patch-2_6_7-vs1_9_1_12
[linux-2.6.git] / mm / vmscan.c
index f789212..e5f0b09 100644 (file)
@@ -153,20 +153,23 @@ static int shrink_slab(unsigned long scanned, unsigned int gfp_mask)
                delta *= (*shrinker->shrinker)(0, gfp_mask);
                do_div(delta, pages + 1);
                shrinker->nr += delta;
-               if (shrinker->nr > SHRINK_BATCH) {
-                       long nr_to_scan = shrinker->nr;
-
-                       shrinker->nr = 0;
-                       mod_page_state(slabs_scanned, nr_to_scan);
-                       while (nr_to_scan) {
-                               long this_scan = nr_to_scan;
-
-                               if (this_scan > 128)
-                                       this_scan = 128;
-                               (*shrinker->shrinker)(this_scan, gfp_mask);
-                               nr_to_scan -= this_scan;
-                               cond_resched();
-                       }
+               if (shrinker->nr < 0)
+                       shrinker->nr = LONG_MAX;        /* It wrapped! */
+
+               if (shrinker->nr <= SHRINK_BATCH)
+                       continue;
+               while (shrinker->nr) {
+                       long this_scan = shrinker->nr;
+                       int shrink_ret;
+
+                       if (this_scan > 128)
+                               this_scan = 128;
+                       shrink_ret = (*shrinker->shrinker)(this_scan, gfp_mask);
+                       mod_page_state(slabs_scanned, this_scan);
+                       shrinker->nr -= this_scan;
+                       if (shrink_ret == -1)
+                               break;
+                       cond_resched();
                }
        }
        up(&shrinker_sem);
@@ -237,23 +240,113 @@ static void handle_write_error(struct address_space *mapping,
        unlock_page(page);
 }
 
+/* possible outcome of pageout() */
+typedef enum {
+       /* failed to write page out, page is locked */
+       PAGE_KEEP,
+       /* move page to the active list, page is locked */
+       PAGE_ACTIVATE,
+       /* page has been sent to the disk successfully, page is unlocked */
+       PAGE_SUCCESS,
+       /* page is clean and locked */
+       PAGE_CLEAN,
+} pageout_t;
+
 /*
- * shrink_list returns the number of reclaimed pages
+ * pageout is called by shrink_list() for each dirty page. Calls ->writepage().
  */
-static int
-shrink_list(struct list_head *page_list, unsigned int gfp_mask,
-               int *nr_scanned, int do_writepage)
+static pageout_t pageout(struct page *page, struct address_space *mapping)
+{
+       /*
+        * If the page is dirty, only perform writeback if that write
+        * will be non-blocking.  To prevent this allocation from being
+        * stalled by pagecache activity.  But note that there may be
+        * stalls if we need to run get_block().  We could test
+        * PagePrivate for that.
+        *
+        * If this process is currently in generic_file_write() against
+        * this page's queue, we can perform writeback even if that
+        * will block.
+        *
+        * If the page is swapcache, write it back even if that would
+        * block, for some throttling. This happens by accident, because
+        * swap_backing_dev_info is bust: it doesn't reflect the
+        * congestion state of the swapdevs.  Easy to fix, if needed.
+        * See swapfile.c:page_queue_congested().
+        */
+       if (!is_page_cache_freeable(page))
+               return PAGE_KEEP;
+       if (!mapping)
+               return PAGE_KEEP;
+       if (mapping->a_ops->writepage == NULL)
+               return PAGE_ACTIVATE;
+       if (!may_write_to_queue(mapping->backing_dev_info))
+               return PAGE_KEEP;
+
+       if (clear_page_dirty_for_io(page)) {
+               int res;
+               struct writeback_control wbc = {
+                       .sync_mode = WB_SYNC_NONE,
+                       .nr_to_write = SWAP_CLUSTER_MAX,
+                       .nonblocking = 1,
+                       .for_reclaim = 1,
+               };
+
+               SetPageReclaim(page);
+               res = mapping->a_ops->writepage(page, &wbc);
+               if (res < 0)
+                       handle_write_error(mapping, page, res);
+               if (res == WRITEPAGE_ACTIVATE) {
+                       ClearPageReclaim(page);
+                       return PAGE_ACTIVATE;
+               }
+               if (!PageWriteback(page)) {
+                       /* synchronous write or broken a_ops? */
+                       ClearPageReclaim(page);
+               }
+
+               return PAGE_SUCCESS;
+       }
+
+       return PAGE_CLEAN;
+}
+
+struct scan_control {
+       /* Ask refill_inactive_zone, or shrink_cache to scan this many pages */
+       unsigned long nr_to_scan;
+
+       /* Incremented by the number of inactive pages that were scanned */
+       unsigned long nr_scanned;
+
+       /* Incremented by the number of pages reclaimed */
+       unsigned long nr_reclaimed;
+
+       unsigned long nr_mapped;        /* From page_state */
+
+       /* Ask shrink_caches, or shrink_zone to scan at this priority */
+       unsigned int priority;
+
+       /* This context's GFP mask */
+       unsigned int gfp_mask;
+
+       int may_writepage;
+};
+
+/*
+ * shrink_list adds the number of reclaimed pages to sc->nr_reclaimed
+ */
+static int shrink_list(struct list_head *page_list, struct scan_control *sc)
 {
-       struct address_space *mapping;
        LIST_HEAD(ret_pages);
        struct pagevec freed_pvec;
        int pgactivate = 0;
-       int ret = 0;
+       int reclaimed = 0;
 
