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 * Add isolated pages on the list back to the LRU under page lock
76 * to avoid leaking evictable pages back onto unevictable list.
78 void putback_lru_pages(struct list_head *l)
83 list_for_each_entry_safe(page, page2, l, lru) {
85 dec_zone_page_state(page, NR_ISOLATED_ANON +
86 page_is_file_cache(page));
87 putback_lru_page(page);
92 * Put previously isolated pages back onto the appropriate lists
93 * from where they were once taken off for compaction/migration.
95 * This function shall be used instead of putback_lru_pages(),
96 * whenever the isolated pageset has been built by isolate_migratepages_range()
98 void putback_movable_pages(struct list_head *l)
103 list_for_each_entry_safe(page, page2, l, lru) {
104 if (unlikely(PageHuge(page))) {
105 putback_active_hugepage(page);
108 list_del(&page->lru);
109 dec_zone_page_state(page, NR_ISOLATED_ANON +
110 page_is_file_cache(page));
111 if (unlikely(isolated_balloon_page(page)))
112 balloon_page_putback(page);
114 putback_lru_page(page);
119 * Restore a potential migration pte to a working pte entry
121 static int remove_migration_pte(struct page *new, struct vm_area_struct *vma,
122 unsigned long addr, void *old)
124 struct mm_struct *mm = vma->vm_mm;
130 if (unlikely(PageHuge(new))) {
131 ptep = huge_pte_offset(mm, addr);
134 ptl = huge_pte_lockptr(hstate_vma(vma), mm, ptep);
136 pmd = mm_find_pmd(mm, addr);
139 if (pmd_trans_huge(*pmd))
142 ptep = pte_offset_map(pmd, addr);
145 * Peek to check is_swap_pte() before taking ptlock? No, we
146 * can race mremap's move_ptes(), which skips anon_vma lock.
149 ptl = pte_lockptr(mm, pmd);
154 if (!is_swap_pte(pte))
157 entry = pte_to_swp_entry(pte);
159 if (!is_migration_entry(entry) ||
160 migration_entry_to_page(entry) != old)
164 pte = pte_mkold(mk_pte(new, vma->vm_page_prot));
165 if (pte_swp_soft_dirty(*ptep))
166 pte = pte_mksoft_dirty(pte);
167 if (is_write_migration_entry(entry))
168 pte = pte_mkwrite(pte);
169 #ifdef CONFIG_HUGETLB_PAGE
171 pte = pte_mkhuge(pte);
172 pte = arch_make_huge_pte(pte, vma, new, 0);
175 flush_dcache_page(new);
176 set_pte_at(mm, addr, ptep, pte);
180 hugepage_add_anon_rmap(new, vma, addr);
183 } else if (PageAnon(new))
184 page_add_anon_rmap(new, vma, addr);
186 page_add_file_rmap(new);
188 /* No need to invalidate - it was non-present before */
189 update_mmu_cache(vma, addr, ptep);
191 pte_unmap_unlock(ptep, ptl);
197 * Get rid of all migration entries and replace them by
198 * references to the indicated page.
200 static void remove_migration_ptes(struct page *old, struct page *new)
202 struct rmap_walk_control rwc = {
203 .rmap_one = remove_migration_pte,
207 rmap_walk(new, &rwc);
211 * Something used the pte of a page under migration. We need to
212 * get to the page and wait until migration is finished.
213 * When we return from this function the fault will be retried.
215 static void __migration_entry_wait(struct mm_struct *mm, pte_t *ptep,
224 if (!is_swap_pte(pte))
227 entry = pte_to_swp_entry(pte);
228 if (!is_migration_entry(entry))
231 page = migration_entry_to_page(entry);
234 * Once radix-tree replacement of page migration started, page_count
235 * *must* be zero. And, we don't want to call wait_on_page_locked()
236 * against a page without get_page().
237 * So, we use get_page_unless_zero(), here. Even failed, page fault
240 if (!get_page_unless_zero(page))
242 pte_unmap_unlock(ptep, ptl);
243 wait_on_page_locked(page);
247 pte_unmap_unlock(ptep, ptl);
250 void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
251 unsigned long address)
253 spinlock_t *ptl = pte_lockptr(mm, pmd);
254 pte_t *ptep = pte_offset_map(pmd, address);
255 __migration_entry_wait(mm, ptep, ptl);
258 void migration_entry_wait_huge(struct vm_area_struct *vma,
259 struct mm_struct *mm, pte_t *pte)
261 spinlock_t *ptl = huge_pte_lockptr(hstate_vma(vma), mm, pte);
262 __migration_entry_wait(mm, pte, ptl);
266 /* Returns true if all buffers are successfully locked */
267 static bool buffer_migrate_lock_buffers(struct buffer_head *head,
268 enum migrate_mode mode)
270 struct buffer_head *bh = head;
272 /* Simple case, sync compaction */
273 if (mode != MIGRATE_ASYNC) {
277 bh = bh->b_this_page;
279 } while (bh != head);
284 /* async case, we cannot block on lock_buffer so use trylock_buffer */
287 if (!trylock_buffer(bh)) {
289 * We failed to lock the buffer and cannot stall in
290 * async migration. Release the taken locks
292 struct buffer_head *failed_bh = bh;
295 while (bh != failed_bh) {
298 bh = bh->b_this_page;
303 bh = bh->b_this_page;
304 } while (bh != head);
308 static inline bool buffer_migrate_lock_buffers(struct buffer_head *head,
309 enum migrate_mode mode)
313 #endif /* CONFIG_BLOCK */
316 * Replace the page in the mapping.
318 * The number of remaining references must be:
319 * 1 for anonymous pages without a mapping
320 * 2 for pages with a mapping
321 * 3 for pages with a mapping and PagePrivate/PagePrivate2 set.
323 int migrate_page_move_mapping(struct address_space *mapping,
324 struct page *newpage, struct page *page,
325 struct buffer_head *head, enum migrate_mode mode,
328 int expected_count = 1 + extra_count;
332 /* Anonymous page without mapping */
333 if (page_count(page) != expected_count)
335 return MIGRATEPAGE_SUCCESS;
338 spin_lock_irq(&mapping->tree_lock);
340 pslot = radix_tree_lookup_slot(&mapping->page_tree,
343 expected_count += 1 + page_has_private(page);
344 if (page_count(page) != expected_count ||
345 radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) {
346 spin_unlock_irq(&mapping->tree_lock);
350 if (!page_freeze_refs(page, expected_count)) {
351 spin_unlock_irq(&mapping->tree_lock);
356 * In the async migration case of moving a page with buffers, lock the
357 * buffers using trylock before the mapping is moved. If the mapping
358 * was moved, we later failed to lock the buffers and could not move
359 * the mapping back due to an elevated page count, we would have to
360 * block waiting on other references to be dropped.
362 if (mode == MIGRATE_ASYNC && head &&
363 !buffer_migrate_lock_buffers(head, mode)) {
364 page_unfreeze_refs(page, expected_count);
365 spin_unlock_irq(&mapping->tree_lock);
370 * Now we know that no one else is looking at the page.
372 get_page(newpage); /* add cache reference */
373 if (PageSwapCache(page)) {
374 SetPageSwapCache(newpage);
375 set_page_private(newpage, page_private(page));
378 radix_tree_replace_slot(pslot, newpage);
381 * Drop cache reference from old page by unfreezing
382 * to one less reference.
383 * We know this isn't the last reference.
385 page_unfreeze_refs(page, expected_count - 1);
388 * If moved to a different zone then also account
389 * the page for that zone. Other VM counters will be
390 * taken care of when we establish references to the
391 * new page and drop references to the old page.
393 * Note that anonymous pages are accounted for
394 * via NR_FILE_PAGES and NR_ANON_PAGES if they
395 * are mapped to swap space.
397 __dec_zone_page_state(page, NR_FILE_PAGES);
398 __inc_zone_page_state(newpage, NR_FILE_PAGES);
399 if (!PageSwapCache(page) && PageSwapBacked(page)) {
400 __dec_zone_page_state(page, NR_SHMEM);
401 __inc_zone_page_state(newpage, NR_SHMEM);
403 spin_unlock_irq(&mapping->tree_lock);
405 return MIGRATEPAGE_SUCCESS;
409 * The expected number of remaining references is the same as that
410 * of migrate_page_move_mapping().
412 int migrate_huge_page_move_mapping(struct address_space *mapping,
413 struct page *newpage, struct page *page)
419 if (page_count(page) != 1)
421 return MIGRATEPAGE_SUCCESS;
424 spin_lock_irq(&mapping->tree_lock);
426 pslot = radix_tree_lookup_slot(&mapping->page_tree,
429 expected_count = 2 + page_has_private(page);
430 if (page_count(page) != expected_count ||
431 radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) {
432 spin_unlock_irq(&mapping->tree_lock);
436 if (!page_freeze_refs(page, expected_count)) {
437 spin_unlock_irq(&mapping->tree_lock);
443 radix_tree_replace_slot(pslot, newpage);
445 page_unfreeze_refs(page, expected_count - 1);
447 spin_unlock_irq(&mapping->tree_lock);
448 return MIGRATEPAGE_SUCCESS;
452 * Gigantic pages are so large that we do not guarantee that page++ pointer
453 * arithmetic will work across the entire page. We need something more
456 static void __copy_gigantic_page(struct page *dst, struct page *src,
460 struct page *dst_base = dst;
461 struct page *src_base = src;
463 for (i = 0; i < nr_pages; ) {
465 copy_highpage(dst, src);
468 dst = mem_map_next(dst, dst_base, i);
469 src = mem_map_next(src, src_base, i);
473 static void copy_huge_page(struct page *dst, struct page *src)
480 struct hstate *h = page_hstate(src);
481 nr_pages = pages_per_huge_page(h);
483 if (unlikely(nr_pages > MAX_ORDER_NR_PAGES)) {
484 __copy_gigantic_page(dst, src, nr_pages);
489 BUG_ON(!PageTransHuge(src));
490 nr_pages = hpage_nr_pages(src);
493 for (i = 0; i < nr_pages; i++) {
495 copy_highpage(dst + i, src + i);
500 * Copy the page to its new location
502 void migrate_page_copy(struct page *newpage, struct page *page)
506 if (PageHuge(page) || PageTransHuge(page))
507 copy_huge_page(newpage, page);
509 copy_highpage(newpage, page);
512 SetPageError(newpage);
513 if (PageReferenced(page))
514 SetPageReferenced(newpage);
515 if (PageUptodate(page))
516 SetPageUptodate(newpage);
517 if (TestClearPageActive(page)) {
518 VM_BUG_ON(PageUnevictable(page));
519 SetPageActive(newpage);
520 } else if (TestClearPageUnevictable(page))
521 SetPageUnevictable(newpage);
522 if (PageChecked(page))
523 SetPageChecked(newpage);
524 if (PageMappedToDisk(page))
525 SetPageMappedToDisk(newpage);
527 if (PageDirty(page)) {
528 clear_page_dirty_for_io(page);
530 * Want to mark the page and the radix tree as dirty, and
531 * redo the accounting that clear_page_dirty_for_io undid,
532 * but we can't use set_page_dirty because that function
533 * is actually a signal that all of the page has become dirty.
534 * Whereas only part of our page may be dirty.
536 if (PageSwapBacked(page))
537 SetPageDirty(newpage);
539 __set_page_dirty_nobuffers(newpage);
543 * Copy NUMA information to the new page, to prevent over-eager
544 * future migrations of this same page.
