Re: Problem with "tcp: bind() fix when many ports are bound" commit

Re: Problem with "tcp: bind() fix when many ports are bound" commit

[David Miller]

From: Grzegorz Chwesewicz <grzegorz.chwesewicz@retis.net.pl>
Date: Wed, 28 Apr 2010 14:33:45 +0200

> 	Hi, I have a problem with binding to port with the latest git kernel
> (my HEAD is at 1600f9def09de07c5dbeb539e978fa73880690dd). Please CC to
> me as I'm not subscribed to the list.

Please send networking bug reports CC:'d to netdev and the people
who wrote and signed off on the commit you've narrowed down the
problem to.

Otherwise the appropriate people will take longer to find out about
your bug, and therefore the bug will take longer to fix than
necessary.

Thanks.

> 
> Example with buggy kernel:
> 
> ensima-hp ~ # /etc/init.d/apache2 start
>  * Starting apache2 ...
> (98)Address already in use: make_sock: could not bind to address
> 127.0.0.1:80
> no listening sockets available, shutting down
> Unable to open logs
> 
> As you can see nothing is listening on port 80, but there are old
> connections to port 80 with CLOSE_WAIT and FIN_WAIT2 state.
> 
> ensima-hp ~ # netstat -pan --inet|grep 80
> netstat: no support for `AF INET (sctp)' on this system.
> tcp        0      0 127.0.0.1:631           0.0.0.0:*
> LISTEN      4806/cupsd
> tcp        1      0 127.0.0.1:54040         127.0.0.1:80
> CLOSE_WAIT  5814/konquerorHk573
> tcp        0      0 127.0.0.1:80            127.0.0.1:54042
> FIN_WAIT2   -
> tcp        0      0 127.0.0.1:80            127.0.0.1:54040
> FIN_WAIT2   -
> tcp        1      0 127.0.0.1:54042         127.0.0.1:80
> CLOSE_WAIT  6175/konquerordx573
> 
> So I can't start apache as long as these connections are not fully
> closed, after that apache starts without problems.
> 
> ensima-hp ~ # netstat -pan --inet|grep 80
> netstat: no support for `AF INET (sctp)' on this system.
> tcp        0      0 127.0.0.1:631           0.0.0.0:*
> LISTEN      4806/cupsd
> 
> ensima-hp ~ # /etc/init.d/apache2 start
>  * Starting apache2 ...		[OK]
> 
> Problem occured between 2.6.34-rc4 and latest git, bisect shows that the
> problem is caused by:
> 
> commit fda48a0d7a8412cedacda46a9c0bf8ef9cd13559
> Author: Eric Dumazet <eric.dumazet@gmail.com>
> Date:   Wed Apr 21 09:26:15 2010 +0000
> 
> tcp: bind() fix when many ports are bound
> 
> Reverting this commit from current HEAD, resolving conflict in
> 'net/ipv6/inet6_connection_sock.c' file, and compiling new kernel solves
> the problem.
> 
> 'net/ipv6/inet6_connection_sock.c' before resolving conflict:
> 
>  41         sk_for_each_bound(sk2, node, &tb->owners) {
>  42                 if (sk != sk2 &&
>  43                     (!sk->sk_bound_dev_if ||
>  44                      !sk2->sk_bound_dev_if ||
>  45 <<<<<<< HEAD
>  46                      sk->sk_bound_dev_if == sk2->sk_bound_dev_if)) {
>  47                         if ((!sk->sk_reuse || !sk2->sk_reuse ||
>  48                              sk2->sk_state == TCP_LISTEN) &&
>  49                              ipv6_rcv_saddr_equal(sk, sk2))
>  50                                 break;
>  51                         else if (sk->sk_reuse && sk2->sk_reuse &&
>  52                                 !ipv6_addr_any(inet6_rcv_saddr(sk)) &&
>  53                                 ipv6_rcv_saddr_equal(sk, sk2))
>  54                                 break;
>  55                 }
>  56 =======
>  57                      sk->sk_bound_dev_if == sk2->sk_bound_dev_if) &&
>  58                     (!sk->sk_reuse || !sk2->sk_reuse ||
>  59                      sk2->sk_state == TCP_LISTEN) &&
>  60                      ipv6_rcv_saddr_equal(sk, sk2))
>  61                         break;
>  62 >>>>>>> fda48a0... tcp: bind() fix when many ports are bound
>  63         }
>  64
>  65         return node != NULL;
> 
> 
> 
>  66 }
> 
> 'net/ipv6/inet6_connection_sock.c' after resolving conflict:
> 
>  41         sk_for_each_bound(sk2, node, &tb->owners) {
>  42                 if (sk != sk2 &&
>  43                     (!sk->sk_bound_dev_if ||
>  44                      !sk2->sk_bound_dev_if ||
>  45                      sk->sk_bound_dev_if == sk2->sk_bound_dev_if) &&
>  46                     (!sk->sk_reuse || !sk2->sk_reuse ||
>  47                      sk2->sk_state == TCP_LISTEN) &&
>  48                      ipv6_rcv_saddr_equal(sk, sk2))
>  49                         break;
>  50         }
> 
> 
> 
>  51
>  52         return node != NULL;
>  53 }
> 
> -- 
> Greetings
> Grzegorz Chwesewicz
> mailto:grzegorz.chwesewicz@retis.net.pl
> --
> To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html
> Please read the FAQ at  http://www.tux.org/lkml/

Re: Problem with "tcp: bind() fix when many ports are bound" commit

[Eric Dumazet]

Le mercredi 28 avril 2010 à 09:50 -0700, David Miller a écrit :
> From: Grzegorz Chwesewicz <grzegorz.chwesewicz@retis.net.pl>
> Date: Wed, 28 Apr 2010 14:33:45 +0200
> 
> > 	Hi, I have a problem with binding to port with the latest git kernel
> > (my HEAD is at 1600f9def09de07c5dbeb539e978fa73880690dd). Please CC to
> > me as I'm not subscribed to the list.
> 
> Please send networking bug reports CC:'d to netdev and the people
> who wrote and signed off on the commit you've narrowed down the
> problem to.
> 
> Otherwise the appropriate people will take longer to find out about
> your bug, and therefore the bug will take longer to fix than
> necessary.
> 
> Thanks.
> 
> > 
> > Example with buggy kernel:
> > 
> > ensima-hp ~ # /etc/init.d/apache2 start
> >  * Starting apache2 ...
> > (98)Address already in use: make_sock: could not bind to address
> > 127.0.0.1:80
> > no listening sockets available, shutting down
> > Unable to open logs
> > 
> > As you can see nothing is listening on port 80, but there are old
> > connections to port 80 with CLOSE_WAIT and FIN_WAIT2 state.
> > 
> > ensima-hp ~ # netstat -pan --inet|grep 80
> > netstat: no support for `AF INET (sctp)' on this system.
> > tcp        0      0 127.0.0.1:631           0.0.0.0:*
> > LISTEN      4806/cupsd
> > tcp        1      0 127.0.0.1:54040         127.0.0.1:80
> > CLOSE_WAIT  5814/konquerorHk573
> > tcp        0      0 127.0.0.1:80            127.0.0.1:54042
> > FIN_WAIT2   -
> > tcp        0      0 127.0.0.1:80            127.0.0.1:54040
> > FIN_WAIT2   -
> > tcp        1      0 127.0.0.1:54042         127.0.0.1:80
> > CLOSE_WAIT  6175/konquerordx573
> > 
> > So I can't start apache as long as these connections are not fully
> > closed, after that apache starts without problems.
> > 
> > ensima-hp ~ # netstat -pan --inet|grep 80
> > netstat: no support for `AF INET (sctp)' on this system.
> > tcp        0      0 127.0.0.1:631           0.0.0.0:*
> > LISTEN      4806/cupsd
> > 
> > ensima-hp ~ # /etc/init.d/apache2 start
> >  * Starting apache2 ...		[OK]
> > 
> > Problem occured between 2.6.34-rc4 and latest git, bisect shows that the
> > problem is caused by:
> > 
> > commit fda48a0d7a8412cedacda46a9c0bf8ef9cd13559
> > Author: Eric Dumazet <eric.dumazet@gmail.com>
> > Date:   Wed Apr 21 09:26:15 2010 +0000


Oh well...

