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 newpage->index = page->index;
330 newpage->mapping = page->mapping;
331 if (PageSwapBacked(page))
332 SetPageSwapBacked(newpage);
334 mem_cgroup_migrate(page, newpage);
336 return MIGRATEPAGE_SUCCESS;
339 oldzone = page_zone(page);
340 newzone = page_zone(newpage);
342 spin_lock_irq(&mapping->tree_lock);
344 pslot = radix_tree_lookup_slot(&mapping->page_tree,
347 expected_count += 1 + page_has_private(page);
348 if (page_count(page) != expected_count ||
349 radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) {
350 spin_unlock_irq(&mapping->tree_lock);
354 if (!page_freeze_refs(page, expected_count)) {
355 spin_unlock_irq(&mapping->tree_lock);
360 * In the async migration case of moving a page with buffers, lock the
361 * buffers using trylock before the mapping is moved. If the mapping
362 * was moved, we later failed to lock the buffers and could not move
363 * the mapping back due to an elevated page count, we would have to
364 * block waiting on other references to be dropped.
366 if (mode == MIGRATE_ASYNC && head &&
367 !buffer_migrate_lock_buffers(head, mode)) {
368 page_unfreeze_refs(page, expected_count);
369 spin_unlock_irq(&mapping->tree_lock);
374 * Now we know that no one else is looking at the page:
375 * no turning back from here.
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 mem_cgroup_migrate(page, newpage);
433 return MIGRATEPAGE_SUCCESS;
437 * The expected number of remaining references is the same as that
438 * of migrate_page_move_mapping().
440 int migrate_huge_page_move_mapping(struct address_space *mapping,
441 struct page *newpage, struct page *page)
446 spin_lock_irq(&mapping->tree_lock);
448 pslot = radix_tree_lookup_slot(&mapping->page_tree,
451 expected_count = 2 + page_has_private(page);
452 if (page_count(page) != expected_count ||
453 radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) {
454 spin_unlock_irq(&mapping->tree_lock);
458 if (!page_freeze_refs(page, expected_count)) {
459 spin_unlock_irq(&mapping->tree_lock);
463 newpage->index = page->index;
464 newpage->mapping = page->mapping;
468 radix_tree_replace_slot(pslot, newpage);
470 page_unfreeze_refs(page, expected_count - 1);
472 spin_unlock_irq(&mapping->tree_lock);
474 mem_cgroup_migrate(page, newpage);
476 return MIGRATEPAGE_SUCCESS;
480 * Gigantic pages are so large that we do not guarantee that page++ pointer
481 * arithmetic will work across the entire page. We need something more
484 static void __copy_gigantic_page(struct page *dst, struct page *src,
488 struct page *dst_base = dst;
489 struct page *src_base = src;
491 for (i = 0; i < nr_pages; ) {
493 copy_highpage(dst, src);
496 dst = mem_map_next(dst, dst_base, i);
497 src = mem_map_next(src, src_base, i);
501 static void copy_huge_page(struct page *dst, struct page *src)
508 struct hstate *h = page_hstate(src);
509 nr_pages = pages_per_huge_page(h);
511 if (unlikely(nr_pages > MAX_ORDER_NR_PAGES)) {
512 __copy_gigantic_page(dst, src, nr_pages);
517 BUG_ON(!PageTransHuge(src));
518 nr_pages = hpage_nr_pages(src);
521 for (i = 0; i < nr_pages; i++) {
523 copy_highpage(dst + i, src + i);
528 * Copy the page to its new location
530 void migrate_page_copy(struct page *newpage, struct page *page)
534 if (PageHuge(page) || PageTransHuge(page))
535 copy_huge_page(newpage, page);
537 copy_highpage(newpage, page);
540 SetPageError(newpage);
541 if (PageReferenced(page))
542 SetPageReferenced(newpage);
543 if (PageUptodate(page))
544 SetPageUptodate(newpage);
545 if (TestClearPageActive(page)) {
546 VM_BUG_ON_PAGE(PageUnevictable(page), page);
547 SetPageActive(newpage);
548 } else if (TestClearPageUnevictable(page))
549 SetPageUnevictable(newpage);
550 if (PageChecked(page))
551 SetPageChecked(newpage);
552 if (PageMappedToDisk(page))
553 SetPageMappedToDisk(newpage);
555 /* Move dirty on pages not done by migrate_page_move_mapping() */
557 SetPageDirty(newpage);
559 if (page_is_young(page))
560 set_page_young(newpage);
561 if (page_is_idle(page))
562 set_page_idle(newpage);
565 * Copy NUMA information to the new page, to prevent over-eager
566 * future migrations of this same page.
568 cpupid = page_cpupid_xchg_last(page, -1);
569 page_cpupid_xchg_last(newpage, cpupid);
571 ksm_migrate_page(newpage, page);
573 * Please do not reorder this without considering how mm/ksm.c's
574 * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache().
