2 * Memory Migration functionality - linux/mm/migrate.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/syscalls.h>
34 #include <linux/hugetlb.h>
35 #include <linux/hugetlb_cgroup.h>
36 #include <linux/gfp.h>
37 #include <linux/balloon_compaction.h>
38 #include <linux/mmu_notifier.h>
39 #include <linux/page_idle.h>
41 #include <asm/tlbflush.h>
43 #define CREATE_TRACE_POINTS
44 #include <trace/events/migrate.h>
49 * migrate_prep() needs to be called before we start compiling a list of pages
50 * to be migrated using isolate_lru_page(). If scheduling work on other CPUs is
51 * undesirable, use migrate_prep_local()
53 int migrate_prep(void)
56 * Clear the LRU lists so pages can be isolated.
57 * Note that pages may be moved off the LRU after we have
58 * drained them. Those pages will fail to migrate like other
59 * pages that may be busy.
66 /* Do the necessary work of migrate_prep but not if it involves other CPUs */
67 int migrate_prep_local(void)
75 * Put previously isolated pages back onto the appropriate lists
76 * from where they were once taken off for compaction/migration.
78 * This function shall be used whenever the isolated pageset has been
79 * built from lru, balloon, hugetlbfs page. See isolate_migratepages_range()
80 * and isolate_huge_page().
82 void putback_movable_pages(struct list_head *l)
87 list_for_each_entry_safe(page, page2, l, lru) {
88 if (unlikely(PageHuge(page))) {
89 putback_active_hugepage(page);
93 dec_zone_page_state(page, NR_ISOLATED_ANON +
94 page_is_file_cache(page));
95 if (unlikely(isolated_balloon_page(page)))
96 balloon_page_putback(page);
98 putback_lru_page(page);
103 * Restore a potential migration pte to a working pte entry
105 static int remove_migration_pte(struct page *new, struct vm_area_struct *vma,
106 unsigned long addr, void *old)
108 struct mm_struct *mm = vma->vm_mm;
114 if (unlikely(PageHuge(new))) {
115 ptep = huge_pte_offset(mm, addr);
118 ptl = huge_pte_lockptr(hstate_vma(vma), mm, ptep);
120 pmd = mm_find_pmd(mm, addr);
124 ptep = pte_offset_map(pmd, addr);
127 * Peek to check is_swap_pte() before taking ptlock? No, we
128 * can race mremap's move_ptes(), which skips anon_vma lock.
131 ptl = pte_lockptr(mm, pmd);
136 if (!is_swap_pte(pte))
139 entry = pte_to_swp_entry(pte);
141 if (!is_migration_entry(entry) ||
142 migration_entry_to_page(entry) != old)
146 pte = pte_mkold(mk_pte(new, vma->vm_page_prot));
147 if (pte_swp_soft_dirty(*ptep))
148 pte = pte_mksoft_dirty(pte);
150 /* Recheck VMA as permissions can change since migration started */
151 if (is_write_migration_entry(entry))
152 pte = maybe_mkwrite(pte, vma);
154 #ifdef CONFIG_HUGETLB_PAGE
156 pte = pte_mkhuge(pte);
157 pte = arch_make_huge_pte(pte, vma, new, 0);
160 flush_dcache_page(new);
161 set_pte_at(mm, addr, ptep, pte);
165 hugepage_add_anon_rmap(new, vma, addr);
168 } else if (PageAnon(new))
169 page_add_anon_rmap(new, vma, addr);
171 page_add_file_rmap(new);
173 if (vma->vm_flags & VM_LOCKED)
176 /* No need to invalidate - it was non-present before */
177 update_mmu_cache(vma, addr, ptep);
179 pte_unmap_unlock(ptep, ptl);
185 * Get rid of all migration entries and replace them by
186 * references to the indicated page.
188 static void remove_migration_ptes(struct page *old, struct page *new)
190 struct rmap_walk_control rwc = {
191 .rmap_one = remove_migration_pte,
195 rmap_walk(new, &rwc);
199 * Something used the pte of a page under migration. We need to
200 * get to the page and wait until migration is finished.
201 * When we return from this function the fault will be retried.
203 void __migration_entry_wait(struct mm_struct *mm, pte_t *ptep,
212 if (!is_swap_pte(pte))
215 entry = pte_to_swp_entry(pte);
216 if (!is_migration_entry(entry))
219 page = migration_entry_to_page(entry);
222 * Once radix-tree replacement of page migration started, page_count
223 * *must* be zero. And, we don't want to call wait_on_page_locked()
224 * against a page without get_page().
225 * So, we use get_page_unless_zero(), here. Even failed, page fault
228 if (!get_page_unless_zero(page))
230 pte_unmap_unlock(ptep, ptl);
231 wait_on_page_locked(page);
235 pte_unmap_unlock(ptep, ptl);
238 void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
239 unsigned long address)
241 spinlock_t *ptl = pte_lockptr(mm, pmd);
242 pte_t *ptep = pte_offset_map(pmd, address);
243 __migration_entry_wait(mm, ptep, ptl);
246 void migration_entry_wait_huge(struct vm_area_struct *vma,
247 struct mm_struct *mm, pte_t *pte)
249 spinlock_t *ptl = huge_pte_lockptr(hstate_vma(vma), mm, pte);
250 __migration_entry_wait(mm, pte, ptl);
254 /* Returns true if all buffers are successfully locked */
255 static bool buffer_migrate_lock_buffers(struct buffer_head *head,
256 enum migrate_mode mode)
258 struct buffer_head *bh = head;
260 /* Simple case, sync compaction */
261 if (mode != MIGRATE_ASYNC) {
265 bh = bh->b_this_page;
267 } while (bh != head);
272 /* async case, we cannot block on lock_buffer so use trylock_buffer */
275 if (!trylock_buffer(bh)) {
277 * We failed to lock the buffer and cannot stall in
278 * async migration. Release the taken locks
280 struct buffer_head *failed_bh = bh;
283 while (bh != failed_bh) {
286 bh = bh->b_this_page;
291 bh = bh->b_this_page;
292 } while (bh != head);
296 static inline bool buffer_migrate_lock_buffers(struct buffer_head *head,
297 enum migrate_mode mode)
301 #endif /* CONFIG_BLOCK */
304 * Replace the page in the mapping.
306 * The number of remaining references must be:
307 * 1 for anonymous pages without a mapping
308 * 2 for pages with a mapping
309 * 3 for pages with a mapping and PagePrivate/PagePrivate2 set.
311 int migrate_page_move_mapping(struct address_space *mapping,
312 struct page *newpage, struct page *page,
313 struct buffer_head *head, enum migrate_mode mode,
316 int expected_count = 1 + extra_count;
320 /* Anonymous page without mapping */
321 if (page_count(page) != expected_count)
323 return MIGRATEPAGE_SUCCESS;
326 spin_lock_irq(&mapping->tree_lock);
328 pslot = radix_tree_lookup_slot(&mapping->page_tree,
331 expected_count += 1 + page_has_private(page);
332 if (page_count(page) != expected_count ||
333 radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) {
334 spin_unlock_irq(&mapping->tree_lock);
338 if (!page_freeze_refs(page, expected_count)) {
339 spin_unlock_irq(&mapping->tree_lock);
344 * In the async migration case of moving a page with buffers, lock the
345 * buffers using trylock before the mapping is moved. If the mapping
346 * was moved, we later failed to lock the buffers and could not move
347 * the mapping back due to an elevated page count, we would have to
348 * block waiting on other references to be dropped.
350 if (mode == MIGRATE_ASYNC && head &&
351 !buffer_migrate_lock_buffers(head, mode)) {
352 page_unfreeze_refs(page, expected_count);
353 spin_unlock_irq(&mapping->tree_lock);
358 * Now we know that no one else is looking at the page.
360 get_page(newpage); /* add cache reference */
361 if (PageSwapCache(page)) {
362 SetPageSwapCache(newpage);
363 set_page_private(newpage, page_private(page));
366 radix_tree_replace_slot(pslot, newpage);
369 * Drop cache reference from old page by unfreezing
370 * to one less reference.
371 * We know this isn't the last reference.
373 page_unfreeze_refs(page, expected_count - 1);
376 * If moved to a different zone then also account
377 * the page for that zone. Other VM counters will be
378 * taken care of when we establish references to the
379 * new page and drop references to the old page.
381 * Note that anonymous pages are accounted for
382 * via NR_FILE_PAGES and NR_ANON_PAGES if they
383 * are mapped to swap space.
385 __dec_zone_page_state(page, NR_FILE_PAGES);
386 __inc_zone_page_state(newpage, NR_FILE_PAGES);
387 if (!PageSwapCache(page) && PageSwapBacked(page)) {
388 __dec_zone_page_state(page, NR_SHMEM);
389 __inc_zone_page_state(newpage, NR_SHMEM);
391 spin_unlock_irq(&mapping->tree_lock);
393 return MIGRATEPAGE_SUCCESS;
397 * The expected number of remaining references is the same as that
398 * of migrate_page_move_mapping().
400 int migrate_huge_page_move_mapping(struct address_space *mapping,
401 struct page *newpage, struct page *page)
407 if (page_count(page) != 1)
409 return MIGRATEPAGE_SUCCESS;
412 spin_lock_irq(&mapping->tree_lock);
414 pslot = radix_tree_lookup_slot(&mapping->page_tree,
417 expected_count = 2 + page_has_private(page);
418 if (page_count(page) != expected_count ||
419 radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) {
420 spin_unlock_irq(&mapping->tree_lock);
424 if (!page_freeze_refs(page, expected_count)) {
425 spin_unlock_irq(&mapping->tree_lock);
431 radix_tree_replace_slot(pslot, newpage);
433 page_unfreeze_refs(page, expected_count - 1);
435 spin_unlock_irq(&mapping->tree_lock);
436 return MIGRATEPAGE_SUCCESS;
440 * Gigantic pages are so large that we do not guarantee that page++ pointer
441 * arithmetic will work across the entire page. We need something more
444 static void __copy_gigantic_page(struct page *dst, struct page *src,
448 struct page *dst_base = dst;
449 struct page *src_base = src;
451 for (i = 0; i < nr_pages; ) {
453 copy_highpage(dst, src);
456 dst = mem_map_next(dst, dst_base, i);
457 src = mem_map_next(src, src_base, i);
461 static void copy_huge_page(struct page *dst, struct page *src)
468 struct hstate *h = page_hstate(src);
469 nr_pages = pages_per_huge_page(h);
471 if (unlikely(nr_pages > MAX_ORDER_NR_PAGES)) {
472 __copy_gigantic_page(dst, src, nr_pages);
477 BUG_ON(!PageTransHuge(src));
478 nr_pages = hpage_nr_pages(src);
481 for (i = 0; i < nr_pages; i++) {
483 copy_highpage(dst + i, src + i);
488 * Copy the page to its new location
490 void migrate_page_copy(struct page *newpage, struct page *page)
494 if (PageHuge(page) || PageTransHuge(page))
495 copy_huge_page(newpage, page);
497 copy_highpage(newpage, page);
500 SetPageError(newpage);
501 if (PageReferenced(page))
502 SetPageReferenced(newpage);
503 if (PageUptodate(page))
504 SetPageUptodate(newpage);
505 if (TestClearPageActive(page)) {
506 VM_BUG_ON_PAGE(PageUnevictable(page), page);
507 SetPageActive(newpage);
508 } else if (TestClearPageUnevictable(page))
509 SetPageUnevictable(newpage);
510 if (PageChecked(page))
511 SetPageChecked(newpage);
512 if (PageMappedToDisk(page))
513 SetPageMappedToDisk(newpage);
515 if (PageDirty(page)) {
516 clear_page_dirty_for_io(page);
518 * Want to mark the page and the radix tree as dirty, and
519 * redo the accounting that clear_page_dirty_for_io undid,
520 * but we can't use set_page_dirty because that function
521 * is actually a signal that all of the page has become dirty.
522 * Whereas only part of our page may be dirty.
524 if (PageSwapBacked(page))
525 SetPageDirty(newpage);
527 __set_page_dirty_nobuffers(newpage);
530 if (page_is_young(page))
531 set_page_young(newpage);
532 if (page_is_idle(page))
533 set_page_idle(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 ksm_migrate_page(newpage, page);
544 * Please do not reorder this without considering how mm/ksm.c's
545 * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache().
547 if (PageSwapCache(page))
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 ***********************************************************/
565 * Common logic to directly migrate a single page suitable for
566 * pages that do not use PagePrivate/PagePrivate2.
568 * Pages are locked upon entry and exit.
570 int migrate_page(struct address_space *mapping,
571 struct page *newpage, struct page *page,
572 enum migrate_mode mode)
576 BUG_ON(PageWriteback(page)); /* Writeback must be complete */
578 rc = migrate_page_move_mapping(mapping, newpage, page, NULL, mode, 0);
580 if (rc != MIGRATEPAGE_SUCCESS)
583 migrate_page_copy(newpage, page);
584 return MIGRATEPAGE_SUCCESS;
586 EXPORT_SYMBOL(migrate_page);
590 * Migration function for pages with buffers. This function can only be used
591 * if the underlying filesystem guarantees that no other references to "page"
594 int buffer_migrate_page(struct address_space *mapping,
595 struct page *newpage, struct page *page, enum migrate_mode mode)
597 struct buffer_head *bh, *head;
600 if (!page_has_buffers(page))
601 return migrate_page(mapping, newpage, page, mode);
603 head = page_buffers(page);
605 rc = migrate_page_move_mapping(mapping, newpage, page, head, mode, 0);
607 if (rc != MIGRATEPAGE_SUCCESS)
611 * In the async case, migrate_page_move_mapping locked the buffers
612 * with an IRQ-safe spinlock held. In the sync case, the buffers
613 * need to be locked now
615 if (mode != MIGRATE_ASYNC)
616 BUG_ON(!buffer_migrate_lock_buffers(head, mode));
618 ClearPagePrivate(page);
619 set_page_private(newpage, page_private(page));
620 set_page_private(page, 0);
626 set_bh_page(bh, newpage, bh_offset(bh));
627 bh = bh->b_this_page;
629 } while (bh != head);
631 SetPagePrivate(newpage);
633 migrate_page_copy(newpage, page);
639 bh = bh->b_this_page;
641 } while (bh != head);
643 return MIGRATEPAGE_SUCCESS;
645 EXPORT_SYMBOL(buffer_migrate_page);
649 * Writeback a page to clean the dirty state
651 static int writeout(struct address_space *mapping, struct page *page)
653 struct writeback_control wbc = {
654 .sync_mode = WB_SYNC_NONE,
657 .range_end = LLONG_MAX,
662 if (!mapping->a_ops->writepage)
663 /* No write method for the address space */
666 if (!clear_page_dirty_for_io(page))
667 /* Someone else already triggered a write */
671 * A dirty page may imply that the underlying filesystem has
672 * the page on some queue. So the page must be clean for
673 * migration. Writeout may mean we loose the lock and the
674 * page state is no longer what we checked for earlier.
675 * At this point we know that the migration attempt cannot
678 remove_migration_ptes(page, page);
680 rc = mapping->a_ops->writepage(page, &wbc);
682 if (rc != AOP_WRITEPAGE_ACTIVATE)
683 /* unlocked. Relock */
686 return (rc < 0) ? -EIO : -EAGAIN;
690 * Default handling if a filesystem does not provide a migration function.
692 static int fallback_migrate_page(struct address_space *mapping,
693 struct page *newpage, struct page *page, enum migrate_mode mode)
695 if (PageDirty(page)) {
696 /* Only writeback pages in full synchronous migration */
697 if (mode != MIGRATE_SYNC)
699 return writeout(mapping, page);
703 * Buffers may be managed in a filesystem specific way.
704 * We must have no buffers or drop them.
706 if (page_has_private(page) &&
707 !try_to_release_page(page, GFP_KERNEL))
710 return migrate_page(mapping, newpage, page, mode);
714 * Move a page to a newly allocated page
715 * The page is locked and all ptes have been successfully removed.
717 * The new page will have replaced the old page if this function
722 * MIGRATEPAGE_SUCCESS - success
724 static int move_to_new_page(struct page *newpage, struct page *page,
725 int page_was_mapped, enum migrate_mode mode)
727 struct address_space *mapping;
731 * Block others from accessing the page when we get around to
732 * establishing additional references. We are the only one
733 * holding a reference to the new page at this point.
735 if (!trylock_page(newpage))
738 /* Prepare mapping for the new page.*/
739 newpage->index = page->index;
740 newpage->mapping = page->mapping;
741 if (PageSwapBacked(page))
742 SetPageSwapBacked(newpage);
745 * Indirectly called below, migrate_page_copy() copies PG_dirty and thus
746 * needs newpage's memcg set to transfer memcg dirty page accounting.
747 * So perform memcg migration in two steps:
748 * 1. set newpage->mem_cgroup (here)
749 * 2. clear page->mem_cgroup (below)
751 set_page_memcg(newpage, page_memcg(page));
753 mapping = page_mapping(page);
755 rc = migrate_page(mapping, newpage, page, mode);
756 else if (mapping->a_ops->migratepage)
758 * Most pages have a mapping and most filesystems provide a
759 * migratepage callback. Anonymous pages are part of swap
760 * space which also has its own migratepage callback. This
761 * is the most common path for page migration.
763 rc = mapping->a_ops->migratepage(mapping,
764 newpage, page, mode);
766 rc = fallback_migrate_page(mapping, newpage, page, mode);
768 if (rc != MIGRATEPAGE_SUCCESS) {
769 set_page_memcg(newpage, NULL);
770 newpage->mapping = NULL;
772 set_page_memcg(page, NULL);
774 remove_migration_ptes(page, newpage);
775 page->mapping = NULL;
778 unlock_page(newpage);
783 static int __unmap_and_move(struct page *page, struct page *newpage,
784 int force, enum migrate_mode mode)
787 int page_was_mapped = 0;
788 struct anon_vma *anon_vma = NULL;
790 if (!trylock_page(page)) {
791 if (!force || mode == MIGRATE_ASYNC)
795 * It's not safe for direct compaction to call lock_page.
796 * For example, during page readahead pages are added locked
797 * to the LRU. Later, when the IO completes the pages are
798 * marked uptodate and unlocked. However, the queueing
799 * could be merging multiple pages for one bio (e.g.
800 * mpage_readpages). If an allocation happens for the
801 * second or third page, the process can end up locking
802 * the same page twice and deadlocking. Rather than
803 * trying to be clever about what pages can be locked,
804 * avoid the use of lock_page for direct compaction
807 if (current->flags & PF_MEMALLOC)
813 if (PageWriteback(page)) {
815 * Only in the case of a full synchronous migration is it
816 * necessary to wait for PageWriteback. In the async case,
817 * the retry loop is too short and in the sync-light case,
818 * the overhead of stalling is too much
820 if (mode != MIGRATE_SYNC) {
826 wait_on_page_writeback(page);
829 * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
830 * we cannot notice that anon_vma is freed while we migrates a page.
831 * This get_anon_vma() delays freeing anon_vma pointer until the end
832 * of migration. File cache pages are no problem because of page_lock()
833 * File Caches may use write_page() or lock_page() in migration, then,
834 * just care Anon page here.
836 if (PageAnon(page) && !PageKsm(page)) {
838 * Only page_lock_anon_vma_read() understands the subtleties of
839 * getting a hold on an anon_vma from outside one of its mms.
841 anon_vma = page_get_anon_vma(page);
846 } else if (PageSwapCache(page)) {
848 * We cannot be sure that the anon_vma of an unmapped
849 * swapcache page is safe to use because we don't
850 * know in advance if the VMA that this page belonged
851 * to still exists. If the VMA and others sharing the
852 * data have been freed, then the anon_vma could
853 * already be invalid.
855 * To avoid this possibility, swapcache pages get
856 * migrated but are not remapped when migration
864 if (unlikely(isolated_balloon_page(page))) {
866 * A ballooned page does not need any special attention from
867 * physical to virtual reverse mapping procedures.
868 * Skip any attempt to unmap PTEs or to remap swap cache,
869 * in order to avoid burning cycles at rmap level, and perform
870 * the page migration right away (proteced by page lock).
872 rc = balloon_page_migrate(newpage, page, mode);
877 * Corner case handling:
878 * 1. When a new swap-cache page is read into, it is added to the LRU
879 * and treated as swapcache but it has no rmap yet.
880 * Calling try_to_unmap() against a page->mapping==NULL page will
881 * trigger a BUG. So handle it here.
882 * 2. An orphaned page (see truncate_complete_page) might have
883 * fs-private metadata. The page can be picked up due to memory
884 * offlining. Everywhere else except page reclaim, the page is
885 * invisible to the vm, so the page can not be migrated. So try to
886 * free the metadata, so the page can be freed.
888 if (!page->mapping) {
889 VM_BUG_ON_PAGE(PageAnon(page), page);
890 if (page_has_private(page)) {
891 try_to_free_buffers(page);
897 /* Establish migration ptes or remove ptes */
898 if (page_mapped(page)) {
900 TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
905 if (!page_mapped(page))
906 rc = move_to_new_page(newpage, page, page_was_mapped, mode);
908 if (rc && page_was_mapped)
909 remove_migration_ptes(page, page);
911 /* Drop an anon_vma reference if we took one */
913 put_anon_vma(anon_vma);
922 * gcc 4.7 and 4.8 on arm get an ICEs when inlining unmap_and_move(). Work
925 #if (GCC_VERSION >= 40700 && GCC_VERSION < 40900) && defined(CONFIG_ARM)
926 #define ICE_noinline noinline
932 * Obtain the lock on page, remove all ptes and migrate the page
933 * to the newly allocated page in newpage.
935 static ICE_noinline int unmap_and_move(new_page_t get_new_page,
936 free_page_t put_new_page,
937 unsigned long private, struct page *page,
938 int force, enum migrate_mode mode,
939 enum migrate_reason reason)
941 int rc = MIGRATEPAGE_SUCCESS;
943 struct page *newpage;
945 newpage = get_new_page(page, private, &result);
949 if (page_count(page) == 1) {
950 /* page was freed from under us. So we are done. */
954 if (unlikely(PageTransHuge(page)))
955 if (unlikely(split_huge_page(page)))
958 rc = __unmap_and_move(page, newpage, force, mode);
959 if (rc == MIGRATEPAGE_SUCCESS)
965 * A page that has been migrated has all references
966 * removed and will be freed. A page that has not been
967 * migrated will have kepts its references and be
970 list_del(&page->lru);
971 dec_zone_page_state(page, NR_ISOLATED_ANON +
972 page_is_file_cache(page));
973 /* Soft-offlined page shouldn't go through lru cache list */
974 if (reason == MR_MEMORY_FAILURE) {
976 if (!test_set_page_hwpoison(page))
977 num_poisoned_pages_inc();
979 putback_lru_page(page);
983 * If migration was not successful and there's a freeing callback, use
984 * it. Otherwise, putback_lru_page() will drop the reference grabbed
988 ClearPageSwapBacked(newpage);
989 put_new_page(newpage, private);
990 } else if (unlikely(__is_movable_balloon_page(newpage))) {
991 /* drop our reference, page already in the balloon */
994 putback_lru_page(newpage);
1000 *result = page_to_nid(newpage);
1006 * Counterpart of unmap_and_move_page() for hugepage migration.
