2 * Memory Migration functionality - linux/mm/migration.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/memcontrol.h>
34 #include <linux/syscalls.h>
35 #include <linux/hugetlb.h>
36 #include <linux/hugetlb_cgroup.h>
37 #include <linux/gfp.h>
38 #include <linux/balloon_compaction.h>
39 #include <linux/mmu_notifier.h>
41 #include <asm/tlbflush.h>
43 #define CREATE_TRACE_POINTS
44 #include <trace/events/migrate.h>
49 * migrate_prep() needs to be called before we start compiling a list of pages
50 * to be migrated using isolate_lru_page(). If scheduling work on other CPUs is
51 * undesirable, use migrate_prep_local()
53 int migrate_prep(void)
56 * Clear the LRU lists so pages can be isolated.
57 * Note that pages may be moved off the LRU after we have
58 * drained them. Those pages will fail to migrate like other
59 * pages that may be busy.
66 /* Do the necessary work of migrate_prep but not if it involves other CPUs */
67 int migrate_prep_local(void)
75 * Put previously isolated pages back onto the appropriate lists
76 * from where they were once taken off for compaction/migration.
78 * This function shall be used whenever the isolated pageset has been
79 * built from lru, balloon, hugetlbfs page. See isolate_migratepages_range()
80 * and isolate_huge_page().
82 void putback_movable_pages(struct list_head *l)
87 list_for_each_entry_safe(page, page2, l, lru) {
88 if (unlikely(PageHuge(page))) {
89 putback_active_hugepage(page);
93 dec_zone_page_state(page, NR_ISOLATED_ANON +
94 page_is_file_cache(page));
95 if (unlikely(isolated_balloon_page(page)))
96 balloon_page_putback(page);
98 putback_lru_page(page);
103 * Restore a potential migration pte to a working pte entry
105 static int remove_migration_pte(struct page *new, struct vm_area_struct *vma,
106 unsigned long addr, void *old)
108 struct mm_struct *mm = vma->vm_mm;
114 if (unlikely(PageHuge(new))) {
115 ptep = huge_pte_offset(mm, addr);
118 ptl = huge_pte_lockptr(hstate_vma(vma), mm, ptep);
120 pmd = mm_find_pmd(mm, addr);
123 if (pmd_trans_huge(*pmd))
126 ptep = pte_offset_map(pmd, addr);
129 * Peek to check is_swap_pte() before taking ptlock? No, we
130 * can race mremap's move_ptes(), which skips anon_vma lock.
133 ptl = pte_lockptr(mm, pmd);
138 if (!is_swap_pte(pte))
141 entry = pte_to_swp_entry(pte);
143 if (!is_migration_entry(entry) ||
144 migration_entry_to_page(entry) != old)
148 pte = pte_mkold(mk_pte(new, vma->vm_page_prot));
149 if (pte_swp_soft_dirty(*ptep))
150 pte = pte_mksoft_dirty(pte);
151 if (is_write_migration_entry(entry))
152 pte = pte_mkwrite(pte);
153 #ifdef CONFIG_HUGETLB_PAGE
155 pte = pte_mkhuge(pte);
156 pte = arch_make_huge_pte(pte, vma, new, 0);
159 flush_dcache_page(new);
160 set_pte_at(mm, addr, ptep, pte);
164 hugepage_add_anon_rmap(new, vma, addr);
167 } else if (PageAnon(new))
168 page_add_anon_rmap(new, vma, addr);
170 page_add_file_rmap(new);
172 /* No need to invalidate - it was non-present before */
173 update_mmu_cache(vma, addr, ptep);
175 pte_unmap_unlock(ptep, ptl);
181 * Get rid of all migration entries and replace them by
182 * references to the indicated page.
184 static void remove_migration_ptes(struct page *old, struct page *new)
186 struct rmap_walk_control rwc = {
187 .rmap_one = remove_migration_pte,
191 rmap_walk(new, &rwc);
195 * Something used the pte of a page under migration. We need to
196 * get to the page and wait until migration is finished.
197 * When we return from this function the fault will be retried.
199 static void __migration_entry_wait(struct mm_struct *mm, pte_t *ptep,
208 if (!is_swap_pte(pte))
211 entry = pte_to_swp_entry(pte);
212 if (!is_migration_entry(entry))
215 page = migration_entry_to_page(entry);
218 * Once radix-tree replacement of page migration started, page_count
219 * *must* be zero. And, we don't want to call wait_on_page_locked()
220 * against a page without get_page().
221 * So, we use get_page_unless_zero(), here. Even failed, page fault
224 if (!get_page_unless_zero(page))
226 pte_unmap_unlock(ptep, ptl);
227 wait_on_page_locked(page);
231 pte_unmap_unlock(ptep, ptl);
234 void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
235 unsigned long address)
237 spinlock_t *ptl = pte_lockptr(mm, pmd);
238 pte_t *ptep = pte_offset_map(pmd, address);
239 __migration_entry_wait(mm, ptep, ptl);
242 void migration_entry_wait_huge(struct vm_area_struct *vma,
243 struct mm_struct *mm, pte_t *pte)
245 spinlock_t *ptl = huge_pte_lockptr(hstate_vma(vma), mm, pte);
246 __migration_entry_wait(mm, pte, ptl);
250 /* Returns true if all buffers are successfully locked */
251 static bool buffer_migrate_lock_buffers(struct buffer_head *head,
252 enum migrate_mode mode)
254 struct buffer_head *bh = head;
256 /* Simple case, sync compaction */
257 if (mode != MIGRATE_ASYNC) {
261 bh = bh->b_this_page;
263 } while (bh != head);
268 /* async case, we cannot block on lock_buffer so use trylock_buffer */
271 if (!trylock_buffer(bh)) {
273 * We failed to lock the buffer and cannot stall in
274 * async migration. Release the taken locks
276 struct buffer_head *failed_bh = bh;
279 while (bh != failed_bh) {
282 bh = bh->b_this_page;
287 bh = bh->b_this_page;
288 } while (bh != head);
292 static inline bool buffer_migrate_lock_buffers(struct buffer_head *head,
293 enum migrate_mode mode)
297 #endif /* CONFIG_BLOCK */
300 * Replace the page in the mapping.
