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/export.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/ksm.h>
25 #include <linux/rmap.h>
26 #include <linux/topology.h>
27 #include <linux/cpu.h>
28 #include <linux/cpuset.h>
29 #include <linux/writeback.h>
30 #include <linux/mempolicy.h>
31 #include <linux/vmalloc.h>
32 #include <linux/security.h>
33 #include <linux/memcontrol.h>
34 #include <linux/syscalls.h>
35 #include <linux/hugetlb.h>
36 #include <linux/gfp.h>
37 #include <linux/ptrace.h>
39 #include <asm/tlbflush.h>
43 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
46 * migrate_prep() needs to be called before we start compiling a list of pages
47 * to be migrated using isolate_lru_page(). If scheduling work on other CPUs is
48 * undesirable, use migrate_prep_local()
50 int migrate_prep(void)
53 * Clear the LRU lists so pages can be isolated.
54 * Note that pages may be moved off the LRU after we have
55 * drained them. Those pages will fail to migrate like other
56 * pages that may be busy.
63 /* Do the necessary work of migrate_prep but not if it involves other CPUs */
64 int migrate_prep_local(void)
72 * Add isolated pages on the list back to the LRU under page lock
73 * to avoid leaking evictable pages back onto unevictable list.
75 void putback_lru_pages(struct list_head *l)
80 list_for_each_entry_safe(page, page2, l, lru) {
82 dec_zone_page_state(page, NR_ISOLATED_ANON +
83 page_is_file_cache(page));
84 putback_lru_page(page);
89 * Restore a potential migration pte to a working pte entry
91 static int remove_migration_pte(struct page *new, struct vm_area_struct *vma,
92 unsigned long addr, void *old)
94 struct mm_struct *mm = vma->vm_mm;
102 if (unlikely(PageHuge(new))) {
103 ptep = huge_pte_offset(mm, addr);
106 ptl = &mm->page_table_lock;
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_trans_huge(*pmd))
119 if (!pmd_present(*pmd))
122 ptep = pte_offset_map(pmd, addr);
125 * Peek to check is_swap_pte() before taking ptlock? No, we
126 * can race mremap's move_ptes(), which skips anon_vma lock.
129 ptl = pte_lockptr(mm, pmd);
134 if (!is_swap_pte(pte))
137 entry = pte_to_swp_entry(pte);
139 if (!is_migration_entry(entry) ||
140 migration_entry_to_page(entry) != old)
144 pte = pte_mkold(mk_pte(new, vma->vm_page_prot));
146 /* Recheck VMA as permissions can change since migration started */
147 if (is_write_migration_entry(entry))
148 pte = maybe_mkwrite(pte, vma);
150 #ifdef CONFIG_HUGETLB_PAGE
152 pte = pte_mkhuge(pte);
154 flush_dcache_page(new);
155 set_pte_at(mm, addr, ptep, pte);
159 hugepage_add_anon_rmap(new, vma, addr);
162 } else if (PageAnon(new))
163 page_add_anon_rmap(new, vma, addr);
165 page_add_file_rmap(new);
167 /* No need to invalidate - it was non-present before */
168 update_mmu_cache(vma, addr, ptep);
170 pte_unmap_unlock(ptep, ptl);
176 * Get rid of all migration entries and replace them by
177 * references to the indicated page.
179 static void remove_migration_ptes(struct page *old, struct page *new)
181 rmap_walk(new, remove_migration_pte, old);
185 * Something used the pte of a page under migration. We need to
186 * get to the page and wait until migration is finished.
187 * When we return from this function the fault will be retried.
189 * This function is called from do_swap_page().
191 static void __migration_entry_wait(struct mm_struct *mm, pte_t *ptep,
200 if (!is_swap_pte(pte))
203 entry = pte_to_swp_entry(pte);
204 if (!is_migration_entry(entry))
207 page = migration_entry_to_page(entry);
210 * Once radix-tree replacement of page migration started, page_count
211 * *must* be zero. And, we don't want to call wait_on_page_locked()
212 * against a page without get_page().
