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 * Isolate one page from the LRU lists. If successful put it onto
41 * the indicated list with elevated page count.
44 * -EBUSY: page not on LRU list
45 * 0: page removed from LRU list and added to the specified list.
47 int isolate_lru_page(struct page *page, struct list_head *pagelist)
52 struct zone *zone = page_zone(page);
54 spin_lock_irq(&zone->lru_lock);
55 if (PageLRU(page) && get_page_unless_zero(page)) {
59 del_page_from_active_list(zone, page);
61 del_page_from_inactive_list(zone, page);
62 list_add_tail(&page->lru, pagelist);
64 spin_unlock_irq(&zone->lru_lock);
70 * migrate_prep() needs to be called before we start compiling a list of pages
71 * to be migrated using isolate_lru_page().
73 int migrate_prep(void)
76 * Clear the LRU lists so pages can be isolated.
77 * Note that pages may be moved off the LRU after we have
78 * drained them. Those pages will fail to migrate like other
79 * pages that may be busy.
86 static inline void move_to_lru(struct page *page)
88 if (PageActive(page)) {
90 * lru_cache_add_active checks that
91 * the PG_active bit is off.
93 ClearPageActive(page);
94 lru_cache_add_active(page);
102 * Add isolated pages on the list back to the LRU.
104 * returns the number of pages put back.
106 int putback_lru_pages(struct list_head *l)
112 list_for_each_entry_safe(page, page2, l, lru) {
113 list_del(&page->lru);
121 * Restore a potential migration pte to a working pte entry
123 static void remove_migration_pte(struct vm_area_struct *vma,
124 struct page *old, struct page *new)
126 struct mm_struct *mm = vma->vm_mm;
133 unsigned long addr = page_address_in_vma(new, vma);
138 pgd = pgd_offset(mm, addr);
139 if (!pgd_present(*pgd))
142 pud = pud_offset(pgd, addr);
143 if (!pud_present(*pud))
146 pmd = pmd_offset(pud, addr);
147 if (!pmd_present(*pmd))
150 ptep = pte_offset_map(pmd, addr);
152 if (!is_swap_pte(*ptep)) {
157 ptl = pte_lockptr(mm, pmd);
160 if (!is_swap_pte(pte))
163 entry = pte_to_swp_entry(pte);
165 if (!is_migration_entry(entry) || migration_entry_to_page(entry) != old)
169 * Yes, ignore the return value from a GFP_ATOMIC mem_cgroup_charge.
170 * Failure is not an option here: we're now expected to remove every
171 * migration pte, and will cause crashes otherwise. Normally this
172 * is not an issue: mem_cgroup_prepare_migration bumped up the old
173 * page_cgroup count for safety, that's now attached to the new page,
174 * so this charge should just be another incrementation of the count,
175 * to keep in balance with rmap.c's mem_cgroup_uncharging. But if
176 * there's been a force_empty, those reference counts may no longer
177 * be reliable, and this charge can actually fail: oh well, we don't
178 * make the situation any worse by proceeding as if it had succeeded.
180 mem_cgroup_charge(new, mm, GFP_ATOMIC);
183 pte = pte_mkold(mk_pte(new, vma->vm_page_prot));
184 if (is_write_migration_entry(entry))
185 pte = pte_mkwrite(pte);
186 flush_cache_page(vma, addr, pte_pfn(pte));
187 set_pte_at(mm, addr, ptep, pte);
190 page_add_anon_rmap(new, vma, addr);
192 page_add_file_rmap(new);
194 /* No need to invalidate - it was non-present before */
195 update_mmu_cache(vma, addr, pte);
198 pte_unmap_unlock(ptep, ptl);
202 * Note that remove_file_migration_ptes will only work on regular mappings,
203 * Nonlinear mappings do not use migration entries.
205 static void remove_file_migration_ptes(struct page *old, struct page *new)
207 struct vm_area_struct *vma;
208 struct address_space *mapping = page_mapping(new);
209 struct prio_tree_iter iter;
210 pgoff_t pgoff = new->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
215 spin_lock(&mapping->i_mmap_lock);
217 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff)
218 remove_migration_pte(vma, old, new);
220 spin_unlock(&mapping->i_mmap_lock);
224 * Must hold mmap_sem lock on at least one of the vmas containing
225 * the page so that the anon_vma cannot vanish.
