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>
42 #define CREATE_TRACE_POINTS
43 #include <trace/events/migrate.h>
48 * migrate_prep() needs to be called before we start compiling a list of pages
49 * to be migrated using isolate_lru_page(). If scheduling work on other CPUs is
50 * undesirable, use migrate_prep_local()
52 int migrate_prep(void)
55 * Clear the LRU lists so pages can be isolated.
56 * Note that pages may be moved off the LRU after we have
57 * drained them. Those pages will fail to migrate like other
58 * pages that may be busy.
65 /* Do the necessary work of migrate_prep but not if it involves other CPUs */
66 int migrate_prep_local(void)
74 * Add isolated pages on the list back to the LRU under page lock
75 * to avoid leaking evictable pages back onto unevictable list.
77 void putback_lru_pages(struct list_head
*l
)
82 list_for_each_entry_safe(page
, page2
, l
, lru
) {
84 dec_zone_page_state(page
, NR_ISOLATED_ANON
+
85 page_is_file_cache(page
));
86 putback_lru_page(page
);
91 * Put previously isolated pages back onto the appropriate lists
92 * from where they were once taken off for compaction/migration.
94 * This function shall be used instead of putback_lru_pages(),
95 * whenever the isolated pageset has been built by isolate_migratepages_range()
97 void putback_movable_pages(struct list_head
*l
)
102 list_for_each_entry_safe(page
, page2
, l
, lru
) {
103 if (unlikely(PageHuge(page
))) {
104 putback_active_hugepage(page
);
107 list_del(&page
->lru
);
108 dec_zone_page_state(page
, NR_ISOLATED_ANON
+
109 page_is_file_cache(page
));
110 if (unlikely(isolated_balloon_page(page
)))
111 balloon_page_putback(page
);
113 putback_lru_page(page
);
118 * Restore a potential migration pte to a working pte entry
120 static int remove_migration_pte(struct page
*new, struct vm_area_struct
*vma
,
121 unsigned long addr
, void *old
)
123 struct mm_struct
*mm
= vma
->vm_mm
;
129 if (unlikely(PageHuge(new))) {
130 ptep
= huge_pte_offset(mm
, addr
);
133 ptl
= huge_pte_lockptr(hstate_vma(vma
), mm
, ptep
);
135 pmd
= mm_find_pmd(mm
, addr
);
138 if (pmd_trans_huge(*pmd
))
141 ptep
= pte_offset_map(pmd
, addr
);
144 * Peek to check is_swap_pte() before taking ptlock? No, we
145 * can race mremap's move_ptes(), which skips anon_vma lock.
148 ptl
= pte_lockptr(mm
, pmd
);
153 if (!is_swap_pte(pte
))
156 entry
= pte_to_swp_entry(pte
);
158 if (!is_migration_entry(entry
) ||
159 migration_entry_to_page(entry
) != old
)
163 pte
= pte_mkold(mk_pte(new, vma
->vm_page_prot
));
164 if (pte_swp_soft_dirty(*ptep
))
165 pte
= pte_mksoft_dirty(pte
);
166 if (is_write_migration_entry(entry
))
167 pte
= pte_mkwrite(pte
);
168 #ifdef CONFIG_HUGETLB_PAGE
170 pte
= pte_mkhuge(pte
);
171 pte
= arch_make_huge_pte(pte
, vma
, new, 0);
174 flush_dcache_page(new);
175 set_pte_at(mm
, addr
, ptep
, pte
);
179 hugepage_add_anon_rmap(new, vma
, addr
);
182 } else if (PageAnon(new))
183 page_add_anon_rmap(new, vma
, addr
);
185 page_add_file_rmap(new);
187 /* No need to invalidate - it was non-present before */
188 update_mmu_cache(vma
, addr
, ptep
);
190 pte_unmap_unlock(ptep
, ptl
);
196 * Get rid of all migration entries and replace them by
197 * references to the indicated page.
199 static void remove_migration_ptes(struct page
*old
, struct page
*new)
201 rmap_walk(new, remove_migration_pte
, old
);
205 * Something used the pte of a page under migration. We need to
206 * get to the page and wait until migration is finished.
207 * When we return from this function the fault will be retried.
209 static void __migration_entry_wait(struct mm_struct
*mm
, pte_t
*ptep
,
218 if (!is_swap_pte(pte
))
221 entry
= pte_to_swp_entry(pte
);
222 if (!is_migration_entry(entry
))
225 page
= migration_entry_to_page(entry
);
228 * Once radix-tree replacement of page migration started, page_count
229 * *must* be zero. And, we don't want to call wait_on_page_locked()
230 * against a page without get_page().
231 * So, we use get_page_unless_zero(), here. Even failed, page fault
234 if (!get_page_unless_zero(page
))
236 pte_unmap_unlock(ptep
, ptl
);
237 wait_on_page_locked(page
);
241 pte_unmap_unlock(ptep
, ptl
);
244 void migration_entry_wait(struct mm_struct
*mm
, pmd_t
*pmd
,
245 unsigned long address
)
247 spinlock_t
*ptl
= pte_lockptr(mm
, pmd
);
248 pte_t
*ptep
= pte_offset_map(pmd
, address
);
249 __migration_entry_wait(mm
, ptep
, ptl
);
252 void migration_entry_wait_huge(struct vm_area_struct
*vma
,
253 struct mm_struct
*mm
, pte_t
*pte
)
255 spinlock_t
*ptl
= huge_pte_lockptr(hstate_vma(vma
), mm
, pte
);
256 __migration_entry_wait(mm
, pte
, ptl
);
260 /* Returns true if all buffers are successfully locked */
261 static bool buffer_migrate_lock_buffers(struct buffer_head
*head
,
262 enum migrate_mode mode
)
264 struct buffer_head
*bh
= head
;
266 /* Simple case, sync compaction */
267 if (mode
!= MIGRATE_ASYNC
) {
271 bh
= bh
->b_this_page
;
273 } while (bh
!= head
);
278 /* async case, we cannot block on lock_buffer so use trylock_buffer */
281 if (!trylock_buffer(bh
)) {
283 * We failed to lock the buffer and cannot stall in
284 * async migration. Release the taken locks
286 struct buffer_head
*failed_bh
= bh
;
289 while (bh
!= failed_bh
) {
292 bh
= bh
->b_this_page
;
297 bh
= bh
->b_this_page
;
298 } while (bh
!= head
);
302 static inline bool buffer_migrate_lock_buffers(struct buffer_head
*head
,
303 enum migrate_mode mode
)
307 #endif /* CONFIG_BLOCK */
310 * Replace the page in the mapping.
