On 3/6/20 7:35 AM, Dan Carpenter wrote: > > There a bunch of similar bugs. It's seems a common anti-pattern. > > block/blk-cgroup.c:85 blkg_free() warn: freeing 'blkg' which has percpu_ref_exit() > block/blk-core.c:558 blk_alloc_queue_node() warn: freeing 'q' which has percpu_ref_exit() > drivers/md/md.c:5528 md_free() warn: freeing 'mddev' which has percpu_ref_exit() > drivers/target/target_core_transport.c:583 transport_free_session() warn: freeing 'se_sess' which has percpu_ref_exit() > fs/aio.c:592 free_ioctx() warn: freeing 'ctx' which has percpu_ref_exit() > fs/aio.c:806 ioctx_alloc() warn: freeing 'ctx' which has percpu_ref_exit() > fs/io_uring.c:6115 io_sqe_files_unregister() warn: freeing 'data' which has percpu_ref_exit() > fs/io_uring.c:6431 io_sqe_files_register() warn: freeing 'ctx->file_data' which has percpu_ref_exit() > fs/io_uring.c:7134 io_ring_ctx_free() warn: freeing 'ctx' which has percpu_ref_exit() > kernel/cgroup/cgroup.c:4948 css_free_rwork_fn() warn: freeing 'css' which has percpu_ref_exit() > mm/backing-dev.c:615 cgwb_create() warn: freeing 'wb' which has percpu_ref_exit() The file table io_uring issue is using the ref in a funky way, switching in and out of atomic if we need to quiesce it. That's different from other use cases, that just use it as a "normal" reference. Hence for the funky use case, you can potentially have a switch in progress when you exit the ref. You really want to wait for that, the easiest solution is to punt the exit + free to an RCU callback, if there's nothing else you need to handle once the switch is done. So I would not be so quick to assume that similar patterns (exit + free) have similar issues. -- Jens Axboe
