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>
39 #include <linux/mmu_notifier.h>
40 #include <linux/page_idle.h>
42 #include <asm/tlbflush.h>
44 #define CREATE_TRACE_POINTS
45 #include <trace/events/migrate.h>
50 * migrate_prep() needs to be called before we start compiling a list of pages
51 * to be migrated using isolate_lru_page(). If scheduling work on other CPUs is
52 * undesirable, use migrate_prep_local()
54 int migrate_prep(void)
57 * Clear the LRU lists so pages can be isolated.
58 * Note that pages may be moved off the LRU after we have
59 * drained them. Those pages will fail to migrate like other
60 * pages that may be busy.
67 /* Do the necessary work of migrate_prep but not if it involves other CPUs */
68 int migrate_prep_local(void)
76 * Put previously isolated pages back onto the appropriate lists
77 * from where they were once taken off for compaction/migration.
79 * This function shall be used whenever the isolated pageset has been
80 * built from lru, balloon, hugetlbfs page. See isolate_migratepages_range()
81 * and isolate_huge_page().
83 void putback_movable_pages(struct list_head *l)
88 list_for_each_entry_safe(page, page2, l, lru) {
89 if (unlikely(PageHuge(page))) {
90 putback_active_hugepage(page);
94 dec_zone_page_state(page, NR_ISOLATED_ANON +
95 page_is_file_cache(page));
96 if (unlikely(isolated_balloon_page(page)))
97 balloon_page_putback(page);
99 putback_lru_page(page);
104 * Restore a potential migration pte to a working pte entry
106 static int remove_migration_pte(struct page *new, struct vm_area_struct *vma,
107 unsigned long addr, void *old)
109 struct mm_struct *mm = vma->vm_mm;
115 if (unlikely(PageHuge(new))) {
116 ptep = huge_pte_offset(mm, addr);
119 ptl = huge_pte_lockptr(hstate_vma(vma), mm, ptep);
121 pmd = mm_find_pmd(mm, addr);
125 ptep = pte_offset_map(pmd, addr);
128 * Peek to check is_swap_pte() before taking ptlock? No, we
129 * can race mremap's move_ptes(), which skips anon_vma lock.
132 ptl = pte_lockptr(mm, pmd);
137 if (!is_swap_pte(pte))
140 entry = pte_to_swp_entry(pte);
142 if (!is_migration_entry(entry) ||
143 migration_entry_to_page(entry) != old)
147 pte = pte_mkold(mk_pte(new, vma->vm_page_prot));
148 if (pte_swp_soft_dirty(*ptep))
149 pte = pte_mksoft_dirty(pte);
151 /* Recheck VMA as permissions can change since migration started */
152 if (is_write_migration_entry(entry))
153 pte = maybe_mkwrite(pte, vma);
155 #ifdef CONFIG_HUGETLB_PAGE
157 pte = pte_mkhuge(pte);
158 pte = arch_make_huge_pte(pte, vma, new, 0);
161 flush_dcache_page(new);
162 set_pte_at(mm, addr, ptep, pte);
166 hugepage_add_anon_rmap(new, vma, addr);
169 } else if (PageAnon(new))
170 page_add_anon_rmap(new, vma, addr);
172 page_add_file_rmap(new);
174 /* No need to invalidate - it was non-present before */
175 update_mmu_cache(vma, addr, ptep);
177 pte_unmap_unlock(ptep, ptl);
183 * Get rid of all migration entries and replace them by
184 * references to the indicated page.
186 static void remove_migration_ptes(struct page *old, struct page *new)
188 struct rmap_walk_control rwc = {
189 .rmap_one = remove_migration_pte,
193 rmap_walk(new, &rwc);
197 * Something used the pte of a page under migration. We need to
198 * get to the page and wait until migration is finished.
199 * When we return from this function the fault will be retried.
201 void __migration_entry_wait(struct mm_struct *mm, pte_t *ptep,
210 if (!is_swap_pte(pte))
213 entry = pte_to_swp_entry(pte);
214 if (!is_migration_entry(entry))
217 page = migration_entry_to_page(entry);
220 * Once radix-tree replacement of page migration started, page_count
221 * *must* be zero. And, we don't want to call wait_on_page_locked()
222 * against a page without get_page().
223 * So, we use get_page_unless_zero(), here. Even failed, page fault
226 if (!get_page_unless_zero(page))
228 pte_unmap_unlock(ptep, ptl);
229 wait_on_page_locked(page);
233 pte_unmap_unlock(ptep, ptl);
236 void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
237 unsigned long address)
239 spinlock_t *ptl = pte_lockptr(mm, pmd);
240 pte_t *ptep = pte_offset_map(pmd, address);
241 __migration_entry_wait(mm, ptep, ptl);
244 void migration_entry_wait_huge(struct vm_area_struct *vma,
245 struct mm_struct *mm, pte_t *pte)
247 spinlock_t *ptl = huge_pte_lockptr(hstate_vma(vma), mm, pte);
248 __migration_entry_wait(mm, pte, ptl);
252 /* Returns true if all buffers are successfully locked */
253 static bool buffer_migrate_lock_buffers(struct buffer_head *head,
254 enum migrate_mode mode)
256 struct buffer_head *bh = head;
258 /* Simple case, sync compaction */
259 if (mode != MIGRATE_ASYNC) {
263 bh = bh->b_this_page;
265 } while (bh != head);
270 /* async case, we cannot block on lock_buffer so use trylock_buffer */
273 if (!trylock_buffer(bh)) {
275 * We failed to lock the buffer and cannot stall in
276 * async migration. Release the taken locks
278 struct buffer_head *failed_bh = bh;
281 while (bh != failed_bh) {
284 bh = bh->b_this_page;
289 bh = bh->b_this_page;
290 } while (bh != head);
294 static inline bool buffer_migrate_lock_buffers(struct buffer_head *head,
295 enum migrate_mode mode)
299 #endif /* CONFIG_BLOCK */
302 * Replace the page in the mapping.
304 * The number of remaining references must be:
305 * 1 for anonymous pages without a mapping
306 * 2 for pages with a mapping
307 * 3 for pages with a mapping and PagePrivate/PagePrivate2 set.
309 int migrate_page_move_mapping(struct address_space *mapping,
310 struct page *newpage, struct page *page,
311 struct buffer_head *head, enum migrate_mode mode,
314 int expected_count = 1 + extra_count;
318 /* Anonymous page without mapping */
319 if (page_count(page) != expected_count)
321 return MIGRATEPAGE_SUCCESS;
324 spin_lock_irq(&mapping->tree_lock);
326 pslot = radix_tree_lookup_slot(&mapping->page_tree,
329 expected_count += 1 + page_has_private(page);
330 if (page_count(page) != expected_count ||
331 radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) {
332 spin_unlock_irq(&mapping->tree_lock);
336 if (!page_freeze_refs(page, expected_count)) {
337 spin_unlock_irq(&mapping->tree_lock);
342 * In the async migration case of moving a page with buffers, lock the
343 * buffers using trylock before the mapping is moved. If the mapping
344 * was moved, we later failed to lock the buffers and could not move
345 * the mapping back due to an elevated page count, we would have to
346 * block waiting on other references to be dropped.
348 if (mode == MIGRATE_ASYNC && head &&
349 !buffer_migrate_lock_buffers(head, mode)) {
350 page_unfreeze_refs(page, expected_count);
351 spin_unlock_irq(&mapping->tree_lock);
356 * Now we know that no one else is looking at the page.
358 get_page(newpage); /* add cache reference */
359 if (PageSwapCache(page)) {
360 SetPageSwapCache(newpage);
361 set_page_private(newpage, page_private(page));
364 radix_tree_replace_slot(pslot, newpage);
367 * Drop cache reference from old page by unfreezing
368 * to one less reference.
369 * We know this isn't the last reference.
371 page_unfreeze_refs(page, expected_count - 1);
374 * If moved to a different zone then also account
375 * the page for that zone. Other VM counters will be
376 * taken care of when we establish references to the
377 * new page and drop references to the old page.
379 * Note that anonymous pages are accounted for
380 * via NR_FILE_PAGES and NR_ANON_PAGES if they
381 * are mapped to swap space.
383 __dec_zone_page_state(page, NR_FILE_PAGES);
384 __inc_zone_page_state(newpage, NR_FILE_PAGES);
385 if (!PageSwapCache(page) && PageSwapBacked(page)) {
386 __dec_zone_page_state(page, NR_SHMEM);
387 __inc_zone_page_state(newpage, NR_SHMEM);
389 spin_unlock_irq(&mapping->tree_lock);
391 return MIGRATEPAGE_SUCCESS;
395 * The expected number of remaining references is the same as that
396 * of migrate_page_move_mapping().
398 int migrate_huge_page_move_mapping(struct address_space *mapping,
399 struct page *newpage, struct page *page)
405 if (page_count(page) != 1)
407 return MIGRATEPAGE_SUCCESS;
410 spin_lock_irq(&mapping->tree_lock);
412 pslot = radix_tree_lookup_slot(&mapping->page_tree,
415 expected_count = 2 + page_has_private(page);
416 if (page_count(page) != expected_count ||
417 radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) {
418 spin_unlock_irq(&mapping->tree_lock);
422 if (!page_freeze_refs(page, expected_count)) {
423 spin_unlock_irq(&mapping->tree_lock);
429 radix_tree_replace_slot(pslot, newpage);
431 page_unfreeze_refs(page, expected_count - 1);
433 spin_unlock_irq(&mapping->tree_lock);
434 return MIGRATEPAGE_SUCCESS;
438 * Gigantic pages are so large that we do not guarantee that page++ pointer
439 * arithmetic will work across the entire page. We need something more
442 static void __copy_gigantic_page(struct page *dst, struct page *src,
446 struct page *dst_base = dst;
447 struct page *src_base = src;
449 for (i = 0; i < nr_pages; ) {
451 copy_highpage(dst, src);
454 dst = mem_map_next(dst, dst_base, i);
455 src = mem_map_next(src, src_base, i);
459 static void copy_huge_page(struct page *dst, struct page *src)
466 struct hstate *h = page_hstate(src);
467 nr_pages = pages_per_huge_page(h);
469 if (unlikely(nr_pages > MAX_ORDER_NR_PAGES)) {
470 __copy_gigantic_page(dst, src, nr_pages);
475 BUG_ON(!PageTransHuge(src));
476 nr_pages = hpage_nr_pages(src);
479 for (i = 0; i < nr_pages; i++) {
481 copy_highpage(dst + i, src + i);
486 * Copy the page to its new location
488 void migrate_page_copy(struct page *newpage, struct page *page)
492 if (PageHuge(page) || PageTransHuge(page))
493 copy_huge_page(newpage, page);
495 copy_highpage(newpage, page);
498 SetPageError(newpage);
499 if (PageReferenced(page))
500 SetPageReferenced(newpage);
501 if (PageUptodate(page))
502 SetPageUptodate(newpage);
503 if (TestClearPageActive(page)) {
504 VM_BUG_ON_PAGE(PageUnevictable(page), page);
505 SetPageActive(newpage);
506 } else if (TestClearPageUnevictable(page))
507 SetPageUnevictable(newpage);
508 if (PageChecked(page))
509 SetPageChecked(newpage);
510 if (PageMappedToDisk(page))
511 SetPageMappedToDisk(newpage);
513 if (PageDirty(page)) {
514 clear_page_dirty_for_io(page);
516 * Want to mark the page and the radix tree as dirty, and
517 * redo the accounting that clear_page_dirty_for_io undid,
518 * but we can't use set_page_dirty because that function
519 * is actually a signal that all of the page has become dirty.
