2 * Memory Migration functionality - linux/mm/migration.c
4 * Copyright (C) 2006 Silicon Graphics, Inc., Christoph Lameter
6 * Page migration was first developed in the context of the memory hotplug
7 * project. The main authors of the migration code are:
9 * IWAMOTO Toshihiro <iwamoto@valinux.co.jp>
10 * Hirokazu Takahashi <taka@valinux.co.jp>
11 * Dave Hansen <haveblue@us.ibm.com>
12 * Christoph Lameter <clameter@sgi.com>
15 #include <linux/migrate.h>
16 #include <linux/module.h>
17 #include <linux/swap.h>
18 #include <linux/swapops.h>
19 #include <linux/pagemap.h>
20 #include <linux/buffer_head.h>
21 #include <linux/mm_inline.h>
22 #include <linux/pagevec.h>
23 #include <linux/rmap.h>
24 #include <linux/topology.h>
25 #include <linux/cpu.h>
26 #include <linux/cpuset.h>
27 #include <linux/writeback.h>
28 #include <linux/mempolicy.h>
29 #include <linux/vmalloc.h>
30 #include <linux/security.h>
34 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
37 * Isolate one page from the LRU lists. If successful put it onto
38 * the indicated list with elevated page count.
41 * -EBUSY: page not on LRU list
42 * 0: page removed from LRU list and added to the specified list.
44 int isolate_lru_page(struct page *page, struct list_head *pagelist)
49 struct zone *zone = page_zone(page);
51 spin_lock_irq(&zone->lru_lock);
52 if (PageLRU(page) && get_page_unless_zero(page)) {
56 del_page_from_active_list(zone, page);
58 del_page_from_inactive_list(zone, page);
59 list_add_tail(&page->lru, pagelist);
61 spin_unlock_irq(&zone->lru_lock);
67 * migrate_prep() needs to be called before we start compiling a list of pages
68 * to be migrated using isolate_lru_page().
70 int migrate_prep(void)
73 * Clear the LRU lists so pages can be isolated.
74 * Note that pages may be moved off the LRU after we have
75 * drained them. Those pages will fail to migrate like other
76 * pages that may be busy.
83 static inline void move_to_lru(struct page *page)
85 if (PageActive(page)) {
87 * lru_cache_add_active checks that
88 * the PG_active bit is off.
90 ClearPageActive(page);
91 lru_cache_add_active(page);
99 * Add isolated pages on the list back to the LRU.
101 * returns the number of pages put back.
103 int putback_lru_pages(struct list_head *l)
109 list_for_each_entry_safe(page, page2, l, lru) {
110 list_del(&page->lru);
117 static inline int is_swap_pte(pte_t pte)
119 return !pte_none(pte) && !pte_present(pte) && !pte_file(pte);
123 * Restore a potential migration pte to a working pte entry
125 static void remove_migration_pte(struct vm_area_struct *vma,
126 struct page *old, struct page *new)
128 struct mm_struct *mm = vma->vm_mm;
135 unsigned long addr = page_address_in_vma(new, vma);
140 pgd = pgd_offset(mm, addr);
141 if (!pgd_present(*pgd))
144 pud = pud_offset(pgd, addr);
145 if (!pud_present(*pud))
148 pmd = pmd_offset(pud, addr);
149 if (!pmd_present(*pmd))
152 ptep = pte_offset_map(pmd, addr);
154 if (!is_swap_pte(*ptep)) {
159 ptl = pte_lockptr(mm, pmd);
162 if (!is_swap_pte(pte))
165 entry = pte_to_swp_entry(pte);
167 if (!is_migration_entry(entry) || migration_entry_to_page(entry) != old)
171 pte = pte_mkold(mk_pte(new, vma->vm_page_prot));
172 if (is_write_migration_entry(entry))
173 pte = pte_mkwrite(pte);
174 set_pte_at(mm, addr, ptep, pte);
177 page_add_anon_rmap(new, vma, addr);
179 page_add_file_rmap(new);
181 /* No need to invalidate - it was non-present before */
182 update_mmu_cache(vma, addr, pte);
183 lazy_mmu_prot_update(pte);
186 pte_unmap_unlock(ptep, ptl);
190 * Note that remove_file_migration_ptes will only work on regular mappings,
191 * Nonlinear mappings do not use migration entries.
193 static void remove_file_migration_ptes(struct page *old, struct page *new)
195 struct vm_area_struct *vma;
196 struct address_space *mapping = page_mapping(new);
197 struct prio_tree_iter iter;
198 pgoff_t pgoff = new->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
203 spin_lock(&mapping->i_mmap_lock);
205 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff)
206 remove_migration_pte(vma, old, new);
208 spin_unlock(&mapping->i_mmap_lock);
212 * Must hold mmap_sem lock on at least one of the vmas containing
213 * the page so that the anon_vma cannot vanish.
215 static void remove_anon_migration_ptes(struct page *old, struct page *new)
217 struct anon_vma *anon_vma;
218 struct vm_area_struct *vma;
219 unsigned long mapping;
221 mapping = (unsigned long)new->mapping;
223 if (!mapping || (mapping & PAGE_MAPPING_ANON) == 0)
227 * We hold the mmap_sem lock. So no need to call page_lock_anon_vma.
229 anon_vma = (struct anon_vma *) (mapping - PAGE_MAPPING_ANON);
230 spin_lock(&anon_vma->lock);
232 list_for_each_entry(vma, &anon_vma->head, anon_vma_node)
233 remove_migration_pte(vma, old, new);
235 spin_unlock(&anon_vma->lock);
239 * Get rid of all migration entries and replace them by
240 * references to the indicated page.
242 static void remove_migration_ptes(struct page *old, struct page *new)
245 remove_anon_migration_ptes(old, new);
247 remove_file_migration_ptes(old, new);
251 * Something used the pte of a page under migration. We need to
252 * get to the page and wait until migration is finished.
253 * When we return from this function the fault will be retried.
255 * This function is called from do_swap_page().
257 void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
258 unsigned long address)
265 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
267 if (!is_swap_pte(pte))
270 entry = pte_to_swp_entry(pte);
271 if (!is_migration_entry(entry))
274 page = migration_entry_to_page(entry);
277 pte_unmap_unlock(ptep, ptl);
278 wait_on_page_locked(page);
282 pte_unmap_unlock(ptep, ptl);
286 * Replace the page in the mapping.
288 * The number of remaining references must be:
289 * 1 for anonymous pages without a mapping
290 * 2 for pages with a mapping
291 * 3 for pages with a mapping and PagePrivate set.
293 static int migrate_page_move_mapping(struct address_space *mapping,
294 struct page *newpage, struct page *page)
299 /* Anonymous page without mapping */
300 if (page_count(page) != 1)
305 write_lock_irq(&mapping->tree_lock);
307 pslot = radix_tree_lookup_slot(&mapping->page_tree,
310 if (page_count(page) != 2 + !!PagePrivate(page) ||
311 (struct page *)radix_tree_deref_slot(pslot) != page) {
312 write_unlock_irq(&mapping->tree_lock);
317 * Now we know that no one else is looking at the page.
319 get_page(newpage); /* add cache reference */
321 if (PageSwapCache(page)) {
322 SetPageSwapCache(newpage);
323 set_page_private(newpage, page_private(page));
327 radix_tree_replace_slot(pslot, newpage);
330 * Drop cache reference from old page.
331 * We know this isn't the last reference.
336 * If moved to a different zone then also account
337 * the page for that zone. Other VM counters will be
338 * taken care of when we establish references to the
339 * new page and drop references to the old page.
341 * Note that anonymous pages are accounted for
342 * via NR_FILE_PAGES and NR_ANON_PAGES if they
343 * are mapped to swap space.
345 __dec_zone_page_state(page, NR_FILE_PAGES);
346 __inc_zone_page_state(newpage, NR_FILE_PAGES);
348 write_unlock_irq(&mapping->tree_lock);
354 * Copy the page to its new location
356 static void migrate_page_copy(struct page *newpage, struct page *page)
358 copy_highpage(newpage, page);
361 SetPageError(newpage);
362 if (PageReferenced(page))
363 SetPageReferenced(newpage);
364 if (PageUptodate(page))
365 SetPageUptodate(newpage);
366 if (PageActive(page))
367 SetPageActive(newpage);
368 if (PageChecked(page))
369 SetPageChecked(newpage);
370 if (PageMappedToDisk(page))
371 SetPageMappedToDisk(newpage);
373 if (PageDirty(page)) {
374 clear_page_dirty_for_io(page);
375 set_page_dirty(newpage);
379 ClearPageSwapCache(page);
381 ClearPageActive(page);
382 ClearPagePrivate(page);
383 set_page_private(page, 0);
384 page->mapping = NULL;
387 * If any waiters have accumulated on the new page then
390 if (PageWriteback(newpage))
391 end_page_writeback(newpage);
394 /************************************************************
395 * Migration functions
396 ***********************************************************/
398 /* Always fail migration. Used for mappings that are not movable */
399 int fail_migrate_page(struct address_space *mapping,
400 struct page *newpage, struct page *page)
404 EXPORT_SYMBOL(fail_migrate_page);
407 * Common logic to directly migrate a single page suitable for
408 * pages that do not use PagePrivate.
410 * Pages are locked upon entry and exit.
412 int migrate_page(struct address_space *mapping,
413 struct page *newpage, struct page *page)
417 BUG_ON(PageWriteback(page)); /* Writeback must be complete */
419 rc = migrate_page_move_mapping(mapping, newpage, page);
424 migrate_page_copy(newpage, page);
427 EXPORT_SYMBOL(migrate_page);
431 * Migration function for pages with buffers. This function can only be used
432 * if the underlying filesystem guarantees that no other references to "page"
435 int buffer_migrate_page(struct address_space *mapping,
436 struct page *newpage, struct page *page)
438 struct buffer_head *bh, *head;
441 if (!page_has_buffers(page))
442 return migrate_page(mapping, newpage, page);
444 head = page_buffers(page);
446 rc = migrate_page_move_mapping(mapping, newpage, page);
455 bh = bh->b_this_page;
457 } while (bh != head);
459 ClearPagePrivate(page);
460 set_page_private(newpage, page_private(page));
461 set_page_private(page, 0);
467 set_bh_page(bh, newpage, bh_offset(bh));
468 bh = bh->b_this_page;
470 } while (bh != head);
472 SetPagePrivate(newpage);
474 migrate_page_copy(newpage, page);
480 bh = bh->b_this_page;
482 } while (bh != head);
486 EXPORT_SYMBOL(buffer_migrate_page);
490 * Writeback a page to clean the dirty state
492 static int writeout(struct address_space *mapping, struct page *page)
494 struct writeback_control wbc = {
495 .sync_mode = WB_SYNC_NONE,
498 .range_end = LLONG_MAX,
504 if (!mapping->a_ops->writepage)
505 /* No write method for the address space */
508 if (!clear_page_dirty_for_io(page))
509 /* Someone else already triggered a write */
513 * A dirty page may imply that the underlying filesystem has
514 * the page on some queue. So the page must be clean for
515 * migration. Writeout may mean we loose the lock and the
516 * page state is no longer what we checked for earlier.
517 * At this point we know that the migration attempt cannot
520 remove_migration_ptes(page, page);
522 rc = mapping->a_ops->writepage(page, &wbc);
524 /* I/O Error writing */
527 if (rc != AOP_WRITEPAGE_ACTIVATE)
528 /* unlocked. Relock */
535 * Default handling if a filesystem does not provide a migration function.
537 static int fallback_migrate_page(struct address_space *mapping,
538 struct page *newpage, struct page *page)
541 return writeout(mapping, page);
544 * Buffers may be managed in a filesystem specific way.
545 * We must have no buffers or drop them.
547 if (PagePrivate(page) &&
548 !try_to_release_page(page, GFP_KERNEL))
551 return migrate_page(mapping, newpage, page);
555 * Move a page to a newly allocated page
556 * The page is locked and all ptes have been successfully removed.
558 * The new page will have replaced the old page if this function
561 static int move_to_new_page(struct page *newpage, struct page *page)
563 struct address_space *mapping;
567 * Block others from accessing the page when we get around to
568 * establishing additional references. We are the only one
569 * holding a reference to the new page at this point.
