Re: [PATCH v2] drm/amdgpu: refine amdgpu_bo_create_kernel_at()

["Lazar, Lijo" <lijo.lazar@xxxxxxx>]

On 7/28/2023 11:34 AM, Lang Yu wrote:
> Use amdgpu_bo_create_reserved() to create a BO in VRAM
> domain would fail if requested VRAM size is large(>128MB)
> on APU which usually has a default 512MB VRAM.
>
> That's because VRAM is framgented after several allocations.
>
> The approach is using amdgpu_bo_create_reserved() to
> create a BO in CPU domain first, it will always succeed.
>
> v2: Don't overwrite the contents at specific offset.
>
> Signed-off-by: Lang Yu <Lang.Yu@xxxxxxx>
> ---
>   drivers/gpu/drm/amd/amdgpu/amdgpu_object.c | 24 ++++++++++++++++------
>   drivers/gpu/drm/amd/amdgpu/amdgpu_object.h |  3 ++-
>   drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c    | 15 +++++++-------
>   drivers/gpu/drm/amd/amdgpu/amdgpu_virt.c   |  2 +-
>   4 files changed, 28 insertions(+), 16 deletions(-)
>
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_object.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_object.c
> index ff73cc11d47e..fa30e96f27d0 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_object.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_object.c
> @@ -358,6 +358,7 @@ int amdgpu_bo_create_kernel(struct amdgpu_device *adev,
>    * @offset: offset of the BO
>    * @size: size of the BO
>    * @bo_ptr:  used to initialize BOs in structures
> + * @gpu_addr: GPU addr of the pinned BO
>    * @cpu_addr: optional CPU address mapping
>    *
>    * Creates a kernel BO at a specific offset in VRAM.
> @@ -367,7 +368,8 @@ int amdgpu_bo_create_kernel(struct amdgpu_device *adev,
>    */
>   int amdgpu_bo_create_kernel_at(struct amdgpu_device *adev,
>   			       uint64_t offset, uint64_t size,
> -			       struct amdgpu_bo **bo_ptr, void **cpu_addr)
> +			       struct amdgpu_bo **bo_ptr,
> +			       u64 *gpu_addr, void **cpu_addr)

A generic question (not considering other details in this patch) - this 
API is literally asking to create BO at a particular GPU VRAM offset. 
The offset goes in as the input, so why does it need to return a GPU offset?

