tcp: improve handling of snd_nxt

- avoid changing snd_nxt when doing fast retransmits
- use snd_una_max only to keep track of the max seq number sent
- simplify future ack testing

Change-Id: I3580ad3aefe30128486c3375d0ac3f3f62c04c5e
Signed-off-by: Florin Coras <fcoras@cisco.com>

5 files changed, 93 insertions, 126 deletions

diff --git a/src/plugins/unittest/tcp_test.c b/src/plugins/unittest/tcp_test.c
index c144de4398d..fb01adb6c1f 100644
--- a/src/plugins/unittest/tcp_test.c
+++ b/src/plugins/unittest/tcp_test.c
@@ -239,7 +239,7 @@ tcp_test_sack_rx (vlib_main_t * vm, unformat_input_t * input)
    */
   vec_reset_length (tc->rcv_opts.sacks);
   tc->snd_una += sb->snd_una_adv;
-  tc->snd_una_max = 1900;
+  tc->snd_nxt = tc->snd_una_max = 1900;
   for (i = 0; i < 5; i++)
     {
       block.start = i * 100 + 1200;
@@ -416,7 +416,7 @@ tcp_test_sack_rx (vlib_main_t * vm, unformat_input_t * input)
    * snd_una = 1000 and snd_una_max = 1600
    */
   tc->snd_una = 1000;
-  tc->snd_una_max = 1600;
+  tc->snd_nxt = tc->snd_una_max = 1600;
   vec_reset_length (tc->rcv_opts.sacks);
   block.start = 1200;
   block.end = 1500;
diff --git a/src/vnet/tcp/tcp.c b/src/vnet/tcp/tcp.c
index 215bcbb159e..39d683c78e7 100644
--- a/src/vnet/tcp/tcp.c
+++ b/src/vnet/tcp/tcp.c
@@ -1164,7 +1164,7 @@ tcp_session_flush_data (transport_connection_t * tconn)
   if (tc->flags & TCP_CONN_PSH_PENDING)
     return;
   tc->flags |= TCP_CONN_PSH_PENDING;
-  tc->psh_seq = tc->snd_una_max + transport_max_tx_dequeue (tconn) - 1;
+  tc->psh_seq = tc->snd_una + transport_max_tx_dequeue (tconn) - 1;
 }
 
 /* *INDENT-OFF* */
@@ -1296,7 +1296,7 @@ tcp_timer_waitclose_handler (u32 conn_index)
        * and switch to LAST_ACK. */
       tcp_cong_recovery_off (tc);
       /* Make sure we don't try to send unsent data */
-      tc->snd_una_max = tc->snd_nxt = tc->snd_una;
+      tc->snd_nxt = tc->snd_una;
       tcp_send_fin (tc);
       tcp_connection_set_state (tc, TCP_STATE_LAST_ACK);
 
diff --git a/src/vnet/tcp/tcp.h b/src/vnet/tcp/tcp.h
index 07ceb5a67a6..9e2c6d8c4f5 100644
--- a/src/vnet/tcp/tcp.h
+++ b/src/vnet/tcp/tcp.h
@@ -680,7 +680,7 @@ tcp_flight_size (const tcp_connection_t * tc)
 {
   int flight_size;
 
-  flight_size = (int) (tc->snd_una_max - tc->snd_una) - tcp_bytes_out (tc)
+  flight_size = (int) (tc->snd_nxt - tc->snd_una) - tcp_bytes_out (tc)
     + tc->snd_rxt_bytes;
 
   if (flight_size < 0)
@@ -924,7 +924,7 @@ tcp_persist_timer_reset (tcp_connection_t * tc)
 always_inline void
 tcp_retransmit_timer_update (tcp_connection_t * tc)
 {
-  if (tc->snd_una == tc->snd_una_max)
+  if (tc->snd_una == tc->snd_nxt)
     {
       tcp_retransmit_timer_reset (tc);
       if (tc->snd_wnd < tc->snd_mss)
diff --git a/src/vnet/tcp/tcp_input.c b/src/vnet/tcp/tcp_input.c
index 5235c20a565..c27631f3965 100644
--- a/src/vnet/tcp/tcp_input.c
+++ b/src/vnet/tcp/tcp_input.c
@@ -409,7 +409,7 @@ tcp_rcv_ack_is_acceptable (tcp_connection_t * tc0, vlib_buffer_t * tb0)
 {
   /* SND.UNA =< SEG.ACK =< SND.NXT */
   return (seq_leq (tc0->snd_una, vnet_buffer (tb0)->tcp.ack_number)
-	  && seq_leq (vnet_buffer (tb0)->tcp.ack_number, tc0->snd_nxt));
+	  && seq_leq (vnet_buffer (tb0)->tcp.ack_number, tc0->snd_una_max));
 }
 
