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.
56 static inline void move_to_lru(struct page *page)
58 if (PageActive(page)) {
60 * lru_cache_add_active checks that
61 * the PG_active bit is off.
63 ClearPageActive(page);
64 lru_cache_add_active(page);
72 * Add isolated pages on the list back to the LRU.
74 * returns the number of pages put back.
76 int putback_lru_pages(struct list_head *l)
82 list_for_each_entry_safe(page, page2, l, lru) {
91 * Restore a potential migration pte to a working pte entry
93 static void remove_migration_pte(struct vm_area_struct *vma,
94 struct page *old, struct page *new)
96 struct mm_struct *mm = vma->vm_mm;
103 unsigned long addr = page_address_in_vma(new, vma);
108 pgd = pgd_offset(mm, addr);
109 if (!pgd_present(*pgd))
112 pud = pud_offset(pgd, addr);
113 if (!pud_present(*pud))
116 pmd = pmd_offset(pud, addr);
117 if (!pmd_present(*pmd))
120 ptep = pte_offset_map(pmd, addr);
122 if (!is_swap_pte(*ptep)) {
127 ptl = pte_lockptr(mm, pmd);
130 if (!is_swap_pte(pte))
133 entry = pte_to_swp_entry(pte);
135 if (!is_migration_entry(entry) || migration_entry_to_page(entry) != old)
139 * Yes, ignore the return value from a GFP_ATOMIC mem_cgroup_charge.
140 * Failure is not an option here: we're now expected to remove every
141 * migration pte, and will cause crashes otherwise. Normally this
142 * is not an issue: mem_cgroup_prepare_migration bumped up the old
143 * page_cgroup count for safety, that's now attached to the new page,
144 * so this charge should just be another incrementation of the count,
145 * to keep in balance with rmap.c's mem_cgroup_uncharging. But if
146 * there's been a force_empty, those reference counts may no longer
147 * be reliable, and this charge can actually fail: oh well, we don't
148 * make the situation any worse by proceeding as if it had succeeded.
150 mem_cgroup_charge(new, mm, GFP_ATOMIC);
153 pte = pte_mkold(mk_pte(new, vma->vm_page_prot));
154 if (is_write_migration_entry(entry))
155 pte = pte_mkwrite(pte);
156 flush_cache_page(vma, addr, pte_pfn(pte));
157 set_pte_at(mm, addr, ptep, pte);
160 page_add_anon_rmap(new, vma, addr);
162 page_add_file_rmap(new);
164 /* No need to invalidate - it was non-present before */
165 update_mmu_cache(vma, addr, pte);
168 pte_unmap_unlock(ptep, ptl);
172 * Note that remove_file_migration_ptes will only work on regular mappings,
173 * Nonlinear mappings do not use migration entries.
175 static void remove_file_migration_ptes(struct page *old, struct page *new)
177 struct vm_area_struct *vma;
178 struct address_space *mapping = page_mapping(new);
179 struct prio_tree_iter iter;
180 pgoff_t pgoff = new->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
185 spin_lock(&mapping->i_mmap_lock);
187 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff)
188 remove_migration_pte(vma, old, new);
190 spin_unlock(&mapping->i_mmap_lock);
194 * Must hold mmap_sem lock on at least one of the vmas containing
195 * the page so that the anon_vma cannot vanish.
197 static void remove_anon_migration_ptes(struct page *old, struct page *new)
199 struct anon_vma *anon_vma;
200 struct vm_area_struct *vma;
201 unsigned long mapping;
203 mapping = (unsigned long)new->mapping;
205 if (!mapping || (mapping & PAGE_MAPPING_ANON) == 0)
209 * We hold the mmap_sem lock. So no need to call page_lock_anon_vma.
211 anon_vma = (struct anon_vma *) (mapping - PAGE_MAPPING_ANON);
212 spin_lock(&anon_vma->lock);
214 list_for_each_entry(vma, &anon_vma->head, anon_vma_node)
215 remove_migration_pte(vma, old, new);
217 spin_unlock(&anon_vma->lock);
221 * Get rid of all migration entries and replace them by
222 * references to the indicated page.
224 static void remove_migration_ptes(struct page *old, struct page *new)
227 remove_anon_migration_ptes(old, new);
229 remove_file_migration_ptes(old, new);
233 * Something used the pte of a page under migration. We need to
234 * get to the page and wait until migration is finished.
235 * When we return from this function the fault will be retried.
237 * This function is called from do_swap_page().
239 void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
240 unsigned long address)
247 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
249 if (!is_swap_pte(pte))
252 entry = pte_to_swp_entry(pte);
253 if (!is_migration_entry(entry))
256 page = migration_entry_to_page(entry);
259 * Once radix-tree replacement of page migration started, page_count
260 * *must* be zero. And, we don't want to call wait_on_page_locked()
261 * against a page without get_page().
262 * So, we use get_page_unless_zero(), here. Even failed, page fault
265 if (!get_page_unless_zero(page))
267 pte_unmap_unlock(ptep, ptl);
268 wait_on_page_locked(page);
272 pte_unmap_unlock(ptep, ptl);
276 * Replace the page in the mapping.
278 * The number of remaining references must be:
279 * 1 for anonymous pages without a mapping
280 * 2 for pages with a mapping
281 * 3 for pages with a mapping and PagePrivate set.
283 static int migrate_page_move_mapping(struct address_space *mapping,
284 struct page *newpage, struct page *page)
290 /* Anonymous page without mapping */
291 if (page_count(page) != 1)
296 spin_lock_irq(&mapping->tree_lock);
298 pslot = radix_tree_lookup_slot(&mapping->page_tree,
301 expected_count = 2 + !!PagePrivate(page);
302 if (page_count(page) != expected_count ||
303 (struct page *)radix_tree_deref_slot(pslot) != page) {
304 spin_unlock_irq(&mapping->tree_lock);
308 if (!page_freeze_refs(page, expected_count)) {
309 spin_unlock_irq(&mapping->tree_lock);
314 * Now we know that no one else is looking at the page.
316 get_page(newpage); /* add cache reference */
318 if (PageSwapCache(page)) {
319 SetPageSwapCache(newpage);
320 set_page_private(newpage, page_private(page));
324 radix_tree_replace_slot(pslot, newpage);
326 page_unfreeze_refs(page, expected_count);
328 * Drop cache reference from old page.
329 * We know this isn't the last reference.
334 * If moved to a different zone then also account
335 * the page for that zone. Other VM counters will be
336 * taken care of when we establish references to the
337 * new page and drop references to the old page.
339 * Note that anonymous pages are accounted for
340 * via NR_FILE_PAGES and NR_ANON_PAGES if they
341 * are mapped to swap space.
343 __dec_zone_page_state(page, NR_FILE_PAGES);
344 __inc_zone_page_state(newpage, NR_FILE_PAGES);
346 spin_unlock_irq(&mapping->tree_lock);
347 if (!PageSwapCache(newpage))
348 mem_cgroup_uncharge_cache_page(page);
354 * Copy the page to its new location
356 static void migrate_page_copy(struct page *newpage, struct page *page)
358 copy_highpage(newpage, page);
361 SetPageError(newpage);
362 if (PageReferenced(page))
363 SetPageReferenced(newpage);
364 if (PageUptodate(page))
365 SetPageUptodate(newpage);
366 if (PageActive(page))
367 SetPageActive(newpage);
368 if (PageChecked(page))
369 SetPageChecked(newpage);
370 if (PageMappedToDisk(page))
371 SetPageMappedToDisk(newpage);
373 if (PageDirty(page)) {
374 clear_page_dirty_for_io(page);
376 * Want to mark the page and the radix tree as dirty, and
377 * redo the accounting that clear_page_dirty_for_io undid,
378 * but we can't use set_page_dirty because that function
379 * is actually a signal that all of the page has become dirty.
380 * Wheras only part of our page may be dirty.
382 __set_page_dirty_nobuffers(newpage);
386 ClearPageSwapCache(page);
388 ClearPageActive(page);
389 ClearPagePrivate(page);
390 set_page_private(page, 0);
391 page->mapping = NULL;
394 * If any waiters have accumulated on the new page then
397 if (PageWriteback(newpage))
398 end_page_writeback(newpage);
401 /************************************************************
402 * Migration functions
403 ***********************************************************/
405 /* Always fail migration. Used for mappings that are not movable */
406 int fail_migrate_page(struct address_space *mapping,
407 struct page *newpage, struct page *page)
411 EXPORT_SYMBOL(fail_migrate_page);
414 * Common logic to directly migrate a single page suitable for
415 * pages that do not use PagePrivate.
