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/backing-dev.h>
34 #include <linux/memcontrol.h>
35 #include <linux/syscalls.h>
36 #include <linux/hugetlb.h>
37 #include <linux/hugetlb_cgroup.h>
38 #include <linux/gfp.h>
39 #include <linux/balloon_compaction.h>
41 #include <asm/tlbflush.h>
43 #define CREATE_TRACE_POINTS
44 #include <trace/events/migrate.h>
49 * migrate_prep() needs to be called before we start compiling a list of pages
50 * to be migrated using isolate_lru_page(). If scheduling work on other CPUs is
51 * undesirable, use migrate_prep_local()
53 int migrate_prep(void)
56 * Clear the LRU lists so pages can be isolated.
57 * Note that pages may be moved off the LRU after we have
58 * drained them. Those pages will fail to migrate like other
59 * pages that may be busy.
66 /* Do the necessary work of migrate_prep but not if it involves other CPUs */
67 int migrate_prep_local(void)
75 * Add isolated pages on the list back to the LRU under page lock
76 * to avoid leaking evictable pages back onto unevictable list.
78 void putback_lru_pages(struct list_head
*l
)
83 list_for_each_entry_safe(page
, page2
, l
, lru
) {
85 dec_zone_page_state(page
, NR_ISOLATED_ANON
+
86 page_is_file_cache(page
));
87 putback_lru_page(page
);
92 * Put previously isolated pages back onto the appropriate lists
93 * from where they were once taken off for compaction/migration.
95 * This function shall be used instead of putback_lru_pages(),
96 * whenever the isolated pageset has been built by isolate_migratepages_range()
98 void putback_movable_pages(struct list_head
*l
)
103 list_for_each_entry_safe(page
, page2
, l
, lru
) {
104 list_del(&page
->lru
);
105 dec_zone_page_state(page
, NR_ISOLATED_ANON
+
106 page_is_file_cache(page
));
107 if (unlikely(isolated_balloon_page(page
)))
108 balloon_page_putback(page
);
110 putback_lru_page(page
);
115 * Restore a potential migration pte to a working pte entry
117 static int remove_migration_pte(struct page
*new, struct vm_area_struct
*vma
,
118 unsigned long addr
, void *old
)
120 struct mm_struct
*mm
= vma
->vm_mm
;
126 if (unlikely(PageHuge(new))) {
127 ptep
= huge_pte_offset(mm
, addr
);
130 ptl
= &mm
->page_table_lock
;
132 pmd
= mm_find_pmd(mm
, addr
);
135 if (pmd_trans_huge(*pmd
))
138 ptep
= pte_offset_map(pmd
, addr
);
141 * Peek to check is_swap_pte() before taking ptlock? No, we
142 * can race mremap's move_ptes(), which skips anon_vma lock.
145 ptl
= pte_lockptr(mm
, pmd
);
150 if (!is_swap_pte(pte
))
153 entry
= pte_to_swp_entry(pte
);
155 if (!is_migration_entry(entry
) ||
156 migration_entry_to_page(entry
) != old
)
160 pte
= pte_mkold(mk_pte(new, vma
->vm_page_prot
));
161 if (is_write_migration_entry(entry
))
162 pte
= pte_mkwrite(pte
);
163 #ifdef CONFIG_HUGETLB_PAGE
165 pte
= pte_mkhuge(pte
);
166 pte
= arch_make_huge_pte(pte
, vma
, new, 0);
169 flush_dcache_page(new);
170 set_pte_at(mm
, addr
, ptep
, pte
);
174 hugepage_add_anon_rmap(new, vma
, addr
);
177 } else if (PageAnon(new))
178 page_add_anon_rmap(new, vma
, addr
);
180 page_add_file_rmap(new);
182 /* No need to invalidate - it was non-present before */
183 update_mmu_cache(vma
, addr
, ptep
);
185 pte_unmap_unlock(ptep
, ptl
);
191 * Get rid of all migration entries and replace them by
192 * references to the indicated page.
194 static void remove_migration_ptes(struct page
*old
, struct page
*new)
196 rmap_walk(new, remove_migration_pte
, old
);
200 * Something used the pte of a page under migration. We need to
201 * get to the page and wait until migration is finished.
202 * When we return from this function the fault will be retried.
204 static void __migration_entry_wait(struct mm_struct
*mm
, pte_t
*ptep
,
213 if (!is_swap_pte(pte
))
216 entry
= pte_to_swp_entry(pte
);
217 if (!is_migration_entry(entry
))
220 page
= migration_entry_to_page(entry
);
223 * Once radix-tree replacement of page migration started, page_count
224 * *must* be zero. And, we don't want to call wait_on_page_locked()
225 * against a page without get_page().
226 * So, we use get_page_unless_zero(), here. Even failed, page fault
229 if (!get_page_unless_zero(page
))
231 pte_unmap_unlock(ptep
, ptl
);
232 wait_on_page_locked(page
);
236 pte_unmap_unlock(ptep
, ptl
);
239 void migration_entry_wait(struct mm_struct
*mm
, pmd_t
*pmd
,
240 unsigned long address
)
242 spinlock_t
*ptl
= pte_lockptr(mm
, pmd
);
243 pte_t
*ptep
= pte_offset_map(pmd
, address
);
244 __migration_entry_wait(mm
, ptep
, ptl
);
247 void migration_entry_wait_huge(struct mm_struct
*mm
, pte_t
*pte
)
249 spinlock_t
*ptl
= &(mm
)->page_table_lock
;
250 __migration_entry_wait(mm
, pte
, ptl
);
254 /* Returns true if all buffers are successfully locked */
255 static bool buffer_migrate_lock_buffers(struct buffer_head
*head
,
256 enum migrate_mode mode
)
258 struct buffer_head
*bh
= head
;
260 /* Simple case, sync compaction */
261 if (mode
!= MIGRATE_ASYNC
) {
265 bh
= bh
->b_this_page
;
267 } while (bh
!= head
);
272 /* async case, we cannot block on lock_buffer so use trylock_buffer */
275 if (!trylock_buffer(bh
)) {
277 * We failed to lock the buffer and cannot stall in
278 * async migration. Release the taken locks
280 struct buffer_head
*failed_bh
= bh
;
283 while (bh
!= failed_bh
) {
286 bh
= bh
->b_this_page
;
291 bh
= bh
->b_this_page
;
292 } while (bh
!= head
);
296 static inline bool buffer_migrate_lock_buffers(struct buffer_head
*head
,
297 enum migrate_mode mode
)
301 #endif /* CONFIG_BLOCK */
304 * Replace the page in the mapping.
