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 expected_count += hpage_nr_pages(page) + page_has_private(page);
477 if (page_count(page) != expected_count ||
478 radix_tree_deref_slot_protected(pslot,
479 &mapping->i_pages.xa_lock) != page) {
480 xa_unlock_irq(&mapping->i_pages);
484 if (!page_ref_freeze(page, expected_count)) {
485 xa_unlock_irq(&mapping->i_pages);
490 * In the async migration case of moving a page with buffers, lock the
491 * buffers using trylock before the mapping is moved. If the mapping
492 * was moved, we later failed to lock the buffers and could not move
493 * the mapping back due to an elevated page count, we would have to
494 * block waiting on other references to be dropped.
496 if (mode == MIGRATE_ASYNC && head &&
497 !buffer_migrate_lock_buffers(head, mode)) {
498 page_ref_unfreeze(page, expected_count);
499 xa_unlock_irq(&mapping->i_pages);
504 * Now we know that no one else is looking at the page:
505 * no turning back from here.
507 newpage->index = page->index;
508 newpage->mapping = page->mapping;
509 page_ref_add(newpage, hpage_nr_pages(page)); /* add cache reference */
510 if (PageSwapBacked(page)) {
511 __SetPageSwapBacked(newpage);
512 if (PageSwapCache(page)) {
513 SetPageSwapCache(newpage);
514 set_page_private(newpage, page_private(page));
517 VM_BUG_ON_PAGE(PageSwapCache(page), page);
520 /* Move dirty while page refs frozen and newpage not yet exposed */
521 dirty = PageDirty(page);
523 ClearPageDirty(page);
524 SetPageDirty(newpage);
527 radix_tree_replace_slot(&mapping->i_pages, pslot, newpage);
528 if (PageTransHuge(page)) {
530 int index = page_index(page);
532 for (i = 1; i < HPAGE_PMD_NR; i++) {
533 pslot = radix_tree_lookup_slot(&mapping->i_pages,
535 radix_tree_replace_slot(&mapping->i_pages, pslot,
541 * Drop cache reference from old page by unfreezing
542 * to one less reference.
543 * We know this isn't the last reference.
545 page_ref_unfreeze(page, expected_count - hpage_nr_pages(page));
547 xa_unlock(&mapping->i_pages);
548 /* Leave irq disabled to prevent preemption while updating stats */
551 * If moved to a different zone then also account
552 * the page for that zone. Other VM counters will be
553 * taken care of when we establish references to the
554 * new page and drop references to the old page.
556 * Note that anonymous pages are accounted for
557 * via NR_FILE_PAGES and NR_ANON_MAPPED if they
558 * are mapped to swap space.
560 if (newzone != oldzone) {
561 __dec_node_state(oldzone->zone_pgdat, NR_FILE_PAGES);
562 __inc_node_state(newzone->zone_pgdat, NR_FILE_PAGES);
563 if (PageSwapBacked(page) && !PageSwapCache(page)) {
564 __dec_node_state(oldzone->zone_pgdat, NR_SHMEM);
565 __inc_node_state(newzone->zone_pgdat, NR_SHMEM);
567 if (dirty && mapping_cap_account_dirty(mapping)) {
568 __dec_node_state(oldzone->zone_pgdat, NR_FILE_DIRTY);
569 __dec_zone_state(oldzone, NR_ZONE_WRITE_PENDING);
570 __inc_node_state(newzone->zone_pgdat, NR_FILE_DIRTY);
571 __inc_zone_state(newzone, NR_ZONE_WRITE_PENDING);
576 return MIGRATEPAGE_SUCCESS;
578 EXPORT_SYMBOL(migrate_page_move_mapping);
581 * The expected number of remaining references is the same as that
582 * of migrate_page_move_mapping().
584 int migrate_huge_page_move_mapping(struct address_space *mapping,
585 struct page *newpage, struct page *page)
590 xa_lock_irq(&mapping->i_pages);
592 pslot = radix_tree_lookup_slot(&mapping->i_pages, page_index(page));
594 expected_count = 2 + page_has_private(page);
595 if (page_count(page) != expected_count ||
596 radix_tree_deref_slot_protected(pslot, &mapping->i_pages.xa_lock) != page) {
597 xa_unlock_irq(&mapping->i_pages);
601 if (!page_ref_freeze(page, expected_count)) {
602 xa_unlock_irq(&mapping->i_pages);
606 newpage->index = page->index;
607 newpage->mapping = page->mapping;
611 radix_tree_replace_slot(&mapping->i_pages, pslot, newpage);
613 page_ref_unfreeze(page, expected_count - 1);
615 xa_unlock_irq(&mapping->i_pages);
617 return MIGRATEPAGE_SUCCESS;
621 * Gigantic pages are so large that we do not guarantee that page++ pointer
622 * arithmetic will work across the entire page. We need something more
625 static void __copy_gigantic_page(struct page *dst, struct page *src,
629 struct page *dst_base = dst;
630 struct page *src_base = src;
632 for (i = 0; i < nr_pages; ) {
634 copy_highpage(dst, src);
637 dst = mem_map_next(dst, dst_base, i);
638 src = mem_map_next(src, src_base, i);
642 static void copy_huge_page(struct page *dst, struct page *src)
649 struct hstate *h = page_hstate(src);
650 nr_pages = pages_per_huge_page(h);
652 if (unlikely(nr_pages > MAX_ORDER_NR_PAGES)) {
653 __copy_gigantic_page(dst, src, nr_pages);
658 BUG_ON(!PageTransHuge(src));
659 nr_pages = hpage_nr_pages(src);
662 for (i = 0; i < nr_pages; i++) {
664 copy_highpage(dst + i, src + i);
669 * Copy the page to its new location
671 void migrate_page_states(struct page *newpage, struct page *page)
676 SetPageError(newpage);
677 if (PageReferenced(page))
678 SetPageReferenced(newpage);
679 if (PageUptodate(page))
680 SetPageUptodate(newpage);
681 if (TestClearPageActive(page)) {
682 VM_BUG_ON_PAGE(PageUnevictable(page), page);
683 SetPageActive(newpage);
684 } else if (TestClearPageUnevictable(page))
685 SetPageUnevictable(newpage);
686 if (PageChecked(page))
687 SetPageChecked(newpage);
688 if (PageMappedToDisk(page))
689 SetPageMappedToDisk(newpage);
691 /* Move dirty on pages not done by migrate_page_move_mapping() */
693 SetPageDirty(newpage);
695 if (page_is_young(page))
696 set_page_young(newpage);
697 if (page_is_idle(page))
698 set_page_idle(newpage);
701 * Copy NUMA information to the new page, to prevent over-eager
702 * future migrations of this same page.
704 cpupid = page_cpupid_xchg_last(page, -1);
705 page_cpupid_xchg_last(newpage, cpupid);
707 ksm_migrate_page(newpage, page);
709 * Please do not reorder this without considering how mm/ksm.c's
710 * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache().
712 if (PageSwapCache(page))
713 ClearPageSwapCache(page);
714 ClearPagePrivate(page);
715 set_page_private(page, 0);
718 * If any waiters have accumulated on the new page then
721 if (PageWriteback(newpage))
722 end_page_writeback(newpage);
724 copy_page_owner(page, newpage);
726 mem_cgroup_migrate(page, newpage);
728 EXPORT_SYMBOL(migrate_page_states);
730 void migrate_page_copy(struct page *newpage, struct page *page)
732 if (PageHuge(page) || PageTransHuge(page))
733 copy_huge_page(newpage, page);
735 copy_highpage(newpage, page);
737 migrate_page_states(newpage, page);
739 EXPORT_SYMBOL(migrate_page_copy);
741 /************************************************************
742 * Migration functions
743 ***********************************************************/
746 * Common logic to directly migrate a single LRU page suitable for
747 * pages that do not use PagePrivate/PagePrivate2.
749 * Pages are locked upon entry and exit.
