1 // SPDX-License-Identifier: GPL-2.0
3 * Memory Migration functionality - linux/mm/migrate.c
5 * Copyright (C) 2006 Silicon Graphics, Inc., Christoph Lameter
7 * Page migration was first developed in the context of the memory hotplug
8 * project. The main authors of the migration code are:
10 * IWAMOTO Toshihiro <iwamoto@valinux.co.jp>
11 * Hirokazu Takahashi <taka@valinux.co.jp>
12 * Dave Hansen <haveblue@us.ibm.com>
16 #include <linux/migrate.h>
17 #include <linux/export.h>
18 #include <linux/swap.h>
19 #include <linux/swapops.h>
20 #include <linux/pagemap.h>
21 #include <linux/buffer_head.h>
22 #include <linux/mm_inline.h>
23 #include <linux/nsproxy.h>
24 #include <linux/pagevec.h>
25 #include <linux/ksm.h>
26 #include <linux/rmap.h>
27 #include <linux/topology.h>
28 #include <linux/cpu.h>
29 #include <linux/cpuset.h>
30 #include <linux/writeback.h>
31 #include <linux/mempolicy.h>
32 #include <linux/vmalloc.h>
33 #include <linux/security.h>
34 #include <linux/backing-dev.h>
35 #include <linux/compaction.h>
36 #include <linux/syscalls.h>
37 #include <linux/compat.h>
38 #include <linux/hugetlb.h>
39 #include <linux/hugetlb_cgroup.h>
40 #include <linux/gfp.h>
41 #include <linux/pfn_t.h>
42 #include <linux/memremap.h>
43 #include <linux/userfaultfd_k.h>
44 #include <linux/balloon_compaction.h>
45 #include <linux/page_idle.h>
46 #include <linux/page_owner.h>
47 #include <linux/sched/mm.h>
48 #include <linux/ptrace.h>
49 #include <linux/oom.h>
50 #include <linux/memory.h>
51 #include <linux/random.h>
52 #include <linux/sched/sysctl.h>
54 #include <asm/tlbflush.h>
56 #include <trace/events/migrate.h>
60 int isolate_movable_page(struct page *page, isolate_mode_t mode)
62 const struct movable_operations *mops;
65 * Avoid burning cycles with pages that are yet under __free_pages(),
66 * or just got freed under us.
68 * In case we 'win' a race for a movable page being freed under us and
69 * raise its refcount preventing __free_pages() from doing its job
70 * the put_page() at the end of this block will take care of
71 * release this page, thus avoiding a nasty leakage.
73 if (unlikely(!get_page_unless_zero(page)))
77 * Check PageMovable before holding a PG_lock because page's owner
78 * assumes anybody doesn't touch PG_lock of newly allocated page
79 * so unconditionally grabbing the lock ruins page's owner side.
81 if (unlikely(!__PageMovable(page)))
84 * As movable pages are not isolated from LRU lists, concurrent
85 * compaction threads can race against page migration functions
86 * as well as race against the releasing a page.
88 * In order to avoid having an already isolated movable page
89 * being (wrongly) re-isolated while it is under migration,
90 * or to avoid attempting to isolate pages being released,
91 * lets be sure we have the page lock
92 * before proceeding with the movable page isolation steps.
94 if (unlikely(!trylock_page(page)))
97 if (!PageMovable(page) || PageIsolated(page))
100 mops = page_movable_ops(page);
101 VM_BUG_ON_PAGE(!mops, page);
103 if (!mops->isolate_page(page, mode))
104 goto out_no_isolated;
106 /* Driver shouldn't use PG_isolated bit of page->flags */
107 WARN_ON_ONCE(PageIsolated(page));
108 SetPageIsolated(page);
121 static void putback_movable_page(struct page *page)
123 const struct movable_operations *mops = page_movable_ops(page);
125 mops->putback_page(page);
126 ClearPageIsolated(page);
130 * Put previously isolated pages back onto the appropriate lists
131 * from where they were once taken off for compaction/migration.
133 * This function shall be used whenever the isolated pageset has been
134 * built from lru, balloon, hugetlbfs page. See isolate_migratepages_range()
135 * and isolate_hugetlb().
137 void putback_movable_pages(struct list_head *l)
142 list_for_each_entry_safe(page, page2, l, lru) {
143 if (unlikely(PageHuge(page))) {
144 putback_active_hugepage(page);
147 list_del(&page->lru);
149 * We isolated non-lru movable page so here we can use
150 * __PageMovable because LRU page's mapping cannot have
151 * PAGE_MAPPING_MOVABLE.
153 if (unlikely(__PageMovable(page))) {
154 VM_BUG_ON_PAGE(!PageIsolated(page), page);
156 if (PageMovable(page))
157 putback_movable_page(page);
159 ClearPageIsolated(page);
163 mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON +
164 page_is_file_lru(page), -thp_nr_pages(page));
165 putback_lru_page(page);
171 * Restore a potential migration pte to a working pte entry
173 static bool remove_migration_pte(struct folio *folio,
174 struct vm_area_struct *vma, unsigned long addr, void *old)
176 DEFINE_FOLIO_VMA_WALK(pvmw, old, vma, addr, PVMW_SYNC | PVMW_MIGRATION);
178 while (page_vma_mapped_walk(&pvmw)) {
179 rmap_t rmap_flags = RMAP_NONE;
183 unsigned long idx = 0;
185 /* pgoff is invalid for ksm pages, but they are never large */
186 if (folio_test_large(folio) && !folio_test_hugetlb(folio))
187 idx = linear_page_index(vma, pvmw.address) - pvmw.pgoff;
188 new = folio_page(folio, idx);
190 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
191 /* PMD-mapped THP migration entry */
193 VM_BUG_ON_FOLIO(folio_test_hugetlb(folio) ||
194 !folio_test_pmd_mappable(folio), folio);
195 remove_migration_pmd(&pvmw, new);
201 pte = mk_pte(new, READ_ONCE(vma->vm_page_prot));
202 if (pte_swp_soft_dirty(*pvmw.pte))
203 pte = pte_mksoft_dirty(pte);
206 * Recheck VMA as permissions can change since migration started
208 entry = pte_to_swp_entry(*pvmw.pte);
209 if (!is_migration_entry_young(entry))
210 pte = pte_mkold(pte);
211 if (folio_test_dirty(folio) && is_migration_entry_dirty(entry))
212 pte = pte_mkdirty(pte);
213 if (is_writable_migration_entry(entry))
214 pte = maybe_mkwrite(pte, vma);
215 else if (pte_swp_uffd_wp(*pvmw.pte))
216 pte = pte_mkuffd_wp(pte);
218 if (folio_test_anon(folio) && !is_readable_migration_entry(entry))
219 rmap_flags |= RMAP_EXCLUSIVE;
221 if (unlikely(is_device_private_page(new))) {
223 entry = make_writable_device_private_entry(
226 entry = make_readable_device_private_entry(
228 pte = swp_entry_to_pte(entry);
229 if (pte_swp_soft_dirty(*pvmw.pte))
230 pte = pte_swp_mksoft_dirty(pte);
231 if (pte_swp_uffd_wp(*pvmw.pte))
232 pte = pte_swp_mkuffd_wp(pte);
235 #ifdef CONFIG_HUGETLB_PAGE
236 if (folio_test_hugetlb(folio)) {
237 unsigned int shift = huge_page_shift(hstate_vma(vma));
239 pte = pte_mkhuge(pte);
240 pte = arch_make_huge_pte(pte, shift, vma->vm_flags);
241 if (folio_test_anon(folio))
242 hugepage_add_anon_rmap(new, vma, pvmw.address,
245 page_dup_file_rmap(new, true);
246 set_huge_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte);
250 if (folio_test_anon(folio))
251 page_add_anon_rmap(new, vma, pvmw.address,
254 page_add_file_rmap(new, vma, false);
255 set_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte);
257 if (vma->vm_flags & VM_LOCKED)
258 mlock_page_drain_local();
260 trace_remove_migration_pte(pvmw.address, pte_val(pte),
261 compound_order(new));
263 /* No need to invalidate - it was non-present before */
264 update_mmu_cache(vma, pvmw.address, pvmw.pte);
271 * Get rid of all migration entries and replace them by
272 * references to the indicated page.
274 void remove_migration_ptes(struct folio *src, struct folio *dst, bool locked)
276 struct rmap_walk_control rwc = {
277 .rmap_one = remove_migration_pte,
282 rmap_walk_locked(dst, &rwc);
284 rmap_walk(dst, &rwc);
288 * Something used the pte of a page under migration. We need to
289 * get to the page and wait until migration is finished.
290 * When we return from this function the fault will be retried.