1008 * This function doesn't wait the completion of hugepage I/O
1009 * because there is no race between I/O and migration for hugepage.
1010 * Note that currently hugepage I/O occurs only in direct I/O
1011 * where no lock is held and PG_writeback is irrelevant,
1012 * and writeback status of all subpages are counted in the reference
1013 * count of the head page (i.e. if all subpages of a 2MB hugepage are
1014 * under direct I/O, the reference of the head page is 512 and a bit more.)
1015 * This means that when we try to migrate hugepage whose subpages are
1016 * doing direct I/O, some references remain after try_to_unmap() and
1017 * hugepage migration fails without data corruption.
1019 * There is also no race when direct I/O is issued on the page under migration,
1020 * because then pte is replaced with migration swap entry and direct I/O code
1021 * will wait in the page fault for migration to complete.
1023 static int unmap_and_move_huge_page(new_page_t get_new_page,
1024 free_page_t put_new_page, unsigned long private,
1025 struct page *hpage, int force,
1026 enum migrate_mode mode)
1030 int page_was_mapped = 0;
1031 struct page *new_hpage;
1032 struct anon_vma *anon_vma = NULL;
1035 * Movability of hugepages depends on architectures and hugepage size.
1036 * This check is necessary because some callers of hugepage migration
1037 * like soft offline and memory hotremove don't walk through page
1038 * tables or check whether the hugepage is pmd-based or not before
1039 * kicking migration.
1041 if (!hugepage_migration_supported(page_hstate(hpage))) {
1042 putback_active_hugepage(hpage);
1046 new_hpage = get_new_page(hpage, private, &result);
1050 if (!trylock_page(hpage)) {
1051 if (!force || mode != MIGRATE_SYNC)
1056 if (PageAnon(hpage))
1057 anon_vma = page_get_anon_vma(hpage);
1059 if (page_mapped(hpage)) {
1061 TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
1062 page_was_mapped = 1;
1065 if (!page_mapped(hpage))
1066 rc = move_to_new_page(new_hpage, hpage, page_was_mapped, mode);
1068 if (rc != MIGRATEPAGE_SUCCESS && page_was_mapped)
1069 remove_migration_ptes(hpage, hpage);
1072 put_anon_vma(anon_vma);
1074 if (rc == MIGRATEPAGE_SUCCESS) {
1075 hugetlb_cgroup_migrate(hpage, new_hpage);
1076 put_new_page = NULL;
1082 putback_active_hugepage(hpage);
1085 * If migration was not successful and there's a freeing callback, use
1086 * it. Otherwise, put_page() will drop the reference grabbed during
1090 put_new_page(new_hpage, private);
1092 putback_active_hugepage(new_hpage);
1098 *result = page_to_nid(new_hpage);
1104 * migrate_pages - migrate the pages specified in a list, to the free pages
1105 * supplied as the target for the page migration
1107 * @from: The list of pages to be migrated.
1108 * @get_new_page: The function used to allocate free pages to be used
1109 * as the target of the page migration.
1110 * @put_new_page: The function used to free target pages if migration
1111 * fails, or NULL if no special handling is necessary.
1112 * @private: Private data to be passed on to get_new_page()
1113 * @mode: The migration mode that specifies the constraints for
1114 * page migration, if any.
1115 * @reason: The reason for page migration.
1117 * The function returns after 10 attempts or if no pages are movable any more
1118 * because the list has become empty or no retryable pages exist any more.
1119 * The caller should call putback_movable_pages() to return pages to the LRU
1120 * or free list only if ret != 0.
1122 * Returns the number of pages that were not migrated, or an error code.
1124 int migrate_pages(struct list_head *from, new_page_t get_new_page,
1125 free_page_t put_new_page, unsigned long private,
1126 enum migrate_mode mode, int reason)
1130 int nr_succeeded = 0;
1134 int swapwrite = current->flags & PF_SWAPWRITE;
1138 current->flags |= PF_SWAPWRITE;
1140 for(pass = 0; pass < 10 && retry; pass++) {
1143 list_for_each_entry_safe(page, page2, from, lru) {
1147 rc = unmap_and_move_huge_page(get_new_page,
1148 put_new_page, private, page,
1151 rc = unmap_and_move(get_new_page, put_new_page,
1152 private, page, pass > 2, mode,
1161 case MIGRATEPAGE_SUCCESS:
1166 * Permanent failure (-EBUSY, -ENOSYS, etc.):
1167 * unlike -EAGAIN case, the failed page is
1168 * removed from migration page list and not
1169 * retried in the next outer loop.
1180 count_vm_events(PGMIGRATE_SUCCESS, nr_succeeded);
1182 count_vm_events(PGMIGRATE_FAIL, nr_failed);
1183 trace_mm_migrate_pages(nr_succeeded, nr_failed, mode, reason);
1186 current->flags &= ~PF_SWAPWRITE;
1193 * Move a list of individual pages
1195 struct page_to_node {
1202 static struct page *new_page_node(struct page *p, unsigned long private,
1205 struct page_to_node *pm = (struct page_to_node *)private;
1207 while (pm->node != MAX_NUMNODES && pm->page != p)
1210 if (pm->node == MAX_NUMNODES)
1213 *result = &pm->status;
1216 return alloc_huge_page_node(page_hstate(compound_head(p)),
1219 return __alloc_pages_node(pm->node,
1220 GFP_HIGHUSER_MOVABLE | __GFP_THISNODE, 0);
1224 * Move a set of pages as indicated in the pm array. The addr
1225 * field must be set to the virtual address of the page to be moved
1226 * and the node number must contain a valid target node.
