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>
53 #include <asm/tlbflush.h>
55 #define CREATE_TRACE_POINTS
56 #include <trace/events/migrate.h>
61 * migrate_prep() needs to be called before we start compiling a list of pages
62 * to be migrated using isolate_lru_page(). If scheduling work on other CPUs is
63 * undesirable, use migrate_prep_local()
65 int migrate_prep(void)
68 * Clear the LRU lists so pages can be isolated.
69 * Note that pages may be moved off the LRU after we have
70 * drained them. Those pages will fail to migrate like other
71 * pages that may be busy.
78 /* Do the necessary work of migrate_prep but not if it involves other CPUs */
79 int migrate_prep_local(void)
86 int isolate_movable_page(struct page *page, isolate_mode_t mode)
88 struct address_space *mapping;
91 * Avoid burning cycles with pages that are yet under __free_pages(),
92 * or just got freed under us.
94 * In case we 'win' a race for a movable page being freed under us and
95 * raise its refcount preventing __free_pages() from doing its job
96 * the put_page() at the end of this block will take care of
97 * release this page, thus avoiding a nasty leakage.
99 if (unlikely(!get_page_unless_zero(page)))
103 * Check PageMovable before holding a PG_lock because page's owner
104 * assumes anybody doesn't touch PG_lock of newly allocated page
105 * so unconditionally grabbing the lock ruins page's owner side.
107 if (unlikely(!__PageMovable(page)))
110 * As movable pages are not isolated from LRU lists, concurrent
111 * compaction threads can race against page migration functions
112 * as well as race against the releasing a page.
114 * In order to avoid having an already isolated movable page
115 * being (wrongly) re-isolated while it is under migration,
116 * or to avoid attempting to isolate pages being released,
117 * lets be sure we have the page lock
118 * before proceeding with the movable page isolation steps.
120 if (unlikely(!trylock_page(page)))
123 if (!PageMovable(page) || PageIsolated(page))
124 goto out_no_isolated;
126 mapping = page_mapping(page);
127 VM_BUG_ON_PAGE(!mapping, page);
129 if (!mapping->a_ops->isolate_page(page, mode))
130 goto out_no_isolated;
132 /* Driver shouldn't use PG_isolated bit of page->flags */
133 WARN_ON_ONCE(PageIsolated(page));
134 __SetPageIsolated(page);
147 /* It should be called on page which is PG_movable */
148 void putback_movable_page(struct page *page)
150 struct address_space *mapping;
152 VM_BUG_ON_PAGE(!PageLocked(page), page);
153 VM_BUG_ON_PAGE(!PageMovable(page), page);
154 VM_BUG_ON_PAGE(!PageIsolated(page), page);
156 mapping = page_mapping(page);
157 mapping->a_ops->putback_page(page);
158 __ClearPageIsolated(page);
162 * Put previously isolated pages back onto the appropriate lists
163 * from where they were once taken off for compaction/migration.
165 * This function shall be used whenever the isolated pageset has been
166 * built from lru, balloon, hugetlbfs page. See isolate_migratepages_range()
167 * and isolate_huge_page().
169 void putback_movable_pages(struct list_head *l)
174 list_for_each_entry_safe(page, page2, l, lru) {
175 if (unlikely(PageHuge(page))) {
176 putback_active_hugepage(page);
179 list_del(&page->lru);
181 * We isolated non-lru movable page so here we can use
182 * __PageMovable because LRU page's mapping cannot have
183 * PAGE_MAPPING_MOVABLE.
185 if (unlikely(__PageMovable(page))) {
186 VM_BUG_ON_PAGE(!PageIsolated(page), page);
188 if (PageMovable(page))
189 putback_movable_page(page);
191 __ClearPageIsolated(page);
195 mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON +
196 page_is_file_lru(page), -thp_nr_pages(page));
197 putback_lru_page(page);
203 * Restore a potential migration pte to a working pte entry
205 static bool remove_migration_pte(struct page *page, struct vm_area_struct *vma,
206 unsigned long addr, void *old)
208 struct page_vma_mapped_walk pvmw = {
212 .flags = PVMW_SYNC | PVMW_MIGRATION,
218 VM_BUG_ON_PAGE(PageTail(page), page);
219 while (page_vma_mapped_walk(&pvmw)) {
223 new = page - pvmw.page->index +
224 linear_page_index(vma, pvmw.address);
226 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
227 /* PMD-mapped THP migration entry */
229 VM_BUG_ON_PAGE(PageHuge(page) || !PageTransCompound(page), page);
230 remove_migration_pmd(&pvmw, new);
236 pte = pte_mkold(mk_pte(new, READ_ONCE(vma->vm_page_prot)));
237 if (pte_swp_soft_dirty(*pvmw.pte))
238 pte = pte_mksoft_dirty(pte);
241 * Recheck VMA as permissions can change since migration started
243 entry = pte_to_swp_entry(*pvmw.pte);
244 if (is_write_migration_entry(entry))
245 pte = maybe_mkwrite(pte, vma);
246 else if (pte_swp_uffd_wp(*pvmw.pte))
247 pte = pte_mkuffd_wp(pte);
249 if (unlikely(is_device_private_page(new))) {
250 entry = make_device_private_entry(new, pte_write(pte));
251 pte = swp_entry_to_pte(entry);
252 if (pte_swp_soft_dirty(*pvmw.pte))
253 pte = pte_swp_mksoft_dirty(pte);
254 if (pte_swp_uffd_wp(*pvmw.pte))
255 pte = pte_swp_mkuffd_wp(pte);
258 #ifdef CONFIG_HUGETLB_PAGE
260 pte = pte_mkhuge(pte);
261 pte = arch_make_huge_pte(pte, vma, new, 0);
262 set_huge_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte);
264 hugepage_add_anon_rmap(new, vma, pvmw.address);
266 page_dup_rmap(new, true);
270 set_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte);
273 page_add_anon_rmap(new, vma, pvmw.address, false);
275 page_add_file_rmap(new, false);
277 if (vma->vm_flags & VM_LOCKED && !PageTransCompound(new))
280 if (PageTransHuge(page) && PageMlocked(page))
281 clear_page_mlock(page);
283 /* No need to invalidate - it was non-present before */
284 update_mmu_cache(vma, pvmw.address, pvmw.pte);
291 * Get rid of all migration entries and replace them by
292 * references to the indicated page.
294 void remove_migration_ptes(struct page *old, struct page *new, bool locked)
296 struct rmap_walk_control rwc = {
297 .rmap_one = remove_migration_pte,
302 rmap_walk_locked(new, &rwc);
304 rmap_walk(new, &rwc);
308 * Something used the pte of a page under migration. We need to
309 * get to the page and wait until migration is finished.
310 * When we return from this function the fault will be retried.
312 void __migration_entry_wait(struct mm_struct *mm, pte_t *ptep,
321 if (!is_swap_pte(pte))
324 entry = pte_to_swp_entry(pte);
325 if (!is_migration_entry(entry))
328 page = migration_entry_to_page(entry);
331 * Once page cache replacement of page migration started, page_count
332 * is zero; but we must not call put_and_wait_on_page_locked() without
333 * a ref. Use get_page_unless_zero(), and just fault again if it fails.
335 if (!get_page_unless_zero(page))
337 pte_unmap_unlock(ptep, ptl);
338 put_and_wait_on_page_locked(page);
341 pte_unmap_unlock(ptep, ptl);
344 void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
345 unsigned long address)
347 spinlock_t *ptl = pte_lockptr(mm, pmd);
348 pte_t *ptep = pte_offset_map(pmd, address);
349 __migration_entry_wait(mm, ptep, ptl);
352 void migration_entry_wait_huge(struct vm_area_struct *vma,
353 struct mm_struct *mm, pte_t *pte)
355 spinlock_t *ptl = huge_pte_lockptr(hstate_vma(vma), mm, pte);
356 __migration_entry_wait(mm, pte, ptl);
359 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
360 void pmd_migration_entry_wait(struct mm_struct *mm, pmd_t *pmd)
365 ptl = pmd_lock(mm, pmd);
366 if (!is_pmd_migration_entry(*pmd))
368 page = migration_entry_to_page(pmd_to_swp_entry(*pmd));
369 if (!get_page_unless_zero(page))
372 put_and_wait_on_page_locked(page);
379 static int expected_page_refs(struct address_space *mapping, struct page *page)
381 int expected_count = 1;
384 * Device public or private pages have an extra refcount as they are
387 expected_count += is_device_private_page(page);
389 expected_count += thp_nr_pages(page) + page_has_private(page);
391 return expected_count;
395 * Replace the page in the mapping.
397 * The number of remaining references must be:
398 * 1 for anonymous pages without a mapping
399 * 2 for pages with a mapping
400 * 3 for pages with a mapping and PagePrivate/PagePrivate2 set.
402 int migrate_page_move_mapping(struct address_space *mapping,
403 struct page *newpage, struct page *page, int extra_count)
405 XA_STATE(xas, &mapping->i_pages, page_index(page));
406 struct zone *oldzone, *newzone;
408 int expected_count = expected_page_refs(mapping, page) + extra_count;
411 /* Anonymous page without mapping */
412 if (page_count(page) != expected_count)
415 /* No turning back from here */
416 newpage->index = page->index;
417 newpage->mapping = page->mapping;
418 if (PageSwapBacked(page))
419 __SetPageSwapBacked(newpage);
421 return MIGRATEPAGE_SUCCESS;
424 oldzone = page_zone(page);
425 newzone = page_zone(newpage);
428 if (page_count(page) != expected_count || xas_load(&xas) != page) {
429 xas_unlock_irq(&xas);
433 if (!page_ref_freeze(page, expected_count)) {
434 xas_unlock_irq(&xas);
439 * Now we know that no one else is looking at the page:
440 * no turning back from here.
