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>
39 #include <linux/mmu_notifier.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 * Put previously isolated pages back onto the appropriate lists
76 * from where they were once taken off for compaction/migration.
78 * This function shall be used whenever the isolated pageset has been
79 * built from lru, balloon, hugetlbfs page. See isolate_migratepages_range()
80 * and isolate_huge_page().
82 void putback_movable_pages(struct list_head
*l
)
87 list_for_each_entry_safe(page
, page2
, l
, lru
) {
88 if (unlikely(PageHuge(page
))) {
89 putback_active_hugepage(page
);
93 dec_zone_page_state(page
, NR_ISOLATED_ANON
+
94 page_is_file_cache(page
));
95 if (unlikely(isolated_balloon_page(page
)))
96 balloon_page_putback(page
);
98 putback_lru_page(page
);
103 * Restore a potential migration pte to a working pte entry
105 static int remove_migration_pte(struct page
*new, struct vm_area_struct
*vma
,
106 unsigned long addr
, void *old
)
108 struct mm_struct
*mm
= vma
->vm_mm
;
114 if (unlikely(PageHuge(new))) {
115 ptep
= huge_pte_offset(mm
, addr
);
118 ptl
= huge_pte_lockptr(hstate_vma(vma
), mm
, ptep
);
120 pmd
= mm_find_pmd(mm
, addr
);
123 if (pmd_trans_huge(*pmd
))
126 ptep
= pte_offset_map(pmd
, addr
);
129 * Peek to check is_swap_pte() before taking ptlock? No, we
130 * can race mremap's move_ptes(), which skips anon_vma lock.
133 ptl
= pte_lockptr(mm
, pmd
);
138 if (!is_swap_pte(pte
))
141 entry
= pte_to_swp_entry(pte
);
143 if (!is_migration_entry(entry
) ||
144 migration_entry_to_page(entry
) != old
)
148 pte
= pte_mkold(mk_pte(new, vma
->vm_page_prot
));
149 if (pte_swp_soft_dirty(*ptep
))
150 pte
= pte_mksoft_dirty(pte
);
151 if (is_write_migration_entry(entry
))
152 pte
= pte_mkwrite(pte
);
153 #ifdef CONFIG_HUGETLB_PAGE
155 pte
= pte_mkhuge(pte
);
156 pte
= arch_make_huge_pte(pte
, vma
, new, 0);
159 flush_dcache_page(new);
160 set_pte_at(mm
, addr
, ptep
, pte
);
164 hugepage_add_anon_rmap(new, vma
, addr
);
167 } else if (PageAnon(new))
168 page_add_anon_rmap(new, vma
, addr
);
170 page_add_file_rmap(new);
172 /* No need to invalidate - it was non-present before */
173 update_mmu_cache(vma
, addr
, ptep
);
175 pte_unmap_unlock(ptep
, ptl
);
181 * Get rid of all migration entries and replace them by
182 * references to the indicated page.
184 static void remove_migration_ptes(struct page
*old
, struct page
*new)
186 struct rmap_walk_control rwc
= {
187 .rmap_one
= remove_migration_pte
,
191 rmap_walk(new, &rwc
);
195 * Something used the pte of a page under migration. We need to
196 * get to the page and wait until migration is finished.
197 * When we return from this function the fault will be retried.
199 static void __migration_entry_wait(struct mm_struct
*mm
, pte_t
*ptep
,
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
);
234 void migration_entry_wait(struct mm_struct
*mm
, pmd_t
*pmd
,
235 unsigned long address
)
237 spinlock_t
*ptl
= pte_lockptr(mm
, pmd
);
238 pte_t
*ptep
= pte_offset_map(pmd
, address
);
239 __migration_entry_wait(mm
, ptep
, ptl
);
242 void migration_entry_wait_huge(struct vm_area_struct
*vma
,
243 struct mm_struct
*mm
, pte_t
*pte
)
245 spinlock_t
*ptl
= huge_pte_lockptr(hstate_vma(vma
), mm
, pte
);
246 __migration_entry_wait(mm
, pte
, ptl
);
250 /* Returns true if all buffers are successfully locked */
251 static bool buffer_migrate_lock_buffers(struct buffer_head
*head
,
252 enum migrate_mode mode
)
254 struct buffer_head
*bh
= head
;
256 /* Simple case, sync compaction */
257 if (mode
!= MIGRATE_ASYNC
) {
261 bh
= bh
->b_this_page
;
263 } while (bh
!= head
);
268 /* async case, we cannot block on lock_buffer so use trylock_buffer */
271 if (!trylock_buffer(bh
)) {
273 * We failed to lock the buffer and cannot stall in
274 * async migration. Release the taken locks
276 struct buffer_head
*failed_bh
= bh
;
279 while (bh
!= failed_bh
) {
282 bh
= bh
->b_this_page
;
287 bh
= bh
->b_this_page
;
288 } while (bh
!= head
);
292 static inline bool buffer_migrate_lock_buffers(struct buffer_head
*head
,
293 enum migrate_mode mode
)
297 #endif /* CONFIG_BLOCK */
300 * Replace the page in the mapping.
302 * The number of remaining references must be:
303 * 1 for anonymous pages without a mapping
304 * 2 for pages with a mapping
305 * 3 for pages with a mapping and PagePrivate/PagePrivate2 set.
307 int migrate_page_move_mapping(struct address_space
*mapping
,
308 struct page
*newpage
, struct page
*page
,
309 struct buffer_head
*head
, enum migrate_mode mode
,
312 int expected_count
= 1 + extra_count
;
316 /* Anonymous page without mapping */
317 if (page_count(page
) != expected_count
)
319 return MIGRATEPAGE_SUCCESS
;
322 spin_lock_irq(&mapping
->tree_lock
);
324 pslot
= radix_tree_lookup_slot(&mapping
->page_tree
,
327 expected_count
+= 1 + page_has_private(page
);
328 if (page_count(page
) != expected_count
||
329 radix_tree_deref_slot_protected(pslot
, &mapping
->tree_lock
) != page
) {
330 spin_unlock_irq(&mapping
->tree_lock
);
334 if (!page_freeze_refs(page
, expected_count
)) {
335 spin_unlock_irq(&mapping
->tree_lock
);
340 * In the async migration case of moving a page with buffers, lock the
341 * buffers using trylock before the mapping is moved. If the mapping
342 * was moved, we later failed to lock the buffers and could not move
343 * the mapping back due to an elevated page count, we would have to
344 * block waiting on other references to be dropped.
346 if (mode
== MIGRATE_ASYNC
&& head
&&
347 !buffer_migrate_lock_buffers(head
, mode
)) {
348 page_unfreeze_refs(page
, expected_count
);
349 spin_unlock_irq(&mapping
->tree_lock
);
354 * Now we know that no one else is looking at the page.
356 get_page(newpage
); /* add cache reference */
357 if (PageSwapCache(page
)) {
358 SetPageSwapCache(newpage
);
359 set_page_private(newpage
, page_private(page
));
362 radix_tree_replace_slot(pslot
, newpage
);
365 * Drop cache reference from old page by unfreezing
366 * to one less reference.
367 * We know this isn't the last reference.
369 page_unfreeze_refs(page
, expected_count
- 1);
372 * If moved to a different zone then also account
373 * the page for that zone. Other VM counters will be
374 * taken care of when we establish references to the
375 * new page and drop references to the old page.
377 * Note that anonymous pages are accounted for
378 * via NR_FILE_PAGES and NR_ANON_PAGES if they
379 * are mapped to swap space.
381 __dec_zone_page_state(page
, NR_FILE_PAGES
);
382 __inc_zone_page_state(newpage
, NR_FILE_PAGES
);
383 if (!PageSwapCache(page
) && PageSwapBacked(page
)) {
384 __dec_zone_page_state(page
, NR_SHMEM
);
385 __inc_zone_page_state(newpage
, NR_SHMEM
);
387 spin_unlock_irq(&mapping
->tree_lock
);
389 return MIGRATEPAGE_SUCCESS
;
393 * The expected number of remaining references is the same as that
394 * of migrate_page_move_mapping().
396 int migrate_huge_page_move_mapping(struct address_space
*mapping
,
397 struct page
*newpage
, struct page
*page
)
403 if (page_count(page
) != 1)
405 return MIGRATEPAGE_SUCCESS
;
408 spin_lock_irq(&mapping
->tree_lock
);
410 pslot
= radix_tree_lookup_slot(&mapping
->page_tree
,
413 expected_count
= 2 + page_has_private(page
);
414 if (page_count(page
) != expected_count
||
415 radix_tree_deref_slot_protected(pslot
, &mapping
->tree_lock
) != page
) {
416 spin_unlock_irq(&mapping
->tree_lock
);
420 if (!page_freeze_refs(page
, expected_count
)) {
421 spin_unlock_irq(&mapping
->tree_lock
);
427 radix_tree_replace_slot(pslot
, newpage
);
429 page_unfreeze_refs(page
, expected_count
- 1);
431 spin_unlock_irq(&mapping
->tree_lock
);
432 return MIGRATEPAGE_SUCCESS
;
436 * Gigantic pages are so large that we do not guarantee that page++ pointer
437 * arithmetic will work across the entire page. We need something more
440 static void __copy_gigantic_page(struct page
*dst
, struct page
*src
,
444 struct page
*dst_base
= dst
;
445 struct page
*src_base
= src
;
447 for (i
= 0; i
< nr_pages
; ) {
449 copy_highpage(dst
, src
);
452 dst
= mem_map_next(dst
, dst_base
, i
);
453 src
= mem_map_next(src
, src_base
, i
);
457 static void copy_huge_page(struct page
*dst
, struct page
*src
)
464 struct hstate
*h
= page_hstate(src
);
465 nr_pages
= pages_per_huge_page(h
);
467 if (unlikely(nr_pages
> MAX_ORDER_NR_PAGES
)) {
468 __copy_gigantic_page(dst
, src
, nr_pages
);
473 BUG_ON(!PageTransHuge(src
));
474 nr_pages
= hpage_nr_pages(src
);
477 for (i
= 0; i
< nr_pages
; i
++) {
479 copy_highpage(dst
+ i
, src
+ i
);
484 * Copy the page to its new location
486 void migrate_page_copy(struct page
*newpage
, struct page
*page
)
490 if (PageHuge(page
) || PageTransHuge(page
))
491 copy_huge_page(newpage
, page
);
493 copy_highpage(newpage
, page
);
496 SetPageError(newpage
);
497 if (PageReferenced(page
))
498 SetPageReferenced(newpage
);
499 if (PageUptodate(page
))
500 SetPageUptodate(newpage
);
501 if (TestClearPageActive(page
)) {
502 VM_BUG_ON(PageUnevictable(page
));
503 SetPageActive(newpage
);
504 } else if (TestClearPageUnevictable(page
))
505 SetPageUnevictable(newpage
);
506 if (PageChecked(page
))
507 SetPageChecked(newpage
);
508 if (PageMappedToDisk(page
))
509 SetPageMappedToDisk(newpage
);
511 if (PageDirty(page
)) {
512 clear_page_dirty_for_io(page
);
514 * Want to mark the page and the radix tree as dirty, and
515 * redo the accounting that clear_page_dirty_for_io undid,
516 * but we can't use set_page_dirty because that function
517 * is actually a signal that all of the page has become dirty.
518 * Whereas only part of our page may be dirty.
520 if (PageSwapBacked(page
))
521 SetPageDirty(newpage
);
523 __set_page_dirty_nobuffers(newpage
);
527 * Copy NUMA information to the new page, to prevent over-eager
528 * future migrations of this same page.
530 cpupid
= page_cpupid_xchg_last(page
, -1);
531 page_cpupid_xchg_last(newpage
, cpupid
);
533 mlock_migrate_page(newpage
, page
);
534 ksm_migrate_page(newpage
, page
);
536 * Please do not reorder this without considering how mm/ksm.c's
537 * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache().
539 ClearPageSwapCache(page
);
540 ClearPagePrivate(page
);
541 set_page_private(page
, 0);
544 * If any waiters have accumulated on the new page then
547 if (PageWriteback(newpage
))
548 end_page_writeback(newpage
);
551 /************************************************************
552 * Migration functions
553 ***********************************************************/
555 /* Always fail migration. Used for mappings that are not movable */
556 int fail_migrate_page(struct address_space
*mapping
,
557 struct page
*newpage
, struct page
*page
)
561 EXPORT_SYMBOL(fail_migrate_page
);
564 * Common logic to directly migrate a single page suitable for
565 * pages that do not use PagePrivate/PagePrivate2.
