15 files changed, 375 insertions, 120 deletions

diff --git a/src/plugins/dpdk/device/cli.c b/src/plugins/dpdk/device/cli.c
index aeeb772d4fb..fe1c41c279b 100644
--- a/src/plugins/dpdk/device/cli.c
+++ b/src/plugins/dpdk/device/cli.c
@@ -357,6 +357,7 @@ show_dpdk_buffer (vlib_main_t * vm, unformat_input_t * input,
 			   "name=\"%s\"  available = %7d allocated = %7d total = %7d\n",
 			   rmp->name, (u32) count, (u32) free_count,
 			   (u32) (count + free_count));
+	  rte_mempool_dump (stderr, rmp);
 	}
       else
 	{
diff --git a/src/vlib/error.c b/src/vlib/error.c
index e4ed4ee37ee..dec90bbe440 100644
--- a/src/vlib/error.c
+++ b/src/vlib/error.c
@@ -280,7 +280,7 @@ show_errors (vlib_main_t * vm,
 }
 
 /* *INDENT-OFF* */
-VLIB_CLI_COMMAND (cli_show_errors, static) = {
+VLIB_CLI_COMMAND (vlib_cli_show_errors) = {
   .path = "show errors",
   .short_help = "Show error counts",
   .function = show_errors,
diff --git a/src/vlib/node_funcs.h b/src/vlib/node_funcs.h
index c0389b2fab8..c4c06454228 100644
--- a/src/vlib/node_funcs.h
+++ b/src/vlib/node_funcs.h
@@ -410,19 +410,21 @@ vlib_frame_t *vlib_get_frame_to_node (vlib_main_t * vm, u32 to_node_index);
 void vlib_put_frame_to_node (vlib_main_t * vm, u32 to_node_index,
 			     vlib_frame_t * f);
 
-always_inline vlib_process_t *
-vlib_get_current_process (vlib_main_t * vm)
-{
-  vlib_node_main_t *nm = &vm->node_main;
-  return vec_elt (nm->processes, nm->current_process_index);
-}
-
 always_inline uword
 vlib_in_process_context (vlib_main_t * vm)
 {
   return vm->node_main.current_process_index != ~0;
 }
 
+always_inline vlib_process_t *
+vlib_get_current_process (vlib_main_t * vm)
+{
+  vlib_node_main_t *nm = &vm->node_main;
+  if (vlib_in_process_context (vm))
+    return vec_elt (nm->processes, nm->current_process_index);
+  return 0;
+}
+
 always_inline uword
 vlib_current_process (vlib_main_t * vm)
 {
diff --git a/src/vnet/ip/ip4.h b/src/vnet/ip/ip4.h
index 8f9a8e27bf5..74faa05947e 100644
--- a/src/vnet/ip/ip4.h
+++ b/src/vnet/ip/ip4.h
@@ -354,8 +354,6 @@ vlib_buffer_push_ip4 (vlib_main_t * vm, vlib_buffer_t * b,
       ih->checksum = 0;
       b->flags |= VNET_BUFFER_F_OFFLOAD_IP_CKSUM | VNET_BUFFER_F_IS_IP4;
       vnet_buffer (b)->l3_hdr_offset = (u8 *) ih - b->data;
-      vnet_buffer (b)->l4_hdr_offset = vnet_buffer (b)->l3_hdr_offset +
-	sizeof (*ih);
     }
   else
     ih->checksum = ip4_header_checksum (ih);
diff --git a/src/vnet/session/session.c b/src/vnet/session/session.c
index 004c7193b94..4ba152917ac 100644
--- a/src/vnet/session/session.c
+++ b/src/vnet/session/session.c
@@ -759,6 +759,7 @@ session_manager_main_enable (vlib_main_t * vm)
   session_manager_main_t *smm = &session_manager_main;
   vlib_thread_main_t *vtm = vlib_get_thread_main ();
   u32 num_threads;
+  u32 preallocated_sessions_per_worker;
   int i;
 
   num_threads = 1 /* main thread */  + vtm->n_threads;
@@ -795,15 +796,35 @@ session_manager_main_enable (vlib_main_t * vm)
   for (i = 0; i < vec_len (smm->vpp_event_queues); i++)
     session_vpp_event_queue_allocate (smm, i);
 
-  /* $$$$ preallocate hack config parameter */
-  for (i = 0; i < smm->preallocated_sessions; i++)
+  /* Preallocate sessions */
+  if (num_threads == 1)
     {
-      stream_session_t *ss __attribute__ ((unused));
-      pool_get_aligned (smm->sessions[0], ss, CLIB_CACHE_LINE_BYTES);
+      for (i = 0; i < smm->preallocated_sessions; i++)
+	{
+	  stream_session_t *ss __attribute__ ((unused));
+	  pool_get_aligned (smm->sessions[0], ss, CLIB_CACHE_LINE_BYTES);
+	}
+
+      for (i = 0; i < smm->preallocated_sessions; i++)
+	pool_put_index (smm->sessions[0], i);
     }
+  else
+    {
+      int j;
+      preallocated_sessions_per_worker = smm->preallocated_sessions /
+	(num_threads - 1);
 
-  for (i = 0; i < smm->preallocated_sessions; i++)
-    pool_put_index (smm->sessions[0], i);
+      for (j = 1; j < num_threads; j++)
+	{
+	  for (i = 0; i < preallocated_sessions_per_worker; i++)
+	    {
+	      stream_session_t *ss __attribute__ ((unused));
+	      pool_get_aligned (smm->sessions[j], ss, CLIB_CACHE_LINE_BYTES);
+	    }
+	  for (i = 0; i < preallocated_sessions_per_worker; i++)
+	    pool_put_index (smm->sessions[j], i);
+	}
+    }
 
   session_lookup_init ();
 
