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);
233 #ifdef CONFIG_FINEGRAINED_THP
234 if (PageTransHuge(page) && pte_cont(*pvmw.pte)) {
235 VM_BUG_ON_PAGE(PageHuge(page) || !PageTransCompound(page), page);
236 remove_migration_huge_pte(&pvmw, new);
239 #endif /* CONFIG_FINEGRAINED_THP */
243 pte = pte_mkold(mk_pte(new, READ_ONCE(vma->vm_page_prot)));
244 if (pte_swp_soft_dirty(*pvmw.pte))
245 pte = pte_mksoft_dirty(pte);
248 * Recheck VMA as permissions can change since migration started
250 entry = pte_to_swp_entry(*pvmw.pte);
251 if (is_write_migration_entry(entry))
252 pte = maybe_mkwrite(pte, vma);
253 else if (pte_swp_uffd_wp(*pvmw.pte))
254 pte = pte_mkuffd_wp(pte);
256 if (unlikely(is_device_private_page(new))) {
257 entry = make_device_private_entry(new, pte_write(pte));
258 pte = swp_entry_to_pte(entry);
259 if (pte_swp_soft_dirty(*pvmw.pte))
260 pte = pte_swp_mksoft_dirty(pte);
261 if (pte_swp_uffd_wp(*pvmw.pte))
262 pte = pte_swp_mkuffd_wp(pte);
265 #ifdef CONFIG_HUGETLB_PAGE
267 pte = pte_mkhuge(pte);
268 pte = arch_make_huge_pte(pte, vma, new, 0);
269 set_huge_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte);
271 hugepage_add_anon_rmap(new, vma, pvmw.address);
273 page_dup_rmap(new, true);
277 set_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte);
280 page_add_anon_rmap(new, vma, pvmw.address, false);
282 page_add_file_rmap(new, false);
284 if (vma->vm_flags & VM_LOCKED && !PageTransCompound(new))
287 if (PageTransHuge(page) && PageMlocked(page))
288 clear_page_mlock(page);
290 /* No need to invalidate - it was non-present before */
291 update_mmu_cache(vma, pvmw.address, pvmw.pte);
298 * Get rid of all migration entries and replace them by
299 * references to the indicated page.
301 void remove_migration_ptes(struct page *old, struct page *new, bool locked)
303 struct rmap_walk_control rwc = {
304 .rmap_one = remove_migration_pte,
309 rmap_walk_locked(new, &rwc);
311 rmap_walk(new, &rwc);
315 * Something used the pte of a page under migration. We need to
316 * get to the page and wait until migration is finished.
317 * When we return from this function the fault will be retried.
319 void __migration_entry_wait(struct mm_struct *mm, pte_t *ptep,
328 if (!is_swap_pte(pte))
331 entry = pte_to_swp_entry(pte);
332 if (!is_migration_entry(entry))
335 page = migration_entry_to_page(entry);
338 * Once page cache replacement of page migration started, page_count
339 * is zero; but we must not call put_and_wait_on_page_locked() without
340 * a ref. Use get_page_unless_zero(), and just fault again if it fails.
342 if (!get_page_unless_zero(page))
344 pte_unmap_unlock(ptep, ptl);
345 put_and_wait_on_page_locked(page);
348 pte_unmap_unlock(ptep, ptl);
351 void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
352 unsigned long address)
354 spinlock_t *ptl = pte_lockptr(mm, pmd);
355 pte_t *ptep = pte_offset_map(pmd, address);
356 __migration_entry_wait(mm, ptep, ptl);
359 void migration_entry_wait_huge(struct vm_area_struct *vma,
360 struct mm_struct *mm, pte_t *pte)
362 spinlock_t *ptl = huge_pte_lockptr(hstate_vma(vma), mm, pte);
363 __migration_entry_wait(mm, pte, ptl);
366 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
367 void pmd_migration_entry_wait(struct mm_struct *mm, pmd_t *pmd)
372 ptl = pmd_lock(mm, pmd);
373 if (!is_pmd_migration_entry(*pmd))
375 page = migration_entry_to_page(pmd_to_swp_entry(*pmd));
376 if (!get_page_unless_zero(page))
379 put_and_wait_on_page_locked(page);
386 static int expected_page_refs(struct address_space *mapping, struct page *page)
388 int expected_count = 1;
391 * Device private pages have an extra refcount as they are
394 expected_count += is_device_private_page(page);
396 expected_count += thp_nr_pages(page) + page_has_private(page);
398 return expected_count;
402 * Replace the page in the mapping.
404 * The number of remaining references must be:
405 * 1 for anonymous pages without a mapping
406 * 2 for pages with a mapping
407 * 3 for pages with a mapping and PagePrivate/PagePrivate2 set.
409 int migrate_page_move_mapping(struct address_space *mapping,
410 struct page *newpage, struct page *page, int extra_count)
412 XA_STATE(xas, &mapping->i_pages, page_index(page));
413 struct zone *oldzone, *newzone;
415 int expected_count = expected_page_refs(mapping, page) + extra_count;
416 int nr = thp_nr_pages(page);
419 /* Anonymous page without mapping */
420 if (page_count(page) != expected_count)
423 /* No turning back from here */
424 newpage->index = page->index;
425 newpage->mapping = page->mapping;
426 if (PageSwapBacked(page))
427 __SetPageSwapBacked(newpage);
429 return MIGRATEPAGE_SUCCESS;
432 oldzone = page_zone(page);
433 newzone = page_zone(newpage);
436 if (page_count(page) != expected_count || xas_load(&xas) != page) {
437 xas_unlock_irq(&xas);
441 if (!page_ref_freeze(page, expected_count)) {
442 xas_unlock_irq(&xas);
447 * Now we know that no one else is looking at the page:
448 * no turning back from here.
450 newpage->index = page->index;
451 newpage->mapping = page->mapping;
452 page_ref_add(newpage, nr); /* add cache reference */
453 if (PageSwapBacked(page)) {
454 __SetPageSwapBacked(newpage);
455 if (PageSwapCache(page)) {
456 SetPageSwapCache(newpage);
457 set_page_private(newpage, page_private(page));
460 VM_BUG_ON_PAGE(PageSwapCache(page), page);
463 /* Move dirty while page refs frozen and newpage not yet exposed */
464 dirty = PageDirty(page);
466 ClearPageDirty(page);
467 SetPageDirty(newpage);
470 xas_store(&xas, newpage);
471 if (PageTransHuge(page)) {
474 for (i = 1; i < nr; i++) {
476 xas_store(&xas, newpage);
481 * Drop cache reference from old page by unfreezing
482 * to one less reference.
483 * We know this isn't the last reference.
485 page_ref_unfreeze(page, expected_count - nr);
488 /* Leave irq disabled to prevent preemption while updating stats */
491 * If moved to a different zone then also account
492 * the page for that zone. Other VM counters will be
493 * taken care of when we establish references to the
494 * new page and drop references to the old page.
496 * Note that anonymous pages are accounted for
497 * via NR_FILE_PAGES and NR_ANON_MAPPED if they
498 * are mapped to swap space.
500 if (newzone != oldzone) {
501 struct lruvec *old_lruvec, *new_lruvec;
502 struct mem_cgroup *memcg;
504 memcg = page_memcg(page);
505 old_lruvec = mem_cgroup_lruvec(memcg, oldzone->zone_pgdat);
506 new_lruvec = mem_cgroup_lruvec(memcg, newzone->zone_pgdat);
508 __mod_lruvec_state(old_lruvec, NR_FILE_PAGES, -nr);
509 __mod_lruvec_state(new_lruvec, NR_FILE_PAGES, nr);
510 if (PageSwapBacked(page) && !PageSwapCache(page)) {
511 __mod_lruvec_state(old_lruvec, NR_SHMEM, -nr);
512 __mod_lruvec_state(new_lruvec, NR_SHMEM, nr);
514 if (dirty && mapping_can_writeback(mapping)) {
515 __mod_lruvec_state(old_lruvec, NR_FILE_DIRTY, -nr);
516 __mod_zone_page_state(oldzone, NR_ZONE_WRITE_PENDING, -nr);
517 __mod_lruvec_state(new_lruvec, NR_FILE_DIRTY, nr);
518 __mod_zone_page_state(newzone, NR_ZONE_WRITE_PENDING, nr);
523 return MIGRATEPAGE_SUCCESS;
525 EXPORT_SYMBOL(migrate_page_move_mapping);
528 * The expected number of remaining references is the same as that
529 * of migrate_page_move_mapping().
