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)) {
519 SetPageSwapCache(newpage);
520 for (i = 0; i < (1 << compound_order(page)); i++)
521 set_page_private(newpage + i,
522 page_private(page + i));
525 VM_BUG_ON_PAGE(PageSwapCache(page), page);
528 /* Move dirty while page refs frozen and newpage not yet exposed */
529 dirty = PageDirty(page);
531 ClearPageDirty(page);
532 SetPageDirty(newpage);
535 radix_tree_replace_slot(&mapping->i_pages, pslot, newpage);
536 if (PageTransHuge(page)) {
538 int index = page_index(page);
540 for (i = 1; i < HPAGE_PMD_NR; i++) {
541 pslot = radix_tree_lookup_slot(&mapping->i_pages,
543 radix_tree_replace_slot(&mapping->i_pages, pslot,
549 * Drop cache reference from old page by unfreezing
550 * to one less reference.
551 * We know this isn't the last reference.
553 page_ref_unfreeze(page, expected_count - hpage_nr_pages(page));
555 xa_unlock(&mapping->i_pages);
556 /* Leave irq disabled to prevent preemption while updating stats */
559 * If moved to a different zone then also account
560 * the page for that zone. Other VM counters will be
561 * taken care of when we establish references to the
562 * new page and drop references to the old page.
564 * Note that anonymous pages are accounted for
565 * via NR_FILE_PAGES and NR_ANON_MAPPED if they
566 * are mapped to swap space.
568 if (newzone != oldzone) {
569 __dec_node_state(oldzone->zone_pgdat, NR_FILE_PAGES);
570 __inc_node_state(newzone->zone_pgdat, NR_FILE_PAGES);
571 if (PageSwapBacked(page) && !PageSwapCache(page)) {
572 __dec_node_state(oldzone->zone_pgdat, NR_SHMEM);
573 __inc_node_state(newzone->zone_pgdat, NR_SHMEM);
575 if (dirty && mapping_cap_account_dirty(mapping)) {
576 __dec_node_state(oldzone->zone_pgdat, NR_FILE_DIRTY);
577 __dec_zone_state(oldzone, NR_ZONE_WRITE_PENDING);
578 __inc_node_state(newzone->zone_pgdat, NR_FILE_DIRTY);
579 __inc_zone_state(newzone, NR_ZONE_WRITE_PENDING);
584 return MIGRATEPAGE_SUCCESS;
586 EXPORT_SYMBOL(migrate_page_move_mapping);
589 * The expected number of remaining references is the same as that
590 * of migrate_page_move_mapping().
592 int migrate_huge_page_move_mapping(struct address_space *mapping,
593 struct page *newpage, struct page *page)
598 xa_lock_irq(&mapping->i_pages);
600 pslot = radix_tree_lookup_slot(&mapping->i_pages, page_index(page));
602 xa_unlock_irq(&mapping->i_pages);
606 expected_count = 2 + page_has_private(page);
607 if (page_count(page) != expected_count ||
608 radix_tree_deref_slot_protected(pslot, &mapping->i_pages.xa_lock) != page) {
609 xa_unlock_irq(&mapping->i_pages);
613 if (!page_ref_freeze(page, expected_count)) {
614 xa_unlock_irq(&mapping->i_pages);
618 newpage->index = page->index;
619 newpage->mapping = page->mapping;
623 radix_tree_replace_slot(&mapping->i_pages, pslot, newpage);
625 page_ref_unfreeze(page, expected_count - 1);
627 xa_unlock_irq(&mapping->i_pages);
629 return MIGRATEPAGE_SUCCESS;
633 * Gigantic pages are so large that we do not guarantee that page++ pointer
634 * arithmetic will work across the entire page. We need something more
637 static void __copy_gigantic_page(struct page *dst, struct page *src,
641 struct page *dst_base = dst;
642 struct page *src_base = src;
644 for (i = 0; i < nr_pages; ) {
646 copy_highpage(dst, src);
649 dst = mem_map_next(dst, dst_base, i);
650 src = mem_map_next(src, src_base, i);
654 static void copy_huge_page(struct page *dst, struct page *src)
661 struct hstate *h = page_hstate(src);
662 nr_pages = pages_per_huge_page(h);
664 if (unlikely(nr_pages > MAX_ORDER_NR_PAGES)) {
665 __copy_gigantic_page(dst, src, nr_pages);
670 BUG_ON(!PageTransHuge(src));
671 nr_pages = hpage_nr_pages(src);
674 for (i = 0; i < nr_pages; i++) {
676 copy_highpage(dst + i, src + i);
681 * Copy the page to its new location
683 void migrate_page_states(struct page *newpage, struct page *page)
688 SetPageError(newpage);
689 if (PageReferenced(page))
690 SetPageReferenced(newpage);
691 if (PageUptodate(page))
692 SetPageUptodate(newpage);
693 if (TestClearPageActive(page)) {
694 VM_BUG_ON_PAGE(PageUnevictable(page), page);
695 SetPageActive(newpage);
696 } else if (TestClearPageUnevictable(page))
697 SetPageUnevictable(newpage);
698 if (PageChecked(page))
699 SetPageChecked(newpage);
700 if (PageMappedToDisk(page))
701 SetPageMappedToDisk(newpage);
703 /* Move dirty on pages not done by migrate_page_move_mapping() */
705 SetPageDirty(newpage);
707 if (page_is_young(page))
708 set_page_young(newpage);
709 if (page_is_idle(page))
710 set_page_idle(newpage);
713 * Copy NUMA information to the new page, to prevent over-eager
714 * future migrations of this same page.
716 cpupid = page_cpupid_xchg_last(page, -1);
717 page_cpupid_xchg_last(newpage, cpupid);
719 ksm_migrate_page(newpage, page);
721 * Please do not reorder this without considering how mm/ksm.c's
722 * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache().
724 if (PageSwapCache(page))
725 ClearPageSwapCache(page);
726 ClearPagePrivate(page);
727 set_page_private(page, 0);
730 * If any waiters have accumulated on the new page then
733 if (PageWriteback(newpage))
734 end_page_writeback(newpage);
736 copy_page_owner(page, newpage);
738 mem_cgroup_migrate(page, newpage);
740 EXPORT_SYMBOL(migrate_page_states);
742 void migrate_page_copy(struct page *newpage, struct page *page)
744 if (PageHuge(page) || PageTransHuge(page))
745 copy_huge_page(newpage, page);
747 copy_highpage(newpage, page);
749 migrate_page_states(newpage, page);
751 EXPORT_SYMBOL(migrate_page_copy);
753 /************************************************************
754 * Migration functions
755 ***********************************************************/
758 * Common logic to directly migrate a single LRU page suitable for
759 * pages that do not use PagePrivate/PagePrivate2.
761 * Pages are locked upon entry and exit.
