tcp: fix for zero packets_in_flight was too broad
authorIlpo Järvinen <ilpo.jarvinen@helsinki.fi>
Mon, 4 Feb 2013 02:14:25 +0000 (02:14 +0000)
committerDavid S. Miller <davem@davemloft.net>
Wed, 6 Feb 2013 20:53:03 +0000 (15:53 -0500)
There are transients during normal FRTO procedure during which
the packets_in_flight can go to zero between write_queue state
updates and firing the resulting segments out. As FRTO processing
occurs during that window the check must be more precise to
not match "spuriously" :-). More specificly, e.g., when
packets_in_flight is zero but FLAG_DATA_ACKED is true the problematic
branch that set cwnd into zero would not be taken and new segments
might be sent out later.

Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Tested-by: Eric Dumazet <edumazet@google.com>
Acked-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
net/ipv4/tcp_input.c

index 680c422..ad70a96 100644 (file)
@@ -3484,8 +3484,7 @@ static bool tcp_process_frto(struct sock *sk, int flag)
            ((tp->frto_counter >= 2) && (flag & FLAG_RETRANS_DATA_ACKED)))
                tp->undo_marker = 0;
 
-       if (!before(tp->snd_una, tp->frto_highmark) ||
-           !tcp_packets_in_flight(tp)) {
+       if (!before(tp->snd_una, tp->frto_highmark)) {
                tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 2 : 3), flag);
                return true;
        }
@@ -3505,6 +3504,11 @@ static bool tcp_process_frto(struct sock *sk, int flag)
                }
        } else {
                if (!(flag & FLAG_DATA_ACKED) && (tp->frto_counter == 1)) {
+                       if (!tcp_packets_in_flight(tp)) {
+                               tcp_enter_frto_loss(sk, 2, flag);
+                               return true;
+                       }
+
                        /* Prevent sending of new data. */
                        tp->snd_cwnd = min(tp->snd_cwnd,
                                           tcp_packets_in_flight(tp));