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);
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);
377 static int expected_page_refs(struct page *page)
379 int expected_count = 1;
382 * Device public or private pages have an extra refcount as they are
385 expected_count += is_device_private_page(page);
386 expected_count += is_device_public_page(page);
387 if (page_mapping(page))
388 expected_count += hpage_nr_pages(page) + page_has_private(page);
390 return expected_count;
394 * Replace the page in the mapping.
396 * The number of remaining references must be:
397 * 1 for anonymous pages without a mapping
398 * 2 for pages with a mapping
399 * 3 for pages with a mapping and PagePrivate/PagePrivate2 set.
401 int migrate_page_move_mapping(struct address_space *mapping,
402 struct page *newpage, struct page *page,
403 struct buffer_head *head, enum migrate_mode mode,
406 XA_STATE(xas, &mapping->i_pages, page_index(page));
407 struct zone *oldzone, *newzone;
409 int expected_count = expected_page_refs(page) + extra_count;
412 /* Anonymous page without mapping */
413 if (page_count(page) != expected_count)
416 /* No turning back from here */
417 newpage->index = page->index;
418 newpage->mapping = page->mapping;
419 if (PageSwapBacked(page))
420 __SetPageSwapBacked(newpage);
422 return MIGRATEPAGE_SUCCESS;
425 oldzone = page_zone(page);
426 newzone = page_zone(newpage);
429 if (page_count(page) != expected_count || xas_load(&xas) != page) {
430 xas_unlock_irq(&xas);
434 if (!page_ref_freeze(page, expected_count)) {
435 xas_unlock_irq(&xas);
440 * Now we know that no one else is looking at the page:
441 * no turning back from here.
443 newpage->index = page->index;
444 newpage->mapping = page->mapping;
445 page_ref_add(newpage, hpage_nr_pages(page)); /* add cache reference */
446 if (PageSwapBacked(page)) {
447 __SetPageSwapBacked(newpage);
448 if (PageSwapCache(page)) {
449 SetPageSwapCache(newpage);
450 set_page_private(newpage, page_private(page));
453 VM_BUG_ON_PAGE(PageSwapCache(page), page);
456 /* Move dirty while page refs frozen and newpage not yet exposed */
457 dirty = PageDirty(page);
459 ClearPageDirty(page);
460 SetPageDirty(newpage);
463 xas_store(&xas, newpage);
464 if (PageTransHuge(page)) {
467 for (i = 1; i < HPAGE_PMD_NR; i++) {
469 xas_store(&xas, newpage + i);
474 * Drop cache reference from old page by unfreezing
475 * to one less reference.
476 * We know this isn't the last reference.
478 page_ref_unfreeze(page, expected_count - hpage_nr_pages(page));
481 /* Leave irq disabled to prevent preemption while updating stats */
484 * If moved to a different zone then also account
485 * the page for that zone. Other VM counters will be
486 * taken care of when we establish references to the
487 * new page and drop references to the old page.
489 * Note that anonymous pages are accounted for
490 * via NR_FILE_PAGES and NR_ANON_MAPPED if they
491 * are mapped to swap space.
493 if (newzone != oldzone) {
494 __dec_node_state(oldzone->zone_pgdat, NR_FILE_PAGES);
495 __inc_node_state(newzone->zone_pgdat, NR_FILE_PAGES);
496 if (PageSwapBacked(page) && !PageSwapCache(page)) {
497 __dec_node_state(oldzone->zone_pgdat, NR_SHMEM);
498 __inc_node_state(newzone->zone_pgdat, NR_SHMEM);
500 if (dirty && mapping_cap_account_dirty(mapping)) {
501 __dec_node_state(oldzone->zone_pgdat, NR_FILE_DIRTY);
502 __dec_zone_state(oldzone, NR_ZONE_WRITE_PENDING);
503 __inc_node_state(newzone->zone_pgdat, NR_FILE_DIRTY);
504 __inc_zone_state(newzone, NR_ZONE_WRITE_PENDING);
509 return MIGRATEPAGE_SUCCESS;
511 EXPORT_SYMBOL(migrate_page_move_mapping);
514 * The expected number of remaining references is the same as that
515 * of migrate_page_move_mapping().
517 int migrate_huge_page_move_mapping(struct address_space *mapping,
518 struct page *newpage, struct page *page)
520 XA_STATE(xas, &mapping->i_pages, page_index(page));
524 expected_count = 2 + page_has_private(page);
525 if (page_count(page) != expected_count || xas_load(&xas) != page) {
526 xas_unlock_irq(&xas);
530 if (!page_ref_freeze(page, expected_count)) {
531 xas_unlock_irq(&xas);
535 newpage->index = page->index;
536 newpage->mapping = page->mapping;
540 xas_store(&xas, newpage);
542 page_ref_unfreeze(page, expected_count - 1);
544 xas_unlock_irq(&xas);
546 return MIGRATEPAGE_SUCCESS;
550 * Gigantic pages are so large that we do not guarantee that page++ pointer
551 * arithmetic will work across the entire page. We need something more
554 static void __copy_gigantic_page(struct page *dst, struct page *src,
558 struct page *dst_base = dst;
559 struct page *src_base = src;
561 for (i = 0; i < nr_pages; ) {
563 copy_highpage(dst, src);
566 dst = mem_map_next(dst, dst_base, i);
567 src = mem_map_next(src, src_base, i);
571 static void copy_huge_page(struct page *dst, struct page *src)
578 struct hstate *h = page_hstate(src);
579 nr_pages = pages_per_huge_page(h);
581 if (unlikely(nr_pages > MAX_ORDER_NR_PAGES)) {
582 __copy_gigantic_page(dst, src, nr_pages);
587 BUG_ON(!PageTransHuge(src));
588 nr_pages = hpage_nr_pages(src);
591 for (i = 0; i < nr_pages; i++) {
593 copy_highpage(dst + i, src + i);
598 * Copy the page to its new location
600 void migrate_page_states(struct page *newpage, struct page *page)
605 SetPageError(newpage);
606 if (PageReferenced(page))
607 SetPageReferenced(newpage);
608 if (PageUptodate(page))
609 SetPageUptodate(newpage);
610 if (TestClearPageActive(page)) {
611 VM_BUG_ON_PAGE(PageUnevictable(page), page);
612 SetPageActive(newpage);
613 } else if (TestClearPageUnevictable(page))
614 SetPageUnevictable(newpage);
615 if (PageWorkingset(page))
616 SetPageWorkingset(newpage);
617 if (PageChecked(page))
618 SetPageChecked(newpage);
619 if (PageMappedToDisk(page))
620 SetPageMappedToDisk(newpage);
622 /* Move dirty on pages not done by migrate_page_move_mapping() */
624 SetPageDirty(newpage);
626 if (page_is_young(page))
627 set_page_young(newpage);
628 if (page_is_idle(page))
629 set_page_idle(newpage);
632 * Copy NUMA information to the new page, to prevent over-eager
633 * future migrations of this same page.
635 cpupid = page_cpupid_xchg_last(page, -1);
636 page_cpupid_xchg_last(newpage, cpupid);
638 ksm_migrate_page(newpage, page);
640 * Please do not reorder this without considering how mm/ksm.c's
641 * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache().
643 if (PageSwapCache(page))
644 ClearPageSwapCache(page);
645 ClearPagePrivate(page);
646 set_page_private(page, 0);
649 * If any waiters have accumulated on the new page then
652 if (PageWriteback(newpage))
653 end_page_writeback(newpage);
655 copy_page_owner(page, newpage);
657 mem_cgroup_migrate(page, newpage);
659 EXPORT_SYMBOL(migrate_page_states);
661 void migrate_page_copy(struct page *newpage, struct page *page)
663 if (PageHuge(page) || PageTransHuge(page))
664 copy_huge_page(newpage, page);
666 copy_highpage(newpage, page);
668 migrate_page_states(newpage, page);
670 EXPORT_SYMBOL(migrate_page_copy);
672 /************************************************************
673 * Migration functions
674 ***********************************************************/
677 * Common logic to directly migrate a single LRU page suitable for
678 * pages that do not use PagePrivate/PagePrivate2.
680 * Pages are locked upon entry and exit.
