[PATCH] drm/amdkfd: Move dma unmapping after TLB flush

[Felix Kuehling]

On 2023-09-08 13:45, Philip Yang wrote:
> Otherwise GPU may access the mapping to cause IOMMU IO_PAGE_FAULT.
>
> Remove dma mapping before free the mem attachment, to fix potential dma
> mapping leaking if failed unmap from GPUs.
>
> For queues restore path, because FW already flush TLB, it is safe to
> leave this unchanged.
>
> Signed-off-by: Philip Yang <Philip.Yang at amd.com>
> ---
>   drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h     |  1 +
>   .../gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c   | 18 +++++++++++++++---
>   drivers/gpu/drm/amd/amdkfd/kfd_chardev.c       |  4 ++++
>   3 files changed, 20 insertions(+), 3 deletions(-)
>
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h
> index 559f14cc0a99..9a1f8d09b1e9 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h
> @@ -304,6 +304,7 @@ int amdgpu_amdkfd_gpuvm_map_memory_to_gpu(struct amdgpu_device *adev,
>   					  struct kgd_mem *mem, void *drm_priv);
>   int amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu(
>   		struct amdgpu_device *adev, struct kgd_mem *mem, void *drm_priv);
> +void amdgpu_amdkfd_gpuvm_dmaunmap_mem(struct kgd_mem *mem);
>   int amdgpu_amdkfd_gpuvm_sync_memory(
>   		struct amdgpu_device *adev, struct kgd_mem *mem, bool intr);
>   int amdgpu_amdkfd_gpuvm_map_gtt_bo_to_kernel(struct kgd_mem *mem,
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c
> index b5b940485059..3c566050bed8 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c
> @@ -1249,8 +1249,6 @@ static void unmap_bo_from_gpuvm(struct kgd_mem *mem,
>   	amdgpu_vm_clear_freed(adev, vm, &bo_va->last_pt_update);
>   
>   	amdgpu_sync_fence(sync, bo_va->last_pt_update);
> -
> -	kfd_mem_dmaunmap_attachment(mem, entry);

This probably won't work correctly when the same BO is unmapped and then 
mapped again to the same GPU.

Regards,
   Felix


>   }
>   
>   static int update_gpuvm_pte(struct kgd_mem *mem,
> @@ -1910,8 +1908,10 @@ int amdgpu_amdkfd_gpuvm_free_memory_of_gpu(
>   		mem->va + bo_size * (1 + mem->aql_queue));
>   
>   	/* Remove from VM internal data structures */
> -	list_for_each_entry_safe(entry, tmp, &mem->attachments, list)
> +	list_for_each_entry_safe(entry, tmp, &mem->attachments, list) {
> +		kfd_mem_dmaunmap_attachment(mem, entry);
>   		kfd_mem_detach(entry);
> +	}
>   
>   	ret = unreserve_bo_and_vms(&ctx, false, false);
>   
> @@ -2085,6 +2085,18 @@ int amdgpu_amdkfd_gpuvm_map_memory_to_gpu(
>   	return ret;
>   }
>   
> +void amdgpu_amdkfd_gpuvm_dmaunmap_mem(struct kgd_mem *mem)
> +{
> +	struct kfd_mem_attachment *entry;
> +
> +	mutex_lock(&mem->lock);
> +
> +	list_for_each_entry(entry, &mem->attachments, list)
> +		kfd_mem_dmaunmap_attachment(mem, entry);
> +
> +	mutex_unlock(&mem->lock);
> +}
> +
>   int amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu(
>   		struct amdgpu_device *adev, struct kgd_mem *mem, void *drm_priv)
>   {
> diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c
> index 08687ce0aa8b..fdcc55e324ea 100644
> --- a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c
> +++ b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c
> @@ -1443,6 +1443,10 @@ static int kfd_ioctl_unmap_memory_from_gpu(struct file *filep,
>   			kfd_flush_tlb(peer_pdd, TLB_FLUSH_HEAVYWEIGHT);
>   		}
>   	}
> +
> +	/* Remove IOMMU mapping after tlb flush to avoid IO_PAGE_FAULT race */
> +	amdgpu_amdkfd_gpuvm_dmaunmap_mem(mem);
> +
>   	kfree(devices_arr);
>   
>   	return 0;