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/backing-dev.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>
40 #include <linux/page_idle.h>
41 #include <linux/page_owner.h>
43 #include <asm/tlbflush.h>
45 #define CREATE_TRACE_POINTS
46 #include <trace/events/migrate.h>
51 * migrate_prep() needs to be called before we start compiling a list of pages
52 * to be migrated using isolate_lru_page(). If scheduling work on other CPUs is
53 * undesirable, use migrate_prep_local()
55 int migrate_prep(void)
58 * Clear the LRU lists so pages can be isolated.
59 * Note that pages may be moved off the LRU after we have
60 * drained them. Those pages will fail to migrate like other
61 * pages that may be busy.
68 /* Do the necessary work of migrate_prep but not if it involves other CPUs */
69 int migrate_prep_local(void)
77 * Put previously isolated pages back onto the appropriate lists
78 * from where they were once taken off for compaction/migration.
80 * This function shall be used whenever the isolated pageset has been
81 * built from lru, balloon, hugetlbfs page. See isolate_migratepages_range()
82 * and isolate_huge_page().
84 void putback_movable_pages(struct list_head
*l
)
89 list_for_each_entry_safe(page
, page2
, l
, lru
) {
90 if (unlikely(PageHuge(page
))) {
91 putback_active_hugepage(page
);
95 dec_zone_page_state(page
, NR_ISOLATED_ANON
+
96 page_is_file_cache(page
));
97 if (unlikely(isolated_balloon_page(page
)))
98 balloon_page_putback(page
);
100 putback_lru_page(page
);
105 * Restore a potential migration pte to a working pte entry
107 static int remove_migration_pte(struct page
*new, struct vm_area_struct
*vma
,
108 unsigned long addr
, void *old
)
110 struct mm_struct
*mm
= vma
->vm_mm
;
116 if (unlikely(PageHuge(new))) {
117 ptep
= huge_pte_offset(mm
, addr
);
120 ptl
= huge_pte_lockptr(hstate_vma(vma
), mm
, ptep
);
122 pmd
= mm_find_pmd(mm
, addr
);
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
);
152 /* Recheck VMA as permissions can change since migration started */
153 if (is_write_migration_entry(entry
))
154 pte
= maybe_mkwrite(pte
, vma
);
156 #ifdef CONFIG_HUGETLB_PAGE
158 pte
= pte_mkhuge(pte
);
159 pte
= arch_make_huge_pte(pte
, vma
, new, 0);
162 flush_dcache_page(new);
163 set_pte_at(mm
, addr
, ptep
, pte
);
167 hugepage_add_anon_rmap(new, vma
, addr
);
169 page_dup_rmap(new, true);
170 } else if (PageAnon(new))
171 page_add_anon_rmap(new, vma
, addr
, false);
173 page_add_file_rmap(new);
175 if (vma
->vm_flags
& VM_LOCKED
&& !PageTransCompound(new))
178 /* No need to invalidate - it was non-present before */
179 update_mmu_cache(vma
, addr
, ptep
);
181 pte_unmap_unlock(ptep
, ptl
);
187 * Get rid of all migration entries and replace them by
188 * references to the indicated page.
190 void remove_migration_ptes(struct page
*old
, struct page
*new, bool locked
)
192 struct rmap_walk_control rwc
= {
193 .rmap_one
= remove_migration_pte
,
198 rmap_walk_locked(new, &rwc
);
200 rmap_walk(new, &rwc
);
204 * Something used the pte of a page under migration. We need to
205 * get to the page and wait until migration is finished.
206 * When we return from this function the fault will be retried.
208 void __migration_entry_wait(struct mm_struct
*mm
, pte_t
*ptep
,
217 if (!is_swap_pte(pte
))
220 entry
= pte_to_swp_entry(pte
);
221 if (!is_migration_entry(entry
))
224 page
= migration_entry_to_page(entry
);
227 * Once radix-tree replacement of page migration started, page_count
228 * *must* be zero. And, we don't want to call wait_on_page_locked()
229 * against a page without get_page().
230 * So, we use get_page_unless_zero(), here. Even failed, page fault
233 if (!get_page_unless_zero(page
))
235 pte_unmap_unlock(ptep
, ptl
);
236 wait_on_page_locked(page
);
240 pte_unmap_unlock(ptep
, ptl
);
243 void migration_entry_wait(struct mm_struct
*mm
, pmd_t
*pmd
,
244 unsigned long address
)
246 spinlock_t
*ptl
= pte_lockptr(mm
, pmd
);
247 pte_t
*ptep
= pte_offset_map(pmd
, address
);
248 __migration_entry_wait(mm
, ptep
, ptl
);
251 void migration_entry_wait_huge(struct vm_area_struct
*vma
,
252 struct mm_struct
*mm
, pte_t
*pte
)
254 spinlock_t
*ptl
= huge_pte_lockptr(hstate_vma(vma
), mm
, pte
);
255 __migration_entry_wait(mm
, pte
, ptl
);
259 /* Returns true if all buffers are successfully locked */
260 static bool buffer_migrate_lock_buffers(struct buffer_head
*head
,
261 enum migrate_mode mode
)
263 struct buffer_head
*bh
= head
;
265 /* Simple case, sync compaction */
266 if (mode
!= MIGRATE_ASYNC
) {
270 bh
= bh
->b_this_page
;
272 } while (bh
!= head
);
277 /* async case, we cannot block on lock_buffer so use trylock_buffer */
280 if (!trylock_buffer(bh
)) {
282 * We failed to lock the buffer and cannot stall in
283 * async migration. Release the taken locks
285 struct buffer_head
*failed_bh
= bh
;
288 while (bh
!= failed_bh
) {
291 bh
= bh
->b_this_page
;
296 bh
= bh
->b_this_page
;
297 } while (bh
!= head
);
301 static inline bool buffer_migrate_lock_buffers(struct buffer_head
*head
,
302 enum migrate_mode mode
)
306 #endif /* CONFIG_BLOCK */
309 * Replace the page in the mapping.
311 * The number of remaining references must be:
312 * 1 for anonymous pages without a mapping
313 * 2 for pages with a mapping
314 * 3 for pages with a mapping and PagePrivate/PagePrivate2 set.
316 int migrate_page_move_mapping(struct address_space
*mapping
,
317 struct page
*newpage
, struct page
*page
,
318 struct buffer_head
*head
, enum migrate_mode mode
,
321 struct zone
*oldzone
, *newzone
;
323 int expected_count
= 1 + extra_count
;
327 /* Anonymous page without mapping */
328 if (page_count(page
) != expected_count
)
331 /* No turning back from here */
332 newpage
->index
= page
->index
;
333 newpage
->mapping
= page
->mapping
;
334 if (PageSwapBacked(page
))
335 __SetPageSwapBacked(newpage
);
337 return MIGRATEPAGE_SUCCESS
;
340 oldzone
= page_zone(page
);
341 newzone
= page_zone(newpage
);
343 spin_lock_irq(&mapping
->tree_lock
);
345 pslot
= radix_tree_lookup_slot(&mapping
->page_tree
,
348 expected_count
+= 1 + page_has_private(page
);
349 if (page_count(page
) != expected_count
||
350 radix_tree_deref_slot_protected(pslot
, &mapping
->tree_lock
) != page
) {
351 spin_unlock_irq(&mapping
->tree_lock
);
355 if (!page_ref_freeze(page
, expected_count
)) {
356 spin_unlock_irq(&mapping
->tree_lock
);
361 * In the async migration case of moving a page with buffers, lock the
362 * buffers using trylock before the mapping is moved. If the mapping
363 * was moved, we later failed to lock the buffers and could not move
364 * the mapping back due to an elevated page count, we would have to
365 * block waiting on other references to be dropped.
367 if (mode
== MIGRATE_ASYNC
&& head
&&
368 !buffer_migrate_lock_buffers(head
, mode
)) {
369 page_ref_unfreeze(page
, expected_count
);
370 spin_unlock_irq(&mapping
->tree_lock
);
375 * Now we know that no one else is looking at the page:
376 * no turning back from here.
378 newpage
->index
= page
->index
;
379 newpage
->mapping
= page
->mapping
;
380 if (PageSwapBacked(page
))
381 __SetPageSwapBacked(newpage
);
383 get_page(newpage
); /* add cache reference */
384 if (PageSwapCache(page
)) {
385 SetPageSwapCache(newpage
);
386 set_page_private(newpage
, page_private(page
));
389 /* Move dirty while page refs frozen and newpage not yet exposed */
390 dirty
= PageDirty(page
);
392 ClearPageDirty(page
);
393 SetPageDirty(newpage
);
396 radix_tree_replace_slot(pslot
, newpage
);
399 * Drop cache reference from old page by unfreezing
400 * to one less reference.
401 * We know this isn't the last reference.
403 page_ref_unfreeze(page
, expected_count
- 1);
405 spin_unlock(&mapping
->tree_lock
);
406 /* Leave irq disabled to prevent preemption while updating stats */
409 * If moved to a different zone then also account
410 * the page for that zone. Other VM counters will be
411 * taken care of when we establish references to the
412 * new page and drop references to the old page.
414 * Note that anonymous pages are accounted for
415 * via NR_FILE_PAGES and NR_ANON_PAGES if they
416 * are mapped to swap space.
418 if (newzone
!= oldzone
) {
419 __dec_zone_state(oldzone
, NR_FILE_PAGES
);
420 __inc_zone_state(newzone
, NR_FILE_PAGES
);
421 if (PageSwapBacked(page
) && !PageSwapCache(page
)) {
422 __dec_zone_state(oldzone
, NR_SHMEM
);
423 __inc_zone_state(newzone
, NR_SHMEM
);
425 if (dirty
&& mapping_cap_account_dirty(mapping
)) {
426 __dec_zone_state(oldzone
, NR_FILE_DIRTY
);
427 __inc_zone_state(newzone
, NR_FILE_DIRTY
);
432 return MIGRATEPAGE_SUCCESS
;
434 EXPORT_SYMBOL(migrate_page_move_mapping
);
437 * The expected number of remaining references is the same as that
438 * of migrate_page_move_mapping().
440 int migrate_huge_page_move_mapping(struct address_space
*mapping
,
441 struct page
*newpage
, struct page
*page
)
446 spin_lock_irq(&mapping
->tree_lock
);
448 pslot
= radix_tree_lookup_slot(&mapping
->page_tree
,
451 expected_count
= 2 + page_has_private(page
);
452 if (page_count(page
) != expected_count
||
453 radix_tree_deref_slot_protected(pslot
, &mapping
->tree_lock
) != page
) {
454 spin_unlock_irq(&mapping
->tree_lock
);
458 if (!page_ref_freeze(page
, expected_count
)) {
459 spin_unlock_irq(&mapping
->tree_lock
);
463 newpage
->index
= page
->index
;
464 newpage
->mapping
= page
->mapping
;
468 radix_tree_replace_slot(pslot
, newpage
);
470 page_ref_unfreeze(page
, expected_count
- 1);
472 spin_unlock_irq(&mapping
->tree_lock
);
474 return MIGRATEPAGE_SUCCESS
;
478 * Gigantic pages are so large that we do not guarantee that page++ pointer
479 * arithmetic will work across the entire page. We need something more
482 static void __copy_gigantic_page(struct page
*dst
, struct page
*src
,
486 struct page
*dst_base
= dst
;
487 struct page
*src_base
= src
;
489 for (i
= 0; i
< nr_pages
; ) {
491 copy_highpage(dst
, src
);
494 dst
= mem_map_next(dst
, dst_base
, i
);
495 src
= mem_map_next(src
, src_base
, i
);
499 static void copy_huge_page(struct page
*dst
, struct page
*src
)
506 struct hstate
*h
= page_hstate(src
);
507 nr_pages
= pages_per_huge_page(h
);
509 if (unlikely(nr_pages
> MAX_ORDER_NR_PAGES
)) {
510 __copy_gigantic_page(dst
, src
, nr_pages
);
515 BUG_ON(!PageTransHuge(src
));
516 nr_pages
= hpage_nr_pages(src
);
519 for (i
= 0; i
< nr_pages
; i
++) {
521 copy_highpage(dst
+ i
, src
+ i
);
526 * Copy the page to its new location
528 void migrate_page_copy(struct page
*newpage
, struct page
*page
)
532 if (PageHuge(page
) || PageTransHuge(page
))
533 copy_huge_page(newpage
, page
);
535 copy_highpage(newpage
, page
);
538 SetPageError(newpage
);
539 if (PageReferenced(page
))
540 SetPageReferenced(newpage
);
541 if (PageUptodate(page
))
542 SetPageUptodate(newpage
);
543 if (TestClearPageActive(page
)) {
544 VM_BUG_ON_PAGE(PageUnevictable(page
), page
);
545 SetPageActive(newpage
);
546 } else if (TestClearPageUnevictable(page
))
547 SetPageUnevictable(newpage
);
548 if (PageChecked(page
))
549 SetPageChecked(newpage
);
550 if (PageMappedToDisk(page
))
551 SetPageMappedToDisk(newpage
);
553 /* Move dirty on pages not done by migrate_page_move_mapping() */
555 SetPageDirty(newpage
);
557 if (page_is_young(page
))
558 set_page_young(newpage
);
559 if (page_is_idle(page
))
560 set_page_idle(newpage
);
563 * Copy NUMA information to the new page, to prevent over-eager
564 * future migrations of this same page.
