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>
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>
33 #include <linux/syscalls.h>
37 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
40 * migrate_prep() needs to be called before we start compiling a list of pages
41 * to be migrated using isolate_lru_page().
43 int migrate_prep(void)
46 * Clear the LRU lists so pages can be isolated.
47 * Note that pages may be moved off the LRU after we have
48 * drained them. Those pages will fail to migrate like other
49 * pages that may be busy.
57 * Add isolated pages on the list back to the LRU under page lock
58 * to avoid leaking evictable pages back onto unevictable list.
60 * returns the number of pages put back.
62 int putback_lru_pages(struct list_head *l)
68 list_for_each_entry_safe(page, page2, l, lru) {
70 putback_lru_page(page);
77 * Restore a potential migration pte to a working pte entry
79 static void remove_migration_pte(struct vm_area_struct *vma,
80 struct page *old, struct page *new)
82 struct mm_struct *mm = vma->vm_mm;
89 unsigned long addr = page_address_in_vma(new, vma);
94 pgd = pgd_offset(mm, addr);
95 if (!pgd_present(*pgd))
98 pud = pud_offset(pgd, addr);
99 if (!pud_present(*pud))
102 pmd = pmd_offset(pud, addr);
103 if (!pmd_present(*pmd))
106 ptep = pte_offset_map(pmd, addr);
108 if (!is_swap_pte(*ptep)) {
113 ptl = pte_lockptr(mm, pmd);
116 if (!is_swap_pte(pte))
119 entry = pte_to_swp_entry(pte);
121 if (!is_migration_entry(entry) || migration_entry_to_page(entry) != old)
125 pte = pte_mkold(mk_pte(new, vma->vm_page_prot));
126 if (is_write_migration_entry(entry))
127 pte = pte_mkwrite(pte);
128 flush_cache_page(vma, addr, pte_pfn(pte));
129 set_pte_at(mm, addr, ptep, pte);
132 page_add_anon_rmap(new, vma, addr);
134 page_add_file_rmap(new);
136 /* No need to invalidate - it was non-present before */
137 update_mmu_cache(vma, addr, pte);
140 pte_unmap_unlock(ptep, ptl);
144 * Note that remove_file_migration_ptes will only work on regular mappings,
145 * Nonlinear mappings do not use migration entries.
147 static void remove_file_migration_ptes(struct page *old, struct page *new)
149 struct vm_area_struct *vma;
150 struct address_space *mapping = new->mapping;
151 struct prio_tree_iter iter;
152 pgoff_t pgoff = new->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
157 spin_lock(&mapping->i_mmap_lock);
159 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff)
160 remove_migration_pte(vma, old, new);
162 spin_unlock(&mapping->i_mmap_lock);
166 * Must hold mmap_sem lock on at least one of the vmas containing
167 * the page so that the anon_vma cannot vanish.
169 static void remove_anon_migration_ptes(struct page *old, struct page *new)
171 struct anon_vma *anon_vma;
172 struct vm_area_struct *vma;
173 unsigned long mapping;
175 mapping = (unsigned long)new->mapping;
177 if (!mapping || (mapping & PAGE_MAPPING_ANON) == 0)
181 * We hold the mmap_sem lock. So no need to call page_lock_anon_vma.
183 anon_vma = (struct anon_vma *) (mapping - PAGE_MAPPING_ANON);
184 spin_lock(&anon_vma->lock);
186 list_for_each_entry(vma, &anon_vma->head, anon_vma_node)
187 remove_migration_pte(vma, old, new);
189 spin_unlock(&anon_vma->lock);
193 * Get rid of all migration entries and replace them by
194 * references to the indicated page.
196 static void remove_migration_ptes(struct page *old, struct page *new)
199 remove_anon_migration_ptes(old, new);
201 remove_file_migration_ptes(old, new);
205 * Something used the pte of a page under migration. We need to
206 * get to the page and wait until migration is finished.
207 * When we return from this function the fault will be retried.
209 * This function is called from do_swap_page().
211 void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
212 unsigned long address)
219 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
221 if (!is_swap_pte(pte))
224 entry = pte_to_swp_entry(pte);
225 if (!is_migration_entry(entry))
228 page = migration_entry_to_page(entry);
231 * Once radix-tree replacement of page migration started, page_count
232 * *must* be zero. And, we don't want to call wait_on_page_locked()
233 * against a page without get_page().
234 * So, we use get_page_unless_zero(), here. Even failed, page fault
237 if (!get_page_unless_zero(page))
239 pte_unmap_unlock(ptep, ptl);
240 wait_on_page_locked(page);
244 pte_unmap_unlock(ptep, ptl);
248 * Replace the page in the mapping.
