session: cleanup session tx function

- rework the function to declutter and avoid building more than one tx
  frame
- add dual loop although benefits in my tests seem to be minimal
- improve tcp/udp echo external apps. They have slightly better
  throughput than internal echo apps.
- udp bugfixes

Change-Id: Iea4a245b1b1bb407a7f403dedcce2664a49f774b
Signed-off-by: Florin Coras <fcoras@cisco.com>

4 files changed, 394 insertions, 320 deletions

diff --git a/src/vnet/session/application_interface.h b/src/vnet/session/application_interface.h
index 5fd218533dc..49c2862ac45 100644
--- a/src/vnet/session/application_interface.h
+++ b/src/vnet/session/application_interface.h
@@ -204,7 +204,7 @@ app_send_dgram_raw (svm_fifo_t * f, app_session_transport_t * at,
   int rv;
 
   max_enqueue = svm_fifo_max_enqueue (f);
-  if (svm_fifo_max_enqueue (f) <= sizeof (session_dgram_hdr_t))
+  if (max_enqueue <= sizeof (session_dgram_hdr_t))
     return 0;
 
   max_enqueue -= sizeof (session_dgram_hdr_t);
@@ -217,9 +217,6 @@ app_send_dgram_raw (svm_fifo_t * f, app_session_transport_t * at,
   clib_memcpy (&hdr.lcl_ip, &at->lcl_ip, sizeof (ip46_address_t));
   hdr.lcl_port = at->lcl_port;
   rv = svm_fifo_enqueue_nowait (f, sizeof (hdr), (u8 *) & hdr);
-  if (rv <= 0)
-    return 0;
-
   ASSERT (rv == sizeof (hdr));
 
   if ((rv = svm_fifo_enqueue_nowait (f, actual_write, data)) > 0)
@@ -231,6 +228,7 @@ app_send_dgram_raw (svm_fifo_t * f, app_session_transport_t * at,
 	  svm_queue_add (vpp_evt_q, (u8 *) & evt, noblock);
 	}
     }
+  ASSERT (rv);
   return rv;
 }
 
diff --git a/src/vnet/session/session.c b/src/vnet/session/session.c
index 568065ebc0d..a8a9c66ac81 100644
--- a/src/vnet/session/session.c
+++ b/src/vnet/session/session.c
@@ -1272,6 +1272,7 @@ session_manager_main_enable (vlib_main_t * vm)
   vec_validate (smm->free_event_vector, num_threads - 1);
   vec_validate (smm->vpp_event_queues, num_threads - 1);
   vec_validate (smm->peekers_rw_locks, num_threads - 1);
+  vec_validate_aligned (smm->ctx, num_threads - 1, CLIB_CACHE_LINE_BYTES);
 
   for (i = 0; i < TRANSPORT_N_PROTO; i++)
     {
diff --git a/src/vnet/session/session.h b/src/vnet/session/session.h
index b54f46189f3..c5779b4ffd1 100644
--- a/src/vnet/session/session.h
+++ b/src/vnet/session/session.h
@@ -103,14 +103,58 @@ typedef CLIB_PACKED (struct {
 }) session_fifo_event_t;
 /* *INDENT-ON* */
 
+typedef struct session_dgram_pre_hdr_
+{
+  u32 data_length;
+  u32 data_offset;
+} session_dgram_pre_hdr_t;
+
+/* *INDENT-OFF* */
+typedef CLIB_PACKED (struct session_dgram_header_
+{
+  u32 data_length;
+  u32 data_offset;
+  ip46_address_t rmt_ip;
+  ip46_address_t lcl_ip;
+  u16 rmt_port;
+  u16 lcl_port;
+  u8 is_ip4;
+}) session_dgram_hdr_t;
+/* *INDENT-ON* */
+
+#define SESSION_CONN_ID_LEN 37
+#define SESSION_CONN_HDR_LEN 45
+
+STATIC_ASSERT (sizeof (session_dgram_hdr_t) == (SESSION_CONN_ID_LEN + 8),
+	       "session conn id wrong length");
+
+typedef struct session_tx_context_
+{
+  CLIB_CACHE_LINE_ALIGN_MARK (cacheline0);
+  stream_session_t *s;
+  transport_proto_vft_t *transport_vft;
+  transport_connection_t *tc;
+  vlib_buffer_t *b;
+  u32 max_dequeue;
+  u32 snd_space;
+  u32 left_to_snd;
+  u32 tx_offset;
+  u32 max_len_to_snd;
+  u16 deq_per_first_buf;
+  u16 deq_per_buf;
+  u16 snd_mss;
+  u8 n_bufs_per_seg;
+    CLIB_CACHE_LINE_ALIGN_MARK (cacheline1);
+  session_dgram_hdr_t hdr;
+} session_tx_context_t;
+
 /* Forward definition */
 typedef struct _session_manager_main session_manager_main_t;
 
 typedef int
   (session_fifo_rx_fn) (vlib_main_t * vm, vlib_node_runtime_t * node,
-			session_manager_main_t * smm,
 			session_fifo_event_t * e0, stream_session_t * s0,
-			u32 thread_index, int *n_tx_pkts);
+			int *n_tx_pkts);
 
 extern session_fifo_rx_fn session_tx_fifo_peek_and_snd;
 extern session_fifo_rx_fn session_tx_fifo_dequeue_and_snd;
@@ -144,6 +188,9 @@ struct _session_manager_main
   /** per-worker postponed disconnects */
   session_fifo_event_t **pending_disconnects;
 
