2 * Memory Migration functionality - linux/mm/migrate.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/syscalls.h>
34 #include <linux/hugetlb.h>
35 #include <linux/hugetlb_cgroup.h>
36 #include <linux/gfp.h>
37 #include <linux/balloon_compaction.h>
38 #include <linux/mmu_notifier.h>
39 #include <linux/page_idle.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
);
124 ptep
= pte_offset_map(pmd
, addr
);
127 * Peek to check is_swap_pte() before taking ptlock? No, we
128 * can race mremap's move_ptes(), which skips anon_vma lock.
131 ptl
= pte_lockptr(mm
, pmd
);
136 if (!is_swap_pte(pte
))
139 entry
= pte_to_swp_entry(pte
);
141 if (!is_migration_entry(entry
) ||
142 migration_entry_to_page(entry
) != old
)
146 pte
= pte_mkold(mk_pte(new, vma
->vm_page_prot
));
147 if (pte_swp_soft_dirty(*ptep
))
148 pte
= pte_mksoft_dirty(pte
);
150 /* Recheck VMA as permissions can change since migration started */
151 if (is_write_migration_entry(entry
))
152 pte
= maybe_mkwrite(pte
, vma
);
154 #ifdef CONFIG_HUGETLB_PAGE
156 pte
= pte_mkhuge(pte
);
157 pte
= arch_make_huge_pte(pte
, vma
, new, 0);
160 flush_dcache_page(new);
161 set_pte_at(mm
, addr
, ptep
, pte
);
165 hugepage_add_anon_rmap(new, vma
, addr
);
168 } else if (PageAnon(new))
169 page_add_anon_rmap(new, vma
, addr
);
171 page_add_file_rmap(new);
173 if (vma
->vm_flags
& VM_LOCKED
)
176 /* No need to invalidate - it was non-present before */
177 update_mmu_cache(vma
, addr
, ptep
);
179 pte_unmap_unlock(ptep
, ptl
);
185 * Get rid of all migration entries and replace them by
186 * references to the indicated page.
188 static void remove_migration_ptes(struct page
*old
, struct page
*new)
190 struct rmap_walk_control rwc
= {
191 .rmap_one
= remove_migration_pte
,
195 rmap_walk(new, &rwc
);
199 * Something used the pte of a page under migration. We need to
200 * get to the page and wait until migration is finished.
201 * When we return from this function the fault will be retried.
203 void __migration_entry_wait(struct mm_struct
*mm
, pte_t
*ptep
,
212 if (!is_swap_pte(pte
))
215 entry
= pte_to_swp_entry(pte
);
216 if (!is_migration_entry(entry
))
219 page
= migration_entry_to_page(entry
);
222 * Once radix-tree replacement of page migration started, page_count
223 * *must* be zero. And, we don't want to call wait_on_page_locked()
224 * against a page without get_page().
225 * So, we use get_page_unless_zero(), here. Even failed, page fault
228 if (!get_page_unless_zero(page
))
230 pte_unmap_unlock(ptep
, ptl
);
231 wait_on_page_locked(page
);
235 pte_unmap_unlock(ptep
, ptl
);
238 void migration_entry_wait(struct mm_struct
*mm
, pmd_t
*pmd
,
239 unsigned long address
)
241 spinlock_t
*ptl
= pte_lockptr(mm
, pmd
);
242 pte_t
*ptep
= pte_offset_map(pmd
, address
);
243 __migration_entry_wait(mm
, ptep
, ptl
);
246 void migration_entry_wait_huge(struct vm_area_struct
*vma
,
247 struct mm_struct
*mm
, pte_t
*pte
)
249 spinlock_t
*ptl
= huge_pte_lockptr(hstate_vma(vma
), mm
, pte
);
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 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
,
316 int expected_count
= 1 + extra_count
;
320 /* Anonymous page without mapping */
321 if (page_count(page
) != expected_count
)
323 return MIGRATEPAGE_SUCCESS
;
326 spin_lock_irq(&mapping
->tree_lock
);
328 pslot
= radix_tree_lookup_slot(&mapping
->page_tree
,
331 expected_count
+= 1 + page_has_private(page
);
332 if (page_count(page
) != expected_count
||
333 radix_tree_deref_slot_protected(pslot
, &mapping
->tree_lock
) != page
) {
334 spin_unlock_irq(&mapping
->tree_lock
);
338 if (!page_freeze_refs(page
, expected_count
)) {
339 spin_unlock_irq(&mapping
->tree_lock
);
344 * In the async migration case of moving a page with buffers, lock the
345 * buffers using trylock before the mapping is moved. If the mapping
346 * was moved, we later failed to lock the buffers and could not move
347 * the mapping back due to an elevated page count, we would have to
348 * block waiting on other references to be dropped.
350 if (mode
== MIGRATE_ASYNC
&& head
&&
351 !buffer_migrate_lock_buffers(head
, mode
)) {
352 page_unfreeze_refs(page
, expected_count
);
353 spin_unlock_irq(&mapping
->tree_lock
);
358 * Now we know that no one else is looking at the page.
360 get_page(newpage
); /* add cache reference */
361 if (PageSwapCache(page
)) {
362 SetPageSwapCache(newpage
);
363 set_page_private(newpage
, page_private(page
));
366 radix_tree_replace_slot(pslot
, newpage
);
369 * Drop cache reference from old page by unfreezing
370 * to one less reference.
371 * We know this isn't the last reference.
373 page_unfreeze_refs(page
, expected_count
- 1);
376 * If moved to a different zone then also account
377 * the page for that zone. Other VM counters will be
378 * taken care of when we establish references to the
379 * new page and drop references to the old page.
381 * Note that anonymous pages are accounted for
382 * via NR_FILE_PAGES and NR_ANON_PAGES if they
383 * are mapped to swap space.
