Re: [PATCH v3 16/16] block/mq-deadline: Prioritize high-priority requests

[Niklas Cassel <Niklas.Cassel@wdc.com>]

On Mon, Aug 23, 2021 at 10:15:39AM -0700, Bart Van Assche wrote:
> On 8/23/21 12:36 AM, Niklas Cassel wrote:
> > I was mainly thinking that it should be possible to do a generic fix,
> > such that we eventually won't need a similar fix as yours in all the
> > different I/O schedulers.
> 
> Coming up with a generic fix would be great but I have not yet found an
> elegant approach ...
> 
> Another question is what the impact is of scheduler bypass on zoned block
> devices? Is the zone locking performed by the mq-deadline scheduler for
> writes to zoned block devices compatible with I/O scheduler bypass?

If anyone is curious of how the stack trace looks:

#0  dd_finish_request (rq=0xffff8881051b0000) at block/mq-deadline.c:790
#1  0xffffffff81741fcf in blk_mq_free_request (rq=rq@entry=0xffff8881051b0000)
    at block/blk-mq.c:516
#2  0xffffffff8172e4fa in blk_put_request (req=req@entry=0xffff8881051b0000)
    at block/blk-core.c:644
#3  0xffffffff819c51c2 in __scsi_execute (sdev=0xffff888106064000,
    cmd=cmd@entry=0xffffc900002c78d8 "", data_direction=data_direction@entry=3,
    buffer=buffer@entry=0x0 <fixed_percpu_data>, bufflen=bufflen@entry=0,
    sense=sense@entry=0x0 <fixed_percpu_data>, sshdr=0xffffc900002c7878, timeout=30000, retries=5,
    flags=0, rq_flags=0, resid=0x0 <fixed_percpu_data>) at drivers/scsi/scsi_lib.c:260
#4  0xffffffff819e12b6 in scsi_execute_req (resid=0x0 <fixed_percpu_data>, retries=5,
    timeout=30000, sshdr=0xffffc900002c7878, bufflen=0, buffer=0x0 <fixed_percpu_data>,
    data_direction=3, cmd=0xffffc900002c78d8 "", sdev=<optimized out>)
    at ./include/scsi/scsi_device.h:463
#5  sd_spinup_disk (sdkp=<optimized out>) at drivers/scsi/sd.c:2177
#6  sd_revalidate_disk (disk=<optimized out>) at drivers/scsi/sd.c:3302
#7  0xffffffff819e479d in sd_open (bdev=0xffff888102b45800, mode=1) at drivers/scsi/sd.c:1443
#8  0xffffffff81422285 in blkdev_get_whole (bdev=bdev@entry=0xffff888102b45800, mode=mode@entry=1)
    at fs/block_dev.c:1253
#9  0xffffffff8142348a in blkdev_get_by_dev (dev=<optimized out>, mode=mode@entry=1,
    holder=holder@entry=0x0 <fixed_percpu_data>) at fs/block_dev.c:1417
#10 0xffffffff8175632c in disk_scan_partitions (disk=0xffff88810514dc00) at block/genhd.c:388
#11 register_disk (groups=<optimized out>, disk=0xffff88810514dc00, parent=0xffff888106064268)
    at block/genhd.c:435
#12 __device_add_disk (parent=parent@entry=0xffff888106064268, disk=disk@entry=0xffff88810514dc00,
    groups=groups@entry=0x0 <fixed_percpu_data>, register_queue=register_queue@entry=true)
    at block/genhd.c:527
#13 0xffffffff8175640f in device_add_disk (parent=parent@entry=0xffff888106064268,
    disk=disk@entry=0xffff88810514dc00, groups=groups@entry=0x0 <fixed_percpu_data>)
    at block/genhd.c:548
#14 0xffffffff819e4d0f in sd_probe (dev=0xffff888106064268) at drivers/scsi/sd.c:3581


This stack trace is simply the first one, it can be traced back to
sd_revalidate_disk(). All the other dd_finish_request() calls (which doesn't
have a matching insert) also originate from sd_revalidate_disk().

Like we suspected, this is because of scheduler bypass.

E.g.
sd_revalidate_disk() -> read_capacity_16() -> __scsi_execute() ->
blk_execute_rq() -> blk_execute_rq_nowait() -> blk_mq_sched_insert_request()
-> blk_mq_sched_bypass_insert() -> blk_mq_request_bypass_insert()

__scsi_execute() sets req op to REQ_OP_DRV_OUT or REQ_OP_DRV_IN.

blk_mq_sched_insert_request() bypasses the scheduler when
blk_mq_sched_bypass_insert() returns true, which it does if
blk_rq_is_passthrough().
blk_rq_is_passthrough() returns true if req op is REQ_OP_DRV_OUT
or REQ_OP_DRV_IN.

Basically __scsi_execute() is the equivalent of __nvme_submit_sync_cmd(),
but with a worse name :)


"Is the zone locking performed by the mq-deadline scheduler for
writes to zoned block devices compatible with I/O scheduler bypass?"

Since sd_revalidate_disk() doesn't do any writes, everything is fine.
(And like Damien said, if any kernel code did passthrough writes,
we would have seen errors from the drive a long time ago.)

Yes, a user submitting passthrough writes can of course do "bad things",
but that is expected :)


Kind regards,
Niklas