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_lru(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);
246 else if (pte_swp_uffd_wp(*pvmw.pte))
247 pte = pte_mkuffd_wp(pte);
249 if (unlikely(is_zone_device_page(new))) {
250 if (is_device_private_page(new)) {
251 entry = make_device_private_entry(new, pte_write(pte));
252 pte = swp_entry_to_pte(entry);
253 if (pte_swp_uffd_wp(*pvmw.pte))
254 pte = pte_mkuffd_wp(pte);
258 #ifdef CONFIG_HUGETLB_PAGE
260 pte = pte_mkhuge(pte);
261 pte = arch_make_huge_pte(pte, vma, new, 0);
262 set_huge_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte);
264 hugepage_add_anon_rmap(new, vma, pvmw.address);
266 page_dup_rmap(new, true);
270 set_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte);
273 page_add_anon_rmap(new, vma, pvmw.address, false);
275 page_add_file_rmap(new, false);
277 if (vma->vm_flags & VM_LOCKED && !PageTransCompound(new))
280 if (PageTransHuge(page) && PageMlocked(page))
281 clear_page_mlock(page);
283 /* No need to invalidate - it was non-present before */
284 update_mmu_cache(vma, pvmw.address, pvmw.pte);
291 * Get rid of all migration entries and replace them by
292 * references to the indicated page.
294 void remove_migration_ptes(struct page *old, struct page *new, bool locked)
296 struct rmap_walk_control rwc = {
297 .rmap_one = remove_migration_pte,
302 rmap_walk_locked(new, &rwc);
304 rmap_walk(new, &rwc);
308 * Something used the pte of a page under migration. We need to
309 * get to the page and wait until migration is finished.
310 * When we return from this function the fault will be retried.
312 void __migration_entry_wait(struct mm_struct *mm, pte_t *ptep,
321 if (!is_swap_pte(pte))
324 entry = pte_to_swp_entry(pte);
325 if (!is_migration_entry(entry))
328 page = migration_entry_to_page(entry);
331 * Once page cache replacement of page migration started, page_count
332 * is zero; but we must not call put_and_wait_on_page_locked() without
333 * a ref. Use get_page_unless_zero(), and just fault again if it fails.
335 if (!get_page_unless_zero(page))
337 pte_unmap_unlock(ptep, ptl);
338 put_and_wait_on_page_locked(page);
341 pte_unmap_unlock(ptep, ptl);
344 void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
345 unsigned long address)
347 spinlock_t *ptl = pte_lockptr(mm, pmd);
348 pte_t *ptep = pte_offset_map(pmd, address);
349 __migration_entry_wait(mm, ptep, ptl);
352 void migration_entry_wait_huge(struct vm_area_struct *vma,
353 struct mm_struct *mm, pte_t *pte)
355 spinlock_t *ptl = huge_pte_lockptr(hstate_vma(vma), mm, pte);
356 __migration_entry_wait(mm, pte, ptl);
359 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
360 void pmd_migration_entry_wait(struct mm_struct *mm, pmd_t *pmd)
365 ptl = pmd_lock(mm, pmd);
366 if (!is_pmd_migration_entry(*pmd))
368 page = migration_entry_to_page(pmd_to_swp_entry(*pmd));
369 if (!get_page_unless_zero(page))
372 put_and_wait_on_page_locked(page);
379 static int expected_page_refs(struct address_space *mapping, struct page *page)
381 int expected_count = 1;
384 * Device public or private pages have an extra refcount as they are
387 expected_count += is_device_private_page(page);
389 expected_count += hpage_nr_pages(page) + page_has_private(page);
391 return expected_count;
395 * Replace the page in the mapping.
397 * The number of remaining references must be:
398 * 1 for anonymous pages without a mapping
399 * 2 for pages with a mapping
400 * 3 for pages with a mapping and PagePrivate/PagePrivate2 set.
402 int migrate_page_move_mapping(struct address_space *mapping,
403 struct page *newpage, struct page *page, int extra_count)
405 XA_STATE(xas, &mapping->i_pages, page_index(page));
406 struct zone *oldzone, *newzone;
408 int expected_count = expected_page_refs(mapping, page) + extra_count;
411 /* Anonymous page without mapping */
412 if (page_count(page) != expected_count)
415 /* No turning back from here */
416 newpage->index = page->index;
417 newpage->mapping = page->mapping;
418 if (PageSwapBacked(page))
419 __SetPageSwapBacked(newpage);
421 return MIGRATEPAGE_SUCCESS;
424 oldzone = page_zone(page);
425 newzone = page_zone(newpage);
428 if (page_count(page) != expected_count || xas_load(&xas) != page) {
429 xas_unlock_irq(&xas);
433 if (!page_ref_freeze(page, expected_count)) {
434 xas_unlock_irq(&xas);
439 * Now we know that no one else is looking at the page:
440 * no turning back from here.
442 newpage->index = page->index;
443 newpage->mapping = page->mapping;
444 page_ref_add(newpage, hpage_nr_pages(page)); /* add cache reference */
445 if (PageSwapBacked(page)) {
446 __SetPageSwapBacked(newpage);
447 if (PageSwapCache(page)) {
448 SetPageSwapCache(newpage);
449 set_page_private(newpage, page_private(page));
452 VM_BUG_ON_PAGE(PageSwapCache(page), page);
455 /* Move dirty while page refs frozen and newpage not yet exposed */
456 dirty = PageDirty(page);
458 ClearPageDirty(page);
459 SetPageDirty(newpage);
462 xas_store(&xas, newpage);
463 if (PageTransHuge(page)) {
466 for (i = 1; i < HPAGE_PMD_NR; i++) {
468 xas_store(&xas, newpage);
473 * Drop cache reference from old page by unfreezing
474 * to one less reference.
475 * We know this isn't the last reference.
477 page_ref_unfreeze(page, expected_count - hpage_nr_pages(page));
480 /* Leave irq disabled to prevent preemption while updating stats */
483 * If moved to a different zone then also account
484 * the page for that zone. Other VM counters will be
485 * taken care of when we establish references to the
486 * new page and drop references to the old page.
488 * Note that anonymous pages are accounted for
489 * via NR_FILE_PAGES and NR_ANON_MAPPED if they
490 * are mapped to swap space.
492 if (newzone != oldzone) {
493 __dec_node_state(oldzone->zone_pgdat, NR_FILE_PAGES);
494 __inc_node_state(newzone->zone_pgdat, NR_FILE_PAGES);
495 if (PageSwapBacked(page) && !PageSwapCache(page)) {
496 __dec_node_state(oldzone->zone_pgdat, NR_SHMEM);
497 __inc_node_state(newzone->zone_pgdat, NR_SHMEM);
499 if (dirty && mapping_cap_account_dirty(mapping)) {
500 __dec_node_state(oldzone->zone_pgdat, NR_FILE_DIRTY);
501 __dec_zone_state(oldzone, NR_ZONE_WRITE_PENDING);
502 __inc_node_state(newzone->zone_pgdat, NR_FILE_DIRTY);
503 __inc_zone_state(newzone, NR_ZONE_WRITE_PENDING);
508 return MIGRATEPAGE_SUCCESS;
510 EXPORT_SYMBOL(migrate_page_move_mapping);
513 * The expected number of remaining references is the same as that
514 * of migrate_page_move_mapping().
516 int migrate_huge_page_move_mapping(struct address_space *mapping,
517 struct page *newpage, struct page *page)
519 XA_STATE(xas, &mapping->i_pages, page_index(page));
523 expected_count = 2 + page_has_private(page);
524 if (page_count(page) != expected_count || xas_load(&xas) != page) {
525 xas_unlock_irq(&xas);
529 if (!page_ref_freeze(page, expected_count)) {
530 xas_unlock_irq(&xas);
534 newpage->index = page->index;
535 newpage->mapping = page->mapping;
539 xas_store(&xas, newpage);
541 page_ref_unfreeze(page, expected_count - 1);
543 xas_unlock_irq(&xas);
545 return MIGRATEPAGE_SUCCESS;
549 * Gigantic pages are so large that we do not guarantee that page++ pointer
550 * arithmetic will work across the entire page. We need something more
553 static void __copy_gigantic_page(struct page *dst, struct page *src,
557 struct page *dst_base = dst;
558 struct page *src_base = src;
560 for (i = 0; i < nr_pages; ) {
562 copy_highpage(dst, src);
565 dst = mem_map_next(dst, dst_base, i);
566 src = mem_map_next(src, src_base, i);
570 static void copy_huge_page(struct page *dst, struct page *src)
577 struct hstate *h = page_hstate(src);
578 nr_pages = pages_per_huge_page(h);
580 if (unlikely(nr_pages > MAX_ORDER_NR_PAGES)) {
581 __copy_gigantic_page(dst, src, nr_pages);
586 BUG_ON(!PageTransHuge(src));
587 nr_pages = hpage_nr_pages(src);
590 for (i = 0; i < nr_pages; i++) {
592 copy_highpage(dst + i, src + i);
597 * Copy the page to its new location
599 void migrate_page_states(struct page *newpage, struct page *page)
604 SetPageError(newpage);
605 if (PageReferenced(page))
606 SetPageReferenced(newpage);
607 if (PageUptodate(page))
608 SetPageUptodate(newpage);
609 if (TestClearPageActive(page)) {
610 VM_BUG_ON_PAGE(PageUnevictable(page), page);
611 SetPageActive(newpage);
612 } else if (TestClearPageUnevictable(page))
613 SetPageUnevictable(newpage);
614 if (PageWorkingset(page))
615 SetPageWorkingset(newpage);
616 if (PageChecked(page))
617 SetPageChecked(newpage);
618 if (PageMappedToDisk(page))
619 SetPageMappedToDisk(newpage);
621 /* Move dirty on pages not done by migrate_page_move_mapping() */
623 SetPageDirty(newpage);
625 if (page_is_young(page))
626 set_page_young(newpage);
627 if (page_is_idle(page))
628 set_page_idle(newpage);
631 * Copy NUMA information to the new page, to prevent over-eager
632 * future migrations of this same page.
634 cpupid = page_cpupid_xchg_last(page, -1);
635 page_cpupid_xchg_last(newpage, cpupid);
637 ksm_migrate_page(newpage, page);
639 * Please do not reorder this without considering how mm/ksm.c's
640 * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache().
642 if (PageSwapCache(page))
643 ClearPageSwapCache(page);
644 ClearPagePrivate(page);
645 set_page_private(page, 0);
648 * If any waiters have accumulated on the new page then
651 if (PageWriteback(newpage))
652 end_page_writeback(newpage);
655 * PG_readahead shares the same bit with PG_reclaim. The above
656 * end_page_writeback() may clear PG_readahead mistakenly, so set the
659 if (PageReadahead(page))
660 SetPageReadahead(newpage);
662 copy_page_owner(page, newpage);
664 mem_cgroup_migrate(page, newpage);
666 EXPORT_SYMBOL(migrate_page_states);
668 void migrate_page_copy(struct page *newpage, struct page *page)
670 if (PageHuge(page) || PageTransHuge(page))
671 copy_huge_page(newpage, page);
673 copy_highpage(newpage, page);
675 migrate_page_states(newpage, page);
677 EXPORT_SYMBOL(migrate_page_copy);
679 /************************************************************
680 * Migration functions
681 ***********************************************************/
684 * Common logic to directly migrate a single LRU page suitable for
685 * pages that do not use PagePrivate/PagePrivate2.
687 * Pages are locked upon entry and exit.
689 int migrate_page(struct address_space *mapping,
690 struct page *newpage, struct page *page,
691 enum migrate_mode mode)
695 BUG_ON(PageWriteback(page)); /* Writeback must be complete */
697 rc = migrate_page_move_mapping(mapping, newpage, page, 0);
699 if (rc != MIGRATEPAGE_SUCCESS)
702 if (mode != MIGRATE_SYNC_NO_COPY)
703 migrate_page_copy(newpage, page);
705 migrate_page_states(newpage, page);
706 return MIGRATEPAGE_SUCCESS;
708 EXPORT_SYMBOL(migrate_page);
711 /* Returns true if all buffers are successfully locked */
712 static bool buffer_migrate_lock_buffers(struct buffer_head *head,
713 enum migrate_mode mode)
715 struct buffer_head *bh = head;
717 /* Simple case, sync compaction */
718 if (mode != MIGRATE_ASYNC) {
721 bh = bh->b_this_page;
723 } while (bh != head);
728 /* async case, we cannot block on lock_buffer so use trylock_buffer */
730 if (!trylock_buffer(bh)) {
732 * We failed to lock the buffer and cannot stall in
733 * async migration. Release the taken locks
735 struct buffer_head *failed_bh = bh;
737 while (bh != failed_bh) {
739 bh = bh->b_this_page;
744 bh = bh->b_this_page;
745 } while (bh != head);
749 static int __buffer_migrate_page(struct address_space *mapping,
750 struct page *newpage, struct page *page, enum migrate_mode mode,
753 struct buffer_head *bh, *head;
757 if (!page_has_buffers(page))
758 return migrate_page(mapping, newpage, page, mode);
760 /* Check whether page does not have extra refs before we do more work */
761 expected_count = expected_page_refs(mapping, page);
762 if (page_count(page) != expected_count)
765 head = page_buffers(page);
766 if (!buffer_migrate_lock_buffers(head, mode))
771 bool invalidated = false;
775 spin_lock(&mapping->private_lock);
778 if (atomic_read(&bh->b_count)) {
782 bh = bh->b_this_page;
783 } while (bh != head);
789 spin_unlock(&mapping->private_lock);
790 invalidate_bh_lrus();
792 goto recheck_buffers;
796 rc = migrate_page_move_mapping(mapping, newpage, page, 0);
797 if (rc != MIGRATEPAGE_SUCCESS)
800 attach_page_private(newpage, detach_page_private(page));
805 set_bh_page(bh, newpage, bh_offset(bh));
806 bh = bh->b_this_page;
808 } while (bh != head);
810 if (mode != MIGRATE_SYNC_NO_COPY)
811 migrate_page_copy(newpage, page);
813 migrate_page_states(newpage, page);
815 rc = MIGRATEPAGE_SUCCESS;
818 spin_unlock(&mapping->private_lock);
822 bh = bh->b_this_page;
824 } while (bh != head);
830 * Migration function for pages with buffers. This function can only be used
831 * if the underlying filesystem guarantees that no other references to "page"
832 * exist. For example attached buffer heads are accessed only under page lock.
834 int buffer_migrate_page(struct address_space *mapping,
835 struct page *newpage, struct page *page, enum migrate_mode mode)
837 return __buffer_migrate_page(mapping, newpage, page, mode, false);
839 EXPORT_SYMBOL(buffer_migrate_page);
842 * Same as above except that this variant is more careful and checks that there
843 * are also no buffer head references. This function is the right one for
844 * mappings where buffer heads are directly looked up and referenced (such as
845 * block device mappings).
847 int buffer_migrate_page_norefs(struct address_space *mapping,
848 struct page *newpage, struct page *page, enum migrate_mode mode)
850 return __buffer_migrate_page(mapping, newpage, page, mode, true);
855 * Writeback a page to clean the dirty state
857 static int writeout(struct address_space *mapping, struct page *page)
859 struct writeback_control wbc = {
860 .sync_mode = WB_SYNC_NONE,
863 .range_end = LLONG_MAX,
868 if (!mapping->a_ops->writepage)
869 /* No write method for the address space */
872 if (!clear_page_dirty_for_io(page))
873 /* Someone else already triggered a write */
877 * A dirty page may imply that the underlying filesystem has
878 * the page on some queue. So the page must be clean for
879 * migration. Writeout may mean we loose the lock and the
880 * page state is no longer what we checked for earlier.
881 * At this point we know that the migration attempt cannot
884 remove_migration_ptes(page, page, false);
886 rc = mapping->a_ops->writepage(page, &wbc);
888 if (rc != AOP_WRITEPAGE_ACTIVATE)
889 /* unlocked. Relock */
892 return (rc < 0) ? -EIO : -EAGAIN;
896 * Default handling if a filesystem does not provide a migration function.
898 static int fallback_migrate_page(struct address_space *mapping,
899 struct page *newpage, struct page *page, enum migrate_mode mode)
901 if (PageDirty(page)) {
902 /* Only writeback pages in full synchronous migration */
905 case MIGRATE_SYNC_NO_COPY:
910 return writeout(mapping, page);
914 * Buffers may be managed in a filesystem specific way.
915 * We must have no buffers or drop them.
917 if (page_has_private(page) &&
918 !try_to_release_page(page, GFP_KERNEL))
919 return mode == MIGRATE_SYNC ? -EAGAIN : -EBUSY;
921 return migrate_page(mapping, newpage, page, mode);
925 * Move a page to a newly allocated page
926 * The page is locked and all ptes have been successfully removed.
928 * The new page will have replaced the old page if this function
933 * MIGRATEPAGE_SUCCESS - success
935 static int move_to_new_page(struct page *newpage, struct page *page,
936 enum migrate_mode mode)
938 struct address_space *mapping;
940 bool is_lru = !__PageMovable(page);
942 VM_BUG_ON_PAGE(!PageLocked(page), page);
943 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
945 mapping = page_mapping(page);
947 if (likely(is_lru)) {
949 rc = migrate_page(mapping, newpage, page, mode);
950 else if (mapping->a_ops->migratepage)
952 * Most pages have a mapping and most filesystems
953 * provide a migratepage callback. Anonymous pages
954 * are part of swap space which also has its own
955 * migratepage callback. This is the most common path
956 * for page migration.
958 rc = mapping->a_ops->migratepage(mapping, newpage,
961 rc = fallback_migrate_page(mapping, newpage,
965 * In case of non-lru page, it could be released after
966 * isolation step. In that case, we shouldn't try migration.
968 VM_BUG_ON_PAGE(!PageIsolated(page), page);
969 if (!PageMovable(page)) {
970 rc = MIGRATEPAGE_SUCCESS;
971 __ClearPageIsolated(page);
975 rc = mapping->a_ops->migratepage(mapping, newpage,
977 WARN_ON_ONCE(rc == MIGRATEPAGE_SUCCESS &&
978 !PageIsolated(page));
982 * When successful, old pagecache page->mapping must be cleared before
983 * page is freed; but stats require that PageAnon be left as PageAnon.
985 if (rc == MIGRATEPAGE_SUCCESS) {
986 if (__PageMovable(page)) {
987 VM_BUG_ON_PAGE(!PageIsolated(page), page);
990 * We clear PG_movable under page_lock so any compactor
991 * cannot try to migrate this page.
993 __ClearPageIsolated(page);
997 * Anonymous and movable page->mapping will be cleared by
998 * free_pages_prepare so don't reset it here for keeping
999 * the type to work PageAnon, for example.
1001 if (!PageMappingFlags(page))
1002 page->mapping = NULL;
1004 if (likely(!is_zone_device_page(newpage)))
1005 flush_dcache_page(newpage);
1012 static int __unmap_and_move(struct page *page, struct page *newpage,
1013 int force, enum migrate_mode mode)
1016 int page_was_mapped = 0;
1017 struct anon_vma *anon_vma = NULL;
1018 bool is_lru = !__PageMovable(page);
1020 if (!trylock_page(page)) {
1021 if (!force || mode == MIGRATE_ASYNC)
1025 * It's not safe for direct compaction to call lock_page.
1026 * For example, during page readahead pages are added locked
1027 * to the LRU. Later, when the IO completes the pages are
1028 * marked uptodate and unlocked. However, the queueing
1029 * could be merging multiple pages for one bio (e.g.
1030 * mpage_readahead). If an allocation happens for the
1031 * second or third page, the process can end up locking
1032 * the same page twice and deadlocking. Rather than
1033 * trying to be clever about what pages can be locked,
1034 * avoid the use of lock_page for direct compaction
1037 if (current->flags & PF_MEMALLOC)
1043 if (PageWriteback(page)) {
1045 * Only in the case of a full synchronous migration is it
1046 * necessary to wait for PageWriteback. In the async case,
1047 * the retry loop is too short and in the sync-light case,
1048 * the overhead of stalling is too much
1052 case MIGRATE_SYNC_NO_COPY:
1060 wait_on_page_writeback(page);
1064 * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
1065 * we cannot notice that anon_vma is freed while we migrates a page.
1066 * This get_anon_vma() delays freeing anon_vma pointer until the end
1067 * of migration. File cache pages are no problem because of page_lock()
1068 * File Caches may use write_page() or lock_page() in migration, then,
1069 * just care Anon page here.
1071 * Only page_get_anon_vma() understands the subtleties of
1072 * getting a hold on an anon_vma from outside one of its mms.
1073 * But if we cannot get anon_vma, then we won't need it anyway,
1074 * because that implies that the anon page is no longer mapped
1075 * (and cannot be remapped so long as we hold the page lock).
1077 if (PageAnon(page) && !PageKsm(page))
1078 anon_vma = page_get_anon_vma(page);
1081 * Block others from accessing the new page when we get around to
1082 * establishing additional references. We are usually the only one
1083 * holding a reference to newpage at this point. We used to have a BUG
1084 * here if trylock_page(newpage) fails, but would like to allow for
1085 * cases where there might be a race with the previous use of newpage.
1086 * This is much like races on refcount of oldpage: just don't BUG().
1088 if (unlikely(!trylock_page(newpage)))
1091 if (unlikely(!is_lru)) {
1092 rc = move_to_new_page(newpage, page, mode);
1093 goto out_unlock_both;
1097 * Corner case handling:
1098 * 1. When a new swap-cache page is read into, it is added to the LRU
1099 * and treated as swapcache but it has no rmap yet.
1100 * Calling try_to_unmap() against a page->mapping==NULL page will
1101 * trigger a BUG. So handle it here.
1102 * 2. An orphaned page (see truncate_complete_page) might have
1103 * fs-private metadata. The page can be picked up due to memory
1104 * offlining. Everywhere else except page reclaim, the page is
1105 * invisible to the vm, so the page can not be migrated. So try to
1106 * free the metadata, so the page can be freed.
1108 if (!page->mapping) {
1109 VM_BUG_ON_PAGE(PageAnon(page), page);
1110 if (page_has_private(page)) {
1111 try_to_free_buffers(page);
1112 goto out_unlock_both;
1114 } else if (page_mapped(page)) {
1115 /* Establish migration ptes */
1116 VM_BUG_ON_PAGE(PageAnon(page) && !PageKsm(page) && !anon_vma,
1119 TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
1120 page_was_mapped = 1;
1123 if (!page_mapped(page))
1124 rc = move_to_new_page(newpage, page, mode);
1126 if (page_was_mapped)
1127 remove_migration_ptes(page,
1128 rc == MIGRATEPAGE_SUCCESS ? newpage : page, false);
1131 unlock_page(newpage);
1133 /* Drop an anon_vma reference if we took one */
1135 put_anon_vma(anon_vma);
1139 * If migration is successful, decrease refcount of the newpage
1140 * which will not free the page because new page owner increased
1141 * refcounter. As well, if it is LRU page, add the page to LRU
1142 * list in here. Use the old state of the isolated source page to
1143 * determine if we migrated a LRU page. newpage was already unlocked
1144 * and possibly modified by its owner - don't rely on the page
1147 if (rc == MIGRATEPAGE_SUCCESS) {
1148 if (unlikely(!is_lru))
1151 putback_lru_page(newpage);
1158 * gcc 4.7 and 4.8 on arm get an ICEs when inlining unmap_and_move(). Work
1161 #if defined(CONFIG_ARM) && \
1162 defined(GCC_VERSION) && GCC_VERSION < 40900 && GCC_VERSION >= 40700
1163 #define ICE_noinline noinline
1165 #define ICE_noinline
1169 * Obtain the lock on page, remove all ptes and migrate the page
1170 * to the newly allocated page in newpage.
1172 static ICE_noinline int unmap_and_move(new_page_t get_new_page,
1173 free_page_t put_new_page,
1174 unsigned long private, struct page *page,
1175 int force, enum migrate_mode mode,
1176 enum migrate_reason reason)
1178 int rc = MIGRATEPAGE_SUCCESS;
1179 struct page *newpage = NULL;
1181 if (!thp_migration_supported() && PageTransHuge(page))
1184 if (page_count(page) == 1) {
1185 /* page was freed from under us. So we are done. */
1186 ClearPageActive(page);
1187 ClearPageUnevictable(page);
1188 if (unlikely(__PageMovable(page))) {
1190 if (!PageMovable(page))
1191 __ClearPageIsolated(page);
1197 newpage = get_new_page(page, private);
1201 rc = __unmap_and_move(page, newpage, force, mode);
1202 if (rc == MIGRATEPAGE_SUCCESS)
1203 set_page_owner_migrate_reason(newpage, reason);
1206 if (rc != -EAGAIN) {
1208 * A page that has been migrated has all references
1209 * removed and will be freed. A page that has not been
1210 * migrated will have kept its references and be restored.
1212 list_del(&page->lru);
1215 * Compaction can migrate also non-LRU pages which are
1216 * not accounted to NR_ISOLATED_*. They can be recognized
1219 if (likely(!__PageMovable(page)))
1220 mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON +
1221 page_is_file_lru(page), -hpage_nr_pages(page));
1225 * If migration is successful, releases reference grabbed during
1226 * isolation. Otherwise, restore the page to right list unless
1229 if (rc == MIGRATEPAGE_SUCCESS) {
1231 if (reason == MR_MEMORY_FAILURE) {
1233 * Set PG_HWPoison on just freed page
1234 * intentionally. Although it's rather weird,
1235 * it's how HWPoison flag works at the moment.
1237 if (set_hwpoison_free_buddy_page(page))
1238 num_poisoned_pages_inc();
1241 if (rc != -EAGAIN) {
1242 if (likely(!__PageMovable(page))) {
1243 putback_lru_page(page);
1248 if (PageMovable(page))
1249 putback_movable_page(page);
1251 __ClearPageIsolated(page);
1257 put_new_page(newpage, private);
1266 * Counterpart of unmap_and_move_page() for hugepage migration.
1268 * This function doesn't wait the completion of hugepage I/O
1269 * because there is no race between I/O and migration for hugepage.
1270 * Note that currently hugepage I/O occurs only in direct I/O
1271 * where no lock is held and PG_writeback is irrelevant,
1272 * and writeback status of all subpages are counted in the reference
1273 * count of the head page (i.e. if all subpages of a 2MB hugepage are
1274 * under direct I/O, the reference of the head page is 512 and a bit more.)
1275 * This means that when we try to migrate hugepage whose subpages are
1276 * doing direct I/O, some references remain after try_to_unmap() and
1277 * hugepage migration fails without data corruption.
1279 * There is also no race when direct I/O is issued on the page under migration,
1280 * because then pte is replaced with migration swap entry and direct I/O code
1281 * will wait in the page fault for migration to complete.
1283 static int unmap_and_move_huge_page(new_page_t get_new_page,
1284 free_page_t put_new_page, unsigned long private,
1285 struct page *hpage, int force,
1286 enum migrate_mode mode, int reason)
1289 int page_was_mapped = 0;
1290 struct page *new_hpage;
1291 struct anon_vma *anon_vma = NULL;
1292 struct address_space *mapping = NULL;
1295 * Migratability of hugepages depends on architectures and their size.
1296 * This check is necessary because some callers of hugepage migration
1297 * like soft offline and memory hotremove don't walk through page
1298 * tables or check whether the hugepage is pmd-based or not before
1299 * kicking migration.
1301 if (!hugepage_migration_supported(page_hstate(hpage))) {
1302 putback_active_hugepage(hpage);
1306 new_hpage = get_new_page(hpage, private);
1310 if (!trylock_page(hpage)) {
1315 case MIGRATE_SYNC_NO_COPY:
1324 * Check for pages which are in the process of being freed. Without
1325 * page_mapping() set, hugetlbfs specific move page routine will not
1326 * be called and we could leak usage counts for subpools.
1328 if (page_private(hpage) && !page_mapping(hpage)) {
1333 if (PageAnon(hpage))
1334 anon_vma = page_get_anon_vma(hpage);
1336 if (unlikely(!trylock_page(new_hpage)))
1339 if (page_mapped(hpage)) {
1341 * try_to_unmap could potentially call huge_pmd_unshare.
1342 * Because of this, take semaphore in write mode here and
1343 * set TTU_RMAP_LOCKED to let lower levels know we have
1346 mapping = hugetlb_page_mapping_lock_write(hpage);
1347 if (unlikely(!mapping))
1348 goto unlock_put_anon;
1351 TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS|
1353 page_was_mapped = 1;
1355 * Leave mapping locked until after subsequent call to
1356 * remove_migration_ptes()
1360 if (!page_mapped(hpage))
1361 rc = move_to_new_page(new_hpage, hpage, mode);
1363 if (page_was_mapped) {
1364 remove_migration_ptes(hpage,
1365 rc == MIGRATEPAGE_SUCCESS ? new_hpage : hpage, true);
1366 i_mmap_unlock_write(mapping);
1370 unlock_page(new_hpage);
1374 put_anon_vma(anon_vma);
1376 if (rc == MIGRATEPAGE_SUCCESS) {
1377 move_hugetlb_state(hpage, new_hpage, reason);
1378 put_new_page = NULL;
1385 putback_active_hugepage(hpage);
1388 * If migration was not successful and there's a freeing callback, use
1389 * it. Otherwise, put_page() will drop the reference grabbed during
1393 put_new_page(new_hpage, private);
1395 putback_active_hugepage(new_hpage);
1401 * migrate_pages - migrate the pages specified in a list, to the free pages
1402 * supplied as the target for the page migration
1404 * @from: The list of pages to be migrated.
1405 * @get_new_page: The function used to allocate free pages to be used
1406 * as the target of the page migration.
1407 * @put_new_page: The function used to free target pages if migration
1408 * fails, or NULL if no special handling is necessary.
1409 * @private: Private data to be passed on to get_new_page()
1410 * @mode: The migration mode that specifies the constraints for
1411 * page migration, if any.
1412 * @reason: The reason for page migration.
1414 * The function returns after 10 attempts or if no pages are movable any more
1415 * because the list has become empty or no retryable pages exist any more.
1416 * The caller should call putback_movable_pages() to return pages to the LRU
1417 * or free list only if ret != 0.
1419 * Returns the number of pages that were not migrated, or an error code.
1421 int migrate_pages(struct list_head *from, new_page_t get_new_page,
1422 free_page_t put_new_page, unsigned long private,
1423 enum migrate_mode mode, int reason)
1427 int nr_succeeded = 0;
1431 int swapwrite = current->flags & PF_SWAPWRITE;
1435 current->flags |= PF_SWAPWRITE;
1437 for(pass = 0; pass < 10 && retry; pass++) {
1440 list_for_each_entry_safe(page, page2, from, lru) {
1445 rc = unmap_and_move_huge_page(get_new_page,
1446 put_new_page, private, page,
1447 pass > 2, mode, reason);
1449 rc = unmap_and_move(get_new_page, put_new_page,
1450 private, page, pass > 2, mode,
1456 * THP migration might be unsupported or the
1457 * allocation could've failed so we should
1458 * retry on the same page with the THP split
1461 * Head page is retried immediately and tail
1462 * pages are added to the tail of the list so
1463 * we encounter them after the rest of the list
1466 if (PageTransHuge(page) && !PageHuge(page)) {
1468 rc = split_huge_page_to_list(page, from);
1471 list_safe_reset_next(page, page2, lru);
1480 case MIGRATEPAGE_SUCCESS:
1485 * Permanent failure (-EBUSY, -ENOSYS, etc.):
1486 * unlike -EAGAIN case, the failed page is
1487 * removed from migration page list and not
1488 * retried in the next outer loop.
1499 count_vm_events(PGMIGRATE_SUCCESS, nr_succeeded);
1501 count_vm_events(PGMIGRATE_FAIL, nr_failed);
1502 trace_mm_migrate_pages(nr_succeeded, nr_failed, mode, reason);
1505 current->flags &= ~PF_SWAPWRITE;
1512 static int store_status(int __user *status, int start, int value, int nr)
1515 if (put_user(value, status + start))
1523 static int do_move_pages_to_node(struct mm_struct *mm,
1524 struct list_head *pagelist, int node)
1528 err = migrate_pages(pagelist, alloc_new_node_page, NULL, node,
1529 MIGRATE_SYNC, MR_SYSCALL);
1531 putback_movable_pages(pagelist);
1536 * Resolves the given address to a struct page, isolates it from the LRU and
1537 * puts it to the given pagelist.
1539 * errno - if the page cannot be found/isolated
1540 * 0 - when it doesn't have to be migrated because it is already on the
1542 * 1 - when it has been queued
1544 static int add_page_for_migration(struct mm_struct *mm, unsigned long addr,
1545 int node, struct list_head *pagelist, bool migrate_all)
1547 struct vm_area_struct *vma;
1549 unsigned int follflags;
1552 down_read(&mm->mmap_sem);
1554 vma = find_vma(mm, addr);
1555 if (!vma || addr < vma->vm_start || !vma_migratable(vma))
1558 /* FOLL_DUMP to ignore special (like zero) pages */
1559 follflags = FOLL_GET | FOLL_DUMP;
1560 page = follow_page(vma, addr, follflags);
1562 err = PTR_ERR(page);
1571 if (page_to_nid(page) == node)
1575 if (page_mapcount(page) > 1 && !migrate_all)
1578 if (PageHuge(page)) {
1579 if (PageHead(page)) {
1580 isolate_huge_page(page, pagelist);
1586 head = compound_head(page);
1587 err = isolate_lru_page(head);
1592 list_add_tail(&head->lru, pagelist);
1593 mod_node_page_state(page_pgdat(head),
1594 NR_ISOLATED_ANON + page_is_file_lru(head),
1595 hpage_nr_pages(head));
1599 * Either remove the duplicate refcount from
1600 * isolate_lru_page() or drop the page ref if it was
1605 up_read(&mm->mmap_sem);
1609 static int move_pages_and_store_status(struct mm_struct *mm, int node,
1610 struct list_head *pagelist, int __user *status,
1611 int start, int i, unsigned long nr_pages)
1615 if (list_empty(pagelist))
1618 err = do_move_pages_to_node(mm, pagelist, node);
1621 * Positive err means the number of failed
1622 * pages to migrate. Since we are going to
1623 * abort and return the number of non-migrated
1624 * pages, so need to incude the rest of the
1625 * nr_pages that have not been attempted as
1629 err += nr_pages - i - 1;
1632 return store_status(status, start, node, i - start);
1636 * Migrate an array of page address onto an array of nodes and fill
1637 * the corresponding array of status.
1639 static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes,
1640 unsigned long nr_pages,
1641 const void __user * __user *pages,
1642 const int __user *nodes,
1643 int __user *status, int flags)
1645 int current_node = NUMA_NO_NODE;
1646 LIST_HEAD(pagelist);
1652 for (i = start = 0; i < nr_pages; i++) {
1653 const void __user *p;
1658 if (get_user(p, pages + i))
1660 if (get_user(node, nodes + i))
1662 addr = (unsigned long)untagged_addr(p);
1665 if (node < 0 || node >= MAX_NUMNODES)
1667 if (!node_state(node, N_MEMORY))
1671 if (!node_isset(node, task_nodes))
1674 if (current_node == NUMA_NO_NODE) {
1675 current_node = node;
1677 } else if (node != current_node) {
1678 err = move_pages_and_store_status(mm, current_node,
1679 &pagelist, status, start, i, nr_pages);
1683 current_node = node;
1687 * Errors in the page lookup or isolation are not fatal and we simply
1688 * report them via status
1690 err = add_page_for_migration(mm, addr, current_node,
1691 &pagelist, flags & MPOL_MF_MOVE_ALL);
1694 /* The page is successfully queued for migration */
1699 * If the page is already on the target node (!err), store the
1700 * node, otherwise, store the err.
1702 err = store_status(status, i, err ? : current_node, 1);
1706 err = move_pages_and_store_status(mm, current_node, &pagelist,
1707 status, start, i, nr_pages);
1710 current_node = NUMA_NO_NODE;
1713 /* Make sure we do not overwrite the existing error */
1714 err1 = move_pages_and_store_status(mm, current_node, &pagelist,
1715 status, start, i, nr_pages);
1723 * Determine the nodes of an array of pages and store it in an array of status.
1725 static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages,
1726 const void __user **pages, int *status)
1730 down_read(&mm->mmap_sem);
1732 for (i = 0; i < nr_pages; i++) {
1733 unsigned long addr = (unsigned long)(*pages);
1734 struct vm_area_struct *vma;
1738 vma = find_vma(mm, addr);
1739 if (!vma || addr < vma->vm_start)
1742 /* FOLL_DUMP to ignore special (like zero) pages */
1743 page = follow_page(vma, addr, FOLL_DUMP);
1745 err = PTR_ERR(page);
1749 err = page ? page_to_nid(page) : -ENOENT;
1757 up_read(&mm->mmap_sem);
1761 * Determine the nodes of a user array of pages and store it in
1762 * a user array of status.
1764 static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages,
1765 const void __user * __user *pages,
1768 #define DO_PAGES_STAT_CHUNK_NR 16
1769 const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR];
1770 int chunk_status[DO_PAGES_STAT_CHUNK_NR];
1773 unsigned long chunk_nr;
1775 chunk_nr = nr_pages;
1776 if (chunk_nr > DO_PAGES_STAT_CHUNK_NR)
1777 chunk_nr = DO_PAGES_STAT_CHUNK_NR;
1779 if (copy_from_user(chunk_pages, pages, chunk_nr * sizeof(*chunk_pages)))
1782 do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status);
1784 if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status)))
1789 nr_pages -= chunk_nr;
1791 return nr_pages ? -EFAULT : 0;
1795 * Move a list of pages in the address space of the currently executing
1798 static int kernel_move_pages(pid_t pid, unsigned long nr_pages,
1799 const void __user * __user *pages,
1800 const int __user *nodes,
1801 int __user *status, int flags)
1803 struct task_struct *task;
1804 struct mm_struct *mm;
1806 nodemask_t task_nodes;
1809 if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
1812 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1815 /* Find the mm_struct */
1817 task = pid ? find_task_by_vpid(pid) : current;
1822 get_task_struct(task);
1825 * Check if this process has the right to modify the specified
1826 * process. Use the regular "ptrace_may_access()" checks.
1828 if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) {
1835 err = security_task_movememory(task);
1839 task_nodes = cpuset_mems_allowed(task);
1840 mm = get_task_mm(task);
1841 put_task_struct(task);
1847 err = do_pages_move(mm, task_nodes, nr_pages, pages,
1848 nodes, status, flags);
1850 err = do_pages_stat(mm, nr_pages, pages, status);
1856 put_task_struct(task);
1860 SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages,
1861 const void __user * __user *, pages,
1862 const int __user *, nodes,
1863 int __user *, status, int, flags)
1865 return kernel_move_pages(pid, nr_pages, pages, nodes, status, flags);
1868 #ifdef CONFIG_COMPAT
1869 COMPAT_SYSCALL_DEFINE6(move_pages, pid_t, pid, compat_ulong_t, nr_pages,
1870 compat_uptr_t __user *, pages32,
1871 const int __user *, nodes,
1872 int __user *, status,
1875 const void __user * __user *pages;
1878 pages = compat_alloc_user_space(nr_pages * sizeof(void *));
1879 for (i = 0; i < nr_pages; i++) {
1882 if (get_user(p, pages32 + i) ||
1883 put_user(compat_ptr(p), pages + i))
1886 return kernel_move_pages(pid, nr_pages, pages, nodes, status, flags);
1888 #endif /* CONFIG_COMPAT */
1890 #ifdef CONFIG_NUMA_BALANCING
1892 * Returns true if this is a safe migration target node for misplaced NUMA
1893 * pages. Currently it only checks the watermarks which crude
1895 static bool migrate_balanced_pgdat(struct pglist_data *pgdat,
1896 unsigned long nr_migrate_pages)
1900 for (z = pgdat->nr_zones - 1; z >= 0; z--) {
1901 struct zone *zone = pgdat->node_zones + z;
1903 if (!populated_zone(zone))
1906 /* Avoid waking kswapd by allocating pages_to_migrate pages. */
1907 if (!zone_watermark_ok(zone, 0,
1908 high_wmark_pages(zone) +
1917 static struct page *alloc_misplaced_dst_page(struct page *page,
1920 int nid = (int) data;
1921 struct page *newpage;
1923 newpage = __alloc_pages_node(nid,
1924 (GFP_HIGHUSER_MOVABLE |
1925 __GFP_THISNODE | __GFP_NOMEMALLOC |
1926 __GFP_NORETRY | __GFP_NOWARN) &
1932 static int numamigrate_isolate_page(pg_data_t *pgdat, struct page *page)
1936 VM_BUG_ON_PAGE(compound_order(page) && !PageTransHuge(page), page);
1938 /* Avoid migrating to a node that is nearly full */
1939 if (!migrate_balanced_pgdat(pgdat, compound_nr(page)))
1942 if (isolate_lru_page(page))
1946 * migrate_misplaced_transhuge_page() skips page migration's usual
1947 * check on page_count(), so we must do it here, now that the page
1948 * has been isolated: a GUP pin, or any other pin, prevents migration.
1949 * The expected page count is 3: 1 for page's mapcount and 1 for the
1950 * caller's pin and 1 for the reference taken by isolate_lru_page().
1952 if (PageTransHuge(page) && page_count(page) != 3) {
1953 putback_lru_page(page);
1957 page_lru = page_is_file_lru(page);
1958 mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON + page_lru,
1959 hpage_nr_pages(page));
1962 * Isolating the page has taken another reference, so the
1963 * caller's reference can be safely dropped without the page
1964 * disappearing underneath us during migration.
1970 bool pmd_trans_migrating(pmd_t pmd)
1972 struct page *page = pmd_page(pmd);
1973 return PageLocked(page);
1977 * Attempt to migrate a misplaced page to the specified destination
1978 * node. Caller is expected to have an elevated reference count on
1979 * the page that will be dropped by this function before returning.
1981 int migrate_misplaced_page(struct page *page, struct vm_area_struct *vma,
1984 pg_data_t *pgdat = NODE_DATA(node);
1987 LIST_HEAD(migratepages);
1990 * Don't migrate file pages that are mapped in multiple processes
1991 * with execute permissions as they are probably shared libraries.
1993 if (page_mapcount(page) != 1 && page_is_file_lru(page) &&
1994 (vma->vm_flags & VM_EXEC))
1998 * Also do not migrate dirty pages as not all filesystems can move
1999 * dirty pages in MIGRATE_ASYNC mode which is a waste of cycles.
