The cited commit exposed an old retransmits_timed_out() bug
which assumed it could call tcp_model_timeout() with
TCP_RTO_MIN as rto_base for all states.
But flows in SYN_SENT or SYN_RECV state uses a different
RTO base (1 sec instead of 200 ms, unless BPF choses
another value)
This caused a reduction of SYN retransmits from 6 to 4 with
the default /proc/sys/net/ipv4/tcp_syn_retries value.
Fixes: a41e8a88b06e ("tcp: better handle TCP_USER_TIMEOUT in SYN_SENT state")
Signed-off-by: Eric Dumazet <eduma...@google.com>
Cc: Yuchung Cheng <ych...@google.com>
Cc: Marek Majkowski <ma...@cloudflare.com>
---
net/ipv4/tcp_timer.c | 9 +++++++--
1 file changed, 7 insertions(+), 2 deletions(-)
diff --git a/net/ipv4/tcp_timer.c b/net/ipv4/tcp_timer.c
index
40de2d2364a1eca14c259d77ebed361d17829eb9..05be564414e9b4aad64321e381fc0afa10980190
100644
--- a/net/ipv4/tcp_timer.c
+++ b/net/ipv4/tcp_timer.c
@@ -198,8 +198,13 @@ static bool retransmits_timed_out(struct sock *sk,
return false;
start_ts = tcp_sk(sk)->retrans_stamp;
- if (likely(timeout == 0))
- timeout = tcp_model_timeout(sk, boundary, TCP_RTO_MIN);
+ if (likely(timeout == 0)) {
+ unsigned int rto_base = TCP_RTO_MIN;
+
+ if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
+ rto_base = tcp_timeout_init(sk);
+ timeout = tcp_model_timeout(sk, boundary, rto_base);
+ }
return (s32)(tcp_time_stamp(tcp_sk(sk)) - start_ts - timeout) >= 0;
}
--
2.23.0.444.g18eeb5a265-goog
