On Wed, May 17, 2017 at 08:38:01AM +0200, Johannes Thumshirn wrote:

On 05/17/2017 03:27 AM, Ming Lei wrote:

quoted

When one NVMe PCI device is being resetted and found reset failue,
nvme_remove_dead_ctrl() is called to handle the failure: blk-mq hw queues
are put into stopped first, then schedule .remove_work to release the driver.

Unfortunately if the driver is being released via sysfs store
just before the .remove_work is run, del_gendisk() from
nvme_remove() may hang forever because hw queues are stopped and
the submitted writeback IOs from fsync_bdev() can't be completed at all.

This patch fixes the following issue[1][2] by moving nvme_kill_queues()
into nvme_remove_dead_ctrl() to avoid the issue because nvme_remove()
flushs .reset_work, and this way is reasonable and safe because
nvme_dev_disable() has started to suspend queues and canceled requests
already.

[1] test script
	fio -filename=$NVME_DISK -iodepth=1 -thread -rw=randwrite -ioengine=psync \
	    -bssplit=5k/10:9k/10:13k/10:17k/10:21k/10:25k/10:29k/10:33k/10:37k/10:41k/10 \
	    -bs_unaligned -runtime=1200 -size=-group_reporting -name=mytest -numjobs=60

Nit: the actual size after the -size parameter is missing.

Forget to mention, $NVME_DISK has to be one partition of nvme disk, and
the type need to be 'filesystem' type for reproduction, then actual size
isn't needed.

Anyways:
Reviewed-by: Johannes Thumshirn <redacted>

Thanks for review!


Thanks,
Ming