tcp: cc improvements and fixes

Change-Id: I6615bb612bcc3f795b5f822ea55209bb30ef35b5
Signed-off-by: Florin Coras <fcoras@cisco.com>

5 files changed, 38 insertions, 26 deletions

diff --git a/src/vnet/session/session.h b/src/vnet/session/session.h
index b24e22923f1..6908a568a0f 100644
--- a/src/vnet/session/session.h
+++ b/src/vnet/session/session.h
@@ -450,20 +450,27 @@ session_get_from_handle_safe (u64 handle)
 }
 
 always_inline u32
-stream_session_max_rx_enqueue (transport_connection_t * tc)
+transport_max_rx_enqueue (transport_connection_t * tc)
 {
   stream_session_t *s = session_get (tc->s_index, tc->thread_index);
   return svm_fifo_max_enqueue (s->server_rx_fifo);
 }
 
 always_inline u32
-stream_session_rx_fifo_size (transport_connection_t * tc)
+transport_rx_fifo_size (transport_connection_t * tc)
 {
   stream_session_t *s = session_get (tc->s_index, tc->thread_index);
   return s->server_rx_fifo->nitems;
 }
 
 always_inline u32
+transport_tx_fifo_size (transport_connection_t * tc)
+{
+  stream_session_t *s = session_get (tc->s_index, tc->thread_index);
+  return s->server_tx_fifo->nitems;
+}
+
+always_inline u32
 session_get_index (stream_session_t * s)
 {
   return (s - session_manager_main.sessions[s->thread_index]);
diff --git a/src/vnet/tcp/tcp_debug.h b/src/vnet/tcp/tcp_debug.h
index 2e67c84cc33..4af4f2e7052 100755
--- a/src/vnet/tcp/tcp_debug.h
+++ b/src/vnet/tcp/tcp_debug.h
@@ -626,7 +626,7 @@ if (_av > 0) 								\
 {									\
   ELOG_TYPE_DECLARE (_e) =						\
   {									\
-    .format = "cc: %s wnd %u snd_cong %u rxt_bytes %u",			\
+    .format = "cc: %s snd_space %u snd_cong %u rxt_bytes %u",		\
     .format_args = "t4i4i4i4",						\
     .n_enum_strings = 6,						\
     .enum_strings = {                                           	\
diff --git a/src/vnet/tcp/tcp_input.c b/src/vnet/tcp/tcp_input.c
index 42db82e1214..ef6cb8b8fa2 100644
--- a/src/vnet/tcp/tcp_input.c
+++ b/src/vnet/tcp/tcp_input.c
@@ -936,6 +936,12 @@ tcp_update_snd_wnd (tcp_connection_t * tc, u32 seq, u32 ack, u32 snd_wnd)
     }
 }
 
+/**
+ * Init loss recovery/fast recovery.
+ *
+ * Triggered by dup acks as opposed to timer timeout. Note that cwnd is
+ * updated in @ref tcp_cc_handle_event after fast retransmit
+ */
 void
 tcp_cc_init_congestion (tcp_connection_t * tc)
 {
@@ -949,7 +955,6 @@ tcp_cc_init_congestion (tcp_connection_t * tc)
 static void
 tcp_cc_recovery_exit (tcp_connection_t * tc)
 {
-  /* Deflate rto */
   tc->rto_boff = 0;
   tcp_update_rto (tc);
   tc->snd_rxt_ts = 0;
@@ -999,6 +1004,7 @@ tcp_cc_recover (tcp_connection_t * tc)
   ASSERT (tcp_in_cong_recovery (tc));
   if (tcp_cc_is_spurious_retransmit (tc))
     {
+      clib_warning ("here");
       tcp_cc_congestion_undo (tc);
       return 1;
     }
@@ -1061,28 +1067,24 @@ tcp_cc_handle_event (tcp_connection_t * tc, u32 is_dack)
    * Duplicate ACK. Check if we should enter fast recovery, or if already in
    * it account for the bytes that left the network.
    */
-  if (is_dack)
+  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);
 
       tc->rcv_dupacks++;
 
+      /* Pure duplicate ack. If some data got acked, it's handled lower */
       if (tc->rcv_dupacks > TCP_DUPACK_THRESHOLD && !tc->bytes_acked)
 	{
 	  ASSERT (tcp_in_fastrecovery (tc));
-	  /* Pure duplicate ack. If some data got acked, it's handled lower */
 	  tc->cc_algo->rcv_cong_ack (tc, TCP_CC_DUPACK);
 	  return;
 	}
       else if (tcp_should_fastrecover (tc))
 	{
-	  /* Things are already bad */
-	  if (tcp_in_cong_recovery (tc))
-	    {
-	      tc->rcv_dupacks = 0;
-	      goto partial_ack_test;
-	    }
+	  ASSERT (!tcp_in_fastrecovery (tc));
 
 	  /* If of of the two conditions lower hold, reset dupacks because
 	   * we're probably after timeout (RFC6582 heuristics).
@@ -1139,12 +1141,12 @@ tcp_cc_handle_event (tcp_connection_t * tc, u32 is_dack)
 	goto partial_ack;
     }
 
