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:
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 * __folio_test_movable because LRU folio's mapping cannot
161 * have 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);
215 entry = pte_to_swp_entry(old_pte);
216 if (!is_migration_entry_young(entry))
217 pte = pte_mkold(pte);
218 if (folio_test_dirty(folio) && is_migration_entry_dirty(entry))
219 pte = pte_mkdirty(pte);
220 if (pte_swp_soft_dirty(old_pte))
221 pte = pte_mksoft_dirty(pte);
223 pte = pte_clear_soft_dirty(pte);
225 if (is_writable_migration_entry(entry))
226 pte = pte_mkwrite(pte, vma);
227 else if (pte_swp_uffd_wp(old_pte))
228 pte = pte_mkuffd_wp(pte);
230 if (folio_test_anon(folio) && !is_readable_migration_entry(entry))
231 rmap_flags |= RMAP_EXCLUSIVE;
233 if (unlikely(is_device_private_page(new))) {
235 entry = make_writable_device_private_entry(
238 entry = make_readable_device_private_entry(
240 pte = swp_entry_to_pte(entry);
241 if (pte_swp_soft_dirty(old_pte))
242 pte = pte_swp_mksoft_dirty(pte);
243 if (pte_swp_uffd_wp(old_pte))
244 pte = pte_swp_mkuffd_wp(pte);
247 #ifdef CONFIG_HUGETLB_PAGE
248 if (folio_test_hugetlb(folio)) {
249 struct hstate *h = hstate_vma(vma);
250 unsigned int shift = huge_page_shift(h);
251 unsigned long psize = huge_page_size(h);
253 pte = arch_make_huge_pte(pte, shift, vma->vm_flags);
254 if (folio_test_anon(folio))
255 hugetlb_add_anon_rmap(folio, vma, pvmw.address,
258 hugetlb_add_file_rmap(folio);
259 set_huge_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte,
264 if (folio_test_anon(folio))
265 folio_add_anon_rmap_pte(folio, new, vma,
266 pvmw.address, rmap_flags);
268 folio_add_file_rmap_pte(folio, new, vma);
269 set_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte);
271 if (vma->vm_flags & VM_LOCKED)
274 trace_remove_migration_pte(pvmw.address, pte_val(pte),
275 compound_order(new));
277 /* No need to invalidate - it was non-present before */
278 update_mmu_cache(vma, pvmw.address, pvmw.pte);
285 * Get rid of all migration entries and replace them by
286 * references to the indicated page.
288 void remove_migration_ptes(struct folio *src, struct folio *dst, bool locked)
290 struct rmap_walk_control rwc = {
291 .rmap_one = remove_migration_pte,
296 rmap_walk_locked(dst, &rwc);
298 rmap_walk(dst, &rwc);
302 * Something used the pte of a page under migration. We need to
303 * get to the page and wait until migration is finished.
304 * When we return from this function the fault will be retried.
306 void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
307 unsigned long address)
314 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
318 pte = ptep_get(ptep);
321 if (!is_swap_pte(pte))
324 entry = pte_to_swp_entry(pte);
325 if (!is_migration_entry(entry))
328 migration_entry_wait_on_locked(entry, ptl);
334 #ifdef CONFIG_HUGETLB_PAGE
336 * The vma read lock must be held upon entry. Holding that lock prevents either
337 * the pte or the ptl from being freed.
339 * This function will release the vma lock before returning.
341 void migration_entry_wait_huge(struct vm_area_struct *vma, pte_t *ptep)
343 spinlock_t *ptl = huge_pte_lockptr(hstate_vma(vma), vma->vm_mm, ptep);
346 hugetlb_vma_assert_locked(vma);
348 pte = huge_ptep_get(ptep);
350 if (unlikely(!is_hugetlb_entry_migration(pte))) {
352 hugetlb_vma_unlock_read(vma);
355 * If migration entry existed, safe to release vma lock
356 * here because the pgtable page won't be freed without the
357 * pgtable lock released. See comment right above pgtable
358 * lock release in migration_entry_wait_on_locked().
360 hugetlb_vma_unlock_read(vma);
361 migration_entry_wait_on_locked(pte_to_swp_entry(pte), ptl);
366 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
367 void pmd_migration_entry_wait(struct mm_struct *mm, pmd_t *pmd)
371 ptl = pmd_lock(mm, pmd);
372 if (!is_pmd_migration_entry(*pmd))
374 migration_entry_wait_on_locked(pmd_to_swp_entry(*pmd), ptl);
381 static int folio_expected_refs(struct address_space *mapping,
388 refs += folio_nr_pages(folio);
389 if (folio_test_private(folio))
396 * Replace the page in the mapping.
398 * The number of remaining references must be:
399 * 1 for anonymous pages without a mapping
400 * 2 for pages with a mapping
401 * 3 for pages with a mapping and PagePrivate/PagePrivate2 set.
403 int folio_migrate_mapping(struct address_space *mapping,
404 struct folio *newfolio, struct folio *folio, int extra_count)
406 XA_STATE(xas, &mapping->i_pages, folio_index(folio));
407 struct zone *oldzone, *newzone;
409 int expected_count = folio_expected_refs(mapping, folio) + extra_count;
410 long nr = folio_nr_pages(folio);
414 /* Anonymous page without mapping */
415 if (folio_ref_count(folio) != expected_count)
418 /* No turning back from here */
419 newfolio->index = folio->index;
420 newfolio->mapping = folio->mapping;
421 if (folio_test_swapbacked(folio))
422 __folio_set_swapbacked(newfolio);
424 return MIGRATEPAGE_SUCCESS;
427 oldzone = folio_zone(folio);
428 newzone = folio_zone(newfolio);
431 if (!folio_ref_freeze(folio, expected_count)) {
432 xas_unlock_irq(&xas);
437 * Now we know that no one else is looking at the folio:
438 * no turning back from here.
440 newfolio->index = folio->index;
441 newfolio->mapping = folio->mapping;
442 folio_ref_add(newfolio, nr); /* add cache reference */
443 if (folio_test_swapbacked(folio)) {
444 __folio_set_swapbacked(newfolio);
445 if (folio_test_swapcache(folio)) {
446 folio_set_swapcache(newfolio);
447 newfolio->private = folio_get_private(folio);
451 VM_BUG_ON_FOLIO(folio_test_swapcache(folio), folio);
455 /* Move dirty while page refs frozen and newpage not yet exposed */
456 dirty = folio_test_dirty(folio);
458 folio_clear_dirty(folio);
459 folio_set_dirty(newfolio);
462 /* Swap cache still stores N entries instead of a high-order entry */
463 for (i = 0; i < entries; i++) {
464 xas_store(&xas, newfolio);
469 * Drop cache reference from old page by unfreezing
470 * to one less reference.
471 * We know this isn't the last reference.
473 folio_ref_unfreeze(folio, expected_count - nr);
476 /* Leave irq disabled to prevent preemption while updating stats */
479 * If moved to a different zone then also account
480 * the page for that zone. Other VM counters will be
481 * taken care of when we establish references to the
482 * new page and drop references to the old page.
484 * Note that anonymous pages are accounted for
485 * via NR_FILE_PAGES and NR_ANON_MAPPED if they
486 * are mapped to swap space.
488 if (newzone != oldzone) {
489 struct lruvec *old_lruvec, *new_lruvec;
490 struct mem_cgroup *memcg;
492 memcg = folio_memcg(folio);
493 old_lruvec = mem_cgroup_lruvec(memcg, oldzone->zone_pgdat);
494 new_lruvec = mem_cgroup_lruvec(memcg, newzone->zone_pgdat);
496 __mod_lruvec_state(old_lruvec, NR_FILE_PAGES, -nr);
497 __mod_lruvec_state(new_lruvec, NR_FILE_PAGES, nr);
498 if (folio_test_swapbacked(folio) && !folio_test_swapcache(folio)) {
499 __mod_lruvec_state(old_lruvec, NR_SHMEM, -nr);
500 __mod_lruvec_state(new_lruvec, NR_SHMEM, nr);
502 if (folio_test_pmd_mappable(folio)) {
503 __mod_lruvec_state(old_lruvec, NR_SHMEM_THPS, -nr);
504 __mod_lruvec_state(new_lruvec, NR_SHMEM_THPS, nr);
508 if (folio_test_swapcache(folio)) {
509 __mod_lruvec_state(old_lruvec, NR_SWAPCACHE, -nr);
510 __mod_lruvec_state(new_lruvec, NR_SWAPCACHE, nr);
513 if (dirty && mapping_can_writeback(mapping)) {
514 __mod_lruvec_state(old_lruvec, NR_FILE_DIRTY, -nr);
515 __mod_zone_page_state(oldzone, NR_ZONE_WRITE_PENDING, -nr);
516 __mod_lruvec_state(new_lruvec, NR_FILE_DIRTY, nr);
517 __mod_zone_page_state(newzone, NR_ZONE_WRITE_PENDING, nr);
522 return MIGRATEPAGE_SUCCESS;
524 EXPORT_SYMBOL(folio_migrate_mapping);
527 * The expected number of remaining references is the same as that
528 * of folio_migrate_mapping().