763 int migrate_page(struct address_space *mapping,
764 struct page *newpage, struct page *page,
765 enum migrate_mode mode)
769 BUG_ON(PageWriteback(page)); /* Writeback must be complete */
771 rc = migrate_page_move_mapping(mapping, newpage, page, NULL, mode, 0);
773 if (rc != MIGRATEPAGE_SUCCESS)
776 if (mode != MIGRATE_SYNC_NO_COPY)
777 migrate_page_copy(newpage, page);
779 migrate_page_states(newpage, page);
780 return MIGRATEPAGE_SUCCESS;
782 EXPORT_SYMBOL(migrate_page);
786 * Migration function for pages with buffers. This function can only be used
787 * if the underlying filesystem guarantees that no other references to "page"
790 int buffer_migrate_page(struct address_space *mapping,
791 struct page *newpage, struct page *page, enum migrate_mode mode)
793 struct buffer_head *bh, *head;
796 if (!page_has_buffers(page))
797 return migrate_page(mapping, newpage, page, mode);
799 head = page_buffers(page);
801 rc = migrate_page_move_mapping(mapping, newpage, page, head, mode, 0);
803 if (rc != MIGRATEPAGE_SUCCESS)
807 * In the async case, migrate_page_move_mapping locked the buffers
808 * with an IRQ-safe spinlock held. In the sync case, the buffers
809 * need to be locked now
811 if (mode != MIGRATE_ASYNC)
812 BUG_ON(!buffer_migrate_lock_buffers(head, mode));
814 ClearPagePrivate(page);
815 set_page_private(newpage, page_private(page));
816 set_page_private(page, 0);
822 set_bh_page(bh, newpage, bh_offset(bh));
823 bh = bh->b_this_page;
825 } while (bh != head);
827 SetPagePrivate(newpage);
829 if (mode != MIGRATE_SYNC_NO_COPY)
830 migrate_page_copy(newpage, page);
832 migrate_page_states(newpage, page);
838 bh = bh->b_this_page;
840 } while (bh != head);
842 return MIGRATEPAGE_SUCCESS;
844 EXPORT_SYMBOL(buffer_migrate_page);
848 * Writeback a page to clean the dirty state
850 static int writeout(struct address_space *mapping, struct page *page)
852 struct writeback_control wbc = {
853 .sync_mode = WB_SYNC_NONE,
856 .range_end = LLONG_MAX,
861 if (!mapping->a_ops->writepage)
862 /* No write method for the address space */
865 if (!clear_page_dirty_for_io(page))
866 /* Someone else already triggered a write */
870 * A dirty page may imply that the underlying filesystem has
871 * the page on some queue. So the page must be clean for
872 * migration. Writeout may mean we loose the lock and the
873 * page state is no longer what we checked for earlier.
874 * At this point we know that the migration attempt cannot
877 remove_migration_ptes(page, page, false);
879 rc = mapping->a_ops->writepage(page, &wbc);
881 if (rc != AOP_WRITEPAGE_ACTIVATE)
882 /* unlocked. Relock */
885 return (rc < 0) ? -EIO : -EAGAIN;
889 * Default handling if a filesystem does not provide a migration function.
891 static int fallback_migrate_page(struct address_space *mapping,
892 struct page *newpage, struct page *page, enum migrate_mode mode)
894 if (PageDirty(page)) {
895 /* Only writeback pages in full synchronous migration */
898 case MIGRATE_SYNC_NO_COPY:
903 return writeout(mapping, page);
907 * Buffers may be managed in a filesystem specific way.
908 * We must have no buffers or drop them.
910 if (page_has_private(page) &&
911 !try_to_release_page(page, GFP_KERNEL))
914 return migrate_page(mapping, newpage, page, mode);
918 * Move a page to a newly allocated page
919 * The page is locked and all ptes have been successfully removed.
921 * The new page will have replaced the old page if this function
926 * MIGRATEPAGE_SUCCESS - success
928 static int move_to_new_page(struct page *newpage, struct page *page,
929 enum migrate_mode mode)
931 struct address_space *mapping;
933 bool is_lru = !__PageMovable(page);
935 VM_BUG_ON_PAGE(!PageLocked(page), page);
936 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
938 mapping = page_mapping(page);
940 if (likely(is_lru)) {
942 rc = migrate_page(mapping, newpage, page, mode);
943 else if (mapping->a_ops->migratepage)
945 * Most pages have a mapping and most filesystems
946 * provide a migratepage callback. Anonymous pages
947 * are part of swap space which also has its own
948 * migratepage callback. This is the most common path
949 * for page migration.
951 rc = mapping->a_ops->migratepage(mapping, newpage,
954 rc = fallback_migrate_page(mapping, newpage,
958 * In case of non-lru page, it could be released after
959 * isolation step. In that case, we shouldn't try migration.
961 VM_BUG_ON_PAGE(!PageIsolated(page), page);
962 if (!PageMovable(page)) {
963 rc = MIGRATEPAGE_SUCCESS;
964 __ClearPageIsolated(page);
968 rc = mapping->a_ops->migratepage(mapping, newpage,
970 WARN_ON_ONCE(rc == MIGRATEPAGE_SUCCESS &&
971 !PageIsolated(page));
975 * When successful, old pagecache page->mapping must be cleared before
976 * page is freed; but stats require that PageAnon be left as PageAnon.
978 if (rc == MIGRATEPAGE_SUCCESS) {
979 if (__PageMovable(page)) {
980 VM_BUG_ON_PAGE(!PageIsolated(page), page);
983 * We clear PG_movable under page_lock so any compactor
984 * cannot try to migrate this page.
986 __ClearPageIsolated(page);
990 * Anonymous and movable page->mapping will be cleard by
991 * free_pages_prepare so don't reset it here for keeping
992 * the type to work PageAnon, for example.
994 if (!PageMappingFlags(page))
995 page->mapping = NULL;
997 if (unlikely(is_zone_device_page(newpage))) {
998 if (is_device_public_page(newpage))
999 flush_dcache_page(newpage);
1001 flush_dcache_page(newpage);
1008 static int __unmap_and_move(struct page *page, struct page *newpage,
1009 int force, enum migrate_mode mode)
1012 int page_was_mapped = 0;
1013 struct anon_vma *anon_vma = NULL;
1014 bool is_lru = !__PageMovable(page);
1016 if (!trylock_page(page)) {
1017 if (!force || mode == MIGRATE_ASYNC)
1021 * It's not safe for direct compaction to call lock_page.
1022 * For example, during page readahead pages are added locked
1023 * to the LRU. Later, when the IO completes the pages are
1024 * marked uptodate and unlocked. However, the queueing
1025 * could be merging multiple pages for one bio (e.g.
1026 * mpage_readpages). If an allocation happens for the
1027 * second or third page, the process can end up locking
1028 * the same page twice and deadlocking. Rather than
1029 * trying to be clever about what pages can be locked,
1030 * avoid the use of lock_page for direct compaction
1033 if (current->flags & PF_MEMALLOC)
1039 if (PageWriteback(page)) {
1041 * Only in the case of a full synchronous migration is it
1042 * necessary to wait for PageWriteback. In the async case,
1043 * the retry loop is too short and in the sync-light case,
1044 * the overhead of stalling is too much
1048 case MIGRATE_SYNC_NO_COPY:
1056 wait_on_page_writeback(page);
1060 * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
1061 * we cannot notice that anon_vma is freed while we migrates a page.
1062 * This get_anon_vma() delays freeing anon_vma pointer until the end
1063 * of migration. File cache pages are no problem because of page_lock()
1064 * File Caches may use write_page() or lock_page() in migration, then,
1065 * just care Anon page here.
1067 * Only page_get_anon_vma() understands the subtleties of
1068 * getting a hold on an anon_vma from outside one of its mms.
1069 * But if we cannot get anon_vma, then we won't need it anyway,
1070 * because that implies that the anon page is no longer mapped
1071 * (and cannot be remapped so long as we hold the page lock).
1073 if (PageAnon(page) && !PageKsm(page))
1074 anon_vma = page_get_anon_vma(page);
1077 * Block others from accessing the new page when we get around to
1078 * establishing additional references. We are usually the only one
1079 * holding a reference to newpage at this point. We used to have a BUG
1080 * here if trylock_page(newpage) fails, but would like to allow for
1081 * cases where there might be a race with the previous use of newpage.
1082 * This is much like races on refcount of oldpage: just don't BUG().
1084 if (unlikely(!trylock_page(newpage)))
1087 if (unlikely(!is_lru)) {
1088 rc = move_to_new_page(newpage, page, mode);
1089 goto out_unlock_both;
1093 * Corner case handling:
1094 * 1. When a new swap-cache page is read into, it is added to the LRU
1095 * and treated as swapcache but it has no rmap yet.
1096 * Calling try_to_unmap() against a page->mapping==NULL page will
1097 * trigger a BUG. So handle it here.
1098 * 2. An orphaned page (see truncate_complete_page) might have
1099 * fs-private metadata. The page can be picked up due to memory
1100 * offlining. Everywhere else except page reclaim, the page is
1101 * invisible to the vm, so the page can not be migrated. So try to
1102 * free the metadata, so the page can be freed.
1104 if (!page->mapping) {
1105 VM_BUG_ON_PAGE(PageAnon(page), page);
1106 if (page_has_private(page)) {
1107 try_to_free_buffers(page);
1108 goto out_unlock_both;
1110 } else if (page_mapped(page)) {
1111 /* Establish migration ptes */
1112 VM_BUG_ON_PAGE(PageAnon(page) && !PageKsm(page) && !anon_vma,
1115 TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
1116 page_was_mapped = 1;
1119 if (!page_mapped(page))
1120 rc = move_to_new_page(newpage, page, mode);
1122 if (page_was_mapped)
1123 remove_migration_ptes(page,
1124 rc == MIGRATEPAGE_SUCCESS ? newpage : page, false);
1127 unlock_page(newpage);
1129 /* Drop an anon_vma reference if we took one */
1131 put_anon_vma(anon_vma);
1135 * If migration is successful, decrease refcount of the newpage
1136 * which will not free the page because new page owner increased
1137 * refcounter. As well, if it is LRU page, add the page to LRU
1138 * list in here. Use the old state of the isolated source page to
1139 * determine if we migrated a LRU page. newpage was already unlocked
1140 * and possibly modified by its owner - don't rely on the page
1143 if (rc == MIGRATEPAGE_SUCCESS) {
1144 if (unlikely(!is_lru))
1147 putback_lru_page(newpage);
1154 * gcc 4.7 and 4.8 on arm get an ICEs when inlining unmap_and_move(). Work
1157 #if defined(CONFIG_ARM) && \
1158 defined(GCC_VERSION) && GCC_VERSION < 40900 && GCC_VERSION >= 40700
1159 #define ICE_noinline noinline
1161 #define ICE_noinline
1165 * Obtain the lock on page, remove all ptes and migrate the page
1166 * to the newly allocated page in newpage.
1168 static ICE_noinline int unmap_and_move(new_page_t get_new_page,
1169 free_page_t put_new_page,
1170 unsigned long private, struct page *page,
1171 int force, enum migrate_mode mode,
1172 enum migrate_reason reason)
1174 int rc = MIGRATEPAGE_SUCCESS;
1175 struct page *newpage;
1177 if (!thp_migration_supported() && PageTransHuge(page))
1180 newpage = get_new_page(page, private);
1184 if (page_count(page) == 1) {
1185 /* page was freed from under us. So we are done. */
1186 ClearPageActive(page);
1187 ClearPageUnevictable(page);
1188 if (unlikely(__PageMovable(page))) {
1190 if (!PageMovable(page))
1191 __ClearPageIsolated(page);
1195 put_new_page(newpage, private);
1201 rc = __unmap_and_move(page, newpage, force, mode);
1202 if (rc == MIGRATEPAGE_SUCCESS)
1203 set_page_owner_migrate_reason(newpage, reason);
1206 if (rc != -EAGAIN) {
1208 * A page that has been migrated has all references
1209 * removed and will be freed. A page that has not been
1210 * migrated will have kepts its references and be
1213 list_del(&page->lru);
1216 * Compaction can migrate also non-LRU pages which are
1217 * not accounted to NR_ISOLATED_*. They can be recognized
1220 if (likely(!__PageMovable(page)))
1221 mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON +
1222 page_is_file_cache(page), -hpage_nr_pages(page));
1226 * If migration is successful, releases reference grabbed during
1227 * isolation. Otherwise, restore the page to right list unless
1230 if (rc == MIGRATEPAGE_SUCCESS) {
1232 if (reason == MR_MEMORY_FAILURE) {
1234 * Set PG_HWPoison on just freed page
1235 * intentionally. Although it's rather weird,
1236 * it's how HWPoison flag works at the moment.
1238 if (set_hwpoison_free_buddy_page(page))
1239 num_poisoned_pages_inc();
1242 if (rc != -EAGAIN) {
1243 if (likely(!__PageMovable(page))) {
1244 putback_lru_page(page);
1249 if (PageMovable(page))
1250 putback_movable_page(page);
1252 __ClearPageIsolated(page);
1258 put_new_page(newpage, private);
1267 * Counterpart of unmap_and_move_page() for hugepage migration.
1269 * This function doesn't wait the completion of hugepage I/O
1270 * because there is no race between I/O and migration for hugepage.
1271 * Note that currently hugepage I/O occurs only in direct I/O
1272 * where no lock is held and PG_writeback is irrelevant,
1273 * and writeback status of all subpages are counted in the reference
1274 * count of the head page (i.e. if all subpages of a 2MB hugepage are
1275 * under direct I/O, the reference of the head page is 512 and a bit more.)
1276 * This means that when we try to migrate hugepage whose subpages are
1277 * doing direct I/O, some references remain after try_to_unmap() and
1278 * hugepage migration fails without data corruption.
1280 * There is also no race when direct I/O is issued on the page under migration,
1281 * because then pte is replaced with migration swap entry and direct I/O code
1282 * will wait in the page fault for migration to complete.
1284 static int unmap_and_move_huge_page(new_page_t get_new_page,
1285 free_page_t put_new_page, unsigned long private,
1286 struct page *hpage, int force,
1287 enum migrate_mode mode, int reason)
1290 int page_was_mapped = 0;
1291 struct page *new_hpage;
1292 struct anon_vma *anon_vma = NULL;
1295 * Movability of hugepages depends on architectures and hugepage size.
1296 * This check is necessary because some callers of hugepage migration
1297 * like soft offline and memory hotremove don't walk through page
1298 * tables or check whether the hugepage is pmd-based or not before
1299 * kicking migration.
1301 if (!hugepage_migration_supported(page_hstate(hpage))) {
1302 putback_active_hugepage(hpage);
1306 new_hpage = get_new_page(hpage, private);
1310 if (!trylock_page(hpage)) {
1315 case MIGRATE_SYNC_NO_COPY:
1324 * Check for pages which are in the process of being freed. Without
1325 * page_mapping() set, hugetlbfs specific move page routine will not
1326 * be called and we could leak usage counts for subpools.
1328 if (page_private(hpage) && !page_mapping(hpage)) {
1333 if (PageAnon(hpage))
1334 anon_vma = page_get_anon_vma(hpage);
1336 if (unlikely(!trylock_page(new_hpage)))
1339 if (page_mapped(hpage)) {
1341 TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
1342 page_was_mapped = 1;
1345 if (!page_mapped(hpage))
1346 rc = move_to_new_page(new_hpage, hpage, mode);
1348 if (page_was_mapped)
1349 remove_migration_ptes(hpage,
1350 rc == MIGRATEPAGE_SUCCESS ? new_hpage : hpage, false);
1352 unlock_page(new_hpage);
1356 put_anon_vma(anon_vma);
1358 if (rc == MIGRATEPAGE_SUCCESS) {
1359 move_hugetlb_state(hpage, new_hpage, reason);
1360 put_new_page = NULL;
1367 putback_active_hugepage(hpage);
1370 * If migration was not successful and there's a freeing callback, use
1371 * it. Otherwise, put_page() will drop the reference grabbed during
1375 put_new_page(new_hpage, private);
1377 putback_active_hugepage(new_hpage);
1383 * migrate_pages - migrate the pages specified in a list, to the free pages
1384 * supplied as the target for the page migration
1386 * @from: The list of pages to be migrated.
1387 * @get_new_page: The function used to allocate free pages to be used
1388 * as the target of the page migration.
1389 * @put_new_page: The function used to free target pages if migration
1390 * fails, or NULL if no special handling is necessary.
1391 * @private: Private data to be passed on to get_new_page()
1392 * @mode: The migration mode that specifies the constraints for
1393 * page migration, if any.
1394 * @reason: The reason for page migration.
1396 * The function returns after 10 attempts or if no pages are movable any more
1397 * because the list has become empty or no retryable pages exist any more.
1398 * The caller should call putback_movable_pages() to return pages to the LRU
1399 * or free list only if ret != 0.
1401 * Returns the number of pages that were not migrated, or an error code.
