memcg: fix fatal livelock in kswapd

[Balbir Singh]

Sorry, my mailer might have used intelligence to send HTML (that is
what happens when the setup changes, I apologize). Resending in text
format

On Sun, May 8, 2011 at 3:29 AM, Balbir Singh <balbir at linux.vnet.ibm.com> wrote:
>
>
> On Tue, May 3, 2011 at 4:18 AM, Johannes Weiner <hannes at cmpxchg.org> wrote:
>>
>> Hi,
>>
>> On Mon, May 02, 2011 at 03:07:29PM -0500, James Bottomley wrote:
>> > The fatal livelock in kswapd, reported in this thread:
>> >
>> > http://marc.info/?t=130392066000001
>> >
>> > Is mitigateable if we prevent the cgroups code being so aggressive in
>> > its zone shrinking (by reducing it's default shrink from 0 [everything]
>> > to DEF_PRIORITY [some things]).  This will have an obvious knock on
>> > effect to cgroup accounting, but it's better than hanging systems.
>>
>> Actually, it's not that obvious.  At least not to me.  I added Balbir,
>> who added said comment and code in the first place, to CC: Here is the
>> comment in full quote:
>>
>
> I missed this email in my inbox, just saw it and responding
>
>>
>>        /*
>>         * NOTE: Although we can get the priority field, using it
>>         * here is not a good idea, since it limits the pages we can scan.
>>         * if we don't reclaim here, the shrink_zone from balance_pgdat
>>         * will pick up pages from other mem cgroup's as well. We hack
>>         * the priority and make it zero.
>>         */
>>
>> The idea is that if one memcg is above its softlimit, we prefer
>> reducing pages from this memcg over reclaiming random other pages,
>> including those of other memcgs.
>>
>
> My comment and code were based on the observations I saw during my tests.
> With DEF_PRIORITY we see scan >> priority in get_scan_count(), since we know
> how much exactly we are over the soft limit, it makes sense to go after the
> pages, so that normal balancing can be restored.
>
>>
>> But the code flow looks like this:
>>
>>        balance_pgdat
>>          mem_cgroup_soft_limit_reclaim
>>            mem_cgroup_shrink_node_zone
>>              shrink_zone(0, zone, &sc)
>>          shrink_zone(prio, zone, &sc)
>>
>> so the success of the inner memcg shrink_zone does at least not
>> explicitely result in the outer, global shrink_zone steering clear of
>> other memcgs' pages.
>
> Yes, but it allows soft reclaim to know what to target first for success
>
>>
>>  It just tries to move the pressure of balancing
>> the zones to the memcg with the biggest soft limit excess.  That can
>> only really work if the memcg is a large enough contributor to the
>> zone's total number of lru pages, though, and looks very likely to hit
>> the exceeding memcg too hard in other cases.
>>
>> I am very much for removing this hack.  There is still more scan
>> pressure applied to memcgs in excess of their soft limit even if the
>> extra scan is happening at a sane priority level.  And the fact that
>> global reclaim operates completely unaware of memcgs is a different
>> story.
>>
>> However, this code came into place with v2.6.31-8387-g4e41695.  Why is
>> it only now showing up?
>>
>> You also wrote in that thread that this happens on a standard F15
>> installation.  On the F15 I am running here, systemd does not
>> configure memcgs, however.  Did you manually configure memcgs and set
>> soft limits?  Because I wonder how it ended up in soft limit reclaim
>> in the first place.
>>
>
> I am running F15 as well, but never hit the problem so far. I am surprised
> to see the stack posted on the thread, it seemed like you
> never explicitly enabled anything to wake up the memcg beast :)
> Balbir