tcp: improve pacing after idle send periods

Rest pacer on ack reception if we haven't recently sent anything.

Type: feature

Change-Id: I820bacd81b65130052dfafbfcbe6ca4553069fbc
Signed-off-by: Florin Coras <fcoras@cisco.com>
(cherry picked from commit c31dc31f84961033ecb6354811e0c360b6cf5f79)

6 files changed, 124 insertions, 76 deletions

diff --git a/src/vnet/session/session_node.c b/src/vnet/session/session_node.c
index 99fa4bdac91..615095fdc6c 100644
--- a/src/vnet/session/session_node.c
+++ b/src/vnet/session/session_node.c
@@ -843,6 +843,7 @@ session_tx_fifo_read_and_snd_i (session_worker_t * wrk,
     {
       if (ctx->transport_vft->flush_data)
 	ctx->transport_vft->flush_data (ctx->tc);
+      e->event_type = SESSION_IO_EVT_TX;
     }
 
   if (ctx->s->flags & SESSION_F_CUSTOM_TX)
@@ -881,7 +882,12 @@ session_tx_fifo_read_and_snd_i (session_worker_t * wrk,
   session_tx_set_dequeue_params (vm, ctx, max_burst, peek_data);
 
   if (PREDICT_FALSE (!ctx->max_len_to_snd))
-    return SESSION_TX_NO_DATA;
+    {
+      transport_connection_tx_pacer_reset_bucket (ctx->tc,
+						  vm->clib_time.
+						  last_cpu_time);
+      return SESSION_TX_NO_DATA;
+    }
 
   n_bufs_needed = ctx->n_segs_per_evt * ctx->n_bufs_per_seg;
   vec_validate_aligned (wrk->tx_buffers, n_bufs_needed - 1,
diff --git a/src/vnet/session/transport.c b/src/vnet/session/transport.c
index 49a4af71111..8f7e30c8602 100644
--- a/src/vnet/session/transport.c
+++ b/src/vnet/session/transport.c
@@ -47,10 +47,6 @@ static clib_spinlock_t local_endpoints_lock;
  */
 static double transport_pacer_period;
 
-#define TRANSPORT_PACER_MIN_MSS 	1460
-#define TRANSPORT_PACER_MIN_BURST 	TRANSPORT_PACER_MIN_MSS
-#define TRANSPORT_PACER_MAX_BURST	(43 * TRANSPORT_PACER_MIN_MSS)
-
 u8 *
 format_transport_proto (u8 * s, va_list * args)
 {
@@ -583,6 +579,13 @@ spacer_max_burst (spacer_t * pacer, u64 norm_time_now)
   u64 n_periods = norm_time_now - pacer->last_update;
   u64 inc;
 
+  if (PREDICT_FALSE (n_periods > 5e5))
+    {
+      pacer->last_update = norm_time_now;
+      pacer->bucket = TRANSPORT_PACER_MIN_BURST;
+      return TRANSPORT_PACER_MIN_BURST;
+    }
+
   if (n_periods > 0
       && (inc = (f32) n_periods * pacer->tokens_per_period) > 10)
     {
@@ -674,7 +677,7 @@ transport_connection_snd_space (transport_connection_t * tc, u64 time_now,
   u32 snd_space, max_paced_burst;
 
   snd_space = tp_vfts[tc->proto].send_space (tc);
-  if (transport_connection_is_tx_paced (tc))
+  if (snd_space && transport_connection_is_tx_paced (tc))
     {
       time_now >>= SPACER_CPU_TICKS_PER_PERIOD_SHIFT;
       max_paced_burst = spacer_max_burst (&tc->pacer, time_now);
diff --git a/src/vnet/session/transport.h b/src/vnet/session/transport.h
index 5b45be0a0be..914991d5b31 100644
--- a/src/vnet/session/transport.h
+++ b/src/vnet/session/transport.h
@@ -19,6 +19,10 @@
 #include <vnet/vnet.h>
 #include <vnet/session/transport_types.h>
 
