tcp: retry lost retransmits

Add heuristic that detects lost retransmitted segments and retries
sending them.

Type: feature

Change-Id: I34d1bb16799e1993779222eb2bfad4b40704159e
Signed-off-by: Florin Coras <fcoras@cisco.com>
(cherry picked from commit be237bf02382854118986e8ea84c7544e42023f2)

7 files changed, 205 insertions, 155 deletions

diff --git a/src/vnet/session/session.c b/src/vnet/session/session.c
index be62d929363..32d7914a607 100644
--- a/src/vnet/session/session.c
+++ b/src/vnet/session/session.c
@@ -125,7 +125,7 @@ session_add_self_custom_tx_evt (transport_connection_t * tc, u8 has_prio)
 
   s = session_get (tc->s_index, tc->thread_index);
   ASSERT (s->thread_index == vlib_get_thread_index ());
-  ASSERT (s->session_state < SESSION_STATE_TRANSPORT_DELETED);
+  ASSERT (s->session_state != SESSION_STATE_TRANSPORT_DELETED);
   if (!(s->flags & SESSION_F_CUSTOM_TX))
     {
       s->flags |= SESSION_F_CUSTOM_TX;
diff --git a/src/vnet/session/session_node.c b/src/vnet/session/session_node.c
index e2e98eba19f..99fa4bdac91 100644
--- a/src/vnet/session/session_node.c
+++ b/src/vnet/session/session_node.c
@@ -1418,6 +1418,8 @@ session_node_cmp_event (session_event_t * e, svm_fifo_t * f)
     case SESSION_IO_EVT_RX:
     case SESSION_IO_EVT_TX:
     case SESSION_IO_EVT_BUILTIN_RX:
+    case SESSION_IO_EVT_BUILTIN_TX:
+    case SESSION_IO_EVT_TX_FLUSH:
       if (e->session_index == f->master_session_index)
 	return 1;
       break;
@@ -1467,8 +1469,7 @@ session_node_lookup_fifo_event (svm_fifo_t * f, session_event_t * e)
       found = session_node_cmp_event (e, f);
       if (found)
 	return 1;
-      if (++index == mq->q->maxsize)
-	index = 0;
+      index = (index + 1) % mq->q->maxsize;
     }
   /*
    * Search pending events vector
@@ -1476,18 +1477,31 @@ session_node_lookup_fifo_event (svm_fifo_t * f, session_event_t * e)
 
   /* *INDENT-OFF* */
   clib_llist_foreach (wrk->event_elts, evt_list,
-                      pool_elt_at_index (wrk->event_elts, wrk->old_head),
+                      pool_elt_at_index (wrk->event_elts, wrk->new_head),
                       elt, ({
     found = session_node_cmp_event (&elt->evt, f);
     if (found)
       {
 	clib_memcpy_fast (e, &elt->evt, sizeof (*e));
-	break;
+	goto done;
       }
+  }));
+  /* *INDENT-ON* */
 
+  /* *INDENT-OFF* */
+  clib_llist_foreach (wrk->event_elts, evt_list,
+                      pool_elt_at_index (wrk->event_elts, wrk->old_head),
+                      elt, ({
+    found = session_node_cmp_event (&elt->evt, f);
+    if (found)
+      {
+	clib_memcpy_fast (e, &elt->evt, sizeof (*e));
+	goto done;
+      }
   }));
   /* *INDENT-ON* */
 
+done:
   return found;
 }
 
diff --git a/src/vnet/session/transport.c b/src/vnet/session/transport.c
index 3e723994752..49a4af71111 100644
--- a/src/vnet/session/transport.c
+++ b/src/vnet/session/transport.c
@@ -566,9 +566,14 @@ u8 *
 format_transport_pacer (u8 * s, va_list * args)
 {
   spacer_t *pacer = va_arg (*args, spacer_t *);
+  vlib_main_t *vm = vlib_get_main ();
+  u64 now, diff;
 