546 cpupid = page_cpupid_xchg_last(page, -1);
547 page_cpupid_xchg_last(newpage, cpupid);
549 mlock_migrate_page(newpage, page);
550 ksm_migrate_page(newpage, page);
552 * Please do not reorder this without considering how mm/ksm.c's
553 * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache().
555 ClearPageSwapCache(page);
556 ClearPagePrivate(page);
557 set_page_private(page, 0);
560 * If any waiters have accumulated on the new page then
563 if (PageWriteback(newpage))
564 end_page_writeback(newpage);
567 /************************************************************
568 * Migration functions
569 ***********************************************************/
571 /* Always fail migration. Used for mappings that are not movable */
572 int fail_migrate_page(struct address_space *mapping,
573 struct page *newpage, struct page *page)
577 EXPORT_SYMBOL(fail_migrate_page);
580 * Common logic to directly migrate a single page suitable for
581 * pages that do not use PagePrivate/PagePrivate2.
583 * Pages are locked upon entry and exit.
585 int migrate_page(struct address_space *mapping,
586 struct page *newpage, struct page *page,
587 enum migrate_mode mode)
591 BUG_ON(PageWriteback(page)); /* Writeback must be complete */
593 rc = migrate_page_move_mapping(mapping, newpage, page, NULL, mode, 0);
595 if (rc != MIGRATEPAGE_SUCCESS)
598 migrate_page_copy(newpage, page);
599 return MIGRATEPAGE_SUCCESS;
601 EXPORT_SYMBOL(migrate_page);
605 * Migration function for pages with buffers. This function can only be used
606 * if the underlying filesystem guarantees that no other references to "page"
609 int buffer_migrate_page(struct address_space *mapping,
610 struct page *newpage, struct page *page, enum migrate_mode mode)
612 struct buffer_head *bh, *head;
615 if (!page_has_buffers(page))
616 return migrate_page(mapping, newpage, page, mode);
618 head = page_buffers(page);
620 rc = migrate_page_move_mapping(mapping, newpage, page, head, mode, 0);
622 if (rc != MIGRATEPAGE_SUCCESS)
626 * In the async case, migrate_page_move_mapping locked the buffers
627 * with an IRQ-safe spinlock held. In the sync case, the buffers
628 * need to be locked now
630 if (mode != MIGRATE_ASYNC)
631 BUG_ON(!buffer_migrate_lock_buffers(head, mode));
633 ClearPagePrivate(page);
634 set_page_private(newpage, page_private(page));
635 set_page_private(page, 0);
641 set_bh_page(bh, newpage, bh_offset(bh));
642 bh = bh->b_this_page;
644 } while (bh != head);
646 SetPagePrivate(newpage);
648 migrate_page_copy(newpage, page);
654 bh = bh->b_this_page;
656 } while (bh != head);
658 return MIGRATEPAGE_SUCCESS;
660 EXPORT_SYMBOL(buffer_migrate_page);
664 * Writeback a page to clean the dirty state
666 static int writeout(struct address_space *mapping, struct page *page)
668 struct writeback_control wbc = {
669 .sync_mode = WB_SYNC_NONE,
672 .range_end = LLONG_MAX,
677 if (!mapping->a_ops->writepage)
678 /* No write method for the address space */
681 if (!clear_page_dirty_for_io(page))
682 /* Someone else already triggered a write */
686 * A dirty page may imply that the underlying filesystem has
687 * the page on some queue. So the page must be clean for
688 * migration. Writeout may mean we loose the lock and the
689 * page state is no longer what we checked for earlier.
690 * At this point we know that the migration attempt cannot
693 remove_migration_ptes(page, page);
695 rc = mapping->a_ops->writepage(page, &wbc);
697 if (rc != AOP_WRITEPAGE_ACTIVATE)
698 /* unlocked. Relock */
701 return (rc < 0) ? -EIO : -EAGAIN;
705 * Default handling if a filesystem does not provide a migration function.
707 static int fallback_migrate_page(struct address_space *mapping,
708 struct page *newpage, struct page *page, enum migrate_mode mode)
710 if (PageDirty(page)) {
711 /* Only writeback pages in full synchronous migration */
712 if (mode != MIGRATE_SYNC)
714 return writeout(mapping, page);
718 * Buffers may be managed in a filesystem specific way.
719 * We must have no buffers or drop them.
721 if (page_has_private(page) &&
722 !try_to_release_page(page, GFP_KERNEL))
725 return migrate_page(mapping, newpage, page, mode);
729 * Move a page to a newly allocated page
730 * The page is locked and all ptes have been successfully removed.
732 * The new page will have replaced the old page if this function
737 * MIGRATEPAGE_SUCCESS - success
739 static int move_to_new_page(struct page *newpage, struct page *page,
740 int remap_swapcache, enum migrate_mode mode)
742 struct address_space *mapping;
746 * Block others from accessing the page when we get around to
747 * establishing additional references. We are the only one
748 * holding a reference to the new page at this point.
750 if (!trylock_page(newpage))
753 /* Prepare mapping for the new page.*/
754 newpage->index = page->index;
755 newpage->mapping = page->mapping;
756 if (PageSwapBacked(page))
757 SetPageSwapBacked(newpage);
759 mapping = page_mapping(page);
761 rc = migrate_page(mapping, newpage, page, mode);
762 else if (mapping->a_ops->migratepage)
764 * Most pages have a mapping and most filesystems provide a
765 * migratepage callback. Anonymous pages are part of swap
766 * space which also has its own migratepage callback. This
767 * is the most common path for page migration.
