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
)
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 static void remove_migration_ptes(struct page
*old
, struct page
*new)
192 struct rmap_walk_control rwc
= {
193 .rmap_one
= remove_migration_pte
,
197 rmap_walk(new, &rwc
);
201 * Something used the pte of a page under migration. We need to
202 * get to the page and wait until migration is finished.
203 * When we return from this function the fault will be retried.
205 void __migration_entry_wait(struct mm_struct
*mm
, pte_t
*ptep
,
214 if (!is_swap_pte(pte
))
217 entry
= pte_to_swp_entry(pte
);
218 if (!is_migration_entry(entry
))
221 page
= migration_entry_to_page(entry
);
224 * Once radix-tree replacement of page migration started, page_count
225 * *must* be zero. And, we don't want to call wait_on_page_locked()
226 * against a page without get_page().
227 * So, we use get_page_unless_zero(), here. Even failed, page fault
230 if (!get_page_unless_zero(page
))
232 pte_unmap_unlock(ptep
, ptl
);
233 wait_on_page_locked(page
);
237 pte_unmap_unlock(ptep
, ptl
);
240 void migration_entry_wait(struct mm_struct
*mm
, pmd_t
*pmd
,
241 unsigned long address
)
243 spinlock_t
*ptl
= pte_lockptr(mm
, pmd
);
244 pte_t
*ptep
= pte_offset_map(pmd
, address
);
245 __migration_entry_wait(mm
, ptep
, ptl
);
248 void migration_entry_wait_huge(struct vm_area_struct
*vma
,
249 struct mm_struct
*mm
, pte_t
*pte
)
251 spinlock_t
*ptl
= huge_pte_lockptr(hstate_vma(vma
), mm
, pte
);
252 __migration_entry_wait(mm
, pte
, ptl
);
256 /* Returns true if all buffers are successfully locked */
257 static bool buffer_migrate_lock_buffers(struct buffer_head
*head
,
258 enum migrate_mode mode
)
260 struct buffer_head
*bh
= head
;
262 /* Simple case, sync compaction */
263 if (mode
!= MIGRATE_ASYNC
) {
267 bh
= bh
->b_this_page
;
269 } while (bh
!= head
);
274 /* async case, we cannot block on lock_buffer so use trylock_buffer */
277 if (!trylock_buffer(bh
)) {
279 * We failed to lock the buffer and cannot stall in
280 * async migration. Release the taken locks
282 struct buffer_head
*failed_bh
= bh
;
285 while (bh
!= failed_bh
) {
288 bh
= bh
->b_this_page
;
293 bh
= bh
->b_this_page
;
294 } while (bh
!= head
);
298 static inline bool buffer_migrate_lock_buffers(struct buffer_head
*head
,
299 enum migrate_mode mode
)
303 #endif /* CONFIG_BLOCK */
306 * Replace the page in the mapping.
308 * The number of remaining references must be:
309 * 1 for anonymous pages without a mapping
310 * 2 for pages with a mapping
311 * 3 for pages with a mapping and PagePrivate/PagePrivate2 set.
313 int migrate_page_move_mapping(struct address_space
*mapping
,
314 struct page
*newpage
, struct page
*page
,
315 struct buffer_head
*head
, enum migrate_mode mode
,
318 struct zone
*oldzone
, *newzone
;
320 int expected_count
= 1 + extra_count
;
324 /* Anonymous page without mapping */
325 if (page_count(page
) != expected_count
)
328 /* No turning back from here */
329 newpage
->index
= page
->index
;
330 newpage
->mapping
= page
->mapping
;
331 if (PageSwapBacked(page
))
332 SetPageSwapBacked(newpage
);
334 return MIGRATEPAGE_SUCCESS
;
337 oldzone
= page_zone(page
);
338 newzone
= page_zone(newpage
);
340 spin_lock_irq(&mapping
->tree_lock
);
342 pslot
= radix_tree_lookup_slot(&mapping
->page_tree
,
345 expected_count
+= 1 + page_has_private(page
);
346 if (page_count(page
) != expected_count
||
347 radix_tree_deref_slot_protected(pslot
, &mapping
->tree_lock
) != page
) {
348 spin_unlock_irq(&mapping
->tree_lock
);
352 if (!page_freeze_refs(page
, expected_count
)) {
353 spin_unlock_irq(&mapping
->tree_lock
);
358 * In the async migration case of moving a page with buffers, lock the
359 * buffers using trylock before the mapping is moved. If the mapping
360 * was moved, we later failed to lock the buffers and could not move
361 * the mapping back due to an elevated page count, we would have to
362 * block waiting on other references to be dropped.