        cond_resched();
 
        pagevec_init(&freed_pvec, 1);
        while (!list_empty(page_list)) {
+               struct address_space *mapping;
                struct page *page;
                int may_enter_fs;
                int referenced;
@@ -264,26 +357,24 @@ shrink_list(struct list_head *page_list, unsigned int gfp_mask,
                if (TestSetPageLocked(page))
                        goto keep;
 
-               /* Double the slab pressure for mapped and swapcache pages */
-               if (page_mapped(page) || PageSwapCache(page))
-                       (*nr_scanned)++;
-
                BUG_ON(PageActive(page));
 
                if (PageWriteback(page))
                        goto keep_locked;
 
-               rmap_lock(page);
+               sc->nr_scanned++;
+               /* Double the slab pressure for mapped and swapcache pages */
+               if (page_mapped(page) || PageSwapCache(page))
+                       sc->nr_scanned++;
+
+               page_map_lock(page);
                referenced = page_referenced(page);
                if (referenced && page_mapping_inuse(page)) {
                        /* In active use or really unfreeable.  Activate it. */
-                       rmap_unlock(page);
+                       page_map_unlock(page);
                        goto activate_locked;
                }
 
-               mapping = page_mapping(page);
-               may_enter_fs = (gfp_mask & __GFP_FS);
-
 #ifdef CONFIG_SWAP
                /*
                 * Anonymous process memory has backing store?
@@ -292,17 +383,17 @@ shrink_list(struct list_head *page_list, unsigned int gfp_mask,
                 * XXX: implement swap clustering ?
                 */
                if (PageAnon(page) && !PageSwapCache(page)) {
-                       rmap_unlock(page);
+                       page_map_unlock(page);
                        if (!add_to_swap(page))
                                goto activate_locked;
-                       rmap_lock(page);
-               }
-               if (PageSwapCache(page)) {
-                       mapping = &swapper_space;
-                       may_enter_fs = (gfp_mask & __GFP_IO);
+                       page_map_lock(page);
                }
 #endif /* CONFIG_SWAP */
 
+               mapping = page_mapping(page);
+               may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
+                       (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
+
                /*
                 * The page is mapped into the page tables of one or more
                 * processes. Try to unmap it here.
@@ -310,71 +401,45 @@ shrink_list(struct list_head *page_list, unsigned int gfp_mask,
                if (page_mapped(page) && mapping) {
                        switch (try_to_unmap(page)) {
                        case SWAP_FAIL:
-                               rmap_unlock(page);
+                               page_map_unlock(page);
                                goto activate_locked;
                        case SWAP_AGAIN:
-                               rmap_unlock(page);
+                               page_map_unlock(page);
                                goto keep_locked;
                        case SWAP_SUCCESS:
                                ; /* try to free the page below */
                        }
                }
-               rmap_unlock(page);
+               page_map_unlock(page);
 
