2 * Memory Migration functionality - linux/mm/migrate.c
4 * Copyright (C) 2006 Silicon Graphics, Inc., Christoph Lameter
6 * Page migration was first developed in the context of the memory hotplug
7 * project. The main authors of the migration code are:
9 * IWAMOTO Toshihiro <iwamoto@valinux.co.jp>
10 * Hirokazu Takahashi <taka@valinux.co.jp>
11 * Dave Hansen <haveblue@us.ibm.com>
15 #include <linux/migrate.h>
16 #include <linux/export.h>
17 #include <linux/swap.h>
18 #include <linux/swapops.h>
19 #include <linux/pagemap.h>
20 #include <linux/buffer_head.h>
21 #include <linux/mm_inline.h>
22 #include <linux/nsproxy.h>
23 #include <linux/pagevec.h>
24 #include <linux/ksm.h>
25 #include <linux/rmap.h>
26 #include <linux/topology.h>
27 #include <linux/cpu.h>
28 #include <linux/cpuset.h>
29 #include <linux/writeback.h>
30 #include <linux/mempolicy.h>
31 #include <linux/vmalloc.h>
32 #include <linux/security.h>
33 #include <linux/backing-dev.h>
34 #include <linux/compaction.h>
35 #include <linux/syscalls.h>
36 #include <linux/hugetlb.h>
37 #include <linux/hugetlb_cgroup.h>
38 #include <linux/gfp.h>
39 #include <linux/balloon_compaction.h>
40 #include <linux/mmu_notifier.h>
41 #include <linux/page_idle.h>
42 #include <linux/page_owner.h>
43 #include <linux/sched/mm.h>
44 #include <linux/ptrace.h>
46 #include <asm/tlbflush.h>
48 #define CREATE_TRACE_POINTS
49 #include <trace/events/migrate.h>
54 * migrate_prep() needs to be called before we start compiling a list of pages
55 * to be migrated using isolate_lru_page(). If scheduling work on other CPUs is
56 * undesirable, use migrate_prep_local()
58 int migrate_prep(void)
61 * Clear the LRU lists so pages can be isolated.
62 * Note that pages may be moved off the LRU after we have
63 * drained them. Those pages will fail to migrate like other
64 * pages that may be busy.
71 /* Do the necessary work of migrate_prep but not if it involves other CPUs */
72 int migrate_prep_local(void)
79 int isolate_movable_page(struct page *page, isolate_mode_t mode)
81 struct address_space *mapping;
84 * Avoid burning cycles with pages that are yet under __free_pages(),
85 * or just got freed under us.
87 * In case we 'win' a race for a movable page being freed under us and
88 * raise its refcount preventing __free_pages() from doing its job
89 * the put_page() at the end of this block will take care of
90 * release this page, thus avoiding a nasty leakage.
92 if (unlikely(!get_page_unless_zero(page)))
96 * Check PageMovable before holding a PG_lock because page's owner
97 * assumes anybody doesn't touch PG_lock of newly allocated page
98 * so unconditionally grapping the lock ruins page's owner side.
100 if (unlikely(!__PageMovable(page)))
103 * As movable pages are not isolated from LRU lists, concurrent
104 * compaction threads can race against page migration functions
105 * as well as race against the releasing a page.
107 * In order to avoid having an already isolated movable page
108 * being (wrongly) re-isolated while it is under migration,
109 * or to avoid attempting to isolate pages being released,
110 * lets be sure we have the page lock
111 * before proceeding with the movable page isolation steps.
113 if (unlikely(!trylock_page(page)))
116 if (!PageMovable(page) || PageIsolated(page))
117 goto out_no_isolated;
119 mapping = page_mapping(page);
120 VM_BUG_ON_PAGE(!mapping, page);
122 if (!mapping->a_ops->isolate_page(page, mode))
123 goto out_no_isolated;
125 /* Driver shouldn't use PG_isolated bit of page->flags */
126 WARN_ON_ONCE(PageIsolated(page));
127 __SetPageIsolated(page);
140 /* It should be called on page which is PG_movable */
141 void putback_movable_page(struct page *page)
143 struct address_space *mapping;
145 VM_BUG_ON_PAGE(!PageLocked(page), page);
146 VM_BUG_ON_PAGE(!PageMovable(page), page);
147 VM_BUG_ON_PAGE(!PageIsolated(page), page);
149 mapping = page_mapping(page);
150 mapping->a_ops->putback_page(page);
151 __ClearPageIsolated(page);
155 * Put previously isolated pages back onto the appropriate lists
156 * from where they were once taken off for compaction/migration.
158 * This function shall be used whenever the isolated pageset has been
159 * built from lru, balloon, hugetlbfs page. See isolate_migratepages_range()
160 * and isolate_huge_page().
162 void putback_movable_pages(struct list_head *l)
167 list_for_each_entry_safe(page, page2, l, lru) {
168 if (unlikely(PageHuge(page))) {
169 putback_active_hugepage(page);
172 list_del(&page->lru);
174 * We isolated non-lru movable page so here we can use
175 * __PageMovable because LRU page's mapping cannot have
176 * PAGE_MAPPING_MOVABLE.
178 if (unlikely(__PageMovable(page))) {
179 VM_BUG_ON_PAGE(!PageIsolated(page), page);
181 if (PageMovable(page))
182 putback_movable_page(page);
184 __ClearPageIsolated(page);
188 dec_node_page_state(page, NR_ISOLATED_ANON +
189 page_is_file_cache(page));
190 putback_lru_page(page);
196 * Restore a potential migration pte to a working pte entry
198 static bool remove_migration_pte(struct page *page, struct vm_area_struct *vma,
199 unsigned long addr, void *old)
201 struct page_vma_mapped_walk pvmw = {
205 .flags = PVMW_SYNC | PVMW_MIGRATION,
211 VM_BUG_ON_PAGE(PageTail(page), page);
212 while (page_vma_mapped_walk(&pvmw)) {
216 new = page - pvmw.page->index +
217 linear_page_index(vma, pvmw.address);
219 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
220 /* PMD-mapped THP migration entry */
222 VM_BUG_ON_PAGE(PageHuge(page) || !PageTransCompound(page), page);
223 remove_migration_pmd(&pvmw, new);
229 pte = pte_mkold(mk_pte(new, READ_ONCE(vma->vm_page_prot)));
230 if (pte_swp_soft_dirty(*pvmw.pte))
231 pte = pte_mksoft_dirty(pte);
234 * Recheck VMA as permissions can change since migration started
236 entry = pte_to_swp_entry(*pvmw.pte);
237 if (is_write_migration_entry(entry))
238 pte = maybe_mkwrite(pte, vma);
240 flush_dcache_page(new);
241 #ifdef CONFIG_HUGETLB_PAGE
243 pte = pte_mkhuge(pte);
244 pte = arch_make_huge_pte(pte, vma, new, 0);
245 set_huge_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte);
247 hugepage_add_anon_rmap(new, vma, pvmw.address);
249 page_dup_rmap(new, true);
253 set_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte);
256 page_add_anon_rmap(new, vma, pvmw.address, false);
258 page_add_file_rmap(new, false);
260 if (vma->vm_flags & VM_LOCKED && !PageTransCompound(new))
263 /* No need to invalidate - it was non-present before */
264 update_mmu_cache(vma, pvmw.address, pvmw.pte);
271 * Get rid of all migration entries and replace them by
272 * references to the indicated page.
274 void remove_migration_ptes(struct page *old, struct page *new, bool locked)
276 struct rmap_walk_control rwc = {
277 .rmap_one = remove_migration_pte,
282 rmap_walk_locked(new, &rwc);
284 rmap_walk(new, &rwc);
288 * Something used the pte of a page under migration. We need to
289 * get to the page and wait until migration is finished.
290 * When we return from this function the fault will be retried.
292 void __migration_entry_wait(struct mm_struct *mm, pte_t *ptep,
301 if (!is_swap_pte(pte))
304 entry = pte_to_swp_entry(pte);
305 if (!is_migration_entry(entry))
308 page = migration_entry_to_page(entry);
311 * Once radix-tree replacement of page migration started, page_count
312 * *must* be zero. And, we don't want to call wait_on_page_locked()
313 * against a page without get_page().
314 * So, we use get_page_unless_zero(), here. Even failed, page fault
317 if (!get_page_unless_zero(page))
319 pte_unmap_unlock(ptep, ptl);
320 wait_on_page_locked(page);
324 pte_unmap_unlock(ptep, ptl);
327 void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
328 unsigned long address)
330 spinlock_t *ptl = pte_lockptr(mm, pmd);
331 pte_t *ptep = pte_offset_map(pmd, address);
332 __migration_entry_wait(mm, ptep, ptl);
335 void migration_entry_wait_huge(struct vm_area_struct *vma,
336 struct mm_struct *mm, pte_t *pte)
338 spinlock_t *ptl = huge_pte_lockptr(hstate_vma(vma), mm, pte);
339 __migration_entry_wait(mm, pte, ptl);
342 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
343 void pmd_migration_entry_wait(struct mm_struct *mm, pmd_t *pmd)
348 ptl = pmd_lock(mm, pmd);
349 if (!is_pmd_migration_entry(*pmd))
351 page = migration_entry_to_page(pmd_to_swp_entry(*pmd));
352 if (!get_page_unless_zero(page))
355 wait_on_page_locked(page);
364 /* Returns true if all buffers are successfully locked */
365 static bool buffer_migrate_lock_buffers(struct buffer_head *head,
366 enum migrate_mode mode)
368 struct buffer_head *bh = head;
370 /* Simple case, sync compaction */
371 if (mode != MIGRATE_ASYNC) {
375 bh = bh->b_this_page;
377 } while (bh != head);
382 /* async case, we cannot block on lock_buffer so use trylock_buffer */
385 if (!trylock_buffer(bh)) {
387 * We failed to lock the buffer and cannot stall in
388 * async migration. Release the taken locks
390 struct buffer_head *failed_bh = bh;
393 while (bh != failed_bh) {
396 bh = bh->b_this_page;
401 bh = bh->b_this_page;
402 } while (bh != head);
406 static inline bool buffer_migrate_lock_buffers(struct buffer_head *head,
407 enum migrate_mode mode)
411 #endif /* CONFIG_BLOCK */
414 * Replace the page in the mapping.
416 * The number of remaining references must be:
417 * 1 for anonymous pages without a mapping
418 * 2 for pages with a mapping
419 * 3 for pages with a mapping and PagePrivate/PagePrivate2 set.
421 int migrate_page_move_mapping(struct address_space *mapping,
422 struct page *newpage, struct page *page,
423 struct buffer_head *head, enum migrate_mode mode,
426 struct zone *oldzone, *newzone;
428 int expected_count = 1 + extra_count;
432 /* Anonymous page without mapping */
433 if (page_count(page) != expected_count)
436 /* No turning back from here */
437 newpage->index = page->index;
438 newpage->mapping = page->mapping;
439 if (PageSwapBacked(page))
440 __SetPageSwapBacked(newpage);
442 return MIGRATEPAGE_SUCCESS;
445 oldzone = page_zone(page);
446 newzone = page_zone(newpage);
448 spin_lock_irq(&mapping->tree_lock);
450 pslot = radix_tree_lookup_slot(&mapping->page_tree,
453 expected_count += 1 + page_has_private(page);
454 if (page_count(page) != expected_count ||
455 radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) {
456 spin_unlock_irq(&mapping->tree_lock);
460 if (!page_ref_freeze(page, expected_count)) {
461 spin_unlock_irq(&mapping->tree_lock);
466 * In the async migration case of moving a page with buffers, lock the
467 * buffers using trylock before the mapping is moved. If the mapping
468 * was moved, we later failed to lock the buffers and could not move
469 * the mapping back due to an elevated page count, we would have to
470 * block waiting on other references to be dropped.
472 if (mode == MIGRATE_ASYNC && head &&
473 !buffer_migrate_lock_buffers(head, mode)) {
474 page_ref_unfreeze(page, expected_count);
475 spin_unlock_irq(&mapping->tree_lock);
480 * Now we know that no one else is looking at the page:
481 * no turning back from here.
483 newpage->index = page->index;
484 newpage->mapping = page->mapping;
485 get_page(newpage); /* add cache reference */
486 if (PageSwapBacked(page)) {
487 __SetPageSwapBacked(newpage);
488 if (PageSwapCache(page)) {
489 SetPageSwapCache(newpage);
490 set_page_private(newpage, page_private(page));
493 VM_BUG_ON_PAGE(PageSwapCache(page), page);
496 /* Move dirty while page refs frozen and newpage not yet exposed */
497 dirty = PageDirty(page);
499 ClearPageDirty(page);
500 SetPageDirty(newpage);
503 radix_tree_replace_slot(&mapping->page_tree, pslot, newpage);
506 * Drop cache reference from old page by unfreezing
507 * to one less reference.
508 * We know this isn't the last reference.
510 page_ref_unfreeze(page, expected_count - 1);
512 spin_unlock(&mapping->tree_lock);
513 /* Leave irq disabled to prevent preemption while updating stats */
516 * If moved to a different zone then also account
517 * the page for that zone. Other VM counters will be
518 * taken care of when we establish references to the
519 * new page and drop references to the old page.
521 * Note that anonymous pages are accounted for
522 * via NR_FILE_PAGES and NR_ANON_MAPPED if they
523 * are mapped to swap space.
525 if (newzone != oldzone) {
526 __dec_node_state(oldzone->zone_pgdat, NR_FILE_PAGES);
527 __inc_node_state(newzone->zone_pgdat, NR_FILE_PAGES);
528 if (PageSwapBacked(page) && !PageSwapCache(page)) {
529 __dec_node_state(oldzone->zone_pgdat, NR_SHMEM);
530 __inc_node_state(newzone->zone_pgdat, NR_SHMEM);
532 if (dirty && mapping_cap_account_dirty(mapping)) {
533 __dec_node_state(oldzone->zone_pgdat, NR_FILE_DIRTY);
534 __dec_zone_state(oldzone, NR_ZONE_WRITE_PENDING);
535 __inc_node_state(newzone->zone_pgdat, NR_FILE_DIRTY);
536 __inc_zone_state(newzone, NR_ZONE_WRITE_PENDING);
541 return MIGRATEPAGE_SUCCESS;
543 EXPORT_SYMBOL(migrate_page_move_mapping);
546 * The expected number of remaining references is the same as that
547 * of migrate_page_move_mapping().
549 int migrate_huge_page_move_mapping(struct address_space *mapping,
550 struct page *newpage, struct page *page)
555 spin_lock_irq(&mapping->tree_lock);
557 pslot = radix_tree_lookup_slot(&mapping->page_tree,
560 expected_count = 2 + page_has_private(page);
561 if (page_count(page) != expected_count ||
562 radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) {
563 spin_unlock_irq(&mapping->tree_lock);
567 if (!page_ref_freeze(page, expected_count)) {
568 spin_unlock_irq(&mapping->tree_lock);
572 newpage->index = page->index;
573 newpage->mapping = page->mapping;
577 radix_tree_replace_slot(&mapping->page_tree, pslot, newpage);
579 page_ref_unfreeze(page, expected_count - 1);
581 spin_unlock_irq(&mapping->tree_lock);
583 return MIGRATEPAGE_SUCCESS;
587 * Gigantic pages are so large that we do not guarantee that page++ pointer
588 * arithmetic will work across the entire page. We need something more
591 static void __copy_gigantic_page(struct page *dst, struct page *src,
595 struct page *dst_base = dst;
596 struct page *src_base = src;
598 for (i = 0; i < nr_pages; ) {
600 copy_highpage(dst, src);
603 dst = mem_map_next(dst, dst_base, i);
604 src = mem_map_next(src, src_base, i);
608 static void copy_huge_page(struct page *dst, struct page *src)
615 struct hstate *h = page_hstate(src);
616 nr_pages = pages_per_huge_page(h);
618 if (unlikely(nr_pages > MAX_ORDER_NR_PAGES)) {
619 __copy_gigantic_page(dst, src, nr_pages);
624 BUG_ON(!PageTransHuge(src));
625 nr_pages = hpage_nr_pages(src);
628 for (i = 0; i < nr_pages; i++) {
630 copy_highpage(dst + i, src + i);
635 * Copy the page to its new location
637 void migrate_page_copy(struct page *newpage, struct page *page)
641 if (PageHuge(page) || PageTransHuge(page))
642 copy_huge_page(newpage, page);
644 copy_highpage(newpage, page);
647 SetPageError(newpage);
648 if (PageReferenced(page))
649 SetPageReferenced(newpage);
650 if (PageUptodate(page))
651 SetPageUptodate(newpage);
652 if (TestClearPageActive(page)) {
653 VM_BUG_ON_PAGE(PageUnevictable(page), page);
654 SetPageActive(newpage);
655 } else if (TestClearPageUnevictable(page))
656 SetPageUnevictable(newpage);
657 if (PageChecked(page))
658 SetPageChecked(newpage);
659 if (PageMappedToDisk(page))
660 SetPageMappedToDisk(newpage);
662 /* Move dirty on pages not done by migrate_page_move_mapping() */
664 SetPageDirty(newpage);
666 if (page_is_young(page))
667 set_page_young(newpage);
668 if (page_is_idle(page))
669 set_page_idle(newpage);
672 * Copy NUMA information to the new page, to prevent over-eager
673 * future migrations of this same page.
675 cpupid = page_cpupid_xchg_last(page, -1);
676 page_cpupid_xchg_last(newpage, cpupid);
678 ksm_migrate_page(newpage, page);
680 * Please do not reorder this without considering how mm/ksm.c's
681 * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache().
683 if (PageSwapCache(page))
684 ClearPageSwapCache(page);
685 ClearPagePrivate(page);
686 set_page_private(page, 0);
689 * If any waiters have accumulated on the new page then
692 if (PageWriteback(newpage))
693 end_page_writeback(newpage);
695 copy_page_owner(page, newpage);
697 mem_cgroup_migrate(page, newpage);
699 EXPORT_SYMBOL(migrate_page_copy);
701 /************************************************************
702 * Migration functions
703 ***********************************************************/
706 * Common logic to directly migrate a single LRU page suitable for
707 * pages that do not use PagePrivate/PagePrivate2.
709 * Pages are locked upon entry and exit.
711 int migrate_page(struct address_space *mapping,
712 struct page *newpage, struct page *page,
713 enum migrate_mode mode)
717 BUG_ON(PageWriteback(page)); /* Writeback must be complete */
719 rc = migrate_page_move_mapping(mapping, newpage, page, NULL, mode, 0);
721 if (rc != MIGRATEPAGE_SUCCESS)
724 migrate_page_copy(newpage, page);
725 return MIGRATEPAGE_SUCCESS;
727 EXPORT_SYMBOL(migrate_page);
731 * Migration function for pages with buffers. This function can only be used
732 * if the underlying filesystem guarantees that no other references to "page"
735 int buffer_migrate_page(struct address_space *mapping,
736 struct page *newpage, struct page *page, enum migrate_mode mode)
738 struct buffer_head *bh, *head;
741 if (!page_has_buffers(page))
742 return migrate_page(mapping, newpage, page, mode);
744 head = page_buffers(page);
746 rc = migrate_page_move_mapping(mapping, newpage, page, head, mode, 0);
748 if (rc != MIGRATEPAGE_SUCCESS)
752 * In the async case, migrate_page_move_mapping locked the buffers
753 * with an IRQ-safe spinlock held. In the sync case, the buffers
754 * need to be locked now
756 if (mode != MIGRATE_ASYNC)
757 BUG_ON(!buffer_migrate_lock_buffers(head, mode));
759 ClearPagePrivate(page);
760 set_page_private(newpage, page_private(page));
761 set_page_private(page, 0);
767 set_bh_page(bh, newpage, bh_offset(bh));
768 bh = bh->b_this_page;
770 } while (bh != head);
772 SetPagePrivate(newpage);
774 migrate_page_copy(newpage, page);
780 bh = bh->b_this_page;
782 } while (bh != head);
784 return MIGRATEPAGE_SUCCESS;
786 EXPORT_SYMBOL(buffer_migrate_page);
790 * Writeback a page to clean the dirty state
792 static int writeout(struct address_space *mapping, struct page *page)
794 struct writeback_control wbc = {
795 .sync_mode = WB_SYNC_NONE,
798 .range_end = LLONG_MAX,
803 if (!mapping->a_ops->writepage)
804 /* No write method for the address space */
807 if (!clear_page_dirty_for_io(page))
808 /* Someone else already triggered a write */
812 * A dirty page may imply that the underlying filesystem has
813 * the page on some queue. So the page must be clean for
814 * migration. Writeout may mean we loose the lock and the
815 * page state is no longer what we checked for earlier.
816 * At this point we know that the migration attempt cannot
819 remove_migration_ptes(page, page, false);
821 rc = mapping->a_ops->writepage(page, &wbc);
823 if (rc != AOP_WRITEPAGE_ACTIVATE)
824 /* unlocked. Relock */
827 return (rc < 0) ? -EIO : -EAGAIN;
831 * Default handling if a filesystem does not provide a migration function.
833 static int fallback_migrate_page(struct address_space *mapping,
834 struct page *newpage, struct page *page, enum migrate_mode mode)
836 if (PageDirty(page)) {
837 /* Only writeback pages in full synchronous migration */
838 if (mode != MIGRATE_SYNC)
840 return writeout(mapping, page);
844 * Buffers may be managed in a filesystem specific way.
845 * We must have no buffers or drop them.
847 if (page_has_private(page) &&
848 !try_to_release_page(page, GFP_KERNEL))
851 return migrate_page(mapping, newpage, page, mode);
855 * Move a page to a newly allocated page
856 * The page is locked and all ptes have been successfully removed.
858 * The new page will have replaced the old page if this function
863 * MIGRATEPAGE_SUCCESS - success
865 static int move_to_new_page(struct page *newpage, struct page *page,
866 enum migrate_mode mode)
868 struct address_space *mapping;
870 bool is_lru = !__PageMovable(page);
872 VM_BUG_ON_PAGE(!PageLocked(page), page);
873 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
875 mapping = page_mapping(page);
877 if (likely(is_lru)) {
879 rc = migrate_page(mapping, newpage, page, mode);
880 else if (mapping->a_ops->migratepage)
882 * Most pages have a mapping and most filesystems
883 * provide a migratepage callback. Anonymous pages
884 * are part of swap space which also has its own
885 * migratepage callback. This is the most common path
886 * for page migration.
888 rc = mapping->a_ops->migratepage(mapping, newpage,
891 rc = fallback_migrate_page(mapping, newpage,
895 * In case of non-lru page, it could be released after
896 * isolation step. In that case, we shouldn't try migration.
898 VM_BUG_ON_PAGE(!PageIsolated(page), page);
899 if (!PageMovable(page)) {
900 rc = MIGRATEPAGE_SUCCESS;
901 __ClearPageIsolated(page);
905 rc = mapping->a_ops->migratepage(mapping, newpage,
907 WARN_ON_ONCE(rc == MIGRATEPAGE_SUCCESS &&
908 !PageIsolated(page));
912 * When successful, old pagecache page->mapping must be cleared before
913 * page is freed; but stats require that PageAnon be left as PageAnon.
915 if (rc == MIGRATEPAGE_SUCCESS) {
916 if (__PageMovable(page)) {
917 VM_BUG_ON_PAGE(!PageIsolated(page), page);
920 * We clear PG_movable under page_lock so any compactor
921 * cannot try to migrate this page.
923 __ClearPageIsolated(page);
927 * Anonymous and movable page->mapping will be cleard by
928 * free_pages_prepare so don't reset it here for keeping
929 * the type to work PageAnon, for example.
931 if (!PageMappingFlags(page))
932 page->mapping = NULL;
938 static int __unmap_and_move(struct page *page, struct page *newpage,
939 int force, enum migrate_mode mode)
942 int page_was_mapped = 0;
943 struct anon_vma *anon_vma = NULL;
944 bool is_lru = !__PageMovable(page);
946 if (!trylock_page(page)) {
947 if (!force || mode == MIGRATE_ASYNC)
951 * It's not safe for direct compaction to call lock_page.
952 * For example, during page readahead pages are added locked
953 * to the LRU. Later, when the IO completes the pages are
954 * marked uptodate and unlocked. However, the queueing
955 * could be merging multiple pages for one bio (e.g.
956 * mpage_readpages). If an allocation happens for the
957 * second or third page, the process can end up locking
958 * the same page twice and deadlocking. Rather than
959 * trying to be clever about what pages can be locked,
960 * avoid the use of lock_page for direct compaction
963 if (current->flags & PF_MEMALLOC)
969 if (PageWriteback(page)) {
971 * Only in the case of a full synchronous migration is it
972 * necessary to wait for PageWriteback. In the async case,
973 * the retry loop is too short and in the sync-light case,
974 * the overhead of stalling is too much
976 if (mode != MIGRATE_SYNC) {
982 wait_on_page_writeback(page);
986 * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
987 * we cannot notice that anon_vma is freed while we migrates a page.
988 * This get_anon_vma() delays freeing anon_vma pointer until the end
989 * of migration. File cache pages are no problem because of page_lock()
990 * File Caches may use write_page() or lock_page() in migration, then,
991 * just care Anon page here.
993 * Only page_get_anon_vma() understands the subtleties of
994 * getting a hold on an anon_vma from outside one of its mms.
995 * But if we cannot get anon_vma, then we won't need it anyway,
996 * because that implies that the anon page is no longer mapped
997 * (and cannot be remapped so long as we hold the page lock).
999 if (PageAnon(page) && !PageKsm(page))
1000 anon_vma = page_get_anon_vma(page);
1003 * Block others from accessing the new page when we get around to
1004 * establishing additional references. We are usually the only one
1005 * holding a reference to newpage at this point. We used to have a BUG
1006 * here if trylock_page(newpage) fails, but would like to allow for
1007 * cases where there might be a race with the previous use of newpage.
1008 * This is much like races on refcount of oldpage: just don't BUG().
1010 if (unlikely(!trylock_page(newpage)))
1013 if (unlikely(!is_lru)) {
1014 rc = move_to_new_page(newpage, page, mode);
1015 goto out_unlock_both;
1019 * Corner case handling:
1020 * 1. When a new swap-cache page is read into, it is added to the LRU
1021 * and treated as swapcache but it has no rmap yet.
1022 * Calling try_to_unmap() against a page->mapping==NULL page will
1023 * trigger a BUG. So handle it here.
1024 * 2. An orphaned page (see truncate_complete_page) might have
1025 * fs-private metadata. The page can be picked up due to memory
1026 * offlining. Everywhere else except page reclaim, the page is
1027 * invisible to the vm, so the page can not be migrated. So try to
1028 * free the metadata, so the page can be freed.
1030 if (!page->mapping) {
1031 VM_BUG_ON_PAGE(PageAnon(page), page);
1032 if (page_has_private(page)) {
1033 try_to_free_buffers(page);
1034 goto out_unlock_both;
1036 } else if (page_mapped(page)) {
1037 /* Establish migration ptes */
1038 VM_BUG_ON_PAGE(PageAnon(page) && !PageKsm(page) && !anon_vma,
1041 TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
1042 page_was_mapped = 1;
1045 if (!page_mapped(page))
1046 rc = move_to_new_page(newpage, page, mode);
1048 if (page_was_mapped)
1049 remove_migration_ptes(page,
1050 rc == MIGRATEPAGE_SUCCESS ? newpage : page, false);
1053 unlock_page(newpage);
1055 /* Drop an anon_vma reference if we took one */
1057 put_anon_vma(anon_vma);
1061 * If migration is successful, decrease refcount of the newpage
1062 * which will not free the page because new page owner increased
1063 * refcounter. As well, if it is LRU page, add the page to LRU
1066 if (rc == MIGRATEPAGE_SUCCESS) {
1067 if (unlikely(__PageMovable(newpage)))
1070 putback_lru_page(newpage);
1077 * gcc 4.7 and 4.8 on arm get an ICEs when inlining unmap_and_move(). Work
1080 #if (GCC_VERSION >= 40700 && GCC_VERSION < 40900) && defined(CONFIG_ARM)
1081 #define ICE_noinline noinline
1083 #define ICE_noinline
1087 * Obtain the lock on page, remove all ptes and migrate the page
1088 * to the newly allocated page in newpage.
1090 static ICE_noinline int unmap_and_move(new_page_t get_new_page,
1091 free_page_t put_new_page,
1092 unsigned long private, struct page *page,
1093 int force, enum migrate_mode mode,
1094 enum migrate_reason reason)
1096 int rc = MIGRATEPAGE_SUCCESS;
1098 struct page *newpage;
1100 newpage = get_new_page(page, private, &result);
1104 if (page_count(page) == 1) {
1105 /* page was freed from under us. So we are done. */
1106 ClearPageActive(page);
1107 ClearPageUnevictable(page);
1108 if (unlikely(__PageMovable(page))) {
1110 if (!PageMovable(page))
1111 __ClearPageIsolated(page);
1115 put_new_page(newpage, private);
1121 if (unlikely(PageTransHuge(page) && !PageTransHuge(newpage))) {
1123 rc = split_huge_page(page);
1129 rc = __unmap_and_move(page, newpage, force, mode);
1130 if (rc == MIGRATEPAGE_SUCCESS)
1131 set_page_owner_migrate_reason(newpage, reason);
1134 if (rc != -EAGAIN) {
1136 * A page that has been migrated has all references
1137 * removed and will be freed. A page that has not been
1138 * migrated will have kepts its references and be
1141 list_del(&page->lru);
1144 * Compaction can migrate also non-LRU pages which are
1145 * not accounted to NR_ISOLATED_*. They can be recognized
1148 if (likely(!__PageMovable(page)))
1149 dec_node_page_state(page, NR_ISOLATED_ANON +
1150 page_is_file_cache(page));
1154 * If migration is successful, releases reference grabbed during
1155 * isolation. Otherwise, restore the page to right list unless
1158 if (rc == MIGRATEPAGE_SUCCESS) {
1160 if (reason == MR_MEMORY_FAILURE) {
1162 * Set PG_HWPoison on just freed page
1163 * intentionally. Although it's rather weird,
1164 * it's how HWPoison flag works at the moment.
1166 if (!test_set_page_hwpoison(page))
1167 num_poisoned_pages_inc();
1170 if (rc != -EAGAIN) {
1171 if (likely(!__PageMovable(page))) {
1172 putback_lru_page(page);
1177 if (PageMovable(page))
1178 putback_movable_page(page);
1180 __ClearPageIsolated(page);
1186 put_new_page(newpage, private);
1195 *result = page_to_nid(newpage);
1201 * Counterpart of unmap_and_move_page() for hugepage migration.
1203 * This function doesn't wait the completion of hugepage I/O
1204 * because there is no race between I/O and migration for hugepage.
1205 * Note that currently hugepage I/O occurs only in direct I/O
1206 * where no lock is held and PG_writeback is irrelevant,
1207 * and writeback status of all subpages are counted in the reference
1208 * count of the head page (i.e. if all subpages of a 2MB hugepage are
1209 * under direct I/O, the reference of the head page is 512 and a bit more.)
1210 * This means that when we try to migrate hugepage whose subpages are
1211 * doing direct I/O, some references remain after try_to_unmap() and
1212 * hugepage migration fails without data corruption.
1214 * There is also no race when direct I/O is issued on the page under migration,
1215 * because then pte is replaced with migration swap entry and direct I/O code
1216 * will wait in the page fault for migration to complete.
1218 static int unmap_and_move_huge_page(new_page_t get_new_page,
1219 free_page_t put_new_page, unsigned long private,
1220 struct page *hpage, int force,
1221 enum migrate_mode mode, int reason)
1225 int page_was_mapped = 0;
1226 struct page *new_hpage;
1227 struct anon_vma *anon_vma = NULL;
1230 * Movability of hugepages depends on architectures and hugepage size.
1231 * This check is necessary because some callers of hugepage migration
1232 * like soft offline and memory hotremove don't walk through page
1233 * tables or check whether the hugepage is pmd-based or not before
1234 * kicking migration.
1236 if (!hugepage_migration_supported(page_hstate(hpage))) {
1237 putback_active_hugepage(hpage);
1241 new_hpage = get_new_page(hpage, private, &result);
1245 if (!trylock_page(hpage)) {
1246 if (!force || mode != MIGRATE_SYNC)
1251 if (PageAnon(hpage))
1252 anon_vma = page_get_anon_vma(hpage);
1254 if (unlikely(!trylock_page(new_hpage)))
1257 if (page_mapped(hpage)) {
1259 TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
1260 page_was_mapped = 1;
1263 if (!page_mapped(hpage))
1264 rc = move_to_new_page(new_hpage, hpage, mode);
1266 if (page_was_mapped)
1267 remove_migration_ptes(hpage,
1268 rc == MIGRATEPAGE_SUCCESS ? new_hpage : hpage, false);
1270 unlock_page(new_hpage);
1274 put_anon_vma(anon_vma);
1276 if (rc == MIGRATEPAGE_SUCCESS) {
1277 hugetlb_cgroup_migrate(hpage, new_hpage);
1278 put_new_page = NULL;
1279 set_page_owner_migrate_reason(new_hpage, reason);
1285 putback_active_hugepage(hpage);
1286 if (reason == MR_MEMORY_FAILURE && !test_set_page_hwpoison(hpage))
1287 num_poisoned_pages_inc();
1290 * If migration was not successful and there's a freeing callback, use
1291 * it. Otherwise, put_page() will drop the reference grabbed during
1295 put_new_page(new_hpage, private);
1297 putback_active_hugepage(new_hpage);
1303 *result = page_to_nid(new_hpage);
1309 * migrate_pages - migrate the pages specified in a list, to the free pages
1310 * supplied as the target for the page migration
1312 * @from: The list of pages to be migrated.
1313 * @get_new_page: The function used to allocate free pages to be used
1314 * as the target of the page migration.
1315 * @put_new_page: The function used to free target pages if migration
1316 * fails, or NULL if no special handling is necessary.
1317 * @private: Private data to be passed on to get_new_page()
1318 * @mode: The migration mode that specifies the constraints for
1319 * page migration, if any.
1320 * @reason: The reason for page migration.
1322 * The function returns after 10 attempts or if no pages are movable any more
1323 * because the list has become empty or no retryable pages exist any more.
1324 * The caller should call putback_movable_pages() to return pages to the LRU
1325 * or free list only if ret != 0.
1327 * Returns the number of pages that were not migrated, or an error code.
1329 int migrate_pages(struct list_head *from, new_page_t get_new_page,
1330 free_page_t put_new_page, unsigned long private,
1331 enum migrate_mode mode, int reason)
1335 int nr_succeeded = 0;
1339 int swapwrite = current->flags & PF_SWAPWRITE;
1343 current->flags |= PF_SWAPWRITE;
1345 for(pass = 0; pass < 10 && retry; pass++) {
1348 list_for_each_entry_safe(page, page2, from, lru) {
1352 rc = unmap_and_move_huge_page(get_new_page,
1353 put_new_page, private, page,
1354 pass > 2, mode, reason);
1356 rc = unmap_and_move(get_new_page, put_new_page,
1357 private, page, pass > 2, mode,
1367 case MIGRATEPAGE_SUCCESS:
1372 * Permanent failure (-EBUSY, -ENOSYS, etc.):
1373 * unlike -EAGAIN case, the failed page is
1374 * removed from migration page list and not
1375 * retried in the next outer loop.
1386 count_vm_events(PGMIGRATE_SUCCESS, nr_succeeded);
1388 count_vm_events(PGMIGRATE_FAIL, nr_failed);
1389 trace_mm_migrate_pages(nr_succeeded, nr_failed, mode, reason);
1392 current->flags &= ~PF_SWAPWRITE;
1399 * Move a list of individual pages
1401 struct page_to_node {
1408 static struct page *new_page_node(struct page *p, unsigned long private,
1411 struct page_to_node *pm = (struct page_to_node *)private;
1413 while (pm->node != MAX_NUMNODES && pm->page != p)
1416 if (pm->node == MAX_NUMNODES)
1419 *result = &pm->status;
1422 return alloc_huge_page_node(page_hstate(compound_head(p)),
1425 return __alloc_pages_node(pm->node,
1426 GFP_HIGHUSER_MOVABLE | __GFP_THISNODE, 0);
1430 * Move a set of pages as indicated in the pm array. The addr
1431 * field must be set to the virtual address of the page to be moved
1432 * and the node number must contain a valid target node.
1433 * The pm array ends with node = MAX_NUMNODES.
1435 static int do_move_page_to_node_array(struct mm_struct *mm,
1436 struct page_to_node *pm,
1440 struct page_to_node *pp;
1441 LIST_HEAD(pagelist);
1443 down_read(&mm->mmap_sem);
1446 * Build a list of pages to migrate
1448 for (pp = pm; pp->node != MAX_NUMNODES; pp++) {
1449 struct vm_area_struct *vma;
1453 vma = find_vma(mm, pp->addr);
1454 if (!vma || pp->addr < vma->vm_start || !vma_migratable(vma))
1457 /* FOLL_DUMP to ignore special (like zero) pages */
1458 page = follow_page(vma, pp->addr,
1459 FOLL_GET | FOLL_SPLIT | FOLL_DUMP);
1461 err = PTR_ERR(page);
1470 err = page_to_nid(page);
1472 if (err == pp->node)
1474 * Node already in the right place
1479 if (page_mapcount(page) > 1 &&
1483 if (PageHuge(page)) {
1485 isolate_huge_page(page, &pagelist);
1489 err = isolate_lru_page(page);
1491 list_add_tail(&page->lru, &pagelist);
1492 inc_node_page_state(page, NR_ISOLATED_ANON +
1493 page_is_file_cache(page));
1497 * Either remove the duplicate refcount from
1498 * isolate_lru_page() or drop the page ref if it was
1507 if (!list_empty(&pagelist)) {
1508 err = migrate_pages(&pagelist, new_page_node, NULL,
1509 (unsigned long)pm, MIGRATE_SYNC, MR_SYSCALL);
1511 putback_movable_pages(&pagelist);
1514 up_read(&mm->mmap_sem);
1519 * Migrate an array of page address onto an array of nodes and fill
1520 * the corresponding array of status.
1522 static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes,
1523 unsigned long nr_pages,
1524 const void __user * __user *pages,
1525 const int __user *nodes,
1526 int __user *status, int flags)
1528 struct page_to_node *pm;
1529 unsigned long chunk_nr_pages;
1530 unsigned long chunk_start;
1534 pm = (struct page_to_node *)__get_free_page(GFP_KERNEL);
1541 * Store a chunk of page_to_node array in a page,
1542 * but keep the last one as a marker
1544 chunk_nr_pages = (PAGE_SIZE / sizeof(struct page_to_node)) - 1;
1546 for (chunk_start = 0;
1547 chunk_start < nr_pages;
1548 chunk_start += chunk_nr_pages) {
1551 if (chunk_start + chunk_nr_pages > nr_pages)
1552 chunk_nr_pages = nr_pages - chunk_start;
1554 /* fill the chunk pm with addrs and nodes from user-space */
1555 for (j = 0; j < chunk_nr_pages; j++) {
1556 const void __user *p;
1560 if (get_user(p, pages + j + chunk_start))
1562 pm[j].addr = (unsigned long) p;
1564 if (get_user(node, nodes + j + chunk_start))
1568 if (node < 0 || node >= MAX_NUMNODES)
1571 if (!node_state(node, N_MEMORY))
1575 if (!node_isset(node, task_nodes))
1581 /* End marker for this chunk */
1582 pm[chunk_nr_pages].node = MAX_NUMNODES;
1584 /* Migrate this chunk */
1585 err = do_move_page_to_node_array(mm, pm,
1586 flags & MPOL_MF_MOVE_ALL);
1590 /* Return status information */
1591 for (j = 0; j < chunk_nr_pages; j++)
1592 if (put_user(pm[j].status, status + j + chunk_start)) {
1600 free_page((unsigned long)pm);
1606 * Determine the nodes of an array of pages and store it in an array of status.
