[Intel-gfx] [RFC] drm/i915: Infrastructure for supporting different GGTT views per object

[Chris Wilson]

On Mon, Nov 24, 2014 at 12:32:12PM +0000, Tvrtko Ursulin wrote:
> 
> On 11/12/2014 05:02 PM, Daniel Vetter wrote:
> >The change around finish_gtt is actually a leak since if the last view
> >around is the rotate one (which can happen if userspace drops the buffer
> >but leaves it displayed) we won't clean up (and so leak) the iommu
> >mapping.
> 
> So I somehow need to ensure finish_gtt on a "normal" view is done
> last, correct? Move it from VMA destructon to object destruction?
> Would this have any implications for the shrinker?

Just have a obj->dma_pin_count?
-Chris

-- 
Chris Wilson, Intel Open Source Technology Centre