net/tcp: fix permanent FIN-WAIT-1 state with continuous zero window packets

[RFC net v1] net/tcp: fix permanent FIN-WAIT-1 state with continuous zero window packets

Posted by HaiYang Zhong 4 months, 3 weeks ago

When a TCP connection is in FIN-WAIT-1 state with the FIN packet blocked in
the send buffer, and the peer continuously sends zero-window advertisements,
the current implementation reset the zero-window probe timer while maintaining
the current `icsk->icsk_backoff`, causing the connection to remain permanently
in FIN-WAIT-1 state.

Reproduce conditions:
1. Peer's receive window is full and actively sending continuous zero window
   advertisements.
2. Local FIN packet is blocked in send buffer due to peer's zero-window.
3. Local socket has been closed (entered orphan state).

The root cause lies in the tcp_ack_probe() function: when receiving a zero-window ACK,
- It reset the probe timer while keeping the current `icsk->icsk_backoff`.
- This would result in the condition `icsk->icsk_backoff >= max_probes` false.
- Orphaned socket cannot be set to close.

This patch modifies the tcp_ack_probe() logic: when the socket is dead,
upon receiving a zero-window packet, instead of resetting the probe timer,
we maintain the current timer, ensuring the probe interval grows according
to 'icsk->icsk_backoff', thus causing the zero-window probe timer to eventually
timeout and close the socket.

Signed-off-by: HaiYang Zhong <wokezhong@tencent.com>
---
 net/ipv4/tcp_input.c | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/net/ipv4/tcp_input.c b/net/ipv4/tcp_input.c
index 71b76e98371a..22fc82cb6b73 100644
--- a/net/ipv4/tcp_input.c
+++ b/net/ipv4/tcp_input.c
@@ -3440,6 +3440,8 @@ static void tcp_ack_probe(struct sock *sk)
 	} else {
 		unsigned long when = tcp_probe0_when(sk, tcp_rto_max(sk));
 
+		if (sock_flag(sk, SOCK_DEAD) && icsk->icsk_backoff != 0)
+			return;
 		when = tcp_clamp_probe0_to_user_timeout(sk, when);
 		tcp_reset_xmit_timer(sk, ICSK_TIME_PROBE0, when, true);
 	}
-- 
2.43.7

Re: [RFC net v1] net/tcp: fix permanent FIN-WAIT-1 state with continuous zero window packets

Posted by Eric Dumazet 4 months, 3 weeks ago

On Fri, Sep 19, 2025 at 10:30 AM HaiYang Zhong <wokezhong@gmail.com> wrote:
>
> When a TCP connection is in FIN-WAIT-1 state with the FIN packet blocked in
> the send buffer, and the peer continuously sends zero-window advertisements,
> the current implementation reset the zero-window probe timer while maintaining
> the current `icsk->icsk_backoff`, causing the connection to remain permanently
> in FIN-WAIT-1 state.
>
> Reproduce conditions:
> 1. Peer's receive window is full and actively sending continuous zero window
>    advertisements.
> 2. Local FIN packet is blocked in send buffer due to peer's zero-window.
> 3. Local socket has been closed (entered orphan state).
>
> The root cause lies in the tcp_ack_probe() function: when receiving a zero-window ACK,
> - It reset the probe timer while keeping the current `icsk->icsk_backoff`.
> - This would result in the condition `icsk->icsk_backoff >= max_probes` false.
> - Orphaned socket cannot be set to close.
>
> This patch modifies the tcp_ack_probe() logic: when the socket is dead,
> upon receiving a zero-window packet, instead of resetting the probe timer,
> we maintain the current timer, ensuring the probe interval grows according
> to 'icsk->icsk_backoff', thus causing the zero-window probe timer to eventually
> timeout and close the socket.
>
> Signed-off-by: HaiYang Zhong <wokezhong@tencent.com>
> ---
>  net/ipv4/tcp_input.c | 2 ++
>  1 file changed, 2 insertions(+)
>
> diff --git a/net/ipv4/tcp_input.c b/net/ipv4/tcp_input.c
> index 71b76e98371a..22fc82cb6b73 100644
> --- a/net/ipv4/tcp_input.c
> +++ b/net/ipv4/tcp_input.c
> @@ -3440,6 +3440,8 @@ static void tcp_ack_probe(struct sock *sk)
>         } else {
>                 unsigned long when = tcp_probe0_when(sk, tcp_rto_max(sk));
>
> +               if (sock_flag(sk, SOCK_DEAD) && icsk->icsk_backoff != 0)
> +                       return;
>                 when = tcp_clamp_probe0_to_user_timeout(sk, when);
>                 tcp_reset_xmit_timer(sk, ICSK_TIME_PROBE0, when, true);
>         }
> --
> 2.43.7

Hi there. Seems reasonable, but could you provide a packetdrill test ?

Also, what if the FIN was already sent, but the peer retracted its RWIN ?

tcp_ack_probe() would return early (if (!head) return;)

[PATCH v2 0/2] net/tcp: fix permanent FIN-WAIT-1 state and add packetdrill test