@@ -863,6 +884,7 @@ session_config_fn (vlib_main_t * vm, unformat_input_t * input)
 {
   session_manager_main_t *smm = &session_manager_main;
   u32 nitems;
+  uword tmp;
 
   while (unformat_check_input (input) != UNFORMAT_END_OF_INPUT)
     {
@@ -873,9 +895,53 @@ session_config_fn (vlib_main_t * vm, unformat_input_t * input)
 	  else
 	    clib_warning ("event queue length %d too small, ignored", nitems);
 	}
-      if (unformat (input, "preallocated-sessions %d",
-		    &smm->preallocated_sessions))
+      else if (unformat (input, "preallocated-sessions %d",
+			 &smm->preallocated_sessions))
+	;
+      else if (unformat (input, "v4-session-table-buckets %d",
+			 &smm->configured_v4_session_table_buckets))
 	;
+      else if (unformat (input, "v4-halfopen-table-buckets %d",
+			 &smm->configured_v4_halfopen_table_buckets))
+	;
+      else if (unformat (input, "v6-session-table-buckets %d",
+			 &smm->configured_v6_session_table_buckets))
+	;
+      else if (unformat (input, "v6-halfopen-table-buckets %d",
+			 &smm->configured_v6_halfopen_table_buckets))
+	;
+      else if (unformat (input, "v4-session-table-memory %U",
+			 unformat_memory_size, &tmp))
+	{
+	  if (tmp >= 0x100000000)
+	    return clib_error_return (0, "memory size %llx (%lld) too large",
+				      tmp, tmp);
+	  smm->configured_v4_session_table_memory = tmp;
+	}
+      else if (unformat (input, "v4-halfopen-table-memory %U",
+			 unformat_memory_size, &tmp))
+	{
+	  if (tmp >= 0x100000000)
+	    return clib_error_return (0, "memory size %llx (%lld) too large",
+				      tmp, tmp);
+	  smm->configured_v4_halfopen_table_memory = tmp;
+	}
+      else if (unformat (input, "v6-session-table-memory %U",
+			 unformat_memory_size, &tmp))
+	{
+	  if (tmp >= 0x100000000)
+	    return clib_error_return (0, "memory size %llx (%lld) too large",
+				      tmp, tmp);
+	  smm->configured_v6_session_table_memory = tmp;
+	}
+      else if (unformat (input, "v6-halfopen-table-memory %U",
+			 unformat_memory_size, &tmp))
+	{
+	  if (tmp >= 0x100000000)
+	    return clib_error_return (0, "memory size %llx (%lld) too large",
+				      tmp, tmp);
+	  smm->configured_v6_halfopen_table_memory = tmp;
+	}
       else
 	return clib_error_return (0, "unknown input `%U'",
 				  format_unformat_error, input);
diff --git a/src/vnet/session/session.h b/src/vnet/session/session.h
index 180b9f8a496..538433da0f2 100644
--- a/src/vnet/session/session.h
+++ b/src/vnet/session/session.h
@@ -133,6 +133,16 @@ struct _session_manager_main
   /** vpp fifo event queue configured length */
   u32 configured_event_queue_length;
 
+  /** session table size parameters */
+  u32 configured_v4_session_table_buckets;
+  u32 configured_v4_session_table_memory;
+  u32 configured_v4_halfopen_table_buckets;
+  u32 configured_v4_halfopen_table_memory;
+  u32 configured_v6_session_table_buckets;
+  u32 configured_v6_session_table_memory;
+  u32 configured_v6_halfopen_table_buckets;
+  u32 configured_v6_halfopen_table_memory;
+
   /** Unique segment name counter */
   u32 unique_segment_name_counter;
 
diff --git a/src/vnet/session/session_cli.c b/src/vnet/session/session_cli.c
index de564ea7272..9f3d217cf61 100755
--- a/src/vnet/session/session_cli.c
+++ b/src/vnet/session/session_cli.c
@@ -312,7 +312,7 @@ show_session_command_fn (vlib_main_t * vm, unformat_input_t * input,
 }
 
 /* *INDENT-OFF* */
-VLIB_CLI_COMMAND (show_session_command, static) =
+VLIB_CLI_COMMAND (vlib_cli_show_session_command) =
 {
   .path = "show session",
   .short_help = "show session [verbose]",
diff --git a/src/vnet/session/session_lookup.c b/src/vnet/session/session_lookup.c
index 1ce22f808e5..41f9dbf0c1e 100644
--- a/src/vnet/session/session_lookup.c
+++ b/src/vnet/session/session_lookup.c
@@ -569,23 +569,45 @@ stream_session_lookup_transport6 (ip6_address_t * lcl, ip6_address_t * rmt,
   return 0;
 }
 
