2 * Memory Migration functionality - linux/mm/migrate.c
4 * Copyright (C) 2006 Silicon Graphics, Inc., Christoph Lameter
6 * Page migration was first developed in the context of the memory hotplug
7 * project. The main authors of the migration code are:
9 * IWAMOTO Toshihiro <iwamoto@valinux.co.jp>
10 * Hirokazu Takahashi <taka@valinux.co.jp>
11 * Dave Hansen <haveblue@us.ibm.com>
15 #include <linux/migrate.h>
16 #include <linux/export.h>
17 #include <linux/swap.h>
18 #include <linux/swapops.h>
19 #include <linux/pagemap.h>
20 #include <linux/buffer_head.h>
21 #include <linux/mm_inline.h>
22 #include <linux/nsproxy.h>
23 #include <linux/pagevec.h>
24 #include <linux/ksm.h>
25 #include <linux/rmap.h>
26 #include <linux/topology.h>
27 #include <linux/cpu.h>
28 #include <linux/cpuset.h>
29 #include <linux/writeback.h>
30 #include <linux/mempolicy.h>
31 #include <linux/vmalloc.h>
32 #include <linux/security.h>
33 #include <linux/backing-dev.h>
34 #include <linux/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>
40 #include <linux/page_idle.h>
41 #include <linux/page_owner.h>
43 #include <asm/tlbflush.h>
45 #define CREATE_TRACE_POINTS
46 #include <trace/events/migrate.h>
51 * migrate_prep() needs to be called before we start compiling a list of pages
52 * to be migrated using isolate_lru_page(). If scheduling work on other CPUs is
53 * undesirable, use migrate_prep_local()
55 int migrate_prep(void)
58 * Clear the LRU lists so pages can be isolated.
59 * Note that pages may be moved off the LRU after we have
60 * drained them. Those pages will fail to migrate like other
61 * pages that may be busy.
68 /* Do the necessary work of migrate_prep but not if it involves other CPUs */
69 int migrate_prep_local(void)
77 * Put previously isolated pages back onto the appropriate lists
78 * from where they were once taken off for compaction/migration.
80 * This function shall be used whenever the isolated pageset has been
81 * built from lru, balloon, hugetlbfs page. See isolate_migratepages_range()
82 * and isolate_huge_page().
84 void putback_movable_pages(struct list_head *l)
89 list_for_each_entry_safe(page, page2, l, lru) {
90 if (unlikely(PageHuge(page))) {
91 putback_active_hugepage(page);
95 dec_zone_page_state(page, NR_ISOLATED_ANON +
96 page_is_file_cache(page));
97 if (unlikely(isolated_balloon_page(page)))
98 balloon_page_putback(page);
100 putback_lru_page(page);
105 * Restore a potential migration pte to a working pte entry
107 static int remove_migration_pte(struct page *new, struct vm_area_struct *vma,
108 unsigned long addr, void *old)
110 struct mm_struct *mm = vma->vm_mm;
116 if (unlikely(PageHuge(new))) {
117 ptep = huge_pte_offset(mm, addr);
120 ptl = huge_pte_lockptr(hstate_vma(vma), mm, ptep);
122 pmd = mm_find_pmd(mm, addr);
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);
152 /* Recheck VMA as permissions can change since migration started */
153 if (is_write_migration_entry(entry))
154 pte = maybe_mkwrite(pte, vma);
156 #ifdef CONFIG_HUGETLB_PAGE
158 pte = pte_mkhuge(pte);
159 pte = arch_make_huge_pte(pte, vma, new, 0);
162 flush_dcache_page(new);
163 set_pte_at(mm, addr, ptep, pte);
167 hugepage_add_anon_rmap(new, vma, addr);
169 page_dup_rmap(new, true);
170 } else if (PageAnon(new))
171 page_add_anon_rmap(new, vma, addr, false);
173 page_add_file_rmap(new);
175 if (vma->vm_flags & VM_LOCKED)
178 /* No need to invalidate - it was non-present before */
179 update_mmu_cache(vma, addr, ptep);
181 pte_unmap_unlock(ptep, ptl);
187 * Get rid of all migration entries and replace them by
188 * references to the indicated page.
190 static void remove_migration_ptes(struct page *old, struct page *new)
192 struct rmap_walk_control rwc = {
193 .rmap_one = remove_migration_pte,
197 rmap_walk(new, &rwc);
201 * Something used the pte of a page under migration. We need to
202 * get to the page and wait until migration is finished.
203 * When we return from this function the fault will be retried.
205 void __migration_entry_wait(struct mm_struct *mm, pte_t *ptep,
214 if (!is_swap_pte(pte))
217 entry = pte_to_swp_entry(pte);
218 if (!is_migration_entry(entry))
221 page = migration_entry_to_page(entry);
224 * Once radix-tree replacement of page migration started, page_count
225 * *must* be zero. And, we don't want to call wait_on_page_locked()
226 * against a page without get_page().