682 int migrate_page(struct address_space *mapping,
683 struct page *newpage, struct page *page,
684 enum migrate_mode mode)
688 BUG_ON(PageWriteback(page)); /* Writeback must be complete */
690 rc = migrate_page_move_mapping(mapping, newpage, page, NULL, mode, 0);
692 if (rc != MIGRATEPAGE_SUCCESS)
695 if (mode != MIGRATE_SYNC_NO_COPY)
696 migrate_page_copy(newpage, page);
698 migrate_page_states(newpage, page);
699 return MIGRATEPAGE_SUCCESS;
701 EXPORT_SYMBOL(migrate_page);
704 /* Returns true if all buffers are successfully locked */
705 static bool buffer_migrate_lock_buffers(struct buffer_head *head,
706 enum migrate_mode mode)
708 struct buffer_head *bh = head;
710 /* Simple case, sync compaction */
711 if (mode != MIGRATE_ASYNC) {
715 bh = bh->b_this_page;
717 } while (bh != head);
722 /* async case, we cannot block on lock_buffer so use trylock_buffer */
725 if (!trylock_buffer(bh)) {
727 * We failed to lock the buffer and cannot stall in
728 * async migration. Release the taken locks
730 struct buffer_head *failed_bh = bh;
733 while (bh != failed_bh) {
736 bh = bh->b_this_page;
741 bh = bh->b_this_page;
742 } while (bh != head);
747 * Migration function for pages with buffers. This function can only be used
748 * if the underlying filesystem guarantees that no other references to "page"
751 int buffer_migrate_page(struct address_space *mapping,
752 struct page *newpage, struct page *page, enum migrate_mode mode)
754 struct buffer_head *bh, *head;
758 if (!page_has_buffers(page))
759 return migrate_page(mapping, newpage, page, mode);
761 /* Check whether page does not have extra refs before we do more work */
762 expected_count = expected_page_refs(page);
763 if (page_count(page) != expected_count)
766 head = page_buffers(page);
767 if (!buffer_migrate_lock_buffers(head, mode))
770 rc = migrate_page_move_mapping(mapping, newpage, page, NULL, mode, 0);
771 if (rc != MIGRATEPAGE_SUCCESS)
774 ClearPagePrivate(page);
775 set_page_private(newpage, page_private(page));
776 set_page_private(page, 0);
782 set_bh_page(bh, newpage, bh_offset(bh));
783 bh = bh->b_this_page;
785 } while (bh != head);
787 SetPagePrivate(newpage);
789 if (mode != MIGRATE_SYNC_NO_COPY)
790 migrate_page_copy(newpage, page);
792 migrate_page_states(newpage, page);
794 rc = MIGRATEPAGE_SUCCESS;
800 bh = bh->b_this_page;
802 } while (bh != head);
806 EXPORT_SYMBOL(buffer_migrate_page);
810 * Writeback a page to clean the dirty state
812 static int writeout(struct address_space *mapping, struct page *page)
814 struct writeback_control wbc = {
815 .sync_mode = WB_SYNC_NONE,
818 .range_end = LLONG_MAX,
823 if (!mapping->a_ops->writepage)
824 /* No write method for the address space */
827 if (!clear_page_dirty_for_io(page))
828 /* Someone else already triggered a write */
832 * A dirty page may imply that the underlying filesystem has
833 * the page on some queue. So the page must be clean for
834 * migration. Writeout may mean we loose the lock and the
835 * page state is no longer what we checked for earlier.
836 * At this point we know that the migration attempt cannot
839 remove_migration_ptes(page, page, false);
841 rc = mapping->a_ops->writepage(page, &wbc);
843 if (rc != AOP_WRITEPAGE_ACTIVATE)
844 /* unlocked. Relock */
847 return (rc < 0) ? -EIO : -EAGAIN;
851 * Default handling if a filesystem does not provide a migration function.
853 static int fallback_migrate_page(struct address_space *mapping,
854 struct page *newpage, struct page *page, enum migrate_mode mode)
856 if (PageDirty(page)) {
857 /* Only writeback pages in full synchronous migration */
860 case MIGRATE_SYNC_NO_COPY:
865 return writeout(mapping, page);
869 * Buffers may be managed in a filesystem specific way.
870 * We must have no buffers or drop them.
872 if (page_has_private(page) &&
873 !try_to_release_page(page, GFP_KERNEL))
876 return migrate_page(mapping, newpage, page, mode);
880 * Move a page to a newly allocated page
881 * The page is locked and all ptes have been successfully removed.
883 * The new page will have replaced the old page if this function
888 * MIGRATEPAGE_SUCCESS - success
890 static int move_to_new_page(struct page *newpage, struct page *page,
891 enum migrate_mode mode)
893 struct address_space *mapping;
895 bool is_lru = !__PageMovable(page);
897 VM_BUG_ON_PAGE(!PageLocked(page), page);
898 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
900 mapping = page_mapping(page);
902 if (likely(is_lru)) {
904 rc = migrate_page(mapping, newpage, page, mode);
905 else if (mapping->a_ops->migratepage)
907 * Most pages have a mapping and most filesystems
908 * provide a migratepage callback. Anonymous pages
909 * are part of swap space which also has its own
910 * migratepage callback. This is the most common path
911 * for page migration.
913 rc = mapping->a_ops->migratepage(mapping, newpage,
916 rc = fallback_migrate_page(mapping, newpage,
920 * In case of non-lru page, it could be released after
921 * isolation step. In that case, we shouldn't try migration.
923 VM_BUG_ON_PAGE(!PageIsolated(page), page);
924 if (!PageMovable(page)) {
925 rc = MIGRATEPAGE_SUCCESS;
926 __ClearPageIsolated(page);
930 rc = mapping->a_ops->migratepage(mapping, newpage,
932 WARN_ON_ONCE(rc == MIGRATEPAGE_SUCCESS &&
933 !PageIsolated(page));
937 * When successful, old pagecache page->mapping must be cleared before
938 * page is freed; but stats require that PageAnon be left as PageAnon.
940 if (rc == MIGRATEPAGE_SUCCESS) {
941 if (__PageMovable(page)) {
942 VM_BUG_ON_PAGE(!PageIsolated(page), page);
945 * We clear PG_movable under page_lock so any compactor
946 * cannot try to migrate this page.
948 __ClearPageIsolated(page);
952 * Anonymous and movable page->mapping will be cleard by
953 * free_pages_prepare so don't reset it here for keeping
954 * the type to work PageAnon, for example.
956 if (!PageMappingFlags(page))
957 page->mapping = NULL;
963 static int __unmap_and_move(struct page *page, struct page *newpage,
964 int force, enum migrate_mode mode)
967 int page_was_mapped = 0;
968 struct anon_vma *anon_vma = NULL;
969 bool is_lru = !__PageMovable(page);
971 if (!trylock_page(page)) {
972 if (!force || mode == MIGRATE_ASYNC)
976 * It's not safe for direct compaction to call lock_page.
977 * For example, during page readahead pages are added locked
978 * to the LRU. Later, when the IO completes the pages are
979 * marked uptodate and unlocked. However, the queueing
980 * could be merging multiple pages for one bio (e.g.
981 * mpage_readpages). If an allocation happens for the
982 * second or third page, the process can end up locking
983 * the same page twice and deadlocking. Rather than
984 * trying to be clever about what pages can be locked,
985 * avoid the use of lock_page for direct compaction
988 if (current->flags & PF_MEMALLOC)
994 if (PageWriteback(page)) {
996 * Only in the case of a full synchronous migration is it
997 * necessary to wait for PageWriteback. In the async case,
998 * the retry loop is too short and in the sync-light case,
999 * the overhead of stalling is too much
1003 case MIGRATE_SYNC_NO_COPY:
1011 wait_on_page_writeback(page);
1015 * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
1016 * we cannot notice that anon_vma is freed while we migrates a page.
1017 * This get_anon_vma() delays freeing anon_vma pointer until the end
1018 * of migration. File cache pages are no problem because of page_lock()
1019 * File Caches may use write_page() or lock_page() in migration, then,
1020 * just care Anon page here.
1022 * Only page_get_anon_vma() understands the subtleties of
1023 * getting a hold on an anon_vma from outside one of its mms.
1024 * But if we cannot get anon_vma, then we won't need it anyway,
1025 * because that implies that the anon page is no longer mapped
1026 * (and cannot be remapped so long as we hold the page lock).
1028 if (PageAnon(page) && !PageKsm(page))
1029 anon_vma = page_get_anon_vma(page);
1032 * Block others from accessing the new page when we get around to
1033 * establishing additional references. We are usually the only one
1034 * holding a reference to newpage at this point. We used to have a BUG
1035 * here if trylock_page(newpage) fails, but would like to allow for
1036 * cases where there might be a race with the previous use of newpage.
1037 * This is much like races on refcount of oldpage: just don't BUG().
1039 if (unlikely(!trylock_page(newpage)))
1042 if (unlikely(!is_lru)) {
1043 rc = move_to_new_page(newpage, page, mode);
1044 goto out_unlock_both;
1048 * Corner case handling:
1049 * 1. When a new swap-cache page is read into, it is added to the LRU
1050 * and treated as swapcache but it has no rmap yet.
1051 * Calling try_to_unmap() against a page->mapping==NULL page will
1052 * trigger a BUG. So handle it here.
1053 * 2. An orphaned page (see truncate_complete_page) might have
1054 * fs-private metadata. The page can be picked up due to memory
1055 * offlining. Everywhere else except page reclaim, the page is
1056 * invisible to the vm, so the page can not be migrated. So try to
1057 * free the metadata, so the page can be freed.
1059 if (!page->mapping) {
1060 VM_BUG_ON_PAGE(PageAnon(page), page);
1061 if (page_has_private(page)) {
1062 try_to_free_buffers(page);
1063 goto out_unlock_both;
1065 } else if (page_mapped(page)) {
1066 /* Establish migration ptes */
1067 VM_BUG_ON_PAGE(PageAnon(page) && !PageKsm(page) && !anon_vma,
1070 TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
1071 page_was_mapped = 1;
1074 if (!page_mapped(page))
1075 rc = move_to_new_page(newpage, page, mode);
1077 if (page_was_mapped)
1078 remove_migration_ptes(page,
1079 rc == MIGRATEPAGE_SUCCESS ? newpage : page, false);
1082 unlock_page(newpage);
1084 /* Drop an anon_vma reference if we took one */
1086 put_anon_vma(anon_vma);
1090 * If migration is successful, decrease refcount of the newpage
1091 * which will not free the page because new page owner increased
1092 * refcounter. As well, if it is LRU page, add the page to LRU
1095 if (rc == MIGRATEPAGE_SUCCESS) {
1096 if (unlikely(__PageMovable(newpage)))
1099 putback_lru_page(newpage);
1106 * gcc 4.7 and 4.8 on arm get an ICEs when inlining unmap_and_move(). Work
1109 #if defined(CONFIG_ARM) && \
1110 defined(GCC_VERSION) && GCC_VERSION < 40900 && GCC_VERSION >= 40700
1111 #define ICE_noinline noinline
1113 #define ICE_noinline
1117 * Obtain the lock on page, remove all ptes and migrate the page
1118 * to the newly allocated page in newpage.
1120 static ICE_noinline int unmap_and_move(new_page_t get_new_page,
1121 free_page_t put_new_page,
1122 unsigned long private, struct page *page,
1123 int force, enum migrate_mode mode,
1124 enum migrate_reason reason)
1126 int rc = MIGRATEPAGE_SUCCESS;
1127 struct page *newpage;
1129 if (!thp_migration_supported() && PageTransHuge(page))
1132 newpage = get_new_page(page, private);
1136 if (page_count(page) == 1) {
1137 /* page was freed from under us. So we are done. */
1138 ClearPageActive(page);
1139 ClearPageUnevictable(page);
1140 if (unlikely(__PageMovable(page))) {
1142 if (!PageMovable(page))
1143 __ClearPageIsolated(page);
1147 put_new_page(newpage, private);
1153 rc = __unmap_and_move(page, newpage, force, mode);
1154 if (rc == MIGRATEPAGE_SUCCESS)
1155 set_page_owner_migrate_reason(newpage, reason);
1158 if (rc != -EAGAIN) {
1160 * A page that has been migrated has all references
1161 * removed and will be freed. A page that has not been
1162 * migrated will have kepts its references and be
1165 list_del(&page->lru);
1168 * Compaction can migrate also non-LRU pages which are
1169 * not accounted to NR_ISOLATED_*. They can be recognized
1172 if (likely(!__PageMovable(page)))
1173 mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON +
1174 page_is_file_cache(page), -hpage_nr_pages(page));
1178 * If migration is successful, releases reference grabbed during
1179 * isolation. Otherwise, restore the page to right list unless
1182 if (rc == MIGRATEPAGE_SUCCESS) {
1184 if (reason == MR_MEMORY_FAILURE) {
1186 * Set PG_HWPoison on just freed page
1187 * intentionally. Although it's rather weird,
1188 * it's how HWPoison flag works at the moment.
1190 if (set_hwpoison_free_buddy_page(page))
1191 num_poisoned_pages_inc();
1194 if (rc != -EAGAIN) {
1195 if (likely(!__PageMovable(page))) {
1196 putback_lru_page(page);
1201 if (PageMovable(page))
1202 putback_movable_page(page);
1204 __ClearPageIsolated(page);
1210 put_new_page(newpage, private);
1219 * Counterpart of unmap_and_move_page() for hugepage migration.
1221 * This function doesn't wait the completion of hugepage I/O
1222 * because there is no race between I/O and migration for hugepage.
1223 * Note that currently hugepage I/O occurs only in direct I/O
1224 * where no lock is held and PG_writeback is irrelevant,
1225 * and writeback status of all subpages are counted in the reference
1226 * count of the head page (i.e. if all subpages of a 2MB hugepage are
1227 * under direct I/O, the reference of the head page is 512 and a bit more.)
1228 * This means that when we try to migrate hugepage whose subpages are
1229 * doing direct I/O, some references remain after try_to_unmap() and
1230 * hugepage migration fails without data corruption.
1232 * There is also no race when direct I/O is issued on the page under migration,
1233 * because then pte is replaced with migration swap entry and direct I/O code
1234 * will wait in the page fault for migration to complete.
1236 static int unmap_and_move_huge_page(new_page_t get_new_page,
1237 free_page_t put_new_page, unsigned long private,
1238 struct page *hpage, int force,
1239 enum migrate_mode mode, int reason)
1242 int page_was_mapped = 0;
1243 struct page *new_hpage;
1244 struct anon_vma *anon_vma = NULL;
1247 * Movability of hugepages depends on architectures and hugepage size.
1248 * This check is necessary because some callers of hugepage migration
1249 * like soft offline and memory hotremove don't walk through page
1250 * tables or check whether the hugepage is pmd-based or not before
1251 * kicking migration.
1253 if (!hugepage_migration_supported(page_hstate(hpage))) {
1254 putback_active_hugepage(hpage);
1258 new_hpage = get_new_page(hpage, private);
1262 if (!trylock_page(hpage)) {
1267 case MIGRATE_SYNC_NO_COPY:
1275 if (PageAnon(hpage))
1276 anon_vma = page_get_anon_vma(hpage);
1278 if (unlikely(!trylock_page(new_hpage)))
1281 if (page_mapped(hpage)) {
1283 TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
1284 page_was_mapped = 1;
1287 if (!page_mapped(hpage))
1288 rc = move_to_new_page(new_hpage, hpage, mode);
1290 if (page_was_mapped)
1291 remove_migration_ptes(hpage,
1292 rc == MIGRATEPAGE_SUCCESS ? new_hpage : hpage, false);
1294 unlock_page(new_hpage);
1298 put_anon_vma(anon_vma);
1300 if (rc == MIGRATEPAGE_SUCCESS) {
1301 move_hugetlb_state(hpage, new_hpage, reason);
1302 put_new_page = NULL;
1308 putback_active_hugepage(hpage);
1311 * If migration was not successful and there's a freeing callback, use
1312 * it. Otherwise, put_page() will drop the reference grabbed during
1316 put_new_page(new_hpage, private);
1318 putback_active_hugepage(new_hpage);
1324 * migrate_pages - migrate the pages specified in a list, to the free pages
1325 * supplied as the target for the page migration
1327 * @from: The list of pages to be migrated.
