1 // SPDX-License-Identifier: GPL-2.0
3 * Memory Migration functionality - linux/mm/migrate.c
5 * Copyright (C) 2006 Silicon Graphics, Inc., Christoph Lameter
7 * Page migration was first developed in the context of the memory hotplug
8 * project. The main authors of the migration code are:
10 * IWAMOTO Toshihiro <iwamoto@valinux.co.jp>
11 * Hirokazu Takahashi <taka@valinux.co.jp>
12 * Dave Hansen <haveblue@us.ibm.com>
16 #include <linux/migrate.h>
17 #include <linux/export.h>
18 #include <linux/swap.h>
19 #include <linux/swapops.h>
20 #include <linux/pagemap.h>
21 #include <linux/buffer_head.h>
22 #include <linux/mm_inline.h>
23 #include <linux/nsproxy.h>
24 #include <linux/pagevec.h>
25 #include <linux/ksm.h>
26 #include <linux/rmap.h>
27 #include <linux/topology.h>
28 #include <linux/cpu.h>
29 #include <linux/cpuset.h>
30 #include <linux/writeback.h>
31 #include <linux/mempolicy.h>
32 #include <linux/vmalloc.h>
33 #include <linux/security.h>
34 #include <linux/backing-dev.h>
35 #include <linux/compaction.h>
36 #include <linux/syscalls.h>
37 #include <linux/compat.h>
38 #include <linux/hugetlb.h>
39 #include <linux/hugetlb_cgroup.h>
40 #include <linux/gfp.h>
41 #include <linux/pfn_t.h>
42 #include <linux/memremap.h>
43 #include <linux/userfaultfd_k.h>
44 #include <linux/balloon_compaction.h>
45 #include <linux/mmu_notifier.h>
46 #include <linux/page_idle.h>
47 #include <linux/page_owner.h>
48 #include <linux/sched/mm.h>
49 #include <linux/ptrace.h>
51 #include <asm/tlbflush.h>
53 #define CREATE_TRACE_POINTS
54 #include <trace/events/migrate.h>
59 * migrate_prep() needs to be called before we start compiling a list of pages
60 * to be migrated using isolate_lru_page(). If scheduling work on other CPUs is
61 * undesirable, use migrate_prep_local()
63 int migrate_prep(void)
66 * Clear the LRU lists so pages can be isolated.
67 * Note that pages may be moved off the LRU after we have
68 * drained them. Those pages will fail to migrate like other
69 * pages that may be busy.
76 /* Do the necessary work of migrate_prep but not if it involves other CPUs */
77 int migrate_prep_local(void)
84 int isolate_movable_page(struct page *page, isolate_mode_t mode)
86 struct address_space *mapping;
89 * Avoid burning cycles with pages that are yet under __free_pages(),
90 * or just got freed under us.
92 * In case we 'win' a race for a movable page being freed under us and
93 * raise its refcount preventing __free_pages() from doing its job
94 * the put_page() at the end of this block will take care of
95 * release this page, thus avoiding a nasty leakage.
97 if (unlikely(!get_page_unless_zero(page)))
101 * Check PageMovable before holding a PG_lock because page's owner
102 * assumes anybody doesn't touch PG_lock of newly allocated page
103 * so unconditionally grapping the lock ruins page's owner side.
105 if (unlikely(!__PageMovable(page)))
108 * As movable pages are not isolated from LRU lists, concurrent
109 * compaction threads can race against page migration functions
110 * as well as race against the releasing a page.
112 * In order to avoid having an already isolated movable page
113 * being (wrongly) re-isolated while it is under migration,
114 * or to avoid attempting to isolate pages being released,
115 * lets be sure we have the page lock
116 * before proceeding with the movable page isolation steps.
118 if (unlikely(!trylock_page(page)))
121 if (!PageMovable(page) || PageIsolated(page))
122 goto out_no_isolated;
124 mapping = page_mapping(page);
125 VM_BUG_ON_PAGE(!mapping, page);
127 if (!mapping->a_ops->isolate_page(page, mode))
128 goto out_no_isolated;
130 /* Driver shouldn't use PG_isolated bit of page->flags */
131 WARN_ON_ONCE(PageIsolated(page));
132 __SetPageIsolated(page);
145 /* It should be called on page which is PG_movable */
146 void putback_movable_page(struct page *page)
148 struct address_space *mapping;
150 VM_BUG_ON_PAGE(!PageLocked(page), page);
151 VM_BUG_ON_PAGE(!PageMovable(page), page);
152 VM_BUG_ON_PAGE(!PageIsolated(page), page);
154 mapping = page_mapping(page);
155 mapping->a_ops->putback_page(page);
156 __ClearPageIsolated(page);
160 * Put previously isolated pages back onto the appropriate lists
161 * from where they were once taken off for compaction/migration.
163 * This function shall be used whenever the isolated pageset has been
164 * built from lru, balloon, hugetlbfs page. See isolate_migratepages_range()
165 * and isolate_huge_page().
167 void putback_movable_pages(struct list_head *l)
172 list_for_each_entry_safe(page, page2, l, lru) {
173 if (unlikely(PageHuge(page))) {
174 putback_active_hugepage(page);
177 list_del(&page->lru);
179 * We isolated non-lru movable page so here we can use
180 * __PageMovable because LRU page's mapping cannot have
181 * PAGE_MAPPING_MOVABLE.
183 if (unlikely(__PageMovable(page))) {
184 VM_BUG_ON_PAGE(!PageIsolated(page), page);
186 if (PageMovable(page))
187 putback_movable_page(page);
189 __ClearPageIsolated(page);
193 mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON +
194 page_is_file_cache(page), -hpage_nr_pages(page));
195 putback_lru_page(page);
201 * Restore a potential migration pte to a working pte entry
203 static bool remove_migration_pte(struct page *page, struct vm_area_struct *vma,
204 unsigned long addr, void *old)
206 struct page_vma_mapped_walk pvmw = {
210 .flags = PVMW_SYNC | PVMW_MIGRATION,
216 VM_BUG_ON_PAGE(PageTail(page), page);
217 while (page_vma_mapped_walk(&pvmw)) {
221 new = page - pvmw.page->index +
222 linear_page_index(vma, pvmw.address);
224 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
225 /* PMD-mapped THP migration entry */
227 VM_BUG_ON_PAGE(PageHuge(page) || !PageTransCompound(page), page);
228 remove_migration_pmd(&pvmw, new);
234 pte = pte_mkold(mk_pte(new, READ_ONCE(vma->vm_page_prot)));
235 if (pte_swp_soft_dirty(*pvmw.pte))
236 pte = pte_mksoft_dirty(pte);
239 * Recheck VMA as permissions can change since migration started
241 entry = pte_to_swp_entry(*pvmw.pte);
242 if (is_write_migration_entry(entry))
243 pte = maybe_mkwrite(pte, vma);
245 if (unlikely(is_zone_device_page(new))) {
246 if (is_device_private_page(new)) {
247 entry = make_device_private_entry(new, pte_write(pte));
248 pte = swp_entry_to_pte(entry);
249 } else if (is_device_public_page(new)) {
250 pte = pte_mkdevmap(pte);
254 #ifdef CONFIG_HUGETLB_PAGE
256 pte = pte_mkhuge(pte);
257 pte = arch_make_huge_pte(pte, vma, new, 0);
258 set_huge_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte);
260 hugepage_add_anon_rmap(new, vma, pvmw.address);
262 page_dup_rmap(new, true);
266 set_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte);
269 page_add_anon_rmap(new, vma, pvmw.address, false);
271 page_add_file_rmap(new, false);
273 if (vma->vm_flags & VM_LOCKED && !PageTransCompound(new))
276 if (PageTransHuge(page) && PageMlocked(page))
277 clear_page_mlock(page);
279 /* No need to invalidate - it was non-present before */
280 update_mmu_cache(vma, pvmw.address, pvmw.pte);
287 * Get rid of all migration entries and replace them by
288 * references to the indicated page.
290 void remove_migration_ptes(struct page *old, struct page *new, bool locked)
292 struct rmap_walk_control rwc = {
293 .rmap_one = remove_migration_pte,
298 rmap_walk_locked(new, &rwc);
300 rmap_walk(new, &rwc);
304 * Something used the pte of a page under migration. We need to
305 * get to the page and wait until migration is finished.
306 * When we return from this function the fault will be retried.
308 void __migration_entry_wait(struct mm_struct *mm, pte_t *ptep,
317 if (!is_swap_pte(pte))
320 entry = pte_to_swp_entry(pte);
321 if (!is_migration_entry(entry))
324 page = migration_entry_to_page(entry);
327 * Once radix-tree replacement of page migration started, page_count
328 * *must* be zero. And, we don't want to call wait_on_page_locked()
329 * against a page without get_page().