2001 if (page_is_file_lru(page) && PageDirty(page))
2004 isolated = numamigrate_isolate_page(pgdat, page);
2008 list_add(&page->lru, &migratepages);
2009 nr_remaining = migrate_pages(&migratepages, alloc_misplaced_dst_page,
2010 NULL, node, MIGRATE_ASYNC,
2013 if (!list_empty(&migratepages)) {
2014 list_del(&page->lru);
2015 dec_node_page_state(page, NR_ISOLATED_ANON +
2016 page_is_file_lru(page));
2017 putback_lru_page(page);
2021 count_vm_numa_event(NUMA_PAGE_MIGRATE);
2022 BUG_ON(!list_empty(&migratepages));
2029 #endif /* CONFIG_NUMA_BALANCING */
2031 #if defined(CONFIG_NUMA_BALANCING) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
2033 * Migrates a THP to a given target node. page must be locked and is unlocked
2036 int migrate_misplaced_transhuge_page(struct mm_struct *mm,
2037 struct vm_area_struct *vma,
2038 pmd_t *pmd, pmd_t entry,
2039 unsigned long address,
2040 struct page *page, int node)
2043 pg_data_t *pgdat = NODE_DATA(node);
2045 struct page *new_page = NULL;
2046 int page_lru = page_is_file_lru(page);
2047 unsigned long start = address & HPAGE_PMD_MASK;
2049 new_page = alloc_pages_node(node,
2050 (GFP_TRANSHUGE_LIGHT | __GFP_THISNODE),
2054 prep_transhuge_page(new_page);
2056 isolated = numamigrate_isolate_page(pgdat, page);
2062 /* Prepare a page as a migration target */
2063 __SetPageLocked(new_page);
2064 if (PageSwapBacked(page))
2065 __SetPageSwapBacked(new_page);
2067 /* anon mapping, we can simply copy page->mapping to the new page: */
2068 new_page->mapping = page->mapping;
2069 new_page->index = page->index;
2070 /* flush the cache before copying using the kernel virtual address */
2071 flush_cache_range(vma, start, start + HPAGE_PMD_SIZE);
2072 migrate_page_copy(new_page, page);
2073 WARN_ON(PageLRU(new_page));
2075 /* Recheck the target PMD */
2076 ptl = pmd_lock(mm, pmd);
2077 if (unlikely(!pmd_same(*pmd, entry) || !page_ref_freeze(page, 2))) {
2080 /* Reverse changes made by migrate_page_copy() */
2081 if (TestClearPageActive(new_page))
2082 SetPageActive(page);
2083 if (TestClearPageUnevictable(new_page))
2084 SetPageUnevictable(page);
2086 unlock_page(new_page);
2087 put_page(new_page); /* Free it */
2089 /* Retake the callers reference and putback on LRU */
2091 putback_lru_page(page);
2092 mod_node_page_state(page_pgdat(page),
2093 NR_ISOLATED_ANON + page_lru, -HPAGE_PMD_NR);
2098 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
2099 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
2102 * Overwrite the old entry under pagetable lock and establish
2103 * the new PTE. Any parallel GUP will either observe the old
2104 * page blocking on the page lock, block on the page table
2105 * lock or observe the new page. The SetPageUptodate on the
2106 * new page and page_add_new_anon_rmap guarantee the copy is
2107 * visible before the pagetable update.
2109 page_add_anon_rmap(new_page, vma, start, true);
2111 * At this point the pmd is numa/protnone (i.e. non present) and the TLB
2112 * has already been flushed globally. So no TLB can be currently
2113 * caching this non present pmd mapping. There's no need to clear the
2114 * pmd before doing set_pmd_at(), nor to flush the TLB after
2115 * set_pmd_at(). Clearing the pmd here would introduce a race
2116 * condition against MADV_DONTNEED, because MADV_DONTNEED only holds the
2117 * mmap_sem for reading. If the pmd is set to NULL at any given time,
2118 * MADV_DONTNEED won't wait on the pmd lock and it'll skip clearing this
2121 set_pmd_at(mm, start, pmd, entry);
2122 update_mmu_cache_pmd(vma, address, &entry);
2124 page_ref_unfreeze(page, 2);
2125 mlock_migrate_page(new_page, page);
2126 page_remove_rmap(page, true);
2127 set_page_owner_migrate_reason(new_page, MR_NUMA_MISPLACED);
2131 /* Take an "isolate" reference and put new page on the LRU. */
2133 putback_lru_page(new_page);
2135 unlock_page(new_page);
2137 put_page(page); /* Drop the rmap reference */
2138 put_page(page); /* Drop the LRU isolation reference */
2140 count_vm_events(PGMIGRATE_SUCCESS, HPAGE_PMD_NR);
2141 count_vm_numa_events(NUMA_PAGE_MIGRATE, HPAGE_PMD_NR);
2143 mod_node_page_state(page_pgdat(page),
2144 NR_ISOLATED_ANON + page_lru,
2149 count_vm_events(PGMIGRATE_FAIL, HPAGE_PMD_NR);
2150 ptl = pmd_lock(mm, pmd);
2151 if (pmd_same(*pmd, entry)) {
2152 entry = pmd_modify(entry, vma->vm_page_prot);
2153 set_pmd_at(mm, start, pmd, entry);
2154 update_mmu_cache_pmd(vma, address, &entry);
2163 #endif /* CONFIG_NUMA_BALANCING */
2165 #endif /* CONFIG_NUMA */
2167 #ifdef CONFIG_DEVICE_PRIVATE
2168 static int migrate_vma_collect_hole(unsigned long start,
2170 __always_unused int depth,
2171 struct mm_walk *walk)
2173 struct migrate_vma *migrate = walk->private;
2176 for (addr = start; addr < end; addr += PAGE_SIZE) {
2177 migrate->src[migrate->npages] = MIGRATE_PFN_MIGRATE;
2178 migrate->dst[migrate->npages] = 0;
2186 static int migrate_vma_collect_skip(unsigned long start,
2188 struct mm_walk *walk)
2190 struct migrate_vma *migrate = walk->private;
2193 for (addr = start; addr < end; addr += PAGE_SIZE) {
2194 migrate->dst[migrate->npages] = 0;
2195 migrate->src[migrate->npages++] = 0;
2201 static int migrate_vma_collect_pmd(pmd_t *pmdp,
2202 unsigned long start,
2204 struct mm_walk *walk)
2206 struct migrate_vma *migrate = walk->private;
2207 struct vm_area_struct *vma = walk->vma;
2208 struct mm_struct *mm = vma->vm_mm;
2209 unsigned long addr = start, unmapped = 0;
2214 if (pmd_none(*pmdp))
2215 return migrate_vma_collect_hole(start, end, -1, walk);
2217 if (pmd_trans_huge(*pmdp)) {
2220 ptl = pmd_lock(mm, pmdp);
2221 if (unlikely(!pmd_trans_huge(*pmdp))) {
2226 page = pmd_page(*pmdp);
2227 if (is_huge_zero_page(page)) {
2229 split_huge_pmd(vma, pmdp, addr);
2230 if (pmd_trans_unstable(pmdp))
2231 return migrate_vma_collect_skip(start, end,
2238 if (unlikely(!trylock_page(page)))
2239 return migrate_vma_collect_skip(start, end,
2241 ret = split_huge_page(page);
2245 return migrate_vma_collect_skip(start, end,
2247 if (pmd_none(*pmdp))
2248 return migrate_vma_collect_hole(start, end, -1,
2253 if (unlikely(pmd_bad(*pmdp)))
2254 return migrate_vma_collect_skip(start, end, walk);
2256 ptep = pte_offset_map_lock(mm, pmdp, addr, &ptl);
2257 arch_enter_lazy_mmu_mode();
2259 for (; addr < end; addr += PAGE_SIZE, ptep++) {
2260 unsigned long mpfn = 0, pfn;
2267 if (pte_none(pte)) {
2268 mpfn = MIGRATE_PFN_MIGRATE;
2273 if (!pte_present(pte)) {
2275 * Only care about unaddressable device page special
2276 * page table entry. Other special swap entries are not
2277 * migratable, and we ignore regular swapped page.
2279 entry = pte_to_swp_entry(pte);
2280 if (!is_device_private_entry(entry))
2283 page = device_private_entry_to_page(entry);
2284 if (page->pgmap->owner != migrate->src_owner)
2287 mpfn = migrate_pfn(page_to_pfn(page)) |
2288 MIGRATE_PFN_MIGRATE;
2289 if (is_write_device_private_entry(entry))
2290 mpfn |= MIGRATE_PFN_WRITE;
2292 if (migrate->src_owner)
2295 if (is_zero_pfn(pfn)) {
2296 mpfn = MIGRATE_PFN_MIGRATE;
2300 page = vm_normal_page(migrate->vma, addr, pte);
2301 mpfn = migrate_pfn(pfn) | MIGRATE_PFN_MIGRATE;
2302 mpfn |= pte_write(pte) ? MIGRATE_PFN_WRITE : 0;
2305 /* FIXME support THP */
2306 if (!page || !page->mapping || PageTransCompound(page)) {
2312 * By getting a reference on the page we pin it and that blocks
2313 * any kind of migration. Side effect is that it "freezes" the
2316 * We drop this reference after isolating the page from the lru
2317 * for non device page (device page are not on the lru and thus
2318 * can't be dropped from it).
2324 * Optimize for the common case where page is only mapped once
2325 * in one process. If we can lock the page, then we can safely
2326 * set up a special migration page table entry now.
2328 if (trylock_page(page)) {
2331 mpfn |= MIGRATE_PFN_LOCKED;
2332 ptep_get_and_clear(mm, addr, ptep);
2334 /* Setup special migration page table entry */
2335 entry = make_migration_entry(page, mpfn &
2337 swp_pte = swp_entry_to_pte(entry);
2338 if (pte_soft_dirty(pte))
2339 swp_pte = pte_swp_mksoft_dirty(swp_pte);
2340 if (pte_uffd_wp(pte))
2341 swp_pte = pte_swp_mkuffd_wp(swp_pte);
2342 set_pte_at(mm, addr, ptep, swp_pte);
2345 * This is like regular unmap: we remove the rmap and
2346 * drop page refcount. Page won't be freed, as we took
2347 * a reference just above.