-  s = format (s, "bucket %u tokens/period %.3f last_update %x",
-	      pacer->bucket, pacer->tokens_per_period, pacer->last_update);
+  now = vm->clib_time.last_cpu_time;
+  diff = now - (pacer->last_update << SPACER_CPU_TICKS_PER_PERIOD_SHIFT);
+  s = format (s, "rate %u bucket %u t/p %.3f last_update %.3f",
+	      pacer->bytes_per_sec, pacer->bucket, pacer->tokens_per_period,
+	      diff * vm->clib_time.seconds_per_clock);
   return s;
 }
 
diff --git a/src/vnet/tcp/tcp.c b/src/vnet/tcp/tcp.c
index 5ffb1e27ab8..6ef03dc093d 100644
--- a/src/vnet/tcp/tcp.c
+++ b/src/vnet/tcp/tcp.c
@@ -915,17 +915,26 @@ static u8 *
 format_tcp_congestion (u8 * s, va_list * args)
 {
   tcp_connection_t *tc = va_arg (*args, tcp_connection_t *);
-  u32 indent = format_get_indent (s);
+  u32 indent = format_get_indent (s), prr_space = 0;
 
   s = format (s, "%U ", format_tcp_congestion_status, tc);
   s = format (s, "algo %s cwnd %u ssthresh %u bytes_acked %u\n",
 	      tc->cc_algo->name, tc->cwnd, tc->ssthresh, tc->bytes_acked);
-  s = format (s, "%Ucc space %u prev_cwnd %u prev_ssthresh %u rxt_bytes %u\n",
+  s = format (s, "%Ucc space %u prev_cwnd %u prev_ssthresh %u\n",
 	      format_white_space, indent, tcp_available_cc_snd_space (tc),
-	      tc->prev_cwnd, tc->prev_ssthresh, tc->snd_rxt_bytes);
-  s = format (s, "%Usnd_congestion %u dupack %u limited_transmit %u\n",
+	      tc->prev_cwnd, tc->prev_ssthresh);
+  s = format (s, "%Usnd_cong %u dupack %u limited_tx %u\n",
 	      format_white_space, indent, tc->snd_congestion - tc->iss,
 	      tc->rcv_dupacks, tc->limited_transmit - tc->iss);
+  s = format (s, "%Urxt_bytes %u rxt_delivered %u rxt_head %u rxt_ts %u\n",
+	      format_white_space, indent, tc->snd_rxt_bytes,
+	      tc->rxt_delivered, tc->rxt_head - tc->iss,
+	      tcp_time_now_w_thread (tc->c_thread_index) - tc->snd_rxt_ts);
+  if (tcp_in_fastrecovery (tc))
+    prr_space = tcp_fastrecovery_prr_snd_space (tc);
+  s = format (s, "%Uprr_start %u prr_delivered %u prr space %u\n",
+	      format_white_space, indent, tc->prr_start - tc->iss,
+	      tc->prr_delivered, prr_space);
   return s;
 }
 
@@ -1140,10 +1149,11 @@ format_tcp_scoreboard (u8 * s, va_list * args)
   sack_scoreboard_hole_t *hole;
   u32 indent = format_get_indent (s);
 
-  s = format (s, "sacked %u last_sacked %u lost %u last_lost %u\n",
+  s = format (s, "sacked %u last_sacked %u lost %u last_lost %u"
+	      " rxt_sacked %u\n",
 	      sb->sacked_bytes, sb->last_sacked_bytes, sb->lost_bytes,
-	      sb->last_lost_bytes);
-  s = format (s, "%Ulast_bytes_delivered %u high_sacked %u is_reneging %u\n",
+	      sb->last_lost_bytes, sb->rxt_sacked);
+  s = format (s, "%Ulast_delivered %u high_sacked %u is_reneging %u\n",
 	      format_white_space, indent, sb->last_bytes_delivered,
 	      sb->high_sacked - tc->iss, sb->is_reneging);
   s = format (s, "%Ucur_rxt_hole %u high_rxt %u rescue_rxt %u",
diff --git a/src/vnet/tcp/tcp.h b/src/vnet/tcp/tcp.h
index 1bf32818171..a31c46cdacf 100644
--- a/src/vnet/tcp/tcp.h
+++ b/src/vnet/tcp/tcp.h
@@ -168,7 +168,7 @@ typedef struct _sack_scoreboard
   u32 sacked_bytes;			/**< Number of bytes sacked in sb */
   u32 last_sacked_bytes;		/**< Number of bytes last sacked */
   u32 last_bytes_delivered;		/**< Sack bytes delivered to app */
-  u32 rxt_sacked;			/**< Rxt last delivered */
+  u32 rxt_sacked;			/**< Rxt bytes last delivered */
   u32 high_sacked;			/**< Highest byte sacked (fack) */
   u32 high_rxt;				/**< Highest retransmitted sequence */
   u32 rescue_rxt;			/**< Rescue sequence number */
@@ -226,7 +226,7 @@ sack_scoreboard_hole_t *scoreboard_last_hole (sack_scoreboard_t * sb);
 void scoreboard_clear (sack_scoreboard_t * sb);
 void scoreboard_clear_reneging (sack_scoreboard_t * sb, u32 start, u32 end);
 void scoreboard_init (sack_scoreboard_t * sb);
-void scoreboard_init_high_rxt (sack_scoreboard_t * sb, u32 snd_una);
+void scoreboard_init_rxt (sack_scoreboard_t * sb, u32 snd_una);
 u8 *format_tcp_scoreboard (u8 * s, va_list * args);
 