1403 int migrate_pages(struct list_head *from, new_page_t get_new_page,
1404 free_page_t put_new_page, unsigned long private,
1405 enum migrate_mode mode, int reason)
1409 int nr_succeeded = 0;
1413 int swapwrite = current->flags & PF_SWAPWRITE;
1417 current->flags |= PF_SWAPWRITE;
1419 for(pass = 0; pass < 10 && retry; pass++) {
1422 list_for_each_entry_safe(page, page2, from, lru) {
1427 rc = unmap_and_move_huge_page(get_new_page,
1428 put_new_page, private, page,
1429 pass > 2, mode, reason);
1431 rc = unmap_and_move(get_new_page, put_new_page,
1432 private, page, pass > 2, mode,
1438 * THP migration might be unsupported or the
1439 * allocation could've failed so we should
1440 * retry on the same page with the THP split
1443 * Head page is retried immediately and tail
1444 * pages are added to the tail of the list so
1445 * we encounter them after the rest of the list
1448 if (PageTransHuge(page) && !PageHuge(page)) {
1450 rc = split_huge_page_to_list(page, from);
1453 list_safe_reset_next(page, page2, lru);
1462 case MIGRATEPAGE_SUCCESS:
1467 * Permanent failure (-EBUSY, -ENOSYS, etc.):
1468 * unlike -EAGAIN case, the failed page is
1469 * removed from migration page list and not
1470 * retried in the next outer loop.
1481 count_vm_events(PGMIGRATE_SUCCESS, nr_succeeded);
1483 count_vm_events(PGMIGRATE_FAIL, nr_failed);
1484 trace_mm_migrate_pages(nr_succeeded, nr_failed, mode, reason);
1487 current->flags &= ~PF_SWAPWRITE;
1494 static int store_status(int __user *status, int start, int value, int nr)
1497 if (put_user(value, status + start))
1505 static int do_move_pages_to_node(struct mm_struct *mm,
1506 struct list_head *pagelist, int node)
1510 if (list_empty(pagelist))
1513 err = migrate_pages(pagelist, alloc_new_node_page, NULL, node,
1514 MIGRATE_SYNC, MR_SYSCALL);
1516 putback_movable_pages(pagelist);
1521 * Resolves the given address to a struct page, isolates it from the LRU and
1522 * puts it to the given pagelist.
1524 * errno - if the page cannot be found/isolated
1525 * 0 - when it doesn't have to be migrated because it is already on the
1527 * 1 - when it has been queued
1529 static int add_page_for_migration(struct mm_struct *mm, unsigned long addr,
1530 int node, struct list_head *pagelist, bool migrate_all)
1532 struct vm_area_struct *vma;
1534 unsigned int follflags;
1537 down_read(&mm->mmap_sem);
1539 vma = find_vma(mm, addr);
1540 if (!vma || addr < vma->vm_start || !vma_migratable(vma))
1543 /* FOLL_DUMP to ignore special (like zero) pages */
1544 follflags = FOLL_GET | FOLL_DUMP;
1545 page = follow_page(vma, addr, follflags);
1547 err = PTR_ERR(page);
1556 if (page_to_nid(page) == node)
1560 if (page_mapcount(page) > 1 && !migrate_all)
1563 if (PageHuge(page)) {
1564 if (PageHead(page)) {
1565 isolate_huge_page(page, pagelist);
1571 head = compound_head(page);
1572 err = isolate_lru_page(head);
1577 list_add_tail(&head->lru, pagelist);
1578 mod_node_page_state(page_pgdat(head),
1579 NR_ISOLATED_ANON + page_is_file_cache(head),
1580 hpage_nr_pages(head));
1584 * Either remove the duplicate refcount from
1585 * isolate_lru_page() or drop the page ref if it was
1590 up_read(&mm->mmap_sem);
1595 * Migrate an array of page address onto an array of nodes and fill
1596 * the corresponding array of status.
1598 static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes,
1599 unsigned long nr_pages,
1600 const void __user * __user *pages,
1601 const int __user *nodes,
1602 int __user *status, int flags)
1604 int current_node = NUMA_NO_NODE;
1605 LIST_HEAD(pagelist);
1611 for (i = start = 0; i < nr_pages; i++) {
1612 const void __user *p;
1617 if (get_user(p, pages + i))
1619 if (get_user(node, nodes + i))
1621 addr = (unsigned long)p;
1624 if (node < 0 || node >= MAX_NUMNODES)
1626 if (!node_state(node, N_MEMORY))
1630 if (!node_isset(node, task_nodes))
1633 if (current_node == NUMA_NO_NODE) {
1634 current_node = node;
1636 } else if (node != current_node) {
1637 err = do_move_pages_to_node(mm, &pagelist, current_node);
1640 * Positive err means the number of failed
1641 * pages to migrate. Since we are going to
1642 * abort and return the number of non-migrated
1643 * pages, so need to incude the rest of the
1644 * nr_pages that have not been attempted as
1648 err += nr_pages - i - 1;
1651 err = store_status(status, start, current_node, i - start);
1655 current_node = node;
1659 * Errors in the page lookup or isolation are not fatal and we simply
1660 * report them via status
1662 err = add_page_for_migration(mm, addr, current_node,
1663 &pagelist, flags & MPOL_MF_MOVE_ALL);
1666 /* The page is already on the target node */
1667 err = store_status(status, i, current_node, 1);
1671 } else if (err > 0) {
1672 /* The page is successfully queued for migration */
1676 err = store_status(status, i, err, 1);
1680 err = do_move_pages_to_node(mm, &pagelist, current_node);
1683 err += nr_pages - i - 1;
1687 err = store_status(status, start, current_node, i - start);
1691 current_node = NUMA_NO_NODE;
1694 if (list_empty(&pagelist))
1697 /* Make sure we do not overwrite the existing error */
1698 err1 = do_move_pages_to_node(mm, &pagelist, current_node);
1700 * Don't have to report non-attempted pages here since:
1701 * - If the above loop is done gracefully all pages have been
1703 * - If the above loop is aborted it means a fatal error
1704 * happened, should return ret.
1707 err1 = store_status(status, start, current_node, i - start);
1715 * Determine the nodes of an array of pages and store it in an array of status.
1717 static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages,
1718 const void __user **pages, int *status)
1722 down_read(&mm->mmap_sem);
1724 for (i = 0; i < nr_pages; i++) {
1725 unsigned long addr = (unsigned long)(*pages);
1726 struct vm_area_struct *vma;
1730 vma = find_vma(mm, addr);
1731 if (!vma || addr < vma->vm_start)
1734 /* FOLL_DUMP to ignore special (like zero) pages */
1735 page = follow_page(vma, addr, FOLL_DUMP);
1737 err = PTR_ERR(page);
1741 err = page ? page_to_nid(page) : -ENOENT;
1749 up_read(&mm->mmap_sem);
1753 * Determine the nodes of a user array of pages and store it in
1754 * a user array of status.
1756 static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages,
1757 const void __user * __user *pages,
1760 #define DO_PAGES_STAT_CHUNK_NR 16
1761 const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR];
1762 int chunk_status[DO_PAGES_STAT_CHUNK_NR];
1765 unsigned long chunk_nr;
1767 chunk_nr = nr_pages;
1768 if (chunk_nr > DO_PAGES_STAT_CHUNK_NR)
1769 chunk_nr = DO_PAGES_STAT_CHUNK_NR;
1771 if (copy_from_user(chunk_pages, pages, chunk_nr * sizeof(*chunk_pages)))
1774 do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status);
1776 if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status)))
1781 nr_pages -= chunk_nr;
1783 return nr_pages ? -EFAULT : 0;
1787 * Move a list of pages in the address space of the currently executing
1790 static int kernel_move_pages(pid_t pid, unsigned long nr_pages,
1791 const void __user * __user *pages,
1792 const int __user *nodes,
1793 int __user *status, int flags)
1795 struct task_struct *task;
1796 struct mm_struct *mm;
1798 nodemask_t task_nodes;
1801 if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
1804 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1807 /* Find the mm_struct */
1809 task = pid ? find_task_by_vpid(pid) : current;
1814 get_task_struct(task);
1817 * Check if this process has the right to modify the specified
1818 * process. Use the regular "ptrace_may_access()" checks.
