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/pfn_t.h>
42 #include <linux/memremap.h>
43 #include <linux/userfaultfd_k.h>
44 #include <linux/balloon_compaction.h>
45 #include <linux/page_idle.h>
46 #include <linux/page_owner.h>
47 #include <linux/sched/mm.h>
48 #include <linux/ptrace.h>
49 #include <linux/oom.h>
50 #include <linux/memory.h>
51 #include <linux/random.h>
52 #include <linux/sched/sysctl.h>
54 #include <asm/tlbflush.h>
56 #include <trace/events/migrate.h>
60 int isolate_movable_page(struct page *page, isolate_mode_t mode)
62 const struct movable_operations *mops;
65 * Avoid burning cycles with pages that are yet under __free_pages(),
66 * or just got freed under us.
68 * In case we 'win' a race for a movable page being freed under us and
69 * raise its refcount preventing __free_pages() from doing its job
70 * the put_page() at the end of this block will take care of
71 * release this page, thus avoiding a nasty leakage.
73 if (unlikely(!get_page_unless_zero(page)))
77 * Check PageMovable before holding a PG_lock because page's owner
78 * assumes anybody doesn't touch PG_lock of newly allocated page
79 * so unconditionally grabbing the lock ruins page's owner side.
81 if (unlikely(!__PageMovable(page)))
84 * As movable pages are not isolated from LRU lists, concurrent
85 * compaction threads can race against page migration functions
86 * as well as race against the releasing a page.
88 * In order to avoid having an already isolated movable page
89 * being (wrongly) re-isolated while it is under migration,
90 * or to avoid attempting to isolate pages being released,
91 * lets be sure we have the page lock
92 * before proceeding with the movable page isolation steps.
94 if (unlikely(!trylock_page(page)))
97 if (!PageMovable(page) || PageIsolated(page))
100 mops = page_movable_ops(page);
101 VM_BUG_ON_PAGE(!mops, page);
103 if (!mops->isolate_page(page, mode))
104 goto out_no_isolated;
106 /* Driver shouldn't use PG_isolated bit of page->flags */
107 WARN_ON_ONCE(PageIsolated(page));
108 SetPageIsolated(page);
121 static void putback_movable_page(struct page *page)
123 const struct movable_operations *mops = page_movable_ops(page);
125 mops->putback_page(page);
126 ClearPageIsolated(page);
130 * Put previously isolated pages back onto the appropriate lists
131 * from where they were once taken off for compaction/migration.
133 * This function shall be used whenever the isolated pageset has been
134 * built from lru, balloon, hugetlbfs page. See isolate_migratepages_range()
135 * and isolate_huge_page().
137 void putback_movable_pages(struct list_head *l)
142 list_for_each_entry_safe(page, page2, l, lru) {
143 if (unlikely(PageHuge(page))) {
144 putback_active_hugepage(page);
147 list_del(&page->lru);
149 * We isolated non-lru movable page so here we can use
150 * __PageMovable because LRU page's mapping cannot have
151 * PAGE_MAPPING_MOVABLE.
153 if (unlikely(__PageMovable(page))) {
154 VM_BUG_ON_PAGE(!PageIsolated(page), page);
156 if (PageMovable(page))
157 putback_movable_page(page);
159 ClearPageIsolated(page);
163 mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON +
164 page_is_file_lru(page), -thp_nr_pages(page));
165 putback_lru_page(page);
171 * Restore a potential migration pte to a working pte entry
173 static bool remove_migration_pte(struct folio *folio,
174 struct vm_area_struct *vma, unsigned long addr, void *old)
176 DEFINE_FOLIO_VMA_WALK(pvmw, old, vma, addr, PVMW_SYNC | PVMW_MIGRATION);
178 while (page_vma_mapped_walk(&pvmw)) {
179 rmap_t rmap_flags = RMAP_NONE;
183 unsigned long idx = 0;
185 /* pgoff is invalid for ksm pages, but they are never large */
186 if (folio_test_large(folio) && !folio_test_hugetlb(folio))
187 idx = linear_page_index(vma, pvmw.address) - pvmw.pgoff;
188 new = folio_page(folio, idx);
190 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
191 /* PMD-mapped THP migration entry */
193 VM_BUG_ON_FOLIO(folio_test_hugetlb(folio) ||
194 !folio_test_pmd_mappable(folio), folio);
195 remove_migration_pmd(&pvmw, new);
201 pte = pte_mkold(mk_pte(new, READ_ONCE(vma->vm_page_prot)));
202 if (pte_swp_soft_dirty(*pvmw.pte))
203 pte = pte_mksoft_dirty(pte);
206 * Recheck VMA as permissions can change since migration started
208 entry = pte_to_swp_entry(*pvmw.pte);
209 if (is_writable_migration_entry(entry))
210 pte = maybe_mkwrite(pte, vma);
211 else if (pte_swp_uffd_wp(*pvmw.pte))
212 pte = pte_mkuffd_wp(pte);
214 if (folio_test_anon(folio) && !is_readable_migration_entry(entry))
215 rmap_flags |= RMAP_EXCLUSIVE;
217 if (unlikely(is_device_private_page(new))) {
219 entry = make_writable_device_private_entry(
222 entry = make_readable_device_private_entry(
224 pte = swp_entry_to_pte(entry);
225 if (pte_swp_soft_dirty(*pvmw.pte))
226 pte = pte_swp_mksoft_dirty(pte);
227 if (pte_swp_uffd_wp(*pvmw.pte))
228 pte = pte_swp_mkuffd_wp(pte);
231 #ifdef CONFIG_HUGETLB_PAGE
232 if (folio_test_hugetlb(folio)) {
233 unsigned int shift = huge_page_shift(hstate_vma(vma));
235 pte = pte_mkhuge(pte);
236 pte = arch_make_huge_pte(pte, shift, vma->vm_flags);
237 if (folio_test_anon(folio))
238 hugepage_add_anon_rmap(new, vma, pvmw.address,
241 page_dup_file_rmap(new, true);
242 set_huge_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte);
246 if (folio_test_anon(folio))
247 page_add_anon_rmap(new, vma, pvmw.address,
250 page_add_file_rmap(new, vma, false);
251 set_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte);
253 if (vma->vm_flags & VM_LOCKED)
254 mlock_page_drain_local();
256 trace_remove_migration_pte(pvmw.address, pte_val(pte),
257 compound_order(new));
259 /* No need to invalidate - it was non-present before */
260 update_mmu_cache(vma, pvmw.address, pvmw.pte);
267 * Get rid of all migration entries and replace them by
268 * references to the indicated page.
270 void remove_migration_ptes(struct folio *src, struct folio *dst, bool locked)
272 struct rmap_walk_control rwc = {
273 .rmap_one = remove_migration_pte,
278 rmap_walk_locked(dst, &rwc);
280 rmap_walk(dst, &rwc);
284 * Something used the pte of a page under migration. We need to
285 * get to the page and wait until migration is finished.
286 * When we return from this function the fault will be retried.
288 void __migration_entry_wait(struct mm_struct *mm, pte_t *ptep,
296 if (!is_swap_pte(pte))
299 entry = pte_to_swp_entry(pte);
300 if (!is_migration_entry(entry))
303 migration_entry_wait_on_locked(entry, ptep, ptl);
306 pte_unmap_unlock(ptep, ptl);
309 void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
310 unsigned long address)
312 spinlock_t *ptl = pte_lockptr(mm, pmd);
313 pte_t *ptep = pte_offset_map(pmd, address);
314 __migration_entry_wait(mm, ptep, ptl);
317 void migration_entry_wait_huge(struct vm_area_struct *vma,
318 struct mm_struct *mm, pte_t *pte)
320 spinlock_t *ptl = huge_pte_lockptr(hstate_vma(vma), mm, pte);
321 __migration_entry_wait(mm, pte, ptl);
324 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
325 void pmd_migration_entry_wait(struct mm_struct *mm, pmd_t *pmd)
329 ptl = pmd_lock(mm, pmd);
330 if (!is_pmd_migration_entry(*pmd))
332 migration_entry_wait_on_locked(pmd_to_swp_entry(*pmd), NULL, ptl);
339 static int expected_page_refs(struct address_space *mapping, struct page *page)
341 int expected_count = 1;
344 expected_count += compound_nr(page) + page_has_private(page);
345 return expected_count;
349 * Replace the page in the mapping.
351 * The number of remaining references must be:
352 * 1 for anonymous pages without a mapping
353 * 2 for pages with a mapping
354 * 3 for pages with a mapping and PagePrivate/PagePrivate2 set.
