[Intel-gfx] [PATCH] drm/i915: Fix frontbuffer false positve.

[Daniel Vetter]

On Tue, Feb 03, 2015 at 11:14:10AM -0800, Matt Roper wrote:
> On Tue, Feb 03, 2015 at 10:46:49AM -0800, Rodrigo Vivi wrote:
> > On Tue, Feb 3, 2015 at 8:21 AM, Matt Roper <matthew.d.roper at intel.com> wrote:
> > > On Tue, Feb 03, 2015 at 12:57:31PM +0100, Daniel Vetter wrote:
> > >> On Mon, Feb 02, 2015 at 03:38:16PM -0800, Rodrigo Vivi wrote:
> > >> > frontbuffer bits must be updated during commit times not on atomica prepare
> > >> > one, otherwise we have a risk of false positive.
> > >> >
> > >> > Cc Daniel Vetter <daniel.vetter at ffwll.ch>
> > >> > Cc: Sonika Jindal <sonika.jindal at intel.com>
> > >> > Signed-off-by: Rodrigo Vivi <rodrigo.vivi at intel.com>
> > >>
> > >> atomic.fb_bits isn't used at all right now, instead the
> > >> begin_crtc_commit function recomputes them. That looks wrong.
> > >
> > > We build up the collection of bits in atomic.fb_bits while going through
> > > the atomic pipeline for each plane, then do a single call to
> > >
> > >         intel_frontbuffer_flip(dev, intel_crtc->atomic.fb_bits);
> > >
> > > in intel_finish_crtc_commit to flush them all out together (and if
> > > check/prepare fail, we never actually get to that flush).
> > >
> > >> Also for async commits we need
> > >> to do the proper 2-stage flip stuff that current page_flip code does using
> > >> frontbuffer_flip_prepare/complete.
> > >
> > > I need to look at the frontbuffer stuff again...maybe we don't need
> > > atomic.fb_bits at all and should just use intel_frontbuffer_flip_prepare
> > > in the places we set the bits now and intel_frontbuffer_flip_complete
> > > where we're calling the flip mentioned above?
> > >
> > >
> > > Matt
> > >
> > >>
> > >> This patch here should have 0 effect (presuming I'm reading code
> > >> correctly), so what kinf of bug exactly are you seeing?
> > 
> > Yeah, after I sent the patch I was thinking about that: that it should
> > have 0 effect,
> > but the symptom that this apparently fixed here is that PSR wasn't
> > starting at all without
> > doing something like going to fbcon and come back to X.
> > Something similar what Sonika had told on that psr-skl thread where
> > she couldn't get psr working on login screen.
> > After this patch I didn't' have to change back and forth to fbcon to
> > make psr work.
> > 
> > Maybe just a coincidence, but anyway I believe the way it is nowadays
> > it is wrong. I believe frontbuffer bits and calls should be done at
> > commit step, not under prepare.
> 
> I may be misunderstanding what you're saying, but I think this is the
> way things already work?  We plan out which bits we're going to update
> in the 'check' step (and record those bits in atomic.fb_bits), but then
> we only actually set them with intel_frontbuffer_flip() in commit if
> everything looks good.  If we wind up rejecting the update because we
> fail to check/prepare, those bits in atomic.fb_bits should just get
> thrown away, unless I'm missing something.

What we could do is store the fb_bits in intel_plane and then move all the
code to handle them into shared code, out of the cursor/primary/sprite
specific stuff. That helps with new hw, in case we get 2 sprite planes and
so can't use the drm type any more to figure out which one it is.
-Daniel
-- 
Daniel Vetter
Software Engineer, Intel Corporation
+41 (0) 79 365 57 48 - http://blog.ffwll.ch