On 11/23/24 12:08 PM, Bart Van Assche wrote:
> Hi Damien,
>
> If I run the following shell commands:
>
> modprobe -r scsi_debug
> modprobe scsi_debug delay=0 dev_size_mb=256 every_nth=2 max_queue=1 \
> opts=0x8000 sector_size=4096 zbc=host-managed zone_nr_conv=0 zone_size_mb=4
> while true; do
> bdev=$(cd /sys/bus/pseudo/drivers/scsi_debug/adapter*/host*/target*/*/block
> && echo *) 2>/dev/null
> if [ -e /dev/"${bdev}" ]; then break; fi
> sleep .1
> done
Running this script, I never get out of this loop... Weird. I changed the
script to this:
...
while true; do
bdev="$(lsscsi -g | grep scsi_debug | cut -d'/' -f3)"
if [ ! -z "${bdev}" ]; then
break;
fi
sleep .1
done
dev="/dev/${bdev}"
[ -b "${dev}" ]
...
and then it runs, but I do not see anything blocked. All is fine. Could you
share your kernel config to see what I am missing may be ?
> dev=/dev/"${bdev}"
> [ -b "${dev}" ]
> fio --direct=1 --filename=$dev --iodepth=1 --ioengine=io_uring \
> --ioscheduler=none --gtod_reduce=1 --hipri=0 --name="$(basename "${dev}")" \
> --runtime=30 --rw=rw --time_based=1 --zonemode=zbd &
> sleep 2
> echo w > /proc/sysrq-trigger
>
> then the following appears in the kernel log:
>
> sysrq: Show Blocked State
> task:(udev-worker) state:D stack:0 pid:3121 tgid:3121 ppid:2191 fl
> ags:0x00000006
> Call Trace:
> <TASK>
> __schedule+0x3e8/0x1410
> schedule+0x27/0xf0
> blk_mq_freeze_queue_wait+0x6f/0xa0
> queue_attr_store+0x60/0xc0
> kernfs_fop_write_iter+0x13e/0x1f0
> vfs_write+0x25b/0x420
> ksys_write+0x65/0xe0
> do_syscall_64+0x82/0x160
> entry_SYSCALL_64_after_hwframe+0x76/0x7e
>
> Do you agree that the above indicates that blk_mq_freeze_queue_wait()
> hangs? I think it is waiting for a q_usage_counter reference that is
> held by a bio on a zwplug->bio_list.
>
> Do you agree that the best way to solve this is to modify
> blk_mq_submit_bio() by moving the blk_zone_plug_bio() call after the
> blk_crypto_rq_get_keyslot() call and also to change the zwplug bio list
> into a request list?
I am not following... blk_zone_plug_bio() acts on BIOs, not on requests, but
blk_crypto_rq_get_keyslot() is a request operation. Also, a requeue list is a
list of requests, not BIOs. One of the main point of zone write plugs is to
operate on BIOs so that:
1) we have code that is common for regular block devices and for BIO based
device-mapper devices.
2) Blocking writes to preserve per-zone write ordering does not result in
device resources (TAGS !) being held for nothing and potentially
delaying/starving read accesses or other operations that do not need ordering.
How can a requeue list help ? Please provide details.
--
Damien Le Moal
Western Digital Research