2349 page_remove_rmap(page, false);
2352 if (pte_present(pte))
2357 migrate->dst[migrate->npages] = 0;
2358 migrate->src[migrate->npages++] = mpfn;
2360 arch_leave_lazy_mmu_mode();
2361 pte_unmap_unlock(ptep - 1, ptl);
2363 /* Only flush the TLB if we actually modified any entries */
2365 flush_tlb_range(walk->vma, start, end);
2370 static const struct mm_walk_ops migrate_vma_walk_ops = {
2371 .pmd_entry = migrate_vma_collect_pmd,
2372 .pte_hole = migrate_vma_collect_hole,
2376 * migrate_vma_collect() - collect pages over a range of virtual addresses
2377 * @migrate: migrate struct containing all migration information
2379 * This will walk the CPU page table. For each virtual address backed by a
2380 * valid page, it updates the src array and takes a reference on the page, in
2381 * order to pin the page until we lock it and unmap it.
2383 static void migrate_vma_collect(struct migrate_vma *migrate)
2385 struct mmu_notifier_range range;
2387 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, NULL,
2388 migrate->vma->vm_mm, migrate->start, migrate->end);
2389 mmu_notifier_invalidate_range_start(&range);
2391 walk_page_range(migrate->vma->vm_mm, migrate->start, migrate->end,
2392 &migrate_vma_walk_ops, migrate);
2394 mmu_notifier_invalidate_range_end(&range);
2395 migrate->end = migrate->start + (migrate->npages << PAGE_SHIFT);
2399 * migrate_vma_check_page() - check if page is pinned or not
2400 * @page: struct page to check
2402 * Pinned pages cannot be migrated. This is the same test as in
2403 * migrate_page_move_mapping(), except that here we allow migration of a
2406 static bool migrate_vma_check_page(struct page *page)
2409 * One extra ref because caller holds an extra reference, either from
2410 * isolate_lru_page() for a regular page, or migrate_vma_collect() for
2416 * FIXME support THP (transparent huge page), it is bit more complex to
2417 * check them than regular pages, because they can be mapped with a pmd
2418 * or with a pte (split pte mapping).
2420 if (PageCompound(page))
2423 /* Page from ZONE_DEVICE have one extra reference */
2424 if (is_zone_device_page(page)) {
2426 * Private page can never be pin as they have no valid pte and
2427 * GUP will fail for those. Yet if there is a pending migration
2428 * a thread might try to wait on the pte migration entry and
2429 * will bump the page reference count. Sadly there is no way to
2430 * differentiate a regular pin from migration wait. Hence to
2431 * avoid 2 racing thread trying to migrate back to CPU to enter
2432 * infinite loop (one stoping migration because the other is
2433 * waiting on pte migration entry). We always return true here.
2435 * FIXME proper solution is to rework migration_entry_wait() so
2436 * it does not need to take a reference on page.
2438 return is_device_private_page(page);
2441 /* For file back page */
2442 if (page_mapping(page))
2443 extra += 1 + page_has_private(page);
2445 if ((page_count(page) - extra) > page_mapcount(page))
2452 * migrate_vma_prepare() - lock pages and isolate them from the lru
2453 * @migrate: migrate struct containing all migration information
2455 * This locks pages that have been collected by migrate_vma_collect(). Once each
2456 * page is locked it is isolated from the lru (for non-device pages). Finally,
2457 * the ref taken by migrate_vma_collect() is dropped, as locked pages cannot be
2458 * migrated by concurrent kernel threads.
2460 static void migrate_vma_prepare(struct migrate_vma *migrate)
2462 const unsigned long npages = migrate->npages;
2463 const unsigned long start = migrate->start;
2464 unsigned long addr, i, restore = 0;
2465 bool allow_drain = true;
2469 for (i = 0; (i < npages) && migrate->cpages; i++) {
2470 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2476 if (!(migrate->src[i] & MIGRATE_PFN_LOCKED)) {
2478 * Because we are migrating several pages there can be
2479 * a deadlock between 2 concurrent migration where each
2480 * are waiting on each other page lock.
2482 * Make migrate_vma() a best effort thing and backoff
2483 * for any page we can not lock right away.
2485 if (!trylock_page(page)) {
2486 migrate->src[i] = 0;
2492 migrate->src[i] |= MIGRATE_PFN_LOCKED;
2495 /* ZONE_DEVICE pages are not on LRU */
2496 if (!is_zone_device_page(page)) {
2497 if (!PageLRU(page) && allow_drain) {
2498 /* Drain CPU's pagevec */
2499 lru_add_drain_all();
2500 allow_drain = false;
2503 if (isolate_lru_page(page)) {
2505 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2509 migrate->src[i] = 0;
2517 /* Drop the reference we took in collect */
2521 if (!migrate_vma_check_page(page)) {
2523 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2527 if (!is_zone_device_page(page)) {
2529 putback_lru_page(page);
2532 migrate->src[i] = 0;
2536 if (!is_zone_device_page(page))
2537 putback_lru_page(page);
2544 for (i = 0, addr = start; i < npages && restore; i++, addr += PAGE_SIZE) {
2545 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2547 if (!page || (migrate->src[i] & MIGRATE_PFN_MIGRATE))
2550 remove_migration_pte(page, migrate->vma, addr, page);
2552 migrate->src[i] = 0;
2560 * migrate_vma_unmap() - replace page mapping with special migration pte entry
2561 * @migrate: migrate struct containing all migration information
2563 * Replace page mapping (CPU page table pte) with a special migration pte entry
2564 * and check again if it has been pinned. Pinned pages are restored because we
2565 * cannot migrate them.
2567 * This is the last step before we call the device driver callback to allocate
2568 * destination memory and copy contents of original page over to new page.
2570 static void migrate_vma_unmap(struct migrate_vma *migrate)
2572 int flags = TTU_MIGRATION | TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS;
2573 const unsigned long npages = migrate->npages;
2574 const unsigned long start = migrate->start;
2575 unsigned long addr, i, restore = 0;
2577 for (i = 0; i < npages; i++) {
2578 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2580 if (!page || !(migrate->src[i] & MIGRATE_PFN_MIGRATE))
2583 if (page_mapped(page)) {
2584 try_to_unmap(page, flags);
2585 if (page_mapped(page))
2589 if (migrate_vma_check_page(page))
2593 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2598 for (addr = start, i = 0; i < npages && restore; addr += PAGE_SIZE, i++) {
2599 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2601 if (!page || (migrate->src[i] & MIGRATE_PFN_MIGRATE))
2604 remove_migration_ptes(page, page, false);
2606 migrate->src[i] = 0;
2610 if (is_zone_device_page(page))
2613 putback_lru_page(page);
2618 * migrate_vma_setup() - prepare to migrate a range of memory
2619 * @args: contains the vma, start, and and pfns arrays for the migration
2621 * Returns: negative errno on failures, 0 when 0 or more pages were migrated
2624 * Prepare to migrate a range of memory virtual address range by collecting all
2625 * the pages backing each virtual address in the range, saving them inside the
2626 * src array. Then lock those pages and unmap them. Once the pages are locked
2627 * and unmapped, check whether each page is pinned or not. Pages that aren't
2628 * pinned have the MIGRATE_PFN_MIGRATE flag set (by this function) in the
2629 * corresponding src array entry. Then restores any pages that are pinned, by
2630 * remapping and unlocking those pages.
2632 * The caller should then allocate destination memory and copy source memory to
2633 * it for all those entries (ie with MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE
2634 * flag set). Once these are allocated and copied, the caller must update each
2635 * corresponding entry in the dst array with the pfn value of the destination
2636 * page and with the MIGRATE_PFN_VALID and MIGRATE_PFN_LOCKED flags set
2637 * (destination pages must have their struct pages locked, via lock_page()).
2639 * Note that the caller does not have to migrate all the pages that are marked
2640 * with MIGRATE_PFN_MIGRATE flag in src array unless this is a migration from
2641 * device memory to system memory. If the caller cannot migrate a device page
2642 * back to system memory, then it must return VM_FAULT_SIGBUS, which has severe
2643 * consequences for the userspace process, so it must be avoided if at all
2646 * For empty entries inside CPU page table (pte_none() or pmd_none() is true) we
2647 * do set MIGRATE_PFN_MIGRATE flag inside the corresponding source array thus
2648 * allowing the caller to allocate device memory for those unback virtual
2649 * address. For this the caller simply has to allocate device memory and
2650 * properly set the destination entry like for regular migration. Note that
2651 * this can still fails and thus inside the device driver must check if the
2652 * migration was successful for those entries after calling migrate_vma_pages()
2653 * just like for regular migration.
2655 * After that, the callers must call migrate_vma_pages() to go over each entry
2656 * in the src array that has the MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE flag
2657 * set. If the corresponding entry in dst array has MIGRATE_PFN_VALID flag set,
2658 * then migrate_vma_pages() to migrate struct page information from the source
2659 * struct page to the destination struct page. If it fails to migrate the
2660 * struct page information, then it clears the MIGRATE_PFN_MIGRATE flag in the
2663 * At this point all successfully migrated pages have an entry in the src
2664 * array with MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE flag set and the dst
2665 * array entry with MIGRATE_PFN_VALID flag set.
