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/module.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/gfp.h>
38 #include <asm/tlbflush.h>
42 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
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 putback_lru_page(page);
88 * Restore a potential migration pte to a working pte entry
90 static int remove_migration_pte(struct page *new, struct vm_area_struct *vma,
91 unsigned long addr, void *old)
93 struct mm_struct *mm = vma->vm_mm;
101 if (unlikely(PageHuge(new))) {
102 ptep = huge_pte_offset(mm, addr);
105 ptl = &mm->page_table_lock;
107 pgd = pgd_offset(mm, addr);
108 if (!pgd_present(*pgd))
111 pud = pud_offset(pgd, addr);
112 if (!pud_present(*pud))
115 pmd = pmd_offset(pud, addr);
116 if (pmd_trans_huge(*pmd))
118 if (!pmd_present(*pmd))
121 ptep = pte_offset_map(pmd, addr);
124 * Peek to check is_swap_pte() before taking ptlock? No, we
125 * can race mremap's move_ptes(), which skips anon_vma lock.
128 ptl = pte_lockptr(mm, pmd);
133 if (!is_swap_pte(pte))
136 entry = pte_to_swp_entry(pte);
138 if (!is_migration_entry(entry) ||
139 migration_entry_to_page(entry) != old)
143 pte = pte_mkold(mk_pte(new, vma->vm_page_prot));
144 if (is_write_migration_entry(entry))
145 pte = pte_mkwrite(pte);
146 #ifdef CONFIG_HUGETLB_PAGE
148 pte = pte_mkhuge(pte);
150 flush_cache_page(vma, addr, pte_pfn(pte));
151 set_pte_at(mm, addr, ptep, pte);
155 hugepage_add_anon_rmap(new, vma, addr);
158 } else if (PageAnon(new))
159 page_add_anon_rmap(new, vma, addr);
161 page_add_file_rmap(new);
163 /* No need to invalidate - it was non-present before */
164 update_mmu_cache(vma, addr, ptep);
166 pte_unmap_unlock(ptep, ptl);
172 * Get rid of all migration entries and replace them by
173 * references to the indicated page.
175 static void remove_migration_ptes(struct page *old, struct page *new)
177 rmap_walk(new, remove_migration_pte, old);
181 * Something used the pte of a page under migration. We need to
182 * get to the page and wait until migration is finished.
183 * When we return from this function the fault will be retried.
185 * This function is called from do_swap_page().
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)
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);
337 page_unfreeze_refs(page, expected_count);
339 * Drop cache reference from old page.
340 * We know this isn't the last reference.
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);
406 spin_unlock_irq(&mapping->tree_lock);
411 * Copy the page to its new location
413 void migrate_page_copy(struct page *newpage, struct page *page)
416 copy_huge_page(newpage, page);
418 copy_highpage(newpage, page);
421 SetPageError(newpage);
422 if (PageReferenced(page))
423 SetPageReferenced(newpage);
424 if (PageUptodate(page))
425 SetPageUptodate(newpage);
426 if (TestClearPageActive(page)) {
427 VM_BUG_ON(PageUnevictable(page));
428 SetPageActive(newpage);
429 } else if (TestClearPageUnevictable(page))
430 SetPageUnevictable(newpage);
431 if (PageChecked(page))
432 SetPageChecked(newpage);
433 if (PageMappedToDisk(page))
434 SetPageMappedToDisk(newpage);
436 if (PageDirty(page)) {
437 clear_page_dirty_for_io(page);
439 * Want to mark the page and the radix tree as dirty, and
440 * redo the accounting that clear_page_dirty_for_io undid,
441 * but we can't use set_page_dirty because that function
442 * is actually a signal that all of the page has become dirty.
443 * Whereas only part of our page may be dirty.
445 __set_page_dirty_nobuffers(newpage);
448 mlock_migrate_page(newpage, page);
449 ksm_migrate_page(newpage, page);
451 ClearPageSwapCache(page);
452 ClearPagePrivate(page);
453 set_page_private(page, 0);
454 page->mapping = NULL;
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);
678 unlock_page(newpage);
683 static int __unmap_and_move(struct page *page, struct page *newpage,
684 int force, bool offlining, enum migrate_mode mode)
687 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 charge = mem_cgroup_prepare_migration(page, newpage, &mem, GFP_KERNEL);
731 if (charge == -ENOMEM) {
737 if (PageWriteback(page)) {
739 * Only in the case of a full syncronous migration is it
740 * necessary to wait for PageWriteback. In the async case,
741 * the retry loop is too short and in the sync-light case,
742 * the overhead of stalling is too much
744 if (mode != MIGRATE_SYNC) {
750 wait_on_page_writeback(page);
753 * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
754 * we cannot notice that anon_vma is freed while we migrates a page.
