[Intel-gfx] [RFC] drm/i915: Simplify waiting for registers

[Chris Wilson]

On Wed, Apr 12, 2017 at 11:42:55AM +0100, Tvrtko Ursulin wrote:
> 
> On 12/04/2017 11:36, Tvrtko Ursulin wrote:
> >From: Tvrtko Ursulin <tvrtko.ursulin at intel.com>
> >
> >I was thinking if we could get away with simplifying the API
> >a bit by getting rid of the _fw variant and only have these
> >three functions with a common implementation:
> >
> >  intel_wait_for_register
> >  intel_wait_for_register_atomic
> >  __intel_wait_for_register
> >
> >The fast/busy loop in all cases grabs it's own forcewake and
> >is done under the uncore lock. The extra overhead for call
> >sites which already have the forcewake, or do not need it is
> >there, but not sure that it matters for where wait_for_register
> >functions are used.
> 
> This is probably quite bad for pcode, since AFAIR those can be quite
> slow. So scratch this idea I think..

There's definitely some merit here. I'd wait until Ville presents his
gt/de split since that's going to be quite a major paradigm shift for us
and then re-evaluate.
-Chris

-- 
Chris Wilson, Intel Open Source Technology Centre