1 // SPDX-License-Identifier: GPL-2.0
3 * Memory Migration functionality - linux/mm/migrate.c
5 * Copyright (C) 2006 Silicon Graphics, Inc., Christoph Lameter
7 * Page migration was first developed in the context of the memory hotplug
8 * project. The main authors of the migration code are:
10 * IWAMOTO Toshihiro <iwamoto@valinux.co.jp>
11 * Hirokazu Takahashi <taka@valinux.co.jp>
12 * Dave Hansen <haveblue@us.ibm.com>
16 #include <linux/migrate.h>
17 #include <linux/export.h>
18 #include <linux/swap.h>
19 #include <linux/swapops.h>
20 #include <linux/pagemap.h>
21 #include <linux/buffer_head.h>
22 #include <linux/mm_inline.h>
23 #include <linux/nsproxy.h>
24 #include <linux/pagevec.h>
25 #include <linux/ksm.h>
26 #include <linux/rmap.h>
27 #include <linux/topology.h>
28 #include <linux/cpu.h>
29 #include <linux/cpuset.h>
30 #include <linux/writeback.h>
31 #include <linux/mempolicy.h>
32 #include <linux/vmalloc.h>
33 #include <linux/security.h>
34 #include <linux/backing-dev.h>
35 #include <linux/compaction.h>
36 #include <linux/syscalls.h>
37 #include <linux/compat.h>
38 #include <linux/hugetlb.h>
39 #include <linux/hugetlb_cgroup.h>
40 #include <linux/gfp.h>
41 #include <linux/pfn_t.h>
42 #include <linux/memremap.h>
43 #include <linux/userfaultfd_k.h>
44 #include <linux/balloon_compaction.h>
45 #include <linux/page_idle.h>
46 #include <linux/page_owner.h>
47 #include <linux/sched/mm.h>
48 #include <linux/ptrace.h>
49 #include <linux/oom.h>
50 #include <linux/memory.h>
51 #include <linux/random.h>
52 #include <linux/sched/sysctl.h>
54 #include <asm/tlbflush.h>
56 #include <trace/events/migrate.h>
60 int isolate_movable_page(struct page *page, isolate_mode_t mode)
62 const struct movable_operations *mops;
65 * Avoid burning cycles with pages that are yet under __free_pages(),
66 * or just got freed under us.
68 * In case we 'win' a race for a movable page being freed under us and
69 * raise its refcount preventing __free_pages() from doing its job
70 * the put_page() at the end of this block will take care of
71 * release this page, thus avoiding a nasty leakage.
73 if (unlikely(!get_page_unless_zero(page)))
77 * Check PageMovable before holding a PG_lock because page's owner
78 * assumes anybody doesn't touch PG_lock of newly allocated page
79 * so unconditionally grabbing the lock ruins page's owner side.
81 if (unlikely(!__PageMovable(page)))
84 * As movable pages are not isolated from LRU lists, concurrent
85 * compaction threads can race against page migration functions
86 * as well as race against the releasing a page.
88 * In order to avoid having an already isolated movable page
89 * being (wrongly) re-isolated while it is under migration,
90 * or to avoid attempting to isolate pages being released,
91 * lets be sure we have the page lock
92 * before proceeding with the movable page isolation steps.
94 if (unlikely(!trylock_page(page)))
97 if (!PageMovable(page) || PageIsolated(page))
100 mops = page_movable_ops(page);
101 VM_BUG_ON_PAGE(!mops, page);
103 if (!mops->isolate_page(page, mode))
104 goto out_no_isolated;
106 /* Driver shouldn't use PG_isolated bit of page->flags */
107 WARN_ON_ONCE(PageIsolated(page));
108 SetPageIsolated(page);
121 static void putback_movable_page(struct page *page)
123 const struct movable_operations *mops = page_movable_ops(page);
125 mops->putback_page(page);
126 ClearPageIsolated(page);
130 * Put previously isolated pages back onto the appropriate lists
131 * from where they were once taken off for compaction/migration.
133 * This function shall be used whenever the isolated pageset has been
134 * built from lru, balloon, hugetlbfs page. See isolate_migratepages_range()
135 * and isolate_huge_page().
137 void putback_movable_pages(struct list_head *l)
142 list_for_each_entry_safe(page, page2, l, lru) {
143 if (unlikely(PageHuge(page))) {
144 putback_active_hugepage(page);
147 list_del(&page->lru);
149 * We isolated non-lru movable page so here we can use
150 * __PageMovable because LRU page's mapping cannot have
151 * PAGE_MAPPING_MOVABLE.
153 if (unlikely(__PageMovable(page))) {
154 VM_BUG_ON_PAGE(!PageIsolated(page), page);
156 if (PageMovable(page))
157 putback_movable_page(page);
159 ClearPageIsolated(page);
163 mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON +
164 page_is_file_lru(page), -thp_nr_pages(page));
165 putback_lru_page(page);
171 * Restore a potential migration pte to a working pte entry
173 static bool remove_migration_pte(struct folio *folio,
174 struct vm_area_struct *vma, unsigned long addr, void *old)
176 DEFINE_FOLIO_VMA_WALK(pvmw, old, vma, addr, PVMW_SYNC | PVMW_MIGRATION);
178 while (page_vma_mapped_walk(&pvmw)) {
179 rmap_t rmap_flags = RMAP_NONE;
183 unsigned long idx = 0;
185 /* pgoff is invalid for ksm pages, but they are never large */
186 if (folio_test_large(folio) && !folio_test_hugetlb(folio))
187 idx = linear_page_index(vma, pvmw.address) - pvmw.pgoff;
188 new = folio_page(folio, idx);
190 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
191 /* PMD-mapped THP migration entry */
193 VM_BUG_ON_FOLIO(folio_test_hugetlb(folio) ||
194 !folio_test_pmd_mappable(folio), folio);
195 remove_migration_pmd(&pvmw, new);
201 pte = pte_mkold(mk_pte(new, READ_ONCE(vma->vm_page_prot)));
202 if (pte_swp_soft_dirty(*pvmw.pte))
203 pte = pte_mksoft_dirty(pte);
206 * Recheck VMA as permissions can change since migration started
208 entry = pte_to_swp_entry(*pvmw.pte);
209 if (is_writable_migration_entry(entry))
210 pte = maybe_mkwrite(pte, vma);
211 else if (pte_swp_uffd_wp(*pvmw.pte))
212 pte = pte_mkuffd_wp(pte);
214 if (folio_test_anon(folio) && !is_readable_migration_entry(entry))
215 rmap_flags |= RMAP_EXCLUSIVE;
217 if (unlikely(is_device_private_page(new))) {
219 entry = make_writable_device_private_entry(
222 entry = make_readable_device_private_entry(
224 pte = swp_entry_to_pte(entry);
225 if (pte_swp_soft_dirty(*pvmw.pte))
226 pte = pte_swp_mksoft_dirty(pte);
227 if (pte_swp_uffd_wp(*pvmw.pte))
228 pte = pte_swp_mkuffd_wp(pte);
231 #ifdef CONFIG_HUGETLB_PAGE
232 if (folio_test_hugetlb(folio)) {
233 unsigned int shift = huge_page_shift(hstate_vma(vma));
235 pte = pte_mkhuge(pte);
236 pte = arch_make_huge_pte(pte, shift, vma->vm_flags);
237 if (folio_test_anon(folio))
238 hugepage_add_anon_rmap(new, vma, pvmw.address,
241 page_dup_file_rmap(new, true);
242 set_huge_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte);
246 if (folio_test_anon(folio))
247 page_add_anon_rmap(new, vma, pvmw.address,
250 page_add_file_rmap(new, vma, false);
251 set_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte);
253 if (vma->vm_flags & VM_LOCKED)
254 mlock_page_drain_local();
256 trace_remove_migration_pte(pvmw.address, pte_val(pte),
257 compound_order(new));
259 /* No need to invalidate - it was non-present before */
260 update_mmu_cache(vma, pvmw.address, pvmw.pte);
267 * Get rid of all migration entries and replace them by
268 * references to the indicated page.
270 void remove_migration_ptes(struct folio *src, struct folio *dst, bool locked)
272 struct rmap_walk_control rwc = {
273 .rmap_one = remove_migration_pte,
278 rmap_walk_locked(dst, &rwc);
280 rmap_walk(dst, &rwc);
284 * Something used the pte of a page under migration. We need to
285 * get to the page and wait until migration is finished.
286 * When we return from this function the fault will be retried.
288 void __migration_entry_wait(struct mm_struct *mm, pte_t *ptep,
296 if (!is_swap_pte(pte))
299 entry = pte_to_swp_entry(pte);
300 if (!is_migration_entry(entry))
303 migration_entry_wait_on_locked(entry, ptep, ptl);
306 pte_unmap_unlock(ptep, ptl);
309 void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
310 unsigned long address)
312 spinlock_t *ptl = pte_lockptr(mm, pmd);
313 pte_t *ptep = pte_offset_map(pmd, address);
314 __migration_entry_wait(mm, ptep, ptl);
317 void migration_entry_wait_huge(struct vm_area_struct *vma,
318 struct mm_struct *mm, pte_t *pte)
320 spinlock_t *ptl = huge_pte_lockptr(hstate_vma(vma), mm, pte);
321 __migration_entry_wait(mm, pte, ptl);
324 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
325 void pmd_migration_entry_wait(struct mm_struct *mm, pmd_t *pmd)
329 ptl = pmd_lock(mm, pmd);
330 if (!is_pmd_migration_entry(*pmd))
332 migration_entry_wait_on_locked(pmd_to_swp_entry(*pmd), NULL, ptl);
339 static int folio_expected_refs(struct address_space *mapping,
346 refs += folio_nr_pages(folio);
347 if (folio_test_private(folio))
354 * Replace the page in the mapping.
356 * The number of remaining references must be:
357 * 1 for anonymous pages without a mapping
358 * 2 for pages with a mapping
359 * 3 for pages with a mapping and PagePrivate/PagePrivate2 set.
361 int folio_migrate_mapping(struct address_space *mapping,
362 struct folio *newfolio, struct folio *folio, int extra_count)
364 XA_STATE(xas, &mapping->i_pages, folio_index(folio));
365 struct zone *oldzone, *newzone;
367 int expected_count = folio_expected_refs(mapping, folio) + extra_count;
368 long nr = folio_nr_pages(folio);
371 /* Anonymous page without mapping */
372 if (folio_ref_count(folio) != expected_count)
375 /* No turning back from here */
376 newfolio->index = folio->index;
377 newfolio->mapping = folio->mapping;
378 if (folio_test_swapbacked(folio))
379 __folio_set_swapbacked(newfolio);
381 return MIGRATEPAGE_SUCCESS;
384 oldzone = folio_zone(folio);
385 newzone = folio_zone(newfolio);
388 if (!folio_ref_freeze(folio, expected_count)) {
389 xas_unlock_irq(&xas);
394 * Now we know that no one else is looking at the folio:
395 * no turning back from here.
397 newfolio->index = folio->index;
398 newfolio->mapping = folio->mapping;
399 folio_ref_add(newfolio, nr); /* add cache reference */
400 if (folio_test_swapbacked(folio)) {
401 __folio_set_swapbacked(newfolio);
402 if (folio_test_swapcache(folio)) {
403 folio_set_swapcache(newfolio);
404 newfolio->private = folio_get_private(folio);
407 VM_BUG_ON_FOLIO(folio_test_swapcache(folio), folio);
410 /* Move dirty while page refs frozen and newpage not yet exposed */
411 dirty = folio_test_dirty(folio);
413 folio_clear_dirty(folio);
414 folio_set_dirty(newfolio);
417 xas_store(&xas, newfolio);
420 * Drop cache reference from old page by unfreezing
421 * to one less reference.
422 * We know this isn't the last reference.
424 folio_ref_unfreeze(folio, expected_count - nr);
427 /* Leave irq disabled to prevent preemption while updating stats */
430 * If moved to a different zone then also account
431 * the page for that zone. Other VM counters will be
432 * taken care of when we establish references to the
433 * new page and drop references to the old page.
435 * Note that anonymous pages are accounted for
436 * via NR_FILE_PAGES and NR_ANON_MAPPED if they
437 * are mapped to swap space.
439 if (newzone != oldzone) {
440 struct lruvec *old_lruvec, *new_lruvec;
441 struct mem_cgroup *memcg;
443 memcg = folio_memcg(folio);
444 old_lruvec = mem_cgroup_lruvec(memcg, oldzone->zone_pgdat);
445 new_lruvec = mem_cgroup_lruvec(memcg, newzone->zone_pgdat);
447 __mod_lruvec_state(old_lruvec, NR_FILE_PAGES, -nr);
448 __mod_lruvec_state(new_lruvec, NR_FILE_PAGES, nr);
449 if (folio_test_swapbacked(folio) && !folio_test_swapcache(folio)) {
450 __mod_lruvec_state(old_lruvec, NR_SHMEM, -nr);
451 __mod_lruvec_state(new_lruvec, NR_SHMEM, nr);
454 if (folio_test_swapcache(folio)) {
455 __mod_lruvec_state(old_lruvec, NR_SWAPCACHE, -nr);
456 __mod_lruvec_state(new_lruvec, NR_SWAPCACHE, nr);
459 if (dirty && mapping_can_writeback(mapping)) {
460 __mod_lruvec_state(old_lruvec, NR_FILE_DIRTY, -nr);
461 __mod_zone_page_state(oldzone, NR_ZONE_WRITE_PENDING, -nr);
462 __mod_lruvec_state(new_lruvec, NR_FILE_DIRTY, nr);
463 __mod_zone_page_state(newzone, NR_ZONE_WRITE_PENDING, nr);
468 return MIGRATEPAGE_SUCCESS;
470 EXPORT_SYMBOL(folio_migrate_mapping);
473 * The expected number of remaining references is the same as that
474 * of folio_migrate_mapping().
476 int migrate_huge_page_move_mapping(struct address_space *mapping,
477 struct page *newpage, struct page *page)
479 XA_STATE(xas, &mapping->i_pages, page_index(page));
483 expected_count = 2 + page_has_private(page);
484 if (!page_ref_freeze(page, expected_count)) {
485 xas_unlock_irq(&xas);
489 newpage->index = page->index;
490 newpage->mapping = page->mapping;
494 xas_store(&xas, newpage);
496 page_ref_unfreeze(page, expected_count - 1);
498 xas_unlock_irq(&xas);
500 return MIGRATEPAGE_SUCCESS;
504 * Copy the flags and some other ancillary information
506 void folio_migrate_flags(struct folio *newfolio, struct folio *folio)
510 if (folio_test_error(folio))
511 folio_set_error(newfolio);
512 if (folio_test_referenced(folio))
513 folio_set_referenced(newfolio);
514 if (folio_test_uptodate(folio))
515 folio_mark_uptodate(newfolio);
516 if (folio_test_clear_active(folio)) {
517 VM_BUG_ON_FOLIO(folio_test_unevictable(folio), folio);
518 folio_set_active(newfolio);
519 } else if (folio_test_clear_unevictable(folio))
520 folio_set_unevictable(newfolio);
521 if (folio_test_workingset(folio))
522 folio_set_workingset(newfolio);
523 if (folio_test_checked(folio))
524 folio_set_checked(newfolio);
526 * PG_anon_exclusive (-> PG_mappedtodisk) is always migrated via
527 * migration entries. We can still have PG_anon_exclusive set on an
528 * effectively unmapped and unreferenced first sub-pages of an
529 * anonymous THP: we can simply copy it here via PG_mappedtodisk.
531 if (folio_test_mappedtodisk(folio))
532 folio_set_mappedtodisk(newfolio);
534 /* Move dirty on pages not done by folio_migrate_mapping() */
535 if (folio_test_dirty(folio))
536 folio_set_dirty(newfolio);
538 if (folio_test_young(folio))
539 folio_set_young(newfolio);
540 if (folio_test_idle(folio))
541 folio_set_idle(newfolio);
544 * Copy NUMA information to the new page, to prevent over-eager
545 * future migrations of this same page.
547 cpupid = page_cpupid_xchg_last(&folio->page, -1);
548 page_cpupid_xchg_last(&newfolio->page, cpupid);
550 folio_migrate_ksm(newfolio, folio);
552 * Please do not reorder this without considering how mm/ksm.c's
553 * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache().
555 if (folio_test_swapcache(folio))
556 folio_clear_swapcache(folio);
557 folio_clear_private(folio);
559 /* page->private contains hugetlb specific flags */
560 if (!folio_test_hugetlb(folio))
561 folio->private = NULL;
564 * If any waiters have accumulated on the new page then
567 if (folio_test_writeback(newfolio))
568 folio_end_writeback(newfolio);
571 * PG_readahead shares the same bit with PG_reclaim. The above
572 * end_page_writeback() may clear PG_readahead mistakenly, so set the
575 if (folio_test_readahead(folio))
576 folio_set_readahead(newfolio);
578 folio_copy_owner(newfolio, folio);
580 if (!folio_test_hugetlb(folio))
581 mem_cgroup_migrate(folio, newfolio);
583 EXPORT_SYMBOL(folio_migrate_flags);
585 void folio_migrate_copy(struct folio *newfolio, struct folio *folio)
587 folio_copy(newfolio, folio);
588 folio_migrate_flags(newfolio, folio);
590 EXPORT_SYMBOL(folio_migrate_copy);
592 /************************************************************
593 * Migration functions
594 ***********************************************************/
597 * Common logic to directly migrate a single LRU page suitable for
598 * pages that do not use PagePrivate/PagePrivate2.
600 * Pages are locked upon entry and exit.
602 int migrate_page(struct address_space *mapping,
603 struct page *newpage, struct page *page,
604 enum migrate_mode mode)
606 struct folio *newfolio = page_folio(newpage);
607 struct folio *folio = page_folio(page);
610 BUG_ON(folio_test_writeback(folio)); /* Writeback must be complete */
612 rc = folio_migrate_mapping(mapping, newfolio, folio, 0);
614 if (rc != MIGRATEPAGE_SUCCESS)
617 if (mode != MIGRATE_SYNC_NO_COPY)
618 folio_migrate_copy(newfolio, folio);
620 folio_migrate_flags(newfolio, folio);
621 return MIGRATEPAGE_SUCCESS;
623 EXPORT_SYMBOL(migrate_page);
626 /* Returns true if all buffers are successfully locked */
627 static bool buffer_migrate_lock_buffers(struct buffer_head *head,
628 enum migrate_mode mode)
630 struct buffer_head *bh = head;
632 /* Simple case, sync compaction */
633 if (mode != MIGRATE_ASYNC) {
636 bh = bh->b_this_page;
638 } while (bh != head);
643 /* async case, we cannot block on lock_buffer so use trylock_buffer */
645 if (!trylock_buffer(bh)) {
647 * We failed to lock the buffer and cannot stall in
648 * async migration. Release the taken locks
650 struct buffer_head *failed_bh = bh;
652 while (bh != failed_bh) {
654 bh = bh->b_this_page;
659 bh = bh->b_this_page;
660 } while (bh != head);
664 static int __buffer_migrate_folio(struct address_space *mapping,
665 struct folio *dst, struct folio *src, enum migrate_mode mode,
668 struct buffer_head *bh, *head;
672 head = folio_buffers(src);
674 return migrate_page(mapping, &dst->page, &src->page, mode);
676 /* Check whether page does not have extra refs before we do more work */
677 expected_count = folio_expected_refs(mapping, src);
678 if (folio_ref_count(src) != expected_count)
681 if (!buffer_migrate_lock_buffers(head, mode))
686 bool invalidated = false;
690 spin_lock(&mapping->private_lock);
693 if (atomic_read(&bh->b_count)) {
697 bh = bh->b_this_page;
698 } while (bh != head);
704 spin_unlock(&mapping->private_lock);
705 invalidate_bh_lrus();
707 goto recheck_buffers;
711 rc = folio_migrate_mapping(mapping, dst, src, 0);
712 if (rc != MIGRATEPAGE_SUCCESS)
715 folio_attach_private(dst, folio_detach_private(src));
719 set_bh_page(bh, &dst->page, bh_offset(bh));
720 bh = bh->b_this_page;
721 } while (bh != head);
723 if (mode != MIGRATE_SYNC_NO_COPY)
724 folio_migrate_copy(dst, src);
726 folio_migrate_flags(dst, src);
728 rc = MIGRATEPAGE_SUCCESS;
731 spin_unlock(&mapping->private_lock);
735 bh = bh->b_this_page;
736 } while (bh != head);
742 * buffer_migrate_folio() - Migration function for folios with buffers.
743 * @mapping: The address space containing @src.
744 * @dst: The folio to migrate to.
745 * @src: The folio to migrate from.
746 * @mode: How to migrate the folio.
748 * This function can only be used if the underlying filesystem guarantees
749 * that no other references to @src exist. For example attached buffer
750 * heads are accessed only under the folio lock. If your filesystem cannot
751 * provide this guarantee, buffer_migrate_folio_norefs() may be more
754 * Return: 0 on success or a negative errno on failure.
756 int buffer_migrate_folio(struct address_space *mapping,
757 struct folio *dst, struct folio *src, enum migrate_mode mode)
759 return __buffer_migrate_folio(mapping, dst, src, mode, false);
761 EXPORT_SYMBOL(buffer_migrate_folio);
764 * buffer_migrate_folio_norefs() - Migration function for folios with buffers.
765 * @mapping: The address space containing @src.
766 * @dst: The folio to migrate to.
767 * @src: The folio to migrate from.
768 * @mode: How to migrate the folio.
770 * Like buffer_migrate_folio() except that this variant is more careful
771 * and checks that there are also no buffer head references. This function
772 * is the right one for mappings where buffer heads are directly looked
773 * up and referenced (such as block device mappings).
775 * Return: 0 on success or a negative errno on failure.
777 int buffer_migrate_folio_norefs(struct address_space *mapping,
778 struct folio *dst, struct folio *src, enum migrate_mode mode)
780 return __buffer_migrate_folio(mapping, dst, src, mode, true);
785 * Writeback a folio to clean the dirty state
787 static int writeout(struct address_space *mapping, struct folio *folio)
789 struct writeback_control wbc = {
790 .sync_mode = WB_SYNC_NONE,
793 .range_end = LLONG_MAX,
798 if (!mapping->a_ops->writepage)
799 /* No write method for the address space */
802 if (!folio_clear_dirty_for_io(folio))
803 /* Someone else already triggered a write */
807 * A dirty folio may imply that the underlying filesystem has
808 * the folio on some queue. So the folio must be clean for
809 * migration. Writeout may mean we lose the lock and the
810 * folio state is no longer what we checked for earlier.
811 * At this point we know that the migration attempt cannot
814 remove_migration_ptes(folio, folio, false);
816 rc = mapping->a_ops->writepage(&folio->page, &wbc);
818 if (rc != AOP_WRITEPAGE_ACTIVATE)
819 /* unlocked. Relock */
822 return (rc < 0) ? -EIO : -EAGAIN;
826 * Default handling if a filesystem does not provide a migration function.
828 static int fallback_migrate_folio(struct address_space *mapping,
829 struct folio *dst, struct folio *src, enum migrate_mode mode)
831 if (folio_test_dirty(src)) {
832 /* Only writeback folios in full synchronous migration */
835 case MIGRATE_SYNC_NO_COPY:
840 return writeout(mapping, src);
844 * Buffers may be managed in a filesystem specific way.
845 * We must have no buffers or drop them.
847 if (folio_test_private(src) &&
848 !filemap_release_folio(src, GFP_KERNEL))
849 return mode == MIGRATE_SYNC ? -EAGAIN : -EBUSY;
851 return migrate_page(mapping, &dst->page, &src->page, mode);
855 * Move a page to a newly allocated page
856 * The page is locked and all ptes have been successfully removed.
858 * The new page will have replaced the old page if this function
863 * MIGRATEPAGE_SUCCESS - success
865 static int move_to_new_folio(struct folio *dst, struct folio *src,
866 enum migrate_mode mode)
869 bool is_lru = !__PageMovable(&src->page);
871 VM_BUG_ON_FOLIO(!folio_test_locked(src), src);
872 VM_BUG_ON_FOLIO(!folio_test_locked(dst), dst);
874 if (likely(is_lru)) {
875 struct address_space *mapping = folio_mapping(src);
878 rc = migrate_page(mapping, &dst->page, &src->page, mode);
879 else if (mapping->a_ops->migrate_folio)
881 * Most folios have a mapping and most filesystems
882 * provide a migrate_folio callback. Anonymous folios
883 * are part of swap space which also has its own
884 * migrate_folio callback. This is the most common path
885 * for page migration.
887 rc = mapping->a_ops->migrate_folio(mapping, dst, src,
889 else if (mapping->a_ops->migratepage)
890 rc = mapping->a_ops->migratepage(mapping, &dst->page,
893 rc = fallback_migrate_folio(mapping, dst, src, mode);
895 const struct movable_operations *mops;
898 * In case of non-lru page, it could be released after
899 * isolation step. In that case, we shouldn't try migration.
901 VM_BUG_ON_FOLIO(!folio_test_isolated(src), src);
902 if (!folio_test_movable(src)) {
903 rc = MIGRATEPAGE_SUCCESS;
904 folio_clear_isolated(src);
908 mops = page_movable_ops(&src->page);
909 rc = mops->migrate_page(&dst->page, &src->page, mode);
910 WARN_ON_ONCE(rc == MIGRATEPAGE_SUCCESS &&
911 !folio_test_isolated(src));
915 * When successful, old pagecache src->mapping must be cleared before
916 * src is freed; but stats require that PageAnon be left as PageAnon.
918 if (rc == MIGRATEPAGE_SUCCESS) {
919 if (__PageMovable(&src->page)) {
920 VM_BUG_ON_FOLIO(!folio_test_isolated(src), src);
923 * We clear PG_movable under page_lock so any compactor
924 * cannot try to migrate this page.
926 folio_clear_isolated(src);
930 * Anonymous and movable src->mapping will be cleared by
931 * free_pages_prepare so don't reset it here for keeping
932 * the type to work PageAnon, for example.
934 if (!folio_mapping_flags(src))
937 if (likely(!folio_is_zone_device(dst)))
938 flush_dcache_folio(dst);
944 static int __unmap_and_move(struct page *page, struct page *newpage,
945 int force, enum migrate_mode mode)
947 struct folio *folio = page_folio(page);
948 struct folio *dst = page_folio(newpage);
950 bool page_was_mapped = false;
951 struct anon_vma *anon_vma = NULL;
952 bool is_lru = !__PageMovable(page);
954 if (!trylock_page(page)) {
955 if (!force || mode == MIGRATE_ASYNC)
959 * It's not safe for direct compaction to call lock_page.
960 * For example, during page readahead pages are added locked
961 * to the LRU. Later, when the IO completes the pages are
962 * marked uptodate and unlocked. However, the queueing
963 * could be merging multiple pages for one bio (e.g.
964 * mpage_readahead). If an allocation happens for the
965 * second or third page, the process can end up locking
966 * the same page twice and deadlocking. Rather than
967 * trying to be clever about what pages can be locked,
968 * avoid the use of lock_page for direct compaction
971 if (current->flags & PF_MEMALLOC)
977 if (PageWriteback(page)) {
979 * Only in the case of a full synchronous migration is it
980 * necessary to wait for PageWriteback. In the async case,
981 * the retry loop is too short and in the sync-light case,
982 * the overhead of stalling is too much
986 case MIGRATE_SYNC_NO_COPY:
994 wait_on_page_writeback(page);
998 * By try_to_migrate(), page->mapcount goes down to 0 here. In this case,
999 * we cannot notice that anon_vma is freed while we migrates a page.
1000 * This get_anon_vma() delays freeing anon_vma pointer until the end
1001 * of migration. File cache pages are no problem because of page_lock()
1002 * File Caches may use write_page() or lock_page() in migration, then,
1003 * just care Anon page here.
1005 * Only page_get_anon_vma() understands the subtleties of
1006 * getting a hold on an anon_vma from outside one of its mms.
1007 * But if we cannot get anon_vma, then we won't need it anyway,
1008 * because that implies that the anon page is no longer mapped
1009 * (and cannot be remapped so long as we hold the page lock).
1011 if (PageAnon(page) && !PageKsm(page))
1012 anon_vma = page_get_anon_vma(page);
1015 * Block others from accessing the new page when we get around to
1016 * establishing additional references. We are usually the only one
1017 * holding a reference to newpage at this point. We used to have a BUG
1018 * here if trylock_page(newpage) fails, but would like to allow for
1019 * cases where there might be a race with the previous use of newpage.
1020 * This is much like races on refcount of oldpage: just don't BUG().
1022 if (unlikely(!trylock_page(newpage)))
1025 if (unlikely(!is_lru)) {
1026 rc = move_to_new_folio(dst, folio, mode);
1027 goto out_unlock_both;
1031 * Corner case handling:
1032 * 1. When a new swap-cache page is read into, it is added to the LRU
1033 * and treated as swapcache but it has no rmap yet.
1034 * Calling try_to_unmap() against a page->mapping==NULL page will
1035 * trigger a BUG. So handle it here.
1036 * 2. An orphaned page (see truncate_cleanup_page) might have
1037 * fs-private metadata. The page can be picked up due to memory
1038 * offlining. Everywhere else except page reclaim, the page is
1039 * invisible to the vm, so the page can not be migrated. So try to
1040 * free the metadata, so the page can be freed.
1042 if (!page->mapping) {
1043 VM_BUG_ON_PAGE(PageAnon(page), page);
1044 if (page_has_private(page)) {
1045 try_to_free_buffers(folio);
1046 goto out_unlock_both;
1048 } else if (page_mapped(page)) {
1049 /* Establish migration ptes */
1050 VM_BUG_ON_PAGE(PageAnon(page) && !PageKsm(page) && !anon_vma,
1052 try_to_migrate(folio, 0);
1053 page_was_mapped = true;
1056 if (!page_mapped(page))
1057 rc = move_to_new_folio(dst, folio, mode);
1060 * When successful, push newpage to LRU immediately: so that if it
1061 * turns out to be an mlocked page, remove_migration_ptes() will
1062 * automatically build up the correct newpage->mlock_count for it.
1064 * We would like to do something similar for the old page, when
1065 * unsuccessful, and other cases when a page has been temporarily
1066 * isolated from the unevictable LRU: but this case is the easiest.
1068 if (rc == MIGRATEPAGE_SUCCESS) {
1069 lru_cache_add(newpage);
1070 if (page_was_mapped)
1074 if (page_was_mapped)
1075 remove_migration_ptes(folio,
1076 rc == MIGRATEPAGE_SUCCESS ? dst : folio, false);
1079 unlock_page(newpage);
1081 /* Drop an anon_vma reference if we took one */
1083 put_anon_vma(anon_vma);
1087 * If migration is successful, decrease refcount of the newpage,
1088 * which will not free the page because new page owner increased
1091 if (rc == MIGRATEPAGE_SUCCESS)
1098 * Obtain the lock on page, remove all ptes and migrate the page
1099 * to the newly allocated page in newpage.
1101 static int unmap_and_move(new_page_t get_new_page,
1102 free_page_t put_new_page,
1103 unsigned long private, struct page *page,
1104 int force, enum migrate_mode mode,
1105 enum migrate_reason reason,
1106 struct list_head *ret)
1108 int rc = MIGRATEPAGE_SUCCESS;
1109 struct page *newpage = NULL;
1111 if (!thp_migration_supported() && PageTransHuge(page))
1114 if (page_count(page) == 1) {
1115 /* page was freed from under us. So we are done. */
1116 ClearPageActive(page);
1117 ClearPageUnevictable(page);
1118 if (unlikely(__PageMovable(page))) {
1120 if (!PageMovable(page))
1121 ClearPageIsolated(page);
1127 newpage = get_new_page(page, private);
1131 newpage->private = 0;
1132 rc = __unmap_and_move(page, newpage, force, mode);
1133 if (rc == MIGRATEPAGE_SUCCESS)
1134 set_page_owner_migrate_reason(newpage, reason);
1137 if (rc != -EAGAIN) {
1139 * A page that has been migrated has all references
1140 * removed and will be freed. A page that has not been
1141 * migrated will have kept its references and be restored.
1143 list_del(&page->lru);
1147 * If migration is successful, releases reference grabbed during
1148 * isolation. Otherwise, restore the page to right list unless
1151 if (rc == MIGRATEPAGE_SUCCESS) {
1153 * Compaction can migrate also non-LRU pages which are
1154 * not accounted to NR_ISOLATED_*. They can be recognized
1157 if (likely(!__PageMovable(page)))
1158 mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON +
1159 page_is_file_lru(page), -thp_nr_pages(page));
1161 if (reason != MR_MEMORY_FAILURE)
1163 * We release the page in page_handle_poison.
1168 list_add_tail(&page->lru, ret);
1171 put_new_page(newpage, private);
1180 * Counterpart of unmap_and_move_page() for hugepage migration.
1182 * This function doesn't wait the completion of hugepage I/O
1183 * because there is no race between I/O and migration for hugepage.
1184 * Note that currently hugepage I/O occurs only in direct I/O
1185 * where no lock is held and PG_writeback is irrelevant,
1186 * and writeback status of all subpages are counted in the reference
1187 * count of the head page (i.e. if all subpages of a 2MB hugepage are
1188 * under direct I/O, the reference of the head page is 512 and a bit more.)
1189 * This means that when we try to migrate hugepage whose subpages are
1190 * doing direct I/O, some references remain after try_to_unmap() and
1191 * hugepage migration fails without data corruption.
1193 * There is also no race when direct I/O is issued on the page under migration,
1194 * because then pte is replaced with migration swap entry and direct I/O code
1195 * will wait in the page fault for migration to complete.
1197 static int unmap_and_move_huge_page(new_page_t get_new_page,
1198 free_page_t put_new_page, unsigned long private,
1199 struct page *hpage, int force,
1200 enum migrate_mode mode, int reason,
1201 struct list_head *ret)
1203 struct folio *dst, *src = page_folio(hpage);
1205 int page_was_mapped = 0;
1206 struct page *new_hpage;
1207 struct anon_vma *anon_vma = NULL;
1208 struct address_space *mapping = NULL;
1211 * Migratability of hugepages depends on architectures and their size.
1212 * This check is necessary because some callers of hugepage migration
1213 * like soft offline and memory hotremove don't walk through page
1214 * tables or check whether the hugepage is pmd-based or not before
1215 * kicking migration.
1217 if (!hugepage_migration_supported(page_hstate(hpage))) {
1218 list_move_tail(&hpage->lru, ret);
1222 if (page_count(hpage) == 1) {
1223 /* page was freed from under us. So we are done. */
1224 putback_active_hugepage(hpage);
1225 return MIGRATEPAGE_SUCCESS;
1228 new_hpage = get_new_page(hpage, private);
1231 dst = page_folio(new_hpage);
1233 if (!trylock_page(hpage)) {
1238 case MIGRATE_SYNC_NO_COPY:
1247 * Check for pages which are in the process of being freed. Without
1248 * page_mapping() set, hugetlbfs specific move page routine will not
1249 * be called and we could leak usage counts for subpools.
1251 if (hugetlb_page_subpool(hpage) && !page_mapping(hpage)) {
1256 if (PageAnon(hpage))
1257 anon_vma = page_get_anon_vma(hpage);
1259 if (unlikely(!trylock_page(new_hpage)))
1262 if (page_mapped(hpage)) {
1263 enum ttu_flags ttu = 0;
1265 if (!PageAnon(hpage)) {
1267 * In shared mappings, try_to_unmap could potentially
1268 * call huge_pmd_unshare. Because of this, take
1269 * semaphore in write mode here and set TTU_RMAP_LOCKED
1270 * to let lower levels know we have taken the lock.
1272 mapping = hugetlb_page_mapping_lock_write(hpage);
1273 if (unlikely(!mapping))
1274 goto unlock_put_anon;
1276 ttu = TTU_RMAP_LOCKED;
1279 try_to_migrate(src, ttu);
1280 page_was_mapped = 1;
1282 if (ttu & TTU_RMAP_LOCKED)
1283 i_mmap_unlock_write(mapping);
1286 if (!page_mapped(hpage))
1287 rc = move_to_new_folio(dst, src, mode);
1289 if (page_was_mapped)
1290 remove_migration_ptes(src,
1291 rc == MIGRATEPAGE_SUCCESS ? dst : src, false);
1294 unlock_page(new_hpage);
1298 put_anon_vma(anon_vma);
1300 if (rc == MIGRATEPAGE_SUCCESS) {
1301 move_hugetlb_state(hpage, new_hpage, reason);
1302 put_new_page = NULL;
1308 if (rc == MIGRATEPAGE_SUCCESS)
1309 putback_active_hugepage(hpage);
1310 else if (rc != -EAGAIN)
1311 list_move_tail(&hpage->lru, ret);
1314 * If migration was not successful and there's a freeing callback, use
1315 * it. Otherwise, put_page() will drop the reference grabbed during
1319 put_new_page(new_hpage, private);
1321 putback_active_hugepage(new_hpage);
1326 static inline int try_split_thp(struct page *page, struct page **page2,
1327 struct list_head *from)
1332 rc = split_huge_page_to_list(page, from);
1335 list_safe_reset_next(page, *page2, lru);
1341 * migrate_pages - migrate the pages specified in a list, to the free pages
1342 * supplied as the target for the page migration
1344 * @from: The list of pages to be migrated.
1345 * @get_new_page: The function used to allocate free pages to be used
1346 * as the target of the page migration.
1347 * @put_new_page: The function used to free target pages if migration
1348 * fails, or NULL if no special handling is necessary.
1349 * @private: Private data to be passed on to get_new_page()
1350 * @mode: The migration mode that specifies the constraints for
1351 * page migration, if any.
1352 * @reason: The reason for page migration.
1353 * @ret_succeeded: Set to the number of normal pages migrated successfully if
1354 * the caller passes a non-NULL pointer.
1356 * The function returns after 10 attempts or if no pages are movable any more
1357 * because the list has become empty or no retryable pages exist any more.
1358 * It is caller's responsibility to call putback_movable_pages() to return pages
1359 * to the LRU or free list only if ret != 0.
1361 * Returns the number of {normal page, THP, hugetlb} that were not migrated, or
1362 * an error code. The number of THP splits will be considered as the number of
1363 * non-migrated THP, no matter how many subpages of the THP are migrated successfully.
1365 int migrate_pages(struct list_head *from, new_page_t get_new_page,
1366 free_page_t put_new_page, unsigned long private,
1367 enum migrate_mode mode, int reason, unsigned int *ret_succeeded)
1372 int nr_failed_pages = 0;
1373 int nr_succeeded = 0;
1374 int nr_thp_succeeded = 0;
1375 int nr_thp_failed = 0;
1376 int nr_thp_split = 0;
1378 bool is_thp = false;
1381 int rc, nr_subpages;
1382 LIST_HEAD(ret_pages);
1383 LIST_HEAD(thp_split_pages);
1384 bool nosplit = (reason == MR_NUMA_MISPLACED);
1385 bool no_subpage_counting = false;
1387 trace_mm_migrate_pages_start(mode, reason);
1389 thp_subpage_migration:
1390 for (pass = 0; pass < 10 && (retry || thp_retry); pass++) {
1394 list_for_each_entry_safe(page, page2, from, lru) {
1397 * THP statistics is based on the source huge page.
1398 * Capture required information that might get lost
1401 is_thp = PageTransHuge(page) && !PageHuge(page);
1402 nr_subpages = compound_nr(page);
1406 rc = unmap_and_move_huge_page(get_new_page,
1407 put_new_page, private, page,
1408 pass > 2, mode, reason,
1411 rc = unmap_and_move(get_new_page, put_new_page,
1412 private, page, pass > 2, mode,
1413 reason, &ret_pages);
1416 * Success: non hugetlb page will be freed, hugetlb
1417 * page will be put back
1418 * -EAGAIN: stay on the from list
1419 * -ENOMEM: stay on the from list
1420 * Other errno: put on ret_pages list then splice to
1425 * THP migration might be unsupported or the
1426 * allocation could've failed so we should
1427 * retry on the same page with the THP split
1430 * Head page is retried immediately and tail
1431 * pages are added to the tail of the list so
1432 * we encounter them after the rest of the list
1436 /* THP migration is unsupported */
1439 if (!try_split_thp(page, &page2, &thp_split_pages)) {
1443 /* Hugetlb migration is unsupported */
1444 } else if (!no_subpage_counting) {
1448 nr_failed_pages += nr_subpages;
1452 * When memory is low, don't bother to try to migrate
1453 * other pages, just exit.
1454 * THP NUMA faulting doesn't split THP to retry.
1456 if (is_thp && !nosplit) {
1458 if (!try_split_thp(page, &page2, &thp_split_pages)) {
1462 } else if (!no_subpage_counting) {
1466 nr_failed_pages += nr_subpages;
1468 * There might be some subpages of fail-to-migrate THPs
1469 * left in thp_split_pages list. Move them back to migration
1470 * list so that they could be put back to the right list by
1471 * the caller otherwise the page refcnt will be leaked.
1473 list_splice_init(&thp_split_pages, from);
1474 nr_thp_failed += thp_retry;
1482 case MIGRATEPAGE_SUCCESS:
1483 nr_succeeded += nr_subpages;
1489 * Permanent failure (-EBUSY, etc.):
1490 * unlike -EAGAIN case, the failed page is
1491 * removed from migration page list and not
1492 * retried in the next outer loop.
1496 else if (!no_subpage_counting)
1499 nr_failed_pages += nr_subpages;
1505 nr_thp_failed += thp_retry;
1507 * Try to migrate subpages of fail-to-migrate THPs, no nr_failed
1508 * counting in this round, since all subpages of a THP is counted
1509 * as 1 failure in the first round.
1511 if (!list_empty(&thp_split_pages)) {
1513 * Move non-migrated pages (after 10 retries) to ret_pages
1514 * to avoid migrating them again.
1516 list_splice_init(from, &ret_pages);
1517 list_splice_init(&thp_split_pages, from);
1518 no_subpage_counting = true;
1520 goto thp_subpage_migration;
1523 rc = nr_failed + nr_thp_failed;
1526 * Put the permanent failure page back to migration list, they
1527 * will be put back to the right list by the caller.
1529 list_splice(&ret_pages, from);
1531 count_vm_events(PGMIGRATE_SUCCESS, nr_succeeded);
1532 count_vm_events(PGMIGRATE_FAIL, nr_failed_pages);
1533 count_vm_events(THP_MIGRATION_SUCCESS, nr_thp_succeeded);
1534 count_vm_events(THP_MIGRATION_FAIL, nr_thp_failed);
1535 count_vm_events(THP_MIGRATION_SPLIT, nr_thp_split);
1536 trace_mm_migrate_pages(nr_succeeded, nr_failed_pages, nr_thp_succeeded,
1537 nr_thp_failed, nr_thp_split, mode, reason);
1540 *ret_succeeded = nr_succeeded;
1545 struct page *alloc_migration_target(struct page *page, unsigned long private)
1547 struct folio *folio = page_folio(page);
1548 struct migration_target_control *mtc;
1550 unsigned int order = 0;
1551 struct folio *new_folio = NULL;
1555 mtc = (struct migration_target_control *)private;
1556 gfp_mask = mtc->gfp_mask;
1558 if (nid == NUMA_NO_NODE)
1559 nid = folio_nid(folio);
1561 if (folio_test_hugetlb(folio)) {
1562 struct hstate *h = page_hstate(&folio->page);
1564 gfp_mask = htlb_modify_alloc_mask(h, gfp_mask);
1565 return alloc_huge_page_nodemask(h, nid, mtc->nmask, gfp_mask);
1568 if (folio_test_large(folio)) {
1570 * clear __GFP_RECLAIM to make the migration callback
1571 * consistent with regular THP allocations.
1573 gfp_mask &= ~__GFP_RECLAIM;
1574 gfp_mask |= GFP_TRANSHUGE;
1575 order = folio_order(folio);
1577 zidx = zone_idx(folio_zone(folio));
1578 if (is_highmem_idx(zidx) || zidx == ZONE_MOVABLE)
1579 gfp_mask |= __GFP_HIGHMEM;
1581 new_folio = __folio_alloc(gfp_mask, order, nid, mtc->nmask);
1583 return &new_folio->page;
1588 static int store_status(int __user *status, int start, int value, int nr)
1591 if (put_user(value, status + start))
1599 static int do_move_pages_to_node(struct mm_struct *mm,
1600 struct list_head *pagelist, int node)
1603 struct migration_target_control mtc = {
1605 .gfp_mask = GFP_HIGHUSER_MOVABLE | __GFP_THISNODE,
1608 err = migrate_pages(pagelist, alloc_migration_target, NULL,
1609 (unsigned long)&mtc, MIGRATE_SYNC, MR_SYSCALL, NULL);
1611 putback_movable_pages(pagelist);
1616 * Resolves the given address to a struct page, isolates it from the LRU and
1617 * puts it to the given pagelist.
1619 * errno - if the page cannot be found/isolated
1620 * 0 - when it doesn't have to be migrated because it is already on the
1622 * 1 - when it has been queued
1624 static int add_page_for_migration(struct mm_struct *mm, unsigned long addr,
1625 int node, struct list_head *pagelist, bool migrate_all)
1627 struct vm_area_struct *vma;
1633 vma = vma_lookup(mm, addr);
1634 if (!vma || !vma_migratable(vma))
1637 /* FOLL_DUMP to ignore special (like zero) pages */
1638 page = follow_page(vma, addr, FOLL_GET | FOLL_DUMP);
1640 err = PTR_ERR(page);
1649 if (page_to_nid(page) == node)
1653 if (page_mapcount(page) > 1 && !migrate_all)
1656 if (PageHuge(page)) {
1657 if (PageHead(page)) {
1658 isolate_huge_page(page, pagelist);
1664 head = compound_head(page);
1665 err = isolate_lru_page(head);
1670 list_add_tail(&head->lru, pagelist);
1671 mod_node_page_state(page_pgdat(head),
1672 NR_ISOLATED_ANON + page_is_file_lru(head),
1673 thp_nr_pages(head));
1677 * Either remove the duplicate refcount from
1678 * isolate_lru_page() or drop the page ref if it was
1683 mmap_read_unlock(mm);
1687 static int move_pages_and_store_status(struct mm_struct *mm, int node,
1688 struct list_head *pagelist, int __user *status,
1689 int start, int i, unsigned long nr_pages)
1693 if (list_empty(pagelist))
1696 err = do_move_pages_to_node(mm, pagelist, node);
1699 * Positive err means the number of failed
1700 * pages to migrate. Since we are going to
1701 * abort and return the number of non-migrated
1702 * pages, so need to include the rest of the
1703 * nr_pages that have not been attempted as
1707 err += nr_pages - i - 1;
1710 return store_status(status, start, node, i - start);
1714 * Migrate an array of page address onto an array of nodes and fill
1715 * the corresponding array of status.
1717 static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes,
1718 unsigned long nr_pages,
1719 const void __user * __user *pages,
1720 const int __user *nodes,
1721 int __user *status, int flags)
1723 int current_node = NUMA_NO_NODE;
1724 LIST_HEAD(pagelist);
1728 lru_cache_disable();
1730 for (i = start = 0; i < nr_pages; i++) {
1731 const void __user *p;
1736 if (get_user(p, pages + i))
1738 if (get_user(node, nodes + i))
1740 addr = (unsigned long)untagged_addr(p);
1743 if (node < 0 || node >= MAX_NUMNODES)
1745 if (!node_state(node, N_MEMORY))
1749 if (!node_isset(node, task_nodes))
1752 if (current_node == NUMA_NO_NODE) {
1753 current_node = node;
1755 } else if (node != current_node) {
1756 err = move_pages_and_store_status(mm, current_node,
1757 &pagelist, status, start, i, nr_pages);
1761 current_node = node;
1765 * Errors in the page lookup or isolation are not fatal and we simply
1766 * report them via status
1768 err = add_page_for_migration(mm, addr, current_node,
1769 &pagelist, flags & MPOL_MF_MOVE_ALL);
1772 /* The page is successfully queued for migration */
1777 * The move_pages() man page does not have an -EEXIST choice, so
1778 * use -EFAULT instead.
1784 * If the page is already on the target node (!err), store the
1785 * node, otherwise, store the err.
1787 err = store_status(status, i, err ? : current_node, 1);
1791 err = move_pages_and_store_status(mm, current_node, &pagelist,
1792 status, start, i, nr_pages);
1795 current_node = NUMA_NO_NODE;
1798 /* Make sure we do not overwrite the existing error */
1799 err1 = move_pages_and_store_status(mm, current_node, &pagelist,
1800 status, start, i, nr_pages);
1809 * Determine the nodes of an array of pages and store it in an array of status.
1811 static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages,
1812 const void __user **pages, int *status)
1818 for (i = 0; i < nr_pages; i++) {
1819 unsigned long addr = (unsigned long)(*pages);
1820 struct vm_area_struct *vma;
1824 vma = vma_lookup(mm, addr);
1828 /* FOLL_DUMP to ignore special (like zero) pages */
1829 page = follow_page(vma, addr, FOLL_GET | FOLL_DUMP);
1831 err = PTR_ERR(page);
1836 err = page_to_nid(page);
1848 mmap_read_unlock(mm);
1851 static int get_compat_pages_array(const void __user *chunk_pages[],
1852 const void __user * __user *pages,
1853 unsigned long chunk_nr)
1855 compat_uptr_t __user *pages32 = (compat_uptr_t __user *)pages;
1859 for (i = 0; i < chunk_nr; i++) {
1860 if (get_user(p, pages32 + i))
1862 chunk_pages[i] = compat_ptr(p);
1869 * Determine the nodes of a user array of pages and store it in
1870 * a user array of status.
1872 static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages,
1873 const void __user * __user *pages,
1876 #define DO_PAGES_STAT_CHUNK_NR 16UL
1877 const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR];
1878 int chunk_status[DO_PAGES_STAT_CHUNK_NR];
1881 unsigned long chunk_nr = min(nr_pages, DO_PAGES_STAT_CHUNK_NR);
1883 if (in_compat_syscall()) {
1884 if (get_compat_pages_array(chunk_pages, pages,
1888 if (copy_from_user(chunk_pages, pages,
1889 chunk_nr * sizeof(*chunk_pages)))
1893 do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status);
1895 if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status)))
1900 nr_pages -= chunk_nr;
1902 return nr_pages ? -EFAULT : 0;
1905 static struct mm_struct *find_mm_struct(pid_t pid, nodemask_t *mem_nodes)
1907 struct task_struct *task;
1908 struct mm_struct *mm;
1911 * There is no need to check if current process has the right to modify
1912 * the specified process when they are same.
1916 *mem_nodes = cpuset_mems_allowed(current);
1920 /* Find the mm_struct */
1922 task = find_task_by_vpid(pid);
1925 return ERR_PTR(-ESRCH);
1927 get_task_struct(task);
1930 * Check if this process has the right to modify the specified
1931 * process. Use the regular "ptrace_may_access()" checks.
1933 if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) {
1935 mm = ERR_PTR(-EPERM);
1940 mm = ERR_PTR(security_task_movememory(task));
1943 *mem_nodes = cpuset_mems_allowed(task);
1944 mm = get_task_mm(task);
1946 put_task_struct(task);
1948 mm = ERR_PTR(-EINVAL);
1953 * Move a list of pages in the address space of the currently executing
1956 static int kernel_move_pages(pid_t pid, unsigned long nr_pages,
1957 const void __user * __user *pages,
1958 const int __user *nodes,
1959 int __user *status, int flags)
1961 struct mm_struct *mm;
1963 nodemask_t task_nodes;
1966 if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
1969 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1972 mm = find_mm_struct(pid, &task_nodes);
1977 err = do_pages_move(mm, task_nodes, nr_pages, pages,
1978 nodes, status, flags);
1980 err = do_pages_stat(mm, nr_pages, pages, status);
1986 SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages,
1987 const void __user * __user *, pages,
1988 const int __user *, nodes,
1989 int __user *, status, int, flags)
1991 return kernel_move_pages(pid, nr_pages, pages, nodes, status, flags);
1994 #ifdef CONFIG_NUMA_BALANCING
1996 * Returns true if this is a safe migration target node for misplaced NUMA
1997 * pages. Currently it only checks the watermarks which is crude.
1999 static bool migrate_balanced_pgdat(struct pglist_data *pgdat,
2000 unsigned long nr_migrate_pages)
2004 for (z = pgdat->nr_zones - 1; z >= 0; z--) {
2005 struct zone *zone = pgdat->node_zones + z;
2007 if (!managed_zone(zone))
2010 /* Avoid waking kswapd by allocating pages_to_migrate pages. */
2011 if (!zone_watermark_ok(zone, 0,
2012 high_wmark_pages(zone) +
2021 static struct page *alloc_misplaced_dst_page(struct page *page,
2024 int nid = (int) data;
2025 int order = compound_order(page);
2026 gfp_t gfp = __GFP_THISNODE;
2030 gfp |= GFP_TRANSHUGE_LIGHT;
2032 gfp |= GFP_HIGHUSER_MOVABLE | __GFP_NOMEMALLOC | __GFP_NORETRY |
2034 gfp &= ~__GFP_RECLAIM;
2036 new = __folio_alloc_node(gfp, order, nid);
2041 static int numamigrate_isolate_page(pg_data_t *pgdat, struct page *page)
2043 int nr_pages = thp_nr_pages(page);
2044 int order = compound_order(page);
2046 VM_BUG_ON_PAGE(order && !PageTransHuge(page), page);
2048 /* Do not migrate THP mapped by multiple processes */
2049 if (PageTransHuge(page) && total_mapcount(page) > 1)
2052 /* Avoid migrating to a node that is nearly full */
2053 if (!migrate_balanced_pgdat(pgdat, nr_pages)) {
2056 if (!(sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING))
2058 for (z = pgdat->nr_zones - 1; z >= 0; z--) {
2059 if (managed_zone(pgdat->node_zones + z))
2062 wakeup_kswapd(pgdat->node_zones + z, 0, order, ZONE_MOVABLE);
2066 if (isolate_lru_page(page))
2069 mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON + page_is_file_lru(page),
2073 * Isolating the page has taken another reference, so the
2074 * caller's reference can be safely dropped without the page
2075 * disappearing underneath us during migration.
2082 * Attempt to migrate a misplaced page to the specified destination
2083 * node. Caller is expected to have an elevated reference count on
2084 * the page that will be dropped by this function before returning.
2086 int migrate_misplaced_page(struct page *page, struct vm_area_struct *vma,
2089 pg_data_t *pgdat = NODE_DATA(node);
2092 unsigned int nr_succeeded;
2093 LIST_HEAD(migratepages);
2094 int nr_pages = thp_nr_pages(page);
2097 * Don't migrate file pages that are mapped in multiple processes
2098 * with execute permissions as they are probably shared libraries.
2100 if (page_mapcount(page) != 1 && page_is_file_lru(page) &&
2101 (vma->vm_flags & VM_EXEC))
2105 * Also do not migrate dirty pages as not all filesystems can move
2106 * dirty pages in MIGRATE_ASYNC mode which is a waste of cycles.
2108 if (page_is_file_lru(page) && PageDirty(page))
2111 isolated = numamigrate_isolate_page(pgdat, page);
2115 list_add(&page->lru, &migratepages);
2116 nr_remaining = migrate_pages(&migratepages, alloc_misplaced_dst_page,
2117 NULL, node, MIGRATE_ASYNC,
2118 MR_NUMA_MISPLACED, &nr_succeeded);
2120 if (!list_empty(&migratepages)) {
2121 list_del(&page->lru);
2122 mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON +
2123 page_is_file_lru(page), -nr_pages);
2124 putback_lru_page(page);
2129 count_vm_numa_events(NUMA_PAGE_MIGRATE, nr_succeeded);
2130 if (!node_is_toptier(page_to_nid(page)) && node_is_toptier(node))
2131 mod_node_page_state(pgdat, PGPROMOTE_SUCCESS,
2134 BUG_ON(!list_empty(&migratepages));
2141 #endif /* CONFIG_NUMA_BALANCING */
2144 * node_demotion[] example:
2146 * Consider a system with two sockets. Each socket has
2147 * three classes of memory attached: fast, medium and slow.
2148 * Each memory class is placed in its own NUMA node. The
2149 * CPUs are placed in the node with the "fast" memory. The
2150 * 6 NUMA nodes (0-5) might be split among the sockets like
2156 * When Node 0 fills up, its memory should be migrated to
2157 * Node 1. When Node 1 fills up, it should be migrated to
2158 * Node 2. The migration path start on the nodes with the
2159 * processors (since allocations default to this node) and
2160 * fast memory, progress through medium and end with the
2163 * 0 -> 1 -> 2 -> stop
2164 * 3 -> 4 -> 5 -> stop
2166 * This is represented in the node_demotion[] like this:
2168 * { nr=1, nodes[0]=1 }, // Node 0 migrates to 1
2169 * { nr=1, nodes[0]=2 }, // Node 1 migrates to 2
2170 * { nr=0, nodes[0]=-1 }, // Node 2 does not migrate
2171 * { nr=1, nodes[0]=4 }, // Node 3 migrates to 4
2172 * { nr=1, nodes[0]=5 }, // Node 4 migrates to 5
2173 * { nr=0, nodes[0]=-1 }, // Node 5 does not migrate
2175 * Moreover some systems may have multiple slow memory nodes.
2176 * Suppose a system has one socket with 3 memory nodes, node 0
2177 * is fast memory type, and node 1/2 both are slow memory
2178 * type, and the distance between fast memory node and slow
2179 * memory node is same. So the migration path should be:
2183 * This is represented in the node_demotion[] like this:
2184 * { nr=2, {nodes[0]=1, nodes[1]=2} }, // Node 0 migrates to node 1 and node 2
2185 * { nr=0, nodes[0]=-1, }, // Node 1 dose not migrate
2186 * { nr=0, nodes[0]=-1, }, // Node 2 does not migrate
2190 * Writes to this array occur without locking. Cycles are
2191 * not allowed: Node X demotes to Y which demotes to X...
2193 * If multiple reads are performed, a single rcu_read_lock()
2194 * must be held over all reads to ensure that no cycles are
2197 #define DEFAULT_DEMOTION_TARGET_NODES 15
2199 #if MAX_NUMNODES < DEFAULT_DEMOTION_TARGET_NODES
2200 #define DEMOTION_TARGET_NODES (MAX_NUMNODES - 1)
2202 #define DEMOTION_TARGET_NODES DEFAULT_DEMOTION_TARGET_NODES
2205 struct demotion_nodes {
2207 short nodes[DEMOTION_TARGET_NODES];
2210 static struct demotion_nodes *node_demotion __read_mostly;
2213 * next_demotion_node() - Get the next node in the demotion path
2214 * @node: The starting node to lookup the next node
2216 * Return: node id for next memory node in the demotion path hierarchy
2217 * from @node; NUMA_NO_NODE if @node is terminal. This does not keep
2218 * @node online or guarantee that it *continues* to be the next demotion
2221 int next_demotion_node(int node)
2223 struct demotion_nodes *nd;
2224 unsigned short target_nr, index;
2228 return NUMA_NO_NODE;
2230 nd = &node_demotion[node];
2233 * node_demotion[] is updated without excluding this
2234 * function from running. RCU doesn't provide any
2235 * compiler barriers, so the READ_ONCE() is required
2236 * to avoid compiler reordering or read merging.
2238 * Make sure to use RCU over entire code blocks if
2239 * node_demotion[] reads need to be consistent.
2242 target_nr = READ_ONCE(nd->nr);
2244 switch (target_nr) {
2246 target = NUMA_NO_NODE;
2253 * If there are multiple target nodes, just select one
2254 * target node randomly.