385 __dec_zone_page_state(page
, NR_FILE_PAGES
);
386 __inc_zone_page_state(newpage
, NR_FILE_PAGES
);
387 if (!PageSwapCache(page
) && PageSwapBacked(page
)) {
388 __dec_zone_page_state(page
, NR_SHMEM
);
389 __inc_zone_page_state(newpage
, NR_SHMEM
);
391 spin_unlock_irq(&mapping
->tree_lock
);
393 return MIGRATEPAGE_SUCCESS
;
397 * The expected number of remaining references is the same as that
398 * of migrate_page_move_mapping().
400 int migrate_huge_page_move_mapping(struct address_space
*mapping
,
401 struct page
*newpage
, struct page
*page
)
407 if (page_count(page
) != 1)
409 return MIGRATEPAGE_SUCCESS
;
412 spin_lock_irq(&mapping
->tree_lock
);
414 pslot
= radix_tree_lookup_slot(&mapping
->page_tree
,
417 expected_count
= 2 + page_has_private(page
);
418 if (page_count(page
) != expected_count
||
419 radix_tree_deref_slot_protected(pslot
, &mapping
->tree_lock
) != page
) {
420 spin_unlock_irq(&mapping
->tree_lock
);
424 if (!page_freeze_refs(page
, expected_count
)) {
425 spin_unlock_irq(&mapping
->tree_lock
);
431 radix_tree_replace_slot(pslot
, newpage
);
433 page_unfreeze_refs(page
, expected_count
- 1);
435 spin_unlock_irq(&mapping
->tree_lock
);
436 return MIGRATEPAGE_SUCCESS
;
440 * Gigantic pages are so large that we do not guarantee that page++ pointer
441 * arithmetic will work across the entire page. We need something more
444 static void __copy_gigantic_page(struct page
*dst
, struct page
*src
,
448 struct page
*dst_base
= dst
;
449 struct page
*src_base
= src
;
451 for (i
= 0; i
< nr_pages
; ) {
453 copy_highpage(dst
, src
);
456 dst
= mem_map_next(dst
, dst_base
, i
);
457 src
= mem_map_next(src
, src_base
, i
);
461 static void copy_huge_page(struct page
*dst
, struct page
*src
)
468 struct hstate
*h
= page_hstate(src
);
469 nr_pages
= pages_per_huge_page(h
);
471 if (unlikely(nr_pages
> MAX_ORDER_NR_PAGES
)) {
472 __copy_gigantic_page(dst
, src
, nr_pages
);
477 BUG_ON(!PageTransHuge(src
));
478 nr_pages
= hpage_nr_pages(src
);
481 for (i
= 0; i
< nr_pages
; i
++) {
483 copy_highpage(dst
+ i
, src
+ i
);
488 * Copy the page to its new location
490 void migrate_page_copy(struct page
*newpage
, struct page
*page
)
494 if (PageHuge(page
) || PageTransHuge(page
))
495 copy_huge_page(newpage
, page
);
497 copy_highpage(newpage
, page
);
500 SetPageError(newpage
);
501 if (PageReferenced(page
))
502 SetPageReferenced(newpage
);
503 if (PageUptodate(page
))
504 SetPageUptodate(newpage
);
505 if (TestClearPageActive(page
)) {
506 VM_BUG_ON_PAGE(PageUnevictable(page
), page
);
507 SetPageActive(newpage
);
508 } else if (TestClearPageUnevictable(page
))
509 SetPageUnevictable(newpage
);
510 if (PageChecked(page
))
511 SetPageChecked(newpage
);
512 if (PageMappedToDisk(page
))
513 SetPageMappedToDisk(newpage
);
515 if (PageDirty(page
)) {
516 clear_page_dirty_for_io(page
);
518 * Want to mark the page and the radix tree as dirty, and
519 * redo the accounting that clear_page_dirty_for_io undid,
520 * but we can't use set_page_dirty because that function
521 * is actually a signal that all of the page has become dirty.
522 * Whereas only part of our page may be dirty.
524 if (PageSwapBacked(page
))
525 SetPageDirty(newpage
);
527 __set_page_dirty_nobuffers(newpage
);
530 if (page_is_young(page
))
531 set_page_young(newpage
);
532 if (page_is_idle(page
))
533 set_page_idle(newpage
);
536 * Copy NUMA information to the new page, to prevent over-eager
537 * future migrations of this same page.
539 cpupid
= page_cpupid_xchg_last(page
, -1);
540 page_cpupid_xchg_last(newpage
, cpupid
);
542 ksm_migrate_page(newpage
, page
);
544 * Please do not reorder this without considering how mm/ksm.c's
545 * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache().
547 if (PageSwapCache(page
))
548 ClearPageSwapCache(page
);
549 ClearPagePrivate(page
);
550 set_page_private(page
, 0);
553 * If any waiters have accumulated on the new page then
556 if (PageWriteback(newpage
))
557 end_page_writeback(newpage
);
560 /************************************************************
561 * Migration functions
562 ***********************************************************/
565 * Common logic to directly migrate a single page suitable for
566 * pages that do not use PagePrivate/PagePrivate2.
568 * Pages are locked upon entry and exit.
