[PATCH] page migration reorg
Christoph Lameter [Wed, 22 Mar 2006 08:09:12 +0000 (00:09 -0800)]
Centralize the page migration functions in anticipation of additional
tinkering.  Creates a new file mm/migrate.c

1. Extract buffer_migrate_page() from fs/buffer.c

2. Extract central migration code from vmscan.c

3. Extract some components from mempolicy.c

4. Export pageout() and remove_from_swap() from vmscan.c

5. Make it possible to configure NUMA systems without page migration
   and non-NUMA systems with page migration.

I had to so some #ifdeffing in mempolicy.c that may need a cleanup.

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

fs/buffer.c
fs/xfs/linux-2.6/xfs_buf.c
include/linux/migrate.h [new file with mode: 0644]
include/linux/swap.h
mm/Kconfig
mm/Makefile
mm/mempolicy.c
mm/migrate.c [new file with mode: 0644]
mm/swap_state.c
mm/vmscan.c

index a9b3994..1d3683d 100644 (file)
@@ -3051,68 +3051,6 @@ asmlinkage long sys_bdflush(int func, long data)
 }
 
 /*
- * Migration function for pages with buffers. This function can only be used
- * if the underlying filesystem guarantees that no other references to "page"
- * exist.
- */
-#ifdef CONFIG_MIGRATION
-int buffer_migrate_page(struct page *newpage, struct page *page)
-{
-       struct address_space *mapping = page->mapping;
-       struct buffer_head *bh, *head;
-       int rc;
-
-       if (!mapping)
-               return -EAGAIN;
-
-       if (!page_has_buffers(page))
-               return migrate_page(newpage, page);
-
-       head = page_buffers(page);
-
-       rc = migrate_page_remove_references(newpage, page, 3);
-       if (rc)
-               return rc;
-
-       bh = head;
-       do {
-               get_bh(bh);
-               lock_buffer(bh);
-               bh = bh->b_this_page;
-
-       } while (bh != head);
-
-       ClearPagePrivate(page);
-       set_page_private(newpage, page_private(page));
-       set_page_private(page, 0);
-       put_page(page);
-       get_page(newpage);
-
-       bh = head;
-       do {
-               set_bh_page(bh, newpage, bh_offset(bh));
-               bh = bh->b_this_page;
-
-       } while (bh != head);
-
-       SetPagePrivate(newpage);
-
-       migrate_page_copy(newpage, page);
-
-       bh = head;
-       do {
-               unlock_buffer(bh);
-               put_bh(bh);
-               bh = bh->b_this_page;
-
-       } while (bh != head);
-
-       return 0;
-}
-EXPORT_SYMBOL(buffer_migrate_page);
-#endif
-
-/*
  * Buffer-head allocation
  */
 static kmem_cache_t *bh_cachep;
index bfb4f29..8cdfa41 100644 (file)
@@ -29,6 +29,7 @@
 #include <linux/blkdev.h>
 #include <linux/hash.h>
 #include <linux/kthread.h>
+#include <linux/migrate.h>
 #include "xfs_linux.h"
 
 STATIC kmem_zone_t *xfs_buf_zone;
diff --git a/include/linux/migrate.h b/include/linux/migrate.h
new file mode 100644 (file)
index 0000000..7d09962
--- /dev/null
@@ -0,0 +1,36 @@
+#ifndef _LINUX_MIGRATE_H
+#define _LINUX_MIGRATE_H
+
+#include <linux/config.h>
+#include <linux/mm.h>
+
+#ifdef CONFIG_MIGRATION
+extern int isolate_lru_page(struct page *p, struct list_head *pagelist);
+extern int putback_lru_pages(struct list_head *l);
+extern int migrate_page(struct page *, struct page *);
+extern void migrate_page_copy(struct page *, struct page *);
+extern int migrate_page_remove_references(struct page *, struct page *, int);
+extern int migrate_pages(struct list_head *l, struct list_head *t,
+               struct list_head *moved, struct list_head *failed);
+int migrate_pages_to(struct list_head *pagelist,
+                       struct vm_area_struct *vma, int dest);
+extern int fail_migrate_page(struct page *, struct page *);
+
+extern int migrate_prep(void);
+
+#else
+
+static inline int isolate_lru_page(struct page *p, struct list_head *list)
+                                       { return -ENOSYS; }
+static inline int putback_lru_pages(struct list_head *l) { return 0; }
+static inline int migrate_pages(struct list_head *l, struct list_head *t,
+       struct list_head *moved, struct list_head *failed) { return -ENOSYS; }
+
+static inline int migrate_prep(void) { return -ENOSYS; }
+
+/* Possible settings for the migrate_page() method in address_operations */
+#define migrate_page NULL
+#define fail_migrate_page NULL
+
+#endif /* CONFIG_MIGRATION */
+#endif /* _LINUX_MIGRATE_H */
index 3dc6c89..12415dd 100644 (file)
@@ -175,6 +175,21 @@ extern void swap_setup(void);
 extern unsigned long try_to_free_pages(struct zone **, gfp_t);
 extern unsigned long shrink_all_memory(unsigned long nr_pages);
 extern int vm_swappiness;
+extern int remove_mapping(struct address_space *mapping, struct page *page);
+
+/* possible outcome of pageout() */
+typedef enum {
+       /* failed to write page out, page is locked */
+       PAGE_KEEP,
+       /* move page to the active list, page is locked */
+       PAGE_ACTIVATE,
+       /* page has been sent to the disk successfully, page is unlocked */
+       PAGE_SUCCESS,
+       /* page is clean and locked */
+       PAGE_CLEAN,
+} pageout_t;
+
+extern pageout_t pageout(struct page *page, struct address_space *mapping);
 