+#define foreach_hash_table_parameter            \
+  _(v4,session,buckets,20000)                   \
+  _(v4,session,memory,(64<<20))                 \
+  _(v6,session,buckets,20000)                   \
+  _(v6,session,memory,(64<<20))                 \
+  _(v4,halfopen,buckets,20000)                  \
+  _(v4,halfopen,memory,(64<<20))                \
+  _(v6,halfopen,buckets,20000)                  \
+  _(v6,halfopen,memory,(64<<20))
+
 void
 session_lookup_init (void)
 {
   session_lookup_t *sl = &session_lookup;
+
+#define _(af,table,parm,value) \
+  u32 configured_##af##_##table##_table_##parm = value;
+  foreach_hash_table_parameter;
+#undef _
+
+#define _(af,table,parm,value)                                          \
+  if (session_manager_main.configured_##af##_##table##_table_##parm)    \
+    configured_##af##_##table##_table_##parm =                          \
+      session_manager_main.configured_##af##_##table##_table_##parm;
+  foreach_hash_table_parameter;
+#undef _
+
   clib_bihash_init_16_8 (&sl->v4_session_hash, "v4 session table",
-			 200000 /* $$$$ config parameter nbuckets */ ,
-			 (64 << 20) /*$$$ config parameter table size */ );
+			 configured_v4_session_table_buckets,
+			 configured_v4_session_table_memory);
   clib_bihash_init_48_8 (&sl->v6_session_hash, "v6 session table",
-			 200000 /* $$$$ config parameter nbuckets */ ,
-			 (64 << 20) /*$$$ config parameter table size */ );
-
+			 configured_v6_session_table_buckets,
+			 configured_v6_session_table_memory);
   clib_bihash_init_16_8 (&sl->v4_half_open_hash, "v4 half-open table",
-			 200000 /* $$$$ config parameter nbuckets */ ,
-			 (64 << 20) /*$$$ config parameter table size */ );
+			 configured_v4_halfopen_table_buckets,
+			 configured_v4_halfopen_table_memory);
   clib_bihash_init_48_8 (&sl->v6_half_open_hash, "v6 half-open table",
-			 200000 /* $$$$ config parameter nbuckets */ ,
-			 (64 << 20) /*$$$ config parameter table size */ );
+			 configured_v6_halfopen_table_buckets,
+			 configured_v6_halfopen_table_memory);
 }
 
 /*
diff --git a/src/vnet/tcp/builtin_client.c b/src/vnet/tcp/builtin_client.c
index 27e20f8e8e5..48daffb41de 100644
--- a/src/vnet/tcp/builtin_client.c
+++ b/src/vnet/tcp/builtin_client.c
@@ -597,8 +597,9 @@ clients_connect (vlib_main_t * vm, u8 * uri, u32 n_clients)
       a->mp = 0;
       vnet_connect_uri (a);
 
-      /* Crude pacing for call setups, 100k/sec  */
-      vlib_process_suspend (vm, 10e-6);
+      /* Crude pacing for call setups  */
+      if ((i % 4) == 0)
+	vlib_process_suspend (vm, 10e-6);
     }
 }
 
@@ -612,8 +613,10 @@ test_tcp_clients_command_fn (vlib_main_t * vm,
   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;
+  f64 test_timeout = 20.0, syn_timeout = 20.0, delta;
+  f64 time_before_connects;
   u32 n_clients = 1;
+  int preallocate_sessions = 0;
   char *transfer_type;
   int i;
 
@@ -640,7 +643,9 @@ test_tcp_clients_command_fn (vlib_main_t * vm,
 	;
       else if (unformat (input, "uri %s", &tm->connect_uri))
 	;
-      else if (unformat (input, "cli-timeout %f", &cli_timeout))
+      else if (unformat (input, "test-timeout %f", &test_timeout))
+	;
+      else if (unformat (input, "syn-timeout %f", &syn_timeout))
 	;
       else if (unformat (input, "no-return"))
 	tm->no_return = 1;
@@ -657,6 +662,8 @@ test_tcp_clients_command_fn (vlib_main_t * vm,
 	tm->private_segment_size = tmp;
       else if (unformat (input, "preallocate-fifos"))
 	tm->prealloc_fifos = 1;
+      else if (unformat (input, "preallocate-sessions"))
+	preallocate_sessions = 1;
       else
 	if (unformat (input, "client-batch %d", &tm->connections_per_batch))
 	;
@@ -674,6 +681,7 @@ test_tcp_clients_command_fn (vlib_main_t * vm,
 	return clib_error_return (0, "failed init");
     }
 
+
   tm->ready_connections = 0;
   tm->expected_connections = n_clients;
   tm->rx_total = 0;
@@ -705,11 +713,21 @@ test_tcp_clients_command_fn (vlib_main_t * vm,
     vlib_node_set_state (vlib_mains[i], builtin_client_node.index,
 			 VLIB_NODE_STATE_POLLING);
 
+  if (preallocate_sessions)
+    {
+      session_t *sp __attribute__ ((unused));
+      for (i = 0; i < n_clients; i++)
+	pool_get (tm->sessions, sp);
+      for (i = 0; i < n_clients; i++)
+	pool_put_index (tm->sessions, i);
+    }
+
   /* Fire off connect requests */
+  time_before_connects = vlib_time_now (vm);
   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 /* timeout, seconds */ );
+  vlib_process_wait_for_event_or_clock (vm, syn_timeout);
   event_type = vlib_process_get_events (vm, &event_data);
   switch (event_type)
     {
@@ -719,6 +737,15 @@ test_tcp_clients_command_fn (vlib_main_t * vm,
       goto cleanup;
 
     case 1:
+      delta = vlib_time_now (vm) - time_before_connects;
+
+      if (delta != 0.0)
+	{
+	  vlib_cli_output
+	    (vm, "%d three-way handshakes in %.2f seconds, %.2f/sec",
+	     n_clients, delta, ((f64) n_clients) / delta);
+	}
+
       tm->test_start_time = vlib_time_now (tm->vlib_main);
       vlib_cli_output (vm, "Test started at %.6f", tm->test_start_time);
       break;
@@ -729,7 +756,7 @@ test_tcp_clients_command_fn (vlib_main_t * vm,
     }
 
