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>
45 * migrate_prep() needs to be called before we start compiling a list of pages
46 * to be migrated using isolate_lru_page(). If scheduling work on other CPUs is
47 * undesirable, use migrate_prep_local()
49 int migrate_prep(void)
52 * Clear the LRU lists so pages can be isolated.
53 * Note that pages may be moved off the LRU after we have
54 * drained them. Those pages will fail to migrate like other
55 * pages that may be busy.
62 /* Do the necessary work of migrate_prep but not if it involves other CPUs */
63 int migrate_prep_local(void)
71 * Add isolated pages on the list back to the LRU under page lock
72 * to avoid leaking evictable pages back onto unevictable list.
74 void putback_lru_pages(struct list_head *l)
79 list_for_each_entry_safe(page, page2, l, lru) {
81 dec_zone_page_state(page, NR_ISOLATED_ANON +
82 page_is_file_cache(page));
83 if (unlikely(balloon_page_movable(page)))
84 balloon_page_putback(page);
86 putback_lru_page(page);
91 * Restore a potential migration pte to a working pte entry
93 static int remove_migration_pte(struct page *new, struct vm_area_struct *vma,
94 unsigned long addr, void *old)
96 struct mm_struct *mm = vma->vm_mm;
102 if (unlikely(PageHuge(new))) {
103 ptep = huge_pte_offset(mm, addr);
106 ptl = &mm->page_table_lock;
108 pmd = mm_find_pmd(mm, addr);
111 if (pmd_trans_huge(*pmd))
114 ptep = pte_offset_map(pmd, addr);
117 * Peek to check is_swap_pte() before taking ptlock? No, we
118 * can race mremap's move_ptes(), which skips anon_vma lock.
121 ptl = pte_lockptr(mm, pmd);
126 if (!is_swap_pte(pte))
129 entry = pte_to_swp_entry(pte);
131 if (!is_migration_entry(entry) ||
132 migration_entry_to_page(entry) != old)
136 pte = pte_mkold(mk_pte(new, vma->vm_page_prot));
137 if (is_write_migration_entry(entry))
138 pte = pte_mkwrite(pte);
139 #ifdef CONFIG_HUGETLB_PAGE
141 pte = pte_mkhuge(pte);
143 flush_cache_page(vma, addr, pte_pfn(pte));
144 set_pte_at(mm, addr, ptep, pte);
148 hugepage_add_anon_rmap(new, vma, addr);
151 } else if (PageAnon(new))
152 page_add_anon_rmap(new, vma, addr);
154 page_add_file_rmap(new);
156 /* No need to invalidate - it was non-present before */
157 update_mmu_cache(vma, addr, ptep);
159 pte_unmap_unlock(ptep, ptl);
165 * Get rid of all migration entries and replace them by
166 * references to the indicated page.
168 static void remove_migration_ptes(struct page *old, struct page *new)
170 rmap_walk(new, remove_migration_pte, old);
174 * Something used the pte of a page under migration. We need to
175 * get to the page and wait until migration is finished.
176 * When we return from this function the fault will be retried.
178 void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
179 unsigned long address)
186 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
188 if (!is_swap_pte(pte))
191 entry = pte_to_swp_entry(pte);
192 if (!is_migration_entry(entry))
195 page = migration_entry_to_page(entry);
198 * Once radix-tree replacement of page migration started, page_count
199 * *must* be zero. And, we don't want to call wait_on_page_locked()
200 * against a page without get_page().
201 * So, we use get_page_unless_zero(), here. Even failed, page fault
204 if (!get_page_unless_zero(page))
206 pte_unmap_unlock(ptep, ptl);
207 wait_on_page_locked(page);
211 pte_unmap_unlock(ptep, ptl);
215 /* Returns true if all buffers are successfully locked */
216 static bool buffer_migrate_lock_buffers(struct buffer_head *head,
217 enum migrate_mode mode)
219 struct buffer_head *bh = head;
221 /* Simple case, sync compaction */
222 if (mode != MIGRATE_ASYNC) {
226 bh = bh->b_this_page;
228 } while (bh != head);
233 /* async case, we cannot block on lock_buffer so use trylock_buffer */
236 if (!trylock_buffer(bh)) {
238 * We failed to lock the buffer and cannot stall in
239 * async migration. Release the taken locks
241 struct buffer_head *failed_bh = bh;
244 while (bh != failed_bh) {
247 bh = bh->b_this_page;
252 bh = bh->b_this_page;
253 } while (bh != head);
257 static inline bool buffer_migrate_lock_buffers(struct buffer_head *head,
258 enum migrate_mode mode)
262 #endif /* CONFIG_BLOCK */
265 * Replace the page in the mapping.
267 * The number of remaining references must be:
268 * 1 for anonymous pages without a mapping
269 * 2 for pages with a mapping
270 * 3 for pages with a mapping and PagePrivate/PagePrivate2 set.
272 static int migrate_page_move_mapping(struct address_space *mapping,
273 struct page *newpage, struct page *page,
274 struct buffer_head *head, enum migrate_mode mode)
280 /* Anonymous page without mapping */
281 if (page_count(page) != 1)
283 return MIGRATEPAGE_SUCCESS;
286 spin_lock_irq(&mapping->tree_lock);
288 pslot = radix_tree_lookup_slot(&mapping->page_tree,
291 expected_count = 2 + page_has_private(page);
292 if (page_count(page) != expected_count ||
293 radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) {
294 spin_unlock_irq(&mapping->tree_lock);
298 if (!page_freeze_refs(page, expected_count)) {
299 spin_unlock_irq(&mapping->tree_lock);
304 * In the async migration case of moving a page with buffers, lock the
305 * buffers using trylock before the mapping is moved. If the mapping
306 * was moved, we later failed to lock the buffers and could not move
307 * the mapping back due to an elevated page count, we would have to
308 * block waiting on other references to be dropped.
310 if (mode == MIGRATE_ASYNC && head &&
311 !buffer_migrate_lock_buffers(head, mode)) {
312 page_unfreeze_refs(page, expected_count);
313 spin_unlock_irq(&mapping->tree_lock);
318 * Now we know that no one else is looking at the page.
320 get_page(newpage); /* add cache reference */
321 if (PageSwapCache(page)) {
322 SetPageSwapCache(newpage);
323 set_page_private(newpage, page_private(page));
326 radix_tree_replace_slot(pslot, newpage);
329 * Drop cache reference from old page by unfreezing
330 * to one less reference.
331 * We know this isn't the last reference.
333 page_unfreeze_refs(page, expected_count - 1);
336 * If moved to a different zone then also account
337 * the page for that zone. Other VM counters will be
338 * taken care of when we establish references to the
339 * new page and drop references to the old page.
341 * Note that anonymous pages are accounted for
342 * via NR_FILE_PAGES and NR_ANON_PAGES if they
343 * are mapped to swap space.
345 __dec_zone_page_state(page, NR_FILE_PAGES);
346 __inc_zone_page_state(newpage, NR_FILE_PAGES);
347 if (!PageSwapCache(page) && PageSwapBacked(page)) {
348 __dec_zone_page_state(page, NR_SHMEM);
349 __inc_zone_page_state(newpage, NR_SHMEM);
351 spin_unlock_irq(&mapping->tree_lock);
353 return MIGRATEPAGE_SUCCESS;
357 * The expected number of remaining references is the same as that
358 * of migrate_page_move_mapping().
