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));
145 /* Recheck VMA as permissions can change since migration started */
146 if (is_write_migration_entry(entry))
147 pte = maybe_mkwrite(pte, vma);
149 #ifdef CONFIG_HUGETLB_PAGE
151 pte = pte_mkhuge(pte);
153 flush_dcache_page(new);
154 set_pte_at(mm, addr, ptep, pte);
158 hugepage_add_anon_rmap(new, vma, addr);
161 } else if (PageAnon(new))
162 page_add_anon_rmap(new, vma, addr);
164 page_add_file_rmap(new);
166 /* No need to invalidate - it was non-present before */
167 update_mmu_cache(vma, addr, ptep);
169 pte_unmap_unlock(ptep, ptl);
175 * Get rid of all migration entries and replace them by
176 * references to the indicated page.
178 static void remove_migration_ptes(struct page *old, struct page *new)
180 rmap_walk(new, remove_migration_pte, old);
184 * Something used the pte of a page under migration. We need to
185 * get to the page and wait until migration is finished.
186 * When we return from this function the fault will be retried.
188 * This function is called from do_swap_page().
190 static void __migration_entry_wait(struct mm_struct *mm, pte_t *ptep,
199 if (!is_swap_pte(pte))
202 entry = pte_to_swp_entry(pte);
203 if (!is_migration_entry(entry))
206 page = migration_entry_to_page(entry);
209 * Once radix-tree replacement of page migration started, page_count
210 * *must* be zero. And, we don't want to call wait_on_page_locked()
211 * against a page without get_page().
212 * So, we use get_page_unless_zero(), here. Even failed, page fault
215 if (!get_page_unless_zero(page))
217 pte_unmap_unlock(ptep, ptl);
218 wait_on_page_locked(page);
222 pte_unmap_unlock(ptep, ptl);
225 void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
226 unsigned long address)
228 spinlock_t *ptl = pte_lockptr(mm, pmd);
229 pte_t *ptep = pte_offset_map(pmd, address);
230 __migration_entry_wait(mm, ptep, ptl);
233 void migration_entry_wait_huge(struct mm_struct *mm, pte_t *pte)
235 spinlock_t *ptl = &(mm)->page_table_lock;
236 __migration_entry_wait(mm, pte, ptl);
240 /* Returns true if all buffers are successfully locked */
241 static bool buffer_migrate_lock_buffers(struct buffer_head *head,
242 enum migrate_mode mode)
244 struct buffer_head *bh = head;
246 /* Simple case, sync compaction */
247 if (mode != MIGRATE_ASYNC) {
251 bh = bh->b_this_page;
253 } while (bh != head);
258 /* async case, we cannot block on lock_buffer so use trylock_buffer */
261 if (!trylock_buffer(bh)) {
263 * We failed to lock the buffer and cannot stall in
264 * async migration. Release the taken locks
266 struct buffer_head *failed_bh = bh;
269 while (bh != failed_bh) {
272 bh = bh->b_this_page;
277 bh = bh->b_this_page;
278 } while (bh != head);
282 static inline bool buffer_migrate_lock_buffers(struct buffer_head *head,
283 enum migrate_mode mode)
287 #endif /* CONFIG_BLOCK */
290 * Replace the page in the mapping.
292 * The number of remaining references must be:
293 * 1 for anonymous pages without a mapping
294 * 2 for pages with a mapping
295 * 3 for pages with a mapping and PagePrivate/PagePrivate2 set.
297 int migrate_page_move_mapping(struct address_space *mapping,
298 struct page *newpage, struct page *page,
299 struct buffer_head *head, enum migrate_mode mode)
305 /* Anonymous page without mapping */
306 if (page_count(page) != 1)
311 spin_lock_irq(&mapping->tree_lock);
313 pslot = radix_tree_lookup_slot(&mapping->page_tree,
316 expected_count = 2 + page_has_private(page);
317 if (page_count(page) != expected_count ||
318 radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) {
319 spin_unlock_irq(&mapping->tree_lock);
323 if (!page_freeze_refs(page, expected_count)) {
324 spin_unlock_irq(&mapping->tree_lock);
329 * In the async migration case of moving a page with buffers, lock the
330 * buffers using trylock before the mapping is moved. If the mapping
331 * was moved, we later failed to lock the buffers and could not move
332 * the mapping back due to an elevated page count, we would have to
333 * block waiting on other references to be dropped.