769 rc = mapping->a_ops->migratepage(mapping,
770 newpage, page, mode);
772 rc = fallback_migrate_page(mapping, newpage, page, mode);
774 if (rc != MIGRATEPAGE_SUCCESS) {
775 newpage->mapping = NULL;
778 remove_migration_ptes(page, newpage);
779 page->mapping = NULL;
782 unlock_page(newpage);
787 static int __unmap_and_move(struct page *page, struct page *newpage,
788 int force, enum migrate_mode mode)
791 int remap_swapcache = 1;
792 struct mem_cgroup *mem;
793 struct anon_vma *anon_vma = NULL;
795 if (!trylock_page(page)) {
796 if (!force || mode == MIGRATE_ASYNC)
800 * It's not safe for direct compaction to call lock_page.
801 * For example, during page readahead pages are added locked
802 * to the LRU. Later, when the IO completes the pages are
803 * marked uptodate and unlocked. However, the queueing
804 * could be merging multiple pages for one bio (e.g.
805 * mpage_readpages). If an allocation happens for the
806 * second or third page, the process can end up locking
807 * the same page twice and deadlocking. Rather than
808 * trying to be clever about what pages can be locked,
809 * avoid the use of lock_page for direct compaction
812 if (current->flags & PF_MEMALLOC)
818 /* charge against new page */
819 mem_cgroup_prepare_migration(page, newpage, &mem);
821 if (PageWriteback(page)) {
823 * Only in the case of a full synchronous migration is it
824 * necessary to wait for PageWriteback. In the async case,
825 * the retry loop is too short and in the sync-light case,
826 * the overhead of stalling is too much
828 if (mode != MIGRATE_SYNC) {
834 wait_on_page_writeback(page);
837 * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
838 * we cannot notice that anon_vma is freed while we migrates a page.
839 * This get_anon_vma() delays freeing anon_vma pointer until the end
840 * of migration. File cache pages are no problem because of page_lock()
841 * File Caches may use write_page() or lock_page() in migration, then,
842 * just care Anon page here.
844 if (PageAnon(page) && !PageKsm(page)) {
846 * Only page_lock_anon_vma_read() understands the subtleties of
847 * getting a hold on an anon_vma from outside one of its mms.
849 anon_vma = page_get_anon_vma(page);
854 } else if (PageSwapCache(page)) {
856 * We cannot be sure that the anon_vma of an unmapped
857 * swapcache page is safe to use because we don't
858 * know in advance if the VMA that this page belonged
859 * to still exists. If the VMA and others sharing the
860 * data have been freed, then the anon_vma could
861 * already be invalid.
863 * To avoid this possibility, swapcache pages get
864 * migrated but are not remapped when migration
873 if (unlikely(balloon_page_movable(page))) {
875 * A ballooned page does not need any special attention from
876 * physical to virtual reverse mapping procedures.
877 * Skip any attempt to unmap PTEs or to remap swap cache,
878 * in order to avoid burning cycles at rmap level, and perform
879 * the page migration right away (proteced by page lock).
881 rc = balloon_page_migrate(newpage, page, mode);
886 * Corner case handling:
887 * 1. When a new swap-cache page is read into, it is added to the LRU
888 * and treated as swapcache but it has no rmap yet.
889 * Calling try_to_unmap() against a page->mapping==NULL page will
890 * trigger a BUG. So handle it here.
891 * 2. An orphaned page (see truncate_complete_page) might have
892 * fs-private metadata. The page can be picked up due to memory
893 * offlining. Everywhere else except page reclaim, the page is
894 * invisible to the vm, so the page can not be migrated. So try to
895 * free the metadata, so the page can be freed.
897 if (!page->mapping) {
898 VM_BUG_ON(PageAnon(page));
899 if (page_has_private(page)) {
900 try_to_free_buffers(page);
906 /* Establish migration ptes or remove ptes */
907 try_to_unmap(page, TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
910 if (!page_mapped(page))
911 rc = move_to_new_page(newpage, page, remap_swapcache, mode);
913 if (rc && remap_swapcache)
914 remove_migration_ptes(page, page);
916 /* Drop an anon_vma reference if we took one */
918 put_anon_vma(anon_vma);
921 mem_cgroup_end_migration(mem, page, newpage,
922 (rc == MIGRATEPAGE_SUCCESS ||
923 rc == MIGRATEPAGE_BALLOON_SUCCESS));
930 * Obtain the lock on page, remove all ptes and migrate the page
931 * to the newly allocated page in newpage.
933 static int unmap_and_move(new_page_t get_new_page, unsigned long private,
934 struct page *page, int force, enum migrate_mode mode)
938 struct page *newpage = get_new_page(page, private, &result);
943 if (page_count(page) == 1) {
944 /* page was freed from under us. So we are done. */
948 if (unlikely(PageTransHuge(page)))
949 if (unlikely(split_huge_page(page)))
952 rc = __unmap_and_move(page, newpage, force, mode);
954 if (unlikely(rc == MIGRATEPAGE_BALLOON_SUCCESS)) {
956 * A ballooned page has been migrated already.
957 * Now, it's the time to wrap-up counters,
958 * handle the page back to Buddy and return.
960 dec_zone_page_state(page, NR_ISOLATED_ANON +
961 page_is_file_cache(page));
962 balloon_page_free(page);
963 return MIGRATEPAGE_SUCCESS;
968 * A page that has been migrated has all references
969 * removed and will be freed. A page that has not been
970 * migrated will have kepts its references and be
973 list_del(&page->lru);
974 dec_zone_page_state(page, NR_ISOLATED_ANON +
975 page_is_file_cache(page));
976 putback_lru_page(page);
979 * Move the new page to the LRU. If migration was not successful
980 * then this will free the page.
982 putback_lru_page(newpage);
987 *result = page_to_nid(newpage);
993 * Counterpart of unmap_and_move_page() for hugepage migration.
995 * This function doesn't wait the completion of hugepage I/O
996 * because there is no race between I/O and migration for hugepage.
997 * Note that currently hugepage I/O occurs only in direct I/O
998 * where no lock is held and PG_writeback is irrelevant,
999 * and writeback status of all subpages are counted in the reference
1000 * count of the head page (i.e. if all subpages of a 2MB hugepage are
1001 * under direct I/O, the reference of the head page is 512 and a bit more.)