 /**
@@ -604,7 +604,7 @@ 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_una_max, tc->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));
 }
@@ -913,7 +913,7 @@ tcp_scoreboard_is_sane_post_recovery (tcp_connection_t * tc)
   sack_scoreboard_hole_t *hole;
   hole = scoreboard_first_hole (&tc->sack_sb);
   return (!hole || (seq_geq (hole->start, tc->snd_una)
-		    && seq_lt (hole->end, tc->snd_una_max)));
+		    && seq_lt (hole->end, tc->snd_nxt)));
 }
 
 #ifndef CLIB_MARCH_VARIANT
@@ -943,8 +943,8 @@ tcp_rcv_sacks (tcp_connection_t * tc, u32 ack)
       if (seq_lt (blk->start, blk->end)
 	  && seq_gt (blk->start, tc->snd_una)
 	  && seq_gt (blk->start, ack)
-	  && seq_lt (blk->start, tc->snd_una_max)
-	  && seq_leq (blk->end, tc->snd_una_max))
+	  && seq_lt (blk->start, tc->snd_nxt)
+	  && seq_leq (blk->end, tc->snd_nxt))
 	{
 	  blk++;
 	  continue;
@@ -979,7 +979,7 @@ tcp_rcv_sacks (tcp_connection_t * tc, u32 ack)
     {
       /* If no holes, insert the first that covers all outstanding bytes */
       last_hole = scoreboard_insert_hole (sb, TCP_INVALID_SACK_HOLE_INDEX,
-					  tc->snd_una, tc->snd_una_max);
+					  tc->snd_una, tc->snd_nxt);
       sb->tail = scoreboard_hole_index (sb, last_hole);
       tmp = tc->rcv_opts.sacks[vec_len (tc->rcv_opts.sacks) - 1];
       sb->high_sacked = tmp.end;
@@ -990,17 +990,17 @@ tcp_rcv_sacks (tcp_connection_t * tc, u32 ack)
        * last hole end */
       tmp = tc->rcv_opts.sacks[vec_len (tc->rcv_opts.sacks) - 1];
       last_hole = scoreboard_last_hole (sb);
-      if (seq_gt (tc->snd_una_max, last_hole->end))
+      if (seq_gt (tc->snd_nxt, last_hole->end))
 	{
 	  if (seq_geq (last_hole->start, sb->high_sacked))
 	    {
-	      last_hole->end = tc->snd_una_max;
+	      last_hole->end = tc->snd_nxt;
 	    }
 	  /* New hole after high sacked block */
-	  else if (seq_lt (sb->high_sacked, tc->snd_una_max))
+	  else if (seq_lt (sb->high_sacked, tc->snd_nxt))
 	    {
 	      scoreboard_insert_hole (sb, sb->tail, sb->high_sacked,
-				      tc->snd_una_max);
+				      tc->snd_nxt);
 	    }
 	}
       /* Keep track of max byte sacked for when the last hole
@@ -1078,8 +1078,7 @@ tcp_rcv_sacks (tcp_connection_t * tc, u32 ack)
   if (pool_elts (sb->holes) == 1)
     {
       hole = scoreboard_first_hole (sb);
-      if (hole->start == ack + sb->snd_una_adv
-	  && hole->end == tc->snd_una_max)
+      if (hole->start == ack + sb->snd_una_adv && hole->end == tc->snd_nxt)
 	scoreboard_remove_hole (sb, hole);
     }
 
