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/nsproxy.h>
23 #include <linux/pagevec.h>
24 #include <linux/rmap.h>
25 #include <linux/topology.h>
26 #include <linux/cpu.h>
27 #include <linux/cpuset.h>
28 #include <linux/writeback.h>
29 #include <linux/mempolicy.h>
30 #include <linux/vmalloc.h>
31 #include <linux/security.h>
35 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
38 * Isolate one page from the LRU lists. If successful put it onto
39 * the indicated list with elevated page count.
42 * -EBUSY: page not on LRU list
43 * 0: page removed from LRU list and added to the specified list.
45 int isolate_lru_page(struct page *page, struct list_head *pagelist)
50 struct zone *zone = page_zone(page);
52 spin_lock_irq(&zone->lru_lock);
53 if (PageLRU(page) && get_page_unless_zero(page)) {
57 del_page_from_active_list(zone, page);
59 del_page_from_inactive_list(zone, page);
60 list_add_tail(&page->lru, pagelist);
62 spin_unlock_irq(&zone->lru_lock);
68 * migrate_prep() needs to be called before we start compiling a list of pages
69 * to be migrated using isolate_lru_page().
71 int migrate_prep(void)
74 * Clear the LRU lists so pages can be isolated.
75 * Note that pages may be moved off the LRU after we have
76 * drained them. Those pages will fail to migrate like other
77 * pages that may be busy.
84 static inline void move_to_lru(struct page *page)
86 if (PageActive(page)) {
88 * lru_cache_add_active checks that
89 * the PG_active bit is off.
91 ClearPageActive(page);
92 lru_cache_add_active(page);
100 * Add isolated pages on the list back to the LRU.
102 * returns the number of pages put back.
104 int putback_lru_pages(struct list_head *l)
110 list_for_each_entry_safe(page, page2, l, lru) {
111 list_del(&page->lru);
119 * Restore a potential migration pte to a working pte entry
121 static void remove_migration_pte(struct vm_area_struct *vma,
122 struct page *old, struct page *new)
124 struct mm_struct *mm = vma->vm_mm;
131 unsigned long addr = page_address_in_vma(new, vma);
136 pgd = pgd_offset(mm, addr);
137 if (!pgd_present(*pgd))
140 pud = pud_offset(pgd, addr);
141 if (!pud_present(*pud))
144 pmd = pmd_offset(pud, addr);
145 if (!pmd_present(*pmd))
148 ptep = pte_offset_map(pmd, addr);
150 if (!is_swap_pte(*ptep)) {
155 ptl = pte_lockptr(mm, pmd);
158 if (!is_swap_pte(pte))
161 entry = pte_to_swp_entry(pte);
163 if (!is_migration_entry(entry) || migration_entry_to_page(entry) != old)
167 pte = pte_mkold(mk_pte(new, vma->vm_page_prot));
168 if (is_write_migration_entry(entry))
169 pte = pte_mkwrite(pte);
170 flush_cache_page(vma, addr, pte_pfn(pte));
171 set_pte_at(mm, addr, ptep, pte);
174 page_add_anon_rmap(new, vma, addr);
176 page_add_file_rmap(new);
178 /* No need to invalidate - it was non-present before */
179 update_mmu_cache(vma, addr, pte);
182 pte_unmap_unlock(ptep, ptl);
186 * Note that remove_file_migration_ptes will only work on regular mappings,
187 * Nonlinear mappings do not use migration entries.
189 static void remove_file_migration_ptes(struct page *old, struct page *new)
191 struct vm_area_struct *vma;
192 struct address_space *mapping = page_mapping(new);
193 struct prio_tree_iter iter;
194 pgoff_t pgoff = new->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
199 spin_lock(&mapping->i_mmap_lock);
201 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff)
202 remove_migration_pte(vma, old, new);
204 spin_unlock(&mapping->i_mmap_lock);
208 * Must hold mmap_sem lock on at least one of the vmas containing
209 * the page so that the anon_vma cannot vanish.
