Re: [PATCH] fix-flush_workqueue-vs-cpu_dead-race-update

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

On 01/08, Srivatsa Vaddagiri wrote:
>
> On Mon, Jan 08, 2007 at 12:51:03AM +0300, Oleg Nesterov wrote:
> > Change flush_workqueue() to use for_each_possible_cpu(). This means that
> > flush_cpu_workqueue() may hit CPU which is already dead. However in that
> > case
> > 
> > 	if (!list_empty(&cwq->worklist) || cwq->current_work != NULL)
> > 
> > means that CPU_DEAD in progress, it will do kthread_stop() + take_over_work()
> > so we can proceed and insert a barrier. We hold cwq->lock, so we are safe.
> > 
> > This patch replaces fix-flush_workqueue-vs-cpu_dead-race.patch which was
> > broken by switching to preempt_disable (now we don't need locking at all).
> > Note that migrate_sequence (was hotplug_sequence) is incremented under
> > cwq->lock. Since flush_workqueue does lock/unlock of cwq->lock on all CPUs,
> > it must see the new value if take_over_work() happened before we checked
> > this cwq, and this is the case we should worry about: otherwise we added
> > a barrier.
> > 
> > Srivatsa?
> 
> This is head-spinning :)

Thank you for review!

> Spotted atleast these problems:
> 
> 1. run_workqueue()->work.func()->flush_work()->mutex_lock(workqueue_mutex)
>    deadlocks if we are blocked in cleanup_workqueue_thread()->kthread_stop() 
>    for the same worker thread to exit.
> 
>    Looks possible in practice to me.

Yes, this is the same (old) problem as we have/had with flush_workqueue().
We can convert flush_work() to use preempt_disable (this is not straightforward,
but easy), or forbid to call flush_work() from work.func().

This patch doesn't touch this problem.

> 2. 
>      
> CPU_DEAD->cleanup_workqueue_thread->(cwq->thread = NULL)->kthread_stop() ..
> 				    ^^^^^^^^^^^^^^^^^^^^
> 						|___ Problematic

Hmm... This should not be possible? cwq->thread != NULL on CPU_DEAD event.
Event IF it was NULL, we don't call kthread_stop() in that case.

> Now while we are blocked here, if a work->func() calls
> flush_workqueue->flush_cpu_workqueue, we clearly cant identify that event 
> thread is trying to flush its own queue (cwq->thread == current test
> fails) and hence we will deadlock.

Could you clarify? I believe cwq->thread == current test always works, we never
"substitute" cwq->thread.

> A lock_cpu_hotplug(), or any other ability to block concurrent hotplug 
> operations from happening, in run_workqueue would have avoided both the above
> races.

I still don't think this is a good idea. We also need
	is_cpu_down_waits_for_lock_cpu_hotplug()

helper, otherwise we have a deadlock if work->func() sleeps and re-queues itself.

> Alternatively, for the second race, I guess we can avoid setting 
> cwq->thread = NULL in cleanup_workqueue_thread() till the thread has exited,

Yes, http://marc.theaimsgroup.com/?l=linux-kernel&m=116818097927685, I believe
we can do this later. This way workqueue will have almost zero interaction
with cpu-hotplug, and cpu UP/DOWN event won't be delayed by sleeping work.func().
take_over_work() can go away, this also allows us to simplify things.

Thanks!

Oleg.