[Intel-gfx] [PATCH 14/17] drm/i915: Nonblocking request submission

[Joonas Lahtinen]

On ti, 2016-08-30 at 09:43 +0100, Chris Wilson wrote:
> On Tue, Aug 30, 2016 at 11:35:14AM +0300, Joonas Lahtinen wrote:
> > 
> > On su, 2016-08-28 at 21:46 +0100, Chris Wilson wrote:
> > > 
> > > @@ -442,6 +442,18 @@ i915_gem_request_alloc(struct intel_engine_cs *engine,
> > >  	 */
> > >  	req->head = req->ring->tail;
> > >  
> > > +	prev = i915_gem_active_peek(&engine->last_request,
> > > +				    &req->i915->drm.struct_mutex);
> > > +	if (prev) {
> > > +		ret = i915_sw_fence_await_sw_fence(&req->submit,
> > > +						   &prev->submit,
> > > +						   GFP_KERNEL);
> > > +		if (ret < 0) {
> > > +			i915_add_request(req);
> > As discussed in IRC, this should not be necessary at all. We're still
> > allocating the request, and the fence_await call failed (not setting up
> > any dependencies to our request implied) so nobody should know of our
> > request yet.
> At this point in the request alloc, we are already exposed (by the setup
> of the context), it's just if we do the await before we do the context
> setup then we are ok to fail. The problem being unwinding the fence if
> the context setup fails... Ugh, now I remember why I chose the
> ordering...

Hmm, true so, maybe drop a comment here?

Regards, Joonas

> -Chris
> 
-- 
Joonas Lahtinen
Open Source Technology Center
Intel Corporation