1328 * @get_new_page: The function used to allocate free pages to be used
1329 * as the target of the page migration.
1330 * @put_new_page: The function used to free target pages if migration
1331 * fails, or NULL if no special handling is necessary.
1332 * @private: Private data to be passed on to get_new_page()
1333 * @mode: The migration mode that specifies the constraints for
1334 * page migration, if any.
1335 * @reason: The reason for page migration.
1337 * The function returns after 10 attempts or if no pages are movable any more
1338 * because the list has become empty or no retryable pages exist any more.
1339 * The caller should call putback_movable_pages() to return pages to the LRU
1340 * or free list only if ret != 0.
1342 * Returns the number of pages that were not migrated, or an error code.
1344 int migrate_pages(struct list_head *from, new_page_t get_new_page,
1345 free_page_t put_new_page, unsigned long private,
1346 enum migrate_mode mode, int reason)
1350 int nr_succeeded = 0;
1354 int swapwrite = current->flags & PF_SWAPWRITE;
1358 current->flags |= PF_SWAPWRITE;
1360 for(pass = 0; pass < 10 && retry; pass++) {
1363 list_for_each_entry_safe(page, page2, from, lru) {
1368 rc = unmap_and_move_huge_page(get_new_page,
1369 put_new_page, private, page,
1370 pass > 2, mode, reason);
1372 rc = unmap_and_move(get_new_page, put_new_page,
1373 private, page, pass > 2, mode,
1379 * THP migration might be unsupported or the
1380 * allocation could've failed so we should
1381 * retry on the same page with the THP split
1384 * Head page is retried immediately and tail
1385 * pages are added to the tail of the list so
1386 * we encounter them after the rest of the list
1389 if (PageTransHuge(page) && !PageHuge(page)) {
1391 rc = split_huge_page_to_list(page, from);
1394 list_safe_reset_next(page, page2, lru);
1403 case MIGRATEPAGE_SUCCESS:
1408 * Permanent failure (-EBUSY, -ENOSYS, etc.):
1409 * unlike -EAGAIN case, the failed page is
1410 * removed from migration page list and not
1411 * retried in the next outer loop.
1422 count_vm_events(PGMIGRATE_SUCCESS, nr_succeeded);
1424 count_vm_events(PGMIGRATE_FAIL, nr_failed);
1425 trace_mm_migrate_pages(nr_succeeded, nr_failed, mode, reason);
1428 current->flags &= ~PF_SWAPWRITE;
1435 static int store_status(int __user *status, int start, int value, int nr)
1438 if (put_user(value, status + start))
1446 static int do_move_pages_to_node(struct mm_struct *mm,
1447 struct list_head *pagelist, int node)
1451 if (list_empty(pagelist))
1454 err = migrate_pages(pagelist, alloc_new_node_page, NULL, node,
1455 MIGRATE_SYNC, MR_SYSCALL);
1457 putback_movable_pages(pagelist);
1462 * Resolves the given address to a struct page, isolates it from the LRU and
1463 * puts it to the given pagelist.
1464 * Returns -errno if the page cannot be found/isolated or 0 when it has been
1465 * queued or the page doesn't need to be migrated because it is already on
1468 static int add_page_for_migration(struct mm_struct *mm, unsigned long addr,
1469 int node, struct list_head *pagelist, bool migrate_all)
1471 struct vm_area_struct *vma;
1473 unsigned int follflags;
1476 down_read(&mm->mmap_sem);
1478 vma = find_vma(mm, addr);
1479 if (!vma || addr < vma->vm_start || !vma_migratable(vma))
1482 /* FOLL_DUMP to ignore special (like zero) pages */
1483 follflags = FOLL_GET | FOLL_DUMP;
1484 page = follow_page(vma, addr, follflags);
1486 err = PTR_ERR(page);
1495 if (page_to_nid(page) == node)
1499 if (page_mapcount(page) > 1 && !migrate_all)
1502 if (PageHuge(page)) {
1503 if (PageHead(page)) {
1504 isolate_huge_page(page, pagelist);
1510 head = compound_head(page);
1511 err = isolate_lru_page(head);
1516 list_add_tail(&head->lru, pagelist);
1517 mod_node_page_state(page_pgdat(head),
1518 NR_ISOLATED_ANON + page_is_file_cache(head),
1519 hpage_nr_pages(head));
1523 * Either remove the duplicate refcount from
1524 * isolate_lru_page() or drop the page ref if it was
1529 up_read(&mm->mmap_sem);
1534 * Migrate an array of page address onto an array of nodes and fill
1535 * the corresponding array of status.
1537 static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes,
1538 unsigned long nr_pages,
1539 const void __user * __user *pages,
1540 const int __user *nodes,
1541 int __user *status, int flags)
1543 int current_node = NUMA_NO_NODE;
1544 LIST_HEAD(pagelist);
1550 for (i = start = 0; i < nr_pages; i++) {
1551 const void __user *p;
1556 if (get_user(p, pages + i))
1558 if (get_user(node, nodes + i))
1560 addr = (unsigned long)p;
1563 if (node < 0 || node >= MAX_NUMNODES)
1565 if (!node_state(node, N_MEMORY))
1569 if (!node_isset(node, task_nodes))
1572 if (current_node == NUMA_NO_NODE) {
1573 current_node = node;
1575 } else if (node != current_node) {
1576 err = do_move_pages_to_node(mm, &pagelist, current_node);
1579 err = store_status(status, start, current_node, i - start);
1583 current_node = node;
1587 * Errors in the page lookup or isolation are not fatal and we simply
1588 * report them via status
1590 err = add_page_for_migration(mm, addr, current_node,
1591 &pagelist, flags & MPOL_MF_MOVE_ALL);
1595 err = store_status(status, i, err, 1);
1599 err = do_move_pages_to_node(mm, &pagelist, current_node);
1603 err = store_status(status, start, current_node, i - start);
1607 current_node = NUMA_NO_NODE;
1610 if (list_empty(&pagelist))
1613 /* Make sure we do not overwrite the existing error */
1614 err1 = do_move_pages_to_node(mm, &pagelist, current_node);
1616 err1 = store_status(status, start, current_node, i - start);
1624 * Determine the nodes of an array of pages and store it in an array of status.
1626 static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages,
1627 const void __user **pages, int *status)
1631 down_read(&mm->mmap_sem);
1633 for (i = 0; i < nr_pages; i++) {
1634 unsigned long addr = (unsigned long)(*pages);
1635 struct vm_area_struct *vma;
1639 vma = find_vma(mm, addr);
1640 if (!vma || addr < vma->vm_start)
1643 /* FOLL_DUMP to ignore special (like zero) pages */
1644 page = follow_page(vma, addr, FOLL_DUMP);
1646 err = PTR_ERR(page);
1650 err = page ? page_to_nid(page) : -ENOENT;
1658 up_read(&mm->mmap_sem);
1662 * Determine the nodes of a user array of pages and store it in
1663 * a user array of status.
1665 static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages,
1666 const void __user * __user *pages,
1669 #define DO_PAGES_STAT_CHUNK_NR 16
1670 const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR];
1671 int chunk_status[DO_PAGES_STAT_CHUNK_NR];
1674 unsigned long chunk_nr;
1676 chunk_nr = nr_pages;
1677 if (chunk_nr > DO_PAGES_STAT_CHUNK_NR)
1678 chunk_nr = DO_PAGES_STAT_CHUNK_NR;
1680 if (copy_from_user(chunk_pages, pages, chunk_nr * sizeof(*chunk_pages)))
1683 do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status);
1685 if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status)))
1690 nr_pages -= chunk_nr;
1692 return nr_pages ? -EFAULT : 0;
1696 * Move a list of pages in the address space of the currently executing
1699 static int kernel_move_pages(pid_t pid, unsigned long nr_pages,
1700 const void __user * __user *pages,
1701 const int __user *nodes,
1702 int __user *status, int flags)
1704 struct task_struct *task;
1705 struct mm_struct *mm;
1707 nodemask_t task_nodes;
1710 if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
1713 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1716 /* Find the mm_struct */
1718 task = pid ? find_task_by_vpid(pid) : current;
1723 get_task_struct(task);
1726 * Check if this process has the right to modify the specified
1727 * process. Use the regular "ptrace_may_access()" checks.
