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>
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 /* No need to invalidate - it was non-present before */
174 update_mmu_cache(vma, addr, ptep);
176 pte_unmap_unlock(ptep, ptl);
182 * Get rid of all migration entries and replace them by
183 * references to the indicated page.
185 static void remove_migration_ptes(struct page *old, struct page *new)
187 struct rmap_walk_control rwc = {
188 .rmap_one = remove_migration_pte,
192 rmap_walk(new, &rwc);
196 * Something used the pte of a page under migration. We need to
197 * get to the page and wait until migration is finished.
198 * When we return from this function the fault will be retried.
200 void __migration_entry_wait(struct mm_struct *mm, pte_t *ptep,
209 if (!is_swap_pte(pte))
212 entry = pte_to_swp_entry(pte);
213 if (!is_migration_entry(entry))
216 page = migration_entry_to_page(entry);
219 * Once radix-tree replacement of page migration started, page_count
220 * *must* be zero. And, we don't want to call wait_on_page_locked()
221 * against a page without get_page().
222 * So, we use get_page_unless_zero(), here. Even failed, page fault
225 if (!get_page_unless_zero(page))
227 pte_unmap_unlock(ptep, ptl);
228 wait_on_page_locked(page);
232 pte_unmap_unlock(ptep, ptl);
235 void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
236 unsigned long address)
238 spinlock_t *ptl = pte_lockptr(mm, pmd);
239 pte_t *ptep = pte_offset_map(pmd, address);
240 __migration_entry_wait(mm, ptep, ptl);
243 void migration_entry_wait_huge(struct vm_area_struct *vma,
244 struct mm_struct *mm, pte_t *pte)
246 spinlock_t *ptl = huge_pte_lockptr(hstate_vma(vma), mm, pte);
247 __migration_entry_wait(mm, pte, ptl);
251 /* Returns true if all buffers are successfully locked */
252 static bool buffer_migrate_lock_buffers(struct buffer_head *head,
253 enum migrate_mode mode)
255 struct buffer_head *bh = head;
257 /* Simple case, sync compaction */
258 if (mode != MIGRATE_ASYNC) {
262 bh = bh->b_this_page;
264 } while (bh != head);
269 /* async case, we cannot block on lock_buffer so use trylock_buffer */
272 if (!trylock_buffer(bh)) {
274 * We failed to lock the buffer and cannot stall in
275 * async migration. Release the taken locks
277 struct buffer_head *failed_bh = bh;
280 while (bh != failed_bh) {
283 bh = bh->b_this_page;
288 bh = bh->b_this_page;
289 } while (bh != head);
293 static inline bool buffer_migrate_lock_buffers(struct buffer_head *head,
294 enum migrate_mode mode)
298 #endif /* CONFIG_BLOCK */
301 * Replace the page in the mapping.
303 * The number of remaining references must be:
304 * 1 for anonymous pages without a mapping
305 * 2 for pages with a mapping
306 * 3 for pages with a mapping and PagePrivate/PagePrivate2 set.
308 int migrate_page_move_mapping(struct address_space *mapping,
309 struct page *newpage, struct page *page,
310 struct buffer_head *head, enum migrate_mode mode,
313 int expected_count = 1 + extra_count;
317 /* Anonymous page without mapping */
318 if (page_count(page) != expected_count)
320 return MIGRATEPAGE_SUCCESS;
323 spin_lock_irq(&mapping->tree_lock);
325 pslot = radix_tree_lookup_slot(&mapping->page_tree,
328 expected_count += 1 + page_has_private(page);
329 if (page_count(page) != expected_count ||
330 radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) {
331 spin_unlock_irq(&mapping->tree_lock);
335 if (!page_freeze_refs(page, expected_count)) {
336 spin_unlock_irq(&mapping->tree_lock);
341 * In the async migration case of moving a page with buffers, lock the
342 * buffers using trylock before the mapping is moved. If the mapping
343 * was moved, we later failed to lock the buffers and could not move
344 * the mapping back due to an elevated page count, we would have to
345 * block waiting on other references to be dropped.
347 if (mode == MIGRATE_ASYNC && head &&
348 !buffer_migrate_lock_buffers(head, mode)) {
349 page_unfreeze_refs(page, expected_count);
350 spin_unlock_irq(&mapping->tree_lock);
355 * Now we know that no one else is looking at the page.
357 get_page(newpage); /* add cache reference */
358 if (PageSwapCache(page)) {
359 SetPageSwapCache(newpage);
360 set_page_private(newpage, page_private(page));
363 radix_tree_replace_slot(pslot, newpage);
366 * Drop cache reference from old page by unfreezing
367 * to one less reference.
368 * We know this isn't the last reference.
370 page_unfreeze_refs(page, expected_count - 1);
373 * If moved to a different zone then also account
374 * the page for that zone. Other VM counters will be
375 * taken care of when we establish references to the
376 * new page and drop references to the old page.
378 * Note that anonymous pages are accounted for
379 * via NR_FILE_PAGES and NR_ANON_PAGES if they
380 * are mapped to swap space.
382 __dec_zone_page_state(page, NR_FILE_PAGES);
383 __inc_zone_page_state(newpage, NR_FILE_PAGES);
384 if (!PageSwapCache(page) && PageSwapBacked(page)) {
385 __dec_zone_page_state(page, NR_SHMEM);
386 __inc_zone_page_state(newpage, NR_SHMEM);
388 spin_unlock_irq(&mapping->tree_lock);
390 return MIGRATEPAGE_SUCCESS;
394 * The expected number of remaining references is the same as that
395 * of migrate_page_move_mapping().
397 int migrate_huge_page_move_mapping(struct address_space *mapping,
398 struct page *newpage, struct page *page)
404 if (page_count(page) != 1)
406 return MIGRATEPAGE_SUCCESS;
409 spin_lock_irq(&mapping->tree_lock);
411 pslot = radix_tree_lookup_slot(&mapping->page_tree,
414 expected_count = 2 + page_has_private(page);
415 if (page_count(page) != expected_count ||
416 radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) {
417 spin_unlock_irq(&mapping->tree_lock);
421 if (!page_freeze_refs(page, expected_count)) {
422 spin_unlock_irq(&mapping->tree_lock);
428 radix_tree_replace_slot(pslot, newpage);
430 page_unfreeze_refs(page, expected_count - 1);
432 spin_unlock_irq(&mapping->tree_lock);
433 return MIGRATEPAGE_SUCCESS;
437 * Gigantic pages are so large that we do not guarantee that page++ pointer
438 * arithmetic will work across the entire page. We need something more
441 static void __copy_gigantic_page(struct page *dst, struct page *src,
445 struct page *dst_base = dst;
446 struct page *src_base = src;
448 for (i = 0; i < nr_pages; ) {
450 copy_highpage(dst, src);
453 dst = mem_map_next(dst, dst_base, i);
454 src = mem_map_next(src, src_base, i);
458 static void copy_huge_page(struct page *dst, struct page *src)
465 struct hstate *h = page_hstate(src);
466 nr_pages = pages_per_huge_page(h);
468 if (unlikely(nr_pages > MAX_ORDER_NR_PAGES)) {
469 __copy_gigantic_page(dst, src, nr_pages);
474 BUG_ON(!PageTransHuge(src));
475 nr_pages = hpage_nr_pages(src);
478 for (i = 0; i < nr_pages; i++) {
480 copy_highpage(dst + i, src + i);
485 * Copy the page to its new location
487 void migrate_page_copy(struct page *newpage, struct page *page)
491 if (PageHuge(page) || PageTransHuge(page))
492 copy_huge_page(newpage, page);
494 copy_highpage(newpage, page);
497 SetPageError(newpage);
498 if (PageReferenced(page))
499 SetPageReferenced(newpage);
500 if (PageUptodate(page))
501 SetPageUptodate(newpage);
502 if (TestClearPageActive(page)) {
503 VM_BUG_ON_PAGE(PageUnevictable(page), page);
504 SetPageActive(newpage);
505 } else if (TestClearPageUnevictable(page))
506 SetPageUnevictable(newpage);
507 if (PageChecked(page))
508 SetPageChecked(newpage);
509 if (PageMappedToDisk(page))
510 SetPageMappedToDisk(newpage);
512 if (PageDirty(page)) {
513 clear_page_dirty_for_io(page);
515 * Want to mark the page and the radix tree as dirty, and
516 * redo the accounting that clear_page_dirty_for_io undid,
517 * but we can't use set_page_dirty because that function
518 * is actually a signal that all of the page has become dirty.
519 * Whereas only part of our page may be dirty.
521 if (PageSwapBacked(page))
522 SetPageDirty(newpage);
524 __set_page_dirty_nobuffers(newpage);
528 * Copy NUMA information to the new page, to prevent over-eager
529 * future migrations of this same page.
531 cpupid = page_cpupid_xchg_last(page, -1);
532 page_cpupid_xchg_last(newpage, cpupid);
534 mlock_migrate_page(newpage, page);
535 ksm_migrate_page(newpage, page);
537 * Please do not reorder this without considering how mm/ksm.c's
538 * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache().
540 if (PageSwapCache(page))
541 ClearPageSwapCache(page);
542 ClearPagePrivate(page);
543 set_page_private(page, 0);
546 * If any waiters have accumulated on the new page then
549 if (PageWriteback(newpage))
550 end_page_writeback(newpage);
553 /************************************************************
554 * Migration functions
555 ***********************************************************/
558 * Common logic to directly migrate a single page suitable for
559 * pages that do not use PagePrivate/PagePrivate2.
561 * Pages are locked upon entry and exit.
