2 * Memory Migration functionality - linux/mm/migration.c
4 * Copyright (C) 2006 Silicon Graphics, Inc., Christoph Lameter
6 * Page migration was first developed in the context of the memory hotplug
7 * project. The main authors of the migration code are:
15 #include <linux/migrate.h>
16 #include <linux/export.h>
17 #include <linux/swap.h>
18 #include <linux/swapops.h>
19 #include <linux/pagemap.h>
20 #include <linux/buffer_head.h>
21 #include <linux/mm_inline.h>
22 #include <linux/nsproxy.h>
23 #include <linux/pagevec.h>
24 #include <linux/ksm.h>
25 #include <linux/rmap.h>
26 #include <linux/topology.h>
27 #include <linux/cpu.h>
28 #include <linux/cpuset.h>
29 #include <linux/writeback.h>
30 #include <linux/mempolicy.h>
31 #include <linux/vmalloc.h>
32 #include <linux/security.h>
33 #include <linux/memcontrol.h>
34 #include <linux/syscalls.h>
35 #include <linux/hugetlb.h>
36 #include <linux/hugetlb_cgroup.h>
37 #include <linux/gfp.h>
39 #include <asm/tlbflush.h>
41 #define CREATE_TRACE_POINTS
42 #include <trace/events/migrate.h>
47 * migrate_prep() needs to be called before we start compiling a list of pages
48 * to be migrated using isolate_lru_page(). If scheduling work on other CPUs is
49 * undesirable, use migrate_prep_local()
51 int migrate_prep(void)
54 * Clear the LRU lists so pages can be isolated.
55 * Note that pages may be moved off the LRU after we have
56 * drained them. Those pages will fail to migrate like other
57 * pages that may be busy.
64 /* Do the necessary work of migrate_prep but not if it involves other CPUs */
65 int migrate_prep_local(void)
73 * Add isolated pages on the list back to the LRU under page lock
74 * to avoid leaking evictable pages back onto unevictable list.
76 void putback_lru_pages(struct list_head *l)
81 list_for_each_entry_safe(page, page2, l, lru) {
83 dec_zone_page_state(page, NR_ISOLATED_ANON +
84 page_is_file_cache(page));
85 putback_lru_page(page);
90 * Restore a potential migration pte to a working pte entry
92 static int remove_migration_pte(struct page *new, struct vm_area_struct *vma,
93 unsigned long addr, void *old)
95 struct mm_struct *mm = vma->vm_mm;
103 if (unlikely(PageHuge(new))) {
104 ptep = huge_pte_offset(mm, addr);
107 ptl = &mm->page_table_lock;
109 pgd = pgd_offset(mm, addr);
110 if (!pgd_present(*pgd))
113 pud = pud_offset(pgd, addr);
114 if (!pud_present(*pud))
117 pmd = pmd_offset(pud, addr);
118 if (pmd_trans_huge(*pmd))
120 if (!pmd_present(*pmd))
123 ptep = pte_offset_map(pmd, addr);
126 * Peek to check is_swap_pte() before taking ptlock? No, we
127 * can race mremap's move_ptes(), which skips anon_vma lock.
130 ptl = pte_lockptr(mm, pmd);
135 if (!is_swap_pte(pte))
138 entry = pte_to_swp_entry(pte);
140 if (!is_migration_entry(entry) ||
141 migration_entry_to_page(entry) != old)
145 pte = pte_mkold(mk_pte(new, vma->vm_page_prot));
146 if (is_write_migration_entry(entry))
147 pte = pte_mkwrite(pte);
148 #ifdef CONFIG_HUGETLB_PAGE
150 pte = pte_mkhuge(pte);
152 flush_cache_page(vma, addr, pte_pfn(pte));
153 set_pte_at(mm, addr, ptep, pte);
157 hugepage_add_anon_rmap(new, vma, addr);
160 } else if (PageAnon(new))
161 page_add_anon_rmap(new, vma, addr);
163 page_add_file_rmap(new);
165 /* No need to invalidate - it was non-present before */
166 update_mmu_cache(vma, addr, ptep);
168 pte_unmap_unlock(ptep, ptl);
174 * Get rid of all migration entries and replace them by
175 * references to the indicated page.
177 static void remove_migration_ptes(struct page *old, struct page *new)
179 rmap_walk(new, remove_migration_pte, old);
183 * Something used the pte of a page under migration. We need to
184 * get to the page and wait until migration is finished.
185 * When we return from this function the fault will be retried.
187 void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
188 unsigned long address)
195 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
197 if (!is_swap_pte(pte))
200 entry = pte_to_swp_entry(pte);
201 if (!is_migration_entry(entry))
204 page = migration_entry_to_page(entry);
207 * Once radix-tree replacement of page migration started, page_count
208 * *must* be zero. And, we don't want to call wait_on_page_locked()
209 * against a page without get_page().
210 * So, we use get_page_unless_zero(), here. Even failed, page fault
213 if (!get_page_unless_zero(page))
215 pte_unmap_unlock(ptep, ptl);
216 wait_on_page_locked(page);
220 pte_unmap_unlock(ptep, ptl);
224 /* Returns true if all buffers are successfully locked */
225 static bool buffer_migrate_lock_buffers(struct buffer_head *head,
226 enum migrate_mode mode)
228 struct buffer_head *bh = head;
230 /* Simple case, sync compaction */
231 if (mode != MIGRATE_ASYNC) {
235 bh = bh->b_this_page;
237 } while (bh != head);
242 /* async case, we cannot block on lock_buffer so use trylock_buffer */
245 if (!trylock_buffer(bh)) {
247 * We failed to lock the buffer and cannot stall in
248 * async migration. Release the taken locks
250 struct buffer_head *failed_bh = bh;
253 while (bh != failed_bh) {
256 bh = bh->b_this_page;
261 bh = bh->b_this_page;
262 } while (bh != head);
266 static inline bool buffer_migrate_lock_buffers(struct buffer_head *head,
267 enum migrate_mode mode)
271 #endif /* CONFIG_BLOCK */
274 * Replace the page in the mapping.
276 * The number of remaining references must be:
277 * 1 for anonymous pages without a mapping
278 * 2 for pages with a mapping
279 * 3 for pages with a mapping and PagePrivate/PagePrivate2 set.
281 static int migrate_page_move_mapping(struct address_space *mapping,
282 struct page *newpage, struct page *page,
283 struct buffer_head *head, enum migrate_mode mode)
285 int expected_count = 0;
289 /* Anonymous page without mapping */
290 if (page_count(page) != 1)
295 spin_lock_irq(&mapping->tree_lock);
297 pslot = radix_tree_lookup_slot(&mapping->page_tree,
300 expected_count = 2 + page_has_private(page);
301 if (page_count(page) != expected_count ||
302 radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) {
303 spin_unlock_irq(&mapping->tree_lock);
307 if (!page_freeze_refs(page, expected_count)) {
308 spin_unlock_irq(&mapping->tree_lock);
313 * In the async migration case of moving a page with buffers, lock the
314 * buffers using trylock before the mapping is moved. If the mapping
315 * was moved, we later failed to lock the buffers and could not move
316 * the mapping back due to an elevated page count, we would have to
317 * block waiting on other references to be dropped.
319 if (mode == MIGRATE_ASYNC && head &&
320 !buffer_migrate_lock_buffers(head, mode)) {
321 page_unfreeze_refs(page, expected_count);
322 spin_unlock_irq(&mapping->tree_lock);
327 * Now we know that no one else is looking at the page.
329 get_page(newpage); /* add cache reference */
330 if (PageSwapCache(page)) {
331 SetPageSwapCache(newpage);
332 set_page_private(newpage, page_private(page));
335 radix_tree_replace_slot(pslot, newpage);
338 * Drop cache reference from old page by unfreezing
339 * to one less reference.
340 * We know this isn't the last reference.
342 page_unfreeze_refs(page, expected_count - 1);
345 * If moved to a different zone then also account
346 * the page for that zone. Other VM counters will be
347 * taken care of when we establish references to the
348 * new page and drop references to the old page.
350 * Note that anonymous pages are accounted for
351 * via NR_FILE_PAGES and NR_ANON_PAGES if they
352 * are mapped to swap space.
354 __dec_zone_page_state(page, NR_FILE_PAGES);
355 __inc_zone_page_state(newpage, NR_FILE_PAGES);
356 if (!PageSwapCache(page) && PageSwapBacked(page)) {
357 __dec_zone_page_state(page, NR_SHMEM);
358 __inc_zone_page_state(newpage, NR_SHMEM);
360 spin_unlock_irq(&mapping->tree_lock);
366 * The expected number of remaining references is the same as that
367 * of migrate_page_move_mapping().
369 int migrate_huge_page_move_mapping(struct address_space *mapping,
370 struct page *newpage, struct page *page)
376 if (page_count(page) != 1)
381 spin_lock_irq(&mapping->tree_lock);
383 pslot = radix_tree_lookup_slot(&mapping->page_tree,
386 expected_count = 2 + page_has_private(page);
387 if (page_count(page) != expected_count ||
388 radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) {
389 spin_unlock_irq(&mapping->tree_lock);
393 if (!page_freeze_refs(page, expected_count)) {
394 spin_unlock_irq(&mapping->tree_lock);
400 radix_tree_replace_slot(pslot, newpage);
402 page_unfreeze_refs(page, expected_count - 1);
404 spin_unlock_irq(&mapping->tree_lock);
409 * Copy the page to its new location
411 void migrate_page_copy(struct page *newpage, struct page *page)
413 if (PageHuge(page) || PageTransHuge(page))
414 copy_huge_page(newpage, page);
416 copy_highpage(newpage, page);
419 SetPageError(newpage);
420 if (PageReferenced(page))
421 SetPageReferenced(newpage);
422 if (PageUptodate(page))
423 SetPageUptodate(newpage);
424 if (TestClearPageActive(page)) {
425 VM_BUG_ON(PageUnevictable(page));
426 SetPageActive(newpage);
427 } else if (TestClearPageUnevictable(page))
428 SetPageUnevictable(newpage);
429 if (PageChecked(page))
430 SetPageChecked(newpage);
431 if (PageMappedToDisk(page))
432 SetPageMappedToDisk(newpage);
434 if (PageDirty(page)) {
435 clear_page_dirty_for_io(page);
437 * Want to mark the page and the radix tree as dirty, and
438 * redo the accounting that clear_page_dirty_for_io undid,
439 * but we can't use set_page_dirty because that function
440 * is actually a signal that all of the page has become dirty.
