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>
60 int isolate_movable_page(struct page *page, isolate_mode_t mode)
62 struct address_space *mapping;
65 * Avoid burning cycles with pages that are yet under __free_pages(),
66 * or just got freed under us.
68 * In case we 'win' a race for a movable page being freed under us and
69 * raise its refcount preventing __free_pages() from doing its job
70 * the put_page() at the end of this block will take care of
71 * release this page, thus avoiding a nasty leakage.
73 if (unlikely(!get_page_unless_zero(page)))
77 * Check PageMovable before holding a PG_lock because page's owner
78 * assumes anybody doesn't touch PG_lock of newly allocated page
79 * so unconditionally grabbing the lock ruins page's owner side.
81 if (unlikely(!__PageMovable(page)))
84 * As movable pages are not isolated from LRU lists, concurrent
85 * compaction threads can race against page migration functions
86 * as well as race against the releasing a page.
88 * In order to avoid having an already isolated movable page
89 * being (wrongly) re-isolated while it is under migration,
90 * or to avoid attempting to isolate pages being released,
91 * lets be sure we have the page lock
92 * before proceeding with the movable page isolation steps.
94 if (unlikely(!trylock_page(page)))
97 if (!PageMovable(page) || PageIsolated(page))
100 mapping = page_mapping(page);
101 VM_BUG_ON_PAGE(!mapping, page);
103 if (!mapping->a_ops->isolate_page(page, mode))
104 goto out_no_isolated;
106 /* Driver shouldn't use PG_isolated bit of page->flags */
107 WARN_ON_ONCE(PageIsolated(page));
108 __SetPageIsolated(page);
121 static void putback_movable_page(struct page *page)
123 struct address_space *mapping;
125 mapping = page_mapping(page);
126 mapping->a_ops->putback_page(page);
127 __ClearPageIsolated(page);
131 * Put previously isolated pages back onto the appropriate lists
132 * from where they were once taken off for compaction/migration.
134 * This function shall be used whenever the isolated pageset has been
135 * built from lru, balloon, hugetlbfs page. See isolate_migratepages_range()
136 * and isolate_huge_page().
138 void putback_movable_pages(struct list_head *l)
143 list_for_each_entry_safe(page, page2, l, lru) {
144 if (unlikely(PageHuge(page))) {
145 putback_active_hugepage(page);
148 list_del(&page->lru);
150 * We isolated non-lru movable page so here we can use
151 * __PageMovable because LRU page's mapping cannot have
152 * PAGE_MAPPING_MOVABLE.
154 if (unlikely(__PageMovable(page))) {
155 VM_BUG_ON_PAGE(!PageIsolated(page), page);
157 if (PageMovable(page))
158 putback_movable_page(page);
160 __ClearPageIsolated(page);
164 mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON +
165 page_is_file_lru(page), -thp_nr_pages(page));
166 putback_lru_page(page);
172 * Restore a potential migration pte to a working pte entry
174 static bool remove_migration_pte(struct page *page, struct vm_area_struct *vma,
175 unsigned long addr, void *old)
177 struct page_vma_mapped_walk pvmw = {
181 .flags = PVMW_SYNC | PVMW_MIGRATION,
187 VM_BUG_ON_PAGE(PageTail(page), page);
188 while (page_vma_mapped_walk(&pvmw)) {
192 new = page - pvmw.page->index +
193 linear_page_index(vma, pvmw.address);
195 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
196 /* PMD-mapped THP migration entry */
198 VM_BUG_ON_PAGE(PageHuge(page) || !PageTransCompound(page), page);
199 remove_migration_pmd(&pvmw, new);
205 pte = pte_mkold(mk_pte(new, READ_ONCE(vma->vm_page_prot)));
206 if (pte_swp_soft_dirty(*pvmw.pte))
207 pte = pte_mksoft_dirty(pte);
210 * Recheck VMA as permissions can change since migration started
212 entry = pte_to_swp_entry(*pvmw.pte);
213 if (is_write_migration_entry(entry))
214 pte = maybe_mkwrite(pte, vma);
215 else if (pte_swp_uffd_wp(*pvmw.pte))
216 pte = pte_mkuffd_wp(pte);
218 if (unlikely(is_device_private_page(new))) {
219 entry = make_device_private_entry(new, pte_write(pte));
220 pte = swp_entry_to_pte(entry);
221 if (pte_swp_soft_dirty(*pvmw.pte))
222 pte = pte_swp_mksoft_dirty(pte);
223 if (pte_swp_uffd_wp(*pvmw.pte))
224 pte = pte_swp_mkuffd_wp(pte);
227 #ifdef CONFIG_HUGETLB_PAGE
229 pte = pte_mkhuge(pte);
230 pte = arch_make_huge_pte(pte, vma, new, 0);
231 set_huge_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte);
233 hugepage_add_anon_rmap(new, vma, pvmw.address);
235 page_dup_rmap(new, true);
239 set_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte);
242 page_add_anon_rmap(new, vma, pvmw.address, false);
244 page_add_file_rmap(new, false);
246 if (vma->vm_flags & VM_LOCKED && !PageTransCompound(new))
249 if (PageTransHuge(page) && PageMlocked(page))
250 clear_page_mlock(page);
252 /* No need to invalidate - it was non-present before */
253 update_mmu_cache(vma, pvmw.address, pvmw.pte);
260 * Get rid of all migration entries and replace them by
261 * references to the indicated page.
263 void remove_migration_ptes(struct page *old, struct page *new, bool locked)
265 struct rmap_walk_control rwc = {
266 .rmap_one = remove_migration_pte,
271 rmap_walk_locked(new, &rwc);
273 rmap_walk(new, &rwc);
277 * Something used the pte of a page under migration. We need to
278 * get to the page and wait until migration is finished.
279 * When we return from this function the fault will be retried.
281 void __migration_entry_wait(struct mm_struct *mm, pte_t *ptep,
290 if (!is_swap_pte(pte))
293 entry = pte_to_swp_entry(pte);
294 if (!is_migration_entry(entry))
297 page = migration_entry_to_page(entry);
298 page = compound_head(page);
301 * Once page cache replacement of page migration started, page_count
302 * is zero; but we must not call put_and_wait_on_page_locked() without
303 * a ref. Use get_page_unless_zero(), and just fault again if it fails.
305 if (!get_page_unless_zero(page))
307 pte_unmap_unlock(ptep, ptl);
308 put_and_wait_on_page_locked(page, TASK_UNINTERRUPTIBLE);
311 pte_unmap_unlock(ptep, ptl);
314 void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
315 unsigned long address)
317 spinlock_t *ptl = pte_lockptr(mm, pmd);
318 pte_t *ptep = pte_offset_map(pmd, address);
319 __migration_entry_wait(mm, ptep, ptl);
322 void migration_entry_wait_huge(struct vm_area_struct *vma,
323 struct mm_struct *mm, pte_t *pte)
325 spinlock_t *ptl = huge_pte_lockptr(hstate_vma(vma), mm, pte);
326 __migration_entry_wait(mm, pte, ptl);
329 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
330 void pmd_migration_entry_wait(struct mm_struct *mm, pmd_t *pmd)
335 ptl = pmd_lock(mm, pmd);
336 if (!is_pmd_migration_entry(*pmd))
338 page = migration_entry_to_page(pmd_to_swp_entry(*pmd));
339 if (!get_page_unless_zero(page))
342 put_and_wait_on_page_locked(page, TASK_UNINTERRUPTIBLE);
349 static int expected_page_refs(struct address_space *mapping, struct page *page)
351 int expected_count = 1;
354 * Device private pages have an extra refcount as they are
357 expected_count += is_device_private_page(page);
359 expected_count += thp_nr_pages(page) + page_has_private(page);
361 return expected_count;
365 * Replace the page in the mapping.
367 * The number of remaining references must be:
368 * 1 for anonymous pages without a mapping
369 * 2 for pages with a mapping
370 * 3 for pages with a mapping and PagePrivate/PagePrivate2 set.
372 int migrate_page_move_mapping(struct address_space *mapping,
373 struct page *newpage, struct page *page, int extra_count)
375 XA_STATE(xas, &mapping->i_pages, page_index(page));
376 struct zone *oldzone, *newzone;
378 int expected_count = expected_page_refs(mapping, page) + extra_count;
379 int nr = thp_nr_pages(page);
382 /* Anonymous page without mapping */
383 if (page_count(page) != expected_count)
386 /* No turning back from here */
387 newpage->index = page->index;
388 newpage->mapping = page->mapping;
389 if (PageSwapBacked(page))
390 __SetPageSwapBacked(newpage);
392 return MIGRATEPAGE_SUCCESS;
395 oldzone = page_zone(page);
396 newzone = page_zone(newpage);
399 if (page_count(page) != expected_count || xas_load(&xas) != page) {
400 xas_unlock_irq(&xas);
404 if (!page_ref_freeze(page, expected_count)) {
405 xas_unlock_irq(&xas);
410 * Now we know that no one else is looking at the page:
411 * no turning back from here.
413 newpage->index = page->index;
414 newpage->mapping = page->mapping;
415 page_ref_add(newpage, nr); /* add cache reference */
416 if (PageSwapBacked(page)) {
417 __SetPageSwapBacked(newpage);
418 if (PageSwapCache(page)) {
419 SetPageSwapCache(newpage);
420 set_page_private(newpage, page_private(page));
423 VM_BUG_ON_PAGE(PageSwapCache(page), page);
426 /* Move dirty while page refs frozen and newpage not yet exposed */
427 dirty = PageDirty(page);
429 ClearPageDirty(page);
430 SetPageDirty(newpage);
433 xas_store(&xas, newpage);
434 if (PageTransHuge(page)) {
437 for (i = 1; i < nr; i++) {
439 xas_store(&xas, newpage);
444 * Drop cache reference from old page by unfreezing
445 * to one less reference.
446 * We know this isn't the last reference.
448 page_ref_unfreeze(page, expected_count - nr);
451 /* Leave irq disabled to prevent preemption while updating stats */
454 * If moved to a different zone then also account
455 * the page for that zone. Other VM counters will be
456 * taken care of when we establish references to the
457 * new page and drop references to the old page.