1227 * The pm array ends with node = MAX_NUMNODES.
1229 static int do_move_page_to_node_array(struct mm_struct *mm,
1230 struct page_to_node *pm,
1234 struct page_to_node *pp;
1235 LIST_HEAD(pagelist);
1237 down_read(&mm->mmap_sem);
1240 * Build a list of pages to migrate
1242 for (pp = pm; pp->node != MAX_NUMNODES; pp++) {
1243 struct vm_area_struct *vma;
1247 vma = find_vma(mm, pp->addr);
1248 if (!vma || pp->addr < vma->vm_start || !vma_migratable(vma))
1251 /* FOLL_DUMP to ignore special (like zero) pages */
1252 page = follow_page(vma, pp->addr,
1253 FOLL_GET | FOLL_SPLIT | FOLL_DUMP);
1255 err = PTR_ERR(page);
1264 err = page_to_nid(page);
1266 if (err == pp->node)
1268 * Node already in the right place
1273 if (page_mapcount(page) > 1 &&
1277 if (PageHuge(page)) {
1279 isolate_huge_page(page, &pagelist);
1283 err = isolate_lru_page(page);
1285 list_add_tail(&page->lru, &pagelist);
1286 inc_zone_page_state(page, NR_ISOLATED_ANON +
1287 page_is_file_cache(page));
1291 * Either remove the duplicate refcount from
1292 * isolate_lru_page() or drop the page ref if it was
1301 if (!list_empty(&pagelist)) {
1302 err = migrate_pages(&pagelist, new_page_node, NULL,
1303 (unsigned long)pm, MIGRATE_SYNC, MR_SYSCALL);
1305 putback_movable_pages(&pagelist);
1308 up_read(&mm->mmap_sem);
1313 * Migrate an array of page address onto an array of nodes and fill
1314 * the corresponding array of status.
1316 static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes,
1317 unsigned long nr_pages,
1318 const void __user * __user *pages,
1319 const int __user *nodes,
1320 int __user *status, int flags)
1322 struct page_to_node *pm;
1323 unsigned long chunk_nr_pages;
1324 unsigned long chunk_start;
1328 pm = (struct page_to_node *)__get_free_page(GFP_KERNEL);
1335 * Store a chunk of page_to_node array in a page,
1336 * but keep the last one as a marker
1338 chunk_nr_pages = (PAGE_SIZE / sizeof(struct page_to_node)) - 1;
1340 for (chunk_start = 0;
1341 chunk_start < nr_pages;
1342 chunk_start += chunk_nr_pages) {
1345 if (chunk_start + chunk_nr_pages > nr_pages)
1346 chunk_nr_pages = nr_pages - chunk_start;
1348 /* fill the chunk pm with addrs and nodes from user-space */
1349 for (j = 0; j < chunk_nr_pages; j++) {
1350 const void __user *p;
1354 if (get_user(p, pages + j + chunk_start))
1356 pm[j].addr = (unsigned long) p;
1358 if (get_user(node, nodes + j + chunk_start))
1362 if (node < 0 || node >= MAX_NUMNODES)
1365 if (!node_state(node, N_MEMORY))
1369 if (!node_isset(node, task_nodes))
1375 /* End marker for this chunk */
1376 pm[chunk_nr_pages].node = MAX_NUMNODES;
1378 /* Migrate this chunk */
1379 err = do_move_page_to_node_array(mm, pm,
1380 flags & MPOL_MF_MOVE_ALL);
1384 /* Return status information */
1385 for (j = 0; j < chunk_nr_pages; j++)
1386 if (put_user(pm[j].status, status + j + chunk_start)) {
1394 free_page((unsigned long)pm);
1400 * Determine the nodes of an array of pages and store it in an array of status.
1402 static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages,
1403 const void __user **pages, int *status)
1407 down_read(&mm->mmap_sem);
1409 for (i = 0; i < nr_pages; i++) {
1410 unsigned long addr = (unsigned long)(*pages);
1411 struct vm_area_struct *vma;
1415 vma = find_vma(mm, addr);
1416 if (!vma || addr < vma->vm_start)
1419 /* FOLL_DUMP to ignore special (like zero) pages */
1420 page = follow_page(vma, addr, FOLL_DUMP);
1422 err = PTR_ERR(page);
1426 err = page ? page_to_nid(page) : -ENOENT;
1434 up_read(&mm->mmap_sem);
1438 * Determine the nodes of a user array of pages and store it in
1439 * a user array of status.
1441 static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages,
1442 const void __user * __user *pages,
1445 #define DO_PAGES_STAT_CHUNK_NR 16
1446 const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR];
1447 int chunk_status[DO_PAGES_STAT_CHUNK_NR];
1450 unsigned long chunk_nr;
1452 chunk_nr = nr_pages;
1453 if (chunk_nr > DO_PAGES_STAT_CHUNK_NR)
1454 chunk_nr = DO_PAGES_STAT_CHUNK_NR;
1456 if (copy_from_user(chunk_pages, pages, chunk_nr * sizeof(*chunk_pages)))
1459 do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status);
1461 if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status)))
1466 nr_pages -= chunk_nr;
1468 return nr_pages ? -EFAULT : 0;
1472 * Move a list of pages in the address space of the currently executing
1475 SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages,
1476 const void __user * __user *, pages,
1477 const int __user *, nodes,
1478 int __user *, status, int, flags)
1480 const struct cred *cred = current_cred(), *tcred;
1481 struct task_struct *task;
1482 struct mm_struct *mm;
1484 nodemask_t task_nodes;
1487 if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
1490 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1493 /* Find the mm_struct */
1495 task = pid ? find_task_by_vpid(pid) : current;
1500 get_task_struct(task);
1503 * Check if this process has the right to modify the specified
1504 * process. The right exists if the process has administrative
1505 * capabilities, superuser privileges or the same
1506 * userid as the target process.