   /* Now wait for the sessions to finish... */
-  vlib_process_wait_for_event_or_clock (vm, cli_timeout);
+  vlib_process_wait_for_event_or_clock (vm, test_timeout);
   event_type = vlib_process_get_events (vm, &event_data);
   switch (event_type)
     {
diff --git a/src/vnet/tcp/tcp.c b/src/vnet/tcp/tcp.c
index 59b20747da6..8e2eb9f4e3d 100644
--- a/src/vnet/tcp/tcp.c
+++ b/src/vnet/tcp/tcp.c
@@ -173,7 +173,7 @@ tcp_connection_cleanup (tcp_connection_t * tc)
 
   /* Cleanup local endpoint if this was an active connect */
   tepi = transport_endpoint_lookup (&tm->local_endpoints_table, &tc->c_lcl_ip,
-				    tc->c_lcl_port);
+				    clib_net_to_host_u16 (tc->c_lcl_port));
   if (tepi != TRANSPORT_ENDPOINT_INVALID_INDEX)
     {
       tep = pool_elt_at_index (tm->local_endpoints, tepi);
@@ -367,25 +367,24 @@ tcp_allocate_local_port (ip46_address_t * ip)
 {
   tcp_main_t *tm = vnet_get_tcp_main ();
   transport_endpoint_t *tep;
-  u32 time_now, tei;
+  u32 tei;
   u16 min = 1024, max = 65535;	/* XXX configurable ? */
-  int tries;
+  int tries, limit;
 
-  tries = max - min;
-  time_now = tcp_time_now ();
+  limit = max - min;
 
   /* Only support active opens from thread 0 */
   ASSERT (vlib_get_thread_index () == 0);
 
   /* Search for first free slot */
-  for (; tries >= 0; tries--)
+  for (tries = 0; tries < limit; tries++)
     {
       u16 port = 0;
 
       /* Find a port in the specified range */
       while (1)
 	{
-	  port = random_u32 (&time_now) & PORT_MASK;
+	  port = random_u32 (&tm->port_allocator_seed) & PORT_MASK;
 	  if (PREDICT_TRUE (port >= min && port < max))
 	    break;
 	}
@@ -1189,8 +1188,9 @@ tcp_main_enable (vlib_main_t * vm)
   vlib_thread_main_t *vtm = vlib_get_thread_main ();
   clib_error_t *error = 0;
   u32 num_threads;
-  int thread, i;
+  int i, thread;
   tcp_connection_t *tc __attribute__ ((unused));
+  u32 preallocated_connections_per_thread;
 
   if ((error = vlib_call_init_function (vm, ip_main_init)))
     return error;
@@ -1224,14 +1224,26 @@ tcp_main_enable (vlib_main_t * vm)
   vec_validate (tm->connections, num_threads - 1);
 
   /*
-   * Preallocate connections
+   * Preallocate connections. Assume that thread 0 won't
+   * use preallocated threads when running multi-core
    */
-  for (thread = 0; thread < num_threads; thread++)
+  if (num_threads == 1)
     {
-      for (i = 0; i < tm->preallocated_connections; i++)
+      thread = 0;
+      preallocated_connections_per_thread = tm->preallocated_connections;
+    }
+  else
+    {
+      thread = 1;
+      preallocated_connections_per_thread =
+	tm->preallocated_connections / (num_threads - 1);
+    }
+  for (; thread < num_threads; thread++)
+    {
+      for (i = 0; i < preallocated_connections_per_thread; i++)
 	pool_get (tm->connections[thread], tc);
 
-      for (i = 0; i < tm->preallocated_connections; i++)
+      for (i = 0; i < preallocated_connections_per_thread; i++)
 	pool_put_index (tm->connections[thread], i);
     }
 
@@ -1257,13 +1269,21 @@ tcp_main_enable (vlib_main_t * vm)
     / TCP_TSTAMP_RESOLUTION;
 
   clib_bihash_init_24_8 (&tm->local_endpoints_table, "local endpoint table",
-			 200000 /* $$$$ config parameter nbuckets */ ,
-			 (64 << 20) /*$$$ config parameter table size */ );
+			 1000000 /* $$$$ config parameter nbuckets */ ,
+			 (512 << 20) /*$$$ config parameter table size */ );
+
+  /* Initialize [port-allocator] random number seed */
+  tm->port_allocator_seed = (u32) clib_cpu_time_now ();
+
   if (num_threads > 1)
     {
       clib_spinlock_init (&tm->half_open_lock);
       clib_spinlock_init (&tm->local_endpoints_lock);
     }
+
+  vec_validate (tm->tx_frames[0], num_threads - 1);
+  vec_validate (tm->tx_frames[1], num_threads - 1);
+
   return error;
 }
 
@@ -1289,16 +1309,12 @@ clib_error_t *
 tcp_init (vlib_main_t * vm)
 {
   tcp_main_t *tm = vnet_get_tcp_main ();
-
-  tm->vnet_main = vnet_get_main ();
   tm->is_enabled = 0;
-
   return 0;
 }
 
 VLIB_INIT_FUNCTION (tcp_init);
 