213 * So, we use get_page_unless_zero(), here. Even failed, page fault
216 if (!get_page_unless_zero(page))
218 pte_unmap_unlock(ptep, ptl);
219 wait_on_page_locked(page);
223 pte_unmap_unlock(ptep, ptl);
226 void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
227 unsigned long address)
229 spinlock_t *ptl = pte_lockptr(mm, pmd);
230 pte_t *ptep = pte_offset_map(pmd, address);
231 __migration_entry_wait(mm, ptep, ptl);
234 void migration_entry_wait_huge(struct mm_struct *mm, pte_t *pte)
236 spinlock_t *ptl = &(mm)->page_table_lock;
237 __migration_entry_wait(mm, pte, ptl);
241 /* Returns true if all buffers are successfully locked */
242 static bool buffer_migrate_lock_buffers(struct buffer_head *head,
243 enum migrate_mode mode)
245 struct buffer_head *bh = head;
247 /* Simple case, sync compaction */
248 if (mode != MIGRATE_ASYNC) {
252 bh = bh->b_this_page;
254 } while (bh != head);
259 /* async case, we cannot block on lock_buffer so use trylock_buffer */
262 if (!trylock_buffer(bh)) {
264 * We failed to lock the buffer and cannot stall in
265 * async migration. Release the taken locks
267 struct buffer_head *failed_bh = bh;
270 while (bh != failed_bh) {
273 bh = bh->b_this_page;
278 bh = bh->b_this_page;
279 } while (bh != head);
283 static inline bool buffer_migrate_lock_buffers(struct buffer_head *head,
284 enum migrate_mode mode)
288 #endif /* CONFIG_BLOCK */
291 * Replace the page in the mapping.
293 * The number of remaining references must be:
294 * 1 for anonymous pages without a mapping
295 * 2 for pages with a mapping
296 * 3 for pages with a mapping and PagePrivate/PagePrivate2 set.
298 int migrate_page_move_mapping(struct address_space *mapping,
299 struct page *newpage, struct page *page,
300 struct buffer_head *head, enum migrate_mode mode)
306 /* Anonymous page without mapping */
307 if (page_count(page) != 1)
312 spin_lock_irq(&mapping->tree_lock);
314 pslot = radix_tree_lookup_slot(&mapping->page_tree,
317 expected_count = 2 + page_has_private(page);
318 if (page_count(page) != expected_count ||
319 radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) {
320 spin_unlock_irq(&mapping->tree_lock);
324 if (!page_freeze_refs(page, expected_count)) {
325 spin_unlock_irq(&mapping->tree_lock);
330 * In the async migration case of moving a page with buffers, lock the
331 * buffers using trylock before the mapping is moved. If the mapping
332 * was moved, we later failed to lock the buffers and could not move
333 * the mapping back due to an elevated page count, we would have to
334 * block waiting on other references to be dropped.
336 if (mode == MIGRATE_ASYNC && head &&
337 !buffer_migrate_lock_buffers(head, mode)) {
338 page_unfreeze_refs(page, expected_count);
339 spin_unlock_irq(&mapping->tree_lock);
344 * Now we know that no one else is looking at the page.
346 get_page(newpage); /* add cache reference */
347 if (PageSwapCache(page)) {
348 SetPageSwapCache(newpage);
349 set_page_private(newpage, page_private(page));
352 radix_tree_replace_slot(pslot, newpage);
354 page_unfreeze_refs(page, expected_count);
356 * Drop cache reference from old page.
357 * We know this isn't the last reference.
362 * If moved to a different zone then also account
363 * the page for that zone. Other VM counters will be
364 * taken care of when we establish references to the
365 * new page and drop references to the old page.
367 * Note that anonymous pages are accounted for
368 * via NR_FILE_PAGES and NR_ANON_PAGES if they
369 * are mapped to swap space.
