tcp: disable Fast Open on timeouts after handshake
authorYuchung Cheng <ycheng@google.com>
Thu, 19 Nov 2015 02:17:30 +0000 (18:17 -0800)
committerDavid S. Miller <davem@davemloft.net>
Fri, 20 Nov 2015 15:51:12 +0000 (10:51 -0500)
Some middle-boxes black-hole the data after the Fast Open handshake
(https://www.ietf.org/proceedings/94/slides/slides-94-tcpm-13.pdf).
The exact reason is unknown. The work-around is to disable Fast Open
temporarily after multiple recurring timeouts with few or no data
delivered in the established state.

Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reported-by: Christoph Paasch <cpaasch@apple.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
net/ipv4/tcp_timer.c

index c9c716a483e457e8b4fe063776ae3d7a855889cc..448603a819668bb5fb0abb02f86207987fd8ce81 100644 (file)
@@ -176,6 +176,18 @@ static int tcp_write_timeout(struct sock *sk)
                syn_set = true;
        } else {
                if (retransmits_timed_out(sk, sysctl_tcp_retries1, 0, 0)) {
+                       /* Some middle-boxes may black-hole Fast Open _after_
+                        * the handshake. Therefore we conservatively disable
+                        * Fast Open on this path on recurring timeouts with
+                        * few or zero bytes acked after Fast Open.
+                        */
+                       if (tp->syn_data_acked &&
+                           tp->bytes_acked <= tp->rx_opt.mss_clamp) {
+                               tcp_fastopen_cache_set(sk, 0, NULL, true, 0);
+                               if (icsk->icsk_retransmits == sysctl_tcp_retries1)
+                                       NET_INC_STATS_BH(sock_net(sk),
+                                                        LINUX_MIB_TCPFASTOPENACTIVEFAIL);
+                       }
                        /* Black hole detection */
                        tcp_mtu_probing(icsk, sk);