520 * Whereas only part of our page may be dirty.
522 if (PageSwapBacked(page))
523 SetPageDirty(newpage);
525 __set_page_dirty_nobuffers(newpage);
528 if (page_is_young(page))
529 set_page_young(newpage);
530 if (page_is_idle(page))
531 set_page_idle(newpage);
534 * Copy NUMA information to the new page, to prevent over-eager
535 * future migrations of this same page.
537 cpupid = page_cpupid_xchg_last(page, -1);
538 page_cpupid_xchg_last(newpage, cpupid);
540 mlock_migrate_page(newpage, page);
541 ksm_migrate_page(newpage, page);
543 * Please do not reorder this without considering how mm/ksm.c's
544 * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache().
546 if (PageSwapCache(page))
547 ClearPageSwapCache(page);
548 ClearPagePrivate(page);
549 set_page_private(page, 0);
552 * If any waiters have accumulated on the new page then
555 if (PageWriteback(newpage))
556 end_page_writeback(newpage);
559 /************************************************************
560 * Migration functions
561 ***********************************************************/
564 * Common logic to directly migrate a single page suitable for
565 * pages that do not use PagePrivate/PagePrivate2.
567 * Pages are locked upon entry and exit.
569 int migrate_page(struct address_space *mapping,
570 struct page *newpage, struct page *page,
571 enum migrate_mode mode)
575 BUG_ON(PageWriteback(page)); /* Writeback must be complete */
577 rc = migrate_page_move_mapping(mapping, newpage, page, NULL, mode, 0);
579 if (rc != MIGRATEPAGE_SUCCESS)
582 migrate_page_copy(newpage, page);
583 return MIGRATEPAGE_SUCCESS;
585 EXPORT_SYMBOL(migrate_page);
589 * Migration function for pages with buffers. This function can only be used
590 * if the underlying filesystem guarantees that no other references to "page"
593 int buffer_migrate_page(struct address_space *mapping,
594 struct page *newpage, struct page *page, enum migrate_mode mode)
596 struct buffer_head *bh, *head;
599 if (!page_has_buffers(page))
600 return migrate_page(mapping, newpage, page, mode);
602 head = page_buffers(page);
604 rc = migrate_page_move_mapping(mapping, newpage, page, head, mode, 0);
606 if (rc != MIGRATEPAGE_SUCCESS)
610 * In the async case, migrate_page_move_mapping locked the buffers
611 * with an IRQ-safe spinlock held. In the sync case, the buffers
612 * need to be locked now
614 if (mode != MIGRATE_ASYNC)
615 BUG_ON(!buffer_migrate_lock_buffers(head, mode));
617 ClearPagePrivate(page);
618 set_page_private(newpage, page_private(page));
619 set_page_private(page, 0);
625 set_bh_page(bh, newpage, bh_offset(bh));
626 bh = bh->b_this_page;
628 } while (bh != head);
630 SetPagePrivate(newpage);
632 migrate_page_copy(newpage, page);
638 bh = bh->b_this_page;
640 } while (bh != head);
642 return MIGRATEPAGE_SUCCESS;
644 EXPORT_SYMBOL(buffer_migrate_page);
648 * Writeback a page to clean the dirty state
650 static int writeout(struct address_space *mapping, struct page *page)
652 struct writeback_control wbc = {
653 .sync_mode = WB_SYNC_NONE,
656 .range_end = LLONG_MAX,
661 if (!mapping->a_ops->writepage)
662 /* No write method for the address space */
665 if (!clear_page_dirty_for_io(page))
666 /* Someone else already triggered a write */
670 * A dirty page may imply that the underlying filesystem has
671 * the page on some queue. So the page must be clean for
672 * migration. Writeout may mean we loose the lock and the
673 * page state is no longer what we checked for earlier.
674 * At this point we know that the migration attempt cannot
677 remove_migration_ptes(page, page);
679 rc = mapping->a_ops->writepage(page, &wbc);
681 if (rc != AOP_WRITEPAGE_ACTIVATE)
682 /* unlocked. Relock */
685 return (rc < 0) ? -EIO : -EAGAIN;
689 * Default handling if a filesystem does not provide a migration function.
691 static int fallback_migrate_page(struct address_space *mapping,
692 struct page *newpage, struct page *page, enum migrate_mode mode)
694 if (PageDirty(page)) {
695 /* Only writeback pages in full synchronous migration */
696 if (mode != MIGRATE_SYNC)
698 return writeout(mapping, page);
702 * Buffers may be managed in a filesystem specific way.
703 * We must have no buffers or drop them.
705 if (page_has_private(page) &&
706 !try_to_release_page(page, GFP_KERNEL))
709 return migrate_page(mapping, newpage, page, mode);
713 * Move a page to a newly allocated page
714 * The page is locked and all ptes have been successfully removed.
716 * The new page will have replaced the old page if this function
721 * MIGRATEPAGE_SUCCESS - success
723 static int move_to_new_page(struct page *newpage, struct page *page,
724 int page_was_mapped, enum migrate_mode mode)
726 struct address_space *mapping;
730 * Block others from accessing the page when we get around to
731 * establishing additional references. We are the only one
732 * holding a reference to the new page at this point.
734 if (!trylock_page(newpage))
737 /* Prepare mapping for the new page.*/
738 newpage->index = page->index;
739 newpage->mapping = page->mapping;
740 if (PageSwapBacked(page))
741 SetPageSwapBacked(newpage);
744 * Indirectly called below, migrate_page_copy() copies PG_dirty and thus
745 * needs newpage's memcg set to transfer memcg dirty page accounting.
