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/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_dcache_page(new);
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 static void __migration_entry_wait(struct mm_struct *mm, pte_t *ptep,
196 if (!is_swap_pte(pte))
199 entry = pte_to_swp_entry(pte);
200 if (!is_migration_entry(entry))
203 page = migration_entry_to_page(entry);
206 * Once radix-tree replacement of page migration started, page_count
207 * *must* be zero. And, we don't want to call wait_on_page_locked()
208 * against a page without get_page().
209 * So, we use get_page_unless_zero(), here. Even failed, page fault
212 if (!get_page_unless_zero(page))
214 pte_unmap_unlock(ptep, ptl);
215 wait_on_page_locked(page);
219 pte_unmap_unlock(ptep, ptl);
222 void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
223 unsigned long address)
225 spinlock_t *ptl = pte_lockptr(mm, pmd);
226 pte_t *ptep = pte_offset_map(pmd, address);
227 __migration_entry_wait(mm, ptep, ptl);
230 void migration_entry_wait_huge(struct mm_struct *mm, pte_t *pte)
232 spinlock_t *ptl = &(mm)->page_table_lock;
233 __migration_entry_wait(mm, pte, ptl);
237 /* Returns true if all buffers are successfully locked */
238 static bool buffer_migrate_lock_buffers(struct buffer_head *head,
239 enum migrate_mode mode)
241 struct buffer_head *bh = head;
243 /* Simple case, sync compaction */
244 if (mode != MIGRATE_ASYNC) {
248 bh = bh->b_this_page;
250 } while (bh != head);
255 /* async case, we cannot block on lock_buffer so use trylock_buffer */
258 if (!trylock_buffer(bh)) {
260 * We failed to lock the buffer and cannot stall in
261 * async migration. Release the taken locks
263 struct buffer_head *failed_bh = bh;
266 while (bh != failed_bh) {
269 bh = bh->b_this_page;
274 bh = bh->b_this_page;
275 } while (bh != head);
279 static inline bool buffer_migrate_lock_buffers(struct buffer_head *head,
280 enum migrate_mode mode)
284 #endif /* CONFIG_BLOCK */
287 * Replace the page in the mapping.
289 * The number of remaining references must be:
290 * 1 for anonymous pages without a mapping
291 * 2 for pages with a mapping
292 * 3 for pages with a mapping and PagePrivate/PagePrivate2 set.
294 static int migrate_page_move_mapping(struct address_space *mapping,
295 struct page *newpage, struct page *page,
296 struct buffer_head *head, enum migrate_mode mode)
302 /* Anonymous page without mapping */
303 if (page_count(page) != 1)
308 spin_lock_irq(&mapping->tree_lock);
310 pslot = radix_tree_lookup_slot(&mapping->page_tree,
313 expected_count = 2 + page_has_private(page);
314 if (page_count(page) != expected_count ||
315 radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) {
316 spin_unlock_irq(&mapping->tree_lock);
320 if (!page_freeze_refs(page, expected_count)) {
321 spin_unlock_irq(&mapping->tree_lock);
326 * In the async migration case of moving a page with buffers, lock the
327 * buffers using trylock before the mapping is moved. If the mapping
328 * was moved, we later failed to lock the buffers and could not move
329 * the mapping back due to an elevated page count, we would have to
330 * block waiting on other references to be dropped.
332 if (mode == MIGRATE_ASYNC && head &&
333 !buffer_migrate_lock_buffers(head, mode)) {
334 page_unfreeze_refs(page, expected_count);
335 spin_unlock_irq(&mapping->tree_lock);
340 * Now we know that no one else is looking at the page.
342 get_page(newpage); /* add cache reference */
343 if (PageSwapCache(page)) {
344 SetPageSwapCache(newpage);
345 set_page_private(newpage, page_private(page));
348 radix_tree_replace_slot(pslot, newpage);
350 page_unfreeze_refs(page, expected_count);
352 * Drop cache reference from old page.
353 * We know this isn't the last reference.