211 static void remove_anon_migration_ptes(struct page *old, struct page *new)
213 struct anon_vma *anon_vma;
214 struct vm_area_struct *vma;
215 unsigned long mapping;
217 mapping = (unsigned long)new->mapping;
219 if (!mapping || (mapping & PAGE_MAPPING_ANON) == 0)
223 * We hold the mmap_sem lock. So no need to call page_lock_anon_vma.
225 anon_vma = (struct anon_vma *) (mapping - PAGE_MAPPING_ANON);
226 spin_lock(&anon_vma->lock);
228 list_for_each_entry(vma, &anon_vma->head, anon_vma_node)
229 remove_migration_pte(vma, old, new);
231 spin_unlock(&anon_vma->lock);
235 * Get rid of all migration entries and replace them by
236 * references to the indicated page.
238 static void remove_migration_ptes(struct page *old, struct page *new)
241 remove_anon_migration_ptes(old, new);
243 remove_file_migration_ptes(old, new);
247 * Something used the pte of a page under migration. We need to
248 * get to the page and wait until migration is finished.
249 * When we return from this function the fault will be retried.
251 * This function is called from do_swap_page().
253 void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
254 unsigned long address)
261 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
263 if (!is_swap_pte(pte))
266 entry = pte_to_swp_entry(pte);
267 if (!is_migration_entry(entry))
270 page = migration_entry_to_page(entry);
273 pte_unmap_unlock(ptep, ptl);
274 wait_on_page_locked(page);
278 pte_unmap_unlock(ptep, ptl);
282 * Replace the page in the mapping.
284 * The number of remaining references must be:
285 * 1 for anonymous pages without a mapping
286 * 2 for pages with a mapping
287 * 3 for pages with a mapping and PagePrivate set.
289 static int migrate_page_move_mapping(struct address_space *mapping,
290 struct page *newpage, struct page *page)
295 /* Anonymous page without mapping */
296 if (page_count(page) != 1)
301 write_lock_irq(&mapping->tree_lock);
303 pslot = radix_tree_lookup_slot(&mapping->page_tree,
306 if (page_count(page) != 2 + !!PagePrivate(page) ||
307 (struct page *)radix_tree_deref_slot(pslot) != page) {
308 write_unlock_irq(&mapping->tree_lock);
313 * Now we know that no one else is looking at the page.
315 get_page(newpage); /* add cache reference */
317 if (PageSwapCache(page)) {
318 SetPageSwapCache(newpage);
319 set_page_private(newpage, page_private(page));
323 radix_tree_replace_slot(pslot, newpage);
326 * Drop cache reference from old page.
327 * We know this isn't the last reference.
332 * If moved to a different zone then also account
333 * the page for that zone. Other VM counters will be
334 * taken care of when we establish references to the
335 * new page and drop references to the old page.
337 * Note that anonymous pages are accounted for
338 * via NR_FILE_PAGES and NR_ANON_PAGES if they
339 * are mapped to swap space.
341 __dec_zone_page_state(page, NR_FILE_PAGES);
342 __inc_zone_page_state(newpage, NR_FILE_PAGES);
344 write_unlock_irq(&mapping->tree_lock);
350 * Copy the page to its new location
352 static void migrate_page_copy(struct page *newpage, struct page *page)
354 copy_highpage(newpage, page);
357 SetPageError(newpage);
358 if (PageReferenced(page))
359 SetPageReferenced(newpage);
360 if (PageUptodate(page))
361 SetPageUptodate(newpage);
362 if (PageActive(page))
363 SetPageActive(newpage);
364 if (PageChecked(page))
365 SetPageChecked(newpage);
366 if (PageMappedToDisk(page))
367 SetPageMappedToDisk(newpage);
369 if (PageDirty(page)) {
370 clear_page_dirty_for_io(page);
371 set_page_dirty(newpage);
375 ClearPageSwapCache(page);
377 ClearPageActive(page);
378 ClearPagePrivate(page);
379 set_page_private(page, 0);
380 page->mapping = NULL;
383 * If any waiters have accumulated on the new page then
386 if (PageWriteback(newpage))
387 end_page_writeback(newpage);
390 /************************************************************
391 * Migration functions
392 ***********************************************************/
394 /* Always fail migration. Used for mappings that are not movable */
395 int fail_migrate_page(struct address_space *mapping,
396 struct page *newpage, struct page *page)
400 EXPORT_SYMBOL(fail_migrate_page);
403 * Common logic to directly migrate a single page suitable for
404 * pages that do not use PagePrivate.
