Re: [PATCH, RFC] reimplement flush_workqueue()

[Oleg Nesterov <oleg@tv-sign.ru>]

On 01/04, Srivatsa Vaddagiri wrote:
>
> On Thu, Jan 04, 2007 at 05:29:36PM +0300, Oleg Nesterov wrote:
> > Thanks, I need to think about this.
> > 
> > However I am not sure I fully understand the problem.
> > 
> > First, this deadlock was not introduced by recent changes (including "single
> > threaded flush_workqueue() takes workqueue_mutex too"), yes?
> 
> AFAIK this deadlock originated from Andrew's patch here:
> 
> 	http://lkml.org/lkml/2006/12/7/231

I don't think so. The core problem is not that we are doing unlock/sleep/lock
with this patch. The thing is: work->func() can't take wq_mutex (and thus use
flush_work/workqueue) because it is possible that CPU_DEAD holds this mutex
and waits for us to complete(kthread_stop_info). I believe this bug is old.

> (Yes, your patches didnt introduce this. I was just reiterating here my
> earlier point that workqueue code is broken of late wrt cpu hotplug).
> 
> > Also, it seems to me we have a much more simple scenario for deadlock.
> > 
> > events/0 runs run_workqueue(), work->func() sleeps or takes a preemtion. CPU 0
> > dies, keventd thread migrates to another CPU. CPU_DEAD calls kthread_stop() under
> > workqueue_mutex and waits for until kevents thread exits. Now, if this work (or
> > another work pending on cwq->worklist) takes workqueue_mutex (for example, does
> > flush_workqueue) we have a deadlock.
> > 
> > No?
> 
> Yes, the above scenario also will cause a deadlock. 

Ok, thanks for acknowledgement.

> I supposed one could avoid the deadlock by having a 'workqueue_mutex_held' 
> flag and avoid taking the mutex set under some conditions,

I am thinking about the same right now. Probably we can do something like this:


	int xxx_lock(void)
	{
		for (;;) {
			if (mutex_trylock(wq_mutex))
				return 1;

			// the owner of wq_mutex sleeps, we can proceed
			if (kthread_should_stop())
				return 0;
		}
	}
	void xxx_unlock(int yesno)
	{
		if (yesno)
			mutex_unlock(wq_mutext);
	}

and then do

	locked = xxx_lock();
	...
	xxx_unlock(locked);

in flush_xxx() instead of plain lock/unlock.

Yes, ugly. I'll try to do something else on weekend.

>                                                             but IMHO a
> more neater solution is to provide a cpu-hotplug lock which works under
> all these corner cases. One such proposal was made here:
> 
> 	http://lkml.org/lkml/2006/10/26/65

I'll take a look later, thanks.

Oleg.

-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/