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:
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);
251 flush_dcache_page(new);
254 flush_dcache_page(new);
256 #ifdef CONFIG_HUGETLB_PAGE
258 pte = pte_mkhuge(pte);
259 pte = arch_make_huge_pte(pte, vma, new, 0);
260 set_huge_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte);
262 hugepage_add_anon_rmap(new, vma, pvmw.address);
264 page_dup_rmap(new, true);
268 set_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte);
271 page_add_anon_rmap(new, vma, pvmw.address, false);
273 page_add_file_rmap(new, false);
275 if (vma->vm_flags & VM_LOCKED && !PageTransCompound(new))
278 if (PageTransHuge(page) && PageMlocked(page))
279 clear_page_mlock(page);
281 /* No need to invalidate - it was non-present before */
282 update_mmu_cache(vma, pvmw.address, pvmw.pte);
289 * Get rid of all migration entries and replace them by
290 * references to the indicated page.
292 void remove_migration_ptes(struct page *old, struct page *new, bool locked)
294 struct rmap_walk_control rwc = {
295 .rmap_one = remove_migration_pte,
300 rmap_walk_locked(new, &rwc);
302 rmap_walk(new, &rwc);
306 * Something used the pte of a page under migration. We need to
307 * get to the page and wait until migration is finished.
308 * When we return from this function the fault will be retried.
310 void __migration_entry_wait(struct mm_struct *mm, pte_t *ptep,
319 if (!is_swap_pte(pte))
322 entry = pte_to_swp_entry(pte);
323 if (!is_migration_entry(entry))
326 page = migration_entry_to_page(entry);
329 * Once page cache replacement of page migration started, page_count
330 * is zero; but we must not call put_and_wait_on_page_locked() without
331 * a ref. Use get_page_unless_zero(), and just fault again if it fails.
333 if (!get_page_unless_zero(page))
335 pte_unmap_unlock(ptep, ptl);
336 put_and_wait_on_page_locked(page);
339 pte_unmap_unlock(ptep, ptl);
342 void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
343 unsigned long address)
345 spinlock_t *ptl = pte_lockptr(mm, pmd);
346 pte_t *ptep = pte_offset_map(pmd, address);
347 __migration_entry_wait(mm, ptep, ptl);
350 void migration_entry_wait_huge(struct vm_area_struct *vma,
351 struct mm_struct *mm, pte_t *pte)
353 spinlock_t *ptl = huge_pte_lockptr(hstate_vma(vma), mm, pte);
354 __migration_entry_wait(mm, pte, ptl);
357 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
358 void pmd_migration_entry_wait(struct mm_struct *mm, pmd_t *pmd)
363 ptl = pmd_lock(mm, pmd);
364 if (!is_pmd_migration_entry(*pmd))
366 page = migration_entry_to_page(pmd_to_swp_entry(*pmd));
367 if (!get_page_unless_zero(page))
370 put_and_wait_on_page_locked(page);
378 /* Returns true if all buffers are successfully locked */
379 static bool buffer_migrate_lock_buffers(struct buffer_head *head,
380 enum migrate_mode mode)
382 struct buffer_head *bh = head;
384 /* Simple case, sync compaction */
385 if (mode != MIGRATE_ASYNC) {
389 bh = bh->b_this_page;
391 } while (bh != head);
396 /* async case, we cannot block on lock_buffer so use trylock_buffer */
399 if (!trylock_buffer(bh)) {
401 * We failed to lock the buffer and cannot stall in
402 * async migration. Release the taken locks
404 struct buffer_head *failed_bh = bh;
407 while (bh != failed_bh) {
410 bh = bh->b_this_page;
415 bh = bh->b_this_page;
416 } while (bh != head);
420 static inline bool buffer_migrate_lock_buffers(struct buffer_head *head,
421 enum migrate_mode mode)
425 #endif /* CONFIG_BLOCK */
428 * Replace the page in the mapping.
430 * The number of remaining references must be:
431 * 1 for anonymous pages without a mapping
432 * 2 for pages with a mapping
433 * 3 for pages with a mapping and PagePrivate/PagePrivate2 set.
435 int migrate_page_move_mapping(struct address_space *mapping,
436 struct page *newpage, struct page *page,
437 struct buffer_head *head, enum migrate_mode mode,
440 XA_STATE(xas, &mapping->i_pages, page_index(page));
441 struct zone *oldzone, *newzone;
443 int expected_count = 1 + extra_count;
446 * Device public or private pages have an extra refcount as they are
449 expected_count += is_device_private_page(page);
450 expected_count += is_device_public_page(page);
453 /* Anonymous page without mapping */
454 if (page_count(page) != expected_count)
457 /* No turning back from here */
458 newpage->index = page->index;
459 newpage->mapping = page->mapping;
460 if (PageSwapBacked(page))
461 __SetPageSwapBacked(newpage);
463 return MIGRATEPAGE_SUCCESS;
466 oldzone = page_zone(page);
467 newzone = page_zone(newpage);
471 expected_count += hpage_nr_pages(page) + page_has_private(page);
472 if (page_count(page) != expected_count || xas_load(&xas) != page) {
473 xas_unlock_irq(&xas);
477 if (!page_ref_freeze(page, expected_count)) {
478 xas_unlock_irq(&xas);
483 * In the async migration case of moving a page with buffers, lock the
484 * buffers using trylock before the mapping is moved. If the mapping
485 * was moved, we later failed to lock the buffers and could not move
486 * the mapping back due to an elevated page count, we would have to
487 * block waiting on other references to be dropped.
489 if (mode == MIGRATE_ASYNC && head &&
490 !buffer_migrate_lock_buffers(head, mode)) {
491 page_ref_unfreeze(page, expected_count);
492 xas_unlock_irq(&xas);
497 * Now we know that no one else is looking at the page:
498 * no turning back from here.
500 newpage->index = page->index;
501 newpage->mapping = page->mapping;
502 page_ref_add(newpage, hpage_nr_pages(page)); /* add cache reference */
503 if (PageSwapBacked(page)) {
504 __SetPageSwapBacked(newpage);
505 if (PageSwapCache(page)) {
506 SetPageSwapCache(newpage);
507 set_page_private(newpage, page_private(page));
510 VM_BUG_ON_PAGE(PageSwapCache(page), page);
513 /* Move dirty while page refs frozen and newpage not yet exposed */
514 dirty = PageDirty(page);
516 ClearPageDirty(page);
517 SetPageDirty(newpage);
520 xas_store(&xas, newpage);
521 if (PageTransHuge(page)) {
524 for (i = 1; i < HPAGE_PMD_NR; i++) {
526 xas_store(&xas, newpage + i);
531 * Drop cache reference from old page by unfreezing
532 * to one less reference.
533 * We know this isn't the last reference.
535 page_ref_unfreeze(page, expected_count - hpage_nr_pages(page));
538 /* Leave irq disabled to prevent preemption while updating stats */
541 * If moved to a different zone then also account
542 * the page for that zone. Other VM counters will be
543 * taken care of when we establish references to the
544 * new page and drop references to the old page.
546 * Note that anonymous pages are accounted for
547 * via NR_FILE_PAGES and NR_ANON_MAPPED if they
548 * are mapped to swap space.
550 if (newzone != oldzone) {
551 __dec_node_state(oldzone->zone_pgdat, NR_FILE_PAGES);
552 __inc_node_state(newzone->zone_pgdat, NR_FILE_PAGES);
553 if (PageSwapBacked(page) && !PageSwapCache(page)) {
554 __dec_node_state(oldzone->zone_pgdat, NR_SHMEM);
555 __inc_node_state(newzone->zone_pgdat, NR_SHMEM);
557 if (dirty && mapping_cap_account_dirty(mapping)) {
558 __dec_node_state(oldzone->zone_pgdat, NR_FILE_DIRTY);
559 __dec_zone_state(oldzone, NR_ZONE_WRITE_PENDING);
560 __inc_node_state(newzone->zone_pgdat, NR_FILE_DIRTY);
561 __inc_zone_state(newzone, NR_ZONE_WRITE_PENDING);
566 return MIGRATEPAGE_SUCCESS;
568 EXPORT_SYMBOL(migrate_page_move_mapping);
571 * The expected number of remaining references is the same as that
572 * of migrate_page_move_mapping().
