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,
298 if (!is_swap_pte(pte))
301 entry = pte_to_swp_entry(pte);
302 if (!is_migration_entry(entry))
305 page = pfn_swap_entry_to_page(entry);
306 page = compound_head(page);
309 * Once page cache replacement of page migration started, page_count
310 * is zero; but we must not call put_and_wait_on_page_locked() without
311 * a ref. Use get_page_unless_zero(), and just fault again if it fails.
313 if (!get_page_unless_zero(page))
315 pte_unmap_unlock(ptep, ptl);
316 put_and_wait_on_page_locked(page, TASK_UNINTERRUPTIBLE);
319 pte_unmap_unlock(ptep, ptl);
322 void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
323 unsigned long address)
325 spinlock_t *ptl = pte_lockptr(mm, pmd);
326 pte_t *ptep = pte_offset_map(pmd, address);
327 __migration_entry_wait(mm, ptep, ptl);
330 void migration_entry_wait_huge(struct vm_area_struct *vma,
331 struct mm_struct *mm, pte_t *pte)
333 spinlock_t *ptl = huge_pte_lockptr(hstate_vma(vma), mm, pte);
334 __migration_entry_wait(mm, pte, ptl);
337 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
338 void pmd_migration_entry_wait(struct mm_struct *mm, pmd_t *pmd)
343 ptl = pmd_lock(mm, pmd);
344 if (!is_pmd_migration_entry(*pmd))
346 page = pfn_swap_entry_to_page(pmd_to_swp_entry(*pmd));
347 if (!get_page_unless_zero(page))
350 put_and_wait_on_page_locked(page, TASK_UNINTERRUPTIBLE);
357 static int expected_page_refs(struct address_space *mapping, struct page *page)
359 int expected_count = 1;
362 * Device private pages have an extra refcount as they are
365 expected_count += is_device_private_page(page);
367 expected_count += compound_nr(page) + page_has_private(page);
369 return expected_count;
373 * Replace the page in the mapping.
375 * The number of remaining references must be:
376 * 1 for anonymous pages without a mapping
377 * 2 for pages with a mapping
378 * 3 for pages with a mapping and PagePrivate/PagePrivate2 set.
380 int folio_migrate_mapping(struct address_space *mapping,
381 struct folio *newfolio, struct folio *folio, int extra_count)
383 XA_STATE(xas, &mapping->i_pages, folio_index(folio));
384 struct zone *oldzone, *newzone;
386 int expected_count = expected_page_refs(mapping, &folio->page) + extra_count;
387 long nr = folio_nr_pages(folio);
390 /* Anonymous page without mapping */
391 if (folio_ref_count(folio) != expected_count)
394 /* No turning back from here */
395 newfolio->index = folio->index;
396 newfolio->mapping = folio->mapping;
397 if (folio_test_swapbacked(folio))
398 __folio_set_swapbacked(newfolio);
400 return MIGRATEPAGE_SUCCESS;
403 oldzone = folio_zone(folio);
404 newzone = folio_zone(newfolio);
407 if (!folio_ref_freeze(folio, expected_count)) {
408 xas_unlock_irq(&xas);
413 * Now we know that no one else is looking at the folio:
414 * no turning back from here.
416 newfolio->index = folio->index;
417 newfolio->mapping = folio->mapping;
418 folio_ref_add(newfolio, nr); /* add cache reference */
419 if (folio_test_swapbacked(folio)) {
420 __folio_set_swapbacked(newfolio);
421 if (folio_test_swapcache(folio)) {
422 folio_set_swapcache(newfolio);
423 newfolio->private = folio_get_private(folio);
426 VM_BUG_ON_FOLIO(folio_test_swapcache(folio), folio);
429 /* Move dirty while page refs frozen and newpage not yet exposed */
430 dirty = folio_test_dirty(folio);
432 folio_clear_dirty(folio);
433 folio_set_dirty(newfolio);
436 xas_store(&xas, newfolio);
440 for (i = 1; i < nr; i++) {
442 xas_store(&xas, newfolio);
447 * Drop cache reference from old page by unfreezing
448 * to one less reference.
449 * We know this isn't the last reference.
451 folio_ref_unfreeze(folio, expected_count - nr);
454 /* Leave irq disabled to prevent preemption while updating stats */
457 * If moved to a different zone then also account
458 * the page for that zone. Other VM counters will be
459 * taken care of when we establish references to the
460 * new page and drop references to the old page.
462 * Note that anonymous pages are accounted for
463 * via NR_FILE_PAGES and NR_ANON_MAPPED if they
464 * are mapped to swap space.
466 if (newzone != oldzone) {
467 struct lruvec *old_lruvec, *new_lruvec;
468 struct mem_cgroup *memcg;
470 memcg = folio_memcg(folio);
471 old_lruvec = mem_cgroup_lruvec(memcg, oldzone->zone_pgdat);
472 new_lruvec = mem_cgroup_lruvec(memcg, newzone->zone_pgdat);
474 __mod_lruvec_state(old_lruvec, NR_FILE_PAGES, -nr);
475 __mod_lruvec_state(new_lruvec, NR_FILE_PAGES, nr);
476 if (folio_test_swapbacked(folio) && !folio_test_swapcache(folio)) {
477 __mod_lruvec_state(old_lruvec, NR_SHMEM, -nr);
478 __mod_lruvec_state(new_lruvec, NR_SHMEM, nr);
481 if (folio_test_swapcache(folio)) {
482 __mod_lruvec_state(old_lruvec, NR_SWAPCACHE, -nr);
483 __mod_lruvec_state(new_lruvec, NR_SWAPCACHE, nr);
486 if (dirty && mapping_can_writeback(mapping)) {
487 __mod_lruvec_state(old_lruvec, NR_FILE_DIRTY, -nr);
488 __mod_zone_page_state(oldzone, NR_ZONE_WRITE_PENDING, -nr);
489 __mod_lruvec_state(new_lruvec, NR_FILE_DIRTY, nr);
490 __mod_zone_page_state(newzone, NR_ZONE_WRITE_PENDING, nr);
495 return MIGRATEPAGE_SUCCESS;
497 EXPORT_SYMBOL(folio_migrate_mapping);
500 * The expected number of remaining references is the same as that
501 * of folio_migrate_mapping().
503 int migrate_huge_page_move_mapping(struct address_space *mapping,
504 struct page *newpage, struct page *page)
506 XA_STATE(xas, &mapping->i_pages, page_index(page));
510 expected_count = 2 + page_has_private(page);
511 if (page_count(page) != expected_count || xas_load(&xas) != page) {
512 xas_unlock_irq(&xas);
516 if (!page_ref_freeze(page, expected_count)) {
517 xas_unlock_irq(&xas);
521 newpage->index = page->index;
522 newpage->mapping = page->mapping;
526 xas_store(&xas, newpage);
528 page_ref_unfreeze(page, expected_count - 1);
530 xas_unlock_irq(&xas);
532 return MIGRATEPAGE_SUCCESS;
536 * Copy the flags and some other ancillary information
538 void folio_migrate_flags(struct folio *newfolio, struct folio *folio)
542 if (folio_test_error(folio))
543 folio_set_error(newfolio);
544 if (folio_test_referenced(folio))
545 folio_set_referenced(newfolio);
546 if (folio_test_uptodate(folio))
547 folio_mark_uptodate(newfolio);
548 if (folio_test_clear_active(folio)) {
549 VM_BUG_ON_FOLIO(folio_test_unevictable(folio), folio);
550 folio_set_active(newfolio);
551 } else if (folio_test_clear_unevictable(folio))
552 folio_set_unevictable(newfolio);
553 if (folio_test_workingset(folio))
554 folio_set_workingset(newfolio);
555 if (folio_test_checked(folio))
556 folio_set_checked(newfolio);
557 if (folio_test_mappedtodisk(folio))
558 folio_set_mappedtodisk(newfolio);
560 /* Move dirty on pages not done by folio_migrate_mapping() */
561 if (folio_test_dirty(folio))
562 folio_set_dirty(newfolio);
564 if (folio_test_young(folio))
565 folio_set_young(newfolio);
566 if (folio_test_idle(folio))
567 folio_set_idle(newfolio);
570 * Copy NUMA information to the new page, to prevent over-eager
571 * future migrations of this same page.
573 cpupid = page_cpupid_xchg_last(&folio->page, -1);
574 page_cpupid_xchg_last(&newfolio->page, cpupid);
576 folio_migrate_ksm(newfolio, folio);
578 * Please do not reorder this without considering how mm/ksm.c's
579 * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache().
581 if (folio_test_swapcache(folio))
582 folio_clear_swapcache(folio);
583 folio_clear_private(folio);
585 /* page->private contains hugetlb specific flags */
586 if (!folio_test_hugetlb(folio))
587 folio->private = NULL;
590 * If any waiters have accumulated on the new page then
593 if (folio_test_writeback(newfolio))
594 folio_end_writeback(newfolio);
597 * PG_readahead shares the same bit with PG_reclaim. The above
598 * end_page_writeback() may clear PG_readahead mistakenly, so set the
601 if (folio_test_readahead(folio))
602 folio_set_readahead(newfolio);
604 folio_copy_owner(newfolio, folio);
606 if (!folio_test_hugetlb(folio))
607 mem_cgroup_migrate(folio, newfolio);
609 EXPORT_SYMBOL(folio_migrate_flags);
611 void folio_migrate_copy(struct folio *newfolio, struct folio *folio)
613 folio_copy(newfolio, folio);
614 folio_migrate_flags(newfolio, folio);
616 EXPORT_SYMBOL(folio_migrate_copy);
618 /************************************************************
619 * Migration functions
620 ***********************************************************/
623 * Common logic to directly migrate a single LRU page suitable for
624 * pages that do not use PagePrivate/PagePrivate2.
626 * Pages are locked upon entry and exit.
628 int migrate_page(struct address_space *mapping,
629 struct page *newpage, struct page *page,
630 enum migrate_mode mode)
632 struct folio *newfolio = page_folio(newpage);
633 struct folio *folio = page_folio(page);
636 BUG_ON(folio_test_writeback(folio)); /* Writeback must be complete */
638 rc = folio_migrate_mapping(mapping, newfolio, folio, 0);
640 if (rc != MIGRATEPAGE_SUCCESS)
643 if (mode != MIGRATE_SYNC_NO_COPY)
644 folio_migrate_copy(newfolio, folio);
646 folio_migrate_flags(newfolio, folio);
647 return MIGRATEPAGE_SUCCESS;
649 EXPORT_SYMBOL(migrate_page);
652 /* Returns true if all buffers are successfully locked */
653 static bool buffer_migrate_lock_buffers(struct buffer_head *head,
654 enum migrate_mode mode)
656 struct buffer_head *bh = head;
658 /* Simple case, sync compaction */
659 if (mode != MIGRATE_ASYNC) {
662 bh = bh->b_this_page;
664 } while (bh != head);
669 /* async case, we cannot block on lock_buffer so use trylock_buffer */
671 if (!trylock_buffer(bh)) {
673 * We failed to lock the buffer and cannot stall in
674 * async migration. Release the taken locks
676 struct buffer_head *failed_bh = bh;
678 while (bh != failed_bh) {
680 bh = bh->b_this_page;
685 bh = bh->b_this_page;
686 } while (bh != head);
690 static int __buffer_migrate_page(struct address_space *mapping,
691 struct page *newpage, struct page *page, enum migrate_mode mode,
694 struct buffer_head *bh, *head;
698 if (!page_has_buffers(page))
699 return migrate_page(mapping, newpage, page, mode);
701 /* Check whether page does not have extra refs before we do more work */
702 expected_count = expected_page_refs(mapping, page);
703 if (page_count(page) != expected_count)
706 head = page_buffers(page);
707 if (!buffer_migrate_lock_buffers(head, mode))
712 bool invalidated = false;
716 spin_lock(&mapping->private_lock);
719 if (atomic_read(&bh->b_count)) {
723 bh = bh->b_this_page;
724 } while (bh != head);
730 spin_unlock(&mapping->private_lock);
731 invalidate_bh_lrus();
733 goto recheck_buffers;
737 rc = migrate_page_move_mapping(mapping, newpage, page, 0);
738 if (rc != MIGRATEPAGE_SUCCESS)
741 attach_page_private(newpage, detach_page_private(page));
745 set_bh_page(bh, newpage, bh_offset(bh));
746 bh = bh->b_this_page;
748 } while (bh != head);
750 if (mode != MIGRATE_SYNC_NO_COPY)
751 migrate_page_copy(newpage, page);
753 migrate_page_states(newpage, page);
755 rc = MIGRATEPAGE_SUCCESS;
758 spin_unlock(&mapping->private_lock);
762 bh = bh->b_this_page;
764 } while (bh != head);
770 * Migration function for pages with buffers. This function can only be used
771 * if the underlying filesystem guarantees that no other references to "page"
772 * exist. For example attached buffer heads are accessed only under page lock.
774 int buffer_migrate_page(struct address_space *mapping,
775 struct page *newpage, struct page *page, enum migrate_mode mode)
777 return __buffer_migrate_page(mapping, newpage, page, mode, false);
779 EXPORT_SYMBOL(buffer_migrate_page);
782 * Same as above except that this variant is more careful and checks that there
783 * are also no buffer head references. This function is the right one for
784 * mappings where buffer heads are directly looked up and referenced (such as
785 * block device mappings).
787 int buffer_migrate_page_norefs(struct address_space *mapping,
788 struct page *newpage, struct page *page, enum migrate_mode mode)
790 return __buffer_migrate_page(mapping, newpage, page, mode, true);
795 * Writeback a page to clean the dirty state
797 static int writeout(struct address_space *mapping, struct page *page)
799 struct writeback_control wbc = {
800 .sync_mode = WB_SYNC_NONE,
803 .range_end = LLONG_MAX,
808 if (!mapping->a_ops->writepage)
809 /* No write method for the address space */
812 if (!clear_page_dirty_for_io(page))
813 /* Someone else already triggered a write */
817 * A dirty page may imply that the underlying filesystem has
818 * the page on some queue. So the page must be clean for
819 * migration. Writeout may mean we loose the lock and the
820 * page state is no longer what we checked for earlier.
821 * At this point we know that the migration attempt cannot
824 remove_migration_ptes(page, page, false);
826 rc = mapping->a_ops->writepage(page, &wbc);
828 if (rc != AOP_WRITEPAGE_ACTIVATE)
829 /* unlocked. Relock */
832 return (rc < 0) ? -EIO : -EAGAIN;
836 * Default handling if a filesystem does not provide a migration function.
838 static int fallback_migrate_page(struct address_space *mapping,
839 struct page *newpage, struct page *page, enum migrate_mode mode)
841 if (PageDirty(page)) {
842 /* Only writeback pages in full synchronous migration */
845 case MIGRATE_SYNC_NO_COPY:
850 return writeout(mapping, page);
854 * Buffers may be managed in a filesystem specific way.
855 * We must have no buffers or drop them.
857 if (page_has_private(page) &&
858 !try_to_release_page(page, GFP_KERNEL))
859 return mode == MIGRATE_SYNC ? -EAGAIN : -EBUSY;
861 return migrate_page(mapping, newpage, page, mode);
865 * Move a page to a newly allocated page
866 * The page is locked and all ptes have been successfully removed.
868 * The new page will have replaced the old page if this function
873 * MIGRATEPAGE_SUCCESS - success
875 static int move_to_new_page(struct page *newpage, struct page *page,
876 enum migrate_mode mode)
878 struct address_space *mapping;
880 bool is_lru = !__PageMovable(page);
882 VM_BUG_ON_PAGE(!PageLocked(page), page);
883 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
885 mapping = page_mapping(page);
887 if (likely(is_lru)) {
889 rc = migrate_page(mapping, newpage, page, mode);
890 else if (mapping->a_ops->migratepage)
892 * Most pages have a mapping and most filesystems
893 * provide a migratepage callback. Anonymous pages
894 * are part of swap space which also has its own
895 * migratepage callback. This is the most common path
896 * for page migration.
898 rc = mapping->a_ops->migratepage(mapping, newpage,
901 rc = fallback_migrate_page(mapping, newpage,
905 * In case of non-lru page, it could be released after
906 * isolation step. In that case, we shouldn't try migration.
908 VM_BUG_ON_PAGE(!PageIsolated(page), page);
909 if (!PageMovable(page)) {
910 rc = MIGRATEPAGE_SUCCESS;
911 __ClearPageIsolated(page);
915 rc = mapping->a_ops->migratepage(mapping, newpage,
917 WARN_ON_ONCE(rc == MIGRATEPAGE_SUCCESS &&
918 !PageIsolated(page));
922 * When successful, old pagecache page->mapping must be cleared before
923 * page is freed; but stats require that PageAnon be left as PageAnon.
925 if (rc == MIGRATEPAGE_SUCCESS) {
926 if (__PageMovable(page)) {
927 VM_BUG_ON_PAGE(!PageIsolated(page), page);
930 * We clear PG_movable under page_lock so any compactor
931 * cannot try to migrate this page.
933 __ClearPageIsolated(page);
937 * Anonymous and movable page->mapping will be cleared by
938 * free_pages_prepare so don't reset it here for keeping
939 * the type to work PageAnon, for example.
941 if (!PageMappingFlags(page))
942 page->mapping = NULL;
944 if (likely(!is_zone_device_page(newpage)))
945 flush_dcache_page(newpage);
952 static int __unmap_and_move(struct page *page, struct page *newpage,
953 int force, enum migrate_mode mode)
956 bool page_was_mapped = false;
957 struct anon_vma *anon_vma = NULL;
958 bool is_lru = !__PageMovable(page);
960 if (!trylock_page(page)) {
961 if (!force || mode == MIGRATE_ASYNC)
965 * It's not safe for direct compaction to call lock_page.
966 * For example, during page readahead pages are added locked
967 * to the LRU. Later, when the IO completes the pages are
968 * marked uptodate and unlocked. However, the queueing
969 * could be merging multiple pages for one bio (e.g.
970 * mpage_readahead). If an allocation happens for the
971 * second or third page, the process can end up locking
972 * the same page twice and deadlocking. Rather than
973 * trying to be clever about what pages can be locked,
974 * avoid the use of lock_page for direct compaction
977 if (current->flags & PF_MEMALLOC)
983 if (PageWriteback(page)) {
985 * Only in the case of a full synchronous migration is it
986 * necessary to wait for PageWriteback. In the async case,
987 * the retry loop is too short and in the sync-light case,
988 * the overhead of stalling is too much
992 case MIGRATE_SYNC_NO_COPY:
1000 wait_on_page_writeback(page);
1004 * By try_to_migrate(), page->mapcount goes down to 0 here. In this case,
1005 * we cannot notice that anon_vma is freed while we migrates a page.
1006 * This get_anon_vma() delays freeing anon_vma pointer until the end
1007 * of migration. File cache pages are no problem because of page_lock()
1008 * File Caches may use write_page() or lock_page() in migration, then,
1009 * just care Anon page here.
1011 * Only page_get_anon_vma() understands the subtleties of
1012 * getting a hold on an anon_vma from outside one of its mms.
1013 * But if we cannot get anon_vma, then we won't need it anyway,
1014 * because that implies that the anon page is no longer mapped
1015 * (and cannot be remapped so long as we hold the page lock).
1017 if (PageAnon(page) && !PageKsm(page))
1018 anon_vma = page_get_anon_vma(page);
1021 * Block others from accessing the new page when we get around to
1022 * establishing additional references. We are usually the only one
1023 * holding a reference to newpage at this point. We used to have a BUG
1024 * here if trylock_page(newpage) fails, but would like to allow for
1025 * cases where there might be a race with the previous use of newpage.
1026 * This is much like races on refcount of oldpage: just don't BUG().
1028 if (unlikely(!trylock_page(newpage)))
1031 if (unlikely(!is_lru)) {
1032 rc = move_to_new_page(newpage, page, mode);
1033 goto out_unlock_both;
1037 * Corner case handling:
1038 * 1. When a new swap-cache page is read into, it is added to the LRU
1039 * and treated as swapcache but it has no rmap yet.
1040 * Calling try_to_unmap() against a page->mapping==NULL page will
1041 * trigger a BUG. So handle it here.
1042 * 2. An orphaned page (see truncate_cleanup_page) might have
1043 * fs-private metadata. The page can be picked up due to memory
1044 * offlining. Everywhere else except page reclaim, the page is
1045 * invisible to the vm, so the page can not be migrated. So try to
1046 * free the metadata, so the page can be freed.
1048 if (!page->mapping) {
1049 VM_BUG_ON_PAGE(PageAnon(page), page);
1050 if (page_has_private(page)) {
1051 try_to_free_buffers(page);
1052 goto out_unlock_both;
1054 } else if (page_mapped(page)) {
1055 /* Establish migration ptes */
1056 VM_BUG_ON_PAGE(PageAnon(page) && !PageKsm(page) && !anon_vma,
1058 try_to_migrate(page, 0);
1059 page_was_mapped = true;
1062 if (!page_mapped(page))
1063 rc = move_to_new_page(newpage, page, mode);
1065 if (page_was_mapped)
1066 remove_migration_ptes(page,
1067 rc == MIGRATEPAGE_SUCCESS ? newpage : page, false);
1070 unlock_page(newpage);
1072 /* Drop an anon_vma reference if we took one */
1074 put_anon_vma(anon_vma);
1078 * If migration is successful, decrease refcount of the newpage
1079 * which will not free the page because new page owner increased
1080 * refcounter. As well, if it is LRU page, add the page to LRU
1081 * list in here. Use the old state of the isolated source page to
1082 * determine if we migrated a LRU page. newpage was already unlocked
1083 * and possibly modified by its owner - don't rely on the page
1086 if (rc == MIGRATEPAGE_SUCCESS) {
1087 if (unlikely(!is_lru))
1090 putback_lru_page(newpage);
1098 * node_demotion[] example:
1100 * Consider a system with two sockets. Each socket has
1101 * three classes of memory attached: fast, medium and slow.
1102 * Each memory class is placed in its own NUMA node. The
1103 * CPUs are placed in the node with the "fast" memory. The
1104 * 6 NUMA nodes (0-5) might be split among the sockets like
1110 * When Node 0 fills up, its memory should be migrated to
1111 * Node 1. When Node 1 fills up, it should be migrated to
1112 * Node 2. The migration path start on the nodes with the
1113 * processors (since allocations default to this node) and
1114 * fast memory, progress through medium and end with the
1117 * 0 -> 1 -> 2 -> stop
1118 * 3 -> 4 -> 5 -> stop
1120 * This is represented in the node_demotion[] like this:
1122 * { nr=1, nodes[0]=1 }, // Node 0 migrates to 1
1123 * { nr=1, nodes[0]=2 }, // Node 1 migrates to 2
1124 * { nr=0, nodes[0]=-1 }, // Node 2 does not migrate
1125 * { nr=1, nodes[0]=4 }, // Node 3 migrates to 4
1126 * { nr=1, nodes[0]=5 }, // Node 4 migrates to 5
1127 * { nr=0, nodes[0]=-1 }, // Node 5 does not migrate
1129 * Moreover some systems may have multiple slow memory nodes.
1130 * Suppose a system has one socket with 3 memory nodes, node 0
1131 * is fast memory type, and node 1/2 both are slow memory
1132 * type, and the distance between fast memory node and slow
1133 * memory node is same. So the migration path should be:
1137 * This is represented in the node_demotion[] like this:
1138 * { nr=2, {nodes[0]=1, nodes[1]=2} }, // Node 0 migrates to node 1 and node 2
1139 * { nr=0, nodes[0]=-1, }, // Node 1 dose not migrate
1140 * { nr=0, nodes[0]=-1, }, // Node 2 does not migrate
1144 * Writes to this array occur without locking. Cycles are
1145 * not allowed: Node X demotes to Y which demotes to X...
1147 * If multiple reads are performed, a single rcu_read_lock()
1148 * must be held over all reads to ensure that no cycles are
1151 #define DEFAULT_DEMOTION_TARGET_NODES 15
1153 #if MAX_NUMNODES < DEFAULT_DEMOTION_TARGET_NODES
1154 #define DEMOTION_TARGET_NODES (MAX_NUMNODES - 1)
1156 #define DEMOTION_TARGET_NODES DEFAULT_DEMOTION_TARGET_NODES
1159 struct demotion_nodes {
1161 short nodes[DEMOTION_TARGET_NODES];
1164 static struct demotion_nodes *node_demotion __read_mostly;
1167 * next_demotion_node() - Get the next node in the demotion path
1168 * @node: The starting node to lookup the next node
1170 * Return: node id for next memory node in the demotion path hierarchy
1171 * from @node; NUMA_NO_NODE if @node is terminal. This does not keep
1172 * @node online or guarantee that it *continues* to be the next demotion
1175 int next_demotion_node(int node)
1177 struct demotion_nodes *nd;
1178 unsigned short target_nr, index;
1182 return NUMA_NO_NODE;
1184 nd = &node_demotion[node];
1187 * node_demotion[] is updated without excluding this
1188 * function from running. RCU doesn't provide any
1189 * compiler barriers, so the READ_ONCE() is required
1190 * to avoid compiler reordering or read merging.
1192 * Make sure to use RCU over entire code blocks if
1193 * node_demotion[] reads need to be consistent.
1196 target_nr = READ_ONCE(nd->nr);
1198 switch (target_nr) {
1200 target = NUMA_NO_NODE;
1207 * If there are multiple target nodes, just select one
1208 * target node randomly.
1210 * In addition, we can also use round-robin to select
1211 * target node, but we should introduce another variable
1212 * for node_demotion[] to record last selected target node,
1213 * that may cause cache ping-pong due to the changing of
1214 * last target node. Or introducing per-cpu data to avoid
1215 * caching issue, which seems more complicated. So selecting
1216 * target node randomly seems better until now.
1218 index = get_random_int() % target_nr;
1222 target = READ_ONCE(nd->nodes[index]);
1230 * Obtain the lock on page, remove all ptes and migrate the page
1231 * to the newly allocated page in newpage.
1233 static int unmap_and_move(new_page_t get_new_page,
1234 free_page_t put_new_page,
1235 unsigned long private, struct page *page,
1236 int force, enum migrate_mode mode,
1237 enum migrate_reason reason,
1238 struct list_head *ret)
1240 int rc = MIGRATEPAGE_SUCCESS;
1241 struct page *newpage = NULL;
1243 if (!thp_migration_supported() && PageTransHuge(page))
1246 if (page_count(page) == 1) {
1247 /* page was freed from under us. So we are done. */
1248 ClearPageActive(page);
1249 ClearPageUnevictable(page);
1250 if (unlikely(__PageMovable(page))) {
1252 if (!PageMovable(page))
1253 __ClearPageIsolated(page);
1259 newpage = get_new_page(page, private);
1263 rc = __unmap_and_move(page, newpage, force, mode);
1264 if (rc == MIGRATEPAGE_SUCCESS)
1265 set_page_owner_migrate_reason(newpage, reason);
1268 if (rc != -EAGAIN) {
1270 * A page that has been migrated has all references
1271 * removed and will be freed. A page that has not been
1272 * migrated will have kept its references and be restored.
1274 list_del(&page->lru);
1278 * If migration is successful, releases reference grabbed during
1279 * isolation. Otherwise, restore the page to right list unless
1282 if (rc == MIGRATEPAGE_SUCCESS) {
1284 * Compaction can migrate also non-LRU pages which are
1285 * not accounted to NR_ISOLATED_*. They can be recognized
1288 if (likely(!__PageMovable(page)))
1289 mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON +
1290 page_is_file_lru(page), -thp_nr_pages(page));
1292 if (reason != MR_MEMORY_FAILURE)
1294 * We release the page in page_handle_poison.
1299 list_add_tail(&page->lru, ret);
1302 put_new_page(newpage, private);
1311 * Counterpart of unmap_and_move_page() for hugepage migration.
1313 * This function doesn't wait the completion of hugepage I/O
1314 * because there is no race between I/O and migration for hugepage.
1315 * Note that currently hugepage I/O occurs only in direct I/O
1316 * where no lock is held and PG_writeback is irrelevant,
1317 * and writeback status of all subpages are counted in the reference
1318 * count of the head page (i.e. if all subpages of a 2MB hugepage are
1319 * under direct I/O, the reference of the head page is 512 and a bit more.)
1320 * This means that when we try to migrate hugepage whose subpages are
1321 * doing direct I/O, some references remain after try_to_unmap() and
1322 * hugepage migration fails without data corruption.
1324 * There is also no race when direct I/O is issued on the page under migration,
1325 * because then pte is replaced with migration swap entry and direct I/O code
1326 * will wait in the page fault for migration to complete.
1328 static int unmap_and_move_huge_page(new_page_t get_new_page,
1329 free_page_t put_new_page, unsigned long private,
1330 struct page *hpage, int force,
1331 enum migrate_mode mode, int reason,
1332 struct list_head *ret)
1335 int page_was_mapped = 0;
1336 struct page *new_hpage;
1337 struct anon_vma *anon_vma = NULL;
1338 struct address_space *mapping = NULL;
1341 * Migratability of hugepages depends on architectures and their size.
1342 * This check is necessary because some callers of hugepage migration
1343 * like soft offline and memory hotremove don't walk through page
1344 * tables or check whether the hugepage is pmd-based or not before
1345 * kicking migration.
1347 if (!hugepage_migration_supported(page_hstate(hpage))) {
1348 list_move_tail(&hpage->lru, ret);
1352 if (page_count(hpage) == 1) {
1353 /* page was freed from under us. So we are done. */
1354 putback_active_hugepage(hpage);
1355 return MIGRATEPAGE_SUCCESS;
1358 new_hpage = get_new_page(hpage, private);
1362 if (!trylock_page(hpage)) {
1367 case MIGRATE_SYNC_NO_COPY:
1376 * Check for pages which are in the process of being freed. Without
1377 * page_mapping() set, hugetlbfs specific move page routine will not
1378 * be called and we could leak usage counts for subpools.
1380 if (hugetlb_page_subpool(hpage) && !page_mapping(hpage)) {
1385 if (PageAnon(hpage))
1386 anon_vma = page_get_anon_vma(hpage);
1388 if (unlikely(!trylock_page(new_hpage)))
1391 if (page_mapped(hpage)) {
1392 bool mapping_locked = false;
1393 enum ttu_flags ttu = 0;
1395 if (!PageAnon(hpage)) {
1397 * In shared mappings, try_to_unmap could potentially
1398 * call huge_pmd_unshare. Because of this, take
1399 * semaphore in write mode here and set TTU_RMAP_LOCKED
1400 * to let lower levels know we have taken the lock.
1402 mapping = hugetlb_page_mapping_lock_write(hpage);
1403 if (unlikely(!mapping))
1404 goto unlock_put_anon;
1406 mapping_locked = true;
1407 ttu |= TTU_RMAP_LOCKED;
1410 try_to_migrate(hpage, ttu);
1411 page_was_mapped = 1;
1414 i_mmap_unlock_write(mapping);
1417 if (!page_mapped(hpage))
1418 rc = move_to_new_page(new_hpage, hpage, mode);
1420 if (page_was_mapped)
1421 remove_migration_ptes(hpage,
1422 rc == MIGRATEPAGE_SUCCESS ? new_hpage : hpage, false);
1425 unlock_page(new_hpage);
1429 put_anon_vma(anon_vma);
1431 if (rc == MIGRATEPAGE_SUCCESS) {
1432 move_hugetlb_state(hpage, new_hpage, reason);
1433 put_new_page = NULL;
1439 if (rc == MIGRATEPAGE_SUCCESS)
1440 putback_active_hugepage(hpage);
1441 else if (rc != -EAGAIN)
1442 list_move_tail(&hpage->lru, ret);
1445 * If migration was not successful and there's a freeing callback, use
1446 * it. Otherwise, put_page() will drop the reference grabbed during
1450 put_new_page(new_hpage, private);
1452 putback_active_hugepage(new_hpage);
1457 static inline int try_split_thp(struct page *page, struct page **page2,
1458 struct list_head *from)
1463 rc = split_huge_page_to_list(page, from);
1466 list_safe_reset_next(page, *page2, lru);
1472 * migrate_pages - migrate the pages specified in a list, to the free pages
1473 * supplied as the target for the page migration
1475 * @from: The list of pages to be migrated.
1476 * @get_new_page: The function used to allocate free pages to be used
1477 * as the target of the page migration.
1478 * @put_new_page: The function used to free target pages if migration
1479 * fails, or NULL if no special handling is necessary.
1480 * @private: Private data to be passed on to get_new_page()
1481 * @mode: The migration mode that specifies the constraints for
1482 * page migration, if any.
1483 * @reason: The reason for page migration.
1484 * @ret_succeeded: Set to the number of normal pages migrated successfully if
1485 * the caller passes a non-NULL pointer.
1487 * The function returns after 10 attempts or if no pages are movable any more
1488 * because the list has become empty or no retryable pages exist any more.
1489 * It is caller's responsibility to call putback_movable_pages() to return pages
1490 * to the LRU or free list only if ret != 0.
1492 * Returns the number of {normal page, THP, hugetlb} that were not migrated, or
1493 * an error code. The number of THP splits will be considered as the number of
1494 * non-migrated THP, no matter how many subpages of the THP are migrated successfully.
1496 int migrate_pages(struct list_head *from, new_page_t get_new_page,
1497 free_page_t put_new_page, unsigned long private,
1498 enum migrate_mode mode, int reason, unsigned int *ret_succeeded)
1503 int nr_failed_pages = 0;
1504 int nr_succeeded = 0;
1505 int nr_thp_succeeded = 0;
1506 int nr_thp_failed = 0;
1507 int nr_thp_split = 0;
1509 bool is_thp = false;
1512 int swapwrite = current->flags & PF_SWAPWRITE;
1513 int rc, nr_subpages;
1514 LIST_HEAD(ret_pages);
1515 LIST_HEAD(thp_split_pages);
1516 bool nosplit = (reason == MR_NUMA_MISPLACED);
1517 bool no_subpage_counting = false;
1519 trace_mm_migrate_pages_start(mode, reason);
1522 current->flags |= PF_SWAPWRITE;
1524 thp_subpage_migration:
1525 for (pass = 0; pass < 10 && (retry || thp_retry); pass++) {
1529 list_for_each_entry_safe(page, page2, from, lru) {
1532 * THP statistics is based on the source huge page.
1533 * Capture required information that might get lost
1536 is_thp = PageTransHuge(page) && !PageHuge(page);
1537 nr_subpages = compound_nr(page);
1541 rc = unmap_and_move_huge_page(get_new_page,
1542 put_new_page, private, page,
1543 pass > 2, mode, reason,
1546 rc = unmap_and_move(get_new_page, put_new_page,
1547 private, page, pass > 2, mode,
1548 reason, &ret_pages);
1551 * Success: non hugetlb page will be freed, hugetlb
1552 * page will be put back
1553 * -EAGAIN: stay on the from list
1554 * -ENOMEM: stay on the from list
1555 * Other errno: put on ret_pages list then splice to
1560 * THP migration might be unsupported or the
1561 * allocation could've failed so we should
1562 * retry on the same page with the THP split
1565 * Head page is retried immediately and tail
1566 * pages are added to the tail of the list so
1567 * we encounter them after the rest of the list
1571 /* THP migration is unsupported */
1574 if (!try_split_thp(page, &page2, &thp_split_pages)) {
1579 nr_failed_pages += nr_subpages;
1583 /* Hugetlb migration is unsupported */
1584 if (!no_subpage_counting)
1586 nr_failed_pages += nr_subpages;
1590 * When memory is low, don't bother to try to migrate
1591 * other pages, just exit.
1592 * THP NUMA faulting doesn't split THP to retry.
1594 if (is_thp && !nosplit) {
1596 if (!try_split_thp(page, &page2, &thp_split_pages)) {
1601 nr_failed_pages += nr_subpages;
1605 if (!no_subpage_counting)
1607 nr_failed_pages += nr_subpages;
1616 case MIGRATEPAGE_SUCCESS:
1617 nr_succeeded += nr_subpages;
1625 * Permanent failure (-EBUSY, etc.):
1626 * unlike -EAGAIN case, the failed page is
1627 * removed from migration page list and not
1628 * retried in the next outer loop.
1632 nr_failed_pages += nr_subpages;
1636 if (!no_subpage_counting)
1638 nr_failed_pages += nr_subpages;
1644 nr_thp_failed += thp_retry;
1646 * Try to migrate subpages of fail-to-migrate THPs, no nr_failed
1647 * counting in this round, since all subpages of a THP is counted
1648 * as 1 failure in the first round.
1650 if (!list_empty(&thp_split_pages)) {
1652 * Move non-migrated pages (after 10 retries) to ret_pages
1653 * to avoid migrating them again.
1655 list_splice_init(from, &ret_pages);
1656 list_splice_init(&thp_split_pages, from);
1657 no_subpage_counting = true;
1659 goto thp_subpage_migration;
1662 rc = nr_failed + nr_thp_failed;
1665 * Put the permanent failure page back to migration list, they
1666 * will be put back to the right list by the caller.
1668 list_splice(&ret_pages, from);
1670 count_vm_events(PGMIGRATE_SUCCESS, nr_succeeded);
1671 count_vm_events(PGMIGRATE_FAIL, nr_failed_pages);
1672 count_vm_events(THP_MIGRATION_SUCCESS, nr_thp_succeeded);
1673 count_vm_events(THP_MIGRATION_FAIL, nr_thp_failed);
1674 count_vm_events(THP_MIGRATION_SPLIT, nr_thp_split);
1675 trace_mm_migrate_pages(nr_succeeded, nr_failed_pages, nr_thp_succeeded,
1676 nr_thp_failed, nr_thp_split, mode, reason);
1679 current->flags &= ~PF_SWAPWRITE;
1682 *ret_succeeded = nr_succeeded;
1687 struct page *alloc_migration_target(struct page *page, unsigned long private)
1689 struct migration_target_control *mtc;
1691 unsigned int order = 0;
1692 struct page *new_page = NULL;
1696 mtc = (struct migration_target_control *)private;
1697 gfp_mask = mtc->gfp_mask;
1699 if (nid == NUMA_NO_NODE)
1700 nid = page_to_nid(page);
1702 if (PageHuge(page)) {
1703 struct hstate *h = page_hstate(compound_head(page));
1705 gfp_mask = htlb_modify_alloc_mask(h, gfp_mask);
1706 return alloc_huge_page_nodemask(h, nid, mtc->nmask, gfp_mask);
1709 if (PageTransHuge(page)) {
1711 * clear __GFP_RECLAIM to make the migration callback
1712 * consistent with regular THP allocations.
1714 gfp_mask &= ~__GFP_RECLAIM;
1715 gfp_mask |= GFP_TRANSHUGE;
1716 order = HPAGE_PMD_ORDER;
1718 zidx = zone_idx(page_zone(page));
1719 if (is_highmem_idx(zidx) || zidx == ZONE_MOVABLE)
1720 gfp_mask |= __GFP_HIGHMEM;
1722 new_page = __alloc_pages(gfp_mask, order, nid, mtc->nmask);
1724 if (new_page && PageTransHuge(new_page))
1725 prep_transhuge_page(new_page);
1732 static int store_status(int __user *status, int start, int value, int nr)
1735 if (put_user(value, status + start))
1743 static int do_move_pages_to_node(struct mm_struct *mm,
1744 struct list_head *pagelist, int node)
1747 struct migration_target_control mtc = {
1749 .gfp_mask = GFP_HIGHUSER_MOVABLE | __GFP_THISNODE,
1752 err = migrate_pages(pagelist, alloc_migration_target, NULL,
1753 (unsigned long)&mtc, MIGRATE_SYNC, MR_SYSCALL, NULL);
1755 putback_movable_pages(pagelist);
1760 * Resolves the given address to a struct page, isolates it from the LRU and
1761 * puts it to the given pagelist.
1763 * errno - if the page cannot be found/isolated
1764 * 0 - when it doesn't have to be migrated because it is already on the
1766 * 1 - when it has been queued
1768 static int add_page_for_migration(struct mm_struct *mm, unsigned long addr,
1769 int node, struct list_head *pagelist, bool migrate_all)
1771 struct vm_area_struct *vma;
1773 unsigned int follflags;
1778 vma = find_vma(mm, addr);
1779 if (!vma || addr < vma->vm_start || !vma_migratable(vma))
1782 /* FOLL_DUMP to ignore special (like zero) pages */
1783 follflags = FOLL_GET | FOLL_DUMP;
1784 page = follow_page(vma, addr, follflags);
1786 err = PTR_ERR(page);
1795 if (page_to_nid(page) == node)
1799 if (page_mapcount(page) > 1 && !migrate_all)
1802 if (PageHuge(page)) {
1803 if (PageHead(page)) {
1804 isolate_huge_page(page, pagelist);
1810 head = compound_head(page);
1811 err = isolate_lru_page(head);
1816 list_add_tail(&head->lru, pagelist);
1817 mod_node_page_state(page_pgdat(head),
1818 NR_ISOLATED_ANON + page_is_file_lru(head),
1819 thp_nr_pages(head));
1823 * Either remove the duplicate refcount from
1824 * isolate_lru_page() or drop the page ref if it was
1829 mmap_read_unlock(mm);
1833 static int move_pages_and_store_status(struct mm_struct *mm, int node,
1834 struct list_head *pagelist, int __user *status,
1835 int start, int i, unsigned long nr_pages)
1839 if (list_empty(pagelist))
1842 err = do_move_pages_to_node(mm, pagelist, node);
1845 * Positive err means the number of failed
1846 * pages to migrate. Since we are going to
1847 * abort and return the number of non-migrated
1848 * pages, so need to include the rest of the
1849 * nr_pages that have not been attempted as
1853 err += nr_pages - i - 1;
1856 return store_status(status, start, node, i - start);
1860 * Migrate an array of page address onto an array of nodes and fill
1861 * the corresponding array of status.
1863 static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes,
1864 unsigned long nr_pages,
1865 const void __user * __user *pages,
1866 const int __user *nodes,
1867 int __user *status, int flags)
1869 int current_node = NUMA_NO_NODE;
1870 LIST_HEAD(pagelist);
1874 lru_cache_disable();
1876 for (i = start = 0; i < nr_pages; i++) {
1877 const void __user *p;
1882 if (get_user(p, pages + i))
1884 if (get_user(node, nodes + i))
1886 addr = (unsigned long)untagged_addr(p);
1889 if (node < 0 || node >= MAX_NUMNODES)
1891 if (!node_state(node, N_MEMORY))
1895 if (!node_isset(node, task_nodes))
1898 if (current_node == NUMA_NO_NODE) {
1899 current_node = node;
1901 } else if (node != current_node) {
1902 err = move_pages_and_store_status(mm, current_node,
1903 &pagelist, status, start, i, nr_pages);
1907 current_node = node;
1911 * Errors in the page lookup or isolation are not fatal and we simply
1912 * report them via status
1914 err = add_page_for_migration(mm, addr, current_node,
1915 &pagelist, flags & MPOL_MF_MOVE_ALL);
1918 /* The page is successfully queued for migration */
1923 * If the page is already on the target node (!err), store the
1924 * node, otherwise, store the err.
1926 err = store_status(status, i, err ? : current_node, 1);
1930 err = move_pages_and_store_status(mm, current_node, &pagelist,
1931 status, start, i, nr_pages);
1934 current_node = NUMA_NO_NODE;
1937 /* Make sure we do not overwrite the existing error */
1938 err1 = move_pages_and_store_status(mm, current_node, &pagelist,
1939 status, start, i, nr_pages);
1948 * Determine the nodes of an array of pages and store it in an array of status.
1950 static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages,
1951 const void __user **pages, int *status)
1957 for (i = 0; i < nr_pages; i++) {
1958 unsigned long addr = (unsigned long)(*pages);
1959 struct vm_area_struct *vma;
1963 vma = vma_lookup(mm, addr);
1967 /* FOLL_DUMP to ignore special (like zero) pages */
1968 page = follow_page(vma, addr, FOLL_DUMP);
1970 err = PTR_ERR(page);
1974 err = page ? page_to_nid(page) : -ENOENT;
1982 mmap_read_unlock(mm);
1985 static int get_compat_pages_array(const void __user *chunk_pages[],
1986 const void __user * __user *pages,
1987 unsigned long chunk_nr)
1989 compat_uptr_t __user *pages32 = (compat_uptr_t __user *)pages;
1993 for (i = 0; i < chunk_nr; i++) {
1994 if (get_user(p, pages32 + i))
1996 chunk_pages[i] = compat_ptr(p);
2003 * Determine the nodes of a user array of pages and store it in
2004 * a user array of status.
2006 static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages,
2007 const void __user * __user *pages,
2010 #define DO_PAGES_STAT_CHUNK_NR 16
2011 const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR];
2012 int chunk_status[DO_PAGES_STAT_CHUNK_NR];
2015 unsigned long chunk_nr;
2017 chunk_nr = nr_pages;
2018 if (chunk_nr > DO_PAGES_STAT_CHUNK_NR)
2019 chunk_nr = DO_PAGES_STAT_CHUNK_NR;
2021 if (in_compat_syscall()) {
2022 if (get_compat_pages_array(chunk_pages, pages,
2026 if (copy_from_user(chunk_pages, pages,
2027 chunk_nr * sizeof(*chunk_pages)))
2031 do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status);
2033 if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status)))
2038 nr_pages -= chunk_nr;
2040 return nr_pages ? -EFAULT : 0;
2043 static struct mm_struct *find_mm_struct(pid_t pid, nodemask_t *mem_nodes)
2045 struct task_struct *task;
2046 struct mm_struct *mm;
2049 * There is no need to check if current process has the right to modify
2050 * the specified process when they are same.
2054 *mem_nodes = cpuset_mems_allowed(current);
2058 /* Find the mm_struct */
2060 task = find_task_by_vpid(pid);
2063 return ERR_PTR(-ESRCH);
2065 get_task_struct(task);
2068 * Check if this process has the right to modify the specified
2069 * process. Use the regular "ptrace_may_access()" checks.
2071 if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) {
2073 mm = ERR_PTR(-EPERM);
2078 mm = ERR_PTR(security_task_movememory(task));
2081 *mem_nodes = cpuset_mems_allowed(task);
2082 mm = get_task_mm(task);
2084 put_task_struct(task);
2086 mm = ERR_PTR(-EINVAL);
2091 * Move a list of pages in the address space of the currently executing
2094 static int kernel_move_pages(pid_t pid, unsigned long nr_pages,
2095 const void __user * __user *pages,
2096 const int __user *nodes,
2097 int __user *status, int flags)
2099 struct mm_struct *mm;
2101 nodemask_t task_nodes;
2104 if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
2107 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
2110 mm = find_mm_struct(pid, &task_nodes);
2115 err = do_pages_move(mm, task_nodes, nr_pages, pages,
2116 nodes, status, flags);
2118 err = do_pages_stat(mm, nr_pages, pages, status);
2124 SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages,
2125 const void __user * __user *, pages,
2126 const int __user *, nodes,
2127 int __user *, status, int, flags)
2129 return kernel_move_pages(pid, nr_pages, pages, nodes, status, flags);
2132 #ifdef CONFIG_NUMA_BALANCING
2134 * Returns true if this is a safe migration target node for misplaced NUMA
2135 * pages. Currently it only checks the watermarks which crude
2137 static bool migrate_balanced_pgdat(struct pglist_data *pgdat,
2138 unsigned long nr_migrate_pages)
2142 for (z = pgdat->nr_zones - 1; z >= 0; z--) {
2143 struct zone *zone = pgdat->node_zones + z;
2145 if (!populated_zone(zone))
2148 /* Avoid waking kswapd by allocating pages_to_migrate pages. */
2149 if (!zone_watermark_ok(zone, 0,
2150 high_wmark_pages(zone) +
2159 static struct page *alloc_misplaced_dst_page(struct page *page,
2162 int nid = (int) data;
2163 struct page *newpage;
2165 newpage = __alloc_pages_node(nid,
2166 (GFP_HIGHUSER_MOVABLE |
2167 __GFP_THISNODE | __GFP_NOMEMALLOC |
2168 __GFP_NORETRY | __GFP_NOWARN) &
2174 static struct page *alloc_misplaced_dst_page_thp(struct page *page,
2177 int nid = (int) data;
2178 struct page *newpage;
2180 newpage = alloc_pages_node(nid, (GFP_TRANSHUGE_LIGHT | __GFP_THISNODE),
2185 prep_transhuge_page(newpage);
2191 static int numamigrate_isolate_page(pg_data_t *pgdat, struct page *page)
2194 int nr_pages = thp_nr_pages(page);
2196 VM_BUG_ON_PAGE(compound_order(page) && !PageTransHuge(page), page);
2198 /* Do not migrate THP mapped by multiple processes */
2199 if (PageTransHuge(page) && total_mapcount(page) > 1)
2202 /* Avoid migrating to a node that is nearly full */
2203 if (!migrate_balanced_pgdat(pgdat, nr_pages))
2206 if (isolate_lru_page(page))
2209 page_lru = page_is_file_lru(page);
2210 mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON + page_lru,
2214 * Isolating the page has taken another reference, so the
2215 * caller's reference can be safely dropped without the page
2216 * disappearing underneath us during migration.
2223 * Attempt to migrate a misplaced page to the specified destination
2224 * node. Caller is expected to have an elevated reference count on
2225 * the page that will be dropped by this function before returning.
2227 int migrate_misplaced_page(struct page *page, struct vm_area_struct *vma,
2230 pg_data_t *pgdat = NODE_DATA(node);
2233 LIST_HEAD(migratepages);
2236 int nr_pages = thp_nr_pages(page);
2239 * PTE mapped THP or HugeTLB page can't reach here so the page could
2240 * be either base page or THP. And it must be head page if it is
2243 compound = PageTransHuge(page);
2246 new = alloc_misplaced_dst_page_thp;
2248 new = alloc_misplaced_dst_page;
2251 * Don't migrate file pages that are mapped in multiple processes
2252 * with execute permissions as they are probably shared libraries.
2254 if (page_mapcount(page) != 1 && page_is_file_lru(page) &&
2255 (vma->vm_flags & VM_EXEC))
2259 * Also do not migrate dirty pages as not all filesystems can move
2260 * dirty pages in MIGRATE_ASYNC mode which is a waste of cycles.