 #ifdef CONFIG_NUMA
 extern int zone_reclaim_mode;
@@ -188,25 +203,6 @@ static inline int zone_reclaim(struct zone *z, gfp_t mask, unsigned int order)
 }
 #endif
 
-#ifdef CONFIG_MIGRATION
-extern int isolate_lru_page(struct page *p);
-extern unsigned long putback_lru_pages(struct list_head *l);
-extern int migrate_page(struct page *, struct page *);
-extern void migrate_page_copy(struct page *, struct page *);
-extern int migrate_page_remove_references(struct page *, struct page *, int);
-extern unsigned long migrate_pages(struct list_head *l, struct list_head *t,
-               struct list_head *moved, struct list_head *failed);
-extern int fail_migrate_page(struct page *, struct page *);
-#else
-static inline int isolate_lru_page(struct page *p) { return -ENOSYS; }
-static inline int putback_lru_pages(struct list_head *l) { return 0; }
-static inline int migrate_pages(struct list_head *l, struct list_head *t,
-       struct list_head *moved, struct list_head *failed) { return -ENOSYS; }
-/* Possible settings for the migrate_page() method in address_operations */
-#define migrate_page NULL
-#define fail_migrate_page NULL
-#endif
-
 #ifdef CONFIG_MMU
 /* linux/mm/shmem.c */
 extern int shmem_unuse(swp_entry_t entry, struct page *page);
index a9cb80a..bd80460 100644 (file)
@@ -137,5 +137,11 @@ config SPLIT_PTLOCK_CPUS
 # support for page migration
 #
 config MIGRATION
+       bool "Page migration"
        def_bool y if NUMA || SPARSEMEM || DISCONTIGMEM
        depends on SWAP
+       help
+         Allows the migration of the physical location of pages of processes
+         while the virtual addresses are not changed. This is useful for
+         example on NUMA systems to put pages nearer to the processors accessing
+         the page.
index 9aa03fa..f10c753 100644 (file)
@@ -22,3 +22,5 @@ obj-$(CONFIG_SLOB) += slob.o
 obj-$(CONFIG_SLAB) += slab.o
 obj-$(CONFIG_MEMORY_HOTPLUG) += memory_hotplug.o
 obj-$(CONFIG_FS_XIP) += filemap_xip.o
+obj-$(CONFIG_MIGRATION) += migrate.o
+
index 96195dc..e93cc74 100644 (file)
@@ -86,6 +86,7 @@
 #include <linux/swap.h>
 #include <linux/seq_file.h>
 #include <linux/proc_fs.h>
+#include <linux/migrate.h>
 
 #include <asm/tlbflush.h>
 #include <asm/uaccess.h>
@@ -95,9 +96,6 @@
 #define MPOL_MF_INVERT (MPOL_MF_INTERNAL << 1)         /* Invert check for nodemask */
 #define MPOL_MF_STATS (MPOL_MF_INTERNAL << 2)          /* Gather statistics */
 
-/* The number of pages to migrate per call to migrate_pages() */
-#define MIGRATE_CHUNK_SIZE 256
-
 static struct kmem_cache *policy_cache;
 static struct kmem_cache *sn_cache;
 
@@ -331,17 +329,10 @@ check_range(struct mm_struct *mm, unsigned long start, unsigned long end,
        struct vm_area_struct *first, *vma, *prev;
 
        if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
-               /* Must have swap device for migration */
-               if (nr_swap_pages <= 0)
-                       return ERR_PTR(-ENODEV);
 
-               /*
-                * Clear the LRU lists so pages can be isolated.
-                * Note that pages may be moved off the LRU after we have
-                * drained them. Those pages will fail to migrate like other
-                * pages that may be busy.
-                */
-               lru_add_drain_all();
+               err = migrate_prep();
+               if (err)
+                       return ERR_PTR(err);
        }
 
        first = find_vma(mm, start);
@@ -550,92 +541,18 @@ long do_get_mempolicy(int *policy, nodemask_t *nmask,
        return err;
 }
 
+#ifdef CONFIG_MIGRATION
 /*
  * page migration
  */
-
 static void migrate_page_add(struct page *page, struct list_head *pagelist,
                                unsigned long flags)
 {
        /*
         * Avoid migrating a page that is shared with others.
         */
-       if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(page) == 1) {
-               if (isolate_lru_page(page))
-                       list_add_tail(&page->lru, pagelist);
-       }
-}
-
-/*
- * Migrate the list 'pagelist' of pages to a certain destination.
- *
- * Specify destination with either non-NULL vma or dest_node >= 0
- * Return the number of pages not migrated or error code
- */
-static int migrate_pages_to(struct list_head *pagelist,
-                       struct vm_area_struct *vma, int dest)
-{
-       LIST_HEAD(newlist);
-       LIST_HEAD(moved);
-       LIST_HEAD(failed);
-       int err = 0;
-       unsigned long offset = 0;
-       int nr_pages;
-       struct page *page;
-       struct list_head *p;
-
-redo:
-       nr_pages = 0;
-       list_for_each(p, pagelist) {
-               if (vma) {
-                       /*
-                        * The address passed to alloc_page_vma is used to
-                        * generate the proper interleave behavior. We fake
-                        * the address here by an increasing offset in order
-                        * to get the proper distribution of pages.
-                        *
-                        * No decision has been made as to which page
-                        * a certain old page is moved to so we cannot
-                        * specify the correct address.
-                        */
-                       page = alloc_page_vma(GFP_HIGHUSER, vma,
-                                       offset + vma->vm_start);
-                       offset += PAGE_SIZE;
-               }
-               else
-                       page = alloc_pages_node(dest, GFP_HIGHUSER, 0);
-
-               if (!page) {
-                       err = -ENOMEM;
-                       goto out;
-               }
-               list_add_tail(&page->lru, &newlist);
-               nr_pages++;
-               if (nr_pages > MIGRATE_CHUNK_SIZE)
-                       break;
-       }
-       err = migrate_pages(pagelist, &newlist, &moved, &failed);
-
-       putback_lru_pages(&moved);      /* Call release pages instead ?? */
-
-       if (err >= 0 && list_empty(&newlist) && !list_empty(pagelist))
-               goto redo;
-out:
-       /* Return leftover allocated pages */
-       while (!list_empty(&newlist)) {
-               page = list_entry(newlist.next, struct page, lru);
-               list_del(&page->lru);
-               __free_page(page);
-       }
-       list_splice(&failed, pagelist);
-       if (err < 0)
-               return err;
-
-       /* Calculate number of leftover pages */
-       nr_pages = 0;
-       list_for_each(p, pagelist)
-               nr_pages++;
-       return nr_pages;
+       if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(page) == 1)
+               isolate_lru_page(page, pagelist);
 }
 