751 int migrate_page(struct address_space *mapping,
752 struct page *newpage, struct page *page,
753 enum migrate_mode mode)
757 BUG_ON(PageWriteback(page)); /* Writeback must be complete */
759 rc = migrate_page_move_mapping(mapping, newpage, page, NULL, mode, 0);
761 if (rc != MIGRATEPAGE_SUCCESS)
764 if (mode != MIGRATE_SYNC_NO_COPY)
765 migrate_page_copy(newpage, page);
767 migrate_page_states(newpage, page);
768 return MIGRATEPAGE_SUCCESS;
770 EXPORT_SYMBOL(migrate_page);
774 * Migration function for pages with buffers. This function can only be used
775 * if the underlying filesystem guarantees that no other references to "page"
778 int buffer_migrate_page(struct address_space *mapping,
779 struct page *newpage, struct page *page, enum migrate_mode mode)
781 struct buffer_head *bh, *head;
784 if (!page_has_buffers(page))
785 return migrate_page(mapping, newpage, page, mode);
787 head = page_buffers(page);
789 rc = migrate_page_move_mapping(mapping, newpage, page, head, mode, 0);
791 if (rc != MIGRATEPAGE_SUCCESS)
795 * In the async case, migrate_page_move_mapping locked the buffers
796 * with an IRQ-safe spinlock held. In the sync case, the buffers
797 * need to be locked now
799 if (mode != MIGRATE_ASYNC)
800 BUG_ON(!buffer_migrate_lock_buffers(head, mode));
802 ClearPagePrivate(page);
803 set_page_private(newpage, page_private(page));
804 set_page_private(page, 0);
810 set_bh_page(bh, newpage, bh_offset(bh));
811 bh = bh->b_this_page;
813 } while (bh != head);
815 SetPagePrivate(newpage);
817 if (mode != MIGRATE_SYNC_NO_COPY)
818 migrate_page_copy(newpage, page);
820 migrate_page_states(newpage, page);
826 bh = bh->b_this_page;
828 } while (bh != head);
830 return MIGRATEPAGE_SUCCESS;
832 EXPORT_SYMBOL(buffer_migrate_page);
836 * Writeback a page to clean the dirty state
838 static int writeout(struct address_space *mapping, struct page *page)
840 struct writeback_control wbc = {
841 .sync_mode = WB_SYNC_NONE,
844 .range_end = LLONG_MAX,
849 if (!mapping->a_ops->writepage)
850 /* No write method for the address space */
853 if (!clear_page_dirty_for_io(page))
854 /* Someone else already triggered a write */
858 * A dirty page may imply that the underlying filesystem has
859 * the page on some queue. So the page must be clean for
860 * migration. Writeout may mean we loose the lock and the
861 * page state is no longer what we checked for earlier.
862 * At this point we know that the migration attempt cannot
865 remove_migration_ptes(page, page, false);
867 rc = mapping->a_ops->writepage(page, &wbc);
869 if (rc != AOP_WRITEPAGE_ACTIVATE)
870 /* unlocked. Relock */
873 return (rc < 0) ? -EIO : -EAGAIN;
877 * Default handling if a filesystem does not provide a migration function.
879 static int fallback_migrate_page(struct address_space *mapping,
880 struct page *newpage, struct page *page, enum migrate_mode mode)
882 if (PageDirty(page)) {
883 /* Only writeback pages in full synchronous migration */
886 case MIGRATE_SYNC_NO_COPY:
891 return writeout(mapping, page);
895 * Buffers may be managed in a filesystem specific way.
896 * We must have no buffers or drop them.
898 if (page_has_private(page) &&
899 !try_to_release_page(page, GFP_KERNEL))
902 return migrate_page(mapping, newpage, page, mode);
906 * Move a page to a newly allocated page
907 * The page is locked and all ptes have been successfully removed.
909 * The new page will have replaced the old page if this function
914 * MIGRATEPAGE_SUCCESS - success
916 static int move_to_new_page(struct page *newpage, struct page *page,
917 enum migrate_mode mode)
919 struct address_space *mapping;
921 bool is_lru = !__PageMovable(page);
923 VM_BUG_ON_PAGE(!PageLocked(page), page);
924 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
926 mapping = page_mapping(page);
928 if (likely(is_lru)) {
930 rc = migrate_page(mapping, newpage, page, mode);
931 else if (mapping->a_ops->migratepage)
933 * Most pages have a mapping and most filesystems
934 * provide a migratepage callback. Anonymous pages
935 * are part of swap space which also has its own
936 * migratepage callback. This is the most common path
937 * for page migration.
939 rc = mapping->a_ops->migratepage(mapping, newpage,
942 rc = fallback_migrate_page(mapping, newpage,
946 * In case of non-lru page, it could be released after
947 * isolation step. In that case, we shouldn't try migration.
949 VM_BUG_ON_PAGE(!PageIsolated(page), page);
950 if (!PageMovable(page)) {
951 rc = MIGRATEPAGE_SUCCESS;
952 __ClearPageIsolated(page);
956 rc = mapping->a_ops->migratepage(mapping, newpage,
958 WARN_ON_ONCE(rc == MIGRATEPAGE_SUCCESS &&
959 !PageIsolated(page));
963 * When successful, old pagecache page->mapping must be cleared before
964 * page is freed; but stats require that PageAnon be left as PageAnon.
966 if (rc == MIGRATEPAGE_SUCCESS) {
967 if (__PageMovable(page)) {
968 VM_BUG_ON_PAGE(!PageIsolated(page), page);
971 * We clear PG_movable under page_lock so any compactor
972 * cannot try to migrate this page.
974 __ClearPageIsolated(page);
978 * Anonymous and movable page->mapping will be cleard by
979 * free_pages_prepare so don't reset it here for keeping
980 * the type to work PageAnon, for example.
982 if (!PageMappingFlags(page))
983 page->mapping = NULL;
985 if (unlikely(is_zone_device_page(newpage))) {
986 if (is_device_public_page(newpage))
987 flush_dcache_page(newpage);
989 flush_dcache_page(newpage);
996 static int __unmap_and_move(struct page *page, struct page *newpage,
997 int force, enum migrate_mode mode)
1000 int page_was_mapped = 0;
1001 struct anon_vma *anon_vma = NULL;
1002 bool is_lru = !__PageMovable(page);
1004 if (!trylock_page(page)) {
1005 if (!force || mode == MIGRATE_ASYNC)
1009 * It's not safe for direct compaction to call lock_page.
1010 * For example, during page readahead pages are added locked
1011 * to the LRU. Later, when the IO completes the pages are
1012 * marked uptodate and unlocked. However, the queueing
1013 * could be merging multiple pages for one bio (e.g.
1014 * mpage_readpages). If an allocation happens for the
1015 * second or third page, the process can end up locking
1016 * the same page twice and deadlocking. Rather than
1017 * trying to be clever about what pages can be locked,
1018 * avoid the use of lock_page for direct compaction
1021 if (current->flags & PF_MEMALLOC)
1027 if (PageWriteback(page)) {
1029 * Only in the case of a full synchronous migration is it
1030 * necessary to wait for PageWriteback. In the async case,
1031 * the retry loop is too short and in the sync-light case,
1032 * the overhead of stalling is too much
1036 case MIGRATE_SYNC_NO_COPY:
1044 wait_on_page_writeback(page);
1048 * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
1049 * we cannot notice that anon_vma is freed while we migrates a page.
1050 * This get_anon_vma() delays freeing anon_vma pointer until the end
1051 * of migration. File cache pages are no problem because of page_lock()
1052 * File Caches may use write_page() or lock_page() in migration, then,
1053 * just care Anon page here.
1055 * Only page_get_anon_vma() understands the subtleties of
1056 * getting a hold on an anon_vma from outside one of its mms.
1057 * But if we cannot get anon_vma, then we won't need it anyway,
1058 * because that implies that the anon page is no longer mapped
1059 * (and cannot be remapped so long as we hold the page lock).
1061 if (PageAnon(page) && !PageKsm(page))
1062 anon_vma = page_get_anon_vma(page);
1065 * Block others from accessing the new page when we get around to
1066 * establishing additional references. We are usually the only one
1067 * holding a reference to newpage at this point. We used to have a BUG
1068 * here if trylock_page(newpage) fails, but would like to allow for
1069 * cases where there might be a race with the previous use of newpage.
1070 * This is much like races on refcount of oldpage: just don't BUG().
1072 if (unlikely(!trylock_page(newpage)))
1075 if (unlikely(!is_lru)) {
1076 rc = move_to_new_page(newpage, page, mode);
1077 goto out_unlock_both;
1081 * Corner case handling:
1082 * 1. When a new swap-cache page is read into, it is added to the LRU
1083 * and treated as swapcache but it has no rmap yet.
1084 * Calling try_to_unmap() against a page->mapping==NULL page will
1085 * trigger a BUG. So handle it here.
1086 * 2. An orphaned page (see truncate_complete_page) might have
1087 * fs-private metadata. The page can be picked up due to memory
1088 * offlining. Everywhere else except page reclaim, the page is
1089 * invisible to the vm, so the page can not be migrated. So try to
1090 * free the metadata, so the page can be freed.
1092 if (!page->mapping) {
1093 VM_BUG_ON_PAGE(PageAnon(page), page);
1094 if (page_has_private(page)) {
1095 try_to_free_buffers(page);
1096 goto out_unlock_both;
1098 } else if (page_mapped(page)) {
1099 /* Establish migration ptes */
1100 VM_BUG_ON_PAGE(PageAnon(page) && !PageKsm(page) && !anon_vma,
1103 TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
1104 page_was_mapped = 1;
1107 if (!page_mapped(page))
1108 rc = move_to_new_page(newpage, page, mode);
1110 if (page_was_mapped)
1111 remove_migration_ptes(page,
1112 rc == MIGRATEPAGE_SUCCESS ? newpage : page, false);
1115 unlock_page(newpage);
1117 /* Drop an anon_vma reference if we took one */
1119 put_anon_vma(anon_vma);
1123 * If migration is successful, decrease refcount of the newpage
1124 * which will not free the page because new page owner increased
1125 * refcounter. As well, if it is LRU page, add the page to LRU
1126 * list in here. Use the old state of the isolated source page to
1127 * determine if we migrated a LRU page. newpage was already unlocked
1128 * and possibly modified by its owner - don't rely on the page
1131 if (rc == MIGRATEPAGE_SUCCESS) {
1132 if (unlikely(!is_lru))
1135 putback_lru_page(newpage);
1142 * gcc 4.7 and 4.8 on arm get an ICEs when inlining unmap_and_move(). Work
1145 #if defined(CONFIG_ARM) && \
1146 defined(GCC_VERSION) && GCC_VERSION < 40900 && GCC_VERSION >= 40700
1147 #define ICE_noinline noinline
1149 #define ICE_noinline
1153 * Obtain the lock on page, remove all ptes and migrate the page
1154 * to the newly allocated page in newpage.
1156 static ICE_noinline int unmap_and_move(new_page_t get_new_page,
1157 free_page_t put_new_page,
1158 unsigned long private, struct page *page,
1159 int force, enum migrate_mode mode,
1160 enum migrate_reason reason)
1162 int rc = MIGRATEPAGE_SUCCESS;
1163 struct page *newpage;
1165 if (!thp_migration_supported() && PageTransHuge(page))
1168 newpage = get_new_page(page, private);
1172 if (page_count(page) == 1) {
1173 /* page was freed from under us. So we are done. */
1174 ClearPageActive(page);
1175 ClearPageUnevictable(page);
1176 if (unlikely(__PageMovable(page))) {
1178 if (!PageMovable(page))
1179 __ClearPageIsolated(page);
1183 put_new_page(newpage, private);
1189 rc = __unmap_and_move(page, newpage, force, mode);
1190 if (rc == MIGRATEPAGE_SUCCESS)
1191 set_page_owner_migrate_reason(newpage, reason);
1194 if (rc != -EAGAIN) {
1196 * A page that has been migrated has all references
1197 * removed and will be freed. A page that has not been
1198 * migrated will have kepts its references and be
1201 list_del(&page->lru);
1204 * Compaction can migrate also non-LRU pages which are
1205 * not accounted to NR_ISOLATED_*. They can be recognized
1208 if (likely(!__PageMovable(page)))
1209 mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON +
1210 page_is_file_cache(page), -hpage_nr_pages(page));
1214 * If migration is successful, releases reference grabbed during
1215 * isolation. Otherwise, restore the page to right list unless
1218 if (rc == MIGRATEPAGE_SUCCESS) {
1220 if (reason == MR_MEMORY_FAILURE) {
1222 * Set PG_HWPoison on just freed page
1223 * intentionally. Although it's rather weird,
1224 * it's how HWPoison flag works at the moment.
1226 if (set_hwpoison_free_buddy_page(page))
1227 num_poisoned_pages_inc();
1230 if (rc != -EAGAIN) {
1231 if (likely(!__PageMovable(page))) {
1232 putback_lru_page(page);
1237 if (PageMovable(page))
1238 putback_movable_page(page);
1240 __ClearPageIsolated(page);
1246 put_new_page(newpage, private);
1255 * Counterpart of unmap_and_move_page() for hugepage migration.
1257 * This function doesn't wait the completion of hugepage I/O
1258 * because there is no race between I/O and migration for hugepage.
1259 * Note that currently hugepage I/O occurs only in direct I/O
1260 * where no lock is held and PG_writeback is irrelevant,
1261 * and writeback status of all subpages are counted in the reference
1262 * count of the head page (i.e. if all subpages of a 2MB hugepage are
1263 * under direct I/O, the reference of the head page is 512 and a bit more.)
1264 * This means that when we try to migrate hugepage whose subpages are
1265 * doing direct I/O, some references remain after try_to_unmap() and
1266 * hugepage migration fails without data corruption.
1268 * There is also no race when direct I/O is issued on the page under migration,
1269 * because then pte is replaced with migration swap entry and direct I/O code
1270 * will wait in the page fault for migration to complete.
1272 static int unmap_and_move_huge_page(new_page_t get_new_page,
1273 free_page_t put_new_page, unsigned long private,
1274 struct page *hpage, int force,
1275 enum migrate_mode mode, int reason)
1278 int page_was_mapped = 0;
1279 struct page *new_hpage;
1280 struct anon_vma *anon_vma = NULL;
1283 * Movability of hugepages depends on architectures and hugepage size.
1284 * This check is necessary because some callers of hugepage migration
1285 * like soft offline and memory hotremove don't walk through page
1286 * tables or check whether the hugepage is pmd-based or not before
1287 * kicking migration.
1289 if (!hugepage_migration_supported(page_hstate(hpage))) {
1290 putback_active_hugepage(hpage);
1294 new_hpage = get_new_page(hpage, private);
1298 if (!trylock_page(hpage)) {
1303 case MIGRATE_SYNC_NO_COPY:
1312 * Check for pages which are in the process of being freed. Without
1313 * page_mapping() set, hugetlbfs specific move page routine will not
1314 * be called and we could leak usage counts for subpools.
1316 if (page_private(hpage) && !page_mapping(hpage)) {
1321 if (PageAnon(hpage))
1322 anon_vma = page_get_anon_vma(hpage);
1324 if (unlikely(!trylock_page(new_hpage)))
1327 if (page_mapped(hpage)) {
1329 TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
1330 page_was_mapped = 1;
1333 if (!page_mapped(hpage))
1334 rc = move_to_new_page(new_hpage, hpage, mode);
1336 if (page_was_mapped)
1337 remove_migration_ptes(hpage,
1338 rc == MIGRATEPAGE_SUCCESS ? new_hpage : hpage, false);
1340 unlock_page(new_hpage);
1344 put_anon_vma(anon_vma);
1346 if (rc == MIGRATEPAGE_SUCCESS) {
1347 move_hugetlb_state(hpage, new_hpage, reason);
1348 put_new_page = NULL;
1355 putback_active_hugepage(hpage);
1358 * If migration was not successful and there's a freeing callback, use
1359 * it. Otherwise, put_page() will drop the reference grabbed during
1363 put_new_page(new_hpage, private);
1365 putback_active_hugepage(new_hpage);
1371 * migrate_pages - migrate the pages specified in a list, to the free pages
1372 * supplied as the target for the page migration
1374 * @from: The list of pages to be migrated.
1375 * @get_new_page: The function used to allocate free pages to be used
1376 * as the target of the page migration.
1377 * @put_new_page: The function used to free target pages if migration
1378 * fails, or NULL if no special handling is necessary.
1379 * @private: Private data to be passed on to get_new_page()
1380 * @mode: The migration mode that specifies the constraints for
1381 * page migration, if any.
1382 * @reason: The reason for page migration.