-
 static clib_error_t *
 tcp_config_fn (vlib_main_t * vm, unformat_input_t * input)
 {
diff --git a/src/vnet/tcp/tcp.h b/src/vnet/tcp/tcp.h
index 4fa681f8cc1..997df76f545 100644
--- a/src/vnet/tcp/tcp.h
+++ b/src/vnet/tcp/tcp.h
@@ -369,6 +369,8 @@ typedef struct _tcp_main
 
   /** per-worker tx buffer free lists */
   u32 **tx_buffers;
+  /** per-worker tx frames to 4/6 output nodes */
+  vlib_frame_t **tx_frames[2];
 
   /* Per worker-thread timer wheel for connections timers */
   tw_timer_wheel_16t_2w_512sl_t *timer_wheels;
@@ -400,11 +402,8 @@ typedef struct _tcp_main
   u32 last_v6_address_rotor;
   ip6_address_t *ip6_src_addresses;
 
-  /* convenience */
-  vlib_main_t *vlib_main;
-  vnet_main_t *vnet_main;
-  ip4_main_t *ip4_main;
-  ip6_main_t *ip6_main;
+  /** Port allocator random number generator seed */
+  u32 port_allocator_seed;
 } tcp_main_t;
 
 extern tcp_main_t tcp_main;
@@ -493,6 +492,8 @@ void tcp_send_fin (tcp_connection_t * tc);
 void tcp_init_mss (tcp_connection_t * tc);
 void tcp_update_snd_mss (tcp_connection_t * tc);
 void tcp_update_rto (tcp_connection_t * tc);
+void tcp_flush_frame_to_output (vlib_main_t * vm, u8 thread_index, u8 is_ip4);
+void tcp_flush_frames_to_output (u8 thread_index);
 
 always_inline u32
 tcp_end_seq (tcp_header_t * th, u32 len)
@@ -614,6 +615,7 @@ tcp_update_time (f64 now, u32 thread_index)
 {
   tw_timer_expire_timers_16t_2w_512sl (&tcp_main.timer_wheels[thread_index],
 				       now);
+  tcp_flush_frames_to_output (thread_index);
 }
 
 u32 tcp_push_header (transport_connection_t * tconn, vlib_buffer_t * b);
diff --git a/src/vnet/tcp/tcp_input.c b/src/vnet/tcp/tcp_input.c
index 6c59d70f072..29f4f08d72c 100644
--- a/src/vnet/tcp/tcp_input.c
+++ b/src/vnet/tcp/tcp_input.c
@@ -1751,6 +1751,8 @@ tcp46_established_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
 
   errors = session_manager_flush_enqueue_events (my_thread_index);
   tcp_established_inc_counter (vm, is_ip4, TCP_ERROR_EVENT_FIFO_FULL, errors);
+  tcp_flush_frame_to_output (vm, my_thread_index, is_ip4);
+
   return from_frame->n_vectors;
 }
 
diff --git a/src/vnet/tcp/tcp_output.c b/src/vnet/tcp/tcp_output.c
index ad13493a14c..f8fbb8a9e69 100644
--- a/src/vnet/tcp/tcp_output.c
+++ b/src/vnet/tcp/tcp_output.c
@@ -436,34 +436,41 @@ tcp_init_mss (tcp_connection_t * tc)
     tc->snd_mss -= TCP_OPTION_LEN_TIMESTAMP;
 }
 
-#define tcp_get_free_buffer_index(tm, bidx)                             \
-do {                                                                    \
-  u32 *my_tx_buffers, n_free_buffers;                                   \
-  u32 thread_index = vlib_get_thread_index();                           \
-  my_tx_buffers = tm->tx_buffers[thread_index];                         \
-  if (PREDICT_FALSE(vec_len (my_tx_buffers) == 0))                      \
-    {                                                                   \
-      n_free_buffers = 32;      /* TODO config or macro */              \
-      vec_validate (my_tx_buffers, n_free_buffers - 1);                 \
-      _vec_len(my_tx_buffers) = vlib_buffer_alloc_from_free_list (      \
-       vlib_get_main(), my_tx_buffers, n_free_buffers,                  \
-          VLIB_BUFFER_DEFAULT_FREE_LIST_INDEX);                         \
-      tm->tx_buffers[thread_index] = my_tx_buffers;                     \
-    }                                                                   \
-  /* buffer shortage */                                                 \
-  if (PREDICT_FALSE (vec_len (my_tx_buffers) == 0))                     \
-    return;                                                             \
-  *bidx = my_tx_buffers[_vec_len (my_tx_buffers)-1];                    \
-  _vec_len (my_tx_buffers) -= 1;                                        \
-} while (0)
-
-#define tcp_return_buffer(tm)                   \
-do {                                            \
-  u32 *my_tx_buffers;                           \
-  u32 thread_index = vlib_get_thread_index();   \
-  my_tx_buffers = tm->tx_buffers[thread_index]; \
-  _vec_len (my_tx_buffers) +=1;                 \
-} while (0)
+always_inline int
+tcp_get_free_buffer_index (tcp_main_t * tm, u32 * bidx)
+{
+  u32 *my_tx_buffers, n_free_buffers;
+  u32 thread_index = vlib_get_thread_index ();
+  my_tx_buffers = tm->tx_buffers[thread_index];
+  if (PREDICT_FALSE (vec_len (my_tx_buffers) == 0))
+    {
+      n_free_buffers = VLIB_FRAME_SIZE;
+      vec_validate (my_tx_buffers, n_free_buffers - 1);
+      _vec_len (my_tx_buffers) =
+	vlib_buffer_alloc_from_free_list (vlib_get_main (), my_tx_buffers,
+					  n_free_buffers,
+					  VLIB_BUFFER_DEFAULT_FREE_LIST_INDEX);
+      /* buffer shortage, report failure */
+      if (vec_len (my_tx_buffers) == 0)
+	{
+	  clib_warning ("out of buffers");
+	  return -1;
+	}
+      tm->tx_buffers[thread_index] = my_tx_buffers;
+    }
+  *bidx = my_tx_buffers[_vec_len (my_tx_buffers) - 1];
+  _vec_len (my_tx_buffers) -= 1;
+  return 0;
+}
+
+always_inline void
+tcp_return_buffer (tcp_main_t * tm)
+{
+  u32 *my_tx_buffers;
+  u32 thread_index = vlib_get_thread_index ();
+  my_tx_buffers = tm->tx_buffers[thread_index];
+  _vec_len (my_tx_buffers) += 1;
+}
 