292 void __migration_entry_wait(struct mm_struct *mm, pte_t *ptep,
300 if (!is_swap_pte(pte))
303 entry = pte_to_swp_entry(pte);
304 if (!is_migration_entry(entry))
307 migration_entry_wait_on_locked(entry, ptep, ptl);
310 pte_unmap_unlock(ptep, ptl);
313 void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
314 unsigned long address)
316 spinlock_t *ptl = pte_lockptr(mm, pmd);
317 pte_t *ptep = pte_offset_map(pmd, address);
318 __migration_entry_wait(mm, ptep, ptl);
321 #ifdef CONFIG_HUGETLB_PAGE
322 void __migration_entry_wait_huge(pte_t *ptep, spinlock_t *ptl)
327 pte = huge_ptep_get(ptep);
329 if (unlikely(!is_hugetlb_entry_migration(pte)))
332 migration_entry_wait_on_locked(pte_to_swp_entry(pte), NULL, ptl);
335 void migration_entry_wait_huge(struct vm_area_struct *vma, pte_t *pte)
337 spinlock_t *ptl = huge_pte_lockptr(hstate_vma(vma), vma->vm_mm, pte);
339 __migration_entry_wait_huge(pte, ptl);
343 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
344 void pmd_migration_entry_wait(struct mm_struct *mm, pmd_t *pmd)
348 ptl = pmd_lock(mm, pmd);
349 if (!is_pmd_migration_entry(*pmd))
351 migration_entry_wait_on_locked(pmd_to_swp_entry(*pmd), NULL, ptl);
358 static int folio_expected_refs(struct address_space *mapping,
365 refs += folio_nr_pages(folio);
366 if (folio_test_private(folio))
373 * Replace the page in the mapping.
375 * The number of remaining references must be:
376 * 1 for anonymous pages without a mapping
377 * 2 for pages with a mapping
378 * 3 for pages with a mapping and PagePrivate/PagePrivate2 set.
380 int folio_migrate_mapping(struct address_space *mapping,
381 struct folio *newfolio, struct folio *folio, int extra_count)
383 XA_STATE(xas, &mapping->i_pages, folio_index(folio));
384 struct zone *oldzone, *newzone;
386 int expected_count = folio_expected_refs(mapping, folio) + extra_count;
387 long nr = folio_nr_pages(folio);
390 /* Anonymous page without mapping */
391 if (folio_ref_count(folio) != expected_count)
394 /* No turning back from here */
395 newfolio->index = folio->index;
396 newfolio->mapping = folio->mapping;
397 if (folio_test_swapbacked(folio))
398 __folio_set_swapbacked(newfolio);
400 return MIGRATEPAGE_SUCCESS;
403 oldzone = folio_zone(folio);
404 newzone = folio_zone(newfolio);
407 if (!folio_ref_freeze(folio, expected_count)) {
408 xas_unlock_irq(&xas);
413 * Now we know that no one else is looking at the folio:
414 * no turning back from here.
416 newfolio->index = folio->index;
417 newfolio->mapping = folio->mapping;
418 folio_ref_add(newfolio, nr); /* add cache reference */
419 if (folio_test_swapbacked(folio)) {
420 __folio_set_swapbacked(newfolio);
421 if (folio_test_swapcache(folio)) {
422 folio_set_swapcache(newfolio);
423 newfolio->private = folio_get_private(folio);
426 VM_BUG_ON_FOLIO(folio_test_swapcache(folio), folio);
429 /* Move dirty while page refs frozen and newpage not yet exposed */
430 dirty = folio_test_dirty(folio);
432 folio_clear_dirty(folio);
433 folio_set_dirty(newfolio);
436 xas_store(&xas, newfolio);
439 * Drop cache reference from old page by unfreezing
440 * to one less reference.
441 * We know this isn't the last reference.
443 folio_ref_unfreeze(folio, expected_count - nr);
446 /* Leave irq disabled to prevent preemption while updating stats */
449 * If moved to a different zone then also account
450 * the page for that zone. Other VM counters will be
451 * taken care of when we establish references to the
452 * new page and drop references to the old page.
454 * Note that anonymous pages are accounted for
455 * via NR_FILE_PAGES and NR_ANON_MAPPED if they
456 * are mapped to swap space.
458 if (newzone != oldzone) {
459 struct lruvec *old_lruvec, *new_lruvec;
460 struct mem_cgroup *memcg;
462 memcg = folio_memcg(folio);
463 old_lruvec = mem_cgroup_lruvec(memcg, oldzone->zone_pgdat);
464 new_lruvec = mem_cgroup_lruvec(memcg, newzone->zone_pgdat);
466 __mod_lruvec_state(old_lruvec, NR_FILE_PAGES, -nr);
467 __mod_lruvec_state(new_lruvec, NR_FILE_PAGES, nr);
468 if (folio_test_swapbacked(folio) && !folio_test_swapcache(folio)) {
469 __mod_lruvec_state(old_lruvec, NR_SHMEM, -nr);
470 __mod_lruvec_state(new_lruvec, NR_SHMEM, nr);
473 if (folio_test_swapcache(folio)) {
474 __mod_lruvec_state(old_lruvec, NR_SWAPCACHE, -nr);
475 __mod_lruvec_state(new_lruvec, NR_SWAPCACHE, nr);
478 if (dirty && mapping_can_writeback(mapping)) {
479 __mod_lruvec_state(old_lruvec, NR_FILE_DIRTY, -nr);
480 __mod_zone_page_state(oldzone, NR_ZONE_WRITE_PENDING, -nr);
481 __mod_lruvec_state(new_lruvec, NR_FILE_DIRTY, nr);
482 __mod_zone_page_state(newzone, NR_ZONE_WRITE_PENDING, nr);
487 return MIGRATEPAGE_SUCCESS;
489 EXPORT_SYMBOL(folio_migrate_mapping);
492 * The expected number of remaining references is the same as that
493 * of folio_migrate_mapping().
495 int migrate_huge_page_move_mapping(struct address_space *mapping,
496 struct folio *dst, struct folio *src)
498 XA_STATE(xas, &mapping->i_pages, folio_index(src));
502 expected_count = 2 + folio_has_private(src);
503 if (!folio_ref_freeze(src, expected_count)) {
504 xas_unlock_irq(&xas);
508 dst->index = src->index;
509 dst->mapping = src->mapping;
513 xas_store(&xas, dst);
515 folio_ref_unfreeze(src, expected_count - 1);
517 xas_unlock_irq(&xas);
519 return MIGRATEPAGE_SUCCESS;
523 * Copy the flags and some other ancillary information
525 void folio_migrate_flags(struct folio *newfolio, struct folio *folio)
529 if (folio_test_error(folio))
530 folio_set_error(newfolio);
531 if (folio_test_referenced(folio))
532 folio_set_referenced(newfolio);
533 if (folio_test_uptodate(folio))
534 folio_mark_uptodate(newfolio);
535 if (folio_test_clear_active(folio)) {
536 VM_BUG_ON_FOLIO(folio_test_unevictable(folio), folio);
537 folio_set_active(newfolio);
538 } else if (folio_test_clear_unevictable(folio))
539 folio_set_unevictable(newfolio);
540 if (folio_test_workingset(folio))
541 folio_set_workingset(newfolio);
542 if (folio_test_checked(folio))
543 folio_set_checked(newfolio);
545 * PG_anon_exclusive (-> PG_mappedtodisk) is always migrated via
546 * migration entries. We can still have PG_anon_exclusive set on an
547 * effectively unmapped and unreferenced first sub-pages of an
548 * anonymous THP: we can simply copy it here via PG_mappedtodisk.
550 if (folio_test_mappedtodisk(folio))
551 folio_set_mappedtodisk(newfolio);
553 /* Move dirty on pages not done by folio_migrate_mapping() */
554 if (folio_test_dirty(folio))
555 folio_set_dirty(newfolio);
557 if (folio_test_young(folio))
558 folio_set_young(newfolio);
559 if (folio_test_idle(folio))
560 folio_set_idle(newfolio);
563 * Copy NUMA information to the new page, to prevent over-eager
564 * future migrations of this same page.
566 cpupid = page_cpupid_xchg_last(&folio->page, -1);
568 * For memory tiering mode, when migrate between slow and fast
569 * memory node, reset cpupid, because that is used to record
570 * page access time in slow memory node.
572 if (sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING) {
573 bool f_toptier = node_is_toptier(page_to_nid(&folio->page));
574 bool t_toptier = node_is_toptier(page_to_nid(&newfolio->page));
576 if (f_toptier != t_toptier)
579 page_cpupid_xchg_last(&newfolio->page, cpupid);
581 folio_migrate_ksm(newfolio, folio);
583 * Please do not reorder this without considering how mm/ksm.c's
584 * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache().
586 if (folio_test_swapcache(folio))
587 folio_clear_swapcache(folio);
588 folio_clear_private(folio);
590 /* page->private contains hugetlb specific flags */
591 if (!folio_test_hugetlb(folio))
592 folio->private = NULL;
595 * If any waiters have accumulated on the new page then
598 if (folio_test_writeback(newfolio))
599 folio_end_writeback(newfolio);
602 * PG_readahead shares the same bit with PG_reclaim. The above
603 * end_page_writeback() may clear PG_readahead mistakenly, so set the
606 if (folio_test_readahead(folio))
607 folio_set_readahead(newfolio);
609 folio_copy_owner(newfolio, folio);
611 if (!folio_test_hugetlb(folio))
612 mem_cgroup_migrate(folio, newfolio);
614 EXPORT_SYMBOL(folio_migrate_flags);
616 void folio_migrate_copy(struct folio *newfolio, struct folio *folio)
618 folio_copy(newfolio, folio);
619 folio_migrate_flags(newfolio, folio);
621 EXPORT_SYMBOL(folio_migrate_copy);
623 /************************************************************
624 * Migration functions
625 ***********************************************************/
628 * migrate_folio() - Simple folio migration.
629 * @mapping: The address_space containing the folio.
630 * @dst: The folio to migrate the data to.
631 * @src: The folio containing the current data.
632 * @mode: How to migrate the page.
634 * Common logic to directly migrate a single LRU folio suitable for
635 * folios that do not use PagePrivate/PagePrivate2.
637 * Folios are locked upon entry and exit.
639 int migrate_folio(struct address_space *mapping, struct folio *dst,
640 struct folio *src, enum migrate_mode mode)
644 BUG_ON(folio_test_writeback(src)); /* Writeback must be complete */
646 rc = folio_migrate_mapping(mapping, dst, src, 0);
648 if (rc != MIGRATEPAGE_SUCCESS)
651 if (mode != MIGRATE_SYNC_NO_COPY)
652 folio_migrate_copy(dst, src);
654 folio_migrate_flags(dst, src);
655 return MIGRATEPAGE_SUCCESS;
657 EXPORT_SYMBOL(migrate_folio);
660 /* Returns true if all buffers are successfully locked */
661 static bool buffer_migrate_lock_buffers(struct buffer_head *head,
662 enum migrate_mode mode)
664 struct buffer_head *bh = head;
666 /* Simple case, sync compaction */
667 if (mode != MIGRATE_ASYNC) {
670 bh = bh->b_this_page;
672 } while (bh != head);
677 /* async case, we cannot block on lock_buffer so use trylock_buffer */
679 if (!trylock_buffer(bh)) {
681 * We failed to lock the buffer and cannot stall in
682 * async migration. Release the taken locks
684 struct buffer_head *failed_bh = bh;
686 while (bh != failed_bh) {
688 bh = bh->b_this_page;
693 bh = bh->b_this_page;
694 } while (bh != head);
698 static int __buffer_migrate_folio(struct address_space *mapping,
699 struct folio *dst, struct folio *src, enum migrate_mode mode,
702 struct buffer_head *bh, *head;
706 head = folio_buffers(src);
708 return migrate_folio(mapping, dst, src, mode);
710 /* Check whether page does not have extra refs before we do more work */
711 expected_count = folio_expected_refs(mapping, src);
712 if (folio_ref_count(src) != expected_count)
715 if (!buffer_migrate_lock_buffers(head, mode))
720 bool invalidated = false;
724 spin_lock(&mapping->private_lock);
727 if (atomic_read(&bh->b_count)) {
731 bh = bh->b_this_page;
732 } while (bh != head);
738 spin_unlock(&mapping->private_lock);
739 invalidate_bh_lrus();
741 goto recheck_buffers;
745 rc = folio_migrate_mapping(mapping, dst, src, 0);
746 if (rc != MIGRATEPAGE_SUCCESS)
749 folio_attach_private(dst, folio_detach_private(src));
753 set_bh_page(bh, &dst->page, bh_offset(bh));
754 bh = bh->b_this_page;
755 } while (bh != head);
757 if (mode != MIGRATE_SYNC_NO_COPY)
758 folio_migrate_copy(dst, src);
760 folio_migrate_flags(dst, src);
762 rc = MIGRATEPAGE_SUCCESS;
765 spin_unlock(&mapping->private_lock);
769 bh = bh->b_this_page;
770 } while (bh != head);
776 * buffer_migrate_folio() - Migration function for folios with buffers.
777 * @mapping: The address space containing @src.
778 * @dst: The folio to migrate to.
779 * @src: The folio to migrate from.
780 * @mode: How to migrate the folio.
782 * This function can only be used if the underlying filesystem guarantees
783 * that no other references to @src exist. For example attached buffer
784 * heads are accessed only under the folio lock. If your filesystem cannot
785 * provide this guarantee, buffer_migrate_folio_norefs() may be more
788 * Return: 0 on success or a negative errno on failure.
790 int buffer_migrate_folio(struct address_space *mapping,
791 struct folio *dst, struct folio *src, enum migrate_mode mode)
793 return __buffer_migrate_folio(mapping, dst, src, mode, false);
795 EXPORT_SYMBOL(buffer_migrate_folio);
798 * buffer_migrate_folio_norefs() - Migration function for folios with buffers.
799 * @mapping: The address space containing @src.
800 * @dst: The folio to migrate to.
801 * @src: The folio to migrate from.
802 * @mode: How to migrate the folio.
804 * Like buffer_migrate_folio() except that this variant is more careful
805 * and checks that there are also no buffer head references. This function
806 * is the right one for mappings where buffer heads are directly looked
807 * up and referenced (such as block device mappings).
809 * Return: 0 on success or a negative errno on failure.
811 int buffer_migrate_folio_norefs(struct address_space *mapping,
812 struct folio *dst, struct folio *src, enum migrate_mode mode)
814 return __buffer_migrate_folio(mapping, dst, src, mode, true);
818 int filemap_migrate_folio(struct address_space *mapping,
819 struct folio *dst, struct folio *src, enum migrate_mode mode)
823 ret = folio_migrate_mapping(mapping, dst, src, 0);
824 if (ret != MIGRATEPAGE_SUCCESS)
827 if (folio_get_private(src))
828 folio_attach_private(dst, folio_detach_private(src));
830 if (mode != MIGRATE_SYNC_NO_COPY)
831 folio_migrate_copy(dst, src);
833 folio_migrate_flags(dst, src);
834 return MIGRATEPAGE_SUCCESS;
836 EXPORT_SYMBOL_GPL(filemap_migrate_folio);
839 * Writeback a folio to clean the dirty state
841 static int writeout(struct address_space *mapping, struct folio *folio)
843 struct writeback_control wbc = {
844 .sync_mode = WB_SYNC_NONE,
847 .range_end = LLONG_MAX,
852 if (!mapping->a_ops->writepage)
853 /* No write method for the address space */
856 if (!folio_clear_dirty_for_io(folio))
857 /* Someone else already triggered a write */
861 * A dirty folio may imply that the underlying filesystem has
862 * the folio on some queue. So the folio must be clean for
863 * migration. Writeout may mean we lose the lock and the
864 * folio state is no longer what we checked for earlier.
865 * At this point we know that the migration attempt cannot
868 remove_migration_ptes(folio, folio, false);
870 rc = mapping->a_ops->writepage(&folio->page, &wbc);
872 if (rc != AOP_WRITEPAGE_ACTIVATE)
873 /* unlocked. Relock */
876 return (rc < 0) ? -EIO : -EAGAIN;
880 * Default handling if a filesystem does not provide a migration function.
882 static int fallback_migrate_folio(struct address_space *mapping,
883 struct folio *dst, struct folio *src, enum migrate_mode mode)
885 if (folio_test_dirty(src)) {
886 /* Only writeback folios in full synchronous migration */
889 case MIGRATE_SYNC_NO_COPY:
894 return writeout(mapping, src);
898 * Buffers may be managed in a filesystem specific way.
899 * We must have no buffers or drop them.
901 if (folio_test_private(src) &&
902 !filemap_release_folio(src, GFP_KERNEL))
903 return mode == MIGRATE_SYNC ? -EAGAIN : -EBUSY;
905 return migrate_folio(mapping, dst, src, mode);
909 * Move a page to a newly allocated page
910 * The page is locked and all ptes have been successfully removed.
912 * The new page will have replaced the old page if this function
917 * MIGRATEPAGE_SUCCESS - success
919 static int move_to_new_folio(struct folio *dst, struct folio *src,
920 enum migrate_mode mode)
923 bool is_lru = !__PageMovable(&src->page);
925 VM_BUG_ON_FOLIO(!folio_test_locked(src), src);
926 VM_BUG_ON_FOLIO(!folio_test_locked(dst), dst);
928 if (likely(is_lru)) {
929 struct address_space *mapping = folio_mapping(src);
932 rc = migrate_folio(mapping, dst, src, mode);
933 else if (mapping->a_ops->migrate_folio)
935 * Most folios have a mapping and most filesystems
936 * provide a migrate_folio callback. Anonymous folios
937 * are part of swap space which also has its own
938 * migrate_folio callback. This is the most common path
939 * for page migration.
941 rc = mapping->a_ops->migrate_folio(mapping, dst, src,
944 rc = fallback_migrate_folio(mapping, dst, src, mode);
946 const struct movable_operations *mops;
949 * In case of non-lru page, it could be released after
950 * isolation step. In that case, we shouldn't try migration.
952 VM_BUG_ON_FOLIO(!folio_test_isolated(src), src);
953 if (!folio_test_movable(src)) {
954 rc = MIGRATEPAGE_SUCCESS;
955 folio_clear_isolated(src);
959 mops = page_movable_ops(&src->page);
960 rc = mops->migrate_page(&dst->page, &src->page, mode);
961 WARN_ON_ONCE(rc == MIGRATEPAGE_SUCCESS &&
962 !folio_test_isolated(src));
966 * When successful, old pagecache src->mapping must be cleared before
967 * src is freed; but stats require that PageAnon be left as PageAnon.
969 if (rc == MIGRATEPAGE_SUCCESS) {
970 if (__PageMovable(&src->page)) {
971 VM_BUG_ON_FOLIO(!folio_test_isolated(src), src);
974 * We clear PG_movable under page_lock so any compactor
975 * cannot try to migrate this page.
977 folio_clear_isolated(src);
981 * Anonymous and movable src->mapping will be cleared by
982 * free_pages_prepare so don't reset it here for keeping
983 * the type to work PageAnon, for example.
985 if (!folio_mapping_flags(src))
988 if (likely(!folio_is_zone_device(dst)))
989 flush_dcache_folio(dst);
995 static int __unmap_and_move(struct page *page, struct page *newpage,
996 int force, enum migrate_mode mode)
998 struct folio *folio = page_folio(page);
999 struct folio *dst = page_folio(newpage);
1001 bool page_was_mapped = false;
1002 struct anon_vma *anon_vma = NULL;
1003 bool is_lru = !__PageMovable(page);
1005 if (!trylock_page(page)) {
1006 if (!force || mode == MIGRATE_ASYNC)
1010 * It's not safe for direct compaction to call lock_page.
1011 * For example, during page readahead pages are added locked
1012 * to the LRU. Later, when the IO completes the pages are
1013 * marked uptodate and unlocked. However, the queueing
1014 * could be merging multiple pages for one bio (e.g.
1015 * mpage_readahead). If an allocation happens for the
1016 * second or third page, the process can end up locking
1017 * the same page twice and deadlocking. Rather than
1018 * trying to be clever about what pages can be locked,
1019 * avoid the use of lock_page for direct compaction
1022 if (current->flags & PF_MEMALLOC)
1028 if (PageWriteback(page)) {
1030 * Only in the case of a full synchronous migration is it
1031 * necessary to wait for PageWriteback. In the async case,
1032 * the retry loop is too short and in the sync-light case,
1033 * the overhead of stalling is too much
1037 case MIGRATE_SYNC_NO_COPY:
1045 wait_on_page_writeback(page);
1049 * By try_to_migrate(), page->mapcount goes down to 0 here. In this case,
1050 * we cannot notice that anon_vma is freed while we migrates a page.
1051 * This get_anon_vma() delays freeing anon_vma pointer until the end
1052 * of migration. File cache pages are no problem because of page_lock()
1053 * File Caches may use write_page() or lock_page() in migration, then,
1054 * just care Anon page here.
1056 * Only page_get_anon_vma() understands the subtleties of
1057 * getting a hold on an anon_vma from outside one of its mms.
1058 * But if we cannot get anon_vma, then we won't need it anyway,
1059 * because that implies that the anon page is no longer mapped
1060 * (and cannot be remapped so long as we hold the page lock).
1062 if (PageAnon(page) && !PageKsm(page))
1063 anon_vma = page_get_anon_vma(page);
1066 * Block others from accessing the new page when we get around to
1067 * establishing additional references. We are usually the only one
1068 * holding a reference to newpage at this point. We used to have a BUG
1069 * here if trylock_page(newpage) fails, but would like to allow for
1070 * cases where there might be a race with the previous use of newpage.
1071 * This is much like races on refcount of oldpage: just don't BUG().
1073 if (unlikely(!trylock_page(newpage)))
1076 if (unlikely(!is_lru)) {
1077 rc = move_to_new_folio(dst, folio, mode);
1078 goto out_unlock_both;
1082 * Corner case handling:
1083 * 1. When a new swap-cache page is read into, it is added to the LRU
1084 * and treated as swapcache but it has no rmap yet.
1085 * Calling try_to_unmap() against a page->mapping==NULL page will
1086 * trigger a BUG. So handle it here.
1087 * 2. An orphaned page (see truncate_cleanup_page) might have
1088 * fs-private metadata. The page can be picked up due to memory
1089 * offlining. Everywhere else except page reclaim, the page is
1090 * invisible to the vm, so the page can not be migrated. So try to
1091 * free the metadata, so the page can be freed.
1093 if (!page->mapping) {
1094 VM_BUG_ON_PAGE(PageAnon(page), page);
1095 if (page_has_private(page)) {
1096 try_to_free_buffers(folio);
1097 goto out_unlock_both;
1099 } else if (page_mapped(page)) {
1100 /* Establish migration ptes */
1101 VM_BUG_ON_PAGE(PageAnon(page) && !PageKsm(page) && !anon_vma,
1103 try_to_migrate(folio, 0);
1104 page_was_mapped = true;
1107 if (!page_mapped(page))
1108 rc = move_to_new_folio(dst, folio, mode);
1111 * When successful, push newpage to LRU immediately: so that if it
1112 * turns out to be an mlocked page, remove_migration_ptes() will
1113 * automatically build up the correct newpage->mlock_count for it.
1115 * We would like to do something similar for the old page, when
1116 * unsuccessful, and other cases when a page has been temporarily
1117 * isolated from the unevictable LRU: but this case is the easiest.
1119 if (rc == MIGRATEPAGE_SUCCESS) {
1120 lru_cache_add(newpage);
1121 if (page_was_mapped)
1125 if (page_was_mapped)
1126 remove_migration_ptes(folio,
1127 rc == MIGRATEPAGE_SUCCESS ? dst : folio, false);
1130 unlock_page(newpage);
1132 /* Drop an anon_vma reference if we took one */
1134 put_anon_vma(anon_vma);
1138 * If migration is successful, decrease refcount of the newpage,
1139 * which will not free the page because new page owner increased
1142 if (rc == MIGRATEPAGE_SUCCESS)
1149 * Obtain the lock on page, remove all ptes and migrate the page
1150 * to the newly allocated page in newpage.
1152 static int unmap_and_move(new_page_t get_new_page,
1153 free_page_t put_new_page,
1154 unsigned long private, struct page *page,
1155 int force, enum migrate_mode mode,
1156 enum migrate_reason reason,
1157 struct list_head *ret)
1159 int rc = MIGRATEPAGE_SUCCESS;
1160 struct page *newpage = NULL;
1162 if (!thp_migration_supported() && PageTransHuge(page))
1165 if (page_count(page) == 1) {
1166 /* Page was freed from under us. So we are done. */
1167 ClearPageActive(page);
1168 ClearPageUnevictable(page);
1169 /* free_pages_prepare() will clear PG_isolated. */
1173 newpage = get_new_page(page, private);
1177 newpage->private = 0;
1178 rc = __unmap_and_move(page, newpage, force, mode);
1179 if (rc == MIGRATEPAGE_SUCCESS)
1180 set_page_owner_migrate_reason(newpage, reason);
1183 if (rc != -EAGAIN) {
1185 * A page that has been migrated has all references
1186 * removed and will be freed. A page that has not been
1187 * migrated will have kept its references and be restored.
1189 list_del(&page->lru);
1193 * If migration is successful, releases reference grabbed during
1194 * isolation. Otherwise, restore the page to right list unless
1197 if (rc == MIGRATEPAGE_SUCCESS) {
1199 * Compaction can migrate also non-LRU pages which are
1200 * not accounted to NR_ISOLATED_*. They can be recognized
1203 if (likely(!__PageMovable(page)))
1204 mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON +
1205 page_is_file_lru(page), -thp_nr_pages(page));
1207 if (reason != MR_MEMORY_FAILURE)
1209 * We release the page in page_handle_poison.
1214 list_add_tail(&page->lru, ret);
1217 put_new_page(newpage, private);
1226 * Counterpart of unmap_and_move_page() for hugepage migration.
1228 * This function doesn't wait the completion of hugepage I/O
1229 * because there is no race between I/O and migration for hugepage.
1230 * Note that currently hugepage I/O occurs only in direct I/O
1231 * where no lock is held and PG_writeback is irrelevant,
1232 * and writeback status of all subpages are counted in the reference
1233 * count of the head page (i.e. if all subpages of a 2MB hugepage are
1234 * under direct I/O, the reference of the head page is 512 and a bit more.)
1235 * This means that when we try to migrate hugepage whose subpages are
1236 * doing direct I/O, some references remain after try_to_unmap() and
1237 * hugepage migration fails without data corruption.
1239 * There is also no race when direct I/O is issued on the page under migration,
1240 * because then pte is replaced with migration swap entry and direct I/O code
1241 * will wait in the page fault for migration to complete.
1243 static int unmap_and_move_huge_page(new_page_t get_new_page,
1244 free_page_t put_new_page, unsigned long private,
1245 struct page *hpage, int force,
1246 enum migrate_mode mode, int reason,
1247 struct list_head *ret)
1249 struct folio *dst, *src = page_folio(hpage);
1251 int page_was_mapped = 0;
1252 struct page *new_hpage;
1253 struct anon_vma *anon_vma = NULL;
1254 struct address_space *mapping = NULL;
1257 * Migratability of hugepages depends on architectures and their size.
1258 * This check is necessary because some callers of hugepage migration
1259 * like soft offline and memory hotremove don't walk through page
1260 * tables or check whether the hugepage is pmd-based or not before
1261 * kicking migration.
1263 if (!hugepage_migration_supported(page_hstate(hpage))) {
1264 list_move_tail(&hpage->lru, ret);
1268 if (page_count(hpage) == 1) {
1269 /* page was freed from under us. So we are done. */
1270 putback_active_hugepage(hpage);
1271 return MIGRATEPAGE_SUCCESS;
1274 new_hpage = get_new_page(hpage, private);
1277 dst = page_folio(new_hpage);
1279 if (!trylock_page(hpage)) {
1284 case MIGRATE_SYNC_NO_COPY:
1293 * Check for pages which are in the process of being freed. Without
1294 * page_mapping() set, hugetlbfs specific move page routine will not
1295 * be called and we could leak usage counts for subpools.
1297 if (hugetlb_page_subpool(hpage) && !page_mapping(hpage)) {
1302 if (PageAnon(hpage))
1303 anon_vma = page_get_anon_vma(hpage);
1305 if (unlikely(!trylock_page(new_hpage)))
1308 if (page_mapped(hpage)) {
1309 enum ttu_flags ttu = 0;
1311 if (!PageAnon(hpage)) {
1313 * In shared mappings, try_to_unmap could potentially
1314 * call huge_pmd_unshare. Because of this, take
1315 * semaphore in write mode here and set TTU_RMAP_LOCKED
1316 * to let lower levels know we have taken the lock.
1318 mapping = hugetlb_page_mapping_lock_write(hpage);
1319 if (unlikely(!mapping))
1320 goto unlock_put_anon;
1322 ttu = TTU_RMAP_LOCKED;
1325 try_to_migrate(src, ttu);
1326 page_was_mapped = 1;
1328 if (ttu & TTU_RMAP_LOCKED)
1329 i_mmap_unlock_write(mapping);
1332 if (!page_mapped(hpage))
1333 rc = move_to_new_folio(dst, src, mode);
1335 if (page_was_mapped)
1336 remove_migration_ptes(src,
1337 rc == MIGRATEPAGE_SUCCESS ? dst : src, false);
1340 unlock_page(new_hpage);
1344 put_anon_vma(anon_vma);
1346 if (rc == MIGRATEPAGE_SUCCESS) {
1347 move_hugetlb_state(hpage, new_hpage, reason);
1348 put_new_page = NULL;
1354 if (rc == MIGRATEPAGE_SUCCESS)
1355 putback_active_hugepage(hpage);
1356 else if (rc != -EAGAIN)
1357 list_move_tail(&hpage->lru, ret);
1360 * If migration was not successful and there's a freeing callback, use
1361 * it. Otherwise, put_page() will drop the reference grabbed during
1365 put_new_page(new_hpage, private);
1367 putback_active_hugepage(new_hpage);
1372 static inline int try_split_thp(struct page *page, struct list_head *split_pages)
1377 rc = split_huge_page_to_list(page, split_pages);
1384 * migrate_pages - migrate the pages specified in a list, to the free pages
1385 * supplied as the target for the page migration
1387 * @from: The list of pages to be migrated.
1388 * @get_new_page: The function used to allocate free pages to be used
1389 * as the target of the page migration.
1390 * @put_new_page: The function used to free target pages if migration
1391 * fails, or NULL if no special handling is necessary.
1392 * @private: Private data to be passed on to get_new_page()
1393 * @mode: The migration mode that specifies the constraints for
1394 * page migration, if any.
1395 * @reason: The reason for page migration.
1396 * @ret_succeeded: Set to the number of normal pages migrated successfully if
1397 * the caller passes a non-NULL pointer.
1399 * The function returns after 10 attempts or if no pages are movable any more
1400 * because the list has become empty or no retryable pages exist any more.
1401 * It is caller's responsibility to call putback_movable_pages() to return pages
1402 * to the LRU or free list only if ret != 0.
1404 * Returns the number of {normal page, THP, hugetlb} that were not migrated, or
1405 * an error code. The number of THP splits will be considered as the number of
1406 * non-migrated THP, no matter how many subpages of the THP are migrated successfully.
1408 int migrate_pages(struct list_head *from, new_page_t get_new_page,
1409 free_page_t put_new_page, unsigned long private,
1410 enum migrate_mode mode, int reason, unsigned int *ret_succeeded)
1415 int nr_failed_pages = 0;
1416 int nr_succeeded = 0;
1417 int nr_thp_succeeded = 0;
1418 int nr_thp_failed = 0;
1419 int nr_thp_split = 0;
1421 bool is_thp = false;
1424 int rc, nr_subpages;
1425 LIST_HEAD(ret_pages);
1426 LIST_HEAD(thp_split_pages);
1427 bool nosplit = (reason == MR_NUMA_MISPLACED);
1428 bool no_subpage_counting = false;
1430 trace_mm_migrate_pages_start(mode, reason);
1432 thp_subpage_migration:
1433 for (pass = 0; pass < 10 && (retry || thp_retry); pass++) {
1437 list_for_each_entry_safe(page, page2, from, lru) {
1440 * THP statistics is based on the source huge page.
1441 * Capture required information that might get lost
1444 is_thp = PageTransHuge(page) && !PageHuge(page);
1445 nr_subpages = compound_nr(page);
1449 rc = unmap_and_move_huge_page(get_new_page,
1450 put_new_page, private, page,
1451 pass > 2, mode, reason,
1454 rc = unmap_and_move(get_new_page, put_new_page,
1455 private, page, pass > 2, mode,
1456 reason, &ret_pages);
1459 * Success: non hugetlb page will be freed, hugetlb
1460 * page will be put back
1461 * -EAGAIN: stay on the from list
1462 * -ENOMEM: stay on the from list
1463 * Other errno: put on ret_pages list then splice to
1468 * THP migration might be unsupported or the
1469 * allocation could've failed so we should
1470 * retry on the same page with the THP split
1473 * Head page is retried immediately and tail
1474 * pages are added to the tail of the list so
1475 * we encounter them after the rest of the list
1479 /* THP migration is unsupported */
1482 if (!try_split_thp(page, &thp_split_pages)) {
1486 /* Hugetlb migration is unsupported */
1487 } else if (!no_subpage_counting) {
1491 nr_failed_pages += nr_subpages;
1495 * When memory is low, don't bother to try to migrate
1496 * other pages, just exit.
1500 /* THP NUMA faulting doesn't split THP to retry. */
1501 if (!nosplit && !try_split_thp(page, &thp_split_pages)) {
1505 } else if (!no_subpage_counting) {
1509 nr_failed_pages += nr_subpages;
1511 * There might be some subpages of fail-to-migrate THPs
1512 * left in thp_split_pages list. Move them back to migration
1513 * list so that they could be put back to the right list by
1514 * the caller otherwise the page refcnt will be leaked.
1516 list_splice_init(&thp_split_pages, from);
1517 /* nr_failed isn't updated for not used */
1518 nr_thp_failed += thp_retry;
1526 case MIGRATEPAGE_SUCCESS:
1527 nr_succeeded += nr_subpages;
1533 * Permanent failure (-EBUSY, etc.):
1534 * unlike -EAGAIN case, the failed page is
1535 * removed from migration page list and not
1536 * retried in the next outer loop.
1540 else if (!no_subpage_counting)
1543 nr_failed_pages += nr_subpages;
1549 nr_thp_failed += thp_retry;
1551 * Try to migrate subpages of fail-to-migrate THPs, no nr_failed
1552 * counting in this round, since all subpages of a THP is counted
1553 * as 1 failure in the first round.
1555 if (!list_empty(&thp_split_pages)) {
1557 * Move non-migrated pages (after 10 retries) to ret_pages
1558 * to avoid migrating them again.
1560 list_splice_init(from, &ret_pages);
1561 list_splice_init(&thp_split_pages, from);
1562 no_subpage_counting = true;
1564 goto thp_subpage_migration;
1567 rc = nr_failed + nr_thp_failed;
1570 * Put the permanent failure page back to migration list, they
1571 * will be put back to the right list by the caller.
