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>
52 #include <linux/memory.h>
53 #include <linux/random.h>
55 #include <asm/tlbflush.h>
57 #define CREATE_TRACE_POINTS
58 #include <trace/events/migrate.h>
62 int isolate_movable_page(struct page *page, isolate_mode_t mode)
64 struct address_space *mapping;
67 * Avoid burning cycles with pages that are yet under __free_pages(),
68 * or just got freed under us.
70 * In case we 'win' a race for a movable page being freed under us and
71 * raise its refcount preventing __free_pages() from doing its job
72 * the put_page() at the end of this block will take care of
73 * release this page, thus avoiding a nasty leakage.
75 if (unlikely(!get_page_unless_zero(page)))
79 * Check PageMovable before holding a PG_lock because page's owner
80 * assumes anybody doesn't touch PG_lock of newly allocated page
81 * so unconditionally grabbing the lock ruins page's owner side.
83 if (unlikely(!__PageMovable(page)))
86 * As movable pages are not isolated from LRU lists, concurrent
87 * compaction threads can race against page migration functions
88 * as well as race against the releasing a page.
90 * In order to avoid having an already isolated movable page
91 * being (wrongly) re-isolated while it is under migration,
92 * or to avoid attempting to isolate pages being released,
93 * lets be sure we have the page lock
94 * before proceeding with the movable page isolation steps.
96 if (unlikely(!trylock_page(page)))
99 if (!PageMovable(page) || PageIsolated(page))
100 goto out_no_isolated;
102 mapping = page_mapping(page);
103 VM_BUG_ON_PAGE(!mapping, page);
105 if (!mapping->a_ops->isolate_page(page, mode))
106 goto out_no_isolated;
108 /* Driver shouldn't use PG_isolated bit of page->flags */
109 WARN_ON_ONCE(PageIsolated(page));
110 __SetPageIsolated(page);
123 static void putback_movable_page(struct page *page)
125 struct address_space *mapping;
127 mapping = page_mapping(page);
128 mapping->a_ops->putback_page(page);
129 __ClearPageIsolated(page);
133 * Put previously isolated pages back onto the appropriate lists
134 * from where they were once taken off for compaction/migration.
136 * This function shall be used whenever the isolated pageset has been
137 * built from lru, balloon, hugetlbfs page. See isolate_migratepages_range()
138 * and isolate_huge_page().
140 void putback_movable_pages(struct list_head *l)
145 list_for_each_entry_safe(page, page2, l, lru) {
146 if (unlikely(PageHuge(page))) {
147 putback_active_hugepage(page);
150 list_del(&page->lru);
152 * We isolated non-lru movable page so here we can use
153 * __PageMovable because LRU page's mapping cannot have
154 * PAGE_MAPPING_MOVABLE.
156 if (unlikely(__PageMovable(page))) {
157 VM_BUG_ON_PAGE(!PageIsolated(page), page);
159 if (PageMovable(page))
160 putback_movable_page(page);
162 __ClearPageIsolated(page);
166 mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON +
167 page_is_file_lru(page), -thp_nr_pages(page));
168 putback_lru_page(page);
174 * Restore a potential migration pte to a working pte entry
176 static bool remove_migration_pte(struct page *page, struct vm_area_struct *vma,
177 unsigned long addr, void *old)
179 struct page_vma_mapped_walk pvmw = {
183 .flags = PVMW_SYNC | PVMW_MIGRATION,
189 VM_BUG_ON_PAGE(PageTail(page), page);
190 while (page_vma_mapped_walk(&pvmw)) {
194 new = page - pvmw.page->index +
195 linear_page_index(vma, pvmw.address);
197 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
198 /* PMD-mapped THP migration entry */
200 VM_BUG_ON_PAGE(PageHuge(page) || !PageTransCompound(page), page);
201 remove_migration_pmd(&pvmw, new);
207 pte = pte_mkold(mk_pte(new, READ_ONCE(vma->vm_page_prot)));
208 if (pte_swp_soft_dirty(*pvmw.pte))
209 pte = pte_mksoft_dirty(pte);
212 * Recheck VMA as permissions can change since migration started
214 entry = pte_to_swp_entry(*pvmw.pte);
215 if (is_writable_migration_entry(entry))
216 pte = maybe_mkwrite(pte, vma);
217 else if (pte_swp_uffd_wp(*pvmw.pte))
218 pte = pte_mkuffd_wp(pte);
220 if (unlikely(is_device_private_page(new))) {
222 entry = make_writable_device_private_entry(
225 entry = make_readable_device_private_entry(
227 pte = swp_entry_to_pte(entry);
228 if (pte_swp_soft_dirty(*pvmw.pte))
229 pte = pte_swp_mksoft_dirty(pte);
230 if (pte_swp_uffd_wp(*pvmw.pte))
231 pte = pte_swp_mkuffd_wp(pte);
234 #ifdef CONFIG_HUGETLB_PAGE
236 unsigned int shift = huge_page_shift(hstate_vma(vma));
238 pte = pte_mkhuge(pte);
239 pte = arch_make_huge_pte(pte, shift, vma->vm_flags);
241 hugepage_add_anon_rmap(new, vma, pvmw.address);
243 page_dup_rmap(new, true);
244 set_huge_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte);
249 page_add_anon_rmap(new, vma, pvmw.address, false);
251 page_add_file_rmap(new, false);
252 set_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte);
254 if (vma->vm_flags & VM_LOCKED && !PageTransCompound(new))
257 if (PageTransHuge(page) && PageMlocked(page))
258 clear_page_mlock(page);
260 /* No need to invalidate - it was non-present before */
261 update_mmu_cache(vma, pvmw.address, pvmw.pte);
268 * Get rid of all migration entries and replace them by
269 * references to the indicated page.
271 void remove_migration_ptes(struct page *old, struct page *new, bool locked)
273 struct rmap_walk_control rwc = {
274 .rmap_one = remove_migration_pte,
279 rmap_walk_locked(new, &rwc);
281 rmap_walk(new, &rwc);
285 * Something used the pte of a page under migration. We need to
286 * get to the page and wait until migration is finished.
287 * When we return from this function the fault will be retried.
289 void __migration_entry_wait(struct mm_struct *mm, pte_t *ptep,
298 if (!is_swap_pte(pte))
301 entry = pte_to_swp_entry(pte);
302 if (!is_migration_entry(entry))
305 folio = page_folio(pfn_swap_entry_to_page(entry));
308 * Once page cache replacement of page migration started, page_count
309 * is zero; but we must not call folio_put_wait_locked() without
310 * a ref. Use folio_try_get(), and just fault again if it fails.
312 if (!folio_try_get(folio))
314 pte_unmap_unlock(ptep, ptl);
315 folio_put_wait_locked(folio, TASK_UNINTERRUPTIBLE);
318 pte_unmap_unlock(ptep, ptl);
321 void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
322 unsigned long address)
324 spinlock_t *ptl = pte_lockptr(mm, pmd);
325 pte_t *ptep = pte_offset_map(pmd, address);
326 __migration_entry_wait(mm, ptep, ptl);
329 void migration_entry_wait_huge(struct vm_area_struct *vma,
330 struct mm_struct *mm, pte_t *pte)
332 spinlock_t *ptl = huge_pte_lockptr(hstate_vma(vma), mm, pte);
333 __migration_entry_wait(mm, pte, ptl);
336 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
337 void pmd_migration_entry_wait(struct mm_struct *mm, pmd_t *pmd)
342 ptl = pmd_lock(mm, pmd);
343 if (!is_pmd_migration_entry(*pmd))
345 folio = page_folio(pfn_swap_entry_to_page(pmd_to_swp_entry(*pmd)));
346 if (!folio_try_get(folio))
349 folio_put_wait_locked(folio, TASK_UNINTERRUPTIBLE);
356 static int expected_page_refs(struct address_space *mapping, struct page *page)
358 int expected_count = 1;
361 * Device private pages have an extra refcount as they are
364 expected_count += is_device_private_page(page);
366 expected_count += compound_nr(page) + page_has_private(page);
368 return expected_count;
372 * Replace the page in the mapping.
374 * The number of remaining references must be:
375 * 1 for anonymous pages without a mapping
376 * 2 for pages with a mapping
377 * 3 for pages with a mapping and PagePrivate/PagePrivate2 set.
379 int folio_migrate_mapping(struct address_space *mapping,
380 struct folio *newfolio, struct folio *folio, int extra_count)
382 XA_STATE(xas, &mapping->i_pages, folio_index(folio));
383 struct zone *oldzone, *newzone;
385 int expected_count = expected_page_refs(mapping, &folio->page) + extra_count;
386 long nr = folio_nr_pages(folio);
389 /* Anonymous page without mapping */
390 if (folio_ref_count(folio) != expected_count)
393 /* No turning back from here */
394 newfolio->index = folio->index;
395 newfolio->mapping = folio->mapping;
396 if (folio_test_swapbacked(folio))
397 __folio_set_swapbacked(newfolio);
399 return MIGRATEPAGE_SUCCESS;
402 oldzone = folio_zone(folio);
403 newzone = folio_zone(newfolio);
406 if (!folio_ref_freeze(folio, expected_count)) {
407 xas_unlock_irq(&xas);
412 * Now we know that no one else is looking at the folio:
413 * no turning back from here.
415 newfolio->index = folio->index;
416 newfolio->mapping = folio->mapping;
417 folio_ref_add(newfolio, nr); /* add cache reference */
418 if (folio_test_swapbacked(folio)) {
419 __folio_set_swapbacked(newfolio);
420 if (folio_test_swapcache(folio)) {
421 folio_set_swapcache(newfolio);
422 newfolio->private = folio_get_private(folio);
425 VM_BUG_ON_FOLIO(folio_test_swapcache(folio), folio);
428 /* Move dirty while page refs frozen and newpage not yet exposed */
429 dirty = folio_test_dirty(folio);
431 folio_clear_dirty(folio);
432 folio_set_dirty(newfolio);
435 xas_store(&xas, newfolio);
438 * Drop cache reference from old page by unfreezing
439 * to one less reference.
440 * We know this isn't the last reference.
442 folio_ref_unfreeze(folio, expected_count - nr);
445 /* Leave irq disabled to prevent preemption while updating stats */
448 * If moved to a different zone then also account
449 * the page for that zone. Other VM counters will be
450 * taken care of when we establish references to the
451 * new page and drop references to the old page.
453 * Note that anonymous pages are accounted for
454 * via NR_FILE_PAGES and NR_ANON_MAPPED if they
455 * are mapped to swap space.