-               /*
-                * If the page is dirty, only perform writeback if that write
-                * will be non-blocking.  To prevent this allocation from being
-                * stalled by pagecache activity.  But note that there may be
-                * stalls if we need to run get_block().  We could test
-                * PagePrivate for that.
-                *
-                * If this process is currently in generic_file_write() against
-                * this page's queue, we can perform writeback even if that
-                * will block.
-                *
-                * If the page is swapcache, write it back even if that would
-                * block, for some throttling. This happens by accident, because
-                * swap_backing_dev_info is bust: it doesn't reflect the
-                * congestion state of the swapdevs.  Easy to fix, if needed.
-                * See swapfile.c:page_queue_congested().
-                */
                if (PageDirty(page)) {
                        if (referenced)
                                goto keep_locked;
-                       if (!is_page_cache_freeable(page))
-                               goto keep_locked;
-                       if (!mapping)
-                               goto keep_locked;
-                       if (mapping->a_ops->writepage == NULL)
-                               goto activate_locked;
                        if (!may_enter_fs)
                                goto keep_locked;
-                       if (!may_write_to_queue(mapping->backing_dev_info))
+                       if (laptop_mode && !sc->may_writepage)
                                goto keep_locked;
-                       if (laptop_mode && !do_writepage)
+
+                       /* Page is dirty, try to write it out here */
+                       switch(pageout(page, mapping)) {
+                       case PAGE_KEEP:
                                goto keep_locked;
-                       if (clear_page_dirty_for_io(page)) {
-                               int res;
-                               struct writeback_control wbc = {
-                                       .sync_mode = WB_SYNC_NONE,
-                                       .nr_to_write = SWAP_CLUSTER_MAX,
-                                       .nonblocking = 1,
-                                       .for_reclaim = 1,
-                               };
-
-                               SetPageReclaim(page);
-                               res = mapping->a_ops->writepage(page, &wbc);
-                               if (res < 0)
-                                       handle_write_error(mapping, page, res);
-                               if (res == WRITEPAGE_ACTIVATE) {
-                                       ClearPageReclaim(page);
-                                       goto activate_locked;
-                               }
-                               if (!PageWriteback(page)) {
-                                       /* synchronous write or broken a_ops? */
-                                       ClearPageReclaim(page);
-                               }
-                               goto keep;
+                       case PAGE_ACTIVATE:
+                               goto activate_locked;
+                       case PAGE_SUCCESS:
+                               if (PageWriteback(page) || PageDirty(page))
+                                       goto keep;
+                               /*
+                                * A synchronous write - probably a ramdisk.  Go
+                                * ahead and try to reclaim the page.
+                                */
+                               if (TestSetPageLocked(page))
+                                       goto keep;
+                               if (PageDirty(page) || PageWriteback(page))
+                                       goto keep_locked;
+                               mapping = page_mapping(page);
+                       case PAGE_CLEAN:
+                               ; /* try to free the page below */
                        }
                }
 
@@ -396,11 +461,11 @@ shrink_list(struct list_head *page_list, unsigned int gfp_mask,
                 * the pages which were not successfully invalidated in
                 * truncate_complete_page().  We try to drop those buffers here
                 * and if that worked, and the page is no longer mapped into
-                * process address space (page_count == 0) it can be freed.
+                * process address space (page_count == 1) it can be freed.
                 * Otherwise, leave the page on the LRU so it is swappable.
                 */
                if (PagePrivate(page)) {
-                       if (!try_to_release_page(page, gfp_mask))
+                       if (!try_to_release_page(page, sc->gfp_mask))
                                goto activate_locked;
                        if (!mapping && page_count(page) == 1)
                                goto free_it;
@@ -438,7 +503,7 @@ shrink_list(struct list_head *page_list, unsigned int gfp_mask,
 
