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>

17 files changed, 664 insertions, 426 deletions

diff --git a/src/vnet/session/application.c b/src/vnet/session/application.c
index c679b1f53f8..4bdb10271ab 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 07eeae82915..c0ab1bf096f 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 caf8eaa308a..bf571963b0e 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 534598d637f..fe198044db9 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 d9c38bd1ad8..5fa4225c4ed 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 f772cb9fdb4..60f764af48a 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 e5f788bee08..04bd5ca0f29 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 7238cda3e60..6f8be082d95 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 d5d79e53a22..3462e0eeee8 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 8bd2f3605e6..775bfc26e47 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 e0b67a8e5e5..5c554bac5a9 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 071f1ab185f..e83987182fa 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 3a16cf63194..ae68ad1b264 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 ff2229b3792..a2e6dad1298 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 c66250e41ea..c825e952c9b 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 47c94e6daca..554a981d924 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 3f8afa40e54..a461e3b8b7d 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));