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>
54 #include <asm/tlbflush.h>
56 #define CREATE_TRACE_POINTS
57 #include <trace/events/migrate.h>
61 int isolate_movable_page(struct page *page, isolate_mode_t mode)
63 struct address_space *mapping;
66 * Avoid burning cycles with pages that are yet under __free_pages(),
67 * or just got freed under us.
69 * In case we 'win' a race for a movable page being freed under us and
70 * raise its refcount preventing __free_pages() from doing its job
71 * the put_page() at the end of this block will take care of
72 * release this page, thus avoiding a nasty leakage.
74 if (unlikely(!get_page_unless_zero(page)))
78 * Check PageMovable before holding a PG_lock because page's owner
79 * assumes anybody doesn't touch PG_lock of newly allocated page
80 * so unconditionally grabbing the lock ruins page's owner side.
82 if (unlikely(!__PageMovable(page)))
85 * As movable pages are not isolated from LRU lists, concurrent
86 * compaction threads can race against page migration functions
87 * as well as race against the releasing a page.
89 * In order to avoid having an already isolated movable page
90 * being (wrongly) re-isolated while it is under migration,
91 * or to avoid attempting to isolate pages being released,
92 * lets be sure we have the page lock
93 * before proceeding with the movable page isolation steps.
95 if (unlikely(!trylock_page(page)))
98 if (!PageMovable(page) || PageIsolated(page))
101 mapping = page_mapping(page);
102 VM_BUG_ON_PAGE(!mapping, page);
104 if (!mapping->a_ops->isolate_page(page, mode))
105 goto out_no_isolated;
107 /* Driver shouldn't use PG_isolated bit of page->flags */
108 WARN_ON_ONCE(PageIsolated(page));
109 __SetPageIsolated(page);
122 static void putback_movable_page(struct page *page)
124 struct address_space *mapping;
126 mapping = page_mapping(page);
127 mapping->a_ops->putback_page(page);
128 __ClearPageIsolated(page);
132 * Put previously isolated pages back onto the appropriate lists
133 * from where they were once taken off for compaction/migration.
135 * This function shall be used whenever the isolated pageset has been
136 * built from lru, balloon, hugetlbfs page. See isolate_migratepages_range()
137 * and isolate_huge_page().
139 void putback_movable_pages(struct list_head *l)
144 list_for_each_entry_safe(page, page2, l, lru) {
145 if (unlikely(PageHuge(page))) {
146 putback_active_hugepage(page);
149 list_del(&page->lru);
151 * We isolated non-lru movable page so here we can use
152 * __PageMovable because LRU page's mapping cannot have
153 * PAGE_MAPPING_MOVABLE.
155 if (unlikely(__PageMovable(page))) {
156 VM_BUG_ON_PAGE(!PageIsolated(page), page);
158 if (PageMovable(page))
159 putback_movable_page(page);
161 __ClearPageIsolated(page);
165 mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON +
166 page_is_file_lru(page), -thp_nr_pages(page));
167 putback_lru_page(page);
173 * Restore a potential migration pte to a working pte entry
175 static bool remove_migration_pte(struct page *page, struct vm_area_struct *vma,
176 unsigned long addr, void *old)
178 struct page_vma_mapped_walk pvmw = {
182 .flags = PVMW_SYNC | PVMW_MIGRATION,
188 VM_BUG_ON_PAGE(PageTail(page), page);
189 while (page_vma_mapped_walk(&pvmw)) {
193 new = page - pvmw.page->index +
194 linear_page_index(vma, pvmw.address);
196 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
197 /* PMD-mapped THP migration entry */
199 VM_BUG_ON_PAGE(PageHuge(page) || !PageTransCompound(page), page);
200 remove_migration_pmd(&pvmw, new);
206 pte = pte_mkold(mk_pte(new, READ_ONCE(vma->vm_page_prot)));
207 if (pte_swp_soft_dirty(*pvmw.pte))
208 pte = pte_mksoft_dirty(pte);
211 * Recheck VMA as permissions can change since migration started
213 entry = pte_to_swp_entry(*pvmw.pte);
214 if (is_writable_migration_entry(entry))
215 pte = maybe_mkwrite(pte, vma);
216 else if (pte_swp_uffd_wp(*pvmw.pte))
217 pte = pte_mkuffd_wp(pte);
219 if (unlikely(is_device_private_page(new))) {
221 entry = make_writable_device_private_entry(
224 entry = make_readable_device_private_entry(
226 pte = swp_entry_to_pte(entry);
227 if (pte_swp_soft_dirty(*pvmw.pte))
228 pte = pte_swp_mksoft_dirty(pte);
229 if (pte_swp_uffd_wp(*pvmw.pte))
230 pte = pte_swp_mkuffd_wp(pte);
233 #ifdef CONFIG_HUGETLB_PAGE
235 unsigned int shift = huge_page_shift(hstate_vma(vma));
237 pte = pte_mkhuge(pte);
238 pte = arch_make_huge_pte(pte, shift, vma->vm_flags);
239 set_huge_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte);
241 hugepage_add_anon_rmap(new, vma, pvmw.address);
243 page_dup_rmap(new, true);
247 set_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte);
250 page_add_anon_rmap(new, vma, pvmw.address, false);
252 page_add_file_rmap(new, false);
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);
1089 * node_demotion[] example:
1091 * Consider a system with two sockets. Each socket has
1092 * three classes of memory attached: fast, medium and slow.
1093 * Each memory class is placed in its own NUMA node. The
1094 * CPUs are placed in the node with the "fast" memory. The
1095 * 6 NUMA nodes (0-5) might be split among the sockets like
1101 * When Node 0 fills up, its memory should be migrated to
1102 * Node 1. When Node 1 fills up, it should be migrated to
1103 * Node 2. The migration path start on the nodes with the
1104 * processors (since allocations default to this node) and
1105 * fast memory, progress through medium and end with the
1108 * 0 -> 1 -> 2 -> stop
1109 * 3 -> 4 -> 5 -> stop
1111 * This is represented in the node_demotion[] like this:
1113 * { 1, // Node 0 migrates to 1
1114 * 2, // Node 1 migrates to 2
1115 * -1, // Node 2 does not migrate
1116 * 4, // Node 3 migrates to 4
1117 * 5, // Node 4 migrates to 5
1118 * -1} // Node 5 does not migrate
1122 * Writes to this array occur without locking. Cycles are
1123 * not allowed: Node X demotes to Y which demotes to X...
1125 * If multiple reads are performed, a single rcu_read_lock()
1126 * must be held over all reads to ensure that no cycles are
1129 static int node_demotion[MAX_NUMNODES] __read_mostly =
1130 {[0 ... MAX_NUMNODES - 1] = NUMA_NO_NODE};
1133 * next_demotion_node() - Get the next node in the demotion path
1134 * @node: The starting node to lookup the next node
1136 * Return: node id for next memory node in the demotion path hierarchy
1137 * from @node; NUMA_NO_NODE if @node is terminal. This does not keep
1138 * @node online or guarantee that it *continues* to be the next demotion
1141 int next_demotion_node(int node)
1146 * node_demotion[] is updated without excluding this
1147 * function from running. RCU doesn't provide any
1148 * compiler barriers, so the READ_ONCE() is required
1149 * to avoid compiler reordering or read merging.
1151 * Make sure to use RCU over entire code blocks if
1152 * node_demotion[] reads need to be consistent.
1155 target = READ_ONCE(node_demotion[node]);
1162 * Obtain the lock on page, remove all ptes and migrate the page
1163 * to the newly allocated page in newpage.
