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/hugetlb_cgroup.h>
37 #include <linux/gfp.h>
38 #include <linux/balloon_compaction.h>
40 #include <asm/tlbflush.h>
42 #define CREATE_TRACE_POINTS
43 #include <trace/events/migrate.h>
48 * migrate_prep() needs to be called before we start compiling a list of pages
49 * to be migrated using isolate_lru_page(). If scheduling work on other CPUs is
50 * undesirable, use migrate_prep_local()
52 int migrate_prep(void)
55 * Clear the LRU lists so pages can be isolated.
56 * Note that pages may be moved off the LRU after we have
57 * drained them. Those pages will fail to migrate like other
58 * pages that may be busy.
65 /* Do the necessary work of migrate_prep but not if it involves other CPUs */
66 int migrate_prep_local(void)
74 * Add isolated pages on the list back to the LRU under page lock
75 * to avoid leaking evictable pages back onto unevictable list.
77 void putback_lru_pages(struct list_head *l)
82 list_for_each_entry_safe(page, page2, l, lru) {
84 dec_zone_page_state(page, NR_ISOLATED_ANON +
85 page_is_file_cache(page));
86 putback_lru_page(page);
91 * Put previously isolated pages back onto the appropriate lists
92 * from where they were once taken off for compaction/migration.
94 * This function shall be used instead of putback_lru_pages(),
95 * whenever the isolated pageset has been built by isolate_migratepages_range()
97 void putback_movable_pages(struct list_head *l)
102 list_for_each_entry_safe(page, page2, l, lru) {
103 if (unlikely(PageHuge(page))) {
104 putback_active_hugepage(page);
107 list_del(&page->lru);
108 dec_zone_page_state(page, NR_ISOLATED_ANON +
109 page_is_file_cache(page));
110 if (unlikely(isolated_balloon_page(page)))
111 balloon_page_putback(page);
113 putback_lru_page(page);
118 * Restore a potential migration pte to a working pte entry
120 static int remove_migration_pte(struct page *new, struct vm_area_struct *vma,
121 unsigned long addr, void *old)
123 struct mm_struct *mm = vma->vm_mm;
129 if (unlikely(PageHuge(new))) {
130 ptep = huge_pte_offset(mm, addr);
133 ptl = huge_pte_lockptr(hstate_vma(vma), mm, ptep);
135 pmd = mm_find_pmd(mm, addr);
138 if (pmd_trans_huge(*pmd))
141 ptep = pte_offset_map(pmd, addr);
144 * Peek to check is_swap_pte() before taking ptlock? No, we
145 * can race mremap's move_ptes(), which skips anon_vma lock.
148 ptl = pte_lockptr(mm, pmd);
153 if (!is_swap_pte(pte))
156 entry = pte_to_swp_entry(pte);
158 if (!is_migration_entry(entry) ||
159 migration_entry_to_page(entry) != old)
163 pte = pte_mkold(mk_pte(new, vma->vm_page_prot));
164 if (pte_swp_soft_dirty(*ptep))
165 pte = pte_mksoft_dirty(pte);
166 if (is_write_migration_entry(entry))
167 pte = pte_mkwrite(pte);
168 #ifdef CONFIG_HUGETLB_PAGE
170 pte = pte_mkhuge(pte);
171 pte = arch_make_huge_pte(pte, vma, new, 0);
174 flush_dcache_page(new);
175 set_pte_at(mm, addr, ptep, pte);
179 hugepage_add_anon_rmap(new, vma, addr);
182 } else if (PageAnon(new))
183 page_add_anon_rmap(new, vma, addr);
185 page_add_file_rmap(new);
187 /* No need to invalidate - it was non-present before */
188 update_mmu_cache(vma, addr, ptep);
190 pte_unmap_unlock(ptep, ptl);
196 * Get rid of all migration entries and replace them by
197 * references to the indicated page.
199 static void remove_migration_ptes(struct page *old, struct page *new)
201 rmap_walk(new, remove_migration_pte, old);
205 * Something used the pte of a page under migration. We need to
206 * get to the page and wait until migration is finished.
207 * When we return from this function the fault will be retried.
209 static void __migration_entry_wait(struct mm_struct *mm, pte_t *ptep,
218 if (!is_swap_pte(pte))
221 entry = pte_to_swp_entry(pte);
222 if (!is_migration_entry(entry))
225 page = migration_entry_to_page(entry);
228 * Once radix-tree replacement of page migration started, page_count
229 * *must* be zero. And, we don't want to call wait_on_page_locked()
230 * against a page without get_page().
231 * So, we use get_page_unless_zero(), here. Even failed, page fault
234 if (!get_page_unless_zero(page))
236 pte_unmap_unlock(ptep, ptl);
237 wait_on_page_locked(page);
241 pte_unmap_unlock(ptep, ptl);
244 void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
245 unsigned long address)
247 spinlock_t *ptl = pte_lockptr(mm, pmd);
248 pte_t *ptep = pte_offset_map(pmd, address);
249 __migration_entry_wait(mm, ptep, ptl);
252 void migration_entry_wait_huge(struct vm_area_struct *vma,
253 struct mm_struct *mm, pte_t *pte)
255 spinlock_t *ptl = huge_pte_lockptr(hstate_vma(vma), mm, pte);
256 __migration_entry_wait(mm, pte, ptl);
260 /* Returns true if all buffers are successfully locked */
261 static bool buffer_migrate_lock_buffers(struct buffer_head *head,
262 enum migrate_mode mode)
264 struct buffer_head *bh = head;
266 /* Simple case, sync compaction */
267 if (mode != MIGRATE_ASYNC) {
271 bh = bh->b_this_page;
273 } while (bh != head);
278 /* async case, we cannot block on lock_buffer so use trylock_buffer */
281 if (!trylock_buffer(bh)) {
283 * We failed to lock the buffer and cannot stall in
284 * async migration. Release the taken locks
286 struct buffer_head *failed_bh = bh;
289 while (bh != failed_bh) {
292 bh = bh->b_this_page;
297 bh = bh->b_this_page;
298 } while (bh != head);
302 static inline bool buffer_migrate_lock_buffers(struct buffer_head *head,
303 enum migrate_mode mode)
307 #endif /* CONFIG_BLOCK */
310 * Replace the page in the mapping.
312 * The number of remaining references must be:
313 * 1 for anonymous pages without a mapping
314 * 2 for pages with a mapping
315 * 3 for pages with a mapping and PagePrivate/PagePrivate2 set.
317 int migrate_page_move_mapping(struct address_space *mapping,
318 struct page *newpage, struct page *page,
319 struct buffer_head *head, enum migrate_mode mode)
321 int expected_count = 0;
325 /* Anonymous page without mapping */
326 if (page_count(page) != 1)
328 return MIGRATEPAGE_SUCCESS;
331 spin_lock_irq(&mapping->tree_lock);
333 pslot = radix_tree_lookup_slot(&mapping->page_tree,
336 expected_count = 2 + page_has_private(page);
337 if (page_count(page) != expected_count ||
338 radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) {
339 spin_unlock_irq(&mapping->tree_lock);
343 if (!page_freeze_refs(page, expected_count)) {
344 spin_unlock_irq(&mapping->tree_lock);
349 * In the async migration case of moving a page with buffers, lock the
350 * buffers using trylock before the mapping is moved. If the mapping
351 * was moved, we later failed to lock the buffers and could not move
352 * the mapping back due to an elevated page count, we would have to
353 * block waiting on other references to be dropped.
355 if (mode == MIGRATE_ASYNC && head &&
356 !buffer_migrate_lock_buffers(head, mode)) {
357 page_unfreeze_refs(page, expected_count);
358 spin_unlock_irq(&mapping->tree_lock);
363 * Now we know that no one else is looking at the page.
365 get_page(newpage); /* add cache reference */
366 if (PageSwapCache(page)) {
367 SetPageSwapCache(newpage);
368 set_page_private(newpage, page_private(page));
371 radix_tree_replace_slot(pslot, newpage);
374 * Drop cache reference from old page by unfreezing
375 * to one less reference.
376 * We know this isn't the last reference.
378 page_unfreeze_refs(page, expected_count - 1);
381 * If moved to a different zone then also account
382 * the page for that zone. Other VM counters will be
383 * taken care of when we establish references to the
384 * new page and drop references to the old page.
386 * Note that anonymous pages are accounted for
387 * via NR_FILE_PAGES and NR_ANON_PAGES if they
388 * are mapped to swap space.
390 __dec_zone_page_state(page, NR_FILE_PAGES);
391 __inc_zone_page_state(newpage, NR_FILE_PAGES);
392 if (!PageSwapCache(page) && PageSwapBacked(page)) {
393 __dec_zone_page_state(page, NR_SHMEM);
394 __inc_zone_page_state(newpage, NR_SHMEM);
396 spin_unlock_irq(&mapping->tree_lock);
398 return MIGRATEPAGE_SUCCESS;
402 * The expected number of remaining references is the same as that
403 * of migrate_page_move_mapping().
405 int migrate_huge_page_move_mapping(struct address_space *mapping,
406 struct page *newpage, struct page *page)
412 if (page_count(page) != 1)
414 return MIGRATEPAGE_SUCCESS;
417 spin_lock_irq(&mapping->tree_lock);
419 pslot = radix_tree_lookup_slot(&mapping->page_tree,
422 expected_count = 2 + page_has_private(page);
423 if (page_count(page) != expected_count ||
424 radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) {
425 spin_unlock_irq(&mapping->tree_lock);
429 if (!page_freeze_refs(page, expected_count)) {
430 spin_unlock_irq(&mapping->tree_lock);
436 radix_tree_replace_slot(pslot, newpage);
438 page_unfreeze_refs(page, expected_count - 1);
440 spin_unlock_irq(&mapping->tree_lock);
441 return MIGRATEPAGE_SUCCESS;
445 * Gigantic pages are so large that we do not guarantee that page++ pointer
446 * arithmetic will work across the entire page. We need something more
449 static void __copy_gigantic_page(struct page *dst, struct page *src,
453 struct page *dst_base = dst;
454 struct page *src_base = src;
456 for (i = 0; i < nr_pages; ) {
458 copy_highpage(dst, src);
461 dst = mem_map_next(dst, dst_base, i);
462 src = mem_map_next(src, src_base, i);
466 static void copy_huge_page(struct page *dst, struct page *src)
473 struct hstate *h = page_hstate(src);
474 nr_pages = pages_per_huge_page(h);
476 if (unlikely(nr_pages > MAX_ORDER_NR_PAGES)) {
477 __copy_gigantic_page(dst, src, nr_pages);
482 BUG_ON(!PageTransHuge(src));
483 nr_pages = hpage_nr_pages(src);
486 for (i = 0; i < nr_pages; i++) {
488 copy_highpage(dst + i, src + i);
493 * Copy the page to its new location
495 void migrate_page_copy(struct page *newpage, struct page *page)
499 if (PageHuge(page) || PageTransHuge(page))
500 copy_huge_page(newpage, page);
502 copy_highpage(newpage, page);
505 SetPageError(newpage);
506 if (PageReferenced(page))
507 SetPageReferenced(newpage);
508 if (PageUptodate(page))
509 SetPageUptodate(newpage);
510 if (TestClearPageActive(page)) {
511 VM_BUG_ON(PageUnevictable(page));
512 SetPageActive(newpage);
513 } else if (TestClearPageUnevictable(page))
514 SetPageUnevictable(newpage);
515 if (PageChecked(page))
516 SetPageChecked(newpage);
517 if (PageMappedToDisk(page))
518 SetPageMappedToDisk(newpage);
520 if (PageDirty(page)) {
521 clear_page_dirty_for_io(page);
523 * Want to mark the page and the radix tree as dirty, and
524 * redo the accounting that clear_page_dirty_for_io undid,
525 * but we can't use set_page_dirty because that function
526 * is actually a signal that all of the page has become dirty.