566 cpupid
= page_cpupid_xchg_last(page
, -1);
567 page_cpupid_xchg_last(newpage
, cpupid
);
569 ksm_migrate_page(newpage
, page
);
571 * Please do not reorder this without considering how mm/ksm.c's
572 * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache().
574 if (PageSwapCache(page
))
575 ClearPageSwapCache(page
);
576 ClearPagePrivate(page
);
577 set_page_private(page
, 0);
580 * If any waiters have accumulated on the new page then
583 if (PageWriteback(newpage
))
584 end_page_writeback(newpage
);
586 copy_page_owner(page
, newpage
);
588 mem_cgroup_migrate(page
, newpage
);
590 EXPORT_SYMBOL(migrate_page_copy
);
592 /************************************************************
593 * Migration functions
594 ***********************************************************/
597 * Common logic to directly migrate a single page suitable for
598 * pages that do not use PagePrivate/PagePrivate2.
600 * Pages are locked upon entry and exit.
602 int migrate_page(struct address_space
*mapping
,
603 struct page
*newpage
, struct page
*page
,
604 enum migrate_mode mode
)
608 BUG_ON(PageWriteback(page
)); /* Writeback must be complete */
610 rc
= migrate_page_move_mapping(mapping
, newpage
, page
, NULL
, mode
, 0);
612 if (rc
!= MIGRATEPAGE_SUCCESS
)
615 migrate_page_copy(newpage
, page
);
616 return MIGRATEPAGE_SUCCESS
;
618 EXPORT_SYMBOL(migrate_page
);
622 * Migration function for pages with buffers. This function can only be used
623 * if the underlying filesystem guarantees that no other references to "page"
626 int buffer_migrate_page(struct address_space
*mapping
,
627 struct page
*newpage
, struct page
*page
, enum migrate_mode mode
)
629 struct buffer_head
*bh
, *head
;
632 if (!page_has_buffers(page
))
633 return migrate_page(mapping
, newpage
, page
, mode
);
635 head
= page_buffers(page
);
637 rc
= migrate_page_move_mapping(mapping
, newpage
, page
, head
, mode
, 0);
639 if (rc
!= MIGRATEPAGE_SUCCESS
)
643 * In the async case, migrate_page_move_mapping locked the buffers
644 * with an IRQ-safe spinlock held. In the sync case, the buffers
645 * need to be locked now
647 if (mode
!= MIGRATE_ASYNC
)
648 BUG_ON(!buffer_migrate_lock_buffers(head
, mode
));
650 ClearPagePrivate(page
);
651 set_page_private(newpage
, page_private(page
));
652 set_page_private(page
, 0);
658 set_bh_page(bh
, newpage
, bh_offset(bh
));
659 bh
= bh
->b_this_page
;
661 } while (bh
!= head
);
663 SetPagePrivate(newpage
);
665 migrate_page_copy(newpage
, page
);
671 bh
= bh
->b_this_page
;
673 } while (bh
!= head
);
675 return MIGRATEPAGE_SUCCESS
;
677 EXPORT_SYMBOL(buffer_migrate_page
);
681 * Writeback a page to clean the dirty state
683 static int writeout(struct address_space
*mapping
, struct page
*page
)
685 struct writeback_control wbc
= {
686 .sync_mode
= WB_SYNC_NONE
,
689 .range_end
= LLONG_MAX
,
694 if (!mapping
->a_ops
->writepage
)
695 /* No write method for the address space */
698 if (!clear_page_dirty_for_io(page
))
699 /* Someone else already triggered a write */
703 * A dirty page may imply that the underlying filesystem has
704 * the page on some queue. So the page must be clean for
705 * migration. Writeout may mean we loose the lock and the
706 * page state is no longer what we checked for earlier.
707 * At this point we know that the migration attempt cannot
710 remove_migration_ptes(page
, page
, false);
712 rc
= mapping
->a_ops
->writepage(page
, &wbc
);
714 if (rc
!= AOP_WRITEPAGE_ACTIVATE
)
715 /* unlocked. Relock */
718 return (rc
< 0) ? -EIO
: -EAGAIN
;
722 * Default handling if a filesystem does not provide a migration function.
724 static int fallback_migrate_page(struct address_space
*mapping
,
725 struct page
*newpage
, struct page
*page
, enum migrate_mode mode
)
727 if (PageDirty(page
)) {
728 /* Only writeback pages in full synchronous migration */
729 if (mode
!= MIGRATE_SYNC
)
731 return writeout(mapping
, page
);
735 * Buffers may be managed in a filesystem specific way.
736 * We must have no buffers or drop them.
738 if (page_has_private(page
) &&
739 !try_to_release_page(page
, GFP_KERNEL
))
742 return migrate_page(mapping
, newpage
, page
, mode
);
746 * Move a page to a newly allocated page
747 * The page is locked and all ptes have been successfully removed.
749 * The new page will have replaced the old page if this function
754 * MIGRATEPAGE_SUCCESS - success
756 static int move_to_new_page(struct page
*newpage
, struct page
*page
,
757 enum migrate_mode mode
)
759 struct address_space
*mapping
;
762 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
763 VM_BUG_ON_PAGE(!PageLocked(newpage
), newpage
);
765 mapping
= page_mapping(page
);
767 rc
= migrate_page(mapping
, newpage
, page
, mode
);
768 else if (mapping
->a_ops
->migratepage
)
770 * Most pages have a mapping and most filesystems provide a
771 * migratepage callback. Anonymous pages are part of swap
772 * space which also has its own migratepage callback. This
773 * is the most common path for page migration.
775 rc
= mapping
->a_ops
->migratepage(mapping
, newpage
, page
, mode
);
777 rc
= fallback_migrate_page(mapping
, newpage
, page
, mode
);
780 * When successful, old pagecache page->mapping must be cleared before
781 * page is freed; but stats require that PageAnon be left as PageAnon.