1165 static int unmap_and_move(new_page_t get_new_page,
1166 free_page_t put_new_page,
1167 unsigned long private, struct page *page,
1168 int force, enum migrate_mode mode,
1169 enum migrate_reason reason,
1170 struct list_head *ret)
1172 int rc = MIGRATEPAGE_SUCCESS;
1173 struct page *newpage = NULL;
1175 if (!thp_migration_supported() && PageTransHuge(page))
1178 if (page_count(page) == 1) {
1179 /* page was freed from under us. So we are done. */
1180 ClearPageActive(page);
1181 ClearPageUnevictable(page);
1182 if (unlikely(__PageMovable(page))) {
1184 if (!PageMovable(page))
1185 __ClearPageIsolated(page);
1191 newpage = get_new_page(page, private);
1195 rc = __unmap_and_move(page, newpage, force, mode);
1196 if (rc == MIGRATEPAGE_SUCCESS)
1197 set_page_owner_migrate_reason(newpage, reason);
1200 if (rc != -EAGAIN) {
1202 * A page that has been migrated has all references
1203 * removed and will be freed. A page that has not been
1204 * migrated will have kept its references and be restored.
1206 list_del(&page->lru);
1210 * If migration is successful, releases reference grabbed during
1211 * isolation. Otherwise, restore the page to right list unless
1214 if (rc == MIGRATEPAGE_SUCCESS) {
1216 * Compaction can migrate also non-LRU pages which are
1217 * not accounted to NR_ISOLATED_*. They can be recognized
1220 if (likely(!__PageMovable(page)))
1221 mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON +
1222 page_is_file_lru(page), -thp_nr_pages(page));
1224 if (reason != MR_MEMORY_FAILURE)
1226 * We release the page in page_handle_poison.
1231 list_add_tail(&page->lru, ret);
1234 put_new_page(newpage, private);
1243 * Counterpart of unmap_and_move_page() for hugepage migration.
1245 * This function doesn't wait the completion of hugepage I/O
1246 * because there is no race between I/O and migration for hugepage.
1247 * Note that currently hugepage I/O occurs only in direct I/O
1248 * where no lock is held and PG_writeback is irrelevant,
1249 * and writeback status of all subpages are counted in the reference
1250 * count of the head page (i.e. if all subpages of a 2MB hugepage are
1251 * under direct I/O, the reference of the head page is 512 and a bit more.)
1252 * This means that when we try to migrate hugepage whose subpages are
1253 * doing direct I/O, some references remain after try_to_unmap() and
1254 * hugepage migration fails without data corruption.
1256 * There is also no race when direct I/O is issued on the page under migration,
1257 * because then pte is replaced with migration swap entry and direct I/O code
1258 * will wait in the page fault for migration to complete.
1260 static int unmap_and_move_huge_page(new_page_t get_new_page,
1261 free_page_t put_new_page, unsigned long private,
1262 struct page *hpage, int force,
1263 enum migrate_mode mode, int reason,
1264 struct list_head *ret)
1267 int page_was_mapped = 0;
1268 struct page *new_hpage;
1269 struct anon_vma *anon_vma = NULL;
1270 struct address_space *mapping = NULL;
1273 * Migratability of hugepages depends on architectures and their size.
1274 * This check is necessary because some callers of hugepage migration
1275 * like soft offline and memory hotremove don't walk through page
1276 * tables or check whether the hugepage is pmd-based or not before
1277 * kicking migration.
1279 if (!hugepage_migration_supported(page_hstate(hpage))) {
1280 list_move_tail(&hpage->lru, ret);
1284 if (page_count(hpage) == 1) {
1285 /* page was freed from under us. So we are done. */
1286 putback_active_hugepage(hpage);
1287 return MIGRATEPAGE_SUCCESS;
1290 new_hpage = get_new_page(hpage, private);
1294 if (!trylock_page(hpage)) {
1299 case MIGRATE_SYNC_NO_COPY:
1308 * Check for pages which are in the process of being freed. Without
1309 * page_mapping() set, hugetlbfs specific move page routine will not
1310 * be called and we could leak usage counts for subpools.
1312 if (hugetlb_page_subpool(hpage) && !page_mapping(hpage)) {
1317 if (PageAnon(hpage))
1318 anon_vma = page_get_anon_vma(hpage);
1320 if (unlikely(!trylock_page(new_hpage)))
1323 if (page_mapped(hpage)) {
1324 bool mapping_locked = false;
1325 enum ttu_flags ttu = 0;
1327 if (!PageAnon(hpage)) {
1329 * In shared mappings, try_to_unmap could potentially
1330 * call huge_pmd_unshare. Because of this, take
1331 * semaphore in write mode here and set TTU_RMAP_LOCKED
1332 * to let lower levels know we have taken the lock.
1334 mapping = hugetlb_page_mapping_lock_write(hpage);
1335 if (unlikely(!mapping))
1336 goto unlock_put_anon;
1338 mapping_locked = true;
1339 ttu |= TTU_RMAP_LOCKED;
1342 try_to_migrate(hpage, ttu);
1343 page_was_mapped = 1;
1346 i_mmap_unlock_write(mapping);
1349 if (!page_mapped(hpage))
1350 rc = move_to_new_page(new_hpage, hpage, mode);
1352 if (page_was_mapped)
1353 remove_migration_ptes(hpage,
1354 rc == MIGRATEPAGE_SUCCESS ? new_hpage : hpage, false);
1357 unlock_page(new_hpage);
1361 put_anon_vma(anon_vma);
1363 if (rc == MIGRATEPAGE_SUCCESS) {
1364 move_hugetlb_state(hpage, new_hpage, reason);
1365 put_new_page = NULL;
1371 if (rc == MIGRATEPAGE_SUCCESS)
1372 putback_active_hugepage(hpage);
1373 else if (rc != -EAGAIN)
1374 list_move_tail(&hpage->lru, ret);
1377 * If migration was not successful and there's a freeing callback, use
1378 * it. Otherwise, put_page() will drop the reference grabbed during
1382 put_new_page(new_hpage, private);
1384 putback_active_hugepage(new_hpage);
1389 static inline int try_split_thp(struct page *page, struct page **page2,
1390 struct list_head *from)
1395 rc = split_huge_page_to_list(page, from);
1398 list_safe_reset_next(page, *page2, lru);
1404 * migrate_pages - migrate the pages specified in a list, to the free pages
1405 * supplied as the target for the page migration
1407 * @from: The list of pages to be migrated.
1408 * @get_new_page: The function used to allocate free pages to be used
1409 * as the target of the page migration.
1410 * @put_new_page: The function used to free target pages if migration
1411 * fails, or NULL if no special handling is necessary.
1412 * @private: Private data to be passed on to get_new_page()
1413 * @mode: The migration mode that specifies the constraints for
1414 * page migration, if any.
1415 * @reason: The reason for page migration.
1416 * @ret_succeeded: Set to the number of pages migrated successfully if
1417 * the caller passes a non-NULL pointer.
1419 * The function returns after 10 attempts or if no pages are movable any more
1420 * because the list has become empty or no retryable pages exist any more.
1421 * It is caller's responsibility to call putback_movable_pages() to return pages
1422 * to the LRU or free list only if ret != 0.
1424 * Returns the number of pages that were not migrated, or an error code.
1426 int migrate_pages(struct list_head *from, new_page_t get_new_page,
1427 free_page_t put_new_page, unsigned long private,
1428 enum migrate_mode mode, int reason, unsigned int *ret_succeeded)
1433 int nr_succeeded = 0;
1434 int nr_thp_succeeded = 0;
1435 int nr_thp_failed = 0;
1436 int nr_thp_split = 0;
1438 bool is_thp = false;
1441 int swapwrite = current->flags & PF_SWAPWRITE;
1442 int rc, nr_subpages;
1443 LIST_HEAD(ret_pages);
1444 bool nosplit = (reason == MR_NUMA_MISPLACED);
1446 trace_mm_migrate_pages_start(mode, reason);
1449 current->flags |= PF_SWAPWRITE;
1451 for (pass = 0; pass < 10 && (retry || thp_retry); pass++) {
1455 list_for_each_entry_safe(page, page2, from, lru) {
1458 * THP statistics is based on the source huge page.
