2012-10-16 07:21:39

by Kamezawa Hiroyuki

[permalink] [raw]
Subject: Re: [PATCH] memory cgroup: update root memory cgroup when node is onlined

(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 <[email protected]>
> Signed-off-by: Hugh Dickins <[email protected]>

Acked-by: KAMEZAWA Hiroyuki <[email protected]>