@@ -1088,8 +1087,8 @@ tcp_rcv_sacks (tcp_connection_t * tc, u32 ack)
     - (old_sacked_bytes - sb->last_bytes_delivered);
   ASSERT (sb->last_sacked_bytes <= sb->sacked_bytes || tcp_in_recovery (tc));
   ASSERT (sb->sacked_bytes == 0 || tcp_in_recovery (tc)
-	  || sb->sacked_bytes < tc->snd_una_max - seq_max (tc->snd_una, ack));
-  ASSERT (sb->last_sacked_bytes + sb->lost_bytes <= tc->snd_una_max
+	  || sb->sacked_bytes < tc->snd_nxt - seq_max (tc->snd_una, ack));
+  ASSERT (sb->last_sacked_bytes + sb->lost_bytes <= tc->snd_nxt
 	  - seq_max (tc->snd_una, ack) || tcp_in_recovery (tc));
   ASSERT (sb->head == TCP_INVALID_SACK_HOLE_INDEX || tcp_in_recovery (tc)
 	  || sb->holes[sb->head].start == ack + sb->snd_una_adv);
@@ -1146,7 +1145,7 @@ void
 tcp_cc_init_congestion (tcp_connection_t * tc)
 {
   tcp_fastrecovery_on (tc);
-  tc->snd_congestion = tc->snd_una_max;
+  tc->snd_congestion = tc->snd_nxt;
   tc->cwnd_acc_bytes = 0;
   tc->snd_rxt_bytes = 0;
   tc->prev_ssthresh = tc->ssthresh;
@@ -1162,7 +1161,6 @@ tcp_cc_recovery_exit (tcp_connection_t * tc)
   tc->rto_boff = 0;
   tcp_update_rto (tc);
   tc->snd_rxt_ts = 0;
-  tc->snd_nxt = tc->snd_una_max;
   tc->rtt_ts = 0;
   tcp_recovery_off (tc);
   TCP_EVT_DBG (TCP_EVT_CC_EVT, tc, 3);
@@ -1175,7 +1173,6 @@ tcp_cc_fastrecovery_exit (tcp_connection_t * tc)
   tc->cc_algo->recovered (tc);
   tc->snd_rxt_bytes = 0;
   tc->rcv_dupacks = 0;
-  tc->snd_nxt = tc->snd_una_max;
   tc->snd_rxt_bytes = 0;
   tc->rtt_ts = 0;
 
@@ -1191,12 +1188,16 @@ tcp_cc_congestion_undo (tcp_connection_t * tc)
 {
   tc->cwnd = tc->prev_cwnd;
   tc->ssthresh = tc->prev_ssthresh;
-  tc->snd_nxt = tc->snd_una_max;
   tc->rcv_dupacks = 0;
   if (tcp_in_recovery (tc))
-    tcp_cc_recovery_exit (tc);
+    {
+      tcp_cc_recovery_exit (tc);
+      tc->snd_nxt = seq_max (tc->snd_nxt, tc->snd_congestion);
+    }
   else if (tcp_in_fastrecovery (tc))
-    tcp_cc_fastrecovery_exit (tc);
+    {
+      tcp_cc_fastrecovery_exit (tc);
+    }
   ASSERT (tc->rto_boff == 0);
   TCP_EVT_DBG (TCP_EVT_CC_EVT, tc, 5);
 }
@@ -1374,8 +1375,7 @@ tcp_cc_handle_event (tcp_connection_t * tc, u32 is_dack)
   else if (is_dack && !tcp_in_recovery (tc))
     {
       TCP_EVT_DBG (TCP_EVT_DUPACK_RCVD, tc, 1);
-      ASSERT (tc->snd_una != tc->snd_una_max
-	      || tc->sack_sb.last_sacked_bytes);
+      ASSERT (tc->snd_una != tc->snd_nxt || tc->sack_sb.last_sacked_bytes);
 
       tc->rcv_dupacks++;
 
@@ -1482,8 +1482,6 @@ partial_ack:
 	  return;
 	}
 
-      tc->snd_nxt = tc->snd_una_max;
-
       /* Treat as congestion avoidance ack */
       tcp_cc_rcv_ack (tc);
       return;
@@ -1565,38 +1563,18 @@ tcp_rcv_ack (tcp_worker_ctx_t * wrk, tcp_connection_t * tc, vlib_buffer_t * b,
   /* If the ACK acks something not yet sent (SEG.ACK > SND.NXT) */
   if (PREDICT_FALSE (seq_gt (vnet_buffer (b)->tcp.ack_number, tc->snd_nxt)))
     {
-      /* When we entered cong recovery, we reset snd_nxt to snd_una. Seems
-       * peer still has the data so accept the ack */
-      if (tcp_in_cong_recovery (tc)
-	  && seq_leq (vnet_buffer (b)->tcp.ack_number,
-		      tc->snd_una + tc->snd_wnd))
+      /* We've probably entered recovery and the peer still has some
+       * of the data we've sent. Update snd_nxt and accept the ack */
+      if (seq_leq (vnet_buffer (b)->tcp.ack_number, tc->snd_una_max))
 	{
 	  tc->snd_nxt = vnet_buffer (b)->tcp.ack_number;
-	  if (seq_gt (tc->snd_nxt, tc->snd_una_max))
-	    tc->snd_una_max = tc->snd_nxt;
 	  goto process_ack;
 	}
 