1459 * Capture required information that might get lost
1462 is_thp = PageTransHuge(page) && !PageHuge(page);
1463 nr_subpages = thp_nr_pages(page);
1467 rc = unmap_and_move_huge_page(get_new_page,
1468 put_new_page, private, page,
1469 pass > 2, mode, reason,
1472 rc = unmap_and_move(get_new_page, put_new_page,
1473 private, page, pass > 2, mode,
1474 reason, &ret_pages);
1477 * Success: non hugetlb page will be freed, hugetlb
1478 * page will be put back
1479 * -EAGAIN: stay on the from list
1480 * -ENOMEM: stay on the from list
1481 * Other errno: put on ret_pages list then splice to
1486 * THP migration might be unsupported or the
1487 * allocation could've failed so we should
1488 * retry on the same page with the THP split
1491 * Head page is retried immediately and tail
1492 * pages are added to the tail of the list so
1493 * we encounter them after the rest of the list
1497 /* THP migration is unsupported */
1499 if (!try_split_thp(page, &page2, from)) {
1505 nr_failed += nr_subpages;
1509 /* Hugetlb migration is unsupported */
1514 * When memory is low, don't bother to try to migrate
1515 * other pages, just exit.
1516 * THP NUMA faulting doesn't split THP to retry.
1518 if (is_thp && !nosplit) {
1519 if (!try_split_thp(page, &page2, from)) {
1525 nr_failed += nr_subpages;
1537 case MIGRATEPAGE_SUCCESS:
1540 nr_succeeded += nr_subpages;
1547 * Permanent failure (-EBUSY, etc.):
1548 * unlike -EAGAIN case, the failed page is
1549 * removed from migration page list and not
1550 * retried in the next outer loop.
1554 nr_failed += nr_subpages;
1562 nr_failed += retry + thp_retry;
1563 nr_thp_failed += thp_retry;
1567 * Put the permanent failure page back to migration list, they
1568 * will be put back to the right list by the caller.
1570 list_splice(&ret_pages, from);
1572 count_vm_events(PGMIGRATE_SUCCESS, nr_succeeded);
1573 count_vm_events(PGMIGRATE_FAIL, nr_failed);
1574 count_vm_events(THP_MIGRATION_SUCCESS, nr_thp_succeeded);
1575 count_vm_events(THP_MIGRATION_FAIL, nr_thp_failed);
1576 count_vm_events(THP_MIGRATION_SPLIT, nr_thp_split);
1577 trace_mm_migrate_pages(nr_succeeded, nr_failed, nr_thp_succeeded,
1578 nr_thp_failed, nr_thp_split, mode, reason);
1581 current->flags &= ~PF_SWAPWRITE;
1584 *ret_succeeded = nr_succeeded;
1589 struct page *alloc_migration_target(struct page *page, unsigned long private)
1591 struct migration_target_control *mtc;
1593 unsigned int order = 0;
1594 struct page *new_page = NULL;
1598 mtc = (struct migration_target_control *)private;
1599 gfp_mask = mtc->gfp_mask;
1601 if (nid == NUMA_NO_NODE)
1602 nid = page_to_nid(page);
1604 if (PageHuge(page)) {
1605 struct hstate *h = page_hstate(compound_head(page));
1607 gfp_mask = htlb_modify_alloc_mask(h, gfp_mask);
1608 return alloc_huge_page_nodemask(h, nid, mtc->nmask, gfp_mask);
1611 if (PageTransHuge(page)) {
1613 * clear __GFP_RECLAIM to make the migration callback
1614 * consistent with regular THP allocations.
1616 gfp_mask &= ~__GFP_RECLAIM;
1617 gfp_mask |= GFP_TRANSHUGE;
1618 order = HPAGE_PMD_ORDER;
1620 zidx = zone_idx(page_zone(page));
1621 if (is_highmem_idx(zidx) || zidx == ZONE_MOVABLE)
1622 gfp_mask |= __GFP_HIGHMEM;
1624 new_page = __alloc_pages(gfp_mask, order, nid, mtc->nmask);
1626 if (new_page && PageTransHuge(new_page))
1627 prep_transhuge_page(new_page);
1634 static int store_status(int __user *status, int start, int value, int nr)
1637 if (put_user(value, status + start))
1645 static int do_move_pages_to_node(struct mm_struct *mm,
1646 struct list_head *pagelist, int node)
1649 struct migration_target_control mtc = {
1651 .gfp_mask = GFP_HIGHUSER_MOVABLE | __GFP_THISNODE,
1654 err = migrate_pages(pagelist, alloc_migration_target, NULL,
1655 (unsigned long)&mtc, MIGRATE_SYNC, MR_SYSCALL, NULL);
1657 putback_movable_pages(pagelist);
1662 * Resolves the given address to a struct page, isolates it from the LRU and
1663 * puts it to the given pagelist.
1665 * errno - if the page cannot be found/isolated
1666 * 0 - when it doesn't have to be migrated because it is already on the
1668 * 1 - when it has been queued
1670 static int add_page_for_migration(struct mm_struct *mm, unsigned long addr,
1671 int node, struct list_head *pagelist, bool migrate_all)
1673 struct vm_area_struct *vma;
1675 unsigned int follflags;
1680 vma = find_vma(mm, addr);
1681 if (!vma || addr < vma->vm_start || !vma_migratable(vma))
1684 /* FOLL_DUMP to ignore special (like zero) pages */
1685 follflags = FOLL_GET | FOLL_DUMP;
1686 page = follow_page(vma, addr, follflags);
1688 err = PTR_ERR(page);
1697 if (page_to_nid(page) == node)
1701 if (page_mapcount(page) > 1 && !migrate_all)
1704 if (PageHuge(page)) {
1705 if (PageHead(page)) {
1706 isolate_huge_page(page, pagelist);
1712 head = compound_head(page);
1713 err = isolate_lru_page(head);
1718 list_add_tail(&head->lru, pagelist);
1719 mod_node_page_state(page_pgdat(head),
1720 NR_ISOLATED_ANON + page_is_file_lru(head),
1721 thp_nr_pages(head));
1725 * Either remove the duplicate refcount from
1726 * isolate_lru_page() or drop the page ref if it was
1731 mmap_read_unlock(mm);
1735 static int move_pages_and_store_status(struct mm_struct *mm, int node,
1736 struct list_head *pagelist, int __user *status,
1737 int start, int i, unsigned long nr_pages)
1741 if (list_empty(pagelist))
1744 err = do_move_pages_to_node(mm, pagelist, node);
1747 * Positive err means the number of failed
1748 * pages to migrate. Since we are going to
1749 * abort and return the number of non-migrated
1750 * pages, so need to include the rest of the
1751 * nr_pages that have not been attempted as
1755 err += nr_pages - i - 1;
1758 return store_status(status, start, node, i - start);
1762 * Migrate an array of page address onto an array of nodes and fill
1763 * the corresponding array of status.
1765 static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes,
1766 unsigned long nr_pages,
1767 const void __user * __user *pages,
1768 const int __user *nodes,
1769 int __user *status, int flags)
1771 int current_node = NUMA_NO_NODE;
1772 LIST_HEAD(pagelist);
1776 lru_cache_disable();
1778 for (i = start = 0; i < nr_pages; i++) {
1779 const void __user *p;
1784 if (get_user(p, pages + i))
1786 if (get_user(node, nodes + i))
1788 addr = (unsigned long)untagged_addr(p);
1791 if (node < 0 || node >= MAX_NUMNODES)
1793 if (!node_state(node, N_MEMORY))
1797 if (!node_isset(node, task_nodes))
1800 if (current_node == NUMA_NO_NODE) {
1801 current_node = node;
1803 } else if (node != current_node) {
1804 err = move_pages_and_store_status(mm, current_node,
1805 &pagelist, status, start, i, nr_pages);
1809 current_node = node;
1813 * Errors in the page lookup or isolation are not fatal and we simply
1814 * report them via status
1816 err = add_page_for_migration(mm, addr, current_node,
1817 &pagelist, flags & MPOL_MF_MOVE_ALL);
1820 /* The page is successfully queued for migration */
1825 * If the page is already on the target node (!err), store the
1826 * node, otherwise, store the err.
1828 err = store_status(status, i, err ? : current_node, 1);
1832 err = move_pages_and_store_status(mm, current_node, &pagelist,
1833 status, start, i, nr_pages);
1836 current_node = NUMA_NO_NODE;
1839 /* Make sure we do not overwrite the existing error */
1840 err1 = move_pages_and_store_status(mm, current_node, &pagelist,
1841 status, start, i, nr_pages);
1850 * Determine the nodes of an array of pages and store it in an array of status.
1852 static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages,
1853 const void __user **pages, int *status)
1859 for (i = 0; i < nr_pages; i++) {
1860 unsigned long addr = (unsigned long)(*pages);
1861 struct vm_area_struct *vma;
1865 vma = vma_lookup(mm, addr);
1869 /* FOLL_DUMP to ignore special (like zero) pages */
1870 page = follow_page(vma, addr, FOLL_DUMP);
1872 err = PTR_ERR(page);
1876 err = page ? page_to_nid(page) : -ENOENT;
1884 mmap_read_unlock(mm);
1887 static int get_compat_pages_array(const void __user *chunk_pages[],
1888 const void __user * __user *pages,
1889 unsigned long chunk_nr)
1891 compat_uptr_t __user *pages32 = (compat_uptr_t __user *)pages;
1895 for (i = 0; i < chunk_nr; i++) {
1896 if (get_user(p, pages32 + i))
1898 chunk_pages[i] = compat_ptr(p);
1905 * Determine the nodes of a user array of pages and store it in
1906 * a user array of status.
1908 static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages,
1909 const void __user * __user *pages,
1912 #define DO_PAGES_STAT_CHUNK_NR 16
1913 const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR];
1914 int chunk_status[DO_PAGES_STAT_CHUNK_NR];
1917 unsigned long chunk_nr;
1919 chunk_nr = nr_pages;
1920 if (chunk_nr > DO_PAGES_STAT_CHUNK_NR)
1921 chunk_nr = DO_PAGES_STAT_CHUNK_NR;
1923 if (in_compat_syscall()) {
1924 if (get_compat_pages_array(chunk_pages, pages,
1928 if (copy_from_user(chunk_pages, pages,
1929 chunk_nr * sizeof(*chunk_pages)))
1933 do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status);
1935 if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status)))
1940 nr_pages -= chunk_nr;
1942 return nr_pages ? -EFAULT : 0;
1945 static struct mm_struct *find_mm_struct(pid_t pid, nodemask_t *mem_nodes)
1947 struct task_struct *task;
1948 struct mm_struct *mm;
1951 * There is no need to check if current process has the right to modify
1952 * the specified process when they are same.
1956 *mem_nodes = cpuset_mems_allowed(current);
1960 /* Find the mm_struct */
1962 task = find_task_by_vpid(pid);
1965 return ERR_PTR(-ESRCH);
1967 get_task_struct(task);
1970 * Check if this process has the right to modify the specified
1971 * process. Use the regular "ptrace_may_access()" checks.
1973 if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) {
1975 mm = ERR_PTR(-EPERM);
1980 mm = ERR_PTR(security_task_movememory(task));
1983 *mem_nodes = cpuset_mems_allowed(task);
1984 mm = get_task_mm(task);
1986 put_task_struct(task);
1988 mm = ERR_PTR(-EINVAL);
1993 * Move a list of pages in the address space of the currently executing
1996 static int kernel_move_pages(pid_t pid, unsigned long nr_pages,
1997 const void __user * __user *pages,
1998 const int __user *nodes,
1999 int __user *status, int flags)
2001 struct mm_struct *mm;
2003 nodemask_t task_nodes;
2006 if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
2009 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
2012 mm = find_mm_struct(pid, &task_nodes);
2017 err = do_pages_move(mm, task_nodes, nr_pages, pages,
2018 nodes, status, flags);
2020 err = do_pages_stat(mm, nr_pages, pages, status);
2026 SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages,
2027 const void __user * __user *, pages,
2028 const int __user *, nodes,
2029 int __user *, status, int, flags)
2031 return kernel_move_pages(pid, nr_pages, pages, nodes, status, flags);
2034 #ifdef CONFIG_NUMA_BALANCING
2036 * Returns true if this is a safe migration target node for misplaced NUMA
2037 * pages. Currently it only checks the watermarks which crude
2039 static bool migrate_balanced_pgdat(struct pglist_data *pgdat,
2040 unsigned long nr_migrate_pages)
2044 for (z = pgdat->nr_zones - 1; z >= 0; z--) {
2045 struct zone *zone = pgdat->node_zones + z;
2047 if (!populated_zone(zone))
2050 /* Avoid waking kswapd by allocating pages_to_migrate pages. */
2051 if (!zone_watermark_ok(zone, 0,
2052 high_wmark_pages(zone) +
2061 static struct page *alloc_misplaced_dst_page(struct page *page,
2064 int nid = (int) data;
2065 struct page *newpage;
2067 newpage = __alloc_pages_node(nid,
2068 (GFP_HIGHUSER_MOVABLE |
2069 __GFP_THISNODE | __GFP_NOMEMALLOC |
2070 __GFP_NORETRY | __GFP_NOWARN) &
2076 static struct page *alloc_misplaced_dst_page_thp(struct page *page,
2079 int nid = (int) data;
2080 struct page *newpage;
2082 newpage = alloc_pages_node(nid, (GFP_TRANSHUGE_LIGHT | __GFP_THISNODE),
2087 prep_transhuge_page(newpage);
2093 static int numamigrate_isolate_page(pg_data_t *pgdat, struct page *page)
2096 int nr_pages = thp_nr_pages(page);
2098 VM_BUG_ON_PAGE(compound_order(page) && !PageTransHuge(page), page);
2100 /* Do not migrate THP mapped by multiple processes */
2101 if (PageTransHuge(page) && total_mapcount(page) > 1)
2104 /* Avoid migrating to a node that is nearly full */
2105 if (!migrate_balanced_pgdat(pgdat, nr_pages))
2108 if (isolate_lru_page(page))
2111 page_lru = page_is_file_lru(page);
2112 mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON + page_lru,
2116 * Isolating the page has taken another reference, so the
2117 * caller's reference can be safely dropped without the page
2118 * disappearing underneath us during migration.
2125 * Attempt to migrate a misplaced page to the specified destination
2126 * node. Caller is expected to have an elevated reference count on
2127 * the page that will be dropped by this function before returning.
2129 int migrate_misplaced_page(struct page *page, struct vm_area_struct *vma,
2132 pg_data_t *pgdat = NODE_DATA(node);
2135 LIST_HEAD(migratepages);
2138 int nr_pages = thp_nr_pages(page);
2141 * PTE mapped THP or HugeTLB page can't reach here so the page could
2142 * be either base page or THP. And it must be head page if it is
2145 compound = PageTransHuge(page);
2148 new = alloc_misplaced_dst_page_thp;
2150 new = alloc_misplaced_dst_page;
2153 * Don't migrate file pages that are mapped in multiple processes
2154 * with execute permissions as they are probably shared libraries.
2156 if (page_mapcount(page) != 1 && page_is_file_lru(page) &&
2157 (vma->vm_flags & VM_EXEC))
2161 * Also do not migrate dirty pages as not all filesystems can move
2162 * dirty pages in MIGRATE_ASYNC mode which is a waste of cycles.