1384 * The function returns after 10 attempts or if no pages are movable any more
1385 * because the list has become empty or no retryable pages exist any more.
1386 * The caller should call putback_movable_pages() to return pages to the LRU
1387 * or free list only if ret != 0.
1389 * Returns the number of pages that were not migrated, or an error code.
1391 int migrate_pages(struct list_head *from, new_page_t get_new_page,
1392 free_page_t put_new_page, unsigned long private,
1393 enum migrate_mode mode, int reason)
1397 int nr_succeeded = 0;
1401 int swapwrite = current->flags & PF_SWAPWRITE;
1405 current->flags |= PF_SWAPWRITE;
1407 for(pass = 0; pass < 10 && retry; pass++) {
1410 list_for_each_entry_safe(page, page2, from, lru) {
1415 rc = unmap_and_move_huge_page(get_new_page,
1416 put_new_page, private, page,
1417 pass > 2, mode, reason);
1419 rc = unmap_and_move(get_new_page, put_new_page,
1420 private, page, pass > 2, mode,
1426 * THP migration might be unsupported or the
1427 * allocation could've failed so we should
1428 * retry on the same page with the THP split
1431 * Head page is retried immediately and tail
1432 * pages are added to the tail of the list so
1433 * we encounter them after the rest of the list
1436 if (PageTransHuge(page) && !PageHuge(page)) {
1438 rc = split_huge_page_to_list(page, from);
1441 list_safe_reset_next(page, page2, lru);
1450 case MIGRATEPAGE_SUCCESS:
1455 * Permanent failure (-EBUSY, -ENOSYS, etc.):
1456 * unlike -EAGAIN case, the failed page is
1457 * removed from migration page list and not
1458 * retried in the next outer loop.
1469 count_vm_events(PGMIGRATE_SUCCESS, nr_succeeded);
1471 count_vm_events(PGMIGRATE_FAIL, nr_failed);
1472 trace_mm_migrate_pages(nr_succeeded, nr_failed, mode, reason);
1475 current->flags &= ~PF_SWAPWRITE;
1482 static int store_status(int __user *status, int start, int value, int nr)
1485 if (put_user(value, status + start))
1493 static int do_move_pages_to_node(struct mm_struct *mm,
1494 struct list_head *pagelist, int node)
1498 if (list_empty(pagelist))
1501 err = migrate_pages(pagelist, alloc_new_node_page, NULL, node,
1502 MIGRATE_SYNC, MR_SYSCALL);
1504 putback_movable_pages(pagelist);
1509 * Resolves the given address to a struct page, isolates it from the LRU and
1510 * puts it to the given pagelist.
1512 * errno - if the page cannot be found/isolated
1513 * 0 - when it doesn't have to be migrated because it is already on the
1515 * 1 - when it has been queued
1517 static int add_page_for_migration(struct mm_struct *mm, unsigned long addr,
1518 int node, struct list_head *pagelist, bool migrate_all)
1520 struct vm_area_struct *vma;
1522 unsigned int follflags;
1525 down_read(&mm->mmap_sem);
1527 vma = find_vma(mm, addr);
1528 if (!vma || addr < vma->vm_start || !vma_migratable(vma))
1531 /* FOLL_DUMP to ignore special (like zero) pages */
1532 follflags = FOLL_GET | FOLL_DUMP;
1533 page = follow_page(vma, addr, follflags);
1535 err = PTR_ERR(page);
1544 if (page_to_nid(page) == node)
1548 if (page_mapcount(page) > 1 && !migrate_all)
1551 if (PageHuge(page)) {
1552 if (PageHead(page)) {
1553 isolate_huge_page(page, pagelist);
1559 head = compound_head(page);
1560 err = isolate_lru_page(head);
1565 list_add_tail(&head->lru, pagelist);
1566 mod_node_page_state(page_pgdat(head),
1567 NR_ISOLATED_ANON + page_is_file_cache(head),
1568 hpage_nr_pages(head));
1572 * Either remove the duplicate refcount from
1573 * isolate_lru_page() or drop the page ref if it was
1578 up_read(&mm->mmap_sem);
1583 * Migrate an array of page address onto an array of nodes and fill
1584 * the corresponding array of status.
1586 static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes,
1587 unsigned long nr_pages,
1588 const void __user * __user *pages,
1589 const int __user *nodes,
1590 int __user *status, int flags)
1592 int current_node = NUMA_NO_NODE;
1593 LIST_HEAD(pagelist);
1599 for (i = start = 0; i < nr_pages; i++) {
1600 const void __user *p;
1605 if (get_user(p, pages + i))
1607 if (get_user(node, nodes + i))
1609 addr = (unsigned long)p;
1612 if (node < 0 || node >= MAX_NUMNODES)
1614 if (!node_state(node, N_MEMORY))
1618 if (!node_isset(node, task_nodes))
1621 if (current_node == NUMA_NO_NODE) {
1622 current_node = node;
1624 } else if (node != current_node) {
1625 err = do_move_pages_to_node(mm, &pagelist, current_node);
1628 * Positive err means the number of failed
1629 * pages to migrate. Since we are going to
1630 * abort and return the number of non-migrated
1631 * pages, so need to incude the rest of the
1632 * nr_pages that have not been attempted as
1636 err += nr_pages - i - 1;
1639 err = store_status(status, start, current_node, i - start);
1643 current_node = node;
1647 * Errors in the page lookup or isolation are not fatal and we simply
1648 * report them via status
1650 err = add_page_for_migration(mm, addr, current_node,
1651 &pagelist, flags & MPOL_MF_MOVE_ALL);
1654 /* The page is already on the target node */
1655 err = store_status(status, i, current_node, 1);
1659 } else if (err > 0) {
1660 /* The page is successfully queued for migration */
1664 err = store_status(status, i, err, 1);
1668 err = do_move_pages_to_node(mm, &pagelist, current_node);
1671 err += nr_pages - i - 1;
1675 err = store_status(status, start, current_node, i - start);
1679 current_node = NUMA_NO_NODE;
1682 if (list_empty(&pagelist))
1685 /* Make sure we do not overwrite the existing error */
1686 err1 = do_move_pages_to_node(mm, &pagelist, current_node);
1688 * Don't have to report non-attempted pages here since:
1689 * - If the above loop is done gracefully all pages have been
1691 * - If the above loop is aborted it means a fatal error
1692 * happened, should return ret.
1695 err1 = store_status(status, start, current_node, i - start);
1703 * Determine the nodes of an array of pages and store it in an array of status.
1705 static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages,
1706 const void __user **pages, int *status)
1710 down_read(&mm->mmap_sem);
1712 for (i = 0; i < nr_pages; i++) {
1713 unsigned long addr = (unsigned long)(*pages);
1714 struct vm_area_struct *vma;
1718 vma = find_vma(mm, addr);
1719 if (!vma || addr < vma->vm_start)
1722 /* FOLL_DUMP to ignore special (like zero) pages */
1723 page = follow_page(vma, addr, FOLL_DUMP);
1725 err = PTR_ERR(page);
1729 err = page ? page_to_nid(page) : -ENOENT;
1737 up_read(&mm->mmap_sem);
1741 * Determine the nodes of a user array of pages and store it in
1742 * a user array of status.
1744 static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages,
1745 const void __user * __user *pages,
1748 #define DO_PAGES_STAT_CHUNK_NR 16
1749 const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR];
1750 int chunk_status[DO_PAGES_STAT_CHUNK_NR];
1753 unsigned long chunk_nr;
1755 chunk_nr = nr_pages;
1756 if (chunk_nr > DO_PAGES_STAT_CHUNK_NR)
1757 chunk_nr = DO_PAGES_STAT_CHUNK_NR;
1759 if (copy_from_user(chunk_pages, pages, chunk_nr * sizeof(*chunk_pages)))
1762 do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status);
1764 if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status)))
1769 nr_pages -= chunk_nr;
1771 return nr_pages ? -EFAULT : 0;
1775 * Move a list of pages in the address space of the currently executing
1778 static int kernel_move_pages(pid_t pid, unsigned long nr_pages,
1779 const void __user * __user *pages,
1780 const int __user *nodes,
1781 int __user *status, int flags)
1783 struct task_struct *task;
1784 struct mm_struct *mm;
1786 nodemask_t task_nodes;
1789 if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
1792 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1795 /* Find the mm_struct */
1797 task = pid ? find_task_by_vpid(pid) : current;
1802 get_task_struct(task);
1805 * Check if this process has the right to modify the specified
1806 * process. Use the regular "ptrace_may_access()" checks.