1729 if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) {
1736 err = security_task_movememory(task);
1740 task_nodes = cpuset_mems_allowed(task);
1741 mm = get_task_mm(task);
1742 put_task_struct(task);
1748 err = do_pages_move(mm, task_nodes, nr_pages, pages,
1749 nodes, status, flags);
1751 err = do_pages_stat(mm, nr_pages, pages, status);
1757 put_task_struct(task);
1761 SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages,
1762 const void __user * __user *, pages,
1763 const int __user *, nodes,
1764 int __user *, status, int, flags)
1766 return kernel_move_pages(pid, nr_pages, pages, nodes, status, flags);
1769 #ifdef CONFIG_COMPAT
1770 COMPAT_SYSCALL_DEFINE6(move_pages, pid_t, pid, compat_ulong_t, nr_pages,
1771 compat_uptr_t __user *, pages32,
1772 const int __user *, nodes,
1773 int __user *, status,
1776 const void __user * __user *pages;
1779 pages = compat_alloc_user_space(nr_pages * sizeof(void *));
1780 for (i = 0; i < nr_pages; i++) {
1783 if (get_user(p, pages32 + i) ||
1784 put_user(compat_ptr(p), pages + i))
1787 return kernel_move_pages(pid, nr_pages, pages, nodes, status, flags);
1789 #endif /* CONFIG_COMPAT */
1791 #ifdef CONFIG_NUMA_BALANCING
1793 * Returns true if this is a safe migration target node for misplaced NUMA
1794 * pages. Currently it only checks the watermarks which crude
1796 static bool migrate_balanced_pgdat(struct pglist_data *pgdat,
1797 unsigned long nr_migrate_pages)
1801 for (z = pgdat->nr_zones - 1; z >= 0; z--) {
1802 struct zone *zone = pgdat->node_zones + z;
1804 if (!populated_zone(zone))
1807 /* Avoid waking kswapd by allocating pages_to_migrate pages. */
1808 if (!zone_watermark_ok(zone, 0,
1809 high_wmark_pages(zone) +
1818 static struct page *alloc_misplaced_dst_page(struct page *page,
1821 int nid = (int) data;
1822 struct page *newpage;
1824 newpage = __alloc_pages_node(nid,
1825 (GFP_HIGHUSER_MOVABLE |
1826 __GFP_THISNODE | __GFP_NOMEMALLOC |
1827 __GFP_NORETRY | __GFP_NOWARN) &
1833 static int numamigrate_isolate_page(pg_data_t *pgdat, struct page *page)
1837 VM_BUG_ON_PAGE(compound_order(page) && !PageTransHuge(page), page);
1839 /* Avoid migrating to a node that is nearly full */
1840 if (!migrate_balanced_pgdat(pgdat, 1UL << compound_order(page)))
1843 if (isolate_lru_page(page))
1847 * migrate_misplaced_transhuge_page() skips page migration's usual
1848 * check on page_count(), so we must do it here, now that the page
1849 * has been isolated: a GUP pin, or any other pin, prevents migration.
1850 * The expected page count is 3: 1 for page's mapcount and 1 for the
1851 * caller's pin and 1 for the reference taken by isolate_lru_page().
1853 if (PageTransHuge(page) && page_count(page) != 3) {
1854 putback_lru_page(page);
1858 page_lru = page_is_file_cache(page);
1859 mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON + page_lru,
1860 hpage_nr_pages(page));
1863 * Isolating the page has taken another reference, so the
1864 * caller's reference can be safely dropped without the page
1865 * disappearing underneath us during migration.
1871 bool pmd_trans_migrating(pmd_t pmd)
1873 struct page *page = pmd_page(pmd);
1874 return PageLocked(page);
1878 * Attempt to migrate a misplaced page to the specified destination
1879 * node. Caller is expected to have an elevated reference count on
1880 * the page that will be dropped by this function before returning.
1882 int migrate_misplaced_page(struct page *page, struct vm_area_struct *vma,
1885 pg_data_t *pgdat = NODE_DATA(node);
1888 LIST_HEAD(migratepages);
1891 * Don't migrate file pages that are mapped in multiple processes
1892 * with execute permissions as they are probably shared libraries.
1894 if (page_mapcount(page) != 1 && page_is_file_cache(page) &&
1895 (vma->vm_flags & VM_EXEC))
1899 * Also do not migrate dirty pages as not all filesystems can move
1900 * dirty pages in MIGRATE_ASYNC mode which is a waste of cycles.
1902 if (page_is_file_cache(page) && PageDirty(page))
1905 isolated = numamigrate_isolate_page(pgdat, page);
1909 list_add(&page->lru, &migratepages);
1910 nr_remaining = migrate_pages(&migratepages, alloc_misplaced_dst_page,
1911 NULL, node, MIGRATE_ASYNC,
1914 if (!list_empty(&migratepages)) {
1915 list_del(&page->lru);
1916 dec_node_page_state(page, NR_ISOLATED_ANON +
1917 page_is_file_cache(page));
1918 putback_lru_page(page);
1922 count_vm_numa_event(NUMA_PAGE_MIGRATE);
1923 BUG_ON(!list_empty(&migratepages));
1930 #endif /* CONFIG_NUMA_BALANCING */
1932 #if defined(CONFIG_NUMA_BALANCING) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
1934 * Migrates a THP to a given target node. page must be locked and is unlocked
1937 int migrate_misplaced_transhuge_page(struct mm_struct *mm,
1938 struct vm_area_struct *vma,
1939 pmd_t *pmd, pmd_t entry,
1940 unsigned long address,
1941 struct page *page, int node)
1944 pg_data_t *pgdat = NODE_DATA(node);
1946 struct page *new_page = NULL;
1947 int page_lru = page_is_file_cache(page);
1948 unsigned long start = address & HPAGE_PMD_MASK;
1950 new_page = alloc_pages_node(node,
1951 (GFP_TRANSHUGE_LIGHT | __GFP_THISNODE),
1955 prep_transhuge_page(new_page);
1957 isolated = numamigrate_isolate_page(pgdat, page);
1963 /* Prepare a page as a migration target */
1964 __SetPageLocked(new_page);
1965 if (PageSwapBacked(page))
1966 __SetPageSwapBacked(new_page);
1968 /* anon mapping, we can simply copy page->mapping to the new page: */
1969 new_page->mapping = page->mapping;
1970 new_page->index = page->index;
1971 /* flush the cache before copying using the kernel virtual address */
1972 flush_cache_range(vma, start, start + HPAGE_PMD_SIZE);
1973 migrate_page_copy(new_page, page);
1974 WARN_ON(PageLRU(new_page));
1976 /* Recheck the target PMD */
1977 ptl = pmd_lock(mm, pmd);
1978 if (unlikely(!pmd_same(*pmd, entry) || !page_ref_freeze(page, 2))) {
1981 /* Reverse changes made by migrate_page_copy() */
1982 if (TestClearPageActive(new_page))
1983 SetPageActive(page);
1984 if (TestClearPageUnevictable(new_page))
1985 SetPageUnevictable(page);
1987 unlock_page(new_page);
1988 put_page(new_page); /* Free it */
1990 /* Retake the callers reference and putback on LRU */
1992 putback_lru_page(page);
1993 mod_node_page_state(page_pgdat(page),
1994 NR_ISOLATED_ANON + page_lru, -HPAGE_PMD_NR);
1999 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
2000 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
2003 * Overwrite the old entry under pagetable lock and establish
2004 * the new PTE. Any parallel GUP will either observe the old
2005 * page blocking on the page lock, block on the page table
2006 * lock or observe the new page. The SetPageUptodate on the
2007 * new page and page_add_new_anon_rmap guarantee the copy is
2008 * visible before the pagetable update.