 /*
@@ -742,8 +659,23 @@ int do_migrate_pages(struct mm_struct *mm,
        if (err < 0)
                return err;
        return busy;
+
 }
 
+#else
+
+static void migrate_page_add(struct page *page, struct list_head *pagelist,
+                               unsigned long flags)
+{
+}
+
+int do_migrate_pages(struct mm_struct *mm,
+       const nodemask_t *from_nodes, const nodemask_t *to_nodes, int flags)
+{
+       return -ENOSYS;
+}
+#endif
+
 long do_mbind(unsigned long start, unsigned long len,
                unsigned long mode, nodemask_t *nmask, unsigned long flags)
 {
@@ -808,6 +740,7 @@ long do_mbind(unsigned long start, unsigned long len,
                if (!err && nr_failed && (flags & MPOL_MF_STRICT))
                        err = -EIO;
        }
+
        if (!list_empty(&pagelist))
                putback_lru_pages(&pagelist);
 
diff --git a/mm/migrate.c b/mm/migrate.c
new file mode 100644 (file)
index 0000000..09f6e4a
--- /dev/null
@@ -0,0 +1,655 @@
+/*
+ * Memory Migration functionality - linux/mm/migration.c
+ *
+ * Copyright (C) 2006 Silicon Graphics, Inc., Christoph Lameter
+ *
+ * Page migration was first developed in the context of the memory hotplug
+ * project. The main authors of the migration code are:
+ *
+ * IWAMOTO Toshihiro <iwamoto@valinux.co.jp>
+ * Hirokazu Takahashi <taka@valinux.co.jp>
+ * Dave Hansen <haveblue@us.ibm.com>
+ * Christoph Lameter <clameter@sgi.com>
+ */
+
+#include <linux/migrate.h>
+#include <linux/module.h>
+#include <linux/swap.h>
+#include <linux/pagemap.h>
+#include <linux/buffer_head.h> /* for try_to_release_page(),
+                                       buffer_heads_over_limit */
+#include <linux/mm_inline.h>
+#include <linux/pagevec.h>
+#include <linux/rmap.h>
+#include <linux/topology.h>
+#include <linux/cpu.h>
+#include <linux/cpuset.h>
+#include <linux/swapops.h>
+
+#include "internal.h"
+
+#include "internal.h"
+
+/* The maximum number of pages to take off the LRU for migration */
+#define MIGRATE_CHUNK_SIZE 256
+
+#define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
+
+/*
+ * Isolate one page from the LRU lists. If successful put it onto
+ * the indicated list with elevated page count.
+ *
+ * Result:
+ *  -EBUSY: page not on LRU list
+ *  0: page removed from LRU list and added to the specified list.
+ */
+int isolate_lru_page(struct page *page, struct list_head *pagelist)
+{
+       int ret = -EBUSY;
+
+       if (PageLRU(page)) {
+               struct zone *zone = page_zone(page);
+
+               spin_lock_irq(&zone->lru_lock);
+               if (PageLRU(page)) {
+                       ret = 0;
+                       get_page(page);
+                       ClearPageLRU(page);
+                       if (PageActive(page))
+                               del_page_from_active_list(zone, page);
+                       else
+                               del_page_from_inactive_list(zone, page);
+                       list_add_tail(&page->lru, pagelist);
+               }
+               spin_unlock_irq(&zone->lru_lock);
+       }
+       return ret;
+}
+
+/*
+ * migrate_prep() needs to be called after we have compiled the list of pages
+ * to be migrated using isolate_lru_page() but before we begin a series of calls
+ * to migrate_pages().
+ */
+int migrate_prep(void)
+{
+       /* Must have swap device for migration */
+       if (nr_swap_pages <= 0)
+               return -ENODEV;
+
+       /*
+        * Clear the LRU lists so pages can be isolated.
+        * Note that pages may be moved off the LRU after we have
+        * drained them. Those pages will fail to migrate like other
+        * pages that may be busy.
+        */
+       lru_add_drain_all();
+
+       return 0;
+}
+
+static inline void move_to_lru(struct page *page)
+{
+       list_del(&page->lru);
+       if (PageActive(page)) {
+               /*
+                * lru_cache_add_active checks that
+                * the PG_active bit is off.
+                */
+               ClearPageActive(page);
+               lru_cache_add_active(page);
+       } else {
+               lru_cache_add(page);
+       }
+       put_page(page);
+}
+
+/*
+ * Add isolated pages on the list back to the LRU.
+ *
+ * returns the number of pages put back.
+ */
+int putback_lru_pages(struct list_head *l)
+{
+       struct page *page;
+       struct page *page2;
+       int count = 0;
+
+       list_for_each_entry_safe(page, page2, l, lru) {
+               move_to_lru(page);
+               count++;
+       }
+       return count;
+}
+
+/*
+ * Non migratable page
+ */
+int fail_migrate_page(struct page *newpage, struct page *page)
+{
+       return -EIO;
+}
+EXPORT_SYMBOL(fail_migrate_page);
+
+/*
+ * swapout a single page
+ * page is locked upon entry, unlocked on exit
+ */
+static int swap_page(struct page *page)
+{
+       struct address_space *mapping = page_mapping(page);
+
+       if (page_mapped(page) && mapping)
+               if (try_to_unmap(page, 1) != SWAP_SUCCESS)
+                       goto unlock_retry;
+
+       if (PageDirty(page)) {
+               /* Page is dirty, try to write it out here */
+               switch(pageout(page, mapping)) {
+               case PAGE_KEEP:
+               case PAGE_ACTIVATE:
+                       goto unlock_retry;
+
+               case PAGE_SUCCESS:
+                       goto retry;
+
+               case PAGE_CLEAN:
+                       ; /* try to free the page below */
+               }
+       }
+
+       if (PagePrivate(page)) {
+               if (!try_to_release_page(page, GFP_KERNEL) ||
+                   (!mapping && page_count(page) == 1))
+                       goto unlock_retry;
+       }
+
+       if (remove_mapping(mapping, page)) {
+               /* Success */
+               unlock_page(page);
+               return 0;
+       }
+
+unlock_retry:
+       unlock_page(page);
+
+retry:
+       return -EAGAIN;
+}
+EXPORT_SYMBOL(swap_page);
+
+/*
+ * Remove references for a page and establish the new page with the correct
+ * basic settings to be able to stop accesses to the page.
+ */
+int migrate_page_remove_references(struct page *newpage,
+                               struct page *page, int nr_refs)
+{
+       struct address_space *mapping = page_mapping(page);
+       struct page **radix_pointer;
+
+       /*
+        * Avoid doing any of the following work if the page count
+        * indicates that the page is in use or truncate has removed
+        * the page.
+        */
+       if (!mapping || page_mapcount(page) + nr_refs != page_count(page))
+               return -EAGAIN;
+
+       /*
+        * 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.
+        */
+       if (try_to_unmap(page, 1) == SWAP_FAIL)
+               /* A vma has VM_LOCKED set -> permanent failure */
