Re: Fwd: Re: [BUG] Fatal exception in interrupt - nf_nat_cleanup_conntrack during IPv6 tests

[Patrick McHardy <kaber@trash.net>]

On Wed, Apr 10, 2013 at 04:56:21PM +0200, Florian Westphal wrote:
> Patrick McHardy <kaber@trash.net> wrote:
> > On Wed, Apr 10, 2013 at 11:32:04AM +0200, Florian Westphal wrote:
> > > Patrick McHardy <kaber@trash.net> wrote:
> > > > On Wed, Apr 10, 2013 at 11:04:36AM +0200, Florian Westphal wrote:
> > > > > > [ 3599.241868] Code: 83 ec 08 0f b6 58 11 84 db 74 43 48 01 c3 48 83 7b 20 00 74 39 48 c7 c7 b8 65 32 a0 e8 98 fc 2e e1 48 8b 03 48 8b 53 08 48 85 c0 <48> 89 02 74 04 48 89 50 08 48 ba 00 02 20 00 00 00 ad de 48 c7  
> > > > > > [ 3599.337037] RIP  [<ffffffffa03227f2>] nf_nat_cleanup_conntrack+0x42/0x70 [nf_nat] 
> > > > > 
> > > > > Looks like we tried to remove bysource hash twice (rdx is
> > > > > LIST_POISON_2).
> > > > > 
> > > > > I wonder if this would explain it:
> > > > > 
> > > > > static void nf_nat_l4proto_clean(u8 l3proto, u8 l4proto)
> > > > > {
> > > > > [..]
> > > > >       /* Step 1 - remove from bysource hash */
> > > > >       clean.hash = true;
> > > > >       for_each_net(net)
> > > > >                 nf_ct_iterate_cleanup(net, nf_nat_proto_clean, &clean);
> > > > > 
> > > > > A nfct->timer fires and a conntrack is free'd before step 2 memsets the
> > > > > nat extension.  In that case, we would try to delete nat->bysource
> > > > > again?
> > > > 
> > > > Not sure I follow, we only invoke nf_nat_l4proto_clean() through
> > > > nf_nat_l4proto_unregister(), right?
> > > >
> > > > Did this happen during module unload?
> > > 
> > > Looks like it, nf_nat_ipv4 is listed as F- in the oops trace. (afaics,
> > > "-" means "module going away").
> > 
> > Yes, that seems like a real race condition. We probably could extend the
> > nf_nat_lock sections to avoid this, but I wonder wether we should just kill
> > those conntracks, the connections are not going to work after being
> > "de-nated" anymore anyway.
> 
> I like it, just killing them would make it a lot more simple.
> 
> The clear-nat-extension-on-module-unload dance is getting out of hand,
> and, as you point out, the connections are not going to work anyway...

Yeah, lets just do that. Do you want to take care of this?