2010 page_add_anon_rmap(new_page, vma, start, true);
2012 * At this point the pmd is numa/protnone (i.e. non present) and the TLB
2013 * has already been flushed globally. So no TLB can be currently
2014 * caching this non present pmd mapping. There's no need to clear the
2015 * pmd before doing set_pmd_at(), nor to flush the TLB after
2016 * set_pmd_at(). Clearing the pmd here would introduce a race
2017 * condition against MADV_DONTNEED, because MADV_DONTNEED only holds the
2018 * mmap_sem for reading. If the pmd is set to NULL at any given time,
2019 * MADV_DONTNEED won't wait on the pmd lock and it'll skip clearing this
2022 set_pmd_at(mm, start, pmd, entry);
2023 update_mmu_cache_pmd(vma, address, &entry);
2025 page_ref_unfreeze(page, 2);
2026 mlock_migrate_page(new_page, page);
2027 page_remove_rmap(page, true);
2028 set_page_owner_migrate_reason(new_page, MR_NUMA_MISPLACED);
2032 /* Take an "isolate" reference and put new page on the LRU. */
2034 putback_lru_page(new_page);
2036 unlock_page(new_page);
2038 put_page(page); /* Drop the rmap reference */
2039 put_page(page); /* Drop the LRU isolation reference */
2041 count_vm_events(PGMIGRATE_SUCCESS, HPAGE_PMD_NR);
2042 count_vm_numa_events(NUMA_PAGE_MIGRATE, HPAGE_PMD_NR);
2044 mod_node_page_state(page_pgdat(page),
2045 NR_ISOLATED_ANON + page_lru,
2050 count_vm_events(PGMIGRATE_FAIL, HPAGE_PMD_NR);
2051 ptl = pmd_lock(mm, pmd);
2052 if (pmd_same(*pmd, entry)) {
2053 entry = pmd_modify(entry, vma->vm_page_prot);
2054 set_pmd_at(mm, start, pmd, entry);
2055 update_mmu_cache_pmd(vma, address, &entry);
2064 #endif /* CONFIG_NUMA_BALANCING */
2066 #endif /* CONFIG_NUMA */
2068 #if defined(CONFIG_MIGRATE_VMA_HELPER)
2069 struct migrate_vma {
2070 struct vm_area_struct *vma;
2073 unsigned long cpages;
2074 unsigned long npages;
2075 unsigned long start;
2079 static int migrate_vma_collect_hole(unsigned long start,
2081 struct mm_walk *walk)
2083 struct migrate_vma *migrate = walk->private;
2086 for (addr = start & PAGE_MASK; addr < end; addr += PAGE_SIZE) {
2087 migrate->src[migrate->npages] = MIGRATE_PFN_MIGRATE;
2088 migrate->dst[migrate->npages] = 0;
2096 static int migrate_vma_collect_skip(unsigned long start,
2098 struct mm_walk *walk)
2100 struct migrate_vma *migrate = walk->private;
2103 for (addr = start & PAGE_MASK; addr < end; addr += PAGE_SIZE) {
2104 migrate->dst[migrate->npages] = 0;
2105 migrate->src[migrate->npages++] = 0;
2111 static int migrate_vma_collect_pmd(pmd_t *pmdp,
2112 unsigned long start,
2114 struct mm_walk *walk)
2116 struct migrate_vma *migrate = walk->private;
2117 struct vm_area_struct *vma = walk->vma;
2118 struct mm_struct *mm = vma->vm_mm;
2119 unsigned long addr = start, unmapped = 0;
2124 if (pmd_none(*pmdp))
2125 return migrate_vma_collect_hole(start, end, walk);
2127 if (pmd_trans_huge(*pmdp)) {
2130 ptl = pmd_lock(mm, pmdp);
2131 if (unlikely(!pmd_trans_huge(*pmdp))) {
2136 page = pmd_page(*pmdp);
2137 if (is_huge_zero_page(page)) {
2139 split_huge_pmd(vma, pmdp, addr);
2140 if (pmd_trans_unstable(pmdp))
2141 return migrate_vma_collect_skip(start, end,
2148 if (unlikely(!trylock_page(page)))
2149 return migrate_vma_collect_skip(start, end,
2151 ret = split_huge_page(page);
2155 return migrate_vma_collect_skip(start, end,
2157 if (pmd_none(*pmdp))
2158 return migrate_vma_collect_hole(start, end,
2163 if (unlikely(pmd_bad(*pmdp)))
2164 return migrate_vma_collect_skip(start, end, walk);
2166 ptep = pte_offset_map_lock(mm, pmdp, addr, &ptl);
2167 arch_enter_lazy_mmu_mode();
2169 for (; addr < end; addr += PAGE_SIZE, ptep++) {
2170 unsigned long mpfn, pfn;
2178 if (pte_none(pte)) {
2179 mpfn = MIGRATE_PFN_MIGRATE;
2185 if (!pte_present(pte)) {
2189 * Only care about unaddressable device page special
2190 * page table entry. Other special swap entries are not
2191 * migratable, and we ignore regular swapped page.
2193 entry = pte_to_swp_entry(pte);
2194 if (!is_device_private_entry(entry))
2197 page = device_private_entry_to_page(entry);
2198 mpfn = migrate_pfn(page_to_pfn(page))|
2199 MIGRATE_PFN_DEVICE | MIGRATE_PFN_MIGRATE;
2200 if (is_write_device_private_entry(entry))
2201 mpfn |= MIGRATE_PFN_WRITE;
2203 if (is_zero_pfn(pfn)) {
2204 mpfn = MIGRATE_PFN_MIGRATE;
2209 page = _vm_normal_page(migrate->vma, addr, pte, true);
2210 mpfn = migrate_pfn(pfn) | MIGRATE_PFN_MIGRATE;
2211 mpfn |= pte_write(pte) ? MIGRATE_PFN_WRITE : 0;
2214 /* FIXME support THP */
2215 if (!page || !page->mapping || PageTransCompound(page)) {
2219 pfn = page_to_pfn(page);
2222 * By getting a reference on the page we pin it and that blocks
2223 * any kind of migration. Side effect is that it "freezes" the
2226 * We drop this reference after isolating the page from the lru
2227 * for non device page (device page are not on the lru and thus
2228 * can't be dropped from it).
2234 * Optimize for the common case where page is only mapped once
2235 * in one process. If we can lock the page, then we can safely
2236 * set up a special migration page table entry now.
2238 if (trylock_page(page)) {
2241 mpfn |= MIGRATE_PFN_LOCKED;
2242 ptep_get_and_clear(mm, addr, ptep);
2244 /* Setup special migration page table entry */
2245 entry = make_migration_entry(page, mpfn &
2247 swp_pte = swp_entry_to_pte(entry);
2248 if (pte_soft_dirty(pte))
2249 swp_pte = pte_swp_mksoft_dirty(swp_pte);
2250 set_pte_at(mm, addr, ptep, swp_pte);
2253 * This is like regular unmap: we remove the rmap and
2254 * drop page refcount. Page won't be freed, as we took
2255 * a reference just above.
2257 page_remove_rmap(page, false);
2260 if (pte_present(pte))
2265 migrate->dst[migrate->npages] = 0;
2266 migrate->src[migrate->npages++] = mpfn;
2268 arch_leave_lazy_mmu_mode();
2269 pte_unmap_unlock(ptep - 1, ptl);
2271 /* Only flush the TLB if we actually modified any entries */
2273 flush_tlb_range(walk->vma, start, end);
2279 * migrate_vma_collect() - collect pages over a range of virtual addresses
2280 * @migrate: migrate struct containing all migration information
2282 * This will walk the CPU page table. For each virtual address backed by a
2283 * valid page, it updates the src array and takes a reference on the page, in
2284 * order to pin the page until we lock it and unmap it.
2286 static void migrate_vma_collect(struct migrate_vma *migrate)
2288 struct mmu_notifier_range range;
2289 struct mm_walk mm_walk;
2291 mm_walk.pmd_entry = migrate_vma_collect_pmd;
2292 mm_walk.pte_entry = NULL;
2293 mm_walk.pte_hole = migrate_vma_collect_hole;
2294 mm_walk.hugetlb_entry = NULL;
2295 mm_walk.test_walk = NULL;
2296 mm_walk.vma = migrate->vma;
2297 mm_walk.mm = migrate->vma->vm_mm;
2298 mm_walk.private = migrate;
2300 mmu_notifier_range_init(&range, mm_walk.mm, migrate->start,
2302 mmu_notifier_invalidate_range_start(&range);
2303 walk_page_range(migrate->start, migrate->end, &mm_walk);
2304 mmu_notifier_invalidate_range_end(&range);
2306 migrate->end = migrate->start + (migrate->npages << PAGE_SHIFT);
2310 * migrate_vma_check_page() - check if page is pinned or not
2311 * @page: struct page to check
2313 * Pinned pages cannot be migrated. This is the same test as in
2314 * migrate_page_move_mapping(), except that here we allow migration of a
2317 static bool migrate_vma_check_page(struct page *page)
2320 * One extra ref because caller holds an extra reference, either from
2321 * isolate_lru_page() for a regular page, or migrate_vma_collect() for
2327 * FIXME support THP (transparent huge page), it is bit more complex to
2328 * check them than regular pages, because they can be mapped with a pmd
2329 * or with a pte (split pte mapping).