563 int migrate_page(struct address_space *mapping,
564 struct page *newpage, struct page *page,
565 enum migrate_mode mode)
569 BUG_ON(PageWriteback(page)); /* Writeback must be complete */
571 rc = migrate_page_move_mapping(mapping, newpage, page, NULL, mode, 0);
573 if (rc != MIGRATEPAGE_SUCCESS)
576 migrate_page_copy(newpage, page);
577 return MIGRATEPAGE_SUCCESS;
579 EXPORT_SYMBOL(migrate_page);
583 * Migration function for pages with buffers. This function can only be used
584 * if the underlying filesystem guarantees that no other references to "page"
587 int buffer_migrate_page(struct address_space *mapping,
588 struct page *newpage, struct page *page, enum migrate_mode mode)
590 struct buffer_head *bh, *head;
593 if (!page_has_buffers(page))
594 return migrate_page(mapping, newpage, page, mode);
596 head = page_buffers(page);
598 rc = migrate_page_move_mapping(mapping, newpage, page, head, mode, 0);
600 if (rc != MIGRATEPAGE_SUCCESS)
604 * In the async case, migrate_page_move_mapping locked the buffers
605 * with an IRQ-safe spinlock held. In the sync case, the buffers
606 * need to be locked now
608 if (mode != MIGRATE_ASYNC)
609 BUG_ON(!buffer_migrate_lock_buffers(head, mode));
611 ClearPagePrivate(page);
612 set_page_private(newpage, page_private(page));
613 set_page_private(page, 0);
619 set_bh_page(bh, newpage, bh_offset(bh));
620 bh = bh->b_this_page;
622 } while (bh != head);
624 SetPagePrivate(newpage);
626 migrate_page_copy(newpage, page);
632 bh = bh->b_this_page;
634 } while (bh != head);
636 return MIGRATEPAGE_SUCCESS;
638 EXPORT_SYMBOL(buffer_migrate_page);
642 * Writeback a page to clean the dirty state
644 static int writeout(struct address_space *mapping, struct page *page)
646 struct writeback_control wbc = {
647 .sync_mode = WB_SYNC_NONE,
650 .range_end = LLONG_MAX,
655 if (!mapping->a_ops->writepage)
656 /* No write method for the address space */
659 if (!clear_page_dirty_for_io(page))
660 /* Someone else already triggered a write */
664 * A dirty page may imply that the underlying filesystem has
665 * the page on some queue. So the page must be clean for
666 * migration. Writeout may mean we loose the lock and the
667 * page state is no longer what we checked for earlier.
668 * At this point we know that the migration attempt cannot
671 remove_migration_ptes(page, page);
673 rc = mapping->a_ops->writepage(page, &wbc);
675 if (rc != AOP_WRITEPAGE_ACTIVATE)
676 /* unlocked. Relock */
679 return (rc < 0) ? -EIO : -EAGAIN;
683 * Default handling if a filesystem does not provide a migration function.
685 static int fallback_migrate_page(struct address_space *mapping,
686 struct page *newpage, struct page *page, enum migrate_mode mode)
688 if (PageDirty(page)) {
689 /* Only writeback pages in full synchronous migration */
690 if (mode != MIGRATE_SYNC)
692 return writeout(mapping, page);
696 * Buffers may be managed in a filesystem specific way.
697 * We must have no buffers or drop them.
699 if (page_has_private(page) &&
700 !try_to_release_page(page, GFP_KERNEL))
703 return migrate_page(mapping, newpage, page, mode);
707 * Move a page to a newly allocated page
708 * The page is locked and all ptes have been successfully removed.
710 * The new page will have replaced the old page if this function
715 * MIGRATEPAGE_SUCCESS - success
717 static int move_to_new_page(struct page *newpage, struct page *page,
718 int page_was_mapped, enum migrate_mode mode)
720 struct address_space *mapping;
724 * Block others from accessing the page when we get around to
725 * establishing additional references. We are the only one
726 * holding a reference to the new page at this point.
728 if (!trylock_page(newpage))
731 /* Prepare mapping for the new page.*/
732 newpage->index = page->index;
733 newpage->mapping = page->mapping;
734 if (PageSwapBacked(page))
735 SetPageSwapBacked(newpage);
737 mapping = page_mapping(page);
739 rc = migrate_page(mapping, newpage, page, mode);
740 else if (mapping->a_ops->migratepage)
742 * Most pages have a mapping and most filesystems provide a
743 * migratepage callback. Anonymous pages are part of swap
744 * space which also has its own migratepage callback. This
745 * is the most common path for page migration.
747 rc = mapping->a_ops->migratepage(mapping,
748 newpage, page, mode);
750 rc = fallback_migrate_page(mapping, newpage, page, mode);
752 if (rc != MIGRATEPAGE_SUCCESS) {
753 newpage->mapping = NULL;
755 mem_cgroup_migrate(page, newpage, false);
757 remove_migration_ptes(page, newpage);
758 page->mapping = NULL;
761 unlock_page(newpage);
766 static int __unmap_and_move(struct page *page, struct page *newpage,
767 int force, enum migrate_mode mode)
770 int page_was_mapped = 0;
771 struct anon_vma *anon_vma = NULL;
773 if (!trylock_page(page)) {
774 if (!force || mode == MIGRATE_ASYNC)
778 * It's not safe for direct compaction to call lock_page.
779 * For example, during page readahead pages are added locked
780 * to the LRU. Later, when the IO completes the pages are
781 * marked uptodate and unlocked. However, the queueing
782 * could be merging multiple pages for one bio (e.g.
783 * mpage_readpages). If an allocation happens for the
784 * second or third page, the process can end up locking
785 * the same page twice and deadlocking. Rather than
786 * trying to be clever about what pages can be locked,
787 * avoid the use of lock_page for direct compaction
790 if (current->flags & PF_MEMALLOC)
796 if (PageWriteback(page)) {
798 * Only in the case of a full synchronous migration is it
799 * necessary to wait for PageWriteback. In the async case,
800 * the retry loop is too short and in the sync-light case,
801 * the overhead of stalling is too much
803 if (mode != MIGRATE_SYNC) {
809 wait_on_page_writeback(page);
812 * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
813 * we cannot notice that anon_vma is freed while we migrates a page.
814 * This get_anon_vma() delays freeing anon_vma pointer until the end
815 * of migration. File cache pages are no problem because of page_lock()
816 * File Caches may use write_page() or lock_page() in migration, then,
817 * just care Anon page here.
