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/pagewalk.h>
42 #include <linux/pfn_t.h>
43 #include <linux/memremap.h>
44 #include <linux/userfaultfd_k.h>
45 #include <linux/balloon_compaction.h>
46 #include <linux/mmu_notifier.h>
47 #include <linux/page_idle.h>
48 #include <linux/page_owner.h>
49 #include <linux/sched/mm.h>
50 #include <linux/ptrace.h>
51 #include <linux/oom.h>
53 #include <asm/tlbflush.h>
55 #define CREATE_TRACE_POINTS
56 #include <trace/events/migrate.h>
61 * migrate_prep() needs to be called before we start compiling a list of pages
62 * to be migrated using isolate_lru_page(). If scheduling work on other CPUs is
63 * undesirable, use migrate_prep_local()
65 int migrate_prep(void)
68 * Clear the LRU lists so pages can be isolated.
69 * Note that pages may be moved off the LRU after we have
70 * drained them. Those pages will fail to migrate like other
71 * pages that may be busy.
78 /* Do the necessary work of migrate_prep but not if it involves other CPUs */
79 int migrate_prep_local(void)
86 int isolate_movable_page(struct page *page, isolate_mode_t mode)
88 struct address_space *mapping;
91 * Avoid burning cycles with pages that are yet under __free_pages(),
92 * or just got freed under us.
94 * In case we 'win' a race for a movable page being freed under us and
95 * raise its refcount preventing __free_pages() from doing its job
96 * the put_page() at the end of this block will take care of
97 * release this page, thus avoiding a nasty leakage.
99 if (unlikely(!get_page_unless_zero(page)))
103 * Check PageMovable before holding a PG_lock because page's owner
104 * assumes anybody doesn't touch PG_lock of newly allocated page
105 * so unconditionally grabbing the lock ruins page's owner side.
107 if (unlikely(!__PageMovable(page)))
110 * As movable pages are not isolated from LRU lists, concurrent
111 * compaction threads can race against page migration functions
112 * as well as race against the releasing a page.
114 * In order to avoid having an already isolated movable page
115 * being (wrongly) re-isolated while it is under migration,
116 * or to avoid attempting to isolate pages being released,
117 * lets be sure we have the page lock
118 * before proceeding with the movable page isolation steps.
120 if (unlikely(!trylock_page(page)))
123 if (!PageMovable(page) || PageIsolated(page))
124 goto out_no_isolated;
126 mapping = page_mapping(page);
127 VM_BUG_ON_PAGE(!mapping, page);
129 if (!mapping->a_ops->isolate_page(page, mode))
130 goto out_no_isolated;
132 /* Driver shouldn't use PG_isolated bit of page->flags */
133 WARN_ON_ONCE(PageIsolated(page));
134 __SetPageIsolated(page);
147 /* It should be called on page which is PG_movable */
148 void putback_movable_page(struct page *page)
150 struct address_space *mapping;
152 VM_BUG_ON_PAGE(!PageLocked(page), page);
153 VM_BUG_ON_PAGE(!PageMovable(page), page);
154 VM_BUG_ON_PAGE(!PageIsolated(page), page);
156 mapping = page_mapping(page);
157 mapping->a_ops->putback_page(page);
158 __ClearPageIsolated(page);
162 * Put previously isolated pages back onto the appropriate lists
163 * from where they were once taken off for compaction/migration.
165 * This function shall be used whenever the isolated pageset has been
166 * built from lru, balloon, hugetlbfs page. See isolate_migratepages_range()
167 * and isolate_huge_page().
169 void putback_movable_pages(struct list_head *l)
174 list_for_each_entry_safe(page, page2, l, lru) {
175 if (unlikely(PageHuge(page))) {
176 putback_active_hugepage(page);
179 list_del(&page->lru);
181 * We isolated non-lru movable page so here we can use
182 * __PageMovable because LRU page's mapping cannot have
183 * PAGE_MAPPING_MOVABLE.
185 if (unlikely(__PageMovable(page))) {
186 VM_BUG_ON_PAGE(!PageIsolated(page), page);
188 if (PageMovable(page))
189 putback_movable_page(page);
191 __ClearPageIsolated(page);
195 mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON +
196 page_is_file_cache(page), -hpage_nr_pages(page));
197 putback_lru_page(page);
203 * Restore a potential migration pte to a working pte entry
205 static bool remove_migration_pte(struct page *page, struct vm_area_struct *vma,
206 unsigned long addr, void *old)
208 struct page_vma_mapped_walk pvmw = {
212 .flags = PVMW_SYNC | PVMW_MIGRATION,
218 VM_BUG_ON_PAGE(PageTail(page), page);
219 while (page_vma_mapped_walk(&pvmw)) {
223 new = page - pvmw.page->index +
224 linear_page_index(vma, pvmw.address);
226 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
227 /* PMD-mapped THP migration entry */
229 VM_BUG_ON_PAGE(PageHuge(page) || !PageTransCompound(page), page);
230 remove_migration_pmd(&pvmw, new);
236 pte = pte_mkold(mk_pte(new, READ_ONCE(vma->vm_page_prot)));
237 if (pte_swp_soft_dirty(*pvmw.pte))
238 pte = pte_mksoft_dirty(pte);
241 * Recheck VMA as permissions can change since migration started
243 entry = pte_to_swp_entry(*pvmw.pte);
244 if (is_write_migration_entry(entry))
245 pte = maybe_mkwrite(pte, vma);
247 if (unlikely(is_zone_device_page(new))) {
248 if (is_device_private_page(new)) {
249 entry = make_device_private_entry(new, pte_write(pte));
250 pte = swp_entry_to_pte(entry);
254 #ifdef CONFIG_HUGETLB_PAGE
256 pte = pte_mkhuge(pte);
257 pte = arch_make_huge_pte(pte, vma, new, 0);
258 set_huge_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte);
260 hugepage_add_anon_rmap(new, vma, pvmw.address);
262 page_dup_rmap(new, true);
266 set_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte);
269 page_add_anon_rmap(new, vma, pvmw.address, false);
271 page_add_file_rmap(new, false);
273 if (vma->vm_flags & VM_LOCKED && !PageTransCompound(new))
276 if (PageTransHuge(page) && PageMlocked(page))
277 clear_page_mlock(page);
279 /* No need to invalidate - it was non-present before */
280 update_mmu_cache(vma, pvmw.address, pvmw.pte);
287 * Get rid of all migration entries and replace them by
288 * references to the indicated page.
290 void remove_migration_ptes(struct page *old, struct page *new, bool locked)
292 struct rmap_walk_control rwc = {
293 .rmap_one = remove_migration_pte,
298 rmap_walk_locked(new, &rwc);
300 rmap_walk(new, &rwc);
304 * Something used the pte of a page under migration. We need to
305 * get to the page and wait until migration is finished.
306 * When we return from this function the fault will be retried.
308 void __migration_entry_wait(struct mm_struct *mm, pte_t *ptep,
317 if (!is_swap_pte(pte))
320 entry = pte_to_swp_entry(pte);
321 if (!is_migration_entry(entry))
324 page = migration_entry_to_page(entry);
327 * Once page cache replacement of page migration started, page_count
328 * is zero; but we must not call put_and_wait_on_page_locked() without
329 * a ref. Use get_page_unless_zero(), and just fault again if it fails.
331 if (!get_page_unless_zero(page))
333 pte_unmap_unlock(ptep, ptl);
334 put_and_wait_on_page_locked(page);
337 pte_unmap_unlock(ptep, ptl);
340 void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
341 unsigned long address)
343 spinlock_t *ptl = pte_lockptr(mm, pmd);
344 pte_t *ptep = pte_offset_map(pmd, address);
345 __migration_entry_wait(mm, ptep, ptl);
348 void migration_entry_wait_huge(struct vm_area_struct *vma,
349 struct mm_struct *mm, pte_t *pte)
351 spinlock_t *ptl = huge_pte_lockptr(hstate_vma(vma), mm, pte);
352 __migration_entry_wait(mm, pte, ptl);
355 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
356 void pmd_migration_entry_wait(struct mm_struct *mm, pmd_t *pmd)
361 ptl = pmd_lock(mm, pmd);
362 if (!is_pmd_migration_entry(*pmd))
364 page = migration_entry_to_page(pmd_to_swp_entry(*pmd));
365 if (!get_page_unless_zero(page))
368 put_and_wait_on_page_locked(page);
375 static int expected_page_refs(struct address_space *mapping, struct page *page)
377 int expected_count = 1;
380 * Device public or private pages have an extra refcount as they are
383 expected_count += is_device_private_page(page);
385 expected_count += hpage_nr_pages(page) + page_has_private(page);
387 return expected_count;
391 * Replace the page in the mapping.
393 * The number of remaining references must be:
394 * 1 for anonymous pages without a mapping
395 * 2 for pages with a mapping
396 * 3 for pages with a mapping and PagePrivate/PagePrivate2 set.
398 int migrate_page_move_mapping(struct address_space *mapping,
399 struct page *newpage, struct page *page, int extra_count)
401 XA_STATE(xas, &mapping->i_pages, page_index(page));
402 struct zone *oldzone, *newzone;
404 int expected_count = expected_page_refs(mapping, page) + extra_count;
407 /* Anonymous page without mapping */
408 if (page_count(page) != expected_count)
411 /* No turning back from here */
412 newpage->index = page->index;
413 newpage->mapping = page->mapping;
414 if (PageSwapBacked(page))
415 __SetPageSwapBacked(newpage);
417 return MIGRATEPAGE_SUCCESS;
420 oldzone = page_zone(page);
421 newzone = page_zone(newpage);
424 if (page_count(page) != expected_count || xas_load(&xas) != page) {
425 xas_unlock_irq(&xas);
429 if (!page_ref_freeze(page, expected_count)) {
430 xas_unlock_irq(&xas);
435 * Now we know that no one else is looking at the page:
436 * no turning back from here.