783 if (rc
== MIGRATEPAGE_SUCCESS
) {
785 page
->mapping
= NULL
;
790 static int __unmap_and_move(struct page
*page
, struct page
*newpage
,
791 int force
, enum migrate_mode mode
)
794 int page_was_mapped
= 0;
795 struct anon_vma
*anon_vma
= NULL
;
797 if (!trylock_page(page
)) {
798 if (!force
|| mode
== MIGRATE_ASYNC
)
802 * It's not safe for direct compaction to call lock_page.
803 * For example, during page readahead pages are added locked
804 * to the LRU. Later, when the IO completes the pages are
805 * marked uptodate and unlocked. However, the queueing
806 * could be merging multiple pages for one bio (e.g.
807 * mpage_readpages). If an allocation happens for the
808 * second or third page, the process can end up locking
809 * the same page twice and deadlocking. Rather than
810 * trying to be clever about what pages can be locked,
811 * avoid the use of lock_page for direct compaction
814 if (current
->flags
& PF_MEMALLOC
)
820 if (PageWriteback(page
)) {
822 * Only in the case of a full synchronous migration is it
823 * necessary to wait for PageWriteback. In the async case,
824 * the retry loop is too short and in the sync-light case,
825 * the overhead of stalling is too much
827 if (mode
!= MIGRATE_SYNC
) {
833 wait_on_page_writeback(page
);
837 * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
838 * we cannot notice that anon_vma is freed while we migrates a page.
839 * This get_anon_vma() delays freeing anon_vma pointer until the end
840 * of migration. File cache pages are no problem because of page_lock()
841 * File Caches may use write_page() or lock_page() in migration, then,
842 * just care Anon page here.
844 * Only page_get_anon_vma() understands the subtleties of
845 * getting a hold on an anon_vma from outside one of its mms.
846 * But if we cannot get anon_vma, then we won't need it anyway,
847 * because that implies that the anon page is no longer mapped
848 * (and cannot be remapped so long as we hold the page lock).
850 if (PageAnon(page
) && !PageKsm(page
))
851 anon_vma
= page_get_anon_vma(page
);
854 * Block others from accessing the new page when we get around to
855 * establishing additional references. We are usually the only one
856 * holding a reference to newpage at this point. We used to have a BUG
857 * here if trylock_page(newpage) fails, but would like to allow for
858 * cases where there might be a race with the previous use of newpage.
859 * This is much like races on refcount of oldpage: just don't BUG().
861 if (unlikely(!trylock_page(newpage
)))
864 if (unlikely(isolated_balloon_page(page
))) {
866 * A ballooned page does not need any special attention from
867 * physical to virtual reverse mapping procedures.
868 * Skip any attempt to unmap PTEs or to remap swap cache,
869 * in order to avoid burning cycles at rmap level, and perform
870 * the page migration right away (proteced by page lock).
872 rc
= balloon_page_migrate(newpage
, page
, mode
);
873 goto out_unlock_both
;
877 * Corner case handling:
878 * 1. When a new swap-cache page is read into, it is added to the LRU
879 * and treated as swapcache but it has no rmap yet.
880 * Calling try_to_unmap() against a page->mapping==NULL page will
881 * trigger a BUG. So handle it here.
882 * 2. An orphaned page (see truncate_complete_page) might have
883 * fs-private metadata. The page can be picked up due to memory
884 * offlining. Everywhere else except page reclaim, the page is
885 * invisible to the vm, so the page can not be migrated. So try to
886 * free the metadata, so the page can be freed.
888 if (!page
->mapping
) {
889 VM_BUG_ON_PAGE(PageAnon(page
), page
);
890 if (page_has_private(page
)) {
891 try_to_free_buffers(page
);
892 goto out_unlock_both
;
894 } else if (page_mapped(page
)) {
895 /* Establish migration ptes */
896 VM_BUG_ON_PAGE(PageAnon(page
) && !PageKsm(page
) && !anon_vma
,
899 TTU_MIGRATION
|TTU_IGNORE_MLOCK
|TTU_IGNORE_ACCESS
);
903 if (!page_mapped(page
))
904 rc
= move_to_new_page(newpage
, page
, mode
);
907 remove_migration_ptes(page
,
908 rc
== MIGRATEPAGE_SUCCESS
? newpage
: page
, false);
911 unlock_page(newpage
);
913 /* Drop an anon_vma reference if we took one */
915 put_anon_vma(anon_vma
);
922 * gcc 4.7 and 4.8 on arm get an ICEs when inlining unmap_and_move(). Work
925 #if (GCC_VERSION >= 40700 && GCC_VERSION < 40900) && defined(CONFIG_ARM)
926 #define ICE_noinline noinline
932 * Obtain the lock on page, remove all ptes and migrate the page
933 * to the newly allocated page in newpage.
935 static ICE_noinline
int unmap_and_move(new_page_t get_new_page
,
936 free_page_t put_new_page
,
937 unsigned long private, struct page
*page
,
938 int force
, enum migrate_mode mode
,
939 enum migrate_reason reason
)
941 int rc
= MIGRATEPAGE_SUCCESS
;
943 struct page
*newpage
;
945 newpage
= get_new_page(page
, private, &result
);
949 if (page_count(page
) == 1) {
950 /* page was freed from under us. So we are done. */
954 if (unlikely(PageTransHuge(page
))) {
956 rc
= split_huge_page(page
);
962 rc
= __unmap_and_move(page
, newpage
, force
, mode
);
963 if (rc
== MIGRATEPAGE_SUCCESS
) {
965 set_page_owner_migrate_reason(newpage
, reason
);
971 * A page that has been migrated has all references
972 * removed and will be freed. A page that has not been
973 * migrated will have kepts its references and be
976 list_del(&page
->lru
);
977 dec_zone_page_state(page
, NR_ISOLATED_ANON
+
978 page_is_file_cache(page
));
979 /* Soft-offlined page shouldn't go through lru cache list */
980 if (reason
== MR_MEMORY_FAILURE
&& rc
== MIGRATEPAGE_SUCCESS
) {
982 * With this release, we free successfully migrated
983 * page and set PG_HWPoison on just freed page
984 * intentionally. Although it's rather weird, it's how
985 * HWPoison flag works at the moment.
988 if (!test_set_page_hwpoison(page
))
989 num_poisoned_pages_inc();
991 putback_lru_page(page
);
995 * If migration was not successful and there's a freeing callback, use
996 * it. Otherwise, putback_lru_page() will drop the reference grabbed
1000 put_new_page(newpage
, private);
1001 else if (unlikely(__is_movable_balloon_page(newpage
))) {
1002 /* drop our reference, page already in the balloon */
1005 putback_lru_page(newpage
);
1011 *result
= page_to_nid(newpage
);
1017 * Counterpart of unmap_and_move_page() for hugepage migration.
1019 * This function doesn't wait the completion of hugepage I/O
1020 * because there is no race between I/O and migration for hugepage.
1021 * Note that currently hugepage I/O occurs only in direct I/O
1022 * where no lock is held and PG_writeback is irrelevant,
1023 * and writeback status of all subpages are counted in the reference
1024 * count of the head page (i.e. if all subpages of a 2MB hugepage are
1025 * under direct I/O, the reference of the head page is 512 and a bit more.)
1026 * This means that when we try to migrate hugepage whose subpages are
1027 * doing direct I/O, some references remain after try_to_unmap() and
1028 * hugepage migration fails without data corruption.
1030 * There is also no race when direct I/O is issued on the page under migration,
1031 * because then pte is replaced with migration swap entry and direct I/O code
1032 * will wait in the page fault for migration to complete.
1034 static int unmap_and_move_huge_page(new_page_t get_new_page
,
1035 free_page_t put_new_page
, unsigned long private,
1036 struct page
*hpage
, int force
,
1037 enum migrate_mode mode
, int reason
)
1041 int page_was_mapped
= 0;
1042 struct page
*new_hpage
;
1043 struct anon_vma
*anon_vma
= NULL
;
1046 * Movability of hugepages depends on architectures and hugepage size.
1047 * This check is necessary because some callers of hugepage migration
1048 * like soft offline and memory hotremove don't walk through page
1049 * tables or check whether the hugepage is pmd-based or not before
1050 * kicking migration.
1052 if (!hugepage_migration_supported(page_hstate(hpage
))) {
1053 putback_active_hugepage(hpage
);
1057 new_hpage
= get_new_page(hpage
, private, &result
);
1061 if (!trylock_page(hpage
)) {
1062 if (!force
|| mode
!= MIGRATE_SYNC
)
1067 if (PageAnon(hpage
))
1068 anon_vma
= page_get_anon_vma(hpage
);
1070 if (unlikely(!trylock_page(new_hpage
)))
1073 if (page_mapped(hpage
)) {
1075 TTU_MIGRATION
|TTU_IGNORE_MLOCK
|TTU_IGNORE_ACCESS
);
1076 page_was_mapped
= 1;
1079 if (!page_mapped(hpage
))
1080 rc
= move_to_new_page(new_hpage
, hpage
, mode
);
1082 if (page_was_mapped
)
1083 remove_migration_ptes(hpage
,
1084 rc
== MIGRATEPAGE_SUCCESS
? new_hpage
: hpage
, false);
1086 unlock_page(new_hpage
);
1090 put_anon_vma(anon_vma
);
1092 if (rc
== MIGRATEPAGE_SUCCESS
) {
1093 hugetlb_cgroup_migrate(hpage
, new_hpage
);
1094 put_new_page
= NULL
;
1095 set_page_owner_migrate_reason(new_hpage
, reason
);
1101 putback_active_hugepage(hpage
);
1104 * If migration was not successful and there's a freeing callback, use
1105 * it. Otherwise, put_page() will drop the reference grabbed during
1109 put_new_page(new_hpage
, private);
1111 putback_active_hugepage(new_hpage
);
1117 *result
= page_to_nid(new_hpage
);
1123 * migrate_pages - migrate the pages specified in a list, to the free pages
1124 * supplied as the target for the page migration
1126 * @from: The list of pages to be migrated.
1127 * @get_new_page: The function used to allocate free pages to be used
1128 * as the target of the page migration.
1129 * @put_new_page: The function used to free target pages if migration
1130 * fails, or NULL if no special handling is necessary.
1131 * @private: Private data to be passed on to get_new_page()
1132 * @mode: The migration mode that specifies the constraints for
1133 * page migration, if any.
1134 * @reason: The reason for page migration.
1136 * The function returns after 10 attempts or if no pages are movable any more
1137 * because the list has become empty or no retryable pages exist any more.
1138 * The caller should call putback_movable_pages() to return pages to the LRU
1139 * or free list only if ret != 0.
1141 * Returns the number of pages that were not migrated, or an error code.
1143 int migrate_pages(struct list_head
*from
, new_page_t get_new_page
,
1144 free_page_t put_new_page
, unsigned long private,
1145 enum migrate_mode mode
, int reason
)
1149 int nr_succeeded
= 0;
1153 int swapwrite
= current
->flags
& PF_SWAPWRITE
;
1157 current
->flags
|= PF_SWAPWRITE
;
1159 for(pass
= 0; pass
< 10 && retry
; pass
++) {
1162 list_for_each_entry_safe(page
, page2
, from
, lru
) {
1166 rc
= unmap_and_move_huge_page(get_new_page
,
1167 put_new_page
, private, page
,
1168 pass
> 2, mode
, reason
);
1170 rc
= unmap_and_move(get_new_page
, put_new_page
,
1171 private, page
, pass
> 2, mode
,
1181 case MIGRATEPAGE_SUCCESS
:
1186 * Permanent failure (-EBUSY, -ENOSYS, etc.):
1187 * unlike -EAGAIN case, the failed page is
1188 * removed from migration page list and not
1189 * retried in the next outer loop.
1200 count_vm_events(PGMIGRATE_SUCCESS
, nr_succeeded
);
1202 count_vm_events(PGMIGRATE_FAIL
, nr_failed
);
1203 trace_mm_migrate_pages(nr_succeeded
, nr_failed
, mode
, reason
);
1206 current
->flags
&= ~PF_SWAPWRITE
;
1213 * Move a list of individual pages
1215 struct page_to_node
{
1222 static struct page
*new_page_node(struct page
*p
, unsigned long private,
1225 struct page_to_node
*pm
= (struct page_to_node
*)private;
1227 while (pm
->node
!= MAX_NUMNODES
&& pm
->page
!= p
)
1230 if (pm
->node
== MAX_NUMNODES
)
1233 *result
= &pm
->status
;
1236 return alloc_huge_page_node(page_hstate(compound_head(p
)),
1239 return __alloc_pages_node(pm
->node
,
1240 GFP_HIGHUSER_MOVABLE
| __GFP_THISNODE
, 0);
1244 * Move a set of pages as indicated in the pm array. The addr
1245 * field must be set to the virtual address of the page to be moved
1246 * and the node number must contain a valid target node.
1247 * The pm array ends with node = MAX_NUMNODES.
