Re: [RFC][-mm] Memory controller add mm->owner

[Balbir Singh <balbir@linux.vnet.ibm.com>]

Paul Menage wrote:
> On Mon, Mar 24, 2008 at 9:21 AM, Balbir Singh <balbir@linux.vnet.ibm.com> wrote:
>>  > Also, if mm->owner exits but mm is still alive (unlikely, but could
>>  > happen with weird custom threading libraries?) then we need to
>>  > reassign mm->owner to one of the other users of the mm (by looking
>>  > first in the thread group, then among the parents/siblings/children,
>>  > and then among all processes as a last resort?)
>>  >
>>
>>  The comment in __exit_signal states that
>>
>>  "The group leader stays around as a zombie as long
>>   as there are other threads.  When it gets reaped,
>>   the exit.c code will add its counts into these totals."
> 
> Ah, that's useful to know.
> 
>>  Given that the thread group leader stays around, do we need to reassign
>>  mm->owner? Do you do anything special in cgroups like cleanup the
>>  task_struct->css->subsys_state on exit?
>>
> 
> OK, so we don't need to handle this for NPTL apps - but for anything
> still using LinuxThreads or manually constructed clone() calls that
> use CLONE_VM without CLONE_PID, this could still be an issue. 

CLONE_PID?? Do you mean CLONE_THREAD?

For the case you mentioned, mm->owner is a moving target and we don't want to
spend time finding the successor, that can be expensive when threads start
exiting one-by-one quickly and when the number of threads are high. I wonder if
there is an efficient way to find mm->owner in that case.

(Also I
> guess there's the case of someone holding a reference to the mm via a
> /proc file?)
> 

Yes, but in that case we'll not be charging/uncharging anything to that mm or
the cgroup to which the mm belongs.

>>  >>  -       rcu_read_lock();
>>  >>  -       mem = rcu_dereference(mm->mem_cgroup);
>>  >>  +       mem = mem_cgroup_from_task(mm->owner);
>>  >
>>  > I think we still need the rcu_read_lock(), since mm->owner can move
>>  > cgroups any time.
>>  >
>>
>>  OK, so cgroup task movement is protected by RCU, right? I'll check for all
>>  mm->owner uses.
>>
> 
> Yes - cgroup_attach() uses synchronize_rcu() before release the cgroup
> mutex. So although you can't guarantee that the cgroup set won't
> change if you're just using RCU, you can't guarantee that you're
> addressing a still-valid non-destroyed (and of course non-freed)
> cgroup set.
> 

Yes, I understand that part of RCU.

-- 
	Warm Regards,
	Balbir Singh
	Linux Technology Center
	IBM, ISTL