On 2023-12-20 8:58, Christian König wrote:
Am 19.12.23 um 23:43 schrieb Felix Kuehling:
On 2023-12-19 3:10, Christian König wrote:
Am 15.12.23 um 16:19 schrieb Felix Kuehling:

On 2023-12-15 07:30, Christian König wrote:
@@ -1425,11 +1451,21 @@ int amdgpu_vm_handle_moved(struct amdgpu_device *adev,
            r = amdgpu_vm_bo_update(adev, bo_va, clear);
-        if (r)
-            return r;
            if (unlock)
+        if (r)
+            return r;
+        /* Remember evicted DMABuf imports in compute VMs for later
+         * validation
+         */
+        if (vm->is_compute_context && bo_va->base.bo &&
+ bo_va->base.bo->tbo.base.import_attach &&
+            (!bo_va->base.bo->tbo.resource ||
+ bo_va->base.bo->tbo.resource->mem_type == TTM_PL_SYSTEM))
+            amdgpu_vm_bo_evicted(&bo_va->base);

The change looks mostly good now. Just one thing which worries me is that when GFX and compute is mixed in the same VM this here might cause problems when we run into an error during command submission.

E.g. GFX validates the VM BOs, but then the IOCTL fails before calling amdgpu_vm_handle_moved().

In this case the DMA-buf wouldn't be validated any more.

This code path shouldn't be relevant for command submission, but for the amdgpu_vm_handle_moved call in amdgpu_dma_buf_move_notify. That's where the BO is first found to be evicted and its PTEs invalidated. That's where we need to remember it so it can be validated in the next call to amdgpu_vm_validate.

Currently the amdgpu_cs code path calls amdgpu_vm_validate with ticket=NULL, so it won't validate these imports. The only place that auto-validates evicted imports is amdgpu_amdkfd_restore_process_bos. So none of this should affect amdgpu_cs command submission.

Yeah, but ticket=NULL will result in removing those imports from the validation list.

I have a comment for that in amdgpu_vm_validate:

                         if (!ticket) {
                                 /* We need to move the BO out of the evicted
                                  * list to avoid an infinite loop. It will be
                                  * moved back to evicted in the next
                                  * amdgpu_vm_handle_moved.

The net result is that the BO is still tracked as evicted.

Yeah, that's exactly what I mean:

What happens if amdgpu_vm_validate() is called, removes the BOs from the evicted list, but then an error happens (or just an interrupted system call) and amdgpu_vm_handle_moved is never called?

In this case the DMA-bufs would be on the moved list and amdgpu_vm_handle moved would have to be called once before we can validate them again.

And that would be a good argument for tracking evicted user BOs with a separate state after all. That way we could just leave that list alone if ticket==NULL in amdgpu_vm_validate.



This could potentially result in not validating them should we ever mix GFX and Compute submissions in the same VM.

My eviction test does exactly this, and it's working just fine.


For now that works, but we need to clean that up more thoughtfully I think.


The flow for amdgpu_amdkfd_restore_process_bos is:

 * amdgpu_vm_validate validates the evicted dmabuf imports (moves the
   bo_vas from "evicted" to "invalidated")
 * amdgpu_vm_handle_moved iterates over invalidated bo_vas and updates
   the PTEs with valid entries (moves the bo_vas to "done")



Reply via email to