 #define TCP_BTS_INVALID_INDEX	((u32)~0)
@@ -369,7 +369,9 @@ typedef struct _tcp_connection
   u32 snd_rxt_bytes;	/**< Retransmitted bytes during current cc event */
   u32 snd_rxt_ts;	/**< Timestamp when first packet is retransmitted */
   u32 prr_delivered;	/**< RFC6937 bytes delivered during current event */
+  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 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 */
@@ -957,8 +959,7 @@ tcp_is_lost_fin (tcp_connection_t * tc)
 }
 
 u32 tcp_snd_space (tcp_connection_t * tc);
-//void tcp_cc_init_congestion (tcp_connection_t * tc);
-//void tcp_cc_fastrecovery_clear (tcp_connection_t * tc);
+int tcp_fastrecovery_prr_snd_space (tcp_connection_t * tc);
 
 fib_node_index_t tcp_lookup_rmt_in_fib (tcp_connection_t * tc);
 
diff --git a/src/vnet/tcp/tcp_input.c b/src/vnet/tcp/tcp_input.c
index b49b8e8fd77..f177f5f4753 100755
--- a/src/vnet/tcp/tcp_input.c
+++ b/src/vnet/tcp/tcp_input.c
@@ -629,34 +629,6 @@ tcp_program_dequeue (tcp_worker_ctx_t * wrk, tcp_connection_t * tc)
   tc->burst_acked += tc->bytes_acked;
 }
 
-/**
- * Check if duplicate ack as per RFC5681 Sec. 2
- */
-static u8
-tcp_ack_is_dupack (tcp_connection_t * tc, vlib_buffer_t * b, u32 prev_snd_wnd,
-		   u32 prev_snd_una)
-{
-  return ((vnet_buffer (b)->tcp.ack_number == prev_snd_una)
-	  && seq_gt (tc->snd_nxt, tc->snd_una)
-	  && (vnet_buffer (b)->tcp.seq_end == vnet_buffer (b)->tcp.seq_number)
-	  && (prev_snd_wnd == tc->snd_wnd));
-}
-
-/**
- * Checks if ack is a congestion control event.
- */
-static u8
-tcp_ack_is_cc_event (tcp_connection_t * tc, vlib_buffer_t * b,
-		     u32 prev_snd_wnd, u32 prev_snd_una, u8 * is_dack)
-{
-  /* Check if ack is duplicate. Per RFC 6675, ACKs that SACK new data are
-   * defined to be 'duplicate' */
-  *is_dack = tc->sack_sb.last_sacked_bytes
-    || tcp_ack_is_dupack (tc, b, prev_snd_wnd, prev_snd_una);
-
-  return ((*is_dack || tcp_in_cong_recovery (tc)) && !tcp_is_lost_fin (tc));
-}
-
 #ifndef CLIB_MARCH_VARIANT
 static u32
 scoreboard_hole_index (sack_scoreboard_t * sb, sack_scoreboard_hole_t * hole)
@@ -920,7 +892,7 @@ scoreboard_next_rxt_hole (sack_scoreboard_t * sb,
 }
 
 void
-scoreboard_init_high_rxt (sack_scoreboard_t * sb, u32 snd_una)
+scoreboard_init_rxt (sack_scoreboard_t * sb, u32 snd_una)
 {
   sack_scoreboard_hole_t *hole;
   hole = scoreboard_first_hole (sb);
@@ -975,7 +947,7 @@ scoreboard_clear_reneging (sack_scoreboard_t * sb, u32 start, u32 end)
   last_hole->is_lost = 1;
   sb->tail = scoreboard_hole_index (sb, last_hole);
   sb->high_sacked = start;
-  scoreboard_init_high_rxt (sb, start);
+  scoreboard_init_rxt (sb, start);
 }
 