442 newpage->index = page->index;
443 newpage->mapping = page->mapping;
444 page_ref_add(newpage, thp_nr_pages(page)); /* add cache reference */
445 if (PageSwapBacked(page)) {
446 __SetPageSwapBacked(newpage);
447 if (PageSwapCache(page)) {
448 SetPageSwapCache(newpage);
449 set_page_private(newpage, page_private(page));
452 VM_BUG_ON_PAGE(PageSwapCache(page), page);
455 /* Move dirty while page refs frozen and newpage not yet exposed */
456 dirty = PageDirty(page);
458 ClearPageDirty(page);
459 SetPageDirty(newpage);
462 xas_store(&xas, newpage);
463 if (PageTransHuge(page)) {
466 for (i = 1; i < HPAGE_PMD_NR; i++) {
468 xas_store(&xas, newpage);
473 * Drop cache reference from old page by unfreezing
474 * to one less reference.
475 * We know this isn't the last reference.
477 page_ref_unfreeze(page, expected_count - thp_nr_pages(page));
480 /* Leave irq disabled to prevent preemption while updating stats */
483 * If moved to a different zone then also account
484 * the page for that zone. Other VM counters will be
485 * taken care of when we establish references to the
486 * new page and drop references to the old page.
488 * Note that anonymous pages are accounted for
489 * via NR_FILE_PAGES and NR_ANON_MAPPED if they
490 * are mapped to swap space.
492 if (newzone != oldzone) {
493 struct lruvec *old_lruvec, *new_lruvec;
494 struct mem_cgroup *memcg;
496 memcg = page_memcg(page);
497 old_lruvec = mem_cgroup_lruvec(memcg, oldzone->zone_pgdat);
498 new_lruvec = mem_cgroup_lruvec(memcg, newzone->zone_pgdat);
500 __dec_lruvec_state(old_lruvec, NR_FILE_PAGES);
501 __inc_lruvec_state(new_lruvec, NR_FILE_PAGES);
502 if (PageSwapBacked(page) && !PageSwapCache(page)) {
503 __dec_lruvec_state(old_lruvec, NR_SHMEM);
504 __inc_lruvec_state(new_lruvec, NR_SHMEM);
506 if (dirty && mapping_cap_account_dirty(mapping)) {
507 __dec_node_state(oldzone->zone_pgdat, NR_FILE_DIRTY);
508 __dec_zone_state(oldzone, NR_ZONE_WRITE_PENDING);
509 __inc_node_state(newzone->zone_pgdat, NR_FILE_DIRTY);
510 __inc_zone_state(newzone, NR_ZONE_WRITE_PENDING);
515 return MIGRATEPAGE_SUCCESS;
517 EXPORT_SYMBOL(migrate_page_move_mapping);
520 * The expected number of remaining references is the same as that
521 * of migrate_page_move_mapping().
523 int migrate_huge_page_move_mapping(struct address_space *mapping,
524 struct page *newpage, struct page *page)
526 XA_STATE(xas, &mapping->i_pages, page_index(page));
530 expected_count = 2 + page_has_private(page);
531 if (page_count(page) != expected_count || xas_load(&xas) != page) {
532 xas_unlock_irq(&xas);
536 if (!page_ref_freeze(page, expected_count)) {
537 xas_unlock_irq(&xas);
541 newpage->index = page->index;
542 newpage->mapping = page->mapping;
546 xas_store(&xas, newpage);
548 page_ref_unfreeze(page, expected_count - 1);
550 xas_unlock_irq(&xas);
552 return MIGRATEPAGE_SUCCESS;
556 * Gigantic pages are so large that we do not guarantee that page++ pointer
557 * arithmetic will work across the entire page. We need something more
560 static void __copy_gigantic_page(struct page *dst, struct page *src,
564 struct page *dst_base = dst;
565 struct page *src_base = src;
567 for (i = 0; i < nr_pages; ) {
569 copy_highpage(dst, src);
572 dst = mem_map_next(dst, dst_base, i);
573 src = mem_map_next(src, src_base, i);
577 static void copy_huge_page(struct page *dst, struct page *src)
584 struct hstate *h = page_hstate(src);
585 nr_pages = pages_per_huge_page(h);
587 if (unlikely(nr_pages > MAX_ORDER_NR_PAGES)) {
588 __copy_gigantic_page(dst, src, nr_pages);
593 BUG_ON(!PageTransHuge(src));
594 nr_pages = thp_nr_pages(src);
597 for (i = 0; i < nr_pages; i++) {
599 copy_highpage(dst + i, src + i);
604 * Copy the page to its new location
606 void migrate_page_states(struct page *newpage, struct page *page)
611 SetPageError(newpage);
612 if (PageReferenced(page))
613 SetPageReferenced(newpage);
614 if (PageUptodate(page))
615 SetPageUptodate(newpage);
616 if (TestClearPageActive(page)) {
617 VM_BUG_ON_PAGE(PageUnevictable(page), page);
618 SetPageActive(newpage);
619 } else if (TestClearPageUnevictable(page))
620 SetPageUnevictable(newpage);
621 if (PageWorkingset(page))
622 SetPageWorkingset(newpage);
623 if (PageChecked(page))
624 SetPageChecked(newpage);
625 if (PageMappedToDisk(page))
626 SetPageMappedToDisk(newpage);
628 /* Move dirty on pages not done by migrate_page_move_mapping() */
630 SetPageDirty(newpage);
632 if (page_is_young(page))
633 set_page_young(newpage);
634 if (page_is_idle(page))
635 set_page_idle(newpage);
638 * Copy NUMA information to the new page, to prevent over-eager
639 * future migrations of this same page.
641 cpupid = page_cpupid_xchg_last(page, -1);
642 page_cpupid_xchg_last(newpage, cpupid);
644 ksm_migrate_page(newpage, page);
646 * Please do not reorder this without considering how mm/ksm.c's
647 * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache().
649 if (PageSwapCache(page))
650 ClearPageSwapCache(page);
651 ClearPagePrivate(page);
652 set_page_private(page, 0);
655 * If any waiters have accumulated on the new page then
658 if (PageWriteback(newpage))
659 end_page_writeback(newpage);
662 * PG_readahead shares the same bit with PG_reclaim. The above
663 * end_page_writeback() may clear PG_readahead mistakenly, so set the
666 if (PageReadahead(page))
667 SetPageReadahead(newpage);
669 copy_page_owner(page, newpage);
672 mem_cgroup_migrate(page, newpage);
674 EXPORT_SYMBOL(migrate_page_states);
676 void migrate_page_copy(struct page *newpage, struct page *page)
678 if (PageHuge(page) || PageTransHuge(page))
679 copy_huge_page(newpage, page);
681 copy_highpage(newpage, page);
683 migrate_page_states(newpage, page);
685 EXPORT_SYMBOL(migrate_page_copy);
687 /************************************************************
688 * Migration functions
689 ***********************************************************/
692 * Common logic to directly migrate a single LRU page suitable for
693 * pages that do not use PagePrivate/PagePrivate2.
695 * Pages are locked upon entry and exit.
697 int migrate_page(struct address_space *mapping,
698 struct page *newpage, struct page *page,
699 enum migrate_mode mode)
703 BUG_ON(PageWriteback(page)); /* Writeback must be complete */
705 rc = migrate_page_move_mapping(mapping, newpage, page, 0);
707 if (rc != MIGRATEPAGE_SUCCESS)
710 if (mode != MIGRATE_SYNC_NO_COPY)
711 migrate_page_copy(newpage, page);
713 migrate_page_states(newpage, page);
714 return MIGRATEPAGE_SUCCESS;
716 EXPORT_SYMBOL(migrate_page);
719 /* Returns true if all buffers are successfully locked */
720 static bool buffer_migrate_lock_buffers(struct buffer_head *head,
721 enum migrate_mode mode)
723 struct buffer_head *bh = head;
725 /* Simple case, sync compaction */
726 if (mode != MIGRATE_ASYNC) {
729 bh = bh->b_this_page;
731 } while (bh != head);
736 /* async case, we cannot block on lock_buffer so use trylock_buffer */
738 if (!trylock_buffer(bh)) {
740 * We failed to lock the buffer and cannot stall in
741 * async migration. Release the taken locks
743 struct buffer_head *failed_bh = bh;
745 while (bh != failed_bh) {
747 bh = bh->b_this_page;
752 bh = bh->b_this_page;
753 } while (bh != head);
757 static int __buffer_migrate_page(struct address_space *mapping,
758 struct page *newpage, struct page *page, enum migrate_mode mode,
761 struct buffer_head *bh, *head;
765 if (!page_has_buffers(page))
766 return migrate_page(mapping, newpage, page, mode);
768 /* Check whether page does not have extra refs before we do more work */
769 expected_count = expected_page_refs(mapping, page);
770 if (page_count(page) != expected_count)
773 head = page_buffers(page);
774 if (!buffer_migrate_lock_buffers(head, mode))
779 bool invalidated = false;
783 spin_lock(&mapping->private_lock);
786 if (atomic_read(&bh->b_count)) {
790 bh = bh->b_this_page;
791 } while (bh != head);
797 spin_unlock(&mapping->private_lock);
798 invalidate_bh_lrus();
800 goto recheck_buffers;
804 rc = migrate_page_move_mapping(mapping, newpage, page, 0);
805 if (rc != MIGRATEPAGE_SUCCESS)
808 attach_page_private(newpage, detach_page_private(page));
812 set_bh_page(bh, newpage, bh_offset(bh));
813 bh = bh->b_this_page;
815 } while (bh != head);
817 if (mode != MIGRATE_SYNC_NO_COPY)
818 migrate_page_copy(newpage, page);
820 migrate_page_states(newpage, page);
822 rc = MIGRATEPAGE_SUCCESS;
825 spin_unlock(&mapping->private_lock);
829 bh = bh->b_this_page;
831 } while (bh != head);
837 * Migration function for pages with buffers. This function can only be used
838 * if the underlying filesystem guarantees that no other references to "page"
839 * exist. For example attached buffer heads are accessed only under page lock.
841 int buffer_migrate_page(struct address_space *mapping,
842 struct page *newpage, struct page *page, enum migrate_mode mode)
844 return __buffer_migrate_page(mapping, newpage, page, mode, false);
846 EXPORT_SYMBOL(buffer_migrate_page);
849 * Same as above except that this variant is more careful and checks that there
850 * are also no buffer head references. This function is the right one for
851 * mappings where buffer heads are directly looked up and referenced (such as
852 * block device mappings).
854 int buffer_migrate_page_norefs(struct address_space *mapping,
855 struct page *newpage, struct page *page, enum migrate_mode mode)
857 return __buffer_migrate_page(mapping, newpage, page, mode, true);
862 * Writeback a page to clean the dirty state
864 static int writeout(struct address_space *mapping, struct page *page)
866 struct writeback_control wbc = {
867 .sync_mode = WB_SYNC_NONE,
870 .range_end = LLONG_MAX,
875 if (!mapping->a_ops->writepage)
876 /* No write method for the address space */
879 if (!clear_page_dirty_for_io(page))
880 /* Someone else already triggered a write */
884 * A dirty page may imply that the underlying filesystem has
885 * the page on some queue. So the page must be clean for
886 * migration. Writeout may mean we loose the lock and the
887 * page state is no longer what we checked for earlier.
888 * At this point we know that the migration attempt cannot
891 remove_migration_ptes(page, page, false);
893 rc = mapping->a_ops->writepage(page, &wbc);
895 if (rc != AOP_WRITEPAGE_ACTIVATE)
896 /* unlocked. Relock */
899 return (rc < 0) ? -EIO : -EAGAIN;
903 * Default handling if a filesystem does not provide a migration function.
905 static int fallback_migrate_page(struct address_space *mapping,
906 struct page *newpage, struct page *page, enum migrate_mode mode)
908 if (PageDirty(page)) {
909 /* Only writeback pages in full synchronous migration */
912 case MIGRATE_SYNC_NO_COPY:
917 return writeout(mapping, page);
921 * Buffers may be managed in a filesystem specific way.
922 * We must have no buffers or drop them.
924 if (page_has_private(page) &&
925 !try_to_release_page(page, GFP_KERNEL))
926 return mode == MIGRATE_SYNC ? -EAGAIN : -EBUSY;
928 return migrate_page(mapping, newpage, page, mode);
932 * Move a page to a newly allocated page
933 * The page is locked and all ptes have been successfully removed.
935 * The new page will have replaced the old page if this function
940 * MIGRATEPAGE_SUCCESS - success
942 static int move_to_new_page(struct page *newpage, struct page *page,
943 enum migrate_mode mode)
945 struct address_space *mapping;
947 bool is_lru = !__PageMovable(page);
949 VM_BUG_ON_PAGE(!PageLocked(page), page);
950 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
952 mapping = page_mapping(page);
954 if (likely(is_lru)) {
956 rc = migrate_page(mapping, newpage, page, mode);
957 else if (mapping->a_ops->migratepage)
959 * Most pages have a mapping and most filesystems
960 * provide a migratepage callback. Anonymous pages
961 * are part of swap space which also has its own
962 * migratepage callback. This is the most common path
963 * for page migration.
965 rc = mapping->a_ops->migratepage(mapping, newpage,
968 rc = fallback_migrate_page(mapping, newpage,
972 * In case of non-lru page, it could be released after
973 * isolation step. In that case, we shouldn't try migration.
975 VM_BUG_ON_PAGE(!PageIsolated(page), page);
976 if (!PageMovable(page)) {
977 rc = MIGRATEPAGE_SUCCESS;
978 __ClearPageIsolated(page);
982 rc = mapping->a_ops->migratepage(mapping, newpage,
984 WARN_ON_ONCE(rc == MIGRATEPAGE_SUCCESS &&
985 !PageIsolated(page));
989 * When successful, old pagecache page->mapping must be cleared before
990 * page is freed; but stats require that PageAnon be left as PageAnon.
992 if (rc == MIGRATEPAGE_SUCCESS) {
993 if (__PageMovable(page)) {
994 VM_BUG_ON_PAGE(!PageIsolated(page), page);
997 * We clear PG_movable under page_lock so any compactor
998 * cannot try to migrate this page.
1000 __ClearPageIsolated(page);
1004 * Anonymous and movable page->mapping will be cleared by
1005 * free_pages_prepare so don't reset it here for keeping
1006 * the type to work PageAnon, for example.
1008 if (!PageMappingFlags(page))
1009 page->mapping = NULL;
1011 if (likely(!is_zone_device_page(newpage)))
1012 flush_dcache_page(newpage);
1019 static int __unmap_and_move(struct page *page, struct page *newpage,
1020 int force, enum migrate_mode mode)
1023 int page_was_mapped = 0;
1024 struct anon_vma *anon_vma = NULL;
1025 bool is_lru = !__PageMovable(page);
1027 if (!trylock_page(page)) {
1028 if (!force || mode == MIGRATE_ASYNC)
1032 * It's not safe for direct compaction to call lock_page.
1033 * For example, during page readahead pages are added locked
1034 * to the LRU. Later, when the IO completes the pages are
1035 * marked uptodate and unlocked. However, the queueing
1036 * could be merging multiple pages for one bio (e.g.
1037 * mpage_readahead). If an allocation happens for the
1038 * second or third page, the process can end up locking
1039 * the same page twice and deadlocking. Rather than
1040 * trying to be clever about what pages can be locked,
1041 * avoid the use of lock_page for direct compaction
1044 if (current->flags & PF_MEMALLOC)
1050 if (PageWriteback(page)) {
1052 * Only in the case of a full synchronous migration is it
1053 * necessary to wait for PageWriteback. In the async case,
1054 * the retry loop is too short and in the sync-light case,
1055 * the overhead of stalling is too much
1059 case MIGRATE_SYNC_NO_COPY:
1067 wait_on_page_writeback(page);
1071 * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
1072 * we cannot notice that anon_vma is freed while we migrates a page.
1073 * This get_anon_vma() delays freeing anon_vma pointer until the end
1074 * of migration. File cache pages are no problem because of page_lock()
1075 * File Caches may use write_page() or lock_page() in migration, then,
1076 * just care Anon page here.
1078 * Only page_get_anon_vma() understands the subtleties of
1079 * getting a hold on an anon_vma from outside one of its mms.
1080 * But if we cannot get anon_vma, then we won't need it anyway,
1081 * because that implies that the anon page is no longer mapped
1082 * (and cannot be remapped so long as we hold the page lock).
1084 if (PageAnon(page) && !PageKsm(page))
1085 anon_vma = page_get_anon_vma(page);
1088 * Block others from accessing the new page when we get around to
1089 * establishing additional references. We are usually the only one
1090 * holding a reference to newpage at this point. We used to have a BUG
1091 * here if trylock_page(newpage) fails, but would like to allow for
1092 * cases where there might be a race with the previous use of newpage.
1093 * This is much like races on refcount of oldpage: just don't BUG().
1095 if (unlikely(!trylock_page(newpage)))
1098 if (unlikely(!is_lru)) {
1099 rc = move_to_new_page(newpage, page, mode);
1100 goto out_unlock_both;
1104 * Corner case handling:
1105 * 1. When a new swap-cache page is read into, it is added to the LRU
1106 * and treated as swapcache but it has no rmap yet.
1107 * Calling try_to_unmap() against a page->mapping==NULL page will
1108 * trigger a BUG. So handle it here.
1109 * 2. An orphaned page (see truncate_complete_page) might have
1110 * fs-private metadata. The page can be picked up due to memory
1111 * offlining. Everywhere else except page reclaim, the page is
1112 * invisible to the vm, so the page can not be migrated. So try to
1113 * free the metadata, so the page can be freed.
1115 if (!page->mapping) {
1116 VM_BUG_ON_PAGE(PageAnon(page), page);
1117 if (page_has_private(page)) {
1118 try_to_free_buffers(page);
1119 goto out_unlock_both;
1121 } else if (page_mapped(page)) {
1122 /* Establish migration ptes */
1123 VM_BUG_ON_PAGE(PageAnon(page) && !PageKsm(page) && !anon_vma,
1126 TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
1127 page_was_mapped = 1;
1130 if (!page_mapped(page))
1131 rc = move_to_new_page(newpage, page, mode);
1133 if (page_was_mapped)
1134 remove_migration_ptes(page,
1135 rc == MIGRATEPAGE_SUCCESS ? newpage : page, false);
1138 unlock_page(newpage);
1140 /* Drop an anon_vma reference if we took one */
1142 put_anon_vma(anon_vma);
1146 * If migration is successful, decrease refcount of the newpage
1147 * which will not free the page because new page owner increased
1148 * refcounter. As well, if it is LRU page, add the page to LRU
1149 * list in here. Use the old state of the isolated source page to
1150 * determine if we migrated a LRU page. newpage was already unlocked
1151 * and possibly modified by its owner - don't rely on the page
1154 if (rc == MIGRATEPAGE_SUCCESS) {
1155 if (unlikely(!is_lru))
1158 putback_lru_page(newpage);
1165 * Obtain the lock on page, remove all ptes and migrate the page
1166 * to the newly allocated page in newpage.
1168 static int unmap_and_move(new_page_t get_new_page,
1169 free_page_t put_new_page,
1170 unsigned long private, struct page *page,
1171 int force, enum migrate_mode mode,
1172 enum migrate_reason reason)
1174 int rc = MIGRATEPAGE_SUCCESS;
1175 struct page *newpage = NULL;
1177 if (!thp_migration_supported() && PageTransHuge(page))
1180 if (page_count(page) == 1) {
1181 /* page was freed from under us. So we are done. */
1182 ClearPageActive(page);
1183 ClearPageUnevictable(page);
1184 if (unlikely(__PageMovable(page))) {
1186 if (!PageMovable(page))
1187 __ClearPageIsolated(page);
1193 newpage = get_new_page(page, private);
1197 rc = __unmap_and_move(page, newpage, force, mode);
1198 if (rc == MIGRATEPAGE_SUCCESS)
1199 set_page_owner_migrate_reason(newpage, reason);
1202 if (rc != -EAGAIN) {
1204 * A page that has been migrated has all references
1205 * removed and will be freed. A page that has not been
1206 * migrated will have kept its references and be restored.
