Am 2022-10-05 um 07:03 schrieb Danijel Slivka:
CPU pagetable updates have issues with HDP flush as VF MMIO access
protection is not allowing write during sriov runtime to
mmBIF_BX_DEV0_EPF0_VF0_HDP_MEM_COHERENCY_FLUSH_CNTL

Signed-off-by: Danijel Slivka <danijel.sli...@amd.com>
---
  drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 4 +++-
  1 file changed, 3 insertions(+), 1 deletion(-)

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c 
b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
index 83b0c5d86e48..32088ac0666c 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
@@ -2338,7 +2338,9 @@ void amdgpu_vm_manager_init(struct amdgpu_device *adev)
         */
  #ifdef CONFIG_X86_64
        if (amdgpu_vm_update_mode == -1) {
-               if (amdgpu_gmc_vram_full_visible(&adev->gmc))
+               if (amdgpu_gmc_vram_full_visible(&adev->gmc) &&
+                   !(adev->asic_type == CHIP_SIENNA_CICHLID &&
+                   amdgpu_sriov_vf(adev)))

This would need at least a code comment. But I'd prefer a more general solution that expresses that some ASICs don't allow any MMIO access under SRIOV.

I found that there is this function defined in amdgpu_virt.c/h: bool amdgpu_virt_mmio_blocked(struct amdgpu_device *adev). Would this return the correct result and could you use it here instead of a hard-coded asic_type?

Or maybe this could be added as a flag in (adev)->virt.caps and get initialized in some ASIC-specific code path.

Regards,
  Felix


                        adev->vm_manager.vm_update_mode =
                                AMDGPU_VM_USE_CPU_FOR_COMPUTE;
                else

Reply via email to