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 lru_cache_add_lru(page, page_lru(page));
63 * Add isolated pages on the list back to the LRU.
65 * returns the number of pages put back.
67 int putback_lru_pages(struct list_head *l)
73 list_for_each_entry_safe(page, page2, l, lru) {
82 * Restore a potential migration pte to a working pte entry
84 static void remove_migration_pte(struct vm_area_struct *vma,
85 struct page *old, struct page *new)
87 struct mm_struct *mm = vma->vm_mm;
94 unsigned long addr = page_address_in_vma(new, vma);
99 pgd = pgd_offset(mm, addr);
100 if (!pgd_present(*pgd))
103 pud = pud_offset(pgd, addr);
104 if (!pud_present(*pud))
107 pmd = pmd_offset(pud, addr);
108 if (!pmd_present(*pmd))
111 ptep = pte_offset_map(pmd, addr);
113 if (!is_swap_pte(*ptep)) {
118 ptl = pte_lockptr(mm, pmd);
121 if (!is_swap_pte(pte))
124 entry = pte_to_swp_entry(pte);
126 if (!is_migration_entry(entry) || migration_entry_to_page(entry) != old)
130 * Yes, ignore the return value from a GFP_ATOMIC mem_cgroup_charge.
131 * Failure is not an option here: we're now expected to remove every
132 * migration pte, and will cause crashes otherwise. Normally this
133 * is not an issue: mem_cgroup_prepare_migration bumped up the old
134 * page_cgroup count for safety, that's now attached to the new page,
135 * so this charge should just be another incrementation of the count,
136 * to keep in balance with rmap.c's mem_cgroup_uncharging. But if
137 * there's been a force_empty, those reference counts may no longer
138 * be reliable, and this charge can actually fail: oh well, we don't
139 * make the situation any worse by proceeding as if it had succeeded.
141 mem_cgroup_charge(new, mm, GFP_ATOMIC);
144 pte = pte_mkold(mk_pte(new, vma->vm_page_prot));
145 if (is_write_migration_entry(entry))
146 pte = pte_mkwrite(pte);
147 flush_cache_page(vma, addr, pte_pfn(pte));
148 set_pte_at(mm, addr, ptep, pte);
151 page_add_anon_rmap(new, vma, addr);
153 page_add_file_rmap(new);
155 /* No need to invalidate - it was non-present before */
156 update_mmu_cache(vma, addr, pte);
159 pte_unmap_unlock(ptep, ptl);
163 * Note that remove_file_migration_ptes will only work on regular mappings,
164 * Nonlinear mappings do not use migration entries.
166 static void remove_file_migration_ptes(struct page *old, struct page *new)
168 struct vm_area_struct *vma;
169 struct address_space *mapping = page_mapping(new);
170 struct prio_tree_iter iter;
171 pgoff_t pgoff = new->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
176 spin_lock(&mapping->i_mmap_lock);
178 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff)
179 remove_migration_pte(vma, old, new);
181 spin_unlock(&mapping->i_mmap_lock);
185 * Must hold mmap_sem lock on at least one of the vmas containing
186 * the page so that the anon_vma cannot vanish.
188 static void remove_anon_migration_ptes(struct page *old, struct page *new)
190 struct anon_vma *anon_vma;
191 struct vm_area_struct *vma;
192 unsigned long mapping;
194 mapping = (unsigned long)new->mapping;
196 if (!mapping || (mapping & PAGE_MAPPING_ANON) == 0)
200 * We hold the mmap_sem lock. So no need to call page_lock_anon_vma.
202 anon_vma = (struct anon_vma *) (mapping - PAGE_MAPPING_ANON);
203 spin_lock(&anon_vma->lock);
205 list_for_each_entry(vma, &anon_vma->head, anon_vma_node)
206 remove_migration_pte(vma, old, new);
208 spin_unlock(&anon_vma->lock);
212 * Get rid of all migration entries and replace them by
213 * references to the indicated page.
215 static void remove_migration_ptes(struct page *old, struct page *new)
218 remove_anon_migration_ptes(old, new);
220 remove_file_migration_ptes(old, new);
224 * Something used the pte of a page under migration. We need to
225 * get to the page and wait until migration is finished.
226 * When we return from this function the fault will be retried.
228 * This function is called from do_swap_page().
230 void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
231 unsigned long address)
238 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
240 if (!is_swap_pte(pte))
243 entry = pte_to_swp_entry(pte);
244 if (!is_migration_entry(entry))
247 page = migration_entry_to_page(entry);
250 * Once radix-tree replacement of page migration started, page_count
251 * *must* be zero. And, we don't want to call wait_on_page_locked()
252 * against a page without get_page().
253 * So, we use get_page_unless_zero(), here. Even failed, page fault
256 if (!get_page_unless_zero(page))
258 pte_unmap_unlock(ptep, ptl);
259 wait_on_page_locked(page);
263 pte_unmap_unlock(ptep, ptl);
267 * Replace the page in the mapping.
269 * The number of remaining references must be:
270 * 1 for anonymous pages without a mapping
271 * 2 for pages with a mapping
272 * 3 for pages with a mapping and PagePrivate set.
274 static int migrate_page_move_mapping(struct address_space *mapping,
275 struct page *newpage, struct page *page)
281 /* Anonymous page without mapping */
282 if (page_count(page) != 1)
287 spin_lock_irq(&mapping->tree_lock);
289 pslot = radix_tree_lookup_slot(&mapping->page_tree,
292 expected_count = 2 + !!PagePrivate(page);
293 if (page_count(page) != expected_count ||
294 (struct page *)radix_tree_deref_slot(pslot) != page) {
295 spin_unlock_irq(&mapping->tree_lock);
299 if (!page_freeze_refs(page, expected_count)) {
300 spin_unlock_irq(&mapping->tree_lock);
305 * Now we know that no one else is looking at the page.
307 get_page(newpage); /* add cache reference */
309 if (PageSwapCache(page)) {
310 SetPageSwapCache(newpage);
311 set_page_private(newpage, page_private(page));
315 radix_tree_replace_slot(pslot, newpage);
317 page_unfreeze_refs(page, expected_count);
319 * Drop cache reference from old page.
320 * We know this isn't the last reference.
325 * If moved to a different zone then also account
326 * the page for that zone. Other VM counters will be
327 * taken care of when we establish references to the
328 * new page and drop references to the old page.
330 * Note that anonymous pages are accounted for
331 * via NR_FILE_PAGES and NR_ANON_PAGES if they
332 * are mapped to swap space.
334 __dec_zone_page_state(page, NR_FILE_PAGES);
335 __inc_zone_page_state(newpage, NR_FILE_PAGES);
337 spin_unlock_irq(&mapping->tree_lock);
338 if (!PageSwapCache(newpage))
339 mem_cgroup_uncharge_cache_page(page);
345 * Copy the page to its new location
347 static void migrate_page_copy(struct page *newpage, struct page *page)
349 copy_highpage(newpage, page);
352 SetPageError(newpage);
353 if (PageReferenced(page))
354 SetPageReferenced(newpage);
355 if (PageUptodate(page))
356 SetPageUptodate(newpage);
357 if (PageActive(page))
358 SetPageActive(newpage);
359 if (PageChecked(page))
360 SetPageChecked(newpage);
361 if (PageMappedToDisk(page))
362 SetPageMappedToDisk(newpage);
364 if (PageDirty(page)) {
365 clear_page_dirty_for_io(page);
367 * Want to mark the page and the radix tree as dirty, and
368 * redo the accounting that clear_page_dirty_for_io undid,
369 * but we can't use set_page_dirty because that function
370 * is actually a signal that all of the page has become dirty.
371 * Wheras only part of our page may be dirty.
373 __set_page_dirty_nobuffers(newpage);
377 ClearPageSwapCache(page);
379 ClearPageActive(page);
380 ClearPagePrivate(page);
381 set_page_private(page, 0);
382 page->mapping = NULL;
385 * If any waiters have accumulated on the new page then
388 if (PageWriteback(newpage))
389 end_page_writeback(newpage);
392 /************************************************************
393 * Migration functions
394 ***********************************************************/
396 /* Always fail migration. Used for mappings that are not movable */
397 int fail_migrate_page(struct address_space *mapping,
398 struct page *newpage, struct page *page)
402 EXPORT_SYMBOL(fail_migrate_page);
405 * Common logic to directly migrate a single page suitable for
406 * pages that do not use PagePrivate.