1208 list_del(&page->lru);
1211 * Compaction can migrate also non-LRU pages which are
1212 * not accounted to NR_ISOLATED_*. They can be recognized
1215 if (likely(!__PageMovable(page)))
1216 mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON +
1217 page_is_file_lru(page), -thp_nr_pages(page));
1221 * If migration is successful, releases reference grabbed during
1222 * isolation. Otherwise, restore the page to right list unless
1225 if (rc == MIGRATEPAGE_SUCCESS) {
1227 if (reason == MR_MEMORY_FAILURE) {
1229 * Set PG_HWPoison on just freed page
1230 * intentionally. Although it's rather weird,
1231 * it's how HWPoison flag works at the moment.
1233 if (set_hwpoison_free_buddy_page(page))
1234 num_poisoned_pages_inc();
1237 if (rc != -EAGAIN) {
1238 if (likely(!__PageMovable(page))) {
1239 putback_lru_page(page);
1244 if (PageMovable(page))
1245 putback_movable_page(page);
1247 __ClearPageIsolated(page);
1253 put_new_page(newpage, private);
1262 * Counterpart of unmap_and_move_page() for hugepage migration.
1264 * This function doesn't wait the completion of hugepage I/O
1265 * because there is no race between I/O and migration for hugepage.
1266 * Note that currently hugepage I/O occurs only in direct I/O
1267 * where no lock is held and PG_writeback is irrelevant,
1268 * and writeback status of all subpages are counted in the reference
1269 * count of the head page (i.e. if all subpages of a 2MB hugepage are
1270 * under direct I/O, the reference of the head page is 512 and a bit more.)
1271 * This means that when we try to migrate hugepage whose subpages are
1272 * doing direct I/O, some references remain after try_to_unmap() and
1273 * hugepage migration fails without data corruption.
1275 * There is also no race when direct I/O is issued on the page under migration,
1276 * because then pte is replaced with migration swap entry and direct I/O code
1277 * will wait in the page fault for migration to complete.
1279 static int unmap_and_move_huge_page(new_page_t get_new_page,
1280 free_page_t put_new_page, unsigned long private,
1281 struct page *hpage, int force,
1282 enum migrate_mode mode, int reason)
1285 int page_was_mapped = 0;
1286 struct page *new_hpage;
1287 struct anon_vma *anon_vma = NULL;
1288 struct address_space *mapping = NULL;
1291 * Migratability of hugepages depends on architectures and their size.
1292 * This check is necessary because some callers of hugepage migration
1293 * like soft offline and memory hotremove don't walk through page
1294 * tables or check whether the hugepage is pmd-based or not before
1295 * kicking migration.
1297 if (!hugepage_migration_supported(page_hstate(hpage))) {
1298 putback_active_hugepage(hpage);
1302 new_hpage = get_new_page(hpage, private);
1306 if (!trylock_page(hpage)) {
1311 case MIGRATE_SYNC_NO_COPY:
1320 * Check for pages which are in the process of being freed. Without
1321 * page_mapping() set, hugetlbfs specific move page routine will not
1322 * be called and we could leak usage counts for subpools.
1324 if (page_private(hpage) && !page_mapping(hpage)) {
1329 if (PageAnon(hpage))
1330 anon_vma = page_get_anon_vma(hpage);
1332 if (unlikely(!trylock_page(new_hpage)))
1335 if (page_mapped(hpage)) {
1337 * try_to_unmap could potentially call huge_pmd_unshare.
1338 * Because of this, take semaphore in write mode here and
1339 * set TTU_RMAP_LOCKED to let lower levels know we have
1342 mapping = hugetlb_page_mapping_lock_write(hpage);
1343 if (unlikely(!mapping))
1344 goto unlock_put_anon;
1347 TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS|
1349 page_was_mapped = 1;
1351 * Leave mapping locked until after subsequent call to
1352 * remove_migration_ptes()
1356 if (!page_mapped(hpage))
1357 rc = move_to_new_page(new_hpage, hpage, mode);
1359 if (page_was_mapped) {
1360 remove_migration_ptes(hpage,
1361 rc == MIGRATEPAGE_SUCCESS ? new_hpage : hpage, true);
1362 i_mmap_unlock_write(mapping);
1366 unlock_page(new_hpage);
1370 put_anon_vma(anon_vma);
1372 if (rc == MIGRATEPAGE_SUCCESS) {
1373 move_hugetlb_state(hpage, new_hpage, reason);
1374 put_new_page = NULL;
1381 putback_active_hugepage(hpage);
1384 * If migration was not successful and there's a freeing callback, use
1385 * it. Otherwise, put_page() will drop the reference grabbed during
1389 put_new_page(new_hpage, private);
1391 putback_active_hugepage(new_hpage);
1397 * migrate_pages - migrate the pages specified in a list, to the free pages
1398 * supplied as the target for the page migration
1400 * @from: The list of pages to be migrated.
1401 * @get_new_page: The function used to allocate free pages to be used
1402 * as the target of the page migration.
1403 * @put_new_page: The function used to free target pages if migration
1404 * fails, or NULL if no special handling is necessary.
1405 * @private: Private data to be passed on to get_new_page()
1406 * @mode: The migration mode that specifies the constraints for
1407 * page migration, if any.
1408 * @reason: The reason for page migration.
1410 * The function returns after 10 attempts or if no pages are movable any more
1411 * because the list has become empty or no retryable pages exist any more.
1412 * The caller should call putback_movable_pages() to return pages to the LRU
1413 * or free list only if ret != 0.
1415 * Returns the number of pages that were not migrated, or an error code.
1417 int migrate_pages(struct list_head *from, new_page_t get_new_page,
1418 free_page_t put_new_page, unsigned long private,
1419 enum migrate_mode mode, int reason)
1424 int nr_succeeded = 0;
1425 int nr_thp_succeeded = 0;
1426 int nr_thp_failed = 0;
1427 int nr_thp_split = 0;
1429 bool is_thp = false;
1432 int swapwrite = current->flags & PF_SWAPWRITE;
1433 int rc, nr_subpages;
1436 current->flags |= PF_SWAPWRITE;
1438 for (pass = 0; pass < 10 && (retry || thp_retry); pass++) {
1442 list_for_each_entry_safe(page, page2, from, lru) {
1445 * THP statistics is based on the source huge page.
1446 * Capture required information that might get lost
1449 is_thp = PageTransHuge(page);
1450 nr_subpages = thp_nr_pages(page);
1454 rc = unmap_and_move_huge_page(get_new_page,
1455 put_new_page, private, page,
1456 pass > 2, mode, reason);
1458 rc = unmap_and_move(get_new_page, put_new_page,
1459 private, page, pass > 2, mode,
1465 * THP migration might be unsupported or the
1466 * allocation could've failed so we should
1467 * retry on the same page with the THP split
1470 * Head page is retried immediately and tail
1471 * pages are added to the tail of the list so
1472 * we encounter them after the rest of the list
1475 if (PageTransHuge(page) && !PageHuge(page)) {
1477 rc = split_huge_page_to_list(page, from);
1480 list_safe_reset_next(page, page2, lru);
1487 nr_failed += nr_subpages;
1499 case MIGRATEPAGE_SUCCESS:
1502 nr_succeeded += nr_subpages;
1509 * Permanent failure (-EBUSY, -ENOSYS, etc.):
1510 * unlike -EAGAIN case, the failed page is
1511 * removed from migration page list and not
1512 * retried in the next outer loop.
1516 nr_failed += nr_subpages;
1524 nr_failed += retry + thp_retry;
1525 nr_thp_failed += thp_retry;
1528 count_vm_events(PGMIGRATE_SUCCESS, nr_succeeded);
1529 count_vm_events(PGMIGRATE_FAIL, nr_failed);
1530 count_vm_events(THP_MIGRATION_SUCCESS, nr_thp_succeeded);
1531 count_vm_events(THP_MIGRATION_FAIL, nr_thp_failed);
1532 count_vm_events(THP_MIGRATION_SPLIT, nr_thp_split);
1533 trace_mm_migrate_pages(nr_succeeded, nr_failed, nr_thp_succeeded,
1534 nr_thp_failed, nr_thp_split, mode, reason);
1537 current->flags &= ~PF_SWAPWRITE;
1542 struct page *alloc_migration_target(struct page *page, unsigned long private)
1544 struct migration_target_control *mtc;
1546 unsigned int order = 0;
1547 struct page *new_page = NULL;
1551 mtc = (struct migration_target_control *)private;
1552 gfp_mask = mtc->gfp_mask;
1554 if (nid == NUMA_NO_NODE)
1555 nid = page_to_nid(page);
1557 if (PageHuge(page)) {
1558 struct hstate *h = page_hstate(compound_head(page));
1560 gfp_mask = htlb_modify_alloc_mask(h, gfp_mask);
1561 return alloc_huge_page_nodemask(h, nid, mtc->nmask, gfp_mask);
1564 if (PageTransHuge(page)) {
1566 * clear __GFP_RECLAIM to make the migration callback
1567 * consistent with regular THP allocations.
1569 gfp_mask &= ~__GFP_RECLAIM;
1570 gfp_mask |= GFP_TRANSHUGE;
1571 order = HPAGE_PMD_ORDER;
1573 zidx = zone_idx(page_zone(page));
1574 if (is_highmem_idx(zidx) || zidx == ZONE_MOVABLE)
1575 gfp_mask |= __GFP_HIGHMEM;
1577 new_page = __alloc_pages_nodemask(gfp_mask, order, nid, mtc->nmask);
1579 if (new_page && PageTransHuge(new_page))
1580 prep_transhuge_page(new_page);
1587 static int store_status(int __user *status, int start, int value, int nr)
1590 if (put_user(value, status + start))
1598 static int do_move_pages_to_node(struct mm_struct *mm,
1599 struct list_head *pagelist, int node)
1602 struct migration_target_control mtc = {
1604 .gfp_mask = GFP_HIGHUSER_MOVABLE | __GFP_THISNODE,
1607 err = migrate_pages(pagelist, alloc_migration_target, NULL,
1608 (unsigned long)&mtc, MIGRATE_SYNC, MR_SYSCALL);
1610 putback_movable_pages(pagelist);
1615 * Resolves the given address to a struct page, isolates it from the LRU and
1616 * puts it to the given pagelist.
1618 * errno - if the page cannot be found/isolated
1619 * 0 - when it doesn't have to be migrated because it is already on the
1621 * 1 - when it has been queued
1623 static int add_page_for_migration(struct mm_struct *mm, unsigned long addr,
1624 int node, struct list_head *pagelist, bool migrate_all)
1626 struct vm_area_struct *vma;
1628 unsigned int follflags;
1633 vma = find_vma(mm, addr);
1634 if (!vma || addr < vma->vm_start || !vma_migratable(vma))
1637 /* FOLL_DUMP to ignore special (like zero) pages */
1638 follflags = FOLL_GET | FOLL_DUMP;
1639 page = follow_page(vma, addr, follflags);
1641 err = PTR_ERR(page);
1650 if (page_to_nid(page) == node)
1654 if (page_mapcount(page) > 1 && !migrate_all)
1657 if (PageHuge(page)) {
1658 if (PageHead(page)) {
1659 isolate_huge_page(page, pagelist);
1665 head = compound_head(page);
1666 err = isolate_lru_page(head);
1671 list_add_tail(&head->lru, pagelist);
1672 mod_node_page_state(page_pgdat(head),
1673 NR_ISOLATED_ANON + page_is_file_lru(head),
1674 thp_nr_pages(head));
1678 * Either remove the duplicate refcount from
1679 * isolate_lru_page() or drop the page ref if it was
1684 mmap_read_unlock(mm);
1688 static int move_pages_and_store_status(struct mm_struct *mm, int node,
1689 struct list_head *pagelist, int __user *status,
1690 int start, int i, unsigned long nr_pages)
1694 if (list_empty(pagelist))
1697 err = do_move_pages_to_node(mm, pagelist, node);
1700 * Positive err means the number of failed
1701 * pages to migrate. Since we are going to
1702 * abort and return the number of non-migrated
1703 * pages, so need to incude the rest of the
1704 * nr_pages that have not been attempted as
1708 err += nr_pages - i - 1;
1711 return store_status(status, start, node, i - start);
1715 * Migrate an array of page address onto an array of nodes and fill
1716 * the corresponding array of status.
1718 static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes,
1719 unsigned long nr_pages,
1720 const void __user * __user *pages,
1721 const int __user *nodes,
1722 int __user *status, int flags)
1724 int current_node = NUMA_NO_NODE;
1725 LIST_HEAD(pagelist);
1731 for (i = start = 0; i < nr_pages; i++) {
1732 const void __user *p;
1737 if (get_user(p, pages + i))
1739 if (get_user(node, nodes + i))
1741 addr = (unsigned long)untagged_addr(p);
1744 if (node < 0 || node >= MAX_NUMNODES)
1746 if (!node_state(node, N_MEMORY))
1750 if (!node_isset(node, task_nodes))
1753 if (current_node == NUMA_NO_NODE) {
1754 current_node = node;
1756 } else if (node != current_node) {
1757 err = move_pages_and_store_status(mm, current_node,
1758 &pagelist, status, start, i, nr_pages);
1762 current_node = node;
1766 * Errors in the page lookup or isolation are not fatal and we simply
1767 * report them via status
1769 err = add_page_for_migration(mm, addr, current_node,
1770 &pagelist, flags & MPOL_MF_MOVE_ALL);
1773 /* The page is successfully queued for migration */
1778 * If the page is already on the target node (!err), store the
1779 * node, otherwise, store the err.
1781 err = store_status(status, i, err ? : current_node, 1);
1785 err = move_pages_and_store_status(mm, current_node, &pagelist,
1786 status, start, i, nr_pages);
1789 current_node = NUMA_NO_NODE;
1792 /* Make sure we do not overwrite the existing error */
1793 err1 = move_pages_and_store_status(mm, current_node, &pagelist,
1794 status, start, i, nr_pages);
1802 * Determine the nodes of an array of pages and store it in an array of status.
1804 static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages,
1805 const void __user **pages, int *status)
1811 for (i = 0; i < nr_pages; i++) {
1812 unsigned long addr = (unsigned long)(*pages);
1813 struct vm_area_struct *vma;
1817 vma = find_vma(mm, addr);
1818 if (!vma || addr < vma->vm_start)
1821 /* FOLL_DUMP to ignore special (like zero) pages */
1822 page = follow_page(vma, addr, FOLL_DUMP);
1824 err = PTR_ERR(page);
1828 err = page ? page_to_nid(page) : -ENOENT;
1836 mmap_read_unlock(mm);
1840 * Determine the nodes of a user array of pages and store it in
1841 * a user array of status.
1843 static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages,
1844 const void __user * __user *pages,
1847 #define DO_PAGES_STAT_CHUNK_NR 16
1848 const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR];
1849 int chunk_status[DO_PAGES_STAT_CHUNK_NR];
1852 unsigned long chunk_nr;
1854 chunk_nr = nr_pages;
1855 if (chunk_nr > DO_PAGES_STAT_CHUNK_NR)
1856 chunk_nr = DO_PAGES_STAT_CHUNK_NR;
1858 if (copy_from_user(chunk_pages, pages, chunk_nr * sizeof(*chunk_pages)))
1861 do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status);
1863 if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status)))
1868 nr_pages -= chunk_nr;
1870 return nr_pages ? -EFAULT : 0;
1874 * Move a list of pages in the address space of the currently executing
1877 static int kernel_move_pages(pid_t pid, unsigned long nr_pages,
1878 const void __user * __user *pages,
1879 const int __user *nodes,
1880 int __user *status, int flags)
1882 struct task_struct *task;
1883 struct mm_struct *mm;
1885 nodemask_t task_nodes;
1888 if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
1891 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1894 /* Find the mm_struct */
1896 task = pid ? find_task_by_vpid(pid) : current;
1901 get_task_struct(task);
1904 * Check if this process has the right to modify the specified
1905 * process. Use the regular "ptrace_may_access()" checks.
1907 if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) {
1914 err = security_task_movememory(task);
1918 task_nodes = cpuset_mems_allowed(task);
1919 mm = get_task_mm(task);
1920 put_task_struct(task);
1926 err = do_pages_move(mm, task_nodes, nr_pages, pages,
1927 nodes, status, flags);
1929 err = do_pages_stat(mm, nr_pages, pages, status);
1935 put_task_struct(task);
1939 SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages,
1940 const void __user * __user *, pages,
1941 const int __user *, nodes,
1942 int __user *, status, int, flags)
1944 return kernel_move_pages(pid, nr_pages, pages, nodes, status, flags);
1947 #ifdef CONFIG_COMPAT
1948 COMPAT_SYSCALL_DEFINE6(move_pages, pid_t, pid, compat_ulong_t, nr_pages,
1949 compat_uptr_t __user *, pages32,
1950 const int __user *, nodes,
1951 int __user *, status,
1954 const void __user * __user *pages;
1957 pages = compat_alloc_user_space(nr_pages * sizeof(void *));
1958 for (i = 0; i < nr_pages; i++) {
1961 if (get_user(p, pages32 + i) ||
1962 put_user(compat_ptr(p), pages + i))
1965 return kernel_move_pages(pid, nr_pages, pages, nodes, status, flags);
1967 #endif /* CONFIG_COMPAT */
1969 #ifdef CONFIG_NUMA_BALANCING
1971 * Returns true if this is a safe migration target node for misplaced NUMA
1972 * pages. Currently it only checks the watermarks which crude
1974 static bool migrate_balanced_pgdat(struct pglist_data *pgdat,
1975 unsigned long nr_migrate_pages)
1979 for (z = pgdat->nr_zones - 1; z >= 0; z--) {
1980 struct zone *zone = pgdat->node_zones + z;
1982 if (!populated_zone(zone))
1985 /* Avoid waking kswapd by allocating pages_to_migrate pages. */
1986 if (!zone_watermark_ok(zone, 0,
1987 high_wmark_pages(zone) +
1996 static struct page *alloc_misplaced_dst_page(struct page *page,
1999 int nid = (int) data;
2000 struct page *newpage;
2002 newpage = __alloc_pages_node(nid,
2003 (GFP_HIGHUSER_MOVABLE |
2004 __GFP_THISNODE | __GFP_NOMEMALLOC |
2005 __GFP_NORETRY | __GFP_NOWARN) &
2011 static int numamigrate_isolate_page(pg_data_t *pgdat, struct page *page)
2015 VM_BUG_ON_PAGE(compound_order(page) && !PageTransHuge(page), page);
2017 /* Avoid migrating to a node that is nearly full */
2018 if (!migrate_balanced_pgdat(pgdat, compound_nr(page)))
2021 if (isolate_lru_page(page))
2025 * migrate_misplaced_transhuge_page() skips page migration's usual
2026 * check on page_count(), so we must do it here, now that the page
2027 * has been isolated: a GUP pin, or any other pin, prevents migration.
2028 * The expected page count is 3: 1 for page's mapcount and 1 for the
2029 * caller's pin and 1 for the reference taken by isolate_lru_page().
2031 if (PageTransHuge(page) && page_count(page) != 3) {
2032 putback_lru_page(page);
2036 page_lru = page_is_file_lru(page);
2037 mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON + page_lru,
2038 thp_nr_pages(page));
2041 * Isolating the page has taken another reference, so the
2042 * caller's reference can be safely dropped without the page
2043 * disappearing underneath us during migration.