250 * The number of remaining references must be:
251 * 1 for anonymous pages without a mapping
252 * 2 for pages with a mapping
253 * 3 for pages with a mapping and PagePrivate/PagePrivate2 set.
255 static int migrate_page_move_mapping(struct address_space *mapping,
256 struct page *newpage, struct page *page)
262 /* Anonymous page without mapping */
263 if (page_count(page) != 1)
268 spin_lock_irq(&mapping->tree_lock);
270 pslot = radix_tree_lookup_slot(&mapping->page_tree,
273 expected_count = 2 + !!page_has_private(page);
274 if (page_count(page) != expected_count ||
275 (struct page *)radix_tree_deref_slot(pslot) != page) {
276 spin_unlock_irq(&mapping->tree_lock);
280 if (!page_freeze_refs(page, expected_count)) {
281 spin_unlock_irq(&mapping->tree_lock);
286 * Now we know that no one else is looking at the page.
288 get_page(newpage); /* add cache reference */
289 if (PageSwapCache(page)) {
290 SetPageSwapCache(newpage);
291 set_page_private(newpage, page_private(page));
294 radix_tree_replace_slot(pslot, newpage);
296 page_unfreeze_refs(page, expected_count);
298 * Drop cache reference from old page.
299 * We know this isn't the last reference.
304 * If moved to a different zone then also account
305 * the page for that zone. Other VM counters will be
306 * taken care of when we establish references to the
307 * new page and drop references to the old page.
309 * Note that anonymous pages are accounted for
310 * via NR_FILE_PAGES and NR_ANON_PAGES if they
311 * are mapped to swap space.
313 __dec_zone_page_state(page, NR_FILE_PAGES);
314 __inc_zone_page_state(newpage, NR_FILE_PAGES);
315 if (PageSwapBacked(page)) {
316 __dec_zone_page_state(page, NR_SHMEM);
317 __inc_zone_page_state(newpage, NR_SHMEM);
319 spin_unlock_irq(&mapping->tree_lock);
325 * Copy the page to its new location
327 static void migrate_page_copy(struct page *newpage, struct page *page)
331 copy_highpage(newpage, page);
334 SetPageError(newpage);
335 if (PageReferenced(page))
336 SetPageReferenced(newpage);
337 if (PageUptodate(page))
338 SetPageUptodate(newpage);
339 if (TestClearPageActive(page)) {
340 VM_BUG_ON(PageUnevictable(page));
341 SetPageActive(newpage);
343 unevictable_migrate_page(newpage, page);
344 if (PageChecked(page))
345 SetPageChecked(newpage);
346 if (PageMappedToDisk(page))
347 SetPageMappedToDisk(newpage);
349 if (PageDirty(page)) {
350 clear_page_dirty_for_io(page);
352 * Want to mark the page and the radix tree as dirty, and
353 * redo the accounting that clear_page_dirty_for_io undid,
354 * but we can't use set_page_dirty because that function
355 * is actually a signal that all of the page has become dirty.
356 * Wheras only part of our page may be dirty.
358 __set_page_dirty_nobuffers(newpage);
361 mlock_migrate_page(newpage, page);
363 ClearPageSwapCache(page);
364 ClearPagePrivate(page);
365 set_page_private(page, 0);
366 /* page->mapping contains a flag for PageAnon() */
367 anon = PageAnon(page);
368 page->mapping = NULL;
371 * If any waiters have accumulated on the new page then
374 if (PageWriteback(newpage))
375 end_page_writeback(newpage);
378 /************************************************************
379 * Migration functions
380 ***********************************************************/
382 /* Always fail migration. Used for mappings that are not movable */
383 int fail_migrate_page(struct address_space *mapping,
384 struct page *newpage, struct page *page)
388 EXPORT_SYMBOL(fail_migrate_page);
391 * Common logic to directly migrate a single page suitable for
392 * pages that do not use PagePrivate/PagePrivate2.
394 * Pages are locked upon entry and exit.