527 * Whereas only part of our page may be dirty.
529 if (PageSwapBacked(page))
530 SetPageDirty(newpage);
532 __set_page_dirty_nobuffers(newpage);
536 * Copy NUMA information to the new page, to prevent over-eager
537 * future migrations of this same page.
539 cpupid = page_cpupid_xchg_last(page, -1);
540 page_cpupid_xchg_last(newpage, cpupid);
542 mlock_migrate_page(newpage, page);
543 ksm_migrate_page(newpage, page);
545 * Please do not reorder this without considering how mm/ksm.c's
546 * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache().
548 ClearPageSwapCache(page);
549 ClearPagePrivate(page);
550 set_page_private(page, 0);
553 * If any waiters have accumulated on the new page then
556 if (PageWriteback(newpage))
557 end_page_writeback(newpage);
560 /************************************************************
561 * Migration functions
562 ***********************************************************/
564 /* Always fail migration. Used for mappings that are not movable */
565 int fail_migrate_page(struct address_space *mapping,
566 struct page *newpage, struct page *page)
570 EXPORT_SYMBOL(fail_migrate_page);
573 * Common logic to directly migrate a single page suitable for
574 * pages that do not use PagePrivate/PagePrivate2.
576 * Pages are locked upon entry and exit.
578 int migrate_page(struct address_space *mapping,
579 struct page *newpage, struct page *page,
580 enum migrate_mode mode)
584 BUG_ON(PageWriteback(page)); /* Writeback must be complete */
586 rc = migrate_page_move_mapping(mapping, newpage, page, NULL, mode);
588 if (rc != MIGRATEPAGE_SUCCESS)
591 migrate_page_copy(newpage, page);
592 return MIGRATEPAGE_SUCCESS;
594 EXPORT_SYMBOL(migrate_page);
598 * Migration function for pages with buffers. This function can only be used
599 * if the underlying filesystem guarantees that no other references to "page"
602 int buffer_migrate_page(struct address_space *mapping,
603 struct page *newpage, struct page *page, enum migrate_mode mode)
605 struct buffer_head *bh, *head;
608 if (!page_has_buffers(page))
609 return migrate_page(mapping, newpage, page, mode);
611 head = page_buffers(page);
613 rc = migrate_page_move_mapping(mapping, newpage, page, head, mode);
615 if (rc != MIGRATEPAGE_SUCCESS)
619 * In the async case, migrate_page_move_mapping locked the buffers
620 * with an IRQ-safe spinlock held. In the sync case, the buffers
621 * need to be locked now
623 if (mode != MIGRATE_ASYNC)
624 BUG_ON(!buffer_migrate_lock_buffers(head, mode));
626 ClearPagePrivate(page);
627 set_page_private(newpage, page_private(page));
628 set_page_private(page, 0);
634 set_bh_page(bh, newpage, bh_offset(bh));
635 bh = bh->b_this_page;
637 } while (bh != head);
639 SetPagePrivate(newpage);
641 migrate_page_copy(newpage, page);
647 bh = bh->b_this_page;
649 } while (bh != head);
651 return MIGRATEPAGE_SUCCESS;
653 EXPORT_SYMBOL(buffer_migrate_page);
657 * Writeback a page to clean the dirty state
659 static int writeout(struct address_space *mapping, struct page *page)
661 struct writeback_control wbc = {
662 .sync_mode = WB_SYNC_NONE,
665 .range_end = LLONG_MAX,
670 if (!mapping->a_ops->writepage)
671 /* No write method for the address space */
674 if (!clear_page_dirty_for_io(page))
675 /* Someone else already triggered a write */
679 * A dirty page may imply that the underlying filesystem has
680 * the page on some queue. So the page must be clean for
681 * migration. Writeout may mean we loose the lock and the
682 * page state is no longer what we checked for earlier.
683 * At this point we know that the migration attempt cannot
686 remove_migration_ptes(page, page);
688 rc = mapping->a_ops->writepage(page, &wbc);
690 if (rc != AOP_WRITEPAGE_ACTIVATE)
691 /* unlocked. Relock */
694 return (rc < 0) ? -EIO : -EAGAIN;
698 * Default handling if a filesystem does not provide a migration function.
700 static int fallback_migrate_page(struct address_space *mapping,
701 struct page *newpage, struct page *page, enum migrate_mode mode)
703 if (PageDirty(page)) {
704 /* Only writeback pages in full synchronous migration */
705 if (mode != MIGRATE_SYNC)
707 return writeout(mapping, page);
711 * Buffers may be managed in a filesystem specific way.
712 * We must have no buffers or drop them.
714 if (page_has_private(page) &&
715 !try_to_release_page(page, GFP_KERNEL))
718 return migrate_page(mapping, newpage, page, mode);
722 * Move a page to a newly allocated page
723 * The page is locked and all ptes have been successfully removed.
725 * The new page will have replaced the old page if this function
730 * MIGRATEPAGE_SUCCESS - success
732 static int move_to_new_page(struct page *newpage, struct page *page,
733 int remap_swapcache, enum migrate_mode mode)
735 struct address_space *mapping;
739 * Block others from accessing the page when we get around to
740 * establishing additional references. We are the only one
741 * holding a reference to the new page at this point.