 free_it:
                unlock_page(page);
-               ret++;
+               reclaimed++;
                if (!pagevec_add(&freed_pvec, page))
                        __pagevec_release_nonlru(&freed_pvec);
                continue;
@@ -456,7 +521,8 @@ keep:
        if (pagevec_count(&freed_pvec))
                __pagevec_release_nonlru(&freed_pvec);
        mod_page_state(pgactivate, pgactivate);
-       return ret;
+       sc->nr_reclaimed += reclaimed;
+       return reclaimed;
 }
 
 /*
@@ -464,19 +530,16 @@ keep:
  * a batch of pages and working on them outside the lock.  Any pages which were
  * not freed will be added back to the LRU.
  *
- * shrink_cache() is passed the number of pages to scan and returns the number
- * of pages which were reclaimed.
+ * shrink_cache() adds the number of pages reclaimed to sc->nr_reclaimed
  *
  * For pagecache intensive workloads, the first loop here is the hottest spot
  * in the kernel (apart from the copy_*_user functions).
  */
-static int
-shrink_cache(struct zone *zone, unsigned int gfp_mask,
-               int max_scan, int *total_scanned, int do_writepage)
+static void shrink_cache(struct zone *zone, struct scan_control *sc)
 {
        LIST_HEAD(page_list);
        struct pagevec pvec;
-       int ret = 0;
+       int max_scan = sc->nr_to_scan;
 
        pagevec_init(&pvec, 1);
 
@@ -498,14 +561,16 @@ shrink_cache(struct zone *zone, unsigned int gfp_mask,
                        if (!TestClearPageLRU(page))
                                BUG();
                        list_del(&page->lru);
-                       if (page_count(page) == 0) {
-                               /* It is currently in pagevec_release() */
+                       if (get_page_testone(page)) {
+                               /*
+                                * It is being freed elsewhere
+                                */
+                               __put_page(page);
                                SetPageLRU(page);
                                list_add(&page->lru, &zone->inactive_list);
                                continue;
                        }
                        list_add(&page->lru, &page_list);
-                       page_cache_get(page);
                        nr_taken++;
                }
                zone->nr_inactive -= nr_taken;
@@ -520,17 +585,11 @@ shrink_cache(struct zone *zone, unsigned int gfp_mask,
                        mod_page_state_zone(zone, pgscan_kswapd, nr_scan);
                else
                        mod_page_state_zone(zone, pgscan_direct, nr_scan);
-               nr_freed = shrink_list(&page_list, gfp_mask,
-                                       total_scanned, do_writepage);
-               *total_scanned += nr_taken;
+               nr_freed = shrink_list(&page_list, sc);
                if (current_is_kswapd())
                        mod_page_state(kswapd_steal, nr_freed);
                mod_page_state_zone(zone, pgsteal, nr_freed);
 
-               ret += nr_freed;
-               if (nr_freed <= 0 && list_empty(&page_list))
-                       goto done;
-
                spin_lock_irq(&zone->lru_lock);
                /*
                 * Put back any unfreeable pages.
@@ -554,7 +613,6 @@ shrink_cache(struct zone *zone, unsigned int gfp_mask,
        spin_unlock_irq(&zone->lru_lock);
 done:
        pagevec_release(&pvec);
-       return ret;
 }
 