438 newpage->index = page->index;
439 newpage->mapping = page->mapping;
440 page_ref_add(newpage, hpage_nr_pages(page)); /* add cache reference */
441 if (PageSwapBacked(page)) {
442 __SetPageSwapBacked(newpage);
443 if (PageSwapCache(page)) {
444 SetPageSwapCache(newpage);
445 set_page_private(newpage, page_private(page));
448 VM_BUG_ON_PAGE(PageSwapCache(page), page);
451 /* Move dirty while page refs frozen and newpage not yet exposed */
452 dirty = PageDirty(page);
454 ClearPageDirty(page);
455 SetPageDirty(newpage);
458 xas_store(&xas, newpage);
459 if (PageTransHuge(page)) {
462 for (i = 1; i < HPAGE_PMD_NR; i++) {
464 xas_store(&xas, newpage);
469 * Drop cache reference from old page by unfreezing
470 * to one less reference.
471 * We know this isn't the last reference.
473 page_ref_unfreeze(page, expected_count - hpage_nr_pages(page));
476 /* Leave irq disabled to prevent preemption while updating stats */
479 * If moved to a different zone then also account
480 * the page for that zone. Other VM counters will be
481 * taken care of when we establish references to the
482 * new page and drop references to the old page.
484 * Note that anonymous pages are accounted for
485 * via NR_FILE_PAGES and NR_ANON_MAPPED if they
486 * are mapped to swap space.
488 if (newzone != oldzone) {
489 __dec_node_state(oldzone->zone_pgdat, NR_FILE_PAGES);
490 __inc_node_state(newzone->zone_pgdat, NR_FILE_PAGES);
491 if (PageSwapBacked(page) && !PageSwapCache(page)) {
492 __dec_node_state(oldzone->zone_pgdat, NR_SHMEM);
493 __inc_node_state(newzone->zone_pgdat, NR_SHMEM);
495 if (dirty && mapping_cap_account_dirty(mapping)) {
496 __dec_node_state(oldzone->zone_pgdat, NR_FILE_DIRTY);
497 __dec_zone_state(oldzone, NR_ZONE_WRITE_PENDING);
498 __inc_node_state(newzone->zone_pgdat, NR_FILE_DIRTY);
499 __inc_zone_state(newzone, NR_ZONE_WRITE_PENDING);
504 return MIGRATEPAGE_SUCCESS;
506 EXPORT_SYMBOL(migrate_page_move_mapping);
509 * The expected number of remaining references is the same as that
510 * of migrate_page_move_mapping().
512 int migrate_huge_page_move_mapping(struct address_space *mapping,
513 struct page *newpage, struct page *page)
515 XA_STATE(xas, &mapping->i_pages, page_index(page));
519 expected_count = 2 + page_has_private(page);
520 if (page_count(page) != expected_count || xas_load(&xas) != page) {
521 xas_unlock_irq(&xas);
525 if (!page_ref_freeze(page, expected_count)) {
526 xas_unlock_irq(&xas);
530 newpage->index = page->index;
531 newpage->mapping = page->mapping;
535 xas_store(&xas, newpage);
537 page_ref_unfreeze(page, expected_count - 1);
539 xas_unlock_irq(&xas);
541 return MIGRATEPAGE_SUCCESS;
545 * Gigantic pages are so large that we do not guarantee that page++ pointer
546 * arithmetic will work across the entire page. We need something more
549 static void __copy_gigantic_page(struct page *dst, struct page *src,
553 struct page *dst_base = dst;
554 struct page *src_base = src;
556 for (i = 0; i < nr_pages; ) {
558 copy_highpage(dst, src);
561 dst = mem_map_next(dst, dst_base, i);
562 src = mem_map_next(src, src_base, i);
566 static void copy_huge_page(struct page *dst, struct page *src)
573 struct hstate *h = page_hstate(src);
574 nr_pages = pages_per_huge_page(h);
576 if (unlikely(nr_pages > MAX_ORDER_NR_PAGES)) {
577 __copy_gigantic_page(dst, src, nr_pages);
582 BUG_ON(!PageTransHuge(src));
583 nr_pages = hpage_nr_pages(src);
586 for (i = 0; i < nr_pages; i++) {
588 copy_highpage(dst + i, src + i);
593 * Copy the page to its new location
595 void migrate_page_states(struct page *newpage, struct page *page)
600 SetPageError(newpage);
601 if (PageReferenced(page))
602 SetPageReferenced(newpage);
603 if (PageUptodate(page))
604 SetPageUptodate(newpage);
605 if (TestClearPageActive(page)) {
606 VM_BUG_ON_PAGE(PageUnevictable(page), page);
607 SetPageActive(newpage);
608 } else if (TestClearPageUnevictable(page))
609 SetPageUnevictable(newpage);
610 if (PageWorkingset(page))
611 SetPageWorkingset(newpage);
612 if (PageChecked(page))
613 SetPageChecked(newpage);
614 if (PageMappedToDisk(page))
615 SetPageMappedToDisk(newpage);
617 /* Move dirty on pages not done by migrate_page_move_mapping() */
619 SetPageDirty(newpage);
621 if (page_is_young(page))
622 set_page_young(newpage);
623 if (page_is_idle(page))
624 set_page_idle(newpage);
627 * Copy NUMA information to the new page, to prevent over-eager
628 * future migrations of this same page.
630 cpupid = page_cpupid_xchg_last(page, -1);
631 page_cpupid_xchg_last(newpage, cpupid);
633 ksm_migrate_page(newpage, page);
635 * Please do not reorder this without considering how mm/ksm.c's
636 * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache().
638 if (PageSwapCache(page))
639 ClearPageSwapCache(page);
640 ClearPagePrivate(page);
641 set_page_private(page, 0);
644 * If any waiters have accumulated on the new page then
647 if (PageWriteback(newpage))
648 end_page_writeback(newpage);
651 * PG_readahead shares the same bit with PG_reclaim. The above
652 * end_page_writeback() may clear PG_readahead mistakenly, so set the
655 if (PageReadahead(page))
656 SetPageReadahead(newpage);
658 copy_page_owner(page, newpage);
660 mem_cgroup_migrate(page, newpage);
662 EXPORT_SYMBOL(migrate_page_states);
664 void migrate_page_copy(struct page *newpage, struct page *page)
666 if (PageHuge(page) || PageTransHuge(page))
667 copy_huge_page(newpage, page);
669 copy_highpage(newpage, page);
671 migrate_page_states(newpage, page);
673 EXPORT_SYMBOL(migrate_page_copy);
675 /************************************************************
676 * Migration functions
677 ***********************************************************/
680 * Common logic to directly migrate a single LRU page suitable for
681 * pages that do not use PagePrivate/PagePrivate2.
683 * Pages are locked upon entry and exit.
685 int migrate_page(struct address_space *mapping,
686 struct page *newpage, struct page *page,
687 enum migrate_mode mode)
691 BUG_ON(PageWriteback(page)); /* Writeback must be complete */
693 rc = migrate_page_move_mapping(mapping, newpage, page, 0);
695 if (rc != MIGRATEPAGE_SUCCESS)
698 if (mode != MIGRATE_SYNC_NO_COPY)
699 migrate_page_copy(newpage, page);
701 migrate_page_states(newpage, page);
702 return MIGRATEPAGE_SUCCESS;
704 EXPORT_SYMBOL(migrate_page);
707 /* Returns true if all buffers are successfully locked */
708 static bool buffer_migrate_lock_buffers(struct buffer_head *head,
709 enum migrate_mode mode)
711 struct buffer_head *bh = head;
713 /* Simple case, sync compaction */
714 if (mode != MIGRATE_ASYNC) {
717 bh = bh->b_this_page;
719 } while (bh != head);
724 /* async case, we cannot block on lock_buffer so use trylock_buffer */
726 if (!trylock_buffer(bh)) {
728 * We failed to lock the buffer and cannot stall in
729 * async migration. Release the taken locks
731 struct buffer_head *failed_bh = bh;
733 while (bh != failed_bh) {
735 bh = bh->b_this_page;
740 bh = bh->b_this_page;
741 } while (bh != head);
745 static int __buffer_migrate_page(struct address_space *mapping,
746 struct page *newpage, struct page *page, enum migrate_mode mode,
749 struct buffer_head *bh, *head;
753 if (!page_has_buffers(page))
754 return migrate_page(mapping, newpage, page, mode);
756 /* Check whether page does not have extra refs before we do more work */
757 expected_count = expected_page_refs(mapping, page);
758 if (page_count(page) != expected_count)
761 head = page_buffers(page);
762 if (!buffer_migrate_lock_buffers(head, mode))
767 bool invalidated = false;
771 spin_lock(&mapping->private_lock);
774 if (atomic_read(&bh->b_count)) {
778 bh = bh->b_this_page;
779 } while (bh != head);
785 spin_unlock(&mapping->private_lock);
786 invalidate_bh_lrus();
788 goto recheck_buffers;
792 rc = migrate_page_move_mapping(mapping, newpage, page, 0);
793 if (rc != MIGRATEPAGE_SUCCESS)
796 ClearPagePrivate(page);
797 set_page_private(newpage, page_private(page));
798 set_page_private(page, 0);
804 set_bh_page(bh, newpage, bh_offset(bh));
805 bh = bh->b_this_page;
807 } while (bh != head);
809 SetPagePrivate(newpage);
811 if (mode != MIGRATE_SYNC_NO_COPY)
812 migrate_page_copy(newpage, page);
814 migrate_page_states(newpage, page);
816 rc = MIGRATEPAGE_SUCCESS;
819 spin_unlock(&mapping->private_lock);
823 bh = bh->b_this_page;
825 } while (bh != head);
831 * Migration function for pages with buffers. This function can only be used
832 * if the underlying filesystem guarantees that no other references to "page"
833 * exist. For example attached buffer heads are accessed only under page lock.
