format_svm_fifo, dummy_fifo, 1);
+    }
+
+  s = format (s, "result: %U", format_svm_fifo, dummy_fifo, 1);
+
+  return s;
+}
+
 u8 *
 format_ooo_list (u8 * s, va_list * args)
 {
@@ -73,6 +157,7 @@ format_ooo_list (u8 * s, va_list * args)
       s = format (s, "  %U\n", format_ooo_segment, seg);
       ooo_segment_index = seg->next;
     }
+
   return s;
 }
 
@@ -94,10 +179,10 @@ format_svm_fifo (u8 * s, va_list * args)
 
   if (verbose)
     {
-      s = format (s, " ooo pool %d active elts\n",
-		  pool_elts (f->ooo_segments));
+      s = format (s, " ooo pool %d active elts newest %u\n",
+		  pool_elts (f->ooo_segments), f->ooos_newest);
       if (svm_fifo_has_ooo_data (f))
-	s = format (s, " %U", format_ooo_list, f);
+	s = format (s, " %U", format_ooo_list, f, verbose);
     }
   return s;
 }
@@ -116,7 +201,6 @@ svm_fifo_create (u32 data_size_in_bytes)
   memset (f, 0, sizeof (*f));
   f->nitems = data_size_in_bytes;
   f->ooos_list_head = OOO_SEGMENT_INVALID_INDEX;
-
   return (f);
 }
 
@@ -178,6 +262,7 @@ ooo_segment_add (svm_fifo_t * f, u32 offset, u32 length)
   u32 new_index, s_end_pos, s_index;
   u32 normalized_position, normalized_end_position;
 
+  ASSERT (offset + length <= ooo_segment_distance_from_tail (f, f->head));
   normalized_position = (f->tail + offset) % f->nitems;
   normalized_end_position = (f->tail + offset + length) % f->nitems;
 
@@ -205,17 +290,9 @@ ooo_segment_add (svm_fifo_t * f, u32 offset, u32 length)
       s = prev;
       s_end_pos = ooo_segment_end_pos (f, s);
 
-      /* Check head and tail now since segment may be wider at both ends so
-       * merge tests lower won't work */
-      if (position_lt (f, normalized_position, s->start))
-	{
-	  s->start = normalized_position;
-	  s->length = position_diff (f, s_end_pos, s->start);
-	}
-      if (position_gt (f, normalized_end_position, s_end_pos))
-	{
-	  s->length = position_diff (f, normalized_end_position, s->start);
-	}
+      /* Since we have previous, normalized start position cannot be smaller
+       * than prev->start. Check tail */
+      ASSERT (position_lt (f, s->start, normalized_position));
       goto check_tail;
     }
 
@@ -256,6 +333,7 @@ ooo_segment_add (svm_fifo_t * f, u32 offset, u32 length)
       /* Pool might've moved, get segment again */
       s = pool_elt_at_index (f->ooo_segments, s_index);
 
+      /* Needs to be last */
       ASSERT (s->next == OOO_SEGMENT_INVALID_INDEX);
 
       new_s->prev = s_index;
@@ -274,32 +352,22 @@ ooo_segment_add (svm_fifo_t * f, u32 offset, u32 length)
     {
       s->start = normalized_position;
       s->length = position_diff (f, s_end_pos, s->start);
-    }
-  /* Overlapping tail */
-  else if (position_gt (f, normalized_end_position, s_end_pos))
-    {
-      s->length = position_diff (f, normalized_end_position, s->start);
-    }
-  /* New segment completely covered by current one */
-  else
-    {
-      /* Do Nothing */
-      s = 0;
-      goto done;
+      f->ooos_newest = s - f->ooo_segments;
     }
 
 check_tail:
-  /* The new segment's tail may cover multiple smaller ones */
+
+  /* Overlapping tail */
   if (position_gt (f, normalized_end_position, s_end_pos))
     {
-      /* Remove the completely overlapped segments */
-      it = (s->next != OOO_SEGMENT_INVALID_INDEX) ?
-	pool_elt_at_index (f->ooo_segments, s->next) : 0;
+      s->length = position_diff (f, normalized_end_position, s->start);
+
+      /* Remove the completely overlapped segments in the tail */
+      it = ooo_segment_next (f, s);
       while (it && position_leq (f, ooo_segment_end_pos (f, it),
 				 normalized_end_position))
 	{
-	  next = (it->next != OOO_SEGMENT_INVALID_INDEX) ?
-	    pool_elt_at_index (f->ooo_segments, it->next) : 0;
+	  next = ooo_segment_next (f, it);
 	  ooo_segment_del (f, it - f->ooo_segments);
 	  it = next;
 	}
@@ -307,16 +375,12 @@ check_tail:
       /* If partial overlap with last, merge */
       if (it && position_leq (f, it->start, normalized_end_position))
 	{
-	  s->length =
-	    position_diff (f, ooo_segment_end_pos (f, it), s->start);
+	  s->length = position_diff (f, ooo_segment_end_pos (f, it),
+				     s->start);
 	  ooo_segment_del (f, it - f->ooo_segments);
 	}
+      f->ooos_newest = s - f->ooo_segments;
     }
-
-done:
-  /* Most recently updated segment */
-  if (s)
-    f->ooos_newest = s - f->ooo_segments;
 }
 
 /**
@@ -422,6 +486,8 @@ svm_fifo_enqueue_internal (svm_fifo_t * f, u32 max_bytes, u8 * copy_from_here)
       total_copy_bytes = max_bytes;
     }
 
+  svm_fifo_trace_add (f, f->head, total_copy_bytes, 2);
+
   /* Any out-of-order segments to collect? */
   if (PREDICT_FALSE (f->ooos_list_head != OOO_SEGMENT_INVALID_INDEX))
     total_copy_bytes += ooo_segment_try_collect (f, total_copy_bytes);
@@ -499,6 +565,8 @@ svm_fifo_enqueue_with_offset_internal (svm_fifo_t * f,
   if ((required_bytes + offset_from_tail) > (nitems - cursize))
     return -1;
 
+  svm_fifo_trace_add (f, offset, required_bytes, 1);
+
   ooo_segment_add (f, offset, required_bytes);
 
   /* Number of bytes we're going to copy */
@@ -707,6 +775,8 @@ svm_fifo_dequeue_drop (svm_fifo_t * f, u32 max_bytes)
   /* Number of bytes we're going to drop */
   total_drop_bytes = (cursize < max_bytes) ? cursize : max_bytes;
 
+  svm_fifo_trace_add (f, f->tail, total_drop_bytes, 3);
+
   /* Number of bytes in first copy segment */
   first_drop_bytes =
     ((nitems - f->head) < total_drop_bytes) ?
diff --git a/src/svm/svm_fifo.h b/src/svm/svm_fifo.h
index fe21de47..a83cd858 100644
--- a/src/svm/svm_fifo.h
+++ b/src/svm/svm_fifo.h
@@ -36,8 +36,16 @@ typedef struct
 format_function_t format_ooo_segment;
 format_function_t format_ooo_list;
 
+#define SVM_FIFO_TRACE (0)
 #define OOO_SEGMENT_INVALID_INDEX ((u32)~0)
 
+typedef struct
+{
+  u32 offset;
+  u32 len;
+  u32 action;
+} svm_fifo_trace_elem_t;
+
 typedef struct _svm_fifo
 {
   volatile u32 cursize;		/**< current fifo size */
@@ -64,9 +72,28 @@ typedef struct _svm_fifo
   u32 ooos_newest;		/**< Last segment to have been updated */
   struct _svm_fifo *next;	/**< next in freelist/active chain */
   struct _svm_fifo *prev;	/**< prev in active chain */
+#if SVM_FIFO_TRACE
+  svm_fifo_trace_elem_t *trace;
+#endif
     CLIB_CACHE_LINE_ALIGN_MARK (data);
 } svm_fifo_t;
 
+#if SVM_FIFO_TRACE
+#define svm_fifo_trace_add(_f, _s, _l, _t)		\
+{							\
+  svm_fifo_trace_elem_t *trace_elt;			\
+  vec_add2(_f->trace, trace_elt, 1);			\
+  trace_elt->offset = _s;				\
+  trace_elt->len = _l;					\
+  trace_elt->action = _t;				\
+}
+#else
+#define svm_fifo_trace_add(_f, _s, _l, _t)
+#endif
+
+u8 *svm_fifo_dump_trace (u8 * s, svm_fifo_t * f);
+u8 *svm_fifo_replay (u8 * s, svm_fifo_t * f, u8 no_read, u8 verbose);
+
 static inline u32
 svm_fifo_max_dequeue (svm_fifo_t * f)
 {
@@ -132,6 +159,12 @@ svm_fifo_newest_ooo_segment (svm_fifo_t * f)
   return pool_elt_at_index (f->ooo_segments, f->ooos_newest);
 }
 
+always_inline void
+svm_fifo_newest_ooo_segment_reset (svm_fifo_t * f)
+{
+  f->ooos_newest = OOO_SEGMENT_INVALID_INDEX;
+}
+
 always_inline u32
 ooo_segment_distance_from_tail (svm_fifo_t * f, u32 pos)
 {
@@ -174,6 +207,14 @@ ooo_segment_get_prev (svm_fifo_t * f, ooo_segment_t * s)
   return pool_elt_at_index (f->ooo_segments, s->prev);
 }
 
+always_inline ooo_segment_t *
+ooo_segment_next (svm_fifo_t * f, ooo_segment_t * s)
+{
+  if (s->next == OOO_SEGMENT_INVALID_INDEX)
+    return 0;
+  return pool_elt_at_index (f->ooo_segments, s->next);
+}
+
 #endif /* __included_ssvm_fifo_h__ */
 
 /*
diff --git a/src/vnet/session/session.c b/src/vnet/session/session.c
index 0a86d563..2c2a27c1 100644
--- a/src/vnet/session/session.c
+++ b/src/vnet/session/session.c
@@ -325,9 +325,9 @@ stream_session_half_open_lookup (session_manager_main_t * smm,
 }
 
 transport_connection_t *
-stream_session_lookup_transport4 (ip4_address_t * lcl, ip4_address_t * rmt,
-				  u16 lcl_port, u16 rmt_port, u8 proto,
-				  u32 my_thread_index)
+stream_session_lookup_transport_wt4 (ip4_address_t * lcl, ip4_address_t * rmt,
+				     u16 lcl_port, u16 rmt_port, u8 proto,
+				     u32 my_thread_index)
 {
   session_manager_main_t *smm = &session_manager_main;
   session_kv4_t kv4;
@@ -358,9 +358,40 @@ stream_session_lookup_transport4 (ip4_address_t * lcl, ip4_address_t * rmt,
 }
 
 transport_connection_t *
-stream_session_lookup_transport6 (ip6_address_t * lcl, ip6_address_t * rmt,
-				  u16 lcl_port, u16 rmt_port, u8 proto,
-				  u32 my_thread_index)
+stream_session_lookup_transport4 (ip4_address_t * lcl, ip4_address_t * rmt,
+				  u16 lcl_port, u16 rmt_port, u8 proto)
+{
+  session_manager_main_t *smm = &session_manager_main;
+  session_kv4_t kv4;
+  stream_session_t *s;
+  int rv;
+
+  /* Lookup session amongst established ones */
+  make_v4_ss_kv (&kv4, lcl, rmt, lcl_port, rmt_port, proto);
+  rv = clib_bihash_search_inline_16_8 (&smm->v4_session_hash, &kv4);
+  if (rv == 0)
+    {
+      s = stream_session_get_from_handle (kv4.value);
+      return tp_vfts[s->session_type].get_connection (s->connection_index,
+						      s->thread_index);
+    }
+
+  /* If nothing is found, check if any listener is available */
+  s = stream_session_lookup_listener4 (lcl, lcl_port, proto);
+  if (s)
+    return tp_vfts[s->session_type].get_listener (s->connection_index);
+
+  /* Finally, try half-open connections */
+  rv = clib_bihash_search_inline_16_8 (&smm->v4_half_open_hash, &kv4);
+  if (rv == 0)
+    return tp_vfts[proto].get_half_open (kv4.value & 0xFFFFFFFF);
+  return 0;
+}
+
+transport_connection_t *
+stream_session_lookup_transport_wt6 (ip6_address_t * lcl, ip6_address_t * rmt,
+				     u16 lcl_port, u16 rmt_port, u8 proto,
+				     u32 my_thread_index)
 {
   session_manager_main_t *smm = &session_manager_main;
   stream_session_t *s;
@@ -390,6 +421,37 @@ stream_session_lookup_transport6 (ip6_address_t * lcl, ip6_address_t * rmt,
   return 0;
 }
 
+transport_connection_t *
+stream_session_lookup_transport6 (ip6_address_t * lcl, ip6_address_t * rmt,
+				  u16 lcl_port, u16 rmt_port, u8 proto)
+{
+  session_manager_main_t *smm = &session_manager_main;
+  stream_session_t *s;
+  session_kv6_t kv6;
+  int rv;
+
+  make_v6_ss_kv (&kv6, lcl, rmt, lcl_port, rmt_port, proto);
+  rv = clib_bihash_search_inline_48_8 (&smm->v6_session_hash, &kv6);
+  if (rv == 0)
+    {
+      s = stream_session_get_from_handle (kv6.value);
+      return tp_vfts[s->session_type].get_connection (s->connection_index,
+						      s->thread_index);
+    }
+
+  /* If nothing is found, check if any listener is available */
+  s = stream_session_lookup_listener6 (lcl, lcl_port, proto);
+  if (s)
+    return tp_vfts[s->session_type].get_listener (s->connection_index);
+
+  /* Finally, try half-open connections */
+  rv = clib_bihash_search_inline_48_8 (&smm->v6_half_open_hash, &kv6);
+  if (rv == 0)
+    return tp_vfts[proto].get_half_open (kv6.value & 0xFFFFFFFF);
+
+  return 0;
+}
+
 int
 stream_session_create_i (segment_manager_t * sm, transport_connection_t * tc,
 			 stream_session_t ** ret_s)
diff --git a/src/vnet/session/session.h b/src/vnet/session/session.h
index b4507d4e..6069c574 100644
--- a/src/vnet/session/session.h
+++ b/src/vnet/session/session.h
@@ -263,15 +263,30 @@ stream_session_t *stream_session_lookup6 (ip6_address_t * lcl,
 					  ip6_address_t * rmt, u16 lcl_port,
 					  u16 rmt_port, u8 proto);
 transport_connection_t
-  * stream_session_lookup_transport4 (ip4_address_t * lcl,
-				      ip4_address_t * rmt, u16 lcl_port,
-				      u16 rmt_port, u8 proto,
-				      u32 thread_index);
-transport_connection_t
-  * stream_session_lookup_transport6 (ip6_address_t * lcl,
-				      ip6_address_t * rmt, u16 lcl_port,
-				      u16 rmt_port, u8 proto,
-				      u32 thread_index);
+  * stream_session_lookup_transport_wt4 (ip4_address_t * lcl,
+					 ip4_address_t * rmt, u16 lcl_port,
+					 u16 rmt_port, u8 proto,
+					 u32 thread_index);
+transport_connection_t *stream_session_lookup_transport4 (ip4_address_t * lcl,
+							  ip4_address_t * rmt,
+							  u16 lcl_port,
+							  u16 rmt_port,
+							  u8 proto);
+transport_connection_t *stream_session_lookup_transport_wt6 (ip6_address_t *
+							     lcl,
+							     ip6_address_t *
+							     rmt,
+							     u16 lcl_port,
+							     u16 rmt_port,
+							     u8 proto,
+							     u32
+							     thread_index);
+transport_connection_t *stream_session_lookup_transport6 (ip6_address_t * lcl,
+							  ip6_address_t * rmt,
+							  u16 lcl_port,
+							  u16 rmt_port,
+							  u8 proto);
+
 stream_session_t *stream_session_lookup_listener (ip46_address_t * lcl,
 						  u16 lcl_port, u8 proto);
 void stream_session_table_add_for_tc (transport_connection_t * tc, u64 value);
@@ -415,7 +430,12 @@ void stream_session_cleanup (stream_session_t * s);
 void session_send_session_evt_to_thread (u64 session_handle,
 					 fifo_event_type_t evt_type,
 					 u32 thread_index);
+
 u8 *format_stream_session (u8 * s, va_list * args);
+uword unformat_stream_session (unformat_input_t * input, va_list * args);
+uword unformat_transport_connection (unformat_input_t * input,
+				     va_list * args);
+
 int
 send_session_connected_callback (u32 app_index, u32 api_context,
 				 stream_session_t * s, u8 is_fail);
diff --git a/src/vnet/session/session_cli.c b/src/vnet/session/session_cli.c
index e06bc586..e8e6f99c 100755
--- a/src/vnet/session/session_cli.c
+++ b/src/vnet/session/session_cli.c
@@ -81,12 +81,141 @@ format_stream_session (u8 * s, va_list * args)
     {
       clib_warning ("Session in state: %d!", ss->session_state);
     }
-
   vec_free (str);
 
   return s;
 }
 
+uword
+unformat_stream_session_id (unformat_input_t * input, va_list * args)
+{
+  u8 *proto = va_arg (*args, u8 *);
+  ip46_address_t *lcl = va_arg (*args, ip46_address_t *);
+  ip46_address_t *rmt = va_arg (*args, ip46_address_t *);
+  u16 *lcl_port = va_arg (*args, u16 *);
+  u16 *rmt_port = va_arg (*args, u16 *);
+  u8 *is_ip4 = va_arg (*args, u8 *);
+  u8 tuple_is_set = 0;
+
+  memset (lcl, 0, sizeof (*lcl));
+  memset (rmt, 0, sizeof (*rmt));
+
+  if (unformat (input, "tcp"))
+    {
+      *proto = TRANSPORT_PROTO_TCP;
+    }
+  if (unformat (input, "udp"))
+    {
+      *proto = TRANSPORT_PROTO_UDP;
+    }
+  else if (unformat (input, "%U:%d->%U:%d", unformat_ip4_address, &lcl->ip4,
+		     lcl_port, unformat_ip4_address, &rmt->ip4, rmt_port))
+    {
+      *is_ip4 = 1;
+      tuple_is_set = 1;
+    }
+  else if (unformat (input, "%U:%d->%U:%d", unformat_ip6_address, &lcl->ip6,
+		     lcl_port, unformat_ip6_address, &rmt->ip6, rmt_port))
+    {
+      *is_ip4 = 0;
+      tuple_is_set = 1;
+    }
+  else
+    return 0;
+
+  if (tuple_is_set)
+    return 1;
+
+  return 0;
+}
+
+uword
+unformat_stream_session (unformat_input_t * input, va_list * args)
+{
+  stream_session_t **result = va_arg (*args, stream_session_t **);
+  stream_session_t *s;
+  u8 proto = ~0;
+  ip46_address_t lcl, rmt;
+  u32 lcl_port = 0, rmt_port = 0;
+  u8 is_ip4 = 0, s_type = ~0, id_is_set = 0;
+
+  if (unformat (input, "%U", unformat_stream_session_id, &proto, &lcl, &rmt,
+		&lcl_port, &rmt_port, &is_ip4))
+    {
+      id_is_set = 1;
+    }
+  else
+    return 0;
+
+  if (!id_is_set)
+    {
+      return 0;
+    }
+
+  s_type = session_type_from_proto_and_ip (proto, is_ip4);
+  if (is_ip4)
+    s = stream_session_lookup4 (&lcl.ip4, &rmt.ip4,
+				clib_host_to_net_u16 (lcl_port),
+				clib_host_to_net_u16 (rmt_port), s_type);
+  else
+    s = stream_session_lookup6 (&lcl.ip6, &rmt.ip6,
+				clib_host_to_net_u16 (lcl_port),
+				clib_host_to_net_u16 (rmt_port), s_type);
+  if (s)
+    {
+      *result = s;
+      return 1;
+    }
+  return 0;
+}
+
+uword
+unformat_transport_connection (unformat_input_t * input, va_list * args)
+{
+  transport_connection_t **result = va_arg (*args, transport_connection_t **);
+  u32 suggested_proto = va_arg (*args, u32);
+  transport_connection_t *tc;
+  u8 proto = ~0;
+  ip46_address_t lcl, rmt;
+  u32 lcl_port = 0, rmt_port = 0;
+  u8 is_ip4 = 0, s_type = ~0, id_is_set = 0;
+
+  if (unformat (input, "%U", unformat_stream_session_id, &proto, &lcl, &rmt,
+		&lcl_port, &rmt_port, &is_ip4))
+    {
+      id_is_set = 1;
+    }
+  else
+    return 0;
+
+  if (!id_is_set)
+    {
+      return 0;
+    }
+
+  proto = (proto == (u8) ~ 0) ? suggested_proto : proto;
+  if (proto == (u8) ~ 0)
+    return 0;
+  s_type = session_type_from_proto_and_ip (proto, is_ip4);
+  if (is_ip4)
+    tc = stream_session_lookup_transport4 (&lcl.ip4, &rmt.ip4,
+					   clib_host_to_net_u16 (lcl_port),
+					   clib_host_to_net_u16 (rmt_port),
+					   s_type);
+  else
+    tc = stream_session_lookup_transport6 (&lcl.ip6, &rmt.ip6,
+					   clib_host_to_net_u16 (lcl_port),
+					   clib_host_to_net_u16 (rmt_port),
+					   s_type);
+
+  if (tc)
+    {
+      *result = tc;
+      return 1;
+    }
+  return 0;
+}
+
 static clib_error_t *
 show_session_command_fn (vlib_main_t * vm, unformat_input_t * input,
 			 vlib_cli_command_t * cmd)
@@ -95,13 +224,7 @@ show_session_command_fn (vlib_main_t * vm, unformat_input_t * input,
   int verbose = 0, i;
   stream_session_t *pool;
   stream_session_t *s;
-  u8 *str = 0, one_session = 0, proto_set = 0, proto = 0;
-  u8 is_ip4 = 0, s_type = 0;
-  ip4_address_t lcl_ip4, rmt_ip4;
-  u32 lcl_port = 0, rmt_port = 0;
-
-  memset (&lcl_ip4, 0, sizeof (lcl_ip4));
-  memset (&rmt_ip4, 0, sizeof (rmt_ip4));
+  u8 *str = 0, one_session = 0;
 
   if (!smm->is_enabled)
     {
@@ -114,40 +237,18 @@ show_session_command_fn (vlib_main_t * vm, unformat_input_t * input,
 	;
       else if (unformat (input, "verbose"))
 	verbose = 1;
-      else if (unformat (input, "tcp"))
-	{
-	  proto_set = 1;
-	  proto = TRANSPORT_PROTO_TCP;
-	}
-      else if (unformat (input, "%U:%d->%U:%d",
-			 unformat_ip4_address, &lcl_ip4, &lcl_port,
-			 unformat_ip4_address, &rmt_ip4, &rmt_port))
+      else if (unformat (input, "%U", unformat_stream_session, &s))
 	{
 	  one_session = 1;
-	  is_ip4 = 1;
 	}
-
       else
-	break;
+	return clib_error_return (0, "unknown input `%U'",
+				  format_unformat_error, input);
     }
 
   if (one_session)
     {
-      if (!proto_set)
-	{
-	  vlib_cli_output (vm, "proto not set");
-	  return clib_error_return (0, "proto not set");
-	}
-
-      s_type = session_type_from_proto_and_ip (proto, is_ip4);
-      s = stream_session_lookup4 (&lcl_ip4, &rmt_ip4,
-				  clib_host_to_net_u16 (lcl_port),
-				  clib_host_to_net_u16 (rmt_port), s_type);
-      if (s)
-	vlib_cli_output (vm, "%U", format_stream_session, s, 2);
-      else
-	vlib_cli_output (vm, "session does not exist");
-
+      vlib_cli_output (vm, "%U", format_stream_session, s, 2);
       return 0;
     }
 
@@ -274,6 +375,103 @@ VLIB_CLI_COMMAND (clear_session_command, static) =
 };
 /* *INDENT-ON* */
 