356 int folio_migrate_mapping(struct address_space *mapping,
357 struct folio *newfolio, struct folio *folio, int extra_count)
359 XA_STATE(xas, &mapping->i_pages, folio_index(folio));
360 struct zone *oldzone, *newzone;
362 int expected_count = expected_page_refs(mapping, &folio->page) + extra_count;
363 long nr = folio_nr_pages(folio);
366 /* Anonymous page without mapping */
367 if (folio_ref_count(folio) != expected_count)
370 /* No turning back from here */
371 newfolio->index = folio->index;
372 newfolio->mapping = folio->mapping;
373 if (folio_test_swapbacked(folio))
374 __folio_set_swapbacked(newfolio);
376 return MIGRATEPAGE_SUCCESS;
379 oldzone = folio_zone(folio);
380 newzone = folio_zone(newfolio);
383 if (!folio_ref_freeze(folio, expected_count)) {
384 xas_unlock_irq(&xas);
389 * Now we know that no one else is looking at the folio:
390 * no turning back from here.
392 newfolio->index = folio->index;
393 newfolio->mapping = folio->mapping;
394 folio_ref_add(newfolio, nr); /* add cache reference */
395 if (folio_test_swapbacked(folio)) {
396 __folio_set_swapbacked(newfolio);
397 if (folio_test_swapcache(folio)) {
398 folio_set_swapcache(newfolio);
399 newfolio->private = folio_get_private(folio);
402 VM_BUG_ON_FOLIO(folio_test_swapcache(folio), folio);
405 /* Move dirty while page refs frozen and newpage not yet exposed */
406 dirty = folio_test_dirty(folio);
408 folio_clear_dirty(folio);
409 folio_set_dirty(newfolio);
412 xas_store(&xas, newfolio);
415 * Drop cache reference from old page by unfreezing
416 * to one less reference.
417 * We know this isn't the last reference.
419 folio_ref_unfreeze(folio, expected_count - nr);
422 /* Leave irq disabled to prevent preemption while updating stats */
425 * If moved to a different zone then also account
426 * the page for that zone. Other VM counters will be
427 * taken care of when we establish references to the
428 * new page and drop references to the old page.
430 * Note that anonymous pages are accounted for
431 * via NR_FILE_PAGES and NR_ANON_MAPPED if they
432 * are mapped to swap space.
434 if (newzone != oldzone) {
435 struct lruvec *old_lruvec, *new_lruvec;
436 struct mem_cgroup *memcg;
438 memcg = folio_memcg(folio);
439 old_lruvec = mem_cgroup_lruvec(memcg, oldzone->zone_pgdat);
440 new_lruvec = mem_cgroup_lruvec(memcg, newzone->zone_pgdat);
442 __mod_lruvec_state(old_lruvec, NR_FILE_PAGES, -nr);
443 __mod_lruvec_state(new_lruvec, NR_FILE_PAGES, nr);
444 if (folio_test_swapbacked(folio) && !folio_test_swapcache(folio)) {
445 __mod_lruvec_state(old_lruvec, NR_SHMEM, -nr);
446 __mod_lruvec_state(new_lruvec, NR_SHMEM, nr);
449 if (folio_test_swapcache(folio)) {
450 __mod_lruvec_state(old_lruvec, NR_SWAPCACHE, -nr);
451 __mod_lruvec_state(new_lruvec, NR_SWAPCACHE, nr);
454 if (dirty && mapping_can_writeback(mapping)) {
455 __mod_lruvec_state(old_lruvec, NR_FILE_DIRTY, -nr);
456 __mod_zone_page_state(oldzone, NR_ZONE_WRITE_PENDING, -nr);
457 __mod_lruvec_state(new_lruvec, NR_FILE_DIRTY, nr);
458 __mod_zone_page_state(newzone, NR_ZONE_WRITE_PENDING, nr);
463 return MIGRATEPAGE_SUCCESS;
465 EXPORT_SYMBOL(folio_migrate_mapping);
468 * The expected number of remaining references is the same as that
469 * of folio_migrate_mapping().
471 int migrate_huge_page_move_mapping(struct address_space *mapping,
472 struct page *newpage, struct page *page)
474 XA_STATE(xas, &mapping->i_pages, page_index(page));
478 expected_count = 2 + page_has_private(page);
479 if (!page_ref_freeze(page, expected_count)) {
480 xas_unlock_irq(&xas);
484 newpage->index = page->index;
485 newpage->mapping = page->mapping;
489 xas_store(&xas, newpage);
491 page_ref_unfreeze(page, expected_count - 1);
493 xas_unlock_irq(&xas);
495 return MIGRATEPAGE_SUCCESS;
499 * Copy the flags and some other ancillary information
501 void folio_migrate_flags(struct folio *newfolio, struct folio *folio)
505 if (folio_test_error(folio))
506 folio_set_error(newfolio);
507 if (folio_test_referenced(folio))
508 folio_set_referenced(newfolio);
509 if (folio_test_uptodate(folio))
510 folio_mark_uptodate(newfolio);
511 if (folio_test_clear_active(folio)) {
512 VM_BUG_ON_FOLIO(folio_test_unevictable(folio), folio);
513 folio_set_active(newfolio);
514 } else if (folio_test_clear_unevictable(folio))
515 folio_set_unevictable(newfolio);
516 if (folio_test_workingset(folio))
517 folio_set_workingset(newfolio);
518 if (folio_test_checked(folio))
519 folio_set_checked(newfolio);
521 * PG_anon_exclusive (-> PG_mappedtodisk) is always migrated via
522 * migration entries. We can still have PG_anon_exclusive set on an
523 * effectively unmapped and unreferenced first sub-pages of an
524 * anonymous THP: we can simply copy it here via PG_mappedtodisk.
526 if (folio_test_mappedtodisk(folio))
527 folio_set_mappedtodisk(newfolio);
529 /* Move dirty on pages not done by folio_migrate_mapping() */
530 if (folio_test_dirty(folio))
531 folio_set_dirty(newfolio);
533 if (folio_test_young(folio))
534 folio_set_young(newfolio);
535 if (folio_test_idle(folio))
536 folio_set_idle(newfolio);
539 * Copy NUMA information to the new page, to prevent over-eager
540 * future migrations of this same page.
542 cpupid = page_cpupid_xchg_last(&folio->page, -1);
543 page_cpupid_xchg_last(&newfolio->page, cpupid);
545 folio_migrate_ksm(newfolio, folio);
547 * Please do not reorder this without considering how mm/ksm.c's
548 * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache().
550 if (folio_test_swapcache(folio))
551 folio_clear_swapcache(folio);
552 folio_clear_private(folio);
554 /* page->private contains hugetlb specific flags */
555 if (!folio_test_hugetlb(folio))
556 folio->private = NULL;
559 * If any waiters have accumulated on the new page then
562 if (folio_test_writeback(newfolio))
563 folio_end_writeback(newfolio);
566 * PG_readahead shares the same bit with PG_reclaim. The above
567 * end_page_writeback() may clear PG_readahead mistakenly, so set the
570 if (folio_test_readahead(folio))
571 folio_set_readahead(newfolio);
573 folio_copy_owner(newfolio, folio);
575 if (!folio_test_hugetlb(folio))
576 mem_cgroup_migrate(folio, newfolio);
578 EXPORT_SYMBOL(folio_migrate_flags);
580 void folio_migrate_copy(struct folio *newfolio, struct folio *folio)
582 folio_copy(newfolio, folio);
583 folio_migrate_flags(newfolio, folio);
585 EXPORT_SYMBOL(folio_migrate_copy);
587 /************************************************************
588 * Migration functions
589 ***********************************************************/
592 * Common logic to directly migrate a single LRU page suitable for
593 * pages that do not use PagePrivate/PagePrivate2.
595 * Pages are locked upon entry and exit.
