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>
39 #include <asm/tlbflush.h>
44 * migrate_prep() needs to be called before we start compiling a list of pages
45 * to be migrated using isolate_lru_page(). If scheduling work on other CPUs is
46 * undesirable, use migrate_prep_local()
48 int migrate_prep(void)
51 * Clear the LRU lists so pages can be isolated.
52 * Note that pages may be moved off the LRU after we have
53 * drained them. Those pages will fail to migrate like other
54 * pages that may be busy.
61 /* Do the necessary work of migrate_prep but not if it involves other CPUs */
62 int migrate_prep_local(void)
70 * Add isolated pages on the list back to the LRU under page lock
71 * to avoid leaking evictable pages back onto unevictable list.
73 void putback_lru_pages(struct list_head *l)
78 list_for_each_entry_safe(page, page2, l, lru) {
80 dec_zone_page_state(page, NR_ISOLATED_ANON +
81 page_is_file_cache(page));
82 putback_lru_page(page);
87 * Restore a potential migration pte to a working pte entry
89 static int remove_migration_pte(struct page *new, struct vm_area_struct *vma,
90 unsigned long addr, void *old)
92 struct mm_struct *mm = vma->vm_mm;
100 if (unlikely(PageHuge(new))) {
101 ptep = huge_pte_offset(mm, addr);
104 ptl = &mm->page_table_lock;
106 pgd = pgd_offset(mm, addr);
107 if (!pgd_present(*pgd))
110 pud = pud_offset(pgd, addr);
111 if (!pud_present(*pud))
114 pmd = pmd_offset(pud, addr);
115 if (pmd_trans_huge(*pmd))
117 if (!pmd_present(*pmd))
120 ptep = pte_offset_map(pmd, addr);
123 * Peek to check is_swap_pte() before taking ptlock? No, we
124 * can race mremap's move_ptes(), which skips anon_vma lock.
127 ptl = pte_lockptr(mm, pmd);
132 if (!is_swap_pte(pte))
135 entry = pte_to_swp_entry(pte);
137 if (!is_migration_entry(entry) ||
138 migration_entry_to_page(entry) != old)
142 pte = pte_mkold(mk_pte(new, vma->vm_page_prot));
143 if (is_write_migration_entry(entry))
144 pte = pte_mkwrite(pte);
145 #ifdef CONFIG_HUGETLB_PAGE
147 pte = pte_mkhuge(pte);
149 flush_cache_page(vma, addr, pte_pfn(pte));
150 set_pte_at(mm, addr, ptep, pte);
154 hugepage_add_anon_rmap(new, vma, addr);
157 } else if (PageAnon(new))
158 page_add_anon_rmap(new, vma, addr);
160 page_add_file_rmap(new);
162 /* No need to invalidate - it was non-present before */
163 update_mmu_cache(vma, addr, ptep);
165 pte_unmap_unlock(ptep, ptl);
171 * Get rid of all migration entries and replace them by
172 * references to the indicated page.
174 static void remove_migration_ptes(struct page *old, struct page *new)
176 rmap_walk(new, remove_migration_pte, old);
180 * Something used the pte of a page under migration. We need to
181 * get to the page and wait until migration is finished.
182 * When we return from this function the fault will be retried.
184 void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
185 unsigned long address)
192 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
194 if (!is_swap_pte(pte))
197 entry = pte_to_swp_entry(pte);
198 if (!is_migration_entry(entry))
201 page = migration_entry_to_page(entry);
204 * Once radix-tree replacement of page migration started, page_count
205 * *must* be zero. And, we don't want to call wait_on_page_locked()
206 * against a page without get_page().
207 * So, we use get_page_unless_zero(), here. Even failed, page fault
210 if (!get_page_unless_zero(page))
212 pte_unmap_unlock(ptep, ptl);
213 wait_on_page_locked(page);
217 pte_unmap_unlock(ptep, ptl);
221 /* Returns true if all buffers are successfully locked */
222 static bool buffer_migrate_lock_buffers(struct buffer_head *head,
223 enum migrate_mode mode)
225 struct buffer_head *bh = head;
227 /* Simple case, sync compaction */
228 if (mode != MIGRATE_ASYNC) {
232 bh = bh->b_this_page;
234 } while (bh != head);
239 /* async case, we cannot block on lock_buffer so use trylock_buffer */
242 if (!trylock_buffer(bh)) {
244 * We failed to lock the buffer and cannot stall in
245 * async migration. Release the taken locks
247 struct buffer_head *failed_bh = bh;
250 while (bh != failed_bh) {
253 bh = bh->b_this_page;
258 bh = bh->b_this_page;
259 } while (bh != head);
263 static inline bool buffer_migrate_lock_buffers(struct buffer_head *head,
264 enum migrate_mode mode)
268 #endif /* CONFIG_BLOCK */
271 * Replace the page in the mapping.