1249 static int do_move_page_to_node_array(struct mm_struct
*mm
,
1250 struct page_to_node
*pm
,
1254 struct page_to_node
*pp
;
1255 LIST_HEAD(pagelist
);
1257 down_read(&mm
->mmap_sem
);
1260 * Build a list of pages to migrate
1262 for (pp
= pm
; pp
->node
!= MAX_NUMNODES
; pp
++) {
1263 struct vm_area_struct
*vma
;
1267 vma
= find_vma(mm
, pp
->addr
);
1268 if (!vma
|| pp
->addr
< vma
->vm_start
|| !vma_migratable(vma
))
1271 /* FOLL_DUMP to ignore special (like zero) pages */
1272 page
= follow_page(vma
, pp
->addr
,
1273 FOLL_GET
| FOLL_SPLIT
| FOLL_DUMP
);
1275 err
= PTR_ERR(page
);
1284 err
= page_to_nid(page
);
1286 if (err
== pp
->node
)
1288 * Node already in the right place
1293 if (page_mapcount(page
) > 1 &&
1297 if (PageHuge(page
)) {
1299 isolate_huge_page(page
, &pagelist
);
1303 err
= isolate_lru_page(page
);
1305 list_add_tail(&page
->lru
, &pagelist
);
1306 inc_zone_page_state(page
, NR_ISOLATED_ANON
+
1307 page_is_file_cache(page
));
1311 * Either remove the duplicate refcount from
1312 * isolate_lru_page() or drop the page ref if it was
1321 if (!list_empty(&pagelist
)) {
1322 err
= migrate_pages(&pagelist
, new_page_node
, NULL
,
1323 (unsigned long)pm
, MIGRATE_SYNC
, MR_SYSCALL
);
1325 putback_movable_pages(&pagelist
);
1328 up_read(&mm
->mmap_sem
);
1333 * Migrate an array of page address onto an array of nodes and fill
1334 * the corresponding array of status.
1336 static int do_pages_move(struct mm_struct
*mm
, nodemask_t task_nodes
,
1337 unsigned long nr_pages
,
1338 const void __user
* __user
*pages
,
1339 const int __user
*nodes
,
1340 int __user
*status
, int flags
)
1342 struct page_to_node
*pm
;
1343 unsigned long chunk_nr_pages
;
1344 unsigned long chunk_start
;
1348 pm
= (struct page_to_node
*)__get_free_page(GFP_KERNEL
);
1355 * Store a chunk of page_to_node array in a page,
1356 * but keep the last one as a marker
1358 chunk_nr_pages
= (PAGE_SIZE
/ sizeof(struct page_to_node
)) - 1;
1360 for (chunk_start
= 0;
1361 chunk_start
< nr_pages
;
1362 chunk_start
+= chunk_nr_pages
) {
1365 if (chunk_start
+ chunk_nr_pages
> nr_pages
)
1366 chunk_nr_pages
= nr_pages
- chunk_start
;
1368 /* fill the chunk pm with addrs and nodes from user-space */
1369 for (j
= 0; j
< chunk_nr_pages
; j
++) {
1370 const void __user
*p
;
1374 if (get_user(p
, pages
+ j
+ chunk_start
))
1376 pm
[j
].addr
= (unsigned long) p
;
1378 if (get_user(node
, nodes
+ j
+ chunk_start
))
1382 if (node
< 0 || node
>= MAX_NUMNODES
)
1385 if (!node_state(node
, N_MEMORY
))
1389 if (!node_isset(node
, task_nodes
))
1395 /* End marker for this chunk */
1396 pm
[chunk_nr_pages
].node
= MAX_NUMNODES
;
1398 /* Migrate this chunk */
1399 err
= do_move_page_to_node_array(mm
, pm
,
1400 flags
& MPOL_MF_MOVE_ALL
);
1404 /* Return status information */
1405 for (j
= 0; j
< chunk_nr_pages
; j
++)
1406 if (put_user(pm
[j
].status
, status
+ j
+ chunk_start
)) {
1414 free_page((unsigned long)pm
);
1420 * Determine the nodes of an array of pages and store it in an array of status.
1422 static void do_pages_stat_array(struct mm_struct
*mm
, unsigned long nr_pages
,
1423 const void __user
**pages
, int *status
)
1427 down_read(&mm
->mmap_sem
);
1429 for (i
= 0; i
< nr_pages
; i
++) {
1430 unsigned long addr
= (unsigned long)(*pages
);
1431 struct vm_area_struct
*vma
;
1435 vma
= find_vma(mm
, addr
);
1436 if (!vma
|| addr
< vma
->vm_start
)
1439 /* FOLL_DUMP to ignore special (like zero) pages */
1440 page
= follow_page(vma
, addr
, FOLL_DUMP
);
1442 err
= PTR_ERR(page
);
1446 err
= page
? page_to_nid(page
) : -ENOENT
;
1454 up_read(&mm
->mmap_sem
);
1458 * Determine the nodes of a user array of pages and store it in
1459 * a user array of status.
1461 static int do_pages_stat(struct mm_struct
*mm
, unsigned long nr_pages
,
1462 const void __user
* __user
*pages
,
1465 #define DO_PAGES_STAT_CHUNK_NR 16
1466 const void __user
*chunk_pages
[DO_PAGES_STAT_CHUNK_NR
];
1467 int chunk_status
[DO_PAGES_STAT_CHUNK_NR
];
1470 unsigned long chunk_nr
;
1472 chunk_nr
= nr_pages
;
1473 if (chunk_nr
> DO_PAGES_STAT_CHUNK_NR
)
1474 chunk_nr
= DO_PAGES_STAT_CHUNK_NR
;
1476 if (copy_from_user(chunk_pages
, pages
, chunk_nr
* sizeof(*chunk_pages
)))
1479 do_pages_stat_array(mm
, chunk_nr
, chunk_pages
, chunk_status
);
1481 if (copy_to_user(status
, chunk_status
, chunk_nr
* sizeof(*status
)))
1486 nr_pages
-= chunk_nr
;
1488 return nr_pages
? -EFAULT
: 0;
1492 * Move a list of pages in the address space of the currently executing
1495 SYSCALL_DEFINE6(move_pages
, pid_t
, pid
, unsigned long, nr_pages
,
1496 const void __user
* __user
*, pages
,
1497 const int __user
*, nodes
,
1498 int __user
*, status
, int, flags
)
1500 const struct cred
*cred
= current_cred(), *tcred
;
1501 struct task_struct
*task
;
1502 struct mm_struct
*mm
;
1504 nodemask_t task_nodes
;
1507 if (flags
& ~(MPOL_MF_MOVE
|MPOL_MF_MOVE_ALL
))
1510 if ((flags
& MPOL_MF_MOVE_ALL
) && !capable(CAP_SYS_NICE
))
1513 /* Find the mm_struct */
1515 task
= pid
? find_task_by_vpid(pid
) : current
;
1520 get_task_struct(task
);
1523 * Check if this process has the right to modify the specified
1524 * process. The right exists if the process has administrative
1525 * capabilities, superuser privileges or the same
1526 * userid as the target process.
