[PATCH] gpu: drm: i915: display: Avoid null values intel_plane_atomic_check_with_state

[Ville Syrjälä]

On Fri, Sep 27, 2024 at 04:45:44PM +0300, Ville Syrjälä wrote:
> On Fri, Sep 27, 2024 at 04:14:17PM +0300, Jani Nikula wrote:
> > On Fri, 27 Sep 2024, Ville Syrjälä <ville.syrjala at linux.intel.com> wrote:
> > > On Fri, Sep 27, 2024 at 11:20:32AM +0300, Jani Nikula wrote:
> > >> On Fri, 27 Sep 2024, Alessandro Zanni <alessandro.zanni87 at gmail.com> wrote:
> > >> > This fix solves multiple Smatch errors:
> > >> >
> > >> > drivers/gpu/drm/i915/display/intel_atomic_plane.c:660
> > >> > intel_plane_atomic_check_with_state() error:
> > >> > we previously assumed 'fb' could be null (see line 648)
> > >> >
> > >> > drivers/gpu/drm/i915/display/intel_atomic_plane.c:664
> > >> > intel_plane_atomic_check_with_state()
> > >> > error: we previously assumed 'fb' could be null (see line 659)
> > >> >
> > >> > drivers/gpu/drm/i915/display/intel_atomic_plane.c:671
> > >> > intel_plane_atomic_check_with_state()
> > >> > error: we previously assumed 'fb' could be null (see line 663)
> > >> >
> > >> > We should check first if fb is not null before to access its properties.
> > >> 
> > >> new_plane_state->uapi.visible && !fb should not be possible, but it's
> > >> probably too hard for smatch to figure out. It's not exactly trivial for
> > >> humans to figure out either.
> > >> 
> > >> I'm thinking something like below to help both.
> > >> 
> > >> Ville, thoughts?
> > >> 
> > >> 
> > >> BR,
> > >> Jani.
> > >> 
> > >> 
> > >> diff --git a/drivers/gpu/drm/i915/display/intel_atomic_plane.c b/drivers/gpu/drm/i915/display/intel_atomic_plane.c
> > >> index 3505a5b52eb9..d9da47aed55d 100644
> > >> --- a/drivers/gpu/drm/i915/display/intel_atomic_plane.c
> > >> +++ b/drivers/gpu/drm/i915/display/intel_atomic_plane.c
> > >> @@ -629,6 +629,9 @@ int intel_plane_atomic_check_with_state(const struct intel_crtc_state *old_crtc_
> > >>  	if (ret)
> > >>  		return ret;
> > >>  
> > >> +	if (drm_WARN_ON(display->drm, new_plane_state->uapi.visible && !fb))
> > >> +		return -EINVAL;
> > >> +
> > >
> > > We have probably 100 places that would need this. So it's going
> > > to be extremely ugly.
> > >
> > > One approach I could maybe tolerate is something like
> > > intel_plane_is_visible(plane_state) 
> > > {
> > > 	if (drm_WARN_ON(visible && !fb))
> > > 		return false;
> > >
> > > 	return plane_state->visible;
> > > }
> > >
> > > + s/plane_state->visible/intel_plane_is_visible(plane_state)/
> > >
> > > But is that going to help these obtuse tools?
> > 
> > That does help people, which is more important. :)
> > 
> > I think the problem is first checking if fb is NULL, and then
> > dereferencing it anyway.
> > 
> > visible always means fb != NULL, but I forget, is the reverse true? Can
> > we have fb != NULL and !visible? I mean could we change the fb check to
> > visible check?
> 
> No, the reverse does not hold. A plane can be invisible
> while still having a valid fb. Eg. the plane could be
> positioned completely offscreen, or the entire crtc may
> be inactive (DPMS off).
> 
> And whenever we have an fb we want to do all the check to make sure
> it satisfies all the requirements, whether the plane is visible or
> not. Otherwise we could end up confusing userspace with something
> like this:
> 
> 1. Usespace assigns some unsupported fb to the plane
>    but positions the plane offscreen -> success
> 2. Userspace moves the plane to somewhere onscreen -> fail


Basically planes should have three different "enabled" states:

logically_enabled: fb!=NULL (also the crtc must be logically enabled,
                             but drm_atomic_plane_check() guarantees
			     this for us)
visible: logically_enabled && dst rectangle is at least
         partially within pipe_src rectangle
active: visible && crtc_is_active

Currently we try to make the proper distinction between
logically_enabled vs. invisible, but we do not properly
handle the visible vs. active case. That is, we currently
mark the plane as invisible if the crtc is inactive.

That means we eg. calculate watermarks as if the plane was
invisible. That may cause a subsequent "DPMS on" operation
to fail unexpectedly because all of a sudden we realize 
that we don't have enough FIFO space for this particular
plane configuration. There's a FIXME somewhere in the plane
code about this.

-- 
Ville Syrjälä
Intel