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/syscalls.h>
34 #include <linux/hugetlb.h>
35 #include <linux/hugetlb_cgroup.h>
36 #include <linux/gfp.h>
37 #include <linux/balloon_compaction.h>
38 #include <linux/mmu_notifier.h>
39 #include <linux/page_idle.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);
124 ptep = pte_offset_map(pmd, addr);
127 * Peek to check is_swap_pte() before taking ptlock? No, we
128 * can race mremap's move_ptes(), which skips anon_vma lock.
131 ptl = pte_lockptr(mm, pmd);
136 if (!is_swap_pte(pte))
139 entry = pte_to_swp_entry(pte);
141 if (!is_migration_entry(entry) ||
142 migration_entry_to_page(entry) != old)
146 pte = pte_mkold(mk_pte(new, vma->vm_page_prot));
147 if (pte_swp_soft_dirty(*ptep))
148 pte = pte_mksoft_dirty(pte);
150 /* Recheck VMA as permissions can change since migration started */
151 if (is_write_migration_entry(entry))
152 pte = maybe_mkwrite(pte, vma);
154 #ifdef CONFIG_HUGETLB_PAGE
156 pte = pte_mkhuge(pte);
157 pte = arch_make_huge_pte(pte, vma, new, 0);
160 flush_dcache_page(new);
161 set_pte_at(mm, addr, ptep, pte);
165 hugepage_add_anon_rmap(new, vma, addr);
168 } else if (PageAnon(new))
169 page_add_anon_rmap(new, vma, addr);
171 page_add_file_rmap(new);
173 if (vma->vm_flags & VM_LOCKED)
176 /* No need to invalidate - it was non-present before */
177 update_mmu_cache(vma, addr, ptep);
179 pte_unmap_unlock(ptep, ptl);
185 * Get rid of all migration entries and replace them by
186 * references to the indicated page.
188 static void remove_migration_ptes(struct page *old, struct page *new)
190 struct rmap_walk_control rwc = {
191 .rmap_one = remove_migration_pte,
195 rmap_walk(new, &rwc);
199 * Something used the pte of a page under migration. We need to
200 * get to the page and wait until migration is finished.
201 * When we return from this function the fault will be retried.
203 void __migration_entry_wait(struct mm_struct *mm, pte_t *ptep,
212 if (!is_swap_pte(pte))
215 entry = pte_to_swp_entry(pte);
216 if (!is_migration_entry(entry))
219 page = migration_entry_to_page(entry);
222 * Once radix-tree replacement of page migration started, page_count
223 * *must* be zero. And, we don't want to call wait_on_page_locked()
224 * against a page without get_page().
225 * So, we use get_page_unless_zero(), here. Even failed, page fault
228 if (!get_page_unless_zero(page))
230 pte_unmap_unlock(ptep, ptl);
231 wait_on_page_locked(page);
235 pte_unmap_unlock(ptep, ptl);
238 void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
239 unsigned long address)
241 spinlock_t *ptl = pte_lockptr(mm, pmd);
242 pte_t *ptep = pte_offset_map(pmd, address);
243 __migration_entry_wait(mm, ptep, ptl);
246 void migration_entry_wait_huge(struct vm_area_struct *vma,
247 struct mm_struct *mm, pte_t *pte)
249 spinlock_t *ptl = huge_pte_lockptr(hstate_vma(vma), mm, pte);
250 __migration_entry_wait(mm, pte, ptl);
254 /* Returns true if all buffers are successfully locked */
255 static bool buffer_migrate_lock_buffers(struct buffer_head *head,
256 enum migrate_mode mode)
258 struct buffer_head *bh = head;
260 /* Simple case, sync compaction */
261 if (mode != MIGRATE_ASYNC) {
265 bh = bh->b_this_page;
267 } while (bh != head);
272 /* async case, we cannot block on lock_buffer so use trylock_buffer */
275 if (!trylock_buffer(bh)) {
277 * We failed to lock the buffer and cannot stall in
278 * async migration. Release the taken locks
280 struct buffer_head *failed_bh = bh;
283 while (bh != failed_bh) {
286 bh = bh->b_this_page;
291 bh = bh->b_this_page;
292 } while (bh != head);
296 static inline bool buffer_migrate_lock_buffers(struct buffer_head *head,
297 enum migrate_mode mode)
301 #endif /* CONFIG_BLOCK */
304 * Replace the page in the mapping.
306 * The number of remaining references must be:
307 * 1 for anonymous pages without a mapping
308 * 2 for pages with a mapping
309 * 3 for pages with a mapping and PagePrivate/PagePrivate2 set.