335 if (mode == MIGRATE_ASYNC && head &&
336 !buffer_migrate_lock_buffers(head, mode)) {
337 page_unfreeze_refs(page, expected_count);
338 spin_unlock_irq(&mapping->tree_lock);
343 * Now we know that no one else is looking at the page.
345 get_page(newpage); /* add cache reference */
346 if (PageSwapCache(page)) {
347 SetPageSwapCache(newpage);
348 set_page_private(newpage, page_private(page));
351 radix_tree_replace_slot(pslot, newpage);
353 page_unfreeze_refs(page, expected_count);
355 * Drop cache reference from old page.
356 * We know this isn't the last reference.
361 * If moved to a different zone then also account
362 * the page for that zone. Other VM counters will be
363 * taken care of when we establish references to the
364 * new page and drop references to the old page.
366 * Note that anonymous pages are accounted for
367 * via NR_FILE_PAGES and NR_ANON_PAGES if they
368 * are mapped to swap space.
370 __dec_zone_page_state(page, NR_FILE_PAGES);
371 __inc_zone_page_state(newpage, NR_FILE_PAGES);
372 if (!PageSwapCache(page) && PageSwapBacked(page)) {
373 __dec_zone_page_state(page, NR_SHMEM);
374 __inc_zone_page_state(newpage, NR_SHMEM);
376 spin_unlock_irq(&mapping->tree_lock);
380 EXPORT_SYMBOL(migrate_page_move_mapping);
383 * The expected number of remaining references is the same as that
384 * of migrate_page_move_mapping().
386 int migrate_huge_page_move_mapping(struct address_space *mapping,
387 struct page *newpage, struct page *page)
393 if (page_count(page) != 1)
398 spin_lock_irq(&mapping->tree_lock);
400 pslot = radix_tree_lookup_slot(&mapping->page_tree,
403 expected_count = 2 + page_has_private(page);
404 if (page_count(page) != expected_count ||
405 radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) {
406 spin_unlock_irq(&mapping->tree_lock);
410 if (!page_freeze_refs(page, expected_count)) {
411 spin_unlock_irq(&mapping->tree_lock);
417 radix_tree_replace_slot(pslot, newpage);
419 page_unfreeze_refs(page, expected_count);
423 spin_unlock_irq(&mapping->tree_lock);
428 * Copy the page to its new location
430 void migrate_page_copy(struct page *newpage, struct page *page)
433 copy_huge_page(newpage, page);
435 copy_highpage(newpage, page);
438 SetPageError(newpage);
439 if (PageReferenced(page))
440 SetPageReferenced(newpage);
441 if (PageUptodate(page))
442 SetPageUptodate(newpage);
443 if (TestClearPageActive(page)) {
444 VM_BUG_ON(PageUnevictable(page));
445 SetPageActive(newpage);
446 } else if (TestClearPageUnevictable(page))
447 SetPageUnevictable(newpage);
448 if (PageChecked(page))
449 SetPageChecked(newpage);
450 if (PageMappedToDisk(page))
451 SetPageMappedToDisk(newpage);
453 if (PageDirty(page)) {
454 clear_page_dirty_for_io(page);
456 * Want to mark the page and the radix tree as dirty, and
457 * redo the accounting that clear_page_dirty_for_io undid,
458 * but we can't use set_page_dirty because that function
459 * is actually a signal that all of the page has become dirty.
460 * Whereas only part of our page may be dirty.
462 __set_page_dirty_nobuffers(newpage);
465 mlock_migrate_page(newpage, page);
466 ksm_migrate_page(newpage, page);
468 ClearPageSwapCache(page);
469 ClearPagePrivate(page);
470 set_page_private(page, 0);
471 page->mapping = NULL;
474 * If any waiters have accumulated on the new page then
477 if (PageWriteback(newpage))
478 end_page_writeback(newpage);
480 EXPORT_SYMBOL(migrate_page_copy);
482 /************************************************************
483 * Migration functions
484 ***********************************************************/
486 /* Always fail migration. Used for mappings that are not movable */
487 int fail_migrate_page(struct address_space *mapping,
488 struct page *newpage, struct page *page)
492 EXPORT_SYMBOL(fail_migrate_page);
495 * Common logic to directly migrate a single page suitable for
496 * pages that do not use PagePrivate/PagePrivate2.