570 int migrate_page(struct address_space
*mapping
,
571 struct page
*newpage
, struct page
*page
,
572 enum migrate_mode mode
)
576 BUG_ON(PageWriteback(page
)); /* Writeback must be complete */
578 rc
= migrate_page_move_mapping(mapping
, newpage
, page
, NULL
, mode
, 0);
580 if (rc
!= MIGRATEPAGE_SUCCESS
)
583 migrate_page_copy(newpage
, page
);
584 return MIGRATEPAGE_SUCCESS
;
586 EXPORT_SYMBOL(migrate_page
);
590 * Migration function for pages with buffers. This function can only be used
591 * if the underlying filesystem guarantees that no other references to "page"
594 int buffer_migrate_page(struct address_space
*mapping
,
595 struct page
*newpage
, struct page
*page
, enum migrate_mode mode
)
597 struct buffer_head
*bh
, *head
;
600 if (!page_has_buffers(page
))
601 return migrate_page(mapping
, newpage
, page
, mode
);
603 head
= page_buffers(page
);
605 rc
= migrate_page_move_mapping(mapping
, newpage
, page
, head
, mode
, 0);
607 if (rc
!= MIGRATEPAGE_SUCCESS
)
611 * In the async case, migrate_page_move_mapping locked the buffers
612 * with an IRQ-safe spinlock held. In the sync case, the buffers
613 * need to be locked now
615 if (mode
!= MIGRATE_ASYNC
)
616 BUG_ON(!buffer_migrate_lock_buffers(head
, mode
));
618 ClearPagePrivate(page
);
619 set_page_private(newpage
, page_private(page
));
620 set_page_private(page
, 0);
626 set_bh_page(bh
, newpage
, bh_offset(bh
));
627 bh
= bh
->b_this_page
;
629 } while (bh
!= head
);
631 SetPagePrivate(newpage
);
633 migrate_page_copy(newpage
, page
);
639 bh
= bh
->b_this_page
;
641 } while (bh
!= head
);
643 return MIGRATEPAGE_SUCCESS
;
645 EXPORT_SYMBOL(buffer_migrate_page
);
649 * Writeback a page to clean the dirty state
651 static int writeout(struct address_space
*mapping
, struct page
*page
)
653 struct writeback_control wbc
= {
654 .sync_mode
= WB_SYNC_NONE
,
657 .range_end
= LLONG_MAX
,
662 if (!mapping
->a_ops
->writepage
)
663 /* No write method for the address space */
666 if (!clear_page_dirty_for_io(page
))
667 /* Someone else already triggered a write */
671 * A dirty page may imply that the underlying filesystem has
672 * the page on some queue. So the page must be clean for
673 * migration. Writeout may mean we loose the lock and the
674 * page state is no longer what we checked for earlier.
675 * At this point we know that the migration attempt cannot
678 remove_migration_ptes(page
, page
);
680 rc
= mapping
->a_ops
->writepage(page
, &wbc
);
682 if (rc
!= AOP_WRITEPAGE_ACTIVATE
)
683 /* unlocked. Relock */
686 return (rc
< 0) ? -EIO
: -EAGAIN
;
690 * Default handling if a filesystem does not provide a migration function.
692 static int fallback_migrate_page(struct address_space
*mapping
,
693 struct page
*newpage
, struct page
*page
, enum migrate_mode mode
)
695 if (PageDirty(page
)) {
696 /* Only writeback pages in full synchronous migration */
697 if (mode
!= MIGRATE_SYNC
)
699 return writeout(mapping
, page
);
703 * Buffers may be managed in a filesystem specific way.
704 * We must have no buffers or drop them.
706 if (page_has_private(page
) &&
707 !try_to_release_page(page
, GFP_KERNEL
))
710 return migrate_page(mapping
, newpage
, page
, mode
);
714 * Move a page to a newly allocated page
715 * The page is locked and all ptes have been successfully removed.
717 * The new page will have replaced the old page if this function
722 * MIGRATEPAGE_SUCCESS - success
724 static int move_to_new_page(struct page
*newpage
, struct page
*page
,
725 enum migrate_mode mode
)
727 struct address_space
*mapping
;
730 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
731 VM_BUG_ON_PAGE(!PageLocked(newpage
), newpage
);
733 /* Prepare mapping for the new page.*/
734 newpage
->index
= page
->index
;
735 newpage
->mapping
= page
->mapping
;
736 if (PageSwapBacked(page
))
737 SetPageSwapBacked(newpage
);
740 * Indirectly called below, migrate_page_copy() copies PG_dirty and thus
741 * needs newpage's memcg set to transfer memcg dirty page accounting.
742 * So perform memcg migration in two steps:
743 * 1. set newpage->mem_cgroup (here)
744 * 2. clear page->mem_cgroup (below)
746 set_page_memcg(newpage
, page_memcg(page
));
748 mapping
= page_mapping(page
);
750 rc
= migrate_page(mapping
, newpage
, page
, mode
);
751 else if (mapping
->a_ops
->migratepage
)
753 * Most pages have a mapping and most filesystems provide a
754 * migratepage callback. Anonymous pages are part of swap
755 * space which also has its own migratepage callback. This
756 * is the most common path for page migration.
758 rc
= mapping
->a_ops
->migratepage(mapping
, newpage
, page
, mode
);
760 rc
= fallback_migrate_page(mapping
, newpage
, page
, mode
);
763 * When successful, old pagecache page->mapping must be cleared before
764 * page is freed; but stats require that PageAnon be left as PageAnon.
766 if (rc
== MIGRATEPAGE_SUCCESS
) {
767 set_page_memcg(page
, NULL
);
769 page
->mapping
= NULL
;
771 set_page_memcg(newpage
, NULL
);
772 newpage
->mapping
= NULL
;
777 static int __unmap_and_move(struct page
*page
, struct page
*newpage
,
778 int force
, enum migrate_mode mode
)
781 int page_was_mapped
= 0;
782 struct anon_vma
*anon_vma
= NULL
;
784 if (!trylock_page(page
)) {
785 if (!force
|| mode
== MIGRATE_ASYNC
)
789 * It's not safe for direct compaction to call lock_page.
790 * For example, during page readahead pages are added locked
791 * to the LRU. Later, when the IO completes the pages are
792 * marked uptodate and unlocked. However, the queueing
793 * could be merging multiple pages for one bio (e.g.
794 * mpage_readpages). If an allocation happens for the
795 * second or third page, the process can end up locking
796 * the same page twice and deadlocking. Rather than
797 * trying to be clever about what pages can be locked,
798 * avoid the use of lock_page for direct compaction
801 if (current
->flags
& PF_MEMALLOC
)
807 if (PageWriteback(page
)) {
809 * Only in the case of a full synchronous migration is it
810 * necessary to wait for PageWriteback. In the async case,
811 * the retry loop is too short and in the sync-light case,
812 * the overhead of stalling is too much
814 if (mode
!= MIGRATE_SYNC
) {
820 wait_on_page_writeback(page
);
824 * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
825 * we cannot notice that anon_vma is freed while we migrates a page.
826 * This get_anon_vma() delays freeing anon_vma pointer until the end
827 * of migration. File cache pages are no problem because of page_lock()
828 * File Caches may use write_page() or lock_page() in migration, then,
829 * just care Anon page here.