364 if (mode
== MIGRATE_ASYNC
&& head
&&
365 !buffer_migrate_lock_buffers(head
, mode
)) {
366 page_unfreeze_refs(page
, expected_count
);
367 spin_unlock_irq(&mapping
->tree_lock
);
372 * Now we know that no one else is looking at the page:
373 * no turning back from here.
375 newpage
->index
= page
->index
;
376 newpage
->mapping
= page
->mapping
;
377 if (PageSwapBacked(page
))
378 SetPageSwapBacked(newpage
);
380 get_page(newpage
); /* add cache reference */
381 if (PageSwapCache(page
)) {
382 SetPageSwapCache(newpage
);
383 set_page_private(newpage
, page_private(page
));
386 /* Move dirty while page refs frozen and newpage not yet exposed */
387 dirty
= PageDirty(page
);
389 ClearPageDirty(page
);
390 SetPageDirty(newpage
);
393 radix_tree_replace_slot(pslot
, newpage
);
396 * Drop cache reference from old page by unfreezing
397 * to one less reference.
398 * We know this isn't the last reference.
400 page_unfreeze_refs(page
, expected_count
- 1);
402 spin_unlock(&mapping
->tree_lock
);
403 /* Leave irq disabled to prevent preemption while updating stats */
406 * If moved to a different zone then also account
407 * the page for that zone. Other VM counters will be
408 * taken care of when we establish references to the
409 * new page and drop references to the old page.
411 * Note that anonymous pages are accounted for
412 * via NR_FILE_PAGES and NR_ANON_PAGES if they
413 * are mapped to swap space.
415 if (newzone
!= oldzone
) {
416 __dec_zone_state(oldzone
, NR_FILE_PAGES
);
417 __inc_zone_state(newzone
, NR_FILE_PAGES
);
418 if (PageSwapBacked(page
) && !PageSwapCache(page
)) {
419 __dec_zone_state(oldzone
, NR_SHMEM
);
420 __inc_zone_state(newzone
, NR_SHMEM
);
422 if (dirty
&& mapping_cap_account_dirty(mapping
)) {
423 __dec_zone_state(oldzone
, NR_FILE_DIRTY
);
424 __inc_zone_state(newzone
, NR_FILE_DIRTY
);
429 return MIGRATEPAGE_SUCCESS
;
433 * The expected number of remaining references is the same as that
434 * of migrate_page_move_mapping().
436 int migrate_huge_page_move_mapping(struct address_space
*mapping
,
437 struct page
*newpage
, struct page
*page
)
442 spin_lock_irq(&mapping
->tree_lock
);
444 pslot
= radix_tree_lookup_slot(&mapping
->page_tree
,
447 expected_count
= 2 + page_has_private(page
);
448 if (page_count(page
) != expected_count
||
449 radix_tree_deref_slot_protected(pslot
, &mapping
->tree_lock
) != page
) {
450 spin_unlock_irq(&mapping
->tree_lock
);
454 if (!page_freeze_refs(page
, expected_count
)) {
455 spin_unlock_irq(&mapping
->tree_lock
);
459 newpage
->index
= page
->index
;
460 newpage
->mapping
= page
->mapping
;
464 radix_tree_replace_slot(pslot
, newpage
);
466 page_unfreeze_refs(page
, expected_count
- 1);
468 spin_unlock_irq(&mapping
->tree_lock
);
470 return MIGRATEPAGE_SUCCESS
;
474 * Gigantic pages are so large that we do not guarantee that page++ pointer
475 * arithmetic will work across the entire page. We need something more
478 static void __copy_gigantic_page(struct page
*dst
, struct page
*src
,
482 struct page
*dst_base
= dst
;
483 struct page
*src_base
= src
;
485 for (i
= 0; i
< nr_pages
; ) {
487 copy_highpage(dst
, src
);
490 dst
= mem_map_next(dst
, dst_base
, i
);
491 src
= mem_map_next(src
, src_base
, i
);
495 static void copy_huge_page(struct page
*dst
, struct page
*src
)
502 struct hstate
*h
= page_hstate(src
);
503 nr_pages
= pages_per_huge_page(h
);
505 if (unlikely(nr_pages
> MAX_ORDER_NR_PAGES
)) {
506 __copy_gigantic_page(dst
, src
, nr_pages
);
511 BUG_ON(!PageTransHuge(src
));
512 nr_pages
= hpage_nr_pages(src
);
515 for (i
= 0; i
< nr_pages
; i
++) {
517 copy_highpage(dst
+ i
, src
+ i
);
522 * Copy the page to its new location
524 void migrate_page_copy(struct page
*newpage
, struct page
*page
)
528 if (PageHuge(page
) || PageTransHuge(page
))
529 copy_huge_page(newpage
, page
);
531 copy_highpage(newpage
, page
);
534 SetPageError(newpage
);
535 if (PageReferenced(page
))
536 SetPageReferenced(newpage
);
537 if (PageUptodate(page
))
538 SetPageUptodate(newpage
);
539 if (TestClearPageActive(page
)) {
540 VM_BUG_ON_PAGE(PageUnevictable(page
), page
);
541 SetPageActive(newpage
);
542 } else if (TestClearPageUnevictable(page
))
543 SetPageUnevictable(newpage
);
544 if (PageChecked(page
))
545 SetPageChecked(newpage
);
546 if (PageMappedToDisk(page
))
547 SetPageMappedToDisk(newpage
);
549 /* Move dirty on pages not done by migrate_page_move_mapping() */
551 SetPageDirty(newpage
);
553 if (page_is_young(page
))
554 set_page_young(newpage
);
555 if (page_is_idle(page
))
556 set_page_idle(newpage
);
559 * Copy NUMA information to the new page, to prevent over-eager
560 * future migrations of this same page.
562 cpupid
= page_cpupid_xchg_last(page
, -1);
563 page_cpupid_xchg_last(newpage
, cpupid
);
565 ksm_migrate_page(newpage
, page
);
567 * Please do not reorder this without considering how mm/ksm.c's
568 * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache().
570 if (PageSwapCache(page
))
571 ClearPageSwapCache(page
);
572 ClearPagePrivate(page
);
573 set_page_private(page
, 0);
576 * If any waiters have accumulated on the new page then
579 if (PageWriteback(newpage
))
580 end_page_writeback(newpage
);
582 copy_page_owner(page
, newpage
);
584 mem_cgroup_migrate(page
, newpage
);
587 /************************************************************
588 * Migration functions
589 ***********************************************************/
592 * Common logic to directly migrate a single page suitable for
593 * pages that do not use PagePrivate/PagePrivate2.
595 * Pages are locked upon entry and exit.
597 int migrate_page(struct address_space
*mapping
,
598 struct page
*newpage
, struct page
*page
,
599 enum migrate_mode mode
)
603 BUG_ON(PageWriteback(page
)); /* Writeback must be complete */
605 rc
= migrate_page_move_mapping(mapping
, newpage
, page
, NULL
, mode
, 0);
607 if (rc
!= MIGRATEPAGE_SUCCESS
)
610 migrate_page_copy(newpage
, page
);
611 return MIGRATEPAGE_SUCCESS
;
613 EXPORT_SYMBOL(migrate_page
);
617 * Migration function for pages with buffers. This function can only be used
618 * if the underlying filesystem guarantees that no other references to "page"
621 int buffer_migrate_page(struct address_space
*mapping
,
622 struct page
*newpage
, struct page
*page
, enum migrate_mode mode
)
624 struct buffer_head
*bh
, *head
;
627 if (!page_has_buffers(page
))
628 return migrate_page(mapping
, newpage
, page
, mode
);
630 head
= page_buffers(page
);
632 rc
= migrate_page_move_mapping(mapping
, newpage
, page
, head
, mode
, 0);
634 if (rc
!= MIGRATEPAGE_SUCCESS
)
638 * In the async case, migrate_page_move_mapping locked the buffers
639 * with an IRQ-safe spinlock held. In the sync case, the buffers
640 * need to be locked now
642 if (mode
!= MIGRATE_ASYNC
)
643 BUG_ON(!buffer_migrate_lock_buffers(head
, mode
));
645 ClearPagePrivate(page
);
646 set_page_private(newpage
, page_private(page
));
647 set_page_private(page
, 0);
653 set_bh_page(bh
, newpage
, bh_offset(bh
));
654 bh
= bh
->b_this_page
;
656 } while (bh
!= head
);
658 SetPagePrivate(newpage
);
660 migrate_page_copy(newpage
, page
);
666 bh
= bh
->b_this_page
;
668 } while (bh
!= head
);
670 return MIGRATEPAGE_SUCCESS
;
672 EXPORT_SYMBOL(buffer_migrate_page
);
676 * Writeback a page to clean the dirty state
678 static int writeout(struct address_space
*mapping
, struct page
*page
)
680 struct writeback_control wbc
= {
681 .sync_mode
= WB_SYNC_NONE
,
684 .range_end
= LLONG_MAX
,
689 if (!mapping
->a_ops
->writepage
)
690 /* No write method for the address space */
693 if (!clear_page_dirty_for_io(page
))
694 /* Someone else already triggered a write */
698 * A dirty page may imply that the underlying filesystem has
699 * the page on some queue. So the page must be clean for
700 * migration. Writeout may mean we loose the lock and the
701 * page state is no longer what we checked for earlier.
702 * At this point we know that the migration attempt cannot
705 remove_migration_ptes(page
, page
);
707 rc
= mapping
->a_ops
->writepage(page
, &wbc
);
709 if (rc
!= AOP_WRITEPAGE_ACTIVATE
)
710 /* unlocked. Relock */
713 return (rc
< 0) ? -EIO
: -EAGAIN
;
717 * Default handling if a filesystem does not provide a migration function.
719 static int fallback_migrate_page(struct address_space
*mapping
,
720 struct page
*newpage
, struct page
*page
, enum migrate_mode mode
)
722 if (PageDirty(page
)) {
723 /* Only writeback pages in full synchronous migration */
724 if (mode
!= MIGRATE_SYNC
)
726 return writeout(mapping
, page
);
730 * Buffers may be managed in a filesystem specific way.
731 * We must have no buffers or drop them.
733 if (page_has_private(page
) &&
734 !try_to_release_page(page
, GFP_KERNEL
))
737 return migrate_page(mapping
, newpage
, page
, mode
);
741 * Move a page to a newly allocated page
742 * The page is locked and all ptes have been successfully removed.
744 * The new page will have replaced the old page if this function
749 * MIGRATEPAGE_SUCCESS - success
751 static int move_to_new_page(struct page
*newpage
, struct page
*page
,
752 enum migrate_mode mode
)
754 struct address_space
*mapping
;
757 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
758 VM_BUG_ON_PAGE(!PageLocked(newpage
), newpage
);
760 mapping
= page_mapping(page
);
762 rc
= migrate_page(mapping
, newpage
, page
, mode
);
763 else if (mapping
->a_ops
->migratepage
)
765 * Most pages have a mapping and most filesystems provide a
766 * migratepage callback. Anonymous pages are part of swap
767 * space which also has its own migratepage callback. This
768 * is the most common path for page migration.
770 rc
= mapping
->a_ops
->migratepage(mapping
, newpage
, page
, mode
);
772 rc
= fallback_migrate_page(mapping
, newpage
, page
, mode
);
775 * When successful, old pagecache page->mapping must be cleared before
776 * page is freed; but stats require that PageAnon be left as PageAnon.
778 if (rc
== MIGRATEPAGE_SUCCESS
) {
780 page
->mapping
= NULL
;
785 static int __unmap_and_move(struct page
*page
, struct page
*newpage
,
786 int force
, enum migrate_mode mode
)
789 int page_was_mapped
= 0;
790 struct anon_vma
*anon_vma
= NULL
;
792 if (!trylock_page(page
)) {
793 if (!force
|| mode
== MIGRATE_ASYNC
)
797 * It's not safe for direct compaction to call lock_page.
798 * For example, during page readahead pages are added locked
799 * to the LRU. Later, when the IO completes the pages are
800 * marked uptodate and unlocked. However, the queueing
801 * could be merging multiple pages for one bio (e.g.
802 * mpage_readpages). If an allocation happens for the
803 * second or third page, the process can end up locking
804 * the same page twice and deadlocking. Rather than
805 * trying to be clever about what pages can be locked,
806 * avoid the use of lock_page for direct compaction
809 if (current
->flags
& PF_MEMALLOC
)
815 if (PageWriteback(page
)) {
817 * Only in the case of a full synchronous migration is it
818 * necessary to wait for PageWriteback. In the async case,
819 * the retry loop is too short and in the sync-light case,
820 * the overhead of stalling is too much
822 if (mode
!= MIGRATE_SYNC
) {
828 wait_on_page_writeback(page
);
832 * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
833 * we cannot notice that anon_vma is freed while we migrates a page.
834 * This get_anon_vma() delays freeing anon_vma pointer until the end
835 * of migration. File cache pages are no problem because of page_lock()
836 * File Caches may use write_page() or lock_page() in migration, then,
837 * just care Anon page here.
839 * Only page_get_anon_vma() understands the subtleties of
840 * getting a hold on an anon_vma from outside one of its mms.
841 * But if we cannot get anon_vma, then we won't need it anyway,
842 * because that implies that the anon page is no longer mapped
843 * (and cannot be remapped so long as we hold the page lock).
845 if (PageAnon(page
) && !PageKsm(page
))
846 anon_vma
= page_get_anon_vma(page
);
849 * Block others from accessing the new page when we get around to
850 * establishing additional references. We are usually the only one
851 * holding a reference to newpage at this point. We used to have a BUG
852 * here if trylock_page(newpage) fails, but would like to allow for
853 * cases where there might be a race with the previous use of newpage.
854 * This is much like races on refcount of oldpage: just don't BUG().
856 if (unlikely(!trylock_page(newpage
)))
859 if (unlikely(isolated_balloon_page(page
))) {
861 * A ballooned page does not need any special attention from
862 * physical to virtual reverse mapping procedures.
863 * Skip any attempt to unmap PTEs or to remap swap cache,
864 * in order to avoid burning cycles at rmap level, and perform
865 * the page migration right away (proteced by page lock).
867 rc
= balloon_page_migrate(newpage
, page
, mode
);
868 goto out_unlock_both
;
872 * Corner case handling:
873 * 1. When a new swap-cache page is read into, it is added to the LRU
874 * and treated as swapcache but it has no rmap yet.
875 * Calling try_to_unmap() against a page->mapping==NULL page will
876 * trigger a BUG. So handle it here.
877 * 2. An orphaned page (see truncate_complete_page) might have
878 * fs-private metadata. The page can be picked up due to memory
879 * offlining. Everywhere else except page reclaim, the page is
880 * invisible to the vm, so the page can not be migrated. So try to
881 * free the metadata, so the page can be freed.
