Re: [PATCH] cgroups: don't cache common ancestor in task counter subsys

[Li Zefan <lizf@cn.fujitsu.com>]

Frederic Weisbecker wrote:
> On Mon, Oct 17, 2011 at 01:27:53PM -0400, Ben Blum wrote:
>> On Mon, Oct 17, 2011 at 07:25:41PM +0200, Frederic Weisbecker wrote:
>>> When we rollback there, we are dealing with oldcgrp of the last thread
>>> we have treated. All threads in the rollback list don't necessary belonged
>>> to that old_cgroup.
>>>
>>> And we can't try to retrieve these old_cgroup through task_cgroup_from_root() because
>>> the threads might have exited and thus could be assigned to the init cgroup.
>>>
>>> I believe we need to cache these old cgroups in the flex array.
>>>
>>
>> Doing this would fulfill the "TODO" in cgroup_attach_proc for being able
>> to pass the oldcgrp for each task in the loop around the ss->attach
>> call, so if you do that, remove the corresponding comment. :)
>>
>> -- Ben
> 
> I don't know, looking at the code, there is a separate issue to solve there.
> The old cgroup passed in ->attach() is the one of a random task from the thread
> group. The migrations that happen right before leave oldcgrp with the
> old cgroup of the last thread in the group.
> 
> This doesn't seem to make any sense. As far as I checked this is only used
> by the cpuset subsystem that does a migration in cpuset_attach() from
> the old cgroup to the new.
> 
> I'm not sure about the details but that looks buggy.
> 

I think you're right, and Tejun's patchset happens to fix this bug in
cpuset_attach() - it gets the mm of the thread leader and do migration.