312 * The number of remaining references must be:
313 * 1 for anonymous pages without a mapping
314 * 2 for pages with a mapping
315 * 3 for pages with a mapping and PagePrivate/PagePrivate2 set.
317 int migrate_page_move_mapping(struct address_space
*mapping
,
318 struct page
*newpage
, struct page
*page
,
319 struct buffer_head
*head
, enum migrate_mode mode
)
321 int expected_count
= 0;
325 /* Anonymous page without mapping */
326 if (page_count(page
) != 1)
328 return MIGRATEPAGE_SUCCESS
;
331 spin_lock_irq(&mapping
->tree_lock
);
333 pslot
= radix_tree_lookup_slot(&mapping
->page_tree
,
336 expected_count
= 2 + page_has_private(page
);
337 if (page_count(page
) != expected_count
||
338 radix_tree_deref_slot_protected(pslot
, &mapping
->tree_lock
) != page
) {
339 spin_unlock_irq(&mapping
->tree_lock
);
343 if (!page_freeze_refs(page
, expected_count
)) {
344 spin_unlock_irq(&mapping
->tree_lock
);
349 * In the async migration case of moving a page with buffers, lock the
350 * buffers using trylock before the mapping is moved. If the mapping
351 * was moved, we later failed to lock the buffers and could not move
352 * the mapping back due to an elevated page count, we would have to
353 * block waiting on other references to be dropped.
355 if (mode
== MIGRATE_ASYNC
&& head
&&
356 !buffer_migrate_lock_buffers(head
, mode
)) {
357 page_unfreeze_refs(page
, expected_count
);
358 spin_unlock_irq(&mapping
->tree_lock
);
363 * Now we know that no one else is looking at the page.
365 get_page(newpage
); /* add cache reference */
366 if (PageSwapCache(page
)) {
367 SetPageSwapCache(newpage
);
368 set_page_private(newpage
, page_private(page
));
371 radix_tree_replace_slot(pslot
, newpage
);
374 * Drop cache reference from old page by unfreezing
375 * to one less reference.
376 * We know this isn't the last reference.
378 page_unfreeze_refs(page
, expected_count
- 1);
381 * If moved to a different zone then also account
382 * the page for that zone. Other VM counters will be
383 * taken care of when we establish references to the
384 * new page and drop references to the old page.
386 * Note that anonymous pages are accounted for
387 * via NR_FILE_PAGES and NR_ANON_PAGES if they
388 * are mapped to swap space.
390 __dec_zone_page_state(page
, NR_FILE_PAGES
);
391 __inc_zone_page_state(newpage
, NR_FILE_PAGES
);
392 if (!PageSwapCache(page
) && PageSwapBacked(page
)) {
393 __dec_zone_page_state(page
, NR_SHMEM
);
394 __inc_zone_page_state(newpage
, NR_SHMEM
);
396 spin_unlock_irq(&mapping
->tree_lock
);
398 return MIGRATEPAGE_SUCCESS
;
402 * The expected number of remaining references is the same as that
403 * of migrate_page_move_mapping().
405 int migrate_huge_page_move_mapping(struct address_space
*mapping
,
406 struct page
*newpage
, struct page
*page
)
412 if (page_count(page
) != 1)
414 return MIGRATEPAGE_SUCCESS
;
417 spin_lock_irq(&mapping
->tree_lock
);
419 pslot
= radix_tree_lookup_slot(&mapping
->page_tree
,
422 expected_count
= 2 + page_has_private(page
);
423 if (page_count(page
) != expected_count
||
424 radix_tree_deref_slot_protected(pslot
, &mapping
->tree_lock
) != page
) {
425 spin_unlock_irq(&mapping
->tree_lock
);
429 if (!page_freeze_refs(page
, expected_count
)) {
430 spin_unlock_irq(&mapping
->tree_lock
);
436 radix_tree_replace_slot(pslot
, newpage
);
438 page_unfreeze_refs(page
, expected_count
- 1);
440 spin_unlock_irq(&mapping
->tree_lock
);
441 return MIGRATEPAGE_SUCCESS
;
445 * Copy the page to its new location
447 void migrate_page_copy(struct page
*newpage
, struct page
*page
)
451 if (PageHuge(page
) || PageTransHuge(page
))
452 copy_huge_page(newpage
, page
);
454 copy_highpage(newpage
, page
);
457 SetPageError(newpage
);
458 if (PageReferenced(page
))
459 SetPageReferenced(newpage
);
460 if (PageUptodate(page
))
461 SetPageUptodate(newpage
);
462 if (TestClearPageActive(page
)) {
463 VM_BUG_ON(PageUnevictable(page
));
464 SetPageActive(newpage
);
465 } else if (TestClearPageUnevictable(page
))
466 SetPageUnevictable(newpage
);
467 if (PageChecked(page
))
468 SetPageChecked(newpage
);
469 if (PageMappedToDisk(page
))
470 SetPageMappedToDisk(newpage
);
472 if (PageDirty(page
)) {
473 clear_page_dirty_for_io(page
);
475 * Want to mark the page and the radix tree as dirty, and
476 * redo the accounting that clear_page_dirty_for_io undid,
477 * but we can't use set_page_dirty because that function
478 * is actually a signal that all of the page has become dirty.
