[PATCH 7/8] drm/amdgpu: track evicted page tables

[Felix Kuehling]

On 2017-08-10 10:50 AM, Christian König wrote:
> @@ -2384,11 +2360,24 @@ void amdgpu_vm_bo_rmv(struct amdgpu_device *adev,
>   * Mark @bo as invalid.
>   */
>  void amdgpu_vm_bo_invalidate(struct amdgpu_device *adev,
> -			     struct amdgpu_bo *bo)
> +			     struct amdgpu_bo *bo, bool evicted)
>  {
>  	struct amdgpu_vm_bo_base *bo_base;
>  
>  	list_for_each_entry(bo_base, &bo->va, bo_list) {
> +		struct amdgpu_vm *vm = bo_base->vm;
> +
> +		if (evicted && bo->tbo.resv == vm->root.base.bo->tbo.resv) {
> +			spin_lock(&bo_base->vm->status_lock);
> +			list_move(&bo_base->vm_status, &vm->evicted);
> +			spin_unlock(&bo_base->vm->status_lock);
> +			continue;
> +		}
> +
> +		/* Don't add page tables to the moved state */
> +		if (bo->tbo.type == ttm_bo_type_kernel)
> +			continue;
> +

It looks like you're using two different conditions to detect page
tables. When evicted is true, you use (bo->tbo.resv ==
vm->root.base.bo->tbo.resv). Otherwise you use (bo->tbo.type ==
ttm_bo_type_kernel). Are these really the same or is there a subtle
difference I'm missing? I'd find the following logic easier to read, but
I'm not sure if that's really your intention:

	if (bo->tbo.resv == vm->root.base.bo->tbo.resv) {
		spin_lock(&bo_base->vm->status_lock);

		if (evicted)
			list_move(&bo_base->vm_status, &vm->evicted);
		else if (list_empty(&bo_base->vm_status))
			/* this is from the next commit */
			list_add(&bo_base->vm_status, &vm->relocated);

		spin_unlock(&bo_base->vm->status_lock);
		continue;
	}

Regards,
  Felix

>  		spin_lock(&bo_base->vm->status_lock);
>  		if (list_empty(&bo_base->vm_status))
>  			list_add(&bo_base->vm_status,