2164 if (page_is_file_lru(page) && PageDirty(page))
2167 isolated = numamigrate_isolate_page(pgdat, page);
2171 list_add(&page->lru, &migratepages);
2172 nr_remaining = migrate_pages(&migratepages, *new, NULL, node,
2173 MIGRATE_ASYNC, MR_NUMA_MISPLACED, NULL);
2175 if (!list_empty(&migratepages)) {
2176 list_del(&page->lru);
2177 mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON +
2178 page_is_file_lru(page), -nr_pages);
2179 putback_lru_page(page);
2183 count_vm_numa_events(NUMA_PAGE_MIGRATE, nr_pages);
2184 BUG_ON(!list_empty(&migratepages));
2191 #endif /* CONFIG_NUMA_BALANCING */
2192 #endif /* CONFIG_NUMA */
2194 #ifdef CONFIG_DEVICE_PRIVATE
2195 static int migrate_vma_collect_skip(unsigned long start,
2197 struct mm_walk *walk)
2199 struct migrate_vma *migrate = walk->private;
2202 for (addr = start; addr < end; addr += PAGE_SIZE) {
2203 migrate->dst[migrate->npages] = 0;
2204 migrate->src[migrate->npages++] = 0;
2210 static int migrate_vma_collect_hole(unsigned long start,
2212 __always_unused int depth,
2213 struct mm_walk *walk)
2215 struct migrate_vma *migrate = walk->private;
2218 /* Only allow populating anonymous memory. */
2219 if (!vma_is_anonymous(walk->vma))
2220 return migrate_vma_collect_skip(start, end, walk);
2222 for (addr = start; addr < end; addr += PAGE_SIZE) {
2223 migrate->src[migrate->npages] = MIGRATE_PFN_MIGRATE;
2224 migrate->dst[migrate->npages] = 0;
2232 static int migrate_vma_collect_pmd(pmd_t *pmdp,
2233 unsigned long start,
2235 struct mm_walk *walk)
2237 struct migrate_vma *migrate = walk->private;
2238 struct vm_area_struct *vma = walk->vma;
2239 struct mm_struct *mm = vma->vm_mm;
2240 unsigned long addr = start, unmapped = 0;
2245 if (pmd_none(*pmdp))
2246 return migrate_vma_collect_hole(start, end, -1, walk);
2248 if (pmd_trans_huge(*pmdp)) {
2251 ptl = pmd_lock(mm, pmdp);
2252 if (unlikely(!pmd_trans_huge(*pmdp))) {
2257 page = pmd_page(*pmdp);
2258 if (is_huge_zero_page(page)) {
2260 split_huge_pmd(vma, pmdp, addr);
2261 if (pmd_trans_unstable(pmdp))
2262 return migrate_vma_collect_skip(start, end,
2269 if (unlikely(!trylock_page(page)))
2270 return migrate_vma_collect_skip(start, end,
2272 ret = split_huge_page(page);
2276 return migrate_vma_collect_skip(start, end,
2278 if (pmd_none(*pmdp))
2279 return migrate_vma_collect_hole(start, end, -1,
2284 if (unlikely(pmd_bad(*pmdp)))
2285 return migrate_vma_collect_skip(start, end, walk);
2287 ptep = pte_offset_map_lock(mm, pmdp, addr, &ptl);
2288 arch_enter_lazy_mmu_mode();
2290 for (; addr < end; addr += PAGE_SIZE, ptep++) {
2291 unsigned long mpfn = 0, pfn;
2298 if (pte_none(pte)) {
2299 if (vma_is_anonymous(vma)) {
2300 mpfn = MIGRATE_PFN_MIGRATE;
2306 if (!pte_present(pte)) {
2308 * Only care about unaddressable device page special
2309 * page table entry. Other special swap entries are not
2310 * migratable, and we ignore regular swapped page.
2312 entry = pte_to_swp_entry(pte);
2313 if (!is_device_private_entry(entry))
2316 page = pfn_swap_entry_to_page(entry);
2317 if (!(migrate->flags &
2318 MIGRATE_VMA_SELECT_DEVICE_PRIVATE) ||
2319 page->pgmap->owner != migrate->pgmap_owner)
2322 mpfn = migrate_pfn(page_to_pfn(page)) |
2323 MIGRATE_PFN_MIGRATE;
2324 if (is_writable_device_private_entry(entry))
2325 mpfn |= MIGRATE_PFN_WRITE;
2327 if (!(migrate->flags & MIGRATE_VMA_SELECT_SYSTEM))
2330 if (is_zero_pfn(pfn)) {
2331 mpfn = MIGRATE_PFN_MIGRATE;
2335 page = vm_normal_page(migrate->vma, addr, pte);
2336 mpfn = migrate_pfn(pfn) | MIGRATE_PFN_MIGRATE;
2337 mpfn |= pte_write(pte) ? MIGRATE_PFN_WRITE : 0;
2340 /* FIXME support THP */
2341 if (!page || !page->mapping || PageTransCompound(page)) {
2347 * By getting a reference on the page we pin it and that blocks
2348 * any kind of migration. Side effect is that it "freezes" the
2351 * We drop this reference after isolating the page from the lru
2352 * for non device page (device page are not on the lru and thus
2353 * can't be dropped from it).
2358 * Optimize for the common case where page is only mapped once
2359 * in one process. If we can lock the page, then we can safely
2360 * set up a special migration page table entry now.
2362 if (trylock_page(page)) {
2366 ptep_get_and_clear(mm, addr, ptep);
2368 /* Setup special migration page table entry */
2369 if (mpfn & MIGRATE_PFN_WRITE)
2370 entry = make_writable_migration_entry(
2373 entry = make_readable_migration_entry(
2375 swp_pte = swp_entry_to_pte(entry);
2376 if (pte_present(pte)) {
2377 if (pte_soft_dirty(pte))
2378 swp_pte = pte_swp_mksoft_dirty(swp_pte);
2379 if (pte_uffd_wp(pte))
2380 swp_pte = pte_swp_mkuffd_wp(swp_pte);
2382 if (pte_swp_soft_dirty(pte))
2383 swp_pte = pte_swp_mksoft_dirty(swp_pte);
2384 if (pte_swp_uffd_wp(pte))
2385 swp_pte = pte_swp_mkuffd_wp(swp_pte);
2387 set_pte_at(mm, addr, ptep, swp_pte);
2390 * This is like regular unmap: we remove the rmap and
2391 * drop page refcount. Page won't be freed, as we took
2392 * a reference just above.
2394 page_remove_rmap(page, false);
2397 if (pte_present(pte))
2405 migrate->dst[migrate->npages] = 0;
2406 migrate->src[migrate->npages++] = mpfn;
2408 arch_leave_lazy_mmu_mode();
2409 pte_unmap_unlock(ptep - 1, ptl);
2411 /* Only flush the TLB if we actually modified any entries */
2413 flush_tlb_range(walk->vma, start, end);
2418 static const struct mm_walk_ops migrate_vma_walk_ops = {
2419 .pmd_entry = migrate_vma_collect_pmd,
2420 .pte_hole = migrate_vma_collect_hole,
2424 * migrate_vma_collect() - collect pages over a range of virtual addresses
2425 * @migrate: migrate struct containing all migration information
2427 * This will walk the CPU page table. For each virtual address backed by a
2428 * valid page, it updates the src array and takes a reference on the page, in
2429 * order to pin the page until we lock it and unmap it.
2431 static void migrate_vma_collect(struct migrate_vma *migrate)
2433 struct mmu_notifier_range range;
2436 * Note that the pgmap_owner is passed to the mmu notifier callback so
2437 * that the registered device driver can skip invalidating device
2438 * private page mappings that won't be migrated.
