]> git.karo-electronics.de Git - mv-sheeva.git/blobdiff - mm/swap.c
mm: deactivate invalidated pages
[mv-sheeva.git] / mm / swap.c
index 3f4854205b16ba0c39a8b815e00fe1b48f3472f1..4aea806d0d441c7d58d4bc8be83103ff743e5289 100644 (file)
--- a/mm/swap.c
+++ b/mm/swap.c
@@ -39,6 +39,7 @@ int page_cluster;
 
 static DEFINE_PER_CPU(struct pagevec[NR_LRU_LISTS], lru_add_pvecs);
 static DEFINE_PER_CPU(struct pagevec, lru_rotate_pvecs);
+static DEFINE_PER_CPU(struct pagevec, lru_deactivate_pvecs);
 
 /*
  * This path almost never happens for VM activity - pages are normally
@@ -56,17 +57,97 @@ static void __page_cache_release(struct page *page)
                del_page_from_lru(zone, page);
                spin_unlock_irqrestore(&zone->lru_lock, flags);
        }
+}
+
+static void __put_single_page(struct page *page)
+{
+       __page_cache_release(page);
        free_hot_cold_page(page, 0);
 }
 
-static void put_compound_page(struct page *page)
+static void __put_compound_page(struct page *page)
 {
-       page = compound_head(page);
-       if (put_page_testzero(page)) {
-               compound_page_dtor *dtor;
+       compound_page_dtor *dtor;
 
-               dtor = get_compound_page_dtor(page);
-               (*dtor)(page);
+       __page_cache_release(page);
+       dtor = get_compound_page_dtor(page);
+       (*dtor)(page);
+}
+
+static void put_compound_page(struct page *page)
+{
+       if (unlikely(PageTail(page))) {
+               /* __split_huge_page_refcount can run under us */
+               struct page *page_head = page->first_page;
+               smp_rmb();
+               /*
+                * If PageTail is still set after smp_rmb() we can be sure
+                * that the page->first_page we read wasn't a dangling pointer.
+                * See __split_huge_page_refcount() smp_wmb().
+                */
+               if (likely(PageTail(page) && get_page_unless_zero(page_head))) {
+                       unsigned long flags;
+                       /*
+                        * Verify that our page_head wasn't converted
+                        * to a a regular page before we got a
+                        * reference on it.
+                        */
+                       if (unlikely(!PageHead(page_head))) {
+                               /* PageHead is cleared after PageTail */
+                               smp_rmb();
+                               VM_BUG_ON(PageTail(page));
+                               goto out_put_head;
+                       }
+                       /*
+                        * Only run compound_lock on a valid PageHead,
+                        * after having it pinned with
+                        * get_page_unless_zero() above.
+                        */
+                       smp_mb();
+                       /* page_head wasn't a dangling pointer */
+                       flags = compound_lock_irqsave(page_head);
+                       if (unlikely(!PageTail(page))) {
+                               /* __split_huge_page_refcount run before us */
+                               compound_unlock_irqrestore(page_head, flags);
+                               VM_BUG_ON(PageHead(page_head));
+                       out_put_head:
+                               if (put_page_testzero(page_head))
+                                       __put_single_page(page_head);
+                       out_put_single:
+                               if (put_page_testzero(page))
+                                       __put_single_page(page);
+                               return;
+                       }
+                       VM_BUG_ON(page_head != page->first_page);
+                       /*
+                        * We can release the refcount taken by
+                        * get_page_unless_zero now that
+                        * split_huge_page_refcount is blocked on the
+                        * compound_lock.
+                        */
+                       if (put_page_testzero(page_head))
+                               VM_BUG_ON(1);
+                       /* __split_huge_page_refcount will wait now */
+                       VM_BUG_ON(atomic_read(&page->_count) <= 0);
+                       atomic_dec(&page->_count);
+                       VM_BUG_ON(atomic_read(&page_head->_count) <= 0);
+                       compound_unlock_irqrestore(page_head, flags);
+                       if (put_page_testzero(page_head)) {
+                               if (PageHead(page_head))
+                                       __put_compound_page(page_head);
+                               else
+                                       __put_single_page(page_head);
+                       }
+               } else {
+                       /* page_head is a dangling pointer */
+                       VM_BUG_ON(PageTail(page));
+                       goto out_put_single;
+               }
+       } else if (put_page_testzero(page)) {
+               if (PageHead(page))
+                       __put_compound_page(page);
+               else
+                       __put_single_page(page);
        }
 }
 