2667 * Once migrate_vma_pages() returns the caller may inspect which pages were
2668 * successfully migrated, and which were not. Successfully migrated pages will
2669 * have the MIGRATE_PFN_MIGRATE flag set for their src array entry.
2671 * It is safe to update device page table after migrate_vma_pages() because
2672 * both destination and source page are still locked, and the mmap_sem is held
2673 * in read mode (hence no one can unmap the range being migrated).
2675 * Once the caller is done cleaning up things and updating its page table (if it
2676 * chose to do so, this is not an obligation) it finally calls
2677 * migrate_vma_finalize() to update the CPU page table to point to new pages
2678 * for successfully migrated pages or otherwise restore the CPU page table to
2679 * point to the original source pages.
2681 int migrate_vma_setup(struct migrate_vma *args)
2683 long nr_pages = (args->end - args->start) >> PAGE_SHIFT;
2685 args->start &= PAGE_MASK;
2686 args->end &= PAGE_MASK;
2687 if (!args->vma || is_vm_hugetlb_page(args->vma) ||
2688 (args->vma->vm_flags & VM_SPECIAL) || vma_is_dax(args->vma))
2692 if (args->start < args->vma->vm_start ||
2693 args->start >= args->vma->vm_end)
2695 if (args->end <= args->vma->vm_start || args->end > args->vma->vm_end)
2697 if (!args->src || !args->dst)
2700 memset(args->src, 0, sizeof(*args->src) * nr_pages);
2704 migrate_vma_collect(args);
2707 migrate_vma_prepare(args);
2709 migrate_vma_unmap(args);
2712 * At this point pages are locked and unmapped, and thus they have
2713 * stable content and can safely be copied to destination memory that
2714 * is allocated by the drivers.
2719 EXPORT_SYMBOL(migrate_vma_setup);
2722 * This code closely matches the code in:
2723 * __handle_mm_fault()
2724 * handle_pte_fault()
2725 * do_anonymous_page()
2726 * to map in an anonymous zero page but the struct page will be a ZONE_DEVICE
2729 static void migrate_vma_insert_page(struct migrate_vma *migrate,
2735 struct vm_area_struct *vma = migrate->vma;
2736 struct mm_struct *mm = vma->vm_mm;
2737 struct mem_cgroup *memcg;
2747 /* Only allow populating anonymous memory */
2748 if (!vma_is_anonymous(vma))
2751 pgdp = pgd_offset(mm, addr);
2752 p4dp = p4d_alloc(mm, pgdp, addr);
2755 pudp = pud_alloc(mm, p4dp, addr);
2758 pmdp = pmd_alloc(mm, pudp, addr);
2762 if (pmd_trans_huge(*pmdp) || pmd_devmap(*pmdp))
2766 * Use pte_alloc() instead of pte_alloc_map(). We can't run
2767 * pte_offset_map() on pmds where a huge pmd might be created
2768 * from a different thread.
2770 * pte_alloc_map() is safe to use under down_write(mmap_sem) or when
2771 * parallel threads are excluded by other means.
2773 * Here we only have down_read(mmap_sem).
2775 if (pte_alloc(mm, pmdp))
2778 /* See the comment in pte_alloc_one_map() */
2779 if (unlikely(pmd_trans_unstable(pmdp)))
2782 if (unlikely(anon_vma_prepare(vma)))
2784 if (mem_cgroup_try_charge(page, vma->vm_mm, GFP_KERNEL, &memcg, false))
2788 * The memory barrier inside __SetPageUptodate makes sure that
2789 * preceding stores to the page contents become visible before
2790 * the set_pte_at() write.
2792 __SetPageUptodate(page);
2794 if (is_zone_device_page(page)) {
2795 if (is_device_private_page(page)) {
2796 swp_entry_t swp_entry;
2798 swp_entry = make_device_private_entry(page, vma->vm_flags & VM_WRITE);
2799 entry = swp_entry_to_pte(swp_entry);
2802 entry = mk_pte(page, vma->vm_page_prot);
2803 if (vma->vm_flags & VM_WRITE)
2804 entry = pte_mkwrite(pte_mkdirty(entry));
2807 ptep = pte_offset_map_lock(mm, pmdp, addr, &ptl);
2809 if (check_stable_address_space(mm))
2812 if (pte_present(*ptep)) {
2813 unsigned long pfn = pte_pfn(*ptep);
2815 if (!is_zero_pfn(pfn))
2818 } else if (!pte_none(*ptep))
2822 * Check for userfaultfd but do not deliver the fault. Instead,
2825 if (userfaultfd_missing(vma))
2828 inc_mm_counter(mm, MM_ANONPAGES);
2829 page_add_new_anon_rmap(page, vma, addr, false);
2830 mem_cgroup_commit_charge(page, memcg, false, false);
2831 if (!is_zone_device_page(page))
2832 lru_cache_add_active_or_unevictable(page, vma);
2836 flush_cache_page(vma, addr, pte_pfn(*ptep));
2837 ptep_clear_flush_notify(vma, addr, ptep);
2838 set_pte_at_notify(mm, addr, ptep, entry);
2839 update_mmu_cache(vma, addr, ptep);
2841 /* No need to invalidate - it was non-present before */
2842 set_pte_at(mm, addr, ptep, entry);
2843 update_mmu_cache(vma, addr, ptep);
2846 pte_unmap_unlock(ptep, ptl);
2847 *src = MIGRATE_PFN_MIGRATE;
2851 pte_unmap_unlock(ptep, ptl);
2852 mem_cgroup_cancel_charge(page, memcg, false);
2854 *src &= ~MIGRATE_PFN_MIGRATE;
2858 * migrate_vma_pages() - migrate meta-data from src page to dst page
2859 * @migrate: migrate struct containing all migration information
2861 * This migrates struct page meta-data from source struct page to destination
2862 * struct page. This effectively finishes the migration from source page to the
2865 void migrate_vma_pages(struct migrate_vma *migrate)
2867 const unsigned long npages = migrate->npages;
2868 const unsigned long start = migrate->start;
2869 struct mmu_notifier_range range;
2870 unsigned long addr, i;
2871 bool notified = false;
2873 for (i = 0, addr = start; i < npages; addr += PAGE_SIZE, i++) {
2874 struct page *newpage = migrate_pfn_to_page(migrate->dst[i]);
2875 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2876 struct address_space *mapping;
2880 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2885 if (!(migrate->src[i] & MIGRATE_PFN_MIGRATE))
2890 mmu_notifier_range_init(&range,
2891 MMU_NOTIFY_CLEAR, 0,
2893 migrate->vma->vm_mm,
2894 addr, migrate->end);
2895 mmu_notifier_invalidate_range_start(&range);
2897 migrate_vma_insert_page(migrate, addr, newpage,
2903 mapping = page_mapping(page);
2905 if (is_zone_device_page(newpage)) {
2906 if (is_device_private_page(newpage)) {
2908 * For now only support private anonymous when
2909 * migrating to un-addressable device memory.
2912 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2917 * Other types of ZONE_DEVICE page are not
2920 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2925 r = migrate_page(mapping, newpage, page, MIGRATE_SYNC_NO_COPY);
2926 if (r != MIGRATEPAGE_SUCCESS)
2927 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2931 * No need to double call mmu_notifier->invalidate_range() callback as
2932 * the above ptep_clear_flush_notify() inside migrate_vma_insert_page()
2933 * did already call it.
2936 mmu_notifier_invalidate_range_only_end(&range);
2938 EXPORT_SYMBOL(migrate_vma_pages);
2941 * migrate_vma_finalize() - restore CPU page table entry
2942 * @migrate: migrate struct containing all migration information
2944 * This replaces the special migration pte entry with either a mapping to the
2945 * new page if migration was successful for that page, or to the original page
2948 * This also unlocks the pages and puts them back on the lru, or drops the extra
2949 * refcount, for device pages.
2951 void migrate_vma_finalize(struct migrate_vma *migrate)
2953 const unsigned long npages = migrate->npages;
2956 for (i = 0; i < npages; i++) {
2957 struct page *newpage = migrate_pfn_to_page(migrate->dst[i]);
2958 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2962 unlock_page(newpage);
2968 if (!(migrate->src[i] & MIGRATE_PFN_MIGRATE) || !newpage) {
2970 unlock_page(newpage);
2976 remove_migration_ptes(page, newpage, false);
2980 if (is_zone_device_page(page))
2983 putback_lru_page(page);
2985 if (newpage != page) {
2986 unlock_page(newpage);
2987 if (is_zone_device_page(newpage))
2990 putback_lru_page(newpage);
2994 EXPORT_SYMBOL(migrate_vma_finalize);
2995 #endif /* CONFIG_DEVICE_PRIVATE */