743 if (!trylock_page(newpage))
746 /* Prepare mapping for the new page.*/
747 newpage->index = page->index;
748 newpage->mapping = page->mapping;
749 if (PageSwapBacked(page))
750 SetPageSwapBacked(newpage);
752 mapping = page_mapping(page);
754 rc = migrate_page(mapping, newpage, page, mode);
755 else if (mapping->a_ops->migratepage)
757 * Most pages have a mapping and most filesystems provide a
758 * migratepage callback. Anonymous pages are part of swap
759 * space which also has its own migratepage callback. This
760 * is the most common path for page migration.
762 rc = mapping->a_ops->migratepage(mapping,
763 newpage, page, mode);
765 rc = fallback_migrate_page(mapping, newpage, page, mode);
767 if (rc != MIGRATEPAGE_SUCCESS) {
768 newpage->mapping = NULL;
771 remove_migration_ptes(page, newpage);
772 page->mapping = NULL;
775 unlock_page(newpage);
780 static int __unmap_and_move(struct page *page, struct page *newpage,
781 int force, enum migrate_mode mode)
784 int remap_swapcache = 1;
785 struct mem_cgroup *mem;
786 struct anon_vma *anon_vma = NULL;
788 if (!trylock_page(page)) {
789 if (!force || mode == MIGRATE_ASYNC)
793 * It's not safe for direct compaction to call lock_page.
794 * For example, during page readahead pages are added locked
795 * to the LRU. Later, when the IO completes the pages are
796 * marked uptodate and unlocked. However, the queueing
797 * could be merging multiple pages for one bio (e.g.
798 * mpage_readpages). If an allocation happens for the
799 * second or third page, the process can end up locking
800 * the same page twice and deadlocking. Rather than
801 * trying to be clever about what pages can be locked,
802 * avoid the use of lock_page for direct compaction
805 if (current->flags & PF_MEMALLOC)
811 /* charge against new page */
812 mem_cgroup_prepare_migration(page, newpage, &mem);
814 if (PageWriteback(page)) {
816 * Only in the case of a full synchronous migration is it
817 * necessary to wait for PageWriteback. In the async case,
818 * the retry loop is too short and in the sync-light case,
819 * the overhead of stalling is too much
821 if (mode != MIGRATE_SYNC) {
827 wait_on_page_writeback(page);
830 * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
831 * we cannot notice that anon_vma is freed while we migrates a page.
832 * This get_anon_vma() delays freeing anon_vma pointer until the end
833 * of migration. File cache pages are no problem because of page_lock()
834 * File Caches may use write_page() or lock_page() in migration, then,
835 * just care Anon page here.
837 if (PageAnon(page) && !PageKsm(page)) {
839 * Only page_lock_anon_vma_read() understands the subtleties of
840 * getting a hold on an anon_vma from outside one of its mms.
842 anon_vma = page_get_anon_vma(page);
847 } else if (PageSwapCache(page)) {
849 * We cannot be sure that the anon_vma of an unmapped
850 * swapcache page is safe to use because we don't
851 * know in advance if the VMA that this page belonged
852 * to still exists. If the VMA and others sharing the
853 * data have been freed, then the anon_vma could
854 * already be invalid.
856 * To avoid this possibility, swapcache pages get
857 * migrated but are not remapped when migration
866 if (unlikely(balloon_page_movable(page))) {
868 * A ballooned page does not need any special attention from
869 * physical to virtual reverse mapping procedures.
870 * Skip any attempt to unmap PTEs or to remap swap cache,
871 * in order to avoid burning cycles at rmap level, and perform
872 * the page migration right away (proteced by page lock).
874 rc = balloon_page_migrate(newpage, page, mode);
879 * Corner case handling:
880 * 1. When a new swap-cache page is read into, it is added to the LRU
881 * and treated as swapcache but it has no rmap yet.
882 * Calling try_to_unmap() against a page->mapping==NULL page will
883 * trigger a BUG. So handle it here.
884 * 2. An orphaned page (see truncate_complete_page) might have
885 * fs-private metadata. The page can be picked up due to memory
886 * offlining. Everywhere else except page reclaim, the page is
887 * invisible to the vm, so the page can not be migrated. So try to
888 * free the metadata, so the page can be freed.
890 if (!page->mapping) {
891 VM_BUG_ON(PageAnon(page));
892 if (page_has_private(page)) {
893 try_to_free_buffers(page);
899 /* Establish migration ptes or remove ptes */
900 try_to_unmap(page, TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
903 if (!page_mapped(page))
904 rc = move_to_new_page(newpage, page, remap_swapcache, mode);
906 if (rc && remap_swapcache)
907 remove_migration_ptes(page, page);
909 /* Drop an anon_vma reference if we took one */
911 put_anon_vma(anon_vma);
914 mem_cgroup_end_migration(mem, page, newpage,
915 (rc == MIGRATEPAGE_SUCCESS ||
916 rc == MIGRATEPAGE_BALLOON_SUCCESS));
923 * Obtain the lock on page, remove all ptes and migrate the page
924 * to the newly allocated page in newpage.
926 static int unmap_and_move(new_page_t get_new_page, unsigned long private,
927 struct page *page, int force, enum migrate_mode mode)
931 struct page *newpage = get_new_page(page, private, &result);
936 if (page_count(page) == 1) {
937 /* page was freed from under us. So we are done. */
941 if (unlikely(PageTransHuge(page)))
942 if (unlikely(split_huge_page(page)))
945 rc = __unmap_and_move(page, newpage, force, mode);
947 if (unlikely(rc == MIGRATEPAGE_BALLOON_SUCCESS)) {
949 * A ballooned page has been migrated already.
