mm: use pagevec to rotate reclaimable page
authorHisashi Hifumi <hifumi.hisashi@oss.ntt.co.jp>
Tue, 16 Oct 2007 08:24:52 +0000 (01:24 -0700)
committerLinus Torvalds <torvalds@woody.linux-foundation.org>
Tue, 16 Oct 2007 16:42:54 +0000 (09:42 -0700)
While running some memory intensive load, system response deteriorated just
after swap-out started.

The cause of this problem is that when a PG_reclaim page is moved to the tail
of the inactive LRU list in rotate_reclaimable_page(), lru_lock spin lock is
acquired every page writeback .  This deteriorates system performance and
makes interrupt hold off time longer when swap-out started.

Following patch solves this problem.  I use pagevec in rotating reclaimable
pages to mitigate LRU spin lock contention and reduce interrupt hold off time.

I did a test that allocating and touching pages in multiple processes, and
pinging to the test machine in flooding mode to measure response under memory
intensive load.

The test result is:

-2.6.23-rc5
--- testmachine ping statistics ---
3000 packets transmitted, 3000 received, 0% packet loss, time 53222ms
rtt min/avg/max/mdev = 0.074/0.652/172.228/7.176 ms, pipe 11, ipg/ewma
17.746/0.092 ms

-2.6.23-rc5-patched
--- testmachine ping statistics ---
3000 packets transmitted, 3000 received, 0% packet loss, time 51924ms
rtt min/avg/max/mdev = 0.072/0.108/3.884/0.114 ms, pipe 2, ipg/ewma
17.314/0.091 ms

Max round-trip-time was improved.

The test machine spec is that 4CPU(3.16GHz, Hyper-threading enabled)
8GB memory , 8GB swap.

I did ping test again to observe performance deterioration caused by taking
a ref.

-2.6.23-rc6-with-modifiedpatch
--- testmachine ping statistics ---
3000 packets transmitted, 3000 received, 0% packet loss, time 53386ms
rtt min/avg/max/mdev = 0.074/0.110/4.716/0.147 ms, pipe 2, ipg/ewma 17.801/0.129 ms

The result for my original patch is as follows.

-2.6.23-rc5-with-originalpatch
--- testmachine ping statistics ---
3000 packets transmitted, 3000 received, 0% packet loss, time 51924ms
rtt min/avg/max/mdev = 0.072/0.108/3.884/0.114 ms, pipe 2, ipg/ewma 17.314/0.091 ms

The influence to response was small.

[akpm@linux-foundation.org: fix uninitalised var warning]
[hugh@veritas.com: fix locking]
[randy.dunlap@oracle.com: fix function declaration]
[hugh@veritas.com: fix BUG at include/linux/mm.h:220!]
[hugh@veritas.com: kill redundancy in rotate_reclaimable_page]
[hugh@veritas.com: move_tail_pages into lru_add_drain]
Signed-off-by: Hisashi Hifumi <hifumi.hisashi@oss.ntt.co.jp>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
mm/swap.c

index d93fa1e..d034b21 100644 (file)
--- a/mm/swap.c
+++ b/mm/swap.c
 /* How many pages do we try to swap or page in/out together? */
 int page_cluster;
 
+static DEFINE_PER_CPU(struct pagevec, lru_add_pvecs) = { 0, };
+static DEFINE_PER_CPU(struct pagevec, lru_add_active_pvecs) = { 0, };
+static DEFINE_PER_CPU(struct pagevec, lru_rotate_pvecs) = { 0, };
+
 /*
  * This path almost never happens for VM activity - pages are normally
  * freed via pagevecs.  But it gets used by networking.
@@ -92,23 +96,47 @@ void put_pages_list(struct list_head *pages)
 EXPORT_SYMBOL(put_pages_list);
 
 /*
+ * pagevec_move_tail() must be called with IRQ disabled.
+ * Otherwise this may cause nasty races.
+ */
+static void pagevec_move_tail(struct pagevec *pvec)
+{
+       int i;
+       int pgmoved = 0;
+       struct zone *zone = NULL;
+
+       for (i = 0; i < pagevec_count(pvec); i++) {
+               struct page *page = pvec->pages[i];
+               struct zone *pagezone = page_zone(page);
+
+               if (pagezone != zone) {
+                       if (zone)
+                               spin_unlock(&zone->lru_lock);
+                       zone = pagezone;
+                       spin_lock(&zone->lru_lock);
+               }
+               if (PageLRU(page) && !PageActive(page)) {
+                       list_move_tail(&page->lru, &zone->inactive_list);
+                       pgmoved++;
+               }
+       }
+       if (zone)
+               spin_unlock(&zone->lru_lock);
+       __count_vm_events(PGROTATED, pgmoved);
+       release_pages(pvec->pages, pvec->nr, pvec->cold);
+       pagevec_reinit(pvec);
+}
+
+/*
  * Writeback is about to end against a page which has been marked for immediate
  * reclaim.  If it still appears to be reclaimable, move it to the tail of the
- * inactive list.  The page still has PageWriteback set, which will pin it.
- *
- * We don't expect many pages to come through here, so don't bother batching
- * things up.
- *
- * To avoid placing the page at the tail of the LRU while PG_writeback is still
- * set, this function will clear PG_writeback before performing the page
- * motion.  Do that inside the lru lock because once PG_writeback is cleared
- * we may not touch the page.
+ * inactive list.
  *
  * Returns zero if it cleared PG_writeback.
  */
 int rotate_reclaimable_page(struct page *page)
 {
-       struct zone *zone;
+       struct pagevec *pvec;
        unsigned long flags;
 
        if (PageLocked(page))
@@ -120,15 +148,16 @@ int rotate_reclaimable_page(struct page *page)
        if (!PageLRU(page))
                return 1;
 
