2 * Memory Migration functionality - linux/mm/migration.c
4 * Copyright (C) 2006 Silicon Graphics, Inc., Christoph Lameter
6 * Page migration was first developed in the context of the memory hotplug
7 * project. The main authors of the migration code are:
9 * IWAMOTO Toshihiro <iwamoto@valinux.co.jp>
10 * Hirokazu Takahashi <taka@valinux.co.jp>
11 * Dave Hansen <haveblue@us.ibm.com>
15 #include <linux/migrate.h>
16 #include <linux/export.h>
17 #include <linux/swap.h>
18 #include <linux/swapops.h>
19 #include <linux/pagemap.h>
20 #include <linux/buffer_head.h>
21 #include <linux/mm_inline.h>
22 #include <linux/nsproxy.h>
23 #include <linux/pagevec.h>
24 #include <linux/ksm.h>
25 #include <linux/rmap.h>
26 #include <linux/topology.h>
27 #include <linux/cpu.h>
28 #include <linux/cpuset.h>
29 #include <linux/writeback.h>
30 #include <linux/mempolicy.h>
31 #include <linux/vmalloc.h>
32 #include <linux/security.h>
33 #include <linux/memcontrol.h>
34 #include <linux/syscalls.h>
35 #include <linux/hugetlb.h>
36 #include <linux/hugetlb_cgroup.h>
37 #include <linux/gfp.h>
38 #include <linux/balloon_compaction.h>
40 #include <asm/tlbflush.h>
42 #define CREATE_TRACE_POINTS
43 #include <trace/events/migrate.h>
48 * migrate_prep() needs to be called before we start compiling a list of pages
49 * to be migrated using isolate_lru_page(). If scheduling work on other CPUs is
50 * undesirable, use migrate_prep_local()
52 int migrate_prep(void)
55 * Clear the LRU lists so pages can be isolated.
56 * Note that pages may be moved off the LRU after we have
57 * drained them. Those pages will fail to migrate like other
58 * pages that may be busy.
65 /* Do the necessary work of migrate_prep but not if it involves other CPUs */
66 int migrate_prep_local(void)
74 * Add isolated pages on the list back to the LRU under page lock
75 * to avoid leaking evictable pages back onto unevictable list.
77 void putback_lru_pages(struct list_head *l)
82 list_for_each_entry_safe(page, page2, l, lru) {
84 dec_zone_page_state(page, NR_ISOLATED_ANON +
85 page_is_file_cache(page));
86 putback_lru_page(page);
91 * Put previously isolated pages back onto the appropriate lists
92 * from where they were once taken off for compaction/migration.
94 * This function shall be used instead of putback_lru_pages(),
95 * whenever the isolated pageset has been built by isolate_migratepages_range()
97 void putback_movable_pages(struct list_head *l)
102 list_for_each_entry_safe(page, page2, l, lru) {
103 list_del(&page->lru);
104 dec_zone_page_state(page, NR_ISOLATED_ANON +
105 page_is_file_cache(page));
106 if (unlikely(balloon_page_movable(page)))
107 balloon_page_putback(page);
109 putback_lru_page(page);
114 * Restore a potential migration pte to a working pte entry
116 static int remove_migration_pte(struct page *new, struct vm_area_struct *vma,
117 unsigned long addr, void *old)
119 struct mm_struct *mm = vma->vm_mm;
125 if (unlikely(PageHuge(new))) {
126 ptep = huge_pte_offset(mm, addr);
129 ptl = &mm->page_table_lock;
131 pmd = mm_find_pmd(mm, addr);
134 if (pmd_trans_huge(*pmd))
137 ptep = pte_offset_map(pmd, addr);
140 * Peek to check is_swap_pte() before taking ptlock? No, we
141 * can race mremap's move_ptes(), which skips anon_vma lock.
144 ptl = pte_lockptr(mm, pmd);
149 if (!is_swap_pte(pte))
152 entry = pte_to_swp_entry(pte);
154 if (!is_migration_entry(entry) ||
155 migration_entry_to_page(entry) != old)
159 pte = pte_mkold(mk_pte(new, vma->vm_page_prot));
160 if (is_write_migration_entry(entry))
161 pte = pte_mkwrite(pte);
162 #ifdef CONFIG_HUGETLB_PAGE
164 pte = pte_mkhuge(pte);
166 flush_cache_page(vma, addr, pte_pfn(pte));
167 set_pte_at(mm, addr, ptep, pte);
171 hugepage_add_anon_rmap(new, vma, addr);
174 } else if (PageAnon(new))
175 page_add_anon_rmap(new, vma, addr);
177 page_add_file_rmap(new);
179 /* No need to invalidate - it was non-present before */
180 update_mmu_cache(vma, addr, ptep);
182 pte_unmap_unlock(ptep, ptl);
188 * Get rid of all migration entries and replace them by
189 * references to the indicated page.
191 static void remove_migration_ptes(struct page *old, struct page *new)
193 rmap_walk(new, remove_migration_pte, old);
197 * Something used the pte of a page under migration. We need to
198 * get to the page and wait until migration is finished.
199 * When we return from this function the fault will be retried.
201 void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
202 unsigned long address)
209 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
211 if (!is_swap_pte(pte))
214 entry = pte_to_swp_entry(pte);
215 if (!is_migration_entry(entry))
218 page = migration_entry_to_page(entry);
221 * Once radix-tree replacement of page migration started, page_count
222 * *must* be zero. And, we don't want to call wait_on_page_locked()
223 * against a page without get_page().
