X25: Move interrupt flag to bitfield
[linux-2.6.git] / mm / swapfile.c
index 2a97faf..6cd0a8f 100644 (file)
 #include <linux/namei.h>
 #include <linux/shm.h>
 #include <linux/blkdev.h>
+#include <linux/random.h>
 #include <linux/writeback.h>
 #include <linux/proc_fs.h>
 #include <linux/seq_file.h>
 #include <linux/init.h>
 #include <linux/module.h>
+#include <linux/ksm.h>
 #include <linux/rmap.h>
 #include <linux/security.h>
 #include <linux/backing-dev.h>
 #include <asm/pgtable.h>
 #include <asm/tlbflush.h>
 #include <linux/swapops.h>
+#include <linux/page_cgroup.h>
+
+static bool swap_count_continued(struct swap_info_struct *, pgoff_t,
+                                unsigned char);
+static void free_swap_count_continuations(struct swap_info_struct *);
+static sector_t map_swap_entry(swp_entry_t, struct block_device**);
 
 static DEFINE_SPINLOCK(swap_lock);
 static unsigned int nr_swapfiles;
+long nr_swap_pages;
 long total_swap_pages;
-static int swap_overflow;
 static int least_priority;
 
 static const char Bad_file[] = "Bad swap file entry ";
@@ -46,10 +54,41 @@ static const char Unused_offset[] = "Unused swap offset entry ";
 
 static struct swap_list_t swap_list = {-1, -1};
 
-static struct swap_info_struct swap_info[MAX_SWAPFILES];
+static struct swap_info_struct *swap_info[MAX_SWAPFILES];
 
 static DEFINE_MUTEX(swapon_mutex);
 
+static inline unsigned char swap_count(unsigned char ent)
+{
+       return ent & ~SWAP_HAS_CACHE;   /* may include SWAP_HAS_CONT flag */
+}
+
+/* returns 1 if swap entry is freed */
+static int
+__try_to_reclaim_swap(struct swap_info_struct *si, unsigned long offset)
+{
+       swp_entry_t entry = swp_entry(si->type, offset);
+       struct page *page;
+       int ret = 0;
+
+       page = find_get_page(&swapper_space, entry.val);
+       if (!page)
+               return 0;
+       /*
+        * This function is called from scan_swap_map() and it's called
+        * by vmscan.c at reclaiming pages. So, we hold a lock on a page, here.
+        * We have to use trylock for avoiding deadlock. This is a special
+        * case and you should use try_to_free_swap() with explicit lock_page()
+        * in usual operations.
+        */
+       if (trylock_page(page)) {
+               ret = try_to_free_swap(page);
+               unlock_page(page);
+       }
+       page_cache_release(page);
+       return ret;
+}
+
 /*
  * We need this because the bdev->unplug_fn can sleep and we cannot
  * hold swap_lock while calling the unplug_fn. And swap_lock
@@ -64,7 +103,7 @@ void swap_unplug_io_fn(struct backing_dev_info *unused_bdi, struct page *page)
        down_read(&swap_unplug_sem);
        entry.val = page_private(page);
        if (PageSwapCache(page)) {
-               struct block_device *bdev = swap_info[swp_type(entry)].bdev;
+               struct block_device *bdev = swap_info[swp_type(entry)]->bdev;
                struct backing_dev_info *bdi;
 
                /*
@@ -83,15 +122,101 @@ void swap_unplug_io_fn(struct backing_dev_info *unused_bdi, struct page *page)
        up_read(&swap_unplug_sem);
 }
 
+/*
+ * swapon tell device that all the old swap contents can be discarded,
+ * to allow the swap device to optimize its wear-levelling.
+ */
+static int discard_swap(struct swap_info_struct *si)
+{
+       struct swap_extent *se;
+       sector_t start_block;
+       sector_t nr_blocks;
+       int err = 0;
+
+       /* Do not discard the swap header page! */
+       se = &si->first_swap_extent;
+       start_block = (se->start_block + 1) << (PAGE_SHIFT - 9);
+       nr_blocks = ((sector_t)se->nr_pages - 1) << (PAGE_SHIFT - 9);
+       if (nr_blocks) {
+               err = blkdev_issue_discard(si->bdev, start_block,
+                               nr_blocks, GFP_KERNEL, DISCARD_FL_BARRIER);
+               if (err)
+                       return err;
+               cond_resched();
+       }
+
+       list_for_each_entry(se, &si->first_swap_extent.list, list) {
+               start_block = se->start_block << (PAGE_SHIFT - 9);
+               nr_blocks = (sector_t)se->nr_pages << (PAGE_SHIFT - 9);
+
+               err = blkdev_issue_discard(si->bdev, start_block,
+                               nr_blocks, GFP_KERNEL, DISCARD_FL_BARRIER);
+               if (err)
+                       break;
+
+               cond_resched();
+       }
+       return err;             /* That will often be -EOPNOTSUPP */
+}
+
+/*
+ * swap allocation tell device that a cluster of swap can now be discarded,
+ * to allow the swap device to optimize its wear-levelling.
+ */
+static void discard_swap_cluster(struct swap_info_struct *si,
+                                pgoff_t start_page, pgoff_t nr_pages)
+{
+       struct swap_extent *se = si->curr_swap_extent;
+       int found_extent = 0;
+
+       while (nr_pages) {
+               struct list_head *lh;
+
+               if (se->start_page <= start_page &&
+                   start_page < se->start_page + se->nr_pages) {
+                       pgoff_t offset = start_page - se->start_page;
+                       sector_t start_block = se->start_block + offset;
+                       sector_t nr_blocks = se->nr_pages - offset;
+
+                       if (nr_blocks > nr_pages)
+                               nr_blocks = nr_pages;
+                       start_page += nr_blocks;
+                       nr_pages -= nr_blocks;
+
+                       if (!found_extent++)
+                               si->curr_swap_extent = se;
+
+                       start_block <<= PAGE_SHIFT - 9;
+                       nr_blocks <<= PAGE_SHIFT - 9;
+                       if (blkdev_issue_discard(si->bdev, start_block,
+                                   nr_blocks, GFP_NOIO, DISCARD_FL_BARRIER))
+                               break;
+               }
+
+               lh = se->list.next;
+               se = list_entry(lh, struct swap_extent, list);
+       }
+}
+
+static int wait_for_discard(void *word)
+{
+       schedule();
+       return 0;
+}
+
 #define SWAPFILE_CLUSTER       256
 #define LATENCY_LIMIT          256
 
-static inline unsigned long scan_swap_map(struct swap_info_struct *si)
+static inline unsigned long scan_swap_map(struct swap_info_struct *si,
+                                         unsigned char usage)
 {
-       unsigned long offset, last_in_cluster;
+       unsigned long offset;
+       unsigned long scan_base;
+       unsigned long last_in_cluster = 0;
        int latency_ration = LATENCY_LIMIT;
+       int found_free_cluster = 0;
 
-       /* 
+       /*
         * We try to cluster swap pages by allocating them sequentially
         * in swap.  Once we've allocated SWAPFILE_CLUSTER pages this
         * way, however, we resort to first-free allocation, starting
@@ -99,16 +224,42 @@ static inline unsigned long scan_swap_map(struct swap_info_struct *si)
         * all over the entire swap partition, so that we reduce
         * overall disk seek times between swap pages.  -- sct
         * But we do now try to find an empty cluster.  -Andrea
+        * And we let swap pages go all over an SSD partition.  Hugh
         */
 
        si->flags += SWP_SCANNING;
-       if (unlikely(!si->cluster_nr)) {
-               si->cluster_nr = SWAPFILE_CLUSTER - 1;
-               if (si->pages - si->inuse_pages < SWAPFILE_CLUSTER)
-                       goto lowest;
+       scan_base = offset = si->cluster_next;
+
+       if (unlikely(!si->cluster_nr--)) {
+               if (si->pages - si->inuse_pages < SWAPFILE_CLUSTER) {
+                       si->cluster_nr = SWAPFILE_CLUSTER - 1;
+                       goto checks;
+               }
+               if (si->flags & SWP_DISCARDABLE) {
+                       /*
+                        * Start range check on racing allocations, in case
+                        * they overlap the cluster we eventually decide on
+                        * (we scan without swap_lock to allow preemption).
+                        * It's hardly conceivable that cluster_nr could be
+                        * wrapped during our scan, but don't depend on it.
+                        */
+                       if (si->lowest_alloc)
+                               goto checks;
+                       si->lowest_alloc = si->max;
+                       si->highest_alloc = 0;
+               }
                spin_unlock(&swap_lock);
 
-               offset = si->lowest_bit;
+               /*
+                * If seek is expensive, start searching for new cluster from
+                * start of partition, to minimize the span of allocated swap.
+                * But if seek is cheap, search from our current position, so
+                * that swap is allocated from all over the partition: if the
+                * Flash Translation Layer only remaps within limited zones,
+                * we don't want to wear out the first zone too quickly.
+                */
+               if (!(si->flags & SWP_SOLIDSTATE))
+                       scan_base = offset = si->lowest_bit;
                last_in_cluster = offset + SWAPFILE_CLUSTER - 1;
 
                /* Locate the first empty (unaligned) cluster */
@@ -117,56 +268,168 @@ static inline unsigned long scan_swap_map(struct swap_info_struct *si)
                                last_in_cluster = offset + SWAPFILE_CLUSTER;
                        else if (offset == last_in_cluster) {
                                spin_lock(&swap_lock);
-                               si->cluster_next = offset-SWAPFILE_CLUSTER+1;
-                               goto cluster;
+                               offset -= SWAPFILE_CLUSTER - 1;
+                               si->cluster_next = offset;
+                               si->cluster_nr = SWAPFILE_CLUSTER - 1;
+                               found_free_cluster = 1;
+                               goto checks;
                        }
                        if (unlikely(--latency_ration < 0)) {
                                cond_resched();
                                latency_ration = LATENCY_LIMIT;
                        }
                }
+
+               offset = si->lowest_bit;
+               last_in_cluster = offset + SWAPFILE_CLUSTER - 1;
+
+               /* Locate the first empty (unaligned) cluster */
+               for (; last_in_cluster < scan_base; offset++) {
+                       if (si->swap_map[offset])
+                               last_in_cluster = offset + SWAPFILE_CLUSTER;
+                       else if (offset == last_in_cluster) {
+                               spin_lock(&swap_lock);
+                               offset -= SWAPFILE_CLUSTER - 1;
+                               si->cluster_next = offset;
+                               si->cluster_nr = SWAPFILE_CLUSTER - 1;
+                               found_free_cluster = 1;
+                               goto checks;
+                       }
+                       if (unlikely(--latency_ration < 0)) {
+                               cond_resched();
+                               latency_ration = LATENCY_LIMIT;
+                       }
+               }
+
+               offset = scan_base;
                spin_lock(&swap_lock);
-               goto lowest;
+               si->cluster_nr = SWAPFILE_CLUSTER - 1;
+               si->lowest_alloc = 0;
        }
 
