Re: [PATCH] memory cgroup: update root memory cgroup when node isonlined

From: Kamezawa Hiroyuki
Date: Tue Oct 16 2012 - 03:21:29 EST


(2012/09/14 10:36), Hugh Dickins wrote:
On Thu, 13 Sep 2012, Johannes Weiner wrote:
On Thu, Sep 13, 2012 at 03:14:28PM +0800, Wen Congyang wrote:
root_mem_cgroup->info.nodeinfo is initialized when the system boots.
But NODE_DATA(nid) is null if the node is not onlined, so
root_mem_cgroup->info.nodeinfo[nid]->zoneinfo[zone].lruvec.zone contains
an invalid pointer. If we use numactl to bind a program to the node
after onlining the node and its memory, it will cause the kernel
panicked:

Is there any chance we could get rid of the zone backpointer in lruvec
again instead?

It could be done, but it would make me sad :(

Adding new nodes is a rare event and so updating every
single memcg in the system might be just borderline crazy.

Not horribly crazy, but rather ugly, yes.

But can't
we just go back to passing the zone along with the lruvec down
vmscan.c paths? I agree it's ugly to pass both, given their
relationship. But I don't think the backpointer is any cleaner but in
addition less robust.

It's like how we use vma->mm: we could change everywhere to pass mm with
vma, but it looks cleaner and cuts down on long arglists to have mm in vma.
From past experience, one of the things I worried about was adding extra
args to the reclaim stack.


That being said, the crashing code in particular makes me wonder:

static __always_inline void add_page_to_lru_list(struct page *page,
struct lruvec *lruvec, enum lru_list lru)
{
int nr_pages = hpage_nr_pages(page);
mem_cgroup_update_lru_size(lruvec, lru, nr_pages);
list_add(&page->lru, &lruvec->lists[lru]);
__mod_zone_page_state(lruvec_zone(lruvec), NR_LRU_BASE + lru, nr_pages);
}

Why did we ever pass zone in here and then felt the need to replace it
with lruvec->zone in fa9add6 "mm/memcg: apply add/del_page to lruvec"?
A page does not roam between zones, its zone is a static property that
can be retrieved with page_zone().

Just as in vmscan.c, we have the lruvec to hand, and that's what we
mainly want to operate upon, but there is also some need for zone.

(Both Konstantin and I were looking towards the day when we move the
lru_lock into the lruvec, removing more dependence on "zone". Pretty
much the only reason that hasn't happened yet, is that we have not found
time to make a performance case convincingly - but that's another topic.)

Yes, page_zone(page) is a static property of the page, but it's not
necessarily cheap to evaluate: depends on how complex the memory model
and the spare page flags space, doesn't it? We both preferred to
derive zone from lruvec where convenient.

How do you feel about this patch, and does it work for you guys?

You'd be right if you guessed that I started out without the
mem_cgroup_zone_lruvec part of it, but oops in get_scan_count
told me that's needed too.

Description to be filled in later: would it be needed for -stable,
or is onlining already broken in other ways that you're now fixing up?

Reported-by: Tang Chen <tangchen@xxxxxxxxxxxxxx>
Signed-off-by: Hugh Dickins <hughd@xxxxxxxxxx>

Acked-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@xxxxxxxxxxxxxx>



--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/