+               return -EPERM;
+
+       /*
+        * Give up if we were unable to remove all mappings.
+        */
+       if (page_mapcount(page))
+               return -EAGAIN;
+
+       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) != nr_refs ||
+                       *radix_pointer != page) {
+               write_unlock_irq(&mapping->tree_lock);
+               return 1;
+       }
+
+       /*
+        * Now we know that no one else is looking at the page.
+        *
+        * Certain minimal information about a page must be available
+        * in order for other subsystems to properly handle the page if they
+        * find it through the radix tree update before we are finished
+        * copying the page.
+        */
+       get_page(newpage);
+       newpage->index = page->index;
+       newpage->mapping = page->mapping;
+       if (PageSwapCache(page)) {
+               SetPageSwapCache(newpage);
+               set_page_private(newpage, page_private(page));
+       }
+
+       *radix_pointer = newpage;
+       __put_page(page);
+       write_unlock_irq(&mapping->tree_lock);
+
+       return 0;
+}
+EXPORT_SYMBOL(migrate_page_remove_references);
+
+/*
+ * Copy the page to its new location
+ */
+void migrate_page_copy(struct page *newpage, struct page *page)
+{
+       copy_highpage(newpage, page);
+
+       if (PageError(page))
+               SetPageError(newpage);
+       if (PageReferenced(page))
+               SetPageReferenced(newpage);
+       if (PageUptodate(page))
+               SetPageUptodate(newpage);
+       if (PageActive(page))
+               SetPageActive(newpage);
+       if (PageChecked(page))
+               SetPageChecked(newpage);
+       if (PageMappedToDisk(page))
+               SetPageMappedToDisk(newpage);
+
+       if (PageDirty(page)) {
+               clear_page_dirty_for_io(page);
+               set_page_dirty(newpage);
+       }
+
+       ClearPageSwapCache(page);
+       ClearPageActive(page);
+       ClearPagePrivate(page);
+       set_page_private(page, 0);
+       page->mapping = NULL;
+
+       /*
+        * If any waiters have accumulated on the new page then
+        * wake them up.
+        */
+       if (PageWriteback(newpage))
+               end_page_writeback(newpage);
+}
+EXPORT_SYMBOL(migrate_page_copy);
+
+/*
+ * Common logic to directly migrate a single page suitable for
+ * pages that do not use PagePrivate.
+ *
+ * Pages are locked upon entry and exit.
+ */
+int migrate_page(struct page *newpage, struct page *page)
+{
+       int rc;
+
+       BUG_ON(PageWriteback(page));    /* Writeback must be complete */
+
+       rc = migrate_page_remove_references(newpage, page, 2);
+
+       if (rc)
+               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);
+
+/*
+ * migrate_pages
+ *
+ * Two lists are passed to this function. The first list
+ * contains the pages isolated from the LRU to be migrated.
+ * The second list contains new pages that the pages isolated
+ * can be moved to. If the second list is NULL then all
+ * pages are swapped out.
+ *
+ * The function returns after 10 attempts or if no pages
+ * are movable anymore because to has become empty
+ * or no retryable pages exist anymore.
+ *
+ * Return: Number of pages not migrated when "to" ran empty.
+ */
+int migrate_pages(struct list_head *from, struct list_head *to,
+                 struct list_head *moved, struct list_head *failed)
+{
+       int retry;
+       int nr_failed = 0;
+       int pass = 0;
+       struct page *page;
+       struct page *page2;
+       int swapwrite = current->flags & PF_SWAPWRITE;
+       int rc;
+
+       if (!swapwrite)
+               current->flags |= PF_SWAPWRITE;
+
+redo:
+       retry = 0;
+
+       list_for_each_entry_safe(page, page2, from, lru) {
+               struct page *newpage = NULL;
+               struct address_space *mapping;
+
+               cond_resched();
+
+               rc = 0;
+               if (page_count(page) == 1)
+                       /* page was freed from under us. So we are done. */
+                       goto next;
+
+               if (to && list_empty(to))
+                       break;
+
+               /*
+                * Skip locked pages during the first two passes to give the
+                * functions holding the lock time to release the page. Later we
+                * use lock_page() to have a higher chance of acquiring the
+                * lock.
+                */
+               rc = -EAGAIN;
+               if (pass > 2)
+                       lock_page(page);
+               else
+                       if (TestSetPageLocked(page))
+                               goto next;
+
+               /*
+                * Only wait on writeback if we have already done a pass where
+                * we we may have triggered writeouts for lots of pages.
+                */
+               if (pass > 0) {
+                       wait_on_page_writeback(page);
+               } else {
+                       if (PageWriteback(page))
+                               goto unlock_page;
+               }
+
+               /*
+                * Anonymous pages must have swap cache references otherwise
+                * the information contained in the page maps cannot be
+                * preserved.
+                */
+               if (PageAnon(page) && !PageSwapCache(page)) {
+                       if (!add_to_swap(page, GFP_KERNEL)) {
+                               rc = -ENOMEM;
+                               goto unlock_page;
+                       }
+               }
+
+               if (!to) {
+                       rc = swap_page(page);
+                       goto next;
+               }
+
+               newpage = lru_to_page(to);
+               lock_page(newpage);
+
+               /*
+                * Pages are properly locked and writeback is complete.
+                * Try to migrate the page.
+                */
+               mapping = page_mapping(page);
+               if (!mapping)
+                       goto unlock_both;
+
+               if (mapping->a_ops->migratepage) {
+                       /*
+                        * Most pages have a mapping and most filesystems
+                        * should provide a migration function. Anonymous
+                        * pages are part of swap space which also has its
+                        * own migration function. This is the most common
+                        * path for page migration.
+                        */
+                       rc = mapping->a_ops->migratepage(newpage, page);
+                       goto unlock_both;
+                }
+
+               /*
+                * Default handling if a filesystem does not provide
