tcp: when rearming RTO, if RTO time is in past then fire RTO ASAP
authorNeal Cardwell <ncardwell@google.com>
Wed, 16 Aug 2017 21:53:36 +0000 (17:53 -0400)
committerDavid S. Miller <davem@davemloft.net>
Fri, 18 Aug 2017 23:07:43 +0000 (16:07 -0700)
In some situations tcp_send_loss_probe() can realize that it's unable
to send a loss probe (TLP), and falls back to calling tcp_rearm_rto()
to schedule an RTO timer. In such cases, sometimes tcp_rearm_rto()
realizes that the RTO was eligible to fire immediately or at some
point in the past (delta_us <= 0). Previously in such cases
tcp_rearm_rto() was scheduling such "overdue" RTOs to happen at now +
icsk_rto, which caused needless delays of hundreds of milliseconds
(and non-linear behavior that made reproducible testing
difficult). This commit changes the logic to schedule "overdue" RTOs
ASAP, rather than at now + icsk_rto.

Fixes: 6ba8a3b19e76 ("tcp: Tail loss probe (TLP)")
Suggested-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
net/ipv4/tcp_input.c

index 53de1424c13cda5d1fec826b97cacf4f95adc99a..bab7f0493098c6521f445923d721d296f326f7e1 100644 (file)
@@ -3009,8 +3009,7 @@ void tcp_rearm_rto(struct sock *sk)
                        /* delta_us may not be positive if the socket is locked
                         * when the retrans timer fires and is rescheduled.
                         */
-                       if (delta_us > 0)
-                               rto = usecs_to_jiffies(delta_us);
+                       rto = usecs_to_jiffies(max_t(int, delta_us, 1));
                }
                inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, rto,
                                          TCP_RTO_MAX);