227 static void remove_anon_migration_ptes(struct page *old, struct page *new)
229 struct anon_vma *anon_vma;
230 struct vm_area_struct *vma;
231 unsigned long mapping;
233 mapping = (unsigned long)new->mapping;
235 if (!mapping || (mapping & PAGE_MAPPING_ANON) == 0)
239 * We hold the mmap_sem lock. So no need to call page_lock_anon_vma.
241 anon_vma = (struct anon_vma *) (mapping - PAGE_MAPPING_ANON);
242 spin_lock(&anon_vma->lock);
244 list_for_each_entry(vma, &anon_vma->head, anon_vma_node)
245 remove_migration_pte(vma, old, new);
247 spin_unlock(&anon_vma->lock);
251 * Get rid of all migration entries and replace them by
252 * references to the indicated page.
254 static void remove_migration_ptes(struct page *old, struct page *new)
257 remove_anon_migration_ptes(old, new);
259 remove_file_migration_ptes(old, new);
263 * Something used the pte of a page under migration. We need to
264 * get to the page and wait until migration is finished.
265 * When we return from this function the fault will be retried.
267 * This function is called from do_swap_page().
269 void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
270 unsigned long address)
277 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
279 if (!is_swap_pte(pte))
282 entry = pte_to_swp_entry(pte);
283 if (!is_migration_entry(entry))
286 page = migration_entry_to_page(entry);
289 pte_unmap_unlock(ptep, ptl);
290 wait_on_page_locked(page);
294 pte_unmap_unlock(ptep, ptl);
298 * Replace the page in the mapping.
300 * The number of remaining references must be:
301 * 1 for anonymous pages without a mapping
302 * 2 for pages with a mapping
303 * 3 for pages with a mapping and PagePrivate set.
305 static int migrate_page_move_mapping(struct address_space *mapping,
306 struct page *newpage, struct page *page)
311 /* Anonymous page without mapping */
312 if (page_count(page) != 1)
317 write_lock_irq(&mapping->tree_lock);
319 pslot = radix_tree_lookup_slot(&mapping->page_tree,
322 if (page_count(page) != 2 + !!PagePrivate(page) ||
323 (struct page *)radix_tree_deref_slot(pslot) != page) {
324 write_unlock_irq(&mapping->tree_lock);
329 * Now we know that no one else is looking at the page.
331 get_page(newpage); /* add cache reference */
333 if (PageSwapCache(page)) {
334 SetPageSwapCache(newpage);
335 set_page_private(newpage, page_private(page));
339 radix_tree_replace_slot(pslot, newpage);
342 * Drop cache reference from old page.
343 * We know this isn't the last reference.
348 * If moved to a different zone then also account
349 * the page for that zone. Other VM counters will be
350 * taken care of when we establish references to the
351 * new page and drop references to the old page.
353 * Note that anonymous pages are accounted for
354 * via NR_FILE_PAGES and NR_ANON_PAGES if they
355 * are mapped to swap space.
357 __dec_zone_page_state(page, NR_FILE_PAGES);
358 __inc_zone_page_state(newpage, NR_FILE_PAGES);
360 write_unlock_irq(&mapping->tree_lock);
361 if (!PageSwapCache(newpage)) {
362 mem_cgroup_uncharge_page(page);
363 mem_cgroup_getref(newpage);
370 * Copy the page to its new location
372 static void migrate_page_copy(struct page *newpage, struct page *page)
374 copy_highpage(newpage, page);
377 SetPageError(newpage);
378 if (PageReferenced(page))
379 SetPageReferenced(newpage);
380 if (PageUptodate(page))
381 SetPageUptodate(newpage);
382 if (PageActive(page))
383 SetPageActive(newpage);
384 if (PageChecked(page))
385 SetPageChecked(newpage);
386 if (PageMappedToDisk(page))
387 SetPageMappedToDisk(newpage);
389 if (PageDirty(page)) {
390 clear_page_dirty_for_io(page);
392 * Want to mark the page and the radix tree as dirty, and
393 * redo the accounting that clear_page_dirty_for_io undid,
394 * but we can't use set_page_dirty because that function
395 * is actually a signal that all of the page has become dirty.
396 * Wheras only part of our page may be dirty.