+                * a migration function. We can only migrate clean
+                * pages so try to write out any dirty pages first.
+                */
+               if (PageDirty(page)) {
+                       switch (pageout(page, mapping)) {
+                       case PAGE_KEEP:
+                       case PAGE_ACTIVATE:
+                               goto unlock_both;
+
+                       case PAGE_SUCCESS:
+                               unlock_page(newpage);
+                               goto next;
+
+                       case PAGE_CLEAN:
+                               ; /* try to migrate the page below */
+                       }
+                }
+
+               /*
+                * Buffers are managed in a filesystem specific way.
+                * We must have no buffers or drop them.
+                */
+               if (!page_has_buffers(page) ||
+                   try_to_release_page(page, GFP_KERNEL)) {
+                       rc = migrate_page(newpage, page);
+                       goto unlock_both;
+               }
+
+               /*
+                * On early passes with mapped pages simply
+                * retry. There may be a lock held for some
+                * buffers that may go away. Later
+                * swap them out.
+                */
+               if (pass > 4) {
+                       /*
+                        * Persistently unable to drop buffers..... As a
+                        * measure of last resort we fall back to
+                        * swap_page().
+                        */
+                       unlock_page(newpage);
+                       newpage = NULL;
+                       rc = swap_page(page);
+                       goto next;
+               }
+
+unlock_both:
+               unlock_page(newpage);
+
+unlock_page:
+               unlock_page(page);
+
+next:
+               if (rc == -EAGAIN) {
+                       retry++;
+               } else if (rc) {
+                       /* Permanent failure */
+                       list_move(&page->lru, failed);
+                       nr_failed++;
+               } else {
+                       if (newpage) {
+                               /* Successful migration. Return page to LRU */
+                               move_to_lru(newpage);
+                       }
+                       list_move(&page->lru, moved);
+               }
+       }
+       if (retry && pass++ < 10)
+               goto redo;
+
+       if (!swapwrite)
+               current->flags &= ~PF_SWAPWRITE;
+
+       return nr_failed + retry;
+}
+
+/*
+ * Migration function for pages with buffers. This function can only be used
+ * if the underlying filesystem guarantees that no other references to "page"
+ * exist.
+ */
+int buffer_migrate_page(struct page *newpage, struct page *page)
+{
+       struct address_space *mapping = page->mapping;
+       struct buffer_head *bh, *head;
+       int rc;
+
+       if (!mapping)
+               return -EAGAIN;
+
+       if (!page_has_buffers(page))
+               return migrate_page(newpage, page);
+
+       head = page_buffers(page);
+
+       rc = migrate_page_remove_references(newpage, page, 3);
+
+       if (rc)
+               return rc;
+
+       bh = head;
+       do {
+               get_bh(bh);
+               lock_buffer(bh);
+               bh = bh->b_this_page;
+
+       } while (bh != head);
+
+       ClearPagePrivate(page);
+       set_page_private(newpage, page_private(page));
+       set_page_private(page, 0);
+       put_page(page);
+       get_page(newpage);
+
+       bh = head;
+       do {
+               set_bh_page(bh, newpage, bh_offset(bh));
+               bh = bh->b_this_page;
+
+       } while (bh != head);
+
+       SetPagePrivate(newpage);
+
+       migrate_page_copy(newpage, page);
+
+       bh = head;
+       do {
+               unlock_buffer(bh);
+               put_bh(bh);
+               bh = bh->b_this_page;
+
+       } while (bh != head);
+
+       return 0;
+}
+EXPORT_SYMBOL(buffer_migrate_page);
+
+/*
+ * Migrate the list 'pagelist' of pages to a certain destination.
+ *
+ * Specify destination with either non-NULL vma or dest_node >= 0
+ * Return the number of pages not migrated or error code
+ */
+int migrate_pages_to(struct list_head *pagelist,
+                       struct vm_area_struct *vma, int dest)
+{
+       LIST_HEAD(newlist);
+       LIST_HEAD(moved);
+       LIST_HEAD(failed);
+       int err = 0;
+       unsigned long offset = 0;
+       int nr_pages;
+       struct page *page;
+       struct list_head *p;
+
+redo:
+       nr_pages = 0;
+       list_for_each(p, pagelist) {
+               if (vma) {
+                       /*
+                        * The address passed to alloc_page_vma is used to
+                        * generate the proper interleave behavior. We fake
+                        * the address here by an increasing offset in order
+                        * to get the proper distribution of pages.
+                        *
+                        * No decision has been made as to which page
+                        * a certain old page is moved to so we cannot
+                        * specify the correct address.
+                        */
+                       page = alloc_page_vma(GFP_HIGHUSER, vma,
+                                       offset + vma->vm_start);
+                       offset += PAGE_SIZE;
+               }
+               else
+                       page = alloc_pages_node(dest, GFP_HIGHUSER, 0);
+
+               if (!page) {
+                       err = -ENOMEM;
+                       goto out;
+               }
+               list_add_tail(&page->lru, &newlist);
+               nr_pages++;
+               if (nr_pages > MIGRATE_CHUNK_SIZE)
+                       break;
+       }
+       err = migrate_pages(pagelist, &newlist, &moved, &failed);
+
+       putback_lru_pages(&moved);      /* Call release pages instead ?? */
+
+       if (err >= 0 && list_empty(&newlist) && !list_empty(pagelist))
+               goto redo;
+out:
+       /* Return leftover allocated pages */
+       while (!list_empty(&newlist)) {
+               page = list_entry(newlist.next, struct page, lru);
+               list_del(&page->lru);
+               __free_page(page);
+       }
+       list_splice(&failed, pagelist);
+       if (err < 0)
+               return err;
+
+       /* Calculate number of leftover pages */
+       nr_pages = 0;
+       list_for_each(p, pagelist)
+               nr_pages++;
+       return nr_pages;
+}
index db8a3d3..d7af296 100644 (file)
@@ -15,6 +15,7 @@
 #include <linux/buffer_head.h>
 #include <linux/backing-dev.h>
 #include <linux/pagevec.h>
+#include <linux/migrate.h>
 
