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

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

[Qiliang Yuan]

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@gmail.com>
---
Zoned ublk devices were not tested at runtime: none of the ublk servers
at hand (ublksrv, the kublk selftest server) serves zoned I/O.
---
 drivers/block/ublk_drv.c | 1 +
 1 file changed, 1 insertion(+)

diff --git a/drivers/block/ublk_drv.c b/drivers/block/ublk_drv.c
index 66eb55e7162e5..8053666c04482 100644
--- a/drivers/block/ublk_drv.c
+++ b/drivers/block/ublk_drv.c
@@ -4389,6 +4389,7 @@ static int ublk_add_tag_set(struct ublk_device *ub)
 	ub->tag_set.nr_hw_queues = ub->dev_info.nr_hw_queues;
 	ub->tag_set.queue_depth = ub->dev_info.queue_depth;
 	ub->tag_set.numa_node = NUMA_NO_NODE;
+	ub->tag_set.flags = BLK_MQ_F_NO_SCHED_BY_DEFAULT;
 	ub->tag_set.driver_data = ub;
 	return blk_mq_alloc_tag_set(&ub->tag_set);
 }

---
base-commit: 551c722f40809618230001baccf219193e22fc5a
change-id: 20261001-bug-ublk-no-sched-by-default-6f53fa4889e6

Best regards,
-- 
Qiliang Yuan <odys.yuan@gmail.com>

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@gmail.com>

Default none does make sense:

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

> ---
> 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