[Intel-gfx] [RFC 1/2] drm/i915: Move ioremap_wc tracking onto VMA

[Chris Wilson]

On Wed, Apr 13, 2016 at 10:14:50AM +0100, Tvrtko Ursulin wrote:
> On 13/04/16 10:04, Chris Wilson wrote:
> >On Wed, Apr 13, 2016 at 09:45:03AM +0100, Tvrtko Ursulin wrote:
> >>I could have two separate series to simplify dependencies a bit:
> >>
> >>  1. GuC premature unpin and
> >>  2. execlist no retired queue.
> >>
> >>The stack for the first one could look like (not as patches but
> >>conceptually):
> >>
> >>  1. default context cleanup
> >>  2. any io_mapping patches of yours
> >>  3. i915_vma_iomap or WC pin_map as you suggested in the other reply.
> >>  4. previous / pinned context
> >>
> >>Execlist no retired queue would then be:
> >>
> >>  1. stable ctx id
> >>  2. removal of i915_gem_request_unreference__unlocked
> >>  3. execlist no retired queue
> >>
> >>If I did not forget about something.
> >>
> >>At any point in any series we could add things like
> >>i915_gem_request.c or those patches which move around the context
> >>init.
> >
> >Could I see you some drm_mm.c patches after that?
> 
> If I am competent enough to help with them sure.

There's a nice tweak in there to help with the aperture thrashing
scenario (speeds up the aperture thrash tests in gem_concurrent_blit by
~60x). You mentioned you were interested in similar scenarios.

Besides which they actually fix the u64 alignment issues and some
residual u32 problems.
 
> Does this mean above sounds like a plan to you? Two series
> containing the listed items?

3. I'll get io_mapping changes ratified, and that can happen in parallel
to these two series.
-Chris
>
-- 
Chris Wilson, Intel Open Source Technology Centre