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 #define CREATE_TRACE_POINTS
57 #include <trace/events/migrate.h>
61 int isolate_movable_page(struct page *page, isolate_mode_t mode)
63 struct address_space *mapping;
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 mapping = page_mapping(page);
102 VM_BUG_ON_PAGE(!mapping, page);
104 if (!mapping->a_ops->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 struct address_space *mapping;
126 mapping = page_mapping(page);
127 mapping->a_ops->putback_page(page);
128 ClearPageIsolated(page);
132 * Put previously isolated pages back onto the appropriate lists
133 * from where they were once taken off for compaction/migration.
135 * This function shall be used whenever the isolated pageset has been
136 * built from lru, balloon, hugetlbfs page. See isolate_migratepages_range()
137 * and isolate_huge_page().
139 void putback_movable_pages(struct list_head *l)
144 list_for_each_entry_safe(page, page2, l, lru) {
145 if (unlikely(PageHuge(page))) {
146 putback_active_hugepage(page);
149 list_del(&page->lru);
151 * We isolated non-lru movable page so here we can use
152 * __PageMovable because LRU page's mapping cannot have
153 * PAGE_MAPPING_MOVABLE.
155 if (unlikely(__PageMovable(page))) {
156 VM_BUG_ON_PAGE(!PageIsolated(page), page);
158 if (PageMovable(page))
159 putback_movable_page(page);
161 ClearPageIsolated(page);
165 mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON +
166 page_is_file_lru(page), -thp_nr_pages(page));
167 putback_lru_page(page);
173 * Restore a potential migration pte to a working pte entry
175 static bool remove_migration_pte(struct folio *folio,
176 struct vm_area_struct *vma, unsigned long addr, void *old)
178 DEFINE_FOLIO_VMA_WALK(pvmw, old, vma, addr, PVMW_SYNC | PVMW_MIGRATION);
180 while (page_vma_mapped_walk(&pvmw)) {
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 = pte_mkold(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_writable_migration_entry(entry))
211 pte = maybe_mkwrite(pte, vma);
212 else if (pte_swp_uffd_wp(*pvmw.pte))
213 pte = pte_mkuffd_wp(pte);
215 if (unlikely(is_device_private_page(new))) {
217 entry = make_writable_device_private_entry(
220 entry = make_readable_device_private_entry(
222 pte = swp_entry_to_pte(entry);
223 if (pte_swp_soft_dirty(*pvmw.pte))
224 pte = pte_swp_mksoft_dirty(pte);
225 if (pte_swp_uffd_wp(*pvmw.pte))
226 pte = pte_swp_mkuffd_wp(pte);
229 #ifdef CONFIG_HUGETLB_PAGE
230 if (folio_test_hugetlb(folio)) {
231 unsigned int shift = huge_page_shift(hstate_vma(vma));
233 pte = pte_mkhuge(pte);
234 pte = arch_make_huge_pte(pte, shift, vma->vm_flags);
235 if (folio_test_anon(folio))
236 hugepage_add_anon_rmap(new, vma, pvmw.address);
238 page_dup_rmap(new, true);
239 set_huge_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte);
243 if (folio_test_anon(folio))
244 page_add_anon_rmap(new, vma, pvmw.address, false);
246 page_add_file_rmap(new, vma, false);
247 set_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte);
249 if (vma->vm_flags & VM_LOCKED)
250 mlock_page_drain(smp_processor_id());
252 /* No need to invalidate - it was non-present before */
253 update_mmu_cache(vma, pvmw.address, pvmw.pte);
260 * Get rid of all migration entries and replace them by
261 * references to the indicated page.
263 void remove_migration_ptes(struct folio *src, struct folio *dst, bool locked)
265 struct rmap_walk_control rwc = {
266 .rmap_one = remove_migration_pte,
271 rmap_walk_locked(dst, &rwc);
273 rmap_walk(dst, &rwc);
277 * Something used the pte of a page under migration. We need to
278 * get to the page and wait until migration is finished.
279 * When we return from this function the fault will be retried.
281 void __migration_entry_wait(struct mm_struct *mm, pte_t *ptep,
289 if (!is_swap_pte(pte))
292 entry = pte_to_swp_entry(pte);
293 if (!is_migration_entry(entry))
296 migration_entry_wait_on_locked(entry, ptep, ptl);
299 pte_unmap_unlock(ptep, ptl);
302 void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
303 unsigned long address)
305 spinlock_t *ptl = pte_lockptr(mm, pmd);
306 pte_t *ptep = pte_offset_map(pmd, address);
307 __migration_entry_wait(mm, ptep, ptl);
310 void migration_entry_wait_huge(struct vm_area_struct *vma,
311 struct mm_struct *mm, pte_t *pte)
313 spinlock_t *ptl = huge_pte_lockptr(hstate_vma(vma), mm, pte);
314 __migration_entry_wait(mm, pte, ptl);
317 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
318 void pmd_migration_entry_wait(struct mm_struct *mm, pmd_t *pmd)
322 ptl = pmd_lock(mm, pmd);
323 if (!is_pmd_migration_entry(*pmd))
325 migration_entry_wait_on_locked(pmd_to_swp_entry(*pmd), NULL, ptl);
332 static int expected_page_refs(struct address_space *mapping, struct page *page)
334 int expected_count = 1;
337 expected_count += compound_nr(page) + page_has_private(page);
338 return expected_count;
342 * Replace the page in the mapping.
344 * The number of remaining references must be:
345 * 1 for anonymous pages without a mapping
346 * 2 for pages with a mapping
347 * 3 for pages with a mapping and PagePrivate/PagePrivate2 set.
349 int folio_migrate_mapping(struct address_space *mapping,
350 struct folio *newfolio, struct folio *folio, int extra_count)
352 XA_STATE(xas, &mapping->i_pages, folio_index(folio));
353 struct zone *oldzone, *newzone;
355 int expected_count = expected_page_refs(mapping, &folio->page) + extra_count;
356 long nr = folio_nr_pages(folio);
359 /* Anonymous page without mapping */
360 if (folio_ref_count(folio) != expected_count)
363 /* No turning back from here */
364 newfolio->index = folio->index;
365 newfolio->mapping = folio->mapping;
366 if (folio_test_swapbacked(folio))
367 __folio_set_swapbacked(newfolio);
369 return MIGRATEPAGE_SUCCESS;
372 oldzone = folio_zone(folio);
373 newzone = folio_zone(newfolio);
376 if (!folio_ref_freeze(folio, expected_count)) {
377 xas_unlock_irq(&xas);
382 * Now we know that no one else is looking at the folio:
383 * no turning back from here.
385 newfolio->index = folio->index;
386 newfolio->mapping = folio->mapping;
387 folio_ref_add(newfolio, nr); /* add cache reference */
388 if (folio_test_swapbacked(folio)) {
389 __folio_set_swapbacked(newfolio);
390 if (folio_test_swapcache(folio)) {
391 folio_set_swapcache(newfolio);
392 newfolio->private = folio_get_private(folio);
395 VM_BUG_ON_FOLIO(folio_test_swapcache(folio), folio);
398 /* Move dirty while page refs frozen and newpage not yet exposed */
399 dirty = folio_test_dirty(folio);
401 folio_clear_dirty(folio);
402 folio_set_dirty(newfolio);
405 xas_store(&xas, newfolio);
408 * Drop cache reference from old page by unfreezing
409 * to one less reference.
410 * We know this isn't the last reference.
412 folio_ref_unfreeze(folio, expected_count - nr);
415 /* Leave irq disabled to prevent preemption while updating stats */
418 * If moved to a different zone then also account
419 * the page for that zone. Other VM counters will be
420 * taken care of when we establish references to the
421 * new page and drop references to the old page.
423 * Note that anonymous pages are accounted for
424 * via NR_FILE_PAGES and NR_ANON_MAPPED if they
425 * are mapped to swap space.
