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>
53 #include <linux/memory-tiers.h>
55 #include <asm/tlbflush.h>
57 #include <trace/events/migrate.h>
61 int isolate_movable_page(struct page *page, isolate_mode_t mode)
63 const struct movable_operations *mops;
66 * Avoid burning cycles with pages that are yet under __free_pages(),
67 * or just got freed under us.
69 * In case we 'win' a race for a movable page being freed under us and
70 * raise its refcount preventing __free_pages() from doing its job
71 * the put_page() at the end of this block will take care of
72 * release this page, thus avoiding a nasty leakage.
74 if (unlikely(!get_page_unless_zero(page)))
78 * Check PageMovable before holding a PG_lock because page's owner
79 * assumes anybody doesn't touch PG_lock of newly allocated page
80 * so unconditionally grabbing the lock ruins page's owner side.
82 if (unlikely(!__PageMovable(page)))
85 * As movable pages are not isolated from LRU lists, concurrent
86 * compaction threads can race against page migration functions
87 * as well as race against the releasing a page.
89 * In order to avoid having an already isolated movable page
90 * being (wrongly) re-isolated while it is under migration,
91 * or to avoid attempting to isolate pages being released,
92 * lets be sure we have the page lock
93 * before proceeding with the movable page isolation steps.
95 if (unlikely(!trylock_page(page)))
98 if (!PageMovable(page) || PageIsolated(page))
101 mops = page_movable_ops(page);
102 VM_BUG_ON_PAGE(!mops, page);
104 if (!mops->isolate_page(page, mode))
105 goto out_no_isolated;
107 /* Driver shouldn't use PG_isolated bit of page->flags */
108 WARN_ON_ONCE(PageIsolated(page));
109 SetPageIsolated(page);
122 static void putback_movable_page(struct page *page)
124 const struct movable_operations *mops = page_movable_ops(page);
126 mops->putback_page(page);
127 ClearPageIsolated(page);
131 * Put previously isolated pages back onto the appropriate lists
132 * from where they were once taken off for compaction/migration.
134 * This function shall be used whenever the isolated pageset has been
135 * built from lru, balloon, hugetlbfs page. See isolate_migratepages_range()
136 * and isolate_hugetlb().
138 void putback_movable_pages(struct list_head *l)
143 list_for_each_entry_safe(page, page2, l, lru) {
144 if (unlikely(PageHuge(page))) {
145 putback_active_hugepage(page);
148 list_del(&page->lru);
150 * We isolated non-lru movable page so here we can use
151 * __PageMovable because LRU page's mapping cannot have
152 * PAGE_MAPPING_MOVABLE.
154 if (unlikely(__PageMovable(page))) {
155 VM_BUG_ON_PAGE(!PageIsolated(page), page);
157 if (PageMovable(page))
158 putback_movable_page(page);
160 ClearPageIsolated(page);
164 mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON +
165 page_is_file_lru(page), -thp_nr_pages(page));
166 putback_lru_page(page);
172 * Restore a potential migration pte to a working pte entry
174 static bool remove_migration_pte(struct folio *folio,
175 struct vm_area_struct *vma, unsigned long addr, void *old)
177 DEFINE_FOLIO_VMA_WALK(pvmw, old, vma, addr, PVMW_SYNC | PVMW_MIGRATION);
179 while (page_vma_mapped_walk(&pvmw)) {
180 rmap_t rmap_flags = RMAP_NONE;
184 unsigned long idx = 0;
186 /* pgoff is invalid for ksm pages, but they are never large */
187 if (folio_test_large(folio) && !folio_test_hugetlb(folio))
188 idx = linear_page_index(vma, pvmw.address) - pvmw.pgoff;
189 new = folio_page(folio, idx);
191 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
192 /* PMD-mapped THP migration entry */
194 VM_BUG_ON_FOLIO(folio_test_hugetlb(folio) ||
195 !folio_test_pmd_mappable(folio), folio);
196 remove_migration_pmd(&pvmw, new);
202 pte = mk_pte(new, READ_ONCE(vma->vm_page_prot));
203 if (pte_swp_soft_dirty(*pvmw.pte))
204 pte = pte_mksoft_dirty(pte);
207 * Recheck VMA as permissions can change since migration started
209 entry = pte_to_swp_entry(*pvmw.pte);
210 if (!is_migration_entry_young(entry))
211 pte = pte_mkold(pte);
212 if (folio_test_dirty(folio) && is_migration_entry_dirty(entry))
213 pte = pte_mkdirty(pte);
214 if (is_writable_migration_entry(entry))
215 pte = maybe_mkwrite(pte, vma);
216 else if (pte_swp_uffd_wp(*pvmw.pte))
217 pte = pte_mkuffd_wp(pte);
219 if (folio_test_anon(folio) && !is_readable_migration_entry(entry))
220 rmap_flags |= RMAP_EXCLUSIVE;
222 if (unlikely(is_device_private_page(new))) {
224 entry = make_writable_device_private_entry(
227 entry = make_readable_device_private_entry(
229 pte = swp_entry_to_pte(entry);
230 if (pte_swp_soft_dirty(*pvmw.pte))
231 pte = pte_swp_mksoft_dirty(pte);
232 if (pte_swp_uffd_wp(*pvmw.pte))
233 pte = pte_swp_mkuffd_wp(pte);
236 #ifdef CONFIG_HUGETLB_PAGE
237 if (folio_test_hugetlb(folio)) {
238 unsigned int shift = huge_page_shift(hstate_vma(vma));
240 pte = pte_mkhuge(pte);
241 pte = arch_make_huge_pte(pte, shift, vma->vm_flags);
242 if (folio_test_anon(folio))
243 hugepage_add_anon_rmap(new, vma, pvmw.address,
246 page_dup_file_rmap(new, true);
247 set_huge_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte);
251 if (folio_test_anon(folio))
252 page_add_anon_rmap(new, vma, pvmw.address,
255 page_add_file_rmap(new, vma, false);
256 set_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte);
258 if (vma->vm_flags & VM_LOCKED)
259 mlock_page_drain_local();
261 trace_remove_migration_pte(pvmw.address, pte_val(pte),
262 compound_order(new));
264 /* No need to invalidate - it was non-present before */
265 update_mmu_cache(vma, pvmw.address, pvmw.pte);
272 * Get rid of all migration entries and replace them by
273 * references to the indicated page.
275 void remove_migration_ptes(struct folio *src, struct folio *dst, bool locked)
277 struct rmap_walk_control rwc = {
278 .rmap_one = remove_migration_pte,
283 rmap_walk_locked(dst, &rwc);
285 rmap_walk(dst, &rwc);
289 * Something used the pte of a page under migration. We need to
290 * get to the page and wait until migration is finished.
