On 2023-10-10 10:32, Francis, David
wrote:
[AMD Official Use Only - General]
[AMD Official Use Only - General]
By the intended semantics of the EXT_COHERENT parameter,
it should cause the driver to use MTYPE_UC instead of
MTYPE_NC on non-local gfx9.4.3 APU memory.
Signed-off-by: David Francis <David.Francis@xxxxxxx>
---
drivers/gpu/drm/amd/amdkfd/kfd_svm.c | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c
index 7b81233bc9ae..d9ef937308ed 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c
@@ -1286,7 +1286,7 @@ svm_range_get_pte_flags(struct kfd_node *node,
if (num_possible_nodes() <= 1)
mapping_flags |= mtype_local;
else
- mapping_flags |= AMDGPU_VM_MTYPE_NC;
+ mapping_flags |= ext_coherent ? AMDGPU_VM_MTYPE_UC : AMDGPU_VM_MTYPE_NC;
Thanks for finding this bug, I think on NUMA system with ext_coherent, we should use CC for local and UC for remote memory from another NUMA node.
The NUMA locality check is done in amdgpu_gmc_override_vm_pte_flag now, seems we need refactor and move the check here in order to use the prange->flags.
Regards,
Philip
/* system memory accessed by the dGPU */
} else {
mapping_flags |= AMDGPU_VM_MTYPE_UC;
--
2.34.1
