Nvidia-prop -> AMDGPU sharing does not work

[hoshinolina]

If the GBM BO is allocated on the Nvidia side, AMDGPU fails all submissions that use the BO with:

amdgpu 0000:06:00.0: amdgpu: [drm] *ERROR* Not enough memory for command submission!

If the BO is allocated by an AMDGPU GBM instance, it works fine (and imports back on the nvidia for blitting), but that's not really how you're supposed to do it.

[jpietek]

It this a problem only for the Nvidia Proprietary driver, and it works on NVK/Nouveau? Generally I wonder if it does make sense to do some testing of libfunnel on Nvidia Proprietary, or is it still bound to break?

[hoshinolina]

It's a problem only on nvidia prop, yeah (but the actual failure happens on the AMD side after buffer import).

libfunnel generally works on nvidia-prop without cross-GPU (since I added explicit sync support), I just don't know what's going on with this re cross-GPU.

AMD -> nvidia-prop with the usual use case (libfunnel->OBS) does not work for a different reason (nvidia-prop does not support importing linear textures into OpenGL with the usual APIs, requiring a whole mess of alternate workaround code that nobody wants to implement).