-       zone = page_zone(page);
-       spin_lock_irqsave(&zone->lru_lock, flags);
-       if (PageLRU(page) && !PageActive(page)) {
-               list_move_tail(&page->lru, &zone->inactive_list);
-               __count_vm_event(PGROTATED);
-       }
+       page_cache_get(page);
+       local_irq_save(flags);
+       pvec = &__get_cpu_var(lru_rotate_pvecs);
+       if (!pagevec_add(pvec, page))
+               pagevec_move_tail(pvec);
+       local_irq_restore(flags);
+
        if (!test_clear_page_writeback(page))
                BUG();
-       spin_unlock_irqrestore(&zone->lru_lock, flags);
+
        return 0;
 }
 
@@ -172,9 +201,6 @@ EXPORT_SYMBOL(mark_page_accessed);
  * lru_cache_add: add a page to the page lists
  * @page: the page to add
  */
-static DEFINE_PER_CPU(struct pagevec, lru_add_pvecs) = { 0, };
-static DEFINE_PER_CPU(struct pagevec, lru_add_active_pvecs) = { 0, };
-
 void fastcall lru_cache_add(struct page *page)
 {
        struct pagevec *pvec = &get_cpu_var(lru_add_pvecs);
@@ -195,21 +221,37 @@ void fastcall lru_cache_add_active(struct page *page)
        put_cpu_var(lru_add_active_pvecs);
 }
 
-static void __lru_add_drain(int cpu)
+/*
+ * Drain pages out of the cpu's pagevecs.
+ * Either "cpu" is the current CPU, and preemption has already been
+ * disabled; or "cpu" is being hot-unplugged, and is already dead.
+ */
+static void drain_cpu_pagevecs(int cpu)
 {
-       struct pagevec *pvec = &per_cpu(lru_add_pvecs, cpu);
+       struct pagevec *pvec;
 
-       /* CPU is dead, so no locking needed. */
+       pvec = &per_cpu(lru_add_pvecs, cpu);
        if (pagevec_count(pvec))
                __pagevec_lru_add(pvec);
+
        pvec = &per_cpu(lru_add_active_pvecs, cpu);
        if (pagevec_count(pvec))
                __pagevec_lru_add_active(pvec);
+
+       pvec = &per_cpu(lru_rotate_pvecs, cpu);
+       if (pagevec_count(pvec)) {
+               unsigned long flags;
+
+               /* No harm done if a racing interrupt already did this */
+               local_irq_save(flags);
+               pagevec_move_tail(pvec);
+               local_irq_restore(flags);
+       }
 }
 
 void lru_add_drain(void)
 {
-       __lru_add_drain(get_cpu());
+       drain_cpu_pagevecs(get_cpu());
        put_cpu();
 }
 
@@ -256,6 +298,7 @@ void release_pages(struct page **pages, int nr, int cold)
        int i;
        struct pagevec pages_to_free;
        struct zone *zone = NULL;
+       unsigned long uninitialized_var(flags);
 
        pagevec_init(&pages_to_free, cold);
        for (i = 0; i < nr; i++) {
@@ -263,7 +306,7 @@ void release_pages(struct page **pages, int nr, int cold)
 
                if (unlikely(PageCompound(page))) {
                        if (zone) {
-                               spin_unlock_irq(&zone->lru_lock);
+                               spin_unlock_irqrestore(&zone->lru_lock, flags);
                                zone = NULL;
                        }
                        put_compound_page(page);
@@ -277,9 +320,10 @@ void release_pages(struct page **pages, int nr, int cold)
                        struct zone *pagezone = page_zone(page);
                        if (pagezone != zone) {
                                if (zone)
-                                       spin_unlock_irq(&zone->lru_lock);
+                                       spin_unlock_irqrestore(&zone->lru_lock,
+                                                                       flags);
                                zone = pagezone;
-                               spin_lock_irq(&zone->lru_lock);
+                               spin_lock_irqsave(&zone->lru_lock, flags);
                        }
                        VM_BUG_ON(!PageLRU(page));
                        __ClearPageLRU(page);
@@ -288,7 +332,7 @@ void release_pages(struct page **pages, int nr, int cold)
 
                if (!pagevec_add(&pages_to_free, page)) {
                        if (zone) {
-                               spin_unlock_irq(&zone->lru_lock);
+                               spin_unlock_irqrestore(&zone->lru_lock, flags);
                                zone = NULL;
                        }
                        __pagevec_free(&pages_to_free);
@@ -296,7 +340,7 @@ void release_pages(struct page **pages, int nr, int cold)
                }
        }
        if (zone)
-               spin_unlock_irq(&zone->lru_lock);
+               spin_unlock_irqrestore(&zone->lru_lock, flags);
 
        pagevec_free(&pages_to_free);
 }
@@ -489,7 +533,7 @@ static int cpu_swap_callback(struct notifier_block *nfb,
        if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
                atomic_add(*committed, &vm_committed_space);
                *committed = 0;
-               __lru_add_drain((long)hcpu);
+               drain_cpu_pagevecs((long)hcpu);
        }
        return NOTIFY_OK;
 }