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), -thp_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_device_private_page(new))) {
250 entry = make_device_private_entry(new, pte_write(pte));
251 pte = swp_entry_to_pte(entry);
252 if (pte_swp_soft_dirty(*pvmw.pte))
253 pte = pte_swp_mksoft_dirty(pte);
254 if (pte_swp_uffd_wp(*pvmw.pte))
255 pte = pte_swp_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 private pages have an extra refcount as they are
387 expected_count += is_device_private_page(page);
389 expected_count += thp_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;
409 int nr = thp_nr_pages(page);
412 /* Anonymous page without mapping */
413 if (page_count(page) != expected_count)
416 /* No turning back from here */
417 newpage->index = page->index;
418 newpage->mapping = page->mapping;
419 if (PageSwapBacked(page))
420 __SetPageSwapBacked(newpage);
422 return MIGRATEPAGE_SUCCESS;
425 oldzone = page_zone(page);
426 newzone = page_zone(newpage);
429 if (page_count(page) != expected_count || xas_load(&xas) != page) {
430 xas_unlock_irq(&xas);
434 if (!page_ref_freeze(page, expected_count)) {
435 xas_unlock_irq(&xas);
440 * Now we know that no one else is looking at the page:
441 * no turning back from here.
443 newpage->index = page->index;
444 newpage->mapping = page->mapping;
445 page_ref_add(newpage, nr); /* add cache reference */
446 if (PageSwapBacked(page)) {
447 __SetPageSwapBacked(newpage);
448 if (PageSwapCache(page)) {
449 SetPageSwapCache(newpage);
450 set_page_private(newpage, page_private(page));
453 VM_BUG_ON_PAGE(PageSwapCache(page), page);
456 /* Move dirty while page refs frozen and newpage not yet exposed */
457 dirty = PageDirty(page);
459 ClearPageDirty(page);
460 SetPageDirty(newpage);
463 xas_store(&xas, newpage);
464 if (PageTransHuge(page)) {
467 for (i = 1; i < nr; i++) {
469 xas_store(&xas, newpage);
474 * Drop cache reference from old page by unfreezing
475 * to one less reference.
476 * We know this isn't the last reference.
478 page_ref_unfreeze(page, expected_count - nr);
481 /* Leave irq disabled to prevent preemption while updating stats */
484 * If moved to a different zone then also account
485 * the page for that zone. Other VM counters will be
486 * taken care of when we establish references to the
487 * new page and drop references to the old page.
489 * Note that anonymous pages are accounted for
490 * via NR_FILE_PAGES and NR_ANON_MAPPED if they
491 * are mapped to swap space.
493 if (newzone != oldzone) {
494 struct lruvec *old_lruvec, *new_lruvec;
495 struct mem_cgroup *memcg;
497 memcg = page_memcg(page);
498 old_lruvec = mem_cgroup_lruvec(memcg, oldzone->zone_pgdat);
499 new_lruvec = mem_cgroup_lruvec(memcg, newzone->zone_pgdat);
501 __mod_lruvec_state(old_lruvec, NR_FILE_PAGES, -nr);
502 __mod_lruvec_state(new_lruvec, NR_FILE_PAGES, nr);
503 if (PageSwapBacked(page) && !PageSwapCache(page)) {
504 __mod_lruvec_state(old_lruvec, NR_SHMEM, -nr);
505 __mod_lruvec_state(new_lruvec, NR_SHMEM, nr);
507 if (dirty && mapping_can_writeback(mapping)) {
508 __mod_lruvec_state(old_lruvec, NR_FILE_DIRTY, -nr);
509 __mod_zone_page_state(oldzone, NR_ZONE_WRITE_PENDING, -nr);
510 __mod_lruvec_state(new_lruvec, NR_FILE_DIRTY, nr);
511 __mod_zone_page_state(newzone, NR_ZONE_WRITE_PENDING, nr);
516 return MIGRATEPAGE_SUCCESS;
518 EXPORT_SYMBOL(migrate_page_move_mapping);
521 * The expected number of remaining references is the same as that
522 * of migrate_page_move_mapping().
524 int migrate_huge_page_move_mapping(struct address_space *mapping,
525 struct page *newpage, struct page *page)
527 XA_STATE(xas, &mapping->i_pages, page_index(page));
531 expected_count = 2 + page_has_private(page);
532 if (page_count(page) != expected_count || xas_load(&xas) != page) {
533 xas_unlock_irq(&xas);
537 if (!page_ref_freeze(page, expected_count)) {
538 xas_unlock_irq(&xas);
542 newpage->index = page->index;
543 newpage->mapping = page->mapping;
547 xas_store(&xas, newpage);
549 page_ref_unfreeze(page, expected_count - 1);
551 xas_unlock_irq(&xas);
553 return MIGRATEPAGE_SUCCESS;
557 * Gigantic pages are so large that we do not guarantee that page++ pointer
558 * arithmetic will work across the entire page. We need something more
561 static void __copy_gigantic_page(struct page *dst, struct page *src,
565 struct page *dst_base = dst;
566 struct page *src_base = src;
568 for (i = 0; i < nr_pages; ) {
570 copy_highpage(dst, src);
573 dst = mem_map_next(dst, dst_base, i);
574 src = mem_map_next(src, src_base, i);
578 static void copy_huge_page(struct page *dst, struct page *src)
585 struct hstate *h = page_hstate(src);
586 nr_pages = pages_per_huge_page(h);
588 if (unlikely(nr_pages > MAX_ORDER_NR_PAGES)) {
589 __copy_gigantic_page(dst, src, nr_pages);
594 BUG_ON(!PageTransHuge(src));
595 nr_pages = thp_nr_pages(src);
598 for (i = 0; i < nr_pages; i++) {
600 copy_highpage(dst + i, src + i);
605 * Copy the page to its new location
607 void migrate_page_states(struct page *newpage, struct page *page)
612 SetPageError(newpage);
613 if (PageReferenced(page))
614 SetPageReferenced(newpage);
615 if (PageUptodate(page))
616 SetPageUptodate(newpage);
617 if (TestClearPageActive(page)) {
618 VM_BUG_ON_PAGE(PageUnevictable(page), page);
619 SetPageActive(newpage);
620 } else if (TestClearPageUnevictable(page))
621 SetPageUnevictable(newpage);
622 if (PageWorkingset(page))
623 SetPageWorkingset(newpage);
624 if (PageChecked(page))
625 SetPageChecked(newpage);
626 if (PageMappedToDisk(page))
627 SetPageMappedToDisk(newpage);
629 /* Move dirty on pages not done by migrate_page_move_mapping() */
631 SetPageDirty(newpage);
633 if (page_is_young(page))
634 set_page_young(newpage);
635 if (page_is_idle(page))
636 set_page_idle(newpage);
639 * Copy NUMA information to the new page, to prevent over-eager
640 * future migrations of this same page.
642 cpupid = page_cpupid_xchg_last(page, -1);
643 page_cpupid_xchg_last(newpage, cpupid);
645 ksm_migrate_page(newpage, page);
647 * Please do not reorder this without considering how mm/ksm.c's
648 * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache().
650 if (PageSwapCache(page))
651 ClearPageSwapCache(page);
652 ClearPagePrivate(page);
653 set_page_private(page, 0);
656 * If any waiters have accumulated on the new page then
659 if (PageWriteback(newpage))
660 end_page_writeback(newpage);
663 * PG_readahead shares the same bit with PG_reclaim. The above
664 * end_page_writeback() may clear PG_readahead mistakenly, so set the
667 if (PageReadahead(page))
668 SetPageReadahead(newpage);
670 copy_page_owner(page, newpage);
673 mem_cgroup_migrate(page, newpage);
675 EXPORT_SYMBOL(migrate_page_states);
677 void migrate_page_copy(struct page *newpage, struct page *page)
679 if (PageHuge(page) || PageTransHuge(page))
680 copy_huge_page(newpage, page);
682 copy_highpage(newpage, page);
684 migrate_page_states(newpage, page);
686 EXPORT_SYMBOL(migrate_page_copy);
688 /************************************************************
689 * Migration functions
690 ***********************************************************/
693 * Common logic to directly migrate a single LRU page suitable for
694 * pages that do not use PagePrivate/PagePrivate2.
696 * Pages are locked upon entry and exit.
