Re: [GIT PULL] cgroup changes for v3.15-rc1

[Tejun Heo <tj@kernel.org>]

Hello,

On Thu, Apr 03, 2014 at 12:01:23PM -0700, Linus Torvalds wrote:
> [ Extending the participants list a bit ]
> 
> On Thu, Apr 3, 2014 at 11:34 AM, Tejun Heo <tj@kernel.org> wrote:
> > On the road so sending from phone. Iirc the param is necessary to
> > distinguishe when a new sb is created so that it can be put properly later.
> > I think cgroup is leaking super ref now and li was planning to send a fix
> > once things are merged.
> 
> So as far as I can tell, cgroup is fine, because the superblock itself
> is properly refcounted by the mounting code. It's the magic hidden

Ah, I remembered the other way around.  We could leak cgroup_root
reference, not the other way around.  cgroup_mount() can be called
multiple times for the same sb and we inc cgroup_root's ref each time
but cgroup_kill_sb() only happens when the sb is released, so if we do
the following,

 # mkdir cpuset cpuset1
 # mount -t cgroup -o cpuset cgroup cpuset
 # mount -t cgroup -o cpuset cgroup cpuset1
 # umount cpuset
 # umount cpuset1

The cgroup_root should be destroyed but it isn't, I think.  We'd need
to bump cgroup_root's refcnt only when a new sb is created.  It's
kinda ugly.  Hmmm...

As for using specific type for ns tag, yeah, that'd be better
regardless of this.  The opaqueness is a bit extreme now.

Thanks.

-- 
tejun