576 if (PageSwapCache(page))
577 ClearPageSwapCache(page);
578 ClearPagePrivate(page);
579 set_page_private(page, 0);
582 * If any waiters have accumulated on the new page then
585 if (PageWriteback(newpage))
586 end_page_writeback(newpage);
588 copy_page_owner(page, newpage);
591 /************************************************************
592 * Migration functions
593 ***********************************************************/
596 * Common logic to directly migrate a single page suitable for
597 * pages that do not use PagePrivate/PagePrivate2.
599 * Pages are locked upon entry and exit.
601 int migrate_page(struct address_space *mapping,
602 struct page *newpage, struct page *page,
603 enum migrate_mode mode)
607 BUG_ON(PageWriteback(page)); /* Writeback must be complete */
609 rc = migrate_page_move_mapping(mapping, newpage, page, NULL, mode, 0);
611 if (rc != MIGRATEPAGE_SUCCESS)
614 migrate_page_copy(newpage, page);
615 return MIGRATEPAGE_SUCCESS;
617 EXPORT_SYMBOL(migrate_page);
621 * Migration function for pages with buffers. This function can only be used
622 * if the underlying filesystem guarantees that no other references to "page"
625 int buffer_migrate_page(struct address_space *mapping,
626 struct page *newpage, struct page *page, enum migrate_mode mode)
628 struct buffer_head *bh, *head;
631 if (!page_has_buffers(page))
632 return migrate_page(mapping, newpage, page, mode);
634 head = page_buffers(page);
636 rc = migrate_page_move_mapping(mapping, newpage, page, head, mode, 0);
638 if (rc != MIGRATEPAGE_SUCCESS)
642 * In the async case, migrate_page_move_mapping locked the buffers
643 * with an IRQ-safe spinlock held. In the sync case, the buffers
644 * need to be locked now
646 if (mode != MIGRATE_ASYNC)
647 BUG_ON(!buffer_migrate_lock_buffers(head, mode));
649 ClearPagePrivate(page);
650 set_page_private(newpage, page_private(page));
651 set_page_private(page, 0);
657 set_bh_page(bh, newpage, bh_offset(bh));
658 bh = bh->b_this_page;
660 } while (bh != head);
662 SetPagePrivate(newpage);
664 migrate_page_copy(newpage, page);
670 bh = bh->b_this_page;
672 } while (bh != head);
674 return MIGRATEPAGE_SUCCESS;
676 EXPORT_SYMBOL(buffer_migrate_page);
680 * Writeback a page to clean the dirty state
682 static int writeout(struct address_space *mapping, struct page *page)
684 struct writeback_control wbc = {
685 .sync_mode = WB_SYNC_NONE,
688 .range_end = LLONG_MAX,
693 if (!mapping->a_ops->writepage)
694 /* No write method for the address space */
697 if (!clear_page_dirty_for_io(page))
698 /* Someone else already triggered a write */
702 * A dirty page may imply that the underlying filesystem has
703 * the page on some queue. So the page must be clean for
704 * migration. Writeout may mean we loose the lock and the
705 * page state is no longer what we checked for earlier.
706 * At this point we know that the migration attempt cannot
709 remove_migration_ptes(page, page);
711 rc = mapping->a_ops->writepage(page, &wbc);
713 if (rc != AOP_WRITEPAGE_ACTIVATE)
714 /* unlocked. Relock */
717 return (rc < 0) ? -EIO : -EAGAIN;
721 * Default handling if a filesystem does not provide a migration function.
723 static int fallback_migrate_page(struct address_space *mapping,
724 struct page *newpage, struct page *page, enum migrate_mode mode)
726 if (PageDirty(page)) {
727 /* Only writeback pages in full synchronous migration */
728 if (mode != MIGRATE_SYNC)
730 return writeout(mapping, page);
734 * Buffers may be managed in a filesystem specific way.
735 * We must have no buffers or drop them.
737 if (page_has_private(page) &&
738 !try_to_release_page(page, GFP_KERNEL))
741 return migrate_page(mapping, newpage, page, mode);
745 * Move a page to a newly allocated page
746 * The page is locked and all ptes have been successfully removed.
748 * The new page will have replaced the old page if this function
753 * MIGRATEPAGE_SUCCESS - success
755 static int move_to_new_page(struct page *newpage, struct page *page,
756 enum migrate_mode mode)
758 struct address_space *mapping;
761 VM_BUG_ON_PAGE(!PageLocked(page), page);
762 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
764 mapping = page_mapping(page);
766 rc = migrate_page(mapping, newpage, page, mode);
767 else if (mapping->a_ops->migratepage)
769 * Most pages have a mapping and most filesystems provide a
770 * migratepage callback. Anonymous pages are part of swap
771 * space which also has its own migratepage callback. This
772 * is the most common path for page migration.
774 rc = mapping->a_ops->migratepage(mapping, newpage, page, mode);
776 rc = fallback_migrate_page(mapping, newpage, page, mode);
779 * When successful, old pagecache page->mapping must be cleared before
780 * page is freed; but stats require that PageAnon be left as PageAnon.
782 if (rc == MIGRATEPAGE_SUCCESS) {
784 page->mapping = NULL;
789 static int __unmap_and_move(struct page *page, struct page *newpage,
790 int force, enum migrate_mode mode)
793 int page_was_mapped = 0;
794 struct anon_vma *anon_vma = NULL;
796 if (!trylock_page(page)) {
797 if (!force || mode == MIGRATE_ASYNC)
801 * It's not safe for direct compaction to call lock_page.
802 * For example, during page readahead pages are added locked
803 * to the LRU. Later, when the IO completes the pages are
804 * marked uptodate and unlocked. However, the queueing
805 * could be merging multiple pages for one bio (e.g.
806 * mpage_readpages). If an allocation happens for the
807 * second or third page, the process can end up locking
808 * the same page twice and deadlocking. Rather than
809 * trying to be clever about what pages can be locked,
810 * avoid the use of lock_page for direct compaction
813 if (current->flags & PF_MEMALLOC)
819 if (PageWriteback(page)) {
821 * Only in the case of a full synchronous migration is it
822 * necessary to wait for PageWriteback. In the async case,
823 * the retry loop is too short and in the sync-light case,
824 * the overhead of stalling is too much
826 if (mode != MIGRATE_SYNC) {
832 wait_on_page_writeback(page);
836 * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
837 * we cannot notice that anon_vma is freed while we migrates a page.
838 * This get_anon_vma() delays freeing anon_vma pointer until the end
839 * of migration. File cache pages are no problem because of page_lock()
840 * File Caches may use write_page() or lock_page() in migration, then,
841 * just care Anon page here.
843 * Only page_get_anon_vma() understands the subtleties of
844 * getting a hold on an anon_vma from outside one of its mms.
845 * But if we cannot get anon_vma, then we won't need it anyway,
846 * because that implies that the anon page is no longer mapped
847 * (and cannot be remapped so long as we hold the page lock).
849 if (PageAnon(page) && !PageKsm(page))
850 anon_vma = page_get_anon_vma(page);
853 * Block others from accessing the new page when we get around to
854 * establishing additional references. We are usually the only one
855 * holding a reference to newpage at this point. We used to have a BUG
856 * here if trylock_page(newpage) fails, but would like to allow for
857 * cases where there might be a race with the previous use of newpage.
858 * This is much like races on refcount of oldpage: just don't BUG().
860 if (unlikely(!trylock_page(newpage)))
863 if (unlikely(isolated_balloon_page(page))) {
865 * A ballooned page does not need any special attention from
866 * physical to virtual reverse mapping procedures.
867 * Skip any attempt to unmap PTEs or to remap swap cache,
868 * in order to avoid burning cycles at rmap level, and perform
869 * the page migration right away (proteced by page lock).
871 rc = balloon_page_migrate(newpage, page, mode);
872 goto out_unlock_both;
876 * Corner case handling:
877 * 1. When a new swap-cache page is read into, it is added to the LRU
878 * and treated as swapcache but it has no rmap yet.
879 * Calling try_to_unmap() against a page->mapping==NULL page will
880 * trigger a BUG. So handle it here.
881 * 2. An orphaned page (see truncate_complete_page) might have
882 * fs-private metadata. The page can be picked up due to memory
883 * offlining. Everywhere else except page reclaim, the page is
884 * invisible to the vm, so the page can not be migrated. So try to
885 * free the metadata, so the page can be freed.
887 if (!page->mapping) {
888 VM_BUG_ON_PAGE(PageAnon(page), page);
889 if (page_has_private(page)) {
890 try_to_free_buffers(page);
891 goto out_unlock_both;
893 } else if (page_mapped(page)) {
894 /* Establish migration ptes */
895 VM_BUG_ON_PAGE(PageAnon(page) && !PageKsm(page) && !anon_vma,
898 TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
902 if (!page_mapped(page))
903 rc = move_to_new_page(newpage, page, mode);
906 remove_migration_ptes(page,
907 rc == MIGRATEPAGE_SUCCESS ? newpage : page);
910 unlock_page(newpage);
912 /* Drop an anon_vma reference if we took one */
914 put_anon_vma(anon_vma);
921 * gcc 4.7 and 4.8 on arm get an ICEs when inlining unmap_and_move(). Work
924 #if (GCC_VERSION >= 40700 && GCC_VERSION < 40900) && defined(CONFIG_ARM)
925 #define ICE_noinline noinline
931 * Obtain the lock on page, remove all ptes and migrate the page
932 * to the newly allocated page in newpage.