302 * The number of remaining references must be:
303 * 1 for anonymous pages without a mapping
304 * 2 for pages with a mapping
305 * 3 for pages with a mapping and PagePrivate/PagePrivate2 set.
307 int migrate_page_move_mapping(struct address_space *mapping,
308 struct page *newpage, struct page *page,
309 struct buffer_head *head, enum migrate_mode mode,
312 int expected_count = 1 + extra_count;
316 /* Anonymous page without mapping */
317 if (page_count(page) != expected_count)
319 return MIGRATEPAGE_SUCCESS;
322 spin_lock_irq(&mapping->tree_lock);
324 pslot = radix_tree_lookup_slot(&mapping->page_tree,
327 expected_count += 1 + page_has_private(page);
328 if (page_count(page) != expected_count ||
329 radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) {
330 spin_unlock_irq(&mapping->tree_lock);
334 if (!page_freeze_refs(page, expected_count)) {
335 spin_unlock_irq(&mapping->tree_lock);
340 * In the async migration case of moving a page with buffers, lock the
341 * buffers using trylock before the mapping is moved. If the mapping
342 * was moved, we later failed to lock the buffers and could not move
343 * the mapping back due to an elevated page count, we would have to
344 * block waiting on other references to be dropped.
346 if (mode == MIGRATE_ASYNC && head &&
347 !buffer_migrate_lock_buffers(head, mode)) {
348 page_unfreeze_refs(page, expected_count);
349 spin_unlock_irq(&mapping->tree_lock);
354 * Now we know that no one else is looking at the page.
356 get_page(newpage); /* add cache reference */
357 if (PageSwapCache(page)) {
358 SetPageSwapCache(newpage);
359 set_page_private(newpage, page_private(page));
362 radix_tree_replace_slot(pslot, newpage);
365 * Drop cache reference from old page by unfreezing
366 * to one less reference.
367 * We know this isn't the last reference.
369 page_unfreeze_refs(page, expected_count - 1);
372 * If moved to a different zone then also account
373 * the page for that zone. Other VM counters will be
374 * taken care of when we establish references to the
375 * new page and drop references to the old page.
377 * Note that anonymous pages are accounted for
378 * via NR_FILE_PAGES and NR_ANON_PAGES if they
379 * are mapped to swap space.
381 __dec_zone_page_state(page, NR_FILE_PAGES);
382 __inc_zone_page_state(newpage, NR_FILE_PAGES);
383 if (!PageSwapCache(page) && PageSwapBacked(page)) {
384 __dec_zone_page_state(page, NR_SHMEM);
385 __inc_zone_page_state(newpage, NR_SHMEM);
387 spin_unlock_irq(&mapping->tree_lock);
389 return MIGRATEPAGE_SUCCESS;
393 * The expected number of remaining references is the same as that
394 * of migrate_page_move_mapping().
396 int migrate_huge_page_move_mapping(struct address_space *mapping,
397 struct page *newpage, struct page *page)
403 if (page_count(page) != 1)
405 return MIGRATEPAGE_SUCCESS;
408 spin_lock_irq(&mapping->tree_lock);
410 pslot = radix_tree_lookup_slot(&mapping->page_tree,
413 expected_count = 2 + page_has_private(page);
414 if (page_count(page) != expected_count ||
415 radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) {
416 spin_unlock_irq(&mapping->tree_lock);
420 if (!page_freeze_refs(page, expected_count)) {
421 spin_unlock_irq(&mapping->tree_lock);
427 radix_tree_replace_slot(pslot, newpage);
429 page_unfreeze_refs(page, expected_count - 1);
431 spin_unlock_irq(&mapping->tree_lock);
432 return MIGRATEPAGE_SUCCESS;
436 * Gigantic pages are so large that we do not guarantee that page++ pointer
437 * arithmetic will work across the entire page. We need something more
440 static void __copy_gigantic_page(struct page *dst, struct page *src,
444 struct page *dst_base = dst;
445 struct page *src_base = src;
447 for (i = 0; i < nr_pages; ) {
449 copy_highpage(dst, src);
452 dst = mem_map_next(dst, dst_base, i);
453 src = mem_map_next(src, src_base, i);
457 static void copy_huge_page(struct page *dst, struct page *src)
464 struct hstate *h = page_hstate(src);
465 nr_pages = pages_per_huge_page(h);
467 if (unlikely(nr_pages > MAX_ORDER_NR_PAGES)) {
468 __copy_gigantic_page(dst, src, nr_pages);
473 BUG_ON(!PageTransHuge(src));
474 nr_pages = hpage_nr_pages(src);
477 for (i = 0; i < nr_pages; i++) {
479 copy_highpage(dst + i, src + i);
484 * Copy the page to its new location
486 void migrate_page_copy(struct page *newpage, struct page *page)
490 if (PageHuge(page) || PageTransHuge(page))
491 copy_huge_page(newpage, page);
493 copy_highpage(newpage, page);
496 SetPageError(newpage);
497 if (PageReferenced(page))
498 SetPageReferenced(newpage);
499 if (PageUptodate(page))
500 SetPageUptodate(newpage);
501 if (TestClearPageActive(page)) {
502 VM_BUG_ON_PAGE(PageUnevictable(page), page);
503 SetPageActive(newpage);
504 } else if (TestClearPageUnevictable(page))
505 SetPageUnevictable(newpage);
506 if (PageChecked(page))
507 SetPageChecked(newpage);
508 if (PageMappedToDisk(page))
509 SetPageMappedToDisk(newpage);
511 if (PageDirty(page)) {
512 clear_page_dirty_for_io(page);
514 * Want to mark the page and the radix tree as dirty, and
515 * redo the accounting that clear_page_dirty_for_io undid,
516 * but we can't use set_page_dirty because that function
517 * is actually a signal that all of the page has become dirty.