459 * Note that anonymous pages are accounted for
460 * via NR_FILE_PAGES and NR_ANON_MAPPED if they
461 * are mapped to swap space.
463 if (newzone != oldzone) {
464 struct lruvec *old_lruvec, *new_lruvec;
465 struct mem_cgroup *memcg;
467 memcg = page_memcg(page);
468 old_lruvec = mem_cgroup_lruvec(memcg, oldzone->zone_pgdat);
469 new_lruvec = mem_cgroup_lruvec(memcg, newzone->zone_pgdat);
471 __mod_lruvec_state(old_lruvec, NR_FILE_PAGES, -nr);
472 __mod_lruvec_state(new_lruvec, NR_FILE_PAGES, nr);
473 if (PageSwapBacked(page) && !PageSwapCache(page)) {
474 __mod_lruvec_state(old_lruvec, NR_SHMEM, -nr);
475 __mod_lruvec_state(new_lruvec, NR_SHMEM, nr);
478 if (PageSwapCache(page)) {
479 __mod_lruvec_state(old_lruvec, NR_SWAPCACHE, -nr);
480 __mod_lruvec_state(new_lruvec, NR_SWAPCACHE, nr);
483 if (dirty && mapping_can_writeback(mapping)) {
484 __mod_lruvec_state(old_lruvec, NR_FILE_DIRTY, -nr);
485 __mod_zone_page_state(oldzone, NR_ZONE_WRITE_PENDING, -nr);
486 __mod_lruvec_state(new_lruvec, NR_FILE_DIRTY, nr);
487 __mod_zone_page_state(newzone, NR_ZONE_WRITE_PENDING, nr);
492 return MIGRATEPAGE_SUCCESS;
494 EXPORT_SYMBOL(migrate_page_move_mapping);
497 * The expected number of remaining references is the same as that
498 * of migrate_page_move_mapping().
500 int migrate_huge_page_move_mapping(struct address_space *mapping,
501 struct page *newpage, struct page *page)
503 XA_STATE(xas, &mapping->i_pages, page_index(page));
507 expected_count = 2 + page_has_private(page);
508 if (page_count(page) != expected_count || xas_load(&xas) != page) {
509 xas_unlock_irq(&xas);
513 if (!page_ref_freeze(page, expected_count)) {
514 xas_unlock_irq(&xas);
518 newpage->index = page->index;
519 newpage->mapping = page->mapping;
523 xas_store(&xas, newpage);
525 page_ref_unfreeze(page, expected_count - 1);
527 xas_unlock_irq(&xas);
529 return MIGRATEPAGE_SUCCESS;
533 * Gigantic pages are so large that we do not guarantee that page++ pointer
534 * arithmetic will work across the entire page. We need something more
537 static void __copy_gigantic_page(struct page *dst, struct page *src,
541 struct page *dst_base = dst;
542 struct page *src_base = src;
544 for (i = 0; i < nr_pages; ) {
546 copy_highpage(dst, src);
549 dst = mem_map_next(dst, dst_base, i);
550 src = mem_map_next(src, src_base, i);
554 static void copy_huge_page(struct page *dst, struct page *src)
561 struct hstate *h = page_hstate(src);
562 nr_pages = pages_per_huge_page(h);
564 if (unlikely(nr_pages > MAX_ORDER_NR_PAGES)) {
565 __copy_gigantic_page(dst, src, nr_pages);
570 BUG_ON(!PageTransHuge(src));
571 nr_pages = thp_nr_pages(src);
574 for (i = 0; i < nr_pages; i++) {
576 copy_highpage(dst + i, src + i);
581 * Copy the page to its new location
583 void migrate_page_states(struct page *newpage, struct page *page)
588 SetPageError(newpage);
589 if (PageReferenced(page))
590 SetPageReferenced(newpage);
591 if (PageUptodate(page))
592 SetPageUptodate(newpage);
593 if (TestClearPageActive(page)) {
594 VM_BUG_ON_PAGE(PageUnevictable(page), page);
595 SetPageActive(newpage);
596 } else if (TestClearPageUnevictable(page))
597 SetPageUnevictable(newpage);
598 if (PageWorkingset(page))
599 SetPageWorkingset(newpage);
600 if (PageChecked(page))
601 SetPageChecked(newpage);
602 if (PageMappedToDisk(page))
603 SetPageMappedToDisk(newpage);
605 /* Move dirty on pages not done by migrate_page_move_mapping() */
607 SetPageDirty(newpage);
609 if (page_is_young(page))
610 set_page_young(newpage);
611 if (page_is_idle(page))
612 set_page_idle(newpage);
615 * Copy NUMA information to the new page, to prevent over-eager
616 * future migrations of this same page.
618 cpupid = page_cpupid_xchg_last(page, -1);
619 page_cpupid_xchg_last(newpage, cpupid);
621 ksm_migrate_page(newpage, page);
623 * Please do not reorder this without considering how mm/ksm.c's
624 * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache().
626 if (PageSwapCache(page))
627 ClearPageSwapCache(page);
628 ClearPagePrivate(page);
630 /* page->private contains hugetlb specific flags */
632 set_page_private(page, 0);
635 * If any waiters have accumulated on the new page then
638 if (PageWriteback(newpage))
639 end_page_writeback(newpage);
642 * PG_readahead shares the same bit with PG_reclaim. The above
643 * end_page_writeback() may clear PG_readahead mistakenly, so set the
646 if (PageReadahead(page))
647 SetPageReadahead(newpage);
649 copy_page_owner(page, newpage);
652 mem_cgroup_migrate(page, newpage);
654 EXPORT_SYMBOL(migrate_page_states);
656 void migrate_page_copy(struct page *newpage, struct page *page)
658 if (PageHuge(page) || PageTransHuge(page))
659 copy_huge_page(newpage, page);
661 copy_highpage(newpage, page);
663 migrate_page_states(newpage, page);
665 EXPORT_SYMBOL(migrate_page_copy);
667 /************************************************************
668 * Migration functions
669 ***********************************************************/
672 * Common logic to directly migrate a single LRU page suitable for
673 * pages that do not use PagePrivate/PagePrivate2.
675 * Pages are locked upon entry and exit.
677 int migrate_page(struct address_space *mapping,
678 struct page *newpage, struct page *page,
679 enum migrate_mode mode)
683 BUG_ON(PageWriteback(page)); /* Writeback must be complete */
685 rc = migrate_page_move_mapping(mapping, newpage, page, 0);
687 if (rc != MIGRATEPAGE_SUCCESS)
690 if (mode != MIGRATE_SYNC_NO_COPY)
691 migrate_page_copy(newpage, page);
693 migrate_page_states(newpage, page);
694 return MIGRATEPAGE_SUCCESS;
696 EXPORT_SYMBOL(migrate_page);
699 /* Returns true if all buffers are successfully locked */
700 static bool buffer_migrate_lock_buffers(struct buffer_head *head,
701 enum migrate_mode mode)
703 struct buffer_head *bh = head;
705 /* Simple case, sync compaction */
706 if (mode != MIGRATE_ASYNC) {
709 bh = bh->b_this_page;
711 } while (bh != head);
716 /* async case, we cannot block on lock_buffer so use trylock_buffer */
718 if (!trylock_buffer(bh)) {
720 * We failed to lock the buffer and cannot stall in
721 * async migration. Release the taken locks
723 struct buffer_head *failed_bh = bh;
725 while (bh != failed_bh) {
727 bh = bh->b_this_page;
732 bh = bh->b_this_page;
733 } while (bh != head);
737 static int __buffer_migrate_page(struct address_space *mapping,
738 struct page *newpage, struct page *page, enum migrate_mode mode,
741 struct buffer_head *bh, *head;
745 if (!page_has_buffers(page))
746 return migrate_page(mapping, newpage, page, mode);
748 /* Check whether page does not have extra refs before we do more work */
749 expected_count = expected_page_refs(mapping, page);
750 if (page_count(page) != expected_count)
753 head = page_buffers(page);
754 if (!buffer_migrate_lock_buffers(head, mode))
759 bool invalidated = false;
763 spin_lock(&mapping->private_lock);
766 if (atomic_read(&bh->b_count)) {
770 bh = bh->b_this_page;
771 } while (bh != head);
777 spin_unlock(&mapping->private_lock);
778 invalidate_bh_lrus();
780 goto recheck_buffers;
784 rc = migrate_page_move_mapping(mapping, newpage, page, 0);
785 if (rc != MIGRATEPAGE_SUCCESS)
788 attach_page_private(newpage, detach_page_private(page));
792 set_bh_page(bh, newpage, bh_offset(bh));
793 bh = bh->b_this_page;
795 } while (bh != head);
797 if (mode != MIGRATE_SYNC_NO_COPY)
798 migrate_page_copy(newpage, page);
800 migrate_page_states(newpage, page);
802 rc = MIGRATEPAGE_SUCCESS;
805 spin_unlock(&mapping->private_lock);
809 bh = bh->b_this_page;
811 } while (bh != head);
817 * Migration function for pages with buffers. This function can only be used
818 * if the underlying filesystem guarantees that no other references to "page"
819 * exist. For example attached buffer heads are accessed only under page lock.
821 int buffer_migrate_page(struct address_space *mapping,
822 struct page *newpage, struct page *page, enum migrate_mode mode)
824 return __buffer_migrate_page(mapping, newpage, page, mode, false);
826 EXPORT_SYMBOL(buffer_migrate_page);
829 * Same as above except that this variant is more careful and checks that there
830 * are also no buffer head references. This function is the right one for
831 * mappings where buffer heads are directly looked up and referenced (such as
832 * block device mappings).
834 int buffer_migrate_page_norefs(struct address_space *mapping,
835 struct page *newpage, struct page *page, enum migrate_mode mode)
837 return __buffer_migrate_page(mapping, newpage, page, mode, true);
842 * Writeback a page to clean the dirty state
844 static int writeout(struct address_space *mapping, struct page *page)
846 struct writeback_control wbc = {
847 .sync_mode = WB_SYNC_NONE,
850 .range_end = LLONG_MAX,
855 if (!mapping->a_ops->writepage)
856 /* No write method for the address space */
859 if (!clear_page_dirty_for_io(page))
860 /* Someone else already triggered a write */
864 * A dirty page may imply that the underlying filesystem has
865 * the page on some queue. So the page must be clean for
866 * migration. Writeout may mean we loose the lock and the
867 * page state is no longer what we checked for earlier.