 #endif /* CLIB_MARCH_VARIANT */
@@ -1203,6 +1175,8 @@ tcp_rcv_sacks (tcp_connection_t * tc, u32 ack)
   ASSERT (sb->last_lost_bytes <= sb->lost_bytes);
   ASSERT ((ack - tc->snd_una) + sb->last_sacked_bytes
 	  - sb->last_bytes_delivered >= sb->rxt_sacked);
+  ASSERT ((ack - tc->snd_una) >= tc->sack_sb.last_bytes_delivered
+	  || (tc->flags & TCP_CONN_FINSNT));
 
   TCP_EVT (TCP_EVT_CC_SCOREBOARD, tc);
 }
@@ -1261,9 +1235,11 @@ tcp_cc_init_congestion (tcp_connection_t * tc)
   tc->snd_rxt_bytes = 0;
   tc->rxt_delivered = 0;
   tc->prr_delivered = 0;
+  tc->prr_start = tc->snd_una;
   tc->prev_ssthresh = tc->ssthresh;
   tc->prev_cwnd = tc->cwnd;
 
+  tc->snd_rxt_ts = tcp_tstamp (tc);
   tcp_cc_congestion (tc);
 
   /* Post retransmit update cwnd to ssthresh and account for the
@@ -1310,8 +1286,29 @@ tcp_should_fastrecover_sack (tcp_connection_t * tc)
 }
 
 static inline u8
-tcp_should_fastrecover (tcp_connection_t * tc)
+tcp_should_fastrecover (tcp_connection_t * tc, u8 has_sack)
 {
+  if (!has_sack)
+    {
+      /* If of of the two conditions lower hold, reset dupacks because
+       * we're probably after timeout (RFC6582 heuristics).
+       * If Cumulative ack does not cover more than congestion threshold,
+       * and:
+       * 1) The following doesn't hold: The congestion window is greater
+       *    than SMSS bytes and the difference between highest_ack
+       *    and prev_highest_ack is at most 4*SMSS bytes
+       * 2) Echoed timestamp in the last non-dup ack does not equal the
+       *    stored timestamp
+       */
+      if (seq_leq (tc->snd_una, tc->snd_congestion)
+	  && ((!(tc->cwnd > tc->snd_mss
+		 && tc->bytes_acked <= 4 * tc->snd_mss))
+	      || (tc->rcv_opts.tsecr != tc->tsecr_last_ack)))
+	{
+	  tc->rcv_dupacks = 0;
+	  return 0;
+	}
+    }
   return ((tc->rcv_dupacks == TCP_DUPACK_THRESHOLD)
 	  || tcp_should_fastrecover_sack (tc));
 }
@@ -1330,6 +1327,7 @@ tcp_cc_recover (tcp_connection_t * tc)
       is_spurious = 1;
     }
 
+  tc->rcv_dupacks = 0;
   tc->prr_delivered = 0;
   tc->rxt_delivered = 0;
   tc->snd_rxt_bytes = 0;
@@ -1342,8 +1340,9 @@ tcp_cc_recover (tcp_connection_t * tc)
   if (tc->sack_sb.sacked_bytes)
     {
       tc->snd_congestion = tc->snd_nxt;
-      if (tcp_opts_sack_permitted (&tc->rcv_opts))
-	scoreboard_init_high_rxt (&tc->sack_sb, tc->snd_una);
+      tc->snd_rxt_ts = tcp_tstamp (tc);
+      tc->prr_start = tc->snd_una;
+      scoreboard_init_rxt (&tc->sack_sb, tc->snd_una);
       tcp_program_retransmit (tc);
       return is_spurious;
     }
@@ -1355,6 +1354,8 @@ 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);
@@ -1387,30 +1388,6 @@ tcp_cc_update (tcp_connection_t * tc, tcp_rate_sample_t * rs)
     tc->snd_congestion = tc->snd_una - 1;
 }
 
-static void
-tcp_update_delivered (tcp_connection_t * tc, u8 is_dack, u8 has_sack)
-{
-  if (has_sack)
-    {
-      ASSERT (tc->bytes_acked >= tc->sack_sb.last_bytes_delivered
-	      || (tc->flags & TCP_CONN_FINSNT));
-
-      tc->rxt_delivered += tc->sack_sb.rxt_sacked;
-      tc->prr_delivered += tc->bytes_acked + tc->sack_sb.last_sacked_bytes
-	- tc->sack_sb.last_bytes_delivered;
-    }
-  else
-    {
-      tcp_fastrecovery_first_on (tc);
-      tc->rxt_delivered = clib_max (tc->rxt_delivered + tc->bytes_acked,
-				    tc->snd_rxt_bytes);
-      if (is_dack)
-	tc->prr_delivered += is_dack;
-      else
-	tc->prr_delivered += tc->bytes_acked - tc->snd_mss * tc->rcv_dupacks;
-    }
-}
-
 /**
  * One function to rule them all ... and in the darkness bind them
  */
@@ -1421,92 +1398,34 @@ tcp_cc_handle_event (tcp_connection_t * tc, tcp_rate_sample_t * rs,
   u8 has_sack = tcp_opts_sack_permitted (&tc->rcv_opts);
 
   /*
-   * Already in fast recovery. Return if no data acked, partial acks
-   * and accounting for segments that have left the network are done
-   * lower.
+   * If not in recovery, figure out if we should enter
    */
-  if (tcp_in_cong_recovery (tc))
+  if (!tcp_in_cong_recovery (tc))
     {
-      if (!has_sack)
-	tc->rcv_dupacks += is_dack;
-
-      if (!tc->bytes_acked)
-	{
-	  tcp_update_delivered (tc, is_dack, has_sack);
-	  tcp_program_retransmit (tc);
-	  if (!has_sack)
-	    tcp_cc_rcv_cong_ack (tc, TCP_CC_DUPACK, rs);
-	  return;
-	}
-    }
-  /*
-   * Duplicate ACK. Check if we should enter fast recovery
-   */
-  else if (is_dack)
-    {
-      TCP_EVT (TCP_EVT_DUPACK_RCVD, tc, 1);
-      ASSERT (tc->snd_una != tc->snd_nxt || tc->sack_sb.last_sacked_bytes);
-
-      /* Heuristic to catch potential late dupacks */
-      if (!tc->sack_sb.sacked_bytes && tc->snd_una == tc->snd_congestion
-	  && timestamp_leq (tc->rcv_opts.tsecr, tc->tsecr_last_ack))
-	return;
+      ASSERT (is_dack);
 
       tc->rcv_dupacks++;
+      TCP_EVT (TCP_EVT_DUPACK_RCVD, tc, 1);
+      tcp_cc_rcv_cong_ack (tc, TCP_CC_DUPACK, rs);
 
-      if (tcp_should_fastrecover (tc))
+      if (tcp_should_fastrecover (tc, has_sack))
 	{
-	  ASSERT (!tcp_in_fastrecovery (tc));
-
 	  tcp_cc_init_congestion (tc);
 
 	  if (has_sack)
-	    scoreboard_init_high_rxt (&tc->sack_sb, tc->snd_una);
+	    scoreboard_init_rxt (&tc->sack_sb, tc->snd_una);
 
 	  tcp_connection_tx_pacer_reset (tc, tc->cwnd, 0 /* start bucket */ );
 	  tcp_program_retransmit (tc);
-	  return;
-	}
-      else
-	{
-	  tcp_cc_rcv_cong_ack (tc, TCP_CC_DUPACK, rs);
-	  return;
-	}
-    }
-  /* Don't allow entry in fast recovery if still in recovery, for now */
-  else if (0 && is_dack && tcp_in_recovery (tc))
-    {
-      /* If of of the two conditions lower hold, reset dupacks because
-       * we're probably after timeout (RFC6582 heuristics).
-       * If Cumulative ack does not cover more than congestion threshold,
-       * and:
-       * 1) The following doesn't hold: The congestion window is greater
-       *    than SMSS bytes and the difference between highest_ack
-       *    and prev_highest_ack is at most 4*SMSS bytes
-       * 2) Echoed timestamp in the last non-dup ack does not equal the
-       *    stored timestamp
-       */
-      if (seq_leq (tc->snd_una, tc->snd_congestion)
-	  && ((!(tc->cwnd > tc->snd_mss
-		 && tc->bytes_acked <= 4 * tc->snd_mss))
-	      || (tc->rcv_opts.tsecr != tc->tsecr_last_ack)))
-	{
-	  tc->rcv_dupacks = 0;
-	  return;
 	}
-    }
 