457 if (newzone != oldzone) {
458 struct lruvec *old_lruvec, *new_lruvec;
459 struct mem_cgroup *memcg;
461 memcg = folio_memcg(folio);
462 old_lruvec = mem_cgroup_lruvec(memcg, oldzone->zone_pgdat);
463 new_lruvec = mem_cgroup_lruvec(memcg, newzone->zone_pgdat);
465 __mod_lruvec_state(old_lruvec, NR_FILE_PAGES, -nr);
466 __mod_lruvec_state(new_lruvec, NR_FILE_PAGES, nr);
467 if (folio_test_swapbacked(folio) && !folio_test_swapcache(folio)) {
468 __mod_lruvec_state(old_lruvec, NR_SHMEM, -nr);
469 __mod_lruvec_state(new_lruvec, NR_SHMEM, nr);
472 if (folio_test_swapcache(folio)) {
473 __mod_lruvec_state(old_lruvec, NR_SWAPCACHE, -nr);
474 __mod_lruvec_state(new_lruvec, NR_SWAPCACHE, nr);
477 if (dirty && mapping_can_writeback(mapping)) {
478 __mod_lruvec_state(old_lruvec, NR_FILE_DIRTY, -nr);
479 __mod_zone_page_state(oldzone, NR_ZONE_WRITE_PENDING, -nr);
480 __mod_lruvec_state(new_lruvec, NR_FILE_DIRTY, nr);
481 __mod_zone_page_state(newzone, NR_ZONE_WRITE_PENDING, nr);
486 return MIGRATEPAGE_SUCCESS;
488 EXPORT_SYMBOL(folio_migrate_mapping);
491 * The expected number of remaining references is the same as that
492 * of folio_migrate_mapping().
494 int migrate_huge_page_move_mapping(struct address_space *mapping,
495 struct page *newpage, struct page *page)
497 XA_STATE(xas, &mapping->i_pages, page_index(page));
501 expected_count = 2 + page_has_private(page);
502 if (page_count(page) != expected_count || xas_load(&xas) != page) {
503 xas_unlock_irq(&xas);
507 if (!page_ref_freeze(page, expected_count)) {
508 xas_unlock_irq(&xas);
512 newpage->index = page->index;
513 newpage->mapping = page->mapping;
517 xas_store(&xas, newpage);
519 page_ref_unfreeze(page, expected_count - 1);
521 xas_unlock_irq(&xas);
523 return MIGRATEPAGE_SUCCESS;
527 * Copy the flags and some other ancillary information
529 void folio_migrate_flags(struct folio *newfolio, struct folio *folio)
533 if (folio_test_error(folio))
534 folio_set_error(newfolio);
535 if (folio_test_referenced(folio))
536 folio_set_referenced(newfolio);
537 if (folio_test_uptodate(folio))
538 folio_mark_uptodate(newfolio);
539 if (folio_test_clear_active(folio)) {
540 VM_BUG_ON_FOLIO(folio_test_unevictable(folio), folio);
541 folio_set_active(newfolio);
542 } else if (folio_test_clear_unevictable(folio))
543 folio_set_unevictable(newfolio);
544 if (folio_test_workingset(folio))
545 folio_set_workingset(newfolio);
546 if (folio_test_checked(folio))
547 folio_set_checked(newfolio);
548 if (folio_test_mappedtodisk(folio))
549 folio_set_mappedtodisk(newfolio);
551 /* Move dirty on pages not done by folio_migrate_mapping() */
552 if (folio_test_dirty(folio))
553 folio_set_dirty(newfolio);
555 if (folio_test_young(folio))
556 folio_set_young(newfolio);
557 if (folio_test_idle(folio))
558 folio_set_idle(newfolio);
561 * Copy NUMA information to the new page, to prevent over-eager
562 * future migrations of this same page.
564 cpupid = page_cpupid_xchg_last(&folio->page, -1);
565 page_cpupid_xchg_last(&newfolio->page, cpupid);
567 folio_migrate_ksm(newfolio, folio);
569 * Please do not reorder this without considering how mm/ksm.c's
570 * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache().
572 if (folio_test_swapcache(folio))
573 folio_clear_swapcache(folio);
574 folio_clear_private(folio);
576 /* page->private contains hugetlb specific flags */
577 if (!folio_test_hugetlb(folio))
578 folio->private = NULL;
581 * If any waiters have accumulated on the new page then
584 if (folio_test_writeback(newfolio))
585 folio_end_writeback(newfolio);
588 * PG_readahead shares the same bit with PG_reclaim. The above
589 * end_page_writeback() may clear PG_readahead mistakenly, so set the
592 if (folio_test_readahead(folio))
593 folio_set_readahead(newfolio);
595 folio_copy_owner(newfolio, folio);
597 if (!folio_test_hugetlb(folio))
598 mem_cgroup_migrate(folio, newfolio);
600 EXPORT_SYMBOL(folio_migrate_flags);
602 void folio_migrate_copy(struct folio *newfolio, struct folio *folio)
604 folio_copy(newfolio, folio);
605 folio_migrate_flags(newfolio, folio);
607 EXPORT_SYMBOL(folio_migrate_copy);
609 /************************************************************
610 * Migration functions
611 ***********************************************************/
614 * Common logic to directly migrate a single LRU page suitable for
615 * pages that do not use PagePrivate/PagePrivate2.
617 * Pages are locked upon entry and exit.
619 int migrate_page(struct address_space *mapping,
620 struct page *newpage, struct page *page,
621 enum migrate_mode mode)
623 struct folio *newfolio = page_folio(newpage);
624 struct folio *folio = page_folio(page);
627 BUG_ON(folio_test_writeback(folio)); /* Writeback must be complete */
629 rc = folio_migrate_mapping(mapping, newfolio, folio, 0);
631 if (rc != MIGRATEPAGE_SUCCESS)
634 if (mode != MIGRATE_SYNC_NO_COPY)
635 folio_migrate_copy(newfolio, folio);
637 folio_migrate_flags(newfolio, folio);
638 return MIGRATEPAGE_SUCCESS;
640 EXPORT_SYMBOL(migrate_page);
643 /* Returns true if all buffers are successfully locked */
644 static bool buffer_migrate_lock_buffers(struct buffer_head *head,
645 enum migrate_mode mode)
647 struct buffer_head *bh = head;
649 /* Simple case, sync compaction */
650 if (mode != MIGRATE_ASYNC) {
653 bh = bh->b_this_page;
655 } while (bh != head);
660 /* async case, we cannot block on lock_buffer so use trylock_buffer */
662 if (!trylock_buffer(bh)) {
664 * We failed to lock the buffer and cannot stall in
665 * async migration. Release the taken locks
667 struct buffer_head *failed_bh = bh;
669 while (bh != failed_bh) {
671 bh = bh->b_this_page;
676 bh = bh->b_this_page;
677 } while (bh != head);
681 static int __buffer_migrate_page(struct address_space *mapping,
682 struct page *newpage, struct page *page, enum migrate_mode mode,
685 struct buffer_head *bh, *head;
689 if (!page_has_buffers(page))
690 return migrate_page(mapping, newpage, page, mode);
692 /* Check whether page does not have extra refs before we do more work */
693 expected_count = expected_page_refs(mapping, page);
694 if (page_count(page) != expected_count)
697 head = page_buffers(page);
698 if (!buffer_migrate_lock_buffers(head, mode))
703 bool invalidated = false;
707 spin_lock(&mapping->private_lock);
710 if (atomic_read(&bh->b_count)) {
714 bh = bh->b_this_page;
715 } while (bh != head);
721 spin_unlock(&mapping->private_lock);
722 invalidate_bh_lrus();
724 goto recheck_buffers;
728 rc = migrate_page_move_mapping(mapping, newpage, page, 0);
729 if (rc != MIGRATEPAGE_SUCCESS)
732 attach_page_private(newpage, detach_page_private(page));
736 set_bh_page(bh, newpage, bh_offset(bh));
737 bh = bh->b_this_page;
739 } while (bh != head);
741 if (mode != MIGRATE_SYNC_NO_COPY)
742 migrate_page_copy(newpage, page);
744 migrate_page_states(newpage, page);
746 rc = MIGRATEPAGE_SUCCESS;
749 spin_unlock(&mapping->private_lock);
753 bh = bh->b_this_page;
755 } while (bh != head);
761 * Migration function for pages with buffers. This function can only be used
762 * if the underlying filesystem guarantees that no other references to "page"
763 * exist. For example attached buffer heads are accessed only under page lock.
765 int buffer_migrate_page(struct address_space *mapping,
766 struct page *newpage, struct page *page, enum migrate_mode mode)
768 return __buffer_migrate_page(mapping, newpage, page, mode, false);
770 EXPORT_SYMBOL(buffer_migrate_page);
773 * Same as above except that this variant is more careful and checks that there
774 * are also no buffer head references. This function is the right one for
775 * mappings where buffer heads are directly looked up and referenced (such as
776 * block device mappings).
778 int buffer_migrate_page_norefs(struct address_space *mapping,
779 struct page *newpage, struct page *page, enum migrate_mode mode)
781 return __buffer_migrate_page(mapping, newpage, page, mode, true);
786 * Writeback a page to clean the dirty state
788 static int writeout(struct address_space *mapping, struct page *page)
790 struct writeback_control wbc = {
791 .sync_mode = WB_SYNC_NONE,
794 .range_end = LLONG_MAX,
799 if (!mapping->a_ops->writepage)
800 /* No write method for the address space */
803 if (!clear_page_dirty_for_io(page))
804 /* Someone else already triggered a write */
808 * A dirty page may imply that the underlying filesystem has
809 * the page on some queue. So the page must be clean for
810 * migration. Writeout may mean we loose the lock and the
811 * page state is no longer what we checked for earlier.
812 * At this point we know that the migration attempt cannot
815 remove_migration_ptes(page, page, false);
817 rc = mapping->a_ops->writepage(page, &wbc);
819 if (rc != AOP_WRITEPAGE_ACTIVATE)
820 /* unlocked. Relock */
823 return (rc < 0) ? -EIO : -EAGAIN;
827 * Default handling if a filesystem does not provide a migration function.
829 static int fallback_migrate_page(struct address_space *mapping,
830 struct page *newpage, struct page *page, enum migrate_mode mode)
832 if (PageDirty(page)) {
833 /* Only writeback pages in full synchronous migration */
836 case MIGRATE_SYNC_NO_COPY:
841 return writeout(mapping, page);
845 * Buffers may be managed in a filesystem specific way.
846 * We must have no buffers or drop them.
848 if (page_has_private(page) &&
849 !try_to_release_page(page, GFP_KERNEL))
850 return mode == MIGRATE_SYNC ? -EAGAIN : -EBUSY;
852 return migrate_page(mapping, newpage, page, mode);
856 * Move a page to a newly allocated page
857 * The page is locked and all ptes have been successfully removed.
859 * The new page will have replaced the old page if this function
864 * MIGRATEPAGE_SUCCESS - success
866 static int move_to_new_page(struct page *newpage, struct page *page,
867 enum migrate_mode mode)
869 struct address_space *mapping;
871 bool is_lru = !__PageMovable(page);
873 VM_BUG_ON_PAGE(!PageLocked(page), page);
874 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
876 mapping = page_mapping(page);
878 if (likely(is_lru)) {
880 rc = migrate_page(mapping, newpage, page, mode);
881 else if (mapping->a_ops->migratepage)
883 * Most pages have a mapping and most filesystems
884 * provide a migratepage callback. Anonymous pages
885 * are part of swap space which also has its own
886 * migratepage callback. This is the most common path
887 * for page migration.
