Quoting Chris Wilson (2018-02-07 15:13:50)
> Reduce the window of opportunity for set-wedged being called
> concurrently with reset (after i915_reset() has performed the
> i915_gem_unset_wedged()) by moving the set_bit(I915_WEDGED) to before we
> complete the inflight requests. When i915_reset() is being blocked on a
> request, such completion may allow it to start and beginning resetting
> the GPU before i915_gem_set_wedged() has finished (and so before
> set-wedge will have marked the device as wedged). As such,
> i915_gem_init_hw() may see a wedged device even from inside
> i915_reset().
So I'm 99% certain this is the problem on blb/pnv. As we break the
modeset deadlock using set-wedged from a timer, the reset springs into
action on the other cpu and races with set_bit(I915_WEDGED). Flaging
I915_WEDGED first will force i915_reset() to serialise via
i915_gem_unset_wedged(). (Well that's the plan at least.)
> References: 36703e79a982 ("drm/i915: Break modeset deadlocks on reset")
> Signed-off-by: Chris Wilson <chris@xxxxxxxxxxxxxxxxxx>
> Cc: Mika Kuoppala <mika.kuoppala@xxxxxxxxxxxxxxx>
> Cc: Joonas Lahtinen <joonas.lahtinen@xxxxxxxxxxxxxxx>
> Cc: Tvrtko Ursulin <tvrtko.ursulin@xxxxxxxxx>
> ---
> drivers/gpu/drm/i915/i915_gem.c | 7 +++++--
> 1 file changed, 5 insertions(+), 2 deletions(-)
>
> diff --git a/drivers/gpu/drm/i915/i915_gem.c b/drivers/gpu/drm/i915/i915_gem.c
> index c1b80cd52f9e..06f0456699af 100644
> --- a/drivers/gpu/drm/i915/i915_gem.c
> +++ b/drivers/gpu/drm/i915/i915_gem.c
> @@ -3205,6 +3205,9 @@ void i915_gem_set_wedged(struct drm_i915_private *i915)
> intel_engine_dump(engine, &p, "%s\n", engine->name);
> }
>
> + set_bit(I915_WEDGED, &i915->gpu_error.flags);
> + smp_mb__after_atomic();
> +
> /*
> * First, stop submission to hw, but do not yet complete requests by
> * rolling the global seqno forward (since this would complete requests
> @@ -3241,7 +3244,8 @@ void i915_gem_set_wedged(struct drm_i915_private *i915)
> for_each_engine(engine, i915, id) {
> unsigned long flags;
>
> - /* Mark all pending requests as complete so that any concurrent
> + /*
> + * Mark all pending requests as complete so that any concurrent
> * (lockless) lookup doesn't try and wait upon the request as we
> * reset it.
> */
> @@ -3251,7 +3255,6 @@ void i915_gem_set_wedged(struct drm_i915_private *i915)
> spin_unlock_irqrestore(&engine->timeline->lock, flags);
> }
>
> - set_bit(I915_WEDGED, &i915->gpu_error.flags);
> wake_up_all(&i915->gpu_error.reset_queue);
> }
>
> --
> 2.16.1
>
_______________________________________________
Intel-gfx mailing list
Intel-gfx@xxxxxxxxxxxxxxxxxxxxx
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
