Re: [PATCH RFC 1/2] memcg: use percpu_counter for statistics

[Vladimir Davydov <vdavydov-bzQdu9zFT3WakBO8gow8eQ@public.gmane.org>]

On Fri, Sep 12, 2014 at 10:10:52AM +0900, Kamezawa Hiroyuki wrote:
> (2014/09/12 0:41), Vladimir Davydov wrote:
> > In the next patch I need a quick way to get a value of
> > MEM_CGROUP_STAT_RSS. The current procedure (mem_cgroup_read_stat) is
> > slow (iterates over all cpus) and may sleep (uses get/put_online_cpus),
> > so it's a no-go.
> > 
> > This patch converts memory cgroup statistics to use percpu_counter so
> > that percpu_counter_read will do the trick.
> > 
> > Signed-off-by: Vladimir Davydov <vdavydov-bzQdu9zFT3WakBO8gow8eQ@public.gmane.org>
> 
> 
> I have no strong objections but you need performance comparison to go with this.
> 
> I thought percpu counter was messy to be used for "array".
> I can't understand why you started from fixing future performance problem before
> merging new feature.

Because the present implementation of mem_cgroup_read_stat may sleep
(get/put_online_cpus) while I need to call it from atomic context in the
next patch.

I didn't do any performance comparisons, because it's just an RFC. It
exists only to attract attention to the problem. Using percpu counters
was the quickest way to implement a draft version, that's why I chose
them. It may have performance impact though, so it shouldn't be merged
w/o performance analysis.

Thanks,
Vladimir