1820 if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) {
1827 err = security_task_movememory(task);
1831 task_nodes = cpuset_mems_allowed(task);
1832 mm = get_task_mm(task);
1833 put_task_struct(task);
1839 err = do_pages_move(mm, task_nodes, nr_pages, pages,
1840 nodes, status, flags);
1842 err = do_pages_stat(mm, nr_pages, pages, status);
1848 put_task_struct(task);
1852 SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages,
1853 const void __user * __user *, pages,
1854 const int __user *, nodes,
1855 int __user *, status, int, flags)
1857 return kernel_move_pages(pid, nr_pages, pages, nodes, status, flags);
1860 #ifdef CONFIG_COMPAT
1861 COMPAT_SYSCALL_DEFINE6(move_pages, pid_t, pid, compat_ulong_t, nr_pages,
1862 compat_uptr_t __user *, pages32,
1863 const int __user *, nodes,
1864 int __user *, status,
1867 const void __user * __user *pages;
1870 pages = compat_alloc_user_space(nr_pages * sizeof(void *));
1871 for (i = 0; i < nr_pages; i++) {
1874 if (get_user(p, pages32 + i) ||
1875 put_user(compat_ptr(p), pages + i))
1878 return kernel_move_pages(pid, nr_pages, pages, nodes, status, flags);
1880 #endif /* CONFIG_COMPAT */
1882 #ifdef CONFIG_NUMA_BALANCING
1884 * Returns true if this is a safe migration target node for misplaced NUMA
1885 * pages. Currently it only checks the watermarks which crude
1887 static bool migrate_balanced_pgdat(struct pglist_data *pgdat,
1888 unsigned long nr_migrate_pages)
1892 for (z = pgdat->nr_zones - 1; z >= 0; z--) {
1893 struct zone *zone = pgdat->node_zones + z;
1895 if (!populated_zone(zone))
1898 /* Avoid waking kswapd by allocating pages_to_migrate pages. */
1899 if (!zone_watermark_ok(zone, 0,
1900 high_wmark_pages(zone) +
1909 static struct page *alloc_misplaced_dst_page(struct page *page,
1912 int nid = (int) data;
1913 struct page *newpage;
1915 newpage = __alloc_pages_node(nid,
1916 (GFP_HIGHUSER_MOVABLE |
1917 __GFP_THISNODE | __GFP_NOMEMALLOC |
1918 __GFP_NORETRY | __GFP_NOWARN) &
1924 static int numamigrate_isolate_page(pg_data_t *pgdat, struct page *page)
1928 VM_BUG_ON_PAGE(compound_order(page) && !PageTransHuge(page), page);
1930 /* Avoid migrating to a node that is nearly full */
1931 if (!migrate_balanced_pgdat(pgdat, 1UL << compound_order(page)))
1934 if (isolate_lru_page(page))
1938 * migrate_misplaced_transhuge_page() skips page migration's usual
1939 * check on page_count(), so we must do it here, now that the page
1940 * has been isolated: a GUP pin, or any other pin, prevents migration.
1941 * The expected page count is 3: 1 for page's mapcount and 1 for the
1942 * caller's pin and 1 for the reference taken by isolate_lru_page().
1944 if (PageTransHuge(page) && page_count(page) != 3) {
1945 putback_lru_page(page);
1949 page_lru = page_is_file_cache(page);
1950 mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON + page_lru,
1951 hpage_nr_pages(page));
1954 * Isolating the page has taken another reference, so the
1955 * caller's reference can be safely dropped without the page
1956 * disappearing underneath us during migration.
1962 bool pmd_trans_migrating(pmd_t pmd)
1964 struct page *page = pmd_page(pmd);
1965 return PageLocked(page);
1969 * Attempt to migrate a misplaced page to the specified destination
1970 * node. Caller is expected to have an elevated reference count on
1971 * the page that will be dropped by this function before returning.
1973 int migrate_misplaced_page(struct page *page, struct vm_area_struct *vma,
1976 pg_data_t *pgdat = NODE_DATA(node);
1979 LIST_HEAD(migratepages);
1982 * Don't migrate file pages that are mapped in multiple processes
1983 * with execute permissions as they are probably shared libraries.
1985 if (page_mapcount(page) != 1 && page_is_file_cache(page) &&
1986 (vma->vm_flags & VM_EXEC))
1990 * Also do not migrate dirty pages as not all filesystems can move
1991 * dirty pages in MIGRATE_ASYNC mode which is a waste of cycles.
1993 if (page_is_file_cache(page) && PageDirty(page))
1996 isolated = numamigrate_isolate_page(pgdat, page);
2000 list_add(&page->lru, &migratepages);
2001 nr_remaining = migrate_pages(&migratepages, alloc_misplaced_dst_page,
2002 NULL, node, MIGRATE_ASYNC,
2005 if (!list_empty(&migratepages)) {
2006 list_del(&page->lru);
2007 dec_node_page_state(page, NR_ISOLATED_ANON +
2008 page_is_file_cache(page));
2009 putback_lru_page(page);
2013 count_vm_numa_event(NUMA_PAGE_MIGRATE);
2014 BUG_ON(!list_empty(&migratepages));
2021 #endif /* CONFIG_NUMA_BALANCING */
2023 #if defined(CONFIG_NUMA_BALANCING) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
2025 * Migrates a THP to a given target node. page must be locked and is unlocked
2028 int migrate_misplaced_transhuge_page(struct mm_struct *mm,
2029 struct vm_area_struct *vma,
2030 pmd_t *pmd, pmd_t entry,
2031 unsigned long address,
2032 struct page *page, int node)
2035 pg_data_t *pgdat = NODE_DATA(node);
2037 struct page *new_page = NULL;
2038 int page_lru = page_is_file_cache(page);
2039 unsigned long mmun_start = address & HPAGE_PMD_MASK;
2040 unsigned long mmun_end = mmun_start + HPAGE_PMD_SIZE;
2042 new_page = alloc_pages_node(node,
2043 (GFP_TRANSHUGE_LIGHT | __GFP_THISNODE),
2047 prep_transhuge_page(new_page);
2049 isolated = numamigrate_isolate_page(pgdat, page);
2055 /* Prepare a page as a migration target */
2056 __SetPageLocked(new_page);
2057 if (PageSwapBacked(page))
2058 __SetPageSwapBacked(new_page);
2060 /* anon mapping, we can simply copy page->mapping to the new page: */
2061 new_page->mapping = page->mapping;
2062 new_page->index = page->index;
2063 migrate_page_copy(new_page, page);
2064 WARN_ON(PageLRU(new_page));
2066 /* Recheck the target PMD */
2067 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2068 ptl = pmd_lock(mm, pmd);
2069 if (unlikely(!pmd_same(*pmd, entry) || !page_ref_freeze(page, 2))) {
2071 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2073 /* Reverse changes made by migrate_page_copy() */
2074 if (TestClearPageActive(new_page))
2075 SetPageActive(page);
2076 if (TestClearPageUnevictable(new_page))
2077 SetPageUnevictable(page);
2079 unlock_page(new_page);
2080 put_page(new_page); /* Free it */
2082 /* Retake the callers reference and putback on LRU */
2084 putback_lru_page(page);
2085 mod_node_page_state(page_pgdat(page),
2086 NR_ISOLATED_ANON + page_lru, -HPAGE_PMD_NR);
2091 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
2092 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
2095 * Overwrite the old entry under pagetable lock and establish
2096 * the new PTE. Any parallel GUP will either observe the old
2097 * page blocking on the page lock, block on the page table
2098 * lock or observe the new page. The SetPageUptodate on the
2099 * new page and page_add_new_anon_rmap guarantee the copy is
2100 * visible before the pagetable update.
2102 flush_cache_range(vma, mmun_start, mmun_end);
2103 page_add_anon_rmap(new_page, vma, mmun_start, true);
2105 * At this point the pmd is numa/protnone (i.e. non present) and the TLB
2106 * has already been flushed globally. So no TLB can be currently
2107 * caching this non present pmd mapping. There's no need to clear the
2108 * pmd before doing set_pmd_at(), nor to flush the TLB after
2109 * set_pmd_at(). Clearing the pmd here would introduce a race
2110 * condition against MADV_DONTNEED, because MADV_DONTNEED only holds the
2111 * mmap_sem for reading. If the pmd is set to NULL at any given time,
2112 * MADV_DONTNEED won't wait on the pmd lock and it'll skip clearing this
2115 set_pmd_at(mm, mmun_start, pmd, entry);
2116 update_mmu_cache_pmd(vma, address, &entry);
2118 page_ref_unfreeze(page, 2);
2119 mlock_migrate_page(new_page, page);
2120 page_remove_rmap(page, true);
2121 set_page_owner_migrate_reason(new_page, MR_NUMA_MISPLACED);
2125 * No need to double call mmu_notifier->invalidate_range() callback as
2126 * the above pmdp_huge_clear_flush_notify() did already call it.