2049 bool pmd_trans_migrating(pmd_t pmd)
2051 struct page *page = pmd_page(pmd);
2052 return PageLocked(page);
2056 * Attempt to migrate a misplaced page to the specified destination
2057 * node. Caller is expected to have an elevated reference count on
2058 * the page that will be dropped by this function before returning.
2060 int migrate_misplaced_page(struct page *page, struct vm_area_struct *vma,
2063 pg_data_t *pgdat = NODE_DATA(node);
2066 LIST_HEAD(migratepages);
2069 * Don't migrate file pages that are mapped in multiple processes
2070 * with execute permissions as they are probably shared libraries.
2072 if (page_mapcount(page) != 1 && page_is_file_lru(page) &&
2073 (vma->vm_flags & VM_EXEC))
2077 * Also do not migrate dirty pages as not all filesystems can move
2078 * dirty pages in MIGRATE_ASYNC mode which is a waste of cycles.
2080 if (page_is_file_lru(page) && PageDirty(page))
2083 isolated = numamigrate_isolate_page(pgdat, page);
2087 list_add(&page->lru, &migratepages);
2088 nr_remaining = migrate_pages(&migratepages, alloc_misplaced_dst_page,
2089 NULL, node, MIGRATE_ASYNC,
2092 if (!list_empty(&migratepages)) {
2093 list_del(&page->lru);
2094 dec_node_page_state(page, NR_ISOLATED_ANON +
2095 page_is_file_lru(page));
2096 putback_lru_page(page);
2100 count_vm_numa_event(NUMA_PAGE_MIGRATE);
2101 BUG_ON(!list_empty(&migratepages));
2108 #endif /* CONFIG_NUMA_BALANCING */
2110 #if defined(CONFIG_NUMA_BALANCING) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
2112 * Migrates a THP to a given target node. page must be locked and is unlocked
2115 int migrate_misplaced_transhuge_page(struct mm_struct *mm,
2116 struct vm_area_struct *vma,
2117 pmd_t *pmd, pmd_t entry,
2118 unsigned long address,
2119 struct page *page, int node)
2122 pg_data_t *pgdat = NODE_DATA(node);
2124 struct page *new_page = NULL;
2125 int page_lru = page_is_file_lru(page);
2126 unsigned long start = address & HPAGE_PMD_MASK;
2128 new_page = alloc_pages_node(node,
2129 (GFP_TRANSHUGE_LIGHT | __GFP_THISNODE),
2133 prep_transhuge_page(new_page);
2135 isolated = numamigrate_isolate_page(pgdat, page);
2141 /* Prepare a page as a migration target */
2142 __SetPageLocked(new_page);
2143 if (PageSwapBacked(page))
2144 __SetPageSwapBacked(new_page);
2146 /* anon mapping, we can simply copy page->mapping to the new page: */
2147 new_page->mapping = page->mapping;
2148 new_page->index = page->index;
2149 /* flush the cache before copying using the kernel virtual address */
2150 flush_cache_range(vma, start, start + HPAGE_PMD_SIZE);
2151 migrate_page_copy(new_page, page);
2152 WARN_ON(PageLRU(new_page));
2154 /* Recheck the target PMD */
2155 ptl = pmd_lock(mm, pmd);
2156 if (unlikely(!pmd_same(*pmd, entry) || !page_ref_freeze(page, 2))) {
2159 /* Reverse changes made by migrate_page_copy() */
2160 if (TestClearPageActive(new_page))
2161 SetPageActive(page);
2162 if (TestClearPageUnevictable(new_page))
2163 SetPageUnevictable(page);
2165 unlock_page(new_page);
2166 put_page(new_page); /* Free it */
2168 /* Retake the callers reference and putback on LRU */
2170 putback_lru_page(page);
2171 mod_node_page_state(page_pgdat(page),
2172 NR_ISOLATED_ANON + page_lru, -HPAGE_PMD_NR);
2177 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
2178 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
2181 * Overwrite the old entry under pagetable lock and establish
2182 * the new PTE. Any parallel GUP will either observe the old
2183 * page blocking on the page lock, block on the page table
2184 * lock or observe the new page. The SetPageUptodate on the
2185 * new page and page_add_new_anon_rmap guarantee the copy is
2186 * visible before the pagetable update.
2188 page_add_anon_rmap(new_page, vma, start, true);
2190 * At this point the pmd is numa/protnone (i.e. non present) and the TLB
2191 * has already been flushed globally. So no TLB can be currently
2192 * caching this non present pmd mapping. There's no need to clear the
2193 * pmd before doing set_pmd_at(), nor to flush the TLB after
2194 * set_pmd_at(). Clearing the pmd here would introduce a race
2195 * condition against MADV_DONTNEED, because MADV_DONTNEED only holds the
2196 * mmap_lock for reading. If the pmd is set to NULL at any given time,
2197 * MADV_DONTNEED won't wait on the pmd lock and it'll skip clearing this
2200 set_pmd_at(mm, start, pmd, entry);
2201 update_mmu_cache_pmd(vma, address, &entry);
2203 page_ref_unfreeze(page, 2);
2204 mlock_migrate_page(new_page, page);
2205 page_remove_rmap(page, true);
2206 set_page_owner_migrate_reason(new_page, MR_NUMA_MISPLACED);
2210 /* Take an "isolate" reference and put new page on the LRU. */
2212 putback_lru_page(new_page);
2214 unlock_page(new_page);
2216 put_page(page); /* Drop the rmap reference */
2217 put_page(page); /* Drop the LRU isolation reference */
2219 count_vm_events(PGMIGRATE_SUCCESS, HPAGE_PMD_NR);
2220 count_vm_numa_events(NUMA_PAGE_MIGRATE, HPAGE_PMD_NR);
2222 mod_node_page_state(page_pgdat(page),
2223 NR_ISOLATED_ANON + page_lru,
2228 count_vm_events(PGMIGRATE_FAIL, HPAGE_PMD_NR);
2229 ptl = pmd_lock(mm, pmd);
2230 if (pmd_same(*pmd, entry)) {
2231 entry = pmd_modify(entry, vma->vm_page_prot);
2232 set_pmd_at(mm, start, pmd, entry);
2233 update_mmu_cache_pmd(vma, address, &entry);
2242 #endif /* CONFIG_NUMA_BALANCING */
2244 #endif /* CONFIG_NUMA */
2246 #ifdef CONFIG_DEVICE_PRIVATE
2247 static int migrate_vma_collect_hole(unsigned long start,
2249 __always_unused int depth,
2250 struct mm_walk *walk)
2252 struct migrate_vma *migrate = walk->private;
2255 /* Only allow populating anonymous memory. */
2256 if (!vma_is_anonymous(walk->vma)) {
2257 for (addr = start; addr < end; addr += PAGE_SIZE) {
2258 migrate->src[migrate->npages] = 0;
2259 migrate->dst[migrate->npages] = 0;
2265 for (addr = start; addr < end; addr += PAGE_SIZE) {
2266 migrate->src[migrate->npages] = MIGRATE_PFN_MIGRATE;
2267 migrate->dst[migrate->npages] = 0;
2275 static int migrate_vma_collect_skip(unsigned long start,
2277 struct mm_walk *walk)
2279 struct migrate_vma *migrate = walk->private;
2282 for (addr = start; addr < end; addr += PAGE_SIZE) {
2283 migrate->dst[migrate->npages] = 0;
2284 migrate->src[migrate->npages++] = 0;
2290 static int migrate_vma_collect_pmd(pmd_t *pmdp,
2291 unsigned long start,
2293 struct mm_walk *walk)
2295 struct migrate_vma *migrate = walk->private;
2296 struct vm_area_struct *vma = walk->vma;
2297 struct mm_struct *mm = vma->vm_mm;
2298 unsigned long addr = start, unmapped = 0;
2303 if (pmd_none(*pmdp))
2304 return migrate_vma_collect_hole(start, end, -1, walk);
2306 if (pmd_trans_huge(*pmdp)) {
2309 ptl = pmd_lock(mm, pmdp);
2310 if (unlikely(!pmd_trans_huge(*pmdp))) {
2315 page = pmd_page(*pmdp);
2316 if (is_huge_zero_page(page)) {
2318 split_huge_pmd(vma, pmdp, addr);
2319 if (pmd_trans_unstable(pmdp))
2320 return migrate_vma_collect_skip(start, end,
2327 if (unlikely(!trylock_page(page)))
2328 return migrate_vma_collect_skip(start, end,
2330 ret = split_huge_page(page);
2334 return migrate_vma_collect_skip(start, end,
2336 if (pmd_none(*pmdp))
2337 return migrate_vma_collect_hole(start, end, -1,
2342 if (unlikely(pmd_bad(*pmdp)))
2343 return migrate_vma_collect_skip(start, end, walk);
2345 ptep = pte_offset_map_lock(mm, pmdp, addr, &ptl);
2346 arch_enter_lazy_mmu_mode();
2348 for (; addr < end; addr += PAGE_SIZE, ptep++) {
2349 unsigned long mpfn = 0, pfn;
2356 if (pte_none(pte)) {
2357 if (vma_is_anonymous(vma)) {
2358 mpfn = MIGRATE_PFN_MIGRATE;
2364 if (!pte_present(pte)) {
2366 * Only care about unaddressable device page special
2367 * page table entry. Other special swap entries are not
2368 * migratable, and we ignore regular swapped page.