2440 mmu_notifier_range_init_owner(&range, MMU_NOTIFY_MIGRATE, 0,
2441 migrate->vma, migrate->vma->vm_mm, migrate->start, migrate->end,
2442 migrate->pgmap_owner);
2443 mmu_notifier_invalidate_range_start(&range);
2445 walk_page_range(migrate->vma->vm_mm, migrate->start, migrate->end,
2446 &migrate_vma_walk_ops, migrate);
2448 mmu_notifier_invalidate_range_end(&range);
2449 migrate->end = migrate->start + (migrate->npages << PAGE_SHIFT);
2453 * migrate_vma_check_page() - check if page is pinned or not
2454 * @page: struct page to check
2456 * Pinned pages cannot be migrated. This is the same test as in
2457 * folio_migrate_mapping(), except that here we allow migration of a
2460 static bool migrate_vma_check_page(struct page *page)
2463 * One extra ref because caller holds an extra reference, either from
2464 * isolate_lru_page() for a regular page, or migrate_vma_collect() for
2470 * FIXME support THP (transparent huge page), it is bit more complex to
2471 * check them than regular pages, because they can be mapped with a pmd
2472 * or with a pte (split pte mapping).
2474 if (PageCompound(page))
2477 /* Page from ZONE_DEVICE have one extra reference */
2478 if (is_zone_device_page(page)) {
2480 * Private page can never be pin as they have no valid pte and
2481 * GUP will fail for those. Yet if there is a pending migration
2482 * a thread might try to wait on the pte migration entry and
2483 * will bump the page reference count. Sadly there is no way to
2484 * differentiate a regular pin from migration wait. Hence to
2485 * avoid 2 racing thread trying to migrate back to CPU to enter
2486 * infinite loop (one stopping migration because the other is
2487 * waiting on pte migration entry). We always return true here.
2489 * FIXME proper solution is to rework migration_entry_wait() so
2490 * it does not need to take a reference on page.
2492 return is_device_private_page(page);
2495 /* For file back page */
2496 if (page_mapping(page))
2497 extra += 1 + page_has_private(page);
2499 if ((page_count(page) - extra) > page_mapcount(page))
2506 * migrate_vma_unmap() - replace page mapping with special migration pte entry
2507 * @migrate: migrate struct containing all migration information
2509 * Isolate pages from the LRU and replace mappings (CPU page table pte) with a
2510 * special migration pte entry and check if it has been pinned. Pinned pages are
2511 * restored because we cannot migrate them.
2513 * This is the last step before we call the device driver callback to allocate
2514 * destination memory and copy contents of original page over to new page.
2516 static void migrate_vma_unmap(struct migrate_vma *migrate)
2518 const unsigned long npages = migrate->npages;
2519 const unsigned long start = migrate->start;
2520 unsigned long addr, i, restore = 0;
2521 bool allow_drain = true;
2525 for (i = 0; i < npages; i++) {
2526 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2531 /* ZONE_DEVICE pages are not on LRU */
2532 if (!is_zone_device_page(page)) {
2533 if (!PageLRU(page) && allow_drain) {
2534 /* Drain CPU's pagevec */
2535 lru_add_drain_all();
2536 allow_drain = false;
2539 if (isolate_lru_page(page)) {
2540 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2546 /* Drop the reference we took in collect */
2550 if (page_mapped(page))
2551 try_to_migrate(page, 0);
2553 if (page_mapped(page) || !migrate_vma_check_page(page)) {
2554 if (!is_zone_device_page(page)) {
2556 putback_lru_page(page);
2559 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2566 for (addr = start, i = 0; i < npages && restore; addr += PAGE_SIZE, i++) {
2567 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2569 if (!page || (migrate->src[i] & MIGRATE_PFN_MIGRATE))
2572 remove_migration_ptes(page, page, false);
2574 migrate->src[i] = 0;
2582 * migrate_vma_setup() - prepare to migrate a range of memory
2583 * @args: contains the vma, start, and pfns arrays for the migration
2585 * Returns: negative errno on failures, 0 when 0 or more pages were migrated
2588 * Prepare to migrate a range of memory virtual address range by collecting all
2589 * the pages backing each virtual address in the range, saving them inside the
2590 * src array. Then lock those pages and unmap them. Once the pages are locked
2591 * and unmapped, check whether each page is pinned or not. Pages that aren't
2592 * pinned have the MIGRATE_PFN_MIGRATE flag set (by this function) in the
2593 * corresponding src array entry. Then restores any pages that are pinned, by
2594 * remapping and unlocking those pages.
2596 * The caller should then allocate destination memory and copy source memory to
2597 * it for all those entries (ie with MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE
2598 * flag set). Once these are allocated and copied, the caller must update each
2599 * corresponding entry in the dst array with the pfn value of the destination
2600 * page and with MIGRATE_PFN_VALID. Destination pages must be locked via
2603 * Note that the caller does not have to migrate all the pages that are marked
2604 * with MIGRATE_PFN_MIGRATE flag in src array unless this is a migration from
2605 * device memory to system memory. If the caller cannot migrate a device page
2606 * back to system memory, then it must return VM_FAULT_SIGBUS, which has severe
2607 * consequences for the userspace process, so it must be avoided if at all
2610 * For empty entries inside CPU page table (pte_none() or pmd_none() is true) we
2611 * do set MIGRATE_PFN_MIGRATE flag inside the corresponding source array thus
2612 * allowing the caller to allocate device memory for those unbacked virtual
2613 * addresses. For this the caller simply has to allocate device memory and
2614 * properly set the destination entry like for regular migration. Note that
2615 * this can still fail, and thus inside the device driver you must check if the
2616 * migration was successful for those entries after calling migrate_vma_pages(),
2617 * just like for regular migration.
2619 * After that, the callers must call migrate_vma_pages() to go over each entry
2620 * in the src array that has the MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE flag
2621 * set. If the corresponding entry in dst array has MIGRATE_PFN_VALID flag set,
2622 * then migrate_vma_pages() to migrate struct page information from the source
2623 * struct page to the destination struct page. If it fails to migrate the
2624 * struct page information, then it clears the MIGRATE_PFN_MIGRATE flag in the
2627 * At this point all successfully migrated pages have an entry in the src
2628 * array with MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE flag set and the dst
2629 * array entry with MIGRATE_PFN_VALID flag set.
2631 * Once migrate_vma_pages() returns the caller may inspect which pages were
2632 * successfully migrated, and which were not. Successfully migrated pages will
2633 * have the MIGRATE_PFN_MIGRATE flag set for their src array entry.
2635 * It is safe to update device page table after migrate_vma_pages() because
2636 * both destination and source page are still locked, and the mmap_lock is held
2637 * in read mode (hence no one can unmap the range being migrated).
2639 * Once the caller is done cleaning up things and updating its page table (if it
2640 * chose to do so, this is not an obligation) it finally calls
2641 * migrate_vma_finalize() to update the CPU page table to point to new pages
2642 * for successfully migrated pages or otherwise restore the CPU page table to
2643 * point to the original source pages.
2645 int migrate_vma_setup(struct migrate_vma *args)
2647 long nr_pages = (args->end - args->start) >> PAGE_SHIFT;
2649 args->start &= PAGE_MASK;
2650 args->end &= PAGE_MASK;
2651 if (!args->vma || is_vm_hugetlb_page(args->vma) ||
2652 (args->vma->vm_flags & VM_SPECIAL) || vma_is_dax(args->vma))
2656 if (args->start < args->vma->vm_start ||
2657 args->start >= args->vma->vm_end)
2659 if (args->end <= args->vma->vm_start || args->end > args->vma->vm_end)
2661 if (!args->src || !args->dst)
2664 memset(args->src, 0, sizeof(*args->src) * nr_pages);
2668 migrate_vma_collect(args);
2671 migrate_vma_unmap(args);
2674 * At this point pages are locked and unmapped, and thus they have
2675 * stable content and can safely be copied to destination memory that
2676 * is allocated by the drivers.