311 int migrate_page_move_mapping(struct address_space *mapping,
312 struct page *newpage, struct page *page,
313 struct buffer_head *head, enum migrate_mode mode,
316 int expected_count = 1 + extra_count;
320 /* Anonymous page without mapping */
321 if (page_count(page) != expected_count)
324 /* No turning back from here */
325 set_page_memcg(newpage, page_memcg(page));
326 newpage->index = page->index;
327 newpage->mapping = page->mapping;
328 if (PageSwapBacked(page))
329 SetPageSwapBacked(newpage);
331 return MIGRATEPAGE_SUCCESS;
334 spin_lock_irq(&mapping->tree_lock);
336 pslot = radix_tree_lookup_slot(&mapping->page_tree,
339 expected_count += 1 + page_has_private(page);
340 if (page_count(page) != expected_count ||
341 radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) {
342 spin_unlock_irq(&mapping->tree_lock);
346 if (!page_freeze_refs(page, expected_count)) {
347 spin_unlock_irq(&mapping->tree_lock);
352 * In the async migration case of moving a page with buffers, lock the
353 * buffers using trylock before the mapping is moved. If the mapping
354 * was moved, we later failed to lock the buffers and could not move
355 * the mapping back due to an elevated page count, we would have to
356 * block waiting on other references to be dropped.
358 if (mode == MIGRATE_ASYNC && head &&
359 !buffer_migrate_lock_buffers(head, mode)) {
360 page_unfreeze_refs(page, expected_count);
361 spin_unlock_irq(&mapping->tree_lock);
366 * Now we know that no one else is looking at the page:
367 * no turning back from here.
369 set_page_memcg(newpage, page_memcg(page));
370 newpage->index = page->index;
371 newpage->mapping = page->mapping;
372 if (PageSwapBacked(page))
373 SetPageSwapBacked(newpage);
375 get_page(newpage); /* add cache reference */
376 if (PageSwapCache(page)) {
377 SetPageSwapCache(newpage);
378 set_page_private(newpage, page_private(page));
381 radix_tree_replace_slot(pslot, newpage);
384 * Drop cache reference from old page by unfreezing
385 * to one less reference.
386 * We know this isn't the last reference.
388 page_unfreeze_refs(page, expected_count - 1);
391 * If moved to a different zone then also account
392 * the page for that zone. Other VM counters will be
393 * taken care of when we establish references to the
394 * new page and drop references to the old page.
396 * Note that anonymous pages are accounted for
397 * via NR_FILE_PAGES and NR_ANON_PAGES if they
398 * are mapped to swap space.
400 __dec_zone_page_state(page, NR_FILE_PAGES);
401 __inc_zone_page_state(newpage, NR_FILE_PAGES);
402 if (!PageSwapCache(page) && PageSwapBacked(page)) {
403 __dec_zone_page_state(page, NR_SHMEM);
404 __inc_zone_page_state(newpage, NR_SHMEM);
406 spin_unlock_irq(&mapping->tree_lock);
408 return MIGRATEPAGE_SUCCESS;
412 * The expected number of remaining references is the same as that
413 * of migrate_page_move_mapping().
415 int migrate_huge_page_move_mapping(struct address_space *mapping,
416 struct page *newpage, struct page *page)
421 spin_lock_irq(&mapping->tree_lock);
423 pslot = radix_tree_lookup_slot(&mapping->page_tree,
426 expected_count = 2 + page_has_private(page);
427 if (page_count(page) != expected_count ||
428 radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) {
429 spin_unlock_irq(&mapping->tree_lock);
433 if (!page_freeze_refs(page, expected_count)) {
434 spin_unlock_irq(&mapping->tree_lock);
438 set_page_memcg(newpage, page_memcg(page));
439 newpage->index = page->index;
440 newpage->mapping = page->mapping;
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_PAGE(PageUnevictable(page), 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);
542 if (page_is_young(page))
543 set_page_young(newpage);
544 if (page_is_idle(page))
545 set_page_idle(newpage);
548 * Copy NUMA information to the new page, to prevent over-eager
549 * future migrations of this same page.
551 cpupid = page_cpupid_xchg_last(page, -1);
552 page_cpupid_xchg_last(newpage, cpupid);
554 ksm_migrate_page(newpage, page);
556 * Please do not reorder this without considering how mm/ksm.c's
557 * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache().
559 if (PageSwapCache(page))
560 ClearPageSwapCache(page);
561 ClearPagePrivate(page);
562 set_page_private(page, 0);
565 * If any waiters have accumulated on the new page then
568 if (PageWriteback(newpage))
569 end_page_writeback(newpage);
572 /************************************************************
573 * Migration functions
574 ***********************************************************/
577 * Common logic to directly migrate a single page suitable for
578 * pages that do not use PagePrivate/PagePrivate2.
580 * Pages are locked upon entry and exit.
582 int migrate_page(struct address_space *mapping,
583 struct page *newpage, struct page *page,
584 enum migrate_mode mode)
588 BUG_ON(PageWriteback(page)); /* Writeback must be complete */
590 rc = migrate_page_move_mapping(mapping, newpage, page, NULL, mode, 0);
592 if (rc != MIGRATEPAGE_SUCCESS)
595 migrate_page_copy(newpage, page);
596 return MIGRATEPAGE_SUCCESS;
598 EXPORT_SYMBOL(migrate_page);
602 * Migration function for pages with buffers. This function can only be used
603 * if the underlying filesystem guarantees that no other references to "page"
606 int buffer_migrate_page(struct address_space *mapping,
607 struct page *newpage, struct page *page, enum migrate_mode mode)
609 struct buffer_head *bh, *head;
612 if (!page_has_buffers(page))
613 return migrate_page(mapping, newpage, page, mode);
615 head = page_buffers(page);
617 rc = migrate_page_move_mapping(mapping, newpage, page, head, mode, 0);
619 if (rc != MIGRATEPAGE_SUCCESS)
623 * In the async case, migrate_page_move_mapping locked the buffers
624 * with an IRQ-safe spinlock held. In the sync case, the buffers
625 * need to be locked now
627 if (mode != MIGRATE_ASYNC)
628 BUG_ON(!buffer_migrate_lock_buffers(head, mode));
630 ClearPagePrivate(page);
631 set_page_private(newpage, page_private(page));
632 set_page_private(page, 0);
638 set_bh_page(bh, newpage, bh_offset(bh));
639 bh = bh->b_this_page;
641 } while (bh != head);
643 SetPagePrivate(newpage);
645 migrate_page_copy(newpage, page);
651 bh = bh->b_this_page;
653 } while (bh != head);
655 return MIGRATEPAGE_SUCCESS;
657 EXPORT_SYMBOL(buffer_migrate_page);
661 * Writeback a page to clean the dirty state
663 static int writeout(struct address_space *mapping, struct page *page)
665 struct writeback_control wbc = {
666 .sync_mode = WB_SYNC_NONE,
669 .range_end = LLONG_MAX,
674 if (!mapping->a_ops->writepage)
675 /* No write method for the address space */
678 if (!clear_page_dirty_for_io(page))
679 /* Someone else already triggered a write */
683 * A dirty page may imply that the underlying filesystem has
684 * the page on some queue. So the page must be clean for
685 * migration. Writeout may mean we loose the lock and the
686 * page state is no longer what we checked for earlier.
687 * At this point we know that the migration attempt cannot
690 remove_migration_ptes(page, page);
692 rc = mapping->a_ops->writepage(page, &wbc);
694 if (rc != AOP_WRITEPAGE_ACTIVATE)
695 /* unlocked. Relock */
698 return (rc < 0) ? -EIO : -EAGAIN;
702 * Default handling if a filesystem does not provide a migration function.
704 static int fallback_migrate_page(struct address_space *mapping,
705 struct page *newpage, struct page *page, enum migrate_mode mode)
707 if (PageDirty(page)) {
708 /* Only writeback pages in full synchronous migration */
709 if (mode != MIGRATE_SYNC)
711 return writeout(mapping, page);
715 * Buffers may be managed in a filesystem specific way.
716 * We must have no buffers or drop them.
718 if (page_has_private(page) &&
719 !try_to_release_page(page, GFP_KERNEL))
722 return migrate_page(mapping, newpage, page, mode);
726 * Move a page to a newly allocated page
727 * The page is locked and all ptes have been successfully removed.
729 * The new page will have replaced the old page if this function
734 * MIGRATEPAGE_SUCCESS - success
736 static int move_to_new_page(struct page *newpage, struct page *page,
737 enum migrate_mode mode)
739 struct address_space *mapping;
742 VM_BUG_ON_PAGE(!PageLocked(page), page);
743 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
745 mapping = page_mapping(page);
747 rc = migrate_page(mapping, newpage, page, mode);
748 else if (mapping->a_ops->migratepage)
750 * Most pages have a mapping and most filesystems provide a
751 * migratepage callback. Anonymous pages are part of swap
752 * space which also has its own migratepage callback. This
753 * is the most common path for page migration.
755 rc = mapping->a_ops->migratepage(mapping, newpage, page, mode);
757 rc = fallback_migrate_page(mapping, newpage, page, mode);
760 * When successful, old pagecache page->mapping must be cleared before
761 * page is freed; but stats require that PageAnon be left as PageAnon.
763 if (rc == MIGRATEPAGE_SUCCESS) {
764 set_page_memcg(page, NULL);
766 page->mapping = NULL;
771 static int __unmap_and_move(struct page *page, struct page *newpage,
772 int force, enum migrate_mode mode)
775 int page_was_mapped = 0;
776 struct anon_vma *anon_vma = NULL;
778 if (!trylock_page(page)) {
779 if (!force || mode == MIGRATE_ASYNC)
783 * It's not safe for direct compaction to call lock_page.
784 * For example, during page readahead pages are added locked
785 * to the LRU. Later, when the IO completes the pages are
786 * marked uptodate and unlocked. However, the queueing
787 * could be merging multiple pages for one bio (e.g.
788 * mpage_readpages). If an allocation happens for the
789 * second or third page, the process can end up locking
790 * the same page twice and deadlocking. Rather than
791 * trying to be clever about what pages can be locked,
792 * avoid the use of lock_page for direct compaction
795 if (current->flags & PF_MEMALLOC)
801 if (PageWriteback(page)) {
803 * Only in the case of a full synchronous migration is it
804 * necessary to wait for PageWriteback. In the async case,
805 * the retry loop is too short and in the sync-light case,
806 * the overhead of stalling is too much
808 if (mode != MIGRATE_SYNC) {
814 wait_on_page_writeback(page);
818 * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
819 * we cannot notice that anon_vma is freed while we migrates a page.