358 * If moved to a different zone then also account
359 * the page for that zone. Other VM counters will be
360 * taken care of when we establish references to the
361 * new page and drop references to the old page.
363 * Note that anonymous pages are accounted for
364 * via NR_FILE_PAGES and NR_ANON_PAGES if they
365 * are mapped to swap space.
367 __dec_zone_page_state(page, NR_FILE_PAGES);
368 __inc_zone_page_state(newpage, NR_FILE_PAGES);
369 if (!PageSwapCache(page) && PageSwapBacked(page)) {
370 __dec_zone_page_state(page, NR_SHMEM);
371 __inc_zone_page_state(newpage, NR_SHMEM);
373 spin_unlock_irq(&mapping->tree_lock);
379 * The expected number of remaining references is the same as that
380 * of migrate_page_move_mapping().
382 int migrate_huge_page_move_mapping(struct address_space *mapping,
383 struct page *newpage, struct page *page)
389 if (page_count(page) != 1)
394 spin_lock_irq(&mapping->tree_lock);
396 pslot = radix_tree_lookup_slot(&mapping->page_tree,
399 expected_count = 2 + page_has_private(page);
400 if (page_count(page) != expected_count ||
401 radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) {
402 spin_unlock_irq(&mapping->tree_lock);
406 if (!page_freeze_refs(page, expected_count)) {
407 spin_unlock_irq(&mapping->tree_lock);
413 radix_tree_replace_slot(pslot, newpage);
415 page_unfreeze_refs(page, expected_count);
419 spin_unlock_irq(&mapping->tree_lock);
424 * Copy the page to its new location
426 void migrate_page_copy(struct page *newpage, struct page *page)
429 copy_huge_page(newpage, page);
431 copy_highpage(newpage, page);
434 SetPageError(newpage);
435 if (PageReferenced(page))
436 SetPageReferenced(newpage);
437 if (PageUptodate(page))
438 SetPageUptodate(newpage);
439 if (TestClearPageActive(page)) {
440 VM_BUG_ON(PageUnevictable(page));
441 SetPageActive(newpage);
442 } else if (TestClearPageUnevictable(page))
443 SetPageUnevictable(newpage);
444 if (PageChecked(page))
445 SetPageChecked(newpage);
446 if (PageMappedToDisk(page))
447 SetPageMappedToDisk(newpage);
449 if (PageDirty(page)) {
450 clear_page_dirty_for_io(page);
452 * Want to mark the page and the radix tree as dirty, and
453 * redo the accounting that clear_page_dirty_for_io undid,
454 * but we can't use set_page_dirty because that function
455 * is actually a signal that all of the page has become dirty.
456 * Whereas only part of our page may be dirty.
458 __set_page_dirty_nobuffers(newpage);
461 mlock_migrate_page(newpage, page);
462 ksm_migrate_page(newpage, page);
464 ClearPageSwapCache(page);
465 ClearPagePrivate(page);
466 set_page_private(page, 0);
467 page->mapping = NULL;
470 * If any waiters have accumulated on the new page then
473 if (PageWriteback(newpage))
474 end_page_writeback(newpage);
477 /************************************************************
478 * Migration functions
479 ***********************************************************/
481 /* Always fail migration. Used for mappings that are not movable */
482 int fail_migrate_page(struct address_space *mapping,
483 struct page *newpage, struct page *page)
487 EXPORT_SYMBOL(fail_migrate_page);
490 * Common logic to directly migrate a single page suitable for
491 * pages that do not use PagePrivate/PagePrivate2.
493 * Pages are locked upon entry and exit.
495 int migrate_page(struct address_space *mapping,
496 struct page *newpage, struct page *page,
497 enum migrate_mode mode)
501 BUG_ON(PageWriteback(page)); /* Writeback must be complete */
503 rc = migrate_page_move_mapping(mapping, newpage, page, NULL, mode);
508 migrate_page_copy(newpage, page);
511 EXPORT_SYMBOL(migrate_page);
515 * Migration function for pages with buffers. This function can only be used
516 * if the underlying filesystem guarantees that no other references to "page"
519 int buffer_migrate_page(struct address_space *mapping,
520 struct page *newpage, struct page *page, enum migrate_mode mode)
522 struct buffer_head *bh, *head;
525 if (!page_has_buffers(page))
526 return migrate_page(mapping, newpage, page, mode);
528 head = page_buffers(page);
530 rc = migrate_page_move_mapping(mapping, newpage, page, head, mode);
536 * In the async case, migrate_page_move_mapping locked the buffers
537 * with an IRQ-safe spinlock held. In the sync case, the buffers
538 * need to be locked now
540 if (mode != MIGRATE_ASYNC)
541 BUG_ON(!buffer_migrate_lock_buffers(head, mode));
543 ClearPagePrivate(page);
544 set_page_private(newpage, page_private(page));
545 set_page_private(page, 0);
551 set_bh_page(bh, newpage, bh_offset(bh));
552 bh = bh->b_this_page;
554 } while (bh != head);
556 SetPagePrivate(newpage);
558 migrate_page_copy(newpage, page);
564 bh = bh->b_this_page;
566 } while (bh != head);
570 EXPORT_SYMBOL(buffer_migrate_page);
574 * Writeback a page to clean the dirty state
576 static int writeout(struct address_space *mapping, struct page *page)
578 struct writeback_control wbc = {
579 .sync_mode = WB_SYNC_NONE,
582 .range_end = LLONG_MAX,
587 if (!mapping->a_ops->writepage)
588 /* No write method for the address space */
591 if (!clear_page_dirty_for_io(page))
592 /* Someone else already triggered a write */
596 * A dirty page may imply that the underlying filesystem has
597 * the page on some queue. So the page must be clean for
598 * migration. Writeout may mean we loose the lock and the
599 * page state is no longer what we checked for earlier.
600 * At this point we know that the migration attempt cannot
603 remove_migration_ptes(page, page);
605 rc = mapping->a_ops->writepage(page, &wbc);
607 if (rc != AOP_WRITEPAGE_ACTIVATE)
608 /* unlocked. Relock */
611 return (rc < 0) ? -EIO : -EAGAIN;
615 * Default handling if a filesystem does not provide a migration function.
617 static int fallback_migrate_page(struct address_space *mapping,
618 struct page *newpage, struct page *page, enum migrate_mode mode)
620 if (PageDirty(page)) {
621 /* Only writeback pages in full synchronous migration */
622 if (mode != MIGRATE_SYNC)
624 return writeout(mapping, page);
628 * Buffers may be managed in a filesystem specific way.
629 * We must have no buffers or drop them.
631 if (page_has_private(page) &&
632 !try_to_release_page(page, GFP_KERNEL))
635 return migrate_page(mapping, newpage, page, mode);
639 * Move a page to a newly allocated page
640 * The page is locked and all ptes have been successfully removed.
642 * The new page will have replaced the old page if this function
649 static int move_to_new_page(struct page *newpage, struct page *page,
650 int remap_swapcache, enum migrate_mode mode)
652 struct address_space *mapping;
656 * Block others from accessing the page when we get around to
657 * establishing additional references. We are the only one
658 * holding a reference to the new page at this point.
660 if (!trylock_page(newpage))
663 /* Prepare mapping for the new page.*/
664 newpage->index = page->index;
665 newpage->mapping = page->mapping;
666 if (PageSwapBacked(page))
667 SetPageSwapBacked(newpage);
669 mapping = page_mapping(page);
671 rc = migrate_page(mapping, newpage, page, mode);
672 else if (mapping->a_ops->migratepage)
674 * Most pages have a mapping and most filesystems provide a
675 * migratepage callback. Anonymous pages are part of swap
676 * space which also has its own migratepage callback. This
677 * is the most common path for page migration.
679 rc = mapping->a_ops->migratepage(mapping,
680 newpage, page, mode);
682 rc = fallback_migrate_page(mapping, newpage, page, mode);
685 newpage->mapping = NULL;
688 remove_migration_ptes(page, newpage);
691 unlock_page(newpage);
696 static int __unmap_and_move(struct page *page, struct page *newpage,
697 int force, bool offlining, enum migrate_mode mode)
700 int remap_swapcache = 1;
702 struct mem_cgroup *mem;
703 struct anon_vma *anon_vma = NULL;
705 if (!trylock_page(page)) {
706 if (!force || mode == MIGRATE_ASYNC)
710 * It's not safe for direct compaction to call lock_page.
711 * For example, during page readahead pages are added locked
712 * to the LRU. Later, when the IO completes the pages are
713 * marked uptodate and unlocked. However, the queueing
714 * could be merging multiple pages for one bio (e.g.
715 * mpage_readpages). If an allocation happens for the
716 * second or third page, the process can end up locking
717 * the same page twice and deadlocking. Rather than
718 * trying to be clever about what pages can be locked,
719 * avoid the use of lock_page for direct compaction
722 if (current->flags & PF_MEMALLOC)
729 * Only memory hotplug's offline_pages() caller has locked out KSM,
730 * and can safely migrate a KSM page. The other cases have skipped
731 * PageKsm along with PageReserved - but it is only now when we have
732 * the page lock that we can be certain it will not go KSM beneath us
733 * (KSM will not upgrade a page from PageAnon to PageKsm when it sees
734 * its pagecount raised, but only here do we take the page lock which
737 if (PageKsm(page) && !offlining) {
742 /* charge against new page */
743 charge = mem_cgroup_prepare_migration(page, newpage, &mem, GFP_KERNEL);
744 if (charge == -ENOMEM) {
750 if (PageWriteback(page)) {
752 * Only in the case of a full syncronous migration is it
753 * necessary to wait for PageWriteback. In the async case,
754 * the retry loop is too short and in the sync-light case,
755 * the overhead of stalling is too much
757 if (mode != MIGRATE_SYNC) {
763 wait_on_page_writeback(page);
766 * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
767 * we cannot notice that anon_vma is freed while we migrates a page.
768 * This get_anon_vma() delays freeing anon_vma pointer until the end
769 * of migration. File cache pages are no problem because of page_lock()
770 * File Caches may use write_page() or lock_page() in migration, then,
771 * just care Anon page here.
773 if (PageAnon(page)) {
775 * Only page_lock_anon_vma() understands the subtleties of
776 * getting a hold on an anon_vma from outside one of its mms.
778 anon_vma = page_get_anon_vma(page);
783 } else if (PageSwapCache(page)) {
785 * We cannot be sure that the anon_vma of an unmapped
786 * swapcache page is safe to use because we don't
787 * know in advance if the VMA that this page belonged
788 * to still exists. If the VMA and others sharing the
789 * data have been freed, then the anon_vma could
790 * already be invalid.
792 * To avoid this possibility, swapcache pages get
793 * migrated but are not remapped when migration
803 * Corner case handling:
804 * 1. When a new swap-cache page is read into, it is added to the LRU
805 * and treated as swapcache but it has no rmap yet.
806 * Calling try_to_unmap() against a page->mapping==NULL page will
807 * trigger a BUG. So handle it here.
808 * 2. An orphaned page (see truncate_complete_page) might have
809 * fs-private metadata. The page can be picked up due to memory
810 * offlining. Everywhere else except page reclaim, the page is
811 * invisible to the vm, so the page can not be migrated. So try to
812 * free the metadata, so the page can be freed.
814 if (!page->mapping) {
815 VM_BUG_ON(PageAnon(page));
816 if (page_has_private(page)) {
817 try_to_free_buffers(page);
823 /* Establish migration ptes or remove ptes */
824 try_to_unmap(page, TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
827 if (!page_mapped(page))
828 rc = move_to_new_page(newpage, page, remap_swapcache, mode);
830 if (rc && remap_swapcache)
831 remove_migration_ptes(page, page);
833 /* Drop an anon_vma reference if we took one */
835 put_anon_vma(anon_vma);
839 mem_cgroup_end_migration(mem, page, newpage, rc == 0);
847 * Obtain the lock on page, remove all ptes and migrate the page
848 * to the newly allocated page in newpage.