 /*
@@ -571,16 +629,16 @@ done:
  * It is safe to rely on PG_active against the non-LRU pages in here because
  * nobody will play with that bit on a non-LRU page.
  *
- * The downside is that we have to touch page->count against each page.
+ * The downside is that we have to touch page->_count against each page.
  * But we had to alter page->flags anyway.
  */
 static void
-refill_inactive_zone(struct zone *zone, const int nr_pages_in,
-                       struct page_state *ps)
+refill_inactive_zone(struct zone *zone, struct scan_control *sc)
 {
        int pgmoved;
        int pgdeactivate = 0;
-       int nr_pages = nr_pages_in;
+       int pgscanned = 0;
+       int nr_pages = sc->nr_to_scan;
        LIST_HEAD(l_hold);      /* The pages which were snipped off */
        LIST_HEAD(l_inactive);  /* Pages to go onto the inactive_list */
        LIST_HEAD(l_active);    /* Pages to go onto the active_list */
@@ -594,22 +652,27 @@ refill_inactive_zone(struct zone *zone, const int nr_pages_in,
        lru_add_drain();
        pgmoved = 0;
        spin_lock_irq(&zone->lru_lock);
-       while (nr_pages && !list_empty(&zone->active_list)) {
+       while (pgscanned < nr_pages && !list_empty(&zone->active_list)) {
                page = lru_to_page(&zone->active_list);
                prefetchw_prev_lru_page(page, &zone->active_list, flags);
                if (!TestClearPageLRU(page))
                        BUG();
                list_del(&page->lru);
-               if (page_count(page) == 0) {
-                       /* It is currently in pagevec_release() */
+               if (get_page_testone(page)) {
+                       /*
+                        * It was already free!  release_pages() or put_page()
+                        * are about to remove it from the LRU and free it. So
+                        * put the refcount back and put the page back on the
+                        * LRU
+                        */
+                       __put_page(page);
                        SetPageLRU(page);
                        list_add(&page->lru, &zone->active_list);
                } else {
-                       page_cache_get(page);
                        list_add(&page->lru, &l_hold);
                        pgmoved++;
                }
-               nr_pages--;
+               pgscanned++;
        }
        zone->nr_active -= pgmoved;
        spin_unlock_irq(&zone->lru_lock);
@@ -625,7 +688,7 @@ refill_inactive_zone(struct zone *zone, const int nr_pages_in,
         * mapped memory instead of just pagecache.  Work out how much memory
         * is mapped.
         */
-       mapped_ratio = (ps->nr_mapped * 100) / total_memory;
+       mapped_ratio = (sc->nr_mapped * 100) / total_memory;
 
        /*
         * Now decide how much we really want to unmap some pages.  The mapped
@@ -653,13 +716,13 @@ refill_inactive_zone(struct zone *zone, const int nr_pages_in,
                                list_add(&page->lru, &l_active);
                                continue;
                        }
-                       rmap_lock(page);
+                       page_map_lock(page);
                        if (page_referenced(page)) {
-                               rmap_unlock(page);
+                               page_map_unlock(page);
                                list_add(&page->lru, &l_active);
                                continue;
                        }
-                       rmap_unlock(page);
+                       page_map_unlock(page);
                }
                /*
                 * FIXME: need to consider page_count(page) here if/when we
@@ -724,7 +787,7 @@ refill_inactive_zone(struct zone *zone, const int nr_pages_in,
        spin_unlock_irq(&zone->lru_lock);
        pagevec_release(&pvec);
 
-       mod_page_state_zone(zone, pgrefill, nr_pages_in - nr_pages);
+       mod_page_state_zone(zone, pgrefill, pgscanned);
        mod_page_state(pgdeactivate, pgdeactivate);
 }
 
@@ -732,41 +795,52 @@ refill_inactive_zone(struct zone *zone, const int nr_pages_in,
  * Scan `nr_pages' from this zone.  Returns the number of reclaimed pages.
  * This is a basic per-zone page freer.  Used by both kswapd and direct reclaim.
  */
-static int
-shrink_zone(struct zone *zone, int max_scan, unsigned int gfp_mask,
-               int *total_scanned, struct page_state *ps, int do_writepage)
+static void
+shrink_zone(struct zone *zone, struct scan_control *sc)
 {
-       unsigned long ratio;
+       unsigned long scan_active, scan_inactive;
        int count;
 