360 int migrate_huge_page_move_mapping(struct address_space *mapping,
361 struct page *newpage, struct page *page)
367 if (page_count(page) != 1)
369 return MIGRATEPAGE_SUCCESS;
372 spin_lock_irq(&mapping->tree_lock);
374 pslot = radix_tree_lookup_slot(&mapping->page_tree,
377 expected_count = 2 + page_has_private(page);
378 if (page_count(page) != expected_count ||
379 radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) {
380 spin_unlock_irq(&mapping->tree_lock);
384 if (!page_freeze_refs(page, expected_count)) {
385 spin_unlock_irq(&mapping->tree_lock);
391 radix_tree_replace_slot(pslot, newpage);
393 page_unfreeze_refs(page, expected_count - 1);
395 spin_unlock_irq(&mapping->tree_lock);
396 return MIGRATEPAGE_SUCCESS;
400 * Copy the page to its new location
402 void migrate_page_copy(struct page *newpage, struct page *page)
405 copy_huge_page(newpage, page);
407 copy_highpage(newpage, page);
410 SetPageError(newpage);
411 if (PageReferenced(page))
412 SetPageReferenced(newpage);
413 if (PageUptodate(page))
414 SetPageUptodate(newpage);
415 if (TestClearPageActive(page)) {
416 VM_BUG_ON(PageUnevictable(page));
417 SetPageActive(newpage);
418 } else if (TestClearPageUnevictable(page))
419 SetPageUnevictable(newpage);
420 if (PageChecked(page))
421 SetPageChecked(newpage);
422 if (PageMappedToDisk(page))
423 SetPageMappedToDisk(newpage);
425 if (PageDirty(page)) {
426 clear_page_dirty_for_io(page);
428 * Want to mark the page and the radix tree as dirty, and
429 * redo the accounting that clear_page_dirty_for_io undid,
430 * but we can't use set_page_dirty because that function
431 * is actually a signal that all of the page has become dirty.
432 * Whereas only part of our page may be dirty.
434 if (PageSwapBacked(page))
435 SetPageDirty(newpage);
437 __set_page_dirty_nobuffers(newpage);
440 mlock_migrate_page(newpage, page);
441 ksm_migrate_page(newpage, page);
443 ClearPageSwapCache(page);
444 ClearPagePrivate(page);
445 set_page_private(page, 0);
448 * If any waiters have accumulated on the new page then
451 if (PageWriteback(newpage))
452 end_page_writeback(newpage);
455 /************************************************************
456 * Migration functions
457 ***********************************************************/
459 /* Always fail migration. Used for mappings that are not movable */
460 int fail_migrate_page(struct address_space *mapping,
461 struct page *newpage, struct page *page)
465 EXPORT_SYMBOL(fail_migrate_page);
468 * Common logic to directly migrate a single page suitable for
469 * pages that do not use PagePrivate/PagePrivate2.
471 * Pages are locked upon entry and exit.
473 int migrate_page(struct address_space *mapping,
474 struct page *newpage, struct page *page,
475 enum migrate_mode mode)
479 BUG_ON(PageWriteback(page)); /* Writeback must be complete */
481 rc = migrate_page_move_mapping(mapping, newpage, page, NULL, mode);
483 if (rc != MIGRATEPAGE_SUCCESS)
486 migrate_page_copy(newpage, page);
487 return MIGRATEPAGE_SUCCESS;
489 EXPORT_SYMBOL(migrate_page);
493 * Migration function for pages with buffers. This function can only be used
494 * if the underlying filesystem guarantees that no other references to "page"
497 int buffer_migrate_page(struct address_space *mapping,
498 struct page *newpage, struct page *page, enum migrate_mode mode)
500 struct buffer_head *bh, *head;
503 if (!page_has_buffers(page))
504 return migrate_page(mapping, newpage, page, mode);
506 head = page_buffers(page);
508 rc = migrate_page_move_mapping(mapping, newpage, page, head, mode);
510 if (rc != MIGRATEPAGE_SUCCESS)
514 * In the async case, migrate_page_move_mapping locked the buffers
515 * with an IRQ-safe spinlock held. In the sync case, the buffers
516 * need to be locked now
518 if (mode != MIGRATE_ASYNC)
519 BUG_ON(!buffer_migrate_lock_buffers(head, mode));
521 ClearPagePrivate(page);
522 set_page_private(newpage, page_private(page));
523 set_page_private(page, 0);
529 set_bh_page(bh, newpage, bh_offset(bh));
530 bh = bh->b_this_page;
532 } while (bh != head);
534 SetPagePrivate(newpage);
536 migrate_page_copy(newpage, page);
542 bh = bh->b_this_page;
544 } while (bh != head);
546 return MIGRATEPAGE_SUCCESS;
548 EXPORT_SYMBOL(buffer_migrate_page);
552 * Writeback a page to clean the dirty state
554 static int writeout(struct address_space *mapping, struct page *page)
556 struct writeback_control wbc = {
557 .sync_mode = WB_SYNC_NONE,
560 .range_end = LLONG_MAX,
565 if (!mapping->a_ops->writepage)
566 /* No write method for the address space */
569 if (!clear_page_dirty_for_io(page))
570 /* Someone else already triggered a write */
574 * A dirty page may imply that the underlying filesystem has
575 * the page on some queue. So the page must be clean for
576 * migration. Writeout may mean we loose the lock and the
577 * page state is no longer what we checked for earlier.