2331 if (PageCompound(page))
2334 /* Page from ZONE_DEVICE have one extra reference */
2335 if (is_zone_device_page(page)) {
2337 * Private page can never be pin as they have no valid pte and
2338 * GUP will fail for those. Yet if there is a pending migration
2339 * a thread might try to wait on the pte migration entry and
2340 * will bump the page reference count. Sadly there is no way to
2341 * differentiate a regular pin from migration wait. Hence to
2342 * avoid 2 racing thread trying to migrate back to CPU to enter
2343 * infinite loop (one stoping migration because the other is
2344 * waiting on pte migration entry). We always return true here.
2346 * FIXME proper solution is to rework migration_entry_wait() so
2347 * it does not need to take a reference on page.
2349 if (is_device_private_page(page))
2353 * Only allow device public page to be migrated and account for
2354 * the extra reference count imply by ZONE_DEVICE pages.
2356 if (!is_device_public_page(page))
2361 /* For file back page */
2362 if (page_mapping(page))
2363 extra += 1 + page_has_private(page);
2365 if ((page_count(page) - extra) > page_mapcount(page))
2372 * migrate_vma_prepare() - lock pages and isolate them from the lru
2373 * @migrate: migrate struct containing all migration information
2375 * This locks pages that have been collected by migrate_vma_collect(). Once each
2376 * page is locked it is isolated from the lru (for non-device pages). Finally,
2377 * the ref taken by migrate_vma_collect() is dropped, as locked pages cannot be
2378 * migrated by concurrent kernel threads.
2380 static void migrate_vma_prepare(struct migrate_vma *migrate)
2382 const unsigned long npages = migrate->npages;
2383 const unsigned long start = migrate->start;
2384 unsigned long addr, i, restore = 0;
2385 bool allow_drain = true;
2389 for (i = 0; (i < npages) && migrate->cpages; i++) {
2390 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2396 if (!(migrate->src[i] & MIGRATE_PFN_LOCKED)) {
2398 * Because we are migrating several pages there can be
2399 * a deadlock between 2 concurrent migration where each
2400 * are waiting on each other page lock.
2402 * Make migrate_vma() a best effort thing and backoff
2403 * for any page we can not lock right away.
2405 if (!trylock_page(page)) {
2406 migrate->src[i] = 0;
2412 migrate->src[i] |= MIGRATE_PFN_LOCKED;
2415 /* ZONE_DEVICE pages are not on LRU */
2416 if (!is_zone_device_page(page)) {
2417 if (!PageLRU(page) && allow_drain) {
2418 /* Drain CPU's pagevec */
2419 lru_add_drain_all();
2420 allow_drain = false;
2423 if (isolate_lru_page(page)) {
2425 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2429 migrate->src[i] = 0;
2437 /* Drop the reference we took in collect */
2441 if (!migrate_vma_check_page(page)) {
2443 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2447 if (!is_zone_device_page(page)) {
2449 putback_lru_page(page);
2452 migrate->src[i] = 0;
2456 if (!is_zone_device_page(page))
2457 putback_lru_page(page);
2464 for (i = 0, addr = start; i < npages && restore; i++, addr += PAGE_SIZE) {
2465 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2467 if (!page || (migrate->src[i] & MIGRATE_PFN_MIGRATE))
2470 remove_migration_pte(page, migrate->vma, addr, page);
2472 migrate->src[i] = 0;
2480 * migrate_vma_unmap() - replace page mapping with special migration pte entry
2481 * @migrate: migrate struct containing all migration information
2483 * Replace page mapping (CPU page table pte) with a special migration pte entry
2484 * and check again if it has been pinned. Pinned pages are restored because we
2485 * cannot migrate them.
2487 * This is the last step before we call the device driver callback to allocate
2488 * destination memory and copy contents of original page over to new page.
2490 static void migrate_vma_unmap(struct migrate_vma *migrate)
2492 int flags = TTU_MIGRATION | TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS;
2493 const unsigned long npages = migrate->npages;
2494 const unsigned long start = migrate->start;
2495 unsigned long addr, i, restore = 0;
2497 for (i = 0; i < npages; i++) {
2498 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2500 if (!page || !(migrate->src[i] & MIGRATE_PFN_MIGRATE))
2503 if (page_mapped(page)) {
2504 try_to_unmap(page, flags);
2505 if (page_mapped(page))
2509 if (migrate_vma_check_page(page))
2513 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2518 for (addr = start, i = 0; i < npages && restore; addr += PAGE_SIZE, i++) {
2519 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2521 if (!page || (migrate->src[i] & MIGRATE_PFN_MIGRATE))
2524 remove_migration_ptes(page, page, false);
2526 migrate->src[i] = 0;
2530 if (is_zone_device_page(page))
2533 putback_lru_page(page);
2537 static void migrate_vma_insert_page(struct migrate_vma *migrate,
2543 struct vm_area_struct *vma = migrate->vma;
2544 struct mm_struct *mm = vma->vm_mm;
2545 struct mem_cgroup *memcg;
2555 /* Only allow populating anonymous memory */
2556 if (!vma_is_anonymous(vma))
2559 pgdp = pgd_offset(mm, addr);
2560 p4dp = p4d_alloc(mm, pgdp, addr);
2563 pudp = pud_alloc(mm, p4dp, addr);
2566 pmdp = pmd_alloc(mm, pudp, addr);
2570 if (pmd_trans_huge(*pmdp) || pmd_devmap(*pmdp))
2574 * Use pte_alloc() instead of pte_alloc_map(). We can't run
2575 * pte_offset_map() on pmds where a huge pmd might be created
2576 * from a different thread.
2578 * pte_alloc_map() is safe to use under down_write(mmap_sem) or when
2579 * parallel threads are excluded by other means.
2581 * Here we only have down_read(mmap_sem).
2583 if (pte_alloc(mm, pmdp, addr))
2586 /* See the comment in pte_alloc_one_map() */
2587 if (unlikely(pmd_trans_unstable(pmdp)))
2590 if (unlikely(anon_vma_prepare(vma)))
2592 if (mem_cgroup_try_charge(page, vma->vm_mm, GFP_KERNEL, &memcg, false))
2596 * The memory barrier inside __SetPageUptodate makes sure that
2597 * preceding stores to the page contents become visible before
2598 * the set_pte_at() write.
