Comment 24 for bug 1933128

Speaking for LXD, LXD itself doesn't perform any cgroup mounts, it just sets up an empty cgroup namespace.

The cgroup setup you see in the container are the result of what systemd and other userspace tools have setup. So it's technically impossible for it to be a LXD issue as LXD doesn't have any logic related to those mounts ;)

As the host appears to be 20.04, it's unlikely that we're dealing with a mixed cgroup2/cgroup1 situation here. Instead it looks more like the cgroup mounts just disappearing from the child mount namespace.

I don't know exactly what would cause that, but we've observed in the past that systemd will perform a full cgroup reset (delete all cgroups) whenever it believes none of the scope/slice units require specific controllers. Maybe it's that same logic which is triggering a reset on those mounts and subsequent modification of the mount namespaces?