[Intel-gfx] [PATCH] drm/i915: Assert that the request is indeed complete when signaled from irq

[Chris Wilson]

Quoting Tvrtko Ursulin (2018-03-05 11:12:45)
> 
> On 05/03/2018 10:41, Chris Wilson wrote:
> > After we call dma_fence_signal(), confirm that the request was indeed
> > complete.
> > 
> > Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk>
> > Cc: Tvrtko Ursulin <tvrtko.ursulin at intel.com>
> > ---
> >   drivers/gpu/drm/i915/i915_irq.c | 1 +
> >   1 file changed, 1 insertion(+)
> > 
> > diff --git a/drivers/gpu/drm/i915/i915_irq.c b/drivers/gpu/drm/i915/i915_irq.c
> > index ce16003ef048..633c18785c1e 100644
> > --- a/drivers/gpu/drm/i915/i915_irq.c
> > +++ b/drivers/gpu/drm/i915/i915_irq.c
> > @@ -1123,6 +1123,7 @@ static void notify_ring(struct intel_engine_cs *engine)
> >   
> >       if (rq) {
> >               dma_fence_signal(&rq->fence);
> > +             GEM_BUG_ON(!i915_request_completed(rq));
> >               i915_request_put(rq);
> >       }
> >   
> > 
> 
> What's the motivation? There is a i915_seqno_passed check a few lines 

The seqno check is on wait.seqno, this is to confirm it all ties
together with the request and our preemption avoidance is solid. The
motivation was the bug in the signaler along the same lines.

> above. So there would have to be a confusion in internal breadcrumbs 
> state for this to be possible. In which case I'd rather put the assert 
> in breadcrumbs code. For instance in intel_wait_check_request, asserting 
> that the seqno in wait matches the seqno in wait->request.

The entire point of that check is to say when they don't match so that
we know when the request was unsubmitted during the wait.
-Chris