Re: [PATCH 4/9] perf/x86/intel: Support hardware TopDown metrics

From: Peter Zijlstra
Date: Wed May 29 2019 - 03:57:58 EST


On Tue, May 28, 2019 at 02:24:38PM -0400, Liang, Kan wrote:
>
>
> On 5/28/2019 9:43 AM, Peter Zijlstra wrote:
> > On Tue, May 21, 2019 at 02:40:50PM -0700, kan.liang@xxxxxxxxxxxxxxx wrote:
> > > @@ -3287,6 +3304,13 @@ static int core_pmu_hw_config(struct perf_event *event)
> > > return intel_pmu_bts_config(event);
> > > }
> > > +#define EVENT_CODE(e) (e->attr.config & INTEL_ARCH_EVENT_MASK)
> > > +#define is_slots_event(e) (EVENT_CODE(e) == 0x0400)
> > > +#define is_perf_metrics_event(e) \
> > > + (((EVENT_CODE(e) & 0xff) == 0xff) && \
> > > + (EVENT_CODE(e) >= 0x01ff) && \
> > > + (EVENT_CODE(e) <= 0x04ff))
> > > +
> >
> > That is horrific.. (e & INTEL_ARCH_EVENT_MASK) & 0xff is just daft,
> > that should be: (e & ARCH_PERFMON_EVENTSEL_EVENT).
> >
> > Also, we already have fake events for event=0, see FIXED2, why are we
> > now using event=0xff ?
>
> I think event=0 is for genuine fixed events. Metrics events are fake events.
> I didn't find FIXED2 in the code. Could you please share more hints?

cd09c0c40a97 ("perf events: Enable raw event support for Intel unhalted_reference_cycles event")

We used the fake event=0x00, umask=0x03 for CPU_CLK_UNHALTED.REF_TSC,
because that was not available as a generic event, *until now* it seems.
I see ICL actually has it as a generic event, which means we need to fix
up the constraint mask for that differently.

But note that for all previous uarchs this event did not in fact exist.

It appears the TOPDOWN.SLOTS thing, which is available in in FIXED3 is
event=0x00, umask=0x04, is indeed a generic event too.