+  /** per-worker session context */
+  session_tx_context_t *ctx;
+
   /** vpp fifo event queue */
   svm_queue_t **vpp_event_queues;
 
@@ -204,31 +251,6 @@ struct _session_manager_main
 
 };
 
-typedef struct session_dgram_pre_hdr_
-{
-  u32 data_length;
-  u32 data_offset;
-} session_dgram_pre_hdr_t;
-
-/* *INDENT-OFF* */
-typedef CLIB_PACKED (struct session_dgram_header_
-{
-  u32 data_length;
-  u32 data_offset;
-  ip46_address_t rmt_ip;
-  ip46_address_t lcl_ip;
-  u16 rmt_port;
-  u16 lcl_port;
-  u8 is_ip4;
-}) session_dgram_hdr_t;
-/* *INDENT-ON* */
-
-#define SESSION_CONN_ID_LEN 37
-#define SESSION_CONN_HDR_LEN 45
-
-STATIC_ASSERT (sizeof (session_dgram_hdr_t) == (SESSION_CONN_ID_LEN + 8),
-	       "session conn id wrong length");
-
 extern session_manager_main_t session_manager_main;
 extern vlib_node_registration_t session_queue_node;
 
diff --git a/src/vnet/session/session_node.c b/src/vnet/session/session_node.c
index 14716965547..269e2fb591e 100644
--- a/src/vnet/session/session_node.c
+++ b/src/vnet/session/session_node.c
@@ -18,6 +18,7 @@
 #include <vnet/vnet.h>
 #include <vppinfra/elog.h>
 #include <vnet/session/transport.h>
+#include <vnet/session/session.h>
 #include <vnet/session/application.h>
 #include <vnet/session/session_debug.h>
 #include <svm/queue.h>
@@ -65,420 +66,473 @@ static char *session_queue_error_strings[] = {
 };
 
 always_inline void
-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 * tx_offset, u16 deq_per_buf,
-			    u8 peek_data, transport_tx_fn_type_t tx_type)
+session_tx_trace_buffer (vlib_main_t * vm, vlib_node_runtime_t * node,
+			 u32 next_index, vlib_buffer_t * b,
+			 stream_session_t * s, u32 * n_trace)
 {
-  vlib_buffer_t *chain_b0, *prev_b0;
-  u32 chain_bi0, to_deq;
-  u16 len_to_deq0, n_bytes_read;
-  u8 *data0, j;
-
-  b0->flags |= VLIB_BUFFER_TOTAL_LENGTH_VALID;
-  b0->total_length_not_including_first_buffer = 0;
+  session_queue_trace_t *t;
+  vlib_trace_buffer (vm, node, next_index, b, 1 /* follow_chain */ );
+  vlib_set_trace_count (vm, node, --*n_trace);
+  t = vlib_add_trace (vm, node, b, sizeof (*t));
+  t->session_index = s->session_index;
+  t->server_thread_index = s->thread_index;
+}
 
-  chain_bi0 = bi0;
-  chain_b0 = b0;
+always_inline void
+session_tx_fifo_chain_tail (vlib_main_t * vm, session_tx_context_t * ctx,
+			    vlib_buffer_t * b, u16 * n_bufs, u8 peek_data)
+{
+  session_manager_main_t *smm = &session_manager_main;
+  vlib_buffer_t *chain_b, *prev_b;
+  u32 chain_bi0, to_deq, left_from_seg;
+  u16 len_to_deq, n_bytes_read;
+  u8 *data, j;
+
+  b->flags |= VLIB_BUFFER_TOTAL_LENGTH_VALID;
+  b->total_length_not_including_first_buffer = 0;
+
+  chain_b = b;
+  left_from_seg = clib_min (ctx->snd_mss - b->current_length,
+			    ctx->left_to_snd);
   to_deq = left_from_seg;
-  for (j = 1; j < n_bufs_per_seg; j++)
+  for (j = 1; j < ctx->n_bufs_per_seg; j++)
     {
-      prev_b0 = chain_b0;
-      len_to_deq0 = clib_min (to_deq, deq_per_buf);
+      prev_b = chain_b;
+      len_to_deq = clib_min (to_deq, ctx->deq_per_buf);
 
       *n_bufs -= 1;
-      chain_bi0 = smm->tx_buffers[thread_index][*n_bufs];
-      _vec_len (smm->tx_buffers[thread_index]) = *n_bufs;
+      chain_bi0 = smm->tx_buffers[ctx->s->thread_index][*n_bufs];
+      _vec_len (smm->tx_buffers[ctx->s->thread_index]) = *n_bufs;
 