417 * Pages are locked upon entry and exit.
419 int migrate_page(struct address_space *mapping,
420 struct page *newpage, struct page *page)
424 BUG_ON(PageWriteback(page)); /* Writeback must be complete */
426 rc = migrate_page_move_mapping(mapping, newpage, page);
431 migrate_page_copy(newpage, page);
434 EXPORT_SYMBOL(migrate_page);
438 * Migration function for pages with buffers. This function can only be used
439 * if the underlying filesystem guarantees that no other references to "page"
442 int buffer_migrate_page(struct address_space *mapping,
443 struct page *newpage, struct page *page)
445 struct buffer_head *bh, *head;
448 if (!page_has_buffers(page))
449 return migrate_page(mapping, newpage, page);
451 head = page_buffers(page);
453 rc = migrate_page_move_mapping(mapping, newpage, page);
462 bh = bh->b_this_page;
464 } while (bh != head);
466 ClearPagePrivate(page);
467 set_page_private(newpage, page_private(page));
468 set_page_private(page, 0);
474 set_bh_page(bh, newpage, bh_offset(bh));
475 bh = bh->b_this_page;
477 } while (bh != head);
479 SetPagePrivate(newpage);
481 migrate_page_copy(newpage, page);
487 bh = bh->b_this_page;
489 } while (bh != head);
493 EXPORT_SYMBOL(buffer_migrate_page);
497 * Writeback a page to clean the dirty state
499 static int writeout(struct address_space *mapping, struct page *page)
501 struct writeback_control wbc = {
502 .sync_mode = WB_SYNC_NONE,
505 .range_end = LLONG_MAX,
511 if (!mapping->a_ops->writepage)
512 /* No write method for the address space */
515 if (!clear_page_dirty_for_io(page))
516 /* Someone else already triggered a write */
520 * A dirty page may imply that the underlying filesystem has
521 * the page on some queue. So the page must be clean for
522 * migration. Writeout may mean we loose the lock and the
523 * page state is no longer what we checked for earlier.
524 * At this point we know that the migration attempt cannot
527 remove_migration_ptes(page, page);
529 rc = mapping->a_ops->writepage(page, &wbc);
531 /* I/O Error writing */
534 if (rc != AOP_WRITEPAGE_ACTIVATE)
535 /* unlocked. Relock */
542 * Default handling if a filesystem does not provide a migration function.
544 static int fallback_migrate_page(struct address_space *mapping,
545 struct page *newpage, struct page *page)
548 return writeout(mapping, page);
551 * Buffers may be managed in a filesystem specific way.
552 * We must have no buffers or drop them.
554 if (PagePrivate(page) &&
555 !try_to_release_page(page, GFP_KERNEL))
558 return migrate_page(mapping, newpage, page);
562 * Move a page to a newly allocated page
563 * The page is locked and all ptes have been successfully removed.
565 * The new page will have replaced the old page if this function
568 static int move_to_new_page(struct page *newpage, struct page *page)
570 struct address_space *mapping;
574 * Block others from accessing the page when we get around to
575 * establishing additional references. We are the only one
576 * holding a reference to the new page at this point.
578 if (!trylock_page(newpage))
581 /* Prepare mapping for the new page.*/
582 newpage->index = page->index;
583 newpage->mapping = page->mapping;
585 mapping = page_mapping(page);
587 rc = migrate_page(mapping, newpage, page);
588 else if (mapping->a_ops->migratepage)
590 * Most pages have a mapping and most filesystems
591 * should provide a migration function. Anonymous
592 * pages are part of swap space which also has its
593 * own migration function. This is the most common
594 * path for page migration.
596 rc = mapping->a_ops->migratepage(mapping,
599 rc = fallback_migrate_page(mapping, newpage, page);
602 remove_migration_ptes(page, newpage);
604 newpage->mapping = NULL;
606 unlock_page(newpage);
612 * Obtain the lock on page, remove all ptes and migrate the page
613 * to the newly allocated page in newpage.