224 * So, we use get_page_unless_zero(), here. Even failed, page fault
227 if (!get_page_unless_zero(page))
229 pte_unmap_unlock(ptep, ptl);
230 wait_on_page_locked(page);
234 pte_unmap_unlock(ptep, ptl);
238 /* Returns true if all buffers are successfully locked */
239 static bool buffer_migrate_lock_buffers(struct buffer_head *head,
240 enum migrate_mode mode)
242 struct buffer_head *bh = head;
244 /* Simple case, sync compaction */
245 if (mode != MIGRATE_ASYNC) {
249 bh = bh->b_this_page;
251 } while (bh != head);
256 /* async case, we cannot block on lock_buffer so use trylock_buffer */
259 if (!trylock_buffer(bh)) {
261 * We failed to lock the buffer and cannot stall in
262 * async migration. Release the taken locks
264 struct buffer_head *failed_bh = bh;
267 while (bh != failed_bh) {
270 bh = bh->b_this_page;
275 bh = bh->b_this_page;
276 } while (bh != head);
280 static inline bool buffer_migrate_lock_buffers(struct buffer_head *head,
281 enum migrate_mode mode)
285 #endif /* CONFIG_BLOCK */
288 * Replace the page in the mapping.
290 * The number of remaining references must be:
291 * 1 for anonymous pages without a mapping
292 * 2 for pages with a mapping
293 * 3 for pages with a mapping and PagePrivate/PagePrivate2 set.
295 static int migrate_page_move_mapping(struct address_space *mapping,
296 struct page *newpage, struct page *page,
297 struct buffer_head *head, enum migrate_mode mode)
299 int expected_count = 0;
303 /* Anonymous page without mapping */
304 if (page_count(page) != 1)
306 return MIGRATEPAGE_SUCCESS;
309 spin_lock_irq(&mapping->tree_lock);
311 pslot = radix_tree_lookup_slot(&mapping->page_tree,
314 expected_count = 2 + page_has_private(page);
315 if (page_count(page) != expected_count ||
316 radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) {
317 spin_unlock_irq(&mapping->tree_lock);
321 if (!page_freeze_refs(page, expected_count)) {
322 spin_unlock_irq(&mapping->tree_lock);
327 * In the async migration case of moving a page with buffers, lock the
328 * buffers using trylock before the mapping is moved. If the mapping
329 * was moved, we later failed to lock the buffers and could not move
330 * the mapping back due to an elevated page count, we would have to
331 * block waiting on other references to be dropped.
333 if (mode == MIGRATE_ASYNC && head &&
334 !buffer_migrate_lock_buffers(head, mode)) {
335 page_unfreeze_refs(page, expected_count);
336 spin_unlock_irq(&mapping->tree_lock);
341 * Now we know that no one else is looking at the page.
343 get_page(newpage); /* add cache reference */
344 if (PageSwapCache(page)) {
345 SetPageSwapCache(newpage);
346 set_page_private(newpage, page_private(page));
349 radix_tree_replace_slot(pslot, newpage);
352 * Drop cache reference from old page by unfreezing
353 * to one less reference.
354 * We know this isn't the last reference.
356 page_unfreeze_refs(page, expected_count - 1);
359 * If moved to a different zone then also account
360 * the page for that zone. Other VM counters will be
361 * taken care of when we establish references to the
362 * new page and drop references to the old page.
364 * Note that anonymous pages are accounted for
365 * via NR_FILE_PAGES and NR_ANON_PAGES if they
366 * are mapped to swap space.
368 __dec_zone_page_state(page, NR_FILE_PAGES);
369 __inc_zone_page_state(newpage, NR_FILE_PAGES);
370 if (!PageSwapCache(page) && PageSwapBacked(page)) {
371 __dec_zone_page_state(page, NR_SHMEM);
372 __inc_zone_page_state(newpage, NR_SHMEM);
374 spin_unlock_irq(&mapping->tree_lock);
376 return MIGRATEPAGE_SUCCESS;
380 * The expected number of remaining references is the same as that
381 * of migrate_page_move_mapping().
383 int migrate_huge_page_move_mapping(struct address_space *mapping,
384 struct page *newpage, struct page *page)
390 if (page_count(page) != 1)
392 return MIGRATEPAGE_SUCCESS;
395 spin_lock_irq(&mapping->tree_lock);
397 pslot = radix_tree_lookup_slot(&mapping->page_tree,
400 expected_count = 2 + page_has_private(page);
401 if (page_count(page) != expected_count ||
402 radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) {
403 spin_unlock_irq(&mapping->tree_lock);
407 if (!page_freeze_refs(page, expected_count)) {
408 spin_unlock_irq(&mapping->tree_lock);
414 radix_tree_replace_slot(pslot, newpage);
416 page_unfreeze_refs(page, expected_count - 1);
418 spin_unlock_irq(&mapping->tree_lock);
419 return MIGRATEPAGE_SUCCESS;
423 * Copy the page to its new location
425 void migrate_page_copy(struct page *newpage, struct page *page)
427 if (PageHuge(page) || PageTransHuge(page))
428 copy_huge_page(newpage, page);
430 copy_highpage(newpage, page);
433 SetPageError(newpage);
434 if (PageReferenced(page))
435 SetPageReferenced(newpage);
436 if (PageUptodate(page))
437 SetPageUptodate(newpage);
438 if (TestClearPageActive(page)) {
439 VM_BUG_ON(PageUnevictable(page));
440 SetPageActive(newpage);
441 } else if (TestClearPageUnevictable(page))
442 SetPageUnevictable(newpage);
443 if (PageChecked(page))
444 SetPageChecked(newpage);
445 if (PageMappedToDisk(page))
446 SetPageMappedToDisk(newpage);
448 if (PageDirty(page)) {
449 clear_page_dirty_for_io(page);
451 * Want to mark the page and the radix tree as dirty, and
452 * redo the accounting that clear_page_dirty_for_io undid,
453 * but we can't use set_page_dirty because that function
454 * is actually a signal that all of the page has become dirty.
