Re: [arc-linux-dev] Re: [PATCH v3 3/6] ARCv2: perf: Support sampling events using overflow interrupts

[Alexey Brodkin <Alexey.Brodkin@synopsys.com>]

Hi Peter,

On Wed, 2015-08-26 at 16:35 +0200, Peter Zijlstra wrote:
> On Wed, Aug 26, 2015 at 01:17:20PM +0000, Alexey Brodkin wrote:
> > Hi Peter,
> > 
> > On Wed, 2015-08-26 at 15:07 +0200, Peter Zijlstra wrote:
> > > On Mon, Aug 24, 2015 at 05:20:20PM +0300, Alexey Brodkin wrote:
> > > > @@ -139,9 +141,11 @@ static int arc_pmu_event_init(struct perf_event *event)
> > > >  	struct hw_perf_event *hwc = &event->hw;
> > > >  	int ret;
> > > >  
> > > > -	hwc->sample_period  = arc_pmu->max_period;
> > > > -	hwc->last_period = hwc->sample_period;
> > > > -	local64_set(&hwc->period_left, hwc->sample_period);
> > > > +	if (!is_sampling_event(event)) {
> > > > +		hwc->sample_period  = arc_pmu->max_period;
> > > > +		hwc->last_period = hwc->sample_period;
> > > > +		local64_set(&hwc->period_left, hwc->sample_period);
> > > > +	}
> > > 
> > > So here we set a max_period sample period for !sampling events such that
> > > we can properly deal with (short) counter overflow and accumulate into a
> > > 64bit value.
> > > 
> > > >  	switch (event->attr.type) {
> > > >  	case PERF_TYPE_HARDWARE:
> > > > @@ -243,6 +247,11 @@ static void arc_pmu_start(struct perf_event *event, int flags)
> > > >  
> > > >  	arc_pmu_event_set_period(event);
> > > >  
> > > > +	/* Enable interrupt for this counter */
> > > > +	if (is_sampling_event(event))
> > > > +		write_aux_reg(ARC_REG_PCT_INT_CTRL,
> > > > +			      read_aux_reg(ARC_REG_PCT_INT_CTRL) | (1 << idx));
> > > > +
> > > 
> > > Yet here you fail to actually enable the interrupt for the non sampling
> > > events, which makes the above not work.
> > 
> > Indeed we intentionally leave interrupts disabled for non-sampling events.
> >  [1] We have quite large counters so we don't expect to overflow normally
> >  [2] We may re-use the same code for hardware that lacks support of IRQs in PCT.
> >      See we check if IRQs are available and if not set PERF_PMU_CAP_NO_INTERRUPT
> >      that will guarantee we won't get sampling event and for non-sampling events
> >      we won't use IRQs.
> 
> Tricky, I was seeing is_isa_arcv2() calls elsewhere, so I figured you'd
> make it conditional on that.
> 
> But sure, if you think you can live with 1 this'll work.

Well indeed there's a room for improvement always.
But from our current experience existing implementation works pretty fine.

Moreover having now PCT IRQs we mostly use sampling events that allow doing
real profiling.

-Alexey