406 * Pages are locked upon entry and exit.
408 int migrate_page(struct address_space *mapping,
409 struct page *newpage, struct page *page)
413 BUG_ON(PageWriteback(page)); /* Writeback must be complete */
415 rc = migrate_page_move_mapping(mapping, newpage, page);
420 migrate_page_copy(newpage, page);
423 EXPORT_SYMBOL(migrate_page);
427 * Migration function for pages with buffers. This function can only be used
428 * if the underlying filesystem guarantees that no other references to "page"
431 int buffer_migrate_page(struct address_space *mapping,
432 struct page *newpage, struct page *page)
434 struct buffer_head *bh, *head;
437 if (!page_has_buffers(page))
438 return migrate_page(mapping, newpage, page);
440 head = page_buffers(page);
442 rc = migrate_page_move_mapping(mapping, newpage, page);
451 bh = bh->b_this_page;
453 } while (bh != head);
455 ClearPagePrivate(page);
456 set_page_private(newpage, page_private(page));
457 set_page_private(page, 0);
463 set_bh_page(bh, newpage, bh_offset(bh));
464 bh = bh->b_this_page;
466 } while (bh != head);
468 SetPagePrivate(newpage);
470 migrate_page_copy(newpage, page);
476 bh = bh->b_this_page;
478 } while (bh != head);
482 EXPORT_SYMBOL(buffer_migrate_page);
486 * Writeback a page to clean the dirty state
488 static int writeout(struct address_space *mapping, struct page *page)
490 struct writeback_control wbc = {
491 .sync_mode = WB_SYNC_NONE,
494 .range_end = LLONG_MAX,
500 if (!mapping->a_ops->writepage)
501 /* No write method for the address space */
504 if (!clear_page_dirty_for_io(page))
505 /* Someone else already triggered a write */
509 * A dirty page may imply that the underlying filesystem has
510 * the page on some queue. So the page must be clean for
511 * migration. Writeout may mean we loose the lock and the
512 * page state is no longer what we checked for earlier.
513 * At this point we know that the migration attempt cannot
516 remove_migration_ptes(page, page);
518 rc = mapping->a_ops->writepage(page, &wbc);
520 /* I/O Error writing */
523 if (rc != AOP_WRITEPAGE_ACTIVATE)
524 /* unlocked. Relock */
531 * Default handling if a filesystem does not provide a migration function.
533 static int fallback_migrate_page(struct address_space *mapping,
534 struct page *newpage, struct page *page)
537 return writeout(mapping, page);
540 * Buffers may be managed in a filesystem specific way.
541 * We must have no buffers or drop them.
543 if (PagePrivate(page) &&
544 !try_to_release_page(page, GFP_KERNEL))
547 return migrate_page(mapping, newpage, page);
551 * Move a page to a newly allocated page
552 * The page is locked and all ptes have been successfully removed.
554 * The new page will have replaced the old page if this function
557 static int move_to_new_page(struct page *newpage, struct page *page)
559 struct address_space *mapping;
563 * Block others from accessing the page when we get around to
564 * establishing additional references. We are the only one
565 * holding a reference to the new page at this point.
567 if (TestSetPageLocked(newpage))
570 /* Prepare mapping for the new page.*/
571 newpage->index = page->index;
572 newpage->mapping = page->mapping;
574 mapping = page_mapping(page);
576 rc = migrate_page(mapping, newpage, page);
577 else if (mapping->a_ops->migratepage)
579 * Most pages have a mapping and most filesystems
580 * should provide a migration function. Anonymous
581 * pages are part of swap space which also has its
582 * own migration function. This is the most common
583 * path for page migration.
585 rc = mapping->a_ops->migratepage(mapping,
588 rc = fallback_migrate_page(mapping, newpage, page);
591 remove_migration_ptes(page, newpage);
593 newpage->mapping = NULL;
595 unlock_page(newpage);
601 * Obtain the lock on page, remove all ptes and migrate the page
602 * to the newly allocated page in newpage.