820 * This get_anon_vma() delays freeing anon_vma pointer until the end
821 * of migration. File cache pages are no problem because of page_lock()
822 * File Caches may use write_page() or lock_page() in migration, then,
823 * just care Anon page here.
825 * Only page_get_anon_vma() understands the subtleties of
826 * getting a hold on an anon_vma from outside one of its mms.
827 * But if we cannot get anon_vma, then we won't need it anyway,
828 * because that implies that the anon page is no longer mapped
829 * (and cannot be remapped so long as we hold the page lock).
831 if (PageAnon(page) && !PageKsm(page))
832 anon_vma = page_get_anon_vma(page);
835 * Block others from accessing the new page when we get around to
836 * establishing additional references. We are usually the only one
837 * holding a reference to newpage at this point. We used to have a BUG
838 * here if trylock_page(newpage) fails, but would like to allow for
839 * cases where there might be a race with the previous use of newpage.
840 * This is much like races on refcount of oldpage: just don't BUG().
842 if (unlikely(!trylock_page(newpage)))
845 if (unlikely(isolated_balloon_page(page))) {
847 * A ballooned page does not need any special attention from
848 * physical to virtual reverse mapping procedures.
849 * Skip any attempt to unmap PTEs or to remap swap cache,
850 * in order to avoid burning cycles at rmap level, and perform
851 * the page migration right away (proteced by page lock).
853 rc = balloon_page_migrate(newpage, page, mode);
854 goto out_unlock_both;
858 * Corner case handling:
859 * 1. When a new swap-cache page is read into, it is added to the LRU
860 * and treated as swapcache but it has no rmap yet.
861 * Calling try_to_unmap() against a page->mapping==NULL page will
862 * trigger a BUG. So handle it here.
863 * 2. An orphaned page (see truncate_complete_page) might have
864 * fs-private metadata. The page can be picked up due to memory
865 * offlining. Everywhere else except page reclaim, the page is
866 * invisible to the vm, so the page can not be migrated. So try to
867 * free the metadata, so the page can be freed.
869 if (!page->mapping) {
870 VM_BUG_ON_PAGE(PageAnon(page), page);
871 if (page_has_private(page)) {
872 try_to_free_buffers(page);
873 goto out_unlock_both;
875 } else if (page_mapped(page)) {
876 /* Establish migration ptes */
877 VM_BUG_ON_PAGE(PageAnon(page) && !PageKsm(page) && !anon_vma,
880 TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
884 if (!page_mapped(page))
885 rc = move_to_new_page(newpage, page, mode);
888 remove_migration_ptes(page,
889 rc == MIGRATEPAGE_SUCCESS ? newpage : page);
892 unlock_page(newpage);
894 /* Drop an anon_vma reference if we took one */
896 put_anon_vma(anon_vma);
903 * gcc 4.7 and 4.8 on arm get an ICEs when inlining unmap_and_move(). Work
906 #if (GCC_VERSION >= 40700 && GCC_VERSION < 40900) && defined(CONFIG_ARM)
907 #define ICE_noinline noinline
913 * Obtain the lock on page, remove all ptes and migrate the page
914 * to the newly allocated page in newpage.
916 static ICE_noinline int unmap_and_move(new_page_t get_new_page,
917 free_page_t put_new_page,
918 unsigned long private, struct page *page,
919 int force, enum migrate_mode mode,
920 enum migrate_reason reason)
922 int rc = MIGRATEPAGE_SUCCESS;
924 struct page *newpage;
926 newpage = get_new_page(page, private, &result);
930 if (page_count(page) == 1) {
931 /* page was freed from under us. So we are done. */
935 if (unlikely(PageTransHuge(page)))
936 if (unlikely(split_huge_page(page)))
939 rc = __unmap_and_move(page, newpage, force, mode);
940 if (rc == MIGRATEPAGE_SUCCESS)
946 * A page that has been migrated has all references
947 * removed and will be freed. A page that has not been
948 * migrated will have kepts its references and be
951 list_del(&page->lru);
952 dec_zone_page_state(page, NR_ISOLATED_ANON +
953 page_is_file_cache(page));
954 /* Soft-offlined page shouldn't go through lru cache list */
955 if (reason == MR_MEMORY_FAILURE) {
957 if (!test_set_page_hwpoison(page))
958 num_poisoned_pages_inc();
960 putback_lru_page(page);
964 * If migration was not successful and there's a freeing callback, use
965 * it. Otherwise, putback_lru_page() will drop the reference grabbed
969 put_new_page(newpage, private);
970 else if (unlikely(__is_movable_balloon_page(newpage))) {
971 /* drop our reference, page already in the balloon */
974 putback_lru_page(newpage);
980 *result = page_to_nid(newpage);
986 * Counterpart of unmap_and_move_page() for hugepage migration.
988 * This function doesn't wait the completion of hugepage I/O
989 * because there is no race between I/O and migration for hugepage.