-       si->cluster_nr--;
-cluster:
-       offset = si->cluster_next;
-       if (offset > si->highest_bit)
-lowest:                offset = si->lowest_bit;
-checks:        if (!(si->flags & SWP_WRITEOK))
+checks:
+       if (!(si->flags & SWP_WRITEOK))
                goto no_page;
        if (!si->highest_bit)
                goto no_page;
-       if (!si->swap_map[offset]) {
-               if (offset == si->lowest_bit)
-                       si->lowest_bit++;
-               if (offset == si->highest_bit)
-                       si->highest_bit--;
-               si->inuse_pages++;
-               if (si->inuse_pages == si->pages) {
-                       si->lowest_bit = si->max;
-                       si->highest_bit = 0;
+       if (offset > si->highest_bit)
+               scan_base = offset = si->lowest_bit;
+
+       /* reuse swap entry of cache-only swap if not busy. */
+       if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
+               int swap_was_freed;
+               spin_unlock(&swap_lock);
+               swap_was_freed = __try_to_reclaim_swap(si, offset);
+               spin_lock(&swap_lock);
+               /* entry was freed successfully, try to use this again */
+               if (swap_was_freed)
+                       goto checks;
+               goto scan; /* check next one */
+       }
+
+       if (si->swap_map[offset])
+               goto scan;
+
+       if (offset == si->lowest_bit)
+               si->lowest_bit++;
+       if (offset == si->highest_bit)
+               si->highest_bit--;
+       si->inuse_pages++;
+       if (si->inuse_pages == si->pages) {
+               si->lowest_bit = si->max;
+               si->highest_bit = 0;
+       }
+       si->swap_map[offset] = usage;
+       si->cluster_next = offset + 1;
+       si->flags -= SWP_SCANNING;
+
+       if (si->lowest_alloc) {
+               /*
+                * Only set when SWP_DISCARDABLE, and there's a scan
+                * for a free cluster in progress or just completed.
+                */
+               if (found_free_cluster) {
+                       /*
+                        * To optimize wear-levelling, discard the
+                        * old data of the cluster, taking care not to
+                        * discard any of its pages that have already
+                        * been allocated by racing tasks (offset has
+                        * already stepped over any at the beginning).
+                        */
+                       if (offset < si->highest_alloc &&
+                           si->lowest_alloc <= last_in_cluster)
+                               last_in_cluster = si->lowest_alloc - 1;
+                       si->flags |= SWP_DISCARDING;
+                       spin_unlock(&swap_lock);
+
+                       if (offset < last_in_cluster)
+                               discard_swap_cluster(si, offset,
+                                       last_in_cluster - offset + 1);
+
+                       spin_lock(&swap_lock);
+                       si->lowest_alloc = 0;
+                       si->flags &= ~SWP_DISCARDING;
+
+                       smp_mb();       /* wake_up_bit advises this */
+                       wake_up_bit(&si->flags, ilog2(SWP_DISCARDING));
+
+               } else if (si->flags & SWP_DISCARDING) {
+                       /*
+                        * Delay using pages allocated by racing tasks
+                        * until the whole discard has been issued. We
+                        * could defer that delay until swap_writepage,
+                        * but it's easier to keep this self-contained.
+                        */
+                       spin_unlock(&swap_lock);
+                       wait_on_bit(&si->flags, ilog2(SWP_DISCARDING),
+                               wait_for_discard, TASK_UNINTERRUPTIBLE);
+                       spin_lock(&swap_lock);
+               } else {
+                       /*
+                        * Note pages allocated by racing tasks while
+                        * scan for a free cluster is in progress, so
+                        * that its final discard can exclude them.
+                        */
+                       if (offset < si->lowest_alloc)
+                               si->lowest_alloc = offset;
+                       if (offset > si->highest_alloc)
+                               si->highest_alloc = offset;
                }
-               si->swap_map[offset] = 1;
-               si->cluster_next = offset + 1;
-               si->flags -= SWP_SCANNING;
-               return offset;
        }
+       return offset;
 
+scan:
        spin_unlock(&swap_lock);
        while (++offset <= si->highest_bit) {
                if (!si->swap_map[offset]) {
                        spin_lock(&swap_lock);
                        goto checks;
                }
+               if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
+                       spin_lock(&swap_lock);
+                       goto checks;
+               }
+               if (unlikely(--latency_ration < 0)) {
+                       cond_resched();
+                       latency_ration = LATENCY_LIMIT;
+               }
+       }
+       offset = si->lowest_bit;
+       while (++offset < scan_base) {
+               if (!si->swap_map[offset]) {
+                       spin_lock(&swap_lock);
+                       goto checks;
+               }
+               if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
+                       spin_lock(&swap_lock);
+                       goto checks;
+               }
                if (unlikely(--latency_ration < 0)) {
                        cond_resched();
                        latency_ration = LATENCY_LIMIT;
                }
        }
        spin_lock(&swap_lock);
-       goto lowest;
 
 no_page:
        si->flags -= SWP_SCANNING;
@@ -186,10 +449,10 @@ swp_entry_t get_swap_page(void)
        nr_swap_pages--;
 
        for (type = swap_list.next; type >= 0 && wrapped < 2; type = next) {
-               si = swap_info + type;
+               si = swap_info[type];
                next = si->next;
                if (next < 0 ||
-                   (!wrapped && si->prio != swap_info[next].prio)) {
+                   (!wrapped && si->prio != swap_info[next]->prio)) {
                        next = swap_list.head;
                        wrapped++;
                }
@@ -200,7 +463,8 @@ swp_entry_t get_swap_page(void)
                        continue;
 
                swap_list.next = next;
-               offset = scan_swap_map(si);
+               /* This is called for allocating swap entry for cache */
+               offset = scan_swap_map(si, SWAP_HAS_CACHE);
                if (offset) {
                        spin_unlock(&swap_lock);
                        return swp_entry(type, offset);
@@ -214,16 +478,18 @@ noswap:
        return (swp_entry_t) {0};
 }
 
+/* The only caller of this function is now susupend routine */
 swp_entry_t get_swap_page_of_type(int type)
 {
        struct swap_info_struct *si;
        pgoff_t offset;
 
        spin_lock(&swap_lock);
-       si = swap_info + type;
-       if (si->flags & SWP_WRITEOK) {
+       si = swap_info[type];
+       if (si && (si->flags & SWP_WRITEOK)) {
                nr_swap_pages--;
-               offset = scan_swap_map(si);
+               /* This is called for allocating swap entry, not cache */
+               offset = scan_swap_map(si, 1);
                if (offset) {
                        spin_unlock(&swap_lock);
                        return swp_entry(type, offset);
@@ -234,9 +500,9 @@ swp_entry_t get_swap_page_of_type(int type)
        return (swp_entry_t) {0};
 }
 
-static struct swap_info_struct * swap_info_get(swp_entry_t entry)
+static struct swap_info_struct *swap_info_get(swp_entry_t entry)
 {
-       struct swap_info_struct * p;
+       struct swap_info_struct *p;
        unsigned long offset, type;
 
        if (!entry.val)
@@ -244,7 +510,7 @@ static struct swap_info_struct * swap_info_get(swp_entry_t entry)
        type = swp_type(entry);
        if (type >= nr_swapfiles)
                goto bad_nofile;
-       p = & swap_info[type];
+       p = swap_info[type];
        if (!(p->flags & SWP_USED))
                goto bad_device;
        offset = swp_offset(entry);
@@ -268,27 +534,58 @@ bad_nofile:
        printk(KERN_ERR "swap_free: %s%08lx\n", Bad_file, entry.val);
 out:
        return NULL;
-}      
+}
 
-static int swap_entry_free(struct swap_info_struct *p, unsigned long offset)
+static unsigned char swap_entry_free(struct swap_info_struct *p,
+                                    swp_entry_t entry, unsigned char usage)
 {
-       int count = p->swap_map[offset];
-
-       if (count < SWAP_MAP_MAX) {
-               count--;
-               p->swap_map[offset] = count;
-               if (!count) {
-                       if (offset < p->lowest_bit)
-                               p->lowest_bit = offset;
-                       if (offset > p->highest_bit)
-                               p->highest_bit = offset;
-                       if (p->prio > swap_info[swap_list.next].prio)
-                               swap_list.next = p - swap_info;
-                       nr_swap_pages++;
-                       p->inuse_pages--;
-               }
+       unsigned long offset = swp_offset(entry);
+       unsigned char count;
+       unsigned char has_cache;
+
+       count = p->swap_map[offset];
+       has_cache = count & SWAP_HAS_CACHE;
+       count &= ~SWAP_HAS_CACHE;
+
+       if (usage == SWAP_HAS_CACHE) {
+               VM_BUG_ON(!has_cache);
+               has_cache = 0;
+       } else if (count == SWAP_MAP_SHMEM) {
+               /*
+                * Or we could insist on shmem.c using a special
+                * swap_shmem_free() and free_shmem_swap_and_cache()...
+                */
+               count = 0;
+       } else if ((count & ~COUNT_CONTINUED) <= SWAP_MAP_MAX) {
+               if (count == COUNT_CONTINUED) {
+                       if (swap_count_continued(p, offset, count))
+                               count = SWAP_MAP_MAX | COUNT_CONTINUED;
+                       else
+                               count = SWAP_MAP_MAX;
+               } else
+                       count--;
        }
-       return count;
+
+       if (!count)
+               mem_cgroup_uncharge_swap(entry);
+
+       usage = count | has_cache;
+       p->swap_map[offset] = usage;
+
+       /* free if no reference */
+       if (!usage) {
+               if (offset < p->lowest_bit)
+                       p->lowest_bit = offset;
+               if (offset > p->highest_bit)
+                       p->highest_bit = offset;
+               if (swap_list.next >= 0 &&
+                   p->prio > swap_info[swap_list.next]->prio)
+                       swap_list.next = p->type;
+               nr_swap_pages++;
+               p->inuse_pages--;
+       }
+
+       return usage;
 }
 
