Semantics of the 'dumb' interface

[Dave Airlie]

On Fri, Jun 3, 2011 at 4:37 AM, Alan Cox <alan at lxorguk.ukuu.org.uk> wrote:
> I have GEM allocation working on the GMA 500 and I can scribble in a
> framebuffer (minus an odd 'last page' bug which is an off by one
> somewhere I imagine) and display it with nice modetest stripes.
>
> If I kill the program however it all goes kerblam
>
> drm_release calls into fb_release which duely destroys the user frame
> buffer. This drops the gem object reference and we start to release all
> the resources but it blows up because the GEM object is still pinned in
> the GTT as it's still being used as scanout.
>
> Where I'm a bit lost right now is understanding what is supposed to
> happen at this point, and how the scanout is supposed to have ended up
> cleaned up and reset to the system frame buffer beforehand thus unpinning
> the scanout from the GTT.

I think the intel driver forces a reset to the system scanout on release. I've
actually never found a test to indicate it was completely necessary on
other GPUs
since they scan out of real VRAM which isn't going to get unbound from the card.

Dave.