On Fri, Aug 20, 2010 at 6:35 PM, Mel Gorman <mel@xxxxxxxxx> wrote:
> On Fri, Aug 20, 2010 at 01:34:47PM +0800, Wu Fengguang wrote:
>> You do run lots of tasks: kernel_stack=1880kB.
>>
>> And you have lots of free memory, page reclaim has never run, so
>> inactive_anon=0. This is where compaction is different from vmscan.
>> In vmscan, inactive_anon is reasonably large, and will only be
>> compared directly with isolated_anon.
>>
>
> True, the key observation here was that compaction is being run via the
> proc trigger. Normally it would be run as part of the direct reclaim
> path when kswapd would already be awake. too_many_isolated() needs to be
> different for compaction to take the whole system into account. What
> would be the best alternative? Here is one possibility. A reasonable
> alternative would be that when inactive < active that isolated can't be
> more than num_online_cpus() * 2 (i.e. one compactor per online cpu).
>
> diff --git a/mm/compaction.c b/mm/compaction.c
> index 94cce51..1e000b7 100644
> --- a/mm/compaction.c
> +++ b/mm/compaction.c
> @@ -215,14 +215,16 @@ static void acct_isolated(struct zone *zone, struct compact_control *cc)
> static bool too_many_isolated(struct zone *zone)
> {
>
> - unsigned long inactive, isolated;
> + unsigned long active, inactive, isolated;
>
> + active = zone_page_state(zone, NR_ACTIVE_FILE) +
> + zone_page_state(zone, NR_INACTIVE_ANON);
> inactive = zone_page_state(zone, NR_INACTIVE_FILE) +
> zone_page_state(zone, NR_INACTIVE_ANON);
> isolated = zone_page_state(zone, NR_ISOLATED_FILE) +
> zone_page_state(zone, NR_ISOLATED_ANON);
>
> - return isolated > inactive;
> + return (inactive > active) ? isolated > inactive : false;
> }
>
> /*
>
1. active : 1000 inactive : 1000
2. parallel reclaiming -> active : 1000 inactive : 500 isolated : 500
3. too_many_isolated return false.
But in this case, there are already many isolated pages. So it should
return true.
How about this?
too_many_isolated()
{
return (isolated > nr_zones * nr_nodes * nr_online_cpu *
SWAP_CLUSTER_MAX);
}
--
Kind regards,
Minchan Kim
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@xxxxxxxxxx For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href