398 __set_page_dirty_nobuffers(newpage);
402 ClearPageSwapCache(page);
404 ClearPageActive(page);
405 ClearPagePrivate(page);
406 set_page_private(page, 0);
407 page->mapping = NULL;
410 * If any waiters have accumulated on the new page then
413 if (PageWriteback(newpage))
414 end_page_writeback(newpage);
417 /************************************************************
418 * Migration functions
419 ***********************************************************/
421 /* Always fail migration. Used for mappings that are not movable */
422 int fail_migrate_page(struct address_space *mapping,
423 struct page *newpage, struct page *page)
427 EXPORT_SYMBOL(fail_migrate_page);
430 * Common logic to directly migrate a single page suitable for
431 * pages that do not use PagePrivate.
433 * Pages are locked upon entry and exit.
435 int migrate_page(struct address_space *mapping,
436 struct page *newpage, struct page *page)
440 BUG_ON(PageWriteback(page)); /* Writeback must be complete */
442 rc = migrate_page_move_mapping(mapping, newpage, page);
447 migrate_page_copy(newpage, page);
450 EXPORT_SYMBOL(migrate_page);
454 * Migration function for pages with buffers. This function can only be used
455 * if the underlying filesystem guarantees that no other references to "page"
458 int buffer_migrate_page(struct address_space *mapping,
459 struct page *newpage, struct page *page)
461 struct buffer_head *bh, *head;
464 if (!page_has_buffers(page))
465 return migrate_page(mapping, newpage, page);
467 head = page_buffers(page);
469 rc = migrate_page_move_mapping(mapping, newpage, page);
478 bh = bh->b_this_page;
480 } while (bh != head);
482 ClearPagePrivate(page);
483 set_page_private(newpage, page_private(page));
484 set_page_private(page, 0);
490 set_bh_page(bh, newpage, bh_offset(bh));
491 bh = bh->b_this_page;
493 } while (bh != head);
495 SetPagePrivate(newpage);
497 migrate_page_copy(newpage, page);
503 bh = bh->b_this_page;
505 } while (bh != head);
509 EXPORT_SYMBOL(buffer_migrate_page);
513 * Writeback a page to clean the dirty state
515 static int writeout(struct address_space *mapping, struct page *page)
517 struct writeback_control wbc = {
518 .sync_mode = WB_SYNC_NONE,
521 .range_end = LLONG_MAX,
527 if (!mapping->a_ops->writepage)
528 /* No write method for the address space */
531 if (!clear_page_dirty_for_io(page))
532 /* Someone else already triggered a write */
536 * A dirty page may imply that the underlying filesystem has
537 * the page on some queue. So the page must be clean for
538 * migration. Writeout may mean we loose the lock and the
539 * page state is no longer what we checked for earlier.
540 * At this point we know that the migration attempt cannot
543 remove_migration_ptes(page, page);
545 rc = mapping->a_ops->writepage(page, &wbc);
547 /* I/O Error writing */
550 if (rc != AOP_WRITEPAGE_ACTIVATE)
551 /* unlocked. Relock */
558 * Default handling if a filesystem does not provide a migration function.
560 static int fallback_migrate_page(struct address_space *mapping,
561 struct page *newpage, struct page *page)
564 return writeout(mapping, page);
567 * Buffers may be managed in a filesystem specific way.
568 * We must have no buffers or drop them.
570 if (PagePrivate(page) &&
571 !try_to_release_page(page, GFP_KERNEL))
574 return migrate_page(mapping, newpage, page);
578 * Move a page to a newly allocated page
579 * The page is locked and all ptes have been successfully removed.
581 * The new page will have replaced the old page if this function
584 static int move_to_new_page(struct page *newpage, struct page *page)
586 struct address_space *mapping;
590 * Block others from accessing the page when we get around to
591 * establishing additional references. We are the only one
592 * holding a reference to the new page at this point.
594 if (TestSetPageLocked(newpage))
597 /* Prepare mapping for the new page.*/
598 newpage->index = page->index;
599 newpage->mapping = page->mapping;
601 mapping = page_mapping(page);
603 rc = migrate_page(mapping, newpage, page);
604 else if (mapping->a_ops->migratepage)
606 * Most pages have a mapping and most filesystems
607 * should provide a migration function. Anonymous
608 * pages are part of swap space which also has its
609 * own migration function. This is the most common
610 * path for page migration.
612 rc = mapping->a_ops->migratepage(mapping,
615 rc = fallback_migrate_page(mapping, newpage, page);
618 remove_migration_ptes(page, newpage);
620 newpage->mapping = NULL;
622 unlock_page(newpage);
628 * Obtain the lock on page, remove all ptes and migrate the page
629 * to the newly allocated page in newpage.