 /*
@@ -297,17 +594,36 @@ static int swap_entry_free(struct swap_info_struct *p, unsigned long offset)
  */
 void swap_free(swp_entry_t entry)
 {
-       struct swap_info_struct * p;
+       struct swap_info_struct *p;
 
        p = swap_info_get(entry);
        if (p) {
-               swap_entry_free(p, swp_offset(entry));
+               swap_entry_free(p, entry, 1);
+               spin_unlock(&swap_lock);
+       }
+}
+
+/*
+ * Called after dropping swapcache to decrease refcnt to swap entries.
+ */
+void swapcache_free(swp_entry_t entry, struct page *page)
+{
+       struct swap_info_struct *p;
+       unsigned char count;
+
+       p = swap_info_get(entry);
+       if (p) {
+               count = swap_entry_free(p, entry, SWAP_HAS_CACHE);
+               if (page)
+                       mem_cgroup_uncharge_swapcache(page, entry, count != 0);
                spin_unlock(&swap_lock);
        }
 }
 
 /*
  * How many references to page are currently swapped out?
+ * This does not give an exact answer when swap count is continued,
+ * but does include the high COUNT_CONTINUED flag to allow for that.
  */
 static inline int page_swapcount(struct page *page)
 {
@@ -318,111 +634,73 @@ static inline int page_swapcount(struct page *page)
        entry.val = page_private(page);
        p = swap_info_get(entry);
        if (p) {
-               /* Subtract the 1 for the swap cache itself */
-               count = p->swap_map[swp_offset(entry)] - 1;
+               count = swap_count(p->swap_map[swp_offset(entry)]);
                spin_unlock(&swap_lock);
        }
        return count;
 }
 
 /*
- * We can use this swap cache entry directly
- * if there are no other references to it.
+ * We can write to an anon page without COW if there are no other references
+ * to it.  And as a side-effect, free up its swap: because the old content
+ * on disk will never be read, and seeking back there to write new content
+ * later would only waste time away from clustering.
  */
-int can_share_swap_page(struct page *page)
+int reuse_swap_page(struct page *page)
 {
        int count;
 
-       BUG_ON(!PageLocked(page));
+       VM_BUG_ON(!PageLocked(page));
+       if (unlikely(PageKsm(page)))
+               return 0;
        count = page_mapcount(page);
-       if (count <= 1 && PageSwapCache(page))
+       if (count <= 1 && PageSwapCache(page)) {
                count += page_swapcount(page);
-       return count == 1;
+               if (count == 1 && !PageWriteback(page)) {
+                       delete_from_swap_cache(page);
+                       SetPageDirty(page);
+               }
+       }
+       return count <= 1;
 }
 
 /*
- * Work out if there are any other processes sharing this
- * swap cache page. Free it if you can. Return success.
+ * If swap is getting full, or if there are no more mappings of this page,
+ * then try_to_free_swap is called to free its swap space.
  */
-static int remove_exclusive_swap_page_count(struct page *page, int count)
+int try_to_free_swap(struct page *page)
 {
-       int retval;
-       struct swap_info_struct * p;
-       swp_entry_t entry;
-
-       BUG_ON(PagePrivate(page));
-       BUG_ON(!PageLocked(page));
+       VM_BUG_ON(!PageLocked(page));
 
        if (!PageSwapCache(page))
                return 0;
        if (PageWriteback(page))
                return 0;
-       if (page_count(page) != count) /* us + cache + ptes */
+       if (page_swapcount(page))
                return 0;
 
-       entry.val = page_private(page);
-       p = swap_info_get(entry);
-       if (!p)
-               return 0;
-
-       /* Is the only swap cache user the cache itself? */
-       retval = 0;
-       if (p->swap_map[swp_offset(entry)] == 1) {
-               /* Recheck the page count with the swapcache lock held.. */
-               spin_lock_irq(&swapper_space.tree_lock);
-               if ((page_count(page) == count) && !PageWriteback(page)) {
-                       __delete_from_swap_cache(page);
-                       SetPageDirty(page);
-                       retval = 1;
-               }
-               spin_unlock_irq(&swapper_space.tree_lock);
-       }
-       spin_unlock(&swap_lock);
-
-       if (retval) {
-               swap_free(entry);
-               page_cache_release(page);
-       }
-
-       return retval;
-}
-
-/*
- * Most of the time the page should have two references: one for the
- * process and one for the swap cache.
- */
-int remove_exclusive_swap_page(struct page *page)
-{
-       return remove_exclusive_swap_page_count(page, 2);
-}
-
-/*
- * The pageout code holds an extra reference to the page.  That raises
- * the reference count to test for to 2 for a page that is only in the
- * swap cache plus 1 for each process that maps the page.
- */
-int remove_exclusive_swap_page_ref(struct page *page)
-{
-       return remove_exclusive_swap_page_count(page, 2 + page_mapcount(page));
+       delete_from_swap_cache(page);
+       SetPageDirty(page);
+       return 1;
 }
 
 /*
  * Free the swap entry like above, but also try to
  * free the page cache entry if it is the last user.
  */
-void free_swap_and_cache(swp_entry_t entry)
+int free_swap_and_cache(swp_entry_t entry)
 {
-       struct swap_info_struct * p;
+       struct swap_info_struct *p;
        struct page *page = NULL;
 
-       if (is_migration_entry(entry))
-               return;
+       if (non_swap_entry(entry))
+               return 1;
 
        p = swap_info_get(entry);
        if (p) {
-               if (swap_entry_free(p, swp_offset(entry)) == 1) {
+               if (swap_entry_free(p, entry, 1) == SWAP_HAS_CACHE) {
                        page = find_get_page(&swapper_space, entry.val);
-                       if (page && unlikely(!trylock_page(page))) {
+                       if (page && !trylock_page(page)) {
                                page_cache_release(page);
                                page = NULL;
                        }
@@ -430,22 +708,52 @@ void free_swap_and_cache(swp_entry_t entry)
                spin_unlock(&swap_lock);
        }
        if (page) {
-               int one_user;
-
-               BUG_ON(PagePrivate(page));
-               one_user = (page_count(page) == 2);
-               /* Only cache user (+us), or swap space full? Free it! */
-               /* Also recheck PageSwapCache after page is locked (above) */
+               /*
+                * Not mapped elsewhere, or swap space full? Free it!
+                * Also recheck PageSwapCache now page is locked (above).
+                */
                if (PageSwapCache(page) && !PageWriteback(page) &&
-                                       (one_user || vm_swap_full())) {
+                               (!page_mapped(page) || vm_swap_full())) {
                        delete_from_swap_cache(page);
                        SetPageDirty(page);
                }
                unlock_page(page);
                page_cache_release(page);
        }
+       return p != NULL;
 }
 
+#ifdef CONFIG_CGROUP_MEM_RES_CTLR
+/**
+ * mem_cgroup_count_swap_user - count the user of a swap entry
+ * @ent: the swap entry to be checked
+ * @pagep: the pointer for the swap cache page of the entry to be stored
+ *
+ * Returns the number of the user of the swap entry. The number is valid only
+ * for swaps of anonymous pages.
+ * If the entry is found on swap cache, the page is stored to pagep with
+ * refcount of it being incremented.
+ */
+int mem_cgroup_count_swap_user(swp_entry_t ent, struct page **pagep)
+{
+       struct page *page;
+       struct swap_info_struct *p;
+       int count = 0;
+
+       page = find_get_page(&swapper_space, ent.val);
+       if (page)
+               count += page_mapcount(page);
+       p = swap_info_get(ent);
+       if (p) {
+               count += swap_count(p->swap_map[swp_offset(ent)]);
+               spin_unlock(&swap_lock);
+       }
+
+       *pagep = page;
+       return count;
+}
+#endif
+
 #ifdef CONFIG_HIBERNATION
 /*
  * Find the swap type that corresponds to given device (if any).
@@ -458,37 +766,35 @@ void free_swap_and_cache(swp_entry_t entry)
 int swap_type_of(dev_t device, sector_t offset, struct block_device **bdev_p)
 {
        struct block_device *bdev = NULL;
-       int i;
+       int type;
 
        if (device)
                bdev = bdget(device);
 
        spin_lock(&swap_lock);
-       for (i = 0; i < nr_swapfiles; i++) {
-               struct swap_info_struct *sis = swap_info + i;
+       for (type = 0; type < nr_swapfiles; type++) {
+               struct swap_info_struct *sis = swap_info[type];
 
                if (!(sis->flags & SWP_WRITEOK))
                        continue;
 
                if (!bdev) {
                        if (bdev_p)
-                               *bdev_p = sis->bdev;
+                               *bdev_p = bdgrab(sis->bdev);
 
                        spin_unlock(&swap_lock);
-                       return i;
+                       return type;
                }
                if (bdev == sis->bdev) {
-                       struct swap_extent *se;
+                       struct swap_extent *se = &sis->first_swap_extent;
 
-                       se = list_entry(sis->extent_list.next,
-                                       struct swap_extent, list);
                        if (se->start_block == offset) {
                                if (bdev_p)
-                                       *bdev_p = sis->bdev;
+                                       *bdev_p = bdgrab(sis->bdev);
 
                                spin_unlock(&swap_lock);
                                bdput(bdev);
-                               return i;
+                               return type;
                        }
                }
        }
@@ -500,6 +806,21 @@ int swap_type_of(dev_t device, sector_t offset, struct block_device **bdev_p)
 }
 
 /*
+ * Get the (PAGE_SIZE) block corresponding to given offset on the swapdev
+ * corresponding to given index in swap_info (swap type).
+ */
+sector_t swapdev_block(int type, pgoff_t offset)
+{
+       struct block_device *bdev;
+
+       if ((unsigned int)type >= nr_swapfiles)
+               return 0;
+       if (!(swap_info[type]->flags & SWP_WRITEOK))
+               return 0;
+       return map_swap_entry(swp_entry(type, offset), &bdev);
+}
+
+/*
  * Return either the total number of swap pages of given type, or the number
  * of free pages of that type (depending on @free)
  *
@@ -509,18 +830,20 @@ unsigned int count_swap_pages(int type, int free)
 {
        unsigned int n = 0;
 
-       if (type < nr_swapfiles) {
-               spin_lock(&swap_lock);
-               if (swap_info[type].flags & SWP_WRITEOK) {
-                       n = swap_info[type].pages;
+       spin_lock(&swap_lock);
+       if ((unsigned int)type < nr_swapfiles) {
+               struct swap_info_struct *sis = swap_info[type];
+
+               if (sis->flags & SWP_WRITEOK) {
+                       n = sis->pages;
                        if (free)
-                               n -= swap_info[type].inuse_pages;
+                               n -= sis->inuse_pages;
                }
-               spin_unlock(&swap_lock);
        }
+       spin_unlock(&swap_lock);
        return n;
 }
-#endif
+#endif /* CONFIG_HIBERNATION */
 