+       scan_inactive = (zone->nr_active + zone->nr_inactive) >> sc->priority;
+
        /*
         * Try to keep the active list 2/3 of the size of the cache.  And
         * make sure that refill_inactive is given a decent number of pages.
         *
-        * The "ratio+1" here is important.  With pagecache-intensive workloads
-        * the inactive list is huge, and `ratio' evaluates to zero all the
-        * time.  Which pins the active list memory.  So we add one to `ratio'
-        * just to make sure that the kernel will slowly sift through the
-        * active list.
+        * The "scan_active + 1" here is important.  With pagecache-intensive
+        * workloads the inactive list is huge, and `ratio' evaluates to zero
+        * all the time.  Which pins the active list memory.  So we add one to
+        * `scan_active' just to make sure that the kernel will slowly sift
+        * through the active list.
         */
-       ratio = (unsigned long)SWAP_CLUSTER_MAX * zone->nr_active /
-                               ((zone->nr_inactive | 1) * 2);
+       if (zone->nr_active >= 4*(zone->nr_inactive*2 + 1)) {
+               /* Don't scan more than 4 times the inactive list scan size */
+               scan_active = 4*scan_inactive;
+       } else {
+               unsigned long long tmp;
+
+               /* Cast to long long so the multiply doesn't overflow */
+
+               tmp = (unsigned long long)scan_inactive * zone->nr_active;
+               do_div(tmp, zone->nr_inactive*2 + 1);
+               scan_active = (unsigned long)tmp;
+       }
 
-       atomic_add(ratio+1, &zone->nr_scan_active);
+       atomic_add(scan_active + 1, &zone->nr_scan_active);
        count = atomic_read(&zone->nr_scan_active);
        if (count >= SWAP_CLUSTER_MAX) {
                atomic_set(&zone->nr_scan_active, 0);
-               refill_inactive_zone(zone, count, ps);
+               sc->nr_to_scan = count;
+               refill_inactive_zone(zone, sc);
        }
 
-       atomic_add(max_scan, &zone->nr_scan_inactive);
+       atomic_add(scan_inactive, &zone->nr_scan_inactive);
        count = atomic_read(&zone->nr_scan_inactive);
        if (count >= SWAP_CLUSTER_MAX) {
                atomic_set(&zone->nr_scan_inactive, 0);
-               return shrink_cache(zone, gfp_mask, count,
-                                       total_scanned, do_writepage);
+               sc->nr_to_scan = count;
+               shrink_cache(zone, sc);
        }
-       return 0;
 }
 
 /*
@@ -785,28 +859,23 @@ shrink_zone(struct zone *zone, int max_scan, unsigned int gfp_mask,
  * If a zone is deemed to be full of pinned pages then just give it a light
  * scan then give up on it.
  */
-static int
-shrink_caches(struct zone **zones, int priority, int *total_scanned,
-               int gfp_mask, struct page_state *ps, int do_writepage)
+static void
+shrink_caches(struct zone **zones, struct scan_control *sc)
 {
-       int ret = 0;
        int i;
 
        for (i = 0; zones[i] != NULL; i++) {
                struct zone *zone = zones[i];
-               int max_scan;
 
-               if (zone->free_pages < zone->pages_high)
-                       zone->temp_priority = priority;
+               zone->temp_priority = sc->priority;
+               if (zone->prev_priority > sc->priority)
+                       zone->prev_priority = sc->priority;
 
-               if (zone->all_unreclaimable && priority != DEF_PRIORITY)
+               if (zone->all_unreclaimable && sc->priority != DEF_PRIORITY)
                        continue;       /* Let kswapd poll it */
 
-               max_scan = zone->nr_inactive >> priority;
-               ret += shrink_zone(zone, max_scan, gfp_mask,
-                                       total_scanned, ps, do_writepage);
+               shrink_zone(zone, sc);
        }
-       return ret;
 }
  
