tcp: don't use F-RTO on non-recurring timeouts
authorYuchung Cheng <ycheng@google.com>
Mon, 13 Jul 2015 19:10:20 +0000 (12:10 -0700)
committerDavid S. Miller <davem@davemloft.net>
Thu, 16 Jul 2015 00:17:21 +0000 (17:17 -0700)
Currently F-RTO may repeatedly send new data packets on non-recurring
timeouts in CA_Loss mode. This is a bug because F-RTO (RFC5682)
should only be used on either new recovery or recurring timeouts.

This exacerbates the recovery progress during frequent timeout &
repair, because we prioritize sending new data packets instead of
repairing the holes when the bandwidth is already scarce.

Fix it by correcting the test of a new recovery episode.

Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
net/ipv4/tcp_input.c

index 684f095d196e20333adb235fc96a8fb8f0dd691c..728f5b3d3c64197bb526240a078744d5a950c8ea 100644 (file)
@@ -1917,14 +1917,13 @@ void tcp_enter_loss(struct sock *sk)
        const struct inet_connection_sock *icsk = inet_csk(sk);
        struct tcp_sock *tp = tcp_sk(sk);
        struct sk_buff *skb;
-       bool new_recovery = false;
+       bool new_recovery = icsk->icsk_ca_state < TCP_CA_Recovery;
        bool is_reneg;                  /* is receiver reneging on SACKs? */
 
        /* Reduce ssthresh if it has not yet been made inside this window. */
        if (icsk->icsk_ca_state <= TCP_CA_Disorder ||
            !after(tp->high_seq, tp->snd_una) ||
            (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
-               new_recovery = true;
                tp->prior_ssthresh = tcp_current_ssthresh(sk);
                tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
                tcp_ca_event(sk, CA_EVENT_LOSS);