868 * At this point we know that the migration attempt cannot
871 remove_migration_ptes(page, page, false);
873 rc = mapping->a_ops->writepage(page, &wbc);
875 if (rc != AOP_WRITEPAGE_ACTIVATE)
876 /* unlocked. Relock */
879 return (rc < 0) ? -EIO : -EAGAIN;
883 * Default handling if a filesystem does not provide a migration function.
885 static int fallback_migrate_page(struct address_space *mapping,
886 struct page *newpage, struct page *page, enum migrate_mode mode)
888 if (PageDirty(page)) {
889 /* Only writeback pages in full synchronous migration */
892 case MIGRATE_SYNC_NO_COPY:
897 return writeout(mapping, page);
901 * Buffers may be managed in a filesystem specific way.
902 * We must have no buffers or drop them.
904 if (page_has_private(page) &&
905 !try_to_release_page(page, GFP_KERNEL))
906 return mode == MIGRATE_SYNC ? -EAGAIN : -EBUSY;
908 return migrate_page(mapping, newpage, page, mode);
912 * Move a page to a newly allocated page
913 * The page is locked and all ptes have been successfully removed.
915 * The new page will have replaced the old page if this function
920 * MIGRATEPAGE_SUCCESS - success
922 static int move_to_new_page(struct page *newpage, struct page *page,
923 enum migrate_mode mode)
925 struct address_space *mapping;
927 bool is_lru = !__PageMovable(page);
929 VM_BUG_ON_PAGE(!PageLocked(page), page);
930 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
932 mapping = page_mapping(page);
934 if (likely(is_lru)) {
936 rc = migrate_page(mapping, newpage, page, mode);
937 else if (mapping->a_ops->migratepage)
939 * Most pages have a mapping and most filesystems
940 * provide a migratepage callback. Anonymous pages
941 * are part of swap space which also has its own
942 * migratepage callback. This is the most common path
943 * for page migration.
945 rc = mapping->a_ops->migratepage(mapping, newpage,
948 rc = fallback_migrate_page(mapping, newpage,
952 * In case of non-lru page, it could be released after
953 * isolation step. In that case, we shouldn't try migration.
955 VM_BUG_ON_PAGE(!PageIsolated(page), page);
956 if (!PageMovable(page)) {
957 rc = MIGRATEPAGE_SUCCESS;
958 __ClearPageIsolated(page);
962 rc = mapping->a_ops->migratepage(mapping, newpage,
964 WARN_ON_ONCE(rc == MIGRATEPAGE_SUCCESS &&
965 !PageIsolated(page));
969 * When successful, old pagecache page->mapping must be cleared before
970 * page is freed; but stats require that PageAnon be left as PageAnon.
972 if (rc == MIGRATEPAGE_SUCCESS) {
973 if (__PageMovable(page)) {
974 VM_BUG_ON_PAGE(!PageIsolated(page), page);
977 * We clear PG_movable under page_lock so any compactor
978 * cannot try to migrate this page.
980 __ClearPageIsolated(page);
984 * Anonymous and movable page->mapping will be cleared by
985 * free_pages_prepare so don't reset it here for keeping
986 * the type to work PageAnon, for example.
988 if (!PageMappingFlags(page))
989 page->mapping = NULL;
991 if (likely(!is_zone_device_page(newpage)))
992 flush_dcache_page(newpage);
999 static int __unmap_and_move(struct page *page, struct page *newpage,
1000 int force, enum migrate_mode mode)
1003 int page_was_mapped = 0;
1004 struct anon_vma *anon_vma = NULL;
1005 bool is_lru = !__PageMovable(page);
1007 if (!trylock_page(page)) {
1008 if (!force || mode == MIGRATE_ASYNC)
1012 * It's not safe for direct compaction to call lock_page.
1013 * For example, during page readahead pages are added locked
1014 * to the LRU. Later, when the IO completes the pages are
1015 * marked uptodate and unlocked. However, the queueing
1016 * could be merging multiple pages for one bio (e.g.
1017 * mpage_readahead). If an allocation happens for the
1018 * second or third page, the process can end up locking
1019 * the same page twice and deadlocking. Rather than
1020 * trying to be clever about what pages can be locked,
1021 * avoid the use of lock_page for direct compaction
1024 if (current->flags & PF_MEMALLOC)
1030 if (PageWriteback(page)) {
1032 * Only in the case of a full synchronous migration is it
1033 * necessary to wait for PageWriteback. In the async case,
1034 * the retry loop is too short and in the sync-light case,
1035 * the overhead of stalling is too much
1039 case MIGRATE_SYNC_NO_COPY:
1047 wait_on_page_writeback(page);
1051 * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
1052 * we cannot notice that anon_vma is freed while we migrates a page.
1053 * This get_anon_vma() delays freeing anon_vma pointer until the end
1054 * of migration. File cache pages are no problem because of page_lock()
1055 * File Caches may use write_page() or lock_page() in migration, then,
1056 * just care Anon page here.
1058 * Only page_get_anon_vma() understands the subtleties of
1059 * getting a hold on an anon_vma from outside one of its mms.
1060 * But if we cannot get anon_vma, then we won't need it anyway,
1061 * because that implies that the anon page is no longer mapped
1062 * (and cannot be remapped so long as we hold the page lock).
1064 if (PageAnon(page) && !PageKsm(page))
1065 anon_vma = page_get_anon_vma(page);
1068 * Block others from accessing the new page when we get around to
1069 * establishing additional references. We are usually the only one
1070 * holding a reference to newpage at this point. We used to have a BUG
1071 * here if trylock_page(newpage) fails, but would like to allow for
1072 * cases where there might be a race with the previous use of newpage.
1073 * This is much like races on refcount of oldpage: just don't BUG().
1075 if (unlikely(!trylock_page(newpage)))
1078 if (unlikely(!is_lru)) {
1079 rc = move_to_new_page(newpage, page, mode);
1080 goto out_unlock_both;
1084 * Corner case handling:
1085 * 1. When a new swap-cache page is read into, it is added to the LRU
1086 * and treated as swapcache but it has no rmap yet.
1087 * Calling try_to_unmap() against a page->mapping==NULL page will
1088 * trigger a BUG. So handle it here.
1089 * 2. An orphaned page (see truncate_cleanup_page) might have
1090 * fs-private metadata. The page can be picked up due to memory
1091 * offlining. Everywhere else except page reclaim, the page is
1092 * invisible to the vm, so the page can not be migrated. So try to
1093 * free the metadata, so the page can be freed.
1095 if (!page->mapping) {
1096 VM_BUG_ON_PAGE(PageAnon(page), page);
1097 if (page_has_private(page)) {
1098 try_to_free_buffers(page);
1099 goto out_unlock_both;
1101 } else if (page_mapped(page)) {
1102 /* Establish migration ptes */
1103 VM_BUG_ON_PAGE(PageAnon(page) && !PageKsm(page) && !anon_vma,
1105 try_to_unmap(page, TTU_MIGRATION|TTU_IGNORE_MLOCK);
1106 page_was_mapped = 1;
1109 if (!page_mapped(page))
1110 rc = move_to_new_page(newpage, page, mode);
1112 if (page_was_mapped)
1113 remove_migration_ptes(page,
1114 rc == MIGRATEPAGE_SUCCESS ? newpage : page, false);
1117 unlock_page(newpage);
1119 /* Drop an anon_vma reference if we took one */
1121 put_anon_vma(anon_vma);
1125 * If migration is successful, decrease refcount of the newpage
1126 * which will not free the page because new page owner increased
1127 * refcounter. As well, if it is LRU page, add the page to LRU
1128 * list in here. Use the old state of the isolated source page to
1129 * determine if we migrated a LRU page. newpage was already unlocked
1130 * and possibly modified by its owner - don't rely on the page
1133 if (rc == MIGRATEPAGE_SUCCESS) {
1134 if (unlikely(!is_lru))
1137 putback_lru_page(newpage);
1144 * Obtain the lock on page, remove all ptes and migrate the page
1145 * to the newly allocated page in newpage.
1147 static int unmap_and_move(new_page_t get_new_page,
1148 free_page_t put_new_page,
1149 unsigned long private, struct page *page,
1150 int force, enum migrate_mode mode,
1151 enum migrate_reason reason,
1152 struct list_head *ret)
1154 int rc = MIGRATEPAGE_SUCCESS;
1155 struct page *newpage = NULL;
1157 if (!thp_migration_supported() && PageTransHuge(page))
1160 if (page_count(page) == 1) {
1161 /* page was freed from under us. So we are done. */
1162 ClearPageActive(page);
1163 ClearPageUnevictable(page);
1164 if (unlikely(__PageMovable(page))) {
1166 if (!PageMovable(page))
1167 __ClearPageIsolated(page);
1173 newpage = get_new_page(page, private);
1177 rc = __unmap_and_move(page, newpage, force, mode);
1178 if (rc == MIGRATEPAGE_SUCCESS)
1179 set_page_owner_migrate_reason(newpage, reason);
1182 if (rc != -EAGAIN) {
1184 * A page that has been migrated has all references
1185 * removed and will be freed. A page that has not been
1186 * migrated will have kept its references and be restored.
1188 list_del(&page->lru);
1192 * If migration is successful, releases reference grabbed during
1193 * isolation. Otherwise, restore the page to right list unless
1196 if (rc == MIGRATEPAGE_SUCCESS) {
1198 * Compaction can migrate also non-LRU pages which are
1199 * not accounted to NR_ISOLATED_*. They can be recognized
1202 if (likely(!__PageMovable(page)))
1203 mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON +
1204 page_is_file_lru(page), -thp_nr_pages(page));
1206 if (reason != MR_MEMORY_FAILURE)
1208 * We release the page in page_handle_poison.
1213 list_add_tail(&page->lru, ret);
1216 put_new_page(newpage, private);
1225 * Counterpart of unmap_and_move_page() for hugepage migration.
1227 * This function doesn't wait the completion of hugepage I/O
1228 * because there is no race between I/O and migration for hugepage.
1229 * Note that currently hugepage I/O occurs only in direct I/O
1230 * where no lock is held and PG_writeback is irrelevant,
1231 * and writeback status of all subpages are counted in the reference
1232 * count of the head page (i.e. if all subpages of a 2MB hugepage are
1233 * under direct I/O, the reference of the head page is 512 and a bit more.)
1234 * This means that when we try to migrate hugepage whose subpages are
1235 * doing direct I/O, some references remain after try_to_unmap() and
1236 * hugepage migration fails without data corruption.
1238 * There is also no race when direct I/O is issued on the page under migration,
1239 * because then pte is replaced with migration swap entry and direct I/O code
1240 * will wait in the page fault for migration to complete.
1242 static int unmap_and_move_huge_page(new_page_t get_new_page,
1243 free_page_t put_new_page, unsigned long private,
1244 struct page *hpage, int force,
1245 enum migrate_mode mode, int reason,
1246 struct list_head *ret)
1249 int page_was_mapped = 0;
1250 struct page *new_hpage;
1251 struct anon_vma *anon_vma = NULL;
1252 struct address_space *mapping = NULL;
1255 * Migratability of hugepages depends on architectures and their size.
1256 * This check is necessary because some callers of hugepage migration
1257 * like soft offline and memory hotremove don't walk through page
1258 * tables or check whether the hugepage is pmd-based or not before
1259 * kicking migration.
1261 if (!hugepage_migration_supported(page_hstate(hpage))) {
1262 list_move_tail(&hpage->lru, ret);
1266 if (page_count(hpage) == 1) {
1267 /* page was freed from under us. So we are done. */
1268 putback_active_hugepage(hpage);
1269 return MIGRATEPAGE_SUCCESS;
1272 new_hpage = get_new_page(hpage, private);
1276 if (!trylock_page(hpage)) {
1281 case MIGRATE_SYNC_NO_COPY:
1290 * Check for pages which are in the process of being freed. Without
1291 * page_mapping() set, hugetlbfs specific move page routine will not
1292 * be called and we could leak usage counts for subpools.
1294 if (page_private(hpage) && !page_mapping(hpage)) {
1299 if (PageAnon(hpage))
1300 anon_vma = page_get_anon_vma(hpage);
1302 if (unlikely(!trylock_page(new_hpage)))
1305 if (page_mapped(hpage)) {
1306 bool mapping_locked = false;
1307 enum ttu_flags ttu = TTU_MIGRATION|TTU_IGNORE_MLOCK;
1309 if (!PageAnon(hpage)) {
1311 * In shared mappings, try_to_unmap could potentially
1312 * call huge_pmd_unshare. Because of this, take
1313 * semaphore in write mode here and set TTU_RMAP_LOCKED
1314 * to let lower levels know we have taken the lock.
1316 mapping = hugetlb_page_mapping_lock_write(hpage);
1317 if (unlikely(!mapping))
1318 goto unlock_put_anon;
1320 mapping_locked = true;
1321 ttu |= TTU_RMAP_LOCKED;
1324 try_to_unmap(hpage, ttu);
1325 page_was_mapped = 1;
1328 i_mmap_unlock_write(mapping);
1331 if (!page_mapped(hpage))
1332 rc = move_to_new_page(new_hpage, hpage, mode);
1334 if (page_was_mapped)
1335 remove_migration_ptes(hpage,
1336 rc == MIGRATEPAGE_SUCCESS ? new_hpage : hpage, false);
1339 unlock_page(new_hpage);
1343 put_anon_vma(anon_vma);
1345 if (rc == MIGRATEPAGE_SUCCESS) {
1346 move_hugetlb_state(hpage, new_hpage, reason);
1347 put_new_page = NULL;
1353 if (rc == MIGRATEPAGE_SUCCESS)
1354 putback_active_hugepage(hpage);
1355 else if (rc != -EAGAIN)
1356 list_move_tail(&hpage->lru, ret);
1359 * If migration was not successful and there's a freeing callback, use
1360 * it. Otherwise, put_page() will drop the reference grabbed during
1364 put_new_page(new_hpage, private);
1366 putback_active_hugepage(new_hpage);
1371 static inline int try_split_thp(struct page *page, struct page **page2,
1372 struct list_head *from)
1377 rc = split_huge_page_to_list(page, from);
1380 list_safe_reset_next(page, *page2, lru);
1386 * migrate_pages - migrate the pages specified in a list, to the free pages
1387 * supplied as the target for the page migration
1389 * @from: The list of pages to be migrated.
1390 * @get_new_page: The function used to allocate free pages to be used
1391 * as the target of the page migration.
1392 * @put_new_page: The function used to free target pages if migration
1393 * fails, or NULL if no special handling is necessary.
1394 * @private: Private data to be passed on to get_new_page()
1395 * @mode: The migration mode that specifies the constraints for
1396 * page migration, if any.
1397 * @reason: The reason for page migration.
1399 * The function returns after 10 attempts or if no pages are movable any more
1400 * because the list has become empty or no retryable pages exist any more.
1401 * It is caller's responsibility to call putback_movable_pages() to return pages
1402 * to the LRU or free list only if ret != 0.
1404 * Returns the number of pages that were not migrated, or an error code.
1406 int migrate_pages(struct list_head *from, new_page_t get_new_page,
1407 free_page_t put_new_page, unsigned long private,
1408 enum migrate_mode mode, int reason)
1413 int nr_succeeded = 0;
1414 int nr_thp_succeeded = 0;
1415 int nr_thp_failed = 0;
1416 int nr_thp_split = 0;
1418 bool is_thp = false;
1421 int swapwrite = current->flags & PF_SWAPWRITE;
1422 int rc, nr_subpages;
1423 LIST_HEAD(ret_pages);
1425 trace_mm_migrate_pages_start(mode, reason);
1428 current->flags |= PF_SWAPWRITE;
1430 for (pass = 0; pass < 10 && (retry || thp_retry); pass++) {
1434 list_for_each_entry_safe(page, page2, from, lru) {
1437 * THP statistics is based on the source huge page.
1438 * Capture required information that might get lost
1441 is_thp = PageTransHuge(page) && !PageHuge(page);
1442 nr_subpages = thp_nr_pages(page);
1446 rc = unmap_and_move_huge_page(get_new_page,
1447 put_new_page, private, page,
1448 pass > 2, mode, reason,
1451 rc = unmap_and_move(get_new_page, put_new_page,
1452 private, page, pass > 2, mode,
1453 reason, &ret_pages);
1456 * Success: non hugetlb page will be freed, hugetlb
1457 * page will be put back
1458 * -EAGAIN: stay on the from list
1459 * -ENOMEM: stay on the from list
1460 * Other errno: put on ret_pages list then splice to
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
1476 /* THP migration is unsupported */
1478 if (!try_split_thp(page, &page2, from)) {
1484 nr_failed += nr_subpages;
1488 /* Hugetlb migration is unsupported */
1493 * When memory is low, don't bother to try to migrate
1494 * other pages, just exit.
1497 if (!try_split_thp(page, &page2, from)) {
1503 nr_failed += nr_subpages;
1515 case MIGRATEPAGE_SUCCESS:
1518 nr_succeeded += nr_subpages;
1525 * Permanent failure (-EBUSY, etc.):
1526 * unlike -EAGAIN case, the failed page is
1527 * removed from migration page list and not
1528 * retried in the next outer loop.
1532 nr_failed += nr_subpages;
1540 nr_failed += retry + thp_retry;
1541 nr_thp_failed += thp_retry;
1545 * Put the permanent failure page back to migration list, they
1546 * will be put back to the right list by the caller.
1548 list_splice(&ret_pages, from);
1550 count_vm_events(PGMIGRATE_SUCCESS, nr_succeeded);
1551 count_vm_events(PGMIGRATE_FAIL, nr_failed);
1552 count_vm_events(THP_MIGRATION_SUCCESS, nr_thp_succeeded);
1553 count_vm_events(THP_MIGRATION_FAIL, nr_thp_failed);
1554 count_vm_events(THP_MIGRATION_SPLIT, nr_thp_split);
1555 trace_mm_migrate_pages(nr_succeeded, nr_failed, nr_thp_succeeded,
1556 nr_thp_failed, nr_thp_split, mode, reason);
1559 current->flags &= ~PF_SWAPWRITE;
1564 struct page *alloc_migration_target(struct page *page, unsigned long private)
1566 struct migration_target_control *mtc;
1568 unsigned int order = 0;
1569 struct page *new_page = NULL;
1573 mtc = (struct migration_target_control *)private;
1574 gfp_mask = mtc->gfp_mask;
1576 if (nid == NUMA_NO_NODE)
1577 nid = page_to_nid(page);
1579 if (PageHuge(page)) {
1580 struct hstate *h = page_hstate(compound_head(page));
1582 gfp_mask = htlb_modify_alloc_mask(h, gfp_mask);
1583 return alloc_huge_page_nodemask(h, nid, mtc->nmask, gfp_mask);
1586 if (PageTransHuge(page)) {
1588 * clear __GFP_RECLAIM to make the migration callback
1589 * consistent with regular THP allocations.
1591 gfp_mask &= ~__GFP_RECLAIM;
1592 gfp_mask |= GFP_TRANSHUGE;
1593 order = HPAGE_PMD_ORDER;
1595 zidx = zone_idx(page_zone(page));
1596 if (is_highmem_idx(zidx) || zidx == ZONE_MOVABLE)
1597 gfp_mask |= __GFP_HIGHMEM;
1599 new_page = __alloc_pages(gfp_mask, order, nid, mtc->nmask);
1601 if (new_page && PageTransHuge(new_page))
1602 prep_transhuge_page(new_page);
1609 static int store_status(int __user *status, int start, int value, int nr)
1612 if (put_user(value, status + start))
1620 static int do_move_pages_to_node(struct mm_struct *mm,
1621 struct list_head *pagelist, int node)
1624 struct migration_target_control mtc = {
1626 .gfp_mask = GFP_HIGHUSER_MOVABLE | __GFP_THISNODE,
1629 err = migrate_pages(pagelist, alloc_migration_target, NULL,
1630 (unsigned long)&mtc, MIGRATE_SYNC, MR_SYSCALL);
1632 putback_movable_pages(pagelist);
1637 * Resolves the given address to a struct page, isolates it from the LRU and
1638 * puts it to the given pagelist.
1640 * errno - if the page cannot be found/isolated
1641 * 0 - when it doesn't have to be migrated because it is already on the
1643 * 1 - when it has been queued
1645 static int add_page_for_migration(struct mm_struct *mm, unsigned long addr,
1646 int node, struct list_head *pagelist, bool migrate_all)
1648 struct vm_area_struct *vma;
1650 unsigned int follflags;
1655 vma = find_vma(mm, addr);
1656 if (!vma || addr < vma->vm_start || !vma_migratable(vma))
1659 /* FOLL_DUMP to ignore special (like zero) pages */
1660 follflags = FOLL_GET | FOLL_DUMP;
1661 page = follow_page(vma, addr, follflags);
1663 err = PTR_ERR(page);
1672 if (page_to_nid(page) == node)
1676 if (page_mapcount(page) > 1 && !migrate_all)
1679 if (PageHuge(page)) {
1680 if (PageHead(page)) {
1681 isolate_huge_page(page, pagelist);
1687 head = compound_head(page);
1688 err = isolate_lru_page(head);
1693 list_add_tail(&head->lru, pagelist);
1694 mod_node_page_state(page_pgdat(head),
1695 NR_ISOLATED_ANON + page_is_file_lru(head),
1696 thp_nr_pages(head));
1700 * Either remove the duplicate refcount from
1701 * isolate_lru_page() or drop the page ref if it was
1706 mmap_read_unlock(mm);
1710 static int move_pages_and_store_status(struct mm_struct *mm, int node,
1711 struct list_head *pagelist, int __user *status,
1712 int start, int i, unsigned long nr_pages)
1716 if (list_empty(pagelist))
1719 err = do_move_pages_to_node(mm, pagelist, node);
1722 * Positive err means the number of failed
1723 * pages to migrate. Since we are going to
1724 * abort and return the number of non-migrated
1725 * pages, so need to include the rest of the
1726 * nr_pages that have not been attempted as
1730 err += nr_pages - i - 1;
1733 return store_status(status, start, node, i - start);
1737 * Migrate an array of page address onto an array of nodes and fill
1738 * the corresponding array of status.
1740 static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes,
1741 unsigned long nr_pages,
1742 const void __user * __user *pages,
1743 const int __user *nodes,
1744 int __user *status, int flags)
1746 int current_node = NUMA_NO_NODE;
1747 LIST_HEAD(pagelist);
1751 lru_cache_disable();
1753 for (i = start = 0; i < nr_pages; i++) {
1754 const void __user *p;
1759 if (get_user(p, pages + i))
1761 if (get_user(node, nodes + i))
1763 addr = (unsigned long)untagged_addr(p);
1766 if (node < 0 || node >= MAX_NUMNODES)
1768 if (!node_state(node, N_MEMORY))
1772 if (!node_isset(node, task_nodes))
1775 if (current_node == NUMA_NO_NODE) {
1776 current_node = node;
1778 } else if (node != current_node) {
1779 err = move_pages_and_store_status(mm, current_node,
1780 &pagelist, status, start, i, nr_pages);
1784 current_node = node;
1788 * Errors in the page lookup or isolation are not fatal and we simply
1789 * report them via status
1791 err = add_page_for_migration(mm, addr, current_node,
1792 &pagelist, flags & MPOL_MF_MOVE_ALL);
1795 /* The page is successfully queued for migration */
1800 * If the page is already on the target node (!err), store the
1801 * node, otherwise, store the err.
1803 err = store_status(status, i, err ? : current_node, 1);
1807 err = move_pages_and_store_status(mm, current_node, &pagelist,
1808 status, start, i, nr_pages);
1811 current_node = NUMA_NO_NODE;
1814 /* Make sure we do not overwrite the existing error */
1815 err1 = move_pages_and_store_status(mm, current_node, &pagelist,
1816 status, start, i, nr_pages);
1825 * Determine the nodes of an array of pages and store it in an array of status.
1827 static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages,
1828 const void __user **pages, int *status)
1834 for (i = 0; i < nr_pages; i++) {
1835 unsigned long addr = (unsigned long)(*pages);
1836 struct vm_area_struct *vma;
1840 vma = vma_lookup(mm, addr);
1844 /* FOLL_DUMP to ignore special (like zero) pages */
1845 page = follow_page(vma, addr, FOLL_DUMP);
1847 err = PTR_ERR(page);
1851 err = page ? page_to_nid(page) : -ENOENT;
1859 mmap_read_unlock(mm);
1863 * Determine the nodes of a user array of pages and store it in
1864 * a user array of status.
1866 static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages,
1867 const void __user * __user *pages,
1870 #define DO_PAGES_STAT_CHUNK_NR 16
1871 const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR];
1872 int chunk_status[DO_PAGES_STAT_CHUNK_NR];
1875 unsigned long chunk_nr;
1877 chunk_nr = nr_pages;
1878 if (chunk_nr > DO_PAGES_STAT_CHUNK_NR)
1879 chunk_nr = DO_PAGES_STAT_CHUNK_NR;
1881 if (copy_from_user(chunk_pages, pages, chunk_nr * sizeof(*chunk_pages)))
1884 do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status);
1886 if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status)))
1891 nr_pages -= chunk_nr;
1893 return nr_pages ? -EFAULT : 0;
1896 static struct mm_struct *find_mm_struct(pid_t pid, nodemask_t *mem_nodes)
1898 struct task_struct *task;
1899 struct mm_struct *mm;
1902 * There is no need to check if current process has the right to modify
1903 * the specified process when they are same.
1907 *mem_nodes = cpuset_mems_allowed(current);
1911 /* Find the mm_struct */
1913 task = find_task_by_vpid(pid);
1916 return ERR_PTR(-ESRCH);
1918 get_task_struct(task);
1921 * Check if this process has the right to modify the specified
1922 * process. Use the regular "ptrace_may_access()" checks.
1924 if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) {
1926 mm = ERR_PTR(-EPERM);
1931 mm = ERR_PTR(security_task_movememory(task));
1934 *mem_nodes = cpuset_mems_allowed(task);
1935 mm = get_task_mm(task);
1937 put_task_struct(task);
1939 mm = ERR_PTR(-EINVAL);
1944 * Move a list of pages in the address space of the currently executing
1947 static int kernel_move_pages(pid_t pid, unsigned long nr_pages,
1948 const void __user * __user *pages,
1949 const int __user *nodes,
1950 int __user *status, int flags)
1952 struct mm_struct *mm;
1954 nodemask_t task_nodes;
1957 if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
1960 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1963 mm = find_mm_struct(pid, &task_nodes);
1968 err = do_pages_move(mm, task_nodes, nr_pages, pages,
1969 nodes, status, flags);
1971 err = do_pages_stat(mm, nr_pages, pages, status);
1977 SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages,
1978 const void __user * __user *, pages,
1979 const int __user *, nodes,
1980 int __user *, status, int, flags)
1982 return kernel_move_pages(pid, nr_pages, pages, nodes, status, flags);
1985 #ifdef CONFIG_COMPAT
1986 COMPAT_SYSCALL_DEFINE6(move_pages, pid_t, pid, compat_ulong_t, nr_pages,
1987 compat_uptr_t __user *, pages32,
1988 const int __user *, nodes,
1989 int __user *, status,
1992 const void __user * __user *pages;
1995 pages = compat_alloc_user_space(nr_pages * sizeof(void *));
1996 for (i = 0; i < nr_pages; i++) {
1999 if (get_user(p, pages32 + i) ||
2000 put_user(compat_ptr(p), pages + i))
2003 return kernel_move_pages(pid, nr_pages, pages, nodes, status, flags);
2005 #endif /* CONFIG_COMPAT */
2007 #ifdef CONFIG_NUMA_BALANCING
2009 * Returns true if this is a safe migration target node for misplaced NUMA
2010 * pages. Currently it only checks the watermarks which crude
2012 static bool migrate_balanced_pgdat(struct pglist_data *pgdat,
2013 unsigned long nr_migrate_pages)
2017 for (z = pgdat->nr_zones - 1; z >= 0; z--) {
2018 struct zone *zone = pgdat->node_zones + z;
2020 if (!populated_zone(zone))
2023 /* Avoid waking kswapd by allocating pages_to_migrate pages. */
2024 if (!zone_watermark_ok(zone, 0,
2025 high_wmark_pages(zone) +
2034 static struct page *alloc_misplaced_dst_page(struct page *page,
2037 int nid = (int) data;
2038 struct page *newpage;
2040 newpage = __alloc_pages_node(nid,
2041 (GFP_HIGHUSER_MOVABLE |
2042 __GFP_THISNODE | __GFP_NOMEMALLOC |
2043 __GFP_NORETRY | __GFP_NOWARN) &
2049 static int numamigrate_isolate_page(pg_data_t *pgdat, struct page *page)
2053 VM_BUG_ON_PAGE(compound_order(page) && !PageTransHuge(page), page);
2055 /* Avoid migrating to a node that is nearly full */
2056 if (!migrate_balanced_pgdat(pgdat, compound_nr(page)))
2059 if (isolate_lru_page(page))
2063 * migrate_misplaced_transhuge_page() skips page migration's usual
2064 * check on page_count(), so we must do it here, now that the page
2065 * has been isolated: a GUP pin, or any other pin, prevents migration.
2066 * The expected page count is 3: 1 for page's mapcount and 1 for the
2067 * caller's pin and 1 for the reference taken by isolate_lru_page().
2069 if (PageTransHuge(page) && page_count(page) != 3) {
2070 putback_lru_page(page);
2074 page_lru = page_is_file_lru(page);
2075 mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON + page_lru,
2076 thp_nr_pages(page));
2079 * Isolating the page has taken another reference, so the
2080 * caller's reference can be safely dropped without the page
2081 * disappearing underneath us during migration.