2262 if (page_is_file_lru(page) && PageDirty(page))
2265 isolated = numamigrate_isolate_page(pgdat, page);
2269 list_add(&page->lru, &migratepages);
2270 nr_remaining = migrate_pages(&migratepages, *new, NULL, node,
2271 MIGRATE_ASYNC, MR_NUMA_MISPLACED, NULL);
2273 if (!list_empty(&migratepages)) {
2274 list_del(&page->lru);
2275 mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON +
2276 page_is_file_lru(page), -nr_pages);
2277 putback_lru_page(page);
2281 count_vm_numa_events(NUMA_PAGE_MIGRATE, nr_pages);
2282 BUG_ON(!list_empty(&migratepages));
2289 #endif /* CONFIG_NUMA_BALANCING */
2290 #endif /* CONFIG_NUMA */
2292 #ifdef CONFIG_DEVICE_PRIVATE
2293 static int migrate_vma_collect_skip(unsigned long start,
2295 struct mm_walk *walk)
2297 struct migrate_vma *migrate = walk->private;
2300 for (addr = start; addr < end; addr += PAGE_SIZE) {
2301 migrate->dst[migrate->npages] = 0;
2302 migrate->src[migrate->npages++] = 0;
2308 static int migrate_vma_collect_hole(unsigned long start,
2310 __always_unused int depth,
2311 struct mm_walk *walk)
2313 struct migrate_vma *migrate = walk->private;
2316 /* Only allow populating anonymous memory. */
2317 if (!vma_is_anonymous(walk->vma))
2318 return migrate_vma_collect_skip(start, end, walk);
2320 for (addr = start; addr < end; addr += PAGE_SIZE) {
2321 migrate->src[migrate->npages] = MIGRATE_PFN_MIGRATE;
2322 migrate->dst[migrate->npages] = 0;
2330 static int migrate_vma_collect_pmd(pmd_t *pmdp,
2331 unsigned long start,
2333 struct mm_walk *walk)
2335 struct migrate_vma *migrate = walk->private;
2336 struct vm_area_struct *vma = walk->vma;
2337 struct mm_struct *mm = vma->vm_mm;
2338 unsigned long addr = start, unmapped = 0;
2343 if (pmd_none(*pmdp))
2344 return migrate_vma_collect_hole(start, end, -1, walk);
2346 if (pmd_trans_huge(*pmdp)) {
2349 ptl = pmd_lock(mm, pmdp);
2350 if (unlikely(!pmd_trans_huge(*pmdp))) {
2355 page = pmd_page(*pmdp);
2356 if (is_huge_zero_page(page)) {
2358 split_huge_pmd(vma, pmdp, addr);
2359 if (pmd_trans_unstable(pmdp))
2360 return migrate_vma_collect_skip(start, end,
2367 if (unlikely(!trylock_page(page)))
2368 return migrate_vma_collect_skip(start, end,
2370 ret = split_huge_page(page);
2374 return migrate_vma_collect_skip(start, end,
2376 if (pmd_none(*pmdp))
2377 return migrate_vma_collect_hole(start, end, -1,
2382 if (unlikely(pmd_bad(*pmdp)))
2383 return migrate_vma_collect_skip(start, end, walk);
2385 ptep = pte_offset_map_lock(mm, pmdp, addr, &ptl);
2386 arch_enter_lazy_mmu_mode();
2388 for (; addr < end; addr += PAGE_SIZE, ptep++) {
2389 unsigned long mpfn = 0, pfn;
2396 if (pte_none(pte)) {
2397 if (vma_is_anonymous(vma)) {
2398 mpfn = MIGRATE_PFN_MIGRATE;
2404 if (!pte_present(pte)) {
2406 * Only care about unaddressable device page special
2407 * page table entry. Other special swap entries are not
2408 * migratable, and we ignore regular swapped page.
2410 entry = pte_to_swp_entry(pte);
2411 if (!is_device_private_entry(entry))
2414 page = pfn_swap_entry_to_page(entry);
2415 if (!(migrate->flags &
2416 MIGRATE_VMA_SELECT_DEVICE_PRIVATE) ||
2417 page->pgmap->owner != migrate->pgmap_owner)
2420 mpfn = migrate_pfn(page_to_pfn(page)) |
2421 MIGRATE_PFN_MIGRATE;
2422 if (is_writable_device_private_entry(entry))
2423 mpfn |= MIGRATE_PFN_WRITE;
2425 if (!(migrate->flags & MIGRATE_VMA_SELECT_SYSTEM))
2428 if (is_zero_pfn(pfn)) {
2429 mpfn = MIGRATE_PFN_MIGRATE;
2433 page = vm_normal_page(migrate->vma, addr, pte);
2434 mpfn = migrate_pfn(pfn) | MIGRATE_PFN_MIGRATE;
2435 mpfn |= pte_write(pte) ? MIGRATE_PFN_WRITE : 0;
2438 /* FIXME support THP */
2439 if (!page || !page->mapping || PageTransCompound(page)) {
2445 * By getting a reference on the page we pin it and that blocks
2446 * any kind of migration. Side effect is that it "freezes" the
2449 * We drop this reference after isolating the page from the lru
2450 * for non device page (device page are not on the lru and thus
2451 * can't be dropped from it).
2456 * Optimize for the common case where page is only mapped once
2457 * in one process. If we can lock the page, then we can safely
2458 * set up a special migration page table entry now.
2460 if (trylock_page(page)) {
2464 ptep_get_and_clear(mm, addr, ptep);
2466 /* Setup special migration page table entry */
2467 if (mpfn & MIGRATE_PFN_WRITE)
2468 entry = make_writable_migration_entry(
2471 entry = make_readable_migration_entry(
2473 swp_pte = swp_entry_to_pte(entry);
2474 if (pte_present(pte)) {
2475 if (pte_soft_dirty(pte))
2476 swp_pte = pte_swp_mksoft_dirty(swp_pte);
2477 if (pte_uffd_wp(pte))
2478 swp_pte = pte_swp_mkuffd_wp(swp_pte);
2480 if (pte_swp_soft_dirty(pte))
2481 swp_pte = pte_swp_mksoft_dirty(swp_pte);
2482 if (pte_swp_uffd_wp(pte))
2483 swp_pte = pte_swp_mkuffd_wp(swp_pte);
2485 set_pte_at(mm, addr, ptep, swp_pte);
2488 * This is like regular unmap: we remove the rmap and
2489 * drop page refcount. Page won't be freed, as we took
2490 * a reference just above.
2492 page_remove_rmap(page, false);
2495 if (pte_present(pte))
2503 migrate->dst[migrate->npages] = 0;
2504 migrate->src[migrate->npages++] = mpfn;
2506 arch_leave_lazy_mmu_mode();
2507 pte_unmap_unlock(ptep - 1, ptl);
2509 /* Only flush the TLB if we actually modified any entries */
2511 flush_tlb_range(walk->vma, start, end);
2516 static const struct mm_walk_ops migrate_vma_walk_ops = {
2517 .pmd_entry = migrate_vma_collect_pmd,
2518 .pte_hole = migrate_vma_collect_hole,
2522 * migrate_vma_collect() - collect pages over a range of virtual addresses
2523 * @migrate: migrate struct containing all migration information
2525 * This will walk the CPU page table. For each virtual address backed by a
2526 * valid page, it updates the src array and takes a reference on the page, in
2527 * order to pin the page until we lock it and unmap it.
2529 static void migrate_vma_collect(struct migrate_vma *migrate)
2531 struct mmu_notifier_range range;
2534 * Note that the pgmap_owner is passed to the mmu notifier callback so
2535 * that the registered device driver can skip invalidating device
2536 * private page mappings that won't be migrated.
2538 mmu_notifier_range_init_owner(&range, MMU_NOTIFY_MIGRATE, 0,
2539 migrate->vma, migrate->vma->vm_mm, migrate->start, migrate->end,
2540 migrate->pgmap_owner);
2541 mmu_notifier_invalidate_range_start(&range);
2543 walk_page_range(migrate->vma->vm_mm, migrate->start, migrate->end,
2544 &migrate_vma_walk_ops, migrate);
2546 mmu_notifier_invalidate_range_end(&range);
2547 migrate->end = migrate->start + (migrate->npages << PAGE_SHIFT);
2551 * migrate_vma_check_page() - check if page is pinned or not
2552 * @page: struct page to check
2554 * Pinned pages cannot be migrated. This is the same test as in
2555 * folio_migrate_mapping(), except that here we allow migration of a
2558 static bool migrate_vma_check_page(struct page *page)
2561 * One extra ref because caller holds an extra reference, either from
2562 * isolate_lru_page() for a regular page, or migrate_vma_collect() for
2568 * FIXME support THP (transparent huge page), it is bit more complex to
2569 * check them than regular pages, because they can be mapped with a pmd
2570 * or with a pte (split pte mapping).
2572 if (PageCompound(page))
2575 /* Page from ZONE_DEVICE have one extra reference */
2576 if (is_zone_device_page(page)) {
2578 * Private page can never be pin as they have no valid pte and
2579 * GUP will fail for those. Yet if there is a pending migration
2580 * a thread might try to wait on the pte migration entry and
2581 * will bump the page reference count. Sadly there is no way to
2582 * differentiate a regular pin from migration wait. Hence to
2583 * avoid 2 racing thread trying to migrate back to CPU to enter
2584 * infinite loop (one stopping migration because the other is
2585 * waiting on pte migration entry). We always return true here.
2587 * FIXME proper solution is to rework migration_entry_wait() so
2588 * it does not need to take a reference on page.
2590 return is_device_private_page(page);
2593 /* For file back page */
2594 if (page_mapping(page))
2595 extra += 1 + page_has_private(page);
2597 if ((page_count(page) - extra) > page_mapcount(page))
2604 * migrate_vma_unmap() - replace page mapping with special migration pte entry
2605 * @migrate: migrate struct containing all migration information
2607 * Isolate pages from the LRU and replace mappings (CPU page table pte) with a
2608 * special migration pte entry and check if it has been pinned. Pinned pages are
2609 * restored because we cannot migrate them.
2611 * This is the last step before we call the device driver callback to allocate
2612 * destination memory and copy contents of original page over to new page.
2614 static void migrate_vma_unmap(struct migrate_vma *migrate)
2616 const unsigned long npages = migrate->npages;
2617 const unsigned long start = migrate->start;
2618 unsigned long addr, i, restore = 0;
2619 bool allow_drain = true;
2623 for (i = 0; i < npages; i++) {
2624 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2629 /* ZONE_DEVICE pages are not on LRU */
2630 if (!is_zone_device_page(page)) {
2631 if (!PageLRU(page) && allow_drain) {
2632 /* Drain CPU's pagevec */
2633 lru_add_drain_all();
2634 allow_drain = false;
2637 if (isolate_lru_page(page)) {
2638 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2644 /* Drop the reference we took in collect */
2648 if (page_mapped(page))
2649 try_to_migrate(page, 0);
2651 if (page_mapped(page) || !migrate_vma_check_page(page)) {
2652 if (!is_zone_device_page(page)) {
2654 putback_lru_page(page);
2657 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2664 for (addr = start, i = 0; i < npages && restore; addr += PAGE_SIZE, i++) {
2665 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2667 if (!page || (migrate->src[i] & MIGRATE_PFN_MIGRATE))
2670 remove_migration_ptes(page, page, false);
2672 migrate->src[i] = 0;
2680 * migrate_vma_setup() - prepare to migrate a range of memory
2681 * @args: contains the vma, start, and pfns arrays for the migration
2683 * Returns: negative errno on failures, 0 when 0 or more pages were migrated
2686 * Prepare to migrate a range of memory virtual address range by collecting all
2687 * the pages backing each virtual address in the range, saving them inside the
2688 * src array. Then lock those pages and unmap them. Once the pages are locked
2689 * and unmapped, check whether each page is pinned or not. Pages that aren't
2690 * pinned have the MIGRATE_PFN_MIGRATE flag set (by this function) in the
2691 * corresponding src array entry. Then restores any pages that are pinned, by
2692 * remapping and unlocking those pages.
