From: Song Wang <wangsin...@gmail.com>
Hi Mike,
Sorry for interruption.
For cgroup v2 blkio writeback throttle, bio uses its blkcg->css and bi_bdev to
calc out blkg. after the commit ca522482e3ea ("dm: pass NULL bdev to
bio_alloc_clone"), when DM targets call bio_set_dev()->bio_associate_blkg(),
the bio->bi_blkg is NULL, so css is another new blkcg_css(), and associated
blkg is root_blkg. this makes cgv2 blkio writeback throttle not work on the
underlying device.
It's very easy to reproduce this issue, firstly create and mount the lvm:
sda 8:0 0 20G 0 disk
└─vg-lv 252:0 0 16G 0 lvm /mnt
vg-lv: 0 33554432 linear 8:0 2048
Secondly, setup the cgv2 blkio writeback throttle:
mkdir /sys/fs/cgroup/cg
cd /sys/fs/cgroup/cg
echo 8:0 wbps=$((10*1024*1024)) >> io.max
echo $$ > cgroup.procs
Lastly, use dd and iostat to test and monitor:
dd if=/dev/urandom of=/mnt/test bs=1M
iostat -xmd 3 sda
Now you can see that the throttle is not working, and the throttle can work
before this commit (you need to backport the commit 320fb0f9, which fixes a
former throttle-not-working bug).
However, cgv2 blkio writeback throttle can still work well if you limit vg-lv
and test vg-lv, or if you limit sda and test sda without lvm.
So, I was wondering if this is a bug or a feature-correction？
Thanks
Song
