Re: memcg: add mlock statistic in memory.stat

[KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>]

On Thu, 12 Jan 2012 11:13:00 -0800
Ying Han <yinghan@google.com> wrote:

> On Wed, Jan 11, 2012 at 7:21 PM, KAMEZAWA Hiroyuki
> <kamezawa.hiroyu@jp.fujitsu.com> wrote:
> > On Wed, 11 Jan 2012 16:50:09 -0800
> > Ying Han <yinghan@google.com> wrote:
> >
> >> On Wed, Jan 11, 2012 at 3:59 PM, KAMEZAWA Hiroyuki
> >> <kamezawa.hiroyu@jp.fujitsu.com> wrote:
> >> > On Wed, 11 Jan 2012 15:17:42 -0800 (PST)
> >> > Hugh Dickins <hughd@google.com> wrote:
> >> >
> >> >> On Wed, 11 Jan 2012, Ying Han wrote:
> >> >>
> >> >> > We have the nr_mlock stat both in meminfo as well as vmstat system wide, this
> >> >> > patch adds the mlock field into per-memcg memory stat. The stat itself enhances
> >> >> > the metrics exported by memcg, especially is used together with "uneivctable"
> >> >> > lru stat.
> >> >> >
> >> >> > --- a/include/linux/page_cgroup.h
> >> >> > +++ b/include/linux/page_cgroup.h
> >> >> > @@ -10,6 +10,7 @@ enum {
> >> >> > A  A  /* flags for mem_cgroup and file and I/O status */
> >> >> > A  A  PCG_MOVE_LOCK, /* For race between move_account v.s. following bits */
> >> >> > A  A  PCG_FILE_MAPPED, /* page is accounted as "mapped" */
> >> >> > + A  PCG_MLOCK, /* page is accounted as "mlock" */
> >> >> > A  A  /* No lock in page_cgroup */
> >> >> > A  A  PCG_ACCT_LRU, /* page has been accounted for (under lru_lock) */
> >> >> > A  A  __NR_PCG_FLAGS,
> >> >>
> >> >> Is this really necessary? A KAMEZAWA-san is engaged in trying to reduce
> >> >> the number of PageCgroup flags, and I expect that in due course we shall
> >> >> want to merge them in with Page flags, so adding more is unwelcome.
> >> >> I'd A have thought that with memcg_ hooks in the right places,
> >> >> a separate flag would not be necessary?
> >> >>
> >> >
> >> > Please don't ;)
> >> >
> >> > NR_UNEIVCTABLE_LRU is not enough ?
> >>
> >> Seems not.
> >>
> >> The unevictable lru includes more than mlock()'d pages ( SHM_LOCK'd
> >> etc). There are use cases where we like to know the mlock-ed size
> >> per-cgroup. We used to archived that in fake-numa based container by
> >> reading the value from per-node meminfo, however we miss that
> >> information in memcg. What do you think?
> >>
> >
> > Hm. The # of mlocked pages can be got sum of /proc/<pid>/? ?
> 
> That is tough. Then we have to do the calculation by adding up all the
> pids within a cgroup.
> 
> > BTW, Roughly..
> >
> > (inactive_anon + active_anon) - rss = # of unlocked shm.
> >
> > cache - (inactive_file + active_file) = total # of shm
> >
> > Then,
> >
> > (cache - A (inactive_file + active_file)) - ((inactive_anon + active_anon) - rss)
> > = cache + rss - (sum of inactive/actige lru)
> > = locked shm.
> >
> > Hm, but this works only when unmapped swapcache is A small ;)
> 
> We might be getting a rough number. But we have use cases relying on
> more accurate output. Thoughts?
> 
If we need mega-byte order accuracy, above will work enough.
But ok, having stats seems useful because meminfo has it ;)

For your input, I'd like to post an updated RFC patch to do page state accounting
without additional bits to pc->flags, today. With that, you can rely on PG_mlocked.

By that patch, I know we can make use of page-flags by some logic but am still
looking for more efficient way...

Thanks,
-Kame

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Fight unfair telecom internet charges in Canada: sign http://stopthemeter.ca/
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>