227 * So, we use get_page_unless_zero(), here. Even failed, page fault
230 if (!get_page_unless_zero(page))
232 pte_unmap_unlock(ptep, ptl);
233 wait_on_page_locked(page);
237 pte_unmap_unlock(ptep, ptl);
240 void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
241 unsigned long address)
243 spinlock_t *ptl = pte_lockptr(mm, pmd);
244 pte_t *ptep = pte_offset_map(pmd, address);
245 __migration_entry_wait(mm, ptep, ptl);
248 void migration_entry_wait_huge(struct vm_area_struct *vma,
249 struct mm_struct *mm, pte_t *pte)
251 spinlock_t *ptl = huge_pte_lockptr(hstate_vma(vma), mm, pte);
252 __migration_entry_wait(mm, pte, ptl);
256 /* Returns true if all buffers are successfully locked */
257 static bool buffer_migrate_lock_buffers(struct buffer_head *head,
258 enum migrate_mode mode)
260 struct buffer_head *bh = head;
262 /* Simple case, sync compaction */
263 if (mode != MIGRATE_ASYNC) {
267 bh = bh->b_this_page;
269 } while (bh != head);
274 /* async case, we cannot block on lock_buffer so use trylock_buffer */
277 if (!trylock_buffer(bh)) {
279 * We failed to lock the buffer and cannot stall in
280 * async migration. Release the taken locks
282 struct buffer_head *failed_bh = bh;
285 while (bh != failed_bh) {
288 bh = bh->b_this_page;
293 bh = bh->b_this_page;
294 } while (bh != head);
298 static inline bool buffer_migrate_lock_buffers(struct buffer_head *head,
299 enum migrate_mode mode)
303 #endif /* CONFIG_BLOCK */
306 * Replace the page in the mapping.
308 * The number of remaining references must be:
309 * 1 for anonymous pages without a mapping
310 * 2 for pages with a mapping
311 * 3 for pages with a mapping and PagePrivate/PagePrivate2 set.
313 int migrate_page_move_mapping(struct address_space *mapping,
314 struct page *newpage, struct page *page,
315 struct buffer_head *head, enum migrate_mode mode,
318 struct zone *oldzone, *newzone;
320 int expected_count = 1 + extra_count;
324 /* Anonymous page without mapping */
325 if (page_count(page) != expected_count)
328 /* No turning back from here */
329 set_page_memcg(newpage, page_memcg(page));
330 newpage->index = page->index;
331 newpage->mapping = page->mapping;
332 if (PageSwapBacked(page))
333 SetPageSwapBacked(newpage);
335 return MIGRATEPAGE_SUCCESS;
338 oldzone = page_zone(page);
339 newzone = page_zone(newpage);
341 spin_lock_irq(&mapping->tree_lock);
343 pslot = radix_tree_lookup_slot(&mapping->page_tree,
346 expected_count += 1 + page_has_private(page);
347 if (page_count(page) != expected_count ||
348 radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) {
349 spin_unlock_irq(&mapping->tree_lock);
353 if (!page_freeze_refs(page, expected_count)) {
354 spin_unlock_irq(&mapping->tree_lock);
359 * In the async migration case of moving a page with buffers, lock the
360 * buffers using trylock before the mapping is moved. If the mapping
361 * was moved, we later failed to lock the buffers and could not move
362 * the mapping back due to an elevated page count, we would have to
363 * block waiting on other references to be dropped.
365 if (mode == MIGRATE_ASYNC && head &&
366 !buffer_migrate_lock_buffers(head, mode)) {
367 page_unfreeze_refs(page, expected_count);
368 spin_unlock_irq(&mapping->tree_lock);
373 * Now we know that no one else is looking at the page:
374 * no turning back from here.
376 set_page_memcg(newpage, page_memcg(page));
377 newpage->index = page->index;
378 newpage->mapping = page->mapping;
379 if (PageSwapBacked(page))
380 SetPageSwapBacked(newpage);
382 get_page(newpage); /* add cache reference */
383 if (PageSwapCache(page)) {
384 SetPageSwapCache(newpage);
385 set_page_private(newpage, page_private(page));
388 /* Move dirty while page refs frozen and newpage not yet exposed */
389 dirty = PageDirty(page);
391 ClearPageDirty(page);
392 SetPageDirty(newpage);
395 radix_tree_replace_slot(pslot, newpage);
398 * Drop cache reference from old page by unfreezing
399 * to one less reference.
400 * We know this isn't the last reference.
402 page_unfreeze_refs(page, expected_count - 1);
404 spin_unlock(&mapping->tree_lock);
405 /* Leave irq disabled to prevent preemption while updating stats */
408 * If moved to a different zone then also account
409 * the page for that zone. Other VM counters will be
410 * taken care of when we establish references to the
411 * new page and drop references to the old page.
413 * Note that anonymous pages are accounted for
414 * via NR_FILE_PAGES and NR_ANON_PAGES if they
415 * are mapped to swap space.
417 if (newzone != oldzone) {
418 __dec_zone_state(oldzone, NR_FILE_PAGES);
419 __inc_zone_state(newzone, NR_FILE_PAGES);
420 if (PageSwapBacked(page) && !PageSwapCache(page)) {
421 __dec_zone_state(oldzone, NR_SHMEM);
422 __inc_zone_state(newzone, NR_SHMEM);
424 if (dirty && mapping_cap_account_dirty(mapping)) {
425 __dec_zone_state(oldzone, NR_FILE_DIRTY);
426 __inc_zone_state(newzone, NR_FILE_DIRTY);
431 return MIGRATEPAGE_SUCCESS;
435 * The expected number of remaining references is the same as that
436 * of migrate_page_move_mapping().
