[Intel-gfx] [igt-dev] [PATCH i-g-t] i915/perf_pmu: Idle before measuring idle

[Chris Wilson]

Quoting Tvrtko Ursulin (2019-04-01 09:15:18)
> 
> On 31/03/2019 22:09, Chris Wilson wrote:
> > Before trying to measure the busyness reporting for the idle state, make
> > sure the gpu is actually idle and not still running any backgound system
> > task.
> > 
> > Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk>
> > Cc: Tvrtko Ursulin <tvrtko.ursulin at intel.com>
> > ---
> >   tests/perf_pmu.c | 1 +
> >   1 file changed, 1 insertion(+)
> > 
> > diff --git a/tests/perf_pmu.c b/tests/perf_pmu.c
> > index 1a08f564b..ef61c4d40 100644
> > --- a/tests/perf_pmu.c
> > +++ b/tests/perf_pmu.c
> > @@ -264,6 +264,7 @@ single(int gem_fd, const struct intel_execution_engine2 *e, unsigned int flags)
> >       uint64_t val;
> >       int fd;
> >   
> > +     gem_quiescent_gpu(gem_fd);
> >       fd = open_pmu(I915_PMU_ENGINE_BUSY(e->class, e->instance));
> >   
> >       if (flags & TEST_BUSY)
> > 
> 
> It's okay but where did it fail? Because subtest has quiescent at the 
> end, and IGT has it at start.

Honestly, not sure. I suspect it was just another idle worker. Other
than the selftest who's mission it is to detect such errors, this might
be the only test that is sensitive to stray requests being run in the
background. So it might be just shooting the canary.
-Chris