Re: [PATCH V6] powercap/drivers/idle_injection: Add an idle injection framework

[Peter Zijlstra <peterz@infradead.org>]

On Tue, Jun 12, 2018 at 04:37:17PM +0200, Daniel Lezcano wrote:
> On 12/06/2018 16:06, Peter Zijlstra wrote:
> > On Tue, Jun 12, 2018 at 03:02:14PM +0200, Daniel Lezcano wrote:
> >> On 12/06/2018 14:52, Peter Zijlstra wrote:
> >>> In this case, you can do:
> >>
> >> That is what we had before but we change the code to set the count
> >> before waking up the task, so compute the cpumask_weight of the
> >> resulting AND right before this loop.
> >>
> >>> +       for_each_cpu_and(cpu, &ii_dev->cpumask, cpu_online_mask) {
> >>> +               iit = per_cpu_ptr(&idle_injection_thread, cpu);
> >>> +               iit->should_run = 1;
> >>> +               wake_up_process(iit->tsk);
> >>> +       }
> > 
> > 
> > Ah, I see, but since you do:
> > 
> > 	if (atomic_dec_and_test())
> > 	  last_man()
> > 
> > where that last_man() thing will start a timer, there is no real problem
> > with doing atomic_inc() with before wake_up_process().
> 
> Viresh was worried about the scenario:
> 
> https://lkml.org/lkml/2018/6/5/276

Ah, but I think you have more races, for instance look at wakeup vs
park, what if wakeup sets should_run after you've just checked it?

Then you have an inc without a dec.

> > Also, last_man() uses @run_duration, but the way I read it, the timer is
> > for waking things up again, this means it is in fact the sleep duration,
> > no?
> 
> No, it is the next idle injection deadline, meanwhile we let the system
> continue running.
> 
> The sleep duration is managed by another timer in play_idle().

No, that's the idle duration. Maybe avoid the issue entire by having a
{period,idle} tuple, where your old run := period - idle.

> > Furthermore, should you not be using hrtimer_forward(&timer,
> > idle_duration + run_duration) instead? AFAICT the current scheme is
> > prone to drifting.
> 
> (I assume you meant setting the timer in the wakeup task function).
> 
> Yes, drifting is not an issue if that happens. This scheme is simpler
> and safer than setting the timer ahead before waking up the tasks with
> the risk it expires before all the tasks ended their idle cycles.

sloppy though..