-      /* If we have outstanding data and this is within the window, accept it,
-       * probably retransmit has timed out. Otherwise ACK segment and then
-       * drop it */
-      if (seq_gt (vnet_buffer (b)->tcp.ack_number, tc->snd_una_max))
-	{
-	  tcp_program_ack (wrk, tc);
-	  *error = TCP_ERROR_ACK_FUTURE;
-	  TCP_EVT_DBG (TCP_EVT_ACK_RCV_ERR, tc, 0,
-		       vnet_buffer (b)->tcp.ack_number);
-	  return -1;
-	}
-
-      TCP_EVT_DBG (TCP_EVT_ACK_RCV_ERR, tc, 2,
+      *error = TCP_ERROR_ACK_FUTURE;
+      TCP_EVT_DBG (TCP_EVT_ACK_RCV_ERR, tc, 0,
 		   vnet_buffer (b)->tcp.ack_number);
-
-      tc->snd_nxt = vnet_buffer (b)->tcp.ack_number;
-      if (seq_gt (tc->snd_nxt, tc->snd_una_max))
-	tc->snd_una_max = tc->snd_nxt;
-
-      goto process_ack;
+      return -1;
     }
 
   /* If old ACK, probably it's an old dupack */
@@ -1611,10 +1589,11 @@ tcp_rcv_ack (tcp_worker_ctx_t * wrk, tcp_connection_t * tc, vlib_buffer_t * b,
       return 0;
     }
 
+process_ack:
+
   /*
    * Looks okay, process feedback
    */
-process_ack:
   if (tcp_opts_sack_permitted (&tc->rcv_opts))
     tcp_rcv_sacks (tc, vnet_buffer (b)->tcp.ack_number);
 
@@ -1767,13 +1746,7 @@ tcp_update_sack_list (tcp_connection_t * tc, u32 start, u32 end)
 
       /* Save to new SACK list if we have space. */
       if (vec_len (new_list) < TCP_MAX_SACK_BLOCKS)
-	{
-	  vec_add1 (new_list, tc->snd_sacks[i]);
-	}
-      else
-	{
-	  clib_warning ("sack discarded");
-	}
+	vec_add1 (new_list, tc->snd_sacks[i]);
     }
 
   ASSERT (vec_len (new_list) <= TCP_MAX_SACK_BLOCKS);
@@ -2799,7 +2772,7 @@ tcp46_rcv_process_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
 		tcp_send_fin (tc0);
 	    }
 	  /* If FIN is ACKed */
-	  else if (tc0->snd_una == tc0->snd_una_max)
+	  else if (tc0->snd_una == tc0->snd_nxt)
 	    {
 	      tcp_connection_set_state (tc0, TCP_STATE_FIN_WAIT_2);
 
@@ -2845,8 +2818,21 @@ tcp46_rcv_process_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
 	  /* In addition to the processing for the ESTABLISHED state, if
 	   * the ACK acknowledges our FIN then enter the TIME-WAIT state,
 	   * otherwise ignore the segment. */
-	  if (tcp_rcv_ack (wrk, tc0, b0, tcp0, &error0))
-	    goto drop;
+	  if (!tcp_rcv_ack_is_acceptable (tc0, b0))
+	    {
+	      error0 = TCP_ERROR_ACK_INVALID;
+	      goto drop;
+	    }
+
+	  error0 = TCP_ERROR_ACK_OK;
+	  tc0->snd_una = vnet_buffer (b0)->tcp.ack_number;
+	  /* Ack moved snd_una beyond snd_nxt so reprogram fin */
+	  if (seq_gt (tc0->snd_una, tc0->snd_nxt))
+	    {
+	      tc0->snd_nxt = tc0->snd_una;
+	      tc0->flags &= ~TCP_CONN_FINSNT;
+	      goto drop;
+	    }
 