597 int migrate_page(struct address_space *mapping,
598 struct page *newpage, struct page *page,
599 enum migrate_mode mode)
601 struct folio *newfolio = page_folio(newpage);
602 struct folio *folio = page_folio(page);
605 BUG_ON(folio_test_writeback(folio)); /* Writeback must be complete */
607 rc = folio_migrate_mapping(mapping, newfolio, folio, 0);
609 if (rc != MIGRATEPAGE_SUCCESS)
612 if (mode != MIGRATE_SYNC_NO_COPY)
613 folio_migrate_copy(newfolio, folio);
615 folio_migrate_flags(newfolio, folio);
616 return MIGRATEPAGE_SUCCESS;
618 EXPORT_SYMBOL(migrate_page);
621 /* Returns true if all buffers are successfully locked */
622 static bool buffer_migrate_lock_buffers(struct buffer_head *head,
623 enum migrate_mode mode)
625 struct buffer_head *bh = head;
627 /* Simple case, sync compaction */
628 if (mode != MIGRATE_ASYNC) {
631 bh = bh->b_this_page;
633 } while (bh != head);
638 /* async case, we cannot block on lock_buffer so use trylock_buffer */
640 if (!trylock_buffer(bh)) {
642 * We failed to lock the buffer and cannot stall in
643 * async migration. Release the taken locks
645 struct buffer_head *failed_bh = bh;
647 while (bh != failed_bh) {
649 bh = bh->b_this_page;
654 bh = bh->b_this_page;
655 } while (bh != head);
659 static int __buffer_migrate_page(struct address_space *mapping,
660 struct page *newpage, struct page *page, enum migrate_mode mode,
663 struct buffer_head *bh, *head;
667 if (!page_has_buffers(page))
668 return migrate_page(mapping, newpage, page, mode);
670 /* Check whether page does not have extra refs before we do more work */
671 expected_count = expected_page_refs(mapping, page);
672 if (page_count(page) != expected_count)
675 head = page_buffers(page);
676 if (!buffer_migrate_lock_buffers(head, mode))
681 bool invalidated = false;
685 spin_lock(&mapping->private_lock);
688 if (atomic_read(&bh->b_count)) {
692 bh = bh->b_this_page;
693 } while (bh != head);
699 spin_unlock(&mapping->private_lock);
700 invalidate_bh_lrus();
702 goto recheck_buffers;
706 rc = migrate_page_move_mapping(mapping, newpage, page, 0);
707 if (rc != MIGRATEPAGE_SUCCESS)
710 attach_page_private(newpage, detach_page_private(page));
714 set_bh_page(bh, newpage, bh_offset(bh));
715 bh = bh->b_this_page;
717 } while (bh != head);
719 if (mode != MIGRATE_SYNC_NO_COPY)
720 migrate_page_copy(newpage, page);
722 migrate_page_states(newpage, page);
724 rc = MIGRATEPAGE_SUCCESS;
727 spin_unlock(&mapping->private_lock);
731 bh = bh->b_this_page;
733 } while (bh != head);
739 * Migration function for pages with buffers. This function can only be used
740 * if the underlying filesystem guarantees that no other references to "page"
741 * exist. For example attached buffer heads are accessed only under page lock.
743 int buffer_migrate_page(struct address_space *mapping,
744 struct page *newpage, struct page *page, enum migrate_mode mode)
746 return __buffer_migrate_page(mapping, newpage, page, mode, false);
748 EXPORT_SYMBOL(buffer_migrate_page);
751 * Same as above except that this variant is more careful and checks that there
752 * are also no buffer head references. This function is the right one for
753 * mappings where buffer heads are directly looked up and referenced (such as
754 * block device mappings).
756 int buffer_migrate_page_norefs(struct address_space *mapping,
757 struct page *newpage, struct page *page, enum migrate_mode mode)
759 return __buffer_migrate_page(mapping, newpage, page, mode, true);
764 * Writeback a folio to clean the dirty state
766 static int writeout(struct address_space *mapping, struct folio *folio)
768 struct writeback_control wbc = {
769 .sync_mode = WB_SYNC_NONE,
772 .range_end = LLONG_MAX,
777 if (!mapping->a_ops->writepage)
778 /* No write method for the address space */
781 if (!folio_clear_dirty_for_io(folio))
782 /* Someone else already triggered a write */
786 * A dirty folio may imply that the underlying filesystem has
787 * the folio on some queue. So the folio must be clean for
788 * migration. Writeout may mean we lose the lock and the
789 * folio state is no longer what we checked for earlier.
790 * At this point we know that the migration attempt cannot
793 remove_migration_ptes(folio, folio, false);
795 rc = mapping->a_ops->writepage(&folio->page, &wbc);
797 if (rc != AOP_WRITEPAGE_ACTIVATE)
798 /* unlocked. Relock */
801 return (rc < 0) ? -EIO : -EAGAIN;
805 * Default handling if a filesystem does not provide a migration function.
807 static int fallback_migrate_folio(struct address_space *mapping,
808 struct folio *dst, struct folio *src, enum migrate_mode mode)
810 if (folio_test_dirty(src)) {
811 /* Only writeback folios in full synchronous migration */
814 case MIGRATE_SYNC_NO_COPY:
819 return writeout(mapping, src);
823 * Buffers may be managed in a filesystem specific way.
824 * We must have no buffers or drop them.
826 if (folio_test_private(src) &&
827 !filemap_release_folio(src, GFP_KERNEL))
828 return mode == MIGRATE_SYNC ? -EAGAIN : -EBUSY;
830 return migrate_page(mapping, &dst->page, &src->page, mode);
834 * Move a page to a newly allocated page
835 * The page is locked and all ptes have been successfully removed.
837 * The new page will have replaced the old page if this function
842 * MIGRATEPAGE_SUCCESS - success
844 static int move_to_new_folio(struct folio *dst, struct folio *src,
845 enum migrate_mode mode)
848 bool is_lru = !__PageMovable(&src->page);
850 VM_BUG_ON_FOLIO(!folio_test_locked(src), src);
851 VM_BUG_ON_FOLIO(!folio_test_locked(dst), dst);
853 if (likely(is_lru)) {
854 struct address_space *mapping = folio_mapping(src);
857 rc = migrate_page(mapping, &dst->page, &src->page, mode);
858 else if (mapping->a_ops->migrate_folio)
860 * Most folios have a mapping and most filesystems
861 * provide a migrate_folio callback. Anonymous folios
862 * are part of swap space which also has its own
863 * migrate_folio callback. This is the most common path
864 * for page migration.
866 rc = mapping->a_ops->migrate_folio(mapping, dst, src,
868 else if (mapping->a_ops->migratepage)
869 rc = mapping->a_ops->migratepage(mapping, &dst->page,
872 rc = fallback_migrate_folio(mapping, dst, src, mode);
874 const struct movable_operations *mops;
877 * In case of non-lru page, it could be released after
878 * isolation step. In that case, we shouldn't try migration.
880 VM_BUG_ON_FOLIO(!folio_test_isolated(src), src);
881 if (!folio_test_movable(src)) {
882 rc = MIGRATEPAGE_SUCCESS;
883 folio_clear_isolated(src);
887 mops = page_movable_ops(&src->page);
888 rc = mops->migrate_page(&dst->page, &src->page, mode);
889 WARN_ON_ONCE(rc == MIGRATEPAGE_SUCCESS &&
890 !folio_test_isolated(src));
894 * When successful, old pagecache src->mapping must be cleared before
895 * src is freed; but stats require that PageAnon be left as PageAnon.
897 if (rc == MIGRATEPAGE_SUCCESS) {
898 if (__PageMovable(&src->page)) {
899 VM_BUG_ON_FOLIO(!folio_test_isolated(src), src);
902 * We clear PG_movable under page_lock so any compactor
903 * cannot try to migrate this page.
905 folio_clear_isolated(src);
909 * Anonymous and movable src->mapping will be cleared by
910 * free_pages_prepare so don't reset it here for keeping
911 * the type to work PageAnon, for example.
913 if (!folio_mapping_flags(src))
916 if (likely(!folio_is_zone_device(dst)))
917 flush_dcache_folio(dst);
923 static int __unmap_and_move(struct page *page, struct page *newpage,
924 int force, enum migrate_mode mode)
926 struct folio *folio = page_folio(page);
927 struct folio *dst = page_folio(newpage);
929 bool page_was_mapped = false;
930 struct anon_vma *anon_vma = NULL;
931 bool is_lru = !__PageMovable(page);
933 if (!trylock_page(page)) {
934 if (!force || mode == MIGRATE_ASYNC)
938 * It's not safe for direct compaction to call lock_page.
939 * For example, during page readahead pages are added locked
940 * to the LRU. Later, when the IO completes the pages are
941 * marked uptodate and unlocked. However, the queueing
942 * could be merging multiple pages for one bio (e.g.
943 * mpage_readahead). If an allocation happens for the
944 * second or third page, the process can end up locking
945 * the same page twice and deadlocking. Rather than
946 * trying to be clever about what pages can be locked,
947 * avoid the use of lock_page for direct compaction
950 if (current->flags & PF_MEMALLOC)
956 if (PageWriteback(page)) {
958 * Only in the case of a full synchronous migration is it
959 * necessary to wait for PageWriteback. In the async case,
960 * the retry loop is too short and in the sync-light case,
961 * the overhead of stalling is too much
965 case MIGRATE_SYNC_NO_COPY:
973 wait_on_page_writeback(page);
977 * By try_to_migrate(), page->mapcount goes down to 0 here. In this case,
978 * we cannot notice that anon_vma is freed while we migrates a page.
979 * This get_anon_vma() delays freeing anon_vma pointer until the end
980 * of migration. File cache pages are no problem because of page_lock()
981 * File Caches may use write_page() or lock_page() in migration, then,
982 * just care Anon page here.
984 * Only page_get_anon_vma() understands the subtleties of
985 * getting a hold on an anon_vma from outside one of its mms.
986 * But if we cannot get anon_vma, then we won't need it anyway,
987 * because that implies that the anon page is no longer mapped
988 * (and cannot be remapped so long as we hold the page lock).
990 if (PageAnon(page) && !PageKsm(page))
991 anon_vma = page_get_anon_vma(page);
994 * Block others from accessing the new page when we get around to
995 * establishing additional references. We are usually the only one
996 * holding a reference to newpage at this point. We used to have a BUG
997 * here if trylock_page(newpage) fails, but would like to allow for
998 * cases where there might be a race with the previous use of newpage.
999 * This is much like races on refcount of oldpage: just don't BUG().
1001 if (unlikely(!trylock_page(newpage)))
1004 if (unlikely(!is_lru)) {
1005 rc = move_to_new_folio(dst, folio, mode);
1006 goto out_unlock_both;
1010 * Corner case handling:
1011 * 1. When a new swap-cache page is read into, it is added to the LRU
1012 * and treated as swapcache but it has no rmap yet.
1013 * Calling try_to_unmap() against a page->mapping==NULL page will
1014 * trigger a BUG. So handle it here.
1015 * 2. An orphaned page (see truncate_cleanup_page) might have
1016 * fs-private metadata. The page can be picked up due to memory
1017 * offlining. Everywhere else except page reclaim, the page is
1018 * invisible to the vm, so the page can not be migrated. So try to
1019 * free the metadata, so the page can be freed.
1021 if (!page->mapping) {
1022 VM_BUG_ON_PAGE(PageAnon(page), page);
1023 if (page_has_private(page)) {
1024 try_to_free_buffers(folio);
1025 goto out_unlock_both;
1027 } else if (page_mapped(page)) {
1028 /* Establish migration ptes */
1029 VM_BUG_ON_PAGE(PageAnon(page) && !PageKsm(page) && !anon_vma,
1031 try_to_migrate(folio, 0);
1032 page_was_mapped = true;
1035 if (!page_mapped(page))
1036 rc = move_to_new_folio(dst, folio, mode);
1039 * When successful, push newpage to LRU immediately: so that if it
1040 * turns out to be an mlocked page, remove_migration_ptes() will
1041 * automatically build up the correct newpage->mlock_count for it.
1043 * We would like to do something similar for the old page, when
1044 * unsuccessful, and other cases when a page has been temporarily
1045 * isolated from the unevictable LRU: but this case is the easiest.
1047 if (rc == MIGRATEPAGE_SUCCESS) {
1048 lru_cache_add(newpage);
1049 if (page_was_mapped)
1053 if (page_was_mapped)
1054 remove_migration_ptes(folio,
1055 rc == MIGRATEPAGE_SUCCESS ? dst : folio, false);
1058 unlock_page(newpage);
1060 /* Drop an anon_vma reference if we took one */
1062 put_anon_vma(anon_vma);
1066 * If migration is successful, decrease refcount of the newpage,
1067 * which will not free the page because new page owner increased
1070 if (rc == MIGRATEPAGE_SUCCESS)
1077 * Obtain the lock on page, remove all ptes and migrate the page
1078 * to the newly allocated page in newpage.
1080 static int unmap_and_move(new_page_t get_new_page,
1081 free_page_t put_new_page,
1082 unsigned long private, struct page *page,
1083 int force, enum migrate_mode mode,
1084 enum migrate_reason reason,
1085 struct list_head *ret)
1087 int rc = MIGRATEPAGE_SUCCESS;
1088 struct page *newpage = NULL;
1090 if (!thp_migration_supported() && PageTransHuge(page))
1093 if (page_count(page) == 1) {
1094 /* page was freed from under us. So we are done. */
1095 ClearPageActive(page);
1096 ClearPageUnevictable(page);
1097 if (unlikely(__PageMovable(page))) {
1099 if (!PageMovable(page))
1100 ClearPageIsolated(page);
1106 newpage = get_new_page(page, private);
1110 newpage->private = 0;
1111 rc = __unmap_and_move(page, newpage, force, mode);
1112 if (rc == MIGRATEPAGE_SUCCESS)
1113 set_page_owner_migrate_reason(newpage, reason);
1116 if (rc != -EAGAIN) {
1118 * A page that has been migrated has all references
1119 * removed and will be freed. A page that has not been
1120 * migrated will have kept its references and be restored.
1122 list_del(&page->lru);
1126 * If migration is successful, releases reference grabbed during
1127 * isolation. Otherwise, restore the page to right list unless
1130 if (rc == MIGRATEPAGE_SUCCESS) {
1132 * Compaction can migrate also non-LRU pages which are
1133 * not accounted to NR_ISOLATED_*. They can be recognized
1136 if (likely(!__PageMovable(page)))
1137 mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON +
1138 page_is_file_lru(page), -thp_nr_pages(page));
1140 if (reason != MR_MEMORY_FAILURE)
1142 * We release the page in page_handle_poison.
1147 list_add_tail(&page->lru, ret);
1150 put_new_page(newpage, private);
1159 * Counterpart of unmap_and_move_page() for hugepage migration.
1161 * This function doesn't wait the completion of hugepage I/O
1162 * because there is no race between I/O and migration for hugepage.
1163 * Note that currently hugepage I/O occurs only in direct I/O
1164 * where no lock is held and PG_writeback is irrelevant,
1165 * and writeback status of all subpages are counted in the reference
1166 * count of the head page (i.e. if all subpages of a 2MB hugepage are
1167 * under direct I/O, the reference of the head page is 512 and a bit more.)
1168 * This means that when we try to migrate hugepage whose subpages are
1169 * doing direct I/O, some references remain after try_to_unmap() and
1170 * hugepage migration fails without data corruption.
1172 * There is also no race when direct I/O is issued on the page under migration,
1173 * because then pte is replaced with migration swap entry and direct I/O code
1174 * will wait in the page fault for migration to complete.
1176 static int unmap_and_move_huge_page(new_page_t get_new_page,
1177 free_page_t put_new_page, unsigned long private,
1178 struct page *hpage, int force,
1179 enum migrate_mode mode, int reason,
1180 struct list_head *ret)
1182 struct folio *dst, *src = page_folio(hpage);
1184 int page_was_mapped = 0;
1185 struct page *new_hpage;
1186 struct anon_vma *anon_vma = NULL;
1187 struct address_space *mapping = NULL;
1190 * Migratability of hugepages depends on architectures and their size.
1191 * This check is necessary because some callers of hugepage migration
1192 * like soft offline and memory hotremove don't walk through page
1193 * tables or check whether the hugepage is pmd-based or not before
1194 * kicking migration.
1196 if (!hugepage_migration_supported(page_hstate(hpage))) {
1197 list_move_tail(&hpage->lru, ret);
1201 if (page_count(hpage) == 1) {
1202 /* page was freed from under us. So we are done. */
1203 putback_active_hugepage(hpage);
1204 return MIGRATEPAGE_SUCCESS;
1207 new_hpage = get_new_page(hpage, private);
1210 dst = page_folio(new_hpage);
1212 if (!trylock_page(hpage)) {
1217 case MIGRATE_SYNC_NO_COPY:
1226 * Check for pages which are in the process of being freed. Without
1227 * page_mapping() set, hugetlbfs specific move page routine will not
1228 * be called and we could leak usage counts for subpools.
1230 if (hugetlb_page_subpool(hpage) && !page_mapping(hpage)) {
1235 if (PageAnon(hpage))
1236 anon_vma = page_get_anon_vma(hpage);
1238 if (unlikely(!trylock_page(new_hpage)))
1241 if (page_mapped(hpage)) {
1242 enum ttu_flags ttu = 0;
1244 if (!PageAnon(hpage)) {
1246 * In shared mappings, try_to_unmap could potentially
1247 * call huge_pmd_unshare. Because of this, take
1248 * semaphore in write mode here and set TTU_RMAP_LOCKED
1249 * to let lower levels know we have taken the lock.
1251 mapping = hugetlb_page_mapping_lock_write(hpage);
1252 if (unlikely(!mapping))
1253 goto unlock_put_anon;
1255 ttu = TTU_RMAP_LOCKED;
1258 try_to_migrate(src, ttu);
1259 page_was_mapped = 1;
1261 if (ttu & TTU_RMAP_LOCKED)
1262 i_mmap_unlock_write(mapping);
1265 if (!page_mapped(hpage))
1266 rc = move_to_new_folio(dst, src, mode);
1268 if (page_was_mapped)
1269 remove_migration_ptes(src,
1270 rc == MIGRATEPAGE_SUCCESS ? dst : src, false);
1273 unlock_page(new_hpage);
1277 put_anon_vma(anon_vma);
1279 if (rc == MIGRATEPAGE_SUCCESS) {
1280 move_hugetlb_state(hpage, new_hpage, reason);
1281 put_new_page = NULL;
1287 if (rc == MIGRATEPAGE_SUCCESS)
1288 putback_active_hugepage(hpage);
1289 else if (rc != -EAGAIN)
1290 list_move_tail(&hpage->lru, ret);
1293 * If migration was not successful and there's a freeing callback, use
1294 * it. Otherwise, put_page() will drop the reference grabbed during
1298 put_new_page(new_hpage, private);
1300 putback_active_hugepage(new_hpage);
1305 static inline int try_split_thp(struct page *page, struct page **page2,
1306 struct list_head *from)
1311 rc = split_huge_page_to_list(page, from);
1314 list_safe_reset_next(page, *page2, lru);
1320 * migrate_pages - migrate the pages specified in a list, to the free pages
1321 * supplied as the target for the page migration
1323 * @from: The list of pages to be migrated.
1324 * @get_new_page: The function used to allocate free pages to be used
1325 * as the target of the page migration.
1326 * @put_new_page: The function used to free target pages if migration
1327 * fails, or NULL if no special handling is necessary.
1328 * @private: Private data to be passed on to get_new_page()
1329 * @mode: The migration mode that specifies the constraints for
1330 * page migration, if any.
1331 * @reason: The reason for page migration.
1332 * @ret_succeeded: Set to the number of normal pages migrated successfully if
1333 * the caller passes a non-NULL pointer.
1335 * The function returns after 10 attempts or if no pages are movable any more
1336 * because the list has become empty or no retryable pages exist any more.
1337 * It is caller's responsibility to call putback_movable_pages() to return pages
1338 * to the LRU or free list only if ret != 0.
1340 * Returns the number of {normal page, THP, hugetlb} that were not migrated, or
1341 * an error code. The number of THP splits will be considered as the number of
1342 * non-migrated THP, no matter how many subpages of the THP are migrated successfully.
1344 int migrate_pages(struct list_head *from, new_page_t get_new_page,
1345 free_page_t put_new_page, unsigned long private,
1346 enum migrate_mode mode, int reason, unsigned int *ret_succeeded)
1351 int nr_failed_pages = 0;
1352 int nr_succeeded = 0;
1353 int nr_thp_succeeded = 0;
1354 int nr_thp_failed = 0;
1355 int nr_thp_split = 0;
1357 bool is_thp = false;
1360 int rc, nr_subpages;
1361 LIST_HEAD(ret_pages);
1362 LIST_HEAD(thp_split_pages);
1363 bool nosplit = (reason == MR_NUMA_MISPLACED);
1364 bool no_subpage_counting = false;
1366 trace_mm_migrate_pages_start(mode, reason);
1368 thp_subpage_migration:
1369 for (pass = 0; pass < 10 && (retry || thp_retry); pass++) {
1373 list_for_each_entry_safe(page, page2, from, lru) {
1376 * THP statistics is based on the source huge page.
1377 * Capture required information that might get lost
1380 is_thp = PageTransHuge(page) && !PageHuge(page);
1381 nr_subpages = compound_nr(page);
1385 rc = unmap_and_move_huge_page(get_new_page,
1386 put_new_page, private, page,
1387 pass > 2, mode, reason,
1390 rc = unmap_and_move(get_new_page, put_new_page,
1391 private, page, pass > 2, mode,
1392 reason, &ret_pages);
1395 * Success: non hugetlb page will be freed, hugetlb
1396 * page will be put back
1397 * -EAGAIN: stay on the from list
1398 * -ENOMEM: stay on the from list
1399 * Other errno: put on ret_pages list then splice to
1404 * THP migration might be unsupported or the
1405 * allocation could've failed so we should
1406 * retry on the same page with the THP split
1409 * Head page is retried immediately and tail
1410 * pages are added to the tail of the list so
1411 * we encounter them after the rest of the list
1415 /* THP migration is unsupported */
1418 if (!try_split_thp(page, &page2, &thp_split_pages)) {
1422 /* Hugetlb migration is unsupported */
1423 } else if (!no_subpage_counting) {
1427 nr_failed_pages += nr_subpages;
1431 * When memory is low, don't bother to try to migrate
1432 * other pages, just exit.
1433 * THP NUMA faulting doesn't split THP to retry.
1435 if (is_thp && !nosplit) {
1437 if (!try_split_thp(page, &page2, &thp_split_pages)) {
1441 } else if (!no_subpage_counting) {
1445 nr_failed_pages += nr_subpages;
1447 * There might be some subpages of fail-to-migrate THPs
1448 * left in thp_split_pages list. Move them back to migration
1449 * list so that they could be put back to the right list by
1450 * the caller otherwise the page refcnt will be leaked.
1452 list_splice_init(&thp_split_pages, from);
1453 nr_thp_failed += thp_retry;
1461 case MIGRATEPAGE_SUCCESS:
1462 nr_succeeded += nr_subpages;
1468 * Permanent failure (-EBUSY, etc.):
1469 * unlike -EAGAIN case, the failed page is
1470 * removed from migration page list and not
1471 * retried in the next outer loop.
1475 else if (!no_subpage_counting)
1478 nr_failed_pages += nr_subpages;
1484 nr_thp_failed += thp_retry;
1486 * Try to migrate subpages of fail-to-migrate THPs, no nr_failed
1487 * counting in this round, since all subpages of a THP is counted
1488 * as 1 failure in the first round.
1490 if (!list_empty(&thp_split_pages)) {
1492 * Move non-migrated pages (after 10 retries) to ret_pages
1493 * to avoid migrating them again.
1495 list_splice_init(from, &ret_pages);
1496 list_splice_init(&thp_split_pages, from);
1497 no_subpage_counting = true;
1499 goto thp_subpage_migration;
1502 rc = nr_failed + nr_thp_failed;
1505 * Put the permanent failure page back to migration list, they
1506 * will be put back to the right list by the caller.
1508 list_splice(&ret_pages, from);
1510 count_vm_events(PGMIGRATE_SUCCESS, nr_succeeded);
1511 count_vm_events(PGMIGRATE_FAIL, nr_failed_pages);
1512 count_vm_events(THP_MIGRATION_SUCCESS, nr_thp_succeeded);
1513 count_vm_events(THP_MIGRATION_FAIL, nr_thp_failed);
1514 count_vm_events(THP_MIGRATION_SPLIT, nr_thp_split);
1515 trace_mm_migrate_pages(nr_succeeded, nr_failed_pages, nr_thp_succeeded,
1516 nr_thp_failed, nr_thp_split, mode, reason);
1519 *ret_succeeded = nr_succeeded;
1524 struct page *alloc_migration_target(struct page *page, unsigned long private)
1526 struct folio *folio = page_folio(page);
1527 struct migration_target_control *mtc;
1529 unsigned int order = 0;
1530 struct folio *new_folio = NULL;
1534 mtc = (struct migration_target_control *)private;
1535 gfp_mask = mtc->gfp_mask;
1537 if (nid == NUMA_NO_NODE)
1538 nid = folio_nid(folio);
1540 if (folio_test_hugetlb(folio)) {
1541 struct hstate *h = page_hstate(&folio->page);
1543 gfp_mask = htlb_modify_alloc_mask(h, gfp_mask);
1544 return alloc_huge_page_nodemask(h, nid, mtc->nmask, gfp_mask);
1547 if (folio_test_large(folio)) {
1549 * clear __GFP_RECLAIM to make the migration callback
1550 * consistent with regular THP allocations.
1552 gfp_mask &= ~__GFP_RECLAIM;
1553 gfp_mask |= GFP_TRANSHUGE;
1554 order = folio_order(folio);
1556 zidx = zone_idx(folio_zone(folio));
1557 if (is_highmem_idx(zidx) || zidx == ZONE_MOVABLE)
1558 gfp_mask |= __GFP_HIGHMEM;
1560 new_folio = __folio_alloc(gfp_mask, order, nid, mtc->nmask);
1562 return &new_folio->page;
1567 static int store_status(int __user *status, int start, int value, int nr)
1570 if (put_user(value, status + start))
1578 static int do_move_pages_to_node(struct mm_struct *mm,
1579 struct list_head *pagelist, int node)
1582 struct migration_target_control mtc = {
1584 .gfp_mask = GFP_HIGHUSER_MOVABLE | __GFP_THISNODE,
1587 err = migrate_pages(pagelist, alloc_migration_target, NULL,
1588 (unsigned long)&mtc, MIGRATE_SYNC, MR_SYSCALL, NULL);
1590 putback_movable_pages(pagelist);
1595 * Resolves the given address to a struct page, isolates it from the LRU and
1596 * puts it to the given pagelist.
1598 * errno - if the page cannot be found/isolated
1599 * 0 - when it doesn't have to be migrated because it is already on the
1601 * 1 - when it has been queued
1603 static int add_page_for_migration(struct mm_struct *mm, unsigned long addr,
1604 int node, struct list_head *pagelist, bool migrate_all)
1606 struct vm_area_struct *vma;
1612 vma = vma_lookup(mm, addr);
1613 if (!vma || !vma_migratable(vma))
1616 /* FOLL_DUMP to ignore special (like zero) pages */
1617 page = follow_page(vma, addr, FOLL_GET | FOLL_DUMP);
1619 err = PTR_ERR(page);
1628 if (page_to_nid(page) == node)
1632 if (page_mapcount(page) > 1 && !migrate_all)
1635 if (PageHuge(page)) {
1636 if (PageHead(page)) {
1637 isolate_huge_page(page, pagelist);
1643 head = compound_head(page);
1644 err = isolate_lru_page(head);
1649 list_add_tail(&head->lru, pagelist);
1650 mod_node_page_state(page_pgdat(head),
1651 NR_ISOLATED_ANON + page_is_file_lru(head),
1652 thp_nr_pages(head));
1656 * Either remove the duplicate refcount from
1657 * isolate_lru_page() or drop the page ref if it was
1662 mmap_read_unlock(mm);
1666 static int move_pages_and_store_status(struct mm_struct *mm, int node,
1667 struct list_head *pagelist, int __user *status,
1668 int start, int i, unsigned long nr_pages)
1672 if (list_empty(pagelist))
1675 err = do_move_pages_to_node(mm, pagelist, node);
1678 * Positive err means the number of failed
1679 * pages to migrate. Since we are going to
1680 * abort and return the number of non-migrated
1681 * pages, so need to include the rest of the
1682 * nr_pages that have not been attempted as
1686 err += nr_pages - i - 1;
1689 return store_status(status, start, node, i - start);
1693 * Migrate an array of page address onto an array of nodes and fill
1694 * the corresponding array of status.
1696 static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes,
1697 unsigned long nr_pages,
1698 const void __user * __user *pages,
1699 const int __user *nodes,
1700 int __user *status, int flags)
1702 int current_node = NUMA_NO_NODE;
1703 LIST_HEAD(pagelist);
1707 lru_cache_disable();
1709 for (i = start = 0; i < nr_pages; i++) {
1710 const void __user *p;
1715 if (get_user(p, pages + i))
1717 if (get_user(node, nodes + i))
1719 addr = (unsigned long)untagged_addr(p);
1722 if (node < 0 || node >= MAX_NUMNODES)
1724 if (!node_state(node, N_MEMORY))
1728 if (!node_isset(node, task_nodes))
1731 if (current_node == NUMA_NO_NODE) {
1732 current_node = node;
1734 } else if (node != current_node) {
1735 err = move_pages_and_store_status(mm, current_node,
1736 &pagelist, status, start, i, nr_pages);
1740 current_node = node;
1744 * Errors in the page lookup or isolation are not fatal and we simply
1745 * report them via status
1747 err = add_page_for_migration(mm, addr, current_node,
1748 &pagelist, flags & MPOL_MF_MOVE_ALL);
1751 /* The page is successfully queued for migration */
1756 * The move_pages() man page does not have an -EEXIST choice, so
1757 * use -EFAULT instead.
1763 * If the page is already on the target node (!err), store the
1764 * node, otherwise, store the err.
1766 err = store_status(status, i, err ? : current_node, 1);
1770 err = move_pages_and_store_status(mm, current_node, &pagelist,
1771 status, start, i, nr_pages);
1774 current_node = NUMA_NO_NODE;
1777 /* Make sure we do not overwrite the existing error */
1778 err1 = move_pages_and_store_status(mm, current_node, &pagelist,
1779 status, start, i, nr_pages);
1788 * Determine the nodes of an array of pages and store it in an array of status.
1790 static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages,
1791 const void __user **pages, int *status)
1797 for (i = 0; i < nr_pages; i++) {
1798 unsigned long addr = (unsigned long)(*pages);
1799 struct vm_area_struct *vma;
1803 vma = vma_lookup(mm, addr);
1807 /* FOLL_DUMP to ignore special (like zero) pages */
1808 page = follow_page(vma, addr, FOLL_GET | FOLL_DUMP);
1810 err = PTR_ERR(page);
1815 err = page_to_nid(page);
1827 mmap_read_unlock(mm);
1830 static int get_compat_pages_array(const void __user *chunk_pages[],
1831 const void __user * __user *pages,
1832 unsigned long chunk_nr)
1834 compat_uptr_t __user *pages32 = (compat_uptr_t __user *)pages;
1838 for (i = 0; i < chunk_nr; i++) {
1839 if (get_user(p, pages32 + i))
1841 chunk_pages[i] = compat_ptr(p);
1848 * Determine the nodes of a user array of pages and store it in
1849 * a user array of status.
1851 static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages,
1852 const void __user * __user *pages,
1855 #define DO_PAGES_STAT_CHUNK_NR 16UL
1856 const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR];
1857 int chunk_status[DO_PAGES_STAT_CHUNK_NR];
1860 unsigned long chunk_nr = min(nr_pages, DO_PAGES_STAT_CHUNK_NR);
1862 if (in_compat_syscall()) {
1863 if (get_compat_pages_array(chunk_pages, pages,
1867 if (copy_from_user(chunk_pages, pages,
1868 chunk_nr * sizeof(*chunk_pages)))
1872 do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status);
1874 if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status)))
1879 nr_pages -= chunk_nr;
1881 return nr_pages ? -EFAULT : 0;
1884 static struct mm_struct *find_mm_struct(pid_t pid, nodemask_t *mem_nodes)
1886 struct task_struct *task;
1887 struct mm_struct *mm;
1890 * There is no need to check if current process has the right to modify
1891 * the specified process when they are same.
1895 *mem_nodes = cpuset_mems_allowed(current);
1899 /* Find the mm_struct */
1901 task = find_task_by_vpid(pid);
1904 return ERR_PTR(-ESRCH);
1906 get_task_struct(task);
1909 * Check if this process has the right to modify the specified
1910 * process. Use the regular "ptrace_may_access()" checks.
1912 if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) {
1914 mm = ERR_PTR(-EPERM);
1919 mm = ERR_PTR(security_task_movememory(task));
1922 *mem_nodes = cpuset_mems_allowed(task);
1923 mm = get_task_mm(task);
1925 put_task_struct(task);
1927 mm = ERR_PTR(-EINVAL);
1932 * Move a list of pages in the address space of the currently executing
1935 static int kernel_move_pages(pid_t pid, unsigned long nr_pages,
1936 const void __user * __user *pages,
1937 const int __user *nodes,
1938 int __user *status, int flags)
1940 struct mm_struct *mm;
1942 nodemask_t task_nodes;
1945 if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
1948 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1951 mm = find_mm_struct(pid, &task_nodes);
1956 err = do_pages_move(mm, task_nodes, nr_pages, pages,
1957 nodes, status, flags);
1959 err = do_pages_stat(mm, nr_pages, pages, status);
1965 SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages,
1966 const void __user * __user *, pages,
1967 const int __user *, nodes,
1968 int __user *, status, int, flags)
1970 return kernel_move_pages(pid, nr_pages, pages, nodes, status, flags);
1973 #ifdef CONFIG_NUMA_BALANCING
1975 * Returns true if this is a safe migration target node for misplaced NUMA
1976 * pages. Currently it only checks the watermarks which is crude.
1978 static bool migrate_balanced_pgdat(struct pglist_data *pgdat,
1979 unsigned long nr_migrate_pages)
1983 for (z = pgdat->nr_zones - 1; z >= 0; z--) {
1984 struct zone *zone = pgdat->node_zones + z;
1986 if (!managed_zone(zone))
1989 /* Avoid waking kswapd by allocating pages_to_migrate pages. */
1990 if (!zone_watermark_ok(zone, 0,
1991 high_wmark_pages(zone) +
2000 static struct page *alloc_misplaced_dst_page(struct page *page,
2003 int nid = (int) data;
2004 int order = compound_order(page);
2005 gfp_t gfp = __GFP_THISNODE;
2009 gfp |= GFP_TRANSHUGE_LIGHT;
2011 gfp |= GFP_HIGHUSER_MOVABLE | __GFP_NOMEMALLOC | __GFP_NORETRY |
2013 gfp &= ~__GFP_RECLAIM;
2015 new = __folio_alloc_node(gfp, order, nid);
2020 static int numamigrate_isolate_page(pg_data_t *pgdat, struct page *page)
2022 int nr_pages = thp_nr_pages(page);
2023 int order = compound_order(page);
2025 VM_BUG_ON_PAGE(order && !PageTransHuge(page), page);
2027 /* Do not migrate THP mapped by multiple processes */
2028 if (PageTransHuge(page) && total_mapcount(page) > 1)
2031 /* Avoid migrating to a node that is nearly full */
2032 if (!migrate_balanced_pgdat(pgdat, nr_pages)) {
2035 if (!(sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING))
2037 for (z = pgdat->nr_zones - 1; z >= 0; z--) {
2038 if (managed_zone(pgdat->node_zones + z))
2041 wakeup_kswapd(pgdat->node_zones + z, 0, order, ZONE_MOVABLE);
2045 if (isolate_lru_page(page))
2048 mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON + page_is_file_lru(page),
2052 * Isolating the page has taken another reference, so the
2053 * caller's reference can be safely dropped without the page
2054 * disappearing underneath us during migration.
2061 * Attempt to migrate a misplaced page to the specified destination
2062 * node. Caller is expected to have an elevated reference count on
2063 * the page that will be dropped by this function before returning.
2065 int migrate_misplaced_page(struct page *page, struct vm_area_struct *vma,
2068 pg_data_t *pgdat = NODE_DATA(node);
2071 unsigned int nr_succeeded;
2072 LIST_HEAD(migratepages);
2073 int nr_pages = thp_nr_pages(page);
2076 * Don't migrate file pages that are mapped in multiple processes
2077 * with execute permissions as they are probably shared libraries.
2079 if (page_mapcount(page) != 1 && page_is_file_lru(page) &&
2080 (vma->vm_flags & VM_EXEC))
2084 * Also do not migrate dirty pages as not all filesystems can move
2085 * dirty pages in MIGRATE_ASYNC mode which is a waste of cycles.
2087 if (page_is_file_lru(page) && PageDirty(page))
2090 isolated = numamigrate_isolate_page(pgdat, page);
2094 list_add(&page->lru, &migratepages);
2095 nr_remaining = migrate_pages(&migratepages, alloc_misplaced_dst_page,
2096 NULL, node, MIGRATE_ASYNC,
2097 MR_NUMA_MISPLACED, &nr_succeeded);
2099 if (!list_empty(&migratepages)) {
2100 list_del(&page->lru);
2101 mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON +
2102 page_is_file_lru(page), -nr_pages);
2103 putback_lru_page(page);
2108 count_vm_numa_events(NUMA_PAGE_MIGRATE, nr_succeeded);
2109 if (!node_is_toptier(page_to_nid(page)) && node_is_toptier(node))
2110 mod_node_page_state(pgdat, PGPROMOTE_SUCCESS,
2113 BUG_ON(!list_empty(&migratepages));
2120 #endif /* CONFIG_NUMA_BALANCING */
2123 * node_demotion[] example:
2125 * Consider a system with two sockets. Each socket has
2126 * three classes of memory attached: fast, medium and slow.
2127 * Each memory class is placed in its own NUMA node. The
2128 * CPUs are placed in the node with the "fast" memory. The
2129 * 6 NUMA nodes (0-5) might be split among the sockets like
2135 * When Node 0 fills up, its memory should be migrated to
2136 * Node 1. When Node 1 fills up, it should be migrated to
2137 * Node 2. The migration path start on the nodes with the
2138 * processors (since allocations default to this node) and
2139 * fast memory, progress through medium and end with the
2142 * 0 -> 1 -> 2 -> stop
2143 * 3 -> 4 -> 5 -> stop
2145 * This is represented in the node_demotion[] like this:
2147 * { nr=1, nodes[0]=1 }, // Node 0 migrates to 1
2148 * { nr=1, nodes[0]=2 }, // Node 1 migrates to 2
2149 * { nr=0, nodes[0]=-1 }, // Node 2 does not migrate
2150 * { nr=1, nodes[0]=4 }, // Node 3 migrates to 4
2151 * { nr=1, nodes[0]=5 }, // Node 4 migrates to 5
2152 * { nr=0, nodes[0]=-1 }, // Node 5 does not migrate
2154 * Moreover some systems may have multiple slow memory nodes.
2155 * Suppose a system has one socket with 3 memory nodes, node 0
2156 * is fast memory type, and node 1/2 both are slow memory
2157 * type, and the distance between fast memory node and slow
2158 * memory node is same. So the migration path should be:
2162 * This is represented in the node_demotion[] like this:
2163 * { nr=2, {nodes[0]=1, nodes[1]=2} }, // Node 0 migrates to node 1 and node 2
2164 * { nr=0, nodes[0]=-1, }, // Node 1 dose not migrate
2165 * { nr=0, nodes[0]=-1, }, // Node 2 does not migrate
2169 * Writes to this array occur without locking. Cycles are
2170 * not allowed: Node X demotes to Y which demotes to X...
2172 * If multiple reads are performed, a single rcu_read_lock()
2173 * must be held over all reads to ensure that no cycles are
2176 #define DEFAULT_DEMOTION_TARGET_NODES 15
2178 #if MAX_NUMNODES < DEFAULT_DEMOTION_TARGET_NODES
2179 #define DEMOTION_TARGET_NODES (MAX_NUMNODES - 1)
2181 #define DEMOTION_TARGET_NODES DEFAULT_DEMOTION_TARGET_NODES
2184 struct demotion_nodes {
2186 short nodes[DEMOTION_TARGET_NODES];
2189 static struct demotion_nodes *node_demotion __read_mostly;
2192 * next_demotion_node() - Get the next node in the demotion path
2193 * @node: The starting node to lookup the next node
2195 * Return: node id for next memory node in the demotion path hierarchy
2196 * from @node; NUMA_NO_NODE if @node is terminal. This does not keep
2197 * @node online or guarantee that it *continues* to be the next demotion
2200 int next_demotion_node(int node)
2202 struct demotion_nodes *nd;
2203 unsigned short target_nr, index;
2207 return NUMA_NO_NODE;
2209 nd = &node_demotion[node];
2212 * node_demotion[] is updated without excluding this
2213 * function from running. RCU doesn't provide any
2214 * compiler barriers, so the READ_ONCE() is required
2215 * to avoid compiler reordering or read merging.
2217 * Make sure to use RCU over entire code blocks if
2218 * node_demotion[] reads need to be consistent.
2221 target_nr = READ_ONCE(nd->nr);
2223 switch (target_nr) {
2225 target = NUMA_NO_NODE;
2232 * If there are multiple target nodes, just select one
2233 * target node randomly.
2235 * In addition, we can also use round-robin to select
2236 * target node, but we should introduce another variable
2237 * for node_demotion[] to record last selected target node,
2238 * that may cause cache ping-pong due to the changing of
2239 * last target node. Or introducing per-cpu data to avoid
2240 * caching issue, which seems more complicated. So selecting
2241 * target node randomly seems better until now.
2243 index = get_random_int() % target_nr;
2247 target = READ_ONCE(nd->nodes[index]);
2254 /* Disable reclaim-based migration. */
2255 static void __disable_all_migrate_targets(void)
2262 for_each_online_node(node) {
2263 node_demotion[node].nr = 0;
2264 for (i = 0; i < DEMOTION_TARGET_NODES; i++)
2265 node_demotion[node].nodes[i] = NUMA_NO_NODE;
2269 static void disable_all_migrate_targets(void)
2271 __disable_all_migrate_targets();
2274 * Ensure that the "disable" is visible across the system.
2275 * Readers will see either a combination of before+disable
2276 * state or disable+after. They will never see before and
2277 * after state together.
2279 * The before+after state together might have cycles and
2280 * could cause readers to do things like loop until this
2281 * function finishes. This ensures they can only see a
2282 * single "bad" read and would, for instance, only loop
2289 * Find an automatic demotion target for 'node'.
2290 * Failing here is OK. It might just indicate
2291 * being at the end of a chain.
2293 static int establish_migrate_target(int node, nodemask_t *used,
2296 int migration_target, index, val;
2297 struct demotion_nodes *nd;
2300 return NUMA_NO_NODE;
2302 nd = &node_demotion[node];
2304 migration_target = find_next_best_node(node, used);
2305 if (migration_target == NUMA_NO_NODE)
2306 return NUMA_NO_NODE;
2309 * If the node has been set a migration target node before,
2310 * which means it's the best distance between them. Still
2311 * check if this node can be demoted to other target nodes
2312 * if they have a same best distance.
2314 if (best_distance != -1) {
2315 val = node_distance(node, migration_target);
2316 if (val > best_distance)
2321 if (WARN_ONCE(index >= DEMOTION_TARGET_NODES,
2322 "Exceeds maximum demotion target nodes\n"))
2325 nd->nodes[index] = migration_target;
2328 return migration_target;
2330 node_clear(migration_target, *used);
2331 return NUMA_NO_NODE;
2335 * When memory fills up on a node, memory contents can be
2336 * automatically migrated to another node instead of
2337 * discarded at reclaim.
2339 * Establish a "migration path" which will start at nodes
2340 * with CPUs and will follow the priorities used to build the
2341 * page allocator zonelists.
2343 * The difference here is that cycles must be avoided. If
2344 * node0 migrates to node1, then neither node1, nor anything
2345 * node1 migrates to can migrate to node0. Also one node can
2346 * be migrated to multiple nodes if the target nodes all have
2347 * a same best-distance against the source node.
2349 * This function can run simultaneously with readers of
2350 * node_demotion[]. However, it can not run simultaneously
2351 * with itself. Exclusion is provided by memory hotplug events
2352 * being single-threaded.
2354 static void __set_migration_target_nodes(void)
2356 nodemask_t next_pass;
2357 nodemask_t this_pass;
2358 nodemask_t used_targets = NODE_MASK_NONE;
2359 int node, best_distance;
2362 * Avoid any oddities like cycles that could occur
2363 * from changes in the topology. This will leave
2364 * a momentary gap when migration is disabled.
2366 disable_all_migrate_targets();
2369 * Allocations go close to CPUs, first. Assume that
2370 * the migration path starts at the nodes with CPUs.
2372 next_pass = node_states[N_CPU];
2374 this_pass = next_pass;
2375 next_pass = NODE_MASK_NONE;
2377 * To avoid cycles in the migration "graph", ensure
2378 * that migration sources are not future targets by
2379 * setting them in 'used_targets'. Do this only
2380 * once per pass so that multiple source nodes can
2381 * share a target node.
2383 * 'used_targets' will become unavailable in future
2384 * passes. This limits some opportunities for
2385 * multiple source nodes to share a destination.
2387 nodes_or(used_targets, used_targets, this_pass);
2389 for_each_node_mask(node, this_pass) {
2393 * Try to set up the migration path for the node, and the target
2394 * migration nodes can be multiple, so doing a loop to find all
2395 * the target nodes if they all have a best node distance.
2399 establish_migrate_target(node, &used_targets,
2402 if (target_node == NUMA_NO_NODE)
2405 if (best_distance == -1)
2406 best_distance = node_distance(node, target_node);
2409 * Visit targets from this pass in the next pass.
2410 * Eventually, every node will have been part of
2411 * a pass, and will become set in 'used_targets'.
2413 node_set(target_node, next_pass);
2417 * 'next_pass' contains nodes which became migration
2418 * targets in this pass. Make additional passes until
2419 * no more migrations targets are available.
2421 if (!nodes_empty(next_pass))
2426 * For callers that do not hold get_online_mems() already.
2428 void set_migration_target_nodes(void)
2431 __set_migration_target_nodes();
2436 * This leaves migrate-on-reclaim transiently disabled between
2437 * the MEM_GOING_OFFLINE and MEM_OFFLINE events. This runs
2438 * whether reclaim-based migration is enabled or not, which
2439 * ensures that the user can turn reclaim-based migration at
2440 * any time without needing to recalculate migration targets.
2442 * These callbacks already hold get_online_mems(). That is why
2443 * __set_migration_target_nodes() can be used as opposed to
2444 * set_migration_target_nodes().
2446 #ifdef CONFIG_MEMORY_HOTPLUG
2447 static int __meminit migrate_on_reclaim_callback(struct notifier_block *self,
2448 unsigned long action, void *_arg)
2450 struct memory_notify *arg = _arg;
2453 * Only update the node migration order when a node is
2454 * changing status, like online->offline. This avoids
2455 * the overhead of synchronize_rcu() in most cases.
2457 if (arg->status_change_nid < 0)
2458 return notifier_from_errno(0);
2461 case MEM_GOING_OFFLINE:
2463 * Make sure there are not transient states where
2464 * an offline node is a migration target. This
2465 * will leave migration disabled until the offline
2466 * completes and the MEM_OFFLINE case below runs.
2468 disable_all_migrate_targets();
2473 * Recalculate the target nodes once the node
2474 * reaches its final state (online or offline).
2476 __set_migration_target_nodes();
2478 case MEM_CANCEL_OFFLINE:
2480 * MEM_GOING_OFFLINE disabled all the migration
2481 * targets. Reenable them.
2483 __set_migration_target_nodes();
2485 case MEM_GOING_ONLINE:
2486 case MEM_CANCEL_ONLINE:
2490 return notifier_from_errno(0);
2494 void __init migrate_on_reclaim_init(void)
2496 node_demotion = kcalloc(nr_node_ids,
2497 sizeof(struct demotion_nodes),
2499 WARN_ON(!node_demotion);
2500 #ifdef CONFIG_MEMORY_HOTPLUG
2501 hotplug_memory_notifier(migrate_on_reclaim_callback, 100);
2504 * At this point, all numa nodes with memory/CPus have their state
2505 * properly set, so we can build the demotion order now.
2506 * Let us hold the cpu_hotplug lock just, as we could possibily have
2507 * CPU hotplug events during boot.
2510 set_migration_target_nodes();
2514 bool numa_demotion_enabled = false;
2517 static ssize_t numa_demotion_enabled_show(struct kobject *kobj,
2518 struct kobj_attribute *attr, char *buf)
2520 return sysfs_emit(buf, "%s\n",
2521 numa_demotion_enabled ? "true" : "false");
2524 static ssize_t numa_demotion_enabled_store(struct kobject *kobj,
2525 struct kobj_attribute *attr,
2526 const char *buf, size_t count)
2530 ret = kstrtobool(buf, &numa_demotion_enabled);
2537 static struct kobj_attribute numa_demotion_enabled_attr =
2538 __ATTR(demotion_enabled, 0644, numa_demotion_enabled_show,
2539 numa_demotion_enabled_store);
2541 static struct attribute *numa_attrs[] = {
2542 &numa_demotion_enabled_attr.attr,
2546 static const struct attribute_group numa_attr_group = {
2547 .attrs = numa_attrs,
2550 static int __init numa_init_sysfs(void)
2553 struct kobject *numa_kobj;
2555 numa_kobj = kobject_create_and_add("numa", mm_kobj);
2557 pr_err("failed to create numa kobject\n");
2560 err = sysfs_create_group(numa_kobj, &numa_attr_group);
2562 pr_err("failed to register numa group\n");
2568 kobject_put(numa_kobj);
2571 subsys_initcall(numa_init_sysfs);
2572 #endif /* CONFIG_SYSFS */
2573 #endif /* CONFIG_NUMA */