531 int migrate_huge_page_move_mapping(struct address_space *mapping,
532 struct page *newpage, struct page *page)
534 XA_STATE(xas, &mapping->i_pages, page_index(page));
538 expected_count = 2 + page_has_private(page);
539 if (page_count(page) != expected_count || xas_load(&xas) != page) {
540 xas_unlock_irq(&xas);
544 if (!page_ref_freeze(page, expected_count)) {
545 xas_unlock_irq(&xas);
549 newpage->index = page->index;
550 newpage->mapping = page->mapping;
554 xas_store(&xas, newpage);
556 page_ref_unfreeze(page, expected_count - 1);
558 xas_unlock_irq(&xas);
560 return MIGRATEPAGE_SUCCESS;
564 * Gigantic pages are so large that we do not guarantee that page++ pointer
565 * arithmetic will work across the entire page. We need something more
568 static void __copy_gigantic_page(struct page *dst, struct page *src,
572 struct page *dst_base = dst;
573 struct page *src_base = src;
575 for (i = 0; i < nr_pages; ) {
577 copy_highpage(dst, src);
580 dst = mem_map_next(dst, dst_base, i);
581 src = mem_map_next(src, src_base, i);
585 static void copy_huge_page(struct page *dst, struct page *src)
592 struct hstate *h = page_hstate(src);
593 nr_pages = pages_per_huge_page(h);
595 if (unlikely(nr_pages > MAX_ORDER_NR_PAGES)) {
596 __copy_gigantic_page(dst, src, nr_pages);
601 BUG_ON(!PageTransHuge(src));
602 nr_pages = thp_nr_pages(src);
605 for (i = 0; i < nr_pages; i++) {
607 copy_highpage(dst + i, src + i);
612 * Copy the page to its new location
614 void migrate_page_states(struct page *newpage, struct page *page)
619 SetPageError(newpage);
620 if (PageReferenced(page))
621 SetPageReferenced(newpage);
622 if (PageUptodate(page))
623 SetPageUptodate(newpage);
624 if (TestClearPageActive(page)) {
625 VM_BUG_ON_PAGE(PageUnevictable(page), page);
626 SetPageActive(newpage);
627 } else if (TestClearPageUnevictable(page))
628 SetPageUnevictable(newpage);
629 if (PageWorkingset(page))
630 SetPageWorkingset(newpage);
631 if (PageChecked(page))
632 SetPageChecked(newpage);
633 if (PageMappedToDisk(page))
634 SetPageMappedToDisk(newpage);
636 /* Move dirty on pages not done by migrate_page_move_mapping() */
638 SetPageDirty(newpage);
640 if (page_is_young(page))
641 set_page_young(newpage);
642 if (page_is_idle(page))
643 set_page_idle(newpage);
646 * Copy NUMA information to the new page, to prevent over-eager
647 * future migrations of this same page.
649 cpupid = page_cpupid_xchg_last(page, -1);
650 page_cpupid_xchg_last(newpage, cpupid);
652 ksm_migrate_page(newpage, page);
654 * Please do not reorder this without considering how mm/ksm.c's
655 * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache().
657 if (PageSwapCache(page))
658 ClearPageSwapCache(page);
659 ClearPagePrivate(page);
660 set_page_private(page, 0);
663 * If any waiters have accumulated on the new page then
666 if (PageWriteback(newpage))
667 end_page_writeback(newpage);
670 * PG_readahead shares the same bit with PG_reclaim. The above
671 * end_page_writeback() may clear PG_readahead mistakenly, so set the
674 if (PageReadahead(page))
675 SetPageReadahead(newpage);
677 copy_page_owner(page, newpage);
680 mem_cgroup_migrate(page, newpage);
682 EXPORT_SYMBOL(migrate_page_states);
684 void migrate_page_copy(struct page *newpage, struct page *page)
686 if (PageHuge(page) || PageTransHuge(page))
687 copy_huge_page(newpage, page);
689 copy_highpage(newpage, page);
691 migrate_page_states(newpage, page);
693 EXPORT_SYMBOL(migrate_page_copy);
695 /************************************************************
696 * Migration functions
697 ***********************************************************/
700 * Common logic to directly migrate a single LRU page suitable for
701 * pages that do not use PagePrivate/PagePrivate2.
703 * Pages are locked upon entry and exit.
705 int migrate_page(struct address_space *mapping,
706 struct page *newpage, struct page *page,
707 enum migrate_mode mode)
711 BUG_ON(PageWriteback(page)); /* Writeback must be complete */
713 rc = migrate_page_move_mapping(mapping, newpage, page, 0);
715 if (rc != MIGRATEPAGE_SUCCESS)
718 if (mode != MIGRATE_SYNC_NO_COPY)
719 migrate_page_copy(newpage, page);
721 migrate_page_states(newpage, page);
722 return MIGRATEPAGE_SUCCESS;
724 EXPORT_SYMBOL(migrate_page);
727 /* Returns true if all buffers are successfully locked */
728 static bool buffer_migrate_lock_buffers(struct buffer_head *head,
729 enum migrate_mode mode)
731 struct buffer_head *bh = head;
733 /* Simple case, sync compaction */
734 if (mode != MIGRATE_ASYNC) {
737 bh = bh->b_this_page;
739 } while (bh != head);
744 /* async case, we cannot block on lock_buffer so use trylock_buffer */
746 if (!trylock_buffer(bh)) {
748 * We failed to lock the buffer and cannot stall in
749 * async migration. Release the taken locks
751 struct buffer_head *failed_bh = bh;
753 while (bh != failed_bh) {
755 bh = bh->b_this_page;
760 bh = bh->b_this_page;
761 } while (bh != head);
765 static int __buffer_migrate_page(struct address_space *mapping,
766 struct page *newpage, struct page *page, enum migrate_mode mode,
769 struct buffer_head *bh, *head;
773 if (!page_has_buffers(page))
774 return migrate_page(mapping, newpage, page, mode);
776 /* Check whether page does not have extra refs before we do more work */
777 expected_count = expected_page_refs(mapping, page);
778 if (page_count(page) != expected_count)
781 head = page_buffers(page);
782 if (!buffer_migrate_lock_buffers(head, mode))
787 bool invalidated = false;
791 spin_lock(&mapping->private_lock);
794 if (atomic_read(&bh->b_count)) {
798 bh = bh->b_this_page;
799 } while (bh != head);
805 spin_unlock(&mapping->private_lock);
806 invalidate_bh_lrus();
808 goto recheck_buffers;
812 rc = migrate_page_move_mapping(mapping, newpage, page, 0);
813 if (rc != MIGRATEPAGE_SUCCESS)
816 attach_page_private(newpage, detach_page_private(page));
820 set_bh_page(bh, newpage, bh_offset(bh));
821 bh = bh->b_this_page;
823 } while (bh != head);
825 if (mode != MIGRATE_SYNC_NO_COPY)
826 migrate_page_copy(newpage, page);
828 migrate_page_states(newpage, page);
830 rc = MIGRATEPAGE_SUCCESS;
833 spin_unlock(&mapping->private_lock);
837 bh = bh->b_this_page;
839 } while (bh != head);
845 * Migration function for pages with buffers. This function can only be used
846 * if the underlying filesystem guarantees that no other references to "page"
847 * exist. For example attached buffer heads are accessed only under page lock.
849 int buffer_migrate_page(struct address_space *mapping,
850 struct page *newpage, struct page *page, enum migrate_mode mode)
852 return __buffer_migrate_page(mapping, newpage, page, mode, false);
854 EXPORT_SYMBOL(buffer_migrate_page);
857 * Same as above except that this variant is more careful and checks that there
858 * are also no buffer head references. This function is the right one for
859 * mappings where buffer heads are directly looked up and referenced (such as
860 * block device mappings).
862 int buffer_migrate_page_norefs(struct address_space *mapping,
863 struct page *newpage, struct page *page, enum migrate_mode mode)
865 return __buffer_migrate_page(mapping, newpage, page, mode, true);
870 * Writeback a page to clean the dirty state
872 static int writeout(struct address_space *mapping, struct page *page)
874 struct writeback_control wbc = {
875 .sync_mode = WB_SYNC_NONE,
878 .range_end = LLONG_MAX,
883 if (!mapping->a_ops->writepage)
884 /* No write method for the address space */
887 if (!clear_page_dirty_for_io(page))
888 /* Someone else already triggered a write */
892 * A dirty page may imply that the underlying filesystem has
893 * the page on some queue. So the page must be clean for
894 * migration. Writeout may mean we loose the lock and the
895 * page state is no longer what we checked for earlier.
896 * At this point we know that the migration attempt cannot
899 remove_migration_ptes(page, page, false);
901 rc = mapping->a_ops->writepage(page, &wbc);
903 if (rc != AOP_WRITEPAGE_ACTIVATE)
904 /* unlocked. Relock */
907 return (rc < 0) ? -EIO : -EAGAIN;
911 * Default handling if a filesystem does not provide a migration function.
913 static int fallback_migrate_page(struct address_space *mapping,
914 struct page *newpage, struct page *page, enum migrate_mode mode)
916 if (PageDirty(page)) {
917 /* Only writeback pages in full synchronous migration */
920 case MIGRATE_SYNC_NO_COPY:
925 return writeout(mapping, page);
929 * Buffers may be managed in a filesystem specific way.
930 * We must have no buffers or drop them.
932 if (page_has_private(page) &&
933 !try_to_release_page(page, GFP_KERNEL))
934 return mode == MIGRATE_SYNC ? -EAGAIN : -EBUSY;
936 return migrate_page(mapping, newpage, page, mode);
940 * Move a page to a newly allocated page
941 * The page is locked and all ptes have been successfully removed.
943 * The new page will have replaced the old page if this function
948 * MIGRATEPAGE_SUCCESS - success
950 static int move_to_new_page(struct page *newpage, struct page *page,
951 enum migrate_mode mode)
953 struct address_space *mapping;
955 bool is_lru = !__PageMovable(page);
957 VM_BUG_ON_PAGE(!PageLocked(page), page);
958 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
960 mapping = page_mapping(page);
962 if (likely(is_lru)) {
964 rc = migrate_page(mapping, newpage, page, mode);
965 else if (mapping->a_ops->migratepage)
967 * Most pages have a mapping and most filesystems
968 * provide a migratepage callback. Anonymous pages
969 * are part of swap space which also has its own
970 * migratepage callback. This is the most common path
971 * for page migration.
973 rc = mapping->a_ops->migratepage(mapping, newpage,
976 rc = fallback_migrate_page(mapping, newpage,
980 * In case of non-lru page, it could be released after
981 * isolation step. In that case, we shouldn't try migration.
983 VM_BUG_ON_PAGE(!PageIsolated(page), page);
984 if (!PageMovable(page)) {
985 rc = MIGRATEPAGE_SUCCESS;
986 __ClearPageIsolated(page);
990 rc = mapping->a_ops->migratepage(mapping, newpage,
992 WARN_ON_ONCE(rc == MIGRATEPAGE_SUCCESS &&
993 !PageIsolated(page));
997 * When successful, old pagecache page->mapping must be cleared before
998 * page is freed; but stats require that PageAnon be left as PageAnon.
1000 if (rc == MIGRATEPAGE_SUCCESS) {
1001 if (__PageMovable(page)) {
1002 VM_BUG_ON_PAGE(!PageIsolated(page), page);
1005 * We clear PG_movable under page_lock so any compactor
1006 * cannot try to migrate this page.
1008 __ClearPageIsolated(page);
1012 * Anonymous and movable page->mapping will be cleared by
1013 * free_pages_prepare so don't reset it here for keeping
1014 * the type to work PageAnon, for example.
1016 if (!PageMappingFlags(page))
1017 page->mapping = NULL;
1019 if (likely(!is_zone_device_page(newpage)))
1020 flush_dcache_page(newpage);
1027 static int __unmap_and_move(struct page *page, struct page *newpage,
1028 int force, enum migrate_mode mode)
1031 int page_was_mapped = 0;
1032 struct anon_vma *anon_vma = NULL;
1033 bool is_lru = !__PageMovable(page);
1035 if (!trylock_page(page)) {
1036 if (!force || mode == MIGRATE_ASYNC)
1040 * It's not safe for direct compaction to call lock_page.
1041 * For example, during page readahead pages are added locked
1042 * to the LRU. Later, when the IO completes the pages are
1043 * marked uptodate and unlocked. However, the queueing
1044 * could be merging multiple pages for one bio (e.g.
1045 * mpage_readahead). If an allocation happens for the
1046 * second or third page, the process can end up locking
1047 * the same page twice and deadlocking. Rather than
1048 * trying to be clever about what pages can be locked,
1049 * avoid the use of lock_page for direct compaction
1052 if (current->flags & PF_MEMALLOC)
1058 if (PageWriteback(page)) {
1060 * Only in the case of a full synchronous migration is it
1061 * necessary to wait for PageWriteback. In the async case,
1062 * the retry loop is too short and in the sync-light case,
1063 * the overhead of stalling is too much
1067 case MIGRATE_SYNC_NO_COPY:
1075 wait_on_page_writeback(page);
1079 * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
1080 * we cannot notice that anon_vma is freed while we migrates a page.
1081 * This get_anon_vma() delays freeing anon_vma pointer until the end
1082 * of migration. File cache pages are no problem because of page_lock()
1083 * File Caches may use write_page() or lock_page() in migration, then,
1084 * just care Anon page here.
1086 * Only page_get_anon_vma() understands the subtleties of
1087 * getting a hold on an anon_vma from outside one of its mms.
1088 * But if we cannot get anon_vma, then we won't need it anyway,
1089 * because that implies that the anon page is no longer mapped
1090 * (and cannot be remapped so long as we hold the page lock).
1092 if (PageAnon(page) && !PageKsm(page))
1093 anon_vma = page_get_anon_vma(page);
1096 * Block others from accessing the new page when we get around to
1097 * establishing additional references. We are usually the only one
1098 * holding a reference to newpage at this point. We used to have a BUG
1099 * here if trylock_page(newpage) fails, but would like to allow for
1100 * cases where there might be a race with the previous use of newpage.
1101 * This is much like races on refcount of oldpage: just don't BUG().
1103 if (unlikely(!trylock_page(newpage)))
1106 if (unlikely(!is_lru)) {
1107 rc = move_to_new_page(newpage, page, mode);
1108 goto out_unlock_both;
1112 * Corner case handling:
1113 * 1. When a new swap-cache page is read into, it is added to the LRU
1114 * and treated as swapcache but it has no rmap yet.
1115 * Calling try_to_unmap() against a page->mapping==NULL page will
1116 * trigger a BUG. So handle it here.
1117 * 2. An orphaned page (see truncate_complete_page) might have
1118 * fs-private metadata. The page can be picked up due to memory
1119 * offlining. Everywhere else except page reclaim, the page is
1120 * invisible to the vm, so the page can not be migrated. So try to
1121 * free the metadata, so the page can be freed.
1123 if (!page->mapping) {
1124 VM_BUG_ON_PAGE(PageAnon(page), page);
1125 if (page_has_private(page)) {
1126 try_to_free_buffers(page);
1127 goto out_unlock_both;
1129 } else if (page_mapped(page)) {
1130 /* Establish migration ptes */
1131 VM_BUG_ON_PAGE(PageAnon(page) && !PageKsm(page) && !anon_vma,
1133 try_to_unmap(page, TTU_MIGRATION|TTU_IGNORE_MLOCK);
1134 page_was_mapped = 1;
1137 if (!page_mapped(page))
1138 rc = move_to_new_page(newpage, page, mode);
1140 if (page_was_mapped)
1141 remove_migration_ptes(page,
1142 rc == MIGRATEPAGE_SUCCESS ? newpage : page, false);
1145 unlock_page(newpage);
1147 /* Drop an anon_vma reference if we took one */
1149 put_anon_vma(anon_vma);
1153 * If migration is successful, decrease refcount of the newpage
1154 * which will not free the page because new page owner increased
1155 * refcounter. As well, if it is LRU page, add the page to LRU
1156 * list in here. Use the old state of the isolated source page to
1157 * determine if we migrated a LRU page. newpage was already unlocked
1158 * and possibly modified by its owner - don't rely on the page
1161 if (rc == MIGRATEPAGE_SUCCESS) {
1162 if (unlikely(!is_lru))
1165 putback_lru_page(newpage);
1172 * Obtain the lock on page, remove all ptes and migrate the page
1173 * to the newly allocated page in newpage.
1175 static int unmap_and_move(new_page_t get_new_page,
1176 free_page_t put_new_page,
1177 unsigned long private, struct page *page,
1178 int force, enum migrate_mode mode,
1179 enum migrate_reason reason)
1181 int rc = MIGRATEPAGE_SUCCESS;
1182 struct page *newpage = NULL;
1184 if (!thp_migration_supported() && PageTransHuge(page))
1187 if (page_count(page) == 1) {
1188 /* page was freed from under us. So we are done. */
1189 ClearPageActive(page);
1190 ClearPageUnevictable(page);
1191 if (unlikely(__PageMovable(page))) {
1193 if (!PageMovable(page))
1194 __ClearPageIsolated(page);
1200 newpage = get_new_page(page, private);
1204 rc = __unmap_and_move(page, newpage, force, mode);
1205 if (rc == MIGRATEPAGE_SUCCESS)
1206 set_page_owner_migrate_reason(newpage, reason);
1209 if (rc != -EAGAIN) {
1211 * A page that has been migrated has all references
1212 * removed and will be freed. A page that has not been
1213 * migrated will have kept its references and be restored.
1215 list_del(&page->lru);
1218 * Compaction can migrate also non-LRU pages which are
1219 * not accounted to NR_ISOLATED_*. They can be recognized
1222 if (likely(!__PageMovable(page)))
1223 mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON +
1224 page_is_file_lru(page), -thp_nr_pages(page));
1228 * If migration is successful, releases reference grabbed during
1229 * isolation. Otherwise, restore the page to right list unless
1232 if (rc == MIGRATEPAGE_SUCCESS) {
1233 if (reason != MR_MEMORY_FAILURE)
1235 * We release the page in page_handle_poison.
1239 if (rc != -EAGAIN) {
1240 if (likely(!__PageMovable(page))) {
1241 putback_lru_page(page);
1246 if (PageMovable(page))
1247 putback_movable_page(page);
1249 __ClearPageIsolated(page);
1255 put_new_page(newpage, private);
1264 * Counterpart of unmap_and_move_page() for hugepage migration.
1266 * This function doesn't wait the completion of hugepage I/O
1267 * because there is no race between I/O and migration for hugepage.
1268 * Note that currently hugepage I/O occurs only in direct I/O
1269 * where no lock is held and PG_writeback is irrelevant,
1270 * and writeback status of all subpages are counted in the reference
1271 * count of the head page (i.e. if all subpages of a 2MB hugepage are
1272 * under direct I/O, the reference of the head page is 512 and a bit more.)
1273 * This means that when we try to migrate hugepage whose subpages are
1274 * doing direct I/O, some references remain after try_to_unmap() and
1275 * hugepage migration fails without data corruption.
1277 * There is also no race when direct I/O is issued on the page under migration,
1278 * because then pte is replaced with migration swap entry and direct I/O code
1279 * will wait in the page fault for migration to complete.
1281 static int unmap_and_move_huge_page(new_page_t get_new_page,
1282 free_page_t put_new_page, unsigned long private,
1283 struct page *hpage, int force,
1284 enum migrate_mode mode, int reason)
1287 int page_was_mapped = 0;
1288 struct page *new_hpage;
1289 struct anon_vma *anon_vma = NULL;
1290 struct address_space *mapping = NULL;
1293 * Migratability of hugepages depends on architectures and their size.
1294 * This check is necessary because some callers of hugepage migration
1295 * like soft offline and memory hotremove don't walk through page
1296 * tables or check whether the hugepage is pmd-based or not before
1297 * kicking migration.
1299 if (!hugepage_migration_supported(page_hstate(hpage))) {
1300 putback_active_hugepage(hpage);
1304 new_hpage = get_new_page(hpage, private);
1308 if (!trylock_page(hpage)) {
1313 case MIGRATE_SYNC_NO_COPY:
1322 * Check for pages which are in the process of being freed. Without
1323 * page_mapping() set, hugetlbfs specific move page routine will not
1324 * be called and we could leak usage counts for subpools.
1326 if (page_private(hpage) && !page_mapping(hpage)) {
1331 if (PageAnon(hpage))
1332 anon_vma = page_get_anon_vma(hpage);
1334 if (unlikely(!trylock_page(new_hpage)))
1337 if (page_mapped(hpage)) {
1338 bool mapping_locked = false;
1339 enum ttu_flags ttu = TTU_MIGRATION|TTU_IGNORE_MLOCK;
1341 if (!PageAnon(hpage)) {
1343 * In shared mappings, try_to_unmap could potentially
1344 * call huge_pmd_unshare. Because of this, take
1345 * semaphore in write mode here and set TTU_RMAP_LOCKED
1346 * to let lower levels know we have taken the lock.
1348 mapping = hugetlb_page_mapping_lock_write(hpage);
1349 if (unlikely(!mapping))
1350 goto unlock_put_anon;
1352 mapping_locked = true;
1353 ttu |= TTU_RMAP_LOCKED;
1356 try_to_unmap(hpage, ttu);
1357 page_was_mapped = 1;
1360 i_mmap_unlock_write(mapping);
1363 if (!page_mapped(hpage))
1364 rc = move_to_new_page(new_hpage, hpage, mode);
1366 if (page_was_mapped)
1367 remove_migration_ptes(hpage,
1368 rc == MIGRATEPAGE_SUCCESS ? new_hpage : hpage, false);
1371 unlock_page(new_hpage);
1375 put_anon_vma(anon_vma);
1377 if (rc == MIGRATEPAGE_SUCCESS) {
1378 move_hugetlb_state(hpage, new_hpage, reason);
1379 put_new_page = NULL;
1386 putback_active_hugepage(hpage);
1389 * If migration was not successful and there's a freeing callback, use
1390 * it. Otherwise, put_page() will drop the reference grabbed during
1394 put_new_page(new_hpage, private);
1396 putback_active_hugepage(new_hpage);
1402 * migrate_pages - migrate the pages specified in a list, to the free pages
1403 * supplied as the target for the page migration
1405 * @from: The list of pages to be migrated.
1406 * @get_new_page: The function used to allocate free pages to be used
1407 * as the target of the page migration.
1408 * @put_new_page: The function used to free target pages if migration
1409 * fails, or NULL if no special handling is necessary.
1410 * @private: Private data to be passed on to get_new_page()
1411 * @mode: The migration mode that specifies the constraints for
1412 * page migration, if any.
1413 * @reason: The reason for page migration.
1415 * The function returns after 10 attempts or if no pages are movable any more
1416 * because the list has become empty or no retryable pages exist any more.
1417 * The caller should call putback_movable_pages() to return pages to the LRU
1418 * or free list only if ret != 0.
1420 * Returns the number of pages that were not migrated, or an error code.
1422 int migrate_pages(struct list_head *from, new_page_t get_new_page,
1423 free_page_t put_new_page, unsigned long private,
1424 enum migrate_mode mode, int reason)
1429 int nr_succeeded = 0;
1430 int nr_thp_succeeded = 0;
1431 int nr_thp_failed = 0;
1432 int nr_thp_split = 0;
1434 bool is_thp = false;
1437 int swapwrite = current->flags & PF_SWAPWRITE;
1438 int rc, nr_subpages;
1440 trace_mm_migrate_pages_start(mode, reason);
1443 current->flags |= PF_SWAPWRITE;
1445 for (pass = 0; pass < 10 && (retry || thp_retry); pass++) {
1449 list_for_each_entry_safe(page, page2, from, lru) {
1452 * THP statistics is based on the source huge page.
1453 * Capture required information that might get lost
1456 is_thp = PageTransHuge(page) && !PageHuge(page);
1457 nr_subpages = thp_nr_pages(page);
1461 rc = unmap_and_move_huge_page(get_new_page,
1462 put_new_page, private, page,
1463 pass > 2, mode, reason);
1465 rc = unmap_and_move(get_new_page, put_new_page,
1466 private, page, pass > 2, mode,
1472 * THP migration might be unsupported or the
1473 * allocation could've failed so we should
1474 * retry on the same page with the THP split
1477 * Head page is retried immediately and tail
1478 * pages are added to the tail of the list so
1479 * we encounter them after the rest of the list
1484 rc = split_huge_page_to_list(page, from);
1487 list_safe_reset_next(page, page2, lru);
1493 nr_failed += nr_subpages;
1505 case MIGRATEPAGE_SUCCESS:
1508 nr_succeeded += nr_subpages;
1515 * Permanent failure (-EBUSY, -ENOSYS, etc.):
1516 * unlike -EAGAIN case, the failed page is
1517 * removed from migration page list and not
1518 * retried in the next outer loop.
1522 nr_failed += nr_subpages;
1530 nr_failed += retry + thp_retry;
1531 nr_thp_failed += thp_retry;
1534 count_vm_events(PGMIGRATE_SUCCESS, nr_succeeded);
1535 count_vm_events(PGMIGRATE_FAIL, nr_failed);
1536 count_vm_events(THP_MIGRATION_SUCCESS, nr_thp_succeeded);
1537 count_vm_events(THP_MIGRATION_FAIL, nr_thp_failed);
1538 count_vm_events(THP_MIGRATION_SPLIT, nr_thp_split);
1539 trace_mm_migrate_pages(nr_succeeded, nr_failed, nr_thp_succeeded,
1540 nr_thp_failed, nr_thp_split, mode, reason);
1543 current->flags &= ~PF_SWAPWRITE;
1548 struct page *alloc_migration_target(struct page *page, unsigned long private)
1550 struct migration_target_control *mtc;
1552 unsigned int order = 0;
1553 struct page *new_page = NULL;
1557 mtc = (struct migration_target_control *)private;
1558 gfp_mask = mtc->gfp_mask;
1560 if (nid == NUMA_NO_NODE)
1561 nid = page_to_nid(page);
1563 if (PageHuge(page)) {
1564 struct hstate *h = page_hstate(compound_head(page));
1566 gfp_mask = htlb_modify_alloc_mask(h, gfp_mask);
1567 return alloc_huge_page_nodemask(h, nid, mtc->nmask, gfp_mask);
1570 if (PageTransHuge(page)) {
1572 * clear __GFP_RECLAIM to make the migration callback
1573 * consistent with regular THP allocations.
1575 gfp_mask &= ~__GFP_RECLAIM;
1576 gfp_mask |= GFP_TRANSHUGE;
1577 order = HPAGE_PMD_ORDER;
1579 zidx = zone_idx(page_zone(page));
1580 if (is_highmem_idx(zidx) || zidx == ZONE_MOVABLE)
1581 gfp_mask |= __GFP_HIGHMEM;
1583 new_page = __alloc_pages_nodemask(gfp_mask, order, nid, mtc->nmask);
1585 if (new_page && PageTransHuge(new_page))
1586 prep_transhuge_page(new_page);
1593 static int store_status(int __user *status, int start, int value, int nr)
1596 if (put_user(value, status + start))
1604 static int do_move_pages_to_node(struct mm_struct *mm,
1605 struct list_head *pagelist, int node)
1608 struct migration_target_control mtc = {
1610 .gfp_mask = GFP_HIGHUSER_MOVABLE | __GFP_THISNODE,
1613 err = migrate_pages(pagelist, alloc_migration_target, NULL,
1614 (unsigned long)&mtc, MIGRATE_SYNC, MR_SYSCALL);
1616 putback_movable_pages(pagelist);
1621 * Resolves the given address to a struct page, isolates it from the LRU and
1622 * puts it to the given pagelist.
1624 * errno - if the page cannot be found/isolated
1625 * 0 - when it doesn't have to be migrated because it is already on the
1627 * 1 - when it has been queued
1629 static int add_page_for_migration(struct mm_struct *mm, unsigned long addr,
1630 int node, struct list_head *pagelist, bool migrate_all)
1632 struct vm_area_struct *vma;
1634 unsigned int follflags;
1639 vma = find_vma(mm, addr);
1640 if (!vma || addr < vma->vm_start || !vma_migratable(vma))
1643 /* FOLL_DUMP to ignore special (like zero) pages */
1644 follflags = FOLL_GET | FOLL_DUMP;
1645 page = follow_page(vma, addr, follflags);
1647 err = PTR_ERR(page);
1656 if (page_to_nid(page) == node)
1660 if (page_mapcount(page) > 1 && !migrate_all)
1663 if (PageHuge(page)) {
1664 if (PageHead(page)) {
1665 isolate_huge_page(page, pagelist);
1671 head = compound_head(page);
1672 err = isolate_lru_page(head);
1677 list_add_tail(&head->lru, pagelist);
1678 mod_node_page_state(page_pgdat(head),
1679 NR_ISOLATED_ANON + page_is_file_lru(head),
1680 thp_nr_pages(head));
1684 * Either remove the duplicate refcount from
1685 * isolate_lru_page() or drop the page ref if it was
1690 mmap_read_unlock(mm);
1694 static int move_pages_and_store_status(struct mm_struct *mm, int node,
1695 struct list_head *pagelist, int __user *status,
1696 int start, int i, unsigned long nr_pages)
1700 if (list_empty(pagelist))
1703 err = do_move_pages_to_node(mm, pagelist, node);
1706 * Positive err means the number of failed
1707 * pages to migrate. Since we are going to
1708 * abort and return the number of non-migrated
1709 * pages, so need to incude the rest of the
1710 * nr_pages that have not been attempted as
1714 err += nr_pages - i - 1;
1717 return store_status(status, start, node, i - start);
1721 * Migrate an array of page address onto an array of nodes and fill
1722 * the corresponding array of status.