-  ASSERT (tc->bytes_acked);
-  TCP_EVT (TCP_EVT_CC_PACK, tc);
+      return;
+    }
 
   /*
-   * Legitimate ACK. 1) See if we can exit recovery
+   * Already in recovery. See if we can exit and stop retransmitting
    */
 
-  /* RFC6675: If the incoming ACK is a cumulative acknowledgment,
-   * reset dupacks to 0. Also needed if in congestion recovery */
-  tc->rcv_dupacks = 0;
-
   if (seq_geq (tc->snd_una, tc->snd_congestion))
     {
       /* If spurious return, we've already updated everything */
@@ -1522,21 +1441,88 @@ tcp_cc_handle_event (tcp_connection_t * tc, tcp_rate_sample_t * rs,
     }
 
   /*
-   * Legitimate ACK. 2) If PARTIAL ACK try to retransmit
+   * Process (re)transmit feedback. Output path uses this to decide how much
+   * more data to release into the network
    */
+  if (has_sack)
+    {
+      tc->rxt_delivered += tc->sack_sb.rxt_sacked;
+      tc->prr_delivered += tc->bytes_acked + tc->sack_sb.last_sacked_bytes
+	- tc->sack_sb.last_bytes_delivered;
+
+      tcp_program_retransmit (tc);
+    }
+  else
+    {
+      if (is_dack)
+	{
+	  tc->rcv_dupacks += 1;
+	  TCP_EVT (TCP_EVT_DUPACK_RCVD, tc, 1);
+	}
+      tc->rxt_delivered = clib_max (tc->rxt_delivered + tc->bytes_acked,
+				    tc->snd_rxt_bytes);
+      if (is_dack)
+	tc->prr_delivered += 1;
+      else
+	tc->prr_delivered += tc->bytes_acked - tc->snd_mss * tc->rcv_dupacks;
 
-  /* Remove retransmitted bytes that have been delivered */
-  tcp_update_delivered (tc, is_dack, has_sack);
+      /* If partial ack, assume that the first un-acked segment was lost */
+      if (tc->bytes_acked || tc->rcv_dupacks == TCP_DUPACK_THRESHOLD)
+	tcp_fastrecovery_first_on (tc);
+
+      tcp_program_retransmit (tc);
+    }
+
+  /*
+   * Notify cc of the event
+   */
+
+  if (!tc->bytes_acked)
+    {
+      tcp_cc_rcv_cong_ack (tc, TCP_CC_DUPACK, rs);
+      return;
+    }
+
+  /* RFC6675: If the incoming ACK is a cumulative acknowledgment,
+   * reset dupacks to 0. Also needed if in congestion recovery */
+  tc->rcv_dupacks = 0;
 
   if (tcp_in_recovery (tc))
     tcp_cc_rcv_ack (tc, rs);
   else
     tcp_cc_rcv_cong_ack (tc, TCP_CC_PARTIALACK, rs);
+}
 
-  /*
-   * Since this was a partial ack, try to retransmit some more data
-   */
-  tcp_program_retransmit (tc);
+/**
+ * Check if duplicate ack as per RFC5681 Sec. 2
+ */
+always_inline u8
+tcp_ack_is_dupack (tcp_connection_t * tc, vlib_buffer_t * b, u32 prev_snd_wnd,
+		   u32 prev_snd_una)
+{
+  return ((vnet_buffer (b)->tcp.ack_number == prev_snd_una)
+	  && seq_gt (tc->snd_nxt, tc->snd_una)
+	  && (vnet_buffer (b)->tcp.seq_end == vnet_buffer (b)->tcp.seq_number)
+	  && (prev_snd_wnd == tc->snd_wnd));
+}
+
+/**
+ * Checks if ack is a congestion control event.
+ */
+static u8
+tcp_ack_is_cc_event (tcp_connection_t * tc, vlib_buffer_t * b,
+		     u32 prev_snd_wnd, u32 prev_snd_una, u8 * is_dack)
+{
+  /* Check if ack is duplicate. Per RFC 6675, ACKs that SACK new data are
+   * defined to be 'duplicate' as well */
+  *is_dack = tc->sack_sb.last_sacked_bytes
+    || tcp_ack_is_dupack (tc, b, prev_snd_wnd, prev_snd_una);
+
+  /* If reneging, wait for timer based retransmits */
+  if (PREDICT_FALSE (tcp_is_lost_fin (tc) || tc->sack_sb.is_reneging))
+    return 0;
+
+  return (*is_dack || tcp_in_cong_recovery (tc));
 }
 