 	  tcp_connection_timers_reset (tc0);
 	  tcp_connection_set_state (tc0, TCP_STATE_TIME_WAIT);
@@ -2867,7 +2853,7 @@ tcp46_rcv_process_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
 	  error0 = TCP_ERROR_ACK_OK;
 	  tc0->snd_una = vnet_buffer (b0)->tcp.ack_number;
 	  /* Apparently our ACK for the peer's FIN was lost */
-	  if (is_fin0 && tc0->snd_una != tc0->snd_una_max)
+	  if (is_fin0 && tc0->snd_una != tc0->snd_nxt)
 	    {
 	      tcp_send_fin (tc0);
 	      goto drop;
@@ -2971,7 +2957,7 @@ tcp46_rcv_process_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
 		  tcp_fastrecovery_off (tc0);
 		  tcp_recovery_off (tc0);
 		  tcp_connection_timers_reset (tc0);
-		  tc0->snd_nxt = tc0->snd_una_max = tc0->snd_una;
+		  tc0->snd_nxt = tc0->snd_una;
 		}
 	      tcp_send_fin (tc0);
 	    }
diff --git a/src/vnet/tcp/tcp_output.c b/src/vnet/tcp/tcp_output.c
index ce7b50ac3e6..4b38649c171 100644
--- a/src/vnet/tcp/tcp_output.c
+++ b/src/vnet/tcp/tcp_output.c
@@ -183,7 +183,7 @@ tcp_update_rcv_wnd (tcp_connection_t * tc)
 /**
  * Compute and return window to advertise, scaled as per RFC1323
  */
-static u32
+static inline u32
 tcp_window_to_advertise (tcp_connection_t * tc, tcp_state_t state)
 {
   if (state < TCP_STATE_ESTABLISHED)
@@ -1052,14 +1052,14 @@ tcp_send_fin (tcp_connection_t * tc)
 
   fin_snt = tc->flags & TCP_CONN_FINSNT;
   if (fin_snt)
-    tc->snd_nxt = tc->snd_una;
+    tc->snd_nxt -= 1;
 
   if (PREDICT_FALSE (!vlib_buffer_alloc (vm, &bi, 1)))
     {
       /* Out of buffers so program fin retransmit ASAP */
       tcp_timer_update (tc, TCP_TIMER_RETRANSMIT, 1);
       if (fin_snt)
-	tc->snd_nxt = tc->snd_una_max;
+	tc->snd_nxt += 1;
       else
 	/* Make sure retransmit retries a fin not data */
 	tc->flags |= TCP_CONN_FINSNT;
@@ -1072,18 +1072,13 @@ tcp_send_fin (tcp_connection_t * tc)
   tcp_make_fin (tc, b);
   tcp_enqueue_to_output_now (wrk, b, bi, tc->c_is_ip4);
   TCP_EVT_DBG (TCP_EVT_FIN_SENT, tc);
-
+  /* Account for the FIN */
+  tc->snd_nxt += 1;
   if (!fin_snt)
     {
       tc->flags |= TCP_CONN_FINSNT;
       tc->flags &= ~TCP_CONN_FINPNDG;
-      /* Account for the FIN */
-      tc->snd_una_max += 1;
-      tc->snd_nxt = tc->snd_una_max;
-    }
-  else
-    {
-      tc->snd_nxt = tc->snd_una_max;
+      tc->snd_una_max = seq_max (tc->snd_una_max, tc->snd_nxt);
     }
 }
 
@@ -1092,8 +1087,8 @@ tcp_send_fin (tcp_connection_t * tc)
  * for segments with data, not for 'control' packets.
  */
 always_inline void
-tcp_push_hdr_i (tcp_connection_t * tc, vlib_buffer_t * b,
-		tcp_state_t next_state, u8 compute_opts, u8 maybe_burst)
+tcp_push_hdr_i (tcp_connection_t * tc, vlib_buffer_t * b, u32 snd_nxt,
+		u8 compute_opts, u8 maybe_burst, u8 update_snd_nxt)
 {
   u8 tcp_hdr_opts_len, flags = TCP_FLAG_ACK;
   u32 advertise_wnd, data_len;
@@ -1115,15 +1110,15 @@ tcp_push_hdr_i (tcp_connection_t * tc, vlib_buffer_t * b,
   if (maybe_burst)
     advertise_wnd = tc->rcv_wnd >> tc->rcv_wscale;
   else
-    advertise_wnd = tcp_window_to_advertise (tc, next_state);
+    advertise_wnd = tcp_window_to_advertise (tc, TCP_STATE_ESTABLISHED);
 
   if (PREDICT_FALSE (tc->flags & TCP_CONN_PSH_PENDING))
     {
-      if (seq_geq (tc->psh_seq, tc->snd_nxt)
-	  && seq_lt (tc->psh_seq, tc->snd_nxt + data_len))
+      if (seq_geq (tc->psh_seq, snd_nxt)
+	  && seq_lt (tc->psh_seq, snd_nxt + data_len))
 	flags |= TCP_FLAG_PSH;
     }
-  th = vlib_buffer_push_tcp (b, tc->c_lcl_port, tc->c_rmt_port, tc->snd_nxt,
+  th = vlib_buffer_push_tcp (b, tc->c_lcl_port, tc->c_rmt_port, snd_nxt,
 			     tc->rcv_nxt, tcp_hdr_opts_len, flags,
 			     advertise_wnd);
 
