Re: [PATCH 1/2] memcg: dirty pages accounting and limiting infrastructure

[Minchan Kim <minchan.kim@gmail.com>]

On Fri, Feb 26, 2010 at 2:01 PM, KAMEZAWA Hiroyuki
<kamezawa.hiroyu@jp.fujitsu.com> wrote:
> Hi,
>
> On Fri, 26 Feb 2010 13:50:04 +0900
> Minchan Kim <minchan.kim@gmail.com> wrote:
>
>> > Hm ? I don't read the whole thread but can_attach() is called under
>> > cgroup_mutex(). So, it doesn't need to use RCU.
>>
>> Vivek mentioned memcg is protected by RCU if I understand his intention right.
>> So I commented that without enough knowledge of memcg.
>> After your comment, I dive into the code.
>>
>> Just out of curiosity.
>>
>> Really, memcg is protected by RCU?
> yes. All cgroup subsystem is protected by RCU.
>
>> I think most of RCU around memcg is for protecting task_struct and
>> cgroup_subsys_state.
>> The memcg is protected by cgroup_mutex as you mentioned.
>> Am I missing something?
>
> There are several levels of protections.
>
> cgroup subsystem's ->destroy() call back is finally called by
>
> As this.
>
>  768                 synchronize_rcu();
>  769
>  770                 mutex_lock(&cgroup_mutex);
>  771                 /*
>  772                  * Release the subsystem state objects.
>  773                  */
>  774                 for_each_subsys(cgrp->root, ss)
>  775                         ss->destroy(ss, cgrp);
>  776
>  777                 cgrp->root->number_of_cgroups--;
>  778                 mutex_unlock(&cgroup_mutex);
>
> Before here,
>        - there are no tasks under this cgroup (cgroup's refcnt is 0)
>          && cgroup is marked as REMOVED.
>
> Then, this access
>        rcu_read_lock();
>        mem = mem_cgroup_from_task(task);
>        if (css_tryget(mem->css))   <===============checks cgroup refcnt

If it it, do we always need css_tryget after mem_cgroup_from_task
without cgroup_mutex to make sure css is vaild?

But I found several cases that don't call css_tryget

1. mm_match_cgroup
It's used by page_referenced_xxx. so we I think we don't grab
cgroup_mutex at that time.

2. mem_cgroup_oom_called
I think in here we don't grab cgroup_mutex, too.

I guess some design would cover that problems.
Could you tell me if you don't mind?
Sorry for bothering you.

Thanks, Kame.



-- 
Kind regards,
Minchan Kim