438 int migrate_huge_page_move_mapping(struct address_space *mapping,
439 struct page *newpage, struct page *page)
444 spin_lock_irq(&mapping->tree_lock);
446 pslot = radix_tree_lookup_slot(&mapping->page_tree,
449 expected_count = 2 + page_has_private(page);
450 if (page_count(page) != expected_count ||
451 radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) {
452 spin_unlock_irq(&mapping->tree_lock);
456 if (!page_freeze_refs(page, expected_count)) {
457 spin_unlock_irq(&mapping->tree_lock);
461 set_page_memcg(newpage, page_memcg(page));
462 newpage->index = page->index;
463 newpage->mapping = page->mapping;
466 radix_tree_replace_slot(pslot, newpage);
468 page_unfreeze_refs(page, expected_count - 1);
470 spin_unlock_irq(&mapping->tree_lock);
471 return MIGRATEPAGE_SUCCESS;
475 * Gigantic pages are so large that we do not guarantee that page++ pointer
476 * arithmetic will work across the entire page. We need something more
479 static void __copy_gigantic_page(struct page *dst, struct page *src,
483 struct page *dst_base = dst;
484 struct page *src_base = src;
486 for (i = 0; i < nr_pages; ) {
488 copy_highpage(dst, src);
491 dst = mem_map_next(dst, dst_base, i);
492 src = mem_map_next(src, src_base, i);
496 static void copy_huge_page(struct page *dst, struct page *src)
503 struct hstate *h = page_hstate(src);
504 nr_pages = pages_per_huge_page(h);
506 if (unlikely(nr_pages > MAX_ORDER_NR_PAGES)) {
507 __copy_gigantic_page(dst, src, nr_pages);
512 BUG_ON(!PageTransHuge(src));
513 nr_pages = hpage_nr_pages(src);
516 for (i = 0; i < nr_pages; i++) {
518 copy_highpage(dst + i, src + i);
523 * Copy the page to its new location
525 void migrate_page_copy(struct page *newpage, struct page *page)
529 if (PageHuge(page) || PageTransHuge(page))
530 copy_huge_page(newpage, page);
532 copy_highpage(newpage, page);
535 SetPageError(newpage);
536 if (PageReferenced(page))
537 SetPageReferenced(newpage);
538 if (PageUptodate(page))
539 SetPageUptodate(newpage);
540 if (TestClearPageActive(page)) {
541 VM_BUG_ON_PAGE(PageUnevictable(page), page);
542 SetPageActive(newpage);
543 } else if (TestClearPageUnevictable(page))
544 SetPageUnevictable(newpage);
545 if (PageChecked(page))
546 SetPageChecked(newpage);
547 if (PageMappedToDisk(page))
548 SetPageMappedToDisk(newpage);
550 /* Move dirty on pages not done by migrate_page_move_mapping() */
552 SetPageDirty(newpage);
554 if (page_is_young(page))
555 set_page_young(newpage);
556 if (page_is_idle(page))
557 set_page_idle(newpage);
560 * Copy NUMA information to the new page, to prevent over-eager
561 * future migrations of this same page.
563 cpupid = page_cpupid_xchg_last(page, -1);
564 page_cpupid_xchg_last(newpage, cpupid);
566 ksm_migrate_page(newpage, page);
568 * Please do not reorder this without considering how mm/ksm.c's
569 * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache().
571 if (PageSwapCache(page))
572 ClearPageSwapCache(page);
573 ClearPagePrivate(page);
574 set_page_private(page, 0);
577 * If any waiters have accumulated on the new page then
580 if (PageWriteback(newpage))
581 end_page_writeback(newpage);
583 copy_page_owner(page, newpage);
586 /************************************************************
587 * Migration functions
588 ***********************************************************/
591 * Common logic to directly migrate a single page suitable for
592 * pages that do not use PagePrivate/PagePrivate2.
594 * Pages are locked upon entry and exit.
596 int migrate_page(struct address_space *mapping,
597 struct page *newpage, struct page *page,
598 enum migrate_mode mode)
602 BUG_ON(PageWriteback(page)); /* Writeback must be complete */
604 rc = migrate_page_move_mapping(mapping, newpage, page, NULL, mode, 0);
606 if (rc != MIGRATEPAGE_SUCCESS)
609 migrate_page_copy(newpage, page);
610 return MIGRATEPAGE_SUCCESS;
612 EXPORT_SYMBOL(migrate_page);
616 * Migration function for pages with buffers. This function can only be used
617 * if the underlying filesystem guarantees that no other references to "page"
620 int buffer_migrate_page(struct address_space *mapping,
621 struct page *newpage, struct page *page, enum migrate_mode mode)
623 struct buffer_head *bh, *head;
626 if (!page_has_buffers(page))
627 return migrate_page(mapping, newpage, page, mode);
629 head = page_buffers(page);
631 rc = migrate_page_move_mapping(mapping, newpage, page, head, mode, 0);
633 if (rc != MIGRATEPAGE_SUCCESS)
637 * In the async case, migrate_page_move_mapping locked the buffers
638 * with an IRQ-safe spinlock held. In the sync case, the buffers
639 * need to be locked now
641 if (mode != MIGRATE_ASYNC)
642 BUG_ON(!buffer_migrate_lock_buffers(head, mode));
644 ClearPagePrivate(page);
645 set_page_private(newpage, page_private(page));
646 set_page_private(page, 0);
652 set_bh_page(bh, newpage, bh_offset(bh));
653 bh = bh->b_this_page;
655 } while (bh != head);
657 SetPagePrivate(newpage);
659 migrate_page_copy(newpage, page);
665 bh = bh->b_this_page;
667 } while (bh != head);
669 return MIGRATEPAGE_SUCCESS;
671 EXPORT_SYMBOL(buffer_migrate_page);
675 * Writeback a page to clean the dirty state
677 static int writeout(struct address_space *mapping, struct page *page)
679 struct writeback_control wbc = {
680 .sync_mode = WB_SYNC_NONE,
683 .range_end = LLONG_MAX,
688 if (!mapping->a_ops->writepage)
689 /* No write method for the address space */
692 if (!clear_page_dirty_for_io(page))
693 /* Someone else already triggered a write */
697 * A dirty page may imply that the underlying filesystem has
698 * the page on some queue. So the page must be clean for
699 * migration. Writeout may mean we loose the lock and the
700 * page state is no longer what we checked for earlier.
701 * At this point we know that the migration attempt cannot
704 remove_migration_ptes(page, page);
706 rc = mapping->a_ops->writepage(page, &wbc);
708 if (rc != AOP_WRITEPAGE_ACTIVATE)
709 /* unlocked. Relock */
712 return (rc < 0) ? -EIO : -EAGAIN;
716 * Default handling if a filesystem does not provide a migration function.
718 static int fallback_migrate_page(struct address_space *mapping,
719 struct page *newpage, struct page *page, enum migrate_mode mode)
721 if (PageDirty(page)) {
722 /* Only writeback pages in full synchronous migration */
723 if (mode != MIGRATE_SYNC)
725 return writeout(mapping, page);
729 * Buffers may be managed in a filesystem specific way.
730 * We must have no buffers or drop them.
732 if (page_has_private(page) &&
733 !try_to_release_page(page, GFP_KERNEL))
736 return migrate_page(mapping, newpage, page, mode);
740 * Move a page to a newly allocated page
741 * The page is locked and all ptes have been successfully removed.
743 * The new page will have replaced the old page if this function
748 * MIGRATEPAGE_SUCCESS - success
750 static int move_to_new_page(struct page *newpage, struct page *page,
751 enum migrate_mode mode)
753 struct address_space *mapping;
756 VM_BUG_ON_PAGE(!PageLocked(page), page);
757 VM_BUG_ON_PAGE(!PageLocked(newpage), 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, newpage, page, mode);
771 rc = fallback_migrate_page(mapping, newpage, page, mode);
774 * When successful, old pagecache page->mapping must be cleared before
775 * page is freed; but stats require that PageAnon be left as PageAnon.
777 if (rc == MIGRATEPAGE_SUCCESS) {
778 set_page_memcg(page, NULL);
780 page->mapping = NULL;
785 static int __unmap_and_move(struct page *page, struct page *newpage,
786 int force, enum migrate_mode mode)
789 int page_was_mapped = 0;
790 struct anon_vma *anon_vma = NULL;
792 if (!trylock_page(page)) {
793 if (!force || mode == MIGRATE_ASYNC)
797 * It's not safe for direct compaction to call lock_page.
798 * For example, during page readahead pages are added locked
799 * to the LRU. Later, when the IO completes the pages are
800 * marked uptodate and unlocked. However, the queueing
801 * could be merging multiple pages for one bio (e.g.
802 * mpage_readpages). If an allocation happens for the
803 * second or third page, the process can end up locking
804 * the same page twice and deadlocking. Rather than
805 * trying to be clever about what pages can be locked,
806 * avoid the use of lock_page for direct compaction
809 if (current->flags & PF_MEMALLOC)
815 if (PageWriteback(page)) {
817 * Only in the case of a full synchronous migration is it
818 * necessary to wait for PageWriteback. In the async case,
819 * the retry loop is too short and in the sync-light case,
820 * the overhead of stalling is too much
822 if (mode != MIGRATE_SYNC) {
828 wait_on_page_writeback(page);
832 * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
833 * we cannot notice that anon_vma is freed while we migrates a page.
834 * This get_anon_vma() delays freeing anon_vma pointer until the end
835 * of migration. File cache pages are no problem because of page_lock()
836 * File Caches may use write_page() or lock_page() in migration, then,
837 * just care Anon page here.
839 * Only page_get_anon_vma() understands the subtleties of
840 * getting a hold on an anon_vma from outside one of its mms.
841 * But if we cannot get anon_vma, then we won't need it anyway,
842 * because that implies that the anon page is no longer mapped
843 * (and cannot be remapped so long as we hold the page lock).
845 if (PageAnon(page) && !PageKsm(page))
846 anon_vma = page_get_anon_vma(page);
849 * Block others from accessing the new page when we get around to
850 * establishing additional references. We are usually the only one
851 * holding a reference to newpage at this point. We used to have a BUG
852 * here if trylock_page(newpage) fails, but would like to allow for
853 * cases where there might be a race with the previous use of newpage.
854 * This is much like races on refcount of oldpage: just don't BUG().
856 if (unlikely(!trylock_page(newpage)))
859 if (unlikely(isolated_balloon_page(page))) {
861 * A ballooned page does not need any special attention from
862 * physical to virtual reverse mapping procedures.
863 * Skip any attempt to unmap PTEs or to remap swap cache,
864 * in order to avoid burning cycles at rmap level, and perform
865 * the page migration right away (proteced by page lock).
867 rc = balloon_page_migrate(newpage, page, mode);
868 goto out_unlock_both;
872 * Corner case handling:
873 * 1. When a new swap-cache page is read into, it is added to the LRU
874 * and treated as swapcache but it has no rmap yet.
875 * Calling try_to_unmap() against a page->mapping==NULL page will
876 * trigger a BUG. So handle it here.
877 * 2. An orphaned page (see truncate_complete_page) might have
878 * fs-private metadata. The page can be picked up due to memory
879 * offlining. Everywhere else except page reclaim, the page is
880 * invisible to the vm, so the page can not be migrated. So try to
881 * free the metadata, so the page can be freed.
883 if (!page->mapping) {
884 VM_BUG_ON_PAGE(PageAnon(page), page);
885 if (page_has_private(page)) {
886 try_to_free_buffers(page);
887 goto out_unlock_both;
889 } else if (page_mapped(page)) {
890 /* Establish migration ptes */
891 VM_BUG_ON_PAGE(PageAnon(page) && !PageKsm(page) && !anon_vma,
894 TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
898 if (!page_mapped(page))
899 rc = move_to_new_page(newpage, page, mode);
902 remove_migration_ptes(page,
903 rc == MIGRATEPAGE_SUCCESS ? newpage : page);
906 unlock_page(newpage);
908 /* Drop an anon_vma reference if we took one */
910 put_anon_vma(anon_vma);
917 * gcc 4.7 and 4.8 on arm get an ICEs when inlining unmap_and_move(). Work
920 #if (GCC_VERSION >= 40700 && GCC_VERSION < 40900) && defined(CONFIG_ARM)
921 #define ICE_noinline noinline
927 * Obtain the lock on page, remove all ptes and migrate the page
928 * to the newly allocated page in newpage.
930 static ICE_noinline int unmap_and_move(new_page_t get_new_page,
931 free_page_t put_new_page,
932 unsigned long private, struct page *page,
933 int force, enum migrate_mode mode,
934 enum migrate_reason reason)
936 int rc = MIGRATEPAGE_SUCCESS;
938 struct page *newpage;
940 newpage = get_new_page(page, private, &result);
944 if (page_count(page) == 1) {
945 /* page was freed from under us. So we are done. */
949 if (unlikely(PageTransHuge(page))) {
951 rc = split_huge_page(page);
957 rc = __unmap_and_move(page, newpage, force, mode);
958 if (rc == MIGRATEPAGE_SUCCESS)
964 * A page that has been migrated has all references
965 * removed and will be freed. A page that has not been
966 * migrated will have kepts its references and be
969 list_del(&page->lru);
970 dec_zone_page_state(page, NR_ISOLATED_ANON +
971 page_is_file_cache(page));
972 /* Soft-offlined page shouldn't go through lru cache list */
973 if (reason == MR_MEMORY_FAILURE) {
975 if (!test_set_page_hwpoison(page))
976 num_poisoned_pages_inc();
978 putback_lru_page(page);
982 * If migration was not successful and there's a freeing callback, use
983 * it. Otherwise, putback_lru_page() will drop the reference grabbed
987 put_new_page(newpage, private);
988 else if (unlikely(__is_movable_balloon_page(newpage))) {
989 /* drop our reference, page already in the balloon */
992 putback_lru_page(newpage);
998 *result = page_to_nid(newpage);
1004 * Counterpart of unmap_and_move_page() for hugepage migration.
1006 * This function doesn't wait the completion of hugepage I/O
1007 * because there is no race between I/O and migration for hugepage.
1008 * Note that currently hugepage I/O occurs only in direct I/O
1009 * where no lock is held and PG_writeback is irrelevant,
1010 * and writeback status of all subpages are counted in the reference
1011 * count of the head page (i.e. if all subpages of a 2MB hugepage are
1012 * under direct I/O, the reference of the head page is 512 and a bit more.)
1013 * This means that when we try to migrate hugepage whose subpages are
1014 * doing direct I/O, some references remain after try_to_unmap() and
1015 * hugepage migration fails without data corruption.
1017 * There is also no race when direct I/O is issued on the page under migration,
1018 * because then pte is replaced with migration swap entry and direct I/O code
1019 * will wait in the page fault for migration to complete.
1021 static int unmap_and_move_huge_page(new_page_t get_new_page,
1022 free_page_t put_new_page, unsigned long private,
1023 struct page *hpage, int force,
1024 enum migrate_mode mode)
1028 int page_was_mapped = 0;
1029 struct page *new_hpage;
1030 struct anon_vma *anon_vma = NULL;
1033 * Movability of hugepages depends on architectures and hugepage size.
1034 * This check is necessary because some callers of hugepage migration
1035 * like soft offline and memory hotremove don't walk through page
1036 * tables or check whether the hugepage is pmd-based or not before
1037 * kicking migration.
1039 if (!hugepage_migration_supported(page_hstate(hpage))) {
1040 putback_active_hugepage(hpage);
1044 new_hpage = get_new_page(hpage, private, &result);
1048 if (!trylock_page(hpage)) {
1049 if (!force || mode != MIGRATE_SYNC)
1054 if (PageAnon(hpage))
1055 anon_vma = page_get_anon_vma(hpage);
1057 if (unlikely(!trylock_page(new_hpage)))
1060 if (page_mapped(hpage)) {
1062 TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
1063 page_was_mapped = 1;
1066 if (!page_mapped(hpage))
1067 rc = move_to_new_page(new_hpage, hpage, mode);
1069 if (page_was_mapped)
1070 remove_migration_ptes(hpage,
1071 rc == MIGRATEPAGE_SUCCESS ? new_hpage : hpage);
1073 unlock_page(new_hpage);
1077 put_anon_vma(anon_vma);
1079 if (rc == MIGRATEPAGE_SUCCESS) {
1080 hugetlb_cgroup_migrate(hpage, new_hpage);
1081 put_new_page = NULL;
1087 putback_active_hugepage(hpage);
1090 * If migration was not successful and there's a freeing callback, use
1091 * it. Otherwise, put_page() will drop the reference grabbed during
1095 put_new_page(new_hpage, private);
1097 putback_active_hugepage(new_hpage);
1103 *result = page_to_nid(new_hpage);
1109 * migrate_pages - migrate the pages specified in a list, to the free pages
1110 * supplied as the target for the page migration
1112 * @from: The list of pages to be migrated.
1113 * @get_new_page: The function used to allocate free pages to be used
1114 * as the target of the page migration.