1724 static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes,
1725 unsigned long nr_pages,
1726 const void __user * __user *pages,
1727 const int __user *nodes,
1728 int __user *status, int flags)
1730 int current_node = NUMA_NO_NODE;
1731 LIST_HEAD(pagelist);
1737 for (i = start = 0; i < nr_pages; i++) {
1738 const void __user *p;
1743 if (get_user(p, pages + i))
1745 if (get_user(node, nodes + i))
1747 addr = (unsigned long)untagged_addr(p);
1750 if (node < 0 || node >= MAX_NUMNODES)
1752 if (!node_state(node, N_MEMORY))
1756 if (!node_isset(node, task_nodes))
1759 if (current_node == NUMA_NO_NODE) {
1760 current_node = node;
1762 } else if (node != current_node) {
1763 err = move_pages_and_store_status(mm, current_node,
1764 &pagelist, status, start, i, nr_pages);
1768 current_node = node;
1772 * Errors in the page lookup or isolation are not fatal and we simply
1773 * report them via status
1775 err = add_page_for_migration(mm, addr, current_node,
1776 &pagelist, flags & MPOL_MF_MOVE_ALL);
1779 /* The page is successfully queued for migration */
1784 * If the page is already on the target node (!err), store the
1785 * node, otherwise, store the err.
1787 err = store_status(status, i, err ? : current_node, 1);
1791 err = move_pages_and_store_status(mm, current_node, &pagelist,
1792 status, start, i, nr_pages);
1795 current_node = NUMA_NO_NODE;
1798 /* Make sure we do not overwrite the existing error */
1799 err1 = move_pages_and_store_status(mm, current_node, &pagelist,
1800 status, start, i, nr_pages);
1808 * Determine the nodes of an array of pages and store it in an array of status.
1810 static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages,
1811 const void __user **pages, int *status)
1817 for (i = 0; i < nr_pages; i++) {
1818 unsigned long addr = (unsigned long)(*pages);
1819 struct vm_area_struct *vma;
1823 vma = find_vma(mm, addr);
1824 if (!vma || addr < vma->vm_start)
1827 /* FOLL_DUMP to ignore special (like zero) pages */
1828 page = follow_page(vma, addr, FOLL_DUMP);
1830 err = PTR_ERR(page);
1834 err = page ? page_to_nid(page) : -ENOENT;
1842 mmap_read_unlock(mm);
1846 * Determine the nodes of a user array of pages and store it in
1847 * a user array of status.
1849 static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages,
1850 const void __user * __user *pages,
1853 #define DO_PAGES_STAT_CHUNK_NR 16
1854 const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR];
1855 int chunk_status[DO_PAGES_STAT_CHUNK_NR];
1858 unsigned long chunk_nr;
1860 chunk_nr = nr_pages;
1861 if (chunk_nr > DO_PAGES_STAT_CHUNK_NR)
1862 chunk_nr = DO_PAGES_STAT_CHUNK_NR;
1864 if (copy_from_user(chunk_pages, pages, chunk_nr * sizeof(*chunk_pages)))
1867 do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status);
1869 if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status)))
1874 nr_pages -= chunk_nr;
1876 return nr_pages ? -EFAULT : 0;
1879 static struct mm_struct *find_mm_struct(pid_t pid, nodemask_t *mem_nodes)
1881 struct task_struct *task;
1882 struct mm_struct *mm;
1885 * There is no need to check if current process has the right to modify
1886 * the specified process when they are same.
1890 *mem_nodes = cpuset_mems_allowed(current);
1894 /* Find the mm_struct */
1896 task = find_task_by_vpid(pid);
1899 return ERR_PTR(-ESRCH);
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)) {
1909 mm = ERR_PTR(-EPERM);
1914 mm = ERR_PTR(security_task_movememory(task));
1917 *mem_nodes = cpuset_mems_allowed(task);
1918 mm = get_task_mm(task);
1920 put_task_struct(task);
1922 mm = ERR_PTR(-EINVAL);
1927 * Move a list of pages in the address space of the currently executing
1930 static int kernel_move_pages(pid_t pid, unsigned long nr_pages,
1931 const void __user * __user *pages,
1932 const int __user *nodes,
1933 int __user *status, int flags)
1935 struct mm_struct *mm;
1937 nodemask_t task_nodes;
1940 if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
1943 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1946 mm = find_mm_struct(pid, &task_nodes);
1951 err = do_pages_move(mm, task_nodes, nr_pages, pages,
1952 nodes, status, flags);
1954 err = do_pages_stat(mm, nr_pages, pages, status);
1960 SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages,
1961 const void __user * __user *, pages,
1962 const int __user *, nodes,
1963 int __user *, status, int, flags)
1965 return kernel_move_pages(pid, nr_pages, pages, nodes, status, flags);
1968 #ifdef CONFIG_COMPAT
1969 COMPAT_SYSCALL_DEFINE6(move_pages, pid_t, pid, compat_ulong_t, nr_pages,
1970 compat_uptr_t __user *, pages32,
1971 const int __user *, nodes,
1972 int __user *, status,
1975 const void __user * __user *pages;
1978 pages = compat_alloc_user_space(nr_pages * sizeof(void *));
1979 for (i = 0; i < nr_pages; i++) {
1982 if (get_user(p, pages32 + i) ||
1983 put_user(compat_ptr(p), pages + i))
1986 return kernel_move_pages(pid, nr_pages, pages, nodes, status, flags);
1988 #endif /* CONFIG_COMPAT */
1990 #ifdef CONFIG_NUMA_BALANCING
1992 * Returns true if this is a safe migration target node for misplaced NUMA
1993 * pages. Currently it only checks the watermarks which crude
1995 static bool migrate_balanced_pgdat(struct pglist_data *pgdat,
1996 unsigned long nr_migrate_pages)
2000 for (z = pgdat->nr_zones - 1; z >= 0; z--) {
2001 struct zone *zone = pgdat->node_zones + z;
2003 if (!populated_zone(zone))
2006 /* Avoid waking kswapd by allocating pages_to_migrate pages. */
2007 if (!zone_watermark_ok(zone, 0,
2008 high_wmark_pages(zone) +
2017 static struct page *alloc_misplaced_dst_page(struct page *page,
2020 int nid = (int) data;
2021 struct page *newpage;
2023 newpage = __alloc_pages_node(nid,
2024 (GFP_HIGHUSER_MOVABLE |
2025 __GFP_THISNODE | __GFP_NOMEMALLOC |
2026 __GFP_NORETRY | __GFP_NOWARN) &
2032 static int numamigrate_isolate_page(pg_data_t *pgdat, struct page *page)
2036 VM_BUG_ON_PAGE(compound_order(page) && !PageTransHuge(page), page);
2038 /* Avoid migrating to a node that is nearly full */
2039 if (!migrate_balanced_pgdat(pgdat, compound_nr(page)))
2042 if (isolate_lru_page(page))
2046 * migrate_misplaced_transhuge_page() skips page migration's usual
2047 * check on page_count(), so we must do it here, now that the page
2048 * has been isolated: a GUP pin, or any other pin, prevents migration.
2049 * The expected page count is 3: 1 for page's mapcount and 1 for the
2050 * caller's pin and 1 for the reference taken by isolate_lru_page().
2052 if (PageTransHuge(page) && page_count(page) != 3) {
2053 putback_lru_page(page);
2057 page_lru = page_is_file_lru(page);
2058 mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON + page_lru,
2059 thp_nr_pages(page));
2062 * Isolating the page has taken another reference, so the
2063 * caller's reference can be safely dropped without the page
2064 * disappearing underneath us during migration.
2070 bool pmd_trans_migrating(pmd_t pmd)
2072 struct page *page = pmd_page(pmd);
2073 return PageLocked(page);
2077 * Attempt to migrate a misplaced page to the specified destination
2078 * node. Caller is expected to have an elevated reference count on
2079 * the page that will be dropped by this function before returning.
2081 int migrate_misplaced_page(struct page *page, struct vm_area_struct *vma,
2084 pg_data_t *pgdat = NODE_DATA(node);
2087 LIST_HEAD(migratepages);
2090 * Don't migrate file pages that are mapped in multiple processes
2091 * with execute permissions as they are probably shared libraries.
2093 if (page_mapcount(page) != 1 && page_is_file_lru(page) &&
2094 (vma->vm_flags & VM_EXEC))
2098 * Also do not migrate dirty pages as not all filesystems can move
2099 * dirty pages in MIGRATE_ASYNC mode which is a waste of cycles.