831 * Only page_get_anon_vma() understands the subtleties of
832 * getting a hold on an anon_vma from outside one of its mms.
833 * But if we cannot get anon_vma, then we won't need it anyway,
834 * because that implies that the anon page is no longer mapped
835 * (and cannot be remapped so long as we hold the page lock).
837 if (PageAnon(page
) && !PageKsm(page
))
838 anon_vma
= page_get_anon_vma(page
);
841 * Block others from accessing the new page when we get around to
842 * establishing additional references. We are usually the only one
843 * holding a reference to newpage at this point. We used to have a BUG
844 * here if trylock_page(newpage) fails, but would like to allow for
845 * cases where there might be a race with the previous use of newpage.
846 * This is much like races on refcount of oldpage: just don't BUG().
848 if (unlikely(!trylock_page(newpage
)))
851 if (unlikely(isolated_balloon_page(page
))) {
853 * A ballooned page does not need any special attention from
854 * physical to virtual reverse mapping procedures.
855 * Skip any attempt to unmap PTEs or to remap swap cache,
856 * in order to avoid burning cycles at rmap level, and perform
857 * the page migration right away (proteced by page lock).
859 rc
= balloon_page_migrate(newpage
, page
, mode
);
860 goto out_unlock_both
;
864 * Corner case handling:
865 * 1. When a new swap-cache page is read into, it is added to the LRU
866 * and treated as swapcache but it has no rmap yet.
867 * Calling try_to_unmap() against a page->mapping==NULL page will
868 * trigger a BUG. So handle it here.
869 * 2. An orphaned page (see truncate_complete_page) might have
870 * fs-private metadata. The page can be picked up due to memory
871 * offlining. Everywhere else except page reclaim, the page is
872 * invisible to the vm, so the page can not be migrated. So try to
873 * free the metadata, so the page can be freed.
875 if (!page
->mapping
) {
876 VM_BUG_ON_PAGE(PageAnon(page
), page
);
877 if (page_has_private(page
)) {
878 try_to_free_buffers(page
);
879 goto out_unlock_both
;
881 } else if (page_mapped(page
)) {
882 /* Establish migration ptes */
883 VM_BUG_ON_PAGE(PageAnon(page
) && !PageKsm(page
) && !anon_vma
,
886 TTU_MIGRATION
|TTU_IGNORE_MLOCK
|TTU_IGNORE_ACCESS
);
890 if (!page_mapped(page
))
891 rc
= move_to_new_page(newpage
, page
, mode
);
894 remove_migration_ptes(page
,
895 rc
== MIGRATEPAGE_SUCCESS
? newpage
: page
);
898 unlock_page(newpage
);
900 /* Drop an anon_vma reference if we took one */
902 put_anon_vma(anon_vma
);
909 * gcc 4.7 and 4.8 on arm get an ICEs when inlining unmap_and_move(). Work
912 #if (GCC_VERSION >= 40700 && GCC_VERSION < 40900) && defined(CONFIG_ARM)
913 #define ICE_noinline noinline
919 * Obtain the lock on page, remove all ptes and migrate the page
920 * to the newly allocated page in newpage.
922 static ICE_noinline
int unmap_and_move(new_page_t get_new_page
,
923 free_page_t put_new_page
,
924 unsigned long private, struct page
*page
,
925 int force
, enum migrate_mode mode
,
926 enum migrate_reason reason
)
928 int rc
= MIGRATEPAGE_SUCCESS
;
930 struct page
*newpage
;
932 newpage
= get_new_page(page
, private, &result
);
936 if (page_count(page
) == 1) {
937 /* page was freed from under us. So we are done. */
941 if (unlikely(PageTransHuge(page
)))
942 if (unlikely(split_huge_page(page
)))
945 rc
= __unmap_and_move(page
, newpage
, force
, mode
);
946 if (rc
== 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 /* Soft-offlined page shouldn't go through lru cache list */
961 if (reason
== MR_MEMORY_FAILURE
) {
963 if (!test_set_page_hwpoison(page
))
964 num_poisoned_pages_inc();
966 putback_lru_page(page
);
970 * If migration was not successful and there's a freeing callback, use
971 * it. Otherwise, putback_lru_page() will drop the reference grabbed
975 ClearPageSwapBacked(newpage
);
976 put_new_page(newpage
, private);
977 } else if (unlikely(__is_movable_balloon_page(newpage
))) {
978 /* drop our reference, page already in the balloon */
981 putback_lru_page(newpage
);
987 *result
= page_to_nid(newpage
);
993 * Counterpart of unmap_and_move_page() for hugepage migration.
995 * This function doesn't wait the completion of hugepage I/O
996 * because there is no race between I/O and migration for hugepage.
997 * Note that currently hugepage I/O occurs only in direct I/O
998 * where no lock is held and PG_writeback is irrelevant,
999 * and writeback status of all subpages are counted in the reference
1000 * count of the head page (i.e. if all subpages of a 2MB hugepage are
1001 * under direct I/O, the reference of the head page is 512 and a bit more.)
1002 * This means that when we try to migrate hugepage whose subpages are
1003 * doing direct I/O, some references remain after try_to_unmap() and
1004 * hugepage migration fails without data corruption.
1006 * There is also no race when direct I/O is issued on the page under migration,
1007 * because then pte is replaced with migration swap entry and direct I/O code
1008 * will wait in the page fault for migration to complete.
1010 static int unmap_and_move_huge_page(new_page_t get_new_page
,
1011 free_page_t put_new_page
, unsigned long private,
1012 struct page
*hpage
, int force
,
1013 enum migrate_mode mode
)
1017 int page_was_mapped
= 0;
1018 struct page
*new_hpage
;
1019 struct anon_vma
*anon_vma
= NULL
;
1022 * Movability of hugepages depends on architectures and hugepage size.
1023 * This check is necessary because some callers of hugepage migration
1024 * like soft offline and memory hotremove don't walk through page
1025 * tables or check whether the hugepage is pmd-based or not before
1026 * kicking migration.