-      chain_b0 = vlib_get_buffer (vm, chain_bi0);
-      chain_b0->current_data = 0;
-      data0 = vlib_buffer_get_current (chain_b0);
+      chain_b = vlib_get_buffer (vm, chain_bi0);
+      chain_b->current_data = 0;
+      data = vlib_buffer_get_current (chain_b);
       if (peek_data)
 	{
-	  n_bytes_read = svm_fifo_peek (fifo, *tx_offset, len_to_deq0, data0);
-	  *tx_offset += n_bytes_read;
+	  n_bytes_read = svm_fifo_peek (ctx->s->server_tx_fifo,
+					ctx->tx_offset, len_to_deq, data);
+	  ctx->tx_offset += n_bytes_read;
 	}
       else
 	{
-	  if (tx_type == TRANSPORT_TX_DGRAM)
+	  if (ctx->transport_vft->tx_type == TRANSPORT_TX_DGRAM)
 	    {
-	      session_dgram_hdr_t *hdr;
+	      svm_fifo_t *f = ctx->s->server_tx_fifo;
+	      session_dgram_hdr_t *hdr = &ctx->hdr;
 	      u16 deq_now;
-	      hdr = (session_dgram_hdr_t *) svm_fifo_head (fifo);
 	      deq_now = clib_min (hdr->data_length - hdr->data_offset,
-				  len_to_deq0);
-	      n_bytes_read = svm_fifo_peek (fifo, hdr->data_offset, deq_now,
-					    data0);
+				  len_to_deq);
+	      n_bytes_read = svm_fifo_peek (f, hdr->data_offset, deq_now,
+					    data);
 	      ASSERT (n_bytes_read > 0);
 
 	      hdr->data_offset += n_bytes_read;
 	      if (hdr->data_offset == hdr->data_length)
-		svm_fifo_dequeue_drop (fifo, hdr->data_length);
+		svm_fifo_dequeue_drop (f, hdr->data_length);
 	    }
 	  else
-	    n_bytes_read = svm_fifo_dequeue_nowait (fifo, len_to_deq0, data0);
+	    n_bytes_read = svm_fifo_dequeue_nowait (ctx->s->server_tx_fifo,
+						    len_to_deq, data);
 	}
-      ASSERT (n_bytes_read == len_to_deq0);
-      chain_b0->current_length = n_bytes_read;
-      b0->total_length_not_including_first_buffer += chain_b0->current_length;
+      ASSERT (n_bytes_read == len_to_deq);
+      chain_b->current_length = n_bytes_read;
+      b->total_length_not_including_first_buffer += chain_b->current_length;
 
       /* update previous buffer */
-      prev_b0->next_buffer = chain_bi0;
-      prev_b0->flags |= VLIB_BUFFER_NEXT_PRESENT;
+      prev_b->next_buffer = chain_bi0;
+      prev_b->flags |= VLIB_BUFFER_NEXT_PRESENT;
 
       /* update current buffer */
-      chain_b0->next_buffer = 0;
+      chain_b->next_buffer = 0;
 
       to_deq -= n_bytes_read;
       if (to_deq == 0)
 	break;
     }
   ASSERT (to_deq == 0
-	  && b0->total_length_not_including_first_buffer == left_from_seg);
-  *left_to_snd0 -= left_from_seg;
+	  && b->total_length_not_including_first_buffer == left_from_seg);
+  ctx->left_to_snd -= left_from_seg;
 }
 
