[Intel-gfx] [PATCH 46/64] drm/i915: Count how many VMA are bound for an object

[Chris Wilson]

On Tue, Jul 12, 2016 at 04:12:43PM +0100, Tvrtko Ursulin wrote:
> 
> On 12/07/16 15:38, Chris Wilson wrote:
> >On Tue, Jul 12, 2016 at 03:30:25PM +0100, Tvrtko Ursulin wrote:
> >>On 07/07/16 09:41, Chris Wilson wrote:
> >>>@@ -3684,6 +3684,9 @@ i915_gem_object_pin_to_display_plane(struct drm_i915_gem_object *obj,
> >>>  					    old_read_domains,
> >>>  					    old_write_domain);
> >>>
> >>>+	/* Increment the pages_pin_count to guard against the shrinker */
> >>>+	obj->pages_pin_count++;
> >>
> >>Would it be clearer to look at obj->pin_display in the shrinker?
> >>Although it looks like special casing out of the cleanliness of the
> >>design in both case so I am not sure.
> >
> >Yeah. I liked the mechanism of telling the shrinker to avoid certain
> >pages by only having to control the pages_pin_count. It feels easier to
> >explain to others "the shrinker may reap any object that hasn't pinned
> >its pages" (explicitly called i915_gem_object_pin_pages for its own use)
> >rather than that + plus a list of exceptions known by the shrinker.
> 
> Hm but wait a minute, framebuffer already has the VMA pinned. So how
> will the shrinker reap it?

It won't, but it would try. Worse it would include the object in its
estimates for shrinkable pages.
-Chris

-- 
Chris Wilson, Intel Open Source Technology Centre