455 * Whereas only part of our page may be dirty.
457 if (PageSwapBacked(page))
458 SetPageDirty(newpage);
460 __set_page_dirty_nobuffers(newpage);
463 mlock_migrate_page(newpage, page);
464 ksm_migrate_page(newpage, page);
466 ClearPageSwapCache(page);
467 ClearPagePrivate(page);
468 set_page_private(page, 0);
471 * If any waiters have accumulated on the new page then
474 if (PageWriteback(newpage))
475 end_page_writeback(newpage);
478 /************************************************************
479 * Migration functions
480 ***********************************************************/
482 /* Always fail migration. Used for mappings that are not movable */
483 int fail_migrate_page(struct address_space *mapping,
484 struct page *newpage, struct page *page)
488 EXPORT_SYMBOL(fail_migrate_page);
491 * Common logic to directly migrate a single page suitable for
492 * pages that do not use PagePrivate/PagePrivate2.
494 * Pages are locked upon entry and exit.
496 int migrate_page(struct address_space *mapping,
497 struct page *newpage, struct page *page,
498 enum migrate_mode mode)
502 BUG_ON(PageWriteback(page)); /* Writeback must be complete */
504 rc = migrate_page_move_mapping(mapping, newpage, page, NULL, mode);
506 if (rc != MIGRATEPAGE_SUCCESS)
509 migrate_page_copy(newpage, page);
510 return MIGRATEPAGE_SUCCESS;
512 EXPORT_SYMBOL(migrate_page);
516 * Migration function for pages with buffers. This function can only be used
517 * if the underlying filesystem guarantees that no other references to "page"
520 int buffer_migrate_page(struct address_space *mapping,
521 struct page *newpage, struct page *page, enum migrate_mode mode)
523 struct buffer_head *bh, *head;
526 if (!page_has_buffers(page))
527 return migrate_page(mapping, newpage, page, mode);
529 head = page_buffers(page);
531 rc = migrate_page_move_mapping(mapping, newpage, page, head, mode);
533 if (rc != MIGRATEPAGE_SUCCESS)
537 * In the async case, migrate_page_move_mapping locked the buffers
538 * with an IRQ-safe spinlock held. In the sync case, the buffers
539 * need to be locked now
541 if (mode != MIGRATE_ASYNC)
542 BUG_ON(!buffer_migrate_lock_buffers(head, mode));
544 ClearPagePrivate(page);
545 set_page_private(newpage, page_private(page));
546 set_page_private(page, 0);
552 set_bh_page(bh, newpage, bh_offset(bh));
553 bh = bh->b_this_page;
555 } while (bh != head);
557 SetPagePrivate(newpage);
559 migrate_page_copy(newpage, page);
565 bh = bh->b_this_page;
567 } while (bh != head);
569 return MIGRATEPAGE_SUCCESS;
571 EXPORT_SYMBOL(buffer_migrate_page);
575 * Writeback a page to clean the dirty state
577 static int writeout(struct address_space *mapping, struct page *page)
579 struct writeback_control wbc = {
580 .sync_mode = WB_SYNC_NONE,
583 .range_end = LLONG_MAX,
588 if (!mapping->a_ops->writepage)
589 /* No write method for the address space */
592 if (!clear_page_dirty_for_io(page))
593 /* Someone else already triggered a write */
597 * A dirty page may imply that the underlying filesystem has
598 * the page on some queue. So the page must be clean for
599 * migration. Writeout may mean we loose the lock and the
600 * page state is no longer what we checked for earlier.
601 * At this point we know that the migration attempt cannot
604 remove_migration_ptes(page, page);
606 rc = mapping->a_ops->writepage(page, &wbc);
608 if (rc != AOP_WRITEPAGE_ACTIVATE)
609 /* unlocked. Relock */
612 return (rc < 0) ? -EIO : -EAGAIN;
616 * Default handling if a filesystem does not provide a migration function.
618 static int fallback_migrate_page(struct address_space *mapping,
619 struct page *newpage, struct page *page, enum migrate_mode mode)
621 if (PageDirty(page)) {
622 /* Only writeback pages in full synchronous migration */
623 if (mode != MIGRATE_SYNC)
625 return writeout(mapping, page);
629 * Buffers may be managed in a filesystem specific way.
630 * We must have no buffers or drop them.
632 if (page_has_private(page) &&
633 !try_to_release_page(page, GFP_KERNEL))
636 return migrate_page(mapping, newpage, page, mode);
640 * Move a page to a newly allocated page
641 * The page is locked and all ptes have been successfully removed.
643 * The new page will have replaced the old page if this function
648 * MIGRATEPAGE_SUCCESS - success
650 static int move_to_new_page(struct page *newpage, struct page *page,
651 int remap_swapcache, enum migrate_mode mode)
653 struct address_space *mapping;
657 * Block others from accessing the page when we get around to
658 * establishing additional references. We are the only one
659 * holding a reference to the new page at this point.