 always_inline int
-session_tx_fifo_read_and_snd_i (vlib_main_t * vm, vlib_node_runtime_t * node,
+session_output_try_get_buffers (vlib_main_t * vm,
 				session_manager_main_t * smm,
-				session_fifo_event_t * e0,
-				stream_session_t * s0, u32 thread_index,
-				int *n_tx_packets, u8 peek_data)
+				u32 thread_index, u16 * n_bufs, u32 wanted)
 {
-  u32 n_trace = vlib_get_trace_count (vm, node);
-  u32 left_to_snd0, max_len_to_snd0, len_to_deq0, snd_space0;
-  u32 n_bufs_per_evt, n_frames_per_evt, n_bufs_per_frame;
-  transport_connection_t *tc0;
-  transport_proto_vft_t *transport_vft;
-  transport_proto_t tp;
-  u32 next_index, next0, *to_next, n_left_to_next, bi0;
-  vlib_buffer_t *b0;
-  u32 tx_offset = 0, max_dequeue0, n_bytes_per_seg, left_for_seg;
-  u16 snd_mss0, n_bufs_per_seg, n_bufs;
+  u32 bufs_alloc = 0, bufs_now;
+  vec_validate_aligned (smm->tx_buffers[thread_index], *n_bufs + wanted - 1,
+			CLIB_CACHE_LINE_BYTES);
+  do
+    {
+      bufs_now =
+	vlib_buffer_alloc (vm,
+			   &smm->tx_buffers[thread_index][*n_bufs +
+							  bufs_alloc],
+			   wanted - bufs_alloc);
+      bufs_alloc += bufs_now;
+    }
+  while (bufs_now > 0 && ((bufs_alloc + *n_bufs < wanted)));
+
+  *n_bufs += bufs_alloc;
+  _vec_len (smm->tx_buffers[thread_index]) = *n_bufs;
+  return bufs_alloc;
+}
+
+always_inline void
+session_tx_fill_buffer (vlib_main_t * vm, session_tx_context_t * ctx,
+			vlib_buffer_t * b, u16 * n_bufs, u8 peek_data)
+{
+  u32 len_to_deq;
   u8 *data0;
-  int i, n_bytes_read;
-  u32 n_bytes_per_buf, deq_per_buf, deq_per_first_buf;
-  u32 bufs_alloc, bufs_now;
-  session_dgram_hdr_t hdr;
-
-  next_index = next0 = smm->session_type_to_next[s0->session_type];
-  tp = session_get_transport_proto (s0);
-  transport_vft = transport_protocol_get_vft (tp);
+  int n_bytes_read;
+
+  /*
+   * Start with the first buffer in chain
+   */
+  b->error = 0;
+  b->flags = VNET_BUFFER_F_LOCALLY_ORIGINATED;
+  b->current_data = 0;
+  b->total_length_not_including_first_buffer = 0;
+
+  data0 = vlib_buffer_make_headroom (b, MAX_HDRS_LEN);
+  len_to_deq = clib_min (ctx->left_to_snd, ctx->deq_per_first_buf);
+
   if (peek_data)
     {
-      if (PREDICT_FALSE (s0->session_state < SESSION_STATE_READY))
-	{
-	  /* Can retransmit for closed sessions but can't send new data if
-	   * session is not ready or closed */
-	  vec_add1 (smm->pending_event_vector[thread_index], *e0);
-	  return 0;
-	}
-      tc0 =
-	transport_vft->get_connection (s0->connection_index, thread_index);
+      n_bytes_read = svm_fifo_peek (ctx->s->server_tx_fifo, ctx->tx_offset,
+				    len_to_deq, data0);
+      ASSERT (n_bytes_read > 0);
+      /* Keep track of progress locally, transport is also supposed to
+       * increment it independently when pushing the header */
+      ctx->tx_offset += n_bytes_read;
     }
   else
     {
-      if (s0->session_state == SESSION_STATE_LISTENING)
-	tc0 = transport_vft->get_listener (s0->connection_index);
+      if (ctx->transport_vft->tx_type == TRANSPORT_TX_DGRAM)
+	{
+	  session_dgram_hdr_t *hdr = &ctx->hdr;
+	  svm_fifo_t *f = ctx->s->server_tx_fifo;
+	  u16 deq_now;
+	  u32 offset;
+
+	  ASSERT (hdr->data_length > hdr->data_offset);
+	  deq_now = clib_min (hdr->data_length - hdr->data_offset,
+			      len_to_deq);
+	  offset = hdr->data_offset + SESSION_CONN_HDR_LEN;
+	  n_bytes_read = svm_fifo_peek (f, offset, deq_now, data0);
+	  ASSERT (n_bytes_read > 0);
+
+	  if (ctx->s->session_state == SESSION_STATE_LISTENING)
+	    {
+	      ip_copy (&ctx->tc->rmt_ip, &hdr->rmt_ip, ctx->tc->is_ip4);
+	      ctx->tc->rmt_port = hdr->rmt_port;
+	    }
+	  hdr->data_offset += n_bytes_read;
+	  if (hdr->data_offset == hdr->data_length)
+	    {
+	      offset = hdr->data_length + SESSION_CONN_HDR_LEN;
+	      svm_fifo_dequeue_drop (f, offset);
+	    }
+	}
       else
 	{
-	  if (PREDICT_FALSE (s0->session_state == SESSION_STATE_CLOSED))
-	    return 0;
-	  tc0 = transport_vft->get_connection (s0->connection_index,
-					       thread_index);
+	  n_bytes_read = svm_fifo_dequeue_nowait (ctx->s->server_tx_fifo,
+						  len_to_deq, data0);
+	  ASSERT (n_bytes_read > 0);
 	}
     }
+  b->current_length = n_bytes_read;
+  ctx->left_to_snd -= n_bytes_read;
 
-  /* Make sure we have space to send and there's something to dequeue */
-  snd_mss0 = transport_vft->send_mss (tc0);
-  snd_space0 = transport_vft->send_space (tc0);
+  /*
+   * Fill in the remaining buffers in the chain, if any
+   */
+  if (PREDICT_FALSE (ctx->n_bufs_per_seg > 1 && ctx->left_to_snd))
+    session_tx_fifo_chain_tail (vm, ctx, b, n_bufs, peek_data);
+
+  /* *INDENT-OFF* */
+  SESSION_EVT_DBG(SESSION_EVT_DEQ, s, ({
+	ed->data[0] = e->event_type;
+	ed->data[1] = max_dequeue;
+	ed->data[2] = len_to_deq;
+	ed->data[3] = left_to_snd;
+  }));
+  /* *INDENT-ON* */
+}
 
-  /* Can't make any progress */
-  if (snd_space0 == 0 || snd_mss0 == 0)
+always_inline u8
+session_tx_not_ready (stream_session_t * s, u8 peek_data)
+{
+  if (peek_data)
     {
-      vec_add1 (smm->pending_event_vector[thread_index], *e0);
-      return 0;
+      /* Can retransmit for closed sessions but can't send new data if
+       * session is not ready or closed */
+      if (s->session_state < SESSION_STATE_READY)
+	return 1;
     }
+  return 0;
+}
 
-  /* Allow enqueuing of a new event */
-  svm_fifo_unset_event (s0->server_tx_fifo);
+always_inline transport_connection_t *
+session_tx_get_transport (session_tx_context_t * ctx, u8 peek_data)
+{
+  if (peek_data)
+    {
+      return ctx->transport_vft->get_connection (ctx->s->connection_index,
+						 ctx->s->thread_index);
+    }
+  else
+    {
+      if (ctx->s->session_state == SESSION_STATE_LISTENING)
+	return ctx->transport_vft->get_listener (ctx->s->connection_index);
+      else
+	{
+	  return ctx->transport_vft->get_connection (ctx->s->connection_index,
+						     ctx->s->thread_index);
+	}
+    }
+}
 