698 int migrate_page(struct address_space *mapping,
699 struct page *newpage, struct page *page,
700 enum migrate_mode mode)
704 BUG_ON(PageWriteback(page)); /* Writeback must be complete */
706 rc = migrate_page_move_mapping(mapping, newpage, page, 0);
708 if (rc != MIGRATEPAGE_SUCCESS)
711 if (mode != MIGRATE_SYNC_NO_COPY)
712 migrate_page_copy(newpage, page);
714 migrate_page_states(newpage, page);
715 return MIGRATEPAGE_SUCCESS;
717 EXPORT_SYMBOL(migrate_page);
720 /* Returns true if all buffers are successfully locked */
721 static bool buffer_migrate_lock_buffers(struct buffer_head *head,
722 enum migrate_mode mode)
724 struct buffer_head *bh = head;
726 /* Simple case, sync compaction */
727 if (mode != MIGRATE_ASYNC) {
730 bh = bh->b_this_page;
732 } while (bh != head);
737 /* async case, we cannot block on lock_buffer so use trylock_buffer */
739 if (!trylock_buffer(bh)) {
741 * We failed to lock the buffer and cannot stall in
742 * async migration. Release the taken locks
744 struct buffer_head *failed_bh = bh;
746 while (bh != failed_bh) {
748 bh = bh->b_this_page;
753 bh = bh->b_this_page;
754 } while (bh != head);
758 static int __buffer_migrate_page(struct address_space *mapping,
759 struct page *newpage, struct page *page, enum migrate_mode mode,
762 struct buffer_head *bh, *head;
766 if (!page_has_buffers(page))
767 return migrate_page(mapping, newpage, page, mode);
769 /* Check whether page does not have extra refs before we do more work */
770 expected_count = expected_page_refs(mapping, page);
771 if (page_count(page) != expected_count)
774 head = page_buffers(page);
775 if (!buffer_migrate_lock_buffers(head, mode))
780 bool invalidated = false;
784 spin_lock(&mapping->private_lock);
787 if (atomic_read(&bh->b_count)) {
791 bh = bh->b_this_page;
792 } while (bh != head);
798 spin_unlock(&mapping->private_lock);
799 invalidate_bh_lrus();
801 goto recheck_buffers;
805 rc = migrate_page_move_mapping(mapping, newpage, page, 0);
806 if (rc != MIGRATEPAGE_SUCCESS)
809 attach_page_private(newpage, detach_page_private(page));
813 set_bh_page(bh, newpage, bh_offset(bh));
814 bh = bh->b_this_page;
816 } while (bh != head);
818 if (mode != MIGRATE_SYNC_NO_COPY)
819 migrate_page_copy(newpage, page);
821 migrate_page_states(newpage, page);
823 rc = MIGRATEPAGE_SUCCESS;
826 spin_unlock(&mapping->private_lock);
830 bh = bh->b_this_page;
832 } while (bh != head);
838 * Migration function for pages with buffers. This function can only be used
839 * if the underlying filesystem guarantees that no other references to "page"
840 * exist. For example attached buffer heads are accessed only under page lock.
842 int buffer_migrate_page(struct address_space *mapping,
843 struct page *newpage, struct page *page, enum migrate_mode mode)
845 return __buffer_migrate_page(mapping, newpage, page, mode, false);
847 EXPORT_SYMBOL(buffer_migrate_page);
850 * Same as above except that this variant is more careful and checks that there
851 * are also no buffer head references. This function is the right one for
852 * mappings where buffer heads are directly looked up and referenced (such as
853 * block device mappings).
855 int buffer_migrate_page_norefs(struct address_space *mapping,
856 struct page *newpage, struct page *page, enum migrate_mode mode)
858 return __buffer_migrate_page(mapping, newpage, page, mode, true);
863 * Writeback a page to clean the dirty state
865 static int writeout(struct address_space *mapping, struct page *page)
867 struct writeback_control wbc = {
868 .sync_mode = WB_SYNC_NONE,
871 .range_end = LLONG_MAX,
876 if (!mapping->a_ops->writepage)
877 /* No write method for the address space */
880 if (!clear_page_dirty_for_io(page))
881 /* Someone else already triggered a write */
885 * A dirty page may imply that the underlying filesystem has
886 * the page on some queue. So the page must be clean for
887 * migration. Writeout may mean we loose the lock and the
888 * page state is no longer what we checked for earlier.
889 * At this point we know that the migration attempt cannot
892 remove_migration_ptes(page, page, false);
894 rc = mapping->a_ops->writepage(page, &wbc);
896 if (rc != AOP_WRITEPAGE_ACTIVATE)
897 /* unlocked. Relock */
900 return (rc < 0) ? -EIO : -EAGAIN;
904 * Default handling if a filesystem does not provide a migration function.
906 static int fallback_migrate_page(struct address_space *mapping,
907 struct page *newpage, struct page *page, enum migrate_mode mode)
909 if (PageDirty(page)) {
910 /* Only writeback pages in full synchronous migration */
913 case MIGRATE_SYNC_NO_COPY:
918 return writeout(mapping, page);
922 * Buffers may be managed in a filesystem specific way.
923 * We must have no buffers or drop them.
925 if (page_has_private(page) &&
926 !try_to_release_page(page, GFP_KERNEL))
927 return mode == MIGRATE_SYNC ? -EAGAIN : -EBUSY;
929 return migrate_page(mapping, newpage, page, mode);
933 * Move a page to a newly allocated page
934 * The page is locked and all ptes have been successfully removed.
936 * The new page will have replaced the old page if this function
941 * MIGRATEPAGE_SUCCESS - success
943 static int move_to_new_page(struct page *newpage, struct page *page,
944 enum migrate_mode mode)
946 struct address_space *mapping;
948 bool is_lru = !__PageMovable(page);
950 VM_BUG_ON_PAGE(!PageLocked(page), page);
951 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
953 mapping = page_mapping(page);
955 if (likely(is_lru)) {
957 rc = migrate_page(mapping, newpage, page, mode);
958 else if (mapping->a_ops->migratepage)
960 * Most pages have a mapping and most filesystems
961 * provide a migratepage callback. Anonymous pages
962 * are part of swap space which also has its own
963 * migratepage callback. This is the most common path
964 * for page migration.
966 rc = mapping->a_ops->migratepage(mapping, newpage,
969 rc = fallback_migrate_page(mapping, newpage,
973 * In case of non-lru page, it could be released after
974 * isolation step. In that case, we shouldn't try migration.
976 VM_BUG_ON_PAGE(!PageIsolated(page), page);
977 if (!PageMovable(page)) {
978 rc = MIGRATEPAGE_SUCCESS;
979 __ClearPageIsolated(page);
983 rc = mapping->a_ops->migratepage(mapping, newpage,
985 WARN_ON_ONCE(rc == MIGRATEPAGE_SUCCESS &&
986 !PageIsolated(page));
990 * When successful, old pagecache page->mapping must be cleared before
991 * page is freed; but stats require that PageAnon be left as PageAnon.
993 if (rc == MIGRATEPAGE_SUCCESS) {
994 if (__PageMovable(page)) {
995 VM_BUG_ON_PAGE(!PageIsolated(page), page);
998 * We clear PG_movable under page_lock so any compactor
999 * cannot try to migrate this page.
1001 __ClearPageIsolated(page);
1005 * Anonymous and movable page->mapping will be cleared by
1006 * free_pages_prepare so don't reset it here for keeping
1007 * the type to work PageAnon, for example.
1009 if (!PageMappingFlags(page))
1010 page->mapping = NULL;
1012 if (likely(!is_zone_device_page(newpage)))
1013 flush_dcache_page(newpage);
1020 static int __unmap_and_move(struct page *page, struct page *newpage,
1021 int force, enum migrate_mode mode)
1024 int page_was_mapped = 0;
1025 struct anon_vma *anon_vma = NULL;
1026 bool is_lru = !__PageMovable(page);
1028 if (!trylock_page(page)) {
1029 if (!force || mode == MIGRATE_ASYNC)
1033 * It's not safe for direct compaction to call lock_page.
1034 * For example, during page readahead pages are added locked
1035 * to the LRU. Later, when the IO completes the pages are
1036 * marked uptodate and unlocked. However, the queueing
1037 * could be merging multiple pages for one bio (e.g.
1038 * mpage_readahead). If an allocation happens for the
1039 * second or third page, the process can end up locking
1040 * the same page twice and deadlocking. Rather than
1041 * trying to be clever about what pages can be locked,
1042 * avoid the use of lock_page for direct compaction
1045 if (current->flags & PF_MEMALLOC)
1051 if (PageWriteback(page)) {
1053 * Only in the case of a full synchronous migration is it
1054 * necessary to wait for PageWriteback. In the async case,
1055 * the retry loop is too short and in the sync-light case,
1056 * the overhead of stalling is too much
1060 case MIGRATE_SYNC_NO_COPY:
1068 wait_on_page_writeback(page);
1072 * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
1073 * we cannot notice that anon_vma is freed while we migrates a page.
1074 * This get_anon_vma() delays freeing anon_vma pointer until the end
1075 * of migration. File cache pages are no problem because of page_lock()
1076 * File Caches may use write_page() or lock_page() in migration, then,
1077 * just care Anon page here.
1079 * Only page_get_anon_vma() understands the subtleties of
1080 * getting a hold on an anon_vma from outside one of its mms.
1081 * But if we cannot get anon_vma, then we won't need it anyway,
1082 * because that implies that the anon page is no longer mapped
1083 * (and cannot be remapped so long as we hold the page lock).
1085 if (PageAnon(page) && !PageKsm(page))
1086 anon_vma = page_get_anon_vma(page);
1089 * Block others from accessing the new page when we get around to
1090 * establishing additional references. We are usually the only one
1091 * holding a reference to newpage at this point. We used to have a BUG
1092 * here if trylock_page(newpage) fails, but would like to allow for
1093 * cases where there might be a race with the previous use of newpage.
1094 * This is much like races on refcount of oldpage: just don't BUG().
1096 if (unlikely(!trylock_page(newpage)))
1099 if (unlikely(!is_lru)) {
1100 rc = move_to_new_page(newpage, page, mode);
1101 goto out_unlock_both;
1105 * Corner case handling:
1106 * 1. When a new swap-cache page is read into, it is added to the LRU
1107 * and treated as swapcache but it has no rmap yet.
1108 * Calling try_to_unmap() against a page->mapping==NULL page will
1109 * trigger a BUG. So handle it here.
1110 * 2. An orphaned page (see truncate_complete_page) might have
1111 * fs-private metadata. The page can be picked up due to memory
1112 * offlining. Everywhere else except page reclaim, the page is
1113 * invisible to the vm, so the page can not be migrated. So try to
1114 * free the metadata, so the page can be freed.
1116 if (!page->mapping) {
1117 VM_BUG_ON_PAGE(PageAnon(page), page);
1118 if (page_has_private(page)) {
1119 try_to_free_buffers(page);
1120 goto out_unlock_both;
1122 } else if (page_mapped(page)) {
1123 /* Establish migration ptes */
1124 VM_BUG_ON_PAGE(PageAnon(page) && !PageKsm(page) && !anon_vma,
1126 try_to_unmap(page, TTU_MIGRATION|TTU_IGNORE_MLOCK);
1127 page_was_mapped = 1;
1130 if (!page_mapped(page))
1131 rc = move_to_new_page(newpage, page, mode);
1133 if (page_was_mapped)
1134 remove_migration_ptes(page,
1135 rc == MIGRATEPAGE_SUCCESS ? newpage : page, false);
1138 unlock_page(newpage);
1140 /* Drop an anon_vma reference if we took one */
1142 put_anon_vma(anon_vma);
1146 * If migration is successful, decrease refcount of the newpage
1147 * which will not free the page because new page owner increased
1148 * refcounter. As well, if it is LRU page, add the page to LRU
1149 * list in here. Use the old state of the isolated source page to
1150 * determine if we migrated a LRU page. newpage was already unlocked
1151 * and possibly modified by its owner - don't rely on the page
1154 if (rc == MIGRATEPAGE_SUCCESS) {
1155 if (unlikely(!is_lru))
1158 putback_lru_page(newpage);
1165 * Obtain the lock on page, remove all ptes and migrate the page
1166 * to the newly allocated page in newpage.
1168 static int unmap_and_move(new_page_t get_new_page,
1169 free_page_t put_new_page,
1170 unsigned long private, struct page *page,
1171 int force, enum migrate_mode mode,
1172 enum migrate_reason reason)
1174 int rc = MIGRATEPAGE_SUCCESS;
1175 struct page *newpage = NULL;
1177 if (!thp_migration_supported() && PageTransHuge(page))
1180 if (page_count(page) == 1) {
1181 /* page was freed from under us. So we are done. */
1182 ClearPageActive(page);
1183 ClearPageUnevictable(page);
1184 if (unlikely(__PageMovable(page))) {
1186 if (!PageMovable(page))
1187 __ClearPageIsolated(page);
1193 newpage = get_new_page(page, private);
1197 rc = __unmap_and_move(page, newpage, force, mode);
1198 if (rc == MIGRATEPAGE_SUCCESS)
1199 set_page_owner_migrate_reason(newpage, reason);
1202 if (rc != -EAGAIN) {
1204 * A page that has been migrated has all references
1205 * removed and will be freed. A page that has not been
1206 * migrated will have kept its references and be restored.
1208 list_del(&page->lru);
1211 * Compaction can migrate also non-LRU pages which are
1212 * not accounted to NR_ISOLATED_*. They can be recognized
1215 if (likely(!__PageMovable(page)))
1216 mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON +
1217 page_is_file_lru(page), -thp_nr_pages(page));
1221 * If migration is successful, releases reference grabbed during
1222 * isolation. Otherwise, restore the page to right list unless
1225 if (rc == MIGRATEPAGE_SUCCESS) {
1226 if (reason != MR_MEMORY_FAILURE)
1228 * We release the page in page_handle_poison.
1232 if (rc != -EAGAIN) {
1233 if (likely(!__PageMovable(page))) {
1234 putback_lru_page(page);
1239 if (PageMovable(page))
1240 putback_movable_page(page);
1242 __ClearPageIsolated(page);
1248 put_new_page(newpage, private);
1257 * Counterpart of unmap_and_move_page() for hugepage migration.
1259 * This function doesn't wait the completion of hugepage I/O
1260 * because there is no race between I/O and migration for hugepage.
1261 * Note that currently hugepage I/O occurs only in direct I/O
1262 * where no lock is held and PG_writeback is irrelevant,
1263 * and writeback status of all subpages are counted in the reference
1264 * count of the head page (i.e. if all subpages of a 2MB hugepage are
1265 * under direct I/O, the reference of the head page is 512 and a bit more.)
1266 * This means that when we try to migrate hugepage whose subpages are
1267 * doing direct I/O, some references remain after try_to_unmap() and
1268 * hugepage migration fails without data corruption.
1270 * There is also no race when direct I/O is issued on the page under migration,
1271 * because then pte is replaced with migration swap entry and direct I/O code
1272 * will wait in the page fault for migration to complete.
1274 static int unmap_and_move_huge_page(new_page_t get_new_page,
1275 free_page_t put_new_page, unsigned long private,
1276 struct page *hpage, int force,
1277 enum migrate_mode mode, int reason)
1280 int page_was_mapped = 0;
1281 struct page *new_hpage;
1282 struct anon_vma *anon_vma = NULL;
1283 struct address_space *mapping = NULL;
1286 * Migratability of hugepages depends on architectures and their size.
1287 * This check is necessary because some callers of hugepage migration
1288 * like soft offline and memory hotremove don't walk through page
1289 * tables or check whether the hugepage is pmd-based or not before
1290 * kicking migration.
1292 if (!hugepage_migration_supported(page_hstate(hpage))) {
1293 putback_active_hugepage(hpage);
1297 new_hpage = get_new_page(hpage, private);
1301 if (!trylock_page(hpage)) {
1306 case MIGRATE_SYNC_NO_COPY:
1315 * Check for pages which are in the process of being freed. Without
1316 * page_mapping() set, hugetlbfs specific move page routine will not
1317 * be called and we could leak usage counts for subpools.
1319 if (page_private(hpage) && !page_mapping(hpage)) {
1324 if (PageAnon(hpage))
1325 anon_vma = page_get_anon_vma(hpage);
1327 if (unlikely(!trylock_page(new_hpage)))
1330 if (page_mapped(hpage)) {
1331 bool mapping_locked = false;
1332 enum ttu_flags ttu = TTU_MIGRATION|TTU_IGNORE_MLOCK;
1334 if (!PageAnon(hpage)) {
1336 * In shared mappings, try_to_unmap could potentially
1337 * call huge_pmd_unshare. Because of this, take
1338 * semaphore in write mode here and set TTU_RMAP_LOCKED
1339 * to let lower levels know we have taken the lock.
1341 mapping = hugetlb_page_mapping_lock_write(hpage);
1342 if (unlikely(!mapping))
1343 goto unlock_put_anon;
1345 mapping_locked = true;
1346 ttu |= TTU_RMAP_LOCKED;
1349 try_to_unmap(hpage, ttu);
1350 page_was_mapped = 1;
1353 i_mmap_unlock_write(mapping);
1356 if (!page_mapped(hpage))
1357 rc = move_to_new_page(new_hpage, hpage, mode);
1359 if (page_was_mapped)
1360 remove_migration_ptes(hpage,
1361 rc == MIGRATEPAGE_SUCCESS ? new_hpage : hpage, false);
1364 unlock_page(new_hpage);
1368 put_anon_vma(anon_vma);
1370 if (rc == MIGRATEPAGE_SUCCESS) {
1371 move_hugetlb_state(hpage, new_hpage, reason);
1372 put_new_page = NULL;
1379 putback_active_hugepage(hpage);
1382 * If migration was not successful and there's a freeing callback, use
1383 * it. Otherwise, put_page() will drop the reference grabbed during
1387 put_new_page(new_hpage, private);
1389 putback_active_hugepage(new_hpage);
1395 * migrate_pages - migrate the pages specified in a list, to the free pages
1396 * supplied as the target for the page migration
1398 * @from: The list of pages to be migrated.
1399 * @get_new_page: The function used to allocate free pages to be used
1400 * as the target of the page migration.
1401 * @put_new_page: The function used to free target pages if migration
1402 * fails, or NULL if no special handling is necessary.
1403 * @private: Private data to be passed on to get_new_page()
1404 * @mode: The migration mode that specifies the constraints for
1405 * page migration, if any.
1406 * @reason: The reason for page migration.
1408 * The function returns after 10 attempts or if no pages are movable any more
1409 * because the list has become empty or no retryable pages exist any more.
1410 * The caller should call putback_movable_pages() to return pages to the LRU
1411 * or free list only if ret != 0.
1413 * Returns the number of pages that were not migrated, or an error code.
1415 int migrate_pages(struct list_head *from, new_page_t get_new_page,
1416 free_page_t put_new_page, unsigned long private,
1417 enum migrate_mode mode, int reason)
1422 int nr_succeeded = 0;
1423 int nr_thp_succeeded = 0;
1424 int nr_thp_failed = 0;
1425 int nr_thp_split = 0;
1427 bool is_thp = false;
1430 int swapwrite = current->flags & PF_SWAPWRITE;
1431 int rc, nr_subpages;
1434 current->flags |= PF_SWAPWRITE;
1436 for (pass = 0; pass < 10 && (retry || thp_retry); pass++) {
1440 list_for_each_entry_safe(page, page2, from, lru) {
1443 * THP statistics is based on the source huge page.
1444 * Capture required information that might get lost
1447 is_thp = PageTransHuge(page) && !PageHuge(page);
1448 nr_subpages = thp_nr_pages(page);
1452 rc = unmap_and_move_huge_page(get_new_page,
1453 put_new_page, private, page,
1454 pass > 2, mode, reason);
1456 rc = unmap_and_move(get_new_page, put_new_page,
1457 private, page, pass > 2, mode,
1463 * THP migration might be unsupported or the
1464 * allocation could've failed so we should
1465 * retry on the same page with the THP split
1468 * Head page is retried immediately and tail
1469 * pages are added to the tail of the list so
1470 * we encounter them after the rest of the list
1475 rc = split_huge_page_to_list(page, from);
1478 list_safe_reset_next(page, page2, lru);
1484 nr_failed += nr_subpages;
1496 case MIGRATEPAGE_SUCCESS:
1499 nr_succeeded += nr_subpages;
1506 * Permanent failure (-EBUSY, -ENOSYS, etc.):
1507 * unlike -EAGAIN case, the failed page is
1508 * removed from migration page list and not
1509 * retried in the next outer loop.
1513 nr_failed += nr_subpages;
1521 nr_failed += retry + thp_retry;
1522 nr_thp_failed += thp_retry;
1525 count_vm_events(PGMIGRATE_SUCCESS, nr_succeeded);
1526 count_vm_events(PGMIGRATE_FAIL, nr_failed);
1527 count_vm_events(THP_MIGRATION_SUCCESS, nr_thp_succeeded);
1528 count_vm_events(THP_MIGRATION_FAIL, nr_thp_failed);
1529 count_vm_events(THP_MIGRATION_SPLIT, nr_thp_split);
1530 trace_mm_migrate_pages(nr_succeeded, nr_failed, nr_thp_succeeded,
1531 nr_thp_failed, nr_thp_split, mode, reason);
1534 current->flags &= ~PF_SWAPWRITE;
1539 struct page *alloc_migration_target(struct page *page, unsigned long private)
1541 struct migration_target_control *mtc;
1543 unsigned int order = 0;
1544 struct page *new_page = NULL;
1548 mtc = (struct migration_target_control *)private;
1549 gfp_mask = mtc->gfp_mask;
1551 if (nid == NUMA_NO_NODE)
1552 nid = page_to_nid(page);
1554 if (PageHuge(page)) {
1555 struct hstate *h = page_hstate(compound_head(page));
1557 gfp_mask = htlb_modify_alloc_mask(h, gfp_mask);
1558 return alloc_huge_page_nodemask(h, nid, mtc->nmask, gfp_mask);
1561 if (PageTransHuge(page)) {
1563 * clear __GFP_RECLAIM to make the migration callback
1564 * consistent with regular THP allocations.
1566 gfp_mask &= ~__GFP_RECLAIM;
1567 gfp_mask |= GFP_TRANSHUGE;
1568 order = HPAGE_PMD_ORDER;
1570 zidx = zone_idx(page_zone(page));
1571 if (is_highmem_idx(zidx) || zidx == ZONE_MOVABLE)
1572 gfp_mask |= __GFP_HIGHMEM;
1574 new_page = __alloc_pages_nodemask(gfp_mask, order, nid, mtc->nmask);
1576 if (new_page && PageTransHuge(new_page))
1577 prep_transhuge_page(new_page);
1584 static int store_status(int __user *status, int start, int value, int nr)
1587 if (put_user(value, status + start))
1595 static int do_move_pages_to_node(struct mm_struct *mm,
1596 struct list_head *pagelist, int node)
1599 struct migration_target_control mtc = {
1601 .gfp_mask = GFP_HIGHUSER_MOVABLE | __GFP_THISNODE,
1604 err = migrate_pages(pagelist, alloc_migration_target, NULL,
1605 (unsigned long)&mtc, MIGRATE_SYNC, MR_SYSCALL);
1607 putback_movable_pages(pagelist);
1612 * Resolves the given address to a struct page, isolates it from the LRU and
1613 * puts it to the given pagelist.
1615 * errno - if the page cannot be found/isolated
1616 * 0 - when it doesn't have to be migrated because it is already on the
1618 * 1 - when it has been queued
1620 static int add_page_for_migration(struct mm_struct *mm, unsigned long addr,
1621 int node, struct list_head *pagelist, bool migrate_all)
1623 struct vm_area_struct *vma;
1625 unsigned int follflags;
1630 vma = find_vma(mm, addr);
1631 if (!vma || addr < vma->vm_start || !vma_migratable(vma))
1634 /* FOLL_DUMP to ignore special (like zero) pages */
1635 follflags = FOLL_GET | FOLL_DUMP;
1636 page = follow_page(vma, addr, follflags);
1638 err = PTR_ERR(page);
1647 if (page_to_nid(page) == node)
1651 if (page_mapcount(page) > 1 && !migrate_all)
1654 if (PageHuge(page)) {
1655 if (PageHead(page)) {
1656 isolate_huge_page(page, pagelist);
1662 head = compound_head(page);
1663 err = isolate_lru_page(head);
1668 list_add_tail(&head->lru, pagelist);
1669 mod_node_page_state(page_pgdat(head),
1670 NR_ISOLATED_ANON + page_is_file_lru(head),
1671 thp_nr_pages(head));
1675 * Either remove the duplicate refcount from
1676 * isolate_lru_page() or drop the page ref if it was
1681 mmap_read_unlock(mm);
1685 static int move_pages_and_store_status(struct mm_struct *mm, int node,
1686 struct list_head *pagelist, int __user *status,
1687 int start, int i, unsigned long nr_pages)
1691 if (list_empty(pagelist))
1694 err = do_move_pages_to_node(mm, pagelist, node);
1697 * Positive err means the number of failed
1698 * pages to migrate. Since we are going to
1699 * abort and return the number of non-migrated
1700 * pages, so need to incude the rest of the
1701 * nr_pages that have not been attempted as
1705 err += nr_pages - i - 1;
1708 return store_status(status, start, node, i - start);
1712 * Migrate an array of page address onto an array of nodes and fill
1713 * the corresponding array of status.
1715 static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes,
1716 unsigned long nr_pages,
1717 const void __user * __user *pages,
1718 const int __user *nodes,
1719 int __user *status, int flags)
1721 int current_node = NUMA_NO_NODE;
1722 LIST_HEAD(pagelist);
1728 for (i = start = 0; i < nr_pages; i++) {
1729 const void __user *p;
1734 if (get_user(p, pages + i))
1736 if (get_user(node, nodes + i))
1738 addr = (unsigned long)untagged_addr(p);
1741 if (node < 0 || node >= MAX_NUMNODES)
1743 if (!node_state(node, N_MEMORY))
1747 if (!node_isset(node, task_nodes))
1750 if (current_node == NUMA_NO_NODE) {
1751 current_node = node;
1753 } else if (node != current_node) {
1754 err = move_pages_and_store_status(mm, current_node,
1755 &pagelist, status, start, i, nr_pages);
1759 current_node = node;
1763 * Errors in the page lookup or isolation are not fatal and we simply
1764 * report them via status
1766 err = add_page_for_migration(mm, addr, current_node,
1767 &pagelist, flags & MPOL_MF_MOVE_ALL);
1770 /* The page is successfully queued for migration */
1775 * If the page is already on the target node (!err), store the
1776 * node, otherwise, store the err.
1778 err = store_status(status, i, err ? : current_node, 1);
1782 err = move_pages_and_store_status(mm, current_node, &pagelist,
1783 status, start, i, nr_pages);
1786 current_node = NUMA_NO_NODE;
1789 /* Make sure we do not overwrite the existing error */
1790 err1 = move_pages_and_store_status(mm, current_node, &pagelist,
1791 status, start, i, nr_pages);
1799 * Determine the nodes of an array of pages and store it in an array of status.