883 if (!page
->mapping
) {
884 VM_BUG_ON_PAGE(PageAnon(page
), page
);
885 if (page_has_private(page
)) {
886 try_to_free_buffers(page
);
887 goto out_unlock_both
;
889 } else if (page_mapped(page
)) {
890 /* Establish migration ptes */
891 VM_BUG_ON_PAGE(PageAnon(page
) && !PageKsm(page
) && !anon_vma
,
894 TTU_MIGRATION
|TTU_IGNORE_MLOCK
|TTU_IGNORE_ACCESS
);
898 if (!page_mapped(page
))
899 rc
= move_to_new_page(newpage
, page
, mode
);
902 remove_migration_ptes(page
,
903 rc
== MIGRATEPAGE_SUCCESS
? newpage
: page
);
906 unlock_page(newpage
);
908 /* Drop an anon_vma reference if we took one */
910 put_anon_vma(anon_vma
);
917 * gcc 4.7 and 4.8 on arm get an ICEs when inlining unmap_and_move(). Work
920 #if (GCC_VERSION >= 40700 && GCC_VERSION < 40900) && defined(CONFIG_ARM)
921 #define ICE_noinline noinline
927 * Obtain the lock on page, remove all ptes and migrate the page
928 * to the newly allocated page in newpage.
930 static ICE_noinline
int unmap_and_move(new_page_t get_new_page
,
931 free_page_t put_new_page
,
932 unsigned long private, struct page
*page
,
933 int force
, enum migrate_mode mode
,
934 enum migrate_reason reason
)
936 int rc
= MIGRATEPAGE_SUCCESS
;
938 struct page
*newpage
;
940 newpage
= get_new_page(page
, private, &result
);
944 if (page_count(page
) == 1) {
945 /* page was freed from under us. So we are done. */
949 if (unlikely(PageTransHuge(page
))) {
951 rc
= split_huge_page(page
);
957 rc
= __unmap_and_move(page
, newpage
, force
, mode
);
958 if (rc
== MIGRATEPAGE_SUCCESS
) {
960 set_page_owner_migrate_reason(newpage
, reason
);
966 * A page that has been migrated has all references
967 * removed and will be freed. A page that has not been
968 * migrated will have kepts its references and be
971 list_del(&page
->lru
);
972 dec_zone_page_state(page
, NR_ISOLATED_ANON
+
973 page_is_file_cache(page
));
974 /* Soft-offlined page shouldn't go through lru cache list */
975 if (reason
== MR_MEMORY_FAILURE
) {
977 if (!test_set_page_hwpoison(page
))
978 num_poisoned_pages_inc();
980 putback_lru_page(page
);
984 * If migration was not successful and there's a freeing callback, use
985 * it. Otherwise, putback_lru_page() will drop the reference grabbed
989 put_new_page(newpage
, private);
990 else if (unlikely(__is_movable_balloon_page(newpage
))) {
991 /* drop our reference, page already in the balloon */
994 putback_lru_page(newpage
);
1000 *result
= page_to_nid(newpage
);
1006 * Counterpart of unmap_and_move_page() for hugepage migration.
1008 * This function doesn't wait the completion of hugepage I/O
1009 * because there is no race between I/O and migration for hugepage.
1010 * Note that currently hugepage I/O occurs only in direct I/O
1011 * where no lock is held and PG_writeback is irrelevant,
1012 * and writeback status of all subpages are counted in the reference
1013 * count of the head page (i.e. if all subpages of a 2MB hugepage are
1014 * under direct I/O, the reference of the head page is 512 and a bit more.)
1015 * This means that when we try to migrate hugepage whose subpages are
1016 * doing direct I/O, some references remain after try_to_unmap() and
1017 * hugepage migration fails without data corruption.
1019 * There is also no race when direct I/O is issued on the page under migration,
1020 * because then pte is replaced with migration swap entry and direct I/O code
1021 * will wait in the page fault for migration to complete.
1023 static int unmap_and_move_huge_page(new_page_t get_new_page
,
1024 free_page_t put_new_page
, unsigned long private,
1025 struct page
*hpage
, int force
,
1026 enum migrate_mode mode
, int reason
)
1030 int page_was_mapped
= 0;
1031 struct page
*new_hpage
;
1032 struct anon_vma
*anon_vma
= NULL
;
1035 * Movability of hugepages depends on architectures and hugepage size.