1028 if (!hugepage_migration_supported(page_hstate(hpage
))) {
1029 putback_active_hugepage(hpage
);
1033 new_hpage
= get_new_page(hpage
, private, &result
);
1037 if (!trylock_page(hpage
)) {
1038 if (!force
|| mode
!= MIGRATE_SYNC
)
1043 if (PageAnon(hpage
))
1044 anon_vma
= page_get_anon_vma(hpage
);
1046 if (unlikely(!trylock_page(new_hpage
)))
1049 if (page_mapped(hpage
)) {
1051 TTU_MIGRATION
|TTU_IGNORE_MLOCK
|TTU_IGNORE_ACCESS
);
1052 page_was_mapped
= 1;
1055 if (!page_mapped(hpage
))
1056 rc
= move_to_new_page(new_hpage
, hpage
, mode
);
1058 if (page_was_mapped
)
1059 remove_migration_ptes(hpage
,
1060 rc
== MIGRATEPAGE_SUCCESS
? new_hpage
: hpage
);
1062 unlock_page(new_hpage
);
1066 put_anon_vma(anon_vma
);
1068 if (rc
== MIGRATEPAGE_SUCCESS
) {
1069 hugetlb_cgroup_migrate(hpage
, new_hpage
);
1070 put_new_page
= NULL
;
1076 putback_active_hugepage(hpage
);
1079 * If migration was not successful and there's a freeing callback, use
1080 * it. Otherwise, put_page() will drop the reference grabbed during
1084 put_new_page(new_hpage
, private);
1086 putback_active_hugepage(new_hpage
);
1092 *result
= page_to_nid(new_hpage
);
1098 * migrate_pages - migrate the pages specified in a list, to the free pages
1099 * supplied as the target for the page migration
1101 * @from: The list of pages to be migrated.
1102 * @get_new_page: The function used to allocate free pages to be used
1103 * as the target of the page migration.
1104 * @put_new_page: The function used to free target pages if migration
1105 * fails, or NULL if no special handling is necessary.
1106 * @private: Private data to be passed on to get_new_page()
1107 * @mode: The migration mode that specifies the constraints for
1108 * page migration, if any.
1109 * @reason: The reason for page migration.
1111 * The function returns after 10 attempts or if no pages are movable any more
1112 * because the list has become empty or no retryable pages exist any more.
1113 * The caller should call putback_movable_pages() to return pages to the LRU
1114 * or free list only if ret != 0.
1116 * Returns the number of pages that were not migrated, or an error code.
1118 int migrate_pages(struct list_head
*from
, new_page_t get_new_page
,
1119 free_page_t put_new_page
, unsigned long private,
1120 enum migrate_mode mode
, int reason
)
1124 int nr_succeeded
= 0;
1128 int swapwrite
= current
->flags
& PF_SWAPWRITE
;
1132 current
->flags
|= PF_SWAPWRITE
;
1134 for(pass
= 0; pass
< 10 && retry
; pass
++) {
1137 list_for_each_entry_safe(page
, page2
, from
, lru
) {
1141 rc
= unmap_and_move_huge_page(get_new_page
,
1142 put_new_page
, private, page
,
1145 rc
= unmap_and_move(get_new_page
, put_new_page
,
1146 private, page
, pass
> 2, mode
,
1155 case MIGRATEPAGE_SUCCESS
:
1160 * Permanent failure (-EBUSY, -ENOSYS, etc.):
1161 * unlike -EAGAIN case, the failed page is
1162 * removed from migration page list and not
1163 * retried in the next outer loop.
1174 count_vm_events(PGMIGRATE_SUCCESS
, nr_succeeded
);
1176 count_vm_events(PGMIGRATE_FAIL
, nr_failed
);
1177 trace_mm_migrate_pages(nr_succeeded
, nr_failed
, mode
, reason
);
1180 current
->flags
&= ~PF_SWAPWRITE
;
1187 * Move a list of individual pages
1189 struct page_to_node
{
1196 static struct page
*new_page_node(struct page
*p
, unsigned long private,
1199 struct page_to_node
*pm
= (struct page_to_node
*)private;
1201 while (pm
->node
!= MAX_NUMNODES
&& pm
->page
!= p
)
1204 if (pm
->node
== MAX_NUMNODES
)
1207 *result
= &pm
->status
;
1210 return alloc_huge_page_node(page_hstate(compound_head(p
)),
1213 return __alloc_pages_node(pm
->node
,
1214 GFP_HIGHUSER_MOVABLE
| __GFP_THISNODE
, 0);
1218 * Move a set of pages as indicated in the pm array. The addr
1219 * field must be set to the virtual address of the page to be moved
1220 * and the node number must contain a valid target node.
1221 * The pm array ends with node = MAX_NUMNODES.
1223 static int do_move_page_to_node_array(struct mm_struct
*mm
,
1224 struct page_to_node
*pm
,
1228 struct page_to_node
*pp
;
1229 LIST_HEAD(pagelist
);
1231 down_read(&mm
->mmap_sem
);
1234 * Build a list of pages to migrate
1236 for (pp
= pm
; pp
->node
!= MAX_NUMNODES
; pp
++) {
1237 struct vm_area_struct
*vma
;
1241 vma
= find_vma(mm
, pp
->addr
);
1242 if (!vma
|| pp
->addr
< vma
->vm_start
|| !vma_migratable(vma
))
1245 /* FOLL_DUMP to ignore special (like zero) pages */
1246 page
= follow_page(vma
, pp
->addr
,
1247 FOLL_GET
| FOLL_SPLIT
| FOLL_DUMP
);
1249 err
= PTR_ERR(page
);
1258 err
= page_to_nid(page
);
1260 if (err
== pp
->node
)
1262 * Node already in the right place
1267 if (page_mapcount(page
) > 1 &&
1271 if (PageHuge(page
)) {
1273 isolate_huge_page(page
, &pagelist
);
1277 err
= isolate_lru_page(page
);
1279 list_add_tail(&page
->lru
, &pagelist
);
1280 inc_zone_page_state(page
, NR_ISOLATED_ANON
+
1281 page_is_file_cache(page
));
1285 * Either remove the duplicate refcount from
1286 * isolate_lru_page() or drop the page ref if it was
1295 if (!list_empty(&pagelist
)) {
1296 err
= migrate_pages(&pagelist
, new_page_node
, NULL
,
1297 (unsigned long)pm
, MIGRATE_SYNC
, MR_SYSCALL
);
1299 putback_movable_pages(&pagelist
);
1302 up_read(&mm
->mmap_sem
);
1307 * Migrate an array of page address onto an array of nodes and fill
1308 * the corresponding array of status.