571 if (TestSetPageLocked(newpage))
574 /* Prepare mapping for the new page.*/
575 newpage->index = page->index;
576 newpage->mapping = page->mapping;
578 mapping = page_mapping(page);
580 rc = migrate_page(mapping, newpage, page);
581 else if (mapping->a_ops->migratepage)
583 * Most pages have a mapping and most filesystems
584 * should provide a migration function. Anonymous
585 * pages are part of swap space which also has its
586 * own migration function. This is the most common
587 * path for page migration.
589 rc = mapping->a_ops->migratepage(mapping,
592 rc = fallback_migrate_page(mapping, newpage, page);
595 remove_migration_ptes(page, newpage);
597 newpage->mapping = NULL;
599 unlock_page(newpage);
605 * Obtain the lock on page, remove all ptes and migrate the page
606 * to the newly allocated page in newpage.
608 static int unmap_and_move(new_page_t get_new_page, unsigned long private,
609 struct page *page, int force)
613 struct page *newpage = get_new_page(page, private, &result);
619 if (page_count(page) == 1)
620 /* page was freed from under us. So we are done. */
624 if (TestSetPageLocked(page)) {
630 if (PageWriteback(page)) {
633 wait_on_page_writeback(page);
636 * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
637 * we cannot notice that anon_vma is freed while we migrates a page.
638 * This rcu_read_lock() delays freeing anon_vma pointer until the end
639 * of migration. File cache pages are no problem because of page_lock()
640 * File Caches may use write_page() or lock_page() in migration, then,
641 * just care Anon page here.
643 if (PageAnon(page)) {
648 * This is a corner case handling.
649 * When a new swap-cache is read into, it is linked to LRU
650 * and treated as swapcache but has no rmap yet.
651 * Calling try_to_unmap() against a page->mapping==NULL page is
652 * BUG. So handle it here.
656 /* Establish migration ptes or remove ptes */
657 try_to_unmap(page, 1);
659 if (!page_mapped(page))
660 rc = move_to_new_page(newpage, page);
663 remove_migration_ptes(page, page);
674 * A page that has been migrated has all references
675 * removed and will be freed. A page that has not been
676 * migrated will have kepts its references and be
679 list_del(&page->lru);
685 * Move the new page to the LRU. If migration was not successful
686 * then this will free the page.
688 move_to_lru(newpage);
693 *result = page_to_nid(newpage);
701 * The function takes one list of pages to migrate and a function
702 * that determines from the page to be migrated and the private data
703 * the target of the move and allocates the page.
705 * The function returns after 10 attempts or if no pages
706 * are movable anymore because to has become empty
707 * or no retryable pages exist anymore. All pages will be
708 * retruned to the LRU or freed.
710 * Return: Number of pages not migrated or error code.
712 int migrate_pages(struct list_head *from,
713 new_page_t get_new_page, unsigned long private)
720 int swapwrite = current->flags & PF_SWAPWRITE;
724 current->flags |= PF_SWAPWRITE;
726 for(pass = 0; pass < 10 && retry; pass++) {
729 list_for_each_entry_safe(page, page2, from, lru) {
732 rc = unmap_and_move(get_new_page, private,
744 /* Permanent failure */
753 current->flags &= ~PF_SWAPWRITE;
755 putback_lru_pages(from);
760 return nr_failed + retry;
765 * Move a list of individual pages
767 struct page_to_node {
774 static struct page *new_page_node(struct page *p, unsigned long private,
777 struct page_to_node *pm = (struct page_to_node *)private;
779 while (pm->node != MAX_NUMNODES && pm->page != p)
782 if (pm->node == MAX_NUMNODES)
785 *result = &pm->status;
787 return alloc_pages_node(pm->node,
788 GFP_HIGHUSER_MOVABLE | GFP_THISNODE, 0);
792 * Move a set of pages as indicated in the pm array. The addr
793 * field must be set to the virtual address of the page to be moved
794 * and the node number must contain a valid target node.
796 static int do_move_pages(struct mm_struct *mm, struct page_to_node *pm,
800 struct page_to_node *pp;
803 down_read(&mm->mmap_sem);
806 * Build a list of pages to migrate
809 for (pp = pm; pp->node != MAX_NUMNODES; pp++) {
810 struct vm_area_struct *vma;
814 * A valid page pointer that will not match any of the
815 * pages that will be moved.
