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);
118 static inline int is_swap_pte(pte_t pte)
120 return !pte_none(pte) && !pte_present(pte) && !pte_file(pte);
124 * Restore a potential migration pte to a working pte entry
126 static void remove_migration_pte(struct vm_area_struct *vma,
127 struct page *old, struct page *new)
129 struct mm_struct *mm = vma->vm_mm;
136 unsigned long addr = page_address_in_vma(new, vma);
141 pgd = pgd_offset(mm, addr);
142 if (!pgd_present(*pgd))
145 pud = pud_offset(pgd, addr);
146 if (!pud_present(*pud))
149 pmd = pmd_offset(pud, addr);
150 if (!pmd_present(*pmd))
153 ptep = pte_offset_map(pmd, addr);
155 if (!is_swap_pte(*ptep)) {
160 ptl = pte_lockptr(mm, pmd);
163 if (!is_swap_pte(pte))
166 entry = pte_to_swp_entry(pte);
168 if (!is_migration_entry(entry) || migration_entry_to_page(entry) != old)
172 pte = pte_mkold(mk_pte(new, vma->vm_page_prot));
173 if (is_write_migration_entry(entry))
174 pte = pte_mkwrite(pte);
175 flush_cache_page(vma, addr, pte_pfn(pte));
176 set_pte_at(mm, addr, ptep, pte);
179 page_add_anon_rmap(new, vma, addr);
181 page_add_file_rmap(new);
183 /* No need to invalidate - it was non-present before */
184 update_mmu_cache(vma, addr, pte);
187 pte_unmap_unlock(ptep, ptl);
191 * Note that remove_file_migration_ptes will only work on regular mappings,
192 * Nonlinear mappings do not use migration entries.
194 static void remove_file_migration_ptes(struct page *old, struct page *new)
196 struct vm_area_struct *vma;
197 struct address_space *mapping = page_mapping(new);
198 struct prio_tree_iter iter;
199 pgoff_t pgoff = new->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
204 spin_lock(&mapping->i_mmap_lock);
206 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff)
207 remove_migration_pte(vma, old, new);
209 spin_unlock(&mapping->i_mmap_lock);
213 * Must hold mmap_sem lock on at least one of the vmas containing
214 * the page so that the anon_vma cannot vanish.
216 static void remove_anon_migration_ptes(struct page *old, struct page *new)
218 struct anon_vma *anon_vma;
219 struct vm_area_struct *vma;
220 unsigned long mapping;
222 mapping = (unsigned long)new->mapping;
224 if (!mapping || (mapping & PAGE_MAPPING_ANON) == 0)
228 * We hold the mmap_sem lock. So no need to call page_lock_anon_vma.
230 anon_vma = (struct anon_vma *) (mapping - PAGE_MAPPING_ANON);
231 spin_lock(&anon_vma->lock);
233 list_for_each_entry(vma, &anon_vma->head, anon_vma_node)
234 remove_migration_pte(vma, old, new);
236 spin_unlock(&anon_vma->lock);
240 * Get rid of all migration entries and replace them by
241 * references to the indicated page.
243 static void remove_migration_ptes(struct page *old, struct page *new)
246 remove_anon_migration_ptes(old, new);
248 remove_file_migration_ptes(old, new);
252 * Something used the pte of a page under migration. We need to
253 * get to the page and wait until migration is finished.
254 * When we return from this function the fault will be retried.
256 * This function is called from do_swap_page().
258 void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
259 unsigned long address)
266 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
268 if (!is_swap_pte(pte))
271 entry = pte_to_swp_entry(pte);
272 if (!is_migration_entry(entry))
275 page = migration_entry_to_page(entry);
278 pte_unmap_unlock(ptep, ptl);
279 wait_on_page_locked(page);
283 pte_unmap_unlock(ptep, ptl);
287 * Replace the page in the mapping.
289 * The number of remaining references must be:
290 * 1 for anonymous pages without a mapping
291 * 2 for pages with a mapping
292 * 3 for pages with a mapping and PagePrivate set.
294 static int migrate_page_move_mapping(struct address_space *mapping,
295 struct page *newpage, struct page *page)
300 /* Anonymous page without mapping */
301 if (page_count(page) != 1)
306 write_lock_irq(&mapping->tree_lock);
308 pslot = radix_tree_lookup_slot(&mapping->page_tree,
311 if (page_count(page) != 2 + !!PagePrivate(page) ||
312 (struct page *)radix_tree_deref_slot(pslot) != page) {
313 write_unlock_irq(&mapping->tree_lock);
318 * Now we know that no one else is looking at the page.
320 get_page(newpage); /* add cache reference */
322 if (PageSwapCache(page)) {
323 SetPageSwapCache(newpage);
324 set_page_private(newpage, page_private(page));
328 radix_tree_replace_slot(pslot, newpage);
331 * Drop cache reference from old page.
332 * We know this isn't the last reference.
337 * If moved to a different zone then also account
338 * the page for that zone. Other VM counters will be
339 * taken care of when we establish references to the
340 * new page and drop references to the old page.
342 * Note that anonymous pages are accounted for
343 * via NR_FILE_PAGES and NR_ANON_PAGES if they
344 * are mapped to swap space.
346 __dec_zone_page_state(page, NR_FILE_PAGES);
347 __inc_zone_page_state(newpage, NR_FILE_PAGES);
349 write_unlock_irq(&mapping->tree_lock);
355 * Copy the page to its new location
357 static void migrate_page_copy(struct page *newpage, struct page *page)
359 copy_highpage(newpage, page);
362 SetPageError(newpage);
363 if (PageReferenced(page))
364 SetPageReferenced(newpage);
365 if (PageUptodate(page))
366 SetPageUptodate(newpage);
367 if (PageActive(page))
368 SetPageActive(newpage);
369 if (PageChecked(page))
370 SetPageChecked(newpage);
371 if (PageMappedToDisk(page))
372 SetPageMappedToDisk(newpage);
374 if (PageDirty(page)) {
375 clear_page_dirty_for_io(page);
376 set_page_dirty(newpage);
380 ClearPageSwapCache(page);
382 ClearPageActive(page);
383 ClearPagePrivate(page);
384 set_page_private(page, 0);
385 page->mapping = NULL;
388 * If any waiters have accumulated on the new page then
391 if (PageWriteback(newpage))
392 end_page_writeback(newpage);
395 /************************************************************
396 * Migration functions
397 ***********************************************************/
399 /* Always fail migration. Used for mappings that are not movable */
400 int fail_migrate_page(struct address_space *mapping,
401 struct page *newpage, struct page *page)
405 EXPORT_SYMBOL(fail_migrate_page);
408 * Common logic to directly migrate a single page suitable for
409 * pages that do not use PagePrivate.
411 * Pages are locked upon entry and exit.
413 int migrate_page(struct address_space *mapping,
414 struct page *newpage, struct page *page)
418 BUG_ON(PageWriteback(page)); /* Writeback must be complete */
420 rc = migrate_page_move_mapping(mapping, newpage, page);
425 migrate_page_copy(newpage, page);
428 EXPORT_SYMBOL(migrate_page);
432 * Migration function for pages with buffers. This function can only be used
433 * if the underlying filesystem guarantees that no other references to "page"
436 int buffer_migrate_page(struct address_space *mapping,
437 struct page *newpage, struct page *page)
439 struct buffer_head *bh, *head;
442 if (!page_has_buffers(page))
443 return migrate_page(mapping, newpage, page);
445 head = page_buffers(page);
447 rc = migrate_page_move_mapping(mapping, newpage, page);
456 bh = bh->b_this_page;
458 } while (bh != head);
460 ClearPagePrivate(page);
461 set_page_private(newpage, page_private(page));
462 set_page_private(page, 0);
468 set_bh_page(bh, newpage, bh_offset(bh));
469 bh = bh->b_this_page;
471 } while (bh != head);
473 SetPagePrivate(newpage);
475 migrate_page_copy(newpage, page);
481 bh = bh->b_this_page;
483 } while (bh != head);
487 EXPORT_SYMBOL(buffer_migrate_page);
491 * Writeback a page to clean the dirty state
493 static int writeout(struct address_space *mapping, struct page *page)
495 struct writeback_control wbc = {
496 .sync_mode = WB_SYNC_NONE,
499 .range_end = LLONG_MAX,
505 if (!mapping->a_ops->writepage)
506 /* No write method for the address space */
509 if (!clear_page_dirty_for_io(page))
510 /* Someone else already triggered a write */
514 * A dirty page may imply that the underlying filesystem has
515 * the page on some queue. So the page must be clean for
516 * migration. Writeout may mean we loose the lock and the
517 * page state is no longer what we checked for earlier.
518 * At this point we know that the migration attempt cannot
521 remove_migration_ptes(page, page);
523 rc = mapping->a_ops->writepage(page, &wbc);
525 /* I/O Error writing */
528 if (rc != AOP_WRITEPAGE_ACTIVATE)
529 /* unlocked. Relock */
536 * Default handling if a filesystem does not provide a migration function.
538 static int fallback_migrate_page(struct address_space *mapping,
539 struct page *newpage, struct page *page)
542 return writeout(mapping, page);
545 * Buffers may be managed in a filesystem specific way.
546 * We must have no buffers or drop them.
548 if (PagePrivate(page) &&
549 !try_to_release_page(page, GFP_KERNEL))
552 return migrate_page(mapping, newpage, page);
556 * Move a page to a newly allocated page
557 * The page is locked and all ptes have been successfully removed.
559 * The new page will have replaced the old page if this function
562 static int move_to_new_page(struct page *newpage, struct page *page)
564 struct address_space *mapping;
568 * Block others from accessing the page when we get around to
569 * establishing additional references. We are the only one
570 * holding a reference to the new page at this point.
572 if (TestSetPageLocked(newpage))
575 /* Prepare mapping for the new page.*/
576 newpage->index = page->index;
577 newpage->mapping = page->mapping;
579 mapping = page_mapping(page);
581 rc = migrate_page(mapping, newpage, page);
582 else if (mapping->a_ops->migratepage)
584 * Most pages have a mapping and most filesystems
585 * should provide a migration function. Anonymous
586 * pages are part of swap space which also has its
587 * own migration function. This is the most common
588 * path for page migration.
590 rc = mapping->a_ops->migratepage(mapping,
593 rc = fallback_migrate_page(mapping, newpage, page);
596 remove_migration_ptes(page, newpage);
598 newpage->mapping = NULL;
600 unlock_page(newpage);
606 * Obtain the lock on page, remove all ptes and migrate the page
607 * to the newly allocated page in newpage.
609 static int unmap_and_move(new_page_t get_new_page, unsigned long private,
610 struct page *page, int force)
614 struct page *newpage = get_new_page(page, private, &result);
620 if (page_count(page) == 1)
621 /* page was freed from under us. So we are done. */
625 if (TestSetPageLocked(page)) {
631 if (PageWriteback(page)) {
634 wait_on_page_writeback(page);
637 * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
638 * we cannot notice that anon_vma is freed while we migrates a page.
639 * This rcu_read_lock() delays freeing anon_vma pointer until the end
640 * of migration. File cache pages are no problem because of page_lock()
641 * File Caches may use write_page() or lock_page() in migration, then,
642 * just care Anon page here.
644 if (PageAnon(page)) {
649 * This is a corner case handling.
650 * When a new swap-cache is read into, it is linked to LRU
651 * and treated as swapcache but has no rmap yet.
652 * Calling try_to_unmap() against a page->mapping==NULL page is
653 * BUG. So handle it here.
657 /* Establish migration ptes or remove ptes */
658 try_to_unmap(page, 1);
660 if (!page_mapped(page))
661 rc = move_to_new_page(newpage, page);
664 remove_migration_ptes(page, page);
675 * A page that has been migrated has all references
676 * removed and will be freed. A page that has not been
677 * migrated will have kepts its references and be
680 list_del(&page->lru);
686 * Move the new page to the LRU. If migration was not successful
687 * then this will free the page.
689 move_to_lru(newpage);
694 *result = page_to_nid(newpage);
702 * The function takes one list of pages to migrate and a function
703 * that determines from the page to be migrated and the private data
704 * the target of the move and allocates the page.
706 * The function returns after 10 attempts or if no pages
707 * are movable anymore because to has become empty
708 * or no retryable pages exist anymore. All pages will be
709 * returned to the LRU or freed.
711 * Return: Number of pages not migrated or error code.
713 int migrate_pages(struct list_head *from,
714 new_page_t get_new_page, unsigned long private)
721 int swapwrite = current->flags & PF_SWAPWRITE;
725 current->flags |= PF_SWAPWRITE;
727 for(pass = 0; pass < 10 && retry; pass++) {
730 list_for_each_entry_safe(page, page2, from, lru) {
733 rc = unmap_and_move(get_new_page, private,
745 /* Permanent failure */
754 current->flags &= ~PF_SWAPWRITE;
756 putback_lru_pages(from);
761 return nr_failed + retry;
766 * Move a list of individual pages
768 struct page_to_node {
775 static struct page *new_page_node(struct page *p, unsigned long private,
778 struct page_to_node *pm = (struct page_to_node *)private;
780 while (pm->node != MAX_NUMNODES && pm->page != p)
783 if (pm->node == MAX_NUMNODES)
786 *result = &pm->status;
788 return alloc_pages_node(pm->node,
789 GFP_HIGHUSER_MOVABLE | GFP_THISNODE, 0);
793 * Move a set of pages as indicated in the pm array. The addr
794 * field must be set to the virtual address of the page to be moved
795 * and the node number must contain a valid target node.