1310 static int do_pages_move(struct mm_struct
*mm
, nodemask_t task_nodes
,
1311 unsigned long nr_pages
,
1312 const void __user
* __user
*pages
,
1313 const int __user
*nodes
,
1314 int __user
*status
, int flags
)
1316 struct page_to_node
*pm
;
1317 unsigned long chunk_nr_pages
;
1318 unsigned long chunk_start
;
1322 pm
= (struct page_to_node
*)__get_free_page(GFP_KERNEL
);
1329 * Store a chunk of page_to_node array in a page,
1330 * but keep the last one as a marker
1332 chunk_nr_pages
= (PAGE_SIZE
/ sizeof(struct page_to_node
)) - 1;
1334 for (chunk_start
= 0;
1335 chunk_start
< nr_pages
;
1336 chunk_start
+= chunk_nr_pages
) {
1339 if (chunk_start
+ chunk_nr_pages
> nr_pages
)
1340 chunk_nr_pages
= nr_pages
- chunk_start
;
1342 /* fill the chunk pm with addrs and nodes from user-space */
1343 for (j
= 0; j
< chunk_nr_pages
; j
++) {
1344 const void __user
*p
;
1348 if (get_user(p
, pages
+ j
+ chunk_start
))
1350 pm
[j
].addr
= (unsigned long) p
;
1352 if (get_user(node
, nodes
+ j
+ chunk_start
))
1356 if (node
< 0 || node
>= MAX_NUMNODES
)
1359 if (!node_state(node
, N_MEMORY
))
1363 if (!node_isset(node
, task_nodes
))
1369 /* End marker for this chunk */
1370 pm
[chunk_nr_pages
].node
= MAX_NUMNODES
;
1372 /* Migrate this chunk */
1373 err
= do_move_page_to_node_array(mm
, pm
,
1374 flags
& MPOL_MF_MOVE_ALL
);
1378 /* Return status information */
1379 for (j
= 0; j
< chunk_nr_pages
; j
++)
1380 if (put_user(pm
[j
].status
, status
+ j
+ chunk_start
)) {
1388 free_page((unsigned long)pm
);
1394 * Determine the nodes of an array of pages and store it in an array of status.
1396 static void do_pages_stat_array(struct mm_struct
*mm
, unsigned long nr_pages
,
1397 const void __user
**pages
, int *status
)
1401 down_read(&mm
->mmap_sem
);
1403 for (i
= 0; i
< nr_pages
; i
++) {
1404 unsigned long addr
= (unsigned long)(*pages
);
1405 struct vm_area_struct
*vma
;
1409 vma
= find_vma(mm
, addr
);
1410 if (!vma
|| addr
< vma
->vm_start
)
1413 /* FOLL_DUMP to ignore special (like zero) pages */
1414 page
= follow_page(vma
, addr
, FOLL_DUMP
);
1416 err
= PTR_ERR(page
);
1420 err
= page
? page_to_nid(page
) : -ENOENT
;
1428 up_read(&mm
->mmap_sem
);
1432 * Determine the nodes of a user array of pages and store it in
1433 * a user array of status.
1435 static int do_pages_stat(struct mm_struct
*mm
, unsigned long nr_pages
,
1436 const void __user
* __user
*pages
,
1439 #define DO_PAGES_STAT_CHUNK_NR 16
1440 const void __user
*chunk_pages
[DO_PAGES_STAT_CHUNK_NR
];
1441 int chunk_status
[DO_PAGES_STAT_CHUNK_NR
];
1444 unsigned long chunk_nr
;
1446 chunk_nr
= nr_pages
;
1447 if (chunk_nr
> DO_PAGES_STAT_CHUNK_NR
)
1448 chunk_nr
= DO_PAGES_STAT_CHUNK_NR
;
1450 if (copy_from_user(chunk_pages
, pages
, chunk_nr
* sizeof(*chunk_pages
)))
1453 do_pages_stat_array(mm
, chunk_nr
, chunk_pages
, chunk_status
);
1455 if (copy_to_user(status
, chunk_status
, chunk_nr
* sizeof(*status
)))
1460 nr_pages
-= chunk_nr
;
1462 return nr_pages
? -EFAULT
: 0;
1466 * Move a list of pages in the address space of the currently executing
1469 SYSCALL_DEFINE6(move_pages
, pid_t
, pid
, unsigned long, nr_pages
,
1470 const void __user
* __user
*, pages
,
1471 const int __user
*, nodes
,
1472 int __user
*, status
, int, flags
)
1474 const struct cred
*cred
= current_cred(), *tcred
;
1475 struct task_struct
*task
;
1476 struct mm_struct
*mm
;
1478 nodemask_t task_nodes
;
1481 if (flags
& ~(MPOL_MF_MOVE
|MPOL_MF_MOVE_ALL
))
1484 if ((flags
& MPOL_MF_MOVE_ALL
) && !capable(CAP_SYS_NICE
))
1487 /* Find the mm_struct */
1489 task
= pid
? find_task_by_vpid(pid
) : current
;
1494 get_task_struct(task
);
1497 * Check if this process has the right to modify the specified
1498 * process. The right exists if the process has administrative
1499 * capabilities, superuser privileges or the same
1500 * userid as the target process.