578 * At this point we know that the migration attempt cannot
581 remove_migration_ptes(page, page);
583 rc = mapping->a_ops->writepage(page, &wbc);
585 if (rc != AOP_WRITEPAGE_ACTIVATE)
586 /* unlocked. Relock */
589 return (rc < 0) ? -EIO : -EAGAIN;
593 * Default handling if a filesystem does not provide a migration function.
595 static int fallback_migrate_page(struct address_space *mapping,
596 struct page *newpage, struct page *page, enum migrate_mode mode)
598 if (PageDirty(page)) {
599 /* Only writeback pages in full synchronous migration */
600 if (mode != MIGRATE_SYNC)
602 return writeout(mapping, page);
606 * Buffers may be managed in a filesystem specific way.
607 * We must have no buffers or drop them.
609 if (page_has_private(page) &&
610 !try_to_release_page(page, GFP_KERNEL))
613 return migrate_page(mapping, newpage, page, mode);
617 * Move a page to a newly allocated page
618 * The page is locked and all ptes have been successfully removed.
620 * The new page will have replaced the old page if this function
625 * MIGRATEPAGE_SUCCESS - success
627 static int move_to_new_page(struct page *newpage, struct page *page,
628 int remap_swapcache, enum migrate_mode mode)
630 struct address_space *mapping;
634 * Block others from accessing the page when we get around to
635 * establishing additional references. We are the only one
636 * holding a reference to the new page at this point.
638 if (!trylock_page(newpage))
641 /* Prepare mapping for the new page.*/
642 newpage->index = page->index;
643 newpage->mapping = page->mapping;
644 if (PageSwapBacked(page))
645 SetPageSwapBacked(newpage);
647 mapping = page_mapping(page);
649 rc = migrate_page(mapping, newpage, page, mode);
650 else if (mapping->a_ops->migratepage)
652 * Most pages have a mapping and most filesystems provide a
653 * migratepage callback. Anonymous pages are part of swap
654 * space which also has its own migratepage callback. This
655 * is the most common path for page migration.
657 rc = mapping->a_ops->migratepage(mapping,
658 newpage, page, mode);
660 rc = fallback_migrate_page(mapping, newpage, page, mode);
662 if (rc != MIGRATEPAGE_SUCCESS) {
663 newpage->mapping = NULL;
666 remove_migration_ptes(page, newpage);
667 page->mapping = NULL;
670 unlock_page(newpage);
675 static int __unmap_and_move(struct page *page, struct page *newpage,
676 int force, bool offlining, enum migrate_mode mode)
679 int remap_swapcache = 1;
680 struct mem_cgroup *mem;
681 struct anon_vma *anon_vma = NULL;
683 if (!trylock_page(page)) {
684 if (!force || mode == MIGRATE_ASYNC)
688 * It's not safe for direct compaction to call lock_page.
689 * For example, during page readahead pages are added locked
690 * to the LRU. Later, when the IO completes the pages are
691 * marked uptodate and unlocked. However, the queueing
692 * could be merging multiple pages for one bio (e.g.
