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/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>
36 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
39 * Isolate one page from the LRU lists. If successful put it onto
40 * the indicated list with elevated page count.
43 * -EBUSY: page not on LRU list
44 * 0: page removed from LRU list and added to the specified list.
46 int isolate_lru_page(struct page *page, struct list_head *pagelist)
51 struct zone *zone = page_zone(page);
53 spin_lock_irq(&zone->lru_lock);
54 if (PageLRU(page) && get_page_unless_zero(page)) {
58 del_page_from_active_list(zone, page);
60 del_page_from_inactive_list(zone, page);
61 list_add_tail(&page->lru, pagelist);
63 spin_unlock_irq(&zone->lru_lock);
69 * migrate_prep() needs to be called before we start compiling a list of pages
70 * to be migrated using isolate_lru_page().
72 int migrate_prep(void)
75 * Clear the LRU lists so pages can be isolated.
76 * Note that pages may be moved off the LRU after we have
77 * drained them. Those pages will fail to migrate like other
78 * pages that may be busy.
85 static inline void move_to_lru(struct page *page)
87 if (PageActive(page)) {
89 * lru_cache_add_active checks that
90 * the PG_active bit is off.
92 ClearPageActive(page);
93 lru_cache_add_active(page);
101 * Add isolated pages on the list back to the LRU.
103 * returns the number of pages put back.
105 int putback_lru_pages(struct list_head *l)
111 list_for_each_entry_safe(page, page2, l, lru) {
112 list_del(&page->lru);
120 * Restore a potential migration pte to a working pte entry
122 static void remove_migration_pte(struct vm_area_struct *vma,
123 struct page *old, struct page *new)
125 struct mm_struct *mm = vma->vm_mm;
132 unsigned long addr = page_address_in_vma(new, vma);
137 pgd = pgd_offset(mm, addr);
138 if (!pgd_present(*pgd))
141 pud = pud_offset(pgd, addr);
142 if (!pud_present(*pud))
145 pmd = pmd_offset(pud, addr);
146 if (!pmd_present(*pmd))
149 ptep = pte_offset_map(pmd, addr);
151 if (!is_swap_pte(*ptep)) {
156 ptl = pte_lockptr(mm, pmd);
159 if (!is_swap_pte(pte))
162 entry = pte_to_swp_entry(pte);
164 if (!is_migration_entry(entry) || migration_entry_to_page(entry) != old)
168 * Yes, ignore the return value from a GFP_ATOMIC mem_cgroup_charge.
169 * Failure is not an option here: we're now expected to remove every
170 * migration pte, and will cause crashes otherwise. Normally this
171 * is not an issue: mem_cgroup_prepare_migration bumped up the old
172 * page_cgroup count for safety, that's now attached to the new page,
173 * so this charge should just be another incrementation of the count,
174 * to keep in balance with rmap.c's mem_cgroup_uncharging. But if
175 * there's been a force_empty, those reference counts may no longer
176 * be reliable, and this charge can actually fail: oh well, we don't
177 * make the situation any worse by proceeding as if it had succeeded.
179 mem_cgroup_charge(new, mm, GFP_ATOMIC);
182 pte = pte_mkold(mk_pte(new, vma->vm_page_prot));
183 if (is_write_migration_entry(entry))
184 pte = pte_mkwrite(pte);
185 flush_cache_page(vma, addr, pte_pfn(pte));
186 set_pte_at(mm, addr, ptep, pte);
189 page_add_anon_rmap(new, vma, addr);
191 page_add_file_rmap(new);
193 /* No need to invalidate - it was non-present before */
194 update_mmu_cache(vma, addr, pte);
197 pte_unmap_unlock(ptep, ptl);
201 * Note that remove_file_migration_ptes will only work on regular mappings,
202 * Nonlinear mappings do not use migration entries.
204 static void remove_file_migration_ptes(struct page *old, struct page *new)
206 struct vm_area_struct *vma;
207 struct address_space *mapping = page_mapping(new);
208 struct prio_tree_iter iter;
209 pgoff_t pgoff = new->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
214 spin_lock(&mapping->i_mmap_lock);
216 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff)
217 remove_migration_pte(vma, old, new);
219 spin_unlock(&mapping->i_mmap_lock);
223 * Must hold mmap_sem lock on at least one of the vmas containing
224 * the page so that the anon_vma cannot vanish.
226 static void remove_anon_migration_ptes(struct page *old, struct page *new)
228 struct anon_vma *anon_vma;
229 struct vm_area_struct *vma;
230 unsigned long mapping;
232 mapping = (unsigned long)new->mapping;
234 if (!mapping || (mapping & PAGE_MAPPING_ANON) == 0)
238 * We hold the mmap_sem lock. So no need to call page_lock_anon_vma.
