Revert "xen/pv-on-hvm kexec: add xs_reset_watches to shutdown watches from old kernel"
[pandora-kernel.git] / mm / vmscan.c
index dd5fc86..a1893c0 100644 (file)
@@ -633,13 +633,14 @@ redo:
                lru = LRU_UNEVICTABLE;
                add_page_to_unevictable_list(page);
                /*
-                * When racing with an mlock clearing (page is
-                * unlocked), make sure that if the other thread does
-                * not observe our setting of PG_lru and fails
-                * isolation, we see PG_mlocked cleared below and move
+                * When racing with an mlock or AS_UNEVICTABLE clearing
+                * (page is unlocked) make sure that if the other thread
+                * does not observe our setting of PG_lru and fails
+                * isolation/check_move_unevictable_page,
+                * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
                 * the page back to the evictable list.
                 *
-                * The other side is TestClearPageMlocked().
+                * The other side is TestClearPageMlocked() or shmem_lock().
                 */
                smp_mb();
        }
@@ -1766,7 +1767,7 @@ static int inactive_anon_is_low(struct zone *zone, struct scan_control *sc)
        if (scanning_global_lru(sc))
                low = inactive_anon_is_low_global(zone);
        else
-               low = mem_cgroup_inactive_anon_is_low(sc->mem_cgroup);
+               low = mem_cgroup_inactive_anon_is_low(sc->mem_cgroup, zone);
        return low;
 }
 #else
@@ -1809,7 +1810,7 @@ static int inactive_file_is_low(struct zone *zone, struct scan_control *sc)
        if (scanning_global_lru(sc))
                low = inactive_file_is_low_global(zone);
        else
-               low = mem_cgroup_inactive_file_is_low(sc->mem_cgroup);
+               low = mem_cgroup_inactive_file_is_low(sc->mem_cgroup, zone);
        return low;
 }
 
@@ -2102,14 +2103,19 @@ restart:
  *
  * If a zone is deemed to be full of pinned pages then just give it a light
  * scan then give up on it.
+ *
+ * This function returns true if a zone is being reclaimed for a costly
+ * high-order allocation and compaction is either ready to begin or deferred.
+ * This indicates to the caller that it should retry the allocation or fail.
  */
-static void shrink_zones(int priority, struct zonelist *zonelist,
+static bool shrink_zones(int priority, struct zonelist *zonelist,
                                        struct scan_control *sc)
 {
        struct zoneref *z;
        struct zone *zone;
        unsigned long nr_soft_reclaimed;
        unsigned long nr_soft_scanned;
+       bool should_abort_reclaim = false;
 
        for_each_zone_zonelist_nodemask(zone, z, zonelist,
                                        gfp_zone(sc->gfp_mask), sc->nodemask) {
@@ -2124,6 +2130,23 @@ static void shrink_zones(int priority, struct zonelist *zonelist,
                                continue;
                        if (zone->all_unreclaimable && priority != DEF_PRIORITY)
                                continue;       /* Let kswapd poll it */
+                       if (COMPACTION_BUILD) {
+                               /*
+                                * If we already have plenty of memory free for
+                                * compaction in this zone, don't free any more.
+                                * Even though compaction is invoked for any
+                                * non-zero order, only frequent costly order
+                                * reclamation is disruptive enough to become a
+                                * noticable problem, like transparent huge page
+                                * allocations.
+                                */
+                               if (sc->order > PAGE_ALLOC_COSTLY_ORDER &&
+                                       (compaction_suitable(zone, sc->order) ||
+                                        compaction_deferred(zone))) {
+                                       should_abort_reclaim = true;
+                                       continue;
+                               }
+                       }
                        /*
                         * This steals pages from memory cgroups over softlimit
                         * and returns the number of reclaimed pages and
@@ -2141,6 +2164,8 @@ static void shrink_zones(int priority, struct zonelist *zonelist,
 
                shrink_zone(priority, zone, sc);
        }
+
+       return should_abort_reclaim;
 }
 
 static bool zone_reclaimable(struct zone *zone)
@@ -2205,7 +2230,9 @@ static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
                sc->nr_scanned = 0;
                if (!priority)
                        disable_swap_token(sc->mem_cgroup);
-               shrink_zones(priority, zonelist, sc);
+               if (shrink_zones(priority, zonelist, sc))
+                       break;
+
                /*
                 * Don't shrink slabs when reclaiming memory from
                 * over limit cgroups
@@ -2239,7 +2266,8 @@ static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
                 */
                writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
                if (total_scanned > writeback_threshold) {
-                       wakeup_flusher_threads(laptop_mode ? 0 : total_scanned);
+                       wakeup_flusher_threads(laptop_mode ? 0 : total_scanned,
+                                               WB_REASON_TRY_TO_FREE_PAGES);
                        sc->may_writepage = 1;
                }
 
