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 ClearPageSwapCache(page);
541 ClearPagePrivate(page);
542 set_page_private(page, 0);
545 * If any waiters have accumulated on the new page then
548 if (PageWriteback(newpage))
549 end_page_writeback(newpage);
552 /************************************************************
553 * Migration functions
554 ***********************************************************/
557 * Common logic to directly migrate a single page suitable for
558 * pages that do not use PagePrivate/PagePrivate2.
560 * Pages are locked upon entry and exit.
562 int migrate_page(struct address_space *mapping,
563 struct page *newpage, struct page *page,
564 enum migrate_mode mode)
568 BUG_ON(PageWriteback(page)); /* Writeback must be complete */
570 rc = migrate_page_move_mapping(mapping, newpage, page, NULL, mode, 0);
572 if (rc != MIGRATEPAGE_SUCCESS)
575 migrate_page_copy(newpage, page);
576 return MIGRATEPAGE_SUCCESS;
578 EXPORT_SYMBOL(migrate_page);
582 * Migration function for pages with buffers. This function can only be used
583 * if the underlying filesystem guarantees that no other references to "page"
586 int buffer_migrate_page(struct address_space *mapping,
587 struct page *newpage, struct page *page, enum migrate_mode mode)
589 struct buffer_head *bh, *head;
592 if (!page_has_buffers(page))
593 return migrate_page(mapping, newpage, page, mode);
595 head = page_buffers(page);
597 rc = migrate_page_move_mapping(mapping, newpage, page, head, mode, 0);
599 if (rc != MIGRATEPAGE_SUCCESS)
603 * In the async case, migrate_page_move_mapping locked the buffers
604 * with an IRQ-safe spinlock held. In the sync case, the buffers
605 * need to be locked now
607 if (mode != MIGRATE_ASYNC)
608 BUG_ON(!buffer_migrate_lock_buffers(head, mode));
610 ClearPagePrivate(page);
611 set_page_private(newpage, page_private(page));
612 set_page_private(page, 0);
618 set_bh_page(bh, newpage, bh_offset(bh));
619 bh = bh->b_this_page;
621 } while (bh != head);
623 SetPagePrivate(newpage);
625 migrate_page_copy(newpage, page);
631 bh = bh->b_this_page;
633 } while (bh != head);
635 return MIGRATEPAGE_SUCCESS;
637 EXPORT_SYMBOL(buffer_migrate_page);
641 * Writeback a page to clean the dirty state
643 static int writeout(struct address_space *mapping, struct page *page)
645 struct writeback_control wbc = {
646 .sync_mode = WB_SYNC_NONE,
649 .range_end = LLONG_MAX,
654 if (!mapping->a_ops->writepage)
655 /* No write method for the address space */
658 if (!clear_page_dirty_for_io(page))
659 /* Someone else already triggered a write */
663 * A dirty page may imply that the underlying filesystem has
664 * the page on some queue. So the page must be clean for
665 * migration. Writeout may mean we loose the lock and the
666 * page state is no longer what we checked for earlier.
667 * At this point we know that the migration attempt cannot
670 remove_migration_ptes(page, page);
672 rc = mapping->a_ops->writepage(page, &wbc);
674 if (rc != AOP_WRITEPAGE_ACTIVATE)
675 /* unlocked. Relock */
678 return (rc < 0) ? -EIO : -EAGAIN;
682 * Default handling if a filesystem does not provide a migration function.
684 static int fallback_migrate_page(struct address_space *mapping,
685 struct page *newpage, struct page *page, enum migrate_mode mode)
687 if (PageDirty(page)) {
688 /* Only writeback pages in full synchronous migration */
689 if (mode != MIGRATE_SYNC)
691 return writeout(mapping, page);
695 * Buffers may be managed in a filesystem specific way.
696 * We must have no buffers or drop them.
698 if (page_has_private(page) &&
699 !try_to_release_page(page, GFP_KERNEL))
702 return migrate_page(mapping, newpage, page, mode);
706 * Move a page to a newly allocated page
707 * The page is locked and all ptes have been successfully removed.