479 * Whereas only part of our page may be dirty.
481 if (PageSwapBacked(page
))
482 SetPageDirty(newpage
);
484 __set_page_dirty_nobuffers(newpage
);
488 * Copy NUMA information to the new page, to prevent over-eager
489 * future migrations of this same page.
491 cpupid
= page_cpupid_xchg_last(page
, -1);
492 page_cpupid_xchg_last(newpage
, cpupid
);
494 mlock_migrate_page(newpage
, page
);
495 ksm_migrate_page(newpage
, page
);
497 * Please do not reorder this without considering how mm/ksm.c's
498 * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache().
500 ClearPageSwapCache(page
);
501 ClearPagePrivate(page
);
502 set_page_private(page
, 0);
505 * If any waiters have accumulated on the new page then
508 if (PageWriteback(newpage
))
509 end_page_writeback(newpage
);
512 /************************************************************
513 * Migration functions
514 ***********************************************************/
516 /* Always fail migration. Used for mappings that are not movable */
517 int fail_migrate_page(struct address_space
*mapping
,
518 struct page
*newpage
, struct page
*page
)
522 EXPORT_SYMBOL(fail_migrate_page
);
525 * Common logic to directly migrate a single page suitable for
526 * pages that do not use PagePrivate/PagePrivate2.
528 * Pages are locked upon entry and exit.
530 int migrate_page(struct address_space
*mapping
,
531 struct page
*newpage
, struct page
*page
,
532 enum migrate_mode mode
)
536 BUG_ON(PageWriteback(page
)); /* Writeback must be complete */
538 rc
= migrate_page_move_mapping(mapping
, newpage
, page
, NULL
, mode
);
540 if (rc
!= MIGRATEPAGE_SUCCESS
)
543 migrate_page_copy(newpage
, page
);
544 return MIGRATEPAGE_SUCCESS
;
546 EXPORT_SYMBOL(migrate_page
);
550 * Migration function for pages with buffers. This function can only be used
551 * if the underlying filesystem guarantees that no other references to "page"
554 int buffer_migrate_page(struct address_space
*mapping
,
555 struct page
*newpage
, struct page
*page
, enum migrate_mode mode
)
557 struct buffer_head
*bh
, *head
;
560 if (!page_has_buffers(page
))
561 return migrate_page(mapping
, newpage
, page
, mode
);
563 head
= page_buffers(page
);
565 rc
= migrate_page_move_mapping(mapping
, newpage
, page
, head
, mode
);
567 if (rc
!= MIGRATEPAGE_SUCCESS
)
571 * In the async case, migrate_page_move_mapping locked the buffers
572 * with an IRQ-safe spinlock held. In the sync case, the buffers
573 * need to be locked now
575 if (mode
!= MIGRATE_ASYNC
)
576 BUG_ON(!buffer_migrate_lock_buffers(head
, mode
));
578 ClearPagePrivate(page
);
579 set_page_private(newpage
, page_private(page
));
580 set_page_private(page
, 0);
586 set_bh_page(bh
, newpage
, bh_offset(bh
));
587 bh
= bh
->b_this_page
;
589 } while (bh
!= head
);
591 SetPagePrivate(newpage
);
593 migrate_page_copy(newpage
, page
);
599 bh
= bh
->b_this_page
;
601 } while (bh
!= head
);
603 return MIGRATEPAGE_SUCCESS
;
605 EXPORT_SYMBOL(buffer_migrate_page
);
609 * Writeback a page to clean the dirty state
611 static int writeout(struct address_space
*mapping
, struct page
*page
)
613 struct writeback_control wbc
= {
614 .sync_mode
= WB_SYNC_NONE
,
617 .range_end
= LLONG_MAX
,
622 if (!mapping
->a_ops
->writepage
)
623 /* No write method for the address space */
626 if (!clear_page_dirty_for_io(page
))
627 /* Someone else already triggered a write */
631 * A dirty page may imply that the underlying filesystem has
632 * the page on some queue. So the page must be clean for
633 * migration. Writeout may mean we loose the lock and the
634 * page state is no longer what we checked for earlier.
635 * At this point we know that the migration attempt cannot
638 remove_migration_ptes(page
, page
);
640 rc
= mapping
->a_ops
->writepage(page
, &wbc
);
642 if (rc
!= AOP_WRITEPAGE_ACTIVATE
)
643 /* unlocked. Relock */
646 return (rc
< 0) ? -EIO
: -EAGAIN
;
650 * Default handling if a filesystem does not provide a migration function.
652 static int fallback_migrate_page(struct address_space
*mapping
,
653 struct page
*newpage
, struct page
*page
, enum migrate_mode mode
)
655 if (PageDirty(page
)) {
656 /* Only writeback pages in full synchronous migration */
657 if (mode
!= MIGRATE_SYNC
)
659 return writeout(mapping
, page
);
663 * Buffers may be managed in a filesystem specific way.
664 * We must have no buffers or drop them.
666 if (page_has_private(page
) &&
667 !try_to_release_page(page
, GFP_KERNEL
))
670 return migrate_page(mapping
, newpage
, page
, mode
);
674 * Move a page to a newly allocated page
675 * The page is locked and all ptes have been successfully removed.
677 * The new page will have replaced the old page if this function
682 * MIGRATEPAGE_SUCCESS - success
684 static int move_to_new_page(struct page
*newpage
, struct page
*page
,
685 int remap_swapcache
, enum migrate_mode mode
)
687 struct address_space
*mapping
;
691 * Block others from accessing the page when we get around to
692 * establishing additional references. We are the only one
693 * holding a reference to the new page at this point.