Thanks,
Lijo

>   {
>   	struct ttm_operation_ctx ctx = { false, false };
>   	unsigned int i;
> @@ -377,32 +379,42 @@ int amdgpu_bo_create_kernel_at(struct amdgpu_device *adev,
>   	size = ALIGN(size, PAGE_SIZE);
>   
>   	r = amdgpu_bo_create_reserved(adev, size, PAGE_SIZE,
> -				      AMDGPU_GEM_DOMAIN_VRAM, bo_ptr, NULL,
> -				      cpu_addr);
> +				      AMDGPU_GEM_DOMAIN_CPU,
> +				      bo_ptr, NULL, NULL);
>   	if (r)
>   		return r;
>   
>   	if ((*bo_ptr) == NULL)
>   		return 0;
>   
> +	(*bo_ptr)->flags = AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS;
> +	(*bo_ptr)->flags |= cpu_addr ? AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED
> +		: AMDGPU_GEM_CREATE_NO_CPU_ACCESS;
> +
>   	/*
>   	 * Remove the original mem node and create a new one at the request
>   	 * position.
>   	 */
> -	if (cpu_addr)
> -		amdgpu_bo_kunmap(*bo_ptr);
> -
>   	ttm_resource_free(&(*bo_ptr)->tbo, &(*bo_ptr)->tbo.resource);
>   
>   	for (i = 0; i < (*bo_ptr)->placement.num_placement; ++i) {
>   		(*bo_ptr)->placements[i].fpfn = offset >> PAGE_SHIFT;
>   		(*bo_ptr)->placements[i].lpfn = (offset + size) >> PAGE_SHIFT;
> +		(*bo_ptr)->placements[i].mem_type = TTM_PL_VRAM;
> +		(*bo_ptr)->placements[i].flags = TTM_PL_FLAG_CONTIGUOUS;
> +
> +		if (!((*bo_ptr)->flags & AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED))
> +			(*bo_ptr)->placements[i].flags |= TTM_PL_FLAG_TOPDOWN;
>   	}
> +
>   	r = ttm_bo_mem_space(&(*bo_ptr)->tbo, &(*bo_ptr)->placement,
>   			     &(*bo_ptr)->tbo.resource, &ctx);
>   	if (r)
>   		goto error;
>   
> +	if (gpu_addr)
> +		*gpu_addr = amdgpu_bo_gpu_offset(*bo_ptr);
> +
>   	if (cpu_addr) {
>   		r = amdgpu_bo_kmap(*bo_ptr, cpu_addr);
>   		if (r)
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_object.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_object.h
> index 5d3440d719e4..8f5b5664a1b6 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_object.h
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_object.h
> @@ -315,7 +315,8 @@ int amdgpu_bo_create_kernel(struct amdgpu_device *adev,
>   			    u64 *gpu_addr, void **cpu_addr);
>   int amdgpu_bo_create_kernel_at(struct amdgpu_device *adev,
>   			       uint64_t offset, uint64_t size,
> -			       struct amdgpu_bo **bo_ptr, void **cpu_addr);
> +			       struct amdgpu_bo **bo_ptr,
> +			       u64 *gpu_addr, void **cpu_addr);
>   int amdgpu_bo_create_user(struct amdgpu_device *adev,
>   			  struct amdgpu_bo_param *bp,
>   			  struct amdgpu_bo_user **ubo_ptr);
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c
> index 7c6dd3de1867..a210c243dac0 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c
> @@ -1619,7 +1619,7 @@ static int amdgpu_ttm_fw_reserve_vram_init(struct amdgpu_device *adev)
>   					  adev->mman.fw_vram_usage_start_offset,
>   					  adev->mman.fw_vram_usage_size,
>   					  &adev->mman.fw_vram_usage_reserved_bo,
> -					  &adev->mman.fw_vram_usage_va);
> +					  NULL, &adev->mman.fw_vram_usage_va);
>   }
>   
>   /**
> @@ -1644,7 +1644,7 @@ static int amdgpu_ttm_drv_reserve_vram_init(struct amdgpu_device *adev)
>   					  adev->mman.drv_vram_usage_start_offset,
>   					  adev->mman.drv_vram_usage_size,
>   					  &adev->mman.drv_vram_usage_reserved_bo,
> -					  &adev->mman.drv_vram_usage_va);
> +					  NULL, &adev->mman.drv_vram_usage_va);
>   }
>   
>   /*
> @@ -1729,8 +1729,7 @@ static int amdgpu_ttm_reserve_tmr(struct amdgpu_device *adev)
>   		ret = amdgpu_bo_create_kernel_at(adev,
>   						 ctx->c2p_train_data_offset,
>   						 ctx->train_data_size,
> -						 &ctx->c2p_bo,
> -						 NULL);
> +						 &ctx->c2p_bo, NULL, NULL);
>   		if (ret) {
>   			DRM_ERROR("alloc c2p_bo failed(%d)!\n", ret);
>   			amdgpu_ttm_training_reserve_vram_fini(adev);
> @@ -1742,7 +1741,7 @@ static int amdgpu_ttm_reserve_tmr(struct amdgpu_device *adev)
>   	if (!adev->gmc.is_app_apu) {
>   		ret = amdgpu_bo_create_kernel_at(
>   			adev, adev->gmc.real_vram_size - reserve_size,
> -			reserve_size, &adev->mman.fw_reserved_memory, NULL);
> +			reserve_size, &adev->mman.fw_reserved_memory, NULL, NULL);
>   		if (ret) {
>   			DRM_ERROR("alloc tmr failed(%d)!\n", ret);
>   			amdgpu_bo_free_kernel(&adev->mman.fw_reserved_memory,
> @@ -1885,14 +1884,14 @@ int amdgpu_ttm_init(struct amdgpu_device *adev)
>   		r = amdgpu_bo_create_kernel_at(adev, 0,
>   					       adev->mman.stolen_vga_size,
>   					       &adev->mman.stolen_vga_memory,
> -					       NULL);
> +					       NULL, NULL);
>   		if (r)
>   			return r;
>   
>   		r = amdgpu_bo_create_kernel_at(adev, adev->mman.stolen_vga_size,
>   					       adev->mman.stolen_extended_size,
>   					       &adev->mman.stolen_extended_memory,
> -					       NULL);
> +					       NULL, NULL);
>   
>   		if (r)
>   			return r;
> @@ -1901,7 +1900,7 @@ int amdgpu_ttm_init(struct amdgpu_device *adev)
>   					       adev->mman.stolen_reserved_offset,
>   					       adev->mman.stolen_reserved_size,
>   					       &adev->mman.stolen_reserved_memory,
> -					       NULL);
> +					       NULL, NULL);
>   		if (r)
>   			return r;
>   	} else {
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_virt.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_virt.c
> index 41aa853a07d2..b93b42b916ce 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_virt.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_virt.c
> @@ -397,7 +397,7 @@ static void amdgpu_virt_ras_reserve_bps(struct amdgpu_device *adev)
>   		 */
>   		if (amdgpu_bo_create_kernel_at(adev, bp << AMDGPU_GPU_PAGE_SHIFT,
>   					       AMDGPU_GPU_PAGE_SIZE,
> -					       &bo, NULL))
> +					       &bo, NULL, NULL))
>   			DRM_DEBUG("RAS WARN: reserve vram for retired page %llx fail\n", bp);
>   
>   		data->bps_bo[i] = bo;