273 * The number of remaining references must be:
274 * 1 for anonymous pages without a mapping
275 * 2 for pages with a mapping
276 * 3 for pages with a mapping and PagePrivate/PagePrivate2 set.
278 static int migrate_page_move_mapping(struct address_space *mapping,
279 struct page *newpage, struct page *page,
280 struct buffer_head *head, enum migrate_mode mode)
286 /* Anonymous page without mapping */
287 if (page_count(page) != 1)
292 spin_lock_irq(&mapping->tree_lock);
294 pslot = radix_tree_lookup_slot(&mapping->page_tree,
297 expected_count = 2 + page_has_private(page);
298 if (page_count(page) != expected_count ||
299 radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) {
300 spin_unlock_irq(&mapping->tree_lock);
304 if (!page_freeze_refs(page, expected_count)) {
305 spin_unlock_irq(&mapping->tree_lock);
310 * In the async migration case of moving a page with buffers, lock the
311 * buffers using trylock before the mapping is moved. If the mapping
312 * was moved, we later failed to lock the buffers and could not move
313 * the mapping back due to an elevated page count, we would have to
314 * block waiting on other references to be dropped.
316 if (mode == MIGRATE_ASYNC && head &&
317 !buffer_migrate_lock_buffers(head, mode)) {
318 page_unfreeze_refs(page, expected_count);
319 spin_unlock_irq(&mapping->tree_lock);
324 * Now we know that no one else is looking at the page.
326 get_page(newpage); /* add cache reference */
327 if (PageSwapCache(page)) {
328 SetPageSwapCache(newpage);
329 set_page_private(newpage, page_private(page));
332 radix_tree_replace_slot(pslot, newpage);
335 * Drop cache reference from old page by unfreezing
336 * to one less reference.
337 * We know this isn't the last reference.
339 page_unfreeze_refs(page, expected_count - 1);
342 * If moved to a different zone then also account
343 * the page for that zone. Other VM counters will be
344 * taken care of when we establish references to the
345 * new page and drop references to the old page.
347 * Note that anonymous pages are accounted for
348 * via NR_FILE_PAGES and NR_ANON_PAGES if they
349 * are mapped to swap space.
351 __dec_zone_page_state(page, NR_FILE_PAGES);
352 __inc_zone_page_state(newpage, NR_FILE_PAGES);
353 if (!PageSwapCache(page) && PageSwapBacked(page)) {
354 __dec_zone_page_state(page, NR_SHMEM);
355 __inc_zone_page_state(newpage, NR_SHMEM);
357 spin_unlock_irq(&mapping->tree_lock);
363 * The expected number of remaining references is the same as that
364 * of migrate_page_move_mapping().
366 int migrate_huge_page_move_mapping(struct address_space *mapping,
367 struct page *newpage, struct page *page)
373 if (page_count(page) != 1)
378 spin_lock_irq(&mapping->tree_lock);
380 pslot = radix_tree_lookup_slot(&mapping->page_tree,
383 expected_count = 2 + page_has_private(page);
384 if (page_count(page) != expected_count ||
385 radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) {
386 spin_unlock_irq(&mapping->tree_lock);
390 if (!page_freeze_refs(page, expected_count)) {
391 spin_unlock_irq(&mapping->tree_lock);
397 radix_tree_replace_slot(pslot, newpage);
399 page_unfreeze_refs(page, expected_count - 1);
401 spin_unlock_irq(&mapping->tree_lock);
406 * Copy the page to its new location
408 void migrate_page_copy(struct page *newpage, struct page *page)
411 copy_huge_page(newpage, page);
413 copy_highpage(newpage, page);
416 SetPageError(newpage);
417 if (PageReferenced(page))
418 SetPageReferenced(newpage);
419 if (PageUptodate(page))
420 SetPageUptodate(newpage);
421 if (TestClearPageActive(page)) {
422 VM_BUG_ON(PageUnevictable(page));
423 SetPageActive(newpage);
424 } else if (TestClearPageUnevictable(page))
425 SetPageUnevictable(newpage);
426 if (PageChecked(page))
427 SetPageChecked(newpage);
428 if (PageMappedToDisk(page))
429 SetPageMappedToDisk(newpage);
431 if (PageDirty(page)) {
432 clear_page_dirty_for_io(page);
434 * Want to mark the page and the radix tree as dirty, and
435 * redo the accounting that clear_page_dirty_for_io undid,
436 * but we can't use set_page_dirty because that function
437 * is actually a signal that all of the page has become dirty.
