[PATCH] page invalidation cleanup
authorNick Piggin <nickpiggin@yahoo.com.au>
Wed, 27 Sep 2006 08:50:02 +0000 (01:50 -0700)
committerLinus Torvalds <torvalds@g5.osdl.org>
Wed, 27 Sep 2006 15:26:12 +0000 (08:26 -0700)
Clean up the invalidate code, and use a common function to safely remove
the page from pagecache.

Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Hugh Dickins <hugh@veritas.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
mm/truncate.c
mm/vmscan.c

index c6ab55e..a654928 100644 (file)
@@ -9,6 +9,7 @@
 
 #include <linux/kernel.h>
 #include <linux/mm.h>
+#include <linux/swap.h>
 #include <linux/module.h>
 #include <linux/pagemap.h>
 #include <linux/pagevec.h>
@@ -52,36 +53,26 @@ truncate_complete_page(struct address_space *mapping, struct page *page)
 /*
  * This is for invalidate_inode_pages().  That function can be called at
  * any time, and is not supposed to throw away dirty pages.  But pages can
- * be marked dirty at any time too.  So we re-check the dirtiness inside
- * ->tree_lock.  That provides exclusion against the __set_page_dirty
- * functions.
+ * be marked dirty at any time too, so use remove_mapping which safely
+ * discards clean, unused pages.
  *
  * Returns non-zero if the page was successfully invalidated.
  */
 static int
 invalidate_complete_page(struct address_space *mapping, struct page *page)
 {
+       int ret;
+
        if (page->mapping != mapping)
                return 0;
 
        if (PagePrivate(page) && !try_to_release_page(page, 0))
                return 0;
 
-       write_lock_irq(&mapping->tree_lock);
-       if (PageDirty(page))
-               goto failed;
-       if (page_count(page) != 2)      /* caller's ref + pagecache ref */
-               goto failed;
-
-       BUG_ON(PagePrivate(page));
-       __remove_from_page_cache(page);
-       write_unlock_irq(&mapping->tree_lock);
+       ret = remove_mapping(mapping, page);
        ClearPageUptodate(page);
-       page_cache_release(page);       /* pagecache ref */
-       return 1;
-failed:
-       write_unlock_irq(&mapping->tree_lock);
-       return 0;
+
+       return ret;
 }
 
 /**
index 87d3409..eca7031 100644 (file)
@@ -384,11 +384,30 @@ int remove_mapping(struct address_space *mapping, struct page *page)
        BUG_ON(mapping != page_mapping(page));
 
        write_lock_irq(&mapping->tree_lock);
-
        /*
-        * The non-racy check for busy page.  It is critical to check
-        * PageDirty _after_ making sure that the page is freeable and
-        * not in use by anybody.       (pagecache + us == 2)
+        * The non racy check for a busy page.
+        *
+        * Must be careful with the order of the tests. When someone has
+        * a ref to the page, it may be possible that they dirty it then
+        * drop the reference. So if PageDirty is tested before page_count
+        * here, then the following race may occur:
+        *
+        * get_user_pages(&page);
+        * [user mapping goes away]
+        * write_to(page);
+        *                              !PageDirty(page)    [good]
+        * SetPageDirty(page);
+        * put_page(page);
+        *                              !page_count(page)   [good, discard it]
+        *
+        * [oops, our write_to data is lost]
+        *
+        * Reversing the order of the tests ensures such a situation cannot
+        * escape unnoticed. The smp_rmb is needed to ensure the page->flags
+        * load is not satisfied before that of page->_count.
+        *
+        * Note that if SetPageDirty is always performed via set_page_dirty,
+        * and thus under tree_lock, then this ordering is not required.
         */
        if (unlikely(page_count(page) != 2))
                goto cannot_free;