574 int migrate_huge_page_move_mapping(struct address_space *mapping,
575 struct page *newpage, struct page *page)
577 XA_STATE(xas, &mapping->i_pages, page_index(page));
581 expected_count = 2 + page_has_private(page);
582 if (page_count(page) != expected_count || xas_load(&xas) != page) {
583 xas_unlock_irq(&xas);
587 if (!page_ref_freeze(page, expected_count)) {
588 xas_unlock_irq(&xas);
592 newpage->index = page->index;
593 newpage->mapping = page->mapping;
597 xas_store(&xas, newpage);
599 page_ref_unfreeze(page, expected_count - 1);
601 xas_unlock_irq(&xas);
603 return MIGRATEPAGE_SUCCESS;
607 * Gigantic pages are so large that we do not guarantee that page++ pointer
608 * arithmetic will work across the entire page. We need something more
611 static void __copy_gigantic_page(struct page *dst, struct page *src,
615 struct page *dst_base = dst;
616 struct page *src_base = src;
618 for (i = 0; i < nr_pages; ) {
620 copy_highpage(dst, src);
623 dst = mem_map_next(dst, dst_base, i);
624 src = mem_map_next(src, src_base, i);
628 static void copy_huge_page(struct page *dst, struct page *src)
635 struct hstate *h = page_hstate(src);
636 nr_pages = pages_per_huge_page(h);
638 if (unlikely(nr_pages > MAX_ORDER_NR_PAGES)) {
639 __copy_gigantic_page(dst, src, nr_pages);
644 BUG_ON(!PageTransHuge(src));
645 nr_pages = hpage_nr_pages(src);
648 for (i = 0; i < nr_pages; i++) {
650 copy_highpage(dst + i, src + i);
655 * Copy the page to its new location
657 void migrate_page_states(struct page *newpage, struct page *page)
662 SetPageError(newpage);
663 if (PageReferenced(page))
664 SetPageReferenced(newpage);
665 if (PageUptodate(page))
666 SetPageUptodate(newpage);
667 if (TestClearPageActive(page)) {
668 VM_BUG_ON_PAGE(PageUnevictable(page), page);
669 SetPageActive(newpage);
670 } else if (TestClearPageUnevictable(page))
671 SetPageUnevictable(newpage);
672 if (PageWorkingset(page))
673 SetPageWorkingset(newpage);
674 if (PageChecked(page))
675 SetPageChecked(newpage);
676 if (PageMappedToDisk(page))
677 SetPageMappedToDisk(newpage);
679 /* Move dirty on pages not done by migrate_page_move_mapping() */
681 SetPageDirty(newpage);
683 if (page_is_young(page))
684 set_page_young(newpage);
685 if (page_is_idle(page))
686 set_page_idle(newpage);
689 * Copy NUMA information to the new page, to prevent over-eager
690 * future migrations of this same page.
692 cpupid = page_cpupid_xchg_last(page, -1);
693 page_cpupid_xchg_last(newpage, cpupid);
695 ksm_migrate_page(newpage, page);
697 * Please do not reorder this without considering how mm/ksm.c's
698 * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache().
700 if (PageSwapCache(page))
701 ClearPageSwapCache(page);
702 ClearPagePrivate(page);
703 set_page_private(page, 0);
706 * If any waiters have accumulated on the new page then
709 if (PageWriteback(newpage))
710 end_page_writeback(newpage);
712 copy_page_owner(page, newpage);
714 mem_cgroup_migrate(page, newpage);
716 EXPORT_SYMBOL(migrate_page_states);
718 void migrate_page_copy(struct page *newpage, struct page *page)
720 if (PageHuge(page) || PageTransHuge(page))
721 copy_huge_page(newpage, page);
723 copy_highpage(newpage, page);
725 migrate_page_states(newpage, page);
727 EXPORT_SYMBOL(migrate_page_copy);
729 /************************************************************
730 * Migration functions
731 ***********************************************************/
734 * Common logic to directly migrate a single LRU page suitable for
735 * pages that do not use PagePrivate/PagePrivate2.
737 * Pages are locked upon entry and exit.
739 int migrate_page(struct address_space *mapping,
740 struct page *newpage, struct page *page,
741 enum migrate_mode mode)
745 BUG_ON(PageWriteback(page)); /* Writeback must be complete */
747 rc = migrate_page_move_mapping(mapping, newpage, page, NULL, mode, 0);
749 if (rc != MIGRATEPAGE_SUCCESS)
752 if (mode != MIGRATE_SYNC_NO_COPY)
753 migrate_page_copy(newpage, page);
755 migrate_page_states(newpage, page);
756 return MIGRATEPAGE_SUCCESS;
758 EXPORT_SYMBOL(migrate_page);
762 * Migration function for pages with buffers. This function can only be used
763 * if the underlying filesystem guarantees that no other references to "page"
766 int buffer_migrate_page(struct address_space *mapping,
767 struct page *newpage, struct page *page, enum migrate_mode mode)
769 struct buffer_head *bh, *head;
772 if (!page_has_buffers(page))
773 return migrate_page(mapping, newpage, page, mode);
775 head = page_buffers(page);
777 rc = migrate_page_move_mapping(mapping, newpage, page, head, mode, 0);
779 if (rc != MIGRATEPAGE_SUCCESS)
783 * In the async case, migrate_page_move_mapping locked the buffers
784 * with an IRQ-safe spinlock held. In the sync case, the buffers
785 * need to be locked now
787 if (mode != MIGRATE_ASYNC)
788 BUG_ON(!buffer_migrate_lock_buffers(head, mode));
790 ClearPagePrivate(page);
791 set_page_private(newpage, page_private(page));
792 set_page_private(page, 0);
798 set_bh_page(bh, newpage, bh_offset(bh));
799 bh = bh->b_this_page;
801 } while (bh != head);
803 SetPagePrivate(newpage);
805 if (mode != MIGRATE_SYNC_NO_COPY)
806 migrate_page_copy(newpage, page);
808 migrate_page_states(newpage, page);
814 bh = bh->b_this_page;
816 } while (bh != head);
818 return MIGRATEPAGE_SUCCESS;
820 EXPORT_SYMBOL(buffer_migrate_page);
824 * Writeback a page to clean the dirty state
826 static int writeout(struct address_space *mapping, struct page *page)
828 struct writeback_control wbc = {
829 .sync_mode = WB_SYNC_NONE,
832 .range_end = LLONG_MAX,
837 if (!mapping->a_ops->writepage)
838 /* No write method for the address space */
841 if (!clear_page_dirty_for_io(page))
842 /* Someone else already triggered a write */
846 * A dirty page may imply that the underlying filesystem has
847 * the page on some queue. So the page must be clean for
848 * migration. Writeout may mean we loose the lock and the
849 * page state is no longer what we checked for earlier.
850 * At this point we know that the migration attempt cannot
853 remove_migration_ptes(page, page, false);
855 rc = mapping->a_ops->writepage(page, &wbc);
857 if (rc != AOP_WRITEPAGE_ACTIVATE)
858 /* unlocked. Relock */
861 return (rc < 0) ? -EIO : -EAGAIN;
865 * Default handling if a filesystem does not provide a migration function.
867 static int fallback_migrate_page(struct address_space *mapping,
868 struct page *newpage, struct page *page, enum migrate_mode mode)
870 if (PageDirty(page)) {
871 /* Only writeback pages in full synchronous migration */
874 case MIGRATE_SYNC_NO_COPY:
879 return writeout(mapping, page);
883 * Buffers may be managed in a filesystem specific way.
884 * We must have no buffers or drop them.
886 if (page_has_private(page) &&
887 !try_to_release_page(page, GFP_KERNEL))
890 return migrate_page(mapping, newpage, page, mode);
894 * Move a page to a newly allocated page
895 * The page is locked and all ptes have been successfully removed.
897 * The new page will have replaced the old page if this function
902 * MIGRATEPAGE_SUCCESS - success
904 static int move_to_new_page(struct page *newpage, struct page *page,
905 enum migrate_mode mode)
907 struct address_space *mapping;
909 bool is_lru = !__PageMovable(page);
911 VM_BUG_ON_PAGE(!PageLocked(page), page);
912 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
914 mapping = page_mapping(page);
916 if (likely(is_lru)) {
918 rc = migrate_page(mapping, newpage, page, mode);
919 else if (mapping->a_ops->migratepage)
921 * Most pages have a mapping and most filesystems
922 * provide a migratepage callback. Anonymous pages
923 * are part of swap space which also has its own
924 * migratepage callback. This is the most common path
925 * for page migration.
927 rc = mapping->a_ops->migratepage(mapping, newpage,
930 rc = fallback_migrate_page(mapping, newpage,
934 * In case of non-lru page, it could be released after
935 * isolation step. In that case, we shouldn't try migration.
937 VM_BUG_ON_PAGE(!PageIsolated(page), page);
938 if (!PageMovable(page)) {
939 rc = MIGRATEPAGE_SUCCESS;
940 __ClearPageIsolated(page);
944 rc = mapping->a_ops->migratepage(mapping, newpage,
946 WARN_ON_ONCE(rc == MIGRATEPAGE_SUCCESS &&
947 !PageIsolated(page));
951 * When successful, old pagecache page->mapping must be cleared before
952 * page is freed; but stats require that PageAnon be left as PageAnon.
954 if (rc == MIGRATEPAGE_SUCCESS) {
955 if (__PageMovable(page)) {
956 VM_BUG_ON_PAGE(!PageIsolated(page), page);
959 * We clear PG_movable under page_lock so any compactor
960 * cannot try to migrate this page.
962 __ClearPageIsolated(page);
966 * Anonymous and movable page->mapping will be cleard by
967 * free_pages_prepare so don't reset it here for keeping
968 * the type to work PageAnon, for example.
970 if (!PageMappingFlags(page))
971 page->mapping = NULL;
977 static int __unmap_and_move(struct page *page, struct page *newpage,
978 int force, enum migrate_mode mode)
981 int page_was_mapped = 0;
982 struct anon_vma *anon_vma = NULL;
983 bool is_lru = !__PageMovable(page);
985 if (!trylock_page(page)) {
986 if (!force || mode == MIGRATE_ASYNC)
990 * It's not safe for direct compaction to call lock_page.
991 * For example, during page readahead pages are added locked
992 * to the LRU. Later, when the IO completes the pages are
993 * marked uptodate and unlocked. However, the queueing
994 * could be merging multiple pages for one bio (e.g.
995 * mpage_readpages). If an allocation happens for the
996 * second or third page, the process can end up locking
997 * the same page twice and deadlocking. Rather than
998 * trying to be clever about what pages can be locked,
999 * avoid the use of lock_page for direct compaction
1002 if (current->flags & PF_MEMALLOC)
1008 if (PageWriteback(page)) {
1010 * Only in the case of a full synchronous migration is it
1011 * necessary to wait for PageWriteback. In the async case,
1012 * the retry loop is too short and in the sync-light case,
1013 * the overhead of stalling is too much
1017 case MIGRATE_SYNC_NO_COPY:
1025 wait_on_page_writeback(page);
1029 * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
1030 * we cannot notice that anon_vma is freed while we migrates a page.
1031 * This get_anon_vma() delays freeing anon_vma pointer until the end
1032 * of migration. File cache pages are no problem because of page_lock()
1033 * File Caches may use write_page() or lock_page() in migration, then,
1034 * just care Anon page here.
1036 * Only page_get_anon_vma() understands the subtleties of
1037 * getting a hold on an anon_vma from outside one of its mms.
1038 * But if we cannot get anon_vma, then we won't need it anyway,
1039 * because that implies that the anon page is no longer mapped
1040 * (and cannot be remapped so long as we hold the page lock).
1042 if (PageAnon(page) && !PageKsm(page))
1043 anon_vma = page_get_anon_vma(page);
1046 * Block others from accessing the new page when we get around to
1047 * establishing additional references. We are usually the only one
1048 * holding a reference to newpage at this point. We used to have a BUG
1049 * here if trylock_page(newpage) fails, but would like to allow for
1050 * cases where there might be a race with the previous use of newpage.
1051 * This is much like races on refcount of oldpage: just don't BUG().
1053 if (unlikely(!trylock_page(newpage)))
1056 if (unlikely(!is_lru)) {
1057 rc = move_to_new_page(newpage, page, mode);
1058 goto out_unlock_both;
1062 * Corner case handling:
1063 * 1. When a new swap-cache page is read into, it is added to the LRU
1064 * and treated as swapcache but it has no rmap yet.
1065 * Calling try_to_unmap() against a page->mapping==NULL page will
1066 * trigger a BUG. So handle it here.
1067 * 2. An orphaned page (see truncate_complete_page) might have
1068 * fs-private metadata. The page can be picked up due to memory
1069 * offlining. Everywhere else except page reclaim, the page is
1070 * invisible to the vm, so the page can not be migrated. So try to
1071 * free the metadata, so the page can be freed.
1073 if (!page->mapping) {
1074 VM_BUG_ON_PAGE(PageAnon(page), page);
1075 if (page_has_private(page)) {
1076 try_to_free_buffers(page);
1077 goto out_unlock_both;
1079 } else if (page_mapped(page)) {
1080 /* Establish migration ptes */
1081 VM_BUG_ON_PAGE(PageAnon(page) && !PageKsm(page) && !anon_vma,
1084 TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
1085 page_was_mapped = 1;
1088 if (!page_mapped(page))
1089 rc = move_to_new_page(newpage, page, mode);
1091 if (page_was_mapped)
1092 remove_migration_ptes(page,
1093 rc == MIGRATEPAGE_SUCCESS ? newpage : page, false);
1096 unlock_page(newpage);
1098 /* Drop an anon_vma reference if we took one */
1100 put_anon_vma(anon_vma);
1104 * If migration is successful, decrease refcount of the newpage
1105 * which will not free the page because new page owner increased
1106 * refcounter. As well, if it is LRU page, add the page to LRU
1109 if (rc == MIGRATEPAGE_SUCCESS) {
1110 if (unlikely(__PageMovable(newpage)))
1113 putback_lru_page(newpage);
1120 * gcc 4.7 and 4.8 on arm get an ICEs when inlining unmap_and_move(). Work
1123 #if defined(CONFIG_ARM) && \
1124 defined(GCC_VERSION) && GCC_VERSION < 40900 && GCC_VERSION >= 40700
1125 #define ICE_noinline noinline
1127 #define ICE_noinline
1131 * Obtain the lock on page, remove all ptes and migrate the page
1132 * to the newly allocated page in newpage.
1134 static ICE_noinline int unmap_and_move(new_page_t get_new_page,
1135 free_page_t put_new_page,
1136 unsigned long private, struct page *page,
1137 int force, enum migrate_mode mode,
1138 enum migrate_reason reason)
1140 int rc = MIGRATEPAGE_SUCCESS;
1141 struct page *newpage;
1143 if (!thp_migration_supported() && PageTransHuge(page))
1146 newpage = get_new_page(page, private);
1150 if (page_count(page) == 1) {
1151 /* page was freed from under us. So we are done. */
1152 ClearPageActive(page);
1153 ClearPageUnevictable(page);
1154 if (unlikely(__PageMovable(page))) {
1156 if (!PageMovable(page))
1157 __ClearPageIsolated(page);
1161 put_new_page(newpage, private);
1167 rc = __unmap_and_move(page, newpage, force, mode);
1168 if (rc == MIGRATEPAGE_SUCCESS)
1169 set_page_owner_migrate_reason(newpage, reason);
1172 if (rc != -EAGAIN) {
1174 * A page that has been migrated has all references
1175 * removed and will be freed. A page that has not been
1176 * migrated will have kepts its references and be
1179 list_del(&page->lru);
1182 * Compaction can migrate also non-LRU pages which are
1183 * not accounted to NR_ISOLATED_*. They can be recognized
1186 if (likely(!__PageMovable(page)))
1187 mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON +
1188 page_is_file_cache(page), -hpage_nr_pages(page));
1192 * If migration is successful, releases reference grabbed during
1193 * isolation. Otherwise, restore the page to right list unless
1196 if (rc == MIGRATEPAGE_SUCCESS) {
1198 if (reason == MR_MEMORY_FAILURE) {
1200 * Set PG_HWPoison on just freed page
1201 * intentionally. Although it's rather weird,
1202 * it's how HWPoison flag works at the moment.
1204 if (set_hwpoison_free_buddy_page(page))
1205 num_poisoned_pages_inc();
1208 if (rc != -EAGAIN) {
1209 if (likely(!__PageMovable(page))) {
1210 putback_lru_page(page);
1215 if (PageMovable(page))
1216 putback_movable_page(page);
1218 __ClearPageIsolated(page);
1224 put_new_page(newpage, private);
1233 * Counterpart of unmap_and_move_page() for hugepage migration.
1235 * This function doesn't wait the completion of hugepage I/O
1236 * because there is no race between I/O and migration for hugepage.
1237 * Note that currently hugepage I/O occurs only in direct I/O
1238 * where no lock is held and PG_writeback is irrelevant,
1239 * and writeback status of all subpages are counted in the reference
1240 * count of the head page (i.e. if all subpages of a 2MB hugepage are
1241 * under direct I/O, the reference of the head page is 512 and a bit more.)
1242 * This means that when we try to migrate hugepage whose subpages are
1243 * doing direct I/O, some references remain after try_to_unmap() and
1244 * hugepage migration fails without data corruption.
1246 * There is also no race when direct I/O is issued on the page under migration,
1247 * because then pte is replaced with migration swap entry and direct I/O code
1248 * will wait in the page fault for migration to complete.
1250 static int unmap_and_move_huge_page(new_page_t get_new_page,
1251 free_page_t put_new_page, unsigned long private,
1252 struct page *hpage, int force,
1253 enum migrate_mode mode, int reason)
1256 int page_was_mapped = 0;
1257 struct page *new_hpage;
1258 struct anon_vma *anon_vma = NULL;
1261 * Movability of hugepages depends on architectures and hugepage size.
1262 * This check is necessary because some callers of hugepage migration
1263 * like soft offline and memory hotremove don't walk through page
1264 * tables or check whether the hugepage is pmd-based or not before
1265 * kicking migration.
1267 if (!hugepage_migration_supported(page_hstate(hpage))) {
1268 putback_active_hugepage(hpage);
1272 new_hpage = get_new_page(hpage, private);
1276 if (!trylock_page(hpage)) {
1281 case MIGRATE_SYNC_NO_COPY:
1289 if (PageAnon(hpage))
1290 anon_vma = page_get_anon_vma(hpage);
1292 if (unlikely(!trylock_page(new_hpage)))
1295 if (page_mapped(hpage)) {
1297 TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
1298 page_was_mapped = 1;
1301 if (!page_mapped(hpage))
1302 rc = move_to_new_page(new_hpage, hpage, mode);
1304 if (page_was_mapped)
1305 remove_migration_ptes(hpage,
1306 rc == MIGRATEPAGE_SUCCESS ? new_hpage : hpage, false);
1308 unlock_page(new_hpage);
1312 put_anon_vma(anon_vma);
1314 if (rc == MIGRATEPAGE_SUCCESS) {
1315 move_hugetlb_state(hpage, new_hpage, reason);
1316 put_new_page = NULL;
1322 putback_active_hugepage(hpage);
1325 * If migration was not successful and there's a freeing callback, use
1326 * it. Otherwise, put_page() will drop the reference grabbed during
1330 put_new_page(new_hpage, private);
1332 putback_active_hugepage(new_hpage);
1338 * migrate_pages - migrate the pages specified in a list, to the free pages
1339 * supplied as the target for the page migration
1341 * @from: The list of pages to be migrated.
1342 * @get_new_page: The function used to allocate free pages to be used
1343 * as the target of the page migration.
1344 * @put_new_page: The function used to free target pages if migration
1345 * fails, or NULL if no special handling is necessary.
1346 * @private: Private data to be passed on to get_new_page()
1347 * @mode: The migration mode that specifies the constraints for
1348 * page migration, if any.
1349 * @reason: The reason for page migration.
1351 * The function returns after 10 attempts or if no pages are movable any more
1352 * because the list has become empty or no retryable pages exist any more.
1353 * The caller should call putback_movable_pages() to return pages to the LRU
1354 * or free list only if ret != 0.
1356 * Returns the number of pages that were not migrated, or an error code.
1358 int migrate_pages(struct list_head *from, new_page_t get_new_page,
1359 free_page_t put_new_page, unsigned long private,
1360 enum migrate_mode mode, int reason)
1364 int nr_succeeded = 0;
1368 int swapwrite = current->flags & PF_SWAPWRITE;
1372 current->flags |= PF_SWAPWRITE;
1374 for(pass = 0; pass < 10 && retry; pass++) {
1377 list_for_each_entry_safe(page, page2, from, lru) {
1382 rc = unmap_and_move_huge_page(get_new_page,
1383 put_new_page, private, page,
1384 pass > 2, mode, reason);
1386 rc = unmap_and_move(get_new_page, put_new_page,
1387 private, page, pass > 2, mode,
1393 * THP migration might be unsupported or the
1394 * allocation could've failed so we should
1395 * retry on the same page with the THP split
1398 * Head page is retried immediately and tail
1399 * pages are added to the tail of the list so
1400 * we encounter them after the rest of the list
1403 if (PageTransHuge(page) && !PageHuge(page)) {
1405 rc = split_huge_page_to_list(page, from);
1408 list_safe_reset_next(page, page2, lru);
1417 case MIGRATEPAGE_SUCCESS:
1422 * Permanent failure (-EBUSY, -ENOSYS, etc.):
1423 * unlike -EAGAIN case, the failed page is
1424 * removed from migration page list and not
1425 * retried in the next outer loop.
1436 count_vm_events(PGMIGRATE_SUCCESS, nr_succeeded);
1438 count_vm_events(PGMIGRATE_FAIL, nr_failed);
1439 trace_mm_migrate_pages(nr_succeeded, nr_failed, mode, reason);
1442 current->flags &= ~PF_SWAPWRITE;
1449 static int store_status(int __user *status, int start, int value, int nr)
1452 if (put_user(value, status + start))
1460 static int do_move_pages_to_node(struct mm_struct *mm,
1461 struct list_head *pagelist, int node)
1465 if (list_empty(pagelist))
1468 err = migrate_pages(pagelist, alloc_new_node_page, NULL, node,
1469 MIGRATE_SYNC, MR_SYSCALL);
1471 putback_movable_pages(pagelist);
1476 * Resolves the given address to a struct page, isolates it from the LRU and
1477 * puts it to the given pagelist.
1478 * Returns -errno if the page cannot be found/isolated or 0 when it has been
1479 * queued or the page doesn't need to be migrated because it is already on
1482 static int add_page_for_migration(struct mm_struct *mm, unsigned long addr,
1483 int node, struct list_head *pagelist, bool migrate_all)
1485 struct vm_area_struct *vma;
1487 unsigned int follflags;
1490 down_read(&mm->mmap_sem);
1492 vma = find_vma(mm, addr);
1493 if (!vma || addr < vma->vm_start || !vma_migratable(vma))
1496 /* FOLL_DUMP to ignore special (like zero) pages */
1497 follflags = FOLL_GET | FOLL_DUMP;
1498 page = follow_page(vma, addr, follflags);
1500 err = PTR_ERR(page);
1509 if (page_to_nid(page) == node)
1513 if (page_mapcount(page) > 1 && !migrate_all)
1516 if (PageHuge(page)) {
1517 if (PageHead(page)) {
1518 isolate_huge_page(page, pagelist);
1524 head = compound_head(page);
1525 err = isolate_lru_page(head);
1530 list_add_tail(&head->lru, pagelist);
1531 mod_node_page_state(page_pgdat(head),
1532 NR_ISOLATED_ANON + page_is_file_cache(head),
1533 hpage_nr_pages(head));
1537 * Either remove the duplicate refcount from
1538 * isolate_lru_page() or drop the page ref if it was
1543 up_read(&mm->mmap_sem);
1548 * Migrate an array of page address onto an array of nodes and fill
1549 * the corresponding array of status.
1551 static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes,
1552 unsigned long nr_pages,
1553 const void __user * __user *pages,
1554 const int __user *nodes,
1555 int __user *status, int flags)
1557 int current_node = NUMA_NO_NODE;
1558 LIST_HEAD(pagelist);
1564 for (i = start = 0; i < nr_pages; i++) {
1565 const void __user *p;
1570 if (get_user(p, pages + i))
1572 if (get_user(node, nodes + i))
1574 addr = (unsigned long)p;
1577 if (node < 0 || node >= MAX_NUMNODES)
1579 if (!node_state(node, N_MEMORY))
1583 if (!node_isset(node, task_nodes))
1586 if (current_node == NUMA_NO_NODE) {
1587 current_node = node;
1589 } else if (node != current_node) {
1590 err = do_move_pages_to_node(mm, &pagelist, current_node);
1593 err = store_status(status, start, current_node, i - start);
1597 current_node = node;
1601 * Errors in the page lookup or isolation are not fatal and we simply
1602 * report them via status
1604 err = add_page_for_migration(mm, addr, current_node,
1605 &pagelist, flags & MPOL_MF_MOVE_ALL);
1609 err = store_status(status, i, err, 1);
1613 err = do_move_pages_to_node(mm, &pagelist, current_node);
1617 err = store_status(status, start, current_node, i - start);
1621 current_node = NUMA_NO_NODE;
1624 if (list_empty(&pagelist))
1627 /* Make sure we do not overwrite the existing error */
1628 err1 = do_move_pages_to_node(mm, &pagelist, current_node);
1630 err1 = store_status(status, start, current_node, i - start);
1638 * Determine the nodes of an array of pages and store it in an array of status.