371 __dec_zone_page_state(page, NR_FILE_PAGES);
372 __inc_zone_page_state(newpage, NR_FILE_PAGES);
373 if (!PageSwapCache(page) && PageSwapBacked(page)) {
374 __dec_zone_page_state(page, NR_SHMEM);
375 __inc_zone_page_state(newpage, NR_SHMEM);
377 spin_unlock_irq(&mapping->tree_lock);
381 EXPORT_SYMBOL(migrate_page_move_mapping);
384 * The expected number of remaining references is the same as that
385 * of migrate_page_move_mapping().
387 int migrate_huge_page_move_mapping(struct address_space *mapping,
388 struct page *newpage, struct page *page)
394 if (page_count(page) != 1)
399 spin_lock_irq(&mapping->tree_lock);
401 pslot = radix_tree_lookup_slot(&mapping->page_tree,
404 expected_count = 2 + page_has_private(page);
405 if (page_count(page) != expected_count ||
406 radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) {
407 spin_unlock_irq(&mapping->tree_lock);
411 if (!page_freeze_refs(page, expected_count)) {
412 spin_unlock_irq(&mapping->tree_lock);
418 radix_tree_replace_slot(pslot, newpage);
420 page_unfreeze_refs(page, expected_count);
424 spin_unlock_irq(&mapping->tree_lock);
429 * Copy the page to its new location
431 void migrate_page_copy(struct page *newpage, struct page *page)
434 copy_huge_page(newpage, page);
436 copy_highpage(newpage, page);
439 SetPageError(newpage);
440 if (PageReferenced(page))
441 SetPageReferenced(newpage);
442 if (PageUptodate(page))
443 SetPageUptodate(newpage);
444 if (TestClearPageActive(page)) {
445 VM_BUG_ON(PageUnevictable(page));
446 SetPageActive(newpage);
447 } else if (TestClearPageUnevictable(page))
448 SetPageUnevictable(newpage);
449 if (PageChecked(page))
450 SetPageChecked(newpage);
451 if (PageMappedToDisk(page))
452 SetPageMappedToDisk(newpage);
454 if (PageDirty(page)) {
455 clear_page_dirty_for_io(page);
457 * Want to mark the page and the radix tree as dirty, and
458 * redo the accounting that clear_page_dirty_for_io undid,
459 * but we can't use set_page_dirty because that function
460 * is actually a signal that all of the page has become dirty.
461 * Whereas only part of our page may be dirty.
463 __set_page_dirty_nobuffers(newpage);
466 mlock_migrate_page(newpage, page);
467 ksm_migrate_page(newpage, page);
469 ClearPageSwapCache(page);
470 ClearPagePrivate(page);
471 set_page_private(page, 0);
472 page->mapping = NULL;
475 * If any waiters have accumulated on the new page then
478 if (PageWriteback(newpage))
479 end_page_writeback(newpage);
481 EXPORT_SYMBOL(migrate_page_copy);
483 /************************************************************
484 * Migration functions
485 ***********************************************************/
487 /* Always fail migration. Used for mappings that are not movable */
488 int fail_migrate_page(struct address_space *mapping,
489 struct page *newpage, struct page *page)
493 EXPORT_SYMBOL(fail_migrate_page);
496 * Common logic to directly migrate a single page suitable for
497 * pages that do not use PagePrivate/PagePrivate2.
499 * Pages are locked upon entry and exit.
