On 2021-02-09 10:08 p.m., Kashyap Desai wrote:
On Wed, Feb 10, 2021 at 4:26 AM Douglas Gilbert <dgilbert@xxxxxxxxxxxx> wrote:
Add a new sdeb_defer_type enumeration: SDEB_DEFER_POLL for requests
that have REQ_HIPRI set in cmd_flags field. It is expected that
these requests will be polled via the mq_poll entry point which
is driven by calls to blk_poll() in the block layer. Therefore
timer events are not 'wired up' in the normal fashion.
There are still cases with short delays (e.g. < 10 microseconds)
where by the time the command response processing occurs, the delay
is already exceeded in which case the code calls scsi_done()
directly. In such cases there is no window for mq_poll() to be
called.
Add 'mq_polls' counter that increments on each scsi_done() called
via the mq_poll entry point. Can be used to show (with 'cat
/proc/scsi/scsi_debug/<host_id>') that blk_poll() is causing
completions rather than some other mechanism.
This patch is against the 5.12/scsi-staging branch which includes
a98d6bdf181eb71bf4686666eaf47c642a061642
scsi_debug : iouring iopoll support
which it alters. So this patch is a fix of that patch.
Changes since version 2 [sent 20210206 to linux-scsi list]
- the sdebug_blk_mq_poll() callback didn't cope with the
uncommon case where sqcp->sd_dp is NULL. Fix.
Hi Doug, I tried this patch on top of below iouring patch series -
"[v3,1/4] add io_uring with IOPOLL support in scsi layer"
After applying patch, I am seeing an IO hang issue - I see
io_wqe_worker goes into a tight loop.
18210 root 20 0 1316348 3552 2376 R 99.1 0.0 0:24.09
fio 18303 root 20 0
0 0 0 D 78.8 0.0 0:01.75 io_wqe_worker-0
18219 root 20 0 0 0 0 R 71.7 0.0
0:06.59 io_wqe_worker-0
I used below command -
insmod drivers/scsi/scsi_debug.ko dev_size_mb=1024 sector_size=512
add_host=24 per_host_store=1 ndelay=10000 host_max_queue=32
submit_queues=76 num_parts=1 poll_queues=8
and here is my fio script -
[global]
ioengine=io_uring
hipri=1
direct=1
runtime=30s
rw=randread
norandommap
bs=4k
iodepth=64
Kashyap,
There is another issue here, namely iodepth > host_max_queue (64 > 32)
and in my setup that is not handled well. So there is a problem with
scsi_debug *** or the mid-level in handling this case.
If you have modified the sd driver to call blk_poll() then perhaps
you could try the above test again with a reduced iodepth.
Doug Gilbert
*** the scsi_debug should either yield a TASK SET FULL status or return
SCSI_MLQUEUE_HOST_BUSY from queuecommand() when it has run out of
slots.
[seqprecon]
filename=/dev/sdd
[seqprecon]
filename=/dev/sde
[seqprecon]
filename=/dev/sdf
I will ask Martin to pick all the patches from "[v3,1/4] add io_uring
with IOPOLL support in the scsi layer" except scsi_debug. We can work
on scsi_debug and send standalone patch.
Kashyap