240 anon_vma = (struct anon_vma *) (mapping - PAGE_MAPPING_ANON);
241 spin_lock(&anon_vma->lock);
243 list_for_each_entry(vma, &anon_vma->head, anon_vma_node)
244 remove_migration_pte(vma, old, new);
246 spin_unlock(&anon_vma->lock);
250 * Get rid of all migration entries and replace them by
251 * references to the indicated page.
253 static void remove_migration_ptes(struct page *old, struct page *new)
256 remove_anon_migration_ptes(old, new);
258 remove_file_migration_ptes(old, new);
262 * Something used the pte of a page under migration. We need to
263 * get to the page and wait until migration is finished.
264 * When we return from this function the fault will be retried.
266 * This function is called from do_swap_page().
268 void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
269 unsigned long address)
276 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
278 if (!is_swap_pte(pte))
281 entry = pte_to_swp_entry(pte);
282 if (!is_migration_entry(entry))
285 page = migration_entry_to_page(entry);
288 pte_unmap_unlock(ptep, ptl);
289 wait_on_page_locked(page);
293 pte_unmap_unlock(ptep, ptl);
297 * Replace the page in the mapping.
299 * The number of remaining references must be:
300 * 1 for anonymous pages without a mapping
301 * 2 for pages with a mapping
302 * 3 for pages with a mapping and PagePrivate set.
304 static int migrate_page_move_mapping(struct address_space *mapping,
305 struct page *newpage, struct page *page)
310 /* Anonymous page without mapping */
311 if (page_count(page) != 1)
316 write_lock_irq(&mapping->tree_lock);
318 pslot = radix_tree_lookup_slot(&mapping->page_tree,
321 if (page_count(page) != 2 + !!PagePrivate(page) ||
322 (struct page *)radix_tree_deref_slot(pslot) != page) {
323 write_unlock_irq(&mapping->tree_lock);
328 * Now we know that no one else is looking at the page.
330 get_page(newpage); /* add cache reference */
332 if (PageSwapCache(page)) {
333 SetPageSwapCache(newpage);
334 set_page_private(newpage, page_private(page));
338 radix_tree_replace_slot(pslot, newpage);
341 * Drop cache reference from old page.
342 * We know this isn't the last reference.
347 * If moved to a different zone then also account
348 * the page for that zone. Other VM counters will be
349 * taken care of when we establish references to the
350 * new page and drop references to the old page.
352 * Note that anonymous pages are accounted for
353 * via NR_FILE_PAGES and NR_ANON_PAGES if they
354 * are mapped to swap space.
356 __dec_zone_page_state(page, NR_FILE_PAGES);
357 __inc_zone_page_state(newpage, NR_FILE_PAGES);
359 write_unlock_irq(&mapping->tree_lock);
365 * Copy the page to its new location
367 static void migrate_page_copy(struct page *newpage, struct page *page)
369 copy_highpage(newpage, page);
372 SetPageError(newpage);
373 if (PageReferenced(page))
374 SetPageReferenced(newpage);
375 if (PageUptodate(page))
376 SetPageUptodate(newpage);
377 if (PageActive(page))
378 SetPageActive(newpage);
379 if (PageChecked(page))
380 SetPageChecked(newpage);
381 if (PageMappedToDisk(page))
382 SetPageMappedToDisk(newpage);
384 if (PageDirty(page)) {
385 clear_page_dirty_for_io(page);
386 set_page_dirty(newpage);
390 ClearPageSwapCache(page);
392 ClearPageActive(page);
393 ClearPagePrivate(page);
394 set_page_private(page, 0);
395 page->mapping = NULL;
398 * If any waiters have accumulated on the new page then
401 if (PageWriteback(newpage))
402 end_page_writeback(newpage);
405 /************************************************************
406 * Migration functions
407 ***********************************************************/
409 /* Always fail migration. Used for mappings that are not movable */
410 int fail_migrate_page(struct address_space *mapping,
411 struct page *newpage, struct page *page)
415 EXPORT_SYMBOL(fail_migrate_page);
418 * Common logic to directly migrate a single page suitable for
419 * pages that do not use PagePrivate.
421 * Pages are locked upon entry and exit.