2128 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2130 /* Take an "isolate" reference and put new page on the LRU. */
2132 putback_lru_page(new_page);
2134 unlock_page(new_page);
2136 put_page(page); /* Drop the rmap reference */
2137 put_page(page); /* Drop the LRU isolation reference */
2139 count_vm_events(PGMIGRATE_SUCCESS, HPAGE_PMD_NR);
2140 count_vm_numa_events(NUMA_PAGE_MIGRATE, HPAGE_PMD_NR);
2142 mod_node_page_state(page_pgdat(page),
2143 NR_ISOLATED_ANON + page_lru,
2148 count_vm_events(PGMIGRATE_FAIL, HPAGE_PMD_NR);
2149 ptl = pmd_lock(mm, pmd);
2150 if (pmd_same(*pmd, entry)) {
2151 entry = pmd_modify(entry, vma->vm_page_prot);
2152 set_pmd_at(mm, mmun_start, pmd, entry);
2153 update_mmu_cache_pmd(vma, address, &entry);
2162 #endif /* CONFIG_NUMA_BALANCING */
2164 #endif /* CONFIG_NUMA */
2166 #if defined(CONFIG_MIGRATE_VMA_HELPER)
2167 struct migrate_vma {
2168 struct vm_area_struct *vma;
2171 unsigned long cpages;
2172 unsigned long npages;
2173 unsigned long start;
2177 static int migrate_vma_collect_hole(unsigned long start,
2179 struct mm_walk *walk)
2181 struct migrate_vma *migrate = walk->private;
2184 for (addr = start & PAGE_MASK; addr < end; addr += PAGE_SIZE) {
2185 migrate->src[migrate->npages] = MIGRATE_PFN_MIGRATE;
2186 migrate->dst[migrate->npages] = 0;
2194 static int migrate_vma_collect_skip(unsigned long start,
2196 struct mm_walk *walk)
2198 struct migrate_vma *migrate = walk->private;
2201 for (addr = start & PAGE_MASK; addr < end; addr += PAGE_SIZE) {
2202 migrate->dst[migrate->npages] = 0;
2203 migrate->src[migrate->npages++] = 0;
2209 static int migrate_vma_collect_pmd(pmd_t *pmdp,
2210 unsigned long start,
2212 struct mm_walk *walk)
2214 struct migrate_vma *migrate = walk->private;
2215 struct vm_area_struct *vma = walk->vma;
2216 struct mm_struct *mm = vma->vm_mm;
2217 unsigned long addr = start, unmapped = 0;
2222 if (pmd_none(*pmdp))
2223 return migrate_vma_collect_hole(start, end, walk);
2225 if (pmd_trans_huge(*pmdp)) {
2228 ptl = pmd_lock(mm, pmdp);
2229 if (unlikely(!pmd_trans_huge(*pmdp))) {
2234 page = pmd_page(*pmdp);
2235 if (is_huge_zero_page(page)) {
2237 split_huge_pmd(vma, pmdp, addr);
2238 if (pmd_trans_unstable(pmdp))
2239 return migrate_vma_collect_skip(start, end,
2246 if (unlikely(!trylock_page(page)))
2247 return migrate_vma_collect_skip(start, end,
2249 ret = split_huge_page(page);
2253 return migrate_vma_collect_skip(start, end,
2255 if (pmd_none(*pmdp))
2256 return migrate_vma_collect_hole(start, end,
2261 if (unlikely(pmd_bad(*pmdp)))
2262 return migrate_vma_collect_skip(start, end, walk);
2264 ptep = pte_offset_map_lock(mm, pmdp, addr, &ptl);
2265 arch_enter_lazy_mmu_mode();
2267 for (; addr < end; addr += PAGE_SIZE, ptep++) {
2268 unsigned long mpfn, pfn;
2276 if (pte_none(pte)) {
2277 mpfn = MIGRATE_PFN_MIGRATE;
2283 if (!pte_present(pte)) {
2287 * Only care about unaddressable device page special
2288 * page table entry. Other special swap entries are not
2289 * migratable, and we ignore regular swapped page.
2291 entry = pte_to_swp_entry(pte);
2292 if (!is_device_private_entry(entry))
2295 page = device_private_entry_to_page(entry);
2296 mpfn = migrate_pfn(page_to_pfn(page))|
2297 MIGRATE_PFN_DEVICE | MIGRATE_PFN_MIGRATE;
2298 if (is_write_device_private_entry(entry))
2299 mpfn |= MIGRATE_PFN_WRITE;
2301 if (is_zero_pfn(pfn)) {
2302 mpfn = MIGRATE_PFN_MIGRATE;
2307 page = _vm_normal_page(migrate->vma, addr, pte, true);
2308 mpfn = migrate_pfn(pfn) | MIGRATE_PFN_MIGRATE;
2309 mpfn |= pte_write(pte) ? MIGRATE_PFN_WRITE : 0;
2312 /* FIXME support THP */
2313 if (!page || !page->mapping || PageTransCompound(page)) {
2317 pfn = page_to_pfn(page);
2320 * By getting a reference on the page we pin it and that blocks
2321 * any kind of migration. Side effect is that it "freezes" the
2324 * We drop this reference after isolating the page from the lru
2325 * for non device page (device page are not on the lru and thus
2326 * can't be dropped from it).
2332 * Optimize for the common case where page is only mapped once
2333 * in one process. If we can lock the page, then we can safely
2334 * set up a special migration page table entry now.
2336 if (trylock_page(page)) {
2339 mpfn |= MIGRATE_PFN_LOCKED;
2340 ptep_get_and_clear(mm, addr, ptep);
2342 /* Setup special migration page table entry */
2343 entry = make_migration_entry(page, mpfn &
2345 swp_pte = swp_entry_to_pte(entry);
2346 if (pte_soft_dirty(pte))
2347 swp_pte = pte_swp_mksoft_dirty(swp_pte);
2348 set_pte_at(mm, addr, ptep, swp_pte);
2351 * This is like regular unmap: we remove the rmap and
2352 * drop page refcount. Page won't be freed, as we took
2353 * a reference just above.
2355 page_remove_rmap(page, false);
2358 if (pte_present(pte))
2363 migrate->dst[migrate->npages] = 0;
2364 migrate->src[migrate->npages++] = mpfn;
2367 /* Only flush the TLB if we actually modified any entries */
2369 flush_tlb_range(walk->vma, start, end);
2371 arch_leave_lazy_mmu_mode();
2372 pte_unmap_unlock(ptep - 1, ptl);
2378 * migrate_vma_collect() - collect pages over a range of virtual addresses
2379 * @migrate: migrate struct containing all migration information
2381 * This will walk the CPU page table. For each virtual address backed by a
2382 * valid page, it updates the src array and takes a reference on the page, in
2383 * order to pin the page until we lock it and unmap it.
2385 static void migrate_vma_collect(struct migrate_vma *migrate)
2387 struct mm_walk mm_walk = {
2388 .pmd_entry = migrate_vma_collect_pmd,
2389 .pte_hole = migrate_vma_collect_hole,
2390 .vma = migrate->vma,
2391 .mm = migrate->vma->vm_mm,
2395 mmu_notifier_invalidate_range_start(mm_walk.mm,
2398 walk_page_range(migrate->start, migrate->end, &mm_walk);
2399 mmu_notifier_invalidate_range_end(mm_walk.mm,
2403 migrate->end = migrate->start + (migrate->npages << PAGE_SHIFT);
2407 * migrate_vma_check_page() - check if page is pinned or not
2408 * @page: struct page to check
2410 * Pinned pages cannot be migrated. This is the same test as in
2411 * migrate_page_move_mapping(), except that here we allow migration of a
2414 static bool migrate_vma_check_page(struct page *page)
2417 * One extra ref because caller holds an extra reference, either from
2418 * isolate_lru_page() for a regular page, or migrate_vma_collect() for
2424 * FIXME support THP (transparent huge page), it is bit more complex to
2425 * check them than regular pages, because they can be mapped with a pmd
2426 * or with a pte (split pte mapping).
