On Thu, Mar 21, 2019 at 08:17:44PM -0700, Bart Van Assche wrote: > On 3/21/19 7:11 PM, Ming Lei wrote: > > On Fri, Mar 22, 2019 at 09:45:54AM +0800, zhengbin wrote: > > > + /* If device state changes to SDEV_RUNNING, we need to run hw queue > > > + * to avoid io hung. > > > + */ > > > + if (ret == 0 && state == SDEV_RUNNING) > > > + blk_mq_run_hw_queues(sdev->request_queue, true); > > > > > > > Thinking of further, this way isn't safe, given the caller should make > > sure that the queue is alive before calling blk_mq_run_hw_queues(). > > > > Now it is in sysfs write path, and the scsi device can become gone > > just before calling blk_mq_run_hw_queues(). > > Hi Ming, > > What makes you think that this approach is unsafe? I think that > __scsi_remove_device() waits until all sysfs callbacks have finished before > it calls blk_cleanup_queue(). You mean sysfs_remove_groups() will wait for completion of the current store? If sysfs would provide such guarantee, it is safe. Looks too tricky even though it is true, so I still suggest to run queue with holding the mutex of state_mutex, given this usage has been done by __scsi_remove_device() already. Thanks, Ming
