[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Date: Wed, 21 Dec 2011 09:13:47 +0900
From: KAMEZAWA Hiroyuki <kamezawa.hiroyu@...fujitsu.com>
To: Hiroyuki Kamezawa <kamezawa.hiroyuki@...il.com>
Cc: Hugh Dickins <hughd@...gle.com>,
"akpm@...ux-foundation.org" <akpm@...ux-foundation.org>,
"linux-mm@...ck.org" <linux-mm@...ck.org>, cgroups@...r.kernel.org,
"linux-kernel@...r.kernel.org" <linux-kernel@...r.kernel.org>,
"hannes@...xchg.org" <hannes@...xchg.org>,
Michal Hocko <mhocko@...e.cz>
Subject: Re: [PATCH] memcg: reset to root_mem_cgroup at bypassing
On Tue, 20 Dec 2011 09:24:47 +0900
Hiroyuki Kamezawa <kamezawa.hiroyuki@...il.com> wrote:
> 2011/12/20 Hugh Dickins <hughd@...gle.com>:
> > On Mon, 19 Dec 2011, KAMEZAWA Hiroyuki wrote:
> >> From d620ff605a3a592c2b1de3a046498ce5cd3d3c50 Mon Sep 17 00:00:00 2001
> >> From: KAMEZAWA Hiroyuki <kamezawa.hiroyu@...fujitsu.com>
> >> Date: Mon, 19 Dec 2011 16:55:10 +0900
> >> Subject: [PATCH 2/2] memcg: reset lru to root_mem_cgroup in special cases.
> >>
> >> This patch is a fix for memcg-simplify-lru-handling-by-new-rule.patch
> >>
> >> After the patch, all pages which will be onto LRU must have sane
> >> pc->mem_cgroup. But, in special case, it's not set.
> >>
> >> If task->mm is NULL or task is TIF_MEMDIE or fatal_signal_pending(),
> >> try_charge() is bypassed and the new charge will not be charged. And
> >> pc->mem_cgroup is unset even if the page will be used/mapped and added
> >> to LRU. To avoid this, this patch charges such pages to root_mem_cgroup,
> >> then, pc->mem_cgroup will be handled correctly.
> >>
> >> Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@...fujitsu.com>
> >> ---
> >> mm/memcontrol.c | 2 +-
> >> 1 files changed, 1 insertions(+), 1 deletions(-)
> >>
> >> diff --git a/mm/memcontrol.c b/mm/memcontrol.c
> >> index 0d6d21c..9268e8e 100644
> >> --- a/mm/memcontrol.c
> >> +++ b/mm/memcontrol.c
> >> @@ -2324,7 +2324,7 @@ nomem:
> >> *ptr = NULL;
> >> return -ENOMEM;
> >> bypass:
> >> - *ptr = NULL;
> >> + *ptr = root_mem_cgroup;
> >> return 0;
> >> }
> >>
> >> --
> >
> Thank you for review.
>
> > I'm dubious about this patch: certainly you have not fully justified it.
> >
> I sometimes see panics (in !pc->mem_cgroup check in lru code)
> when I stops test programs by Ctrl-C or some. That was because
> of this path. I checked this by adding a debug code to make
> pc->mem_cgroup = NULL in prep_new_page.
>
> > I speak from experience: I did *exactly* the same at "bypass" when
> > I introduced our mem_cgroup_reset_page(), which corresponds to your
> > mem_cgroup_reset_owner(); it seemed right to me that a successful
> > (return 0) call to try_charge() should provide a good *ptr.
> >
> ok.
>
> > But others (Ying and Greg) pointed out that it changes the semantics
> > of __mem_cgroup_try_charge() in this case, so you need to justify the
> > change to all those places which do something like "if (ret || !memcg)"
> > after calling it. Perhaps it is a good change everywhere, but that's
> > not obvious, so we chose caution.
> >
>
> > Doesn't it lead to bypass pages being marked as charged to root, so
> > they don't get charged to the right owner next time they're touched?
> >
> Yes. You're right.
> Hm. So, it seems I should add reset_owner() to the !memcg path
> rather than here.
>
Considering this again..
Now, we catch 'charge' event only once in lifetime of anon/file page.
So, it doesn't depend on that it's marked as PCG_USED or not.
> > In our internal kernel, I restored "bypass" to set *ptr = NULL as
> > before, but routed those callers that need it to continue on to
> > __mem_cgroup_commit_charge() when it's NULL, and let that do a
> > quick little mem_cgroup_reset_page() to root_mem_cgroup for this.
> >
> Yes, I'll prepare v2.
>
But ok, I'll go this way with some more description.
Thanks,
-Kame
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
Powered by blists - more mailing lists