2101 if (page_is_file_lru(page) && PageDirty(page))
2104 isolated = numamigrate_isolate_page(pgdat, page);
2108 list_add(&page->lru, &migratepages);
2109 nr_remaining = migrate_pages(&migratepages, alloc_misplaced_dst_page,
2110 NULL, node, MIGRATE_ASYNC,
2113 if (!list_empty(&migratepages)) {
2114 list_del(&page->lru);
2115 dec_node_page_state(page, NR_ISOLATED_ANON +
2116 page_is_file_lru(page));
2117 putback_lru_page(page);
2121 count_vm_numa_event(NUMA_PAGE_MIGRATE);
2122 BUG_ON(!list_empty(&migratepages));
2129 #endif /* CONFIG_NUMA_BALANCING */
2131 #if defined(CONFIG_NUMA_BALANCING) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
2133 * Migrates a THP to a given target node. page must be locked and is unlocked
2136 int migrate_misplaced_transhuge_page(struct mm_struct *mm,
2137 struct vm_area_struct *vma,
2138 pmd_t *pmd, pmd_t entry,
2139 unsigned long address,
2140 struct page *page, int node)
2143 pg_data_t *pgdat = NODE_DATA(node);
2145 struct page *new_page = NULL;
2146 int page_lru = page_is_file_lru(page);
2147 unsigned long start = address & HPAGE_PMD_MASK;
2149 new_page = alloc_pages_node(node,
2150 (GFP_TRANSHUGE_LIGHT | __GFP_THISNODE),
2154 prep_transhuge_page(new_page);
2156 isolated = numamigrate_isolate_page(pgdat, page);
2162 /* Prepare a page as a migration target */
2163 __SetPageLocked(new_page);
2164 if (PageSwapBacked(page))
2165 __SetPageSwapBacked(new_page);
2167 /* anon mapping, we can simply copy page->mapping to the new page: */
2168 new_page->mapping = page->mapping;
2169 new_page->index = page->index;
2170 /* flush the cache before copying using the kernel virtual address */
2171 flush_cache_range(vma, start, start + HPAGE_PMD_SIZE);
2172 migrate_page_copy(new_page, page);
2173 WARN_ON(PageLRU(new_page));
2175 /* Recheck the target PMD */
2176 ptl = pmd_lock(mm, pmd);
2177 if (unlikely(!pmd_same(*pmd, entry) || !page_ref_freeze(page, 2))) {
2180 /* Reverse changes made by migrate_page_copy() */
2181 if (TestClearPageActive(new_page))
2182 SetPageActive(page);
2183 if (TestClearPageUnevictable(new_page))
2184 SetPageUnevictable(page);
2186 unlock_page(new_page);
2187 put_page(new_page); /* Free it */
2189 /* Retake the callers reference and putback on LRU */
2191 putback_lru_page(page);
2192 mod_node_page_state(page_pgdat(page),
2193 NR_ISOLATED_ANON + page_lru, -HPAGE_PMD_NR);
2198 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
2199 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
2202 * Overwrite the old entry under pagetable lock and establish
2203 * the new PTE. Any parallel GUP will either observe the old
2204 * page blocking on the page lock, block on the page table
2205 * lock or observe the new page. The SetPageUptodate on the
2206 * new page and page_add_new_anon_rmap guarantee the copy is
2207 * visible before the pagetable update.
2209 page_add_anon_rmap(new_page, vma, start, true);
2211 * At this point the pmd is numa/protnone (i.e. non present) and the TLB
2212 * has already been flushed globally. So no TLB can be currently
2213 * caching this non present pmd mapping. There's no need to clear the
2214 * pmd before doing set_pmd_at(), nor to flush the TLB after
2215 * set_pmd_at(). Clearing the pmd here would introduce a race
2216 * condition against MADV_DONTNEED, because MADV_DONTNEED only holds the
2217 * mmap_lock for reading. If the pmd is set to NULL at any given time,
2218 * MADV_DONTNEED won't wait on the pmd lock and it'll skip clearing this
2221 set_pmd_at(mm, start, pmd, entry);
2222 update_mmu_cache_pmd(vma, address, &entry);
2224 page_ref_unfreeze(page, 2);
2225 mlock_migrate_page(new_page, page);
2226 page_remove_rmap(page, true);
2227 set_page_owner_migrate_reason(new_page, MR_NUMA_MISPLACED);
2231 /* Take an "isolate" reference and put new page on the LRU. */
2233 putback_lru_page(new_page);
2235 unlock_page(new_page);
2237 put_page(page); /* Drop the rmap reference */
2238 put_page(page); /* Drop the LRU isolation reference */
2240 count_vm_events(PGMIGRATE_SUCCESS, HPAGE_PMD_NR);
2241 count_vm_numa_events(NUMA_PAGE_MIGRATE, HPAGE_PMD_NR);
2243 mod_node_page_state(page_pgdat(page),
2244 NR_ISOLATED_ANON + page_lru,
2249 count_vm_events(PGMIGRATE_FAIL, HPAGE_PMD_NR);
2250 ptl = pmd_lock(mm, pmd);
2251 if (pmd_same(*pmd, entry)) {
2252 entry = pmd_modify(entry, vma->vm_page_prot);
2253 set_pmd_at(mm, start, pmd, entry);
2254 update_mmu_cache_pmd(vma, address, &entry);
2263 #endif /* CONFIG_NUMA_BALANCING */
2265 #endif /* CONFIG_NUMA */
2267 #ifdef CONFIG_DEVICE_PRIVATE
2268 static int migrate_vma_collect_hole(unsigned long start,
2270 __always_unused int depth,
2271 struct mm_walk *walk)
2273 struct migrate_vma *migrate = walk->private;
2276 /* Only allow populating anonymous memory. */
2277 if (!vma_is_anonymous(walk->vma)) {
2278 for (addr = start; addr < end; addr += PAGE_SIZE) {
2279 migrate->src[migrate->npages] = 0;
2280 migrate->dst[migrate->npages] = 0;
2286 for (addr = start; addr < end; addr += PAGE_SIZE) {
2287 migrate->src[migrate->npages] = MIGRATE_PFN_MIGRATE;
2288 migrate->dst[migrate->npages] = 0;
2296 static int migrate_vma_collect_skip(unsigned long start,
2298 struct mm_walk *walk)
2300 struct migrate_vma *migrate = walk->private;
2303 for (addr = start; addr < end; addr += PAGE_SIZE) {
2304 migrate->dst[migrate->npages] = 0;
2305 migrate->src[migrate->npages++] = 0;
2311 static int migrate_vma_collect_pmd(pmd_t *pmdp,
2312 unsigned long start,
2314 struct mm_walk *walk)
2316 struct migrate_vma *migrate = walk->private;
2317 struct vm_area_struct *vma = walk->vma;
2318 struct mm_struct *mm = vma->vm_mm;
2319 unsigned long addr = start, unmapped = 0;
2324 if (pmd_none(*pmdp))
2325 return migrate_vma_collect_hole(start, end, -1, walk);
2327 if (pmd_trans_huge(*pmdp)) {
2330 ptl = pmd_lock(mm, pmdp);
2331 if (unlikely(!pmd_trans_huge(*pmdp))) {
2336 page = pmd_page(*pmdp);
2337 if (is_huge_zero_page(page)) {
2339 split_huge_pmd(vma, pmdp, addr);
2340 if (pmd_trans_unstable(pmdp))
2341 return migrate_vma_collect_skip(start, end,
2348 if (unlikely(!trylock_page(page)))
2349 return migrate_vma_collect_skip(start, end,
2351 ret = split_huge_page(page);
2355 return migrate_vma_collect_skip(start, end,
2357 if (pmd_none(*pmdp))
2358 return migrate_vma_collect_hole(start, end, -1,
2363 if (unlikely(pmd_bad(*pmdp)))
2364 return migrate_vma_collect_skip(start, end, walk);
2366 ptep = pte_offset_map_lock(mm, pmdp, addr, &ptl);
2367 arch_enter_lazy_mmu_mode();
2369 for (; addr < end; addr += PAGE_SIZE, ptep++) {
2370 unsigned long mpfn = 0, pfn;
2377 if (pte_none(pte)) {
2378 if (vma_is_anonymous(vma)) {
2379 mpfn = MIGRATE_PFN_MIGRATE;
2385 if (!pte_present(pte)) {
2387 * Only care about unaddressable device page special
2388 * page table entry. Other special swap entries are not
2389 * migratable, and we ignore regular swapped page.
2391 entry = pte_to_swp_entry(pte);
2392 if (!is_device_private_entry(entry))
2395 page = device_private_entry_to_page(entry);
2396 if (!(migrate->flags &
2397 MIGRATE_VMA_SELECT_DEVICE_PRIVATE) ||
2398 page->pgmap->owner != migrate->pgmap_owner)
2401 mpfn = migrate_pfn(page_to_pfn(page)) |
2402 MIGRATE_PFN_MIGRATE;
2403 if (is_write_device_private_entry(entry))
2404 mpfn |= MIGRATE_PFN_WRITE;
2406 if (!(migrate->flags & MIGRATE_VMA_SELECT_SYSTEM))
2409 if (is_zero_pfn(pfn)) {
2410 mpfn = MIGRATE_PFN_MIGRATE;
2414 page = vm_normal_page(migrate->vma, addr, pte);
2415 mpfn = migrate_pfn(pfn) | MIGRATE_PFN_MIGRATE;
2416 mpfn |= pte_write(pte) ? MIGRATE_PFN_WRITE : 0;
2419 /* FIXME support THP */
2420 if (!page || !page->mapping || PageTransCompound(page)) {
2426 * By getting a reference on the page we pin it and that blocks
2427 * any kind of migration. Side effect is that it "freezes" the
2430 * We drop this reference after isolating the page from the lru
2431 * for non device page (device page are not on the lru and thus
2432 * can't be dropped from it).
2438 * Optimize for the common case where page is only mapped once
2439 * in one process. If we can lock the page, then we can safely
2440 * set up a special migration page table entry now.
