mm: batch activate_page() to reduce lock contention
authorShaohua Li <shaohua.li@intel.com>
Wed, 25 May 2011 00:12:55 +0000 (17:12 -0700)
committerLinus Torvalds <torvalds@linux-foundation.org>
Wed, 25 May 2011 15:39:37 +0000 (08:39 -0700)
The zone->lru_lock is heavily contented in workload where activate_page()
is frequently used.  We could do batch activate_page() to reduce the lock
contention.  The batched pages will be added into zone list when the pool
is full or page reclaim is trying to drain them.

For example, in a 4 socket 64 CPU system, create a sparse file and 64
processes, processes shared map to the file.  Each process read access the
whole file and then exit.  The process exit will do unmap_vmas() and cause
a lot of activate_page() call.  In such workload, we saw about 58% total
time reduction with below patch.  Other workloads with a lot of
activate_page also benefits a lot too.

Andrew Morton suggested activate_page() and putback_lru_pages() should
follow the same path to active pages, but this is hard to implement (see
commit 7a608572a282a ("Revert "mm: batch activate_page() to reduce lock
contention")).  On the other hand, do we really need putback_lru_pages()
to follow the same path?  I tested several FIO/FFSB benchmark (about 20
scripts for each benchmark) in 3 machines here from 2 sockets to 4
sockets.  My test doesn't show anything significant with/without below
patch (there is slight difference but mostly some noise which we found
even without below patch before).  Below patch basically returns to the
same as my first post.

I tested some microbenchmarks:
  case-anon-cow-rand-mt         0.58%
  case-anon-cow-rand           -3.30%
  case-anon-cow-seq-mt         -0.51%
  case-anon-cow-seq            -5.68%
  case-anon-r-rand-mt           0.23%
  case-anon-r-rand              0.81%
  case-anon-r-seq-mt           -0.71%
  case-anon-r-seq              -1.99%
  case-anon-rx-rand-mt          2.11%
  case-anon-rx-seq-mt           3.46%
  case-anon-w-rand-mt          -0.03%
  case-anon-w-rand             -0.50%
  case-anon-w-seq-mt           -1.08%
  case-anon-w-seq              -0.12%
  case-anon-wx-rand-mt         -5.02%
  case-anon-wx-seq-mt          -1.43%
  case-fork                     1.65%
  case-fork-sleep              -0.07%
  case-fork-withmem             1.39%
  case-hugetlb                 -0.59%
  case-lru-file-mmap-read-mt   -0.54%
  case-lru-file-mmap-read       0.61%
  case-lru-file-mmap-read-rand -2.24%
  case-lru-file-readonce       -0.64%
  case-lru-file-readtwice     -11.69%
  case-lru-memcg               -1.35%
  case-mmap-pread-rand-mt       1.88%
  case-mmap-pread-rand        -15.26%
  case-mmap-pread-seq-mt        0.89%
  case-mmap-pread-seq         -69.72%
  case-mmap-xread-rand-mt       0.71%
  case-mmap-xread-seq-mt        0.38%

The most significent are:
  case-lru-file-readtwice     -11.69%
  case-mmap-pread-rand        -15.26%
  case-mmap-pread-seq         -69.72%

which use activate_page a lot.  others are basically variations because
each run has slightly difference.

In UP case, 'size mm/swap.o'
before the two patches:
   text    data     bss     dec     hex filename
   6466     896       4    7366    1cc6 mm/swap.o
after the two patches:
   text    data     bss     dec     hex filename
   6343     896       4    7243    1c4b mm/swap.o

Signed-off-by: Shaohua Li <shaohua.li@intel.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Hiroyuki Kamezawa <kamezawa.hiroyuki@gmail.com>
Cc: Andi Kleen <andi@firstfloor.org>
Cc: Minchan Kim <minchan.kim@gmail.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Hugh Dickins <hughd@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
mm/swap.c

index 2f365d1a4bb2c33be69036121c325f2c4f103604..3a442f18b0b3dab5acfd99b5fc4e85389b9b5439 100644 (file)
--- a/mm/swap.c
+++ b/mm/swap.c
@@ -272,14 +272,10 @@ static void update_page_reclaim_stat(struct zone *zone, struct page *page,
                memcg_reclaim_stat->recent_rotated[file]++;
 }
 
-/*
- * FIXME: speed this up?
- */
-void activate_page(struct page *page)
+static void __activate_page(struct page *page, void *arg)
 {
        struct zone *zone = page_zone(page);
 
-       spin_lock_irq(&zone->lru_lock);
        if (PageLRU(page) && !PageActive(page) && !PageUnevictable(page)) {
                int file = page_is_file_cache(page);
                int lru = page_lru_base_type(page);
@@ -292,8 +288,45 @@ void activate_page(struct page *page)
 
                update_page_reclaim_stat(zone, page, file, 1);
        }
+}
+
+#ifdef CONFIG_SMP
+static DEFINE_PER_CPU(struct pagevec, activate_page_pvecs);
+
+static void activate_page_drain(int cpu)
+{
+       struct pagevec *pvec = &per_cpu(activate_page_pvecs, cpu);
+
+       if (pagevec_count(pvec))
+               pagevec_lru_move_fn(pvec, __activate_page, NULL);
+}
+
+void activate_page(struct page *page)
+{
+       if (PageLRU(page) && !PageActive(page) && !PageUnevictable(page)) {
+               struct pagevec *pvec = &get_cpu_var(activate_page_pvecs);
+
+               page_cache_get(page);
+               if (!pagevec_add(pvec, page))
+                       pagevec_lru_move_fn(pvec, __activate_page, NULL);
+               put_cpu_var(activate_page_pvecs);
+       }
+}
+
+#else
+static inline void activate_page_drain(int cpu)
+{
+}
+
+void activate_page(struct page *page)
+{
+       struct zone *zone = page_zone(page);
+
+       spin_lock_irq(&zone->lru_lock);
+       __activate_page(page, NULL);
        spin_unlock_irq(&zone->lru_lock);
 }
+#endif
 
 /*
  * Mark a page as having seen activity.
@@ -464,6 +497,8 @@ static void drain_cpu_pagevecs(int cpu)
        pvec = &per_cpu(lru_deactivate_pvecs, cpu);
        if (pagevec_count(pvec))
                pagevec_lru_move_fn(pvec, lru_deactivate_fn, NULL);
+
+       activate_page_drain(cpu);
 }
 
 /**