]> err.no Git - linux-2.6/commitdiff
[PATCH] Swapless page migration: modify core logic
authorChristoph Lameter <clameter@sgi.com>
Fri, 23 Jun 2006 09:03:37 +0000 (02:03 -0700)
committerLinus Torvalds <torvalds@g5.osdl.org>
Fri, 23 Jun 2006 14:42:50 +0000 (07:42 -0700)
Use the migration entries for page migration

This modifies the migration code to use the new migration entries.  It now
becomes possible to migrate anonymous pages without having to add a swap
entry.

We add a couple of new functions to replace migration entries with the proper
ptes.

We cannot take the tree_lock for migrating anonymous pages anymore.  However,
we know that we hold the only remaining reference to the page when the page
count reaches 1.

Signed-off-by: Christoph Lameter <clameter@sgi.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
mm/Kconfig
mm/migrate.c

index 332f5c29b53a3c04665e4d727fc50dbfde609fb2..66e65ab3942651e86f2fbbadcf914fd600e41600 100644 (file)
@@ -138,8 +138,8 @@ config SPLIT_PTLOCK_CPUS
 #
 config MIGRATION
        bool "Page migration"
-       def_bool y if NUMA
-       depends on SWAP && NUMA
+       def_bool y
+       depends on NUMA
        help
          Allows the migration of the physical location of pages of processes
          while the virtual addresses are not changed. This is useful for
index 81721a061d50a4711e40a3663bd1a43529d85e45..8f91463eab4e2906018d816f2364f3270eb12131 100644 (file)
@@ -254,14 +254,20 @@ static int migrate_page_move_mapping(struct address_space *mapping,
 {
        struct page **radix_pointer;
 
+       if (!mapping) {
+               /* Anonymous page */
+               if (page_count(page) != 1)
+                       return -EAGAIN;
+               return 0;
+       }
+
        write_lock_irq(&mapping->tree_lock);
 
        radix_pointer = (struct page **)radix_tree_lookup_slot(
                                                &mapping->page_tree,
                                                page_index(page));
 
-       if (!page_mapping(page) ||
-                       page_count(page) != 2 + !!PagePrivate(page) ||
+       if (page_count(page) != 2 + !!PagePrivate(page) ||
                        *radix_pointer != page) {
                write_unlock_irq(&mapping->tree_lock);
                return -EAGAIN;
@@ -271,10 +277,12 @@ static int migrate_page_move_mapping(struct address_space *mapping,
         * Now we know that no one else is looking at the page.
         */
        get_page(newpage);
+#ifdef CONFIG_SWAP
        if (PageSwapCache(page)) {
                SetPageSwapCache(newpage);
                set_page_private(newpage, page_private(page));
        }
+#endif
 
        *radix_pointer = newpage;
        __put_page(page);
@@ -308,7 +316,9 @@ static void migrate_page_copy(struct page *newpage, struct page *page)
                set_page_dirty(newpage);
        }
 
+#ifdef CONFIG_SWAP
        ClearPageSwapCache(page);
+#endif
        ClearPageActive(page);
        ClearPagePrivate(page);
        set_page_private(page, 0);
@@ -353,16 +363,6 @@ int migrate_page(struct address_space *mapping,
                return rc;
 
        migrate_page_copy(newpage, page);
-
-       /*
-        * Remove auxiliary swap entries and replace
-        * them with real ptes.
-        *
-        * Note that a real pte entry will allow processes that are not
-        * waiting on the page lock to use the new page via the page tables
-        * before the new page is unlocked.
-        */
-       remove_from_swap(newpage);
        return 0;
 }
 EXPORT_SYMBOL(migrate_page);
@@ -530,23 +530,7 @@ redo:
                                goto unlock_page;
 
                /*
-                * Establish swap ptes for anonymous pages or destroy pte
-                * maps for files.
-                *
-                * In order to reestablish file backed mappings the fault handlers
-                * will take the radix tree_lock which may then be used to stop
-                * processses from accessing this page until the new page is ready.
-                *
-                * A process accessing via a swap pte (an anonymous page) will take a
-                * page_lock on the old page which will block the process until the
-                * migration attempt is complete. At that time the PageSwapCache bit
-                * will be examined. If the page was migrated then the PageSwapCache
-                * bit will be clear and the operation to retrieve the page will be
-                * retried which will find the new page in the radix tree. Then a new
-                * direct mapping may be generated based on the radix tree contents.
-                *
-                * If the page was not migrated then the PageSwapCache bit
-                * is still set and the operation may continue.
+                * Establish migration ptes or remove ptes
                 */
                rc = -EPERM;
                if (try_to_unmap(page, 1) == SWAP_FAIL)
@@ -569,9 +553,9 @@ redo:
                 */
                mapping = page_mapping(page);
                if (!mapping)
-                       goto unlock_both;
+                       rc = migrate_page(mapping, newpage, page);
 
-               if (mapping->a_ops->migratepage)
+               else if (mapping->a_ops->migratepage)
                        /*
                         * Most pages have a mapping and most filesystems
                         * should provide a migration function. Anonymous
@@ -584,10 +568,15 @@ redo:
                else
                        rc = fallback_migrate_page(mapping, newpage, page);
 
-unlock_both:
+               if (!rc)
+                       remove_migration_ptes(page, newpage);
+
                unlock_page(newpage);
 
 unlock_page:
+               if (rc)
+                       remove_migration_ptes(page, page);
+
                unlock_page(page);
 
 next: