Re: [PATCH] ublk: don't use an I/O scheduler by default

[Ming Lei]

On Thu, Oct 01, 2026 at 11:10:40AM +0800, Qiliang Yuan wrote:
> Requests of a ublk device are handed to the ublk server through
> io_uring, and the server does its own queueing and ordering, so an I/O
> scheduler in front of the device only adds overhead. A single-queue
> ublk device still gets mq-deadline by default, which costs throughput
> on every I/O and an elevator setup on every START_DEV.
> 
> Set BLK_MQ_F_NO_SCHED_BY_DEFAULT so that add_disk() selects "none", as
> loop does since commit 2112f5c1330a ("loop: Select I/O scheduler 'none'
> from inside add_disk()"). Doing it in the kernel also avoids switching
> the scheduler from user space after the device has been added, which
> waits for RCU grace periods. A scheduler can still be selected through
> sysfs. Zoned ublk devices don't need mq-deadline either, zone write
> plugging serializes writes per zone in the block layer since
> commit fde02699c242 ("block: mq-deadline: Remove support for zone write
> locking").
> 
> ublk null target, single-queue devices with depth 16, fio io_uring 4k
> randread at iodepth=16, 16 vCPU KVM guest:
> 
>                                  mq-deadline       none
>   1 device                          111K IOPS   342K IOPS
>   4 devices                         286K IOPS  1070K IOPS
>   START_DEV p50 (1000 devices)        8.53ms      0.36ms
> 
> Signed-off-by: Qiliang Yuan <odys.yuan@xxxxxxxxx>

Default none does make sense:

Reviewed-by: Ming Lei <tom.leiming@xxxxxxxxx>

> ---
> Zoned ublk devices were not tested at runtime: none of the ublk servers
> at hand (ublksrv, the kublk selftest server) serves zoned I/O.

rublk supports zoned:

`cargo install rublk` or https://github.com/ublk-org/rublk


Thanks,
Ming