567 * Pages are locked upon entry and exit.
569 int migrate_page(struct address_space
*mapping
,
570 struct page
*newpage
, struct page
*page
,
571 enum migrate_mode mode
)
575 BUG_ON(PageWriteback(page
)); /* Writeback must be complete */
577 rc
= migrate_page_move_mapping(mapping
, newpage
, page
, NULL
, mode
, 0);
579 if (rc
!= MIGRATEPAGE_SUCCESS
)
582 migrate_page_copy(newpage
, page
);
583 return MIGRATEPAGE_SUCCESS
;
585 EXPORT_SYMBOL(migrate_page
);
589 * Migration function for pages with buffers. This function can only be used
590 * if the underlying filesystem guarantees that no other references to "page"
593 int buffer_migrate_page(struct address_space
*mapping
,
594 struct page
*newpage
, struct page
*page
, enum migrate_mode mode
)
596 struct buffer_head
*bh
, *head
;
599 if (!page_has_buffers(page
))
600 return migrate_page(mapping
, newpage
, page
, mode
);
602 head
= page_buffers(page
);
604 rc
= migrate_page_move_mapping(mapping
, newpage
, page
, head
, mode
, 0);
606 if (rc
!= MIGRATEPAGE_SUCCESS
)
610 * In the async case, migrate_page_move_mapping locked the buffers
611 * with an IRQ-safe spinlock held. In the sync case, the buffers
612 * need to be locked now
614 if (mode
!= MIGRATE_ASYNC
)
615 BUG_ON(!buffer_migrate_lock_buffers(head
, mode
));
617 ClearPagePrivate(page
);
618 set_page_private(newpage
, page_private(page
));
619 set_page_private(page
, 0);
625 set_bh_page(bh
, newpage
, bh_offset(bh
));
626 bh
= bh
->b_this_page
;
628 } while (bh
!= head
);
630 SetPagePrivate(newpage
);
632 migrate_page_copy(newpage
, page
);
638 bh
= bh
->b_this_page
;
640 } while (bh
!= head
);
642 return MIGRATEPAGE_SUCCESS
;
644 EXPORT_SYMBOL(buffer_migrate_page
);
648 * Writeback a page to clean the dirty state
650 static int writeout(struct address_space
*mapping
, struct page
*page
)
652 struct writeback_control wbc
= {
653 .sync_mode
= WB_SYNC_NONE
,
656 .range_end
= LLONG_MAX
,
661 if (!mapping
->a_ops
->writepage
)
662 /* No write method for the address space */
665 if (!clear_page_dirty_for_io(page
))
666 /* Someone else already triggered a write */
670 * A dirty page may imply that the underlying filesystem has
671 * the page on some queue. So the page must be clean for
672 * migration. Writeout may mean we loose the lock and the
673 * page state is no longer what we checked for earlier.
674 * At this point we know that the migration attempt cannot
677 remove_migration_ptes(page
, page
);
679 rc
= mapping
->a_ops
->writepage(page
, &wbc
);
681 if (rc
!= AOP_WRITEPAGE_ACTIVATE
)
682 /* unlocked. Relock */
685 return (rc
< 0) ? -EIO
: -EAGAIN
;
689 * Default handling if a filesystem does not provide a migration function.
691 static int fallback_migrate_page(struct address_space
*mapping
,
692 struct page
*newpage
, struct page
*page
, enum migrate_mode mode
)
694 if (PageDirty(page
)) {
695 /* Only writeback pages in full synchronous migration */
696 if (mode
!= MIGRATE_SYNC
)
698 return writeout(mapping
, page
);
702 * Buffers may be managed in a filesystem specific way.
703 * We must have no buffers or drop them.
705 if (page_has_private(page
) &&
706 !try_to_release_page(page
, GFP_KERNEL
))
709 return migrate_page(mapping
, newpage
, page
, mode
);
713 * Move a page to a newly allocated page
714 * The page is locked and all ptes have been successfully removed.
716 * The new page will have replaced the old page if this function
721 * MIGRATEPAGE_SUCCESS - success
723 static int move_to_new_page(struct page
*newpage
, struct page
*page
,
724 int remap_swapcache
, enum migrate_mode mode
)
726 struct address_space
*mapping
;
730 * Block others from accessing the page when we get around to
731 * establishing additional references. We are the only one
732 * holding a reference to the new page at this point.
734 if (!trylock_page(newpage
))
737 /* Prepare mapping for the new page.*/
738 newpage
->index
= page
->index
;
739 newpage
->mapping
= page
->mapping
;
740 if (PageSwapBacked(page
))
741 SetPageSwapBacked(newpage
);
743 mapping
= page_mapping(page
);
745 rc
= migrate_page(mapping
, newpage
, page
, mode
);
746 else if (mapping
->a_ops
->migratepage
)
748 * Most pages have a mapping and most filesystems provide a
749 * migratepage callback. Anonymous pages are part of swap
750 * space which also has its own migratepage callback. This
751 * is the most common path for page migration.
753 rc
= mapping
->a_ops
->migratepage(mapping
,
754 newpage
, page
, mode
);
756 rc
= fallback_migrate_page(mapping
, newpage
, page
, mode
);
758 if (rc
!= MIGRATEPAGE_SUCCESS
) {
759 newpage
->mapping
= NULL
;
762 remove_migration_ptes(page
, newpage
);
763 page
->mapping
= NULL
;
766 unlock_page(newpage
);
771 static int __unmap_and_move(struct page
*page
, struct page
*newpage
,
772 int force
, enum migrate_mode mode
)
775 int remap_swapcache
= 1;
776 struct mem_cgroup
*mem
;
777 struct anon_vma
*anon_vma
= NULL
;
779 if (!trylock_page(page
)) {
780 if (!force
|| mode
== MIGRATE_ASYNC
)
784 * It's not safe for direct compaction to call lock_page.
785 * For example, during page readahead pages are added locked
786 * to the LRU. Later, when the IO completes the pages are
787 * marked uptodate and unlocked. However, the queueing
788 * could be merging multiple pages for one bio (e.g.
