[PATCH 1/1] drm/amdgpu: cleanup pasid handling

Christian König christian.koenig at amd.com
Wed Jun 23 08:23:58 UTC 2021


Am 23.06.21 um 09:56 schrieb Nirmoy Das:
> Cleanup code related to vm pasid by adding helper functions.
> Also replace idr with xarray as we actually need hash functionality.

That looks quite a bit better than before, but I think we should 
approach it differently.

First of all make a patch which moves amdgpu_pasid_free() outside of the 
VM code.

We don't allocate the pasid inside the VM code for good reasons so we 
shouldn't free it either.

Then in a second patch make a function amdgpu_vm_set_pasid(adev, vm, pasid);

When the pasid is zero we remove the VM from the xarray, otherwise we 
update the entry.

Thanks,
Christian.

>
> Signed-off-by: Nirmoy Das <nirmoy.das at amd.com>
> ---
>   drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 130 ++++++++++++-------------
>   drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h |   3 +-
>   2 files changed, 62 insertions(+), 71 deletions(-)
>
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
> index 63975bda8e76..abba1e2de264 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
> @@ -87,6 +87,45 @@ struct amdgpu_prt_cb {
>   	struct dma_fence_cb cb;
>   };
>   
> +static int amdgpu_vm_pasid_insert(struct amdgpu_device *adev,
> +				  struct amdgpu_vm *vm,
> +				  unsigned long pasid,
> +				  unsigned int *vm_pasid)
> +{
> +	unsigned long flags;
> +	int r;
> +
> +	if (!pasid)
> +		return 0;
> +
> +	xa_lock_irqsave(&adev->vm_manager.pasids, flags);
> +	r = xa_err(__xa_store(&adev->vm_manager.pasids, pasid, vm, GFP_ATOMIC));
> +	xa_unlock_irqrestore(&adev->vm_manager.pasids, flags);
> +	if (r < 0)
> +		return r;
> +	if (vm_pasid)
> +		*vm_pasid = pasid;
> +
> +	return 0;
> +}
> +
> +static void amdgpu_vm_pasid_remove(struct amdgpu_device *adev,
> +				   unsigned long pasid,
> +				   unsigned int *vm_pasid)
> +{
> +	unsigned long flags;
> +
> +	if (!pasid)
> +		return;
> +
> +	xa_lock_irqsave(&adev->vm_manager.pasids, flags);
> +	__xa_erase(&adev->vm_manager.pasids, pasid);
> +	xa_unlock_irqrestore(&adev->vm_manager.pasids, flags);
> +
> +	if (vm_pasid)
> +		*vm_pasid = 0;
> +}
> +
>   /*
>    * vm eviction_lock can be taken in MMU notifiers. Make sure no reclaim-FS
>    * happens while holding this lock anywhere to prevent deadlocks when
> @@ -2940,18 +2979,9 @@ int amdgpu_vm_init(struct amdgpu_device *adev, struct amdgpu_vm *vm, u32 pasid)
>   
>   	amdgpu_bo_unreserve(vm->root.bo);
>   
> -	if (pasid) {
> -		unsigned long flags;
> -
> -		spin_lock_irqsave(&adev->vm_manager.pasid_lock, flags);
> -		r = idr_alloc(&adev->vm_manager.pasid_idr, vm, pasid, pasid + 1,
> -			      GFP_ATOMIC);
> -		spin_unlock_irqrestore(&adev->vm_manager.pasid_lock, flags);
> -		if (r < 0)
> -			goto error_free_root;
> -
> -		vm->pasid = pasid;
> -	}
> +	r = amdgpu_vm_pasid_insert(adev, vm, pasid, &vm->pasid);
> +	if (r)
> +		goto error_free_root;
>   
>   	INIT_KFIFO(vm->faults);
>   
> @@ -3038,19 +3068,9 @@ int amdgpu_vm_make_compute(struct amdgpu_device *adev, struct amdgpu_vm *vm,
>   	r = amdgpu_vm_check_clean_reserved(adev, vm);
>   	if (r)
>   		goto unreserve_bo;
> -
> -	if (pasid) {
> -		unsigned long flags;
> -
> -		spin_lock_irqsave(&adev->vm_manager.pasid_lock, flags);
> -		r = idr_alloc(&adev->vm_manager.pasid_idr, vm, pasid, pasid + 1,
> -			      GFP_ATOMIC);
> -		spin_unlock_irqrestore(&adev->vm_manager.pasid_lock, flags);
> -
> -		if (r == -ENOSPC)
> -			goto unreserve_bo;
> -		r = 0;
> -	}
> +	r = amdgpu_vm_pasid_insert(adev, vm, pasid, NULL);
> +	if (r)
> +		goto unreserve_bo;
>   
>   	/* Check if PD needs to be reinitialized and do it before
>   	 * changing any other state, in case it fails.
> @@ -3089,35 +3109,23 @@ int amdgpu_vm_make_compute(struct amdgpu_device *adev, struct amdgpu_vm *vm,
>   	vm->is_compute_context = true;
>   
>   	if (vm->pasid) {
> -		unsigned long flags;
> -
> -		spin_lock_irqsave(&adev->vm_manager.pasid_lock, flags);
> -		idr_remove(&adev->vm_manager.pasid_idr, vm->pasid);
> -		spin_unlock_irqrestore(&adev->vm_manager.pasid_lock, flags);
> -
>   		/* Free the original amdgpu allocated pasid
>   		 * Will be replaced with kfd allocated pasid
>   		 */
>   		amdgpu_pasid_free(vm->pasid);
> -		vm->pasid = 0;
> +		amdgpu_vm_pasid_remove(adev, vm->pasid, &vm->pasid);
>   	}
>   
>   	/* Free the shadow bo for compute VM */
>   	amdgpu_bo_unref(&to_amdgpu_bo_vm(vm->root.bo)->shadow);
> -
>   	if (pasid)
>   		vm->pasid = pasid;
>   
>   	goto unreserve_bo;
>   
>   free_idr:
> -	if (pasid) {
> -		unsigned long flags;
> +	amdgpu_vm_pasid_remove(adev, pasid, NULL);
>   
> -		spin_lock_irqsave(&adev->vm_manager.pasid_lock, flags);
> -		idr_remove(&adev->vm_manager.pasid_idr, pasid);
> -		spin_unlock_irqrestore(&adev->vm_manager.pasid_lock, flags);
> -	}
>   unreserve_bo:
>   	amdgpu_bo_unreserve(vm->root.bo);
>   	return r;
> @@ -3133,14 +3141,7 @@ int amdgpu_vm_make_compute(struct amdgpu_device *adev, struct amdgpu_vm *vm,
>    */
>   void amdgpu_vm_release_compute(struct amdgpu_device *adev, struct amdgpu_vm *vm)
>   {
> -	if (vm->pasid) {
> -		unsigned long flags;
> -
> -		spin_lock_irqsave(&adev->vm_manager.pasid_lock, flags);
> -		idr_remove(&adev->vm_manager.pasid_idr, vm->pasid);
> -		spin_unlock_irqrestore(&adev->vm_manager.pasid_lock, flags);
> -	}
> -	vm->pasid = 0;
> +	amdgpu_vm_pasid_remove(adev, vm->pasid, &vm->pasid);
>   	vm->is_compute_context = false;
>   }
>   
> @@ -3164,15 +3165,7 @@ void amdgpu_vm_fini(struct amdgpu_device *adev, struct amdgpu_vm *vm)
>   
>   	root = amdgpu_bo_ref(vm->root.bo);
>   	amdgpu_bo_reserve(root, true);
> -	if (vm->pasid) {
> -		unsigned long flags;
> -
> -		spin_lock_irqsave(&adev->vm_manager.pasid_lock, flags);
> -		idr_remove(&adev->vm_manager.pasid_idr, vm->pasid);
> -		spin_unlock_irqrestore(&adev->vm_manager.pasid_lock, flags);
> -		vm->pasid = 0;
> -	}
> -
> +	amdgpu_vm_pasid_remove(adev, vm->pasid, &vm->pasid);
>   	dma_fence_wait(vm->last_unlocked, false);
>   	dma_fence_put(vm->last_unlocked);
>   
> @@ -3254,8 +3247,7 @@ void amdgpu_vm_manager_init(struct amdgpu_device *adev)
>   	adev->vm_manager.vm_update_mode = 0;
>   #endif
>   
> -	idr_init(&adev->vm_manager.pasid_idr);
> -	spin_lock_init(&adev->vm_manager.pasid_lock);
> +	xa_init_flags(&adev->vm_manager.pasids, XA_FLAGS_LOCK_IRQ);
>   }
>   
>   /**
> @@ -3267,8 +3259,8 @@ void amdgpu_vm_manager_init(struct amdgpu_device *adev)
>    */
>   void amdgpu_vm_manager_fini(struct amdgpu_device *adev)
>   {
> -	WARN_ON(!idr_is_empty(&adev->vm_manager.pasid_idr));
> -	idr_destroy(&adev->vm_manager.pasid_idr);
> +	WARN_ON(!xa_empty(&adev->vm_manager.pasids));
> +	xa_destroy(&adev->vm_manager.pasids);
>   
>   	amdgpu_vmid_mgr_fini(adev);
>   }
> @@ -3337,13 +3329,13 @@ void amdgpu_vm_get_task_info(struct amdgpu_device *adev, u32 pasid,
>   	struct amdgpu_vm *vm;
>   	unsigned long flags;
>   
> -	spin_lock_irqsave(&adev->vm_manager.pasid_lock, flags);
> +	xa_lock_irqsave(&adev->vm_manager.pasids, flags);
>   
> -	vm = idr_find(&adev->vm_manager.pasid_idr, pasid);
> +	vm = xa_load(&adev->vm_manager.pasids, pasid);
>   	if (vm)
>   		*task_info = vm->task_info;
>   
> -	spin_unlock_irqrestore(&adev->vm_manager.pasid_lock, flags);
> +	xa_unlock_irqrestore(&adev->vm_manager.pasids, flags);
>   }
>   
>   /**
> @@ -3385,15 +3377,15 @@ bool amdgpu_vm_handle_fault(struct amdgpu_device *adev, u32 pasid,
>   	struct amdgpu_vm *vm;
>   	int r;
>   
> -	spin_lock_irqsave(&adev->vm_manager.pasid_lock, irqflags);
> -	vm = idr_find(&adev->vm_manager.pasid_idr, pasid);
> +	xa_lock_irqsave(&adev->vm_manager.pasids, irqflags);
> +	vm = xa_load(&adev->vm_manager.pasids, pasid);
>   	if (vm) {
>   		root = amdgpu_bo_ref(vm->root.bo);
>   		is_compute_context = vm->is_compute_context;
>   	} else {
>   		root = NULL;
>   	}
> -	spin_unlock_irqrestore(&adev->vm_manager.pasid_lock, irqflags);
> +	xa_unlock_irqrestore(&adev->vm_manager.pasids, irqflags);
>   
>   	if (!root)
>   		return false;
> @@ -3411,11 +3403,11 @@ bool amdgpu_vm_handle_fault(struct amdgpu_device *adev, u32 pasid,
>   		goto error_unref;
>   
>   	/* Double check that the VM still exists */
> -	spin_lock_irqsave(&adev->vm_manager.pasid_lock, irqflags);
> -	vm = idr_find(&adev->vm_manager.pasid_idr, pasid);
> +	xa_lock_irqsave(&adev->vm_manager.pasids, irqflags);
> +	vm = xa_load(&adev->vm_manager.pasids, pasid);
>   	if (vm && vm->root.bo != root)
>   		vm = NULL;
> -	spin_unlock_irqrestore(&adev->vm_manager.pasid_lock, irqflags);
> +	xa_unlock_irqrestore(&adev->vm_manager.pasids, irqflags);
>   	if (!vm)
>   		goto error_unlock;
>   
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h
> index ddb85a85cbba..31c467764162 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h
> @@ -359,8 +359,7 @@ struct amdgpu_vm_manager {
>   	/* PASID to VM mapping, will be used in interrupt context to
>   	 * look up VM of a page fault
>   	 */
> -	struct idr				pasid_idr;
> -	spinlock_t				pasid_lock;
> +	struct xarray				pasids;
>   };
>   
>   struct amdgpu_bo_va_mapping;



More information about the amd-gfx mailing list