2694 * The caller should then allocate destination memory and copy source memory to
2695 * it for all those entries (ie with MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE
2696 * flag set). Once these are allocated and copied, the caller must update each
2697 * corresponding entry in the dst array with the pfn value of the destination
2698 * page and with MIGRATE_PFN_VALID. Destination pages must be locked via
2701 * Note that the caller does not have to migrate all the pages that are marked
2702 * with MIGRATE_PFN_MIGRATE flag in src array unless this is a migration from
2703 * device memory to system memory. If the caller cannot migrate a device page
2704 * back to system memory, then it must return VM_FAULT_SIGBUS, which has severe
2705 * consequences for the userspace process, so it must be avoided if at all
2708 * For empty entries inside CPU page table (pte_none() or pmd_none() is true) we
2709 * do set MIGRATE_PFN_MIGRATE flag inside the corresponding source array thus
2710 * allowing the caller to allocate device memory for those unbacked virtual
2711 * addresses. For this the caller simply has to allocate device memory and
2712 * properly set the destination entry like for regular migration. Note that
2713 * this can still fail, and thus inside the device driver you must check if the
2714 * migration was successful for those entries after calling migrate_vma_pages(),
2715 * just like for regular migration.
2717 * After that, the callers must call migrate_vma_pages() to go over each entry
2718 * in the src array that has the MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE flag
2719 * set. If the corresponding entry in dst array has MIGRATE_PFN_VALID flag set,
2720 * then migrate_vma_pages() to migrate struct page information from the source
2721 * struct page to the destination struct page. If it fails to migrate the
2722 * struct page information, then it clears the MIGRATE_PFN_MIGRATE flag in the
2725 * At this point all successfully migrated pages have an entry in the src
2726 * array with MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE flag set and the dst
2727 * array entry with MIGRATE_PFN_VALID flag set.
2729 * Once migrate_vma_pages() returns the caller may inspect which pages were
2730 * successfully migrated, and which were not. Successfully migrated pages will
2731 * have the MIGRATE_PFN_MIGRATE flag set for their src array entry.
2733 * It is safe to update device page table after migrate_vma_pages() because
2734 * both destination and source page are still locked, and the mmap_lock is held
2735 * in read mode (hence no one can unmap the range being migrated).
2737 * Once the caller is done cleaning up things and updating its page table (if it
2738 * chose to do so, this is not an obligation) it finally calls
2739 * migrate_vma_finalize() to update the CPU page table to point to new pages
2740 * for successfully migrated pages or otherwise restore the CPU page table to
2741 * point to the original source pages.
2743 int migrate_vma_setup(struct migrate_vma *args)
2745 long nr_pages = (args->end - args->start) >> PAGE_SHIFT;
2747 args->start &= PAGE_MASK;
2748 args->end &= PAGE_MASK;
2749 if (!args->vma || is_vm_hugetlb_page(args->vma) ||
2750 (args->vma->vm_flags & VM_SPECIAL) || vma_is_dax(args->vma))
2754 if (args->start < args->vma->vm_start ||
2755 args->start >= args->vma->vm_end)
2757 if (args->end <= args->vma->vm_start || args->end > args->vma->vm_end)
2759 if (!args->src || !args->dst)
2762 memset(args->src, 0, sizeof(*args->src) * nr_pages);
2766 migrate_vma_collect(args);
2769 migrate_vma_unmap(args);
2772 * At this point pages are locked and unmapped, and thus they have
2773 * stable content and can safely be copied to destination memory that
2774 * is allocated by the drivers.
2779 EXPORT_SYMBOL(migrate_vma_setup);
2782 * This code closely matches the code in:
2783 * __handle_mm_fault()
2784 * handle_pte_fault()
2785 * do_anonymous_page()
2786 * to map in an anonymous zero page but the struct page will be a ZONE_DEVICE
2789 static void migrate_vma_insert_page(struct migrate_vma *migrate,
2794 struct vm_area_struct *vma = migrate->vma;
2795 struct mm_struct *mm = vma->vm_mm;
2805 /* Only allow populating anonymous memory */
2806 if (!vma_is_anonymous(vma))
2809 pgdp = pgd_offset(mm, addr);
2810 p4dp = p4d_alloc(mm, pgdp, addr);
2813 pudp = pud_alloc(mm, p4dp, addr);
2816 pmdp = pmd_alloc(mm, pudp, addr);
2820 if (pmd_trans_huge(*pmdp) || pmd_devmap(*pmdp))
2824 * Use pte_alloc() instead of pte_alloc_map(). We can't run
2825 * pte_offset_map() on pmds where a huge pmd might be created
2826 * from a different thread.
2828 * pte_alloc_map() is safe to use under mmap_write_lock(mm) or when
2829 * parallel threads are excluded by other means.
2831 * Here we only have mmap_read_lock(mm).
2833 if (pte_alloc(mm, pmdp))
2836 /* See the comment in pte_alloc_one_map() */
2837 if (unlikely(pmd_trans_unstable(pmdp)))
2840 if (unlikely(anon_vma_prepare(vma)))
2842 if (mem_cgroup_charge(page_folio(page), vma->vm_mm, GFP_KERNEL))
2846 * The memory barrier inside __SetPageUptodate makes sure that
2847 * preceding stores to the page contents become visible before
2848 * the set_pte_at() write.
2850 __SetPageUptodate(page);
2852 if (is_zone_device_page(page)) {
2853 if (is_device_private_page(page)) {
2854 swp_entry_t swp_entry;
2856 if (vma->vm_flags & VM_WRITE)
2857 swp_entry = make_writable_device_private_entry(
2860 swp_entry = make_readable_device_private_entry(
2862 entry = swp_entry_to_pte(swp_entry);
2865 * For now we only support migrating to un-addressable
2868 pr_warn_once("Unsupported ZONE_DEVICE page type.\n");
2872 entry = mk_pte(page, vma->vm_page_prot);
2873 if (vma->vm_flags & VM_WRITE)
2874 entry = pte_mkwrite(pte_mkdirty(entry));
2877 ptep = pte_offset_map_lock(mm, pmdp, addr, &ptl);
2879 if (check_stable_address_space(mm))
2882 if (pte_present(*ptep)) {
2883 unsigned long pfn = pte_pfn(*ptep);
2885 if (!is_zero_pfn(pfn))
2888 } else if (!pte_none(*ptep))
2892 * Check for userfaultfd but do not deliver the fault. Instead,
2895 if (userfaultfd_missing(vma))
2898 inc_mm_counter(mm, MM_ANONPAGES);
2899 page_add_new_anon_rmap(page, vma, addr, false);
2900 if (!is_zone_device_page(page))
2901 lru_cache_add_inactive_or_unevictable(page, vma);
2905 flush_cache_page(vma, addr, pte_pfn(*ptep));
2906 ptep_clear_flush_notify(vma, addr, ptep);
2907 set_pte_at_notify(mm, addr, ptep, entry);
2908 update_mmu_cache(vma, addr, ptep);
2910 /* No need to invalidate - it was non-present before */
2911 set_pte_at(mm, addr, ptep, entry);
2912 update_mmu_cache(vma, addr, ptep);
2915 pte_unmap_unlock(ptep, ptl);
2916 *src = MIGRATE_PFN_MIGRATE;
2920 pte_unmap_unlock(ptep, ptl);
2922 *src &= ~MIGRATE_PFN_MIGRATE;
2926 * migrate_vma_pages() - migrate meta-data from src page to dst page
2927 * @migrate: migrate struct containing all migration information
2929 * This migrates struct page meta-data from source struct page to destination
2930 * struct page. This effectively finishes the migration from source page to the
2933 void migrate_vma_pages(struct migrate_vma *migrate)
2935 const unsigned long npages = migrate->npages;
2936 const unsigned long start = migrate->start;
2937 struct mmu_notifier_range range;
2938 unsigned long addr, i;
2939 bool notified = false;
2941 for (i = 0, addr = start; i < npages; addr += PAGE_SIZE, i++) {
2942 struct page *newpage = migrate_pfn_to_page(migrate->dst[i]);
2943 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2944 struct address_space *mapping;
2948 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2953 if (!(migrate->src[i] & MIGRATE_PFN_MIGRATE))
2958 mmu_notifier_range_init_owner(&range,
2959 MMU_NOTIFY_MIGRATE, 0, migrate->vma,
2960 migrate->vma->vm_mm, addr, migrate->end,
2961 migrate->pgmap_owner);
2962 mmu_notifier_invalidate_range_start(&range);
2964 migrate_vma_insert_page(migrate, addr, newpage,
2969 mapping = page_mapping(page);
2971 if (is_zone_device_page(newpage)) {
2972 if (is_device_private_page(newpage)) {
2974 * For now only support private anonymous when
2975 * migrating to un-addressable device memory.
2978 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2983 * Other types of ZONE_DEVICE page are not
2986 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2991 r = migrate_page(mapping, newpage, page, MIGRATE_SYNC_NO_COPY);
2992 if (r != MIGRATEPAGE_SUCCESS)
2993 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2997 * No need to double call mmu_notifier->invalidate_range() callback as
2998 * the above ptep_clear_flush_notify() inside migrate_vma_insert_page()
2999 * did already call it.
3002 mmu_notifier_invalidate_range_only_end(&range);
3004 EXPORT_SYMBOL(migrate_vma_pages);
3007 * migrate_vma_finalize() - restore CPU page table entry
3008 * @migrate: migrate struct containing all migration information
3010 * This replaces the special migration pte entry with either a mapping to the
3011 * new page if migration was successful for that page, or to the original page
3014 * This also unlocks the pages and puts them back on the lru, or drops the extra
3015 * refcount, for device pages.
