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>
32 #include <linux/memcontrol.h>
36 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
39 * Isolate one page from the LRU lists. If successful put it onto
40 * the indicated list with elevated page count.
43 * -EBUSY: page not on LRU list
44 * 0: page removed from LRU list and added to the specified list.
46 int isolate_lru_page(struct page *page, struct list_head *pagelist)
51 struct zone *zone = page_zone(page);
53 spin_lock_irq(&zone->lru_lock);
54 if (PageLRU(page) && get_page_unless_zero(page)) {
58 del_page_from_active_list(zone, page);
60 del_page_from_inactive_list(zone, page);
61 list_add_tail(&page->lru, pagelist);
63 spin_unlock_irq(&zone->lru_lock);
69 * migrate_prep() needs to be called before we start compiling a list of pages
70 * to be migrated using isolate_lru_page().
72 int migrate_prep(void)
75 * Clear the LRU lists so pages can be isolated.
76 * Note that pages may be moved off the LRU after we have
77 * drained them. Those pages will fail to migrate like other
78 * pages that may be busy.
85 static inline void move_to_lru(struct page *page)
87 if (PageActive(page)) {
89 * lru_cache_add_active checks that
90 * the PG_active bit is off.
92 ClearPageActive(page);
93 lru_cache_add_active(page);
101 * Add isolated pages on the list back to the LRU.
103 * returns the number of pages put back.
105 int putback_lru_pages(struct list_head *l)
111 list_for_each_entry_safe(page, page2, l, lru) {
112 list_del(&page->lru);
120 * Restore a potential migration pte to a working pte entry
122 static void remove_migration_pte(struct vm_area_struct *vma,
123 struct page *old, struct page *new)
125 struct mm_struct *mm = vma->vm_mm;
132 unsigned long addr = page_address_in_vma(new, vma);
137 pgd = pgd_offset(mm, addr);
138 if (!pgd_present(*pgd))
141 pud = pud_offset(pgd, addr);
142 if (!pud_present(*pud))
145 pmd = pmd_offset(pud, addr);
146 if (!pmd_present(*pmd))
149 ptep = pte_offset_map(pmd, addr);
151 if (!is_swap_pte(*ptep)) {
156 if (mem_cgroup_charge(new, mm, GFP_KERNEL)) {
161 ptl = pte_lockptr(mm, pmd);
164 if (!is_swap_pte(pte))
167 entry = pte_to_swp_entry(pte);
169 if (!is_migration_entry(entry) || migration_entry_to_page(entry) != old)
173 pte = pte_mkold(mk_pte(new, vma->vm_page_prot));
174 if (is_write_migration_entry(entry))
175 pte = pte_mkwrite(pte);
176 flush_cache_page(vma, addr, pte_pfn(pte));
177 set_pte_at(mm, addr, ptep, pte);
180 page_add_anon_rmap(new, vma, addr);
182 page_add_file_rmap(new);
184 /* No need to invalidate - it was non-present before */
185 update_mmu_cache(vma, addr, pte);
188 pte_unmap_unlock(ptep, ptl);
192 * Note that remove_file_migration_ptes will only work on regular mappings,
193 * Nonlinear mappings do not use migration entries.
195 static void remove_file_migration_ptes(struct page *old, struct page *new)
197 struct vm_area_struct *vma;
198 struct address_space *mapping = page_mapping(new);
199 struct prio_tree_iter iter;
200 pgoff_t pgoff = new->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
205 spin_lock(&mapping->i_mmap_lock);
207 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff)
208 remove_migration_pte(vma, old, new);
210 spin_unlock(&mapping->i_mmap_lock);
214 * Must hold mmap_sem lock on at least one of the vmas containing
215 * the page so that the anon_vma cannot vanish.
217 static void remove_anon_migration_ptes(struct page *old, struct page *new)
219 struct anon_vma *anon_vma;
220 struct vm_area_struct *vma;
221 unsigned long mapping;
223 mapping = (unsigned long)new->mapping;
225 if (!mapping || (mapping & PAGE_MAPPING_ANON) == 0)
229 * We hold the mmap_sem lock. So no need to call page_lock_anon_vma.