835 int buffer_migrate_page(struct address_space *mapping,
836 struct page *newpage, struct page *page, enum migrate_mode mode)
838 return __buffer_migrate_page(mapping, newpage, page, mode, false);
840 EXPORT_SYMBOL(buffer_migrate_page);
843 * Same as above except that this variant is more careful and checks that there
844 * are also no buffer head references. This function is the right one for
845 * mappings where buffer heads are directly looked up and referenced (such as
846 * block device mappings).
848 int buffer_migrate_page_norefs(struct address_space *mapping,
849 struct page *newpage, struct page *page, enum migrate_mode mode)
851 return __buffer_migrate_page(mapping, newpage, page, mode, true);
856 * Writeback a page to clean the dirty state
858 static int writeout(struct address_space *mapping, struct page *page)
860 struct writeback_control wbc = {
861 .sync_mode = WB_SYNC_NONE,
864 .range_end = LLONG_MAX,
869 if (!mapping->a_ops->writepage)
870 /* No write method for the address space */
873 if (!clear_page_dirty_for_io(page))
874 /* Someone else already triggered a write */
878 * A dirty page may imply that the underlying filesystem has
879 * the page on some queue. So the page must be clean for
880 * migration. Writeout may mean we loose the lock and the
881 * page state is no longer what we checked for earlier.
882 * At this point we know that the migration attempt cannot
885 remove_migration_ptes(page, page, false);
887 rc = mapping->a_ops->writepage(page, &wbc);
889 if (rc != AOP_WRITEPAGE_ACTIVATE)
890 /* unlocked. Relock */
893 return (rc < 0) ? -EIO : -EAGAIN;
897 * Default handling if a filesystem does not provide a migration function.
899 static int fallback_migrate_page(struct address_space *mapping,
900 struct page *newpage, struct page *page, enum migrate_mode mode)
902 if (PageDirty(page)) {
903 /* Only writeback pages in full synchronous migration */
906 case MIGRATE_SYNC_NO_COPY:
911 return writeout(mapping, page);
915 * Buffers may be managed in a filesystem specific way.
916 * We must have no buffers or drop them.
918 if (page_has_private(page) &&
919 !try_to_release_page(page, GFP_KERNEL))
920 return mode == MIGRATE_SYNC ? -EAGAIN : -EBUSY;
922 return migrate_page(mapping, newpage, page, mode);
926 * Move a page to a newly allocated page
927 * The page is locked and all ptes have been successfully removed.
929 * The new page will have replaced the old page if this function
934 * MIGRATEPAGE_SUCCESS - success
936 static int move_to_new_page(struct page *newpage, struct page *page,
937 enum migrate_mode mode)
939 struct address_space *mapping;
941 bool is_lru = !__PageMovable(page);
943 VM_BUG_ON_PAGE(!PageLocked(page), page);
944 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
946 mapping = page_mapping(page);
948 if (likely(is_lru)) {
950 rc = migrate_page(mapping, newpage, page, mode);
951 else if (mapping->a_ops->migratepage)
953 * Most pages have a mapping and most filesystems
954 * provide a migratepage callback. Anonymous pages
955 * are part of swap space which also has its own
956 * migratepage callback. This is the most common path
957 * for page migration.
959 rc = mapping->a_ops->migratepage(mapping, newpage,
962 rc = fallback_migrate_page(mapping, newpage,
966 * In case of non-lru page, it could be released after
967 * isolation step. In that case, we shouldn't try migration.
969 VM_BUG_ON_PAGE(!PageIsolated(page), page);
970 if (!PageMovable(page)) {
971 rc = MIGRATEPAGE_SUCCESS;
972 __ClearPageIsolated(page);
976 rc = mapping->a_ops->migratepage(mapping, newpage,
978 WARN_ON_ONCE(rc == MIGRATEPAGE_SUCCESS &&
979 !PageIsolated(page));
983 * When successful, old pagecache page->mapping must be cleared before
984 * page is freed; but stats require that PageAnon be left as PageAnon.
986 if (rc == MIGRATEPAGE_SUCCESS) {
987 if (__PageMovable(page)) {
988 VM_BUG_ON_PAGE(!PageIsolated(page), page);
991 * We clear PG_movable under page_lock so any compactor
992 * cannot try to migrate this page.
994 __ClearPageIsolated(page);
998 * Anonymous and movable page->mapping will be cleared by
999 * free_pages_prepare so don't reset it here for keeping
1000 * the type to work PageAnon, for example.
1002 if (!PageMappingFlags(page))
1003 page->mapping = NULL;
1005 if (likely(!is_zone_device_page(newpage)))
1006 flush_dcache_page(newpage);
1013 static int __unmap_and_move(struct page *page, struct page *newpage,
1014 int force, enum migrate_mode mode)
1017 int page_was_mapped = 0;
1018 struct anon_vma *anon_vma = NULL;
1019 bool is_lru = !__PageMovable(page);
1021 if (!trylock_page(page)) {
1022 if (!force || mode == MIGRATE_ASYNC)
1026 * It's not safe for direct compaction to call lock_page.
1027 * For example, during page readahead pages are added locked
1028 * to the LRU. Later, when the IO completes the pages are
1029 * marked uptodate and unlocked. However, the queueing
1030 * could be merging multiple pages for one bio (e.g.
1031 * mpage_readpages). If an allocation happens for the
1032 * second or third page, the process can end up locking
1033 * the same page twice and deadlocking. Rather than
1034 * trying to be clever about what pages can be locked,
1035 * avoid the use of lock_page for direct compaction
1038 if (current->flags & PF_MEMALLOC)
1044 if (PageWriteback(page)) {
1046 * Only in the case of a full synchronous migration is it
1047 * necessary to wait for PageWriteback. In the async case,
1048 * the retry loop is too short and in the sync-light case,
1049 * the overhead of stalling is too much
1053 case MIGRATE_SYNC_NO_COPY:
1061 wait_on_page_writeback(page);
1065 * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
1066 * we cannot notice that anon_vma is freed while we migrates a page.
1067 * This get_anon_vma() delays freeing anon_vma pointer until the end
1068 * of migration. File cache pages are no problem because of page_lock()
1069 * File Caches may use write_page() or lock_page() in migration, then,
1070 * just care Anon page here.
1072 * Only page_get_anon_vma() understands the subtleties of
1073 * getting a hold on an anon_vma from outside one of its mms.
1074 * But if we cannot get anon_vma, then we won't need it anyway,
1075 * because that implies that the anon page is no longer mapped
1076 * (and cannot be remapped so long as we hold the page lock).
1078 if (PageAnon(page) && !PageKsm(page))
1079 anon_vma = page_get_anon_vma(page);
1082 * Block others from accessing the new page when we get around to
1083 * establishing additional references. We are usually the only one
1084 * holding a reference to newpage at this point. We used to have a BUG
1085 * here if trylock_page(newpage) fails, but would like to allow for
1086 * cases where there might be a race with the previous use of newpage.
1087 * This is much like races on refcount of oldpage: just don't BUG().
1089 if (unlikely(!trylock_page(newpage)))
1092 if (unlikely(!is_lru)) {
1093 rc = move_to_new_page(newpage, page, mode);
1094 goto out_unlock_both;
1098 * Corner case handling:
1099 * 1. When a new swap-cache page is read into, it is added to the LRU
1100 * and treated as swapcache but it has no rmap yet.
1101 * Calling try_to_unmap() against a page->mapping==NULL page will
1102 * trigger a BUG. So handle it here.
1103 * 2. An orphaned page (see truncate_complete_page) might have
1104 * fs-private metadata. The page can be picked up due to memory
1105 * offlining. Everywhere else except page reclaim, the page is
1106 * invisible to the vm, so the page can not be migrated. So try to
1107 * free the metadata, so the page can be freed.
1109 if (!page->mapping) {
1110 VM_BUG_ON_PAGE(PageAnon(page), page);
1111 if (page_has_private(page)) {
1112 try_to_free_buffers(page);
1113 goto out_unlock_both;
1115 } else if (page_mapped(page)) {
1116 /* Establish migration ptes */
1117 VM_BUG_ON_PAGE(PageAnon(page) && !PageKsm(page) && !anon_vma,
1120 TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
1121 page_was_mapped = 1;
1124 if (!page_mapped(page))
1125 rc = move_to_new_page(newpage, page, mode);
1127 if (page_was_mapped)
1128 remove_migration_ptes(page,
1129 rc == MIGRATEPAGE_SUCCESS ? newpage : page, false);
1132 unlock_page(newpage);
1134 /* Drop an anon_vma reference if we took one */
1136 put_anon_vma(anon_vma);
1140 * If migration is successful, decrease refcount of the newpage
1141 * which will not free the page because new page owner increased
1142 * refcounter. As well, if it is LRU page, add the page to LRU
1143 * list in here. Use the old state of the isolated source page to
1144 * determine if we migrated a LRU page. newpage was already unlocked
1145 * and possibly modified by its owner - don't rely on the page
1148 if (rc == MIGRATEPAGE_SUCCESS) {
1149 if (unlikely(!is_lru))
1152 putback_lru_page(newpage);
1159 * gcc 4.7 and 4.8 on arm get an ICEs when inlining unmap_and_move(). Work
1162 #if defined(CONFIG_ARM) && \
1163 defined(GCC_VERSION) && GCC_VERSION < 40900 && GCC_VERSION >= 40700
1164 #define ICE_noinline noinline
1166 #define ICE_noinline
1170 * Obtain the lock on page, remove all ptes and migrate the page
1171 * to the newly allocated page in newpage.