889 rc = mapping->a_ops->migratepage(mapping, newpage,
892 rc = fallback_migrate_page(mapping, newpage,
896 * In case of non-lru page, it could be released after
897 * isolation step. In that case, we shouldn't try migration.
899 VM_BUG_ON_PAGE(!PageIsolated(page), page);
900 if (!PageMovable(page)) {
901 rc = MIGRATEPAGE_SUCCESS;
902 __ClearPageIsolated(page);
906 rc = mapping->a_ops->migratepage(mapping, newpage,
908 WARN_ON_ONCE(rc == MIGRATEPAGE_SUCCESS &&
909 !PageIsolated(page));
913 * When successful, old pagecache page->mapping must be cleared before
914 * page is freed; but stats require that PageAnon be left as PageAnon.
916 if (rc == MIGRATEPAGE_SUCCESS) {
917 if (__PageMovable(page)) {
918 VM_BUG_ON_PAGE(!PageIsolated(page), page);
921 * We clear PG_movable under page_lock so any compactor
922 * cannot try to migrate this page.
924 __ClearPageIsolated(page);
928 * Anonymous and movable page->mapping will be cleared by
929 * free_pages_prepare so don't reset it here for keeping
930 * the type to work PageAnon, for example.
932 if (!PageMappingFlags(page))
933 page->mapping = NULL;
935 if (likely(!is_zone_device_page(newpage)))
936 flush_dcache_page(newpage);
943 static int __unmap_and_move(struct page *page, struct page *newpage,
944 int force, enum migrate_mode mode)
947 bool page_was_mapped = false;
948 struct anon_vma *anon_vma = NULL;
949 bool is_lru = !__PageMovable(page);
951 if (!trylock_page(page)) {
952 if (!force || mode == MIGRATE_ASYNC)
956 * It's not safe for direct compaction to call lock_page.
957 * For example, during page readahead pages are added locked
958 * to the LRU. Later, when the IO completes the pages are
959 * marked uptodate and unlocked. However, the queueing
960 * could be merging multiple pages for one bio (e.g.
961 * mpage_readahead). If an allocation happens for the
962 * second or third page, the process can end up locking
963 * the same page twice and deadlocking. Rather than
964 * trying to be clever about what pages can be locked,
965 * avoid the use of lock_page for direct compaction
968 if (current->flags & PF_MEMALLOC)
974 if (PageWriteback(page)) {
976 * Only in the case of a full synchronous migration is it
977 * necessary to wait for PageWriteback. In the async case,
978 * the retry loop is too short and in the sync-light case,
979 * the overhead of stalling is too much
983 case MIGRATE_SYNC_NO_COPY:
991 wait_on_page_writeback(page);
995 * By try_to_migrate(), page->mapcount goes down to 0 here. In this case,
996 * we cannot notice that anon_vma is freed while we migrates a page.
997 * This get_anon_vma() delays freeing anon_vma pointer until the end
998 * of migration. File cache pages are no problem because of page_lock()
999 * File Caches may use write_page() or lock_page() in migration, then,
1000 * just care Anon page here.
1002 * Only page_get_anon_vma() understands the subtleties of
1003 * getting a hold on an anon_vma from outside one of its mms.
1004 * But if we cannot get anon_vma, then we won't need it anyway,
1005 * because that implies that the anon page is no longer mapped
1006 * (and cannot be remapped so long as we hold the page lock).
1008 if (PageAnon(page) && !PageKsm(page))
1009 anon_vma = page_get_anon_vma(page);
1012 * Block others from accessing the new page when we get around to
1013 * establishing additional references. We are usually the only one
1014 * holding a reference to newpage at this point. We used to have a BUG
1015 * here if trylock_page(newpage) fails, but would like to allow for
1016 * cases where there might be a race with the previous use of newpage.
1017 * This is much like races on refcount of oldpage: just don't BUG().
1019 if (unlikely(!trylock_page(newpage)))
1022 if (unlikely(!is_lru)) {
1023 rc = move_to_new_page(newpage, page, mode);
1024 goto out_unlock_both;
1028 * Corner case handling:
1029 * 1. When a new swap-cache page is read into, it is added to the LRU
1030 * and treated as swapcache but it has no rmap yet.
1031 * Calling try_to_unmap() against a page->mapping==NULL page will
1032 * trigger a BUG. So handle it here.
1033 * 2. An orphaned page (see truncate_cleanup_page) might have
1034 * fs-private metadata. The page can be picked up due to memory
1035 * offlining. Everywhere else except page reclaim, the page is
1036 * invisible to the vm, so the page can not be migrated. So try to
1037 * free the metadata, so the page can be freed.
1039 if (!page->mapping) {
1040 VM_BUG_ON_PAGE(PageAnon(page), page);
1041 if (page_has_private(page)) {
1042 try_to_free_buffers(page);
1043 goto out_unlock_both;
1045 } else if (page_mapped(page)) {
1046 /* Establish migration ptes */
1047 VM_BUG_ON_PAGE(PageAnon(page) && !PageKsm(page) && !anon_vma,
1049 try_to_migrate(page, 0);
1050 page_was_mapped = true;
1053 if (!page_mapped(page))
1054 rc = move_to_new_page(newpage, page, mode);
1056 if (page_was_mapped)
1057 remove_migration_ptes(page,
1058 rc == MIGRATEPAGE_SUCCESS ? newpage : page, false);
1061 unlock_page(newpage);
1063 /* Drop an anon_vma reference if we took one */
1065 put_anon_vma(anon_vma);
1069 * If migration is successful, decrease refcount of the newpage
1070 * which will not free the page because new page owner increased
1071 * refcounter. As well, if it is LRU page, add the page to LRU
1072 * list in here. Use the old state of the isolated source page to
1073 * determine if we migrated a LRU page. newpage was already unlocked
1074 * and possibly modified by its owner - don't rely on the page
1077 if (rc == MIGRATEPAGE_SUCCESS) {
1078 if (unlikely(!is_lru))
1081 putback_lru_page(newpage);
1088 * Obtain the lock on page, remove all ptes and migrate the page
1089 * to the newly allocated page in newpage.
1091 static int unmap_and_move(new_page_t get_new_page,
1092 free_page_t put_new_page,
1093 unsigned long private, struct page *page,
1094 int force, enum migrate_mode mode,
1095 enum migrate_reason reason,
1096 struct list_head *ret)
1098 int rc = MIGRATEPAGE_SUCCESS;
1099 struct page *newpage = NULL;
1101 if (!thp_migration_supported() && PageTransHuge(page))
1104 if (page_count(page) == 1) {
1105 /* page was freed from under us. So we are done. */
1106 ClearPageActive(page);
1107 ClearPageUnevictable(page);
1108 if (unlikely(__PageMovable(page))) {
1110 if (!PageMovable(page))
1111 __ClearPageIsolated(page);
1117 newpage = get_new_page(page, private);
1121 rc = __unmap_and_move(page, newpage, force, mode);
1122 if (rc == MIGRATEPAGE_SUCCESS)
1123 set_page_owner_migrate_reason(newpage, reason);
1126 if (rc != -EAGAIN) {
1128 * A page that has been migrated has all references
1129 * removed and will be freed. A page that has not been
1130 * migrated will have kept its references and be restored.
1132 list_del(&page->lru);
1136 * If migration is successful, releases reference grabbed during
1137 * isolation. Otherwise, restore the page to right list unless
1140 if (rc == MIGRATEPAGE_SUCCESS) {
1142 * Compaction can migrate also non-LRU pages which are
1143 * not accounted to NR_ISOLATED_*. They can be recognized
1146 if (likely(!__PageMovable(page)))
1147 mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON +
1148 page_is_file_lru(page), -thp_nr_pages(page));
1150 if (reason != MR_MEMORY_FAILURE)
1152 * We release the page in page_handle_poison.
1157 list_add_tail(&page->lru, ret);
1160 put_new_page(newpage, private);
1169 * Counterpart of unmap_and_move_page() for hugepage migration.
1171 * This function doesn't wait the completion of hugepage I/O
1172 * because there is no race between I/O and migration for hugepage.
1173 * Note that currently hugepage I/O occurs only in direct I/O
1174 * where no lock is held and PG_writeback is irrelevant,
1175 * and writeback status of all subpages are counted in the reference
1176 * count of the head page (i.e. if all subpages of a 2MB hugepage are
1177 * under direct I/O, the reference of the head page is 512 and a bit more.)
1178 * This means that when we try to migrate hugepage whose subpages are
1179 * doing direct I/O, some references remain after try_to_unmap() and
1180 * hugepage migration fails without data corruption.
1182 * There is also no race when direct I/O is issued on the page under migration,
1183 * because then pte is replaced with migration swap entry and direct I/O code
1184 * will wait in the page fault for migration to complete.
1186 static int unmap_and_move_huge_page(new_page_t get_new_page,
1187 free_page_t put_new_page, unsigned long private,
1188 struct page *hpage, int force,
1189 enum migrate_mode mode, int reason,
1190 struct list_head *ret)
1193 int page_was_mapped = 0;
1194 struct page *new_hpage;
1195 struct anon_vma *anon_vma = NULL;
1196 struct address_space *mapping = NULL;
1199 * Migratability of hugepages depends on architectures and their size.
1200 * This check is necessary because some callers of hugepage migration
1201 * like soft offline and memory hotremove don't walk through page
1202 * tables or check whether the hugepage is pmd-based or not before
1203 * kicking migration.
1205 if (!hugepage_migration_supported(page_hstate(hpage))) {
1206 list_move_tail(&hpage->lru, ret);
1210 if (page_count(hpage) == 1) {
1211 /* page was freed from under us. So we are done. */
1212 putback_active_hugepage(hpage);
1213 return MIGRATEPAGE_SUCCESS;
1216 new_hpage = get_new_page(hpage, private);
1220 if (!trylock_page(hpage)) {
1225 case MIGRATE_SYNC_NO_COPY:
1234 * Check for pages which are in the process of being freed. Without
1235 * page_mapping() set, hugetlbfs specific move page routine will not
1236 * be called and we could leak usage counts for subpools.
1238 if (hugetlb_page_subpool(hpage) && !page_mapping(hpage)) {
1243 if (PageAnon(hpage))
1244 anon_vma = page_get_anon_vma(hpage);
1246 if (unlikely(!trylock_page(new_hpage)))
1249 if (page_mapped(hpage)) {
1250 bool mapping_locked = false;
1251 enum ttu_flags ttu = 0;
1253 if (!PageAnon(hpage)) {
1255 * In shared mappings, try_to_unmap could potentially
1256 * call huge_pmd_unshare. Because of this, take
1257 * semaphore in write mode here and set TTU_RMAP_LOCKED
1258 * to let lower levels know we have taken the lock.
1260 mapping = hugetlb_page_mapping_lock_write(hpage);
1261 if (unlikely(!mapping))
1262 goto unlock_put_anon;
1264 mapping_locked = true;
1265 ttu |= TTU_RMAP_LOCKED;
1268 try_to_migrate(hpage, ttu);
1269 page_was_mapped = 1;
1272 i_mmap_unlock_write(mapping);
1275 if (!page_mapped(hpage))
1276 rc = move_to_new_page(new_hpage, hpage, mode);
1278 if (page_was_mapped)
1279 remove_migration_ptes(hpage,
1280 rc == MIGRATEPAGE_SUCCESS ? new_hpage : hpage, false);
1283 unlock_page(new_hpage);
1287 put_anon_vma(anon_vma);
1289 if (rc == MIGRATEPAGE_SUCCESS) {
1290 move_hugetlb_state(hpage, new_hpage, reason);
1291 put_new_page = NULL;
1297 if (rc == MIGRATEPAGE_SUCCESS)
1298 putback_active_hugepage(hpage);
1299 else if (rc != -EAGAIN)
1300 list_move_tail(&hpage->lru, ret);
1303 * If migration was not successful and there's a freeing callback, use
1304 * it. Otherwise, put_page() will drop the reference grabbed during
1308 put_new_page(new_hpage, private);
1310 putback_active_hugepage(new_hpage);
1315 static inline int try_split_thp(struct page *page, struct page **page2,
1316 struct list_head *from)
1321 rc = split_huge_page_to_list(page, from);
1324 list_safe_reset_next(page, *page2, lru);
1330 * migrate_pages - migrate the pages specified in a list, to the free pages
1331 * supplied as the target for the page migration
1333 * @from: The list of pages to be migrated.
1334 * @get_new_page: The function used to allocate free pages to be used
1335 * as the target of the page migration.
1336 * @put_new_page: The function used to free target pages if migration
1337 * fails, or NULL if no special handling is necessary.
1338 * @private: Private data to be passed on to get_new_page()
1339 * @mode: The migration mode that specifies the constraints for
1340 * page migration, if any.
1341 * @reason: The reason for page migration.
1342 * @ret_succeeded: Set to the number of normal pages migrated successfully if
1343 * the caller passes a non-NULL pointer.
1345 * The function returns after 10 attempts or if no pages are movable any more
1346 * because the list has become empty or no retryable pages exist any more.
1347 * It is caller's responsibility to call putback_movable_pages() to return pages
1348 * to the LRU or free list only if ret != 0.
1350 * Returns the number of {normal page, THP, hugetlb} that were not migrated, or
1351 * an error code. The number of THP splits will be considered as the number of
1352 * non-migrated THP, no matter how many subpages of the THP are migrated successfully.
1354 int migrate_pages(struct list_head *from, new_page_t get_new_page,
1355 free_page_t put_new_page, unsigned long private,
1356 enum migrate_mode mode, int reason, unsigned int *ret_succeeded)
1361 int nr_failed_pages = 0;
1362 int nr_succeeded = 0;
1363 int nr_thp_succeeded = 0;
1364 int nr_thp_failed = 0;
1365 int nr_thp_split = 0;
1367 bool is_thp = false;
1370 int swapwrite = current->flags & PF_SWAPWRITE;
1371 int rc, nr_subpages;
1372 LIST_HEAD(ret_pages);
1373 LIST_HEAD(thp_split_pages);
1374 bool nosplit = (reason == MR_NUMA_MISPLACED);
1375 bool no_subpage_counting = false;
1377 trace_mm_migrate_pages_start(mode, reason);
1380 current->flags |= PF_SWAPWRITE;
1382 thp_subpage_migration:
1383 for (pass = 0; pass < 10 && (retry || thp_retry); pass++) {
1387 list_for_each_entry_safe(page, page2, from, lru) {
1390 * THP statistics is based on the source huge page.
1391 * Capture required information that might get lost
1394 is_thp = PageTransHuge(page) && !PageHuge(page);
1395 nr_subpages = compound_nr(page);
1399 rc = unmap_and_move_huge_page(get_new_page,
1400 put_new_page, private, page,
1401 pass > 2, mode, reason,
1404 rc = unmap_and_move(get_new_page, put_new_page,
1405 private, page, pass > 2, mode,
1406 reason, &ret_pages);
1409 * Success: non hugetlb page will be freed, hugetlb
1410 * page will be put back
1411 * -EAGAIN: stay on the from list
1412 * -ENOMEM: stay on the from list
1413 * Other errno: put on ret_pages list then splice to
1418 * THP migration might be unsupported or the
1419 * allocation could've failed so we should
1420 * retry on the same page with the THP split
1423 * Head page is retried immediately and tail
1424 * pages are added to the tail of the list so
1425 * we encounter them after the rest of the list
1429 /* THP migration is unsupported */
1432 if (!try_split_thp(page, &page2, &thp_split_pages)) {
1437 nr_failed_pages += nr_subpages;
1441 /* Hugetlb migration is unsupported */
1442 if (!no_subpage_counting)
1444 nr_failed_pages += nr_subpages;
1448 * When memory is low, don't bother to try to migrate
1449 * other pages, just exit.
1450 * THP NUMA faulting doesn't split THP to retry.
1452 if (is_thp && !nosplit) {
1454 if (!try_split_thp(page, &page2, &thp_split_pages)) {
1459 nr_failed_pages += nr_subpages;
1463 if (!no_subpage_counting)
1465 nr_failed_pages += nr_subpages;
1474 case MIGRATEPAGE_SUCCESS:
1475 nr_succeeded += nr_subpages;
1483 * Permanent failure (-EBUSY, etc.):
1484 * unlike -EAGAIN case, the failed page is
1485 * removed from migration page list and not
1486 * retried in the next outer loop.
1490 nr_failed_pages += nr_subpages;
1494 if (!no_subpage_counting)
1496 nr_failed_pages += nr_subpages;
1502 nr_thp_failed += thp_retry;
1504 * Try to migrate subpages of fail-to-migrate THPs, no nr_failed
1505 * counting in this round, since all subpages of a THP is counted
1506 * as 1 failure in the first round.
1508 if (!list_empty(&thp_split_pages)) {
1510 * Move non-migrated pages (after 10 retries) to ret_pages
1511 * to avoid migrating them again.
1513 list_splice_init(from, &ret_pages);
1514 list_splice_init(&thp_split_pages, from);
1515 no_subpage_counting = true;
1517 goto thp_subpage_migration;
1520 rc = nr_failed + nr_thp_failed;
1523 * Put the permanent failure page back to migration list, they
1524 * will be put back to the right list by the caller.
1526 list_splice(&ret_pages, from);
1528 count_vm_events(PGMIGRATE_SUCCESS, nr_succeeded);
1529 count_vm_events(PGMIGRATE_FAIL, nr_failed_pages);
1530 count_vm_events(THP_MIGRATION_SUCCESS, nr_thp_succeeded);
1531 count_vm_events(THP_MIGRATION_FAIL, nr_thp_failed);
1532 count_vm_events(THP_MIGRATION_SPLIT, nr_thp_split);
1533 trace_mm_migrate_pages(nr_succeeded, nr_failed_pages, nr_thp_succeeded,
1534 nr_thp_failed, nr_thp_split, mode, reason);
1537 current->flags &= ~PF_SWAPWRITE;
1540 *ret_succeeded = nr_succeeded;
1545 struct page *alloc_migration_target(struct page *page, unsigned long private)
1547 struct migration_target_control *mtc;
1549 unsigned int order = 0;
1550 struct page *new_page = NULL;
1554 mtc = (struct migration_target_control *)private;
1555 gfp_mask = mtc->gfp_mask;
1557 if (nid == NUMA_NO_NODE)
1558 nid = page_to_nid(page);
1560 if (PageHuge(page)) {
1561 struct hstate *h = page_hstate(compound_head(page));
1563 gfp_mask = htlb_modify_alloc_mask(h, gfp_mask);
1564 return alloc_huge_page_nodemask(h, nid, mtc->nmask, gfp_mask);
1567 if (PageTransHuge(page)) {
1569 * clear __GFP_RECLAIM to make the migration callback
1570 * consistent with regular THP allocations.
1572 gfp_mask &= ~__GFP_RECLAIM;
1573 gfp_mask |= GFP_TRANSHUGE;
1574 order = HPAGE_PMD_ORDER;
1576 zidx = zone_idx(page_zone(page));
1577 if (is_highmem_idx(zidx) || zidx == ZONE_MOVABLE)
1578 gfp_mask |= __GFP_HIGHMEM;
1580 new_page = __alloc_pages(gfp_mask, order, nid, mtc->nmask);
1582 if (new_page && PageTransHuge(new_page))
1583 prep_transhuge_page(new_page);
1590 static int store_status(int __user *status, int start, int value, int nr)
1593 if (put_user(value, status + start))
1601 static int do_move_pages_to_node(struct mm_struct *mm,
1602 struct list_head *pagelist, int node)
1605 struct migration_target_control mtc = {
1607 .gfp_mask = GFP_HIGHUSER_MOVABLE | __GFP_THISNODE,
1610 err = migrate_pages(pagelist, alloc_migration_target, NULL,
1611 (unsigned long)&mtc, MIGRATE_SYNC, MR_SYSCALL, NULL);
1613 putback_movable_pages(pagelist);
1618 * Resolves the given address to a struct page, isolates it from the LRU and
1619 * puts it to the given pagelist.
1621 * errno - if the page cannot be found/isolated
1622 * 0 - when it doesn't have to be migrated because it is already on the
1624 * 1 - when it has been queued
1626 static int add_page_for_migration(struct mm_struct *mm, unsigned long addr,
1627 int node, struct list_head *pagelist, bool migrate_all)
1629 struct vm_area_struct *vma;
1631 unsigned int follflags;
1636 vma = find_vma(mm, addr);
1637 if (!vma || addr < vma->vm_start || !vma_migratable(vma))
1640 /* FOLL_DUMP to ignore special (like zero) pages */
1641 follflags = FOLL_GET | FOLL_DUMP;
1642 page = follow_page(vma, addr, follflags);
1644 err = PTR_ERR(page);
1653 if (page_to_nid(page) == node)
1657 if (page_mapcount(page) > 1 && !migrate_all)
1660 if (PageHuge(page)) {
1661 if (PageHead(page)) {
1662 isolate_huge_page(page, pagelist);
1668 head = compound_head(page);
1669 err = isolate_lru_page(head);
1674 list_add_tail(&head->lru, pagelist);
1675 mod_node_page_state(page_pgdat(head),
1676 NR_ISOLATED_ANON + page_is_file_lru(head),
1677 thp_nr_pages(head));
1681 * Either remove the duplicate refcount from
1682 * isolate_lru_page() or drop the page ref if it was
1687 mmap_read_unlock(mm);
1691 static int move_pages_and_store_status(struct mm_struct *mm, int node,
1692 struct list_head *pagelist, int __user *status,
1693 int start, int i, unsigned long nr_pages)
1697 if (list_empty(pagelist))
1700 err = do_move_pages_to_node(mm, pagelist, node);
1703 * Positive err means the number of failed
1704 * pages to migrate. Since we are going to
1705 * abort and return the number of non-migrated
1706 * pages, so need to include the rest of the
1707 * nr_pages that have not been attempted as
1711 err += nr_pages - i - 1;
1714 return store_status(status, start, node, i - start);
1718 * Migrate an array of page address onto an array of nodes and fill
1719 * the corresponding array of status.
1721 static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes,
1722 unsigned long nr_pages,
1723 const void __user * __user *pages,
1724 const int __user *nodes,
1725 int __user *status, int flags)
1727 int current_node = NUMA_NO_NODE;
1728 LIST_HEAD(pagelist);
1732 lru_cache_disable();
1734 for (i = start = 0; i < nr_pages; i++) {
1735 const void __user *p;
1740 if (get_user(p, pages + i))
1742 if (get_user(node, nodes + i))
1744 addr = (unsigned long)untagged_addr(p);
1747 if (node < 0 || node >= MAX_NUMNODES)
1749 if (!node_state(node, N_MEMORY))
1753 if (!node_isset(node, task_nodes))
1756 if (current_node == NUMA_NO_NODE) {
1757 current_node = node;
1759 } else if (node != current_node) {
1760 err = move_pages_and_store_status(mm, current_node,
1761 &pagelist, status, start, i, nr_pages);
1765 current_node = node;
1769 * Errors in the page lookup or isolation are not fatal and we simply
1770 * report them via status
1772 err = add_page_for_migration(mm, addr, current_node,
1773 &pagelist, flags & MPOL_MF_MOVE_ALL);
1776 /* The page is successfully queued for migration */
1781 * If the page is already on the target node (!err), store the
1782 * node, otherwise, store the err.
1784 err = store_status(status, i, err ? : current_node, 1);
1788 err = move_pages_and_store_status(mm, current_node, &pagelist,
1789 status, start, i, nr_pages);
1792 current_node = NUMA_NO_NODE;
1795 /* Make sure we do not overwrite the existing error */
1796 err1 = move_pages_and_store_status(mm, current_node, &pagelist,
1797 status, start, i, nr_pages);
1806 * Determine the nodes of an array of pages and store it in an array of status.
1808 static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages,
1809 const void __user **pages, int *status)
1815 for (i = 0; i < nr_pages; i++) {
1816 unsigned long addr = (unsigned long)(*pages);
1817 struct vm_area_struct *vma;
1821 vma = vma_lookup(mm, addr);
1825 /* FOLL_DUMP to ignore special (like zero) pages */
1826 page = follow_page(vma, addr, FOLL_DUMP);
1828 err = PTR_ERR(page);
1832 err = page ? page_to_nid(page) : -ENOENT;
1840 mmap_read_unlock(mm);
1843 static int get_compat_pages_array(const void __user *chunk_pages[],
1844 const void __user * __user *pages,
1845 unsigned long chunk_nr)
1847 compat_uptr_t __user *pages32 = (compat_uptr_t __user *)pages;
1851 for (i = 0; i < chunk_nr; i++) {
1852 if (get_user(p, pages32 + i))
1854 chunk_pages[i] = compat_ptr(p);
1861 * Determine the nodes of a user array of pages and store it in
1862 * a user array of status.
1864 static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages,
1865 const void __user * __user *pages,
1868 #define DO_PAGES_STAT_CHUNK_NR 16
1869 const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR];
1870 int chunk_status[DO_PAGES_STAT_CHUNK_NR];
1873 unsigned long chunk_nr;
1875 chunk_nr = nr_pages;
1876 if (chunk_nr > DO_PAGES_STAT_CHUNK_NR)
1877 chunk_nr = DO_PAGES_STAT_CHUNK_NR;
1879 if (in_compat_syscall()) {
1880 if (get_compat_pages_array(chunk_pages, pages,
1884 if (copy_from_user(chunk_pages, pages,
1885 chunk_nr * sizeof(*chunk_pages)))
1889 do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status);
1891 if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status)))
1896 nr_pages -= chunk_nr;
1898 return nr_pages ? -EFAULT : 0;
1901 static struct mm_struct *find_mm_struct(pid_t pid, nodemask_t *mem_nodes)
1903 struct task_struct *task;
1904 struct mm_struct *mm;
1907 * There is no need to check if current process has the right to modify
1908 * the specified process when they are same.
1912 *mem_nodes = cpuset_mems_allowed(current);
1916 /* Find the mm_struct */
1918 task = find_task_by_vpid(pid);
1921 return ERR_PTR(-ESRCH);
1923 get_task_struct(task);
1926 * Check if this process has the right to modify the specified
1927 * process. Use the regular "ptrace_may_access()" checks.
1929 if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) {
1931 mm = ERR_PTR(-EPERM);
1936 mm = ERR_PTR(security_task_movememory(task));
1939 *mem_nodes = cpuset_mems_allowed(task);
1940 mm = get_task_mm(task);
1942 put_task_struct(task);
1944 mm = ERR_PTR(-EINVAL);
1949 * Move a list of pages in the address space of the currently executing
1952 static int kernel_move_pages(pid_t pid, unsigned long nr_pages,
1953 const void __user * __user *pages,
1954 const int __user *nodes,
1955 int __user *status, int flags)
1957 struct mm_struct *mm;
1959 nodemask_t task_nodes;
1962 if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
1965 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1968 mm = find_mm_struct(pid, &task_nodes);
1973 err = do_pages_move(mm, task_nodes, nr_pages, pages,
1974 nodes, status, flags);
1976 err = do_pages_stat(mm, nr_pages, pages, status);
1982 SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages,
1983 const void __user * __user *, pages,
1984 const int __user *, nodes,
1985 int __user *, status, int, flags)
1987 return kernel_move_pages(pid, nr_pages, pages, nodes, status, flags);
1990 #ifdef CONFIG_NUMA_BALANCING
1992 * Returns true if this is a safe migration target node for misplaced NUMA
1993 * pages. Currently it only checks the watermarks which crude
1995 static bool migrate_balanced_pgdat(struct pglist_data *pgdat,
1996 unsigned long nr_migrate_pages)
2000 for (z = pgdat->nr_zones - 1; z >= 0; z--) {
2001 struct zone *zone = pgdat->node_zones + z;
2003 if (!populated_zone(zone))
2006 /* Avoid waking kswapd by allocating pages_to_migrate pages. */
2007 if (!zone_watermark_ok(zone, 0,
2008 high_wmark_pages(zone) +
2017 static struct page *alloc_misplaced_dst_page(struct page *page,
2020 int nid = (int) data;
2021 struct page *newpage;
2023 newpage = __alloc_pages_node(nid,
2024 (GFP_HIGHUSER_MOVABLE |
2025 __GFP_THISNODE | __GFP_NOMEMALLOC |
2026 __GFP_NORETRY | __GFP_NOWARN) &
2032 static struct page *alloc_misplaced_dst_page_thp(struct page *page,
2035 int nid = (int) data;
2036 struct page *newpage;
2038 newpage = alloc_pages_node(nid, (GFP_TRANSHUGE_LIGHT | __GFP_THISNODE),
2043 prep_transhuge_page(newpage);
2049 static int numamigrate_isolate_page(pg_data_t *pgdat, struct page *page)
2052 int nr_pages = thp_nr_pages(page);
2054 VM_BUG_ON_PAGE(compound_order(page) && !PageTransHuge(page), page);
2056 /* Do not migrate THP mapped by multiple processes */
2057 if (PageTransHuge(page) && total_mapcount(page) > 1)
2060 /* Avoid migrating to a node that is nearly full */
2061 if (!migrate_balanced_pgdat(pgdat, nr_pages))
2064 if (isolate_lru_page(page))
2067 page_lru = page_is_file_lru(page);
2068 mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON + page_lru,
2072 * Isolating the page has taken another reference, so the
2073 * caller's reference can be safely dropped without the page
2074 * disappearing underneath us during migration.
2081 * Attempt to migrate a misplaced page to the specified destination
2082 * node. Caller is expected to have an elevated reference count on
2083 * the page that will be dropped by this function before returning.
2085 int migrate_misplaced_page(struct page *page, struct vm_area_struct *vma,
2088 pg_data_t *pgdat = NODE_DATA(node);
2091 LIST_HEAD(migratepages);
2094 int nr_pages = thp_nr_pages(page);
2097 * PTE mapped THP or HugeTLB page can't reach here so the page could
2098 * be either base page or THP. And it must be head page if it is
2101 compound = PageTransHuge(page);
2104 new = alloc_misplaced_dst_page_thp;
2106 new = alloc_misplaced_dst_page;
2109 * Don't migrate file pages that are mapped in multiple processes
2110 * with execute permissions as they are probably shared libraries.
2112 if (page_mapcount(page) != 1 && page_is_file_lru(page) &&
2113 (vma->vm_flags & VM_EXEC))
2117 * Also do not migrate dirty pages as not all filesystems can move
2118 * dirty pages in MIGRATE_ASYNC mode which is a waste of cycles.
2120 if (page_is_file_lru(page) && PageDirty(page))
2123 isolated = numamigrate_isolate_page(pgdat, page);
2127 list_add(&page->lru, &migratepages);
2128 nr_remaining = migrate_pages(&migratepages, *new, NULL, node,
2129 MIGRATE_ASYNC, MR_NUMA_MISPLACED, NULL);
2131 if (!list_empty(&migratepages)) {
2132 list_del(&page->lru);
2133 mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON +
2134 page_is_file_lru(page), -nr_pages);
2135 putback_lru_page(page);
2139 count_vm_numa_events(NUMA_PAGE_MIGRATE, nr_pages);
2140 BUG_ON(!list_empty(&migratepages));
2147 #endif /* CONFIG_NUMA_BALANCING */
2148 #endif /* CONFIG_NUMA */
2150 #ifdef CONFIG_DEVICE_PRIVATE
2151 static int migrate_vma_collect_skip(unsigned long start,
2153 struct mm_walk *walk)
2155 struct migrate_vma *migrate = walk->private;
2158 for (addr = start; addr < end; addr += PAGE_SIZE) {
2159 migrate->dst[migrate->npages] = 0;
2160 migrate->src[migrate->npages++] = 0;
2166 static int migrate_vma_collect_hole(unsigned long start,
2168 __always_unused int depth,
2169 struct mm_walk *walk)
2171 struct migrate_vma *migrate = walk->private;
2174 /* Only allow populating anonymous memory. */
2175 if (!vma_is_anonymous(walk->vma))
2176 return migrate_vma_collect_skip(start, end, walk);
2178 for (addr = start; addr < end; addr += PAGE_SIZE) {
2179 migrate->src[migrate->npages] = MIGRATE_PFN_MIGRATE;
2180 migrate->dst[migrate->npages] = 0;
2188 static int migrate_vma_collect_pmd(pmd_t *pmdp,
2189 unsigned long start,
2191 struct mm_walk *walk)
2193 struct migrate_vma *migrate = walk->private;
2194 struct vm_area_struct *vma = walk->vma;
2195 struct mm_struct *mm = vma->vm_mm;
2196 unsigned long addr = start, unmapped = 0;
2201 if (pmd_none(*pmdp))
2202 return migrate_vma_collect_hole(start, end, -1, walk);
2204 if (pmd_trans_huge(*pmdp)) {
2207 ptl = pmd_lock(mm, pmdp);
2208 if (unlikely(!pmd_trans_huge(*pmdp))) {
2213 page = pmd_page(*pmdp);
2214 if (is_huge_zero_page(page)) {
2216 split_huge_pmd(vma, pmdp, addr);
2217 if (pmd_trans_unstable(pmdp))
2218 return migrate_vma_collect_skip(start, end,
2225 if (unlikely(!trylock_page(page)))
2226 return migrate_vma_collect_skip(start, end,
2228 ret = split_huge_page(page);
2232 return migrate_vma_collect_skip(start, end,
2234 if (pmd_none(*pmdp))
2235 return migrate_vma_collect_hole(start, end, -1,
2240 if (unlikely(pmd_bad(*pmdp)))
2241 return migrate_vma_collect_skip(start, end, walk);
2243 ptep = pte_offset_map_lock(mm, pmdp, addr, &ptl);
2244 arch_enter_lazy_mmu_mode();
2246 for (; addr < end; addr += PAGE_SIZE, ptep++) {
2247 unsigned long mpfn = 0, pfn;
2254 if (pte_none(pte)) {
2255 if (vma_is_anonymous(vma)) {
2256 mpfn = MIGRATE_PFN_MIGRATE;
2262 if (!pte_present(pte)) {
2264 * Only care about unaddressable device page special
2265 * page table entry. Other special swap entries are not
2266 * migratable, and we ignore regular swapped page.
2268 entry = pte_to_swp_entry(pte);
2269 if (!is_device_private_entry(entry))
2272 page = pfn_swap_entry_to_page(entry);
2273 if (!(migrate->flags &
2274 MIGRATE_VMA_SELECT_DEVICE_PRIVATE) ||
2275 page->pgmap->owner != migrate->pgmap_owner)
2278 mpfn = migrate_pfn(page_to_pfn(page)) |
2279 MIGRATE_PFN_MIGRATE;
2280 if (is_writable_device_private_entry(entry))
2281 mpfn |= MIGRATE_PFN_WRITE;
2283 if (!(migrate->flags & MIGRATE_VMA_SELECT_SYSTEM))
2286 if (is_zero_pfn(pfn)) {
2287 mpfn = MIGRATE_PFN_MIGRATE;
2291 page = vm_normal_page(migrate->vma, addr, pte);
2292 mpfn = migrate_pfn(pfn) | MIGRATE_PFN_MIGRATE;
2293 mpfn |= pte_write(pte) ? MIGRATE_PFN_WRITE : 0;
2296 /* FIXME support THP */
2297 if (!page || !page->mapping || PageTransCompound(page)) {
2303 * By getting a reference on the page we pin it and that blocks
2304 * any kind of migration. Side effect is that it "freezes" the
2307 * We drop this reference after isolating the page from the lru
2308 * for non device page (device page are not on the lru and thus
2309 * can't be dropped from it).
2314 * Optimize for the common case where page is only mapped once
2315 * in one process. If we can lock the page, then we can safely
2316 * set up a special migration page table entry now.
2318 if (trylock_page(page)) {
2322 ptep_get_and_clear(mm, addr, ptep);
2324 /* Setup special migration page table entry */
2325 if (mpfn & MIGRATE_PFN_WRITE)
2326 entry = make_writable_migration_entry(
2329 entry = make_readable_migration_entry(
2331 swp_pte = swp_entry_to_pte(entry);
2332 if (pte_present(pte)) {
2333 if (pte_soft_dirty(pte))
2334 swp_pte = pte_swp_mksoft_dirty(swp_pte);
2335 if (pte_uffd_wp(pte))
2336 swp_pte = pte_swp_mkuffd_wp(swp_pte);
2338 if (pte_swp_soft_dirty(pte))
2339 swp_pte = pte_swp_mksoft_dirty(swp_pte);
2340 if (pte_swp_uffd_wp(pte))
2341 swp_pte = pte_swp_mkuffd_wp(swp_pte);
2343 set_pte_at(mm, addr, ptep, swp_pte);
2346 * This is like regular unmap: we remove the rmap and
2347 * drop page refcount. Page won't be freed, as we took
2348 * a reference just above.
2350 page_remove_rmap(page, false);
2353 if (pte_present(pte))
2361 migrate->dst[migrate->npages] = 0;
2362 migrate->src[migrate->npages++] = mpfn;
2364 arch_leave_lazy_mmu_mode();
2365 pte_unmap_unlock(ptep - 1, ptl);
2367 /* Only flush the TLB if we actually modified any entries */
2369 flush_tlb_range(walk->vma, start, end);
2374 static const struct mm_walk_ops migrate_vma_walk_ops = {
2375 .pmd_entry = migrate_vma_collect_pmd,
2376 .pte_hole = migrate_vma_collect_hole,
2380 * migrate_vma_collect() - collect pages over a range of virtual addresses
2381 * @migrate: migrate struct containing all migration information
2383 * This will walk the CPU page table. For each virtual address backed by a
2384 * valid page, it updates the src array and takes a reference on the page, in
2385 * order to pin the page until we lock it and unmap it.
2387 static void migrate_vma_collect(struct migrate_vma *migrate)
2389 struct mmu_notifier_range range;
2392 * Note that the pgmap_owner is passed to the mmu notifier callback so
2393 * that the registered device driver can skip invalidating device
2394 * private page mappings that won't be migrated.
2396 mmu_notifier_range_init_owner(&range, MMU_NOTIFY_MIGRATE, 0,
2397 migrate->vma, migrate->vma->vm_mm, migrate->start, migrate->end,
2398 migrate->pgmap_owner);
2399 mmu_notifier_invalidate_range_start(&range);
2401 walk_page_range(migrate->vma->vm_mm, migrate->start, migrate->end,
2402 &migrate_vma_walk_ops, migrate);
2404 mmu_notifier_invalidate_range_end(&range);
2405 migrate->end = migrate->start + (migrate->npages << PAGE_SHIFT);
2409 * migrate_vma_check_page() - check if page is pinned or not
2410 * @page: struct page to check
2412 * Pinned pages cannot be migrated. This is the same test as in
2413 * folio_migrate_mapping(), except that here we allow migration of a
2416 static bool migrate_vma_check_page(struct page *page)
2419 * One extra ref because caller holds an extra reference, either from
2420 * isolate_lru_page() for a regular page, or migrate_vma_collect() for
2426 * FIXME support THP (transparent huge page), it is bit more complex to
2427 * check them than regular pages, because they can be mapped with a pmd
2428 * or with a pte (split pte mapping).
2430 if (PageCompound(page))
2433 /* Page from ZONE_DEVICE have one extra reference */
2434 if (is_zone_device_page(page)) {
2436 * Private page can never be pin as they have no valid pte and
2437 * GUP will fail for those. Yet if there is a pending migration
2438 * a thread might try to wait on the pte migration entry and
2439 * will bump the page reference count. Sadly there is no way to
2440 * differentiate a regular pin from migration wait. Hence to
2441 * avoid 2 racing thread trying to migrate back to CPU to enter
2442 * infinite loop (one stopping migration because the other is
2443 * waiting on pte migration entry). We always return true here.
2445 * FIXME proper solution is to rework migration_entry_wait() so
2446 * it does not need to take a reference on page.
2448 return is_device_private_page(page);
2451 /* For file back page */
2452 if (page_mapping(page))
2453 extra += 1 + page_has_private(page);
2455 if ((page_count(page) - extra) > page_mapcount(page))
2462 * migrate_vma_unmap() - replace page mapping with special migration pte entry
2463 * @migrate: migrate struct containing all migration information
2465 * Isolate pages from the LRU and replace mappings (CPU page table pte) with a
2466 * special migration pte entry and check if it has been pinned. Pinned pages are
2467 * restored because we cannot migrate them.
2469 * This is the last step before we call the device driver callback to allocate
2470 * destination memory and copy contents of original page over to new page.
2472 static void migrate_vma_unmap(struct migrate_vma *migrate)
2474 const unsigned long npages = migrate->npages;
2475 unsigned long i, restore = 0;
2476 bool allow_drain = true;
2480 for (i = 0; i < npages; i++) {
2481 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2486 /* ZONE_DEVICE pages are not on LRU */
2487 if (!is_zone_device_page(page)) {
2488 if (!PageLRU(page) && allow_drain) {
2489 /* Drain CPU's pagevec */
2490 lru_add_drain_all();
2491 allow_drain = false;
2494 if (isolate_lru_page(page)) {
2495 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2501 /* Drop the reference we took in collect */
2505 if (page_mapped(page))
2506 try_to_migrate(page, 0);
2508 if (page_mapped(page) || !migrate_vma_check_page(page)) {
2509 if (!is_zone_device_page(page)) {
2511 putback_lru_page(page);
2514 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2521 for (i = 0; i < npages && restore; i++) {
2522 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2524 if (!page || (migrate->src[i] & MIGRATE_PFN_MIGRATE))
2527 remove_migration_ptes(page, page, false);
2529 migrate->src[i] = 0;
2537 * migrate_vma_setup() - prepare to migrate a range of memory
2538 * @args: contains the vma, start, and pfns arrays for the migration
2540 * Returns: negative errno on failures, 0 when 0 or more pages were migrated
2543 * Prepare to migrate a range of memory virtual address range by collecting all
2544 * the pages backing each virtual address in the range, saving them inside the
2545 * src array. Then lock those pages and unmap them. Once the pages are locked
2546 * and unmapped, check whether each page is pinned or not. Pages that aren't
2547 * pinned have the MIGRATE_PFN_MIGRATE flag set (by this function) in the
2548 * corresponding src array entry. Then restores any pages that are pinned, by
2549 * remapping and unlocking those pages.
2551 * The caller should then allocate destination memory and copy source memory to
2552 * it for all those entries (ie with MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE
2553 * flag set). Once these are allocated and copied, the caller must update each
2554 * corresponding entry in the dst array with the pfn value of the destination
2555 * page and with MIGRATE_PFN_VALID. Destination pages must be locked via
2558 * Note that the caller does not have to migrate all the pages that are marked
2559 * with MIGRATE_PFN_MIGRATE flag in src array unless this is a migration from
2560 * device memory to system memory. If the caller cannot migrate a device page
2561 * back to system memory, then it must return VM_FAULT_SIGBUS, which has severe
2562 * consequences for the userspace process, so it must be avoided if at all
2565 * For empty entries inside CPU page table (pte_none() or pmd_none() is true) we
2566 * do set MIGRATE_PFN_MIGRATE flag inside the corresponding source array thus
2567 * allowing the caller to allocate device memory for those unbacked virtual
2568 * addresses. For this the caller simply has to allocate device memory and
2569 * properly set the destination entry like for regular migration. Note that
2570 * this can still fail, and thus inside the device driver you must check if the
2571 * migration was successful for those entries after calling migrate_vma_pages(),
2572 * just like for regular migration.
2574 * After that, the callers must call migrate_vma_pages() to go over each entry
2575 * in the src array that has the MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE flag
2576 * set. If the corresponding entry in dst array has MIGRATE_PFN_VALID flag set,
2577 * then migrate_vma_pages() to migrate struct page information from the source
2578 * struct page to the destination struct page. If it fails to migrate the
2579 * struct page information, then it clears the MIGRATE_PFN_MIGRATE flag in the
2582 * At this point all successfully migrated pages have an entry in the src
2583 * array with MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE flag set and the dst
2584 * array entry with MIGRATE_PFN_VALID flag set.