819 if (PageAnon(page) && !PageKsm(page)) {
821 * Only page_lock_anon_vma_read() understands the subtleties of
822 * getting a hold on an anon_vma from outside one of its mms.
824 anon_vma = page_get_anon_vma(page);
829 } else if (PageSwapCache(page)) {
831 * We cannot be sure that the anon_vma of an unmapped
832 * swapcache page is safe to use because we don't
833 * know in advance if the VMA that this page belonged
834 * to still exists. If the VMA and others sharing the
835 * data have been freed, then the anon_vma could
836 * already be invalid.
838 * To avoid this possibility, swapcache pages get
839 * migrated but are not remapped when migration
847 if (unlikely(isolated_balloon_page(page))) {
849 * A ballooned page does not need any special attention from
850 * physical to virtual reverse mapping procedures.
851 * Skip any attempt to unmap PTEs or to remap swap cache,
852 * in order to avoid burning cycles at rmap level, and perform
853 * the page migration right away (proteced by page lock).
855 rc = balloon_page_migrate(newpage, page, mode);
860 * Corner case handling:
861 * 1. When a new swap-cache page is read into, it is added to the LRU
862 * and treated as swapcache but it has no rmap yet.
863 * Calling try_to_unmap() against a page->mapping==NULL page will
864 * trigger a BUG. So handle it here.
865 * 2. An orphaned page (see truncate_complete_page) might have
866 * fs-private metadata. The page can be picked up due to memory
867 * offlining. Everywhere else except page reclaim, the page is
868 * invisible to the vm, so the page can not be migrated. So try to
869 * free the metadata, so the page can be freed.
871 if (!page->mapping) {
872 VM_BUG_ON_PAGE(PageAnon(page), page);
873 if (page_has_private(page)) {
874 try_to_free_buffers(page);
880 /* Establish migration ptes or remove ptes */
881 if (page_mapped(page)) {
883 TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
888 if (!page_mapped(page))
889 rc = move_to_new_page(newpage, page, page_was_mapped, mode);
891 if (rc && page_was_mapped)
892 remove_migration_ptes(page, page);
894 /* Drop an anon_vma reference if we took one */
896 put_anon_vma(anon_vma);
905 * gcc 4.7 and 4.8 on arm get an ICEs when inlining unmap_and_move(). Work
908 #if (GCC_VERSION >= 40700 && GCC_VERSION < 40900) && defined(CONFIG_ARM)
909 #define ICE_noinline noinline
915 * Obtain the lock on page, remove all ptes and migrate the page
916 * to the newly allocated page in newpage.
918 static ICE_noinline int unmap_and_move(new_page_t get_new_page,
919 free_page_t put_new_page,
920 unsigned long private, struct page *page,
921 int force, enum migrate_mode mode)
925 struct page *newpage = get_new_page(page, private, &result);
930 if (page_count(page) == 1) {
931 /* page was freed from under us. So we are done. */
935 if (unlikely(PageTransHuge(page)))
936 if (unlikely(split_huge_page(page)))
939 rc = __unmap_and_move(page, newpage, force, mode);
944 * A page that has been migrated has all references
945 * removed and will be freed. A page that has not been
946 * migrated will have kepts its references and be
949 list_del(&page->lru);
950 dec_zone_page_state(page, NR_ISOLATED_ANON +
951 page_is_file_cache(page));
952 putback_lru_page(page);
956 * If migration was not successful and there's a freeing callback, use
957 * it. Otherwise, putback_lru_page() will drop the reference grabbed
960 if (rc != MIGRATEPAGE_SUCCESS && put_new_page) {
961 ClearPageSwapBacked(newpage);
962 put_new_page(newpage, private);
963 } else if (unlikely(__is_movable_balloon_page(newpage))) {
964 /* drop our reference, page already in the balloon */
967 putback_lru_page(newpage);
973 *result = page_to_nid(newpage);
979 * Counterpart of unmap_and_move_page() for hugepage migration.
981 * This function doesn't wait the completion of hugepage I/O
982 * because there is no race between I/O and migration for hugepage.
983 * Note that currently hugepage I/O occurs only in direct I/O
984 * where no lock is held and PG_writeback is irrelevant,
985 * and writeback status of all subpages are counted in the reference
986 * count of the head page (i.e. if all subpages of a 2MB hugepage are
987 * under direct I/O, the reference of the head page is 512 and a bit more.)
988 * This means that when we try to migrate hugepage whose subpages are
989 * doing direct I/O, some references remain after try_to_unmap() and
990 * hugepage migration fails without data corruption.
992 * There is also no race when direct I/O is issued on the page under migration,
993 * because then pte is replaced with migration swap entry and direct I/O code
994 * will wait in the page fault for migration to complete.
996 static int unmap_and_move_huge_page(new_page_t get_new_page,
997 free_page_t put_new_page, unsigned long private,
998 struct page *hpage, int force,
999 enum migrate_mode mode)
1003 int page_was_mapped = 0;
1004 struct page *new_hpage;
1005 struct anon_vma *anon_vma = NULL;
1008 * Movability of hugepages depends on architectures and hugepage size.
1009 * This check is necessary because some callers of hugepage migration
1010 * like soft offline and memory hotremove don't walk through page
1011 * tables or check whether the hugepage is pmd-based or not before
1012 * kicking migration.