2370 entry = pte_to_swp_entry(pte);
2371 if (!is_device_private_entry(entry))
2374 page = device_private_entry_to_page(entry);
2375 if (!(migrate->flags &
2376 MIGRATE_VMA_SELECT_DEVICE_PRIVATE) ||
2377 page->pgmap->owner != migrate->pgmap_owner)
2380 mpfn = migrate_pfn(page_to_pfn(page)) |
2381 MIGRATE_PFN_MIGRATE;
2382 if (is_write_device_private_entry(entry))
2383 mpfn |= MIGRATE_PFN_WRITE;
2385 if (!(migrate->flags & MIGRATE_VMA_SELECT_SYSTEM))
2388 if (is_zero_pfn(pfn)) {
2389 mpfn = MIGRATE_PFN_MIGRATE;
2393 page = vm_normal_page(migrate->vma, addr, pte);
2394 mpfn = migrate_pfn(pfn) | MIGRATE_PFN_MIGRATE;
2395 mpfn |= pte_write(pte) ? MIGRATE_PFN_WRITE : 0;
2398 /* FIXME support THP */
2399 if (!page || !page->mapping || PageTransCompound(page)) {
2405 * By getting a reference on the page we pin it and that blocks
2406 * any kind of migration. Side effect is that it "freezes" the
2409 * We drop this reference after isolating the page from the lru
2410 * for non device page (device page are not on the lru and thus
2411 * can't be dropped from it).
2417 * Optimize for the common case where page is only mapped once
2418 * in one process. If we can lock the page, then we can safely
2419 * set up a special migration page table entry now.
2421 if (trylock_page(page)) {
2424 mpfn |= MIGRATE_PFN_LOCKED;
2425 ptep_get_and_clear(mm, addr, ptep);
2427 /* Setup special migration page table entry */
2428 entry = make_migration_entry(page, mpfn &
2430 swp_pte = swp_entry_to_pte(entry);
2431 if (pte_present(pte)) {
2432 if (pte_soft_dirty(pte))
2433 swp_pte = pte_swp_mksoft_dirty(swp_pte);
2434 if (pte_uffd_wp(pte))
2435 swp_pte = pte_swp_mkuffd_wp(swp_pte);
2437 if (pte_swp_soft_dirty(pte))
2438 swp_pte = pte_swp_mksoft_dirty(swp_pte);
2439 if (pte_swp_uffd_wp(pte))
2440 swp_pte = pte_swp_mkuffd_wp(swp_pte);
2442 set_pte_at(mm, addr, ptep, swp_pte);
2445 * This is like regular unmap: we remove the rmap and
2446 * drop page refcount. Page won't be freed, as we took
2447 * a reference just above.
2449 page_remove_rmap(page, false);
2452 if (pte_present(pte))
2457 migrate->dst[migrate->npages] = 0;
2458 migrate->src[migrate->npages++] = mpfn;
2460 arch_leave_lazy_mmu_mode();
2461 pte_unmap_unlock(ptep - 1, ptl);
2463 /* Only flush the TLB if we actually modified any entries */
2465 flush_tlb_range(walk->vma, start, end);
2470 static const struct mm_walk_ops migrate_vma_walk_ops = {
2471 .pmd_entry = migrate_vma_collect_pmd,
2472 .pte_hole = migrate_vma_collect_hole,
2476 * migrate_vma_collect() - collect pages over a range of virtual addresses
2477 * @migrate: migrate struct containing all migration information
2479 * This will walk the CPU page table. For each virtual address backed by a
2480 * valid page, it updates the src array and takes a reference on the page, in
2481 * order to pin the page until we lock it and unmap it.
2483 static void migrate_vma_collect(struct migrate_vma *migrate)
2485 struct mmu_notifier_range range;
2488 * Note that the pgmap_owner is passed to the mmu notifier callback so
2489 * that the registered device driver can skip invalidating device
2490 * private page mappings that won't be migrated.
2492 mmu_notifier_range_init_migrate(&range, 0, migrate->vma,
2493 migrate->vma->vm_mm, migrate->start, migrate->end,
2494 migrate->pgmap_owner);
2495 mmu_notifier_invalidate_range_start(&range);
2497 walk_page_range(migrate->vma->vm_mm, migrate->start, migrate->end,
2498 &migrate_vma_walk_ops, migrate);
2500 mmu_notifier_invalidate_range_end(&range);
2501 migrate->end = migrate->start + (migrate->npages << PAGE_SHIFT);
2505 * migrate_vma_check_page() - check if page is pinned or not
2506 * @page: struct page to check
2508 * Pinned pages cannot be migrated. This is the same test as in
2509 * migrate_page_move_mapping(), except that here we allow migration of a
2512 static bool migrate_vma_check_page(struct page *page)
2515 * One extra ref because caller holds an extra reference, either from
2516 * isolate_lru_page() for a regular page, or migrate_vma_collect() for
2522 * FIXME support THP (transparent huge page), it is bit more complex to
2523 * check them than regular pages, because they can be mapped with a pmd
2524 * or with a pte (split pte mapping).
2526 if (PageCompound(page))
2529 /* Page from ZONE_DEVICE have one extra reference */
2530 if (is_zone_device_page(page)) {
2532 * Private page can never be pin as they have no valid pte and
2533 * GUP will fail for those. Yet if there is a pending migration
2534 * a thread might try to wait on the pte migration entry and
2535 * will bump the page reference count. Sadly there is no way to
2536 * differentiate a regular pin from migration wait. Hence to
2537 * avoid 2 racing thread trying to migrate back to CPU to enter
2538 * infinite loop (one stoping migration because the other is
2539 * waiting on pte migration entry). We always return true here.
2541 * FIXME proper solution is to rework migration_entry_wait() so
2542 * it does not need to take a reference on page.
2544 return is_device_private_page(page);
2547 /* For file back page */
2548 if (page_mapping(page))
2549 extra += 1 + page_has_private(page);
2551 if ((page_count(page) - extra) > page_mapcount(page))
2558 * migrate_vma_prepare() - lock pages and isolate them from the lru
2559 * @migrate: migrate struct containing all migration information
2561 * This locks pages that have been collected by migrate_vma_collect(). Once each
2562 * page is locked it is isolated from the lru (for non-device pages). Finally,
2563 * the ref taken by migrate_vma_collect() is dropped, as locked pages cannot be
2564 * migrated by concurrent kernel threads.
2566 static void migrate_vma_prepare(struct migrate_vma *migrate)
2568 const unsigned long npages = migrate->npages;
2569 const unsigned long start = migrate->start;
2570 unsigned long addr, i, restore = 0;
2571 bool allow_drain = true;
2575 for (i = 0; (i < npages) && migrate->cpages; i++) {
2576 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2582 if (!(migrate->src[i] & MIGRATE_PFN_LOCKED)) {
2584 * Because we are migrating several pages there can be
2585 * a deadlock between 2 concurrent migration where each
2586 * are waiting on each other page lock.
2588 * Make migrate_vma() a best effort thing and backoff
2589 * for any page we can not lock right away.
2591 if (!trylock_page(page)) {
2592 migrate->src[i] = 0;
2598 migrate->src[i] |= MIGRATE_PFN_LOCKED;
2601 /* ZONE_DEVICE pages are not on LRU */
2602 if (!is_zone_device_page(page)) {
2603 if (!PageLRU(page) && allow_drain) {
2604 /* Drain CPU's pagevec */
2605 lru_add_drain_all();
2606 allow_drain = false;
2609 if (isolate_lru_page(page)) {
2611 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2615 migrate->src[i] = 0;
2623 /* Drop the reference we took in collect */
2627 if (!migrate_vma_check_page(page)) {
2629 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2633 if (!is_zone_device_page(page)) {
2635 putback_lru_page(page);
2638 migrate->src[i] = 0;
2642 if (!is_zone_device_page(page))
2643 putback_lru_page(page);
2650 for (i = 0, addr = start; i < npages && restore; i++, addr += PAGE_SIZE) {
2651 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2653 if (!page || (migrate->src[i] & MIGRATE_PFN_MIGRATE))
2656 remove_migration_pte(page, migrate->vma, addr, page);
2658 migrate->src[i] = 0;
2666 * migrate_vma_unmap() - replace page mapping with special migration pte entry
2667 * @migrate: migrate struct containing all migration information
2669 * Replace page mapping (CPU page table pte) with a special migration pte entry
2670 * and check again if it has been pinned. Pinned pages are restored because we
2671 * cannot migrate them.
2673 * This is the last step before we call the device driver callback to allocate
2674 * destination memory and copy contents of original page over to new page.
2676 static void migrate_vma_unmap(struct migrate_vma *migrate)
2678 int flags = TTU_MIGRATION | TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS;
2679 const unsigned long npages = migrate->npages;
2680 const unsigned long start = migrate->start;
2681 unsigned long addr, i, restore = 0;
2683 for (i = 0; i < npages; i++) {
2684 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2686 if (!page || !(migrate->src[i] & MIGRATE_PFN_MIGRATE))
2689 if (page_mapped(page)) {
2690 try_to_unmap(page, flags);
2691 if (page_mapped(page))
2695 if (migrate_vma_check_page(page))
2699 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2704 for (addr = start, i = 0; i < npages && restore; addr += PAGE_SIZE, i++) {
2705 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2707 if (!page || (migrate->src[i] & MIGRATE_PFN_MIGRATE))
2710 remove_migration_ptes(page, page, false);
2712 migrate->src[i] = 0;
2716 if (is_zone_device_page(page))
2719 putback_lru_page(page);
2724 * migrate_vma_setup() - prepare to migrate a range of memory
2725 * @args: contains the vma, start, and pfns arrays for the migration
2727 * Returns: negative errno on failures, 0 when 0 or more pages were migrated
2730 * Prepare to migrate a range of memory virtual address range by collecting all
2731 * the pages backing each virtual address in the range, saving them inside the
2732 * src array. Then lock those pages and unmap them. Once the pages are locked
2733 * and unmapped, check whether each page is pinned or not. Pages that aren't
2734 * pinned have the MIGRATE_PFN_MIGRATE flag set (by this function) in the
2735 * corresponding src array entry. Then restores any pages that are pinned, by
2736 * remapping and unlocking those pages.
2738 * The caller should then allocate destination memory and copy source memory to
2739 * it for all those entries (ie with MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE
2740 * flag set). Once these are allocated and copied, the caller must update each
2741 * corresponding entry in the dst array with the pfn value of the destination
2742 * page and with the MIGRATE_PFN_VALID and MIGRATE_PFN_LOCKED flags set
2743 * (destination pages must have their struct pages locked, via lock_page()).