950 * Now, it's the time to wrap-up counters,
951 * handle the page back to Buddy and return.
953 dec_zone_page_state(page, NR_ISOLATED_ANON +
954 page_is_file_cache(page));
955 balloon_page_free(page);
956 return MIGRATEPAGE_SUCCESS;
961 * A page that has been migrated has all references
962 * removed and will be freed. A page that has not been
963 * migrated will have kepts its references and be
966 list_del(&page->lru);
967 dec_zone_page_state(page, NR_ISOLATED_ANON +
968 page_is_file_cache(page));
969 putback_lru_page(page);
972 * Move the new page to the LRU. If migration was not successful
973 * then this will free the page.
975 putback_lru_page(newpage);
980 *result = page_to_nid(newpage);
986 * Counterpart of unmap_and_move_page() for hugepage migration.
988 * This function doesn't wait the completion of hugepage I/O
989 * because there is no race between I/O and migration for hugepage.
990 * Note that currently hugepage I/O occurs only in direct I/O
991 * where no lock is held and PG_writeback is irrelevant,
992 * and writeback status of all subpages are counted in the reference
993 * count of the head page (i.e. if all subpages of a 2MB hugepage are
994 * under direct I/O, the reference of the head page is 512 and a bit more.)
995 * This means that when we try to migrate hugepage whose subpages are
996 * doing direct I/O, some references remain after try_to_unmap() and
997 * hugepage migration fails without data corruption.
999 * There is also no race when direct I/O is issued on the page under migration,
1000 * because then pte is replaced with migration swap entry and direct I/O code
1001 * will wait in the page fault for migration to complete.
1003 static int unmap_and_move_huge_page(new_page_t get_new_page,
1004 unsigned long private, struct page *hpage,
1005 int force, enum migrate_mode mode)
1009 struct page *new_hpage = get_new_page(hpage, private, &result);
1010 struct anon_vma *anon_vma = NULL;
1013 * Movability of hugepages depends on architectures and hugepage size.
1014 * This check is necessary because some callers of hugepage migration
1015 * like soft offline and memory hotremove don't walk through page
1016 * tables or check whether the hugepage is pmd-based or not before
1017 * kicking migration.
1019 if (!hugepage_migration_support(page_hstate(hpage)))
1027 if (!trylock_page(hpage)) {
1028 if (!force || mode != MIGRATE_SYNC)
1033 if (PageAnon(hpage))
1034 anon_vma = page_get_anon_vma(hpage);
1036 try_to_unmap(hpage, TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
1038 if (!page_mapped(hpage))
1039 rc = move_to_new_page(new_hpage, hpage, 1, mode);
1042 remove_migration_ptes(hpage, hpage);
1045 put_anon_vma(anon_vma);
1048 hugetlb_cgroup_migrate(hpage, new_hpage);
1053 putback_active_hugepage(hpage);
1054 put_page(new_hpage);
1059 *result = page_to_nid(new_hpage);
1065 * migrate_pages - migrate the pages specified in a list, to the free pages
1066 * supplied as the target for the page migration
1068 * @from: The list of pages to be migrated.
1069 * @get_new_page: The function used to allocate free pages to be used
1070 * as the target of the page migration.
1071 * @private: Private data to be passed on to get_new_page()
1072 * @mode: The migration mode that specifies the constraints for
1073 * page migration, if any.
1074 * @reason: The reason for page migration.
1076 * The function returns after 10 attempts or if no pages are movable any more
1077 * because the list has become empty or no retryable pages exist any more.
1078 * The caller should call putback_lru_pages() to return pages to the LRU
1079 * or free list only if ret != 0.
1081 * Returns the number of pages that were not migrated, or an error code.
1083 int migrate_pages(struct list_head *from, new_page_t get_new_page,
1084 unsigned long private, enum migrate_mode mode, int reason)
1088 int nr_succeeded = 0;
1092 int swapwrite = current->flags & PF_SWAPWRITE;
1096 current->flags |= PF_SWAPWRITE;
1098 for(pass = 0; pass < 10 && retry; pass++) {
1101 list_for_each_entry_safe(page, page2, from, lru) {
1105 rc = unmap_and_move_huge_page(get_new_page,
1106 private, page, pass > 2, mode);
1108 rc = unmap_and_move(get_new_page, private,
1109 page, pass > 2, mode);
1117 case MIGRATEPAGE_SUCCESS:
1121 /* Permanent failure */
1127 rc = nr_failed + retry;
1130 count_vm_events(PGMIGRATE_SUCCESS, nr_succeeded);
1132 count_vm_events(PGMIGRATE_FAIL, nr_failed);
1133 trace_mm_migrate_pages(nr_succeeded, nr_failed, mode, reason);
1136 current->flags &= ~PF_SWAPWRITE;
1143 * Move a list of individual pages
1145 struct page_to_node {
1152 static struct page *new_page_node(struct page *p, unsigned long private,
1155 struct page_to_node *pm = (struct page_to_node *)private;
1157 while (pm->node != MAX_NUMNODES && pm->page != p)
1160 if (pm->node == MAX_NUMNODES)
1163 *result = &pm->status;
1166 return alloc_huge_page_node(page_hstate(compound_head(p)),
1169 return alloc_pages_exact_node(pm->node,
1170 GFP_HIGHUSER_MOVABLE | GFP_THISNODE, 0);
1174 * Move a set of pages as indicated in the pm array. The addr
1175 * field must be set to the virtual address of the page to be moved
1176 * and the node number must contain a valid target node.
1177 * The pm array ends with node = MAX_NUMNODES.
1179 static int do_move_page_to_node_array(struct mm_struct *mm,
1180 struct page_to_node *pm,
1184 struct page_to_node *pp;
1185 LIST_HEAD(pagelist);
1187 down_read(&mm->mmap_sem);
1190 * Build a list of pages to migrate
1192 for (pp = pm; pp->node != MAX_NUMNODES; pp++) {
1193 struct vm_area_struct *vma;
1197 vma = find_vma(mm, pp->addr);
1198 if (!vma || pp->addr < vma->vm_start || !vma_migratable(vma))
1201 page = follow_page(vma, pp->addr, FOLL_GET|FOLL_SPLIT);
1203 err = PTR_ERR(page);
1211 /* Use PageReserved to check for zero page */
1212 if (PageReserved(page))
1216 err = page_to_nid(page);
1218 if (err == pp->node)
1220 * Node already in the right place
1225 if (page_mapcount(page) > 1 &&
1229 if (PageHuge(page)) {
1230 isolate_huge_page(page, &pagelist);
1234 err = isolate_lru_page(page);
1236 list_add_tail(&page->lru, &pagelist);
1237 inc_zone_page_state(page, NR_ISOLATED_ANON +
1238 page_is_file_cache(page));
1242 * Either remove the duplicate refcount from
1243 * isolate_lru_page() or drop the page ref if it was
1252 if (!list_empty(&pagelist)) {
1253 err = migrate_pages(&pagelist, new_page_node,
1254 (unsigned long)pm, MIGRATE_SYNC, MR_SYSCALL);
1256 putback_movable_pages(&pagelist);
1259 up_read(&mm->mmap_sem);
1264 * Migrate an array of page address onto an array of nodes and fill
1265 * the corresponding array of status.
1267 static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes,
1268 unsigned long nr_pages,
1269 const void __user * __user *pages,
1270 const int __user *nodes,
1271 int __user *status, int flags)
1273 struct page_to_node *pm;
1274 unsigned long chunk_nr_pages;
1275 unsigned long chunk_start;
1279 pm = (struct page_to_node *)__get_free_page(GFP_KERNEL);
1286 * Store a chunk of page_to_node array in a page,
1287 * but keep the last one as a marker
1289 chunk_nr_pages = (PAGE_SIZE / sizeof(struct page_to_node)) - 1;
1291 for (chunk_start = 0;
1292 chunk_start < nr_pages;
1293 chunk_start += chunk_nr_pages) {
1296 if (chunk_start + chunk_nr_pages > nr_pages)
1297 chunk_nr_pages = nr_pages - chunk_start;
1299 /* fill the chunk pm with addrs and nodes from user-space */
1300 for (j = 0; j < chunk_nr_pages; j++) {
1301 const void __user *p;
1305 if (get_user(p, pages + j + chunk_start))
1307 pm[j].addr = (unsigned long) p;
1309 if (get_user(node, nodes + j + chunk_start))
1313 if (node < 0 || node >= MAX_NUMNODES)
1316 if (!node_state(node, N_MEMORY))
1320 if (!node_isset(node, task_nodes))
1326 /* End marker for this chunk */
1327 pm[chunk_nr_pages].node = MAX_NUMNODES;
1329 /* Migrate this chunk */
1330 err = do_move_page_to_node_array(mm, pm,
1331 flags & MPOL_MF_MOVE_ALL);
1335 /* Return status information */
1336 for (j = 0; j < chunk_nr_pages; j++)
1337 if (put_user(pm[j].status, status + j + chunk_start)) {
1345 free_page((unsigned long)pm);
1351 * Determine the nodes of an array of pages and store it in an array of status.
1353 static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages,
1354 const void __user **pages, int *status)
1358 down_read(&mm->mmap_sem);
1360 for (i = 0; i < nr_pages; i++) {
1361 unsigned long addr = (unsigned long)(*pages);
1362 struct vm_area_struct *vma;
1366 vma = find_vma(mm, addr);
1367 if (!vma || addr < vma->vm_start)
1370 page = follow_page(vma, addr, 0);
1372 err = PTR_ERR(page);
1377 /* Use PageReserved to check for zero page */
1378 if (!page || PageReserved(page))
1381 err = page_to_nid(page);
1389 up_read(&mm->mmap_sem);
1393 * Determine the nodes of a user array of pages and store it in
1394 * a user array of status.
1396 static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages,
1397 const void __user * __user *pages,
1400 #define DO_PAGES_STAT_CHUNK_NR 16
1401 const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR];
1402 int chunk_status[DO_PAGES_STAT_CHUNK_NR];
1405 unsigned long chunk_nr;
1407 chunk_nr = nr_pages;
1408 if (chunk_nr > DO_PAGES_STAT_CHUNK_NR)
1409 chunk_nr = DO_PAGES_STAT_CHUNK_NR;
1411 if (copy_from_user(chunk_pages, pages, chunk_nr * sizeof(*chunk_pages)))
1414 do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status);
1416 if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status)))
1421 nr_pages -= chunk_nr;
1423 return nr_pages ? -EFAULT : 0;
1427 * Move a list of pages in the address space of the currently executing
1430 SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages,
1431 const void __user * __user *, pages,
1432 const int __user *, nodes,
1433 int __user *, status, int, flags)
1435 const struct cred *cred = current_cred(), *tcred;
1436 struct task_struct *task;
1437 struct mm_struct *mm;
1439 nodemask_t task_nodes;
1442 if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
1445 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1448 /* Find the mm_struct */
1450 task = pid ? find_task_by_vpid(pid) : current;
1455 get_task_struct(task);
1458 * Check if this process has the right to modify the specified
1459 * process. The right exists if the process has administrative
1460 * capabilities, superuser privileges or the same
1461 * userid as the target process.