746 * So perform memcg migration in two steps:
747 * 1. set newpage->mem_cgroup (here)
748 * 2. clear page->mem_cgroup (below)
750 set_page_memcg(newpage, page_memcg(page));
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 set_page_memcg(newpage, NULL);
769 newpage->mapping = NULL;
771 set_page_memcg(page, NULL);
773 remove_migration_ptes(page, newpage);
774 page->mapping = NULL;
777 unlock_page(newpage);
782 static int __unmap_and_move(struct page *page, struct page *newpage,
783 int force, enum migrate_mode mode)
786 int page_was_mapped = 0;
787 struct anon_vma *anon_vma = NULL;
789 if (!trylock_page(page)) {
790 if (!force || mode == MIGRATE_ASYNC)
794 * It's not safe for direct compaction to call lock_page.
795 * For example, during page readahead pages are added locked
796 * to the LRU. Later, when the IO completes the pages are
797 * marked uptodate and unlocked. However, the queueing
798 * could be merging multiple pages for one bio (e.g.
799 * mpage_readpages). If an allocation happens for the
800 * second or third page, the process can end up locking
801 * the same page twice and deadlocking. Rather than
802 * trying to be clever about what pages can be locked,
803 * avoid the use of lock_page for direct compaction
806 if (current->flags & PF_MEMALLOC)
812 if (PageWriteback(page)) {
814 * Only in the case of a full synchronous migration is it
815 * necessary to wait for PageWriteback. In the async case,
816 * the retry loop is too short and in the sync-light case,
817 * the overhead of stalling is too much
819 if (mode != MIGRATE_SYNC) {
825 wait_on_page_writeback(page);
828 * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
829 * we cannot notice that anon_vma is freed while we migrates a page.
830 * This get_anon_vma() delays freeing anon_vma pointer until the end
831 * of migration. File cache pages are no problem because of page_lock()
832 * File Caches may use write_page() or lock_page() in migration, then,
833 * just care Anon page here.
835 if (PageAnon(page) && !PageKsm(page)) {
837 * Only page_lock_anon_vma_read() understands the subtleties of
838 * getting a hold on an anon_vma from outside one of its mms.
840 anon_vma = page_get_anon_vma(page);
845 } else if (PageSwapCache(page)) {
847 * We cannot be sure that the anon_vma of an unmapped
848 * swapcache page is safe to use because we don't
849 * know in advance if the VMA that this page belonged
850 * to still exists. If the VMA and others sharing the
851 * data have been freed, then the anon_vma could
852 * already be invalid.
854 * To avoid this possibility, swapcache pages get
855 * migrated but are not remapped when migration
863 if (unlikely(isolated_balloon_page(page))) {
865 * A ballooned page does not need any special attention from
866 * physical to virtual reverse mapping procedures.
867 * Skip any attempt to unmap PTEs or to remap swap cache,
868 * in order to avoid burning cycles at rmap level, and perform
869 * the page migration right away (proteced by page lock).
871 rc = balloon_page_migrate(newpage, page, mode);
876 * Corner case handling:
877 * 1. When a new swap-cache page is read into, it is added to the LRU
878 * and treated as swapcache but it has no rmap yet.
879 * Calling try_to_unmap() against a page->mapping==NULL page will
880 * trigger a BUG. So handle it here.
881 * 2. An orphaned page (see truncate_complete_page) might have
882 * fs-private metadata. The page can be picked up due to memory
883 * offlining. Everywhere else except page reclaim, the page is
884 * invisible to the vm, so the page can not be migrated. So try to
885 * free the metadata, so the page can be freed.
887 if (!page->mapping) {
888 VM_BUG_ON_PAGE(PageAnon(page), page);
889 if (page_has_private(page)) {
890 try_to_free_buffers(page);
896 /* Establish migration ptes or remove ptes */
897 if (page_mapped(page)) {
899 TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
904 if (!page_mapped(page))
905 rc = move_to_new_page(newpage, page, page_was_mapped, mode);
907 if (rc && page_was_mapped)
908 remove_migration_ptes(page, page);
910 /* Drop an anon_vma reference if we took one */
912 put_anon_vma(anon_vma);
921 * gcc 4.7 and 4.8 on arm get an ICEs when inlining unmap_and_move(). Work
924 #if (GCC_VERSION >= 40700 && GCC_VERSION < 40900) && defined(CONFIG_ARM)
925 #define ICE_noinline noinline
931 * Obtain the lock on page, remove all ptes and migrate the page
932 * to the newly allocated page in newpage.
934 static ICE_noinline int unmap_and_move(new_page_t get_new_page,
935 free_page_t put_new_page,
936 unsigned long private, struct page *page,
937 int force, enum migrate_mode mode,
938 enum migrate_reason reason)
942 struct page *newpage = get_new_page(page, private, &result);
947 if (page_count(page) == 1) {
948 /* page was freed from under us. So we are done. */
952 if (unlikely(PageTransHuge(page)))
953 if (unlikely(split_huge_page(page)))
956 rc = __unmap_and_move(page, newpage, force, mode);
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 /* Soft-offlined page shouldn't go through lru cache list */
970 if (reason == MR_MEMORY_FAILURE) {
972 if (!test_set_page_hwpoison(page))
973 num_poisoned_pages_inc();
975 putback_lru_page(page);
979 * If migration was not successful and there's a freeing callback, use
980 * it. Otherwise, putback_lru_page() will drop the reference grabbed
983 if (rc != MIGRATEPAGE_SUCCESS && put_new_page) {
984 ClearPageSwapBacked(newpage);
985 put_new_page(newpage, private);
986 } else if (unlikely(__is_movable_balloon_page(newpage))) {
987 /* drop our reference, page already in the balloon */
990 putback_lru_page(newpage);
996 *result = page_to_nid(newpage);
1002 * Counterpart of unmap_and_move_page() for hugepage migration.