-  /* Check how much we can pull. */
-  max_dequeue0 = svm_fifo_max_dequeue (s0->server_tx_fifo);
+always_inline void
+session_tx_set_dequeue_params (vlib_main_t * vm, session_tx_context_t * ctx,
+			       u8 peek_data)
+{
+  u32 n_bytes_per_buf, n_bytes_per_seg;
+  ctx->max_dequeue = svm_fifo_max_dequeue (ctx->s->server_tx_fifo);
   if (peek_data)
     {
       /* Offset in rx fifo from where to peek data */
-      tx_offset = transport_vft->tx_fifo_offset (tc0);
-      if (PREDICT_FALSE (tx_offset >= max_dequeue0))
-	return 0;
-      max_dequeue0 -= tx_offset;
+      ctx->tx_offset = ctx->transport_vft->tx_fifo_offset (ctx->tc);
+      if (PREDICT_FALSE (ctx->tx_offset >= ctx->max_dequeue))
+	{
+	  ctx->max_len_to_snd = 0;
+	  return;
+	}
+      ctx->max_dequeue -= ctx->tx_offset;
     }
   else
     {
-      if (transport_vft->tx_type == TRANSPORT_TX_DGRAM)
+      if (ctx->transport_vft->tx_type == TRANSPORT_TX_DGRAM)
 	{
-	  if (max_dequeue0 < sizeof (hdr))
-	    return 0;
-	  svm_fifo_peek (s0->server_tx_fifo, 0, sizeof (hdr), (u8 *) & hdr);
-	  ASSERT (hdr.data_length > hdr.data_offset);
-	  max_dequeue0 = hdr.data_length - hdr.data_offset;
+	  if (ctx->max_dequeue <= sizeof (ctx->hdr))
+	    {
+	      ctx->max_len_to_snd = 0;
+	      return;
+	    }
+	  svm_fifo_peek (ctx->s->server_tx_fifo, 0, sizeof (ctx->hdr),
+			 (u8 *) & ctx->hdr);
+	  ASSERT (ctx->hdr.data_length > ctx->hdr.data_offset);
+	  ctx->max_dequeue = ctx->hdr.data_length - ctx->hdr.data_offset;
 	}
     }
-  ASSERT (max_dequeue0 > 0);
+  ASSERT (ctx->max_dequeue > 0);
 
   /* Ensure we're not writing more than transport window allows */
-  if (max_dequeue0 < snd_space0)
+  if (ctx->max_dequeue < ctx->snd_space)
     {
       /* 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;
+      ctx->max_len_to_snd =
+	(ctx->max_dequeue > ctx->snd_mss) ?
+	ctx->max_dequeue - ctx->max_dequeue % ctx->snd_mss : ctx->max_dequeue;
       /* TODO Nagle ? */
     }
   else
     {
       /* Expectation is that snd_space0 is already a multiple of snd_mss */
-      max_len_to_snd0 = snd_space0;
+      ctx->max_len_to_snd = ctx->snd_space;
     }
 
-  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);
   ASSERT (n_bytes_per_buf > MAX_HDRS_LEN);
-  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);
-  n_frames_per_evt = ceil ((double) n_bufs_per_evt / VLIB_FRAME_SIZE);
-  n_bufs_per_frame = n_bufs_per_seg * VLIB_FRAME_SIZE;
+  n_bytes_per_seg = MAX_HDRS_LEN + ctx->snd_mss;
+  ctx->n_bufs_per_seg = ceil ((double) n_bytes_per_seg / n_bytes_per_buf);
+  ctx->deq_per_buf = clib_min (ctx->snd_mss, n_bytes_per_buf);
+  ctx->deq_per_first_buf = clib_min (ctx->snd_mss,
+				     n_bytes_per_buf - MAX_HDRS_LEN);
+}
+
+always_inline int
+session_tx_fifo_read_and_snd_i (vlib_main_t * vm, vlib_node_runtime_t * node,
+				session_fifo_event_t * e,
+				stream_session_t * s, int *n_tx_packets,
+				u8 peek_data)
+{
+  u32 next_index, next0, next1, next2, next3, *to_next, n_left_to_next;
+  u32 n_trace = vlib_get_trace_count (vm, node), n_packets = 0, pbi;
+  u32 n_bufs_per_frame, thread_index = s->thread_index;
+  session_manager_main_t *smm = &session_manager_main;
+  session_tx_context_t *ctx = &smm->ctx[thread_index];
+  transport_proto_t tp;
+  vlib_buffer_t *pb;
+  u16 n_bufs;
 
-  deq_per_buf = clib_min (snd_mss0, n_bytes_per_buf);
-  deq_per_first_buf = clib_min (snd_mss0, n_bytes_per_buf - MAX_HDRS_LEN);
+  if (PREDICT_FALSE (session_tx_not_ready (s, peek_data)))
+    {
+      vec_add1 (smm->pending_event_vector[thread_index], *e);
+      return 0;
+    }
+
+  next_index = smm->session_type_to_next[s->session_type];
+  next0 = next1 = next2 = next3 = next_index;
+
+  tp = session_get_transport_proto (s);
+  ctx->s = s;
+  ctx->transport_vft = transport_protocol_get_vft (tp);
+  ctx->tc = session_tx_get_transport (ctx, peek_data);
+  ctx->snd_mss = ctx->transport_vft->send_mss (ctx->tc);
+  ctx->snd_space = ctx->transport_vft->send_space (ctx->tc);
+  if (ctx->snd_space == 0 || ctx->snd_mss == 0)
+    {
+      vec_add1 (smm->pending_event_vector[thread_index], *e);
+      return 0;
+    }
+
+  /* Allow enqueuing of a new event */
+  svm_fifo_unset_event (s->server_tx_fifo);
+
+  /* Check how much we can pull. */
+  session_tx_set_dequeue_params (vm, ctx, peek_data);
+  if (PREDICT_FALSE (!ctx->max_len_to_snd))
+    return 0;
 
   n_bufs = vec_len (smm->tx_buffers[thread_index]);
-  left_to_snd0 = max_len_to_snd0;
-  for (i = 0; i < n_frames_per_evt; i++)
+  ctx->left_to_snd = ctx->max_len_to_snd;
+
+  /*
+   * Make sure we have at least one full frame of buffers ready
+   */
+  n_bufs_per_frame = ctx->n_bufs_per_seg * VLIB_FRAME_SIZE;
+  if (n_bufs < n_bufs_per_frame)
     {
-      /* Make sure we have at least one full frame of buffers ready */
+      session_output_try_get_buffers (vm, smm, thread_index, &n_bufs,
+				      n_bufs_per_frame);
       if (PREDICT_FALSE (n_bufs < n_bufs_per_frame))
 	{
-	  vec_validate (smm->tx_buffers[thread_index],
-			n_bufs + n_bufs_per_frame - 1);
-	  bufs_alloc = 0;
-	  do
-	    {
-	      bufs_now =
-		vlib_buffer_alloc (vm,
-				   &smm->tx_buffers[thread_index][n_bufs +
-								  bufs_alloc],
-				   n_bufs_per_frame - bufs_alloc);
-	      bufs_alloc += bufs_now;
-	    }
-	  while (bufs_now > 0 && ((bufs_alloc + n_bufs < n_bufs_per_frame)));
-
-	  n_bufs += bufs_alloc;
-	  _vec_len (smm->tx_buffers[thread_index]) = n_bufs;
-
-	  if (PREDICT_FALSE (n_bufs < n_bufs_per_frame))
-	    {
-	      vec_add1 (smm->pending_event_vector[thread_index], *e0);
-	      return -1;
-	    }
-	  ASSERT (n_bufs >= n_bufs_per_frame);
+	  vec_add1 (smm->pending_event_vector[thread_index], *e);
+	  return -1;
 	}
+    }
 
-      vlib_get_next_frame (vm, node, next_index, to_next, n_left_to_next);
-      while (left_to_snd0 && n_left_to_next)
+  /*
+   * Write until we fill up a frame
+   */
+  vlib_get_next_frame (vm, node, next_index, to_next, n_left_to_next);
+  while (ctx->left_to_snd && n_left_to_next)
+    {
+      while (ctx->left_to_snd > 3 * ctx->snd_mss && n_left_to_next >= 4)
 	{
-	  /*
-	   * Handle first buffer in chain separately
-	   */
+	  vlib_buffer_t *b0, *b1;
+	  u32 bi0, bi1;
 
-	  len_to_deq0 = clib_min (left_to_snd0, deq_per_first_buf);
-	  if (left_to_snd0 > len_to_deq0 && n_left_to_next > 1)
-	    {
-	      vlib_buffer_t *pb;
-	      u32 pbi = smm->tx_buffers[thread_index][n_bufs - 2];
-	      pb = vlib_get_buffer (vm, pbi);
-	      vlib_prefetch_buffer_header (pb, LOAD);
-	    }
+	  pbi = smm->tx_buffers[thread_index][n_bufs - 3];
+	  pb = vlib_get_buffer (vm, pbi);
+	  vlib_prefetch_buffer_header (pb, STORE);
+	  pbi = smm->tx_buffers[thread_index][n_bufs - 4];
+	  pb = vlib_get_buffer (vm, pbi);
+	  vlib_prefetch_buffer_header (pb, STORE);
 
-	  /* Get free buffer */
-	  ASSERT (n_bufs >= 1);
-	  bi0 = smm->tx_buffers[thread_index][--n_bufs];
-	  _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;
+	  to_next[0] = bi0 = smm->tx_buffers[thread_index][--n_bufs];
+	  to_next[1] = bi1 = smm->tx_buffers[thread_index][--n_bufs];
 
 	  b0 = vlib_get_buffer (vm, bi0);
-	  b0->error = 0;
-	  b0->flags = VNET_BUFFER_F_LOCALLY_ORIGINATED;
-	  b0->current_data = 0;
-	  b0->total_length_not_including_first_buffer = 0;
+	  b1 = vlib_get_buffer (vm, bi1);
 