501 int migrate_page(struct address_space *mapping,
502 struct page *newpage, struct page *page,
503 enum migrate_mode mode)
507 BUG_ON(PageWriteback(page)); /* Writeback must be complete */
509 rc = migrate_page_move_mapping(mapping, newpage, page, NULL, mode);
514 migrate_page_copy(newpage, page);
517 EXPORT_SYMBOL(migrate_page);
521 * Migration function for pages with buffers. This function can only be used
522 * if the underlying filesystem guarantees that no other references to "page"
525 int buffer_migrate_page(struct address_space *mapping,
526 struct page *newpage, struct page *page, enum migrate_mode mode)
528 struct buffer_head *bh, *head;
531 if (!page_has_buffers(page))
532 return migrate_page(mapping, newpage, page, mode);
534 head = page_buffers(page);
536 rc = migrate_page_move_mapping(mapping, newpage, page, head, mode);
542 * In the async case, migrate_page_move_mapping locked the buffers
543 * with an IRQ-safe spinlock held. In the sync case, the buffers
544 * need to be locked now
546 if (mode != MIGRATE_ASYNC)
547 BUG_ON(!buffer_migrate_lock_buffers(head, mode));
549 ClearPagePrivate(page);
550 set_page_private(newpage, page_private(page));
551 set_page_private(page, 0);
557 set_bh_page(bh, newpage, bh_offset(bh));
558 bh = bh->b_this_page;
560 } while (bh != head);
562 SetPagePrivate(newpage);
564 migrate_page_copy(newpage, page);
570 bh = bh->b_this_page;
572 } while (bh != head);
576 EXPORT_SYMBOL(buffer_migrate_page);
580 * Writeback a page to clean the dirty state
582 static int writeout(struct address_space *mapping, struct page *page)
584 struct writeback_control wbc = {
585 .sync_mode = WB_SYNC_NONE,
588 .range_end = LLONG_MAX,
593 if (!mapping->a_ops->writepage)
594 /* No write method for the address space */
597 if (!clear_page_dirty_for_io(page))
598 /* Someone else already triggered a write */
602 * A dirty page may imply that the underlying filesystem has
603 * the page on some queue. So the page must be clean for
604 * migration. Writeout may mean we loose the lock and the
605 * page state is no longer what we checked for earlier.
606 * At this point we know that the migration attempt cannot
609 remove_migration_ptes(page, page);
611 rc = mapping->a_ops->writepage(page, &wbc);
613 if (rc != AOP_WRITEPAGE_ACTIVATE)
614 /* unlocked. Relock */
617 return (rc < 0) ? -EIO : -EAGAIN;
621 * Default handling if a filesystem does not provide a migration function.
623 static int fallback_migrate_page(struct address_space *mapping,
624 struct page *newpage, struct page *page, enum migrate_mode mode)
626 if (PageDirty(page)) {
627 /* Only writeback pages in full synchronous migration */
628 if (mode != MIGRATE_SYNC)
630 return writeout(mapping, page);
634 * Buffers may be managed in a filesystem specific way.
635 * We must have no buffers or drop them.
637 if (page_has_private(page) &&
638 !try_to_release_page(page, GFP_KERNEL))
641 return migrate_page(mapping, newpage, page, mode);
645 * Move a page to a newly allocated page
646 * The page is locked and all ptes have been successfully removed.
648 * The new page will have replaced the old page if this function
655 static int move_to_new_page(struct page *newpage, struct page *page,
656 int remap_swapcache, enum migrate_mode mode)
658 struct address_space *mapping;
662 * Block others from accessing the page when we get around to
663 * establishing additional references. We are the only one
664 * holding a reference to the new page at this point.
666 if (!trylock_page(newpage))
669 /* Prepare mapping for the new page.*/
670 newpage->index = page->index;
671 newpage->mapping = page->mapping;
672 if (PageSwapBacked(page))
673 SetPageSwapBacked(newpage);
675 mapping = page_mapping(page);
677 rc = migrate_page(mapping, newpage, page, mode);
678 else if (mapping->a_ops->migratepage)
680 * Most pages have a mapping and most filesystems provide a
681 * migratepage callback. Anonymous pages are part of swap
682 * space which also has its own migratepage callback. This
683 * is the most common path for page migration.
