Re: [PATCH v4 4/6] cgroup/cpuset: Allow non-top parent partition root to distribute out all CPUs

[Tejun Heo <tj@kernel.org>]

On Wed, Aug 11, 2021 at 02:46:24PM -0400, Waiman Long wrote:
> On 8/11/21 2:21 PM, Tejun Heo wrote:
> > On Wed, Aug 11, 2021 at 02:18:17PM -0400, Waiman Long wrote:
> > > I don't think that is true. A task can reside anywhere in the cgroup
> > > hierarchy. I have encountered no problem moving tasks around.
> > Oh, that shouldn't be happening with controllers enabled. Can you please
> > share a repro?
> 
> I have done further testing. Enabling controllers won't prohibit moving a
> task into a parent cgroup as long as the child cgroups have no tasks. Once

Should be "as long as there's no child cgroups".

  root@test /s/f/cgroup# mkdir test
  root@test /s/f/cgroup# mkdir -p test/A
  root@test /s/f/cgroup# echo +io > test/cgroup.subtree_control 
  root@test /s/f/cgroup# echo $fish_pid > test/cgroup.procs
  write: Device or resource busy

> the child cgroup has task, moving another task to the parent is not allowed
> (-EBUSY). Similarly if a parent cgroup has tasks, you can't put new tasks
> into the child cgroup. I don't realize that we have such constraints as I

You can't enable controller from a populated cgroup:

  root@test /s/f/cgroup# mkdir test
  root@test /s/f/cgroup# echo +io > test/cgroup.subtree_control 
  root@test /s/f/cgroup# echo $fish_pid > test/cgroup.procs

> usually do my testing with a cgroup hierarchy with no tasks initially.
> Anyway, a new lesson learned.

The invariant is that from each controller's POV, all cgroups with processes
in them are leaves. This is all pretty well documented in cgroup-v2.rst.

Thanks.

-- 
tejun