1573 list_splice(&ret_pages, from);
1575 count_vm_events(PGMIGRATE_SUCCESS, nr_succeeded);
1576 count_vm_events(PGMIGRATE_FAIL, nr_failed_pages);
1577 count_vm_events(THP_MIGRATION_SUCCESS, nr_thp_succeeded);
1578 count_vm_events(THP_MIGRATION_FAIL, nr_thp_failed);
1579 count_vm_events(THP_MIGRATION_SPLIT, nr_thp_split);
1580 trace_mm_migrate_pages(nr_succeeded, nr_failed_pages, nr_thp_succeeded,
1581 nr_thp_failed, nr_thp_split, mode, reason);
1584 *ret_succeeded = nr_succeeded;
1589 struct page *alloc_migration_target(struct page *page, unsigned long private)
1591 struct folio *folio = page_folio(page);
1592 struct migration_target_control *mtc;
1594 unsigned int order = 0;
1595 struct folio *new_folio = NULL;
1599 mtc = (struct migration_target_control *)private;
1600 gfp_mask = mtc->gfp_mask;
1602 if (nid == NUMA_NO_NODE)
1603 nid = folio_nid(folio);
1605 if (folio_test_hugetlb(folio)) {
1606 struct hstate *h = page_hstate(&folio->page);
1608 gfp_mask = htlb_modify_alloc_mask(h, gfp_mask);
1609 return alloc_huge_page_nodemask(h, nid, mtc->nmask, gfp_mask);
1612 if (folio_test_large(folio)) {
1614 * clear __GFP_RECLAIM to make the migration callback
1615 * consistent with regular THP allocations.
1617 gfp_mask &= ~__GFP_RECLAIM;
1618 gfp_mask |= GFP_TRANSHUGE;
1619 order = folio_order(folio);
1621 zidx = zone_idx(folio_zone(folio));
1622 if (is_highmem_idx(zidx) || zidx == ZONE_MOVABLE)
1623 gfp_mask |= __GFP_HIGHMEM;
1625 new_folio = __folio_alloc(gfp_mask, order, nid, mtc->nmask);
1627 return &new_folio->page;
1632 static int store_status(int __user *status, int start, int value, int nr)
1635 if (put_user(value, status + start))
1643 static int do_move_pages_to_node(struct mm_struct *mm,
1644 struct list_head *pagelist, int node)
1647 struct migration_target_control mtc = {
1649 .gfp_mask = GFP_HIGHUSER_MOVABLE | __GFP_THISNODE,
1652 err = migrate_pages(pagelist, alloc_migration_target, NULL,
1653 (unsigned long)&mtc, MIGRATE_SYNC, MR_SYSCALL, NULL);
1655 putback_movable_pages(pagelist);
1660 * Resolves the given address to a struct page, isolates it from the LRU and
1661 * puts it to the given pagelist.
1663 * errno - if the page cannot be found/isolated
1664 * 0 - when it doesn't have to be migrated because it is already on the
1666 * 1 - when it has been queued
1668 static int add_page_for_migration(struct mm_struct *mm, unsigned long addr,
1669 int node, struct list_head *pagelist, bool migrate_all)
1671 struct vm_area_struct *vma;
1677 vma = vma_lookup(mm, addr);
1678 if (!vma || !vma_migratable(vma))
1681 /* FOLL_DUMP to ignore special (like zero) pages */
1682 page = follow_page(vma, addr, FOLL_GET | FOLL_DUMP);
1684 err = PTR_ERR(page);
1689 if (!page || is_zone_device_page(page))
1693 if (page_to_nid(page) == node)
1697 if (page_mapcount(page) > 1 && !migrate_all)
1700 if (PageHuge(page)) {
1701 if (PageHead(page)) {
1702 err = isolate_hugetlb(page, pagelist);
1709 head = compound_head(page);
1710 err = isolate_lru_page(head);
1715 list_add_tail(&head->lru, pagelist);
1716 mod_node_page_state(page_pgdat(head),
1717 NR_ISOLATED_ANON + page_is_file_lru(head),
1718 thp_nr_pages(head));
1722 * Either remove the duplicate refcount from
1723 * isolate_lru_page() or drop the page ref if it was
1728 mmap_read_unlock(mm);
1732 static int move_pages_and_store_status(struct mm_struct *mm, int node,
1733 struct list_head *pagelist, int __user *status,
1734 int start, int i, unsigned long nr_pages)
1738 if (list_empty(pagelist))
1741 err = do_move_pages_to_node(mm, pagelist, node);
1744 * Positive err means the number of failed
1745 * pages to migrate. Since we are going to
1746 * abort and return the number of non-migrated
1747 * pages, so need to include the rest of the
1748 * nr_pages that have not been attempted as
1752 err += nr_pages - i;
1755 return store_status(status, start, node, i - start);
1759 * Migrate an array of page address onto an array of nodes and fill
1760 * the corresponding array of status.
1762 static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes,
1763 unsigned long nr_pages,
1764 const void __user * __user *pages,
1765 const int __user *nodes,
1766 int __user *status, int flags)
1768 int current_node = NUMA_NO_NODE;
1769 LIST_HEAD(pagelist);
1773 lru_cache_disable();
1775 for (i = start = 0; i < nr_pages; i++) {
1776 const void __user *p;
1781 if (get_user(p, pages + i))
1783 if (get_user(node, nodes + i))
1785 addr = (unsigned long)untagged_addr(p);
1788 if (node < 0 || node >= MAX_NUMNODES)
1790 if (!node_state(node, N_MEMORY))
1794 if (!node_isset(node, task_nodes))
1797 if (current_node == NUMA_NO_NODE) {
1798 current_node = node;
1800 } else if (node != current_node) {
1801 err = move_pages_and_store_status(mm, current_node,
1802 &pagelist, status, start, i, nr_pages);
1806 current_node = node;
1810 * Errors in the page lookup or isolation are not fatal and we simply
1811 * report them via status
1813 err = add_page_for_migration(mm, addr, current_node,
1814 &pagelist, flags & MPOL_MF_MOVE_ALL);
1817 /* The page is successfully queued for migration */
1822 * The move_pages() man page does not have an -EEXIST choice, so
1823 * use -EFAULT instead.
1829 * If the page is already on the target node (!err), store the
1830 * node, otherwise, store the err.
1832 err = store_status(status, i, err ? : current_node, 1);
1836 err = move_pages_and_store_status(mm, current_node, &pagelist,
1837 status, start, i, nr_pages);
1839 /* We have accounted for page i */
1844 current_node = NUMA_NO_NODE;
1847 /* Make sure we do not overwrite the existing error */
1848 err1 = move_pages_and_store_status(mm, current_node, &pagelist,
1849 status, start, i, nr_pages);
1858 * Determine the nodes of an array of pages and store it in an array of status.
1860 static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages,
1861 const void __user **pages, int *status)
1867 for (i = 0; i < nr_pages; i++) {
1868 unsigned long addr = (unsigned long)(*pages);
1869 unsigned int foll_flags = FOLL_DUMP;
1870 struct vm_area_struct *vma;
1874 vma = vma_lookup(mm, addr);
1878 /* Not all huge page follow APIs support 'FOLL_GET' */
1879 if (!is_vm_hugetlb_page(vma))
1880 foll_flags |= FOLL_GET;
1882 /* FOLL_DUMP to ignore special (like zero) pages */
1883 page = follow_page(vma, addr, foll_flags);
1885 err = PTR_ERR(page);
1889 if (page && !is_zone_device_page(page)) {
1890 err = page_to_nid(page);
1891 if (foll_flags & FOLL_GET)
1903 mmap_read_unlock(mm);
1906 static int get_compat_pages_array(const void __user *chunk_pages[],
1907 const void __user * __user *pages,
1908 unsigned long chunk_nr)
1910 compat_uptr_t __user *pages32 = (compat_uptr_t __user *)pages;
1914 for (i = 0; i < chunk_nr; i++) {
1915 if (get_user(p, pages32 + i))
1917 chunk_pages[i] = compat_ptr(p);
1924 * Determine the nodes of a user array of pages and store it in
1925 * a user array of status.
1927 static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages,
1928 const void __user * __user *pages,
1931 #define DO_PAGES_STAT_CHUNK_NR 16UL
1932 const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR];
1933 int chunk_status[DO_PAGES_STAT_CHUNK_NR];
1936 unsigned long chunk_nr = min(nr_pages, DO_PAGES_STAT_CHUNK_NR);
1938 if (in_compat_syscall()) {
1939 if (get_compat_pages_array(chunk_pages, pages,
1943 if (copy_from_user(chunk_pages, pages,
1944 chunk_nr * sizeof(*chunk_pages)))
1948 do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status);
1950 if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status)))
1955 nr_pages -= chunk_nr;
1957 return nr_pages ? -EFAULT : 0;
1960 static struct mm_struct *find_mm_struct(pid_t pid, nodemask_t *mem_nodes)
1962 struct task_struct *task;
1963 struct mm_struct *mm;
1966 * There is no need to check if current process has the right to modify
1967 * the specified process when they are same.
1971 *mem_nodes = cpuset_mems_allowed(current);
1975 /* Find the mm_struct */
1977 task = find_task_by_vpid(pid);
1980 return ERR_PTR(-ESRCH);
1982 get_task_struct(task);
1985 * Check if this process has the right to modify the specified
1986 * process. Use the regular "ptrace_may_access()" checks.
1988 if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) {
1990 mm = ERR_PTR(-EPERM);
1995 mm = ERR_PTR(security_task_movememory(task));
1998 *mem_nodes = cpuset_mems_allowed(task);
1999 mm = get_task_mm(task);
2001 put_task_struct(task);
2003 mm = ERR_PTR(-EINVAL);
2008 * Move a list of pages in the address space of the currently executing
2011 static int kernel_move_pages(pid_t pid, unsigned long nr_pages,
2012 const void __user * __user *pages,
2013 const int __user *nodes,
2014 int __user *status, int flags)
2016 struct mm_struct *mm;
2018 nodemask_t task_nodes;
2021 if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
2024 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
2027 mm = find_mm_struct(pid, &task_nodes);
2032 err = do_pages_move(mm, task_nodes, nr_pages, pages,
2033 nodes, status, flags);
2035 err = do_pages_stat(mm, nr_pages, pages, status);
2041 SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages,
2042 const void __user * __user *, pages,
2043 const int __user *, nodes,
2044 int __user *, status, int, flags)
2046 return kernel_move_pages(pid, nr_pages, pages, nodes, status, flags);
2049 #ifdef CONFIG_NUMA_BALANCING
2051 * Returns true if this is a safe migration target node for misplaced NUMA
2052 * pages. Currently it only checks the watermarks which is crude.
2054 static bool migrate_balanced_pgdat(struct pglist_data *pgdat,
2055 unsigned long nr_migrate_pages)
2059 for (z = pgdat->nr_zones - 1; z >= 0; z--) {
2060 struct zone *zone = pgdat->node_zones + z;
2062 if (!managed_zone(zone))
2065 /* Avoid waking kswapd by allocating pages_to_migrate pages. */
2066 if (!zone_watermark_ok(zone, 0,
2067 high_wmark_pages(zone) +
2076 static struct page *alloc_misplaced_dst_page(struct page *page,
2079 int nid = (int) data;
2080 int order = compound_order(page);
2081 gfp_t gfp = __GFP_THISNODE;
2085 gfp |= GFP_TRANSHUGE_LIGHT;
2087 gfp |= GFP_HIGHUSER_MOVABLE | __GFP_NOMEMALLOC | __GFP_NORETRY |
2089 gfp &= ~__GFP_RECLAIM;
2091 new = __folio_alloc_node(gfp, order, nid);
2096 static int numamigrate_isolate_page(pg_data_t *pgdat, struct page *page)
2098 int nr_pages = thp_nr_pages(page);
2099 int order = compound_order(page);
2101 VM_BUG_ON_PAGE(order && !PageTransHuge(page), page);
2103 /* Do not migrate THP mapped by multiple processes */
2104 if (PageTransHuge(page) && total_mapcount(page) > 1)
2107 /* Avoid migrating to a node that is nearly full */
2108 if (!migrate_balanced_pgdat(pgdat, nr_pages)) {
2111 if (!(sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING))
2113 for (z = pgdat->nr_zones - 1; z >= 0; z--) {
2114 if (managed_zone(pgdat->node_zones + z))
2117 wakeup_kswapd(pgdat->node_zones + z, 0, order, ZONE_MOVABLE);
2121 if (isolate_lru_page(page))
2124 mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON + page_is_file_lru(page),
2128 * Isolating the page has taken another reference, so the
2129 * caller's reference can be safely dropped without the page
2130 * disappearing underneath us during migration.
2137 * Attempt to migrate a misplaced page to the specified destination
2138 * node. Caller is expected to have an elevated reference count on
2139 * the page that will be dropped by this function before returning.
2141 int migrate_misplaced_page(struct page *page, struct vm_area_struct *vma,
2144 pg_data_t *pgdat = NODE_DATA(node);
2147 unsigned int nr_succeeded;
2148 LIST_HEAD(migratepages);
2149 int nr_pages = thp_nr_pages(page);
2152 * Don't migrate file pages that are mapped in multiple processes
2153 * with execute permissions as they are probably shared libraries.
2155 if (page_mapcount(page) != 1 && page_is_file_lru(page) &&
2156 (vma->vm_flags & VM_EXEC))
2160 * Also do not migrate dirty pages as not all filesystems can move
2161 * dirty pages in MIGRATE_ASYNC mode which is a waste of cycles.
2163 if (page_is_file_lru(page) && PageDirty(page))
2166 isolated = numamigrate_isolate_page(pgdat, page);
2170 list_add(&page->lru, &migratepages);
2171 nr_remaining = migrate_pages(&migratepages, alloc_misplaced_dst_page,
2172 NULL, node, MIGRATE_ASYNC,
2173 MR_NUMA_MISPLACED, &nr_succeeded);
2175 if (!list_empty(&migratepages)) {
2176 list_del(&page->lru);
2177 mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON +
2178 page_is_file_lru(page), -nr_pages);
2179 putback_lru_page(page);
2184 count_vm_numa_events(NUMA_PAGE_MIGRATE, nr_succeeded);
2185 if (!node_is_toptier(page_to_nid(page)) && node_is_toptier(node))
2186 mod_node_page_state(pgdat, PGPROMOTE_SUCCESS,
2189 BUG_ON(!list_empty(&migratepages));
2196 #endif /* CONFIG_NUMA_BALANCING */
2199 * node_demotion[] example:
2201 * Consider a system with two sockets. Each socket has
2202 * three classes of memory attached: fast, medium and slow.
2203 * Each memory class is placed in its own NUMA node. The
2204 * CPUs are placed in the node with the "fast" memory. The
2205 * 6 NUMA nodes (0-5) might be split among the sockets like
2211 * When Node 0 fills up, its memory should be migrated to
2212 * Node 1. When Node 1 fills up, it should be migrated to
2213 * Node 2. The migration path start on the nodes with the
2214 * processors (since allocations default to this node) and
2215 * fast memory, progress through medium and end with the
2218 * 0 -> 1 -> 2 -> stop
2219 * 3 -> 4 -> 5 -> stop
2221 * This is represented in the node_demotion[] like this:
2223 * { nr=1, nodes[0]=1 }, // Node 0 migrates to 1
2224 * { nr=1, nodes[0]=2 }, // Node 1 migrates to 2
2225 * { nr=0, nodes[0]=-1 }, // Node 2 does not migrate
2226 * { nr=1, nodes[0]=4 }, // Node 3 migrates to 4
2227 * { nr=1, nodes[0]=5 }, // Node 4 migrates to 5
2228 * { nr=0, nodes[0]=-1 }, // Node 5 does not migrate
2230 * Moreover some systems may have multiple slow memory nodes.
2231 * Suppose a system has one socket with 3 memory nodes, node 0
2232 * is fast memory type, and node 1/2 both are slow memory
2233 * type, and the distance between fast memory node and slow
2234 * memory node is same. So the migration path should be:
2238 * This is represented in the node_demotion[] like this:
2239 * { nr=2, {nodes[0]=1, nodes[1]=2} }, // Node 0 migrates to node 1 and node 2
2240 * { nr=0, nodes[0]=-1, }, // Node 1 dose not migrate
2241 * { nr=0, nodes[0]=-1, }, // Node 2 does not migrate
2245 * Writes to this array occur without locking. Cycles are
2246 * not allowed: Node X demotes to Y which demotes to X...
2248 * If multiple reads are performed, a single rcu_read_lock()
2249 * must be held over all reads to ensure that no cycles are
2252 #define DEFAULT_DEMOTION_TARGET_NODES 15
2254 #if MAX_NUMNODES < DEFAULT_DEMOTION_TARGET_NODES
2255 #define DEMOTION_TARGET_NODES (MAX_NUMNODES - 1)
2257 #define DEMOTION_TARGET_NODES DEFAULT_DEMOTION_TARGET_NODES
2260 struct demotion_nodes {
2262 short nodes[DEMOTION_TARGET_NODES];
2265 static struct demotion_nodes *node_demotion __read_mostly;
2268 * next_demotion_node() - Get the next node in the demotion path
2269 * @node: The starting node to lookup the next node
2271 * Return: node id for next memory node in the demotion path hierarchy
2272 * from @node; NUMA_NO_NODE if @node is terminal. This does not keep
2273 * @node online or guarantee that it *continues* to be the next demotion
2276 int next_demotion_node(int node)
2278 struct demotion_nodes *nd;
2279 unsigned short target_nr, index;
2283 return NUMA_NO_NODE;
2285 nd = &node_demotion[node];
2288 * node_demotion[] is updated without excluding this
2289 * function from running. RCU doesn't provide any
2290 * compiler barriers, so the READ_ONCE() is required
2291 * to avoid compiler reordering or read merging.
2293 * Make sure to use RCU over entire code blocks if
2294 * node_demotion[] reads need to be consistent.
2297 target_nr = READ_ONCE(nd->nr);
2299 switch (target_nr) {
2301 target = NUMA_NO_NODE;
2308 * If there are multiple target nodes, just select one
2309 * target node randomly.
2311 * In addition, we can also use round-robin to select
2312 * target node, but we should introduce another variable
2313 * for node_demotion[] to record last selected target node,
2314 * that may cause cache ping-pong due to the changing of
2315 * last target node. Or introducing per-cpu data to avoid
2316 * caching issue, which seems more complicated. So selecting
2317 * target node randomly seems better until now.
2319 index = get_random_int() % target_nr;
2323 target = READ_ONCE(nd->nodes[index]);
2330 /* Disable reclaim-based migration. */
2331 static void __disable_all_migrate_targets(void)
2338 for_each_online_node(node) {
2339 node_demotion[node].nr = 0;
2340 for (i = 0; i < DEMOTION_TARGET_NODES; i++)
2341 node_demotion[node].nodes[i] = NUMA_NO_NODE;
2345 static void disable_all_migrate_targets(void)
2347 __disable_all_migrate_targets();
2350 * Ensure that the "disable" is visible across the system.
2351 * Readers will see either a combination of before+disable
2352 * state or disable+after. They will never see before and
2353 * after state together.
2355 * The before+after state together might have cycles and
2356 * could cause readers to do things like loop until this
2357 * function finishes. This ensures they can only see a
2358 * single "bad" read and would, for instance, only loop
2365 * Find an automatic demotion target for 'node'.
2366 * Failing here is OK. It might just indicate
2367 * being at the end of a chain.
2369 static int establish_migrate_target(int node, nodemask_t *used,
2372 int migration_target, index, val;
2373 struct demotion_nodes *nd;
2376 return NUMA_NO_NODE;
2378 nd = &node_demotion[node];
2380 migration_target = find_next_best_node(node, used);
2381 if (migration_target == NUMA_NO_NODE)
2382 return NUMA_NO_NODE;
2385 * If the node has been set a migration target node before,
2386 * which means it's the best distance between them. Still
2387 * check if this node can be demoted to other target nodes
2388 * if they have a same best distance.
2390 if (best_distance != -1) {
2391 val = node_distance(node, migration_target);
2392 if (val > best_distance)
2397 if (WARN_ONCE(index >= DEMOTION_TARGET_NODES,
2398 "Exceeds maximum demotion target nodes\n"))
2401 nd->nodes[index] = migration_target;
2404 return migration_target;
2406 node_clear(migration_target, *used);
2407 return NUMA_NO_NODE;
2411 * When memory fills up on a node, memory contents can be
2412 * automatically migrated to another node instead of
2413 * discarded at reclaim.
2415 * Establish a "migration path" which will start at nodes
2416 * with CPUs and will follow the priorities used to build the
2417 * page allocator zonelists.
2419 * The difference here is that cycles must be avoided. If
2420 * node0 migrates to node1, then neither node1, nor anything
2421 * node1 migrates to can migrate to node0. Also one node can
2422 * be migrated to multiple nodes if the target nodes all have
2423 * a same best-distance against the source node.
2425 * This function can run simultaneously with readers of
2426 * node_demotion[]. However, it can not run simultaneously
2427 * with itself. Exclusion is provided by memory hotplug events
2428 * being single-threaded.
2430 static void __set_migration_target_nodes(void)
2432 nodemask_t next_pass;
2433 nodemask_t this_pass;
2434 nodemask_t used_targets = NODE_MASK_NONE;
2435 int node, best_distance;
2438 * Avoid any oddities like cycles that could occur
2439 * from changes in the topology. This will leave
2440 * a momentary gap when migration is disabled.
2442 disable_all_migrate_targets();
2445 * Allocations go close to CPUs, first. Assume that
2446 * the migration path starts at the nodes with CPUs.
2448 next_pass = node_states[N_CPU];
2450 this_pass = next_pass;
2451 next_pass = NODE_MASK_NONE;
2453 * To avoid cycles in the migration "graph", ensure
2454 * that migration sources are not future targets by
2455 * setting them in 'used_targets'. Do this only
2456 * once per pass so that multiple source nodes can
2457 * share a target node.
2459 * 'used_targets' will become unavailable in future
2460 * passes. This limits some opportunities for
2461 * multiple source nodes to share a destination.
2463 nodes_or(used_targets, used_targets, this_pass);
2465 for_each_node_mask(node, this_pass) {
2469 * Try to set up the migration path for the node, and the target
2470 * migration nodes can be multiple, so doing a loop to find all
2471 * the target nodes if they all have a best node distance.
2475 establish_migrate_target(node, &used_targets,
2478 if (target_node == NUMA_NO_NODE)
2481 if (best_distance == -1)
2482 best_distance = node_distance(node, target_node);
2485 * Visit targets from this pass in the next pass.
2486 * Eventually, every node will have been part of
2487 * a pass, and will become set in 'used_targets'.
2489 node_set(target_node, next_pass);
2493 * 'next_pass' contains nodes which became migration
2494 * targets in this pass. Make additional passes until
2495 * no more migrations targets are available.
2497 if (!nodes_empty(next_pass))
2502 * For callers that do not hold get_online_mems() already.
2504 void set_migration_target_nodes(void)
2507 __set_migration_target_nodes();
2512 * This leaves migrate-on-reclaim transiently disabled between
2513 * the MEM_GOING_OFFLINE and MEM_OFFLINE events. This runs
2514 * whether reclaim-based migration is enabled or not, which
2515 * ensures that the user can turn reclaim-based migration at
2516 * any time without needing to recalculate migration targets.
2518 * These callbacks already hold get_online_mems(). That is why
2519 * __set_migration_target_nodes() can be used as opposed to
2520 * set_migration_target_nodes().
2522 #ifdef CONFIG_MEMORY_HOTPLUG
2523 static int __meminit migrate_on_reclaim_callback(struct notifier_block *self,
2524 unsigned long action, void *_arg)
2526 struct memory_notify *arg = _arg;
2529 * Only update the node migration order when a node is
2530 * changing status, like online->offline. This avoids
2531 * the overhead of synchronize_rcu() in most cases.
2533 if (arg->status_change_nid < 0)
2534 return notifier_from_errno(0);
2537 case MEM_GOING_OFFLINE:
2539 * Make sure there are not transient states where
2540 * an offline node is a migration target. This
2541 * will leave migration disabled until the offline
2542 * completes and the MEM_OFFLINE case below runs.
2544 disable_all_migrate_targets();
2549 * Recalculate the target nodes once the node
2550 * reaches its final state (online or offline).
2552 __set_migration_target_nodes();
2554 case MEM_CANCEL_OFFLINE:
2556 * MEM_GOING_OFFLINE disabled all the migration
2557 * targets. Reenable them.
2559 __set_migration_target_nodes();
2561 case MEM_GOING_ONLINE:
2562 case MEM_CANCEL_ONLINE:
2566 return notifier_from_errno(0);
2570 void __init migrate_on_reclaim_init(void)
2572 node_demotion = kcalloc(nr_node_ids,
2573 sizeof(struct demotion_nodes),
2575 WARN_ON(!node_demotion);
2576 #ifdef CONFIG_MEMORY_HOTPLUG
2577 hotplug_memory_notifier(migrate_on_reclaim_callback, 100);
2580 * At this point, all numa nodes with memory/CPus have their state
2581 * properly set, so we can build the demotion order now.
2582 * Let us hold the cpu_hotplug lock just, as we could possibily have
2583 * CPU hotplug events during boot.
2586 set_migration_target_nodes();
2590 bool numa_demotion_enabled = false;
2593 static ssize_t numa_demotion_enabled_show(struct kobject *kobj,
2594 struct kobj_attribute *attr, char *buf)
2596 return sysfs_emit(buf, "%s\n",
2597 numa_demotion_enabled ? "true" : "false");
2600 static ssize_t numa_demotion_enabled_store(struct kobject *kobj,
2601 struct kobj_attribute *attr,
2602 const char *buf, size_t count)
2606 ret = kstrtobool(buf, &numa_demotion_enabled);
2613 static struct kobj_attribute numa_demotion_enabled_attr =
2614 __ATTR(demotion_enabled, 0644, numa_demotion_enabled_show,
2615 numa_demotion_enabled_store);
2617 static struct attribute *numa_attrs[] = {
2618 &numa_demotion_enabled_attr.attr,
2622 static const struct attribute_group numa_attr_group = {
2623 .attrs = numa_attrs,
2626 static int __init numa_init_sysfs(void)
2629 struct kobject *numa_kobj;
2631 numa_kobj = kobject_create_and_add("numa", mm_kobj);
2633 pr_err("failed to create numa kobject\n");
2636 err = sysfs_create_group(numa_kobj, &numa_attr_group);
2638 pr_err("failed to register numa group\n");
2644 kobject_put(numa_kobj);
2647 subsys_initcall(numa_init_sysfs);
2648 #endif /* CONFIG_SYSFS */
2649 #endif /* CONFIG_NUMA */