850 static int unmap_and_move(new_page_t get_new_page, unsigned long private,
851 struct page *page, int force, bool offlining,
852 enum migrate_mode mode)
856 struct page *newpage = get_new_page(page, private, &result);
861 if (page_count(page) == 1) {
862 /* page was freed from under us. So we are done. */
866 if (unlikely(PageTransHuge(page)))
867 if (unlikely(split_huge_page(page)))
870 rc = __unmap_and_move(page, newpage, force, offlining, mode);
874 * A page that has been migrated has all references
875 * removed and will be freed. A page that has not been
876 * migrated will have kepts its references and be
879 list_del(&page->lru);
880 dec_zone_page_state(page, NR_ISOLATED_ANON +
881 page_is_file_cache(page));
882 putback_lru_page(page);
885 * Move the new page to the LRU. If migration was not successful
886 * then this will free the page.
888 putback_lru_page(newpage);
893 *result = page_to_nid(newpage);
899 * Counterpart of unmap_and_move_page() for hugepage migration.
901 * This function doesn't wait the completion of hugepage I/O
902 * because there is no race between I/O and migration for hugepage.
903 * Note that currently hugepage I/O occurs only in direct I/O
904 * where no lock is held and PG_writeback is irrelevant,
905 * and writeback status of all subpages are counted in the reference
906 * count of the head page (i.e. if all subpages of a 2MB hugepage are
907 * under direct I/O, the reference of the head page is 512 and a bit more.)
908 * This means that when we try to migrate hugepage whose subpages are
909 * doing direct I/O, some references remain after try_to_unmap() and
910 * hugepage migration fails without data corruption.
912 * There is also no race when direct I/O is issued on the page under migration,
913 * because then pte is replaced with migration swap entry and direct I/O code
914 * will wait in the page fault for migration to complete.
916 static int unmap_and_move_huge_page(new_page_t get_new_page,
917 unsigned long private, struct page *hpage,
918 int force, bool offlining,
919 enum migrate_mode mode)
923 struct page *new_hpage = get_new_page(hpage, private, &result);
924 struct anon_vma *anon_vma = NULL;
931 if (!trylock_page(hpage)) {
932 if (!force || mode != MIGRATE_SYNC)
938 anon_vma = page_get_anon_vma(hpage);
940 try_to_unmap(hpage, TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
942 if (!page_mapped(hpage))
943 rc = move_to_new_page(new_hpage, hpage, 1, mode);
946 remove_migration_ptes(hpage, hpage);
949 put_anon_vma(anon_vma);
954 list_del(&hpage->lru);
964 *result = page_to_nid(new_hpage);
972 * The function takes one list of pages to migrate and a function
973 * that determines from the page to be migrated and the private data
974 * the target of the move and allocates the page.
976 * The function returns after 10 attempts or if no pages
977 * are movable anymore because to has become empty
978 * or no retryable pages exist anymore.
979 * Caller should call putback_lru_pages to return pages to the LRU
980 * or free list only if ret != 0.
982 * Return: Number of pages not migrated or error code.