 always_inline void
 tcp_reuse_buffer (vlib_main_t * vm, vlib_buffer_t * b)
@@ -706,7 +713,9 @@ tcp_send_reset (tcp_connection_t * tc, vlib_buffer_t * pkt, u8 is_ip4)
   ip4_header_t *ih4, *pkt_ih4;
   ip6_header_t *ih6, *pkt_ih6;
 
-  tcp_get_free_buffer_index (tm, &bi);
+  if (PREDICT_FALSE (tcp_get_free_buffer_index (tm, &bi)))
+    return;
+
   b = vlib_get_buffer (vm, bi);
 
   /* Leave enough space for headers */
@@ -811,7 +820,9 @@ tcp_send_syn (tcp_connection_t * tc)
   u16 initial_wnd;
   tcp_options_t snd_opts;
 
-  tcp_get_free_buffer_index (tm, &bi);
+  if (PREDICT_FALSE (tcp_get_free_buffer_index (tm, &bi)))
+    return;
+
   b = vlib_get_buffer (vm, bi);
 
   /* Leave enough space for headers */
@@ -854,8 +865,11 @@ tcp_send_syn (tcp_connection_t * tc)
 }
 
 always_inline void
-tcp_enqueue_to_output (vlib_main_t * vm, vlib_buffer_t * b, u32 bi, u8 is_ip4)
+tcp_enqueue_to_output_i (vlib_main_t * vm, vlib_buffer_t * b, u32 bi,
+			 u8 is_ip4, u8 flush)
 {
+  tcp_main_t *tm = vnet_get_tcp_main ();
+  u32 thread_index = vlib_get_thread_index ();
   u32 *to_next, next_index;
   vlib_frame_t *f;
 
@@ -872,12 +886,62 @@ tcp_enqueue_to_output (vlib_main_t * vm, vlib_buffer_t * b, u32 bi, u8 is_ip4)
       b->pre_data[1] = next_index;
     }
 
-  /* Enqueue the packet */
-  f = vlib_get_frame_to_node (vm, next_index);
+  /* Get frame to v4/6 output node */
+  f = tm->tx_frames[!is_ip4][thread_index];
+  if (!f)
+    {
+      f = vlib_get_frame_to_node (vm, next_index);
+      ASSERT (f);
+      tm->tx_frames[!is_ip4][thread_index] = f;
+    }
   to_next = vlib_frame_vector_args (f);
-  to_next[0] = bi;
-  f->n_vectors = 1;
-  vlib_put_frame_to_node (vm, next_index, f);
+  to_next[f->n_vectors] = bi;
+  f->n_vectors += 1;
+  if (flush || f->n_vectors == VLIB_FRAME_SIZE)
+    {
+      vlib_put_frame_to_node (vm, next_index, f);
+      tm->tx_frames[!is_ip4][thread_index] = 0;
+    }
+}
+
+always_inline void
+tcp_enqueue_to_output (vlib_main_t * vm, vlib_buffer_t * b, u32 bi, u8 is_ip4)
+{
+  tcp_enqueue_to_output_i (vm, b, bi, is_ip4, 0);
+}
+
+always_inline void
+tcp_enqueue_to_output_now (vlib_main_t * vm, vlib_buffer_t * b, u32 bi,
+			   u8 is_ip4)
+{
+  tcp_enqueue_to_output_i (vm, b, bi, is_ip4, 1);
+}
+
+/**
+ * Flush tx frame populated by retransmits and timer pops
+ */
+void
+tcp_flush_frame_to_output (vlib_main_t * vm, u8 thread_index, u8 is_ip4)
+{
+  if (tcp_main.tx_frames[!is_ip4][thread_index])
+    {
+      u32 next_index;
+      next_index = is_ip4 ? tcp4_output_node.index : tcp6_output_node.index;
+      vlib_put_frame_to_node (vm, next_index,
+			      tcp_main.tx_frames[!is_ip4][thread_index]);
+      tcp_main.tx_frames[!is_ip4][thread_index] = 0;
+    }
+}
+
+/**
+ * Flush both v4 and v6 tx frames for thread index
+ */
+void
+tcp_flush_frames_to_output (u8 thread_index)
+{
+  vlib_main_t *vm = vlib_get_main ();
+  tcp_flush_frame_to_output (vm, thread_index, 1);
+  tcp_flush_frame_to_output (vm, thread_index, 0);
 }
 
 /**
@@ -891,14 +955,15 @@ tcp_send_fin (tcp_connection_t * tc)
   tcp_main_t *tm = vnet_get_tcp_main ();
   vlib_main_t *vm = vlib_get_main ();
 