1173 static ICE_noinline int unmap_and_move(new_page_t get_new_page,
1174 free_page_t put_new_page,
1175 unsigned long private, struct page *page,
1176 int force, enum migrate_mode mode,
1177 enum migrate_reason reason)
1179 int rc = MIGRATEPAGE_SUCCESS;
1180 struct page *newpage = NULL;
1182 if (!thp_migration_supported() && PageTransHuge(page))
1185 if (page_count(page) == 1) {
1186 /* page was freed from under us. So we are done. */
1187 ClearPageActive(page);
1188 ClearPageUnevictable(page);
1189 if (unlikely(__PageMovable(page))) {
1191 if (!PageMovable(page))
1192 __ClearPageIsolated(page);
1198 newpage = get_new_page(page, private);
1202 rc = __unmap_and_move(page, newpage, force, mode);
1203 if (rc == MIGRATEPAGE_SUCCESS)
1204 set_page_owner_migrate_reason(newpage, reason);
1207 if (rc != -EAGAIN) {
1209 * A page that has been migrated has all references
1210 * removed and will be freed. A page that has not been
1211 * migrated will have kept its references and be restored.
1213 list_del(&page->lru);
1216 * Compaction can migrate also non-LRU pages which are
1217 * not accounted to NR_ISOLATED_*. They can be recognized
1220 if (likely(!__PageMovable(page)))
1221 mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON +
1222 page_is_file_cache(page), -hpage_nr_pages(page));
1226 * If migration is successful, releases reference grabbed during
1227 * isolation. Otherwise, restore the page to right list unless
1230 if (rc == MIGRATEPAGE_SUCCESS) {
1232 if (reason == MR_MEMORY_FAILURE) {
1234 * Set PG_HWPoison on just freed page
1235 * intentionally. Although it's rather weird,
1236 * it's how HWPoison flag works at the moment.
1238 if (set_hwpoison_free_buddy_page(page))
1239 num_poisoned_pages_inc();
1242 if (rc != -EAGAIN) {
1243 if (likely(!__PageMovable(page))) {
1244 putback_lru_page(page);
1249 if (PageMovable(page))
1250 putback_movable_page(page);
1252 __ClearPageIsolated(page);
1258 put_new_page(newpage, private);
1267 * Counterpart of unmap_and_move_page() for hugepage migration.
1269 * This function doesn't wait the completion of hugepage I/O
1270 * because there is no race between I/O and migration for hugepage.
1271 * Note that currently hugepage I/O occurs only in direct I/O
1272 * where no lock is held and PG_writeback is irrelevant,
1273 * and writeback status of all subpages are counted in the reference
1274 * count of the head page (i.e. if all subpages of a 2MB hugepage are
1275 * under direct I/O, the reference of the head page is 512 and a bit more.)
1276 * This means that when we try to migrate hugepage whose subpages are
1277 * doing direct I/O, some references remain after try_to_unmap() and
1278 * hugepage migration fails without data corruption.
1280 * There is also no race when direct I/O is issued on the page under migration,
1281 * because then pte is replaced with migration swap entry and direct I/O code
1282 * will wait in the page fault for migration to complete.
1284 static int unmap_and_move_huge_page(new_page_t get_new_page,
1285 free_page_t put_new_page, unsigned long private,
1286 struct page *hpage, int force,
1287 enum migrate_mode mode, int reason)
1290 int page_was_mapped = 0;
1291 struct page *new_hpage;
1292 struct anon_vma *anon_vma = NULL;
1293 struct address_space *mapping = NULL;
1296 * Migratability of hugepages depends on architectures and their size.
1297 * This check is necessary because some callers of hugepage migration
1298 * like soft offline and memory hotremove don't walk through page
1299 * tables or check whether the hugepage is pmd-based or not before
1300 * kicking migration.
1302 if (!hugepage_migration_supported(page_hstate(hpage))) {
1303 putback_active_hugepage(hpage);
1307 new_hpage = get_new_page(hpage, private);
1311 if (!trylock_page(hpage)) {
1316 case MIGRATE_SYNC_NO_COPY:
1325 * Check for pages which are in the process of being freed. Without
1326 * page_mapping() set, hugetlbfs specific move page routine will not
1327 * be called and we could leak usage counts for subpools.
1329 if (page_private(hpage) && !page_mapping(hpage)) {
1334 if (PageAnon(hpage))
1335 anon_vma = page_get_anon_vma(hpage);
1337 if (unlikely(!trylock_page(new_hpage)))
1340 if (page_mapped(hpage)) {
1342 * try_to_unmap could potentially call huge_pmd_unshare.
1343 * Because of this, take semaphore in write mode here and
1344 * set TTU_RMAP_LOCKED to let lower levels know we have
1347 mapping = hugetlb_page_mapping_lock_write(hpage);
1348 if (unlikely(!mapping))
1349 goto unlock_put_anon;
1352 TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS|
1354 page_was_mapped = 1;
1356 * Leave mapping locked until after subsequent call to
1357 * remove_migration_ptes()
1361 if (!page_mapped(hpage))
1362 rc = move_to_new_page(new_hpage, hpage, mode);
1364 if (page_was_mapped) {
1365 remove_migration_ptes(hpage,
1366 rc == MIGRATEPAGE_SUCCESS ? new_hpage : hpage, true);
1367 i_mmap_unlock_write(mapping);
1371 unlock_page(new_hpage);
1375 put_anon_vma(anon_vma);
1377 if (rc == MIGRATEPAGE_SUCCESS) {
1378 move_hugetlb_state(hpage, new_hpage, reason);
1379 put_new_page = NULL;
1386 putback_active_hugepage(hpage);
1389 * If migration was not successful and there's a freeing callback, use
1390 * it. Otherwise, put_page() will drop the reference grabbed during
1394 put_new_page(new_hpage, private);
1396 putback_active_hugepage(new_hpage);
1402 * migrate_pages - migrate the pages specified in a list, to the free pages
1403 * supplied as the target for the page migration
1405 * @from: The list of pages to be migrated.
1406 * @get_new_page: The function used to allocate free pages to be used
1407 * as the target of the page migration.
1408 * @put_new_page: The function used to free target pages if migration
1409 * fails, or NULL if no special handling is necessary.
1410 * @private: Private data to be passed on to get_new_page()
1411 * @mode: The migration mode that specifies the constraints for
1412 * page migration, if any.
1413 * @reason: The reason for page migration.
1415 * The function returns after 10 attempts or if no pages are movable any more
1416 * because the list has become empty or no retryable pages exist any more.
1417 * The caller should call putback_movable_pages() to return pages to the LRU
1418 * or free list only if ret != 0.
1420 * Returns the number of pages that were not migrated, or an error code.
1422 int migrate_pages(struct list_head *from, new_page_t get_new_page,
1423 free_page_t put_new_page, unsigned long private,
1424 enum migrate_mode mode, int reason)
1428 int nr_succeeded = 0;
1432 int swapwrite = current->flags & PF_SWAPWRITE;
1436 current->flags |= PF_SWAPWRITE;
1438 for(pass = 0; pass < 10 && retry; pass++) {
1441 list_for_each_entry_safe(page, page2, from, lru) {
1446 rc = unmap_and_move_huge_page(get_new_page,
1447 put_new_page, private, page,
1448 pass > 2, mode, reason);
1450 rc = unmap_and_move(get_new_page, put_new_page,
1451 private, page, pass > 2, mode,
1457 * THP migration might be unsupported or the
1458 * allocation could've failed so we should
1459 * retry on the same page with the THP split
1462 * Head page is retried immediately and tail
1463 * pages are added to the tail of the list so
1464 * we encounter them after the rest of the list
1467 if (PageTransHuge(page) && !PageHuge(page)) {
1469 rc = split_huge_page_to_list(page, from);
1472 list_safe_reset_next(page, page2, lru);
1481 case MIGRATEPAGE_SUCCESS:
1486 * Permanent failure (-EBUSY, -ENOSYS, etc.):
1487 * unlike -EAGAIN case, the failed page is
1488 * removed from migration page list and not
1489 * retried in the next outer loop.
1500 count_vm_events(PGMIGRATE_SUCCESS, nr_succeeded);
1502 count_vm_events(PGMIGRATE_FAIL, nr_failed);
1503 trace_mm_migrate_pages(nr_succeeded, nr_failed, mode, reason);
1506 current->flags &= ~PF_SWAPWRITE;
1513 static int store_status(int __user *status, int start, int value, int nr)
1516 if (put_user(value, status + start))
1524 static int do_move_pages_to_node(struct mm_struct *mm,
1525 struct list_head *pagelist, int node)
1529 err = migrate_pages(pagelist, alloc_new_node_page, NULL, node,
1530 MIGRATE_SYNC, MR_SYSCALL);
1532 putback_movable_pages(pagelist);
1537 * Resolves the given address to a struct page, isolates it from the LRU and
1538 * puts it to the given pagelist.
