Message-ID: <BANLkTinYGwRa_7uGzbYq+pW3T7jL-nQ7sA@mail.gmail.com>
Date:	Sun, 15 May 2011 11:59:14 -0400
From:	Andrew Lutomirski <luto@....edu>
To:	Wu Fengguang <fengguang.wu@...el.com>
Cc:	Minchan Kim <minchan.kim@...il.com>,
	Andi Kleen <andi@...stfloor.org>,
	"linux-mm@...ck.org" <linux-mm@...ck.org>,
	LKML <linux-kernel@...r.kernel.org>
Subject: Re: Kernel falls apart under light memory pressure (i.e. linking vmlinux)

On Sun, May 15, 2011 at 11:27 AM, Wu Fengguang <fengguang.wu@...el.com> wrote:
> On Sun, May 15, 2011 at 09:37:58AM +0800, Minchan Kim wrote:
>> On Sun, May 15, 2011 at 2:43 AM, Andi Kleen <andi@...stfloor.org> wrote:
>> > Copying back linux-mm.
>> >
>> >> Recently, we added following patch.
>> >> https://lkml.org/lkml/2011/4/26/129
>> >> If it's a culprit, the patch should solve the problem.
>> >
>> > It would be probably better to not do the allocations at all under
>> > memory pressure.  Even if the RA allocation doesn't go into reclaim
>>
>> Fair enough.
>> I think we can do it easily now.
>> If page_cache_alloc_readahead(ie, GFP_NORETRY) is fail, we can adjust
>> RA window size or turn off a while. The point is that we can use the
>> fail of __do_page_cache_readahead as sign of memory pressure.
>> Wu, What do you think?
>
> No, disabling readahead can hardly help.
>
> The sequential readahead memory consumption can be estimated by
>
>                2 * (number of concurrent read streams) * (readahead window size)
>
> And you can double that when there are two level of readaheads.
>
> Since there are hardly any concurrent read streams in Andy's case,
> the readahead memory consumption will be ignorable.
>
> Typically readahead thrashing will happen long before excessive
> GFP_NORETRY failures, so the reasonable solutions are to
>
> - shrink readahead window on readahead thrashing
>  (current readahead heuristic can somehow do this, and I have patches
>  to further improve it)
>
> - prevent abnormal GFP_NORETRY failures
>  (when there are many reclaimable pages)
>
>
> Andy's OOM memory dump (incorrect_oom_kill.txt.xz) shows that there are
>
> - 8MB   active+inactive file pages
> - 160MB active+inactive anon pages
> - 1GB   shmem pages
> - 1.4GB unevictable pages
>
> Hmm, why are there so many unevictable pages?  How come the shmem
> pages become unevictable when there are plenty of swap space?

I have no clue, but this patch (from Minchan, whitespace-damaged) seems to help:

diff --git a/mm/vmscan.c b/mm/vmscan.c
index f6b435c..4d24828 100644
--- a/mm/vmscan.c
+++ b/mm/vmscan.c
@@ -2251,6 +2251,10 @@ static bool sleeping_prematurely(pg_data_t
*pgdat, int order, long remaining,
       unsigned long balanced = 0;
       bool all_zones_ok = true;

+       /* If kswapd has been running too long, just sleep */
+       if (need_resched())
+               return false;
+
       /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
       if (remaining)
               return true;
@@ -2286,7 +2290,7 @@ static bool sleeping_prematurely(pg_data_t
*pgdat, int order, long remaining,
        * must be balanced
        */
       if (order)
-               return pgdat_balanced(pgdat, balanced, classzone_idx);
+               return !pgdat_balanced(pgdat, balanced, classzone_idx);
       else
               return !all_zones_ok;
 }

I haven't tested it very thoroughly, but it's survived much longer
than an unpatched kernel probably would have under moderate use.

I have no idea what the patch does :)

I'm happy to run any tests.  I'm also planning to upgrade from 2GB to
8GB RAM soon, which might change something.

--Andy
--
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