438 * Whereas only part of our page may be dirty.
440 if (PageSwapBacked(page))
441 SetPageDirty(newpage);
443 __set_page_dirty_nobuffers(newpage);
446 mlock_migrate_page(newpage, page);
447 ksm_migrate_page(newpage, page);
449 ClearPageSwapCache(page);
450 ClearPagePrivate(page);
451 set_page_private(page, 0);
454 * If any waiters have accumulated on the new page then
457 if (PageWriteback(newpage))
458 end_page_writeback(newpage);
461 /************************************************************
462 * Migration functions
463 ***********************************************************/
465 /* Always fail migration. Used for mappings that are not movable */
466 int fail_migrate_page(struct address_space *mapping,
467 struct page *newpage, struct page *page)
471 EXPORT_SYMBOL(fail_migrate_page);
474 * Common logic to directly migrate a single page suitable for
475 * pages that do not use PagePrivate/PagePrivate2.
477 * Pages are locked upon entry and exit.
479 int migrate_page(struct address_space *mapping,
480 struct page *newpage, struct page *page,
481 enum migrate_mode mode)
485 BUG_ON(PageWriteback(page)); /* Writeback must be complete */
487 rc = migrate_page_move_mapping(mapping, newpage, page, NULL, mode);
492 migrate_page_copy(newpage, page);
495 EXPORT_SYMBOL(migrate_page);
499 * Migration function for pages with buffers. This function can only be used
500 * if the underlying filesystem guarantees that no other references to "page"
503 int buffer_migrate_page(struct address_space *mapping,
504 struct page *newpage, struct page *page, enum migrate_mode mode)
506 struct buffer_head *bh, *head;
509 if (!page_has_buffers(page))
510 return migrate_page(mapping, newpage, page, mode);
512 head = page_buffers(page);
514 rc = migrate_page_move_mapping(mapping, newpage, page, head, mode);
520 * In the async case, migrate_page_move_mapping locked the buffers
521 * with an IRQ-safe spinlock held. In the sync case, the buffers
522 * need to be locked now
524 if (mode != MIGRATE_ASYNC)
525 BUG_ON(!buffer_migrate_lock_buffers(head, mode));
527 ClearPagePrivate(page);
528 set_page_private(newpage, page_private(page));
529 set_page_private(page, 0);
535 set_bh_page(bh, newpage, bh_offset(bh));
536 bh = bh->b_this_page;
538 } while (bh != head);
540 SetPagePrivate(newpage);
542 migrate_page_copy(newpage, page);
548 bh = bh->b_this_page;
550 } while (bh != head);
554 EXPORT_SYMBOL(buffer_migrate_page);
558 * Writeback a page to clean the dirty state
560 static int writeout(struct address_space *mapping, struct page *page)
562 struct writeback_control wbc = {
563 .sync_mode = WB_SYNC_NONE,
566 .range_end = LLONG_MAX,
571 if (!mapping->a_ops->writepage)
572 /* No write method for the address space */
575 if (!clear_page_dirty_for_io(page))
576 /* Someone else already triggered a write */
580 * A dirty page may imply that the underlying filesystem has
581 * the page on some queue. So the page must be clean for
582 * migration. Writeout may mean we loose the lock and the
583 * page state is no longer what we checked for earlier.
584 * At this point we know that the migration attempt cannot
587 remove_migration_ptes(page, page);
589 rc = mapping->a_ops->writepage(page, &wbc);
591 if (rc != AOP_WRITEPAGE_ACTIVATE)
592 /* unlocked. Relock */
595 return (rc < 0) ? -EIO : -EAGAIN;
599 * Default handling if a filesystem does not provide a migration function.
601 static int fallback_migrate_page(struct address_space *mapping,
602 struct page *newpage, struct page *page, enum migrate_mode mode)
604 if (PageDirty(page)) {
605 /* Only writeback pages in full synchronous migration */
606 if (mode != MIGRATE_SYNC)
608 return writeout(mapping, page);
612 * Buffers may be managed in a filesystem specific way.
613 * We must have no buffers or drop them.
615 if (page_has_private(page) &&
616 !try_to_release_page(page, GFP_KERNEL))
619 return migrate_page(mapping, newpage, page, mode);
623 * Move a page to a newly allocated page
624 * The page is locked and all ptes have been successfully removed.
626 * The new page will have replaced the old page if this function
633 static int move_to_new_page(struct page *newpage, struct page *page,
634 int remap_swapcache, enum migrate_mode mode)
636 struct address_space *mapping;
640 * Block others from accessing the page when we get around to
641 * establishing additional references. We are the only one
642 * holding a reference to the new page at this point.
644 if (!trylock_page(newpage))
647 /* Prepare mapping for the new page.*/
648 newpage->index = page->index;
649 newpage->mapping = page->mapping;
650 if (PageSwapBacked(page))
651 SetPageSwapBacked(newpage);
653 mapping = page_mapping(page);
655 rc = migrate_page(mapping, newpage, page, mode);
656 else if (mapping->a_ops->migratepage)
658 * Most pages have a mapping and most filesystems provide a
659 * migratepage callback. Anonymous pages are part of swap
660 * space which also has its own migratepage callback. This
661 * is the most common path for page migration.
663 rc = mapping->a_ops->migratepage(mapping,
664 newpage, page, mode);
666 rc = fallback_migrate_page(mapping, newpage, page, mode);
669 newpage->mapping = NULL;
672 remove_migration_ptes(page, newpage);
673 page->mapping = NULL;
676 unlock_page(newpage);
681 static int __unmap_and_move(struct page *page, struct page *newpage,
682 int force, bool offlining, enum migrate_mode mode)
685 int remap_swapcache = 1;
686 struct mem_cgroup *mem;
687 struct anon_vma *anon_vma = NULL;
689 if (!trylock_page(page)) {
690 if (!force || mode == MIGRATE_ASYNC)
694 * It's not safe for direct compaction to call lock_page.
695 * For example, during page readahead pages are added locked
696 * to the LRU. Later, when the IO completes the pages are
697 * marked uptodate and unlocked. However, the queueing
698 * could be merging multiple pages for one bio (e.g.
699 * mpage_readpages). If an allocation happens for the
700 * second or third page, the process can end up locking
701 * the same page twice and deadlocking. Rather than
702 * trying to be clever about what pages can be locked,
703 * avoid the use of lock_page for direct compaction
706 if (current->flags & PF_MEMALLOC)
713 * Only memory hotplug's offline_pages() caller has locked out KSM,
714 * and can safely migrate a KSM page. The other cases have skipped
715 * PageKsm along with PageReserved - but it is only now when we have
716 * the page lock that we can be certain it will not go KSM beneath us
717 * (KSM will not upgrade a page from PageAnon to PageKsm when it sees
718 * its pagecount raised, but only here do we take the page lock which
721 if (PageKsm(page) && !offlining) {
726 /* charge against new page */
727 mem_cgroup_prepare_migration(page, newpage, &mem);
729 if (PageWriteback(page)) {
731 * Only in the case of a full syncronous migration is it
732 * necessary to wait for PageWriteback. In the async case,
733 * the retry loop is too short and in the sync-light case,
734 * the overhead of stalling is too much
736 if (mode != MIGRATE_SYNC) {
742 wait_on_page_writeback(page);
745 * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
746 * we cannot notice that anon_vma is freed while we migrates a page.
747 * This get_anon_vma() delays freeing anon_vma pointer until the end
748 * of migration. File cache pages are no problem because of page_lock()
749 * File Caches may use write_page() or lock_page() in migration, then,
750 * just care Anon page here.
752 if (PageAnon(page)) {
754 * Only page_lock_anon_vma() understands the subtleties of
755 * getting a hold on an anon_vma from outside one of its mms.
757 anon_vma = page_get_anon_vma(page);
762 } else if (PageSwapCache(page)) {
764 * We cannot be sure that the anon_vma of an unmapped
765 * swapcache page is safe to use because we don't
766 * know in advance if the VMA that this page belonged
767 * to still exists. If the VMA and others sharing the
768 * data have been freed, then the anon_vma could
769 * already be invalid.
771 * To avoid this possibility, swapcache pages get
772 * migrated but are not remapped when migration
782 * Corner case handling:
783 * 1. When a new swap-cache page is read into, it is added to the LRU
784 * and treated as swapcache but it has no rmap yet.
785 * Calling try_to_unmap() against a page->mapping==NULL page will
786 * trigger a BUG. So handle it here.
787 * 2. An orphaned page (see truncate_complete_page) might have
788 * fs-private metadata. The page can be picked up due to memory
789 * offlining. Everywhere else except page reclaim, the page is
790 * invisible to the vm, so the page can not be migrated. So try to
791 * free the metadata, so the page can be freed.
793 if (!page->mapping) {
794 VM_BUG_ON(PageAnon(page));
795 if (page_has_private(page)) {
796 try_to_free_buffers(page);
802 /* Establish migration ptes or remove ptes */
803 try_to_unmap(page, TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
806 if (!page_mapped(page))
807 rc = move_to_new_page(newpage, page, remap_swapcache, mode);
809 if (rc && remap_swapcache)
810 remove_migration_ptes(page, page);
812 /* Drop an anon_vma reference if we took one */
814 put_anon_vma(anon_vma);
817 mem_cgroup_end_migration(mem, page, newpage, rc == 0);
825 * Obtain the lock on page, remove all ptes and migrate the page
826 * to the newly allocated page in newpage.
828 static int unmap_and_move(new_page_t get_new_page, unsigned long private,
829 struct page *page, int force, bool offlining,
830 enum migrate_mode mode)
834 struct page *newpage = get_new_page(page, private, &result);
839 if (page_count(page) == 1) {
840 /* page was freed from under us. So we are done. */
844 if (unlikely(PageTransHuge(page)))
845 if (unlikely(split_huge_page(page)))
848 rc = __unmap_and_move(page, newpage, force, offlining, mode);
852 * A page that has been migrated has all references
853 * removed and will be freed. A page that has not been
854 * migrated will have kepts its references and be
857 list_del(&page->lru);
858 dec_zone_page_state(page, NR_ISOLATED_ANON +
859 page_is_file_cache(page));
860 putback_lru_page(page);
863 * Move the new page to the LRU. If migration was not successful
864 * then this will free the page.
866 putback_lru_page(newpage);
871 *result = page_to_nid(newpage);
877 * Counterpart of unmap_and_move_page() for hugepage migration.
879 * This function doesn't wait the completion of hugepage I/O
880 * because there is no race between I/O and migration for hugepage.
881 * Note that currently hugepage I/O occurs only in direct I/O
882 * where no lock is held and PG_writeback is irrelevant,
883 * and writeback status of all subpages are counted in the reference
884 * count of the head page (i.e. if all subpages of a 2MB hugepage are
885 * under direct I/O, the reference of the head page is 512 and a bit more.)
886 * This means that when we try to migrate hugepage whose subpages are
887 * doing direct I/O, some references remain after try_to_unmap() and
888 * hugepage migration fails without data corruption.
890 * There is also no race when direct I/O is issued on the page under migration,
891 * because then pte is replaced with migration swap entry and direct I/O code
892 * will wait in the page fault for migration to complete.
894 static int unmap_and_move_huge_page(new_page_t get_new_page,
895 unsigned long private, struct page *hpage,
896 int force, bool offlining,
897 enum migrate_mode mode)
901 struct page *new_hpage = get_new_page(hpage, private, &result);
902 struct anon_vma *anon_vma = NULL;
909 if (!trylock_page(hpage)) {
910 if (!force || mode != MIGRATE_SYNC)
916 anon_vma = page_get_anon_vma(hpage);
918 try_to_unmap(hpage, TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
920 if (!page_mapped(hpage))
921 rc = move_to_new_page(new_hpage, hpage, 1, mode);
924 remove_migration_ptes(hpage, hpage);
927 put_anon_vma(anon_vma);
930 hugetlb_cgroup_migrate(hpage, new_hpage);
939 *result = page_to_nid(new_hpage);
947 * The function takes one list of pages to migrate and a function
948 * that determines from the page to be migrated and the private data
949 * the target of the move and allocates the page.
951 * The function returns after 10 attempts or if no pages
952 * are movable anymore because to has become empty
953 * or no retryable pages exist anymore.
954 * Caller should call putback_lru_pages to return pages to the LRU
955 * or free list only if ret != 0.
957 * Return: Number of pages not migrated or error code.
959 int migrate_pages(struct list_head *from,
960 new_page_t get_new_page, unsigned long private, bool offlining,
961 enum migrate_mode mode)
968 int swapwrite = current->flags & PF_SWAPWRITE;
972 current->flags |= PF_SWAPWRITE;
974 for(pass = 0; pass < 10 && retry; pass++) {
977 list_for_each_entry_safe(page, page2, from, lru) {
980 rc = unmap_and_move(get_new_page, private,
981 page, pass > 2, offlining,
993 /* Permanent failure */
1002 current->flags &= ~PF_SWAPWRITE;
1007 return nr_failed + retry;
1010 int migrate_huge_page(struct page *hpage, new_page_t get_new_page,
1011 unsigned long private, bool offlining,
1012 enum migrate_mode mode)
1016 for (pass = 0; pass < 10; pass++) {
1017 rc = unmap_and_move_huge_page(get_new_page,
1018 private, hpage, pass > 2, offlining,
1040 * Move a list of individual pages
1042 struct page_to_node {
1049 static struct page *new_page_node(struct page *p, unsigned long private,
1052 struct page_to_node *pm = (struct page_to_node *)private;
1054 while (pm->node != MAX_NUMNODES && pm->page != p)
1057 if (pm->node == MAX_NUMNODES)
1060 *result = &pm->status;
1062 return alloc_pages_exact_node(pm->node,
1063 GFP_HIGHUSER_MOVABLE | GFP_THISNODE, 0);
1067 * Move a set of pages as indicated in the pm array. The addr
1068 * field must be set to the virtual address of the page to be moved
1069 * and the node number must contain a valid target node.
1070 * The pm array ends with node = MAX_NUMNODES.
1072 static int do_move_page_to_node_array(struct mm_struct *mm,
1073 struct page_to_node *pm,
1077 struct page_to_node *pp;
1078 LIST_HEAD(pagelist);
1080 down_read(&mm->mmap_sem);
1083 * Build a list of pages to migrate
1085 for (pp = pm; pp->node != MAX_NUMNODES; pp++) {
1086 struct vm_area_struct *vma;
1090 vma = find_vma(mm, pp->addr);
1091 if (!vma || pp->addr < vma->vm_start || !vma_migratable(vma))
1094 page = follow_page(vma, pp->addr, FOLL_GET|FOLL_SPLIT);
1096 err = PTR_ERR(page);
1104 /* Use PageReserved to check for zero page */
1105 if (PageReserved(page) || PageKsm(page))
1109 err = page_to_nid(page);
1111 if (err == pp->node)
1113 * Node already in the right place
1118 if (page_mapcount(page) > 1 &&
1122 err = isolate_lru_page(page);
1124 list_add_tail(&page->lru, &pagelist);
1125 inc_zone_page_state(page, NR_ISOLATED_ANON +
1126 page_is_file_cache(page));
1130 * Either remove the duplicate refcount from
1131 * isolate_lru_page() or drop the page ref if it was
1140 if (!list_empty(&pagelist)) {
1141 err = migrate_pages(&pagelist, new_page_node,
1142 (unsigned long)pm, 0, MIGRATE_SYNC);
1144 putback_lru_pages(&pagelist);
1147 up_read(&mm->mmap_sem);
1152 * Migrate an array of page address onto an array of nodes and fill
1153 * the corresponding array of status.
1155 static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes,
1156 unsigned long nr_pages,
1157 const void __user * __user *pages,
1158 const int __user *nodes,
1159 int __user *status, int flags)
1161 struct page_to_node *pm;
1162 unsigned long chunk_nr_pages;
1163 unsigned long chunk_start;
1167 pm = (struct page_to_node *)__get_free_page(GFP_KERNEL);
1174 * Store a chunk of page_to_node array in a page,
1175 * but keep the last one as a marker
1177 chunk_nr_pages = (PAGE_SIZE / sizeof(struct page_to_node)) - 1;
1179 for (chunk_start = 0;
1180 chunk_start < nr_pages;
1181 chunk_start += chunk_nr_pages) {
1184 if (chunk_start + chunk_nr_pages > nr_pages)
1185 chunk_nr_pages = nr_pages - chunk_start;
1187 /* fill the chunk pm with addrs and nodes from user-space */
1188 for (j = 0; j < chunk_nr_pages; j++) {
1189 const void __user *p;
1193 if (get_user(p, pages + j + chunk_start))
1195 pm[j].addr = (unsigned long) p;
1197 if (get_user(node, nodes + j + chunk_start))
1201 if (node < 0 || node >= MAX_NUMNODES)
1204 if (!node_state(node, N_HIGH_MEMORY))
1208 if (!node_isset(node, task_nodes))
1214 /* End marker for this chunk */
1215 pm[chunk_nr_pages].node = MAX_NUMNODES;
1217 /* Migrate this chunk */
1218 err = do_move_page_to_node_array(mm, pm,
1219 flags & MPOL_MF_MOVE_ALL);
1223 /* Return status information */
1224 for (j = 0; j < chunk_nr_pages; j++)
1225 if (put_user(pm[j].status, status + j + chunk_start)) {
1233 free_page((unsigned long)pm);
1239 * Determine the nodes of an array of pages and store it in an array of status.
