amdgpu_gem_object_close() calls amdgpu_vm_clear_freed() after deleting a BO VA. 
If vm->freed is empty, that call is a no-op but still allocates sync state and 
walks reservation fences before returning.

Check vm->freed first to avoid the overhead on the GEM close hot path. This 
does not change behavior because the empty-list path leaves the fence unset and 
returns success.

Signed-off-by: Bob Zhou <[email protected]>
---
 drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c 
b/drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c
index 6a0699746fbcd..72811f6963a15 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c
@@ -348,6 +348,8 @@ static void amdgpu_gem_object_close(struct drm_gem_object 
*obj,
        amdgpu_vm_bo_update_shared(bo);
        if (!amdgpu_vm_ready(vm))
                goto out_unlock;
+       if (list_empty(&vm->freed))
+               goto out_unlock;
 
        r = amdgpu_vm_clear_freed(adev, vm, &fence);
        if (unlikely(r < 0) && !drm_dev_is_unplugged(adev_to_drm(adev)))
-- 
2.34.1

Reply via email to