1002 * This means that when we try to migrate hugepage whose subpages are
1003 * doing direct I/O, some references remain after try_to_unmap() and
1004 * hugepage migration fails without data corruption.
1006 * There is also no race when direct I/O is issued on the page under migration,
1007 * because then pte is replaced with migration swap entry and direct I/O code
1008 * will wait in the page fault for migration to complete.
1010 static int unmap_and_move_huge_page(new_page_t get_new_page,
1011 unsigned long private, struct page *hpage,
1012 int force, enum migrate_mode mode)
1016 struct page *new_hpage = get_new_page(hpage, private, &result);
1017 struct anon_vma *anon_vma = NULL;
1020 * Movability of hugepages depends on architectures and hugepage size.
1021 * This check is necessary because some callers of hugepage migration
1022 * like soft offline and memory hotremove don't walk through page
1023 * tables or check whether the hugepage is pmd-based or not before
1024 * kicking migration.
1026 if (!hugepage_migration_support(page_hstate(hpage)))
1034 if (!trylock_page(hpage)) {
1035 if (!force || mode != MIGRATE_SYNC)
1040 if (PageAnon(hpage))
1041 anon_vma = page_get_anon_vma(hpage);
1043 try_to_unmap(hpage, TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
1045 if (!page_mapped(hpage))
1046 rc = move_to_new_page(new_hpage, hpage, 1, mode);
1049 remove_migration_ptes(hpage, hpage);
1052 put_anon_vma(anon_vma);
1055 hugetlb_cgroup_migrate(hpage, new_hpage);
1060 putback_active_hugepage(hpage);
1061 put_page(new_hpage);
1066 *result = page_to_nid(new_hpage);
1072 * migrate_pages - migrate the pages specified in a list, to the free pages
1073 * supplied as the target for the page migration
1075 * @from: The list of pages to be migrated.
1076 * @get_new_page: The function used to allocate free pages to be used
1077 * as the target of the page migration.
1078 * @private: Private data to be passed on to get_new_page()
1079 * @mode: The migration mode that specifies the constraints for
1080 * page migration, if any.
1081 * @reason: The reason for page migration.
1083 * The function returns after 10 attempts or if no pages are movable any more
1084 * because the list has become empty or no retryable pages exist any more.
1085 * The caller should call putback_lru_pages() to return pages to the LRU
1086 * or free list only if ret != 0.
1088 * Returns the number of pages that were not migrated, or an error code.
1090 int migrate_pages(struct list_head *from, new_page_t get_new_page,
1091 unsigned long private, enum migrate_mode mode, int reason)
1095 int nr_succeeded = 0;
1099 int swapwrite = current->flags & PF_SWAPWRITE;
1103 current->flags |= PF_SWAPWRITE;
1105 for(pass = 0; pass < 10 && retry; pass++) {
1108 list_for_each_entry_safe(page, page2, from, lru) {
1112 rc = unmap_and_move_huge_page(get_new_page,
1113 private, page, pass > 2, mode);
1115 rc = unmap_and_move(get_new_page, private,
1116 page, pass > 2, mode);
1124 case MIGRATEPAGE_SUCCESS:
1128 /* Permanent failure */
1134 rc = nr_failed + retry;
1137 count_vm_events(PGMIGRATE_SUCCESS, nr_succeeded);
1139 count_vm_events(PGMIGRATE_FAIL, nr_failed);
1140 trace_mm_migrate_pages(nr_succeeded, nr_failed, mode, reason);
1143 current->flags &= ~PF_SWAPWRITE;
1150 * Move a list of individual pages
1152 struct page_to_node {
1159 static struct page *new_page_node(struct page *p, unsigned long private,
1162 struct page_to_node *pm = (struct page_to_node *)private;
1164 while (pm->node != MAX_NUMNODES && pm->page != p)
1167 if (pm->node == MAX_NUMNODES)
1170 *result = &pm->status;
1173 return alloc_huge_page_node(page_hstate(compound_head(p)),
1176 return alloc_pages_exact_node(pm->node,
1177 GFP_HIGHUSER_MOVABLE | GFP_THISNODE, 0);
1181 * Move a set of pages as indicated in the pm array. The addr
1182 * field must be set to the virtual address of the page to be moved
1183 * and the node number must contain a valid target node.
1184 * The pm array ends with node = MAX_NUMNODES.
1186 static int do_move_page_to_node_array(struct mm_struct *mm,
1187 struct page_to_node *pm,
1191 struct page_to_node *pp;
1192 LIST_HEAD(pagelist);
1194 down_read(&mm->mmap_sem);
1197 * Build a list of pages to migrate
1199 for (pp = pm; pp->node != MAX_NUMNODES; pp++) {
1200 struct vm_area_struct *vma;
1204 vma = find_vma(mm, pp->addr);
1205 if (!vma || pp->addr < vma->vm_start || !vma_migratable(vma))
1208 page = follow_page(vma, pp->addr, FOLL_GET|FOLL_SPLIT);
1210 err = PTR_ERR(page);
1218 /* Use PageReserved to check for zero page */
1219 if (PageReserved(page))
1223 err = page_to_nid(page);
1225 if (err == pp->node)
1227 * Node already in the right place
1232 if (page_mapcount(page) > 1 &&
1236 if (PageHuge(page)) {
1237 isolate_huge_page(page, &pagelist);
1241 err = isolate_lru_page(page);
1243 list_add_tail(&page->lru, &pagelist);
1244 inc_zone_page_state(page, NR_ISOLATED_ANON +
1245 page_is_file_cache(page));
1249 * Either remove the duplicate refcount from
1250 * isolate_lru_page() or drop the page ref if it was
1259 if (!list_empty(&pagelist)) {
1260 err = migrate_pages(&pagelist, new_page_node,
1261 (unsigned long)pm, MIGRATE_SYNC, MR_SYSCALL);
1263 putback_movable_pages(&pagelist);
1266 up_read(&mm->mmap_sem);
1271 * Migrate an array of page address onto an array of nodes and fill
1272 * the corresponding array of status.