1801 static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages,
1802 const void __user **pages, int *status)
1808 for (i = 0; i < nr_pages; i++) {
1809 unsigned long addr = (unsigned long)(*pages);
1810 struct vm_area_struct *vma;
1814 vma = find_vma(mm, addr);
1815 if (!vma || addr < vma->vm_start)
1818 /* FOLL_DUMP to ignore special (like zero) pages */
1819 page = follow_page(vma, addr, FOLL_DUMP);
1821 err = PTR_ERR(page);
1825 err = page ? page_to_nid(page) : -ENOENT;
1833 mmap_read_unlock(mm);
1837 * Determine the nodes of a user array of pages and store it in
1838 * a user array of status.
1840 static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages,
1841 const void __user * __user *pages,
1844 #define DO_PAGES_STAT_CHUNK_NR 16
1845 const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR];
1846 int chunk_status[DO_PAGES_STAT_CHUNK_NR];
1849 unsigned long chunk_nr;
1851 chunk_nr = nr_pages;
1852 if (chunk_nr > DO_PAGES_STAT_CHUNK_NR)
1853 chunk_nr = DO_PAGES_STAT_CHUNK_NR;
1855 if (copy_from_user(chunk_pages, pages, chunk_nr * sizeof(*chunk_pages)))
1858 do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status);
1860 if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status)))
1865 nr_pages -= chunk_nr;
1867 return nr_pages ? -EFAULT : 0;
1870 static struct mm_struct *find_mm_struct(pid_t pid, nodemask_t *mem_nodes)
1872 struct task_struct *task;
1873 struct mm_struct *mm;
1876 * There is no need to check if current process has the right to modify
1877 * the specified process when they are same.
1881 *mem_nodes = cpuset_mems_allowed(current);
1885 /* Find the mm_struct */
1887 task = find_task_by_vpid(pid);
1890 return ERR_PTR(-ESRCH);
1892 get_task_struct(task);
1895 * Check if this process has the right to modify the specified
1896 * process. Use the regular "ptrace_may_access()" checks.
1898 if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) {
1900 mm = ERR_PTR(-EPERM);
1905 mm = ERR_PTR(security_task_movememory(task));
1908 *mem_nodes = cpuset_mems_allowed(task);
1909 mm = get_task_mm(task);
1911 put_task_struct(task);
1913 mm = ERR_PTR(-EINVAL);
1918 * Move a list of pages in the address space of the currently executing
1921 static int kernel_move_pages(pid_t pid, unsigned long nr_pages,
1922 const void __user * __user *pages,
1923 const int __user *nodes,
1924 int __user *status, int flags)
1926 struct mm_struct *mm;
1928 nodemask_t task_nodes;
1931 if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
1934 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1937 mm = find_mm_struct(pid, &task_nodes);
1942 err = do_pages_move(mm, task_nodes, nr_pages, pages,
1943 nodes, status, flags);
1945 err = do_pages_stat(mm, nr_pages, pages, status);
1951 SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages,
1952 const void __user * __user *, pages,
1953 const int __user *, nodes,
1954 int __user *, status, int, flags)
1956 return kernel_move_pages(pid, nr_pages, pages, nodes, status, flags);
1959 #ifdef CONFIG_COMPAT
1960 COMPAT_SYSCALL_DEFINE6(move_pages, pid_t, pid, compat_ulong_t, nr_pages,
1961 compat_uptr_t __user *, pages32,
1962 const int __user *, nodes,
1963 int __user *, status,
1966 const void __user * __user *pages;
1969 pages = compat_alloc_user_space(nr_pages * sizeof(void *));
1970 for (i = 0; i < nr_pages; i++) {
1973 if (get_user(p, pages32 + i) ||
1974 put_user(compat_ptr(p), pages + i))
1977 return kernel_move_pages(pid, nr_pages, pages, nodes, status, flags);
1979 #endif /* CONFIG_COMPAT */
1981 #ifdef CONFIG_NUMA_BALANCING
1983 * Returns true if this is a safe migration target node for misplaced NUMA
1984 * pages. Currently it only checks the watermarks which crude
1986 static bool migrate_balanced_pgdat(struct pglist_data *pgdat,
1987 unsigned long nr_migrate_pages)
1991 for (z = pgdat->nr_zones - 1; z >= 0; z--) {
1992 struct zone *zone = pgdat->node_zones + z;
1994 if (!populated_zone(zone))
1997 /* Avoid waking kswapd by allocating pages_to_migrate pages. */
1998 if (!zone_watermark_ok(zone, 0,
1999 high_wmark_pages(zone) +
2008 static struct page *alloc_misplaced_dst_page(struct page *page,
2011 int nid = (int) data;
2012 struct page *newpage;
2014 newpage = __alloc_pages_node(nid,
2015 (GFP_HIGHUSER_MOVABLE |
2016 __GFP_THISNODE | __GFP_NOMEMALLOC |
2017 __GFP_NORETRY | __GFP_NOWARN) &
2023 static int numamigrate_isolate_page(pg_data_t *pgdat, struct page *page)
2027 VM_BUG_ON_PAGE(compound_order(page) && !PageTransHuge(page), page);
2029 /* Avoid migrating to a node that is nearly full */
2030 if (!migrate_balanced_pgdat(pgdat, compound_nr(page)))
2033 if (isolate_lru_page(page))
2037 * migrate_misplaced_transhuge_page() skips page migration's usual
2038 * check on page_count(), so we must do it here, now that the page
2039 * has been isolated: a GUP pin, or any other pin, prevents migration.
2040 * The expected page count is 3: 1 for page's mapcount and 1 for the
2041 * caller's pin and 1 for the reference taken by isolate_lru_page().
2043 if (PageTransHuge(page) && page_count(page) != 3) {
2044 putback_lru_page(page);
2048 page_lru = page_is_file_lru(page);
2049 mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON + page_lru,
2050 thp_nr_pages(page));
2053 * Isolating the page has taken another reference, so the
2054 * caller's reference can be safely dropped without the page
2055 * disappearing underneath us during migration.
2061 bool pmd_trans_migrating(pmd_t pmd)
2063 struct page *page = pmd_page(pmd);
2064 return PageLocked(page);
2068 * Attempt to migrate a misplaced page to the specified destination
2069 * node. Caller is expected to have an elevated reference count on
2070 * the page that will be dropped by this function before returning.
2072 int migrate_misplaced_page(struct page *page, struct vm_area_struct *vma,
2075 pg_data_t *pgdat = NODE_DATA(node);
2078 LIST_HEAD(migratepages);
2081 * Don't migrate file pages that are mapped in multiple processes
2082 * with execute permissions as they are probably shared libraries.
2084 if (page_mapcount(page) != 1 && page_is_file_lru(page) &&
2085 (vma->vm_flags & VM_EXEC))
2089 * Also do not migrate dirty pages as not all filesystems can move
2090 * dirty pages in MIGRATE_ASYNC mode which is a waste of cycles.
2092 if (page_is_file_lru(page) && PageDirty(page))
2095 isolated = numamigrate_isolate_page(pgdat, page);
2099 list_add(&page->lru, &migratepages);
2100 nr_remaining = migrate_pages(&migratepages, alloc_misplaced_dst_page,
2101 NULL, node, MIGRATE_ASYNC,
2104 if (!list_empty(&migratepages)) {
2105 list_del(&page->lru);
2106 dec_node_page_state(page, NR_ISOLATED_ANON +
2107 page_is_file_lru(page));
2108 putback_lru_page(page);
2112 count_vm_numa_event(NUMA_PAGE_MIGRATE);
2113 BUG_ON(!list_empty(&migratepages));
2120 #endif /* CONFIG_NUMA_BALANCING */
2122 #if defined(CONFIG_NUMA_BALANCING) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
2124 * Migrates a THP to a given target node. page must be locked and is unlocked
2127 int migrate_misplaced_transhuge_page(struct mm_struct *mm,
2128 struct vm_area_struct *vma,
2129 pmd_t *pmd, pmd_t entry,
2130 unsigned long address,
2131 struct page *page, int node)
2134 pg_data_t *pgdat = NODE_DATA(node);
2136 struct page *new_page = NULL;
2137 int page_lru = page_is_file_lru(page);
2138 unsigned long start = address & HPAGE_PMD_MASK;
2140 new_page = alloc_pages_node(node,
2141 (GFP_TRANSHUGE_LIGHT | __GFP_THISNODE),
2145 prep_transhuge_page(new_page);
2147 isolated = numamigrate_isolate_page(pgdat, page);
2153 /* Prepare a page as a migration target */
2154 __SetPageLocked(new_page);
2155 if (PageSwapBacked(page))
2156 __SetPageSwapBacked(new_page);
2158 /* anon mapping, we can simply copy page->mapping to the new page: */
2159 new_page->mapping = page->mapping;
2160 new_page->index = page->index;
2161 /* flush the cache before copying using the kernel virtual address */
2162 flush_cache_range(vma, start, start + HPAGE_PMD_SIZE);
2163 migrate_page_copy(new_page, page);
2164 WARN_ON(PageLRU(new_page));
2166 /* Recheck the target PMD */
2167 ptl = pmd_lock(mm, pmd);
2168 if (unlikely(!pmd_same(*pmd, entry) || !page_ref_freeze(page, 2))) {
2171 /* Reverse changes made by migrate_page_copy() */
2172 if (TestClearPageActive(new_page))
2173 SetPageActive(page);
2174 if (TestClearPageUnevictable(new_page))
2175 SetPageUnevictable(page);
2177 unlock_page(new_page);
2178 put_page(new_page); /* Free it */
2180 /* Retake the callers reference and putback on LRU */
2182 putback_lru_page(page);
2183 mod_node_page_state(page_pgdat(page),
2184 NR_ISOLATED_ANON + page_lru, -HPAGE_PMD_NR);
2189 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
2190 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
2193 * Overwrite the old entry under pagetable lock and establish
2194 * the new PTE. Any parallel GUP will either observe the old
2195 * page blocking on the page lock, block on the page table
2196 * lock or observe the new page. The SetPageUptodate on the
2197 * new page and page_add_new_anon_rmap guarantee the copy is
2198 * visible before the pagetable update.