631 static int unmap_and_move(new_page_t get_new_page, unsigned long private,
632 struct page *page, int force)
636 struct page *newpage = get_new_page(page, private, &result);
643 if (page_count(page) == 1)
644 /* page was freed from under us. So we are done. */
647 charge = mem_cgroup_prepare_migration(page, newpage);
648 if (charge == -ENOMEM) {
652 /* prepare cgroup just returns 0 or -ENOMEM */
656 if (TestSetPageLocked(page)) {
662 if (PageWriteback(page)) {
665 wait_on_page_writeback(page);
668 * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
669 * we cannot notice that anon_vma is freed while we migrates a page.
670 * This rcu_read_lock() delays freeing anon_vma pointer until the end
671 * of migration. File cache pages are no problem because of page_lock()
672 * File Caches may use write_page() or lock_page() in migration, then,
673 * just care Anon page here.
675 if (PageAnon(page)) {
681 * Corner case handling:
682 * 1. When a new swap-cache page is read into, it is added to the LRU
683 * and treated as swapcache but it has no rmap yet.
684 * Calling try_to_unmap() against a page->mapping==NULL page will
685 * trigger a BUG. So handle it here.
686 * 2. An orphaned page (see truncate_complete_page) might have
687 * fs-private metadata. The page can be picked up due to memory
688 * offlining. Everywhere else except page reclaim, the page is
689 * invisible to the vm, so the page can not be migrated. So try to
690 * free the metadata, so the page can be freed.
692 if (!page->mapping) {
693 if (!PageAnon(page) && PagePrivate(page)) {
695 * Go direct to try_to_free_buffers() here because
696 * a) that's what try_to_release_page() would do anyway
697 * b) we may be under rcu_read_lock() here, so we can't
698 * use GFP_KERNEL which is what try_to_release_page()
699 * needs to be effective.
701 try_to_free_buffers(page);
706 /* Establish migration ptes or remove ptes */
707 try_to_unmap(page, 1);
709 if (!page_mapped(page))
710 rc = move_to_new_page(newpage, page);
713 remove_migration_ptes(page, page);
724 * A page that has been migrated has all references
725 * removed and will be freed. A page that has not been
726 * migrated will have kepts its references and be
729 list_del(&page->lru);
735 mem_cgroup_end_migration(newpage);
737 * Move the new page to the LRU. If migration was not successful
738 * then this will free the page.
740 move_to_lru(newpage);
745 *result = page_to_nid(newpage);
753 * The function takes one list of pages to migrate and a function
754 * that determines from the page to be migrated and the private data
755 * the target of the move and allocates the page.
757 * The function returns after 10 attempts or if no pages
758 * are movable anymore because to has become empty
759 * or no retryable pages exist anymore. All pages will be
760 * returned to the LRU or freed.
762 * Return: Number of pages not migrated or error code.
764 int migrate_pages(struct list_head *from,
765 new_page_t get_new_page, unsigned long private)
772 int swapwrite = current->flags & PF_SWAPWRITE;
776 current->flags |= PF_SWAPWRITE;
778 for(pass = 0; pass < 10 && retry; pass++) {
781 list_for_each_entry_safe(page, page2, from, lru) {
784 rc = unmap_and_move(get_new_page, private,
796 /* Permanent failure */
805 current->flags &= ~PF_SWAPWRITE;
807 putback_lru_pages(from);
812 return nr_failed + retry;
817 * Move a list of individual pages
819 struct page_to_node {
826 static struct page *new_page_node(struct page *p, unsigned long private,
829 struct page_to_node *pm = (struct page_to_node *)private;
831 while (pm->node != MAX_NUMNODES && pm->page != p)
834 if (pm->node == MAX_NUMNODES)
837 *result = &pm->status;
839 return alloc_pages_node(pm->node,
840 GFP_HIGHUSER_MOVABLE | GFP_THISNODE, 0);
844 * Move a set of pages as indicated in the pm array. The addr
845 * field must be set to the virtual address of the page to be moved
846 * and the node number must contain a valid target node.
848 static int do_move_pages(struct mm_struct *mm, struct page_to_node *pm,
852 struct page_to_node *pp;
855 down_read(&mm->mmap_sem);
858 * Build a list of pages to migrate
861 for (pp = pm; pp->node != MAX_NUMNODES; pp++) {
862 struct vm_area_struct *vma;
866 * A valid page pointer that will not match any of the
867 * pages that will be moved.
