2 * Memory Migration functionality - linux/mm/migration.c
4 * Copyright (C) 2006 Silicon Graphics, Inc., Christoph Lameter
6 * Page migration was first developed in the context of the memory hotplug
7 * project. The main authors of the migration code are:
9 * IWAMOTO Toshihiro <iwamoto@valinux.co.jp>
10 * Hirokazu Takahashi <taka@valinux.co.jp>
11 * Dave Hansen <haveblue@us.ibm.com>
15 #include <linux/migrate.h>
16 #include <linux/export.h>
17 #include <linux/swap.h>
18 #include <linux/swapops.h>
19 #include <linux/pagemap.h>
20 #include <linux/buffer_head.h>
21 #include <linux/mm_inline.h>
22 #include <linux/nsproxy.h>
23 #include <linux/pagevec.h>
24 #include <linux/ksm.h>
25 #include <linux/rmap.h>
26 #include <linux/topology.h>
27 #include <linux/cpu.h>
28 #include <linux/cpuset.h>
29 #include <linux/writeback.h>
30 #include <linux/mempolicy.h>
31 #include <linux/vmalloc.h>
32 #include <linux/security.h>
33 #include <linux/memcontrol.h>
34 #include <linux/syscalls.h>
35 #include <linux/hugetlb.h>
36 #include <linux/hugetlb_cgroup.h>
37 #include <linux/gfp.h>
38 #include <linux/balloon_compaction.h>
40 #include <asm/tlbflush.h>
45 * migrate_prep() needs to be called before we start compiling a list of pages
46 * to be migrated using isolate_lru_page(). If scheduling work on other CPUs is
47 * undesirable, use migrate_prep_local()
49 int migrate_prep(void)
52 * Clear the LRU lists so pages can be isolated.
53 * Note that pages may be moved off the LRU after we have
54 * drained them. Those pages will fail to migrate like other
55 * pages that may be busy.
62 /* Do the necessary work of migrate_prep but not if it involves other CPUs */
63 int migrate_prep_local(void)
71 * Add isolated pages on the list back to the LRU under page lock
72 * to avoid leaking evictable pages back onto unevictable list.
74 void putback_lru_pages(struct list_head
*l
)
79 list_for_each_entry_safe(page
, page2
, l
, lru
) {
81 dec_zone_page_state(page
, NR_ISOLATED_ANON
+
82 page_is_file_cache(page
));
83 putback_lru_page(page
);
88 * Put previously isolated pages back onto the appropriate lists
89 * from where they were once taken off for compaction/migration.
91 * This function shall be used instead of putback_lru_pages(),
92 * whenever the isolated pageset has been built by isolate_migratepages_range()
94 void putback_movable_pages(struct list_head
*l
)
99 list_for_each_entry_safe(page
, page2
, l
, lru
) {
100 list_del(&page
->lru
);
101 dec_zone_page_state(page
, NR_ISOLATED_ANON
+
102 page_is_file_cache(page
));
103 if (unlikely(balloon_page_movable(page
)))
104 balloon_page_putback(page
);
106 putback_lru_page(page
);
111 * Restore a potential migration pte to a working pte entry
113 static int remove_migration_pte(struct page
*new, struct vm_area_struct
*vma
,
114 unsigned long addr
, void *old
)
116 struct mm_struct
*mm
= vma
->vm_mm
;
122 if (unlikely(PageHuge(new))) {
123 ptep
= huge_pte_offset(mm
, addr
);
126 ptl
= &mm
->page_table_lock
;
128 pmd
= mm_find_pmd(mm
, addr
);
131 if (pmd_trans_huge(*pmd
))
134 ptep
= pte_offset_map(pmd
, addr
);
137 * Peek to check is_swap_pte() before taking ptlock? No, we
138 * can race mremap's move_ptes(), which skips anon_vma lock.
141 ptl
= pte_lockptr(mm
, pmd
);
146 if (!is_swap_pte(pte
))
149 entry
= pte_to_swp_entry(pte
);
151 if (!is_migration_entry(entry
) ||
152 migration_entry_to_page(entry
) != old
)
156 pte
= pte_mkold(mk_pte(new, vma
->vm_page_prot
));
157 if (is_write_migration_entry(entry
))
158 pte
= pte_mkwrite(pte
);
159 #ifdef CONFIG_HUGETLB_PAGE
161 pte
= pte_mkhuge(pte
);
163 flush_cache_page(vma
, addr
, pte_pfn(pte
));
164 set_pte_at(mm
, addr
, ptep
, pte
);
168 hugepage_add_anon_rmap(new, vma
, addr
);
171 } else if (PageAnon(new))
172 page_add_anon_rmap(new, vma
, addr
);
174 page_add_file_rmap(new);
176 /* No need to invalidate - it was non-present before */
177 update_mmu_cache(vma
, addr
, ptep
);
179 pte_unmap_unlock(ptep
, ptl
);
185 * Get rid of all migration entries and replace them by
186 * references to the indicated page.