+#define TRANSPORT_PACER_MIN_MSS 	1460
+#define TRANSPORT_PACER_MIN_BURST 	TRANSPORT_PACER_MIN_MSS
+#define TRANSPORT_PACER_MAX_BURST	(43 * TRANSPORT_PACER_MIN_MSS)
+
 typedef struct _transport_options_t
 {
   transport_tx_fn_type_t tx_type;
diff --git a/src/vnet/tcp/tcp.h b/src/vnet/tcp/tcp.h
index a31c46cdacf..5e683f72dbd 100644
--- a/src/vnet/tcp/tcp.h
+++ b/src/vnet/tcp/tcp.h
@@ -372,6 +372,7 @@ typedef struct _tcp_connection
   u32 prr_start;	/**< snd_una when prr starts */
   u32 rxt_delivered;	/**< Rxt bytes delivered during current cc event */
   u32 rxt_head;		/**< snd_una last time we re rxted the head */
+  u32 prev_dsegs_out;	/**< Number of dsegs after last ack */
   u32 tsecr_last_ack;	/**< Timestamp echoed to us in last healthy ACK */
   u32 snd_congestion;	/**< snd_una_max when congestion is detected */
   u32 tx_fifo_size;	/**< Tx fifo size. Used to constrain cwnd */
diff --git a/src/vnet/tcp/tcp_input.c b/src/vnet/tcp/tcp_input.c
index f177f5f4753..ea64e4ab754 100755
--- a/src/vnet/tcp/tcp_input.c
+++ b/src/vnet/tcp/tcp_input.c
@@ -574,6 +574,15 @@ tcp_estimate_initial_rtt (tcp_connection_t * tc)
   tcp_update_rto (tc);
 }
 
+always_inline u8
+tcp_recovery_no_snd_space (tcp_connection_t * tc)
+{
+  return (tcp_in_fastrecovery (tc)
+	  && tcp_fastrecovery_prr_snd_space (tc) < tc->snd_mss)
+    || (tcp_in_recovery (tc)
+	&& tcp_available_output_snd_space (tc) < tc->snd_mss);
+}
+
 /**
  * Dequeue bytes for connections that have received acks in last burst
  */
@@ -594,26 +603,35 @@ tcp_handle_postponed_dequeues (tcp_worker_ctx_t * wrk)
       tc = tcp_connection_get (pending_deq_acked[i], thread_index);
       tc->flags &= ~TCP_CONN_DEQ_PENDING;
 
-      if (PREDICT_FALSE (!tc->burst_acked))
-	continue;
+      if (tc->burst_acked)
+	{
+	  /* Dequeue the newly ACKed bytes */
+	  session_tx_fifo_dequeue_drop (&tc->connection, tc->burst_acked);
+	  tc->burst_acked = 0;
+	  tcp_validate_txf_size (tc, tc->snd_una_max - tc->snd_una);
 
-      /* Dequeue the newly ACKed bytes */
-      session_tx_fifo_dequeue_drop (&tc->connection, tc->burst_acked);
-      tc->burst_acked = 0;
-      tcp_validate_txf_size (tc, tc->snd_una_max - tc->snd_una);
+	  if (PREDICT_FALSE (tc->flags & TCP_CONN_PSH_PENDING))
+	    {
+	      if (seq_leq (tc->psh_seq, tc->snd_una))
+		tc->flags &= ~TCP_CONN_PSH_PENDING;
+	    }
 
-      if (PREDICT_FALSE (tc->flags & TCP_CONN_PSH_PENDING))
-	{
-	  if (seq_leq (tc->psh_seq, tc->snd_una))
-	    tc->flags &= ~TCP_CONN_PSH_PENDING;
-	}
+	  /* If everything has been acked, stop retransmit timer
+	   * otherwise update. */
+	  tcp_retransmit_timer_update (tc);
 
-      /* If everything has been acked, stop retransmit timer
-       * otherwise update. */
-      tcp_retransmit_timer_update (tc);
+	  /* Update pacer based on our new cwnd estimate */
+	  tcp_connection_tx_pacer_update (tc);
+	}
 
-      /* Update pacer based on our new cwnd estimate */
-      tcp_connection_tx_pacer_update (tc);
+      /* Reset the pacer if we've been idle, i.e., no data sent or if
+       * we're in recovery and snd space constrained */
+      if (tc->data_segs_out == tc->prev_dsegs_out
+	  || tcp_recovery_no_snd_space (tc))
+	transport_connection_tx_pacer_reset_bucket (&tc->connection,
+						    wrk->vm->clib_time.
+						    last_cpu_time);
+      tc->prev_dsegs_out = tc->data_segs_out;
     }
   _vec_len (wrk->pending_deq_acked) = 0;
 }
@@ -1335,6 +1353,8 @@ tcp_cc_recover (tcp_connection_t * tc)
   tc->rtt_ts = 0;
   tc->flags &= ~TCP_CONN_RXT_PENDING;
 
+  tcp_connection_tx_pacer_reset (tc, tc->cwnd, 0 /* start bucket */ );
+
   /* Previous recovery left us congested. Continue sending as part
    * of the current recovery event with an updated snd_congestion */
   if (tc->sack_sb.sacked_bytes)
@@ -1354,8 +1374,6 @@ tcp_cc_recover (tcp_connection_t * tc)
   if (!tcp_in_recovery (tc) && !is_spurious)
     tcp_cc_recovered (tc);
 