1502 tcred
= __task_cred(task
);
1503 if (!uid_eq(cred
->euid
, tcred
->suid
) && !uid_eq(cred
->euid
, tcred
->uid
) &&
1504 !uid_eq(cred
->uid
, tcred
->suid
) && !uid_eq(cred
->uid
, tcred
->uid
) &&
1505 !capable(CAP_SYS_NICE
)) {
1512 err
= security_task_movememory(task
);
1516 task_nodes
= cpuset_mems_allowed(task
);
1517 mm
= get_task_mm(task
);
1518 put_task_struct(task
);
1524 err
= do_pages_move(mm
, task_nodes
, nr_pages
, pages
,
1525 nodes
, status
, flags
);
1527 err
= do_pages_stat(mm
, nr_pages
, pages
, status
);
1533 put_task_struct(task
);
1537 #ifdef CONFIG_NUMA_BALANCING
1539 * Returns true if this is a safe migration target node for misplaced NUMA
1540 * pages. Currently it only checks the watermarks which crude
1542 static bool migrate_balanced_pgdat(struct pglist_data
*pgdat
,
1543 unsigned long nr_migrate_pages
)
1546 for (z
= pgdat
->nr_zones
- 1; z
>= 0; z
--) {
1547 struct zone
*zone
= pgdat
->node_zones
+ z
;
1549 if (!populated_zone(zone
))
1552 if (!zone_reclaimable(zone
))
1555 /* Avoid waking kswapd by allocating pages_to_migrate pages. */
1556 if (!zone_watermark_ok(zone
, 0,
1557 high_wmark_pages(zone
) +
1566 static struct page
*alloc_misplaced_dst_page(struct page
*page
,
1570 int nid
= (int) data
;
1571 struct page
*newpage
;
1573 newpage
= __alloc_pages_node(nid
,
1574 (GFP_HIGHUSER_MOVABLE
|
1575 __GFP_THISNODE
| __GFP_NOMEMALLOC
|
1576 __GFP_NORETRY
| __GFP_NOWARN
) &
1583 * page migration rate limiting control.
1584 * Do not migrate more than @pages_to_migrate in a @migrate_interval_millisecs
1585 * window of time. Default here says do not migrate more than 1280M per second.
1587 static unsigned int migrate_interval_millisecs __read_mostly
= 100;
1588 static unsigned int ratelimit_pages __read_mostly
= 128 << (20 - PAGE_SHIFT
);
1590 /* Returns true if the node is migrate rate-limited after the update */
1591 static bool numamigrate_update_ratelimit(pg_data_t
*pgdat
,
1592 unsigned long nr_pages
)
1595 * Rate-limit the amount of data that is being migrated to a node.
1596 * Optimal placement is no good if the memory bus is saturated and
1597 * all the time is being spent migrating!
1599 if (time_after(jiffies
, pgdat
->numabalancing_migrate_next_window
)) {
1600 spin_lock(&pgdat
->numabalancing_migrate_lock
);
1601 pgdat
->numabalancing_migrate_nr_pages
= 0;
1602 pgdat
->numabalancing_migrate_next_window
= jiffies
+
1603 msecs_to_jiffies(migrate_interval_millisecs
);
1604 spin_unlock(&pgdat
->numabalancing_migrate_lock
);
1606 if (pgdat
->numabalancing_migrate_nr_pages
> ratelimit_pages
) {
1607 trace_mm_numa_migrate_ratelimit(current
, pgdat
->node_id
,
1613 * This is an unlocked non-atomic update so errors are possible.
1614 * The consequences are failing to migrate when we potentiall should
1615 * have which is not severe enough to warrant locking. If it is ever
1616 * a problem, it can be converted to a per-cpu counter.
1618 pgdat
->numabalancing_migrate_nr_pages
+= nr_pages
;
1622 static int numamigrate_isolate_page(pg_data_t
*pgdat
, struct page
*page
)
1626 VM_BUG_ON_PAGE(compound_order(page
) && !PageTransHuge(page
), page
);
1628 /* Avoid migrating to a node that is nearly full */
1629 if (!migrate_balanced_pgdat(pgdat
, 1UL << compound_order(page
)))
1632 if (isolate_lru_page(page
))
1636 * migrate_misplaced_transhuge_page() skips page migration's usual
1637 * check on page_count(), so we must do it here, now that the page
1638 * has been isolated: a GUP pin, or any other pin, prevents migration.
1639 * The expected page count is 3: 1 for page's mapcount and 1 for the
1640 * caller's pin and 1 for the reference taken by isolate_lru_page().
1642 if (PageTransHuge(page
) && page_count(page
) != 3) {
1643 putback_lru_page(page
);
1647 page_lru
= page_is_file_cache(page
);
1648 mod_zone_page_state(page_zone(page
), NR_ISOLATED_ANON
+ page_lru
,
1649 hpage_nr_pages(page
));
1652 * Isolating the page has taken another reference, so the
1653 * caller's reference can be safely dropped without the page
1654 * disappearing underneath us during migration.
1660 bool pmd_trans_migrating(pmd_t pmd
)
1662 struct page
*page
= pmd_page(pmd
);
1663 return PageLocked(page
);
1667 * Attempt to migrate a misplaced page to the specified destination
1668 * node. Caller is expected to have an elevated reference count on
1669 * the page that will be dropped by this function before returning.
1671 int migrate_misplaced_page(struct page
*page
, struct vm_area_struct
*vma
,
1674 pg_data_t
*pgdat
= NODE_DATA(node
);
1677 LIST_HEAD(migratepages
);
1680 * Don't migrate file pages that are mapped in multiple processes
1681 * with execute permissions as they are probably shared libraries.
1683 if (page_mapcount(page
) != 1 && page_is_file_cache(page
) &&
1684 (vma
->vm_flags
& VM_EXEC
))
1688 * Rate-limit the amount of data that is being migrated to a node.
1689 * Optimal placement is no good if the memory bus is saturated and
1690 * all the time is being spent migrating!
