Neal reported that using neper tcp_stream with TCP_TX_DELAY set to 50ms would often lead to flows stuck in a small cwnd mode, regardless of the congestion control. While tcp_stream sets TCP_TX_DELAY too late after the connect(), it highlighted two kernel bugs. The following heuristic in tcp_tso_should_defer() seems wrong for large RTT: delta = tp->tcp_clock_cache - head->tstamp; /* If next ACK is likely to come too late (half srtt), do not defer */ if ((s64)(delta - (u64)NSEC_PER_USEC * (tp->srtt_us >> 4)) < 0) goto send_now; If next ACK is expected to come in more than 1 ms, we should not defer because we prefer a smooth ACK clocking. While blamed commit was a step in the good direction, it was not generic enough. Another patch fixing TCP_TX_DELAY for established flows will be proposed when net-next reopens. Fixes: 50c8339e9299 ("tcp: tso: restore IW10 after TSO autosizing") Reported-by: Neal Cardwell Signed-off-by: Eric Dumazet --- net/ipv4/tcp_output.c | 21 ++++++++++++++++----- 1 file changed, 16 insertions(+), 5 deletions(-) diff --git a/net/ipv4/tcp_output.c b/net/ipv4/tcp_output.c index bb3576ac0ad7d7330ef272e1d9dc1f19bb8f86bb..bbeed379a3c5342c7de0d2416f97ad944e3e35b0 100644 --- a/net/ipv4/tcp_output.c +++ b/net/ipv4/tcp_output.c @@ -2369,7 +2369,8 @@ static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb, u32 max_segs) { const struct inet_connection_sock *icsk = inet_csk(sk); - u32 send_win, cong_win, limit, in_flight; + u32 send_win, cong_win, limit, in_flight, threshold; + u64 srtt_in_ns, expected_ack, how_far_is_the_ack; struct tcp_sock *tp = tcp_sk(sk); struct sk_buff *head; int win_divisor; @@ -2431,10 +2432,20 @@ static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb, head = tcp_rtx_queue_head(sk); if (!head) goto send_now; - delta = tp->tcp_clock_cache - head->tstamp; - /* If next ACK is likely to come too late (half srtt), do not defer */ - if ((s64)(delta - (u64)NSEC_PER_USEC * (tp->srtt_us >> 4)) < 0) - goto send_now; + + srtt_in_ns = (u64)(NSEC_PER_USEC >> 3) * tp->srtt_us; + /* When is the ACK expected ? */ + expected_ack = head->tstamp + srtt_in_ns; + /* How far from now is the ACK expected ? */ + how_far_is_the_ack = expected_ack - tp->tcp_clock_cache; + + /* If next ACK is likely to come too late, + * ie in more than min(1ms, half srtt), do not defer. + */ + threshold = min(srtt_in_ns >> 1, NSEC_PER_MSEC); + + if ((s64)(how_far_is_the_ack - threshold) > 0) + goto send_now; /* Ok, it looks like it is advisable to defer. * Three cases are tracked : -- 2.51.0.740.g6adb054d12-goog