423 int migrate_page(struct address_space *mapping,
424 struct page *newpage, struct page *page)
428 BUG_ON(PageWriteback(page)); /* Writeback must be complete */
430 rc = migrate_page_move_mapping(mapping, newpage, page);
435 migrate_page_copy(newpage, page);
438 EXPORT_SYMBOL(migrate_page);
442 * Migration function for pages with buffers. This function can only be used
443 * if the underlying filesystem guarantees that no other references to "page"
446 int buffer_migrate_page(struct address_space *mapping,
447 struct page *newpage, struct page *page)
449 struct buffer_head *bh, *head;
452 if (!page_has_buffers(page))
453 return migrate_page(mapping, newpage, page);
455 head = page_buffers(page);
457 rc = migrate_page_move_mapping(mapping, newpage, page);
466 bh = bh->b_this_page;
468 } while (bh != head);
470 ClearPagePrivate(page);
471 set_page_private(newpage, page_private(page));
472 set_page_private(page, 0);
478 set_bh_page(bh, newpage, bh_offset(bh));
479 bh = bh->b_this_page;
481 } while (bh != head);
483 SetPagePrivate(newpage);
485 migrate_page_copy(newpage, page);
491 bh = bh->b_this_page;
493 } while (bh != head);
497 EXPORT_SYMBOL(buffer_migrate_page);
501 * Writeback a page to clean the dirty state
503 static int writeout(struct address_space *mapping, struct page *page)
505 struct writeback_control wbc = {
506 .sync_mode = WB_SYNC_NONE,
509 .range_end = LLONG_MAX,
515 if (!mapping->a_ops->writepage)
516 /* No write method for the address space */
519 if (!clear_page_dirty_for_io(page))
520 /* Someone else already triggered a write */
524 * A dirty page may imply that the underlying filesystem has
525 * the page on some queue. So the page must be clean for
526 * migration. Writeout may mean we loose the lock and the
527 * page state is no longer what we checked for earlier.
528 * At this point we know that the migration attempt cannot
531 remove_migration_ptes(page, page);
533 rc = mapping->a_ops->writepage(page, &wbc);
535 /* I/O Error writing */
538 if (rc != AOP_WRITEPAGE_ACTIVATE)
539 /* unlocked. Relock */
546 * Default handling if a filesystem does not provide a migration function.
548 static int fallback_migrate_page(struct address_space *mapping,
549 struct page *newpage, struct page *page)
552 return writeout(mapping, page);
555 * Buffers may be managed in a filesystem specific way.
556 * We must have no buffers or drop them.
558 if (PagePrivate(page) &&
559 !try_to_release_page(page, GFP_KERNEL))
562 return migrate_page(mapping, newpage, page);
566 * Move a page to a newly allocated page
567 * The page is locked and all ptes have been successfully removed.
569 * The new page will have replaced the old page if this function
572 static int move_to_new_page(struct page *newpage, struct page *page)
574 struct address_space *mapping;
578 * Block others from accessing the page when we get around to
579 * establishing additional references. We are the only one
580 * holding a reference to the new page at this point.
582 if (TestSetPageLocked(newpage))
585 /* Prepare mapping for the new page.*/
586 newpage->index = page->index;
587 newpage->mapping = page->mapping;
589 mapping = page_mapping(page);
591 rc = migrate_page(mapping, newpage, page);
592 else if (mapping->a_ops->migratepage)
594 * Most pages have a mapping and most filesystems
595 * should provide a migration function. Anonymous
596 * pages are part of swap space which also has its
597 * own migration function. This is the most common
598 * path for page migration.
600 rc = mapping->a_ops->migratepage(mapping,
603 rc = fallback_migrate_page(mapping, newpage, page);
606 mem_cgroup_page_migration(page, newpage);
607 remove_migration_ptes(page, newpage);
609 newpage->mapping = NULL;
611 unlock_page(newpage);
617 * Obtain the lock on page, remove all ptes and migrate the page
618 * to the newly allocated page in newpage.
620 static int unmap_and_move(new_page_t get_new_page, unsigned long private,
621 struct page *page, int force)
625 struct page *newpage = get_new_page(page, private, &result);
632 if (page_count(page) == 1)
633 /* page was freed from under us. So we are done. */
637 if (TestSetPageLocked(page)) {
643 if (PageWriteback(page)) {
646 wait_on_page_writeback(page);
649 * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
650 * we cannot notice that anon_vma is freed while we migrates a page.
651 * This rcu_read_lock() delays freeing anon_vma pointer until the end
652 * of migration. File cache pages are no problem because of page_lock()
653 * File Caches may use write_page() or lock_page() in migration, then,
654 * just care Anon page here.
656 if (PageAnon(page)) {
662 * Corner case handling:
663 * 1. When a new swap-cache page is read into, it is added to the LRU
664 * and treated as swapcache but it has no rmap yet.
665 * Calling try_to_unmap() against a page->mapping==NULL page will
666 * trigger a BUG. So handle it here.
667 * 2. An orphaned page (see truncate_complete_page) might have
668 * fs-private metadata. The page can be picked up due to memory
669 * offlining. Everywhere else except page reclaim, the page is
670 * invisible to the vm, so the page can not be migrated. So try to
671 * free the metadata, so the page can be freed.
