Unable to remove control groups on 2.6.23-rc8-mm1

[Badari Pulavarty]

On Wed, 2007-09-26 at 08:55 +0530, Balbir Singh wrote:
> Badari Pulavarty wrote:
> > Hi,
> > 
> > I am playing with control groups on 2.6.23-rc8-mm1.
> > 
> > I am able to mount cgroup and create subgroups. I was able
> > to move some tasks into them. But, after killing tasks I am
> > not able to remove the subgroups. Any idea on why ?
> > 
> > Thanks,
> > Badari
> > 
> 
> Badari,
> 
> We account for page cache usage as well now, I suspect you
> most likely have page/swap cache pages charged to the container.

Yep. Even after killing all the tasks and "echo 1 > drop_caches"
some memory still accounted for on this group.

elm3b155:/dev/cgroup # cat zzz/memory.usage_in_bytes
131072


> You can try several options (some documented in Documentation/
> controllers/memory.txt)
> 
> 1. Try executing sync; echo 1 > /proc/sys/vm/drop_caches
>    and then remove the directory
> 2. Prior to assigning tasks, set memory.control_type to 1,
>    that tracks only RSS pages. You'll find memory.usage_in_bytes
>    go to zero as soon as all the tasks exit

This helped. 

> 3. Set notify_on_release and use the release_agent and releasable
>    to free the container once all pages charged to it are freed.
> 
> I wonder if I should provide a force_reclaim (hard to guarantee
> it will work) for each container, so that the container can
> be freed.
> 

BTW, how do I detach a "pid" from the cgroup ?

Thanks,
Badari