2087 bool pmd_trans_migrating(pmd_t pmd)
2089 struct page *page = pmd_page(pmd);
2090 return PageLocked(page);
2094 * Attempt to migrate a misplaced page to the specified destination
2095 * node. Caller is expected to have an elevated reference count on
2096 * the page that will be dropped by this function before returning.
2098 int migrate_misplaced_page(struct page *page, struct vm_area_struct *vma,
2101 pg_data_t *pgdat = NODE_DATA(node);
2104 LIST_HEAD(migratepages);
2107 * Don't migrate file pages that are mapped in multiple processes
2108 * with execute permissions as they are probably shared libraries.
2110 if (page_mapcount(page) != 1 && page_is_file_lru(page) &&
2111 (vma->vm_flags & VM_EXEC))
2115 * Also do not migrate dirty pages as not all filesystems can move
2116 * dirty pages in MIGRATE_ASYNC mode which is a waste of cycles.
2118 if (page_is_file_lru(page) && PageDirty(page))
2121 isolated = numamigrate_isolate_page(pgdat, page);
2125 list_add(&page->lru, &migratepages);
2126 nr_remaining = migrate_pages(&migratepages, alloc_misplaced_dst_page,
2127 NULL, node, MIGRATE_ASYNC,
2130 if (!list_empty(&migratepages)) {
2131 list_del(&page->lru);
2132 dec_node_page_state(page, NR_ISOLATED_ANON +
2133 page_is_file_lru(page));
2134 putback_lru_page(page);
2138 count_vm_numa_event(NUMA_PAGE_MIGRATE);
2139 BUG_ON(!list_empty(&migratepages));
2146 #endif /* CONFIG_NUMA_BALANCING */
2148 #if defined(CONFIG_NUMA_BALANCING) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
2150 * Migrates a THP to a given target node. page must be locked and is unlocked
2153 int migrate_misplaced_transhuge_page(struct mm_struct *mm,
2154 struct vm_area_struct *vma,
2155 pmd_t *pmd, pmd_t entry,
2156 unsigned long address,
2157 struct page *page, int node)
2160 pg_data_t *pgdat = NODE_DATA(node);
2162 struct page *new_page = NULL;
2163 int page_lru = page_is_file_lru(page);
2164 unsigned long start = address & HPAGE_PMD_MASK;
2166 new_page = alloc_pages_node(node,
2167 (GFP_TRANSHUGE_LIGHT | __GFP_THISNODE),
2171 prep_transhuge_page(new_page);
2173 isolated = numamigrate_isolate_page(pgdat, page);
2179 /* Prepare a page as a migration target */
2180 __SetPageLocked(new_page);
2181 if (PageSwapBacked(page))
2182 __SetPageSwapBacked(new_page);
2184 /* anon mapping, we can simply copy page->mapping to the new page: */
2185 new_page->mapping = page->mapping;
2186 new_page->index = page->index;
2187 /* flush the cache before copying using the kernel virtual address */
2188 flush_cache_range(vma, start, start + HPAGE_PMD_SIZE);
2189 migrate_page_copy(new_page, page);
2190 WARN_ON(PageLRU(new_page));
2192 /* Recheck the target PMD */
2193 ptl = pmd_lock(mm, pmd);
2194 if (unlikely(!pmd_same(*pmd, entry) || !page_ref_freeze(page, 2))) {
2197 /* Reverse changes made by migrate_page_copy() */
2198 if (TestClearPageActive(new_page))
2199 SetPageActive(page);
2200 if (TestClearPageUnevictable(new_page))
2201 SetPageUnevictable(page);
2203 unlock_page(new_page);
2204 put_page(new_page); /* Free it */
2206 /* Retake the callers reference and putback on LRU */
2208 putback_lru_page(page);
2209 mod_node_page_state(page_pgdat(page),
2210 NR_ISOLATED_ANON + page_lru, -HPAGE_PMD_NR);
2215 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
2216 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
2219 * Overwrite the old entry under pagetable lock and establish
2220 * the new PTE. Any parallel GUP will either observe the old
2221 * page blocking on the page lock, block on the page table
2222 * lock or observe the new page. The SetPageUptodate on the
2223 * new page and page_add_new_anon_rmap guarantee the copy is
2224 * visible before the pagetable update.
2226 page_add_anon_rmap(new_page, vma, start, true);
2228 * At this point the pmd is numa/protnone (i.e. non present) and the TLB
2229 * has already been flushed globally. So no TLB can be currently
2230 * caching this non present pmd mapping. There's no need to clear the
2231 * pmd before doing set_pmd_at(), nor to flush the TLB after
2232 * set_pmd_at(). Clearing the pmd here would introduce a race
2233 * condition against MADV_DONTNEED, because MADV_DONTNEED only holds the
2234 * mmap_lock for reading. If the pmd is set to NULL at any given time,
2235 * MADV_DONTNEED won't wait on the pmd lock and it'll skip clearing this
2238 set_pmd_at(mm, start, pmd, entry);
2239 update_mmu_cache_pmd(vma, address, &entry);
2241 page_ref_unfreeze(page, 2);
2242 mlock_migrate_page(new_page, page);
2243 page_remove_rmap(page, true);
2244 set_page_owner_migrate_reason(new_page, MR_NUMA_MISPLACED);
2248 /* Take an "isolate" reference and put new page on the LRU. */
2250 putback_lru_page(new_page);
2252 unlock_page(new_page);
2254 put_page(page); /* Drop the rmap reference */
2255 put_page(page); /* Drop the LRU isolation reference */
2257 count_vm_events(PGMIGRATE_SUCCESS, HPAGE_PMD_NR);
2258 count_vm_numa_events(NUMA_PAGE_MIGRATE, HPAGE_PMD_NR);
2260 mod_node_page_state(page_pgdat(page),
2261 NR_ISOLATED_ANON + page_lru,
2266 count_vm_events(PGMIGRATE_FAIL, HPAGE_PMD_NR);
2267 ptl = pmd_lock(mm, pmd);
2268 if (pmd_same(*pmd, entry)) {
2269 entry = pmd_modify(entry, vma->vm_page_prot);
2270 set_pmd_at(mm, start, pmd, entry);
2271 update_mmu_cache_pmd(vma, address, &entry);
2280 #endif /* CONFIG_NUMA_BALANCING */
2282 #endif /* CONFIG_NUMA */
2284 #ifdef CONFIG_DEVICE_PRIVATE
2285 static int migrate_vma_collect_skip(unsigned long start,
2287 struct mm_walk *walk)
2289 struct migrate_vma *migrate = walk->private;
2292 for (addr = start; addr < end; addr += PAGE_SIZE) {
2293 migrate->dst[migrate->npages] = 0;
2294 migrate->src[migrate->npages++] = 0;
2300 static int migrate_vma_collect_hole(unsigned long start,
2302 __always_unused int depth,
2303 struct mm_walk *walk)
2305 struct migrate_vma *migrate = walk->private;
2308 /* Only allow populating anonymous memory. */
2309 if (!vma_is_anonymous(walk->vma))
2310 return migrate_vma_collect_skip(start, end, walk);
2312 for (addr = start; addr < end; addr += PAGE_SIZE) {
2313 migrate->src[migrate->npages] = MIGRATE_PFN_MIGRATE;
2314 migrate->dst[migrate->npages] = 0;
2322 static int migrate_vma_collect_pmd(pmd_t *pmdp,
2323 unsigned long start,
2325 struct mm_walk *walk)
2327 struct migrate_vma *migrate = walk->private;
2328 struct vm_area_struct *vma = walk->vma;
2329 struct mm_struct *mm = vma->vm_mm;
2330 unsigned long addr = start, unmapped = 0;
2335 if (pmd_none(*pmdp))
2336 return migrate_vma_collect_hole(start, end, -1, walk);
2338 if (pmd_trans_huge(*pmdp)) {
2341 ptl = pmd_lock(mm, pmdp);
2342 if (unlikely(!pmd_trans_huge(*pmdp))) {
2347 page = pmd_page(*pmdp);
2348 if (is_huge_zero_page(page)) {
2350 split_huge_pmd(vma, pmdp, addr);
2351 if (pmd_trans_unstable(pmdp))
2352 return migrate_vma_collect_skip(start, end,
2359 if (unlikely(!trylock_page(page)))
2360 return migrate_vma_collect_skip(start, end,
2362 ret = split_huge_page(page);
2366 return migrate_vma_collect_skip(start, end,
2368 if (pmd_none(*pmdp))
2369 return migrate_vma_collect_hole(start, end, -1,
2374 if (unlikely(pmd_bad(*pmdp)))
2375 return migrate_vma_collect_skip(start, end, walk);
2377 ptep = pte_offset_map_lock(mm, pmdp, addr, &ptl);
2378 arch_enter_lazy_mmu_mode();
2380 for (; addr < end; addr += PAGE_SIZE, ptep++) {
2381 unsigned long mpfn = 0, pfn;
2388 if (pte_none(pte)) {
2389 if (vma_is_anonymous(vma)) {
2390 mpfn = MIGRATE_PFN_MIGRATE;
2396 if (!pte_present(pte)) {
2398 * Only care about unaddressable device page special
2399 * page table entry. Other special swap entries are not
2400 * migratable, and we ignore regular swapped page.
2402 entry = pte_to_swp_entry(pte);
2403 if (!is_device_private_entry(entry))
2406 page = device_private_entry_to_page(entry);
2407 if (!(migrate->flags &
2408 MIGRATE_VMA_SELECT_DEVICE_PRIVATE) ||
2409 page->pgmap->owner != migrate->pgmap_owner)
2412 mpfn = migrate_pfn(page_to_pfn(page)) |
2413 MIGRATE_PFN_MIGRATE;
2414 if (is_write_device_private_entry(entry))
2415 mpfn |= MIGRATE_PFN_WRITE;
2417 if (!(migrate->flags & MIGRATE_VMA_SELECT_SYSTEM))
2420 if (is_zero_pfn(pfn)) {
2421 mpfn = MIGRATE_PFN_MIGRATE;
2425 page = vm_normal_page(migrate->vma, addr, pte);
2426 mpfn = migrate_pfn(pfn) | MIGRATE_PFN_MIGRATE;
2427 mpfn |= pte_write(pte) ? MIGRATE_PFN_WRITE : 0;
2430 /* FIXME support THP */
2431 if (!page || !page->mapping || PageTransCompound(page)) {
2437 * By getting a reference on the page we pin it and that blocks
2438 * any kind of migration. Side effect is that it "freezes" the
2441 * We drop this reference after isolating the page from the lru
2442 * for non device page (device page are not on the lru and thus
2443 * can't be dropped from it).