498 * Pages are locked upon entry and exit.
500 int migrate_page(struct address_space *mapping,
501 struct page *newpage, struct page *page,
502 enum migrate_mode mode)
506 BUG_ON(PageWriteback(page)); /* Writeback must be complete */
508 rc = migrate_page_move_mapping(mapping, newpage, page, NULL, mode);
513 migrate_page_copy(newpage, page);
516 EXPORT_SYMBOL(migrate_page);
520 * Migration function for pages with buffers. This function can only be used
521 * if the underlying filesystem guarantees that no other references to "page"
524 int buffer_migrate_page(struct address_space *mapping,
525 struct page *newpage, struct page *page, enum migrate_mode mode)
527 struct buffer_head *bh, *head;
530 if (!page_has_buffers(page))
531 return migrate_page(mapping, newpage, page, mode);
533 head = page_buffers(page);
535 rc = migrate_page_move_mapping(mapping, newpage, page, head, mode);
541 * In the async case, migrate_page_move_mapping locked the buffers
542 * with an IRQ-safe spinlock held. In the sync case, the buffers
543 * need to be locked now
545 if (mode != MIGRATE_ASYNC)
546 BUG_ON(!buffer_migrate_lock_buffers(head, mode));
548 ClearPagePrivate(page);
549 set_page_private(newpage, page_private(page));
550 set_page_private(page, 0);
556 set_bh_page(bh, newpage, bh_offset(bh));
557 bh = bh->b_this_page;
559 } while (bh != head);
561 SetPagePrivate(newpage);
563 migrate_page_copy(newpage, page);
569 bh = bh->b_this_page;
571 } while (bh != head);
575 EXPORT_SYMBOL(buffer_migrate_page);
579 * Writeback a page to clean the dirty state
581 static int writeout(struct address_space *mapping, struct page *page)
583 struct writeback_control wbc = {
584 .sync_mode = WB_SYNC_NONE,
587 .range_end = LLONG_MAX,
592 if (!mapping->a_ops->writepage)
593 /* No write method for the address space */
596 if (!clear_page_dirty_for_io(page))
597 /* Someone else already triggered a write */
601 * A dirty page may imply that the underlying filesystem has
602 * the page on some queue. So the page must be clean for
603 * migration. Writeout may mean we loose the lock and the
604 * page state is no longer what we checked for earlier.
605 * At this point we know that the migration attempt cannot
608 remove_migration_ptes(page, page);
610 rc = mapping->a_ops->writepage(page, &wbc);
612 if (rc != AOP_WRITEPAGE_ACTIVATE)
613 /* unlocked. Relock */
616 return (rc < 0) ? -EIO : -EAGAIN;
620 * Default handling if a filesystem does not provide a migration function.
622 static int fallback_migrate_page(struct address_space *mapping,
623 struct page *newpage, struct page *page, enum migrate_mode mode)
625 if (PageDirty(page)) {
626 /* Only writeback pages in full synchronous migration */
627 if (mode != MIGRATE_SYNC)
629 return writeout(mapping, page);
633 * Buffers may be managed in a filesystem specific way.
634 * We must have no buffers or drop them.
636 if (page_has_private(page) &&
637 !try_to_release_page(page, GFP_KERNEL))
640 return migrate_page(mapping, newpage, page, mode);
644 * Move a page to a newly allocated page
645 * The page is locked and all ptes have been successfully removed.
647 * The new page will have replaced the old page if this function
654 static int move_to_new_page(struct page *newpage, struct page *page,
655 int remap_swapcache, enum migrate_mode mode)
657 struct address_space *mapping;
661 * Block others from accessing the page when we get around to
662 * establishing additional references. We are the only one
663 * holding a reference to the new page at this point.