604 static int unmap_and_move(new_page_t get_new_page, unsigned long private,
605 struct page *page, int force)
609 struct page *newpage = get_new_page(page, private, &result);
615 if (page_count(page) == 1)
616 /* page was freed from under us. So we are done. */
620 if (TestSetPageLocked(page)) {
626 if (PageWriteback(page)) {
629 wait_on_page_writeback(page);
632 * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
633 * we cannot notice that anon_vma is freed while we migrates a page.
634 * This rcu_read_lock() delays freeing anon_vma pointer until the end
635 * of migration. File cache pages are no problem because of page_lock()
636 * File Caches may use write_page() or lock_page() in migration, then,
637 * just care Anon page here.
639 if (PageAnon(page)) {
645 * Corner case handling:
646 * 1. When a new swap-cache page is read into, it is added to the LRU
647 * and treated as swapcache but it has no rmap yet.
648 * Calling try_to_unmap() against a page->mapping==NULL page will
649 * trigger a BUG. So handle it here.
650 * 2. An orphaned page (see truncate_complete_page) might have
651 * fs-private metadata. The page can be picked up due to memory
652 * offlining. Everywhere else except page reclaim, the page is
653 * invisible to the vm, so the page can not be migrated. So try to
654 * free the metadata, so the page can be freed.
656 if (!page->mapping) {
657 if (!PageAnon(page) && PagePrivate(page)) {
659 * Go direct to try_to_free_buffers() here because
660 * a) that's what try_to_release_page() would do anyway
661 * b) we may be under rcu_read_lock() here, so we can't
662 * use GFP_KERNEL which is what try_to_release_page()
663 * needs to be effective.
665 try_to_free_buffers(page);
670 /* Establish migration ptes or remove ptes */
671 try_to_unmap(page, 1);
673 if (!page_mapped(page))
674 rc = move_to_new_page(newpage, page);
677 remove_migration_ptes(page, page);
688 * A page that has been migrated has all references
689 * removed and will be freed. A page that has not been
690 * migrated will have kepts its references and be
693 list_del(&page->lru);
699 * Move the new page to the LRU. If migration was not successful
700 * then this will free the page.
702 move_to_lru(newpage);
707 *result = page_to_nid(newpage);
715 * The function takes one list of pages to migrate and a function
716 * that determines from the page to be migrated and the private data
717 * the target of the move and allocates the page.
719 * The function returns after 10 attempts or if no pages
720 * are movable anymore because to has become empty
721 * or no retryable pages exist anymore. All pages will be
722 * returned to the LRU or freed.
724 * Return: Number of pages not migrated or error code.
726 int migrate_pages(struct list_head *from,
727 new_page_t get_new_page, unsigned long private)
734 int swapwrite = current->flags & PF_SWAPWRITE;
738 current->flags |= PF_SWAPWRITE;
740 for(pass = 0; pass < 10 && retry; pass++) {
743 list_for_each_entry_safe(page, page2, from, lru) {
746 rc = unmap_and_move(get_new_page, private,
758 /* Permanent failure */
767 current->flags &= ~PF_SWAPWRITE;
769 putback_lru_pages(from);
774 return nr_failed + retry;
779 * Move a list of individual pages
781 struct page_to_node {
788 static struct page *new_page_node(struct page *p, unsigned long private,
791 struct page_to_node *pm = (struct page_to_node *)private;
793 while (pm->node != MAX_NUMNODES && pm->page != p)
796 if (pm->node == MAX_NUMNODES)
799 *result = &pm->status;
801 return alloc_pages_node(pm->node,
802 GFP_HIGHUSER_MOVABLE | GFP_THISNODE, 0);
806 * Move a set of pages as indicated in the pm array. The addr
807 * field must be set to the virtual address of the page to be moved
808 * and the node number must contain a valid target node.
810 static int do_move_pages(struct mm_struct *mm, struct page_to_node *pm,
814 struct page_to_node *pp;
817 down_read(&mm->mmap_sem);
820 * Build a list of pages to migrate
823 for (pp = pm; pp->node != MAX_NUMNODES; pp++) {
824 struct vm_area_struct *vma;
828 * A valid page pointer that will not match any of the
829 * pages that will be moved.