1540 * errno - if the page cannot be found/isolated
1541 * 0 - when it doesn't have to be migrated because it is already on the
1543 * 1 - when it has been queued
1545 static int add_page_for_migration(struct mm_struct *mm, unsigned long addr,
1546 int node, struct list_head *pagelist, bool migrate_all)
1548 struct vm_area_struct *vma;
1550 unsigned int follflags;
1553 down_read(&mm->mmap_sem);
1555 vma = find_vma(mm, addr);
1556 if (!vma || addr < vma->vm_start || !vma_migratable(vma))
1559 /* FOLL_DUMP to ignore special (like zero) pages */
1560 follflags = FOLL_GET | FOLL_DUMP;
1561 page = follow_page(vma, addr, follflags);
1563 err = PTR_ERR(page);
1572 if (page_to_nid(page) == node)
1576 if (page_mapcount(page) > 1 && !migrate_all)
1579 if (PageHuge(page)) {
1580 if (PageHead(page)) {
1581 isolate_huge_page(page, pagelist);
1587 head = compound_head(page);
1588 err = isolate_lru_page(head);
1593 list_add_tail(&head->lru, pagelist);
1594 mod_node_page_state(page_pgdat(head),
1595 NR_ISOLATED_ANON + page_is_file_cache(head),
1596 hpage_nr_pages(head));
1600 * Either remove the duplicate refcount from
1601 * isolate_lru_page() or drop the page ref if it was
1606 up_read(&mm->mmap_sem);
1610 static int move_pages_and_store_status(struct mm_struct *mm, int node,
1611 struct list_head *pagelist, int __user *status,
1612 int start, int i, unsigned long nr_pages)
1616 if (list_empty(pagelist))
1619 err = do_move_pages_to_node(mm, pagelist, node);
1622 * Positive err means the number of failed
1623 * pages to migrate. Since we are going to
1624 * abort and return the number of non-migrated
1625 * pages, so need to incude the rest of the
1626 * nr_pages that have not been attempted as
1630 err += nr_pages - i - 1;
1633 return store_status(status, start, node, i - start);
1637 * Migrate an array of page address onto an array of nodes and fill
1638 * the corresponding array of status.
1640 static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes,
1641 unsigned long nr_pages,
1642 const void __user * __user *pages,
1643 const int __user *nodes,
1644 int __user *status, int flags)
1646 int current_node = NUMA_NO_NODE;
1647 LIST_HEAD(pagelist);
1653 for (i = start = 0; i < nr_pages; i++) {
1654 const void __user *p;
1659 if (get_user(p, pages + i))
1661 if (get_user(node, nodes + i))
1663 addr = (unsigned long)untagged_addr(p);
1666 if (node < 0 || node >= MAX_NUMNODES)
1668 if (!node_state(node, N_MEMORY))
1672 if (!node_isset(node, task_nodes))
1675 if (current_node == NUMA_NO_NODE) {
1676 current_node = node;
1678 } else if (node != current_node) {
1679 err = move_pages_and_store_status(mm, current_node,
1680 &pagelist, status, start, i, nr_pages);
1684 current_node = node;
1688 * Errors in the page lookup or isolation are not fatal and we simply
1689 * report them via status
1691 err = add_page_for_migration(mm, addr, current_node,
1692 &pagelist, flags & MPOL_MF_MOVE_ALL);
1695 /* The page is successfully queued for migration */
1700 * If the page is already on the target node (!err), store the
1701 * node, otherwise, store the err.
1703 err = store_status(status, i, err ? : current_node, 1);
1707 err = move_pages_and_store_status(mm, current_node, &pagelist,
1708 status, start, i, nr_pages);
1711 current_node = NUMA_NO_NODE;
1714 /* Make sure we do not overwrite the existing error */
1715 err1 = move_pages_and_store_status(mm, current_node, &pagelist,
1716 status, start, i, nr_pages);
1724 * Determine the nodes of an array of pages and store it in an array of status.
1726 static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages,
1727 const void __user **pages, int *status)
1731 down_read(&mm->mmap_sem);
1733 for (i = 0; i < nr_pages; i++) {
1734 unsigned long addr = (unsigned long)(*pages);
1735 struct vm_area_struct *vma;
1739 vma = find_vma(mm, addr);
1740 if (!vma || addr < vma->vm_start)
1743 /* FOLL_DUMP to ignore special (like zero) pages */
1744 page = follow_page(vma, addr, FOLL_DUMP);
1746 err = PTR_ERR(page);
1750 err = page ? page_to_nid(page) : -ENOENT;
1758 up_read(&mm->mmap_sem);
1762 * Determine the nodes of a user array of pages and store it in
1763 * a user array of status.
1765 static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages,
1766 const void __user * __user *pages,
1769 #define DO_PAGES_STAT_CHUNK_NR 16
1770 const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR];
1771 int chunk_status[DO_PAGES_STAT_CHUNK_NR];
1774 unsigned long chunk_nr;
1776 chunk_nr = nr_pages;
1777 if (chunk_nr > DO_PAGES_STAT_CHUNK_NR)
1778 chunk_nr = DO_PAGES_STAT_CHUNK_NR;
1780 if (copy_from_user(chunk_pages, pages, chunk_nr * sizeof(*chunk_pages)))
1783 do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status);
1785 if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status)))
1790 nr_pages -= chunk_nr;
1792 return nr_pages ? -EFAULT : 0;
1796 * Move a list of pages in the address space of the currently executing
1799 static int kernel_move_pages(pid_t pid, unsigned long nr_pages,
1800 const void __user * __user *pages,
1801 const int __user *nodes,
1802 int __user *status, int flags)
1804 struct task_struct *task;
1805 struct mm_struct *mm;
1807 nodemask_t task_nodes;
1810 if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
1813 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1816 /* Find the mm_struct */
1818 task = pid ? find_task_by_vpid(pid) : current;
1823 get_task_struct(task);
1826 * Check if this process has the right to modify the specified
1827 * process. Use the regular "ptrace_may_access()" checks.
1829 if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) {
1836 err = security_task_movememory(task);
1840 task_nodes = cpuset_mems_allowed(task);
1841 mm = get_task_mm(task);
1842 put_task_struct(task);
1848 err = do_pages_move(mm, task_nodes, nr_pages, pages,
1849 nodes, status, flags);
1851 err = do_pages_stat(mm, nr_pages, pages, status);
1857 put_task_struct(task);
1861 SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages,
1862 const void __user * __user *, pages,
1863 const int __user *, nodes,
1864 int __user *, status, int, flags)
1866 return kernel_move_pages(pid, nr_pages, pages, nodes, status, flags);
1869 #ifdef CONFIG_COMPAT
1870 COMPAT_SYSCALL_DEFINE6(move_pages, pid_t, pid, compat_ulong_t, nr_pages,
1871 compat_uptr_t __user *, pages32,
1872 const int __user *, nodes,
1873 int __user *, status,
1876 const void __user * __user *pages;
1879 pages = compat_alloc_user_space(nr_pages * sizeof(void *));
1880 for (i = 0; i < nr_pages; i++) {
1883 if (get_user(p, pages32 + i) ||
1884 put_user(compat_ptr(p), pages + i))
1887 return kernel_move_pages(pid, nr_pages, pages, nodes, status, flags);
1889 #endif /* CONFIG_COMPAT */
1891 #ifdef CONFIG_NUMA_BALANCING
1893 * Returns true if this is a safe migration target node for misplaced NUMA
1894 * pages. Currently it only checks the watermarks which crude
1896 static bool migrate_balanced_pgdat(struct pglist_data *pgdat,
1897 unsigned long nr_migrate_pages)
1901 for (z = pgdat->nr_zones - 1; z >= 0; z--) {
1902 struct zone *zone = pgdat->node_zones + z;
1904 if (!populated_zone(zone))
1907 /* Avoid waking kswapd by allocating pages_to_migrate pages. */
1908 if (!zone_watermark_ok(zone, 0,
1909 high_wmark_pages(zone) +
1918 static struct page *alloc_misplaced_dst_page(struct page *page,
1921 int nid = (int) data;
1922 struct page *newpage;
1924 newpage = __alloc_pages_node(nid,
1925 (GFP_HIGHUSER_MOVABLE |
1926 __GFP_THISNODE | __GFP_NOMEMALLOC |
1927 __GFP_NORETRY | __GFP_NOWARN) &
1933 static int numamigrate_isolate_page(pg_data_t *pgdat, struct page *page)
1937 VM_BUG_ON_PAGE(compound_order(page) && !PageTransHuge(page), page);
1939 /* Avoid migrating to a node that is nearly full */
1940 if (!migrate_balanced_pgdat(pgdat, compound_nr(page)))
1943 if (isolate_lru_page(page))
1947 * migrate_misplaced_transhuge_page() skips page migration's usual
1948 * check on page_count(), so we must do it here, now that the page
1949 * has been isolated: a GUP pin, or any other pin, prevents migration.
1950 * The expected page count is 3: 1 for page's mapcount and 1 for the
1951 * caller's pin and 1 for the reference taken by isolate_lru_page().
1953 if (PageTransHuge(page) && page_count(page) != 3) {
1954 putback_lru_page(page);
1958 page_lru = page_is_file_cache(page);
1959 mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON + page_lru,
1960 hpage_nr_pages(page));
1963 * Isolating the page has taken another reference, so the
1964 * caller's reference can be safely dropped without the page
1965 * disappearing underneath us during migration.
1971 bool pmd_trans_migrating(pmd_t pmd)
1973 struct page *page = pmd_page(pmd);
1974 return PageLocked(page);
1978 * Attempt to migrate a misplaced page to the specified destination
1979 * node. Caller is expected to have an elevated reference count on
1980 * the page that will be dropped by this function before returning.
