cgroups: pure v1, hybrid, pure v2

[brauner]

Technically, this is neither a feature request nor a bug report. So sorry for that.

We currently support three different cgroup layouts:

1. pure cgroup v1 layout
2. hybrid layout
3. pure cgroup v2 layout

I would like to propose an agreement for the way forward. :)

So, we all agree that cgroup v2 is going to be the future and that we are in the process of making userspace tools and libraries that we maintain compatible with it.

What would be nice though is if we could agree that the hybrid cgroup layout will only be supported when the cgroup v2 controller is mounted at /sys/fs/cgroup/unified and is empty, i.e. that we don't ever support having some controllers mounted into v1 hierarchies and some controllers mounted into v2 hierarchies. This is a level of complexity userspace and kernel wise that I would like to make sure to avoid. In short, hybrid mode should be considered to be somewhat in maintenance-only mode. I wasn't sure if we ever discussed this.

[floppym]

What would be nice though is if we could agree that the hybrid cgroup layout will only be supported when the cgroup v2 controller is That is fine and we (@lxc) at least have made appropriate preparations.

I'm having trouble parsing this sentence... copy/paste fail?

[brauner]

@floppym, yip, thanks for pointing it out. :) Should be fixed.

[evverx]

Just to complete the list, there is another layout called unified_systemd_v232, which is also hybrid and totally incompatible with everything else. I'm not sure anybody uses it in practice, but, it's probably worth noting that Debian Stretch comes with systemd-232. The good news is that it's very hard to accidentally start to use it there: https://salsa.debian.org/systemd-team/systemd/blob/stretch/debian/patches/core-don-t-use-the-unified-hierarchy-for-the-systemd-cgro.patch).

[poettering]

Quite frankly at this point I think doing "hybrid" is a stopgap we should never have added... It blurs the road forward. People should either use full cgroupsv1 or full cgroupsv2 but anything in between is just a maintainance burden.

But anyway, that ship might have sailed, the code is in now.

What would be nice though is if we could agree that the hybrid cgroup layout will only be supported when the cgroup v2 controller is mounted at /sys/fs/cgroup/unified and is empty, i.e. that we don't ever support having some controllers mounted into v1 hierarchies and some controllers mounted into v2 hierarchies.

Yes, this is the intention and the only thing systemd supports. The hybrid mode is dumb and minimal. All it adds is the cgroupv2 core hierarchy itself, but controllers are not attached to it, and remain pinned to the classic cgroupsv1 mounts.

I'll prep a patch that clarifies this in docs/CGROUP_DELEGATION.md.

[brauner]

@poettering, thanks for doing this! Excellent!

[poettering]

Fix waiting in #10161