-  tcp_get_free_buffer_index (tm, &bi);
+  if (PREDICT_FALSE (tcp_get_free_buffer_index (tm, &bi)))
+    return;
   b = vlib_get_buffer (vm, bi);
 
   /* Leave enough space for headers */
   vlib_buffer_make_headroom (b, MAX_HDRS_LEN);
 
   tcp_make_fin (tc, b);
-  tcp_enqueue_to_output (vm, b, bi, tc->c_is_ip4);
+  tcp_enqueue_to_output_now (vm, b, bi, tc->c_is_ip4);
   tc->flags |= TCP_CONN_FINSNT;
   tcp_retransmit_timer_force_update (tc);
   TCP_EVT_DBG (TCP_EVT_FIN_SENT, tc);
@@ -981,7 +1046,8 @@ tcp_send_ack (tcp_connection_t * tc)
   u32 bi;
 
   /* Get buffer */
-  tcp_get_free_buffer_index (tm, &bi);
+  if (PREDICT_FALSE (tcp_get_free_buffer_index (tm, &bi)))
+    return;
   b = vlib_get_buffer (vm, bi);
 
   /* Fill in the ACK */
@@ -1108,7 +1174,9 @@ tcp_timer_retransmit_handler_i (u32 index, u8 is_syn)
   /* Go back to first un-acked byte */
   tc->snd_nxt = tc->snd_una;
 
-  tcp_get_free_buffer_index (tm, &bi);
+  if (PREDICT_FALSE (tcp_get_free_buffer_index (tm, &bi)))
+    return;
+
   b = vlib_get_buffer (vm, bi);
 
   if (tc->state >= TCP_STATE_ESTABLISHED)
@@ -1116,6 +1184,7 @@ tcp_timer_retransmit_handler_i (u32 index, u8 is_syn)
       /* Lost FIN, retransmit and return */
       if (tc->flags & TCP_CONN_FINSNT)
 	{
+	  tcp_return_buffer (tm);
 	  tcp_send_fin (tc);
 	  return;
 	}
@@ -1143,6 +1212,7 @@ tcp_timer_retransmit_handler_i (u32 index, u8 is_syn)
 	  tcp_retransmit_timer_set (tc);
 	  ASSERT (0 || (tc->rto_boff > 1
 			&& tc->snd_una == tc->snd_congestion));
+	  tcp_return_buffer (tm);
 	  return;
 	}
 
@@ -1164,6 +1234,7 @@ tcp_timer_retransmit_handler_i (u32 index, u8 is_syn)
 	      clib_warning ("could not remove half-open connection");
 	      ASSERT (0);
 	    }
+	  tcp_return_buffer (tm);
 	  return;
 	}
 
@@ -1185,6 +1256,7 @@ tcp_timer_retransmit_handler_i (u32 index, u8 is_syn)
     {
       ASSERT (tc->state == TCP_STATE_CLOSED);
       clib_warning ("connection closed ...");
+      tcp_return_buffer (tm);
       return;
     }
 
@@ -1254,7 +1326,9 @@ tcp_timer_persist_handler (u32 index)
   tc->rto = clib_min (tc->rto << 1, TCP_RTO_MAX);
 
   /* Try to force the first unsent segment  */
-  tcp_get_free_buffer_index (tm, &bi);
+  if (PREDICT_FALSE (tcp_get_free_buffer_index (tm, &bi)))
+    return;
+
   b = vlib_get_buffer (vm, bi);
 
   tcp_validate_txf_size (tc, tc->snd_una_max - tc->snd_una);
@@ -1300,7 +1374,9 @@ tcp_retransmit_first_unacked (tcp_connection_t * tc)
   tc->snd_nxt = tc->snd_una;
 
   /* Get buffer */
-  tcp_get_free_buffer_index (tm, &bi);
+  if (PREDICT_FALSE (tcp_get_free_buffer_index (tm, &bi)))
+    return;
+
   b = vlib_get_buffer (vm, bi);
 
   TCP_EVT_DBG (TCP_EVT_CC_EVT, tc, 2);
@@ -1344,9 +1420,10 @@ tcp_fast_retransmit_sack (tcp_connection_t * tc)
   hole = scoreboard_get_hole (sb, sb->cur_rxt_hole);
   while (hole && snd_space > 0)
     {
-      tcp_get_free_buffer_index (tm, &bi);
-      b = vlib_get_buffer (vm, bi);
+      if (PREDICT_FALSE (tcp_get_free_buffer_index (tm, &bi)))
+	return;
 
+      b = vlib_get_buffer (vm, bi);
       hole = scoreboard_next_rxt_hole (sb, hole,
 				       tcp_fastrecovery_sent_1_smss (tc),
 				       &can_rescue, &snd_limited);
@@ -1414,9 +1491,9 @@ tcp_fast_retransmit_no_sack (tcp_connection_t * tc)
 
   while (snd_space > 0)
     {
-      tcp_get_free_buffer_index (tm, &bi);
+      if (PREDICT_FALSE (tcp_get_free_buffer_index (tm, &bi)))
+	return;
       b = vlib_get_buffer (vm, bi);
-
       offset += n_written;
       n_written = tcp_prepare_retransmit_segment (tc, b, offset, snd_space);
 