695 if (!trylock_page(newpage
))
698 /* Prepare mapping for the new page.*/
699 newpage
->index
= page
->index
;
700 newpage
->mapping
= page
->mapping
;
701 if (PageSwapBacked(page
))
702 SetPageSwapBacked(newpage
);
704 mapping
= page_mapping(page
);
706 rc
= migrate_page(mapping
, newpage
, page
, mode
);
707 else if (mapping
->a_ops
->migratepage
)
709 * Most pages have a mapping and most filesystems provide a
710 * migratepage callback. Anonymous pages are part of swap
711 * space which also has its own migratepage callback. This
712 * is the most common path for page migration.
714 rc
= mapping
->a_ops
->migratepage(mapping
,
715 newpage
, page
, mode
);
717 rc
= fallback_migrate_page(mapping
, newpage
, page
, mode
);
719 if (rc
!= MIGRATEPAGE_SUCCESS
) {
720 newpage
->mapping
= NULL
;
723 remove_migration_ptes(page
, newpage
);
724 page
->mapping
= NULL
;
727 unlock_page(newpage
);
732 static int __unmap_and_move(struct page
*page
, struct page
*newpage
,
733 int force
, enum migrate_mode mode
)
736 int remap_swapcache
= 1;
737 struct mem_cgroup
*mem
;
738 struct anon_vma
*anon_vma
= NULL
;
740 if (!trylock_page(page
)) {
741 if (!force
|| mode
== MIGRATE_ASYNC
)
745 * It's not safe for direct compaction to call lock_page.
746 * For example, during page readahead pages are added locked
747 * to the LRU. Later, when the IO completes the pages are
748 * marked uptodate and unlocked. However, the queueing
749 * could be merging multiple pages for one bio (e.g.
750 * mpage_readpages). If an allocation happens for the
751 * second or third page, the process can end up locking
752 * the same page twice and deadlocking. Rather than
753 * trying to be clever about what pages can be locked,
754 * avoid the use of lock_page for direct compaction
757 if (current
->flags
& PF_MEMALLOC
)
763 /* charge against new page */
764 mem_cgroup_prepare_migration(page
, newpage
, &mem
);
766 if (PageWriteback(page
)) {
768 * Only in the case of a full synchronous migration is it
769 * necessary to wait for PageWriteback. In the async case,
770 * the retry loop is too short and in the sync-light case,
771 * the overhead of stalling is too much
773 if (mode
!= MIGRATE_SYNC
) {
779 wait_on_page_writeback(page
);
782 * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
783 * we cannot notice that anon_vma is freed while we migrates a page.
784 * This get_anon_vma() delays freeing anon_vma pointer until the end
785 * of migration. File cache pages are no problem because of page_lock()
786 * File Caches may use write_page() or lock_page() in migration, then,
787 * just care Anon page here.
789 if (PageAnon(page
) && !PageKsm(page
)) {
791 * Only page_lock_anon_vma_read() understands the subtleties of
792 * getting a hold on an anon_vma from outside one of its mms.
794 anon_vma
= page_get_anon_vma(page
);
799 } else if (PageSwapCache(page
)) {
801 * We cannot be sure that the anon_vma of an unmapped
802 * swapcache page is safe to use because we don't
803 * know in advance if the VMA that this page belonged
804 * to still exists. If the VMA and others sharing the
805 * data have been freed, then the anon_vma could
806 * already be invalid.
808 * To avoid this possibility, swapcache pages get
809 * migrated but are not remapped when migration
818 if (unlikely(balloon_page_movable(page
))) {
820 * A ballooned page does not need any special attention from
821 * physical to virtual reverse mapping procedures.
822 * Skip any attempt to unmap PTEs or to remap swap cache,
823 * in order to avoid burning cycles at rmap level, and perform
824 * the page migration right away (proteced by page lock).
826 rc
= balloon_page_migrate(newpage
, page
, mode
);
831 * Corner case handling:
832 * 1. When a new swap-cache page is read into, it is added to the LRU
833 * and treated as swapcache but it has no rmap yet.
834 * Calling try_to_unmap() against a page->mapping==NULL page will
835 * trigger a BUG. So handle it here.
836 * 2. An orphaned page (see truncate_complete_page) might have
837 * fs-private metadata. The page can be picked up due to memory
838 * offlining. Everywhere else except page reclaim, the page is
839 * invisible to the vm, so the page can not be migrated. So try to
840 * free the metadata, so the page can be freed.
842 if (!page
->mapping
) {
843 VM_BUG_ON(PageAnon(page
));
844 if (page_has_private(page
)) {
845 try_to_free_buffers(page
);
851 /* Establish migration ptes or remove ptes */
852 try_to_unmap(page
, TTU_MIGRATION
|TTU_IGNORE_MLOCK
|TTU_IGNORE_ACCESS
);
855 if (!page_mapped(page
))
856 rc
= move_to_new_page(newpage
, page
, remap_swapcache
, mode
);
858 if (rc
&& remap_swapcache
)
859 remove_migration_ptes(page
, page
);
861 /* Drop an anon_vma reference if we took one */
863 put_anon_vma(anon_vma
);
866 mem_cgroup_end_migration(mem
, page
, newpage
,
867 (rc
== MIGRATEPAGE_SUCCESS
||
868 rc
== MIGRATEPAGE_BALLOON_SUCCESS
));
875 * Obtain the lock on page, remove all ptes and migrate the page
876 * to the newly allocated page in newpage.
878 static int unmap_and_move(new_page_t get_new_page
, unsigned long private,
879 struct page
*page
, int force
, enum migrate_mode mode
)
883 struct page
*newpage
= get_new_page(page
, private, &result
);
888 if (page_count(page
) == 1) {
889 /* page was freed from under us. So we are done. */
893 if (unlikely(PageTransHuge(page
)))
894 if (unlikely(split_huge_page(page
)))
897 rc
= __unmap_and_move(page
, newpage
, force
, mode
);
899 if (unlikely(rc
== MIGRATEPAGE_BALLOON_SUCCESS
)) {
901 * A ballooned page has been migrated already.
902 * Now, it's the time to wrap-up counters,
903 * handle the page back to Buddy and return.