330 * So, we use get_page_unless_zero(), here. Even failed, page fault
333 if (!get_page_unless_zero(page))
335 pte_unmap_unlock(ptep, ptl);
336 wait_on_page_locked(page);
340 pte_unmap_unlock(ptep, ptl);
343 void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
344 unsigned long address)
346 spinlock_t *ptl = pte_lockptr(mm, pmd);
347 pte_t *ptep = pte_offset_map(pmd, address);
348 __migration_entry_wait(mm, ptep, ptl);
351 void migration_entry_wait_huge(struct vm_area_struct *vma,
352 struct mm_struct *mm, pte_t *pte)
354 spinlock_t *ptl = huge_pte_lockptr(hstate_vma(vma), mm, pte);
355 __migration_entry_wait(mm, pte, ptl);
358 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
359 void pmd_migration_entry_wait(struct mm_struct *mm, pmd_t *pmd)
364 ptl = pmd_lock(mm, pmd);
365 if (!is_pmd_migration_entry(*pmd))
367 page = migration_entry_to_page(pmd_to_swp_entry(*pmd));
368 if (!get_page_unless_zero(page))
371 wait_on_page_locked(page);
380 /* Returns true if all buffers are successfully locked */
381 static bool buffer_migrate_lock_buffers(struct buffer_head *head,
382 enum migrate_mode mode)
384 struct buffer_head *bh = head;
386 /* Simple case, sync compaction */
387 if (mode != MIGRATE_ASYNC) {
391 bh = bh->b_this_page;
393 } while (bh != head);
398 /* async case, we cannot block on lock_buffer so use trylock_buffer */
401 if (!trylock_buffer(bh)) {
403 * We failed to lock the buffer and cannot stall in
404 * async migration. Release the taken locks
406 struct buffer_head *failed_bh = bh;
409 while (bh != failed_bh) {
412 bh = bh->b_this_page;
417 bh = bh->b_this_page;
418 } while (bh != head);
422 static inline bool buffer_migrate_lock_buffers(struct buffer_head *head,
423 enum migrate_mode mode)
427 #endif /* CONFIG_BLOCK */
430 * Replace the page in the mapping.
432 * The number of remaining references must be:
433 * 1 for anonymous pages without a mapping
434 * 2 for pages with a mapping
435 * 3 for pages with a mapping and PagePrivate/PagePrivate2 set.
437 int migrate_page_move_mapping(struct address_space *mapping,
438 struct page *newpage, struct page *page,
439 struct buffer_head *head, enum migrate_mode mode,
442 struct zone *oldzone, *newzone;
444 int expected_count = 1 + extra_count;
448 * Device public or private pages have an extra refcount as they are
451 expected_count += is_device_private_page(page);
452 expected_count += is_device_public_page(page);
455 /* Anonymous page without mapping */
456 if (page_count(page) != expected_count)
459 /* No turning back from here */
460 newpage->index = page->index;
461 newpage->mapping = page->mapping;
462 if (PageSwapBacked(page))
463 __SetPageSwapBacked(newpage);
465 return MIGRATEPAGE_SUCCESS;
468 oldzone = page_zone(page);
469 newzone = page_zone(newpage);
471 xa_lock_irq(&mapping->i_pages);
473 pslot = radix_tree_lookup_slot(&mapping->i_pages,
476 xa_unlock_irq(&mapping->i_pages);
480 expected_count += hpage_nr_pages(page) + page_has_private(page);
481 if (page_count(page) != expected_count ||
482 radix_tree_deref_slot_protected(pslot,
483 &mapping->i_pages.xa_lock) != page) {
484 xa_unlock_irq(&mapping->i_pages);
488 if (!page_ref_freeze(page, expected_count)) {
489 xa_unlock_irq(&mapping->i_pages);
494 * In the async migration case of moving a page with buffers, lock the
495 * buffers using trylock before the mapping is moved. If the mapping
496 * was moved, we later failed to lock the buffers and could not move
497 * the mapping back due to an elevated page count, we would have to
498 * block waiting on other references to be dropped.
500 if (mode == MIGRATE_ASYNC && head &&
501 !buffer_migrate_lock_buffers(head, mode)) {
502 page_ref_unfreeze(page, expected_count);
503 xa_unlock_irq(&mapping->i_pages);
508 * Now we know that no one else is looking at the page:
509 * no turning back from here.
511 newpage->index = page->index;
512 newpage->mapping = page->mapping;
513 page_ref_add(newpage, hpage_nr_pages(page)); /* add cache reference */
514 if (PageSwapBacked(page)) {
515 __SetPageSwapBacked(newpage);
516 if (PageSwapCache(page)) {
517 SetPageSwapCache(newpage);
518 set_page_private(newpage, page_private(page));
521 VM_BUG_ON_PAGE(PageSwapCache(page), page);
524 /* Move dirty while page refs frozen and newpage not yet exposed */
525 dirty = PageDirty(page);
527 ClearPageDirty(page);
528 SetPageDirty(newpage);
531 radix_tree_replace_slot(&mapping->i_pages, pslot, newpage);
532 if (PageTransHuge(page)) {
534 int index = page_index(page);
536 for (i = 1; i < HPAGE_PMD_NR; i++) {
537 pslot = radix_tree_lookup_slot(&mapping->i_pages,
539 radix_tree_replace_slot(&mapping->i_pages, pslot,
545 * Drop cache reference from old page by unfreezing
546 * to one less reference.
547 * We know this isn't the last reference.
549 page_ref_unfreeze(page, expected_count - hpage_nr_pages(page));
551 xa_unlock(&mapping->i_pages);
552 /* Leave irq disabled to prevent preemption while updating stats */
555 * If moved to a different zone then also account
556 * the page for that zone. Other VM counters will be
557 * taken care of when we establish references to the
558 * new page and drop references to the old page.
560 * Note that anonymous pages are accounted for
561 * via NR_FILE_PAGES and NR_ANON_MAPPED if they
562 * are mapped to swap space.
564 if (newzone != oldzone) {
565 __dec_node_state(oldzone->zone_pgdat, NR_FILE_PAGES);
566 __inc_node_state(newzone->zone_pgdat, NR_FILE_PAGES);
567 if (PageSwapBacked(page) && !PageSwapCache(page)) {
568 __dec_node_state(oldzone->zone_pgdat, NR_SHMEM);
569 __inc_node_state(newzone->zone_pgdat, NR_SHMEM);
571 if (dirty && mapping_cap_account_dirty(mapping)) {
572 __dec_node_state(oldzone->zone_pgdat, NR_FILE_DIRTY);
573 __dec_zone_state(oldzone, NR_ZONE_WRITE_PENDING);
574 __inc_node_state(newzone->zone_pgdat, NR_FILE_DIRTY);
575 __inc_zone_state(newzone, NR_ZONE_WRITE_PENDING);
580 return MIGRATEPAGE_SUCCESS;
582 EXPORT_SYMBOL(migrate_page_move_mapping);
585 * The expected number of remaining references is the same as that
586 * of migrate_page_move_mapping().
588 int migrate_huge_page_move_mapping(struct address_space *mapping,
589 struct page *newpage, struct page *page)
594 xa_lock_irq(&mapping->i_pages);
596 pslot = radix_tree_lookup_slot(&mapping->i_pages, page_index(page));
598 xa_unlock_irq(&mapping->i_pages);
602 expected_count = 2 + page_has_private(page);
603 if (page_count(page) != expected_count ||
604 radix_tree_deref_slot_protected(pslot, &mapping->i_pages.xa_lock) != page) {
605 xa_unlock_irq(&mapping->i_pages);
609 if (!page_ref_freeze(page, expected_count)) {
610 xa_unlock_irq(&mapping->i_pages);
614 newpage->index = page->index;
615 newpage->mapping = page->mapping;
619 radix_tree_replace_slot(&mapping->i_pages, pslot, newpage);
621 page_ref_unfreeze(page, expected_count - 1);
623 xa_unlock_irq(&mapping->i_pages);
625 return MIGRATEPAGE_SUCCESS;
629 * Gigantic pages are so large that we do not guarantee that page++ pointer
630 * arithmetic will work across the entire page. We need something more
633 static void __copy_gigantic_page(struct page *dst, struct page *src,
637 struct page *dst_base = dst;
638 struct page *src_base = src;
640 for (i = 0; i < nr_pages; ) {
642 copy_highpage(dst, src);
645 dst = mem_map_next(dst, dst_base, i);
646 src = mem_map_next(src, src_base, i);
650 static void copy_huge_page(struct page *dst, struct page *src)
657 struct hstate *h = page_hstate(src);
658 nr_pages = pages_per_huge_page(h);
660 if (unlikely(nr_pages > MAX_ORDER_NR_PAGES)) {
661 __copy_gigantic_page(dst, src, nr_pages);
666 BUG_ON(!PageTransHuge(src));
667 nr_pages = hpage_nr_pages(src);
670 for (i = 0; i < nr_pages; i++) {
672 copy_highpage(dst + i, src + i);
677 * Copy the page to its new location
679 void migrate_page_states(struct page *newpage, struct page *page)
684 SetPageError(newpage);
685 if (PageReferenced(page))
686 SetPageReferenced(newpage);
687 if (PageUptodate(page))
688 SetPageUptodate(newpage);
689 if (TestClearPageActive(page)) {
690 VM_BUG_ON_PAGE(PageUnevictable(page), page);
691 SetPageActive(newpage);
692 } else if (TestClearPageUnevictable(page))
693 SetPageUnevictable(newpage);
694 if (PageChecked(page))
695 SetPageChecked(newpage);
696 if (PageMappedToDisk(page))
697 SetPageMappedToDisk(newpage);
699 /* Move dirty on pages not done by migrate_page_move_mapping() */
701 SetPageDirty(newpage);
703 if (page_is_young(page))
704 set_page_young(newpage);
705 if (page_is_idle(page))
706 set_page_idle(newpage);
709 * Copy NUMA information to the new page, to prevent over-eager
710 * future migrations of this same page.
712 cpupid = page_cpupid_xchg_last(page, -1);
713 page_cpupid_xchg_last(newpage, cpupid);
715 ksm_migrate_page(newpage, page);
717 * Please do not reorder this without considering how mm/ksm.c's
718 * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache().
720 if (PageSwapCache(page))
721 ClearPageSwapCache(page);
722 ClearPagePrivate(page);
723 set_page_private(page, 0);
726 * If any waiters have accumulated on the new page then
729 if (PageWriteback(newpage))
730 end_page_writeback(newpage);
732 copy_page_owner(page, newpage);
734 mem_cgroup_migrate(page, newpage);
736 EXPORT_SYMBOL(migrate_page_states);
738 void migrate_page_copy(struct page *newpage, struct page *page)
740 if (PageHuge(page) || PageTransHuge(page))
741 copy_huge_page(newpage, page);
743 copy_highpage(newpage, page);
745 migrate_page_states(newpage, page);
747 EXPORT_SYMBOL(migrate_page_copy);
749 /************************************************************
750 * Migration functions
751 ***********************************************************/
754 * Common logic to directly migrate a single LRU page suitable for
755 * pages that do not use PagePrivate/PagePrivate2.