231 anon_vma = (struct anon_vma *) (mapping - PAGE_MAPPING_ANON);
232 spin_lock(&anon_vma->lock);
234 list_for_each_entry(vma, &anon_vma->head, anon_vma_node)
235 remove_migration_pte(vma, old, new);
237 spin_unlock(&anon_vma->lock);
241 * Get rid of all migration entries and replace them by
242 * references to the indicated page.
244 static void remove_migration_ptes(struct page *old, struct page *new)
247 remove_anon_migration_ptes(old, new);
249 remove_file_migration_ptes(old, new);
253 * Something used the pte of a page under migration. We need to
254 * get to the page and wait until migration is finished.
255 * When we return from this function the fault will be retried.
257 * This function is called from do_swap_page().
259 void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
260 unsigned long address)
267 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
269 if (!is_swap_pte(pte))
272 entry = pte_to_swp_entry(pte);
273 if (!is_migration_entry(entry))
276 page = migration_entry_to_page(entry);
279 pte_unmap_unlock(ptep, ptl);
280 wait_on_page_locked(page);
284 pte_unmap_unlock(ptep, ptl);
288 * Replace the page in the mapping.
290 * The number of remaining references must be:
291 * 1 for anonymous pages without a mapping
292 * 2 for pages with a mapping
293 * 3 for pages with a mapping and PagePrivate set.
295 static int migrate_page_move_mapping(struct address_space *mapping,
296 struct page *newpage, struct page *page)
301 /* Anonymous page without mapping */
302 if (page_count(page) != 1)
307 write_lock_irq(&mapping->tree_lock);
309 pslot = radix_tree_lookup_slot(&mapping->page_tree,
312 if (page_count(page) != 2 + !!PagePrivate(page) ||
313 (struct page *)radix_tree_deref_slot(pslot) != page) {
314 write_unlock_irq(&mapping->tree_lock);
319 * Now we know that no one else is looking at the page.
321 get_page(newpage); /* add cache reference */
323 if (PageSwapCache(page)) {
324 SetPageSwapCache(newpage);
325 set_page_private(newpage, page_private(page));
329 radix_tree_replace_slot(pslot, newpage);
332 * Drop cache reference from old page.
333 * We know this isn't the last reference.
338 * If moved to a different zone then also account
339 * the page for that zone. Other VM counters will be
340 * taken care of when we establish references to the
341 * new page and drop references to the old page.
343 * Note that anonymous pages are accounted for
344 * via NR_FILE_PAGES and NR_ANON_PAGES if they
345 * are mapped to swap space.
347 __dec_zone_page_state(page, NR_FILE_PAGES);
348 __inc_zone_page_state(newpage, NR_FILE_PAGES);
350 write_unlock_irq(&mapping->tree_lock);
356 * Copy the page to its new location
358 static void migrate_page_copy(struct page *newpage, struct page *page)
360 copy_highpage(newpage, page);
363 SetPageError(newpage);
364 if (PageReferenced(page))
365 SetPageReferenced(newpage);
366 if (PageUptodate(page))
367 SetPageUptodate(newpage);
368 if (PageActive(page))
369 SetPageActive(newpage);
370 if (PageChecked(page))
371 SetPageChecked(newpage);
372 if (PageMappedToDisk(page))
373 SetPageMappedToDisk(newpage);
375 if (PageDirty(page)) {
376 clear_page_dirty_for_io(page);
377 set_page_dirty(newpage);
381 ClearPageSwapCache(page);
383 ClearPageActive(page);
384 ClearPagePrivate(page);
385 set_page_private(page, 0);
386 page->mapping = NULL;
389 * If any waiters have accumulated on the new page then
392 if (PageWriteback(newpage))
393 end_page_writeback(newpage);
396 /************************************************************
397 * Migration functions
398 ***********************************************************/
400 /* Always fail migration. Used for mappings that are not movable */
401 int fail_migrate_page(struct address_space *mapping,
402 struct page *newpage, struct page *page)
406 EXPORT_SYMBOL(fail_migrate_page);
409 * Common logic to directly migrate a single page suitable for
410 * pages that do not use PagePrivate.
412 * Pages are locked upon entry and exit.