2256 * In addition, we can also use round-robin to select
2257 * target node, but we should introduce another variable
2258 * for node_demotion[] to record last selected target node,
2259 * that may cause cache ping-pong due to the changing of
2260 * last target node. Or introducing per-cpu data to avoid
2261 * caching issue, which seems more complicated. So selecting
2262 * target node randomly seems better until now.
2264 index = get_random_int() % target_nr;
2268 target = READ_ONCE(nd->nodes[index]);
2275 /* Disable reclaim-based migration. */
2276 static void __disable_all_migrate_targets(void)
2283 for_each_online_node(node) {
2284 node_demotion[node].nr = 0;
2285 for (i = 0; i < DEMOTION_TARGET_NODES; i++)
2286 node_demotion[node].nodes[i] = NUMA_NO_NODE;
2290 static void disable_all_migrate_targets(void)
2292 __disable_all_migrate_targets();
2295 * Ensure that the "disable" is visible across the system.
2296 * Readers will see either a combination of before+disable
2297 * state or disable+after. They will never see before and
2298 * after state together.
2300 * The before+after state together might have cycles and
2301 * could cause readers to do things like loop until this
2302 * function finishes. This ensures they can only see a
2303 * single "bad" read and would, for instance, only loop
2310 * Find an automatic demotion target for 'node'.
2311 * Failing here is OK. It might just indicate
2312 * being at the end of a chain.
2314 static int establish_migrate_target(int node, nodemask_t *used,
2317 int migration_target, index, val;
2318 struct demotion_nodes *nd;
2321 return NUMA_NO_NODE;
2323 nd = &node_demotion[node];
2325 migration_target = find_next_best_node(node, used);
2326 if (migration_target == NUMA_NO_NODE)
2327 return NUMA_NO_NODE;
2330 * If the node has been set a migration target node before,
2331 * which means it's the best distance between them. Still
2332 * check if this node can be demoted to other target nodes
2333 * if they have a same best distance.
2335 if (best_distance != -1) {
2336 val = node_distance(node, migration_target);
2337 if (val > best_distance)
2342 if (WARN_ONCE(index >= DEMOTION_TARGET_NODES,
2343 "Exceeds maximum demotion target nodes\n"))
2346 nd->nodes[index] = migration_target;
2349 return migration_target;
2351 node_clear(migration_target, *used);
2352 return NUMA_NO_NODE;
2356 * When memory fills up on a node, memory contents can be
2357 * automatically migrated to another node instead of
2358 * discarded at reclaim.
2360 * Establish a "migration path" which will start at nodes
2361 * with CPUs and will follow the priorities used to build the
2362 * page allocator zonelists.
2364 * The difference here is that cycles must be avoided. If
2365 * node0 migrates to node1, then neither node1, nor anything
2366 * node1 migrates to can migrate to node0. Also one node can
2367 * be migrated to multiple nodes if the target nodes all have
2368 * a same best-distance against the source node.
2370 * This function can run simultaneously with readers of
2371 * node_demotion[]. However, it can not run simultaneously
2372 * with itself. Exclusion is provided by memory hotplug events
2373 * being single-threaded.
2375 static void __set_migration_target_nodes(void)
2377 nodemask_t next_pass;
2378 nodemask_t this_pass;
2379 nodemask_t used_targets = NODE_MASK_NONE;
2380 int node, best_distance;
2383 * Avoid any oddities like cycles that could occur
2384 * from changes in the topology. This will leave
2385 * a momentary gap when migration is disabled.
2387 disable_all_migrate_targets();
2390 * Allocations go close to CPUs, first. Assume that
2391 * the migration path starts at the nodes with CPUs.
2393 next_pass = node_states[N_CPU];
2395 this_pass = next_pass;
2396 next_pass = NODE_MASK_NONE;
2398 * To avoid cycles in the migration "graph", ensure
2399 * that migration sources are not future targets by
2400 * setting them in 'used_targets'. Do this only
2401 * once per pass so that multiple source nodes can
2402 * share a target node.
2404 * 'used_targets' will become unavailable in future
2405 * passes. This limits some opportunities for
2406 * multiple source nodes to share a destination.
2408 nodes_or(used_targets, used_targets, this_pass);
2410 for_each_node_mask(node, this_pass) {
2414 * Try to set up the migration path for the node, and the target
2415 * migration nodes can be multiple, so doing a loop to find all
2416 * the target nodes if they all have a best node distance.
2420 establish_migrate_target(node, &used_targets,
2423 if (target_node == NUMA_NO_NODE)
2426 if (best_distance == -1)
2427 best_distance = node_distance(node, target_node);
2430 * Visit targets from this pass in the next pass.
2431 * Eventually, every node will have been part of
2432 * a pass, and will become set in 'used_targets'.
2434 node_set(target_node, next_pass);
2438 * 'next_pass' contains nodes which became migration
2439 * targets in this pass. Make additional passes until
2440 * no more migrations targets are available.
2442 if (!nodes_empty(next_pass))
2447 * For callers that do not hold get_online_mems() already.
2449 void set_migration_target_nodes(void)
2452 __set_migration_target_nodes();
2457 * This leaves migrate-on-reclaim transiently disabled between
2458 * the MEM_GOING_OFFLINE and MEM_OFFLINE events. This runs
2459 * whether reclaim-based migration is enabled or not, which
2460 * ensures that the user can turn reclaim-based migration at
2461 * any time without needing to recalculate migration targets.
2463 * These callbacks already hold get_online_mems(). That is why
2464 * __set_migration_target_nodes() can be used as opposed to
2465 * set_migration_target_nodes().
2467 #ifdef CONFIG_MEMORY_HOTPLUG
2468 static int __meminit migrate_on_reclaim_callback(struct notifier_block *self,
2469 unsigned long action, void *_arg)
2471 struct memory_notify *arg = _arg;
2474 * Only update the node migration order when a node is
2475 * changing status, like online->offline. This avoids
2476 * the overhead of synchronize_rcu() in most cases.
2478 if (arg->status_change_nid < 0)
2479 return notifier_from_errno(0);
2482 case MEM_GOING_OFFLINE:
2484 * Make sure there are not transient states where
2485 * an offline node is a migration target. This
2486 * will leave migration disabled until the offline
2487 * completes and the MEM_OFFLINE case below runs.
2489 disable_all_migrate_targets();
2494 * Recalculate the target nodes once the node
2495 * reaches its final state (online or offline).
2497 __set_migration_target_nodes();
2499 case MEM_CANCEL_OFFLINE:
2501 * MEM_GOING_OFFLINE disabled all the migration
2502 * targets. Reenable them.
2504 __set_migration_target_nodes();
2506 case MEM_GOING_ONLINE:
2507 case MEM_CANCEL_ONLINE:
2511 return notifier_from_errno(0);
2515 void __init migrate_on_reclaim_init(void)
2517 node_demotion = kcalloc(nr_node_ids,
2518 sizeof(struct demotion_nodes),
2520 WARN_ON(!node_demotion);
2521 #ifdef CONFIG_MEMORY_HOTPLUG
2522 hotplug_memory_notifier(migrate_on_reclaim_callback, 100);
2525 * At this point, all numa nodes with memory/CPus have their state
2526 * properly set, so we can build the demotion order now.
2527 * Let us hold the cpu_hotplug lock just, as we could possibily have
2528 * CPU hotplug events during boot.
2531 set_migration_target_nodes();
2535 bool numa_demotion_enabled = false;
2538 static ssize_t numa_demotion_enabled_show(struct kobject *kobj,
2539 struct kobj_attribute *attr, char *buf)
2541 return sysfs_emit(buf, "%s\n",
2542 numa_demotion_enabled ? "true" : "false");
2545 static ssize_t numa_demotion_enabled_store(struct kobject *kobj,
2546 struct kobj_attribute *attr,
2547 const char *buf, size_t count)
2551 ret = kstrtobool(buf, &numa_demotion_enabled);
2558 static struct kobj_attribute numa_demotion_enabled_attr =
2559 __ATTR(demotion_enabled, 0644, numa_demotion_enabled_show,
2560 numa_demotion_enabled_store);
2562 static struct attribute *numa_attrs[] = {
2563 &numa_demotion_enabled_attr.attr,
2567 static const struct attribute_group numa_attr_group = {
2568 .attrs = numa_attrs,
2571 static int __init numa_init_sysfs(void)
2574 struct kobject *numa_kobj;
2576 numa_kobj = kobject_create_and_add("numa", mm_kobj);
2578 pr_err("failed to create numa kobject\n");
2581 err = sysfs_create_group(numa_kobj, &numa_attr_group);
2583 pr_err("failed to register numa group\n");
2589 kobject_put(numa_kobj);
2592 subsys_initcall(numa_init_sysfs);
2593 #endif /* CONFIG_SYSFS */
2594 #endif /* CONFIG_NUMA */