757 * Pages are locked upon entry and exit.
759 int migrate_page(struct address_space *mapping,
760 struct page *newpage, struct page *page,
761 enum migrate_mode mode)
765 BUG_ON(PageWriteback(page)); /* Writeback must be complete */
767 rc = migrate_page_move_mapping(mapping, newpage, page, NULL, mode, 0);
769 if (rc != MIGRATEPAGE_SUCCESS)
772 if (mode != MIGRATE_SYNC_NO_COPY)
773 migrate_page_copy(newpage, page);
775 migrate_page_states(newpage, page);
776 return MIGRATEPAGE_SUCCESS;
778 EXPORT_SYMBOL(migrate_page);
782 * Migration function for pages with buffers. This function can only be used
783 * if the underlying filesystem guarantees that no other references to "page"
786 int buffer_migrate_page(struct address_space *mapping,
787 struct page *newpage, struct page *page, enum migrate_mode mode)
789 struct buffer_head *bh, *head;
792 if (!page_has_buffers(page))
793 return migrate_page(mapping, newpage, page, mode);
795 head = page_buffers(page);
797 rc = migrate_page_move_mapping(mapping, newpage, page, head, mode, 0);
799 if (rc != MIGRATEPAGE_SUCCESS)
803 * In the async case, migrate_page_move_mapping locked the buffers
804 * with an IRQ-safe spinlock held. In the sync case, the buffers
805 * need to be locked now
807 if (mode != MIGRATE_ASYNC)
808 BUG_ON(!buffer_migrate_lock_buffers(head, mode));
810 ClearPagePrivate(page);
811 set_page_private(newpage, page_private(page));
812 set_page_private(page, 0);
818 set_bh_page(bh, newpage, bh_offset(bh));
819 bh = bh->b_this_page;
821 } while (bh != head);
823 SetPagePrivate(newpage);
825 if (mode != MIGRATE_SYNC_NO_COPY)
826 migrate_page_copy(newpage, page);
828 migrate_page_states(newpage, page);
834 bh = bh->b_this_page;
836 } while (bh != head);
838 return MIGRATEPAGE_SUCCESS;
840 EXPORT_SYMBOL(buffer_migrate_page);
844 * Writeback a page to clean the dirty state
846 static int writeout(struct address_space *mapping, struct page *page)
848 struct writeback_control wbc = {
849 .sync_mode = WB_SYNC_NONE,
852 .range_end = LLONG_MAX,
857 if (!mapping->a_ops->writepage)
858 /* No write method for the address space */
861 if (!clear_page_dirty_for_io(page))
862 /* Someone else already triggered a write */
866 * A dirty page may imply that the underlying filesystem has
867 * the page on some queue. So the page must be clean for
868 * migration. Writeout may mean we loose the lock and the
869 * page state is no longer what we checked for earlier.
870 * At this point we know that the migration attempt cannot
873 remove_migration_ptes(page, page, false);
875 rc = mapping->a_ops->writepage(page, &wbc);
877 if (rc != AOP_WRITEPAGE_ACTIVATE)
878 /* unlocked. Relock */
881 return (rc < 0) ? -EIO : -EAGAIN;
885 * Default handling if a filesystem does not provide a migration function.
887 static int fallback_migrate_page(struct address_space *mapping,
888 struct page *newpage, struct page *page, enum migrate_mode mode)
890 if (PageDirty(page)) {
891 /* Only writeback pages in full synchronous migration */
894 case MIGRATE_SYNC_NO_COPY:
899 return writeout(mapping, page);
903 * Buffers may be managed in a filesystem specific way.
904 * We must have no buffers or drop them.
906 if (page_has_private(page) &&
907 !try_to_release_page(page, GFP_KERNEL))
910 return migrate_page(mapping, newpage, page, mode);
914 * Move a page to a newly allocated page
915 * The page is locked and all ptes have been successfully removed.
917 * The new page will have replaced the old page if this function
922 * MIGRATEPAGE_SUCCESS - success
924 static int move_to_new_page(struct page *newpage, struct page *page,
925 enum migrate_mode mode)
927 struct address_space *mapping;
929 bool is_lru = !__PageMovable(page);
931 VM_BUG_ON_PAGE(!PageLocked(page), page);
932 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
934 mapping = page_mapping(page);
936 if (likely(is_lru)) {
938 rc = migrate_page(mapping, newpage, page, mode);
939 else if (mapping->a_ops->migratepage)
941 * Most pages have a mapping and most filesystems
942 * provide a migratepage callback. Anonymous pages
943 * are part of swap space which also has its own
944 * migratepage callback. This is the most common path
945 * for page migration.
947 rc = mapping->a_ops->migratepage(mapping, newpage,
950 rc = fallback_migrate_page(mapping, newpage,
954 * In case of non-lru page, it could be released after
955 * isolation step. In that case, we shouldn't try migration.
957 VM_BUG_ON_PAGE(!PageIsolated(page), page);
958 if (!PageMovable(page)) {
959 rc = MIGRATEPAGE_SUCCESS;
960 __ClearPageIsolated(page);
964 rc = mapping->a_ops->migratepage(mapping, newpage,
966 WARN_ON_ONCE(rc == MIGRATEPAGE_SUCCESS &&
967 !PageIsolated(page));
971 * When successful, old pagecache page->mapping must be cleared before
972 * page is freed; but stats require that PageAnon be left as PageAnon.
974 if (rc == MIGRATEPAGE_SUCCESS) {
975 if (__PageMovable(page)) {
976 VM_BUG_ON_PAGE(!PageIsolated(page), page);
979 * We clear PG_movable under page_lock so any compactor
980 * cannot try to migrate this page.
982 __ClearPageIsolated(page);
986 * Anonymous and movable page->mapping will be cleard by
987 * free_pages_prepare so don't reset it here for keeping
988 * the type to work PageAnon, for example.
990 if (!PageMappingFlags(page))
991 page->mapping = NULL;
993 if (unlikely(is_zone_device_page(newpage))) {
994 if (is_device_public_page(newpage))
995 flush_dcache_page(newpage);
997 flush_dcache_page(newpage);
1004 static int __unmap_and_move(struct page *page, struct page *newpage,
1005 int force, enum migrate_mode mode)
1008 int page_was_mapped = 0;
1009 struct anon_vma *anon_vma = NULL;
1010 bool is_lru = !__PageMovable(page);
1012 if (!trylock_page(page)) {
1013 if (!force || mode == MIGRATE_ASYNC)
1017 * It's not safe for direct compaction to call lock_page.
1018 * For example, during page readahead pages are added locked
1019 * to the LRU. Later, when the IO completes the pages are
1020 * marked uptodate and unlocked. However, the queueing
1021 * could be merging multiple pages for one bio (e.g.
1022 * mpage_readpages). If an allocation happens for the
1023 * second or third page, the process can end up locking
1024 * the same page twice and deadlocking. Rather than
1025 * trying to be clever about what pages can be locked,
1026 * avoid the use of lock_page for direct compaction
1029 if (current->flags & PF_MEMALLOC)
1035 if (PageWriteback(page)) {
1037 * Only in the case of a full synchronous migration is it
1038 * necessary to wait for PageWriteback. In the async case,
1039 * the retry loop is too short and in the sync-light case,
1040 * the overhead of stalling is too much
1044 case MIGRATE_SYNC_NO_COPY:
1052 wait_on_page_writeback(page);
1056 * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
1057 * we cannot notice that anon_vma is freed while we migrates a page.
1058 * This get_anon_vma() delays freeing anon_vma pointer until the end
1059 * of migration. File cache pages are no problem because of page_lock()
1060 * File Caches may use write_page() or lock_page() in migration, then,
1061 * just care Anon page here.
1063 * Only page_get_anon_vma() understands the subtleties of
1064 * getting a hold on an anon_vma from outside one of its mms.
1065 * But if we cannot get anon_vma, then we won't need it anyway,
1066 * because that implies that the anon page is no longer mapped
1067 * (and cannot be remapped so long as we hold the page lock).
1069 if (PageAnon(page) && !PageKsm(page))
1070 anon_vma = page_get_anon_vma(page);
1073 * Block others from accessing the new page when we get around to
1074 * establishing additional references. We are usually the only one
1075 * holding a reference to newpage at this point. We used to have a BUG
1076 * here if trylock_page(newpage) fails, but would like to allow for
1077 * cases where there might be a race with the previous use of newpage.
1078 * This is much like races on refcount of oldpage: just don't BUG().
1080 if (unlikely(!trylock_page(newpage)))
1083 if (unlikely(!is_lru)) {
1084 rc = move_to_new_page(newpage, page, mode);
1085 goto out_unlock_both;
1089 * Corner case handling:
1090 * 1. When a new swap-cache page is read into, it is added to the LRU
1091 * and treated as swapcache but it has no rmap yet.
1092 * Calling try_to_unmap() against a page->mapping==NULL page will
1093 * trigger a BUG. So handle it here.
1094 * 2. An orphaned page (see truncate_complete_page) might have
1095 * fs-private metadata. The page can be picked up due to memory
1096 * offlining. Everywhere else except page reclaim, the page is
1097 * invisible to the vm, so the page can not be migrated. So try to
1098 * free the metadata, so the page can be freed.