 #include <asm/pgtable.h>
 
index 548e023..fd572bb 100644 (file)
 
 #include "internal.h"
 
-/* possible outcome of pageout() */
-typedef enum {
-       /* failed to write page out, page is locked */
-       PAGE_KEEP,
-       /* move page to the active list, page is locked */
-       PAGE_ACTIVATE,
-       /* page has been sent to the disk successfully, page is unlocked */
-       PAGE_SUCCESS,
-       /* page is clean and locked */
-       PAGE_CLEAN,
-} pageout_t;
-
 struct scan_control {
        /* Incremented by the number of inactive pages that were scanned */
        unsigned long nr_scanned;
@@ -304,7 +292,7 @@ static void handle_write_error(struct address_space *mapping,
  * pageout is called by shrink_page_list() for each dirty page.
  * Calls ->writepage().
  */
-static pageout_t pageout(struct page *page, struct address_space *mapping)
+pageout_t pageout(struct page *page, struct address_space *mapping)
 {
        /*
         * If the page is dirty, only perform writeback if that write
@@ -372,7 +360,7 @@ static pageout_t pageout(struct page *page, struct address_space *mapping)
        return PAGE_CLEAN;
 }
 
-static int remove_mapping(struct address_space *mapping, struct page *page)
+int remove_mapping(struct address_space *mapping, struct page *page)
 {
        if (!mapping)
                return 0;               /* truncate got there first */
@@ -570,481 +558,6 @@ keep:
        return nr_reclaimed;
 }
 
