Re: [PATCH net 1/2] tcp: avoid false bind exhaustion with odd-sized port ranges

[Eric Dumazet]

Le jeu. 5 nov. 2026 à 12:04, Florin Badita-Nistor
<baditaflorin@xxxxxxxxx> a écrit :
>
> inet_csk_find_open_port() rounds the range length down to an even
> number but keeps the original exclusive upper bound for wrapping.
> For an odd-sized range, the two scans can therefore miss a valid port
> and return EADDRINUSE while an eligible port remains free.
>
> For example, restrict a TCP socket to IP_LOCAL_PORT_RANGE [40000, 40002]
> and hold explicit binds on 40000 and 40001. bind(..., 0) always fails,
> although an explicit bind to 40002 succeeds. Both IPv4 and IPv6 TCP and
> MPTCP are affected. A namespace-wide odd-sized ip_local_port_range can
> also trigger the failure without a per-socket range setting.
>
> Keep the full length for the scan bound and wrapping. The two scans
> then cover every candidate in an odd-sized range, with at most one
> repeated candidate when exhausted. Even-sized candidate sequences
> remain unchanged.
>

Please mention the side effect: with an odd @remaining, the wrap
(port -= remaining) flips parity, so the first scan is no longer
restricted to odd offsets. For odd-sized ranges, bind() will now pick
ports of the parity __inet_hash_connect() prefers.

I think this is fine: same-parity netns ranges already trigger
"ip_local_port_range: prefer different parity for start/end values",
and __inet_hash_connect() ignores parity (step == 1) for per-socket
IP_LOCAL_PORT_RANGE. But the changelog should say so.

Also note that __inet_hash_connect() has the same rounding for
netns-wide ranges, so connect() still misses a port there. This is
fine since the sysctl already warns about this setup, but the
changelog should not suggest netns-wide odd ranges are now fully
handled.


> The issue was triggered while developing port-range selftests with
> Codex assistance, then confirmed using userspace reproducers on
> unmodified kernels. No kernel instrumentation or fault injection was
> needed. Booted x86-64 KVM kernels were tested against the regression
> selftest, the full ip_local_port_range suite, and a Go net.Listen
> reproducer. The Go reproducer fails 100/100 times per IP version before
> the fix and 0/100 afterwards with a three-port namespace range and two
> listeners occupying its first two ports.
>
> Fixes: ea8add2b1903 ("tcp/dccp: better use of ephemeral ports in bind()")
> Cc: stable@xxxxxxxxxxxxxxx

This has been there since 2016 and needs an odd-sized range, I would
target net-next and drop the stable Cc.

> Signed-off-by: Florin Badita-Nistor <baditaflorin@xxxxxxxxx>
> ---
>  net/ipv4/inet_connection_sock.c | 2 --
>  1 file changed, 2 deletions(-)
>
> diff --git a/net/ipv4/inet_connection_sock.c b/net/ipv4/inet_connection_sock.c
> index e56beec43..755fcea35 100644
> --- a/net/ipv4/inet_connection_sock.c
> +++ b/net/ipv4/inet_connection_sock.c
> @@ -343,8 +343,6 @@ inet_csk_find_open_port(const struct sock *sk,
> struct inet_bind_bucket **tb_ret,

Both patches are mangled by your mailer (wrapped hunk header here,
wrapped diff --git line in 2/2), they do not apply. Please use
git send-email for v2.

Also, please trim the cover letter to the problem and the test
results.

The code change itself looks good to me.

pw-bot:cr