1036 * This check is necessary because some callers of hugepage migration
1037 * like soft offline and memory hotremove don't walk through page
1038 * tables or check whether the hugepage is pmd-based or not before
1039 * kicking migration.
1041 if (!hugepage_migration_supported(page_hstate(hpage
))) {
1042 putback_active_hugepage(hpage
);
1046 new_hpage
= get_new_page(hpage
, private, &result
);
1050 if (!trylock_page(hpage
)) {
1051 if (!force
|| mode
!= MIGRATE_SYNC
)
1056 if (PageAnon(hpage
))
1057 anon_vma
= page_get_anon_vma(hpage
);
1059 if (unlikely(!trylock_page(new_hpage
)))
1062 if (page_mapped(hpage
)) {
1064 TTU_MIGRATION
|TTU_IGNORE_MLOCK
|TTU_IGNORE_ACCESS
);
1065 page_was_mapped
= 1;
1068 if (!page_mapped(hpage
))
1069 rc
= move_to_new_page(new_hpage
, hpage
, mode
);
1071 if (page_was_mapped
)
1072 remove_migration_ptes(hpage
,
1073 rc
== MIGRATEPAGE_SUCCESS
? new_hpage
: hpage
);
1075 unlock_page(new_hpage
);
1079 put_anon_vma(anon_vma
);
1081 if (rc
== MIGRATEPAGE_SUCCESS
) {
1082 hugetlb_cgroup_migrate(hpage
, new_hpage
);
1083 put_new_page
= NULL
;
1084 set_page_owner_migrate_reason(new_hpage
, reason
);
1090 putback_active_hugepage(hpage
);
1093 * If migration was not successful and there's a freeing callback, use
1094 * it. Otherwise, put_page() will drop the reference grabbed during
1098 put_new_page(new_hpage
, private);
1100 putback_active_hugepage(new_hpage
);
1106 *result
= page_to_nid(new_hpage
);
1112 * migrate_pages - migrate the pages specified in a list, to the free pages
1113 * supplied as the target for the page migration
1115 * @from: The list of pages to be migrated.
1116 * @get_new_page: The function used to allocate free pages to be used
1117 * as the target of the page migration.
1118 * @put_new_page: The function used to free target pages if migration
1119 * fails, or NULL if no special handling is necessary.
1120 * @private: Private data to be passed on to get_new_page()
1121 * @mode: The migration mode that specifies the constraints for
1122 * page migration, if any.
1123 * @reason: The reason for page migration.
1125 * The function returns after 10 attempts or if no pages are movable any more
1126 * because the list has become empty or no retryable pages exist any more.
1127 * The caller should call putback_movable_pages() to return pages to the LRU
1128 * or free list only if ret != 0.
1130 * Returns the number of pages that were not migrated, or an error code.
1132 int migrate_pages(struct list_head
*from
, new_page_t get_new_page
,
1133 free_page_t put_new_page
, unsigned long private,
1134 enum migrate_mode mode
, int reason
)
1138 int nr_succeeded
= 0;
1142 int swapwrite
= current
->flags
& PF_SWAPWRITE
;
1146 current
->flags
|= PF_SWAPWRITE
;
1148 for(pass
= 0; pass
< 10 && retry
; pass
++) {
1151 list_for_each_entry_safe(page
, page2
, from
, lru
) {
1155 rc
= unmap_and_move_huge_page(get_new_page
,
1156 put_new_page
, private, page
,
1157 pass
> 2, mode
, reason
);
1159 rc
= unmap_and_move(get_new_page
, put_new_page
,
1160 private, page
, pass
> 2, mode
,
1169 case MIGRATEPAGE_SUCCESS
:
1174 * Permanent failure (-EBUSY, -ENOSYS, etc.):
1175 * unlike -EAGAIN case, the failed page is
1176 * removed from migration page list and not
1177 * retried in the next outer loop.