673 if (!page->mapping) {
674 if (!PageAnon(page) && PagePrivate(page)) {
676 * Go direct to try_to_free_buffers() here because
677 * a) that's what try_to_release_page() would do anyway
678 * b) we may be under rcu_read_lock() here, so we can't
679 * use GFP_KERNEL which is what try_to_release_page()
680 * needs to be effective.
682 try_to_free_buffers(page);
687 charge = mem_cgroup_prepare_migration(page);
688 /* Establish migration ptes or remove ptes */
689 try_to_unmap(page, 1);
691 if (!page_mapped(page))
692 rc = move_to_new_page(newpage, page);
695 remove_migration_ptes(page, page);
697 mem_cgroup_end_migration(page);
699 mem_cgroup_end_migration(newpage);
710 * A page that has been migrated has all references
711 * removed and will be freed. A page that has not been
712 * migrated will have kepts its references and be
715 list_del(&page->lru);
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);
865 if (PageReserved(page)) /* Check for zero page */
869 err = page_to_nid(page);
873 * Node already in the right place
878 if (page_mapcount(page) > 1 &&
882 err = isolate_lru_page(page, &pagelist);
885 * Either remove the duplicate refcount from
886 * isolate_lru_page() or drop the page ref if it was
894 if (!list_empty(&pagelist))
895 err = migrate_pages(&pagelist, new_page_node,
900 up_read(&mm->mmap_sem);
905 * Determine the nodes of a list of pages. The addr in the pm array
906 * must have been set to the virtual address of which we want to determine
909 static int do_pages_stat(struct mm_struct *mm, struct page_to_node *pm)
911 down_read(&mm->mmap_sem);
913 for ( ; pm->node != MAX_NUMNODES; pm++) {
914 struct vm_area_struct *vma;
919 vma = find_vma(mm, pm->addr);
923 page = follow_page(vma, pm->addr, 0);
925 /* Use PageReserved to check for zero page */
926 if (!page || PageReserved(page))
929 err = page_to_nid(page);
934 up_read(&mm->mmap_sem);
939 * Move a list of pages in the address space of the currently executing
942 asmlinkage long sys_move_pages(pid_t pid, unsigned long nr_pages,
943 const void __user * __user *pages,
944 const int __user *nodes,
945 int __user *status, int flags)
949 struct task_struct *task;
950 nodemask_t task_nodes;
951 struct mm_struct *mm;
952 struct page_to_node *pm = NULL;
955 if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
958 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
961 /* Find the mm_struct */
962 read_lock(&tasklist_lock);
963 task = pid ? find_task_by_vpid(pid) : current;
965 read_unlock(&tasklist_lock);
968 mm = get_task_mm(task);
969 read_unlock(&tasklist_lock);
975 * Check if this process has the right to modify the specified
976 * process. The right exists if the process has administrative
977 * capabilities, superuser privileges or the same
978 * userid as the target process.
980 if ((current->euid != task->suid) && (current->euid != task->uid) &&
981 (current->uid != task->suid) && (current->uid != task->uid) &&
982 !capable(CAP_SYS_NICE)) {
987 err = security_task_movememory(task);
992 task_nodes = cpuset_mems_allowed(task);
994 /* Limit nr_pages so that the multiplication may not overflow */
995 if (nr_pages >= ULONG_MAX / sizeof(struct page_to_node) - 1) {
1000 pm = vmalloc((nr_pages + 1) * sizeof(struct page_to_node));
1007 * Get parameters from user space and initialize the pm
1008 * array. Return various errors if the user did something wrong.
1010 for (i = 0; i < nr_pages; i++) {
1011 const void __user *p;
1014 if (get_user(p, pages + i))
1017 pm[i].addr = (unsigned long)p;
1021 if (get_user(node, nodes + i))
1025 if (!node_state(node, N_HIGH_MEMORY))
1029 if (!node_isset(node, task_nodes))
1034 pm[i].node = 0; /* anything to not match MAX_NUMNODES */
1037 pm[nr_pages].node = MAX_NUMNODES;
1040 err = do_move_pages(mm, pm, flags & MPOL_MF_MOVE_ALL);
1042 err = do_pages_stat(mm, pm);
1045 /* Return status information */
1046 for (i = 0; i < nr_pages; i++)
1047 if (put_user(pm[i].status, status + i))
1059 * Call migration functions in the vma_ops that may prepare
1060 * memory in a vm for migration. migration functions may perform
1061 * the migration for vmas that do not have an underlying page struct.
1063 int migrate_vmas(struct mm_struct *mm, const nodemask_t *to,
1064 const nodemask_t *from, unsigned long flags)
1066 struct vm_area_struct *vma;
1069 for(vma = mm->mmap; vma->vm_next && !err; vma = vma->vm_next) {
1070 if (vma->vm_ops && vma->vm_ops->migrate) {
1071 err = vma->vm_ops->migrate(vma, to, from, flags);