-  tcp_connection_tx_pacer_reset (tc, tc->cwnd, 0 /* start bucket */ );
-
   tcp_fastrecovery_off (tc);
   tcp_fastrecovery_first_off (tc);
   tcp_recovery_off (tc);
@@ -1589,11 +1607,10 @@ process_ack:
   if (tc->flags & TCP_CONN_RATE_SAMPLE)
     tcp_bt_sample_delivery_rate (tc, &rs);
 
+  tcp_program_dequeue (wrk, tc);
+
   if (tc->bytes_acked)
-    {
-      tcp_program_dequeue (wrk, tc);
-      tcp_update_rtt (tc, &rs, vnet_buffer (b)->tcp.ack_number);
-    }
+    tcp_update_rtt (tc, &rs, vnet_buffer (b)->tcp.ack_number);
 
   TCP_EVT (TCP_EVT_ACK_RCVD, tc);
 
diff --git a/src/vnet/tcp/tcp_output.c b/src/vnet/tcp/tcp_output.c
index 79866aff03a..047247e9ebe 100644
--- a/src/vnet/tcp/tcp_output.c
+++ b/src/vnet/tcp/tcp_output.c
@@ -409,7 +409,10 @@ tcp_update_burst_snd_vars (tcp_connection_t * tc)
     tc->flags |= TCP_CONN_TRACK_BURST;
 
   if (tc->snd_una == tc->snd_nxt)
-    tcp_cc_event (tc, TCP_CC_EVT_START_TX);
+    {
+      tcp_cc_event (tc, TCP_CC_EVT_START_TX);
+      tcp_connection_tx_pacer_reset (tc, tc->cwnd, TRANSPORT_PACER_MIN_MSS);
+    }
 }
 
 #endif /* CLIB_MARCH_VARIANT */
@@ -1435,19 +1438,7 @@ tcp_prepare_retransmit_segment (tcp_worker_ctx_t * wrk,
   max_deq_bytes = clib_min (tc->snd_mss, max_deq_bytes);
   max_deq_bytes = clib_min (available_bytes, max_deq_bytes);
 
-  /* Start is beyond snd_congestion */
   start = tc->snd_una + offset;
-  if (seq_geq (start, tc->snd_congestion))
-    return 0;
-
-  /* Don't overshoot snd_congestion */
-  if (seq_gt (start + max_deq_bytes, tc->snd_congestion))
-    {
-      max_deq_bytes = tc->snd_congestion - start;
-      if (max_deq_bytes == 0)
-	return 0;
-    }
-
   n_bytes = tcp_prepare_segment (wrk, tc, offset, max_deq_bytes, b);
   if (!n_bytes)
     return 0;
@@ -1903,17 +1894,28 @@ static int
 tcp_retransmit_sack (tcp_worker_ctx_t * wrk, tcp_connection_t * tc,
 		     u32 burst_size)
 {
+  u8 snd_limited = 0, can_rescue = 0, reset_pacer = 0;
   u32 n_written = 0, offset, max_bytes, n_segs = 0;
+  u32 bi, max_deq, burst_bytes, sent_bytes;
   sack_scoreboard_hole_t *hole;
   vlib_main_t *vm = wrk->vm;
   vlib_buffer_t *b = 0;
   sack_scoreboard_t *sb;
-  u32 bi, max_deq;
   int snd_space;
-  u8 snd_limited = 0, can_rescue = 0;
+  u64 time_now;
 
   ASSERT (tcp_in_cong_recovery (tc));
 
+  time_now = wrk->vm->clib_time.last_cpu_time;
+  burst_bytes = transport_connection_tx_pacer_burst (&tc->connection,
+						     time_now);
+  burst_size = clib_min (burst_size, burst_bytes / tc->snd_mss);
+  if (!burst_size)
+    {
+      tcp_program_retransmit (tc);
+      return 0;
+    }
+
   if (tcp_in_recovery (tc))
     snd_space = tcp_available_cc_snd_space (tc);
   else
@@ -1921,13 +1923,12 @@ tcp_retransmit_sack (tcp_worker_ctx_t * wrk, tcp_connection_t * tc,
 
   if (snd_space < tc->snd_mss)
     {
-      /* We're cc constrained so don't accumulate tokens */
-      transport_connection_tx_pacer_reset_bucket (&tc->connection,
-						  vm->
-						  clib_time.last_cpu_time);
-      return 0;
+      reset_pacer = burst_bytes > tc->snd_mss;
+      goto done;
     }
 
+  reset_pacer = snd_space < burst_bytes;
+
   sb = &tc->sack_sb;
 