+static clib_error_t *
+show_session_fifo_trace_command_fn (vlib_main_t * vm,
+				    unformat_input_t * input,
+				    vlib_cli_command_t * cmd)
+{
+  stream_session_t *s = 0;
+  u8 is_rx = 0, *str = 0;
+
+  while (unformat_check_input (input) != UNFORMAT_END_OF_INPUT)
+    {
+      if (unformat (input, "%U", unformat_stream_session, &s))
+	;
+      else if (unformat (input, "rx"))
+	is_rx = 1;
+      else if (unformat (input, "tx"))
+	is_rx = 0;
+      else
+	return clib_error_return (0, "unknown input `%U'",
+				  format_unformat_error, input);
+    }
+
+  if (!SVM_FIFO_TRACE)
+    {
+      vlib_cli_output (vm, "fifo tracing not enabled");
+      return 0;
+    }
+
+  if (!s)
+    {
+      vlib_cli_output (vm, "could not find session");
+      return 0;
+    }
+
+  str = is_rx ?
+    svm_fifo_dump_trace (str, s->server_rx_fifo) :
+    svm_fifo_dump_trace (str, s->server_tx_fifo);
+
+  vlib_cli_output (vm, "%v", str);
+  return 0;
+}
+
+/* *INDENT-OFF* */
+VLIB_CLI_COMMAND (show_session_fifo_trace_command, static) =
+{
+  .path = "show session fifo trace",
+  .short_help = "show session fifo trace <session>",
+  .function = show_session_fifo_trace_command_fn,
+};
+/* *INDENT-ON* */
+
+static clib_error_t *
+session_replay_fifo_command_fn (vlib_main_t * vm, unformat_input_t * input,
+				vlib_cli_command_t * cmd)
+{
+  stream_session_t *s = 0;
+  u8 is_rx = 0, *str = 0;
+
+  while (unformat_check_input (input) != UNFORMAT_END_OF_INPUT)
+    {
+      if (unformat (input, "%U", unformat_stream_session, &s))
+	;
+      else if (unformat (input, "rx"))
+	is_rx = 1;
+      else
+	return clib_error_return (0, "unknown input `%U'",
+				  format_unformat_error, input);
+    }
+
+  if (!SVM_FIFO_TRACE)
+    {
+      vlib_cli_output (vm, "fifo tracing not enabled");
+      return 0;
+    }
+
+  if (!s)
+    {
+      vlib_cli_output (vm, "could not find session");
+      return 0;
+    }
+
+  str = is_rx ?
+    svm_fifo_replay (str, s->server_rx_fifo, 0, 1) :
+    svm_fifo_replay (str, s->server_tx_fifo, 0, 1);
+
+  vlib_cli_output (vm, "%v", str);
+  return 0;
+}
+
+/* *INDENT-OFF* */
+VLIB_CLI_COMMAND (session_replay_fifo_trace_command, static) =
+{
+  .path = "session replay fifo",
+  .short_help = "session replay fifo <session>",
+  .function = session_replay_fifo_command_fn,
+};
+/* *INDENT-ON* */
+
 static clib_error_t *
 session_enable_disable_fn (vlib_main_t * vm, unformat_input_t * input,
 			   vlib_cli_command_t * cmd)
diff --git a/src/vnet/tcp/builtin_client.c b/src/vnet/tcp/builtin_client.c
index a6c8a235..a92bacaa 100644
--- a/src/vnet/tcp/builtin_client.c
+++ b/src/vnet/tcp/builtin_client.c
@@ -464,6 +464,8 @@ builtin_session_connected_callback (u32 app_index, u32 api_context,
 static void
 builtin_session_reset_callback (stream_session_t * s)
 {
+  if (s->session_state == SESSION_STATE_READY)
+    clib_warning ("Reset active connection %U", format_stream_session, s, 2);
   return;
 }
 
diff --git a/src/vnet/tcp/builtin_server.c b/src/vnet/tcp/builtin_server.c
index 8e958ac0..4ecaf56a 100644
--- a/src/vnet/tcp/builtin_server.c
+++ b/src/vnet/tcp/builtin_server.c
@@ -99,8 +99,7 @@ builtin_session_disconnect_callback (stream_session_t * s)
 void
 builtin_session_reset_callback (stream_session_t * s)
 {
-  clib_warning ("called.. ");
-
+  clib_warning ("Reset session %U", format_stream_session, s, 2);
   stream_session_cleanup (s);
 }
 
@@ -224,10 +223,6 @@ builtin_server_rx_callback (stream_session_t * s)
 	      clib_warning ("session stuck: %U", format_stream_session, s, 2);
 	    }
 	}
-      else
-	{
-	  bsm->rx_retries[thread_index][s->session_index] = 0;
-	}
 
       return 0;
     }
diff --git a/src/vnet/tcp/tcp.c b/src/vnet/tcp/tcp.c
index f379e699..8ed325d2 100644
--- a/src/vnet/tcp/tcp.c
+++ b/src/vnet/tcp/tcp.c
@@ -732,6 +732,7 @@ format_tcp_connection (u8 * s, va_list * args)
       if (verbose > 1)
 	s = format (s, " %U\n%U", format_tcp_timers, tc, format_tcp_vars, tc);
     }
+
   return s;
 }
 
@@ -791,6 +792,30 @@ format_tcp_sacks (u8 * s, va_list * args)
   return s;
 }
 
+u8 *
+format_tcp_rcv_sacks (u8 * s, va_list * args)
+{
+  tcp_connection_t *tc = va_arg (*args, tcp_connection_t *);
+  sack_block_t *sacks = tc->rcv_opts.sacks;
+  sack_block_t *block;
+  int i, len = 0;
+
+  len = vec_len (sacks);
+  for (i = 0; i < len - 1; i++)
+    {
+      block = &sacks[i];
+      s = format (s, " start %u end %u\n", block->start - tc->iss,
+		  block->end - tc->iss);
+    }
+  if (len)
+    {
+      block = &sacks[len - 1];
+      s = format (s, " start %u end %u", block->start - tc->iss,
+		  block->end - tc->iss);
+    }
+  return s;
+}
+
 u8 *
 format_tcp_sack_hole (u8 * s, va_list * args)
 {
@@ -820,6 +845,7 @@ format_tcp_scoreboard (u8 * s, va_list * args)
       s = format (s, "%U", format_tcp_sack_hole, hole);
       hole = scoreboard_next_hole (sb, hole);
     }
+
   return s;
 }
 
@@ -1304,7 +1330,189 @@ VLIB_CLI_COMMAND (tcp_src_address_command, static) =
 };
 /* *INDENT-ON* */
 
+static u8 *
+tcp_scoreboard_dump_trace (u8 * s, sack_scoreboard_t * sb)
+{
+#if TCP_SCOREBOARD_TRACE
+
+  scoreboard_trace_elt_t *block;
+  int i = 0;
+
+  if (!sb->trace)
+    return s;
+
+  s = format (s, "scoreboard trace:");
+  vec_foreach (block, sb->trace)
+  {
+    s = format (s, "{%u, %u, %u, %u, %u}, ", block->start, block->end,
+		block->ack, block->snd_una_max, block->group);
+    if ((++i % 3) == 0)
+      s = format (s, "\n");
+  }
+  return s;
+#else
+  return 0;
+#endif
+}
+
+static clib_error_t *
+tcp_show_scoreboard_trace_fn (vlib_main_t * vm, unformat_input_t * input,
+			      vlib_cli_command_t * cmd_arg)
+{
+  transport_connection_t *tconn = 0;
+  tcp_connection_t *tc;
+  u8 *s = 0;
+  while (unformat_check_input (input) != UNFORMAT_END_OF_INPUT)
+    {
+      if (unformat (input, "%U", unformat_transport_connection, &tconn,
+		    TRANSPORT_PROTO_TCP))
+	;
+      else
+	return clib_error_return (0, "unknown input `%U'",
+				  format_unformat_error, input);
+    }
+
+  if (!TCP_SCOREBOARD_TRACE)
+    {
+      vlib_cli_output (vm, "scoreboard tracing not enabled");
+      return 0;
+    }
+
+  tc = tcp_get_connection_from_transport (tconn);
+  s = tcp_scoreboard_dump_trace (s, &tc->sack_sb);
+  vlib_cli_output (vm, "%v", s);
+  return 0;
+}
+
+/* *INDENT-OFF* */
+VLIB_CLI_COMMAND (tcp_show_scoreboard_trace_command, static) =
+{
+  .path = "show tcp scoreboard trace",
+  .short_help = "show tcp scoreboard trace <connection>",
+  .function = tcp_show_scoreboard_trace_fn,
+};
+/* *INDENT-ON* */
 
+u8 *
+tcp_scoreboard_replay (u8 * s, tcp_connection_t * tc, u8 verbose)
+{
+  int i, trace_len;
+  scoreboard_trace_elt_t *trace;
+  u32 next_ack, left, group, has_new_ack = 0;
+  tcp_connection_t _dummy_tc, *dummy_tc = &_dummy_tc;
+  sack_block_t *block;
+
+  if (!tc)
+    return s;
+
+  memset (dummy_tc, 0, sizeof (*dummy_tc));
+  tcp_connection_timers_init (dummy_tc);
+  scoreboard_init (&dummy_tc->sack_sb);
+  dummy_tc->rcv_opts.flags |= TCP_OPTS_FLAG_SACK;
+
+#if TCP_SCOREBOARD_TRACE
+  trace = tc->sack_sb.trace;
+  trace_len = vec_len (tc->sack_sb.trace);
+#else
+  trace = 0;
+  trace_len = 0;
+#endif
+
+  for (i = 0; i < trace_len; i++)
+    {
+      if (trace[i].ack != 0)
+	{
+	  dummy_tc->snd_una = trace[i].ack - 1448;
+	  dummy_tc->snd_una_max = trace[i].ack;
+	}
+    }
+
+  left = 0;
+  while (left < trace_len)
+    {
+      group = trace[left].group;
+      vec_reset_length (dummy_tc->rcv_opts.sacks);
+      has_new_ack = 0;
+      while (trace[left].group == group)
+	{
+	  if (trace[left].ack != 0)
+	    {
+	      if (verbose)
+		s = format (s, "Adding ack %u, snd_una_max %u, segs: ",
+			    trace[left].ack, trace[left].snd_una_max);
+	      dummy_tc->snd_una_max = trace[left].snd_una_max;
+	      next_ack = trace[left].ack;
+	      has_new_ack = 1;
+	    }
+	  else
+	    {
+	      if (verbose)
+		s = format (s, "[%u, %u], ", trace[left].start,
+			    trace[left].end);
+	      vec_add2 (dummy_tc->rcv_opts.sacks, block, 1);
+	      block->start = trace[left].start;
+	      block->end = trace[left].end;
+	    }
+	  left++;
+	}
+
+      /* Push segments */
+      tcp_rcv_sacks (dummy_tc, next_ack);
+      if (has_new_ack)
+	dummy_tc->snd_una = next_ack + dummy_tc->sack_sb.snd_una_adv;
+
+      if (verbose)
+	s = format (s, "result: %U", format_tcp_scoreboard,
+		    &dummy_tc->sack_sb);
+
+    }
+  s = format (s, "result: %U", format_tcp_scoreboard, &dummy_tc->sack_sb);
+
+  return s;
+}
+
+static clib_error_t *
+tcp_scoreboard_trace_fn (vlib_main_t * vm, unformat_input_t * input,
+			 vlib_cli_command_t * cmd_arg)
+{
+  transport_connection_t *tconn = 0;
+  tcp_connection_t *tc = 0;
+  u8 *str = 0;
+  while (unformat_check_input (input) != UNFORMAT_END_OF_INPUT)
+    {
+      if (unformat (input, "%U", unformat_transport_connection, &tconn,
+		    TRANSPORT_PROTO_TCP))
+	;
+      else
+	return clib_error_return (0, "unknown input `%U'",
+				  format_unformat_error, input);
+    }
+
+  if (!TCP_SCOREBOARD_TRACE)
+    {
+      vlib_cli_output (vm, "scoreboard tracing not enabled");
+      return 0;
+    }
+
+  tc = tcp_get_connection_from_transport (tconn);
+  if (!tc)
+    {
+      vlib_cli_output (vm, "connection not found");
+      return 0;
+    }
+  str = tcp_scoreboard_replay (str, tc, 1);
+  vlib_cli_output (vm, "%v", str);
+  return 0;
+}
+
+/* *INDENT-OFF* */
+VLIB_CLI_COMMAND (tcp_replay_scoreboard_command, static) =
+{
+  .path = "tcp replay scoreboard",
+  .short_help = "tcp replay scoreboard <connection>",
+  .function = tcp_scoreboard_trace_fn,
+};
+/* *INDENT-ON* */
 
 /*
  * fd.io coding-style-patch-verification: ON
diff --git a/src/vnet/tcp/tcp.h b/src/vnet/tcp/tcp.h
index 37b10fd4..fd0d02b9 100644
--- a/src/vnet/tcp/tcp.h
+++ b/src/vnet/tcp/tcp.h
@@ -62,6 +62,7 @@ typedef enum _tcp_state
 format_function_t format_tcp_state;
 format_function_t format_tcp_flags;
 format_function_t format_tcp_sacks;
+format_function_t format_tcp_rcv_sacks;
 
 /** TCP timers */
 #define foreach_tcp_timer               \
@@ -151,9 +152,19 @@ enum
 #undef _
 };
 
+#define TCP_SCOREBOARD_TRACE (0)
 #define TCP_MAX_SACK_BLOCKS 15	/**< Max number of SACK blocks stored */
 #define TCP_INVALID_SACK_HOLE_INDEX ((u32)~0)
 
+typedef struct _scoreboard_trace_elt
+{
+  u32 start;
+  u32 end;
+  u32 ack;
+  u32 snd_una_max;
+  u32 group;
+} scoreboard_trace_elt_t;
+
 typedef struct _sack_scoreboard_hole
 {
   u32 next;		/**< Index for next entry in linked list */
@@ -177,8 +188,38 @@ typedef struct _sack_scoreboard
   u32 rescue_rxt;			/**< Rescue sequence number */
   u32 lost_bytes;			/**< Bytes lost as per RFC6675 */
   u32 cur_rxt_hole;			/**< Retransmitting from this hole */
+
+#if TCP_SCOREBOARD_TRACE
+  scoreboard_trace_elt_t *trace;
+#endif
+
 } sack_scoreboard_t;
 
+#if TCP_SCOREBOARD_TRACE
+#define tcp_scoreboard_trace_add(_tc, _ack) 				\
+{									\
+    static u64 _group = 0;						\
+    sack_scoreboard_t *_sb = &_tc->sack_sb;				\
+    sack_block_t *_sack, *_sacks;					\
+    scoreboard_trace_elt_t *_elt;					\
+    int i;								\
+    _group++;								\
+    _sacks = _tc->rcv_opts.sacks;					\
+    for (i = 0; i < vec_len (_sacks); i++) 				\
+      {									\
+	_sack = &_sacks[i];						\
+	vec_add2 (_sb->trace, _elt, 1);					\
+	_elt->start = _sack->start;					\
+	_elt->end = _sack->end;						\
+	_elt->ack = _elt->end == _ack ? _ack : 0;			\
+	_elt->snd_una_max = _elt->end == _ack ? _tc->snd_una_max : 0;	\
+	_elt->group = _group;						\
+      }									\
+}
+#else
+#define tcp_scoreboard_trace_add(_tc, _ack)
+#endif
+
 typedef enum _tcp_cc_algorithm_type
 {
   TCP_CC_NEWRENO,
@@ -405,6 +446,12 @@ tcp_connection_get_if_valid (u32 conn_index, u32 thread_index)
   return pool_elt_at_index (tcp_main.connections[thread_index], conn_index);
 }
 
+always_inline tcp_connection_t *
+tcp_get_connection_from_transport (transport_connection_t * tconn)
+{
+  return (tcp_connection_t *) tconn;
+}
+
 void tcp_connection_close (tcp_connection_t * tc);
 void tcp_connection_cleanup (tcp_connection_t * tc);
 void tcp_connection_del (tcp_connection_t * tc);
@@ -414,6 +461,8 @@ u8 *format_tcp_connection_id (u8 * s, va_list * args);
 u8 *format_tcp_connection (u8 * s, va_list * args);
 u8 *format_tcp_scoreboard (u8 * s, va_list * args);
 
+u8 *tcp_scoreboard_replay (u8 * s, tcp_connection_t * tc, u8 verbose);
+
 always_inline tcp_connection_t *
 tcp_listener_get (u32 tli)
 {
@@ -689,7 +738,7 @@ sack_scoreboard_hole_t *scoreboard_next_rxt_hole (sack_scoreboard_t * sb,
 						  start, u8 have_sent_1_smss,
 						  u8 * can_rescue,
 						  u8 * snd_limited);
-void scoreboard_init_high_rxt (sack_scoreboard_t * sb);
+void scoreboard_init_high_rxt (sack_scoreboard_t * sb, u32 seq);
 
 always_inline sack_scoreboard_hole_t *
 scoreboard_get_hole (sack_scoreboard_t * sb, u32 index)
@@ -740,6 +789,7 @@ scoreboard_clear (sack_scoreboard_t * sb)
       scoreboard_remove_hole (sb, hole);
     }
   ASSERT (sb->head == sb->tail && sb->head == TCP_INVALID_SACK_HOLE_INDEX);
+  ASSERT (pool_elts (sb->holes) == 0);
   sb->sacked_bytes = 0;
   sb->last_sacked_bytes = 0;
   sb->last_bytes_delivered = 0;
@@ -759,6 +809,7 @@ scoreboard_hole_bytes (sack_scoreboard_hole_t * hole)
 always_inline u32
 scoreboard_hole_index (sack_scoreboard_t * sb, sack_scoreboard_hole_t * hole)
 {
+  ASSERT (!pool_is_free_index (sb->holes, hole - sb->holes));
   return hole - sb->holes;
 }
 
diff --git a/src/vnet/tcp/tcp_input.c b/src/vnet/tcp/tcp_input.c
index 45db0da6..bc7d9015 100644
--- a/src/vnet/tcp/tcp_input.c
+++ b/src/vnet/tcp/tcp_input.c
@@ -206,8 +206,8 @@ tcp_options_parse (tcp_header_t * th, tcp_options_t * to)
 	  vec_reset_length (to->sacks);
 	  for (j = 0; j < to->n_sack_blocks; j++)
 	    {
-	      b.start = clib_net_to_host_u32 (*(u32 *) (data + 2 + 4 * j));
-	      b.end = clib_net_to_host_u32 (*(u32 *) (data + 6 + 4 * j));
+	      b.start = clib_net_to_host_u32 (*(u32 *) (data + 2 + 8 * j));
+	      b.end = clib_net_to_host_u32 (*(u32 *) (data + 6 + 8 * j));
 	      vec_add1 (to->sacks, b);
 	    }
 	  break;
@@ -540,6 +540,10 @@ scoreboard_remove_hole (sack_scoreboard_t * sb, sack_scoreboard_hole_t * hole)
   if (scoreboard_hole_index (sb, hole) == sb->cur_rxt_hole)
     sb->cur_rxt_hole = TCP_INVALID_SACK_HOLE_INDEX;
 
+  /* Poison the entry */
+  if (CLIB_DEBUG > 0)
+    memset (hole, 0xfe, sizeof (*hole));
+
   pool_put (sb->holes, hole);
 }
 
@@ -555,7 +559,7 @@ scoreboard_insert_hole (sack_scoreboard_t * sb, u32 prev_index,
 
   hole->start = start;
   hole->end = end;
-  hole_index = hole - sb->holes;
+  hole_index = scoreboard_hole_index (sb, hole);
 
   prev = scoreboard_get_hole (sb, prev_index);
   if (prev)
@@ -680,12 +684,30 @@ scoreboard_next_rxt_hole (sack_scoreboard_t * sb,
 }
 
 void
-scoreboard_init_high_rxt (sack_scoreboard_t * sb)
+scoreboard_init_high_rxt (sack_scoreboard_t * sb, u32 seq)
 {
   sack_scoreboard_hole_t *hole;
   hole = scoreboard_first_hole (sb);
-  sb->high_rxt = hole->start;
-  sb->cur_rxt_hole = sb->head;
+  if (hole)
+    {
+      seq = seq_gt (seq, hole->start) ? seq : hole->start;
+      sb->cur_rxt_hole = sb->head;
+    }
+  sb->high_rxt = seq;
+}
+
+/**
+ * Test that scoreboard is sane after recovery
+ *
+ * Returns 1 if scoreboard is empty or if first hole beyond
+ * snd_una.
+ */
+u8
+tcp_scoreboard_is_sane_post_recovery (tcp_connection_t * tc)
+{
+  sack_scoreboard_hole_t *hole;
+  hole = scoreboard_first_hole (&tc->sack_sb);
+  return (!hole || seq_geq (hole->start, tc->snd_una));
 }
 
 void
@@ -712,7 +734,7 @@ tcp_rcv_sacks (tcp_connection_t * tc, u32 ack)
     {
       if (seq_lt (blk->start, blk->end)
 	  && seq_gt (blk->start, tc->snd_una)
-	  && seq_gt (blk->start, ack) && seq_leq (blk->end, tc->snd_nxt))
+	  && seq_gt (blk->start, ack) && seq_leq (blk->end, tc->snd_una_max))
 	{
 	  blk++;
 	  continue;
@@ -731,6 +753,8 @@ tcp_rcv_sacks (tcp_connection_t * tc, u32 ack)
   if (vec_len (tc->rcv_opts.sacks) == 0)
     return;
 
+  tcp_scoreboard_trace_add (tc, ack);
+
   /* Make sure blocks are ordered */
   for (i = 0; i < vec_len (tc->rcv_opts.sacks); i++)
     for (j = i + 1; j < vec_len (tc->rcv_opts.sacks); j++)
@@ -797,7 +821,7 @@ tcp_rcv_sacks (tcp_connection_t * tc, u32 ack)
 		      sb->last_bytes_delivered +=
 			next_hole->start - hole->end;
 		    }
-		  else if (!next_hole)
+		  else
 		    {
 		      ASSERT (seq_geq (sb->high_sacked, ack));
 		      sb->snd_una_adv = sb->high_sacked - ack;
@@ -824,12 +848,14 @@ tcp_rcv_sacks (tcp_connection_t * tc, u32 ack)
 	  if (seq_lt (blk->end, hole->end))
 	    {
 	      hole_index = scoreboard_hole_index (sb, hole);
-	      scoreboard_insert_hole (sb, hole_index, blk->end, hole->end);
+	      next_hole = scoreboard_insert_hole (sb, hole_index, blk->end,
+						  hole->end);
 
 	      /* Pool might've moved */
 	      hole = scoreboard_get_hole (sb, hole_index);
 	      hole->end = blk->start;
 	      blk_index++;
+	      ASSERT (hole->next == scoreboard_hole_index (sb, next_hole));
 	    }
 	  else if (seq_lt (blk->start, hole->end))
 	    {
@@ -957,7 +983,7 @@ tcp_cc_recover (tcp_connection_t * tc)
 
   ASSERT (tc->rto_boff == 0);
   ASSERT (!tcp_in_cong_recovery (tc));
-
+  ASSERT (tcp_scoreboard_is_sane_post_recovery (tc));
   TCP_EVT_DBG (TCP_EVT_CC_EVT, tc, 3);
   return 0;
 }
@@ -965,7 +991,7 @@ tcp_cc_recover (tcp_connection_t * tc)
 static void
 tcp_cc_update (tcp_connection_t * tc, vlib_buffer_t * b)
 {
-  ASSERT (!tcp_in_cong_recovery (tc));
+  ASSERT (!tcp_in_cong_recovery (tc) || tcp_is_lost_fin (tc));
 
   /* Congestion avoidance */
   tc->cc_algo->rcv_ack (tc);
@@ -1064,10 +1090,10 @@ tcp_cc_handle_event (tcp_connection_t * tc, u32 is_dack)
 	  ASSERT (tc->cwnd >= tc->snd_mss);
 