905 dec_zone_page_state(page
, NR_ISOLATED_ANON
+
906 page_is_file_cache(page
));
907 balloon_page_free(page
);
908 return MIGRATEPAGE_SUCCESS
;
913 * A page that has been migrated has all references
914 * removed and will be freed. A page that has not been
915 * migrated will have kepts its references and be
918 list_del(&page
->lru
);
919 dec_zone_page_state(page
, NR_ISOLATED_ANON
+
920 page_is_file_cache(page
));
921 putback_lru_page(page
);
924 * Move the new page to the LRU. If migration was not successful
925 * then this will free the page.
927 putback_lru_page(newpage
);
932 *result
= page_to_nid(newpage
);
938 * Counterpart of unmap_and_move_page() for hugepage migration.
940 * This function doesn't wait the completion of hugepage I/O
941 * because there is no race between I/O and migration for hugepage.
942 * Note that currently hugepage I/O occurs only in direct I/O
943 * where no lock is held and PG_writeback is irrelevant,
944 * and writeback status of all subpages are counted in the reference
945 * count of the head page (i.e. if all subpages of a 2MB hugepage are
946 * under direct I/O, the reference of the head page is 512 and a bit more.)
947 * This means that when we try to migrate hugepage whose subpages are
948 * doing direct I/O, some references remain after try_to_unmap() and
949 * hugepage migration fails without data corruption.
951 * There is also no race when direct I/O is issued on the page under migration,
952 * because then pte is replaced with migration swap entry and direct I/O code
953 * will wait in the page fault for migration to complete.
955 static int unmap_and_move_huge_page(new_page_t get_new_page
,
956 unsigned long private, struct page
*hpage
,
957 int force
, enum migrate_mode mode
)
961 struct page
*new_hpage
= get_new_page(hpage
, private, &result
);
962 struct anon_vma
*anon_vma
= NULL
;
965 * Movability of hugepages depends on architectures and hugepage size.
966 * This check is necessary because some callers of hugepage migration
967 * like soft offline and memory hotremove don't walk through page
968 * tables or check whether the hugepage is pmd-based or not before
971 if (!hugepage_migration_support(page_hstate(hpage
)))
979 if (!trylock_page(hpage
)) {
980 if (!force
|| mode
!= MIGRATE_SYNC
)
986 anon_vma
= page_get_anon_vma(hpage
);
988 try_to_unmap(hpage
, TTU_MIGRATION
|TTU_IGNORE_MLOCK
|TTU_IGNORE_ACCESS
);
990 if (!page_mapped(hpage
))
991 rc
= move_to_new_page(new_hpage
, hpage
, 1, mode
);
994 remove_migration_ptes(hpage
, hpage
);
997 put_anon_vma(anon_vma
);
1000 hugetlb_cgroup_migrate(hpage
, new_hpage
);
1005 putback_active_hugepage(hpage
);
1006 put_page(new_hpage
);
1011 *result
= page_to_nid(new_hpage
);
1017 * migrate_pages - migrate the pages specified in a list, to the free pages
1018 * supplied as the target for the page migration
1020 * @from: The list of pages to be migrated.
1021 * @get_new_page: The function used to allocate free pages to be used
1022 * as the target of the page migration.
1023 * @private: Private data to be passed on to get_new_page()
1024 * @mode: The migration mode that specifies the constraints for
1025 * page migration, if any.
1026 * @reason: The reason for page migration.
1028 * The function returns after 10 attempts or if no pages are movable any more
1029 * because the list has become empty or no retryable pages exist any more.
1030 * The caller should call putback_lru_pages() to return pages to the LRU
1031 * or free list only if ret != 0.
1033 * Returns the number of pages that were not migrated, or an error code.
1035 int migrate_pages(struct list_head
*from
, new_page_t get_new_page
,
1036 unsigned long private, enum migrate_mode mode
, int reason
)
1040 int nr_succeeded
= 0;
1044 int swapwrite
= current
->flags
& PF_SWAPWRITE
;
1048 current
->flags
|= PF_SWAPWRITE
;
1050 for(pass
= 0; pass
< 10 && retry
; pass
++) {
1053 list_for_each_entry_safe(page
, page2
, from
, lru
) {
1057 rc
= unmap_and_move_huge_page(get_new_page
,
1058 private, page
, pass
> 2, mode
);
1060 rc
= unmap_and_move(get_new_page
, private,
1061 page
, pass
> 2, mode
);
1069 case MIGRATEPAGE_SUCCESS
:
1073 /* Permanent failure */
1079 rc
= nr_failed
+ retry
;
1082 count_vm_events(PGMIGRATE_SUCCESS
, nr_succeeded
);
1084 count_vm_events(PGMIGRATE_FAIL
, nr_failed
);
1085 trace_mm_migrate_pages(nr_succeeded
, nr_failed
, mode
, reason
);
1088 current
->flags
&= ~PF_SWAPWRITE
;
1095 * Move a list of individual pages
1097 struct page_to_node
{
1104 static struct page
*new_page_node(struct page
*p
, unsigned long private,
1107 struct page_to_node
*pm
= (struct page_to_node
*)private;
1109 while (pm
->node
!= MAX_NUMNODES
&& pm
->page
!= p
)
1112 if (pm
->node
== MAX_NUMNODES
)
1115 *result
= &pm
->status
;
1118 return alloc_huge_page_node(page_hstate(compound_head(p
)),
1121 return alloc_pages_exact_node(pm
->node
,
1122 GFP_HIGHUSER_MOVABLE
| GFP_THISNODE
, 0);
1126 * Move a set of pages as indicated in the pm array. The addr
1127 * field must be set to the virtual address of the page to be moved
1128 * and the node number must contain a valid target node.
1129 * The pm array ends with node = MAX_NUMNODES.
