Quoting Tvrtko Ursulin (2018-02-05 13:37:42) > > On 05/02/2018 13:18, Chris Wilson wrote: > > Quoting Tvrtko Ursulin (2018-02-05 11:45:17) > >> > >> On 05/02/2018 10:31, Patchwork wrote: > >>> == Series Details == > >>> > >>> Series: RFC drm/i915/pmu: Avoid sleeping rpm_get under atomic PMU read > >>> URL : https://patchwork.freedesktop.org/series/37644/ > >>> State : success > >>> > >>> == Summary == > >>> > >>> Test perf_pmu: > >>> Subgroup busy-double-start-vcs0: > >>> pass -> INCOMPLETE (shard-apl) fdo#103927 > >> > >> This is most probably the race "drm/i915/pmu: Fix PMU enable vs > >> execlists tasklet race" will fix once merged. > > > > I was hoping to see some pmu-rc6 oddity to write off the patch. :| > > Albeit not ideal some fix is needed. > > Maybe move to timer sampling and assume RC6 grows by PERIOD while device > is suspended, then when awake, make sure we don't add the real RC6 > counter until it catches up with the faked value.
Yeah, I think we might go with something like that. It also takes care of the nasty wraparound issue in the counters that we don't handle yet. If rc6 is not available, neither is rpm. We cannot sleep unless rc6 provides the power gating for the GPU, so I think it's a safe assumption that if !gt.awake, rc6 is active. > So to ensure no > backwards jumps? It should catch up eventually, unless there is no RC6 > at all. Hm, can we detect that in the very beginning? HAS_RC6 is updated during init based on whether we can enable rc6. That doesn't tell us if it's completely broken, just that we are trying. -Chris _______________________________________________ Intel-gfx mailing list Intel-gfx@lists.freedesktop.org https://lists.freedesktop.org/mailman/listinfo/intel-gfx