1004 * This function doesn't wait the completion of hugepage I/O
1005 * because there is no race between I/O and migration for hugepage.
1006 * Note that currently hugepage I/O occurs only in direct I/O
1007 * where no lock is held and PG_writeback is irrelevant,
1008 * and writeback status of all subpages are counted in the reference
1009 * count of the head page (i.e. if all subpages of a 2MB hugepage are
1010 * under direct I/O, the reference of the head page is 512 and a bit more.)
1011 * This means that when we try to migrate hugepage whose subpages are
1012 * doing direct I/O, some references remain after try_to_unmap() and
1013 * hugepage migration fails without data corruption.
1015 * There is also no race when direct I/O is issued on the page under migration,
1016 * because then pte is replaced with migration swap entry and direct I/O code
1017 * will wait in the page fault for migration to complete.
1019 static int unmap_and_move_huge_page(new_page_t get_new_page,
1020 free_page_t put_new_page, unsigned long private,
1021 struct page *hpage, int force,
1022 enum migrate_mode mode)
1026 int page_was_mapped = 0;
1027 struct page *new_hpage;
1028 struct anon_vma *anon_vma = NULL;
1031 * Movability of hugepages depends on architectures and hugepage size.
1032 * This check is necessary because some callers of hugepage migration
1033 * like soft offline and memory hotremove don't walk through page
1034 * tables or check whether the hugepage is pmd-based or not before
1035 * kicking migration.
1037 if (!hugepage_migration_supported(page_hstate(hpage))) {
1038 putback_active_hugepage(hpage);
1042 new_hpage = get_new_page(hpage, private, &result);
1048 if (!trylock_page(hpage)) {
1049 if (!force || mode != MIGRATE_SYNC)
1054 if (PageAnon(hpage))
1055 anon_vma = page_get_anon_vma(hpage);
1057 if (page_mapped(hpage)) {
1059 TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
1060 page_was_mapped = 1;
1063 if (!page_mapped(hpage))
1064 rc = move_to_new_page(new_hpage, hpage, page_was_mapped, mode);
1066 if (rc != MIGRATEPAGE_SUCCESS && page_was_mapped)
1067 remove_migration_ptes(hpage, hpage);
1070 put_anon_vma(anon_vma);
1072 if (rc == MIGRATEPAGE_SUCCESS)
1073 hugetlb_cgroup_migrate(hpage, new_hpage);
1078 putback_active_hugepage(hpage);
1081 * If migration was not successful and there's a freeing callback, use
1082 * it. Otherwise, put_page() will drop the reference grabbed during
1085 if (rc != MIGRATEPAGE_SUCCESS && put_new_page)
1086 put_new_page(new_hpage, private);
1088 putback_active_hugepage(new_hpage);
1094 *result = page_to_nid(new_hpage);
1100 * migrate_pages - migrate the pages specified in a list, to the free pages
1101 * supplied as the target for the page migration
1103 * @from: The list of pages to be migrated.
1104 * @get_new_page: The function used to allocate free pages to be used
1105 * as the target of the page migration.
1106 * @put_new_page: The function used to free target pages if migration
1107 * fails, or NULL if no special handling is necessary.
1108 * @private: Private data to be passed on to get_new_page()
1109 * @mode: The migration mode that specifies the constraints for
1110 * page migration, if any.
1111 * @reason: The reason for page migration.
1113 * The function returns after 10 attempts or if no pages are movable any more
1114 * because the list has become empty or no retryable pages exist any more.
1115 * The caller should call putback_lru_pages() to return pages to the LRU
1116 * or free list only if ret != 0.
1118 * Returns the number of pages that were not migrated, or an error code.
1120 int migrate_pages(struct list_head *from, new_page_t get_new_page,
1121 free_page_t put_new_page, unsigned long private,
1122 enum migrate_mode mode, int reason)
1126 int nr_succeeded = 0;
1130 int swapwrite = current->flags & PF_SWAPWRITE;
1134 current->flags |= PF_SWAPWRITE;
1136 for(pass = 0; pass < 10 && retry; pass++) {
1139 list_for_each_entry_safe(page, page2, from, lru) {
1143 rc = unmap_and_move_huge_page(get_new_page,
1144 put_new_page, private, page,
1147 rc = unmap_and_move(get_new_page, put_new_page,
1148 private, page, pass > 2, mode,
1157 case MIGRATEPAGE_SUCCESS:
1162 * Permanent failure (-EBUSY, -ENOSYS, etc.):
1163 * unlike -EAGAIN case, the failed page is
1164 * removed from migration page list and not
1165 * retried in the next outer loop.
1172 rc = nr_failed + retry;
1175 count_vm_events(PGMIGRATE_SUCCESS, nr_succeeded);
1177 count_vm_events(PGMIGRATE_FAIL, nr_failed);
1178 trace_mm_migrate_pages(nr_succeeded, nr_failed, mode, reason);
1181 current->flags &= ~PF_SWAPWRITE;
1188 * Move a list of individual pages
1190 struct page_to_node {
1197 static struct page *new_page_node(struct page *p, unsigned long private,
1200 struct page_to_node *pm = (struct page_to_node *)private;
1202 while (pm->node != MAX_NUMNODES && pm->page != p)
1205 if (pm->node == MAX_NUMNODES)
1208 *result = &pm->status;
1211 return alloc_huge_page_node(page_hstate(compound_head(p)),
1214 return __alloc_pages_node(pm->node,
1215 GFP_HIGHUSER_MOVABLE | __GFP_THISNODE, 0);
1219 * Move a set of pages as indicated in the pm array. The addr
1220 * field must be set to the virtual address of the page to be moved
1221 * and the node number must contain a valid target node.