408 * Pages are locked upon entry and exit.
410 int migrate_page(struct address_space *mapping,
411 struct page *newpage, struct page *page)
415 BUG_ON(PageWriteback(page)); /* Writeback must be complete */
417 rc = migrate_page_move_mapping(mapping, newpage, page);
422 migrate_page_copy(newpage, page);
425 EXPORT_SYMBOL(migrate_page);
429 * Migration function for pages with buffers. This function can only be used
430 * if the underlying filesystem guarantees that no other references to "page"
433 int buffer_migrate_page(struct address_space *mapping,
434 struct page *newpage, struct page *page)
436 struct buffer_head *bh, *head;
439 if (!page_has_buffers(page))
440 return migrate_page(mapping, newpage, page);
442 head = page_buffers(page);
444 rc = migrate_page_move_mapping(mapping, newpage, page);
453 bh = bh->b_this_page;
455 } while (bh != head);
457 ClearPagePrivate(page);
458 set_page_private(newpage, page_private(page));
459 set_page_private(page, 0);
465 set_bh_page(bh, newpage, bh_offset(bh));
466 bh = bh->b_this_page;
468 } while (bh != head);
470 SetPagePrivate(newpage);
472 migrate_page_copy(newpage, page);
478 bh = bh->b_this_page;
480 } while (bh != head);
484 EXPORT_SYMBOL(buffer_migrate_page);
488 * Writeback a page to clean the dirty state
490 static int writeout(struct address_space *mapping, struct page *page)
492 struct writeback_control wbc = {
493 .sync_mode = WB_SYNC_NONE,
496 .range_end = LLONG_MAX,
502 if (!mapping->a_ops->writepage)
503 /* No write method for the address space */
506 if (!clear_page_dirty_for_io(page))
507 /* Someone else already triggered a write */
511 * A dirty page may imply that the underlying filesystem has
512 * the page on some queue. So the page must be clean for
513 * migration. Writeout may mean we loose the lock and the
514 * page state is no longer what we checked for earlier.
515 * At this point we know that the migration attempt cannot
518 remove_migration_ptes(page, page);
520 rc = mapping->a_ops->writepage(page, &wbc);
522 /* I/O Error writing */
525 if (rc != AOP_WRITEPAGE_ACTIVATE)
526 /* unlocked. Relock */
533 * Default handling if a filesystem does not provide a migration function.
535 static int fallback_migrate_page(struct address_space *mapping,
536 struct page *newpage, struct page *page)
539 return writeout(mapping, page);
542 * Buffers may be managed in a filesystem specific way.
543 * We must have no buffers or drop them.
545 if (PagePrivate(page) &&
546 !try_to_release_page(page, GFP_KERNEL))
549 return migrate_page(mapping, newpage, page);
553 * Move a page to a newly allocated page
554 * The page is locked and all ptes have been successfully removed.
556 * The new page will have replaced the old page if this function
559 static int move_to_new_page(struct page *newpage, struct page *page)
561 struct address_space *mapping;
565 * Block others from accessing the page when we get around to
566 * establishing additional references. We are the only one
567 * holding a reference to the new page at this point.
569 if (!trylock_page(newpage))
572 /* Prepare mapping for the new page.*/
573 newpage->index = page->index;
574 newpage->mapping = page->mapping;
576 mapping = page_mapping(page);
578 rc = migrate_page(mapping, newpage, page);
579 else if (mapping->a_ops->migratepage)
581 * Most pages have a mapping and most filesystems
582 * should provide a migration function. Anonymous
583 * pages are part of swap space which also has its
584 * own migration function. This is the most common
585 * path for page migration.
587 rc = mapping->a_ops->migratepage(mapping,
590 rc = fallback_migrate_page(mapping, newpage, page);
593 remove_migration_ptes(page, newpage);
595 newpage->mapping = NULL;
597 unlock_page(newpage);
603 * Obtain the lock on page, remove all ptes and migrate the page
604 * to the newly allocated page in newpage.
