If there is no cached TCP metrics entry for a connection, initialize
tp->snd_ssthresh from the corresponding dst entry. Also move the check
against tp->snd_cwnd_clamp to the common path to ensure that the ssthresh
value is never greater than maximum cwnd, regardless where it came from.

Fixes: 51c5d0c4b169 ("tcp: Maintain dynamic metrics in local cache.")
Signed-off-by: Petr Tesarik <ptesarik@suse.com>
---
 net/ipv4/tcp_metrics.c | 11 +++++++----
 1 file changed, 7 insertions(+), 4 deletions(-)

diff --git a/net/ipv4/tcp_metrics.c b/net/ipv4/tcp_metrics.c
index dd8f3457bd72e..b08920abec0e6 100644
--- a/net/ipv4/tcp_metrics.c
+++ b/net/ipv4/tcp_metrics.c
@@ -479,6 +479,9 @@ void tcp_init_metrics(struct sock *sk)
 
 	if (dst_metric_locked(dst, RTAX_CWND))
 		tp->snd_cwnd_clamp = dst_metric(dst, RTAX_CWND);
+	val = dst_metric(dst, RTAX_SSTHRESH);
+	if (val)
+		tp->snd_ssthresh = val;
 
 	rcu_read_lock();
 	tm = tcp_get_metrics(sk, dst, false);
@@ -489,11 +492,8 @@ void tcp_init_metrics(struct sock *sk)
 
 	val = READ_ONCE(net->ipv4.sysctl_tcp_no_ssthresh_metrics_save) ?
 	      0 : tcp_metric_get(tm, TCP_METRIC_SSTHRESH);
-	if (val) {
+	if (val)
 		tp->snd_ssthresh = val;
-		if (tp->snd_ssthresh > tp->snd_cwnd_clamp)
-			tp->snd_ssthresh = tp->snd_cwnd_clamp;
-	}
 	val = tcp_metric_get(tm, TCP_METRIC_REORDERING);
 	if (val && tp->reordering != val)
 		tp->reordering = val;
@@ -537,6 +537,9 @@ void tcp_init_metrics(struct sock *sk)
 
 		inet_csk(sk)->icsk_rto = TCP_TIMEOUT_FALLBACK;
 	}
+
+	if (tp->snd_ssthresh > tp->snd_cwnd_clamp)
+		tp->snd_ssthresh = tp->snd_cwnd_clamp;
 }
 
 bool tcp_peer_is_proven(struct request_sock *req, struct dst_entry *dst)
-- 
2.49.0

On 6/13/25 12:20 PM, Petr Tesarik wrote:
> @@ -537,6 +537,9 @@ void tcp_init_metrics(struct sock *sk)
>  
>  		inet_csk(sk)->icsk_rto = TCP_TIMEOUT_FALLBACK;
>  	}
> +
> +	if (tp->snd_ssthresh > tp->snd_cwnd_clamp)
> +		tp->snd_ssthresh = tp->snd_cwnd_clamp;

I don't think we can do this unconditionally, as other parts of the TCP
stack check explicitly for TCP_INFINITE_SSTHRESH.

/P

On Tue, 17 Jun 2025 12:48:30 +0200
Paolo Abeni <pabeni@redhat.com> wrote:

> On 6/13/25 12:20 PM, Petr Tesarik wrote:
> > @@ -537,6 +537,9 @@ void tcp_init_metrics(struct sock *sk)
> >  
> >  		inet_csk(sk)->icsk_rto = TCP_TIMEOUT_FALLBACK;
> >  	}
> > +
> > +	if (tp->snd_ssthresh > tp->snd_cwnd_clamp)
> > +		tp->snd_ssthresh = tp->snd_cwnd_clamp;  
> 
> I don't think we can do this unconditionally, as other parts of the TCP
> stack check explicitly for TCP_INFINITE_SSTHRESH.


Good catch! I noticed that the condition can never be true unless the
congestion window is explicitly clamped, but you're right that it is a
valid combination to lock the maximum cwnd but keep the initial TCP Slow
Start.

I'll fix that in v2.

Thank you,
Petr T