1100 if (!page->mapping) {
1101 VM_BUG_ON_PAGE(PageAnon(page), page);
1102 if (page_has_private(page)) {
1103 try_to_free_buffers(page);
1104 goto out_unlock_both;
1106 } else if (page_mapped(page)) {
1107 /* Establish migration ptes */
1108 VM_BUG_ON_PAGE(PageAnon(page) && !PageKsm(page) && !anon_vma,
1111 TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
1112 page_was_mapped = 1;
1115 if (!page_mapped(page))
1116 rc = move_to_new_page(newpage, page, mode);
1118 if (page_was_mapped)
1119 remove_migration_ptes(page,
1120 rc == MIGRATEPAGE_SUCCESS ? newpage : page, false);
1123 unlock_page(newpage);
1125 /* Drop an anon_vma reference if we took one */
1127 put_anon_vma(anon_vma);
1131 * If migration is successful, decrease refcount of the newpage
1132 * which will not free the page because new page owner increased
1133 * refcounter. As well, if it is LRU page, add the page to LRU
1134 * list in here. Use the old state of the isolated source page to
1135 * determine if we migrated a LRU page. newpage was already unlocked
1136 * and possibly modified by its owner - don't rely on the page
1139 if (rc == MIGRATEPAGE_SUCCESS) {
1140 if (unlikely(!is_lru))
1143 putback_lru_page(newpage);
1150 * gcc 4.7 and 4.8 on arm get an ICEs when inlining unmap_and_move(). Work
1153 #if defined(CONFIG_ARM) && \
1154 defined(GCC_VERSION) && GCC_VERSION < 40900 && GCC_VERSION >= 40700
1155 #define ICE_noinline noinline
1157 #define ICE_noinline
1161 * Obtain the lock on page, remove all ptes and migrate the page
1162 * to the newly allocated page in newpage.
1164 static ICE_noinline int unmap_and_move(new_page_t get_new_page,
1165 free_page_t put_new_page,
1166 unsigned long private, struct page *page,
1167 int force, enum migrate_mode mode,
1168 enum migrate_reason reason)
1170 int rc = MIGRATEPAGE_SUCCESS;
1171 struct page *newpage;
1173 if (!thp_migration_supported() && PageTransHuge(page))
1176 newpage = get_new_page(page, private);
1180 if (page_count(page) == 1) {
1181 /* page was freed from under us. So we are done. */
1182 ClearPageActive(page);
1183 ClearPageUnevictable(page);
1184 if (unlikely(__PageMovable(page))) {
1186 if (!PageMovable(page))
1187 __ClearPageIsolated(page);
1191 put_new_page(newpage, private);
1197 rc = __unmap_and_move(page, newpage, force, mode);
1198 if (rc == MIGRATEPAGE_SUCCESS)
1199 set_page_owner_migrate_reason(newpage, reason);
1202 if (rc != -EAGAIN) {
1204 * A page that has been migrated has all references
1205 * removed and will be freed. A page that has not been
1206 * migrated will have kepts its references and be
1209 list_del(&page->lru);
1212 * Compaction can migrate also non-LRU pages which are
1213 * not accounted to NR_ISOLATED_*. They can be recognized
1216 if (likely(!__PageMovable(page)))
1217 mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON +
1218 page_is_file_cache(page), -hpage_nr_pages(page));
1222 * If migration is successful, releases reference grabbed during
1223 * isolation. Otherwise, restore the page to right list unless
1226 if (rc == MIGRATEPAGE_SUCCESS) {
1228 if (reason == MR_MEMORY_FAILURE) {
1230 * Set PG_HWPoison on just freed page
1231 * intentionally. Although it's rather weird,
1232 * it's how HWPoison flag works at the moment.
1234 if (set_hwpoison_free_buddy_page(page))
1235 num_poisoned_pages_inc();
1238 if (rc != -EAGAIN) {
1239 if (likely(!__PageMovable(page))) {
1240 putback_lru_page(page);
1245 if (PageMovable(page))
1246 putback_movable_page(page);
1248 __ClearPageIsolated(page);
1254 put_new_page(newpage, private);
1263 * Counterpart of unmap_and_move_page() for hugepage migration.
1265 * This function doesn't wait the completion of hugepage I/O
1266 * because there is no race between I/O and migration for hugepage.
1267 * Note that currently hugepage I/O occurs only in direct I/O
1268 * where no lock is held and PG_writeback is irrelevant,
1269 * and writeback status of all subpages are counted in the reference
1270 * count of the head page (i.e. if all subpages of a 2MB hugepage are
1271 * under direct I/O, the reference of the head page is 512 and a bit more.)
1272 * This means that when we try to migrate hugepage whose subpages are
1273 * doing direct I/O, some references remain after try_to_unmap() and
1274 * hugepage migration fails without data corruption.
1276 * There is also no race when direct I/O is issued on the page under migration,
1277 * because then pte is replaced with migration swap entry and direct I/O code
1278 * will wait in the page fault for migration to complete.
1280 static int unmap_and_move_huge_page(new_page_t get_new_page,
1281 free_page_t put_new_page, unsigned long private,
1282 struct page *hpage, int force,
1283 enum migrate_mode mode, int reason)
1286 int page_was_mapped = 0;
1287 struct page *new_hpage;
1288 struct anon_vma *anon_vma = NULL;
1291 * Movability of hugepages depends on architectures and hugepage size.
1292 * This check is necessary because some callers of hugepage migration
1293 * like soft offline and memory hotremove don't walk through page
1294 * tables or check whether the hugepage is pmd-based or not before
1295 * kicking migration.
1297 if (!hugepage_migration_supported(page_hstate(hpage))) {
1298 putback_active_hugepage(hpage);
1302 new_hpage = get_new_page(hpage, private);
1306 if (!trylock_page(hpage)) {
1311 case MIGRATE_SYNC_NO_COPY:
1320 * Check for pages which are in the process of being freed. Without
1321 * page_mapping() set, hugetlbfs specific move page routine will not
1322 * be called and we could leak usage counts for subpools.
1324 if (page_private(hpage) && !page_mapping(hpage)) {
1329 if (PageAnon(hpage))
1330 anon_vma = page_get_anon_vma(hpage);
1332 if (unlikely(!trylock_page(new_hpage)))
1335 if (page_mapped(hpage)) {
1337 TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
1338 page_was_mapped = 1;
1341 if (!page_mapped(hpage))
1342 rc = move_to_new_page(new_hpage, hpage, mode);
1344 if (page_was_mapped)
1345 remove_migration_ptes(hpage,
1346 rc == MIGRATEPAGE_SUCCESS ? new_hpage : hpage, false);
1348 unlock_page(new_hpage);
1352 put_anon_vma(anon_vma);
1354 if (rc == MIGRATEPAGE_SUCCESS) {
1355 move_hugetlb_state(hpage, new_hpage, reason);
1356 put_new_page = NULL;
1363 putback_active_hugepage(hpage);
1366 * If migration was not successful and there's a freeing callback, use
1367 * it. Otherwise, put_page() will drop the reference grabbed during
1371 put_new_page(new_hpage, private);
1373 putback_active_hugepage(new_hpage);
1379 * migrate_pages - migrate the pages specified in a list, to the free pages
1380 * supplied as the target for the page migration
1382 * @from: The list of pages to be migrated.
1383 * @get_new_page: The function used to allocate free pages to be used
1384 * as the target of the page migration.
1385 * @put_new_page: The function used to free target pages if migration
1386 * fails, or NULL if no special handling is necessary.
1387 * @private: Private data to be passed on to get_new_page()
1388 * @mode: The migration mode that specifies the constraints for
1389 * page migration, if any.
1390 * @reason: The reason for page migration.
1392 * The function returns after 10 attempts or if no pages are movable any more
1393 * because the list has become empty or no retryable pages exist any more.
1394 * The caller should call putback_movable_pages() to return pages to the LRU
1395 * or free list only if ret != 0.
1397 * Returns the number of pages that were not migrated, or an error code.
1399 int migrate_pages(struct list_head *from, new_page_t get_new_page,
1400 free_page_t put_new_page, unsigned long private,
1401 enum migrate_mode mode, int reason)
1405 int nr_succeeded = 0;
1409 int swapwrite = current->flags & PF_SWAPWRITE;
1413 current->flags |= PF_SWAPWRITE;
1415 for(pass = 0; pass < 10 && retry; pass++) {
1418 list_for_each_entry_safe(page, page2, from, lru) {
1423 rc = unmap_and_move_huge_page(get_new_page,
1424 put_new_page, private, page,
1425 pass > 2, mode, reason);
1427 rc = unmap_and_move(get_new_page, put_new_page,
1428 private, page, pass > 2, mode,
1434 * THP migration might be unsupported or the
1435 * allocation could've failed so we should
1436 * retry on the same page with the THP split
1439 * Head page is retried immediately and tail
1440 * pages are added to the tail of the list so
1441 * we encounter them after the rest of the list
1444 if (PageTransHuge(page) && !PageHuge(page)) {
1446 rc = split_huge_page_to_list(page, from);
1449 list_safe_reset_next(page, page2, lru);
1458 case MIGRATEPAGE_SUCCESS:
1463 * Permanent failure (-EBUSY, -ENOSYS, etc.):
1464 * unlike -EAGAIN case, the failed page is
1465 * removed from migration page list and not
1466 * retried in the next outer loop.
1477 count_vm_events(PGMIGRATE_SUCCESS, nr_succeeded);
1479 count_vm_events(PGMIGRATE_FAIL, nr_failed);
1480 trace_mm_migrate_pages(nr_succeeded, nr_failed, mode, reason);
1483 current->flags &= ~PF_SWAPWRITE;
1490 static int store_status(int __user *status, int start, int value, int nr)
1493 if (put_user(value, status + start))
1501 static int do_move_pages_to_node(struct mm_struct *mm,
1502 struct list_head *pagelist, int node)
1506 if (list_empty(pagelist))
1509 err = migrate_pages(pagelist, alloc_new_node_page, NULL, node,
1510 MIGRATE_SYNC, MR_SYSCALL);
1512 putback_movable_pages(pagelist);
1517 * Resolves the given address to a struct page, isolates it from the LRU and
1518 * puts it to the given pagelist.