990 * Note that currently hugepage I/O occurs only in direct I/O
991 * where no lock is held and PG_writeback is irrelevant,
992 * and writeback status of all subpages are counted in the reference
993 * count of the head page (i.e. if all subpages of a 2MB hugepage are
994 * under direct I/O, the reference of the head page is 512 and a bit more.)
995 * This means that when we try to migrate hugepage whose subpages are
996 * doing direct I/O, some references remain after try_to_unmap() and
997 * hugepage migration fails without data corruption.
999 * There is also no race when direct I/O is issued on the page under migration,
1000 * because then pte is replaced with migration swap entry and direct I/O code
1001 * will wait in the page fault for migration to complete.
1003 static int unmap_and_move_huge_page(new_page_t get_new_page,
1004 free_page_t put_new_page, unsigned long private,
1005 struct page *hpage, int force,
1006 enum migrate_mode mode)
1010 int page_was_mapped = 0;
1011 struct page *new_hpage;
1012 struct anon_vma *anon_vma = NULL;
1015 * Movability of hugepages depends on architectures and hugepage size.
1016 * This check is necessary because some callers of hugepage migration
1017 * like soft offline and memory hotremove don't walk through page
1018 * tables or check whether the hugepage is pmd-based or not before
1019 * kicking migration.
1021 if (!hugepage_migration_supported(page_hstate(hpage))) {
1022 putback_active_hugepage(hpage);
1026 new_hpage = get_new_page(hpage, private, &result);
1030 if (!trylock_page(hpage)) {
1031 if (!force || mode != MIGRATE_SYNC)
1036 if (PageAnon(hpage))
1037 anon_vma = page_get_anon_vma(hpage);
1039 if (unlikely(!trylock_page(new_hpage)))
1042 if (page_mapped(hpage)) {
1044 TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
1045 page_was_mapped = 1;
1048 if (!page_mapped(hpage))
1049 rc = move_to_new_page(new_hpage, hpage, mode);
1051 if (page_was_mapped)
1052 remove_migration_ptes(hpage,
1053 rc == MIGRATEPAGE_SUCCESS ? new_hpage : hpage);
1055 unlock_page(new_hpage);
1059 put_anon_vma(anon_vma);
1061 if (rc == MIGRATEPAGE_SUCCESS) {
1062 hugetlb_cgroup_migrate(hpage, new_hpage);
1063 put_new_page = NULL;
1069 putback_active_hugepage(hpage);
1072 * If migration was not successful and there's a freeing callback, use
1073 * it. Otherwise, put_page() will drop the reference grabbed during
1077 put_new_page(new_hpage, private);
1079 putback_active_hugepage(new_hpage);
1085 *result = page_to_nid(new_hpage);
1091 * migrate_pages - migrate the pages specified in a list, to the free pages
1092 * supplied as the target for the page migration
1094 * @from: The list of pages to be migrated.
1095 * @get_new_page: The function used to allocate free pages to be used
1096 * as the target of the page migration.
1097 * @put_new_page: The function used to free target pages if migration
1098 * fails, or NULL if no special handling is necessary.
1099 * @private: Private data to be passed on to get_new_page()
1100 * @mode: The migration mode that specifies the constraints for
1101 * page migration, if any.
1102 * @reason: The reason for page migration.
1104 * The function returns after 10 attempts or if no pages are movable any more
1105 * because the list has become empty or no retryable pages exist any more.
1106 * The caller should call putback_movable_pages() to return pages to the LRU
1107 * or free list only if ret != 0.
1109 * Returns the number of pages that were not migrated, or an error code.
1111 int migrate_pages(struct list_head *from, new_page_t get_new_page,
1112 free_page_t put_new_page, unsigned long private,
1113 enum migrate_mode mode, int reason)
1117 int nr_succeeded = 0;
1121 int swapwrite = current->flags & PF_SWAPWRITE;
1125 current->flags |= PF_SWAPWRITE;
1127 for(pass = 0; pass < 10 && retry; pass++) {
1130 list_for_each_entry_safe(page, page2, from, lru) {
1134 rc = unmap_and_move_huge_page(get_new_page,
1135 put_new_page, private, page,
1138 rc = unmap_and_move(get_new_page, put_new_page,
1139 private, page, pass > 2, mode,
1148 case MIGRATEPAGE_SUCCESS:
1153 * Permanent failure (-EBUSY, -ENOSYS, etc.):
1154 * unlike -EAGAIN case, the failed page is
1155 * removed from migration page list and not
1156 * retried in the next outer loop.
1167 count_vm_events(PGMIGRATE_SUCCESS, nr_succeeded);
1169 count_vm_events(PGMIGRATE_FAIL, nr_failed);
1170 trace_mm_migrate_pages(nr_succeeded, nr_failed, mode, reason);
1173 current->flags &= ~PF_SWAPWRITE;
1180 * Move a list of individual pages
1182 struct page_to_node {
1189 static struct page *new_page_node(struct page *p, unsigned long private,
1192 struct page_to_node *pm = (struct page_to_node *)private;
1194 while (pm->node != MAX_NUMNODES && pm->page != p)
1197 if (pm->node == MAX_NUMNODES)
1200 *result = &pm->status;
1203 return alloc_huge_page_node(page_hstate(compound_head(p)),
1206 return __alloc_pages_node(pm->node,
1207 GFP_HIGHUSER_MOVABLE | __GFP_THISNODE, 0);
1211 * Move a set of pages as indicated in the pm array. The addr
1212 * field must be set to the virtual address of the page to be moved
1213 * and the node number must contain a valid target node.
1214 * The pm array ends with node = MAX_NUMNODES.
1216 static int do_move_page_to_node_array(struct mm_struct *mm,
1217 struct page_to_node *pm,
1221 struct page_to_node *pp;
1222 LIST_HEAD(pagelist);
1224 down_read(&mm->mmap_sem);
1227 * Build a list of pages to migrate
1229 for (pp = pm; pp->node != MAX_NUMNODES; pp++) {
1230 struct vm_area_struct *vma;
1234 vma = find_vma(mm, pp->addr);
1235 if (!vma || pp->addr < vma->vm_start || !vma_migratable(vma))
1238 /* FOLL_DUMP to ignore special (like zero) pages */
1239 page = follow_page(vma, pp->addr,
1240 FOLL_GET | FOLL_SPLIT | FOLL_DUMP);
1242 err = PTR_ERR(page);
1251 err = page_to_nid(page);
1253 if (err == pp->node)
1255 * Node already in the right place
1260 if (page_mapcount(page) > 1 &&
1264 if (PageHuge(page)) {
1266 isolate_huge_page(page, &pagelist);
1270 err = isolate_lru_page(page);
1272 list_add_tail(&page->lru, &pagelist);
1273 inc_zone_page_state(page, NR_ISOLATED_ANON +
1274 page_is_file_cache(page));
1278 * Either remove the duplicate refcount from
1279 * isolate_lru_page() or drop the page ref if it was
1288 if (!list_empty(&pagelist)) {
1289 err = migrate_pages(&pagelist, new_page_node, NULL,
1290 (unsigned long)pm, MIGRATE_SYNC, MR_SYSCALL);
1292 putback_movable_pages(&pagelist);
1295 up_read(&mm->mmap_sem);
1300 * Migrate an array of page address onto an array of nodes and fill
1301 * the corresponding array of status.
1303 static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes,
1304 unsigned long nr_pages,
1305 const void __user * __user *pages,
1306 const int __user *nodes,
1307 int __user *status, int flags)
1309 struct page_to_node *pm;
1310 unsigned long chunk_nr_pages;
1311 unsigned long chunk_start;
1315 pm = (struct page_to_node *)__get_free_page(GFP_KERNEL);
1322 * Store a chunk of page_to_node array in a page,
1323 * but keep the last one as a marker
1325 chunk_nr_pages = (PAGE_SIZE / sizeof(struct page_to_node)) - 1;
1327 for (chunk_start = 0;
1328 chunk_start < nr_pages;
1329 chunk_start += chunk_nr_pages) {
1332 if (chunk_start + chunk_nr_pages > nr_pages)
1333 chunk_nr_pages = nr_pages - chunk_start;
1335 /* fill the chunk pm with addrs and nodes from user-space */
1336 for (j = 0; j < chunk_nr_pages; j++) {
1337 const void __user *p;
1341 if (get_user(p, pages + j + chunk_start))
1343 pm[j].addr = (unsigned long) p;
1345 if (get_user(node, nodes + j + chunk_start))
1349 if (node < 0 || node >= MAX_NUMNODES)
1352 if (!node_state(node, N_MEMORY))
1356 if (!node_isset(node, task_nodes))
1362 /* End marker for this chunk */
1363 pm[chunk_nr_pages].node = MAX_NUMNODES;
1365 /* Migrate this chunk */
1366 err = do_move_page_to_node_array(mm, pm,
1367 flags & MPOL_MF_MOVE_ALL);
1371 /* Return status information */
1372 for (j = 0; j < chunk_nr_pages; j++)
1373 if (put_user(pm[j].status, status + j + chunk_start)) {
1381 free_page((unsigned long)pm);
1387 * Determine the nodes of an array of pages and store it in an array of status.
1389 static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages,
1390 const void __user **pages, int *status)
1394 down_read(&mm->mmap_sem);
1396 for (i = 0; i < nr_pages; i++) {
1397 unsigned long addr = (unsigned long)(*pages);
1398 struct vm_area_struct *vma;
1402 vma = find_vma(mm, addr);
1403 if (!vma || addr < vma->vm_start)
1406 /* FOLL_DUMP to ignore special (like zero) pages */
1407 page = follow_page(vma, addr, FOLL_DUMP);
1409 err = PTR_ERR(page);
1413 err = page ? page_to_nid(page) : -ENOENT;
1421 up_read(&mm->mmap_sem);
1425 * Determine the nodes of a user array of pages and store it in
1426 * a user array of status.
1428 static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages,
1429 const void __user * __user *pages,
1432 #define DO_PAGES_STAT_CHUNK_NR 16
1433 const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR];
1434 int chunk_status[DO_PAGES_STAT_CHUNK_NR];
1437 unsigned long chunk_nr;
1439 chunk_nr = nr_pages;
1440 if (chunk_nr > DO_PAGES_STAT_CHUNK_NR)
1441 chunk_nr = DO_PAGES_STAT_CHUNK_NR;
1443 if (copy_from_user(chunk_pages, pages, chunk_nr * sizeof(*chunk_pages)))
1446 do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status);
1448 if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status)))
1453 nr_pages -= chunk_nr;
1455 return nr_pages ? -EFAULT : 0;
1459 * Move a list of pages in the address space of the currently executing
1462 SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages,
1463 const void __user * __user *, pages,
1464 const int __user *, nodes,
1465 int __user *, status, int, flags)
1467 const struct cred *cred = current_cred(), *tcred;
1468 struct task_struct *task;
1469 struct mm_struct *mm;
1471 nodemask_t task_nodes;
1474 if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
1477 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1480 /* Find the mm_struct */
1482 task = pid ? find_task_by_vpid(pid) : current;
1487 get_task_struct(task);
1490 * Check if this process has the right to modify the specified
1491 * process. The right exists if the process has administrative
1492 * capabilities, superuser privileges or the same
1493 * userid as the target process.
1495 tcred = __task_cred(task);
1496 if (!uid_eq(cred->euid, tcred->suid) && !uid_eq(cred->euid, tcred->uid) &&
1497 !uid_eq(cred->uid, tcred->suid) && !uid_eq(cred->uid, tcred->uid) &&
1498 !capable(CAP_SYS_NICE)) {
1505 err = security_task_movememory(task);
1509 task_nodes = cpuset_mems_allowed(task);
1510 mm = get_task_mm(task);
1511 put_task_struct(task);
1517 err = do_pages_move(mm, task_nodes, nr_pages, pages,
1518 nodes, status, flags);
1520 err = do_pages_stat(mm, nr_pages, pages, status);
1526 put_task_struct(task);
1530 #ifdef CONFIG_NUMA_BALANCING
1532 * Returns true if this is a safe migration target node for misplaced NUMA
1533 * pages. Currently it only checks the watermarks which crude
1535 static bool migrate_balanced_pgdat(struct pglist_data *pgdat,
1536 unsigned long nr_migrate_pages)
1539 for (z = pgdat->nr_zones - 1; z >= 0; z--) {
1540 struct zone *zone = pgdat->node_zones + z;
1542 if (!populated_zone(zone))
1545 if (!zone_reclaimable(zone))
1548 /* Avoid waking kswapd by allocating pages_to_migrate pages. */
1549 if (!zone_watermark_ok(zone, 0,
1550 high_wmark_pages(zone) +
1559 static struct page *alloc_misplaced_dst_page(struct page *page,
1563 int nid = (int) data;
1564 struct page *newpage;
1566 newpage = __alloc_pages_node(nid,
1567 (GFP_HIGHUSER_MOVABLE |
1568 __GFP_THISNODE | __GFP_NOMEMALLOC |
1569 __GFP_NORETRY | __GFP_NOWARN) &
1576 * page migration rate limiting control.
1577 * Do not migrate more than @pages_to_migrate in a @migrate_interval_millisecs
1578 * window of time. Default here says do not migrate more than 1280M per second.
1580 static unsigned int migrate_interval_millisecs __read_mostly = 100;
1581 static unsigned int ratelimit_pages __read_mostly = 128 << (20 - PAGE_SHIFT);
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);
1660 * Attempt to migrate a misplaced page to the specified destination
1661 * node. Caller is expected to have an elevated reference count on
1662 * the page that will be dropped by this function before returning.
1664 int migrate_misplaced_page(struct page *page, struct vm_area_struct *vma,
1667 pg_data_t *pgdat = NODE_DATA(node);
1670 LIST_HEAD(migratepages);
1673 * Don't migrate file pages that are mapped in multiple processes
1674 * with execute permissions as they are probably shared libraries.
1676 if (page_mapcount(page) != 1 && page_is_file_cache(page) &&
1677 (vma->vm_flags & VM_EXEC))
1681 * Rate-limit the amount of data that is being migrated to a node.
1682 * Optimal placement is no good if the memory bus is saturated and
1683 * all the time is being spent migrating!
