[PATCH] drm/amdgpu: Pin MMIO/DOORBELL BO's in GTT domain

[Felix Kuehling]

On 2021-11-09 2:12 p.m., Ramesh Errabolu wrote:
> MMIO/DOORBELL BOs encode control data and should be pinned in GTT
> domain before enabling PCIe connected peer devices in accessing it
>
> Signed-off-by: Ramesh Errabolu <Ramesh.Errabolu at amd.com>

Reviewed-by: Felix Kuehling <Felix.Kuehling at amd.com>


> ---
>   .../gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c  | 70 +++++++++++++++++++
>   1 file changed, 70 insertions(+)
>
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c
> index 08675f89bfb2..5e063fac0250 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c
> @@ -1297,6 +1297,56 @@ static int init_kfd_vm(struct amdgpu_vm *vm, void **process_info,
>   	return ret;
>   }
>   
> +/**
> + * amdgpu_amdkfd_gpuvm_pin_bo() - Pins a BO using following criteria
> + * @bo: Handle of buffer object being pinned
> + * @domain: Domain into which BO should be pinned
> + *
> + *   - USERPTR BOs are UNPINNABLE and will return error
> + *   - All other BO types (GTT, VRAM, MMIO and DOORBELL) will have their
> + *     PIN count incremented. It is valid to PIN a BO multiple times
> + *
> + * Return: ZERO if successful in pinning, Non-Zero in case of error.
> + * Will return -EINVAL if input BO parameter is a USERPTR type.
> + */
> +static int amdgpu_amdkfd_gpuvm_pin_bo(struct amdgpu_bo *bo, u32 domain)
> +{
> +	int ret = 0;
> +
> +	ret = amdgpu_bo_reserve(bo, false);
> +	if (unlikely(ret))
> +		return ret;
> +
> +	ret = amdgpu_bo_pin_restricted(bo, domain, 0, 0);
> +	if (ret)
> +		pr_err("Error in Pinning BO to domain: %d\n", domain);
> +
> +	amdgpu_bo_sync_wait(bo, AMDGPU_FENCE_OWNER_KFD, false);
> +	amdgpu_bo_unreserve(bo);
> +
> +	return ret;
> +}
> +
> +/**
> + * amdgpu_amdkfd_gpuvm_unpin_bo() - Unpins BO using following criteria
> + * @bo: Handle of buffer object being unpinned
> + *
> + *   - Is a illegal request for USERPTR BOs and is ignored
> + *   - All other BO types (GTT, VRAM, MMIO and DOORBELL) will have their
> + *     PIN count decremented. Calls to UNPIN must balance calls to PIN
> + */
> +void amdgpu_amdkfd_gpuvm_unpin_bo(struct amdgpu_bo *bo)
> +{
> +	int ret = 0;
> +
> +	ret = amdgpu_bo_reserve(bo, false);
> +	if (unlikely(ret))
> +		return;
> +
> +	amdgpu_bo_unpin(bo);
> +	amdgpu_bo_unreserve(bo);
> +}
> +
>   int amdgpu_amdkfd_gpuvm_acquire_process_vm(struct amdgpu_device *adev,
>   					   struct file *filp, u32 pasid,
>   					   void **process_info,
> @@ -1523,10 +1573,22 @@ int amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu(
>   	if (offset)
>   		*offset = amdgpu_bo_mmap_offset(bo);
>   
> +	if (flags & (KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL |
> +			KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP)) {
> +		ret = amdgpu_amdkfd_gpuvm_pin_bo(bo, AMDGPU_GEM_DOMAIN_GTT);
> +		if (ret) {
> +			pr_err("Pinning MMIO/DOORBELL BO during ALLOC FAILED\n");
> +			goto err_pin_bo;
> +		}
> +		bo->allowed_domains = AMDGPU_GEM_DOMAIN_GTT;
> +		bo->preferred_domains = AMDGPU_GEM_DOMAIN_GTT;
> +	}
> +
>   	return 0;
>   
>   allocate_init_user_pages_failed:
>   	remove_kgd_mem_from_kfd_bo_list(*mem, avm->process_info);
> +err_pin_bo:
>   	drm_vma_node_revoke(&gobj->vma_node, drm_priv);
>   err_node_allow:
>   	drm_gem_object_put(gobj);
> @@ -1559,6 +1621,14 @@ int amdgpu_amdkfd_gpuvm_free_memory_of_gpu(
>   	bool is_imported = false;
>   
>   	mutex_lock(&mem->lock);
> +
> +	/* Unpin MMIO/DOORBELL BO's that were pinnned during allocation */
> +	if (mem->alloc_flags &
> +	    (KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL |
> +	     KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP)) {
> +		amdgpu_amdkfd_gpuvm_unpin_bo(mem->bo);
> +	}
> +
>   	mapped_to_gpu_memory = mem->mapped_to_gpu_memory;
>   	is_imported = mem->is_imported;
>   	mutex_unlock(&mem->lock);