1520 * errno - if the page cannot be found/isolated
1521 * 0 - when it doesn't have to be migrated because it is already on the
1523 * 1 - when it has been queued
1525 static int add_page_for_migration(struct mm_struct *mm, unsigned long addr,
1526 int node, struct list_head *pagelist, bool migrate_all)
1528 struct vm_area_struct *vma;
1530 unsigned int follflags;
1533 down_read(&mm->mmap_sem);
1535 vma = find_vma(mm, addr);
1536 if (!vma || addr < vma->vm_start || !vma_migratable(vma))
1539 /* FOLL_DUMP to ignore special (like zero) pages */
1540 follflags = FOLL_GET | FOLL_DUMP;
1541 page = follow_page(vma, addr, follflags);
1543 err = PTR_ERR(page);
1552 if (page_to_nid(page) == node)
1556 if (page_mapcount(page) > 1 && !migrate_all)
1559 if (PageHuge(page)) {
1560 if (PageHead(page)) {
1561 isolate_huge_page(page, pagelist);
1567 head = compound_head(page);
1568 err = isolate_lru_page(head);
1573 list_add_tail(&head->lru, pagelist);
1574 mod_node_page_state(page_pgdat(head),
1575 NR_ISOLATED_ANON + page_is_file_cache(head),
1576 hpage_nr_pages(head));
1580 * Either remove the duplicate refcount from
1581 * isolate_lru_page() or drop the page ref if it was
1586 up_read(&mm->mmap_sem);
1591 * Migrate an array of page address onto an array of nodes and fill
1592 * the corresponding array of status.
1594 static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes,
1595 unsigned long nr_pages,
1596 const void __user * __user *pages,
1597 const int __user *nodes,
1598 int __user *status, int flags)
1600 int current_node = NUMA_NO_NODE;
1601 LIST_HEAD(pagelist);
1607 for (i = start = 0; i < nr_pages; i++) {
1608 const void __user *p;
1613 if (get_user(p, pages + i))
1615 if (get_user(node, nodes + i))
1617 addr = (unsigned long)p;
1620 if (node < 0 || node >= MAX_NUMNODES)
1622 if (!node_state(node, N_MEMORY))
1626 if (!node_isset(node, task_nodes))
1629 if (current_node == NUMA_NO_NODE) {
1630 current_node = node;
1632 } else if (node != current_node) {
1633 err = do_move_pages_to_node(mm, &pagelist, current_node);
1636 * Positive err means the number of failed
1637 * pages to migrate. Since we are going to
1638 * abort and return the number of non-migrated
1639 * pages, so need to incude the rest of the
1640 * nr_pages that have not been attempted as
1644 err += nr_pages - i - 1;
1647 err = store_status(status, start, current_node, i - start);
1651 current_node = node;
1655 * Errors in the page lookup or isolation are not fatal and we simply
1656 * report them via status
1658 err = add_page_for_migration(mm, addr, current_node,
1659 &pagelist, flags & MPOL_MF_MOVE_ALL);
1662 /* The page is already on the target node */
1663 err = store_status(status, i, current_node, 1);
1667 } else if (err > 0) {
1668 /* The page is successfully queued for migration */
1672 err = store_status(status, i, err, 1);
1676 err = do_move_pages_to_node(mm, &pagelist, current_node);
1679 err += nr_pages - i - 1;
1683 err = store_status(status, start, current_node, i - start);
1687 current_node = NUMA_NO_NODE;
1690 if (list_empty(&pagelist))
1693 /* Make sure we do not overwrite the existing error */
1694 err1 = do_move_pages_to_node(mm, &pagelist, current_node);
1696 * Don't have to report non-attempted pages here since:
1697 * - If the above loop is done gracefully all pages have been
1699 * - If the above loop is aborted it means a fatal error
1700 * happened, should return ret.
1703 err1 = store_status(status, start, current_node, i - start);
1711 * Determine the nodes of an array of pages and store it in an array of status.
1713 static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages,
1714 const void __user **pages, int *status)
1718 down_read(&mm->mmap_sem);
1720 for (i = 0; i < nr_pages; i++) {
1721 unsigned long addr = (unsigned long)(*pages);
1722 struct vm_area_struct *vma;
1726 vma = find_vma(mm, addr);
1727 if (!vma || addr < vma->vm_start)
1730 /* FOLL_DUMP to ignore special (like zero) pages */
1731 page = follow_page(vma, addr, FOLL_DUMP);
1733 err = PTR_ERR(page);
1737 err = page ? page_to_nid(page) : -ENOENT;
1745 up_read(&mm->mmap_sem);
1749 * Determine the nodes of a user array of pages and store it in
1750 * a user array of status.
1752 static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages,
1753 const void __user * __user *pages,
1756 #define DO_PAGES_STAT_CHUNK_NR 16
1757 const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR];
1758 int chunk_status[DO_PAGES_STAT_CHUNK_NR];
1761 unsigned long chunk_nr;
1763 chunk_nr = nr_pages;
1764 if (chunk_nr > DO_PAGES_STAT_CHUNK_NR)
1765 chunk_nr = DO_PAGES_STAT_CHUNK_NR;
1767 if (copy_from_user(chunk_pages, pages, chunk_nr * sizeof(*chunk_pages)))
1770 do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status);
1772 if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status)))
1777 nr_pages -= chunk_nr;
1779 return nr_pages ? -EFAULT : 0;
1783 * Move a list of pages in the address space of the currently executing
1786 static int kernel_move_pages(pid_t pid, unsigned long nr_pages,
1787 const void __user * __user *pages,
1788 const int __user *nodes,
1789 int __user *status, int flags)
1791 struct task_struct *task;
1792 struct mm_struct *mm;
1794 nodemask_t task_nodes;
1797 if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
1800 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1803 /* Find the mm_struct */
1805 task = pid ? find_task_by_vpid(pid) : current;
1810 get_task_struct(task);
1813 * Check if this process has the right to modify the specified
1814 * process. Use the regular "ptrace_may_access()" checks.
1816 if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) {
1823 err = security_task_movememory(task);
1827 task_nodes = cpuset_mems_allowed(task);
1828 mm = get_task_mm(task);
1829 put_task_struct(task);
1835 err = do_pages_move(mm, task_nodes, nr_pages, pages,
1836 nodes, status, flags);
1838 err = do_pages_stat(mm, nr_pages, pages, status);
1844 put_task_struct(task);
1848 SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages,
1849 const void __user * __user *, pages,
1850 const int __user *, nodes,
1851 int __user *, status, int, flags)
1853 return kernel_move_pages(pid, nr_pages, pages, nodes, status, flags);
1856 #ifdef CONFIG_COMPAT
1857 COMPAT_SYSCALL_DEFINE6(move_pages, pid_t, pid, compat_ulong_t, nr_pages,
1858 compat_uptr_t __user *, pages32,
1859 const int __user *, nodes,
1860 int __user *, status,
1863 const void __user * __user *pages;
1866 pages = compat_alloc_user_space(nr_pages * sizeof(void *));
1867 for (i = 0; i < nr_pages; i++) {
1870 if (get_user(p, pages32 + i) ||
1871 put_user(compat_ptr(p), pages + i))
1874 return kernel_move_pages(pid, nr_pages, pages, nodes, status, flags);
1876 #endif /* CONFIG_COMPAT */
1878 #ifdef CONFIG_NUMA_BALANCING
1880 * Returns true if this is a safe migration target node for misplaced NUMA
1881 * pages. Currently it only checks the watermarks which crude
1883 static bool migrate_balanced_pgdat(struct pglist_data *pgdat,
1884 unsigned long nr_migrate_pages)
1888 for (z = pgdat->nr_zones - 1; z >= 0; z--) {
1889 struct zone *zone = pgdat->node_zones + z;
1891 if (!populated_zone(zone))
1894 /* Avoid waking kswapd by allocating pages_to_migrate pages. */
1895 if (!zone_watermark_ok(zone, 0,
1896 high_wmark_pages(zone) +
1905 static struct page *alloc_misplaced_dst_page(struct page *page,
1908 int nid = (int) data;
1909 struct page *newpage;
1911 newpage = __alloc_pages_node(nid,
1912 (GFP_HIGHUSER_MOVABLE |
1913 __GFP_THISNODE | __GFP_NOMEMALLOC |
1914 __GFP_NORETRY | __GFP_NOWARN) &
1920 static int numamigrate_isolate_page(pg_data_t *pgdat, struct page *page)
1924 VM_BUG_ON_PAGE(compound_order(page) && !PageTransHuge(page), page);
1926 /* Avoid migrating to a node that is nearly full */
1927 if (!migrate_balanced_pgdat(pgdat, 1UL << compound_order(page)))
1930 if (isolate_lru_page(page))
1934 * migrate_misplaced_transhuge_page() skips page migration's usual
1935 * check on page_count(), so we must do it here, now that the page
1936 * has been isolated: a GUP pin, or any other pin, prevents migration.
1937 * The expected page count is 3: 1 for page's mapcount and 1 for the
1938 * caller's pin and 1 for the reference taken by isolate_lru_page().
1940 if (PageTransHuge(page) && page_count(page) != 3) {
1941 putback_lru_page(page);
1945 page_lru = page_is_file_cache(page);
1946 mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON + page_lru,
1947 hpage_nr_pages(page));
1950 * Isolating the page has taken another reference, so the
1951 * caller's reference can be safely dropped without the page
1952 * disappearing underneath us during migration.