306 * The number of remaining references must be:
307 * 1 for anonymous pages without a mapping
308 * 2 for pages with a mapping
309 * 3 for pages with a mapping and PagePrivate/PagePrivate2 set.
311 static int migrate_page_move_mapping(struct address_space
*mapping
,
312 struct page
*newpage
, struct page
*page
,
313 struct buffer_head
*head
, enum migrate_mode mode
)
315 struct zone
*oldzone
, *newzone
;
317 int expected_count
= 0;
321 /* Anonymous page without mapping */
322 if (page_count(page
) != 1)
324 return MIGRATEPAGE_SUCCESS
;
327 oldzone
= page_zone(page
);
328 newzone
= page_zone(newpage
);
330 spin_lock_irq(&mapping
->tree_lock
);
332 pslot
= radix_tree_lookup_slot(&mapping
->page_tree
,
335 expected_count
= 2 + page_has_private(page
);
336 if (page_count(page
) != expected_count
||
337 radix_tree_deref_slot_protected(pslot
, &mapping
->tree_lock
) != page
) {
338 spin_unlock_irq(&mapping
->tree_lock
);
342 if (!page_freeze_refs(page
, expected_count
)) {
343 spin_unlock_irq(&mapping
->tree_lock
);
348 * In the async migration case of moving a page with buffers, lock the
349 * buffers using trylock before the mapping is moved. If the mapping
350 * was moved, we later failed to lock the buffers and could not move
351 * the mapping back due to an elevated page count, we would have to
352 * block waiting on other references to be dropped.
354 if (mode
== MIGRATE_ASYNC
&& head
&&
355 !buffer_migrate_lock_buffers(head
, mode
)) {
356 page_unfreeze_refs(page
, expected_count
);
357 spin_unlock_irq(&mapping
->tree_lock
);
362 * Now we know that no one else is looking at the page.
364 get_page(newpage
); /* add cache reference */
365 if (PageSwapCache(page
)) {
366 SetPageSwapCache(newpage
);
367 set_page_private(newpage
, page_private(page
));
370 /* Move dirty while page refs frozen and newpage not yet exposed */
371 dirty
= PageDirty(page
);
373 ClearPageDirty(page
);
374 SetPageDirty(newpage
);
377 radix_tree_replace_slot(pslot
, newpage
);
380 * Drop cache reference from old page by unfreezing
381 * to one less reference.
382 * We know this isn't the last reference.
384 page_unfreeze_refs(page
, expected_count
- 1);
386 spin_unlock(&mapping
->tree_lock
);
387 /* Leave irq disabled to prevent preemption while updating stats */
390 * If moved to a different zone then also account
391 * the page for that zone. Other VM counters will be
392 * taken care of when we establish references to the
393 * new page and drop references to the old page.
395 * Note that anonymous pages are accounted for
396 * via NR_FILE_PAGES and NR_ANON_PAGES if they
397 * are mapped to swap space.
399 if (newzone
!= oldzone
) {
400 __dec_zone_state(oldzone
, NR_FILE_PAGES
);
401 __inc_zone_state(newzone
, NR_FILE_PAGES
);
402 if (PageSwapBacked(page
) && !PageSwapCache(page
)) {
403 __dec_zone_state(oldzone
, NR_SHMEM
);
404 __inc_zone_state(newzone
, NR_SHMEM
);
406 if (dirty
&& mapping_cap_account_dirty(mapping
)) {
407 __dec_zone_state(oldzone
, NR_FILE_DIRTY
);
408 __inc_zone_state(newzone
, NR_FILE_DIRTY
);
413 return MIGRATEPAGE_SUCCESS
;
417 * The expected number of remaining references is the same as that
418 * of migrate_page_move_mapping().
420 int migrate_huge_page_move_mapping(struct address_space
*mapping
,
421 struct page
*newpage
, struct page
*page
)
427 if (page_count(page
) != 1)
429 return MIGRATEPAGE_SUCCESS
;
432 spin_lock_irq(&mapping
->tree_lock
);
434 pslot
= radix_tree_lookup_slot(&mapping
->page_tree
,
437 expected_count
= 2 + page_has_private(page
);
438 if (page_count(page
) != expected_count
||
439 radix_tree_deref_slot_protected(pslot
, &mapping
->tree_lock
) != page
) {
440 spin_unlock_irq(&mapping
->tree_lock
);
444 if (!page_freeze_refs(page
, expected_count
)) {
445 spin_unlock_irq(&mapping
->tree_lock
);
451 radix_tree_replace_slot(pslot
, newpage
);
453 page_unfreeze_refs(page
, expected_count
- 1);
455 spin_unlock_irq(&mapping
->tree_lock
);
456 return MIGRATEPAGE_SUCCESS
;
460 * Copy the page to its new location
462 void migrate_page_copy(struct page
*newpage
, struct page
*page
)
464 if (PageHuge(page
) || PageTransHuge(page
))
465 copy_huge_page(newpage
, page
);
467 copy_highpage(newpage
, page
);
470 SetPageError(newpage
);
471 if (PageReferenced(page
))
472 SetPageReferenced(newpage
);
473 if (PageUptodate(page
))
474 SetPageUptodate(newpage
);
475 if (TestClearPageActive(page
)) {
476 VM_BUG_ON(PageUnevictable(page
));
477 SetPageActive(newpage
);
478 } else if (TestClearPageUnevictable(page
))
479 SetPageUnevictable(newpage
);
480 if (PageChecked(page
))
481 SetPageChecked(newpage
);
482 if (PageMappedToDisk(page
))
483 SetPageMappedToDisk(newpage
);
485 /* Move dirty on pages not done by migrate_page_move_mapping() */
487 SetPageDirty(newpage
);
489 mlock_migrate_page(newpage
, page
);
490 ksm_migrate_page(newpage
, page
);
492 * Please do not reorder this without considering how mm/ksm.c's
493 * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache().