@@ -2887,6 +2915,8 @@ static int kswapd(void *p)
                                                balanced_classzone_idx);
                        order = pgdat->kswapd_max_order;
                        classzone_idx = pgdat->classzone_idx;
+                       new_order = order;
+                       new_classzone_idx = classzone_idx;
                        pgdat->kswapd_max_order = 0;
                        pgdat->classzone_idx = pgdat->nr_zones - 1;
                }
@@ -3415,66 +3445,12 @@ void scan_mapping_unevictable_pages(struct address_space *mapping)
 
 }
 
-/**
- * scan_zone_unevictable_pages - check unevictable list for evictable pages
- * @zone - zone of which to scan the unevictable list
- *
- * Scan @zone's unevictable LRU lists to check for pages that have become
- * evictable.  Move those that have to @zone's inactive list where they
- * become candidates for reclaim, unless shrink_inactive_zone() decides
- * to reactivate them.  Pages that are still unevictable are rotated
- * back onto @zone's unevictable list.
- */
-#define SCAN_UNEVICTABLE_BATCH_SIZE 16UL /* arbitrary lock hold batch size */
-static void scan_zone_unevictable_pages(struct zone *zone)
-{
-       struct list_head *l_unevictable = &zone->lru[LRU_UNEVICTABLE].list;
-       unsigned long scan;
-       unsigned long nr_to_scan = zone_page_state(zone, NR_UNEVICTABLE);
-
-       while (nr_to_scan > 0) {
-               unsigned long batch_size = min(nr_to_scan,
-                                               SCAN_UNEVICTABLE_BATCH_SIZE);
-
-               spin_lock_irq(&zone->lru_lock);
-               for (scan = 0;  scan < batch_size; scan++) {
-                       struct page *page = lru_to_page(l_unevictable);
-
-                       if (!trylock_page(page))
-                               continue;
-
-                       prefetchw_prev_lru_page(page, l_unevictable, flags);
-
-                       if (likely(PageLRU(page) && PageUnevictable(page)))
-                               check_move_unevictable_page(page, zone);
-
-                       unlock_page(page);
-               }
-               spin_unlock_irq(&zone->lru_lock);
-
-               nr_to_scan -= batch_size;
-       }
-}
-
-
-/**
- * scan_all_zones_unevictable_pages - scan all unevictable lists for evictable pages
- *
- * A really big hammer:  scan all zones' unevictable LRU lists to check for
- * pages that have become evictable.  Move those back to the zones'
- * inactive list where they become candidates for reclaim.
- * This occurs when, e.g., we have unswappable pages on the unevictable lists,
- * and we add swap to the system.  As such, it runs in the context of a task
- * that has possibly/probably made some previously unevictable pages
- * evictable.
- */
-static void scan_all_zones_unevictable_pages(void)
+static void warn_scan_unevictable_pages(void)
 {
-       struct zone *zone;
-
-       for_each_zone(zone) {
-               scan_zone_unevictable_pages(zone);
-       }
+       printk_once(KERN_WARNING
+                   "The scan_unevictable_pages sysctl/node-interface has been "
+                   "disabled for lack of a legitimate use case.  If you have "
+                   "one, please send an email to linux-mm@kvack.org.\n");
 }
 
 /*
@@ -3487,11 +3463,8 @@ int scan_unevictable_handler(struct ctl_table *table, int write,
                           void __user *buffer,
                           size_t *length, loff_t *ppos)
 {
+       warn_scan_unevictable_pages();
        proc_doulongvec_minmax(table, write, buffer, length, ppos);
-
-       if (write && *(unsigned long *)table->data)
-               scan_all_zones_unevictable_pages();
-
        scan_unevictable_pages = 0;
        return 0;
 }
@@ -3506,6 +3479,7 @@ static ssize_t read_scan_unevictable_node(struct sys_device *dev,
                                          struct sysdev_attribute *attr,
                                          char *buf)
 {
+       warn_scan_unevictable_pages();
        return sprintf(buf, "0\n");     /* always zero; should fit... */
 }
 
@@ -3513,19 +3487,7 @@ static ssize_t write_scan_unevictable_node(struct sys_device *dev,
                                           struct sysdev_attribute *attr,
                                        const char *buf, size_t count)
 {
-       struct zone *node_zones = NODE_DATA(dev->id)->node_zones;
-       struct zone *zone;
-       unsigned long res;
-       unsigned long req = strict_strtoul(buf, 10, &res);
-
-       if (!req)
-               return 1;       /* zero is no-op */
-
-       for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
-               if (!populated_zone(zone))
-                       continue;
-               scan_zone_unevictable_pages(zone);
-       }
+       warn_scan_unevictable_pages();
        return 1;
 }