1808 if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) {
1815 err = security_task_movememory(task);
1819 task_nodes = cpuset_mems_allowed(task);
1820 mm = get_task_mm(task);
1821 put_task_struct(task);
1827 err = do_pages_move(mm, task_nodes, nr_pages, pages,
1828 nodes, status, flags);
1830 err = do_pages_stat(mm, nr_pages, pages, status);
1836 put_task_struct(task);
1840 SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages,
1841 const void __user * __user *, pages,
1842 const int __user *, nodes,
1843 int __user *, status, int, flags)
1845 return kernel_move_pages(pid, nr_pages, pages, nodes, status, flags);
1848 #ifdef CONFIG_COMPAT
1849 COMPAT_SYSCALL_DEFINE6(move_pages, pid_t, pid, compat_ulong_t, nr_pages,
1850 compat_uptr_t __user *, pages32,
1851 const int __user *, nodes,
1852 int __user *, status,
1855 const void __user * __user *pages;
1858 pages = compat_alloc_user_space(nr_pages * sizeof(void *));
1859 for (i = 0; i < nr_pages; i++) {
1862 if (get_user(p, pages32 + i) ||
1863 put_user(compat_ptr(p), pages + i))
1866 return kernel_move_pages(pid, nr_pages, pages, nodes, status, flags);
1868 #endif /* CONFIG_COMPAT */
1870 #ifdef CONFIG_NUMA_BALANCING
1872 * Returns true if this is a safe migration target node for misplaced NUMA
1873 * pages. Currently it only checks the watermarks which crude
1875 static bool migrate_balanced_pgdat(struct pglist_data *pgdat,
1876 unsigned long nr_migrate_pages)
1880 for (z = pgdat->nr_zones - 1; z >= 0; z--) {
1881 struct zone *zone = pgdat->node_zones + z;
1883 if (!populated_zone(zone))
1886 /* Avoid waking kswapd by allocating pages_to_migrate pages. */
1887 if (!zone_watermark_ok(zone, 0,
1888 high_wmark_pages(zone) +
1897 static struct page *alloc_misplaced_dst_page(struct page *page,
1900 int nid = (int) data;
1901 struct page *newpage;
1903 newpage = __alloc_pages_node(nid,
1904 (GFP_HIGHUSER_MOVABLE |
1905 __GFP_THISNODE | __GFP_NOMEMALLOC |
1906 __GFP_NORETRY | __GFP_NOWARN) &
1912 static int numamigrate_isolate_page(pg_data_t *pgdat, struct page *page)
1916 VM_BUG_ON_PAGE(compound_order(page) && !PageTransHuge(page), page);
1918 /* Avoid migrating to a node that is nearly full */
1919 if (!migrate_balanced_pgdat(pgdat, 1UL << compound_order(page)))
1922 if (isolate_lru_page(page))
1926 * migrate_misplaced_transhuge_page() skips page migration's usual
1927 * check on page_count(), so we must do it here, now that the page
1928 * has been isolated: a GUP pin, or any other pin, prevents migration.
1929 * The expected page count is 3: 1 for page's mapcount and 1 for the
1930 * caller's pin and 1 for the reference taken by isolate_lru_page().
1932 if (PageTransHuge(page) && page_count(page) != 3) {
1933 putback_lru_page(page);
1937 page_lru = page_is_file_cache(page);
1938 mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON + page_lru,
1939 hpage_nr_pages(page));
1942 * Isolating the page has taken another reference, so the
1943 * caller's reference can be safely dropped without the page
1944 * disappearing underneath us during migration.
1950 bool pmd_trans_migrating(pmd_t pmd)
1952 struct page *page = pmd_page(pmd);
1953 return PageLocked(page);
1957 * Attempt to migrate a misplaced page to the specified destination
1958 * node. Caller is expected to have an elevated reference count on
1959 * the page that will be dropped by this function before returning.
1961 int migrate_misplaced_page(struct page *page, struct vm_area_struct *vma,
1964 pg_data_t *pgdat = NODE_DATA(node);
1967 LIST_HEAD(migratepages);
1970 * Don't migrate file pages that are mapped in multiple processes
1971 * with execute permissions as they are probably shared libraries.
1973 if (page_mapcount(page) != 1 && page_is_file_cache(page) &&
1974 (vma->vm_flags & VM_EXEC))
1978 * Also do not migrate dirty pages as not all filesystems can move
1979 * dirty pages in MIGRATE_ASYNC mode which is a waste of cycles.
1981 if (page_is_file_cache(page) && PageDirty(page))
1984 isolated = numamigrate_isolate_page(pgdat, page);
1988 list_add(&page->lru, &migratepages);
1989 nr_remaining = migrate_pages(&migratepages, alloc_misplaced_dst_page,
1990 NULL, node, MIGRATE_ASYNC,
1993 if (!list_empty(&migratepages)) {
1994 list_del(&page->lru);
1995 dec_node_page_state(page, NR_ISOLATED_ANON +
1996 page_is_file_cache(page));
1997 putback_lru_page(page);
2001 count_vm_numa_event(NUMA_PAGE_MIGRATE);
2002 BUG_ON(!list_empty(&migratepages));
2009 #endif /* CONFIG_NUMA_BALANCING */
2011 #if defined(CONFIG_NUMA_BALANCING) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
2013 * Migrates a THP to a given target node. page must be locked and is unlocked
2016 int migrate_misplaced_transhuge_page(struct mm_struct *mm,
2017 struct vm_area_struct *vma,
2018 pmd_t *pmd, pmd_t entry,
2019 unsigned long address,
2020 struct page *page, int node)
2023 pg_data_t *pgdat = NODE_DATA(node);
2025 struct page *new_page = NULL;
2026 int page_lru = page_is_file_cache(page);
2027 unsigned long mmun_start = address & HPAGE_PMD_MASK;
2028 unsigned long mmun_end = mmun_start + HPAGE_PMD_SIZE;
2030 new_page = alloc_pages_node(node,
2031 (GFP_TRANSHUGE_LIGHT | __GFP_THISNODE),
2035 prep_transhuge_page(new_page);
2037 isolated = numamigrate_isolate_page(pgdat, page);
2043 /* Prepare a page as a migration target */
2044 __SetPageLocked(new_page);
2045 if (PageSwapBacked(page))
2046 __SetPageSwapBacked(new_page);
2048 /* anon mapping, we can simply copy page->mapping to the new page: */
2049 new_page->mapping = page->mapping;
2050 new_page->index = page->index;
2051 migrate_page_copy(new_page, page);
2052 WARN_ON(PageLRU(new_page));
2054 /* Recheck the target PMD */
2055 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2056 ptl = pmd_lock(mm, pmd);
2057 if (unlikely(!pmd_same(*pmd, entry) || !page_ref_freeze(page, 2))) {
2059 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2061 /* Reverse changes made by migrate_page_copy() */
2062 if (TestClearPageActive(new_page))
2063 SetPageActive(page);
2064 if (TestClearPageUnevictable(new_page))
2065 SetPageUnevictable(page);
2067 unlock_page(new_page);
2068 put_page(new_page); /* Free it */
2070 /* Retake the callers reference and putback on LRU */
2072 putback_lru_page(page);
2073 mod_node_page_state(page_pgdat(page),
2074 NR_ISOLATED_ANON + page_lru, -HPAGE_PMD_NR);
2079 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
2080 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
2083 * Overwrite the old entry under pagetable lock and establish
2084 * the new PTE. Any parallel GUP will either observe the old
2085 * page blocking on the page lock, block on the page table
2086 * lock or observe the new page. The SetPageUptodate on the
2087 * new page and page_add_new_anon_rmap guarantee the copy is
2088 * visible before the pagetable update.