661 if (!trylock_page(newpage))
664 /* Prepare mapping for the new page.*/
665 newpage->index = page->index;
666 newpage->mapping = page->mapping;
667 if (PageSwapBacked(page))
668 SetPageSwapBacked(newpage);
670 mapping = page_mapping(page);
672 rc = migrate_page(mapping, newpage, page, mode);
673 else if (mapping->a_ops->migratepage)
675 * Most pages have a mapping and most filesystems provide a
676 * migratepage callback. Anonymous pages are part of swap
677 * space which also has its own migratepage callback. This
678 * is the most common path for page migration.
680 rc = mapping->a_ops->migratepage(mapping,
681 newpage, page, mode);
683 rc = fallback_migrate_page(mapping, newpage, page, mode);
685 if (rc != MIGRATEPAGE_SUCCESS) {
686 newpage->mapping = NULL;
689 remove_migration_ptes(page, newpage);
690 page->mapping = NULL;
693 unlock_page(newpage);
698 static int __unmap_and_move(struct page *page, struct page *newpage,
699 int force, bool offlining, enum migrate_mode mode)
702 int remap_swapcache = 1;
703 struct mem_cgroup *mem;
704 struct anon_vma *anon_vma = NULL;
706 if (!trylock_page(page)) {
707 if (!force || mode == MIGRATE_ASYNC)
711 * It's not safe for direct compaction to call lock_page.
712 * For example, during page readahead pages are added locked
713 * to the LRU. Later, when the IO completes the pages are
714 * marked uptodate and unlocked. However, the queueing
715 * could be merging multiple pages for one bio (e.g.
716 * mpage_readpages). If an allocation happens for the
717 * second or third page, the process can end up locking
718 * the same page twice and deadlocking. Rather than
719 * trying to be clever about what pages can be locked,
720 * avoid the use of lock_page for direct compaction
723 if (current->flags & PF_MEMALLOC)
730 * Only memory hotplug's offline_pages() caller has locked out KSM,
731 * and can safely migrate a KSM page. The other cases have skipped
732 * PageKsm along with PageReserved - but it is only now when we have
733 * the page lock that we can be certain it will not go KSM beneath us
734 * (KSM will not upgrade a page from PageAnon to PageKsm when it sees
735 * its pagecount raised, but only here do we take the page lock which
738 if (PageKsm(page) && !offlining) {
743 /* charge against new page */
744 mem_cgroup_prepare_migration(page, newpage, &mem);
746 if (PageWriteback(page)) {
748 * Only in the case of a full syncronous migration is it
749 * necessary to wait for PageWriteback. In the async case,
750 * the retry loop is too short and in the sync-light case,
751 * the overhead of stalling is too much
753 if (mode != MIGRATE_SYNC) {
759 wait_on_page_writeback(page);
762 * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
763 * we cannot notice that anon_vma is freed while we migrates a page.
764 * This get_anon_vma() delays freeing anon_vma pointer until the end
765 * of migration. File cache pages are no problem because of page_lock()
766 * File Caches may use write_page() or lock_page() in migration, then,
767 * just care Anon page here.
769 if (PageAnon(page)) {
771 * Only page_lock_anon_vma_read() understands the subtleties of
772 * getting a hold on an anon_vma from outside one of its mms.
774 anon_vma = page_get_anon_vma(page);
779 } else if (PageSwapCache(page)) {
781 * We cannot be sure that the anon_vma of an unmapped
782 * swapcache page is safe to use because we don't
783 * know in advance if the VMA that this page belonged
784 * to still exists. If the VMA and others sharing the
785 * data have been freed, then the anon_vma could
786 * already be invalid.
788 * To avoid this possibility, swapcache pages get
789 * migrated but are not remapped when migration
798 if (unlikely(balloon_page_movable(page))) {
800 * A ballooned page does not need any special attention from
801 * physical to virtual reverse mapping procedures.
802 * Skip any attempt to unmap PTEs or to remap swap cache,
803 * in order to avoid burning cycles at rmap level, and perform
804 * the page migration right away (proteced by page lock).
806 rc = balloon_page_migrate(newpage, page, mode);
811 * Corner case handling:
812 * 1. When a new swap-cache page is read into, it is added to the LRU
813 * and treated as swapcache but it has no rmap yet.
814 * Calling try_to_unmap() against a page->mapping==NULL page will
815 * trigger a BUG. So handle it here.
816 * 2. An orphaned page (see truncate_complete_page) might have
817 * fs-private metadata. The page can be picked up due to memory
818 * offlining. Everywhere else except page reclaim, the page is
819 * invisible to the vm, so the page can not be migrated. So try to
820 * free the metadata, so the page can be freed.
822 if (!page->mapping) {
823 VM_BUG_ON(PageAnon(page));
824 if (page_has_private(page)) {
825 try_to_free_buffers(page);
831 /* Establish migration ptes or remove ptes */
832 try_to_unmap(page, TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
835 if (!page_mapped(page))
836 rc = move_to_new_page(newpage, page, remap_swapcache, mode);
838 if (rc && remap_swapcache)
839 remove_migration_ptes(page, page);
841 /* Drop an anon_vma reference if we took one */
843 put_anon_vma(anon_vma);
846 mem_cgroup_end_migration(mem, page, newpage,
847 (rc == MIGRATEPAGE_SUCCESS ||
848 rc == MIGRATEPAGE_BALLOON_SUCCESS));
856 * Obtain the lock on page, remove all ptes and migrate the page
857 * to the newly allocated page in newpage.