2200 page_add_anon_rmap(new_page, vma, start, true);
2202 * At this point the pmd is numa/protnone (i.e. non present) and the TLB
2203 * has already been flushed globally. So no TLB can be currently
2204 * caching this non present pmd mapping. There's no need to clear the
2205 * pmd before doing set_pmd_at(), nor to flush the TLB after
2206 * set_pmd_at(). Clearing the pmd here would introduce a race
2207 * condition against MADV_DONTNEED, because MADV_DONTNEED only holds the
2208 * mmap_lock for reading. If the pmd is set to NULL at any given time,
2209 * MADV_DONTNEED won't wait on the pmd lock and it'll skip clearing this
2212 set_pmd_at(mm, start, pmd, entry);
2213 update_mmu_cache_pmd(vma, address, &entry);
2215 page_ref_unfreeze(page, 2);
2216 mlock_migrate_page(new_page, page);
2217 page_remove_rmap(page, true);
2218 set_page_owner_migrate_reason(new_page, MR_NUMA_MISPLACED);
2222 /* Take an "isolate" reference and put new page on the LRU. */
2224 putback_lru_page(new_page);
2226 unlock_page(new_page);
2228 put_page(page); /* Drop the rmap reference */
2229 put_page(page); /* Drop the LRU isolation reference */
2231 count_vm_events(PGMIGRATE_SUCCESS, HPAGE_PMD_NR);
2232 count_vm_numa_events(NUMA_PAGE_MIGRATE, HPAGE_PMD_NR);
2234 mod_node_page_state(page_pgdat(page),
2235 NR_ISOLATED_ANON + page_lru,
2240 count_vm_events(PGMIGRATE_FAIL, HPAGE_PMD_NR);
2241 ptl = pmd_lock(mm, pmd);
2242 if (pmd_same(*pmd, entry)) {
2243 entry = pmd_modify(entry, vma->vm_page_prot);
2244 set_pmd_at(mm, start, pmd, entry);
2245 update_mmu_cache_pmd(vma, address, &entry);
2254 #endif /* CONFIG_NUMA_BALANCING */
2256 #endif /* CONFIG_NUMA */
2258 #ifdef CONFIG_DEVICE_PRIVATE
2259 static int migrate_vma_collect_hole(unsigned long start,
2261 __always_unused int depth,
2262 struct mm_walk *walk)
2264 struct migrate_vma *migrate = walk->private;
2267 /* Only allow populating anonymous memory. */
2268 if (!vma_is_anonymous(walk->vma)) {
2269 for (addr = start; addr < end; addr += PAGE_SIZE) {
2270 migrate->src[migrate->npages] = 0;
2271 migrate->dst[migrate->npages] = 0;
2277 for (addr = start; addr < end; addr += PAGE_SIZE) {
2278 migrate->src[migrate->npages] = MIGRATE_PFN_MIGRATE;
2279 migrate->dst[migrate->npages] = 0;
2287 static int migrate_vma_collect_skip(unsigned long start,
2289 struct mm_walk *walk)
2291 struct migrate_vma *migrate = walk->private;
2294 for (addr = start; addr < end; addr += PAGE_SIZE) {
2295 migrate->dst[migrate->npages] = 0;
2296 migrate->src[migrate->npages++] = 0;
2302 static int migrate_vma_collect_pmd(pmd_t *pmdp,
2303 unsigned long start,
2305 struct mm_walk *walk)
2307 struct migrate_vma *migrate = walk->private;
2308 struct vm_area_struct *vma = walk->vma;
2309 struct mm_struct *mm = vma->vm_mm;
2310 unsigned long addr = start, unmapped = 0;
2315 if (pmd_none(*pmdp))
2316 return migrate_vma_collect_hole(start, end, -1, walk);
2318 if (pmd_trans_huge(*pmdp)) {
2321 ptl = pmd_lock(mm, pmdp);
2322 if (unlikely(!pmd_trans_huge(*pmdp))) {
2327 page = pmd_page(*pmdp);
2328 if (is_huge_zero_page(page)) {
2330 split_huge_pmd(vma, pmdp, addr);
2331 if (pmd_trans_unstable(pmdp))
2332 return migrate_vma_collect_skip(start, end,
2339 if (unlikely(!trylock_page(page)))
2340 return migrate_vma_collect_skip(start, end,
2342 ret = split_huge_page(page);
2346 return migrate_vma_collect_skip(start, end,
2348 if (pmd_none(*pmdp))
2349 return migrate_vma_collect_hole(start, end, -1,
2354 if (unlikely(pmd_bad(*pmdp)))
2355 return migrate_vma_collect_skip(start, end, walk);
2357 ptep = pte_offset_map_lock(mm, pmdp, addr, &ptl);
2358 arch_enter_lazy_mmu_mode();
2360 for (; addr < end; addr += PAGE_SIZE, ptep++) {
2361 unsigned long mpfn = 0, pfn;
2368 if (pte_none(pte)) {
2369 if (vma_is_anonymous(vma)) {
2370 mpfn = MIGRATE_PFN_MIGRATE;
2376 if (!pte_present(pte)) {
2378 * Only care about unaddressable device page special
2379 * page table entry. Other special swap entries are not
2380 * migratable, and we ignore regular swapped page.
2382 entry = pte_to_swp_entry(pte);
2383 if (!is_device_private_entry(entry))
2386 page = device_private_entry_to_page(entry);
2387 if (!(migrate->flags &
2388 MIGRATE_VMA_SELECT_DEVICE_PRIVATE) ||
2389 page->pgmap->owner != migrate->pgmap_owner)
2392 mpfn = migrate_pfn(page_to_pfn(page)) |
2393 MIGRATE_PFN_MIGRATE;
2394 if (is_write_device_private_entry(entry))
2395 mpfn |= MIGRATE_PFN_WRITE;
2397 if (!(migrate->flags & MIGRATE_VMA_SELECT_SYSTEM))
2400 if (is_zero_pfn(pfn)) {
2401 mpfn = MIGRATE_PFN_MIGRATE;
2405 page = vm_normal_page(migrate->vma, addr, pte);
2406 mpfn = migrate_pfn(pfn) | MIGRATE_PFN_MIGRATE;
2407 mpfn |= pte_write(pte) ? MIGRATE_PFN_WRITE : 0;
2410 /* FIXME support THP */
2411 if (!page || !page->mapping || PageTransCompound(page)) {
2417 * By getting a reference on the page we pin it and that blocks
2418 * any kind of migration. Side effect is that it "freezes" the
2421 * We drop this reference after isolating the page from the lru
2422 * for non device page (device page are not on the lru and thus
2423 * can't be dropped from it).
2429 * Optimize for the common case where page is only mapped once
2430 * in one process. If we can lock the page, then we can safely
2431 * set up a special migration page table entry now.
2433 if (trylock_page(page)) {
2436 mpfn |= MIGRATE_PFN_LOCKED;
2437 ptep_get_and_clear(mm, addr, ptep);
2439 /* Setup special migration page table entry */
2440 entry = make_migration_entry(page, mpfn &
2442 swp_pte = swp_entry_to_pte(entry);
2443 if (pte_present(pte)) {
2444 if (pte_soft_dirty(pte))
2445 swp_pte = pte_swp_mksoft_dirty(swp_pte);
2446 if (pte_uffd_wp(pte))
2447 swp_pte = pte_swp_mkuffd_wp(swp_pte);
2449 if (pte_swp_soft_dirty(pte))
2450 swp_pte = pte_swp_mksoft_dirty(swp_pte);
2451 if (pte_swp_uffd_wp(pte))
2452 swp_pte = pte_swp_mkuffd_wp(swp_pte);
2454 set_pte_at(mm, addr, ptep, swp_pte);
2457 * This is like regular unmap: we remove the rmap and
2458 * drop page refcount. Page won't be freed, as we took
2459 * a reference just above.
