Re: [RFC] cgroup: deprecate clone_children

[Tejun Heo <tj@kernel.org>]

Hello, Serge.

On Mon, Nov 05, 2012 at 01:17:14PM -0600, Serge Hallyn wrote:
> > As this is unnecessary feature with very limited use and awkward in
> 
> clone_children is currently required by lxc.  Of course lxc could
> manually set the .cpus and .mems in the newly created child, but if
> those interfaces or others like it ever change (i.e. any new ones which
> must be set before a task can join a cgroup) we'll get into yet more of
> a rats nets of code to support different kernel versions.

If lxc is using it, this will have to stay then.  I'll try to make it
a cpuset specific thing.

> (Just as an idea, if there was a way to generically tell from the list
> of files in my cgroups dir which files need to be initialized before a
> task can join, I think that would suffice.  Maybe a cgroups.needssetup
> file which right now contains 'cpuset.mems\ncpuset.cpus'. Then if we
> find a file we don't recognize in there we can throw an intelligent
> error, or guess at duplicating the parent value.)

Hmmm... I think the root problem is that different controllers don't
agree on the way they inherit configurations from parents.  cgroup
really is a trainwreck.  :(

I don't know whether "needssetup" is the right way to deal with it.
I'll think more about it.

> > co-mounted use cases, let's try to deprecate it.  Whine on the mount
> > option and accesses to cgroupfs knobs.
> > 
> > Signed-off-by: Tejun Heo <tj@kernel.org>
> > Cc: Glauber Costa <glommer@parallels.com>
> > Cc: Peter Zijlstra <peterz@infradead.org>
> > ---
> > Glauber, I think this is more befitting change for .post_clone().  If
> 
> What do you mean by that?  That he is working on a patch-set which will
> remove post_clone and this belongs there, or that there is a proposed
> alternative?

The former.  There was a patchset from Glauber updating ->post_clone()
(which didn't change the behavior).

Thanks.

-- 
tejun