441 * Whereas only part of our page may be dirty.
443 if (PageSwapBacked(page))
444 SetPageDirty(newpage);
446 __set_page_dirty_nobuffers(newpage);
449 mlock_migrate_page(newpage, page);
450 ksm_migrate_page(newpage, page);
452 ClearPageSwapCache(page);
453 ClearPagePrivate(page);
454 set_page_private(page, 0);
457 * If any waiters have accumulated on the new page then
460 if (PageWriteback(newpage))
461 end_page_writeback(newpage);
464 /************************************************************
465 * Migration functions
466 ***********************************************************/
468 /* Always fail migration. Used for mappings that are not movable */
469 int fail_migrate_page(struct address_space *mapping,
470 struct page *newpage, struct page *page)
474 EXPORT_SYMBOL(fail_migrate_page);
477 * Common logic to directly migrate a single page suitable for
478 * pages that do not use PagePrivate/PagePrivate2.
480 * Pages are locked upon entry and exit.
482 int migrate_page(struct address_space *mapping,
483 struct page *newpage, struct page *page,
484 enum migrate_mode mode)
488 BUG_ON(PageWriteback(page)); /* Writeback must be complete */
490 rc = migrate_page_move_mapping(mapping, newpage, page, NULL, mode);
495 migrate_page_copy(newpage, page);
498 EXPORT_SYMBOL(migrate_page);
502 * Migration function for pages with buffers. This function can only be used
503 * if the underlying filesystem guarantees that no other references to "page"
506 int buffer_migrate_page(struct address_space *mapping,
507 struct page *newpage, struct page *page, enum migrate_mode mode)
509 struct buffer_head *bh, *head;
512 if (!page_has_buffers(page))
513 return migrate_page(mapping, newpage, page, mode);
515 head = page_buffers(page);
517 rc = migrate_page_move_mapping(mapping, newpage, page, head, mode);
523 * In the async case, migrate_page_move_mapping locked the buffers
524 * with an IRQ-safe spinlock held. In the sync case, the buffers
525 * need to be locked now
527 if (mode != MIGRATE_ASYNC)
528 BUG_ON(!buffer_migrate_lock_buffers(head, mode));
530 ClearPagePrivate(page);
531 set_page_private(newpage, page_private(page));
532 set_page_private(page, 0);
538 set_bh_page(bh, newpage, bh_offset(bh));
539 bh = bh->b_this_page;
541 } while (bh != head);
543 SetPagePrivate(newpage);
545 migrate_page_copy(newpage, page);
551 bh = bh->b_this_page;
553 } while (bh != head);
557 EXPORT_SYMBOL(buffer_migrate_page);
561 * Writeback a page to clean the dirty state
563 static int writeout(struct address_space *mapping, struct page *page)
565 struct writeback_control wbc = {
566 .sync_mode = WB_SYNC_NONE,
569 .range_end = LLONG_MAX,
574 if (!mapping->a_ops->writepage)
575 /* No write method for the address space */
578 if (!clear_page_dirty_for_io(page))
579 /* Someone else already triggered a write */
583 * A dirty page may imply that the underlying filesystem has
584 * the page on some queue. So the page must be clean for
585 * migration. Writeout may mean we loose the lock and the
586 * page state is no longer what we checked for earlier.
587 * At this point we know that the migration attempt cannot
590 remove_migration_ptes(page, page);
592 rc = mapping->a_ops->writepage(page, &wbc);
594 if (rc != AOP_WRITEPAGE_ACTIVATE)
595 /* unlocked. Relock */
598 return (rc < 0) ? -EIO : -EAGAIN;
602 * Default handling if a filesystem does not provide a migration function.
604 static int fallback_migrate_page(struct address_space *mapping,
605 struct page *newpage, struct page *page, enum migrate_mode mode)
607 if (PageDirty(page)) {
608 /* Only writeback pages in full synchronous migration */
609 if (mode != MIGRATE_SYNC)
611 return writeout(mapping, page);
615 * Buffers may be managed in a filesystem specific way.
616 * We must have no buffers or drop them.
618 if (page_has_private(page) &&
619 !try_to_release_page(page, GFP_KERNEL))
622 return migrate_page(mapping, newpage, page, mode);
626 * Move a page to a newly allocated page
627 * The page is locked and all ptes have been successfully removed.
629 * The new page will have replaced the old page if this function
636 static int move_to_new_page(struct page *newpage, struct page *page,
637 int remap_swapcache, enum migrate_mode mode)
639 struct address_space *mapping;
643 * Block others from accessing the page when we get around to
644 * establishing additional references. We are the only one
645 * holding a reference to the new page at this point.
647 if (!trylock_page(newpage))
650 /* Prepare mapping for the new page.*/
651 newpage->index = page->index;
652 newpage->mapping = page->mapping;
653 if (PageSwapBacked(page))
654 SetPageSwapBacked(newpage);
656 mapping = page_mapping(page);
658 rc = migrate_page(mapping, newpage, page, mode);
659 else if (mapping->a_ops->migratepage)
661 * Most pages have a mapping and most filesystems provide a
662 * migratepage callback. Anonymous pages are part of swap
663 * space which also has its own migratepage callback. This
664 * is the most common path for page migration.
666 rc = mapping->a_ops->migratepage(mapping,
667 newpage, page, mode);
669 rc = fallback_migrate_page(mapping, newpage, page, mode);
672 newpage->mapping = NULL;
675 remove_migration_ptes(page, newpage);
676 page->mapping = NULL;
679 unlock_page(newpage);
684 static int __unmap_and_move(struct page *page, struct page *newpage,
685 int force, bool offlining, enum migrate_mode mode)
688 int remap_swapcache = 1;
689 struct mem_cgroup *mem;
690 struct anon_vma *anon_vma = NULL;
692 if (!trylock_page(page)) {
693 if (!force || mode == MIGRATE_ASYNC)
697 * It's not safe for direct compaction to call lock_page.
698 * For example, during page readahead pages are added locked
699 * to the LRU. Later, when the IO completes the pages are
700 * marked uptodate and unlocked. However, the queueing
701 * could be merging multiple pages for one bio (e.g.
702 * mpage_readpages). If an allocation happens for the
703 * second or third page, the process can end up locking
704 * the same page twice and deadlocking. Rather than
705 * trying to be clever about what pages can be locked,
706 * avoid the use of lock_page for direct compaction
709 if (current->flags & PF_MEMALLOC)
716 * Only memory hotplug's offline_pages() caller has locked out KSM,
717 * and can safely migrate a KSM page. The other cases have skipped
718 * PageKsm along with PageReserved - but it is only now when we have
719 * the page lock that we can be certain it will not go KSM beneath us
720 * (KSM will not upgrade a page from PageAnon to PageKsm when it sees
721 * its pagecount raised, but only here do we take the page lock which
724 if (PageKsm(page) && !offlining) {
729 /* charge against new page */
730 mem_cgroup_prepare_migration(page, newpage, &mem);
732 if (PageWriteback(page)) {
734 * Only in the case of a full syncronous migration is it
735 * necessary to wait for PageWriteback. In the async case,
736 * the retry loop is too short and in the sync-light case,
737 * the overhead of stalling is too much
739 if (mode != MIGRATE_SYNC) {
745 wait_on_page_writeback(page);
748 * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
749 * we cannot notice that anon_vma is freed while we migrates a page.
750 * This get_anon_vma() delays freeing anon_vma pointer until the end
751 * of migration. File cache pages are no problem because of page_lock()
752 * File Caches may use write_page() or lock_page() in migration, then,
753 * just care Anon page here.
755 if (PageAnon(page)) {
757 * Only page_lock_anon_vma() understands the subtleties of
758 * getting a hold on an anon_vma from outside one of its mms.
760 anon_vma = page_get_anon_vma(page);
765 } else if (PageSwapCache(page)) {
767 * We cannot be sure that the anon_vma of an unmapped
768 * swapcache page is safe to use because we don't
769 * know in advance if the VMA that this page belonged
770 * to still exists. If the VMA and others sharing the
771 * data have been freed, then the anon_vma could
772 * already be invalid.
774 * To avoid this possibility, swapcache pages get
775 * migrated but are not remapped when migration
785 * Corner case handling:
786 * 1. When a new swap-cache page is read into, it is added to the LRU
787 * and treated as swapcache but it has no rmap yet.
788 * Calling try_to_unmap() against a page->mapping==NULL page will
789 * trigger a BUG. So handle it here.
790 * 2. An orphaned page (see truncate_complete_page) might have
791 * fs-private metadata. The page can be picked up due to memory
792 * offlining. Everywhere else except page reclaim, the page is
793 * invisible to the vm, so the page can not be migrated. So try to
794 * free the metadata, so the page can be freed.
796 if (!page->mapping) {
797 VM_BUG_ON(PageAnon(page));
798 if (page_has_private(page)) {
799 try_to_free_buffers(page);
805 /* Establish migration ptes or remove ptes */
806 try_to_unmap(page, TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
809 if (!page_mapped(page))
810 rc = move_to_new_page(newpage, page, remap_swapcache, mode);
812 if (rc && remap_swapcache)
813 remove_migration_ptes(page, page);
815 /* Drop an anon_vma reference if we took one */
817 put_anon_vma(anon_vma);
820 mem_cgroup_end_migration(mem, page, newpage, rc == 0);
828 * Obtain the lock on page, remove all ptes and migrate the page
829 * to the newly allocated page in newpage.
831 static int unmap_and_move(new_page_t get_new_page, unsigned long private,
832 struct page *page, int force, bool offlining,
833 enum migrate_mode mode)
837 struct page *newpage = get_new_page(page, private, &result);
842 if (page_count(page) == 1) {
843 /* page was freed from under us. So we are done. */
847 if (unlikely(PageTransHuge(page)))
848 if (unlikely(split_huge_page(page)))
851 rc = __unmap_and_move(page, newpage, force, offlining, mode);
855 * A page that has been migrated has all references
856 * removed and will be freed. A page that has not been
857 * migrated will have kepts its references and be
860 list_del(&page->lru);
861 dec_zone_page_state(page, NR_ISOLATED_ANON +
862 page_is_file_cache(page));
863 putback_lru_page(page);
866 * Move the new page to the LRU. If migration was not successful
867 * then this will free the page.
869 putback_lru_page(newpage);
874 *result = page_to_nid(newpage);
880 * Counterpart of unmap_and_move_page() for hugepage migration.
882 * This function doesn't wait the completion of hugepage I/O
883 * because there is no race between I/O and migration for hugepage.
884 * Note that currently hugepage I/O occurs only in direct I/O
885 * where no lock is held and PG_writeback is irrelevant,
886 * and writeback status of all subpages are counted in the reference
887 * count of the head page (i.e. if all subpages of a 2MB hugepage are
888 * under direct I/O, the reference of the head page is 512 and a bit more.)
889 * This means that when we try to migrate hugepage whose subpages are
890 * doing direct I/O, some references remain after try_to_unmap() and
891 * hugepage migration fails without data corruption.
893 * There is also no race when direct I/O is issued on the page under migration,
894 * because then pte is replaced with migration swap entry and direct I/O code
895 * will wait in the page fault for migration to complete.
897 static int unmap_and_move_huge_page(new_page_t get_new_page,
898 unsigned long private, struct page *hpage,
899 int force, bool offlining,
900 enum migrate_mode mode)
904 struct page *new_hpage = get_new_page(hpage, private, &result);
905 struct anon_vma *anon_vma = NULL;
912 if (!trylock_page(hpage)) {
913 if (!force || mode != MIGRATE_SYNC)
919 anon_vma = page_get_anon_vma(hpage);
921 try_to_unmap(hpage, TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
923 if (!page_mapped(hpage))
924 rc = move_to_new_page(new_hpage, hpage, 1, mode);
927 remove_migration_ptes(hpage, hpage);
930 put_anon_vma(anon_vma);
933 hugetlb_cgroup_migrate(hpage, new_hpage);
942 *result = page_to_nid(new_hpage);
950 * The function takes one list of pages to migrate and a function
951 * that determines from the page to be migrated and the private data
952 * the target of the move and allocates the page.
954 * The function returns after 10 attempts or if no pages
955 * are movable anymore because to has become empty
956 * or no retryable pages exist anymore.
957 * Caller should call putback_lru_pages to return pages to the LRU
958 * or free list only if ret != 0.
960 * Return: Number of pages not migrated or error code.
962 int migrate_pages(struct list_head *from,
963 new_page_t get_new_page, unsigned long private, bool offlining,
964 enum migrate_mode mode, int reason)
968 int nr_succeeded = 0;
972 int swapwrite = current->flags & PF_SWAPWRITE;
976 current->flags |= PF_SWAPWRITE;
978 for(pass = 0; pass < 10 && retry; pass++) {
981 list_for_each_entry_safe(page, page2, from, lru) {
984 rc = unmap_and_move(get_new_page, private,
985 page, pass > 2, offlining,
998 /* Permanent failure */
1007 count_vm_events(PGMIGRATE_SUCCESS, nr_succeeded);
1009 count_vm_events(PGMIGRATE_FAIL, nr_failed);
1010 trace_mm_migrate_pages(nr_succeeded, nr_failed, mode, reason);
1013 current->flags &= ~PF_SWAPWRITE;
1018 return nr_failed + retry;
1021 int migrate_huge_page(struct page *hpage, new_page_t get_new_page,
1022 unsigned long private, bool offlining,
1023 enum migrate_mode mode)
1027 for (pass = 0; pass < 10; pass++) {
1028 rc = unmap_and_move_huge_page(get_new_page,
1029 private, hpage, pass > 2, offlining,
1051 * Move a list of individual pages
1053 struct page_to_node {
1060 static struct page *new_page_node(struct page *p, unsigned long private,
1063 struct page_to_node *pm = (struct page_to_node *)private;
1065 while (pm->node != MAX_NUMNODES && pm->page != p)
1068 if (pm->node == MAX_NUMNODES)
1071 *result = &pm->status;
1073 return alloc_pages_exact_node(pm->node,
1074 GFP_HIGHUSER_MOVABLE | GFP_THISNODE, 0);
1078 * Move a set of pages as indicated in the pm array. The addr
1079 * field must be set to the virtual address of the page to be moved
1080 * and the node number must contain a valid target node.
1081 * The pm array ends with node = MAX_NUMNODES.
1083 static int do_move_page_to_node_array(struct mm_struct *mm,
1084 struct page_to_node *pm,
1088 struct page_to_node *pp;
1089 LIST_HEAD(pagelist);
1091 down_read(&mm->mmap_sem);
1094 * Build a list of pages to migrate
1096 for (pp = pm; pp->node != MAX_NUMNODES; pp++) {
1097 struct vm_area_struct *vma;
1101 vma = find_vma(mm, pp->addr);
1102 if (!vma || pp->addr < vma->vm_start || !vma_migratable(vma))
1105 page = follow_page(vma, pp->addr, FOLL_GET|FOLL_SPLIT);
1107 err = PTR_ERR(page);
1115 /* Use PageReserved to check for zero page */
1116 if (PageReserved(page) || PageKsm(page))
1120 err = page_to_nid(page);
1122 if (err == pp->node)
1124 * Node already in the right place
1129 if (page_mapcount(page) > 1 &&
1133 err = isolate_lru_page(page);
1135 list_add_tail(&page->lru, &pagelist);
1136 inc_zone_page_state(page, NR_ISOLATED_ANON +
1137 page_is_file_cache(page));
1141 * Either remove the duplicate refcount from
1142 * isolate_lru_page() or drop the page ref if it was
1151 if (!list_empty(&pagelist)) {
1152 err = migrate_pages(&pagelist, new_page_node,
1153 (unsigned long)pm, 0, MIGRATE_SYNC,
1156 putback_lru_pages(&pagelist);
1159 up_read(&mm->mmap_sem);
1164 * Migrate an array of page address onto an array of nodes and fill
1165 * the corresponding array of status.
1167 static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes,
1168 unsigned long nr_pages,
1169 const void __user * __user *pages,
1170 const int __user *nodes,
1171 int __user *status, int flags)
1173 struct page_to_node *pm;
1174 unsigned long chunk_nr_pages;
1175 unsigned long chunk_start;
1179 pm = (struct page_to_node *)__get_free_page(GFP_KERNEL);
1186 * Store a chunk of page_to_node array in a page,
1187 * but keep the last one as a marker
1189 chunk_nr_pages = (PAGE_SIZE / sizeof(struct page_to_node)) - 1;
1191 for (chunk_start = 0;
1192 chunk_start < nr_pages;
1193 chunk_start += chunk_nr_pages) {
1196 if (chunk_start + chunk_nr_pages > nr_pages)
1197 chunk_nr_pages = nr_pages - chunk_start;
1199 /* fill the chunk pm with addrs and nodes from user-space */
1200 for (j = 0; j < chunk_nr_pages; j++) {
1201 const void __user *p;
1205 if (get_user(p, pages + j + chunk_start))
1207 pm[j].addr = (unsigned long) p;
1209 if (get_user(node, nodes + j + chunk_start))
1213 if (node < 0 || node >= MAX_NUMNODES)
1216 if (!node_state(node, N_HIGH_MEMORY))
1220 if (!node_isset(node, task_nodes))
1226 /* End marker for this chunk */
1227 pm[chunk_nr_pages].node = MAX_NUMNODES;
1229 /* Migrate this chunk */
1230 err = do_move_page_to_node_array(mm, pm,
1231 flags & MPOL_MF_MOVE_ALL);
1235 /* Return status information */
1236 for (j = 0; j < chunk_nr_pages; j++)
1237 if (put_user(pm[j].status, status + j + chunk_start)) {
1245 free_page((unsigned long)pm);
1251 * Determine the nodes of an array of pages and store it in an array of status.
1253 static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages,
1254 const void __user **pages, int *status)
1258 down_read(&mm->mmap_sem);
1260 for (i = 0; i < nr_pages; i++) {
1261 unsigned long addr = (unsigned long)(*pages);
1262 struct vm_area_struct *vma;
1266 vma = find_vma(mm, addr);
1267 if (!vma || addr < vma->vm_start)
1270 page = follow_page(vma, addr, 0);
1272 err = PTR_ERR(page);
1277 /* Use PageReserved to check for zero page */
1278 if (!page || PageReserved(page) || PageKsm(page))
1281 err = page_to_nid(page);
1289 up_read(&mm->mmap_sem);
1293 * Determine the nodes of a user array of pages and store it in
1294 * a user array of status.
1296 static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages,
1297 const void __user * __user *pages,
1300 #define DO_PAGES_STAT_CHUNK_NR 16
1301 const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR];
1302 int chunk_status[DO_PAGES_STAT_CHUNK_NR];
1305 unsigned long chunk_nr;
1307 chunk_nr = nr_pages;
1308 if (chunk_nr > DO_PAGES_STAT_CHUNK_NR)
1309 chunk_nr = DO_PAGES_STAT_CHUNK_NR;
1311 if (copy_from_user(chunk_pages, pages, chunk_nr * sizeof(*chunk_pages)))
1314 do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status);
1316 if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status)))
1321 nr_pages -= chunk_nr;
1323 return nr_pages ? -EFAULT : 0;
1327 * Move a list of pages in the address space of the currently executing
1330 SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages,
1331 const void __user * __user *, pages,
1332 const int __user *, nodes,
1333 int __user *, status, int, flags)
1335 const struct cred *cred = current_cred(), *tcred;
1336 struct task_struct *task;
1337 struct mm_struct *mm;
1339 nodemask_t task_nodes;
1342 if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
1345 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1348 /* Find the mm_struct */
1350 task = pid ? find_task_by_vpid(pid) : current;
1355 get_task_struct(task);
1358 * Check if this process has the right to modify the specified
1359 * process. The right exists if the process has administrative
1360 * capabilities, superuser privileges or the same
1361 * userid as the target process.
1363 tcred = __task_cred(task);
1364 if (!uid_eq(cred->euid, tcred->suid) && !uid_eq(cred->euid, tcred->uid) &&
1365 !uid_eq(cred->uid, tcred->suid) && !uid_eq(cred->uid, tcred->uid) &&
1366 !capable(CAP_SYS_NICE)) {
1373 err = security_task_movememory(task);
1377 task_nodes = cpuset_mems_allowed(task);
1378 mm = get_task_mm(task);
1379 put_task_struct(task);
1385 err = do_pages_move(mm, task_nodes, nr_pages, pages,
1386 nodes, status, flags);
1388 err = do_pages_stat(mm, nr_pages, pages, status);
1394 put_task_struct(task);
1399 * Call migration functions in the vma_ops that may prepare
1400 * memory in a vm for migration. migration functions may perform
1401 * the migration for vmas that do not have an underlying page struct.
1403 int migrate_vmas(struct mm_struct *mm, const nodemask_t *to,
1404 const nodemask_t *from, unsigned long flags)
1406 struct vm_area_struct *vma;
1409 for (vma = mm->mmap; vma && !err; vma = vma->vm_next) {
1410 if (vma->vm_ops && vma->vm_ops->migrate) {
1411 err = vma->vm_ops->migrate(vma, to, from, flags);
1419 #ifdef CONFIG_NUMA_BALANCING
1421 * Returns true if this is a safe migration target node for misplaced NUMA
1422 * pages. Currently it only checks the watermarks which crude
1424 static bool migrate_balanced_pgdat(struct pglist_data *pgdat,
1425 int nr_migrate_pages)
1428 for (z = pgdat->nr_zones - 1; z >= 0; z--) {
1429 struct zone *zone = pgdat->node_zones + z;
1431 if (!populated_zone(zone))
1434 if (zone->all_unreclaimable)
1437 /* Avoid waking kswapd by allocating pages_to_migrate pages. */
1438 if (!zone_watermark_ok(zone, 0,
1439 high_wmark_pages(zone) +
1448 static struct page *alloc_misplaced_dst_page(struct page *page,
1452 int nid = (int) data;
1453 struct page *newpage;
1455 newpage = alloc_pages_exact_node(nid,
1456 (GFP_HIGHUSER_MOVABLE | GFP_THISNODE |
1457 __GFP_NOMEMALLOC | __GFP_NORETRY |
1461 page_xchg_last_nid(newpage, page_last_nid(page));
1467 * page migration rate limiting control.
1468 * Do not migrate more than @pages_to_migrate in a @migrate_interval_millisecs
1469 * window of time. Default here says do not migrate more than 1280M per second.
1470 * If a node is rate-limited then PTE NUMA updates are also rate-limited. However
1471 * as it is faults that reset the window, pte updates will happen unconditionally
1472 * if there has not been a fault since @pteupdate_interval_millisecs after the
1473 * throttle window closed.
1475 static unsigned int migrate_interval_millisecs __read_mostly = 100;
1476 static unsigned int pteupdate_interval_millisecs __read_mostly = 1000;
1477 static unsigned int ratelimit_pages __read_mostly = 128 << (20 - PAGE_SHIFT);
1479 /* Returns true if NUMA migration is currently rate limited */
1480 bool migrate_ratelimited(int node)
1482 pg_data_t *pgdat = NODE_DATA(node);
1484 if (time_after(jiffies, pgdat->numabalancing_migrate_next_window +
1485 msecs_to_jiffies(pteupdate_interval_millisecs)))
1488 if (pgdat->numabalancing_migrate_nr_pages < ratelimit_pages)
1494 /* Returns true if the node is migrate rate-limited after the update */
1495 bool numamigrate_update_ratelimit(pg_data_t *pgdat)
1497 bool rate_limited = false;
1500 * Rate-limit the amount of data that is being migrated to a node.
1501 * Optimal placement is no good if the memory bus is saturated and
1502 * all the time is being spent migrating!
1504 spin_lock(&pgdat->numabalancing_migrate_lock);
1505 if (time_after(jiffies, pgdat->numabalancing_migrate_next_window)) {
1506 pgdat->numabalancing_migrate_nr_pages = 0;
1507 pgdat->numabalancing_migrate_next_window = jiffies +
1508 msecs_to_jiffies(migrate_interval_millisecs);
1510 if (pgdat->numabalancing_migrate_nr_pages > ratelimit_pages)
1511 rate_limited = true;
1513 pgdat->numabalancing_migrate_nr_pages++;
1514 spin_unlock(&pgdat->numabalancing_migrate_lock);
1516 return rate_limited;
1519 int numamigrate_isolate_page(pg_data_t *pgdat, struct page *page)
1523 /* Avoid migrating to a node that is nearly full */
1524 if (migrate_balanced_pgdat(pgdat, 1)) {
1527 if (isolate_lru_page(page)) {
1532 /* Page is isolated */
1534 page_lru = page_is_file_cache(page);
1535 if (!PageTransHuge(page))
1536 inc_zone_page_state(page, NR_ISOLATED_ANON + page_lru);
1538 mod_zone_page_state(page_zone(page),
1539 NR_ISOLATED_ANON + page_lru,
1544 * Page is either isolated or there is not enough space on the target
1545 * node. If isolated, then it has taken a reference count and the
1546 * callers reference can be safely dropped without the page
1547 * disappearing underneath us during migration. Otherwise the page is
1548 * not to be migrated but the callers reference should still be
1549 * dropped so it does not leak.
1557 * Attempt to migrate a misplaced page to the specified destination
1558 * node. Caller is expected to have an elevated reference count on
1559 * the page that will be dropped by this function before returning.
1561 int migrate_misplaced_page(struct page *page, int node)
1563 pg_data_t *pgdat = NODE_DATA(node);
1566 LIST_HEAD(migratepages);
1569 * Don't migrate pages that are mapped in multiple processes.
1570 * TODO: Handle false sharing detection instead of this hammer
1572 if (page_mapcount(page) != 1) {
1578 * Rate-limit the amount of data that is being migrated to a node.
1579 * Optimal placement is no good if the memory bus is saturated and
1580 * all the time is being spent migrating!
1582 if (numamigrate_update_ratelimit(pgdat)) {
1587 isolated = numamigrate_isolate_page(pgdat, page);
1591 list_add(&page->lru, &migratepages);
1592 nr_remaining = migrate_pages(&migratepages,
1593 alloc_misplaced_dst_page,
1594 node, false, MIGRATE_ASYNC,
1597 putback_lru_pages(&migratepages);
1600 count_vm_numa_event(NUMA_PAGE_MIGRATE);
1601 BUG_ON(!list_empty(&migratepages));
1606 int migrate_misplaced_transhuge_page(struct mm_struct *mm,
1607 struct vm_area_struct *vma,
1608 pmd_t *pmd, pmd_t entry,
1609 unsigned long address,
1610 struct page *page, int node)
1612 unsigned long haddr = address & HPAGE_PMD_MASK;
1613 pg_data_t *pgdat = NODE_DATA(node);
1615 struct page *new_page = NULL;
1616 struct mem_cgroup *memcg = NULL;
1617 int page_lru = page_is_file_cache(page);
1620 * Don't migrate pages that are mapped in multiple processes.
1621 * TODO: Handle false sharing detection instead of this hammer
1623 if (page_mapcount(page) != 1)
1627 * Rate-limit the amount of data that is being migrated to a node.
1628 * Optimal placement is no good if the memory bus is saturated and
1629 * all the time is being spent migrating!
1631 if (numamigrate_update_ratelimit(pgdat))
1634 new_page = alloc_pages_node(node,
1635 (GFP_TRANSHUGE | GFP_THISNODE) & ~__GFP_WAIT, HPAGE_PMD_ORDER);
1638 page_xchg_last_nid(new_page, page_last_nid(page));
1640 isolated = numamigrate_isolate_page(pgdat, page);
1643 goto out_keep_locked;
1646 /* Prepare a page as a migration target */
1647 __set_page_locked(new_page);
1648 SetPageSwapBacked(new_page);
1650 /* anon mapping, we can simply copy page->mapping to the new page: */
1651 new_page->mapping = page->mapping;
1652 new_page->index = page->index;
1653 migrate_page_copy(new_page, page);
1654 WARN_ON(PageLRU(new_page));
1656 /* Recheck the target PMD */
1657 spin_lock(&mm->page_table_lock);
1658 if (unlikely(!pmd_same(*pmd, entry))) {
1659 spin_unlock(&mm->page_table_lock);
1661 /* Reverse changes made by migrate_page_copy() */
1662 if (TestClearPageActive(new_page))
1663 SetPageActive(page);
1664 if (TestClearPageUnevictable(new_page))
1665 SetPageUnevictable(page);
1666 mlock_migrate_page(page, new_page);
1668 unlock_page(new_page);
1669 put_page(new_page); /* Free it */
1672 putback_lru_page(page);
1674 count_vm_events(PGMIGRATE_FAIL, HPAGE_PMD_NR);
1679 * Traditional migration needs to prepare the memcg charge
1680 * transaction early to prevent the old page from being
1681 * uncharged when installing migration entries. Here we can
1682 * save the potential rollback and start the charge transfer
1683 * only when migration is already known to end successfully.
1685 mem_cgroup_prepare_migration(page, new_page, &memcg);
1687 entry = mk_pmd(new_page, vma->vm_page_prot);
1688 entry = pmd_mknonnuma(entry);
1689 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1690 entry = pmd_mkhuge(entry);
1692 page_add_new_anon_rmap(new_page, vma, haddr);
1694 set_pmd_at(mm, haddr, pmd, entry);
1695 update_mmu_cache_pmd(vma, address, entry);
1696 page_remove_rmap(page);
1698 * Finish the charge transaction under the page table lock to
1699 * prevent split_huge_page() from dividing up the charge
1700 * before it's fully transferred to the new page.
1702 mem_cgroup_end_migration(memcg, page, new_page, true);
1703 spin_unlock(&mm->page_table_lock);
1705 unlock_page(new_page);
1707 put_page(page); /* Drop the rmap reference */
1708 put_page(page); /* Drop the LRU isolation reference */
1710 count_vm_events(PGMIGRATE_SUCCESS, HPAGE_PMD_NR);
1711 count_vm_numa_events(NUMA_PAGE_MIGRATE, HPAGE_PMD_NR);
1714 mod_zone_page_state(page_zone(page),
1715 NR_ISOLATED_ANON + page_lru,
1724 #endif /* CONFIG_NUMA_BALANCING */
1726 #endif /* CONFIG_NUMA */