Re: [regression] nf_iterate(), BUG: unable to handle kernel NULLpointer dereference

[Krzysztof Oledzki]

On Thu, 24 Jul 2008, Ingo Molnar wrote:

> * Patrick McHardy <kaber@xxxxxxxxx> wrote:
>
> > Ingo Molnar wrote:
> > > Then i tried both suggested fix patches Patrick sent me (a suggested
> > > revert and an netfilter/RCU use-after-free fix), but none of them
> > > solved the crash.
> >
> > Just to make sure - the "netfilter/RCU use-after-free fix" was the
> > patch from Pekka?
>
> yes. You can see it in tip/out-of-tree:
>
>   http://people.redhat.com/mingo/tip.git/README
>
> > > Thus i finally arrived to:
> > >
> > >  # good: [ae6134bd] hdlcdrv: Fix CRC calculation.
> > >  # bad:  [5547cd0d] netfilter: nf_conntrack_sctp: fix sparse warnings
> > >  # bad:  [280763c6] netfilter: xt_time: fix time's time_mt()'s use of
> > >  # good: [07a7c10b] netlink: add NLA_PUT_BE64 macro
> > >  # bad:  [58401573] netfilter: accounting rework: ct_extend + 64bit co
> > >
> > > | 584015727a3b88b46602b20077b46cd04f8b4ab3 is first bad commit
> > > | commit 584015727a3b88b46602b20077b46cd04f8b4ab3
> > > | Author: Krzysztof Piotr Oledzki <ole@xxxxxx>
> > > | AuthorDate: Mon Jul 21 10:01:34 2008 -0700
> > > | Commit:     David S. Miller <davem@xxxxxxxxxxxxx>
> > > | CommitDate: Mon Jul 21 10:10:58 2008 -0700
> > > |
> > > |    netfilter: accounting rework: ct_extend + 64bit counters (v4)
> > > [...]
> > > |     Signed-off-by: Krzysztof Piotr Oledzki <ole@xxxxxx>
> > > |     Signed-off-by: Patrick McHardy <kaber@xxxxxxxxx>
> > > |     Signed-off-by: David S. Miller <davem@xxxxxxxxxxxxx>
> > >
> > > Which i double-checked by reverting that commit from -git as well and
> > > that solved the crash. Find the tested reverter patch below.
> >
> > Thats odd. I don't think anything is wrong with that patch itself, its
> > more likely that its triggering a bug in ct_extend. You config has a
> > few helper enabled (FTP, H.323, TFTP) and the crash is when trying to
> > call the helper functions. Did you actually have traffic of one of
> > these protocols?
>
> no, that's not likely - it's a default distro bootup.

The commit makes ct_extend area to be used *very* frequently. Could you 
try to boot your kernel with nf_conntrack.acct=0 to disable accounting?
Does it help?

Best regards,

				Krzysztof Olędzki