 /*
  * No need to decide whether this PTE shares the swap entry with others,
@@ -530,26 +853,31 @@ unsigned int count_swap_pages(int type, int free)
 static int unuse_pte(struct vm_area_struct *vma, pmd_t *pmd,
                unsigned long addr, swp_entry_t entry, struct page *page)
 {
+       struct mem_cgroup *ptr = NULL;
        spinlock_t *ptl;
        pte_t *pte;
        int ret = 1;
 
-       if (mem_cgroup_charge(page, vma->vm_mm, GFP_KERNEL))
+       if (mem_cgroup_try_charge_swapin(vma->vm_mm, page, GFP_KERNEL, &ptr)) {
                ret = -ENOMEM;
+               goto out_nolock;
+       }
 
        pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
        if (unlikely(!pte_same(*pte, swp_entry_to_pte(entry)))) {
                if (ret > 0)
-                       mem_cgroup_uncharge_page(page);
+                       mem_cgroup_cancel_charge_swapin(ptr);
                ret = 0;
                goto out;
        }
 
-       inc_mm_counter(vma->vm_mm, anon_rss);
+       dec_mm_counter(vma->vm_mm, MM_SWAPENTS);
+       inc_mm_counter(vma->vm_mm, MM_ANONPAGES);
        get_page(page);
        set_pte_at(vma->vm_mm, addr, pte,
                   pte_mkold(mk_pte(page, vma->vm_page_prot)));
        page_add_anon_rmap(page, vma, addr);
+       mem_cgroup_commit_charge_swapin(page, ptr);
        swap_free(entry);
        /*
         * Move the page to the active list so it is not
@@ -558,6 +886,7 @@ static int unuse_pte(struct vm_area_struct *vma, pmd_t *pmd,
        activate_page(page);
 out:
        pte_unmap_unlock(pte, ptl);
+out_nolock:
        return ret;
 }
 
@@ -644,7 +973,7 @@ static int unuse_vma(struct vm_area_struct *vma,
        unsigned long addr, end, next;
        int ret;
 
-       if (page->mapping) {
+       if (page_anon_vma(page)) {
                addr = page_address_in_vma(page, vma);
                if (addr == -EFAULT)
                        return 0;
@@ -700,7 +1029,7 @@ static unsigned int find_next_to_unuse(struct swap_info_struct *si,
 {
        unsigned int max = si->max;
        unsigned int i = prev;
-       int count;
+       unsigned char count;
 
        /*
         * No need for swap_lock here: we're just looking
@@ -723,7 +1052,7 @@ static unsigned int find_next_to_unuse(struct swap_info_struct *si,
                        i = 1;
                }
                count = si->swap_map[i];
-               if (count && count != SWAP_MAP_BAD)
+               if (count && swap_count(count) != SWAP_MAP_BAD)
                        break;
        }
        return i;
@@ -736,16 +1065,14 @@ static unsigned int find_next_to_unuse(struct swap_info_struct *si,
  */
 static int try_to_unuse(unsigned int type)
 {
-       struct swap_info_struct * si = &swap_info[type];
+       struct swap_info_struct *si = swap_info[type];
        struct mm_struct *start_mm;
-       unsigned short *swap_map;
-       unsigned short swcount;
+       unsigned char *swap_map;
+       unsigned char swcount;
        struct page *page;
        swp_entry_t entry;
        unsigned int i = 0;
        int retval = 0;
-       int reset_overflow = 0;
-       int shmem;
 
        /*
         * When searching mms for an entry, a good strategy is to
@@ -759,8 +1086,7 @@ static int try_to_unuse(unsigned int type)
         * together, child after parent.  If we race with dup_mmap(), we
         * prefer to resolve parent before child, lest we miss entries
         * duplicated after we scanned child: using last mm would invert
-        * that.  Though it's only a serious concern when an overflowed
-        * swap count is reset from SWAP_MAP_MAX, preventing a rescan.
+        * that.
         */
        start_mm = &init_mm;
        atomic_inc(&init_mm.mm_users);
@@ -776,10 +1102,10 @@ static int try_to_unuse(unsigned int type)
                        break;
                }
 
-               /* 
+               /*
                 * Get a page for the entry, using the existing swap
                 * cache page if there is one.  Otherwise, get a clean
-                * page and read the swap into it. 
+                * page and read the swap into it.
                 */
                swap_map = &si->swap_map[i];
                entry = swp_entry(type, i);
@@ -822,18 +1148,19 @@ static int try_to_unuse(unsigned int type)
 
                /*
                 * Remove all references to entry.
-                * Whenever we reach init_mm, there's no address space
-                * to search, but use it as a reminder to search shmem.
                 */
-               shmem = 0;
                swcount = *swap_map;
-               if (swcount > 1) {
-                       if (start_mm == &init_mm)
-                               shmem = shmem_unuse(entry, page);
-                       else
-                               retval = unuse_mm(start_mm, entry, page);
+               if (swap_count(swcount) == SWAP_MAP_SHMEM) {
+                       retval = shmem_unuse(entry, page);
+                       /* page has already been unlocked and released */
+                       if (retval < 0)
+                               break;
+                       continue;
                }
-               if (*swap_map > 1) {
+               if (swap_count(swcount) && start_mm != &init_mm)
+                       retval = unuse_mm(start_mm, entry, page);
+
+               if (swap_count(*swap_map)) {
                        int set_start_mm = (*swap_map >= swcount);
                        struct list_head *p = &start_mm->mmlist;
                        struct mm_struct *new_start_mm = start_mm;
@@ -843,7 +1170,7 @@ static int try_to_unuse(unsigned int type)
                        atomic_inc(&new_start_mm->mm_users);
                        atomic_inc(&prev_mm->mm_users);
                        spin_lock(&mmlist_lock);
-                       while (*swap_map > 1 && !retval && !shmem &&
+                       while (swap_count(*swap_map) && !retval &&
                                        (p = p->next) != &start_mm->mmlist) {
                                mm = list_entry(p, struct mm_struct, mmlist);
                                if (!atomic_inc_not_zero(&mm->mm_users))
@@ -855,13 +1182,13 @@ static int try_to_unuse(unsigned int type)
                                cond_resched();
 
                                swcount = *swap_map;
-                               if (swcount <= 1)
+                               if (!swap_count(swcount)) /* any usage ? */
                                        ;
-                               else if (mm == &init_mm) {
+                               else if (mm == &init_mm)
                                        set_start_mm = 1;
-                                       shmem = shmem_unuse(entry, page);
-                               } else
+                               else
                                        retval = unuse_mm(mm, entry, page);
+
                                if (set_start_mm && *swap_map < swcount) {
                                        mmput(new_start_mm);
                                        atomic_inc(&mm->mm_users);
@@ -875,13 +1202,6 @@ static int try_to_unuse(unsigned int type)
                        mmput(start_mm);
                        start_mm = new_start_mm;
                }
-               if (shmem) {
-                       /* page has already been unlocked and released */
-                       if (shmem > 0)
-                               continue;
-                       retval = shmem;
-                       break;
-               }
                if (retval) {
                        unlock_page(page);
                        page_cache_release(page);
@@ -889,26 +1209,6 @@ static int try_to_unuse(unsigned int type)
                }
 
                /*
-                * How could swap count reach 0x7fff when the maximum
-                * pid is 0x7fff, and there's no way to repeat a swap
-                * page within an mm (except in shmem, where it's the
-                * shared object which takes the reference count)?
-                * We believe SWAP_MAP_MAX cannot occur in Linux 2.4.
-                *
-                * If that's wrong, then we should worry more about
-                * exit_mmap() and do_munmap() cases described above:
-                * we might be resetting SWAP_MAP_MAX too early here.
-                * We know "Undead"s can happen, they're okay, so don't
-                * report them; but do report if we reset SWAP_MAP_MAX.
-                */
-               if (*swap_map == SWAP_MAP_MAX) {
-                       spin_lock(&swap_lock);
-                       *swap_map = 1;
-                       spin_unlock(&swap_lock);
-                       reset_overflow = 1;
-               }
-
-               /*
                 * If a reference remains (rare), we would like to leave
                 * the page in the swap cache; but try_to_unmap could
                 * then re-duplicate the entry once we drop page lock,
@@ -920,8 +1220,15 @@ static int try_to_unuse(unsigned int type)
                 * read from disk into another page.  Splitting into two
                 * pages would be incorrect if swap supported "shared
                 * private" pages, but they are handled by tmpfs files.
+                *
+                * Given how unuse_vma() targets one particular offset
+                * in an anon_vma, once the anon_vma has been determined,
+                * this splitting happens to be just what is needed to
+                * handle where KSM pages have been swapped out: re-reading
+                * is unnecessarily slow, but we can fix that later on.
                 */
-               if ((*swap_map > 1) && PageDirty(page) && PageSwapCache(page)) {
+               if (swap_count(*swap_map) &&
+                    PageDirty(page) && PageSwapCache(page)) {
                        struct writeback_control wbc = {
                                .sync_mode = WB_SYNC_NONE,
                        };
@@ -930,7 +1237,16 @@ static int try_to_unuse(unsigned int type)
                        lock_page(page);
                        wait_on_page_writeback(page);
                }
-               if (PageSwapCache(page))
+
+               /*
+                * It is conceivable that a racing task removed this page from
+                * swap cache just before we acquired the page lock at the top,
+                * or while we dropped it in unuse_mm().  The page might even
+                * be back in swap cache on another swap area: that we must not
+                * delete, since it may not have been written out to swap yet.
+                */
+               if (PageSwapCache(page) &&
+                   likely(page_private(page) == entry.val))
                        delete_from_swap_cache(page);
 
                /*
@@ -950,10 +1266,6 @@ static int try_to_unuse(unsigned int type)
        }
 
        mmput(start_mm);
-       if (reset_overflow) {
-               printk(KERN_WARNING "swapoff: cleared swap entry overflow\n");
-               swap_overflow = 0;
-       }
        return retval;
 }
 
@@ -966,10 +1278,10 @@ static int try_to_unuse(unsigned int type)
 static void drain_mmlist(void)
 {
        struct list_head *p, *next;
-       unsigned int i;
+       unsigned int type;
 
-       for (i = 0; i < nr_swapfiles; i++)
-               if (swap_info[i].inuse_pages)
+       for (type = 0; type < nr_swapfiles; type++)
+               if (swap_info[type]->inuse_pages)
                        return;
        spin_lock(&mmlist_lock);
        list_for_each_safe(p, next, &init_mm.mmlist)
@@ -979,12 +1291,23 @@ static void drain_mmlist(void)
 