495 ClearPageSwapCache(page
);
496 ClearPagePrivate(page
);
497 set_page_private(page
, 0);
500 * If any waiters have accumulated on the new page then
503 if (PageWriteback(newpage
))
504 end_page_writeback(newpage
);
507 /************************************************************
508 * Migration functions
509 ***********************************************************/
511 /* Always fail migration. Used for mappings that are not movable */
512 int fail_migrate_page(struct address_space
*mapping
,
513 struct page
*newpage
, struct page
*page
)
517 EXPORT_SYMBOL(fail_migrate_page
);
520 * Common logic to directly migrate a single page suitable for
521 * pages that do not use PagePrivate/PagePrivate2.
523 * Pages are locked upon entry and exit.
525 int migrate_page(struct address_space
*mapping
,
526 struct page
*newpage
, struct page
*page
,
527 enum migrate_mode mode
)
531 BUG_ON(PageWriteback(page
)); /* Writeback must be complete */
533 rc
= migrate_page_move_mapping(mapping
, newpage
, page
, NULL
, mode
);
535 if (rc
!= MIGRATEPAGE_SUCCESS
)
538 migrate_page_copy(newpage
, page
);
539 return MIGRATEPAGE_SUCCESS
;
541 EXPORT_SYMBOL(migrate_page
);
545 * Migration function for pages with buffers. This function can only be used
546 * if the underlying filesystem guarantees that no other references to "page"
549 int buffer_migrate_page(struct address_space
*mapping
,
550 struct page
*newpage
, struct page
*page
, enum migrate_mode mode
)
552 struct buffer_head
*bh
, *head
;
555 if (!page_has_buffers(page
))
556 return migrate_page(mapping
, newpage
, page
, mode
);
558 head
= page_buffers(page
);
560 rc
= migrate_page_move_mapping(mapping
, newpage
, page
, head
, mode
);
562 if (rc
!= MIGRATEPAGE_SUCCESS
)
566 * In the async case, migrate_page_move_mapping locked the buffers
567 * with an IRQ-safe spinlock held. In the sync case, the buffers
568 * need to be locked now
570 if (mode
!= MIGRATE_ASYNC
)
571 BUG_ON(!buffer_migrate_lock_buffers(head
, mode
));
573 ClearPagePrivate(page
);
574 set_page_private(newpage
, page_private(page
));
575 set_page_private(page
, 0);
581 set_bh_page(bh
, newpage
, bh_offset(bh
));
582 bh
= bh
->b_this_page
;
584 } while (bh
!= head
);
586 SetPagePrivate(newpage
);
588 migrate_page_copy(newpage
, page
);
594 bh
= bh
->b_this_page
;
596 } while (bh
!= head
);
598 return MIGRATEPAGE_SUCCESS
;
600 EXPORT_SYMBOL(buffer_migrate_page
);
604 * Writeback a page to clean the dirty state
606 static int writeout(struct address_space
*mapping
, struct page
*page
)
608 struct writeback_control wbc
= {
609 .sync_mode
= WB_SYNC_NONE
,
612 .range_end
= LLONG_MAX
,
617 if (!mapping
->a_ops
->writepage
)
618 /* No write method for the address space */
621 if (!clear_page_dirty_for_io(page
))
622 /* Someone else already triggered a write */
626 * A dirty page may imply that the underlying filesystem has
627 * the page on some queue. So the page must be clean for
628 * migration. Writeout may mean we loose the lock and the
629 * page state is no longer what we checked for earlier.
630 * At this point we know that the migration attempt cannot
633 remove_migration_ptes(page
, page
);
635 rc
= mapping
->a_ops
->writepage(page
, &wbc
);
637 if (rc
!= AOP_WRITEPAGE_ACTIVATE
)
638 /* unlocked. Relock */
641 return (rc
< 0) ? -EIO
: -EAGAIN
;
645 * Default handling if a filesystem does not provide a migration function.
647 static int fallback_migrate_page(struct address_space
*mapping
,
648 struct page
*newpage
, struct page
*page
, enum migrate_mode mode
)
650 if (PageDirty(page
)) {
651 /* Only writeback pages in full synchronous migration */
652 if (mode
!= MIGRATE_SYNC
)
654 return writeout(mapping
, page
);
658 * Buffers may be managed in a filesystem specific way.
659 * We must have no buffers or drop them.
661 if (page_has_private(page
) &&
662 !try_to_release_page(page
, GFP_KERNEL
))
665 return migrate_page(mapping
, newpage
, page
, mode
);
669 * Move a page to a newly allocated page
670 * The page is locked and all ptes have been successfully removed.
672 * The new page will have replaced the old page if this function
677 * MIGRATEPAGE_SUCCESS - success
679 static int move_to_new_page(struct page
*newpage
, struct page
*page
,
680 int remap_swapcache
, enum migrate_mode mode
)
682 struct address_space
*mapping
;
686 * Block others from accessing the page when we get around to
687 * establishing additional references. We are the only one
688 * holding a reference to the new page at this point.
690 if (!trylock_page(newpage
))
693 /* Prepare mapping for the new page.*/
694 newpage
->index
= page
->index
;
695 newpage
->mapping
= page
->mapping
;
696 if (PageSwapBacked(page
))
697 SetPageSwapBacked(newpage
);
699 mapping
= page_mapping(page
);
701 rc
= migrate_page(mapping
, newpage
, page
, mode
);
702 else if (mapping
->a_ops
->migratepage
)
704 * Most pages have a mapping and most filesystems provide a
705 * migratepage callback. Anonymous pages are part of swap
706 * space which also has its own migratepage callback. This
707 * is the most common path for page migration.