685 rc = mapping->a_ops->migratepage(mapping,
686 newpage, page, mode);
688 rc = fallback_migrate_page(mapping, newpage, page, mode);
691 newpage->mapping = NULL;
694 remove_migration_ptes(page, newpage);
697 unlock_page(newpage);
702 static int __unmap_and_move(struct page *page, struct page *newpage,
703 int force, bool offlining, enum migrate_mode mode)
706 int remap_swapcache = 1;
708 struct mem_cgroup *mem;
709 struct anon_vma *anon_vma = NULL;
711 if (!trylock_page(page)) {
712 if (!force || mode == MIGRATE_ASYNC)
716 * It's not safe for direct compaction to call lock_page.
717 * For example, during page readahead pages are added locked
718 * to the LRU. Later, when the IO completes the pages are
719 * marked uptodate and unlocked. However, the queueing
720 * could be merging multiple pages for one bio (e.g.
721 * mpage_readpages). If an allocation happens for the
722 * second or third page, the process can end up locking
723 * the same page twice and deadlocking. Rather than
724 * trying to be clever about what pages can be locked,
725 * avoid the use of lock_page for direct compaction
728 if (current->flags & PF_MEMALLOC)
735 * Only memory hotplug's offline_pages() caller has locked out KSM,
736 * and can safely migrate a KSM page. The other cases have skipped
737 * PageKsm along with PageReserved - but it is only now when we have
738 * the page lock that we can be certain it will not go KSM beneath us
739 * (KSM will not upgrade a page from PageAnon to PageKsm when it sees
740 * its pagecount raised, but only here do we take the page lock which
743 if (PageKsm(page) && !offlining) {
748 /* charge against new page */
749 charge = mem_cgroup_prepare_migration(page, newpage, &mem, GFP_KERNEL);
750 if (charge == -ENOMEM) {
756 if (PageWriteback(page)) {
758 * Only in the case of a full syncronous migration is it
759 * necessary to wait for PageWriteback. In the async case,
760 * the retry loop is too short and in the sync-light case,
761 * the overhead of stalling is too much
763 if (mode != MIGRATE_SYNC) {
769 wait_on_page_writeback(page);
772 * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
773 * we cannot notice that anon_vma is freed while we migrates a page.
774 * This get_anon_vma() delays freeing anon_vma pointer until the end
775 * of migration. File cache pages are no problem because of page_lock()
776 * File Caches may use write_page() or lock_page() in migration, then,
777 * just care Anon page here.
779 if (PageAnon(page)) {
781 * Only page_lock_anon_vma() understands the subtleties of
782 * getting a hold on an anon_vma from outside one of its mms.
784 anon_vma = page_get_anon_vma(page);
789 } else if (PageSwapCache(page)) {
791 * We cannot be sure that the anon_vma of an unmapped
792 * swapcache page is safe to use because we don't
793 * know in advance if the VMA that this page belonged
794 * to still exists. If the VMA and others sharing the
795 * data have been freed, then the anon_vma could
796 * already be invalid.
798 * To avoid this possibility, swapcache pages get
799 * migrated but are not remapped when migration
809 * Corner case handling:
810 * 1. When a new swap-cache page is read into, it is added to the LRU
811 * and treated as swapcache but it has no rmap yet.
812 * Calling try_to_unmap() against a page->mapping==NULL page will
813 * trigger a BUG. So handle it here.
814 * 2. An orphaned page (see truncate_complete_page) might have
815 * fs-private metadata. The page can be picked up due to memory
816 * offlining. Everywhere else except page reclaim, the page is
817 * invisible to the vm, so the page can not be migrated. So try to
818 * free the metadata, so the page can be freed.
820 if (!page->mapping) {
821 VM_BUG_ON(PageAnon(page));
822 if (page_has_private(page)) {
823 try_to_free_buffers(page);
829 /* Establish migration ptes or remove ptes */
830 try_to_unmap(page, TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
833 if (!page_mapped(page))
834 rc = move_to_new_page(newpage, page, remap_swapcache, mode);
836 if (rc && remap_swapcache)
837 remove_migration_ptes(page, page);
839 /* Drop an anon_vma reference if we took one */
841 put_anon_vma(anon_vma);
845 mem_cgroup_end_migration(mem, page, newpage, rc == 0);
853 * Obtain the lock on page, remove all ptes and migrate the page
854 * to the newly allocated page in newpage.