 	  /* If cwnd allows, send more data */
-	  if (tcp_opts_sack_permitted (&tc->rcv_opts)
-	      && scoreboard_first_hole (&tc->sack_sb))
+	  if (tcp_opts_sack_permitted (&tc->rcv_opts))
 	    {
-	      scoreboard_init_high_rxt (&tc->sack_sb);
+	      scoreboard_init_high_rxt (&tc->sack_sb,
+					tc->snd_una + tc->snd_mss);
 	      tcp_fast_retransmit_sack (tc);
 	    }
 	  else
@@ -1134,12 +1160,13 @@ partial_ack:
   /* Remove retransmitted bytes that have been delivered */
   ASSERT (tc->bytes_acked + tc->sack_sb.snd_una_adv
 	  >= tc->sack_sb.last_bytes_delivered);
-  rxt_delivered = tc->bytes_acked + tc->sack_sb.snd_una_adv
-    - tc->sack_sb.last_bytes_delivered;
-  if (0 && rxt_delivered && seq_gt (tc->sack_sb.high_rxt, tc->snd_una))
+
+  if (seq_lt (tc->snd_una, tc->sack_sb.high_rxt))
     {
       /* If we have sacks and we haven't gotten an ack beyond high_rxt,
        * remove sacked bytes delivered */
+      rxt_delivered = tc->bytes_acked + tc->sack_sb.snd_una_adv
+	- tc->sack_sb.last_bytes_delivered;
       ASSERT (tc->snd_rxt_bytes >= rxt_delivered);
       tc->snd_rxt_bytes -= rxt_delivered;
     }
@@ -1256,6 +1283,18 @@ tcp_rcv_ack (tcp_connection_t * tc, vlib_buffer_t * b,
   return 0;
 }
 
+static u8
+tcp_sack_vector_is_sane (sack_block_t * sacks)
+{
+  int i;
+  for (i = 1; i < vec_len (sacks); i++)
+    {
+      if (sacks[i - 1].end == sacks[i].start)
+	return 0;
+    }
+  return 1;
+}
+
 /**
  * Build SACK list as per RFC2018.
  *
@@ -1316,6 +1355,9 @@ tcp_update_sack_list (tcp_connection_t * tc, u32 start, u32 end)
   /* Replace old vector with new one */
   vec_free (tc->snd_sacks);
   tc->snd_sacks = new_list;
+
+  /* Segments should not 'touch' */
+  ASSERT (tcp_sack_vector_is_sane (tc->snd_sacks));
 }
 
 /** Enqueue data for delivery to application */
@@ -1330,7 +1372,6 @@ tcp_session_enqueue_data (tcp_connection_t * tc, vlib_buffer_t * b,
   /* Pure ACK. Update rcv_nxt and be done. */
   if (PREDICT_FALSE (data_len == 0))
     {
-      tc->rcv_nxt = vnet_buffer (b)->tcp.seq_end;
       return TCP_ERROR_PURE_ACK;
     }
 
@@ -1385,7 +1426,7 @@ tcp_session_enqueue_ooo (tcp_connection_t * tc, vlib_buffer_t * b,
 			 u16 data_len)
 {
   stream_session_t *s0;
-  int rv;
+  int rv, offset;
 
   ASSERT (seq_gt (vnet_buffer (b)->tcp.seq_number, tc->rcv_nxt));
 
@@ -1421,12 +1462,12 @@ tcp_session_enqueue_ooo (tcp_connection_t * tc, vlib_buffer_t * b,
       newest = svm_fifo_newest_ooo_segment (s0->server_rx_fifo);
       if (newest)
 	{
-	  start =
-	    tc->rcv_nxt + ooo_segment_offset (s0->server_rx_fifo, newest);
+	  offset = ooo_segment_offset (s0->server_rx_fifo, newest);
+	  ASSERT (offset <= vnet_buffer (b)->tcp.seq_number - tc->rcv_nxt);
+	  start = tc->rcv_nxt + offset;
 	  end = start + ooo_segment_length (s0->server_rx_fifo, newest);
 	  tcp_update_sack_list (tc, start, end);
-
-	  ASSERT (seq_gt (start, tc->rcv_nxt));
+	  svm_fifo_newest_ooo_segment_reset (s0->server_rx_fifo);
 	}
     }
 
@@ -2736,12 +2777,12 @@ tcp46_input_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
 	      /* lookup session */
 	      tc0 =
 		(tcp_connection_t *)
-		stream_session_lookup_transport4 (&ip40->dst_address,
-						  &ip40->src_address,
-						  tcp0->dst_port,
-						  tcp0->src_port,
-						  SESSION_TYPE_IP4_TCP,
-						  my_thread_index);
+		stream_session_lookup_transport_wt4 (&ip40->dst_address,
+						     &ip40->src_address,
+						     tcp0->dst_port,
+						     tcp0->src_port,
+						     SESSION_TYPE_IP4_TCP,
+						     my_thread_index);
 	    }
 	  else
 	    {
@@ -2754,12 +2795,12 @@ tcp46_input_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
 
 	      tc0 =
 		(tcp_connection_t *)
-		stream_session_lookup_transport6 (&ip60->src_address,
-						  &ip60->dst_address,
-						  tcp0->src_port,
-						  tcp0->dst_port,
-						  SESSION_TYPE_IP6_TCP,
-						  my_thread_index);
+		stream_session_lookup_transport_wt6 (&ip60->src_address,
+						     &ip60->dst_address,
+						     tcp0->src_port,
+						     tcp0->dst_port,
+						     SESSION_TYPE_IP6_TCP,
+						     my_thread_index);
 	    }
 
 	  /* Length check */
@@ -2931,6 +2972,8 @@ do {                                                       	\
   _(ESTABLISHED, TCP_FLAG_RST | TCP_FLAG_ACK, TCP_INPUT_NEXT_ESTABLISHED,
     TCP_ERROR_NONE);
   _(ESTABLISHED, TCP_FLAG_SYN, TCP_INPUT_NEXT_ESTABLISHED, TCP_ERROR_NONE);
+  _(ESTABLISHED, TCP_FLAG_SYN | TCP_FLAG_ACK, TCP_INPUT_NEXT_ESTABLISHED,
+    TCP_ERROR_NONE);
   /* ACK or FIN-ACK to our FIN */
   _(FIN_WAIT_1, TCP_FLAG_ACK, TCP_INPUT_NEXT_RCV_PROCESS, TCP_ERROR_NONE);
   _(FIN_WAIT_1, TCP_FLAG_ACK | TCP_FLAG_FIN, TCP_INPUT_NEXT_RCV_PROCESS,
@@ -2954,6 +2997,7 @@ do {                                                       	\
   _(TIME_WAIT, TCP_FLAG_FIN, TCP_INPUT_NEXT_RCV_PROCESS, TCP_ERROR_NONE);
   _(TIME_WAIT, TCP_FLAG_FIN | TCP_FLAG_ACK, TCP_INPUT_NEXT_RCV_PROCESS,
     TCP_ERROR_NONE);
+  _(TIME_WAIT, TCP_FLAG_RST, TCP_INPUT_NEXT_RCV_PROCESS, TCP_ERROR_NONE);
   _(CLOSED, TCP_FLAG_ACK, TCP_INPUT_NEXT_RESET, TCP_ERROR_CONNECTION_CLOSED);
   _(CLOSED, TCP_FLAG_RST, TCP_INPUT_NEXT_DROP, TCP_ERROR_CONNECTION_CLOSED);
   _(CLOSED, TCP_FLAG_FIN | TCP_FLAG_ACK, TCP_INPUT_NEXT_RESET,
diff --git a/src/vnet/tcp/tcp_test.c b/src/vnet/tcp/tcp_test.c
index 510deb4f..f37ba96d 100644
--- a/src/vnet/tcp/tcp_test.c
+++ b/src/vnet/tcp/tcp_test.c
@@ -34,6 +34,38 @@
     }								\
 }
 
+/* *INDENT-OFF* */
+scoreboard_trace_elt_t sb_trace[] = {};
+/* *INDENT-ON* */
+
+static int
+tcp_test_scoreboard_replay (vlib_main_t * vm, unformat_input_t * input)
+{
+  int verbose = 0;
+  tcp_connection_t _tc, *tc = &_tc;
+  u8 *s = 0;
+
+  while (unformat_check_input (input) != UNFORMAT_END_OF_INPUT)
+    {
+      if (unformat (input, "detail"))
+	verbose = 1;
+      else
+	{
+	  clib_error_t *e = clib_error_return
+	    (0, "unknown input `%U'", format_unformat_error, input);
+	  clib_error_report (e);
+	  return -1;
+	}
+    }
+
+#if TCP_SCOREBOARD_TRACE
+  tc->sack_sb.trace = sb_trace;
+#endif
+  s = tcp_scoreboard_replay (s, tc, verbose);
+  vlib_cli_output (vm, "%v", s);
+  return 0;
+}
+
 static int
 tcp_test_sack_rx (vlib_main_t * vm, unformat_input_t * input)
 {
@@ -47,6 +79,8 @@ tcp_test_sack_rx (vlib_main_t * vm, unformat_input_t * input)
     {
       if (unformat (input, "verbose"))
 	verbose = 1;
+      else if (unformat (input, "replay"))
+	return tcp_test_scoreboard_replay (vm, input);
     }
 
   memset (tc, 0, sizeof (*tc));
@@ -282,6 +316,44 @@ tcp_test_sack_rx (vlib_main_t * vm, unformat_input_t * input)
   TCP_TEST ((sb->last_bytes_delivered == 400),
 	    "last bytes delivered %d", sb->last_bytes_delivered);
 
+  /*
+   * One hole close to head, patch head, split in two and start acking
+   * the lowest part
+   */
+  scoreboard_clear (sb);
+  tc->snd_una = 0;
+  tc->snd_una_max = 1000;
+  tc->snd_nxt = 1000;
+
+  block.start = 500;
+  block.end = 1000;
+  vec_add1 (tc->rcv_opts.sacks, block);
+  tc->rcv_opts.n_sack_blocks = vec_len (tc->rcv_opts.sacks);
+
+  tcp_rcv_sacks (tc, 0);
+  if (verbose)
+    vlib_cli_output (vm, "sb added [500, 1000]:\n%U",
+		     format_tcp_scoreboard, sb);
+
+  vec_reset_length (tc->rcv_opts.sacks);
+  block.start = 300;
+  block.end = 400;
+  vec_add1 (tc->rcv_opts.sacks, block);
+  tc->rcv_opts.n_sack_blocks = vec_len (tc->rcv_opts.sacks);
+  tcp_rcv_sacks (tc, 100);
+  if (verbose)
+    vlib_cli_output (vm, "sb added [0, 100] [300, 400]:\n%U",
+		     format_tcp_scoreboard, sb);
+  TCP_TEST ((pool_elts (sb->holes) == 2),
+	    "scoreboard has %d elements", pool_elts (sb->holes));
+
+  tc->snd_una = 100;
+  tcp_rcv_sacks (tc, 200);
+  tcp_rcv_sacks (tc, 300);
+  if (verbose)
+    vlib_cli_output (vm, "sb added [0, 300]:\n%U", format_tcp_scoreboard, sb);
+  TCP_TEST ((sb->sacked_bytes == 500), "sacked bytes %d", sb->sacked_bytes);
+
   return 0;
 }
 
@@ -390,6 +462,37 @@ tcp_test_sack_tx (vlib_main_t * vm, unformat_input_t * input)
     vlib_cli_output (vm, "advance rcv_nxt to 1200\n%U", format_tcp_sacks, tc);
   TCP_TEST ((vec_len (tc->snd_sacks) == 0), "sack blocks %d expected %d",
 	    vec_len (tc->snd_sacks), 0);
+
+
+  /*
+   * Add 2 blocks, overwrite first and update rcv_nxt to also remove it
+   */
+
+  vec_reset_length (tc->snd_sacks);
+  tc->rcv_nxt = 0;
+
+  tcp_update_sack_list (tc, 100, 200);
+  tcp_update_sack_list (tc, 300, 400);
+
+  if (verbose)
+    vlib_cli_output (vm, "add [100, 200] [300, 400]\n%U",
+		     format_tcp_sacks, tc);
+  TCP_TEST ((vec_len (tc->snd_sacks) == 2),
+	    "sack blocks %d expected %d", vec_len (tc->snd_sacks), 2);
+  TCP_TEST ((tc->snd_sacks[0].start == 300),
+	    "first sack block start %u expected %u", tc->snd_sacks[0].start,
+	    300);
+
+  tc->rcv_nxt = 100;
+  tcp_update_sack_list (tc, 100, 100);
+  if (verbose)
+    vlib_cli_output (vm, "add [100, 200] rcv_nxt = 100\n%U",
+		     format_tcp_sacks, tc);
+  TCP_TEST ((vec_len (tc->snd_sacks) == 1),
+	    "sack blocks %d expected %d", vec_len (tc->snd_sacks), 1);
+  TCP_TEST ((tc->snd_sacks[0].start == 300),
+	    "first sack block start %u expected %u", tc->snd_sacks[0].start,
+	    300);
   return 0;
 }
 
@@ -1188,6 +1291,176 @@ tcp_test_fifo4 (vlib_main_t * vm, unformat_input_t * input)
   return 0;
 }
 
+static u32
+fifo_pos (svm_fifo_t * f, u32 pos)
+{
+  return pos % f->nitems;
+}
+
+static int
+tcp_test_fifo5 (vlib_main_t * vm, unformat_input_t * input)
+{
+  svm_fifo_t *f;
+  u32 fifo_size = 400, j = 0, offset = 200;
+  int i, rv, verbose = 0;
+  u8 *test_data = 0, *data_buf = 0;
+  ooo_segment_t *ooo_seg;
+
+  while (unformat_check_input (input) != UNFORMAT_END_OF_INPUT)
+    {
+      if (unformat (input, "verbose"))
+	verbose = 1;
+      else
+	{
+	  clib_error_t *e = clib_error_return
+	    (0, "unknown input `%U'", format_unformat_error, input);
+	  clib_error_report (e);
+	  return -1;
+	}
+    }
+
+  f = fifo_prepare (fifo_size);
+  svm_fifo_init_pointers (f, offset);
+
+  vec_validate (test_data, 399);
+  for (i = 0; i < vec_len (test_data); i++)
+    test_data[i] = i % 0xff;
+
+  /*
+   * Start with [100, 200] and [300, 400]
+   */
+  svm_fifo_enqueue_with_offset (f, 100, 100, &test_data[100]);
+  svm_fifo_enqueue_with_offset (f, 300, 100, &test_data[300]);
+
+  TCP_TEST ((svm_fifo_number_ooo_segments (f) == 2),
+	    "number of ooo segments %u", svm_fifo_number_ooo_segments (f));
+  TCP_TEST ((f->ooos_newest == 1), "newest %u", f->ooos_newest);
+  if (verbose)
+    vlib_cli_output (vm, "fifo after [100, 200] and [300, 400] : %U",
+		     format_svm_fifo, f, 2 /* verbose */ );
+
+  /*
+   * Add [225, 275]
+   */
+
+  rv = svm_fifo_enqueue_with_offset (f, 225, 50, &test_data[200]);
+  if (verbose)
+    vlib_cli_output (vm, "fifo after [225, 275] : %U",
+		     format_svm_fifo, f, 2 /* verbose */ );
+  TCP_TEST ((svm_fifo_number_ooo_segments (f) == 3),
+	    "number of ooo segments %u", svm_fifo_number_ooo_segments (f));
+  ooo_seg = svm_fifo_first_ooo_segment (f);
+  TCP_TEST ((ooo_seg->start == fifo_pos (f, 100 + offset)),
+	    "first seg start %u expected %u", ooo_seg->start,
+	    fifo_pos (f, 100 + offset));
+  TCP_TEST ((ooo_seg->length == 100), "first seg length %u expected %u",
+	    ooo_seg->length, 100);
+  ooo_seg = ooo_segment_next (f, ooo_seg);
+  TCP_TEST ((ooo_seg->start == fifo_pos (f, 225 + offset)),
+	    "second seg start %u expected %u",
+	    ooo_seg->start, fifo_pos (f, 225 + offset));
+  TCP_TEST ((ooo_seg->length == 50), "second seg length %u expected %u",
+	    ooo_seg->length, 50);
+  ooo_seg = ooo_segment_next (f, ooo_seg);
+  TCP_TEST ((ooo_seg->start == fifo_pos (f, 300 + offset)),
+	    "third seg start %u expected %u",
+	    ooo_seg->start, fifo_pos (f, 300 + offset));
+  TCP_TEST ((ooo_seg->length == 100), "third seg length %u expected %u",
+	    ooo_seg->length, 100);
+  TCP_TEST ((f->ooos_newest == 2), "newest %u", f->ooos_newest);
+  /*
+   * Add [190, 310]
+   */
+  rv = svm_fifo_enqueue_with_offset (f, 190, 120, &test_data[190]);
+  if (verbose)
+    vlib_cli_output (vm, "fifo after [190, 310] : %U",
+		     format_svm_fifo, f, 1 /* verbose */ );
+  TCP_TEST ((svm_fifo_number_ooo_segments (f) == 1),
+	    "number of ooo segments %u", svm_fifo_number_ooo_segments (f));
+  ooo_seg = svm_fifo_first_ooo_segment (f);
+  TCP_TEST ((ooo_seg->start == fifo_pos (f, offset + 100)),
+	    "first seg start %u expected %u",
+	    ooo_seg->start, fifo_pos (f, offset + 100));
+  TCP_TEST ((ooo_seg->length == 300), "first seg length %u expected %u",
+	    ooo_seg->length, 300);
+
+  /*
+   * Add [0, 150]
+   */
+  rv = svm_fifo_enqueue_nowait (f, 150, test_data);
+
+  if (verbose)
+    vlib_cli_output (vm, "fifo after [0 150] : %U", format_svm_fifo, f,
+		     2 /* verbose */ );
+
+  TCP_TEST ((rv == 400), "managed to enqueue %u expected %u", rv, 400);
+  TCP_TEST ((svm_fifo_number_ooo_segments (f) == 0),
+	    "number of ooo segments %u", svm_fifo_number_ooo_segments (f));
+
+  vec_validate (data_buf, 399);
+  svm_fifo_peek (f, 0, 400, data_buf);
+  if (compare_data (data_buf, test_data, 0, 400, &j))
+    {
+      TCP_TEST (0, "[%d] peeked %u expected %u", j, data_buf[j],
+		test_data[j]);
+    }
+
+  /*
+   * Add [100 200] and overlap it with [50 250]
+   */
+  svm_fifo_free (f);
+  f = fifo_prepare (fifo_size);
+
+  svm_fifo_enqueue_with_offset (f, 100, 100, &test_data[100]);
+  svm_fifo_enqueue_with_offset (f, 50, 200, &test_data[50]);
+  TCP_TEST ((svm_fifo_number_ooo_segments (f) == 1),
+	    "number of ooo segments %u", svm_fifo_number_ooo_segments (f));
+  ooo_seg = svm_fifo_first_ooo_segment (f);
+  TCP_TEST ((ooo_seg->start == 50), "first seg start %u expected %u",
+	    ooo_seg->start, 50);
+  TCP_TEST ((ooo_seg->length == 200), "first seg length %u expected %u",
+	    ooo_seg->length, 200);
+
+  svm_fifo_free (f);
+  vec_free (test_data);
+  return 0;
+}
+
+/* *INDENT-OFF* */
+svm_fifo_trace_elem_t fifo_trace[] = {};
+/* *INDENT-ON* */
+
+static int
+tcp_test_fifo_replay (vlib_main_t * vm, unformat_input_t * input)
+{
+  svm_fifo_t f;
+  int verbose = 0;
+  u8 no_read = 0, *str = 0;
+
+  while (unformat_check_input (input) != UNFORMAT_END_OF_INPUT)
+    {
+      if (unformat (input, "verbose"))
+	verbose = 1;
+      else if (unformat (input, "no-read"))
+	no_read = 1;
+      else
+	{
+	  clib_error_t *e = clib_error_return
+	    (0, "unknown input `%U'", format_unformat_error, input);
+	  clib_error_report (e);
+	  return -1;
+	}
+    }
+
+#if SVMF_FIFO_TRACE
+  f.trace = fifo_trace;
+#endif
+
+  str = svm_fifo_replay (str, &f, no_read, verbose);
+  vlib_cli_output (vm, "%v", str);
+  return 0;
+}
+
 static int
 tcp_test_fifo (vlib_main_t * vm, unformat_input_t * input)
 {
@@ -1237,6 +1510,14 @@ tcp_test_fifo (vlib_main_t * vm, unformat_input_t * input)
       if (tcp_test_fifo3 (vm, input))
 	return -1;
       unformat_free (input);
+
+      res = tcp_test_fifo4 (vm, input);
+      if (res)
+	return res;
+
+      res = tcp_test_fifo5 (vm, input);
+      if (res)
+	return res;
     }
   else
     {
@@ -1256,6 +1537,14 @@ tcp_test_fifo (vlib_main_t * vm, unformat_input_t * input)
 	{
 	  res = tcp_test_fifo4 (vm, input);
 	}
+      else if (unformat (input, "fifo5"))
+	{
+	  res = tcp_test_fifo5 (vm, input);
+	}
+      else if (unformat (input, "replay"))
+	{
+	  res = tcp_test_fifo_replay (vm, input);
+	}
     }
 
   return res;
-- 
cgit 1.2.3-korg


From 52851e6aa9304054fd1059c8dd284abf8e532bf2 Mon Sep 17 00:00:00 2001
From: Dave Barach <dave@barachs.net>
Date: Mon, 7 Aug 2017 09:35:25 -0400
Subject: TCP proxy prototype

- Clean up internal API client registration
- Add proxy server
- Add a reference count to the svm fifo

Change-Id: I5ace1c85497062ed412d26ae76a9e6741af1e984
Signed-off-by: Dave Barach <dave@barachs.net>
Signed-off-by: Florin Coras <fcoras@cisco.com>
---
 src/svm/svm_fifo.c                       |  10 +-
 src/svm/svm_fifo.h                       |   1 +
 src/svm/svm_fifo_segment.c               |   5 +
 src/vlibmemory/api_common.h              |   1 +
 src/vlibmemory/memory_vlib.c             |  38 ++
 src/vnet.am                              |   1 +
 src/vnet/session/application.c           |   6 +
 src/vnet/session/application.h           |   1 +
 src/vnet/session/application_interface.h |   4 +
 src/vnet/session/session.c               |  43 ++-
 src/vnet/session/session.h               |   8 +-
 src/vnet/session/stream_session.h        |   5 +-
 src/vnet/tcp/builtin_client.c            |  95 +----
 src/vnet/tcp/builtin_proxy.c             | 599 +++++++++++++++++++++++++++++++
 src/vnet/tcp/builtin_proxy.h             | 100 ++++++
 src/vnet/tcp/builtin_server.c            |  96 +----
 16 files changed, 801 insertions(+), 212 deletions(-)
 create mode 100644 src/vnet/tcp/builtin_proxy.c
 create mode 100644 src/vnet/tcp/builtin_proxy.h

(limited to 'src/svm/svm_fifo.c')

diff --git a/src/svm/svm_fifo.c b/src/svm/svm_fifo.c
index fc2189c5..e478c06e 100644
--- a/src/svm/svm_fifo.c
+++ b/src/svm/svm_fifo.c
@@ -201,14 +201,20 @@ svm_fifo_create (u32 data_size_in_bytes)
   memset (f, 0, sizeof (*f));
   f->nitems = data_size_in_bytes;
   f->ooos_list_head = OOO_SEGMENT_INVALID_INDEX;
+  f->refcnt = 1;
   return (f);
 }
 
 void
 svm_fifo_free (svm_fifo_t * f)
 {
-  pool_free (f->ooo_segments);
-  clib_mem_free (f);
+  ASSERT (f->refcnt > 0);
+
+  if (--f->refcnt == 0)
+    {
+      pool_free (f->ooo_segments);
+      clib_mem_free (f);
+    }
 }
 
 always_inline ooo_segment_t *
diff --git a/src/svm/svm_fifo.h b/src/svm/svm_fifo.h
index a83cd858..f10b4d91 100644
--- a/src/svm/svm_fifo.h
+++ b/src/svm/svm_fifo.h
@@ -75,6 +75,7 @@ typedef struct _svm_fifo
 #if SVM_FIFO_TRACE
   svm_fifo_trace_elem_t *trace;
 #endif
+  i8 refcnt;
     CLIB_CACHE_LINE_ALIGN_MARK (data);
 } svm_fifo_t;
 
diff --git a/src/svm/svm_fifo_segment.c b/src/svm/svm_fifo_segment.c
index 86661e51..c04b9d8c 100644
--- a/src/svm/svm_fifo_segment.c
+++ b/src/svm/svm_fifo_segment.c
@@ -296,6 +296,7 @@ svm_fifo_segment_alloc_fifo (svm_fifo_segment_private_t * s,
 	  memset (f, 0, sizeof (*f));
 	  f->nitems = data_size_in_bytes;
 	  f->ooos_list_head = OOO_SEGMENT_INVALID_INDEX;
+	  f->refcnt = 1;
 	  goto found;
 	}
       /* FALLTHROUGH */
@@ -344,6 +345,10 @@ svm_fifo_segment_free_fifo (svm_fifo_segment_private_t * s, svm_fifo_t * f,
   svm_fifo_segment_header_t *fsh;
   void *oldheap;
 
