Re: [PATCH v5 08/20] kthread: Allow to cancel kthread work

[Petr Mladek]

On Fri 2016-02-26 17:25:52, Peter Zijlstra wrote:
> On Fri, Feb 26, 2016 at 04:38:18PM +0100, Petr Mladek wrote:
> > On Thu 2016-02-25 13:59:32, Peter Zijlstra wrote:
> > > On Wed, Feb 24, 2016 at 05:18:05PM +0100, Petr Mladek wrote:
> > > > @@ -770,7 +782,22 @@ void delayed_kthread_work_timer_fn(unsigned long __data)
> > > >  	if (WARN_ON_ONCE(!worker))
> > > >  		return;
> > > >  
> > > > -	spin_lock(&worker->lock);
> > > > +	/*
> > > > +	 * We might be unable to take the lock if someone is trying to
> > > > +	 * cancel this work and calls del_timer_sync() when this callback
> > > > +	 * has already been removed from the timer list.
> > > > +	 */
> > > > +	while (!spin_trylock(&worker->lock)) {
> > > > +		/*
> > > > +		 * Busy wait with spin_is_locked() to avoid cache bouncing.
> > > > +		 * Break when canceling is set to avoid a deadlock.
> > > > +		 */
> > > > +		do {
> > > > +			if (work->canceling)
> > > > +				return;
> > > > +			cpu_relax();
> > > > +		} while (spin_is_locked(&worker->lock));
> > > > +	}
> > > >  	/* Work must not be used with more workers, see queue_kthread_work(). */
> > > >  	WARN_ON_ONCE(work->worker != worker);
> > > >  
> > > 
> > > This is pretty vile; why can't you drop the lock over del_timer_sync() ?
> > 
> > We would need to take the lock later and check if nobody has set the timer
> > again in the meantime.
> 
> Well, if ->cancelling is !0, nobody should be re-queueing, re-arming
> timers etc.., right?

Which rings a bell. __cancel_kthread_work()/del_timer_sync() is used
also from mod_timer_sync() and we do not increment ->canceling there.
See the 9th patch. It is racy and I have to fix it.


> And since you do add_timer() while holding the spinlock, this should all
> work out, no?

Interesting idea. Yes, it should work. But is this really easier? The
try_again/relock/recheck code is not trivial either.

I personally slightly more prefer the current code. I am open
to use your version ff you persist on it. But I also do not want
to end up in a ping-pong re-implementation as Tejun king of suggested
the current code.


Thanks a lot,
Petr