1528 tcred
= __task_cred(task
);
1529 if (!uid_eq(cred
->euid
, tcred
->suid
) && !uid_eq(cred
->euid
, tcred
->uid
) &&
1530 !uid_eq(cred
->uid
, tcred
->suid
) && !uid_eq(cred
->uid
, tcred
->uid
) &&
1531 !capable(CAP_SYS_NICE
)) {
1538 err
= security_task_movememory(task
);
1542 task_nodes
= cpuset_mems_allowed(task
);
1543 mm
= get_task_mm(task
);
1544 put_task_struct(task
);
1550 err
= do_pages_move(mm
, task_nodes
, nr_pages
, pages
,
1551 nodes
, status
, flags
);
1553 err
= do_pages_stat(mm
, nr_pages
, pages
, status
);
1559 put_task_struct(task
);
1563 #ifdef CONFIG_NUMA_BALANCING
1565 * Returns true if this is a safe migration target node for misplaced NUMA
1566 * pages. Currently it only checks the watermarks which crude
1568 static bool migrate_balanced_pgdat(struct pglist_data
*pgdat
,
1569 unsigned long nr_migrate_pages
)
1572 for (z
= pgdat
->nr_zones
- 1; z
>= 0; z
--) {
1573 struct zone
*zone
= pgdat
->node_zones
+ z
;
1575 if (!populated_zone(zone
))
1578 if (!zone_reclaimable(zone
))
1581 /* Avoid waking kswapd by allocating pages_to_migrate pages. */
1582 if (!zone_watermark_ok(zone
, 0,
1583 high_wmark_pages(zone
) +
1592 static struct page
*alloc_misplaced_dst_page(struct page
*page
,
1596 int nid
= (int) data
;
1597 struct page
*newpage
;
1599 newpage
= __alloc_pages_node(nid
,
1600 (GFP_HIGHUSER_MOVABLE
|
1601 __GFP_THISNODE
| __GFP_NOMEMALLOC
|
1602 __GFP_NORETRY
| __GFP_NOWARN
) &
1609 * page migration rate limiting control.
1610 * Do not migrate more than @pages_to_migrate in a @migrate_interval_millisecs
1611 * window of time. Default here says do not migrate more than 1280M per second.
1613 static unsigned int migrate_interval_millisecs __read_mostly
= 100;
1614 static unsigned int ratelimit_pages __read_mostly
= 128 << (20 - PAGE_SHIFT
);
1616 /* Returns true if the node is migrate rate-limited after the update */
1617 static bool numamigrate_update_ratelimit(pg_data_t
*pgdat
,
1618 unsigned long nr_pages
)
1621 * Rate-limit the amount of data that is being migrated to a node.
1622 * Optimal placement is no good if the memory bus is saturated and
1623 * all the time is being spent migrating!
1625 if (time_after(jiffies
, pgdat
->numabalancing_migrate_next_window
)) {
1626 spin_lock(&pgdat
->numabalancing_migrate_lock
);
1627 pgdat
->numabalancing_migrate_nr_pages
= 0;
1628 pgdat
->numabalancing_migrate_next_window
= jiffies
+
1629 msecs_to_jiffies(migrate_interval_millisecs
);
1630 spin_unlock(&pgdat
->numabalancing_migrate_lock
);
1632 if (pgdat
->numabalancing_migrate_nr_pages
> ratelimit_pages
) {
1633 trace_mm_numa_migrate_ratelimit(current
, pgdat
->node_id
,
1639 * This is an unlocked non-atomic update so errors are possible.
1640 * The consequences are failing to migrate when we potentiall should
1641 * have which is not severe enough to warrant locking. If it is ever
1642 * a problem, it can be converted to a per-cpu counter.
1644 pgdat
->numabalancing_migrate_nr_pages
+= nr_pages
;
1648 static int numamigrate_isolate_page(pg_data_t
*pgdat
, struct page
*page
)
1652 VM_BUG_ON_PAGE(compound_order(page
) && !PageTransHuge(page
), page
);
1654 /* Avoid migrating to a node that is nearly full */
1655 if (!migrate_balanced_pgdat(pgdat
, 1UL << compound_order(page
)))
1658 if (isolate_lru_page(page
))
1662 * migrate_misplaced_transhuge_page() skips page migration's usual
1663 * check on page_count(), so we must do it here, now that the page
1664 * has been isolated: a GUP pin, or any other pin, prevents migration.
1665 * The expected page count is 3: 1 for page's mapcount and 1 for the
1666 * caller's pin and 1 for the reference taken by isolate_lru_page().
1668 if (PageTransHuge(page
) && page_count(page
) != 3) {
1669 putback_lru_page(page
);
1673 page_lru
= page_is_file_cache(page
);
1674 mod_zone_page_state(page_zone(page
), NR_ISOLATED_ANON
+ page_lru
,
1675 hpage_nr_pages(page
));
1678 * Isolating the page has taken another reference, so the
1679 * caller's reference can be safely dropped without the page
1680 * disappearing underneath us during migration.
1686 bool pmd_trans_migrating(pmd_t pmd
)
1688 struct page
*page
= pmd_page(pmd
);
1689 return PageLocked(page
);
1693 * Attempt to migrate a misplaced page to the specified destination
1694 * node. Caller is expected to have an elevated reference count on
1695 * the page that will be dropped by this function before returning.
1697 int migrate_misplaced_page(struct page
*page
, struct vm_area_struct
*vma
,
1700 pg_data_t
*pgdat
= NODE_DATA(node
);
1703 LIST_HEAD(migratepages
);
1706 * Don't migrate file pages that are mapped in multiple processes
1707 * with execute permissions as they are probably shared libraries.
1709 if (page_mapcount(page
) != 1 && page_is_file_cache(page
) &&
1710 (vma
->vm_flags
& VM_EXEC
))
1714 * Rate-limit the amount of data that is being migrated to a node.
1715 * Optimal placement is no good if the memory bus is saturated and
1716 * all the time is being spent migrating!