I am going to test following patch in a couple of hours (I have to
run ...)

Always allowing a reuse LISTEN if no other listener is already active on
same IP.

diff --git a/net/ipv4/inet_connection_sock.c b/net/ipv4/inet_connection_sock.c
index 78cbc39..e2076dd 100644
--- a/net/ipv4/inet_connection_sock.c
+++ b/net/ipv4/inet_connection_sock.c
@@ -78,6 +78,7 @@ int inet_csk_bind_conflict(const struct sock *sk,
 				    sk2_rcv_saddr == sk_rcv_saddr)
 					break;
 			} else if (reuse && sk2->sk_reuse &&
+				   sk->sk_state != TCP_LISTEN &&
 				   sk2_rcv_saddr &&
 				   sk2_rcv_saddr == sk_rcv_saddr)
 				break;

Re: Problem with "tcp: bind() fix when many ports are bound" commit

[Eric Dumazet]

Le mercredi 28 avril 2010 à 19:34 +0200, Eric Dumazet a écrit :
> L
> I am going to test following patch in a couple of hours (I have to
> run ...)
> 
> Always allowing a reuse LISTEN if no other listener is already active on
> same IP.
> 
> diff --git a/net/ipv4/inet_connection_sock.c b/net/ipv4/inet_connection_sock.c
> index 78cbc39..e2076dd 100644
> --- a/net/ipv4/inet_connection_sock.c
> +++ b/net/ipv4/inet_connection_sock.c
> @@ -78,6 +78,7 @@ int inet_csk_bind_conflict(const struct sock *sk,
>  				    sk2_rcv_saddr == sk_rcv_saddr)
>  					break;
>  			} else if (reuse && sk2->sk_reuse &&
> +				   sk->sk_state != TCP_LISTEN &&
>  				   sk2_rcv_saddr &&
>  				   sk2_rcv_saddr == sk_rcv_saddr)
>  				break;
> 
> 
> --

Hmm, this wont work because we dont know this sk socket will serve as a
listener.

bind(... port=80);

listen(...) or connect(...)

David, I think we should revert patches for linux-2.6, and work on this
problem more closely on net-next-2.6

Thanks

Re: Problem with "tcp: bind() fix when many ports are bound" commit

[David Miller]

From: Eric Dumazet <eric.dumazet@gmail.com>
Date: Wed, 28 Apr 2010 19:44:57 +0200

> David, I think we should revert patches for linux-2.6, and work on this
> problem more closely on net-next-2.6

Yep, I agree, I'll send the revert to Linus.

Thanks.

Re: Problem with "tcp: bind() fix when many ports are bound" commit

[Grzegorz Chwesewicz]

David Miller pisze:
> From: Grzegorz Chwesewicz <grzegorz.chwesewicz@retis.net.pl>
> Date: Wed, 28 Apr 2010 14:33:45 +0200
> 
>> 	Hi, I have a problem with binding to port with the latest git kernel
>> (my HEAD is at 1600f9def09de07c5dbeb539e978fa73880690dd). Please CC to
>> me as I'm not subscribed to the list.
> 
> Please send networking bug reports CC:'d to netdev and the people
> who wrote and signed off on the commit you've narrowed down the
> problem to.
> 
> Otherwise the appropriate people will take longer to find out about
> your bug, and therefore the bug will take longer to fix than
> necessary.
> 
> Thanks.

OK, thanks, next time I'll do like you said.

-- 
Pozdrawiam
Grzegorz Chwesewicz
mailto:grzegorz.chwesewicz@retis.net.pl