2461 page_remove_rmap(page, false);
2464 if (pte_present(pte))
2469 migrate->dst[migrate->npages] = 0;
2470 migrate->src[migrate->npages++] = mpfn;
2472 arch_leave_lazy_mmu_mode();
2473 pte_unmap_unlock(ptep - 1, ptl);
2475 /* Only flush the TLB if we actually modified any entries */
2477 flush_tlb_range(walk->vma, start, end);
2482 static const struct mm_walk_ops migrate_vma_walk_ops = {
2483 .pmd_entry = migrate_vma_collect_pmd,
2484 .pte_hole = migrate_vma_collect_hole,
2488 * migrate_vma_collect() - collect pages over a range of virtual addresses
2489 * @migrate: migrate struct containing all migration information
2491 * This will walk the CPU page table. For each virtual address backed by a
2492 * valid page, it updates the src array and takes a reference on the page, in
2493 * order to pin the page until we lock it and unmap it.
2495 static void migrate_vma_collect(struct migrate_vma *migrate)
2497 struct mmu_notifier_range range;
2500 * Note that the pgmap_owner is passed to the mmu notifier callback so
2501 * that the registered device driver can skip invalidating device
2502 * private page mappings that won't be migrated.
2504 mmu_notifier_range_init_migrate(&range, 0, migrate->vma,
2505 migrate->vma->vm_mm, migrate->start, migrate->end,
2506 migrate->pgmap_owner);
2507 mmu_notifier_invalidate_range_start(&range);
2509 walk_page_range(migrate->vma->vm_mm, migrate->start, migrate->end,
2510 &migrate_vma_walk_ops, migrate);
2512 mmu_notifier_invalidate_range_end(&range);
2513 migrate->end = migrate->start + (migrate->npages << PAGE_SHIFT);
2517 * migrate_vma_check_page() - check if page is pinned or not
2518 * @page: struct page to check
2520 * Pinned pages cannot be migrated. This is the same test as in
2521 * migrate_page_move_mapping(), except that here we allow migration of a
2524 static bool migrate_vma_check_page(struct page *page)
2527 * One extra ref because caller holds an extra reference, either from
2528 * isolate_lru_page() for a regular page, or migrate_vma_collect() for
2534 * FIXME support THP (transparent huge page), it is bit more complex to
2535 * check them than regular pages, because they can be mapped with a pmd
2536 * or with a pte (split pte mapping).
2538 if (PageCompound(page))
2541 /* Page from ZONE_DEVICE have one extra reference */
2542 if (is_zone_device_page(page)) {
2544 * Private page can never be pin as they have no valid pte and
2545 * GUP will fail for those. Yet if there is a pending migration
2546 * a thread might try to wait on the pte migration entry and
2547 * will bump the page reference count. Sadly there is no way to
2548 * differentiate a regular pin from migration wait. Hence to
2549 * avoid 2 racing thread trying to migrate back to CPU to enter
2550 * infinite loop (one stoping migration because the other is
2551 * waiting on pte migration entry). We always return true here.
2553 * FIXME proper solution is to rework migration_entry_wait() so
2554 * it does not need to take a reference on page.
2556 return is_device_private_page(page);
2559 /* For file back page */
2560 if (page_mapping(page))
2561 extra += 1 + page_has_private(page);
2563 if ((page_count(page) - extra) > page_mapcount(page))
2570 * migrate_vma_prepare() - lock pages and isolate them from the lru
2571 * @migrate: migrate struct containing all migration information
2573 * This locks pages that have been collected by migrate_vma_collect(). Once each
2574 * page is locked it is isolated from the lru (for non-device pages). Finally,
2575 * the ref taken by migrate_vma_collect() is dropped, as locked pages cannot be
2576 * migrated by concurrent kernel threads.
2578 static void migrate_vma_prepare(struct migrate_vma *migrate)
2580 const unsigned long npages = migrate->npages;
2581 const unsigned long start = migrate->start;
2582 unsigned long addr, i, restore = 0;
2583 bool allow_drain = true;
2587 for (i = 0; (i < npages) && migrate->cpages; i++) {
2588 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2594 if (!(migrate->src[i] & MIGRATE_PFN_LOCKED)) {
2596 * Because we are migrating several pages there can be
2597 * a deadlock between 2 concurrent migration where each
2598 * are waiting on each other page lock.
2600 * Make migrate_vma() a best effort thing and backoff
2601 * for any page we can not lock right away.
2603 if (!trylock_page(page)) {
2604 migrate->src[i] = 0;
2610 migrate->src[i] |= MIGRATE_PFN_LOCKED;
2613 /* ZONE_DEVICE pages are not on LRU */
2614 if (!is_zone_device_page(page)) {
2615 if (!PageLRU(page) && allow_drain) {
2616 /* Drain CPU's pagevec */
2617 lru_add_drain_all();
2618 allow_drain = false;
2621 if (isolate_lru_page(page)) {
2623 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2627 migrate->src[i] = 0;
2635 /* Drop the reference we took in collect */
2639 if (!migrate_vma_check_page(page)) {
2641 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2645 if (!is_zone_device_page(page)) {
2647 putback_lru_page(page);
2650 migrate->src[i] = 0;
2654 if (!is_zone_device_page(page))
2655 putback_lru_page(page);
2662 for (i = 0, addr = start; i < npages && restore; i++, addr += PAGE_SIZE) {
2663 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2665 if (!page || (migrate->src[i] & MIGRATE_PFN_MIGRATE))
2668 remove_migration_pte(page, migrate->vma, addr, page);
2670 migrate->src[i] = 0;
2678 * migrate_vma_unmap() - replace page mapping with special migration pte entry
2679 * @migrate: migrate struct containing all migration information
2681 * Replace page mapping (CPU page table pte) with a special migration pte entry
2682 * and check again if it has been pinned. Pinned pages are restored because we
2683 * cannot migrate them.
2685 * This is the last step before we call the device driver callback to allocate
2686 * destination memory and copy contents of original page over to new page.
2688 static void migrate_vma_unmap(struct migrate_vma *migrate)
2690 int flags = TTU_MIGRATION | TTU_IGNORE_MLOCK;
2691 const unsigned long npages = migrate->npages;
2692 const unsigned long start = migrate->start;
2693 unsigned long addr, i, restore = 0;
2695 for (i = 0; i < npages; i++) {
2696 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2698 if (!page || !(migrate->src[i] & MIGRATE_PFN_MIGRATE))
2701 if (page_mapped(page)) {
2702 try_to_unmap(page, flags);
2703 if (page_mapped(page))
2707 if (migrate_vma_check_page(page))
2711 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2716 for (addr = start, i = 0; i < npages && restore; addr += PAGE_SIZE, i++) {
2717 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2719 if (!page || (migrate->src[i] & MIGRATE_PFN_MIGRATE))
2722 remove_migration_ptes(page, page, false);
2724 migrate->src[i] = 0;
2728 if (is_zone_device_page(page))
2731 putback_lru_page(page);
2736 * migrate_vma_setup() - prepare to migrate a range of memory
2737 * @args: contains the vma, start, and pfns arrays for the migration
2739 * Returns: negative errno on failures, 0 when 0 or more pages were migrated
2742 * Prepare to migrate a range of memory virtual address range by collecting all
2743 * the pages backing each virtual address in the range, saving them inside the
2744 * src array. Then lock those pages and unmap them. Once the pages are locked
2745 * and unmapped, check whether each page is pinned or not. Pages that aren't
2746 * pinned have the MIGRATE_PFN_MIGRATE flag set (by this function) in the
2747 * corresponding src array entry. Then restores any pages that are pinned, by
2748 * remapping and unlocking those pages.
2750 * The caller should then allocate destination memory and copy source memory to
2751 * it for all those entries (ie with MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE
2752 * flag set). Once these are allocated and copied, the caller must update each
2753 * corresponding entry in the dst array with the pfn value of the destination
2754 * page and with the MIGRATE_PFN_VALID and MIGRATE_PFN_LOCKED flags set
2755 * (destination pages must have their struct pages locked, via lock_page()).
2757 * Note that the caller does not have to migrate all the pages that are marked
2758 * with MIGRATE_PFN_MIGRATE flag in src array unless this is a migration from
2759 * device memory to system memory. If the caller cannot migrate a device page
2760 * back to system memory, then it must return VM_FAULT_SIGBUS, which has severe
2761 * consequences for the userspace process, so it must be avoided if at all
2764 * For empty entries inside CPU page table (pte_none() or pmd_none() is true) we
2765 * do set MIGRATE_PFN_MIGRATE flag inside the corresponding source array thus
2766 * allowing the caller to allocate device memory for those unback virtual
2767 * address. For this the caller simply has to allocate device memory and
2768 * properly set the destination entry like for regular migration. Note that
2769 * this can still fails and thus inside the device driver must check if the
2770 * migration was successful for those entries after calling migrate_vma_pages()
2771 * just like for regular migration.
