If we keep the BO in the GTT domain, it means it will not be updated if
we validate it again later in kfd_mem_dmamap_dmabuf. This means we'll
use stale DMA addresses when we update the page tables after evictions.
I think we'll need to find a different way to avoid triggering the
eviction fence on the original BO when changing the placement of the
DMABuf import here. If you need help brainstorming here, please share a
backtrace from the eviction generated with the debug_evictions module param.
Regards,
Felix
Am 2023-04-03 um 13:59 schrieb Eric Huang:
dmabuf is allocated/mapped as GTT domain, when dma-unmapping dmabuf
changing placement to CPU will trigger memory eviction after calling
ttm_bo_validate, and the eviction will cause performance drop.
Keeping the correct domain will solve the issue.
Signed-off-by: Eric Huang <jinhuieric.huang@xxxxxxx>
---
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c
index a3b09edfd1bf..17b708acb447 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c
@@ -642,7 +642,7 @@ kfd_mem_dmaunmap_dmabuf(struct kfd_mem_attachment *attachment)
struct ttm_operation_ctx ctx = {.interruptible = true};
struct amdgpu_bo *bo = attachment->bo_va->base.bo;
- amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_CPU);
+ amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_GTT);
ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
}