2449 * Optimize for the common case where page is only mapped once
2450 * in one process. If we can lock the page, then we can safely
2451 * set up a special migration page table entry now.
2453 if (trylock_page(page)) {
2456 mpfn |= MIGRATE_PFN_LOCKED;
2457 ptep_get_and_clear(mm, addr, ptep);
2459 /* Setup special migration page table entry */
2460 entry = make_migration_entry(page, mpfn &
2462 swp_pte = swp_entry_to_pte(entry);
2463 if (pte_present(pte)) {
2464 if (pte_soft_dirty(pte))
2465 swp_pte = pte_swp_mksoft_dirty(swp_pte);
2466 if (pte_uffd_wp(pte))
2467 swp_pte = pte_swp_mkuffd_wp(swp_pte);
2469 if (pte_swp_soft_dirty(pte))
2470 swp_pte = pte_swp_mksoft_dirty(swp_pte);
2471 if (pte_swp_uffd_wp(pte))
2472 swp_pte = pte_swp_mkuffd_wp(swp_pte);
2474 set_pte_at(mm, addr, ptep, swp_pte);
2477 * This is like regular unmap: we remove the rmap and
2478 * drop page refcount. Page won't be freed, as we took
2479 * a reference just above.
2481 page_remove_rmap(page, false);
2484 if (pte_present(pte))
2489 migrate->dst[migrate->npages] = 0;
2490 migrate->src[migrate->npages++] = mpfn;
2492 arch_leave_lazy_mmu_mode();
2493 pte_unmap_unlock(ptep - 1, ptl);
2495 /* Only flush the TLB if we actually modified any entries */
2497 flush_tlb_range(walk->vma, start, end);
2502 static const struct mm_walk_ops migrate_vma_walk_ops = {
2503 .pmd_entry = migrate_vma_collect_pmd,
2504 .pte_hole = migrate_vma_collect_hole,
2508 * migrate_vma_collect() - collect pages over a range of virtual addresses
2509 * @migrate: migrate struct containing all migration information
2511 * This will walk the CPU page table. For each virtual address backed by a
2512 * valid page, it updates the src array and takes a reference on the page, in
2513 * order to pin the page until we lock it and unmap it.
2515 static void migrate_vma_collect(struct migrate_vma *migrate)
2517 struct mmu_notifier_range range;
2520 * Note that the pgmap_owner is passed to the mmu notifier callback so
2521 * that the registered device driver can skip invalidating device
2522 * private page mappings that won't be migrated.
2524 mmu_notifier_range_init_migrate(&range, 0, migrate->vma,
2525 migrate->vma->vm_mm, migrate->start, migrate->end,
2526 migrate->pgmap_owner);
2527 mmu_notifier_invalidate_range_start(&range);
2529 walk_page_range(migrate->vma->vm_mm, migrate->start, migrate->end,
2530 &migrate_vma_walk_ops, migrate);
2532 mmu_notifier_invalidate_range_end(&range);
2533 migrate->end = migrate->start + (migrate->npages << PAGE_SHIFT);
2537 * migrate_vma_check_page() - check if page is pinned or not
2538 * @page: struct page to check
2540 * Pinned pages cannot be migrated. This is the same test as in
2541 * migrate_page_move_mapping(), except that here we allow migration of a
2544 static bool migrate_vma_check_page(struct page *page)
2547 * One extra ref because caller holds an extra reference, either from
2548 * isolate_lru_page() for a regular page, or migrate_vma_collect() for
2554 * FIXME support THP (transparent huge page), it is bit more complex to
2555 * check them than regular pages, because they can be mapped with a pmd
2556 * or with a pte (split pte mapping).
2558 if (PageCompound(page))
2561 /* Page from ZONE_DEVICE have one extra reference */
2562 if (is_zone_device_page(page)) {
2564 * Private page can never be pin as they have no valid pte and
2565 * GUP will fail for those. Yet if there is a pending migration
2566 * a thread might try to wait on the pte migration entry and
2567 * will bump the page reference count. Sadly there is no way to
2568 * differentiate a regular pin from migration wait. Hence to
2569 * avoid 2 racing thread trying to migrate back to CPU to enter
2570 * infinite loop (one stopping migration because the other is
2571 * waiting on pte migration entry). We always return true here.
2573 * FIXME proper solution is to rework migration_entry_wait() so
2574 * it does not need to take a reference on page.
2576 return is_device_private_page(page);
2579 /* For file back page */
2580 if (page_mapping(page))
2581 extra += 1 + page_has_private(page);
2583 if ((page_count(page) - extra) > page_mapcount(page))
2590 * migrate_vma_prepare() - lock pages and isolate them from the lru
2591 * @migrate: migrate struct containing all migration information
2593 * This locks pages that have been collected by migrate_vma_collect(). Once each
2594 * page is locked it is isolated from the lru (for non-device pages). Finally,
2595 * the ref taken by migrate_vma_collect() is dropped, as locked pages cannot be
2596 * migrated by concurrent kernel threads.
2598 static void migrate_vma_prepare(struct migrate_vma *migrate)
2600 const unsigned long npages = migrate->npages;
2601 const unsigned long start = migrate->start;
2602 unsigned long addr, i, restore = 0;
2603 bool allow_drain = true;
2607 for (i = 0; (i < npages) && migrate->cpages; i++) {
2608 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2614 if (!(migrate->src[i] & MIGRATE_PFN_LOCKED)) {
2616 * Because we are migrating several pages there can be
2617 * a deadlock between 2 concurrent migration where each
2618 * are waiting on each other page lock.
2620 * Make migrate_vma() a best effort thing and backoff
2621 * for any page we can not lock right away.
2623 if (!trylock_page(page)) {
2624 migrate->src[i] = 0;
2630 migrate->src[i] |= MIGRATE_PFN_LOCKED;
2633 /* ZONE_DEVICE pages are not on LRU */
2634 if (!is_zone_device_page(page)) {
2635 if (!PageLRU(page) && allow_drain) {
2636 /* Drain CPU's pagevec */
2637 lru_add_drain_all();
2638 allow_drain = false;
2641 if (isolate_lru_page(page)) {
2643 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2647 migrate->src[i] = 0;
2655 /* Drop the reference we took in collect */
2659 if (!migrate_vma_check_page(page)) {
2661 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2665 if (!is_zone_device_page(page)) {
2667 putback_lru_page(page);
2670 migrate->src[i] = 0;
2674 if (!is_zone_device_page(page))
2675 putback_lru_page(page);
2682 for (i = 0, addr = start; i < npages && restore; i++, addr += PAGE_SIZE) {
2683 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2685 if (!page || (migrate->src[i] & MIGRATE_PFN_MIGRATE))
2688 remove_migration_pte(page, migrate->vma, addr, page);
2690 migrate->src[i] = 0;
2698 * migrate_vma_unmap() - replace page mapping with special migration pte entry
2699 * @migrate: migrate struct containing all migration information
2701 * Replace page mapping (CPU page table pte) with a special migration pte entry
2702 * and check again if it has been pinned. Pinned pages are restored because we
2703 * cannot migrate them.
2705 * This is the last step before we call the device driver callback to allocate
2706 * destination memory and copy contents of original page over to new page.
2708 static void migrate_vma_unmap(struct migrate_vma *migrate)
2710 int flags = TTU_MIGRATION | TTU_IGNORE_MLOCK;
2711 const unsigned long npages = migrate->npages;
2712 const unsigned long start = migrate->start;
2713 unsigned long addr, i, restore = 0;
2715 for (i = 0; i < npages; i++) {
2716 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2718 if (!page || !(migrate->src[i] & MIGRATE_PFN_MIGRATE))
2721 if (page_mapped(page)) {
2722 try_to_unmap(page, flags);
2723 if (page_mapped(page))
2727 if (migrate_vma_check_page(page))
2731 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
2736 for (addr = start, i = 0; i < npages && restore; addr += PAGE_SIZE, i++) {
2737 struct page *page = migrate_pfn_to_page(migrate->src[i]);
2739 if (!page || (migrate->src[i] & MIGRATE_PFN_MIGRATE))
2742 remove_migration_ptes(page, page, false);
2744 migrate->src[i] = 0;
2748 if (is_zone_device_page(page))
2751 putback_lru_page(page);
2756 * migrate_vma_setup() - prepare to migrate a range of memory
2757 * @args: contains the vma, start, and pfns arrays for the migration
2759 * Returns: negative errno on failures, 0 when 0 or more pages were migrated
2762 * Prepare to migrate a range of memory virtual address range by collecting all
2763 * the pages backing each virtual address in the range, saving them inside the
2764 * src array. Then lock those pages and unmap them. Once the pages are locked
2765 * and unmapped, check whether each page is pinned or not. Pages that aren't
2766 * pinned have the MIGRATE_PFN_MIGRATE flag set (by this function) in the
2767 * corresponding src array entry. Then restores any pages that are pinned, by
2768 * remapping and unlocking those pages.
2770 * The caller should then allocate destination memory and copy source memory to
2771 * it for all those entries (ie with MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE
2772 * flag set). Once these are allocated and copied, the caller must update each
2773 * corresponding entry in the dst array with the pfn value of the destination
2774 * page and with the MIGRATE_PFN_VALID and MIGRATE_PFN_LOCKED flags set
2775 * (destination pages must have their struct pages locked, via lock_page()).