856 static int unmap_and_move(new_page_t get_new_page, unsigned long private,
857 struct page *page, int force, bool offlining,
858 enum migrate_mode mode)
862 struct page *newpage = get_new_page(page, private, &result);
867 if (page_count(page) == 1) {
868 /* page was freed from under us. So we are done. */
872 if (unlikely(PageTransHuge(page)))
873 if (unlikely(split_huge_page(page)))
876 rc = __unmap_and_move(page, newpage, force, offlining, mode);
880 * A page that has been migrated has all references
881 * removed and will be freed. A page that has not been
882 * migrated will have kepts its references and be
885 list_del(&page->lru);
886 dec_zone_page_state(page, NR_ISOLATED_ANON +
887 page_is_file_cache(page));
888 putback_lru_page(page);
891 * Move the new page to the LRU. If migration was not successful
892 * then this will free the page.
894 putback_lru_page(newpage);
899 *result = page_to_nid(newpage);
905 * Counterpart of unmap_and_move_page() for hugepage migration.
907 * This function doesn't wait the completion of hugepage I/O
908 * because there is no race between I/O and migration for hugepage.
909 * Note that currently hugepage I/O occurs only in direct I/O
910 * where no lock is held and PG_writeback is irrelevant,
911 * and writeback status of all subpages are counted in the reference
912 * count of the head page (i.e. if all subpages of a 2MB hugepage are
913 * under direct I/O, the reference of the head page is 512 and a bit more.)
914 * This means that when we try to migrate hugepage whose subpages are
915 * doing direct I/O, some references remain after try_to_unmap() and
916 * hugepage migration fails without data corruption.
918 * There is also no race when direct I/O is issued on the page under migration,
919 * because then pte is replaced with migration swap entry and direct I/O code
920 * will wait in the page fault for migration to complete.
922 static int unmap_and_move_huge_page(new_page_t get_new_page,
923 unsigned long private, struct page *hpage,
924 int force, bool offlining,
925 enum migrate_mode mode)
929 struct page *new_hpage = get_new_page(hpage, private, &result);
930 struct anon_vma *anon_vma = NULL;
937 if (!trylock_page(hpage)) {
938 if (!force || mode != MIGRATE_SYNC)
944 anon_vma = page_get_anon_vma(hpage);
946 try_to_unmap(hpage, TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
948 if (!page_mapped(hpage))
949 rc = move_to_new_page(new_hpage, hpage, 1, mode);
952 remove_migration_ptes(hpage, hpage);
955 put_anon_vma(anon_vma);
960 list_del(&hpage->lru);
970 *result = page_to_nid(new_hpage);
978 * The function takes one list of pages to migrate and a function
979 * that determines from the page to be migrated and the private data
980 * the target of the move and allocates the page.
982 * The function returns after 10 attempts or if no pages
983 * are movable anymore because to has become empty
984 * or no retryable pages exist anymore.
985 * Caller should call putback_lru_pages to return pages to the LRU
986 * or free list only if ret != 0.
988 * Return: Number of pages not migrated or error code.
