On Wed, 2018-06-20 at 10:09 +0800, jianchao.wang wrote: > It is very easy to reproduce with following scripts. > > script 0 > while true > do > modprobe null_blk queue_mode=2 shared_tags=1 > sleep 0.1 > rmmod null_blk > sleep 0.1 > done > > script 1 > file0="/sys/block/nullb0/mq/0/nr_tags" > file1="/sys/block/nullb0/mq/0/cpu0/rq_list" > while true; > do > if [ -e $file0 ];then > cat $file0 > fi > if [ -e $file1 ];then > cat $file1 > fi > done Hello Jianchao, Thanks for having shared a reproducer. However, the approach of the patch you posted doesn't seem like the right approach to me. I propose to proceed as follows: * Convert the reproducer into a blktests test and submit is as a patch to the blktests project (https://github.com/osandov/blktests). * Remove the mutex_lock/unlock(&sysfs_lock) calls from blk_cleanup_queue(). These calls are useless. Block drivers are required to call del_gendisk() before calling blk_cleanup_queue(). That means that sysfs attributes are removed synchronously before blk_cleanup_queue() is called. The following statement in blk_cleanup_queue() verifies that: WARN_ON_ONCE(q->kobj.state_in_sysfs); BTW, this also means that the blk_queue_dying() checks in various show and store methods are superfluous. * Introduce a new mutex to serialize blk_mq_register_dev() and blk_mq_sysfs_register() and blk_mq_sysfs_unregister(). I think it is wrong that these functions use sysfs_mutex. * Document the purpose of sysfs_mutex in include/linux/blkdev.h, namely to serialize the sysfs .show() and .store() callback functions and also to serialize elevator changes. Thanks, Bart.