1640 static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages,
1641 const void __user **pages, int *status)
1645 down_read(&mm->mmap_sem);
1647 for (i = 0; i < nr_pages; i++) {
1648 unsigned long addr = (unsigned long)(*pages);
1649 struct vm_area_struct *vma;
1653 vma = find_vma(mm, addr);
1654 if (!vma || addr < vma->vm_start)
1657 /* FOLL_DUMP to ignore special (like zero) pages */
1658 page = follow_page(vma, addr, FOLL_DUMP);
1660 err = PTR_ERR(page);
1664 err = page ? page_to_nid(page) : -ENOENT;
1672 up_read(&mm->mmap_sem);
1676 * Determine the nodes of a user array of pages and store it in
1677 * a user array of status.
1679 static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages,
1680 const void __user * __user *pages,
1683 #define DO_PAGES_STAT_CHUNK_NR 16
1684 const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR];
1685 int chunk_status[DO_PAGES_STAT_CHUNK_NR];
1688 unsigned long chunk_nr;
1690 chunk_nr = nr_pages;
1691 if (chunk_nr > DO_PAGES_STAT_CHUNK_NR)
1692 chunk_nr = DO_PAGES_STAT_CHUNK_NR;
1694 if (copy_from_user(chunk_pages, pages, chunk_nr * sizeof(*chunk_pages)))
1697 do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status);
1699 if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status)))
1704 nr_pages -= chunk_nr;
1706 return nr_pages ? -EFAULT : 0;
1710 * Move a list of pages in the address space of the currently executing
1713 static int kernel_move_pages(pid_t pid, unsigned long nr_pages,
1714 const void __user * __user *pages,
1715 const int __user *nodes,
1716 int __user *status, int flags)
1718 struct task_struct *task;
1719 struct mm_struct *mm;
1721 nodemask_t task_nodes;
1724 if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
1727 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1730 /* Find the mm_struct */
1732 task = pid ? find_task_by_vpid(pid) : current;
1737 get_task_struct(task);
1740 * Check if this process has the right to modify the specified
1741 * process. Use the regular "ptrace_may_access()" checks.
1743 if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) {
1750 err = security_task_movememory(task);
1754 task_nodes = cpuset_mems_allowed(task);
1755 mm = get_task_mm(task);
1756 put_task_struct(task);
1762 err = do_pages_move(mm, task_nodes, nr_pages, pages,
1763 nodes, status, flags);
1765 err = do_pages_stat(mm, nr_pages, pages, status);
1771 put_task_struct(task);
1775 SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages,
1776 const void __user * __user *, pages,
1777 const int __user *, nodes,
1778 int __user *, status, int, flags)
1780 return kernel_move_pages(pid, nr_pages, pages, nodes, status, flags);
1783 #ifdef CONFIG_COMPAT
1784 COMPAT_SYSCALL_DEFINE6(move_pages, pid_t, pid, compat_ulong_t, nr_pages,
1785 compat_uptr_t __user *, pages32,
1786 const int __user *, nodes,
1787 int __user *, status,
1790 const void __user * __user *pages;
1793 pages = compat_alloc_user_space(nr_pages * sizeof(void *));
1794 for (i = 0; i < nr_pages; i++) {
1797 if (get_user(p, pages32 + i) ||
1798 put_user(compat_ptr(p), pages + i))
1801 return kernel_move_pages(pid, nr_pages, pages, nodes, status, flags);
1803 #endif /* CONFIG_COMPAT */
1805 #ifdef CONFIG_NUMA_BALANCING
1807 * Returns true if this is a safe migration target node for misplaced NUMA
1808 * pages. Currently it only checks the watermarks which crude
1810 static bool migrate_balanced_pgdat(struct pglist_data *pgdat,
1811 unsigned long nr_migrate_pages)
1815 for (z = pgdat->nr_zones - 1; z >= 0; z--) {
1816 struct zone *zone = pgdat->node_zones + z;
1818 if (!populated_zone(zone))
1821 /* Avoid waking kswapd by allocating pages_to_migrate pages. */
1822 if (!zone_watermark_ok(zone, 0,
1823 high_wmark_pages(zone) +
1832 static struct page *alloc_misplaced_dst_page(struct page *page,
1835 int nid = (int) data;
1836 struct page *newpage;
1838 newpage = __alloc_pages_node(nid,
1839 (GFP_HIGHUSER_MOVABLE |
1840 __GFP_THISNODE | __GFP_NOMEMALLOC |
1841 __GFP_NORETRY | __GFP_NOWARN) &
1847 static int numamigrate_isolate_page(pg_data_t *pgdat, struct page *page)
1851 VM_BUG_ON_PAGE(compound_order(page) && !PageTransHuge(page), page);
1853 /* Avoid migrating to a node that is nearly full */
1854 if (!migrate_balanced_pgdat(pgdat, 1UL << compound_order(page)))
1857 if (isolate_lru_page(page))
1861 * migrate_misplaced_transhuge_page() skips page migration's usual
1862 * check on page_count(), so we must do it here, now that the page
1863 * has been isolated: a GUP pin, or any other pin, prevents migration.
1864 * The expected page count is 3: 1 for page's mapcount and 1 for the
1865 * caller's pin and 1 for the reference taken by isolate_lru_page().
1867 if (PageTransHuge(page) && page_count(page) != 3) {
1868 putback_lru_page(page);
1872 page_lru = page_is_file_cache(page);
1873 mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON + page_lru,
1874 hpage_nr_pages(page));
1877 * Isolating the page has taken another reference, so the
1878 * caller's reference can be safely dropped without the page
1879 * disappearing underneath us during migration.
1885 bool pmd_trans_migrating(pmd_t pmd)
1887 struct page *page = pmd_page(pmd);
1888 return PageLocked(page);
1892 * Attempt to migrate a misplaced page to the specified destination
1893 * node. Caller is expected to have an elevated reference count on
1894 * the page that will be dropped by this function before returning.
1896 int migrate_misplaced_page(struct page *page, struct vm_area_struct *vma,
1899 pg_data_t *pgdat = NODE_DATA(node);
1902 LIST_HEAD(migratepages);
1905 * Don't migrate file pages that are mapped in multiple processes
1906 * with execute permissions as they are probably shared libraries.
1908 if (page_mapcount(page) != 1 && page_is_file_cache(page) &&
1909 (vma->vm_flags & VM_EXEC))
1913 * Also do not migrate dirty pages as not all filesystems can move
1914 * dirty pages in MIGRATE_ASYNC mode which is a waste of cycles.
1916 if (page_is_file_cache(page) && PageDirty(page))
1919 isolated = numamigrate_isolate_page(pgdat, page);
1923 list_add(&page->lru, &migratepages);
1924 nr_remaining = migrate_pages(&migratepages, alloc_misplaced_dst_page,
1925 NULL, node, MIGRATE_ASYNC,
1928 if (!list_empty(&migratepages)) {
1929 list_del(&page->lru);
1930 dec_node_page_state(page, NR_ISOLATED_ANON +
1931 page_is_file_cache(page));
1932 putback_lru_page(page);
1936 count_vm_numa_event(NUMA_PAGE_MIGRATE);
1937 BUG_ON(!list_empty(&migratepages));
1944 #endif /* CONFIG_NUMA_BALANCING */
1946 #if defined(CONFIG_NUMA_BALANCING) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
1948 * Migrates a THP to a given target node. page must be locked and is unlocked
1951 int migrate_misplaced_transhuge_page(struct mm_struct *mm,
1952 struct vm_area_struct *vma,
1953 pmd_t *pmd, pmd_t entry,
1954 unsigned long address,
1955 struct page *page, int node)
1958 pg_data_t *pgdat = NODE_DATA(node);
1960 struct page *new_page = NULL;
1961 int page_lru = page_is_file_cache(page);
1962 unsigned long start = address & HPAGE_PMD_MASK;
1964 new_page = alloc_pages_node(node,
1965 (GFP_TRANSHUGE_LIGHT | __GFP_THISNODE),
1969 prep_transhuge_page(new_page);
1971 isolated = numamigrate_isolate_page(pgdat, page);
1977 /* Prepare a page as a migration target */
1978 __SetPageLocked(new_page);
1979 if (PageSwapBacked(page))
1980 __SetPageSwapBacked(new_page);
1982 /* anon mapping, we can simply copy page->mapping to the new page: */
1983 new_page->mapping = page->mapping;
1984 new_page->index = page->index;
1985 /* flush the cache before copying using the kernel virtual address */
1986 flush_cache_range(vma, start, start + HPAGE_PMD_SIZE);
1987 migrate_page_copy(new_page, page);
1988 WARN_ON(PageLRU(new_page));
1990 /* Recheck the target PMD */
1991 ptl = pmd_lock(mm, pmd);
1992 if (unlikely(!pmd_same(*pmd, entry) || !page_ref_freeze(page, 2))) {
1995 /* Reverse changes made by migrate_page_copy() */
1996 if (TestClearPageActive(new_page))
1997 SetPageActive(page);
1998 if (TestClearPageUnevictable(new_page))
1999 SetPageUnevictable(page);
2001 unlock_page(new_page);
2002 put_page(new_page); /* Free it */
2004 /* Retake the callers reference and putback on LRU */
2006 putback_lru_page(page);
2007 mod_node_page_state(page_pgdat(page),
2008 NR_ISOLATED_ANON + page_lru, -HPAGE_PMD_NR);
2013 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
2014 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
2017 * Overwrite the old entry under pagetable lock and establish
2018 * the new PTE. Any parallel GUP will either observe the old
2019 * page blocking on the page lock, block on the page table
2020 * lock or observe the new page. The SetPageUptodate on the
2021 * new page and page_add_new_anon_rmap guarantee the copy is
2022 * visible before the pagetable update.