-#ifdef CONFIG_MIGRATION
-static inline void move_to_lru(struct page *page)
-{
-       list_del(&page->lru);
-       if (PageActive(page)) {
-               /*
-                * lru_cache_add_active checks that
-                * the PG_active bit is off.
-                */
-               ClearPageActive(page);
-               lru_cache_add_active(page);
-       } else {
-               lru_cache_add(page);
-       }
-       put_page(page);
-}
-
-/*
- * Add isolated pages on the list back to the LRU.
- *
- * returns the number of pages put back.
- */
-unsigned long putback_lru_pages(struct list_head *l)
-{
-       struct page *page;
-       struct page *page2;
-       unsigned long count = 0;
-
-       list_for_each_entry_safe(page, page2, l, lru) {
-               move_to_lru(page);
-               count++;
-       }
-       return count;
-}
-
-/*
- * Non migratable page
- */
-int fail_migrate_page(struct page *newpage, struct page *page)
-{
-       return -EIO;
-}
-EXPORT_SYMBOL(fail_migrate_page);
-
-/*
- * swapout a single page
- * page is locked upon entry, unlocked on exit
- */
-static int swap_page(struct page *page)
-{
-       struct address_space *mapping = page_mapping(page);
-
-       if (page_mapped(page) && mapping)
-               if (try_to_unmap(page, 1) != SWAP_SUCCESS)
-                       goto unlock_retry;
-
-       if (PageDirty(page)) {
-               /* Page is dirty, try to write it out here */
-               switch(pageout(page, mapping)) {
-               case PAGE_KEEP:
-               case PAGE_ACTIVATE:
-                       goto unlock_retry;
-
-               case PAGE_SUCCESS:
-                       goto retry;
-
-               case PAGE_CLEAN:
-                       ; /* try to free the page below */
-               }
-       }
-
-       if (PagePrivate(page)) {
-               if (!try_to_release_page(page, GFP_KERNEL) ||
-                   (!mapping && page_count(page) == 1))
-                       goto unlock_retry;
-       }
-
-       if (remove_mapping(mapping, page)) {
-               /* Success */
-               unlock_page(page);
-               return 0;
-       }
-
-unlock_retry:
-       unlock_page(page);
-
-retry:
-       return -EAGAIN;
-}
-EXPORT_SYMBOL(swap_page);
-
-/*
- * Page migration was first developed in the context of the memory hotplug
- * project. The main authors of the migration code are:
- *
- * IWAMOTO Toshihiro <iwamoto@valinux.co.jp>
- * Hirokazu Takahashi <taka@valinux.co.jp>
- * Dave Hansen <haveblue@us.ibm.com>
- * Christoph Lameter <clameter@sgi.com>
- */
-
-/*
- * Remove references for a page and establish the new page with the correct
- * basic settings to be able to stop accesses to the page.
- */
-int migrate_page_remove_references(struct page *newpage,
-                               struct page *page, int nr_refs)
-{
-       struct address_space *mapping = page_mapping(page);
-       struct page **radix_pointer;
-
-       /*
-        * Avoid doing any of the following work if the page count
-        * indicates that the page is in use or truncate has removed
-        * the page.
-        */
-       if (!mapping || page_mapcount(page) + nr_refs != page_count(page))
-               return -EAGAIN;
-
-       /*
-        * 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.
-        */
-       if (try_to_unmap(page, 1) == SWAP_FAIL)
-               /* A vma has VM_LOCKED set -> Permanent failure */
-               return -EPERM;
-
-       /*
-        * Give up if we were unable to remove all mappings.
-        */
-       if (page_mapcount(page))
-               return -EAGAIN;
-
-       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) != nr_refs ||
-                       *radix_pointer != page) {
-               write_unlock_irq(&mapping->tree_lock);
-               return -EAGAIN;
-       }
-
-       /*
-        * Now we know that no one else is looking at the page.
-        *
-        * Certain minimal information about a page must be available
-        * in order for other subsystems to properly handle the page if they
-        * find it through the radix tree update before we are finished
-        * copying the page.
-        */
-       get_page(newpage);
-       newpage->index = page->index;
-       newpage->mapping = page->mapping;
-       if (PageSwapCache(page)) {
-               SetPageSwapCache(newpage);
-               set_page_private(newpage, page_private(page));
-       }
-
-       *radix_pointer = newpage;
-       __put_page(page);
-       write_unlock_irq(&mapping->tree_lock);
-
-       return 0;
-}
-EXPORT_SYMBOL(migrate_page_remove_references);
-
-/*
- * Copy the page to its new location
- */
-void migrate_page_copy(struct page *newpage, struct page *page)
-{
-       copy_highpage(newpage, page);
-
-       if (PageError(page))
-               SetPageError(newpage);
-       if (PageReferenced(page))
-               SetPageReferenced(newpage);
-       if (PageUptodate(page))
-               SetPageUptodate(newpage);
-       if (PageActive(page))
-               SetPageActive(newpage);
-       if (PageChecked(page))
-               SetPageChecked(newpage);
-       if (PageMappedToDisk(page))
-               SetPageMappedToDisk(newpage);
-
-       if (PageDirty(page)) {
-               clear_page_dirty_for_io(page);
-               set_page_dirty(newpage);
-       }
-
-       ClearPageSwapCache(page);
-       ClearPageActive(page);
-       ClearPagePrivate(page);
-       set_page_private(page, 0);
-       page->mapping = NULL;
-
-       /*
-        * If any waiters have accumulated on the new page then
-        * wake them up.
-        */
-       if (PageWriteback(newpage))
-               end_page_writeback(newpage);
-}
-EXPORT_SYMBOL(migrate_page_copy);
-
-/*
- * Common logic to directly migrate a single page suitable for
- * pages that do not use PagePrivate.
- *
- * Pages are locked upon entry and exit.
- */
-int migrate_page(struct page *newpage, struct page *page)
-{
-       int rc;
-
-       BUG_ON(PageWriteback(page));    /* Writeback must be complete */
-
-       rc = migrate_page_remove_references(newpage, page, 2);
-
-       if (rc)
-               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);
-
-/*
- * migrate_pages
- *
- * Two lists are passed to this function. The first list
- * contains the pages isolated from the LRU to be migrated.
- * The second list contains new pages that the pages isolated
- * can be moved to. If the second list is NULL then all
- * pages are swapped out.
- *
- * The function returns after 10 attempts or if no pages
- * are movable anymore because to has become empty
- * or no retryable pages exist anymore.
- *
- * Return: Number of pages not migrated when "to" ran empty.
- */
-unsigned long migrate_pages(struct list_head *from, struct list_head *to,
-                 struct list_head *moved, struct list_head *failed)
-{
-       unsigned long retry;
-       unsigned long nr_failed = 0;