984 int migrate_pages(struct list_head *from,
985 new_page_t get_new_page, unsigned long private, bool offlining,
986 enum migrate_mode mode)
993 int swapwrite = current->flags & PF_SWAPWRITE;
997 current->flags |= PF_SWAPWRITE;
999 for(pass = 0; pass < 10 && retry; pass++) {
1002 list_for_each_entry_safe(page, page2, from, lru) {
1005 rc = unmap_and_move(get_new_page, private,
1006 page, pass > 2, offlining,
1018 /* Permanent failure */
1027 current->flags &= ~PF_SWAPWRITE;
1032 return nr_failed + retry;
1035 int migrate_huge_pages(struct list_head *from,
1036 new_page_t get_new_page, unsigned long private, bool offlining,
1037 enum migrate_mode mode)
1046 for (pass = 0; pass < 10 && retry; pass++) {
1049 list_for_each_entry_safe(page, page2, from, lru) {
1052 rc = unmap_and_move_huge_page(get_new_page,
1053 private, page, pass > 2, offlining,
1065 /* Permanent failure */
1076 return nr_failed + retry;
1081 * Move a list of individual pages
1083 struct page_to_node {
1090 static struct page *new_page_node(struct page *p, unsigned long private,
1093 struct page_to_node *pm = (struct page_to_node *)private;
1095 while (pm->node != MAX_NUMNODES && pm->page != p)
1098 if (pm->node == MAX_NUMNODES)
1101 *result = &pm->status;
1103 return alloc_pages_exact_node(pm->node,
1104 GFP_HIGHUSER_MOVABLE | GFP_THISNODE, 0);
1108 * Move a set of pages as indicated in the pm array. The addr
1109 * field must be set to the virtual address of the page to be moved
1110 * and the node number must contain a valid target node.
1111 * The pm array ends with node = MAX_NUMNODES.
1113 static int do_move_page_to_node_array(struct mm_struct *mm,
1114 struct page_to_node *pm,
1118 struct page_to_node *pp;
1119 LIST_HEAD(pagelist);
1121 down_read(&mm->mmap_sem);
1124 * Build a list of pages to migrate
1126 for (pp = pm; pp->node != MAX_NUMNODES; pp++) {
1127 struct vm_area_struct *vma;
1131 vma = find_vma(mm, pp->addr);
1132 if (!vma || pp->addr < vma->vm_start || !vma_migratable(vma))
1135 page = follow_page(vma, pp->addr, FOLL_GET|FOLL_SPLIT);
1137 err = PTR_ERR(page);
1145 /* Use PageReserved to check for zero page */
1146 if (PageReserved(page) || PageKsm(page))
1150 err = page_to_nid(page);
1152 if (err == pp->node)
1154 * Node already in the right place
1159 if (page_mapcount(page) > 1 &&
1163 err = isolate_lru_page(page);
1165 list_add_tail(&page->lru, &pagelist);
1166 inc_zone_page_state(page, NR_ISOLATED_ANON +
1167 page_is_file_cache(page));
1171 * Either remove the duplicate refcount from
1172 * isolate_lru_page() or drop the page ref if it was
1181 if (!list_empty(&pagelist)) {
1182 err = migrate_pages(&pagelist, new_page_node,
1183 (unsigned long)pm, 0, MIGRATE_SYNC);
1185 putback_lru_pages(&pagelist);
1188 up_read(&mm->mmap_sem);
1193 * Migrate an array of page address onto an array of nodes and fill
1194 * the corresponding array of status.
1196 static int do_pages_move(struct mm_struct *mm, struct task_struct *task,
1197 unsigned long nr_pages,
1198 const void __user * __user *pages,
1199 const int __user *nodes,
1200 int __user *status, int flags)
1202 struct page_to_node *pm;
1203 nodemask_t task_nodes;
1204 unsigned long chunk_nr_pages;
1205 unsigned long chunk_start;
1208 task_nodes = cpuset_mems_allowed(task);
1211 pm = (struct page_to_node *)__get_free_page(GFP_KERNEL);
1218 * Store a chunk of page_to_node array in a page,
1219 * but keep the last one as a marker
1221 chunk_nr_pages = (PAGE_SIZE / sizeof(struct page_to_node)) - 1;
1223 for (chunk_start = 0;
1224 chunk_start < nr_pages;
1225 chunk_start += chunk_nr_pages) {
1228 if (chunk_start + chunk_nr_pages > nr_pages)
1229 chunk_nr_pages = nr_pages - chunk_start;
1231 /* fill the chunk pm with addrs and nodes from user-space */
1232 for (j = 0; j < chunk_nr_pages; j++) {
1233 const void __user *p;
1237 if (get_user(p, pages + j + chunk_start))
1239 pm[j].addr = (unsigned long) p;
1241 if (get_user(node, nodes + j + chunk_start))
1245 if (node < 0 || node >= MAX_NUMNODES)
1248 if (!node_state(node, N_HIGH_MEMORY))
1252 if (!node_isset(node, task_nodes))
1258 /* End marker for this chunk */
1259 pm[chunk_nr_pages].node = MAX_NUMNODES;
1261 /* Migrate this chunk */
1262 err = do_move_page_to_node_array(mm, pm,
1263 flags & MPOL_MF_MOVE_ALL);
1267 /* Return status information */
1268 for (j = 0; j < chunk_nr_pages; j++)
1269 if (put_user(pm[j].status, status + j + chunk_start)) {
1277 free_page((unsigned long)pm);
1283 * Determine the nodes of an array of pages and store it in an array of status.