709 * The new page will have replaced the old page if this function
714 * MIGRATEPAGE_SUCCESS - success
716 static int move_to_new_page(struct page *newpage, struct page *page,
717 int page_was_mapped, enum migrate_mode mode)
719 struct address_space *mapping;
723 * Block others from accessing the page when we get around to
724 * establishing additional references. We are the only one
725 * holding a reference to the new page at this point.
727 if (!trylock_page(newpage))
730 /* Prepare mapping for the new page.*/
731 newpage->index = page->index;
732 newpage->mapping = page->mapping;
733 if (PageSwapBacked(page))
734 SetPageSwapBacked(newpage);
736 mapping = page_mapping(page);
738 rc = migrate_page(mapping, newpage, page, mode);
739 else if (mapping->a_ops->migratepage)
741 * Most pages have a mapping and most filesystems provide a
742 * migratepage callback. Anonymous pages are part of swap
743 * space which also has its own migratepage callback. This
744 * is the most common path for page migration.
746 rc = mapping->a_ops->migratepage(mapping,
747 newpage, page, mode);
749 rc = fallback_migrate_page(mapping, newpage, page, mode);
751 if (rc != MIGRATEPAGE_SUCCESS) {
752 newpage->mapping = NULL;
754 mem_cgroup_migrate(page, newpage, false);
756 remove_migration_ptes(page, newpage);
757 page->mapping = NULL;
760 unlock_page(newpage);
765 static int __unmap_and_move(struct page *page, struct page *newpage,
766 int force, enum migrate_mode mode)
769 int page_was_mapped = 0;
770 struct anon_vma *anon_vma = NULL;
772 if (!trylock_page(page)) {
773 if (!force || mode == MIGRATE_ASYNC)
777 * It's not safe for direct compaction to call lock_page.
778 * For example, during page readahead pages are added locked
779 * to the LRU. Later, when the IO completes the pages are
780 * marked uptodate and unlocked. However, the queueing
781 * could be merging multiple pages for one bio (e.g.
782 * mpage_readpages). If an allocation happens for the
783 * second or third page, the process can end up locking
784 * the same page twice and deadlocking. Rather than
785 * trying to be clever about what pages can be locked,
786 * avoid the use of lock_page for direct compaction
789 if (current->flags & PF_MEMALLOC)
795 if (PageWriteback(page)) {
797 * Only in the case of a full synchronous migration is it
798 * necessary to wait for PageWriteback. In the async case,
799 * the retry loop is too short and in the sync-light case,
800 * the overhead of stalling is too much
802 if (mode != MIGRATE_SYNC) {
808 wait_on_page_writeback(page);
811 * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
812 * we cannot notice that anon_vma is freed while we migrates a page.
813 * This get_anon_vma() delays freeing anon_vma pointer until the end
814 * of migration. File cache pages are no problem because of page_lock()
815 * File Caches may use write_page() or lock_page() in migration, then,
816 * just care Anon page here.
818 if (PageAnon(page) && !PageKsm(page)) {
820 * Only page_lock_anon_vma_read() understands the subtleties of
821 * getting a hold on an anon_vma from outside one of its mms.
823 anon_vma = page_get_anon_vma(page);
828 } else if (PageSwapCache(page)) {
830 * We cannot be sure that the anon_vma of an unmapped
831 * swapcache page is safe to use because we don't
832 * know in advance if the VMA that this page belonged
833 * to still exists. If the VMA and others sharing the
834 * data have been freed, then the anon_vma could
835 * already be invalid.
837 * To avoid this possibility, swapcache pages get
838 * migrated but are not remapped when migration
846 if (unlikely(isolated_balloon_page(page))) {
848 * A ballooned page does not need any special attention from
849 * physical to virtual reverse mapping procedures.
850 * Skip any attempt to unmap PTEs or to remap swap cache,
851 * in order to avoid burning cycles at rmap level, and perform
852 * the page migration right away (proteced by page lock).
854 rc = balloon_page_migrate(newpage, page, mode);
859 * Corner case handling:
860 * 1. When a new swap-cache page is read into, it is added to the LRU
861 * and treated as swapcache but it has no rmap yet.
862 * Calling try_to_unmap() against a page->mapping==NULL page will
863 * trigger a BUG. So handle it here.
864 * 2. An orphaned page (see truncate_complete_page) might have
865 * fs-private metadata. The page can be picked up due to memory
866 * offlining. Everywhere else except page reclaim, the page is
867 * invisible to the vm, so the page can not be migrated. So try to
868 * free the metadata, so the page can be freed.
870 if (!page->mapping) {
871 VM_BUG_ON_PAGE(PageAnon(page), page);
872 if (page_has_private(page)) {
873 try_to_free_buffers(page);
879 /* Establish migration ptes or remove ptes */
880 if (page_mapped(page)) {
882 TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
887 if (!page_mapped(page))
888 rc = move_to_new_page(newpage, page, page_was_mapped, mode);
890 if (rc && page_was_mapped)
891 remove_migration_ptes(page, page);
893 /* Drop an anon_vma reference if we took one */
895 put_anon_vma(anon_vma);
904 * gcc 4.7 and 4.8 on arm get an ICEs when inlining unmap_and_move(). Work
907 #if (GCC_VERSION >= 40700 && GCC_VERSION < 40900) && defined(CONFIG_ARM)
908 #define ICE_noinline noinline
914 * Obtain the lock on page, remove all ptes and migrate the page
915 * to the newly allocated page in newpage.
917 static ICE_noinline int unmap_and_move(new_page_t get_new_page,
918 free_page_t put_new_page,
919 unsigned long private, struct page *page,
920 int force, enum migrate_mode mode)
924 struct page *newpage = get_new_page(page, private, &result);
929 if (page_count(page) == 1) {
930 /* page was freed from under us. So we are done. */
934 if (unlikely(PageTransHuge(page)))
935 if (unlikely(split_huge_page(page)))
938 rc = __unmap_and_move(page, newpage, force, mode);
943 * A page that has been migrated has all references
944 * removed and will be freed. A page that has not been
945 * migrated will have kepts its references and be
948 list_del(&page->lru);
949 dec_zone_page_state(page, NR_ISOLATED_ANON +
950 page_is_file_cache(page));
951 putback_lru_page(page);
955 * If migration was not successful and there's a freeing callback, use
956 * it. Otherwise, putback_lru_page() will drop the reference grabbed
959 if (rc != MIGRATEPAGE_SUCCESS && put_new_page) {
960 ClearPageSwapBacked(newpage);
961 put_new_page(newpage, private);
962 } else if (unlikely(__is_movable_balloon_page(newpage))) {
963 /* drop our reference, page already in the balloon */
966 putback_lru_page(newpage);
972 *result = page_to_nid(newpage);
978 * Counterpart of unmap_and_move_page() for hugepage migration.
980 * This function doesn't wait the completion of hugepage I/O
981 * because there is no race between I/O and migration for hugepage.
982 * Note that currently hugepage I/O occurs only in direct I/O
983 * where no lock is held and PG_writeback is irrelevant,
984 * and writeback status of all subpages are counted in the reference
985 * count of the head page (i.e. if all subpages of a 2MB hugepage are
986 * under direct I/O, the reference of the head page is 512 and a bit more.)
987 * This means that when we try to migrate hugepage whose subpages are
988 * doing direct I/O, some references remain after try_to_unmap() and
989 * hugepage migration fails without data corruption.
991 * There is also no race when direct I/O is issued on the page under migration,
992 * because then pte is replaced with migration swap entry and direct I/O code
993 * will wait in the page fault for migration to complete.
995 static int unmap_and_move_huge_page(new_page_t get_new_page,
996 free_page_t put_new_page, unsigned long private,
997 struct page *hpage, int force,
998 enum migrate_mode mode)
1002 int page_was_mapped = 0;
1003 struct page *new_hpage;
1004 struct anon_vma *anon_vma = NULL;
1007 * Movability of hugepages depends on architectures and hugepage size.
1008 * This check is necessary because some callers of hugepage migration
1009 * like soft offline and memory hotremove don't walk through page
1010 * tables or check whether the hugepage is pmd-based or not before
1011 * kicking migration.
1013 if (!hugepage_migration_supported(page_hstate(hpage))) {
1014 putback_active_hugepage(hpage);
1018 new_hpage = get_new_page(hpage, private, &result);
1024 if (!trylock_page(hpage)) {
1025 if (!force || mode != MIGRATE_SYNC)
1030 if (PageAnon(hpage))
1031 anon_vma = page_get_anon_vma(hpage);
1033 if (page_mapped(hpage)) {
1035 TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
1036 page_was_mapped = 1;
1039 if (!page_mapped(hpage))
1040 rc = move_to_new_page(new_hpage, hpage, page_was_mapped, mode);
1042 if (rc != MIGRATEPAGE_SUCCESS && page_was_mapped)
1043 remove_migration_ptes(hpage, hpage);
1046 put_anon_vma(anon_vma);
1048 if (rc == MIGRATEPAGE_SUCCESS)
1049 hugetlb_cgroup_migrate(hpage, new_hpage);
1054 putback_active_hugepage(hpage);
1057 * If migration was not successful and there's a freeing callback, use
1058 * it. Otherwise, put_page() will drop the reference grabbed during
1061 if (rc != MIGRATEPAGE_SUCCESS && put_new_page)
1062 put_new_page(new_hpage, private);
1064 put_page(new_hpage);
1070 *result = page_to_nid(new_hpage);
1076 * migrate_pages - migrate the pages specified in a list, to the free pages
1077 * supplied as the target for the page migration
1079 * @from: The list of pages to be migrated.
1080 * @get_new_page: The function used to allocate free pages to be used
1081 * as the target of the page migration.
1082 * @put_new_page: The function used to free target pages if migration
1083 * fails, or NULL if no special handling is necessary.
1084 * @private: Private data to be passed on to get_new_page()
1085 * @mode: The migration mode that specifies the constraints for
1086 * page migration, if any.
1087 * @reason: The reason for page migration.
1089 * The function returns after 10 attempts or if no pages are movable any more
1090 * because the list has become empty or no retryable pages exist any more.
1091 * The caller should call putback_lru_pages() to return pages to the LRU
1092 * or free list only if ret != 0.
1094 * Returns the number of pages that were not migrated, or an error code.
1096 int migrate_pages(struct list_head *from, new_page_t get_new_page,
1097 free_page_t put_new_page, unsigned long private,
1098 enum migrate_mode mode, int reason)
1102 int nr_succeeded = 0;
1106 int swapwrite = current->flags & PF_SWAPWRITE;
1110 current->flags |= PF_SWAPWRITE;
1112 for(pass = 0; pass < 10 && retry; pass++) {
1115 list_for_each_entry_safe(page, page2, from, lru) {
1119 rc = unmap_and_move_huge_page(get_new_page,
1120 put_new_page, private, page,
1123 rc = unmap_and_move(get_new_page, put_new_page,
1124 private, page, pass > 2, mode);
1132 case MIGRATEPAGE_SUCCESS:
1137 * Permanent failure (-EBUSY, -ENOSYS, etc.):
1138 * unlike -EAGAIN case, the failed page is
1139 * removed from migration page list and not
1140 * retried in the next outer loop.
1147 rc = nr_failed + retry;
1150 count_vm_events(PGMIGRATE_SUCCESS, nr_succeeded);
1152 count_vm_events(PGMIGRATE_FAIL, nr_failed);
1153 trace_mm_migrate_pages(nr_succeeded, nr_failed, mode, reason);
1156 current->flags &= ~PF_SWAPWRITE;
1163 * Move a list of individual pages
1165 struct page_to_node {
1172 static struct page *new_page_node(struct page *p, unsigned long private,
1175 struct page_to_node *pm = (struct page_to_node *)private;
1177 while (pm->node != MAX_NUMNODES && pm->page != p)
1180 if (pm->node == MAX_NUMNODES)
1183 *result = &pm->status;
1186 return alloc_huge_page_node(page_hstate(compound_head(p)),
1189 return alloc_pages_exact_node(pm->node,
1190 GFP_HIGHUSER_MOVABLE | __GFP_THISNODE, 0);
1194 * Move a set of pages as indicated in the pm array. The addr
1195 * field must be set to the virtual address of the page to be moved
1196 * and the node number must contain a valid target node.
1197 * The pm array ends with node = MAX_NUMNODES.
1199 static int do_move_page_to_node_array(struct mm_struct *mm,
1200 struct page_to_node *pm,
1204 struct page_to_node *pp;
1205 LIST_HEAD(pagelist);
1207 down_read(&mm->mmap_sem);
1210 * Build a list of pages to migrate
1212 for (pp = pm; pp->node != MAX_NUMNODES; pp++) {
1213 struct vm_area_struct *vma;
1217 vma = find_vma(mm, pp->addr);
1218 if (!vma || pp->addr < vma->vm_start || !vma_migratable(vma))
1221 page = follow_page(vma, pp->addr, FOLL_GET|FOLL_SPLIT);
1223 err = PTR_ERR(page);
1231 /* Use PageReserved to check for zero page */
1232 if (PageReserved(page))
1236 err = page_to_nid(page);
1238 if (err == pp->node)
1240 * Node already in the right place
1245 if (page_mapcount(page) > 1 &&
1249 if (PageHuge(page)) {
1251 isolate_huge_page(page, &pagelist);
1255 err = isolate_lru_page(page);
1257 list_add_tail(&page->lru, &pagelist);
1258 inc_zone_page_state(page, NR_ISOLATED_ANON +
1259 page_is_file_cache(page));
1263 * Either remove the duplicate refcount from
1264 * isolate_lru_page() or drop the page ref if it was
1273 if (!list_empty(&pagelist)) {
1274 err = migrate_pages(&pagelist, new_page_node, NULL,
1275 (unsigned long)pm, MIGRATE_SYNC, MR_SYSCALL);
1277 putback_movable_pages(&pagelist);
1280 up_read(&mm->mmap_sem);
1285 * Migrate an array of page address onto an array of nodes and fill
1286 * the corresponding array of status.
1288 static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes,
1289 unsigned long nr_pages,
1290 const void __user * __user *pages,
1291 const int __user *nodes,
1292 int __user *status, int flags)
1294 struct page_to_node *pm;
1295 unsigned long chunk_nr_pages;
1296 unsigned long chunk_start;
1300 pm = (struct page_to_node *)__get_free_page(GFP_KERNEL);
1307 * Store a chunk of page_to_node array in a page,
1308 * but keep the last one as a marker
1310 chunk_nr_pages = (PAGE_SIZE / sizeof(struct page_to_node)) - 1;
1312 for (chunk_start = 0;
1313 chunk_start < nr_pages;
1314 chunk_start += chunk_nr_pages) {
1317 if (chunk_start + chunk_nr_pages > nr_pages)
1318 chunk_nr_pages = nr_pages - chunk_start;
1320 /* fill the chunk pm with addrs and nodes from user-space */
1321 for (j = 0; j < chunk_nr_pages; j++) {
1322 const void __user *p;
1326 if (get_user(p, pages + j + chunk_start))
1328 pm[j].addr = (unsigned long) p;
1330 if (get_user(node, nodes + j + chunk_start))
1334 if (node < 0 || node >= MAX_NUMNODES)
1337 if (!node_state(node, N_MEMORY))
1341 if (!node_isset(node, task_nodes))
1347 /* End marker for this chunk */
1348 pm[chunk_nr_pages].node = MAX_NUMNODES;
1350 /* Migrate this chunk */
1351 err = do_move_page_to_node_array(mm, pm,
1352 flags & MPOL_MF_MOVE_ALL);
1356 /* Return status information */
1357 for (j = 0; j < chunk_nr_pages; j++)
1358 if (put_user(pm[j].status, status + j + chunk_start)) {
1366 free_page((unsigned long)pm);
1372 * Determine the nodes of an array of pages and store it in an array of status.
1374 static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages,
1375 const void __user **pages, int *status)
1379 down_read(&mm->mmap_sem);
1381 for (i = 0; i < nr_pages; i++) {
1382 unsigned long addr = (unsigned long)(*pages);
1383 struct vm_area_struct *vma;
1387 vma = find_vma(mm, addr);
1388 if (!vma || addr < vma->vm_start)
1391 page = follow_page(vma, addr, 0);
1393 err = PTR_ERR(page);
1398 /* Use PageReserved to check for zero page */
1399 if (!page || PageReserved(page))
1402 err = page_to_nid(page);
1410 up_read(&mm->mmap_sem);
1414 * Determine the nodes of a user array of pages and store it in
1415 * a user array of status.
1417 static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages,
1418 const void __user * __user *pages,
1421 #define DO_PAGES_STAT_CHUNK_NR 16
1422 const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR];
1423 int chunk_status[DO_PAGES_STAT_CHUNK_NR];
1426 unsigned long chunk_nr;
1428 chunk_nr = nr_pages;
1429 if (chunk_nr > DO_PAGES_STAT_CHUNK_NR)
1430 chunk_nr = DO_PAGES_STAT_CHUNK_NR;
1432 if (copy_from_user(chunk_pages, pages, chunk_nr * sizeof(*chunk_pages)))
1435 do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status);
1437 if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status)))
1442 nr_pages -= chunk_nr;
1444 return nr_pages ? -EFAULT : 0;
1448 * Move a list of pages in the address space of the currently executing
1451 SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages,
1452 const void __user * __user *, pages,
1453 const int __user *, nodes,
1454 int __user *, status, int, flags)
1456 const struct cred *cred = current_cred(), *tcred;
1457 struct task_struct *task;
1458 struct mm_struct *mm;
1460 nodemask_t task_nodes;
1463 if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
1466 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1469 /* Find the mm_struct */
1471 task = pid ? find_task_by_vpid(pid) : current;
1476 get_task_struct(task);
1479 * Check if this process has the right to modify the specified
1480 * process. The right exists if the process has administrative
1481 * capabilities, superuser privileges or the same
1482 * userid as the target process.
1484 tcred = __task_cred(task);
1485 if (!uid_eq(cred->euid, tcred->suid) && !uid_eq(cred->euid, tcred->uid) &&
1486 !uid_eq(cred->uid, tcred->suid) && !uid_eq(cred->uid, tcred->uid) &&
1487 !capable(CAP_SYS_NICE)) {
1494 err = security_task_movememory(task);
1498 task_nodes = cpuset_mems_allowed(task);
1499 mm = get_task_mm(task);
1500 put_task_struct(task);
1506 err = do_pages_move(mm, task_nodes, nr_pages, pages,
1507 nodes, status, flags);
1509 err = do_pages_stat(mm, nr_pages, pages, status);
1515 put_task_struct(task);
1519 #ifdef CONFIG_NUMA_BALANCING
1521 * Returns true if this is a safe migration target node for misplaced NUMA
1522 * pages. Currently it only checks the watermarks which crude
1524 static bool migrate_balanced_pgdat(struct pglist_data *pgdat,
1525 unsigned long nr_migrate_pages)
1528 for (z = pgdat->nr_zones - 1; z >= 0; z--) {
1529 struct zone *zone = pgdat->node_zones + z;
1531 if (!populated_zone(zone))
1534 if (!zone_reclaimable(zone))
1537 /* Avoid waking kswapd by allocating pages_to_migrate pages. */
1538 if (!zone_watermark_ok(zone, 0,
1539 high_wmark_pages(zone) +
1548 static struct page *alloc_misplaced_dst_page(struct page *page,
1552 int nid = (int) data;
1553 struct page *newpage;
1555 newpage = alloc_pages_exact_node(nid,
1556 (GFP_HIGHUSER_MOVABLE |
1557 __GFP_THISNODE | __GFP_NOMEMALLOC |
1558 __GFP_NORETRY | __GFP_NOWARN) &
1565 * page migration rate limiting control.
1566 * Do not migrate more than @pages_to_migrate in a @migrate_interval_millisecs
1567 * window of time. Default here says do not migrate more than 1280M per second.
1569 static unsigned int migrate_interval_millisecs __read_mostly = 100;
1570 static unsigned int ratelimit_pages __read_mostly = 128 << (20 - PAGE_SHIFT);
1572 /* Returns true if the node is migrate rate-limited after the update */
1573 static bool numamigrate_update_ratelimit(pg_data_t *pgdat,
1574 unsigned long nr_pages)
1577 * Rate-limit the amount of data that is being migrated to a node.
1578 * Optimal placement is no good if the memory bus is saturated and
1579 * all the time is being spent migrating!
1581 if (time_after(jiffies, pgdat->numabalancing_migrate_next_window)) {
1582 spin_lock(&pgdat->numabalancing_migrate_lock);
1583 pgdat->numabalancing_migrate_nr_pages = 0;
1584 pgdat->numabalancing_migrate_next_window = jiffies +
1585 msecs_to_jiffies(migrate_interval_millisecs);
1586 spin_unlock(&pgdat->numabalancing_migrate_lock);
1588 if (pgdat->numabalancing_migrate_nr_pages > ratelimit_pages) {
1589 trace_mm_numa_migrate_ratelimit(current, pgdat->node_id,
1595 * This is an unlocked non-atomic update so errors are possible.
1596 * The consequences are failing to migrate when we potentiall should
1597 * have which is not severe enough to warrant locking. If it is ever
1598 * a problem, it can be converted to a per-cpu counter.
1600 pgdat->numabalancing_migrate_nr_pages += nr_pages;
1604 static int numamigrate_isolate_page(pg_data_t *pgdat, struct page *page)
1608 VM_BUG_ON_PAGE(compound_order(page) && !PageTransHuge(page), page);
1610 /* Avoid migrating to a node that is nearly full */
1611 if (!migrate_balanced_pgdat(pgdat, 1UL << compound_order(page)))
1614 if (isolate_lru_page(page))
1618 * migrate_misplaced_transhuge_page() skips page migration's usual
1619 * check on page_count(), so we must do it here, now that the page
1620 * has been isolated: a GUP pin, or any other pin, prevents migration.
1621 * The expected page count is 3: 1 for page's mapcount and 1 for the
1622 * caller's pin and 1 for the reference taken by isolate_lru_page().
1624 if (PageTransHuge(page) && page_count(page) != 3) {
1625 putback_lru_page(page);
1629 page_lru = page_is_file_cache(page);
1630 mod_zone_page_state(page_zone(page), NR_ISOLATED_ANON + page_lru,
1631 hpage_nr_pages(page));
1634 * Isolating the page has taken another reference, so the
1635 * caller's reference can be safely dropped without the page
1636 * disappearing underneath us during migration.
1642 bool pmd_trans_migrating(pmd_t pmd)
1644 struct page *page = pmd_page(pmd);
1645 return PageLocked(page);
1649 * Attempt to migrate a misplaced page to the specified destination
1650 * node. Caller is expected to have an elevated reference count on
1651 * the page that will be dropped by this function before returning.
1653 int migrate_misplaced_page(struct page *page, struct vm_area_struct *vma,
1656 pg_data_t *pgdat = NODE_DATA(node);
1659 LIST_HEAD(migratepages);
1662 * Don't migrate file pages that are mapped in multiple processes
1663 * with execute permissions as they are probably shared libraries.
1665 if (page_mapcount(page) != 1 && page_is_file_cache(page) &&
1666 (vma->vm_flags & VM_EXEC))
1670 * Rate-limit the amount of data that is being migrated to a node.
1671 * Optimal placement is no good if the memory bus is saturated and
1672 * all the time is being spent migrating!
1674 if (numamigrate_update_ratelimit(pgdat, 1))
1677 isolated = numamigrate_isolate_page(pgdat, page);
1681 list_add(&page->lru, &migratepages);
1682 nr_remaining = migrate_pages(&migratepages, alloc_misplaced_dst_page,
1683 NULL, node, MIGRATE_ASYNC,
1686 if (!list_empty(&migratepages)) {
1687 list_del(&page->lru);
1688 dec_zone_page_state(page, NR_ISOLATED_ANON +
1689 page_is_file_cache(page));
1690 putback_lru_page(page);
1694 count_vm_numa_event(NUMA_PAGE_MIGRATE);
1695 BUG_ON(!list_empty(&migratepages));
1702 #endif /* CONFIG_NUMA_BALANCING */
1704 #if defined(CONFIG_NUMA_BALANCING) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
1706 * Migrates a THP to a given target node. page must be locked and is unlocked
1709 int migrate_misplaced_transhuge_page(struct mm_struct *mm,
1710 struct vm_area_struct *vma,
1711 pmd_t *pmd, pmd_t entry,
1712 unsigned long address,
1713 struct page *page, int node)
1716 pg_data_t *pgdat = NODE_DATA(node);
1718 struct page *new_page = NULL;
1719 int page_lru = page_is_file_cache(page);
1720 unsigned long mmun_start = address & HPAGE_PMD_MASK;
1721 unsigned long mmun_end = mmun_start + HPAGE_PMD_SIZE;
1725 * Rate-limit the amount of data that is being migrated to a node.
1726 * Optimal placement is no good if the memory bus is saturated and
1727 * all the time is being spent migrating!
1729 if (numamigrate_update_ratelimit(pgdat, HPAGE_PMD_NR))
1732 new_page = alloc_pages_node(node,
1733 (GFP_TRANSHUGE | __GFP_THISNODE) & ~__GFP_WAIT,
1738 isolated = numamigrate_isolate_page(pgdat, page);
1744 if (mm_tlb_flush_pending(mm))
1745 flush_tlb_range(vma, mmun_start, mmun_end);
1747 /* Prepare a page as a migration target */
1748 __set_page_locked(new_page);
1749 SetPageSwapBacked(new_page);
1751 /* anon mapping, we can simply copy page->mapping to the new page: */
1752 new_page->mapping = page->mapping;
1753 new_page->index = page->index;
1754 migrate_page_copy(new_page, page);
1755 WARN_ON(PageLRU(new_page));
1757 /* Recheck the target PMD */
1758 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1759 ptl = pmd_lock(mm, pmd);
1760 if (unlikely(!pmd_same(*pmd, entry) || page_count(page) != 2)) {
1763 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1765 /* Reverse changes made by migrate_page_copy() */
1766 if (TestClearPageActive(new_page))
1767 SetPageActive(page);
1768 if (TestClearPageUnevictable(new_page))
1769 SetPageUnevictable(page);
1770 mlock_migrate_page(page, new_page);
1772 unlock_page(new_page);
1773 put_page(new_page); /* Free it */
1775 /* Retake the callers reference and putback on LRU */
1777 putback_lru_page(page);
1778 mod_zone_page_state(page_zone(page),
1779 NR_ISOLATED_ANON + page_lru, -HPAGE_PMD_NR);
1785 entry = mk_pmd(new_page, vma->vm_page_prot);
1786 entry = pmd_mkhuge(entry);
1787 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1790 * Clear the old entry under pagetable lock and establish the new PTE.
1791 * Any parallel GUP will either observe the old page blocking on the
1792 * page lock, block on the page table lock or observe the new page.
1793 * The SetPageUptodate on the new page and page_add_new_anon_rmap
1794 * guarantee the copy is visible before the pagetable update.
1796 flush_cache_range(vma, mmun_start, mmun_end);
1797 page_add_anon_rmap(new_page, vma, mmun_start);
1798 pmdp_clear_flush_notify(vma, mmun_start, pmd);
1799 set_pmd_at(mm, mmun_start, pmd, entry);
1800 flush_tlb_range(vma, mmun_start, mmun_end);
1801 update_mmu_cache_pmd(vma, address, &entry);
1803 if (page_count(page) != 2) {
1804 set_pmd_at(mm, mmun_start, pmd, orig_entry);
1805 flush_tlb_range(vma, mmun_start, mmun_end);
1806 mmu_notifier_invalidate_range(mm, mmun_start, mmun_end);
1807 update_mmu_cache_pmd(vma, address, &entry);
1808 page_remove_rmap(new_page);
1812 mem_cgroup_migrate(page, new_page, false);
1814 page_remove_rmap(page);
1817 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1819 /* Take an "isolate" reference and put new page on the LRU. */
1821 putback_lru_page(new_page);
1823 unlock_page(new_page);
1825 put_page(page); /* Drop the rmap reference */
1826 put_page(page); /* Drop the LRU isolation reference */
1828 count_vm_events(PGMIGRATE_SUCCESS, HPAGE_PMD_NR);
1829 count_vm_numa_events(NUMA_PAGE_MIGRATE, HPAGE_PMD_NR);
1831 mod_zone_page_state(page_zone(page),
1832 NR_ISOLATED_ANON + page_lru,
1837 count_vm_events(PGMIGRATE_FAIL, HPAGE_PMD_NR);
1839 ptl = pmd_lock(mm, pmd);
1840 if (pmd_same(*pmd, entry)) {
1841 entry = pmd_modify(entry, vma->vm_page_prot);
1842 set_pmd_at(mm, mmun_start, pmd, entry);
1843 update_mmu_cache_pmd(vma, address, &entry);
1852 #endif /* CONFIG_NUMA_BALANCING */
1854 #endif /* CONFIG_NUMA */