@@ -75,7 +156,7 @@ void put_page(struct page *page)
        if (unlikely(PageCompound(page)))
                put_compound_page(page);
        else if (put_page_testzero(page))
-               __page_cache_release(page);
+               __put_single_page(page);
 }
 EXPORT_SYMBOL(put_page);
 
@@ -266,6 +347,60 @@ void add_page_to_unevictable_list(struct page *page)
        spin_unlock_irq(&zone->lru_lock);
 }
 
+/*
+ * If the page can not be invalidated, it is moved to the
+ * inactive list to speed up its reclaim.  It is moved to the
+ * head of the list, rather than the tail, to give the flusher
+ * threads some time to write it out, as this is much more
+ * effective than the single-page writeout from reclaim.
+ */
+static void lru_deactivate(struct page *page, struct zone *zone)
+{
+       int lru, file;
+
+       if (!PageLRU(page) || !PageActive(page))
+               return;
+
+       /* Some processes are using the page */
+       if (page_mapped(page))
+               return;
+
+       file = page_is_file_cache(page);
+       lru = page_lru_base_type(page);
+       del_page_from_lru_list(zone, page, lru + LRU_ACTIVE);
+       ClearPageActive(page);
+       ClearPageReferenced(page);
+       add_page_to_lru_list(zone, page, lru);
+       __count_vm_event(PGDEACTIVATE);
+
+       update_page_reclaim_stat(zone, page, file, 0);
+}
+
+static void ____pagevec_lru_deactivate(struct pagevec *pvec)
+{
+       int i;
+       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_irq(&zone->lru_lock);
+                       zone = pagezone;
+                       spin_lock_irq(&zone->lru_lock);
+               }
+               lru_deactivate(page, zone);
+       }
+       if (zone)
+               spin_unlock_irq(&zone->lru_lock);
+
+       release_pages(pvec->pages, pvec->nr, pvec->cold);
+       pagevec_reinit(pvec);
+}
+
+
 /*
  * Drain pages out of the cpu's pagevecs.
  * Either "cpu" is the current CPU, and preemption has already been
@@ -292,6 +427,29 @@ static void drain_cpu_pagevecs(int cpu)
                pagevec_move_tail(pvec);
                local_irq_restore(flags);
        }
+
+       pvec = &per_cpu(lru_deactivate_pvecs, cpu);
+       if (pagevec_count(pvec))
+               ____pagevec_lru_deactivate(pvec);
+}
+
+/**
+ * deactivate_page - forcefully deactivate a page
+ * @page: page to deactivate
+ *
+ * This function hints the VM that @page is a good reclaim candidate,
+ * for example if its invalidation fails due to the page being dirty
+ * or under writeback.
+ */
+void deactivate_page(struct page *page)
+{
+       if (likely(get_page_unless_zero(page))) {
+               struct pagevec *pvec = &get_cpu_var(lru_deactivate_pvecs);
+
+               if (!pagevec_add(pvec, page))
+                       ____pagevec_lru_deactivate(pvec);
+               put_cpu_var(lru_deactivate_pvecs);
+       }
 }
 
 void lru_add_drain(void)
@@ -399,6 +557,43 @@ void __pagevec_release(struct pagevec *pvec)
 
 EXPORT_SYMBOL(__pagevec_release);
 
+/* used by __split_huge_page_refcount() */
+void lru_add_page_tail(struct zone* zone,
+                      struct page *page, struct page *page_tail)
+{
+       int active;
+       enum lru_list lru;
+       const int file = 0;
+       struct list_head *head;
+
+       VM_BUG_ON(!PageHead(page));
+       VM_BUG_ON(PageCompound(page_tail));
+       VM_BUG_ON(PageLRU(page_tail));
+       VM_BUG_ON(!spin_is_locked(&zone->lru_lock));
+
+       SetPageLRU(page_tail);
+
+       if (page_evictable(page_tail, NULL)) {
+               if (PageActive(page)) {
+                       SetPageActive(page_tail);
+                       active = 1;
+                       lru = LRU_ACTIVE_ANON;
+               } else {
+                       active = 0;
+                       lru = LRU_INACTIVE_ANON;
+               }
+               update_page_reclaim_stat(zone, page_tail, file, active);
+               if (likely(PageLRU(page)))
+                       head = page->lru.prev;
+               else
+                       head = &zone->lru[lru].list;
+               __add_page_to_lru_list(zone, page_tail, lru, head);
+       } else {
+               SetPageUnevictable(page_tail);
+               add_page_to_lru_list(zone, page_tail, LRU_UNEVICTABLE);
+       }
+}
+
 /*
  * Add the passed pages to the LRU, then drop the caller's refcount
  * on them.  Reinitialises the caller's pagevec.