817 pp->page = ZERO_PAGE(0);
820 vma = find_vma(mm, pp->addr);
821 if (!vma || !vma_migratable(vma))
824 page = follow_page(vma, pp->addr, FOLL_GET);
829 if (PageReserved(page)) /* Check for zero page */
833 err = page_to_nid(page);
837 * Node already in the right place
842 if (page_mapcount(page) > 1 &&
846 err = isolate_lru_page(page, &pagelist);
849 * Either remove the duplicate refcount from
850 * isolate_lru_page() or drop the page ref if it was
858 if (!list_empty(&pagelist))
859 err = migrate_pages(&pagelist, new_page_node,
864 up_read(&mm->mmap_sem);
869 * Determine the nodes of a list of pages. The addr in the pm array
870 * must have been set to the virtual address of which we want to determine
873 static int do_pages_stat(struct mm_struct *mm, struct page_to_node *pm)
875 down_read(&mm->mmap_sem);
877 for ( ; pm->node != MAX_NUMNODES; pm++) {
878 struct vm_area_struct *vma;
883 vma = find_vma(mm, pm->addr);
887 page = follow_page(vma, pm->addr, 0);
889 /* Use PageReserved to check for zero page */
890 if (!page || PageReserved(page))
893 err = page_to_nid(page);
898 up_read(&mm->mmap_sem);
903 * Move a list of pages in the address space of the currently executing
906 asmlinkage long sys_move_pages(pid_t pid, unsigned long nr_pages,
907 const void __user * __user *pages,
908 const int __user *nodes,
909 int __user *status, int flags)
913 struct task_struct *task;
914 nodemask_t task_nodes;
915 struct mm_struct *mm;
916 struct page_to_node *pm = NULL;
919 if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
922 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
925 /* Find the mm_struct */
926 read_lock(&tasklist_lock);
927 task = pid ? find_task_by_pid(pid) : current;
929 read_unlock(&tasklist_lock);
932 mm = get_task_mm(task);
933 read_unlock(&tasklist_lock);
939 * Check if this process has the right to modify the specified
940 * process. The right exists if the process has administrative
941 * capabilities, superuser privileges or the same
942 * userid as the target process.
944 if ((current->euid != task->suid) && (current->euid != task->uid) &&
945 (current->uid != task->suid) && (current->uid != task->uid) &&
946 !capable(CAP_SYS_NICE)) {
951 err = security_task_movememory(task);
956 task_nodes = cpuset_mems_allowed(task);
958 /* Limit nr_pages so that the multiplication may not overflow */
959 if (nr_pages >= ULONG_MAX / sizeof(struct page_to_node) - 1) {
964 pm = vmalloc((nr_pages + 1) * sizeof(struct page_to_node));
971 * Get parameters from user space and initialize the pm
972 * array. Return various errors if the user did something wrong.
974 for (i = 0; i < nr_pages; i++) {
975 const void __user *p;
978 if (get_user(p, pages + i))
981 pm[i].addr = (unsigned long)p;
985 if (get_user(node, nodes + i))
989 if (!node_online(node))
993 if (!node_isset(node, task_nodes))
998 pm[i].node = 0; /* anything to not match MAX_NUMNODES */
1001 pm[nr_pages].node = MAX_NUMNODES;
1004 err = do_move_pages(mm, pm, flags & MPOL_MF_MOVE_ALL);
1006 err = do_pages_stat(mm, pm);
1009 /* Return status information */
1010 for (i = 0; i < nr_pages; i++)
1011 if (put_user(pm[i].status, status + i))
1023 * Call migration functions in the vma_ops that may prepare
1024 * memory in a vm for migration. migration functions may perform
1025 * the migration for vmas that do not have an underlying page struct.
1027 int migrate_vmas(struct mm_struct *mm, const nodemask_t *to,
1028 const nodemask_t *from, unsigned long flags)
1030 struct vm_area_struct *vma;
1033 for(vma = mm->mmap; vma->vm_next && !err; vma = vma->vm_next) {
1034 if (vma->vm_ops && vma->vm_ops->migrate) {
1035 err = vma->vm_ops->migrate(vma, to, from, flags);