518 * Whereas only part of our page may be dirty.
520 if (PageSwapBacked(page))
521 SetPageDirty(newpage);
523 __set_page_dirty_nobuffers(newpage);
527 * Copy NUMA information to the new page, to prevent over-eager
528 * future migrations of this same page.
530 cpupid = page_cpupid_xchg_last(page, -1);
531 page_cpupid_xchg_last(newpage, cpupid);
533 mlock_migrate_page(newpage, page);
534 ksm_migrate_page(newpage, page);
536 * Please do not reorder this without considering how mm/ksm.c's
537 * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache().
539 ClearPageSwapCache(page);
540 ClearPagePrivate(page);
541 set_page_private(page, 0);
544 * If any waiters have accumulated on the new page then
547 if (PageWriteback(newpage))
548 end_page_writeback(newpage);
551 /************************************************************
552 * Migration functions
553 ***********************************************************/
556 * Common logic to directly migrate a single page suitable for
557 * pages that do not use PagePrivate/PagePrivate2.
559 * Pages are locked upon entry and exit.
561 int migrate_page(struct address_space *mapping,
562 struct page *newpage, struct page *page,
563 enum migrate_mode mode)
567 BUG_ON(PageWriteback(page)); /* Writeback must be complete */
569 rc = migrate_page_move_mapping(mapping, newpage, page, NULL, mode, 0);
571 if (rc != MIGRATEPAGE_SUCCESS)
574 migrate_page_copy(newpage, page);
575 return MIGRATEPAGE_SUCCESS;
577 EXPORT_SYMBOL(migrate_page);
581 * Migration function for pages with buffers. This function can only be used
582 * if the underlying filesystem guarantees that no other references to "page"
585 int buffer_migrate_page(struct address_space *mapping,
586 struct page *newpage, struct page *page, enum migrate_mode mode)
588 struct buffer_head *bh, *head;
591 if (!page_has_buffers(page))
592 return migrate_page(mapping, newpage, page, mode);
594 head = page_buffers(page);
596 rc = migrate_page_move_mapping(mapping, newpage, page, head, mode, 0);
598 if (rc != MIGRATEPAGE_SUCCESS)
602 * In the async case, migrate_page_move_mapping locked the buffers
603 * with an IRQ-safe spinlock held. In the sync case, the buffers
604 * need to be locked now
606 if (mode != MIGRATE_ASYNC)
607 BUG_ON(!buffer_migrate_lock_buffers(head, mode));
609 ClearPagePrivate(page);
610 set_page_private(newpage, page_private(page));
611 set_page_private(page, 0);
617 set_bh_page(bh, newpage, bh_offset(bh));
618 bh = bh->b_this_page;
620 } while (bh != head);
622 SetPagePrivate(newpage);
624 migrate_page_copy(newpage, page);
630 bh = bh->b_this_page;
632 } while (bh != head);
634 return MIGRATEPAGE_SUCCESS;
636 EXPORT_SYMBOL(buffer_migrate_page);
640 * Writeback a page to clean the dirty state
642 static int writeout(struct address_space *mapping, struct page *page)
644 struct writeback_control wbc = {
645 .sync_mode = WB_SYNC_NONE,
648 .range_end = LLONG_MAX,
653 if (!mapping->a_ops->writepage)
654 /* No write method for the address space */
657 if (!clear_page_dirty_for_io(page))
658 /* Someone else already triggered a write */
662 * A dirty page may imply that the underlying filesystem has
663 * the page on some queue. So the page must be clean for
664 * migration. Writeout may mean we loose the lock and the
665 * page state is no longer what we checked for earlier.
666 * At this point we know that the migration attempt cannot
669 remove_migration_ptes(page, page);
671 rc = mapping->a_ops->writepage(page, &wbc);
673 if (rc != AOP_WRITEPAGE_ACTIVATE)
674 /* unlocked. Relock */
677 return (rc < 0) ? -EIO : -EAGAIN;
681 * Default handling if a filesystem does not provide a migration function.
683 static int fallback_migrate_page(struct address_space *mapping,
684 struct page *newpage, struct page *page, enum migrate_mode mode)
686 if (PageDirty(page)) {
687 /* Only writeback pages in full synchronous migration */
688 if (mode != MIGRATE_SYNC)
690 return writeout(mapping, page);
694 * Buffers may be managed in a filesystem specific way.
695 * We must have no buffers or drop them.
697 if (page_has_private(page) &&
698 !try_to_release_page(page, GFP_KERNEL))
701 return migrate_page(mapping, newpage, page, mode);
705 * Move a page to a newly allocated page
706 * The page is locked and all ptes have been successfully removed.
708 * The new page will have replaced the old page if this function
713 * MIGRATEPAGE_SUCCESS - success
715 static int move_to_new_page(struct page *newpage, struct page *page,
716 int remap_swapcache, enum migrate_mode mode)
718 struct address_space *mapping;
722 * Block others from accessing the page when we get around to
723 * establishing additional references. We are the only one
724 * holding a reference to the new page at this point.
726 if (!trylock_page(newpage))
729 /* Prepare mapping for the new page.*/
730 newpage->index = page->index;
731 newpage->mapping = page->mapping;
732 if (PageSwapBacked(page))
733 SetPageSwapBacked(newpage);
735 mapping = page_mapping(page);
737 rc = migrate_page(mapping, newpage, page, mode);
738 else if (mapping->a_ops->migratepage)
740 * Most pages have a mapping and most filesystems provide a
741 * migratepage callback. Anonymous pages are part of swap
742 * space which also has its own migratepage callback. This
743 * is the most common path for page migration.