530 int migrate_huge_page_move_mapping(struct address_space *mapping,
531 struct folio *dst, struct folio *src)
533 XA_STATE(xas, &mapping->i_pages, folio_index(src));
537 expected_count = folio_expected_refs(mapping, src);
538 if (!folio_ref_freeze(src, expected_count)) {
539 xas_unlock_irq(&xas);
543 dst->index = src->index;
544 dst->mapping = src->mapping;
546 folio_ref_add(dst, folio_nr_pages(dst));
548 xas_store(&xas, dst);
550 folio_ref_unfreeze(src, expected_count - folio_nr_pages(src));
552 xas_unlock_irq(&xas);
554 return MIGRATEPAGE_SUCCESS;
558 * Copy the flags and some other ancillary information
560 void folio_migrate_flags(struct folio *newfolio, struct folio *folio)
564 if (folio_test_error(folio))
565 folio_set_error(newfolio);
566 if (folio_test_referenced(folio))
567 folio_set_referenced(newfolio);
568 if (folio_test_uptodate(folio))
569 folio_mark_uptodate(newfolio);
570 if (folio_test_clear_active(folio)) {
571 VM_BUG_ON_FOLIO(folio_test_unevictable(folio), folio);
572 folio_set_active(newfolio);
573 } else if (folio_test_clear_unevictable(folio))
574 folio_set_unevictable(newfolio);
575 if (folio_test_workingset(folio))
576 folio_set_workingset(newfolio);
577 if (folio_test_checked(folio))
578 folio_set_checked(newfolio);
580 * PG_anon_exclusive (-> PG_mappedtodisk) is always migrated via
581 * migration entries. We can still have PG_anon_exclusive set on an
582 * effectively unmapped and unreferenced first sub-pages of an
583 * anonymous THP: we can simply copy it here via PG_mappedtodisk.
585 if (folio_test_mappedtodisk(folio))
586 folio_set_mappedtodisk(newfolio);
588 /* Move dirty on pages not done by folio_migrate_mapping() */
589 if (folio_test_dirty(folio))
590 folio_set_dirty(newfolio);
592 if (folio_test_young(folio))
593 folio_set_young(newfolio);
594 if (folio_test_idle(folio))
595 folio_set_idle(newfolio);
598 * Copy NUMA information to the new page, to prevent over-eager
599 * future migrations of this same page.
601 cpupid = folio_xchg_last_cpupid(folio, -1);
603 * For memory tiering mode, when migrate between slow and fast
604 * memory node, reset cpupid, because that is used to record
605 * page access time in slow memory node.
607 if (sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING) {
608 bool f_toptier = node_is_toptier(folio_nid(folio));
609 bool t_toptier = node_is_toptier(folio_nid(newfolio));
611 if (f_toptier != t_toptier)
614 folio_xchg_last_cpupid(newfolio, cpupid);
616 folio_migrate_ksm(newfolio, folio);
618 * Please do not reorder this without considering how mm/ksm.c's
619 * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache().
621 if (folio_test_swapcache(folio))
622 folio_clear_swapcache(folio);
623 folio_clear_private(folio);
625 /* page->private contains hugetlb specific flags */
626 if (!folio_test_hugetlb(folio))
627 folio->private = NULL;
630 * If any waiters have accumulated on the new page then
633 if (folio_test_writeback(newfolio))
634 folio_end_writeback(newfolio);
637 * PG_readahead shares the same bit with PG_reclaim. The above
638 * end_page_writeback() may clear PG_readahead mistakenly, so set the
641 if (folio_test_readahead(folio))
642 folio_set_readahead(newfolio);
644 folio_copy_owner(newfolio, folio);
646 mem_cgroup_migrate(folio, newfolio);
648 EXPORT_SYMBOL(folio_migrate_flags);
650 void folio_migrate_copy(struct folio *newfolio, struct folio *folio)
652 folio_copy(newfolio, folio);
653 folio_migrate_flags(newfolio, folio);
655 EXPORT_SYMBOL(folio_migrate_copy);
657 /************************************************************
658 * Migration functions
659 ***********************************************************/
661 int migrate_folio_extra(struct address_space *mapping, struct folio *dst,
662 struct folio *src, enum migrate_mode mode, int extra_count)
666 BUG_ON(folio_test_writeback(src)); /* Writeback must be complete */
668 rc = folio_migrate_mapping(mapping, dst, src, extra_count);
670 if (rc != MIGRATEPAGE_SUCCESS)
673 if (mode != MIGRATE_SYNC_NO_COPY)
674 folio_migrate_copy(dst, src);
676 folio_migrate_flags(dst, src);
677 return MIGRATEPAGE_SUCCESS;
681 * migrate_folio() - Simple folio migration.
682 * @mapping: The address_space containing the folio.
683 * @dst: The folio to migrate the data to.
684 * @src: The folio containing the current data.
685 * @mode: How to migrate the page.
687 * Common logic to directly migrate a single LRU folio suitable for
688 * folios that do not use PagePrivate/PagePrivate2.
690 * Folios are locked upon entry and exit.
692 int migrate_folio(struct address_space *mapping, struct folio *dst,
693 struct folio *src, enum migrate_mode mode)
695 return migrate_folio_extra(mapping, dst, src, mode, 0);
697 EXPORT_SYMBOL(migrate_folio);
699 #ifdef CONFIG_BUFFER_HEAD
700 /* Returns true if all buffers are successfully locked */
701 static bool buffer_migrate_lock_buffers(struct buffer_head *head,
702 enum migrate_mode mode)
704 struct buffer_head *bh = head;
705 struct buffer_head *failed_bh;
708 if (!trylock_buffer(bh)) {
709 if (mode == MIGRATE_ASYNC)
711 if (mode == MIGRATE_SYNC_LIGHT && !buffer_uptodate(bh))
716 bh = bh->b_this_page;
717 } while (bh != head);
722 /* We failed to lock the buffer and cannot stall. */
725 while (bh != failed_bh) {
727 bh = bh->b_this_page;
733 static int __buffer_migrate_folio(struct address_space *mapping,
734 struct folio *dst, struct folio *src, enum migrate_mode mode,
737 struct buffer_head *bh, *head;
741 head = folio_buffers(src);
743 return migrate_folio(mapping, dst, src, mode);
745 /* Check whether page does not have extra refs before we do more work */
746 expected_count = folio_expected_refs(mapping, src);
747 if (folio_ref_count(src) != expected_count)
750 if (!buffer_migrate_lock_buffers(head, mode))
755 bool invalidated = false;
759 spin_lock(&mapping->i_private_lock);
762 if (atomic_read(&bh->b_count)) {
766 bh = bh->b_this_page;
767 } while (bh != head);
773 spin_unlock(&mapping->i_private_lock);
774 invalidate_bh_lrus();
776 goto recheck_buffers;
780 rc = folio_migrate_mapping(mapping, dst, src, 0);
781 if (rc != MIGRATEPAGE_SUCCESS)
784 folio_attach_private(dst, folio_detach_private(src));
788 folio_set_bh(bh, dst, bh_offset(bh));
789 bh = bh->b_this_page;
790 } while (bh != head);
792 if (mode != MIGRATE_SYNC_NO_COPY)
793 folio_migrate_copy(dst, src);
795 folio_migrate_flags(dst, src);
797 rc = MIGRATEPAGE_SUCCESS;
800 spin_unlock(&mapping->i_private_lock);
804 bh = bh->b_this_page;
805 } while (bh != head);
811 * buffer_migrate_folio() - Migration function for folios with buffers.
812 * @mapping: The address space containing @src.
813 * @dst: The folio to migrate to.
814 * @src: The folio to migrate from.
815 * @mode: How to migrate the folio.