427 if (newzone != oldzone) {
428 struct lruvec *old_lruvec, *new_lruvec;
429 struct mem_cgroup *memcg;
431 memcg = folio_memcg(folio);
432 old_lruvec = mem_cgroup_lruvec(memcg, oldzone->zone_pgdat);
433 new_lruvec = mem_cgroup_lruvec(memcg, newzone->zone_pgdat);
435 __mod_lruvec_state(old_lruvec, NR_FILE_PAGES, -nr);
436 __mod_lruvec_state(new_lruvec, NR_FILE_PAGES, nr);
437 if (folio_test_swapbacked(folio) && !folio_test_swapcache(folio)) {
438 __mod_lruvec_state(old_lruvec, NR_SHMEM, -nr);
439 __mod_lruvec_state(new_lruvec, NR_SHMEM, nr);
442 if (folio_test_swapcache(folio)) {
443 __mod_lruvec_state(old_lruvec, NR_SWAPCACHE, -nr);
444 __mod_lruvec_state(new_lruvec, NR_SWAPCACHE, nr);
447 if (dirty && mapping_can_writeback(mapping)) {
448 __mod_lruvec_state(old_lruvec, NR_FILE_DIRTY, -nr);
449 __mod_zone_page_state(oldzone, NR_ZONE_WRITE_PENDING, -nr);
450 __mod_lruvec_state(new_lruvec, NR_FILE_DIRTY, nr);
451 __mod_zone_page_state(newzone, NR_ZONE_WRITE_PENDING, nr);
456 return MIGRATEPAGE_SUCCESS;
458 EXPORT_SYMBOL(folio_migrate_mapping);
461 * The expected number of remaining references is the same as that
462 * of folio_migrate_mapping().
464 int migrate_huge_page_move_mapping(struct address_space *mapping,
465 struct page *newpage, struct page *page)
467 XA_STATE(xas, &mapping->i_pages, page_index(page));
471 expected_count = 2 + page_has_private(page);
472 if (page_count(page) != expected_count || xas_load(&xas) != page) {
473 xas_unlock_irq(&xas);
477 if (!page_ref_freeze(page, expected_count)) {
478 xas_unlock_irq(&xas);
482 newpage->index = page->index;
483 newpage->mapping = page->mapping;
487 xas_store(&xas, newpage);
489 page_ref_unfreeze(page, expected_count - 1);
491 xas_unlock_irq(&xas);
493 return MIGRATEPAGE_SUCCESS;
497 * Copy the flags and some other ancillary information
499 void folio_migrate_flags(struct folio *newfolio, struct folio *folio)
503 if (folio_test_error(folio))
504 folio_set_error(newfolio);
505 if (folio_test_referenced(folio))
506 folio_set_referenced(newfolio);
507 if (folio_test_uptodate(folio))
508 folio_mark_uptodate(newfolio);
509 if (folio_test_clear_active(folio)) {
510 VM_BUG_ON_FOLIO(folio_test_unevictable(folio), folio);
511 folio_set_active(newfolio);
512 } else if (folio_test_clear_unevictable(folio))
513 folio_set_unevictable(newfolio);
514 if (folio_test_workingset(folio))
515 folio_set_workingset(newfolio);
516 if (folio_test_checked(folio))
517 folio_set_checked(newfolio);
518 if (folio_test_mappedtodisk(folio))
519 folio_set_mappedtodisk(newfolio);
521 /* Move dirty on pages not done by folio_migrate_mapping() */
522 if (folio_test_dirty(folio))
523 folio_set_dirty(newfolio);
525 if (folio_test_young(folio))
526 folio_set_young(newfolio);
527 if (folio_test_idle(folio))
528 folio_set_idle(newfolio);
531 * Copy NUMA information to the new page, to prevent over-eager
532 * future migrations of this same page.
534 cpupid = page_cpupid_xchg_last(&folio->page, -1);
535 page_cpupid_xchg_last(&newfolio->page, cpupid);
537 folio_migrate_ksm(newfolio, folio);
539 * Please do not reorder this without considering how mm/ksm.c's
540 * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache().
542 if (folio_test_swapcache(folio))
543 folio_clear_swapcache(folio);
544 folio_clear_private(folio);
546 /* page->private contains hugetlb specific flags */
547 if (!folio_test_hugetlb(folio))
548 folio->private = NULL;
551 * If any waiters have accumulated on the new page then
554 if (folio_test_writeback(newfolio))
555 folio_end_writeback(newfolio);
558 * PG_readahead shares the same bit with PG_reclaim. The above
559 * end_page_writeback() may clear PG_readahead mistakenly, so set the
562 if (folio_test_readahead(folio))
563 folio_set_readahead(newfolio);
565 folio_copy_owner(newfolio, folio);
567 if (!folio_test_hugetlb(folio))
568 mem_cgroup_migrate(folio, newfolio);
570 EXPORT_SYMBOL(folio_migrate_flags);
572 void folio_migrate_copy(struct folio *newfolio, struct folio *folio)
574 folio_copy(newfolio, folio);
575 folio_migrate_flags(newfolio, folio);
577 EXPORT_SYMBOL(folio_migrate_copy);
579 /************************************************************
580 * Migration functions
581 ***********************************************************/
584 * Common logic to directly migrate a single LRU page suitable for
585 * pages that do not use PagePrivate/PagePrivate2.
587 * Pages are locked upon entry and exit.
589 int migrate_page(struct address_space *mapping,
590 struct page *newpage, struct page *page,
591 enum migrate_mode mode)
593 struct folio *newfolio = page_folio(newpage);
594 struct folio *folio = page_folio(page);
597 BUG_ON(folio_test_writeback(folio)); /* Writeback must be complete */
599 rc = folio_migrate_mapping(mapping, newfolio, folio, 0);
601 if (rc != MIGRATEPAGE_SUCCESS)
604 if (mode != MIGRATE_SYNC_NO_COPY)
605 folio_migrate_copy(newfolio, folio);
607 folio_migrate_flags(newfolio, folio);
608 return MIGRATEPAGE_SUCCESS;
610 EXPORT_SYMBOL(migrate_page);
613 /* Returns true if all buffers are successfully locked */
614 static bool buffer_migrate_lock_buffers(struct buffer_head *head,
615 enum migrate_mode mode)
617 struct buffer_head *bh = head;
619 /* Simple case, sync compaction */
620 if (mode != MIGRATE_ASYNC) {
623 bh = bh->b_this_page;
625 } while (bh != head);
630 /* async case, we cannot block on lock_buffer so use trylock_buffer */
632 if (!trylock_buffer(bh)) {
634 * We failed to lock the buffer and cannot stall in
635 * async migration. Release the taken locks
637 struct buffer_head *failed_bh = bh;
639 while (bh != failed_bh) {
641 bh = bh->b_this_page;
646 bh = bh->b_this_page;
647 } while (bh != head);
651 static int __buffer_migrate_page(struct address_space *mapping,
652 struct page *newpage, struct page *page, enum migrate_mode mode,
655 struct buffer_head *bh, *head;
659 if (!page_has_buffers(page))
660 return migrate_page(mapping, newpage, page, mode);
662 /* Check whether page does not have extra refs before we do more work */
663 expected_count = expected_page_refs(mapping, page);
664 if (page_count(page) != expected_count)
667 head = page_buffers(page);
668 if (!buffer_migrate_lock_buffers(head, mode))
673 bool invalidated = false;
677 spin_lock(&mapping->private_lock);
680 if (atomic_read(&bh->b_count)) {
684 bh = bh->b_this_page;
685 } while (bh != head);
691 spin_unlock(&mapping->private_lock);
692 invalidate_bh_lrus();
694 goto recheck_buffers;
698 rc = migrate_page_move_mapping(mapping, newpage, page, 0);
699 if (rc != MIGRATEPAGE_SUCCESS)
702 attach_page_private(newpage, detach_page_private(page));
706 set_bh_page(bh, newpage, bh_offset(bh));
707 bh = bh->b_this_page;
709 } while (bh != head);
711 if (mode != MIGRATE_SYNC_NO_COPY)
712 migrate_page_copy(newpage, page);
714 migrate_page_states(newpage, page);
716 rc = MIGRATEPAGE_SUCCESS;
719 spin_unlock(&mapping->private_lock);
723 bh = bh->b_this_page;
725 } while (bh != head);
731 * Migration function for pages with buffers. This function can only be used
732 * if the underlying filesystem guarantees that no other references to "page"
733 * exist. For example attached buffer heads are accessed only under page lock.