 /*
@@ -817,25 +886,23 @@ shrink_caches(struct zone **zones, int priority, int *total_scanned,
  *
  * If the caller is !__GFP_FS then the probability of a failure is reasonably
  * high - the zone may be full of dirty or under-writeback pages, which this
- * caller can't do much about.  So for !__GFP_FS callers, we just perform a
- * small LRU walk and if that didn't work out, fail the allocation back to the
- * caller.  GFP_NOFS allocators need to know how to deal with it.  Kicking
- * bdflush, waiting and retrying will work.
- *
- * This is a fairly lame algorithm - it can result in excessive CPU burning and
- * excessive rotation of the inactive list, which is _supposed_ to be an LRU,
- * yes?
+ * caller can't do much about.  We kick pdflush and take explicit naps in the
+ * hope that some of these pages can be written.  But if the allocating task
+ * holds filesystem locks which prevent writeout this might not work, and the
+ * allocation attempt will fail.
  */
 int try_to_free_pages(struct zone **zones,
                unsigned int gfp_mask, unsigned int order)
 {
        int priority;
        int ret = 0;
-       int nr_reclaimed = 0;
+       int total_scanned = 0, total_reclaimed = 0;
        struct reclaim_state *reclaim_state = current->reclaim_state;
+       struct scan_control sc;
        int i;
-       unsigned long total_scanned = 0;
-       int do_writepage = 0;
+
+       sc.gfp_mask = gfp_mask;
+       sc.may_writepage = 0;
 
        inc_page_state(allocstall);
 
@@ -843,23 +910,23 @@ int try_to_free_pages(struct zone **zones,
                zones[i]->temp_priority = DEF_PRIORITY;
 
        for (priority = DEF_PRIORITY; priority >= 0; priority--) {
-               int scanned = 0;
-               struct page_state ps;
-
-               get_page_state(&ps);
-               nr_reclaimed += shrink_caches(zones, priority, &scanned,
-                                               gfp_mask, &ps, do_writepage);
-               shrink_slab(scanned, gfp_mask);
+               sc.nr_mapped = read_page_state(nr_mapped);
+               sc.nr_scanned = 0;
+               sc.nr_reclaimed = 0;
+               sc.priority = priority;
+               shrink_caches(zones, &sc);
+               shrink_slab(sc.nr_scanned, gfp_mask);
                if (reclaim_state) {
-                       nr_reclaimed += reclaim_state->reclaimed_slab;
+                       sc.nr_reclaimed += reclaim_state->reclaimed_slab;
                        reclaim_state->reclaimed_slab = 0;
                }
-               if (nr_reclaimed >= SWAP_CLUSTER_MAX) {
+               if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX) {
                        ret = 1;
                        goto out;
                }
-               if (!(gfp_mask & __GFP_FS))
-                       break;          /* Let the caller handle it */
+               total_scanned += sc.nr_scanned;
+               total_reclaimed += sc.nr_reclaimed;
+
                /*
                 * Try to write back as many pages as we just scanned.  This
                 * tends to cause slow streaming writers to write data to the
@@ -867,14 +934,13 @@ int try_to_free_pages(struct zone **zones,
                 * that's undesirable in laptop mode, where we *want* lumpy
                 * writeout.  So in laptop mode, write out the whole world.
                 */
-               total_scanned += scanned;
                if (total_scanned > SWAP_CLUSTER_MAX + SWAP_CLUSTER_MAX/2) {
                        wakeup_bdflush(laptop_mode ? 0 : total_scanned);
-                       do_writepage = 1;
+                       sc.may_writepage = 1;
                }
 
                /* Take a nap, wait for some writeback to complete */
-               if (scanned && priority < DEF_PRIORITY - 2)
+               if (sc.nr_scanned && priority < DEF_PRIORITY - 2)
                        blk_congestion_wait(WRITE, HZ/10);
        }
        if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY))
@@ -910,15 +976,18 @@ out:
  * the page allocator fallback scheme to ensure that aging of pages is balanced
  * across the zones.
  */
-static int balance_pgdat(pg_data_t *pgdat, int nr_pages, struct page_state *ps)
+static int balance_pgdat(pg_data_t *pgdat, int nr_pages)
 {
        int to_free = nr_pages;
        int priority;
        int i;
+       int total_scanned = 0, total_reclaimed = 0;
        struct reclaim_state *reclaim_state = current->reclaim_state;
-       unsigned long total_scanned = 0;
-       unsigned long total_reclaimed = 0;
-       int do_writepage = 0;
+       struct scan_control sc;
+
+       sc.gfp_mask = GFP_KERNEL;
+       sc.may_writepage = 0;
+       sc.nr_mapped = read_page_state(nr_mapped);
 
        inc_page_state(pageoutrun);
 