1274 static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes,
1275 unsigned long nr_pages,
1276 const void __user * __user *pages,
1277 const int __user *nodes,
1278 int __user *status, int flags)
1280 struct page_to_node *pm;
1281 unsigned long chunk_nr_pages;
1282 unsigned long chunk_start;
1286 pm = (struct page_to_node *)__get_free_page(GFP_KERNEL);
1293 * Store a chunk of page_to_node array in a page,
1294 * but keep the last one as a marker
1296 chunk_nr_pages = (PAGE_SIZE / sizeof(struct page_to_node)) - 1;
1298 for (chunk_start = 0;
1299 chunk_start < nr_pages;
1300 chunk_start += chunk_nr_pages) {
1303 if (chunk_start + chunk_nr_pages > nr_pages)
1304 chunk_nr_pages = nr_pages - chunk_start;
1306 /* fill the chunk pm with addrs and nodes from user-space */
1307 for (j = 0; j < chunk_nr_pages; j++) {
1308 const void __user *p;
1312 if (get_user(p, pages + j + chunk_start))
1314 pm[j].addr = (unsigned long) p;
1316 if (get_user(node, nodes + j + chunk_start))
1320 if (node < 0 || node >= MAX_NUMNODES)
1323 if (!node_state(node, N_MEMORY))
1327 if (!node_isset(node, task_nodes))
1333 /* End marker for this chunk */
1334 pm[chunk_nr_pages].node = MAX_NUMNODES;
1336 /* Migrate this chunk */
1337 err = do_move_page_to_node_array(mm, pm,
1338 flags & MPOL_MF_MOVE_ALL);
1342 /* Return status information */
1343 for (j = 0; j < chunk_nr_pages; j++)
1344 if (put_user(pm[j].status, status + j + chunk_start)) {
1352 free_page((unsigned long)pm);
1358 * Determine the nodes of an array of pages and store it in an array of status.
1360 static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages,
1361 const void __user **pages, int *status)
1365 down_read(&mm->mmap_sem);
1367 for (i = 0; i < nr_pages; i++) {
1368 unsigned long addr = (unsigned long)(*pages);
1369 struct vm_area_struct *vma;
1373 vma = find_vma(mm, addr);
1374 if (!vma || addr < vma->vm_start)
1377 page = follow_page(vma, addr, 0);
1379 err = PTR_ERR(page);
1384 /* Use PageReserved to check for zero page */
1385 if (!page || PageReserved(page))
1388 err = page_to_nid(page);
1396 up_read(&mm->mmap_sem);
1400 * Determine the nodes of a user array of pages and store it in
1401 * a user array of status.
1403 static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages,
1404 const void __user * __user *pages,
1407 #define DO_PAGES_STAT_CHUNK_NR 16
1408 const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR];
1409 int chunk_status[DO_PAGES_STAT_CHUNK_NR];
1412 unsigned long chunk_nr;
1414 chunk_nr = nr_pages;
1415 if (chunk_nr > DO_PAGES_STAT_CHUNK_NR)
1416 chunk_nr = DO_PAGES_STAT_CHUNK_NR;
1418 if (copy_from_user(chunk_pages, pages, chunk_nr * sizeof(*chunk_pages)))
1421 do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status);
1423 if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status)))
1428 nr_pages -= chunk_nr;
1430 return nr_pages ? -EFAULT : 0;
1434 * Move a list of pages in the address space of the currently executing
1437 SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages,
1438 const void __user * __user *, pages,
1439 const int __user *, nodes,
1440 int __user *, status, int, flags)
1442 const struct cred *cred = current_cred(), *tcred;
1443 struct task_struct *task;
1444 struct mm_struct *mm;
1446 nodemask_t task_nodes;
1449 if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
1452 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1455 /* Find the mm_struct */
1457 task = pid ? find_task_by_vpid(pid) : current;
1462 get_task_struct(task);
1465 * Check if this process has the right to modify the specified
1466 * process. The right exists if the process has administrative
1467 * capabilities, superuser privileges or the same
1468 * userid as the target process.
1470 tcred = __task_cred(task);
1471 if (!uid_eq(cred->euid, tcred->suid) && !uid_eq(cred->euid, tcred->uid) &&
1472 !uid_eq(cred->uid, tcred->suid) && !uid_eq(cred->uid, tcred->uid) &&
1473 !capable(CAP_SYS_NICE)) {
1480 err = security_task_movememory(task);
1484 task_nodes = cpuset_mems_allowed(task);
1485 mm = get_task_mm(task);
1486 put_task_struct(task);
1492 err = do_pages_move(mm, task_nodes, nr_pages, pages,
1493 nodes, status, flags);
1495 err = do_pages_stat(mm, nr_pages, pages, status);
1501 put_task_struct(task);
1506 * Call migration functions in the vma_ops that may prepare
1507 * memory in a vm for migration. migration functions may perform
1508 * the migration for vmas that do not have an underlying page struct.