1131 static int do_move_page_to_node_array(struct mm_struct
*mm
,
1132 struct page_to_node
*pm
,
1136 struct page_to_node
*pp
;
1137 LIST_HEAD(pagelist
);
1139 down_read(&mm
->mmap_sem
);
1142 * Build a list of pages to migrate
1144 for (pp
= pm
; pp
->node
!= MAX_NUMNODES
; pp
++) {
1145 struct vm_area_struct
*vma
;
1149 vma
= find_vma(mm
, pp
->addr
);
1150 if (!vma
|| pp
->addr
< vma
->vm_start
|| !vma_migratable(vma
))
1153 page
= follow_page(vma
, pp
->addr
, FOLL_GET
|FOLL_SPLIT
);
1155 err
= PTR_ERR(page
);
1163 /* Use PageReserved to check for zero page */
1164 if (PageReserved(page
))
1168 err
= page_to_nid(page
);
1170 if (err
== pp
->node
)
1172 * Node already in the right place
1177 if (page_mapcount(page
) > 1 &&
1181 if (PageHuge(page
)) {
1182 isolate_huge_page(page
, &pagelist
);
1186 err
= isolate_lru_page(page
);
1188 list_add_tail(&page
->lru
, &pagelist
);
1189 inc_zone_page_state(page
, NR_ISOLATED_ANON
+
1190 page_is_file_cache(page
));
1194 * Either remove the duplicate refcount from
1195 * isolate_lru_page() or drop the page ref if it was
1204 if (!list_empty(&pagelist
)) {
1205 err
= migrate_pages(&pagelist
, new_page_node
,
1206 (unsigned long)pm
, MIGRATE_SYNC
, MR_SYSCALL
);
1208 putback_movable_pages(&pagelist
);
1211 up_read(&mm
->mmap_sem
);
1216 * Migrate an array of page address onto an array of nodes and fill
1217 * the corresponding array of status.
1219 static int do_pages_move(struct mm_struct
*mm
, nodemask_t task_nodes
,
1220 unsigned long nr_pages
,
1221 const void __user
* __user
*pages
,
1222 const int __user
*nodes
,
1223 int __user
*status
, int flags
)
1225 struct page_to_node
*pm
;
1226 unsigned long chunk_nr_pages
;
1227 unsigned long chunk_start
;
1231 pm
= (struct page_to_node
*)__get_free_page(GFP_KERNEL
);
1238 * Store a chunk of page_to_node array in a page,
1239 * but keep the last one as a marker
1241 chunk_nr_pages
= (PAGE_SIZE
/ sizeof(struct page_to_node
)) - 1;
1243 for (chunk_start
= 0;
1244 chunk_start
< nr_pages
;
1245 chunk_start
+= chunk_nr_pages
) {
1248 if (chunk_start
+ chunk_nr_pages
> nr_pages
)
1249 chunk_nr_pages
= nr_pages
- chunk_start
;
1251 /* fill the chunk pm with addrs and nodes from user-space */
1252 for (j
= 0; j
< chunk_nr_pages
; j
++) {
1253 const void __user
*p
;
1257 if (get_user(p
, pages
+ j
+ chunk_start
))
1259 pm
[j
].addr
= (unsigned long) p
;
1261 if (get_user(node
, nodes
+ j
+ chunk_start
))
1265 if (node
< 0 || node
>= MAX_NUMNODES
)
1268 if (!node_state(node
, N_MEMORY
))
1272 if (!node_isset(node
, task_nodes
))
1278 /* End marker for this chunk */
1279 pm
[chunk_nr_pages
].node
= MAX_NUMNODES
;
1281 /* Migrate this chunk */
1282 err
= do_move_page_to_node_array(mm
, pm
,
1283 flags
& MPOL_MF_MOVE_ALL
);
1287 /* Return status information */
1288 for (j
= 0; j
< chunk_nr_pages
; j
++)
1289 if (put_user(pm
[j
].status
, status
+ j
+ chunk_start
)) {
1297 free_page((unsigned long)pm
);
1303 * Determine the nodes of an array of pages and store it in an array of status.
1305 static void do_pages_stat_array(struct mm_struct
*mm
, unsigned long nr_pages
,
1306 const void __user
**pages
, int *status
)
1310 down_read(&mm
->mmap_sem
);
1312 for (i
= 0; i
< nr_pages
; i
++) {
1313 unsigned long addr
= (unsigned long)(*pages
);
1314 struct vm_area_struct
*vma
;
1318 vma
= find_vma(mm
, addr
);
1319 if (!vma
|| addr
< vma
->vm_start
)
1322 page
= follow_page(vma
, addr
, 0);
1324 err
= PTR_ERR(page
);
1329 /* Use PageReserved to check for zero page */
1330 if (!page
|| PageReserved(page
))
1333 err
= page_to_nid(page
);
1341 up_read(&mm
->mmap_sem
);
1345 * Determine the nodes of a user array of pages and store it in
1346 * a user array of status.
1348 static int do_pages_stat(struct mm_struct
*mm
, unsigned long nr_pages
,
1349 const void __user
* __user
*pages
,
1352 #define DO_PAGES_STAT_CHUNK_NR 16
1353 const void __user
*chunk_pages
[DO_PAGES_STAT_CHUNK_NR
];
1354 int chunk_status
[DO_PAGES_STAT_CHUNK_NR
];
1357 unsigned long chunk_nr
;
1359 chunk_nr
= nr_pages
;
1360 if (chunk_nr
> DO_PAGES_STAT_CHUNK_NR
)
1361 chunk_nr
= DO_PAGES_STAT_CHUNK_NR
;
1363 if (copy_from_user(chunk_pages
, pages
, chunk_nr
* sizeof(*chunk_pages
)))
1366 do_pages_stat_array(mm
, chunk_nr
, chunk_pages
, chunk_status
);
1368 if (copy_to_user(status
, chunk_status
, chunk_nr
* sizeof(*status
)))
1373 nr_pages
-= chunk_nr
;
1375 return nr_pages
? -EFAULT
: 0;
1379 * Move a list of pages in the address space of the currently executing
1382 SYSCALL_DEFINE6(move_pages
, pid_t
, pid
, unsigned long, nr_pages
,
1383 const void __user
* __user
*, pages
,
1384 const int __user
*, nodes
,
1385 int __user
*, status
, int, flags
)
1387 const struct cred
*cred
= current_cred(), *tcred
;
1388 struct task_struct
*task
;
1389 struct mm_struct
*mm
;
1391 nodemask_t task_nodes
;
1394 if (flags
& ~(MPOL_MF_MOVE
|MPOL_MF_MOVE_ALL
))
1397 if ((flags
& MPOL_MF_MOVE_ALL
) && !capable(CAP_SYS_NICE
))
1400 /* Find the mm_struct */
1402 task
= pid
? find_task_by_vpid(pid
) : current
;
1407 get_task_struct(task
);
1410 * Check if this process has the right to modify the specified
1411 * process. The right exists if the process has administrative
1412 * capabilities, superuser privileges or the same
1413 * userid as the target process.