831 pp->page = ZERO_PAGE(0);
834 vma = find_vma(mm, pp->addr);
835 if (!vma || !vma_migratable(vma))
838 page = follow_page(vma, pp->addr, FOLL_GET);
843 if (PageReserved(page)) /* Check for zero page */
847 err = page_to_nid(page);
851 * Node already in the right place
856 if (page_mapcount(page) > 1 &&
860 err = isolate_lru_page(page, &pagelist);
863 * Either remove the duplicate refcount from
864 * isolate_lru_page() or drop the page ref if it was
872 if (!list_empty(&pagelist))
873 err = migrate_pages(&pagelist, new_page_node,
878 up_read(&mm->mmap_sem);
883 * Determine the nodes of a list of pages. The addr in the pm array
884 * must have been set to the virtual address of which we want to determine
887 static int do_pages_stat(struct mm_struct *mm, struct page_to_node *pm)
889 down_read(&mm->mmap_sem);
891 for ( ; pm->node != MAX_NUMNODES; pm++) {
892 struct vm_area_struct *vma;
897 vma = find_vma(mm, pm->addr);
901 page = follow_page(vma, pm->addr, 0);
903 /* Use PageReserved to check for zero page */
904 if (!page || PageReserved(page))
907 err = page_to_nid(page);
912 up_read(&mm->mmap_sem);
917 * Move a list of pages in the address space of the currently executing
920 asmlinkage long sys_move_pages(pid_t pid, unsigned long nr_pages,
921 const void __user * __user *pages,
922 const int __user *nodes,
923 int __user *status, int flags)
927 struct task_struct *task;
928 nodemask_t task_nodes;
929 struct mm_struct *mm;
930 struct page_to_node *pm = NULL;
933 if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
936 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
939 /* Find the mm_struct */
940 read_lock(&tasklist_lock);
941 task = pid ? find_task_by_vpid(pid) : current;
943 read_unlock(&tasklist_lock);
946 mm = get_task_mm(task);
947 read_unlock(&tasklist_lock);
953 * Check if this process has the right to modify the specified
954 * process. The right exists if the process has administrative
955 * capabilities, superuser privileges or the same
956 * userid as the target process.
958 if ((current->euid != task->suid) && (current->euid != task->uid) &&
959 (current->uid != task->suid) && (current->uid != task->uid) &&
960 !capable(CAP_SYS_NICE)) {
965 err = security_task_movememory(task);
970 task_nodes = cpuset_mems_allowed(task);
972 /* Limit nr_pages so that the multiplication may not overflow */
973 if (nr_pages >= ULONG_MAX / sizeof(struct page_to_node) - 1) {
978 pm = vmalloc((nr_pages + 1) * sizeof(struct page_to_node));
985 * Get parameters from user space and initialize the pm
986 * array. Return various errors if the user did something wrong.
988 for (i = 0; i < nr_pages; i++) {
989 const void __user *p;
992 if (get_user(p, pages + i))
995 pm[i].addr = (unsigned long)p;
999 if (get_user(node, nodes + i))
1003 if (!node_state(node, N_HIGH_MEMORY))
1007 if (!node_isset(node, task_nodes))
1012 pm[i].node = 0; /* anything to not match MAX_NUMNODES */
1015 pm[nr_pages].node = MAX_NUMNODES;
1018 err = do_move_pages(mm, pm, flags & MPOL_MF_MOVE_ALL);
1020 err = do_pages_stat(mm, pm);
1023 /* Return status information */
1024 for (i = 0; i < nr_pages; i++)
1025 if (put_user(pm[i].status, status + i))
1037 * Call migration functions in the vma_ops that may prepare
1038 * memory in a vm for migration. migration functions may perform
1039 * the migration for vmas that do not have an underlying page struct.
1041 int migrate_vmas(struct mm_struct *mm, const nodemask_t *to,
1042 const nodemask_t *from, unsigned long flags)
1044 struct vm_area_struct *vma;
1047 for(vma = mm->mmap; vma->vm_next && !err; vma = vma->vm_next) {
1048 if (vma->vm_ops && vma->vm_ops->migrate) {
1049 err = vma->vm_ops->migrate(vma, to, from, flags);