755 * This get_anon_vma() delays freeing anon_vma pointer until the end
756 * of migration. File cache pages are no problem because of page_lock()
757 * File Caches may use write_page() or lock_page() in migration, then,
758 * just care Anon page here.
760 if (PageAnon(page)) {
762 * Only page_lock_anon_vma() understands the subtleties of
763 * getting a hold on an anon_vma from outside one of its mms.
765 anon_vma = page_get_anon_vma(page);
770 } else if (PageSwapCache(page)) {
772 * We cannot be sure that the anon_vma of an unmapped
773 * swapcache page is safe to use because we don't
774 * know in advance if the VMA that this page belonged
775 * to still exists. If the VMA and others sharing the
776 * data have been freed, then the anon_vma could
777 * already be invalid.
779 * To avoid this possibility, swapcache pages get
780 * migrated but are not remapped when migration
790 * Corner case handling:
791 * 1. When a new swap-cache page is read into, it is added to the LRU
792 * and treated as swapcache but it has no rmap yet.
793 * Calling try_to_unmap() against a page->mapping==NULL page will
794 * trigger a BUG. So handle it here.
795 * 2. An orphaned page (see truncate_complete_page) might have
796 * fs-private metadata. The page can be picked up due to memory
797 * offlining. Everywhere else except page reclaim, the page is
798 * invisible to the vm, so the page can not be migrated. So try to
799 * free the metadata, so the page can be freed.
801 if (!page->mapping) {
802 VM_BUG_ON(PageAnon(page));
803 if (page_has_private(page)) {
804 try_to_free_buffers(page);
810 /* Establish migration ptes or remove ptes */
811 try_to_unmap(page, TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
814 if (!page_mapped(page))
815 rc = move_to_new_page(newpage, page, remap_swapcache, mode);
817 if (rc && remap_swapcache)
818 remove_migration_ptes(page, page);
820 /* Drop an anon_vma reference if we took one */
822 put_anon_vma(anon_vma);
826 mem_cgroup_end_migration(mem, page, newpage, rc == 0);
834 * Obtain the lock on page, remove all ptes and migrate the page
835 * to the newly allocated page in newpage.
837 static int unmap_and_move(new_page_t get_new_page, unsigned long private,
838 struct page *page, int force, bool offlining,
839 enum migrate_mode mode)
843 struct page *newpage = get_new_page(page, private, &result);
848 if (page_count(page) == 1) {
849 /* page was freed from under us. So we are done. */
853 if (unlikely(PageTransHuge(page)))
854 if (unlikely(split_huge_page(page)))
857 rc = __unmap_and_move(page, newpage, force, offlining, mode);
861 * A page that has been migrated has all references
862 * removed and will be freed. A page that has not been
863 * migrated will have kepts its references and be
866 list_del(&page->lru);
867 dec_zone_page_state(page, NR_ISOLATED_ANON +
868 page_is_file_cache(page));
869 putback_lru_page(page);
872 * Move the new page to the LRU. If migration was not successful
873 * then this will free the page.
875 putback_lru_page(newpage);
880 *result = page_to_nid(newpage);
886 * Counterpart of unmap_and_move_page() for hugepage migration.
888 * This function doesn't wait the completion of hugepage I/O
889 * because there is no race between I/O and migration for hugepage.
890 * Note that currently hugepage I/O occurs only in direct I/O
891 * where no lock is held and PG_writeback is irrelevant,
892 * and writeback status of all subpages are counted in the reference
893 * count of the head page (i.e. if all subpages of a 2MB hugepage are
894 * under direct I/O, the reference of the head page is 512 and a bit more.)
895 * This means that when we try to migrate hugepage whose subpages are
896 * doing direct I/O, some references remain after try_to_unmap() and
897 * hugepage migration fails without data corruption.
899 * There is also no race when direct I/O is issued on the page under migration,
900 * because then pte is replaced with migration swap entry and direct I/O code
901 * will wait in the page fault for migration to complete.
903 static int unmap_and_move_huge_page(new_page_t get_new_page,
904 unsigned long private, struct page *hpage,
905 int force, bool offlining,
906 enum migrate_mode mode)
910 struct page *new_hpage = get_new_page(hpage, private, &result);
911 struct anon_vma *anon_vma = NULL;
918 if (!trylock_page(hpage)) {
919 if (!force || mode != MIGRATE_SYNC)
925 anon_vma = page_get_anon_vma(hpage);
927 try_to_unmap(hpage, TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
929 if (!page_mapped(hpage))
930 rc = move_to_new_page(new_hpage, hpage, 1, mode);
933 remove_migration_ptes(hpage, hpage);
936 put_anon_vma(anon_vma);
941 list_del(&hpage->lru);
951 *result = page_to_nid(new_hpage);
959 * The function takes one list of pages to migrate and a function
960 * that determines from the page to be migrated and the private data
961 * the target of the move and allocates the page.