1508 tcred = __task_cred(task);
1509 if (!uid_eq(cred->euid, tcred->suid) && !uid_eq(cred->euid, tcred->uid) &&
1510 !uid_eq(cred->uid, tcred->suid) && !uid_eq(cred->uid, tcred->uid) &&
1511 !capable(CAP_SYS_NICE)) {
1518 err = security_task_movememory(task);
1522 task_nodes = cpuset_mems_allowed(task);
1523 mm = get_task_mm(task);
1524 put_task_struct(task);
1530 err = do_pages_move(mm, task_nodes, nr_pages, pages,
1531 nodes, status, flags);
1533 err = do_pages_stat(mm, nr_pages, pages, status);
1539 put_task_struct(task);
1543 #ifdef CONFIG_NUMA_BALANCING
1545 * Returns true if this is a safe migration target node for misplaced NUMA
1546 * pages. Currently it only checks the watermarks which crude
1548 static bool migrate_balanced_pgdat(struct pglist_data *pgdat,
1549 unsigned long nr_migrate_pages)
1552 for (z = pgdat->nr_zones - 1; z >= 0; z--) {
1553 struct zone *zone = pgdat->node_zones + z;
1555 if (!populated_zone(zone))
1558 if (!zone_reclaimable(zone))
1561 /* Avoid waking kswapd by allocating pages_to_migrate pages. */
1562 if (!zone_watermark_ok(zone, 0,
1563 high_wmark_pages(zone) +
1572 static struct page *alloc_misplaced_dst_page(struct page *page,
1576 int nid = (int) data;
1577 struct page *newpage;
1579 newpage = __alloc_pages_node(nid,
1580 (GFP_HIGHUSER_MOVABLE |
1581 __GFP_THISNODE | __GFP_NOMEMALLOC |
1582 __GFP_NORETRY | __GFP_NOWARN) &
1589 * page migration rate limiting control.
1590 * Do not migrate more than @pages_to_migrate in a @migrate_interval_millisecs
1591 * window of time. Default here says do not migrate more than 1280M per second.
1593 static unsigned int migrate_interval_millisecs __read_mostly = 100;
1594 static unsigned int ratelimit_pages __read_mostly = 128 << (20 - PAGE_SHIFT);
1596 /* Returns true if the node is migrate rate-limited after the update */
1597 static bool numamigrate_update_ratelimit(pg_data_t *pgdat,
1598 unsigned long nr_pages)
1601 * Rate-limit the amount of data that is being migrated to a node.
1602 * Optimal placement is no good if the memory bus is saturated and
1603 * all the time is being spent migrating!
1605 if (time_after(jiffies, pgdat->numabalancing_migrate_next_window)) {
1606 spin_lock(&pgdat->numabalancing_migrate_lock);
1607 pgdat->numabalancing_migrate_nr_pages = 0;
1608 pgdat->numabalancing_migrate_next_window = jiffies +
1609 msecs_to_jiffies(migrate_interval_millisecs);
1610 spin_unlock(&pgdat->numabalancing_migrate_lock);
1612 if (pgdat->numabalancing_migrate_nr_pages > ratelimit_pages) {
1613 trace_mm_numa_migrate_ratelimit(current, pgdat->node_id,
1619 * This is an unlocked non-atomic update so errors are possible.
1620 * The consequences are failing to migrate when we potentiall should
1621 * have which is not severe enough to warrant locking. If it is ever
1622 * a problem, it can be converted to a per-cpu counter.
1624 pgdat->numabalancing_migrate_nr_pages += nr_pages;
1628 static int numamigrate_isolate_page(pg_data_t *pgdat, struct page *page)
1632 VM_BUG_ON_PAGE(compound_order(page) && !PageTransHuge(page), page);
1634 /* Avoid migrating to a node that is nearly full */
1635 if (!migrate_balanced_pgdat(pgdat, 1UL << compound_order(page)))
1638 if (isolate_lru_page(page))
1642 * migrate_misplaced_transhuge_page() skips page migration's usual
1643 * check on page_count(), so we must do it here, now that the page
1644 * has been isolated: a GUP pin, or any other pin, prevents migration.
1645 * The expected page count is 3: 1 for page's mapcount and 1 for the
1646 * caller's pin and 1 for the reference taken by isolate_lru_page().
1648 if (PageTransHuge(page) && page_count(page) != 3) {
1649 putback_lru_page(page);
1653 page_lru = page_is_file_cache(page);
1654 mod_zone_page_state(page_zone(page), NR_ISOLATED_ANON + page_lru,
1655 hpage_nr_pages(page));
1658 * Isolating the page has taken another reference, so the
1659 * caller's reference can be safely dropped without the page
1660 * disappearing underneath us during migration.
1666 bool pmd_trans_migrating(pmd_t pmd)
1668 struct page *page = pmd_page(pmd);
1669 return PageLocked(page);
1673 * Attempt to migrate a misplaced page to the specified destination
1674 * node. Caller is expected to have an elevated reference count on
1675 * the page that will be dropped by this function before returning.
1677 int migrate_misplaced_page(struct page *page, struct vm_area_struct *vma,
1680 pg_data_t *pgdat = NODE_DATA(node);
1683 LIST_HEAD(migratepages);
1686 * Don't migrate file pages that are mapped in multiple processes
1687 * with execute permissions as they are probably shared libraries.
1689 if (page_mapcount(page) != 1 && page_is_file_cache(page) &&
1690 (vma->vm_flags & VM_EXEC))
1694 * Rate-limit the amount of data that is being migrated to a node.
1695 * Optimal placement is no good if the memory bus is saturated and
1696 * all the time is being spent migrating!