1692 if (numamigrate_update_ratelimit(pgdat
, 1))
1695 isolated
= numamigrate_isolate_page(pgdat
, page
);
1699 list_add(&page
->lru
, &migratepages
);
1700 nr_remaining
= migrate_pages(&migratepages
, alloc_misplaced_dst_page
,
1701 NULL
, node
, MIGRATE_ASYNC
,
1704 if (!list_empty(&migratepages
)) {
1705 list_del(&page
->lru
);
1706 dec_zone_page_state(page
, NR_ISOLATED_ANON
+
1707 page_is_file_cache(page
));
1708 putback_lru_page(page
);
1712 count_vm_numa_event(NUMA_PAGE_MIGRATE
);
1713 BUG_ON(!list_empty(&migratepages
));
1720 #endif /* CONFIG_NUMA_BALANCING */
1722 #if defined(CONFIG_NUMA_BALANCING) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
1724 * Migrates a THP to a given target node. page must be locked and is unlocked
1727 int migrate_misplaced_transhuge_page(struct mm_struct
*mm
,
1728 struct vm_area_struct
*vma
,
1729 pmd_t
*pmd
, pmd_t entry
,
1730 unsigned long address
,
1731 struct page
*page
, int node
)
1734 pg_data_t
*pgdat
= NODE_DATA(node
);
1736 struct page
*new_page
= NULL
;
1737 int page_lru
= page_is_file_cache(page
);
1738 unsigned long mmun_start
= address
& HPAGE_PMD_MASK
;
1739 unsigned long mmun_end
= mmun_start
+ HPAGE_PMD_SIZE
;
1743 * Rate-limit the amount of data that is being migrated to a node.
1744 * Optimal placement is no good if the memory bus is saturated and
1745 * all the time is being spent migrating!
1747 if (numamigrate_update_ratelimit(pgdat
, HPAGE_PMD_NR
))
1750 new_page
= alloc_pages_node(node
,
1751 (GFP_TRANSHUGE
| __GFP_THISNODE
) & ~__GFP_WAIT
,
1756 isolated
= numamigrate_isolate_page(pgdat
, page
);
1762 if (mm_tlb_flush_pending(mm
))
1763 flush_tlb_range(vma
, mmun_start
, mmun_end
);
1765 /* Prepare a page as a migration target */
1766 __set_page_locked(new_page
);
1767 SetPageSwapBacked(new_page
);
1769 /* anon mapping, we can simply copy page->mapping to the new page: */
1770 new_page
->mapping
= page
->mapping
;
1771 new_page
->index
= page
->index
;
1772 migrate_page_copy(new_page
, page
);
1773 WARN_ON(PageLRU(new_page
));
1775 /* Recheck the target PMD */
1776 mmu_notifier_invalidate_range_start(mm
, mmun_start
, mmun_end
);
1777 ptl
= pmd_lock(mm
, pmd
);
1778 if (unlikely(!pmd_same(*pmd
, entry
) || page_count(page
) != 2)) {
1781 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
1783 /* Reverse changes made by migrate_page_copy() */
1784 if (TestClearPageActive(new_page
))
1785 SetPageActive(page
);
1786 if (TestClearPageUnevictable(new_page
))
1787 SetPageUnevictable(page
);
1789 unlock_page(new_page
);
1790 put_page(new_page
); /* Free it */
1792 /* Retake the callers reference and putback on LRU */
1794 putback_lru_page(page
);
1795 mod_zone_page_state(page_zone(page
),
1796 NR_ISOLATED_ANON
+ page_lru
, -HPAGE_PMD_NR
);
1802 entry
= mk_pmd(new_page
, vma
->vm_page_prot
);
1803 entry
= pmd_mkhuge(entry
);
1804 entry
= maybe_pmd_mkwrite(pmd_mkdirty(entry
), vma
);
1807 * Clear the old entry under pagetable lock and establish the new PTE.
1808 * Any parallel GUP will either observe the old page blocking on the
1809 * page lock, block on the page table lock or observe the new page.
1810 * The SetPageUptodate on the new page and page_add_new_anon_rmap
1811 * guarantee the copy is visible before the pagetable update.
1813 flush_cache_range(vma
, mmun_start
, mmun_end
);
1814 page_add_anon_rmap(new_page
, vma
, mmun_start
);
1815 pmdp_huge_clear_flush_notify(vma
, mmun_start
, pmd
);
1816 set_pmd_at(mm
, mmun_start
, pmd
, entry
);
1817 flush_tlb_range(vma
, mmun_start
, mmun_end
);
1818 update_mmu_cache_pmd(vma
, address
, &entry
);
1820 if (page_count(page
) != 2) {
1821 set_pmd_at(mm
, mmun_start
, pmd
, orig_entry
);
1822 flush_tlb_range(vma
, mmun_start
, mmun_end
);
1823 mmu_notifier_invalidate_range(mm
, mmun_start
, mmun_end
);
1824 update_mmu_cache_pmd(vma
, address
, &entry
);
1825 page_remove_rmap(new_page
);
1829 mlock_migrate_page(new_page
, page
);
1830 set_page_memcg(new_page
, page_memcg(page
));
1831 set_page_memcg(page
, NULL
);
1832 page_remove_rmap(page
);
1835 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
1837 /* Take an "isolate" reference and put new page on the LRU. */
1839 putback_lru_page(new_page
);
1841 unlock_page(new_page
);
1843 put_page(page
); /* Drop the rmap reference */
1844 put_page(page
); /* Drop the LRU isolation reference */
1846 count_vm_events(PGMIGRATE_SUCCESS
, HPAGE_PMD_NR
);
1847 count_vm_numa_events(NUMA_PAGE_MIGRATE
, HPAGE_PMD_NR
);
1849 mod_zone_page_state(page_zone(page
),
1850 NR_ISOLATED_ANON
+ page_lru
,
1855 count_vm_events(PGMIGRATE_FAIL
, HPAGE_PMD_NR
);
1857 ptl
= pmd_lock(mm
, pmd
);
1858 if (pmd_same(*pmd
, entry
)) {
1859 entry
= pmd_modify(entry
, vma
->vm_page_prot
);
1860 set_pmd_at(mm
, mmun_start
, pmd
, entry
);
1861 update_mmu_cache_pmd(vma
, address
, &entry
);
1870 #endif /* CONFIG_NUMA_BALANCING */
1872 #endif /* CONFIG_NUMA */