963 * The function returns after 10 attempts or if no pages
964 * are movable anymore because to has become empty
965 * or no retryable pages exist anymore.
966 * Caller should call putback_lru_pages to return pages to the LRU
967 * or free list only if ret != 0.
969 * Return: Number of pages not migrated or error code.
971 int migrate_pages(struct list_head *from,
972 new_page_t get_new_page, unsigned long private, bool offlining,
973 enum migrate_mode mode)
980 int swapwrite = current->flags & PF_SWAPWRITE;
984 current->flags |= PF_SWAPWRITE;
986 for(pass = 0; pass < 10 && retry; pass++) {
989 list_for_each_entry_safe(page, page2, from, lru) {
992 rc = unmap_and_move(get_new_page, private,
993 page, pass > 2, offlining,
1005 /* Permanent failure */
1014 current->flags &= ~PF_SWAPWRITE;
1019 return nr_failed + retry;
1022 int migrate_huge_pages(struct list_head *from,
1023 new_page_t get_new_page, unsigned long private, bool offlining,
1024 enum migrate_mode mode)
1033 for (pass = 0; pass < 10 && retry; pass++) {
1036 list_for_each_entry_safe(page, page2, from, lru) {
1039 rc = unmap_and_move_huge_page(get_new_page,
1040 private, page, pass > 2, offlining,
1052 /* Permanent failure */
1063 return nr_failed + retry;
1068 * Move a list of individual pages
1070 struct page_to_node {
1077 static struct page *new_page_node(struct page *p, unsigned long private,
1080 struct page_to_node *pm = (struct page_to_node *)private;
1082 while (pm->node != MAX_NUMNODES && pm->page != p)
1085 if (pm->node == MAX_NUMNODES)
1088 *result = &pm->status;
1090 return alloc_pages_exact_node(pm->node,
1091 GFP_HIGHUSER_MOVABLE | GFP_THISNODE, 0);
1095 * Move a set of pages as indicated in the pm array. The addr
1096 * field must be set to the virtual address of the page to be moved
1097 * and the node number must contain a valid target node.
1098 * The pm array ends with node = MAX_NUMNODES.
1100 static int do_move_page_to_node_array(struct mm_struct *mm,
1101 struct page_to_node *pm,
1105 struct page_to_node *pp;
1106 LIST_HEAD(pagelist);
1108 down_read(&mm->mmap_sem);
1111 * Build a list of pages to migrate
1113 for (pp = pm; pp->node != MAX_NUMNODES; pp++) {
1114 struct vm_area_struct *vma;
1118 vma = find_vma(mm, pp->addr);
1119 if (!vma || pp->addr < vma->vm_start || !vma_migratable(vma))
1122 page = follow_page(vma, pp->addr, FOLL_GET|FOLL_SPLIT);
1124 err = PTR_ERR(page);
1132 /* Use PageReserved to check for zero page */
1133 if (PageReserved(page) || PageKsm(page))
1137 err = page_to_nid(page);
1139 if (err == pp->node)
1141 * Node already in the right place
1146 if (page_mapcount(page) > 1 &&
1150 err = isolate_lru_page(page);
1152 list_add_tail(&page->lru, &pagelist);
1153 inc_zone_page_state(page, NR_ISOLATED_ANON +
1154 page_is_file_cache(page));
1158 * Either remove the duplicate refcount from
1159 * isolate_lru_page() or drop the page ref if it was
1168 if (!list_empty(&pagelist)) {
1169 err = migrate_pages(&pagelist, new_page_node,
1170 (unsigned long)pm, 0, MIGRATE_SYNC);
1172 putback_lru_pages(&pagelist);
1175 up_read(&mm->mmap_sem);
1180 * Migrate an array of page address onto an array of nodes and fill
1181 * the corresponding array of status.
1183 static int do_pages_move(struct mm_struct *mm, struct task_struct *task,
1184 unsigned long nr_pages,
1185 const void __user * __user *pages,
1186 const int __user *nodes,
1187 int __user *status, int flags)
1189 struct page_to_node *pm;
1190 nodemask_t task_nodes;
1191 unsigned long chunk_nr_pages;
1192 unsigned long chunk_start;
1195 task_nodes = cpuset_mems_allowed(task);
1198 pm = (struct page_to_node *)__get_free_page(GFP_KERNEL);
1205 * Store a chunk of page_to_node array in a page,
1206 * but keep the last one as a marker
1208 chunk_nr_pages = (PAGE_SIZE / sizeof(struct page_to_node)) - 1;
1210 for (chunk_start = 0;
1211 chunk_start < nr_pages;
1212 chunk_start += chunk_nr_pages) {
1215 if (chunk_start + chunk_nr_pages > nr_pages)
1216 chunk_nr_pages = nr_pages - chunk_start;
1218 /* fill the chunk pm with addrs and nodes from user-space */
1219 for (j = 0; j < chunk_nr_pages; j++) {
1220 const void __user *p;
1224 if (get_user(p, pages + j + chunk_start))
1226 pm[j].addr = (unsigned long) p;
1228 if (get_user(node, nodes + j + chunk_start))
1232 if (node < 0 || node >= MAX_NUMNODES)
1235 if (!node_state(node, N_HIGH_MEMORY))
1239 if (!node_isset(node, task_nodes))
1245 /* End marker for this chunk */
1246 pm[chunk_nr_pages].node = MAX_NUMNODES;
1248 /* Migrate this chunk */
1249 err = do_move_page_to_node_array(mm, pm,
1250 flags & MPOL_MF_MOVE_ALL);
1254 /* Return status information */
1255 for (j = 0; j < chunk_nr_pages; j++)
1256 if (put_user(pm[j].status, status + j + chunk_start)) {
1264 free_page((unsigned long)pm);
1270 * Determine the nodes of an array of pages and store it in an array of status.
1272 static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages,
1273 const void __user **pages, int *status)
1277 down_read(&mm->mmap_sem);
1279 for (i = 0; i < nr_pages; i++) {
1280 unsigned long addr = (unsigned long)(*pages);
1281 struct vm_area_struct *vma;
1285 vma = find_vma(mm, addr);
1286 if (!vma || addr < vma->vm_start)
1289 page = follow_page(vma, addr, 0);
1291 err = PTR_ERR(page);
1296 /* Use PageReserved to check for zero page */
1297 if (!page || PageReserved(page) || PageKsm(page))
1300 err = page_to_nid(page);
1308 up_read(&mm->mmap_sem);
1312 * Determine the nodes of a user array of pages and store it in
1313 * a user array of status.
1315 static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages,
1316 const void __user * __user *pages,
1319 #define DO_PAGES_STAT_CHUNK_NR 16
1320 const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR];
1321 int chunk_status[DO_PAGES_STAT_CHUNK_NR];
1324 unsigned long chunk_nr;
1326 chunk_nr = nr_pages;
1327 if (chunk_nr > DO_PAGES_STAT_CHUNK_NR)
1328 chunk_nr = DO_PAGES_STAT_CHUNK_NR;
1330 if (copy_from_user(chunk_pages, pages, chunk_nr * sizeof(*chunk_pages)))
1333 do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status);
1335 if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status)))
1340 nr_pages -= chunk_nr;
1342 return nr_pages ? -EFAULT : 0;
1346 * Move a list of pages in the address space of the currently executing
1349 SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages,
1350 const void __user * __user *, pages,
1351 const int __user *, nodes,
1352 int __user *, status, int, flags)
1354 const struct cred *cred = current_cred(), *tcred;
1355 struct task_struct *task;
1356 struct mm_struct *mm;
1360 if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
1363 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1366 /* Find the mm_struct */
1368 task = pid ? find_task_by_vpid(pid) : current;
1373 mm = get_task_mm(task);
1380 * Check if this process has the right to modify the specified
1381 * process. The right exists if the process has administrative
1382 * capabilities, superuser privileges or the same
1383 * userid as the target process.
1386 tcred = __task_cred(task);
1387 if (cred->euid != tcred->suid && cred->euid != tcred->uid &&
1388 cred->uid != tcred->suid && cred->uid != tcred->uid &&
1389 !capable(CAP_SYS_NICE)) {
1396 err = security_task_movememory(task);
1401 err = do_pages_move(mm, task, nr_pages, pages, nodes, status,
1404 err = do_pages_stat(mm, nr_pages, pages, status);
1413 * Call migration functions in the vma_ops that may prepare
1414 * memory in a vm for migration. migration functions may perform
1415 * the migration for vmas that do not have an underlying page struct.
1417 int migrate_vmas(struct mm_struct *mm, const nodemask_t *to,
1418 const nodemask_t *from, unsigned long flags)
1420 struct vm_area_struct *vma;
1423 for (vma = mm->mmap; vma && !err; vma = vma->vm_next) {
1424 if (vma->vm_ops && vma->vm_ops->migrate) {
1425 err = vma->vm_ops->migrate(vma, to, from, flags);