@@ -1143,7 +1138,8 @@ tcp_push_hdr_i (tcp_connection_t * tc, vlib_buffer_t * b,
    * Update connection variables
    */
 
-  tc->snd_nxt += data_len;
+  if (update_snd_nxt)
+    tc->snd_nxt += data_len;
   tc->rcv_las = tc->rcv_nxt;
 
   TCP_EVT_DBG (TCP_EVT_PKTIZE, tc);
@@ -1153,9 +1149,9 @@ u32
 tcp_session_push_header (transport_connection_t * tconn, vlib_buffer_t * b)
 {
   tcp_connection_t *tc = (tcp_connection_t *) tconn;
-  tcp_push_hdr_i (tc, b, TCP_STATE_ESTABLISHED, /* compute opts */ 0,
-		  /* burst */ 1);
-  tc->snd_una_max = tc->snd_nxt;
+  tcp_push_hdr_i (tc, b, tc->snd_nxt, /* compute opts */ 0, /* burst */ 1,
+		  /* update_snd_nxt */ 1);
+  tc->snd_una_max = seq_max (tc->snd_nxt, tc->snd_una_max);
   ASSERT (seq_leq (tc->snd_una_max, tc->snd_una + tc->snd_wnd));
   tcp_validate_txf_size (tc, tc->snd_una_max - tc->snd_una);
   /* If not tracking an ACK, start tracking */
@@ -1322,9 +1318,8 @@ tcp_prepare_segment (tcp_worker_ctx_t * wrk, tcp_connection_t * tc,
 					    max_deq_bytes);
       ASSERT (n_bytes == max_deq_bytes);
       b[0]->current_length = n_bytes;
-      tcp_push_hdr_i (tc, *b, tc->state, /* compute opts */ 0, /* burst */ 0);
-      if (seq_gt (tc->snd_nxt, tc->snd_una_max))
-	tc->snd_una_max = tc->snd_nxt;
+      tcp_push_hdr_i (tc, *b, tc->snd_una + offset, /* compute opts */ 0,
+		      /* burst */ 0, /* update_snd_nxt */ 0);
     }
   /* Split mss into multiple buffers */
   else
@@ -1381,9 +1376,8 @@ tcp_prepare_segment (tcp_worker_ctx_t * wrk, tcp_connection_t * tc,
 	  b[0]->total_length_not_including_first_buffer += n_peeked;
 	}
 
-      tcp_push_hdr_i (tc, *b, tc->state, /* compute opts */ 0, /* burst */ 0);
-      if (seq_gt (tc->snd_nxt, tc->snd_una_max))
-	tc->snd_una_max = tc->snd_nxt;
+      tcp_push_hdr_i (tc, *b, tc->snd_una + offset, /* compute opts */ 0,
+		      /* burst */ 0, /* update_snd_nxt */ 0);
 
       if (PREDICT_FALSE (n_bufs))
 	{
@@ -1475,7 +1469,7 @@ tcp_rxt_timeout_cc (tcp_connection_t * tc)
   /* Start again from the beginning */
   tc->cc_algo->congestion (tc);
   tc->cwnd = tcp_loss_wnd (tc);
-  tc->snd_congestion = tc->snd_una_max;
+  tc->snd_congestion = tc->snd_nxt;
   tc->rtt_ts = 0;
   tc->cwnd_acc_bytes = 0;
   tcp_connection_tx_pacer_reset (tc, tc->cwnd, 2 * tc->snd_mss);
@@ -1527,7 +1521,7 @@ tcp_timer_retransmit_handler_i (u32 index, u8 is_syn)
 