869 pp->page = ZERO_PAGE(0);
872 vma = find_vma(mm, pp->addr);
873 if (!vma || !vma_migratable(vma))
876 page = follow_page(vma, pp->addr, FOLL_GET);
886 if (PageReserved(page)) /* Check for zero page */
890 err = page_to_nid(page);
894 * Node already in the right place
899 if (page_mapcount(page) > 1 &&
903 err = isolate_lru_page(page, &pagelist);
906 * Either remove the duplicate refcount from
907 * isolate_lru_page() or drop the page ref if it was
915 if (!list_empty(&pagelist))
916 err = migrate_pages(&pagelist, new_page_node,
921 up_read(&mm->mmap_sem);
926 * Determine the nodes of a list of pages. The addr in the pm array
927 * must have been set to the virtual address of which we want to determine
930 static int do_pages_stat(struct mm_struct *mm, struct page_to_node *pm)
932 down_read(&mm->mmap_sem);
934 for ( ; pm->node != MAX_NUMNODES; pm++) {
935 struct vm_area_struct *vma;
940 vma = find_vma(mm, pm->addr);
944 page = follow_page(vma, pm->addr, 0);
951 /* Use PageReserved to check for zero page */
952 if (!page || PageReserved(page))
955 err = page_to_nid(page);
960 up_read(&mm->mmap_sem);
965 * Move a list of pages in the address space of the currently executing
968 asmlinkage long sys_move_pages(pid_t pid, unsigned long nr_pages,
969 const void __user * __user *pages,
970 const int __user *nodes,
971 int __user *status, int flags)
975 struct task_struct *task;
976 nodemask_t task_nodes;
977 struct mm_struct *mm;
978 struct page_to_node *pm = NULL;
981 if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
984 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
987 /* Find the mm_struct */
988 read_lock(&tasklist_lock);
989 task = pid ? find_task_by_vpid(pid) : current;
991 read_unlock(&tasklist_lock);
994 mm = get_task_mm(task);
995 read_unlock(&tasklist_lock);
1001 * Check if this process has the right to modify the specified
1002 * process. The right exists if the process has administrative
1003 * capabilities, superuser privileges or the same
1004 * userid as the target process.
1006 if ((current->euid != task->suid) && (current->euid != task->uid) &&
1007 (current->uid != task->suid) && (current->uid != task->uid) &&
1008 !capable(CAP_SYS_NICE)) {
1013 err = security_task_movememory(task);
1018 task_nodes = cpuset_mems_allowed(task);
1020 /* Limit nr_pages so that the multiplication may not overflow */
1021 if (nr_pages >= ULONG_MAX / sizeof(struct page_to_node) - 1) {
1026 pm = vmalloc((nr_pages + 1) * sizeof(struct page_to_node));
1033 * Get parameters from user space and initialize the pm
1034 * array. Return various errors if the user did something wrong.
1036 for (i = 0; i < nr_pages; i++) {
1037 const void __user *p;
1040 if (get_user(p, pages + i))
1043 pm[i].addr = (unsigned long)p;
1047 if (get_user(node, nodes + i))
1051 if (!node_state(node, N_HIGH_MEMORY))
1055 if (!node_isset(node, task_nodes))
1060 pm[i].node = 0; /* anything to not match MAX_NUMNODES */
1063 pm[nr_pages].node = MAX_NUMNODES;
1066 err = do_move_pages(mm, pm, flags & MPOL_MF_MOVE_ALL);
1068 err = do_pages_stat(mm, pm);
1071 /* Return status information */
1072 for (i = 0; i < nr_pages; i++)
1073 if (put_user(pm[i].status, status + i))
1084 * Call migration functions in the vma_ops that may prepare
1085 * memory in a vm for migration. migration functions may perform
1086 * the migration for vmas that do not have an underlying page struct.
1088 int migrate_vmas(struct mm_struct *mm, const nodemask_t *to,
1089 const nodemask_t *from, unsigned long flags)
1091 struct vm_area_struct *vma;
1094 for(vma = mm->mmap; vma->vm_next && !err; vma = vma->vm_next) {
1095 if (vma->vm_ops && vma->vm_ops->migrate) {
1096 err = vma->vm_ops->migrate(vma, to, from, flags);