Re: [PATCH 4/4] memcg: replace cgroup_lock with memcg specific memcg_lock

[Tejun Heo <tj-DgEjT+Ai2ygdnm+yROfE0A@public.gmane.org>]

Hello, Michal.

On Tue, Dec 04, 2012 at 04:14:20PM +0100, Michal Hocko wrote:
> OK, I read this as "generic helper doesn't make much sense". Then I
> would just ask. Does cgroup core really care whether we do
> list_empty test? Is this something that we have to care about in memcg
> and should fix? If yes then just try to do it as simple as possible.

The thing is, what does the test mean when it doesn't have proper
synchronization?  list_empty(&cgroup->children) doesn't really have a
precise meaning if you're not synchronized.  There could be cases
where such correct-most-of-the-time results are okay but depending on
stuff like that is a sure-fire way to subtle bugs.

So, my recommendation would be to bite the bullet and implement
properly synchronized on/offline state and teach the memcg iterator
about it so that memcg can definitively tell what's online and what's
not while holding memcg_mutex, and use such knowledge consistently.

Thanks.

-- 
tejun