Re: [PATCH] Add virtio gpu driver.

[Daniel Vetter]

On Mon, May 30, 2016 at 03:50:36PM +0200, Gerd Hoffmann wrote:
>   Hi,
> 
> > - add a small core function to registerr HOT_X/HOT_Y for a (cursor) plane,
> >   e.g. drm_plane_register_hotspot(). That should allocate the properties
> >   (if they don't exist yet) and then attach those props to the cursor. We
> >   don't want those props everywhere, but only on drivers that support/need
> >   them, aka virtual hw.
> 
> Hmm, why is this special to virtual hw?
> 
> >  	if (crtc->cursor) {
> > -		ret = drm_mode_cursor_universal(crtc, req, file_priv);
> > +		if (drm_core_check_feature(DRIVER_ATOMIC))
> > +			ret = drm_mode_cursor_atomic(crtc, req, file_priv);
> > +		else
> > +			ret = drm_mode_cursor_universal(crtc, req, file_priv);
> >  		goto out;
> 
> >   drm_mode_cursor_atomic would simply be a fusing of
> >   drm_mode_cursor_universal + drm_atomic_helper_update_plane (dump all the
> >   intermediate variables and store directly in the plane state), with the
> >   addition of also storing hot_x/y into the plane state.
> 
> Hmm, that'll either make drm_mode_cursor_atomic a big cut+pasted
> function, or need quite some refactoring to move common code into
> functions callable from both drm_mode_cursor_atomic
> +drm_mode_cursor_universal ...
> 
> Why attach the hotspot to the plane?  Wouldn't it make more sense to
> make it a framebuffer property?

We don't have properties on the framebuffer. I guess you /could/ just add
it internally to struct drm_framebuffer, and not bother exposing to
userspace. I guess that would be a lot simpler, but it also means that
atomic userspace can't use hotspots before we add properties to fbs. And
doing that is a bit tricky since drm_framebuffer objects are meant to be
invariant - this assumption is deeply in-grained into the code all over
the place, everything just compares pointers when semantically it means to
compare the entire fb (including backing storage pointer/offsets and
everything).

So would be a bit more work to wire up for atomic userspace, but indeed a
lot less work to implement. I'm totally happy if you go with that tradeoff
;-)

Cheers, Daniel
-- 
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch