Need a pair decrement for fence's refcount if ttm_bo_add_move_fence failed?

Koenig, Christian Christian.Koenig at amd.com
Mon Apr 8 07:34:53 UTC 2019


Am 07.04.19 um 13:44 schrieb 易林:
> Hi, all:
>      when analyzing v5.1 source code, I notice that in ttm_bo_add_move_fence,
> when reservation_object_reserve_shared failed and return ENOMEM,
> the fence's refcount increased without a pair decrement even after return to ttm_bo_add_move_fence's caller ttm_bo_mem_force_space:
>
> static int ttm_bo_add_move_fence(struct ttm_buffer_object *bo,
> 				 struct ttm_mem_type_manager *man,
> 				 struct ttm_mem_reg *mem)
> {
>          ......
> 	fence = dma_fence_get(man->move);
> 	spin_unlock(&man->move_lock);
>
> 	if (fence) {
> 		reservation_object_add_shared_fence(bo->resv, fence);
>
> 		ret = reservation_object_reserve_shared(bo->resv, 1);
> 		if (unlikely(ret))
> 			return ret;
>
> 		dma_fence_put(bo->moving);
> 		bo->moving = fence;
> 	}
>
> 	return 0;
> }
>
> can this lead to the imbalance of the fence's refcount? though the ENOMEN almost won't be trigger.

Yeah, the fence is leaked in the error path. Feel free to provide a 
patch to fix this.

Otherwise I will provide one in the next merge window.

Thanks,
Christian.

>
> Best Regards
>
> Lin Yi



More information about the dri-devel mailing list