> -----邮件原件-----
> 发件人: Kuniyuki Iwashima <kuniyu@amazon.com> 
> 发送时间: 2024年11月14日 2:56
> 收件人: mengkanglai <mengkanglai2@huawei.com>
> 抄送: davem@davemloft.net; dsahern@kernel.org; edumazet@google.com; Fengtao (fengtao, Euler) <fengtao40@huawei.com>; kuba@kernel.org; linux-kernel@vger.kernel.org; netdev@vger.kernel.org; pabeni@redhat.com; Yanan (Euler) <yanan@huawei.com>; kuniyu@amazon.com
> 主题: Re: kernel tcp sockets stuck in FIN_WAIT1 after call tcp_close
> 
> From: mengkanglai <mengkanglai2@huawei.com>
> Date: Wed, 13 Nov 2024 12:40:34 +0000
> > Hello, Eric:
> > Commit 151c9c724d05 (tcp: properly terminate timers for kernel 
> > sockets) introduce inet_csk_clear_xmit_timers_sync in tcp_close.
> > For kernel sockets it does not hold sk->sk_net_refcnt, if this is 
> > kernel tcp socket it will call tcp_send_fin in __tcp_close to send FIN 
> > packet to remotes server,
> 
> Just curious which subsystem the kernel socket is created by.
> 
> Recently, CIFS and sunrpc are (being) converted to hold net refcnt.
> 
> CIFS: https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=ef7134c7fc48e1441b398e55a862232868a6f0a7
> sunrpc: https://lore.kernel.org/netdev/20241112135434.803890-1-liujian56@huawei.com/
> 
> I remember RDS's listener does not hold refcnt but other client sockets (SMC, RDS, MPTCP, CIFS, sunrpc) do.
> 
> I think all TCP kernel sockets should hold netns refcnt except for one created at pernet_operations.init() hook like RDS.
> 
> > if this fin packet lost due to network faults, tcp should retransmit 
> > this fin packet, but tcp_timer stopped by inet_csk_clear_xmit_timers_sync.
> > tcp sockets state will stuck in FIN_WAIT1 and never go away. I think 
> > it's not right.


I found this problem when testing nfs. sunrpc: https://lore.kernel.org/netdev/20241112135434.803890-1-liujian56@huawei.com/ will solve this problem. 
I agree with that all TCP kernel sockets should hold netns refcnt.
However, for kernel tcp sockets created by other kernel modules through sock_create_kern or sk_alloc(kern=0), it means that they must now hold sk_net_refcnf, otherwise fin will only be sent once and will not be retransmitted when the socket is released.But other use tcp modules may not be aware of hold sk_net_refcnt. should we add a check in tcp_close？

---
diff --git a/net/ipv4/tcp.c b/net/ipv4/tcp.c
index fb920369c..6b92026a4 100644
--- a/net/ipv4/tcp.c
+++ b/net/ipv4/tcp.c
@@ -2804,7 +2804,7 @@ void tcp_close(struct sock *sk, long timeout)
        lock_sock(sk);
        __tcp_close(sk, timeout);
        release_sock(sk);
-       if (!sk->sk_net_refcnt)
+       if (sk->net != &init_net && !sk->sk_net_refcnt)
                inet_csk_clear_xmit_timers_sync(sk);
        sock_put(sk);
 }

From: mengkanglai <mengkanglai2@huawei.com>
Date: Tue, 19 Nov 2024 08:38:26 +0000
> > 
> > From: mengkanglai <mengkanglai2@huawei.com>
> > Date: Wed, 13 Nov 2024 12:40:34 +0000
> > > Hello, Eric:
> > > Commit 151c9c724d05 (tcp: properly terminate timers for kernel 
> > > sockets) introduce inet_csk_clear_xmit_timers_sync in tcp_close.
> > > For kernel sockets it does not hold sk->sk_net_refcnt, if this is 
> > > kernel tcp socket it will call tcp_send_fin in __tcp_close to send FIN 
> > > packet to remotes server,
> > 
> > Just curious which subsystem the kernel socket is created by.
> > 
> > Recently, CIFS and sunrpc are (being) converted to hold net refcnt.
> > 
> > CIFS: https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=ef7134c7fc48e1441b398e55a862232868a6f0a7
> > sunrpc: https://lore.kernel.org/netdev/20241112135434.803890-1-liujian56@huawei.com/
> > 
> > I remember RDS's listener does not hold refcnt but other client sockets (SMC, RDS, MPTCP, CIFS, sunrpc) do.
> > 
> > I think all TCP kernel sockets should hold netns refcnt except for one created at pernet_operations.init() hook like RDS.
> > 
> > > if this fin packet lost due to network faults, tcp should retransmit 
> > > this fin packet, but tcp_timer stopped by inet_csk_clear_xmit_timers_sync.
> > > tcp sockets state will stuck in FIN_WAIT1 and never go away. I think 
> > > it's not right.
> 
> 
> I found this problem when testing nfs. sunrpc: https://lore.kernel.org/netdev/20241112135434.803890-1-liujian56@huawei.com/ will solve this problem. 
> I agree with that all TCP kernel sockets should hold netns refcnt.
> However, for kernel tcp sockets created by other kernel modules through
> sock_create_kern or sk_alloc(kern=0),

In the next cycle, I'll rename sock_create_kern() to sock_create_net_noref()
and add sock_create_net() so that out-of-tree module will fail to build and
such users will notice sock_create_net_noref() would trigger an issue.

https://github.com/q2ven/linux/commits/427_2


> it means that they must now hold
> sk_net_refcnf, otherwise fin will only be sent once and will not be
> retransmitted when the socket is released.But other use tcp modules may
> not be aware of hold sk_net_refcnt. should we add a check in tcp_close？

The check doesn't fix the issue for in-netns users.

I'd rather print the allocator and change it to use
sock_create_net() instead.

---8<---
diff --git a/net/ipv4/tcp.c b/net/ipv4/tcp.c
index 0d704bda6c41..7d6a1faa05a3 100644
--- a/net/ipv4/tcp.c
+++ b/net/ipv4/tcp.c
@@ -3220,8 +3220,12 @@ void tcp_close(struct sock *sk, long timeout)
 	lock_sock(sk);
 	__tcp_close(sk, timeout);
 	release_sock(sk);
+
+#ifdef CONFIG_NET_NS_REFCNT_TRACKER
 	if (!sk->sk_net_refcnt)
-		inet_csk_clear_xmit_timers_sync(sk);
+		stack_depot_print(sk->ns_tracker);
+#endif
+
 	sock_put(sk);
 }
 EXPORT_SYMBOL(tcp_close);
---8<---

> 
> ---
> diff --git a/net/ipv4/tcp.c b/net/ipv4/tcp.c
> index fb920369c..6b92026a4 100644
> --- a/net/ipv4/tcp.c
> +++ b/net/ipv4/tcp.c
> @@ -2804,7 +2804,7 @@ void tcp_close(struct sock *sk, long timeout)
>         lock_sock(sk);
>         __tcp_close(sk, timeout);
>         release_sock(sk);
> -       if (!sk->sk_net_refcnt)
> +       if (sk->net != &init_net && !sk->sk_net_refcnt)
>                 inet_csk_clear_xmit_timers_sync(sk);
>         sock_put(sk);
>  }