Re: [Bug 86851] New: Reproducible panic on heavy UDP traffic

[Nikolay Aleksandrov <nikolay@redhat.com>]

On 10/27/2014 11:59 PM, Patrick McLean wrote:
> On Mon, 27 Oct 2014 09:48:15 +0100
> Nikolay Aleksandrov <nikolay@redhat.com> wrote:
> 
>> On 10/27/2014 01:47 AM, Eric Dumazet wrote:
>>> On Mon, 2014-10-27 at 00:28 +0100, Nikolay Aleksandrov wrote:
>>>
>>>>
>>>> Thanks for CCing me.
>>>> I'll dig in the code tomorrow but my first thought when I saw this
>>>> was could it be possible that we have a race condition between
>>>> ip_frag_queue() and inet_frag_evict(), more precisely between the
>>>> ipq_kill() calls from ip_frag_queue and inet_frag_evict since the
>>>> frag could be found before we have entered the evictor which then
>>>> can add it to its expire list but the ipq_kill() from
>>>> ip_frag_queue() can do a list_del after we release the chain lock
>>>> in the evictor so we may end up like this ?
>>>
>>> Yes, either we use hlist_del_init() but loose poison aid, or test if
>>> frag was evicted :
>>>
>>> Not sure about refcount.
>>>
>>> diff --git a/net/ipv4/inet_fragment.c b/net/ipv4/inet_fragment.c
>>> index 9eb89f3f0ee4..894ec30c5896 100644
>>> --- a/net/ipv4/inet_fragment.c
>>> +++ b/net/ipv4/inet_fragment.c
>>> @@ -285,7 +285,8 @@ static inline void fq_unlink(struct
>>> inet_frag_queue *fq, struct inet_frags *f) struct inet_frag_bucket
>>> *hb; 
>>>  	hb = get_frag_bucket_locked(fq, f);
>>> -	hlist_del(&fq->list);
>>> +	if (!(fq->flags & INET_FRAG_EVICTED))
>>> +		hlist_del(&fq->list);
>>>  	spin_unlock(&hb->chain_lock);
>>>  }
>>>  
>>>
>>>
>>
>> Exactly, I was thinking about a similar fix since the evict flag is
>> only set with the chain lock. IMO the refcount should be fine.
>> CCing the reporter.
>> Patrick could you please try Eric's patch ?
>>
> 
> It no longer panics with that patch, but it does produce a large amount
> of warnings, here is an example of what I am getting. I will attach the
> full log to the bug.
> 

Great! Thanks for testing.
As I said earlier we have a valid case that can hit the WARN_ON in
inet_evict_frag().
Anyhow, Eric would you mind posting the patch officially ?
If you'd like me to remove the WARN_ON() in a separate one just let me
know, otherwise feel free to remove it in the fix for the race.

Cheers,
 Nik