Clarify meaning of --cgroup-parent about runc HOT 9 CLOSED

I think that for relative case it should be

/sys/fs/cgroup/memory/system.slice/docker-65c642b685216ddeb40b3ea34892ac97faa1dc3487e9c8dfcf0fdf7ed05d0941.scope/test/7f299d5c524193189cabcc01df143c88abe95ac1e274b36926340e000a5e786c/

But for abs, it should be

/sys/fs/cgroup/memory/test/7f299d5c524193189cabcc01df143c88abe95ac1e274b36926340e000a5e786c/

from runc.

@mrunalp but it's impossible, I can't leave mount namespace of container to create

/sys/fs/cgroup/memory/test/7f299d5c524193189cabcc01df143c88abe95ac1e274b36926340e000a5e786c/

treat my

/sys/fs/cgroup/memory/system.slice/docker-65c642b685216ddeb40b3ea34892ac97faa1dc3487e9c8dfcf0fdf7ed05d0941.scope

as /sys/fs/cgroup/memory, because it's in container.
We can discuss how it should be on host, without my crappy slices.

from runc.

So, I disabled my systemd noise, now path with parents:

/sys/fs/cgroup/memory/test/c8af2a2a2b82c46e3120034a0cfe9288e4f118b6f7ec0dabeaf489cc25a66fb2

from runc.

So, we basically don't understand for what we need absolute paths. You can't create cgroups in parent namespace anyway and if you want to use existing cgroup from host you can use only your own cgroup, so you always can use relative path.

from runc.

@LK4D4 absolute paths are required for providing QoS tiers. These cgroups can be created from within the root namespace. Unless we provide subcontainer capabilities, I don't see a need for --cgroup-parent. That feature is useful mainly for orchestration.

from runc.

@vishh Could you provide example how it should look? Because now they're(abs and rel) identical.

from runc.

If the path is relative, then the cgroups are expected to be created under the cgroup of the calling process. Ex: cgroups created under /sys/fs/cgroup/memory/docker, in the case of docker daemon.
In the case of runc by itself, both abs and rel paths will be the same.
If there is no need for relative cgroup parent paths, we can eliminate that to avoid ambiguity.

from runc.

Should this be closed? The --cgroup-parent option doesn't even seem to exist anymore.

from runc.

@fabiokung it's still exitst in libcontainer but I still didn't get it :D But I'll delay to personal meeting with Vishnu.

from runc.