2428 if (PageCompound(page))
2431 /* Page from ZONE_DEVICE have one extra reference */
2432 if (is_zone_device_page(page)) {
2434 * Private page can never be pin as they have no valid pte and
2435 * GUP will fail for those. Yet if there is a pending migration
2436 * a thread might try to wait on the pte migration entry and
2437 * will bump the page reference count. Sadly there is no way to
2438 * differentiate a regular pin from migration wait. Hence to
2439 * avoid 2 racing thread trying to migrate back to CPU to enter
2440 * infinite loop (one stoping migration because the other is
2441 * waiting on pte migration entry). We always return true here.
2443 * FIXME proper solution is to rework migration_entry_wait() so
2444 * it does not need to take a reference on page.
2446 if (is_device_private_page(page))
2450 * Only allow device public page to be migrated and account for
2451 * the extra reference count imply by ZONE_DEVICE pages.
2453 if (!is_device_public_page(page))
2458 /* For file back page */
2459 if (page_mapping(page))
2460 extra += 1 + page_has_private(page);
2462 if ((page_count(page) - extra) > page_mapcount(page))
2469 * migrate_vma_prepare() - lock pages and isolate them from the lru
2470 * @migrate: migrate struct containing all migration information
2472 * This locks pages that have been collected by migrate_vma_collect(). Once each
2473 * page is locked it is isolated from the lru (for non-device pages). Finally,
2474 * the ref taken by migrate_vma_collect() is dropped, as locked pages cannot be
2475 * migrated by concurrent kernel threads.
2477 static void migrate_vma_prepare(struct migrate_vma *migrate)
2479 const unsigned long npages = migrate->npages;
2480 const unsigned long start = migrate->start;
2481 unsigned long addr, i, restore = 0;
2482 bool allow_drain = true;
2486 for (i = 0; (i < npages) && migrate->cpages; i++) {
2487 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2493 if (!(migrate->src[i] & MIGRATE_PFN_LOCKED)) {
2495 * Because we are migrating several pages there can be
2496 * a deadlock between 2 concurrent migration where each
2497 * are waiting on each other page lock.
2499 * Make migrate_vma() a best effort thing and backoff
2500 * for any page we can not lock right away.
2502 if (!trylock_page(page)) {
2503 migrate->src[i] = 0;
2509 migrate->src[i] |= MIGRATE_PFN_LOCKED;
2512 /* ZONE_DEVICE pages are not on LRU */
2513 if (!is_zone_device_page(page)) {
2514 if (!PageLRU(page) && allow_drain) {
2515 /* Drain CPU's pagevec */
2516 lru_add_drain_all();
2517 allow_drain = false;
2520 if (isolate_lru_page(page)) {
2522 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2526 migrate->src[i] = 0;
2534 /* Drop the reference we took in collect */
2538 if (!migrate_vma_check_page(page)) {
2540 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2544 if (!is_zone_device_page(page)) {
2546 putback_lru_page(page);
2549 migrate->src[i] = 0;
2553 if (!is_zone_device_page(page))
2554 putback_lru_page(page);
2561 for (i = 0, addr = start; i < npages && restore; i++, addr += PAGE_SIZE) {
2562 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2564 if (!page || (migrate->src[i] & MIGRATE_PFN_MIGRATE))
2567 remove_migration_pte(page, migrate->vma, addr, page);
2569 migrate->src[i] = 0;
2577 * migrate_vma_unmap() - replace page mapping with special migration pte entry
2578 * @migrate: migrate struct containing all migration information
2580 * Replace page mapping (CPU page table pte) with a special migration pte entry
2581 * and check again if it has been pinned. Pinned pages are restored because we
2582 * cannot migrate them.
2584 * This is the last step before we call the device driver callback to allocate
2585 * destination memory and copy contents of original page over to new page.
2587 static void migrate_vma_unmap(struct migrate_vma *migrate)
2589 int flags = TTU_MIGRATION | TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS;
2590 const unsigned long npages = migrate->npages;
2591 const unsigned long start = migrate->start;
2592 unsigned long addr, i, restore = 0;
2594 for (i = 0; i < npages; i++) {
2595 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2597 if (!page || !(migrate->src[i] & MIGRATE_PFN_MIGRATE))
2600 if (page_mapped(page)) {
2601 try_to_unmap(page, flags);
2602 if (page_mapped(page))
2606 if (migrate_vma_check_page(page))
2610 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2615 for (addr = start, i = 0; i < npages && restore; addr += PAGE_SIZE, i++) {
2616 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2618 if (!page || (migrate->src[i] & MIGRATE_PFN_MIGRATE))
2621 remove_migration_ptes(page, page, false);
2623 migrate->src[i] = 0;
2627 if (is_zone_device_page(page))
2630 putback_lru_page(page);
2634 static void migrate_vma_insert_page(struct migrate_vma *migrate,
2640 struct vm_area_struct *vma = migrate->vma;
2641 struct mm_struct *mm = vma->vm_mm;
2642 struct mem_cgroup *memcg;
2652 /* Only allow populating anonymous memory */
2653 if (!vma_is_anonymous(vma))
2656 pgdp = pgd_offset(mm, addr);
2657 p4dp = p4d_alloc(mm, pgdp, addr);
2660 pudp = pud_alloc(mm, p4dp, addr);
2663 pmdp = pmd_alloc(mm, pudp, addr);
2667 if (pmd_trans_huge(*pmdp) || pmd_devmap(*pmdp))
2671 * Use pte_alloc() instead of pte_alloc_map(). We can't run
2672 * pte_offset_map() on pmds where a huge pmd might be created
2673 * from a different thread.
2675 * pte_alloc_map() is safe to use under down_write(mmap_sem) or when
2676 * parallel threads are excluded by other means.
2678 * Here we only have down_read(mmap_sem).
2680 if (pte_alloc(mm, pmdp, addr))
2683 /* See the comment in pte_alloc_one_map() */
2684 if (unlikely(pmd_trans_unstable(pmdp)))
2687 if (unlikely(anon_vma_prepare(vma)))
2689 if (mem_cgroup_try_charge(page, vma->vm_mm, GFP_KERNEL, &memcg, false))
2693 * The memory barrier inside __SetPageUptodate makes sure that
2694 * preceding stores to the page contents become visible before
2695 * the set_pte_at() write.