-	  data0 = vlib_buffer_make_headroom (b0, MAX_HDRS_LEN);
-	  if (peek_data)
-	    {
-	      n_bytes_read = svm_fifo_peek (s0->server_tx_fifo, tx_offset,
-					    len_to_deq0, data0);
-	      if (n_bytes_read <= 0)
-		goto dequeue_fail;
-	      /* Keep track of progress locally, transport is also supposed to
-	       * increment it independently when pushing the header */
-	      tx_offset += n_bytes_read;
-	    }
-	  else
-	    {
-	      if (transport_vft->tx_type == TRANSPORT_TX_DGRAM)
-		{
-		  svm_fifo_t *f = s0->server_tx_fifo;
-		  u16 deq_now;
-		  u32 offset;
-
-		  ASSERT (hdr.data_length > hdr.data_offset);
-		  deq_now = clib_min (hdr.data_length - hdr.data_offset,
-				      len_to_deq0);
-		  offset = hdr.data_offset + SESSION_CONN_HDR_LEN;
-		  n_bytes_read = svm_fifo_peek (f, offset, deq_now, data0);
-		  if (PREDICT_FALSE (n_bytes_read <= 0))
-		    goto dequeue_fail;
-
-		  if (s0->session_state == SESSION_STATE_LISTENING)
-		    {
-		      ip_copy (&tc0->rmt_ip, &hdr.rmt_ip, tc0->is_ip4);
-		      tc0->rmt_port = hdr.rmt_port;
-		    }
-		  hdr.data_offset += n_bytes_read;
-		  if (hdr.data_offset == hdr.data_length)
-		    {
-		      offset = hdr.data_length + SESSION_CONN_HDR_LEN;
-		      svm_fifo_dequeue_drop (f, offset);
-		    }
-		}
-	      else
-		{
-		  n_bytes_read = svm_fifo_dequeue_nowait (s0->server_tx_fifo,
-							  len_to_deq0, data0);
-		  if (n_bytes_read <= 0)
-		    goto dequeue_fail;
-		}
-	    }
+	  session_tx_fill_buffer (vm, ctx, b0, &n_bufs, peek_data);
+	  session_tx_fill_buffer (vm, ctx, b1, &n_bufs, peek_data);
 
-	  b0->current_length = n_bytes_read;
-	  left_to_snd0 -= n_bytes_read;
-	  *n_tx_packets = *n_tx_packets + 1;
+	  ctx->transport_vft->push_header (ctx->tc, b0);
+	  ctx->transport_vft->push_header (ctx->tc, b1);
 
-	  /*
-	   * Fill in the remaining buffers in the chain, if any
-	   */
-	  if (PREDICT_FALSE (n_bufs_per_seg > 1 && left_to_snd0))
+	  to_next += 2;
+	  n_left_to_next -= 2;
+	  n_packets += 2;
+
+	  VLIB_BUFFER_TRACE_TRAJECTORY_INIT (b0);
+	  VLIB_BUFFER_TRACE_TRAJECTORY_INIT (b1);
+
+	  if (PREDICT_FALSE (n_trace > 0))
 	    {
-	      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, &tx_offset,
-					  deq_per_buf, peek_data,
-					  transport_vft->tx_type);
+	      session_tx_trace_buffer (vm, node, next_index, b0, s, &n_trace);
+	      if (n_trace)
+		session_tx_trace_buffer (vm, node, next_index, b1, s,
+					 &n_trace);
 	    }
 
+	  vlib_validate_buffer_enqueue_x2 (vm, node, next_index, to_next,
+					   n_left_to_next, bi0, bi1,
+					   next0, next1);
+	}
+      while (ctx->left_to_snd && n_left_to_next)
+	{
+	  vlib_buffer_t *b0;
+	  u32 bi0;
+
+	  ASSERT (n_bufs >= 1);
+	  to_next[0] = bi0 = smm->tx_buffers[thread_index][--n_bufs];
+	  b0 = vlib_get_buffer (vm, bi0);
+	  session_tx_fill_buffer (vm, ctx, b0, &n_bufs, peek_data);
+
 	  /* Ask transport to push header after current_length and
 	   * total_length_not_including_first_buffer are updated */
-	  transport_vft->push_header (tc0, b0);
+	  ctx->transport_vft->push_header (ctx->tc, b0);
 
-	  /* *INDENT-OFF* */
-	  SESSION_EVT_DBG(SESSION_EVT_DEQ, s0, ({
-	      ed->data[0] = e0->event_type;
-	      ed->data[1] = max_dequeue0;
-	      ed->data[2] = len_to_deq0;
-	      ed->data[3] = left_to_snd0;
-	  }));
-	  /* *INDENT-ON* */
+	  to_next += 1;
+	  n_left_to_next -= 1;
+	  n_packets += 1;
 
 	  VLIB_BUFFER_TRACE_TRAJECTORY_INIT (b0);
-
 	  if (PREDICT_FALSE (n_trace > 0))
-	    {
-	      session_queue_trace_t *t0;
-	      vlib_trace_buffer (vm, node, next_index, b0,
-				 1 /* follow_chain */ );
-	      vlib_set_trace_count (vm, node, --n_trace);
-	      t0 = vlib_add_trace (vm, node, b0, sizeof (*t0));
-	      t0->session_index = s0->session_index;
-	      t0->server_thread_index = s0->thread_index;
-	    }
+	    session_tx_trace_buffer (vm, node, next_index, b0, s, &n_trace);
 
