[PATCH 3/4] drm/amdgpu: Support snooped PTE flag

[Kuehling, Felix]

On 2019-08-26 2:12 p.m., Koenig, Christian wrote:
>
> We should probably just revert back the to doing most of the mapping in
> amdgpu_vm_bo_split_mapping().
>
> Here we already have a whole bunch of ASIC specific handling anyway:
>
>>          if (!(mapping->flags & AMDGPU_PTE_READABLE))
>>                  flags &= ~AMDGPU_PTE_READABLE;
>>          if (!(mapping->flags & AMDGPU_PTE_WRITEABLE))
>>                  flags &= ~AMDGPU_PTE_WRITEABLE;
>>
>>          flags &= ~AMDGPU_PTE_EXECUTABLE;
>>          flags |= mapping->flags & AMDGPU_PTE_EXECUTABLE;
>>
>>          if (adev->asic_type == CHIP_NAVI10) {
>>                  flags &= ~AMDGPU_PTE_MTYPE_NV10_MASK;
>>                  flags |= (mapping->flags & AMDGPU_PTE_MTYPE_NV10_MASK);
>>          } else {
>>                  flags &= ~AMDGPU_PTE_MTYPE_VG10_MASK;
>>                  flags |= (mapping->flags & AMDGPU_PTE_MTYPE_VG10_MASK);
>>          }
>>
>>          if ((mapping->flags & AMDGPU_PTE_PRT) &&
>>              (adev->asic_type >= CHIP_VEGA10)) {
>>                  flags |= AMDGPU_PTE_PRT;
>>                  if (adev->asic_type >= CHIP_NAVI10) {
>>                          flags |= AMDGPU_PTE_SNOOPED;
>>                          flags |= AMDGPU_PTE_LOG;
>>                          flags |= AMDGPU_PTE_SYSTEM;
>>                  }
>>                  flags &= ~AMDGPU_PTE_VALID;
>>          }
> And now that you mentioned it, the code you proposed wouldn't have
> worked anyway because the AMDGPU_PTE_SNOOPED would have been filtered
> out here :)

OK, and that does look like the right place to set the PTE_SNOOPED bit 
for XGMI mappings on Arcturus. I'll send out an updated patch series 
that no longer needs the AMDGPU_VM_PAGE_INVALIDATE_PROBE bit in the UAPI.

Thanks for the pointer,
   Felix


>
> Regards,
> Christian.
>