990 int migrate_pages(struct list_head *from,
991 new_page_t get_new_page, unsigned long private, bool offlining,
992 enum migrate_mode mode)
999 int swapwrite = current->flags & PF_SWAPWRITE;
1003 current->flags |= PF_SWAPWRITE;
1005 for(pass = 0; pass < 10 && retry; pass++) {
1008 list_for_each_entry_safe(page, page2, from, lru) {
1011 rc = unmap_and_move(get_new_page, private,
1012 page, pass > 2, offlining,
1024 /* Permanent failure */
1033 current->flags &= ~PF_SWAPWRITE;
1038 return nr_failed + retry;
1041 int migrate_huge_pages(struct list_head *from,
1042 new_page_t get_new_page, unsigned long private, bool offlining,
1043 enum migrate_mode mode)
1052 for (pass = 0; pass < 10 && retry; pass++) {
1055 list_for_each_entry_safe(page, page2, from, lru) {
1058 rc = unmap_and_move_huge_page(get_new_page,
1059 private, page, pass > 2, offlining,
1071 /* Permanent failure */
1082 return nr_failed + retry;
1087 * Move a list of individual pages
1089 struct page_to_node {
1096 static struct page *new_page_node(struct page *p, unsigned long private,
1099 struct page_to_node *pm = (struct page_to_node *)private;
1101 while (pm->node != MAX_NUMNODES && pm->page != p)
1104 if (pm->node == MAX_NUMNODES)
1107 *result = &pm->status;
1109 return alloc_pages_exact_node(pm->node,
1110 GFP_HIGHUSER_MOVABLE | GFP_THISNODE, 0);
1114 * Move a set of pages as indicated in the pm array. The addr
1115 * field must be set to the virtual address of the page to be moved
1116 * and the node number must contain a valid target node.
1117 * The pm array ends with node = MAX_NUMNODES.
1119 static int do_move_page_to_node_array(struct mm_struct *mm,
1120 struct page_to_node *pm,
1124 struct page_to_node *pp;
1125 LIST_HEAD(pagelist);
1127 down_read(&mm->mmap_sem);
1130 * Build a list of pages to migrate
1132 for (pp = pm; pp->node != MAX_NUMNODES; pp++) {
1133 struct vm_area_struct *vma;
1137 vma = find_vma(mm, pp->addr);
1138 if (!vma || pp->addr < vma->vm_start || !vma_migratable(vma))
1141 page = follow_page(vma, pp->addr, FOLL_GET|FOLL_SPLIT);
1143 err = PTR_ERR(page);
1151 /* Use PageReserved to check for zero page */
1152 if (PageReserved(page) || PageKsm(page))
1156 err = page_to_nid(page);
1158 if (err == pp->node)
1160 * Node already in the right place
1165 if (page_mapcount(page) > 1 &&
1169 err = isolate_lru_page(page);
1171 list_add_tail(&page->lru, &pagelist);
1172 inc_zone_page_state(page, NR_ISOLATED_ANON +
1173 page_is_file_cache(page));
1177 * Either remove the duplicate refcount from
1178 * isolate_lru_page() or drop the page ref if it was
1187 if (!list_empty(&pagelist)) {
1188 err = migrate_pages(&pagelist, new_page_node,
1189 (unsigned long)pm, 0, MIGRATE_SYNC);
1191 putback_lru_pages(&pagelist);
1194 up_read(&mm->mmap_sem);
1199 * Migrate an array of page address onto an array of nodes and fill
1200 * the corresponding array of status.
1202 static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes,
1203 unsigned long nr_pages,
1204 const void __user * __user *pages,
1205 const int __user *nodes,
1206 int __user *status, int flags)
1208 struct page_to_node *pm;
1209 unsigned long chunk_nr_pages;
1210 unsigned long chunk_start;
1214 pm = (struct page_to_node *)__get_free_page(GFP_KERNEL);
1221 * Store a chunk of page_to_node array in a page,
1222 * but keep the last one as a marker
1224 chunk_nr_pages = (PAGE_SIZE / sizeof(struct page_to_node)) - 1;
1226 for (chunk_start = 0;
1227 chunk_start < nr_pages;
1228 chunk_start += chunk_nr_pages) {
1231 if (chunk_start + chunk_nr_pages > nr_pages)
1232 chunk_nr_pages = nr_pages - chunk_start;
1234 /* fill the chunk pm with addrs and nodes from user-space */
1235 for (j = 0; j < chunk_nr_pages; j++) {
1236 const void __user *p;
1240 if (get_user(p, pages + j + chunk_start))
1242 pm[j].addr = (unsigned long) p;
1244 if (get_user(node, nodes + j + chunk_start))
1248 if (node < 0 || node >= MAX_NUMNODES)
1251 if (!node_state(node, N_HIGH_MEMORY))
1255 if (!node_isset(node, task_nodes))
1261 /* End marker for this chunk */
1262 pm[chunk_nr_pages].node = MAX_NUMNODES;
1264 /* Migrate this chunk */
1265 err = do_move_page_to_node_array(mm, pm,
1266 flags & MPOL_MF_MOVE_ALL);
1270 /* Return status information */
1271 for (j = 0; j < chunk_nr_pages; j++)
1272 if (put_user(pm[j].status, status + j + chunk_start)) {
1280 free_page((unsigned long)pm);
1286 * Determine the nodes of an array of pages and store it in an array of status.