291 * When we return from this function the fault will be retried.
293 void __migration_entry_wait(struct mm_struct *mm, pte_t *ptep,
301 if (!is_swap_pte(pte))
304 entry = pte_to_swp_entry(pte);
305 if (!is_migration_entry(entry))
308 migration_entry_wait_on_locked(entry, ptep, ptl);
311 pte_unmap_unlock(ptep, ptl);
314 void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
315 unsigned long address)
317 spinlock_t *ptl = pte_lockptr(mm, pmd);
318 pte_t *ptep = pte_offset_map(pmd, address);
319 __migration_entry_wait(mm, ptep, ptl);
322 #ifdef CONFIG_HUGETLB_PAGE
323 void __migration_entry_wait_huge(pte_t *ptep, spinlock_t *ptl)
328 pte = huge_ptep_get(ptep);
330 if (unlikely(!is_hugetlb_entry_migration(pte)))
333 migration_entry_wait_on_locked(pte_to_swp_entry(pte), NULL, ptl);
336 void migration_entry_wait_huge(struct vm_area_struct *vma, pte_t *pte)
338 spinlock_t *ptl = huge_pte_lockptr(hstate_vma(vma), vma->vm_mm, pte);
340 __migration_entry_wait_huge(pte, ptl);
344 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
345 void pmd_migration_entry_wait(struct mm_struct *mm, pmd_t *pmd)
349 ptl = pmd_lock(mm, pmd);
350 if (!is_pmd_migration_entry(*pmd))
352 migration_entry_wait_on_locked(pmd_to_swp_entry(*pmd), NULL, ptl);
359 static int folio_expected_refs(struct address_space *mapping,
366 refs += folio_nr_pages(folio);
367 if (folio_test_private(folio))
374 * Replace the page in the mapping.
376 * The number of remaining references must be:
377 * 1 for anonymous pages without a mapping
378 * 2 for pages with a mapping
379 * 3 for pages with a mapping and PagePrivate/PagePrivate2 set.
381 int folio_migrate_mapping(struct address_space *mapping,
382 struct folio *newfolio, struct folio *folio, int extra_count)
384 XA_STATE(xas, &mapping->i_pages, folio_index(folio));
385 struct zone *oldzone, *newzone;
387 int expected_count = folio_expected_refs(mapping, folio) + extra_count;
388 long nr = folio_nr_pages(folio);
391 /* Anonymous page without mapping */
392 if (folio_ref_count(folio) != expected_count)
395 /* No turning back from here */
396 newfolio->index = folio->index;
397 newfolio->mapping = folio->mapping;
398 if (folio_test_swapbacked(folio))
399 __folio_set_swapbacked(newfolio);
401 return MIGRATEPAGE_SUCCESS;
404 oldzone = folio_zone(folio);
405 newzone = folio_zone(newfolio);
408 if (!folio_ref_freeze(folio, expected_count)) {
409 xas_unlock_irq(&xas);
414 * Now we know that no one else is looking at the folio:
415 * no turning back from here.
417 newfolio->index = folio->index;
418 newfolio->mapping = folio->mapping;
419 folio_ref_add(newfolio, nr); /* add cache reference */
420 if (folio_test_swapbacked(folio)) {
421 __folio_set_swapbacked(newfolio);
422 if (folio_test_swapcache(folio)) {
423 folio_set_swapcache(newfolio);
424 newfolio->private = folio_get_private(folio);
427 VM_BUG_ON_FOLIO(folio_test_swapcache(folio), folio);
430 /* Move dirty while page refs frozen and newpage not yet exposed */
431 dirty = folio_test_dirty(folio);
433 folio_clear_dirty(folio);
434 folio_set_dirty(newfolio);
437 xas_store(&xas, newfolio);
440 * Drop cache reference from old page by unfreezing
441 * to one less reference.
442 * We know this isn't the last reference.
444 folio_ref_unfreeze(folio, expected_count - nr);
447 /* Leave irq disabled to prevent preemption while updating stats */
450 * If moved to a different zone then also account
451 * the page for that zone. Other VM counters will be
452 * taken care of when we establish references to the
453 * new page and drop references to the old page.
455 * Note that anonymous pages are accounted for
456 * via NR_FILE_PAGES and NR_ANON_MAPPED if they
457 * are mapped to swap space.
459 if (newzone != oldzone) {
460 struct lruvec *old_lruvec, *new_lruvec;
461 struct mem_cgroup *memcg;
463 memcg = folio_memcg(folio);
464 old_lruvec = mem_cgroup_lruvec(memcg, oldzone->zone_pgdat);
465 new_lruvec = mem_cgroup_lruvec(memcg, newzone->zone_pgdat);
467 __mod_lruvec_state(old_lruvec, NR_FILE_PAGES, -nr);
468 __mod_lruvec_state(new_lruvec, NR_FILE_PAGES, nr);
469 if (folio_test_swapbacked(folio) && !folio_test_swapcache(folio)) {
470 __mod_lruvec_state(old_lruvec, NR_SHMEM, -nr);
471 __mod_lruvec_state(new_lruvec, NR_SHMEM, nr);
474 if (folio_test_swapcache(folio)) {
475 __mod_lruvec_state(old_lruvec, NR_SWAPCACHE, -nr);
476 __mod_lruvec_state(new_lruvec, NR_SWAPCACHE, nr);
479 if (dirty && mapping_can_writeback(mapping)) {
480 __mod_lruvec_state(old_lruvec, NR_FILE_DIRTY, -nr);
481 __mod_zone_page_state(oldzone, NR_ZONE_WRITE_PENDING, -nr);
482 __mod_lruvec_state(new_lruvec, NR_FILE_DIRTY, nr);
483 __mod_zone_page_state(newzone, NR_ZONE_WRITE_PENDING, nr);
488 return MIGRATEPAGE_SUCCESS;
490 EXPORT_SYMBOL(folio_migrate_mapping);
493 * The expected number of remaining references is the same as that
494 * of folio_migrate_mapping().
496 int migrate_huge_page_move_mapping(struct address_space *mapping,
497 struct folio *dst, struct folio *src)
499 XA_STATE(xas, &mapping->i_pages, folio_index(src));
503 expected_count = 2 + folio_has_private(src);
504 if (!folio_ref_freeze(src, expected_count)) {
505 xas_unlock_irq(&xas);
509 dst->index = src->index;
510 dst->mapping = src->mapping;
514 xas_store(&xas, dst);
516 folio_ref_unfreeze(src, expected_count - 1);
518 xas_unlock_irq(&xas);
520 return MIGRATEPAGE_SUCCESS;
524 * Copy the flags and some other ancillary information
526 void folio_migrate_flags(struct folio *newfolio, struct folio *folio)
530 if (folio_test_error(folio))
531 folio_set_error(newfolio);
532 if (folio_test_referenced(folio))
533 folio_set_referenced(newfolio);
534 if (folio_test_uptodate(folio))
535 folio_mark_uptodate(newfolio);
536 if (folio_test_clear_active(folio)) {
537 VM_BUG_ON_FOLIO(folio_test_unevictable(folio), folio);
538 folio_set_active(newfolio);
539 } else if (folio_test_clear_unevictable(folio))
540 folio_set_unevictable(newfolio);
541 if (folio_test_workingset(folio))
542 folio_set_workingset(newfolio);
543 if (folio_test_checked(folio))
544 folio_set_checked(newfolio);
546 * PG_anon_exclusive (-> PG_mappedtodisk) is always migrated via
547 * migration entries. We can still have PG_anon_exclusive set on an
548 * effectively unmapped and unreferenced first sub-pages of an
549 * anonymous THP: we can simply copy it here via PG_mappedtodisk.
551 if (folio_test_mappedtodisk(folio))
552 folio_set_mappedtodisk(newfolio);
554 /* Move dirty on pages not done by folio_migrate_mapping() */
555 if (folio_test_dirty(folio))
556 folio_set_dirty(newfolio);
558 if (folio_test_young(folio))
559 folio_set_young(newfolio);
560 if (folio_test_idle(folio))
561 folio_set_idle(newfolio);
564 * Copy NUMA information to the new page, to prevent over-eager
565 * future migrations of this same page.
567 cpupid = page_cpupid_xchg_last(&folio->page, -1);
569 * For memory tiering mode, when migrate between slow and fast
570 * memory node, reset cpupid, because that is used to record
571 * page access time in slow memory node.
573 if (sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING) {
574 bool f_toptier = node_is_toptier(page_to_nid(&folio->page));
575 bool t_toptier = node_is_toptier(page_to_nid(&newfolio->page));
577 if (f_toptier != t_toptier)
580 page_cpupid_xchg_last(&newfolio->page, cpupid);
582 folio_migrate_ksm(newfolio, folio);
584 * Please do not reorder this without considering how mm/ksm.c's
585 * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache().
587 if (folio_test_swapcache(folio))
588 folio_clear_swapcache(folio);
589 folio_clear_private(folio);
591 /* page->private contains hugetlb specific flags */
592 if (!folio_test_hugetlb(folio))
593 folio->private = NULL;
596 * If any waiters have accumulated on the new page then
599 if (folio_test_writeback(newfolio))
600 folio_end_writeback(newfolio);
603 * PG_readahead shares the same bit with PG_reclaim. The above
604 * end_page_writeback() may clear PG_readahead mistakenly, so set the
607 if (folio_test_readahead(folio))
608 folio_set_readahead(newfolio);
610 folio_copy_owner(newfolio, folio);
612 if (!folio_test_hugetlb(folio))
613 mem_cgroup_migrate(folio, newfolio);
615 EXPORT_SYMBOL(folio_migrate_flags);
617 void folio_migrate_copy(struct folio *newfolio, struct folio *folio)
619 folio_copy(newfolio, folio);
620 folio_migrate_flags(newfolio, folio);
622 EXPORT_SYMBOL(folio_migrate_copy);
624 /************************************************************
625 * Migration functions
626 ***********************************************************/
629 * migrate_folio() - Simple folio migration.
630 * @mapping: The address_space containing the folio.
631 * @dst: The folio to migrate the data to.
632 * @src: The folio containing the current data.
633 * @mode: How to migrate the page.
635 * Common logic to directly migrate a single LRU folio suitable for
636 * folios that do not use PagePrivate/PagePrivate2.
638 * Folios are locked upon entry and exit.
640 int migrate_folio(struct address_space *mapping, struct folio *dst,
641 struct folio *src, enum migrate_mode mode)
645 BUG_ON(folio_test_writeback(src)); /* Writeback must be complete */
647 rc = folio_migrate_mapping(mapping, dst, src, 0);
649 if (rc != MIGRATEPAGE_SUCCESS)
652 if (mode != MIGRATE_SYNC_NO_COPY)
653 folio_migrate_copy(dst, src);
655 folio_migrate_flags(dst, src);
656 return MIGRATEPAGE_SUCCESS;
658 EXPORT_SYMBOL(migrate_folio);
661 /* Returns true if all buffers are successfully locked */
662 static bool buffer_migrate_lock_buffers(struct buffer_head *head,
663 enum migrate_mode mode)
665 struct buffer_head *bh = head;
667 /* Simple case, sync compaction */
668 if (mode != MIGRATE_ASYNC) {
671 bh = bh->b_this_page;
673 } while (bh != head);
678 /* async case, we cannot block on lock_buffer so use trylock_buffer */
680 if (!trylock_buffer(bh)) {
682 * We failed to lock the buffer and cannot stall in
683 * async migration. Release the taken locks
685 struct buffer_head *failed_bh = bh;
687 while (bh != failed_bh) {
689 bh = bh->b_this_page;
694 bh = bh->b_this_page;
695 } while (bh != head);
699 static int __buffer_migrate_folio(struct address_space *mapping,
700 struct folio *dst, struct folio *src, enum migrate_mode mode,
703 struct buffer_head *bh, *head;
707 head = folio_buffers(src);
709 return migrate_folio(mapping, dst, src, mode);
711 /* Check whether page does not have extra refs before we do more work */
712 expected_count = folio_expected_refs(mapping, src);
713 if (folio_ref_count(src) != expected_count)
716 if (!buffer_migrate_lock_buffers(head, mode))
721 bool invalidated = false;
725 spin_lock(&mapping->private_lock);
728 if (atomic_read(&bh->b_count)) {
732 bh = bh->b_this_page;
733 } while (bh != head);
739 spin_unlock(&mapping->private_lock);
740 invalidate_bh_lrus();
742 goto recheck_buffers;
746 rc = folio_migrate_mapping(mapping, dst, src, 0);
747 if (rc != MIGRATEPAGE_SUCCESS)
750 folio_attach_private(dst, folio_detach_private(src));
754 set_bh_page(bh, &dst->page, bh_offset(bh));
755 bh = bh->b_this_page;
756 } while (bh != head);
758 if (mode != MIGRATE_SYNC_NO_COPY)
759 folio_migrate_copy(dst, src);
761 folio_migrate_flags(dst, src);
763 rc = MIGRATEPAGE_SUCCESS;
766 spin_unlock(&mapping->private_lock);
770 bh = bh->b_this_page;
771 } while (bh != head);
777 * buffer_migrate_folio() - Migration function for folios with buffers.
778 * @mapping: The address space containing @src.
779 * @dst: The folio to migrate to.
780 * @src: The folio to migrate from.
781 * @mode: How to migrate the folio.
783 * This function can only be used if the underlying filesystem guarantees
784 * that no other references to @src exist. For example attached buffer
785 * heads are accessed only under the folio lock. If your filesystem cannot
786 * provide this guarantee, buffer_migrate_folio_norefs() may be more
789 * Return: 0 on success or a negative errno on failure.
791 int buffer_migrate_folio(struct address_space *mapping,
792 struct folio *dst, struct folio *src, enum migrate_mode mode)
794 return __buffer_migrate_folio(mapping, dst, src, mode, false);
796 EXPORT_SYMBOL(buffer_migrate_folio);
799 * buffer_migrate_folio_norefs() - Migration function for folios with buffers.
800 * @mapping: The address space containing @src.
801 * @dst: The folio to migrate to.
802 * @src: The folio to migrate from.
803 * @mode: How to migrate the folio.
805 * Like buffer_migrate_folio() except that this variant is more careful
806 * and checks that there are also no buffer head references. This function
807 * is the right one for mappings where buffer heads are directly looked
808 * up and referenced (such as block device mappings).
810 * Return: 0 on success or a negative errno on failure.
812 int buffer_migrate_folio_norefs(struct address_space *mapping,
813 struct folio *dst, struct folio *src, enum migrate_mode mode)
815 return __buffer_migrate_folio(mapping, dst, src, mode, true);
819 int filemap_migrate_folio(struct address_space *mapping,
820 struct folio *dst, struct folio *src, enum migrate_mode mode)
824 ret = folio_migrate_mapping(mapping, dst, src, 0);
825 if (ret != MIGRATEPAGE_SUCCESS)
828 if (folio_get_private(src))
829 folio_attach_private(dst, folio_detach_private(src));
831 if (mode != MIGRATE_SYNC_NO_COPY)
832 folio_migrate_copy(dst, src);
834 folio_migrate_flags(dst, src);
835 return MIGRATEPAGE_SUCCESS;
837 EXPORT_SYMBOL_GPL(filemap_migrate_folio);
840 * Writeback a folio to clean the dirty state
842 static int writeout(struct address_space *mapping, struct folio *folio)
844 struct writeback_control wbc = {
845 .sync_mode = WB_SYNC_NONE,
848 .range_end = LLONG_MAX,
853 if (!mapping->a_ops->writepage)
854 /* No write method for the address space */
857 if (!folio_clear_dirty_for_io(folio))
858 /* Someone else already triggered a write */
862 * A dirty folio may imply that the underlying filesystem has
863 * the folio on some queue. So the folio must be clean for
864 * migration. Writeout may mean we lose the lock and the
865 * folio state is no longer what we checked for earlier.
866 * At this point we know that the migration attempt cannot
869 remove_migration_ptes(folio, folio, false);
871 rc = mapping->a_ops->writepage(&folio->page, &wbc);
873 if (rc != AOP_WRITEPAGE_ACTIVATE)
874 /* unlocked. Relock */
877 return (rc < 0) ? -EIO : -EAGAIN;
881 * Default handling if a filesystem does not provide a migration function.
883 static int fallback_migrate_folio(struct address_space *mapping,
884 struct folio *dst, struct folio *src, enum migrate_mode mode)
886 if (folio_test_dirty(src)) {
887 /* Only writeback folios in full synchronous migration */
890 case MIGRATE_SYNC_NO_COPY:
895 return writeout(mapping, src);
899 * Buffers may be managed in a filesystem specific way.
900 * We must have no buffers or drop them.
902 if (folio_test_private(src) &&
903 !filemap_release_folio(src, GFP_KERNEL))
904 return mode == MIGRATE_SYNC ? -EAGAIN : -EBUSY;
906 return migrate_folio(mapping, dst, src, mode);
910 * Move a page to a newly allocated page
911 * The page is locked and all ptes have been successfully removed.
913 * The new page will have replaced the old page if this function
918 * MIGRATEPAGE_SUCCESS - success
920 static int move_to_new_folio(struct folio *dst, struct folio *src,
921 enum migrate_mode mode)
924 bool is_lru = !__PageMovable(&src->page);
926 VM_BUG_ON_FOLIO(!folio_test_locked(src), src);
927 VM_BUG_ON_FOLIO(!folio_test_locked(dst), dst);
929 if (likely(is_lru)) {
930 struct address_space *mapping = folio_mapping(src);
933 rc = migrate_folio(mapping, dst, src, mode);
934 else if (mapping->a_ops->migrate_folio)
936 * Most folios have a mapping and most filesystems
937 * provide a migrate_folio callback. Anonymous folios
938 * are part of swap space which also has its own
939 * migrate_folio callback. This is the most common path
940 * for page migration.
942 rc = mapping->a_ops->migrate_folio(mapping, dst, src,
945 rc = fallback_migrate_folio(mapping, dst, src, mode);
947 const struct movable_operations *mops;
950 * In case of non-lru page, it could be released after
951 * isolation step. In that case, we shouldn't try migration.
953 VM_BUG_ON_FOLIO(!folio_test_isolated(src), src);
954 if (!folio_test_movable(src)) {
955 rc = MIGRATEPAGE_SUCCESS;
956 folio_clear_isolated(src);
960 mops = page_movable_ops(&src->page);
961 rc = mops->migrate_page(&dst->page, &src->page, mode);
962 WARN_ON_ONCE(rc == MIGRATEPAGE_SUCCESS &&
963 !folio_test_isolated(src));
967 * When successful, old pagecache src->mapping must be cleared before
968 * src is freed; but stats require that PageAnon be left as PageAnon.
970 if (rc == MIGRATEPAGE_SUCCESS) {
971 if (__PageMovable(&src->page)) {
972 VM_BUG_ON_FOLIO(!folio_test_isolated(src), src);
975 * We clear PG_movable under page_lock so any compactor
976 * cannot try to migrate this page.
978 folio_clear_isolated(src);
982 * Anonymous and movable src->mapping will be cleared by
983 * free_pages_prepare so don't reset it here for keeping
984 * the type to work PageAnon, for example.
986 if (!folio_mapping_flags(src))
989 if (likely(!folio_is_zone_device(dst)))
990 flush_dcache_folio(dst);
996 static int __unmap_and_move(struct folio *src, struct folio *dst,
997 int force, enum migrate_mode mode)
1000 bool page_was_mapped = false;
1001 struct anon_vma *anon_vma = NULL;
1002 bool is_lru = !__PageMovable(&src->page);
1004 if (!folio_trylock(src)) {
1005 if (!force || mode == MIGRATE_ASYNC)
1009 * It's not safe for direct compaction to call lock_page.
1010 * For example, during page readahead pages are added locked
1011 * to the LRU. Later, when the IO completes the pages are
1012 * marked uptodate and unlocked. However, the queueing
1013 * could be merging multiple pages for one bio (e.g.
1014 * mpage_readahead). If an allocation happens for the
1015 * second or third page, the process can end up locking
1016 * the same page twice and deadlocking. Rather than
1017 * trying to be clever about what pages can be locked,
1018 * avoid the use of lock_page for direct compaction
1021 if (current->flags & PF_MEMALLOC)
1027 if (folio_test_writeback(src)) {
1029 * Only in the case of a full synchronous migration is it
1030 * necessary to wait for PageWriteback. In the async case,
1031 * the retry loop is too short and in the sync-light case,
1032 * the overhead of stalling is too much
1036 case MIGRATE_SYNC_NO_COPY:
1044 folio_wait_writeback(src);
1048 * By try_to_migrate(), src->mapcount goes down to 0 here. In this case,
1049 * we cannot notice that anon_vma is freed while we migrate a page.
1050 * This get_anon_vma() delays freeing anon_vma pointer until the end
1051 * of migration. File cache pages are no problem because of page_lock()
1052 * File Caches may use write_page() or lock_page() in migration, then,
1053 * just care Anon page here.
1055 * Only folio_get_anon_vma() understands the subtleties of
1056 * getting a hold on an anon_vma from outside one of its mms.
1057 * But if we cannot get anon_vma, then we won't need it anyway,
1058 * because that implies that the anon page is no longer mapped
1059 * (and cannot be remapped so long as we hold the page lock).
1061 if (folio_test_anon(src) && !folio_test_ksm(src))
1062 anon_vma = folio_get_anon_vma(src);
1065 * Block others from accessing the new page when we get around to
1066 * establishing additional references. We are usually the only one
1067 * holding a reference to dst at this point. We used to have a BUG
1068 * here if folio_trylock(dst) fails, but would like to allow for
1069 * cases where there might be a race with the previous use of dst.
1070 * This is much like races on refcount of oldpage: just don't BUG().
1072 if (unlikely(!folio_trylock(dst)))
1075 if (unlikely(!is_lru)) {
1076 rc = move_to_new_folio(dst, src, mode);
1077 goto out_unlock_both;
1081 * Corner case handling:
1082 * 1. When a new swap-cache page is read into, it is added to the LRU
1083 * and treated as swapcache but it has no rmap yet.
1084 * Calling try_to_unmap() against a src->mapping==NULL page will
1085 * trigger a BUG. So handle it here.
1086 * 2. An orphaned page (see truncate_cleanup_page) might have
1087 * fs-private metadata. The page can be picked up due to memory
1088 * offlining. Everywhere else except page reclaim, the page is
1089 * invisible to the vm, so the page can not be migrated. So try to
1090 * free the metadata, so the page can be freed.
1092 if (!src->mapping) {
1093 if (folio_test_private(src)) {
1094 try_to_free_buffers(src);
1095 goto out_unlock_both;
1097 } else if (folio_mapped(src)) {
1098 /* Establish migration ptes */
1099 VM_BUG_ON_FOLIO(folio_test_anon(src) &&
1100 !folio_test_ksm(src) && !anon_vma, src);
1101 try_to_migrate(src, 0);
1102 page_was_mapped = true;
1105 if (!folio_mapped(src))
1106 rc = move_to_new_folio(dst, src, mode);
1109 * When successful, push dst to LRU immediately: so that if it
1110 * turns out to be an mlocked page, remove_migration_ptes() will
1111 * automatically build up the correct dst->mlock_count for it.
1113 * We would like to do something similar for the old page, when
1114 * unsuccessful, and other cases when a page has been temporarily
1115 * isolated from the unevictable LRU: but this case is the easiest.
1117 if (rc == MIGRATEPAGE_SUCCESS) {
1119 if (page_was_mapped)
1123 if (page_was_mapped)
1124 remove_migration_ptes(src,
1125 rc == MIGRATEPAGE_SUCCESS ? dst : src, false);
1130 /* Drop an anon_vma reference if we took one */
1132 put_anon_vma(anon_vma);
1136 * If migration is successful, decrease refcount of dst,
1137 * which will not free the page because new page owner increased
1140 if (rc == MIGRATEPAGE_SUCCESS)
1147 * Obtain the lock on page, remove all ptes and migrate the page
1148 * to the newly allocated page in newpage.
1150 static int unmap_and_move(new_page_t get_new_page,
1151 free_page_t put_new_page,
1152 unsigned long private, struct page *page,
1153 int force, enum migrate_mode mode,
1154 enum migrate_reason reason,
1155 struct list_head *ret)
1157 struct folio *dst, *src = page_folio(page);
1158 int rc = MIGRATEPAGE_SUCCESS;
1159 struct page *newpage = NULL;
1161 if (!thp_migration_supported() && PageTransHuge(page))
1164 if (page_count(page) == 1) {
1165 /* Page was freed from under us. So we are done. */
1166 ClearPageActive(page);
1167 ClearPageUnevictable(page);
1168 /* free_pages_prepare() will clear PG_isolated. */
1172 newpage = get_new_page(page, private);
1175 dst = page_folio(newpage);
1177 newpage->private = 0;
1178 rc = __unmap_and_move(src, dst, 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)))
1266 if (folio_ref_count(src) == 1) {
1267 /* page was freed from under us. So we are done. */
1268 putback_active_hugepage(hpage);
1269 return MIGRATEPAGE_SUCCESS;
1272 new_hpage = get_new_page(hpage, private);
1275 dst = page_folio(new_hpage);
1277 if (!folio_trylock(src)) {
1282 case MIGRATE_SYNC_NO_COPY:
1291 * Check for pages which are in the process of being freed. Without
1292 * folio_mapping() set, hugetlbfs specific move page routine will not
1293 * be called and we could leak usage counts for subpools.
1295 if (hugetlb_page_subpool(hpage) && !folio_mapping(src)) {
1300 if (folio_test_anon(src))
1301 anon_vma = folio_get_anon_vma(src);
1303 if (unlikely(!folio_trylock(dst)))
1306 if (folio_mapped(src)) {
1307 enum ttu_flags ttu = 0;
1309 if (!folio_test_anon(src)) {
1311 * In shared mappings, try_to_unmap could potentially
1312 * call huge_pmd_unshare. Because of this, take
1313 * semaphore in write mode here and set TTU_RMAP_LOCKED
1314 * to let lower levels know we have taken the lock.
1316 mapping = hugetlb_page_mapping_lock_write(hpage);
1317 if (unlikely(!mapping))
1318 goto unlock_put_anon;
1320 ttu = TTU_RMAP_LOCKED;
1323 try_to_migrate(src, ttu);
1324 page_was_mapped = 1;
1326 if (ttu & TTU_RMAP_LOCKED)
1327 i_mmap_unlock_write(mapping);
1330 if (!folio_mapped(src))
1331 rc = move_to_new_folio(dst, src, mode);
1333 if (page_was_mapped)
1334 remove_migration_ptes(src,
1335 rc == MIGRATEPAGE_SUCCESS ? dst : src, false);
1342 put_anon_vma(anon_vma);
1344 if (rc == MIGRATEPAGE_SUCCESS) {
1345 move_hugetlb_state(hpage, new_hpage, reason);
1346 put_new_page = NULL;
1352 if (rc == MIGRATEPAGE_SUCCESS)
1353 putback_active_hugepage(hpage);
1354 else if (rc != -EAGAIN)
1355 list_move_tail(&src->lru, ret);
1358 * If migration was not successful and there's a freeing callback, use
1359 * it. Otherwise, put_page() will drop the reference grabbed during
1363 put_new_page(new_hpage, private);
1365 putback_active_hugepage(new_hpage);
1370 static inline int try_split_thp(struct page *page, struct list_head *split_pages)
1375 rc = split_huge_page_to_list(page, split_pages);
1378 list_move_tail(&page->lru, 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_retry_pages = 0;
1417 int nr_succeeded = 0;
1418 int nr_thp_succeeded = 0;
1419 int nr_thp_failed = 0;
1420 int nr_thp_split = 0;
1422 bool is_thp = false;
1425 int rc, nr_subpages;
1426 LIST_HEAD(ret_pages);
1427 LIST_HEAD(thp_split_pages);
1428 bool nosplit = (reason == MR_NUMA_MISPLACED);
1429 bool no_subpage_counting = false;
1431 trace_mm_migrate_pages_start(mode, reason);
1433 thp_subpage_migration:
1434 for (pass = 0; pass < 10 && (retry || thp_retry); pass++) {
1439 list_for_each_entry_safe(page, page2, from, lru) {
1441 * THP statistics is based on the source huge page.
1442 * Capture required information that might get lost
1445 is_thp = PageTransHuge(page) && !PageHuge(page);
1446 nr_subpages = compound_nr(page);
1450 rc = unmap_and_move_huge_page(get_new_page,
1451 put_new_page, private, page,
1452 pass > 2, mode, reason,
1455 rc = unmap_and_move(get_new_page, put_new_page,
1456 private, page, pass > 2, mode,
1457 reason, &ret_pages);
1460 * Success: non hugetlb page will be freed, hugetlb
1461 * page will be put back
1462 * -EAGAIN: stay on the from list
1463 * -ENOMEM: stay on the from list
1464 * -ENOSYS: stay on the from list
1465 * Other errno: put on ret_pages list then splice to
1470 * THP migration might be unsupported or the
1471 * allocation could've failed so we should
1472 * retry on the same page with the THP split
1475 * Sub-pages are put in thp_split_pages, and
1476 * we will migrate them after the rest of the
1477 * list is processed.
1480 /* THP migration is unsupported */
1483 if (!try_split_thp(page, &thp_split_pages)) {
1487 /* Hugetlb migration is unsupported */
1488 } else if (!no_subpage_counting) {
1492 nr_failed_pages += nr_subpages;
1493 list_move_tail(&page->lru, &ret_pages);
1497 * When memory is low, don't bother to try to migrate
1498 * other pages, just exit.
1502 /* THP NUMA faulting doesn't split THP to retry. */
1503 if (!nosplit && !try_split_thp(page, &thp_split_pages)) {
1507 } else if (!no_subpage_counting) {
1511 nr_failed_pages += nr_subpages + nr_retry_pages;
1513 * There might be some subpages of fail-to-migrate THPs
1514 * left in thp_split_pages list. Move them back to migration
1515 * list so that they could be put back to the right list by
1516 * the caller otherwise the page refcnt will be leaked.
1518 list_splice_init(&thp_split_pages, from);
1519 /* nr_failed isn't updated for not used */
1520 nr_thp_failed += thp_retry;
1525 else if (!no_subpage_counting)
1527 nr_retry_pages += nr_subpages;
1529 case MIGRATEPAGE_SUCCESS:
1530 nr_succeeded += nr_subpages;
1536 * Permanent failure (-EBUSY, etc.):
1537 * unlike -EAGAIN case, the failed page is
1538 * removed from migration page list and not
1539 * retried in the next outer loop.
1543 else if (!no_subpage_counting)
1546 nr_failed_pages += nr_subpages;
1552 nr_thp_failed += thp_retry;
1553 nr_failed_pages += nr_retry_pages;
1555 * Try to migrate subpages of fail-to-migrate THPs, no nr_failed
1556 * counting in this round, since all subpages of a THP is counted
1557 * as 1 failure in the first round.
1559 if (!list_empty(&thp_split_pages)) {
1561 * Move non-migrated pages (after 10 retries) to ret_pages
1562 * to avoid migrating them again.
1564 list_splice_init(from, &ret_pages);
1565 list_splice_init(&thp_split_pages, from);
1566 no_subpage_counting = true;
1568 goto thp_subpage_migration;
1571 rc = nr_failed + nr_thp_failed;
1574 * Put the permanent failure page back to migration list, they
1575 * will be put back to the right list by the caller.
1577 list_splice(&ret_pages, from);
1579 count_vm_events(PGMIGRATE_SUCCESS, nr_succeeded);
1580 count_vm_events(PGMIGRATE_FAIL, nr_failed_pages);
1581 count_vm_events(THP_MIGRATION_SUCCESS, nr_thp_succeeded);
1582 count_vm_events(THP_MIGRATION_FAIL, nr_thp_failed);
1583 count_vm_events(THP_MIGRATION_SPLIT, nr_thp_split);
1584 trace_mm_migrate_pages(nr_succeeded, nr_failed_pages, nr_thp_succeeded,
1585 nr_thp_failed, nr_thp_split, mode, reason);
1588 *ret_succeeded = nr_succeeded;
1593 struct page *alloc_migration_target(struct page *page, unsigned long private)
1595 struct folio *folio = page_folio(page);
1596 struct migration_target_control *mtc;
1598 unsigned int order = 0;
1599 struct folio *new_folio = NULL;
1603 mtc = (struct migration_target_control *)private;
1604 gfp_mask = mtc->gfp_mask;
1606 if (nid == NUMA_NO_NODE)
1607 nid = folio_nid(folio);
1609 if (folio_test_hugetlb(folio)) {
1610 struct hstate *h = page_hstate(&folio->page);
1612 gfp_mask = htlb_modify_alloc_mask(h, gfp_mask);
1613 return alloc_huge_page_nodemask(h, nid, mtc->nmask, gfp_mask);
1616 if (folio_test_large(folio)) {
1618 * clear __GFP_RECLAIM to make the migration callback
1619 * consistent with regular THP allocations.