797 static int do_move_pages(struct mm_struct *mm, struct page_to_node *pm,
801 struct page_to_node *pp;
804 down_read(&mm->mmap_sem);
807 * Build a list of pages to migrate
810 for (pp = pm; pp->node != MAX_NUMNODES; pp++) {
811 struct vm_area_struct *vma;
815 * A valid page pointer that will not match any of the
816 * pages that will be moved.
818 pp->page = ZERO_PAGE(0);
821 vma = find_vma(mm, pp->addr);
822 if (!vma || !vma_migratable(vma))
825 page = follow_page(vma, pp->addr, FOLL_GET);
830 if (PageReserved(page)) /* Check for zero page */
834 err = page_to_nid(page);
838 * Node already in the right place
843 if (page_mapcount(page) > 1 &&
847 err = isolate_lru_page(page, &pagelist);
850 * Either remove the duplicate refcount from
851 * isolate_lru_page() or drop the page ref if it was
859 if (!list_empty(&pagelist))
860 err = migrate_pages(&pagelist, new_page_node,
865 up_read(&mm->mmap_sem);
870 * Determine the nodes of a list of pages. The addr in the pm array
871 * must have been set to the virtual address of which we want to determine
874 static int do_pages_stat(struct mm_struct *mm, struct page_to_node *pm)
876 down_read(&mm->mmap_sem);
878 for ( ; pm->node != MAX_NUMNODES; pm++) {
879 struct vm_area_struct *vma;
884 vma = find_vma(mm, pm->addr);
888 page = follow_page(vma, pm->addr, 0);
890 /* Use PageReserved to check for zero page */
891 if (!page || PageReserved(page))
894 err = page_to_nid(page);
899 up_read(&mm->mmap_sem);
904 * Move a list of pages in the address space of the currently executing
907 asmlinkage long sys_move_pages(pid_t pid, unsigned long nr_pages,
908 const void __user * __user *pages,
909 const int __user *nodes,
910 int __user *status, int flags)
914 struct task_struct *task;
915 nodemask_t task_nodes;
916 struct mm_struct *mm;
917 struct page_to_node *pm = NULL;
920 if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
923 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
926 /* Find the mm_struct */
927 read_lock(&tasklist_lock);
928 task = pid ? find_task_by_vpid(pid) : current;
930 read_unlock(&tasklist_lock);
933 mm = get_task_mm(task);
934 read_unlock(&tasklist_lock);
940 * Check if this process has the right to modify the specified
941 * process. The right exists if the process has administrative
942 * capabilities, superuser privileges or the same
943 * userid as the target process.
945 if ((current->euid != task->suid) && (current->euid != task->uid) &&
946 (current->uid != task->suid) && (current->uid != task->uid) &&
947 !capable(CAP_SYS_NICE)) {
952 err = security_task_movememory(task);
957 task_nodes = cpuset_mems_allowed(task);
959 /* Limit nr_pages so that the multiplication may not overflow */
960 if (nr_pages >= ULONG_MAX / sizeof(struct page_to_node) - 1) {
965 pm = vmalloc((nr_pages + 1) * sizeof(struct page_to_node));
972 * Get parameters from user space and initialize the pm
973 * array. Return various errors if the user did something wrong.
975 for (i = 0; i < nr_pages; i++) {
976 const void __user *p;
979 if (get_user(p, pages + i))
982 pm[i].addr = (unsigned long)p;
986 if (get_user(node, nodes + i))
990 if (!node_state(node, N_HIGH_MEMORY))
994 if (!node_isset(node, task_nodes))
999 pm[i].node = 0; /* anything to not match MAX_NUMNODES */
1002 pm[nr_pages].node = MAX_NUMNODES;
1005 err = do_move_pages(mm, pm, flags & MPOL_MF_MOVE_ALL);
1007 err = do_pages_stat(mm, pm);
1010 /* Return status information */
1011 for (i = 0; i < nr_pages; i++)
1012 if (put_user(pm[i].status, status + i))
1024 * Call migration functions in the vma_ops that may prepare
1025 * memory in a vm for migration. migration functions may perform
1026 * the migration for vmas that do not have an underlying page struct.
1028 int migrate_vmas(struct mm_struct *mm, const nodemask_t *to,
1029 const nodemask_t *from, unsigned long flags)
1031 struct vm_area_struct *vma;
1034 for(vma = mm->mmap; vma->vm_next && !err; vma = vma->vm_next) {
1035 if (vma->vm_ops && vma->vm_ops->migrate) {
1036 err = vma->vm_ops->migrate(vma, to, from, flags);