On Wed, 2017-09-27 at 06:23 +0800, Ming Lei wrote: > mutex_lock(&sdev->state_mutex); > err = scsi_device_set_state(sdev, SDEV_QUIESCE); > if (err == 0) > blk_set_preempt_only(q, true); > mutex_unlock(&sdev->state_mutex); > > if (err) > return err; > > blk_mq_freeze_queue(q); > blk_mq_unfreeze_queue(q); > > Any requests allocated before scsi_device_set_state() and > dispatched after scsi_device_set_state() can't be drained up > any more, then blk_mq_freeze_queue() will wait forever for the > drain up, so not only this patchset can't fix the current quiesce > issue, but also introduces new I/O hang in scsi_device_quiesce().
That's a good catch, but fortunately this is very easy to fix: move the blk_mq_freeze_queue() call before the mutex_lock() and scsi_device_set_state() calls. The error path will also need some adjustment. Bart.