2090 flush_cache_range(vma, mmun_start, mmun_end);
2091 page_add_anon_rmap(new_page, vma, mmun_start, true);
2093 * At this point the pmd is numa/protnone (i.e. non present) and the TLB
2094 * has already been flushed globally. So no TLB can be currently
2095 * caching this non present pmd mapping. There's no need to clear the
2096 * pmd before doing set_pmd_at(), nor to flush the TLB after
2097 * set_pmd_at(). Clearing the pmd here would introduce a race
2098 * condition against MADV_DONTNEED, because MADV_DONTNEED only holds the
2099 * mmap_sem for reading. If the pmd is set to NULL at any given time,
2100 * MADV_DONTNEED won't wait on the pmd lock and it'll skip clearing this
2103 set_pmd_at(mm, mmun_start, pmd, entry);
2104 update_mmu_cache_pmd(vma, address, &entry);
2106 page_ref_unfreeze(page, 2);
2107 mlock_migrate_page(new_page, page);
2108 page_remove_rmap(page, true);
2109 set_page_owner_migrate_reason(new_page, MR_NUMA_MISPLACED);
2113 * No need to double call mmu_notifier->invalidate_range() callback as
2114 * the above pmdp_huge_clear_flush_notify() did already call it.
2116 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2118 /* Take an "isolate" reference and put new page on the LRU. */
2120 putback_lru_page(new_page);
2122 unlock_page(new_page);
2124 put_page(page); /* Drop the rmap reference */
2125 put_page(page); /* Drop the LRU isolation reference */
2127 count_vm_events(PGMIGRATE_SUCCESS, HPAGE_PMD_NR);
2128 count_vm_numa_events(NUMA_PAGE_MIGRATE, HPAGE_PMD_NR);
2130 mod_node_page_state(page_pgdat(page),
2131 NR_ISOLATED_ANON + page_lru,
2136 count_vm_events(PGMIGRATE_FAIL, HPAGE_PMD_NR);
2137 ptl = pmd_lock(mm, pmd);
2138 if (pmd_same(*pmd, entry)) {
2139 entry = pmd_modify(entry, vma->vm_page_prot);
2140 set_pmd_at(mm, mmun_start, pmd, entry);
2141 update_mmu_cache_pmd(vma, address, &entry);
2150 #endif /* CONFIG_NUMA_BALANCING */
2152 #endif /* CONFIG_NUMA */
2154 #if defined(CONFIG_MIGRATE_VMA_HELPER)
2155 struct migrate_vma {
2156 struct vm_area_struct *vma;
2159 unsigned long cpages;
2160 unsigned long npages;
2161 unsigned long start;
2165 static int migrate_vma_collect_hole(unsigned long start,
2167 struct mm_walk *walk)
2169 struct migrate_vma *migrate = walk->private;
2172 for (addr = start & PAGE_MASK; addr < end; addr += PAGE_SIZE) {
2173 migrate->src[migrate->npages] = MIGRATE_PFN_MIGRATE;
2174 migrate->dst[migrate->npages] = 0;
2182 static int migrate_vma_collect_skip(unsigned long start,
2184 struct mm_walk *walk)
2186 struct migrate_vma *migrate = walk->private;
2189 for (addr = start & PAGE_MASK; addr < end; addr += PAGE_SIZE) {
2190 migrate->dst[migrate->npages] = 0;
2191 migrate->src[migrate->npages++] = 0;
2197 static int migrate_vma_collect_pmd(pmd_t *pmdp,
2198 unsigned long start,
2200 struct mm_walk *walk)
2202 struct migrate_vma *migrate = walk->private;
2203 struct vm_area_struct *vma = walk->vma;
2204 struct mm_struct *mm = vma->vm_mm;
2205 unsigned long addr = start, unmapped = 0;
2210 if (pmd_none(*pmdp))
2211 return migrate_vma_collect_hole(start, end, walk);
2213 if (pmd_trans_huge(*pmdp)) {
2216 ptl = pmd_lock(mm, pmdp);
2217 if (unlikely(!pmd_trans_huge(*pmdp))) {
2222 page = pmd_page(*pmdp);
2223 if (is_huge_zero_page(page)) {
2225 split_huge_pmd(vma, pmdp, addr);
2226 if (pmd_trans_unstable(pmdp))
2227 return migrate_vma_collect_skip(start, end,
2234 if (unlikely(!trylock_page(page)))
2235 return migrate_vma_collect_skip(start, end,
2237 ret = split_huge_page(page);
2241 return migrate_vma_collect_skip(start, end,
2243 if (pmd_none(*pmdp))
2244 return migrate_vma_collect_hole(start, end,
2249 if (unlikely(pmd_bad(*pmdp)))
2250 return migrate_vma_collect_skip(start, end, walk);
2252 ptep = pte_offset_map_lock(mm, pmdp, addr, &ptl);
2253 arch_enter_lazy_mmu_mode();
2255 for (; addr < end; addr += PAGE_SIZE, ptep++) {
2256 unsigned long mpfn, pfn;
2264 if (pte_none(pte)) {
2265 mpfn = MIGRATE_PFN_MIGRATE;
2271 if (!pte_present(pte)) {
2275 * Only care about unaddressable device page special
2276 * page table entry. Other special swap entries are not
2277 * migratable, and we ignore regular swapped page.
2279 entry = pte_to_swp_entry(pte);
2280 if (!is_device_private_entry(entry))
2283 page = device_private_entry_to_page(entry);
2284 mpfn = migrate_pfn(page_to_pfn(page))|
2285 MIGRATE_PFN_DEVICE | MIGRATE_PFN_MIGRATE;
2286 if (is_write_device_private_entry(entry))
2287 mpfn |= MIGRATE_PFN_WRITE;
2289 if (is_zero_pfn(pfn)) {
2290 mpfn = MIGRATE_PFN_MIGRATE;
2295 page = _vm_normal_page(migrate->vma, addr, pte, true);
2296 mpfn = migrate_pfn(pfn) | MIGRATE_PFN_MIGRATE;
2297 mpfn |= pte_write(pte) ? MIGRATE_PFN_WRITE : 0;
2300 /* FIXME support THP */
2301 if (!page || !page->mapping || PageTransCompound(page)) {
2305 pfn = page_to_pfn(page);
2308 * By getting a reference on the page we pin it and that blocks
2309 * any kind of migration. Side effect is that it "freezes" the
2312 * We drop this reference after isolating the page from the lru
2313 * for non device page (device page are not on the lru and thus
2314 * can't be dropped from it).
2320 * Optimize for the common case where page is only mapped once
2321 * in one process. If we can lock the page, then we can safely
2322 * set up a special migration page table entry now.
2324 if (trylock_page(page)) {
2327 mpfn |= MIGRATE_PFN_LOCKED;
2328 ptep_get_and_clear(mm, addr, ptep);
2330 /* Setup special migration page table entry */
2331 entry = make_migration_entry(page, mpfn &
2333 swp_pte = swp_entry_to_pte(entry);
2334 if (pte_soft_dirty(pte))
2335 swp_pte = pte_swp_mksoft_dirty(swp_pte);
2336 set_pte_at(mm, addr, ptep, swp_pte);
2339 * This is like regular unmap: we remove the rmap and
2340 * drop page refcount. Page won't be freed, as we took
2341 * a reference just above.
2343 page_remove_rmap(page, false);
2346 if (pte_present(pte))
2351 migrate->dst[migrate->npages] = 0;
2352 migrate->src[migrate->npages++] = mpfn;
2354 arch_leave_lazy_mmu_mode();
2355 pte_unmap_unlock(ptep - 1, ptl);
2357 /* Only flush the TLB if we actually modified any entries */
2359 flush_tlb_range(walk->vma, start, end);
2365 * migrate_vma_collect() - collect pages over a range of virtual addresses
2366 * @migrate: migrate struct containing all migration information
2368 * This will walk the CPU page table. For each virtual address backed by a
2369 * valid page, it updates the src array and takes a reference on the page, in
2370 * order to pin the page until we lock it and unmap it.
