Re: [PATCH]: kthread: Fix memory ordering in __kthread_parkme

[Peter Zijlstra <peterz@infradead.org>]

On Fri, Nov 07, 2014 at 07:41:03PM +0100, Oleg Nesterov wrote:
> On 11/07, Peter Zijlstra wrote:
> >
> >  static void __kthread_parkme(struct kthread *self)
> >  {
> > -	__set_current_state(TASK_PARKED);
> > +	set_current_state(TASK_PARKED);
> >  	while (test_bit(KTHREAD_SHOULD_PARK, &self->flags)) {
> >  		if (!test_and_set_bit(KTHREAD_IS_PARKED, &self->flags))
> >  			complete(&self->parked);
> >  		schedule();
> > -		__set_current_state(TASK_PARKED);
> > +		set_current_state(TASK_PARKED);
> >  	}
> 
> Perhaps it makses sense to do set_current_state(PARKED) once at the start
> of "for (;;)" loop, but this is cosmetic.

Yeah, we should probably clean that up, it looks a bit odd. But I didn't
want to do too many changes.

> What if kthread_unpark() is called right after test_bit(KTHREAD_SHOULD_PARK)
> and KTHREAD_IS_PARKED is not set? It seems that __kthread_unpark() should
> call wake_up_state() unconditionally ?

set_current_state(TASK_PARKED)
while (test_bit(KTHREAD_SHOULD_PARK, ..)) {
						clear_bit(KTHREAD_SHOULD_PARK, ..);
						if (test_and_clear_bit(KTHREAD_IS_PARKED, ..) {
							...
							wake_up_state(, TASK_PARKED);
						}
	if (!test_and_set_bit(KTHREAD_IS_PARKED, ..))
		complete(..);
	schedule();


Then yes we'll miss the wakeup, but we also miss the __kthread_bind().

Now I don't think this'll actually happen because kthread_park() waits
for the completion under the hotplug and smpboot_threads_lock lock, and
we do the unpark under the hotplug lock as well, so its fully serialized

But yes, we should probably clean this up as well.