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/ksm.h>
25 #include <linux/rmap.h>
26 #include <linux/topology.h>
27 #include <linux/cpu.h>
28 #include <linux/cpuset.h>
29 #include <linux/writeback.h>
30 #include <linux/mempolicy.h>
31 #include <linux/vmalloc.h>
32 #include <linux/security.h>
33 #include <linux/backing-dev.h>
34 #include <linux/compaction.h>
35 #include <linux/syscalls.h>
36 #include <linux/compat.h>
37 #include <linux/hugetlb.h>
38 #include <linux/hugetlb_cgroup.h>
39 #include <linux/gfp.h>
40 #include <linux/pfn_t.h>
41 #include <linux/memremap.h>
42 #include <linux/userfaultfd_k.h>
43 #include <linux/balloon_compaction.h>
44 #include <linux/page_idle.h>
45 #include <linux/page_owner.h>
46 #include <linux/sched/mm.h>
47 #include <linux/ptrace.h>
48 #include <linux/oom.h>
49 #include <linux/memory.h>
50 #include <linux/random.h>
51 #include <linux/sched/sysctl.h>
52 #include <linux/memory-tiers.h>
54 #include <asm/tlbflush.h>
56 #include <trace/events/migrate.h>
60 bool isolate_movable_page(struct page *page, isolate_mode_t mode)
62 struct folio *folio = folio_get_nontail_page(page);
63 const struct movable_operations *mops;
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.
77 if (unlikely(folio_test_slab(folio)))
79 /* Pairs with smp_wmb() in slab freeing, e.g. SLUB's __free_slab() */
82 * Check movable flag before taking the page lock because
83 * we use non-atomic bitops on newly allocated page flags so
84 * unconditionally grabbing the lock ruins page's owner side.
86 if (unlikely(!__folio_test_movable(folio)))
88 /* Pairs with smp_wmb() in slab allocation, e.g. SLUB's alloc_slab_page() */
90 if (unlikely(folio_test_slab(folio)))
94 * As movable pages are not isolated from LRU lists, concurrent
95 * compaction threads can race against page migration functions
96 * as well as race against the releasing a page.
98 * In order to avoid having an already isolated movable page
99 * being (wrongly) re-isolated while it is under migration,
100 * or to avoid attempting to isolate pages being released,
101 * lets be sure we have the page lock
102 * before proceeding with the movable page isolation steps.
104 if (unlikely(!folio_trylock(folio)))
107 if (!folio_test_movable(folio) || folio_test_isolated(folio))
108 goto out_no_isolated;
110 mops = folio_movable_ops(folio);
111 VM_BUG_ON_FOLIO(!mops, folio);
113 if (!mops->isolate_page(&folio->page, mode))
114 goto out_no_isolated;
116 /* Driver shouldn't use PG_isolated bit of page->flags */
117 WARN_ON_ONCE(folio_test_isolated(folio));
118 folio_set_isolated(folio);
131 static void putback_movable_folio(struct folio *folio)
133 const struct movable_operations *mops = folio_movable_ops(folio);
135 mops->putback_page(&folio->page);
136 folio_clear_isolated(folio);
140 * Put previously isolated pages back onto the appropriate lists
141 * from where they were once taken off for compaction/migration.
143 * This function shall be used whenever the isolated pageset has been
144 * built from lru, balloon, hugetlbfs page. See isolate_migratepages_range()
145 * and isolate_hugetlb().
147 void putback_movable_pages(struct list_head *l)
150 struct folio *folio2;
152 list_for_each_entry_safe(folio, folio2, l, lru) {
153 if (unlikely(folio_test_hugetlb(folio))) {
154 folio_putback_active_hugetlb(folio);
157 list_del(&folio->lru);
159 * We isolated non-lru movable folio so here we can use
160 * __PageMovable because LRU folio's mapping cannot have
161 * PAGE_MAPPING_MOVABLE.
163 if (unlikely(__folio_test_movable(folio))) {
164 VM_BUG_ON_FOLIO(!folio_test_isolated(folio), folio);
166 if (folio_test_movable(folio))
167 putback_movable_folio(folio);
169 folio_clear_isolated(folio);
173 node_stat_mod_folio(folio, NR_ISOLATED_ANON +
174 folio_is_file_lru(folio), -folio_nr_pages(folio));
175 folio_putback_lru(folio);
181 * Restore a potential migration pte to a working pte entry
183 static bool remove_migration_pte(struct folio *folio,
184 struct vm_area_struct *vma, unsigned long addr, void *old)
186 DEFINE_FOLIO_VMA_WALK(pvmw, old, vma, addr, PVMW_SYNC | PVMW_MIGRATION);
188 while (page_vma_mapped_walk(&pvmw)) {
189 rmap_t rmap_flags = RMAP_NONE;
194 unsigned long idx = 0;
196 /* pgoff is invalid for ksm pages, but they are never large */
197 if (folio_test_large(folio) && !folio_test_hugetlb(folio))
198 idx = linear_page_index(vma, pvmw.address) - pvmw.pgoff;
199 new = folio_page(folio, idx);
201 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
202 /* PMD-mapped THP migration entry */
204 VM_BUG_ON_FOLIO(folio_test_hugetlb(folio) ||
205 !folio_test_pmd_mappable(folio), folio);
206 remove_migration_pmd(&pvmw, new);
212 pte = mk_pte(new, READ_ONCE(vma->vm_page_prot));
213 old_pte = ptep_get(pvmw.pte);
214 if (pte_swp_soft_dirty(old_pte))
215 pte = pte_mksoft_dirty(pte);
217 entry = pte_to_swp_entry(old_pte);
218 if (!is_migration_entry_young(entry))
219 pte = pte_mkold(pte);
220 if (folio_test_dirty(folio) && is_migration_entry_dirty(entry))
221 pte = pte_mkdirty(pte);
222 if (is_writable_migration_entry(entry))
223 pte = pte_mkwrite(pte, vma);
224 else if (pte_swp_uffd_wp(old_pte))
225 pte = pte_mkuffd_wp(pte);
227 if (folio_test_anon(folio) && !is_readable_migration_entry(entry))
228 rmap_flags |= RMAP_EXCLUSIVE;
230 if (unlikely(is_device_private_page(new))) {
232 entry = make_writable_device_private_entry(
235 entry = make_readable_device_private_entry(
237 pte = swp_entry_to_pte(entry);
238 if (pte_swp_soft_dirty(old_pte))
239 pte = pte_swp_mksoft_dirty(pte);
240 if (pte_swp_uffd_wp(old_pte))
241 pte = pte_swp_mkuffd_wp(pte);
244 #ifdef CONFIG_HUGETLB_PAGE
245 if (folio_test_hugetlb(folio)) {
246 struct hstate *h = hstate_vma(vma);
247 unsigned int shift = huge_page_shift(h);
248 unsigned long psize = huge_page_size(h);
250 pte = arch_make_huge_pte(pte, shift, vma->vm_flags);
251 if (folio_test_anon(folio))
252 hugepage_add_anon_rmap(new, vma, pvmw.address,
255 page_dup_file_rmap(new, true);
256 set_huge_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte,
261 if (folio_test_anon(folio))
262 page_add_anon_rmap(new, vma, pvmw.address,
265 page_add_file_rmap(new, vma, false);
266 set_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte);
268 if (vma->vm_flags & VM_LOCKED)
271 trace_remove_migration_pte(pvmw.address, pte_val(pte),
272 compound_order(new));
274 /* No need to invalidate - it was non-present before */
275 update_mmu_cache(vma, pvmw.address, pvmw.pte);
282 * Get rid of all migration entries and replace them by
283 * references to the indicated page.
285 void remove_migration_ptes(struct folio *src, struct folio *dst, bool locked)
287 struct rmap_walk_control rwc = {
288 .rmap_one = remove_migration_pte,
293 rmap_walk_locked(dst, &rwc);
295 rmap_walk(dst, &rwc);
299 * Something used the pte of a page under migration. We need to
300 * get to the page and wait until migration is finished.
301 * When we return from this function the fault will be retried.
303 void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
304 unsigned long address)
311 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
315 pte = ptep_get(ptep);
318 if (!is_swap_pte(pte))
321 entry = pte_to_swp_entry(pte);
322 if (!is_migration_entry(entry))
325 migration_entry_wait_on_locked(entry, ptl);
331 #ifdef CONFIG_HUGETLB_PAGE
333 * The vma read lock must be held upon entry. Holding that lock prevents either
334 * the pte or the ptl from being freed.
336 * This function will release the vma lock before returning.
338 void migration_entry_wait_huge(struct vm_area_struct *vma, pte_t *ptep)
340 spinlock_t *ptl = huge_pte_lockptr(hstate_vma(vma), vma->vm_mm, ptep);
343 hugetlb_vma_assert_locked(vma);
345 pte = huge_ptep_get(ptep);
347 if (unlikely(!is_hugetlb_entry_migration(pte))) {
349 hugetlb_vma_unlock_read(vma);
352 * If migration entry existed, safe to release vma lock
353 * here because the pgtable page won't be freed without the
354 * pgtable lock released. See comment right above pgtable
355 * lock release in migration_entry_wait_on_locked().