2773 * After that, the callers must call migrate_vma_pages() to go over each entry
2774 * in the src array that has the MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE flag
2775 * set. If the corresponding entry in dst array has MIGRATE_PFN_VALID flag set,
2776 * then migrate_vma_pages() to migrate struct page information from the source
2777 * struct page to the destination struct page. If it fails to migrate the
2778 * struct page information, then it clears the MIGRATE_PFN_MIGRATE flag in the
2781 * At this point all successfully migrated pages have an entry in the src
2782 * array with MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE flag set and the dst
2783 * array entry with MIGRATE_PFN_VALID flag set.
2785 * Once migrate_vma_pages() returns the caller may inspect which pages were
2786 * successfully migrated, and which were not. Successfully migrated pages will
2787 * have the MIGRATE_PFN_MIGRATE flag set for their src array entry.
2789 * It is safe to update device page table after migrate_vma_pages() because
2790 * both destination and source page are still locked, and the mmap_lock is held
2791 * in read mode (hence no one can unmap the range being migrated).
2793 * Once the caller is done cleaning up things and updating its page table (if it
2794 * chose to do so, this is not an obligation) it finally calls
2795 * migrate_vma_finalize() to update the CPU page table to point to new pages
2796 * for successfully migrated pages or otherwise restore the CPU page table to
2797 * point to the original source pages.
2799 int migrate_vma_setup(struct migrate_vma *args)
2801 long nr_pages = (args->end - args->start) >> PAGE_SHIFT;
2803 args->start &= PAGE_MASK;
2804 args->end &= PAGE_MASK;
2805 if (!args->vma || is_vm_hugetlb_page(args->vma) ||
2806 (args->vma->vm_flags & VM_SPECIAL) || vma_is_dax(args->vma))
2810 if (args->start < args->vma->vm_start ||
2811 args->start >= args->vma->vm_end)
2813 if (args->end <= args->vma->vm_start || args->end > args->vma->vm_end)
2815 if (!args->src || !args->dst)
2818 memset(args->src, 0, sizeof(*args->src) * nr_pages);
2822 migrate_vma_collect(args);
2825 migrate_vma_prepare(args);
2827 migrate_vma_unmap(args);
2830 * At this point pages are locked and unmapped, and thus they have
2831 * stable content and can safely be copied to destination memory that
2832 * is allocated by the drivers.
2837 EXPORT_SYMBOL(migrate_vma_setup);
2840 * This code closely matches the code in:
2841 * __handle_mm_fault()
2842 * handle_pte_fault()
2843 * do_anonymous_page()
2844 * to map in an anonymous zero page but the struct page will be a ZONE_DEVICE
2847 static void migrate_vma_insert_page(struct migrate_vma *migrate,
2853 struct vm_area_struct *vma = migrate->vma;
2854 struct mm_struct *mm = vma->vm_mm;
2864 /* Only allow populating anonymous memory */
2865 if (!vma_is_anonymous(vma))
2868 pgdp = pgd_offset(mm, addr);
2869 p4dp = p4d_alloc(mm, pgdp, addr);
2872 pudp = pud_alloc(mm, p4dp, addr);
2875 pmdp = pmd_alloc(mm, pudp, addr);
2879 if (pmd_trans_huge(*pmdp) || pmd_devmap(*pmdp))
2883 * Use pte_alloc() instead of pte_alloc_map(). We can't run
2884 * pte_offset_map() on pmds where a huge pmd might be created
2885 * from a different thread.
2887 * pte_alloc_map() is safe to use under mmap_write_lock(mm) or when
2888 * parallel threads are excluded by other means.
2890 * Here we only have mmap_read_lock(mm).
2892 if (pte_alloc(mm, pmdp))
2895 /* See the comment in pte_alloc_one_map() */
2896 if (unlikely(pmd_trans_unstable(pmdp)))
2899 if (unlikely(anon_vma_prepare(vma)))
2901 if (mem_cgroup_charge(page, vma->vm_mm, GFP_KERNEL))
2905 * The memory barrier inside __SetPageUptodate makes sure that
2906 * preceding stores to the page contents become visible before
2907 * the set_pte_at() write.
2909 __SetPageUptodate(page);
2911 if (is_zone_device_page(page)) {
2912 if (is_device_private_page(page)) {
2913 swp_entry_t swp_entry;
2915 swp_entry = make_device_private_entry(page, vma->vm_flags & VM_WRITE);
2916 entry = swp_entry_to_pte(swp_entry);
2919 entry = mk_pte(page, vma->vm_page_prot);
2920 if (vma->vm_flags & VM_WRITE)
2921 entry = pte_mkwrite(pte_mkdirty(entry));
2924 ptep = pte_offset_map_lock(mm, pmdp, addr, &ptl);
2926 if (check_stable_address_space(mm))
2929 if (pte_present(*ptep)) {
2930 unsigned long pfn = pte_pfn(*ptep);
2932 if (!is_zero_pfn(pfn))
2935 } else if (!pte_none(*ptep))
2939 * Check for userfaultfd but do not deliver the fault. Instead,
2942 if (userfaultfd_missing(vma))
2945 inc_mm_counter(mm, MM_ANONPAGES);
2946 page_add_new_anon_rmap(page, vma, addr, false);
2947 if (!is_zone_device_page(page))
2948 lru_cache_add_inactive_or_unevictable(page, vma);
2952 flush_cache_page(vma, addr, pte_pfn(*ptep));
2953 ptep_clear_flush_notify(vma, addr, ptep);
2954 set_pte_at_notify(mm, addr, ptep, entry);
2955 update_mmu_cache(vma, addr, ptep);
2957 /* No need to invalidate - it was non-present before */
2958 set_pte_at(mm, addr, ptep, entry);
2959 update_mmu_cache(vma, addr, ptep);
2962 pte_unmap_unlock(ptep, ptl);
2963 *src = MIGRATE_PFN_MIGRATE;
2967 pte_unmap_unlock(ptep, ptl);
2969 *src &= ~MIGRATE_PFN_MIGRATE;
2973 * migrate_vma_pages() - migrate meta-data from src page to dst page
2974 * @migrate: migrate struct containing all migration information
2976 * This migrates struct page meta-data from source struct page to destination
2977 * struct page. This effectively finishes the migration from source page to the
2980 void migrate_vma_pages(struct migrate_vma *migrate)
2982 const unsigned long npages = migrate->npages;
2983 const unsigned long start = migrate->start;
2984 struct mmu_notifier_range range;
2985 unsigned long addr, i;
2986 bool notified = false;
2988 for (i = 0, addr = start; i < npages; addr += PAGE_SIZE, i++) {
2989 struct page *newpage = migrate_pfn_to_page(migrate->dst[i]);
2990 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2991 struct address_space *mapping;
2995 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
3000 if (!(migrate->src[i] & MIGRATE_PFN_MIGRATE))
3005 mmu_notifier_range_init(&range,
3006 MMU_NOTIFY_CLEAR, 0,
3008 migrate->vma->vm_mm,
3009 addr, migrate->end);
3010 mmu_notifier_invalidate_range_start(&range);
3012 migrate_vma_insert_page(migrate, addr, newpage,
3018 mapping = page_mapping(page);
3020 if (is_zone_device_page(newpage)) {
3021 if (is_device_private_page(newpage)) {
3023 * For now only support private anonymous when
3024 * migrating to un-addressable device memory.
3027 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
3032 * Other types of ZONE_DEVICE page are not
3035 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
3040 r = migrate_page(mapping, newpage, page, MIGRATE_SYNC_NO_COPY);
3041 if (r != MIGRATEPAGE_SUCCESS)
3042 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
3046 * No need to double call mmu_notifier->invalidate_range() callback as
3047 * the above ptep_clear_flush_notify() inside migrate_vma_insert_page()
3048 * did already call it.
3051 mmu_notifier_invalidate_range_only_end(&range);
3053 EXPORT_SYMBOL(migrate_vma_pages);
3056 * migrate_vma_finalize() - restore CPU page table entry
3057 * @migrate: migrate struct containing all migration information
3059 * This replaces the special migration pte entry with either a mapping to the
3060 * new page if migration was successful for that page, or to the original page
3063 * This also unlocks the pages and puts them back on the lru, or drops the extra
3064 * refcount, for device pages.
3066 void migrate_vma_finalize(struct migrate_vma *migrate)
3068 const unsigned long npages = migrate->npages;
3071 for (i = 0; i < npages; i++) {
3072 struct page *newpage = migrate_pfn_to_page(migrate->dst[i]);
3073 struct page *page = migrate_pfn_to_page(migrate->src[i]);
3077 unlock_page(newpage);
3083 if (!(migrate->src[i] & MIGRATE_PFN_MIGRATE) || !newpage) {
3085 unlock_page(newpage);
3091 remove_migration_ptes(page, newpage, false);
3094 if (is_zone_device_page(page))
3097 putback_lru_page(page);
3099 if (newpage != page) {
3100 unlock_page(newpage);
3101 if (is_zone_device_page(newpage))
3104 putback_lru_page(newpage);
3108 EXPORT_SYMBOL(migrate_vma_finalize);
3109 #endif /* CONFIG_DEVICE_PRIVATE */