1415 tcred
= __task_cred(task
);
1416 if (!uid_eq(cred
->euid
, tcred
->suid
) && !uid_eq(cred
->euid
, tcred
->uid
) &&
1417 !uid_eq(cred
->uid
, tcred
->suid
) && !uid_eq(cred
->uid
, tcred
->uid
) &&
1418 !capable(CAP_SYS_NICE
)) {
1425 err
= security_task_movememory(task
);
1429 task_nodes
= cpuset_mems_allowed(task
);
1430 mm
= get_task_mm(task
);
1431 put_task_struct(task
);
1437 err
= do_pages_move(mm
, task_nodes
, nr_pages
, pages
,
1438 nodes
, status
, flags
);
1440 err
= do_pages_stat(mm
, nr_pages
, pages
, status
);
1446 put_task_struct(task
);
1451 * Call migration functions in the vma_ops that may prepare
1452 * memory in a vm for migration. migration functions may perform
1453 * the migration for vmas that do not have an underlying page struct.
1455 int migrate_vmas(struct mm_struct
*mm
, const nodemask_t
*to
,
1456 const nodemask_t
*from
, unsigned long flags
)
1458 struct vm_area_struct
*vma
;
1461 for (vma
= mm
->mmap
; vma
&& !err
; vma
= vma
->vm_next
) {
1462 if (vma
->vm_ops
&& vma
->vm_ops
->migrate
) {
1463 err
= vma
->vm_ops
->migrate(vma
, to
, from
, flags
);
1471 #ifdef CONFIG_NUMA_BALANCING
1473 * Returns true if this is a safe migration target node for misplaced NUMA
1474 * pages. Currently it only checks the watermarks which crude
1476 static bool migrate_balanced_pgdat(struct pglist_data
*pgdat
,
1477 unsigned long nr_migrate_pages
)
1480 for (z
= pgdat
->nr_zones
- 1; z
>= 0; z
--) {
1481 struct zone
*zone
= pgdat
->node_zones
+ z
;
1483 if (!populated_zone(zone
))
1486 if (!zone_reclaimable(zone
))
1489 /* Avoid waking kswapd by allocating pages_to_migrate pages. */
1490 if (!zone_watermark_ok(zone
, 0,
1491 high_wmark_pages(zone
) +
1500 static struct page
*alloc_misplaced_dst_page(struct page
*page
,
1504 int nid
= (int) data
;
1505 struct page
*newpage
;
1507 newpage
= alloc_pages_exact_node(nid
,
1508 (GFP_HIGHUSER_MOVABLE
| GFP_THISNODE
|
1509 __GFP_NOMEMALLOC
| __GFP_NORETRY
|
1513 page_cpupid_xchg_last(newpage
, page_cpupid_last(page
));
1519 * page migration rate limiting control.
1520 * Do not migrate more than @pages_to_migrate in a @migrate_interval_millisecs
1521 * window of time. Default here says do not migrate more than 1280M per second.
1522 * If a node is rate-limited then PTE NUMA updates are also rate-limited. However
1523 * as it is faults that reset the window, pte updates will happen unconditionally
1524 * if there has not been a fault since @pteupdate_interval_millisecs after the
1525 * throttle window closed.
1527 static unsigned int migrate_interval_millisecs __read_mostly
= 100;
1528 static unsigned int pteupdate_interval_millisecs __read_mostly
= 1000;
1529 static unsigned int ratelimit_pages __read_mostly
= 128 << (20 - PAGE_SHIFT
);
1531 /* Returns true if NUMA migration is currently rate limited */
1532 bool migrate_ratelimited(int node
)
1534 pg_data_t
*pgdat
= NODE_DATA(node
);
1536 if (time_after(jiffies
, pgdat
->numabalancing_migrate_next_window
+
1537 msecs_to_jiffies(pteupdate_interval_millisecs
)))
1540 if (pgdat
->numabalancing_migrate_nr_pages
< ratelimit_pages
)
1546 /* Returns true if the node is migrate rate-limited after the update */
1547 bool numamigrate_update_ratelimit(pg_data_t
*pgdat
, unsigned long nr_pages
)
1549 bool rate_limited
= false;
1552 * Rate-limit the amount of data that is being migrated to a node.
1553 * Optimal placement is no good if the memory bus is saturated and
1554 * all the time is being spent migrating!