@@ -1506,32 +1583,21 @@ tcp46_output_inline (vlib_main_t * vm,
 
 	  if (is_ip4)
 	    {
-	      ip4_header_t *ih0;
-	      ih0 = vlib_buffer_push_ip4 (vm, b0, &tc0->c_lcl_ip4,
-					  &tc0->c_rmt_ip4, IP_PROTOCOL_TCP,
-					  1);
-	      b0->flags |=
-		VNET_BUFFER_F_IS_IP4 | VNET_BUFFER_F_OFFLOAD_IP_CKSUM |
-		VNET_BUFFER_F_OFFLOAD_TCP_CKSUM;
-	      vnet_buffer (b0)->l3_hdr_offset = (u8 *) ih0 - b0->data;
+	      vlib_buffer_push_ip4 (vm, b0, &tc0->c_lcl_ip4, &tc0->c_rmt_ip4,
+				    IP_PROTOCOL_TCP, 1);
+	      b0->flags |= VNET_BUFFER_F_OFFLOAD_TCP_CKSUM;
 	      vnet_buffer (b0)->l4_hdr_offset = (u8 *) th0 - b0->data;
 	      th0->checksum = 0;
 	    }
 	  else
 	    {
 	      ip6_header_t *ih0;
-	      int bogus = ~0;
-
 	      ih0 = vlib_buffer_push_ip6 (vm, b0, &tc0->c_lcl_ip6,
 					  &tc0->c_rmt_ip6, IP_PROTOCOL_TCP);
-
-	      b0->flags |= VNET_BUFFER_F_IS_IP6 |
-		VNET_BUFFER_F_OFFLOAD_IP_CKSUM |
-		VNET_BUFFER_F_OFFLOAD_TCP_CKSUM;
+	      b0->flags |= VNET_BUFFER_F_OFFLOAD_TCP_CKSUM;
 	      vnet_buffer (b0)->l3_hdr_offset = (u8 *) ih0 - b0->data;
 	      vnet_buffer (b0)->l4_hdr_offset = (u8 *) th0 - b0->data;
 	      th0->checksum = 0;
-	      ASSERT (!bogus);
 	    }
 
 	  /* Filter out DUPACKs if there are no OOO segments left */
diff --git a/src/vnet/unix/gdb_funcs.c b/src/vnet/unix/gdb_funcs.c
index cca2e420aa8..32e22d924ba 100644
--- a/src/vnet/unix/gdb_funcs.c
+++ b/src/vnet/unix/gdb_funcs.c
@@ -21,7 +21,7 @@
 
 #include <vlib/threads.h>
 #include <vnet/vnet.h>
-
+#include <vppinfra/format.h>
 
 /**
  * @brief GDB callable function: vl - Return vector length of vector
@@ -135,6 +135,47 @@ void vlib_runtime_index_to_node_name (u32 index)
   fformat(stderr, "node runtime index %d name %s\n", index, nm->nodes[index]->name);
 }
 
+void gdb_show_errors (int verbose)
+{
+  extern vlib_cli_command_t vlib_cli_show_errors;
+  unformat_input_t input;
+  vlib_main_t * vm = vlib_get_main();
+
+  if (verbose == 0)
+    unformat_init_string (&input, "verbose 0", 9);
+  else if (verbose == 1)
+    unformat_init_string (&input, "verbose 1", 9);
+  else 
+    {
+      fformat(stderr, "verbose not 0 or 1\n");
+      return;
+    }
+
+  vlib_cli_show_errors.function (vm, &input, 0 /* cmd */);
+  unformat_free (&input);
+}  
+
+void gdb_show_session (int verbose)
+{
+  extern vlib_cli_command_t vlib_cli_show_session_command;
+  unformat_input_t input;
+  vlib_main_t * vm = vlib_get_main();
+
+  if (verbose == 0)
+    unformat_init_string (&input, "verbose 0", 9);
+  else if (verbose == 1)
+    unformat_init_string (&input, "verbose 1", 9);
+  else if (verbose == 2)
+    unformat_init_string (&input, "verbose 2", 9);
+  else 
+    {
+      fformat(stderr, "verbose not 0 - 2\n");
+      return;
+    }
+
+  vlib_cli_show_session_command.function (vm, &input, 0 /* cmd */);
+  unformat_free (&input);
+}  
 
 /**
  * @brief GDB callable function: show_gdb_command_fn - show gdb
@@ -151,6 +192,8 @@ show_gdb_command_fn (vlib_main_t * vm,
   vlib_cli_output (vm, "vl(p) returns vec_len(p)");
   vlib_cli_output (vm, "pe(p) returns pool_elts(p)");
   vlib_cli_output (vm, "pifi(p, i) returns pool_is_free_index(p, i)");
+  vlib_cli_output (vm, "gdb_show_errors(0|1) dumps error counters");
+  vlib_cli_output (vm, "gdb_show_session dumps session counters");
   vlib_cli_output (vm, "debug_hex_bytes (ptr, n_bytes) dumps n_bytes in hex");
   vlib_cli_output (vm, "vlib_dump_frame_ownership() does what it says");
   vlib_cli_output (vm, "vlib_runtime_index_to_node_name (index) prints NN");
diff --git a/src/vppinfra/pool.h b/src/vppinfra/pool.h
index 57838e1c609..56536b77963 100644
--- a/src/vppinfra/pool.h
+++ b/src/vppinfra/pool.h
@@ -200,7 +200,7 @@ do {									\
 #define pool_get(P,E) pool_get_aligned(P,E,0)
 
 /** See if pool_get will expand the pool or not */
-#define pool_get_aligned_will_expand (P,YESNO,A)                        \
+#define pool_get_aligned_will_expand(P,YESNO,A)                         \
 do {                                                                    \
   pool_header_t * _pool_var (p) = pool_header (P);                      \
   uword _pool_var (l);                                                  \