396 int migrate_page(struct address_space *mapping,
397 struct page *newpage, struct page *page)
401 BUG_ON(PageWriteback(page)); /* Writeback must be complete */
403 rc = migrate_page_move_mapping(mapping, newpage, page);
408 migrate_page_copy(newpage, page);
411 EXPORT_SYMBOL(migrate_page);
415 * Migration function for pages with buffers. This function can only be used
416 * if the underlying filesystem guarantees that no other references to "page"
419 int buffer_migrate_page(struct address_space *mapping,
420 struct page *newpage, struct page *page)
422 struct buffer_head *bh, *head;
425 if (!page_has_buffers(page))
426 return migrate_page(mapping, newpage, page);
428 head = page_buffers(page);
430 rc = migrate_page_move_mapping(mapping, newpage, page);
439 bh = bh->b_this_page;
441 } while (bh != head);
443 ClearPagePrivate(page);
444 set_page_private(newpage, page_private(page));
445 set_page_private(page, 0);
451 set_bh_page(bh, newpage, bh_offset(bh));
452 bh = bh->b_this_page;
454 } while (bh != head);
456 SetPagePrivate(newpage);
458 migrate_page_copy(newpage, page);
464 bh = bh->b_this_page;
466 } while (bh != head);
470 EXPORT_SYMBOL(buffer_migrate_page);
474 * Writeback a page to clean the dirty state
476 static int writeout(struct address_space *mapping, struct page *page)
478 struct writeback_control wbc = {
479 .sync_mode = WB_SYNC_NONE,
482 .range_end = LLONG_MAX,
488 if (!mapping->a_ops->writepage)
489 /* No write method for the address space */
492 if (!clear_page_dirty_for_io(page))
493 /* Someone else already triggered a write */
497 * A dirty page may imply that the underlying filesystem has
498 * the page on some queue. So the page must be clean for
499 * migration. Writeout may mean we loose the lock and the
500 * page state is no longer what we checked for earlier.
501 * At this point we know that the migration attempt cannot
504 remove_migration_ptes(page, page);
506 rc = mapping->a_ops->writepage(page, &wbc);
508 if (rc != AOP_WRITEPAGE_ACTIVATE)
509 /* unlocked. Relock */
512 return (rc < 0) ? -EIO : -EAGAIN;
516 * Default handling if a filesystem does not provide a migration function.
518 static int fallback_migrate_page(struct address_space *mapping,
519 struct page *newpage, struct page *page)
522 return writeout(mapping, page);
525 * Buffers may be managed in a filesystem specific way.
526 * We must have no buffers or drop them.
528 if (page_has_private(page) &&
529 !try_to_release_page(page, GFP_KERNEL))
532 return migrate_page(mapping, newpage, page);
536 * Move a page to a newly allocated page
537 * The page is locked and all ptes have been successfully removed.
539 * The new page will have replaced the old page if this function
546 static int move_to_new_page(struct page *newpage, struct page *page)
548 struct address_space *mapping;
552 * Block others from accessing the page when we get around to
553 * establishing additional references. We are the only one
554 * holding a reference to the new page at this point.
556 if (!trylock_page(newpage))
559 /* Prepare mapping for the new page.*/
560 newpage->index = page->index;
561 newpage->mapping = page->mapping;
562 if (PageSwapBacked(page))
563 SetPageSwapBacked(newpage);
565 mapping = page_mapping(page);
567 rc = migrate_page(mapping, newpage, page);
568 else if (mapping->a_ops->migratepage)
570 * Most pages have a mapping and most filesystems
571 * should provide a migration function. Anonymous
572 * pages are part of swap space which also has its
573 * own migration function. This is the most common
574 * path for page migration.
576 rc = mapping->a_ops->migratepage(mapping,
579 rc = fallback_migrate_page(mapping, newpage, page);
582 remove_migration_ptes(page, newpage);
584 newpage->mapping = NULL;
586 unlock_page(newpage);
592 * Obtain the lock on page, remove all ptes and migrate the page
593 * to the newly allocated page in newpage.
595 static int unmap_and_move(new_page_t get_new_page, unsigned long private,
596 struct page *page, int force)
600 struct page *newpage = get_new_page(page, private, &result);
603 struct mem_cgroup *mem;
608 if (page_count(page) == 1) {
609 /* page was freed from under us. So we are done. */
613 /* prepare cgroup just returns 0 or -ENOMEM */
616 if (!trylock_page(page)) {
622 /* charge against new page */
623 charge = mem_cgroup_prepare_migration(page, &mem);
624 if (charge == -ENOMEM) {
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)) {
649 * Corner case handling:
650 * 1. When a new swap-cache page is read into, it is added to the LRU
651 * and treated as swapcache but it has no rmap yet.
652 * Calling try_to_unmap() against a page->mapping==NULL page will
653 * trigger a BUG. So handle it here.
654 * 2. An orphaned page (see truncate_complete_page) might have
655 * fs-private metadata. The page can be picked up due to memory
656 * offlining. Everywhere else except page reclaim, the page is
657 * invisible to the vm, so the page can not be migrated. So try to
658 * free the metadata, so the page can be freed.