1556 spin_lock(&pgdat
->numabalancing_migrate_lock
);
1557 if (time_after(jiffies
, pgdat
->numabalancing_migrate_next_window
)) {
1558 pgdat
->numabalancing_migrate_nr_pages
= 0;
1559 pgdat
->numabalancing_migrate_next_window
= jiffies
+
1560 msecs_to_jiffies(migrate_interval_millisecs
);
1562 if (pgdat
->numabalancing_migrate_nr_pages
> ratelimit_pages
)
1563 rate_limited
= true;
1565 pgdat
->numabalancing_migrate_nr_pages
+= nr_pages
;
1566 spin_unlock(&pgdat
->numabalancing_migrate_lock
);
1568 return rate_limited
;
1571 int numamigrate_isolate_page(pg_data_t
*pgdat
, struct page
*page
)
1575 VM_BUG_ON(compound_order(page
) && !PageTransHuge(page
));
1577 /* Avoid migrating to a node that is nearly full */
1578 if (!migrate_balanced_pgdat(pgdat
, 1UL << compound_order(page
)))
1581 if (isolate_lru_page(page
))
1585 * migrate_misplaced_transhuge_page() skips page migration's usual
1586 * check on page_count(), so we must do it here, now that the page
1587 * has been isolated: a GUP pin, or any other pin, prevents migration.
1588 * The expected page count is 3: 1 for page's mapcount and 1 for the
1589 * caller's pin and 1 for the reference taken by isolate_lru_page().
1591 if (PageTransHuge(page
) && page_count(page
) != 3) {
1592 putback_lru_page(page
);
1596 page_lru
= page_is_file_cache(page
);
1597 mod_zone_page_state(page_zone(page
), NR_ISOLATED_ANON
+ page_lru
,
1598 hpage_nr_pages(page
));
1601 * Isolating the page has taken another reference, so the
1602 * caller's reference can be safely dropped without the page
1603 * disappearing underneath us during migration.
1610 * Attempt to migrate a misplaced page to the specified destination
1611 * node. Caller is expected to have an elevated reference count on
1612 * the page that will be dropped by this function before returning.
1614 int migrate_misplaced_page(struct page
*page
, struct vm_area_struct
*vma
,
1617 pg_data_t
*pgdat
= NODE_DATA(node
);
1620 LIST_HEAD(migratepages
);
1623 * Don't migrate file pages that are mapped in multiple processes
1624 * with execute permissions as they are probably shared libraries.
1626 if (page_mapcount(page
) != 1 && page_is_file_cache(page
) &&
1627 (vma
->vm_flags
& VM_EXEC
))
1631 * Rate-limit the amount of data that is being migrated to a node.
1632 * Optimal placement is no good if the memory bus is saturated and
1633 * all the time is being spent migrating!
1635 if (numamigrate_update_ratelimit(pgdat
, 1))
1638 isolated
= numamigrate_isolate_page(pgdat
, page
);
1642 list_add(&page
->lru
, &migratepages
);
1643 nr_remaining
= migrate_pages(&migratepages
, alloc_misplaced_dst_page
,
1644 node
, MIGRATE_ASYNC
, MR_NUMA_MISPLACED
);
1646 putback_lru_pages(&migratepages
);
1649 count_vm_numa_event(NUMA_PAGE_MIGRATE
);
1650 BUG_ON(!list_empty(&migratepages
));
1657 #endif /* CONFIG_NUMA_BALANCING */
1659 #if defined(CONFIG_NUMA_BALANCING) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
1661 * Migrates a THP to a given target node. page must be locked and is unlocked
1664 int migrate_misplaced_transhuge_page(struct mm_struct
*mm
,
1665 struct vm_area_struct
*vma
,
1666 pmd_t
*pmd
, pmd_t entry
,
1667 unsigned long address
,
1668 struct page
*page
, int node
)
1671 unsigned long haddr
= address
& HPAGE_PMD_MASK
;
1672 pg_data_t
*pgdat
= NODE_DATA(node
);
1674 struct page
*new_page
= NULL
;
1675 struct mem_cgroup
*memcg
= NULL
;
1676 int page_lru
= page_is_file_cache(page
);
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_cpupid_xchg_last(new_page
, page_cpupid_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 ptl
= pmd_lock(mm
, pmd
);
1711 if (unlikely(!pmd_same(*pmd
, entry
))) {
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
);
1733 * Traditional migration needs to prepare the memcg charge
1734 * transaction early to prevent the old page from being
1735 * uncharged when installing migration entries. Here we can
1736 * save the potential rollback and start the charge transfer
1737 * only when migration is already known to end successfully.
1739 mem_cgroup_prepare_migration(page
, new_page
, &memcg
);
1741 entry
= mk_pmd(new_page
, vma
->vm_page_prot
);
1742 entry
= pmd_mknonnuma(entry
);
1743 entry
= maybe_pmd_mkwrite(pmd_mkdirty(entry
), vma
);
1744 entry
= pmd_mkhuge(entry
);
1746 pmdp_clear_flush(vma
, haddr
, pmd
);
1747 set_pmd_at(mm
, haddr
, pmd
, entry
);
1748 page_add_new_anon_rmap(new_page
, vma
, haddr
);
1749 update_mmu_cache_pmd(vma
, address
, &entry
);
1750 page_remove_rmap(page
);
1752 * Finish the charge transaction under the page table lock to
1753 * prevent split_huge_page() from dividing up the charge
1754 * before it's fully transferred to the new page.
1756 mem_cgroup_end_migration(memcg
, page
, new_page
, true);
1759 unlock_page(new_page
);
1761 put_page(page
); /* Drop the rmap reference */
1762 put_page(page
); /* Drop the LRU isolation reference */
1764 count_vm_events(PGMIGRATE_SUCCESS
, HPAGE_PMD_NR
);
1765 count_vm_numa_events(NUMA_PAGE_MIGRATE
, HPAGE_PMD_NR
);
1767 mod_zone_page_state(page_zone(page
),
1768 NR_ISOLATED_ANON
+ page_lru
,
1773 count_vm_events(PGMIGRATE_FAIL
, HPAGE_PMD_NR
);
1775 entry
= pmd_mknonnuma(entry
);
1776 set_pmd_at(mm
, haddr
, pmd
, entry
);
1777 update_mmu_cache_pmd(vma
, address
, &entry
);
1783 #endif /* CONFIG_NUMA_BALANCING */
1785 #endif /* CONFIG_NUMA */