934 static ICE_noinline int unmap_and_move(new_page_t get_new_page,
935 free_page_t put_new_page,
936 unsigned long private, struct page *page,
937 int force, enum migrate_mode mode,
938 enum migrate_reason reason)
940 int rc = MIGRATEPAGE_SUCCESS;
942 struct page *newpage;
944 newpage = get_new_page(page, private, &result);
948 if (page_count(page) == 1) {
949 /* page was freed from under us. So we are done. */
953 if (unlikely(PageTransHuge(page))) {
955 rc = split_huge_page(page);
961 rc = __unmap_and_move(page, newpage, force, mode);
962 if (rc == MIGRATEPAGE_SUCCESS) {
964 set_page_owner_migrate_reason(newpage, reason);
970 * A page that has been migrated has all references
971 * removed and will be freed. A page that has not been
972 * migrated will have kepts its references and be
975 list_del(&page->lru);
976 dec_zone_page_state(page, NR_ISOLATED_ANON +
977 page_is_file_cache(page));
978 /* Soft-offlined page shouldn't go through lru cache list */
979 if (reason == MR_MEMORY_FAILURE) {
981 if (!test_set_page_hwpoison(page))
982 num_poisoned_pages_inc();
984 putback_lru_page(page);
988 * If migration was not successful and there's a freeing callback, use
989 * it. Otherwise, putback_lru_page() will drop the reference grabbed
993 put_new_page(newpage, private);
994 else if (unlikely(__is_movable_balloon_page(newpage))) {
995 /* drop our reference, page already in the balloon */
998 putback_lru_page(newpage);
1004 *result = page_to_nid(newpage);
1010 * Counterpart of unmap_and_move_page() for hugepage migration.
1012 * This function doesn't wait the completion of hugepage I/O
1013 * because there is no race between I/O and migration for hugepage.
1014 * Note that currently hugepage I/O occurs only in direct I/O
1015 * where no lock is held and PG_writeback is irrelevant,
1016 * and writeback status of all subpages are counted in the reference
1017 * count of the head page (i.e. if all subpages of a 2MB hugepage are
1018 * under direct I/O, the reference of the head page is 512 and a bit more.)
1019 * This means that when we try to migrate hugepage whose subpages are
1020 * doing direct I/O, some references remain after try_to_unmap() and
1021 * hugepage migration fails without data corruption.
1023 * There is also no race when direct I/O is issued on the page under migration,
1024 * because then pte is replaced with migration swap entry and direct I/O code
1025 * will wait in the page fault for migration to complete.
1027 static int unmap_and_move_huge_page(new_page_t get_new_page,
1028 free_page_t put_new_page, unsigned long private,
1029 struct page *hpage, int force,
1030 enum migrate_mode mode, int reason)
1034 int page_was_mapped = 0;
1035 struct page *new_hpage;
1036 struct anon_vma *anon_vma = NULL;
1039 * Movability of hugepages depends on architectures and hugepage size.
1040 * This check is necessary because some callers of hugepage migration
1041 * like soft offline and memory hotremove don't walk through page
1042 * tables or check whether the hugepage is pmd-based or not before
1043 * kicking migration.
1045 if (!hugepage_migration_supported(page_hstate(hpage))) {
1046 putback_active_hugepage(hpage);
1050 new_hpage = get_new_page(hpage, private, &result);
1054 if (!trylock_page(hpage)) {
1055 if (!force || mode != MIGRATE_SYNC)
1060 if (PageAnon(hpage))
1061 anon_vma = page_get_anon_vma(hpage);
1063 if (unlikely(!trylock_page(new_hpage)))
1066 if (page_mapped(hpage)) {
1068 TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
1069 page_was_mapped = 1;
1072 if (!page_mapped(hpage))
1073 rc = move_to_new_page(new_hpage, hpage, mode);
1075 if (page_was_mapped)
1076 remove_migration_ptes(hpage,
1077 rc == MIGRATEPAGE_SUCCESS ? new_hpage : hpage);
1079 unlock_page(new_hpage);
1083 put_anon_vma(anon_vma);
1085 if (rc == MIGRATEPAGE_SUCCESS) {
1086 hugetlb_cgroup_migrate(hpage, new_hpage);
1087 put_new_page = NULL;
1088 set_page_owner_migrate_reason(new_hpage, reason);
1094 putback_active_hugepage(hpage);
1097 * If migration was not successful and there's a freeing callback, use
1098 * it. Otherwise, put_page() will drop the reference grabbed during
1102 put_new_page(new_hpage, private);
1104 putback_active_hugepage(new_hpage);
1110 *result = page_to_nid(new_hpage);
1116 * migrate_pages - migrate the pages specified in a list, to the free pages
1117 * supplied as the target for the page migration
1119 * @from: The list of pages to be migrated.
1120 * @get_new_page: The function used to allocate free pages to be used
1121 * as the target of the page migration.
1122 * @put_new_page: The function used to free target pages if migration
1123 * fails, or NULL if no special handling is necessary.
1124 * @private: Private data to be passed on to get_new_page()
1125 * @mode: The migration mode that specifies the constraints for
1126 * page migration, if any.
1127 * @reason: The reason for page migration.
1129 * The function returns after 10 attempts or if no pages are movable any more
1130 * because the list has become empty or no retryable pages exist any more.
1131 * The caller should call putback_movable_pages() to return pages to the LRU
1132 * or free list only if ret != 0.
1134 * Returns the number of pages that were not migrated, or an error code.
1136 int migrate_pages(struct list_head *from, new_page_t get_new_page,
1137 free_page_t put_new_page, unsigned long private,
1138 enum migrate_mode mode, int reason)
1142 int nr_succeeded = 0;
1146 int swapwrite = current->flags & PF_SWAPWRITE;
1150 current->flags |= PF_SWAPWRITE;
1152 for(pass = 0; pass < 10 && retry; pass++) {
1155 list_for_each_entry_safe(page, page2, from, lru) {
1159 rc = unmap_and_move_huge_page(get_new_page,
1160 put_new_page, private, page,
1161 pass > 2, mode, reason);
1163 rc = unmap_and_move(get_new_page, put_new_page,
1164 private, page, pass > 2, mode,
1173 case MIGRATEPAGE_SUCCESS:
1178 * Permanent failure (-EBUSY, -ENOSYS, etc.):
1179 * unlike -EAGAIN case, the failed page is
1180 * removed from migration page list and not
1181 * retried in the next outer loop.
1192 count_vm_events(PGMIGRATE_SUCCESS, nr_succeeded);
1194 count_vm_events(PGMIGRATE_FAIL, nr_failed);
1195 trace_mm_migrate_pages(nr_succeeded, nr_failed, mode, reason);
1198 current->flags &= ~PF_SWAPWRITE;
1205 * Move a list of individual pages
1207 struct page_to_node {
1214 static struct page *new_page_node(struct page *p, unsigned long private,
1217 struct page_to_node *pm = (struct page_to_node *)private;
1219 while (pm->node != MAX_NUMNODES && pm->page != p)
1222 if (pm->node == MAX_NUMNODES)
1225 *result = &pm->status;
1228 return alloc_huge_page_node(page_hstate(compound_head(p)),
1231 return __alloc_pages_node(pm->node,
1232 GFP_HIGHUSER_MOVABLE | __GFP_THISNODE, 0);
1236 * Move a set of pages as indicated in the pm array. The addr
1237 * field must be set to the virtual address of the page to be moved
1238 * and the node number must contain a valid target node.
1239 * The pm array ends with node = MAX_NUMNODES.
1241 static int do_move_page_to_node_array(struct mm_struct *mm,
1242 struct page_to_node *pm,
1246 struct page_to_node *pp;
1247 LIST_HEAD(pagelist);
1249 down_read(&mm->mmap_sem);
1252 * Build a list of pages to migrate
1254 for (pp = pm; pp->node != MAX_NUMNODES; pp++) {
1255 struct vm_area_struct *vma;
1259 vma = find_vma(mm, pp->addr);
1260 if (!vma || pp->addr < vma->vm_start || !vma_migratable(vma))
1263 /* FOLL_DUMP to ignore special (like zero) pages */
1264 page = follow_page(vma, pp->addr,
1265 FOLL_GET | FOLL_SPLIT | FOLL_DUMP);
1267 err = PTR_ERR(page);
1276 err = page_to_nid(page);
1278 if (err == pp->node)
1280 * Node already in the right place
1285 if (page_mapcount(page) > 1 &&
1289 if (PageHuge(page)) {
1291 isolate_huge_page(page, &pagelist);
1295 err = isolate_lru_page(page);
1297 list_add_tail(&page->lru, &pagelist);
1298 inc_zone_page_state(page, NR_ISOLATED_ANON +
1299 page_is_file_cache(page));
1303 * Either remove the duplicate refcount from
1304 * isolate_lru_page() or drop the page ref if it was
1313 if (!list_empty(&pagelist)) {
1314 err = migrate_pages(&pagelist, new_page_node, NULL,
1315 (unsigned long)pm, MIGRATE_SYNC, MR_SYSCALL);
1317 putback_movable_pages(&pagelist);
1320 up_read(&mm->mmap_sem);
1325 * Migrate an array of page address onto an array of nodes and fill
1326 * the corresponding array of status.