+  ASSERT (f->refcnt > 0);
+
+  if (--f->refcnt > 0)
+    return;
 
   sh = s->ssvm.sh;
   fsh = (svm_fifo_segment_header_t *) sh->opaque[0];
diff --git a/src/vlibmemory/api_common.h b/src/vlibmemory/api_common.h
index 53909cc4..19daecdf 100644
--- a/src/vlibmemory/api_common.h
+++ b/src/vlibmemory/api_common.h
@@ -126,6 +126,7 @@ int vl_client_connect_to_vlib_no_rx_pthread (const char *svm_name,
 u16 vl_client_get_first_plugin_msg_id (const char *plugin_name);
 
 void vl_api_rpc_call_main_thread (void *fp, u8 * data, u32 data_length);
+u32 vl_api_memclnt_create_internal (char *, unix_shared_memory_queue_t *);
 
 #endif /* included_vlibmemory_api_common_h */
 
diff --git a/src/vlibmemory/memory_vlib.c b/src/vlibmemory/memory_vlib.c
index 004a9974..688ce604 100644
--- a/src/vlibmemory/memory_vlib.c
+++ b/src/vlibmemory/memory_vlib.c
@@ -136,6 +136,44 @@ vl_api_serialize_message_table (api_main_t * am, u8 * vector)
   return serialize_close_vector (sm);
 }
 
+/*
+ * vl_api_memclnt_create_internal
+ */
+
+u32
+vl_api_memclnt_create_internal (char *name, unix_shared_memory_queue_t * q)
+{
+  vl_api_registration_t **regpp;
+  vl_api_registration_t *regp;
+  svm_region_t *svm;
+  void *oldheap;
+  api_main_t *am = &api_main;
+
+  ASSERT (vlib_get_thread_index () == 0);
+  pool_get (am->vl_clients, regpp);
+
+  svm = am->vlib_rp;
+
+  pthread_mutex_lock (&svm->mutex);
+  oldheap = svm_push_data_heap (svm);
+  *regpp = clib_mem_alloc (sizeof (vl_api_registration_t));
+
+  regp = *regpp;
+  memset (regp, 0, sizeof (*regp));
+  regp->registration_type = REGISTRATION_TYPE_SHMEM;
+  regp->vl_api_registration_pool_index = regpp - am->vl_clients;
+
+  regp->vl_input_queue = q;
+  regp->name = format (0, "%s%c", name, 0);
+
+  pthread_mutex_unlock (&svm->mutex);
+  svm_pop_heap (oldheap);
+  return vl_msg_api_handle_from_index_and_epoch
+    (regp->vl_api_registration_pool_index,
+     am->shmem_hdr->application_restarts);
+}
+
+
 /*
  * vl_api_memclnt_create_t_handler
  */
diff --git a/src/vnet.am b/src/vnet.am
index ad84c028..ede0376d 100644
--- a/src/vnet.am
+++ b/src/vnet.am
@@ -475,6 +475,7 @@ libvnet_la_SOURCES +=				\
  vnet/tcp/builtin_client.c			\
  vnet/tcp/builtin_server.c			\
  vnet/tcp/builtin_http_server.c			\
+ vnet/tcp/builtin_proxy.c			\
  vnet/tcp/tcp_test.c				\
  vnet/tcp/tcp.c
 
diff --git a/src/vnet/session/application.c b/src/vnet/session/application.c
index 076c66f1..78c41b93 100644
--- a/src/vnet/session/application.c
+++ b/src/vnet/session/application.c
@@ -367,6 +367,12 @@ app_get_name_from_reg_index (application_t * app)
   return app_name;
 }
 
+int
+application_is_proxy (application_t * app)
+{
+  return !(app->flags & APP_OPTIONS_FLAGS_IS_PROXY);
+}
+
 u8 *
 format_application_listener (u8 * s, va_list * args)
 {
diff --git a/src/vnet/session/application.h b/src/vnet/session/application.h
index 35caae85..29d37a06 100644
--- a/src/vnet/session/application.h
+++ b/src/vnet/session/application.h
@@ -116,6 +116,7 @@ segment_manager_t *application_get_listen_segment_manager (application_t *
 							   s);
 segment_manager_t *application_get_connect_segment_manager (application_t *
 							    app);
+int application_is_proxy (application_t * app);
 
 #endif /* SRC_VNET_SESSION_APPLICATION_H_ */
 
diff --git a/src/vnet/session/application_interface.h b/src/vnet/session/application_interface.h
index ed9f89b3..1d63f6cc 100644
--- a/src/vnet/session/application_interface.h
+++ b/src/vnet/session/application_interface.h
@@ -100,6 +100,9 @@ typedef struct _vnet_connect_args
 
   /* Used for redirects */
   void *mp;
+
+  /* used for proxy connections */
+  u64 server_handle;
 } vnet_connect_args_t;
 
 typedef struct _vnet_disconnect_args_t
@@ -129,6 +132,7 @@ typedef enum
   _(USE_FIFO, "Use FIFO with redirects")			\
   _(ADD_SEGMENT, "Add segment and signal app if needed")	\
   _(BUILTIN_APP, "Application is builtin")			\
+  _(IS_PROXY, "Application is proxying")
 
 typedef enum _app_options
 {
diff --git a/src/vnet/session/session.c b/src/vnet/session/session.c
index 4ba15291..991bcd5a 100644
--- a/src/vnet/session/session.c
+++ b/src/vnet/session/session.c
@@ -30,7 +30,7 @@ extern transport_proto_vft_t *tp_vfts;
 
 int
 stream_session_create_i (segment_manager_t * sm, transport_connection_t * tc,
-			 stream_session_t ** ret_s)
+			 u8 alloc_fifos, stream_session_t ** ret_s)
 {
   session_manager_main_t *smm = &session_manager_main;
   svm_fifo_t *server_rx_fifo = 0, *server_tx_fifo = 0;
@@ -43,31 +43,37 @@ stream_session_create_i (segment_manager_t * sm, transport_connection_t * tc,
 
   ASSERT (thread_index == vlib_get_thread_index ());
 
-  if ((rv = segment_manager_alloc_session_fifos (sm, &server_rx_fifo,
-						 &server_tx_fifo,
-						 &fifo_segment_index)))
-    return rv;
-
   /* Create the session */
   pool_get_aligned (smm->sessions[thread_index], s, CLIB_CACHE_LINE_BYTES);
   memset (s, 0, sizeof (*s));
-
-  /* Initialize backpointers */
   pool_index = s - smm->sessions[thread_index];
-  server_rx_fifo->master_session_index = pool_index;
-  server_rx_fifo->master_thread_index = thread_index;
 
-  server_tx_fifo->master_session_index = pool_index;
-  server_tx_fifo->master_thread_index = thread_index;
+  /* Allocate fifos */
+  if (alloc_fifos)
+    {
+      if ((rv = segment_manager_alloc_session_fifos (sm, &server_rx_fifo,
+						     &server_tx_fifo,
+						     &fifo_segment_index)))
+	{
+	  pool_put (smm->sessions[thread_index], s);
+	  return rv;
+	}
+      /* Initialize backpointers */
+      server_rx_fifo->master_session_index = pool_index;
+      server_rx_fifo->master_thread_index = thread_index;
+
+      server_tx_fifo->master_session_index = pool_index;
+      server_tx_fifo->master_thread_index = thread_index;
 
-  s->server_rx_fifo = server_rx_fifo;
-  s->server_tx_fifo = server_tx_fifo;
+      s->server_rx_fifo = server_rx_fifo;
+      s->server_tx_fifo = server_tx_fifo;
+      s->svm_segment_index = fifo_segment_index;
+    }
 
   /* Initialize state machine, such as it is... */
   s->session_type = session_type_from_proto_and_ip (tc->transport_proto,
 						    tc->is_ip4);
   s->session_state = SESSION_STATE_CONNECTING;
-  s->svm_segment_index = fifo_segment_index;
   s->thread_index = thread_index;
   s->session_index = pool_index;
 
@@ -379,10 +385,11 @@ stream_session_connect_notify (transport_connection_t * tc, u8 is_fail)
   if (!is_fail)
     {
       segment_manager_t *sm;
+      u8 alloc_fifos;
       sm = application_get_connect_segment_manager (app);
-
+      alloc_fifos = application_is_proxy (app);
       /* Create new session (svm segments are allocated if needed) */
-      if (stream_session_create_i (sm, tc, &new_s))
+      if (stream_session_create_i (sm, tc, alloc_fifos, &new_s))
 	{
 	  is_fail = 1;
 	  error = -1;
@@ -515,7 +522,7 @@ stream_session_accept (transport_connection_t * tc, u32 listener_index,
   server = application_get (listener->app_index);
 
   sm = application_get_listen_segment_manager (server, listener);
-  if ((rv = stream_session_create_i (sm, tc, &s)))
+  if ((rv = stream_session_create_i (sm, tc, 1, &s)))
     return rv;
 
   s->app_index = server->index;
diff --git a/src/vnet/session/session.h b/src/vnet/session/session.h
index 538433da..74d82a40 100644
--- a/src/vnet/session/session.h
+++ b/src/vnet/session/session.h
@@ -183,10 +183,10 @@ stream_session_is_valid (u32 si, u8 thread_index)
   stream_session_t *s;
   s = pool_elt_at_index (session_manager_main.sessions[thread_index], si);
   if (s->thread_index != thread_index || s->session_index != si
-      || s->server_rx_fifo->master_session_index != si
-      || s->server_tx_fifo->master_session_index != si
-      || s->server_rx_fifo->master_thread_index != thread_index
-      || s->server_tx_fifo->master_thread_index != thread_index)
+      /* || s->server_rx_fifo->master_session_index != si
+         || s->server_tx_fifo->master_session_index != si
+         || s->server_rx_fifo->master_thread_index != thread_index
+         || s->server_tx_fifo->master_thread_index != thread_index */ )
     return 0;
   return 1;
 }
diff --git a/src/vnet/session/stream_session.h b/src/vnet/session/stream_session.h
index 82bbf521..4c263211 100644
--- a/src/vnet/session/stream_session.h
+++ b/src/vnet/session/stream_session.h
@@ -83,8 +83,11 @@ typedef struct _stream_session_t
 
   u32 opaque2;
 
+  /** connected (server) session handle */
+  u64 server_session_handle;
+
   /** Opaque, pad to a 64-octet boundary */
-  u64 opaque[2];
+  u64 opaque[1];
 } stream_session_t;
 
 #endif /* SRC_VNET_SESSION_STREAM_SESSION_H_ */
diff --git a/src/vnet/tcp/builtin_client.c b/src/vnet/tcp/builtin_client.c
index 0cb9faa8..938e07ba 100644
--- a/src/vnet/tcp/builtin_client.c
+++ b/src/vnet/tcp/builtin_client.c
@@ -24,25 +24,6 @@
 #include <vlibsocket/api.h>
 #include <vpp/app/version.h>
 
-/* define message IDs */
-#include <vpp/api/vpe_msg_enum.h>
-
-/* define message structures */
-#define vl_typedefs
-#include <vpp/api/vpe_all_api_h.h>
-#undef vl_typedefs
-
-/* define generated endian-swappers */
-#define vl_endianfun
-#include <vpp/api/vpe_all_api_h.h>
-#undef vl_endianfun
-
-/* instantiate all the print functions we know about */
-#define vl_print(handle, ...) vlib_cli_output (handle, __VA_ARGS__)
-#define vl_printfun
-#include <vpp/api/vpe_all_api_h.h>
-#undef vl_printfun
-
 #define TCP_BUILTIN_CLIENT_DBG (0)
 
 static void
@@ -308,87 +289,16 @@ VLIB_REGISTER_NODE (builtin_client_node) =
 };
 /* *INDENT-ON* */
 
-/* So we don't get "no handler for... " msgs */
-static void
-vl_api_memclnt_create_reply_t_handler (vl_api_memclnt_create_reply_t * mp)
-{
-  vlib_main_t *vm = vlib_get_main ();
-  tclient_main_t *tm = &tclient_main;
-  tm->my_client_index = mp->index;
-  vlib_process_signal_event (vm, tm->cli_node_index, 1 /* evt */ ,
-			     0 /* data */ );
-}
-
 static int
 create_api_loopback (tclient_main_t * tm)
 {
-  vlib_main_t *vm = vlib_get_main ();
-  vl_api_memclnt_create_t _m, *mp = &_m;
-  extern void vl_api_memclnt_create_t_handler (vl_api_memclnt_create_t *);
   api_main_t *am = &api_main;
   vl_shmem_hdr_t *shmem_hdr;
-  uword *event_data = 0, event_type;
-  int resolved = 0;
-
-  /*
-   * Create a "loopback" API client connection
-   * Don't do things like this unless you know what you're doing...
-   */
 
   shmem_hdr = am->shmem_hdr;
   tm->vl_input_queue = shmem_hdr->vl_input_queue;
-  memset (mp, 0, sizeof (*mp));
-  mp->_vl_msg_id = VL_API_MEMCLNT_CREATE;
-  mp->context = 0xFEEDFACE;
-  mp->input_queue = pointer_to_uword (tm->vl_input_queue);
-  strncpy ((char *) mp->name, "tcp_clients_tester", sizeof (mp->name) - 1);
-
-  vl_api_memclnt_create_t_handler (mp);
-
-  /* Wait for reply */
-  vlib_process_wait_for_event_or_clock (vm, 1.0);
-  event_type = vlib_process_get_events (vm, &event_data);
-  switch (event_type)
-    {
-    case 1:
-      resolved = 1;
-      break;
-    case ~0:
-      /* timed out */
-      break;
-    default:
-      clib_warning ("unknown event_type %d", event_type);
-    }
-  if (!resolved)
-    return -1;
-  return 0;
-}
-
-#define foreach_tclient_static_api_msg       	\
-_(MEMCLNT_CREATE_REPLY, memclnt_create_reply)   \
-
-static clib_error_t *
-tclient_api_hookup (vlib_main_t * vm)
-{
-  vl_msg_api_msg_config_t _c, *c = &_c;
-
-  /* Hook up client-side static APIs to our handlers */
-#define _(N,n) do {                                             \
-    c->id = VL_API_##N;                                         \
-    c->name = #n;                                               \
-    c->handler = vl_api_##n##_t_handler;                        \
-    c->cleanup = vl_noop_handler;                               \
-    c->endian = vl_api_##n##_t_endian;                          \
-    c->print = vl_api_##n##_t_print;                            \
-    c->size = sizeof(vl_api_##n##_t);                           \
-    c->traced = 1; /* trace, so these msgs print */             \
-    c->replay = 0; /* don't replay client create/delete msgs */ \
-    c->message_bounce = 0; /* don't bounce this message */	\
-    vl_msg_api_config(c);} while (0);
-
-  foreach_tclient_static_api_msg;
-#undef _
-
+  tm->my_client_index =
+    vl_api_memclnt_create_internal ("tcp_test_client", tm->vl_input_queue);
   return 0;
 }
 
@@ -400,7 +310,6 @@ tcp_test_clients_init (vlib_main_t * vm)
   u32 num_threads;
   int i;
 
-  tclient_api_hookup (vm);
   if (create_api_loopback (tm))
     return -1;
 
