Re: [PATCHSET cgroup/for-3.8] cpuset: decouple cpuset locking from cgroup core

[Glauber Costa <glommer-bzQdu9zFT3WakBO8gow8eQ@public.gmane.org>]

On 11/30/2012 06:59 PM, Tejun Heo wrote:
> Hello,
> 
> On Fri, Nov 30, 2012 at 02:00:21PM +0400, Glauber Costa wrote:
>> Now, what I am actually seeing with cgroup creation, is that the
>> children will copy a lot of the values from the parent, like swappiness,
>> hierarchy, etc. Once the child copies it, we should no longer be able to
>> change those values in the parent: otherwise we'll get funny things like
>> parent.use_hierarchy = 1, child.use_hierarchy = 0.
> 
> So, the best way to do this is from ->css_online().  If memcg
> synchronizes and inherits from ->css_online(), it can guarantee that
> the new cgroup will be visible in any following iterations.  Just have
> an online flag which is turned on and off from ->css_on/offline() and
> ignore any cgroups w/o online set.
> 
>> One option is to take a global lock in memcg_alloc_css(), and keep it
>> locked until we did all the cgroup bookkeeping, and then unlock it in
>> css_online. But I am guessing Tejun won't like it very much.
> 
> No, please *NEVER* *EVER* do that.  You'll be creating a bunch of
> locking dependencies as cgroup walks through different controllers.
> 
> memcg should be able to synchornize fully both css on/offlining and
> task attachments in memcg proper.  Let's please be boring about
> locking.
> 

Of course, there was a purely rhetorical statement, as indicated by
"Tejun won't like it very much" =p

Take a look at the final result, I just posted a couple of hours ago.
Let me know if there is still something extremely funny, and I'll look
into fixing it.