735 int buffer_migrate_page(struct address_space *mapping,
736 struct page *newpage, struct page *page, enum migrate_mode mode)
738 return __buffer_migrate_page(mapping, newpage, page, mode, false);
740 EXPORT_SYMBOL(buffer_migrate_page);
743 * Same as above except that this variant is more careful and checks that there
744 * are also no buffer head references. This function is the right one for
745 * mappings where buffer heads are directly looked up and referenced (such as
746 * block device mappings).
748 int buffer_migrate_page_norefs(struct address_space *mapping,
749 struct page *newpage, struct page *page, enum migrate_mode mode)
751 return __buffer_migrate_page(mapping, newpage, page, mode, true);
756 * Writeback a page to clean the dirty state
758 static int writeout(struct address_space *mapping, struct page *page)
760 struct folio *folio = page_folio(page);
761 struct writeback_control wbc = {
762 .sync_mode = WB_SYNC_NONE,
765 .range_end = LLONG_MAX,
770 if (!mapping->a_ops->writepage)
771 /* No write method for the address space */
774 if (!clear_page_dirty_for_io(page))
775 /* Someone else already triggered a write */
779 * A dirty page may imply that the underlying filesystem has
780 * the page on some queue. So the page must be clean for
781 * migration. Writeout may mean we loose the lock and the
782 * page state is no longer what we checked for earlier.
783 * At this point we know that the migration attempt cannot
786 remove_migration_ptes(folio, folio, false);
788 rc = mapping->a_ops->writepage(page, &wbc);
790 if (rc != AOP_WRITEPAGE_ACTIVATE)
791 /* unlocked. Relock */
794 return (rc < 0) ? -EIO : -EAGAIN;
798 * Default handling if a filesystem does not provide a migration function.
800 static int fallback_migrate_page(struct address_space *mapping,
801 struct page *newpage, struct page *page, enum migrate_mode mode)
803 if (PageDirty(page)) {
804 /* Only writeback pages in full synchronous migration */
807 case MIGRATE_SYNC_NO_COPY:
812 return writeout(mapping, page);
816 * Buffers may be managed in a filesystem specific way.
817 * We must have no buffers or drop them.
819 if (page_has_private(page) &&
820 !try_to_release_page(page, GFP_KERNEL))
821 return mode == MIGRATE_SYNC ? -EAGAIN : -EBUSY;
823 return migrate_page(mapping, newpage, page, mode);
827 * Move a page to a newly allocated page
828 * The page is locked and all ptes have been successfully removed.
830 * The new page will have replaced the old page if this function
835 * MIGRATEPAGE_SUCCESS - success
837 static int move_to_new_page(struct page *newpage, struct page *page,
838 enum migrate_mode mode)
840 struct address_space *mapping;
842 bool is_lru = !__PageMovable(page);
844 VM_BUG_ON_PAGE(!PageLocked(page), page);
845 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
847 mapping = page_mapping(page);
849 if (likely(is_lru)) {
851 rc = migrate_page(mapping, newpage, page, mode);
852 else if (mapping->a_ops->migratepage)
854 * Most pages have a mapping and most filesystems
855 * provide a migratepage callback. Anonymous pages
856 * are part of swap space which also has its own
857 * migratepage callback. This is the most common path
858 * for page migration.
860 rc = mapping->a_ops->migratepage(mapping, newpage,
863 rc = fallback_migrate_page(mapping, newpage,
867 * In case of non-lru page, it could be released after
868 * isolation step. In that case, we shouldn't try migration.
870 VM_BUG_ON_PAGE(!PageIsolated(page), page);
871 if (!PageMovable(page)) {
872 rc = MIGRATEPAGE_SUCCESS;
873 ClearPageIsolated(page);
877 rc = mapping->a_ops->migratepage(mapping, newpage,
879 WARN_ON_ONCE(rc == MIGRATEPAGE_SUCCESS &&
880 !PageIsolated(page));
884 * When successful, old pagecache page->mapping must be cleared before
885 * page is freed; but stats require that PageAnon be left as PageAnon.
887 if (rc == MIGRATEPAGE_SUCCESS) {
888 if (__PageMovable(page)) {
889 VM_BUG_ON_PAGE(!PageIsolated(page), page);
892 * We clear PG_movable under page_lock so any compactor
893 * cannot try to migrate this page.
895 ClearPageIsolated(page);
899 * Anonymous and movable page->mapping will be cleared by
900 * free_pages_prepare so don't reset it here for keeping
901 * the type to work PageAnon, for example.
903 if (!PageMappingFlags(page))
904 page->mapping = NULL;
906 if (likely(!is_zone_device_page(newpage)))
907 flush_dcache_folio(page_folio(newpage));
913 static int __unmap_and_move(struct page *page, struct page *newpage,
914 int force, enum migrate_mode mode)
916 struct folio *folio = page_folio(page);
917 struct folio *dst = page_folio(newpage);
919 bool page_was_mapped = false;
920 struct anon_vma *anon_vma = NULL;
921 bool is_lru = !__PageMovable(page);
923 if (!trylock_page(page)) {
924 if (!force || mode == MIGRATE_ASYNC)
928 * It's not safe for direct compaction to call lock_page.
929 * For example, during page readahead pages are added locked
930 * to the LRU. Later, when the IO completes the pages are
931 * marked uptodate and unlocked. However, the queueing
932 * could be merging multiple pages for one bio (e.g.
933 * mpage_readahead). If an allocation happens for the
934 * second or third page, the process can end up locking
935 * the same page twice and deadlocking. Rather than
936 * trying to be clever about what pages can be locked,
937 * avoid the use of lock_page for direct compaction
940 if (current->flags & PF_MEMALLOC)
946 if (PageWriteback(page)) {
948 * Only in the case of a full synchronous migration is it
949 * necessary to wait for PageWriteback. In the async case,
950 * the retry loop is too short and in the sync-light case,
951 * the overhead of stalling is too much
955 case MIGRATE_SYNC_NO_COPY:
963 wait_on_page_writeback(page);
967 * By try_to_migrate(), page->mapcount goes down to 0 here. In this case,
968 * we cannot notice that anon_vma is freed while we migrates a page.
969 * This get_anon_vma() delays freeing anon_vma pointer until the end
970 * of migration. File cache pages are no problem because of page_lock()
971 * File Caches may use write_page() or lock_page() in migration, then,
972 * just care Anon page here.
974 * Only page_get_anon_vma() understands the subtleties of
975 * getting a hold on an anon_vma from outside one of its mms.
976 * But if we cannot get anon_vma, then we won't need it anyway,
977 * because that implies that the anon page is no longer mapped
978 * (and cannot be remapped so long as we hold the page lock).
980 if (PageAnon(page) && !PageKsm(page))
981 anon_vma = page_get_anon_vma(page);
984 * Block others from accessing the new page when we get around to
985 * establishing additional references. We are usually the only one
986 * holding a reference to newpage at this point. We used to have a BUG
987 * here if trylock_page(newpage) fails, but would like to allow for
988 * cases where there might be a race with the previous use of newpage.
989 * This is much like races on refcount of oldpage: just don't BUG().
991 if (unlikely(!trylock_page(newpage)))
994 if (unlikely(!is_lru)) {
995 rc = move_to_new_page(newpage, page, mode);
996 goto out_unlock_both;
1000 * Corner case handling:
1001 * 1. When a new swap-cache page is read into, it is added to the LRU
1002 * and treated as swapcache but it has no rmap yet.
1003 * Calling try_to_unmap() against a page->mapping==NULL page will
1004 * trigger a BUG. So handle it here.
1005 * 2. An orphaned page (see truncate_cleanup_page) might have
1006 * fs-private metadata. The page can be picked up due to memory
1007 * offlining. Everywhere else except page reclaim, the page is
1008 * invisible to the vm, so the page can not be migrated. So try to
1009 * free the metadata, so the page can be freed.
1011 if (!page->mapping) {
1012 VM_BUG_ON_PAGE(PageAnon(page), page);
1013 if (page_has_private(page)) {
1014 try_to_free_buffers(page);
1015 goto out_unlock_both;
1017 } else if (page_mapped(page)) {
1018 /* Establish migration ptes */
1019 VM_BUG_ON_PAGE(PageAnon(page) && !PageKsm(page) && !anon_vma,
1021 try_to_migrate(folio, 0);
1022 page_was_mapped = true;
1025 if (!page_mapped(page))
1026 rc = move_to_new_page(newpage, page, mode);
1029 * When successful, push newpage to LRU immediately: so that if it
1030 * turns out to be an mlocked page, remove_migration_ptes() will
1031 * automatically build up the correct newpage->mlock_count for it.
1033 * We would like to do something similar for the old page, when
1034 * unsuccessful, and other cases when a page has been temporarily
1035 * isolated from the unevictable LRU: but this case is the easiest.
1037 if (rc == MIGRATEPAGE_SUCCESS) {
1038 lru_cache_add(newpage);
1039 if (page_was_mapped)
1043 if (page_was_mapped)
1044 remove_migration_ptes(folio,
1045 rc == MIGRATEPAGE_SUCCESS ? dst : folio, false);
1048 unlock_page(newpage);
1050 /* Drop an anon_vma reference if we took one */
1052 put_anon_vma(anon_vma);
1056 * If migration is successful, decrease refcount of the newpage,
1057 * which will not free the page because new page owner increased
1060 if (rc == MIGRATEPAGE_SUCCESS)
1067 * Obtain the lock on page, remove all ptes and migrate the page
1068 * to the newly allocated page in newpage.
1070 static int unmap_and_move(new_page_t get_new_page,
1071 free_page_t put_new_page,
1072 unsigned long private, struct page *page,
1073 int force, enum migrate_mode mode,
1074 enum migrate_reason reason,
1075 struct list_head *ret)
1077 int rc = MIGRATEPAGE_SUCCESS;
1078 struct page *newpage = NULL;
1080 if (!thp_migration_supported() && PageTransHuge(page))
1083 if (page_count(page) == 1) {
1084 /* page was freed from under us. So we are done. */
1085 ClearPageActive(page);
1086 ClearPageUnevictable(page);
1087 if (unlikely(__PageMovable(page))) {
1089 if (!PageMovable(page))
1090 ClearPageIsolated(page);
1096 newpage = get_new_page(page, private);
1100 rc = __unmap_and_move(page, newpage, force, mode);
1101 if (rc == MIGRATEPAGE_SUCCESS)
1102 set_page_owner_migrate_reason(newpage, reason);
1105 if (rc != -EAGAIN) {
1107 * A page that has been migrated has all references
1108 * removed and will be freed. A page that has not been
1109 * migrated will have kept its references and be restored.
1111 list_del(&page->lru);
1115 * If migration is successful, releases reference grabbed during
1116 * isolation. Otherwise, restore the page to right list unless
1119 if (rc == MIGRATEPAGE_SUCCESS) {
1121 * Compaction can migrate also non-LRU pages which are
1122 * not accounted to NR_ISOLATED_*. They can be recognized
1125 if (likely(!__PageMovable(page)))
1126 mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON +
1127 page_is_file_lru(page), -thp_nr_pages(page));
1129 if (reason != MR_MEMORY_FAILURE)
1131 * We release the page in page_handle_poison.
1136 list_add_tail(&page->lru, ret);
1139 put_new_page(newpage, private);
1148 * Counterpart of unmap_and_move_page() for hugepage migration.
1150 * This function doesn't wait the completion of hugepage I/O
1151 * because there is no race between I/O and migration for hugepage.
1152 * Note that currently hugepage I/O occurs only in direct I/O
1153 * where no lock is held and PG_writeback is irrelevant,
1154 * and writeback status of all subpages are counted in the reference
1155 * count of the head page (i.e. if all subpages of a 2MB hugepage are
1156 * under direct I/O, the reference of the head page is 512 and a bit more.)
1157 * This means that when we try to migrate hugepage whose subpages are
1158 * doing direct I/O, some references remain after try_to_unmap() and
1159 * hugepage migration fails without data corruption.
1161 * There is also no race when direct I/O is issued on the page under migration,
1162 * because then pte is replaced with migration swap entry and direct I/O code
1163 * will wait in the page fault for migration to complete.
1165 static int unmap_and_move_huge_page(new_page_t get_new_page,
1166 free_page_t put_new_page, unsigned long private,
1167 struct page *hpage, int force,
1168 enum migrate_mode mode, int reason,
1169 struct list_head *ret)
1171 struct folio *dst, *src = page_folio(hpage);
1173 int page_was_mapped = 0;
1174 struct page *new_hpage;
1175 struct anon_vma *anon_vma = NULL;
1176 struct address_space *mapping = NULL;
1179 * Migratability of hugepages depends on architectures and their size.
1180 * This check is necessary because some callers of hugepage migration
1181 * like soft offline and memory hotremove don't walk through page
1182 * tables or check whether the hugepage is pmd-based or not before
1183 * kicking migration.
1185 if (!hugepage_migration_supported(page_hstate(hpage))) {
1186 list_move_tail(&hpage->lru, ret);
1190 if (page_count(hpage) == 1) {
1191 /* page was freed from under us. So we are done. */
1192 putback_active_hugepage(hpage);
1193 return MIGRATEPAGE_SUCCESS;
1196 new_hpage = get_new_page(hpage, private);
1199 dst = page_folio(new_hpage);
1201 if (!trylock_page(hpage)) {
1206 case MIGRATE_SYNC_NO_COPY:
1215 * Check for pages which are in the process of being freed. Without
1216 * page_mapping() set, hugetlbfs specific move page routine will not
1217 * be called and we could leak usage counts for subpools.
1219 if (hugetlb_page_subpool(hpage) && !page_mapping(hpage)) {
1224 if (PageAnon(hpage))
1225 anon_vma = page_get_anon_vma(hpage);
1227 if (unlikely(!trylock_page(new_hpage)))
1230 if (page_mapped(hpage)) {
1231 bool mapping_locked = false;
1232 enum ttu_flags ttu = 0;
1234 if (!PageAnon(hpage)) {
1236 * In shared mappings, try_to_unmap could potentially
1237 * call huge_pmd_unshare. Because of this, take
1238 * semaphore in write mode here and set TTU_RMAP_LOCKED
1239 * to let lower levels know we have taken the lock.
1241 mapping = hugetlb_page_mapping_lock_write(hpage);
1242 if (unlikely(!mapping))
1243 goto unlock_put_anon;
1245 mapping_locked = true;
1246 ttu |= TTU_RMAP_LOCKED;
1249 try_to_migrate(src, ttu);
1250 page_was_mapped = 1;
1253 i_mmap_unlock_write(mapping);
1256 if (!page_mapped(hpage))
1257 rc = move_to_new_page(new_hpage, hpage, mode);
1259 if (page_was_mapped)
1260 remove_migration_ptes(src,
1261 rc == MIGRATEPAGE_SUCCESS ? dst : src, false);
1264 unlock_page(new_hpage);
1268 put_anon_vma(anon_vma);
1270 if (rc == MIGRATEPAGE_SUCCESS) {
1271 move_hugetlb_state(hpage, new_hpage, reason);
1272 put_new_page = NULL;
1278 if (rc == MIGRATEPAGE_SUCCESS)
1279 putback_active_hugepage(hpage);
1280 else if (rc != -EAGAIN)
1281 list_move_tail(&hpage->lru, ret);
1284 * If migration was not successful and there's a freeing callback, use
1285 * it. Otherwise, put_page() will drop the reference grabbed during
1289 put_new_page(new_hpage, private);
1291 putback_active_hugepage(new_hpage);
1296 static inline int try_split_thp(struct page *page, struct page **page2,
1297 struct list_head *from)
1302 rc = split_huge_page_to_list(page, from);
1305 list_safe_reset_next(page, *page2, lru);
1311 * migrate_pages - migrate the pages specified in a list, to the free pages
1312 * supplied as the target for the page migration
1314 * @from: The list of pages to be migrated.
1315 * @get_new_page: The function used to allocate free pages to be used
1316 * as the target of the page migration.
1317 * @put_new_page: The function used to free target pages if migration
1318 * fails, or NULL if no special handling is necessary.
1319 * @private: Private data to be passed on to get_new_page()
1320 * @mode: The migration mode that specifies the constraints for
1321 * page migration, if any.
1322 * @reason: The reason for page migration.
1323 * @ret_succeeded: Set to the number of normal pages migrated successfully if
1324 * the caller passes a non-NULL pointer.
1326 * The function returns after 10 attempts or if no pages are movable any more
1327 * because the list has become empty or no retryable pages exist any more.
1328 * It is caller's responsibility to call putback_movable_pages() to return pages
1329 * to the LRU or free list only if ret != 0.
1331 * Returns the number of {normal page, THP, hugetlb} that were not migrated, or
1332 * an error code. The number of THP splits will be considered as the number of
1333 * non-migrated THP, no matter how many subpages of the THP are migrated successfully.
1335 int migrate_pages(struct list_head *from, new_page_t get_new_page,
1336 free_page_t put_new_page, unsigned long private,
1337 enum migrate_mode mode, int reason, unsigned int *ret_succeeded)
1342 int nr_failed_pages = 0;
1343 int nr_succeeded = 0;
1344 int nr_thp_succeeded = 0;
1345 int nr_thp_failed = 0;
1346 int nr_thp_split = 0;
1348 bool is_thp = false;
1351 int rc, nr_subpages;
1352 LIST_HEAD(ret_pages);
1353 LIST_HEAD(thp_split_pages);
1354 bool nosplit = (reason == MR_NUMA_MISPLACED);
1355 bool no_subpage_counting = false;
1357 trace_mm_migrate_pages_start(mode, reason);
1359 thp_subpage_migration:
1360 for (pass = 0; pass < 10 && (retry || thp_retry); pass++) {
1364 list_for_each_entry_safe(page, page2, from, lru) {
1367 * THP statistics is based on the source huge page.
1368 * Capture required information that might get lost
1371 is_thp = PageTransHuge(page) && !PageHuge(page);
1372 nr_subpages = compound_nr(page);
1376 rc = unmap_and_move_huge_page(get_new_page,
1377 put_new_page, private, page,
1378 pass > 2, mode, reason,
1381 rc = unmap_and_move(get_new_page, put_new_page,
1382 private, page, pass > 2, mode,
1383 reason, &ret_pages);
1386 * Success: non hugetlb page will be freed, hugetlb
1387 * page will be put back
1388 * -EAGAIN: stay on the from list
1389 * -ENOMEM: stay on the from list
1390 * Other errno: put on ret_pages list then splice to
1395 * THP migration might be unsupported or the
1396 * allocation could've failed so we should
1397 * retry on the same page with the THP split
1400 * Head page is retried immediately and tail
1401 * pages are added to the tail of the list so
1402 * we encounter them after the rest of the list
1406 /* THP migration is unsupported */
1409 if (!try_split_thp(page, &page2, &thp_split_pages)) {
1414 nr_failed_pages += nr_subpages;
1418 /* Hugetlb migration is unsupported */
1419 if (!no_subpage_counting)
1421 nr_failed_pages += nr_subpages;
1425 * When memory is low, don't bother to try to migrate
1426 * other pages, just exit.
1427 * THP NUMA faulting doesn't split THP to retry.
1429 if (is_thp && !nosplit) {
1431 if (!try_split_thp(page, &page2, &thp_split_pages)) {
1436 nr_failed_pages += nr_subpages;
1440 if (!no_subpage_counting)
1442 nr_failed_pages += nr_subpages;
1451 case MIGRATEPAGE_SUCCESS:
1452 nr_succeeded += nr_subpages;
1460 * Permanent failure (-EBUSY, etc.):
1461 * unlike -EAGAIN case, the failed page is
1462 * removed from migration page list and not
1463 * retried in the next outer loop.
1467 nr_failed_pages += nr_subpages;
1471 if (!no_subpage_counting)
1473 nr_failed_pages += nr_subpages;
1479 nr_thp_failed += thp_retry;
1481 * Try to migrate subpages of fail-to-migrate THPs, no nr_failed
1482 * counting in this round, since all subpages of a THP is counted
1483 * as 1 failure in the first round.
1485 if (!list_empty(&thp_split_pages)) {
1487 * Move non-migrated pages (after 10 retries) to ret_pages
1488 * to avoid migrating them again.
1490 list_splice_init(from, &ret_pages);
1491 list_splice_init(&thp_split_pages, from);
1492 no_subpage_counting = true;
1494 goto thp_subpage_migration;
1497 rc = nr_failed + nr_thp_failed;
1500 * Put the permanent failure page back to migration list, they
1501 * will be put back to the right list by the caller.
1503 list_splice(&ret_pages, from);
1505 count_vm_events(PGMIGRATE_SUCCESS, nr_succeeded);
1506 count_vm_events(PGMIGRATE_FAIL, nr_failed_pages);
1507 count_vm_events(THP_MIGRATION_SUCCESS, nr_thp_succeeded);
1508 count_vm_events(THP_MIGRATION_FAIL, nr_thp_failed);
1509 count_vm_events(THP_MIGRATION_SPLIT, nr_thp_split);
1510 trace_mm_migrate_pages(nr_succeeded, nr_failed_pages, nr_thp_succeeded,
1511 nr_thp_failed, nr_thp_split, mode, reason);
1514 *ret_succeeded = nr_succeeded;
1519 struct page *alloc_migration_target(struct page *page, unsigned long private)
1521 struct migration_target_control *mtc;
1523 unsigned int order = 0;
1524 struct page *new_page = NULL;
1528 mtc = (struct migration_target_control *)private;
1529 gfp_mask = mtc->gfp_mask;
1531 if (nid == NUMA_NO_NODE)
1532 nid = page_to_nid(page);
1534 if (PageHuge(page)) {
1535 struct hstate *h = page_hstate(compound_head(page));
1537 gfp_mask = htlb_modify_alloc_mask(h, gfp_mask);
1538 return alloc_huge_page_nodemask(h, nid, mtc->nmask, gfp_mask);
1541 if (PageTransHuge(page)) {
1543 * clear __GFP_RECLAIM to make the migration callback
1544 * consistent with regular THP allocations.
1546 gfp_mask &= ~__GFP_RECLAIM;
1547 gfp_mask |= GFP_TRANSHUGE;
1548 order = HPAGE_PMD_ORDER;
1550 zidx = zone_idx(page_zone(page));
1551 if (is_highmem_idx(zidx) || zidx == ZONE_MOVABLE)
1552 gfp_mask |= __GFP_HIGHMEM;
1554 new_page = __alloc_pages(gfp_mask, order, nid, mtc->nmask);
1556 if (new_page && PageTransHuge(new_page))
1557 prep_transhuge_page(new_page);
1564 static int store_status(int __user *status, int start, int value, int nr)
1567 if (put_user(value, status + start))
1575 static int do_move_pages_to_node(struct mm_struct *mm,
1576 struct list_head *pagelist, int node)
1579 struct migration_target_control mtc = {
1581 .gfp_mask = GFP_HIGHUSER_MOVABLE | __GFP_THISNODE,
1584 err = migrate_pages(pagelist, alloc_migration_target, NULL,
1585 (unsigned long)&mtc, MIGRATE_SYNC, MR_SYSCALL, NULL);
1587 putback_movable_pages(pagelist);
1592 * Resolves the given address to a struct page, isolates it from the LRU and
1593 * puts it to the given pagelist.
1595 * errno - if the page cannot be found/isolated
1596 * 0 - when it doesn't have to be migrated because it is already on the
1598 * 1 - when it has been queued
1600 static int add_page_for_migration(struct mm_struct *mm, unsigned long addr,
1601 int node, struct list_head *pagelist, bool migrate_all)
1603 struct vm_area_struct *vma;
1609 vma = find_vma(mm, addr);
1610 if (!vma || addr < vma->vm_start || !vma_migratable(vma))
1613 /* FOLL_DUMP to ignore special (like zero) pages */
1614 page = follow_page(vma, addr, FOLL_GET | FOLL_DUMP);
1616 err = PTR_ERR(page);
1625 if (page_to_nid(page) == node)
1629 if (page_mapcount(page) > 1 && !migrate_all)
1632 if (PageHuge(page)) {
1633 if (PageHead(page)) {
1634 isolate_huge_page(page, pagelist);
1640 head = compound_head(page);
1641 err = isolate_lru_page(head);
1646 list_add_tail(&head->lru, pagelist);
1647 mod_node_page_state(page_pgdat(head),
1648 NR_ISOLATED_ANON + page_is_file_lru(head),
1649 thp_nr_pages(head));
1653 * Either remove the duplicate refcount from
1654 * isolate_lru_page() or drop the page ref if it was
1659 mmap_read_unlock(mm);
1663 static int move_pages_and_store_status(struct mm_struct *mm, int node,
1664 struct list_head *pagelist, int __user *status,
1665 int start, int i, unsigned long nr_pages)
1669 if (list_empty(pagelist))
1672 err = do_move_pages_to_node(mm, pagelist, node);
1675 * Positive err means the number of failed
1676 * pages to migrate. Since we are going to
1677 * abort and return the number of non-migrated
1678 * pages, so need to include the rest of the
1679 * nr_pages that have not been attempted as
1683 err += nr_pages - i - 1;
1686 return store_status(status, start, node, i - start);
1690 * Migrate an array of page address onto an array of nodes and fill
1691 * the corresponding array of status.
1693 static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes,
1694 unsigned long nr_pages,
1695 const void __user * __user *pages,
1696 const int __user *nodes,
1697 int __user *status, int flags)
1699 int current_node = NUMA_NO_NODE;
1700 LIST_HEAD(pagelist);
1704 lru_cache_disable();
1706 for (i = start = 0; i < nr_pages; i++) {
1707 const void __user *p;
1712 if (get_user(p, pages + i))
1714 if (get_user(node, nodes + i))
1716 addr = (unsigned long)untagged_addr(p);
1719 if (node < 0 || node >= MAX_NUMNODES)
1721 if (!node_state(node, N_MEMORY))
1725 if (!node_isset(node, task_nodes))
1728 if (current_node == NUMA_NO_NODE) {
1729 current_node = node;
1731 } else if (node != current_node) {
1732 err = move_pages_and_store_status(mm, current_node,
1733 &pagelist, status, start, i, nr_pages);
1737 current_node = node;
1741 * Errors in the page lookup or isolation are not fatal and we simply
1742 * report them via status
1744 err = add_page_for_migration(mm, addr, current_node,
1745 &pagelist, flags & MPOL_MF_MOVE_ALL);
1748 /* The page is successfully queued for migration */
1753 * The move_pages() man page does not have an -EEXIST choice, so
1754 * use -EFAULT instead.
1760 * If the page is already on the target node (!err), store the
1761 * node, otherwise, store the err.
1763 err = store_status(status, i, err ? : current_node, 1);
1767 err = move_pages_and_store_status(mm, current_node, &pagelist,
1768 status, start, i, nr_pages);
1771 current_node = NUMA_NO_NODE;
1774 /* Make sure we do not overwrite the existing error */
1775 err1 = move_pages_and_store_status(mm, current_node, &pagelist,
1776 status, start, i, nr_pages);
1785 * Determine the nodes of an array of pages and store it in an array of status.
1787 static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages,
1788 const void __user **pages, int *status)
1794 for (i = 0; i < nr_pages; i++) {
1795 unsigned long addr = (unsigned long)(*pages);
1796 struct vm_area_struct *vma;
1800 vma = vma_lookup(mm, addr);
1804 /* FOLL_DUMP to ignore special (like zero) pages */
1805 page = follow_page(vma, addr, FOLL_DUMP);
1807 err = PTR_ERR(page);
1811 err = page ? page_to_nid(page) : -ENOENT;
1819 mmap_read_unlock(mm);
1822 static int get_compat_pages_array(const void __user *chunk_pages[],
1823 const void __user * __user *pages,
1824 unsigned long chunk_nr)
1826 compat_uptr_t __user *pages32 = (compat_uptr_t __user *)pages;
1830 for (i = 0; i < chunk_nr; i++) {
1831 if (get_user(p, pages32 + i))
1833 chunk_pages[i] = compat_ptr(p);
1840 * Determine the nodes of a user array of pages and store it in
1841 * a user array of status.
1843 static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages,
1844 const void __user * __user *pages,
1847 #define DO_PAGES_STAT_CHUNK_NR 16
1848 const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR];
1849 int chunk_status[DO_PAGES_STAT_CHUNK_NR];
1852 unsigned long chunk_nr;
1854 chunk_nr = nr_pages;
1855 if (chunk_nr > DO_PAGES_STAT_CHUNK_NR)
1856 chunk_nr = DO_PAGES_STAT_CHUNK_NR;
1858 if (in_compat_syscall()) {
1859 if (get_compat_pages_array(chunk_pages, pages,
1863 if (copy_from_user(chunk_pages, pages,
1864 chunk_nr * sizeof(*chunk_pages)))
1868 do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status);
1870 if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status)))
1875 nr_pages -= chunk_nr;
1877 return nr_pages ? -EFAULT : 0;
1880 static struct mm_struct *find_mm_struct(pid_t pid, nodemask_t *mem_nodes)
1882 struct task_struct *task;
1883 struct mm_struct *mm;
1886 * There is no need to check if current process has the right to modify
1887 * the specified process when they are same.
1891 *mem_nodes = cpuset_mems_allowed(current);
1895 /* Find the mm_struct */
1897 task = find_task_by_vpid(pid);
1900 return ERR_PTR(-ESRCH);
1902 get_task_struct(task);
1905 * Check if this process has the right to modify the specified
1906 * process. Use the regular "ptrace_may_access()" checks.
1908 if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) {
1910 mm = ERR_PTR(-EPERM);
1915 mm = ERR_PTR(security_task_movememory(task));
1918 *mem_nodes = cpuset_mems_allowed(task);
1919 mm = get_task_mm(task);
1921 put_task_struct(task);
1923 mm = ERR_PTR(-EINVAL);
1928 * Move a list of pages in the address space of the currently executing
1931 static int kernel_move_pages(pid_t pid, unsigned long nr_pages,
1932 const void __user * __user *pages,
1933 const int __user *nodes,
1934 int __user *status, int flags)
1936 struct mm_struct *mm;
1938 nodemask_t task_nodes;
1941 if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
1944 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1947 mm = find_mm_struct(pid, &task_nodes);
1952 err = do_pages_move(mm, task_nodes, nr_pages, pages,
1953 nodes, status, flags);
1955 err = do_pages_stat(mm, nr_pages, pages, status);
1961 SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages,
1962 const void __user * __user *, pages,
1963 const int __user *, nodes,
1964 int __user *, status, int, flags)
1966 return kernel_move_pages(pid, nr_pages, pages, nodes, status, flags);
1969 #ifdef CONFIG_NUMA_BALANCING
1971 * Returns true if this is a safe migration target node for misplaced NUMA
1972 * pages. Currently it only checks the watermarks which crude
1974 static bool migrate_balanced_pgdat(struct pglist_data *pgdat,
1975 unsigned long nr_migrate_pages)
1979 for (z = pgdat->nr_zones - 1; z >= 0; z--) {
1980 struct zone *zone = pgdat->node_zones + z;
1982 if (!populated_zone(zone))
1985 /* Avoid waking kswapd by allocating pages_to_migrate pages. */
1986 if (!zone_watermark_ok(zone, 0,
1987 high_wmark_pages(zone) +
1996 static struct page *alloc_misplaced_dst_page(struct page *page,
1999 int nid = (int) data;
2000 struct page *newpage;
2002 newpage = __alloc_pages_node(nid,
2003 (GFP_HIGHUSER_MOVABLE |
2004 __GFP_THISNODE | __GFP_NOMEMALLOC |
2005 __GFP_NORETRY | __GFP_NOWARN) &
2011 static struct page *alloc_misplaced_dst_page_thp(struct page *page,
2014 int nid = (int) data;
2015 struct page *newpage;
2017 newpage = alloc_pages_node(nid, (GFP_TRANSHUGE_LIGHT | __GFP_THISNODE),
2022 prep_transhuge_page(newpage);
2028 static int numamigrate_isolate_page(pg_data_t *pgdat, struct page *page)
2031 int nr_pages = thp_nr_pages(page);
2032 int order = compound_order(page);
2034 VM_BUG_ON_PAGE(order && !PageTransHuge(page), page);
2036 /* Do not migrate THP mapped by multiple processes */
2037 if (PageTransHuge(page) && total_mapcount(page) > 1)
2040 /* Avoid migrating to a node that is nearly full */
2041 if (!migrate_balanced_pgdat(pgdat, nr_pages)) {
2044 if (!(sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING))
2046 for (z = pgdat->nr_zones - 1; z >= 0; z--) {
2047 if (populated_zone(pgdat->node_zones + z))
2050 wakeup_kswapd(pgdat->node_zones + z, 0, order, ZONE_MOVABLE);
2054 if (isolate_lru_page(page))
2057 page_lru = page_is_file_lru(page);
2058 mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON + page_lru,
2062 * Isolating the page has taken another reference, so the
2063 * caller's reference can be safely dropped without the page
2064 * disappearing underneath us during migration.
2071 * Attempt to migrate a misplaced page to the specified destination
2072 * node. Caller is expected to have an elevated reference count on
2073 * the page that will be dropped by this function before returning.
2075 int migrate_misplaced_page(struct page *page, struct vm_area_struct *vma,
2078 pg_data_t *pgdat = NODE_DATA(node);
2081 unsigned int nr_succeeded;
2082 LIST_HEAD(migratepages);
2085 int nr_pages = thp_nr_pages(page);
2088 * PTE mapped THP or HugeTLB page can't reach here so the page could
2089 * be either base page or THP. And it must be head page if it is
2092 compound = PageTransHuge(page);
2095 new = alloc_misplaced_dst_page_thp;
2097 new = alloc_misplaced_dst_page;
2100 * Don't migrate file pages that are mapped in multiple processes
2101 * with execute permissions as they are probably shared libraries.
2103 if (page_mapcount(page) != 1 && page_is_file_lru(page) &&
2104 (vma->vm_flags & VM_EXEC))
2108 * Also do not migrate dirty pages as not all filesystems can move
2109 * dirty pages in MIGRATE_ASYNC mode which is a waste of cycles.
2111 if (page_is_file_lru(page) && PageDirty(page))
2114 isolated = numamigrate_isolate_page(pgdat, page);
2118 list_add(&page->lru, &migratepages);
2119 nr_remaining = migrate_pages(&migratepages, *new, NULL, node,
2120 MIGRATE_ASYNC, MR_NUMA_MISPLACED,
2123 if (!list_empty(&migratepages)) {
2124 list_del(&page->lru);
2125 mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON +
2126 page_is_file_lru(page), -nr_pages);
2127 putback_lru_page(page);
2132 count_vm_numa_events(NUMA_PAGE_MIGRATE, nr_succeeded);
2133 if (!node_is_toptier(page_to_nid(page)) && node_is_toptier(node))
2134 mod_node_page_state(pgdat, PGPROMOTE_SUCCESS,
2137 BUG_ON(!list_empty(&migratepages));
2144 #endif /* CONFIG_NUMA_BALANCING */
2145 #endif /* CONFIG_NUMA */
2148 * node_demotion[] example:
2150 * Consider a system with two sockets. Each socket has
2151 * three classes of memory attached: fast, medium and slow.
2152 * Each memory class is placed in its own NUMA node. The
2153 * CPUs are placed in the node with the "fast" memory. The
2154 * 6 NUMA nodes (0-5) might be split among the sockets like
2160 * When Node 0 fills up, its memory should be migrated to
2161 * Node 1. When Node 1 fills up, it should be migrated to
2162 * Node 2. The migration path start on the nodes with the
2163 * processors (since allocations default to this node) and
2164 * fast memory, progress through medium and end with the
2167 * 0 -> 1 -> 2 -> stop
2168 * 3 -> 4 -> 5 -> stop
2170 * This is represented in the node_demotion[] like this:
2172 * { nr=1, nodes[0]=1 }, // Node 0 migrates to 1
2173 * { nr=1, nodes[0]=2 }, // Node 1 migrates to 2
2174 * { nr=0, nodes[0]=-1 }, // Node 2 does not migrate
2175 * { nr=1, nodes[0]=4 }, // Node 3 migrates to 4
2176 * { nr=1, nodes[0]=5 }, // Node 4 migrates to 5
2177 * { nr=0, nodes[0]=-1 }, // Node 5 does not migrate
2179 * Moreover some systems may have multiple slow memory nodes.
2180 * Suppose a system has one socket with 3 memory nodes, node 0
2181 * is fast memory type, and node 1/2 both are slow memory
2182 * type, and the distance between fast memory node and slow
2183 * memory node is same. So the migration path should be:
2187 * This is represented in the node_demotion[] like this:
2188 * { nr=2, {nodes[0]=1, nodes[1]=2} }, // Node 0 migrates to node 1 and node 2
2189 * { nr=0, nodes[0]=-1, }, // Node 1 dose not migrate
2190 * { nr=0, nodes[0]=-1, }, // Node 2 does not migrate
2194 * Writes to this array occur without locking. Cycles are
2195 * not allowed: Node X demotes to Y which demotes to X...
2197 * If multiple reads are performed, a single rcu_read_lock()
2198 * must be held over all reads to ensure that no cycles are
2201 #define DEFAULT_DEMOTION_TARGET_NODES 15
2203 #if MAX_NUMNODES < DEFAULT_DEMOTION_TARGET_NODES
2204 #define DEMOTION_TARGET_NODES (MAX_NUMNODES - 1)
2206 #define DEMOTION_TARGET_NODES DEFAULT_DEMOTION_TARGET_NODES
2209 struct demotion_nodes {
2211 short nodes[DEMOTION_TARGET_NODES];
2214 static struct demotion_nodes *node_demotion __read_mostly;
2217 * next_demotion_node() - Get the next node in the demotion path
2218 * @node: The starting node to lookup the next node
2220 * Return: node id for next memory node in the demotion path hierarchy
2221 * from @node; NUMA_NO_NODE if @node is terminal. This does not keep
2222 * @node online or guarantee that it *continues* to be the next demotion
2225 int next_demotion_node(int node)
2227 struct demotion_nodes *nd;
2228 unsigned short target_nr, index;
2232 return NUMA_NO_NODE;
2234 nd = &node_demotion[node];
2237 * node_demotion[] is updated without excluding this
2238 * function from running. RCU doesn't provide any
2239 * compiler barriers, so the READ_ONCE() is required
2240 * to avoid compiler reordering or read merging.
2242 * Make sure to use RCU over entire code blocks if
2243 * node_demotion[] reads need to be consistent.
2246 target_nr = READ_ONCE(nd->nr);
2248 switch (target_nr) {
2250 target = NUMA_NO_NODE;
2257 * If there are multiple target nodes, just select one
2258 * target node randomly.
2260 * In addition, we can also use round-robin to select
2261 * target node, but we should introduce another variable
2262 * for node_demotion[] to record last selected target node,
2263 * that may cause cache ping-pong due to the changing of
2264 * last target node. Or introducing per-cpu data to avoid
2265 * caching issue, which seems more complicated. So selecting
2266 * target node randomly seems better until now.
2268 index = get_random_int() % target_nr;
2272 target = READ_ONCE(nd->nodes[index]);
2279 #if defined(CONFIG_HOTPLUG_CPU)
2280 /* Disable reclaim-based migration. */
2281 static void __disable_all_migrate_targets(void)
2288 for_each_online_node(node) {
2289 node_demotion[node].nr = 0;
2290 for (i = 0; i < DEMOTION_TARGET_NODES; i++)
2291 node_demotion[node].nodes[i] = NUMA_NO_NODE;
2295 static void disable_all_migrate_targets(void)
2297 __disable_all_migrate_targets();
2300 * Ensure that the "disable" is visible across the system.
2301 * Readers will see either a combination of before+disable
2302 * state or disable+after. They will never see before and
2303 * after state together.
2305 * The before+after state together might have cycles and
2306 * could cause readers to do things like loop until this
2307 * function finishes. This ensures they can only see a
2308 * single "bad" read and would, for instance, only loop
2315 * Find an automatic demotion target for 'node'.
2316 * Failing here is OK. It might just indicate
2317 * being at the end of a chain.
2319 static int establish_migrate_target(int node, nodemask_t *used,
2322 int migration_target, index, val;
2323 struct demotion_nodes *nd;
2326 return NUMA_NO_NODE;
2328 nd = &node_demotion[node];
2330 migration_target = find_next_best_node(node, used);
2331 if (migration_target == NUMA_NO_NODE)
2332 return NUMA_NO_NODE;
2335 * If the node has been set a migration target node before,
2336 * which means it's the best distance between them. Still
2337 * check if this node can be demoted to other target nodes
2338 * if they have a same best distance.
2340 if (best_distance != -1) {
2341 val = node_distance(node, migration_target);
2342 if (val > best_distance)
2347 if (WARN_ONCE(index >= DEMOTION_TARGET_NODES,
2348 "Exceeds maximum demotion target nodes\n"))
2351 nd->nodes[index] = migration_target;
2354 return migration_target;
2356 node_clear(migration_target, *used);
2357 return NUMA_NO_NODE;
2361 * When memory fills up on a node, memory contents can be
2362 * automatically migrated to another node instead of
2363 * discarded at reclaim.
2365 * Establish a "migration path" which will start at nodes
2366 * with CPUs and will follow the priorities used to build the
2367 * page allocator zonelists.
2369 * The difference here is that cycles must be avoided. If
2370 * node0 migrates to node1, then neither node1, nor anything
2371 * node1 migrates to can migrate to node0. Also one node can
2372 * be migrated to multiple nodes if the target nodes all have
2373 * a same best-distance against the source node.
2375 * This function can run simultaneously with readers of
2376 * node_demotion[]. However, it can not run simultaneously
2377 * with itself. Exclusion is provided by memory hotplug events
2378 * being single-threaded.
2380 static void __set_migration_target_nodes(void)
2382 nodemask_t next_pass = NODE_MASK_NONE;
2383 nodemask_t this_pass = NODE_MASK_NONE;
2384 nodemask_t used_targets = NODE_MASK_NONE;
2385 int node, best_distance;
2388 * Avoid any oddities like cycles that could occur
2389 * from changes in the topology. This will leave
2390 * a momentary gap when migration is disabled.
2392 disable_all_migrate_targets();
2395 * Allocations go close to CPUs, first. Assume that
2396 * the migration path starts at the nodes with CPUs.
2398 next_pass = node_states[N_CPU];
2400 this_pass = next_pass;
2401 next_pass = NODE_MASK_NONE;
2403 * To avoid cycles in the migration "graph", ensure
2404 * that migration sources are not future targets by
2405 * setting them in 'used_targets'. Do this only
2406 * once per pass so that multiple source nodes can
2407 * share a target node.
2409 * 'used_targets' will become unavailable in future
2410 * passes. This limits some opportunities for
2411 * multiple source nodes to share a destination.
2413 nodes_or(used_targets, used_targets, this_pass);
2415 for_each_node_mask(node, this_pass) {
2419 * Try to set up the migration path for the node, and the target
2420 * migration nodes can be multiple, so doing a loop to find all
2421 * the target nodes if they all have a best node distance.
2425 establish_migrate_target(node, &used_targets,
2428 if (target_node == NUMA_NO_NODE)
2431 if (best_distance == -1)
2432 best_distance = node_distance(node, target_node);
2435 * Visit targets from this pass in the next pass.
2436 * Eventually, every node will have been part of
2437 * a pass, and will become set in 'used_targets'.
2439 node_set(target_node, next_pass);
2443 * 'next_pass' contains nodes which became migration
2444 * targets in this pass. Make additional passes until
2445 * no more migrations targets are available.
2447 if (!nodes_empty(next_pass))
2452 * For callers that do not hold get_online_mems() already.
2454 void set_migration_target_nodes(void)
2457 __set_migration_target_nodes();
2462 * This leaves migrate-on-reclaim transiently disabled between
2463 * the MEM_GOING_OFFLINE and MEM_OFFLINE events. This runs
2464 * whether reclaim-based migration is enabled or not, which
2465 * ensures that the user can turn reclaim-based migration at
2466 * any time without needing to recalculate migration targets.
2468 * These callbacks already hold get_online_mems(). That is why
2469 * __set_migration_target_nodes() can be used as opposed to
2470 * set_migration_target_nodes().
2472 static int __meminit migrate_on_reclaim_callback(struct notifier_block *self,
2473 unsigned long action, void *_arg)
2475 struct memory_notify *arg = _arg;
2478 * Only update the node migration order when a node is
2479 * changing status, like online->offline. This avoids
2480 * the overhead of synchronize_rcu() in most cases.
2482 if (arg->status_change_nid < 0)
2483 return notifier_from_errno(0);
2486 case MEM_GOING_OFFLINE:
2488 * Make sure there are not transient states where
2489 * an offline node is a migration target. This
2490 * will leave migration disabled until the offline
2491 * completes and the MEM_OFFLINE case below runs.
2493 disable_all_migrate_targets();
2498 * Recalculate the target nodes once the node
2499 * reaches its final state (online or offline).
2501 __set_migration_target_nodes();
2503 case MEM_CANCEL_OFFLINE:
2505 * MEM_GOING_OFFLINE disabled all the migration
2506 * targets. Reenable them.
2508 __set_migration_target_nodes();
2510 case MEM_GOING_ONLINE:
2511 case MEM_CANCEL_ONLINE:
2515 return notifier_from_errno(0);
2518 void __init migrate_on_reclaim_init(void)
2520 node_demotion = kmalloc_array(nr_node_ids,
2521 sizeof(struct demotion_nodes),
2523 WARN_ON(!node_demotion);
2525 hotplug_memory_notifier(migrate_on_reclaim_callback, 100);
2527 * At this point, all numa nodes with memory/CPus have their state
2528 * properly set, so we can build the demotion order now.
2529 * Let us hold the cpu_hotplug lock just, as we could possibily have
2530 * CPU hotplug events during boot.
2533 set_migration_target_nodes();
2536 #endif /* CONFIG_HOTPLUG_CPU */
2538 bool numa_demotion_enabled = false;
2541 static ssize_t numa_demotion_enabled_show(struct kobject *kobj,
2542 struct kobj_attribute *attr, char *buf)
2544 return sysfs_emit(buf, "%s\n",
2545 numa_demotion_enabled ? "true" : "false");
2548 static ssize_t numa_demotion_enabled_store(struct kobject *kobj,
2549 struct kobj_attribute *attr,
2550 const char *buf, size_t count)
2552 if (!strncmp(buf, "true", 4) || !strncmp(buf, "1", 1))
2553 numa_demotion_enabled = true;
2554 else if (!strncmp(buf, "false", 5) || !strncmp(buf, "0", 1))
2555 numa_demotion_enabled = false;
2562 static struct kobj_attribute numa_demotion_enabled_attr =
2563 __ATTR(demotion_enabled, 0644, numa_demotion_enabled_show,
2564 numa_demotion_enabled_store);
2566 static struct attribute *numa_attrs[] = {
2567 &numa_demotion_enabled_attr.attr,
2571 static const struct attribute_group numa_attr_group = {
2572 .attrs = numa_attrs,
2575 static int __init numa_init_sysfs(void)
2578 struct kobject *numa_kobj;
2580 numa_kobj = kobject_create_and_add("numa", mm_kobj);
2582 pr_err("failed to create numa kobject\n");
2585 err = sysfs_create_group(numa_kobj, &numa_attr_group);
2587 pr_err("failed to register numa group\n");
2593 kobject_put(numa_kobj);
2596 subsys_initcall(numa_init_sysfs);