817 * This function can only be used if the underlying filesystem guarantees
818 * that no other references to @src exist. For example attached buffer
819 * heads are accessed only under the folio lock. If your filesystem cannot
820 * provide this guarantee, buffer_migrate_folio_norefs() may be more
823 * Return: 0 on success or a negative errno on failure.
825 int buffer_migrate_folio(struct address_space *mapping,
826 struct folio *dst, struct folio *src, enum migrate_mode mode)
828 return __buffer_migrate_folio(mapping, dst, src, mode, false);
830 EXPORT_SYMBOL(buffer_migrate_folio);
833 * buffer_migrate_folio_norefs() - Migration function for folios with buffers.
834 * @mapping: The address space containing @src.
835 * @dst: The folio to migrate to.
836 * @src: The folio to migrate from.
837 * @mode: How to migrate the folio.
839 * Like buffer_migrate_folio() except that this variant is more careful
840 * and checks that there are also no buffer head references. This function
841 * is the right one for mappings where buffer heads are directly looked
842 * up and referenced (such as block device mappings).
844 * Return: 0 on success or a negative errno on failure.
846 int buffer_migrate_folio_norefs(struct address_space *mapping,
847 struct folio *dst, struct folio *src, enum migrate_mode mode)
849 return __buffer_migrate_folio(mapping, dst, src, mode, true);
851 EXPORT_SYMBOL_GPL(buffer_migrate_folio_norefs);
852 #endif /* CONFIG_BUFFER_HEAD */
854 int filemap_migrate_folio(struct address_space *mapping,
855 struct folio *dst, struct folio *src, enum migrate_mode mode)
859 ret = folio_migrate_mapping(mapping, dst, src, 0);
860 if (ret != MIGRATEPAGE_SUCCESS)
863 if (folio_get_private(src))
864 folio_attach_private(dst, folio_detach_private(src));
866 if (mode != MIGRATE_SYNC_NO_COPY)
867 folio_migrate_copy(dst, src);
869 folio_migrate_flags(dst, src);
870 return MIGRATEPAGE_SUCCESS;
872 EXPORT_SYMBOL_GPL(filemap_migrate_folio);
875 * Writeback a folio to clean the dirty state
877 static int writeout(struct address_space *mapping, struct folio *folio)
879 struct writeback_control wbc = {
880 .sync_mode = WB_SYNC_NONE,
883 .range_end = LLONG_MAX,
888 if (!mapping->a_ops->writepage)
889 /* No write method for the address space */
892 if (!folio_clear_dirty_for_io(folio))
893 /* Someone else already triggered a write */
897 * A dirty folio may imply that the underlying filesystem has
898 * the folio on some queue. So the folio must be clean for
899 * migration. Writeout may mean we lose the lock and the
900 * folio state is no longer what we checked for earlier.
901 * At this point we know that the migration attempt cannot
904 remove_migration_ptes(folio, folio, false);
906 rc = mapping->a_ops->writepage(&folio->page, &wbc);
908 if (rc != AOP_WRITEPAGE_ACTIVATE)
909 /* unlocked. Relock */
912 return (rc < 0) ? -EIO : -EAGAIN;
916 * Default handling if a filesystem does not provide a migration function.
918 static int fallback_migrate_folio(struct address_space *mapping,
919 struct folio *dst, struct folio *src, enum migrate_mode mode)
921 if (folio_test_dirty(src)) {
922 /* Only writeback folios in full synchronous migration */
925 case MIGRATE_SYNC_NO_COPY:
930 return writeout(mapping, src);
934 * Buffers may be managed in a filesystem specific way.
935 * We must have no buffers or drop them.
937 if (!filemap_release_folio(src, GFP_KERNEL))
938 return mode == MIGRATE_SYNC ? -EAGAIN : -EBUSY;
940 return migrate_folio(mapping, dst, src, mode);
944 * Move a page to a newly allocated page
945 * The page is locked and all ptes have been successfully removed.
947 * The new page will have replaced the old page if this function
952 * MIGRATEPAGE_SUCCESS - success
954 static int move_to_new_folio(struct folio *dst, struct folio *src,
955 enum migrate_mode mode)
958 bool is_lru = !__folio_test_movable(src);
960 VM_BUG_ON_FOLIO(!folio_test_locked(src), src);
961 VM_BUG_ON_FOLIO(!folio_test_locked(dst), dst);
963 if (likely(is_lru)) {
964 struct address_space *mapping = folio_mapping(src);
967 rc = migrate_folio(mapping, dst, src, mode);
968 else if (mapping_unmovable(mapping))
970 else if (mapping->a_ops->migrate_folio)
972 * Most folios have a mapping and most filesystems
973 * provide a migrate_folio callback. Anonymous folios
974 * are part of swap space which also has its own
975 * migrate_folio callback. This is the most common path
976 * for page migration.
978 rc = mapping->a_ops->migrate_folio(mapping, dst, src,
981 rc = fallback_migrate_folio(mapping, dst, src, mode);
983 const struct movable_operations *mops;
986 * In case of non-lru page, it could be released after
987 * isolation step. In that case, we shouldn't try migration.
989 VM_BUG_ON_FOLIO(!folio_test_isolated(src), src);
990 if (!folio_test_movable(src)) {
991 rc = MIGRATEPAGE_SUCCESS;
992 folio_clear_isolated(src);
996 mops = folio_movable_ops(src);
997 rc = mops->migrate_page(&dst->page, &src->page, mode);
998 WARN_ON_ONCE(rc == MIGRATEPAGE_SUCCESS &&
999 !folio_test_isolated(src));
1003 * When successful, old pagecache src->mapping must be cleared before
1004 * src is freed; but stats require that PageAnon be left as PageAnon.
1006 if (rc == MIGRATEPAGE_SUCCESS) {
1007 if (__folio_test_movable(src)) {
1008 VM_BUG_ON_FOLIO(!folio_test_isolated(src), src);
1011 * We clear PG_movable under page_lock so any compactor
1012 * cannot try to migrate this page.
1014 folio_clear_isolated(src);
1018 * Anonymous and movable src->mapping will be cleared by
1019 * free_pages_prepare so don't reset it here for keeping
1020 * the type to work PageAnon, for example.
1022 if (!folio_mapping_flags(src))
1023 src->mapping = NULL;
1025 if (likely(!folio_is_zone_device(dst)))
1026 flush_dcache_folio(dst);
1033 * To record some information during migration, we use unused private
1034 * field of struct folio of the newly allocated destination folio.
1035 * This is safe because nobody is using it except us.
1038 PAGE_WAS_MAPPED = BIT(0),
1039 PAGE_WAS_MLOCKED = BIT(1),
1040 PAGE_OLD_STATES = PAGE_WAS_MAPPED | PAGE_WAS_MLOCKED,
1043 static void __migrate_folio_record(struct folio *dst,
1045 struct anon_vma *anon_vma)
1047 dst->private = (void *)anon_vma + old_page_state;
1050 static void __migrate_folio_extract(struct folio *dst,
1051 int *old_page_state,
1052 struct anon_vma **anon_vmap)
1054 unsigned long private = (unsigned long)dst->private;
1056 *anon_vmap = (struct anon_vma *)(private & ~PAGE_OLD_STATES);
1057 *old_page_state = private & PAGE_OLD_STATES;
1058 dst->private = NULL;
1061 /* Restore the source folio to the original state upon failure */
1062 static void migrate_folio_undo_src(struct folio *src,
1063 int page_was_mapped,
1064 struct anon_vma *anon_vma,
1066 struct list_head *ret)
1068 if (page_was_mapped)
1069 remove_migration_ptes(src, src, false);
1070 /* Drop an anon_vma reference if we took one */
1072 put_anon_vma(anon_vma);
1076 list_move_tail(&src->lru, ret);
1079 /* Restore the destination folio to the original state upon failure */
1080 static void migrate_folio_undo_dst(struct folio *dst, bool locked,
1081 free_folio_t put_new_folio, unsigned long private)
1086 put_new_folio(dst, private);
1091 /* Cleanup src folio upon migration success */
1092 static void migrate_folio_done(struct folio *src,
1093 enum migrate_reason reason)
1096 * Compaction can migrate also non-LRU pages which are
1097 * not accounted to NR_ISOLATED_*. They can be recognized
1098 * as __folio_test_movable
1100 if (likely(!__folio_test_movable(src)))
1101 mod_node_page_state(folio_pgdat(src), NR_ISOLATED_ANON +
1102 folio_is_file_lru(src), -folio_nr_pages(src));
1104 if (reason != MR_MEMORY_FAILURE)
1105 /* We release the page in page_handle_poison. */
1109 /* Obtain the lock on page, remove all ptes. */
1110 static int migrate_folio_unmap(new_folio_t get_new_folio,
1111 free_folio_t put_new_folio, unsigned long private,
1112 struct folio *src, struct folio **dstp, enum migrate_mode mode,
1113 enum migrate_reason reason, struct list_head *ret)
1117 int old_page_state = 0;
1118 struct anon_vma *anon_vma = NULL;
1119 bool is_lru = !__folio_test_movable(src);
1120 bool locked = false;
1121 bool dst_locked = false;
1123 if (folio_ref_count(src) == 1) {
1124 /* Folio was freed from under us. So we are done. */
1125 folio_clear_active(src);
1126 folio_clear_unevictable(src);
1127 /* free_pages_prepare() will clear PG_isolated. */
1128 list_del(&src->lru);
1129 migrate_folio_done(src, reason);
1130 return MIGRATEPAGE_SUCCESS;
1133 dst = get_new_folio(src, private);
1138 dst->private = NULL;
1140 if (!folio_trylock(src)) {
1141 if (mode == MIGRATE_ASYNC)
1145 * It's not safe for direct compaction to call lock_page.
1146 * For example, during page readahead pages are added locked
1147 * to the LRU. Later, when the IO completes the pages are
1148 * marked uptodate and unlocked. However, the queueing
1149 * could be merging multiple pages for one bio (e.g.
1150 * mpage_readahead). If an allocation happens for the
1151 * second or third page, the process can end up locking
1152 * the same page twice and deadlocking. Rather than
1153 * trying to be clever about what pages can be locked,
1154 * avoid the use of lock_page for direct compaction
1157 if (current->flags & PF_MEMALLOC)
1161 * In "light" mode, we can wait for transient locks (eg
1162 * inserting a page into the page table), but it's not
1163 * worth waiting for I/O.
1165 if (mode == MIGRATE_SYNC_LIGHT && !folio_test_uptodate(src))
1171 if (folio_test_mlocked(src))
1172 old_page_state |= PAGE_WAS_MLOCKED;
1174 if (folio_test_writeback(src)) {
1176 * Only in the case of a full synchronous migration is it
1177 * necessary to wait for PageWriteback. In the async case,
1178 * the retry loop is too short and in the sync-light case,
1179 * the overhead of stalling is too much
1183 case MIGRATE_SYNC_NO_COPY:
1189 folio_wait_writeback(src);
1193 * By try_to_migrate(), src->mapcount goes down to 0 here. In this case,
1194 * we cannot notice that anon_vma is freed while we migrate a page.
1195 * This get_anon_vma() delays freeing anon_vma pointer until the end
1196 * of migration. File cache pages are no problem because of page_lock()
1197 * File Caches may use write_page() or lock_page() in migration, then,
1198 * just care Anon page here.
1200 * Only folio_get_anon_vma() understands the subtleties of
1201 * getting a hold on an anon_vma from outside one of its mms.
1202 * But if we cannot get anon_vma, then we won't need it anyway,
1203 * because that implies that the anon page is no longer mapped
1204 * (and cannot be remapped so long as we hold the page lock).
1206 if (folio_test_anon(src) && !folio_test_ksm(src))
1207 anon_vma = folio_get_anon_vma(src);
1210 * Block others from accessing the new page when we get around to
1211 * establishing additional references. We are usually the only one
1212 * holding a reference to dst at this point. We used to have a BUG
1213 * here if folio_trylock(dst) fails, but would like to allow for
1214 * cases where there might be a race with the previous use of dst.
1215 * This is much like races on refcount of oldpage: just don't BUG().
1217 if (unlikely(!folio_trylock(dst)))
1221 if (unlikely(!is_lru)) {
1222 __migrate_folio_record(dst, old_page_state, anon_vma);
1223 return MIGRATEPAGE_UNMAP;
1227 * Corner case handling:
1228 * 1. When a new swap-cache page is read into, it is added to the LRU
1229 * and treated as swapcache but it has no rmap yet.
1230 * Calling try_to_unmap() against a src->mapping==NULL page will
1231 * trigger a BUG. So handle it here.
1232 * 2. An orphaned page (see truncate_cleanup_page) might have
1233 * fs-private metadata. The page can be picked up due to memory
1234 * offlining. Everywhere else except page reclaim, the page is
1235 * invisible to the vm, so the page can not be migrated. So try to
1236 * free the metadata, so the page can be freed.
1238 if (!src->mapping) {
1239 if (folio_test_private(src)) {
1240 try_to_free_buffers(src);
1243 } else if (folio_mapped(src)) {
1244 /* Establish migration ptes */
1245 VM_BUG_ON_FOLIO(folio_test_anon(src) &&
1246 !folio_test_ksm(src) && !anon_vma, src);
1247 try_to_migrate(src, mode == MIGRATE_ASYNC ? TTU_BATCH_FLUSH : 0);
1248 old_page_state |= PAGE_WAS_MAPPED;
1251 if (!folio_mapped(src)) {
1252 __migrate_folio_record(dst, old_page_state, anon_vma);
1253 return MIGRATEPAGE_UNMAP;
1258 * A folio that has not been unmapped will be restored to
1259 * right list unless we want to retry.
1264 migrate_folio_undo_src(src, old_page_state & PAGE_WAS_MAPPED,
1265 anon_vma, locked, ret);
1266 migrate_folio_undo_dst(dst, dst_locked, put_new_folio, private);
1271 /* Migrate the folio to the newly allocated folio in dst. */
1272 static int migrate_folio_move(free_folio_t put_new_folio, unsigned long private,
1273 struct folio *src, struct folio *dst,
1274 enum migrate_mode mode, enum migrate_reason reason,
1275 struct list_head *ret)
1278 int old_page_state = 0;
1279 struct anon_vma *anon_vma = NULL;
1280 bool is_lru = !__folio_test_movable(src);
1281 struct list_head *prev;
1283 __migrate_folio_extract(dst, &old_page_state, &anon_vma);
1284 prev = dst->lru.prev;
1285 list_del(&dst->lru);
1287 rc = move_to_new_folio(dst, src, mode);
1291 if (unlikely(!is_lru))
1292 goto out_unlock_both;
1295 * When successful, push dst to LRU immediately: so that if it
1296 * turns out to be an mlocked page, remove_migration_ptes() will
1297 * automatically build up the correct dst->mlock_count for it.
1299 * We would like to do something similar for the old page, when
1300 * unsuccessful, and other cases when a page has been temporarily
1301 * isolated from the unevictable LRU: but this case is the easiest.
1304 if (old_page_state & PAGE_WAS_MLOCKED)
1307 if (old_page_state & PAGE_WAS_MAPPED)
1308 remove_migration_ptes(src, dst, false);
1312 set_page_owner_migrate_reason(&dst->page, reason);
1314 * If migration is successful, decrease refcount of dst,
1315 * which will not free the page because new page owner increased
1321 * A folio that has been migrated has all references removed
1322 * and will be freed.
1324 list_del(&src->lru);
1325 /* Drop an anon_vma reference if we took one */
1327 put_anon_vma(anon_vma);
1329 migrate_folio_done(src, reason);
1334 * A folio that has not been migrated will be restored to
1335 * right list unless we want to retry.
1337 if (rc == -EAGAIN) {
1338 list_add(&dst->lru, prev);
1339 __migrate_folio_record(dst, old_page_state, anon_vma);
1343 migrate_folio_undo_src(src, old_page_state & PAGE_WAS_MAPPED,
1344 anon_vma, true, ret);
1345 migrate_folio_undo_dst(dst, true, put_new_folio, private);
1351 * Counterpart of unmap_and_move_page() for hugepage migration.
1353 * This function doesn't wait the completion of hugepage I/O
1354 * because there is no race between I/O and migration for hugepage.
1355 * Note that currently hugepage I/O occurs only in direct I/O
1356 * where no lock is held and PG_writeback is irrelevant,
1357 * and writeback status of all subpages are counted in the reference
1358 * count of the head page (i.e. if all subpages of a 2MB hugepage are
1359 * under direct I/O, the reference of the head page is 512 and a bit more.)
1360 * This means that when we try to migrate hugepage whose subpages are
1361 * doing direct I/O, some references remain after try_to_unmap() and
1362 * hugepage migration fails without data corruption.
1364 * There is also no race when direct I/O is issued on the page under migration,
1365 * because then pte is replaced with migration swap entry and direct I/O code
1366 * will wait in the page fault for migration to complete.
1368 static int unmap_and_move_huge_page(new_folio_t get_new_folio,
1369 free_folio_t put_new_folio, unsigned long private,
1370 struct folio *src, int force, enum migrate_mode mode,
1371 int reason, struct list_head *ret)
1375 int page_was_mapped = 0;
1376 struct anon_vma *anon_vma = NULL;
1377 struct address_space *mapping = NULL;
1379 if (folio_ref_count(src) == 1) {
1380 /* page was freed from under us. So we are done. */
1381 folio_putback_active_hugetlb(src);
1382 return MIGRATEPAGE_SUCCESS;
1385 dst = get_new_folio(src, private);
1389 if (!folio_trylock(src)) {
1394 case MIGRATE_SYNC_NO_COPY:
1403 * Check for pages which are in the process of being freed. Without
1404 * folio_mapping() set, hugetlbfs specific move page routine will not
1405 * be called and we could leak usage counts for subpools.
1407 if (hugetlb_folio_subpool(src) && !folio_mapping(src)) {
1412 if (folio_test_anon(src))
1413 anon_vma = folio_get_anon_vma(src);
1415 if (unlikely(!folio_trylock(dst)))
1418 if (folio_mapped(src)) {
1419 enum ttu_flags ttu = 0;
1421 if (!folio_test_anon(src)) {
1423 * In shared mappings, try_to_unmap could potentially
1424 * call huge_pmd_unshare. Because of this, take
1425 * semaphore in write mode here and set TTU_RMAP_LOCKED
1426 * to let lower levels know we have taken the lock.
1428 mapping = hugetlb_page_mapping_lock_write(&src->page);
1429 if (unlikely(!mapping))
1430 goto unlock_put_anon;
1432 ttu = TTU_RMAP_LOCKED;
1435 try_to_migrate(src, ttu);
1436 page_was_mapped = 1;
1438 if (ttu & TTU_RMAP_LOCKED)
1439 i_mmap_unlock_write(mapping);
1442 if (!folio_mapped(src))
1443 rc = move_to_new_folio(dst, src, mode);
1445 if (page_was_mapped)
1446 remove_migration_ptes(src,
1447 rc == MIGRATEPAGE_SUCCESS ? dst : src, false);
1454 put_anon_vma(anon_vma);
1456 if (rc == MIGRATEPAGE_SUCCESS) {
1457 move_hugetlb_state(src, dst, reason);
1458 put_new_folio = NULL;
1464 if (rc == MIGRATEPAGE_SUCCESS)
1465 folio_putback_active_hugetlb(src);
1466 else if (rc != -EAGAIN)
1467 list_move_tail(&src->lru, ret);
1470 * If migration was not successful and there's a freeing callback, use
1471 * it. Otherwise, put_page() will drop the reference grabbed during
1475 put_new_folio(dst, private);
1477 folio_putback_active_hugetlb(dst);
1482 static inline int try_split_folio(struct folio *folio, struct list_head *split_folios)
1487 rc = split_folio_to_list(folio, split_folios);
1488 folio_unlock(folio);
1490 list_move_tail(&folio->lru, split_folios);
1495 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1496 #define NR_MAX_BATCHED_MIGRATION HPAGE_PMD_NR
1498 #define NR_MAX_BATCHED_MIGRATION 512
1500 #define NR_MAX_MIGRATE_PAGES_RETRY 10
1501 #define NR_MAX_MIGRATE_ASYNC_RETRY 3
1502 #define NR_MAX_MIGRATE_SYNC_RETRY \
1503 (NR_MAX_MIGRATE_PAGES_RETRY - NR_MAX_MIGRATE_ASYNC_RETRY)
1505 struct migrate_pages_stats {
1506 int nr_succeeded; /* Normal and large folios migrated successfully, in
1507 units of base pages */
1508 int nr_failed_pages; /* Normal and large folios failed to be migrated, in
1509 units of base pages. Untried folios aren't counted */
1510 int nr_thp_succeeded; /* THP migrated successfully */
1511 int nr_thp_failed; /* THP failed to be migrated */
1512 int nr_thp_split; /* THP split before migrating */
1513 int nr_split; /* Large folio (include THP) split before migrating */
1517 * Returns the number of hugetlb folios that were not migrated, or an error code
1518 * after NR_MAX_MIGRATE_PAGES_RETRY attempts or if no hugetlb folios are movable
1519 * any more because the list has become empty or no retryable hugetlb folios
1520 * exist any more. It is caller's responsibility to call putback_movable_pages()
1523 static int migrate_hugetlbs(struct list_head *from, new_folio_t get_new_folio,
1524 free_folio_t put_new_folio, unsigned long private,
1525 enum migrate_mode mode, int reason,
1526 struct migrate_pages_stats *stats,
1527 struct list_head *ret_folios)
1531 int nr_retry_pages = 0;
1533 struct folio *folio, *folio2;
1536 for (pass = 0; pass < NR_MAX_MIGRATE_PAGES_RETRY && retry; pass++) {
1540 list_for_each_entry_safe(folio, folio2, from, lru) {
1541 if (!folio_test_hugetlb(folio))
1544 nr_pages = folio_nr_pages(folio);
1549 * Migratability of hugepages depends on architectures and
1550 * their size. This check is necessary because some callers
1551 * of hugepage migration like soft offline and memory
1552 * hotremove don't walk through page tables or check whether
1553 * the hugepage is pmd-based or not before kicking migration.
1555 if (!hugepage_migration_supported(folio_hstate(folio))) {
1557 stats->nr_failed_pages += nr_pages;
1558 list_move_tail(&folio->lru, ret_folios);
1562 rc = unmap_and_move_huge_page(get_new_folio,
1563 put_new_folio, private,
1564 folio, pass > 2, mode,
1565 reason, ret_folios);
1568 * Success: hugetlb folio will be put back
1569 * -EAGAIN: stay on the from list
1570 * -ENOMEM: stay on the from list
1571 * Other errno: put on ret_folios list
1576 * When memory is low, don't bother to try to migrate
1577 * other folios, just exit.
1579 stats->nr_failed_pages += nr_pages + nr_retry_pages;
1583 nr_retry_pages += nr_pages;
1585 case MIGRATEPAGE_SUCCESS:
1586 stats->nr_succeeded += nr_pages;
1590 * Permanent failure (-EBUSY, etc.):
1591 * unlike -EAGAIN case, the failed folio is
1592 * removed from migration folio list and not
1593 * retried in the next outer loop.
1596 stats->nr_failed_pages += nr_pages;
1602 * nr_failed is number of hugetlb folios failed to be migrated. After
1603 * NR_MAX_MIGRATE_PAGES_RETRY attempts, give up and count retried hugetlb
1607 stats->nr_failed_pages += nr_retry_pages;
1613 * migrate_pages_batch() first unmaps folios in the from list as many as
1614 * possible, then move the unmapped folios.
1616 * We only batch migration if mode == MIGRATE_ASYNC to avoid to wait a
1617 * lock or bit when we have locked more than one folio. Which may cause
1618 * deadlock (e.g., for loop device). So, if mode != MIGRATE_ASYNC, the
1619 * length of the from list must be <= 1.
1621 static int migrate_pages_batch(struct list_head *from,
1622 new_folio_t get_new_folio, free_folio_t put_new_folio,
1623 unsigned long private, enum migrate_mode mode, int reason,
1624 struct list_head *ret_folios, struct list_head *split_folios,
1625 struct migrate_pages_stats *stats, int nr_pass)
1630 int nr_retry_pages = 0;
1632 bool is_thp = false;
1633 bool is_large = false;
1634 struct folio *folio, *folio2, *dst = NULL, *dst2;
1635 int rc, rc_saved = 0, nr_pages;
1636 LIST_HEAD(unmap_folios);
1637 LIST_HEAD(dst_folios);
1638 bool nosplit = (reason == MR_NUMA_MISPLACED);
1640 VM_WARN_ON_ONCE(mode != MIGRATE_ASYNC &&
1641 !list_empty(from) && !list_is_singular(from));
1643 for (pass = 0; pass < nr_pass && retry; pass++) {
1648 list_for_each_entry_safe(folio, folio2, from, lru) {
1649 is_large = folio_test_large(folio);
1650 is_thp = is_large && folio_test_pmd_mappable(folio);
1651 nr_pages = folio_nr_pages(folio);
1656 * Large folio migration might be unsupported or
1657 * the allocation might be failed so we should retry
1658 * on the same folio with the large folio split
1661 * Split folios are put in split_folios, and
1662 * we will migrate them after the rest of the
1663 * list is processed.
1665 if (!thp_migration_supported() && is_thp) {
1667 stats->nr_thp_failed++;
1668 if (!try_split_folio(folio, split_folios)) {
1669 stats->nr_thp_split++;
1673 stats->nr_failed_pages += nr_pages;
1674 list_move_tail(&folio->lru, ret_folios);
1678 rc = migrate_folio_unmap(get_new_folio, put_new_folio,
1679 private, folio, &dst, mode, reason,
1683 * Success: folio will be freed
1684 * Unmap: folio will be put on unmap_folios list,
1685 * dst folio put on dst_folios list
1686 * -EAGAIN: stay on the from list
1687 * -ENOMEM: stay on the from list
1688 * Other errno: put on ret_folios list
1693 * When memory is low, don't bother to try to migrate
1694 * other folios, move unmapped folios, then exit.
1697 stats->nr_thp_failed += is_thp;
1698 /* Large folio NUMA faulting doesn't split to retry. */
1699 if (is_large && !nosplit) {
1700 int ret = try_split_folio(folio, split_folios);
1703 stats->nr_thp_split += is_thp;
1706 } else if (reason == MR_LONGTERM_PIN &&
1709 * Try again to split large folio to
1710 * mitigate the failure of longterm pinning.
1713 thp_retry += is_thp;
1714 nr_retry_pages += nr_pages;
1715 /* Undo duplicated failure counting. */
1717 stats->nr_thp_failed -= is_thp;
1722 stats->nr_failed_pages += nr_pages + nr_retry_pages;
1723 /* nr_failed isn't updated for not used */
1724 stats->nr_thp_failed += thp_retry;
1726 if (list_empty(&unmap_folios))
1732 thp_retry += is_thp;
1733 nr_retry_pages += nr_pages;
1735 case MIGRATEPAGE_SUCCESS:
1736 stats->nr_succeeded += nr_pages;
1737 stats->nr_thp_succeeded += is_thp;
1739 case MIGRATEPAGE_UNMAP:
1740 list_move_tail(&folio->lru, &unmap_folios);
1741 list_add_tail(&dst->lru, &dst_folios);
1745 * Permanent failure (-EBUSY, etc.):
1746 * unlike -EAGAIN case, the failed folio is
1747 * removed from migration folio list and not
1748 * retried in the next outer loop.
1751 stats->nr_thp_failed += is_thp;
1752 stats->nr_failed_pages += nr_pages;
1758 stats->nr_thp_failed += thp_retry;
1759 stats->nr_failed_pages += nr_retry_pages;
1761 /* Flush TLBs for all unmapped folios */
1762 try_to_unmap_flush();
1765 for (pass = 0; pass < nr_pass && retry; pass++) {
1770 dst = list_first_entry(&dst_folios, struct folio, lru);
1771 dst2 = list_next_entry(dst, lru);
1772 list_for_each_entry_safe(folio, folio2, &unmap_folios, lru) {
1773 is_thp = folio_test_large(folio) && folio_test_pmd_mappable(folio);
1774 nr_pages = folio_nr_pages(folio);
1778 rc = migrate_folio_move(put_new_folio, private,
1780 reason, ret_folios);
1783 * Success: folio will be freed
1784 * -EAGAIN: stay on the unmap_folios list
1785 * Other errno: put on ret_folios list
1790 thp_retry += is_thp;
1791 nr_retry_pages += nr_pages;
1793 case MIGRATEPAGE_SUCCESS:
1794 stats->nr_succeeded += nr_pages;
1795 stats->nr_thp_succeeded += is_thp;
1799 stats->nr_thp_failed += is_thp;
1800 stats->nr_failed_pages += nr_pages;
1804 dst2 = list_next_entry(dst, lru);
1808 stats->nr_thp_failed += thp_retry;
1809 stats->nr_failed_pages += nr_retry_pages;
1811 rc = rc_saved ? : nr_failed;
1813 /* Cleanup remaining folios */
1814 dst = list_first_entry(&dst_folios, struct folio, lru);
1815 dst2 = list_next_entry(dst, lru);
1816 list_for_each_entry_safe(folio, folio2, &unmap_folios, lru) {
1817 int old_page_state = 0;
1818 struct anon_vma *anon_vma = NULL;
1820 __migrate_folio_extract(dst, &old_page_state, &anon_vma);
1821 migrate_folio_undo_src(folio, old_page_state & PAGE_WAS_MAPPED,
1822 anon_vma, true, ret_folios);
1823 list_del(&dst->lru);
1824 migrate_folio_undo_dst(dst, true, put_new_folio, private);
1826 dst2 = list_next_entry(dst, lru);
1832 static int migrate_pages_sync(struct list_head *from, new_folio_t get_new_folio,
1833 free_folio_t put_new_folio, unsigned long private,
1834 enum migrate_mode mode, int reason,
1835 struct list_head *ret_folios, struct list_head *split_folios,
1836 struct migrate_pages_stats *stats)
1838 int rc, nr_failed = 0;
1840 struct migrate_pages_stats astats;
1842 memset(&astats, 0, sizeof(astats));
1843 /* Try to migrate in batch with MIGRATE_ASYNC mode firstly */
1844 rc = migrate_pages_batch(from, get_new_folio, put_new_folio, private, MIGRATE_ASYNC,
1845 reason, &folios, split_folios, &astats,
1846 NR_MAX_MIGRATE_ASYNC_RETRY);
1847 stats->nr_succeeded += astats.nr_succeeded;
1848 stats->nr_thp_succeeded += astats.nr_thp_succeeded;
1849 stats->nr_thp_split += astats.nr_thp_split;
1850 stats->nr_split += astats.nr_split;
1852 stats->nr_failed_pages += astats.nr_failed_pages;
1853 stats->nr_thp_failed += astats.nr_thp_failed;
1854 list_splice_tail(&folios, ret_folios);
1857 stats->nr_thp_failed += astats.nr_thp_split;
1859 * Do not count rc, as pages will be retried below.
1860 * Count nr_split only, since it includes nr_thp_split.
1862 nr_failed += astats.nr_split;
1864 * Fall back to migrate all failed folios one by one synchronously. All
1865 * failed folios except split THPs will be retried, so their failure
1868 list_splice_tail_init(&folios, from);
1869 while (!list_empty(from)) {
1870 list_move(from->next, &folios);
1871 rc = migrate_pages_batch(&folios, get_new_folio, put_new_folio,
1872 private, mode, reason, ret_folios,
1873 split_folios, stats, NR_MAX_MIGRATE_SYNC_RETRY);
1874 list_splice_tail_init(&folios, ret_folios);
1884 * migrate_pages - migrate the folios specified in a list, to the free folios
1885 * supplied as the target for the page migration
1887 * @from: The list of folios to be migrated.
1888 * @get_new_folio: The function used to allocate free folios to be used
1889 * as the target of the folio migration.
1890 * @put_new_folio: The function used to free target folios if migration
1891 * fails, or NULL if no special handling is necessary.
1892 * @private: Private data to be passed on to get_new_folio()
1893 * @mode: The migration mode that specifies the constraints for
1894 * folio migration, if any.
1895 * @reason: The reason for folio migration.
1896 * @ret_succeeded: Set to the number of folios migrated successfully if
1897 * the caller passes a non-NULL pointer.
1899 * The function returns after NR_MAX_MIGRATE_PAGES_RETRY attempts or if no folios
1900 * are movable any more because the list has become empty or no retryable folios
1901 * exist any more. It is caller's responsibility to call putback_movable_pages()
1904 * Returns the number of {normal folio, large folio, hugetlb} that were not
1905 * migrated, or an error code. The number of large folio splits will be
1906 * considered as the number of non-migrated large folio, no matter how many
1907 * split folios of the large folio are migrated successfully.
1909 int migrate_pages(struct list_head *from, new_folio_t get_new_folio,
1910 free_folio_t put_new_folio, unsigned long private,
1911 enum migrate_mode mode, int reason, unsigned int *ret_succeeded)
1915 struct folio *folio, *folio2;
1917 LIST_HEAD(ret_folios);
1918 LIST_HEAD(split_folios);
1919 struct migrate_pages_stats stats;
1921 trace_mm_migrate_pages_start(mode, reason);
1923 memset(&stats, 0, sizeof(stats));
1925 rc_gather = migrate_hugetlbs(from, get_new_folio, put_new_folio, private,
1926 mode, reason, &stats, &ret_folios);
1932 list_for_each_entry_safe(folio, folio2, from, lru) {
1933 /* Retried hugetlb folios will be kept in list */
1934 if (folio_test_hugetlb(folio)) {
1935 list_move_tail(&folio->lru, &ret_folios);
1939 nr_pages += folio_nr_pages(folio);
1940 if (nr_pages >= NR_MAX_BATCHED_MIGRATION)
1943 if (nr_pages >= NR_MAX_BATCHED_MIGRATION)
1944 list_cut_before(&folios, from, &folio2->lru);
1946 list_splice_init(from, &folios);
1947 if (mode == MIGRATE_ASYNC)
1948 rc = migrate_pages_batch(&folios, get_new_folio, put_new_folio,
1949 private, mode, reason, &ret_folios,
1950 &split_folios, &stats,
1951 NR_MAX_MIGRATE_PAGES_RETRY);
1953 rc = migrate_pages_sync(&folios, get_new_folio, put_new_folio,
1954 private, mode, reason, &ret_folios,
1955 &split_folios, &stats);
1956 list_splice_tail_init(&folios, &ret_folios);
1959 list_splice_tail(&split_folios, &ret_folios);
1962 if (!list_empty(&split_folios)) {
1964 * Failure isn't counted since all split folios of a large folio
1965 * is counted as 1 failure already. And, we only try to migrate
1966 * with minimal effort, force MIGRATE_ASYNC mode and retry once.
1968 migrate_pages_batch(&split_folios, get_new_folio,
1969 put_new_folio, private, MIGRATE_ASYNC, reason,
1970 &ret_folios, NULL, &stats, 1);
1971 list_splice_tail_init(&split_folios, &ret_folios);
1974 if (!list_empty(from))
1978 * Put the permanent failure folio back to migration list, they
1979 * will be put back to the right list by the caller.
1981 list_splice(&ret_folios, from);
1984 * Return 0 in case all split folios of fail-to-migrate large folios
1985 * are migrated successfully.
1987 if (list_empty(from))
1990 count_vm_events(PGMIGRATE_SUCCESS, stats.nr_succeeded);
1991 count_vm_events(PGMIGRATE_FAIL, stats.nr_failed_pages);
1992 count_vm_events(THP_MIGRATION_SUCCESS, stats.nr_thp_succeeded);
1993 count_vm_events(THP_MIGRATION_FAIL, stats.nr_thp_failed);
1994 count_vm_events(THP_MIGRATION_SPLIT, stats.nr_thp_split);
1995 trace_mm_migrate_pages(stats.nr_succeeded, stats.nr_failed_pages,
1996 stats.nr_thp_succeeded, stats.nr_thp_failed,
1997 stats.nr_thp_split, stats.nr_split, mode,
2001 *ret_succeeded = stats.nr_succeeded;
2006 struct folio *alloc_migration_target(struct folio *src, unsigned long private)
2008 struct migration_target_control *mtc;
2010 unsigned int order = 0;
2014 mtc = (struct migration_target_control *)private;
2015 gfp_mask = mtc->gfp_mask;
2017 if (nid == NUMA_NO_NODE)
2018 nid = folio_nid(src);
2020 if (folio_test_hugetlb(src)) {
2021 struct hstate *h = folio_hstate(src);
2023 gfp_mask = htlb_modify_alloc_mask(h, gfp_mask);
2024 return alloc_hugetlb_folio_nodemask(h, nid,
2025 mtc->nmask, gfp_mask);
2028 if (folio_test_large(src)) {
2030 * clear __GFP_RECLAIM to make the migration callback
2031 * consistent with regular THP allocations.
2033 gfp_mask &= ~__GFP_RECLAIM;
2034 gfp_mask |= GFP_TRANSHUGE;
2035 order = folio_order(src);
2037 zidx = zone_idx(folio_zone(src));
2038 if (is_highmem_idx(zidx) || zidx == ZONE_MOVABLE)
2039 gfp_mask |= __GFP_HIGHMEM;
2041 return __folio_alloc(gfp_mask, order, nid, mtc->nmask);
2046 static int store_status(int __user *status, int start, int value, int nr)
2049 if (put_user(value, status + start))
2057 static int do_move_pages_to_node(struct list_head *pagelist, int node)
2060 struct migration_target_control mtc = {
2062 .gfp_mask = GFP_HIGHUSER_MOVABLE | __GFP_THISNODE,
2065 err = migrate_pages(pagelist, alloc_migration_target, NULL,
2066 (unsigned long)&mtc, MIGRATE_SYNC, MR_SYSCALL, NULL);
2068 putback_movable_pages(pagelist);
2073 * Resolves the given address to a struct page, isolates it from the LRU and
2074 * puts it to the given pagelist.
2076 * errno - if the page cannot be found/isolated
2077 * 0 - when it doesn't have to be migrated because it is already on the
2079 * 1 - when it has been queued
2081 static int add_page_for_migration(struct mm_struct *mm, const void __user *p,
2082 int node, struct list_head *pagelist, bool migrate_all)
2084 struct vm_area_struct *vma;
2087 struct folio *folio;
2091 addr = (unsigned long)untagged_addr_remote(mm, p);
2094 vma = vma_lookup(mm, addr);
2095 if (!vma || !vma_migratable(vma))
2098 /* FOLL_DUMP to ignore special (like zero) pages */
2099 page = follow_page(vma, addr, FOLL_GET | FOLL_DUMP);
2101 err = PTR_ERR(page);
2109 folio = page_folio(page);
2110 if (folio_is_zone_device(folio))
2114 if (folio_nid(folio) == node)
2118 if (page_mapcount(page) > 1 && !migrate_all)
2122 if (folio_test_hugetlb(folio)) {
2123 if (isolate_hugetlb(folio, pagelist))
2126 if (!folio_isolate_lru(folio))
2130 list_add_tail(&folio->lru, pagelist);
2131 node_stat_mod_folio(folio,
2132 NR_ISOLATED_ANON + folio_is_file_lru(folio),
2133 folio_nr_pages(folio));
2137 * Either remove the duplicate refcount from folio_isolate_lru()
2138 * or drop the folio ref if it was not isolated.
2142 mmap_read_unlock(mm);
2146 static int move_pages_and_store_status(int node,
2147 struct list_head *pagelist, int __user *status,
2148 int start, int i, unsigned long nr_pages)
2152 if (list_empty(pagelist))
2155 err = do_move_pages_to_node(pagelist, node);
2158 * Positive err means the number of failed
2159 * pages to migrate. Since we are going to
2160 * abort and return the number of non-migrated
2161 * pages, so need to include the rest of the
2162 * nr_pages that have not been attempted as
2166 err += nr_pages - i;
2169 return store_status(status, start, node, i - start);
2173 * Migrate an array of page address onto an array of nodes and fill
2174 * the corresponding array of status.
2176 static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes,
2177 unsigned long nr_pages,
2178 const void __user * __user *pages,
2179 const int __user *nodes,
2180 int __user *status, int flags)
2182 compat_uptr_t __user *compat_pages = (void __user *)pages;
2183 int current_node = NUMA_NO_NODE;
2184 LIST_HEAD(pagelist);
2188 lru_cache_disable();
2190 for (i = start = 0; i < nr_pages; i++) {
2191 const void __user *p;
2195 if (in_compat_syscall()) {
2198 if (get_user(cp, compat_pages + i))
2203 if (get_user(p, pages + i))
2206 if (get_user(node, nodes + i))
2210 if (node < 0 || node >= MAX_NUMNODES)
2212 if (!node_state(node, N_MEMORY))
2216 if (!node_isset(node, task_nodes))
2219 if (current_node == NUMA_NO_NODE) {
2220 current_node = node;
2222 } else if (node != current_node) {
2223 err = move_pages_and_store_status(current_node,
2224 &pagelist, status, start, i, nr_pages);
2228 current_node = node;
2232 * Errors in the page lookup or isolation are not fatal and we simply
2233 * report them via status
2235 err = add_page_for_migration(mm, p, current_node, &pagelist,
2236 flags & MPOL_MF_MOVE_ALL);
2239 /* The page is successfully queued for migration */
2244 * The move_pages() man page does not have an -EEXIST choice, so
2245 * use -EFAULT instead.
2251 * If the page is already on the target node (!err), store the
2252 * node, otherwise, store the err.
2254 err = store_status(status, i, err ? : current_node, 1);
2258 err = move_pages_and_store_status(current_node, &pagelist,
2259 status, start, i, nr_pages);
2261 /* We have accounted for page i */
2266 current_node = NUMA_NO_NODE;
2269 /* Make sure we do not overwrite the existing error */
2270 err1 = move_pages_and_store_status(current_node, &pagelist,
2271 status, start, i, nr_pages);
2280 * Determine the nodes of an array of pages and store it in an array of status.
2282 static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages,
2283 const void __user **pages, int *status)
2289 for (i = 0; i < nr_pages; i++) {
2290 unsigned long addr = (unsigned long)(*pages);
2291 struct vm_area_struct *vma;
2295 vma = vma_lookup(mm, addr);
2299 /* FOLL_DUMP to ignore special (like zero) pages */
2300 page = follow_page(vma, addr, FOLL_GET | FOLL_DUMP);
2302 err = PTR_ERR(page);
2310 if (!is_zone_device_page(page))
2311 err = page_to_nid(page);
2321 mmap_read_unlock(mm);
2324 static int get_compat_pages_array(const void __user *chunk_pages[],
2325 const void __user * __user *pages,
2326 unsigned long chunk_nr)
2328 compat_uptr_t __user *pages32 = (compat_uptr_t __user *)pages;
2332 for (i = 0; i < chunk_nr; i++) {
2333 if (get_user(p, pages32 + i))
2335 chunk_pages[i] = compat_ptr(p);
2342 * Determine the nodes of a user array of pages and store it in
2343 * a user array of status.
2345 static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages,
2346 const void __user * __user *pages,
2349 #define DO_PAGES_STAT_CHUNK_NR 16UL
2350 const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR];
2351 int chunk_status[DO_PAGES_STAT_CHUNK_NR];
2354 unsigned long chunk_nr = min(nr_pages, DO_PAGES_STAT_CHUNK_NR);
2356 if (in_compat_syscall()) {
2357 if (get_compat_pages_array(chunk_pages, pages,
2361 if (copy_from_user(chunk_pages, pages,
2362 chunk_nr * sizeof(*chunk_pages)))
2366 do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status);
2368 if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status)))
2373 nr_pages -= chunk_nr;
2375 return nr_pages ? -EFAULT : 0;
2378 static struct mm_struct *find_mm_struct(pid_t pid, nodemask_t *mem_nodes)
2380 struct task_struct *task;
2381 struct mm_struct *mm;
2384 * There is no need to check if current process has the right to modify
2385 * the specified process when they are same.
2389 *mem_nodes = cpuset_mems_allowed(current);
2393 /* Find the mm_struct */
2395 task = find_task_by_vpid(pid);
2398 return ERR_PTR(-ESRCH);
2400 get_task_struct(task);
2403 * Check if this process has the right to modify the specified
2404 * process. Use the regular "ptrace_may_access()" checks.
2406 if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) {
2408 mm = ERR_PTR(-EPERM);
2413 mm = ERR_PTR(security_task_movememory(task));
2416 *mem_nodes = cpuset_mems_allowed(task);
2417 mm = get_task_mm(task);
2419 put_task_struct(task);
2421 mm = ERR_PTR(-EINVAL);
2426 * Move a list of pages in the address space of the currently executing
2429 static int kernel_move_pages(pid_t pid, unsigned long nr_pages,
2430 const void __user * __user *pages,
2431 const int __user *nodes,
2432 int __user *status, int flags)
2434 struct mm_struct *mm;
2436 nodemask_t task_nodes;
2439 if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
2442 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
2445 mm = find_mm_struct(pid, &task_nodes);
2450 err = do_pages_move(mm, task_nodes, nr_pages, pages,
2451 nodes, status, flags);
2453 err = do_pages_stat(mm, nr_pages, pages, status);
2459 SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages,
2460 const void __user * __user *, pages,
2461 const int __user *, nodes,
2462 int __user *, status, int, flags)
2464 return kernel_move_pages(pid, nr_pages, pages, nodes, status, flags);
2467 #ifdef CONFIG_NUMA_BALANCING
2469 * Returns true if this is a safe migration target node for misplaced NUMA
2470 * pages. Currently it only checks the watermarks which is crude.
2472 static bool migrate_balanced_pgdat(struct pglist_data *pgdat,
2473 unsigned long nr_migrate_pages)
2477 for (z = pgdat->nr_zones - 1; z >= 0; z--) {
2478 struct zone *zone = pgdat->node_zones + z;
2480 if (!managed_zone(zone))
2483 /* Avoid waking kswapd by allocating pages_to_migrate pages. */
2484 if (!zone_watermark_ok(zone, 0,
2485 high_wmark_pages(zone) +
2494 static struct folio *alloc_misplaced_dst_folio(struct folio *src,
2497 int nid = (int) data;
2498 int order = folio_order(src);
2499 gfp_t gfp = __GFP_THISNODE;
2502 gfp |= GFP_TRANSHUGE_LIGHT;
2504 gfp |= GFP_HIGHUSER_MOVABLE | __GFP_NOMEMALLOC | __GFP_NORETRY |
2506 gfp &= ~__GFP_RECLAIM;
2508 return __folio_alloc_node(gfp, order, nid);
2511 static int numamigrate_isolate_folio(pg_data_t *pgdat, struct folio *folio)
2513 int nr_pages = folio_nr_pages(folio);
2515 /* Avoid migrating to a node that is nearly full */
2516 if (!migrate_balanced_pgdat(pgdat, nr_pages)) {
2519 if (!(sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING))
2521 for (z = pgdat->nr_zones - 1; z >= 0; z--) {
2522 if (managed_zone(pgdat->node_zones + z))
2527 * If there are no managed zones, it should not proceed
2533 wakeup_kswapd(pgdat->node_zones + z, 0,
2534 folio_order(folio), ZONE_MOVABLE);
2538 if (!folio_isolate_lru(folio))
2541 node_stat_mod_folio(folio, NR_ISOLATED_ANON + folio_is_file_lru(folio),
2545 * Isolating the folio has taken another reference, so the
2546 * caller's reference can be safely dropped without the folio
2547 * disappearing underneath us during migration.
2554 * Attempt to migrate a misplaced folio to the specified destination
2555 * node. Caller is expected to have an elevated reference count on
2556 * the folio that will be dropped by this function before returning.
2558 int migrate_misplaced_folio(struct folio *folio, struct vm_area_struct *vma,
2561 pg_data_t *pgdat = NODE_DATA(node);
2564 unsigned int nr_succeeded;
2565 LIST_HEAD(migratepages);
2566 int nr_pages = folio_nr_pages(folio);
2569 * Don't migrate file folios that are mapped in multiple processes
2570 * with execute permissions as they are probably shared libraries.
2571 * To check if the folio is shared, ideally we want to make sure
2572 * every page is mapped to the same process. Doing that is very
2573 * expensive, so check the estimated mapcount of the folio instead.
2575 if (folio_estimated_sharers(folio) != 1 && folio_is_file_lru(folio) &&
2576 (vma->vm_flags & VM_EXEC))
2580 * Also do not migrate dirty folios as not all filesystems can move
2581 * dirty folios in MIGRATE_ASYNC mode which is a waste of cycles.
2583 if (folio_is_file_lru(folio) && folio_test_dirty(folio))
2586 isolated = numamigrate_isolate_folio(pgdat, folio);
2590 list_add(&folio->lru, &migratepages);
2591 nr_remaining = migrate_pages(&migratepages, alloc_misplaced_dst_folio,
2592 NULL, node, MIGRATE_ASYNC,
2593 MR_NUMA_MISPLACED, &nr_succeeded);
2595 if (!list_empty(&migratepages)) {
2596 list_del(&folio->lru);
2597 node_stat_mod_folio(folio, NR_ISOLATED_ANON +
2598 folio_is_file_lru(folio), -nr_pages);
2599 folio_putback_lru(folio);
2604 count_vm_numa_events(NUMA_PAGE_MIGRATE, nr_succeeded);
2605 if (!node_is_toptier(folio_nid(folio)) && node_is_toptier(node))
2606 mod_node_page_state(pgdat, PGPROMOTE_SUCCESS,
2609 BUG_ON(!list_empty(&migratepages));
2616 #endif /* CONFIG_NUMA_BALANCING */
2617 #endif /* CONFIG_NUMA */