 /*
  * Use this swapdev's extent info to locate the (PAGE_SIZE) block which
- * corresponds to page offset `offset'.
+ * corresponds to page offset for the specified swap entry.
+ * Note that the type of this function is sector_t, but it returns page offset
+ * into the bdev, not sector offset.
  */
-sector_t map_swap_page(struct swap_info_struct *sis, pgoff_t offset)
+static sector_t map_swap_entry(swp_entry_t entry, struct block_device **bdev)
 {
-       struct swap_extent *se = sis->curr_swap_extent;
-       struct swap_extent *start_se = se;
+       struct swap_info_struct *sis;
+       struct swap_extent *start_se;
+       struct swap_extent *se;
+       pgoff_t offset;
+
+       sis = swap_info[swp_type(entry)];
+       *bdev = sis->bdev;
+
+       offset = swp_offset(entry);
+       start_se = sis->curr_swap_extent;
+       se = start_se;
 
        for ( ; ; ) {
                struct list_head *lh;
@@ -994,40 +1317,31 @@ sector_t map_swap_page(struct swap_info_struct *sis, pgoff_t offset)
                        return se->start_block + (offset - se->start_page);
                }
                lh = se->list.next;
-               if (lh == &sis->extent_list)
-                       lh = lh->next;
                se = list_entry(lh, struct swap_extent, list);
                sis->curr_swap_extent = se;
                BUG_ON(se == start_se);         /* It *must* be present */
        }
 }
 
-#ifdef CONFIG_HIBERNATION
 /*
- * Get the (PAGE_SIZE) block corresponding to given offset on the swapdev
- * corresponding to given index in swap_info (swap type).
+ * Returns the page offset into bdev for the specified page's swap entry.
  */
-sector_t swapdev_block(int swap_type, pgoff_t offset)
+sector_t map_swap_page(struct page *page, struct block_device **bdev)
 {
-       struct swap_info_struct *sis;
-
-       if (swap_type >= nr_swapfiles)
-               return 0;
-
-       sis = swap_info + swap_type;
-       return (sis->flags & SWP_WRITEOK) ? map_swap_page(sis, offset) : 0;
+       swp_entry_t entry;
+       entry.val = page_private(page);
+       return map_swap_entry(entry, bdev);
 }
-#endif /* CONFIG_HIBERNATION */
 
 /*
  * Free all of a swapdev's extent information
  */
 static void destroy_swap_extents(struct swap_info_struct *sis)
 {
-       while (!list_empty(&sis->extent_list)) {
+       while (!list_empty(&sis->first_swap_extent.list)) {
                struct swap_extent *se;
 
-               se = list_entry(sis->extent_list.next,
+               se = list_entry(sis->first_swap_extent.list.next,
                                struct swap_extent, list);
                list_del(&se->list);
                kfree(se);
@@ -1048,8 +1362,15 @@ add_swap_extent(struct swap_info_struct *sis, unsigned long start_page,
        struct swap_extent *new_se;
        struct list_head *lh;
 
-       lh = sis->extent_list.prev;     /* The highest page extent */
-       if (lh != &sis->extent_list) {
+       if (start_page == 0) {
+               se = &sis->first_swap_extent;
+               sis->curr_swap_extent = se;
+               se->start_page = 0;
+               se->nr_pages = nr_pages;
+               se->start_block = start_block;
+               return 1;
+       } else {
+               lh = sis->first_swap_extent.list.prev;  /* Highest extent */
                se = list_entry(lh, struct swap_extent, list);
                BUG_ON(se->start_page + se->nr_pages != start_page);
                if (se->start_block + se->nr_pages == start_block) {
@@ -1069,7 +1390,7 @@ add_swap_extent(struct swap_info_struct *sis, unsigned long start_page,
        new_se->nr_pages = nr_pages;
        new_se->start_block = start_block;
 
-       list_add_tail(&new_se->list, &sis->extent_list);
+       list_add_tail(&new_se->list, &sis->first_swap_extent.list);
        return 1;
 }
 
@@ -1121,7 +1442,7 @@ static int setup_swap_extents(struct swap_info_struct *sis, sector_t *span)
        if (S_ISBLK(inode->i_mode)) {
                ret = add_swap_extent(sis, 0, sis->max, 0);
                *span = sis->pages;
-               goto done;
+               goto out;
        }
 
        blkbits = inode->i_blkbits;
@@ -1192,48 +1513,25 @@ reprobe:
        sis->max = page_no;
        sis->pages = page_no - 1;
        sis->highest_bit = page_no - 1;
-done:
-       sis->curr_swap_extent = list_entry(sis->extent_list.prev,
-                                       struct swap_extent, list);
-       goto out;
+out:
+       return ret;
 bad_bmap:
        printk(KERN_ERR "swapon: swapfile has holes\n");
        ret = -EINVAL;
-out:
-       return ret;
-}
-
-#if 0  /* We don't need this yet */
-#include <linux/backing-dev.h>
-int page_queue_congested(struct page *page)
-{
-       struct backing_dev_info *bdi;
-
-       BUG_ON(!PageLocked(page));      /* It pins the swap_info_struct */
-
-       if (PageSwapCache(page)) {
-               swp_entry_t entry = { .val = page_private(page) };
-               struct swap_info_struct *sis;
-
-               sis = get_swap_info_struct(swp_type(entry));
-               bdi = sis->bdev->bd_inode->i_mapping->backing_dev_info;
-       } else
-               bdi = page->mapping->backing_dev_info;
-       return bdi_write_congested(bdi);
+       goto out;
 }
-#endif
 
-asmlinkage long sys_swapoff(const char __user * specialfile)
+SYSCALL_DEFINE1(swapoff, const char __user *, specialfile)
 {
-       struct swap_info_struct * p = NULL;
-       unsigned short *swap_map;
+       struct swap_info_struct *p = NULL;
+       unsigned char *swap_map;
        struct file *swap_file, *victim;
        struct address_space *mapping;
        struct inode *inode;
-       char * pathname;
+       char *pathname;
        int i, type, prev;
        int err;
-       
+
        if (!capable(CAP_SYS_ADMIN))
                return -EPERM;
 
@@ -1251,9 +1549,9 @@ asmlinkage long sys_swapoff(const char __user * specialfile)
        mapping = victim->f_mapping;
        prev = -1;
        spin_lock(&swap_lock);
-       for (type = swap_list.head; type >= 0; type = swap_info[type].next) {
-               p = swap_info + type;
-               if ((p->flags & SWP_ACTIVE) == SWP_ACTIVE) {
+       for (type = swap_list.head; type >= 0; type = swap_info[type]->next) {
+               p = swap_info[type];
+               if (p->flags & SWP_WRITEOK) {
                        if (p->swap_file->f_mapping == mapping)
                                break;
                }
@@ -1271,18 +1569,17 @@ asmlinkage long sys_swapoff(const char __user * specialfile)
                spin_unlock(&swap_lock);
                goto out_dput;
        }
-       if (prev < 0) {
+       if (prev < 0)
                swap_list.head = p->next;
-       } else {
-               swap_info[prev].next = p->next;
-       }
+       else
+               swap_info[prev]->next = p->next;
        if (type == swap_list.next) {
                /* just pick something that's safe... */
                swap_list.next = swap_list.head;
        }
        if (p->prio < 0) {
-               for (i = p->next; i >= 0; i = swap_info[i].next)
-                       swap_info[i].prio = p->prio--;
+               for (i = p->next; i >= 0; i = swap_info[i]->next)
+                       swap_info[i]->prio = p->prio--;
                least_priority++;
        }
        nr_swap_pages -= p->pages;
@@ -1290,9 +1587,9 @@ asmlinkage long sys_swapoff(const char __user * specialfile)
        p->flags &= ~SWP_WRITEOK;
        spin_unlock(&swap_lock);
 
-       current->flags |= PF_SWAPOFF;
+       current->flags |= PF_OOM_ORIGIN;
        err = try_to_unuse(type);
-       current->flags &= ~PF_SWAPOFF;
+       current->flags &= ~PF_OOM_ORIGIN;
 
        if (err) {
                /* re-insert swap space back into swap_list */
@@ -1300,16 +1597,16 @@ asmlinkage long sys_swapoff(const char __user * specialfile)
                if (p->prio < 0)
                        p->prio = --least_priority;
                prev = -1;
-               for (i = swap_list.head; i >= 0; i = swap_info[i].next) {
-                       if (p->prio >= swap_info[i].prio)
+               for (i = swap_list.head; i >= 0; i = swap_info[i]->next) {
+                       if (p->prio >= swap_info[i]->prio)
                                break;
                        prev = i;
                }
                p->next = i;
                if (prev < 0)
-                       swap_list.head = swap_list.next = p - swap_info;
+                       swap_list.head = swap_list.next = type;
                else
-                       swap_info[prev].next = p - swap_info;
+                       swap_info[prev]->next = type;
                nr_swap_pages += p->pages;
                total_swap_pages += p->pages;
                p->flags |= SWP_WRITEOK;
@@ -1322,6 +1619,9 @@ asmlinkage long sys_swapoff(const char __user * specialfile)
        up_write(&swap_unplug_sem);
 
        destroy_swap_extents(p);
+       if (p->flags & SWP_CONTINUED)
+               free_swap_count_continuations(p);
+
        mutex_lock(&swapon_mutex);
        spin_lock(&swap_lock);
        drain_mmlist();
@@ -1343,6 +1643,9 @@ asmlinkage long sys_swapoff(const char __user * specialfile)
        spin_unlock(&swap_lock);
        mutex_unlock(&swapon_mutex);
        vfree(swap_map);
+       /* Destroy swap account informatin */
+       swap_cgroup_swapoff(type);
+
        inode = mapping->host;
        if (S_ISBLK(inode->i_mode)) {
                struct block_device *bdev = I_BDEV(inode);
@@ -1366,8 +1669,8 @@ out:
 /* iterator */
 static void *swap_start(struct seq_file *swap, loff_t *pos)
 {
-       struct swap_info_struct *ptr = swap_info;
-       int i;
+       struct swap_info_struct *si;
+       int type;
        loff_t l = *pos;
 
        mutex_lock(&swapon_mutex);
@@ -1375,11 +1678,13 @@ static void *swap_start(struct seq_file *swap, loff_t *pos)
        if (!l)
                return SEQ_START_TOKEN;
 