1510 int migrate_vmas(struct mm_struct *mm, const nodemask_t *to,
1511 const nodemask_t *from, unsigned long flags)
1513 struct vm_area_struct *vma;
1516 for (vma = mm->mmap; vma && !err; vma = vma->vm_next) {
1517 if (vma->vm_ops && vma->vm_ops->migrate) {
1518 err = vma->vm_ops->migrate(vma, to, from, flags);
1526 #ifdef CONFIG_NUMA_BALANCING
1528 * Returns true if this is a safe migration target node for misplaced NUMA
1529 * pages. Currently it only checks the watermarks which crude
1531 static bool migrate_balanced_pgdat(struct pglist_data *pgdat,
1532 unsigned long nr_migrate_pages)
1535 for (z = pgdat->nr_zones - 1; z >= 0; z--) {
1536 struct zone *zone = pgdat->node_zones + z;
1538 if (!populated_zone(zone))
1541 if (!zone_reclaimable(zone))
1544 /* Avoid waking kswapd by allocating pages_to_migrate pages. */
1545 if (!zone_watermark_ok(zone, 0,
1546 high_wmark_pages(zone) +
1555 static struct page *alloc_misplaced_dst_page(struct page *page,
1559 int nid = (int) data;
1560 struct page *newpage;
1562 newpage = alloc_pages_exact_node(nid,
1563 (GFP_HIGHUSER_MOVABLE | GFP_THISNODE |
1564 __GFP_NOMEMALLOC | __GFP_NORETRY |
1568 page_cpupid_xchg_last(newpage, page_cpupid_last(page));
1574 * page migration rate limiting control.
1575 * Do not migrate more than @pages_to_migrate in a @migrate_interval_millisecs
1576 * window of time. Default here says do not migrate more than 1280M per second.
1577 * If a node is rate-limited then PTE NUMA updates are also rate-limited. However
1578 * as it is faults that reset the window, pte updates will happen unconditionally
1579 * if there has not been a fault since @pteupdate_interval_millisecs after the
1580 * throttle window closed.
1582 static unsigned int migrate_interval_millisecs __read_mostly = 100;
1583 static unsigned int pteupdate_interval_millisecs __read_mostly = 1000;
1584 static unsigned int ratelimit_pages __read_mostly = 128 << (20 - PAGE_SHIFT);
1586 /* Returns true if NUMA migration is currently rate limited */
1587 bool migrate_ratelimited(int node)
1589 pg_data_t *pgdat = NODE_DATA(node);
1591 if (time_after(jiffies, pgdat->numabalancing_migrate_next_window +
1592 msecs_to_jiffies(pteupdate_interval_millisecs)))
1595 if (pgdat->numabalancing_migrate_nr_pages < ratelimit_pages)
1601 /* Returns true if the node is migrate rate-limited after the update */
1602 static bool numamigrate_update_ratelimit(pg_data_t *pgdat,
1603 unsigned long nr_pages)
1606 * Rate-limit the amount of data that is being migrated to a node.
1607 * Optimal placement is no good if the memory bus is saturated and
1608 * all the time is being spent migrating!
1610 if (time_after(jiffies, pgdat->numabalancing_migrate_next_window)) {
1611 spin_lock(&pgdat->numabalancing_migrate_lock);
1612 pgdat->numabalancing_migrate_nr_pages = 0;
1613 pgdat->numabalancing_migrate_next_window = jiffies +
1614 msecs_to_jiffies(migrate_interval_millisecs);
1615 spin_unlock(&pgdat->numabalancing_migrate_lock);
1617 if (pgdat->numabalancing_migrate_nr_pages > ratelimit_pages) {
1618 trace_mm_numa_migrate_ratelimit(current, pgdat->node_id,
1624 * This is an unlocked non-atomic update so errors are possible.
1625 * The consequences are failing to migrate when we potentiall should
1626 * have which is not severe enough to warrant locking. If it is ever
1627 * a problem, it can be converted to a per-cpu counter.
1629 pgdat->numabalancing_migrate_nr_pages += nr_pages;
1633 static int numamigrate_isolate_page(pg_data_t *pgdat, struct page *page)
1637 VM_BUG_ON(compound_order(page) && !PageTransHuge(page));
1639 /* Avoid migrating to a node that is nearly full */
1640 if (!migrate_balanced_pgdat(pgdat, 1UL << compound_order(page)))
1643 if (isolate_lru_page(page))
1647 * migrate_misplaced_transhuge_page() skips page migration's usual
1648 * check on page_count(), so we must do it here, now that the page
1649 * has been isolated: a GUP pin, or any other pin, prevents migration.
1650 * The expected page count is 3: 1 for page's mapcount and 1 for the
1651 * caller's pin and 1 for the reference taken by isolate_lru_page().
1653 if (PageTransHuge(page) && page_count(page) != 3) {
1654 putback_lru_page(page);
1658 page_lru = page_is_file_cache(page);
1659 mod_zone_page_state(page_zone(page), NR_ISOLATED_ANON + page_lru,
1660 hpage_nr_pages(page));
1663 * Isolating the page has taken another reference, so the
1664 * caller's reference can be safely dropped without the page
1665 * disappearing underneath us during migration.
1671 bool pmd_trans_migrating(pmd_t pmd)
1673 struct page *page = pmd_page(pmd);
1674 return PageLocked(page);
1677 void wait_migrate_huge_page(struct anon_vma *anon_vma, pmd_t *pmd)
1679 struct page *page = pmd_page(*pmd);
1680 wait_on_page_locked(page);
1684 * Attempt to migrate a misplaced page to the specified destination
1685 * node. Caller is expected to have an elevated reference count on
1686 * the page that will be dropped by this function before returning.
1688 int migrate_misplaced_page(struct page *page, struct vm_area_struct *vma,
1691 pg_data_t *pgdat = NODE_DATA(node);
1694 LIST_HEAD(migratepages);
1697 * Don't migrate file pages that are mapped in multiple processes
1698 * with execute permissions as they are probably shared libraries.
1700 if (page_mapcount(page) != 1 && page_is_file_cache(page) &&
1701 (vma->vm_flags & VM_EXEC))
1705 * Rate-limit the amount of data that is being migrated to a node.
1706 * Optimal placement is no good if the memory bus is saturated and
1707 * all the time is being spent migrating!
1709 if (numamigrate_update_ratelimit(pgdat, 1))
1712 isolated = numamigrate_isolate_page(pgdat, page);
1716 list_add(&page->lru, &migratepages);
1717 nr_remaining = migrate_pages(&migratepages, alloc_misplaced_dst_page,
1718 node, MIGRATE_ASYNC, MR_NUMA_MISPLACED);
1720 putback_lru_pages(&migratepages);
1723 count_vm_numa_event(NUMA_PAGE_MIGRATE);
1724 BUG_ON(!list_empty(&migratepages));
1731 #endif /* CONFIG_NUMA_BALANCING */
1733 #if defined(CONFIG_NUMA_BALANCING) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
1735 * Migrates a THP to a given target node. page must be locked and is unlocked
1738 int migrate_misplaced_transhuge_page(struct mm_struct *mm,
1739 struct vm_area_struct *vma,
1740 pmd_t *pmd, pmd_t entry,
1741 unsigned long address,
1742 struct page *page, int node)
1745 pg_data_t *pgdat = NODE_DATA(node);
1747 struct page *new_page = NULL;
1748 struct mem_cgroup *memcg = NULL;
1749 int page_lru = page_is_file_cache(page);
1750 unsigned long mmun_start = address & HPAGE_PMD_MASK;
1751 unsigned long mmun_end = mmun_start + HPAGE_PMD_SIZE;
1755 * Rate-limit the amount of data that is being migrated to a node.
1756 * Optimal placement is no good if the memory bus is saturated and
1757 * all the time is being spent migrating!
1759 if (numamigrate_update_ratelimit(pgdat, HPAGE_PMD_NR))
1762 new_page = alloc_pages_node(node,
1763 (GFP_TRANSHUGE | GFP_THISNODE) & ~__GFP_WAIT, HPAGE_PMD_ORDER);
1767 page_cpupid_xchg_last(new_page, page_cpupid_last(page));
1769 isolated = numamigrate_isolate_page(pgdat, page);
1775 if (mm_tlb_flush_pending(mm))
1776 flush_tlb_range(vma, mmun_start, mmun_end);
1778 /* Prepare a page as a migration target */
1779 __set_page_locked(new_page);
1780 SetPageSwapBacked(new_page);
1782 /* anon mapping, we can simply copy page->mapping to the new page: */
1783 new_page->mapping = page->mapping;
1784 new_page->index = page->index;
1785 migrate_page_copy(new_page, page);
1786 WARN_ON(PageLRU(new_page));
1788 /* Recheck the target PMD */
1789 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1790 ptl = pmd_lock(mm, pmd);
1791 if (unlikely(!pmd_same(*pmd, entry) || page_count(page) != 2)) {
1794 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1796 /* Reverse changes made by migrate_page_copy() */
1797 if (TestClearPageActive(new_page))
1798 SetPageActive(page);
1799 if (TestClearPageUnevictable(new_page))
1800 SetPageUnevictable(page);
1801 mlock_migrate_page(page, new_page);
1803 unlock_page(new_page);
1804 put_page(new_page); /* Free it */
1806 /* Retake the callers reference and putback on LRU */
1808 putback_lru_page(page);
1809 mod_zone_page_state(page_zone(page),
1810 NR_ISOLATED_ANON + page_lru, -HPAGE_PMD_NR);
1816 * Traditional migration needs to prepare the memcg charge
1817 * transaction early to prevent the old page from being
1818 * uncharged when installing migration entries. Here we can
1819 * save the potential rollback and start the charge transfer
1820 * only when migration is already known to end successfully.
1822 mem_cgroup_prepare_migration(page, new_page, &memcg);
1825 entry = mk_pmd(new_page, vma->vm_page_prot);
1826 entry = pmd_mkhuge(entry);
1827 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1830 * Clear the old entry under pagetable lock and establish the new PTE.
1831 * Any parallel GUP will either observe the old page blocking on the
1832 * page lock, block on the page table lock or observe the new page.
1833 * The SetPageUptodate on the new page and page_add_new_anon_rmap
1834 * guarantee the copy is visible before the pagetable update.
1836 flush_cache_range(vma, mmun_start, mmun_end);
1837 page_add_new_anon_rmap(new_page, vma, mmun_start);
1838 pmdp_clear_flush(vma, mmun_start, pmd);
1839 set_pmd_at(mm, mmun_start, pmd, entry);
1840 flush_tlb_range(vma, mmun_start, mmun_end);
1841 update_mmu_cache_pmd(vma, address, &entry);
1843 if (page_count(page) != 2) {
1844 set_pmd_at(mm, mmun_start, pmd, orig_entry);
1845 flush_tlb_range(vma, mmun_start, mmun_end);
1846 update_mmu_cache_pmd(vma, address, &entry);
1847 page_remove_rmap(new_page);
1851 page_remove_rmap(page);
1854 * Finish the charge transaction under the page table lock to
1855 * prevent split_huge_page() from dividing up the charge
1856 * before it's fully transferred to the new page.
1858 mem_cgroup_end_migration(memcg, page, new_page, true);
1860 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1862 unlock_page(new_page);
1864 put_page(page); /* Drop the rmap reference */
1865 put_page(page); /* Drop the LRU isolation reference */
1867 count_vm_events(PGMIGRATE_SUCCESS, HPAGE_PMD_NR);
1868 count_vm_numa_events(NUMA_PAGE_MIGRATE, HPAGE_PMD_NR);
1870 mod_zone_page_state(page_zone(page),
1871 NR_ISOLATED_ANON + page_lru,
1876 count_vm_events(PGMIGRATE_FAIL, HPAGE_PMD_NR);
1878 ptl = pmd_lock(mm, pmd);
1879 if (pmd_same(*pmd, entry)) {
1880 entry = pmd_mknonnuma(entry);
1881 set_pmd_at(mm, mmun_start, pmd, entry);
1882 update_mmu_cache_pmd(vma, address, &entry);
1891 #endif /* CONFIG_NUMA_BALANCING */
1893 #endif /* CONFIG_NUMA */