   /* Check if snd_una is a lost retransmit */
@@ -1967,9 +1968,12 @@ tcp_retransmit_sack (tcp_worker_ctx_t * wrk, tcp_connection_t * tc,
 	  /* We are out of lost holes to retransmit so send some new data. */
 	  if (max_deq > tc->snd_mss)
 	    {
-	      u32 n_segs_new, av_window;
-	      av_window = tc->snd_wnd - (tc->snd_nxt - tc->snd_una);
-	      snd_space = clib_min (snd_space, av_window);
+	      u32 n_segs_new;
+	      int av_wnd;
+
+	      av_wnd = (int) tc->snd_wnd - (tc->snd_nxt - tc->snd_una);
+	      av_wnd = clib_max (av_wnd, 0);
+	      snd_space = clib_min (snd_space, av_wnd);
 	      snd_space = clib_min (max_deq, snd_space);
 	      burst_size = clib_min (burst_size - n_segs,
 				     snd_space / tc->snd_mss);
@@ -2034,6 +2038,19 @@ tcp_retransmit_sack (tcp_worker_ctx_t * wrk, tcp_connection_t * tc,
 
 done:
 
+  if (reset_pacer)
+    {
+      transport_connection_tx_pacer_reset_bucket (&tc->connection,
+						  vm->clib_time.
+						  last_cpu_time);
+    }
+  else
+    {
+      sent_bytes = clib_min (n_segs * tc->snd_mss, burst_bytes);
+      transport_connection_tx_pacer_update_bytes (&tc->connection,
+						  sent_bytes);
+    }
+
   return n_segs;
 }
 
@@ -2045,14 +2062,28 @@ tcp_retransmit_no_sack (tcp_worker_ctx_t * wrk, tcp_connection_t * tc,
 			u32 burst_size)
 {
   u32 n_written = 0, offset = 0, bi, max_deq, n_segs_now;
+  u32 burst_bytes, sent_bytes;
   vlib_main_t *vm = wrk->vm;
   int snd_space, n_segs = 0;
+  u8 cc_limited = 0;
   vlib_buffer_t *b;
+  u64 time_now;
 
   ASSERT (tcp_in_fastrecovery (tc));
   TCP_EVT (TCP_EVT_CC_EVT, tc, 0);
 
+  time_now = wrk->vm->clib_time.last_cpu_time;
+  burst_bytes = transport_connection_tx_pacer_burst (&tc->connection,
+						     time_now);
+  burst_size = clib_min (burst_size, burst_bytes / tc->snd_mss);
+  if (!burst_size)
+    {
+      tcp_program_retransmit (tc);
+      return 0;
+    }
+
   snd_space = tcp_available_cc_snd_space (tc);
+  cc_limited = snd_space < burst_bytes;
 
   if (!tcp_fastrecovery_first (tc))
     goto send_unsent;
@@ -2098,19 +2129,12 @@ send_unsent:
 
 done:
   tcp_fastrecovery_first_off (tc);
-  return n_segs;
-}
 
-/**
- * Do fast retransmit
- */
-static int
-tcp_retransmit (tcp_worker_ctx_t * wrk, tcp_connection_t * tc, u32 burst_size)
-{
-  if (tcp_opts_sack_permitted (&tc->rcv_opts))
-    return tcp_retransmit_sack (wrk, tc, burst_size);
-  else
-    return tcp_retransmit_no_sack (wrk, tc, burst_size);
+  sent_bytes = clib_min (n_segs * tc->snd_mss, burst_bytes);
+  sent_bytes = cc_limited ? burst_bytes : sent_bytes;
+  transport_connection_tx_pacer_update_bytes (&tc->connection, sent_bytes);
+
+  return n_segs;
 }
 
 static int
@@ -2164,23 +2188,16 @@ tcp_send_acks (tcp_connection_t * tc, u32 max_burst_size)
 static int
 tcp_do_retransmit (tcp_connection_t * tc, u32 max_burst_size)
 {
-  u32 n_segs = 0, burst_size, sent_bytes, burst_bytes;
   tcp_worker_ctx_t *wrk;
+  u32 n_segs;
 
   wrk = tcp_get_worker (tc->c_thread_index);
-  burst_bytes = transport_connection_tx_pacer_burst (&tc->connection,
-						     wrk->vm->
-						     clib_time.last_cpu_time);
-  burst_size = clib_min (max_burst_size, burst_bytes / tc->snd_mss);
-  if (!burst_size)
-    {
-      tcp_program_retransmit (tc);
-      return 0;
-    }
 
-  n_segs = tcp_retransmit (wrk, tc, burst_size);
-  sent_bytes = clib_min (n_segs * tc->snd_mss, burst_bytes);
-  transport_connection_tx_pacer_update_bytes (&tc->connection, sent_bytes);
+  if (tcp_opts_sack_permitted (&tc->rcv_opts))
+    n_segs = tcp_retransmit_sack (wrk, tc, max_burst_size);
+  else
+    n_segs = tcp_retransmit_no_sack (wrk, tc, max_burst_size);
+
   return n_segs;
 }