[Intel-gfx] [PATCH 1/2] drm/i915: Reduce recursive mutex locking from the shrinker

[Chris Wilson]

Quoting Tvrtko Ursulin (2019-01-09 14:12:46)
> From: Tvrtko Ursulin <tvrtko.ursulin at intel.com>
> 
> In two codepaths internal to the shrinker we know we will end up taking
> the resursive mutex path.
> 
> It instead feels more elegant to avoid this altogether and not call
> mutex_trylock_recursive in those cases.
> 
> We achieve this by extracting the shrinking part to __i915_gem_shrink,
> wrapped by struct mutex taking i915_gem_shrink.
> 
> At the same time move the runtime pm reference taking to always be in the
> usual, before struct mutex, order.
> 
> v2:
>  * Don't use flags but split i915_gem_shrink into locked and unlocked.

Yes, I do prefer this to passing flags down, but I feel very
underwhelmed by it. The fact that the recursive lock exists is the
problem and this isn't reducing the problem (imo). I'd rather push the
only call to shrinker_lock down to i915_gem_shrink and remove it from
shrinker_scan and shrinker_vmap which should be doable in a few
patches -- that would get rid of the self-inflicted recursion but still
recursing from our callers.
-Chris