2024 page_add_anon_rmap(new_page, vma, start, true);
2026 * At this point the pmd is numa/protnone (i.e. non present) and the TLB
2027 * has already been flushed globally. So no TLB can be currently
2028 * caching this non present pmd mapping. There's no need to clear the
2029 * pmd before doing set_pmd_at(), nor to flush the TLB after
2030 * set_pmd_at(). Clearing the pmd here would introduce a race
2031 * condition against MADV_DONTNEED, because MADV_DONTNEED only holds the
2032 * mmap_sem for reading. If the pmd is set to NULL at any given time,
2033 * MADV_DONTNEED won't wait on the pmd lock and it'll skip clearing this
2036 set_pmd_at(mm, start, pmd, entry);
2037 update_mmu_cache_pmd(vma, address, &entry);
2039 page_ref_unfreeze(page, 2);
2040 mlock_migrate_page(new_page, page);
2041 page_remove_rmap(page, true);
2042 set_page_owner_migrate_reason(new_page, MR_NUMA_MISPLACED);
2046 /* Take an "isolate" reference and put new page on the LRU. */
2048 putback_lru_page(new_page);
2050 unlock_page(new_page);
2052 put_page(page); /* Drop the rmap reference */
2053 put_page(page); /* Drop the LRU isolation reference */
2055 count_vm_events(PGMIGRATE_SUCCESS, HPAGE_PMD_NR);
2056 count_vm_numa_events(NUMA_PAGE_MIGRATE, HPAGE_PMD_NR);
2058 mod_node_page_state(page_pgdat(page),
2059 NR_ISOLATED_ANON + page_lru,
2064 count_vm_events(PGMIGRATE_FAIL, HPAGE_PMD_NR);
2065 ptl = pmd_lock(mm, pmd);
2066 if (pmd_same(*pmd, entry)) {
2067 entry = pmd_modify(entry, vma->vm_page_prot);
2068 set_pmd_at(mm, start, pmd, entry);
2069 update_mmu_cache_pmd(vma, address, &entry);
2078 #endif /* CONFIG_NUMA_BALANCING */
2080 #endif /* CONFIG_NUMA */
2082 #if defined(CONFIG_MIGRATE_VMA_HELPER)
2083 struct migrate_vma {
2084 struct vm_area_struct *vma;
2087 unsigned long cpages;
2088 unsigned long npages;
2089 unsigned long start;
2093 static int migrate_vma_collect_hole(unsigned long start,
2095 struct mm_walk *walk)
2097 struct migrate_vma *migrate = walk->private;
2100 for (addr = start & PAGE_MASK; addr < end; addr += PAGE_SIZE) {
2101 migrate->src[migrate->npages] = MIGRATE_PFN_MIGRATE;
2102 migrate->dst[migrate->npages] = 0;
2110 static int migrate_vma_collect_skip(unsigned long start,
2112 struct mm_walk *walk)
2114 struct migrate_vma *migrate = walk->private;
2117 for (addr = start & PAGE_MASK; addr < end; addr += PAGE_SIZE) {
2118 migrate->dst[migrate->npages] = 0;
2119 migrate->src[migrate->npages++] = 0;
2125 static int migrate_vma_collect_pmd(pmd_t *pmdp,
2126 unsigned long start,
2128 struct mm_walk *walk)
2130 struct migrate_vma *migrate = walk->private;
2131 struct vm_area_struct *vma = walk->vma;
2132 struct mm_struct *mm = vma->vm_mm;
2133 unsigned long addr = start, unmapped = 0;
2138 if (pmd_none(*pmdp))
2139 return migrate_vma_collect_hole(start, end, walk);
2141 if (pmd_trans_huge(*pmdp)) {
2144 ptl = pmd_lock(mm, pmdp);
2145 if (unlikely(!pmd_trans_huge(*pmdp))) {
2150 page = pmd_page(*pmdp);
2151 if (is_huge_zero_page(page)) {
2153 split_huge_pmd(vma, pmdp, addr);
2154 if (pmd_trans_unstable(pmdp))
2155 return migrate_vma_collect_skip(start, end,
2162 if (unlikely(!trylock_page(page)))
2163 return migrate_vma_collect_skip(start, end,
2165 ret = split_huge_page(page);
2169 return migrate_vma_collect_skip(start, end,
2171 if (pmd_none(*pmdp))
2172 return migrate_vma_collect_hole(start, end,
2177 if (unlikely(pmd_bad(*pmdp)))
2178 return migrate_vma_collect_skip(start, end, walk);
2180 ptep = pte_offset_map_lock(mm, pmdp, addr, &ptl);
2181 arch_enter_lazy_mmu_mode();
2183 for (; addr < end; addr += PAGE_SIZE, ptep++) {
2184 unsigned long mpfn, pfn;
2192 if (pte_none(pte)) {
2193 mpfn = MIGRATE_PFN_MIGRATE;
2199 if (!pte_present(pte)) {
2203 * Only care about unaddressable device page special
2204 * page table entry. Other special swap entries are not
2205 * migratable, and we ignore regular swapped page.
2207 entry = pte_to_swp_entry(pte);
2208 if (!is_device_private_entry(entry))
2211 page = device_private_entry_to_page(entry);
2212 mpfn = migrate_pfn(page_to_pfn(page))|
2213 MIGRATE_PFN_DEVICE | MIGRATE_PFN_MIGRATE;
2214 if (is_write_device_private_entry(entry))
2215 mpfn |= MIGRATE_PFN_WRITE;
2217 if (is_zero_pfn(pfn)) {
2218 mpfn = MIGRATE_PFN_MIGRATE;
2223 page = _vm_normal_page(migrate->vma, addr, pte, true);
2224 mpfn = migrate_pfn(pfn) | MIGRATE_PFN_MIGRATE;
2225 mpfn |= pte_write(pte) ? MIGRATE_PFN_WRITE : 0;
2228 /* FIXME support THP */
2229 if (!page || !page->mapping || PageTransCompound(page)) {
2233 pfn = page_to_pfn(page);
2236 * By getting a reference on the page we pin it and that blocks
2237 * any kind of migration. Side effect is that it "freezes" the
2240 * We drop this reference after isolating the page from the lru
2241 * for non device page (device page are not on the lru and thus
2242 * can't be dropped from it).
2248 * Optimize for the common case where page is only mapped once
2249 * in one process. If we can lock the page, then we can safely
2250 * set up a special migration page table entry now.
2252 if (trylock_page(page)) {
2255 mpfn |= MIGRATE_PFN_LOCKED;
2256 ptep_get_and_clear(mm, addr, ptep);
2258 /* Setup special migration page table entry */
2259 entry = make_migration_entry(page, mpfn &
2261 swp_pte = swp_entry_to_pte(entry);
2262 if (pte_soft_dirty(pte))
2263 swp_pte = pte_swp_mksoft_dirty(swp_pte);
2264 set_pte_at(mm, addr, ptep, swp_pte);
2267 * This is like regular unmap: we remove the rmap and
2268 * drop page refcount. Page won't be freed, as we took
2269 * a reference just above.
2271 page_remove_rmap(page, false);
2274 if (pte_present(pte))
2279 migrate->dst[migrate->npages] = 0;
2280 migrate->src[migrate->npages++] = mpfn;
2282 arch_leave_lazy_mmu_mode();
2283 pte_unmap_unlock(ptep - 1, ptl);
2285 /* Only flush the TLB if we actually modified any entries */
2287 flush_tlb_range(walk->vma, start, end);
2293 * migrate_vma_collect() - collect pages over a range of virtual addresses
2294 * @migrate: migrate struct containing all migration information
2296 * This will walk the CPU page table. For each virtual address backed by a
2297 * valid page, it updates the src array and takes a reference on the page, in
2298 * order to pin the page until we lock it and unmap it.