-       for (i = 0; i < nr_swapfiles; i++, ptr++) {
-               if (!(ptr->flags & SWP_USED) || !ptr->swap_map)
+       for (type = 0; type < nr_swapfiles; type++) {
+               smp_rmb();      /* read nr_swapfiles before swap_info[type] */
+               si = swap_info[type];
+               if (!(si->flags & SWP_USED) || !si->swap_map)
                        continue;
                if (!--l)
-                       return ptr;
+                       return si;
        }
 
        return NULL;
@@ -1387,21 +1692,21 @@ static void *swap_start(struct seq_file *swap, loff_t *pos)
 
 static void *swap_next(struct seq_file *swap, void *v, loff_t *pos)
 {
-       struct swap_info_struct *ptr;
-       struct swap_info_struct *endptr = swap_info + nr_swapfiles;
+       struct swap_info_struct *si = v;
+       int type;
 
        if (v == SEQ_START_TOKEN)
-               ptr = swap_info;
-       else {
-               ptr = v;
-               ptr++;
-       }
+               type = 0;
+       else
+               type = si->type + 1;
 
-       for (; ptr < endptr; ptr++) {
-               if (!(ptr->flags & SWP_USED) || !ptr->swap_map)
+       for (; type < nr_swapfiles; type++) {
+               smp_rmb();      /* read nr_swapfiles before swap_info[type] */
+               si = swap_info[type];
+               if (!(si->flags & SWP_USED) || !si->swap_map)
                        continue;
                ++*pos;
-               return ptr;
+               return si;
        }
 
        return NULL;
@@ -1414,24 +1719,24 @@ static void swap_stop(struct seq_file *swap, void *v)
 
 static int swap_show(struct seq_file *swap, void *v)
 {
-       struct swap_info_struct *ptr = v;
+       struct swap_info_struct *si = v;
        struct file *file;
        int len;
 
-       if (ptr == SEQ_START_TOKEN) {
+       if (si == SEQ_START_TOKEN) {
                seq_puts(swap,"Filename\t\t\t\tType\t\tSize\tUsed\tPriority\n");
                return 0;
        }
 
-       file = ptr->swap_file;
+       file = si->swap_file;
        len = seq_path(swap, &file->f_path, " \t\n\\");
        seq_printf(swap, "%*s%s\t%u\t%u\t%d\n",
-                      len < 40 ? 40 - len : 1, " ",
-                      S_ISBLK(file->f_path.dentry->d_inode->i_mode) ?
+                       len < 40 ? 40 - len : 1, " ",
+                       S_ISBLK(file->f_path.dentry->d_inode->i_mode) ?
                                "partition" : "file\t",
-                      ptr->pages << (PAGE_SHIFT - 10),
-                      ptr->inuse_pages << (PAGE_SHIFT - 10),
-                      ptr->prio);
+                       si->pages << (PAGE_SHIFT - 10),
+                       si->inuse_pages << (PAGE_SHIFT - 10),
+                       si->prio);
        return 0;
 }
 
@@ -1462,14 +1767,23 @@ static int __init procswaps_init(void)
 __initcall(procswaps_init);
 #endif /* CONFIG_PROC_FS */
 
+#ifdef MAX_SWAPFILES_CHECK
+static int __init max_swapfiles_check(void)
+{
+       MAX_SWAPFILES_CHECK();
+       return 0;
+}
+late_initcall(max_swapfiles_check);
+#endif
+
 /*
  * Written 01/25/92 by Simmule Turner, heavily changed by Linus.
  *
  * The swapon system call
  */
-asmlinkage long sys_swapon(const char __user * specialfile, int swap_flags)
+SYSCALL_DEFINE2(swapon, const char __user *, specialfile, int, swap_flags)
 {
-       struct swap_info_struct * p;
+       struct swap_info_struct *p;
        char *name = NULL;
        struct block_device *bdev = NULL;
        struct file *swap_file = NULL;
@@ -1477,37 +1791,58 @@ asmlinkage long sys_swapon(const char __user * specialfile, int swap_flags)
        unsigned int type;
        int i, prev;
        int error;
-       union swap_header *swap_header = NULL;
-       int swap_header_version;
-       unsigned int nr_good_pages = 0;
+       union swap_header *swap_header;
+       unsigned int nr_good_pages;
        int nr_extents = 0;
        sector_t span;
-       unsigned long maxpages = 1;
-       int swapfilesize;
-       unsigned short *swap_map = NULL;
+       unsigned long maxpages;
+       unsigned long swapfilepages;
+       unsigned char *swap_map = NULL;
        struct page *page = NULL;
        struct inode *inode = NULL;
        int did_down = 0;
 
        if (!capable(CAP_SYS_ADMIN))
                return -EPERM;
+
+       p = kzalloc(sizeof(*p), GFP_KERNEL);
+       if (!p)
+               return -ENOMEM;
+
        spin_lock(&swap_lock);
-       p = swap_info;
-       for (type = 0 ; type < nr_swapfiles ; type++,p++)
-               if (!(p->flags & SWP_USED))
+       for (type = 0; type < nr_swapfiles; type++) {
+               if (!(swap_info[type]->flags & SWP_USED))
                        break;
+       }
        error = -EPERM;
        if (type >= MAX_SWAPFILES) {
                spin_unlock(&swap_lock);
+               kfree(p);
                goto out;
        }
-       if (type >= nr_swapfiles)
-               nr_swapfiles = type+1;
-       memset(p, 0, sizeof(*p));
-       INIT_LIST_HEAD(&p->extent_list);
+       if (type >= nr_swapfiles) {
+               p->type = type;
+               swap_info[type] = p;
+               /*
+                * Write swap_info[type] before nr_swapfiles, in case a
+                * racing procfs swap_start() or swap_next() is reading them.
+                * (We never shrink nr_swapfiles, we never free this entry.)
+                */
+               smp_wmb();
+               nr_swapfiles++;
+       } else {
+               kfree(p);
+               p = swap_info[type];
+               /*
+                * Do not memset this entry: a racing procfs swap_next()
+                * would be relying on p->type to remain valid.
+                */
+       }
+       INIT_LIST_HEAD(&p->first_swap_extent.list);
        p->flags = SWP_USED;
        p->next = -1;
        spin_unlock(&swap_lock);
+
        name = getname(specialfile);
        error = PTR_ERR(name);
        if (IS_ERR(name)) {
@@ -1527,7 +1862,7 @@ asmlinkage long sys_swapon(const char __user * specialfile, int swap_flags)
 
        error = -EBUSY;
        for (i = 0; i < nr_swapfiles; i++) {
-               struct swap_info_struct *q = &swap_info[i];
+               struct swap_info_struct *q = swap_info[i];
 
                if (i == type || !q->swap_file)
                        continue;
@@ -1561,7 +1896,7 @@ asmlinkage long sys_swapon(const char __user * specialfile, int swap_flags)
                goto bad_swap;
        }
 
-       swapfilesize = i_size_read(inode) >> PAGE_SHIFT;
+       swapfilepages = i_size_read(inode) >> PAGE_SHIFT;
 
        /*
         * Read the swap header.
@@ -1575,102 +1910,98 @@ asmlinkage long sys_swapon(const char __user * specialfile, int swap_flags)
                error = PTR_ERR(page);
                goto bad_swap;
        }
-       kmap(page);
-       swap_header = page_address(page);
+       swap_header = kmap(page);
 
-       if (!memcmp("SWAP-SPACE",swap_header->magic.magic,10))
-               swap_header_version = 1;
-       else if (!memcmp("SWAPSPACE2",swap_header->magic.magic,10))
-               swap_header_version = 2;
-       else {
+       if (memcmp("SWAPSPACE2", swap_header->magic.magic, 10)) {
                printk(KERN_ERR "Unable to find swap-space signature\n");
                error = -EINVAL;
                goto bad_swap;
        }
-       
-       switch (swap_header_version) {
-       case 1:
-               printk(KERN_ERR "version 0 swap is no longer supported. "
-                       "Use mkswap -v1 %s\n", name);
+
+       /* swap partition endianess hack... */
+       if (swab32(swap_header->info.version) == 1) {
+               swab32s(&swap_header->info.version);
+               swab32s(&swap_header->info.last_page);
+               swab32s(&swap_header->info.nr_badpages);
+               for (i = 0; i < swap_header->info.nr_badpages; i++)
+                       swab32s(&swap_header->info.badpages[i]);
+       }
+       /* Check the swap header's sub-version */
+       if (swap_header->info.version != 1) {
+               printk(KERN_WARNING
+                      "Unable to handle swap header version %d\n",
+                      swap_header->info.version);
                error = -EINVAL;
                goto bad_swap;
-       case 2:
-               /* swap partition endianess hack... */
-               if (swab32(swap_header->info.version) == 1) {
-                       swab32s(&swap_header->info.version);
-                       swab32s(&swap_header->info.last_page);
-                       swab32s(&swap_header->info.nr_badpages);
-                       for (i = 0; i < swap_header->info.nr_badpages; i++)
-                               swab32s(&swap_header->info.badpages[i]);
-               }
-               /* Check the swap header's sub-version and the size of
-                   the swap file and bad block lists */
-               if (swap_header->info.version != 1) {
-                       printk(KERN_WARNING
-                              "Unable to handle swap header version %d\n",
-                              swap_header->info.version);
-                       error = -EINVAL;
-                       goto bad_swap;
-               }
+       }
 
-               p->lowest_bit  = 1;
-               p->cluster_next = 1;
+       p->lowest_bit  = 1;
+       p->cluster_next = 1;
+       p->cluster_nr = 0;
 
-               /*
-                * Find out how many pages are allowed for a single swap
-                * device. There are two limiting factors: 1) the number of
-                * bits for the swap offset in the swp_entry_t type and
-                * 2) the number of bits in the a swap pte as defined by
-                * the different architectures. In order to find the
-                * largest possible bit mask a swap entry with swap type 0
-                * and swap offset ~0UL is created, encoded to a swap pte,
-                * decoded to a swp_entry_t again and finally the swap
-                * offset is extracted. This will mask all the bits from
-                * the initial ~0UL mask that can't be encoded in either
-                * the swp_entry_t or the architecture definition of a
-                * swap pte.
-                */
-               maxpages = swp_offset(pte_to_swp_entry(swp_entry_to_pte(swp_entry(0,~0UL)))) - 1;
-               if (maxpages > swap_header->info.last_page)
-                       maxpages = swap_header->info.last_page;
-               p->highest_bit = maxpages - 1;
+       /*
+        * Find out how many pages are allowed for a single swap
+        * device. There are two limiting factors: 1) the number of
+        * bits for the swap offset in the swp_entry_t type and
+        * 2) the number of bits in the a swap pte as defined by
+        * the different architectures. In order to find the
+        * largest possible bit mask a swap entry with swap type 0
+        * and swap offset ~0UL is created, encoded to a swap pte,
+        * decoded to a swp_entry_t again and finally the swap
+        * offset is extracted. This will mask all the bits from
+        * the initial ~0UL mask that can't be encoded in either
+        * the swp_entry_t or the architecture definition of a
+        * swap pte.
+        */
+       maxpages = swp_offset(pte_to_swp_entry(
+                       swp_entry_to_pte(swp_entry(0, ~0UL)))) + 1;
+       if (maxpages > swap_header->info.last_page) {
+               maxpages = swap_header->info.last_page + 1;
+               /* p->max is an unsigned int: don't overflow it */
+               if ((unsigned int)maxpages == 0)
+                       maxpages = UINT_MAX;
+       }
+       p->highest_bit = maxpages - 1;
 
