Re: [RFC][PATCH 0/3] impelemnt cgroup_(subsys)_disabled in generic.

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

KAMEZAWA Hiroyuki wrote:
> 
> Now, memory cgroup has 'mem_cgroup_disabled()' in memcontrol.h
> 
> I made a brief trial to use static_branch() for that function. At doing that,
> I thought it will be better to implement generic cgroup functions rather
> than having memory cgroup's its own one.
> 
> This series consists of 3 patches
> 1 .... implement cgroup_xxxx_disabled() in generic.
> 2 .... use jump_label for cgroup_xxxx_disabled()
> 3 .... remove mem_cgroup_disabled() in memcontrol.c
> 
> And I post this series for getting review/comments.
> I'm not sure patches for using jump_label is worth to be merged.
> 
> I did a test to run a loop
> 	while(-) {
> 		mmap(1M)
> 		touch all pages
> 		munmap()
> 	}
> 
> and measured performance score in ROOT cgroup. Here,
> 
> (Before patch)
>    182,932,842,128 cycles                    #    0.000 GHz                     [33.33%]
>    192,711,643,877 instructions              #    1.05  insns per cycle         [49.99%]
>        761,483,416 cache-references                                             [49.98%]
>            159,908 cache-misses              #    0.021 % of all cache refs     [50.00%]
>     33,253,084,874 branches                                                     [33.34%]
>        109,796,792 branch-misses             #    0.33% of all branches         [33.34%]
> 
>       58.289265709 seconds time elapsed
> 
> (After patch)
>  Performance counter stats for './malloc 1':
> 
>    183,068,407,487 cycles                    #    0.000 GHz                     [33.33%]
>    191,834,248,678 instructions              #    1.05  insns per cycle         [50.00%]
>        798,635,028 cache-references                                             [49.98%]
>             95,562 cache-misses              #    0.012 % of all cache refs     [50.00%]
>     32,755,318,286 branches                                                     [33.34%]
>         77,774,624 branch-misses             #    0.24% of all branches         [33.34%]
> 
>       58.332356996 seconds time elapsed
> 
> There is no differece in 'time' ;) 
> But I got an impression that 'branch' score gets better in several tests.
> 
> Thanks,
> -Kame
> 
> P.S. maybe I can replace 'do_swap_account' with jump_label, too.
> 

The numbers sugguest using jump label is a win.

However I'm not quite convinced that we make it generic. The subsys.disabled flag
was introduced long ago for memcg, but yet it has no other users.

So maybe for now make use of jump label in memcg only? We probably still needs
a bit help from cgroup core, to provide a subsys->disable() callback.

--
Li Zefan