I can very quickly, almost deterministically, hang the GPU (radeonsi, Cape
Verde) with the following command:

> LIBGL_ALWAYS_SOFTWARE=1 mpv --fs --vo=opengl:sw /path/to/some_video

this works on both 3.16.0 and 3.17rc3. Try seeking, it often happens directly
after a seek. In most cases, the hang is unrecoverable and crashes the kernel
after some "atombios stuck in a loop" messages. Very strange indeed, software
rendered glxgears doesn't cause this.

Can anyone verify? A somewhat reliable test case might be a good start to
finally fixing this.