745 rc = mapping->a_ops->migratepage(mapping,
746 newpage, page, mode);
748 rc = fallback_migrate_page(mapping, newpage, page, mode);
750 if (rc != MIGRATEPAGE_SUCCESS) {
751 newpage->mapping = NULL;
754 remove_migration_ptes(page, newpage);
755 page->mapping = NULL;
758 unlock_page(newpage);
763 static int __unmap_and_move(struct page *page, struct page *newpage,
764 int force, enum migrate_mode mode)
767 int remap_swapcache = 1;
768 struct mem_cgroup *mem;
769 struct anon_vma *anon_vma = NULL;
771 if (!trylock_page(page)) {
772 if (!force || mode == MIGRATE_ASYNC)
776 * It's not safe for direct compaction to call lock_page.
777 * For example, during page readahead pages are added locked
778 * to the LRU. Later, when the IO completes the pages are
779 * marked uptodate and unlocked. However, the queueing
780 * could be merging multiple pages for one bio (e.g.
781 * mpage_readpages). If an allocation happens for the
782 * second or third page, the process can end up locking
783 * the same page twice and deadlocking. Rather than
784 * trying to be clever about what pages can be locked,
785 * avoid the use of lock_page for direct compaction
788 if (current->flags & PF_MEMALLOC)
794 /* charge against new page */
795 mem_cgroup_prepare_migration(page, newpage, &mem);
797 if (PageWriteback(page)) {
799 * Only in the case of a full synchronous migration is it
800 * necessary to wait for PageWriteback. In the async case,
801 * the retry loop is too short and in the sync-light case,
802 * the overhead of stalling is too much
804 if (mode != MIGRATE_SYNC) {
810 wait_on_page_writeback(page);
813 * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
814 * we cannot notice that anon_vma is freed while we migrates a page.
815 * This get_anon_vma() delays freeing anon_vma pointer until the end
816 * of migration. File cache pages are no problem because of page_lock()
817 * File Caches may use write_page() or lock_page() in migration, then,
818 * just care Anon page here.
820 if (PageAnon(page) && !PageKsm(page)) {
822 * Only page_lock_anon_vma_read() understands the subtleties of
823 * getting a hold on an anon_vma from outside one of its mms.
825 anon_vma = page_get_anon_vma(page);
830 } else if (PageSwapCache(page)) {
832 * We cannot be sure that the anon_vma of an unmapped
833 * swapcache page is safe to use because we don't
834 * know in advance if the VMA that this page belonged
835 * to still exists. If the VMA and others sharing the
836 * data have been freed, then the anon_vma could
837 * already be invalid.
839 * To avoid this possibility, swapcache pages get
840 * migrated but are not remapped when migration
849 if (unlikely(balloon_page_movable(page))) {
851 * A ballooned page does not need any special attention from
852 * physical to virtual reverse mapping procedures.
853 * Skip any attempt to unmap PTEs or to remap swap cache,
854 * in order to avoid burning cycles at rmap level, and perform
855 * the page migration right away (proteced by page lock).
857 rc = balloon_page_migrate(newpage, page, mode);
862 * Corner case handling:
863 * 1. When a new swap-cache page is read into, it is added to the LRU
864 * and treated as swapcache but it has no rmap yet.
865 * Calling try_to_unmap() against a page->mapping==NULL page will
866 * trigger a BUG. So handle it here.
867 * 2. An orphaned page (see truncate_complete_page) might have
868 * fs-private metadata. The page can be picked up due to memory
869 * offlining. Everywhere else except page reclaim, the page is
870 * invisible to the vm, so the page can not be migrated. So try to
871 * free the metadata, so the page can be freed.
873 if (!page->mapping) {
874 VM_BUG_ON_PAGE(PageAnon(page), page);
875 if (page_has_private(page)) {
876 try_to_free_buffers(page);
882 /* Establish migration ptes or remove ptes */
883 try_to_unmap(page, TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
886 if (!page_mapped(page))
887 rc = move_to_new_page(newpage, page, remap_swapcache, mode);
889 if (rc && remap_swapcache)
890 remove_migration_ptes(page, page);
892 /* Drop an anon_vma reference if we took one */
894 put_anon_vma(anon_vma);
897 mem_cgroup_end_migration(mem, page, newpage,
898 (rc == MIGRATEPAGE_SUCCESS ||
899 rc == MIGRATEPAGE_BALLOON_SUCCESS));
906 * Obtain the lock on page, remove all ptes and migrate the page
907 * to the newly allocated page in newpage.
909 static int unmap_and_move(new_page_t get_new_page, unsigned long private,
910 struct page *page, int force, enum migrate_mode mode)
914 struct page *newpage = get_new_page(page, private, &result);
919 if (page_count(page) == 1) {
920 /* page was freed from under us. So we are done. */
924 if (unlikely(PageTransHuge(page)))
925 if (unlikely(split_huge_page(page)))
928 rc = __unmap_and_move(page, newpage, force, mode);
930 if (unlikely(rc == MIGRATEPAGE_BALLOON_SUCCESS)) {
932 * A ballooned page has been migrated already.
933 * Now, it's the time to wrap-up counters,
934 * handle the page back to Buddy and return.
936 dec_zone_page_state(page, NR_ISOLATED_ANON +
937 page_is_file_cache(page));
938 balloon_page_free(page);
939 return MIGRATEPAGE_SUCCESS;
944 * A page that has been migrated has all references
945 * removed and will be freed. A page that has not been
946 * migrated will have kepts its references and be
949 list_del(&page->lru);
950 dec_zone_page_state(page, NR_ISOLATED_ANON +
951 page_is_file_cache(page));
952 putback_lru_page(page);
955 * Move the new page to the LRU. If migration was not successful
956 * then this will free the page.
958 putback_lru_page(newpage);
963 *result = page_to_nid(newpage);
969 * Counterpart of unmap_and_move_page() for hugepage migration.
971 * This function doesn't wait the completion of hugepage I/O
972 * because there is no race between I/O and migration for hugepage.
973 * Note that currently hugepage I/O occurs only in direct I/O
974 * where no lock is held and PG_writeback is irrelevant,
975 * and writeback status of all subpages are counted in the reference
976 * count of the head page (i.e. if all subpages of a 2MB hugepage are
977 * under direct I/O, the reference of the head page is 512 and a bit more.)
978 * This means that when we try to migrate hugepage whose subpages are
979 * doing direct I/O, some references remain after try_to_unmap() and
980 * hugepage migration fails without data corruption.
982 * There is also no race when direct I/O is issued on the page under migration,
983 * because then pte is replaced with migration swap entry and direct I/O code
984 * will wait in the page fault for migration to complete.
986 static int unmap_and_move_huge_page(new_page_t get_new_page,
987 unsigned long private, struct page *hpage,
988 int force, enum migrate_mode mode)
992 struct page *new_hpage;
993 struct anon_vma *anon_vma = NULL;
996 * Movability of hugepages depends on architectures and hugepage size.
997 * This check is necessary because some callers of hugepage migration
998 * like soft offline and memory hotremove don't walk through page
999 * tables or check whether the hugepage is pmd-based or not before
1000 * kicking migration.
1002 if (!hugepage_migration_support(page_hstate(hpage))) {
1003 putback_active_hugepage(hpage);
1007 new_hpage = get_new_page(hpage, private, &result);
1013 if (!trylock_page(hpage)) {
1014 if (!force || mode != MIGRATE_SYNC)
1019 if (PageAnon(hpage))
1020 anon_vma = page_get_anon_vma(hpage);
1022 try_to_unmap(hpage, TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
1024 if (!page_mapped(hpage))
1025 rc = move_to_new_page(new_hpage, hpage, 1, mode);
1028 remove_migration_ptes(hpage, hpage);
1031 put_anon_vma(anon_vma);
1034 hugetlb_cgroup_migrate(hpage, new_hpage);
1039 putback_active_hugepage(hpage);
1040 put_page(new_hpage);
1045 *result = page_to_nid(new_hpage);
1051 * migrate_pages - migrate the pages specified in a list, to the free pages
1052 * supplied as the target for the page migration
1054 * @from: The list of pages to be migrated.
1055 * @get_new_page: The function used to allocate free pages to be used
1056 * as the target of the page migration.
1057 * @private: Private data to be passed on to get_new_page()
1058 * @mode: The migration mode that specifies the constraints for
1059 * page migration, if any.
1060 * @reason: The reason for page migration.
1062 * The function returns after 10 attempts or if no pages are movable any more
1063 * because the list has become empty or no retryable pages exist any more.
1064 * The caller should call putback_lru_pages() to return pages to the LRU
1065 * or free list only if ret != 0.
1067 * Returns the number of pages that were not migrated, or an error code.
1069 int migrate_pages(struct list_head *from, new_page_t get_new_page,
1070 unsigned long private, enum migrate_mode mode, int reason)
1074 int nr_succeeded = 0;
1078 int swapwrite = current->flags & PF_SWAPWRITE;
1082 current->flags |= PF_SWAPWRITE;
1084 for(pass = 0; pass < 10 && retry; pass++) {
1087 list_for_each_entry_safe(page, page2, from, lru) {
1091 rc = unmap_and_move_huge_page(get_new_page,
1092 private, page, pass > 2, mode);
1094 rc = unmap_and_move(get_new_page, private,
1095 page, pass > 2, mode);
1103 case MIGRATEPAGE_SUCCESS:
1108 * Permanent failure (-EBUSY, -ENOSYS, etc.):
1109 * unlike -EAGAIN case, the failed page is
1110 * removed from migration page list and not
1111 * retried in the next outer loop.
1118 rc = nr_failed + retry;
1121 count_vm_events(PGMIGRATE_SUCCESS, nr_succeeded);
1123 count_vm_events(PGMIGRATE_FAIL, nr_failed);
1124 trace_mm_migrate_pages(nr_succeeded, nr_failed, mode, reason);
1127 current->flags &= ~PF_SWAPWRITE;
1134 * Move a list of individual pages
1136 struct page_to_node {
1143 static struct page *new_page_node(struct page *p, unsigned long private,
1146 struct page_to_node *pm = (struct page_to_node *)private;
1148 while (pm->node != MAX_NUMNODES && pm->page != p)
1151 if (pm->node == MAX_NUMNODES)
1154 *result = &pm->status;
1157 return alloc_huge_page_node(page_hstate(compound_head(p)),
1160 return alloc_pages_exact_node(pm->node,
1161 GFP_HIGHUSER_MOVABLE | GFP_THISNODE, 0);
1165 * Move a set of pages as indicated in the pm array. The addr
1166 * field must be set to the virtual address of the page to be moved
1167 * and the node number must contain a valid target node.
1168 * The pm array ends with node = MAX_NUMNODES.
1170 static int do_move_page_to_node_array(struct mm_struct *mm,
1171 struct page_to_node *pm,
1175 struct page_to_node *pp;
1176 LIST_HEAD(pagelist);
1178 down_read(&mm->mmap_sem);
1181 * Build a list of pages to migrate
1183 for (pp = pm; pp->node != MAX_NUMNODES; pp++) {
1184 struct vm_area_struct *vma;
1188 vma = find_vma(mm, pp->addr);
1189 if (!vma || pp->addr < vma->vm_start || !vma_migratable(vma))
1192 page = follow_page(vma, pp->addr, FOLL_GET|FOLL_SPLIT);
1194 err = PTR_ERR(page);
1202 /* Use PageReserved to check for zero page */
1203 if (PageReserved(page))
1207 err = page_to_nid(page);
1209 if (err == pp->node)
1211 * Node already in the right place
1216 if (page_mapcount(page) > 1 &&
1220 if (PageHuge(page)) {
1221 isolate_huge_page(page, &pagelist);
1225 err = isolate_lru_page(page);
1227 list_add_tail(&page->lru, &pagelist);
1228 inc_zone_page_state(page, NR_ISOLATED_ANON +
1229 page_is_file_cache(page));
1233 * Either remove the duplicate refcount from
1234 * isolate_lru_page() or drop the page ref if it was
1243 if (!list_empty(&pagelist)) {
1244 err = migrate_pages(&pagelist, new_page_node,
1245 (unsigned long)pm, MIGRATE_SYNC, MR_SYSCALL);
1247 putback_movable_pages(&pagelist);
1250 up_read(&mm->mmap_sem);
1255 * Migrate an array of page address onto an array of nodes and fill
1256 * the corresponding array of status.