1014 if (!hugepage_migration_supported(page_hstate(hpage))) {
1015 putback_active_hugepage(hpage);
1019 new_hpage = get_new_page(hpage, private, &result);
1025 if (!trylock_page(hpage)) {
1026 if (!force || mode != MIGRATE_SYNC)
1031 if (PageAnon(hpage))
1032 anon_vma = page_get_anon_vma(hpage);
1034 if (page_mapped(hpage)) {
1036 TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
1037 page_was_mapped = 1;
1040 if (!page_mapped(hpage))
1041 rc = move_to_new_page(new_hpage, hpage, page_was_mapped, mode);
1043 if (rc != MIGRATEPAGE_SUCCESS && page_was_mapped)
1044 remove_migration_ptes(hpage, hpage);
1047 put_anon_vma(anon_vma);
1049 if (rc == MIGRATEPAGE_SUCCESS)
1050 hugetlb_cgroup_migrate(hpage, new_hpage);
1055 putback_active_hugepage(hpage);
1058 * If migration was not successful and there's a freeing callback, use
1059 * it. Otherwise, put_page() will drop the reference grabbed during
1062 if (rc != MIGRATEPAGE_SUCCESS && put_new_page)
1063 put_new_page(new_hpage, private);
1065 put_page(new_hpage);
1071 *result = page_to_nid(new_hpage);
1077 * migrate_pages - migrate the pages specified in a list, to the free pages
1078 * supplied as the target for the page migration
1080 * @from: The list of pages to be migrated.
1081 * @get_new_page: The function used to allocate free pages to be used
1082 * as the target of the page migration.
1083 * @put_new_page: The function used to free target pages if migration
1084 * fails, or NULL if no special handling is necessary.
1085 * @private: Private data to be passed on to get_new_page()
1086 * @mode: The migration mode that specifies the constraints for
1087 * page migration, if any.
1088 * @reason: The reason for page migration.
1090 * The function returns after 10 attempts or if no pages are movable any more
1091 * because the list has become empty or no retryable pages exist any more.
1092 * The caller should call putback_lru_pages() to return pages to the LRU
1093 * or free list only if ret != 0.
1095 * Returns the number of pages that were not migrated, or an error code.
1097 int migrate_pages(struct list_head *from, new_page_t get_new_page,
1098 free_page_t put_new_page, unsigned long private,
1099 enum migrate_mode mode, int reason)
1103 int nr_succeeded = 0;
1107 int swapwrite = current->flags & PF_SWAPWRITE;
1111 current->flags |= PF_SWAPWRITE;
1113 for(pass = 0; pass < 10 && retry; pass++) {
1116 list_for_each_entry_safe(page, page2, from, lru) {
1120 rc = unmap_and_move_huge_page(get_new_page,
1121 put_new_page, private, page,
1124 rc = unmap_and_move(get_new_page, put_new_page,
1125 private, page, pass > 2, mode);
1133 case MIGRATEPAGE_SUCCESS:
1138 * Permanent failure (-EBUSY, -ENOSYS, etc.):
1139 * unlike -EAGAIN case, the failed page is
1140 * removed from migration page list and not
1141 * retried in the next outer loop.
1148 rc = nr_failed + retry;
1151 count_vm_events(PGMIGRATE_SUCCESS, nr_succeeded);
1153 count_vm_events(PGMIGRATE_FAIL, nr_failed);
1154 trace_mm_migrate_pages(nr_succeeded, nr_failed, mode, reason);
1157 current->flags &= ~PF_SWAPWRITE;
1164 * Move a list of individual pages
1166 struct page_to_node {
1173 static struct page *new_page_node(struct page *p, unsigned long private,
1176 struct page_to_node *pm = (struct page_to_node *)private;
1178 while (pm->node != MAX_NUMNODES && pm->page != p)
1181 if (pm->node == MAX_NUMNODES)
1184 *result = &pm->status;
1187 return alloc_huge_page_node(page_hstate(compound_head(p)),
1190 return alloc_pages_exact_node(pm->node,
1191 GFP_HIGHUSER_MOVABLE | __GFP_THISNODE, 0);
1195 * Move a set of pages as indicated in the pm array. The addr
1196 * field must be set to the virtual address of the page to be moved
1197 * and the node number must contain a valid target node.
1198 * The pm array ends with node = MAX_NUMNODES.
1200 static int do_move_page_to_node_array(struct mm_struct *mm,
1201 struct page_to_node *pm,
1205 struct page_to_node *pp;
1206 LIST_HEAD(pagelist);
1208 down_read(&mm->mmap_sem);
1211 * Build a list of pages to migrate
1213 for (pp = pm; pp->node != MAX_NUMNODES; pp++) {
1214 struct vm_area_struct *vma;
1218 vma = find_vma(mm, pp->addr);
1219 if (!vma || pp->addr < vma->vm_start || !vma_migratable(vma))
1222 page = follow_page(vma, pp->addr, FOLL_GET|FOLL_SPLIT);
1224 err = PTR_ERR(page);
1232 /* Use PageReserved to check for zero page */
1233 if (PageReserved(page))
1237 err = page_to_nid(page);
1239 if (err == pp->node)
1241 * Node already in the right place
1246 if (page_mapcount(page) > 1 &&
1250 if (PageHuge(page)) {
1252 isolate_huge_page(page, &pagelist);
1256 err = isolate_lru_page(page);
1258 list_add_tail(&page->lru, &pagelist);
1259 inc_zone_page_state(page, NR_ISOLATED_ANON +
1260 page_is_file_cache(page));
1264 * Either remove the duplicate refcount from
1265 * isolate_lru_page() or drop the page ref if it was
1274 if (!list_empty(&pagelist)) {
1275 err = migrate_pages(&pagelist, new_page_node, NULL,
1276 (unsigned long)pm, MIGRATE_SYNC, MR_SYSCALL);
1278 putback_movable_pages(&pagelist);
1281 up_read(&mm->mmap_sem);
1286 * Migrate an array of page address onto an array of nodes and fill
1287 * the corresponding array of status.
