[Intel-gfx] [PATCH 3/3] drm: Track framebuffer references at the point of assignment

[Chris Wilson]

On Mon, Nov 28, 2016 at 03:15:12PM +0100, Daniel Vetter wrote:
> On Mon, Nov 28, 2016 at 08:46:11AM +0000, Chris Wilson wrote:
> > On Mon, Nov 28, 2016 at 08:48:34AM +0100, Daniel Vetter wrote:
> > > On Fri, Nov 25, 2016 at 03:32:31PM +0000, Chris Wilson wrote:
> > > > --- a/drivers/gpu/drm/drm_atomic.c
> > > > +++ b/drivers/gpu/drm/drm_atomic.c
> > > > @@ -1253,7 +1253,7 @@ drm_atomic_set_fb_for_plane(struct drm_plane_state *plane_state,
> > > >  		DRM_DEBUG_ATOMIC("Set [NOFB] for plane state %p\n",
> > > >  				 plane_state);
> > > >  
> > > > -	drm_framebuffer_assign(&plane_state->fb, fb);
> > > > +	drm_framebuffer_assign(__mkwrite_drm_framebuffer(&plane_state->fb), fb);
> > > 
> > > Feels like const gone the wrong way round? Or I'm not entirely clear on
> > > what this is supposed to achieve. Just dropping the const would still be a
> > > nice improvement.
> > 
> > No one is supposed to be writing to it. Adding the const generates a
> > compiler warning for incorrect code that doesn't do the refcounting -
> > but doesn't generate a warning for anything simply checking the value.
> > It is the moral equivalent to calling it _fb to catch all the users.
> 
> But then shouldn't all users use the drm_plane_set_fb helper (and that
> maybe gain a comment)? The above escaped the conversion/consolidation,
> which is why it looked funny to me ...

Yes, all users (of plane->fb = fb) are. The above is for
plane_state->fb.

I wasn't sure about adding comments to struct drm_plane, I thought the
crtc/fb there was just a crutch for !atomic.

But adding
+
+       /**
+        * @fb:
+        *
+        * Currently active framebuffer. Do not write this directly, use
+        * drm_plane_set_fb()
+        */
        struct drm_framebuffer * const fb;

is not a hardship.
-Chris

-- 
Chris Wilson, Intel Open Source Technology Centre