693 * mpage_readpages). If an allocation happens for the
694 * second or third page, the process can end up locking
695 * the same page twice and deadlocking. Rather than
696 * trying to be clever about what pages can be locked,
697 * avoid the use of lock_page for direct compaction
700 if (current->flags & PF_MEMALLOC)
707 * Only memory hotplug's offline_pages() caller has locked out KSM,
708 * and can safely migrate a KSM page. The other cases have skipped
709 * PageKsm along with PageReserved - but it is only now when we have
710 * the page lock that we can be certain it will not go KSM beneath us
711 * (KSM will not upgrade a page from PageAnon to PageKsm when it sees
712 * its pagecount raised, but only here do we take the page lock which
715 if (PageKsm(page) && !offlining) {
720 /* charge against new page */
721 mem_cgroup_prepare_migration(page, newpage, &mem);
723 if (PageWriteback(page)) {
725 * Only in the case of a full syncronous migration is it
726 * necessary to wait for PageWriteback. In the async case,
727 * the retry loop is too short and in the sync-light case,
728 * the overhead of stalling is too much
730 if (mode != MIGRATE_SYNC) {
736 wait_on_page_writeback(page);
739 * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
740 * we cannot notice that anon_vma is freed while we migrates a page.
741 * This get_anon_vma() delays freeing anon_vma pointer until the end
742 * of migration. File cache pages are no problem because of page_lock()
743 * File Caches may use write_page() or lock_page() in migration, then,
744 * just care Anon page here.
746 if (PageAnon(page)) {
748 * Only page_lock_anon_vma() understands the subtleties of
749 * getting a hold on an anon_vma from outside one of its mms.
751 anon_vma = page_get_anon_vma(page);
756 } else if (PageSwapCache(page)) {
758 * We cannot be sure that the anon_vma of an unmapped
759 * swapcache page is safe to use because we don't
760 * know in advance if the VMA that this page belonged
761 * to still exists. If the VMA and others sharing the
762 * data have been freed, then the anon_vma could
763 * already be invalid.
765 * To avoid this possibility, swapcache pages get
766 * migrated but are not remapped when migration
775 if (unlikely(balloon_page_movable(page))) {
777 * A ballooned page does not need any special attention from
778 * physical to virtual reverse mapping procedures.
779 * Skip any attempt to unmap PTEs or to remap swap cache,
780 * in order to avoid burning cycles at rmap level, and perform
781 * the page migration right away (proteced by page lock).
783 rc = balloon_page_migrate(newpage, page, mode);
788 * Corner case handling:
789 * 1. When a new swap-cache page is read into, it is added to the LRU
790 * and treated as swapcache but it has no rmap yet.
791 * Calling try_to_unmap() against a page->mapping==NULL page will
792 * trigger a BUG. So handle it here.
793 * 2. An orphaned page (see truncate_complete_page) might have
794 * fs-private metadata. The page can be picked up due to memory
795 * offlining. Everywhere else except page reclaim, the page is
796 * invisible to the vm, so the page can not be migrated. So try to
797 * free the metadata, so the page can be freed.
799 if (!page->mapping) {
800 VM_BUG_ON(PageAnon(page));
801 if (page_has_private(page)) {
802 try_to_free_buffers(page);
808 /* Establish migration ptes or remove ptes */
809 try_to_unmap(page, TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
812 if (!page_mapped(page))
813 rc = move_to_new_page(newpage, page, remap_swapcache, mode);
815 if (rc && remap_swapcache)
816 remove_migration_ptes(page, page);
818 /* Drop an anon_vma reference if we took one */
820 put_anon_vma(anon_vma);
823 mem_cgroup_end_migration(mem, page, newpage,
824 (rc == MIGRATEPAGE_SUCCESS ||
825 rc == MIGRATEPAGE_BALLOON_SUCCESS));
833 * Obtain the lock on page, remove all ptes and migrate the page
834 * to the newly allocated page in newpage.