660 if (!page->mapping) {
661 if (!PageAnon(page) && page_has_private(page)) {
663 * Go direct to try_to_free_buffers() here because
664 * a) that's what try_to_release_page() would do anyway
665 * b) we may be under rcu_read_lock() here, so we can't
666 * use GFP_KERNEL which is what try_to_release_page()
667 * needs to be effective.
669 try_to_free_buffers(page);
675 /* Establish migration ptes or remove ptes */
676 try_to_unmap(page, 1);
679 if (!page_mapped(page))
680 rc = move_to_new_page(newpage, page);
683 remove_migration_ptes(page, page);
689 mem_cgroup_end_migration(mem, page, newpage);
695 * A page that has been migrated has all references
696 * removed and will be freed. A page that has not been
697 * migrated will have kepts its references and be
700 list_del(&page->lru);
701 putback_lru_page(page);
707 * Move the new page to the LRU. If migration was not successful
708 * then this will free the page.
710 putback_lru_page(newpage);
716 *result = page_to_nid(newpage);
724 * The function takes one list of pages to migrate and a function
725 * that determines from the page to be migrated and the private data
726 * the target of the move and allocates the page.
728 * The function returns after 10 attempts or if no pages
729 * are movable anymore because to has become empty
730 * or no retryable pages exist anymore. All pages will be
731 * returned to the LRU or freed.
733 * Return: Number of pages not migrated or error code.
735 int migrate_pages(struct list_head *from,
736 new_page_t get_new_page, unsigned long private)
743 int swapwrite = current->flags & PF_SWAPWRITE;
747 current->flags |= PF_SWAPWRITE;
749 for(pass = 0; pass < 10 && retry; pass++) {
752 list_for_each_entry_safe(page, page2, from, lru) {
755 rc = unmap_and_move(get_new_page, private,
767 /* Permanent failure */
776 current->flags &= ~PF_SWAPWRITE;
778 putback_lru_pages(from);
783 return nr_failed + retry;
788 * Move a list of individual pages
790 struct page_to_node {
797 static struct page *new_page_node(struct page *p, unsigned long private,
800 struct page_to_node *pm = (struct page_to_node *)private;
802 while (pm->node != MAX_NUMNODES && pm->page != p)
805 if (pm->node == MAX_NUMNODES)
808 *result = &pm->status;
810 return alloc_pages_exact_node(pm->node,
811 GFP_HIGHUSER_MOVABLE | GFP_THISNODE, 0);
815 * Move a set of pages as indicated in the pm array. The addr
816 * field must be set to the virtual address of the page to be moved
817 * and the node number must contain a valid target node.
818 * The pm array ends with node = MAX_NUMNODES.
820 static int do_move_page_to_node_array(struct mm_struct *mm,
821 struct page_to_node *pm,
825 struct page_to_node *pp;
828 down_read(&mm->mmap_sem);
831 * Build a list of pages to migrate
833 for (pp = pm; pp->node != MAX_NUMNODES; pp++) {
834 struct vm_area_struct *vma;
838 vma = find_vma(mm, pp->addr);
839 if (!vma || !vma_migratable(vma))
842 page = follow_page(vma, pp->addr, FOLL_GET);
852 if (PageReserved(page)) /* Check for zero page */
856 err = page_to_nid(page);
860 * Node already in the right place
865 if (page_mapcount(page) > 1 &&
869 err = isolate_lru_page(page);
871 list_add_tail(&page->lru, &pagelist);
874 * Either remove the duplicate refcount from
875 * isolate_lru_page() or drop the page ref if it was
884 if (!list_empty(&pagelist))
885 err = migrate_pages(&pagelist, new_page_node,
888 up_read(&mm->mmap_sem);
893 * Migrate an array of page address onto an array of nodes and fill
894 * the corresponding array of status.