2300 static void migrate_vma_collect(struct migrate_vma *migrate)
2302 struct mm_walk mm_walk;
2304 mm_walk.pmd_entry = migrate_vma_collect_pmd;
2305 mm_walk.pte_entry = NULL;
2306 mm_walk.pte_hole = migrate_vma_collect_hole;
2307 mm_walk.hugetlb_entry = NULL;
2308 mm_walk.test_walk = NULL;
2309 mm_walk.vma = migrate->vma;
2310 mm_walk.mm = migrate->vma->vm_mm;
2311 mm_walk.private = migrate;
2313 mmu_notifier_invalidate_range_start(mm_walk.mm,
2316 walk_page_range(migrate->start, migrate->end, &mm_walk);
2317 mmu_notifier_invalidate_range_end(mm_walk.mm,
2321 migrate->end = migrate->start + (migrate->npages << PAGE_SHIFT);
2325 * migrate_vma_check_page() - check if page is pinned or not
2326 * @page: struct page to check
2328 * Pinned pages cannot be migrated. This is the same test as in
2329 * migrate_page_move_mapping(), except that here we allow migration of a
2332 static bool migrate_vma_check_page(struct page *page)
2335 * One extra ref because caller holds an extra reference, either from
2336 * isolate_lru_page() for a regular page, or migrate_vma_collect() for
2342 * FIXME support THP (transparent huge page), it is bit more complex to
2343 * check them than regular pages, because they can be mapped with a pmd
2344 * or with a pte (split pte mapping).
2346 if (PageCompound(page))
2349 /* Page from ZONE_DEVICE have one extra reference */
2350 if (is_zone_device_page(page)) {
2352 * Private page can never be pin as they have no valid pte and
2353 * GUP will fail for those. Yet if there is a pending migration
2354 * a thread might try to wait on the pte migration entry and
2355 * will bump the page reference count. Sadly there is no way to
2356 * differentiate a regular pin from migration wait. Hence to
2357 * avoid 2 racing thread trying to migrate back to CPU to enter
2358 * infinite loop (one stoping migration because the other is
2359 * waiting on pte migration entry). We always return true here.
2361 * FIXME proper solution is to rework migration_entry_wait() so
2362 * it does not need to take a reference on page.
2364 if (is_device_private_page(page))
2368 * Only allow device public page to be migrated and account for
2369 * the extra reference count imply by ZONE_DEVICE pages.
2371 if (!is_device_public_page(page))
2376 /* For file back page */
2377 if (page_mapping(page))
2378 extra += 1 + page_has_private(page);
2380 if ((page_count(page) - extra) > page_mapcount(page))
2387 * migrate_vma_prepare() - lock pages and isolate them from the lru
2388 * @migrate: migrate struct containing all migration information
2390 * This locks pages that have been collected by migrate_vma_collect(). Once each
2391 * page is locked it is isolated from the lru (for non-device pages). Finally,
2392 * the ref taken by migrate_vma_collect() is dropped, as locked pages cannot be
2393 * migrated by concurrent kernel threads.
2395 static void migrate_vma_prepare(struct migrate_vma *migrate)
2397 const unsigned long npages = migrate->npages;
2398 const unsigned long start = migrate->start;
2399 unsigned long addr, i, restore = 0;
2400 bool allow_drain = true;
2404 for (i = 0; (i < npages) && migrate->cpages; i++) {
2405 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2411 if (!(migrate->src[i] & MIGRATE_PFN_LOCKED)) {
2413 * Because we are migrating several pages there can be
2414 * a deadlock between 2 concurrent migration where each
2415 * are waiting on each other page lock.
2417 * Make migrate_vma() a best effort thing and backoff
2418 * for any page we can not lock right away.
2420 if (!trylock_page(page)) {
2421 migrate->src[i] = 0;
2427 migrate->src[i] |= MIGRATE_PFN_LOCKED;
2430 /* ZONE_DEVICE pages are not on LRU */
2431 if (!is_zone_device_page(page)) {
2432 if (!PageLRU(page) && allow_drain) {
2433 /* Drain CPU's pagevec */
2434 lru_add_drain_all();
2435 allow_drain = false;
2438 if (isolate_lru_page(page)) {
2440 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2444 migrate->src[i] = 0;
2452 /* Drop the reference we took in collect */
2456 if (!migrate_vma_check_page(page)) {
2458 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2462 if (!is_zone_device_page(page)) {
2464 putback_lru_page(page);
2467 migrate->src[i] = 0;
2471 if (!is_zone_device_page(page))
2472 putback_lru_page(page);
2479 for (i = 0, addr = start; i < npages && restore; i++, addr += PAGE_SIZE) {
2480 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2482 if (!page || (migrate->src[i] & MIGRATE_PFN_MIGRATE))
2485 remove_migration_pte(page, migrate->vma, addr, page);
2487 migrate->src[i] = 0;
2495 * migrate_vma_unmap() - replace page mapping with special migration pte entry
2496 * @migrate: migrate struct containing all migration information
2498 * Replace page mapping (CPU page table pte) with a special migration pte entry
2499 * and check again if it has been pinned. Pinned pages are restored because we
2500 * cannot migrate them.
2502 * This is the last step before we call the device driver callback to allocate
2503 * destination memory and copy contents of original page over to new page.
2505 static void migrate_vma_unmap(struct migrate_vma *migrate)
2507 int flags = TTU_MIGRATION | TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS;
2508 const unsigned long npages = migrate->npages;
2509 const unsigned long start = migrate->start;
2510 unsigned long addr, i, restore = 0;
2512 for (i = 0; i < npages; i++) {
2513 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2515 if (!page || !(migrate->src[i] & MIGRATE_PFN_MIGRATE))
2518 if (page_mapped(page)) {
2519 try_to_unmap(page, flags);
2520 if (page_mapped(page))
2524 if (migrate_vma_check_page(page))
2528 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2533 for (addr = start, i = 0; i < npages && restore; addr += PAGE_SIZE, i++) {
2534 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2536 if (!page || (migrate->src[i] & MIGRATE_PFN_MIGRATE))
2539 remove_migration_ptes(page, page, false);
2541 migrate->src[i] = 0;
2545 if (is_zone_device_page(page))
2548 putback_lru_page(page);
2552 static void migrate_vma_insert_page(struct migrate_vma *migrate,
2558 struct vm_area_struct *vma = migrate->vma;
2559 struct mm_struct *mm = vma->vm_mm;
2560 struct mem_cgroup *memcg;
2570 /* Only allow populating anonymous memory */
2571 if (!vma_is_anonymous(vma))
2574 pgdp = pgd_offset(mm, addr);
2575 p4dp = p4d_alloc(mm, pgdp, addr);
2578 pudp = pud_alloc(mm, p4dp, addr);
2581 pmdp = pmd_alloc(mm, pudp, addr);
2585 if (pmd_trans_huge(*pmdp) || pmd_devmap(*pmdp))
2589 * Use pte_alloc() instead of pte_alloc_map(). We can't run
2590 * pte_offset_map() on pmds where a huge pmd might be created
2591 * from a different thread.
2593 * pte_alloc_map() is safe to use under down_write(mmap_sem) or when
2594 * parallel threads are excluded by other means.
2596 * Here we only have down_read(mmap_sem).
2598 if (pte_alloc(mm, pmdp, addr))
2601 /* See the comment in pte_alloc_one_map() */
2602 if (unlikely(pmd_trans_unstable(pmdp)))
2605 if (unlikely(anon_vma_prepare(vma)))
2607 if (mem_cgroup_try_charge(page, vma->vm_mm, GFP_KERNEL, &memcg, false))
2611 * The memory barrier inside __SetPageUptodate makes sure that
2612 * preceding stores to the page contents become visible before
2613 * the set_pte_at() write.