357 hugetlb_vma_unlock_read(vma);
358 migration_entry_wait_on_locked(pte_to_swp_entry(pte), ptl);
363 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
364 void pmd_migration_entry_wait(struct mm_struct *mm, pmd_t *pmd)
368 ptl = pmd_lock(mm, pmd);
369 if (!is_pmd_migration_entry(*pmd))
371 migration_entry_wait_on_locked(pmd_to_swp_entry(*pmd), ptl);
378 static int folio_expected_refs(struct address_space *mapping,
385 refs += folio_nr_pages(folio);
386 if (folio_test_private(folio))
393 * Replace the page in the mapping.
395 * The number of remaining references must be:
396 * 1 for anonymous pages without a mapping
397 * 2 for pages with a mapping
398 * 3 for pages with a mapping and PagePrivate/PagePrivate2 set.
400 int folio_migrate_mapping(struct address_space *mapping,
401 struct folio *newfolio, struct folio *folio, int extra_count)
403 XA_STATE(xas, &mapping->i_pages, folio_index(folio));
404 struct zone *oldzone, *newzone;
406 int expected_count = folio_expected_refs(mapping, folio) + extra_count;
407 long nr = folio_nr_pages(folio);
411 /* Anonymous page without mapping */
412 if (folio_ref_count(folio) != expected_count)
415 /* No turning back from here */
416 newfolio->index = folio->index;
417 newfolio->mapping = folio->mapping;
418 if (folio_test_swapbacked(folio))
419 __folio_set_swapbacked(newfolio);
421 return MIGRATEPAGE_SUCCESS;
424 oldzone = folio_zone(folio);
425 newzone = folio_zone(newfolio);
428 if (!folio_ref_freeze(folio, expected_count)) {
429 xas_unlock_irq(&xas);
434 * Now we know that no one else is looking at the folio:
435 * no turning back from here.
437 newfolio->index = folio->index;
438 newfolio->mapping = folio->mapping;
439 folio_ref_add(newfolio, nr); /* add cache reference */
440 if (folio_test_swapbacked(folio)) {
441 __folio_set_swapbacked(newfolio);
442 if (folio_test_swapcache(folio)) {
443 folio_set_swapcache(newfolio);
444 newfolio->private = folio_get_private(folio);
448 VM_BUG_ON_FOLIO(folio_test_swapcache(folio), folio);
452 /* Move dirty while page refs frozen and newpage not yet exposed */
453 dirty = folio_test_dirty(folio);
455 folio_clear_dirty(folio);
456 folio_set_dirty(newfolio);
459 /* Swap cache still stores N entries instead of a high-order entry */
460 for (i = 0; i < entries; i++) {
461 xas_store(&xas, newfolio);
466 * Drop cache reference from old page by unfreezing
467 * to one less reference.
468 * We know this isn't the last reference.
470 folio_ref_unfreeze(folio, expected_count - nr);
473 /* Leave irq disabled to prevent preemption while updating stats */
476 * If moved to a different zone then also account
477 * the page for that zone. Other VM counters will be
478 * taken care of when we establish references to the
479 * new page and drop references to the old page.
481 * Note that anonymous pages are accounted for
482 * via NR_FILE_PAGES and NR_ANON_MAPPED if they
483 * are mapped to swap space.
485 if (newzone != oldzone) {
486 struct lruvec *old_lruvec, *new_lruvec;
487 struct mem_cgroup *memcg;
489 memcg = folio_memcg(folio);
490 old_lruvec = mem_cgroup_lruvec(memcg, oldzone->zone_pgdat);
491 new_lruvec = mem_cgroup_lruvec(memcg, newzone->zone_pgdat);
493 __mod_lruvec_state(old_lruvec, NR_FILE_PAGES, -nr);
494 __mod_lruvec_state(new_lruvec, NR_FILE_PAGES, nr);
495 if (folio_test_swapbacked(folio) && !folio_test_swapcache(folio)) {
496 __mod_lruvec_state(old_lruvec, NR_SHMEM, -nr);
497 __mod_lruvec_state(new_lruvec, NR_SHMEM, nr);
499 if (folio_test_pmd_mappable(folio)) {
500 __mod_lruvec_state(old_lruvec, NR_SHMEM_THPS, -nr);
501 __mod_lruvec_state(new_lruvec, NR_SHMEM_THPS, nr);
505 if (folio_test_swapcache(folio)) {
506 __mod_lruvec_state(old_lruvec, NR_SWAPCACHE, -nr);
507 __mod_lruvec_state(new_lruvec, NR_SWAPCACHE, nr);
510 if (dirty && mapping_can_writeback(mapping)) {
511 __mod_lruvec_state(old_lruvec, NR_FILE_DIRTY, -nr);
512 __mod_zone_page_state(oldzone, NR_ZONE_WRITE_PENDING, -nr);
513 __mod_lruvec_state(new_lruvec, NR_FILE_DIRTY, nr);
514 __mod_zone_page_state(newzone, NR_ZONE_WRITE_PENDING, nr);
519 return MIGRATEPAGE_SUCCESS;
521 EXPORT_SYMBOL(folio_migrate_mapping);
524 * The expected number of remaining references is the same as that
525 * of folio_migrate_mapping().
527 int migrate_huge_page_move_mapping(struct address_space *mapping,
528 struct folio *dst, struct folio *src)
530 XA_STATE(xas, &mapping->i_pages, folio_index(src));
534 expected_count = 2 + folio_has_private(src);
535 if (!folio_ref_freeze(src, expected_count)) {
536 xas_unlock_irq(&xas);
540 dst->index = src->index;
541 dst->mapping = src->mapping;
545 xas_store(&xas, dst);
547 folio_ref_unfreeze(src, expected_count - 1);
549 xas_unlock_irq(&xas);
551 return MIGRATEPAGE_SUCCESS;
555 * Copy the flags and some other ancillary information
557 void folio_migrate_flags(struct folio *newfolio, struct folio *folio)
561 if (folio_test_error(folio))
562 folio_set_error(newfolio);
563 if (folio_test_referenced(folio))
564 folio_set_referenced(newfolio);
565 if (folio_test_uptodate(folio))
566 folio_mark_uptodate(newfolio);
567 if (folio_test_clear_active(folio)) {
568 VM_BUG_ON_FOLIO(folio_test_unevictable(folio), folio);
569 folio_set_active(newfolio);
570 } else if (folio_test_clear_unevictable(folio))
571 folio_set_unevictable(newfolio);
572 if (folio_test_workingset(folio))
573 folio_set_workingset(newfolio);
574 if (folio_test_checked(folio))
575 folio_set_checked(newfolio);
577 * PG_anon_exclusive (-> PG_mappedtodisk) is always migrated via
578 * migration entries. We can still have PG_anon_exclusive set on an
579 * effectively unmapped and unreferenced first sub-pages of an
580 * anonymous THP: we can simply copy it here via PG_mappedtodisk.
582 if (folio_test_mappedtodisk(folio))
583 folio_set_mappedtodisk(newfolio);
585 /* Move dirty on pages not done by folio_migrate_mapping() */
586 if (folio_test_dirty(folio))
587 folio_set_dirty(newfolio);
589 if (folio_test_young(folio))
590 folio_set_young(newfolio);
591 if (folio_test_idle(folio))
592 folio_set_idle(newfolio);
595 * Copy NUMA information to the new page, to prevent over-eager
596 * future migrations of this same page.
598 cpupid = page_cpupid_xchg_last(&folio->page, -1);
600 * For memory tiering mode, when migrate between slow and fast
601 * memory node, reset cpupid, because that is used to record
602 * page access time in slow memory node.
604 if (sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING) {
605 bool f_toptier = node_is_toptier(page_to_nid(&folio->page));
606 bool t_toptier = node_is_toptier(page_to_nid(&newfolio->page));
608 if (f_toptier != t_toptier)
611 page_cpupid_xchg_last(&newfolio->page, cpupid);
613 folio_migrate_ksm(newfolio, folio);
615 * Please do not reorder this without considering how mm/ksm.c's
616 * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache().