1289 static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes,
1290 unsigned long nr_pages,
1291 const void __user * __user *pages,
1292 const int __user *nodes,
1293 int __user *status, int flags)
1295 struct page_to_node *pm;
1296 unsigned long chunk_nr_pages;
1297 unsigned long chunk_start;
1301 pm = (struct page_to_node *)__get_free_page(GFP_KERNEL);
1308 * Store a chunk of page_to_node array in a page,
1309 * but keep the last one as a marker
1311 chunk_nr_pages = (PAGE_SIZE / sizeof(struct page_to_node)) - 1;
1313 for (chunk_start = 0;
1314 chunk_start < nr_pages;
1315 chunk_start += chunk_nr_pages) {
1318 if (chunk_start + chunk_nr_pages > nr_pages)
1319 chunk_nr_pages = nr_pages - chunk_start;
1321 /* fill the chunk pm with addrs and nodes from user-space */
1322 for (j = 0; j < chunk_nr_pages; j++) {
1323 const void __user *p;
1327 if (get_user(p, pages + j + chunk_start))
1329 pm[j].addr = (unsigned long) p;
1331 if (get_user(node, nodes + j + chunk_start))
1335 if (node < 0 || node >= MAX_NUMNODES)
1338 if (!node_state(node, N_MEMORY))
1342 if (!node_isset(node, task_nodes))
1348 /* End marker for this chunk */
1349 pm[chunk_nr_pages].node = MAX_NUMNODES;
1351 /* Migrate this chunk */
1352 err = do_move_page_to_node_array(mm, pm,
1353 flags & MPOL_MF_MOVE_ALL);
1357 /* Return status information */
1358 for (j = 0; j < chunk_nr_pages; j++)
1359 if (put_user(pm[j].status, status + j + chunk_start)) {
1367 free_page((unsigned long)pm);
1373 * Determine the nodes of an array of pages and store it in an array of status.
1375 static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages,
1376 const void __user **pages, int *status)
1380 down_read(&mm->mmap_sem);
1382 for (i = 0; i < nr_pages; i++) {
1383 unsigned long addr = (unsigned long)(*pages);
1384 struct vm_area_struct *vma;
1388 vma = find_vma(mm, addr);
1389 if (!vma || addr < vma->vm_start)
1392 page = follow_page(vma, addr, 0);
1394 err = PTR_ERR(page);
1399 /* Use PageReserved to check for zero page */
1400 if (!page || PageReserved(page))
1403 err = page_to_nid(page);
1411 up_read(&mm->mmap_sem);
1415 * Determine the nodes of a user array of pages and store it in
1416 * a user array of status.
1418 static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages,
1419 const void __user * __user *pages,
1422 #define DO_PAGES_STAT_CHUNK_NR 16
1423 const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR];
1424 int chunk_status[DO_PAGES_STAT_CHUNK_NR];
1427 unsigned long chunk_nr;
1429 chunk_nr = nr_pages;
1430 if (chunk_nr > DO_PAGES_STAT_CHUNK_NR)
1431 chunk_nr = DO_PAGES_STAT_CHUNK_NR;
1433 if (copy_from_user(chunk_pages, pages, chunk_nr * sizeof(*chunk_pages)))
1436 do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status);
1438 if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status)))
1443 nr_pages -= chunk_nr;
1445 return nr_pages ? -EFAULT : 0;
1449 * Move a list of pages in the address space of the currently executing
1452 SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages,
1453 const void __user * __user *, pages,
1454 const int __user *, nodes,
1455 int __user *, status, int, flags)
1457 const struct cred *cred = current_cred(), *tcred;
1458 struct task_struct *task;
1459 struct mm_struct *mm;
1461 nodemask_t task_nodes;
1464 if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
1467 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1470 /* Find the mm_struct */
1472 task = pid ? find_task_by_vpid(pid) : current;
1477 get_task_struct(task);
1480 * Check if this process has the right to modify the specified
1481 * process. The right exists if the process has administrative
1482 * capabilities, superuser privileges or the same
1483 * userid as the target process.