896 static int do_pages_move(struct mm_struct *mm, struct task_struct *task,
897 unsigned long nr_pages,
898 const void __user * __user *pages,
899 const int __user *nodes,
900 int __user *status, int flags)
902 struct page_to_node *pm;
903 nodemask_t task_nodes;
904 unsigned long chunk_nr_pages;
905 unsigned long chunk_start;
908 task_nodes = cpuset_mems_allowed(task);
911 pm = (struct page_to_node *)__get_free_page(GFP_KERNEL);
918 * Store a chunk of page_to_node array in a page,
919 * but keep the last one as a marker
921 chunk_nr_pages = (PAGE_SIZE / sizeof(struct page_to_node)) - 1;
923 for (chunk_start = 0;
924 chunk_start < nr_pages;
925 chunk_start += chunk_nr_pages) {
928 if (chunk_start + chunk_nr_pages > nr_pages)
929 chunk_nr_pages = nr_pages - chunk_start;
931 /* fill the chunk pm with addrs and nodes from user-space */
932 for (j = 0; j < chunk_nr_pages; j++) {
933 const void __user *p;
937 if (get_user(p, pages + j + chunk_start))
939 pm[j].addr = (unsigned long) p;
941 if (get_user(node, nodes + j + chunk_start))
945 if (!node_state(node, N_HIGH_MEMORY))
949 if (!node_isset(node, task_nodes))
955 /* End marker for this chunk */
956 pm[chunk_nr_pages].node = MAX_NUMNODES;
958 /* Migrate this chunk */
959 err = do_move_page_to_node_array(mm, pm,
960 flags & MPOL_MF_MOVE_ALL);
964 /* Return status information */
965 for (j = 0; j < chunk_nr_pages; j++)
966 if (put_user(pm[j].status, status + j + chunk_start)) {
974 free_page((unsigned long)pm);
980 * Determine the nodes of an array of pages and store it in an array of status.
982 static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages,
983 const void __user **pages, int *status)
987 down_read(&mm->mmap_sem);
989 for (i = 0; i < nr_pages; i++) {
990 unsigned long addr = (unsigned long)(*pages);
991 struct vm_area_struct *vma;
995 vma = find_vma(mm, addr);
999 page = follow_page(vma, addr, 0);
1001 err = PTR_ERR(page);
1006 /* Use PageReserved to check for zero page */
1007 if (!page || PageReserved(page))
1010 err = page_to_nid(page);
1018 up_read(&mm->mmap_sem);
1022 * Determine the nodes of a user array of pages and store it in
1023 * a user array of status.
1025 static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages,
1026 const void __user * __user *pages,
1029 #define DO_PAGES_STAT_CHUNK_NR 16
1030 const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR];
1031 int chunk_status[DO_PAGES_STAT_CHUNK_NR];
1032 unsigned long i, chunk_nr = DO_PAGES_STAT_CHUNK_NR;
1035 for (i = 0; i < nr_pages; i += chunk_nr) {
1036 if (chunk_nr + i > nr_pages)
1037 chunk_nr = nr_pages - i;
1039 err = copy_from_user(chunk_pages, &pages[i],
1040 chunk_nr * sizeof(*chunk_pages));
1046 do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status);
1048 err = copy_to_user(&status[i], chunk_status,
1049 chunk_nr * sizeof(*chunk_status));
1062 * Move a list of pages in the address space of the currently executing
1065 SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages,
1066 const void __user * __user *, pages,
1067 const int __user *, nodes,
1068 int __user *, status, int, flags)
1070 const struct cred *cred = current_cred(), *tcred;
1071 struct task_struct *task;
1072 struct mm_struct *mm;
1076 if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
1079 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1082 /* Find the mm_struct */
1083 read_lock(&tasklist_lock);
1084 task = pid ? find_task_by_vpid(pid) : current;
1086 read_unlock(&tasklist_lock);
1089 mm = get_task_mm(task);
1090 read_unlock(&tasklist_lock);
1096 * Check if this process has the right to modify the specified
1097 * process. The right exists if the process has administrative
1098 * capabilities, superuser privileges or the same
1099 * userid as the target process.
1102 tcred = __task_cred(task);
1103 if (cred->euid != tcred->suid && cred->euid != tcred->uid &&
1104 cred->uid != tcred->suid && cred->uid != tcred->uid &&
1105 !capable(CAP_SYS_NICE)) {
1112 err = security_task_movememory(task);
1117 err = do_pages_move(mm, task, nr_pages, pages, nodes, status,
1120 err = do_pages_stat(mm, nr_pages, pages, status);
1129 * Call migration functions in the vma_ops that may prepare
1130 * memory in a vm for migration. migration functions may perform
1131 * the migration for vmas that do not have an underlying page struct.
1133 int migrate_vmas(struct mm_struct *mm, const nodemask_t *to,
1134 const nodemask_t *from, unsigned long flags)
1136 struct vm_area_struct *vma;
1139 for (vma = mm->mmap; vma && !err; vma = vma->vm_next) {
1140 if (vma->vm_ops && vma->vm_ops->migrate) {
1141 err = vma->vm_ops->migrate(vma, to, from, flags);