tcp: send enough dupacks to cover all sack holes

Make sure we send enough dupacks to cover all the holes created in the
last frame received. Also make sure we send all the blocks, not just the
first.

Change-Id: I9597a34ac14473d1cc3ad07d65bc37043e3d0582
Signed-off-by: Florin Coras <fcoras@cisco.com>

1 files changed, 33 insertions, 11 deletions

diff --git a/src/vnet/tcp/tcp_output.c b/src/vnet/tcp/tcp_output.c
index 725ffec0852..4de479c344b 100644
--- a/src/vnet/tcp/tcp_output.c
+++ b/src/vnet/tcp/tcp_output.c
@@ -253,14 +253,12 @@ tcp_options_write (u8 * data, tcp_options_t * opts)
   if (tcp_opts_sack (opts))
     {
       int i;
-      u32 n_sack_blocks = clib_min (vec_len (opts->sacks),
-				    TCP_OPTS_MAX_SACK_BLOCKS);
 
-      if (n_sack_blocks != 0)
+      if (opts->n_sack_blocks != 0)
 	{
 	  *data++ = TCP_OPTION_SACK_BLOCK;
-	  *data++ = 2 + n_sack_blocks * TCP_OPTION_LEN_SACK_BLOCK;
-	  for (i = 0; i < n_sack_blocks; i++)
+	  *data++ = 2 + opts->n_sack_blocks * TCP_OPTION_LEN_SACK_BLOCK;
+	  for (i = 0; i < opts->n_sack_blocks; i++)
 	    {
 	      buf = clib_host_to_net_u32 (opts->sacks[i].start);
 	      clib_memcpy_fast (data, &buf, seq_len);
@@ -269,7 +267,7 @@ tcp_options_write (u8 * data, tcp_options_t * opts)
 	      clib_memcpy_fast (data, &buf, seq_len);
 	      data += seq_len;
 	    }
-	  opts_len += 2 + n_sack_blocks * TCP_OPTION_LEN_SACK_BLOCK;
+	  opts_len += 2 + opts->n_sack_blocks * TCP_OPTION_LEN_SACK_BLOCK;
 	}
     }
 
@@ -372,9 +370,13 @@ tcp_make_established_options (tcp_connection_t * tc, tcp_options_t * opts)
       if (vec_len (tc->snd_sacks))
 	{
 	  opts->flags |= TCP_OPTS_FLAG_SACK;
-	  opts->sacks = tc->snd_sacks;
-	  opts->n_sack_blocks = clib_min (vec_len (tc->snd_sacks),
+	  if (tc->snd_sack_pos >= vec_len (tc->snd_sacks))
+	    tc->snd_sack_pos = 0;
+	  opts->sacks = &tc->snd_sacks[tc->snd_sack_pos];
+	  opts->n_sack_blocks = vec_len (tc->snd_sacks) - tc->snd_sack_pos;
+	  opts->n_sack_blocks = clib_min (opts->n_sack_blocks,
 					  TCP_OPTS_MAX_SACK_BLOCKS);
+	  tc->snd_sack_pos += opts->n_sack_blocks;
 	  len += 2 + TCP_OPTION_LEN_SACK_BLOCK * opts->n_sack_blocks;
 	}
     }
@@ -1250,14 +1252,34 @@ tcp_send_acks (tcp_worker_ctx_t * wrk)
     {
       tc = tcp_connection_get (pending_acks[i], thread_index);
       tc->flags &= ~TCP_CONN_SNDACK;
-      n_acks = clib_max (1, tc->pending_dupacks);
+      if (!tc->pending_dupacks)
+	{
+	  tcp_send_ack (tc);
+	  continue;
+	}
+
       /* If we're supposed to send dupacks but have no ooo data
        * send only one ack */
-      if (tc->pending_dupacks && !vec_len (tc->snd_sacks))
-	n_acks = 1;
+      if (!vec_len (tc->snd_sacks))
+	{
+	  tcp_send_ack (tc);
+	  continue;
+	}
+
+      /* Start with first sack block */
+      tc->snd_sack_pos = 0;
+
+      /* Generate enough dupacks to cover all sack blocks. Do not generate
+       * more sacks than the number of packets received. But do generate at
+       * least 3, i.e., the number needed to signal congestion, if needed. */
+      n_acks = vec_len (tc->snd_sacks) / TCP_OPTS_MAX_SACK_BLOCKS;
+      n_acks = clib_min (n_acks, tc->pending_dupacks);
+      n_acks = clib_max (n_acks, clib_min (tc->pending_dupacks, 3));
       for (j = 0; j < n_acks; j++)
 	tcp_send_ack (tc);
+
       tc->pending_dupacks = 0;
+      tc->snd_sack_pos = 0;
     }
   _vec_len (wrk->pending_acks) = 0;
 }