789 * mpage_readpages). If an allocation happens for the
790 * second or third page, the process can end up locking
791 * the same page twice and deadlocking. Rather than
792 * trying to be clever about what pages can be locked,
793 * avoid the use of lock_page for direct compaction
796 if (current
->flags
& PF_MEMALLOC
)
802 /* charge against new page */
803 mem_cgroup_prepare_migration(page
, newpage
, &mem
);
805 if (PageWriteback(page
)) {
807 * Only in the case of a full synchronous migration is it
808 * necessary to wait for PageWriteback. In the async case,
809 * the retry loop is too short and in the sync-light case,
810 * the overhead of stalling is too much
812 if (mode
!= MIGRATE_SYNC
) {
818 wait_on_page_writeback(page
);
821 * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
822 * we cannot notice that anon_vma is freed while we migrates a page.
823 * This get_anon_vma() delays freeing anon_vma pointer until the end
824 * of migration. File cache pages are no problem because of page_lock()
825 * File Caches may use write_page() or lock_page() in migration, then,
826 * just care Anon page here.
828 if (PageAnon(page
) && !PageKsm(page
)) {
830 * Only page_lock_anon_vma_read() understands the subtleties of
831 * getting a hold on an anon_vma from outside one of its mms.
833 anon_vma
= page_get_anon_vma(page
);
838 } else if (PageSwapCache(page
)) {
840 * We cannot be sure that the anon_vma of an unmapped
841 * swapcache page is safe to use because we don't
842 * know in advance if the VMA that this page belonged
843 * to still exists. If the VMA and others sharing the
844 * data have been freed, then the anon_vma could
845 * already be invalid.
847 * To avoid this possibility, swapcache pages get
848 * migrated but are not remapped when migration
857 if (unlikely(balloon_page_movable(page
))) {
859 * A ballooned page does not need any special attention from
860 * physical to virtual reverse mapping procedures.
861 * Skip any attempt to unmap PTEs or to remap swap cache,
862 * in order to avoid burning cycles at rmap level, and perform
863 * the page migration right away (proteced by page lock).
865 rc
= balloon_page_migrate(newpage
, page
, mode
);
870 * Corner case handling:
871 * 1. When a new swap-cache page is read into, it is added to the LRU
872 * and treated as swapcache but it has no rmap yet.
873 * Calling try_to_unmap() against a page->mapping==NULL page will
874 * trigger a BUG. So handle it here.
875 * 2. An orphaned page (see truncate_complete_page) might have
876 * fs-private metadata. The page can be picked up due to memory
877 * offlining. Everywhere else except page reclaim, the page is
878 * invisible to the vm, so the page can not be migrated. So try to
879 * free the metadata, so the page can be freed.
881 if (!page
->mapping
) {
882 VM_BUG_ON(PageAnon(page
));
883 if (page_has_private(page
)) {
884 try_to_free_buffers(page
);
890 /* Establish migration ptes or remove ptes */
891 try_to_unmap(page
, TTU_MIGRATION
|TTU_IGNORE_MLOCK
|TTU_IGNORE_ACCESS
);
894 if (!page_mapped(page
))
895 rc
= move_to_new_page(newpage
, page
, remap_swapcache
, mode
);
897 if (rc
&& remap_swapcache
)
898 remove_migration_ptes(page
, page
);
900 /* Drop an anon_vma reference if we took one */
902 put_anon_vma(anon_vma
);
905 mem_cgroup_end_migration(mem
, page
, newpage
,
906 (rc
== MIGRATEPAGE_SUCCESS
||
907 rc
== MIGRATEPAGE_BALLOON_SUCCESS
));
914 * Obtain the lock on page, remove all ptes and migrate the page
915 * to the newly allocated page in newpage.
917 static int unmap_and_move(new_page_t get_new_page
, unsigned long private,
918 struct page
*page
, int force
, enum migrate_mode mode
)
922 struct page
*newpage
= get_new_page(page
, private, &result
);
927 if (page_count(page
) == 1) {
928 /* page was freed from under us. So we are done. */
932 if (unlikely(PageTransHuge(page
)))
933 if (unlikely(split_huge_page(page
)))
936 rc
= __unmap_and_move(page
, newpage
, force
, mode
);
938 if (unlikely(rc
== MIGRATEPAGE_BALLOON_SUCCESS
)) {
940 * A ballooned page has been migrated already.
941 * Now, it's the time to wrap-up counters,
942 * handle the page back to Buddy and return.
944 dec_zone_page_state(page
, NR_ISOLATED_ANON
+
945 page_is_file_cache(page
));
946 balloon_page_free(page
);
947 return MIGRATEPAGE_SUCCESS
;
952 * A page that has been migrated has all references
953 * removed and will be freed. A page that has not been
954 * migrated will have kepts its references and be
957 list_del(&page
->lru
);
958 dec_zone_page_state(page
, NR_ISOLATED_ANON
+
959 page_is_file_cache(page
));
960 putback_lru_page(page
);
963 * Move the new page to the LRU. If migration was not successful
964 * then this will free the page.
966 putback_lru_page(newpage
);
971 *result
= page_to_nid(newpage
);
977 * Counterpart of unmap_and_move_page() for hugepage migration.
979 * This function doesn't wait the completion of hugepage I/O
980 * because there is no race between I/O and migration for hugepage.
981 * Note that currently hugepage I/O occurs only in direct I/O
982 * where no lock is held and PG_writeback is irrelevant,
983 * and writeback status of all subpages are counted in the reference
984 * count of the head page (i.e. if all subpages of a 2MB hugepage are
985 * under direct I/O, the reference of the head page is 512 and a bit more.)
986 * This means that when we try to migrate hugepage whose subpages are
987 * doing direct I/O, some references remain after try_to_unmap() and
988 * hugepage migration fails without data corruption.
990 * There is also no race when direct I/O is issued on the page under migration,
991 * because then pte is replaced with migration swap entry and direct I/O code
992 * will wait in the page fault for migration to complete.
994 static int unmap_and_move_huge_page(new_page_t get_new_page
,
995 unsigned long private, struct page
*hpage
,
996 int force
, enum migrate_mode mode
)
1000 struct page
*new_hpage
;
1001 struct anon_vma
*anon_vma
= NULL
;
1004 * Movability of hugepages depends on architectures and hugepage size.
1005 * This check is necessary because some callers of hugepage migration
1006 * like soft offline and memory hotremove don't walk through page
1007 * tables or check whether the hugepage is pmd-based or not before
1008 * kicking migration.