709 rc
= mapping
->a_ops
->migratepage(mapping
,
710 newpage
, page
, mode
);
712 rc
= fallback_migrate_page(mapping
, newpage
, page
, mode
);
714 if (rc
!= MIGRATEPAGE_SUCCESS
) {
715 newpage
->mapping
= NULL
;
718 remove_migration_ptes(page
, newpage
);
719 page
->mapping
= NULL
;
722 unlock_page(newpage
);
727 static int __unmap_and_move(struct page
*page
, struct page
*newpage
,
728 int force
, enum migrate_mode mode
)
731 int remap_swapcache
= 1;
732 struct mem_cgroup
*mem
;
733 struct anon_vma
*anon_vma
= NULL
;
735 if (!trylock_page(page
)) {
736 if (!force
|| mode
== MIGRATE_ASYNC
)
740 * It's not safe for direct compaction to call lock_page.
741 * For example, during page readahead pages are added locked
742 * to the LRU. Later, when the IO completes the pages are
743 * marked uptodate and unlocked. However, the queueing
744 * could be merging multiple pages for one bio (e.g.
745 * mpage_readpages). If an allocation happens for the
746 * second or third page, the process can end up locking
747 * the same page twice and deadlocking. Rather than
748 * trying to be clever about what pages can be locked,
749 * avoid the use of lock_page for direct compaction
752 if (current
->flags
& PF_MEMALLOC
)
758 /* charge against new page */
759 mem_cgroup_prepare_migration(page
, newpage
, &mem
);
761 if (PageWriteback(page
)) {
763 * Only in the case of a full synchronous migration is it
764 * necessary to wait for PageWriteback. In the async case,
765 * the retry loop is too short and in the sync-light case,
766 * the overhead of stalling is too much
768 if (mode
!= MIGRATE_SYNC
) {
774 wait_on_page_writeback(page
);
777 * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
778 * we cannot notice that anon_vma is freed while we migrates a page.
779 * This get_anon_vma() delays freeing anon_vma pointer until the end
780 * of migration. File cache pages are no problem because of page_lock()
781 * File Caches may use write_page() or lock_page() in migration, then,
782 * just care Anon page here.
784 if (PageAnon(page
) && !PageKsm(page
)) {
786 * Only page_lock_anon_vma_read() understands the subtleties of
787 * getting a hold on an anon_vma from outside one of its mms.
789 anon_vma
= page_get_anon_vma(page
);
794 } else if (PageSwapCache(page
)) {
796 * We cannot be sure that the anon_vma of an unmapped
797 * swapcache page is safe to use because we don't
798 * know in advance if the VMA that this page belonged
799 * to still exists. If the VMA and others sharing the
800 * data have been freed, then the anon_vma could
801 * already be invalid.
803 * To avoid this possibility, swapcache pages get
804 * migrated but are not remapped when migration
813 if (unlikely(balloon_page_movable(page
))) {
815 * A ballooned page does not need any special attention from
816 * physical to virtual reverse mapping procedures.
817 * Skip any attempt to unmap PTEs or to remap swap cache,
818 * in order to avoid burning cycles at rmap level, and perform
819 * the page migration right away (proteced by page lock).
821 rc
= balloon_page_migrate(newpage
, page
, mode
);
826 * Corner case handling:
827 * 1. When a new swap-cache page is read into, it is added to the LRU
828 * and treated as swapcache but it has no rmap yet.
829 * Calling try_to_unmap() against a page->mapping==NULL page will
830 * trigger a BUG. So handle it here.
831 * 2. An orphaned page (see truncate_complete_page) might have
832 * fs-private metadata. The page can be picked up due to memory
833 * offlining. Everywhere else except page reclaim, the page is
834 * invisible to the vm, so the page can not be migrated. So try to
835 * free the metadata, so the page can be freed.
837 if (!page
->mapping
) {
838 VM_BUG_ON(PageAnon(page
));
839 if (page_has_private(page
)) {
840 try_to_free_buffers(page
);
846 /* Establish migration ptes or remove ptes */
847 try_to_unmap(page
, TTU_MIGRATION
|TTU_IGNORE_MLOCK
|TTU_IGNORE_ACCESS
);
850 if (!page_mapped(page
))
851 rc
= move_to_new_page(newpage
, page
, remap_swapcache
, mode
);
853 if (rc
&& remap_swapcache
)
854 remove_migration_ptes(page
, page
);
856 /* Drop an anon_vma reference if we took one */
858 put_anon_vma(anon_vma
);
861 mem_cgroup_end_migration(mem
, page
, newpage
,
862 (rc
== MIGRATEPAGE_SUCCESS
||
863 rc
== MIGRATEPAGE_BALLOON_SUCCESS
));
870 * Obtain the lock on page, remove all ptes and migrate the page
871 * to the newly allocated page in newpage.
873 static int unmap_and_move(new_page_t get_new_page
, unsigned long private,
874 struct page
*page
, int force
, enum migrate_mode mode
)
878 struct page
*newpage
= get_new_page(page
, private, &result
);
883 if (page_count(page
) == 1) {
884 /* page was freed from under us. So we are done. */
888 if (unlikely(PageTransHuge(page
)))
889 if (unlikely(split_huge_page(page
)))
892 rc
= __unmap_and_move(page
, newpage
, force
, mode
);
894 if (unlikely(rc
== MIGRATEPAGE_BALLOON_SUCCESS
)) {
896 * A ballooned page has been migrated already.
897 * Now, it's the time to wrap-up counters,
898 * handle the page back to Buddy and return.
900 dec_zone_page_state(page
, NR_ISOLATED_ANON
+
901 page_is_file_cache(page
));
902 balloon_page_free(page
);
903 return MIGRATEPAGE_SUCCESS
;
908 * A page that has been migrated has all references
909 * removed and will be freed. A page that has not been
910 * migrated will have kepts its references and be
913 list_del(&page
->lru
);
914 dec_zone_page_state(page
, NR_ISOLATED_ANON
+
915 page_is_file_cache(page
));
916 putback_lru_page(page
);
919 * Move the new page to the LRU. If migration was not successful
920 * then this will free the page.
922 putback_lru_page(newpage
);
927 *result
= page_to_nid(newpage
);
933 * Counterpart of unmap_and_move_page() for hugepage migration.