1241 static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages,
1242 const void __user **pages, int *status)
1246 down_read(&mm->mmap_sem);
1248 for (i = 0; i < nr_pages; i++) {
1249 unsigned long addr = (unsigned long)(*pages);
1250 struct vm_area_struct *vma;
1254 vma = find_vma(mm, addr);
1255 if (!vma || addr < vma->vm_start)
1258 page = follow_page(vma, addr, 0);
1260 err = PTR_ERR(page);
1265 /* Use PageReserved to check for zero page */
1266 if (!page || PageReserved(page) || PageKsm(page))
1269 err = page_to_nid(page);
1277 up_read(&mm->mmap_sem);
1281 * Determine the nodes of a user array of pages and store it in
1282 * a user array of status.
1284 static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages,
1285 const void __user * __user *pages,
1288 #define DO_PAGES_STAT_CHUNK_NR 16
1289 const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR];
1290 int chunk_status[DO_PAGES_STAT_CHUNK_NR];
1293 unsigned long chunk_nr;
1295 chunk_nr = nr_pages;
1296 if (chunk_nr > DO_PAGES_STAT_CHUNK_NR)
1297 chunk_nr = DO_PAGES_STAT_CHUNK_NR;
1299 if (copy_from_user(chunk_pages, pages, chunk_nr * sizeof(*chunk_pages)))
1302 do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status);
1304 if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status)))
1309 nr_pages -= chunk_nr;
1311 return nr_pages ? -EFAULT : 0;
1315 * Move a list of pages in the address space of the currently executing
1318 SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages,
1319 const void __user * __user *, pages,
1320 const int __user *, nodes,
1321 int __user *, status, int, flags)
1323 const struct cred *cred = current_cred(), *tcred;
1324 struct task_struct *task;
1325 struct mm_struct *mm;
1327 nodemask_t task_nodes;
1330 if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
1333 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1336 /* Find the mm_struct */
1338 task = pid ? find_task_by_vpid(pid) : current;
1343 get_task_struct(task);
1346 * Check if this process has the right to modify the specified
1347 * process. The right exists if the process has administrative
1348 * capabilities, superuser privileges or the same
1349 * userid as the target process.
1351 tcred = __task_cred(task);
1352 if (!uid_eq(cred->euid, tcred->suid) && !uid_eq(cred->euid, tcred->uid) &&
1353 !uid_eq(cred->uid, tcred->suid) && !uid_eq(cred->uid, tcred->uid) &&
1354 !capable(CAP_SYS_NICE)) {
1361 err = security_task_movememory(task);
1365 task_nodes = cpuset_mems_allowed(task);
1366 mm = get_task_mm(task);
1367 put_task_struct(task);
1373 err = do_pages_move(mm, task_nodes, nr_pages, pages,
1374 nodes, status, flags);
1376 err = do_pages_stat(mm, nr_pages, pages, status);
1382 put_task_struct(task);
1387 * Call migration functions in the vma_ops that may prepare
1388 * memory in a vm for migration. migration functions may perform
1389 * the migration for vmas that do not have an underlying page struct.
1391 int migrate_vmas(struct mm_struct *mm, const nodemask_t *to,
1392 const nodemask_t *from, unsigned long flags)
1394 struct vm_area_struct *vma;
1397 for (vma = mm->mmap; vma && !err; vma = vma->vm_next) {
1398 if (vma->vm_ops && vma->vm_ops->migrate) {
1399 err = vma->vm_ops->migrate(vma, to, from, flags);