Re: [PATCH 8/9] hrtimer: Allow hrtimer::function() to free the timer

[Peter Zijlstra <peterz@infradead.org>]

On Thu, Jun 04, 2015 at 12:55:37PM +0200, Peter Zijlstra wrote:
> On Thu, Jun 04, 2015 at 12:49:02PM +0200, Peter Zijlstra wrote:
> > Needs more comments at the very least, its fully of trickery.
> > 
> > @@ -1118,10 +1125,16 @@ static void __run_hrtimer(struct hrtimer
> >  	enum hrtimer_restart (*fn)(struct hrtimer *);
> >  	int restart;
> >  
> > -	WARN_ON(!irqs_disabled());
> > +	lockdep_assert_held(&cpu_base->lock);
> >  
> >  	debug_deactivate(timer);
> > -	base->running = timer;
> > +	cpu_base->running = timer;
> > +
> > +	/*
> > +	 * separate the ->running assignment from the ->state assignment
> > +	 */
> > +	write_seqcount_begin(&cpu_base->seq);
> 
> Maybe these need to be raw_write_seqcount_latch()..

I'm properly confusing my self, so let me write a little more detail;

I didn't think it needed the double wmb because:

	[S] running = timer;
	[S] seq++;
	    WMB
	[S] state = INACTIVE

I don't think it matters if we re-order the first two stores, since at
that point ->state is still ENQUEUED and we're good.

Similar for the case below, you can flip the seq increment and the
->running clear, but at that time ->state should already be ENQUEUED
again (if indeed the timer got re-armed), otherwise its not active
anymore.

> >  	__remove_hrtimer(timer, base, HRTIMER_STATE_INACTIVE, 0);
> >  	timer_stats_account_hrtimer(timer);
> >  	fn = timer->function;
> > @@ -1150,8 +1163,13 @@ static void __run_hrtimer(struct hrtimer
> >  	    !(timer->state & HRTIMER_STATE_ENQUEUED))
> >  		enqueue_hrtimer(timer, base);
> >  
> > -	WARN_ON_ONCE(base->running != timer);
> > -	base->running = NULL;
> > +	/*
> > +	 * separate the ->running assignment from the ->state assignment
> > +	 */
> > +	write_seqcount_end(&cpu_base->seq);
> > +
> > +	WARN_ON_ONCE(cpu_base->running != timer);
> > +	cpu_base->running = NULL;
> >  }
> >  
> >  static void __hrtimer_run_queues(struct hrtimer_cpu_base *cpu_base, ktime_t now)