src/patches/suse-2.6.27.25/patches.fixes/aggressive-zone-reclaim.patch

   1 From: Nick Piggin <npiggin@suse.de>
   2 Subject: be more aggressive with zone reclaims
   3 References: bnc#476525
   4 Patch-mainline: no
   5
   6 The zone reclaim design is not very good for parallel allocations.
   7 The primary problem is that only one thread is allowed to perform
   8 zone-reclaim at a time. If another thread needs memory from that
   9 zone/node, then its zone-reclaim will fail and it will be forced
  10 to fall back to allocating from another zone.
  11
  12 Additionally, the default zone reclaim priority is insufficient
  13 for massively parallel allocations. Lower ZONE_RECLAIM_PRIORITY
  14 to fix it. This can result in higher latency spikes, but similar
  15 kind of page allocation latency can often be encountered as
  16 normal part of page reclaim when pagecache fills memory.
  17
  18 Signed-off-by: Petr Tesarik <ptesarik@suse.cz>
  19
  20 ---
  21  mm/vmscan.c |   12 ++++--------
  22  1 file changed, 4 insertions(+), 8 deletions(-)
  23
  24 --- linux-2.6.27-SLE11_BRANCH.orig/mm/vmscan.c  2008-10-20 17:24:19.000000000 +0200
  25 +++ linux-2.6.27-SLE11_BRANCH/mm/vmscan.c       2009-06-29 12:59:09.000000000 +0200
  26 @@ -1988,7 +1988,7 @@ int zone_reclaim_mode __read_mostly;
  27   * of a node considered for each zone_reclaim. 4 scans 1/16th of
  28   * a zone.
  29   */
  30 -#define ZONE_RECLAIM_PRIORITY 4
  31 +#define ZONE_RECLAIM_PRIORITY 0
  32
  33  /*
  34   * Percentage of pages in a zone that must be unmapped for zone_reclaim to
  35 @@ -2052,6 +2052,8 @@ static int __zone_reclaim(struct zone *z
  36
  37         slab_reclaimable = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
  38         if (slab_reclaimable > zone->min_slab_pages) {
  39 +               unsigned long lru_pages = zone_page_state(zone, NR_ACTIVE)
  40 +                                       + zone_page_state(zone, NR_INACTIVE);
  41                 /*
  42                  * shrink_slab() does not currently allow us to determine how
  43                  * many pages were freed in this zone. So we take the current
  44 @@ -2062,10 +2064,7 @@ static int __zone_reclaim(struct zone *z
  45                  * Note that shrink_slab will free memory on all zones and may
  46                  * take a long time.
  47                  */
  48 -               while (shrink_slab(sc.nr_scanned, gfp_mask, order) &&
  49 -                       zone_page_state(zone, NR_SLAB_RECLAIMABLE) >
  50 -                               slab_reclaimable - nr_pages)
  51 -                       ;
  52 +               shrink_slab(sc.nr_scanned, gfp_mask, lru_pages);
  53
  54                 /*
  55                  * Update nr_reclaimed by the number of slab pages we
  56 @@ -2120,10 +2119,7 @@ int zone_reclaim(struct zone *zone, gfp_
  57         if (node_state(node_id, N_CPU) && node_id != numa_node_id())
  58                 return 0;
  59
  60 -       if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
  61 -               return 0;
  62         ret = __zone_reclaim(zone, gfp_mask, order);
  63 -       zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
  64
  65         return ret;
  66  }