1463 tcred = __task_cred(task);
1464 if (!uid_eq(cred->euid, tcred->suid) && !uid_eq(cred->euid, tcred->uid) &&
1465 !uid_eq(cred->uid, tcred->suid) && !uid_eq(cred->uid, tcred->uid) &&
1466 !capable(CAP_SYS_NICE)) {
1473 err = security_task_movememory(task);
1477 task_nodes = cpuset_mems_allowed(task);
1478 mm = get_task_mm(task);
1479 put_task_struct(task);
1485 err = do_pages_move(mm, task_nodes, nr_pages, pages,
1486 nodes, status, flags);
1488 err = do_pages_stat(mm, nr_pages, pages, status);
1494 put_task_struct(task);
1499 * Call migration functions in the vma_ops that may prepare
1500 * memory in a vm for migration. migration functions may perform
1501 * the migration for vmas that do not have an underlying page struct.
1503 int migrate_vmas(struct mm_struct *mm, const nodemask_t *to,
1504 const nodemask_t *from, unsigned long flags)
1506 struct vm_area_struct *vma;
1509 for (vma = mm->mmap; vma && !err; vma = vma->vm_next) {
1510 if (vma->vm_ops && vma->vm_ops->migrate) {
1511 err = vma->vm_ops->migrate(vma, to, from, flags);
1519 #ifdef CONFIG_NUMA_BALANCING
1521 * Returns true if this is a safe migration target node for misplaced NUMA
1522 * pages. Currently it only checks the watermarks which crude
1524 static bool migrate_balanced_pgdat(struct pglist_data *pgdat,
1525 unsigned long nr_migrate_pages)
1528 for (z = pgdat->nr_zones - 1; z >= 0; z--) {
1529 struct zone *zone = pgdat->node_zones + z;
1531 if (!populated_zone(zone))
1534 if (!zone_reclaimable(zone))
1537 /* Avoid waking kswapd by allocating pages_to_migrate pages. */
1538 if (!zone_watermark_ok(zone, 0,
1539 high_wmark_pages(zone) +
1548 static struct page *alloc_misplaced_dst_page(struct page *page,
1552 int nid = (int) data;
1553 struct page *newpage;
1555 newpage = alloc_pages_exact_node(nid,
1556 (GFP_HIGHUSER_MOVABLE | GFP_THISNODE |
1557 __GFP_NOMEMALLOC | __GFP_NORETRY |
1561 page_cpupid_xchg_last(newpage, page_cpupid_last(page));
1567 * page migration rate limiting control.
1568 * Do not migrate more than @pages_to_migrate in a @migrate_interval_millisecs
1569 * window of time. Default here says do not migrate more than 1280M per second.
1570 * If a node is rate-limited then PTE NUMA updates are also rate-limited. However
1571 * as it is faults that reset the window, pte updates will happen unconditionally
1572 * if there has not been a fault since @pteupdate_interval_millisecs after the
1573 * throttle window closed.
1575 static unsigned int migrate_interval_millisecs __read_mostly = 100;
1576 static unsigned int pteupdate_interval_millisecs __read_mostly = 1000;
1577 static unsigned int ratelimit_pages __read_mostly = 128 << (20 - PAGE_SHIFT);
1579 /* Returns true if NUMA migration is currently rate limited */
1580 bool migrate_ratelimited(int node)
1582 pg_data_t *pgdat = NODE_DATA(node);
1584 if (time_after(jiffies, pgdat->numabalancing_migrate_next_window +
1585 msecs_to_jiffies(pteupdate_interval_millisecs)))
1588 if (pgdat->numabalancing_migrate_nr_pages < ratelimit_pages)
1594 /* Returns true if the node is migrate rate-limited after the update */
1595 bool numamigrate_update_ratelimit(pg_data_t *pgdat, unsigned long nr_pages)
1597 bool rate_limited = false;
1600 * Rate-limit the amount of data that is being migrated to a node.
1601 * Optimal placement is no good if the memory bus is saturated and
1602 * all the time is being spent migrating!
1604 spin_lock(&pgdat->numabalancing_migrate_lock);
1605 if (time_after(jiffies, pgdat->numabalancing_migrate_next_window)) {
1606 pgdat->numabalancing_migrate_nr_pages = 0;
1607 pgdat->numabalancing_migrate_next_window = jiffies +
1608 msecs_to_jiffies(migrate_interval_millisecs);
1610 if (pgdat->numabalancing_migrate_nr_pages > ratelimit_pages)
1611 rate_limited = true;
1613 pgdat->numabalancing_migrate_nr_pages += nr_pages;
1614 spin_unlock(&pgdat->numabalancing_migrate_lock);
1616 return rate_limited;
1619 int numamigrate_isolate_page(pg_data_t *pgdat, struct page *page)
1623 VM_BUG_ON(compound_order(page) && !PageTransHuge(page));
1625 /* Avoid migrating to a node that is nearly full */
1626 if (!migrate_balanced_pgdat(pgdat, 1UL << compound_order(page)))
1629 if (isolate_lru_page(page))
1633 * migrate_misplaced_transhuge_page() skips page migration's usual
1634 * check on page_count(), so we must do it here, now that the page
1635 * has been isolated: a GUP pin, or any other pin, prevents migration.
1636 * The expected page count is 3: 1 for page's mapcount and 1 for the
1637 * caller's pin and 1 for the reference taken by isolate_lru_page().
1639 if (PageTransHuge(page) && page_count(page) != 3) {
1640 putback_lru_page(page);
1644 page_lru = page_is_file_cache(page);
1645 mod_zone_page_state(page_zone(page), NR_ISOLATED_ANON + page_lru,
1646 hpage_nr_pages(page));
1649 * Isolating the page has taken another reference, so the
1650 * caller's reference can be safely dropped without the page
1651 * disappearing underneath us during migration.