1718 if (numamigrate_update_ratelimit(pgdat
, 1))
1721 isolated
= numamigrate_isolate_page(pgdat
, page
);
1725 list_add(&page
->lru
, &migratepages
);
1726 nr_remaining
= migrate_pages(&migratepages
, alloc_misplaced_dst_page
,
1727 NULL
, node
, MIGRATE_ASYNC
,
1730 if (!list_empty(&migratepages
)) {
1731 list_del(&page
->lru
);
1732 dec_zone_page_state(page
, NR_ISOLATED_ANON
+
1733 page_is_file_cache(page
));
1734 putback_lru_page(page
);
1738 count_vm_numa_event(NUMA_PAGE_MIGRATE
);
1739 BUG_ON(!list_empty(&migratepages
));
1746 #endif /* CONFIG_NUMA_BALANCING */
1748 #if defined(CONFIG_NUMA_BALANCING) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
1750 * Migrates a THP to a given target node. page must be locked and is unlocked
1753 int migrate_misplaced_transhuge_page(struct mm_struct
*mm
,
1754 struct vm_area_struct
*vma
,
1755 pmd_t
*pmd
, pmd_t entry
,
1756 unsigned long address
,
1757 struct page
*page
, int node
)
1760 pg_data_t
*pgdat
= NODE_DATA(node
);
1762 struct page
*new_page
= NULL
;
1763 int page_lru
= page_is_file_cache(page
);
1764 unsigned long mmun_start
= address
& HPAGE_PMD_MASK
;
1765 unsigned long mmun_end
= mmun_start
+ HPAGE_PMD_SIZE
;
1769 * Rate-limit the amount of data that is being migrated to a node.
1770 * Optimal placement is no good if the memory bus is saturated and
1771 * all the time is being spent migrating!
1773 if (numamigrate_update_ratelimit(pgdat
, HPAGE_PMD_NR
))
1776 new_page
= alloc_pages_node(node
,
1777 (GFP_TRANSHUGE
| __GFP_THISNODE
) & ~__GFP_RECLAIM
,
1781 prep_transhuge_page(new_page
);
1783 isolated
= numamigrate_isolate_page(pgdat
, page
);
1789 * We are not sure a pending tlb flush here is for a huge page
1790 * mapping or not. Hence use the tlb range variant
1792 if (mm_tlb_flush_pending(mm
))
1793 flush_tlb_range(vma
, mmun_start
, mmun_end
);
1795 /* Prepare a page as a migration target */
1796 __SetPageLocked(new_page
);
1797 __SetPageSwapBacked(new_page
);
1799 /* anon mapping, we can simply copy page->mapping to the new page: */
1800 new_page
->mapping
= page
->mapping
;
1801 new_page
->index
= page
->index
;
1802 migrate_page_copy(new_page
, page
);
1803 WARN_ON(PageLRU(new_page
));
1805 /* Recheck the target PMD */
1806 mmu_notifier_invalidate_range_start(mm
, mmun_start
, mmun_end
);
1807 ptl
= pmd_lock(mm
, pmd
);
1808 if (unlikely(!pmd_same(*pmd
, entry
) || page_count(page
) != 2)) {
1811 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
1813 /* Reverse changes made by migrate_page_copy() */
1814 if (TestClearPageActive(new_page
))
1815 SetPageActive(page
);
1816 if (TestClearPageUnevictable(new_page
))
1817 SetPageUnevictable(page
);
1819 unlock_page(new_page
);
1820 put_page(new_page
); /* Free it */
1822 /* Retake the callers reference and putback on LRU */
1824 putback_lru_page(page
);
1825 mod_zone_page_state(page_zone(page
),
1826 NR_ISOLATED_ANON
+ page_lru
, -HPAGE_PMD_NR
);
1832 entry
= mk_pmd(new_page
, vma
->vm_page_prot
);
1833 entry
= pmd_mkhuge(entry
);
1834 entry
= maybe_pmd_mkwrite(pmd_mkdirty(entry
), vma
);
1837 * Clear the old entry under pagetable lock and establish the new PTE.
1838 * Any parallel GUP will either observe the old page blocking on the
1839 * page lock, block on the page table lock or observe the new page.
1840 * The SetPageUptodate on the new page and page_add_new_anon_rmap
1841 * guarantee the copy is visible before the pagetable update.
1843 flush_cache_range(vma
, mmun_start
, mmun_end
);
1844 page_add_anon_rmap(new_page
, vma
, mmun_start
, true);
1845 pmdp_huge_clear_flush_notify(vma
, mmun_start
, pmd
);
1846 set_pmd_at(mm
, mmun_start
, pmd
, entry
);
1847 update_mmu_cache_pmd(vma
, address
, &entry
);
1849 if (page_count(page
) != 2) {
1850 set_pmd_at(mm
, mmun_start
, pmd
, orig_entry
);
1851 flush_pmd_tlb_range(vma
, mmun_start
, mmun_end
);
1852 mmu_notifier_invalidate_range(mm
, mmun_start
, mmun_end
);
1853 update_mmu_cache_pmd(vma
, address
, &entry
);
1854 page_remove_rmap(new_page
, true);
1858 mlock_migrate_page(new_page
, page
);
1859 page_remove_rmap(page
, true);
1860 set_page_owner_migrate_reason(new_page
, MR_NUMA_MISPLACED
);
1863 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
1865 /* Take an "isolate" reference and put new page on the LRU. */
1867 putback_lru_page(new_page
);
1869 unlock_page(new_page
);
1871 put_page(page
); /* Drop the rmap reference */
1872 put_page(page
); /* Drop the LRU isolation reference */
1874 count_vm_events(PGMIGRATE_SUCCESS
, HPAGE_PMD_NR
);
1875 count_vm_numa_events(NUMA_PAGE_MIGRATE
, HPAGE_PMD_NR
);
1877 mod_zone_page_state(page_zone(page
),
1878 NR_ISOLATED_ANON
+ page_lru
,
1883 count_vm_events(PGMIGRATE_FAIL
, HPAGE_PMD_NR
);
1885 ptl
= pmd_lock(mm
, pmd
);
1886 if (pmd_same(*pmd
, entry
)) {
1887 entry
= pmd_modify(entry
, vma
->vm_page_prot
);
1888 set_pmd_at(mm
, mmun_start
, pmd
, entry
);
1889 update_mmu_cache_pmd(vma
, address
, &entry
);
1898 #endif /* CONFIG_NUMA_BALANCING */
1900 #endif /* CONFIG_NUMA */