1982 int migrate_misplaced_page(struct page *page, struct vm_area_struct *vma,
1985 pg_data_t *pgdat = NODE_DATA(node);
1988 LIST_HEAD(migratepages);
1991 * Don't migrate file pages that are mapped in multiple processes
1992 * with execute permissions as they are probably shared libraries.
1994 if (page_mapcount(page) != 1 && page_is_file_cache(page) &&
1995 (vma->vm_flags & VM_EXEC))
1999 * Also do not migrate dirty pages as not all filesystems can move
2000 * dirty pages in MIGRATE_ASYNC mode which is a waste of cycles.
2002 if (page_is_file_cache(page) && PageDirty(page))
2005 isolated = numamigrate_isolate_page(pgdat, page);
2009 list_add(&page->lru, &migratepages);
2010 nr_remaining = migrate_pages(&migratepages, alloc_misplaced_dst_page,
2011 NULL, node, MIGRATE_ASYNC,
2014 if (!list_empty(&migratepages)) {
2015 list_del(&page->lru);
2016 dec_node_page_state(page, NR_ISOLATED_ANON +
2017 page_is_file_cache(page));
2018 putback_lru_page(page);
2022 count_vm_numa_event(NUMA_PAGE_MIGRATE);
2023 BUG_ON(!list_empty(&migratepages));
2030 #endif /* CONFIG_NUMA_BALANCING */
2032 #if defined(CONFIG_NUMA_BALANCING) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
2034 * Migrates a THP to a given target node. page must be locked and is unlocked
2037 int migrate_misplaced_transhuge_page(struct mm_struct *mm,
2038 struct vm_area_struct *vma,
2039 pmd_t *pmd, pmd_t entry,
2040 unsigned long address,
2041 struct page *page, int node)
2044 pg_data_t *pgdat = NODE_DATA(node);
2046 struct page *new_page = NULL;
2047 int page_lru = page_is_file_cache(page);
2048 unsigned long start = address & HPAGE_PMD_MASK;
2050 new_page = alloc_pages_node(node,
2051 (GFP_TRANSHUGE_LIGHT | __GFP_THISNODE),
2055 prep_transhuge_page(new_page);
2057 isolated = numamigrate_isolate_page(pgdat, page);
2063 /* Prepare a page as a migration target */
2064 __SetPageLocked(new_page);
2065 if (PageSwapBacked(page))
2066 __SetPageSwapBacked(new_page);
2068 /* anon mapping, we can simply copy page->mapping to the new page: */
2069 new_page->mapping = page->mapping;
2070 new_page->index = page->index;
2071 /* flush the cache before copying using the kernel virtual address */
2072 flush_cache_range(vma, start, start + HPAGE_PMD_SIZE);
2073 migrate_page_copy(new_page, page);
2074 WARN_ON(PageLRU(new_page));
2076 /* Recheck the target PMD */
2077 ptl = pmd_lock(mm, pmd);
2078 if (unlikely(!pmd_same(*pmd, entry) || !page_ref_freeze(page, 2))) {
2081 /* Reverse changes made by migrate_page_copy() */
2082 if (TestClearPageActive(new_page))
2083 SetPageActive(page);
2084 if (TestClearPageUnevictable(new_page))
2085 SetPageUnevictable(page);
2087 unlock_page(new_page);
2088 put_page(new_page); /* Free it */
2090 /* Retake the callers reference and putback on LRU */
2092 putback_lru_page(page);
2093 mod_node_page_state(page_pgdat(page),
2094 NR_ISOLATED_ANON + page_lru, -HPAGE_PMD_NR);
2099 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
2100 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
2103 * Overwrite the old entry under pagetable lock and establish
2104 * the new PTE. Any parallel GUP will either observe the old
2105 * page blocking on the page lock, block on the page table
2106 * lock or observe the new page. The SetPageUptodate on the
2107 * new page and page_add_new_anon_rmap guarantee the copy is
2108 * visible before the pagetable update.
2110 page_add_anon_rmap(new_page, vma, start, true);
2112 * At this point the pmd is numa/protnone (i.e. non present) and the TLB
2113 * has already been flushed globally. So no TLB can be currently
2114 * caching this non present pmd mapping. There's no need to clear the
2115 * pmd before doing set_pmd_at(), nor to flush the TLB after
2116 * set_pmd_at(). Clearing the pmd here would introduce a race
2117 * condition against MADV_DONTNEED, because MADV_DONTNEED only holds the
2118 * mmap_sem for reading. If the pmd is set to NULL at any given time,
2119 * MADV_DONTNEED won't wait on the pmd lock and it'll skip clearing this
2122 set_pmd_at(mm, start, pmd, entry);
2123 update_mmu_cache_pmd(vma, address, &entry);
2125 page_ref_unfreeze(page, 2);
2126 mlock_migrate_page(new_page, page);
2127 page_remove_rmap(page, true);
2128 set_page_owner_migrate_reason(new_page, MR_NUMA_MISPLACED);
2132 /* Take an "isolate" reference and put new page on the LRU. */
2134 putback_lru_page(new_page);
2136 unlock_page(new_page);
2138 put_page(page); /* Drop the rmap reference */
2139 put_page(page); /* Drop the LRU isolation reference */
2141 count_vm_events(PGMIGRATE_SUCCESS, HPAGE_PMD_NR);
2142 count_vm_numa_events(NUMA_PAGE_MIGRATE, HPAGE_PMD_NR);
2144 mod_node_page_state(page_pgdat(page),
2145 NR_ISOLATED_ANON + page_lru,
2150 count_vm_events(PGMIGRATE_FAIL, HPAGE_PMD_NR);
2151 ptl = pmd_lock(mm, pmd);
2152 if (pmd_same(*pmd, entry)) {
2153 entry = pmd_modify(entry, vma->vm_page_prot);
2154 set_pmd_at(mm, start, pmd, entry);
2155 update_mmu_cache_pmd(vma, address, &entry);
2164 #endif /* CONFIG_NUMA_BALANCING */
2166 #endif /* CONFIG_NUMA */
2168 #ifdef CONFIG_DEVICE_PRIVATE
2169 static int migrate_vma_collect_hole(unsigned long start,
2171 __always_unused int depth,
2172 struct mm_walk *walk)
2174 struct migrate_vma *migrate = walk->private;
2177 for (addr = start; addr < end; addr += PAGE_SIZE) {
2178 migrate->src[migrate->npages] = MIGRATE_PFN_MIGRATE;
2179 migrate->dst[migrate->npages] = 0;
2187 static int migrate_vma_collect_skip(unsigned long start,
2189 struct mm_walk *walk)
2191 struct migrate_vma *migrate = walk->private;
2194 for (addr = start; addr < end; addr += PAGE_SIZE) {
2195 migrate->dst[migrate->npages] = 0;
2196 migrate->src[migrate->npages++] = 0;
2202 static int migrate_vma_collect_pmd(pmd_t *pmdp,
2203 unsigned long start,
2205 struct mm_walk *walk)
2207 struct migrate_vma *migrate = walk->private;
2208 struct vm_area_struct *vma = walk->vma;
2209 struct mm_struct *mm = vma->vm_mm;
2210 unsigned long addr = start, unmapped = 0;
2215 if (pmd_none(*pmdp))
2216 return migrate_vma_collect_hole(start, end, -1, walk);
2218 if (pmd_trans_huge(*pmdp)) {
2221 ptl = pmd_lock(mm, pmdp);
2222 if (unlikely(!pmd_trans_huge(*pmdp))) {
2227 page = pmd_page(*pmdp);
2228 if (is_huge_zero_page(page)) {
2230 split_huge_pmd(vma, pmdp, addr);
2231 if (pmd_trans_unstable(pmdp))
2232 return migrate_vma_collect_skip(start, end,
2239 if (unlikely(!trylock_page(page)))
2240 return migrate_vma_collect_skip(start, end,
2242 ret = split_huge_page(page);
2246 return migrate_vma_collect_skip(start, end,
2248 if (pmd_none(*pmdp))
2249 return migrate_vma_collect_hole(start, end, -1,
2254 if (unlikely(pmd_bad(*pmdp)))
2255 return migrate_vma_collect_skip(start, end, walk);
2257 ptep = pte_offset_map_lock(mm, pmdp, addr, &ptl);
2258 arch_enter_lazy_mmu_mode();
2260 for (; addr < end; addr += PAGE_SIZE, ptep++) {
2261 unsigned long mpfn = 0, pfn;
2268 if (pte_none(pte)) {
2269 mpfn = MIGRATE_PFN_MIGRATE;
2274 if (!pte_present(pte)) {
2276 * Only care about unaddressable device page special
2277 * page table entry. Other special swap entries are not
2278 * migratable, and we ignore regular swapped page.
2280 entry = pte_to_swp_entry(pte);
2281 if (!is_device_private_entry(entry))
2284 page = device_private_entry_to_page(entry);
2285 if (page->pgmap->owner != migrate->src_owner)
2288 mpfn = migrate_pfn(page_to_pfn(page)) |
2289 MIGRATE_PFN_MIGRATE;
2290 if (is_write_device_private_entry(entry))
2291 mpfn |= MIGRATE_PFN_WRITE;
2293 if (migrate->src_owner)
2296 if (is_zero_pfn(pfn)) {
2297 mpfn = MIGRATE_PFN_MIGRATE;
2301 page = vm_normal_page(migrate->vma, addr, pte);
2302 mpfn = migrate_pfn(pfn) | MIGRATE_PFN_MIGRATE;
2303 mpfn |= pte_write(pte) ? MIGRATE_PFN_WRITE : 0;
2306 /* FIXME support THP */
2307 if (!page || !page->mapping || PageTransCompound(page)) {
2313 * By getting a reference on the page we pin it and that blocks
2314 * any kind of migration. Side effect is that it "freezes" the
2317 * We drop this reference after isolating the page from the lru
2318 * for non device page (device page are not on the lru and thus
2319 * can't be dropped from it).