414 int migrate_page(struct address_space *mapping,
415 struct page *newpage, struct page *page)
419 BUG_ON(PageWriteback(page)); /* Writeback must be complete */
421 rc = migrate_page_move_mapping(mapping, newpage, page);
426 migrate_page_copy(newpage, page);
429 EXPORT_SYMBOL(migrate_page);
433 * Migration function for pages with buffers. This function can only be used
434 * if the underlying filesystem guarantees that no other references to "page"
437 int buffer_migrate_page(struct address_space *mapping,
438 struct page *newpage, struct page *page)
440 struct buffer_head *bh, *head;
443 if (!page_has_buffers(page))
444 return migrate_page(mapping, newpage, page);
446 head = page_buffers(page);
448 rc = migrate_page_move_mapping(mapping, newpage, page);
457 bh = bh->b_this_page;
459 } while (bh != head);
461 ClearPagePrivate(page);
462 set_page_private(newpage, page_private(page));
463 set_page_private(page, 0);
469 set_bh_page(bh, newpage, bh_offset(bh));
470 bh = bh->b_this_page;
472 } while (bh != head);
474 SetPagePrivate(newpage);
476 migrate_page_copy(newpage, page);
482 bh = bh->b_this_page;
484 } while (bh != head);
488 EXPORT_SYMBOL(buffer_migrate_page);
492 * Writeback a page to clean the dirty state
494 static int writeout(struct address_space *mapping, struct page *page)
496 struct writeback_control wbc = {
497 .sync_mode = WB_SYNC_NONE,
500 .range_end = LLONG_MAX,
506 if (!mapping->a_ops->writepage)
507 /* No write method for the address space */
510 if (!clear_page_dirty_for_io(page))
511 /* Someone else already triggered a write */
515 * A dirty page may imply that the underlying filesystem has
516 * the page on some queue. So the page must be clean for
517 * migration. Writeout may mean we loose the lock and the
518 * page state is no longer what we checked for earlier.
519 * At this point we know that the migration attempt cannot
522 remove_migration_ptes(page, page);
524 rc = mapping->a_ops->writepage(page, &wbc);
526 /* I/O Error writing */
529 if (rc != AOP_WRITEPAGE_ACTIVATE)
530 /* unlocked. Relock */
537 * Default handling if a filesystem does not provide a migration function.
539 static int fallback_migrate_page(struct address_space *mapping,
540 struct page *newpage, struct page *page)
543 return writeout(mapping, page);
546 * Buffers may be managed in a filesystem specific way.
547 * We must have no buffers or drop them.
549 if (PagePrivate(page) &&
550 !try_to_release_page(page, GFP_KERNEL))
553 return migrate_page(mapping, newpage, page);
557 * Move a page to a newly allocated page
558 * The page is locked and all ptes have been successfully removed.
560 * The new page will have replaced the old page if this function
563 static int move_to_new_page(struct page *newpage, struct page *page)
565 struct address_space *mapping;
569 * Block others from accessing the page when we get around to
570 * establishing additional references. We are the only one
571 * holding a reference to the new page at this point.
573 if (TestSetPageLocked(newpage))
576 /* Prepare mapping for the new page.*/
577 newpage->index = page->index;
578 newpage->mapping = page->mapping;
580 mapping = page_mapping(page);
582 rc = migrate_page(mapping, newpage, page);
583 else if (mapping->a_ops->migratepage)
585 * Most pages have a mapping and most filesystems
586 * should provide a migration function. Anonymous
587 * pages are part of swap space which also has its
588 * own migration function. This is the most common
589 * path for page migration.
591 rc = mapping->a_ops->migratepage(mapping,
594 rc = fallback_migrate_page(mapping, newpage, page);
597 mem_cgroup_page_migration(page, newpage);
598 remove_migration_ptes(page, newpage);
600 newpage->mapping = NULL;
602 unlock_page(newpage);
608 * Obtain the lock on page, remove all ptes and migrate the page
609 * to the newly allocated page in newpage.