-       int pass = 0;
-       struct page *page;
-       struct page *page2;
-       int swapwrite = current->flags & PF_SWAPWRITE;
-       int rc;
-
-       if (!swapwrite)
-               current->flags |= PF_SWAPWRITE;
-
-redo:
-       retry = 0;
-
-       list_for_each_entry_safe(page, page2, from, lru) {
-               struct page *newpage = NULL;
-               struct address_space *mapping;
-
-               cond_resched();
-
-               rc = 0;
-               if (page_count(page) == 1)
-                       /* page was freed from under us. So we are done. */
-                       goto next;
-
-               if (to && list_empty(to))
-                       break;
-
-               /*
-                * Skip locked pages during the first two passes to give the
-                * functions holding the lock time to release the page. Later we
-                * use lock_page() to have a higher chance of acquiring the
-                * lock.
-                */
-               rc = -EAGAIN;
-               if (pass > 2)
-                       lock_page(page);
-               else
-                       if (TestSetPageLocked(page))
-                               goto next;
-
-               /*
-                * Only wait on writeback if we have already done a pass where
-                * we we may have triggered writeouts for lots of pages.
-                */
-               if (pass > 0) {
-                       wait_on_page_writeback(page);
-               } else {
-                       if (PageWriteback(page))
-                               goto unlock_page;
-               }
-
-               /*
-                * Anonymous pages must have swap cache references otherwise
-                * the information contained in the page maps cannot be
-                * preserved.
-                */
-               if (PageAnon(page) && !PageSwapCache(page)) {
-                       if (!add_to_swap(page, GFP_KERNEL)) {
-                               rc = -ENOMEM;
-                               goto unlock_page;
-                       }
-               }
-
-               if (!to) {
-                       rc = swap_page(page);
-                       goto next;
-               }
-
-               newpage = lru_to_page(to);
-               lock_page(newpage);
-
-               /*
-                * Pages are properly locked and writeback is complete.
-                * Try to migrate the page.
-                */
-               mapping = page_mapping(page);
-               if (!mapping)
-                       goto unlock_both;
-
-               if (mapping->a_ops->migratepage) {
-                       /*
-                        * Most pages have a mapping and most filesystems
-                        * should provide a migration function. Anonymous
-                        * pages are part of swap space which also has its
-                        * own migration function. This is the most common
-                        * path for page migration.
-                        */
-                       rc = mapping->a_ops->migratepage(newpage, page);
-                       goto unlock_both;
-                }
-
-               /*
-                * Default handling if a filesystem does not provide
-                * a migration function. We can only migrate clean
-                * pages so try to write out any dirty pages first.
-                */
-               if (PageDirty(page)) {
-                       switch (pageout(page, mapping)) {
-                       case PAGE_KEEP:
-                       case PAGE_ACTIVATE:
-                               goto unlock_both;
-
-                       case PAGE_SUCCESS:
-                               unlock_page(newpage);
-                               goto next;
-
-                       case PAGE_CLEAN:
-                               ; /* try to migrate the page below */
-                       }
-                }
-
-               /*
-                * Buffers are managed in a filesystem specific way.
-                * We must have no buffers or drop them.
-                */
-               if (!page_has_buffers(page) ||
-                   try_to_release_page(page, GFP_KERNEL)) {
-                       rc = migrate_page(newpage, page);
-                       goto unlock_both;
-               }
-
-               /*
-                * On early passes with mapped pages simply
-                * retry. There may be a lock held for some
-                * buffers that may go away. Later
-                * swap them out.
-                */
-               if (pass > 4) {
-                       /*
-                        * Persistently unable to drop buffers..... As a
-                        * measure of last resort we fall back to
-                        * swap_page().
-                        */
-                       unlock_page(newpage);
-                       newpage = NULL;
-                       rc = swap_page(page);
-                       goto next;
-               }
-
-unlock_both:
-               unlock_page(newpage);
-
-unlock_page:
-               unlock_page(page);
-
-next:
-               if (rc == -EAGAIN) {
-                       retry++;
-               } else if (rc) {
-                       /* Permanent failure */
-                       list_move(&page->lru, failed);
-                       nr_failed++;
-               } else {
-                       if (newpage) {
-                               /* Successful migration. Return page to LRU */
-                               move_to_lru(newpage);
-                       }
-                       list_move(&page->lru, moved);
-               }
-       }
-       if (retry && pass++ < 10)
-               goto redo;
-
-       if (!swapwrite)
-               current->flags &= ~PF_SWAPWRITE;
-
-       return nr_failed + retry;
-}
-
-/*
- * Isolate one page from the LRU lists and put it on the
- * indicated list with elevated refcount.
- *
- * Result:
- *  0 = page not on LRU list
- *  1 = page removed from LRU list and added to the specified list.
- */
-int isolate_lru_page(struct page *page)
-{
-       int ret = 0;
-
-       if (PageLRU(page)) {
-               struct zone *zone = page_zone(page);
-               spin_lock_irq(&zone->lru_lock);
-               if (PageLRU(page)) {
-                       ret = 1;
-                       get_page(page);
-                       ClearPageLRU(page);
-                       if (PageActive(page))
-                               del_page_from_active_list(zone, page);
-                       else
-                               del_page_from_inactive_list(zone, page);
-               }
-               spin_unlock_irq(&zone->lru_lock);
-       }
-
-       return ret;
-}
-#endif
-
 /*
  * zone->lru_lock is heavily contended.  Some of the functions that
  * shrink the lists perform better by taking out a batch of pages