1222 * The pm array ends with node = MAX_NUMNODES.
1224 static int do_move_page_to_node_array(struct mm_struct *mm,
1225 struct page_to_node *pm,
1229 struct page_to_node *pp;
1230 LIST_HEAD(pagelist);
1232 down_read(&mm->mmap_sem);
1235 * Build a list of pages to migrate
1237 for (pp = pm; pp->node != MAX_NUMNODES; pp++) {
1238 struct vm_area_struct *vma;
1242 vma = find_vma(mm, pp->addr);
1243 if (!vma || pp->addr < vma->vm_start || !vma_migratable(vma))
1246 /* FOLL_DUMP to ignore special (like zero) pages */
1247 page = follow_page(vma, pp->addr,
1248 FOLL_GET | FOLL_SPLIT | FOLL_DUMP);
1250 err = PTR_ERR(page);
1259 err = page_to_nid(page);
1261 if (err == pp->node)
1263 * Node already in the right place
1268 if (page_mapcount(page) > 1 &&
1272 if (PageHuge(page)) {
1274 isolate_huge_page(page, &pagelist);
1278 err = isolate_lru_page(page);
1280 list_add_tail(&page->lru, &pagelist);
1281 inc_zone_page_state(page, NR_ISOLATED_ANON +
1282 page_is_file_cache(page));
1286 * Either remove the duplicate refcount from
1287 * isolate_lru_page() or drop the page ref if it was
1296 if (!list_empty(&pagelist)) {
1297 err = migrate_pages(&pagelist, new_page_node, NULL,
1298 (unsigned long)pm, MIGRATE_SYNC, MR_SYSCALL);
1300 putback_movable_pages(&pagelist);
1303 up_read(&mm->mmap_sem);
1308 * Migrate an array of page address onto an array of nodes and fill
1309 * the corresponding array of status.
1311 static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes,
1312 unsigned long nr_pages,
1313 const void __user * __user *pages,
1314 const int __user *nodes,
1315 int __user *status, int flags)
1317 struct page_to_node *pm;
1318 unsigned long chunk_nr_pages;
1319 unsigned long chunk_start;
1323 pm = (struct page_to_node *)__get_free_page(GFP_KERNEL);
1330 * Store a chunk of page_to_node array in a page,
1331 * but keep the last one as a marker
1333 chunk_nr_pages = (PAGE_SIZE / sizeof(struct page_to_node)) - 1;
1335 for (chunk_start = 0;
1336 chunk_start < nr_pages;
1337 chunk_start += chunk_nr_pages) {
1340 if (chunk_start + chunk_nr_pages > nr_pages)
1341 chunk_nr_pages = nr_pages - chunk_start;
1343 /* fill the chunk pm with addrs and nodes from user-space */
1344 for (j = 0; j < chunk_nr_pages; j++) {
1345 const void __user *p;
1349 if (get_user(p, pages + j + chunk_start))
1351 pm[j].addr = (unsigned long) p;
1353 if (get_user(node, nodes + j + chunk_start))
1357 if (node < 0 || node >= MAX_NUMNODES)
1360 if (!node_state(node, N_MEMORY))
1364 if (!node_isset(node, task_nodes))
1370 /* End marker for this chunk */
1371 pm[chunk_nr_pages].node = MAX_NUMNODES;
1373 /* Migrate this chunk */
1374 err = do_move_page_to_node_array(mm, pm,
1375 flags & MPOL_MF_MOVE_ALL);
1379 /* Return status information */
1380 for (j = 0; j < chunk_nr_pages; j++)
1381 if (put_user(pm[j].status, status + j + chunk_start)) {
1389 free_page((unsigned long)pm);
1395 * Determine the nodes of an array of pages and store it in an array of status.
1397 static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages,
1398 const void __user **pages, int *status)
1402 down_read(&mm->mmap_sem);
1404 for (i = 0; i < nr_pages; i++) {
1405 unsigned long addr = (unsigned long)(*pages);
1406 struct vm_area_struct *vma;
1410 vma = find_vma(mm, addr);
1411 if (!vma || addr < vma->vm_start)
1414 /* FOLL_DUMP to ignore special (like zero) pages */
1415 page = follow_page(vma, addr, FOLL_DUMP);
1417 err = PTR_ERR(page);
1421 err = page ? page_to_nid(page) : -ENOENT;
1429 up_read(&mm->mmap_sem);
1433 * Determine the nodes of a user array of pages and store it in
1434 * a user array of status.
1436 static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages,
1437 const void __user * __user *pages,
1440 #define DO_PAGES_STAT_CHUNK_NR 16
1441 const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR];
1442 int chunk_status[DO_PAGES_STAT_CHUNK_NR];
1445 unsigned long chunk_nr;
1447 chunk_nr = nr_pages;
1448 if (chunk_nr > DO_PAGES_STAT_CHUNK_NR)
1449 chunk_nr = DO_PAGES_STAT_CHUNK_NR;
1451 if (copy_from_user(chunk_pages, pages, chunk_nr * sizeof(*chunk_pages)))
1454 do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status);
1456 if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status)))
1461 nr_pages -= chunk_nr;
1463 return nr_pages ? -EFAULT : 0;
1467 * Move a list of pages in the address space of the currently executing
1470 SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages,
1471 const void __user * __user *, pages,
1472 const int __user *, nodes,
1473 int __user *, status, int, flags)
1475 const struct cred *cred = current_cred(), *tcred;
1476 struct task_struct *task;
1477 struct mm_struct *mm;
1479 nodemask_t task_nodes;
1482 if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
1485 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1488 /* Find the mm_struct */
1490 task = pid ? find_task_by_vpid(pid) : current;
1495 get_task_struct(task);
1498 * Check if this process has the right to modify the specified
1499 * process. The right exists if the process has administrative
1500 * capabilities, superuser privileges or the same
1501 * userid as the target process.