diff --git a/src/vnet/tcp/builtin_proxy.c b/src/vnet/tcp/builtin_proxy.c
new file mode 100644
index 00000000..d8cfb11d
--- /dev/null
+++ b/src/vnet/tcp/builtin_proxy.c
@@ -0,0 +1,599 @@
+/*
+* Copyright (c) 2015-2017 Cisco and/or its affiliates.
+* Licensed under the Apache License, Version 2.0 (the "License");
+* you may not use this file except in compliance with the License.
+* You may obtain a copy of the License at:
+*
+*     http://www.apache.org/licenses/LICENSE-2.0
+*
+* Unless required by applicable law or agreed to in writing, software
+* distributed under the License is distributed on an "AS IS" BASIS,
+* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+* See the License for the specific language governing permissions and
+* limitations under the License.
+*/
+
+#include <vnet/vnet.h>
+#include <vlibmemory/api.h>
+#include <vnet/session/application.h>
+#include <vnet/session/application_interface.h>
+#include <vnet/tcp/builtin_proxy.h>
+
+builtin_proxy_main_t builtin_proxy_main;
+
+static void
+delete_proxy_session (stream_session_t * s, int is_active_open)
+{
+  builtin_proxy_main_t *bpm = &builtin_proxy_main;
+  proxy_session_t *ps = 0;
+  vnet_disconnect_args_t _a, *a = &_a;
+  stream_session_t *active_open_session = 0;
+  stream_session_t *server_session = 0;
+  uword *p;
+  u64 handle;
+
+  handle = stream_session_handle (s);
+
+  clib_spinlock_lock_if_init (&bpm->sessions_lock);
+  if (is_active_open)
+    {
+      active_open_session = s;
+
+      p = hash_get (bpm->proxy_session_by_active_open_handle, handle);
+      if (p == 0)
+	{
+	  clib_warning ("proxy session for %s handle %lld (%llx) AWOL",
+			is_active_open ? "active open" : "server",
+			handle, handle);
+	}
+      else
+	{
+	  ps = pool_elt_at_index (bpm->sessions, p[0]);
+	  if (ps->vpp_server_handle != ~0)
+	    server_session = stream_session_get_from_handle
+	      (ps->vpp_server_handle);
+	  else
+	    server_session = 0;
+	}
+    }
+  else
+    {
+      server_session = s;
+
+      p = hash_get (bpm->proxy_session_by_server_handle, handle);
+      if (p == 0)
+	{
+	  clib_warning ("proxy session for %s handle %lld (%llx) AWOL",
+			is_active_open ? "active open" : "server",
+			handle, handle);
+	}
+      else
+	{
+	  ps = pool_elt_at_index (bpm->sessions, p[0]);
+	  if (ps->vpp_server_handle != ~0)
+	    active_open_session = stream_session_get_from_handle
+	      (ps->vpp_server_handle);
+	  else
+	    active_open_session = 0;
+	}
+    }
+
+  if (ps)
+    {
+      if (CLIB_DEBUG > 0)
+	memset (ps, 0xFE, sizeof (*ps));
+      pool_put (bpm->sessions, ps);
+    }
+
+  clib_spinlock_unlock_if_init (&bpm->sessions_lock);
+
+  if (active_open_session)
+    {
+      a->handle = stream_session_handle (active_open_session);
+      a->app_index = bpm->active_open_app_index;
+      hash_unset (bpm->proxy_session_by_active_open_handle,
+		  stream_session_handle (active_open_session));
+      vnet_disconnect_session (a);
+    }
+
+  if (server_session)
+    {
+      a->handle = stream_session_handle (server_session);
+      a->app_index = bpm->server_app_index;
+      hash_unset (bpm->proxy_session_by_server_handle,
+		  stream_session_handle (server_session));
+      vnet_disconnect_session (a);
+    }
+}
+
+static int
+server_accept_callback (stream_session_t * s)
+{
+  builtin_proxy_main_t *bpm = &builtin_proxy_main;
+
+  s->session_state = SESSION_STATE_READY;
+
+  clib_spinlock_lock_if_init (&bpm->sessions_lock);
+
+  return 0;
+}
+
+static void
+server_disconnect_callback (stream_session_t * s)
+{
+  delete_proxy_session (s, 0 /* is_active_open */ );
+}
+
+static void
+server_reset_callback (stream_session_t * s)
+{
+  clib_warning ("Reset session %U", format_stream_session, s, 2);
+  delete_proxy_session (s, 0 /* is_active_open */ );
+}
+
+static int
+server_connected_callback (u32 app_index, u32 api_context,
+			   stream_session_t * s, u8 is_fail)
+{
+  clib_warning ("called...");
+  return -1;
+}
+
+static int
+server_add_segment_callback (u32 client_index,
+			     const u8 * seg_name, u32 seg_size)
+{
+  clib_warning ("called...");
+  return -1;
+}
+
+static int
+server_redirect_connect_callback (u32 client_index, void *mp)
+{
+  clib_warning ("called...");
+  return -1;
+}
+
+static int
+server_rx_callback (stream_session_t * s)
+{
+  u32 max_dequeue;
+  int actual_transfer __attribute__ ((unused));
+  svm_fifo_t *tx_fifo, *rx_fifo;
+  builtin_proxy_main_t *bpm = &builtin_proxy_main;
+  u32 thread_index = vlib_get_thread_index ();
+  vnet_connect_args_t _a, *a = &_a;
+  proxy_session_t *ps;
+  int proxy_index;
+  uword *p;
+  svm_fifo_t *active_open_tx_fifo;
+  session_fifo_event_t evt;
+
+  ASSERT (s->thread_index == thread_index);
+
+  clib_spinlock_lock_if_init (&bpm->sessions_lock);
+  p =
+    hash_get (bpm->proxy_session_by_server_handle, stream_session_handle (s));
+
+  if (PREDICT_TRUE (p != 0))
+    {
+      clib_spinlock_unlock_if_init (&bpm->sessions_lock);
+      active_open_tx_fifo = s->server_rx_fifo;
+
+      /*
+       * Send event for active open tx fifo
+       */
+      if (svm_fifo_set_event (active_open_tx_fifo))
+	{
+	  evt.fifo = active_open_tx_fifo;
+	  evt.event_type = FIFO_EVENT_APP_TX;
+	  if (unix_shared_memory_queue_add
+	      (bpm->active_open_event_queue[thread_index], (u8 *) & evt,
+	       0 /* do wait for mutex */ ))
+	    clib_warning ("failed to enqueue tx evt");
+	}
+    }
+  else
+    {
+      rx_fifo = s->server_rx_fifo;
+      tx_fifo = s->server_tx_fifo;
+
+      ASSERT (rx_fifo->master_thread_index == thread_index);
+      ASSERT (tx_fifo->master_thread_index == thread_index);
+
+      max_dequeue = svm_fifo_max_dequeue (s->server_rx_fifo);
+
+      if (PREDICT_FALSE (max_dequeue == 0))
+	return 0;
+
+      actual_transfer = svm_fifo_peek (rx_fifo, 0 /* relative_offset */ ,
+				       max_dequeue,
+				       bpm->rx_buf[thread_index]);
+
+      /* $$$ your message in this space: parse url, etc. */
+
+      memset (a, 0, sizeof (*a));
+
+      clib_spinlock_lock_if_init (&bpm->sessions_lock);
+      pool_get (bpm->sessions, ps);
+      memset (ps, 0, sizeof (*ps));
+      ps->server_rx_fifo = rx_fifo;
+      ps->server_tx_fifo = tx_fifo;
+      ps->vpp_server_handle = stream_session_handle (s);
+
+      proxy_index = ps - bpm->sessions;
+
+      hash_set (bpm->proxy_session_by_server_handle, ps->vpp_server_handle,
+		proxy_index);
+
+      clib_spinlock_unlock_if_init (&bpm->sessions_lock);
+
+      a->uri = "tcp://6.0.2.2/23";
+      a->api_context = proxy_index;
+      a->app_index = bpm->active_open_app_index;
+      a->mp = 0;
+      vnet_connect_uri (a);
+    }
+
+  return 0;
+}
+
+static session_cb_vft_t builtin_session_cb_vft = {
+  .session_accept_callback = server_accept_callback,
+  .session_disconnect_callback = server_disconnect_callback,
+  .session_connected_callback = server_connected_callback,
+  .add_segment_callback = server_add_segment_callback,
+  .redirect_connect_callback = server_redirect_connect_callback,
+  .builtin_server_rx_callback = server_rx_callback,
+  .session_reset_callback = server_reset_callback
+};
+
+static int
+active_open_connected_callback (u32 app_index, u32 opaque,
+				stream_session_t * s, u8 is_fail)
+{
+  builtin_proxy_main_t *bpm = &builtin_proxy_main;
+  proxy_session_t *ps;
+  u8 thread_index = vlib_get_thread_index ();
+  session_fifo_event_t evt;
+
+  if (is_fail)
+    {
+      clib_warning ("connection %d failed!", opaque);
+      return 0;
+    }
+
+  /*
+   * Setup proxy session handle.
+   */
+  clib_spinlock_lock_if_init (&bpm->sessions_lock);
+
+  ps = pool_elt_at_index (bpm->sessions, opaque);
+  ps->vpp_active_open_handle = stream_session_handle (s);
+
+  s->server_tx_fifo = ps->server_rx_fifo;
+  s->server_rx_fifo = ps->server_tx_fifo;
+
+  /*
+   * Reset the active-open tx-fifo master indices so the active-open session
+   * will receive data, etc.
+   */
+  s->server_tx_fifo->master_session_index = s->session_index;
+  s->server_tx_fifo->master_thread_index = s->thread_index;
+
+  /*
+   * Account for the active-open session's use of the fifos
+   * so they won't disappear until the last session which uses
+   * them disappears
+   */
+  s->server_tx_fifo->refcnt++;
+  s->server_rx_fifo->refcnt++;
+
+  hash_set (bpm->proxy_session_by_active_open_handle,
+	    ps->vpp_active_open_handle, opaque);
+
+  clib_spinlock_unlock_if_init (&bpm->sessions_lock);
+
+  /*
+   * Send event for active open tx fifo
+   */
+  if (svm_fifo_set_event (s->server_tx_fifo))
+    {
+      evt.fifo = s->server_tx_fifo;
+      evt.event_type = FIFO_EVENT_APP_TX;
+      if (unix_shared_memory_queue_add
+	  (bpm->active_open_event_queue[thread_index], (u8 *) & evt,
+	   0 /* do wait for mutex */ ))
+	clib_warning ("failed to enqueue tx evt");
+    }
+
+  return 0;
+}
+
+static void
+active_open_reset_callback (stream_session_t * s)
+{
+  delete_proxy_session (s, 1 /* is_active_open */ );
+}
+
+static int
+active_open_create_callback (stream_session_t * s)
+{
+  return 0;
+}
+
+static void
+active_open_disconnect_callback (stream_session_t * s)
+{
+  delete_proxy_session (s, 1 /* is_active_open */ );
+}
+
+static int
+active_open_rx_callback (stream_session_t * s)
+{
+  builtin_proxy_main_t *bpm = &builtin_proxy_main;
+  session_fifo_event_t evt;
+  svm_fifo_t *server_rx_fifo;
+  u32 thread_index = vlib_get_thread_index ();
+
+  server_rx_fifo = s->server_rx_fifo;
+
+  /*
+   * Send event for server tx fifo
+   */
+  if (svm_fifo_set_event (server_rx_fifo))
+    {
+      evt.fifo = server_rx_fifo;
+      evt.event_type = FIFO_EVENT_APP_TX;
+      if (unix_shared_memory_queue_add
+	  (bpm->server_event_queue[thread_index], (u8 *) & evt,
+	   0 /* do wait for mutex */ ))
+	clib_warning ("failed to enqueue server rx evt");
+    }
+
+  return 0;
+}
+
+/* *INDENT-OFF* */
+static session_cb_vft_t builtin_clients = {
+  .session_reset_callback = active_open_reset_callback,
+  .session_connected_callback = active_open_connected_callback,
+  .session_accept_callback = active_open_create_callback,
+  .session_disconnect_callback = active_open_disconnect_callback,
+  .builtin_server_rx_callback = active_open_rx_callback
+};
+/* *INDENT-ON* */
+
+
+static void
+create_api_loopbacks (vlib_main_t * vm)
+{
+  builtin_proxy_main_t *bpm = &builtin_proxy_main;
+  api_main_t *am = &api_main;
+  vl_shmem_hdr_t *shmem_hdr;
+
+  shmem_hdr = am->shmem_hdr;
+  bpm->vl_input_queue = shmem_hdr->vl_input_queue;
+  bpm->server_client_index =
+    vl_api_memclnt_create_internal ("proxy_server", bpm->vl_input_queue);
+  bpm->active_open_client_index =
+    vl_api_memclnt_create_internal ("proxy_active_open", bpm->vl_input_queue);
+}
+
+static int
+server_attach ()
+{
+  builtin_proxy_main_t *bpm = &builtin_proxy_main;
+  u8 segment_name[128];
+  u64 options[SESSION_OPTIONS_N_OPTIONS];
+  vnet_app_attach_args_t _a, *a = &_a;
+
+  memset (a, 0, sizeof (*a));
+  memset (options, 0, sizeof (options));
+
+  a->api_client_index = bpm->server_client_index;
+  a->session_cb_vft = &builtin_session_cb_vft;
+  a->options = options;
+  a->options[SESSION_OPTIONS_SEGMENT_SIZE] = 512 << 20;
+  a->options[SESSION_OPTIONS_RX_FIFO_SIZE] = bpm->fifo_size;
+  a->options[SESSION_OPTIONS_TX_FIFO_SIZE] = bpm->fifo_size;
+  a->options[APP_OPTIONS_PRIVATE_SEGMENT_COUNT] = bpm->private_segment_count;
+  a->options[APP_OPTIONS_PRIVATE_SEGMENT_SIZE] = bpm->private_segment_size;
+  a->options[APP_OPTIONS_PREALLOC_FIFO_PAIRS] =
+    bpm->prealloc_fifos ? bpm->prealloc_fifos : 1;
+
+  a->options[APP_OPTIONS_FLAGS] = APP_OPTIONS_FLAGS_BUILTIN_APP;
+
+  a->segment_name = segment_name;
+  a->segment_name_length = ARRAY_LEN (segment_name);
+
+  if (vnet_application_attach (a))
+    {
+      clib_warning ("failed to attach server");
+      return -1;
+    }
+  bpm->server_app_index = a->app_index;
+
+  return 0;
+}
+
+static int
+active_open_attach (void)
+{
+  builtin_proxy_main_t *bpm = &builtin_proxy_main;
+  vnet_app_attach_args_t _a, *a = &_a;
+  u8 segment_name[128];
+  u32 segment_name_length;
+  u64 options[16];
+
+  segment_name_length = ARRAY_LEN (segment_name);
+
+  memset (a, 0, sizeof (*a));
+  memset (options, 0, sizeof (options));
+
+  a->api_client_index = bpm->active_open_client_index;
+  a->segment_name = segment_name;
+  a->segment_name_length = segment_name_length;
+  a->session_cb_vft = &builtin_clients;
+
+  options[SESSION_OPTIONS_ACCEPT_COOKIE] = 0x12345678;
+  options[SESSION_OPTIONS_SEGMENT_SIZE] = 512 << 20;
+  options[SESSION_OPTIONS_RX_FIFO_SIZE] = bpm->fifo_size;
+  options[SESSION_OPTIONS_TX_FIFO_SIZE] = bpm->fifo_size;
+  options[APP_OPTIONS_PRIVATE_SEGMENT_COUNT] = bpm->private_segment_count;
+  options[APP_OPTIONS_PRIVATE_SEGMENT_SIZE] = bpm->private_segment_size;
+  options[APP_OPTIONS_PREALLOC_FIFO_PAIRS] =
+    bpm->prealloc_fifos ? bpm->prealloc_fifos : 1;
+
+  options[APP_OPTIONS_FLAGS] = APP_OPTIONS_FLAGS_BUILTIN_APP
+    | APP_OPTIONS_FLAGS_IS_PROXY;
+
+  a->options = options;
+
+  if (vnet_application_attach (a))
+    return -1;
+
+  bpm->active_open_app_index = a->app_index;
+
+  return 0;
+}
+
+static int
+server_listen ()
+{
+  builtin_proxy_main_t *bpm = &builtin_proxy_main;
+  vnet_bind_args_t _a, *a = &_a;
+  memset (a, 0, sizeof (*a));
+  a->app_index = bpm->server_app_index;
+  a->uri = "tcp://0.0.0.0/23";
+  return vnet_bind_uri (a);
+}
+
+static int
+server_create (vlib_main_t * vm)
+{
+  builtin_proxy_main_t *bpm = &builtin_proxy_main;
+  vlib_thread_main_t *vtm = vlib_get_thread_main ();
+  u32 num_threads;
+  int i;
+
+  if (bpm->server_client_index == (u32) ~ 0)
+    create_api_loopbacks (vm);
+
+  num_threads = 1 /* main thread */  + vtm->n_threads;
+  vec_validate (builtin_proxy_main.server_event_queue, num_threads - 1);
+  vec_validate (builtin_proxy_main.active_open_event_queue, num_threads - 1);
+  vec_validate (bpm->rx_buf, num_threads - 1);
+
+  for (i = 0; i < num_threads; i++)
+    vec_validate (bpm->rx_buf[i], bpm->rcv_buffer_size);
+
+  if (server_attach ())
+    {
+      clib_warning ("failed to attach server app");
+      return -1;
+    }
+  if (server_listen ())
+    {
+      clib_warning ("failed to start listening");
+      return -1;
+    }
+  if (active_open_attach ())
+    {
+      clib_warning ("failed to attach active open app");
+      return -1;
+    }
+
+  for (i = 0; i < num_threads; i++)
+    {
+      bpm->active_open_event_queue[i] =
+	session_manager_get_vpp_event_queue (i);
+
+      ASSERT (bpm->active_open_event_queue[i]);
+
+      bpm->server_event_queue[i] = session_manager_get_vpp_event_queue (i);
+    }
+
+  return 0;
+}
+
+static clib_error_t *
+proxy_server_create_command_fn (vlib_main_t * vm, unformat_input_t * input,
+				vlib_cli_command_t * cmd)
+{
+  builtin_proxy_main_t *bpm = &builtin_proxy_main;
+  int rv;
+  u32 tmp;
+
+  bpm->fifo_size = 64 << 10;
+  bpm->rcv_buffer_size = 1024;
+  bpm->prealloc_fifos = 0;
+  bpm->private_segment_count = 0;
+  bpm->private_segment_size = 0;
+
+  while (unformat_check_input (input) != UNFORMAT_END_OF_INPUT)
+    {
+      if (unformat (input, "fifo-size %d", &bpm->fifo_size))
+	bpm->fifo_size <<= 10;
+      else if (unformat (input, "rcv-buf-size %d", &bpm->rcv_buffer_size))
+	;
+      else if (unformat (input, "prealloc-fifos %d", &bpm->prealloc_fifos))
+	;
+      else if (unformat (input, "private-segment-count %d",
+			 &bpm->private_segment_count))
+	;
+      else if (unformat (input, "private-segment-size %dm", &tmp))
+	bpm->private_segment_size = tmp << 20;
+      else if (unformat (input, "private-segment-size %dg", &tmp))
+	bpm->private_segment_size = tmp << 30;
+      else if (unformat (input, "private-segment-size %d", &tmp))
+	bpm->private_segment_size = tmp;
+      else
+	return clib_error_return (0, "unknown input `%U'",
+				  format_unformat_error, input);
+    }
+
+  vnet_session_enable_disable (vm, 1 /* turn on TCP, etc. */ );
+
+  rv = server_create (vm);
+  switch (rv)
+    {
+    case 0:
+      break;
+    default:
+      return clib_error_return (0, "server_create returned %d", rv);
+    }
+
+  return 0;
+}
+
+/* *INDENT-OFF* */
+VLIB_CLI_COMMAND (server_create_command, static) =
+{
+  .path = "test proxy server",
+  .short_help = "test proxy server",
+  .function = proxy_server_create_command_fn,
+};
+/* *INDENT-ON* */
+
+clib_error_t *
+builtin_tcp_proxy_main_init (vlib_main_t * vm)
+{
+  builtin_proxy_main_t *bpm = &builtin_proxy_main;
+  bpm->server_client_index = ~0;
+  bpm->active_open_client_index = ~0;
+  bpm->proxy_session_by_active_open_handle = hash_create (0, sizeof (uword));
+  bpm->proxy_session_by_server_handle = hash_create (0, sizeof (uword));
+
+  return 0;
+}
+
+VLIB_INIT_FUNCTION (builtin_tcp_proxy_main_init);
+
+/*
+* fd.io coding-style-patch-verification: ON
+*
+* Local Variables:
+* eval: (c-set-style "gnu")
+* End:
+*/
diff --git a/src/vnet/tcp/builtin_proxy.h b/src/vnet/tcp/builtin_proxy.h
new file mode 100644
index 00000000..cf707a15
--- /dev/null
+++ b/src/vnet/tcp/builtin_proxy.h
@@ -0,0 +1,100 @@
+
+/*
+ * builtin_proxy.h - skeleton vpp engine plug-in header file
+ *
+ * Copyright (c) <current-year> <your-organization>
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at:
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#ifndef __included_builtin_proxy_h__
+#define __included_builtin_proxy_h__
+
+#include <vnet/vnet.h>
+#include <vnet/ip/ip.h>
+#include <vnet/ethernet/ethernet.h>
+
+#include <vppinfra/hash.h>
+#include <vppinfra/error.h>
+#include <vlibmemory/unix_shared_memory_queue.h>
+#include <svm/svm_fifo_segment.h>
+#include <vnet/session/session.h>
+#include <vnet/session/application_interface.h>
+
+typedef struct
+{
+  svm_fifo_t *server_rx_fifo;
+  svm_fifo_t *server_tx_fifo;
+
+  u64 vpp_server_handle;
+  u64 vpp_active_open_handle;
+} proxy_session_t;
+
+typedef struct
+{
+  unix_shared_memory_queue_t *vl_input_queue;	/**< vpe input queue */
+  /** per-thread vectors */
+  unix_shared_memory_queue_t **server_event_queue;
+  unix_shared_memory_queue_t **active_open_event_queue;
+  u8 **rx_buf;				/**< intermediate rx buffers */
+
+  u32 cli_node_index;			/**< cli process node index */
+  u32 server_client_index;		/**< server API client handle */
+  u32 server_app_index;			/**< server app index */
+  u32 active_open_client_index;		/**< active open API client handle */
+  u32 active_open_app_index;		/**< active open index after attach */
+
+  uword *proxy_session_by_server_handle;
+  uword *proxy_session_by_active_open_handle;
+
+  /*
+   * Configuration params
+   */
+  u8 *connect_uri;			/**< URI for slave's connect */
+  u32 configured_segment_size;
+  u32 fifo_size;
+  u32 private_segment_count;		/**< Number of private fifo segs */
+  u32 private_segment_size;		/**< size of private fifo segs */
+  int rcv_buffer_size;
+
+  /*
+   * Test state variables
+   */
+  proxy_session_t *sessions;		/**< Session pool, shared */
+  clib_spinlock_t sessions_lock;
+  u32 **connection_index_by_thread;
+  pthread_t client_thread_handle;
+
+  /*
+   * Flags
+   */
+  u8 is_init;
+  u8 prealloc_fifos;		/**< Request fifo preallocation */
+
+  /*
+   * Convenience
+   */
+  vlib_main_t *vlib_main;
+  vnet_main_t *vnet_main;
+  ethernet_main_t *ethernet_main;
+} builtin_proxy_main_t;
+
+builtin_proxy_main_t builtin_proxy_main;
+
+#endif /* __included_builtin_proxy_h__ */
+
+/*
+ * fd.io coding-style-patch-verification: ON
+ *
+ * Local Variables:
+ * eval: (c-set-style "gnu")
+ * End:
+ */
diff --git a/src/vnet/tcp/builtin_server.c b/src/vnet/tcp/builtin_server.c
index 3416678e..2454a831 100644
--- a/src/vnet/tcp/builtin_server.c
+++ b/src/vnet/tcp/builtin_server.c
@@ -18,25 +18,6 @@
 #include <vnet/session/application.h>
 #include <vnet/session/application_interface.h>
 
-/* define message IDs */
-#include <vpp/api/vpe_msg_enum.h>
-
-/* define message structures */
-#define vl_typedefs
-#include <vpp/api/vpe_all_api_h.h>
-#undef vl_typedefs
-
-/* define generated endian-swappers */
-#define vl_endianfun
-#include <vpp/api/vpe_all_api_h.h>
-#undef vl_endianfun
-
-/* instantiate all the print functions we know about */
-#define vl_print(handle, ...) vlib_cli_output (handle, __VA_ARGS__)
-#define vl_printfun
-#include <vpp/api/vpe_all_api_h.h>
-#undef vl_printfun
-
 typedef struct
 {
   /*
@@ -279,46 +260,13 @@ static int
 create_api_loopback (vlib_main_t * vm)
 {
   builtin_server_main_t *bsm = &builtin_server_main;
-  vl_api_memclnt_create_t _m, *mp = &_m;
-  extern void vl_api_memclnt_create_t_handler (vl_api_memclnt_create_t *);
   api_main_t *am = &api_main;
   vl_shmem_hdr_t *shmem_hdr;
-  uword *event_data = 0, event_type;
-  int resolved = 0;
-
-  /*
-   * Create a "loopback" API client connection
-   * Don't do things like this unless you know what you're doing...
-   */
 
   shmem_hdr = am->shmem_hdr;
   bsm->vl_input_queue = shmem_hdr->vl_input_queue;
-  memset (mp, 0, sizeof (*mp));
-  mp->_vl_msg_id = VL_API_MEMCLNT_CREATE;
-  mp->context = 0xFEEDFACE;
-  mp->input_queue = pointer_to_uword (bsm->vl_input_queue);
-  strncpy ((char *) mp->name, "tcp_test_server", sizeof (mp->name) - 1);
-
-  vl_api_memclnt_create_t_handler (mp);
-
-  /* Wait for reply */
-  bsm->node_index = vlib_get_current_process (vm)->node_runtime.node_index;
-  vlib_process_wait_for_event_or_clock (vm, 2.0);
-  event_type = vlib_process_get_events (vm, &event_data);
-  switch (event_type)
-    {
-    case 1:
-      resolved = 1;
-      break;
-    case ~0:
-      /* timed out */
-      break;
-    default:
-      clib_warning ("unknown event_type %d", event_type);
-    }
-  if (!resolved)
-    return -1;
-
+  bsm->my_client_index =
+    vl_api_memclnt_create_internal ("tcp_test_server", bsm->vl_input_queue);
   return 0;
 }
 
@@ -413,45 +361,6 @@ server_create (vlib_main_t * vm)
   return 0;
 }
 
-/* Get our api client index */
-static void
-vl_api_memclnt_create_reply_t_handler (vl_api_memclnt_create_reply_t * mp)
-{
-  vlib_main_t *vm = vlib_get_main ();
-  builtin_server_main_t *bsm = &builtin_server_main;
-  bsm->my_client_index = mp->index;
-  vlib_process_signal_event (vm, bsm->node_index, 1 /* evt */ ,
-			     0 /* data */ );
-}
-
-#define foreach_tcp_builtin_server_api_msg      		\
-_(MEMCLNT_CREATE_REPLY, memclnt_create_reply)   		\
-
-static clib_error_t *
-tcp_builtin_server_api_hookup (vlib_main_t * vm)
-{
-  vl_msg_api_msg_config_t _c, *c = &_c;
-
-  /* Hook up client-side static APIs to our handlers */
-#define _(N,n) do {                                             \
-    c->id = VL_API_##N;                                         \
-    c->name = #n;                                               \
-    c->handler = vl_api_##n##_t_handler;                        \
-    c->cleanup = vl_noop_handler;                               \
-    c->endian = vl_api_##n##_t_endian;                          \
-    c->print = vl_api_##n##_t_print;                            \
-    c->size = sizeof(vl_api_##n##_t);                           \
-    c->traced = 1; /* trace, so these msgs print */             \
-    c->replay = 0; /* don't replay client create/delete msgs */ \
-    c->message_bounce = 0; /* don't bounce this message */	\
-    vl_msg_api_config(c);} while (0);
-
-  foreach_tcp_builtin_server_api_msg;
-#undef _
-
-  return 0;
-}
-
 static clib_error_t *
 server_create_command_fn (vlib_main_t * vm, unformat_input_t * input,
 			  vlib_cli_command_t * cmd)
@@ -491,7 +400,6 @@ server_create_command_fn (vlib_main_t * vm, unformat_input_t * input,
 				  format_unformat_error, input);
     }
 
-  tcp_builtin_server_api_hookup (vm);
   vnet_session_enable_disable (vm, 1 /* turn on TCP, etc. */ );
 
   rv = server_create (vm);
-- 
cgit 1.2.3-korg


From 818eb54de01459ed3d823f8a9781bbed0845db82 Mon Sep 17 00:00:00 2001
From: Dave Barach <dbarach@cisco.com>
Date: Wed, 2 Aug 2017 13:56:13 -0400
Subject: Improve the svm fifo allocator

- Round up requested fifo size to the next power of two
- Maintain per-segment power-of-two freelists
- Allocate fifos in chunks, to amortize alignment overhead
- Detach builtin test client application after each run
  so we can use different fifo sizes each time
- Be more suspicious of session / application indices

Useful prep work for dynamically resizing fifos. As far as the svm
fifo code is concerned, it's OK to set fifo->nitems anywhere in the
interval: [0, 1<<(fifo->freelist_index) + FIFO_SEGMENT_MIN_FIFO_SIZE]

It's unlikely that setting nitems below the path MTU will work out
very well...