       /* Shouldn't be here. This condition is tricky because it has to take
        * into account boff > 0 due to persist timeout. */
-      if ((tc->rto_boff == 0 && tc->snd_una == tc->snd_una_max)
+      if ((tc->rto_boff == 0 && tc->snd_una == tc->snd_nxt)
 	  || (tc->rto_boff > 0 && seq_geq (tc->snd_una, tc->snd_congestion)
 	      && !tcp_flight_size (tc)))
 	{
@@ -1555,10 +1549,9 @@ tcp_timer_retransmit_handler_i (u32 index, u8 is_syn)
 	scoreboard_clear (&tc->sack_sb);
 
       /* If we've sent beyond snd_congestion, update it */
-      if (seq_gt (tc->snd_una_max, tc->snd_congestion))
-	tc->snd_congestion = tc->snd_una_max;
+      tc->snd_congestion = seq_max (tc->snd_nxt, tc->snd_congestion);
 
-      tc->snd_una_max = tc->snd_nxt = tc->snd_una;
+      tc->snd_nxt = tc->snd_una;
       tc->rto = clib_min (tc->rto << 1, TCP_RTO_MAX);
 
       /* Send one segment. Note that n_bytes may be zero due to buffer
@@ -1695,7 +1688,7 @@ tcp_timer_persist_handler (u32 index)
     return;
 
   available_bytes = transport_max_tx_dequeue (&tc->connection);
-  offset = tc->snd_una_max - tc->snd_una;
+  offset = tc->snd_nxt - tc->snd_una;
 
   /* Reprogram persist if no new bytes available to send. We may have data
    * next time */
@@ -1737,8 +1730,9 @@ tcp_timer_persist_handler (u32 index)
 			   || tc->snd_nxt == tc->snd_una_max
 			   || tc->rto_boff > 1));
 
-  tcp_push_hdr_i (tc, b, tc->state, /* compute opts */ 0, /* burst */ 0);
-  tc->snd_una_max = tc->snd_nxt;
+  tcp_push_hdr_i (tc, b, tc->snd_nxt, /* compute opts */ 0,
+		  /* burst */ 0, /* update_snd_nxt */ 1);
+  tc->snd_una_max = seq_max (tc->snd_nxt, tc->snd_una_max);
   tcp_validate_txf_size (tc, tc->snd_una_max - tc->snd_una);
   tcp_enqueue_to_output (wrk, b, bi, tc->c_is_ip4);
 
@@ -1752,12 +1746,9 @@ tcp_timer_persist_handler (u32 index)
 int
 tcp_retransmit_first_unacked (tcp_worker_ctx_t * wrk, tcp_connection_t * tc)
 {
-  u32 bi, old_snd_nxt, n_bytes;
   vlib_main_t *vm = wrk->vm;
   vlib_buffer_t *b;
-
-  old_snd_nxt = tc->snd_nxt;
-  tc->snd_nxt = tc->snd_una;
+  u32 bi, n_bytes;
 
   TCP_EVT_DBG (TCP_EVT_CC_EVT, tc, 1);
 
@@ -1767,7 +1758,6 @@ tcp_retransmit_first_unacked (tcp_worker_ctx_t * wrk, tcp_connection_t * tc)
 
   bi = vlib_get_buffer_index (vm, b);
   tcp_enqueue_to_output (wrk, b, bi, tc->c_is_ip4);
-  tc->snd_nxt = old_snd_nxt;
 
   return 0;
 }
@@ -1780,8 +1770,7 @@ tcp_fast_retransmit_unsent (tcp_worker_ctx_t * wrk, tcp_connection_t * tc,
   vlib_main_t *vm = wrk->vm;
   vlib_buffer_t *b = 0;
 
-  tc->snd_nxt = tc->snd_una_max;
-  offset = tc->snd_una_max - tc->snd_una;
+  offset = tc->snd_nxt - tc->snd_una;
   while (n_segs < burst_size)
     {
       n_written = tcp_prepare_segment (wrk, tc, offset, tc->snd_mss, &b);
@@ -1792,6 +1781,9 @@ tcp_fast_retransmit_unsent (tcp_worker_ctx_t * wrk, tcp_connection_t * tc,
       tcp_enqueue_to_output (wrk, b, bi, tc->c_is_ip4);
       offset += n_written;
       n_segs += 1;
+
+      tc->snd_nxt += n_written;
+      tc->snd_una_max = seq_max (tc->snd_nxt, tc->snd_una_max);
     }
 
 done:
@@ -1814,9 +1806,8 @@ tcp_fast_retransmit_sack (tcp_worker_ctx_t * wrk, tcp_connection_t * tc,
   vlib_main_t *vm = wrk->vm;
   vlib_buffer_t *b = 0;
   sack_scoreboard_t *sb;
-  u32 bi, old_snd_nxt;
+  u32 bi, max_deq;
   int snd_space;
-  u32 max_deq;
   u8 snd_limited = 0, can_rescue = 0;
 