@@ -928,7 +997,7 @@ static int balance_pgdat(pg_data_t *pgdat, int nr_pages, struct page_state *ps)
                zone->temp_priority = DEF_PRIORITY;
        }
 
-       for (priority = DEF_PRIORITY; priority; priority--) {
+       for (priority = DEF_PRIORITY; priority >= 0; priority--) {
                int all_zones_ok = 1;
                int end_zone = 0;       /* Inclusive.  0 = ZONE_DMA */
 
@@ -966,9 +1035,6 @@ scan:
                 */
                for (i = 0; i <= end_zone; i++) {
                        struct zone *zone = pgdat->node_zones + i;
-                       int max_scan;
-                       int reclaimed;
-                       int scanned = 0;
 
                        if (zone->all_unreclaimable && priority != DEF_PRIORITY)
                                continue;
@@ -978,15 +1044,16 @@ scan:
                                        all_zones_ok = 0;
                        }
                        zone->temp_priority = priority;
-                       max_scan = zone->nr_inactive >> priority;
-                       reclaimed = shrink_zone(zone, max_scan, GFP_KERNEL,
-                                       &scanned, ps, do_writepage);
-                       total_scanned += scanned;
+                       if (zone->prev_priority > priority)
+                               zone->prev_priority = priority;
+                       sc.nr_scanned = 0;
+                       sc.nr_reclaimed = 0;
+                       sc.priority = priority;
+                       shrink_zone(zone, &sc);
                        reclaim_state->reclaimed_slab = 0;
-                       shrink_slab(scanned, GFP_KERNEL);
-                       reclaimed += reclaim_state->reclaimed_slab;
-                       total_reclaimed += reclaimed;
-                       to_free -= reclaimed;
+                       shrink_slab(sc.nr_scanned, GFP_KERNEL);
+                       sc.nr_reclaimed += reclaim_state->reclaimed_slab;
+                       total_reclaimed += sc.nr_reclaimed;
                        if (zone->all_unreclaimable)
                                continue;
                        if (zone->pages_scanned > zone->present_pages * 2)
@@ -998,9 +1065,9 @@ scan:
                         */
                        if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
                            total_scanned > total_reclaimed+total_reclaimed/2)
-                               do_writepage = 1;
+                               sc.may_writepage = 1;
                }
-               if (nr_pages && to_free > 0)
+               if (nr_pages && to_free > total_reclaimed)
                        continue;       /* swsusp: need to do more work */
                if (all_zones_ok)
                        break;          /* kswapd: all done */
@@ -1064,15 +1131,13 @@ int kswapd(void *p)
        tsk->flags |= PF_MEMALLOC|PF_KSWAPD;
 
        for ( ; ; ) {
-               struct page_state ps;
-
                if (current->flags & PF_FREEZE)
                        refrigerator(PF_FREEZE);
                prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
                schedule();
                finish_wait(&pgdat->kswapd_wait, &wait);
-               get_page_state(&ps);
-               balance_pgdat(pgdat, 0, &ps);
+
+               balance_pgdat(pgdat, 0);
        }
 }
 
@@ -1105,10 +1170,7 @@ int shrink_all_memory(int nr_pages)
        current->reclaim_state = &reclaim_state;
        for_each_pgdat(pgdat) {
                int freed;
-               struct page_state ps;
-
-               get_page_state(&ps);
-               freed = balance_pgdat(pgdat, nr_to_free, &ps);
+               freed = balance_pgdat(pgdat, nr_to_free);
                ret += freed;
                nr_to_free -= freed;
                if (nr_to_free <= 0)