2586 * Once migrate_vma_pages() returns the caller may inspect which pages were
2587 * successfully migrated, and which were not. Successfully migrated pages will
2588 * have the MIGRATE_PFN_MIGRATE flag set for their src array entry.
2590 * It is safe to update device page table after migrate_vma_pages() because
2591 * both destination and source page are still locked, and the mmap_lock is held
2592 * in read mode (hence no one can unmap the range being migrated).
2594 * Once the caller is done cleaning up things and updating its page table (if it
2595 * chose to do so, this is not an obligation) it finally calls
2596 * migrate_vma_finalize() to update the CPU page table to point to new pages
2597 * for successfully migrated pages or otherwise restore the CPU page table to
2598 * point to the original source pages.
2600 int migrate_vma_setup(struct migrate_vma *args)
2602 long nr_pages = (args->end - args->start) >> PAGE_SHIFT;
2604 args->start &= PAGE_MASK;
2605 args->end &= PAGE_MASK;
2606 if (!args->vma || is_vm_hugetlb_page(args->vma) ||
2607 (args->vma->vm_flags & VM_SPECIAL) || vma_is_dax(args->vma))
2611 if (args->start < args->vma->vm_start ||
2612 args->start >= args->vma->vm_end)
2614 if (args->end <= args->vma->vm_start || args->end > args->vma->vm_end)
2616 if (!args->src || !args->dst)
2619 memset(args->src, 0, sizeof(*args->src) * nr_pages);
2623 migrate_vma_collect(args);
2626 migrate_vma_unmap(args);
2629 * At this point pages are locked and unmapped, and thus they have
2630 * stable content and can safely be copied to destination memory that
2631 * is allocated by the drivers.
2636 EXPORT_SYMBOL(migrate_vma_setup);
2639 * This code closely matches the code in:
2640 * __handle_mm_fault()
2641 * handle_pte_fault()
2642 * do_anonymous_page()
2643 * to map in an anonymous zero page but the struct page will be a ZONE_DEVICE
2646 static void migrate_vma_insert_page(struct migrate_vma *migrate,
2651 struct vm_area_struct *vma = migrate->vma;
2652 struct mm_struct *mm = vma->vm_mm;
2662 /* Only allow populating anonymous memory */
2663 if (!vma_is_anonymous(vma))
2666 pgdp = pgd_offset(mm, addr);
2667 p4dp = p4d_alloc(mm, pgdp, addr);
2670 pudp = pud_alloc(mm, p4dp, addr);
2673 pmdp = pmd_alloc(mm, pudp, addr);
2677 if (pmd_trans_huge(*pmdp) || pmd_devmap(*pmdp))
2681 * Use pte_alloc() instead of pte_alloc_map(). We can't run
2682 * pte_offset_map() on pmds where a huge pmd might be created
2683 * from a different thread.
2685 * pte_alloc_map() is safe to use under mmap_write_lock(mm) or when
2686 * parallel threads are excluded by other means.
2688 * Here we only have mmap_read_lock(mm).
2690 if (pte_alloc(mm, pmdp))
2693 /* See the comment in pte_alloc_one_map() */
2694 if (unlikely(pmd_trans_unstable(pmdp)))
2697 if (unlikely(anon_vma_prepare(vma)))
2699 if (mem_cgroup_charge(page_folio(page), vma->vm_mm, GFP_KERNEL))
2703 * The memory barrier inside __SetPageUptodate makes sure that
2704 * preceding stores to the page contents become visible before
2705 * the set_pte_at() write.
2707 __SetPageUptodate(page);
2709 if (is_zone_device_page(page)) {
2710 if (is_device_private_page(page)) {
2711 swp_entry_t swp_entry;
2713 if (vma->vm_flags & VM_WRITE)
2714 swp_entry = make_writable_device_private_entry(
2717 swp_entry = make_readable_device_private_entry(
2719 entry = swp_entry_to_pte(swp_entry);
2722 * For now we only support migrating to un-addressable
2725 pr_warn_once("Unsupported ZONE_DEVICE page type.\n");
2729 entry = mk_pte(page, vma->vm_page_prot);
2730 if (vma->vm_flags & VM_WRITE)
2731 entry = pte_mkwrite(pte_mkdirty(entry));
2734 ptep = pte_offset_map_lock(mm, pmdp, addr, &ptl);
2736 if (check_stable_address_space(mm))
2739 if (pte_present(*ptep)) {
2740 unsigned long pfn = pte_pfn(*ptep);
2742 if (!is_zero_pfn(pfn))
2745 } else if (!pte_none(*ptep))
2749 * Check for userfaultfd but do not deliver the fault. Instead,
2752 if (userfaultfd_missing(vma))
2755 inc_mm_counter(mm, MM_ANONPAGES);
2756 page_add_new_anon_rmap(page, vma, addr, false);
2757 if (!is_zone_device_page(page))
2758 lru_cache_add_inactive_or_unevictable(page, vma);
2762 flush_cache_page(vma, addr, pte_pfn(*ptep));
2763 ptep_clear_flush_notify(vma, addr, ptep);
2764 set_pte_at_notify(mm, addr, ptep, entry);
2765 update_mmu_cache(vma, addr, ptep);
2767 /* No need to invalidate - it was non-present before */
2768 set_pte_at(mm, addr, ptep, entry);
2769 update_mmu_cache(vma, addr, ptep);
2772 pte_unmap_unlock(ptep, ptl);
2773 *src = MIGRATE_PFN_MIGRATE;
2777 pte_unmap_unlock(ptep, ptl);
2779 *src &= ~MIGRATE_PFN_MIGRATE;
2783 * migrate_vma_pages() - migrate meta-data from src page to dst page
2784 * @migrate: migrate struct containing all migration information
2786 * This migrates struct page meta-data from source struct page to destination
2787 * struct page. This effectively finishes the migration from source page to the
2790 void migrate_vma_pages(struct migrate_vma *migrate)
2792 const unsigned long npages = migrate->npages;
2793 const unsigned long start = migrate->start;
2794 struct mmu_notifier_range range;
2795 unsigned long addr, i;
2796 bool notified = false;
2798 for (i = 0, addr = start; i < npages; addr += PAGE_SIZE, i++) {
2799 struct page *newpage = migrate_pfn_to_page(migrate->dst[i]);
2800 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2801 struct address_space *mapping;
2805 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2810 if (!(migrate->src[i] & MIGRATE_PFN_MIGRATE))
2815 mmu_notifier_range_init_owner(&range,
2816 MMU_NOTIFY_MIGRATE, 0, migrate->vma,
2817 migrate->vma->vm_mm, addr, migrate->end,
2818 migrate->pgmap_owner);
2819 mmu_notifier_invalidate_range_start(&range);
2821 migrate_vma_insert_page(migrate, addr, newpage,
2826 mapping = page_mapping(page);
2828 if (is_zone_device_page(newpage)) {
2829 if (is_device_private_page(newpage)) {
2831 * For now only support private anonymous when
2832 * migrating to un-addressable device memory.
2835 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2840 * Other types of ZONE_DEVICE page are not
2843 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2848 r = migrate_page(mapping, newpage, page, MIGRATE_SYNC_NO_COPY);
2849 if (r != MIGRATEPAGE_SUCCESS)
2850 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2854 * No need to double call mmu_notifier->invalidate_range() callback as
2855 * the above ptep_clear_flush_notify() inside migrate_vma_insert_page()
2856 * did already call it.
2859 mmu_notifier_invalidate_range_only_end(&range);
2861 EXPORT_SYMBOL(migrate_vma_pages);
2864 * migrate_vma_finalize() - restore CPU page table entry
2865 * @migrate: migrate struct containing all migration information
2867 * This replaces the special migration pte entry with either a mapping to the
2868 * new page if migration was successful for that page, or to the original page
2871 * This also unlocks the pages and puts them back on the lru, or drops the extra
2872 * refcount, for device pages.
2874 void migrate_vma_finalize(struct migrate_vma *migrate)
2876 const unsigned long npages = migrate->npages;
2879 for (i = 0; i < npages; i++) {
2880 struct page *newpage = migrate_pfn_to_page(migrate->dst[i]);
2881 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2885 unlock_page(newpage);
2891 if (!(migrate->src[i] & MIGRATE_PFN_MIGRATE) || !newpage) {
2893 unlock_page(newpage);
2899 remove_migration_ptes(page, newpage, false);
2902 if (is_zone_device_page(page))
2905 putback_lru_page(page);
2907 if (newpage != page) {
2908 unlock_page(newpage);
2909 if (is_zone_device_page(newpage))
2912 putback_lru_page(newpage);
2916 EXPORT_SYMBOL(migrate_vma_finalize);
2917 #endif /* CONFIG_DEVICE_PRIVATE */
2920 * node_demotion[] example:
2922 * Consider a system with two sockets. Each socket has
2923 * three classes of memory attached: fast, medium and slow.
2924 * Each memory class is placed in its own NUMA node. The
2925 * CPUs are placed in the node with the "fast" memory. The
2926 * 6 NUMA nodes (0-5) might be split among the sockets like
2932 * When Node 0 fills up, its memory should be migrated to
2933 * Node 1. When Node 1 fills up, it should be migrated to
2934 * Node 2. The migration path start on the nodes with the
2935 * processors (since allocations default to this node) and
2936 * fast memory, progress through medium and end with the
2939 * 0 -> 1 -> 2 -> stop
2940 * 3 -> 4 -> 5 -> stop
2942 * This is represented in the node_demotion[] like this:
2944 * { nr=1, nodes[0]=1 }, // Node 0 migrates to 1
2945 * { nr=1, nodes[0]=2 }, // Node 1 migrates to 2
2946 * { nr=0, nodes[0]=-1 }, // Node 2 does not migrate
2947 * { nr=1, nodes[0]=4 }, // Node 3 migrates to 4
2948 * { nr=1, nodes[0]=5 }, // Node 4 migrates to 5
2949 * { nr=0, nodes[0]=-1 }, // Node 5 does not migrate
2951 * Moreover some systems may have multiple slow memory nodes.
2952 * Suppose a system has one socket with 3 memory nodes, node 0
2953 * is fast memory type, and node 1/2 both are slow memory
2954 * type, and the distance between fast memory node and slow
2955 * memory node is same. So the migration path should be:
2959 * This is represented in the node_demotion[] like this:
2960 * { nr=2, {nodes[0]=1, nodes[1]=2} }, // Node 0 migrates to node 1 and node 2
2961 * { nr=0, nodes[0]=-1, }, // Node 1 dose not migrate
2962 * { nr=0, nodes[0]=-1, }, // Node 2 does not migrate
2966 * Writes to this array occur without locking. Cycles are
2967 * not allowed: Node X demotes to Y which demotes to X...
