tcp: refine tcp_pacing_delay() for very low pacing rates

With the addition of horizon feature to sch_fq, we noticed some
suboptimal behavior of extremely low pacing rate TCP flows, especially
when TCP is not aware of a drop happening in lower stacks.

Back in commit 3f80e08f40cd ("tcp: add tcp_reset_xmit_timer() helper"),
tcp_pacing_delay() was added to estimate an extra delay to add to standard
rto timers.

This patch removes the skb argument from this helper and
tcp_reset_xmit_timer() because it makes more sense to simply
consider the time at which next packet is allowed to be sent,
instead of the time of whatever packet has been sent.

This avoids arming RTO timer too soon and removes
spurious horizon drops.

Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

1 files changed, 3 insertions, 5 deletions

diff --git a/net/ipv4/tcp_output.c b/net/ipv4/tcp_output.c
index c414aeb1efa9..32c9db902f18 100644
--- a/net/ipv4/tcp_output.c
+++ b/net/ipv4/tcp_output.c
@@ -2593,8 +2593,7 @@ bool tcp_schedule_loss_probe(struct sock *sk, bool advancing_rto)
 	if (rto_delta_us > 0)
 		timeout = min_t(u32, timeout, usecs_to_jiffies(rto_delta_us));
 
-	tcp_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout,
-			     TCP_RTO_MAX, NULL);
+	tcp_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout, TCP_RTO_MAX);
 	return true;
 }
 
@@ -3174,8 +3173,7 @@ void tcp_xmit_retransmit_queue(struct sock *sk)
 		    icsk->icsk_pending != ICSK_TIME_REO_TIMEOUT)
 			tcp_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
 					     inet_csk(sk)->icsk_rto,
-					     TCP_RTO_MAX,
-					     skb);
+					     TCP_RTO_MAX);
 	}
 }
 
@@ -3907,7 +3905,7 @@ void tcp_send_probe0(struct sock *sk)
 		 */
 		timeout = TCP_RESOURCE_PROBE_INTERVAL;
 	}
-	tcp_reset_xmit_timer(sk, ICSK_TIME_PROBE0, timeout, TCP_RTO_MAX, NULL);
+	tcp_reset_xmit_timer(sk, ICSK_TIME_PROBE0, timeout, TCP_RTO_MAX);
 }
 
 int tcp_rtx_synack(const struct sock *sk, struct request_sock *req)