2325 * Optimize for the common case where page is only mapped once
2326 * in one process. If we can lock the page, then we can safely
2327 * set up a special migration page table entry now.
2329 if (trylock_page(page)) {
2332 mpfn |= MIGRATE_PFN_LOCKED;
2333 ptep_get_and_clear(mm, addr, ptep);
2335 /* Setup special migration page table entry */
2336 entry = make_migration_entry(page, mpfn &
2338 swp_pte = swp_entry_to_pte(entry);
2339 if (pte_soft_dirty(pte))
2340 swp_pte = pte_swp_mksoft_dirty(swp_pte);
2341 set_pte_at(mm, addr, ptep, swp_pte);
2344 * This is like regular unmap: we remove the rmap and
2345 * drop page refcount. Page won't be freed, as we took
2346 * a reference just above.
2348 page_remove_rmap(page, false);
2351 if (pte_present(pte))
2356 migrate->dst[migrate->npages] = 0;
2357 migrate->src[migrate->npages++] = mpfn;
2359 arch_leave_lazy_mmu_mode();
2360 pte_unmap_unlock(ptep - 1, ptl);
2362 /* Only flush the TLB if we actually modified any entries */
2364 flush_tlb_range(walk->vma, start, end);
2369 static const struct mm_walk_ops migrate_vma_walk_ops = {
2370 .pmd_entry = migrate_vma_collect_pmd,
2371 .pte_hole = migrate_vma_collect_hole,
2375 * migrate_vma_collect() - collect pages over a range of virtual addresses
2376 * @migrate: migrate struct containing all migration information
2378 * This will walk the CPU page table. For each virtual address backed by a
2379 * valid page, it updates the src array and takes a reference on the page, in
2380 * order to pin the page until we lock it and unmap it.
2382 static void migrate_vma_collect(struct migrate_vma *migrate)
2384 struct mmu_notifier_range range;
2386 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, NULL,
2387 migrate->vma->vm_mm, migrate->start, migrate->end);
2388 mmu_notifier_invalidate_range_start(&range);
2390 walk_page_range(migrate->vma->vm_mm, migrate->start, migrate->end,
2391 &migrate_vma_walk_ops, migrate);
2393 mmu_notifier_invalidate_range_end(&range);
2394 migrate->end = migrate->start + (migrate->npages << PAGE_SHIFT);
2398 * migrate_vma_check_page() - check if page is pinned or not
2399 * @page: struct page to check
2401 * Pinned pages cannot be migrated. This is the same test as in
2402 * migrate_page_move_mapping(), except that here we allow migration of a
2405 static bool migrate_vma_check_page(struct page *page)
2408 * One extra ref because caller holds an extra reference, either from
2409 * isolate_lru_page() for a regular page, or migrate_vma_collect() for
2415 * FIXME support THP (transparent huge page), it is bit more complex to
2416 * check them than regular pages, because they can be mapped with a pmd
2417 * or with a pte (split pte mapping).
2419 if (PageCompound(page))
2422 /* Page from ZONE_DEVICE have one extra reference */
2423 if (is_zone_device_page(page)) {
2425 * Private page can never be pin as they have no valid pte and
2426 * GUP will fail for those. Yet if there is a pending migration
2427 * a thread might try to wait on the pte migration entry and
2428 * will bump the page reference count. Sadly there is no way to
2429 * differentiate a regular pin from migration wait. Hence to
2430 * avoid 2 racing thread trying to migrate back to CPU to enter
2431 * infinite loop (one stoping migration because the other is
2432 * waiting on pte migration entry). We always return true here.
2434 * FIXME proper solution is to rework migration_entry_wait() so
2435 * it does not need to take a reference on page.
2437 return is_device_private_page(page);
2440 /* For file back page */
2441 if (page_mapping(page))
2442 extra += 1 + page_has_private(page);
2444 if ((page_count(page) - extra) > page_mapcount(page))
2451 * migrate_vma_prepare() - lock pages and isolate them from the lru
2452 * @migrate: migrate struct containing all migration information
2454 * This locks pages that have been collected by migrate_vma_collect(). Once each
2455 * page is locked it is isolated from the lru (for non-device pages). Finally,
2456 * the ref taken by migrate_vma_collect() is dropped, as locked pages cannot be
2457 * migrated by concurrent kernel threads.
2459 static void migrate_vma_prepare(struct migrate_vma *migrate)
2461 const unsigned long npages = migrate->npages;
2462 const unsigned long start = migrate->start;
2463 unsigned long addr, i, restore = 0;
2464 bool allow_drain = true;
2468 for (i = 0; (i < npages) && migrate->cpages; i++) {
2469 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2475 if (!(migrate->src[i] & MIGRATE_PFN_LOCKED)) {
2477 * Because we are migrating several pages there can be
2478 * a deadlock between 2 concurrent migration where each
2479 * are waiting on each other page lock.
2481 * Make migrate_vma() a best effort thing and backoff
2482 * for any page we can not lock right away.
2484 if (!trylock_page(page)) {
2485 migrate->src[i] = 0;
2491 migrate->src[i] |= MIGRATE_PFN_LOCKED;
2494 /* ZONE_DEVICE pages are not on LRU */
2495 if (!is_zone_device_page(page)) {
2496 if (!PageLRU(page) && allow_drain) {
2497 /* Drain CPU's pagevec */
2498 lru_add_drain_all();
2499 allow_drain = false;
2502 if (isolate_lru_page(page)) {
2504 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2508 migrate->src[i] = 0;
2516 /* Drop the reference we took in collect */
2520 if (!migrate_vma_check_page(page)) {
2522 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2526 if (!is_zone_device_page(page)) {
2528 putback_lru_page(page);
2531 migrate->src[i] = 0;
2535 if (!is_zone_device_page(page))
2536 putback_lru_page(page);
2543 for (i = 0, addr = start; i < npages && restore; i++, addr += PAGE_SIZE) {
2544 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2546 if (!page || (migrate->src[i] & MIGRATE_PFN_MIGRATE))
2549 remove_migration_pte(page, migrate->vma, addr, page);
2551 migrate->src[i] = 0;
2559 * migrate_vma_unmap() - replace page mapping with special migration pte entry
2560 * @migrate: migrate struct containing all migration information
2562 * Replace page mapping (CPU page table pte) with a special migration pte entry
2563 * and check again if it has been pinned. Pinned pages are restored because we
2564 * cannot migrate them.
2566 * This is the last step before we call the device driver callback to allocate
2567 * destination memory and copy contents of original page over to new page.
2569 static void migrate_vma_unmap(struct migrate_vma *migrate)
2571 int flags = TTU_MIGRATION | TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS;
2572 const unsigned long npages = migrate->npages;
2573 const unsigned long start = migrate->start;
2574 unsigned long addr, i, restore = 0;
2576 for (i = 0; i < npages; i++) {
2577 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2579 if (!page || !(migrate->src[i] & MIGRATE_PFN_MIGRATE))
2582 if (page_mapped(page)) {
2583 try_to_unmap(page, flags);
2584 if (page_mapped(page))
2588 if (migrate_vma_check_page(page))
2592 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2597 for (addr = start, i = 0; i < npages && restore; addr += PAGE_SIZE, i++) {
2598 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2600 if (!page || (migrate->src[i] & MIGRATE_PFN_MIGRATE))
2603 remove_migration_ptes(page, page, false);
2605 migrate->src[i] = 0;
2609 if (is_zone_device_page(page))
2612 putback_lru_page(page);
2617 * migrate_vma_setup() - prepare to migrate a range of memory
2618 * @args: contains the vma, start, and and pfns arrays for the migration
2620 * Returns: negative errno on failures, 0 when 0 or more pages were migrated
2623 * Prepare to migrate a range of memory virtual address range by collecting all
2624 * the pages backing each virtual address in the range, saving them inside the
2625 * src array. Then lock those pages and unmap them. Once the pages are locked
2626 * and unmapped, check whether each page is pinned or not. Pages that aren't
2627 * pinned have the MIGRATE_PFN_MIGRATE flag set (by this function) in the
2628 * corresponding src array entry. Then restores any pages that are pinned, by
2629 * remapping and unlocking those pages.
2631 * The caller should then allocate destination memory and copy source memory to
2632 * it for all those entries (ie with MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE
2633 * flag set). Once these are allocated and copied, the caller must update each
2634 * corresponding entry in the dst array with the pfn value of the destination
2635 * page and with the MIGRATE_PFN_VALID and MIGRATE_PFN_LOCKED flags set
2636 * (destination pages must have their struct pages locked, via lock_page()).
2638 * Note that the caller does not have to migrate all the pages that are marked
2639 * with MIGRATE_PFN_MIGRATE flag in src array unless this is a migration from
2640 * device memory to system memory. If the caller cannot migrate a device page
2641 * back to system memory, then it must return VM_FAULT_SIGBUS, which has severe
2642 * consequences for the userspace process, so it must be avoided if at all
2645 * For empty entries inside CPU page table (pte_none() or pmd_none() is true) we
2646 * do set MIGRATE_PFN_MIGRATE flag inside the corresponding source array thus
2647 * allowing the caller to allocate device memory for those unback virtual
2648 * address. For this the caller simply has to allocate device memory and
2649 * properly set the destination entry like for regular migration. Note that
2650 * this can still fails and thus inside the device driver must check if the
2651 * migration was successful for those entries after calling migrate_vma_pages()
2652 * just like for regular migration.