188 static void remove_migration_ptes(struct page
*old
, struct page
*new)
190 rmap_walk(new, remove_migration_pte
, old
);
194 * Something used the pte of a page under migration. We need to
195 * get to the page and wait until migration is finished.
196 * When we return from this function the fault will be retried.
198 void migration_entry_wait(struct mm_struct
*mm
, pmd_t
*pmd
,
199 unsigned long address
)
206 ptep
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
208 if (!is_swap_pte(pte
))
211 entry
= pte_to_swp_entry(pte
);
212 if (!is_migration_entry(entry
))
215 page
= migration_entry_to_page(entry
);
218 * Once radix-tree replacement of page migration started, page_count
219 * *must* be zero. And, we don't want to call wait_on_page_locked()
220 * against a page without get_page().
221 * So, we use get_page_unless_zero(), here. Even failed, page fault
224 if (!get_page_unless_zero(page
))
226 pte_unmap_unlock(ptep
, ptl
);
227 wait_on_page_locked(page
);
231 pte_unmap_unlock(ptep
, ptl
);
235 /* Returns true if all buffers are successfully locked */
236 static bool buffer_migrate_lock_buffers(struct buffer_head
*head
,
237 enum migrate_mode mode
)
239 struct buffer_head
*bh
= head
;
241 /* Simple case, sync compaction */
242 if (mode
!= MIGRATE_ASYNC
) {
246 bh
= bh
->b_this_page
;
248 } while (bh
!= head
);
253 /* async case, we cannot block on lock_buffer so use trylock_buffer */
256 if (!trylock_buffer(bh
)) {
258 * We failed to lock the buffer and cannot stall in
259 * async migration. Release the taken locks
261 struct buffer_head
*failed_bh
= bh
;
264 while (bh
!= failed_bh
) {
267 bh
= bh
->b_this_page
;
272 bh
= bh
->b_this_page
;
273 } while (bh
!= head
);
277 static inline bool buffer_migrate_lock_buffers(struct buffer_head
*head
,
278 enum migrate_mode mode
)
282 #endif /* CONFIG_BLOCK */
285 * Replace the page in the mapping.
287 * The number of remaining references must be:
288 * 1 for anonymous pages without a mapping
289 * 2 for pages with a mapping
290 * 3 for pages with a mapping and PagePrivate/PagePrivate2 set.
292 static int migrate_page_move_mapping(struct address_space
*mapping
,
293 struct page
*newpage
, struct page
*page
,
294 struct buffer_head
*head
, enum migrate_mode mode
)
300 /* Anonymous page without mapping */
301 if (page_count(page
) != 1)
303 return MIGRATEPAGE_SUCCESS
;
306 spin_lock_irq(&mapping
->tree_lock
);
308 pslot
= radix_tree_lookup_slot(&mapping
->page_tree
,
311 expected_count
= 2 + page_has_private(page
);
312 if (page_count(page
) != expected_count
||
313 radix_tree_deref_slot_protected(pslot
, &mapping
->tree_lock
) != page
) {
314 spin_unlock_irq(&mapping
->tree_lock
);
318 if (!page_freeze_refs(page
, expected_count
)) {
319 spin_unlock_irq(&mapping
->tree_lock
);
324 * In the async migration case of moving a page with buffers, lock the
325 * buffers using trylock before the mapping is moved. If the mapping
326 * was moved, we later failed to lock the buffers and could not move
327 * the mapping back due to an elevated page count, we would have to
328 * block waiting on other references to be dropped.
330 if (mode
== MIGRATE_ASYNC
&& head
&&
331 !buffer_migrate_lock_buffers(head
, mode
)) {
332 page_unfreeze_refs(page
, expected_count
);
333 spin_unlock_irq(&mapping
->tree_lock
);
338 * Now we know that no one else is looking at the page.
340 get_page(newpage
); /* add cache reference */
341 if (PageSwapCache(page
)) {
342 SetPageSwapCache(newpage
);
343 set_page_private(newpage
, page_private(page
));
346 radix_tree_replace_slot(pslot
, newpage
);
349 * Drop cache reference from old page by unfreezing
350 * to one less reference.
351 * We know this isn't the last reference.
353 page_unfreeze_refs(page
, expected_count
- 1);
356 * If moved to a different zone then also account
357 * the page for that zone. Other VM counters will be
358 * taken care of when we establish references to the
359 * new page and drop references to the old page.
361 * Note that anonymous pages are accounted for
362 * via NR_FILE_PAGES and NR_ANON_PAGES if they
363 * are mapped to swap space.
365 __dec_zone_page_state(page
, NR_FILE_PAGES
);
366 __inc_zone_page_state(newpage
, NR_FILE_PAGES
);
367 if (!PageSwapCache(page
) && PageSwapBacked(page
)) {
368 __dec_zone_page_state(page
, NR_SHMEM
);
369 __inc_zone_page_state(newpage
, NR_SHMEM
);
371 spin_unlock_irq(&mapping
->tree_lock
);
373 return MIGRATEPAGE_SUCCESS
;
377 * The expected number of remaining references is the same as that
378 * of migrate_page_move_mapping().
380 int migrate_huge_page_move_mapping(struct address_space
*mapping
,
381 struct page
*newpage
, struct page
*page
)
387 if (page_count(page
) != 1)
389 return MIGRATEPAGE_SUCCESS
;
392 spin_lock_irq(&mapping
->tree_lock
);
394 pslot
= radix_tree_lookup_slot(&mapping
->page_tree
,
397 expected_count
= 2 + page_has_private(page
);
398 if (page_count(page
) != expected_count
||
399 radix_tree_deref_slot_protected(pslot
, &mapping
->tree_lock
) != page
) {
400 spin_unlock_irq(&mapping
->tree_lock
);
404 if (!page_freeze_refs(page
, expected_count
)) {
405 spin_unlock_irq(&mapping
->tree_lock
);
411 radix_tree_replace_slot(pslot
, newpage
);
413 page_unfreeze_refs(page
, expected_count
- 1);
415 spin_unlock_irq(&mapping
->tree_lock
);
416 return MIGRATEPAGE_SUCCESS
;
420 * Copy the page to its new location
422 void migrate_page_copy(struct page
*newpage
, struct page
*page
)
425 copy_huge_page(newpage
, page
);
427 copy_highpage(newpage
, page
);
430 SetPageError(newpage
);
431 if (PageReferenced(page
))
432 SetPageReferenced(newpage
);
433 if (PageUptodate(page
))
434 SetPageUptodate(newpage
);
435 if (TestClearPageActive(page
)) {
436 VM_BUG_ON(PageUnevictable(page
));
437 SetPageActive(newpage
);
438 } else if (TestClearPageUnevictable(page
))
439 SetPageUnevictable(newpage
);
440 if (PageChecked(page
))
441 SetPageChecked(newpage
);
442 if (PageMappedToDisk(page
))
443 SetPageMappedToDisk(newpage
);
445 if (PageDirty(page
)) {
446 clear_page_dirty_for_io(page
);
448 * Want to mark the page and the radix tree as dirty, and
449 * redo the accounting that clear_page_dirty_for_io undid,
450 * but we can't use set_page_dirty because that function
451 * is actually a signal that all of the page has become dirty.
452 * Whereas only part of our page may be dirty.
454 if (PageSwapBacked(page
))
455 SetPageDirty(newpage
);
457 __set_page_dirty_nobuffers(newpage
);
460 mlock_migrate_page(newpage
, page
);
461 ksm_migrate_page(newpage
, page
);
463 ClearPageSwapCache(page
);
464 ClearPagePrivate(page
);
465 set_page_private(page
, 0);
468 * If any waiters have accumulated on the new page then
471 if (PageWriteback(newpage
))
472 end_page_writeback(newpage
);
475 /************************************************************
476 * Migration functions
477 ***********************************************************/
479 /* Always fail migration. Used for mappings that are not movable */
480 int fail_migrate_page(struct address_space
*mapping
,
481 struct page
*newpage
, struct page
*page
)
485 EXPORT_SYMBOL(fail_migrate_page
);
488 * Common logic to directly migrate a single page suitable for
489 * pages that do not use PagePrivate/PagePrivate2.
491 * Pages are locked upon entry and exit.
493 int migrate_page(struct address_space
*mapping
,
494 struct page
*newpage
, struct page
*page
,
495 enum migrate_mode mode
)
499 BUG_ON(PageWriteback(page
)); /* Writeback must be complete */
501 rc
= migrate_page_move_mapping(mapping
, newpage
, page
, NULL
, mode
);
503 if (rc
!= MIGRATEPAGE_SUCCESS
)
506 migrate_page_copy(newpage
, page
);
507 return MIGRATEPAGE_SUCCESS
;
509 EXPORT_SYMBOL(migrate_page
);
513 * Migration function for pages with buffers. This function can only be used
514 * if the underlying filesystem guarantees that no other references to "page"
517 int buffer_migrate_page(struct address_space
*mapping
,
518 struct page
*newpage
, struct page
*page
, enum migrate_mode mode
)
520 struct buffer_head
*bh
, *head
;
523 if (!page_has_buffers(page
))
524 return migrate_page(mapping
, newpage
, page
, mode
);
526 head
= page_buffers(page
);
528 rc
= migrate_page_move_mapping(mapping
, newpage
, page
, head
, mode
);
530 if (rc
!= MIGRATEPAGE_SUCCESS
)
534 * In the async case, migrate_page_move_mapping locked the buffers
535 * with an IRQ-safe spinlock held. In the sync case, the buffers
536 * need to be locked now
538 if (mode
!= MIGRATE_ASYNC
)
539 BUG_ON(!buffer_migrate_lock_buffers(head
, mode
));
541 ClearPagePrivate(page
);
542 set_page_private(newpage
, page_private(page
));
543 set_page_private(page
, 0);
549 set_bh_page(bh
, newpage
, bh_offset(bh
));
550 bh
= bh
->b_this_page
;
552 } while (bh
!= head
);
554 SetPagePrivate(newpage
);
556 migrate_page_copy(newpage
, page
);
562 bh
= bh
->b_this_page
;
564 } while (bh
!= head
);
566 return MIGRATEPAGE_SUCCESS
;
568 EXPORT_SYMBOL(buffer_migrate_page
);
572 * Writeback a page to clean the dirty state
574 static int writeout(struct address_space
*mapping
, struct page
*page
)
576 struct writeback_control wbc
= {
577 .sync_mode
= WB_SYNC_NONE
,
580 .range_end
= LLONG_MAX
,
585 if (!mapping
->a_ops
->writepage
)
586 /* No write method for the address space */
589 if (!clear_page_dirty_for_io(page
))
590 /* Someone else already triggered a write */
594 * A dirty page may imply that the underlying filesystem has
595 * the page on some queue. So the page must be clean for
596 * migration. Writeout may mean we loose the lock and the
597 * page state is no longer what we checked for earlier.
598 * At this point we know that the migration attempt cannot
601 remove_migration_ptes(page
, page
);
603 rc
= mapping
->a_ops
->writepage(page
, &wbc
);
605 if (rc
!= AOP_WRITEPAGE_ACTIVATE
)
606 /* unlocked. Relock */
609 return (rc
< 0) ? -EIO
: -EAGAIN
;
613 * Default handling if a filesystem does not provide a migration function.
615 static int fallback_migrate_page(struct address_space
*mapping
,
616 struct page
*newpage
, struct page
*page
, enum migrate_mode mode
)
618 if (PageDirty(page
)) {
619 /* Only writeback pages in full synchronous migration */
620 if (mode
!= MIGRATE_SYNC
)
622 return writeout(mapping
, page
);
626 * Buffers may be managed in a filesystem specific way.
627 * We must have no buffers or drop them.
629 if (page_has_private(page
) &&
630 !try_to_release_page(page
, GFP_KERNEL
))
633 return migrate_page(mapping
, newpage
, page
, mode
);
637 * Move a page to a newly allocated page
638 * The page is locked and all ptes have been successfully removed.
640 * The new page will have replaced the old page if this function
645 * MIGRATEPAGE_SUCCESS - success
647 static int move_to_new_page(struct page
*newpage
, struct page
*page
,
648 int remap_swapcache
, enum migrate_mode mode
)
650 struct address_space
*mapping
;
654 * Block others from accessing the page when we get around to
655 * establishing additional references. We are the only one
656 * holding a reference to the new page at this point.
658 if (!trylock_page(newpage
))
661 /* Prepare mapping for the new page.*/
662 newpage
->index
= page
->index
;
663 newpage
->mapping
= page
->mapping
;
664 if (PageSwapBacked(page
))
665 SetPageSwapBacked(newpage
);
667 mapping
= page_mapping(page
);
669 rc
= migrate_page(mapping
, newpage
, page
, mode
);
670 else if (mapping
->a_ops
->migratepage
)
672 * Most pages have a mapping and most filesystems provide a
673 * migratepage callback. Anonymous pages are part of swap
674 * space which also has its own migratepage callback. This
675 * is the most common path for page migration.
677 rc
= mapping
->a_ops
->migratepage(mapping
,
678 newpage
, page
, mode
);
680 rc
= fallback_migrate_page(mapping
, newpage
, page
, mode
);
682 if (rc
!= MIGRATEPAGE_SUCCESS
) {
683 newpage
->mapping
= NULL
;
686 remove_migration_ptes(page
, newpage
);
687 page
->mapping
= NULL
;
690 unlock_page(newpage
);
695 static int __unmap_and_move(struct page
*page
, struct page
*newpage
,
696 int force
, bool offlining
, enum migrate_mode mode
)
699 int remap_swapcache
= 1;
700 struct mem_cgroup
*mem
;
701 struct anon_vma
*anon_vma
= NULL
;
703 if (!trylock_page(page
)) {
704 if (!force
|| mode
== MIGRATE_ASYNC
)
708 * It's not safe for direct compaction to call lock_page.
709 * For example, during page readahead pages are added locked
710 * to the LRU. Later, when the IO completes the pages are
711 * marked uptodate and unlocked. However, the queueing
712 * could be merging multiple pages for one bio (e.g.
713 * mpage_readpages). If an allocation happens for the
714 * second or third page, the process can end up locking
715 * the same page twice and deadlocking. Rather than
716 * trying to be clever about what pages can be locked,
717 * avoid the use of lock_page for direct compaction
720 if (current
->flags
& PF_MEMALLOC
)
727 * Only memory hotplug's offline_pages() caller has locked out KSM,
728 * and can safely migrate a KSM page. The other cases have skipped
729 * PageKsm along with PageReserved - but it is only now when we have
730 * the page lock that we can be certain it will not go KSM beneath us
731 * (KSM will not upgrade a page from PageAnon to PageKsm when it sees
732 * its pagecount raised, but only here do we take the page lock which
735 if (PageKsm(page
) && !offlining
) {
740 /* charge against new page */
741 mem_cgroup_prepare_migration(page
, newpage
, &mem
);
743 if (PageWriteback(page
)) {
745 * Only in the case of a full syncronous migration is it
746 * necessary to wait for PageWriteback. In the async case,
747 * the retry loop is too short and in the sync-light case,
748 * the overhead of stalling is too much
750 if (mode
!= MIGRATE_SYNC
) {
756 wait_on_page_writeback(page
);
759 * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
760 * we cannot notice that anon_vma is freed while we migrates a page.
761 * This get_anon_vma() delays freeing anon_vma pointer until the end
762 * of migration. File cache pages are no problem because of page_lock()
763 * File Caches may use write_page() or lock_page() in migration, then,
764 * just care Anon page here.
766 if (PageAnon(page
)) {
768 * Only page_lock_anon_vma() understands the subtleties of
769 * getting a hold on an anon_vma from outside one of its mms.
771 anon_vma
= page_get_anon_vma(page
);
776 } else if (PageSwapCache(page
)) {
778 * We cannot be sure that the anon_vma of an unmapped
779 * swapcache page is safe to use because we don't
780 * know in advance if the VMA that this page belonged
781 * to still exists. If the VMA and others sharing the
782 * data have been freed, then the anon_vma could
783 * already be invalid.
785 * To avoid this possibility, swapcache pages get
786 * migrated but are not remapped when migration
795 if (unlikely(balloon_page_movable(page
))) {
797 * A ballooned page does not need any special attention from
798 * physical to virtual reverse mapping procedures.
799 * Skip any attempt to unmap PTEs or to remap swap cache,
800 * in order to avoid burning cycles at rmap level, and perform
801 * the page migration right away (proteced by page lock).
803 rc
= balloon_page_migrate(newpage
, page
, mode
);
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
);
843 mem_cgroup_end_migration(mem
, page
, newpage
,
844 (rc
== MIGRATEPAGE_SUCCESS
||
845 rc
== MIGRATEPAGE_BALLOON_SUCCESS
));
853 * Obtain the lock on page, remove all ptes and migrate the page
854 * to the newly allocated page in newpage.
856 static int unmap_and_move(new_page_t get_new_page
, unsigned long private,
857 struct page
*page
, int force
, bool offlining
,
858 enum migrate_mode mode
)
862 struct page
*newpage
= get_new_page(page
, private, &result
);
867 if (page_count(page
) == 1) {
868 /* page was freed from under us. So we are done. */
872 if (unlikely(PageTransHuge(page
)))
873 if (unlikely(split_huge_page(page
)))
876 rc
= __unmap_and_move(page
, newpage
, force
, offlining
, mode
);
878 if (unlikely(rc
== MIGRATEPAGE_BALLOON_SUCCESS
)) {
880 * A ballooned page has been migrated already.
881 * Now, it's the time to wrap-up counters,
882 * handle the page back to Buddy and return.
884 dec_zone_page_state(page
, NR_ISOLATED_ANON
+
885 page_is_file_cache(page
));
886 balloon_page_free(page
);
887 return MIGRATEPAGE_SUCCESS
;
892 * A page that has been migrated has all references
893 * removed and will be freed. A page that has not been
894 * migrated will have kepts its references and be
897 list_del(&page
->lru
);
898 dec_zone_page_state(page
, NR_ISOLATED_ANON
+
899 page_is_file_cache(page
));
900 putback_lru_page(page
);
903 * Move the new page to the LRU. If migration was not successful
904 * then this will free the page.
906 putback_lru_page(newpage
);
911 *result
= page_to_nid(newpage
);
917 * Counterpart of unmap_and_move_page() for hugepage migration.
919 * This function doesn't wait the completion of hugepage I/O
920 * because there is no race between I/O and migration for hugepage.
921 * Note that currently hugepage I/O occurs only in direct I/O
922 * where no lock is held and PG_writeback is irrelevant,
923 * and writeback status of all subpages are counted in the reference
924 * count of the head page (i.e. if all subpages of a 2MB hugepage are
925 * under direct I/O, the reference of the head page is 512 and a bit more.)
926 * This means that when we try to migrate hugepage whose subpages are
927 * doing direct I/O, some references remain after try_to_unmap() and
928 * hugepage migration fails without data corruption.
930 * There is also no race when direct I/O is issued on the page under migration,
931 * because then pte is replaced with migration swap entry and direct I/O code
932 * will wait in the page fault for migration to complete.
934 static int unmap_and_move_huge_page(new_page_t get_new_page
,
935 unsigned long private, struct page
*hpage
,
936 int force
, bool offlining
,
937 enum migrate_mode mode
)
941 struct page
*new_hpage
= get_new_page(hpage
, private, &result
);
942 struct anon_vma
*anon_vma
= NULL
;
949 if (!trylock_page(hpage
)) {
950 if (!force
|| mode
!= MIGRATE_SYNC
)
956 anon_vma
= page_get_anon_vma(hpage
);
958 try_to_unmap(hpage
, TTU_MIGRATION
|TTU_IGNORE_MLOCK
|TTU_IGNORE_ACCESS
);
960 if (!page_mapped(hpage
))
961 rc
= move_to_new_page(new_hpage
, hpage
, 1, mode
);
964 remove_migration_ptes(hpage
, hpage
);
967 put_anon_vma(anon_vma
);
970 hugetlb_cgroup_migrate(hpage
, new_hpage
);
979 *result
= page_to_nid(new_hpage
);
987 * The function takes one list of pages to migrate and a function
988 * that determines from the page to be migrated and the private data
989 * the target of the move and allocates the page.
991 * The function returns after 10 attempts or if no pages
992 * are movable anymore because to has become empty
993 * or no retryable pages exist anymore.
994 * Caller should call putback_lru_pages to return pages to the LRU
995 * or free list only if ret != 0.
997 * Return: Number of pages not migrated or error code.
999 int migrate_pages(struct list_head
*from
,
1000 new_page_t get_new_page
, unsigned long private, bool offlining
,
1001 enum migrate_mode mode
)
1008 int swapwrite
= current
->flags
& PF_SWAPWRITE
;
1012 current
->flags
|= PF_SWAPWRITE
;
1014 for(pass
= 0; pass
< 10 && retry
; pass
++) {
1017 list_for_each_entry_safe(page
, page2
, from
, lru
) {
1020 rc
= unmap_and_move(get_new_page
, private,
1021 page
, pass
> 2, offlining
,
1030 case MIGRATEPAGE_SUCCESS
:
1033 /* Permanent failure */
1039 rc
= nr_failed
+ retry
;
1042 current
->flags
&= ~PF_SWAPWRITE
;
1047 int migrate_huge_page(struct page
*hpage
, new_page_t get_new_page
,
1048 unsigned long private, bool offlining
,
1049 enum migrate_mode mode
)
1053 for (pass
= 0; pass
< 10; pass
++) {
1054 rc
= unmap_and_move_huge_page(get_new_page
,
1055 private, hpage
, pass
> 2, offlining
,
1064 case MIGRATEPAGE_SUCCESS
:
1077 * Move a list of individual pages
1079 struct page_to_node
{
1086 static struct page
*new_page_node(struct page
*p
, unsigned long private,
1089 struct page_to_node
*pm
= (struct page_to_node
*)private;
1091 while (pm
->node
!= MAX_NUMNODES
&& pm
->page
!= p
)
1094 if (pm
->node
== MAX_NUMNODES
)
1097 *result
= &pm
->status
;
1099 return alloc_pages_exact_node(pm
->node
,
1100 GFP_HIGHUSER_MOVABLE
| GFP_THISNODE
, 0);
1104 * Move a set of pages as indicated in the pm array. The addr
1105 * field must be set to the virtual address of the page to be moved
1106 * and the node number must contain a valid target node.
1107 * The pm array ends with node = MAX_NUMNODES.
1109 static int do_move_page_to_node_array(struct mm_struct
*mm
,
1110 struct page_to_node
*pm
,
1114 struct page_to_node
*pp
;
1115 LIST_HEAD(pagelist
);
1117 down_read(&mm
->mmap_sem
);
1120 * Build a list of pages to migrate
1122 for (pp
= pm
; pp
->node
!= MAX_NUMNODES
; pp
++) {
1123 struct vm_area_struct
*vma
;
1127 vma
= find_vma(mm
, pp
->addr
);
1128 if (!vma
|| pp
->addr
< vma
->vm_start
|| !vma_migratable(vma
))
1131 page
= follow_page(vma
, pp
->addr
, FOLL_GET
|FOLL_SPLIT
);
1133 err
= PTR_ERR(page
);
1141 /* Use PageReserved to check for zero page */
1142 if (PageReserved(page
) || PageKsm(page
))
1146 err
= page_to_nid(page
);
1148 if (err
== pp
->node
)
1150 * Node already in the right place
1155 if (page_mapcount(page
) > 1 &&
1159 err
= isolate_lru_page(page
);
1161 list_add_tail(&page
->lru
, &pagelist
);
1162 inc_zone_page_state(page
, NR_ISOLATED_ANON
+
1163 page_is_file_cache(page
));
1167 * Either remove the duplicate refcount from
1168 * isolate_lru_page() or drop the page ref if it was
1177 if (!list_empty(&pagelist
)) {
1178 err
= migrate_pages(&pagelist
, new_page_node
,
1179 (unsigned long)pm
, 0, MIGRATE_SYNC
);
1181 putback_lru_pages(&pagelist
);
1184 up_read(&mm
->mmap_sem
);
1189 * Migrate an array of page address onto an array of nodes and fill
1190 * the corresponding array of status.
1192 static int do_pages_move(struct mm_struct
*mm
, nodemask_t task_nodes
,
1193 unsigned long nr_pages
,
1194 const void __user
* __user
*pages
,
1195 const int __user
*nodes
,
1196 int __user
*status
, int flags
)
1198 struct page_to_node
*pm
;
1199 unsigned long chunk_nr_pages
;
1200 unsigned long chunk_start
;
1204 pm
= (struct page_to_node
*)__get_free_page(GFP_KERNEL
);
1211 * Store a chunk of page_to_node array in a page,
1212 * but keep the last one as a marker
1214 chunk_nr_pages
= (PAGE_SIZE
/ sizeof(struct page_to_node
)) - 1;
1216 for (chunk_start
= 0;
1217 chunk_start
< nr_pages
;
1218 chunk_start
+= chunk_nr_pages
) {
1221 if (chunk_start
+ chunk_nr_pages
> nr_pages
)
1222 chunk_nr_pages
= nr_pages
- chunk_start
;
1224 /* fill the chunk pm with addrs and nodes from user-space */
1225 for (j
= 0; j
< chunk_nr_pages
; j
++) {
1226 const void __user
*p
;
1230 if (get_user(p
, pages
+ j
+ chunk_start
))
1232 pm
[j
].addr
= (unsigned long) p
;
1234 if (get_user(node
, nodes
+ j
+ chunk_start
))
1238 if (node
< 0 || node
>= MAX_NUMNODES
)
1241 if (!node_state(node
, N_MEMORY
))
1245 if (!node_isset(node
, task_nodes
))
1251 /* End marker for this chunk */
1252 pm
[chunk_nr_pages
].node
= MAX_NUMNODES
;
1254 /* Migrate this chunk */
1255 err
= do_move_page_to_node_array(mm
, pm
,
1256 flags
& MPOL_MF_MOVE_ALL
);
1260 /* Return status information */
1261 for (j
= 0; j
< chunk_nr_pages
; j
++)
1262 if (put_user(pm
[j
].status
, status
+ j
+ chunk_start
)) {
1270 free_page((unsigned long)pm
);
1276 * Determine the nodes of an array of pages and store it in an array of status.
1278 static void do_pages_stat_array(struct mm_struct
*mm
, unsigned long nr_pages
,
1279 const void __user
**pages
, int *status
)
1283 down_read(&mm
->mmap_sem
);
1285 for (i
= 0; i
< nr_pages
; i
++) {
1286 unsigned long addr
= (unsigned long)(*pages
);
1287 struct vm_area_struct
*vma
;
1291 vma
= find_vma(mm
, addr
);
1292 if (!vma
|| addr
< vma
->vm_start
)
1295 page
= follow_page(vma
, addr
, 0);
1297 err
= PTR_ERR(page
);
1302 /* Use PageReserved to check for zero page */
1303 if (!page
|| PageReserved(page
) || PageKsm(page
))
1306 err
= page_to_nid(page
);
1314 up_read(&mm
->mmap_sem
);
1318 * Determine the nodes of a user array of pages and store it in
1319 * a user array of status.
1321 static int do_pages_stat(struct mm_struct
*mm
, unsigned long nr_pages
,
1322 const void __user
* __user
*pages
,
1325 #define DO_PAGES_STAT_CHUNK_NR 16
1326 const void __user
*chunk_pages
[DO_PAGES_STAT_CHUNK_NR
];
1327 int chunk_status
[DO_PAGES_STAT_CHUNK_NR
];
1330 unsigned long chunk_nr
;
1332 chunk_nr
= nr_pages
;
1333 if (chunk_nr
> DO_PAGES_STAT_CHUNK_NR
)
1334 chunk_nr
= DO_PAGES_STAT_CHUNK_NR
;
1336 if (copy_from_user(chunk_pages
, pages
, chunk_nr
* sizeof(*chunk_pages
)))
1339 do_pages_stat_array(mm
, chunk_nr
, chunk_pages
, chunk_status
);
1341 if (copy_to_user(status
, chunk_status
, chunk_nr
* sizeof(*status
)))
1346 nr_pages
-= chunk_nr
;
1348 return nr_pages
? -EFAULT
: 0;
1352 * Move a list of pages in the address space of the currently executing
1355 SYSCALL_DEFINE6(move_pages
, pid_t
, pid
, unsigned long, nr_pages
,
1356 const void __user
* __user
*, pages
,
1357 const int __user
*, nodes
,
1358 int __user
*, status
, int, flags
)
1360 const struct cred
*cred
= current_cred(), *tcred
;
1361 struct task_struct
*task
;
1362 struct mm_struct
*mm
;
1364 nodemask_t task_nodes
;
1367 if (flags
& ~(MPOL_MF_MOVE
|MPOL_MF_MOVE_ALL
))
1370 if ((flags
& MPOL_MF_MOVE_ALL
) && !capable(CAP_SYS_NICE
))
1373 /* Find the mm_struct */
1375 task
= pid
? find_task_by_vpid(pid
) : current
;
1380 get_task_struct(task
);
1383 * Check if this process has the right to modify the specified
1384 * process. The right exists if the process has administrative
1385 * capabilities, superuser privileges or the same
1386 * userid as the target process.
1388 tcred
= __task_cred(task
);
1389 if (!uid_eq(cred
->euid
, tcred
->suid
) && !uid_eq(cred
->euid
, tcred
->uid
) &&
1390 !uid_eq(cred
->uid
, tcred
->suid
) && !uid_eq(cred
->uid
, tcred
->uid
) &&
1391 !capable(CAP_SYS_NICE
)) {
1398 err
= security_task_movememory(task
);
1402 task_nodes
= cpuset_mems_allowed(task
);
1403 mm
= get_task_mm(task
);
1404 put_task_struct(task
);
1410 err
= do_pages_move(mm
, task_nodes
, nr_pages
, pages
,
1411 nodes
, status
, flags
);
1413 err
= do_pages_stat(mm
, nr_pages
, pages
, status
);
1419 put_task_struct(task
);
1424 * Call migration functions in the vma_ops that may prepare
1425 * memory in a vm for migration. migration functions may perform
1426 * the migration for vmas that do not have an underlying page struct.
1428 int migrate_vmas(struct mm_struct
*mm
, const nodemask_t
*to
,
1429 const nodemask_t
*from
, unsigned long flags
)
1431 struct vm_area_struct
*vma
;
1434 for (vma
= mm
->mmap
; vma
&& !err
; vma
= vma
->vm_next
) {
1435 if (vma
->vm_ops
&& vma
->vm_ops
->migrate
) {
1436 err
= vma
->vm_ops
->migrate(vma
, to
, from
, flags
);