[PATCH] drm/amdgpu: Enable XGMI mapping for peer device

[Christian König]

Am 05.03.19 um 00:27 schrieb Liu, Shaoyun:
> Adjust vram base offset for XGMI mapping when update the PT entry so
> the address will fall into correct XGMI aperture for peer device
>
> Change-Id: I78bdf244da699d2559481ef5afe9663b3e752236
> Signed-off-by: shaoyunl <shaoyun.liu at amd.com>
> ---
>   drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 21 +++++++++++++++++----
>   1 file changed, 17 insertions(+), 4 deletions(-)
>
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
> index 12d51d9..0bc32c9 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
> @@ -1876,6 +1876,7 @@ static int amdgpu_vm_bo_update_mapping(struct amdgpu_device *adev,
>    * @vm: requested vm
>    * @mapping: mapped range and flags to use for the update
>    * @flags: HW flags for the mapping
> + * @bo_adev: amdgpu_device pointer that bo actually been allocated
>    * @nodes: array of drm_mm_nodes with the MC addresses
>    * @fence: optional resulting fence
>    *
> @@ -1891,11 +1892,13 @@ static int amdgpu_vm_bo_split_mapping(struct amdgpu_device *adev,
>   				      struct amdgpu_vm *vm,
>   				      struct amdgpu_bo_va_mapping *mapping,
>   				      uint64_t flags,
> +				      struct amdgpu_device *bo_adev,
>   				      struct drm_mm_node *nodes,
>   				      struct dma_fence **fence)
>   {
>   	unsigned min_linear_pages = 1 << adev->vm_manager.fragment_size;
>   	uint64_t pfn, start = mapping->start;
> +	bool is_xgmi = false;
>   	int r;
>   
>   	/* normally,bo_va->flags only contians READABLE and WIRTEABLE bit go here
> @@ -1917,6 +1920,10 @@ static int amdgpu_vm_bo_split_mapping(struct amdgpu_device *adev,
>   		flags |= AMDGPU_PTE_PRT;
>   		flags &= ~AMDGPU_PTE_VALID;
>   	}
> +	if (adev != bo_adev &&
> +	    adev->gmc.xgmi.hive_id &&
> +	    adev->gmc.xgmi.hive_id == bo_adev->gmc.xgmi.hive_id)
> +		is_xgmi = true;
>   
>   	trace_amdgpu_vm_bo_update(mapping);
>   
> @@ -1965,7 +1972,11 @@ static int amdgpu_vm_bo_split_mapping(struct amdgpu_device *adev,
>   			}
>   
>   		} else if (flags & AMDGPU_PTE_VALID) {
> -			addr += adev->vm_manager.vram_base_offset;
> +			if (is_xgmi == false) {
> +				DRM_ERROR("Failed to map the VRAM for other device access.\n");
> +				return -EINVAL;
> +			}

Please drop that checking and error message.

It is completely unrelated to the calculation done here and so certainly 
the wrong place for that kind of checking.

> +			addr += bo_adev->vm_manager.vram_base_offset;
>   			addr += pfn << PAGE_SHIFT;
>   		}
>   
> @@ -2012,6 +2023,7 @@ int amdgpu_vm_bo_update(struct amdgpu_device *adev,
>   	struct drm_mm_node *nodes;
>   	struct dma_fence *exclusive, **last_update;
>   	uint64_t flags;
> +	struct amdgpu_device *bo_adev = adev;
>   	int r;
>   
>   	if (clear || !bo) {
> @@ -2030,9 +2042,10 @@ int amdgpu_vm_bo_update(struct amdgpu_device *adev,
>   		exclusive = reservation_object_get_excl(bo->tbo.resv);
>   	}
>   
> -	if (bo)
> +	if (bo) {
>   		flags = amdgpu_ttm_tt_pte_flags(adev, bo->tbo.ttm, mem);
> -	else
> +		bo_adev = amdgpu_ttm_adev(bo->tbo.bdev);
> +	} else
>   		flags = 0x0;

Coding style says that you should always use "} else {".

Apart from that looks good to me,
Christian.

>   
>   	if (clear || (bo && bo->tbo.resv == vm->root.base.bo->tbo.resv))
> @@ -2050,7 +2063,7 @@ int amdgpu_vm_bo_update(struct amdgpu_device *adev,
>   
>   	list_for_each_entry(mapping, &bo_va->invalids, list) {
>   		r = amdgpu_vm_bo_split_mapping(adev, exclusive, pages_addr, vm,
> -					       mapping, flags, nodes,
> +					       mapping, flags, bo_adev, nodes,
>   					       last_update);
>   		if (r)
>   			return r;