Re: [RFC][PATCH 13/19] perf: Per cpu-context rotation timer

[Peter Zijlstra <peterz@infradead.org>]

On Tue, 2010-09-07 at 19:33 +0200, Thomas Gleixner wrote:
> On Tue, 7 Sep 2010, Peter Zijlstra wrote:
> 
> > On Tue, 2010-09-07 at 18:46 +0200, Peter Zijlstra wrote:
> > 
> > > @@ -5904,6 +5930,9 @@ static void __init perf_event_init_all_c
> > >  
> > >                 cpuctx = &per_cpu(perf_cpu_context, cpu);
> > >                 __perf_event_init_context(&cpuctx->ctx, NULL);
> > > +               cpuctx->timer_interval = TICK_NSEC;
> > > +               hrtimer_init(&cpuctx->timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
> > > +               cpuctx->timer.function = perf_event_context_tick;
> > >         }
> > 
> > > +static void perf_pmu_rotate_start(struct pmu *pmu)
> > > +{
> > > +       struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
> > > +
> > > +       if (hrtimer_active(&cpuctx->timer))
> > > +               return;
> > > +
> > > +       __hrtimer_start_range_ns(&cpuctx->timer,
> > > +                       ns_to_ktime(cpuctx->timer_interval), 0,
> > > +                       HRTIMER_MODE_REL, 0);
> > >  } 
> > 
> > This probably wants a fuzz factor that lets it fold into the tick we
> > already have. Thomas what's the easiest way to do that, give it a soft
> > limit of 1ns and hardlimit of TICK_NSEC?
> 
> Hmm, why don't you hang it off the tick right away ?

Because some people (Corey) want to be able to have different rotation
periods per pmu driver, so a timer per context was the easy way out.