-	  vlib_validate_buffer_enqueue_x1 (vm, node, next_index,
-					   to_next, n_left_to_next,
-					   bi0, next0);
+	  vlib_validate_buffer_enqueue_x1 (vm, node, next_index, to_next,
+					   n_left_to_next, bi0, next0);
 	}
-      vlib_put_next_frame (vm, node, next_index, n_left_to_next);
     }
+  _vec_len (smm->tx_buffers[thread_index]) = n_bufs;
+  *n_tx_packets += n_packets;
+  vlib_put_next_frame (vm, node, next_index, n_left_to_next);
 
   /* If we couldn't dequeue all bytes mark as partially read */
-  if (max_len_to_snd0 < max_dequeue0)
-    if (svm_fifo_set_event (s0->server_tx_fifo))
-      vec_add1 (smm->pending_event_vector[thread_index], *e0);
+  if (ctx->max_len_to_snd < ctx->max_dequeue)
+    if (svm_fifo_set_event (s->server_tx_fifo))
+      vec_add1 (smm->pending_event_vector[thread_index], *e);
 
-  if (!peek_data && transport_vft->tx_type == TRANSPORT_TX_DGRAM)
+  if (!peek_data && ctx->transport_vft->tx_type == TRANSPORT_TX_DGRAM)
     {
       /* Fix dgram pre header */
-      if (max_len_to_snd0 < max_dequeue0)
-	svm_fifo_overwrite_head (s0->server_tx_fifo, (u8 *) & hdr,
+      if (ctx->max_len_to_snd < ctx->max_dequeue)
+	svm_fifo_overwrite_head (s->server_tx_fifo, (u8 *) & ctx->hdr,
 				 sizeof (session_dgram_pre_hdr_t));
       /* More data needs to be read */
-      else if (svm_fifo_max_dequeue (s0->server_tx_fifo) > 0)
-	vec_add1 (smm->pending_event_vector[thread_index], *e0);
+      else if (svm_fifo_max_dequeue (s->server_tx_fifo) > 0)
+	if (svm_fifo_set_event (s->server_tx_fifo))
+	  vec_add1 (smm->pending_event_vector[thread_index], *e);
     }
   return 0;
-
-dequeue_fail:
-  /*
-   * 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");
-  if (svm_fifo_set_event (s0->server_tx_fifo))
-    {
-      vec_add1 (smm->pending_event_vector[thread_index], *e0);
-    }
-  vlib_put_next_frame (vm, node, next_index, n_left_to_next + 1);
-  _vec_len (smm->tx_buffers[thread_index]) += 1;
-
-  return 0;
 }
 
 int
 session_tx_fifo_peek_and_snd (vlib_main_t * vm, vlib_node_runtime_t * node,
-			      session_manager_main_t * smm,
 			      session_fifo_event_t * e0,
-			      stream_session_t * s0, u32 thread_index,
-			      int *n_tx_pkts)
+			      stream_session_t * s0, int *n_tx_pkts)
 {
-  return session_tx_fifo_read_and_snd_i (vm, node, smm, e0, s0, thread_index,
-					 n_tx_pkts, 1);
+  return session_tx_fifo_read_and_snd_i (vm, node, e0, s0, n_tx_pkts, 1);
 }
 
 int
 session_tx_fifo_dequeue_and_snd (vlib_main_t * vm, vlib_node_runtime_t * node,
-				 session_manager_main_t * smm,
 				 session_fifo_event_t * e0,
-				 stream_session_t * s0, u32 thread_index,
-				 int *n_tx_pkts)
+				 stream_session_t * s0, int *n_tx_pkts)
 {
-  return session_tx_fifo_read_and_snd_i (vm, node, smm, e0, s0, thread_index,
-					 n_tx_pkts, 0);
+  return session_tx_fifo_read_and_snd_i (vm, node, e0, s0, n_tx_pkts, 0);
 }
 
 int
 session_tx_fifo_dequeue_internal (vlib_main_t * vm,
 				  vlib_node_runtime_t * node,
-				  session_manager_main_t * smm,
 				  session_fifo_event_t * e0,
-				  stream_session_t * s0, u32 thread_index,
-				  int *n_tx_pkts)
+				  stream_session_t * s0, int *n_tx_pkts)
 {
   application_t *app;
   app = application_get (s0->opaque);
@@ -719,8 +773,7 @@ skip_dequeue:
 	    }
 	  /* 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,
-							my_thread_index,
+	  rv = (smm->session_tx_fns[s0->session_type]) (vm, node, e0, s0,
 							&n_tx_packets);
 	  /* Out of buffers */
 	  if (PREDICT_FALSE (rv < 0))
@@ -732,13 +785,13 @@ skip_dequeue:
 	  break;
 	case FIFO_EVENT_DISCONNECT:
 	  /* Make sure disconnects run after the pending list is drained */
-	  if (!e0->postponed)
+	  s0 = session_get_from_handle (e0->session_handle);
+	  if (!e0->postponed || svm_fifo_max_dequeue (s0->server_tx_fifo))
 	    {
 	      e0->postponed = 1;
 	      vec_add1 (smm->pending_disconnects[my_thread_index], *e0);
 	      continue;
 	    }
-	  s0 = session_get_from_handle (e0->session_handle);
 	  stream_session_disconnect_transport (s0);
 	  break;
 	case FIFO_EVENT_BUILTIN_RX: