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/pagewalk.h>
42 #include <linux/pfn_t.h>
43 #include <linux/memremap.h>
44 #include <linux/userfaultfd_k.h>
45 #include <linux/balloon_compaction.h>
46 #include <linux/mmu_notifier.h>
47 #include <linux/page_idle.h>
48 #include <linux/page_owner.h>
49 #include <linux/sched/mm.h>
50 #include <linux/ptrace.h>
51 #include <linux/oom.h>
52 #include <linux/memory.h>
53 #include <linux/random.h>
55 #include <asm/tlbflush.h>
57 #define CREATE_TRACE_POINTS
58 #include <trace/events/migrate.h>
62 int isolate_movable_page(struct page *page, isolate_mode_t mode)
64 struct address_space *mapping;
67 * Avoid burning cycles with pages that are yet under __free_pages(),
68 * or just got freed under us.
70 * In case we 'win' a race for a movable page being freed under us and
71 * raise its refcount preventing __free_pages() from doing its job
72 * the put_page() at the end of this block will take care of
73 * release this page, thus avoiding a nasty leakage.
75 if (unlikely(!get_page_unless_zero(page)))
79 * Check PageMovable before holding a PG_lock because page's owner
80 * assumes anybody doesn't touch PG_lock of newly allocated page
81 * so unconditionally grabbing the lock ruins page's owner side.
83 if (unlikely(!__PageMovable(page)))
86 * As movable pages are not isolated from LRU lists, concurrent
87 * compaction threads can race against page migration functions
88 * as well as race against the releasing a page.
90 * In order to avoid having an already isolated movable page
91 * being (wrongly) re-isolated while it is under migration,
92 * or to avoid attempting to isolate pages being released,
93 * lets be sure we have the page lock
94 * before proceeding with the movable page isolation steps.
96 if (unlikely(!trylock_page(page)))
99 if (!PageMovable(page) || PageIsolated(page))
100 goto out_no_isolated;
102 mapping = page_mapping(page);
103 VM_BUG_ON_PAGE(!mapping, page);
105 if (!mapping->a_ops->isolate_page(page, mode))
106 goto out_no_isolated;
108 /* Driver shouldn't use PG_isolated bit of page->flags */
109 WARN_ON_ONCE(PageIsolated(page));
110 SetPageIsolated(page);
123 static void putback_movable_page(struct page *page)
125 struct address_space *mapping;
127 mapping = page_mapping(page);
128 mapping->a_ops->putback_page(page);
129 ClearPageIsolated(page);
133 * Put previously isolated pages back onto the appropriate lists
134 * from where they were once taken off for compaction/migration.
136 * This function shall be used whenever the isolated pageset has been
137 * built from lru, balloon, hugetlbfs page. See isolate_migratepages_range()
138 * and isolate_huge_page().
140 void putback_movable_pages(struct list_head *l)
145 list_for_each_entry_safe(page, page2, l, lru) {
146 if (unlikely(PageHuge(page))) {
147 putback_active_hugepage(page);
150 list_del(&page->lru);
152 * We isolated non-lru movable page so here we can use
153 * __PageMovable because LRU page's mapping cannot have
154 * PAGE_MAPPING_MOVABLE.
156 if (unlikely(__PageMovable(page))) {
157 VM_BUG_ON_PAGE(!PageIsolated(page), page);
159 if (PageMovable(page))
160 putback_movable_page(page);
162 ClearPageIsolated(page);
166 mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON +
167 page_is_file_lru(page), -thp_nr_pages(page));
168 putback_lru_page(page);
174 * Restore a potential migration pte to a working pte entry
176 static bool remove_migration_pte(struct page *page, struct vm_area_struct *vma,
177 unsigned long addr, void *old)
179 struct page_vma_mapped_walk pvmw = {
183 .flags = PVMW_SYNC | PVMW_MIGRATION,
189 VM_BUG_ON_PAGE(PageTail(page), page);
190 while (page_vma_mapped_walk(&pvmw)) {
194 new = page - pvmw.page->index +
195 linear_page_index(vma, pvmw.address);
197 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
198 /* PMD-mapped THP migration entry */
200 VM_BUG_ON_PAGE(PageHuge(page) || !PageTransCompound(page), page);
201 remove_migration_pmd(&pvmw, new);
207 pte = pte_mkold(mk_pte(new, READ_ONCE(vma->vm_page_prot)));
208 if (pte_swp_soft_dirty(*pvmw.pte))
209 pte = pte_mksoft_dirty(pte);
212 * Recheck VMA as permissions can change since migration started
214 entry = pte_to_swp_entry(*pvmw.pte);
215 if (is_writable_migration_entry(entry))
216 pte = maybe_mkwrite(pte, vma);
217 else if (pte_swp_uffd_wp(*pvmw.pte))
218 pte = pte_mkuffd_wp(pte);
220 if (unlikely(is_device_private_page(new))) {
222 entry = make_writable_device_private_entry(
225 entry = make_readable_device_private_entry(
227 pte = swp_entry_to_pte(entry);
228 if (pte_swp_soft_dirty(*pvmw.pte))
229 pte = pte_swp_mksoft_dirty(pte);
230 if (pte_swp_uffd_wp(*pvmw.pte))
231 pte = pte_swp_mkuffd_wp(pte);
234 #ifdef CONFIG_HUGETLB_PAGE
236 unsigned int shift = huge_page_shift(hstate_vma(vma));
238 pte = pte_mkhuge(pte);
239 pte = arch_make_huge_pte(pte, shift, vma->vm_flags);
241 hugepage_add_anon_rmap(new, vma, pvmw.address);
243 page_dup_rmap(new, true);
244 set_huge_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte);
249 page_add_anon_rmap(new, vma, pvmw.address, false);
251 page_add_file_rmap(new, false);
252 set_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte);
254 if (vma->vm_flags & VM_LOCKED && !PageTransCompound(new))
257 if (PageTransHuge(page) && PageMlocked(page))
258 clear_page_mlock(page);
260 /* No need to invalidate - it was non-present before */
261 update_mmu_cache(vma, pvmw.address, pvmw.pte);
268 * Get rid of all migration entries and replace them by
269 * references to the indicated page.
271 void remove_migration_ptes(struct page *old, struct page *new, bool locked)
273 struct rmap_walk_control rwc = {
274 .rmap_one = remove_migration_pte,
279 rmap_walk_locked(new, &rwc);
281 rmap_walk(new, &rwc);
285 * Something used the pte of a page under migration. We need to
286 * get to the page and wait until migration is finished.
287 * When we return from this function the fault will be retried.
289 void __migration_entry_wait(struct mm_struct *mm, pte_t *ptep,
297 if (!is_swap_pte(pte))
300 entry = pte_to_swp_entry(pte);
301 if (!is_migration_entry(entry))
304 migration_entry_wait_on_locked(entry, ptep, ptl);
307 pte_unmap_unlock(ptep, ptl);
310 void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
311 unsigned long address)
313 spinlock_t *ptl = pte_lockptr(mm, pmd);
314 pte_t *ptep = pte_offset_map(pmd, address);
315 __migration_entry_wait(mm, ptep, ptl);
318 void migration_entry_wait_huge(struct vm_area_struct *vma,
319 struct mm_struct *mm, pte_t *pte)
321 spinlock_t *ptl = huge_pte_lockptr(hstate_vma(vma), mm, pte);
322 __migration_entry_wait(mm, pte, ptl);
325 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
326 void pmd_migration_entry_wait(struct mm_struct *mm, pmd_t *pmd)
330 ptl = pmd_lock(mm, pmd);
331 if (!is_pmd_migration_entry(*pmd))
333 migration_entry_wait_on_locked(pmd_to_swp_entry(*pmd), NULL, ptl);
340 static int expected_page_refs(struct address_space *mapping, struct page *page)
342 int expected_count = 1;
345 * Device private pages have an extra refcount as they are
348 expected_count += is_device_private_page(page);
350 expected_count += compound_nr(page) + page_has_private(page);
352 return expected_count;
356 * Replace the page in the mapping.
358 * The number of remaining references must be:
359 * 1 for anonymous pages without a mapping
360 * 2 for pages with a mapping
361 * 3 for pages with a mapping and PagePrivate/PagePrivate2 set.
363 int folio_migrate_mapping(struct address_space *mapping,
364 struct folio *newfolio, struct folio *folio, int extra_count)
366 XA_STATE(xas, &mapping->i_pages, folio_index(folio));
367 struct zone *oldzone, *newzone;
369 int expected_count = expected_page_refs(mapping, &folio->page) + extra_count;
370 long nr = folio_nr_pages(folio);
373 /* Anonymous page without mapping */
374 if (folio_ref_count(folio) != expected_count)
377 /* No turning back from here */
378 newfolio->index = folio->index;
379 newfolio->mapping = folio->mapping;
380 if (folio_test_swapbacked(folio))
381 __folio_set_swapbacked(newfolio);
383 return MIGRATEPAGE_SUCCESS;
386 oldzone = folio_zone(folio);
387 newzone = folio_zone(newfolio);
390 if (!folio_ref_freeze(folio, expected_count)) {
391 xas_unlock_irq(&xas);
396 * Now we know that no one else is looking at the folio:
397 * no turning back from here.
399 newfolio->index = folio->index;
400 newfolio->mapping = folio->mapping;
401 folio_ref_add(newfolio, nr); /* add cache reference */
402 if (folio_test_swapbacked(folio)) {
403 __folio_set_swapbacked(newfolio);
404 if (folio_test_swapcache(folio)) {
405 folio_set_swapcache(newfolio);
406 newfolio->private = folio_get_private(folio);
409 VM_BUG_ON_FOLIO(folio_test_swapcache(folio), folio);
412 /* Move dirty while page refs frozen and newpage not yet exposed */
413 dirty = folio_test_dirty(folio);
415 folio_clear_dirty(folio);
416 folio_set_dirty(newfolio);
419 xas_store(&xas, newfolio);
422 * Drop cache reference from old page by unfreezing
423 * to one less reference.
424 * We know this isn't the last reference.
426 folio_ref_unfreeze(folio, expected_count - nr);
429 /* Leave irq disabled to prevent preemption while updating stats */
432 * If moved to a different zone then also account
433 * the page for that zone. Other VM counters will be
434 * taken care of when we establish references to the
435 * new page and drop references to the old page.
437 * Note that anonymous pages are accounted for
438 * via NR_FILE_PAGES and NR_ANON_MAPPED if they
439 * are mapped to swap space.
441 if (newzone != oldzone) {
442 struct lruvec *old_lruvec, *new_lruvec;
443 struct mem_cgroup *memcg;
445 memcg = folio_memcg(folio);
446 old_lruvec = mem_cgroup_lruvec(memcg, oldzone->zone_pgdat);
447 new_lruvec = mem_cgroup_lruvec(memcg, newzone->zone_pgdat);
449 __mod_lruvec_state(old_lruvec, NR_FILE_PAGES, -nr);
450 __mod_lruvec_state(new_lruvec, NR_FILE_PAGES, nr);
451 if (folio_test_swapbacked(folio) && !folio_test_swapcache(folio)) {
452 __mod_lruvec_state(old_lruvec, NR_SHMEM, -nr);
453 __mod_lruvec_state(new_lruvec, NR_SHMEM, nr);
456 if (folio_test_swapcache(folio)) {
457 __mod_lruvec_state(old_lruvec, NR_SWAPCACHE, -nr);
458 __mod_lruvec_state(new_lruvec, NR_SWAPCACHE, nr);
461 if (dirty && mapping_can_writeback(mapping)) {
462 __mod_lruvec_state(old_lruvec, NR_FILE_DIRTY, -nr);
463 __mod_zone_page_state(oldzone, NR_ZONE_WRITE_PENDING, -nr);
464 __mod_lruvec_state(new_lruvec, NR_FILE_DIRTY, nr);
465 __mod_zone_page_state(newzone, NR_ZONE_WRITE_PENDING, nr);
470 return MIGRATEPAGE_SUCCESS;
472 EXPORT_SYMBOL(folio_migrate_mapping);
475 * The expected number of remaining references is the same as that
476 * of folio_migrate_mapping().
478 int migrate_huge_page_move_mapping(struct address_space *mapping,
479 struct page *newpage, struct page *page)
481 XA_STATE(xas, &mapping->i_pages, page_index(page));
485 expected_count = 2 + page_has_private(page);
486 if (page_count(page) != expected_count || xas_load(&xas) != page) {
487 xas_unlock_irq(&xas);
491 if (!page_ref_freeze(page, expected_count)) {
492 xas_unlock_irq(&xas);
496 newpage->index = page->index;
497 newpage->mapping = page->mapping;
501 xas_store(&xas, newpage);
503 page_ref_unfreeze(page, expected_count - 1);
505 xas_unlock_irq(&xas);
507 return MIGRATEPAGE_SUCCESS;
511 * Copy the flags and some other ancillary information
513 void folio_migrate_flags(struct folio *newfolio, struct folio *folio)
517 if (folio_test_error(folio))
518 folio_set_error(newfolio);
519 if (folio_test_referenced(folio))
520 folio_set_referenced(newfolio);
521 if (folio_test_uptodate(folio))
522 folio_mark_uptodate(newfolio);
523 if (folio_test_clear_active(folio)) {
524 VM_BUG_ON_FOLIO(folio_test_unevictable(folio), folio);
525 folio_set_active(newfolio);
526 } else if (folio_test_clear_unevictable(folio))
527 folio_set_unevictable(newfolio);
528 if (folio_test_workingset(folio))
529 folio_set_workingset(newfolio);
530 if (folio_test_checked(folio))
531 folio_set_checked(newfolio);
532 if (folio_test_mappedtodisk(folio))
533 folio_set_mappedtodisk(newfolio);
535 /* Move dirty on pages not done by folio_migrate_mapping() */
536 if (folio_test_dirty(folio))
537 folio_set_dirty(newfolio);
539 if (folio_test_young(folio))
540 folio_set_young(newfolio);
541 if (folio_test_idle(folio))
542 folio_set_idle(newfolio);
545 * Copy NUMA information to the new page, to prevent over-eager
546 * future migrations of this same page.
548 cpupid = page_cpupid_xchg_last(&folio->page, -1);
549 page_cpupid_xchg_last(&newfolio->page, cpupid);
551 folio_migrate_ksm(newfolio, folio);
553 * Please do not reorder this without considering how mm/ksm.c's
554 * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache().
556 if (folio_test_swapcache(folio))
557 folio_clear_swapcache(folio);
558 folio_clear_private(folio);
560 /* page->private contains hugetlb specific flags */
561 if (!folio_test_hugetlb(folio))
562 folio->private = NULL;
565 * If any waiters have accumulated on the new page then
568 if (folio_test_writeback(newfolio))
569 folio_end_writeback(newfolio);
572 * PG_readahead shares the same bit with PG_reclaim. The above
573 * end_page_writeback() may clear PG_readahead mistakenly, so set the
576 if (folio_test_readahead(folio))
577 folio_set_readahead(newfolio);
579 folio_copy_owner(newfolio, folio);
581 if (!folio_test_hugetlb(folio))
582 mem_cgroup_migrate(folio, newfolio);
584 EXPORT_SYMBOL(folio_migrate_flags);
586 void folio_migrate_copy(struct folio *newfolio, struct folio *folio)
588 folio_copy(newfolio, folio);
589 folio_migrate_flags(newfolio, folio);
591 EXPORT_SYMBOL(folio_migrate_copy);
593 /************************************************************
594 * Migration functions
595 ***********************************************************/
598 * Common logic to directly migrate a single LRU page suitable for
599 * pages that do not use PagePrivate/PagePrivate2.
601 * Pages are locked upon entry and exit.
603 int migrate_page(struct address_space *mapping,
604 struct page *newpage, struct page *page,
605 enum migrate_mode mode)
607 struct folio *newfolio = page_folio(newpage);
608 struct folio *folio = page_folio(page);
611 BUG_ON(folio_test_writeback(folio)); /* Writeback must be complete */
613 rc = folio_migrate_mapping(mapping, newfolio, folio, 0);
615 if (rc != MIGRATEPAGE_SUCCESS)
618 if (mode != MIGRATE_SYNC_NO_COPY)
619 folio_migrate_copy(newfolio, folio);
621 folio_migrate_flags(newfolio, folio);
622 return MIGRATEPAGE_SUCCESS;
624 EXPORT_SYMBOL(migrate_page);
627 /* Returns true if all buffers are successfully locked */
628 static bool buffer_migrate_lock_buffers(struct buffer_head *head,
629 enum migrate_mode mode)
631 struct buffer_head *bh = head;
633 /* Simple case, sync compaction */
634 if (mode != MIGRATE_ASYNC) {
637 bh = bh->b_this_page;
639 } while (bh != head);
644 /* async case, we cannot block on lock_buffer so use trylock_buffer */
646 if (!trylock_buffer(bh)) {
648 * We failed to lock the buffer and cannot stall in
649 * async migration. Release the taken locks
651 struct buffer_head *failed_bh = bh;
653 while (bh != failed_bh) {
655 bh = bh->b_this_page;
660 bh = bh->b_this_page;
661 } while (bh != head);
665 static int __buffer_migrate_page(struct address_space *mapping,
666 struct page *newpage, struct page *page, enum migrate_mode mode,
669 struct buffer_head *bh, *head;
673 if (!page_has_buffers(page))
674 return migrate_page(mapping, newpage, page, mode);
676 /* Check whether page does not have extra refs before we do more work */
677 expected_count = expected_page_refs(mapping, page);
678 if (page_count(page) != expected_count)
681 head = page_buffers(page);
682 if (!buffer_migrate_lock_buffers(head, mode))
687 bool invalidated = false;
691 spin_lock(&mapping->private_lock);
694 if (atomic_read(&bh->b_count)) {
698 bh = bh->b_this_page;
699 } while (bh != head);
705 spin_unlock(&mapping->private_lock);
706 invalidate_bh_lrus();
708 goto recheck_buffers;
712 rc = migrate_page_move_mapping(mapping, newpage, page, 0);
713 if (rc != MIGRATEPAGE_SUCCESS)
716 attach_page_private(newpage, detach_page_private(page));
720 set_bh_page(bh, newpage, bh_offset(bh));
721 bh = bh->b_this_page;
723 } while (bh != head);
725 if (mode != MIGRATE_SYNC_NO_COPY)
726 migrate_page_copy(newpage, page);
728 migrate_page_states(newpage, page);
730 rc = MIGRATEPAGE_SUCCESS;
733 spin_unlock(&mapping->private_lock);
737 bh = bh->b_this_page;
739 } while (bh != head);
745 * Migration function for pages with buffers. This function can only be used
746 * if the underlying filesystem guarantees that no other references to "page"
747 * exist. For example attached buffer heads are accessed only under page lock.
749 int buffer_migrate_page(struct address_space *mapping,
750 struct page *newpage, struct page *page, enum migrate_mode mode)
752 return __buffer_migrate_page(mapping, newpage, page, mode, false);
754 EXPORT_SYMBOL(buffer_migrate_page);
757 * Same as above except that this variant is more careful and checks that there
758 * are also no buffer head references. This function is the right one for
759 * mappings where buffer heads are directly looked up and referenced (such as
760 * block device mappings).
762 int buffer_migrate_page_norefs(struct address_space *mapping,
763 struct page *newpage, struct page *page, enum migrate_mode mode)
765 return __buffer_migrate_page(mapping, newpage, page, mode, true);
770 * Writeback a page to clean the dirty state
772 static int writeout(struct address_space *mapping, struct page *page)
774 struct writeback_control wbc = {
775 .sync_mode = WB_SYNC_NONE,
778 .range_end = LLONG_MAX,
783 if (!mapping->a_ops->writepage)
784 /* No write method for the address space */
787 if (!clear_page_dirty_for_io(page))
788 /* Someone else already triggered a write */
792 * A dirty page may imply that the underlying filesystem has
793 * the page on some queue. So the page must be clean for
794 * migration. Writeout may mean we loose the lock and the
795 * page state is no longer what we checked for earlier.
796 * At this point we know that the migration attempt cannot
799 remove_migration_ptes(page, page, false);
801 rc = mapping->a_ops->writepage(page, &wbc);
803 if (rc != AOP_WRITEPAGE_ACTIVATE)
804 /* unlocked. Relock */
807 return (rc < 0) ? -EIO : -EAGAIN;
811 * Default handling if a filesystem does not provide a migration function.
813 static int fallback_migrate_page(struct address_space *mapping,
814 struct page *newpage, struct page *page, enum migrate_mode mode)
816 if (PageDirty(page)) {
817 /* Only writeback pages in full synchronous migration */
820 case MIGRATE_SYNC_NO_COPY:
825 return writeout(mapping, page);
829 * Buffers may be managed in a filesystem specific way.
830 * We must have no buffers or drop them.
832 if (page_has_private(page) &&
833 !try_to_release_page(page, GFP_KERNEL))
834 return mode == MIGRATE_SYNC ? -EAGAIN : -EBUSY;
836 return migrate_page(mapping, newpage, page, mode);
840 * Move a page to a newly allocated page
841 * The page is locked and all ptes have been successfully removed.
843 * The new page will have replaced the old page if this function
848 * MIGRATEPAGE_SUCCESS - success
850 static int move_to_new_page(struct page *newpage, struct page *page,
851 enum migrate_mode mode)
853 struct address_space *mapping;
855 bool is_lru = !__PageMovable(page);
857 VM_BUG_ON_PAGE(!PageLocked(page), page);
858 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
860 mapping = page_mapping(page);
862 if (likely(is_lru)) {
864 rc = migrate_page(mapping, newpage, page, mode);
865 else if (mapping->a_ops->migratepage)
867 * Most pages have a mapping and most filesystems
868 * provide a migratepage callback. Anonymous pages
869 * are part of swap space which also has its own
870 * migratepage callback. This is the most common path
871 * for page migration.
873 rc = mapping->a_ops->migratepage(mapping, newpage,
876 rc = fallback_migrate_page(mapping, newpage,
880 * In case of non-lru page, it could be released after
881 * isolation step. In that case, we shouldn't try migration.
883 VM_BUG_ON_PAGE(!PageIsolated(page), page);
884 if (!PageMovable(page)) {
885 rc = MIGRATEPAGE_SUCCESS;
886 ClearPageIsolated(page);
890 rc = mapping->a_ops->migratepage(mapping, newpage,
892 WARN_ON_ONCE(rc == MIGRATEPAGE_SUCCESS &&
893 !PageIsolated(page));
897 * When successful, old pagecache page->mapping must be cleared before
898 * page is freed; but stats require that PageAnon be left as PageAnon.
900 if (rc == MIGRATEPAGE_SUCCESS) {
901 if (__PageMovable(page)) {
902 VM_BUG_ON_PAGE(!PageIsolated(page), page);
905 * We clear PG_movable under page_lock so any compactor
906 * cannot try to migrate this page.
908 ClearPageIsolated(page);
912 * Anonymous and movable page->mapping will be cleared by
913 * free_pages_prepare so don't reset it here for keeping
914 * the type to work PageAnon, for example.
916 if (!PageMappingFlags(page))
917 page->mapping = NULL;
919 if (likely(!is_zone_device_page(newpage)))
920 flush_dcache_folio(page_folio(newpage));
926 static int __unmap_and_move(struct page *page, struct page *newpage,
927 int force, enum migrate_mode mode)
930 bool page_was_mapped = false;
931 struct anon_vma *anon_vma = NULL;
932 bool is_lru = !__PageMovable(page);
934 if (!trylock_page(page)) {
935 if (!force || mode == MIGRATE_ASYNC)
939 * It's not safe for direct compaction to call lock_page.
940 * For example, during page readahead pages are added locked
941 * to the LRU. Later, when the IO completes the pages are
942 * marked uptodate and unlocked. However, the queueing
943 * could be merging multiple pages for one bio (e.g.
944 * mpage_readahead). If an allocation happens for the
945 * second or third page, the process can end up locking
946 * the same page twice and deadlocking. Rather than
947 * trying to be clever about what pages can be locked,
948 * avoid the use of lock_page for direct compaction
951 if (current->flags & PF_MEMALLOC)
957 if (PageWriteback(page)) {
959 * Only in the case of a full synchronous migration is it
960 * necessary to wait for PageWriteback. In the async case,
961 * the retry loop is too short and in the sync-light case,
962 * the overhead of stalling is too much
966 case MIGRATE_SYNC_NO_COPY:
974 wait_on_page_writeback(page);
978 * By try_to_migrate(), page->mapcount goes down to 0 here. In this case,
979 * we cannot notice that anon_vma is freed while we migrates a page.
980 * This get_anon_vma() delays freeing anon_vma pointer until the end
981 * of migration. File cache pages are no problem because of page_lock()
982 * File Caches may use write_page() or lock_page() in migration, then,
983 * just care Anon page here.
985 * Only page_get_anon_vma() understands the subtleties of
986 * getting a hold on an anon_vma from outside one of its mms.
987 * But if we cannot get anon_vma, then we won't need it anyway,
988 * because that implies that the anon page is no longer mapped
989 * (and cannot be remapped so long as we hold the page lock).
991 if (PageAnon(page) && !PageKsm(page))
992 anon_vma = page_get_anon_vma(page);
995 * Block others from accessing the new page when we get around to
996 * establishing additional references. We are usually the only one
997 * holding a reference to newpage at this point. We used to have a BUG
998 * here if trylock_page(newpage) fails, but would like to allow for
999 * cases where there might be a race with the previous use of newpage.
1000 * This is much like races on refcount of oldpage: just don't BUG().
1002 if (unlikely(!trylock_page(newpage)))
1005 if (unlikely(!is_lru)) {
1006 rc = move_to_new_page(newpage, page, mode);
1007 goto out_unlock_both;
1011 * Corner case handling:
1012 * 1. When a new swap-cache page is read into, it is added to the LRU
1013 * and treated as swapcache but it has no rmap yet.
1014 * Calling try_to_unmap() against a page->mapping==NULL page will
1015 * trigger a BUG. So handle it here.
1016 * 2. An orphaned page (see truncate_cleanup_page) might have
1017 * fs-private metadata. The page can be picked up due to memory
1018 * offlining. Everywhere else except page reclaim, the page is
1019 * invisible to the vm, so the page can not be migrated. So try to
1020 * free the metadata, so the page can be freed.
1022 if (!page->mapping) {
1023 VM_BUG_ON_PAGE(PageAnon(page), page);
1024 if (page_has_private(page)) {
1025 try_to_free_buffers(page);
1026 goto out_unlock_both;
1028 } else if (page_mapped(page)) {
1029 /* Establish migration ptes */
1030 VM_BUG_ON_PAGE(PageAnon(page) && !PageKsm(page) && !anon_vma,
1032 try_to_migrate(page, 0);
1033 page_was_mapped = true;
1036 if (!page_mapped(page))
1037 rc = move_to_new_page(newpage, page, mode);
1039 if (page_was_mapped)
1040 remove_migration_ptes(page,
1041 rc == MIGRATEPAGE_SUCCESS ? newpage : page, false);
1044 unlock_page(newpage);
1046 /* Drop an anon_vma reference if we took one */
1048 put_anon_vma(anon_vma);
1052 * If migration is successful, decrease refcount of the newpage
1053 * which will not free the page because new page owner increased
1054 * refcounter. As well, if it is LRU page, add the page to LRU
1055 * list in here. Use the old state of the isolated source page to
1056 * determine if we migrated a LRU page. newpage was already unlocked
1057 * and possibly modified by its owner - don't rely on the page
1060 if (rc == MIGRATEPAGE_SUCCESS) {
1061 if (unlikely(!is_lru))
1064 putback_lru_page(newpage);
1071 * Obtain the lock on page, remove all ptes and migrate the page
1072 * to the newly allocated page in newpage.
1074 static int unmap_and_move(new_page_t get_new_page,
1075 free_page_t put_new_page,
1076 unsigned long private, struct page *page,
1077 int force, enum migrate_mode mode,
1078 enum migrate_reason reason,
1079 struct list_head *ret)
1081 int rc = MIGRATEPAGE_SUCCESS;
1082 struct page *newpage = NULL;
1084 if (!thp_migration_supported() && PageTransHuge(page))
1087 if (page_count(page) == 1) {
1088 /* page was freed from under us. So we are done. */
1089 ClearPageActive(page);
1090 ClearPageUnevictable(page);
1091 if (unlikely(__PageMovable(page))) {
1093 if (!PageMovable(page))
1094 ClearPageIsolated(page);
1100 newpage = get_new_page(page, private);
1104 rc = __unmap_and_move(page, newpage, force, mode);
1105 if (rc == MIGRATEPAGE_SUCCESS)
1106 set_page_owner_migrate_reason(newpage, reason);
1109 if (rc != -EAGAIN) {
1111 * A page that has been migrated has all references
1112 * removed and will be freed. A page that has not been
1113 * migrated will have kept its references and be restored.
1115 list_del(&page->lru);
1119 * If migration is successful, releases reference grabbed during
1120 * isolation. Otherwise, restore the page to right list unless
1123 if (rc == MIGRATEPAGE_SUCCESS) {
1125 * Compaction can migrate also non-LRU pages which are
1126 * not accounted to NR_ISOLATED_*. They can be recognized
1129 if (likely(!__PageMovable(page)))
1130 mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON +
1131 page_is_file_lru(page), -thp_nr_pages(page));
1133 if (reason != MR_MEMORY_FAILURE)
1135 * We release the page in page_handle_poison.
1140 list_add_tail(&page->lru, ret);
1143 put_new_page(newpage, private);
1152 * Counterpart of unmap_and_move_page() for hugepage migration.
1154 * This function doesn't wait the completion of hugepage I/O
1155 * because there is no race between I/O and migration for hugepage.
1156 * Note that currently hugepage I/O occurs only in direct I/O
1157 * where no lock is held and PG_writeback is irrelevant,
1158 * and writeback status of all subpages are counted in the reference
1159 * count of the head page (i.e. if all subpages of a 2MB hugepage are
1160 * under direct I/O, the reference of the head page is 512 and a bit more.)
1161 * This means that when we try to migrate hugepage whose subpages are
1162 * doing direct I/O, some references remain after try_to_unmap() and
1163 * hugepage migration fails without data corruption.
1165 * There is also no race when direct I/O is issued on the page under migration,
1166 * because then pte is replaced with migration swap entry and direct I/O code
1167 * will wait in the page fault for migration to complete.
1169 static int unmap_and_move_huge_page(new_page_t get_new_page,
1170 free_page_t put_new_page, unsigned long private,
1171 struct page *hpage, int force,
1172 enum migrate_mode mode, int reason,
1173 struct list_head *ret)
1176 int page_was_mapped = 0;
1177 struct page *new_hpage;
1178 struct anon_vma *anon_vma = NULL;
1179 struct address_space *mapping = NULL;
1182 * Migratability of hugepages depends on architectures and their size.
1183 * This check is necessary because some callers of hugepage migration
1184 * like soft offline and memory hotremove don't walk through page
1185 * tables or check whether the hugepage is pmd-based or not before
1186 * kicking migration.
1188 if (!hugepage_migration_supported(page_hstate(hpage))) {
1189 list_move_tail(&hpage->lru, ret);
1193 if (page_count(hpage) == 1) {
1194 /* page was freed from under us. So we are done. */
1195 putback_active_hugepage(hpage);
1196 return MIGRATEPAGE_SUCCESS;
1199 new_hpage = get_new_page(hpage, private);
1203 if (!trylock_page(hpage)) {
1208 case MIGRATE_SYNC_NO_COPY:
1217 * Check for pages which are in the process of being freed. Without
1218 * page_mapping() set, hugetlbfs specific move page routine will not
1219 * be called and we could leak usage counts for subpools.
1221 if (hugetlb_page_subpool(hpage) && !page_mapping(hpage)) {
1226 if (PageAnon(hpage))
1227 anon_vma = page_get_anon_vma(hpage);
1229 if (unlikely(!trylock_page(new_hpage)))
1232 if (page_mapped(hpage)) {
1233 bool mapping_locked = false;
1234 enum ttu_flags ttu = 0;
1236 if (!PageAnon(hpage)) {
1238 * In shared mappings, try_to_unmap could potentially
1239 * call huge_pmd_unshare. Because of this, take
1240 * semaphore in write mode here and set TTU_RMAP_LOCKED
1241 * to let lower levels know we have taken the lock.
1243 mapping = hugetlb_page_mapping_lock_write(hpage);
1244 if (unlikely(!mapping))
1245 goto unlock_put_anon;
1247 mapping_locked = true;
1248 ttu |= TTU_RMAP_LOCKED;
1251 try_to_migrate(hpage, ttu);
1252 page_was_mapped = 1;
1255 i_mmap_unlock_write(mapping);
1258 if (!page_mapped(hpage))
1259 rc = move_to_new_page(new_hpage, hpage, mode);
1261 if (page_was_mapped)
1262 remove_migration_ptes(hpage,
1263 rc == MIGRATEPAGE_SUCCESS ? new_hpage : hpage, false);
1266 unlock_page(new_hpage);
1270 put_anon_vma(anon_vma);
1272 if (rc == MIGRATEPAGE_SUCCESS) {
1273 move_hugetlb_state(hpage, new_hpage, reason);
1274 put_new_page = NULL;
1280 if (rc == MIGRATEPAGE_SUCCESS)
1281 putback_active_hugepage(hpage);
1282 else if (rc != -EAGAIN)
1283 list_move_tail(&hpage->lru, ret);
1286 * If migration was not successful and there's a freeing callback, use
1287 * it. Otherwise, put_page() will drop the reference grabbed during
1291 put_new_page(new_hpage, private);
1293 putback_active_hugepage(new_hpage);
1298 static inline int try_split_thp(struct page *page, struct page **page2,
1299 struct list_head *from)
1304 rc = split_huge_page_to_list(page, from);
1307 list_safe_reset_next(page, *page2, lru);
1313 * migrate_pages - migrate the pages specified in a list, to the free pages
1314 * supplied as the target for the page migration
1316 * @from: The list of pages to be migrated.
1317 * @get_new_page: The function used to allocate free pages to be used
1318 * as the target of the page migration.
1319 * @put_new_page: The function used to free target pages if migration
1320 * fails, or NULL if no special handling is necessary.
1321 * @private: Private data to be passed on to get_new_page()
1322 * @mode: The migration mode that specifies the constraints for
1323 * page migration, if any.
1324 * @reason: The reason for page migration.
1325 * @ret_succeeded: Set to the number of normal pages migrated successfully if
1326 * the caller passes a non-NULL pointer.
1328 * The function returns after 10 attempts or if no pages are movable any more
1329 * because the list has become empty or no retryable pages exist any more.
1330 * It is caller's responsibility to call putback_movable_pages() to return pages
1331 * to the LRU or free list only if ret != 0.
1333 * Returns the number of {normal page, THP, hugetlb} that were not migrated, or
1334 * an error code. The number of THP splits will be considered as the number of
1335 * non-migrated THP, no matter how many subpages of the THP are migrated successfully.
1337 int migrate_pages(struct list_head *from, new_page_t get_new_page,
1338 free_page_t put_new_page, unsigned long private,
1339 enum migrate_mode mode, int reason, unsigned int *ret_succeeded)
1344 int nr_failed_pages = 0;
1345 int nr_succeeded = 0;
1346 int nr_thp_succeeded = 0;
1347 int nr_thp_failed = 0;
1348 int nr_thp_split = 0;
1350 bool is_thp = false;
1353 int rc, nr_subpages;
1354 LIST_HEAD(ret_pages);
1355 LIST_HEAD(thp_split_pages);
1356 bool nosplit = (reason == MR_NUMA_MISPLACED);
1357 bool no_subpage_counting = false;
1359 trace_mm_migrate_pages_start(mode, reason);
1361 thp_subpage_migration:
1362 for (pass = 0; pass < 10 && (retry || thp_retry); pass++) {
1366 list_for_each_entry_safe(page, page2, from, lru) {
1369 * THP statistics is based on the source huge page.
1370 * Capture required information that might get lost
1373 is_thp = PageTransHuge(page) && !PageHuge(page);
1374 nr_subpages = compound_nr(page);
1378 rc = unmap_and_move_huge_page(get_new_page,
1379 put_new_page, private, page,
1380 pass > 2, mode, reason,
1383 rc = unmap_and_move(get_new_page, put_new_page,
1384 private, page, pass > 2, mode,
1385 reason, &ret_pages);
1388 * Success: non hugetlb page will be freed, hugetlb
1389 * page will be put back
1390 * -EAGAIN: stay on the from list
1391 * -ENOMEM: stay on the from list
1392 * Other errno: put on ret_pages list then splice to
1397 * THP migration might be unsupported or the
1398 * allocation could've failed so we should
1399 * retry on the same page with the THP split
1402 * Head page is retried immediately and tail
1403 * pages are added to the tail of the list so
1404 * we encounter them after the rest of the list
1408 /* THP migration is unsupported */
1411 if (!try_split_thp(page, &page2, &thp_split_pages)) {
1416 nr_failed_pages += nr_subpages;
1420 /* Hugetlb migration is unsupported */
1421 if (!no_subpage_counting)
1423 nr_failed_pages += nr_subpages;
1427 * When memory is low, don't bother to try to migrate
1428 * other pages, just exit.
1429 * THP NUMA faulting doesn't split THP to retry.
1431 if (is_thp && !nosplit) {
1433 if (!try_split_thp(page, &page2, &thp_split_pages)) {
1438 nr_failed_pages += nr_subpages;
1442 if (!no_subpage_counting)
1444 nr_failed_pages += nr_subpages;
1453 case MIGRATEPAGE_SUCCESS:
1454 nr_succeeded += nr_subpages;
1462 * Permanent failure (-EBUSY, etc.):
1463 * unlike -EAGAIN case, the failed page is
1464 * removed from migration page list and not
1465 * retried in the next outer loop.
1469 nr_failed_pages += nr_subpages;
1473 if (!no_subpage_counting)
1475 nr_failed_pages += nr_subpages;
1481 nr_thp_failed += thp_retry;
1483 * Try to migrate subpages of fail-to-migrate THPs, no nr_failed
1484 * counting in this round, since all subpages of a THP is counted
1485 * as 1 failure in the first round.
1487 if (!list_empty(&thp_split_pages)) {
1489 * Move non-migrated pages (after 10 retries) to ret_pages
1490 * to avoid migrating them again.
1492 list_splice_init(from, &ret_pages);
1493 list_splice_init(&thp_split_pages, from);
1494 no_subpage_counting = true;
1496 goto thp_subpage_migration;
1499 rc = nr_failed + nr_thp_failed;
1502 * Put the permanent failure page back to migration list, they
1503 * will be put back to the right list by the caller.
1505 list_splice(&ret_pages, from);
1507 count_vm_events(PGMIGRATE_SUCCESS, nr_succeeded);
1508 count_vm_events(PGMIGRATE_FAIL, nr_failed_pages);
1509 count_vm_events(THP_MIGRATION_SUCCESS, nr_thp_succeeded);
1510 count_vm_events(THP_MIGRATION_FAIL, nr_thp_failed);
1511 count_vm_events(THP_MIGRATION_SPLIT, nr_thp_split);
1512 trace_mm_migrate_pages(nr_succeeded, nr_failed_pages, nr_thp_succeeded,
1513 nr_thp_failed, nr_thp_split, mode, reason);
1516 *ret_succeeded = nr_succeeded;
1521 struct page *alloc_migration_target(struct page *page, unsigned long private)
1523 struct migration_target_control *mtc;
1525 unsigned int order = 0;
1526 struct page *new_page = NULL;
1530 mtc = (struct migration_target_control *)private;
1531 gfp_mask = mtc->gfp_mask;
1533 if (nid == NUMA_NO_NODE)
1534 nid = page_to_nid(page);
1536 if (PageHuge(page)) {
1537 struct hstate *h = page_hstate(compound_head(page));
1539 gfp_mask = htlb_modify_alloc_mask(h, gfp_mask);
1540 return alloc_huge_page_nodemask(h, nid, mtc->nmask, gfp_mask);
1543 if (PageTransHuge(page)) {
1545 * clear __GFP_RECLAIM to make the migration callback
1546 * consistent with regular THP allocations.
1548 gfp_mask &= ~__GFP_RECLAIM;
1549 gfp_mask |= GFP_TRANSHUGE;
1550 order = HPAGE_PMD_ORDER;
1552 zidx = zone_idx(page_zone(page));
1553 if (is_highmem_idx(zidx) || zidx == ZONE_MOVABLE)
1554 gfp_mask |= __GFP_HIGHMEM;
1556 new_page = __alloc_pages(gfp_mask, order, nid, mtc->nmask);
1558 if (new_page && PageTransHuge(new_page))
1559 prep_transhuge_page(new_page);
1566 static int store_status(int __user *status, int start, int value, int nr)
1569 if (put_user(value, status + start))
1577 static int do_move_pages_to_node(struct mm_struct *mm,
1578 struct list_head *pagelist, int node)
1581 struct migration_target_control mtc = {
1583 .gfp_mask = GFP_HIGHUSER_MOVABLE | __GFP_THISNODE,
1586 err = migrate_pages(pagelist, alloc_migration_target, NULL,
1587 (unsigned long)&mtc, MIGRATE_SYNC, MR_SYSCALL, NULL);
1589 putback_movable_pages(pagelist);
1594 * Resolves the given address to a struct page, isolates it from the LRU and
1595 * puts it to the given pagelist.
1597 * errno - if the page cannot be found/isolated
1598 * 0 - when it doesn't have to be migrated because it is already on the
1600 * 1 - when it has been queued
1602 static int add_page_for_migration(struct mm_struct *mm, unsigned long addr,
1603 int node, struct list_head *pagelist, bool migrate_all)
1605 struct vm_area_struct *vma;
1611 vma = find_vma(mm, addr);
1612 if (!vma || addr < vma->vm_start || !vma_migratable(vma))
1615 /* FOLL_DUMP to ignore special (like zero) pages */
1616 page = follow_page(vma, addr, FOLL_GET | FOLL_DUMP);
1618 err = PTR_ERR(page);
1627 if (page_to_nid(page) == node)
1631 if (page_mapcount(page) > 1 && !migrate_all)
1634 if (PageHuge(page)) {
1635 if (PageHead(page)) {
1636 isolate_huge_page(page, pagelist);
1642 head = compound_head(page);
1643 err = isolate_lru_page(head);
1648 list_add_tail(&head->lru, pagelist);
1649 mod_node_page_state(page_pgdat(head),
1650 NR_ISOLATED_ANON + page_is_file_lru(head),
1651 thp_nr_pages(head));
1655 * Either remove the duplicate refcount from
1656 * isolate_lru_page() or drop the page ref if it was
1661 mmap_read_unlock(mm);
1665 static int move_pages_and_store_status(struct mm_struct *mm, int node,
1666 struct list_head *pagelist, int __user *status,
1667 int start, int i, unsigned long nr_pages)
1671 if (list_empty(pagelist))
1674 err = do_move_pages_to_node(mm, pagelist, node);
1677 * Positive err means the number of failed
1678 * pages to migrate. Since we are going to
1679 * abort and return the number of non-migrated
1680 * pages, so need to include the rest of the
1681 * nr_pages that have not been attempted as
1685 err += nr_pages - i - 1;
1688 return store_status(status, start, node, i - start);
1692 * Migrate an array of page address onto an array of nodes and fill
1693 * the corresponding array of status.
1695 static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes,
1696 unsigned long nr_pages,
1697 const void __user * __user *pages,
1698 const int __user *nodes,
1699 int __user *status, int flags)
1701 int current_node = NUMA_NO_NODE;
1702 LIST_HEAD(pagelist);
1706 lru_cache_disable();
1708 for (i = start = 0; i < nr_pages; i++) {
1709 const void __user *p;
1714 if (get_user(p, pages + i))
1716 if (get_user(node, nodes + i))
1718 addr = (unsigned long)untagged_addr(p);
1721 if (node < 0 || node >= MAX_NUMNODES)
1723 if (!node_state(node, N_MEMORY))
1727 if (!node_isset(node, task_nodes))
1730 if (current_node == NUMA_NO_NODE) {
1731 current_node = node;
1733 } else if (node != current_node) {
1734 err = move_pages_and_store_status(mm, current_node,
1735 &pagelist, status, start, i, nr_pages);
1739 current_node = node;
1743 * Errors in the page lookup or isolation are not fatal and we simply
1744 * report them via status
1746 err = add_page_for_migration(mm, addr, current_node,
1747 &pagelist, flags & MPOL_MF_MOVE_ALL);
1750 /* The page is successfully queued for migration */
1755 * The move_pages() man page does not have an -EEXIST choice, so
1756 * use -EFAULT instead.
1762 * If the page is already on the target node (!err), store the
1763 * node, otherwise, store the err.
1765 err = store_status(status, i, err ? : current_node, 1);
1769 err = move_pages_and_store_status(mm, current_node, &pagelist,
1770 status, start, i, nr_pages);
1773 current_node = NUMA_NO_NODE;
1776 /* Make sure we do not overwrite the existing error */
1777 err1 = move_pages_and_store_status(mm, current_node, &pagelist,
1778 status, start, i, nr_pages);
1787 * Determine the nodes of an array of pages and store it in an array of status.
1789 static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages,
1790 const void __user **pages, int *status)
1796 for (i = 0; i < nr_pages; i++) {
1797 unsigned long addr = (unsigned long)(*pages);
1798 struct vm_area_struct *vma;
1802 vma = vma_lookup(mm, addr);
1806 /* FOLL_DUMP to ignore special (like zero) pages */
1807 page = follow_page(vma, addr, FOLL_DUMP);
1809 err = PTR_ERR(page);
1813 err = page ? page_to_nid(page) : -ENOENT;
1821 mmap_read_unlock(mm);
1824 static int get_compat_pages_array(const void __user *chunk_pages[],
1825 const void __user * __user *pages,
1826 unsigned long chunk_nr)
1828 compat_uptr_t __user *pages32 = (compat_uptr_t __user *)pages;
1832 for (i = 0; i < chunk_nr; i++) {
1833 if (get_user(p, pages32 + i))
1835 chunk_pages[i] = compat_ptr(p);
1842 * Determine the nodes of a user array of pages and store it in
1843 * a user array of status.
1845 static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages,
1846 const void __user * __user *pages,
1849 #define DO_PAGES_STAT_CHUNK_NR 16
1850 const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR];
1851 int chunk_status[DO_PAGES_STAT_CHUNK_NR];
1854 unsigned long chunk_nr;
1856 chunk_nr = nr_pages;
1857 if (chunk_nr > DO_PAGES_STAT_CHUNK_NR)
1858 chunk_nr = DO_PAGES_STAT_CHUNK_NR;
1860 if (in_compat_syscall()) {
1861 if (get_compat_pages_array(chunk_pages, pages,
1865 if (copy_from_user(chunk_pages, pages,
1866 chunk_nr * sizeof(*chunk_pages)))
1870 do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status);
1872 if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status)))
1877 nr_pages -= chunk_nr;
1879 return nr_pages ? -EFAULT : 0;
1882 static struct mm_struct *find_mm_struct(pid_t pid, nodemask_t *mem_nodes)
1884 struct task_struct *task;
1885 struct mm_struct *mm;
1888 * There is no need to check if current process has the right to modify
1889 * the specified process when they are same.
1893 *mem_nodes = cpuset_mems_allowed(current);
1897 /* Find the mm_struct */
1899 task = find_task_by_vpid(pid);
1902 return ERR_PTR(-ESRCH);
1904 get_task_struct(task);
1907 * Check if this process has the right to modify the specified
1908 * process. Use the regular "ptrace_may_access()" checks.
1910 if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) {
1912 mm = ERR_PTR(-EPERM);
1917 mm = ERR_PTR(security_task_movememory(task));
1920 *mem_nodes = cpuset_mems_allowed(task);
1921 mm = get_task_mm(task);
1923 put_task_struct(task);
1925 mm = ERR_PTR(-EINVAL);
1930 * Move a list of pages in the address space of the currently executing
1933 static int kernel_move_pages(pid_t pid, unsigned long nr_pages,
1934 const void __user * __user *pages,
1935 const int __user *nodes,
1936 int __user *status, int flags)
1938 struct mm_struct *mm;
1940 nodemask_t task_nodes;
1943 if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
1946 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1949 mm = find_mm_struct(pid, &task_nodes);
1954 err = do_pages_move(mm, task_nodes, nr_pages, pages,
1955 nodes, status, flags);
1957 err = do_pages_stat(mm, nr_pages, pages, status);
1963 SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages,
1964 const void __user * __user *, pages,
1965 const int __user *, nodes,
1966 int __user *, status, int, flags)
1968 return kernel_move_pages(pid, nr_pages, pages, nodes, status, flags);
1971 #ifdef CONFIG_NUMA_BALANCING
1973 * Returns true if this is a safe migration target node for misplaced NUMA
1974 * pages. Currently it only checks the watermarks which crude
1976 static bool migrate_balanced_pgdat(struct pglist_data *pgdat,
1977 unsigned long nr_migrate_pages)
1981 for (z = pgdat->nr_zones - 1; z >= 0; z--) {
1982 struct zone *zone = pgdat->node_zones + z;
1984 if (!populated_zone(zone))
1987 /* Avoid waking kswapd by allocating pages_to_migrate pages. */
1988 if (!zone_watermark_ok(zone, 0,
1989 high_wmark_pages(zone) +
1998 static struct page *alloc_misplaced_dst_page(struct page *page,
2001 int nid = (int) data;
2002 struct page *newpage;
2004 newpage = __alloc_pages_node(nid,
2005 (GFP_HIGHUSER_MOVABLE |
2006 __GFP_THISNODE | __GFP_NOMEMALLOC |
2007 __GFP_NORETRY | __GFP_NOWARN) &
2013 static struct page *alloc_misplaced_dst_page_thp(struct page *page,
2016 int nid = (int) data;
2017 struct page *newpage;
2019 newpage = alloc_pages_node(nid, (GFP_TRANSHUGE_LIGHT | __GFP_THISNODE),
2024 prep_transhuge_page(newpage);
2030 static int numamigrate_isolate_page(pg_data_t *pgdat, struct page *page)
2033 int nr_pages = thp_nr_pages(page);
2035 VM_BUG_ON_PAGE(compound_order(page) && !PageTransHuge(page), page);
2037 /* Do not migrate THP mapped by multiple processes */
2038 if (PageTransHuge(page) && total_mapcount(page) > 1)
2041 /* Avoid migrating to a node that is nearly full */
2042 if (!migrate_balanced_pgdat(pgdat, nr_pages))
2045 if (isolate_lru_page(page))
2048 page_lru = page_is_file_lru(page);
2049 mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON + page_lru,
2053 * Isolating the page has taken another reference, so the
2054 * caller's reference can be safely dropped without the page
2055 * disappearing underneath us during migration.
2062 * Attempt to migrate a misplaced page to the specified destination
2063 * node. Caller is expected to have an elevated reference count on
2064 * the page that will be dropped by this function before returning.
2066 int migrate_misplaced_page(struct page *page, struct vm_area_struct *vma,
2069 pg_data_t *pgdat = NODE_DATA(node);
2072 LIST_HEAD(migratepages);
2075 int nr_pages = thp_nr_pages(page);
2078 * PTE mapped THP or HugeTLB page can't reach here so the page could
2079 * be either base page or THP. And it must be head page if it is
2082 compound = PageTransHuge(page);
2085 new = alloc_misplaced_dst_page_thp;
2087 new = alloc_misplaced_dst_page;
2090 * Don't migrate file pages that are mapped in multiple processes
2091 * with execute permissions as they are probably shared libraries.
2093 if (page_mapcount(page) != 1 && page_is_file_lru(page) &&
2094 (vma->vm_flags & VM_EXEC))
2098 * Also do not migrate dirty pages as not all filesystems can move
2099 * dirty pages in MIGRATE_ASYNC mode which is a waste of cycles.
2101 if (page_is_file_lru(page) && PageDirty(page))
2104 isolated = numamigrate_isolate_page(pgdat, page);
2108 list_add(&page->lru, &migratepages);
2109 nr_remaining = migrate_pages(&migratepages, *new, NULL, node,
2110 MIGRATE_ASYNC, MR_NUMA_MISPLACED, NULL);
2112 if (!list_empty(&migratepages)) {
2113 list_del(&page->lru);
2114 mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON +
2115 page_is_file_lru(page), -nr_pages);
2116 putback_lru_page(page);
2120 count_vm_numa_events(NUMA_PAGE_MIGRATE, nr_pages);
2121 BUG_ON(!list_empty(&migratepages));
2128 #endif /* CONFIG_NUMA_BALANCING */
2129 #endif /* CONFIG_NUMA */
2131 #ifdef CONFIG_DEVICE_PRIVATE
2132 static int migrate_vma_collect_skip(unsigned long start,
2134 struct mm_walk *walk)
2136 struct migrate_vma *migrate = walk->private;
2139 for (addr = start; addr < end; addr += PAGE_SIZE) {
2140 migrate->dst[migrate->npages] = 0;
2141 migrate->src[migrate->npages++] = 0;
2147 static int migrate_vma_collect_hole(unsigned long start,
2149 __always_unused int depth,
2150 struct mm_walk *walk)
2152 struct migrate_vma *migrate = walk->private;
2155 /* Only allow populating anonymous memory. */
2156 if (!vma_is_anonymous(walk->vma))
2157 return migrate_vma_collect_skip(start, end, walk);
2159 for (addr = start; addr < end; addr += PAGE_SIZE) {
2160 migrate->src[migrate->npages] = MIGRATE_PFN_MIGRATE;
2161 migrate->dst[migrate->npages] = 0;
2169 static int migrate_vma_collect_pmd(pmd_t *pmdp,
2170 unsigned long start,
2172 struct mm_walk *walk)
2174 struct migrate_vma *migrate = walk->private;
2175 struct vm_area_struct *vma = walk->vma;
2176 struct mm_struct *mm = vma->vm_mm;
2177 unsigned long addr = start, unmapped = 0;
2182 if (pmd_none(*pmdp))
2183 return migrate_vma_collect_hole(start, end, -1, walk);
2185 if (pmd_trans_huge(*pmdp)) {
2188 ptl = pmd_lock(mm, pmdp);
2189 if (unlikely(!pmd_trans_huge(*pmdp))) {
2194 page = pmd_page(*pmdp);
2195 if (is_huge_zero_page(page)) {
2197 split_huge_pmd(vma, pmdp, addr);
2198 if (pmd_trans_unstable(pmdp))
2199 return migrate_vma_collect_skip(start, end,
2206 if (unlikely(!trylock_page(page)))
2207 return migrate_vma_collect_skip(start, end,
2209 ret = split_huge_page(page);
2213 return migrate_vma_collect_skip(start, end,
2215 if (pmd_none(*pmdp))
2216 return migrate_vma_collect_hole(start, end, -1,
2221 if (unlikely(pmd_bad(*pmdp)))
2222 return migrate_vma_collect_skip(start, end, walk);
2224 ptep = pte_offset_map_lock(mm, pmdp, addr, &ptl);
2225 arch_enter_lazy_mmu_mode();
2227 for (; addr < end; addr += PAGE_SIZE, ptep++) {
2228 unsigned long mpfn = 0, pfn;
2235 if (pte_none(pte)) {
2236 if (vma_is_anonymous(vma)) {
2237 mpfn = MIGRATE_PFN_MIGRATE;
2243 if (!pte_present(pte)) {
2245 * Only care about unaddressable device page special
2246 * page table entry. Other special swap entries are not
2247 * migratable, and we ignore regular swapped page.
2249 entry = pte_to_swp_entry(pte);
2250 if (!is_device_private_entry(entry))
2253 page = pfn_swap_entry_to_page(entry);
2254 if (!(migrate->flags &
2255 MIGRATE_VMA_SELECT_DEVICE_PRIVATE) ||
2256 page->pgmap->owner != migrate->pgmap_owner)
2259 mpfn = migrate_pfn(page_to_pfn(page)) |
2260 MIGRATE_PFN_MIGRATE;
2261 if (is_writable_device_private_entry(entry))
2262 mpfn |= MIGRATE_PFN_WRITE;
2264 if (!(migrate->flags & MIGRATE_VMA_SELECT_SYSTEM))
2267 if (is_zero_pfn(pfn)) {
2268 mpfn = MIGRATE_PFN_MIGRATE;
2272 page = vm_normal_page(migrate->vma, addr, pte);
2273 mpfn = migrate_pfn(pfn) | MIGRATE_PFN_MIGRATE;
2274 mpfn |= pte_write(pte) ? MIGRATE_PFN_WRITE : 0;
2277 /* FIXME support THP */
2278 if (!page || !page->mapping || PageTransCompound(page)) {
2284 * By getting a reference on the page we pin it and that blocks
2285 * any kind of migration. Side effect is that it "freezes" the
2288 * We drop this reference after isolating the page from the lru
2289 * for non device page (device page are not on the lru and thus
2290 * can't be dropped from it).
2295 * Optimize for the common case where page is only mapped once
2296 * in one process. If we can lock the page, then we can safely
2297 * set up a special migration page table entry now.
2299 if (trylock_page(page)) {
2303 ptep_get_and_clear(mm, addr, ptep);
2305 /* Setup special migration page table entry */
2306 if (mpfn & MIGRATE_PFN_WRITE)
2307 entry = make_writable_migration_entry(
2310 entry = make_readable_migration_entry(
2312 swp_pte = swp_entry_to_pte(entry);
2313 if (pte_present(pte)) {
2314 if (pte_soft_dirty(pte))
2315 swp_pte = pte_swp_mksoft_dirty(swp_pte);
2316 if (pte_uffd_wp(pte))
2317 swp_pte = pte_swp_mkuffd_wp(swp_pte);
2319 if (pte_swp_soft_dirty(pte))
2320 swp_pte = pte_swp_mksoft_dirty(swp_pte);
2321 if (pte_swp_uffd_wp(pte))
2322 swp_pte = pte_swp_mkuffd_wp(swp_pte);
2324 set_pte_at(mm, addr, ptep, swp_pte);
2327 * This is like regular unmap: we remove the rmap and
2328 * drop page refcount. Page won't be freed, as we took
2329 * a reference just above.
2331 page_remove_rmap(page, false);
2334 if (pte_present(pte))
2342 migrate->dst[migrate->npages] = 0;
2343 migrate->src[migrate->npages++] = mpfn;
2345 arch_leave_lazy_mmu_mode();
2346 pte_unmap_unlock(ptep - 1, ptl);
2348 /* Only flush the TLB if we actually modified any entries */
2350 flush_tlb_range(walk->vma, start, end);
2355 static const struct mm_walk_ops migrate_vma_walk_ops = {
2356 .pmd_entry = migrate_vma_collect_pmd,
2357 .pte_hole = migrate_vma_collect_hole,
2361 * migrate_vma_collect() - collect pages over a range of virtual addresses
2362 * @migrate: migrate struct containing all migration information
2364 * This will walk the CPU page table. For each virtual address backed by a
2365 * valid page, it updates the src array and takes a reference on the page, in
2366 * order to pin the page until we lock it and unmap it.
2368 static void migrate_vma_collect(struct migrate_vma *migrate)
2370 struct mmu_notifier_range range;
2373 * Note that the pgmap_owner is passed to the mmu notifier callback so
2374 * that the registered device driver can skip invalidating device
2375 * private page mappings that won't be migrated.
2377 mmu_notifier_range_init_owner(&range, MMU_NOTIFY_MIGRATE, 0,
2378 migrate->vma, migrate->vma->vm_mm, migrate->start, migrate->end,
2379 migrate->pgmap_owner);
2380 mmu_notifier_invalidate_range_start(&range);
2382 walk_page_range(migrate->vma->vm_mm, migrate->start, migrate->end,
2383 &migrate_vma_walk_ops, migrate);
2385 mmu_notifier_invalidate_range_end(&range);
2386 migrate->end = migrate->start + (migrate->npages << PAGE_SHIFT);
2390 * migrate_vma_check_page() - check if page is pinned or not
2391 * @page: struct page to check
2393 * Pinned pages cannot be migrated. This is the same test as in
2394 * folio_migrate_mapping(), except that here we allow migration of a
2397 static bool migrate_vma_check_page(struct page *page)
2400 * One extra ref because caller holds an extra reference, either from
2401 * isolate_lru_page() for a regular page, or migrate_vma_collect() for
2407 * FIXME support THP (transparent huge page), it is bit more complex to
2408 * check them than regular pages, because they can be mapped with a pmd
2409 * or with a pte (split pte mapping).
2411 if (PageCompound(page))
2414 /* Page from ZONE_DEVICE have one extra reference */
2415 if (is_zone_device_page(page))
2418 /* For file back page */
2419 if (page_mapping(page))
2420 extra += 1 + page_has_private(page);
2422 if ((page_count(page) - extra) > page_mapcount(page))
2429 * migrate_vma_unmap() - replace page mapping with special migration pte entry
2430 * @migrate: migrate struct containing all migration information
2432 * Isolate pages from the LRU and replace mappings (CPU page table pte) with a
2433 * special migration pte entry and check if it has been pinned. Pinned pages are
2434 * restored because we cannot migrate them.
2436 * This is the last step before we call the device driver callback to allocate
2437 * destination memory and copy contents of original page over to new page.
2439 static void migrate_vma_unmap(struct migrate_vma *migrate)
2441 const unsigned long npages = migrate->npages;
2442 unsigned long i, restore = 0;
2443 bool allow_drain = true;
2447 for (i = 0; i < npages; i++) {
2448 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2453 /* ZONE_DEVICE pages are not on LRU */
2454 if (!is_zone_device_page(page)) {
2455 if (!PageLRU(page) && allow_drain) {
2456 /* Drain CPU's pagevec */
2457 lru_add_drain_all();
2458 allow_drain = false;
2461 if (isolate_lru_page(page)) {
2462 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2468 /* Drop the reference we took in collect */
2472 if (page_mapped(page))
2473 try_to_migrate(page, 0);
2475 if (page_mapped(page) || !migrate_vma_check_page(page)) {
2476 if (!is_zone_device_page(page)) {
2478 putback_lru_page(page);
2481 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2488 for (i = 0; i < npages && restore; i++) {
2489 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2491 if (!page || (migrate->src[i] & MIGRATE_PFN_MIGRATE))
2494 remove_migration_ptes(page, page, false);
2496 migrate->src[i] = 0;
2504 * migrate_vma_setup() - prepare to migrate a range of memory
2505 * @args: contains the vma, start, and pfns arrays for the migration
2507 * Returns: negative errno on failures, 0 when 0 or more pages were migrated
2510 * Prepare to migrate a range of memory virtual address range by collecting all
2511 * the pages backing each virtual address in the range, saving them inside the
2512 * src array. Then lock those pages and unmap them. Once the pages are locked
2513 * and unmapped, check whether each page is pinned or not. Pages that aren't
2514 * pinned have the MIGRATE_PFN_MIGRATE flag set (by this function) in the
2515 * corresponding src array entry. Then restores any pages that are pinned, by
2516 * remapping and unlocking those pages.
2518 * The caller should then allocate destination memory and copy source memory to
2519 * it for all those entries (ie with MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE
2520 * flag set). Once these are allocated and copied, the caller must update each
2521 * corresponding entry in the dst array with the pfn value of the destination
2522 * page and with MIGRATE_PFN_VALID. Destination pages must be locked via
2525 * Note that the caller does not have to migrate all the pages that are marked
2526 * with MIGRATE_PFN_MIGRATE flag in src array unless this is a migration from
2527 * device memory to system memory. If the caller cannot migrate a device page
2528 * back to system memory, then it must return VM_FAULT_SIGBUS, which has severe
2529 * consequences for the userspace process, so it must be avoided if at all
2532 * For empty entries inside CPU page table (pte_none() or pmd_none() is true) we
2533 * do set MIGRATE_PFN_MIGRATE flag inside the corresponding source array thus
2534 * allowing the caller to allocate device memory for those unbacked virtual
2535 * addresses. For this the caller simply has to allocate device memory and
2536 * properly set the destination entry like for regular migration. Note that
2537 * this can still fail, and thus inside the device driver you must check if the
2538 * migration was successful for those entries after calling migrate_vma_pages(),
2539 * just like for regular migration.
2541 * After that, the callers must call migrate_vma_pages() to go over each entry
2542 * in the src array that has the MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE flag
2543 * set. If the corresponding entry in dst array has MIGRATE_PFN_VALID flag set,
2544 * then migrate_vma_pages() to migrate struct page information from the source
2545 * struct page to the destination struct page. If it fails to migrate the
2546 * struct page information, then it clears the MIGRATE_PFN_MIGRATE flag in the
2549 * At this point all successfully migrated pages have an entry in the src
2550 * array with MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE flag set and the dst
2551 * array entry with MIGRATE_PFN_VALID flag set.
