<NinjaTrappeur>
What would be the perk for you to have that as a list instead of an attrset?
<andi->
I commonly split network definitions across multiple files
<andi->
or only add one more knob as part of another module.
<andi->
What if I want to override the match expression? You want to soley rely on the order of the rule that comes first by doing that?
<andi->
In other places we are actively trying to get away from lists because of their not so economic override situation.
<arianvp>
andi-: it doesn't merge
<arianvp>
because merging also isn't something that networkd supports upstream
<arianvp>
I think the merging behaviour is confusing either way
<arianvp>
if you have
<andi->
what
<andi->
next thing you tell me is to do network.confDir = ./my-manually-managed-scripts;
<arianvp>
hmm maybe not
<andi->
I mean maybe you should reiterate on your motivation
<andi->
I see issues with our current networkd configuration but lists vs sets isn't one of them
<arianvp>
andi-: odd I cant reproduce this IPAccounting issue on my desktop but I can on my laptop
<arianvp>
(Desktop running 246.6 and laptop running 246.4)
<andi->
arianvp: I can do on the laptop, desktop and my servers..
<andi->
laptop/desktop being on unstable as of a few days ago
<arianvp>
Yeh my desktop is on unstable; laptop on 20.09
<arianvp>
I'm gonna say obscure systemd bug
<arianvp>
=)
<arianvp>
happy hunting through the bpf stack
<arianvp>
so my server (unstable) doesnt reproduce either. It works fine
<arianvp>
so only 1/3 devices are bugged for me so far..
<flokli>
arianvp: cgroup hierarchies?
<flokli>
I recall gchristensen asking about IO accounting in here some time ago
<gchristensen>
yeah that is some shit
<gchristensen>
do idea what that is about, and #systemd didn't bite eiter
<flokli>
I don't /really/ want to reboot, but can we beef up nixos/tests/systemd.nix to have a unit doing some networking traffic and IO, and verify it's not all zeroes?
<flokli>
and then check with both hierarchiers?
<flokli>
s/iers/ies/
<flokli>
should be a suffiently sophisticated integration test ;-)
<gchristensen>
fwiw I think the reason I get all zeros is the io.stat files are all empty
<gchristensen>
I've read it, but it doesn't seem to have anything related to the issue
<flokli>
so, I spinned up a vm test with systemd.unified_cgroup_hierarchy in kernelParams, and can confirm I only have io.stat nodes there, not if I boot it with our default cgroup hierarchy
<flokli>
"systemctl status ….service" also has IP: and IO: fields
<flokli>
they're currently still all zeroes.
<gchristensen>
do those io.stat files contain anything?
<flokli>
/sys/fs/cgroup/user.slice/io.stat does
<flokli>
same for system.slice
<gchristensen>
yeah min edoes too
<flokli>
systemctl status user.slice and system.slice also show this
<flokli>
both IP and IO
<NinjaTrappeur>
I can't repro this 0 bug on any machine (tried on desktop, server, laptop and router). I don't get it :/
<flokli>
now let me try on a box without setting this kernel param
<flokli>
I have IP, but not IO in the output. and no "io.stat" files in /sys/fs/cgroup either
<flokli>
so I'd say, for IO accounting we need systemd.unified_cgroup_hierarchy
<flokli>
one more reason on why we should flip that default, like fedora and most other distros do
<andi->
unified is still compatible with that docker crap?
<flokli>
people still wanting to run docker just need to reboot, or build a nixos with that kernelParam while installing
<flokli>
docker/moby 20 seems to support (parts of?) the unified cgroup hierarchy, but it's beta, or was in september
<andi->
ugh
<andi->
the host where I noticed it also runs docker containers for gitlab :D
<flokli>
My point is. There's workarounds. add the kernel cmdline if you want the old hierarchy, or use podman or another container runtime.
<andi->
ok, so lets do that for 21.03 with proper docs etc?
<flokli>
yes, I'd like to do the switch now-ish already, so things can be sorted out till 21.03 becomes a thing
<andi->
Shall we try a bi-weekly systemd call that I talked about >6 months ago already? I feel like that makes it much easier to coordinate thatn random ramblings here :D
<flokli>
yes, that sounds good
<flokli>
we should set up a doodle (in UTC) for signalling interest and availabilities, and then we can try to come up with a recurring ical invite
<andi->
yeah, what is the ethically acceptable doodle alternative these days?
<andi->
framadate it is
* andi-
starts to draft something
<gchristensen>
okay but also, the io.stat behavior is imo broken still under cgroups v2
<flokli>
andi-: you write some discourse post too? ;-)
<andi->
no
<andi->
I would only send it out to those actually doing systemd work for now
<flokli>
ok.
<andi->
and if we have a date for those working on it we can invite others
<andi->
otherwise we end up with a nixos-security mumble that isn't about security in nixos ;)
<flokli>
I'll draft a PR :-)
<flokli>
gchristensen: yeah, that might be another bug. wanna sketch a vm test and bisect nixpkgs with it? ;-)
<gchristensen>
I don't have evidence that it has ever worked
<flokli>
I've definitely seen it working once, I'm not sure how/when it broke
<flokli>
okay, I saw it working on another distro at least ;-)
<gchristensen>
it is a bummer, I wanted to have a button to pop up a HUD that shows IP/disk/cpu activity for the focused cgroup=
<andi->
aanderse, arianvp, flokli, Mic92, NinjaTrappeur, …: Here is a thing where we can try to find a rough weekday and time where most of us are available. The idea is to pick the best date from this poll and then use that for weekday & time for every other week to sync up on systemd. https://www.when2meet.com/?10369248-hdnZc
<andi->
I remove everythin in the late night and morning in europe as a bunch of us are already in europe and I do not think anything after midnight is reasonable for regular meetings.