Change-Id: Idad73a027dfb7412056cb02988b77e300fa7e8a7
Signed-off-by: Dave Barach <dave@barachs.net>
---
 src/svm/svm_fifo.c            |   5 +-
 src/svm/svm_fifo.h            |   3 +-
 src/svm/svm_fifo_segment.c    | 130 +++++++++++++++++++++++++++++++++++++-----
 src/svm/svm_fifo_segment.h    |   6 +-
 src/vnet/session/session.c    |  13 ++++-
 src/vnet/tcp/builtin_client.c |  14 +++++
 6 files changed, 151 insertions(+), 20 deletions(-)

(limited to 'src/svm/svm_fifo.c')

diff --git a/src/svm/svm_fifo.c b/src/svm/svm_fifo.c
index e478c06e..7f8127cf 100644
--- a/src/svm/svm_fifo.c
+++ b/src/svm/svm_fifo.c
@@ -192,8 +192,11 @@ svm_fifo_t *
 svm_fifo_create (u32 data_size_in_bytes)
 {
   svm_fifo_t *f;
+  u32 rounded_data_size;
 
-  f = clib_mem_alloc_aligned_or_null (sizeof (*f) + data_size_in_bytes,
+  /* always round fifo data size to the next highest power-of-two */
+  rounded_data_size = (1 << (max_log2 (data_size_in_bytes)));
+  f = clib_mem_alloc_aligned_or_null (sizeof (*f) + rounded_data_size,
 				      CLIB_CACHE_LINE_BYTES);
   if (f == 0)
     return 0;
diff --git a/src/svm/svm_fifo.h b/src/svm/svm_fifo.h
index f10b4d91..84901d02 100644
--- a/src/svm/svm_fifo.h
+++ b/src/svm/svm_fifo.h
@@ -75,7 +75,8 @@ typedef struct _svm_fifo
 #if SVM_FIFO_TRACE
   svm_fifo_trace_elem_t *trace;
 #endif
-  i8 refcnt;
+  u32 freelist_index;		/**< aka log2(allocated_size) - const. */
+  i8 refcnt;			/**< reference count  */
     CLIB_CACHE_LINE_ALIGN_MARK (data);
 } svm_fifo_t;
 
diff --git a/src/svm/svm_fifo_segment.c b/src/svm/svm_fifo_segment.c
index c04b9d8c..2094ba74 100644
--- a/src/svm/svm_fifo_segment.c
+++ b/src/svm/svm_fifo_segment.c
@@ -17,13 +17,54 @@
 
 svm_fifo_segment_main_t svm_fifo_segment_main;
 
+static void
+allocate_new_fifo_chunk (svm_fifo_segment_header_t * fsh,
+			 u32 data_size_in_bytes, int chunk_size)
+{
+  int freelist_index;
+  u32 size;
+  u8 *fifo_space;
+  u32 rounded_data_size;
+  svm_fifo_t *f;
+  int i;
+
+  rounded_data_size = (1 << (max_log2 (data_size_in_bytes)));
+  freelist_index = max_log2 (rounded_data_size)
+    - max_log2 (FIFO_SEGMENT_MIN_FIFO_SIZE);
+
+  /* Calculate space requirement $$$ round-up data_size_in_bytes */
+  size = (sizeof (*f) + rounded_data_size) * chunk_size;
+
+  /* Allocate fifo space. May fail. */
+  fifo_space = clib_mem_alloc_aligned_at_offset
+    (size, CLIB_CACHE_LINE_BYTES, 0 /* align_offset */ ,
+     0 /* os_out_of_memory */ );
+
+  /* Out of space.. */
+  if (fifo_space == 0)
+    return;
+
+  /* Carve fifo space */
+  f = (svm_fifo_t *) fifo_space;
+  for (i = 0; i < chunk_size; i++)
+    {
+      f->freelist_index = freelist_index;
+      f->next = fsh->free_fifos[freelist_index];
+      fsh->free_fifos[freelist_index] = f;
+      fifo_space += sizeof (*f) + rounded_data_size;
+      f = (svm_fifo_t *) fifo_space;
+    }
+}
+
 static void
 preallocate_fifo_pairs (svm_fifo_segment_header_t * fsh,
 			svm_fifo_segment_create_args_t * a)
 {
   u32 rx_fifo_size, tx_fifo_size;
+  u32 rx_rounded_data_size, tx_rounded_data_size;
   svm_fifo_t *f;
   u8 *rx_fifo_space, *tx_fifo_space;
+  int rx_freelist_index, tx_freelist_index;
   int i;
 
   /* Parameter check */
@@ -31,10 +72,39 @@ preallocate_fifo_pairs (svm_fifo_segment_header_t * fsh,
       || a->preallocated_fifo_pairs == 0)
     return;
 
-  /* Calculate space requirements */
-  rx_fifo_size = (sizeof (*f) + a->rx_fifo_size) * a->preallocated_fifo_pairs;
-  tx_fifo_size = (sizeof (*f) + a->tx_fifo_size) * a->preallocated_fifo_pairs;
+  if (a->rx_fifo_size < FIFO_SEGMENT_MIN_FIFO_SIZE ||
+      a->rx_fifo_size > FIFO_SEGMENT_MAX_FIFO_SIZE)
+    {
+      clib_warning ("rx fifo_size out of range %d", a->rx_fifo_size);
+      return;
+    }
+
+  if (a->tx_fifo_size < FIFO_SEGMENT_MIN_FIFO_SIZE ||
+      a->tx_fifo_size > FIFO_SEGMENT_MAX_FIFO_SIZE)
+    {
+      clib_warning ("tx fifo_size out of range %d", a->rx_fifo_size);
+      return;
+    }
+
+  rx_rounded_data_size = (1 << (max_log2 (a->rx_fifo_size)));
+
+  rx_freelist_index = max_log2 (a->rx_fifo_size)
+    - max_log2 (FIFO_SEGMENT_MIN_FIFO_SIZE);
+
+  tx_rounded_data_size = (1 << (max_log2 (a->rx_fifo_size)));
+
+  tx_freelist_index = max_log2 (a->tx_fifo_size)
+    - max_log2 (FIFO_SEGMENT_MIN_FIFO_SIZE);
 
+  /* Calculate space requirements */
+  rx_fifo_size = (sizeof (*f) + rx_rounded_data_size)
+    * a->preallocated_fifo_pairs;
+  tx_fifo_size = (sizeof (*f) + tx_rounded_data_size)
+    * a->preallocated_fifo_pairs;
+
+  vec_validate_init_empty (fsh->free_fifos,
+			   clib_max (rx_freelist_index, tx_freelist_index),
+			   0);
   if (0)
     clib_warning ("rx_fifo_size %u (%d mb), tx_fifo_size %u (%d mb)",
 		  rx_fifo_size, rx_fifo_size >> 20,
@@ -71,18 +141,20 @@ preallocate_fifo_pairs (svm_fifo_segment_header_t * fsh,
   f = (svm_fifo_t *) rx_fifo_space;
   for (i = 0; i < a->preallocated_fifo_pairs; i++)
     {
-      f->next = fsh->free_fifos[FIFO_SEGMENT_RX_FREELIST];
-      fsh->free_fifos[FIFO_SEGMENT_RX_FREELIST] = f;
-      rx_fifo_space += sizeof (*f) + a->rx_fifo_size;
+      f->freelist_index = rx_freelist_index;
+      f->next = fsh->free_fifos[rx_freelist_index];
+      fsh->free_fifos[rx_freelist_index] = f;
+      rx_fifo_space += sizeof (*f) + rx_rounded_data_size;
       f = (svm_fifo_t *) rx_fifo_space;
     }
   /* Carve tx fifo space */
   f = (svm_fifo_t *) tx_fifo_space;
   for (i = 0; i < a->preallocated_fifo_pairs; i++)
     {
-      f->next = fsh->free_fifos[FIFO_SEGMENT_TX_FREELIST];
-      fsh->free_fifos[FIFO_SEGMENT_TX_FREELIST] = f;
-      tx_fifo_space += sizeof (*f) + a->tx_fifo_size;
+      f->freelist_index = tx_freelist_index;
+      f->next = fsh->free_fifos[tx_freelist_index];
+      fsh->free_fifos[tx_freelist_index] = f;
+      tx_fifo_space += sizeof (*f) + tx_rounded_data_size;
       f = (svm_fifo_t *) tx_fifo_space;
     }
 }
@@ -277,6 +349,21 @@ svm_fifo_segment_alloc_fifo (svm_fifo_segment_private_t * s,
   svm_fifo_segment_header_t *fsh;
   svm_fifo_t *f;
   void *oldheap;
+  int freelist_index;
+
+  /*
+   * 2K minimum. It's not likely that anything good will happen
+   * with a 1K FIFO.
+   */
+  if (data_size_in_bytes < FIFO_SEGMENT_MIN_FIFO_SIZE ||
+      data_size_in_bytes > FIFO_SEGMENT_MAX_FIFO_SIZE)
+    {
+      clib_warning ("fifo size out of range %d", data_size_in_bytes);
+      return 0;
+    }
+
+  freelist_index = max_log2 (data_size_in_bytes)
+    - max_log2 (FIFO_SEGMENT_MIN_FIFO_SIZE);
 
   sh = s->ssvm.sh;
   fsh = (svm_fifo_segment_header_t *) sh->opaque[0];
@@ -288,15 +375,24 @@ svm_fifo_segment_alloc_fifo (svm_fifo_segment_private_t * s,
     {
     case FIFO_SEGMENT_RX_FREELIST:
     case FIFO_SEGMENT_TX_FREELIST:
-      f = fsh->free_fifos[list_index];
-      if (f)
+      vec_validate_init_empty (fsh->free_fifos, freelist_index, 0);
+
+      f = fsh->free_fifos[freelist_index];
+      if (PREDICT_FALSE (f == 0))
 	{
-	  fsh->free_fifos[list_index] = f->next;
+	  allocate_new_fifo_chunk (fsh, data_size_in_bytes,
+				   FIFO_SEGMENT_ALLOC_CHUNK_SIZE);
+	  f = fsh->free_fifos[freelist_index];
+	}
+      if (PREDICT_TRUE (f != 0))
+	{
+	  fsh->free_fifos[freelist_index] = f->next;
 	  /* (re)initialize the fifo, as in svm_fifo_create */
 	  memset (f, 0, sizeof (*f));
 	  f->nitems = data_size_in_bytes;
 	  f->ooos_list_head = OOO_SEGMENT_INVALID_INDEX;
 	  f->refcnt = 1;
+	  f->freelist_index = freelist_index;
 	  goto found;
 	}
       /* FALLTHROUGH */
@@ -316,6 +412,7 @@ svm_fifo_segment_alloc_fifo (svm_fifo_segment_private_t * s,
       ssvm_unlock_non_recursive (sh);
       return (0);
     }
+  f->freelist_index = freelist_index;
 
 found:
   /* If rx_freelist add to active fifos list. When cleaning up segment,
@@ -344,6 +441,7 @@ svm_fifo_segment_free_fifo (svm_fifo_segment_private_t * s, svm_fifo_t * f,
   ssvm_shared_header_t *sh;
   svm_fifo_segment_header_t *fsh;
   void *oldheap;
+  int freelist_index;
 
   ASSERT (f->refcnt > 0);
 
@@ -353,6 +451,10 @@ svm_fifo_segment_free_fifo (svm_fifo_segment_private_t * s, svm_fifo_t * f,
   sh = s->ssvm.sh;
   fsh = (svm_fifo_segment_header_t *) sh->opaque[0];
 
+  freelist_index = f->freelist_index;
+
+  ASSERT (freelist_index > 0 && freelist_index < vec_len (fsh->free_fifos));
+
   ssvm_lock_non_recursive (sh, 2);
   oldheap = ssvm_push_heap (sh);
 
@@ -369,9 +471,9 @@ svm_fifo_segment_free_fifo (svm_fifo_segment_private_t * s, svm_fifo_t * f,
       /* Fall through: we add only rx fifos to active pool */
     case FIFO_SEGMENT_TX_FREELIST:
       /* Add to free list */
-      f->next = fsh->free_fifos[list_index];
+      f->next = fsh->free_fifos[freelist_index];
       f->prev = 0;
-      fsh->free_fifos[list_index] = f;
+      fsh->free_fifos[freelist_index] = f;
       break;
     case FIFO_SEGMENT_FREELIST_NONE:
       break;
diff --git a/src/svm/svm_fifo_segment.h b/src/svm/svm_fifo_segment.h
index a7a3f469..68bb4d3d 100644
--- a/src/svm/svm_fifo_segment.h
+++ b/src/svm/svm_fifo_segment.h
@@ -27,11 +27,15 @@ typedef enum
   FIFO_SEGMENT_N_FREELISTS
 } svm_fifo_segment_freelist_t;
 
+#define FIFO_SEGMENT_MIN_FIFO_SIZE 2048
+#define FIFO_SEGMENT_MAX_FIFO_SIZE (8<<20)	/* 8mb max fifo size */
+#define FIFO_SEGMENT_ALLOC_CHUNK_SIZE 32	/* allocation quantum */
+
 typedef struct
 {
   svm_fifo_t *fifos;		/**< Linked list of active RX fifos */
   u8 *segment_name;		/**< Segment name */
-  svm_fifo_t *free_fifos[FIFO_SEGMENT_N_FREELISTS];	/**< Free lists */
+  svm_fifo_t **free_fifos;	/**< Freelists, by fifo size  */
 } svm_fifo_segment_header_t;
 
 typedef struct
diff --git a/src/vnet/session/session.c b/src/vnet/session/session.c
index 991bcd5a..533a6c22 100644
--- a/src/vnet/session/session.c
+++ b/src/vnet/session/session.c
@@ -266,7 +266,13 @@ stream_session_enqueue_notify (stream_session_t * s, u8 block)
     return 0;
 
   /* Get session's server */
-  app = application_get (s->app_index);
+  app = application_get_if_valid (s->app_index);
+
+  if (PREDICT_FALSE (app == 0))
+    {
+      clib_warning ("invalid s->app_index = %d", s->app_index);
+      return 0;
+    }
 
   /* Built-in server? Hand event to the callback... */
   if (app->cb_fns.builtin_server_rx_callback)
@@ -327,8 +333,9 @@ session_manager_flush_enqueue_events (u32 thread_index)
       stream_session_t *s0;
 
       /* Get session */
-      s0 = stream_session_get (session_indices_to_enqueue[i], thread_index);
-      if (stream_session_enqueue_notify (s0, 0 /* don't block */ ))
+      s0 = stream_session_get_if_valid (session_indices_to_enqueue[i],
+					thread_index);
+      if (s0 == 0 || stream_session_enqueue_notify (s0, 0 /* don't block */ ))
 	{
 	  errors++;
 	}
diff --git a/src/vnet/tcp/builtin_client.c b/src/vnet/tcp/builtin_client.c
index 938e07ba..5fa54460 100644
--- a/src/vnet/tcp/builtin_client.c
+++ b/src/vnet/tcp/builtin_client.c
@@ -713,6 +713,20 @@ cleanup:
 
   pool_free (tm->sessions);
 
+  /* Detach the application, so we can use different fifo sizes next time */
+  if (tm->test_client_attached)
+    {
+      vnet_app_detach_args_t _da, *da = &_da;
+      int rv;
+
+      da->app_index = tm->app_index;
+
+      rv = vnet_application_detach (da);
+      if (rv)
+	vlib_cli_output (vm, "WARNING: app detach failed...");
+      tm->test_client_attached = 0;
+      tm->app_index = ~0;
+    }
   return 0;
 }
 
-- 
cgit 1.2.3-korg


From 1f152cd6faf96b524b6b7071b5cffe1916f9c5cc Mon Sep 17 00:00:00 2001
From: Florin Coras <fcoras@cisco.com>
Date: Fri, 18 Aug 2017 19:28:03 -0700
Subject: tcp: retransmit and multi-buffer segment fixes and improvements

- set session state as closed on session manager delete
- enable retransmit as opposed to persist timer after persist timer completes
- properly discard buffer chain bytes when new data overlaps ooo
  segments
- don't use rxt bytes in snd space estimate used on tx path

Change-Id: Id9cab686e532e5fe70c775d5440260e8eb890a9f
Signed-off-by: Florin Coras <fcoras@cisco.com>
---
 src/svm/svm_fifo.c                 |  14 +-
 src/vnet/session/segment_manager.c |   1 +
 src/vnet/session/session.c         | 105 ++++++++++++---
 src/vnet/session/session_node.c    |  26 ++--
 src/vnet/session/stream_session.h  |   2 +-
 src/vnet/tcp/tcp.c                 |  39 +++---
 src/vnet/tcp/tcp.h                 |  23 +++-
 src/vnet/tcp/tcp_input.c           |  27 ++--
 src/vnet/tcp/tcp_output.c          | 264 +++++++++++++++++++++----------------
 9 files changed, 319 insertions(+), 182 deletions(-)

(limited to 'src/svm/svm_fifo.c')

diff --git a/src/svm/svm_fifo.c b/src/svm/svm_fifo.c
index 7f8127cf..8fe82f56 100644
--- a/src/svm/svm_fifo.c
+++ b/src/svm/svm_fifo.c
@@ -53,10 +53,12 @@ ooo_segment_end_pos (svm_fifo_t * f, ooo_segment_t * s)
 u8 *
 format_ooo_segment (u8 * s, va_list * args)
 {
+  svm_fifo_t *f = va_arg (*args, svm_fifo_t *);
   ooo_segment_t *seg = va_arg (*args, ooo_segment_t *);
-
-  s = format (s, "pos %u, len %u, next %d, prev %d",
-	      seg->start, seg->length, seg->next, seg->prev);
+  u32 normalized_start = (seg->start + f->nitems - f->tail) % f->nitems;
+  s = format (s, "[%u, %u], len %u, next %d, prev %d", normalized_start,
+	      (normalized_start + seg->length) % f->nitems, seg->length,
+	      seg->next, seg->prev);
   return s;
 }
 
@@ -154,7 +156,7 @@ format_ooo_list (u8 * s, va_list * args)
   while (ooo_segment_index != OOO_SEGMENT_INVALID_INDEX)
     {
       seg = pool_elt_at_index (f->ooo_segments, ooo_segment_index);
-      s = format (s, "  %U\n", format_ooo_segment, seg);
+      s = format (s, "  %U\n", format_ooo_segment, f, seg);
       ooo_segment_index = seg->next;
     }
 
@@ -557,7 +559,6 @@ svm_fifo_enqueue_with_offset_internal (svm_fifo_t * f,
 {
   u32 total_copy_bytes, first_copy_bytes, second_copy_bytes;
   u32 cursize, nitems, normalized_offset;
-  u32 offset_from_tail;
 
   f->ooos_newest = OOO_SEGMENT_INVALID_INDEX;
 
@@ -570,8 +571,7 @@ svm_fifo_enqueue_with_offset_internal (svm_fifo_t * f,
   normalized_offset = (f->tail + offset) % nitems;
 
   /* Will this request fit? */
-  offset_from_tail = (nitems + normalized_offset - f->tail) % nitems;
-  if ((required_bytes + offset_from_tail) > (nitems - cursize))
+  if ((required_bytes + offset) > (nitems - cursize))
     return -1;
 
   svm_fifo_trace_add (f, offset, required_bytes, 1);
diff --git a/src/vnet/session/segment_manager.c b/src/vnet/session/segment_manager.c
index 262b7faa..43977063 100644
--- a/src/vnet/session/segment_manager.c
+++ b/src/vnet/session/segment_manager.c
@@ -224,6 +224,7 @@ segment_manager_del (segment_manager_t * sm)
 	  session = stream_session_get (session_index, thread_index);
 
 	  /* Instead of directly removing the session call disconnect */
+	  session->session_state = SESSION_STATE_CLOSED;
 	  session_send_session_evt_to_thread (stream_session_handle (session),
 					      FIFO_EVENT_DISCONNECT,
 					      thread_index);
diff --git a/src/vnet/session/session.c b/src/vnet/session/session.c
index 70a5cd83..6fe99047 100644
--- a/src/vnet/session/session.c
+++ b/src/vnet/session/session.c
@@ -92,38 +92,104 @@ stream_session_create_i (segment_manager_t * sm, transport_connection_t * tc,
   return 0;
 }
 
-/** Enqueue buffer chain tail */
+/**
+ * Discards bytes from buffer chain
+ *
+ * It discards n_bytes_to_drop starting at first buffer after chain_b
+ */
+always_inline void
+session_enqueue_discard_chain_bytes (vlib_main_t * vm, vlib_buffer_t * b,
+				     vlib_buffer_t ** chain_b,
+				     u32 n_bytes_to_drop)
+{
+  vlib_buffer_t *next = *chain_b;
+  u32 to_drop = n_bytes_to_drop;
+  ASSERT (b->flags & VLIB_BUFFER_NEXT_PRESENT);
+  while (to_drop && (next->flags & VLIB_BUFFER_NEXT_PRESENT))
+    {
+      next = vlib_get_buffer (vm, next->next_buffer);
+      if (next->current_length > to_drop)
+	{
+	  vlib_buffer_advance (next, to_drop);
+	  to_drop = 0;
+	}
+      else
+	{
+	  to_drop -= next->current_length;
+	  next->current_length = 0;
+	}
+    }
+  *chain_b = next;
+
+  if (to_drop == 0)
+    b->total_length_not_including_first_buffer -= n_bytes_to_drop;
+}
+
+/**
+ * Enqueue buffer chain tail
+ */
 always_inline int
 session_enqueue_chain_tail (stream_session_t * s, vlib_buffer_t * b,
 			    u32 offset, u8 is_in_order)
 {
   vlib_buffer_t *chain_b;
-  u32 chain_bi = b->next_buffer, len;
+  u32 chain_bi, len, diff;
   vlib_main_t *vm = vlib_get_main ();
   u8 *data;
-  u16 written = 0;
+  u32 written = 0;
   int rv = 0;
 
+  if (is_in_order && offset)
+    {
+      diff = offset - b->current_length;
+      if (diff > b->total_length_not_including_first_buffer)
+	return 0;
+      chain_b = b;
+      session_enqueue_discard_chain_bytes (vm, b, &chain_b, diff);
+      chain_bi = vlib_get_buffer_index (vm, chain_b);
+    }
+  else
+    chain_bi = b->next_buffer;
+
   do
     {
       chain_b = vlib_get_buffer (vm, chain_bi);
       data = vlib_buffer_get_current (chain_b);
       len = chain_b->current_length;
+      if (!len)
+	continue;
       if (is_in_order)
 	{
 	  rv = svm_fifo_enqueue_nowait (s->server_rx_fifo, len, data);
-	  if (rv < len)
+	  if (rv == len)
+	    {
+	      written += rv;
+	    }
+	  else if (rv < len)
 	    {
 	      return (rv > 0) ? (written + rv) : written;
 	    }
-	  written += rv;
+	  else if (rv > len)
+	    {
+	      written += rv;
+
+	      /* written more than what was left in chain */
+	      if (written > b->total_length_not_including_first_buffer)
+		return written;
+
+	      /* drop the bytes that have already been delivered */
+	      session_enqueue_discard_chain_bytes (vm, b, &chain_b, rv - len);
+	    }
 	}
       else
 	{
 	  rv = svm_fifo_enqueue_with_offset (s->server_rx_fifo, offset, len,
 					     data);
 	  if (rv)
-	    return -1;
+	    {
+	      clib_warning ("failed to enqueue multi-buffer seg");
+	      return -1;
+	    }
 	  offset += len;
 	}
     }
@@ -155,22 +221,22 @@ stream_session_enqueue_data (transport_connection_t * tc, vlib_buffer_t * b,
 			     u32 offset, u8 queue_event, u8 is_in_order)
 {
   stream_session_t *s;
-  int enqueued = 0, rv;
+  int enqueued = 0, rv, in_order_off;
 
   s = stream_session_get (tc->s_index, tc->thread_index);
 
   if (is_in_order)
     {
-      enqueued =
-	svm_fifo_enqueue_nowait (s->server_rx_fifo, b->current_length,
-				 vlib_buffer_get_current (b));
-      if (PREDICT_FALSE
-	  ((b->flags & VLIB_BUFFER_NEXT_PRESENT) && enqueued > 0))
+      enqueued = svm_fifo_enqueue_nowait (s->server_rx_fifo,
+					  b->current_length,
+					  vlib_buffer_get_current (b));
+      if (PREDICT_FALSE ((b->flags & VLIB_BUFFER_NEXT_PRESENT)
+			 && enqueued >= 0))
 	{
-	  rv = session_enqueue_chain_tail (s, b, 0, 1);
-	  if (rv <= 0)
-	    return enqueued;
-	  enqueued += rv;
+	  in_order_off = enqueued > b->current_length ? enqueued : 0;
+	  rv = session_enqueue_chain_tail (s, b, in_order_off, 1);
+	  if (rv > 0)
+	    enqueued += rv;
 	}
     }
   else
@@ -179,9 +245,10 @@ stream_session_enqueue_data (transport_connection_t * tc, vlib_buffer_t * b,
 					 b->current_length,
 					 vlib_buffer_get_current (b));
       if (PREDICT_FALSE ((b->flags & VLIB_BUFFER_NEXT_PRESENT) && !rv))
-	rv = session_enqueue_chain_tail (s, b, offset + b->current_length, 0);
-      if (rv)
-	return -1;
+	session_enqueue_chain_tail (s, b, offset + b->current_length, 0);
+      /* if something was enqueued, report even this as success for ooo
+       * segment handling */
+      return rv;
     }
 
   if (queue_event)
diff --git a/src/vnet/session/session_node.c b/src/vnet/session/session_node.c
index fac2b852..cd52742b 100644
--- a/src/vnet/session/session_node.c
+++ b/src/vnet/session/session_node.c
@@ -76,7 +76,7 @@ session_tx_fifo_chain_tail (session_manager_main_t * smm, vlib_main_t * vm,
 			    u8 thread_index, svm_fifo_t * fifo,
 			    vlib_buffer_t * b0, u32 bi0, u8 n_bufs_per_seg,
 			    u32 left_from_seg, u32 * left_to_snd0,
-			    u16 * n_bufs, u32 * rx_offset, u16 deq_per_buf,
+			    u16 * n_bufs, u32 * tx_offset, u16 deq_per_buf,
 			    u8 peek_data)
 {
   vlib_buffer_t *chain_b0, *prev_b0;
@@ -104,8 +104,8 @@ session_tx_fifo_chain_tail (session_manager_main_t * smm, vlib_main_t * vm,
       data0 = vlib_buffer_get_current (chain_b0);
       if (peek_data)
 	{
-	  n_bytes_read = svm_fifo_peek (fifo, *rx_offset, len_to_deq0, data0);
-	  *rx_offset += n_bytes_read;
+	  n_bytes_read = svm_fifo_peek (fifo, *tx_offset, len_to_deq0, data0);
+	  *tx_offset += n_bytes_read;
 	}
       else
 	{
@@ -126,7 +126,8 @@ session_tx_fifo_chain_tail (session_manager_main_t * smm, vlib_main_t * vm,
       if (to_deq == 0)
 	break;
     }
-  ASSERT (to_deq == 0);
+  ASSERT (to_deq == 0
+	  && b0->total_length_not_including_first_buffer == left_from_seg);
   *left_to_snd0 -= left_from_seg;
 }
 