1328 static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes,
1329 unsigned long nr_pages,
1330 const void __user * __user *pages,
1331 const int __user *nodes,
1332 int __user *status, int flags)
1334 struct page_to_node *pm;
1335 unsigned long chunk_nr_pages;
1336 unsigned long chunk_start;
1340 pm = (struct page_to_node *)__get_free_page(GFP_KERNEL);
1347 * Store a chunk of page_to_node array in a page,
1348 * but keep the last one as a marker
1350 chunk_nr_pages = (PAGE_SIZE / sizeof(struct page_to_node)) - 1;
1352 for (chunk_start = 0;
1353 chunk_start < nr_pages;
1354 chunk_start += chunk_nr_pages) {
1357 if (chunk_start + chunk_nr_pages > nr_pages)
1358 chunk_nr_pages = nr_pages - chunk_start;
1360 /* fill the chunk pm with addrs and nodes from user-space */
1361 for (j = 0; j < chunk_nr_pages; j++) {
1362 const void __user *p;
1366 if (get_user(p, pages + j + chunk_start))
1368 pm[j].addr = (unsigned long) p;
1370 if (get_user(node, nodes + j + chunk_start))
1374 if (node < 0 || node >= MAX_NUMNODES)
1377 if (!node_state(node, N_MEMORY))
1381 if (!node_isset(node, task_nodes))
1387 /* End marker for this chunk */
1388 pm[chunk_nr_pages].node = MAX_NUMNODES;
1390 /* Migrate this chunk */
1391 err = do_move_page_to_node_array(mm, pm,
1392 flags & MPOL_MF_MOVE_ALL);
1396 /* Return status information */
1397 for (j = 0; j < chunk_nr_pages; j++)
1398 if (put_user(pm[j].status, status + j + chunk_start)) {
1406 free_page((unsigned long)pm);
1412 * Determine the nodes of an array of pages and store it in an array of status.
1414 static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages,
1415 const void __user **pages, int *status)
1419 down_read(&mm->mmap_sem);
1421 for (i = 0; i < nr_pages; i++) {
1422 unsigned long addr = (unsigned long)(*pages);
1423 struct vm_area_struct *vma;
1427 vma = find_vma(mm, addr);
1428 if (!vma || addr < vma->vm_start)
1431 /* FOLL_DUMP to ignore special (like zero) pages */
1432 page = follow_page(vma, addr, FOLL_DUMP);
1434 err = PTR_ERR(page);
1438 err = page ? page_to_nid(page) : -ENOENT;
1446 up_read(&mm->mmap_sem);
1450 * Determine the nodes of a user array of pages and store it in
1451 * a user array of status.
1453 static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages,
1454 const void __user * __user *pages,
1457 #define DO_PAGES_STAT_CHUNK_NR 16
1458 const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR];
1459 int chunk_status[DO_PAGES_STAT_CHUNK_NR];
1462 unsigned long chunk_nr;
1464 chunk_nr = nr_pages;
1465 if (chunk_nr > DO_PAGES_STAT_CHUNK_NR)
1466 chunk_nr = DO_PAGES_STAT_CHUNK_NR;
1468 if (copy_from_user(chunk_pages, pages, chunk_nr * sizeof(*chunk_pages)))
1471 do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status);
1473 if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status)))
1478 nr_pages -= chunk_nr;
1480 return nr_pages ? -EFAULT : 0;
1484 * Move a list of pages in the address space of the currently executing
1487 SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages,
1488 const void __user * __user *, pages,
1489 const int __user *, nodes,
1490 int __user *, status, int, flags)
1492 const struct cred *cred = current_cred(), *tcred;
1493 struct task_struct *task;
1494 struct mm_struct *mm;
1496 nodemask_t task_nodes;
1499 if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
1502 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1505 /* Find the mm_struct */
1507 task = pid ? find_task_by_vpid(pid) : current;
1512 get_task_struct(task);
1515 * Check if this process has the right to modify the specified
1516 * process. The right exists if the process has administrative
1517 * capabilities, superuser privileges or the same
1518 * userid as the target process.
1520 tcred = __task_cred(task);
1521 if (!uid_eq(cred->euid, tcred->suid) && !uid_eq(cred->euid, tcred->uid) &&
1522 !uid_eq(cred->uid, tcred->suid) && !uid_eq(cred->uid, tcred->uid) &&
1523 !capable(CAP_SYS_NICE)) {
1530 err = security_task_movememory(task);
1534 task_nodes = cpuset_mems_allowed(task);
1535 mm = get_task_mm(task);
1536 put_task_struct(task);
1542 err = do_pages_move(mm, task_nodes, nr_pages, pages,
1543 nodes, status, flags);
1545 err = do_pages_stat(mm, nr_pages, pages, status);
1551 put_task_struct(task);
1555 #ifdef CONFIG_NUMA_BALANCING
1557 * Returns true if this is a safe migration target node for misplaced NUMA
1558 * pages. Currently it only checks the watermarks which crude
1560 static bool migrate_balanced_pgdat(struct pglist_data *pgdat,
1561 unsigned long nr_migrate_pages)
1564 for (z = pgdat->nr_zones - 1; z >= 0; z--) {
1565 struct zone *zone = pgdat->node_zones + z;
1567 if (!populated_zone(zone))
1570 if (!zone_reclaimable(zone))
1573 /* Avoid waking kswapd by allocating pages_to_migrate pages. */
1574 if (!zone_watermark_ok(zone, 0,
1575 high_wmark_pages(zone) +
1584 static struct page *alloc_misplaced_dst_page(struct page *page,
1588 int nid = (int) data;
1589 struct page *newpage;
1591 newpage = __alloc_pages_node(nid,
1592 (GFP_HIGHUSER_MOVABLE |
1593 __GFP_THISNODE | __GFP_NOMEMALLOC |
1594 __GFP_NORETRY | __GFP_NOWARN) &
1601 * page migration rate limiting control.
