Re: System freeze on reboot - general protection fault

[Eric Dumazet <eric.dumazet@gmail.com>]

Zdenek Kabelac a écrit :
> 2009/8/17 Patrick McHardy <kaber@trash.net>:
>> Eric Dumazet wrote:
>>> Zdenek Kabelac a écrit :
>>>>  [<ffffffffa02c502f>] nf_conntrack_ftp_fini+0x2f/0x70 [nf_conntrack_ftp]
>>>>  [<ffffffff8027bcc5>] sys_delete_module+0x1a5/0x270
>>>>  [<ffffffff8020d329>] ? retint_swapgs+0xe/0x13
>>>>  [<ffffffff80271bf2>] ? trace_hardirqs_on_caller+0x162/0x1b0
>>>>  [<ffffffff80292121>] ? audit_syscall_entry+0x191/0x1c0
>>>>  [<ffffffff80526dae>] ? trace_hardirqs_on_thunk+0x3a/0x3f
>>>>  [<ffffffff8020c84b>] system_call_fastpath+0x16/0x1b
>>>> Code: c6 00 00 0f 82 66 ff ff ff 49 8b 9e d8 05 00 00 48 85 db 75 16
>>>> e9 8e 00 00 00 0f 1f 44 00 00 48 85 c0 0f 84 80 00 00 00 48 89 c3 <0f>
>>>> b6 4b 37 48 8b 03 48 8d 14 cd 00 00 00 00 0f 18 08 48 29 ca
>>>> RIP  [<ffffffffa02b2c2c>] nf_conntrack_helper_unregister+0x16c/0x320
>>>> [nf_conntrack]
>>>>  RSP <ffff88013982fe68>
>>>> CR2: 0000000000000038
>>>> ---[ end trace bc3a0ede3d0084db ]---
>>>>
>>> I am currently traveling and wont be able to help you before next week.
>>>
>>> I added netdev, Patrick, and netfilter-devel in CC so that more eyes can take a look.
>> Thanks for the report, I'll have a look at this. Zdenek, please
>> send me the nf_conntrack.ko file used in the above oops. Thanks.
>>
> 
> Ok
> 
> I've found the solution for my problem.
> 
> http://thread.gmane.org/gmane.comp.security.firewalls.netfilter.devel/30483
> 
> I've made this small fix from this thread:
> 
> diff --git a/net/netfilter/nf_conntrack_core.c b/net/netfilter/nf_conntrack_core
> index b5869b9..68488f8 100644
> --- a/net/netfilter/nf_conntrack_core.c
> +++ b/net/netfilter/nf_conntrack_core.c
> @@ -1108,6 +1108,7 @@ static void nf_conntrack_cleanup_init_net(void)
>  {
>         nf_conntrack_helper_fini();
>         nf_conntrack_proto_fini();
> +       rcu_barrier();
>         kmem_cache_destroy(nf_conntrack_cachep);
>  }
> 
> @@ -1266,7 +1267,7 @@ static int nf_conntrack_init_init_net(void)
> 
>         nf_conntrack_cachep = kmem_cache_create("nf_conntrack",
>                                                 sizeof(struct nf_conn),
> -                                               0, SLAB_DESTROY_BY_RCU, NULL);
> +                                               0, 0, NULL);
>         if (!nf_conntrack_cachep) {
>                 printk(KERN_ERR "Unable to create nf_conn slab cache\n");
>                 ret = -ENOMEM;
> 
> 
> As the thread nf_conntrack: Use rcu_barrier() and fix kmem_cache_create flags
> seems to be samewhat 'unfinished'  and already a bit old and I've no
> idea whether it actually fixes problem completely or just hides it in
> my case - I'm leaving it to some RCU gurus to fix this issue.
> 
> All I could say is - this this extra rcu_barrier() and removal of
> SLAB_DESTROY removes my GPF on reboot.
> 
> Zdenek

Ouch..

Dont think such a patch makes your kernel better, it'll crash too.

You cannot remove SLAB_DESTROY_BY_RCU like this, it's there for very good reasons.