611 static int unmap_and_move(new_page_t get_new_page, unsigned long private,
612 struct page *page, int force)
616 struct page *newpage = get_new_page(page, private, &result);
623 if (page_count(page) == 1)
624 /* page was freed from under us. So we are done. */
628 if (TestSetPageLocked(page)) {
634 if (PageWriteback(page)) {
637 wait_on_page_writeback(page);
640 * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
641 * we cannot notice that anon_vma is freed while we migrates a page.
642 * This rcu_read_lock() delays freeing anon_vma pointer until the end
643 * of migration. File cache pages are no problem because of page_lock()
644 * File Caches may use write_page() or lock_page() in migration, then,
645 * just care Anon page here.
647 if (PageAnon(page)) {
653 * Corner case handling:
654 * 1. When a new swap-cache page is read into, it is added to the LRU
655 * and treated as swapcache but it has no rmap yet.
656 * Calling try_to_unmap() against a page->mapping==NULL page will
657 * trigger a BUG. So handle it here.
658 * 2. An orphaned page (see truncate_complete_page) might have
659 * fs-private metadata. The page can be picked up due to memory
660 * offlining. Everywhere else except page reclaim, the page is
661 * invisible to the vm, so the page can not be migrated. So try to
662 * free the metadata, so the page can be freed.
664 if (!page->mapping) {
665 if (!PageAnon(page) && PagePrivate(page)) {
667 * Go direct to try_to_free_buffers() here because
668 * a) that's what try_to_release_page() would do anyway
669 * b) we may be under rcu_read_lock() here, so we can't
670 * use GFP_KERNEL which is what try_to_release_page()
671 * needs to be effective.
673 try_to_free_buffers(page);
678 charge = mem_cgroup_prepare_migration(page);
679 /* Establish migration ptes or remove ptes */
680 try_to_unmap(page, 1);
682 if (!page_mapped(page))
683 rc = move_to_new_page(newpage, page);
686 remove_migration_ptes(page, page);
688 mem_cgroup_end_migration(page);
690 mem_cgroup_end_migration(newpage);
701 * A page that has been migrated has all references
702 * removed and will be freed. A page that has not been
703 * migrated will have kepts its references and be
706 list_del(&page->lru);
712 * Move the new page to the LRU. If migration was not successful
713 * then this will free the page.
715 move_to_lru(newpage);
720 *result = page_to_nid(newpage);
728 * The function takes one list of pages to migrate and a function
729 * that determines from the page to be migrated and the private data
730 * the target of the move and allocates the page.
732 * The function returns after 10 attempts or if no pages
733 * are movable anymore because to has become empty
734 * or no retryable pages exist anymore. All pages will be
735 * returned to the LRU or freed.
737 * Return: Number of pages not migrated or error code.
739 int migrate_pages(struct list_head *from,
740 new_page_t get_new_page, unsigned long private)
747 int swapwrite = current->flags & PF_SWAPWRITE;
751 current->flags |= PF_SWAPWRITE;
753 for(pass = 0; pass < 10 && retry; pass++) {
756 list_for_each_entry_safe(page, page2, from, lru) {
759 rc = unmap_and_move(get_new_page, private,
771 /* Permanent failure */
780 current->flags &= ~PF_SWAPWRITE;
782 putback_lru_pages(from);
787 return nr_failed + retry;
792 * Move a list of individual pages
794 struct page_to_node {
801 static struct page *new_page_node(struct page *p, unsigned long private,
804 struct page_to_node *pm = (struct page_to_node *)private;
806 while (pm->node != MAX_NUMNODES && pm->page != p)
809 if (pm->node == MAX_NUMNODES)
812 *result = &pm->status;
814 return alloc_pages_node(pm->node,
815 GFP_HIGHUSER_MOVABLE | GFP_THISNODE, 0);
819 * Move a set of pages as indicated in the pm array. The addr
820 * field must be set to the virtual address of the page to be moved
821 * and the node number must contain a valid target node.
823 static int do_move_pages(struct mm_struct *mm, struct page_to_node *pm,
827 struct page_to_node *pp;
830 down_read(&mm->mmap_sem);
833 * Build a list of pages to migrate
836 for (pp = pm; pp->node != MAX_NUMNODES; pp++) {
837 struct vm_area_struct *vma;
841 * A valid page pointer that will not match any of the
842 * pages that will be moved.
