Hi guys, I finally found time to work on queue reset a bit more and also gave it some more testing. The per VMID reset actually seems to work 100% reliable at least on gfx9. What still could be is that an application is using multiple VMIDs on the graphics ring or that we re-use the same VMID for a different application. Preventing this will need some more work on the VMID management. The compute queue reset as well as the gfx11 and gfx12 code unfortunately still has some calls to amdgpu_bo_reserv/amdgpu_bo_kmap in their path which are totally illegal inside the reset handling, but that can be fixed later on. Please test, comment and review, Christian.