2600 __SetPageUptodate(page);
2602 if (is_zone_device_page(page)) {
2603 if (is_device_private_page(page)) {
2604 swp_entry_t swp_entry;
2606 swp_entry = make_device_private_entry(page, vma->vm_flags & VM_WRITE);
2607 entry = swp_entry_to_pte(swp_entry);
2608 } else if (is_device_public_page(page)) {
2609 entry = pte_mkold(mk_pte(page, READ_ONCE(vma->vm_page_prot)));
2610 if (vma->vm_flags & VM_WRITE)
2611 entry = pte_mkwrite(pte_mkdirty(entry));
2612 entry = pte_mkdevmap(entry);
2615 entry = mk_pte(page, vma->vm_page_prot);
2616 if (vma->vm_flags & VM_WRITE)
2617 entry = pte_mkwrite(pte_mkdirty(entry));
2620 ptep = pte_offset_map_lock(mm, pmdp, addr, &ptl);
2622 if (pte_present(*ptep)) {
2623 unsigned long pfn = pte_pfn(*ptep);
2625 if (!is_zero_pfn(pfn)) {
2626 pte_unmap_unlock(ptep, ptl);
2627 mem_cgroup_cancel_charge(page, memcg, false);
2631 } else if (!pte_none(*ptep)) {
2632 pte_unmap_unlock(ptep, ptl);
2633 mem_cgroup_cancel_charge(page, memcg, false);
2638 * Check for usefaultfd but do not deliver the fault. Instead,
2641 if (userfaultfd_missing(vma)) {
2642 pte_unmap_unlock(ptep, ptl);
2643 mem_cgroup_cancel_charge(page, memcg, false);
2647 inc_mm_counter(mm, MM_ANONPAGES);
2648 page_add_new_anon_rmap(page, vma, addr, false);
2649 mem_cgroup_commit_charge(page, memcg, false, false);
2650 if (!is_zone_device_page(page))
2651 lru_cache_add_active_or_unevictable(page, vma);
2655 flush_cache_page(vma, addr, pte_pfn(*ptep));
2656 ptep_clear_flush_notify(vma, addr, ptep);
2657 set_pte_at_notify(mm, addr, ptep, entry);
2658 update_mmu_cache(vma, addr, ptep);
2660 /* No need to invalidate - it was non-present before */
2661 set_pte_at(mm, addr, ptep, entry);
2662 update_mmu_cache(vma, addr, ptep);
2665 pte_unmap_unlock(ptep, ptl);
2666 *src = MIGRATE_PFN_MIGRATE;
2670 *src &= ~MIGRATE_PFN_MIGRATE;
2674 * migrate_vma_pages() - migrate meta-data from src page to dst page
2675 * @migrate: migrate struct containing all migration information
2677 * This migrates struct page meta-data from source struct page to destination
2678 * struct page. This effectively finishes the migration from source page to the
2681 static void migrate_vma_pages(struct migrate_vma *migrate)
2683 const unsigned long npages = migrate->npages;
2684 const unsigned long start = migrate->start;
2685 struct mmu_notifier_range range;
2686 unsigned long addr, i;
2687 bool notified = false;
2689 for (i = 0, addr = start; i < npages; addr += PAGE_SIZE, i++) {
2690 struct page *newpage = migrate_pfn_to_page(migrate->dst[i]);
2691 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2692 struct address_space *mapping;
2696 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2701 if (!(migrate->src[i] & MIGRATE_PFN_MIGRATE)) {
2707 mmu_notifier_range_init(&range,
2708 migrate->vma->vm_mm,
2709 addr, migrate->end);
2710 mmu_notifier_invalidate_range_start(&range);
2712 migrate_vma_insert_page(migrate, addr, newpage,
2718 mapping = page_mapping(page);
2720 if (is_zone_device_page(newpage)) {
2721 if (is_device_private_page(newpage)) {
2723 * For now only support private anonymous when
2724 * migrating to un-addressable device memory.
2727 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2730 } else if (!is_device_public_page(newpage)) {
2732 * Other types of ZONE_DEVICE page are not
2735 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2740 r = migrate_page(mapping, newpage, page, MIGRATE_SYNC_NO_COPY);
2741 if (r != MIGRATEPAGE_SUCCESS)
2742 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2746 * No need to double call mmu_notifier->invalidate_range() callback as
2747 * the above ptep_clear_flush_notify() inside migrate_vma_insert_page()
2748 * did already call it.
2751 mmu_notifier_invalidate_range_only_end(&range);
2755 * migrate_vma_finalize() - restore CPU page table entry
2756 * @migrate: migrate struct containing all migration information
2758 * This replaces the special migration pte entry with either a mapping to the
2759 * new page if migration was successful for that page, or to the original page
2762 * This also unlocks the pages and puts them back on the lru, or drops the extra
2763 * refcount, for device pages.
2765 static void migrate_vma_finalize(struct migrate_vma *migrate)
2767 const unsigned long npages = migrate->npages;
2770 for (i = 0; i < npages; i++) {
2771 struct page *newpage = migrate_pfn_to_page(migrate->dst[i]);
2772 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2776 unlock_page(newpage);
2782 if (!(migrate->src[i] & MIGRATE_PFN_MIGRATE) || !newpage) {
2784 unlock_page(newpage);
2790 remove_migration_ptes(page, newpage, false);
2794 if (is_zone_device_page(page))
2797 putback_lru_page(page);
2799 if (newpage != page) {
2800 unlock_page(newpage);
2801 if (is_zone_device_page(newpage))
2804 putback_lru_page(newpage);
2810 * migrate_vma() - migrate a range of memory inside vma
2812 * @ops: migration callback for allocating destination memory and copying
2813 * @vma: virtual memory area containing the range to be migrated
2814 * @start: start address of the range to migrate (inclusive)
2815 * @end: end address of the range to migrate (exclusive)
2816 * @src: array of hmm_pfn_t containing source pfns
2817 * @dst: array of hmm_pfn_t containing destination pfns
2818 * @private: pointer passed back to each of the callback
2819 * Returns: 0 on success, error code otherwise
2821 * This function tries to migrate a range of memory virtual address range, using
2822 * callbacks to allocate and copy memory from source to destination. First it
2823 * collects all the pages backing each virtual address in the range, saving this
2824 * inside the src array. Then it locks those pages and unmaps them. Once the pages
2825 * are locked and unmapped, it checks whether each page is pinned or not. Pages
2826 * that aren't pinned have the MIGRATE_PFN_MIGRATE flag set (by this function)
2827 * in the corresponding src array entry. It then restores any pages that are
2828 * pinned, by remapping and unlocking those pages.
2830 * At this point it calls the alloc_and_copy() callback. For documentation on
2831 * what is expected from that callback, see struct migrate_vma_ops comments in
2832 * include/linux/migrate.h
2834 * After the alloc_and_copy() callback, this function goes over each entry in
2835 * the src array that has the MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE flag
2836 * set. If the corresponding entry in dst array has MIGRATE_PFN_VALID flag set,
2837 * then the function tries to migrate struct page information from the source
2838 * struct page to the destination struct page. If it fails to migrate the struct
2839 * page information, then it clears the MIGRATE_PFN_MIGRATE flag in the src
2842 * At this point all successfully migrated pages have an entry in the src
2843 * array with MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE flag set and the dst
2844 * array entry with MIGRATE_PFN_VALID flag set.
2846 * It then calls the finalize_and_map() callback. See comments for "struct
2847 * migrate_vma_ops", in include/linux/migrate.h for details about
2848 * finalize_and_map() behavior.
2850 * After the finalize_and_map() callback, for successfully migrated pages, this
2851 * function updates the CPU page table to point to new pages, otherwise it
2852 * restores the CPU page table to point to the original source pages.
2854 * Function returns 0 after the above steps, even if no pages were migrated
2855 * (The function only returns an error if any of the arguments are invalid.)
2857 * Both src and dst array must be big enough for (end - start) >> PAGE_SHIFT
2858 * unsigned long entries.
2860 int migrate_vma(const struct migrate_vma_ops *ops,
2861 struct vm_area_struct *vma,
2862 unsigned long start,
2868 struct migrate_vma migrate;
2870 /* Sanity check the arguments */
2873 if (!vma || is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_SPECIAL) ||
2876 if (start < vma->vm_start || start >= vma->vm_end)
2878 if (end <= vma->vm_start || end > vma->vm_end)
2880 if (!ops || !src || !dst || start >= end)
2883 memset(src, 0, sizeof(*src) * ((end - start) >> PAGE_SHIFT));
2886 migrate.start = start;
2892 /* Collect, and try to unmap source pages */
2893 migrate_vma_collect(&migrate);
2894 if (!migrate.cpages)
2897 /* Lock and isolate page */
2898 migrate_vma_prepare(&migrate);
2899 if (!migrate.cpages)
2903 migrate_vma_unmap(&migrate);
2904 if (!migrate.cpages)
2908 * At this point pages are locked and unmapped, and thus they have
2909 * stable content and can safely be copied to destination memory that
2910 * is allocated by the callback.
2912 * Note that migration can fail in migrate_vma_struct_page() for each
2915 ops->alloc_and_copy(vma, src, dst, start, end, private);
2917 /* This does the real migration of struct page */
2918 migrate_vma_pages(&migrate);
2920 ops->finalize_and_map(vma, src, dst, start, end, private);
2922 /* Unlock and remap pages */
2923 migrate_vma_finalize(&migrate);
2927 EXPORT_SYMBOL(migrate_vma);
2928 #endif /* defined(MIGRATE_VMA_HELPER) */