1698 if (numamigrate_update_ratelimit(pgdat, 1))
1701 isolated = numamigrate_isolate_page(pgdat, page);
1705 list_add(&page->lru, &migratepages);
1706 nr_remaining = migrate_pages(&migratepages, alloc_misplaced_dst_page,
1707 NULL, node, MIGRATE_ASYNC,
1710 if (!list_empty(&migratepages)) {
1711 list_del(&page->lru);
1712 dec_zone_page_state(page, NR_ISOLATED_ANON +
1713 page_is_file_cache(page));
1714 putback_lru_page(page);
1718 count_vm_numa_event(NUMA_PAGE_MIGRATE);
1719 BUG_ON(!list_empty(&migratepages));
1726 #endif /* CONFIG_NUMA_BALANCING */
1728 #if defined(CONFIG_NUMA_BALANCING) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
1730 * Migrates a THP to a given target node. page must be locked and is unlocked
1733 int migrate_misplaced_transhuge_page(struct mm_struct *mm,
1734 struct vm_area_struct *vma,
1735 pmd_t *pmd, pmd_t entry,
1736 unsigned long address,
1737 struct page *page, int node)
1740 pg_data_t *pgdat = NODE_DATA(node);
1742 struct page *new_page = NULL;
1743 int page_lru = page_is_file_cache(page);
1744 unsigned long mmun_start = address & HPAGE_PMD_MASK;
1745 unsigned long mmun_end = mmun_start + HPAGE_PMD_SIZE;
1749 * Rate-limit the amount of data that is being migrated to a node.
1750 * Optimal placement is no good if the memory bus is saturated and
1751 * all the time is being spent migrating!
1753 if (numamigrate_update_ratelimit(pgdat, HPAGE_PMD_NR))
1756 new_page = alloc_pages_node(node,
1757 (GFP_TRANSHUGE | __GFP_THISNODE) & ~__GFP_WAIT,
1762 isolated = numamigrate_isolate_page(pgdat, page);
1768 if (mm_tlb_flush_pending(mm))
1769 flush_tlb_range(vma, mmun_start, mmun_end);
1771 /* Prepare a page as a migration target */
1772 __set_page_locked(new_page);
1773 SetPageSwapBacked(new_page);
1775 /* anon mapping, we can simply copy page->mapping to the new page: */
1776 new_page->mapping = page->mapping;
1777 new_page->index = page->index;
1778 migrate_page_copy(new_page, page);
1779 WARN_ON(PageLRU(new_page));
1781 /* Recheck the target PMD */
1782 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1783 ptl = pmd_lock(mm, pmd);
1784 if (unlikely(!pmd_same(*pmd, entry) || page_count(page) != 2)) {
1787 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1789 /* Reverse changes made by migrate_page_copy() */
1790 if (TestClearPageActive(new_page))
1791 SetPageActive(page);
1792 if (TestClearPageUnevictable(new_page))
1793 SetPageUnevictable(page);
1795 unlock_page(new_page);
1796 put_page(new_page); /* Free it */
1798 /* Retake the callers reference and putback on LRU */
1800 putback_lru_page(page);
1801 mod_zone_page_state(page_zone(page),
1802 NR_ISOLATED_ANON + page_lru, -HPAGE_PMD_NR);
1808 entry = mk_pmd(new_page, vma->vm_page_prot);
1809 entry = pmd_mkhuge(entry);
1810 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1813 * Clear the old entry under pagetable lock and establish the new PTE.
1814 * Any parallel GUP will either observe the old page blocking on the
1815 * page lock, block on the page table lock or observe the new page.
1816 * The SetPageUptodate on the new page and page_add_new_anon_rmap
1817 * guarantee the copy is visible before the pagetable update.
1819 flush_cache_range(vma, mmun_start, mmun_end);
1820 page_add_anon_rmap(new_page, vma, mmun_start);
1821 pmdp_huge_clear_flush_notify(vma, mmun_start, pmd);
1822 set_pmd_at(mm, mmun_start, pmd, entry);
1823 flush_tlb_range(vma, mmun_start, mmun_end);
1824 update_mmu_cache_pmd(vma, address, &entry);
1826 if (page_count(page) != 2) {
1827 set_pmd_at(mm, mmun_start, pmd, orig_entry);
1828 flush_tlb_range(vma, mmun_start, mmun_end);
1829 mmu_notifier_invalidate_range(mm, mmun_start, mmun_end);
1830 update_mmu_cache_pmd(vma, address, &entry);
1831 page_remove_rmap(new_page);
1835 mlock_migrate_page(new_page, page);
1836 set_page_memcg(new_page, page_memcg(page));
1837 set_page_memcg(page, NULL);
1838 page_remove_rmap(page);
1841 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1843 /* Take an "isolate" reference and put new page on the LRU. */
1845 putback_lru_page(new_page);
1847 unlock_page(new_page);
1849 put_page(page); /* Drop the rmap reference */
1850 put_page(page); /* Drop the LRU isolation reference */
1852 count_vm_events(PGMIGRATE_SUCCESS, HPAGE_PMD_NR);
1853 count_vm_numa_events(NUMA_PAGE_MIGRATE, HPAGE_PMD_NR);
1855 mod_zone_page_state(page_zone(page),
1856 NR_ISOLATED_ANON + page_lru,
1861 count_vm_events(PGMIGRATE_FAIL, HPAGE_PMD_NR);
1863 ptl = pmd_lock(mm, pmd);
1864 if (pmd_same(*pmd, entry)) {
1865 entry = pmd_modify(entry, vma->vm_page_prot);
1866 set_pmd_at(mm, mmun_start, pmd, entry);
1867 update_mmu_cache_pmd(vma, address, &entry);
1876 #endif /* CONFIG_NUMA_BALANCING */
1878 #endif /* CONFIG_NUMA */