2372 static void migrate_vma_collect(struct migrate_vma *migrate)
2374 struct mm_walk mm_walk = {
2375 .pmd_entry = migrate_vma_collect_pmd,
2376 .pte_hole = migrate_vma_collect_hole,
2377 .vma = migrate->vma,
2378 .mm = migrate->vma->vm_mm,
2382 mmu_notifier_invalidate_range_start(mm_walk.mm,
2385 walk_page_range(migrate->start, migrate->end, &mm_walk);
2386 mmu_notifier_invalidate_range_end(mm_walk.mm,
2390 migrate->end = migrate->start + (migrate->npages << PAGE_SHIFT);
2394 * migrate_vma_check_page() - check if page is pinned or not
2395 * @page: struct page to check
2397 * Pinned pages cannot be migrated. This is the same test as in
2398 * migrate_page_move_mapping(), except that here we allow migration of a
2401 static bool migrate_vma_check_page(struct page *page)
2404 * One extra ref because caller holds an extra reference, either from
2405 * isolate_lru_page() for a regular page, or migrate_vma_collect() for
2411 * FIXME support THP (transparent huge page), it is bit more complex to
2412 * check them than regular pages, because they can be mapped with a pmd
2413 * or with a pte (split pte mapping).
2415 if (PageCompound(page))
2418 /* Page from ZONE_DEVICE have one extra reference */
2419 if (is_zone_device_page(page)) {
2421 * Private page can never be pin as they have no valid pte and
2422 * GUP will fail for those. Yet if there is a pending migration
2423 * a thread might try to wait on the pte migration entry and
2424 * will bump the page reference count. Sadly there is no way to
2425 * differentiate a regular pin from migration wait. Hence to
2426 * avoid 2 racing thread trying to migrate back to CPU to enter
2427 * infinite loop (one stoping migration because the other is
2428 * waiting on pte migration entry). We always return true here.
2430 * FIXME proper solution is to rework migration_entry_wait() so
2431 * it does not need to take a reference on page.
2433 if (is_device_private_page(page))
2437 * Only allow device public page to be migrated and account for
2438 * the extra reference count imply by ZONE_DEVICE pages.
2440 if (!is_device_public_page(page))
2445 /* For file back page */
2446 if (page_mapping(page))
2447 extra += 1 + page_has_private(page);
2449 if ((page_count(page) - extra) > page_mapcount(page))
2456 * migrate_vma_prepare() - lock pages and isolate them from the lru
2457 * @migrate: migrate struct containing all migration information
2459 * This locks pages that have been collected by migrate_vma_collect(). Once each
2460 * page is locked it is isolated from the lru (for non-device pages). Finally,
2461 * the ref taken by migrate_vma_collect() is dropped, as locked pages cannot be
2462 * migrated by concurrent kernel threads.
2464 static void migrate_vma_prepare(struct migrate_vma *migrate)
2466 const unsigned long npages = migrate->npages;
2467 const unsigned long start = migrate->start;
2468 unsigned long addr, i, restore = 0;
2469 bool allow_drain = true;
2473 for (i = 0; (i < npages) && migrate->cpages; i++) {
2474 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2480 if (!(migrate->src[i] & MIGRATE_PFN_LOCKED)) {
2482 * Because we are migrating several pages there can be
2483 * a deadlock between 2 concurrent migration where each
2484 * are waiting on each other page lock.
2486 * Make migrate_vma() a best effort thing and backoff
2487 * for any page we can not lock right away.
2489 if (!trylock_page(page)) {
2490 migrate->src[i] = 0;
2496 migrate->src[i] |= MIGRATE_PFN_LOCKED;
2499 /* ZONE_DEVICE pages are not on LRU */
2500 if (!is_zone_device_page(page)) {
2501 if (!PageLRU(page) && allow_drain) {
2502 /* Drain CPU's pagevec */
2503 lru_add_drain_all();
2504 allow_drain = false;
2507 if (isolate_lru_page(page)) {
2509 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2513 migrate->src[i] = 0;
2521 /* Drop the reference we took in collect */
2525 if (!migrate_vma_check_page(page)) {
2527 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2531 if (!is_zone_device_page(page)) {
2533 putback_lru_page(page);
2536 migrate->src[i] = 0;
2540 if (!is_zone_device_page(page))
2541 putback_lru_page(page);
2548 for (i = 0, addr = start; i < npages && restore; i++, addr += PAGE_SIZE) {
2549 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2551 if (!page || (migrate->src[i] & MIGRATE_PFN_MIGRATE))
2554 remove_migration_pte(page, migrate->vma, addr, page);
2556 migrate->src[i] = 0;
2564 * migrate_vma_unmap() - replace page mapping with special migration pte entry
2565 * @migrate: migrate struct containing all migration information
2567 * Replace page mapping (CPU page table pte) with a special migration pte entry
2568 * and check again if it has been pinned. Pinned pages are restored because we
2569 * cannot migrate them.
2571 * This is the last step before we call the device driver callback to allocate
2572 * destination memory and copy contents of original page over to new page.
2574 static void migrate_vma_unmap(struct migrate_vma *migrate)
2576 int flags = TTU_MIGRATION | TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS;
2577 const unsigned long npages = migrate->npages;
2578 const unsigned long start = migrate->start;
2579 unsigned long addr, i, restore = 0;
2581 for (i = 0; i < npages; i++) {
2582 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2584 if (!page || !(migrate->src[i] & MIGRATE_PFN_MIGRATE))
2587 if (page_mapped(page)) {
2588 try_to_unmap(page, flags);
2589 if (page_mapped(page))
2593 if (migrate_vma_check_page(page))
2597 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2602 for (addr = start, i = 0; i < npages && restore; addr += PAGE_SIZE, i++) {
2603 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2605 if (!page || (migrate->src[i] & MIGRATE_PFN_MIGRATE))
2608 remove_migration_ptes(page, page, false);
2610 migrate->src[i] = 0;
2614 if (is_zone_device_page(page))
2617 putback_lru_page(page);
2621 static void migrate_vma_insert_page(struct migrate_vma *migrate,
2627 struct vm_area_struct *vma = migrate->vma;
2628 struct mm_struct *mm = vma->vm_mm;
2629 struct mem_cgroup *memcg;
2639 /* Only allow populating anonymous memory */
2640 if (!vma_is_anonymous(vma))
2643 pgdp = pgd_offset(mm, addr);
2644 p4dp = p4d_alloc(mm, pgdp, addr);
2647 pudp = pud_alloc(mm, p4dp, addr);
2650 pmdp = pmd_alloc(mm, pudp, addr);
2654 if (pmd_trans_huge(*pmdp) || pmd_devmap(*pmdp))
2658 * Use pte_alloc() instead of pte_alloc_map(). We can't run
2659 * pte_offset_map() on pmds where a huge pmd might be created
2660 * from a different thread.
2662 * pte_alloc_map() is safe to use under down_write(mmap_sem) or when
2663 * parallel threads are excluded by other means.
2665 * Here we only have down_read(mmap_sem).
2667 if (pte_alloc(mm, pmdp, addr))
2670 /* See the comment in pte_alloc_one_map() */
2671 if (unlikely(pmd_trans_unstable(pmdp)))
2674 if (unlikely(anon_vma_prepare(vma)))
2676 if (mem_cgroup_try_charge(page, vma->vm_mm, GFP_KERNEL, &memcg, false))
2680 * The memory barrier inside __SetPageUptodate makes sure that
2681 * preceding stores to the page contents become visible before
2682 * the set_pte_at() write.