665 if (!trylock_page(newpage))
668 /* Prepare mapping for the new page.*/
669 newpage->index = page->index;
670 newpage->mapping = page->mapping;
671 if (PageSwapBacked(page))
672 SetPageSwapBacked(newpage);
674 mapping = page_mapping(page);
676 rc = migrate_page(mapping, newpage, page, mode);
677 else if (mapping->a_ops->migratepage)
679 * Most pages have a mapping and most filesystems provide a
680 * migratepage callback. Anonymous pages are part of swap
681 * space which also has its own migratepage callback. This
682 * is the most common path for page migration.
684 rc = mapping->a_ops->migratepage(mapping,
685 newpage, page, mode);
687 rc = fallback_migrate_page(mapping, newpage, page, mode);
690 newpage->mapping = NULL;
693 remove_migration_ptes(page, newpage);
696 unlock_page(newpage);
701 static int __unmap_and_move(struct page *page, struct page *newpage,
702 int force, bool offlining, enum migrate_mode mode)
705 int remap_swapcache = 1;
707 struct mem_cgroup *mem;
708 struct anon_vma *anon_vma = NULL;
710 if (!trylock_page(page)) {
711 if (!force || mode == MIGRATE_ASYNC)
715 * It's not safe for direct compaction to call lock_page.
716 * For example, during page readahead pages are added locked
717 * to the LRU. Later, when the IO completes the pages are
718 * marked uptodate and unlocked. However, the queueing
719 * could be merging multiple pages for one bio (e.g.
720 * mpage_readpages). If an allocation happens for the
721 * second or third page, the process can end up locking
722 * the same page twice and deadlocking. Rather than
723 * trying to be clever about what pages can be locked,
724 * avoid the use of lock_page for direct compaction
727 if (current->flags & PF_MEMALLOC)
734 * Only memory hotplug's offline_pages() caller has locked out KSM,
735 * and can safely migrate a KSM page. The other cases have skipped
736 * PageKsm along with PageReserved - but it is only now when we have
737 * the page lock that we can be certain it will not go KSM beneath us
738 * (KSM will not upgrade a page from PageAnon to PageKsm when it sees
739 * its pagecount raised, but only here do we take the page lock which
742 if (PageKsm(page) && !offlining) {
747 /* charge against new page */
748 charge = mem_cgroup_prepare_migration(page, newpage, &mem, GFP_KERNEL);
749 if (charge == -ENOMEM) {
755 if (PageWriteback(page)) {
757 * Only in the case of a full syncronous migration is it
758 * necessary to wait for PageWriteback. In the async case,
759 * the retry loop is too short and in the sync-light case,
760 * the overhead of stalling is too much
762 if (mode != MIGRATE_SYNC) {
768 wait_on_page_writeback(page);
771 * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
772 * we cannot notice that anon_vma is freed while we migrates a page.
773 * This get_anon_vma() delays freeing anon_vma pointer until the end
774 * of migration. File cache pages are no problem because of page_lock()
775 * File Caches may use write_page() or lock_page() in migration, then,
776 * just care Anon page here.
778 if (PageAnon(page)) {
780 * Only page_lock_anon_vma() understands the subtleties of
781 * getting a hold on an anon_vma from outside one of its mms.
783 anon_vma = page_get_anon_vma(page);
788 } else if (PageSwapCache(page)) {
790 * We cannot be sure that the anon_vma of an unmapped
791 * swapcache page is safe to use because we don't
792 * know in advance if the VMA that this page belonged
793 * to still exists. If the VMA and others sharing the
794 * data have been freed, then the anon_vma could
795 * already be invalid.
797 * To avoid this possibility, swapcache pages get
798 * migrated but are not remapped when migration
808 * Corner case handling:
809 * 1. When a new swap-cache page is read into, it is added to the LRU
810 * and treated as swapcache but it has no rmap yet.
811 * Calling try_to_unmap() against a page->mapping==NULL page will
812 * trigger a BUG. So handle it here.
813 * 2. An orphaned page (see truncate_complete_page) might have
814 * fs-private metadata. The page can be picked up due to memory
815 * offlining. Everywhere else except page reclaim, the page is
816 * invisible to the vm, so the page can not be migrated. So try to
817 * free the metadata, so the page can be freed.