1485 tcred = __task_cred(task);
1486 if (!uid_eq(cred->euid, tcred->suid) && !uid_eq(cred->euid, tcred->uid) &&
1487 !uid_eq(cred->uid, tcred->suid) && !uid_eq(cred->uid, tcred->uid) &&
1488 !capable(CAP_SYS_NICE)) {
1495 err = security_task_movememory(task);
1499 task_nodes = cpuset_mems_allowed(task);
1500 mm = get_task_mm(task);
1501 put_task_struct(task);
1507 err = do_pages_move(mm, task_nodes, nr_pages, pages,
1508 nodes, status, flags);
1510 err = do_pages_stat(mm, nr_pages, pages, status);
1516 put_task_struct(task);
1520 #ifdef CONFIG_NUMA_BALANCING
1522 * Returns true if this is a safe migration target node for misplaced NUMA
1523 * pages. Currently it only checks the watermarks which crude
1525 static bool migrate_balanced_pgdat(struct pglist_data *pgdat,
1526 unsigned long nr_migrate_pages)
1529 for (z = pgdat->nr_zones - 1; z >= 0; z--) {
1530 struct zone *zone = pgdat->node_zones + z;
1532 if (!populated_zone(zone))
1535 if (!zone_reclaimable(zone))
1538 /* Avoid waking kswapd by allocating pages_to_migrate pages. */
1539 if (!zone_watermark_ok(zone, 0,
1540 high_wmark_pages(zone) +
1549 static struct page *alloc_misplaced_dst_page(struct page *page,
1553 int nid = (int) data;
1554 struct page *newpage;
1556 newpage = alloc_pages_exact_node(nid,
1557 (GFP_HIGHUSER_MOVABLE |
1558 __GFP_THISNODE | __GFP_NOMEMALLOC |
1559 __GFP_NORETRY | __GFP_NOWARN) &
1566 * page migration rate limiting control.
1567 * Do not migrate more than @pages_to_migrate in a @migrate_interval_millisecs
1568 * window of time. Default here says do not migrate more than 1280M per second.
1570 static unsigned int migrate_interval_millisecs __read_mostly = 100;
1571 static unsigned int ratelimit_pages __read_mostly = 128 << (20 - PAGE_SHIFT);
1573 /* Returns true if the node is migrate rate-limited after the update */
1574 static bool numamigrate_update_ratelimit(pg_data_t *pgdat,
1575 unsigned long nr_pages)
1578 * Rate-limit the amount of data that is being migrated to a node.
1579 * Optimal placement is no good if the memory bus is saturated and
1580 * all the time is being spent migrating!
1582 if (time_after(jiffies, pgdat->numabalancing_migrate_next_window)) {
1583 spin_lock(&pgdat->numabalancing_migrate_lock);
1584 pgdat->numabalancing_migrate_nr_pages = 0;
1585 pgdat->numabalancing_migrate_next_window = jiffies +
1586 msecs_to_jiffies(migrate_interval_millisecs);
1587 spin_unlock(&pgdat->numabalancing_migrate_lock);
1589 if (pgdat->numabalancing_migrate_nr_pages > ratelimit_pages) {
1590 trace_mm_numa_migrate_ratelimit(current, pgdat->node_id,
1596 * This is an unlocked non-atomic update so errors are possible.
1597 * The consequences are failing to migrate when we potentiall should
1598 * have which is not severe enough to warrant locking. If it is ever
1599 * a problem, it can be converted to a per-cpu counter.
1601 pgdat->numabalancing_migrate_nr_pages += nr_pages;
1605 static int numamigrate_isolate_page(pg_data_t *pgdat, struct page *page)
1609 VM_BUG_ON_PAGE(compound_order(page) && !PageTransHuge(page), page);
1611 /* Avoid migrating to a node that is nearly full */
1612 if (!migrate_balanced_pgdat(pgdat, 1UL << compound_order(page)))
1615 if (isolate_lru_page(page))
1619 * migrate_misplaced_transhuge_page() skips page migration's usual
1620 * check on page_count(), so we must do it here, now that the page
1621 * has been isolated: a GUP pin, or any other pin, prevents migration.
1622 * The expected page count is 3: 1 for page's mapcount and 1 for the
1623 * caller's pin and 1 for the reference taken by isolate_lru_page().
1625 if (PageTransHuge(page) && page_count(page) != 3) {
1626 putback_lru_page(page);
1630 page_lru = page_is_file_cache(page);
1631 mod_zone_page_state(page_zone(page), NR_ISOLATED_ANON + page_lru,
1632 hpage_nr_pages(page));
1635 * Isolating the page has taken another reference, so the
1636 * caller's reference can be safely dropped without the page
1637 * disappearing underneath us during migration.
1643 bool pmd_trans_migrating(pmd_t pmd)
1645 struct page *page = pmd_page(pmd);
1646 return PageLocked(page);
1650 * Attempt to migrate a misplaced page to the specified destination
1651 * node. Caller is expected to have an elevated reference count on
1652 * the page that will be dropped by this function before returning.
1654 int migrate_misplaced_page(struct page *page, struct vm_area_struct *vma,
1657 pg_data_t *pgdat = NODE_DATA(node);
1660 LIST_HEAD(migratepages);
1663 * Don't migrate file pages that are mapped in multiple processes
1664 * with execute permissions as they are probably shared libraries.
1666 if (page_mapcount(page) != 1 && page_is_file_cache(page) &&
1667 (vma->vm_flags & VM_EXEC))
1671 * Rate-limit the amount of data that is being migrated to a node.
1672 * Optimal placement is no good if the memory bus is saturated and
1673 * all the time is being spent migrating!
