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);
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 set_pte_at(mm, addr, ptep, pte);
178 page_add_anon_rmap(new, vma, addr);
180 page_add_file_rmap(new);
182 /* No need to invalidate - it was non-present before */
183 update_mmu_cache(vma, addr, pte);
184 lazy_mmu_prot_update(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)
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.
336 write_unlock_irq(&mapping->tree_lock);
342 * Copy the page to its new location
344 static void migrate_page_copy(struct page *newpage, struct page *page)
346 copy_highpage(newpage, page);
349 SetPageError(newpage);
350 if (PageReferenced(page))
351 SetPageReferenced(newpage);
352 if (PageUptodate(page))
353 SetPageUptodate(newpage);
354 if (PageActive(page))
355 SetPageActive(newpage);
356 if (PageChecked(page))
357 SetPageChecked(newpage);
358 if (PageMappedToDisk(page))
359 SetPageMappedToDisk(newpage);
361 if (PageDirty(page)) {
362 clear_page_dirty_for_io(page);
363 set_page_dirty(newpage);
367 ClearPageSwapCache(page);
369 ClearPageActive(page);
370 ClearPagePrivate(page);
371 set_page_private(page, 0);
372 page->mapping = NULL;
375 * If any waiters have accumulated on the new page then
378 if (PageWriteback(newpage))
379 end_page_writeback(newpage);
382 /************************************************************
383 * Migration functions
384 ***********************************************************/
386 /* Always fail migration. Used for mappings that are not movable */
387 int fail_migrate_page(struct address_space *mapping,
388 struct page *newpage, struct page *page)
392 EXPORT_SYMBOL(fail_migrate_page);
395 * Common logic to directly migrate a single page suitable for
396 * pages that do not use PagePrivate.
398 * Pages are locked upon entry and exit.
400 int migrate_page(struct address_space *mapping,
401 struct page *newpage, struct page *page)
405 BUG_ON(PageWriteback(page)); /* Writeback must be complete */
407 rc = migrate_page_move_mapping(mapping, newpage, page);
412 migrate_page_copy(newpage, page);
415 EXPORT_SYMBOL(migrate_page);
419 * Migration function for pages with buffers. This function can only be used
420 * if the underlying filesystem guarantees that no other references to "page"
423 int buffer_migrate_page(struct address_space *mapping,
424 struct page *newpage, struct page *page)
426 struct buffer_head *bh, *head;
429 if (!page_has_buffers(page))
430 return migrate_page(mapping, newpage, page);
432 head = page_buffers(page);
434 rc = migrate_page_move_mapping(mapping, newpage, page);
443 bh = bh->b_this_page;
445 } while (bh != head);
447 ClearPagePrivate(page);
448 set_page_private(newpage, page_private(page));
449 set_page_private(page, 0);
455 set_bh_page(bh, newpage, bh_offset(bh));
456 bh = bh->b_this_page;
458 } while (bh != head);
460 SetPagePrivate(newpage);
462 migrate_page_copy(newpage, page);
468 bh = bh->b_this_page;
470 } while (bh != head);
474 EXPORT_SYMBOL(buffer_migrate_page);
478 * Writeback a page to clean the dirty state
480 static int writeout(struct address_space *mapping, struct page *page)
482 struct writeback_control wbc = {
483 .sync_mode = WB_SYNC_NONE,
486 .range_end = LLONG_MAX,
492 if (!mapping->a_ops->writepage)
493 /* No write method for the address space */
496 if (!clear_page_dirty_for_io(page))
497 /* Someone else already triggered a write */
501 * A dirty page may imply that the underlying filesystem has
502 * the page on some queue. So the page must be clean for
503 * migration. Writeout may mean we loose the lock and the
504 * page state is no longer what we checked for earlier.
505 * At this point we know that the migration attempt cannot
508 remove_migration_ptes(page, page);
510 rc = mapping->a_ops->writepage(page, &wbc);
512 /* I/O Error writing */
515 if (rc != AOP_WRITEPAGE_ACTIVATE)
516 /* unlocked. Relock */
523 * Default handling if a filesystem does not provide a migration function.
525 static int fallback_migrate_page(struct address_space *mapping,
526 struct page *newpage, struct page *page)
529 return writeout(mapping, page);
532 * Buffers may be managed in a filesystem specific way.
533 * We must have no buffers or drop them.
535 if (PagePrivate(page) &&
536 !try_to_release_page(page, GFP_KERNEL))
539 return migrate_page(mapping, newpage, page);
543 * Move a page to a newly allocated page
544 * The page is locked and all ptes have been successfully removed.
546 * The new page will have replaced the old page if this function
549 static int move_to_new_page(struct page *newpage, struct page *page)
551 struct address_space *mapping;
555 * Block others from accessing the page when we get around to
556 * establishing additional references. We are the only one
557 * holding a reference to the new page at this point.
559 if (TestSetPageLocked(newpage))
562 /* Prepare mapping for the new page.*/
563 newpage->index = page->index;
564 newpage->mapping = page->mapping;
566 mapping = page_mapping(page);
568 rc = migrate_page(mapping, newpage, page);
569 else if (mapping->a_ops->migratepage)
571 * Most pages have a mapping and most filesystems
572 * should provide a migration function. Anonymous
573 * pages are part of swap space which also has its
574 * own migration function. This is the most common
575 * path for page migration.
577 rc = mapping->a_ops->migratepage(mapping,
580 rc = fallback_migrate_page(mapping, newpage, page);
583 remove_migration_ptes(page, newpage);
585 newpage->mapping = NULL;
587 unlock_page(newpage);
593 * Obtain the lock on page, remove all ptes and migrate the page
594 * to the newly allocated page in newpage.
596 static int unmap_and_move(new_page_t get_new_page, unsigned long private,
597 struct page *page, int force)
601 struct page *newpage = get_new_page(page, private, &result);
606 if (page_count(page) == 1)
607 /* page was freed from under us. So we are done. */
611 if (TestSetPageLocked(page)) {
617 if (PageWriteback(page)) {
620 wait_on_page_writeback(page);
624 * Establish migration ptes or remove ptes
626 try_to_unmap(page, 1);
627 if (!page_mapped(page))
628 rc = move_to_new_page(newpage, page);
631 remove_migration_ptes(page, page);
638 * A page that has been migrated has all references
639 * removed and will be freed. A page that has not been
640 * migrated will have kepts its references and be
643 list_del(&page->lru);
649 * Move the new page to the LRU. If migration was not successful
650 * then this will free the page.
652 move_to_lru(newpage);
657 *result = page_to_nid(newpage);
665 * The function takes one list of pages to migrate and a function
666 * that determines from the page to be migrated and the private data
667 * the target of the move and allocates the page.
669 * The function returns after 10 attempts or if no pages
670 * are movable anymore because to has become empty
671 * or no retryable pages exist anymore. All pages will be
672 * retruned to the LRU or freed.
674 * Return: Number of pages not migrated or error code.
676 int migrate_pages(struct list_head *from,
677 new_page_t get_new_page, unsigned long private)
684 int swapwrite = current->flags & PF_SWAPWRITE;
688 current->flags |= PF_SWAPWRITE;
690 for(pass = 0; pass < 10 && retry; pass++) {
693 list_for_each_entry_safe(page, page2, from, lru) {
696 rc = unmap_and_move(get_new_page, private,
708 /* Permanent failure */
717 current->flags &= ~PF_SWAPWRITE;
719 putback_lru_pages(from);
724 return nr_failed + retry;
729 * Move a list of individual pages
731 struct page_to_node {
738 static struct page *new_page_node(struct page *p, unsigned long private,
741 struct page_to_node *pm = (struct page_to_node *)private;
743 while (pm->node != MAX_NUMNODES && pm->page != p)
746 if (pm->node == MAX_NUMNODES)
749 *result = &pm->status;
751 return alloc_pages_node(pm->node, GFP_HIGHUSER | GFP_THISNODE, 0);
755 * Move a set of pages as indicated in the pm array. The addr
756 * field must be set to the virtual address of the page to be moved
757 * and the node number must contain a valid target node.
759 static int do_move_pages(struct mm_struct *mm, struct page_to_node *pm,
763 struct page_to_node *pp;
766 down_read(&mm->mmap_sem);
769 * Build a list of pages to migrate
772 for (pp = pm; pp->node != MAX_NUMNODES; pp++) {
773 struct vm_area_struct *vma;
777 * A valid page pointer that will not match any of the
778 * pages that will be moved.
780 pp->page = ZERO_PAGE(0);
783 vma = find_vma(mm, pp->addr);
787 page = follow_page(vma, pp->addr, FOLL_GET);
792 if (PageReserved(page)) /* Check for zero page */
796 err = page_to_nid(page);
800 * Node already in the right place
805 if (page_mapcount(page) > 1 &&
809 err = isolate_lru_page(page, &pagelist);
812 * Either remove the duplicate refcount from
813 * isolate_lru_page() or drop the page ref if it was
821 if (!list_empty(&pagelist))
822 err = migrate_pages(&pagelist, new_page_node,
827 up_read(&mm->mmap_sem);
832 * Determine the nodes of a list of pages. The addr in the pm array
833 * must have been set to the virtual address of which we want to determine
836 static int do_pages_stat(struct mm_struct *mm, struct page_to_node *pm)
838 down_read(&mm->mmap_sem);
840 for ( ; pm->node != MAX_NUMNODES; pm++) {
841 struct vm_area_struct *vma;
846 vma = find_vma(mm, pm->addr);
850 page = follow_page(vma, pm->addr, 0);
852 /* Use PageReserved to check for zero page */
853 if (!page || PageReserved(page))
856 err = page_to_nid(page);
861 up_read(&mm->mmap_sem);
866 * Move a list of pages in the address space of the currently executing
869 asmlinkage long sys_move_pages(pid_t pid, unsigned long nr_pages,
870 const void __user * __user *pages,
871 const int __user *nodes,
872 int __user *status, int flags)
876 struct task_struct *task;
877 nodemask_t task_nodes;
878 struct mm_struct *mm;
879 struct page_to_node *pm = NULL;
882 if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
885 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
888 /* Find the mm_struct */
889 read_lock(&tasklist_lock);
890 task = pid ? find_task_by_pid(pid) : current;
892 read_unlock(&tasklist_lock);
895 mm = get_task_mm(task);
896 read_unlock(&tasklist_lock);
902 * Check if this process has the right to modify the specified
903 * process. The right exists if the process has administrative
904 * capabilities, superuser privileges or the same
905 * userid as the target process.
907 if ((current->euid != task->suid) && (current->euid != task->uid) &&
908 (current->uid != task->suid) && (current->uid != task->uid) &&
909 !capable(CAP_SYS_NICE)) {
914 err = security_task_movememory(task);
919 task_nodes = cpuset_mems_allowed(task);
921 /* Limit nr_pages so that the multiplication may not overflow */
922 if (nr_pages >= ULONG_MAX / sizeof(struct page_to_node) - 1) {
927 pm = vmalloc((nr_pages + 1) * sizeof(struct page_to_node));
934 * Get parameters from user space and initialize the pm
935 * array. Return various errors if the user did something wrong.
937 for (i = 0; i < nr_pages; i++) {
941 if (get_user(p, pages + i))
944 pm[i].addr = (unsigned long)p;
948 if (get_user(node, nodes + i))
952 if (!node_online(node))
956 if (!node_isset(node, task_nodes))
961 pm[i].node = 0; /* anything to not match MAX_NUMNODES */
964 pm[nr_pages].node = MAX_NUMNODES;
967 err = do_move_pages(mm, pm, flags & MPOL_MF_MOVE_ALL);
969 err = do_pages_stat(mm, pm);
972 /* Return status information */
973 for (i = 0; i < nr_pages; i++)
974 if (put_user(pm[i].status, status + i))
986 * Call migration functions in the vma_ops that may prepare
987 * memory in a vm for migration. migration functions may perform
988 * the migration for vmas that do not have an underlying page struct.
990 int migrate_vmas(struct mm_struct *mm, const nodemask_t *to,
991 const nodemask_t *from, unsigned long flags)
993 struct vm_area_struct *vma;
996 for(vma = mm->mmap; vma->vm_next && !err; vma = vma->vm_next) {
997 if (vma->vm_ops && vma->vm_ops->migrate) {
998 err = vma->vm_ops->migrate(vma, to, from, flags);