2697 __SetPageUptodate(page);
2699 if (is_zone_device_page(page)) {
2700 if (is_device_private_page(page)) {
2701 swp_entry_t swp_entry;
2703 swp_entry = make_device_private_entry(page, vma->vm_flags & VM_WRITE);
2704 entry = swp_entry_to_pte(swp_entry);
2705 } else if (is_device_public_page(page)) {
2706 entry = pte_mkold(mk_pte(page, READ_ONCE(vma->vm_page_prot)));
2707 if (vma->vm_flags & VM_WRITE)
2708 entry = pte_mkwrite(pte_mkdirty(entry));
2709 entry = pte_mkdevmap(entry);
2712 entry = mk_pte(page, vma->vm_page_prot);
2713 if (vma->vm_flags & VM_WRITE)
2714 entry = pte_mkwrite(pte_mkdirty(entry));
2717 ptep = pte_offset_map_lock(mm, pmdp, addr, &ptl);
2719 if (pte_present(*ptep)) {
2720 unsigned long pfn = pte_pfn(*ptep);
2722 if (!is_zero_pfn(pfn)) {
2723 pte_unmap_unlock(ptep, ptl);
2724 mem_cgroup_cancel_charge(page, memcg, false);
2728 } else if (!pte_none(*ptep)) {
2729 pte_unmap_unlock(ptep, ptl);
2730 mem_cgroup_cancel_charge(page, memcg, false);
2735 * Check for usefaultfd but do not deliver the fault. Instead,
2738 if (userfaultfd_missing(vma)) {
2739 pte_unmap_unlock(ptep, ptl);
2740 mem_cgroup_cancel_charge(page, memcg, false);
2744 inc_mm_counter(mm, MM_ANONPAGES);
2745 page_add_new_anon_rmap(page, vma, addr, false);
2746 mem_cgroup_commit_charge(page, memcg, false, false);
2747 if (!is_zone_device_page(page))
2748 lru_cache_add_active_or_unevictable(page, vma);
2752 flush_cache_page(vma, addr, pte_pfn(*ptep));
2753 ptep_clear_flush_notify(vma, addr, ptep);
2754 set_pte_at_notify(mm, addr, ptep, entry);
2755 update_mmu_cache(vma, addr, ptep);
2757 /* No need to invalidate - it was non-present before */
2758 set_pte_at(mm, addr, ptep, entry);
2759 update_mmu_cache(vma, addr, ptep);
2762 pte_unmap_unlock(ptep, ptl);
2763 *src = MIGRATE_PFN_MIGRATE;
2767 *src &= ~MIGRATE_PFN_MIGRATE;
2771 * migrate_vma_pages() - migrate meta-data from src page to dst page
2772 * @migrate: migrate struct containing all migration information
2774 * This migrates struct page meta-data from source struct page to destination
2775 * struct page. This effectively finishes the migration from source page to the
2778 static void migrate_vma_pages(struct migrate_vma *migrate)
2780 const unsigned long npages = migrate->npages;
2781 const unsigned long start = migrate->start;
2782 struct vm_area_struct *vma = migrate->vma;
2783 struct mm_struct *mm = vma->vm_mm;
2784 unsigned long addr, i, mmu_start;
2785 bool notified = false;
2787 for (i = 0, addr = start; i < npages; addr += PAGE_SIZE, i++) {
2788 struct page *newpage = migrate_pfn_to_page(migrate->dst[i]);
2789 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2790 struct address_space *mapping;
2794 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2799 if (!(migrate->src[i] & MIGRATE_PFN_MIGRATE)) {
2805 mmu_notifier_invalidate_range_start(mm,
2809 migrate_vma_insert_page(migrate, addr, newpage,
2815 mapping = page_mapping(page);
2817 if (is_zone_device_page(newpage)) {
2818 if (is_device_private_page(newpage)) {
2820 * For now only support private anonymous when
2821 * migrating to un-addressable device memory.
2824 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2827 } else if (!is_device_public_page(newpage)) {
2829 * Other types of ZONE_DEVICE page are not
2832 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2837 r = migrate_page(mapping, newpage, page, MIGRATE_SYNC_NO_COPY);
2838 if (r != MIGRATEPAGE_SUCCESS)
2839 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2843 * No need to double call mmu_notifier->invalidate_range() callback as
2844 * the above ptep_clear_flush_notify() inside migrate_vma_insert_page()
2845 * did already call it.
2848 mmu_notifier_invalidate_range_only_end(mm, mmu_start,
2853 * migrate_vma_finalize() - restore CPU page table entry
2854 * @migrate: migrate struct containing all migration information
2856 * This replaces the special migration pte entry with either a mapping to the
2857 * new page if migration was successful for that page, or to the original page
2860 * This also unlocks the pages and puts them back on the lru, or drops the extra
2861 * refcount, for device pages.
2863 static void migrate_vma_finalize(struct migrate_vma *migrate)
2865 const unsigned long npages = migrate->npages;
2868 for (i = 0; i < npages; i++) {
2869 struct page *newpage = migrate_pfn_to_page(migrate->dst[i]);
2870 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2874 unlock_page(newpage);
2880 if (!(migrate->src[i] & MIGRATE_PFN_MIGRATE) || !newpage) {
2882 unlock_page(newpage);
2888 remove_migration_ptes(page, newpage, false);
2892 if (is_zone_device_page(page))
2895 putback_lru_page(page);
2897 if (newpage != page) {
2898 unlock_page(newpage);
2899 if (is_zone_device_page(newpage))
2902 putback_lru_page(newpage);
2908 * migrate_vma() - migrate a range of memory inside vma
2910 * @ops: migration callback for allocating destination memory and copying
2911 * @vma: virtual memory area containing the range to be migrated
2912 * @start: start address of the range to migrate (inclusive)
2913 * @end: end address of the range to migrate (exclusive)
2914 * @src: array of hmm_pfn_t containing source pfns
2915 * @dst: array of hmm_pfn_t containing destination pfns
2916 * @private: pointer passed back to each of the callback
2917 * Returns: 0 on success, error code otherwise
2919 * This function tries to migrate a range of memory virtual address range, using
2920 * callbacks to allocate and copy memory from source to destination. First it
2921 * collects all the pages backing each virtual address in the range, saving this
2922 * inside the src array. Then it locks those pages and unmaps them. Once the pages
2923 * are locked and unmapped, it checks whether each page is pinned or not. Pages
2924 * that aren't pinned have the MIGRATE_PFN_MIGRATE flag set (by this function)
2925 * in the corresponding src array entry. It then restores any pages that are
2926 * pinned, by remapping and unlocking those pages.
2928 * At this point it calls the alloc_and_copy() callback. For documentation on
2929 * what is expected from that callback, see struct migrate_vma_ops comments in
2930 * include/linux/migrate.h
2932 * After the alloc_and_copy() callback, this function goes over each entry in
2933 * the src array that has the MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE flag
2934 * set. If the corresponding entry in dst array has MIGRATE_PFN_VALID flag set,
2935 * then the function tries to migrate struct page information from the source
2936 * struct page to the destination struct page. If it fails to migrate the struct
2937 * page information, then it clears the MIGRATE_PFN_MIGRATE flag in the src
2940 * At this point all successfully migrated pages have an entry in the src
2941 * array with MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE flag set and the dst
2942 * array entry with MIGRATE_PFN_VALID flag set.
2944 * It then calls the finalize_and_map() callback. See comments for "struct
2945 * migrate_vma_ops", in include/linux/migrate.h for details about
2946 * finalize_and_map() behavior.
2948 * After the finalize_and_map() callback, for successfully migrated pages, this
2949 * function updates the CPU page table to point to new pages, otherwise it
2950 * restores the CPU page table to point to the original source pages.
2952 * Function returns 0 after the above steps, even if no pages were migrated
2953 * (The function only returns an error if any of the arguments are invalid.)
2955 * Both src and dst array must be big enough for (end - start) >> PAGE_SHIFT
2956 * unsigned long entries.
2958 int migrate_vma(const struct migrate_vma_ops *ops,
2959 struct vm_area_struct *vma,
2960 unsigned long start,
2966 struct migrate_vma migrate;
2968 /* Sanity check the arguments */
2971 if (!vma || is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_SPECIAL) ||
2974 if (start < vma->vm_start || start >= vma->vm_end)
2976 if (end <= vma->vm_start || end > vma->vm_end)
2978 if (!ops || !src || !dst || start >= end)
2981 memset(src, 0, sizeof(*src) * ((end - start) >> PAGE_SHIFT));
2984 migrate.start = start;
2990 /* Collect, and try to unmap source pages */
2991 migrate_vma_collect(&migrate);
2992 if (!migrate.cpages)
2995 /* Lock and isolate page */
2996 migrate_vma_prepare(&migrate);
2997 if (!migrate.cpages)
3001 migrate_vma_unmap(&migrate);
3002 if (!migrate.cpages)
3006 * At this point pages are locked and unmapped, and thus they have
3007 * stable content and can safely be copied to destination memory that
3008 * is allocated by the callback.
3010 * Note that migration can fail in migrate_vma_struct_page() for each
3013 ops->alloc_and_copy(vma, src, dst, start, end, private);
3015 /* This does the real migration of struct page */
3016 migrate_vma_pages(&migrate);
3018 ops->finalize_and_map(vma, src, dst, start, end, private);
3020 /* Unlock and remap pages */
3021 migrate_vma_finalize(&migrate);
3025 EXPORT_SYMBOL(migrate_vma);
3026 #endif /* defined(MIGRATE_VMA_HELPER) */