819 if (!page->mapping) {
820 VM_BUG_ON(PageAnon(page));
821 if (page_has_private(page)) {
822 try_to_free_buffers(page);
828 /* Establish migration ptes or remove ptes */
829 try_to_unmap(page, TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
832 if (!page_mapped(page))
833 rc = move_to_new_page(newpage, page, remap_swapcache, mode);
835 if (rc && remap_swapcache)
836 remove_migration_ptes(page, page);
838 /* Drop an anon_vma reference if we took one */
840 put_anon_vma(anon_vma);
844 mem_cgroup_end_migration(mem, page, newpage, rc == 0);
852 * Obtain the lock on page, remove all ptes and migrate the page
853 * to the newly allocated page in newpage.
855 static int unmap_and_move(new_page_t get_new_page, unsigned long private,
856 struct page *page, int force, bool offlining,
857 enum migrate_mode mode)
861 struct page *newpage = get_new_page(page, private, &result);
866 if (page_count(page) == 1) {
867 /* page was freed from under us. So we are done. */
871 if (unlikely(PageTransHuge(page)))
872 if (unlikely(split_huge_page(page)))
875 rc = __unmap_and_move(page, newpage, force, offlining, mode);
879 * A page that has been migrated has all references
880 * removed and will be freed. A page that has not been
881 * migrated will have kepts its references and be
884 list_del(&page->lru);
885 dec_zone_page_state(page, NR_ISOLATED_ANON +
886 page_is_file_cache(page));
887 putback_lru_page(page);
890 * Move the new page to the LRU. If migration was not successful
891 * then this will free the page.
893 putback_lru_page(newpage);
898 *result = page_to_nid(newpage);
904 * Counterpart of unmap_and_move_page() for hugepage migration.
906 * This function doesn't wait the completion of hugepage I/O
907 * because there is no race between I/O and migration for hugepage.
908 * Note that currently hugepage I/O occurs only in direct I/O
909 * where no lock is held and PG_writeback is irrelevant,
910 * and writeback status of all subpages are counted in the reference
911 * count of the head page (i.e. if all subpages of a 2MB hugepage are
912 * under direct I/O, the reference of the head page is 512 and a bit more.)
913 * This means that when we try to migrate hugepage whose subpages are
914 * doing direct I/O, some references remain after try_to_unmap() and
915 * hugepage migration fails without data corruption.
917 * There is also no race when direct I/O is issued on the page under migration,
918 * because then pte is replaced with migration swap entry and direct I/O code
919 * will wait in the page fault for migration to complete.
921 static int unmap_and_move_huge_page(new_page_t get_new_page,
922 unsigned long private, struct page *hpage,
923 int force, bool offlining,
924 enum migrate_mode mode)
928 struct page *new_hpage = get_new_page(hpage, private, &result);
929 struct anon_vma *anon_vma = NULL;
936 if (!trylock_page(hpage)) {
937 if (!force || mode != MIGRATE_SYNC)
943 anon_vma = page_get_anon_vma(hpage);
945 try_to_unmap(hpage, TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
947 if (!page_mapped(hpage))
948 rc = move_to_new_page(new_hpage, hpage, 1, mode);
951 remove_migration_ptes(hpage, hpage);
954 put_anon_vma(anon_vma);
959 list_del(&hpage->lru);
969 *result = page_to_nid(new_hpage);
977 * The function takes one list of pages to migrate and a function
978 * that determines from the page to be migrated and the private data
979 * the target of the move and allocates the page.
981 * The function returns after 10 attempts or if no pages
982 * are movable anymore because to has become empty
983 * or no retryable pages exist anymore.
984 * Caller should call putback_lru_pages to return pages to the LRU
985 * or free list only if ret != 0.
987 * Return: Number of pages not migrated or error code.
989 int migrate_pages(struct list_head *from,
990 new_page_t get_new_page, unsigned long private, bool offlining,
991 enum migrate_mode mode)
998 int swapwrite = current->flags & PF_SWAPWRITE;
1002 current->flags |= PF_SWAPWRITE;
1004 for(pass = 0; pass < 10 && retry; pass++) {
1007 list_for_each_entry_safe(page, page2, from, lru) {
1010 rc = unmap_and_move(get_new_page, private,
1011 page, pass > 2, offlining,
1023 /* Permanent failure */
1032 current->flags &= ~PF_SWAPWRITE;
1037 return nr_failed + retry;
1040 int migrate_huge_pages(struct list_head *from,
1041 new_page_t get_new_page, unsigned long private, bool offlining,
1042 enum migrate_mode mode)
1051 for (pass = 0; pass < 10 && retry; pass++) {
1054 list_for_each_entry_safe(page, page2, from, lru) {
1057 rc = unmap_and_move_huge_page(get_new_page,
1058 private, page, pass > 2, offlining,
1070 /* Permanent failure */
1081 return nr_failed + retry;
1086 * Move a list of individual pages
1088 struct page_to_node {
1095 static struct page *new_page_node(struct page *p, unsigned long private,
1098 struct page_to_node *pm = (struct page_to_node *)private;
1100 while (pm->node != MAX_NUMNODES && pm->page != p)
1103 if (pm->node == MAX_NUMNODES)
1106 *result = &pm->status;
1108 return alloc_pages_exact_node(pm->node,
1109 GFP_HIGHUSER_MOVABLE | GFP_THISNODE, 0);
1113 * Move a set of pages as indicated in the pm array. The addr
1114 * field must be set to the virtual address of the page to be moved
1115 * and the node number must contain a valid target node.
1116 * The pm array ends with node = MAX_NUMNODES.
1118 static int do_move_page_to_node_array(struct mm_struct *mm,
1119 struct page_to_node *pm,
1123 struct page_to_node *pp;
1124 LIST_HEAD(pagelist);
1126 down_read(&mm->mmap_sem);
1129 * Build a list of pages to migrate
1131 for (pp = pm; pp->node != MAX_NUMNODES; pp++) {
1132 struct vm_area_struct *vma;
1136 vma = find_vma(mm, pp->addr);
1137 if (!vma || pp->addr < vma->vm_start || !vma_migratable(vma))
1140 page = follow_page(vma, pp->addr, FOLL_GET|FOLL_SPLIT);
1142 err = PTR_ERR(page);
1150 /* Use PageReserved to check for zero page */
1151 if (PageReserved(page) || PageKsm(page))
1155 err = page_to_nid(page);
1157 if (err == pp->node)
1159 * Node already in the right place
1164 if (page_mapcount(page) > 1 &&
1168 err = isolate_lru_page(page);
1170 list_add_tail(&page->lru, &pagelist);
1171 inc_zone_page_state(page, NR_ISOLATED_ANON +
1172 page_is_file_cache(page));
1176 * Either remove the duplicate refcount from
1177 * isolate_lru_page() or drop the page ref if it was
1186 if (!list_empty(&pagelist)) {
1187 err = migrate_pages(&pagelist, new_page_node,
1188 (unsigned long)pm, 0, MIGRATE_SYNC);
1190 putback_lru_pages(&pagelist);
1193 up_read(&mm->mmap_sem);
1198 * Migrate an array of page address onto an array of nodes and fill
1199 * the corresponding array of status.
1201 static int do_pages_move(struct mm_struct *mm, struct task_struct *task,
1202 unsigned long nr_pages,
1203 const void __user * __user *pages,
1204 const int __user *nodes,
1205 int __user *status, int flags)
1207 struct page_to_node *pm;
1208 nodemask_t task_nodes;
1209 unsigned long chunk_nr_pages;
1210 unsigned long chunk_start;
1213 task_nodes = cpuset_mems_allowed(task);
1216 pm = (struct page_to_node *)__get_free_page(GFP_KERNEL);
1223 * Store a chunk of page_to_node array in a page,
1224 * but keep the last one as a marker
1226 chunk_nr_pages = (PAGE_SIZE / sizeof(struct page_to_node)) - 1;
1228 for (chunk_start = 0;
1229 chunk_start < nr_pages;
1230 chunk_start += chunk_nr_pages) {
1233 if (chunk_start + chunk_nr_pages > nr_pages)
1234 chunk_nr_pages = nr_pages - chunk_start;
1236 /* fill the chunk pm with addrs and nodes from user-space */
1237 for (j = 0; j < chunk_nr_pages; j++) {
1238 const void __user *p;
1242 if (get_user(p, pages + j + chunk_start))
1244 pm[j].addr = (unsigned long) p;
1246 if (get_user(node, nodes + j + chunk_start))
1250 if (node < 0 || node >= MAX_NUMNODES)
1253 if (!node_state(node, N_HIGH_MEMORY))
1257 if (!node_isset(node, task_nodes))
1263 /* End marker for this chunk */
1264 pm[chunk_nr_pages].node = MAX_NUMNODES;
1266 /* Migrate this chunk */
1267 err = do_move_page_to_node_array(mm, pm,
1268 flags & MPOL_MF_MOVE_ALL);
1272 /* Return status information */
1273 for (j = 0; j < chunk_nr_pages; j++)
1274 if (put_user(pm[j].status, status + j + chunk_start)) {
1282 free_page((unsigned long)pm);
1288 * Determine the nodes of an array of pages and store it in an array of status.