2681 EXPORT_SYMBOL(migrate_vma_setup);
2684 * This code closely matches the code in:
2685 * __handle_mm_fault()
2686 * handle_pte_fault()
2687 * do_anonymous_page()
2688 * to map in an anonymous zero page but the struct page will be a ZONE_DEVICE
2691 static void migrate_vma_insert_page(struct migrate_vma *migrate,
2696 struct vm_area_struct *vma = migrate->vma;
2697 struct mm_struct *mm = vma->vm_mm;
2707 /* Only allow populating anonymous memory */
2708 if (!vma_is_anonymous(vma))
2711 pgdp = pgd_offset(mm, addr);
2712 p4dp = p4d_alloc(mm, pgdp, addr);
2715 pudp = pud_alloc(mm, p4dp, addr);
2718 pmdp = pmd_alloc(mm, pudp, addr);
2722 if (pmd_trans_huge(*pmdp) || pmd_devmap(*pmdp))
2726 * Use pte_alloc() instead of pte_alloc_map(). We can't run
2727 * pte_offset_map() on pmds where a huge pmd might be created
2728 * from a different thread.
2730 * pte_alloc_map() is safe to use under mmap_write_lock(mm) or when
2731 * parallel threads are excluded by other means.
2733 * Here we only have mmap_read_lock(mm).
2735 if (pte_alloc(mm, pmdp))
2738 /* See the comment in pte_alloc_one_map() */
2739 if (unlikely(pmd_trans_unstable(pmdp)))
2742 if (unlikely(anon_vma_prepare(vma)))
2744 if (mem_cgroup_charge(page_folio(page), vma->vm_mm, GFP_KERNEL))
2748 * The memory barrier inside __SetPageUptodate makes sure that
2749 * preceding stores to the page contents become visible before
2750 * the set_pte_at() write.
2752 __SetPageUptodate(page);
2754 if (is_zone_device_page(page)) {
2755 if (is_device_private_page(page)) {
2756 swp_entry_t swp_entry;
2758 if (vma->vm_flags & VM_WRITE)
2759 swp_entry = make_writable_device_private_entry(
2762 swp_entry = make_readable_device_private_entry(
2764 entry = swp_entry_to_pte(swp_entry);
2767 * For now we only support migrating to un-addressable
2770 pr_warn_once("Unsupported ZONE_DEVICE page type.\n");
2774 entry = mk_pte(page, vma->vm_page_prot);
2775 if (vma->vm_flags & VM_WRITE)
2776 entry = pte_mkwrite(pte_mkdirty(entry));
2779 ptep = pte_offset_map_lock(mm, pmdp, addr, &ptl);
2781 if (check_stable_address_space(mm))
2784 if (pte_present(*ptep)) {
2785 unsigned long pfn = pte_pfn(*ptep);
2787 if (!is_zero_pfn(pfn))
2790 } else if (!pte_none(*ptep))
2794 * Check for userfaultfd but do not deliver the fault. Instead,
2797 if (userfaultfd_missing(vma))
2800 inc_mm_counter(mm, MM_ANONPAGES);
2801 page_add_new_anon_rmap(page, vma, addr, false);
2802 if (!is_zone_device_page(page))
2803 lru_cache_add_inactive_or_unevictable(page, vma);
2807 flush_cache_page(vma, addr, pte_pfn(*ptep));
2808 ptep_clear_flush_notify(vma, addr, ptep);
2809 set_pte_at_notify(mm, addr, ptep, entry);
2810 update_mmu_cache(vma, addr, ptep);
2812 /* No need to invalidate - it was non-present before */
2813 set_pte_at(mm, addr, ptep, entry);
2814 update_mmu_cache(vma, addr, ptep);
2817 pte_unmap_unlock(ptep, ptl);
2818 *src = MIGRATE_PFN_MIGRATE;
2822 pte_unmap_unlock(ptep, ptl);
2824 *src &= ~MIGRATE_PFN_MIGRATE;
2828 * migrate_vma_pages() - migrate meta-data from src page to dst page
2829 * @migrate: migrate struct containing all migration information
2831 * This migrates struct page meta-data from source struct page to destination
2832 * struct page. This effectively finishes the migration from source page to the
2835 void migrate_vma_pages(struct migrate_vma *migrate)
2837 const unsigned long npages = migrate->npages;
2838 const unsigned long start = migrate->start;
2839 struct mmu_notifier_range range;
2840 unsigned long addr, i;
2841 bool notified = false;
2843 for (i = 0, addr = start; i < npages; addr += PAGE_SIZE, i++) {
2844 struct page *newpage = migrate_pfn_to_page(migrate->dst[i]);
2845 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2846 struct address_space *mapping;
2850 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2855 if (!(migrate->src[i] & MIGRATE_PFN_MIGRATE))
2860 mmu_notifier_range_init_owner(&range,
2861 MMU_NOTIFY_MIGRATE, 0, migrate->vma,
2862 migrate->vma->vm_mm, addr, migrate->end,
2863 migrate->pgmap_owner);
2864 mmu_notifier_invalidate_range_start(&range);
2866 migrate_vma_insert_page(migrate, addr, newpage,
2871 mapping = page_mapping(page);
2873 if (is_zone_device_page(newpage)) {
2874 if (is_device_private_page(newpage)) {
2876 * For now only support private anonymous when
2877 * migrating to un-addressable device memory.
2880 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2885 * Other types of ZONE_DEVICE page are not
2888 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2893 r = migrate_page(mapping, newpage, page, MIGRATE_SYNC_NO_COPY);
2894 if (r != MIGRATEPAGE_SUCCESS)
2895 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2899 * No need to double call mmu_notifier->invalidate_range() callback as
2900 * the above ptep_clear_flush_notify() inside migrate_vma_insert_page()
2901 * did already call it.
2904 mmu_notifier_invalidate_range_only_end(&range);
2906 EXPORT_SYMBOL(migrate_vma_pages);
2909 * migrate_vma_finalize() - restore CPU page table entry
2910 * @migrate: migrate struct containing all migration information
2912 * This replaces the special migration pte entry with either a mapping to the
2913 * new page if migration was successful for that page, or to the original page
2916 * This also unlocks the pages and puts them back on the lru, or drops the extra
2917 * refcount, for device pages.
2919 void migrate_vma_finalize(struct migrate_vma *migrate)
2921 const unsigned long npages = migrate->npages;
2924 for (i = 0; i < npages; i++) {
2925 struct page *newpage = migrate_pfn_to_page(migrate->dst[i]);
2926 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2930 unlock_page(newpage);
2936 if (!(migrate->src[i] & MIGRATE_PFN_MIGRATE) || !newpage) {
2938 unlock_page(newpage);
2944 remove_migration_ptes(page, newpage, false);
2947 if (is_zone_device_page(page))
2950 putback_lru_page(page);
2952 if (newpage != page) {
2953 unlock_page(newpage);
2954 if (is_zone_device_page(newpage))
2957 putback_lru_page(newpage);
2961 EXPORT_SYMBOL(migrate_vma_finalize);
2962 #endif /* CONFIG_DEVICE_PRIVATE */
2964 #if defined(CONFIG_HOTPLUG_CPU)
2965 /* Disable reclaim-based migration. */
2966 static void __disable_all_migrate_targets(void)
2970 for_each_online_node(node)
2971 node_demotion[node] = NUMA_NO_NODE;
2974 static void disable_all_migrate_targets(void)
2976 __disable_all_migrate_targets();
2979 * Ensure that the "disable" is visible across the system.
2980 * Readers will see either a combination of before+disable
2981 * state or disable+after. They will never see before and
2982 * after state together.
