tmpfs: convert shmem_writepage and enable swap
authorHugh Dickins <hughd@google.com>
Wed, 3 Aug 2011 23:21:25 +0000 (16:21 -0700)
committerLinus Torvalds <torvalds@linux-foundation.org>
Thu, 4 Aug 2011 00:25:24 +0000 (14:25 -1000)
Convert shmem_writepage() to use shmem_delete_from_page_cache() to use
shmem_radix_tree_replace() to substitute swap entry for page pointer
atomically in the radix tree.

As with shmem_add_to_page_cache(), it's not entirely satisfactory to be
copying such code from delete_from_swap_cache, but again judged easier
to sell than making its other callers go through the extras.

Remove the toy implementation's shmem_put_swap() and shmem_get_swap(),
now unreferenced, and the hack to disable swap: it's now good to go.

The way things have worked out, info->lock no longer helps to guard the
shmem_swaplist: we increment swapped under shmem_swaplist_mutex only.
That global mutex exclusion between shmem_writepage() and shmem_unuse()
is not pretty, and we ought to find another way; but it's been forced on
us by recent race discoveries, not a consequence of this patchset.

And what has become of the WARN_ON_ONCE(1) free_swap_and_cache() if a
swap entry was found already present? That's no longer possible, the
(unknown) one inserting this page into filecache would hit the swap
entry occupying that slot.

Signed-off-by: Hugh Dickins <hughd@google.com>
Acked-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
mm/shmem.c

index 13ef2d7e912df21ea69316a0ef34371ca106cb27..0f094a2585262f6fceb1d289e17f747c2a745438 100644 (file)
@@ -6,7 +6,8 @@
  *              2000-2001 Christoph Rohland
  *              2000-2001 SAP AG
  *              2002 Red Hat Inc.
- * Copyright (C) 2002-2005 Hugh Dickins.
+ * Copyright (C) 2002-2011 Hugh Dickins.
+ * Copyright (C) 2011 Google Inc.
  * Copyright (C) 2002-2005 VERITAS Software Corporation.
  * Copyright (C) 2004 Andi Kleen, SuSE Labs
  *
@@ -219,19 +220,6 @@ static void shmem_recalc_inode(struct inode *inode)
        }
 }
 
-static void shmem_put_swap(struct shmem_inode_info *info, pgoff_t index,
-                          swp_entry_t swap)
-{
-       if (index < SHMEM_NR_DIRECT)
-               info->i_direct[index] = swap;
-}
-
-static swp_entry_t shmem_get_swap(struct shmem_inode_info *info, pgoff_t index)
-{
-       return (index < SHMEM_NR_DIRECT) ?
-               info->i_direct[index] : (swp_entry_t){0};
-}
-
 /*
  * Replace item expected in radix tree by a new item, while holding tree lock.
  */
@@ -299,6 +287,25 @@ static int shmem_add_to_page_cache(struct page *page,
        return error;
 }
 
+/*
+ * Like delete_from_page_cache, but substitutes swap for page.
+ */
+static void shmem_delete_from_page_cache(struct page *page, void *radswap)
+{
+       struct address_space *mapping = page->mapping;
+       int error;
+
+       spin_lock_irq(&mapping->tree_lock);
+       error = shmem_radix_tree_replace(mapping, page->index, page, radswap);
+       page->mapping = NULL;
+       mapping->nrpages--;
+       __dec_zone_page_state(page, NR_FILE_PAGES);
+       __dec_zone_page_state(page, NR_SHMEM);
+       spin_unlock_irq(&mapping->tree_lock);
+       page_cache_release(page);
+       BUG_ON(error);
+}
+
 /*
  * Like find_get_pages, but collecting swap entries as well as pages.
  */
@@ -664,14 +671,10 @@ int shmem_unuse(swp_entry_t swap, struct page *page)
        mutex_lock(&shmem_swaplist_mutex);
        list_for_each_safe(this, next, &shmem_swaplist) {
                info = list_entry(this, struct shmem_inode_info, swaplist);
-               if (!info->swapped) {
-                       spin_lock(&info->lock);
-                       if (!info->swapped)
-                               list_del_init(&info->swaplist);
-                       spin_unlock(&info->lock);
-               }
                if (info->swapped)
                        found = shmem_unuse_inode(info, swap, page);
+               else
+                       list_del_init(&info->swaplist);
                cond_resched();
                if (found)
                        break;
@@ -694,10 +697,10 @@ out:
 static int shmem_writepage(struct page *page, struct writeback_control *wbc)
 {
        struct shmem_inode_info *info;
-       swp_entry_t swap, oswap;
        struct address_space *mapping;
-       pgoff_t index;
        struct inode *inode;
+       swp_entry_t swap;
+       pgoff_t index;
 
        BUG_ON(!PageLocked(page));
        mapping = page->mapping;
@@ -720,55 +723,38 @@ static int shmem_writepage(struct page *page, struct writeback_control *wbc)
                WARN_ON_ONCE(1);        /* Still happens? Tell us about it! */
                goto redirty;
        }
-
-       /*
-        * Disable even the toy swapping implementation, while we convert
-        * functions one by one to having swap entries in the radix tree.
-        */
-       if (index < ULONG_MAX)
-               goto redirty;
-
        swap = get_swap_page();
        if (!swap.val)
                goto redirty;
 
        /*
         * Add inode to shmem_unuse()'s list of swapped-out inodes,
-        * if it's not already there.  Do it now because we cannot take
-        * mutex while holding spinlock, and must do so before the page
-        * is moved to swap cache, when its pagelock no longer protects
+        * if it's not already there.  Do it now before the page is
+        * moved to swap cache, when its pagelock no longer protects
         * the inode from eviction.  But don't unlock the mutex until
-        * we've taken the spinlock, because shmem_unuse_inode() will
-        * prune a !swapped inode from the swaplist under both locks.
+        * we've incremented swapped, because shmem_unuse_inode() will
+        * prune a !swapped inode from the swaplist under this mutex.
         */
        mutex_lock(&shmem_swaplist_mutex);
        if (list_empty(&info->swaplist))
                list_add_tail(&info->swaplist, &shmem_swaplist);
 
-       spin_lock(&info->lock);
-       mutex_unlock(&shmem_swaplist_mutex);
-
-       oswap = shmem_get_swap(info, index);
-       if (oswap.val) {
-               WARN_ON_ONCE(1);        /* Still happens? Tell us about it! */
-               free_swap_and_cache(oswap);
-               shmem_put_swap(info, index, (swp_entry_t){0});
-               info->swapped--;
-       }
-       shmem_recalc_inode(inode);
-
        if (add_to_swap_cache(page, swap, GFP_ATOMIC) == 0) {
-               delete_from_page_cache(page);
-               shmem_put_swap(info, index, swap);
-               info->swapped++;
                swap_shmem_alloc(swap);
+               shmem_delete_from_page_cache(page, swp_to_radix_entry(swap));
+
+               spin_lock(&info->lock);
+               info->swapped++;
+               shmem_recalc_inode(inode);
                spin_unlock(&info->lock);
+
+               mutex_unlock(&shmem_swaplist_mutex);
                BUG_ON(page_mapped(page));
                swap_writepage(page, wbc);
                return 0;
        }
 
-       spin_unlock(&info->lock);
+       mutex_unlock(&shmem_swaplist_mutex);
        swapcache_free(swap, NULL);
 redirty:
        set_page_dirty(page);