[Intel-gfx] [PATCH] drm/i915: Quietly reject attempts to create non-pagealigned stolen objects

[Chris Wilson]

On Wed, Dec 10, 2014 at 11:23:44AM +0100, Daniel Vetter wrote:
> On Wed, Dec 10, 2014 at 08:17:11AM +0000, Chris Wilson wrote:
> > This added as a BUG_ON as it considered that no one would ever request
> > an unaligned object. However, it turns out that some BIOSes will
> > allocate a scanout that is offset from 0 and not aligned to a page
> > boundary, and we were passing this through and hitting the BUG_ON during
> > boot.
> > 
> > Quietly reject such a request to reserve the unaligned stolen object and
> > let the boot continue, restoring previous behaviour (i.e. no BIOS
> > framebuffer preservation).
> > 
> > Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=86883
> > Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk>
> > Cc: stable at vger.kernel.org
> > ---
> >  drivers/gpu/drm/i915/i915_gem_stolen.c | 10 ++++++----
> >  1 file changed, 6 insertions(+), 4 deletions(-)
> > 
> > diff --git a/drivers/gpu/drm/i915/i915_gem_stolen.c b/drivers/gpu/drm/i915/i915_gem_stolen.c
> > index 5c616ec2c5c8..a3bc0fa07c6c 100644
> > --- a/drivers/gpu/drm/i915/i915_gem_stolen.c
> > +++ b/drivers/gpu/drm/i915/i915_gem_stolen.c
> > @@ -646,13 +646,15 @@ i915_gem_object_create_stolen_for_preallocated(struct drm_device *dev,
> >  	DRM_DEBUG_KMS("creating preallocated stolen object: stolen_offset=%x, gtt_offset=%x, size=%x\n",
> >  			stolen_offset, gtt_offset, size);
> >  
> > -	/* KISS and expect everything to be page-aligned */
> > -	BUG_ON(stolen_offset & 4095);
> > -	BUG_ON(size & 4095);
> > -
> >  	if (WARN_ON(size == 0))
> >  		return NULL;
> >  
> > +	/* KISS and expect everything to be GTT page-aligned */
> > +	if ((stolen_offset | size) & 4095) {
> 
> Imo we should stil WARN_ON and fixup up the takeover code to align things
> properly ...

You shot down my idea for storing deltas into objects in the past...

The BIOS scanout is properly aligned to the rules of the display engine,
just not according to our mm restrictions. The bigger question is
whether our 1:1 offset-to-stolen mapping is correct. It could well be
that that the framebuffer is at stolen address 0, but just has a GTT
offset.

So the only question is whether we reject the object reservation at the
stolen layer or at the plane config layer. I decided that stolen was
better, because it is failing to meet our mm restrictions not
hardware restrictions.
-Chris

-- 
Chris Wilson, Intel Open Source Technology Centre