1608 static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages,
1609 const void __user **pages, int *status)
1613 down_read(&mm->mmap_sem);
1615 for (i = 0; i < nr_pages; i++) {
1616 unsigned long addr = (unsigned long)(*pages);
1617 struct vm_area_struct *vma;
1621 vma = find_vma(mm, addr);
1622 if (!vma || addr < vma->vm_start)
1625 /* FOLL_DUMP to ignore special (like zero) pages */
1626 page = follow_page(vma, addr, FOLL_DUMP);
1628 err = PTR_ERR(page);
1632 err = page ? page_to_nid(page) : -ENOENT;
1640 up_read(&mm->mmap_sem);
1644 * Determine the nodes of a user array of pages and store it in
1645 * a user array of status.
1647 static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages,
1648 const void __user * __user *pages,
1651 #define DO_PAGES_STAT_CHUNK_NR 16
1652 const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR];
1653 int chunk_status[DO_PAGES_STAT_CHUNK_NR];
1656 unsigned long chunk_nr;
1658 chunk_nr = nr_pages;
1659 if (chunk_nr > DO_PAGES_STAT_CHUNK_NR)
1660 chunk_nr = DO_PAGES_STAT_CHUNK_NR;
1662 if (copy_from_user(chunk_pages, pages, chunk_nr * sizeof(*chunk_pages)))
1665 do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status);
1667 if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status)))
1672 nr_pages -= chunk_nr;
1674 return nr_pages ? -EFAULT : 0;
1678 * Move a list of pages in the address space of the currently executing
1681 SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages,
1682 const void __user * __user *, pages,
1683 const int __user *, nodes,
1684 int __user *, status, int, flags)
1686 struct task_struct *task;
1687 struct mm_struct *mm;
1689 nodemask_t task_nodes;
1692 if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
1695 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1698 /* Find the mm_struct */
1700 task = pid ? find_task_by_vpid(pid) : current;
1705 get_task_struct(task);
1708 * Check if this process has the right to modify the specified
1709 * process. Use the regular "ptrace_may_access()" checks.
1711 if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) {
1718 err = security_task_movememory(task);
1722 task_nodes = cpuset_mems_allowed(task);
1723 mm = get_task_mm(task);
1724 put_task_struct(task);
1730 err = do_pages_move(mm, task_nodes, nr_pages, pages,
1731 nodes, status, flags);
1733 err = do_pages_stat(mm, nr_pages, pages, status);
1739 put_task_struct(task);
1743 #ifdef CONFIG_NUMA_BALANCING
1745 * Returns true if this is a safe migration target node for misplaced NUMA
1746 * pages. Currently it only checks the watermarks which crude
1748 static bool migrate_balanced_pgdat(struct pglist_data *pgdat,
1749 unsigned long nr_migrate_pages)
1753 for (z = pgdat->nr_zones - 1; z >= 0; z--) {
1754 struct zone *zone = pgdat->node_zones + z;
1756 if (!populated_zone(zone))
1759 /* Avoid waking kswapd by allocating pages_to_migrate pages. */
1760 if (!zone_watermark_ok(zone, 0,
1761 high_wmark_pages(zone) +
1770 static struct page *alloc_misplaced_dst_page(struct page *page,
1774 int nid = (int) data;
1775 struct page *newpage;
1777 newpage = __alloc_pages_node(nid,
1778 (GFP_HIGHUSER_MOVABLE |
1779 __GFP_THISNODE | __GFP_NOMEMALLOC |
1780 __GFP_NORETRY | __GFP_NOWARN) &
1787 * page migration rate limiting control.
1788 * Do not migrate more than @pages_to_migrate in a @migrate_interval_millisecs
1789 * window of time. Default here says do not migrate more than 1280M per second.
1791 static unsigned int migrate_interval_millisecs __read_mostly = 100;
1792 static unsigned int ratelimit_pages __read_mostly = 128 << (20 - PAGE_SHIFT);
1794 /* Returns true if the node is migrate rate-limited after the update */
1795 static bool numamigrate_update_ratelimit(pg_data_t *pgdat,
1796 unsigned long nr_pages)
1799 * Rate-limit the amount of data that is being migrated to a node.
1800 * Optimal placement is no good if the memory bus is saturated and
1801 * all the time is being spent migrating!
1803 if (time_after(jiffies, pgdat->numabalancing_migrate_next_window)) {
1804 spin_lock(&pgdat->numabalancing_migrate_lock);
1805 pgdat->numabalancing_migrate_nr_pages = 0;
1806 pgdat->numabalancing_migrate_next_window = jiffies +
1807 msecs_to_jiffies(migrate_interval_millisecs);
1808 spin_unlock(&pgdat->numabalancing_migrate_lock);
1810 if (pgdat->numabalancing_migrate_nr_pages > ratelimit_pages) {
1811 trace_mm_numa_migrate_ratelimit(current, pgdat->node_id,
1817 * This is an unlocked non-atomic update so errors are possible.
1818 * The consequences are failing to migrate when we potentiall should
1819 * have which is not severe enough to warrant locking. If it is ever
1820 * a problem, it can be converted to a per-cpu counter.
1822 pgdat->numabalancing_migrate_nr_pages += nr_pages;
1826 static int numamigrate_isolate_page(pg_data_t *pgdat, struct page *page)
1830 VM_BUG_ON_PAGE(compound_order(page) && !PageTransHuge(page), page);
1832 /* Avoid migrating to a node that is nearly full */
1833 if (!migrate_balanced_pgdat(pgdat, 1UL << compound_order(page)))
1836 if (isolate_lru_page(page))
1840 * migrate_misplaced_transhuge_page() skips page migration's usual
1841 * check on page_count(), so we must do it here, now that the page
1842 * has been isolated: a GUP pin, or any other pin, prevents migration.
1843 * The expected page count is 3: 1 for page's mapcount and 1 for the
1844 * caller's pin and 1 for the reference taken by isolate_lru_page().
1846 if (PageTransHuge(page) && page_count(page) != 3) {
1847 putback_lru_page(page);
1851 page_lru = page_is_file_cache(page);
1852 mod_node_page_state(page_pgdat(page), NR_ISOLATED_ANON + page_lru,
1853 hpage_nr_pages(page));
1856 * Isolating the page has taken another reference, so the
1857 * caller's reference can be safely dropped without the page
1858 * disappearing underneath us during migration.
1864 bool pmd_trans_migrating(pmd_t pmd)
1866 struct page *page = pmd_page(pmd);
1867 return PageLocked(page);
1871 * Attempt to migrate a misplaced page to the specified destination
1872 * node. Caller is expected to have an elevated reference count on
1873 * the page that will be dropped by this function before returning.
1875 int migrate_misplaced_page(struct page *page, struct vm_area_struct *vma,
1878 pg_data_t *pgdat = NODE_DATA(node);
1881 LIST_HEAD(migratepages);
1884 * Don't migrate file pages that are mapped in multiple processes
1885 * with execute permissions as they are probably shared libraries.
1887 if (page_mapcount(page) != 1 && page_is_file_cache(page) &&
1888 (vma->vm_flags & VM_EXEC))
1892 * Rate-limit the amount of data that is being migrated to a node.
1893 * Optimal placement is no good if the memory bus is saturated and
1894 * all the time is being spent migrating!
1896 if (numamigrate_update_ratelimit(pgdat, 1))
1899 isolated = numamigrate_isolate_page(pgdat, page);
1903 list_add(&page->lru, &migratepages);
1904 nr_remaining = migrate_pages(&migratepages, alloc_misplaced_dst_page,
1905 NULL, node, MIGRATE_ASYNC,
1908 if (!list_empty(&migratepages)) {
1909 list_del(&page->lru);
1910 dec_node_page_state(page, NR_ISOLATED_ANON +
1911 page_is_file_cache(page));
1912 putback_lru_page(page);
1916 count_vm_numa_event(NUMA_PAGE_MIGRATE);
1917 BUG_ON(!list_empty(&migratepages));
1924 #endif /* CONFIG_NUMA_BALANCING */
1926 #if defined(CONFIG_NUMA_BALANCING) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
1928 * Migrates a THP to a given target node. page must be locked and is unlocked
1931 int migrate_misplaced_transhuge_page(struct mm_struct *mm,
1932 struct vm_area_struct *vma,
1933 pmd_t *pmd, pmd_t entry,
1934 unsigned long address,
1935 struct page *page, int node)
1938 pg_data_t *pgdat = NODE_DATA(node);
1940 struct page *new_page = NULL;
1941 int page_lru = page_is_file_cache(page);
1942 unsigned long mmun_start = address & HPAGE_PMD_MASK;
1943 unsigned long mmun_end = mmun_start + HPAGE_PMD_SIZE;
1946 * Rate-limit the amount of data that is being migrated to a node.
1947 * Optimal placement is no good if the memory bus is saturated and
1948 * all the time is being spent migrating!
1950 if (numamigrate_update_ratelimit(pgdat, HPAGE_PMD_NR))
1953 new_page = alloc_pages_node(node,
1954 (GFP_TRANSHUGE_LIGHT | __GFP_THISNODE),
1958 prep_transhuge_page(new_page);
1960 isolated = numamigrate_isolate_page(pgdat, page);
1966 /* Prepare a page as a migration target */
1967 __SetPageLocked(new_page);
1968 if (PageSwapBacked(page))
1969 __SetPageSwapBacked(new_page);
1971 /* anon mapping, we can simply copy page->mapping to the new page: */
1972 new_page->mapping = page->mapping;
1973 new_page->index = page->index;
1974 migrate_page_copy(new_page, page);
1975 WARN_ON(PageLRU(new_page));
1977 /* Recheck the target PMD */
1978 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1979 ptl = pmd_lock(mm, pmd);
1980 if (unlikely(!pmd_same(*pmd, entry) || !page_ref_freeze(page, 2))) {
1982 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1984 /* Reverse changes made by migrate_page_copy() */
1985 if (TestClearPageActive(new_page))
1986 SetPageActive(page);
1987 if (TestClearPageUnevictable(new_page))
1988 SetPageUnevictable(page);
1990 unlock_page(new_page);
1991 put_page(new_page); /* Free it */
1993 /* Retake the callers reference and putback on LRU */
1995 putback_lru_page(page);
1996 mod_node_page_state(page_pgdat(page),
1997 NR_ISOLATED_ANON + page_lru, -HPAGE_PMD_NR);
2002 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
2003 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
2006 * Clear the old entry under pagetable lock and establish the new PTE.
2007 * Any parallel GUP will either observe the old page blocking on the
2008 * page lock, block on the page table lock or observe the new page.
2009 * The SetPageUptodate on the new page and page_add_new_anon_rmap
2010 * guarantee the copy is visible before the pagetable update.
2012 flush_cache_range(vma, mmun_start, mmun_end);
2013 page_add_anon_rmap(new_page, vma, mmun_start, true);
2014 pmdp_huge_clear_flush_notify(vma, mmun_start, pmd);
2015 set_pmd_at(mm, mmun_start, pmd, entry);
2016 update_mmu_cache_pmd(vma, address, &entry);
2018 page_ref_unfreeze(page, 2);
2019 mlock_migrate_page(new_page, page);
2020 page_remove_rmap(page, true);
2021 set_page_owner_migrate_reason(new_page, MR_NUMA_MISPLACED);
2024 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2026 /* Take an "isolate" reference and put new page on the LRU. */
2028 putback_lru_page(new_page);
2030 unlock_page(new_page);
2032 put_page(page); /* Drop the rmap reference */
2033 put_page(page); /* Drop the LRU isolation reference */
2035 count_vm_events(PGMIGRATE_SUCCESS, HPAGE_PMD_NR);
2036 count_vm_numa_events(NUMA_PAGE_MIGRATE, HPAGE_PMD_NR);
2038 mod_node_page_state(page_pgdat(page),
2039 NR_ISOLATED_ANON + page_lru,
2044 count_vm_events(PGMIGRATE_FAIL, HPAGE_PMD_NR);
2046 ptl = pmd_lock(mm, pmd);
2047 if (pmd_same(*pmd, entry)) {
2048 entry = pmd_modify(entry, vma->vm_page_prot);
2049 set_pmd_at(mm, mmun_start, pmd, entry);
2050 update_mmu_cache_pmd(vma, address, &entry);
2059 #endif /* CONFIG_NUMA_BALANCING */
2061 #endif /* CONFIG_NUMA */