859 static int unmap_and_move(new_page_t get_new_page, unsigned long private,
860 struct page *page, int force, bool offlining,
861 enum migrate_mode mode)
865 struct page *newpage = get_new_page(page, private, &result);
870 if (page_count(page) == 1) {
871 /* page was freed from under us. So we are done. */
875 if (unlikely(PageTransHuge(page)))
876 if (unlikely(split_huge_page(page)))
879 rc = __unmap_and_move(page, newpage, force, offlining, mode);
881 if (unlikely(rc == MIGRATEPAGE_BALLOON_SUCCESS)) {
883 * A ballooned page has been migrated already.
884 * Now, it's the time to wrap-up counters,
885 * handle the page back to Buddy and return.
887 dec_zone_page_state(page, NR_ISOLATED_ANON +
888 page_is_file_cache(page));
889 balloon_page_free(page);
890 return MIGRATEPAGE_SUCCESS;
895 * A page that has been migrated has all references
896 * removed and will be freed. A page that has not been
897 * migrated will have kepts its references and be
900 list_del(&page->lru);
901 dec_zone_page_state(page, NR_ISOLATED_ANON +
902 page_is_file_cache(page));
903 putback_lru_page(page);
906 * Move the new page to the LRU. If migration was not successful
907 * then this will free the page.
909 putback_lru_page(newpage);
914 *result = page_to_nid(newpage);
920 * Counterpart of unmap_and_move_page() for hugepage migration.
922 * This function doesn't wait the completion of hugepage I/O
923 * because there is no race between I/O and migration for hugepage.
924 * Note that currently hugepage I/O occurs only in direct I/O
925 * where no lock is held and PG_writeback is irrelevant,
926 * and writeback status of all subpages are counted in the reference
927 * count of the head page (i.e. if all subpages of a 2MB hugepage are
928 * under direct I/O, the reference of the head page is 512 and a bit more.)
929 * This means that when we try to migrate hugepage whose subpages are
930 * doing direct I/O, some references remain after try_to_unmap() and
931 * hugepage migration fails without data corruption.
933 * There is also no race when direct I/O is issued on the page under migration,
934 * because then pte is replaced with migration swap entry and direct I/O code
935 * will wait in the page fault for migration to complete.
937 static int unmap_and_move_huge_page(new_page_t get_new_page,
938 unsigned long private, struct page *hpage,
939 int force, bool offlining,
940 enum migrate_mode mode)
944 struct page *new_hpage = get_new_page(hpage, private, &result);
945 struct anon_vma *anon_vma = NULL;
952 if (!trylock_page(hpage)) {
953 if (!force || mode != MIGRATE_SYNC)
959 anon_vma = page_get_anon_vma(hpage);
961 try_to_unmap(hpage, TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
963 if (!page_mapped(hpage))
964 rc = move_to_new_page(new_hpage, hpage, 1, mode);
967 remove_migration_ptes(hpage, hpage);
970 put_anon_vma(anon_vma);
973 hugetlb_cgroup_migrate(hpage, new_hpage);
982 *result = page_to_nid(new_hpage);
990 * The function takes one list of pages to migrate and a function
991 * that determines from the page to be migrated and the private data
992 * the target of the move and allocates the page.
994 * The function returns after 10 attempts or if no pages
995 * are movable anymore because to has become empty
996 * or no retryable pages exist anymore.
997 * Caller should call putback_lru_pages to return pages to the LRU
998 * or free list only if ret != 0.
1000 * Return: Number of pages not migrated or error code.
1002 int migrate_pages(struct list_head *from,
1003 new_page_t get_new_page, unsigned long private, bool offlining,
1004 enum migrate_mode mode, int reason)
1008 int nr_succeeded = 0;
1012 int swapwrite = current->flags & PF_SWAPWRITE;
1016 current->flags |= PF_SWAPWRITE;
1018 for(pass = 0; pass < 10 && retry; pass++) {
1021 list_for_each_entry_safe(page, page2, from, lru) {
1024 rc = unmap_and_move(get_new_page, private,
1025 page, pass > 2, offlining,
1034 case MIGRATEPAGE_SUCCESS:
1038 /* Permanent failure */
1044 rc = nr_failed + retry;
1047 count_vm_events(PGMIGRATE_SUCCESS, nr_succeeded);
1049 count_vm_events(PGMIGRATE_FAIL, nr_failed);
1050 trace_mm_migrate_pages(nr_succeeded, nr_failed, mode, reason);
1053 current->flags &= ~PF_SWAPWRITE;
1058 int migrate_huge_page(struct page *hpage, new_page_t get_new_page,
1059 unsigned long private, bool offlining,
1060 enum migrate_mode mode)
1064 for (pass = 0; pass < 10; pass++) {
1065 rc = unmap_and_move_huge_page(get_new_page,
1066 private, hpage, pass > 2, offlining,
1075 case MIGRATEPAGE_SUCCESS:
1088 * Move a list of individual pages
1090 struct page_to_node {
1097 static struct page *new_page_node(struct page *p, unsigned long private,
1100 struct page_to_node *pm = (struct page_to_node *)private;
1102 while (pm->node != MAX_NUMNODES && pm->page != p)
1105 if (pm->node == MAX_NUMNODES)
1108 *result = &pm->status;
1110 return alloc_pages_exact_node(pm->node,
1111 GFP_HIGHUSER_MOVABLE | GFP_THISNODE, 0);
1115 * Move a set of pages as indicated in the pm array. The addr
1116 * field must be set to the virtual address of the page to be moved
1117 * and the node number must contain a valid target node.
1118 * The pm array ends with node = MAX_NUMNODES.
1120 static int do_move_page_to_node_array(struct mm_struct *mm,
1121 struct page_to_node *pm,
1125 struct page_to_node *pp;
1126 LIST_HEAD(pagelist);
1128 down_read(&mm->mmap_sem);
1131 * Build a list of pages to migrate
1133 for (pp = pm; pp->node != MAX_NUMNODES; pp++) {
1134 struct vm_area_struct *vma;
1138 vma = find_vma(mm, pp->addr);
1139 if (!vma || pp->addr < vma->vm_start || !vma_migratable(vma))
1142 page = follow_page(vma, pp->addr, FOLL_GET|FOLL_SPLIT);
1144 err = PTR_ERR(page);
1152 /* Use PageReserved to check for zero page */
1153 if (PageReserved(page) || PageKsm(page))
1157 err = page_to_nid(page);
1159 if (err == pp->node)
1161 * Node already in the right place
1166 if (page_mapcount(page) > 1 &&
1170 err = isolate_lru_page(page);
1172 list_add_tail(&page->lru, &pagelist);
1173 inc_zone_page_state(page, NR_ISOLATED_ANON +
1174 page_is_file_cache(page));
1178 * Either remove the duplicate refcount from
1179 * isolate_lru_page() or drop the page ref if it was
1188 if (!list_empty(&pagelist)) {
1189 err = migrate_pages(&pagelist, new_page_node,
1190 (unsigned long)pm, 0, MIGRATE_SYNC,
1193 putback_lru_pages(&pagelist);
1196 up_read(&mm->mmap_sem);
1201 * Migrate an array of page address onto an array of nodes and fill
1202 * the corresponding array of status.
1204 static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes,
1205 unsigned long nr_pages,
1206 const void __user * __user *pages,
1207 const int __user *nodes,
1208 int __user *status, int flags)
1210 struct page_to_node *pm;
1211 unsigned long chunk_nr_pages;
1212 unsigned long chunk_start;
1216 pm = (struct page_to_node *)__get_free_page(GFP_KERNEL);
1223 * Store a chunk of page_to_node array in a page,
1224 * but keep the last one as a marker
1226 chunk_nr_pages = (PAGE_SIZE / sizeof(struct page_to_node)) - 1;
1228 for (chunk_start = 0;
1229 chunk_start < nr_pages;
1230 chunk_start += chunk_nr_pages) {
1233 if (chunk_start + chunk_nr_pages > nr_pages)
1234 chunk_nr_pages = nr_pages - chunk_start;
1236 /* fill the chunk pm with addrs and nodes from user-space */
1237 for (j = 0; j < chunk_nr_pages; j++) {
1238 const void __user *p;
1242 if (get_user(p, pages + j + chunk_start))
1244 pm[j].addr = (unsigned long) p;
1246 if (get_user(node, nodes + j + chunk_start))
1250 if (node < 0 || node >= MAX_NUMNODES)
1253 if (!node_state(node, N_MEMORY))
1257 if (!node_isset(node, task_nodes))
1263 /* End marker for this chunk */
1264 pm[chunk_nr_pages].node = MAX_NUMNODES;
1266 /* Migrate this chunk */
1267 err = do_move_page_to_node_array(mm, pm,
1268 flags & MPOL_MF_MOVE_ALL);
1272 /* Return status information */
1273 for (j = 0; j < chunk_nr_pages; j++)
1274 if (put_user(pm[j].status, status + j + chunk_start)) {
1282 free_page((unsigned long)pm);
1288 * Determine the nodes of an array of pages and store it in an array of status.
1290 static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages,
1291 const void __user **pages, int *status)
1295 down_read(&mm->mmap_sem);
1297 for (i = 0; i < nr_pages; i++) {
1298 unsigned long addr = (unsigned long)(*pages);
1299 struct vm_area_struct *vma;
1303 vma = find_vma(mm, addr);
1304 if (!vma || addr < vma->vm_start)
1307 page = follow_page(vma, addr, 0);
1309 err = PTR_ERR(page);
1314 /* Use PageReserved to check for zero page */
1315 if (!page || PageReserved(page) || PageKsm(page))
1318 err = page_to_nid(page);
1326 up_read(&mm->mmap_sem);
1330 * Determine the nodes of a user array of pages and store it in
1331 * a user array of status.
1333 static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages,
1334 const void __user * __user *pages,
1337 #define DO_PAGES_STAT_CHUNK_NR 16
1338 const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR];
1339 int chunk_status[DO_PAGES_STAT_CHUNK_NR];
1342 unsigned long chunk_nr;
1344 chunk_nr = nr_pages;
1345 if (chunk_nr > DO_PAGES_STAT_CHUNK_NR)
1346 chunk_nr = DO_PAGES_STAT_CHUNK_NR;
1348 if (copy_from_user(chunk_pages, pages, chunk_nr * sizeof(*chunk_pages)))
1351 do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status);
1353 if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status)))
1358 nr_pages -= chunk_nr;
1360 return nr_pages ? -EFAULT : 0;
1364 * Move a list of pages in the address space of the currently executing
1367 SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages,
1368 const void __user * __user *, pages,
1369 const int __user *, nodes,
1370 int __user *, status, int, flags)
1372 const struct cred *cred = current_cred(), *tcred;
1373 struct task_struct *task;
1374 struct mm_struct *mm;
1376 nodemask_t task_nodes;
1379 if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
1382 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1385 /* Find the mm_struct */
1387 task = pid ? find_task_by_vpid(pid) : current;
1392 get_task_struct(task);
1395 * Check if this process has the right to modify the specified
1396 * process. The right exists if the process has administrative
1397 * capabilities, superuser privileges or the same
1398 * userid as the target process.
1400 tcred = __task_cred(task);
1401 if (!uid_eq(cred->euid, tcred->suid) && !uid_eq(cred->euid, tcred->uid) &&
1402 !uid_eq(cred->uid, tcred->suid) && !uid_eq(cred->uid, tcred->uid) &&
1403 !capable(CAP_SYS_NICE)) {
1410 err = security_task_movememory(task);
1414 task_nodes = cpuset_mems_allowed(task);
1415 mm = get_task_mm(task);
1416 put_task_struct(task);
1422 err = do_pages_move(mm, task_nodes, nr_pages, pages,
1423 nodes, status, flags);
1425 err = do_pages_stat(mm, nr_pages, pages, status);
1431 put_task_struct(task);
1436 * Call migration functions in the vma_ops that may prepare
1437 * memory in a vm for migration. migration functions may perform
1438 * the migration for vmas that do not have an underlying page struct.
1440 int migrate_vmas(struct mm_struct *mm, const nodemask_t *to,
1441 const nodemask_t *from, unsigned long flags)
1443 struct vm_area_struct *vma;
1446 for (vma = mm->mmap; vma && !err; vma = vma->vm_next) {
1447 if (vma->vm_ops && vma->vm_ops->migrate) {
1448 err = vma->vm_ops->migrate(vma, to, from, flags);
1456 #ifdef CONFIG_NUMA_BALANCING
1458 * Returns true if this is a safe migration target node for misplaced NUMA
1459 * pages. Currently it only checks the watermarks which crude
1461 static bool migrate_balanced_pgdat(struct pglist_data *pgdat,
1462 int nr_migrate_pages)
1465 for (z = pgdat->nr_zones - 1; z >= 0; z--) {
1466 struct zone *zone = pgdat->node_zones + z;
1468 if (!populated_zone(zone))
1471 if (zone->all_unreclaimable)
1474 /* Avoid waking kswapd by allocating pages_to_migrate pages. */
1475 if (!zone_watermark_ok(zone, 0,
1476 high_wmark_pages(zone) +
1485 static struct page *alloc_misplaced_dst_page(struct page *page,
1489 int nid = (int) data;
1490 struct page *newpage;
1492 newpage = alloc_pages_exact_node(nid,
1493 (GFP_HIGHUSER_MOVABLE | GFP_THISNODE |
1494 __GFP_NOMEMALLOC | __GFP_NORETRY |
1498 page_xchg_last_nid(newpage, page_last_nid(page));
1504 * page migration rate limiting control.
1505 * Do not migrate more than @pages_to_migrate in a @migrate_interval_millisecs
1506 * window of time. Default here says do not migrate more than 1280M per second.
1507 * If a node is rate-limited then PTE NUMA updates are also rate-limited. However
1508 * as it is faults that reset the window, pte updates will happen unconditionally
1509 * if there has not been a fault since @pteupdate_interval_millisecs after the
1510 * throttle window closed.
1512 static unsigned int migrate_interval_millisecs __read_mostly = 100;
1513 static unsigned int pteupdate_interval_millisecs __read_mostly = 1000;
1514 static unsigned int ratelimit_pages __read_mostly = 128 << (20 - PAGE_SHIFT);
1516 /* Returns true if NUMA migration is currently rate limited */
1517 bool migrate_ratelimited(int node)
1519 pg_data_t *pgdat = NODE_DATA(node);
1521 if (time_after(jiffies, pgdat->numabalancing_migrate_next_window +
1522 msecs_to_jiffies(pteupdate_interval_millisecs)))
1525 if (pgdat->numabalancing_migrate_nr_pages < ratelimit_pages)
1531 /* Returns true if the node is migrate rate-limited after the update */
1532 bool numamigrate_update_ratelimit(pg_data_t *pgdat, unsigned long nr_pages)
1534 bool rate_limited = false;
1537 * Rate-limit the amount of data that is being migrated to a node.
1538 * Optimal placement is no good if the memory bus is saturated and
1539 * all the time is being spent migrating!
1541 spin_lock(&pgdat->numabalancing_migrate_lock);
1542 if (time_after(jiffies, pgdat->numabalancing_migrate_next_window)) {
1543 pgdat->numabalancing_migrate_nr_pages = 0;
1544 pgdat->numabalancing_migrate_next_window = jiffies +
1545 msecs_to_jiffies(migrate_interval_millisecs);
1547 if (pgdat->numabalancing_migrate_nr_pages > ratelimit_pages)
1548 rate_limited = true;
1550 pgdat->numabalancing_migrate_nr_pages += nr_pages;
1551 spin_unlock(&pgdat->numabalancing_migrate_lock);
1553 return rate_limited;
1556 int numamigrate_isolate_page(pg_data_t *pgdat, struct page *page)
1560 /* Avoid migrating to a node that is nearly full */
1561 if (migrate_balanced_pgdat(pgdat, 1)) {
1564 if (isolate_lru_page(page)) {
1569 /* Page is isolated */
1571 page_lru = page_is_file_cache(page);
1572 if (!PageTransHuge(page))
1573 inc_zone_page_state(page, NR_ISOLATED_ANON + page_lru);
1575 mod_zone_page_state(page_zone(page),
1576 NR_ISOLATED_ANON + page_lru,
1581 * Page is either isolated or there is not enough space on the target
1582 * node. If isolated, then it has taken a reference count and the
1583 * callers reference can be safely dropped without the page
1584 * disappearing underneath us during migration. Otherwise the page is
1585 * not to be migrated but the callers reference should still be
1586 * dropped so it does not leak.
1594 * Attempt to migrate a misplaced page to the specified destination
1595 * node. Caller is expected to have an elevated reference count on
1596 * the page that will be dropped by this function before returning.
1598 int migrate_misplaced_page(struct page *page, int node)
1600 pg_data_t *pgdat = NODE_DATA(node);
1603 LIST_HEAD(migratepages);
1606 * Don't migrate pages that are mapped in multiple processes.
1607 * TODO: Handle false sharing detection instead of this hammer
1609 if (page_mapcount(page) != 1) {
1615 * Rate-limit the amount of data that is being migrated to a node.
1616 * Optimal placement is no good if the memory bus is saturated and
1617 * all the time is being spent migrating!
1619 if (numamigrate_update_ratelimit(pgdat, 1)) {
1624 isolated = numamigrate_isolate_page(pgdat, page);
1628 list_add(&page->lru, &migratepages);
1629 nr_remaining = migrate_pages(&migratepages,
1630 alloc_misplaced_dst_page,
1631 node, false, MIGRATE_ASYNC,
1634 putback_lru_pages(&migratepages);
1637 count_vm_numa_event(NUMA_PAGE_MIGRATE);
1638 BUG_ON(!list_empty(&migratepages));
1642 #endif /* CONFIG_NUMA_BALANCING */
1644 #if defined(CONFIG_NUMA_BALANCING) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
1645 int migrate_misplaced_transhuge_page(struct mm_struct *mm,
1646 struct vm_area_struct *vma,
1647 pmd_t *pmd, pmd_t entry,
1648 unsigned long address,
1649 struct page *page, int node)
1651 unsigned long haddr = address & HPAGE_PMD_MASK;
1652 pg_data_t *pgdat = NODE_DATA(node);
1654 struct page *new_page = NULL;
1655 struct mem_cgroup *memcg = NULL;
1656 int page_lru = page_is_file_cache(page);
1659 * Don't migrate pages that are mapped in multiple processes.
1660 * TODO: Handle false sharing detection instead of this hammer
1662 if (page_mapcount(page) != 1)
1666 * Rate-limit the amount of data that is being migrated to a node.
1667 * Optimal placement is no good if the memory bus is saturated and
1668 * all the time is being spent migrating!
1670 if (numamigrate_update_ratelimit(pgdat, HPAGE_PMD_NR))
1673 new_page = alloc_pages_node(node,
1674 (GFP_TRANSHUGE | GFP_THISNODE) & ~__GFP_WAIT, HPAGE_PMD_ORDER);
1676 count_vm_events(PGMIGRATE_FAIL, HPAGE_PMD_NR);
1679 page_xchg_last_nid(new_page, page_last_nid(page));
1681 isolated = numamigrate_isolate_page(pgdat, page);
1683 count_vm_events(PGMIGRATE_FAIL, HPAGE_PMD_NR);
1685 goto out_keep_locked;
1688 /* Prepare a page as a migration target */
1689 __set_page_locked(new_page);
1690 SetPageSwapBacked(new_page);
1692 /* anon mapping, we can simply copy page->mapping to the new page: */
1693 new_page->mapping = page->mapping;
1694 new_page->index = page->index;
1695 migrate_page_copy(new_page, page);
1696 WARN_ON(PageLRU(new_page));
1698 /* Recheck the target PMD */
1699 spin_lock(&mm->page_table_lock);
1700 if (unlikely(!pmd_same(*pmd, entry))) {
1701 spin_unlock(&mm->page_table_lock);
1703 /* Reverse changes made by migrate_page_copy() */
1704 if (TestClearPageActive(new_page))
1705 SetPageActive(page);
1706 if (TestClearPageUnevictable(new_page))
1707 SetPageUnevictable(page);
1708 mlock_migrate_page(page, new_page);
1710 unlock_page(new_page);
1711 put_page(new_page); /* Free it */
1714 putback_lru_page(page);
1716 count_vm_events(PGMIGRATE_FAIL, HPAGE_PMD_NR);
1721 * Traditional migration needs to prepare the memcg charge
1722 * transaction early to prevent the old page from being
1723 * uncharged when installing migration entries. Here we can
1724 * save the potential rollback and start the charge transfer
1725 * only when migration is already known to end successfully.
1727 mem_cgroup_prepare_migration(page, new_page, &memcg);
1729 entry = mk_pmd(new_page, vma->vm_page_prot);
1730 entry = pmd_mknonnuma(entry);
1731 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1732 entry = pmd_mkhuge(entry);
1734 page_add_new_anon_rmap(new_page, vma, haddr);
1736 set_pmd_at(mm, haddr, pmd, entry);
1737 update_mmu_cache_pmd(vma, address, &entry);
1738 page_remove_rmap(page);
1740 * Finish the charge transaction under the page table lock to
1741 * prevent split_huge_page() from dividing up the charge
1742 * before it's fully transferred to the new page.
1744 mem_cgroup_end_migration(memcg, page, new_page, true);
1745 spin_unlock(&mm->page_table_lock);
1747 unlock_page(new_page);
1749 put_page(page); /* Drop the rmap reference */
1750 put_page(page); /* Drop the LRU isolation reference */
1752 count_vm_events(PGMIGRATE_SUCCESS, HPAGE_PMD_NR);
1753 count_vm_numa_events(NUMA_PAGE_MIGRATE, HPAGE_PMD_NR);
1756 mod_zone_page_state(page_zone(page),
1757 NR_ISOLATED_ANON + page_lru,
1766 #endif /* CONFIG_NUMA_BALANCING */
1768 #endif /* CONFIG_NUMA */