618 if (folio_test_swapcache(folio))
619 folio_clear_swapcache(folio);
620 folio_clear_private(folio);
622 /* page->private contains hugetlb specific flags */
623 if (!folio_test_hugetlb(folio))
624 folio->private = NULL;
627 * If any waiters have accumulated on the new page then
630 if (folio_test_writeback(newfolio))
631 folio_end_writeback(newfolio);
634 * PG_readahead shares the same bit with PG_reclaim. The above
635 * end_page_writeback() may clear PG_readahead mistakenly, so set the
638 if (folio_test_readahead(folio))
639 folio_set_readahead(newfolio);
641 folio_copy_owner(newfolio, folio);
643 if (!folio_test_hugetlb(folio))
644 mem_cgroup_migrate(folio, newfolio);
646 EXPORT_SYMBOL(folio_migrate_flags);
648 void folio_migrate_copy(struct folio *newfolio, struct folio *folio)
650 folio_copy(newfolio, folio);
651 folio_migrate_flags(newfolio, folio);
653 EXPORT_SYMBOL(folio_migrate_copy);
655 /************************************************************
656 * Migration functions
657 ***********************************************************/
659 int migrate_folio_extra(struct address_space *mapping, struct folio *dst,
660 struct folio *src, enum migrate_mode mode, int extra_count)
664 BUG_ON(folio_test_writeback(src)); /* Writeback must be complete */
666 rc = folio_migrate_mapping(mapping, dst, src, extra_count);
668 if (rc != MIGRATEPAGE_SUCCESS)
671 if (mode != MIGRATE_SYNC_NO_COPY)
672 folio_migrate_copy(dst, src);
674 folio_migrate_flags(dst, src);
675 return MIGRATEPAGE_SUCCESS;
679 * migrate_folio() - Simple folio migration.
680 * @mapping: The address_space containing the folio.
681 * @dst: The folio to migrate the data to.
682 * @src: The folio containing the current data.
683 * @mode: How to migrate the page.
685 * Common logic to directly migrate a single LRU folio suitable for
686 * folios that do not use PagePrivate/PagePrivate2.
688 * Folios are locked upon entry and exit.
690 int migrate_folio(struct address_space *mapping, struct folio *dst,
691 struct folio *src, enum migrate_mode mode)
693 return migrate_folio_extra(mapping, dst, src, mode, 0);
695 EXPORT_SYMBOL(migrate_folio);
697 #ifdef CONFIG_BUFFER_HEAD
698 /* Returns true if all buffers are successfully locked */
699 static bool buffer_migrate_lock_buffers(struct buffer_head *head,
700 enum migrate_mode mode)
702 struct buffer_head *bh = head;
703 struct buffer_head *failed_bh;
706 if (!trylock_buffer(bh)) {
707 if (mode == MIGRATE_ASYNC)
709 if (mode == MIGRATE_SYNC_LIGHT && !buffer_uptodate(bh))
714 bh = bh->b_this_page;
715 } while (bh != head);
720 /* We failed to lock the buffer and cannot stall. */
723 while (bh != failed_bh) {
725 bh = bh->b_this_page;
731 static int __buffer_migrate_folio(struct address_space *mapping,
732 struct folio *dst, struct folio *src, enum migrate_mode mode,
735 struct buffer_head *bh, *head;
739 head = folio_buffers(src);
741 return migrate_folio(mapping, dst, src, mode);
743 /* Check whether page does not have extra refs before we do more work */
744 expected_count = folio_expected_refs(mapping, src);
745 if (folio_ref_count(src) != expected_count)
748 if (!buffer_migrate_lock_buffers(head, mode))
753 bool invalidated = false;
757 spin_lock(&mapping->private_lock);
760 if (atomic_read(&bh->b_count)) {
764 bh = bh->b_this_page;
765 } while (bh != head);
771 spin_unlock(&mapping->private_lock);
772 invalidate_bh_lrus();
774 goto recheck_buffers;
778 rc = folio_migrate_mapping(mapping, dst, src, 0);
779 if (rc != MIGRATEPAGE_SUCCESS)
782 folio_attach_private(dst, folio_detach_private(src));
786 folio_set_bh(bh, dst, bh_offset(bh));
787 bh = bh->b_this_page;
788 } while (bh != head);
790 if (mode != MIGRATE_SYNC_NO_COPY)
791 folio_migrate_copy(dst, src);
793 folio_migrate_flags(dst, src);
795 rc = MIGRATEPAGE_SUCCESS;
798 spin_unlock(&mapping->private_lock);
802 bh = bh->b_this_page;
803 } while (bh != head);
809 * buffer_migrate_folio() - Migration function for folios with buffers.
810 * @mapping: The address space containing @src.
811 * @dst: The folio to migrate to.
812 * @src: The folio to migrate from.
813 * @mode: How to migrate the folio.
815 * This function can only be used if the underlying filesystem guarantees
816 * that no other references to @src exist. For example attached buffer
817 * heads are accessed only under the folio lock. If your filesystem cannot
818 * provide this guarantee, buffer_migrate_folio_norefs() may be more
821 * Return: 0 on success or a negative errno on failure.
823 int buffer_migrate_folio(struct address_space *mapping,
824 struct folio *dst, struct folio *src, enum migrate_mode mode)
826 return __buffer_migrate_folio(mapping, dst, src, mode, false);
828 EXPORT_SYMBOL(buffer_migrate_folio);
831 * buffer_migrate_folio_norefs() - Migration function for folios with buffers.
832 * @mapping: The address space containing @src.
833 * @dst: The folio to migrate to.
834 * @src: The folio to migrate from.
835 * @mode: How to migrate the folio.
837 * Like buffer_migrate_folio() except that this variant is more careful
838 * and checks that there are also no buffer head references. This function
839 * is the right one for mappings where buffer heads are directly looked
840 * up and referenced (such as block device mappings).
842 * Return: 0 on success or a negative errno on failure.
844 int buffer_migrate_folio_norefs(struct address_space *mapping,
845 struct folio *dst, struct folio *src, enum migrate_mode mode)
847 return __buffer_migrate_folio(mapping, dst, src, mode, true);
849 EXPORT_SYMBOL_GPL(buffer_migrate_folio_norefs);
850 #endif /* CONFIG_BUFFER_HEAD */
852 int filemap_migrate_folio(struct address_space *mapping,
853 struct folio *dst, struct folio *src, enum migrate_mode mode)
857 ret = folio_migrate_mapping(mapping, dst, src, 0);
858 if (ret != MIGRATEPAGE_SUCCESS)
861 if (folio_get_private(src))
862 folio_attach_private(dst, folio_detach_private(src));
864 if (mode != MIGRATE_SYNC_NO_COPY)
865 folio_migrate_copy(dst, src);
867 folio_migrate_flags(dst, src);
868 return MIGRATEPAGE_SUCCESS;
870 EXPORT_SYMBOL_GPL(filemap_migrate_folio);
873 * Writeback a folio to clean the dirty state
875 static int writeout(struct address_space *mapping, struct folio *folio)
877 struct writeback_control wbc = {
878 .sync_mode = WB_SYNC_NONE,
881 .range_end = LLONG_MAX,
886 if (!mapping->a_ops->writepage)
887 /* No write method for the address space */
890 if (!folio_clear_dirty_for_io(folio))
891 /* Someone else already triggered a write */
895 * A dirty folio may imply that the underlying filesystem has
896 * the folio on some queue. So the folio must be clean for
897 * migration. Writeout may mean we lose the lock and the
898 * folio state is no longer what we checked for earlier.
899 * At this point we know that the migration attempt cannot
902 remove_migration_ptes(folio, folio, false);
904 rc = mapping->a_ops->writepage(&folio->page, &wbc);
906 if (rc != AOP_WRITEPAGE_ACTIVATE)
907 /* unlocked. Relock */
910 return (rc < 0) ? -EIO : -EAGAIN;
914 * Default handling if a filesystem does not provide a migration function.
916 static int fallback_migrate_folio(struct address_space *mapping,
917 struct folio *dst, struct folio *src, enum migrate_mode mode)
919 if (folio_test_dirty(src)) {
920 /* Only writeback folios in full synchronous migration */
923 case MIGRATE_SYNC_NO_COPY:
928 return writeout(mapping, src);
932 * Buffers may be managed in a filesystem specific way.
933 * We must have no buffers or drop them.
935 if (!filemap_release_folio(src, GFP_KERNEL))
936 return mode == MIGRATE_SYNC ? -EAGAIN : -EBUSY;
938 return migrate_folio(mapping, dst, src, mode);
942 * Move a page to a newly allocated page
943 * The page is locked and all ptes have been successfully removed.
945 * The new page will have replaced the old page if this function
950 * MIGRATEPAGE_SUCCESS - success
952 static int move_to_new_folio(struct folio *dst, struct folio *src,
953 enum migrate_mode mode)
956 bool is_lru = !__PageMovable(&src->page);
958 VM_BUG_ON_FOLIO(!folio_test_locked(src), src);
959 VM_BUG_ON_FOLIO(!folio_test_locked(dst), dst);
961 if (likely(is_lru)) {
962 struct address_space *mapping = folio_mapping(src);
965 rc = migrate_folio(mapping, dst, src, mode);
966 else if (mapping->a_ops->migrate_folio)
968 * Most folios have a mapping and most filesystems
969 * provide a migrate_folio callback. Anonymous folios
970 * are part of swap space which also has its own
971 * migrate_folio callback. This is the most common path
972 * for page migration.
974 rc = mapping->a_ops->migrate_folio(mapping, dst, src,
977 rc = fallback_migrate_folio(mapping, dst, src, mode);
979 const struct movable_operations *mops;
982 * In case of non-lru page, it could be released after
983 * isolation step. In that case, we shouldn't try migration.
985 VM_BUG_ON_FOLIO(!folio_test_isolated(src), src);
986 if (!folio_test_movable(src)) {
987 rc = MIGRATEPAGE_SUCCESS;
988 folio_clear_isolated(src);
992 mops = folio_movable_ops(src);
993 rc = mops->migrate_page(&dst->page, &src->page, mode);
994 WARN_ON_ONCE(rc == MIGRATEPAGE_SUCCESS &&
995 !folio_test_isolated(src));
999 * When successful, old pagecache src->mapping must be cleared before
1000 * src is freed; but stats require that PageAnon be left as PageAnon.
1002 if (rc == MIGRATEPAGE_SUCCESS) {
1003 if (__PageMovable(&src->page)) {
1004 VM_BUG_ON_FOLIO(!folio_test_isolated(src), src);
1007 * We clear PG_movable under page_lock so any compactor
1008 * cannot try to migrate this page.
1010 folio_clear_isolated(src);
1014 * Anonymous and movable src->mapping will be cleared by
1015 * free_pages_prepare so don't reset it here for keeping
1016 * the type to work PageAnon, for example.
1018 if (!folio_mapping_flags(src))
1019 src->mapping = NULL;
1021 if (likely(!folio_is_zone_device(dst)))
1022 flush_dcache_folio(dst);
1029 * To record some information during migration, we use some unused
1030 * fields (mapping and private) of struct folio of the newly allocated
1031 * destination folio. This is safe because nobody is using them
1034 union migration_ptr {
1035 struct anon_vma *anon_vma;
1036 struct address_space *mapping;
1038 static void __migrate_folio_record(struct folio *dst,
1039 unsigned long page_was_mapped,
1040 struct anon_vma *anon_vma)
1042 union migration_ptr ptr = { .anon_vma = anon_vma };
1043 dst->mapping = ptr.mapping;
1044 dst->private = (void *)page_was_mapped;
1047 static void __migrate_folio_extract(struct folio *dst,
1048 int *page_was_mappedp,
1049 struct anon_vma **anon_vmap)
1051 union migration_ptr ptr = { .mapping = dst->mapping };
1052 *anon_vmap = ptr.anon_vma;
1053 *page_was_mappedp = (unsigned long)dst->private;
1054 dst->mapping = NULL;
1055 dst->private = NULL;
1058 /* Restore the source folio to the original state upon failure */
1059 static void migrate_folio_undo_src(struct folio *src,
1060 int page_was_mapped,
1061 struct anon_vma *anon_vma,
1063 struct list_head *ret)
1065 if (page_was_mapped)
1066 remove_migration_ptes(src, src, false);
1067 /* Drop an anon_vma reference if we took one */
1069 put_anon_vma(anon_vma);
1073 list_move_tail(&src->lru, ret);
1076 /* Restore the destination folio to the original state upon failure */
1077 static void migrate_folio_undo_dst(struct folio *dst, bool locked,
1078 free_folio_t put_new_folio, unsigned long private)
1083 put_new_folio(dst, private);
1088 /* Cleanup src folio upon migration success */
1089 static void migrate_folio_done(struct folio *src,
1090 enum migrate_reason reason)
1093 * Compaction can migrate also non-LRU pages which are
1094 * not accounted to NR_ISOLATED_*. They can be recognized
1097 if (likely(!__folio_test_movable(src)))
1098 mod_node_page_state(folio_pgdat(src), NR_ISOLATED_ANON +
1099 folio_is_file_lru(src), -folio_nr_pages(src));
1101 if (reason != MR_MEMORY_FAILURE)
1102 /* We release the page in page_handle_poison. */
1106 /* Obtain the lock on page, remove all ptes. */
1107 static int migrate_folio_unmap(new_folio_t get_new_folio,
1108 free_folio_t put_new_folio, unsigned long private,
1109 struct folio *src, struct folio **dstp, enum migrate_mode mode,
1110 enum migrate_reason reason, struct list_head *ret)
1114 int page_was_mapped = 0;
1115 struct anon_vma *anon_vma = NULL;
1116 bool is_lru = !__PageMovable(&src->page);
1117 bool locked = false;
1118 bool dst_locked = false;
1120 if (folio_ref_count(src) == 1) {
1121 /* Folio was freed from under us. So we are done. */
1122 folio_clear_active(src);
1123 folio_clear_unevictable(src);
1124 /* free_pages_prepare() will clear PG_isolated. */
1125 list_del(&src->lru);
1126 migrate_folio_done(src, reason);
1127 return MIGRATEPAGE_SUCCESS;
1130 dst = get_new_folio(src, private);
1135 dst->private = NULL;
1137 if (!folio_trylock(src)) {
1138 if (mode == MIGRATE_ASYNC)
1142 * It's not safe for direct compaction to call lock_page.
1143 * For example, during page readahead pages are added locked
1144 * to the LRU. Later, when the IO completes the pages are
1145 * marked uptodate and unlocked. However, the queueing
1146 * could be merging multiple pages for one bio (e.g.
1147 * mpage_readahead). If an allocation happens for the
1148 * second or third page, the process can end up locking
1149 * the same page twice and deadlocking. Rather than
1150 * trying to be clever about what pages can be locked,
1151 * avoid the use of lock_page for direct compaction
1154 if (current->flags & PF_MEMALLOC)
1158 * In "light" mode, we can wait for transient locks (eg
1159 * inserting a page into the page table), but it's not
1160 * worth waiting for I/O.
1162 if (mode == MIGRATE_SYNC_LIGHT && !folio_test_uptodate(src))
1169 if (folio_test_writeback(src)) {
1171 * Only in the case of a full synchronous migration is it
1172 * necessary to wait for PageWriteback. In the async case,
1173 * the retry loop is too short and in the sync-light case,
1174 * the overhead of stalling is too much
1178 case MIGRATE_SYNC_NO_COPY:
1184 folio_wait_writeback(src);
1188 * By try_to_migrate(), src->mapcount goes down to 0 here. In this case,
1189 * we cannot notice that anon_vma is freed while we migrate a page.
1190 * This get_anon_vma() delays freeing anon_vma pointer until the end
1191 * of migration. File cache pages are no problem because of page_lock()
1192 * File Caches may use write_page() or lock_page() in migration, then,
1193 * just care Anon page here.
1195 * Only folio_get_anon_vma() understands the subtleties of
1196 * getting a hold on an anon_vma from outside one of its mms.
1197 * But if we cannot get anon_vma, then we won't need it anyway,
1198 * because that implies that the anon page is no longer mapped
1199 * (and cannot be remapped so long as we hold the page lock).
1201 if (folio_test_anon(src) && !folio_test_ksm(src))
1202 anon_vma = folio_get_anon_vma(src);
1205 * Block others from accessing the new page when we get around to
1206 * establishing additional references. We are usually the only one
1207 * holding a reference to dst at this point. We used to have a BUG
1208 * here if folio_trylock(dst) fails, but would like to allow for
1209 * cases where there might be a race with the previous use of dst.
1210 * This is much like races on refcount of oldpage: just don't BUG().
1212 if (unlikely(!folio_trylock(dst)))
1216 if (unlikely(!is_lru)) {
1217 __migrate_folio_record(dst, page_was_mapped, anon_vma);
1218 return MIGRATEPAGE_UNMAP;
1222 * Corner case handling:
1223 * 1. When a new swap-cache page is read into, it is added to the LRU
1224 * and treated as swapcache but it has no rmap yet.
1225 * Calling try_to_unmap() against a src->mapping==NULL page will
1226 * trigger a BUG. So handle it here.
1227 * 2. An orphaned page (see truncate_cleanup_page) might have
1228 * fs-private metadata. The page can be picked up due to memory
1229 * offlining. Everywhere else except page reclaim, the page is
1230 * invisible to the vm, so the page can not be migrated. So try to
1231 * free the metadata, so the page can be freed.
1233 if (!src->mapping) {
1234 if (folio_test_private(src)) {
1235 try_to_free_buffers(src);
1238 } else if (folio_mapped(src)) {
1239 /* Establish migration ptes */
1240 VM_BUG_ON_FOLIO(folio_test_anon(src) &&
1241 !folio_test_ksm(src) && !anon_vma, src);
1242 try_to_migrate(src, mode == MIGRATE_ASYNC ? TTU_BATCH_FLUSH : 0);
1243 page_was_mapped = 1;
1246 if (!folio_mapped(src)) {
1247 __migrate_folio_record(dst, page_was_mapped, anon_vma);
1248 return MIGRATEPAGE_UNMAP;
1253 * A folio that has not been unmapped will be restored to
1254 * right list unless we want to retry.
1259 migrate_folio_undo_src(src, page_was_mapped, anon_vma, locked, ret);
1260 migrate_folio_undo_dst(dst, dst_locked, put_new_folio, private);
1265 /* Migrate the folio to the newly allocated folio in dst. */
1266 static int migrate_folio_move(free_folio_t put_new_folio, unsigned long private,
1267 struct folio *src, struct folio *dst,
1268 enum migrate_mode mode, enum migrate_reason reason,
1269 struct list_head *ret)
1272 int page_was_mapped = 0;
1273 struct anon_vma *anon_vma = NULL;
1274 bool is_lru = !__PageMovable(&src->page);
1275 struct list_head *prev;
1277 __migrate_folio_extract(dst, &page_was_mapped, &anon_vma);
1278 prev = dst->lru.prev;
1279 list_del(&dst->lru);
1281 rc = move_to_new_folio(dst, src, mode);
1285 if (unlikely(!is_lru))
1286 goto out_unlock_both;
1289 * When successful, push dst to LRU immediately: so that if it
1290 * turns out to be an mlocked page, remove_migration_ptes() will
1291 * automatically build up the correct dst->mlock_count for it.
1293 * We would like to do something similar for the old page, when
1294 * unsuccessful, and other cases when a page has been temporarily
1295 * isolated from the unevictable LRU: but this case is the easiest.
1298 if (page_was_mapped)
1301 if (page_was_mapped)
1302 remove_migration_ptes(src, dst, false);
1306 set_page_owner_migrate_reason(&dst->page, reason);
1308 * If migration is successful, decrease refcount of dst,
1309 * which will not free the page because new page owner increased
1315 * A folio that has been migrated has all references removed
1316 * and will be freed.
1318 list_del(&src->lru);
1319 /* Drop an anon_vma reference if we took one */
1321 put_anon_vma(anon_vma);
1323 migrate_folio_done(src, reason);
1328 * A folio that has not been migrated will be restored to
1329 * right list unless we want to retry.
1331 if (rc == -EAGAIN) {
1332 list_add(&dst->lru, prev);
1333 __migrate_folio_record(dst, page_was_mapped, anon_vma);
1337 migrate_folio_undo_src(src, page_was_mapped, anon_vma, true, ret);
1338 migrate_folio_undo_dst(dst, true, put_new_folio, private);
1344 * Counterpart of unmap_and_move_page() for hugepage migration.
1346 * This function doesn't wait the completion of hugepage I/O
1347 * because there is no race between I/O and migration for hugepage.
1348 * Note that currently hugepage I/O occurs only in direct I/O
1349 * where no lock is held and PG_writeback is irrelevant,
1350 * and writeback status of all subpages are counted in the reference
1351 * count of the head page (i.e. if all subpages of a 2MB hugepage are
1352 * under direct I/O, the reference of the head page is 512 and a bit more.)
1353 * This means that when we try to migrate hugepage whose subpages are
1354 * doing direct I/O, some references remain after try_to_unmap() and
1355 * hugepage migration fails without data corruption.
1357 * There is also no race when direct I/O is issued on the page under migration,
1358 * because then pte is replaced with migration swap entry and direct I/O code
1359 * will wait in the page fault for migration to complete.
1361 static int unmap_and_move_huge_page(new_folio_t get_new_folio,
1362 free_folio_t put_new_folio, unsigned long private,
1363 struct folio *src, int force, enum migrate_mode mode,
1364 int reason, struct list_head *ret)
1368 int page_was_mapped = 0;
1369 struct anon_vma *anon_vma = NULL;
1370 struct address_space *mapping = NULL;
1372 if (folio_ref_count(src) == 1) {
1373 /* page was freed from under us. So we are done. */
1374 folio_putback_active_hugetlb(src);
1375 return MIGRATEPAGE_SUCCESS;
1378 dst = get_new_folio(src, private);
1382 if (!folio_trylock(src)) {
1387 case MIGRATE_SYNC_NO_COPY:
1396 * Check for pages which are in the process of being freed. Without
1397 * folio_mapping() set, hugetlbfs specific move page routine will not
1398 * be called and we could leak usage counts for subpools.
1400 if (hugetlb_folio_subpool(src) && !folio_mapping(src)) {
1405 if (folio_test_anon(src))
1406 anon_vma = folio_get_anon_vma(src);
1408 if (unlikely(!folio_trylock(dst)))
1411 if (folio_mapped(src)) {
1412 enum ttu_flags ttu = 0;
1414 if (!folio_test_anon(src)) {
1416 * In shared mappings, try_to_unmap could potentially
1417 * call huge_pmd_unshare. Because of this, take
1418 * semaphore in write mode here and set TTU_RMAP_LOCKED
1419 * to let lower levels know we have taken the lock.
1421 mapping = hugetlb_page_mapping_lock_write(&src->page);
1422 if (unlikely(!mapping))
1423 goto unlock_put_anon;
1425 ttu = TTU_RMAP_LOCKED;
1428 try_to_migrate(src, ttu);
1429 page_was_mapped = 1;
1431 if (ttu & TTU_RMAP_LOCKED)
1432 i_mmap_unlock_write(mapping);
1435 if (!folio_mapped(src))
1436 rc = move_to_new_folio(dst, src, mode);
1438 if (page_was_mapped)
1439 remove_migration_ptes(src,
1440 rc == MIGRATEPAGE_SUCCESS ? dst : src, false);
1447 put_anon_vma(anon_vma);
1449 if (rc == MIGRATEPAGE_SUCCESS) {
1450 move_hugetlb_state(src, dst, reason);
1451 put_new_folio = NULL;
1457 if (rc == MIGRATEPAGE_SUCCESS)
1458 folio_putback_active_hugetlb(src);
1459 else if (rc != -EAGAIN)
1460 list_move_tail(&src->lru, ret);
1463 * If migration was not successful and there's a freeing callback, use
1464 * it. Otherwise, put_page() will drop the reference grabbed during
1468 put_new_folio(dst, private);
1470 folio_putback_active_hugetlb(dst);
1475 static inline int try_split_folio(struct folio *folio, struct list_head *split_folios)
1480 rc = split_folio_to_list(folio, split_folios);
1481 folio_unlock(folio);
1483 list_move_tail(&folio->lru, split_folios);
1488 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1489 #define NR_MAX_BATCHED_MIGRATION HPAGE_PMD_NR
1491 #define NR_MAX_BATCHED_MIGRATION 512
1493 #define NR_MAX_MIGRATE_PAGES_RETRY 10
1494 #define NR_MAX_MIGRATE_ASYNC_RETRY 3
1495 #define NR_MAX_MIGRATE_SYNC_RETRY \
1496 (NR_MAX_MIGRATE_PAGES_RETRY - NR_MAX_MIGRATE_ASYNC_RETRY)
1498 struct migrate_pages_stats {
1499 int nr_succeeded; /* Normal and large folios migrated successfully, in
1500 units of base pages */
1501 int nr_failed_pages; /* Normal and large folios failed to be migrated, in
1502 units of base pages. Untried folios aren't counted */
1503 int nr_thp_succeeded; /* THP migrated successfully */
1504 int nr_thp_failed; /* THP failed to be migrated */
1505 int nr_thp_split; /* THP split before migrating */
1509 * Returns the number of hugetlb folios that were not migrated, or an error code
1510 * after NR_MAX_MIGRATE_PAGES_RETRY attempts or if no hugetlb folios are movable
1511 * any more because the list has become empty or no retryable hugetlb folios
1512 * exist any more. It is caller's responsibility to call putback_movable_pages()
1515 static int migrate_hugetlbs(struct list_head *from, new_folio_t get_new_folio,
1516 free_folio_t put_new_folio, unsigned long private,
1517 enum migrate_mode mode, int reason,
1518 struct migrate_pages_stats *stats,
1519 struct list_head *ret_folios)
1523 int nr_retry_pages = 0;
1525 struct folio *folio, *folio2;
1528 for (pass = 0; pass < NR_MAX_MIGRATE_PAGES_RETRY && retry; pass++) {
1532 list_for_each_entry_safe(folio, folio2, from, lru) {
1533 if (!folio_test_hugetlb(folio))
1536 nr_pages = folio_nr_pages(folio);
1541 * Migratability of hugepages depends on architectures and
1542 * their size. This check is necessary because some callers
1543 * of hugepage migration like soft offline and memory
1544 * hotremove don't walk through page tables or check whether
1545 * the hugepage is pmd-based or not before kicking migration.
1547 if (!hugepage_migration_supported(folio_hstate(folio))) {
1549 stats->nr_failed_pages += nr_pages;
1550 list_move_tail(&folio->lru, ret_folios);
1554 rc = unmap_and_move_huge_page(get_new_folio,
1555 put_new_folio, private,
1556 folio, pass > 2, mode,
1557 reason, ret_folios);
1560 * Success: hugetlb folio will be put back
1561 * -EAGAIN: stay on the from list
1562 * -ENOMEM: stay on the from list
1563 * Other errno: put on ret_folios list
1568 * When memory is low, don't bother to try to migrate
1569 * other folios, just exit.
1571 stats->nr_failed_pages += nr_pages + nr_retry_pages;
1575 nr_retry_pages += nr_pages;
1577 case MIGRATEPAGE_SUCCESS:
1578 stats->nr_succeeded += nr_pages;
1582 * Permanent failure (-EBUSY, etc.):
1583 * unlike -EAGAIN case, the failed folio is
1584 * removed from migration folio list and not
1585 * retried in the next outer loop.
1588 stats->nr_failed_pages += nr_pages;
1594 * nr_failed is number of hugetlb folios failed to be migrated. After
1595 * NR_MAX_MIGRATE_PAGES_RETRY attempts, give up and count retried hugetlb
1599 stats->nr_failed_pages += nr_retry_pages;
1605 * migrate_pages_batch() first unmaps folios in the from list as many as
1606 * possible, then move the unmapped folios.
1608 * We only batch migration if mode == MIGRATE_ASYNC to avoid to wait a
1609 * lock or bit when we have locked more than one folio. Which may cause
1610 * deadlock (e.g., for loop device). So, if mode != MIGRATE_ASYNC, the
1611 * length of the from list must be <= 1.
1613 static int migrate_pages_batch(struct list_head *from,
1614 new_folio_t get_new_folio, free_folio_t put_new_folio,
1615 unsigned long private, enum migrate_mode mode, int reason,
1616 struct list_head *ret_folios, struct list_head *split_folios,
1617 struct migrate_pages_stats *stats, int nr_pass)
1622 int nr_retry_pages = 0;
1624 bool is_thp = false;
1625 struct folio *folio, *folio2, *dst = NULL, *dst2;
1626 int rc, rc_saved = 0, nr_pages;
1627 LIST_HEAD(unmap_folios);
1628 LIST_HEAD(dst_folios);
1629 bool nosplit = (reason == MR_NUMA_MISPLACED);
1631 VM_WARN_ON_ONCE(mode != MIGRATE_ASYNC &&
1632 !list_empty(from) && !list_is_singular(from));
1634 for (pass = 0; pass < nr_pass && retry; pass++) {
1639 list_for_each_entry_safe(folio, folio2, from, lru) {
1640 is_thp = folio_test_large(folio) && folio_test_pmd_mappable(folio);
1641 nr_pages = folio_nr_pages(folio);
1646 * Large folio migration might be unsupported or
1647 * the allocation might be failed so we should retry
1648 * on the same folio with the large folio split
1651 * Split folios are put in split_folios, and
1652 * we will migrate them after the rest of the
1653 * list is processed.
1655 if (!thp_migration_supported() && is_thp) {
1657 stats->nr_thp_failed++;
1658 if (!try_split_folio(folio, split_folios)) {
1659 stats->nr_thp_split++;
1662 stats->nr_failed_pages += nr_pages;
1663 list_move_tail(&folio->lru, ret_folios);
1667 rc = migrate_folio_unmap(get_new_folio, put_new_folio,
1668 private, folio, &dst, mode, reason,
1672 * Success: folio will be freed
1673 * Unmap: folio will be put on unmap_folios list,
1674 * dst folio put on dst_folios list
1675 * -EAGAIN: stay on the from list
1676 * -ENOMEM: stay on the from list
1677 * Other errno: put on ret_folios list
1682 * When memory is low, don't bother to try to migrate
1683 * other folios, move unmapped folios, then exit.
1686 stats->nr_thp_failed += is_thp;
1687 /* Large folio NUMA faulting doesn't split to retry. */
1688 if (folio_test_large(folio) && !nosplit) {
1689 int ret = try_split_folio(folio, split_folios);
1692 stats->nr_thp_split += is_thp;
1694 } else if (reason == MR_LONGTERM_PIN &&
1697 * Try again to split large folio to
1698 * mitigate the failure of longterm pinning.
1701 thp_retry += is_thp;
1702 nr_retry_pages += nr_pages;
1703 /* Undo duplicated failure counting. */
1705 stats->nr_thp_failed -= is_thp;
1710 stats->nr_failed_pages += nr_pages + nr_retry_pages;
1711 /* nr_failed isn't updated for not used */
1712 stats->nr_thp_failed += thp_retry;
1714 if (list_empty(&unmap_folios))
1720 thp_retry += is_thp;
1721 nr_retry_pages += nr_pages;
1723 case MIGRATEPAGE_SUCCESS:
1724 stats->nr_succeeded += nr_pages;
1725 stats->nr_thp_succeeded += is_thp;
1727 case MIGRATEPAGE_UNMAP:
1728 list_move_tail(&folio->lru, &unmap_folios);
1729 list_add_tail(&dst->lru, &dst_folios);
1733 * Permanent failure (-EBUSY, etc.):
1734 * unlike -EAGAIN case, the failed folio is
1735 * removed from migration folio list and not
1736 * retried in the next outer loop.
1739 stats->nr_thp_failed += is_thp;
1740 stats->nr_failed_pages += nr_pages;
1746 stats->nr_thp_failed += thp_retry;
1747 stats->nr_failed_pages += nr_retry_pages;
1749 /* Flush TLBs for all unmapped folios */
1750 try_to_unmap_flush();
1753 for (pass = 0; pass < nr_pass && retry; pass++) {
1758 dst = list_first_entry(&dst_folios, struct folio, lru);
1759 dst2 = list_next_entry(dst, lru);
1760 list_for_each_entry_safe(folio, folio2, &unmap_folios, lru) {
1761 is_thp = folio_test_large(folio) && folio_test_pmd_mappable(folio);
1762 nr_pages = folio_nr_pages(folio);
1766 rc = migrate_folio_move(put_new_folio, private,
1768 reason, ret_folios);
1771 * Success: folio will be freed
1772 * -EAGAIN: stay on the unmap_folios list
1773 * Other errno: put on ret_folios list
1778 thp_retry += is_thp;
1779 nr_retry_pages += nr_pages;
1781 case MIGRATEPAGE_SUCCESS:
1782 stats->nr_succeeded += nr_pages;
1783 stats->nr_thp_succeeded += is_thp;
1787 stats->nr_thp_failed += is_thp;
1788 stats->nr_failed_pages += nr_pages;
1792 dst2 = list_next_entry(dst, lru);
1796 stats->nr_thp_failed += thp_retry;
1797 stats->nr_failed_pages += nr_retry_pages;
1799 rc = rc_saved ? : nr_failed;
1801 /* Cleanup remaining folios */
1802 dst = list_first_entry(&dst_folios, struct folio, lru);
1803 dst2 = list_next_entry(dst, lru);
1804 list_for_each_entry_safe(folio, folio2, &unmap_folios, lru) {
1805 int page_was_mapped = 0;
1806 struct anon_vma *anon_vma = NULL;
1808 __migrate_folio_extract(dst, &page_was_mapped, &anon_vma);
1809 migrate_folio_undo_src(folio, page_was_mapped, anon_vma,
1811 list_del(&dst->lru);
1812 migrate_folio_undo_dst(dst, true, put_new_folio, private);
1814 dst2 = list_next_entry(dst, lru);
1820 static int migrate_pages_sync(struct list_head *from, new_folio_t get_new_folio,
1821 free_folio_t put_new_folio, unsigned long private,
1822 enum migrate_mode mode, int reason,
1823 struct list_head *ret_folios, struct list_head *split_folios,
1824 struct migrate_pages_stats *stats)
1826 int rc, nr_failed = 0;
1828 struct migrate_pages_stats astats;
1830 memset(&astats, 0, sizeof(astats));
1831 /* Try to migrate in batch with MIGRATE_ASYNC mode firstly */
1832 rc = migrate_pages_batch(from, get_new_folio, put_new_folio, private, MIGRATE_ASYNC,
1833 reason, &folios, split_folios, &astats,
1834 NR_MAX_MIGRATE_ASYNC_RETRY);
1835 stats->nr_succeeded += astats.nr_succeeded;
1836 stats->nr_thp_succeeded += astats.nr_thp_succeeded;
1837 stats->nr_thp_split += astats.nr_thp_split;
1839 stats->nr_failed_pages += astats.nr_failed_pages;
1840 stats->nr_thp_failed += astats.nr_thp_failed;
1841 list_splice_tail(&folios, ret_folios);
1844 stats->nr_thp_failed += astats.nr_thp_split;
1845 nr_failed += astats.nr_thp_split;
1847 * Fall back to migrate all failed folios one by one synchronously. All
1848 * failed folios except split THPs will be retried, so their failure
1851 list_splice_tail_init(&folios, from);
1852 while (!list_empty(from)) {
1853 list_move(from->next, &folios);
1854 rc = migrate_pages_batch(&folios, get_new_folio, put_new_folio,
1855 private, mode, reason, ret_folios,
1856 split_folios, stats, NR_MAX_MIGRATE_SYNC_RETRY);
1857 list_splice_tail_init(&folios, ret_folios);
1867 * migrate_pages - migrate the folios specified in a list, to the free folios
1868 * supplied as the target for the page migration
1870 * @from: The list of folios to be migrated.
1871 * @get_new_folio: The function used to allocate free folios to be used
1872 * as the target of the folio migration.
1873 * @put_new_folio: The function used to free target folios if migration
1874 * fails, or NULL if no special handling is necessary.
1875 * @private: Private data to be passed on to get_new_folio()
1876 * @mode: The migration mode that specifies the constraints for
1877 * folio migration, if any.
1878 * @reason: The reason for folio migration.
1879 * @ret_succeeded: Set to the number of folios migrated successfully if
1880 * the caller passes a non-NULL pointer.
1882 * The function returns after NR_MAX_MIGRATE_PAGES_RETRY attempts or if no folios
1883 * are movable any more because the list has become empty or no retryable folios
1884 * exist any more. It is caller's responsibility to call putback_movable_pages()
1887 * Returns the number of {normal folio, large folio, hugetlb} that were not
1888 * migrated, or an error code. The number of large folio splits will be
1889 * considered as the number of non-migrated large folio, no matter how many
1890 * split folios of the large folio are migrated successfully.
1892 int migrate_pages(struct list_head *from, new_folio_t get_new_folio,
1893 free_folio_t put_new_folio, unsigned long private,
1894 enum migrate_mode mode, int reason, unsigned int *ret_succeeded)
1898 struct folio *folio, *folio2;
1900 LIST_HEAD(ret_folios);
1901 LIST_HEAD(split_folios);
1902 struct migrate_pages_stats stats;
1904 trace_mm_migrate_pages_start(mode, reason);
1906 memset(&stats, 0, sizeof(stats));
1908 rc_gather = migrate_hugetlbs(from, get_new_folio, put_new_folio, private,
1909 mode, reason, &stats, &ret_folios);
1915 list_for_each_entry_safe(folio, folio2, from, lru) {
1916 /* Retried hugetlb folios will be kept in list */
1917 if (folio_test_hugetlb(folio)) {
1918 list_move_tail(&folio->lru, &ret_folios);
1922 nr_pages += folio_nr_pages(folio);
1923 if (nr_pages >= NR_MAX_BATCHED_MIGRATION)
1926 if (nr_pages >= NR_MAX_BATCHED_MIGRATION)
1927 list_cut_before(&folios, from, &folio2->lru);
1929 list_splice_init(from, &folios);
1930 if (mode == MIGRATE_ASYNC)
1931 rc = migrate_pages_batch(&folios, get_new_folio, put_new_folio,
1932 private, mode, reason, &ret_folios,
1933 &split_folios, &stats,
1934 NR_MAX_MIGRATE_PAGES_RETRY);
1936 rc = migrate_pages_sync(&folios, get_new_folio, put_new_folio,
1937 private, mode, reason, &ret_folios,
1938 &split_folios, &stats);
1939 list_splice_tail_init(&folios, &ret_folios);
1942 list_splice_tail(&split_folios, &ret_folios);
1945 if (!list_empty(&split_folios)) {
1947 * Failure isn't counted since all split folios of a large folio
1948 * is counted as 1 failure already. And, we only try to migrate
1949 * with minimal effort, force MIGRATE_ASYNC mode and retry once.
1951 migrate_pages_batch(&split_folios, get_new_folio,
1952 put_new_folio, private, MIGRATE_ASYNC, reason,
1953 &ret_folios, NULL, &stats, 1);
1954 list_splice_tail_init(&split_folios, &ret_folios);
1957 if (!list_empty(from))
1961 * Put the permanent failure folio back to migration list, they
1962 * will be put back to the right list by the caller.
1964 list_splice(&ret_folios, from);
1967 * Return 0 in case all split folios of fail-to-migrate large folios
1968 * are migrated successfully.
1970 if (list_empty(from))
1973 count_vm_events(PGMIGRATE_SUCCESS, stats.nr_succeeded);
1974 count_vm_events(PGMIGRATE_FAIL, stats.nr_failed_pages);
1975 count_vm_events(THP_MIGRATION_SUCCESS, stats.nr_thp_succeeded);
1976 count_vm_events(THP_MIGRATION_FAIL, stats.nr_thp_failed);
1977 count_vm_events(THP_MIGRATION_SPLIT, stats.nr_thp_split);
1978 trace_mm_migrate_pages(stats.nr_succeeded, stats.nr_failed_pages,
1979 stats.nr_thp_succeeded, stats.nr_thp_failed,
1980 stats.nr_thp_split, mode, reason);
1983 *ret_succeeded = stats.nr_succeeded;
1988 struct folio *alloc_migration_target(struct folio *src, unsigned long private)
1990 struct migration_target_control *mtc;
1992 unsigned int order = 0;
1996 mtc = (struct migration_target_control *)private;
1997 gfp_mask = mtc->gfp_mask;
1999 if (nid == NUMA_NO_NODE)
2000 nid = folio_nid(src);
2002 if (folio_test_hugetlb(src)) {
2003 struct hstate *h = folio_hstate(src);
2005 gfp_mask = htlb_modify_alloc_mask(h, gfp_mask);
2006 return alloc_hugetlb_folio_nodemask(h, nid,
2007 mtc->nmask, gfp_mask);
2010 if (folio_test_large(src)) {
2012 * clear __GFP_RECLAIM to make the migration callback
2013 * consistent with regular THP allocations.
2015 gfp_mask &= ~__GFP_RECLAIM;
2016 gfp_mask |= GFP_TRANSHUGE;
2017 order = folio_order(src);
2019 zidx = zone_idx(folio_zone(src));
2020 if (is_highmem_idx(zidx) || zidx == ZONE_MOVABLE)
2021 gfp_mask |= __GFP_HIGHMEM;
2023 return __folio_alloc(gfp_mask, order, nid, mtc->nmask);
2028 static int store_status(int __user *status, int start, int value, int nr)
2031 if (put_user(value, status + start))
2039 static int do_move_pages_to_node(struct mm_struct *mm,
2040 struct list_head *pagelist, int node)
2043 struct migration_target_control mtc = {
2045 .gfp_mask = GFP_HIGHUSER_MOVABLE | __GFP_THISNODE,
2048 err = migrate_pages(pagelist, alloc_migration_target, NULL,
2049 (unsigned long)&mtc, MIGRATE_SYNC, MR_SYSCALL, NULL);
2051 putback_movable_pages(pagelist);
2056 * Resolves the given address to a struct page, isolates it from the LRU and
2057 * puts it to the given pagelist.
2059 * errno - if the page cannot be found/isolated
2060 * 0 - when it doesn't have to be migrated because it is already on the
2062 * 1 - when it has been queued
2064 static int add_page_for_migration(struct mm_struct *mm, const void __user *p,
2065 int node, struct list_head *pagelist, bool migrate_all)
2067 struct vm_area_struct *vma;
2074 addr = (unsigned long)untagged_addr_remote(mm, p);
2077 vma = vma_lookup(mm, addr);
2078 if (!vma || !vma_migratable(vma))
2081 /* FOLL_DUMP to ignore special (like zero) pages */
2082 page = follow_page(vma, addr, FOLL_GET | FOLL_DUMP);
2084 err = PTR_ERR(page);
2092 if (is_zone_device_page(page))
2096 if (page_to_nid(page) == node)
2100 if (page_mapcount(page) > 1 && !migrate_all)
2103 if (PageHuge(page)) {
2104 if (PageHead(page)) {
2105 isolated = isolate_hugetlb(page_folio(page), pagelist);
2106 err = isolated ? 1 : -EBUSY;
2111 head = compound_head(page);
2112 isolated = isolate_lru_page(head);
2119 list_add_tail(&head->lru, pagelist);
2120 mod_node_page_state(page_pgdat(head),
2121 NR_ISOLATED_ANON + page_is_file_lru(head),
2122 thp_nr_pages(head));
2126 * Either remove the duplicate refcount from
2127 * isolate_lru_page() or drop the page ref if it was
2132 mmap_read_unlock(mm);
2136 static int move_pages_and_store_status(struct mm_struct *mm, int node,
2137 struct list_head *pagelist, int __user *status,
2138 int start, int i, unsigned long nr_pages)
2142 if (list_empty(pagelist))
2145 err = do_move_pages_to_node(mm, pagelist, node);
2148 * Positive err means the number of failed
2149 * pages to migrate. Since we are going to
2150 * abort and return the number of non-migrated
2151 * pages, so need to include the rest of the
2152 * nr_pages that have not been attempted as
2156 err += nr_pages - i;
2159 return store_status(status, start, node, i - start);
2163 * Migrate an array of page address onto an array of nodes and fill
2164 * the corresponding array of status.
2166 static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes,
2167 unsigned long nr_pages,
2168 const void __user * __user *pages,
2169 const int __user *nodes,
2170 int __user *status, int flags)
2172 compat_uptr_t __user *compat_pages = (void __user *)pages;
2173 int current_node = NUMA_NO_NODE;
2174 LIST_HEAD(pagelist);
2178 lru_cache_disable();
2180 for (i = start = 0; i < nr_pages; i++) {
2181 const void __user *p;
2185 if (in_compat_syscall()) {
2188 if (get_user(cp, compat_pages + i))
2193 if (get_user(p, pages + i))
2196 if (get_user(node, nodes + i))
2200 if (node < 0 || node >= MAX_NUMNODES)
2202 if (!node_state(node, N_MEMORY))
2206 if (!node_isset(node, task_nodes))
2209 if (current_node == NUMA_NO_NODE) {
2210 current_node = node;
2212 } else if (node != current_node) {
2213 err = move_pages_and_store_status(mm, current_node,
2214 &pagelist, status, start, i, nr_pages);
2218 current_node = node;
2222 * Errors in the page lookup or isolation are not fatal and we simply
2223 * report them via status
2225 err = add_page_for_migration(mm, p, current_node, &pagelist,
2226 flags & MPOL_MF_MOVE_ALL);
2229 /* The page is successfully queued for migration */
2234 * The move_pages() man page does not have an -EEXIST choice, so
2235 * use -EFAULT instead.
2241 * If the page is already on the target node (!err), store the
2242 * node, otherwise, store the err.
2244 err = store_status(status, i, err ? : current_node, 1);
2248 err = move_pages_and_store_status(mm, current_node, &pagelist,
2249 status, start, i, nr_pages);
2251 /* We have accounted for page i */
2256 current_node = NUMA_NO_NODE;
2259 /* Make sure we do not overwrite the existing error */
2260 err1 = move_pages_and_store_status(mm, current_node, &pagelist,
2261 status, start, i, nr_pages);
2270 * Determine the nodes of an array of pages and store it in an array of status.
2272 static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages,
2273 const void __user **pages, int *status)
2279 for (i = 0; i < nr_pages; i++) {
2280 unsigned long addr = (unsigned long)(*pages);
2281 struct vm_area_struct *vma;
2285 vma = vma_lookup(mm, addr);
2289 /* FOLL_DUMP to ignore special (like zero) pages */
2290 page = follow_page(vma, addr, FOLL_GET | FOLL_DUMP);
2292 err = PTR_ERR(page);
2300 if (!is_zone_device_page(page))
2301 err = page_to_nid(page);
2311 mmap_read_unlock(mm);
2314 static int get_compat_pages_array(const void __user *chunk_pages[],
2315 const void __user * __user *pages,
2316 unsigned long chunk_nr)
2318 compat_uptr_t __user *pages32 = (compat_uptr_t __user *)pages;
2322 for (i = 0; i < chunk_nr; i++) {
2323 if (get_user(p, pages32 + i))
2325 chunk_pages[i] = compat_ptr(p);
2332 * Determine the nodes of a user array of pages and store it in
2333 * a user array of status.
2335 static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages,
2336 const void __user * __user *pages,
2339 #define DO_PAGES_STAT_CHUNK_NR 16UL
2340 const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR];
2341 int chunk_status[DO_PAGES_STAT_CHUNK_NR];
2344 unsigned long chunk_nr = min(nr_pages, DO_PAGES_STAT_CHUNK_NR);
2346 if (in_compat_syscall()) {
2347 if (get_compat_pages_array(chunk_pages, pages,
2351 if (copy_from_user(chunk_pages, pages,
2352 chunk_nr * sizeof(*chunk_pages)))
2356 do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status);
2358 if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status)))
2363 nr_pages -= chunk_nr;
2365 return nr_pages ? -EFAULT : 0;
2368 static struct mm_struct *find_mm_struct(pid_t pid, nodemask_t *mem_nodes)
2370 struct task_struct *task;
2371 struct mm_struct *mm;
2374 * There is no need to check if current process has the right to modify
2375 * the specified process when they are same.
2379 *mem_nodes = cpuset_mems_allowed(current);
2383 /* Find the mm_struct */
2385 task = find_task_by_vpid(pid);
2388 return ERR_PTR(-ESRCH);
2390 get_task_struct(task);
2393 * Check if this process has the right to modify the specified
2394 * process. Use the regular "ptrace_may_access()" checks.
2396 if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) {
2398 mm = ERR_PTR(-EPERM);
2403 mm = ERR_PTR(security_task_movememory(task));
2406 *mem_nodes = cpuset_mems_allowed(task);
2407 mm = get_task_mm(task);
2409 put_task_struct(task);
2411 mm = ERR_PTR(-EINVAL);
2416 * Move a list of pages in the address space of the currently executing
2419 static int kernel_move_pages(pid_t pid, unsigned long nr_pages,
2420 const void __user * __user *pages,
2421 const int __user *nodes,
2422 int __user *status, int flags)
2424 struct mm_struct *mm;
2426 nodemask_t task_nodes;
2429 if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
2432 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
2435 mm = find_mm_struct(pid, &task_nodes);
2440 err = do_pages_move(mm, task_nodes, nr_pages, pages,
2441 nodes, status, flags);
2443 err = do_pages_stat(mm, nr_pages, pages, status);
2449 SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages,
2450 const void __user * __user *, pages,
2451 const int __user *, nodes,
2452 int __user *, status, int, flags)
2454 return kernel_move_pages(pid, nr_pages, pages, nodes, status, flags);
2457 #ifdef CONFIG_NUMA_BALANCING
2459 * Returns true if this is a safe migration target node for misplaced NUMA
2460 * pages. Currently it only checks the watermarks which is crude.
2462 static bool migrate_balanced_pgdat(struct pglist_data *pgdat,
2463 unsigned long nr_migrate_pages)
2467 for (z = pgdat->nr_zones - 1; z >= 0; z--) {
2468 struct zone *zone = pgdat->node_zones + z;
2470 if (!managed_zone(zone))
2473 /* Avoid waking kswapd by allocating pages_to_migrate pages. */
2474 if (!zone_watermark_ok(zone, 0,
2475 high_wmark_pages(zone) +
2484 static struct folio *alloc_misplaced_dst_folio(struct folio *src,
2487 int nid = (int) data;
2488 int order = folio_order(src);
2489 gfp_t gfp = __GFP_THISNODE;
2492 gfp |= GFP_TRANSHUGE_LIGHT;
2494 gfp |= GFP_HIGHUSER_MOVABLE | __GFP_NOMEMALLOC | __GFP_NORETRY |
2496 gfp &= ~__GFP_RECLAIM;
2498 return __folio_alloc_node(gfp, order, nid);
2501 static int numamigrate_isolate_page(pg_data_t *pgdat, struct page *page)
2503 int nr_pages = thp_nr_pages(page);
2504 int order = compound_order(page);
2506 VM_BUG_ON_PAGE(order && !PageTransHuge(page), page);
2508 /* Do not migrate THP mapped by multiple processes */
2509 if (PageTransHuge(page) && total_mapcount(page) > 1)
2512 /* Avoid migrating to a node that is nearly full */
2513 if (!migrate_balanced_pgdat(pgdat, nr_pages)) {
2516 if (!(sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING))
2518 for (z = pgdat->nr_zones - 1; z >= 0; z--) {
2519 if (managed_zone(pgdat->node_zones + z))
2522 wakeup_kswapd(pgdat->node_zones + z, 0, order, ZONE_MOVABLE);
2526 if (!isolate_lru_page(page))
2529 mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON + page_is_file_lru(page),
2533 * Isolating the page has taken another reference, so the
2534 * caller's reference can be safely dropped without the page
2535 * disappearing underneath us during migration.
2542 * Attempt to migrate a misplaced page to the specified destination
2543 * node. Caller is expected to have an elevated reference count on
2544 * the page that will be dropped by this function before returning.
2546 int migrate_misplaced_page(struct page *page, struct vm_area_struct *vma,
2549 pg_data_t *pgdat = NODE_DATA(node);
2552 unsigned int nr_succeeded;
2553 LIST_HEAD(migratepages);
2554 int nr_pages = thp_nr_pages(page);
2557 * Don't migrate file pages that are mapped in multiple processes
2558 * with execute permissions as they are probably shared libraries.
2560 if (page_mapcount(page) != 1 && page_is_file_lru(page) &&
2561 (vma->vm_flags & VM_EXEC))
2565 * Also do not migrate dirty pages as not all filesystems can move
2566 * dirty pages in MIGRATE_ASYNC mode which is a waste of cycles.
2568 if (page_is_file_lru(page) && PageDirty(page))
2571 isolated = numamigrate_isolate_page(pgdat, page);
2575 list_add(&page->lru, &migratepages);
2576 nr_remaining = migrate_pages(&migratepages, alloc_misplaced_dst_folio,
2577 NULL, node, MIGRATE_ASYNC,
2578 MR_NUMA_MISPLACED, &nr_succeeded);
2580 if (!list_empty(&migratepages)) {
2581 list_del(&page->lru);
2582 mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON +
2583 page_is_file_lru(page), -nr_pages);
2584 putback_lru_page(page);
2589 count_vm_numa_events(NUMA_PAGE_MIGRATE, nr_succeeded);
2590 if (!node_is_toptier(page_to_nid(page)) && node_is_toptier(node))
2591 mod_node_page_state(pgdat, PGPROMOTE_SUCCESS,
2594 BUG_ON(!list_empty(&migratepages));
2601 #endif /* CONFIG_NUMA_BALANCING */
2602 #endif /* CONFIG_NUMA */