 /**
diff --git a/src/vnet/tcp/tcp_output.c b/src/vnet/tcp/tcp_output.c
index a7b0e398d36..79866aff03a 100644
--- a/src/vnet/tcp/tcp_output.c
+++ b/src/vnet/tcp/tcp_output.c
@@ -1599,7 +1599,7 @@ tcp_timer_retransmit_handler (u32 tc_index)
 	}
 
       if (tcp_opts_sack_permitted (&tc->rcv_opts))
-	scoreboard_init_high_rxt (&tc->sack_sb, tc->snd_una + tc->snd_mss);
+	scoreboard_init_rxt (&tc->sack_sb, tc->snd_una + n_bytes);
 
       tcp_program_retransmit (tc);
     }
@@ -1858,7 +1858,7 @@ done:
 /**
  * Estimate send space using proportional rate reduction (RFC6937)
  */
-static int
+int
 tcp_fastrecovery_prr_snd_space (tcp_connection_t * tc)
 {
   u32 pipe, prr_out;
@@ -1874,13 +1874,24 @@ tcp_fastrecovery_prr_snd_space (tcp_connection_t * tc)
     }
   else
     {
-      int limit = tc->prr_delivered - prr_out + tc->snd_mss;
+      int limit;
+      limit = clib_max ((int) (tc->prr_delivered - prr_out), 0) + tc->snd_mss;
       space = clib_min (tc->ssthresh - pipe, limit);
     }
   space = clib_max (space, prr_out ? 0 : tc->snd_mss);
   return space;
 }
 
+static inline u8
+tcp_retransmit_should_retry_head (tcp_connection_t * tc,
+				  sack_scoreboard_t * sb)
+{
+  u32 tx_adv_sack = sb->high_sacked - tc->snd_congestion;
+  f64 rr = (f64) tc->ssthresh / tc->prev_cwnd;
+
+  return (tx_adv_sack > (tc->snd_una - tc->prr_start) * rr);
+}
+
 #define scoreboard_rescue_rxt_valid(_sb, _tc)			\
     (seq_geq (_sb->rescue_rxt, _tc->snd_una) 			\
 	&& seq_leq (_sb->rescue_rxt, _tc->snd_congestion))
@@ -1917,8 +1928,31 @@ tcp_retransmit_sack (tcp_worker_ctx_t * wrk, tcp_connection_t * tc,
       return 0;
     }
 
-  TCP_EVT (TCP_EVT_CC_EVT, tc, 0);
   sb = &tc->sack_sb;
+
+  /* Check if snd_una is a lost retransmit */
+  if (seq_gt (sb->high_sacked, tc->snd_congestion)
+      && tc->rxt_head != tc->snd_una
+      && tcp_retransmit_should_retry_head (tc, sb))
+    {
+      n_written = tcp_prepare_retransmit_segment (wrk, tc, 0, tc->snd_mss,
+						  &b);
+      if (!n_written)
+	{
+	  tcp_program_retransmit (tc);
+	  goto done;
+	}
+      bi = vlib_get_buffer_index (vm, b);
+      tcp_enqueue_to_output (wrk, b, bi, tc->c_is_ip4);
+      n_segs = 1;
+
+      tc->rxt_head = tc->snd_una;
+      tc->rxt_delivered += n_written;
+      tc->prr_delivered += n_written;
+      ASSERT (tc->rxt_delivered <= tc->snd_rxt_bytes);
+    }
+
+  TCP_EVT (TCP_EVT_CC_EVT, tc, 0);
   hole = scoreboard_get_hole (sb, sb->cur_rxt_hole);
 
   max_deq = transport_max_tx_dequeue (&tc->connection);
@@ -1931,7 +1965,7 @@ tcp_retransmit_sack (tcp_worker_ctx_t * wrk, tcp_connection_t * tc,
       if (!hole)
 	{
 	  /* We are out of lost holes to retransmit so send some new data. */
-	  if (max_deq)
+	  if (max_deq > tc->snd_mss)
 	    {
 	      u32 n_segs_new, av_window;
 	      av_window = tc->snd_wnd - (tc->snd_nxt - tc->snd_una);