1258 static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes,
1259 unsigned long nr_pages,
1260 const void __user * __user *pages,
1261 const int __user *nodes,
1262 int __user *status, int flags)
1264 struct page_to_node *pm;
1265 unsigned long chunk_nr_pages;
1266 unsigned long chunk_start;
1270 pm = (struct page_to_node *)__get_free_page(GFP_KERNEL);
1277 * Store a chunk of page_to_node array in a page,
1278 * but keep the last one as a marker
1280 chunk_nr_pages = (PAGE_SIZE / sizeof(struct page_to_node)) - 1;
1282 for (chunk_start = 0;
1283 chunk_start < nr_pages;
1284 chunk_start += chunk_nr_pages) {
1287 if (chunk_start + chunk_nr_pages > nr_pages)
1288 chunk_nr_pages = nr_pages - chunk_start;
1290 /* fill the chunk pm with addrs and nodes from user-space */
1291 for (j = 0; j < chunk_nr_pages; j++) {
1292 const void __user *p;
1296 if (get_user(p, pages + j + chunk_start))
1298 pm[j].addr = (unsigned long) p;
1300 if (get_user(node, nodes + j + chunk_start))
1304 if (node < 0 || node >= MAX_NUMNODES)
1307 if (!node_state(node, N_MEMORY))
1311 if (!node_isset(node, task_nodes))
1317 /* End marker for this chunk */
1318 pm[chunk_nr_pages].node = MAX_NUMNODES;
1320 /* Migrate this chunk */
1321 err = do_move_page_to_node_array(mm, pm,
1322 flags & MPOL_MF_MOVE_ALL);
1326 /* Return status information */
1327 for (j = 0; j < chunk_nr_pages; j++)
1328 if (put_user(pm[j].status, status + j + chunk_start)) {
1336 free_page((unsigned long)pm);
1342 * Determine the nodes of an array of pages and store it in an array of status.
1344 static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages,
1345 const void __user **pages, int *status)
1349 down_read(&mm->mmap_sem);
1351 for (i = 0; i < nr_pages; i++) {
1352 unsigned long addr = (unsigned long)(*pages);
1353 struct vm_area_struct *vma;
1357 vma = find_vma(mm, addr);
1358 if (!vma || addr < vma->vm_start)
1361 page = follow_page(vma, addr, 0);
1363 err = PTR_ERR(page);
1368 /* Use PageReserved to check for zero page */
1369 if (!page || PageReserved(page))
1372 err = page_to_nid(page);
1380 up_read(&mm->mmap_sem);
1384 * Determine the nodes of a user array of pages and store it in
1385 * a user array of status.
1387 static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages,
1388 const void __user * __user *pages,
1391 #define DO_PAGES_STAT_CHUNK_NR 16
1392 const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR];
1393 int chunk_status[DO_PAGES_STAT_CHUNK_NR];
1396 unsigned long chunk_nr;
1398 chunk_nr = nr_pages;
1399 if (chunk_nr > DO_PAGES_STAT_CHUNK_NR)
1400 chunk_nr = DO_PAGES_STAT_CHUNK_NR;
1402 if (copy_from_user(chunk_pages, pages, chunk_nr * sizeof(*chunk_pages)))
1405 do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status);
1407 if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status)))
1412 nr_pages -= chunk_nr;
1414 return nr_pages ? -EFAULT : 0;
1418 * Move a list of pages in the address space of the currently executing
1421 SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages,
1422 const void __user * __user *, pages,
1423 const int __user *, nodes,
1424 int __user *, status, int, flags)
1426 const struct cred *cred = current_cred(), *tcred;
1427 struct task_struct *task;
1428 struct mm_struct *mm;
1430 nodemask_t task_nodes;
1433 if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
1436 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1439 /* Find the mm_struct */
1441 task = pid ? find_task_by_vpid(pid) : current;
1446 get_task_struct(task);
1449 * Check if this process has the right to modify the specified
1450 * process. The right exists if the process has administrative
1451 * capabilities, superuser privileges or the same
1452 * userid as the target process.
1454 tcred = __task_cred(task);
1455 if (!uid_eq(cred->euid, tcred->suid) && !uid_eq(cred->euid, tcred->uid) &&
1456 !uid_eq(cred->uid, tcred->suid) && !uid_eq(cred->uid, tcred->uid) &&
1457 !capable(CAP_SYS_NICE)) {
1464 err = security_task_movememory(task);
1468 task_nodes = cpuset_mems_allowed(task);
1469 mm = get_task_mm(task);
1470 put_task_struct(task);
1476 err = do_pages_move(mm, task_nodes, nr_pages, pages,
1477 nodes, status, flags);
1479 err = do_pages_stat(mm, nr_pages, pages, status);
1485 put_task_struct(task);
1490 * Call migration functions in the vma_ops that may prepare
1491 * memory in a vm for migration. migration functions may perform
1492 * the migration for vmas that do not have an underlying page struct.
1494 int migrate_vmas(struct mm_struct *mm, const nodemask_t *to,
1495 const nodemask_t *from, unsigned long flags)
1497 struct vm_area_struct *vma;
1500 for (vma = mm->mmap; vma && !err; vma = vma->vm_next) {
1501 if (vma->vm_ops && vma->vm_ops->migrate) {
1502 err = vma->vm_ops->migrate(vma, to, from, flags);
1510 #ifdef CONFIG_NUMA_BALANCING
1512 * Returns true if this is a safe migration target node for misplaced NUMA
1513 * pages. Currently it only checks the watermarks which crude
1515 static bool migrate_balanced_pgdat(struct pglist_data *pgdat,
1516 unsigned long nr_migrate_pages)
1519 for (z = pgdat->nr_zones - 1; z >= 0; z--) {
1520 struct zone *zone = pgdat->node_zones + z;
1522 if (!populated_zone(zone))
1525 if (!zone_reclaimable(zone))
1528 /* Avoid waking kswapd by allocating pages_to_migrate pages. */
1529 if (!zone_watermark_ok(zone, 0,
1530 high_wmark_pages(zone) +
1539 static struct page *alloc_misplaced_dst_page(struct page *page,
1543 int nid = (int) data;
1544 struct page *newpage;
1546 newpage = alloc_pages_exact_node(nid,
1547 (GFP_HIGHUSER_MOVABLE | GFP_THISNODE |
1548 __GFP_NOMEMALLOC | __GFP_NORETRY |
1556 * page migration rate limiting control.
1557 * Do not migrate more than @pages_to_migrate in a @migrate_interval_millisecs
1558 * window of time. Default here says do not migrate more than 1280M per second.
1559 * If a node is rate-limited then PTE NUMA updates are also rate-limited. However
1560 * as it is faults that reset the window, pte updates will happen unconditionally
1561 * if there has not been a fault since @pteupdate_interval_millisecs after the
1562 * throttle window closed.
1564 static unsigned int migrate_interval_millisecs __read_mostly = 100;
1565 static unsigned int pteupdate_interval_millisecs __read_mostly = 1000;
1566 static unsigned int ratelimit_pages __read_mostly = 128 << (20 - PAGE_SHIFT);
1568 /* Returns true if NUMA migration is currently rate limited */
1569 bool migrate_ratelimited(int node)
1571 pg_data_t *pgdat = NODE_DATA(node);
1573 if (time_after(jiffies, pgdat->numabalancing_migrate_next_window +
1574 msecs_to_jiffies(pteupdate_interval_millisecs)))
1577 if (pgdat->numabalancing_migrate_nr_pages < ratelimit_pages)
1583 /* Returns true if the node is migrate rate-limited after the update */
1584 static bool numamigrate_update_ratelimit(pg_data_t *pgdat,
1585 unsigned long nr_pages)
1588 * Rate-limit the amount of data that is being migrated to a node.
1589 * Optimal placement is no good if the memory bus is saturated and
1590 * all the time is being spent migrating!
1592 if (time_after(jiffies, pgdat->numabalancing_migrate_next_window)) {
1593 spin_lock(&pgdat->numabalancing_migrate_lock);
1594 pgdat->numabalancing_migrate_nr_pages = 0;
1595 pgdat->numabalancing_migrate_next_window = jiffies +
1596 msecs_to_jiffies(migrate_interval_millisecs);
1597 spin_unlock(&pgdat->numabalancing_migrate_lock);
1599 if (pgdat->numabalancing_migrate_nr_pages > ratelimit_pages) {
1600 trace_mm_numa_migrate_ratelimit(current, pgdat->node_id,
1606 * This is an unlocked non-atomic update so errors are possible.
1607 * The consequences are failing to migrate when we potentiall should
1608 * have which is not severe enough to warrant locking. If it is ever
1609 * a problem, it can be converted to a per-cpu counter.
1611 pgdat->numabalancing_migrate_nr_pages += nr_pages;
1615 static int numamigrate_isolate_page(pg_data_t *pgdat, struct page *page)
1619 VM_BUG_ON_PAGE(compound_order(page) && !PageTransHuge(page), page);
1621 /* Avoid migrating to a node that is nearly full */
1622 if (!migrate_balanced_pgdat(pgdat, 1UL << compound_order(page)))
1625 if (isolate_lru_page(page))
1629 * migrate_misplaced_transhuge_page() skips page migration's usual
1630 * check on page_count(), so we must do it here, now that the page
1631 * has been isolated: a GUP pin, or any other pin, prevents migration.
1632 * The expected page count is 3: 1 for page's mapcount and 1 for the
1633 * caller's pin and 1 for the reference taken by isolate_lru_page().
1635 if (PageTransHuge(page) && page_count(page) != 3) {
1636 putback_lru_page(page);
1640 page_lru = page_is_file_cache(page);
1641 mod_zone_page_state(page_zone(page), NR_ISOLATED_ANON + page_lru,
1642 hpage_nr_pages(page));
1645 * Isolating the page has taken another reference, so the
1646 * caller's reference can be safely dropped without the page
1647 * disappearing underneath us during migration.
1653 bool pmd_trans_migrating(pmd_t pmd)
1655 struct page *page = pmd_page(pmd);
1656 return PageLocked(page);
1659 void wait_migrate_huge_page(struct anon_vma *anon_vma, pmd_t *pmd)
1661 struct page *page = pmd_page(*pmd);
1662 wait_on_page_locked(page);
1666 * Attempt to migrate a misplaced page to the specified destination
1667 * node. Caller is expected to have an elevated reference count on
1668 * the page that will be dropped by this function before returning.
1670 int migrate_misplaced_page(struct page *page, struct vm_area_struct *vma,
1673 pg_data_t *pgdat = NODE_DATA(node);
1676 LIST_HEAD(migratepages);
1679 * Don't migrate file pages that are mapped in multiple processes
1680 * with execute permissions as they are probably shared libraries.
1682 if (page_mapcount(page) != 1 && page_is_file_cache(page) &&
1683 (vma->vm_flags & VM_EXEC))
1687 * Rate-limit the amount of data that is being migrated to a node.
1688 * Optimal placement is no good if the memory bus is saturated and
1689 * all the time is being spent migrating!
