[PATCH net] tcp: do not change rcv_ssthresh in tcp_measure_rcv_mss()

[PATCH net] tcp: do not change rcv_ssthresh in tcp_measure_rcv_mss()

[Nathan Gao]

Commit f5da7c45188e ("tcp: adjust rcvq_space after updating scaling
ratio") replaced the direct window_clamp update in tcp_measure_rcv_mss()
with a call to tcp_set_window_clamp(), a helper that implements the
TCP_WINDOW_CLAMP setsockopt. As a side effect, the helper also shrinks
rcv_ssthresh via __tcp_adjust_rcv_ssthresh().

As a result, each scaling_ratio decrease detected by
tcp_measure_rcv_mss() also cuts rcv_ssthresh. Elsewhere in TCP,
rcv_ssthresh is usually cut under memory pressure and grows via
tcp_grow_window().

Flows whose segment sizes vary keep scaling_ratio oscillating, which
leads to an unstable rcv_ssthresh: a dip of rcv_ssthresh only recovers
via tcp_grow_window(), keeping the advertised window at a relatively
low level even after the ratio itself has recovered, and can even stall
the sender.

Observed on a customer's proxy gateway after upgrading from kernel 6.1
to 6.12: in the worst case, rcv_ssthresh was cut in half by a
scaling_ratio dip. P99 latency jumped from <10ms on 6.1 to ~100ms on
6.12, and almost returned to the 6.1 level with this patch applied.

Restore the plain WRITE_ONCE() update of window_clamp, as introduced
in commit a2cbb1603943 ("tcp: Update window clamping condition"), and
keep the rcvq_space.space adjustment. Now rcv_ssthresh is decoupled from
scaling_ratio changes in tcp_measure_rcv_mss().

Fixes: f5da7c45188e ("tcp: adjust rcvq_space after updating scaling ratio")
Signed-off-by: Nathan Gao <zcgao@amazon.com>
---
 net/ipv4/tcp_input.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/net/ipv4/tcp_input.c b/net/ipv4/tcp_input.c
index daff93d513428..5b6378b94701e 100644
--- a/net/ipv4/tcp_input.c
+++ b/net/ipv4/tcp_input.c
@@ -252,7 +252,7 @@ static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb)
 				struct tcp_sock *tp = tcp_sk(sk);
 
 				val = tcp_win_from_space(sk, sk->sk_rcvbuf);
-				tcp_set_window_clamp(sk, val);
+				WRITE_ONCE(tp->window_clamp, val);
 
 				if (tp->window_clamp < tp->rcvq_space.space)
 					tp->rcvq_space.space = tp->window_clamp;

base-commit: 78f75d632f74b8de0f081a128588f7c37d0d1164
-- 
2.47.3

Re: [PATCH net] tcp: do not change rcv_ssthresh in tcp_measure_rcv_mss()

[Jakub Kicinski]

On Fri, 24 Jul 2026 20:08:06 -0700 Nathan Gao wrote:
> Commit f5da7c45188e ("tcp: adjust rcvq_space after updating scaling
> ratio") replaced the direct window_clamp update in tcp_measure_rcv_mss()
> with a call to tcp_set_window_clamp(), a helper that implements the
> TCP_WINDOW_CLAMP setsockopt. As a side effect, the helper also shrinks
> rcv_ssthresh via __tcp_adjust_rcv_ssthresh().
> 
> As a result, each scaling_ratio decrease detected by
> tcp_measure_rcv_mss() also cuts rcv_ssthresh. Elsewhere in TCP,
> rcv_ssthresh is usually cut under memory pressure and grows via
> tcp_grow_window().
> 
> Flows whose segment sizes vary keep scaling_ratio oscillating, which
> leads to an unstable rcv_ssthresh: a dip of rcv_ssthresh only recovers
> via tcp_grow_window(), keeping the advertised window at a relatively
> low level even after the ratio itself has recovered, and can even stall
> the sender.
> 
> Observed on a customer's proxy gateway after upgrading from kernel 6.1
> to 6.12: in the worst case, rcv_ssthresh was cut in half by a
> scaling_ratio dip. P99 latency jumped from <10ms on 6.1 to ~100ms on
> 6.12, and almost returned to the 6.1 level with this patch applied.
> 
> Restore the plain WRITE_ONCE() update of window_clamp, as introduced
> in commit a2cbb1603943 ("tcp: Update window clamping condition"), and
> keep the rcvq_space.space adjustment. Now rcv_ssthresh is decoupled from
> scaling_ratio changes in tcp_measure_rcv_mss().
> 
> Fixes: f5da7c45188e ("tcp: adjust rcvq_space after updating scaling ratio")
> Signed-off-by: Nathan Gao <zcgao@amazon.com>

Not sure, I mean regression is a regression, but also the previous
behavior seems to have just been lucky rather than correct in principle?

Looks like Eric and Neal are AFK, Kuniyuki, Paolo, any opinion on this
patch?

> diff --git a/net/ipv4/tcp_input.c b/net/ipv4/tcp_input.c
> index daff93d513428..5b6378b94701e 100644
> --- a/net/ipv4/tcp_input.c
> +++ b/net/ipv4/tcp_input.c
> @@ -252,7 +252,7 @@ static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb)
>  				struct tcp_sock *tp = tcp_sk(sk);
>  
>  				val = tcp_win_from_space(sk, sk->sk_rcvbuf);
> -				tcp_set_window_clamp(sk, val);
> +				WRITE_ONCE(tp->window_clamp, val);
>  
>  				if (tp->window_clamp < tp->rcvq_space.space)
>  					tp->rcvq_space.space = tp->window_clamp;
> 
> base-commit: 78f75d632f74b8de0f081a128588f7c37d0d1164