2969 * If multiple reads are performed, a single rcu_read_lock()
2970 * must be held over all reads to ensure that no cycles are
2973 #define DEFAULT_DEMOTION_TARGET_NODES 15
2975 #if MAX_NUMNODES < DEFAULT_DEMOTION_TARGET_NODES
2976 #define DEMOTION_TARGET_NODES (MAX_NUMNODES - 1)
2978 #define DEMOTION_TARGET_NODES DEFAULT_DEMOTION_TARGET_NODES
2981 struct demotion_nodes {
2983 short nodes[DEMOTION_TARGET_NODES];
2986 static struct demotion_nodes *node_demotion __read_mostly;
2989 * next_demotion_node() - Get the next node in the demotion path
2990 * @node: The starting node to lookup the next node
2992 * Return: node id for next memory node in the demotion path hierarchy
2993 * from @node; NUMA_NO_NODE if @node is terminal. This does not keep
2994 * @node online or guarantee that it *continues* to be the next demotion
2997 int next_demotion_node(int node)
2999 struct demotion_nodes *nd;
3000 unsigned short target_nr, index;
3004 return NUMA_NO_NODE;
3006 nd = &node_demotion[node];
3009 * node_demotion[] is updated without excluding this
3010 * function from running. RCU doesn't provide any
3011 * compiler barriers, so the READ_ONCE() is required
3012 * to avoid compiler reordering or read merging.
3014 * Make sure to use RCU over entire code blocks if
3015 * node_demotion[] reads need to be consistent.
3018 target_nr = READ_ONCE(nd->nr);
3020 switch (target_nr) {
3022 target = NUMA_NO_NODE;
3029 * If there are multiple target nodes, just select one
3030 * target node randomly.
3032 * In addition, we can also use round-robin to select
3033 * target node, but we should introduce another variable
3034 * for node_demotion[] to record last selected target node,
3035 * that may cause cache ping-pong due to the changing of
3036 * last target node. Or introducing per-cpu data to avoid
3037 * caching issue, which seems more complicated. So selecting
3038 * target node randomly seems better until now.
3040 index = get_random_int() % target_nr;
3044 target = READ_ONCE(nd->nodes[index]);
3051 #if defined(CONFIG_HOTPLUG_CPU)
3052 /* Disable reclaim-based migration. */
3053 static void __disable_all_migrate_targets(void)
3060 for_each_online_node(node) {
3061 node_demotion[node].nr = 0;
3062 for (i = 0; i < DEMOTION_TARGET_NODES; i++)
3063 node_demotion[node].nodes[i] = NUMA_NO_NODE;
3067 static void disable_all_migrate_targets(void)
3069 __disable_all_migrate_targets();
3072 * Ensure that the "disable" is visible across the system.
3073 * Readers will see either a combination of before+disable
3074 * state or disable+after. They will never see before and
3075 * after state together.
3077 * The before+after state together might have cycles and
3078 * could cause readers to do things like loop until this
3079 * function finishes. This ensures they can only see a
3080 * single "bad" read and would, for instance, only loop
3087 * Find an automatic demotion target for 'node'.
3088 * Failing here is OK. It might just indicate
3089 * being at the end of a chain.
3091 static int establish_migrate_target(int node, nodemask_t *used,
3094 int migration_target, index, val;
3095 struct demotion_nodes *nd;
3098 return NUMA_NO_NODE;
3100 nd = &node_demotion[node];
3102 migration_target = find_next_best_node(node, used);
3103 if (migration_target == NUMA_NO_NODE)
3104 return NUMA_NO_NODE;
3107 * If the node has been set a migration target node before,
3108 * which means it's the best distance between them. Still
3109 * check if this node can be demoted to other target nodes
3110 * if they have a same best distance.
3112 if (best_distance != -1) {
3113 val = node_distance(node, migration_target);
3114 if (val > best_distance)
3115 return NUMA_NO_NODE;
3119 if (WARN_ONCE(index >= DEMOTION_TARGET_NODES,
3120 "Exceeds maximum demotion target nodes\n"))
3121 return NUMA_NO_NODE;
3123 nd->nodes[index] = migration_target;
3126 return migration_target;
3130 * When memory fills up on a node, memory contents can be
3131 * automatically migrated to another node instead of
3132 * discarded at reclaim.
3134 * Establish a "migration path" which will start at nodes
3135 * with CPUs and will follow the priorities used to build the
3136 * page allocator zonelists.
3138 * The difference here is that cycles must be avoided. If
3139 * node0 migrates to node1, then neither node1, nor anything
3140 * node1 migrates to can migrate to node0. Also one node can
3141 * be migrated to multiple nodes if the target nodes all have
3142 * a same best-distance against the source node.
3144 * This function can run simultaneously with readers of
3145 * node_demotion[]. However, it can not run simultaneously
3146 * with itself. Exclusion is provided by memory hotplug events
3147 * being single-threaded.
3149 static void __set_migration_target_nodes(void)
3151 nodemask_t next_pass = NODE_MASK_NONE;
3152 nodemask_t this_pass = NODE_MASK_NONE;
3153 nodemask_t used_targets = NODE_MASK_NONE;
3154 int node, best_distance;
3157 * Avoid any oddities like cycles that could occur
3158 * from changes in the topology. This will leave
3159 * a momentary gap when migration is disabled.
3161 disable_all_migrate_targets();
3164 * Allocations go close to CPUs, first. Assume that
3165 * the migration path starts at the nodes with CPUs.
3167 next_pass = node_states[N_CPU];
3169 this_pass = next_pass;
3170 next_pass = NODE_MASK_NONE;
3172 * To avoid cycles in the migration "graph", ensure
3173 * that migration sources are not future targets by
3174 * setting them in 'used_targets'. Do this only
3175 * once per pass so that multiple source nodes can
3176 * share a target node.
3178 * 'used_targets' will become unavailable in future
3179 * passes. This limits some opportunities for
3180 * multiple source nodes to share a destination.
3182 nodes_or(used_targets, used_targets, this_pass);
3184 for_each_node_mask(node, this_pass) {
3188 * Try to set up the migration path for the node, and the target
3189 * migration nodes can be multiple, so doing a loop to find all
3190 * the target nodes if they all have a best node distance.
3194 establish_migrate_target(node, &used_targets,
3197 if (target_node == NUMA_NO_NODE)
3200 if (best_distance == -1)
3201 best_distance = node_distance(node, target_node);
3204 * Visit targets from this pass in the next pass.
3205 * Eventually, every node will have been part of
3206 * a pass, and will become set in 'used_targets'.
3208 node_set(target_node, next_pass);
3212 * 'next_pass' contains nodes which became migration
3213 * targets in this pass. Make additional passes until
3214 * no more migrations targets are available.
3216 if (!nodes_empty(next_pass))
3221 * For callers that do not hold get_online_mems() already.
3223 static void set_migration_target_nodes(void)
3226 __set_migration_target_nodes();
3231 * This leaves migrate-on-reclaim transiently disabled between
3232 * the MEM_GOING_OFFLINE and MEM_OFFLINE events. This runs
3233 * whether reclaim-based migration is enabled or not, which
3234 * ensures that the user can turn reclaim-based migration at
3235 * any time without needing to recalculate migration targets.
3237 * These callbacks already hold get_online_mems(). That is why
3238 * __set_migration_target_nodes() can be used as opposed to
3239 * set_migration_target_nodes().
3241 static int __meminit migrate_on_reclaim_callback(struct notifier_block *self,
3242 unsigned long action, void *_arg)
3244 struct memory_notify *arg = _arg;
3247 * Only update the node migration order when a node is
3248 * changing status, like online->offline. This avoids
3249 * the overhead of synchronize_rcu() in most cases.
3251 if (arg->status_change_nid < 0)
3252 return notifier_from_errno(0);
3255 case MEM_GOING_OFFLINE:
3257 * Make sure there are not transient states where
3258 * an offline node is a migration target. This
3259 * will leave migration disabled until the offline
3260 * completes and the MEM_OFFLINE case below runs.
3262 disable_all_migrate_targets();
3267 * Recalculate the target nodes once the node
3268 * reaches its final state (online or offline).
3270 __set_migration_target_nodes();
3272 case MEM_CANCEL_OFFLINE:
3274 * MEM_GOING_OFFLINE disabled all the migration
3275 * targets. Reenable them.
3277 __set_migration_target_nodes();
3279 case MEM_GOING_ONLINE:
3280 case MEM_CANCEL_ONLINE:
3284 return notifier_from_errno(0);
3288 * React to hotplug events that might affect the migration targets
3289 * like events that online or offline NUMA nodes.
3291 * The ordering is also currently dependent on which nodes have
3292 * CPUs. That means we need CPU on/offline notification too.
3294 static int migration_online_cpu(unsigned int cpu)
3296 set_migration_target_nodes();
3300 static int migration_offline_cpu(unsigned int cpu)
3302 set_migration_target_nodes();
3306 static int __init migrate_on_reclaim_init(void)
3310 node_demotion = kmalloc_array(nr_node_ids,
3311 sizeof(struct demotion_nodes),
3313 WARN_ON(!node_demotion);
3315 ret = cpuhp_setup_state_nocalls(CPUHP_MM_DEMOTION_DEAD, "mm/demotion:offline",
3316 NULL, migration_offline_cpu);
3318 * In the unlikely case that this fails, the automatic
3319 * migration targets may become suboptimal for nodes
3320 * where N_CPU changes. With such a small impact in a
3321 * rare case, do not bother trying to do anything special.
3324 ret = cpuhp_setup_state(CPUHP_AP_MM_DEMOTION_ONLINE, "mm/demotion:online",
3325 migration_online_cpu, NULL);
3328 hotplug_memory_notifier(migrate_on_reclaim_callback, 100);
3331 late_initcall(migrate_on_reclaim_init);
3332 #endif /* CONFIG_HOTPLUG_CPU */
3334 bool numa_demotion_enabled = false;
3337 static ssize_t numa_demotion_enabled_show(struct kobject *kobj,
3338 struct kobj_attribute *attr, char *buf)
3340 return sysfs_emit(buf, "%s\n",
3341 numa_demotion_enabled ? "true" : "false");
3344 static ssize_t numa_demotion_enabled_store(struct kobject *kobj,
3345 struct kobj_attribute *attr,
3346 const char *buf, size_t count)
3348 if (!strncmp(buf, "true", 4) || !strncmp(buf, "1", 1))
3349 numa_demotion_enabled = true;
3350 else if (!strncmp(buf, "false", 5) || !strncmp(buf, "0", 1))
3351 numa_demotion_enabled = false;
3358 static struct kobj_attribute numa_demotion_enabled_attr =
3359 __ATTR(demotion_enabled, 0644, numa_demotion_enabled_show,
3360 numa_demotion_enabled_store);
3362 static struct attribute *numa_attrs[] = {
3363 &numa_demotion_enabled_attr.attr,
3367 static const struct attribute_group numa_attr_group = {
3368 .attrs = numa_attrs,
3371 static int __init numa_init_sysfs(void)
3374 struct kobject *numa_kobj;
3376 numa_kobj = kobject_create_and_add("numa", mm_kobj);
3378 pr_err("failed to create numa kobject\n");
3381 err = sysfs_create_group(numa_kobj, &numa_attr_group);
3383 pr_err("failed to register numa group\n");
3389 kobject_put(numa_kobj);
3392 subsys_initcall(numa_init_sysfs);