[Intel-gfx] [PATCH igt] igt/perf_pmu: Combine rc6/rc6p tests

[Chris Wilson]

Quoting Tvrtko Ursulin (2017-11-24 10:01:35)
> 
> On 24/11/2017 09:24, Chris Wilson wrote:
> > Quoting Chris Wilson (2017-11-24 09:23:22)
> >> How? How do you tell if rc6 is not incrementing because (a) the hw is
> >> busy, (b) the hw is idle but in a different rc6 state, or (c) the pmu is
> >> broken.
> > 
> > Case in point is rc6p on shard-snb, where rc6p is disabled but the
> > pmu event still exists. It disabled by the kernel itself, but since the
> > pmu exists we expect it to increment.
> 
> You mean I should be checking for more than HAS_RC6p when allowing an 
> event to be opened?

No worries, it's going to become static and always correct very shortly.
And HAS_RC6pp, which breaks your test :|

> If I can fix that, then I was thinking the rc6p subtest would just need 
> to be updated to only open the RC6p counter, and rc6 subtest would only 
> open the RC6 one.
> 
> Or RC6 not incrementing while in RC6p is only true on some platforms?

On my snb, the rc6 test fails because whilst idle, rc6_residency doesn't
increase, only rc6p_residency.

On the farm's snb, it is reversed.

The moral of the story is that the rc6 modes are currently exclusive;
and you cannot dictate which the engine will drop into.
-Chris