   ASSERT (tcp_in_fastrecovery (tc));
@@ -1829,12 +1820,11 @@ tcp_fast_retransmit_sack (tcp_worker_ctx_t * wrk, tcp_connection_t * tc,
     }
 
   TCP_EVT_DBG (TCP_EVT_CC_EVT, tc, 0);
-  old_snd_nxt = tc->snd_nxt;
   sb = &tc->sack_sb;
   hole = scoreboard_get_hole (sb, sb->cur_rxt_hole);
 
   max_deq = transport_max_tx_dequeue (&tc->connection);
-  max_deq -= tc->snd_una_max - tc->snd_una;
+  max_deq -= tc->snd_nxt - tc->snd_una;
 
   while (snd_space > 0 && n_segs < burst_size)
     {
@@ -1867,7 +1857,6 @@ tcp_fast_retransmit_sack (tcp_worker_ctx_t * wrk, tcp_connection_t * tc,
 	  max_bytes = clib_min (max_bytes, snd_space);
 	  offset = tc->snd_congestion - tc->snd_una - max_bytes;
 	  sb->rescue_rxt = tc->snd_congestion;
-	  tc->snd_nxt = tc->snd_una + offset;
 	  n_written = tcp_prepare_retransmit_segment (wrk, tc, offset,
 						      max_bytes, &b);
 	  if (!n_written)
@@ -1885,7 +1874,6 @@ tcp_fast_retransmit_sack (tcp_worker_ctx_t * wrk, tcp_connection_t * tc,
 	break;
 
       offset = sb->high_rxt - tc->snd_una;
-      tc->snd_nxt = sb->high_rxt;
       n_written = tcp_prepare_retransmit_segment (wrk, tc, offset, max_bytes,
 						  &b);
       ASSERT (n_written <= snd_space);
@@ -1906,8 +1894,6 @@ tcp_fast_retransmit_sack (tcp_worker_ctx_t * wrk, tcp_connection_t * tc,
     tcp_program_fastretransmit (wrk, tc);
 
 done:
-  /* If window allows, send 1 SMSS of new data */
-  tc->snd_nxt = old_snd_nxt;
   return n_segs;
 }
 
@@ -1918,14 +1904,13 @@ int
 tcp_fast_retransmit_no_sack (tcp_worker_ctx_t * wrk, tcp_connection_t * tc,
 			     u32 burst_size)
 {
-  u32 n_written = 0, offset = 0, bi, old_snd_nxt, max_deq, n_segs_now;
+  u32 n_written = 0, offset = 0, bi, max_deq, n_segs_now;
   vlib_main_t *vm = wrk->vm;
   int snd_space, n_segs = 0;
   vlib_buffer_t *b;
 
   ASSERT (tcp_in_fastrecovery (tc));
   TCP_EVT_DBG (TCP_EVT_CC_EVT, tc, 0);
-  old_snd_nxt = tc->snd_nxt;
 
   if (!tcp_fastrecovery_first (tc))
     goto send_unsent;
@@ -1933,7 +1918,6 @@ tcp_fast_retransmit_no_sack (tcp_worker_ctx_t * wrk, tcp_connection_t * tc,
   /* RFC 6582: [If a partial ack], retransmit the first unacknowledged
    * segment. */
   snd_space = tc->sack_sb.last_bytes_delivered;
-  tc->snd_nxt = tc->snd_una;
   while (snd_space > 0 && n_segs < burst_size)
     {
       n_written = tcp_prepare_retransmit_segment (wrk, tc, offset,
@@ -1961,7 +1945,7 @@ send_unsent:
     goto done;
 
   max_deq = transport_max_tx_dequeue (&tc->connection);
-  max_deq -= tc->snd_una_max - tc->snd_una;
+  max_deq -= tc->snd_nxt - tc->snd_una;
   if (max_deq)
     {
       snd_space = clib_min (max_deq, snd_space);
@@ -1972,9 +1956,6 @@ send_unsent:
       n_segs += n_segs_now;
     }
 
-  /* Restore snd_nxt */
-  tc->snd_nxt = old_snd_nxt;
-
 done:
   tcp_fastrecovery_first_off (tc);
   return n_segs;