@@ -144,7 +145,7 @@ session_tx_fifo_read_and_snd_i (vlib_main_t * vm, vlib_node_runtime_t * node,
   transport_proto_vft_t *transport_vft;
   u32 next_index, next0, *to_next, n_left_to_next, bi0;
   vlib_buffer_t *b0;
-  u32 rx_offset = 0, max_dequeue0, n_bytes_per_seg;
+  u32 tx_offset = 0, max_dequeue0, n_bytes_per_seg, left_for_seg;
   u16 snd_mss0, n_bufs_per_seg, n_bufs;
   u8 *data0;
   int i, n_bytes_read;
@@ -170,11 +171,11 @@ session_tx_fifo_read_and_snd_i (vlib_main_t * vm, vlib_node_runtime_t * node,
   if (peek_data)
     {
       /* Offset in rx fifo from where to peek data  */
-      rx_offset = transport_vft->tx_fifo_offset (tc0);
+      tx_offset = transport_vft->tx_fifo_offset (tc0);
     }
 
   /* Check how much we can pull. If buffering, subtract the offset */
-  max_dequeue0 = svm_fifo_max_dequeue (s0->server_tx_fifo) - rx_offset;
+  max_dequeue0 = svm_fifo_max_dequeue (s0->server_tx_fifo) - tx_offset;
 
   /* Nothing to read return */
   if (max_dequeue0 == 0)
@@ -193,6 +194,7 @@ session_tx_fifo_read_and_snd_i (vlib_main_t * vm, vlib_node_runtime_t * node,
     }
   else
     {
+      /* Expectation is that snd_space0 is already a multiple of snd_mss */
       max_len_to_snd0 = snd_space0;
     }
 
@@ -265,8 +267,7 @@ session_tx_fifo_read_and_snd_i (vlib_main_t * vm, vlib_node_runtime_t * node,
 
 	  b0 = vlib_get_buffer (vm, bi0);
 	  b0->error = 0;
-	  b0->flags = VLIB_BUFFER_TOTAL_LENGTH_VALID
-	    | VNET_BUFFER_F_LOCALLY_ORIGINATED;
+	  b0->flags = VNET_BUFFER_F_LOCALLY_ORIGINATED;
 	  b0->current_data = 0;
 	  b0->total_length_not_including_first_buffer = 0;
 
@@ -274,11 +275,11 @@ session_tx_fifo_read_and_snd_i (vlib_main_t * vm, vlib_node_runtime_t * node,
 	  data0 = vlib_buffer_make_headroom (b0, MAX_HDRS_LEN);
 	  if (peek_data)
 	    {
-	      n_bytes_read = svm_fifo_peek (s0->server_tx_fifo, rx_offset,
+	      n_bytes_read = svm_fifo_peek (s0->server_tx_fifo, tx_offset,
 					    len_to_deq0, data0);
 	      /* Keep track of progress locally, transport is also supposed to
 	       * increment it independently when pushing the header */
-	      rx_offset += n_bytes_read;
+	      tx_offset += n_bytes_read;
 	    }
 	  else
 	    {
@@ -299,12 +300,11 @@ session_tx_fifo_read_and_snd_i (vlib_main_t * vm, vlib_node_runtime_t * node,
 	   */
 	  if (PREDICT_FALSE (n_bufs_per_seg > 1 && left_to_snd0))
 	    {
-	      u32 left_for_seg;
 	      left_for_seg = clib_min (snd_mss0 - n_bytes_read, left_to_snd0);
 	      session_tx_fifo_chain_tail (smm, vm, thread_index,
 					  s0->server_tx_fifo, b0, bi0,
 					  n_bufs_per_seg, left_for_seg,
-					  &left_to_snd0, &n_bufs, &rx_offset,
+					  &left_to_snd0, &n_bufs, &tx_offset,
 					  deq_per_buf, peek_data);
 	    }
 
diff --git a/src/vnet/session/stream_session.h b/src/vnet/session/stream_session.h
index 533cf97f..275052d3 100644
--- a/src/vnet/session/stream_session.h
+++ b/src/vnet/session/stream_session.h
@@ -56,7 +56,7 @@ typedef struct _stream_session_t
   u8 session_type;
 
   /** State */
-  u8 session_state;
+  volatile u8 session_state;
 
   u8 thread_index;
 
diff --git a/src/vnet/tcp/tcp.c b/src/vnet/tcp/tcp.c
index 6edf52af..197fff96 100644
--- a/src/vnet/tcp/tcp.c
+++ b/src/vnet/tcp/tcp.c
@@ -798,7 +798,8 @@ format_tcp_vars (u8 * s, va_list * args)
   s = format (s, "rtt_seq %u\n", tc->rtt_seq);
   s = format (s, " tsval_recent %u tsval_recent_age %u\n", tc->tsval_recent,
 	      tcp_time_now () - tc->tsval_recent_age);
-  s = format (s, " scoreboard: %U\n", format_tcp_scoreboard, &tc->sack_sb);
+  s = format (s, " scoreboard: %U\n", format_tcp_scoreboard, &tc->sack_sb,
+	      tc);
   if (vec_len (tc->snd_sacks))
     s = format (s, " sacks tx: %U\n", format_tcp_sacks, tc);
 
@@ -858,7 +859,7 @@ format_tcp_session (u8 * s, va_list * args)
   if (tc)
     s = format (s, "%U", format_tcp_connection, tc, verbose);
   else
-    s = format (s, "empty");
+    s = format (s, "empty\n");
   return s;
 }
 
@@ -930,7 +931,11 @@ u8 *
 format_tcp_sack_hole (u8 * s, va_list * args)
 {
   sack_scoreboard_hole_t *hole = va_arg (*args, sack_scoreboard_hole_t *);
-  s = format (s, "[%u, %u]", hole->start, hole->end);
+  tcp_connection_t *tc = va_arg (*args, tcp_connection_t *);
+  if (tc)
+    s = format (s, "  [%u, %u]", hole->start - tc->iss, hole->end - tc->iss);
+  else
+    s = format (s, "  [%u, %u]", hole->start, hole->end);
   return s;
 }
 
@@ -938,6 +943,7 @@ u8 *
 format_tcp_scoreboard (u8 * s, va_list * args)
 {
   sack_scoreboard_t *sb = va_arg (*args, sack_scoreboard_t *);
+  tcp_connection_t *tc = va_arg (*args, tcp_connection_t *);
   sack_scoreboard_hole_t *hole;
   s = format (s, "sacked_bytes %u last_sacked_bytes %u lost_bytes %u\n",
 	      sb->sacked_bytes, sb->last_sacked_bytes, sb->lost_bytes);
@@ -952,7 +958,7 @@ format_tcp_scoreboard (u8 * s, va_list * args)
 
   while (hole)
     {
-      s = format (s, "%U", format_tcp_sack_hole, hole);
+      s = format (s, "%U", format_tcp_sack_hole, hole, tc);
       hole = scoreboard_next_hole (sb, hole);
     }
 
@@ -1001,13 +1007,10 @@ tcp_round_snd_space (tcp_connection_t * tc, u32 snd_space)
       return tc->snd_wnd <= snd_space ? tc->snd_wnd : 0;
     }
 
-  /* If we can't write at least a segment, don't try at all */
+  /* If not snd_wnd constrained and we can't write at least a segment,
+   * don't try at all */
   if (PREDICT_FALSE (snd_space < tc->snd_mss))
-    {
-      if (snd_space > clib_min (tc->mss, tc->rcv_opts.mss) - TCP_HDR_LEN_MAX)
-	return snd_space;
-      return 0;
-    }
+    return 0;
 
   /* round down to mss multiple */
   return snd_space - (snd_space % tc->snd_mss);
@@ -1030,7 +1033,7 @@ tcp_snd_space (tcp_connection_t * tc)
 
   if (PREDICT_TRUE (tcp_in_cong_recovery (tc) == 0))
     {
-      snd_space = tcp_available_snd_space (tc);
+      snd_space = tcp_available_output_snd_space (tc);
 
       /* If we haven't gotten dupacks or if we did and have gotten sacked
        * bytes then we can still send as per Limited Transmit (RFC3042) */
@@ -1051,17 +1054,20 @@ tcp_snd_space (tcp_connection_t * tc)
   if (tcp_in_recovery (tc))
     {
       tc->snd_nxt = tc->snd_una_max;
-      snd_space = tcp_available_wnd (tc) - tc->snd_rxt_bytes
+      snd_space = tcp_available_snd_wnd (tc) - tc->snd_rxt_bytes
 	- (tc->snd_una_max - tc->snd_congestion);
       if (snd_space <= 0 || (tc->snd_una_max - tc->snd_una) >= tc->snd_wnd)
 	return 0;
       return tcp_round_snd_space (tc, snd_space);
     }
 
-  /* If in fast recovery, send 1 SMSS if wnd allows */
-  if (tcp_in_fastrecovery (tc)
-      && tcp_available_snd_space (tc) && !tcp_fastrecovery_sent_1_smss (tc))
+  /* RFC 5681: When previously unsent data is available and the new value of
+   * cwnd and the receiver's advertised window allow, a TCP SHOULD send 1*SMSS
+   * bytes of previously unsent data. */
+  if (tcp_in_fastrecovery (tc) && !tcp_fastrecovery_sent_1_smss (tc))
     {
+      if (tcp_available_output_snd_space (tc) < tc->snd_mss)
+	return 0;
       tcp_fastrecovery_1_smss_on (tc);
       return tc->snd_mss;
     }
@@ -1073,7 +1079,8 @@ u32
 tcp_session_send_space (transport_connection_t * trans_conn)
 {
   tcp_connection_t *tc = (tcp_connection_t *) trans_conn;
-  return tcp_snd_space (tc);
+  return clib_min (tcp_snd_space (tc),
+		   tc->snd_wnd - (tc->snd_nxt - tc->snd_una));
 }
 
 i32
diff --git a/src/vnet/tcp/tcp.h b/src/vnet/tcp/tcp.h
index 097cc8cf..9e4660b8 100644
--- a/src/vnet/tcp/tcp.h
+++ b/src/vnet/tcp/tcp.h
@@ -498,7 +498,9 @@ tcp_half_open_connection_get (u32 conn_index)
 void tcp_make_ack (tcp_connection_t * ts, vlib_buffer_t * b);
 void tcp_make_fin (tcp_connection_t * tc, vlib_buffer_t * b);
 void tcp_make_synack (tcp_connection_t * ts, vlib_buffer_t * b);
-void tcp_send_reset (tcp_connection_t * tc, vlib_buffer_t * pkt, u8 is_ip4);
+void tcp_send_reset_w_pkt (tcp_connection_t * tc, vlib_buffer_t * pkt,
+			   u8 is_ip4);
+void tcp_send_reset (tcp_connection_t * tc);
 void tcp_send_syn (tcp_connection_t * tc);
 void tcp_send_fin (tcp_connection_t * tc);
 void tcp_init_mss (tcp_connection_t * tc);
@@ -582,15 +584,30 @@ tcp_loss_wnd (const tcp_connection_t * tc)
 }
 
 always_inline u32
-tcp_available_wnd (const tcp_connection_t * tc)
+tcp_available_snd_wnd (const tcp_connection_t * tc)
 {
   return clib_min (tc->cwnd, tc->snd_wnd);
 }
 
+always_inline u32
+tcp_available_output_snd_space (const tcp_connection_t * tc)
+{
+  u32 available_wnd = tcp_available_snd_wnd (tc);
+  int flight_size = (int) (tc->snd_nxt - tc->snd_una);
+
+  if (available_wnd <= flight_size)
+    return 0;
+
+  return available_wnd - flight_size;
+}
+
+/**
+ * Estimate of how many bytes we can still push into the network
+ */
 always_inline u32
 tcp_available_snd_space (const tcp_connection_t * tc)
 {
-  u32 available_wnd = tcp_available_wnd (tc);
+  u32 available_wnd = tcp_available_snd_wnd (tc);
   u32 flight_size = tcp_flight_size (tc);
 
   if (available_wnd <= flight_size)
diff --git a/src/vnet/tcp/tcp_input.c b/src/vnet/tcp/tcp_input.c
index 6f9e4c7a..95f9ade1 100644
--- a/src/vnet/tcp/tcp_input.c
+++ b/src/vnet/tcp/tcp_input.c
@@ -1363,7 +1363,7 @@ always_inline int
 tcp_session_enqueue_data (tcp_connection_t * tc, vlib_buffer_t * b,
 			  u16 data_len)
 {
-  int written;
+  int written, error = TCP_ERROR_ENQUEUED;
 
   ASSERT (seq_geq (vnet_buffer (b)->tcp.seq_number, tc->rcv_nxt));
 
@@ -1381,12 +1381,12 @@ tcp_session_enqueue_data (tcp_connection_t * tc, vlib_buffer_t * b,
   /* Update rcv_nxt */
   if (PREDICT_TRUE (written == data_len))
     {
-      tc->rcv_nxt = vnet_buffer (b)->tcp.seq_end;
+      tc->rcv_nxt += written;
     }
   /* If more data written than expected, account for out-of-order bytes. */
   else if (written > data_len)
     {
-      tc->rcv_nxt = vnet_buffer (b)->tcp.seq_end + written - data_len;
+      tc->rcv_nxt += written;
 
       /* Send ACK confirming the update */
       tc->flags |= TCP_CONN_SNDACK;
@@ -1400,7 +1400,7 @@ tcp_session_enqueue_data (tcp_connection_t * tc, vlib_buffer_t * b,
        * not be enqueued. Inform peer */
       tc->flags |= TCP_CONN_SNDACK;
 
-      return TCP_ERROR_PARTIALLY_ENQUEUED;
+      error = TCP_ERROR_PARTIALLY_ENQUEUED;
     }
   else
     {
@@ -1415,7 +1415,7 @@ tcp_session_enqueue_data (tcp_connection_t * tc, vlib_buffer_t * b,
       tcp_update_sack_list (tc, tc->rcv_nxt, tc->rcv_nxt);
     }
 
-  return TCP_ERROR_ENQUEUED;
+  return error;
 }
 
 /** Enqueue out-of-order data */
@@ -1495,10 +1495,10 @@ tcp_can_delack (tcp_connection_t * tc)
 static int
 tcp_buffer_discard_bytes (vlib_buffer_t * b, u32 n_bytes_to_drop)
 {
-  u32 discard;
+  u32 discard, first = b->current_length;
   vlib_main_t *vm = vlib_get_main ();
 
-  /* Handle multi segment packets */
+  /* Handle multi-buffer segments */
   if (n_bytes_to_drop > b->current_length)
     {
       if (!(b->flags & VLIB_BUFFER_NEXT_PRESENT))
@@ -1511,7 +1511,12 @@ tcp_buffer_discard_bytes (vlib_buffer_t * b, u32 n_bytes_to_drop)
 	  n_bytes_to_drop -= discard;
 	}
       while (n_bytes_to_drop);
+      if (n_bytes_to_drop > first)
+	b->total_length_not_including_first_buffer -= n_bytes_to_drop - first;
     }
+  else
+    vlib_buffer_advance (b, n_bytes_to_drop);
+  vnet_buffer (b)->tcp.data_len -= n_bytes_to_drop;
   return 0;
 }
 
@@ -1908,7 +1913,7 @@ tcp46_syn_sent_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
 	      if (ack0 <= tc0->iss || ack0 > tc0->snd_nxt)
 		{
 		  if (!tcp_rst (tcp0))
-		    tcp_send_reset (tc0, b0, is_ip4);
+		    tcp_send_reset_w_pkt (tc0, b0, is_ip4);
 		  goto drop;
 		}
 
@@ -1995,7 +2000,7 @@ tcp46_syn_sent_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
 	       * allocate session send reset */
 	      if (stream_session_connect_notify (&new_tc0->connection, 0))
 		{
-		  tcp_send_reset (new_tc0, b0, is_ip4);
+		  tcp_send_reset_w_pkt (new_tc0, b0, is_ip4);
 		  tcp_connection_cleanup (new_tc0);
 		  goto drop;
 		}
@@ -2017,7 +2022,7 @@ tcp46_syn_sent_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
 	      if (stream_session_connect_notify (&new_tc0->connection, 0))
 		{
 		  tcp_connection_cleanup (new_tc0);
-		  tcp_send_reset (tc0, b0, is_ip4);
+		  tcp_send_reset_w_pkt (tc0, b0, is_ip4);
 		  TCP_EVT_DBG (TCP_EVT_RST_SENT, tc0);
 		  goto drop;
 		}
@@ -2221,7 +2226,7 @@ tcp46_rcv_process_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
 	       */
 	      if (!tcp_rcv_ack_is_acceptable (tc0, b0))
 		{
-		  tcp_send_reset (tc0, b0, is_ip4);
+		  tcp_send_reset_w_pkt (tc0, b0, is_ip4);
 		  goto drop;
 		}
 
diff --git a/src/vnet/tcp/tcp_output.c b/src/vnet/tcp/tcp_output.c
index 7da0c073..c56eadf8 100644
--- a/src/vnet/tcp/tcp_output.c
+++ b/src/vnet/tcp/tcp_output.c
@@ -440,7 +440,8 @@ tcp_init_mss (tcp_connection_t * tc)
 always_inline int
 tcp_alloc_tx_buffers (tcp_main_t * tm, u8 thread_index, u32 n_free_buffers)
 {
-  vec_validate (tm->tx_buffers[thread_index], n_free_buffers - 1);
+  vec_validate (tm->tx_buffers[thread_index],
+		vec_len (tm->tx_buffers[thread_index]) + n_free_buffers - 1);
   _vec_len (tm->tx_buffers[thread_index]) =
     vlib_buffer_alloc_from_free_list (vlib_get_main (),
 				      tm->tx_buffers[thread_index],
@@ -480,27 +481,31 @@ tcp_return_buffer (tcp_main_t * tm)
   _vec_len (my_tx_buffers) += 1;
 }
 
-always_inline void
+always_inline void *
 tcp_reuse_buffer (vlib_main_t * vm, vlib_buffer_t * b)
 {
-  vlib_buffer_t *it = b;
-  u32 save_free_list = b->flags & VLIB_BUFFER_FREE_LIST_INDEX_MASK;
-  do
-    {
-      it->current_data = 0;
-      it->current_length = 0;
-      it->total_length_not_including_first_buffer = 0;
-    }
-  while ((it->flags & VLIB_BUFFER_NEXT_PRESENT)
-	 && (it = vlib_get_buffer (vm, it->next_buffer)));
-
   if (b->flags & VLIB_BUFFER_NEXT_PRESENT)
     vlib_buffer_free_one (vm, b->next_buffer);
-  b->flags = save_free_list;
+  b->flags = 0;
+  b->current_data = 0;
+  b->current_length = 0;
+  b->total_length_not_including_first_buffer = 0;
+  vnet_buffer (b)->tcp.flags = 0;
 
   /* Leave enough space for headers */
-  vlib_buffer_make_headroom (b, MAX_HDRS_LEN);
+  return vlib_buffer_make_headroom (b, MAX_HDRS_LEN);
+}
+
+always_inline void *
+tcp_init_buffer (vlib_main_t * vm, vlib_buffer_t * b)
+{
+  ASSERT ((b->flags & VLIB_BUFFER_NEXT_PRESENT) == 0);
+  b->flags = VNET_BUFFER_F_LOCALLY_ORIGINATED;
+  b->total_length_not_including_first_buffer = 0;
   vnet_buffer (b)->tcp.flags = 0;
+
+  /* Leave enough space for headers */
+  return vlib_buffer_make_headroom (b, MAX_HDRS_LEN);
 }
 
 /**
@@ -632,6 +637,59 @@ tcp_enqueue_to_ip_lookup (vlib_main_t * vm, vlib_buffer_t * b, u32 bi,
   vlib_put_frame_to_node (vm, next_index, f);
 }
 
+always_inline void
+tcp_enqueue_to_output_i (vlib_main_t * vm, vlib_buffer_t * b, u32 bi,
+			 u8 is_ip4, u8 flush)
+{
+  tcp_main_t *tm = vnet_get_tcp_main ();
+  u32 thread_index = vlib_get_thread_index ();
+  u32 *to_next, next_index;
+  vlib_frame_t *f;
+
+  b->flags |= VNET_BUFFER_F_LOCALLY_ORIGINATED;
+  b->error = 0;
+
+  /* Decide where to send the packet */
+  next_index = is_ip4 ? tcp4_output_node.index : tcp6_output_node.index;
+
+  /* Initialize the trajectory trace, if configured */
+  if (VLIB_BUFFER_TRACE_TRAJECTORY > 0)
+    {
+      b->pre_data[0] = 1;
+      b->pre_data[1] = next_index;
+    }
+
+  /* Get frame to v4/6 output node */
+  f = tm->tx_frames[!is_ip4][thread_index];
+  if (!f)
+    {
+      f = vlib_get_frame_to_node (vm, next_index);
+      ASSERT (f);
+      tm->tx_frames[!is_ip4][thread_index] = f;
+    }
+  to_next = vlib_frame_vector_args (f);
+  to_next[f->n_vectors] = bi;
+  f->n_vectors += 1;
+  if (flush || f->n_vectors == VLIB_FRAME_SIZE)
+    {
+      vlib_put_frame_to_node (vm, next_index, f);
+      tm->tx_frames[!is_ip4][thread_index] = 0;
+    }
+}
+
+always_inline void
+tcp_enqueue_to_output (vlib_main_t * vm, vlib_buffer_t * b, u32 bi, u8 is_ip4)
+{
+  tcp_enqueue_to_output_i (vm, b, bi, is_ip4, 0);
+}
+
+always_inline void
+tcp_enqueue_to_output_now (vlib_main_t * vm, vlib_buffer_t * b, u32 bi,
+			   u8 is_ip4)
+{
+  tcp_enqueue_to_output_i (vm, b, bi, is_ip4, 1);
+}
+
 int
 tcp_make_reset_in_place (vlib_main_t * vm, vlib_buffer_t * b0,
 			 tcp_state_t state, u8 thread_index, u8 is_ip4)
@@ -712,9 +770,11 @@ tcp_make_reset_in_place (vlib_main_t * vm, vlib_buffer_t * b0,
 
 /**
  *  Send reset without reusing existing buffer
+ *
+ *  It extracts connection info out of original packet
  */
 void
-tcp_send_reset (tcp_connection_t * tc, vlib_buffer_t * pkt, u8 is_ip4)
+tcp_send_reset_w_pkt (tcp_connection_t * tc, vlib_buffer_t * pkt, u8 is_ip4)
 {
   vlib_buffer_t *b;
   u32 bi;
@@ -730,9 +790,7 @@ tcp_send_reset (tcp_connection_t * tc, vlib_buffer_t * pkt, u8 is_ip4)
     return;
 
   b = vlib_get_buffer (vm, bi);
-
-  /* Leave enough space for headers */
-  vlib_buffer_make_headroom (b, MAX_HDRS_LEN);
+  tcp_init_buffer (vm, b);
 
   /* Make and write options */
   tcp_hdr_len = sizeof (tcp_header_t);
@@ -787,6 +845,38 @@ tcp_send_reset (tcp_connection_t * tc, vlib_buffer_t * pkt, u8 is_ip4)
   TCP_EVT_DBG (TCP_EVT_RST_SENT, tc);
 }
 