1010 if (!hugepage_migration_support(page_hstate(hpage
))) {
1011 putback_active_hugepage(hpage
);
1015 new_hpage
= get_new_page(hpage
, private, &result
);
1021 if (!trylock_page(hpage
)) {
1022 if (!force
|| mode
!= MIGRATE_SYNC
)
1027 if (PageAnon(hpage
))
1028 anon_vma
= page_get_anon_vma(hpage
);
1030 try_to_unmap(hpage
, TTU_MIGRATION
|TTU_IGNORE_MLOCK
|TTU_IGNORE_ACCESS
);
1032 if (!page_mapped(hpage
))
1033 rc
= move_to_new_page(new_hpage
, hpage
, 1, mode
);
1036 remove_migration_ptes(hpage
, hpage
);
1039 put_anon_vma(anon_vma
);
1042 hugetlb_cgroup_migrate(hpage
, new_hpage
);
1047 putback_active_hugepage(hpage
);
1048 put_page(new_hpage
);
1053 *result
= page_to_nid(new_hpage
);
1059 * migrate_pages - migrate the pages specified in a list, to the free pages
1060 * supplied as the target for the page migration
1062 * @from: The list of pages to be migrated.
1063 * @get_new_page: The function used to allocate free pages to be used
1064 * as the target of the page migration.
1065 * @private: Private data to be passed on to get_new_page()
1066 * @mode: The migration mode that specifies the constraints for
1067 * page migration, if any.
1068 * @reason: The reason for page migration.
1070 * The function returns after 10 attempts or if no pages are movable any more
1071 * because the list has become empty or no retryable pages exist any more.
1072 * The caller should call putback_lru_pages() to return pages to the LRU
1073 * or free list only if ret != 0.
1075 * Returns the number of pages that were not migrated, or an error code.
1077 int migrate_pages(struct list_head
*from
, new_page_t get_new_page
,
1078 unsigned long private, enum migrate_mode mode
, int reason
)
1082 int nr_succeeded
= 0;
1086 int swapwrite
= current
->flags
& PF_SWAPWRITE
;
1090 current
->flags
|= PF_SWAPWRITE
;
1092 for(pass
= 0; pass
< 10 && retry
; pass
++) {
1095 list_for_each_entry_safe(page
, page2
, from
, lru
) {
1099 rc
= unmap_and_move_huge_page(get_new_page
,
1100 private, page
, pass
> 2, mode
);
1102 rc
= unmap_and_move(get_new_page
, private,
1103 page
, pass
> 2, mode
);
1111 case MIGRATEPAGE_SUCCESS
:
1116 * Permanent failure (-EBUSY, -ENOSYS, etc.):
1117 * unlike -EAGAIN case, the failed page is
1118 * removed from migration page list and not
1119 * retried in the next outer loop.
1126 rc
= nr_failed
+ retry
;
1129 count_vm_events(PGMIGRATE_SUCCESS
, nr_succeeded
);
1131 count_vm_events(PGMIGRATE_FAIL
, nr_failed
);
1132 trace_mm_migrate_pages(nr_succeeded
, nr_failed
, mode
, reason
);
1135 current
->flags
&= ~PF_SWAPWRITE
;
1142 * Move a list of individual pages
1144 struct page_to_node
{
1151 static struct page
*new_page_node(struct page
*p
, unsigned long private,
1154 struct page_to_node
*pm
= (struct page_to_node
*)private;
1156 while (pm
->node
!= MAX_NUMNODES
&& pm
->page
!= p
)
1159 if (pm
->node
== MAX_NUMNODES
)
1162 *result
= &pm
->status
;
1165 return alloc_huge_page_node(page_hstate(compound_head(p
)),
1168 return alloc_pages_exact_node(pm
->node
,
1169 GFP_HIGHUSER_MOVABLE
| GFP_THISNODE
, 0);
1173 * Move a set of pages as indicated in the pm array. The addr
1174 * field must be set to the virtual address of the page to be moved
1175 * and the node number must contain a valid target node.
1176 * The pm array ends with node = MAX_NUMNODES.
1178 static int do_move_page_to_node_array(struct mm_struct
*mm
,
1179 struct page_to_node
*pm
,
1183 struct page_to_node
*pp
;
1184 LIST_HEAD(pagelist
);
1186 down_read(&mm
->mmap_sem
);
1189 * Build a list of pages to migrate
1191 for (pp
= pm
; pp
->node
!= MAX_NUMNODES
; pp
++) {
1192 struct vm_area_struct
*vma
;
1196 vma
= find_vma(mm
, pp
->addr
);
1197 if (!vma
|| pp
->addr
< vma
->vm_start
|| !vma_migratable(vma
))
1200 page
= follow_page(vma
, pp
->addr
, FOLL_GET
|FOLL_SPLIT
);
1202 err
= PTR_ERR(page
);
1210 /* Use PageReserved to check for zero page */
1211 if (PageReserved(page
))
1215 err
= page_to_nid(page
);
1217 if (err
== pp
->node
)
1219 * Node already in the right place
1224 if (page_mapcount(page
) > 1 &&
1228 if (PageHuge(page
)) {
1229 isolate_huge_page(page
, &pagelist
);
1233 err
= isolate_lru_page(page
);
1235 list_add_tail(&page
->lru
, &pagelist
);
1236 inc_zone_page_state(page
, NR_ISOLATED_ANON
+
1237 page_is_file_cache(page
));
1241 * Either remove the duplicate refcount from
1242 * isolate_lru_page() or drop the page ref if it was
1251 if (!list_empty(&pagelist
)) {
1252 err
= migrate_pages(&pagelist
, new_page_node
,
1253 (unsigned long)pm
, MIGRATE_SYNC
, MR_SYSCALL
);
1255 putback_movable_pages(&pagelist
);
1258 up_read(&mm
->mmap_sem
);
1263 * Migrate an array of page address onto an array of nodes and fill
1264 * the corresponding array of status.