844 pp->page = ZERO_PAGE(0);
847 vma = find_vma(mm, pp->addr);
848 if (!vma || !vma_migratable(vma))
851 page = follow_page(vma, pp->addr, FOLL_GET);
856 if (PageReserved(page)) /* Check for zero page */
860 err = page_to_nid(page);
864 * Node already in the right place
869 if (page_mapcount(page) > 1 &&
873 err = isolate_lru_page(page, &pagelist);
876 * Either remove the duplicate refcount from
877 * isolate_lru_page() or drop the page ref if it was
885 if (!list_empty(&pagelist))
886 err = migrate_pages(&pagelist, new_page_node,
891 up_read(&mm->mmap_sem);
896 * Determine the nodes of a list of pages. The addr in the pm array
897 * must have been set to the virtual address of which we want to determine
900 static int do_pages_stat(struct mm_struct *mm, struct page_to_node *pm)
902 down_read(&mm->mmap_sem);
904 for ( ; pm->node != MAX_NUMNODES; pm++) {
905 struct vm_area_struct *vma;
910 vma = find_vma(mm, pm->addr);
914 page = follow_page(vma, pm->addr, 0);
916 /* Use PageReserved to check for zero page */
917 if (!page || PageReserved(page))
920 err = page_to_nid(page);
925 up_read(&mm->mmap_sem);
930 * Move a list of pages in the address space of the currently executing
933 asmlinkage long sys_move_pages(pid_t pid, unsigned long nr_pages,
934 const void __user * __user *pages,
935 const int __user *nodes,
936 int __user *status, int flags)
940 struct task_struct *task;
941 nodemask_t task_nodes;
942 struct mm_struct *mm;
943 struct page_to_node *pm = NULL;
946 if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
949 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
952 /* Find the mm_struct */
953 read_lock(&tasklist_lock);
954 task = pid ? find_task_by_vpid(pid) : current;
956 read_unlock(&tasklist_lock);
959 mm = get_task_mm(task);
960 read_unlock(&tasklist_lock);
966 * Check if this process has the right to modify the specified
967 * process. The right exists if the process has administrative
968 * capabilities, superuser privileges or the same
969 * userid as the target process.
971 if ((current->euid != task->suid) && (current->euid != task->uid) &&
972 (current->uid != task->suid) && (current->uid != task->uid) &&
973 !capable(CAP_SYS_NICE)) {
978 err = security_task_movememory(task);
983 task_nodes = cpuset_mems_allowed(task);
985 /* Limit nr_pages so that the multiplication may not overflow */
986 if (nr_pages >= ULONG_MAX / sizeof(struct page_to_node) - 1) {
991 pm = vmalloc((nr_pages + 1) * sizeof(struct page_to_node));
998 * Get parameters from user space and initialize the pm
999 * array. Return various errors if the user did something wrong.
1001 for (i = 0; i < nr_pages; i++) {
1002 const void __user *p;
1005 if (get_user(p, pages + i))
1008 pm[i].addr = (unsigned long)p;
1012 if (get_user(node, nodes + i))
1016 if (!node_state(node, N_HIGH_MEMORY))
1020 if (!node_isset(node, task_nodes))
1025 pm[i].node = 0; /* anything to not match MAX_NUMNODES */
1028 pm[nr_pages].node = MAX_NUMNODES;
1031 err = do_move_pages(mm, pm, flags & MPOL_MF_MOVE_ALL);
1033 err = do_pages_stat(mm, pm);
1036 /* Return status information */
1037 for (i = 0; i < nr_pages; i++)
1038 if (put_user(pm[i].status, status + i))
1050 * Call migration functions in the vma_ops that may prepare
1051 * memory in a vm for migration. migration functions may perform
1052 * the migration for vmas that do not have an underlying page struct.
1054 int migrate_vmas(struct mm_struct *mm, const nodemask_t *to,
1055 const nodemask_t *from, unsigned long flags)
1057 struct vm_area_struct *vma;
1060 for(vma = mm->mmap; vma->vm_next && !err; vma = vma->vm_next) {
1061 if (vma->vm_ops && vma->vm_ops->migrate) {
1062 err = vma->vm_ops->migrate(vma, to, from, flags);