1621 gfp_mask &= ~__GFP_RECLAIM;
1622 gfp_mask |= GFP_TRANSHUGE;
1623 order = folio_order(folio);
1625 zidx = zone_idx(folio_zone(folio));
1626 if (is_highmem_idx(zidx) || zidx == ZONE_MOVABLE)
1627 gfp_mask |= __GFP_HIGHMEM;
1629 new_folio = __folio_alloc(gfp_mask, order, nid, mtc->nmask);
1631 return &new_folio->page;
1636 static int store_status(int __user *status, int start, int value, int nr)
1639 if (put_user(value, status + start))
1647 static int do_move_pages_to_node(struct mm_struct *mm,
1648 struct list_head *pagelist, int node)
1651 struct migration_target_control mtc = {
1653 .gfp_mask = GFP_HIGHUSER_MOVABLE | __GFP_THISNODE,
1656 err = migrate_pages(pagelist, alloc_migration_target, NULL,
1657 (unsigned long)&mtc, MIGRATE_SYNC, MR_SYSCALL, NULL);
1659 putback_movable_pages(pagelist);
1664 * Resolves the given address to a struct page, isolates it from the LRU and
1665 * puts it to the given pagelist.
1667 * errno - if the page cannot be found/isolated
1668 * 0 - when it doesn't have to be migrated because it is already on the
1670 * 1 - when it has been queued
1672 static int add_page_for_migration(struct mm_struct *mm, unsigned long addr,
1673 int node, struct list_head *pagelist, bool migrate_all)
1675 struct vm_area_struct *vma;
1681 vma = vma_lookup(mm, addr);
1682 if (!vma || !vma_migratable(vma))
1685 /* FOLL_DUMP to ignore special (like zero) pages */
1686 page = follow_page(vma, addr, FOLL_GET | FOLL_DUMP);
1688 err = PTR_ERR(page);
1696 if (is_zone_device_page(page))
1700 if (page_to_nid(page) == node)
1704 if (page_mapcount(page) > 1 && !migrate_all)
1707 if (PageHuge(page)) {
1708 if (PageHead(page)) {
1709 err = isolate_hugetlb(page, pagelist);
1716 head = compound_head(page);
1717 err = isolate_lru_page(head);
1722 list_add_tail(&head->lru, pagelist);
1723 mod_node_page_state(page_pgdat(head),
1724 NR_ISOLATED_ANON + page_is_file_lru(head),
1725 thp_nr_pages(head));
1729 * Either remove the duplicate refcount from
1730 * isolate_lru_page() or drop the page ref if it was
1735 mmap_read_unlock(mm);
1739 static int move_pages_and_store_status(struct mm_struct *mm, int node,
1740 struct list_head *pagelist, int __user *status,
1741 int start, int i, unsigned long nr_pages)
1745 if (list_empty(pagelist))
1748 err = do_move_pages_to_node(mm, pagelist, node);
1751 * Positive err means the number of failed
1752 * pages to migrate. Since we are going to
1753 * abort and return the number of non-migrated
1754 * pages, so need to include the rest of the
1755 * nr_pages that have not been attempted as
1759 err += nr_pages - i;
1762 return store_status(status, start, node, i - start);
1766 * Migrate an array of page address onto an array of nodes and fill
1767 * the corresponding array of status.
1769 static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes,
1770 unsigned long nr_pages,
1771 const void __user * __user *pages,
1772 const int __user *nodes,
1773 int __user *status, int flags)
1775 int current_node = NUMA_NO_NODE;
1776 LIST_HEAD(pagelist);
1780 lru_cache_disable();
1782 for (i = start = 0; i < nr_pages; i++) {
1783 const void __user *p;
1788 if (get_user(p, pages + i))
1790 if (get_user(node, nodes + i))
1792 addr = (unsigned long)untagged_addr(p);
1795 if (node < 0 || node >= MAX_NUMNODES)
1797 if (!node_state(node, N_MEMORY))
1801 if (!node_isset(node, task_nodes))
1804 if (current_node == NUMA_NO_NODE) {
1805 current_node = node;
1807 } else if (node != current_node) {
1808 err = move_pages_and_store_status(mm, current_node,
1809 &pagelist, status, start, i, nr_pages);
1813 current_node = node;
1817 * Errors in the page lookup or isolation are not fatal and we simply
1818 * report them via status
1820 err = add_page_for_migration(mm, addr, current_node,
1821 &pagelist, flags & MPOL_MF_MOVE_ALL);
1824 /* The page is successfully queued for migration */
1829 * The move_pages() man page does not have an -EEXIST choice, so
1830 * use -EFAULT instead.
1836 * If the page is already on the target node (!err), store the
1837 * node, otherwise, store the err.
1839 err = store_status(status, i, err ? : current_node, 1);
1843 err = move_pages_and_store_status(mm, current_node, &pagelist,
1844 status, start, i, nr_pages);
1846 /* We have accounted for page i */
1851 current_node = NUMA_NO_NODE;
1854 /* Make sure we do not overwrite the existing error */
1855 err1 = move_pages_and_store_status(mm, current_node, &pagelist,
1856 status, start, i, nr_pages);
1865 * Determine the nodes of an array of pages and store it in an array of status.
1867 static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages,
1868 const void __user **pages, int *status)
1874 for (i = 0; i < nr_pages; i++) {
1875 unsigned long addr = (unsigned long)(*pages);
1876 unsigned int foll_flags = FOLL_DUMP;
1877 struct vm_area_struct *vma;
1881 vma = vma_lookup(mm, addr);
1885 /* Not all huge page follow APIs support 'FOLL_GET' */
1886 if (!is_vm_hugetlb_page(vma))
1887 foll_flags |= FOLL_GET;
1889 /* FOLL_DUMP to ignore special (like zero) pages */
1890 page = follow_page(vma, addr, foll_flags);
1892 err = PTR_ERR(page);
1900 if (!is_zone_device_page(page))
1901 err = page_to_nid(page);
1903 if (foll_flags & FOLL_GET)
1912 mmap_read_unlock(mm);
1915 static int get_compat_pages_array(const void __user *chunk_pages[],
1916 const void __user * __user *pages,
1917 unsigned long chunk_nr)
1919 compat_uptr_t __user *pages32 = (compat_uptr_t __user *)pages;
1923 for (i = 0; i < chunk_nr; i++) {
1924 if (get_user(p, pages32 + i))
1926 chunk_pages[i] = compat_ptr(p);
1933 * Determine the nodes of a user array of pages and store it in
1934 * a user array of status.
1936 static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages,
1937 const void __user * __user *pages,
1940 #define DO_PAGES_STAT_CHUNK_NR 16UL
1941 const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR];
1942 int chunk_status[DO_PAGES_STAT_CHUNK_NR];
1945 unsigned long chunk_nr = min(nr_pages, DO_PAGES_STAT_CHUNK_NR);
1947 if (in_compat_syscall()) {
1948 if (get_compat_pages_array(chunk_pages, pages,
1952 if (copy_from_user(chunk_pages, pages,
1953 chunk_nr * sizeof(*chunk_pages)))
1957 do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status);
1959 if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status)))
1964 nr_pages -= chunk_nr;
1966 return nr_pages ? -EFAULT : 0;
1969 static struct mm_struct *find_mm_struct(pid_t pid, nodemask_t *mem_nodes)
1971 struct task_struct *task;
1972 struct mm_struct *mm;
1975 * There is no need to check if current process has the right to modify
1976 * the specified process when they are same.
1980 *mem_nodes = cpuset_mems_allowed(current);
1984 /* Find the mm_struct */
1986 task = find_task_by_vpid(pid);
1989 return ERR_PTR(-ESRCH);
1991 get_task_struct(task);
1994 * Check if this process has the right to modify the specified
1995 * process. Use the regular "ptrace_may_access()" checks.
1997 if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) {
1999 mm = ERR_PTR(-EPERM);
2004 mm = ERR_PTR(security_task_movememory(task));
2007 *mem_nodes = cpuset_mems_allowed(task);
2008 mm = get_task_mm(task);
2010 put_task_struct(task);
2012 mm = ERR_PTR(-EINVAL);
2017 * Move a list of pages in the address space of the currently executing
2020 static int kernel_move_pages(pid_t pid, unsigned long nr_pages,
2021 const void __user * __user *pages,
2022 const int __user *nodes,
2023 int __user *status, int flags)
2025 struct mm_struct *mm;
2027 nodemask_t task_nodes;
2030 if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
2033 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
2036 mm = find_mm_struct(pid, &task_nodes);
2041 err = do_pages_move(mm, task_nodes, nr_pages, pages,
2042 nodes, status, flags);
2044 err = do_pages_stat(mm, nr_pages, pages, status);
2050 SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages,
2051 const void __user * __user *, pages,
2052 const int __user *, nodes,
2053 int __user *, status, int, flags)
2055 return kernel_move_pages(pid, nr_pages, pages, nodes, status, flags);
2058 #ifdef CONFIG_NUMA_BALANCING
2060 * Returns true if this is a safe migration target node for misplaced NUMA
2061 * pages. Currently it only checks the watermarks which is crude.
2063 static bool migrate_balanced_pgdat(struct pglist_data *pgdat,
2064 unsigned long nr_migrate_pages)
2068 for (z = pgdat->nr_zones - 1; z >= 0; z--) {
2069 struct zone *zone = pgdat->node_zones + z;
2071 if (!managed_zone(zone))
2074 /* Avoid waking kswapd by allocating pages_to_migrate pages. */
2075 if (!zone_watermark_ok(zone, 0,
2076 high_wmark_pages(zone) +
2085 static struct page *alloc_misplaced_dst_page(struct page *page,
2088 int nid = (int) data;
2089 int order = compound_order(page);
2090 gfp_t gfp = __GFP_THISNODE;
2094 gfp |= GFP_TRANSHUGE_LIGHT;
2096 gfp |= GFP_HIGHUSER_MOVABLE | __GFP_NOMEMALLOC | __GFP_NORETRY |
2098 gfp &= ~__GFP_RECLAIM;
2100 new = __folio_alloc_node(gfp, order, nid);
2105 static int numamigrate_isolate_page(pg_data_t *pgdat, struct page *page)
2107 int nr_pages = thp_nr_pages(page);
2108 int order = compound_order(page);
2110 VM_BUG_ON_PAGE(order && !PageTransHuge(page), page);
2112 /* Do not migrate THP mapped by multiple processes */
2113 if (PageTransHuge(page) && total_mapcount(page) > 1)
2116 /* Avoid migrating to a node that is nearly full */
2117 if (!migrate_balanced_pgdat(pgdat, nr_pages)) {
2120 if (!(sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING))
2122 for (z = pgdat->nr_zones - 1; z >= 0; z--) {
2123 if (managed_zone(pgdat->node_zones + z))
2126 wakeup_kswapd(pgdat->node_zones + z, 0, order, ZONE_MOVABLE);
2130 if (isolate_lru_page(page))
2133 mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON + page_is_file_lru(page),
2137 * Isolating the page has taken another reference, so the
2138 * caller's reference can be safely dropped without the page
2139 * disappearing underneath us during migration.
2146 * Attempt to migrate a misplaced page to the specified destination
2147 * node. Caller is expected to have an elevated reference count on
2148 * the page that will be dropped by this function before returning.
2150 int migrate_misplaced_page(struct page *page, struct vm_area_struct *vma,
2153 pg_data_t *pgdat = NODE_DATA(node);
2156 unsigned int nr_succeeded;
2157 LIST_HEAD(migratepages);
2158 int nr_pages = thp_nr_pages(page);
2161 * Don't migrate file pages that are mapped in multiple processes
2162 * with execute permissions as they are probably shared libraries.
2164 if (page_mapcount(page) != 1 && page_is_file_lru(page) &&
2165 (vma->vm_flags & VM_EXEC))
2169 * Also do not migrate dirty pages as not all filesystems can move
2170 * dirty pages in MIGRATE_ASYNC mode which is a waste of cycles.
2172 if (page_is_file_lru(page) && PageDirty(page))
2175 isolated = numamigrate_isolate_page(pgdat, page);
2179 list_add(&page->lru, &migratepages);
2180 nr_remaining = migrate_pages(&migratepages, alloc_misplaced_dst_page,
2181 NULL, node, MIGRATE_ASYNC,
2182 MR_NUMA_MISPLACED, &nr_succeeded);
2184 if (!list_empty(&migratepages)) {
2185 list_del(&page->lru);
2186 mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON +
2187 page_is_file_lru(page), -nr_pages);
2188 putback_lru_page(page);
2193 count_vm_numa_events(NUMA_PAGE_MIGRATE, nr_succeeded);
2194 if (!node_is_toptier(page_to_nid(page)) && node_is_toptier(node))
2195 mod_node_page_state(pgdat, PGPROMOTE_SUCCESS,
2198 BUG_ON(!list_empty(&migratepages));
2205 #endif /* CONFIG_NUMA_BALANCING */
2206 #endif /* CONFIG_NUMA */