Re: [PATCH] the dm-loop target

[Mikulas Patocka <mpatocka@redhat.com>]

On Mon, 3 Mar 2025, Christoph Hellwig wrote:

> On Mon, Mar 03, 2025 at 11:24:27AM +0100, Mikulas Patocka wrote:
> > This is the dm-loop target - a replacement for the regular loop driver 
> > with better performance. The dm-loop target builds a map of the file in 
> > the constructor and it just remaps bios according to this map.
> 
> Using ->bmap is broken and a no-go for new code.

What should I use instead of bmap? Is fiemap exported for use in the 
kernel?

ext4_bmap flushes the journal if journaled data are used. Is there some 
equivalent function that would provide the same guarantee w.r.t. journaled 
data?

> If you have any real
> performance issues with the loop driver document and report them so that
> they can be fixed instead of working around them by duplicating the code
> (and in this case making the new code completely broken).

Would Jens Axboe agree to merge the dm-loop logic into the existing loop 
driver?


Dm-loop is significantly faster than the regular loop:

# modprobe brd rd_size=1048576
# dd if=/dev/zero of=/dev/ram0 bs=1048576
# mkfs.ext4 /dev/ram0
# mount -t ext4 /dev/ram0 /mnt/test
# dd if=/dev/zero of=/mnt/test/test bs=1048576 count=512

dm-loop (on /mnt/test/test):
# dmsetup create loop --table '0 1048576 loop /mnt/test/test 0'
# fio --direct=1 --bs=4k --runtime=10 --time_based --numjobs=16 
--ioengine=psync --iodepth=1 --group_reporting=1 
--filename=/dev/mapper/loop --name=job --rw=rw

   READ: bw=2428MiB/s (2546MB/s), 2428MiB/s-2428MiB/s (2546MB/s-2546MB/s), io=23.7GiB (25.5GB), run=10001-10001msec
  WRITE: bw=2429MiB/s (2547MB/s), 2429MiB/s-2429MiB/s (2547MB/s-2547MB/s), io=23.7GiB (25.5GB), run=10001-10001msec

regular loop (on /mnt/test/test):
# losetup /dev/loop0 /mnt/test/test
# fio --direct=1 --bs=4k --runtime=10 --time_based --numjobs=16 
--ioengine=psync --iodepth=1 --group_reporting=1 --filename=/dev/loop0 
--name=job --rw=rw

   READ: bw=326MiB/s (342MB/s), 326MiB/s-326MiB/s (342MB/s-342MB/s), io=3259MiB (3417MB), run=10003-10003msec
  WRITE: bw=326MiB/s (342MB/s), 326MiB/s-326MiB/s (342MB/s-342MB/s), io=3260MiB (3418MB), run=10003-10003msec


dm-loop is even slightly faster than running fio directly on the regular 
file:

# fio --direct=1 --bs=4k --runtime=10 --time_based --numjobs=16 
--ioengine=psync --iodepth=1 --group_reporting=1 --filename=/mnt/test/test 
--name=job --rw=rw
   READ: bw=2005MiB/s (2103MB/s), 2005MiB/s-2005MiB/s (2103MB/s-2103MB/s), io=19.6GiB (21.0GB), run=10002-10002msec
  WRITE: bw=2007MiB/s (2104MB/s), 2007MiB/s-2007MiB/s (2104MB/s-2104MB/s), io=19.6GiB (21.0GB), run=10002-10002msec

Mikulas