[Intel-gfx] [PATCH 07/27] drm/i915: Squash repeated awaits on the same fence

[Chris Wilson]

On Wed, Apr 26, 2017 at 03:36:19PM +0100, Tvrtko Ursulin wrote:
> 
> On 26/04/2017 13:23, Chris Wilson wrote:
> >Too early, it's the timeline (and syncs along it) that's interesting.
> >For our contexts, we can hook into context-close, but we still have some
> >foreign dma-fence-contexts to worry about. I was thinking of walking all
> >timelines from the idle_worker. And possibly forcibly prune across
> >suspend.
> 
> Hm I don't see why it is too early. If request is getting freed,
> there is no one waiting on it any longer, so how can it be OK to
> keep that seqno in the map?

The fence->seqno represents a known synchronisation point from our timeline to the
fence->timeline. In the future, we know that we can skip all
synchronisations onto that timeline that are older than the fence->seqno
because we already have synchronised. That coupling persists after the
fence itself is destroyed. It is why tracking it between timelines is
more effective than just squashing repeats within a request.

(This is perhaps more significant when we expend ring space to emit GPU
commands for each sync point, i.e. hw semaphores.)
-Chris

-- 
Chris Wilson, Intel Open Source Technology Centre