2777 * Note that the caller does not have to migrate all the pages that are marked
2778 * with MIGRATE_PFN_MIGRATE flag in src array unless this is a migration from
2779 * device memory to system memory. If the caller cannot migrate a device page
2780 * back to system memory, then it must return VM_FAULT_SIGBUS, which has severe
2781 * consequences for the userspace process, so it must be avoided if at all
2784 * For empty entries inside CPU page table (pte_none() or pmd_none() is true) we
2785 * do set MIGRATE_PFN_MIGRATE flag inside the corresponding source array thus
2786 * allowing the caller to allocate device memory for those unbacked virtual
2787 * addresses. For this the caller simply has to allocate device memory and
2788 * properly set the destination entry like for regular migration. Note that
2789 * this can still fail, and thus inside the device driver you must check if the
2790 * migration was successful for those entries after calling migrate_vma_pages(),
2791 * just like for regular migration.
2793 * After that, the callers must call migrate_vma_pages() to go over each entry
2794 * in the src array that has the MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE flag
2795 * set. If the corresponding entry in dst array has MIGRATE_PFN_VALID flag set,
2796 * then migrate_vma_pages() to migrate struct page information from the source
2797 * struct page to the destination struct page. If it fails to migrate the
2798 * struct page information, then it clears the MIGRATE_PFN_MIGRATE flag in the
2801 * At this point all successfully migrated pages have an entry in the src
2802 * array with MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE flag set and the dst
2803 * array entry with MIGRATE_PFN_VALID flag set.
2805 * Once migrate_vma_pages() returns the caller may inspect which pages were
2806 * successfully migrated, and which were not. Successfully migrated pages will
2807 * have the MIGRATE_PFN_MIGRATE flag set for their src array entry.
2809 * It is safe to update device page table after migrate_vma_pages() because
2810 * both destination and source page are still locked, and the mmap_lock is held
2811 * in read mode (hence no one can unmap the range being migrated).
2813 * Once the caller is done cleaning up things and updating its page table (if it
2814 * chose to do so, this is not an obligation) it finally calls
2815 * migrate_vma_finalize() to update the CPU page table to point to new pages
2816 * for successfully migrated pages or otherwise restore the CPU page table to
2817 * point to the original source pages.
2819 int migrate_vma_setup(struct migrate_vma *args)
2821 long nr_pages = (args->end - args->start) >> PAGE_SHIFT;
2823 args->start &= PAGE_MASK;
2824 args->end &= PAGE_MASK;
2825 if (!args->vma || is_vm_hugetlb_page(args->vma) ||
2826 (args->vma->vm_flags & VM_SPECIAL) || vma_is_dax(args->vma))
2830 if (args->start < args->vma->vm_start ||
2831 args->start >= args->vma->vm_end)
2833 if (args->end <= args->vma->vm_start || args->end > args->vma->vm_end)
2835 if (!args->src || !args->dst)
2838 memset(args->src, 0, sizeof(*args->src) * nr_pages);
2842 migrate_vma_collect(args);
2845 migrate_vma_prepare(args);
2847 migrate_vma_unmap(args);
2850 * At this point pages are locked and unmapped, and thus they have
2851 * stable content and can safely be copied to destination memory that
2852 * is allocated by the drivers.
2857 EXPORT_SYMBOL(migrate_vma_setup);
2860 * This code closely matches the code in:
2861 * __handle_mm_fault()
2862 * handle_pte_fault()
2863 * do_anonymous_page()
2864 * to map in an anonymous zero page but the struct page will be a ZONE_DEVICE
2867 static void migrate_vma_insert_page(struct migrate_vma *migrate,
2872 struct vm_area_struct *vma = migrate->vma;
2873 struct mm_struct *mm = vma->vm_mm;
2883 /* Only allow populating anonymous memory */
2884 if (!vma_is_anonymous(vma))
2887 pgdp = pgd_offset(mm, addr);
2888 p4dp = p4d_alloc(mm, pgdp, addr);
2891 pudp = pud_alloc(mm, p4dp, addr);
2894 pmdp = pmd_alloc(mm, pudp, addr);
2898 if (pmd_trans_huge(*pmdp) || pmd_devmap(*pmdp))
2902 * Use pte_alloc() instead of pte_alloc_map(). We can't run
2903 * pte_offset_map() on pmds where a huge pmd might be created
2904 * from a different thread.
2906 * pte_alloc_map() is safe to use under mmap_write_lock(mm) or when
2907 * parallel threads are excluded by other means.
2909 * Here we only have mmap_read_lock(mm).
2911 if (pte_alloc(mm, pmdp))
2914 /* See the comment in pte_alloc_one_map() */
2915 if (unlikely(pmd_trans_unstable(pmdp)))
2918 if (unlikely(anon_vma_prepare(vma)))
2920 if (mem_cgroup_charge(page, vma->vm_mm, GFP_KERNEL))
2924 * The memory barrier inside __SetPageUptodate makes sure that
2925 * preceding stores to the page contents become visible before
2926 * the set_pte_at() write.
2928 __SetPageUptodate(page);
2930 if (is_zone_device_page(page)) {
2931 if (is_device_private_page(page)) {
2932 swp_entry_t swp_entry;
2934 swp_entry = make_device_private_entry(page, vma->vm_flags & VM_WRITE);
2935 entry = swp_entry_to_pte(swp_entry);
2938 * For now we only support migrating to un-addressable
2941 pr_warn_once("Unsupported ZONE_DEVICE page type.\n");
2945 entry = mk_pte(page, vma->vm_page_prot);
2946 if (vma->vm_flags & VM_WRITE)
2947 entry = pte_mkwrite(pte_mkdirty(entry));
2950 ptep = pte_offset_map_lock(mm, pmdp, addr, &ptl);
2952 if (check_stable_address_space(mm))
2955 if (pte_present(*ptep)) {
2956 unsigned long pfn = pte_pfn(*ptep);
2958 if (!is_zero_pfn(pfn))
2961 } else if (!pte_none(*ptep))
2965 * Check for userfaultfd but do not deliver the fault. Instead,
2968 if (userfaultfd_missing(vma))
2971 inc_mm_counter(mm, MM_ANONPAGES);
2972 page_add_new_anon_rmap(page, vma, addr, false);
2973 if (!is_zone_device_page(page))
2974 lru_cache_add_inactive_or_unevictable(page, vma);
2978 flush_cache_page(vma, addr, pte_pfn(*ptep));
2979 ptep_clear_flush_notify(vma, addr, ptep);
2980 set_pte_at_notify(mm, addr, ptep, entry);
2981 update_mmu_cache(vma, addr, ptep);
2983 /* No need to invalidate - it was non-present before */
2984 set_pte_at(mm, addr, ptep, entry);
2985 update_mmu_cache(vma, addr, ptep);
2988 pte_unmap_unlock(ptep, ptl);
2989 *src = MIGRATE_PFN_MIGRATE;
2993 pte_unmap_unlock(ptep, ptl);
2995 *src &= ~MIGRATE_PFN_MIGRATE;
2999 * migrate_vma_pages() - migrate meta-data from src page to dst page
3000 * @migrate: migrate struct containing all migration information
3002 * This migrates struct page meta-data from source struct page to destination
3003 * struct page. This effectively finishes the migration from source page to the
3006 void migrate_vma_pages(struct migrate_vma *migrate)
3008 const unsigned long npages = migrate->npages;
3009 const unsigned long start = migrate->start;
3010 struct mmu_notifier_range range;
3011 unsigned long addr, i;
3012 bool notified = false;
3014 for (i = 0, addr = start; i < npages; addr += PAGE_SIZE, i++) {
3015 struct page *newpage = migrate_pfn_to_page(migrate->dst[i]);
3016 struct page *page = migrate_pfn_to_page(migrate->src[i]);
3017 struct address_space *mapping;
3021 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
3026 if (!(migrate->src[i] & MIGRATE_PFN_MIGRATE))
3031 mmu_notifier_range_init_migrate(&range, 0,
3032 migrate->vma, migrate->vma->vm_mm,
3034 migrate->pgmap_owner);
3035 mmu_notifier_invalidate_range_start(&range);
3037 migrate_vma_insert_page(migrate, addr, newpage,
3042 mapping = page_mapping(page);
3044 if (is_zone_device_page(newpage)) {
3045 if (is_device_private_page(newpage)) {
3047 * For now only support private anonymous when
3048 * migrating to un-addressable device memory.
3051 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
3056 * Other types of ZONE_DEVICE page are not
3059 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
3064 r = migrate_page(mapping, newpage, page, MIGRATE_SYNC_NO_COPY);
3065 if (r != MIGRATEPAGE_SUCCESS)
3066 migrate->src[i] &= ~MIGRATE_PFN_MIGRATE;
3070 * No need to double call mmu_notifier->invalidate_range() callback as
3071 * the above ptep_clear_flush_notify() inside migrate_vma_insert_page()
3072 * did already call it.
3075 mmu_notifier_invalidate_range_only_end(&range);
3077 EXPORT_SYMBOL(migrate_vma_pages);
3080 * migrate_vma_finalize() - restore CPU page table entry
3081 * @migrate: migrate struct containing all migration information
3083 * This replaces the special migration pte entry with either a mapping to the
3084 * new page if migration was successful for that page, or to the original page
3087 * This also unlocks the pages and puts them back on the lru, or drops the extra
3088 * refcount, for device pages.
3090 void migrate_vma_finalize(struct migrate_vma *migrate)
3092 const unsigned long npages = migrate->npages;
3095 for (i = 0; i < npages; i++) {
3096 struct page *newpage = migrate_pfn_to_page(migrate->dst[i]);
3097 struct page *page = migrate_pfn_to_page(migrate->src[i]);
3101 unlock_page(newpage);
3107 if (!(migrate->src[i] & MIGRATE_PFN_MIGRATE) || !newpage) {
3109 unlock_page(newpage);
3115 remove_migration_ptes(page, newpage, false);
3118 if (is_zone_device_page(page))
3121 putback_lru_page(page);
3123 if (newpage != page) {
3124 unlock_page(newpage);
3125 if (is_zone_device_page(newpage))
3128 putback_lru_page(newpage);
3132 EXPORT_SYMBOL(migrate_vma_finalize);
3133 #endif /* CONFIG_DEVICE_PRIVATE */