606 static int unmap_and_move(new_page_t get_new_page, unsigned long private,
607 struct page *page, int force)
611 struct page *newpage = get_new_page(page, private, &result);
618 if (page_count(page) == 1)
619 /* page was freed from under us. So we are done. */
622 charge = mem_cgroup_prepare_migration(page, newpage);
623 if (charge == -ENOMEM) {
627 /* prepare cgroup just returns 0 or -ENOMEM */
631 if (!trylock_page(page)) {
637 if (PageWriteback(page)) {
640 wait_on_page_writeback(page);
643 * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
644 * we cannot notice that anon_vma is freed while we migrates a page.
645 * This rcu_read_lock() delays freeing anon_vma pointer until the end
646 * of migration. File cache pages are no problem because of page_lock()
647 * File Caches may use write_page() or lock_page() in migration, then,
648 * just care Anon page here.
650 if (PageAnon(page)) {
656 * Corner case handling:
657 * 1. When a new swap-cache page is read into, it is added to the LRU
658 * and treated as swapcache but it has no rmap yet.
659 * Calling try_to_unmap() against a page->mapping==NULL page will
660 * trigger a BUG. So handle it here.
661 * 2. An orphaned page (see truncate_complete_page) might have
662 * fs-private metadata. The page can be picked up due to memory
663 * offlining. Everywhere else except page reclaim, the page is
664 * invisible to the vm, so the page can not be migrated. So try to
665 * free the metadata, so the page can be freed.
667 if (!page->mapping) {
668 if (!PageAnon(page) && PagePrivate(page)) {
670 * Go direct to try_to_free_buffers() here because
671 * a) that's what try_to_release_page() would do anyway
672 * b) we may be under rcu_read_lock() here, so we can't
673 * use GFP_KERNEL which is what try_to_release_page()
674 * needs to be effective.
676 try_to_free_buffers(page);
681 /* Establish migration ptes or remove ptes */
682 try_to_unmap(page, 1);
684 if (!page_mapped(page))
685 rc = move_to_new_page(newpage, page);
688 remove_migration_ptes(page, page);
699 * A page that has been migrated has all references
700 * removed and will be freed. A page that has not been
701 * migrated will have kepts its references and be
704 list_del(&page->lru);
710 mem_cgroup_end_migration(newpage);
712 * Move the new page to the LRU. If migration was not successful
713 * then this will free the page.
715 move_to_lru(newpage);
720 *result = page_to_nid(newpage);
728 * The function takes one list of pages to migrate and a function
729 * that determines from the page to be migrated and the private data
730 * the target of the move and allocates the page.
732 * The function returns after 10 attempts or if no pages
733 * are movable anymore because to has become empty
734 * or no retryable pages exist anymore. All pages will be
735 * returned to the LRU or freed.
737 * Return: Number of pages not migrated or error code.
739 int migrate_pages(struct list_head *from,
740 new_page_t get_new_page, unsigned long private)
747 int swapwrite = current->flags & PF_SWAPWRITE;
751 current->flags |= PF_SWAPWRITE;
753 for(pass = 0; pass < 10 && retry; pass++) {
756 list_for_each_entry_safe(page, page2, from, lru) {
759 rc = unmap_and_move(get_new_page, private,
771 /* Permanent failure */
780 current->flags &= ~PF_SWAPWRITE;
782 putback_lru_pages(from);
787 return nr_failed + retry;
792 * Move a list of individual pages
794 struct page_to_node {
801 static struct page *new_page_node(struct page *p, unsigned long private,
804 struct page_to_node *pm = (struct page_to_node *)private;
806 while (pm->node != MAX_NUMNODES && pm->page != p)
809 if (pm->node == MAX_NUMNODES)
812 *result = &pm->status;
814 return alloc_pages_node(pm->node,
815 GFP_HIGHUSER_MOVABLE | GFP_THISNODE, 0);
819 * Move a set of pages as indicated in the pm array. The addr
820 * field must be set to the virtual address of the page to be moved
821 * and the node number must contain a valid target node.
823 static int do_move_pages(struct mm_struct *mm, struct page_to_node *pm,
827 struct page_to_node *pp;
830 down_read(&mm->mmap_sem);
833 * Build a list of pages to migrate
836 for (pp = pm; pp->node != MAX_NUMNODES; pp++) {
837 struct vm_area_struct *vma;
841 * A valid page pointer that will not match any of the
842 * pages that will be moved.
