[PATCH 4/6] drm/amdgpu: Attach eviction fence on alloc

[Chen, Xiaogang]

Reviewed-by: Xiaogang Chen <Xiaoganng.Chen at amd.com>

Regards

Xiaogang

On 1/11/2023 7:31 PM, Felix Kuehling wrote:
> Instead of attaching the eviction fence when a KFD BO is first mapped,
> attach it when it is allocated or imported. This in preparation to allow
> KFD BOs to be mapped using the render node API.
>
> Signed-off-by: Felix Kuehling <Felix.Kuehling at amd.com>
> ---
>   .../gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c  | 63 ++++++++++---------
>   1 file changed, 32 insertions(+), 31 deletions(-)
>
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c
> index 5645103beed0..79213f476493 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c
> @@ -360,6 +360,24 @@ static int amdgpu_amdkfd_bo_validate(struct amdgpu_bo *bo, uint32_t domain,
>   	return ret;
>   }
>   
> +static int amdgpu_amdkfd_bo_validate_and_fence(struct amdgpu_bo *bo,
> +					       uint32_t domain,
> +					       struct dma_fence *fence)
> +{
> +	int ret = amdgpu_bo_reserve(bo, false);
> +
> +	if (ret)
> +		return ret;
> +
> +	ret = amdgpu_amdkfd_bo_validate(bo, domain, true);
> +	if (!ret)
> +		dma_resv_add_fence(bo->tbo.base.resv, fence,
> +				   DMA_RESV_USAGE_BOOKKEEP);
> +	amdgpu_bo_unreserve(bo);
> +
> +	return ret;
> +}
> +
>   static int amdgpu_amdkfd_validate_vm_bo(void *_unused, struct amdgpu_bo *bo)
>   {
>   	return amdgpu_amdkfd_bo_validate(bo, bo->allowed_domains, false);
> @@ -1720,6 +1738,11 @@ int amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu(
>   		}
>   		bo->allowed_domains = AMDGPU_GEM_DOMAIN_GTT;
>   		bo->preferred_domains = AMDGPU_GEM_DOMAIN_GTT;
> +	} else {
> +		ret = amdgpu_amdkfd_bo_validate_and_fence(bo, domain,
> +				&avm->process_info->eviction_fence->base);
> +		if (ret)
> +			goto err_validate_bo;
>   	}
>   
>   	if (offset)
> @@ -1729,6 +1752,7 @@ int amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu(
>   
>   allocate_init_user_pages_failed:
>   err_pin_bo:
> +err_validate_bo:
>   	remove_kgd_mem_from_kfd_bo_list(*mem, avm->process_info);
>   	drm_vma_node_revoke(&gobj->vma_node, drm_priv);
>   err_node_allow:
> @@ -1804,10 +1828,6 @@ int amdgpu_amdkfd_gpuvm_free_memory_of_gpu(
>   	if (unlikely(ret))
>   		return ret;
>   
> -	/* The eviction fence should be removed by the last unmap.
> -	 * TODO: Log an error condition if the bo still has the eviction fence
> -	 * attached
> -	 */
>   	amdgpu_amdkfd_remove_eviction_fence(mem->bo,
>   					process_info->eviction_fence);
>   	pr_debug("Release VA 0x%llx - 0x%llx\n", mem->va,
> @@ -1931,19 +1951,6 @@ int amdgpu_amdkfd_gpuvm_map_memory_to_gpu(
>   	if (unlikely(ret))
>   		goto out_unreserve;
>   
> -	if (mem->mapped_to_gpu_memory == 0 &&
> -	    !amdgpu_ttm_tt_get_usermm(bo->tbo.ttm)) {
> -		/* Validate BO only once. The eviction fence gets added to BO
> -		 * the first time it is mapped. Validate will wait for all
> -		 * background evictions to complete.
> -		 */
> -		ret = amdgpu_amdkfd_bo_validate(bo, domain, true);
> -		if (ret) {
> -			pr_debug("Validate failed\n");
> -			goto out_unreserve;
> -		}
> -	}
> -
>   	list_for_each_entry(entry, &mem->attachments, list) {
>   		if (entry->bo_va->base.vm != avm || entry->is_mapped)
>   			continue;
> @@ -1970,10 +1977,6 @@ int amdgpu_amdkfd_gpuvm_map_memory_to_gpu(
>   			 mem->mapped_to_gpu_memory);
>   	}
>   
> -	if (!amdgpu_ttm_tt_get_usermm(bo->tbo.ttm) && !bo->tbo.pin_count)
> -		dma_resv_add_fence(bo->tbo.base.resv,
> -				   &avm->process_info->eviction_fence->base,
> -				   DMA_RESV_USAGE_BOOKKEEP);
>   	ret = unreserve_bo_and_vms(&ctx, false, false);
>   
>   	goto out;
> @@ -1990,7 +1993,6 @@ int amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu(
>   		struct amdgpu_device *adev, struct kgd_mem *mem, void *drm_priv)
>   {
>   	struct amdgpu_vm *avm = drm_priv_to_vm(drm_priv);
> -	struct amdkfd_process_info *process_info = avm->process_info;
>   	unsigned long bo_size = mem->bo->tbo.base.size;
>   	struct kfd_mem_attachment *entry;
>   	struct bo_vm_reservation_context ctx;
> @@ -2031,15 +2033,6 @@ int amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu(
>   			 mem->mapped_to_gpu_memory);
>   	}
>   
> -	/* If BO is unmapped from all VMs, unfence it. It can be evicted if
> -	 * required.
> -	 */
> -	if (mem->mapped_to_gpu_memory == 0 &&
> -	    !amdgpu_ttm_tt_get_usermm(mem->bo->tbo.ttm) &&
> -	    !mem->bo->tbo.pin_count)
> -		amdgpu_amdkfd_remove_eviction_fence(mem->bo,
> -						process_info->eviction_fence);
> -
>   unreserve_out:
>   	unreserve_bo_and_vms(&ctx, false, false);
>   out:
> @@ -2266,8 +2259,16 @@ int amdgpu_amdkfd_gpuvm_import_dmabuf(struct amdgpu_device *adev,
>   	amdgpu_sync_create(&(*mem)->sync);
>   	(*mem)->is_imported = true;
>   
> +	ret = amdgpu_amdkfd_bo_validate_and_fence(bo, (*mem)->domain,
> +				&avm->process_info->eviction_fence->base);
> +	if (ret)
> +		goto err_remove_mem;
> +
>   	return 0;
>   
> +err_remove_mem:
> +	remove_kgd_mem_from_kfd_bo_list(*mem, avm->process_info);
> +	drm_vma_node_revoke(&obj->vma_node, drm_priv);
>   err_free_mem:
>   	kfree(*mem);
>   err_put_obj: