[PATCH driver/intel] Allow copy_front() to fail and clean up gracefully if it does

[Chris Wilson]

On Thu, Sep 25, 2014 at 04:10:26PM +0100, Chris Wilson wrote:
> On Thu, Sep 25, 2014 at 03:16:23PM +0200, Egbert Eich wrote:
> > Chris Wilson writes:
> >  > On Wed, Sep 24, 2014 at 01:23:26PM +0200, Egbert Eich wrote:
> >  > > Chris Wilson writes:
> >  > >  > On Wed, Sep 24, 2014 at 12:34:29PM +0200, Egbert Eich wrote:
> >  > >  > > 
> >  > >  > > For panning one needs to be able to draw into the screen pixmap outside 
> >  > >  > > of the area exposed by the crtc.
> >  > >  > 
> >  > >  > Note that the screen pixmap still exists and is drawn into by the
> >  > >  > clients. What should happen with the panning is that the CRTC is updated
> >  > >  > through set_mode_major which invalidates the new CRTC location in the big
> >  > >  > screen Pixmap and the contents copied during the next shadow update.
> >  > >  > 
> >  > >  > That's the theory at least...
> >  > > 
> >  > > Ok, so this isn't working. I will look into this as it's easy for me
> >  > > to trigger this situation.
> >  > 
> >  > Are you using an OpenGL compositor? So far bare X, awesome, dwm,
> >  > enlightenment and kwin/xrender behave correctly. But gnome-shell and
> >  > kwin/gl are broken, which is not an issue with the ddx.
> >  > -Chris
> > 
> > With that patch from me which you accepted, yes, but even without it, 
> > a fallback should be used but this doesn't work.
> > I found this:
> > In sna_pixmap_move_to_gpu() you move a CPU bo to the GPU if available
> > and feasable. Shouldn't you check for the proper pitch alignment as well?
> 
> Oh, it fails to test for scanout bo, right. That's a nasty little corner
> case that only befalls a linear scanout.
> 
> > Without it, use_shadow() fails in sna_crtc_attach(), thus fb_get() is
> > called for a bo that was previously created for the cpu which may not
> > be what is expected by the GPU.
> > Quickly hacking in a test for 64bit alignment (which isn't correct for
> > all use cases) fixes the panning issue for me.
> 
> Good to know, thanks

Also realised that even this should be fixed up as a last resort by
falling back to the shadow CRTC allocation. That band-aid should be
working again.

Hopefully

commit 9f7c1a4c4f2a6352263c36e75a984ed4095adbc0
Author: Chris Wilson <chris at chris-wilson.co.uk>
Date:   Thu Sep 25 16:29:14 2014 +0100

    sna: Check for scanout pitch restrictions on linear GPU bo
    
    When converting a linear cached CPU bo into an uncached GPU bo, we must
    be careful to adhere to the scanout restrictions if they apply for this
    transfer or this Pixmap.
    
    Reported-by: Egbert Eich <eich at suse.com>
    Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk>

catches all the cases where we need to check the alignment on the pitch
before changing the cache level. If not, it now explicitly converts the
bo before making the framebuffer.
-Chris

-- 
Chris Wilson, Intel Open Source Technology Centre