3017 void migrate_vma_finalize(struct migrate_vma *migrate)
3019 const unsigned long npages = migrate->npages;
3022 for (i = 0; i < npages; i++) {
3023 struct page *newpage = migrate_pfn_to_page(migrate->dst[i]);
3024 struct page *page = migrate_pfn_to_page(migrate->src[i]);
3028 unlock_page(newpage);
3034 if (!(migrate->src[i] & MIGRATE_PFN_MIGRATE) || !newpage) {
3036 unlock_page(newpage);
3042 remove_migration_ptes(page, newpage, false);
3045 if (is_zone_device_page(page))
3048 putback_lru_page(page);
3050 if (newpage != page) {
3051 unlock_page(newpage);
3052 if (is_zone_device_page(newpage))
3055 putback_lru_page(newpage);
3059 EXPORT_SYMBOL(migrate_vma_finalize);
3060 #endif /* CONFIG_DEVICE_PRIVATE */
3062 #if defined(CONFIG_HOTPLUG_CPU)
3063 /* Disable reclaim-based migration. */
3064 static void __disable_all_migrate_targets(void)
3071 for_each_online_node(node) {
3072 node_demotion[node].nr = 0;
3073 for (i = 0; i < DEMOTION_TARGET_NODES; i++)
3074 node_demotion[node].nodes[i] = NUMA_NO_NODE;
3078 static void disable_all_migrate_targets(void)
3080 __disable_all_migrate_targets();
3083 * Ensure that the "disable" is visible across the system.
3084 * Readers will see either a combination of before+disable
3085 * state or disable+after. They will never see before and
3086 * after state together.
3088 * The before+after state together might have cycles and
3089 * could cause readers to do things like loop until this
3090 * function finishes. This ensures they can only see a
3091 * single "bad" read and would, for instance, only loop
3098 * Find an automatic demotion target for 'node'.
3099 * Failing here is OK. It might just indicate
3100 * being at the end of a chain.
3102 static int establish_migrate_target(int node, nodemask_t *used,
3105 int migration_target, index, val;
3106 struct demotion_nodes *nd;
3109 return NUMA_NO_NODE;
3111 nd = &node_demotion[node];
3113 migration_target = find_next_best_node(node, used);
3114 if (migration_target == NUMA_NO_NODE)
3115 return NUMA_NO_NODE;
3118 * If the node has been set a migration target node before,
3119 * which means it's the best distance between them. Still
3120 * check if this node can be demoted to other target nodes
3121 * if they have a same best distance.
3123 if (best_distance != -1) {
3124 val = node_distance(node, migration_target);
3125 if (val > best_distance)
3126 return NUMA_NO_NODE;
3130 if (WARN_ONCE(index >= DEMOTION_TARGET_NODES,
3131 "Exceeds maximum demotion target nodes\n"))
3132 return NUMA_NO_NODE;
3134 nd->nodes[index] = migration_target;
3137 return migration_target;
3141 * When memory fills up on a node, memory contents can be
3142 * automatically migrated to another node instead of
3143 * discarded at reclaim.
3145 * Establish a "migration path" which will start at nodes
3146 * with CPUs and will follow the priorities used to build the
3147 * page allocator zonelists.
3149 * The difference here is that cycles must be avoided. If
3150 * node0 migrates to node1, then neither node1, nor anything
3151 * node1 migrates to can migrate to node0. Also one node can
3152 * be migrated to multiple nodes if the target nodes all have
3153 * a same best-distance against the source node.
3155 * This function can run simultaneously with readers of
3156 * node_demotion[]. However, it can not run simultaneously
3157 * with itself. Exclusion is provided by memory hotplug events
3158 * being single-threaded.
3160 static void __set_migration_target_nodes(void)
3162 nodemask_t next_pass = NODE_MASK_NONE;
3163 nodemask_t this_pass = NODE_MASK_NONE;
3164 nodemask_t used_targets = NODE_MASK_NONE;
3165 int node, best_distance;
3168 * Avoid any oddities like cycles that could occur
3169 * from changes in the topology. This will leave
3170 * a momentary gap when migration is disabled.
3172 disable_all_migrate_targets();
3175 * Allocations go close to CPUs, first. Assume that
3176 * the migration path starts at the nodes with CPUs.
3178 next_pass = node_states[N_CPU];
3180 this_pass = next_pass;
3181 next_pass = NODE_MASK_NONE;
3183 * To avoid cycles in the migration "graph", ensure
3184 * that migration sources are not future targets by
3185 * setting them in 'used_targets'. Do this only
3186 * once per pass so that multiple source nodes can
3187 * share a target node.
3189 * 'used_targets' will become unavailable in future
3190 * passes. This limits some opportunities for
3191 * multiple source nodes to share a destination.
3193 nodes_or(used_targets, used_targets, this_pass);
3195 for_each_node_mask(node, this_pass) {
3199 * Try to set up the migration path for the node, and the target
3200 * migration nodes can be multiple, so doing a loop to find all
3201 * the target nodes if they all have a best node distance.
3205 establish_migrate_target(node, &used_targets,
3208 if (target_node == NUMA_NO_NODE)
3211 if (best_distance == -1)
3212 best_distance = node_distance(node, target_node);
3215 * Visit targets from this pass in the next pass.
3216 * Eventually, every node will have been part of
3217 * a pass, and will become set in 'used_targets'.
3219 node_set(target_node, next_pass);
3223 * 'next_pass' contains nodes which became migration
3224 * targets in this pass. Make additional passes until
3225 * no more migrations targets are available.
3227 if (!nodes_empty(next_pass))
3232 * For callers that do not hold get_online_mems() already.
3234 static void set_migration_target_nodes(void)
3237 __set_migration_target_nodes();
3242 * This leaves migrate-on-reclaim transiently disabled between
3243 * the MEM_GOING_OFFLINE and MEM_OFFLINE events. This runs
3244 * whether reclaim-based migration is enabled or not, which
3245 * ensures that the user can turn reclaim-based migration at
3246 * any time without needing to recalculate migration targets.
3248 * These callbacks already hold get_online_mems(). That is why
3249 * __set_migration_target_nodes() can be used as opposed to
3250 * set_migration_target_nodes().
3252 static int __meminit migrate_on_reclaim_callback(struct notifier_block *self,
3253 unsigned long action, void *_arg)
3255 struct memory_notify *arg = _arg;
3258 * Only update the node migration order when a node is
3259 * changing status, like online->offline. This avoids
3260 * the overhead of synchronize_rcu() in most cases.
3262 if (arg->status_change_nid < 0)
3263 return notifier_from_errno(0);
3266 case MEM_GOING_OFFLINE:
3268 * Make sure there are not transient states where
3269 * an offline node is a migration target. This
3270 * will leave migration disabled until the offline
3271 * completes and the MEM_OFFLINE case below runs.
3273 disable_all_migrate_targets();
3278 * Recalculate the target nodes once the node
3279 * reaches its final state (online or offline).
3281 __set_migration_target_nodes();
3283 case MEM_CANCEL_OFFLINE:
3285 * MEM_GOING_OFFLINE disabled all the migration
3286 * targets. Reenable them.
3288 __set_migration_target_nodes();
3290 case MEM_GOING_ONLINE:
3291 case MEM_CANCEL_ONLINE:
3295 return notifier_from_errno(0);
3299 * React to hotplug events that might affect the migration targets
3300 * like events that online or offline NUMA nodes.
3302 * The ordering is also currently dependent on which nodes have
3303 * CPUs. That means we need CPU on/offline notification too.
3305 static int migration_online_cpu(unsigned int cpu)
3307 set_migration_target_nodes();
3311 static int migration_offline_cpu(unsigned int cpu)
3313 set_migration_target_nodes();
3317 static int __init migrate_on_reclaim_init(void)
3321 node_demotion = kmalloc_array(nr_node_ids,
3322 sizeof(struct demotion_nodes),
3324 WARN_ON(!node_demotion);
3326 ret = cpuhp_setup_state_nocalls(CPUHP_MM_DEMOTION_DEAD, "mm/demotion:offline",
3327 NULL, migration_offline_cpu);
3329 * In the unlikely case that this fails, the automatic
3330 * migration targets may become suboptimal for nodes
3331 * where N_CPU changes. With such a small impact in a
3332 * rare case, do not bother trying to do anything special.
3335 ret = cpuhp_setup_state(CPUHP_AP_MM_DEMOTION_ONLINE, "mm/demotion:online",
3336 migration_online_cpu, NULL);
3339 hotplug_memory_notifier(migrate_on_reclaim_callback, 100);
3342 late_initcall(migrate_on_reclaim_init);
3343 #endif /* CONFIG_HOTPLUG_CPU */
3345 bool numa_demotion_enabled = false;
3348 static ssize_t numa_demotion_enabled_show(struct kobject *kobj,
3349 struct kobj_attribute *attr, char *buf)
3351 return sysfs_emit(buf, "%s\n",
3352 numa_demotion_enabled ? "true" : "false");
3355 static ssize_t numa_demotion_enabled_store(struct kobject *kobj,
3356 struct kobj_attribute *attr,
3357 const char *buf, size_t count)
3359 if (!strncmp(buf, "true", 4) || !strncmp(buf, "1", 1))
3360 numa_demotion_enabled = true;
3361 else if (!strncmp(buf, "false", 5) || !strncmp(buf, "0", 1))
3362 numa_demotion_enabled = false;
3369 static struct kobj_attribute numa_demotion_enabled_attr =
3370 __ATTR(demotion_enabled, 0644, numa_demotion_enabled_show,
3371 numa_demotion_enabled_store);
3373 static struct attribute *numa_attrs[] = {
3374 &numa_demotion_enabled_attr.attr,
3378 static const struct attribute_group numa_attr_group = {
3379 .attrs = numa_attrs,
3382 static int __init numa_init_sysfs(void)
3385 struct kobject *numa_kobj;
3387 numa_kobj = kobject_create_and_add("numa", mm_kobj);
3389 pr_err("failed to create numa kobject\n");
3392 err = sysfs_create_group(numa_kobj, &numa_attr_group);
3394 pr_err("failed to register numa group\n");
3400 kobject_put(numa_kobj);
3403 subsys_initcall(numa_init_sysfs);