2654 * After that, the callers must call migrate_vma_pages() to go over each entry
2655 * in the src array that has the MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE flag
2656 * set. If the corresponding entry in dst array has MIGRATE_PFN_VALID flag set,
2657 * then migrate_vma_pages() to migrate struct page information from the source
2658 * struct page to the destination struct page. If it fails to migrate the
2659 * struct page information, then it clears the MIGRATE_PFN_MIGRATE flag in the
2662 * At this point all successfully migrated pages have an entry in the src
2663 * array with MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE flag set and the dst
2664 * array entry with MIGRATE_PFN_VALID flag set.
2666 * Once migrate_vma_pages() returns the caller may inspect which pages were
2667 * successfully migrated, and which were not. Successfully migrated pages will
2668 * have the MIGRATE_PFN_MIGRATE flag set for their src array entry.
2670 * It is safe to update device page table after migrate_vma_pages() because
2671 * both destination and source page are still locked, and the mmap_sem is held
2672 * in read mode (hence no one can unmap the range being migrated).
2674 * Once the caller is done cleaning up things and updating its page table (if it
2675 * chose to do so, this is not an obligation) it finally calls
2676 * migrate_vma_finalize() to update the CPU page table to point to new pages
2677 * for successfully migrated pages or otherwise restore the CPU page table to
2678 * point to the original source pages.
2680 int migrate_vma_setup(struct migrate_vma *args)
2682 long nr_pages = (args->end - args->start) >> PAGE_SHIFT;
2684 args->start &= PAGE_MASK;
2685 args->end &= PAGE_MASK;
2686 if (!args->vma || is_vm_hugetlb_page(args->vma) ||
2687 (args->vma->vm_flags & VM_SPECIAL) || vma_is_dax(args->vma))
2691 if (args->start < args->vma->vm_start ||
2692 args->start >= args->vma->vm_end)
2694 if (args->end <= args->vma->vm_start || args->end > args->vma->vm_end)
2696 if (!args->src || !args->dst)
2699 memset(args->src, 0, sizeof(*args->src) * nr_pages);
2703 migrate_vma_collect(args);
2706 migrate_vma_prepare(args);
2708 migrate_vma_unmap(args);
2711 * At this point pages are locked and unmapped, and thus they have
2712 * stable content and can safely be copied to destination memory that
2713 * is allocated by the drivers.
2718 EXPORT_SYMBOL(migrate_vma_setup);
2721 * This code closely matches the code in:
2722 * __handle_mm_fault()
2723 * handle_pte_fault()
2724 * do_anonymous_page()
2725 * to map in an anonymous zero page but the struct page will be a ZONE_DEVICE
2728 static void migrate_vma_insert_page(struct migrate_vma *migrate,
2734 struct vm_area_struct *vma = migrate->vma;
2735 struct mm_struct *mm = vma->vm_mm;
2736 struct mem_cgroup *memcg;
2746 /* Only allow populating anonymous memory */
2747 if (!vma_is_anonymous(vma))
2750 pgdp = pgd_offset(mm, addr);
2751 p4dp = p4d_alloc(mm, pgdp, addr);
2754 pudp = pud_alloc(mm, p4dp, addr);
2757 pmdp = pmd_alloc(mm, pudp, addr);
2761 if (pmd_trans_huge(*pmdp) || pmd_devmap(*pmdp))
2765 * Use pte_alloc() instead of pte_alloc_map(). We can't run
2766 * pte_offset_map() on pmds where a huge pmd might be created
2767 * from a different thread.
2769 * pte_alloc_map() is safe to use under down_write(mmap_sem) or when
2770 * parallel threads are excluded by other means.
2772 * Here we only have down_read(mmap_sem).
2774 if (pte_alloc(mm, pmdp))
2777 /* See the comment in pte_alloc_one_map() */
2778 if (unlikely(pmd_trans_unstable(pmdp)))
2781 if (unlikely(anon_vma_prepare(vma)))
2783 if (mem_cgroup_try_charge(page, vma->vm_mm, GFP_KERNEL, &memcg, false))
2787 * The memory barrier inside __SetPageUptodate makes sure that
2788 * preceding stores to the page contents become visible before
2789 * the set_pte_at() write.
2791 __SetPageUptodate(page);
2793 if (is_zone_device_page(page)) {
2794 if (is_device_private_page(page)) {
2795 swp_entry_t swp_entry;
2797 swp_entry = make_device_private_entry(page, vma->vm_flags & VM_WRITE);
2798 entry = swp_entry_to_pte(swp_entry);
2801 entry = mk_pte(page, vma->vm_page_prot);
2802 if (vma->vm_flags & VM_WRITE)
2803 entry = pte_mkwrite(pte_mkdirty(entry));
2806 ptep = pte_offset_map_lock(mm, pmdp, addr, &ptl);
2808 if (check_stable_address_space(mm))
2811 if (pte_present(*ptep)) {
2812 unsigned long pfn = pte_pfn(*ptep);
2814 if (!is_zero_pfn(pfn))
2817 } else if (!pte_none(*ptep))
2821 * Check for userfaultfd but do not deliver the fault. Instead,
2824 if (userfaultfd_missing(vma))
2827 inc_mm_counter(mm, MM_ANONPAGES);
2828 page_add_new_anon_rmap(page, vma, addr, false);
2829 mem_cgroup_commit_charge(page, memcg, false, false);
2830 if (!is_zone_device_page(page))
2831 lru_cache_add_active_or_unevictable(page, vma);
2835 flush_cache_page(vma, addr, pte_pfn(*ptep));
2836 ptep_clear_flush_notify(vma, addr, ptep);
2837 set_pte_at_notify(mm, addr, ptep, entry);
2838 update_mmu_cache(vma, addr, ptep);
2840 /* No need to invalidate - it was non-present before */
2841 set_pte_at(mm, addr, ptep, entry);
2842 update_mmu_cache(vma, addr, ptep);
2845 pte_unmap_unlock(ptep, ptl);
2846 *src = MIGRATE_PFN_MIGRATE;
2850 pte_unmap_unlock(ptep, ptl);
2851 mem_cgroup_cancel_charge(page, memcg, false);
2853 *src &= ~MIGRATE_PFN_MIGRATE;
2857 * migrate_vma_pages() - migrate meta-data from src page to dst page
2858 * @migrate: migrate struct containing all migration information
2860 * This migrates struct page meta-data from source struct page to destination
2861 * struct page. This effectively finishes the migration from source page to the
2864 void migrate_vma_pages(struct migrate_vma *migrate)
2866 const unsigned long npages = migrate->npages;
2867 const unsigned long start = migrate->start;
2868 struct mmu_notifier_range range;
2869 unsigned long addr, i;
2870 bool notified = false;
2872 for (i = 0, addr = start; i < npages; addr += PAGE_SIZE, i++) {
2873 struct page *newpage = migrate_pfn_to_page(migrate->dst[i]);
2874 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2875 struct address_space *mapping;
2879 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2884 if (!(migrate->src[i] & MIGRATE_PFN_MIGRATE))
2889 mmu_notifier_range_init(&range,
2890 MMU_NOTIFY_CLEAR, 0,
2892 migrate->vma->vm_mm,
2893 addr, migrate->end);
2894 mmu_notifier_invalidate_range_start(&range);
2896 migrate_vma_insert_page(migrate, addr, newpage,
2902 mapping = page_mapping(page);
2904 if (is_zone_device_page(newpage)) {
2905 if (is_device_private_page(newpage)) {
2907 * For now only support private anonymous when
2908 * migrating to un-addressable device memory.
2911 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2916 * Other types of ZONE_DEVICE page are not
2919 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2924 r = migrate_page(mapping, newpage, page, MIGRATE_SYNC_NO_COPY);
2925 if (r != MIGRATEPAGE_SUCCESS)
2926 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2930 * No need to double call mmu_notifier->invalidate_range() callback as
2931 * the above ptep_clear_flush_notify() inside migrate_vma_insert_page()
2932 * did already call it.
2935 mmu_notifier_invalidate_range_only_end(&range);
2937 EXPORT_SYMBOL(migrate_vma_pages);
2940 * migrate_vma_finalize() - restore CPU page table entry
2941 * @migrate: migrate struct containing all migration information
2943 * This replaces the special migration pte entry with either a mapping to the
2944 * new page if migration was successful for that page, or to the original page
2947 * This also unlocks the pages and puts them back on the lru, or drops the extra
2948 * refcount, for device pages.
2950 void migrate_vma_finalize(struct migrate_vma *migrate)
2952 const unsigned long npages = migrate->npages;
2955 for (i = 0; i < npages; i++) {
2956 struct page *newpage = migrate_pfn_to_page(migrate->dst[i]);
2957 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2961 unlock_page(newpage);
2967 if (!(migrate->src[i] & MIGRATE_PFN_MIGRATE) || !newpage) {
2969 unlock_page(newpage);
2975 remove_migration_ptes(page, newpage, false);
2979 if (is_zone_device_page(page))
2982 putback_lru_page(page);
2984 if (newpage != page) {
2985 unlock_page(newpage);
2986 if (is_zone_device_page(newpage))
2989 putback_lru_page(newpage);
2993 EXPORT_SYMBOL(migrate_vma_finalize);
2994 #endif /* CONFIG_DEVICE_PRIVATE */