Re: [PATCH net] netfilter: nf_conntrack: prevent uninit-value in gc_worker

[Dmitry Vyukov <dvyukov@google.com>]

On Thu, Jul 12, 2018 at 2:11 PM, Eric Dumazet <eric.dumazet@gmail.com> wrote:
>
>
> On 07/12/2018 02:00 AM, Florian Westphal wrote:
>> Eric Dumazet <edumazet@google.com> wrote:
>>> KMSAN reported use of uninit-value in gc_worker [1]
>>>
>>> We need to clear ct->timeout in __nf_conntrack_alloc()
>>> otherwise __nf_conntrack_confirm() might propagate garbage when
>>> adding nfct_time_stamp to ct->timeout :
>>>
>>>      ct->timeout += nfct_time_stamp;
>>>
>>> [1]
>>> BUG: KMSAN: uninit-value in gc_worker+0x89e/0x1530 net/netfilter/nf_conntrack_core.c:1028
>>> CPU: 1 PID: 19 Comm: kworker/1:0 Not tainted 4.18.0-rc4+ #24
>>> Hardware name: Google Google Compute Engine/Google Compute Engine, BIOS Google 01/01/2011
>>> Workqueue: events_power_efficient gc_worker
>>> Call Trace:
>>>  __dump_stack lib/dump_stack.c:77 [inline]
>>>  dump_stack+0x185/0x1e0 lib/dump_stack.c:113
>>>  kmsan_report+0x195/0x2c0 mm/kmsan/kmsan.c:1092
>>>  __msan_warning_32+0x7d/0xe0 mm/kmsan/kmsan_instr.c:640
>>>  gc_worker+0x89e/0x1530 net/netfilter/nf_conntrack_core.c:1028
>>
>> I wonder how this can happen.
>>
>> All trackers are supposed to set ->timeout to the correct value,
>> otherwise (assuming init-to-0), we add a ct entry to global hash that
>> is expired.
>>
>> For instance, tcp calls
>> nf_ct_refresh_acct() at end of its ->packet() callback to set
>> a timeout based on the connection state.
>>
>> That being said, I don't see any harm in initing to 0 of course.
>>
>
> Yeah, unfortunately there is no repro yet, all the info I have I put it
> in the changelog.


What should have been initialized it?
I assume it should have been happened in between init_conntrack and
nf_conntrack_confirm, because nf_conntrack_confirm already adds to an
uninit timeout value.
Since we got only 3 such reports and no reproducer, I would suspect
that there is some race involved. Is it possible that timeout
initialization (presumably a call to nf_ct_refresh_acct) happens after
and non-atomically with the corresponding connection state update, so
that the call to nf_conntrack_confirm sneaks before it?