Re: ath5k/mac80211: Reproducible deadlock with 64-stations.

[Ben Greear <greearb@candelatech.com>]

On 11/12/2010 10:13 AM, Tejun Heo wrote:
> Hello,
>
> On 11/12/2010 07:06 PM, Ben Greear wrote:
>> On 11/12/2010 02:15 AM, Tejun Heo wrote:
>>> Please note that under those circumstances, what's guaranteed is
>>> forward-progress for workqueues which are used during memory reclaim.
>>> Continuously scheduling works which will in turn pile up on rtnl_lock
>>> is akin to constantly allocating memory while something holding
>>> rtnl_lock is blocked due to memory pressure.  Correctness-wise, it
>>> isn't necessarily deadlock but the only possible recourse is OOM.
>>
>>  From looking at the wireless code, since sdata is stopped, the
>> 'work' isn't going to actually do anything anyway.
>>
>> Is there a way to clear the work from the work-queue w/out
>> requiring any locks that a running worker thread might hold?
>> (So instead of flush_work, we could call something like "remove_all_work"
>> and not block on the worker thread that may currently be trying to
>> acquire rtnl?)
>
> Hmmm... there's cancel_work_sync().  It'll cancel if the work is
> pending and wait for completion if it's already running.  BTW, which

That would help, but it *might* be possible that the worker thread is
currently active.  That work shouldn't be asking for rtnl, as far as I can
tell, so maybe that's OK.  I'll give that a try in a bit.

> part of code are we talking about?  Can you please attach full thread
> dump at deadlock?

The problem code seems to be flush_work() call in ieee80211_do_stop()
in net/mac80211/iface.c.  RTNL is held when the flush_work() method
is called.

I've seen this apparently deadlock with a worker process trying to
call the wireless_nlevent_process method in net/wireless/wext-core.c
(it acquires rtnl).

Please let me know if you saw the previous thread dumps I sent in this thread.
Those were only for processes blocked > 120secs.  I can probably get
a full sysrq dump if you want that instead.

Thanks,
Ben

-- 
Ben Greear <greearb@candelatech.com>
Candela Technologies Inc  http://www.candelatech.com