+/**
+ * Build and set reset packet for connection
+ */
+void
+tcp_send_reset (tcp_connection_t * tc)
+{
+  vlib_main_t *vm = vlib_get_main ();
+  tcp_main_t *tm = vnet_get_tcp_main ();
+  vlib_buffer_t *b;
+  u32 bi;
+  tcp_header_t *th;
+  u16 tcp_hdr_opts_len, advertise_wnd, opts_write_len;
+  u8 flags;
+
+  if (PREDICT_FALSE (tcp_get_free_buffer_index (tm, &bi)))
+    return;
+  b = vlib_get_buffer (vm, bi);
+  tcp_init_buffer (vm, b);
+
+  tc->snd_opts_len = tcp_make_options (tc, &tc->snd_opts, tc->state);
+  tcp_hdr_opts_len = tc->snd_opts_len + sizeof (tcp_header_t);
+  advertise_wnd = tcp_window_to_advertise (tc, TCP_STATE_ESTABLISHED);
+  flags = TCP_FLAG_RST;
+  th = vlib_buffer_push_tcp (b, tc->c_lcl_port, tc->c_rmt_port, tc->snd_nxt,
+			     tc->rcv_nxt, tcp_hdr_opts_len, flags,
+			     advertise_wnd);
+  opts_write_len = tcp_options_write ((u8 *) (th + 1), &tc->snd_opts);
+  ASSERT (opts_write_len == tc->snd_opts_len);
+  vnet_buffer (b)->tcp.connection_index = tc->c_c_index;
+  tcp_enqueue_to_output_now (vm, b, bi, tc->c_is_ip4);
+}
+
 void
 tcp_push_ip_hdr (tcp_main_t * tm, tcp_connection_t * tc, vlib_buffer_t * b)
 {
@@ -835,9 +925,7 @@ tcp_send_syn (tcp_connection_t * tc)
     return;
 
   b = vlib_get_buffer (vm, bi);
-
-  /* Leave enough space for headers */
-  vlib_buffer_make_headroom (b, MAX_HDRS_LEN);
+  tcp_init_buffer (vm, b);
 
   /* Set random initial sequence */
   time_now = tcp_time_now ();
@@ -875,59 +963,6 @@ tcp_send_syn (tcp_connection_t * tc)
   TCP_EVT_DBG (TCP_EVT_SYN_SENT, tc);
 }
 
-always_inline void
-tcp_enqueue_to_output_i (vlib_main_t * vm, vlib_buffer_t * b, u32 bi,
-			 u8 is_ip4, u8 flush)
-{
-  tcp_main_t *tm = vnet_get_tcp_main ();
-  u32 thread_index = vlib_get_thread_index ();
-  u32 *to_next, next_index;
-  vlib_frame_t *f;
-
-  b->flags |= VNET_BUFFER_F_LOCALLY_ORIGINATED;
-  b->error = 0;
-
-  /* Decide where to send the packet */
-  next_index = is_ip4 ? tcp4_output_node.index : tcp6_output_node.index;
-
-  /* Initialize the trajectory trace, if configured */
-  if (VLIB_BUFFER_TRACE_TRAJECTORY > 0)
-    {
-      b->pre_data[0] = 1;
-      b->pre_data[1] = next_index;
-    }
-
-  /* Get frame to v4/6 output node */
-  f = tm->tx_frames[!is_ip4][thread_index];
-  if (!f)
-    {
-      f = vlib_get_frame_to_node (vm, next_index);
-      ASSERT (f);
-      tm->tx_frames[!is_ip4][thread_index] = f;
-    }
-  to_next = vlib_frame_vector_args (f);
-  to_next[f->n_vectors] = bi;
-  f->n_vectors += 1;
-  if (flush || f->n_vectors == VLIB_FRAME_SIZE)
-    {
-      vlib_put_frame_to_node (vm, next_index, f);
-      tm->tx_frames[!is_ip4][thread_index] = 0;
-    }
-}
-
-always_inline void
-tcp_enqueue_to_output (vlib_main_t * vm, vlib_buffer_t * b, u32 bi, u8 is_ip4)
-{
-  tcp_enqueue_to_output_i (vm, b, bi, is_ip4, 0);
-}
-
-always_inline void
-tcp_enqueue_to_output_now (vlib_main_t * vm, vlib_buffer_t * b, u32 bi,
-			   u8 is_ip4)
-{
-  tcp_enqueue_to_output_i (vm, b, bi, is_ip4, 1);
-}
-
 /**
  * Flush tx frame populated by retransmits and timer pops
  */
@@ -969,7 +1004,7 @@ tcp_send_fin (tcp_connection_t * tc)
   if (PREDICT_FALSE (tcp_get_free_buffer_index (tm, &bi)))
     return;
   b = vlib_get_buffer (vm, bi);
-
+  /* buffer will be initialized by in tcp_make_fin */
   tcp_make_fin (tc, b);
   tcp_enqueue_to_output_now (vm, b, bi, tc->c_is_ip4);
   tc->flags |= TCP_CONN_FINSNT;
@@ -1013,6 +1048,8 @@ tcp_push_hdr_i (tcp_connection_t * tc, vlib_buffer_t * b,
   tcp_header_t *th;
 
   data_len = b->current_length + b->total_length_not_including_first_buffer;
+  ASSERT (!b->total_length_not_including_first_buffer
+	  || (b->flags & VLIB_BUFFER_NEXT_PRESENT));
   vnet_buffer (b)->tcp.flags = 0;
 
   if (compute_opts)
@@ -1106,29 +1143,27 @@ tcp_prepare_retransmit_segment (tcp_connection_t * tc, u32 offset,
    * Make sure we can retransmit something
    */
   available_bytes = stream_session_tx_fifo_max_dequeue (&tc->connection);
+  available_bytes -= offset;
   if (!available_bytes)
     return 0;
   max_deq_bytes = clib_min (tc->snd_mss, max_deq_bytes);
   max_deq_bytes = clib_min (available_bytes, max_deq_bytes);
-  seg_size = max_deq_bytes + MAX_HDRS_LEN;
 
   /* Start is beyond snd_congestion */
   start = tc->snd_una + offset;
   if (seq_geq (start, tc->snd_congestion))
-    {
-      goto done;
-    }
+    goto done;
 
   /* Don't overshoot snd_congestion */
   if (seq_gt (start + max_deq_bytes, tc->snd_congestion))
     {
       max_deq_bytes = tc->snd_congestion - start;
       if (max_deq_bytes == 0)
-	{
-	  goto done;
-	}
+	goto done;
     }
 
+  seg_size = max_deq_bytes + MAX_HDRS_LEN;
+
   /*
    * Prepare options
    */
@@ -1141,7 +1176,7 @@ tcp_prepare_retransmit_segment (tcp_connection_t * tc, u32 offset,
   if (PREDICT_FALSE (tcp_get_free_buffer_index (tm, &bi)))
     return 0;
   *b = vlib_get_buffer (vm, bi);
-  data = vlib_buffer_make_headroom (*b, MAX_HDRS_LEN);
+  data = tcp_init_buffer (vm, *b);
 
   /* Easy case, buffer size greater than mss */
   if (PREDICT_TRUE (seg_size <= tm->bytes_per_buffer))
@@ -1162,7 +1197,6 @@ tcp_prepare_retransmit_segment (tcp_connection_t * tc, u32 offset,
       int i;
 
       n_bufs_per_seg = ceil ((double) seg_size / tm->bytes_per_buffer);
-      ASSERT (available_bytes >= max_deq_bytes);
 
       /* Make sure we have enough buffers */
       available_bufs = vec_len (tm->tx_buffers[thread_index]);
@@ -1182,8 +1216,6 @@ tcp_prepare_retransmit_segment (tcp_connection_t * tc, u32 offset,
       b[0]->current_length = n_bytes;
       b[0]->flags |= VLIB_BUFFER_TOTAL_LENGTH_VALID;
       b[0]->total_length_not_including_first_buffer = 0;
-
-      tcp_push_hdr_i (tc, *b, tc->state, 0);
       max_deq_bytes -= n_bytes;
 
       chain_b = *b;
@@ -1197,22 +1229,22 @@ tcp_prepare_retransmit_segment (tcp_connection_t * tc, u32 offset,
 	  chain_b->current_data = 0;
 	  data = vlib_buffer_get_current (chain_b);
 	  n_peeked = stream_session_peek_bytes (&tc->connection, data,
-						n_bytes, len_to_deq);
-	  n_bytes += n_peeked;
+						offset + n_bytes, len_to_deq);
 	  ASSERT (n_peeked == len_to_deq);
+	  n_bytes += n_peeked;
 	  chain_b->current_length = n_peeked;
-	  b[0]->total_length_not_including_first_buffer +=
-	    chain_b->current_length;
+	  chain_b->flags = 0;
+	  chain_b->next_buffer = 0;
 
 	  /* update previous buffer */
 	  prev_b->next_buffer = chain_bi;
 	  prev_b->flags |= VLIB_BUFFER_NEXT_PRESENT;
 
-	  /* update current buffer */
-	  chain_b->next_buffer = 0;
-
 	  max_deq_bytes -= n_peeked;
+	  b[0]->total_length_not_including_first_buffer += n_peeked;
 	}
+
+      tcp_push_hdr_i (tc, *b, tc->state, 0);
     }
 
   ASSERT (n_bytes > 0);
@@ -1348,7 +1380,7 @@ tcp_timer_retransmit_handler_i (u32 index, u8 is_syn)
       if (PREDICT_FALSE (tcp_get_free_buffer_index (tm, &bi)))
 	return;
       b = vlib_get_buffer (vm, bi);
-      vlib_buffer_make_headroom (b, MAX_HDRS_LEN);
+      tcp_init_buffer (vm, b);
       tcp_push_hdr_i (tc, b, tc->state, 1);
 
       /* Account for the SYN */
@@ -1409,8 +1441,9 @@ tcp_timer_persist_handler (u32 index)
   u32 thread_index = vlib_get_thread_index ();
   tcp_connection_t *tc;
   vlib_buffer_t *b;
-  u32 bi, old_snd_nxt, snd_bytes = 0, available_bytes = 0;
+  u32 bi, old_snd_nxt, max_snd_bytes, available_bytes, offset;
   int n_bytes = 0;
+  u8 *data;
 
   tc = tcp_connection_get_if_valid (index, thread_index);
 
@@ -1419,12 +1452,13 @@ tcp_timer_persist_handler (u32 index)
 
   /* Make sure timer handle is set to invalid */
   tc->timers[TCP_TIMER_PERSIST] = TCP_TIMER_HANDLE_INVALID;
+  offset = tc->snd_una_max - tc->snd_una;
 
   /* Problem already solved or worse */
   available_bytes = stream_session_tx_fifo_max_dequeue (&tc->connection);
   if (tc->state == TCP_STATE_CLOSED || tc->state > TCP_STATE_ESTABLISHED
       || tc->snd_wnd > tc->snd_mss || tcp_in_recovery (tc)
-      || !available_bytes)
+      || !available_bytes || available_bytes <= offset)
     return;
 
   /* Increment RTO backoff */
@@ -1437,18 +1471,17 @@ tcp_timer_persist_handler (u32 index)
   if (PREDICT_FALSE (tcp_get_free_buffer_index (tm, &bi)))
     return;
   b = vlib_get_buffer (vm, bi);
+  data = tcp_init_buffer (vm, b);
 
-  tcp_validate_txf_size (tc, tc->snd_una_max - tc->snd_una);
+  tcp_validate_txf_size (tc, offset);
   tc->snd_opts_len = tcp_make_options (tc, &tc->snd_opts, tc->state);
-  snd_bytes = clib_min (tc->snd_mss, tm->bytes_per_buffer);
-  n_bytes = stream_session_peek_bytes (&tc->connection,
-				       vlib_buffer_get_current (b),
-				       tc->snd_una_max - tc->snd_una,
-				       snd_bytes);
-  ASSERT (n_bytes != 0);
+  max_snd_bytes = clib_min (tc->snd_mss, tm->bytes_per_buffer - MAX_HDRS_LEN);
+  n_bytes = stream_session_peek_bytes (&tc->connection, data, offset,
+				       max_snd_bytes);
   b->current_length = n_bytes;
-  ASSERT (tc->snd_nxt == tc->snd_una_max || tc->rto_boff > 1
-	  || tcp_timer_is_active (tc, TCP_TIMER_RETRANSMIT));
+  ASSERT (n_bytes != 0 && (tc->snd_nxt == tc->snd_una_max || tc->rto_boff > 1
+			   || tcp_timer_is_active (tc,
+						   TCP_TIMER_RETRANSMIT)));
 
   /* Allow updating of snd_una_max but don't update snd_nxt */
   old_snd_nxt = tc->snd_nxt;
@@ -1456,8 +1489,8 @@ tcp_timer_persist_handler (u32 index)
   tc->snd_nxt = old_snd_nxt;
   tcp_enqueue_to_output (vm, b, bi, tc->c_is_ip4);
 
-  /* Re-enable persist timer */
-  tcp_persist_timer_set (tc);
+  /* Just sent new data, enable retransmit */
+  tcp_retransmit_timer_update (tc);
 }
 
 /**
@@ -1490,7 +1523,7 @@ void
 tcp_fast_retransmit_sack (tcp_connection_t * tc)
 {
   vlib_main_t *vm = vlib_get_main ();
-  u32 n_written = 0, offset = 0, max_bytes;
+  u32 n_written = 0, offset, max_bytes;
   vlib_buffer_t *b = 0;
   sack_scoreboard_hole_t *hole;
   sack_scoreboard_t *sb;
@@ -1523,7 +1556,9 @@ tcp_fast_retransmit_sack (tcp_connection_t * tc)
 	   * unSACKed sequence number SHOULD be returned, and RescueRxt set to
 	   * RecoveryPoint. HighRxt MUST NOT be updated.
 	   */
-	  max_bytes = clib_min (tc->snd_mss, snd_space);
+	  max_bytes = clib_min (tc->snd_mss,
+				tc->snd_congestion - tc->snd_una);
+	  max_bytes = clib_min (max_bytes, snd_space);
 	  offset = tc->snd_congestion - tc->snd_una - max_bytes;
 	  sb->rescue_rxt = tc->snd_congestion;
 	  tc->snd_nxt = tc->snd_una + offset;
@@ -1535,9 +1570,12 @@ tcp_fast_retransmit_sack (tcp_connection_t * tc)
 	  break;
 	}
 
-      max_bytes = snd_limited ? tc->snd_mss : hole->end - sb->high_rxt;
+      max_bytes = clib_min (hole->end - sb->high_rxt, snd_space);
+      max_bytes = snd_limited ? clib_min (max_bytes, tc->snd_mss) : max_bytes;
+      if (max_bytes == 0)
+	break;
       offset = sb->high_rxt - tc->snd_una;
-      tc->snd_nxt = tc->snd_una + offset;
+      tc->snd_nxt = sb->high_rxt;
       n_written = tcp_prepare_retransmit_segment (tc, offset, max_bytes, &b);
 
       /* Nothing left to retransmit */
@@ -1547,6 +1585,7 @@ tcp_fast_retransmit_sack (tcp_connection_t * tc)
       bi = vlib_get_buffer_index (vm, b);
       sb->high_rxt += n_written;
       tcp_enqueue_to_output (vm, b, bi, tc->c_is_ip4);
+      ASSERT (n_written <= snd_space);
       snd_space -= n_written;
     }
 
@@ -1835,6 +1874,7 @@ tcp_push_header (transport_connection_t * tconn, vlib_buffer_t * b)
 
   tc = (tcp_connection_t *) tconn;
   tcp_push_hdr_i (tc, b, TCP_STATE_ESTABLISHED, 0);
+  ASSERT (seq_leq (tc->snd_una_max, tc->snd_una + tc->snd_wnd));
 
   if (tc->rtt_ts == 0 && !tcp_in_cong_recovery (tc))
     {
-- 
cgit 1.2.3-korg


From ab7b8d93cf1098970bc17fb4937376bb1ff33a21 Mon Sep 17 00:00:00 2001
From: Chris Luke <chrisy@flirble.org>
Date: Thu, 7 Sep 2017 07:40:13 -0400
Subject: Fixes for issues reported by Coverity (VPP-972)

Change-Id: I25238debb7081b4467aec4620dfdef33fbef3295
Signed-off-by: Chris Luke <chrisy@flirble.org>
---
 src/svm/svm_fifo.c         |  2 +-
 src/uri/sock_test_client.c |  2 ++
 src/uri/sock_test_server.c | 12 +++++++++---
 src/uri/vppcom.c           | 17 ++++++++---------
 src/vlib/unix/main.c       |  8 ++++----
 src/vnet/ipsec/ikev2.c     | 11 ++++++++++-
 src/vppinfra/socket.c      | 16 ++++++++++++++--
 7 files changed, 48 insertions(+), 20 deletions(-)

(limited to 'src/svm/svm_fifo.c')

diff --git a/src/svm/svm_fifo.c b/src/svm/svm_fifo.c
index 8fe82f56..42eb1ee8 100644
--- a/src/svm/svm_fifo.c
+++ b/src/svm/svm_fifo.c
@@ -443,7 +443,7 @@ ooo_segment_try_collect (svm_fifo_t * f, u32 n_bytes_enqueued)
 	}
     }
 
-  ASSERT (bytes >= 0 && bytes <= f->nitems);
+  ASSERT (bytes <= f->nitems);
   return bytes;
 }
 
diff --git a/src/uri/sock_test_client.c b/src/uri/sock_test_client.c
index 4319f01b..ab8e5a0e 100644
--- a/src/uri/sock_test_client.c
+++ b/src/uri/sock_test_client.c
@@ -895,6 +895,8 @@ main (int argc, char **argv)
 	  case 'w':
 	    fprintf (stderr, "ERROR: Option -%c requires an argument.\n",
 		     optopt);
+	    break;
+
 	  default:
 	    if (isprint (optopt))
 	      fprintf (stderr, "ERROR: Unknown option `-%c'.\n", optopt);
diff --git a/src/uri/sock_test_server.c b/src/uri/sock_test_server.c
index f703b177..29adea25 100644
--- a/src/uri/sock_test_server.c
+++ b/src/uri/sock_test_server.c
@@ -514,9 +514,15 @@ main (int argc, char **argv)
 		      continue;
 		    }
 
-		  else if (strlen ((char *) conn->buf))
-		    printf ("\nSERVER (fd %d): RX (%d bytes) - '%s'\n",
-			    conn->fd, rx_bytes, conn->buf);
+		  else if (((char *) conn->buf)[0] != 0)
+		    {
+		      // If it looks vaguely like a string, make sure it's terminated
+		      ((char *) conn->buf)[rx_bytes <
+					   conn->buf_size ? rx_bytes :
+					   conn->buf_size - 1] = 0;
+		      printf ("\nSERVER (fd %d): RX (%d bytes) - '%s'\n",
+			      conn->fd, rx_bytes, conn->buf);
+		    }
 		}
 	      else		// rx_bytes < 0
 		{
diff --git a/src/uri/vppcom.c b/src/uri/vppcom.c
index aec1295f..aa307f1d 100644
--- a/src/uri/vppcom.c
+++ b/src/uri/vppcom.c
@@ -1369,23 +1369,18 @@ vppcom_cfg_heapsize (char *conf_fname)
 	  argc++;
 	  char **tmp = realloc (argv, argc * sizeof (char *));
 	  if (tmp == NULL)
-	    {
-	      fclose (fp);
-	      goto defaulted;
-	    }
+	    goto defaulted;
 	  argv = tmp;
 	  arg = strndup (p, 1024);
 	  if (arg == NULL)
-	    {
-	      fclose (fp);
-	      goto defaulted;
-	    }
+	    goto defaulted;
 	  argv[argc - 1] = arg;
 	  p = strtok (NULL, " \t\n");
 	}
     }
 
   fclose (fp);
+  fp = NULL;
 
   char **tmp = realloc (argv, (argc + 1) * sizeof (char *));
   if (tmp == NULL)
@@ -1438,6 +1433,10 @@ vppcom_cfg_heapsize (char *conf_fname)
     }
 
 defaulted:
+  if (fp != NULL)
+    fclose (fp);
+  if (argv != NULL)
+    free (argv);
   if (!clib_mem_init (0, vcl_cfg->heapsize))
     clib_warning ("[%d] vppcom heap allocation failure!", vcm->my_pid);
   else if (VPPCOM_DEBUG > 0)
@@ -1687,7 +1686,7 @@ input_done:
   unformat_free (input);
 
 file_done:
-  if (fd > 0)
+  if (fd >= 0)
     close (fd);
 }
 
diff --git a/src/vlib/unix/main.c b/src/vlib/unix/main.c
index c90e1331..3a92b2e3 100644
--- a/src/vlib/unix/main.c
+++ b/src/vlib/unix/main.c
@@ -434,6 +434,10 @@ unix_config (vlib_main_t * vm, unformat_input_t * input)
 				  vlib_default_runtime_dir, 0);
     }
 
+  error = setup_signal_handlers (um);
+  if (error)
+    return error;
+
   if (um->pidfile)
     {
       if ((error = vlib_unix_validate_runtime_file (um,
@@ -448,10 +452,6 @@ unix_config (vlib_main_t * vm, unformat_input_t * input)
 	}
     }
 
-  error = setup_signal_handlers (um);
-  if (error)
-    return error;
-
   if (!(um->flags & UNIX_FLAG_INTERACTIVE))
     {
       openlog (vm->name, LOG_CONS | LOG_PERROR | LOG_PID, LOG_DAEMON);
diff --git a/src/vnet/ipsec/ikev2.c b/src/vnet/ipsec/ikev2.c
index 296654ec..a3dc7b87 100644
--- a/src/vnet/ipsec/ikev2.c
+++ b/src/vnet/ipsec/ikev2.c
@@ -17,6 +17,7 @@
 #include <vnet/vnet.h>
 #include <vnet/pg/pg.h>
 #include <vppinfra/error.h>
+#include <vppinfra/random.h>
 #include <vnet/udp/udp.h>
 #include <vnet/ipsec/ipsec.h>
 #include <vnet/ipsec/ikev2.h>
@@ -1595,8 +1596,16 @@ ikev2_create_tunnel_interface (vnet_main_t * vnm, ikev2_sa_t * sa,
 	+ sa->profile->lifetime;
       if (sa->profile->lifetime_jitter)
 	{
+	  // This is not much better than rand(3), which Coverity warns
+	  // is unsuitable for security applications; random_u32 is
+	  // however fast. If this perturbance to the expiration time
+	  // needs to use a better RNG then we may need to use something
+	  // like /dev/urandom which has significant overhead.
+	  u32 rnd = (u32) (vlib_time_now (vnm->vlib_main) * 1e6);
+	  rnd = random_u32 (&rnd);
+
 	  child->time_to_expiration +=
-	    1 + (rand () % sa->profile->lifetime_jitter);
+	    1 + (rnd % sa->profile->lifetime_jitter);
 	}
     }
 
diff --git a/src/vppinfra/socket.c b/src/vppinfra/socket.c
index 7ade440c..37dcbbfd 100644
--- a/src/vppinfra/socket.c
+++ b/src/vppinfra/socket.c
@@ -359,9 +359,21 @@ clib_socket_init (clib_socket_t * s)
 	  && s->flags & SOCKET_ALLOW_GROUP_WRITE)
 	{
 	  struct stat st = { 0 };
-	  stat (((struct sockaddr_un *) &addr)->sun_path, &st);
+	  if (stat (((struct sockaddr_un *) &addr)->sun_path, &st) < 0)
+	    {
+	      error = clib_error_return_unix (0, "stat (fd %d, '%s')",
+					      s->fd, s->config);
+	      goto done;
+	    }
 	  st.st_mode |= S_IWGRP;
-	  chmod (((struct sockaddr_un *) &addr)->sun_path, st.st_mode);
+	  if (chmod (((struct sockaddr_un *) &addr)->sun_path, st.st_mode) <
+	      0)
+	    {
+	      error =
+		clib_error_return_unix (0, "chmod (fd %d, '%s', mode %o)",
+					s->fd, s->config, st.st_mode);
+	      goto done;
+	    }
 	}
     }
   else
-- 
cgit 1.2.3-korg