1503 tcred = __task_cred(task);
1504 if (!uid_eq(cred->euid, tcred->suid) && !uid_eq(cred->euid, tcred->uid) &&
1505 !uid_eq(cred->uid, tcred->suid) && !uid_eq(cred->uid, tcred->uid) &&
1506 !capable(CAP_SYS_NICE)) {
1513 err = security_task_movememory(task);
1517 task_nodes = cpuset_mems_allowed(task);
1518 mm = get_task_mm(task);
1519 put_task_struct(task);
1525 err = do_pages_move(mm, task_nodes, nr_pages, pages,
1526 nodes, status, flags);
1528 err = do_pages_stat(mm, nr_pages, pages, status);
1534 put_task_struct(task);
1538 #ifdef CONFIG_NUMA_BALANCING
1540 * Returns true if this is a safe migration target node for misplaced NUMA
1541 * pages. Currently it only checks the watermarks which crude
1543 static bool migrate_balanced_pgdat(struct pglist_data *pgdat,
1544 unsigned long nr_migrate_pages)
1547 for (z = pgdat->nr_zones - 1; z >= 0; z--) {
1548 struct zone *zone = pgdat->node_zones + z;
1550 if (!populated_zone(zone))
1553 if (!zone_reclaimable(zone))
1556 /* Avoid waking kswapd by allocating pages_to_migrate pages. */
1557 if (!zone_watermark_ok(zone, 0,
1558 high_wmark_pages(zone) +
1567 static struct page *alloc_misplaced_dst_page(struct page *page,
1571 int nid = (int) data;
1572 struct page *newpage;
1574 newpage = __alloc_pages_node(nid,
1575 (GFP_HIGHUSER_MOVABLE |
1576 __GFP_THISNODE | __GFP_NOMEMALLOC |
1577 __GFP_NORETRY | __GFP_NOWARN) &
1584 * page migration rate limiting control.
1585 * Do not migrate more than @pages_to_migrate in a @migrate_interval_millisecs
1586 * window of time. Default here says do not migrate more than 1280M per second.
1588 static unsigned int migrate_interval_millisecs __read_mostly = 100;
1589 static unsigned int ratelimit_pages __read_mostly = 128 << (20 - PAGE_SHIFT);
1591 /* Returns true if the node is migrate rate-limited after the update */
1592 static bool numamigrate_update_ratelimit(pg_data_t *pgdat,
1593 unsigned long nr_pages)
1596 * Rate-limit the amount of data that is being migrated to a node.
1597 * Optimal placement is no good if the memory bus is saturated and
1598 * all the time is being spent migrating!
1600 if (time_after(jiffies, pgdat->numabalancing_migrate_next_window)) {
1601 spin_lock(&pgdat->numabalancing_migrate_lock);
1602 pgdat->numabalancing_migrate_nr_pages = 0;
1603 pgdat->numabalancing_migrate_next_window = jiffies +
1604 msecs_to_jiffies(migrate_interval_millisecs);
1605 spin_unlock(&pgdat->numabalancing_migrate_lock);
1607 if (pgdat->numabalancing_migrate_nr_pages > ratelimit_pages) {
1608 trace_mm_numa_migrate_ratelimit(current, pgdat->node_id,
1614 * This is an unlocked non-atomic update so errors are possible.
1615 * The consequences are failing to migrate when we potentiall should
1616 * have which is not severe enough to warrant locking. If it is ever
1617 * a problem, it can be converted to a per-cpu counter.
1619 pgdat->numabalancing_migrate_nr_pages += nr_pages;
1623 static int numamigrate_isolate_page(pg_data_t *pgdat, struct page *page)
1627 VM_BUG_ON_PAGE(compound_order(page) && !PageTransHuge(page), page);
1629 /* Avoid migrating to a node that is nearly full */
1630 if (!migrate_balanced_pgdat(pgdat, 1UL << compound_order(page)))
1633 if (isolate_lru_page(page))
1637 * migrate_misplaced_transhuge_page() skips page migration's usual
1638 * check on page_count(), so we must do it here, now that the page
1639 * has been isolated: a GUP pin, or any other pin, prevents migration.
1640 * The expected page count is 3: 1 for page's mapcount and 1 for the
1641 * caller's pin and 1 for the reference taken by isolate_lru_page().
1643 if (PageTransHuge(page) && page_count(page) != 3) {
1644 putback_lru_page(page);
1648 page_lru = page_is_file_cache(page);
1649 mod_zone_page_state(page_zone(page), NR_ISOLATED_ANON + page_lru,
1650 hpage_nr_pages(page));
1653 * Isolating the page has taken another reference, so the
1654 * caller's reference can be safely dropped without the page
1655 * disappearing underneath us during migration.
1661 bool pmd_trans_migrating(pmd_t pmd)
1663 struct page *page = pmd_page(pmd);
1664 return PageLocked(page);
1668 * Attempt to migrate a misplaced page to the specified destination
1669 * node. Caller is expected to have an elevated reference count on
1670 * the page that will be dropped by this function before returning.
1672 int migrate_misplaced_page(struct page *page, struct vm_area_struct *vma,
1675 pg_data_t *pgdat = NODE_DATA(node);
1678 LIST_HEAD(migratepages);
1681 * Don't migrate file pages that are mapped in multiple processes
1682 * with execute permissions as they are probably shared libraries.
1684 if (page_mapcount(page) != 1 && page_is_file_cache(page) &&
1685 (vma->vm_flags & VM_EXEC))
1689 * Rate-limit the amount of data that is being migrated to a node.
1690 * Optimal placement is no good if the memory bus is saturated and
1691 * all the time is being spent migrating!