1288 static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages,
1289 const void __user **pages, int *status)
1293 down_read(&mm->mmap_sem);
1295 for (i = 0; i < nr_pages; i++) {
1296 unsigned long addr = (unsigned long)(*pages);
1297 struct vm_area_struct *vma;
1301 vma = find_vma(mm, addr);
1302 if (!vma || addr < vma->vm_start)
1305 page = follow_page(vma, addr, 0);
1307 err = PTR_ERR(page);
1312 /* Use PageReserved to check for zero page */
1313 if (!page || PageReserved(page) || PageKsm(page))
1316 err = page_to_nid(page);
1324 up_read(&mm->mmap_sem);
1328 * Determine the nodes of a user array of pages and store it in
1329 * a user array of status.
1331 static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages,
1332 const void __user * __user *pages,
1335 #define DO_PAGES_STAT_CHUNK_NR 16
1336 const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR];
1337 int chunk_status[DO_PAGES_STAT_CHUNK_NR];
1340 unsigned long chunk_nr;
1342 chunk_nr = nr_pages;
1343 if (chunk_nr > DO_PAGES_STAT_CHUNK_NR)
1344 chunk_nr = DO_PAGES_STAT_CHUNK_NR;
1346 if (copy_from_user(chunk_pages, pages, chunk_nr * sizeof(*chunk_pages)))
1349 do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status);
1351 if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status)))
1356 nr_pages -= chunk_nr;
1358 return nr_pages ? -EFAULT : 0;
1362 * Move a list of pages in the address space of the currently executing
1365 SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages,
1366 const void __user * __user *, pages,
1367 const int __user *, nodes,
1368 int __user *, status, int, flags)
1370 struct task_struct *task;
1371 struct mm_struct *mm;
1373 nodemask_t task_nodes;
1376 if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
1379 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1382 /* Find the mm_struct */
1384 task = pid ? find_task_by_vpid(pid) : current;
1389 get_task_struct(task);
1392 * Check if this process has the right to modify the specified
1393 * process. Use the regular "ptrace_may_access()" checks.
1395 if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) {
1402 err = security_task_movememory(task);
1406 task_nodes = cpuset_mems_allowed(task);
1407 mm = get_task_mm(task);
1408 put_task_struct(task);
1414 err = do_pages_move(mm, task_nodes, nr_pages, pages,
1415 nodes, status, flags);
1417 err = do_pages_stat(mm, nr_pages, pages, status);
1423 put_task_struct(task);
1428 * Call migration functions in the vma_ops that may prepare
1429 * memory in a vm for migration. migration functions may perform
1430 * the migration for vmas that do not have an underlying page struct.
1432 int migrate_vmas(struct mm_struct *mm, const nodemask_t *to,
1433 const nodemask_t *from, unsigned long flags)
1435 struct vm_area_struct *vma;
1438 for (vma = mm->mmap; vma && !err; vma = vma->vm_next) {
1439 if (vma->vm_ops && vma->vm_ops->migrate) {
1440 err = vma->vm_ops->migrate(vma, to, from, flags);