Re: PROBLEM with summary: Re: [PATCH] netfilter: per netns nf_conntrack_cachep

[Patrick McHardy <kaber@trash.net>]

Jon Masters wrote:
> On Tue, 2010-02-02 at 11:46 -0500, Jon Masters wrote:
>> On Tue, 2010-02-02 at 06:35 -0500, Jon Masters wrote:
>>
>>> I think there's something more fundamental going on here.
>> What happens is the conntrack code attempts to free
>> nf_conntrack_untracked back into the SL[U]B cache from which it
>> allocates other ct's. There's just one problem...that's a static struct
>> not from the cache. So, this is why we end up with the SLAB being
>> corrupted and the address immediately following the
>> nf_conntrack_untracked being corrupted.
>>
>> I shoved some debug comments into the destroy code to see if we were
>> trying to free nf_conntrack_untracked, and bingo. I have shoved a panic
>> in there now, will send you a backtrace.
> 
> So we attach after starting a VM due to icmpv6_error setting the ct in
> some incoming skb to the "untracked" catchall one. Then we panic when I
> force a panic if attempting to free the nf_conntrack_untracked static
> struct with this backtrace:
> 
> #5  0xffffffff81455884 in panic (
>     fmt=0xffffffff81ec51e0 "JCM: nf_conntrack_destroy: trying to destroy
> nf_conntrack_untracked!\n")
>     at kernel/panic.c:73
> #6  0xffffffff813d266c in nf_conntrack_destroy (nfct=<value optimized
> out>) at net/netfilter/core.c:244
> #7  0xffffffff813abd97 in nf_conntrack_put (skb=0xffff880223b8b700) at
> include/linux/skbuff.h:1924
> #8  skb_release_head_state (skb=0xffff880223b8b700) at
> net/core/skbuff.c:402
> #9  0xffffffff813abaf9 in skb_release_all (skb=0xffff880223b8b700) at
> net/core/skbuff.c:420
> #10 __kfree_skb (skb=0xffff880223b8b700) at net/core/skbuff.c:435
> #11 0xffffffff813abbfe in kfree_skb (skb=0xffff880223b8b700) at
> net/core/skbuff.c:456
> 
> Could you please add (or recommend for me to test) some logic to catch
> attempts to free the nf_conntrack_untracked and prevent it? Also, maybe
> in the init_net code you could remove the re-initialization of the
> untracked conntrack each time a namespace is created?

Ah nice catch, that seems to be the problem. When the untracked
conntrack is already attached to an skb and thus has refcnt > 1
and we re-initalize the refcnt, it will get freed.

The question is whether the ct_net pointer of the untracked conntrack
is actually required. If so, we need one instance per namespace,
otherwise we can just move initialization and cleanup to the init_net
init/cleanup functions. Alexey, do you happen to know this?