2442 if (trylock_page(page)) {
2445 mpfn |= MIGRATE_PFN_LOCKED;
2446 ptep_get_and_clear(mm, addr, ptep);
2448 /* Setup special migration page table entry */
2449 entry = make_migration_entry(page, mpfn &
2451 swp_pte = swp_entry_to_pte(entry);
2452 if (pte_present(pte)) {
2453 if (pte_soft_dirty(pte))
2454 swp_pte = pte_swp_mksoft_dirty(swp_pte);
2455 if (pte_uffd_wp(pte))
2456 swp_pte = pte_swp_mkuffd_wp(swp_pte);
2458 if (pte_swp_soft_dirty(pte))
2459 swp_pte = pte_swp_mksoft_dirty(swp_pte);
2460 if (pte_swp_uffd_wp(pte))
2461 swp_pte = pte_swp_mkuffd_wp(swp_pte);
2463 set_pte_at(mm, addr, ptep, swp_pte);
2466 * This is like regular unmap: we remove the rmap and
2467 * drop page refcount. Page won't be freed, as we took
2468 * a reference just above.
2470 page_remove_rmap(page, false);
2473 if (pte_present(pte))
2478 migrate->dst[migrate->npages] = 0;
2479 migrate->src[migrate->npages++] = mpfn;
2481 arch_leave_lazy_mmu_mode();
2482 pte_unmap_unlock(ptep - 1, ptl);
2484 /* Only flush the TLB if we actually modified any entries */
2486 flush_tlb_range(walk->vma, start, end);
2491 static const struct mm_walk_ops migrate_vma_walk_ops = {
2492 .pmd_entry = migrate_vma_collect_pmd,
2493 .pte_hole = migrate_vma_collect_hole,
2497 * migrate_vma_collect() - collect pages over a range of virtual addresses
2498 * @migrate: migrate struct containing all migration information
2500 * This will walk the CPU page table. For each virtual address backed by a
2501 * valid page, it updates the src array and takes a reference on the page, in
2502 * order to pin the page until we lock it and unmap it.
2504 static void migrate_vma_collect(struct migrate_vma *migrate)
2506 struct mmu_notifier_range range;
2509 * Note that the pgmap_owner is passed to the mmu notifier callback so
2510 * that the registered device driver can skip invalidating device
2511 * private page mappings that won't be migrated.
2513 mmu_notifier_range_init_migrate(&range, 0, migrate->vma,
2514 migrate->vma->vm_mm, migrate->start, migrate->end,
2515 migrate->pgmap_owner);
2516 mmu_notifier_invalidate_range_start(&range);
2518 walk_page_range(migrate->vma->vm_mm, migrate->start, migrate->end,
2519 &migrate_vma_walk_ops, migrate);
2521 mmu_notifier_invalidate_range_end(&range);
2522 migrate->end = migrate->start + (migrate->npages << PAGE_SHIFT);
2526 * migrate_vma_check_page() - check if page is pinned or not
2527 * @page: struct page to check
2529 * Pinned pages cannot be migrated. This is the same test as in
2530 * migrate_page_move_mapping(), except that here we allow migration of a
2533 static bool migrate_vma_check_page(struct page *page)
2536 * One extra ref because caller holds an extra reference, either from
2537 * isolate_lru_page() for a regular page, or migrate_vma_collect() for
2543 * FIXME support THP (transparent huge page), it is bit more complex to
2544 * check them than regular pages, because they can be mapped with a pmd
2545 * or with a pte (split pte mapping).
2547 if (PageCompound(page))
2550 /* Page from ZONE_DEVICE have one extra reference */
2551 if (is_zone_device_page(page)) {
2553 * Private page can never be pin as they have no valid pte and
2554 * GUP will fail for those. Yet if there is a pending migration
2555 * a thread might try to wait on the pte migration entry and
2556 * will bump the page reference count. Sadly there is no way to
2557 * differentiate a regular pin from migration wait. Hence to
2558 * avoid 2 racing thread trying to migrate back to CPU to enter
2559 * infinite loop (one stoping migration because the other is
2560 * waiting on pte migration entry). We always return true here.
2562 * FIXME proper solution is to rework migration_entry_wait() so
2563 * it does not need to take a reference on page.
2565 return is_device_private_page(page);
2568 /* For file back page */
2569 if (page_mapping(page))
2570 extra += 1 + page_has_private(page);
2572 if ((page_count(page) - extra) > page_mapcount(page))
2579 * migrate_vma_prepare() - lock pages and isolate them from the lru
2580 * @migrate: migrate struct containing all migration information
2582 * This locks pages that have been collected by migrate_vma_collect(). Once each
2583 * page is locked it is isolated from the lru (for non-device pages). Finally,
2584 * the ref taken by migrate_vma_collect() is dropped, as locked pages cannot be
2585 * migrated by concurrent kernel threads.
2587 static void migrate_vma_prepare(struct migrate_vma *migrate)
2589 const unsigned long npages = migrate->npages;
2590 const unsigned long start = migrate->start;
2591 unsigned long addr, i, restore = 0;
2592 bool allow_drain = true;
2596 for (i = 0; (i < npages) && migrate->cpages; i++) {
2597 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2603 if (!(migrate->src[i] & MIGRATE_PFN_LOCKED)) {
2605 * Because we are migrating several pages there can be
2606 * a deadlock between 2 concurrent migration where each
2607 * are waiting on each other page lock.
2609 * Make migrate_vma() a best effort thing and backoff
2610 * for any page we can not lock right away.
2612 if (!trylock_page(page)) {
2613 migrate->src[i] = 0;
2619 migrate->src[i] |= MIGRATE_PFN_LOCKED;
2622 /* ZONE_DEVICE pages are not on LRU */
2623 if (!is_zone_device_page(page)) {
2624 if (!PageLRU(page) && allow_drain) {
2625 /* Drain CPU's pagevec */
2626 lru_add_drain_all();
2627 allow_drain = false;
2630 if (isolate_lru_page(page)) {
2632 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2636 migrate->src[i] = 0;
2644 /* Drop the reference we took in collect */
2648 if (!migrate_vma_check_page(page)) {
2650 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2654 if (!is_zone_device_page(page)) {
2656 putback_lru_page(page);
2659 migrate->src[i] = 0;
2663 if (!is_zone_device_page(page))
2664 putback_lru_page(page);
2671 for (i = 0, addr = start; i < npages && restore; i++, addr += PAGE_SIZE) {
2672 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2674 if (!page || (migrate->src[i] & MIGRATE_PFN_MIGRATE))
2677 remove_migration_pte(page, migrate->vma, addr, page);
2679 migrate->src[i] = 0;
2687 * migrate_vma_unmap() - replace page mapping with special migration pte entry
2688 * @migrate: migrate struct containing all migration information
2690 * Replace page mapping (CPU page table pte) with a special migration pte entry
2691 * and check again if it has been pinned. Pinned pages are restored because we
2692 * cannot migrate them.
2694 * This is the last step before we call the device driver callback to allocate
2695 * destination memory and copy contents of original page over to new page.
2697 static void migrate_vma_unmap(struct migrate_vma *migrate)
2699 int flags = TTU_MIGRATION | TTU_IGNORE_MLOCK;
2700 const unsigned long npages = migrate->npages;
2701 const unsigned long start = migrate->start;
2702 unsigned long addr, i, restore = 0;
2704 for (i = 0; i < npages; i++) {
2705 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2707 if (!page || !(migrate->src[i] & MIGRATE_PFN_MIGRATE))
2710 if (page_mapped(page)) {
2711 try_to_unmap(page, flags);
2712 if (page_mapped(page))
2716 if (migrate_vma_check_page(page))
2720 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2725 for (addr = start, i = 0; i < npages && restore; addr += PAGE_SIZE, i++) {
2726 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2728 if (!page || (migrate->src[i] & MIGRATE_PFN_MIGRATE))
2731 remove_migration_ptes(page, page, false);
2733 migrate->src[i] = 0;
2737 if (is_zone_device_page(page))
2740 putback_lru_page(page);
2745 * migrate_vma_setup() - prepare to migrate a range of memory
2746 * @args: contains the vma, start, and pfns arrays for the migration
2748 * Returns: negative errno on failures, 0 when 0 or more pages were migrated
2751 * Prepare to migrate a range of memory virtual address range by collecting all
2752 * the pages backing each virtual address in the range, saving them inside the
2753 * src array. Then lock those pages and unmap them. Once the pages are locked
2754 * and unmapped, check whether each page is pinned or not. Pages that aren't
2755 * pinned have the MIGRATE_PFN_MIGRATE flag set (by this function) in the
2756 * corresponding src array entry. Then restores any pages that are pinned, by
2757 * remapping and unlocking those pages.
2759 * The caller should then allocate destination memory and copy source memory to
2760 * it for all those entries (ie with MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE
2761 * flag set). Once these are allocated and copied, the caller must update each
2762 * corresponding entry in the dst array with the pfn value of the destination
2763 * page and with the MIGRATE_PFN_VALID and MIGRATE_PFN_LOCKED flags set
2764 * (destination pages must have their struct pages locked, via lock_page()).
2766 * Note that the caller does not have to migrate all the pages that are marked
2767 * with MIGRATE_PFN_MIGRATE flag in src array unless this is a migration from
2768 * device memory to system memory. If the caller cannot migrate a device page
2769 * back to system memory, then it must return VM_FAULT_SIGBUS, which has severe
2770 * consequences for the userspace process, so it must be avoided if at all
2773 * For empty entries inside CPU page table (pte_none() or pmd_none() is true) we
2774 * do set MIGRATE_PFN_MIGRATE flag inside the corresponding source array thus
2775 * allowing the caller to allocate device memory for those unback virtual
2776 * address. For this the caller simply has to allocate device memory and
2777 * properly set the destination entry like for regular migration. Note that
2778 * this can still fails and thus inside the device driver must check if the
2779 * migration was successful for those entries after calling migrate_vma_pages()
2780 * just like for regular migration.