2684 __SetPageUptodate(page);
2686 if (is_zone_device_page(page)) {
2687 if (is_device_private_page(page)) {
2688 swp_entry_t swp_entry;
2690 swp_entry = make_device_private_entry(page, vma->vm_flags & VM_WRITE);
2691 entry = swp_entry_to_pte(swp_entry);
2692 } else if (is_device_public_page(page)) {
2693 entry = pte_mkold(mk_pte(page, READ_ONCE(vma->vm_page_prot)));
2694 if (vma->vm_flags & VM_WRITE)
2695 entry = pte_mkwrite(pte_mkdirty(entry));
2696 entry = pte_mkdevmap(entry);
2699 entry = mk_pte(page, vma->vm_page_prot);
2700 if (vma->vm_flags & VM_WRITE)
2701 entry = pte_mkwrite(pte_mkdirty(entry));
2704 ptep = pte_offset_map_lock(mm, pmdp, addr, &ptl);
2706 if (pte_present(*ptep)) {
2707 unsigned long pfn = pte_pfn(*ptep);
2709 if (!is_zero_pfn(pfn)) {
2710 pte_unmap_unlock(ptep, ptl);
2711 mem_cgroup_cancel_charge(page, memcg, false);
2715 } else if (!pte_none(*ptep)) {
2716 pte_unmap_unlock(ptep, ptl);
2717 mem_cgroup_cancel_charge(page, memcg, false);
2722 * Check for usefaultfd but do not deliver the fault. Instead,
2725 if (userfaultfd_missing(vma)) {
2726 pte_unmap_unlock(ptep, ptl);
2727 mem_cgroup_cancel_charge(page, memcg, false);
2731 inc_mm_counter(mm, MM_ANONPAGES);
2732 page_add_new_anon_rmap(page, vma, addr, false);
2733 mem_cgroup_commit_charge(page, memcg, false, false);
2734 if (!is_zone_device_page(page))
2735 lru_cache_add_active_or_unevictable(page, vma);
2739 flush_cache_page(vma, addr, pte_pfn(*ptep));
2740 ptep_clear_flush_notify(vma, addr, ptep);
2741 set_pte_at_notify(mm, addr, ptep, entry);
2742 update_mmu_cache(vma, addr, ptep);
2744 /* No need to invalidate - it was non-present before */
2745 set_pte_at(mm, addr, ptep, entry);
2746 update_mmu_cache(vma, addr, ptep);
2749 pte_unmap_unlock(ptep, ptl);
2750 *src = MIGRATE_PFN_MIGRATE;
2754 *src &= ~MIGRATE_PFN_MIGRATE;
2758 * migrate_vma_pages() - migrate meta-data from src page to dst page
2759 * @migrate: migrate struct containing all migration information
2761 * This migrates struct page meta-data from source struct page to destination
2762 * struct page. This effectively finishes the migration from source page to the
2765 static void migrate_vma_pages(struct migrate_vma *migrate)
2767 const unsigned long npages = migrate->npages;
2768 const unsigned long start = migrate->start;
2769 struct vm_area_struct *vma = migrate->vma;
2770 struct mm_struct *mm = vma->vm_mm;
2771 unsigned long addr, i, mmu_start;
2772 bool notified = false;
2774 for (i = 0, addr = start; i < npages; addr += PAGE_SIZE, i++) {
2775 struct page *newpage = migrate_pfn_to_page(migrate->dst[i]);
2776 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2777 struct address_space *mapping;
2781 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2786 if (!(migrate->src[i] & MIGRATE_PFN_MIGRATE)) {
2792 mmu_notifier_invalidate_range_start(mm,
2796 migrate_vma_insert_page(migrate, addr, newpage,
2802 mapping = page_mapping(page);
2804 if (is_zone_device_page(newpage)) {
2805 if (is_device_private_page(newpage)) {
2807 * For now only support private anonymous when
2808 * migrating to un-addressable device memory.
2811 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2814 } else if (!is_device_public_page(newpage)) {
2816 * Other types of ZONE_DEVICE page are not
2819 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2824 r = migrate_page(mapping, newpage, page, MIGRATE_SYNC_NO_COPY);
2825 if (r != MIGRATEPAGE_SUCCESS)
2826 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2830 * No need to double call mmu_notifier->invalidate_range() callback as
2831 * the above ptep_clear_flush_notify() inside migrate_vma_insert_page()
2832 * did already call it.
2835 mmu_notifier_invalidate_range_only_end(mm, mmu_start,
2840 * migrate_vma_finalize() - restore CPU page table entry
2841 * @migrate: migrate struct containing all migration information
2843 * This replaces the special migration pte entry with either a mapping to the
2844 * new page if migration was successful for that page, or to the original page
2847 * This also unlocks the pages and puts them back on the lru, or drops the extra
2848 * refcount, for device pages.
2850 static void migrate_vma_finalize(struct migrate_vma *migrate)
2852 const unsigned long npages = migrate->npages;
2855 for (i = 0; i < npages; i++) {
2856 struct page *newpage = migrate_pfn_to_page(migrate->dst[i]);
2857 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2861 unlock_page(newpage);
2867 if (!(migrate->src[i] & MIGRATE_PFN_MIGRATE) || !newpage) {
2869 unlock_page(newpage);
2875 remove_migration_ptes(page, newpage, false);
2879 if (is_zone_device_page(page))
2882 putback_lru_page(page);
2884 if (newpage != page) {
2885 unlock_page(newpage);
2886 if (is_zone_device_page(newpage))
2889 putback_lru_page(newpage);
2895 * migrate_vma() - migrate a range of memory inside vma
2897 * @ops: migration callback for allocating destination memory and copying
2898 * @vma: virtual memory area containing the range to be migrated
2899 * @start: start address of the range to migrate (inclusive)
2900 * @end: end address of the range to migrate (exclusive)
2901 * @src: array of hmm_pfn_t containing source pfns
2902 * @dst: array of hmm_pfn_t containing destination pfns
2903 * @private: pointer passed back to each of the callback
2904 * Returns: 0 on success, error code otherwise
2906 * This function tries to migrate a range of memory virtual address range, using
2907 * callbacks to allocate and copy memory from source to destination. First it
2908 * collects all the pages backing each virtual address in the range, saving this
2909 * inside the src array. Then it locks those pages and unmaps them. Once the pages
2910 * are locked and unmapped, it checks whether each page is pinned or not. Pages
2911 * that aren't pinned have the MIGRATE_PFN_MIGRATE flag set (by this function)
2912 * in the corresponding src array entry. It then restores any pages that are
2913 * pinned, by remapping and unlocking those pages.
2915 * At this point it calls the alloc_and_copy() callback. For documentation on
2916 * what is expected from that callback, see struct migrate_vma_ops comments in
2917 * include/linux/migrate.h
2919 * After the alloc_and_copy() callback, this function goes over each entry in
2920 * the src array that has the MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE flag
2921 * set. If the corresponding entry in dst array has MIGRATE_PFN_VALID flag set,
2922 * then the function tries to migrate struct page information from the source
2923 * struct page to the destination struct page. If it fails to migrate the struct
2924 * page information, then it clears the MIGRATE_PFN_MIGRATE flag in the src
2927 * At this point all successfully migrated pages have an entry in the src
2928 * array with MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE flag set and the dst
2929 * array entry with MIGRATE_PFN_VALID flag set.
2931 * It then calls the finalize_and_map() callback. See comments for "struct
2932 * migrate_vma_ops", in include/linux/migrate.h for details about
2933 * finalize_and_map() behavior.
2935 * After the finalize_and_map() callback, for successfully migrated pages, this
2936 * function updates the CPU page table to point to new pages, otherwise it
2937 * restores the CPU page table to point to the original source pages.
2939 * Function returns 0 after the above steps, even if no pages were migrated
2940 * (The function only returns an error if any of the arguments are invalid.)
2942 * Both src and dst array must be big enough for (end - start) >> PAGE_SHIFT
2943 * unsigned long entries.
2945 int migrate_vma(const struct migrate_vma_ops *ops,
2946 struct vm_area_struct *vma,
2947 unsigned long start,
2953 struct migrate_vma migrate;
2955 /* Sanity check the arguments */
2958 if (!vma || is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_SPECIAL) ||
2961 if (start < vma->vm_start || start >= vma->vm_end)
2963 if (end <= vma->vm_start || end > vma->vm_end)
2965 if (!ops || !src || !dst || start >= end)
2968 memset(src, 0, sizeof(*src) * ((end - start) >> PAGE_SHIFT));
2971 migrate.start = start;
2977 /* Collect, and try to unmap source pages */
2978 migrate_vma_collect(&migrate);
2979 if (!migrate.cpages)
2982 /* Lock and isolate page */
2983 migrate_vma_prepare(&migrate);
2984 if (!migrate.cpages)
2988 migrate_vma_unmap(&migrate);
2989 if (!migrate.cpages)
2993 * At this point pages are locked and unmapped, and thus they have
2994 * stable content and can safely be copied to destination memory that
2995 * is allocated by the callback.
2997 * Note that migration can fail in migrate_vma_struct_page() for each
3000 ops->alloc_and_copy(vma, src, dst, start, end, private);
3002 /* This does the real migration of struct page */
3003 migrate_vma_pages(&migrate);
3005 ops->finalize_and_map(vma, src, dst, start, end, private);
3007 /* Unlock and remap pages */
3008 migrate_vma_finalize(&migrate);
3012 EXPORT_SYMBOL(migrate_vma);
3013 #endif /* defined(MIGRATE_VMA_HELPER) */