1693 if (numamigrate_update_ratelimit(pgdat, 1))
1696 isolated = numamigrate_isolate_page(pgdat, page);
1700 list_add(&page->lru, &migratepages);
1701 nr_remaining = migrate_pages(&migratepages, alloc_misplaced_dst_page,
1702 NULL, node, MIGRATE_ASYNC,
1705 if (!list_empty(&migratepages)) {
1706 list_del(&page->lru);
1707 dec_zone_page_state(page, NR_ISOLATED_ANON +
1708 page_is_file_cache(page));
1709 putback_lru_page(page);
1713 count_vm_numa_event(NUMA_PAGE_MIGRATE);
1714 BUG_ON(!list_empty(&migratepages));
1721 #endif /* CONFIG_NUMA_BALANCING */
1723 #if defined(CONFIG_NUMA_BALANCING) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
1725 * Migrates a THP to a given target node. page must be locked and is unlocked
1728 int migrate_misplaced_transhuge_page(struct mm_struct *mm,
1729 struct vm_area_struct *vma,
1730 pmd_t *pmd, pmd_t entry,
1731 unsigned long address,
1732 struct page *page, int node)
1735 pg_data_t *pgdat = NODE_DATA(node);
1737 struct page *new_page = NULL;
1738 int page_lru = page_is_file_cache(page);
1739 unsigned long mmun_start = address & HPAGE_PMD_MASK;
1740 unsigned long mmun_end = mmun_start + HPAGE_PMD_SIZE;
1744 * Rate-limit the amount of data that is being migrated to a node.
1745 * Optimal placement is no good if the memory bus is saturated and
1746 * all the time is being spent migrating!
1748 if (numamigrate_update_ratelimit(pgdat, HPAGE_PMD_NR))
1751 new_page = alloc_pages_node(node,
1752 (GFP_TRANSHUGE | __GFP_THISNODE) & ~__GFP_WAIT,
1757 isolated = numamigrate_isolate_page(pgdat, page);
1763 if (mm_tlb_flush_pending(mm))
1764 flush_tlb_range(vma, mmun_start, mmun_end);
1766 /* Prepare a page as a migration target */
1767 __set_page_locked(new_page);
1768 SetPageSwapBacked(new_page);
1770 /* anon mapping, we can simply copy page->mapping to the new page: */
1771 new_page->mapping = page->mapping;
1772 new_page->index = page->index;
1773 migrate_page_copy(new_page, page);
1774 WARN_ON(PageLRU(new_page));
1776 /* Recheck the target PMD */
1777 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1778 ptl = pmd_lock(mm, pmd);
1779 if (unlikely(!pmd_same(*pmd, entry) || page_count(page) != 2)) {
1782 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1784 /* Reverse changes made by migrate_page_copy() */
1785 if (TestClearPageActive(new_page))
1786 SetPageActive(page);
1787 if (TestClearPageUnevictable(new_page))
1788 SetPageUnevictable(page);
1789 mlock_migrate_page(page, new_page);
1791 unlock_page(new_page);
1792 put_page(new_page); /* Free it */
1794 /* Retake the callers reference and putback on LRU */
1796 putback_lru_page(page);
1797 mod_zone_page_state(page_zone(page),
1798 NR_ISOLATED_ANON + page_lru, -HPAGE_PMD_NR);
1804 entry = mk_pmd(new_page, vma->vm_page_prot);
1805 entry = pmd_mkhuge(entry);
1806 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1809 * Clear the old entry under pagetable lock and establish the new PTE.
1810 * Any parallel GUP will either observe the old page blocking on the
1811 * page lock, block on the page table lock or observe the new page.
1812 * The SetPageUptodate on the new page and page_add_new_anon_rmap
1813 * guarantee the copy is visible before the pagetable update.
1815 flush_cache_range(vma, mmun_start, mmun_end);
1816 page_add_anon_rmap(new_page, vma, mmun_start);
1817 pmdp_huge_clear_flush_notify(vma, mmun_start, pmd);
1818 set_pmd_at(mm, mmun_start, pmd, entry);
1819 flush_tlb_range(vma, mmun_start, mmun_end);
1820 update_mmu_cache_pmd(vma, address, &entry);
1822 if (page_count(page) != 2) {
1823 set_pmd_at(mm, mmun_start, pmd, orig_entry);
1824 flush_tlb_range(vma, mmun_start, mmun_end);
1825 mmu_notifier_invalidate_range(mm, mmun_start, mmun_end);
1826 update_mmu_cache_pmd(vma, address, &entry);
1827 page_remove_rmap(new_page);
1831 mem_cgroup_migrate(page, new_page, false);
1833 page_remove_rmap(page);
1836 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1838 /* Take an "isolate" reference and put new page on the LRU. */
1840 putback_lru_page(new_page);
1842 unlock_page(new_page);
1844 put_page(page); /* Drop the rmap reference */
1845 put_page(page); /* Drop the LRU isolation reference */
1847 count_vm_events(PGMIGRATE_SUCCESS, HPAGE_PMD_NR);
1848 count_vm_numa_events(NUMA_PAGE_MIGRATE, HPAGE_PMD_NR);
1850 mod_zone_page_state(page_zone(page),
1851 NR_ISOLATED_ANON + page_lru,
1856 count_vm_events(PGMIGRATE_FAIL, HPAGE_PMD_NR);
1858 ptl = pmd_lock(mm, pmd);
1859 if (pmd_same(*pmd, entry)) {
1860 entry = pmd_modify(entry, vma->vm_page_prot);
1861 set_pmd_at(mm, mmun_start, pmd, entry);
1862 update_mmu_cache_pmd(vma, address, &entry);
1871 #endif /* CONFIG_NUMA_BALANCING */
1873 #endif /* CONFIG_NUMA */