Re: [PATCH 2/4] perf/x86/intel: fix event update for auto-reload

[Liang, Kan]

On 12/19/2017 3:08 PM, Liang, Kan wrote:
> On 12/19/2017 1:58 PM, Peter Zijlstra wrote:
> > On Mon, Dec 18, 2017 at 03:34:49AM -0800, kan.liang@xxxxxxxxxxxxxxx 
> > wrote:
> > > Â arch/x86/events/core.cÂÂÂÂ | 14 ++++++++++++++
> > > Â arch/x86/events/intel/ds.c |Â 8 +++++++-
> > > Â 2 files changed, 21 insertions(+), 1 deletion(-)
> > >
> > > diff --git a/arch/x86/events/core.c b/arch/x86/events/core.c
> > > index 35552ea..f74e21d 100644
> > > --- a/arch/x86/events/core.c
> > > +++ b/arch/x86/events/core.c
> > > @@ -100,6 +100,20 @@ u64 x86_perf_event_update(struct perf_event *event,
> > > ÂÂÂÂÂÂ * of the count.
> > > ÂÂÂÂÂÂ */
> > > ÂÂÂÂÂ delta = (new_raw_count << shift) - (prev_raw_count << shift);
> > > +
> > > +ÂÂÂ /*
> > > +ÂÂÂÂ * Take auto-reload into account
> > > +ÂÂÂÂ * For the auto-reload before the last time, it went through the
> > > +ÂÂÂÂ * whole period (reload_val) every time.
> > > +ÂÂÂÂ * Just simply add period * times to the event.
> > > +ÂÂÂÂ *
> > > +ÂÂÂÂ * For the last load, the elapsed delta (event-)time need to be
> > > +ÂÂÂÂ * corrected by adding the period. Because the start point is 
> > > -period.
> > > +ÂÂÂÂ */
> > > +ÂÂÂ if (reload_times > 0) {
> > > +ÂÂÂÂÂÂÂ delta += (reload_val << shift);
> > > +ÂÂÂÂÂÂÂ local64_add(reload_val * (reload_times - 1), &event->count);
> > > +ÂÂÂ }
> > > ÂÂÂÂÂ delta >>= shift;
> > > ÂÂÂÂÂ local64_add(delta, &event->count);
> > > diff --git a/arch/x86/events/intel/ds.c b/arch/x86/events/intel/ds.c
> > > index 0b693b7..f0f6026 100644
> > > --- a/arch/x86/events/intel/ds.c
> > > +++ b/arch/x86/events/intel/ds.c
> > > @@ -1256,11 +1256,17 @@ static void __intel_pmu_pebs_event(struct 
> > > perf_event *event,
> > > ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ void *base, void *top,
> > > ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ int bit, int count)
> > > Â {
> > > +ÂÂÂ struct hw_perf_event *hwc = &event->hw;
> > > ÂÂÂÂÂ struct perf_sample_data data;
> > > ÂÂÂÂÂ struct pt_regs regs;
> > > ÂÂÂÂÂ void *at = get_next_pebs_record_by_bit(base, top, bit);
> > > -ÂÂÂ if (!intel_pmu_save_and_restart(event, 0, 0) &&
> > > +ÂÂÂ /*
> > > +ÂÂÂÂ * Now, auto-reload is only enabled in fixed period mode.
> > > +ÂÂÂÂ * The reload value is always hwc->sample_period.
> > > +ÂÂÂÂ * May need to change it, if auto-reload is enabled in freq mode 
> > > later.
> > > +ÂÂÂÂ */
> > > +ÂÂÂ if (!intel_pmu_save_and_restart(event, hwc->sample_period, count 
> > > - 1) &&
> > > ÂÂÂÂÂÂÂÂÂ !(event->hw.flags & PERF_X86_EVENT_AUTO_RELOAD))
> > > ÂÂÂÂÂÂÂÂÂ return;
> >
> > This all looks very wrong... In auto reload we should never call
> > intel_pmu_save_and_restore() in the first place I think.
> >
> > Things like x86_perf_event_update() and x86_perf_event_set_period()
> > simply _cannot_ do the right thing when we auto reload the counter.
>
> I think it should be OK to call it in first place.
> For x86_perf_event_update(), the reload_times will tell if it's auto 
> reload. Both period_left and event->count are carefully recalculated for 
> auto reload.

Think a bit more. You are right. We cannot rely on count to tell us if 
it's auto reload.
The count could also be 1 if auto reload is enabled.

I will fix it in V2.

Thanks,
Kan

> For x86_perf_event_set_period(), there is nothing special needed for 
> auto reload. The period is fixed. The period_left from 
> x86_perf_event_update() is already handled.
>
>
> BTW: It should be 'count' not 'count - 1' which pass to
> intel_pmu_save_and_restart(). I just found the issue. I will fix it in
> V2 with other improvements if there are any.
>
> Thanks,
> Kan