Re: [PATCH] the dm-loop target

[Ming Lei <ming.lei@redhat.com>]

On Fri, Mar 07, 2025 at 04:21:58PM +0100, Mikulas Patocka wrote:
> > I didn't say you were. I said the concept that dm-loop is based on
> > is fundamentally flawed and that your benchmark setup does not
> > reflect real world usage of loop devices.
> 
> > Where are the bug reports about the loop device being slow and the
> > analysis that indicates that it is unfixable?
> 
> So, I did benchmarks on an enterprise nvme drive (SAMSUNG 
> MZPLJ1T6HBJR-00007). I stacked ext4/loop/ext4, xfs/loop/xfs (using losetup 
> --direct-io=on), ext4/dm-loop/ext4 and xfs/dm-loop/xfs. And loop is slow.
> 
> synchronous I/O:
> fio --direct=1 --bs=4k --runtime=10 --time_based --numjobs=12 --ioengine=psync --iodepth=1 --group_reporting=1 --filename=/mnt/test2/l -name=job --rw=rw
> raw block device:
>    READ: bw=399MiB/s (418MB/s), 399MiB/s-399MiB/s (418MB/s-418MB/s), io=3985MiB (4179MB), run=10001-10001msec
>   WRITE: bw=399MiB/s (418MB/s), 399MiB/s-399MiB/s (418MB/s-418MB/s), io=3990MiB (4184MB), run=10001-10001msec
> ext4/loop/ext4:
>    READ: bw=223MiB/s (234MB/s), 223MiB/s-223MiB/s (234MB/s-234MB/s), io=2232MiB (2341MB), run=10002-10002msec
>   WRITE: bw=223MiB/s (234MB/s), 223MiB/s-223MiB/s (234MB/s-234MB/s), io=2231MiB (2339MB), run=10002-10002msec
> xfs/loop/xfs:
>    READ: bw=220MiB/s (230MB/s), 220MiB/s-220MiB/s (230MB/s-230MB/s), io=2196MiB (2303MB), run=10001-10001msec
>   WRITE: bw=219MiB/s (230MB/s), 219MiB/s-219MiB/s (230MB/s-230MB/s), io=2193MiB (2300MB), run=10001-10001msec
> ext4/dm-loop/ext4:
>    READ: bw=338MiB/s (355MB/s), 338MiB/s-338MiB/s (355MB/s-355MB/s), io=3383MiB (3547MB), run=10002-10002msec
>   WRITE: bw=338MiB/s (355MB/s), 338MiB/s-338MiB/s (355MB/s-355MB/s), io=3385MiB (3549MB), run=10002-10002msec
> xfs/dm-loop/xfs:
>    READ: bw=375MiB/s (393MB/s), 375MiB/s-375MiB/s (393MB/s-393MB/s), io=3752MiB (3934MB), run=10002-10002msec
>   WRITE: bw=376MiB/s (394MB/s), 376MiB/s-376MiB/s (394MB/s-394MB/s), io=3756MiB (3938MB), run=10002-10002msec
> 
> asynchronous I/O:
> fio --direct=1 --bs=4k --runtime=10 --time_based --numjobs=12 --ioengine=libaio --iodepth=16 --group_reporting=1 --filename=/mnt/test2/l -name=job --rw=rw
> raw block device:
>    READ: bw=1246MiB/s (1306MB/s), 1246MiB/s-1246MiB/s (1306MB/s-1306MB/s), io=12.2GiB (13.1GB), run=10001-10001msec
>   WRITE: bw=1247MiB/s (1308MB/s), 1247MiB/s-1247MiB/s (1308MB/s-1308MB/s), io=12.2GiB (13.1GB), run=10001-10001msec
> ext4/loop/ext4:
>    READ: bw=274MiB/s (288MB/s), 274MiB/s-274MiB/s (288MB/s-288MB/s), io=2743MiB (2877MB), run=10001-10001msec
>   WRITE: bw=275MiB/s (288MB/s), 275MiB/s-275MiB/s (288MB/s-288MB/s), io=2747MiB (2880MB), run=10001-10001msec
> xfs/loop/xfs:
>    READ: bw=276MiB/s (289MB/s), 276MiB/s-276MiB/s (289MB/s-289MB/s), io=2761MiB (2896MB), run=10002-10002msec
>   WRITE: bw=276MiB/s (290MB/s), 276MiB/s-276MiB/s (290MB/s-290MB/s), io=2765MiB (2899MB), run=10002-10002msec
> ext4/dm-loop/ext4:
>    READ: bw=1189MiB/s (1247MB/s), 1189MiB/s-1189MiB/s (1247MB/s-1247MB/s), io=11.6GiB (12.5GB), run=10002-10002msec
>   WRITE: bw=1190MiB/s (1248MB/s), 1190MiB/s-1190MiB/s (1248MB/s-1248MB/s), io=11.6GiB (12.5GB), run=10002-10002msec
> xfs/dm-loop/xfs:
>    READ: bw=1209MiB/s (1268MB/s), 1209MiB/s-1209MiB/s (1268MB/s-1268MB/s), io=11.8GiB (12.7GB), run=10001-10001msec
>   WRITE: bw=1210MiB/s (1269MB/s), 1210MiB/s-1210MiB/s (1269MB/s-1269MB/s), io=11.8GiB (12.7GB), run=10001-10001msec

Hi Mikulas,

Please try the following patchset:

https://lore.kernel.org/linux-block/20250308162312.1640828-1-ming.lei@redhat.com/

which tries to handle IO in current context directly via NOWAIT, and
supports MQ for loop since 12 io jobs are applied in your test. With this
change, I can observe similar perf data on raw block device and loop/xfs
over mq-virtio-scsi & nvme in my test VM.

1) try single queue first by `modprobe loop`

2) then try MQ by 'modprobe loop nr_hw_queues=4'

If it still doesn't work, please provide fio log for both `raw block
device` and 'loop/xfs', which may provide some clue for the big perf
gap.



Thanks,
Ming