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

[Ding Tianhong <dthxman@gmail.com>]

于 2013/12/18 18:21, Hannes Frederic Sowa 写道:
> On Wed, Dec 18, 2013 at 06:02:33PM +0800, Ding Tianhong wrote:
>> On 2013/12/18 17:28, Hannes Frederic Sowa wrote:
>>> On Wed, Dec 18, 2013 at 04:57:01PM +0800, Ding Tianhong wrote:
>>>> On 2013/12/18 16:41, Hannes Frederic Sowa wrote:
>>>>> On Wed, Dec 18, 2013 at 04:19:43PM +0800, Ding Tianhong wrote:
>>>>>> 0xffffffff812f8e29 <neigh_timer_handler+265>:   mov    0xe8(%rbx),%rax
>>>>>> 0xffffffff812f8e30 <neigh_timer_handler+272>:   mov    %rbp,%rsi
>>>>>> 0xffffffff812f8e33 <neigh_timer_handler+275>:   mov    %rbx,%rdi
>>>>>> 0xffffffff812f8e36 <neigh_timer_handler+278>:   callq  *0x8(%rax)						<-----crash
>>>>>> /usr/src/linux/net/core/neighbour.c: 877
>>>>>> 0xffffffff812f8e39 <neigh_timer_handler+281>:   lea    0x3c(%rbx),%rax
>>>>>
>>>>> For me it looks like this:
>>>>>
>>>>> %rax is neigh->ops and the function pointer solicit is NULL and causes the the
>>>>> page fault.
>>>>>
>>>>>
>>>> yes, it is. So I was trying to find the situation that may free the neighbour when
>>>> the timer is running, but I could not yet.
>>>
>>> Hm. Ok. It is actually ops which is NULL, not the function pointer, may bad.
>>>
>>> Could you try to follow param or table links and check if this is an arp or
>>> ndisc one? Maybe some interactions with arp.c or ndisc.c causes this bug?
>>>
>>>
>>
>> David and Eric has said that someone may called neigh_release in a wrong place, I agree with that,
>> and review the code which calling the function in the kernel, I could not find any obvious problem,
>> and doubt with the situation:
> 
> Maybe it could also be a reference count overflow and we wrap around to zero
> again? Otherwise I agree, it really looks like this is the case.
> 
>> CPU0					  CPU1					  CPU2
>>       --------		      			--------                		---------
>> neigh_timer_handler				
>> write_lock(n->lock);		
>> 	...
>> write_unlock(n->lock);
>> n->ref_cnt = 2 or 3(if mode_time)				
> 
> 
> 
>> 	...					neigh_flush_dev
>> 						write_lock(n->lock);
>> 						n->ref_cnt = 2;
>> 						n->nud_state = NUD_NONE;
>> 						write_unlock(n->lock);
>> 						neigh_release()
>> 						n->ref_cnt = 1;
>> 						...					neigh_periodic_work
>> 											write_lock(n->lock);
>> 											write_unlock(n->lock);
>> 											neigh_release();
>> 											kfree(n)
>> n->ops->solicit()									...
> 
> On CPU0 the neigh_release happens after solicit. So the timer_handler should
> still be guarded to not touch already freed memory. The table lock should make
> sure that we either see a reference from the hash table or we don't (with
> appropriate reference count). It looks consistent for me for now.
> 
> I guess you cannot reproduce this?
> 
> Greetings,
> 
>   Hannes
> 

yes, I cannot repruduce the bug again.

Regards
Ding
> --
> To unsubscribe from this list: send the line "unsubscribe netdev" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html
>