Control: tag -1 moreinfo On Sun, 23 Mar 2025 08:46:28 +0100 Salvatore Bonaccorso <car...@debian.org> wrote: [...] > it has passed a substantial amount of time and we did not respond, > apologies about that. > > In my opinion we should leave it as module, people wanting to use that > setup can load the module by own means and do additional configuration > selecting the bfq scheduler, and think resepctive units will as well > need to enable the IOAccounting (correct me if I'm wrong)? > > Given this reply now, and time permitting we might discuss it in our > next kernel-team weekly meeting.
We discussed this in a kernel team meeting and I agreed to look at what
was going on here.
In a test VM with 6.12.16-amd64:
- When I enable the "io" controller for a cgroup, I see an "io.weight"
attribute in that cgroup. So the originally reported problem has been
fixed.
- With bfq not yet loaded, I can still set it as a scheduler and the
module is auto-loaded:
# cd /sys/class/block/vda/queue
# cat scheduler
[none] mq-deadline
# echo bfq > scheduler
# cat scheduler
none mq-deadline [bfq]
- After that, the cgroup also has an "io.bfq.weight" attribute.
systemd tries to set both attributes, which should cover devices with
BFQ and other schedulers. But I see a problem with this sequence of
events:
1. systemd creates the cgroup for a service and sets its "io.weight"
attribute only, as bfq is not yet loaded.
2. The scheduler of a block device that the service will use is changed
to "bfq". The bfq module is loaded and this cgroup gets an
"io.bfq.weight" attribute, but with a default value.
So far as I can see, neither the kernel nor systemd will synchronise the
added "io.bfq.weight" attribute with the "io.weight" attribute in step
2. Whereas if bfq was built-in, this wouldn't be a problem.
Ryutaroh, do you consider your bug report to be fixed, or is the problem
I described above part of the bug?
Ben.
--
Ben Hutchings
The two most common things in the universe are hydrogen and stupidity.
signature.asc
Description: This is a digitally signed message part