2615 __SetPageUptodate(page);
2617 if (is_zone_device_page(page)) {
2618 if (is_device_private_page(page)) {
2619 swp_entry_t swp_entry;
2621 swp_entry = make_device_private_entry(page, vma->vm_flags & VM_WRITE);
2622 entry = swp_entry_to_pte(swp_entry);
2623 } else if (is_device_public_page(page)) {
2624 entry = pte_mkold(mk_pte(page, READ_ONCE(vma->vm_page_prot)));
2625 if (vma->vm_flags & VM_WRITE)
2626 entry = pte_mkwrite(pte_mkdirty(entry));
2627 entry = pte_mkdevmap(entry);
2630 entry = mk_pte(page, vma->vm_page_prot);
2631 if (vma->vm_flags & VM_WRITE)
2632 entry = pte_mkwrite(pte_mkdirty(entry));
2635 ptep = pte_offset_map_lock(mm, pmdp, addr, &ptl);
2637 if (pte_present(*ptep)) {
2638 unsigned long pfn = pte_pfn(*ptep);
2640 if (!is_zero_pfn(pfn)) {
2641 pte_unmap_unlock(ptep, ptl);
2642 mem_cgroup_cancel_charge(page, memcg, false);
2646 } else if (!pte_none(*ptep)) {
2647 pte_unmap_unlock(ptep, ptl);
2648 mem_cgroup_cancel_charge(page, memcg, false);
2653 * Check for usefaultfd but do not deliver the fault. Instead,
2656 if (userfaultfd_missing(vma)) {
2657 pte_unmap_unlock(ptep, ptl);
2658 mem_cgroup_cancel_charge(page, memcg, false);
2662 inc_mm_counter(mm, MM_ANONPAGES);
2663 page_add_new_anon_rmap(page, vma, addr, false);
2664 mem_cgroup_commit_charge(page, memcg, false, false);
2665 if (!is_zone_device_page(page))
2666 lru_cache_add_active_or_unevictable(page, vma);
2670 flush_cache_page(vma, addr, pte_pfn(*ptep));
2671 ptep_clear_flush_notify(vma, addr, ptep);
2672 set_pte_at_notify(mm, addr, ptep, entry);
2673 update_mmu_cache(vma, addr, ptep);
2675 /* No need to invalidate - it was non-present before */
2676 set_pte_at(mm, addr, ptep, entry);
2677 update_mmu_cache(vma, addr, ptep);
2680 pte_unmap_unlock(ptep, ptl);
2681 *src = MIGRATE_PFN_MIGRATE;
2685 *src &= ~MIGRATE_PFN_MIGRATE;
2689 * migrate_vma_pages() - migrate meta-data from src page to dst page
2690 * @migrate: migrate struct containing all migration information
2692 * This migrates struct page meta-data from source struct page to destination
2693 * struct page. This effectively finishes the migration from source page to the
2696 static void migrate_vma_pages(struct migrate_vma *migrate)
2698 const unsigned long npages = migrate->npages;
2699 const unsigned long start = migrate->start;
2700 struct vm_area_struct *vma = migrate->vma;
2701 struct mm_struct *mm = vma->vm_mm;
2702 unsigned long addr, i, mmu_start;
2703 bool notified = false;
2705 for (i = 0, addr = start; i < npages; addr += PAGE_SIZE, i++) {
2706 struct page *newpage = migrate_pfn_to_page(migrate->dst[i]);
2707 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2708 struct address_space *mapping;
2712 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2717 if (!(migrate->src[i] & MIGRATE_PFN_MIGRATE)) {
2723 mmu_notifier_invalidate_range_start(mm,
2727 migrate_vma_insert_page(migrate, addr, newpage,
2733 mapping = page_mapping(page);
2735 if (is_zone_device_page(newpage)) {
2736 if (is_device_private_page(newpage)) {
2738 * For now only support private anonymous when
2739 * migrating to un-addressable device memory.
2742 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2745 } else if (!is_device_public_page(newpage)) {
2747 * Other types of ZONE_DEVICE page are not
2750 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2755 r = migrate_page(mapping, newpage, page, MIGRATE_SYNC_NO_COPY);
2756 if (r != MIGRATEPAGE_SUCCESS)
2757 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2761 * No need to double call mmu_notifier->invalidate_range() callback as
2762 * the above ptep_clear_flush_notify() inside migrate_vma_insert_page()
2763 * did already call it.
2766 mmu_notifier_invalidate_range_only_end(mm, mmu_start,
2771 * migrate_vma_finalize() - restore CPU page table entry
2772 * @migrate: migrate struct containing all migration information
2774 * This replaces the special migration pte entry with either a mapping to the
2775 * new page if migration was successful for that page, or to the original page
2778 * This also unlocks the pages and puts them back on the lru, or drops the extra
2779 * refcount, for device pages.
2781 static void migrate_vma_finalize(struct migrate_vma *migrate)
2783 const unsigned long npages = migrate->npages;
2786 for (i = 0; i < npages; i++) {
2787 struct page *newpage = migrate_pfn_to_page(migrate->dst[i]);
2788 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2792 unlock_page(newpage);
2798 if (!(migrate->src[i] & MIGRATE_PFN_MIGRATE) || !newpage) {
2800 unlock_page(newpage);
2806 remove_migration_ptes(page, newpage, false);
2810 if (is_zone_device_page(page))
2813 putback_lru_page(page);
2815 if (newpage != page) {
2816 unlock_page(newpage);
2817 if (is_zone_device_page(newpage))
2820 putback_lru_page(newpage);
2826 * migrate_vma() - migrate a range of memory inside vma
2828 * @ops: migration callback for allocating destination memory and copying
2829 * @vma: virtual memory area containing the range to be migrated
2830 * @start: start address of the range to migrate (inclusive)
2831 * @end: end address of the range to migrate (exclusive)
2832 * @src: array of hmm_pfn_t containing source pfns
2833 * @dst: array of hmm_pfn_t containing destination pfns
2834 * @private: pointer passed back to each of the callback
2835 * Returns: 0 on success, error code otherwise
2837 * This function tries to migrate a range of memory virtual address range, using
2838 * callbacks to allocate and copy memory from source to destination. First it
2839 * collects all the pages backing each virtual address in the range, saving this
2840 * inside the src array. Then it locks those pages and unmaps them. Once the pages
2841 * are locked and unmapped, it checks whether each page is pinned or not. Pages
2842 * that aren't pinned have the MIGRATE_PFN_MIGRATE flag set (by this function)
2843 * in the corresponding src array entry. It then restores any pages that are
2844 * pinned, by remapping and unlocking those pages.
2846 * At this point it calls the alloc_and_copy() callback. For documentation on
2847 * what is expected from that callback, see struct migrate_vma_ops comments in
2848 * include/linux/migrate.h
2850 * After the alloc_and_copy() callback, this function goes over each entry in
2851 * the src array that has the MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE flag
2852 * set. If the corresponding entry in dst array has MIGRATE_PFN_VALID flag set,
2853 * then the function tries to migrate struct page information from the source
2854 * struct page to the destination struct page. If it fails to migrate the struct
2855 * page information, then it clears the MIGRATE_PFN_MIGRATE flag in the src
2858 * At this point all successfully migrated pages have an entry in the src
2859 * array with MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE flag set and the dst
2860 * array entry with MIGRATE_PFN_VALID flag set.
2862 * It then calls the finalize_and_map() callback. See comments for "struct
2863 * migrate_vma_ops", in include/linux/migrate.h for details about
2864 * finalize_and_map() behavior.
2866 * After the finalize_and_map() callback, for successfully migrated pages, this
2867 * function updates the CPU page table to point to new pages, otherwise it
2868 * restores the CPU page table to point to the original source pages.
2870 * Function returns 0 after the above steps, even if no pages were migrated
2871 * (The function only returns an error if any of the arguments are invalid.)
2873 * Both src and dst array must be big enough for (end - start) >> PAGE_SHIFT
2874 * unsigned long entries.
2876 int migrate_vma(const struct migrate_vma_ops *ops,
2877 struct vm_area_struct *vma,
2878 unsigned long start,
2884 struct migrate_vma migrate;
2886 /* Sanity check the arguments */
2889 if (!vma || is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_SPECIAL) ||
2892 if (start < vma->vm_start || start >= vma->vm_end)
2894 if (end <= vma->vm_start || end > vma->vm_end)
2896 if (!ops || !src || !dst || start >= end)
2899 memset(src, 0, sizeof(*src) * ((end - start) >> PAGE_SHIFT));
2902 migrate.start = start;
2908 /* Collect, and try to unmap source pages */
2909 migrate_vma_collect(&migrate);
2910 if (!migrate.cpages)
2913 /* Lock and isolate page */
2914 migrate_vma_prepare(&migrate);
2915 if (!migrate.cpages)
2919 migrate_vma_unmap(&migrate);
2920 if (!migrate.cpages)
2924 * At this point pages are locked and unmapped, and thus they have
2925 * stable content and can safely be copied to destination memory that
2926 * is allocated by the callback.
2928 * Note that migration can fail in migrate_vma_struct_page() for each
2931 ops->alloc_and_copy(vma, src, dst, start, end, private);
2933 /* This does the real migration of struct page */
2934 migrate_vma_pages(&migrate);
2936 ops->finalize_and_map(vma, src, dst, start, end, private);
2938 /* Unlock and remap pages */
2939 migrate_vma_finalize(&migrate);
2943 EXPORT_SYMBOL(migrate_vma);
2944 #endif /* defined(MIGRATE_VMA_HELPER) */