2745 * Note that the caller does not have to migrate all the pages that are marked
2746 * with MIGRATE_PFN_MIGRATE flag in src array unless this is a migration from
2747 * device memory to system memory. If the caller cannot migrate a device page
2748 * back to system memory, then it must return VM_FAULT_SIGBUS, which has severe
2749 * consequences for the userspace process, so it must be avoided if at all
2752 * For empty entries inside CPU page table (pte_none() or pmd_none() is true) we
2753 * do set MIGRATE_PFN_MIGRATE flag inside the corresponding source array thus
2754 * allowing the caller to allocate device memory for those unback virtual
2755 * address. For this the caller simply has to allocate device memory and
2756 * properly set the destination entry like for regular migration. Note that
2757 * this can still fails and thus inside the device driver must check if the
2758 * migration was successful for those entries after calling migrate_vma_pages()
2759 * just like for regular migration.
2761 * After that, the callers must call migrate_vma_pages() to go over each entry
2762 * in the src array that has the MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE flag
2763 * set. If the corresponding entry in dst array has MIGRATE_PFN_VALID flag set,
2764 * then migrate_vma_pages() to migrate struct page information from the source
2765 * struct page to the destination struct page. If it fails to migrate the
2766 * struct page information, then it clears the MIGRATE_PFN_MIGRATE flag in the
2769 * At this point all successfully migrated pages have an entry in the src
2770 * array with MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE flag set and the dst
2771 * array entry with MIGRATE_PFN_VALID flag set.
2773 * Once migrate_vma_pages() returns the caller may inspect which pages were
2774 * successfully migrated, and which were not. Successfully migrated pages will
2775 * have the MIGRATE_PFN_MIGRATE flag set for their src array entry.
2777 * It is safe to update device page table after migrate_vma_pages() because
2778 * both destination and source page are still locked, and the mmap_lock is held
2779 * in read mode (hence no one can unmap the range being migrated).
2781 * Once the caller is done cleaning up things and updating its page table (if it
2782 * chose to do so, this is not an obligation) it finally calls
2783 * migrate_vma_finalize() to update the CPU page table to point to new pages
2784 * for successfully migrated pages or otherwise restore the CPU page table to
2785 * point to the original source pages.
2787 int migrate_vma_setup(struct migrate_vma *args)
2789 long nr_pages = (args->end - args->start) >> PAGE_SHIFT;
2791 args->start &= PAGE_MASK;
2792 args->end &= PAGE_MASK;
2793 if (!args->vma || is_vm_hugetlb_page(args->vma) ||
2794 (args->vma->vm_flags & VM_SPECIAL) || vma_is_dax(args->vma))
2798 if (args->start < args->vma->vm_start ||
2799 args->start >= args->vma->vm_end)
2801 if (args->end <= args->vma->vm_start || args->end > args->vma->vm_end)
2803 if (!args->src || !args->dst)
2806 memset(args->src, 0, sizeof(*args->src) * nr_pages);
2810 migrate_vma_collect(args);
2813 migrate_vma_prepare(args);
2815 migrate_vma_unmap(args);
2818 * At this point pages are locked and unmapped, and thus they have
2819 * stable content and can safely be copied to destination memory that
2820 * is allocated by the drivers.
2825 EXPORT_SYMBOL(migrate_vma_setup);
2828 * This code closely matches the code in:
2829 * __handle_mm_fault()
2830 * handle_pte_fault()
2831 * do_anonymous_page()
2832 * to map in an anonymous zero page but the struct page will be a ZONE_DEVICE
2835 static void migrate_vma_insert_page(struct migrate_vma *migrate,
2841 struct vm_area_struct *vma = migrate->vma;
2842 struct mm_struct *mm = vma->vm_mm;
2852 /* Only allow populating anonymous memory */
2853 if (!vma_is_anonymous(vma))
2856 pgdp = pgd_offset(mm, addr);
2857 p4dp = p4d_alloc(mm, pgdp, addr);
2860 pudp = pud_alloc(mm, p4dp, addr);
2863 pmdp = pmd_alloc(mm, pudp, addr);
2867 if (pmd_trans_huge(*pmdp) || pmd_devmap(*pmdp))
2871 * Use pte_alloc() instead of pte_alloc_map(). We can't run
2872 * pte_offset_map() on pmds where a huge pmd might be created
2873 * from a different thread.
2875 * pte_alloc_map() is safe to use under mmap_write_lock(mm) or when
2876 * parallel threads are excluded by other means.
2878 * Here we only have mmap_read_lock(mm).
2880 if (pte_alloc(mm, pmdp))
2883 /* See the comment in pte_alloc_one_map() */
2884 if (unlikely(pmd_trans_unstable(pmdp)))
2887 if (unlikely(anon_vma_prepare(vma)))
2889 if (mem_cgroup_charge(page, vma->vm_mm, GFP_KERNEL))
2893 * The memory barrier inside __SetPageUptodate makes sure that
2894 * preceding stores to the page contents become visible before
2895 * the set_pte_at() write.
2897 __SetPageUptodate(page);
2899 if (is_zone_device_page(page)) {
2900 if (is_device_private_page(page)) {
2901 swp_entry_t swp_entry;
2903 swp_entry = make_device_private_entry(page, vma->vm_flags & VM_WRITE);
2904 entry = swp_entry_to_pte(swp_entry);
2907 entry = mk_pte(page, vma->vm_page_prot);
2908 if (vma->vm_flags & VM_WRITE)
2909 entry = pte_mkwrite(pte_mkdirty(entry));
2912 ptep = pte_offset_map_lock(mm, pmdp, addr, &ptl);
2914 if (check_stable_address_space(mm))
2917 if (pte_present(*ptep)) {
2918 unsigned long pfn = pte_pfn(*ptep);
2920 if (!is_zero_pfn(pfn))
2923 } else if (!pte_none(*ptep))
2927 * Check for userfaultfd but do not deliver the fault. Instead,
2930 if (userfaultfd_missing(vma))
2933 inc_mm_counter(mm, MM_ANONPAGES);
2934 page_add_new_anon_rmap(page, vma, addr, false);
2935 if (!is_zone_device_page(page))
2936 lru_cache_add_inactive_or_unevictable(page, vma);
2940 flush_cache_page(vma, addr, pte_pfn(*ptep));
2941 ptep_clear_flush_notify(vma, addr, ptep);
2942 set_pte_at_notify(mm, addr, ptep, entry);
2943 update_mmu_cache(vma, addr, ptep);
2945 /* No need to invalidate - it was non-present before */
2946 set_pte_at(mm, addr, ptep, entry);
2947 update_mmu_cache(vma, addr, ptep);
2950 pte_unmap_unlock(ptep, ptl);
2951 *src = MIGRATE_PFN_MIGRATE;
2955 pte_unmap_unlock(ptep, ptl);
2957 *src &= ~MIGRATE_PFN_MIGRATE;
2961 * migrate_vma_pages() - migrate meta-data from src page to dst page
2962 * @migrate: migrate struct containing all migration information
2964 * This migrates struct page meta-data from source struct page to destination
2965 * struct page. This effectively finishes the migration from source page to the
2968 void migrate_vma_pages(struct migrate_vma *migrate)
2970 const unsigned long npages = migrate->npages;
2971 const unsigned long start = migrate->start;
2972 struct mmu_notifier_range range;
2973 unsigned long addr, i;
2974 bool notified = false;
2976 for (i = 0, addr = start; i < npages; addr += PAGE_SIZE, i++) {
2977 struct page *newpage = migrate_pfn_to_page(migrate->dst[i]);
2978 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2979 struct address_space *mapping;
2983 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2988 if (!(migrate->src[i] & MIGRATE_PFN_MIGRATE))
2993 mmu_notifier_range_init(&range,
2994 MMU_NOTIFY_CLEAR, 0,
2996 migrate->vma->vm_mm,
2997 addr, migrate->end);
2998 mmu_notifier_invalidate_range_start(&range);
3000 migrate_vma_insert_page(migrate, addr, newpage,
3006 mapping = page_mapping(page);
3008 if (is_zone_device_page(newpage)) {
3009 if (is_device_private_page(newpage)) {
3011 * For now only support private anonymous when
3012 * migrating to un-addressable device memory.
3015 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
3020 * Other types of ZONE_DEVICE page are not
3023 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
3028 r = migrate_page(mapping, newpage, page, MIGRATE_SYNC_NO_COPY);
3029 if (r != MIGRATEPAGE_SUCCESS)
3030 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
3034 * No need to double call mmu_notifier->invalidate_range() callback as
3035 * the above ptep_clear_flush_notify() inside migrate_vma_insert_page()
3036 * did already call it.
3039 mmu_notifier_invalidate_range_only_end(&range);
3041 EXPORT_SYMBOL(migrate_vma_pages);
3044 * migrate_vma_finalize() - restore CPU page table entry
3045 * @migrate: migrate struct containing all migration information
3047 * This replaces the special migration pte entry with either a mapping to the
3048 * new page if migration was successful for that page, or to the original page
3051 * This also unlocks the pages and puts them back on the lru, or drops the extra
3052 * refcount, for device pages.
3054 void migrate_vma_finalize(struct migrate_vma *migrate)
3056 const unsigned long npages = migrate->npages;
3059 for (i = 0; i < npages; i++) {
3060 struct page *newpage = migrate_pfn_to_page(migrate->dst[i]);
3061 struct page *page = migrate_pfn_to_page(migrate->src[i]);
3065 unlock_page(newpage);
3071 if (!(migrate->src[i] & MIGRATE_PFN_MIGRATE) || !newpage) {
3073 unlock_page(newpage);
3079 remove_migration_ptes(page, newpage, false);
3083 if (is_zone_device_page(page))
3086 putback_lru_page(page);
3088 if (newpage != page) {
3089 unlock_page(newpage);
3090 if (is_zone_device_page(newpage))
3093 putback_lru_page(newpage);
3097 EXPORT_SYMBOL(migrate_vma_finalize);
3098 #endif /* CONFIG_DEVICE_PRIVATE */