[Intel-gfx] [PATCH 2/6] drm/i915: Check current i915_vma.pin_count status first on unbind

[Chris Wilson]

Quoting Chris Wilson (2020-01-27 14:20:21)
> Quoting Tvrtko Ursulin (2020-01-27 14:15:57)
> > 
> > On 26/01/2020 10:23, Chris Wilson wrote:
> > > Do an early rejection of a i915_vma_unbind() attempt if the i915_vma is
> > > currently pinned, without waiting to see if the inflight operations may
> > > unpin it. We see this problem with the shrinker trying to unbind the
> > > active vma from inside its bind worker:
> > 
> > What is the actual problem? flush_work from a worker?
> 
> We hit the shrinker from the inside the worker (which is intended). But
> what is not intended is for the shrinker to wait on any of the workers,
> because of the potential recursion of waiting on one's self. The
> intention was that we avoid the shrinker waiting on the vma by keeping
> the vma pinned... But I only pinned the pages to prevent them being
> freed and not the worker to prevent the recursive wait.
> 
> Heh.

Which then ties back into the whole i915_vma is not krefed, so to
complete the destroy at the moment we need to sync with the workers.
-Chris