844 pp->page = ZERO_PAGE(0);
847 vma = find_vma(mm, pp->addr);
848 if (!vma || !vma_migratable(vma))
851 page = follow_page(vma, pp->addr, FOLL_GET);
861 if (PageReserved(page)) /* Check for zero page */
865 err = page_to_nid(page);
869 * Node already in the right place
874 if (page_mapcount(page) > 1 &&
878 err = isolate_lru_page(page);
880 list_add_tail(&page->lru, &pagelist);
883 * Either remove the duplicate refcount from
884 * isolate_lru_page() or drop the page ref if it was
892 if (!list_empty(&pagelist))
893 err = migrate_pages(&pagelist, new_page_node,
898 up_read(&mm->mmap_sem);
903 * Determine the nodes of a list of pages. The addr in the pm array
904 * must have been set to the virtual address of which we want to determine
907 static int do_pages_stat(struct mm_struct *mm, struct page_to_node *pm)
909 down_read(&mm->mmap_sem);
911 for ( ; pm->node != MAX_NUMNODES; pm++) {
912 struct vm_area_struct *vma;
917 vma = find_vma(mm, pm->addr);
921 page = follow_page(vma, pm->addr, 0);
928 /* Use PageReserved to check for zero page */
929 if (!page || PageReserved(page))
932 err = page_to_nid(page);
937 up_read(&mm->mmap_sem);
942 * Move a list of pages in the address space of the currently executing
945 asmlinkage long sys_move_pages(pid_t pid, unsigned long nr_pages,
946 const void __user * __user *pages,
947 const int __user *nodes,
948 int __user *status, int flags)
952 struct task_struct *task;
953 nodemask_t task_nodes;
954 struct mm_struct *mm;
955 struct page_to_node *pm = NULL;
958 if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
961 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
964 /* Find the mm_struct */
965 read_lock(&tasklist_lock);
966 task = pid ? find_task_by_vpid(pid) : current;
968 read_unlock(&tasklist_lock);
971 mm = get_task_mm(task);
972 read_unlock(&tasklist_lock);
978 * Check if this process has the right to modify the specified
979 * process. The right exists if the process has administrative
980 * capabilities, superuser privileges or the same
981 * userid as the target process.
983 if ((current->euid != task->suid) && (current->euid != task->uid) &&
984 (current->uid != task->suid) && (current->uid != task->uid) &&
985 !capable(CAP_SYS_NICE)) {
990 err = security_task_movememory(task);
995 task_nodes = cpuset_mems_allowed(task);
997 /* Limit nr_pages so that the multiplication may not overflow */
998 if (nr_pages >= ULONG_MAX / sizeof(struct page_to_node) - 1) {
1003 pm = vmalloc((nr_pages + 1) * sizeof(struct page_to_node));
1010 * Get parameters from user space and initialize the pm
1011 * array. Return various errors if the user did something wrong.
1013 for (i = 0; i < nr_pages; i++) {
1014 const void __user *p;
1017 if (get_user(p, pages + i))
1020 pm[i].addr = (unsigned long)p;
1024 if (get_user(node, nodes + i))
1028 if (!node_state(node, N_HIGH_MEMORY))
1032 if (!node_isset(node, task_nodes))
1037 pm[i].node = 0; /* anything to not match MAX_NUMNODES */
1040 pm[nr_pages].node = MAX_NUMNODES;
1043 err = do_move_pages(mm, pm, flags & MPOL_MF_MOVE_ALL);
1045 err = do_pages_stat(mm, pm);
1048 /* Return status information */
1049 for (i = 0; i < nr_pages; i++)
1050 if (put_user(pm[i].status, status + i))
1061 * Call migration functions in the vma_ops that may prepare
1062 * memory in a vm for migration. migration functions may perform
1063 * the migration for vmas that do not have an underlying page struct.
1065 int migrate_vmas(struct mm_struct *mm, const nodemask_t *to,
1066 const nodemask_t *from, unsigned long flags)
1068 struct vm_area_struct *vma;
1071 for(vma = mm->mmap; vma->vm_next && !err; vma = vma->vm_next) {
1072 if (vma->vm_ops && vma->vm_ops->migrate) {
1073 err = vma->vm_ops->migrate(vma, to, from, flags);