lkml.org 
[lkml]   [2008]   [Aug]   [20]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [RFC][PATCH -mm 0/7] memcg: lockless page_cgroup v1
KAMEZAWA Hiroyuki wrote:
> On Wed, 20 Aug 2008 20:00:06 +0900
> KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> wrote:
>
>> On Wed, 20 Aug 2008 19:41:08 +0900
>> KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> wrote:
>>
>>> On Wed, 20 Aug 2008 18:53:06 +0900
>>> KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> wrote:
>>>
>>>> Hi, this is a patch set for lockless page_cgroup.
>>>>
>>>> dropped patches related to mem+swap controller for easy review.
>>>> (I'm rewriting it, too.)
>>>>
>>>> Changes from current -mm is.
>>>> - page_cgroup->flags operations is set to be atomic.
>>>> - lock_page_cgroup() is removed.
>>>> - page->page_cgroup is changed from unsigned long to struct page_cgroup*
>>>> - page_cgroup is freed by RCU.
>>>> - For avoiding race, charge/uncharge against mm/memory.c::insert_page() is
>>>> omitted. This is ususally used for mapping device's page. (I think...)
>>>>
>>>> In my quick test, perfomance is improved a little. But the benefit of this
>>>> patch is to allow access page_cgroup without lock. I think this is good
>>>> for Yamamoto's Dirty page tracking for memcg.
>>>> For I/O tracking people, I added a header file for allowing access to
>>>> page_cgroup from out of memcontrol.c
>>>>
>>>> The base kernel is recent mmtom. Any comments are welcome.
>>>> This is still under test. I have to do long-run test before removing "RFC".
>>>>
>>> Known problem: force_emtpy is broken...so rmdir will struck into nightmare.
>>> It's because of patch 2/7.
>>> will be fixed in the next version.
>>>
>> This is a quick fix but I think I can find some better solution..
>> ==
>> Because removal from LRU is delayed, mz->lru will never be empty until
>> someone kick drain. This patch rotate LRU while force_empty and makes
>> page_cgroup will be freed.
>>
>
> I'd like to rewrite force_empty to move all usage to "default" cgroup.
> There are some reasons.
>
> 1. current force_empty creates an alive page which has no page_cgroup.
> This is bad for routine which want to access page_cgroup from page.
> And this behavior will be an issue of race condition in future.
> 2. We can see amount of out-of-control usage in default cgroup.
>
> But to do this, I'll have to avoid "hitting limit" in default cgroup.
> I'm now wondering to make it impossible to set limit to default cgroup.
> (will show as a patch in the next version of series.)
> Does anyone have an idea ?
>

Hi, Kamezawa-San,

The definition of default-cgroup would be root cgroup right? I would like to
implement hierarchies correctly in order to define the default-cgroup (it could
be a parent of the child cgroup for example).


--
Balbir


\
 
 \ /
  Last update: 2008-08-21 05:41    [W:0.073 / U:11.776 seconds]
©2003-2018 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site