615 static int unmap_and_move(new_page_t get_new_page, unsigned long private,
616 struct page *page, int force)
620 struct page *newpage = get_new_page(page, private, &result);
627 if (page_count(page) == 1)
628 /* page was freed from under us. So we are done. */
631 charge = mem_cgroup_prepare_migration(page, newpage);
632 if (charge == -ENOMEM) {
636 /* prepare cgroup just returns 0 or -ENOMEM */
640 if (!trylock_page(page)) {
646 if (PageWriteback(page)) {
649 wait_on_page_writeback(page);
652 * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
653 * we cannot notice that anon_vma is freed while we migrates a page.
654 * This rcu_read_lock() delays freeing anon_vma pointer until the end
655 * of migration. File cache pages are no problem because of page_lock()
656 * File Caches may use write_page() or lock_page() in migration, then,
657 * just care Anon page here.
659 if (PageAnon(page)) {
665 * Corner case handling:
666 * 1. When a new swap-cache page is read into, it is added to the LRU
667 * and treated as swapcache but it has no rmap yet.
668 * Calling try_to_unmap() against a page->mapping==NULL page will
669 * trigger a BUG. So handle it here.
670 * 2. An orphaned page (see truncate_complete_page) might have
671 * fs-private metadata. The page can be picked up due to memory
672 * offlining. Everywhere else except page reclaim, the page is
673 * invisible to the vm, so the page can not be migrated. So try to
674 * free the metadata, so the page can be freed.
676 if (!page->mapping) {
677 if (!PageAnon(page) && PagePrivate(page)) {
679 * Go direct to try_to_free_buffers() here because
680 * a) that's what try_to_release_page() would do anyway
681 * b) we may be under rcu_read_lock() here, so we can't
682 * use GFP_KERNEL which is what try_to_release_page()
683 * needs to be effective.
685 try_to_free_buffers(page);
690 /* Establish migration ptes or remove ptes */
691 try_to_unmap(page, 1);
693 if (!page_mapped(page))
694 rc = move_to_new_page(newpage, page);
697 remove_migration_ptes(page, page);
708 * A page that has been migrated has all references
709 * removed and will be freed. A page that has not been
710 * migrated will have kepts its references and be
713 list_del(&page->lru);
719 mem_cgroup_end_migration(newpage);
721 * Move the new page to the LRU. If migration was not successful
722 * then this will free the page.
724 move_to_lru(newpage);
729 *result = page_to_nid(newpage);
737 * The function takes one list of pages to migrate and a function
738 * that determines from the page to be migrated and the private data
739 * the target of the move and allocates the page.
741 * The function returns after 10 attempts or if no pages
742 * are movable anymore because to has become empty
743 * or no retryable pages exist anymore. All pages will be
744 * returned to the LRU or freed.
746 * Return: Number of pages not migrated or error code.
748 int migrate_pages(struct list_head *from,
749 new_page_t get_new_page, unsigned long private)
756 int swapwrite = current->flags & PF_SWAPWRITE;
760 current->flags |= PF_SWAPWRITE;
762 for(pass = 0; pass < 10 && retry; pass++) {
765 list_for_each_entry_safe(page, page2, from, lru) {
768 rc = unmap_and_move(get_new_page, private,
780 /* Permanent failure */
789 current->flags &= ~PF_SWAPWRITE;
791 putback_lru_pages(from);
796 return nr_failed + retry;
801 * Move a list of individual pages
803 struct page_to_node {
810 static struct page *new_page_node(struct page *p, unsigned long private,
813 struct page_to_node *pm = (struct page_to_node *)private;
815 while (pm->node != MAX_NUMNODES && pm->page != p)
818 if (pm->node == MAX_NUMNODES)
821 *result = &pm->status;
823 return alloc_pages_node(pm->node,
824 GFP_HIGHUSER_MOVABLE | GFP_THISNODE, 0);
828 * Move a set of pages as indicated in the pm array. The addr
829 * field must be set to the virtual address of the page to be moved
830 * and the node number must contain a valid target node.
832 static int do_move_pages(struct mm_struct *mm, struct page_to_node *pm,
836 struct page_to_node *pp;
839 down_read(&mm->mmap_sem);
842 * Build a list of pages to migrate
845 for (pp = pm; pp->node != MAX_NUMNODES; pp++) {
846 struct vm_area_struct *vma;
850 * A valid page pointer that will not match any of the
851 * pages that will be moved.
