[Intel-gfx] [PATCH 1/8] drm/i915: Serialise concurrent calls to i915_gem_set_wedged()

[Mika Kuoppala]

Chris Wilson <chris at chris-wilson.co.uk> writes:

> Quoting Chris Wilson (2019-01-15 12:05:27)
>> Quoting Mika Kuoppala (2019-01-15 11:56:11)
>> > Chris Wilson <chris at chris-wilson.co.uk> writes:
>> > 
>> > > Make i915_gem_set_wedged() and i915_gem_unset_wedged() behaviour more
>> > > consistently if called concurrently.
>> > 
>> > More is needed in here. The purpose is to make them wait in turns
>> > on top of mutex, instead of racing on the bit? Where is
>> > the inconsistency tho.
>> 
>> We report set-wedged multiple times on failure paths. Worse is when we
>> report set-wedged multiple times simultaneously.
>
> I've been contemplating just moving the reporting inside the test-bit
> serialisation, but I kept resisting. This issue has been nagging at me
> ever since using the bit for loose serialisation; you either fix a race
> or live to regret it.

That part I can agree with. Now it is all contained so if it coalesces
on reports, it will be alarming :)

Reviewed-by: Mika Kuoppala <mika.kuoppala at linux.intel.com>