-partial_ack_test:
-
   if (!tc->bytes_acked)
     return;
 
 partial_ack:
+  TCP_EVT_DBG (TCP_EVT_CC_PACK, tc);
+
   /*
    * Legitimate ACK. 1) See if we can exit recovery
    */
@@ -1171,17 +1173,18 @@ partial_ack:
   /*
    * Legitimate ACK. 2) If PARTIAL ACK try to retransmit
    */
-  TCP_EVT_DBG (TCP_EVT_CC_PACK, tc);
 
   /* RFC6675: If the incoming ACK is a cumulative acknowledgment,
-   * reset dupacks to 0 */
+   * reset dupacks to 0. Also needed if in congestion recovery */
   tc->rcv_dupacks = 0;
 
-  tcp_retransmit_first_unacked (tc);
-
   /* Post RTO timeout don't try anything fancy */
   if (tcp_in_recovery (tc))
-    return;
+    {
+      tc->cc_algo->rcv_ack (tc);
+      tc->tsecr_last_ack = tc->rcv_opts.tsecr;
+      return;
+    }
 
   /* Remove retransmitted bytes that have been delivered */
   ASSERT (tc->bytes_acked + tc->sack_sb.snd_una_adv
@@ -1262,7 +1265,6 @@ tcp_rcv_ack (tcp_connection_t * tc, vlib_buffer_t * b,
 		   vnet_buffer (b)->tcp.ack_number);
       if (tcp_in_fastrecovery (tc) && tc->rcv_dupacks == TCP_DUPACK_THRESHOLD)
 	{
-	  TCP_EVT_DBG (TCP_EVT_DUPACK_RCVD, tc);
 	  tcp_cc_handle_event (tc, 1);
 	}
       /* Don't drop yet */
@@ -1300,7 +1302,6 @@ tcp_rcv_ack (tcp_connection_t * tc, vlib_buffer_t * b,
       if (!tcp_in_cong_recovery (tc))
 	return 0;
       *error = TCP_ERROR_ACK_DUP;
-      TCP_EVT_DBG (TCP_EVT_DUPACK_RCVD, tc, 1);
       return vnet_buffer (b)->tcp.data_len ? 0 : -1;
     }
 
diff --git a/src/vnet/tcp/tcp_newreno.c b/src/vnet/tcp/tcp_newreno.c
index 7ae7f484b56..0f43d21dfde 100644
--- a/src/vnet/tcp/tcp_newreno.c
+++ b/src/vnet/tcp/tcp_newreno.c
@@ -44,6 +44,8 @@ newreno_rcv_ack (tcp_connection_t * tc)
 	  tc->cwnd += inc * tc->snd_mss;
 	  tc->cwnd_acc_bytes -= inc * tc->cwnd;
 	}
+      tc->cwnd = clib_min (tc->cwnd,
+			   transport_tx_fifo_size (&tc->connection));
     }
 }
 
diff --git a/src/vnet/tcp/tcp_output.c b/src/vnet/tcp/tcp_output.c
index 5932ed406d9..27450654f71 100644
--- a/src/vnet/tcp/tcp_output.c
+++ b/src/vnet/tcp/tcp_output.c
@@ -163,8 +163,8 @@ tcp_update_rcv_wnd (tcp_connection_t * tc)
   /*
    * Figure out how much space we have available
    */
-  available_space = stream_session_max_rx_enqueue (&tc->connection);
-  max_fifo = stream_session_rx_fifo_size (&tc->connection);
+  available_space = transport_max_rx_enqueue (&tc->connection);
+  max_fifo = transport_rx_fifo_size (&tc->connection);
 
   ASSERT (tc->rcv_opts.mss < max_fifo);
   if (available_space < tc->rcv_opts.mss && available_space < max_fifo >> 3)
@@ -1347,10 +1347,12 @@ tcp_rtx_timeout_cc (tcp_connection_t * tc)
     tcp_cc_fastrecovery_exit (tc);
 
   /* Start again from the beginning */
-  tc->ssthresh = clib_max (tcp_flight_size (tc) / 2, 2 * tc->snd_mss);
+  tc->cc_algo->congestion (tc);
   tc->cwnd = tcp_loss_wnd (tc);
   tc->snd_congestion = tc->snd_una_max;
   tc->rtt_ts = 0;
+  tc->cwnd_acc_bytes = 0;
+
   tcp_recovery_on (tc);
 }
 
@@ -1393,7 +1395,7 @@ tcp_timer_retransmit_handler_i (u32 index, u8 is_syn)
 	}
 
       /* Shouldn't be here */
-      if (tc->snd_una == tc->snd_una_max)
+      if (seq_geq (tc->snd_una, tc->snd_congestion))
 	{
 	  tcp_recovery_off (tc);
 	  return;
@@ -1414,7 +1416,7 @@ tcp_timer_retransmit_handler_i (u32 index, u8 is_syn)
       if (tc->rto_boff == 1)
 	tcp_rtx_timeout_cc (tc);
 
-      tc->snd_nxt = tc->snd_una;
+      tc->snd_una_max = tc->snd_nxt = tc->snd_una;
       tc->rto = clib_min (tc->rto << 1, TCP_RTO_MAX);
 
       TCP_EVT_DBG (TCP_EVT_CC_EVT, tc, 1);