853 pp->page = ZERO_PAGE(0);
856 vma = find_vma(mm, pp->addr);
857 if (!vma || !vma_migratable(vma))
860 page = follow_page(vma, pp->addr, FOLL_GET);
870 if (PageReserved(page)) /* Check for zero page */
874 err = page_to_nid(page);
878 * Node already in the right place
883 if (page_mapcount(page) > 1 &&
887 err = isolate_lru_page(page);
889 list_add_tail(&page->lru, &pagelist);
892 * Either remove the duplicate refcount from
893 * isolate_lru_page() or drop the page ref if it was
901 if (!list_empty(&pagelist))
902 err = migrate_pages(&pagelist, new_page_node,
907 up_read(&mm->mmap_sem);
912 * Determine the nodes of a list of pages. The addr in the pm array
913 * must have been set to the virtual address of which we want to determine
916 static int do_pages_stat(struct mm_struct *mm, struct page_to_node *pm)
918 down_read(&mm->mmap_sem);
920 for ( ; pm->node != MAX_NUMNODES; pm++) {
921 struct vm_area_struct *vma;
926 vma = find_vma(mm, pm->addr);
930 page = follow_page(vma, pm->addr, 0);
937 /* Use PageReserved to check for zero page */
938 if (!page || PageReserved(page))
941 err = page_to_nid(page);
946 up_read(&mm->mmap_sem);
951 * Move a list of pages in the address space of the currently executing
954 asmlinkage long sys_move_pages(pid_t pid, unsigned long nr_pages,
955 const void __user * __user *pages,
956 const int __user *nodes,
957 int __user *status, int flags)
961 struct task_struct *task;
962 nodemask_t task_nodes;
963 struct mm_struct *mm;
964 struct page_to_node *pm = NULL;
967 if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
970 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
973 /* Find the mm_struct */
974 read_lock(&tasklist_lock);
975 task = pid ? find_task_by_vpid(pid) : current;
977 read_unlock(&tasklist_lock);
980 mm = get_task_mm(task);
981 read_unlock(&tasklist_lock);
987 * Check if this process has the right to modify the specified
988 * process. The right exists if the process has administrative
989 * capabilities, superuser privileges or the same
990 * userid as the target process.
992 if ((current->euid != task->suid) && (current->euid != task->uid) &&
993 (current->uid != task->suid) && (current->uid != task->uid) &&
994 !capable(CAP_SYS_NICE)) {
999 err = security_task_movememory(task);
1004 task_nodes = cpuset_mems_allowed(task);
1006 /* Limit nr_pages so that the multiplication may not overflow */
1007 if (nr_pages >= ULONG_MAX / sizeof(struct page_to_node) - 1) {
1012 pm = vmalloc((nr_pages + 1) * sizeof(struct page_to_node));
1019 * Get parameters from user space and initialize the pm
1020 * array. Return various errors if the user did something wrong.
1022 for (i = 0; i < nr_pages; i++) {
1023 const void __user *p;
1026 if (get_user(p, pages + i))
1029 pm[i].addr = (unsigned long)p;
1033 if (get_user(node, nodes + i))
1037 if (!node_state(node, N_HIGH_MEMORY))
1041 if (!node_isset(node, task_nodes))
1046 pm[i].node = 0; /* anything to not match MAX_NUMNODES */
1049 pm[nr_pages].node = MAX_NUMNODES;
1052 err = do_move_pages(mm, pm, flags & MPOL_MF_MOVE_ALL);
1054 err = do_pages_stat(mm, pm);
1057 /* Return status information */
1058 for (i = 0; i < nr_pages; i++)
1059 if (put_user(pm[i].status, status + i))
1070 * Call migration functions in the vma_ops that may prepare
1071 * memory in a vm for migration. migration functions may perform
1072 * the migration for vmas that do not have an underlying page struct.
1074 int migrate_vmas(struct mm_struct *mm, const nodemask_t *to,
1075 const nodemask_t *from, unsigned long flags)
1077 struct vm_area_struct *vma;
1080 for(vma = mm->mmap; vma->vm_next && !err; vma = vma->vm_next) {
1081 if (vma->vm_ops && vma->vm_ops->migrate) {
1082 err = vma->vm_ops->migrate(vma, to, from, flags);