[Intel-gfx] [PATCH 5/9] drm/i915: also do frontbuffer tracking on pwrites

[Daniel Vetter]

On Mon, Feb 09, 2015 at 07:41:17PM +0100, Daniel Vetter wrote:
> On Mon, Feb 09, 2015 at 02:46:31PM -0200, Paulo Zanoni wrote:
> > From: Paulo Zanoni <paulo.r.zanoni at intel.com>
> > 
> > We need this for FBC, and possibly for PSR too.
> > 
> > Reviewed-by: Rodrigo Vivi <rodrigo.vivi at intel.com>
> > Signed-off-by: Paulo Zanoni <paulo.r.zanoni at intel.com>
> > ---
> >  drivers/gpu/drm/i915/i915_gem.c | 4 ++++
> >  1 file changed, 4 insertions(+)
> > 
> > diff --git a/drivers/gpu/drm/i915/i915_gem.c b/drivers/gpu/drm/i915/i915_gem.c
> > index 3d198f8..15910fa 100644
> > --- a/drivers/gpu/drm/i915/i915_gem.c
> > +++ b/drivers/gpu/drm/i915/i915_gem.c
> > @@ -1111,6 +1111,10 @@ i915_gem_pwrite_ioctl(struct drm_device *dev, void *data,
> >  			ret = i915_gem_phys_pwrite(obj, args, file);
> >  		else
> >  			ret = i915_gem_shmem_pwrite(dev, obj, args, file);
> > +
> > +		intel_fb_obj_flush(obj, false, ORIGIN_CPU);
> > +	} else {
> > +		intel_fb_obj_flush(obj, false, ORIGIN_GTT);
> 
> A flush alone does nothing. Well should, but you're kinda not quite using
> it correctly in the next patch to convert fbc over to frontbuffer
> tracking.
> 
> I guess the docs aren't perfect, so let me try again. There are two kinds
> of events the frontbuffer tracking code supplies to tell its consumers
> that screen changes are happening:
> - invalidate/flush: Invalidate denotes the start of the frontbuffer
>   rendering, from that point on psr/fbc/drrs must update the screen with
>   the usual refresh rate and not cache anything anywhere. When the flush
>   happens (which could easily be after a _very_ long time, e.g. fbcon)
>   then only can caching recomence. Caching = enable fbc, allow psr or
>   reduce refresh rate.
> - flip events: That's an instantenous event (well at least for consumer,
>   internally we need to track it as prepare/complete for async flips), and
>   mostly just interesting when the hw doesn't notice flips (some psr modes
>   and drrs).
> 
> So if you want to add frontbuffer tracking to pwrite then we need both an
> invalidate (before the actual pwrite) and a flush (after the pwrite, like
> you've added here).
> 
> The other issue is that there's a bug with the origin assignemnt:
> phys_pwrite also goes through the gtt. I think it would be best if we push
> the fb_obj_invalidate/flush into the relevant pwrite functions. That
> should make it easier to review, since the invalidate/flush will be next
> to the write op.

btw what's the use-case here? We don't upload stuff to X-tiled buffers
with pwrite, so this isn't really relevant for fbc I think.

It is a real gap for psr though, since cursor updates are done with
pwrite. But that probably gets papered over by X also updating the
position when the image changes, which means we'll get another (hw)
flip combo. But that's just X, and we indeed don't have a pwrite cursor
case yet in the psr testcase.
-Daniel
-- 
Daniel Vetter
Software Engineer, Intel Corporation
+41 (0) 79 365 57 48 - http://blog.ffwll.ch