1266 static int do_pages_move(struct mm_struct
*mm
, nodemask_t task_nodes
,
1267 unsigned long nr_pages
,
1268 const void __user
* __user
*pages
,
1269 const int __user
*nodes
,
1270 int __user
*status
, int flags
)
1272 struct page_to_node
*pm
;
1273 unsigned long chunk_nr_pages
;
1274 unsigned long chunk_start
;
1278 pm
= (struct page_to_node
*)__get_free_page(GFP_KERNEL
);
1285 * Store a chunk of page_to_node array in a page,
1286 * but keep the last one as a marker
1288 chunk_nr_pages
= (PAGE_SIZE
/ sizeof(struct page_to_node
)) - 1;
1290 for (chunk_start
= 0;
1291 chunk_start
< nr_pages
;
1292 chunk_start
+= chunk_nr_pages
) {
1295 if (chunk_start
+ chunk_nr_pages
> nr_pages
)
1296 chunk_nr_pages
= nr_pages
- chunk_start
;
1298 /* fill the chunk pm with addrs and nodes from user-space */
1299 for (j
= 0; j
< chunk_nr_pages
; j
++) {
1300 const void __user
*p
;
1304 if (get_user(p
, pages
+ j
+ chunk_start
))
1306 pm
[j
].addr
= (unsigned long) p
;
1308 if (get_user(node
, nodes
+ j
+ chunk_start
))
1312 if (node
< 0 || node
>= MAX_NUMNODES
)
1315 if (!node_state(node
, N_MEMORY
))
1319 if (!node_isset(node
, task_nodes
))
1325 /* End marker for this chunk */
1326 pm
[chunk_nr_pages
].node
= MAX_NUMNODES
;
1328 /* Migrate this chunk */
1329 err
= do_move_page_to_node_array(mm
, pm
,
1330 flags
& MPOL_MF_MOVE_ALL
);
1334 /* Return status information */
1335 for (j
= 0; j
< chunk_nr_pages
; j
++)
1336 if (put_user(pm
[j
].status
, status
+ j
+ chunk_start
)) {
1344 free_page((unsigned long)pm
);
1350 * Determine the nodes of an array of pages and store it in an array of status.
1352 static void do_pages_stat_array(struct mm_struct
*mm
, unsigned long nr_pages
,
1353 const void __user
**pages
, int *status
)
1357 down_read(&mm
->mmap_sem
);
1359 for (i
= 0; i
< nr_pages
; i
++) {
1360 unsigned long addr
= (unsigned long)(*pages
);
1361 struct vm_area_struct
*vma
;
1365 vma
= find_vma(mm
, addr
);
1366 if (!vma
|| addr
< vma
->vm_start
)
1369 page
= follow_page(vma
, addr
, 0);
1371 err
= PTR_ERR(page
);
1376 /* Use PageReserved to check for zero page */
1377 if (!page
|| PageReserved(page
))
1380 err
= page_to_nid(page
);
1388 up_read(&mm
->mmap_sem
);
1392 * Determine the nodes of a user array of pages and store it in
1393 * a user array of status.
1395 static int do_pages_stat(struct mm_struct
*mm
, unsigned long nr_pages
,
1396 const void __user
* __user
*pages
,
1399 #define DO_PAGES_STAT_CHUNK_NR 16
1400 const void __user
*chunk_pages
[DO_PAGES_STAT_CHUNK_NR
];
1401 int chunk_status
[DO_PAGES_STAT_CHUNK_NR
];
1404 unsigned long chunk_nr
;
1406 chunk_nr
= nr_pages
;
1407 if (chunk_nr
> DO_PAGES_STAT_CHUNK_NR
)
1408 chunk_nr
= DO_PAGES_STAT_CHUNK_NR
;
1410 if (copy_from_user(chunk_pages
, pages
, chunk_nr
* sizeof(*chunk_pages
)))
1413 do_pages_stat_array(mm
, chunk_nr
, chunk_pages
, chunk_status
);
1415 if (copy_to_user(status
, chunk_status
, chunk_nr
* sizeof(*status
)))
1420 nr_pages
-= chunk_nr
;
1422 return nr_pages
? -EFAULT
: 0;
1426 * Move a list of pages in the address space of the currently executing
1429 SYSCALL_DEFINE6(move_pages
, pid_t
, pid
, unsigned long, nr_pages
,
1430 const void __user
* __user
*, pages
,
1431 const int __user
*, nodes
,
1432 int __user
*, status
, int, flags
)
1434 const struct cred
*cred
= current_cred(), *tcred
;
1435 struct task_struct
*task
;
1436 struct mm_struct
*mm
;
1438 nodemask_t task_nodes
;
1441 if (flags
& ~(MPOL_MF_MOVE
|MPOL_MF_MOVE_ALL
))
1444 if ((flags
& MPOL_MF_MOVE_ALL
) && !capable(CAP_SYS_NICE
))
1447 /* Find the mm_struct */
1449 task
= pid
? find_task_by_vpid(pid
) : current
;
1454 get_task_struct(task
);
1457 * Check if this process has the right to modify the specified
1458 * process. The right exists if the process has administrative
1459 * capabilities, superuser privileges or the same
1460 * userid as the target process.
1462 tcred
= __task_cred(task
);
1463 if (!uid_eq(cred
->euid
, tcred
->suid
) && !uid_eq(cred
->euid
, tcred
->uid
) &&
1464 !uid_eq(cred
->uid
, tcred
->suid
) && !uid_eq(cred
->uid
, tcred
->uid
) &&
1465 !capable(CAP_SYS_NICE
)) {
1472 err
= security_task_movememory(task
);
1476 task_nodes
= cpuset_mems_allowed(task
);
1477 mm
= get_task_mm(task
);
1478 put_task_struct(task
);
1484 err
= do_pages_move(mm
, task_nodes
, nr_pages
, pages
,
1485 nodes
, status
, flags
);
1487 err
= do_pages_stat(mm
, nr_pages
, pages
, status
);
1493 put_task_struct(task
);
1498 * Call migration functions in the vma_ops that may prepare
1499 * memory in a vm for migration. migration functions may perform
1500 * the migration for vmas that do not have an underlying page struct.
