Re: [PATCH 17/25] hrtimer: Implementation of softirq hrtimer handling

[Peter Zijlstra]

On Tue, Dec 19, 2017 at 09:58:44AM +0100, Sebastian Andrzej Siewior wrote:
> this is late I knowâ
> 
> On 2017-09-27 18:40:26 [+0200], Peter Zijlstra wrote:
> >  - removed superfluous local_bh_disable(), since local_irq_disable()
> >    already implies much the same.
> 
> it is not superfluous.
> 
> > Please consider...
> > 
> > @@ -1768,7 +1786,6 @@ int hrtimers_dead_cpu(unsigned int scpu)
> >  	BUG_ON(cpu_online(scpu));
> >  	tick_cancel_sched_timer(scpu);
> >  
> > -	local_bh_disable();
> >  	local_irq_disable();
> >  	old_base = &per_cpu(hrtimer_bases, scpu);
> >  	new_base = this_cpu_ptr(&hrtimer_bases);
> > @@ -1796,7 +1813,6 @@ int hrtimers_dead_cpu(unsigned int scpu)
> >  	/* Check, if we got expired work to do */
> >  	__hrtimer_peek_ahead_timers();
> >  	local_irq_enable();
> > -	local_bh_enable();
> >  	return 0;
> >  }
> 
> we need in there. That local_bh_disable() is required in order to let
> raise_softirq_irqoff() not do anything stupid in particular we need
> !in_interrupt() defer wakeup_softirqd() until local_bh_enable().
> Otherwise wakeup_softirqd() might actually try to wakeup the process and
> go after the pi_lock which can't happen while holding cpu_base->lock.

Argh, that's horrible and definitely needs a comment.