1115 * @put_new_page: The function used to free target pages if migration
1116 * fails, or NULL if no special handling is necessary.
1117 * @private: Private data to be passed on to get_new_page()
1118 * @mode: The migration mode that specifies the constraints for
1119 * page migration, if any.
1120 * @reason: The reason for page migration.
1122 * The function returns after 10 attempts or if no pages are movable any more
1123 * because the list has become empty or no retryable pages exist any more.
1124 * The caller should call putback_movable_pages() to return pages to the LRU
1125 * or free list only if ret != 0.
1127 * Returns the number of pages that were not migrated, or an error code.
1129 int migrate_pages(struct list_head *from, new_page_t get_new_page,
1130 free_page_t put_new_page, unsigned long private,
1131 enum migrate_mode mode, int reason)
1135 int nr_succeeded = 0;
1139 int swapwrite = current->flags & PF_SWAPWRITE;
1143 current->flags |= PF_SWAPWRITE;
1145 for(pass = 0; pass < 10 && retry; pass++) {
1148 list_for_each_entry_safe(page, page2, from, lru) {
1152 rc = unmap_and_move_huge_page(get_new_page,
1153 put_new_page, private, page,
1156 rc = unmap_and_move(get_new_page, put_new_page,
1157 private, page, pass > 2, mode,
1166 case MIGRATEPAGE_SUCCESS:
1171 * Permanent failure (-EBUSY, -ENOSYS, etc.):
1172 * unlike -EAGAIN case, the failed page is
1173 * removed from migration page list and not
1174 * retried in the next outer loop.
1185 count_vm_events(PGMIGRATE_SUCCESS, nr_succeeded);
1187 count_vm_events(PGMIGRATE_FAIL, nr_failed);
1188 trace_mm_migrate_pages(nr_succeeded, nr_failed, mode, reason);
1191 current->flags &= ~PF_SWAPWRITE;
1198 * Move a list of individual pages
1200 struct page_to_node {
1207 static struct page *new_page_node(struct page *p, unsigned long private,
1210 struct page_to_node *pm = (struct page_to_node *)private;
1212 while (pm->node != MAX_NUMNODES && pm->page != p)
1215 if (pm->node == MAX_NUMNODES)
1218 *result = &pm->status;
1221 return alloc_huge_page_node(page_hstate(compound_head(p)),
1224 return __alloc_pages_node(pm->node,
1225 GFP_HIGHUSER_MOVABLE | __GFP_THISNODE, 0);
1229 * Move a set of pages as indicated in the pm array. The addr
1230 * field must be set to the virtual address of the page to be moved
1231 * and the node number must contain a valid target node.
1232 * The pm array ends with node = MAX_NUMNODES.
1234 static int do_move_page_to_node_array(struct mm_struct *mm,
1235 struct page_to_node *pm,
1239 struct page_to_node *pp;
1240 LIST_HEAD(pagelist);
1242 down_read(&mm->mmap_sem);
1245 * Build a list of pages to migrate
1247 for (pp = pm; pp->node != MAX_NUMNODES; pp++) {
1248 struct vm_area_struct *vma;
1252 vma = find_vma(mm, pp->addr);
1253 if (!vma || pp->addr < vma->vm_start || !vma_migratable(vma))
1256 /* FOLL_DUMP to ignore special (like zero) pages */
1257 page = follow_page(vma, pp->addr,
1258 FOLL_GET | FOLL_SPLIT | FOLL_DUMP);
1260 err = PTR_ERR(page);
1269 err = page_to_nid(page);
1271 if (err == pp->node)
1273 * Node already in the right place
1278 if (page_mapcount(page) > 1 &&
1282 if (PageHuge(page)) {
1284 isolate_huge_page(page, &pagelist);
1288 err = isolate_lru_page(page);
1290 list_add_tail(&page->lru, &pagelist);
1291 inc_zone_page_state(page, NR_ISOLATED_ANON +
1292 page_is_file_cache(page));
1296 * Either remove the duplicate refcount from
1297 * isolate_lru_page() or drop the page ref if it was
1306 if (!list_empty(&pagelist)) {
1307 err = migrate_pages(&pagelist, new_page_node, NULL,
1308 (unsigned long)pm, MIGRATE_SYNC, MR_SYSCALL);
1310 putback_movable_pages(&pagelist);
1313 up_read(&mm->mmap_sem);
1318 * Migrate an array of page address onto an array of nodes and fill
1319 * the corresponding array of status.
1321 static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes,
1322 unsigned long nr_pages,
1323 const void __user * __user *pages,
1324 const int __user *nodes,
1325 int __user *status, int flags)
1327 struct page_to_node *pm;
1328 unsigned long chunk_nr_pages;
1329 unsigned long chunk_start;
1333 pm = (struct page_to_node *)__get_free_page(GFP_KERNEL);
1340 * Store a chunk of page_to_node array in a page,
1341 * but keep the last one as a marker
1343 chunk_nr_pages = (PAGE_SIZE / sizeof(struct page_to_node)) - 1;
1345 for (chunk_start = 0;
1346 chunk_start < nr_pages;
1347 chunk_start += chunk_nr_pages) {
1350 if (chunk_start + chunk_nr_pages > nr_pages)
1351 chunk_nr_pages = nr_pages - chunk_start;
1353 /* fill the chunk pm with addrs and nodes from user-space */
1354 for (j = 0; j < chunk_nr_pages; j++) {
1355 const void __user *p;
1359 if (get_user(p, pages + j + chunk_start))
1361 pm[j].addr = (unsigned long) p;
1363 if (get_user(node, nodes + j + chunk_start))
1367 if (node < 0 || node >= MAX_NUMNODES)
1370 if (!node_state(node, N_MEMORY))
1374 if (!node_isset(node, task_nodes))
1380 /* End marker for this chunk */
1381 pm[chunk_nr_pages].node = MAX_NUMNODES;
1383 /* Migrate this chunk */
1384 err = do_move_page_to_node_array(mm, pm,
1385 flags & MPOL_MF_MOVE_ALL);
1389 /* Return status information */
1390 for (j = 0; j < chunk_nr_pages; j++)
1391 if (put_user(pm[j].status, status + j + chunk_start)) {
1399 free_page((unsigned long)pm);
1405 * Determine the nodes of an array of pages and store it in an array of status.
1407 static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages,
1408 const void __user **pages, int *status)
1412 down_read(&mm->mmap_sem);
1414 for (i = 0; i < nr_pages; i++) {
1415 unsigned long addr = (unsigned long)(*pages);
1416 struct vm_area_struct *vma;
1420 vma = find_vma(mm, addr);
1421 if (!vma || addr < vma->vm_start)
1424 /* FOLL_DUMP to ignore special (like zero) pages */
1425 page = follow_page(vma, addr, FOLL_DUMP);
1427 err = PTR_ERR(page);
1431 err = page ? page_to_nid(page) : -ENOENT;
1439 up_read(&mm->mmap_sem);
1443 * Determine the nodes of a user array of pages and store it in
1444 * a user array of status.
1446 static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages,
1447 const void __user * __user *pages,
1450 #define DO_PAGES_STAT_CHUNK_NR 16
1451 const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR];
1452 int chunk_status[DO_PAGES_STAT_CHUNK_NR];
1455 unsigned long chunk_nr;
1457 chunk_nr = nr_pages;
1458 if (chunk_nr > DO_PAGES_STAT_CHUNK_NR)
1459 chunk_nr = DO_PAGES_STAT_CHUNK_NR;
1461 if (copy_from_user(chunk_pages, pages, chunk_nr * sizeof(*chunk_pages)))
1464 do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status);
1466 if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status)))
1471 nr_pages -= chunk_nr;
1473 return nr_pages ? -EFAULT : 0;
1477 * Move a list of pages in the address space of the currently executing
1480 SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages,
1481 const void __user * __user *, pages,
1482 const int __user *, nodes,
1483 int __user *, status, int, flags)
1485 const struct cred *cred = current_cred(), *tcred;
1486 struct task_struct *task;
1487 struct mm_struct *mm;
1489 nodemask_t task_nodes;
1492 if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
1495 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1498 /* Find the mm_struct */
1500 task = pid ? find_task_by_vpid(pid) : current;
1505 get_task_struct(task);
1508 * Check if this process has the right to modify the specified
1509 * process. The right exists if the process has administrative
1510 * capabilities, superuser privileges or the same
1511 * userid as the target process.
1513 tcred = __task_cred(task);
1514 if (!uid_eq(cred->euid, tcred->suid) && !uid_eq(cred->euid, tcred->uid) &&
1515 !uid_eq(cred->uid, tcred->suid) && !uid_eq(cred->uid, tcred->uid) &&
1516 !capable(CAP_SYS_NICE)) {
1523 err = security_task_movememory(task);
1527 task_nodes = cpuset_mems_allowed(task);
1528 mm = get_task_mm(task);
1529 put_task_struct(task);
1535 err = do_pages_move(mm, task_nodes, nr_pages, pages,
1536 nodes, status, flags);
1538 err = do_pages_stat(mm, nr_pages, pages, status);
1544 put_task_struct(task);
1548 #ifdef CONFIG_NUMA_BALANCING
1550 * Returns true if this is a safe migration target node for misplaced NUMA
1551 * pages. Currently it only checks the watermarks which crude
1553 static bool migrate_balanced_pgdat(struct pglist_data *pgdat,
1554 unsigned long nr_migrate_pages)
1557 for (z = pgdat->nr_zones - 1; z >= 0; z--) {
1558 struct zone *zone = pgdat->node_zones + z;
1560 if (!populated_zone(zone))
1563 if (!zone_reclaimable(zone))
1566 /* Avoid waking kswapd by allocating pages_to_migrate pages. */
1567 if (!zone_watermark_ok(zone, 0,
1568 high_wmark_pages(zone) +
1577 static struct page *alloc_misplaced_dst_page(struct page *page,
1581 int nid = (int) data;
1582 struct page *newpage;
1584 newpage = __alloc_pages_node(nid,
1585 (GFP_HIGHUSER_MOVABLE |
1586 __GFP_THISNODE | __GFP_NOMEMALLOC |
1587 __GFP_NORETRY | __GFP_NOWARN) &
1594 * page migration rate limiting control.
1595 * Do not migrate more than @pages_to_migrate in a @migrate_interval_millisecs
1596 * window of time. Default here says do not migrate more than 1280M per second.
1598 static unsigned int migrate_interval_millisecs __read_mostly = 100;
1599 static unsigned int ratelimit_pages __read_mostly = 128 << (20 - PAGE_SHIFT);
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_PAGE(compound_order(page) && !PageTransHuge(page), 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);
1678 * Attempt to migrate a misplaced page to the specified destination
1679 * node. Caller is expected to have an elevated reference count on
1680 * the page that will be dropped by this function before returning.
1682 int migrate_misplaced_page(struct page *page, struct vm_area_struct *vma,
1685 pg_data_t *pgdat = NODE_DATA(node);
1688 LIST_HEAD(migratepages);
1691 * Don't migrate file pages that are mapped in multiple processes
1692 * with execute permissions as they are probably shared libraries.