836 static int unmap_and_move(new_page_t get_new_page, unsigned long private,
837 struct page *page, int force, bool offlining,
838 enum migrate_mode mode)
842 struct page *newpage = get_new_page(page, private, &result);
847 if (page_count(page) == 1) {
848 /* page was freed from under us. So we are done. */
852 if (unlikely(PageTransHuge(page)))
853 if (unlikely(split_huge_page(page)))
856 rc = __unmap_and_move(page, newpage, force, offlining, mode);
858 if (unlikely(rc == MIGRATEPAGE_BALLOON_SUCCESS)) {
860 * A ballooned page has been migrated already.
861 * Now, it's the time to wrap-up counters,
862 * handle the page back to Buddy and return.
864 dec_zone_page_state(page, NR_ISOLATED_ANON +
865 page_is_file_cache(page));
866 balloon_page_free(page);
867 return MIGRATEPAGE_SUCCESS;
872 * A page that has been migrated has all references
873 * removed and will be freed. A page that has not been
874 * migrated will have kepts its references and be
877 list_del(&page->lru);
878 dec_zone_page_state(page, NR_ISOLATED_ANON +
879 page_is_file_cache(page));
880 putback_lru_page(page);
883 * Move the new page to the LRU. If migration was not successful
884 * then this will free the page.
886 putback_lru_page(newpage);
891 *result = page_to_nid(newpage);
897 * Counterpart of unmap_and_move_page() for hugepage migration.
899 * This function doesn't wait the completion of hugepage I/O
900 * because there is no race between I/O and migration for hugepage.
901 * Note that currently hugepage I/O occurs only in direct I/O
902 * where no lock is held and PG_writeback is irrelevant,
903 * and writeback status of all subpages are counted in the reference
904 * count of the head page (i.e. if all subpages of a 2MB hugepage are
905 * under direct I/O, the reference of the head page is 512 and a bit more.)
906 * This means that when we try to migrate hugepage whose subpages are
907 * doing direct I/O, some references remain after try_to_unmap() and
908 * hugepage migration fails without data corruption.
910 * There is also no race when direct I/O is issued on the page under migration,
911 * because then pte is replaced with migration swap entry and direct I/O code
912 * will wait in the page fault for migration to complete.
914 static int unmap_and_move_huge_page(new_page_t get_new_page,
915 unsigned long private, struct page *hpage,
916 int force, bool offlining,
917 enum migrate_mode mode)
921 struct page *new_hpage = get_new_page(hpage, private, &result);
922 struct anon_vma *anon_vma = NULL;
929 if (!trylock_page(hpage)) {
930 if (!force || mode != MIGRATE_SYNC)
936 anon_vma = page_get_anon_vma(hpage);
938 try_to_unmap(hpage, TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
940 if (!page_mapped(hpage))
941 rc = move_to_new_page(new_hpage, hpage, 1, mode);
944 remove_migration_ptes(hpage, hpage);
947 put_anon_vma(anon_vma);
950 hugetlb_cgroup_migrate(hpage, new_hpage);
959 *result = page_to_nid(new_hpage);
967 * The function takes one list of pages to migrate and a function
968 * that determines from the page to be migrated and the private data
969 * the target of the move and allocates the page.
971 * The function returns after 10 attempts or if no pages
972 * are movable anymore because to has become empty
973 * or no retryable pages exist anymore.
974 * Caller should call putback_lru_pages to return pages to the LRU
975 * or free list only if ret != 0.
977 * Return: Number of pages not migrated or error code.
979 int migrate_pages(struct list_head *from,
980 new_page_t get_new_page, unsigned long private, bool offlining,
981 enum migrate_mode mode)
988 int swapwrite = current->flags & PF_SWAPWRITE;
992 current->flags |= PF_SWAPWRITE;
994 for(pass = 0; pass < 10 && retry; pass++) {
997 list_for_each_entry_safe(page, page2, from, lru) {
1000 rc = unmap_and_move(get_new_page, private,
1001 page, pass > 2, offlining,
1010 case MIGRATEPAGE_SUCCESS:
1013 /* Permanent failure */
1019 rc = nr_failed + retry;
1022 current->flags &= ~PF_SWAPWRITE;
1027 int migrate_huge_page(struct page *hpage, new_page_t get_new_page,
1028 unsigned long private, bool offlining,
1029 enum migrate_mode mode)
1033 for (pass = 0; pass < 10; pass++) {
1034 rc = unmap_and_move_huge_page(get_new_page,
1035 private, hpage, pass > 2, offlining,
1044 case MIGRATEPAGE_SUCCESS:
1057 * Move a list of individual pages
1059 struct page_to_node {
1066 static struct page *new_page_node(struct page *p, unsigned long private,
1069 struct page_to_node *pm = (struct page_to_node *)private;
1071 while (pm->node != MAX_NUMNODES && pm->page != p)
1074 if (pm->node == MAX_NUMNODES)
1077 *result = &pm->status;
1079 return alloc_pages_exact_node(pm->node,
1080 GFP_HIGHUSER_MOVABLE | GFP_THISNODE, 0);
1084 * Move a set of pages as indicated in the pm array. The addr
1085 * field must be set to the virtual address of the page to be moved
1086 * and the node number must contain a valid target node.
1087 * The pm array ends with node = MAX_NUMNODES.
1089 static int do_move_page_to_node_array(struct mm_struct *mm,
1090 struct page_to_node *pm,
1094 struct page_to_node *pp;
1095 LIST_HEAD(pagelist);
1097 down_read(&mm->mmap_sem);
1100 * Build a list of pages to migrate
1102 for (pp = pm; pp->node != MAX_NUMNODES; pp++) {
1103 struct vm_area_struct *vma;
1107 vma = find_vma(mm, pp->addr);
1108 if (!vma || pp->addr < vma->vm_start || !vma_migratable(vma))
1111 page = follow_page(vma, pp->addr, FOLL_GET|FOLL_SPLIT);
1113 err = PTR_ERR(page);
1121 /* Use PageReserved to check for zero page */
1122 if (PageReserved(page) || PageKsm(page))
1126 err = page_to_nid(page);
1128 if (err == pp->node)
1130 * Node already in the right place
1135 if (page_mapcount(page) > 1 &&
1139 err = isolate_lru_page(page);
1141 list_add_tail(&page->lru, &pagelist);
1142 inc_zone_page_state(page, NR_ISOLATED_ANON +
1143 page_is_file_cache(page));
1147 * Either remove the duplicate refcount from
1148 * isolate_lru_page() or drop the page ref if it was
1157 if (!list_empty(&pagelist)) {
1158 err = migrate_pages(&pagelist, new_page_node,
1159 (unsigned long)pm, 0, MIGRATE_SYNC);
1161 putback_lru_pages(&pagelist);
1164 up_read(&mm->mmap_sem);
1169 * Migrate an array of page address onto an array of nodes and fill
1170 * the corresponding array of status.
1172 static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes,
1173 unsigned long nr_pages,
1174 const void __user * __user *pages,
1175 const int __user *nodes,
1176 int __user *status, int flags)
1178 struct page_to_node *pm;
1179 unsigned long chunk_nr_pages;
1180 unsigned long chunk_start;
1184 pm = (struct page_to_node *)__get_free_page(GFP_KERNEL);
1191 * Store a chunk of page_to_node array in a page,
1192 * but keep the last one as a marker
1194 chunk_nr_pages = (PAGE_SIZE / sizeof(struct page_to_node)) - 1;
1196 for (chunk_start = 0;
1197 chunk_start < nr_pages;
1198 chunk_start += chunk_nr_pages) {
1201 if (chunk_start + chunk_nr_pages > nr_pages)
1202 chunk_nr_pages = nr_pages - chunk_start;
1204 /* fill the chunk pm with addrs and nodes from user-space */
1205 for (j = 0; j < chunk_nr_pages; j++) {
1206 const void __user *p;
1210 if (get_user(p, pages + j + chunk_start))
1212 pm[j].addr = (unsigned long) p;
1214 if (get_user(node, nodes + j + chunk_start))
1218 if (node < 0 || node >= MAX_NUMNODES)
1221 if (!node_state(node, N_HIGH_MEMORY))
1225 if (!node_isset(node, task_nodes))
1231 /* End marker for this chunk */
1232 pm[chunk_nr_pages].node = MAX_NUMNODES;
1234 /* Migrate this chunk */
1235 err = do_move_page_to_node_array(mm, pm,
1236 flags & MPOL_MF_MOVE_ALL);
1240 /* Return status information */
1241 for (j = 0; j < chunk_nr_pages; j++)
1242 if (put_user(pm[j].status, status + j + chunk_start)) {
1250 free_page((unsigned long)pm);
1256 * Determine the nodes of an array of pages and store it in an array of status.