1685 if (numamigrate_update_ratelimit(pgdat, 1))
1688 isolated = numamigrate_isolate_page(pgdat, page);
1692 list_add(&page->lru, &migratepages);
1693 nr_remaining = migrate_pages(&migratepages, alloc_misplaced_dst_page,
1694 NULL, node, MIGRATE_ASYNC,
1697 if (!list_empty(&migratepages)) {
1698 list_del(&page->lru);
1699 dec_zone_page_state(page, NR_ISOLATED_ANON +
1700 page_is_file_cache(page));
1701 putback_lru_page(page);
1705 count_vm_numa_event(NUMA_PAGE_MIGRATE);
1706 BUG_ON(!list_empty(&migratepages));
1713 #endif /* CONFIG_NUMA_BALANCING */
1715 #if defined(CONFIG_NUMA_BALANCING) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
1717 * Migrates a THP to a given target node. page must be locked and is unlocked
1720 int migrate_misplaced_transhuge_page(struct mm_struct *mm,
1721 struct vm_area_struct *vma,
1722 pmd_t *pmd, pmd_t entry,
1723 unsigned long address,
1724 struct page *page, int node)
1727 pg_data_t *pgdat = NODE_DATA(node);
1729 struct page *new_page = NULL;
1730 int page_lru = page_is_file_cache(page);
1731 unsigned long mmun_start = address & HPAGE_PMD_MASK;
1732 unsigned long mmun_end = mmun_start + HPAGE_PMD_SIZE;
1736 * Rate-limit the amount of data that is being migrated to a node.
1737 * Optimal placement is no good if the memory bus is saturated and
1738 * all the time is being spent migrating!
1740 if (numamigrate_update_ratelimit(pgdat, HPAGE_PMD_NR))
1743 new_page = alloc_pages_node(node,
1744 (GFP_TRANSHUGE | __GFP_THISNODE) & ~__GFP_WAIT,
1749 isolated = numamigrate_isolate_page(pgdat, page);
1755 if (mm_tlb_flush_pending(mm))
1756 flush_tlb_range(vma, mmun_start, mmun_end);
1758 /* Prepare a page as a migration target */
1759 __set_page_locked(new_page);
1760 SetPageSwapBacked(new_page);
1762 /* anon mapping, we can simply copy page->mapping to the new page: */
1763 new_page->mapping = page->mapping;
1764 new_page->index = page->index;
1765 migrate_page_copy(new_page, page);
1766 WARN_ON(PageLRU(new_page));
1768 /* Recheck the target PMD */
1769 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1770 ptl = pmd_lock(mm, pmd);
1771 if (unlikely(!pmd_same(*pmd, entry) || page_count(page) != 2)) {
1774 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1776 /* Reverse changes made by migrate_page_copy() */
1777 if (TestClearPageActive(new_page))
1778 SetPageActive(page);
1779 if (TestClearPageUnevictable(new_page))
1780 SetPageUnevictable(page);
1782 unlock_page(new_page);
1783 put_page(new_page); /* Free it */
1785 /* Retake the callers reference and putback on LRU */
1787 putback_lru_page(page);
1788 mod_zone_page_state(page_zone(page),
1789 NR_ISOLATED_ANON + page_lru, -HPAGE_PMD_NR);
1795 entry = mk_pmd(new_page, vma->vm_page_prot);
1796 entry = pmd_mkhuge(entry);
1797 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1800 * Clear the old entry under pagetable lock and establish the new PTE.
1801 * Any parallel GUP will either observe the old page blocking on the
1802 * page lock, block on the page table lock or observe the new page.
1803 * The SetPageUptodate on the new page and page_add_new_anon_rmap
1804 * guarantee the copy is visible before the pagetable update.
1806 flush_cache_range(vma, mmun_start, mmun_end);
1807 page_add_anon_rmap(new_page, vma, mmun_start);
1808 pmdp_huge_clear_flush_notify(vma, mmun_start, pmd);
1809 set_pmd_at(mm, mmun_start, pmd, entry);
1810 flush_tlb_range(vma, mmun_start, mmun_end);
1811 update_mmu_cache_pmd(vma, address, &entry);
1813 if (page_count(page) != 2) {
1814 set_pmd_at(mm, mmun_start, pmd, orig_entry);
1815 flush_tlb_range(vma, mmun_start, mmun_end);
1816 mmu_notifier_invalidate_range(mm, mmun_start, mmun_end);
1817 update_mmu_cache_pmd(vma, address, &entry);
1818 page_remove_rmap(new_page);
1822 mlock_migrate_page(new_page, page);
1823 set_page_memcg(new_page, page_memcg(page));
1824 set_page_memcg(page, NULL);
1825 page_remove_rmap(page);
1828 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1830 /* Take an "isolate" reference and put new page on the LRU. */
1832 putback_lru_page(new_page);
1834 unlock_page(new_page);
1836 put_page(page); /* Drop the rmap reference */
1837 put_page(page); /* Drop the LRU isolation reference */
1839 count_vm_events(PGMIGRATE_SUCCESS, HPAGE_PMD_NR);
1840 count_vm_numa_events(NUMA_PAGE_MIGRATE, HPAGE_PMD_NR);
1842 mod_zone_page_state(page_zone(page),
1843 NR_ISOLATED_ANON + page_lru,
1848 count_vm_events(PGMIGRATE_FAIL, HPAGE_PMD_NR);
1850 ptl = pmd_lock(mm, pmd);
1851 if (pmd_same(*pmd, entry)) {
1852 entry = pmd_modify(entry, vma->vm_page_prot);
1853 set_pmd_at(mm, mmun_start, pmd, entry);
1854 update_mmu_cache_pmd(vma, address, &entry);
1863 #endif /* CONFIG_NUMA_BALANCING */
1865 #endif /* CONFIG_NUMA */