On 11/27/24 09:30, Kasireddy, Vivek wrote: >> Note: While verifying move_notify(), I noticed that AMD/TTM driver moves >> same shared FB GEMs after each framebuffer update when it renders into >> FB, despite of the 32GB BAR. This should be rather inefficient. I'd >> expect dma-buf staying static if there is no need to evict it. Something >> to check how it works with DG2. > IIUC, I think the exporting driver (AMD GPU driver) in the Guest VM migrates the > FB to System RAM because during export it determines that virtio-gpu is not P2P > compatible. This behavior is expected and seen with other dGPU drivers (i915/Xe) > as well. However, I am trying to fix this in Xe driver for specific use-cases (SRIOV) > where we know for sure that the importer on the Host is another dGPU: > https://patchwork.kernel.org/project/dri-devel/ > cover/20241012024524.1377836-1-vivek.kasireddy@xxxxxxxxx/ > https://patchwork.kernel.org/project/dri-devel/ > cover/20240624065552.1572580-1-vivek.kasireddy@xxxxxxxxx/ Are the dmabufs static with a disabled MOVE_NOTIFY or TTM uses bounce buffer in this case? -- Best regards, Dmitry