1658 * Attempt to migrate a misplaced page to the specified destination
1659 * node. Caller is expected to have an elevated reference count on
1660 * the page that will be dropped by this function before returning.
1662 int migrate_misplaced_page(struct page *page, struct vm_area_struct *vma,
1665 pg_data_t *pgdat = NODE_DATA(node);
1668 LIST_HEAD(migratepages);
1671 * Don't migrate file pages that are mapped in multiple processes
1672 * with execute permissions as they are probably shared libraries.
1674 if (page_mapcount(page) != 1 && page_is_file_cache(page) &&
1675 (vma->vm_flags & VM_EXEC))
1679 * Rate-limit the amount of data that is being migrated to a node.
1680 * Optimal placement is no good if the memory bus is saturated and
1681 * all the time is being spent migrating!
1683 if (numamigrate_update_ratelimit(pgdat, 1))
1686 isolated = numamigrate_isolate_page(pgdat, page);
1690 list_add(&page->lru, &migratepages);
1691 nr_remaining = migrate_pages(&migratepages, alloc_misplaced_dst_page,
1692 node, MIGRATE_ASYNC, MR_NUMA_MISPLACED);
1694 putback_lru_pages(&migratepages);
1697 count_vm_numa_event(NUMA_PAGE_MIGRATE);
1698 BUG_ON(!list_empty(&migratepages));
1705 #endif /* CONFIG_NUMA_BALANCING */
1707 #if defined(CONFIG_NUMA_BALANCING) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
1709 * Migrates a THP to a given target node. page must be locked and is unlocked
1712 int migrate_misplaced_transhuge_page(struct mm_struct *mm,
1713 struct vm_area_struct *vma,
1714 pmd_t *pmd, pmd_t entry,
1715 unsigned long address,
1716 struct page *page, int node)
1719 unsigned long haddr = address & HPAGE_PMD_MASK;
1720 pg_data_t *pgdat = NODE_DATA(node);
1722 struct page *new_page = NULL;
1723 struct mem_cgroup *memcg = NULL;
1724 int page_lru = page_is_file_cache(page);
1727 * Rate-limit the amount of data that is being migrated to a node.
1728 * Optimal placement is no good if the memory bus is saturated and
1729 * all the time is being spent migrating!
1731 if (numamigrate_update_ratelimit(pgdat, HPAGE_PMD_NR))
1734 new_page = alloc_pages_node(node,
1735 (GFP_TRANSHUGE | GFP_THISNODE) & ~__GFP_WAIT, HPAGE_PMD_ORDER);
1739 page_cpupid_xchg_last(new_page, page_cpupid_last(page));
1741 isolated = numamigrate_isolate_page(pgdat, page);
1747 /* Prepare a page as a migration target */
1748 __set_page_locked(new_page);
1749 SetPageSwapBacked(new_page);
1751 /* anon mapping, we can simply copy page->mapping to the new page: */
1752 new_page->mapping = page->mapping;
1753 new_page->index = page->index;
1754 migrate_page_copy(new_page, page);
1755 WARN_ON(PageLRU(new_page));
1757 /* Recheck the target PMD */
1758 ptl = pmd_lock(mm, pmd);
1759 if (unlikely(!pmd_same(*pmd, entry))) {
1762 /* Reverse changes made by migrate_page_copy() */
1763 if (TestClearPageActive(new_page))
1764 SetPageActive(page);
1765 if (TestClearPageUnevictable(new_page))
1766 SetPageUnevictable(page);
1767 mlock_migrate_page(page, new_page);
1769 unlock_page(new_page);
1770 put_page(new_page); /* Free it */
1772 /* Retake the callers reference and putback on LRU */
1774 putback_lru_page(page);
1775 mod_zone_page_state(page_zone(page),
1776 NR_ISOLATED_ANON + page_lru, -HPAGE_PMD_NR);
1781 * Traditional migration needs to prepare the memcg charge
1782 * transaction early to prevent the old page from being
1783 * uncharged when installing migration entries. Here we can
1784 * save the potential rollback and start the charge transfer
1785 * only when migration is already known to end successfully.
1787 mem_cgroup_prepare_migration(page, new_page, &memcg);
1789 entry = mk_pmd(new_page, vma->vm_page_prot);
1790 entry = pmd_mknonnuma(entry);
1791 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1792 entry = pmd_mkhuge(entry);
1794 pmdp_clear_flush(vma, haddr, pmd);
1795 set_pmd_at(mm, haddr, pmd, entry);
1796 page_add_new_anon_rmap(new_page, vma, haddr);
1797 update_mmu_cache_pmd(vma, address, &entry);
1798 page_remove_rmap(page);
1800 * Finish the charge transaction under the page table lock to
1801 * prevent split_huge_page() from dividing up the charge
1802 * before it's fully transferred to the new page.
1804 mem_cgroup_end_migration(memcg, page, new_page, true);
1807 unlock_page(new_page);
1809 put_page(page); /* Drop the rmap reference */
1810 put_page(page); /* Drop the LRU isolation reference */
1812 count_vm_events(PGMIGRATE_SUCCESS, HPAGE_PMD_NR);
1813 count_vm_numa_events(NUMA_PAGE_MIGRATE, HPAGE_PMD_NR);
1815 mod_zone_page_state(page_zone(page),
1816 NR_ISOLATED_ANON + page_lru,
1821 count_vm_events(PGMIGRATE_FAIL, HPAGE_PMD_NR);
1823 entry = pmd_mknonnuma(entry);
1824 set_pmd_at(mm, haddr, pmd, entry);
1825 update_mmu_cache_pmd(vma, address, &entry);
1831 #endif /* CONFIG_NUMA_BALANCING */
1833 #endif /* CONFIG_NUMA */