-               error = -EINVAL;
-               if (!maxpages)
-                       goto bad_swap;
-               if (swapfilesize && maxpages > swapfilesize) {
-                       printk(KERN_WARNING
-                              "Swap area shorter than signature indicates\n");
-                       goto bad_swap;
-               }
-               if (swap_header->info.nr_badpages && S_ISREG(inode->i_mode))
-                       goto bad_swap;
-               if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
-                       goto bad_swap;
+       error = -EINVAL;
+       if (!maxpages)
+               goto bad_swap;
+       if (swapfilepages && maxpages > swapfilepages) {
+               printk(KERN_WARNING
+                      "Swap area shorter than signature indicates\n");
+               goto bad_swap;
+       }
+       if (swap_header->info.nr_badpages && S_ISREG(inode->i_mode))
+               goto bad_swap;
+       if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
+               goto bad_swap;
+
+       /* OK, set up the swap map and apply the bad block list */
+       swap_map = vmalloc(maxpages);
+       if (!swap_map) {
+               error = -ENOMEM;
+               goto bad_swap;
+       }
 
-               /* OK, set up the swap map and apply the bad block list */
-               swap_map = vmalloc(maxpages * sizeof(short));
-               if (!swap_map) {
-                       error = -ENOMEM;
+       memset(swap_map, 0, maxpages);
+       nr_good_pages = maxpages - 1;   /* omit header page */
+
+       for (i = 0; i < swap_header->info.nr_badpages; i++) {
+               unsigned int page_nr = swap_header->info.badpages[i];
+               if (page_nr == 0 || page_nr > swap_header->info.last_page) {
+                       error = -EINVAL;
                        goto bad_swap;
                }
-
-               error = 0;
-               memset(swap_map, 0, maxpages * sizeof(short));
-               for (i = 0; i < swap_header->info.nr_badpages; i++) {
-                       int page_nr = swap_header->info.badpages[i];
-                       if (page_nr <= 0 || page_nr >= swap_header->info.last_page)
-                               error = -EINVAL;
-                       else
-                               swap_map[page_nr] = SWAP_MAP_BAD;
+               if (page_nr < maxpages) {
+                       swap_map[page_nr] = SWAP_MAP_BAD;
+                       nr_good_pages--;
                }
-               nr_good_pages = swap_header->info.last_page -
-                               swap_header->info.nr_badpages -
-                               1 /* header page */;
-               if (error)
-                       goto bad_swap;
        }
 
+       error = swap_cgroup_swapon(type, maxpages);
+       if (error)
+               goto bad_swap;
+
        if (nr_good_pages) {
                swap_map[0] = SWAP_MAP_BAD;
                p->max = maxpages;
@@ -1688,6 +2019,15 @@ asmlinkage long sys_swapon(const char __user * specialfile, int swap_flags)
                goto bad_swap;
        }
 
+       if (p->bdev) {
+               if (blk_queue_nonrot(bdev_get_queue(p->bdev))) {
+                       p->flags |= SWP_SOLIDSTATE;
+                       p->cluster_next = 1 + (random32() % p->highest_bit);
+               }
+               if (discard_swap(p) == 0)
+                       p->flags |= SWP_DISCARDABLE;
+       }
+
        mutex_lock(&swapon_mutex);
        spin_lock(&swap_lock);
        if (swap_flags & SWAP_FLAG_PREFER)
@@ -1696,29 +2036,29 @@ asmlinkage long sys_swapon(const char __user * specialfile, int swap_flags)
        else
                p->prio = --least_priority;
        p->swap_map = swap_map;
-       p->flags = SWP_ACTIVE;
+       p->flags |= SWP_WRITEOK;
        nr_swap_pages += nr_good_pages;
        total_swap_pages += nr_good_pages;
 
        printk(KERN_INFO "Adding %uk swap on %s.  "
-                       "Priority:%d extents:%d across:%lluk\n",
+                       "Priority:%d extents:%d across:%lluk %s%s\n",
                nr_good_pages<<(PAGE_SHIFT-10), name, p->prio,
-               nr_extents, (unsigned long long)span<<(PAGE_SHIFT-10));
+               nr_extents, (unsigned long long)span<<(PAGE_SHIFT-10),
+               (p->flags & SWP_SOLIDSTATE) ? "SS" : "",
+               (p->flags & SWP_DISCARDABLE) ? "D" : "");
 
        /* insert swap space into swap_list: */
        prev = -1;
-       for (i = swap_list.head; i >= 0; i = swap_info[i].next) {
-               if (p->prio >= swap_info[i].prio) {
+       for (i = swap_list.head; i >= 0; i = swap_info[i]->next) {
+               if (p->prio >= swap_info[i]->prio)
                        break;
-               }
                prev = i;
        }
        p->next = i;
-       if (prev < 0) {
-               swap_list.head = swap_list.next = p - swap_info;
-       } else {
-               swap_info[prev].next = p - swap_info;
-       }
+       if (prev < 0)
+               swap_list.head = swap_list.next = type;
+       else
+               swap_info[prev]->next = type;
        spin_unlock(&swap_lock);
        mutex_unlock(&swapon_mutex);
        error = 0;
@@ -1729,6 +2069,7 @@ bad_swap:
                bd_release(bdev);
        }
        destroy_swap_extents(p);
+       swap_cgroup_swapoff(type);
 bad_swap_2:
        spin_lock(&swap_lock);
        p->swap_file = NULL;
@@ -1754,15 +2095,15 @@ out:
 
 void si_swapinfo(struct sysinfo *val)
 {
-       unsigned int i;
+       unsigned int type;
        unsigned long nr_to_be_unused = 0;
 
        spin_lock(&swap_lock);
-       for (i = 0; i < nr_swapfiles; i++) {
-               if (!(swap_info[i].flags & SWP_USED) ||
-                    (swap_info[i].flags & SWP_WRITEOK))
-                       continue;
-               nr_to_be_unused += swap_info[i].inuse_pages;
+       for (type = 0; type < nr_swapfiles; type++) {
+               struct swap_info_struct *si = swap_info[type];
+
+               if ((si->flags & SWP_USED) && !(si->flags & SWP_WRITEOK))
+                       nr_to_be_unused += si->inuse_pages;
        }
        val->freeswap = nr_swap_pages + nr_to_be_unused;
        val->totalswap = total_swap_pages + nr_to_be_unused;
@@ -1772,49 +2113,111 @@ void si_swapinfo(struct sysinfo *val)
 /*
  * Verify that a swap entry is valid and increment its swap map count.
  *
- * Note: if swap_map[] reaches SWAP_MAP_MAX the entries are treated as
- * "permanent", but will be reclaimed by the next swapoff.
+ * Returns error code in following case.
+ * - success -> 0
+ * - swp_entry is invalid -> EINVAL
+ * - swp_entry is migration entry -> EINVAL
+ * - swap-cache reference is requested but there is already one. -> EEXIST
+ * - swap-cache reference is requested but the entry is not used. -> ENOENT
+ * - swap-mapped reference requested but needs continued swap count. -> ENOMEM
  */
-int swap_duplicate(swp_entry_t entry)
+static int __swap_duplicate(swp_entry_t entry, unsigned char usage)
 {
-       struct swap_info_struct * p;
+       struct swap_info_struct *p;
        unsigned long offset, type;
-       int result = 0;
+       unsigned char count;
+       unsigned char has_cache;
+       int err = -EINVAL;
 
-       if (is_migration_entry(entry))
-               return 1;
+       if (non_swap_entry(entry))
+               goto out;
 
        type = swp_type(entry);
        if (type >= nr_swapfiles)
                goto bad_file;
-       p = type + swap_info;
+       p = swap_info[type];
        offset = swp_offset(entry);
 
        spin_lock(&swap_lock);
-       if (offset < p->max && p->swap_map[offset]) {
-               if (p->swap_map[offset] < SWAP_MAP_MAX - 1) {
-                       p->swap_map[offset]++;
-                       result = 1;
-               } else if (p->swap_map[offset] <= SWAP_MAP_MAX) {
-                       if (swap_overflow++ < 5)
-                               printk(KERN_WARNING "swap_dup: swap entry overflow\n");
-                       p->swap_map[offset] = SWAP_MAP_MAX;
-                       result = 1;
-               }
-       }
+       if (unlikely(offset >= p->max))
+               goto unlock_out;
+
+       count = p->swap_map[offset];
+       has_cache = count & SWAP_HAS_CACHE;
+       count &= ~SWAP_HAS_CACHE;
+       err = 0;
+
+       if (usage == SWAP_HAS_CACHE) {
+
+               /* set SWAP_HAS_CACHE if there is no cache and entry is used */
+               if (!has_cache && count)
+                       has_cache = SWAP_HAS_CACHE;
+               else if (has_cache)             /* someone else added cache */
+                       err = -EEXIST;
+               else                            /* no users remaining */
+                       err = -ENOENT;
+
+       } else if (count || has_cache) {
+
+               if ((count & ~COUNT_CONTINUED) < SWAP_MAP_MAX)
+                       count += usage;
+               else if ((count & ~COUNT_CONTINUED) > SWAP_MAP_MAX)
+                       err = -EINVAL;
+               else if (swap_count_continued(p, offset, count))
+                       count = COUNT_CONTINUED;
+               else
+                       err = -ENOMEM;
+       } else
+               err = -ENOENT;                  /* unused swap entry */
+
+       p->swap_map[offset] = count | has_cache;
+
+unlock_out:
        spin_unlock(&swap_lock);
 out:
-       return result;
+       return err;
 
 bad_file:
        printk(KERN_ERR "swap_dup: %s%08lx\n", Bad_file, entry.val);
        goto out;
 }
 