2553 * Once migrate_vma_pages() returns the caller may inspect which pages were
2554 * successfully migrated, and which were not. Successfully migrated pages will
2555 * have the MIGRATE_PFN_MIGRATE flag set for their src array entry.
2557 * It is safe to update device page table after migrate_vma_pages() because
2558 * both destination and source page are still locked, and the mmap_lock is held
2559 * in read mode (hence no one can unmap the range being migrated).
2561 * Once the caller is done cleaning up things and updating its page table (if it
2562 * chose to do so, this is not an obligation) it finally calls
2563 * migrate_vma_finalize() to update the CPU page table to point to new pages
2564 * for successfully migrated pages or otherwise restore the CPU page table to
2565 * point to the original source pages.
2567 int migrate_vma_setup(struct migrate_vma *args)
2569 long nr_pages = (args->end - args->start) >> PAGE_SHIFT;
2571 args->start &= PAGE_MASK;
2572 args->end &= PAGE_MASK;
2573 if (!args->vma || is_vm_hugetlb_page(args->vma) ||
2574 (args->vma->vm_flags & VM_SPECIAL) || vma_is_dax(args->vma))
2578 if (args->start < args->vma->vm_start ||
2579 args->start >= args->vma->vm_end)
2581 if (args->end <= args->vma->vm_start || args->end > args->vma->vm_end)
2583 if (!args->src || !args->dst)
2586 memset(args->src, 0, sizeof(*args->src) * nr_pages);
2590 migrate_vma_collect(args);
2593 migrate_vma_unmap(args);
2596 * At this point pages are locked and unmapped, and thus they have
2597 * stable content and can safely be copied to destination memory that
2598 * is allocated by the drivers.
2603 EXPORT_SYMBOL(migrate_vma_setup);
2606 * This code closely matches the code in:
2607 * __handle_mm_fault()
2608 * handle_pte_fault()
2609 * do_anonymous_page()
2610 * to map in an anonymous zero page but the struct page will be a ZONE_DEVICE
2613 static void migrate_vma_insert_page(struct migrate_vma *migrate,
2618 struct vm_area_struct *vma = migrate->vma;
2619 struct mm_struct *mm = vma->vm_mm;
2629 /* Only allow populating anonymous memory */
2630 if (!vma_is_anonymous(vma))
2633 pgdp = pgd_offset(mm, addr);
2634 p4dp = p4d_alloc(mm, pgdp, addr);
2637 pudp = pud_alloc(mm, p4dp, addr);
2640 pmdp = pmd_alloc(mm, pudp, addr);
2644 if (pmd_trans_huge(*pmdp) || pmd_devmap(*pmdp))
2648 * Use pte_alloc() instead of pte_alloc_map(). We can't run
2649 * pte_offset_map() on pmds where a huge pmd might be created
2650 * from a different thread.
2652 * pte_alloc_map() is safe to use under mmap_write_lock(mm) or when
2653 * parallel threads are excluded by other means.
2655 * Here we only have mmap_read_lock(mm).
2657 if (pte_alloc(mm, pmdp))
2660 /* See the comment in pte_alloc_one_map() */
2661 if (unlikely(pmd_trans_unstable(pmdp)))
2664 if (unlikely(anon_vma_prepare(vma)))
2666 if (mem_cgroup_charge(page_folio(page), vma->vm_mm, GFP_KERNEL))
2670 * The memory barrier inside __SetPageUptodate makes sure that
2671 * preceding stores to the page contents become visible before
2672 * the set_pte_at() write.
2674 __SetPageUptodate(page);
2676 if (is_zone_device_page(page)) {
2677 if (is_device_private_page(page)) {
2678 swp_entry_t swp_entry;
2680 if (vma->vm_flags & VM_WRITE)
2681 swp_entry = make_writable_device_private_entry(
2684 swp_entry = make_readable_device_private_entry(
2686 entry = swp_entry_to_pte(swp_entry);
2689 * For now we only support migrating to un-addressable
2692 pr_warn_once("Unsupported ZONE_DEVICE page type.\n");
2696 entry = mk_pte(page, vma->vm_page_prot);
2697 if (vma->vm_flags & VM_WRITE)
2698 entry = pte_mkwrite(pte_mkdirty(entry));
2701 ptep = pte_offset_map_lock(mm, pmdp, addr, &ptl);
2703 if (check_stable_address_space(mm))
2706 if (pte_present(*ptep)) {
2707 unsigned long pfn = pte_pfn(*ptep);
2709 if (!is_zero_pfn(pfn))
2712 } else if (!pte_none(*ptep))
2716 * Check for userfaultfd but do not deliver the fault. Instead,
2719 if (userfaultfd_missing(vma))
2722 inc_mm_counter(mm, MM_ANONPAGES);
2723 page_add_new_anon_rmap(page, vma, addr, false);
2724 if (!is_zone_device_page(page))
2725 lru_cache_add_inactive_or_unevictable(page, vma);
2729 flush_cache_page(vma, addr, pte_pfn(*ptep));
2730 ptep_clear_flush_notify(vma, addr, ptep);
2731 set_pte_at_notify(mm, addr, ptep, entry);
2732 update_mmu_cache(vma, addr, ptep);
2734 /* No need to invalidate - it was non-present before */
2735 set_pte_at(mm, addr, ptep, entry);
2736 update_mmu_cache(vma, addr, ptep);
2739 pte_unmap_unlock(ptep, ptl);
2740 *src = MIGRATE_PFN_MIGRATE;
2744 pte_unmap_unlock(ptep, ptl);
2746 *src &= ~MIGRATE_PFN_MIGRATE;
2750 * migrate_vma_pages() - migrate meta-data from src page to dst page
2751 * @migrate: migrate struct containing all migration information
2753 * This migrates struct page meta-data from source struct page to destination
2754 * struct page. This effectively finishes the migration from source page to the
2757 void migrate_vma_pages(struct migrate_vma *migrate)
2759 const unsigned long npages = migrate->npages;
2760 const unsigned long start = migrate->start;
2761 struct mmu_notifier_range range;
2762 unsigned long addr, i;
2763 bool notified = false;
2765 for (i = 0, addr = start; i < npages; addr += PAGE_SIZE, i++) {
2766 struct page *newpage = migrate_pfn_to_page(migrate->dst[i]);
2767 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2768 struct address_space *mapping;
2772 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2777 if (!(migrate->src[i] & MIGRATE_PFN_MIGRATE))
2782 mmu_notifier_range_init_owner(&range,
2783 MMU_NOTIFY_MIGRATE, 0, migrate->vma,
2784 migrate->vma->vm_mm, addr, migrate->end,
2785 migrate->pgmap_owner);
2786 mmu_notifier_invalidate_range_start(&range);
2788 migrate_vma_insert_page(migrate, addr, newpage,
2793 mapping = page_mapping(page);
2795 if (is_zone_device_page(newpage)) {
2796 if (is_device_private_page(newpage)) {
2798 * For now only support private anonymous when
2799 * migrating to un-addressable device memory.
2802 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2807 * Other types of ZONE_DEVICE page are not
2810 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2815 r = migrate_page(mapping, newpage, page, MIGRATE_SYNC_NO_COPY);
2816 if (r != MIGRATEPAGE_SUCCESS)
2817 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2821 * No need to double call mmu_notifier->invalidate_range() callback as
2822 * the above ptep_clear_flush_notify() inside migrate_vma_insert_page()
2823 * did already call it.
2826 mmu_notifier_invalidate_range_only_end(&range);
2828 EXPORT_SYMBOL(migrate_vma_pages);
2831 * migrate_vma_finalize() - restore CPU page table entry
2832 * @migrate: migrate struct containing all migration information
2834 * This replaces the special migration pte entry with either a mapping to the
2835 * new page if migration was successful for that page, or to the original page
2838 * This also unlocks the pages and puts them back on the lru, or drops the extra
2839 * refcount, for device pages.
2841 void migrate_vma_finalize(struct migrate_vma *migrate)
2843 const unsigned long npages = migrate->npages;
2846 for (i = 0; i < npages; i++) {
2847 struct page *newpage = migrate_pfn_to_page(migrate->dst[i]);
2848 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2852 unlock_page(newpage);
2858 if (!(migrate->src[i] & MIGRATE_PFN_MIGRATE) || !newpage) {
2860 unlock_page(newpage);
2866 remove_migration_ptes(page, newpage, false);
2869 if (is_zone_device_page(page))
2872 putback_lru_page(page);
2874 if (newpage != page) {
2875 unlock_page(newpage);
2876 if (is_zone_device_page(newpage))
2879 putback_lru_page(newpage);
2883 EXPORT_SYMBOL(migrate_vma_finalize);
2884 #endif /* CONFIG_DEVICE_PRIVATE */
2887 * node_demotion[] example:
2889 * Consider a system with two sockets. Each socket has
2890 * three classes of memory attached: fast, medium and slow.
2891 * Each memory class is placed in its own NUMA node. The
2892 * CPUs are placed in the node with the "fast" memory. The
2893 * 6 NUMA nodes (0-5) might be split among the sockets like
2899 * When Node 0 fills up, its memory should be migrated to
2900 * Node 1. When Node 1 fills up, it should be migrated to
2901 * Node 2. The migration path start on the nodes with the
2902 * processors (since allocations default to this node) and
2903 * fast memory, progress through medium and end with the
2906 * 0 -> 1 -> 2 -> stop
2907 * 3 -> 4 -> 5 -> stop
2909 * This is represented in the node_demotion[] like this:
2911 * { nr=1, nodes[0]=1 }, // Node 0 migrates to 1
2912 * { nr=1, nodes[0]=2 }, // Node 1 migrates to 2
2913 * { nr=0, nodes[0]=-1 }, // Node 2 does not migrate
2914 * { nr=1, nodes[0]=4 }, // Node 3 migrates to 4
2915 * { nr=1, nodes[0]=5 }, // Node 4 migrates to 5
2916 * { nr=0, nodes[0]=-1 }, // Node 5 does not migrate
2918 * Moreover some systems may have multiple slow memory nodes.
2919 * Suppose a system has one socket with 3 memory nodes, node 0
2920 * is fast memory type, and node 1/2 both are slow memory
2921 * type, and the distance between fast memory node and slow
2922 * memory node is same. So the migration path should be:
2926 * This is represented in the node_demotion[] like this:
2927 * { nr=2, {nodes[0]=1, nodes[1]=2} }, // Node 0 migrates to node 1 and node 2
2928 * { nr=0, nodes[0]=-1, }, // Node 1 dose not migrate
2929 * { nr=0, nodes[0]=-1, }, // Node 2 does not migrate
2933 * Writes to this array occur without locking. Cycles are
2934 * not allowed: Node X demotes to Y which demotes to X...
