Re: [PATCH net] net: neighbour: add neighbour dead check for neigh_timer_handler()

[Ding Tianhong <dingtianhong@huawei.com>]

On 2013/12/18 23:46, Hannes Frederic Sowa wrote:
> On Wed, Dec 18, 2013 at 11:12:51PM +0800, Ding Tianhong wrote:
>> 于 2013/12/18 22:27, Hannes Frederic Sowa 写道:
>>> On Wed, Dec 18, 2013 at 07:57:40PM +0800, Ding Tianhong wrote:
>>>> yes, I cannot repruduce the bug again.
>>>
>>> Hmm, it actually seems hard to hit even if the race happens. Even if slab
>>> poisoning is active it would only hit if ->solicit would be called again,
>>> because that is the only pointer dereference directly used in the old memory.
>>>
>>> neigh_alloc allocates memory with kzalloc, so it would null out that memory,
>>> so the race would not only have to race with kfree, the memory needs to be
>>> reallocated in the mean time.
>>>
>>> I would suggest adding some poisoning manually in neigh_release before kfree
>>> and check for this in all periodic called functions. Maybe we can see it
>>> again?
>>>
>> Great, thanks for your help, I think make the neigh_release not kfree neighbour until
>> the timer is over is a clear way to  fix this, maybe you could another idea, glad to
>> hear your opinion.
> 
> But I don't suggest this as an fix, just as a help for debugging this issue.
> 
> Maybe you could also store the _RET_IP_ in the to be freed struct neighbour
> (just before kfree) and thus have it available in case the machine panics (or
> simply print it with printk).
> 
> Maybe it would make sense to use kmem_cache_create and kmem_cache_alloc for
> struct neighs so we can better utilize the slub debugging features.
> 
> Greetings,
> 
>   Hannes
> 
> 

Good idea, I will try it, but I still could not make it happen again.

I can repeat the process that the problem happed:

(1).A: xxx.xxx.xxx.83, B:xxx.xxx.xxx.84
(2). down A, B instead of A, ifconfig B xxx.xxx.xxx.83
(3).use "/sbin/arping -I %s -U -b -c 1 -w 4 %s "to tell vlan B is xxx.xxx.xxx.83,
(4). then it happened. 

Regards
Ding
> .
>