-struct swap_info_struct *
-get_swap_info_struct(unsigned type)
+/*
+ * Help swapoff by noting that swap entry belongs to shmem/tmpfs
+ * (in which case its reference count is never incremented).
+ */
+void swap_shmem_alloc(swp_entry_t entry)
+{
+       __swap_duplicate(entry, SWAP_MAP_SHMEM);
+}
+
+/*
+ * Increase reference count of swap entry by 1.
+ * Returns 0 for success, or -ENOMEM if a swap_count_continuation is required
+ * but could not be atomically allocated.  Returns 0, just as if it succeeded,
+ * if __swap_duplicate() fails for another reason (-EINVAL or -ENOENT), which
+ * might occur if a page table entry has got corrupted.
+ */
+int swap_duplicate(swp_entry_t entry)
+{
+       int err = 0;
+
+       while (!err && __swap_duplicate(entry, 1) == -ENOMEM)
+               err = add_swap_count_continuation(entry, GFP_ATOMIC);
+       return err;
+}
+
+/*
+ * @entry: swap entry for which we allocate swap cache.
+ *
+ * Called when allocating swap cache for existing swap entry,
+ * This can return error codes. Returns 0 at success.
+ * -EBUSY means there is a swap cache.
+ * Note: return code is different from swap_duplicate().
+ */
+int swapcache_prepare(swp_entry_t entry)
 {
-       return &swap_info[type];
+       return __swap_duplicate(entry, SWAP_HAS_CACHE);
 }
 
 /*
@@ -1832,7 +2235,7 @@ int valid_swaphandles(swp_entry_t entry, unsigned long *offset)
        if (!our_page_cluster)  /* no readahead */
                return 0;
 
-       si = &swap_info[swp_type(entry)];
+       si = swap_info[swp_type(entry)];
        target = swp_offset(entry);
        base = (target >> our_page_cluster) << our_page_cluster;
        end = base + (1 << our_page_cluster);
@@ -1848,7 +2251,7 @@ int valid_swaphandles(swp_entry_t entry, unsigned long *offset)
                /* Don't read in free or bad pages */
                if (!si->swap_map[toff])
                        break;
-               if (si->swap_map[toff] == SWAP_MAP_BAD)
+               if (swap_count(si->swap_map[toff]) == SWAP_MAP_BAD)
                        break;
        }
        /* Count contiguous allocated slots below our target */
@@ -1856,7 +2259,7 @@ int valid_swaphandles(swp_entry_t entry, unsigned long *offset)
                /* Don't read in free or bad pages */
                if (!si->swap_map[toff])
                        break;
-               if (si->swap_map[toff] == SWAP_MAP_BAD)
+               if (swap_count(si->swap_map[toff]) == SWAP_MAP_BAD)
                        break;
        }
        spin_unlock(&swap_lock);
@@ -1868,3 +2271,219 @@ int valid_swaphandles(swp_entry_t entry, unsigned long *offset)
        *offset = ++toff;
        return nr_pages? ++nr_pages: 0;
 }
+
+/*
+ * add_swap_count_continuation - called when a swap count is duplicated
+ * beyond SWAP_MAP_MAX, it allocates a new page and links that to the entry's
+ * page of the original vmalloc'ed swap_map, to hold the continuation count
+ * (for that entry and for its neighbouring PAGE_SIZE swap entries).  Called
+ * again when count is duplicated beyond SWAP_MAP_MAX * SWAP_CONT_MAX, etc.
+ *
+ * These continuation pages are seldom referenced: the common paths all work
+ * on the original swap_map, only referring to a continuation page when the
+ * low "digit" of a count is incremented or decremented through SWAP_MAP_MAX.
+ *
+ * add_swap_count_continuation(, GFP_ATOMIC) can be called while holding
+ * page table locks; if it fails, add_swap_count_continuation(, GFP_KERNEL)
+ * can be called after dropping locks.
+ */
+int add_swap_count_continuation(swp_entry_t entry, gfp_t gfp_mask)
+{
+       struct swap_info_struct *si;
+       struct page *head;
+       struct page *page;
+       struct page *list_page;
+       pgoff_t offset;
+       unsigned char count;
+
+       /*
+        * When debugging, it's easier to use __GFP_ZERO here; but it's better
+        * for latency not to zero a page while GFP_ATOMIC and holding locks.
+        */
+       page = alloc_page(gfp_mask | __GFP_HIGHMEM);
+
+       si = swap_info_get(entry);
+       if (!si) {
+               /*
+                * An acceptable race has occurred since the failing
+                * __swap_duplicate(): the swap entry has been freed,
+                * perhaps even the whole swap_map cleared for swapoff.
+                */
+               goto outer;
+       }
+
+       offset = swp_offset(entry);
+       count = si->swap_map[offset] & ~SWAP_HAS_CACHE;
+
+       if ((count & ~COUNT_CONTINUED) != SWAP_MAP_MAX) {
+               /*
+                * The higher the swap count, the more likely it is that tasks
+                * will race to add swap count continuation: we need to avoid
+                * over-provisioning.
+                */
+               goto out;
+       }
+
+       if (!page) {
+               spin_unlock(&swap_lock);
+               return -ENOMEM;
+       }
+
+       /*
+        * We are fortunate that although vmalloc_to_page uses pte_offset_map,
+        * no architecture is using highmem pages for kernel pagetables: so it
+        * will not corrupt the GFP_ATOMIC caller's atomic pagetable kmaps.
+        */
+       head = vmalloc_to_page(si->swap_map + offset);
+       offset &= ~PAGE_MASK;
+
+       /*
+        * Page allocation does not initialize the page's lru field,
+        * but it does always reset its private field.
+        */
+       if (!page_private(head)) {
+               BUG_ON(count & COUNT_CONTINUED);
+               INIT_LIST_HEAD(&head->lru);
+               set_page_private(head, SWP_CONTINUED);
+               si->flags |= SWP_CONTINUED;
+       }
+
+       list_for_each_entry(list_page, &head->lru, lru) {
+               unsigned char *map;
+
+               /*
+                * If the previous map said no continuation, but we've found
+                * a continuation page, free our allocation and use this one.
+                */
+               if (!(count & COUNT_CONTINUED))
+                       goto out;
+
+               map = kmap_atomic(list_page, KM_USER0) + offset;
+               count = *map;
+               kunmap_atomic(map, KM_USER0);
+
+               /*
+                * If this continuation count now has some space in it,
+                * free our allocation and use this one.
+                */
+               if ((count & ~COUNT_CONTINUED) != SWAP_CONT_MAX)
+                       goto out;
+       }
+
+       list_add_tail(&page->lru, &head->lru);
+       page = NULL;                    /* now it's attached, don't free it */
+out:
+       spin_unlock(&swap_lock);
+outer:
+       if (page)
+               __free_page(page);
+       return 0;
+}
+
+/*
+ * swap_count_continued - when the original swap_map count is incremented
+ * from SWAP_MAP_MAX, check if there is already a continuation page to carry
+ * into, carry if so, or else fail until a new continuation page is allocated;
+ * when the original swap_map count is decremented from 0 with continuation,
+ * borrow from the continuation and report whether it still holds more.
+ * Called while __swap_duplicate() or swap_entry_free() holds swap_lock.
+ */
+static bool swap_count_continued(struct swap_info_struct *si,
+                                pgoff_t offset, unsigned char count)
+{
+       struct page *head;
+       struct page *page;
+       unsigned char *map;
+
+       head = vmalloc_to_page(si->swap_map + offset);
+       if (page_private(head) != SWP_CONTINUED) {
+               BUG_ON(count & COUNT_CONTINUED);
+               return false;           /* need to add count continuation */
+       }
+
+       offset &= ~PAGE_MASK;
+       page = list_entry(head->lru.next, struct page, lru);
+       map = kmap_atomic(page, KM_USER0) + offset;
+
+       if (count == SWAP_MAP_MAX)      /* initial increment from swap_map */
+               goto init_map;          /* jump over SWAP_CONT_MAX checks */
+
+       if (count == (SWAP_MAP_MAX | COUNT_CONTINUED)) { /* incrementing */
+               /*
+                * Think of how you add 1 to 999
+                */
+               while (*map == (SWAP_CONT_MAX | COUNT_CONTINUED)) {
+                       kunmap_atomic(map, KM_USER0);
+                       page = list_entry(page->lru.next, struct page, lru);
+                       BUG_ON(page == head);
+                       map = kmap_atomic(page, KM_USER0) + offset;
+               }
+               if (*map == SWAP_CONT_MAX) {
+                       kunmap_atomic(map, KM_USER0);
+                       page = list_entry(page->lru.next, struct page, lru);
+                       if (page == head)
+                               return false;   /* add count continuation */
+                       map = kmap_atomic(page, KM_USER0) + offset;
+init_map:              *map = 0;               /* we didn't zero the page */
+               }
+               *map += 1;
+               kunmap_atomic(map, KM_USER0);
+               page = list_entry(page->lru.prev, struct page, lru);
+               while (page != head) {
+                       map = kmap_atomic(page, KM_USER0) + offset;
+                       *map = COUNT_CONTINUED;
+                       kunmap_atomic(map, KM_USER0);
+                       page = list_entry(page->lru.prev, struct page, lru);
+               }
+               return true;                    /* incremented */
+
+       } else {                                /* decrementing */
+               /*
+                * Think of how you subtract 1 from 1000
+                */
+               BUG_ON(count != COUNT_CONTINUED);
+               while (*map == COUNT_CONTINUED) {
+                       kunmap_atomic(map, KM_USER0);
+                       page = list_entry(page->lru.next, struct page, lru);
+                       BUG_ON(page == head);
+                       map = kmap_atomic(page, KM_USER0) + offset;
+               }
+               BUG_ON(*map == 0);
+               *map -= 1;
+               if (*map == 0)
+                       count = 0;
+               kunmap_atomic(map, KM_USER0);
+               page = list_entry(page->lru.prev, struct page, lru);
+               while (page != head) {
+                       map = kmap_atomic(page, KM_USER0) + offset;
+                       *map = SWAP_CONT_MAX | count;
+                       count = COUNT_CONTINUED;
+                       kunmap_atomic(map, KM_USER0);
+                       page = list_entry(page->lru.prev, struct page, lru);
+               }
+               return count == COUNT_CONTINUED;
+       }
+}
+
+/*
+ * free_swap_count_continuations - swapoff free all the continuation pages
+ * appended to the swap_map, after swap_map is quiesced, before vfree'ing it.
+ */
+static void free_swap_count_continuations(struct swap_info_struct *si)
+{
+       pgoff_t offset;
+
+       for (offset = 0; offset < si->max; offset += PAGE_SIZE) {
+               struct page *head;
+               head = vmalloc_to_page(si->swap_map + offset);
+               if (page_private(head)) {
+                       struct list_head *this, *next;
+                       list_for_each_safe(this, next, &head->lru) {
+                               struct page *page;
+                               page = list_entry(this, struct page, lru);
+                               list_del(this);
+                               __free_page(page);
+                       }
+               }
+       }
+}