1285 static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages,
1286 const void __user **pages, int *status)
1290 down_read(&mm->mmap_sem);
1292 for (i = 0; i < nr_pages; i++) {
1293 unsigned long addr = (unsigned long)(*pages);
1294 struct vm_area_struct *vma;
1298 vma = find_vma(mm, addr);
1299 if (!vma || addr < vma->vm_start)
1302 page = follow_page(vma, addr, 0);
1304 err = PTR_ERR(page);
1309 /* Use PageReserved to check for zero page */
1310 if (!page || PageReserved(page) || PageKsm(page))
1313 err = page_to_nid(page);
1321 up_read(&mm->mmap_sem);
1325 * Determine the nodes of a user array of pages and store it in
1326 * a user array of status.
1328 static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages,
1329 const void __user * __user *pages,
1332 #define DO_PAGES_STAT_CHUNK_NR 16
1333 const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR];
1334 int chunk_status[DO_PAGES_STAT_CHUNK_NR];
1337 unsigned long chunk_nr;
1339 chunk_nr = nr_pages;
1340 if (chunk_nr > DO_PAGES_STAT_CHUNK_NR)
1341 chunk_nr = DO_PAGES_STAT_CHUNK_NR;
1343 if (copy_from_user(chunk_pages, pages, chunk_nr * sizeof(*chunk_pages)))
1346 do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status);
1348 if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status)))
1353 nr_pages -= chunk_nr;
1355 return nr_pages ? -EFAULT : 0;
1359 * Move a list of pages in the address space of the currently executing
1362 SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages,
1363 const void __user * __user *, pages,
1364 const int __user *, nodes,
1365 int __user *, status, int, flags)
1367 const struct cred *cred = current_cred(), *tcred;
1368 struct task_struct *task;
1369 struct mm_struct *mm;
1373 if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
1376 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1379 /* Find the mm_struct */
1381 task = pid ? find_task_by_vpid(pid) : current;
1386 mm = get_task_mm(task);
1393 * Check if this process has the right to modify the specified
1394 * process. The right exists if the process has administrative
1395 * capabilities, superuser privileges or the same
1396 * userid as the target process.
1399 tcred = __task_cred(task);
1400 if (cred->euid != tcred->suid && cred->euid != tcred->uid &&
1401 cred->uid != tcred->suid && cred->uid != tcred->uid &&
1402 !capable(CAP_SYS_NICE)) {
1409 err = security_task_movememory(task);
1414 err = do_pages_move(mm, task, nr_pages, pages, nodes, status,
1417 err = do_pages_stat(mm, nr_pages, pages, status);
1426 * Call migration functions in the vma_ops that may prepare
1427 * memory in a vm for migration. migration functions may perform
1428 * the migration for vmas that do not have an underlying page struct.
1430 int migrate_vmas(struct mm_struct *mm, const nodemask_t *to,
1431 const nodemask_t *from, unsigned long flags)
1433 struct vm_area_struct *vma;
1436 for (vma = mm->mmap; vma && !err; vma = vma->vm_next) {
1437 if (vma->vm_ops && vma->vm_ops->migrate) {
1438 err = vma->vm_ops->migrate(vma, to, from, flags);