1290 static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages,
1291 const void __user **pages, int *status)
1295 down_read(&mm->mmap_sem);
1297 for (i = 0; i < nr_pages; i++) {
1298 unsigned long addr = (unsigned long)(*pages);
1299 struct vm_area_struct *vma;
1303 vma = find_vma(mm, addr);
1304 if (!vma || addr < vma->vm_start)
1307 page = follow_page(vma, addr, 0);
1309 err = PTR_ERR(page);
1314 /* Use PageReserved to check for zero page */
1315 if (!page || PageReserved(page) || PageKsm(page))
1318 err = page_to_nid(page);
1326 up_read(&mm->mmap_sem);
1330 * Determine the nodes of a user array of pages and store it in
1331 * a user array of status.
1333 static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages,
1334 const void __user * __user *pages,
1337 #define DO_PAGES_STAT_CHUNK_NR 16
1338 const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR];
1339 int chunk_status[DO_PAGES_STAT_CHUNK_NR];
1342 unsigned long chunk_nr;
1344 chunk_nr = nr_pages;
1345 if (chunk_nr > DO_PAGES_STAT_CHUNK_NR)
1346 chunk_nr = DO_PAGES_STAT_CHUNK_NR;
1348 if (copy_from_user(chunk_pages, pages, chunk_nr * sizeof(*chunk_pages)))
1351 do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status);
1353 if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status)))
1358 nr_pages -= chunk_nr;
1360 return nr_pages ? -EFAULT : 0;
1364 * Move a list of pages in the address space of the currently executing
1367 SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages,
1368 const void __user * __user *, pages,
1369 const int __user *, nodes,
1370 int __user *, status, int, flags)
1372 const struct cred *cred = current_cred(), *tcred;
1373 struct task_struct *task;
1374 struct mm_struct *mm;
1378 if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
1381 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1384 /* Find the mm_struct */
1386 task = pid ? find_task_by_vpid(pid) : current;
1391 mm = get_task_mm(task);
1398 * Check if this process has the right to modify the specified
1399 * process. The right exists if the process has administrative
1400 * capabilities, superuser privileges or the same
1401 * userid as the target process.
1404 tcred = __task_cred(task);
1405 if (cred->euid != tcred->suid && cred->euid != tcred->uid &&
1406 cred->uid != tcred->suid && cred->uid != tcred->uid &&
1407 !capable(CAP_SYS_NICE)) {
1414 err = security_task_movememory(task);
1419 err = do_pages_move(mm, task, nr_pages, pages, nodes, status,
1422 err = do_pages_stat(mm, nr_pages, pages, status);
1431 * Call migration functions in the vma_ops that may prepare
1432 * memory in a vm for migration. migration functions may perform
1433 * the migration for vmas that do not have an underlying page struct.
1435 int migrate_vmas(struct mm_struct *mm, const nodemask_t *to,
1436 const nodemask_t *from, unsigned long flags)
1438 struct vm_area_struct *vma;
1441 for (vma = mm->mmap; vma && !err; vma = vma->vm_next) {
1442 if (vma->vm_ops && vma->vm_ops->migrate) {
1443 err = vma->vm_ops->migrate(vma, to, from, flags);