1502 int migrate_vmas(struct mm_struct
*mm
, const nodemask_t
*to
,
1503 const nodemask_t
*from
, unsigned long flags
)
1505 struct vm_area_struct
*vma
;
1508 for (vma
= mm
->mmap
; vma
&& !err
; vma
= vma
->vm_next
) {
1509 if (vma
->vm_ops
&& vma
->vm_ops
->migrate
) {
1510 err
= vma
->vm_ops
->migrate(vma
, to
, from
, flags
);
1518 #ifdef CONFIG_NUMA_BALANCING
1520 * Returns true if this is a safe migration target node for misplaced NUMA
1521 * pages. Currently it only checks the watermarks which crude
1523 static bool migrate_balanced_pgdat(struct pglist_data
*pgdat
,
1524 unsigned long nr_migrate_pages
)
1527 for (z
= pgdat
->nr_zones
- 1; z
>= 0; z
--) {
1528 struct zone
*zone
= pgdat
->node_zones
+ z
;
1530 if (!populated_zone(zone
))
1533 if (!zone_reclaimable(zone
))
1536 /* Avoid waking kswapd by allocating pages_to_migrate pages. */
1537 if (!zone_watermark_ok(zone
, 0,
1538 high_wmark_pages(zone
) +
1547 static struct page
*alloc_misplaced_dst_page(struct page
*page
,
1551 int nid
= (int) data
;
1552 struct page
*newpage
;
1554 newpage
= alloc_pages_exact_node(nid
,
1555 (GFP_HIGHUSER_MOVABLE
| GFP_THISNODE
|
1556 __GFP_NOMEMALLOC
| __GFP_NORETRY
|
1560 page_cpupid_xchg_last(newpage
, page_cpupid_last(page
));
1566 * page migration rate limiting control.
1567 * Do not migrate more than @pages_to_migrate in a @migrate_interval_millisecs
1568 * window of time. Default here says do not migrate more than 1280M per second.
1569 * If a node is rate-limited then PTE NUMA updates are also rate-limited. However
1570 * as it is faults that reset the window, pte updates will happen unconditionally
1571 * if there has not been a fault since @pteupdate_interval_millisecs after the
1572 * throttle window closed.
1574 static unsigned int migrate_interval_millisecs __read_mostly
= 100;
1575 static unsigned int pteupdate_interval_millisecs __read_mostly
= 1000;
1576 static unsigned int ratelimit_pages __read_mostly
= 128 << (20 - PAGE_SHIFT
);
1578 /* Returns true if NUMA migration is currently rate limited */
1579 bool migrate_ratelimited(int node
)
1581 pg_data_t
*pgdat
= NODE_DATA(node
);
1583 if (time_after(jiffies
, pgdat
->numabalancing_migrate_next_window
+
1584 msecs_to_jiffies(pteupdate_interval_millisecs
)))
1587 if (pgdat
->numabalancing_migrate_nr_pages
< ratelimit_pages
)
1593 /* Returns true if the node is migrate rate-limited after the update */
1594 static bool numamigrate_update_ratelimit(pg_data_t
*pgdat
,
1595 unsigned long nr_pages
)
1598 * Rate-limit the amount of data that is being migrated to a node.
1599 * Optimal placement is no good if the memory bus is saturated and
1600 * all the time is being spent migrating!
1602 if (time_after(jiffies
, pgdat
->numabalancing_migrate_next_window
)) {
1603 spin_lock(&pgdat
->numabalancing_migrate_lock
);
1604 pgdat
->numabalancing_migrate_nr_pages
= 0;
1605 pgdat
->numabalancing_migrate_next_window
= jiffies
+
1606 msecs_to_jiffies(migrate_interval_millisecs
);
1607 spin_unlock(&pgdat
->numabalancing_migrate_lock
);
1609 if (pgdat
->numabalancing_migrate_nr_pages
> ratelimit_pages
) {
1610 trace_mm_numa_migrate_ratelimit(current
, pgdat
->node_id
,
1616 * This is an unlocked non-atomic update so errors are possible.
1617 * The consequences are failing to migrate when we potentiall should
1618 * have which is not severe enough to warrant locking. If it is ever
1619 * a problem, it can be converted to a per-cpu counter.
1621 pgdat
->numabalancing_migrate_nr_pages
+= nr_pages
;
1625 static int numamigrate_isolate_page(pg_data_t
*pgdat
, struct page
*page
)
1629 VM_BUG_ON(compound_order(page
) && !PageTransHuge(page
));
1631 /* Avoid migrating to a node that is nearly full */
1632 if (!migrate_balanced_pgdat(pgdat
, 1UL << compound_order(page
)))
1635 if (isolate_lru_page(page
))
1639 * migrate_misplaced_transhuge_page() skips page migration's usual
1640 * check on page_count(), so we must do it here, now that the page
1641 * has been isolated: a GUP pin, or any other pin, prevents migration.
1642 * The expected page count is 3: 1 for page's mapcount and 1 for the
1643 * caller's pin and 1 for the reference taken by isolate_lru_page().
1645 if (PageTransHuge(page
) && page_count(page
) != 3) {
1646 putback_lru_page(page
);
1650 page_lru
= page_is_file_cache(page
);
1651 mod_zone_page_state(page_zone(page
), NR_ISOLATED_ANON
+ page_lru
,
1652 hpage_nr_pages(page
));
1655 * Isolating the page has taken another reference, so the
1656 * caller's reference can be safely dropped without the page
1657 * disappearing underneath us during migration.
1663 bool pmd_trans_migrating(pmd_t pmd
)
1665 struct page
*page
= pmd_page(pmd
);
1666 return PageLocked(page
);
1669 void wait_migrate_huge_page(struct anon_vma
*anon_vma
, pmd_t
*pmd
)
1671 struct page
*page
= pmd_page(*pmd
);
1672 wait_on_page_locked(page
);
1676 * Attempt to migrate a misplaced page to the specified destination
1677 * node. Caller is expected to have an elevated reference count on
1678 * the page that will be dropped by this function before returning.
1680 int migrate_misplaced_page(struct page
*page
, struct vm_area_struct
*vma
,
1683 pg_data_t
*pgdat
= NODE_DATA(node
);
1686 LIST_HEAD(migratepages
);
1689 * Don't migrate file pages that are mapped in multiple processes
1690 * with execute permissions as they are probably shared libraries.
1692 if (page_mapcount(page
) != 1 && page_is_file_cache(page
) &&
1693 (vma
->vm_flags
& VM_EXEC
))
1697 * Rate-limit the amount of data that is being migrated to a node.
1698 * Optimal placement is no good if the memory bus is saturated and
1699 * all the time is being spent migrating!
