tcp: tsq: fix nonagle handling
authorJohn Ogness <john.ogness@linutronix.de>
Mon, 10 Feb 2014 02:40:11 +0000 (18:40 -0800)
committerDavid S. Miller <davem@davemloft.net>
Mon, 10 Feb 2014 23:23:39 +0000 (15:23 -0800)
Commit 46d3ceabd8d9 ("tcp: TCP Small Queues") introduced a possible
regression for applications using TCP_NODELAY.

If TCP session is throttled because of tsq, we should consult
tp->nonagle when TX completion is done and allow us to send additional
segment, especially if this segment is not a full MSS.
Otherwise this segment is sent after an RTO.

[edumazet] : Cooked the changelog, added another fix about testing
sk_wmem_alloc twice because TX completion can happen right before
setting TSQ_THROTTLED bit.

This problem is particularly visible with recent auto corking,
but might also be triggered with low tcp_limit_output_bytes
values or NIC drivers delaying TX completion by hundred of usec,
and very low rtt.

Thomas Glanzmann for example reported an iscsi regression, caused
by tcp auto corking making this bug quite visible.

Fixes: 46d3ceabd8d9 ("tcp: TCP Small Queues")
Signed-off-by: John Ogness <john.ogness@linutronix.de>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reported-by: Thomas Glanzmann <thomas@glanzmann.de>
Signed-off-by: David S. Miller <davem@davemloft.net>
net/ipv4/tcp_output.c

index 10435b3b9d0f9b0ee2f3b9186329deecd6ea80e4..3be16727f058b191d305366f9cf5851f03c7c417 100644 (file)
@@ -698,7 +698,8 @@ static void tcp_tsq_handler(struct sock *sk)
        if ((1 << sk->sk_state) &
            (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING |
             TCPF_CLOSE_WAIT  | TCPF_LAST_ACK))
-               tcp_write_xmit(sk, tcp_current_mss(sk), 0, 0, GFP_ATOMIC);
+               tcp_write_xmit(sk, tcp_current_mss(sk), tcp_sk(sk)->nonagle,
+                              0, GFP_ATOMIC);
 }
 /*
  * One tasklet per cpu tries to send more skbs.
@@ -1904,7 +1905,15 @@ static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
 
                if (atomic_read(&sk->sk_wmem_alloc) > limit) {
                        set_bit(TSQ_THROTTLED, &tp->tsq_flags);
-                       break;
+                       /* It is possible TX completion already happened
+                        * before we set TSQ_THROTTLED, so we must
+                        * test again the condition.
+                        * We abuse smp_mb__after_clear_bit() because
+                        * there is no smp_mb__after_set_bit() yet
+                        */
+                       smp_mb__after_clear_bit();
+                       if (atomic_read(&sk->sk_wmem_alloc) > limit)
+                               break;
                }
 
                limit = mss_now;