1694 if (page_mapcount(page) != 1 && page_is_file_cache(page) &&
1695 (vma->vm_flags & VM_EXEC))
1699 * Rate-limit the amount of data that is being migrated to a node.
1700 * Optimal placement is no good if the memory bus is saturated and
1701 * all the time is being spent migrating!
1703 if (numamigrate_update_ratelimit(pgdat, 1))
1706 isolated = numamigrate_isolate_page(pgdat, page);
1710 list_add(&page->lru, &migratepages);
1711 nr_remaining = migrate_pages(&migratepages, alloc_misplaced_dst_page,
1712 NULL, node, MIGRATE_ASYNC,
1715 if (!list_empty(&migratepages)) {
1716 list_del(&page->lru);
1717 dec_zone_page_state(page, NR_ISOLATED_ANON +
1718 page_is_file_cache(page));
1719 putback_lru_page(page);
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 int page_lru = page_is_file_cache(page);
1749 unsigned long mmun_start = address & HPAGE_PMD_MASK;
1750 unsigned long mmun_end = mmun_start + HPAGE_PMD_SIZE;
1754 * Rate-limit the amount of data that is being migrated to a node.
1755 * Optimal placement is no good if the memory bus is saturated and
1756 * all the time is being spent migrating!
1758 if (numamigrate_update_ratelimit(pgdat, HPAGE_PMD_NR))
1761 new_page = alloc_pages_node(node,
1762 (GFP_TRANSHUGE | __GFP_THISNODE) & ~__GFP_RECLAIM,
1766 prep_transhuge_page(new_page);
1768 isolated = numamigrate_isolate_page(pgdat, page);
1774 if (mm_tlb_flush_pending(mm))
1775 flush_tlb_range(vma, mmun_start, mmun_end);
1777 /* Prepare a page as a migration target */
1778 __SetPageLocked(new_page);
1779 SetPageSwapBacked(new_page);
1781 /* anon mapping, we can simply copy page->mapping to the new page: */
1782 new_page->mapping = page->mapping;
1783 new_page->index = page->index;
1784 migrate_page_copy(new_page, page);
1785 WARN_ON(PageLRU(new_page));
1787 /* Recheck the target PMD */
1788 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1789 ptl = pmd_lock(mm, pmd);
1790 if (unlikely(!pmd_same(*pmd, entry) || page_count(page) != 2)) {
1793 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1795 /* Reverse changes made by migrate_page_copy() */
1796 if (TestClearPageActive(new_page))
1797 SetPageActive(page);
1798 if (TestClearPageUnevictable(new_page))
1799 SetPageUnevictable(page);
1801 unlock_page(new_page);
1802 put_page(new_page); /* Free it */
1804 /* Retake the callers reference and putback on LRU */
1806 putback_lru_page(page);
1807 mod_zone_page_state(page_zone(page),
1808 NR_ISOLATED_ANON + page_lru, -HPAGE_PMD_NR);
1814 entry = mk_pmd(new_page, vma->vm_page_prot);
1815 entry = pmd_mkhuge(entry);
1816 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1819 * Clear the old entry under pagetable lock and establish the new PTE.
1820 * Any parallel GUP will either observe the old page blocking on the
1821 * page lock, block on the page table lock or observe the new page.
1822 * The SetPageUptodate on the new page and page_add_new_anon_rmap
1823 * guarantee the copy is visible before the pagetable update.
1825 flush_cache_range(vma, mmun_start, mmun_end);
1826 page_add_anon_rmap(new_page, vma, mmun_start, true);
1827 pmdp_huge_clear_flush_notify(vma, mmun_start, pmd);
1828 set_pmd_at(mm, mmun_start, pmd, entry);
1829 flush_tlb_range(vma, mmun_start, mmun_end);
1830 update_mmu_cache_pmd(vma, address, &entry);
1832 if (page_count(page) != 2) {
1833 set_pmd_at(mm, mmun_start, pmd, orig_entry);
1834 flush_tlb_range(vma, mmun_start, mmun_end);
1835 mmu_notifier_invalidate_range(mm, mmun_start, mmun_end);
1836 update_mmu_cache_pmd(vma, address, &entry);
1837 page_remove_rmap(new_page, true);
1841 mlock_migrate_page(new_page, page);
1842 set_page_memcg(new_page, page_memcg(page));
1843 set_page_memcg(page, NULL);
1844 page_remove_rmap(page, true);
1847 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1849 /* Take an "isolate" reference and put new page on the LRU. */
1851 putback_lru_page(new_page);
1853 unlock_page(new_page);
1855 put_page(page); /* Drop the rmap reference */
1856 put_page(page); /* Drop the LRU isolation reference */
1858 count_vm_events(PGMIGRATE_SUCCESS, HPAGE_PMD_NR);
1859 count_vm_numa_events(NUMA_PAGE_MIGRATE, HPAGE_PMD_NR);
1861 mod_zone_page_state(page_zone(page),
1862 NR_ISOLATED_ANON + page_lru,
1867 count_vm_events(PGMIGRATE_FAIL, HPAGE_PMD_NR);
1869 ptl = pmd_lock(mm, pmd);
1870 if (pmd_same(*pmd, entry)) {
1871 entry = pmd_modify(entry, vma->vm_page_prot);
1872 set_pmd_at(mm, mmun_start, pmd, entry);
1873 update_mmu_cache_pmd(vma, address, &entry);
1882 #endif /* CONFIG_NUMA_BALANCING */
1884 #endif /* CONFIG_NUMA */