2984 * The before+after state together might have cycles and
2985 * could cause readers to do things like loop until this
2986 * function finishes. This ensures they can only see a
2987 * single "bad" read and would, for instance, only loop
2994 * Find an automatic demotion target for 'node'.
2995 * Failing here is OK. It might just indicate
2996 * being at the end of a chain.
2998 static int establish_migrate_target(int node, nodemask_t *used)
3000 int migration_target;
3003 * Can not set a migration target on a
3004 * node with it already set.
3006 * No need for READ_ONCE() here since this
3007 * in the write path for node_demotion[].
3008 * This should be the only thread writing.
3010 if (node_demotion[node] != NUMA_NO_NODE)
3011 return NUMA_NO_NODE;
3013 migration_target = find_next_best_node(node, used);
3014 if (migration_target == NUMA_NO_NODE)
3015 return NUMA_NO_NODE;
3017 node_demotion[node] = migration_target;
3019 return migration_target;
3023 * When memory fills up on a node, memory contents can be
3024 * automatically migrated to another node instead of
3025 * discarded at reclaim.
3027 * Establish a "migration path" which will start at nodes
3028 * with CPUs and will follow the priorities used to build the
3029 * page allocator zonelists.
3031 * The difference here is that cycles must be avoided. If
3032 * node0 migrates to node1, then neither node1, nor anything
3033 * node1 migrates to can migrate to node0.
3035 * This function can run simultaneously with readers of
3036 * node_demotion[]. However, it can not run simultaneously
3037 * with itself. Exclusion is provided by memory hotplug events
3038 * being single-threaded.
3040 static void __set_migration_target_nodes(void)
3042 nodemask_t next_pass = NODE_MASK_NONE;
3043 nodemask_t this_pass = NODE_MASK_NONE;
3044 nodemask_t used_targets = NODE_MASK_NONE;
3048 * Avoid any oddities like cycles that could occur
3049 * from changes in the topology. This will leave
3050 * a momentary gap when migration is disabled.
3052 disable_all_migrate_targets();
3055 * Allocations go close to CPUs, first. Assume that
3056 * the migration path starts at the nodes with CPUs.
3058 next_pass = node_states[N_CPU];
3060 this_pass = next_pass;
3061 next_pass = NODE_MASK_NONE;
3063 * To avoid cycles in the migration "graph", ensure
3064 * that migration sources are not future targets by
3065 * setting them in 'used_targets'. Do this only
3066 * once per pass so that multiple source nodes can
3067 * share a target node.
3069 * 'used_targets' will become unavailable in future
3070 * passes. This limits some opportunities for
3071 * multiple source nodes to share a destination.
3073 nodes_or(used_targets, used_targets, this_pass);
3074 for_each_node_mask(node, this_pass) {
3075 int target_node = establish_migrate_target(node, &used_targets);
3077 if (target_node == NUMA_NO_NODE)
3081 * Visit targets from this pass in the next pass.
3082 * Eventually, every node will have been part of
3083 * a pass, and will become set in 'used_targets'.
3085 node_set(target_node, next_pass);
3088 * 'next_pass' contains nodes which became migration
3089 * targets in this pass. Make additional passes until
3090 * no more migrations targets are available.
3092 if (!nodes_empty(next_pass))
3097 * For callers that do not hold get_online_mems() already.
3099 static void set_migration_target_nodes(void)
3102 __set_migration_target_nodes();
3107 * This leaves migrate-on-reclaim transiently disabled between
3108 * the MEM_GOING_OFFLINE and MEM_OFFLINE events. This runs
3109 * whether reclaim-based migration is enabled or not, which
3110 * ensures that the user can turn reclaim-based migration at
3111 * any time without needing to recalculate migration targets.
3113 * These callbacks already hold get_online_mems(). That is why
3114 * __set_migration_target_nodes() can be used as opposed to
3115 * set_migration_target_nodes().
3117 static int __meminit migrate_on_reclaim_callback(struct notifier_block *self,
3118 unsigned long action, void *_arg)
3120 struct memory_notify *arg = _arg;
3123 * Only update the node migration order when a node is
3124 * changing status, like online->offline. This avoids
3125 * the overhead of synchronize_rcu() in most cases.
3127 if (arg->status_change_nid < 0)
3128 return notifier_from_errno(0);
3131 case MEM_GOING_OFFLINE:
3133 * Make sure there are not transient states where
3134 * an offline node is a migration target. This
3135 * will leave migration disabled until the offline
3136 * completes and the MEM_OFFLINE case below runs.
3138 disable_all_migrate_targets();
3143 * Recalculate the target nodes once the node
3144 * reaches its final state (online or offline).
3146 __set_migration_target_nodes();
3148 case MEM_CANCEL_OFFLINE:
3150 * MEM_GOING_OFFLINE disabled all the migration
3151 * targets. Reenable them.
3153 __set_migration_target_nodes();
3155 case MEM_GOING_ONLINE:
3156 case MEM_CANCEL_ONLINE:
3160 return notifier_from_errno(0);
3164 * React to hotplug events that might affect the migration targets
3165 * like events that online or offline NUMA nodes.
3167 * The ordering is also currently dependent on which nodes have
3168 * CPUs. That means we need CPU on/offline notification too.
3170 static int migration_online_cpu(unsigned int cpu)
3172 set_migration_target_nodes();
3176 static int migration_offline_cpu(unsigned int cpu)
3178 set_migration_target_nodes();
3182 static int __init migrate_on_reclaim_init(void)
3186 ret = cpuhp_setup_state_nocalls(CPUHP_MM_DEMOTION_DEAD, "mm/demotion:offline",
3187 NULL, migration_offline_cpu);
3189 * In the unlikely case that this fails, the automatic
3190 * migration targets may become suboptimal for nodes
3191 * where N_CPU changes. With such a small impact in a
3192 * rare case, do not bother trying to do anything special.
3195 ret = cpuhp_setup_state(CPUHP_AP_MM_DEMOTION_ONLINE, "mm/demotion:online",
3196 migration_online_cpu, NULL);
3199 hotplug_memory_notifier(migrate_on_reclaim_callback, 100);
3202 late_initcall(migrate_on_reclaim_init);
3203 #endif /* CONFIG_HOTPLUG_CPU */
3205 bool numa_demotion_enabled = false;
3208 static ssize_t numa_demotion_enabled_show(struct kobject *kobj,
3209 struct kobj_attribute *attr, char *buf)
3211 return sysfs_emit(buf, "%s\n",
3212 numa_demotion_enabled ? "true" : "false");
3215 static ssize_t numa_demotion_enabled_store(struct kobject *kobj,
3216 struct kobj_attribute *attr,
3217 const char *buf, size_t count)
3219 if (!strncmp(buf, "true", 4) || !strncmp(buf, "1", 1))
3220 numa_demotion_enabled = true;
3221 else if (!strncmp(buf, "false", 5) || !strncmp(buf, "0", 1))
3222 numa_demotion_enabled = false;
3229 static struct kobj_attribute numa_demotion_enabled_attr =
3230 __ATTR(demotion_enabled, 0644, numa_demotion_enabled_show,
3231 numa_demotion_enabled_store);
3233 static struct attribute *numa_attrs[] = {
3234 &numa_demotion_enabled_attr.attr,
3238 static const struct attribute_group numa_attr_group = {
3239 .attrs = numa_attrs,
3242 static int __init numa_init_sysfs(void)
3245 struct kobject *numa_kobj;
3247 numa_kobj = kobject_create_and_add("numa", mm_kobj);
3249 pr_err("failed to create numa kobject\n");
3252 err = sysfs_create_group(numa_kobj, &numa_attr_group);
3254 pr_err("failed to register numa group\n");
3260 kobject_put(numa_kobj);
3263 subsys_initcall(numa_init_sysfs);