On Fri, 17 Feb 2012 18:24:26 +0900
KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> wrote:

> 
> This patch set is for removing 2 flags PCG_FILE_MAPPED and PCG_MOVE_LOCK on
> page_cgroup->flags. After this, page_cgroup has only 3bits of flags.
> And, this set introduces a new method to update page status accounting per memcg.
> With it, we don't have to add new flags onto page_cgroup if 'struct page' has
> information. This will be good for avoiding a new flag for page_cgroup.
> 
> Fixed pointed out parts.
>  - added more comments
>  - fixed texts
>  - removed redundant arguments.
> 
> Passed some tests on 3.3.0-rc3-next-20120216.
> 

Here is a micro benchmark test before/after this series.
mmap 1G bytes twice and repeat fault->drop repeatedly. (test program is attached)

== Before == 3 runs after 1st run
[root@bluextal test]# time ./mmap 1G

real    0m21.053s
user    0m6.046s
sys     0m14.743s
[root@bluextal test]# time ./mmap 1G

real    0m21.302s
user    0m6.027s
sys     0m14.979s
[root@bluextal test]# time ./mmap 1G

real    0m21.061s
user    0m6.020s
sys     0m14.722s

== After == 3 runs after 1st run
[root@bluextal test]# time ./mmap 1G

real    0m20.969s
user    0m5.960s
sys     0m14.777s
[root@bluextal test]# time ./mmap 1G

real    0m20.968s
user    0m6.069s
sys     0m14.650s
[root@bluextal test]# time ./mmap 1G

real    0m21.164s
user    0m6.152s
sys     0m14.707s


I think there is no regression.


Thanks,
-Kame