1675 if (numamigrate_update_ratelimit(pgdat, 1))
1678 isolated = numamigrate_isolate_page(pgdat, page);
1682 list_add(&page->lru, &migratepages);
1683 nr_remaining = migrate_pages(&migratepages, alloc_misplaced_dst_page,
1684 NULL, node, MIGRATE_ASYNC,
1687 if (!list_empty(&migratepages)) {
1688 list_del(&page->lru);
1689 dec_zone_page_state(page, NR_ISOLATED_ANON +
1690 page_is_file_cache(page));
1691 putback_lru_page(page);
1695 count_vm_numa_event(NUMA_PAGE_MIGRATE);
1696 BUG_ON(!list_empty(&migratepages));
1703 #endif /* CONFIG_NUMA_BALANCING */
1705 #if defined(CONFIG_NUMA_BALANCING) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
1707 * Migrates a THP to a given target node. page must be locked and is unlocked
1710 int migrate_misplaced_transhuge_page(struct mm_struct *mm,
1711 struct vm_area_struct *vma,
1712 pmd_t *pmd, pmd_t entry,
1713 unsigned long address,
1714 struct page *page, int node)
1717 pg_data_t *pgdat = NODE_DATA(node);
1719 struct page *new_page = NULL;
1720 int page_lru = page_is_file_cache(page);
1721 unsigned long mmun_start = address & HPAGE_PMD_MASK;
1722 unsigned long mmun_end = mmun_start + HPAGE_PMD_SIZE;
1726 * Rate-limit the amount of data that is being migrated to a node.
1727 * Optimal placement is no good if the memory bus is saturated and
1728 * all the time is being spent migrating!
1730 if (numamigrate_update_ratelimit(pgdat, HPAGE_PMD_NR))
1733 new_page = alloc_pages_node(node,
1734 (GFP_TRANSHUGE | __GFP_THISNODE) & ~__GFP_WAIT,
1739 isolated = numamigrate_isolate_page(pgdat, page);
1745 if (mm_tlb_flush_pending(mm))
1746 flush_tlb_range(vma, mmun_start, mmun_end);
1748 /* Prepare a page as a migration target */
1749 __set_page_locked(new_page);
1750 SetPageSwapBacked(new_page);
1752 /* anon mapping, we can simply copy page->mapping to the new page: */
1753 new_page->mapping = page->mapping;
1754 new_page->index = page->index;
1755 migrate_page_copy(new_page, page);
1756 WARN_ON(PageLRU(new_page));
1758 /* Recheck the target PMD */
1759 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1760 ptl = pmd_lock(mm, pmd);
1761 if (unlikely(!pmd_same(*pmd, entry) || page_count(page) != 2)) {
1764 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1766 /* Reverse changes made by migrate_page_copy() */
1767 if (TestClearPageActive(new_page))
1768 SetPageActive(page);
1769 if (TestClearPageUnevictable(new_page))
1770 SetPageUnevictable(page);
1771 mlock_migrate_page(page, new_page);
1773 unlock_page(new_page);
1774 put_page(new_page); /* Free it */
1776 /* Retake the callers reference and putback on LRU */
1778 putback_lru_page(page);
1779 mod_zone_page_state(page_zone(page),
1780 NR_ISOLATED_ANON + page_lru, -HPAGE_PMD_NR);
1786 entry = mk_pmd(new_page, vma->vm_page_prot);
1787 entry = pmd_mkhuge(entry);
1788 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1791 * Clear the old entry under pagetable lock and establish the new PTE.
1792 * Any parallel GUP will either observe the old page blocking on the
1793 * page lock, block on the page table lock or observe the new page.
1794 * The SetPageUptodate on the new page and page_add_new_anon_rmap
1795 * guarantee the copy is visible before the pagetable update.
1797 flush_cache_range(vma, mmun_start, mmun_end);
1798 page_add_anon_rmap(new_page, vma, mmun_start);
1799 pmdp_clear_flush_notify(vma, mmun_start, pmd);
1800 set_pmd_at(mm, mmun_start, pmd, entry);
1801 flush_tlb_range(vma, mmun_start, mmun_end);
1802 update_mmu_cache_pmd(vma, address, &entry);
1804 if (page_count(page) != 2) {
1805 set_pmd_at(mm, mmun_start, pmd, orig_entry);
1806 flush_tlb_range(vma, mmun_start, mmun_end);
1807 mmu_notifier_invalidate_range(mm, mmun_start, mmun_end);
1808 update_mmu_cache_pmd(vma, address, &entry);
1809 page_remove_rmap(new_page);
1813 mem_cgroup_migrate(page, new_page, false);
1815 page_remove_rmap(page);
1818 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1820 /* Take an "isolate" reference and put new page on the LRU. */
1822 putback_lru_page(new_page);
1824 unlock_page(new_page);
1826 put_page(page); /* Drop the rmap reference */
1827 put_page(page); /* Drop the LRU isolation reference */
1829 count_vm_events(PGMIGRATE_SUCCESS, HPAGE_PMD_NR);
1830 count_vm_numa_events(NUMA_PAGE_MIGRATE, HPAGE_PMD_NR);
1832 mod_zone_page_state(page_zone(page),
1833 NR_ISOLATED_ANON + page_lru,
1838 count_vm_events(PGMIGRATE_FAIL, HPAGE_PMD_NR);
1840 ptl = pmd_lock(mm, pmd);
1841 if (pmd_same(*pmd, entry)) {
1842 entry = pmd_modify(entry, vma->vm_page_prot);
1843 set_pmd_at(mm, mmun_start, pmd, entry);
1844 update_mmu_cache_pmd(vma, address, &entry);
1853 #endif /* CONFIG_NUMA_BALANCING */
1855 #endif /* CONFIG_NUMA */