1691 if (numamigrate_update_ratelimit(pgdat, 1))
1694 isolated = numamigrate_isolate_page(pgdat, page);
1698 list_add(&page->lru, &migratepages);
1699 nr_remaining = migrate_pages(&migratepages, alloc_misplaced_dst_page,
1700 node, MIGRATE_ASYNC, MR_NUMA_MISPLACED);
1702 if (!list_empty(&migratepages)) {
1703 list_del(&page->lru);
1704 dec_zone_page_state(page, NR_ISOLATED_ANON +
1705 page_is_file_cache(page));
1706 putback_lru_page(page);
1710 count_vm_numa_event(NUMA_PAGE_MIGRATE);
1711 BUG_ON(!list_empty(&migratepages));
1718 #endif /* CONFIG_NUMA_BALANCING */
1720 #if defined(CONFIG_NUMA_BALANCING) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
1722 * Migrates a THP to a given target node. page must be locked and is unlocked
1725 int migrate_misplaced_transhuge_page(struct mm_struct *mm,
1726 struct vm_area_struct *vma,
1727 pmd_t *pmd, pmd_t entry,
1728 unsigned long address,
1729 struct page *page, int node)
1732 pg_data_t *pgdat = NODE_DATA(node);
1734 struct page *new_page = NULL;
1735 struct mem_cgroup *memcg = NULL;
1736 int page_lru = page_is_file_cache(page);
1737 unsigned long mmun_start = address & HPAGE_PMD_MASK;
1738 unsigned long mmun_end = mmun_start + HPAGE_PMD_SIZE;
1742 * Rate-limit the amount of data that is being migrated to a node.
1743 * Optimal placement is no good if the memory bus is saturated and
1744 * all the time is being spent migrating!
1746 if (numamigrate_update_ratelimit(pgdat, HPAGE_PMD_NR))
1749 new_page = alloc_pages_node(node,
1750 (GFP_TRANSHUGE | GFP_THISNODE) & ~__GFP_WAIT, HPAGE_PMD_ORDER);
1754 isolated = numamigrate_isolate_page(pgdat, page);
1760 if (mm_tlb_flush_pending(mm))
1761 flush_tlb_range(vma, mmun_start, mmun_end);
1763 /* Prepare a page as a migration target */
1764 __set_page_locked(new_page);
1765 SetPageSwapBacked(new_page);
1767 /* anon mapping, we can simply copy page->mapping to the new page: */
1768 new_page->mapping = page->mapping;
1769 new_page->index = page->index;
1770 migrate_page_copy(new_page, page);
1771 WARN_ON(PageLRU(new_page));
1773 /* Recheck the target PMD */
1774 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1775 ptl = pmd_lock(mm, pmd);
1776 if (unlikely(!pmd_same(*pmd, entry) || page_count(page) != 2)) {
1779 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1781 /* Reverse changes made by migrate_page_copy() */
1782 if (TestClearPageActive(new_page))
1783 SetPageActive(page);
1784 if (TestClearPageUnevictable(new_page))
1785 SetPageUnevictable(page);
1786 mlock_migrate_page(page, new_page);
1788 unlock_page(new_page);
1789 put_page(new_page); /* Free it */
1791 /* Retake the callers reference and putback on LRU */
1793 putback_lru_page(page);
1794 mod_zone_page_state(page_zone(page),
1795 NR_ISOLATED_ANON + page_lru, -HPAGE_PMD_NR);
1801 * Traditional migration needs to prepare the memcg charge
1802 * transaction early to prevent the old page from being
1803 * uncharged when installing migration entries. Here we can
1804 * save the potential rollback and start the charge transfer
1805 * only when migration is already known to end successfully.
1807 mem_cgroup_prepare_migration(page, new_page, &memcg);
1810 entry = mk_pmd(new_page, vma->vm_page_prot);
1811 entry = pmd_mkhuge(entry);
1812 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1815 * Clear the old entry under pagetable lock and establish the new PTE.
1816 * Any parallel GUP will either observe the old page blocking on the
1817 * page lock, block on the page table lock or observe the new page.
1818 * The SetPageUptodate on the new page and page_add_new_anon_rmap
1819 * guarantee the copy is visible before the pagetable update.
1821 flush_cache_range(vma, mmun_start, mmun_end);
1822 page_add_new_anon_rmap(new_page, vma, mmun_start);
1823 pmdp_clear_flush(vma, mmun_start, pmd);
1824 set_pmd_at(mm, mmun_start, pmd, entry);
1825 flush_tlb_range(vma, mmun_start, mmun_end);
1826 update_mmu_cache_pmd(vma, address, &entry);
1828 if (page_count(page) != 2) {
1829 set_pmd_at(mm, mmun_start, pmd, orig_entry);
1830 flush_tlb_range(vma, mmun_start, mmun_end);
1831 update_mmu_cache_pmd(vma, address, &entry);
1832 page_remove_rmap(new_page);
1836 page_remove_rmap(page);
1839 * Finish the charge transaction under the page table lock to
1840 * prevent split_huge_page() from dividing up the charge
1841 * before it's fully transferred to the new page.
1843 mem_cgroup_end_migration(memcg, page, new_page, true);
1845 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1847 unlock_page(new_page);
1849 put_page(page); /* Drop the rmap reference */
1850 put_page(page); /* Drop the LRU isolation reference */
1852 count_vm_events(PGMIGRATE_SUCCESS, HPAGE_PMD_NR);
1853 count_vm_numa_events(NUMA_PAGE_MIGRATE, HPAGE_PMD_NR);
1855 mod_zone_page_state(page_zone(page),
1856 NR_ISOLATED_ANON + page_lru,
1861 count_vm_events(PGMIGRATE_FAIL, HPAGE_PMD_NR);
1863 ptl = pmd_lock(mm, pmd);
1864 if (pmd_same(*pmd, entry)) {
1865 entry = pmd_mknonnuma(entry);
1866 set_pmd_at(mm, mmun_start, pmd, entry);
1867 update_mmu_cache_pmd(vma, address, &entry);
1876 #endif /* CONFIG_NUMA_BALANCING */
1878 #endif /* CONFIG_NUMA */