1701 if (numamigrate_update_ratelimit(pgdat
, 1))
1704 isolated
= numamigrate_isolate_page(pgdat
, page
);
1708 list_add(&page
->lru
, &migratepages
);
1709 nr_remaining
= migrate_pages(&migratepages
, alloc_misplaced_dst_page
,
1710 node
, MIGRATE_ASYNC
, MR_NUMA_MISPLACED
);
1712 if (!list_empty(&migratepages
)) {
1713 list_del(&page
->lru
);
1714 dec_zone_page_state(page
, NR_ISOLATED_ANON
+
1715 page_is_file_cache(page
));
1716 putback_lru_page(page
);
1720 count_vm_numa_event(NUMA_PAGE_MIGRATE
);
1721 BUG_ON(!list_empty(&migratepages
));
1728 #endif /* CONFIG_NUMA_BALANCING */
1730 #if defined(CONFIG_NUMA_BALANCING) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
1732 * Migrates a THP to a given target node. page must be locked and is unlocked
1735 int migrate_misplaced_transhuge_page(struct mm_struct
*mm
,
1736 struct vm_area_struct
*vma
,
1737 pmd_t
*pmd
, pmd_t entry
,
1738 unsigned long address
,
1739 struct page
*page
, int node
)
1742 pg_data_t
*pgdat
= NODE_DATA(node
);
1744 struct page
*new_page
= NULL
;
1745 struct mem_cgroup
*memcg
= NULL
;
1746 int page_lru
= page_is_file_cache(page
);
1747 unsigned long mmun_start
= address
& HPAGE_PMD_MASK
;
1748 unsigned long mmun_end
= mmun_start
+ HPAGE_PMD_SIZE
;
1752 * Rate-limit the amount of data that is being migrated to a node.
1753 * Optimal placement is no good if the memory bus is saturated and
1754 * all the time is being spent migrating!
1756 if (numamigrate_update_ratelimit(pgdat
, HPAGE_PMD_NR
))
1759 new_page
= alloc_pages_node(node
,
1760 (GFP_TRANSHUGE
| GFP_THISNODE
) & ~__GFP_WAIT
, HPAGE_PMD_ORDER
);
1764 page_cpupid_xchg_last(new_page
, page_cpupid_last(page
));
1766 isolated
= numamigrate_isolate_page(pgdat
, page
);
1772 if (mm_tlb_flush_pending(mm
))
1773 flush_tlb_range(vma
, mmun_start
, mmun_end
);
1775 /* Prepare a page as a migration target */
1776 __set_page_locked(new_page
);
1777 SetPageSwapBacked(new_page
);
1779 /* anon mapping, we can simply copy page->mapping to the new page: */
1780 new_page
->mapping
= page
->mapping
;
1781 new_page
->index
= page
->index
;
1782 migrate_page_copy(new_page
, page
);
1783 WARN_ON(PageLRU(new_page
));
1785 /* Recheck the target PMD */
1786 mmu_notifier_invalidate_range_start(mm
, mmun_start
, mmun_end
);
1787 ptl
= pmd_lock(mm
, pmd
);
1788 if (unlikely(!pmd_same(*pmd
, entry
) || page_count(page
) != 2)) {
1791 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
1793 /* Reverse changes made by migrate_page_copy() */
1794 if (TestClearPageActive(new_page
))
1795 SetPageActive(page
);
1796 if (TestClearPageUnevictable(new_page
))
1797 SetPageUnevictable(page
);
1798 mlock_migrate_page(page
, new_page
);
1800 unlock_page(new_page
);
1801 put_page(new_page
); /* Free it */
1803 /* Retake the callers reference and putback on LRU */
1805 putback_lru_page(page
);
1806 mod_zone_page_state(page_zone(page
),
1807 NR_ISOLATED_ANON
+ page_lru
, -HPAGE_PMD_NR
);
1813 * Traditional migration needs to prepare the memcg charge
1814 * transaction early to prevent the old page from being
1815 * uncharged when installing migration entries. Here we can
1816 * save the potential rollback and start the charge transfer
1817 * only when migration is already known to end successfully.
1819 mem_cgroup_prepare_migration(page
, new_page
, &memcg
);
1822 entry
= mk_pmd(new_page
, vma
->vm_page_prot
);
1823 entry
= pmd_mkhuge(entry
);
1824 entry
= maybe_pmd_mkwrite(pmd_mkdirty(entry
), vma
);
1827 * Clear the old entry under pagetable lock and establish the new PTE.
1828 * Any parallel GUP will either observe the old page blocking on the
1829 * page lock, block on the page table lock or observe the new page.
1830 * The SetPageUptodate on the new page and page_add_new_anon_rmap
1831 * guarantee the copy is visible before the pagetable update.
1833 flush_cache_range(vma
, mmun_start
, mmun_end
);
1834 page_add_new_anon_rmap(new_page
, vma
, mmun_start
);
1835 pmdp_clear_flush(vma
, mmun_start
, pmd
);
1836 set_pmd_at(mm
, mmun_start
, pmd
, entry
);
1837 flush_tlb_range(vma
, mmun_start
, mmun_end
);
1838 update_mmu_cache_pmd(vma
, address
, &entry
);
1840 if (page_count(page
) != 2) {
1841 set_pmd_at(mm
, mmun_start
, pmd
, orig_entry
);
1842 flush_tlb_range(vma
, mmun_start
, mmun_end
);
1843 update_mmu_cache_pmd(vma
, address
, &entry
);
1844 page_remove_rmap(new_page
);
1848 page_remove_rmap(page
);
1851 * Finish the charge transaction under the page table lock to
1852 * prevent split_huge_page() from dividing up the charge
1853 * before it's fully transferred to the new page.
1855 mem_cgroup_end_migration(memcg
, page
, new_page
, true);
1857 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
1859 unlock_page(new_page
);
1861 put_page(page
); /* Drop the rmap reference */
1862 put_page(page
); /* Drop the LRU isolation reference */
1864 count_vm_events(PGMIGRATE_SUCCESS
, HPAGE_PMD_NR
);
1865 count_vm_numa_events(NUMA_PAGE_MIGRATE
, HPAGE_PMD_NR
);
1867 mod_zone_page_state(page_zone(page
),
1868 NR_ISOLATED_ANON
+ page_lru
,
1873 count_vm_events(PGMIGRATE_FAIL
, HPAGE_PMD_NR
);
1875 ptl
= pmd_lock(mm
, pmd
);
1876 if (pmd_same(*pmd
, entry
)) {
1877 entry
= pmd_mknonnuma(entry
);
1878 set_pmd_at(mm
, mmun_start
, pmd
, entry
);
1879 update_mmu_cache_pmd(vma
, address
, &entry
);
1888 #endif /* CONFIG_NUMA_BALANCING */
1890 #endif /* CONFIG_NUMA */