935 * This function doesn't wait the completion of hugepage I/O
936 * because there is no race between I/O and migration for hugepage.
937 * Note that currently hugepage I/O occurs only in direct I/O
938 * where no lock is held and PG_writeback is irrelevant,
939 * and writeback status of all subpages are counted in the reference
940 * count of the head page (i.e. if all subpages of a 2MB hugepage are
941 * under direct I/O, the reference of the head page is 512 and a bit more.)
942 * This means that when we try to migrate hugepage whose subpages are
943 * doing direct I/O, some references remain after try_to_unmap() and
944 * hugepage migration fails without data corruption.
946 * There is also no race when direct I/O is issued on the page under migration,
947 * because then pte is replaced with migration swap entry and direct I/O code
948 * will wait in the page fault for migration to complete.
950 static int unmap_and_move_huge_page(new_page_t get_new_page
,
951 unsigned long private, struct page
*hpage
,
952 int force
, enum migrate_mode mode
)
956 struct page
*new_hpage
= get_new_page(hpage
, private, &result
);
957 struct anon_vma
*anon_vma
= NULL
;
964 if (!trylock_page(hpage
)) {
965 if (!force
|| mode
!= MIGRATE_SYNC
)
971 anon_vma
= page_get_anon_vma(hpage
);
973 try_to_unmap(hpage
, TTU_MIGRATION
|TTU_IGNORE_MLOCK
|TTU_IGNORE_ACCESS
);
975 if (!page_mapped(hpage
))
976 rc
= move_to_new_page(new_hpage
, hpage
, 1, mode
);
979 remove_migration_ptes(hpage
, hpage
);
982 put_anon_vma(anon_vma
);
985 hugetlb_cgroup_migrate(hpage
, new_hpage
);
994 *result
= page_to_nid(new_hpage
);
1000 * migrate_pages - migrate the pages specified in a list, to the free pages
1001 * supplied as the target for the page migration
1003 * @from: The list of pages to be migrated.
1004 * @get_new_page: The function used to allocate free pages to be used
1005 * as the target of the page migration.
1006 * @private: Private data to be passed on to get_new_page()
1007 * @mode: The migration mode that specifies the constraints for
1008 * page migration, if any.
1009 * @reason: The reason for page migration.
1011 * The function returns after 10 attempts or if no pages are movable any more
1012 * because the list has become empty or no retryable pages exist any more.
1013 * The caller should call putback_lru_pages() to return pages to the LRU
1014 * or free list only if ret != 0.
1016 * Returns the number of pages that were not migrated, or an error code.
1018 int migrate_pages(struct list_head
*from
, new_page_t get_new_page
,
1019 unsigned long private, enum migrate_mode mode
, int reason
)
1023 int nr_succeeded
= 0;
1027 int swapwrite
= current
->flags
& PF_SWAPWRITE
;
1031 current
->flags
|= PF_SWAPWRITE
;
1033 for(pass
= 0; pass
< 10 && retry
; pass
++) {
1036 list_for_each_entry_safe(page
, page2
, from
, lru
) {
1039 rc
= unmap_and_move(get_new_page
, private,
1040 page
, pass
> 2, mode
);
1048 case MIGRATEPAGE_SUCCESS
:
1052 /* Permanent failure */
1058 rc
= nr_failed
+ retry
;
1061 count_vm_events(PGMIGRATE_SUCCESS
, nr_succeeded
);
1063 count_vm_events(PGMIGRATE_FAIL
, nr_failed
);
1064 trace_mm_migrate_pages(nr_succeeded
, nr_failed
, mode
, reason
);
1067 current
->flags
&= ~PF_SWAPWRITE
;
1072 int migrate_huge_page(struct page
*hpage
, new_page_t get_new_page
,
1073 unsigned long private, enum migrate_mode mode
)
1077 for (pass
= 0; pass
< 10; pass
++) {
1078 rc
= unmap_and_move_huge_page(get_new_page
, private,
1079 hpage
, pass
> 2, mode
);
1087 case MIGRATEPAGE_SUCCESS
:
1100 * Move a list of individual pages
1102 struct page_to_node
{
1109 static struct page
*new_page_node(struct page
*p
, unsigned long private,
1112 struct page_to_node
*pm
= (struct page_to_node
*)private;
1114 while (pm
->node
!= MAX_NUMNODES
&& pm
->page
!= p
)
1117 if (pm
->node
== MAX_NUMNODES
)
1120 *result
= &pm
->status
;
1122 return alloc_pages_exact_node(pm
->node
,
1123 GFP_HIGHUSER_MOVABLE
| GFP_THISNODE
, 0);
1127 * Move a set of pages as indicated in the pm array. The addr
1128 * field must be set to the virtual address of the page to be moved
1129 * and the node number must contain a valid target node.
1130 * The pm array ends with node = MAX_NUMNODES.
1132 static int do_move_page_to_node_array(struct mm_struct
*mm
,
1133 struct page_to_node
*pm
,
1137 struct page_to_node
*pp
;
1138 LIST_HEAD(pagelist
);
1140 down_read(&mm
->mmap_sem
);
1143 * Build a list of pages to migrate
1145 for (pp
= pm
; pp
->node
!= MAX_NUMNODES
; pp
++) {
1146 struct vm_area_struct
*vma
;
1150 vma
= find_vma(mm
, pp
->addr
);
1151 if (!vma
|| pp
->addr
< vma
->vm_start
|| !vma_migratable(vma
))
1154 page
= follow_page(vma
, pp
->addr
, FOLL_GET
|FOLL_SPLIT
);
1156 err
= PTR_ERR(page
);
1164 /* Use PageReserved to check for zero page */
1165 if (PageReserved(page
))
1169 err
= page_to_nid(page
);
1171 if (err
== pp
->node
)
1173 * Node already in the right place
1178 if (page_mapcount(page
) > 1 &&
1182 err
= isolate_lru_page(page
);
1184 list_add_tail(&page
->lru
, &pagelist
);
1185 inc_zone_page_state(page
, NR_ISOLATED_ANON
+
1186 page_is_file_cache(page
));
1190 * Either remove the duplicate refcount from
1191 * isolate_lru_page() or drop the page ref if it was
1200 if (!list_empty(&pagelist
)) {
1201 err
= migrate_pages(&pagelist
, new_page_node
,
1202 (unsigned long)pm
, MIGRATE_SYNC
, MR_SYSCALL
);
1204 putback_lru_pages(&pagelist
);
1207 up_read(&mm
->mmap_sem
);
1212 * Migrate an array of page address onto an array of nodes and fill
1213 * the corresponding array of status.
1215 static int do_pages_move(struct mm_struct
*mm
, nodemask_t task_nodes
,
1216 unsigned long nr_pages
,
1217 const void __user
* __user
*pages
,
1218 const int __user
*nodes
,
1219 int __user
*status
, int flags
)
1221 struct page_to_node
*pm
;
1222 unsigned long chunk_nr_pages
;
1223 unsigned long chunk_start
;
1227 pm
= (struct page_to_node
*)__get_free_page(GFP_KERNEL
);
1234 * Store a chunk of page_to_node array in a page,
1235 * but keep the last one as a marker
1237 chunk_nr_pages
= (PAGE_SIZE
/ sizeof(struct page_to_node
)) - 1;
1239 for (chunk_start
= 0;
1240 chunk_start
< nr_pages
;
1241 chunk_start
+= chunk_nr_pages
) {
1244 if (chunk_start
+ chunk_nr_pages
> nr_pages
)
1245 chunk_nr_pages
= nr_pages
- chunk_start
;
1247 /* fill the chunk pm with addrs and nodes from user-space */
1248 for (j
= 0; j
< chunk_nr_pages
; j
++) {
1249 const void __user
*p
;
1253 if (get_user(p
, pages
+ j
+ chunk_start
))
1255 pm
[j
].addr
= (unsigned long) p
;
1257 if (get_user(node
, nodes
+ j
+ chunk_start
))
1261 if (node
< 0 || node
>= MAX_NUMNODES
)
1264 if (!node_state(node
, N_MEMORY
))
1268 if (!node_isset(node
, task_nodes
))
1274 /* End marker for this chunk */
1275 pm
[chunk_nr_pages
].node
= MAX_NUMNODES
;
1277 /* Migrate this chunk */
1278 err
= do_move_page_to_node_array(mm
, pm
,
1279 flags
& MPOL_MF_MOVE_ALL
);
1283 /* Return status information */
1284 for (j
= 0; j
< chunk_nr_pages
; j
++)
1285 if (put_user(pm
[j
].status
, status
+ j
+ chunk_start
)) {
1293 free_page((unsigned long)pm
);
1299 * Determine the nodes of an array of pages and store it in an array of status.
1301 static void do_pages_stat_array(struct mm_struct
*mm
, unsigned long nr_pages
,
1302 const void __user
**pages
, int *status
)
1306 down_read(&mm
->mmap_sem
);
1308 for (i
= 0; i
< nr_pages
; i
++) {
1309 unsigned long addr
= (unsigned long)(*pages
);
1310 struct vm_area_struct
*vma
;
1314 vma
= find_vma(mm
, addr
);
1315 if (!vma
|| addr
< vma
->vm_start
)
1318 page
= follow_page(vma
, addr
, 0);
1320 err
= PTR_ERR(page
);
1325 /* Use PageReserved to check for zero page */
1326 if (!page
|| PageReserved(page
))
1329 err
= page_to_nid(page
);
1337 up_read(&mm
->mmap_sem
);
1341 * Determine the nodes of a user array of pages and store it in
1342 * a user array of status.
1344 static int do_pages_stat(struct mm_struct
*mm
, unsigned long nr_pages
,
1345 const void __user
* __user
*pages
,
1348 #define DO_PAGES_STAT_CHUNK_NR 16
1349 const void __user
*chunk_pages
[DO_PAGES_STAT_CHUNK_NR
];
1350 int chunk_status
[DO_PAGES_STAT_CHUNK_NR
];
1353 unsigned long chunk_nr
;
1355 chunk_nr
= nr_pages
;
1356 if (chunk_nr
> DO_PAGES_STAT_CHUNK_NR
)
1357 chunk_nr
= DO_PAGES_STAT_CHUNK_NR
;
1359 if (copy_from_user(chunk_pages
, pages
, chunk_nr
* sizeof(*chunk_pages
)))
1362 do_pages_stat_array(mm
, chunk_nr
, chunk_pages
, chunk_status
);
1364 if (copy_to_user(status
, chunk_status
, chunk_nr
* sizeof(*status
)))
1369 nr_pages
-= chunk_nr
;
1371 return nr_pages
? -EFAULT
: 0;
1375 * Move a list of pages in the address space of the currently executing
1378 SYSCALL_DEFINE6(move_pages
, pid_t
, pid
, unsigned long, nr_pages
,
1379 const void __user
* __user
*, pages
,
1380 const int __user
*, nodes
,
1381 int __user
*, status
, int, flags
)
1383 const struct cred
*cred
= current_cred(), *tcred
;
1384 struct task_struct
*task
;
1385 struct mm_struct
*mm
;
1387 nodemask_t task_nodes
;
1390 if (flags
& ~(MPOL_MF_MOVE
|MPOL_MF_MOVE_ALL
))
1393 if ((flags
& MPOL_MF_MOVE_ALL
) && !capable(CAP_SYS_NICE
))
1396 /* Find the mm_struct */
1398 task
= pid
? find_task_by_vpid(pid
) : current
;
1403 get_task_struct(task
);
1406 * Check if this process has the right to modify the specified
1407 * process. The right exists if the process has administrative
1408 * capabilities, superuser privileges or the same
1409 * userid as the target process.
1411 tcred
= __task_cred(task
);
1412 if (!uid_eq(cred
->euid
, tcred
->suid
) && !uid_eq(cred
->euid
, tcred
->uid
) &&
1413 !uid_eq(cred
->uid
, tcred
->suid
) && !uid_eq(cred
->uid
, tcred
->uid
) &&
1414 !capable(CAP_SYS_NICE
)) {
1421 err
= security_task_movememory(task
);
1425 task_nodes
= cpuset_mems_allowed(task
);
1426 mm
= get_task_mm(task
);
1427 put_task_struct(task
);
1433 err
= do_pages_move(mm
, task_nodes
, nr_pages
, pages
,
1434 nodes
, status
, flags
);
1436 err
= do_pages_stat(mm
, nr_pages
, pages
, status
);
1442 put_task_struct(task
);
1447 * Call migration functions in the vma_ops that may prepare
1448 * memory in a vm for migration. migration functions may perform
1449 * the migration for vmas that do not have an underlying page struct.
1451 int migrate_vmas(struct mm_struct
*mm
, const nodemask_t
*to
,
1452 const nodemask_t
*from
, unsigned long flags
)
1454 struct vm_area_struct
*vma
;
1457 for (vma
= mm
->mmap
; vma
&& !err
; vma
= vma
->vm_next
) {
1458 if (vma
->vm_ops
&& vma
->vm_ops
->migrate
) {
1459 err
= vma
->vm_ops
->migrate(vma
, to
, from
, flags
);
1467 #ifdef CONFIG_NUMA_BALANCING
1469 * Returns true if this is a safe migration target node for misplaced NUMA
1470 * pages. Currently it only checks the watermarks which crude
1472 static bool migrate_balanced_pgdat(struct pglist_data
*pgdat
,
1473 unsigned long nr_migrate_pages
)
1476 for (z
= pgdat
->nr_zones
- 1; z
>= 0; z
--) {
1477 struct zone
*zone
= pgdat
->node_zones
+ z
;
1479 if (!populated_zone(zone
))
1482 if (zone
->all_unreclaimable
)
1485 /* Avoid waking kswapd by allocating pages_to_migrate pages. */
1486 if (!zone_watermark_ok(zone
, 0,
1487 high_wmark_pages(zone
) +
1496 static struct page
*alloc_misplaced_dst_page(struct page
*page
,
1500 int nid
= (int) data
;
1501 struct page
*newpage
;
1503 newpage
= alloc_pages_exact_node(nid
,
1504 (GFP_HIGHUSER_MOVABLE
| GFP_THISNODE
|
1505 __GFP_NOMEMALLOC
| __GFP_NORETRY
|
1509 page_nid_xchg_last(newpage
, page_nid_last(page
));
1515 * page migration rate limiting control.
1516 * Do not migrate more than @pages_to_migrate in a @migrate_interval_millisecs
1517 * window of time. Default here says do not migrate more than 1280M per second.
1518 * If a node is rate-limited then PTE NUMA updates are also rate-limited. However
1519 * as it is faults that reset the window, pte updates will happen unconditionally
1520 * if there has not been a fault since @pteupdate_interval_millisecs after the
1521 * throttle window closed.
1523 static unsigned int migrate_interval_millisecs __read_mostly
= 100;
1524 static unsigned int pteupdate_interval_millisecs __read_mostly
= 1000;
1525 static unsigned int ratelimit_pages __read_mostly
= 128 << (20 - PAGE_SHIFT
);
1527 /* Returns true if NUMA migration is currently rate limited */
1528 bool migrate_ratelimited(int node
)
1530 pg_data_t
*pgdat
= NODE_DATA(node
);
1532 if (time_after(jiffies
, pgdat
->numabalancing_migrate_next_window
+
1533 msecs_to_jiffies(pteupdate_interval_millisecs
)))
1536 if (pgdat
->numabalancing_migrate_nr_pages
< ratelimit_pages
)
1542 /* Returns true if the node is migrate rate-limited after the update */
1543 bool numamigrate_update_ratelimit(pg_data_t
*pgdat
, unsigned long nr_pages
)
1545 bool rate_limited
= false;
1548 * Rate-limit the amount of data that is being migrated to a node.
1549 * Optimal placement is no good if the memory bus is saturated and
1550 * all the time is being spent migrating!
1552 spin_lock(&pgdat
->numabalancing_migrate_lock
);
1553 if (time_after(jiffies
, pgdat
->numabalancing_migrate_next_window
)) {
1554 pgdat
->numabalancing_migrate_nr_pages
= 0;
1555 pgdat
->numabalancing_migrate_next_window
= jiffies
+
1556 msecs_to_jiffies(migrate_interval_millisecs
);
1558 if (pgdat
->numabalancing_migrate_nr_pages
> ratelimit_pages
)
1559 rate_limited
= true;
1561 pgdat
->numabalancing_migrate_nr_pages
+= nr_pages
;
1562 spin_unlock(&pgdat
->numabalancing_migrate_lock
);
1564 return rate_limited
;
1567 int numamigrate_isolate_page(pg_data_t
*pgdat
, struct page
*page
)
1571 VM_BUG_ON(compound_order(page
) && !PageTransHuge(page
));
1573 /* Avoid migrating to a node that is nearly full */
1574 if (!migrate_balanced_pgdat(pgdat
, 1UL << compound_order(page
)))
1577 if (isolate_lru_page(page
))
1581 * migrate_misplaced_transhuge_page() skips page migration's usual
1582 * check on page_count(), so we must do it here, now that the page
1583 * has been isolated: a GUP pin, or any other pin, prevents migration.
1584 * The expected page count is 3: 1 for page's mapcount and 1 for the
1585 * caller's pin and 1 for the reference taken by isolate_lru_page().
1587 if (PageTransHuge(page
) && page_count(page
) != 3) {
1588 putback_lru_page(page
);
1592 page_lru
= page_is_file_cache(page
);
1593 mod_zone_page_state(page_zone(page
), NR_ISOLATED_ANON
+ page_lru
,
1594 hpage_nr_pages(page
));
1597 * Isolating the page has taken another reference, so the
1598 * caller's reference can be safely dropped without the page
1599 * disappearing underneath us during migration.
1606 * Attempt to migrate a misplaced page to the specified destination
1607 * node. Caller is expected to have an elevated reference count on
1608 * the page that will be dropped by this function before returning.
1610 int migrate_misplaced_page(struct page
*page
, int node
)
1612 pg_data_t
*pgdat
= NODE_DATA(node
);
1615 LIST_HEAD(migratepages
);
1618 * Don't migrate pages that are mapped in multiple processes.
1619 * TODO: Handle false sharing detection instead of this hammer
1621 if (page_mapcount(page
) != 1)
1625 * Rate-limit the amount of data that is being migrated to a node.
1626 * Optimal placement is no good if the memory bus is saturated and
1627 * all the time is being spent migrating!
1629 if (numamigrate_update_ratelimit(pgdat
, 1))
1632 isolated
= numamigrate_isolate_page(pgdat
, page
);
1636 list_add(&page
->lru
, &migratepages
);
1637 nr_remaining
= migrate_pages(&migratepages
, alloc_misplaced_dst_page
,
1638 node
, MIGRATE_ASYNC
, MR_NUMA_MISPLACED
);
1640 putback_lru_pages(&migratepages
);
1643 count_vm_numa_event(NUMA_PAGE_MIGRATE
);
1644 BUG_ON(!list_empty(&migratepages
));
1651 #endif /* CONFIG_NUMA_BALANCING */
1653 #if defined(CONFIG_NUMA_BALANCING) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
1655 * Migrates a THP to a given target node. page must be locked and is unlocked
1658 int migrate_misplaced_transhuge_page(struct mm_struct
*mm
,
1659 struct vm_area_struct
*vma
,
1660 pmd_t
*pmd
, pmd_t entry
,
1661 unsigned long address
,
1662 struct page
*page
, int node
)
1664 unsigned long haddr
= address
& HPAGE_PMD_MASK
;
1665 pg_data_t
*pgdat
= NODE_DATA(node
);
1667 struct page
*new_page
= NULL
;
1668 struct mem_cgroup
*memcg
= NULL
;
1669 int page_lru
= page_is_file_cache(page
);
1672 * Don't migrate pages that are mapped in multiple processes.
1673 * TODO: Handle false sharing detection instead of this hammer
1675 if (page_mapcount(page
) != 1)
1679 * Rate-limit the amount of data that is being migrated to a node.
1680 * Optimal placement is no good if the memory bus is saturated and
1681 * all the time is being spent migrating!
1683 if (numamigrate_update_ratelimit(pgdat
, HPAGE_PMD_NR
))
1686 new_page
= alloc_pages_node(node
,
1687 (GFP_TRANSHUGE
| GFP_THISNODE
) & ~__GFP_WAIT
, HPAGE_PMD_ORDER
);
1691 page_nid_xchg_last(new_page
, page_nid_last(page
));
1693 isolated
= numamigrate_isolate_page(pgdat
, page
);
1699 /* Prepare a page as a migration target */
1700 __set_page_locked(new_page
);
1701 SetPageSwapBacked(new_page
);
1703 /* anon mapping, we can simply copy page->mapping to the new page: */
1704 new_page
->mapping
= page
->mapping
;
1705 new_page
->index
= page
->index
;
1706 migrate_page_copy(new_page
, page
);
1707 WARN_ON(PageLRU(new_page
));
1709 /* Recheck the target PMD */
1710 spin_lock(&mm
->page_table_lock
);
1711 if (unlikely(!pmd_same(*pmd
, entry
))) {
1712 spin_unlock(&mm
->page_table_lock
);
1714 /* Reverse changes made by migrate_page_copy() */
1715 if (TestClearPageActive(new_page
))
1716 SetPageActive(page
);
1717 if (TestClearPageUnevictable(new_page
))
1718 SetPageUnevictable(page
);
1719 mlock_migrate_page(page
, new_page
);
1721 unlock_page(new_page
);
1722 put_page(new_page
); /* Free it */
1724 /* Retake the callers reference and putback on LRU */
1726 putback_lru_page(page
);
1727 mod_zone_page_state(page_zone(page
),
1728 NR_ISOLATED_ANON
+ page_lru
, -HPAGE_PMD_NR
);
1734 * Traditional migration needs to prepare the memcg charge
1735 * transaction early to prevent the old page from being
1736 * uncharged when installing migration entries. Here we can
1737 * save the potential rollback and start the charge transfer
1738 * only when migration is already known to end successfully.
1740 mem_cgroup_prepare_migration(page
, new_page
, &memcg
);
1742 entry
= mk_pmd(new_page
, vma
->vm_page_prot
);
1743 entry
= pmd_mknonnuma(entry
);
1744 entry
= maybe_pmd_mkwrite(pmd_mkdirty(entry
), vma
);
1745 entry
= pmd_mkhuge(entry
);
1747 pmdp_clear_flush(vma
, haddr
, pmd
);
1748 set_pmd_at(mm
, haddr
, pmd
, entry
);
1749 page_add_new_anon_rmap(new_page
, vma
, haddr
);
1750 update_mmu_cache_pmd(vma
, address
, &entry
);
1751 page_remove_rmap(page
);
1753 * Finish the charge transaction under the page table lock to
1754 * prevent split_huge_page() from dividing up the charge
1755 * before it's fully transferred to the new page.
1757 mem_cgroup_end_migration(memcg
, page
, new_page
, true);
1758 spin_unlock(&mm
->page_table_lock
);
1760 unlock_page(new_page
);
1762 put_page(page
); /* Drop the rmap reference */
1763 put_page(page
); /* Drop the LRU isolation reference */
1765 count_vm_events(PGMIGRATE_SUCCESS
, HPAGE_PMD_NR
);
1766 count_vm_numa_events(NUMA_PAGE_MIGRATE
, HPAGE_PMD_NR
);
1768 mod_zone_page_state(page_zone(page
),
1769 NR_ISOLATED_ANON
+ page_lru
,
1774 count_vm_events(PGMIGRATE_FAIL
, HPAGE_PMD_NR
);
1776 entry
= pmd_mknonnuma(entry
);
1777 set_pmd_at(mm
, haddr
, pmd
, entry
);
1778 update_mmu_cache_pmd(vma
, address
, &entry
);
1785 #endif /* CONFIG_NUMA_BALANCING */
1787 #endif /* CONFIG_NUMA */