2936 * If multiple reads are performed, a single rcu_read_lock()
2937 * must be held over all reads to ensure that no cycles are
2940 #define DEFAULT_DEMOTION_TARGET_NODES 15
2942 #if MAX_NUMNODES < DEFAULT_DEMOTION_TARGET_NODES
2943 #define DEMOTION_TARGET_NODES (MAX_NUMNODES - 1)
2945 #define DEMOTION_TARGET_NODES DEFAULT_DEMOTION_TARGET_NODES
2948 struct demotion_nodes {
2950 short nodes[DEMOTION_TARGET_NODES];
2953 static struct demotion_nodes *node_demotion __read_mostly;
2956 * next_demotion_node() - Get the next node in the demotion path
2957 * @node: The starting node to lookup the next node
2959 * Return: node id for next memory node in the demotion path hierarchy
2960 * from @node; NUMA_NO_NODE if @node is terminal. This does not keep
2961 * @node online or guarantee that it *continues* to be the next demotion
2964 int next_demotion_node(int node)
2966 struct demotion_nodes *nd;
2967 unsigned short target_nr, index;
2971 return NUMA_NO_NODE;
2973 nd = &node_demotion[node];
2976 * node_demotion[] is updated without excluding this
2977 * function from running. RCU doesn't provide any
2978 * compiler barriers, so the READ_ONCE() is required
2979 * to avoid compiler reordering or read merging.
2981 * Make sure to use RCU over entire code blocks if
2982 * node_demotion[] reads need to be consistent.
2985 target_nr = READ_ONCE(nd->nr);
2987 switch (target_nr) {
2989 target = NUMA_NO_NODE;
2996 * If there are multiple target nodes, just select one
2997 * target node randomly.
2999 * In addition, we can also use round-robin to select
3000 * target node, but we should introduce another variable
3001 * for node_demotion[] to record last selected target node,
3002 * that may cause cache ping-pong due to the changing of
3003 * last target node. Or introducing per-cpu data to avoid
3004 * caching issue, which seems more complicated. So selecting
3005 * target node randomly seems better until now.
3007 index = get_random_int() % target_nr;
3011 target = READ_ONCE(nd->nodes[index]);
3018 #if defined(CONFIG_HOTPLUG_CPU)
3019 /* Disable reclaim-based migration. */
3020 static void __disable_all_migrate_targets(void)
3027 for_each_online_node(node) {
3028 node_demotion[node].nr = 0;
3029 for (i = 0; i < DEMOTION_TARGET_NODES; i++)
3030 node_demotion[node].nodes[i] = NUMA_NO_NODE;
3034 static void disable_all_migrate_targets(void)
3036 __disable_all_migrate_targets();
3039 * Ensure that the "disable" is visible across the system.
3040 * Readers will see either a combination of before+disable
3041 * state or disable+after. They will never see before and
3042 * after state together.
3044 * The before+after state together might have cycles and
3045 * could cause readers to do things like loop until this
3046 * function finishes. This ensures they can only see a
3047 * single "bad" read and would, for instance, only loop
3054 * Find an automatic demotion target for 'node'.
3055 * Failing here is OK. It might just indicate
3056 * being at the end of a chain.
3058 static int establish_migrate_target(int node, nodemask_t *used,
3061 int migration_target, index, val;
3062 struct demotion_nodes *nd;
3065 return NUMA_NO_NODE;
3067 nd = &node_demotion[node];
3069 migration_target = find_next_best_node(node, used);
3070 if (migration_target == NUMA_NO_NODE)
3071 return NUMA_NO_NODE;
3074 * If the node has been set a migration target node before,
3075 * which means it's the best distance between them. Still
3076 * check if this node can be demoted to other target nodes
3077 * if they have a same best distance.
3079 if (best_distance != -1) {
3080 val = node_distance(node, migration_target);
3081 if (val > best_distance)
3086 if (WARN_ONCE(index >= DEMOTION_TARGET_NODES,
3087 "Exceeds maximum demotion target nodes\n"))
3090 nd->nodes[index] = migration_target;
3093 return migration_target;
3095 node_clear(migration_target, *used);
3096 return NUMA_NO_NODE;
3100 * When memory fills up on a node, memory contents can be
3101 * automatically migrated to another node instead of
3102 * discarded at reclaim.
3104 * Establish a "migration path" which will start at nodes
3105 * with CPUs and will follow the priorities used to build the
3106 * page allocator zonelists.
3108 * The difference here is that cycles must be avoided. If
3109 * node0 migrates to node1, then neither node1, nor anything
3110 * node1 migrates to can migrate to node0. Also one node can
3111 * be migrated to multiple nodes if the target nodes all have
3112 * a same best-distance against the source node.
3114 * This function can run simultaneously with readers of
3115 * node_demotion[]. However, it can not run simultaneously
3116 * with itself. Exclusion is provided by memory hotplug events
3117 * being single-threaded.
3119 static void __set_migration_target_nodes(void)
3121 nodemask_t next_pass = NODE_MASK_NONE;
3122 nodemask_t this_pass = NODE_MASK_NONE;
3123 nodemask_t used_targets = NODE_MASK_NONE;
3124 int node, best_distance;
3127 * Avoid any oddities like cycles that could occur
3128 * from changes in the topology. This will leave
3129 * a momentary gap when migration is disabled.
3131 disable_all_migrate_targets();
3134 * Allocations go close to CPUs, first. Assume that
3135 * the migration path starts at the nodes with CPUs.
3137 next_pass = node_states[N_CPU];
3139 this_pass = next_pass;
3140 next_pass = NODE_MASK_NONE;
3142 * To avoid cycles in the migration "graph", ensure
3143 * that migration sources are not future targets by
3144 * setting them in 'used_targets'. Do this only
3145 * once per pass so that multiple source nodes can
3146 * share a target node.
3148 * 'used_targets' will become unavailable in future
3149 * passes. This limits some opportunities for
3150 * multiple source nodes to share a destination.
3152 nodes_or(used_targets, used_targets, this_pass);
3154 for_each_node_mask(node, this_pass) {
3158 * Try to set up the migration path for the node, and the target
3159 * migration nodes can be multiple, so doing a loop to find all
3160 * the target nodes if they all have a best node distance.
3164 establish_migrate_target(node, &used_targets,
3167 if (target_node == NUMA_NO_NODE)
3170 if (best_distance == -1)
3171 best_distance = node_distance(node, target_node);
3174 * Visit targets from this pass in the next pass.
3175 * Eventually, every node will have been part of
3176 * a pass, and will become set in 'used_targets'.
3178 node_set(target_node, next_pass);
3182 * 'next_pass' contains nodes which became migration
3183 * targets in this pass. Make additional passes until
3184 * no more migrations targets are available.
3186 if (!nodes_empty(next_pass))
3191 * For callers that do not hold get_online_mems() already.
3193 static void set_migration_target_nodes(void)
3196 __set_migration_target_nodes();
3201 * This leaves migrate-on-reclaim transiently disabled between
3202 * the MEM_GOING_OFFLINE and MEM_OFFLINE events. This runs
3203 * whether reclaim-based migration is enabled or not, which
3204 * ensures that the user can turn reclaim-based migration at
3205 * any time without needing to recalculate migration targets.
3207 * These callbacks already hold get_online_mems(). That is why
3208 * __set_migration_target_nodes() can be used as opposed to
3209 * set_migration_target_nodes().
3211 static int __meminit migrate_on_reclaim_callback(struct notifier_block *self,
3212 unsigned long action, void *_arg)
3214 struct memory_notify *arg = _arg;
3217 * Only update the node migration order when a node is
3218 * changing status, like online->offline. This avoids
3219 * the overhead of synchronize_rcu() in most cases.
3221 if (arg->status_change_nid < 0)
3222 return notifier_from_errno(0);
3225 case MEM_GOING_OFFLINE:
3227 * Make sure there are not transient states where
3228 * an offline node is a migration target. This
3229 * will leave migration disabled until the offline
3230 * completes and the MEM_OFFLINE case below runs.
3232 disable_all_migrate_targets();
3237 * Recalculate the target nodes once the node
3238 * reaches its final state (online or offline).
3240 __set_migration_target_nodes();
3242 case MEM_CANCEL_OFFLINE:
3244 * MEM_GOING_OFFLINE disabled all the migration
3245 * targets. Reenable them.
3247 __set_migration_target_nodes();
3249 case MEM_GOING_ONLINE:
3250 case MEM_CANCEL_ONLINE:
3254 return notifier_from_errno(0);
3258 * React to hotplug events that might affect the migration targets
3259 * like events that online or offline NUMA nodes.
3261 * The ordering is also currently dependent on which nodes have
3262 * CPUs. That means we need CPU on/offline notification too.
3264 static int migration_online_cpu(unsigned int cpu)
3266 set_migration_target_nodes();
3270 static int migration_offline_cpu(unsigned int cpu)
3272 set_migration_target_nodes();
3276 static int __init migrate_on_reclaim_init(void)
3280 node_demotion = kmalloc_array(nr_node_ids,
3281 sizeof(struct demotion_nodes),
3283 WARN_ON(!node_demotion);
3285 ret = cpuhp_setup_state_nocalls(CPUHP_MM_DEMOTION_DEAD, "mm/demotion:offline",
3286 NULL, migration_offline_cpu);
3288 * In the unlikely case that this fails, the automatic
3289 * migration targets may become suboptimal for nodes
3290 * where N_CPU changes. With such a small impact in a
3291 * rare case, do not bother trying to do anything special.
3294 ret = cpuhp_setup_state(CPUHP_AP_MM_DEMOTION_ONLINE, "mm/demotion:online",
3295 migration_online_cpu, NULL);
3298 hotplug_memory_notifier(migrate_on_reclaim_callback, 100);
3301 late_initcall(migrate_on_reclaim_init);
3302 #endif /* CONFIG_HOTPLUG_CPU */
3304 bool numa_demotion_enabled = false;
3307 static ssize_t numa_demotion_enabled_show(struct kobject *kobj,
3308 struct kobj_attribute *attr, char *buf)
3310 return sysfs_emit(buf, "%s\n",
3311 numa_demotion_enabled ? "true" : "false");
3314 static ssize_t numa_demotion_enabled_store(struct kobject *kobj,
3315 struct kobj_attribute *attr,
3316 const char *buf, size_t count)
3318 if (!strncmp(buf, "true", 4) || !strncmp(buf, "1", 1))
3319 numa_demotion_enabled = true;
3320 else if (!strncmp(buf, "false", 5) || !strncmp(buf, "0", 1))
3321 numa_demotion_enabled = false;
3328 static struct kobj_attribute numa_demotion_enabled_attr =
3329 __ATTR(demotion_enabled, 0644, numa_demotion_enabled_show,
3330 numa_demotion_enabled_store);
3332 static struct attribute *numa_attrs[] = {
3333 &numa_demotion_enabled_attr.attr,
3337 static const struct attribute_group numa_attr_group = {
3338 .attrs = numa_attrs,
3341 static int __init numa_init_sysfs(void)
3344 struct kobject *numa_kobj;
3346 numa_kobj = kobject_create_and_add("numa", mm_kobj);
3348 pr_err("failed to create numa kobject\n");
3351 err = sysfs_create_group(numa_kobj, &numa_attr_group);
3353 pr_err("failed to register numa group\n");
3359 kobject_put(numa_kobj);
3362 subsys_initcall(numa_init_sysfs);