1602 * Do not migrate more than @pages_to_migrate in a @migrate_interval_millisecs
1603 * window of time. Default here says do not migrate more than 1280M per second.
1605 static unsigned int migrate_interval_millisecs __read_mostly = 100;
1606 static unsigned int ratelimit_pages __read_mostly = 128 << (20 - PAGE_SHIFT);
1608 /* Returns true if the node is migrate rate-limited after the update */
1609 static bool numamigrate_update_ratelimit(pg_data_t *pgdat,
1610 unsigned long nr_pages)
1613 * Rate-limit the amount of data that is being migrated to a node.
1614 * Optimal placement is no good if the memory bus is saturated and
1615 * all the time is being spent migrating!
1617 if (time_after(jiffies, pgdat->numabalancing_migrate_next_window)) {
1618 spin_lock(&pgdat->numabalancing_migrate_lock);
1619 pgdat->numabalancing_migrate_nr_pages = 0;
1620 pgdat->numabalancing_migrate_next_window = jiffies +
1621 msecs_to_jiffies(migrate_interval_millisecs);
1622 spin_unlock(&pgdat->numabalancing_migrate_lock);
1624 if (pgdat->numabalancing_migrate_nr_pages > ratelimit_pages) {
1625 trace_mm_numa_migrate_ratelimit(current, pgdat->node_id,
1631 * This is an unlocked non-atomic update so errors are possible.
1632 * The consequences are failing to migrate when we potentiall should
1633 * have which is not severe enough to warrant locking. If it is ever
1634 * a problem, it can be converted to a per-cpu counter.
1636 pgdat->numabalancing_migrate_nr_pages += nr_pages;
1640 static int numamigrate_isolate_page(pg_data_t *pgdat, struct page *page)
1644 VM_BUG_ON_PAGE(compound_order(page) && !PageTransHuge(page), page);
1646 /* Avoid migrating to a node that is nearly full */
1647 if (!migrate_balanced_pgdat(pgdat, 1UL << compound_order(page)))
1650 if (isolate_lru_page(page))
1654 * migrate_misplaced_transhuge_page() skips page migration's usual
1655 * check on page_count(), so we must do it here, now that the page
1656 * has been isolated: a GUP pin, or any other pin, prevents migration.
1657 * The expected page count is 3: 1 for page's mapcount and 1 for the
1658 * caller's pin and 1 for the reference taken by isolate_lru_page().
1660 if (PageTransHuge(page) && page_count(page) != 3) {
1661 putback_lru_page(page);
1665 page_lru = page_is_file_cache(page);
1666 mod_zone_page_state(page_zone(page), NR_ISOLATED_ANON + page_lru,
1667 hpage_nr_pages(page));
1670 * Isolating the page has taken another reference, so the
1671 * caller's reference can be safely dropped without the page
1672 * disappearing underneath us during migration.
1678 bool pmd_trans_migrating(pmd_t pmd)
1680 struct page *page = pmd_page(pmd);
1681 return PageLocked(page);
1685 * Attempt to migrate a misplaced page to the specified destination
1686 * node. Caller is expected to have an elevated reference count on
1687 * the page that will be dropped by this function before returning.
1689 int migrate_misplaced_page(struct page *page, struct vm_area_struct *vma,
1692 pg_data_t *pgdat = NODE_DATA(node);
1695 LIST_HEAD(migratepages);
1698 * Don't migrate file pages that are mapped in multiple processes
1699 * with execute permissions as they are probably shared libraries.
1701 if (page_mapcount(page) != 1 && page_is_file_cache(page) &&
1702 (vma->vm_flags & VM_EXEC))
1706 * Rate-limit the amount of data that is being migrated to a node.
1707 * Optimal placement is no good if the memory bus is saturated and
1708 * all the time is being spent migrating!
1710 if (numamigrate_update_ratelimit(pgdat, 1))
1713 isolated = numamigrate_isolate_page(pgdat, page);
1717 list_add(&page->lru, &migratepages);
1718 nr_remaining = migrate_pages(&migratepages, alloc_misplaced_dst_page,
1719 NULL, node, MIGRATE_ASYNC,
1722 if (!list_empty(&migratepages)) {
1723 list_del(&page->lru);
1724 dec_zone_page_state(page, NR_ISOLATED_ANON +
1725 page_is_file_cache(page));
1726 putback_lru_page(page);
1730 count_vm_numa_event(NUMA_PAGE_MIGRATE);
1731 BUG_ON(!list_empty(&migratepages));
1738 #endif /* CONFIG_NUMA_BALANCING */
1740 #if defined(CONFIG_NUMA_BALANCING) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
1742 * Migrates a THP to a given target node. page must be locked and is unlocked
1745 int migrate_misplaced_transhuge_page(struct mm_struct *mm,
1746 struct vm_area_struct *vma,
1747 pmd_t *pmd, pmd_t entry,
1748 unsigned long address,
1749 struct page *page, int node)
1752 pg_data_t *pgdat = NODE_DATA(node);
1754 struct page *new_page = NULL;
1755 int page_lru = page_is_file_cache(page);
1756 unsigned long mmun_start = address & HPAGE_PMD_MASK;
1757 unsigned long mmun_end = mmun_start + HPAGE_PMD_SIZE;
1761 * Rate-limit the amount of data that is being migrated to a node.
1762 * Optimal placement is no good if the memory bus is saturated and
1763 * all the time is being spent migrating!
1765 if (numamigrate_update_ratelimit(pgdat, HPAGE_PMD_NR))
1768 new_page = alloc_pages_node(node,
1769 (GFP_TRANSHUGE | __GFP_THISNODE) & ~__GFP_RECLAIM,
1773 prep_transhuge_page(new_page);
1775 isolated = numamigrate_isolate_page(pgdat, page);
1781 if (mm_tlb_flush_pending(mm))
1782 flush_tlb_range(vma, mmun_start, mmun_end);
1784 /* Prepare a page as a migration target */
1785 __SetPageLocked(new_page);
1786 SetPageSwapBacked(new_page);
1788 /* anon mapping, we can simply copy page->mapping to the new page: */
1789 new_page->mapping = page->mapping;
1790 new_page->index = page->index;
1791 migrate_page_copy(new_page, page);
1792 WARN_ON(PageLRU(new_page));
1794 /* Recheck the target PMD */
1795 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1796 ptl = pmd_lock(mm, pmd);
1797 if (unlikely(!pmd_same(*pmd, entry) || page_count(page) != 2)) {
1800 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1802 /* Reverse changes made by migrate_page_copy() */
1803 if (TestClearPageActive(new_page))
1804 SetPageActive(page);
1805 if (TestClearPageUnevictable(new_page))
1806 SetPageUnevictable(page);
1808 unlock_page(new_page);
1809 put_page(new_page); /* Free it */
1811 /* Retake the callers reference and putback on LRU */
1813 putback_lru_page(page);
1814 mod_zone_page_state(page_zone(page),
1815 NR_ISOLATED_ANON + page_lru, -HPAGE_PMD_NR);
1821 entry = mk_pmd(new_page, vma->vm_page_prot);
1822 entry = pmd_mkhuge(entry);
1823 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1826 * Clear the old entry under pagetable lock and establish the new PTE.
1827 * Any parallel GUP will either observe the old page blocking on the
1828 * page lock, block on the page table lock or observe the new page.
1829 * The SetPageUptodate on the new page and page_add_new_anon_rmap
1830 * guarantee the copy is visible before the pagetable update.
1832 flush_cache_range(vma, mmun_start, mmun_end);
1833 page_add_anon_rmap(new_page, vma, mmun_start, true);
1834 pmdp_huge_clear_flush_notify(vma, mmun_start, pmd);
1835 set_pmd_at(mm, mmun_start, pmd, entry);
1836 flush_tlb_range(vma, mmun_start, mmun_end);
1837 update_mmu_cache_pmd(vma, address, &entry);
1839 if (page_count(page) != 2) {
1840 set_pmd_at(mm, mmun_start, pmd, orig_entry);
1841 flush_tlb_range(vma, mmun_start, mmun_end);
1842 mmu_notifier_invalidate_range(mm, mmun_start, mmun_end);
1843 update_mmu_cache_pmd(vma, address, &entry);
1844 page_remove_rmap(new_page, true);
1848 mlock_migrate_page(new_page, page);
1849 mem_cgroup_migrate(page, new_page);
1850 page_remove_rmap(page, true);
1851 set_page_owner_migrate_reason(new_page, MR_NUMA_MISPLACED);
1854 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1856 /* Take an "isolate" reference and put new page on the LRU. */
1858 putback_lru_page(new_page);
1860 unlock_page(new_page);
1862 put_page(page); /* Drop the rmap reference */
1863 put_page(page); /* Drop the LRU isolation reference */
1865 count_vm_events(PGMIGRATE_SUCCESS, HPAGE_PMD_NR);
1866 count_vm_numa_events(NUMA_PAGE_MIGRATE, HPAGE_PMD_NR);
1868 mod_zone_page_state(page_zone(page),
1869 NR_ISOLATED_ANON + page_lru,
1874 count_vm_events(PGMIGRATE_FAIL, HPAGE_PMD_NR);
1876 ptl = pmd_lock(mm, pmd);
1877 if (pmd_same(*pmd, entry)) {
1878 entry = pmd_modify(entry, vma->vm_page_prot);
1879 set_pmd_at(mm, mmun_start, pmd, entry);
1880 update_mmu_cache_pmd(vma, address, &entry);
1889 #endif /* CONFIG_NUMA_BALANCING */
1891 #endif /* CONFIG_NUMA */