1958 bool pmd_trans_migrating(pmd_t pmd)
1960 struct page *page = pmd_page(pmd);
1961 return PageLocked(page);
1965 * Attempt to migrate a misplaced page to the specified destination
1966 * node. Caller is expected to have an elevated reference count on
1967 * the page that will be dropped by this function before returning.
1969 int migrate_misplaced_page(struct page *page, struct vm_area_struct *vma,
1972 pg_data_t *pgdat = NODE_DATA(node);
1975 LIST_HEAD(migratepages);
1978 * Don't migrate file pages that are mapped in multiple processes
1979 * with execute permissions as they are probably shared libraries.
1981 if (page_mapcount(page) != 1 && page_is_file_cache(page) &&
1982 (vma->vm_flags & VM_EXEC))
1986 * Also do not migrate dirty pages as not all filesystems can move
1987 * dirty pages in MIGRATE_ASYNC mode which is a waste of cycles.
1989 if (page_is_file_cache(page) && PageDirty(page))
1992 isolated = numamigrate_isolate_page(pgdat, page);
1996 list_add(&page->lru, &migratepages);
1997 nr_remaining = migrate_pages(&migratepages, alloc_misplaced_dst_page,
1998 NULL, node, MIGRATE_ASYNC,
2001 if (!list_empty(&migratepages)) {
2002 list_del(&page->lru);
2003 dec_node_page_state(page, NR_ISOLATED_ANON +
2004 page_is_file_cache(page));
2005 putback_lru_page(page);
2009 count_vm_numa_event(NUMA_PAGE_MIGRATE);
2010 BUG_ON(!list_empty(&migratepages));
2017 #endif /* CONFIG_NUMA_BALANCING */
2019 #if defined(CONFIG_NUMA_BALANCING) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
2021 * Migrates a THP to a given target node. page must be locked and is unlocked
2024 int migrate_misplaced_transhuge_page(struct mm_struct *mm,
2025 struct vm_area_struct *vma,
2026 pmd_t *pmd, pmd_t entry,
2027 unsigned long address,
2028 struct page *page, int node)
2031 pg_data_t *pgdat = NODE_DATA(node);
2033 struct page *new_page = NULL;
2034 int page_lru = page_is_file_cache(page);
2035 unsigned long mmun_start = address & HPAGE_PMD_MASK;
2036 unsigned long mmun_end = mmun_start + HPAGE_PMD_SIZE;
2038 new_page = alloc_pages_node(node,
2039 (GFP_TRANSHUGE_LIGHT | __GFP_THISNODE),
2043 prep_transhuge_page(new_page);
2045 isolated = numamigrate_isolate_page(pgdat, page);
2051 /* Prepare a page as a migration target */
2052 __SetPageLocked(new_page);
2053 if (PageSwapBacked(page))
2054 __SetPageSwapBacked(new_page);
2056 /* anon mapping, we can simply copy page->mapping to the new page: */
2057 new_page->mapping = page->mapping;
2058 new_page->index = page->index;
2059 migrate_page_copy(new_page, page);
2060 WARN_ON(PageLRU(new_page));
2062 /* Recheck the target PMD */
2063 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2064 ptl = pmd_lock(mm, pmd);
2065 if (unlikely(!pmd_same(*pmd, entry) || !page_ref_freeze(page, 2))) {
2067 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2069 /* Reverse changes made by migrate_page_copy() */
2070 if (TestClearPageActive(new_page))
2071 SetPageActive(page);
2072 if (TestClearPageUnevictable(new_page))
2073 SetPageUnevictable(page);
2075 unlock_page(new_page);
2076 put_page(new_page); /* Free it */
2078 /* Retake the callers reference and putback on LRU */
2080 putback_lru_page(page);
2081 mod_node_page_state(page_pgdat(page),
2082 NR_ISOLATED_ANON + page_lru, -HPAGE_PMD_NR);
2087 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
2088 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
2091 * Overwrite the old entry under pagetable lock and establish
2092 * the new PTE. Any parallel GUP will either observe the old
2093 * page blocking on the page lock, block on the page table
2094 * lock or observe the new page. The SetPageUptodate on the
2095 * new page and page_add_new_anon_rmap guarantee the copy is
2096 * visible before the pagetable update.
2098 flush_cache_range(vma, mmun_start, mmun_end);
2099 page_add_anon_rmap(new_page, vma, mmun_start, true);
2101 * At this point the pmd is numa/protnone (i.e. non present) and the TLB
2102 * has already been flushed globally. So no TLB can be currently
2103 * caching this non present pmd mapping. There's no need to clear the
2104 * pmd before doing set_pmd_at(), nor to flush the TLB after
2105 * set_pmd_at(). Clearing the pmd here would introduce a race
2106 * condition against MADV_DONTNEED, because MADV_DONTNEED only holds the
2107 * mmap_sem for reading. If the pmd is set to NULL at any given time,
2108 * MADV_DONTNEED won't wait on the pmd lock and it'll skip clearing this
2111 set_pmd_at(mm, mmun_start, pmd, entry);
2112 update_mmu_cache_pmd(vma, address, &entry);
2114 page_ref_unfreeze(page, 2);
2115 mlock_migrate_page(new_page, page);
2116 page_remove_rmap(page, true);
2117 set_page_owner_migrate_reason(new_page, MR_NUMA_MISPLACED);
2121 * No need to double call mmu_notifier->invalidate_range() callback as
2122 * the above pmdp_huge_clear_flush_notify() did already call it.
2124 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2126 /* Take an "isolate" reference and put new page on the LRU. */
2128 putback_lru_page(new_page);
2130 unlock_page(new_page);
2132 put_page(page); /* Drop the rmap reference */
2133 put_page(page); /* Drop the LRU isolation reference */
2135 count_vm_events(PGMIGRATE_SUCCESS, HPAGE_PMD_NR);
2136 count_vm_numa_events(NUMA_PAGE_MIGRATE, HPAGE_PMD_NR);
2138 mod_node_page_state(page_pgdat(page),
2139 NR_ISOLATED_ANON + page_lru,
2144 count_vm_events(PGMIGRATE_FAIL, HPAGE_PMD_NR);
2145 ptl = pmd_lock(mm, pmd);
2146 if (pmd_same(*pmd, entry)) {
2147 entry = pmd_modify(entry, vma->vm_page_prot);
2148 set_pmd_at(mm, mmun_start, pmd, entry);
2149 update_mmu_cache_pmd(vma, address, &entry);
2158 #endif /* CONFIG_NUMA_BALANCING */
2160 #endif /* CONFIG_NUMA */
2162 #if defined(CONFIG_MIGRATE_VMA_HELPER)
2163 struct migrate_vma {
2164 struct vm_area_struct *vma;
2167 unsigned long cpages;
2168 unsigned long npages;
2169 unsigned long start;
2173 static int migrate_vma_collect_hole(unsigned long start,
2175 struct mm_walk *walk)
2177 struct migrate_vma *migrate = walk->private;
2180 for (addr = start & PAGE_MASK; addr < end; addr += PAGE_SIZE) {
2181 migrate->src[migrate->npages] = MIGRATE_PFN_MIGRATE;
2182 migrate->dst[migrate->npages] = 0;
2190 static int migrate_vma_collect_skip(unsigned long start,
2192 struct mm_walk *walk)
2194 struct migrate_vma *migrate = walk->private;
2197 for (addr = start & PAGE_MASK; addr < end; addr += PAGE_SIZE) {
2198 migrate->dst[migrate->npages] = 0;
2199 migrate->src[migrate->npages++] = 0;
2205 static int migrate_vma_collect_pmd(pmd_t *pmdp,
2206 unsigned long start,
2208 struct mm_walk *walk)
2210 struct migrate_vma *migrate = walk->private;
2211 struct vm_area_struct *vma = walk->vma;
2212 struct mm_struct *mm = vma->vm_mm;
2213 unsigned long addr = start, unmapped = 0;
2218 if (pmd_none(*pmdp))
2219 return migrate_vma_collect_hole(start, end, walk);
2221 if (pmd_trans_huge(*pmdp)) {
2224 ptl = pmd_lock(mm, pmdp);
2225 if (unlikely(!pmd_trans_huge(*pmdp))) {
2230 page = pmd_page(*pmdp);
2231 if (is_huge_zero_page(page)) {
2233 split_huge_pmd(vma, pmdp, addr);
2234 if (pmd_trans_unstable(pmdp))
2235 return migrate_vma_collect_skip(start, end,
2242 if (unlikely(!trylock_page(page)))
2243 return migrate_vma_collect_skip(start, end,
2245 ret = split_huge_page(page);
2249 return migrate_vma_collect_skip(start, end,
2251 if (pmd_none(*pmdp))
2252 return migrate_vma_collect_hole(start, end,
2257 if (unlikely(pmd_bad(*pmdp)))
2258 return migrate_vma_collect_skip(start, end, walk);
2260 ptep = pte_offset_map_lock(mm, pmdp, addr, &ptl);
2261 arch_enter_lazy_mmu_mode();
2263 for (; addr < end; addr += PAGE_SIZE, ptep++) {
2264 unsigned long mpfn, pfn;
2272 if (pte_none(pte)) {
2273 mpfn = MIGRATE_PFN_MIGRATE;
2279 if (!pte_present(pte)) {
2283 * Only care about unaddressable device page special
2284 * page table entry. Other special swap entries are not
2285 * migratable, and we ignore regular swapped page.
2287 entry = pte_to_swp_entry(pte);
2288 if (!is_device_private_entry(entry))
2291 page = device_private_entry_to_page(entry);
2292 mpfn = migrate_pfn(page_to_pfn(page))|
2293 MIGRATE_PFN_DEVICE | MIGRATE_PFN_MIGRATE;
2294 if (is_write_device_private_entry(entry))
2295 mpfn |= MIGRATE_PFN_WRITE;
2297 if (is_zero_pfn(pfn)) {
2298 mpfn = MIGRATE_PFN_MIGRATE;
2303 page = _vm_normal_page(migrate->vma, addr, pte, true);
2304 mpfn = migrate_pfn(pfn) | MIGRATE_PFN_MIGRATE;
2305 mpfn |= pte_write(pte) ? MIGRATE_PFN_WRITE : 0;
2308 /* FIXME support THP */
2309 if (!page || !page->mapping || PageTransCompound(page)) {
2313 pfn = page_to_pfn(page);
2316 * By getting a reference on the page we pin it and that blocks
2317 * any kind of migration. Side effect is that it "freezes" the
2320 * We drop this reference after isolating the page from the lru
2321 * for non device page (device page are not on the lru and thus
2322 * can't be dropped from it).
2328 * Optimize for the common case where page is only mapped once
2329 * in one process. If we can lock the page, then we can safely
2330 * set up a special migration page table entry now.
2332 if (trylock_page(page)) {
2335 mpfn |= MIGRATE_PFN_LOCKED;
2336 ptep_get_and_clear(mm, addr, ptep);
2338 /* Setup special migration page table entry */
2339 entry = make_migration_entry(page, mpfn &
2341 swp_pte = swp_entry_to_pte(entry);
2342 if (pte_soft_dirty(pte))
2343 swp_pte = pte_swp_mksoft_dirty(swp_pte);
2344 set_pte_at(mm, addr, ptep, swp_pte);
2347 * This is like regular unmap: we remove the rmap and
2348 * drop page refcount. Page won't be freed, as we took
2349 * a reference just above.
2351 page_remove_rmap(page, false);
2354 if (pte_present(pte))
2359 migrate->dst[migrate->npages] = 0;
2360 migrate->src[migrate->npages++] = mpfn;
2362 arch_leave_lazy_mmu_mode();
2363 pte_unmap_unlock(ptep - 1, ptl);
2365 /* Only flush the TLB if we actually modified any entries */
2367 flush_tlb_range(walk->vma, start, end);
2373 * migrate_vma_collect() - collect pages over a range of virtual addresses
2374 * @migrate: migrate struct containing all migration information
2376 * This will walk the CPU page table. For each virtual address backed by a
2377 * valid page, it updates the src array and takes a reference on the page, in
2378 * order to pin the page until we lock it and unmap it.
2380 static void migrate_vma_collect(struct migrate_vma *migrate)
2382 struct mm_walk mm_walk = {
2383 .pmd_entry = migrate_vma_collect_pmd,
2384 .pte_hole = migrate_vma_collect_hole,
2385 .vma = migrate->vma,
2386 .mm = migrate->vma->vm_mm,
2390 mmu_notifier_invalidate_range_start(mm_walk.mm,
2393 walk_page_range(migrate->start, migrate->end, &mm_walk);
2394 mmu_notifier_invalidate_range_end(mm_walk.mm,
2398 migrate->end = migrate->start + (migrate->npages << PAGE_SHIFT);
2402 * migrate_vma_check_page() - check if page is pinned or not
2403 * @page: struct page to check
2405 * Pinned pages cannot be migrated. This is the same test as in
2406 * migrate_page_move_mapping(), except that here we allow migration of a
2409 static bool migrate_vma_check_page(struct page *page)
2412 * One extra ref because caller holds an extra reference, either from
2413 * isolate_lru_page() for a regular page, or migrate_vma_collect() for
2419 * FIXME support THP (transparent huge page), it is bit more complex to
2420 * check them than regular pages, because they can be mapped with a pmd
2421 * or with a pte (split pte mapping).
2423 if (PageCompound(page))
2426 /* Page from ZONE_DEVICE have one extra reference */
2427 if (is_zone_device_page(page)) {
2429 * Private page can never be pin as they have no valid pte and
2430 * GUP will fail for those. Yet if there is a pending migration
2431 * a thread might try to wait on the pte migration entry and
2432 * will bump the page reference count. Sadly there is no way to
2433 * differentiate a regular pin from migration wait. Hence to
2434 * avoid 2 racing thread trying to migrate back to CPU to enter
2435 * infinite loop (one stoping migration because the other is
2436 * waiting on pte migration entry). We always return true here.
2438 * FIXME proper solution is to rework migration_entry_wait() so
2439 * it does not need to take a reference on page.
2441 if (is_device_private_page(page))
2445 * Only allow device public page to be migrated and account for
2446 * the extra reference count imply by ZONE_DEVICE pages.
2448 if (!is_device_public_page(page))
2453 /* For file back page */
2454 if (page_mapping(page))
2455 extra += 1 + page_has_private(page);
2457 if ((page_count(page) - extra) > page_mapcount(page))
2464 * migrate_vma_prepare() - lock pages and isolate them from the lru
2465 * @migrate: migrate struct containing all migration information
2467 * This locks pages that have been collected by migrate_vma_collect(). Once each
2468 * page is locked it is isolated from the lru (for non-device pages). Finally,
2469 * the ref taken by migrate_vma_collect() is dropped, as locked pages cannot be
2470 * migrated by concurrent kernel threads.
2472 static void migrate_vma_prepare(struct migrate_vma *migrate)
2474 const unsigned long npages = migrate->npages;
2475 const unsigned long start = migrate->start;
2476 unsigned long addr, i, restore = 0;
2477 bool allow_drain = true;
2481 for (i = 0; (i < npages) && migrate->cpages; i++) {
2482 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2488 if (!(migrate->src[i] & MIGRATE_PFN_LOCKED)) {
2490 * Because we are migrating several pages there can be
2491 * a deadlock between 2 concurrent migration where each
2492 * are waiting on each other page lock.
2494 * Make migrate_vma() a best effort thing and backoff
2495 * for any page we can not lock right away.
2497 if (!trylock_page(page)) {
2498 migrate->src[i] = 0;
2504 migrate->src[i] |= MIGRATE_PFN_LOCKED;
2507 /* ZONE_DEVICE pages are not on LRU */
2508 if (!is_zone_device_page(page)) {
2509 if (!PageLRU(page) && allow_drain) {
2510 /* Drain CPU's pagevec */
2511 lru_add_drain_all();
2512 allow_drain = false;
2515 if (isolate_lru_page(page)) {
2517 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2521 migrate->src[i] = 0;
2529 /* Drop the reference we took in collect */
2533 if (!migrate_vma_check_page(page)) {
2535 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2539 if (!is_zone_device_page(page)) {
2541 putback_lru_page(page);
2544 migrate->src[i] = 0;
2548 if (!is_zone_device_page(page))
2549 putback_lru_page(page);
2556 for (i = 0, addr = start; i < npages && restore; i++, addr += PAGE_SIZE) {
2557 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2559 if (!page || (migrate->src[i] & MIGRATE_PFN_MIGRATE))
2562 remove_migration_pte(page, migrate->vma, addr, page);
2564 migrate->src[i] = 0;
2572 * migrate_vma_unmap() - replace page mapping with special migration pte entry
2573 * @migrate: migrate struct containing all migration information
2575 * Replace page mapping (CPU page table pte) with a special migration pte entry
2576 * and check again if it has been pinned. Pinned pages are restored because we
2577 * cannot migrate them.
2579 * This is the last step before we call the device driver callback to allocate
2580 * destination memory and copy contents of original page over to new page.
2582 static void migrate_vma_unmap(struct migrate_vma *migrate)
2584 int flags = TTU_MIGRATION | TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS;
2585 const unsigned long npages = migrate->npages;
2586 const unsigned long start = migrate->start;
2587 unsigned long addr, i, restore = 0;
2589 for (i = 0; i < npages; i++) {
2590 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2592 if (!page || !(migrate->src[i] & MIGRATE_PFN_MIGRATE))
2595 if (page_mapped(page)) {
2596 try_to_unmap(page, flags);
2597 if (page_mapped(page))
2601 if (migrate_vma_check_page(page))
2605 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2610 for (addr = start, i = 0; i < npages && restore; addr += PAGE_SIZE, i++) {
2611 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2613 if (!page || (migrate->src[i] & MIGRATE_PFN_MIGRATE))
2616 remove_migration_ptes(page, page, false);
2618 migrate->src[i] = 0;
2622 if (is_zone_device_page(page))
2625 putback_lru_page(page);
2629 static void migrate_vma_insert_page(struct migrate_vma *migrate,
2635 struct vm_area_struct *vma = migrate->vma;
2636 struct mm_struct *mm = vma->vm_mm;
2637 struct mem_cgroup *memcg;
2647 /* Only allow populating anonymous memory */
2648 if (!vma_is_anonymous(vma))
2651 pgdp = pgd_offset(mm, addr);
2652 p4dp = p4d_alloc(mm, pgdp, addr);
2655 pudp = pud_alloc(mm, p4dp, addr);
2658 pmdp = pmd_alloc(mm, pudp, addr);
2662 if (pmd_trans_huge(*pmdp) || pmd_devmap(*pmdp))
2666 * Use pte_alloc() instead of pte_alloc_map(). We can't run
2667 * pte_offset_map() on pmds where a huge pmd might be created
2668 * from a different thread.
2670 * pte_alloc_map() is safe to use under down_write(mmap_sem) or when
2671 * parallel threads are excluded by other means.
2673 * Here we only have down_read(mmap_sem).
2675 if (pte_alloc(mm, pmdp, addr))
2678 /* See the comment in pte_alloc_one_map() */
2679 if (unlikely(pmd_trans_unstable(pmdp)))
2682 if (unlikely(anon_vma_prepare(vma)))
2684 if (mem_cgroup_try_charge(page, vma->vm_mm, GFP_KERNEL, &memcg, false))
2688 * The memory barrier inside __SetPageUptodate makes sure that
2689 * preceding stores to the page contents become visible before
2690 * the set_pte_at() write.
2692 __SetPageUptodate(page);
2694 if (is_zone_device_page(page)) {
2695 if (is_device_private_page(page)) {
2696 swp_entry_t swp_entry;
2698 swp_entry = make_device_private_entry(page, vma->vm_flags & VM_WRITE);
2699 entry = swp_entry_to_pte(swp_entry);
2700 } else if (is_device_public_page(page)) {
2701 entry = pte_mkold(mk_pte(page, READ_ONCE(vma->vm_page_prot)));
2702 if (vma->vm_flags & VM_WRITE)
2703 entry = pte_mkwrite(pte_mkdirty(entry));
2704 entry = pte_mkdevmap(entry);
2707 entry = mk_pte(page, vma->vm_page_prot);
2708 if (vma->vm_flags & VM_WRITE)
2709 entry = pte_mkwrite(pte_mkdirty(entry));
2712 ptep = pte_offset_map_lock(mm, pmdp, addr, &ptl);
2714 if (pte_present(*ptep)) {
2715 unsigned long pfn = pte_pfn(*ptep);
2717 if (!is_zero_pfn(pfn)) {
2718 pte_unmap_unlock(ptep, ptl);
2719 mem_cgroup_cancel_charge(page, memcg, false);
2723 } else if (!pte_none(*ptep)) {
2724 pte_unmap_unlock(ptep, ptl);
2725 mem_cgroup_cancel_charge(page, memcg, false);
2730 * Check for usefaultfd but do not deliver the fault. Instead,
2733 if (userfaultfd_missing(vma)) {
2734 pte_unmap_unlock(ptep, ptl);
2735 mem_cgroup_cancel_charge(page, memcg, false);
2739 inc_mm_counter(mm, MM_ANONPAGES);
2740 page_add_new_anon_rmap(page, vma, addr, false);
2741 mem_cgroup_commit_charge(page, memcg, false, false);
2742 if (!is_zone_device_page(page))
2743 lru_cache_add_active_or_unevictable(page, vma);
2747 flush_cache_page(vma, addr, pte_pfn(*ptep));
2748 ptep_clear_flush_notify(vma, addr, ptep);
2749 set_pte_at_notify(mm, addr, ptep, entry);
2750 update_mmu_cache(vma, addr, ptep);
2752 /* No need to invalidate - it was non-present before */
2753 set_pte_at(mm, addr, ptep, entry);
2754 update_mmu_cache(vma, addr, ptep);
2757 pte_unmap_unlock(ptep, ptl);
2758 *src = MIGRATE_PFN_MIGRATE;
2762 *src &= ~MIGRATE_PFN_MIGRATE;
2766 * migrate_vma_pages() - migrate meta-data from src page to dst page
2767 * @migrate: migrate struct containing all migration information
2769 * This migrates struct page meta-data from source struct page to destination
2770 * struct page. This effectively finishes the migration from source page to the
2773 static void migrate_vma_pages(struct migrate_vma *migrate)
2775 const unsigned long npages = migrate->npages;
2776 const unsigned long start = migrate->start;
2777 struct vm_area_struct *vma = migrate->vma;
2778 struct mm_struct *mm = vma->vm_mm;
2779 unsigned long addr, i, mmu_start;
2780 bool notified = false;
2782 for (i = 0, addr = start; i < npages; addr += PAGE_SIZE, i++) {
2783 struct page *newpage = migrate_pfn_to_page(migrate->dst[i]);
2784 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2785 struct address_space *mapping;
2789 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2794 if (!(migrate->src[i] & MIGRATE_PFN_MIGRATE)) {
2800 mmu_notifier_invalidate_range_start(mm,
2804 migrate_vma_insert_page(migrate, addr, newpage,
2810 mapping = page_mapping(page);
2812 if (is_zone_device_page(newpage)) {
2813 if (is_device_private_page(newpage)) {
2815 * For now only support private anonymous when
2816 * migrating to un-addressable device memory.
2819 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2822 } else if (!is_device_public_page(newpage)) {
2824 * Other types of ZONE_DEVICE page are not
2827 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2832 r = migrate_page(mapping, newpage, page, MIGRATE_SYNC_NO_COPY);
2833 if (r != MIGRATEPAGE_SUCCESS)
2834 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2838 * No need to double call mmu_notifier->invalidate_range() callback as
2839 * the above ptep_clear_flush_notify() inside migrate_vma_insert_page()
2840 * did already call it.
2843 mmu_notifier_invalidate_range_only_end(mm, mmu_start,
2848 * migrate_vma_finalize() - restore CPU page table entry
2849 * @migrate: migrate struct containing all migration information
2851 * This replaces the special migration pte entry with either a mapping to the
2852 * new page if migration was successful for that page, or to the original page
2855 * This also unlocks the pages and puts them back on the lru, or drops the extra
2856 * refcount, for device pages.
2858 static void migrate_vma_finalize(struct migrate_vma *migrate)
2860 const unsigned long npages = migrate->npages;
2863 for (i = 0; i < npages; i++) {
2864 struct page *newpage = migrate_pfn_to_page(migrate->dst[i]);
2865 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2869 unlock_page(newpage);
2875 if (!(migrate->src[i] & MIGRATE_PFN_MIGRATE) || !newpage) {
2877 unlock_page(newpage);
2883 remove_migration_ptes(page, newpage, false);
2887 if (is_zone_device_page(page))
2890 putback_lru_page(page);
2892 if (newpage != page) {
2893 unlock_page(newpage);
2894 if (is_zone_device_page(newpage))
2897 putback_lru_page(newpage);
2903 * migrate_vma() - migrate a range of memory inside vma
2905 * @ops: migration callback for allocating destination memory and copying
2906 * @vma: virtual memory area containing the range to be migrated
2907 * @start: start address of the range to migrate (inclusive)
2908 * @end: end address of the range to migrate (exclusive)
2909 * @src: array of hmm_pfn_t containing source pfns
2910 * @dst: array of hmm_pfn_t containing destination pfns
2911 * @private: pointer passed back to each of the callback
2912 * Returns: 0 on success, error code otherwise
2914 * This function tries to migrate a range of memory virtual address range, using
2915 * callbacks to allocate and copy memory from source to destination. First it
2916 * collects all the pages backing each virtual address in the range, saving this
2917 * inside the src array. Then it locks those pages and unmaps them. Once the pages
2918 * are locked and unmapped, it checks whether each page is pinned or not. Pages
2919 * that aren't pinned have the MIGRATE_PFN_MIGRATE flag set (by this function)
2920 * in the corresponding src array entry. It then restores any pages that are
2921 * pinned, by remapping and unlocking those pages.
2923 * At this point it calls the alloc_and_copy() callback. For documentation on
2924 * what is expected from that callback, see struct migrate_vma_ops comments in
2925 * include/linux/migrate.h
2927 * After the alloc_and_copy() callback, this function goes over each entry in
2928 * the src array that has the MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE flag
2929 * set. If the corresponding entry in dst array has MIGRATE_PFN_VALID flag set,
2930 * then the function tries to migrate struct page information from the source
2931 * struct page to the destination struct page. If it fails to migrate the struct
2932 * page information, then it clears the MIGRATE_PFN_MIGRATE flag in the src
2935 * At this point all successfully migrated pages have an entry in the src
2936 * array with MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE flag set and the dst
2937 * array entry with MIGRATE_PFN_VALID flag set.
2939 * It then calls the finalize_and_map() callback. See comments for "struct
2940 * migrate_vma_ops", in include/linux/migrate.h for details about
2941 * finalize_and_map() behavior.
2943 * After the finalize_and_map() callback, for successfully migrated pages, this
2944 * function updates the CPU page table to point to new pages, otherwise it
2945 * restores the CPU page table to point to the original source pages.
2947 * Function returns 0 after the above steps, even if no pages were migrated
2948 * (The function only returns an error if any of the arguments are invalid.)
2950 * Both src and dst array must be big enough for (end - start) >> PAGE_SHIFT
2951 * unsigned long entries.
2953 int migrate_vma(const struct migrate_vma_ops *ops,
2954 struct vm_area_struct *vma,
2955 unsigned long start,
2961 struct migrate_vma migrate;
2963 /* Sanity check the arguments */
2966 if (!vma || is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_SPECIAL) ||
2969 if (start < vma->vm_start || start >= vma->vm_end)
2971 if (end <= vma->vm_start || end > vma->vm_end)
2973 if (!ops || !src || !dst || start >= end)
2976 memset(src, 0, sizeof(*src) * ((end - start) >> PAGE_SHIFT));
2979 migrate.start = start;
2985 /* Collect, and try to unmap source pages */
2986 migrate_vma_collect(&migrate);
2987 if (!migrate.cpages)
2990 /* Lock and isolate page */
2991 migrate_vma_prepare(&migrate);
2992 if (!migrate.cpages)
2996 migrate_vma_unmap(&migrate);
2997 if (!migrate.cpages)
3001 * At this point pages are locked and unmapped, and thus they have
3002 * stable content and can safely be copied to destination memory that
3003 * is allocated by the callback.
3005 * Note that migration can fail in migrate_vma_struct_page() for each
3008 ops->alloc_and_copy(vma, src, dst, start, end, private);
3010 /* This does the real migration of struct page */
3011 migrate_vma_pages(&migrate);
3013 ops->finalize_and_map(vma, src, dst, start, end, private);
3015 /* Unlock and remap pages */
3016 migrate_vma_finalize(&migrate);
3020 EXPORT_SYMBOL(migrate_vma);
3021 #endif /* defined(MIGRATE_VMA_HELPER) */