[lkml]   [2007]   [Sep]   [5]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
SubjectRe: [PATCH] prevent kswapd from freeing excessive amounts of lowmem
> On Wed, 05 Sep 2007 19:01:25 -0400 Rik van Riel <> wrote:
> The current VM can get itself into trouble fairly easily on systems
> with a small ZONE_HIGHMEM, which is common on i686 computers with
> 1GB of memory.
> On one side, page_alloc() will allocate down to zone->pages_low,
> while on the other side, kswapd() and balance_pgdat() will try
> to free memory from every zone, until every zone has more free
> pages than zone->pages_high.
> Highmem can be filled up to zone->pages_low with page tables,
> ramfs, vmalloc allocations and other unswappable things quite
> easily and without many bad side effects, since we still have
> a huge ZONE_NORMAL to do future allocations from.
> However, as long as the number of free pages in the highmem
> zone is below zone->pages_high, kswapd will continue swapping
> things out from ZONE_NORMAL, too!

crap. I guess suitably-fashioned mlock could do the same thing.

> Sami Farin managed to get his system into a stage where kswapd
> had freed about 700MB of low memory and was still "going strong".
> The attached patch will make kswapd stop paging out data from
> zones when there is more than enough memory free.

hm. Did highmem's all_unreclaimable get set? If so perhaps we could use
that in some way.

> We do go above
> zone->pages_high in order to keep pressure between zones equal
> in normal circumstances, but the patch should prevent the kind
> of excesses that made Sami's computer totally unusable.
> Please merge this into -mm.
> Signed-off-by: Rik van Riel <>
> [linux-2.6-excessive-pageout.patch text/x-patch (715B)]
> --- linux-2.6.22.noarch/mm/vmscan.c.excessive 2007-09-05 12:19:49.000000000 -0400
> +++ linux-2.6.22.noarch/mm/vmscan.c 2007-09-05 12:21:40.000000000 -0400
> @@ -1371,7 +1371,13 @@ loop_again:
> temp_priority[i] = priority;
> sc.nr_scanned = 0;
> note_zone_scanning_priority(zone, priority);
> - nr_reclaimed += shrink_zone(priority, zone, &sc);
> + /*
> + * We put equal pressure on every zone, unless one
> + * zone has way too many pages free already.
> + */
> + if (!zone_watermark_ok(zone, order, 8*zone->pages_high,
> + end_zone, 0))
> + nr_reclaimed += shrink_zone(priority, zone, &sc);
> reclaim_state->reclaimed_slab = 0;
> nr_slab = shrink_slab(sc.nr_scanned, GFP_KERNEL,
> lru_pages);

I guess for a very small upper zone and a very large lower zone this could
still put the scan balancing out of whack, fixable by a smarter version of
"8*zone->pages_high" but it doesn't seem very likely that this will affect
things much.

Why doesn't direct reclaim need similar treatment?

To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to
More majordomo info at
Please read the FAQ at

 \ /
  Last update: 2007-09-06 03:27    [W:0.065 / U:1.428 seconds]
©2003-2018 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site