1258 static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages,
1259 const void __user **pages, int *status)
1263 down_read(&mm->mmap_sem);
1265 for (i = 0; i < nr_pages; i++) {
1266 unsigned long addr = (unsigned long)(*pages);
1267 struct vm_area_struct *vma;
1271 vma = find_vma(mm, addr);
1272 if (!vma || addr < vma->vm_start)
1275 page = follow_page(vma, addr, 0);
1277 err = PTR_ERR(page);
1282 /* Use PageReserved to check for zero page */
1283 if (!page || PageReserved(page) || PageKsm(page))
1286 err = page_to_nid(page);
1294 up_read(&mm->mmap_sem);
1298 * Determine the nodes of a user array of pages and store it in
1299 * a user array of status.
1301 static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages,
1302 const void __user * __user *pages,
1305 #define DO_PAGES_STAT_CHUNK_NR 16
1306 const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR];
1307 int chunk_status[DO_PAGES_STAT_CHUNK_NR];
1310 unsigned long chunk_nr;
1312 chunk_nr = nr_pages;
1313 if (chunk_nr > DO_PAGES_STAT_CHUNK_NR)
1314 chunk_nr = DO_PAGES_STAT_CHUNK_NR;
1316 if (copy_from_user(chunk_pages, pages, chunk_nr * sizeof(*chunk_pages)))
1319 do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status);
1321 if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status)))
1326 nr_pages -= chunk_nr;
1328 return nr_pages ? -EFAULT : 0;
1332 * Move a list of pages in the address space of the currently executing
1335 SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages,
1336 const void __user * __user *, pages,
1337 const int __user *, nodes,
1338 int __user *, status, int, flags)
1340 const struct cred *cred = current_cred(), *tcred;
1341 struct task_struct *task;
1342 struct mm_struct *mm;
1344 nodemask_t task_nodes;
1347 if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
1350 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1353 /* Find the mm_struct */
1355 task = pid ? find_task_by_vpid(pid) : current;
1360 get_task_struct(task);
1363 * Check if this process has the right to modify the specified
1364 * process. The right exists if the process has administrative
1365 * capabilities, superuser privileges or the same
1366 * userid as the target process.
1368 tcred = __task_cred(task);
1369 if (!uid_eq(cred->euid, tcred->suid) && !uid_eq(cred->euid, tcred->uid) &&
1370 !uid_eq(cred->uid, tcred->suid) && !uid_eq(cred->uid, tcred->uid) &&
1371 !capable(CAP_SYS_NICE)) {
1378 err = security_task_movememory(task);
1382 task_nodes = cpuset_mems_allowed(task);
1383 mm = get_task_mm(task);
1384 put_task_struct(task);
1390 err = do_pages_move(mm, task_nodes, nr_pages, pages,
1391 nodes, status, flags);
1393 err = do_pages_stat(mm, nr_pages, pages, status);
1399 put_task_struct(task);
1404 * Call migration functions in the vma_ops that may prepare
1405 * memory in a vm for migration. migration functions may perform
1406 * the migration for vmas that do not have an underlying page struct.
1408 int migrate_vmas(struct mm_struct *mm, const nodemask_t *to,
1409 const nodemask_t *from, unsigned long flags)
1411 struct vm_area_struct *vma;
1414 for (vma = mm->mmap; vma && !err; vma = vma->vm_next) {
1415 if (vma->vm_ops && vma->vm_ops->migrate) {
1416 err = vma->vm_ops->migrate(vma, to, from, flags);