2782 * After that, the callers must call migrate_vma_pages() to go over each entry
2783 * in the src array that has the MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE flag
2784 * set. If the corresponding entry in dst array has MIGRATE_PFN_VALID flag set,
2785 * then migrate_vma_pages() to migrate struct page information from the source
2786 * struct page to the destination struct page. If it fails to migrate the
2787 * struct page information, then it clears the MIGRATE_PFN_MIGRATE flag in the
2790 * At this point all successfully migrated pages have an entry in the src
2791 * array with MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE flag set and the dst
2792 * array entry with MIGRATE_PFN_VALID flag set.
2794 * Once migrate_vma_pages() returns the caller may inspect which pages were
2795 * successfully migrated, and which were not. Successfully migrated pages will
2796 * have the MIGRATE_PFN_MIGRATE flag set for their src array entry.
2798 * It is safe to update device page table after migrate_vma_pages() because
2799 * both destination and source page are still locked, and the mmap_lock is held
2800 * in read mode (hence no one can unmap the range being migrated).
2802 * Once the caller is done cleaning up things and updating its page table (if it
2803 * chose to do so, this is not an obligation) it finally calls
2804 * migrate_vma_finalize() to update the CPU page table to point to new pages
2805 * for successfully migrated pages or otherwise restore the CPU page table to
2806 * point to the original source pages.
2808 int migrate_vma_setup(struct migrate_vma *args)
2810 long nr_pages = (args->end - args->start) >> PAGE_SHIFT;
2812 args->start &= PAGE_MASK;
2813 args->end &= PAGE_MASK;
2814 if (!args->vma || is_vm_hugetlb_page(args->vma) ||
2815 (args->vma->vm_flags & VM_SPECIAL) || vma_is_dax(args->vma))
2819 if (args->start < args->vma->vm_start ||
2820 args->start >= args->vma->vm_end)
2822 if (args->end <= args->vma->vm_start || args->end > args->vma->vm_end)
2824 if (!args->src || !args->dst)
2827 memset(args->src, 0, sizeof(*args->src) * nr_pages);
2831 migrate_vma_collect(args);
2834 migrate_vma_prepare(args);
2836 migrate_vma_unmap(args);
2839 * At this point pages are locked and unmapped, and thus they have
2840 * stable content and can safely be copied to destination memory that
2841 * is allocated by the drivers.
2846 EXPORT_SYMBOL(migrate_vma_setup);
2849 * This code closely matches the code in:
2850 * __handle_mm_fault()
2851 * handle_pte_fault()
2852 * do_anonymous_page()
2853 * to map in an anonymous zero page but the struct page will be a ZONE_DEVICE
2856 static void migrate_vma_insert_page(struct migrate_vma *migrate,
2862 struct vm_area_struct *vma = migrate->vma;
2863 struct mm_struct *mm = vma->vm_mm;
2873 /* Only allow populating anonymous memory */
2874 if (!vma_is_anonymous(vma))
2877 pgdp = pgd_offset(mm, addr);
2878 p4dp = p4d_alloc(mm, pgdp, addr);
2881 pudp = pud_alloc(mm, p4dp, addr);
2884 pmdp = pmd_alloc(mm, pudp, addr);
2888 if (pmd_trans_huge(*pmdp) || pmd_devmap(*pmdp))
2892 * Use pte_alloc() instead of pte_alloc_map(). We can't run
2893 * pte_offset_map() on pmds where a huge pmd might be created
2894 * from a different thread.
2896 * pte_alloc_map() is safe to use under mmap_write_lock(mm) or when
2897 * parallel threads are excluded by other means.
2899 * Here we only have mmap_read_lock(mm).
2901 if (pte_alloc(mm, pmdp))
2904 /* See the comment in pte_alloc_one_map() */
2905 if (unlikely(pmd_trans_unstable(pmdp)))
2908 if (unlikely(anon_vma_prepare(vma)))
2910 if (mem_cgroup_charge(page, vma->vm_mm, GFP_KERNEL))
2914 * The memory barrier inside __SetPageUptodate makes sure that
2915 * preceding stores to the page contents become visible before
2916 * the set_pte_at() write.
2918 __SetPageUptodate(page);
2920 if (is_zone_device_page(page)) {
2921 if (is_device_private_page(page)) {
2922 swp_entry_t swp_entry;
2924 swp_entry = make_device_private_entry(page, vma->vm_flags & VM_WRITE);
2925 entry = swp_entry_to_pte(swp_entry);
2928 entry = mk_pte(page, vma->vm_page_prot);
2929 if (vma->vm_flags & VM_WRITE)
2930 entry = pte_mkwrite(pte_mkdirty(entry));
2933 ptep = pte_offset_map_lock(mm, pmdp, addr, &ptl);
2935 if (check_stable_address_space(mm))
2938 if (pte_present(*ptep)) {
2939 unsigned long pfn = pte_pfn(*ptep);
2941 if (!is_zero_pfn(pfn))
2944 } else if (!pte_none(*ptep))
2948 * Check for userfaultfd but do not deliver the fault. Instead,
2951 if (userfaultfd_missing(vma))
2954 inc_mm_counter(mm, MM_ANONPAGES);
2955 page_add_new_anon_rmap(page, vma, addr, false);
2956 if (!is_zone_device_page(page))
2957 lru_cache_add_inactive_or_unevictable(page, vma);
2961 flush_cache_page(vma, addr, pte_pfn(*ptep));
2962 ptep_clear_flush_notify(vma, addr, ptep);
2963 set_pte_at_notify(mm, addr, ptep, entry);
2964 update_mmu_cache(vma, addr, ptep);
2966 /* No need to invalidate - it was non-present before */
2967 set_pte_at(mm, addr, ptep, entry);
2968 update_mmu_cache(vma, addr, ptep);
2971 pte_unmap_unlock(ptep, ptl);
2972 *src = MIGRATE_PFN_MIGRATE;
2976 pte_unmap_unlock(ptep, ptl);
2978 *src &= ~MIGRATE_PFN_MIGRATE;
2982 * migrate_vma_pages() - migrate meta-data from src page to dst page
2983 * @migrate: migrate struct containing all migration information
2985 * This migrates struct page meta-data from source struct page to destination
2986 * struct page. This effectively finishes the migration from source page to the
2989 void migrate_vma_pages(struct migrate_vma *migrate)
2991 const unsigned long npages = migrate->npages;
2992 const unsigned long start = migrate->start;
2993 struct mmu_notifier_range range;
2994 unsigned long addr, i;
2995 bool notified = false;
2997 for (i = 0, addr = start; i < npages; addr += PAGE_SIZE, i++) {
2998 struct page *newpage = migrate_pfn_to_page(migrate->dst[i]);
2999 struct page *page = migrate_pfn_to_page(migrate->src[i]);
3000 struct address_space *mapping;
3004 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
3009 if (!(migrate->src[i] & MIGRATE_PFN_MIGRATE))
3014 mmu_notifier_range_init(&range,
3015 MMU_NOTIFY_CLEAR, 0,
3017 migrate->vma->vm_mm,
3018 addr, migrate->end);
3019 mmu_notifier_invalidate_range_start(&range);
3021 migrate_vma_insert_page(migrate, addr, newpage,
3027 mapping = page_mapping(page);
3029 if (is_zone_device_page(newpage)) {
3030 if (is_device_private_page(newpage)) {
3032 * For now only support private anonymous when
3033 * migrating to un-addressable device memory.
3036 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
3041 * Other types of ZONE_DEVICE page are not
3044 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
3049 r = migrate_page(mapping, newpage, page, MIGRATE_SYNC_NO_COPY);
3050 if (r != MIGRATEPAGE_SUCCESS)
3051 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
3055 * No need to double call mmu_notifier->invalidate_range() callback as
3056 * the above ptep_clear_flush_notify() inside migrate_vma_insert_page()
3057 * did already call it.
3060 mmu_notifier_invalidate_range_only_end(&range);
3062 EXPORT_SYMBOL(migrate_vma_pages);
3065 * migrate_vma_finalize() - restore CPU page table entry
3066 * @migrate: migrate struct containing all migration information
3068 * This replaces the special migration pte entry with either a mapping to the
3069 * new page if migration was successful for that page, or to the original page
3072 * This also unlocks the pages and puts them back on the lru, or drops the extra
3073 * refcount, for device pages.
3075 void migrate_vma_finalize(struct migrate_vma *migrate)
3077 const unsigned long npages = migrate->npages;
3080 for (i = 0; i < npages; i++) {
3081 struct page *newpage = migrate_pfn_to_page(migrate->dst[i]);
3082 struct page *page = migrate_pfn_to_page(migrate->src[i]);
3086 unlock_page(newpage);
3092 if (!(migrate->src[i] & MIGRATE_PFN_MIGRATE) || !newpage) {
3094 unlock_page(newpage);
3100 remove_migration_ptes(page, newpage, false);
3103 if (is_zone_device_page(page))
3106 putback_lru_page(page);
3108 if (newpage != page) {
3109 unlock_page(newpage);
3110 if (is_zone_device_page(newpage))
3113 putback_lru_page(newpage);
3117 EXPORT_SYMBOL(migrate_vma_finalize);
3118 #endif /* CONFIG_DEVICE_PRIVATE */