1188 count_vm_events(PGMIGRATE_SUCCESS
, nr_succeeded
);
1190 count_vm_events(PGMIGRATE_FAIL
, nr_failed
);
1191 trace_mm_migrate_pages(nr_succeeded
, nr_failed
, mode
, reason
);
1194 current
->flags
&= ~PF_SWAPWRITE
;
1201 * Move a list of individual pages
1203 struct page_to_node
{
1210 static struct page
*new_page_node(struct page
*p
, unsigned long private,
1213 struct page_to_node
*pm
= (struct page_to_node
*)private;
1215 while (pm
->node
!= MAX_NUMNODES
&& pm
->page
!= p
)
1218 if (pm
->node
== MAX_NUMNODES
)
1221 *result
= &pm
->status
;
1224 return alloc_huge_page_node(page_hstate(compound_head(p
)),
1227 return __alloc_pages_node(pm
->node
,
1228 GFP_HIGHUSER_MOVABLE
| __GFP_THISNODE
, 0);
1232 * Move a set of pages as indicated in the pm array. The addr
1233 * field must be set to the virtual address of the page to be moved
1234 * and the node number must contain a valid target node.
1235 * The pm array ends with node = MAX_NUMNODES.
1237 static int do_move_page_to_node_array(struct mm_struct
*mm
,
1238 struct page_to_node
*pm
,
1242 struct page_to_node
*pp
;
1243 LIST_HEAD(pagelist
);
1245 down_read(&mm
->mmap_sem
);
1248 * Build a list of pages to migrate
1250 for (pp
= pm
; pp
->node
!= MAX_NUMNODES
; pp
++) {
1251 struct vm_area_struct
*vma
;
1255 vma
= find_vma(mm
, pp
->addr
);
1256 if (!vma
|| pp
->addr
< vma
->vm_start
|| !vma_migratable(vma
))
1259 /* FOLL_DUMP to ignore special (like zero) pages */
1260 page
= follow_page(vma
, pp
->addr
,
1261 FOLL_GET
| FOLL_SPLIT
| FOLL_DUMP
);
1263 err
= PTR_ERR(page
);
1272 err
= page_to_nid(page
);
1274 if (err
== pp
->node
)
1276 * Node already in the right place
1281 if (page_mapcount(page
) > 1 &&
1285 if (PageHuge(page
)) {
1287 isolate_huge_page(page
, &pagelist
);
1291 err
= isolate_lru_page(page
);
1293 list_add_tail(&page
->lru
, &pagelist
);
1294 inc_zone_page_state(page
, NR_ISOLATED_ANON
+
1295 page_is_file_cache(page
));
1299 * Either remove the duplicate refcount from
1300 * isolate_lru_page() or drop the page ref if it was
1309 if (!list_empty(&pagelist
)) {
1310 err
= migrate_pages(&pagelist
, new_page_node
, NULL
,
1311 (unsigned long)pm
, MIGRATE_SYNC
, MR_SYSCALL
);
1313 putback_movable_pages(&pagelist
);
1316 up_read(&mm
->mmap_sem
);
1321 * Migrate an array of page address onto an array of nodes and fill
1322 * the corresponding array of status.
1324 static int do_pages_move(struct mm_struct
*mm
, nodemask_t task_nodes
,
1325 unsigned long nr_pages
,
1326 const void __user
* __user
*pages
,
1327 const int __user
*nodes
,
1328 int __user
*status
, int flags
)
1330 struct page_to_node
*pm
;
1331 unsigned long chunk_nr_pages
;
1332 unsigned long chunk_start
;
1336 pm
= (struct page_to_node
*)__get_free_page(GFP_KERNEL
);
1343 * Store a chunk of page_to_node array in a page,
1344 * but keep the last one as a marker
1346 chunk_nr_pages
= (PAGE_SIZE
/ sizeof(struct page_to_node
)) - 1;
1348 for (chunk_start
= 0;
1349 chunk_start
< nr_pages
;
1350 chunk_start
+= chunk_nr_pages
) {
1353 if (chunk_start
+ chunk_nr_pages
> nr_pages
)
1354 chunk_nr_pages
= nr_pages
- chunk_start
;
1356 /* fill the chunk pm with addrs and nodes from user-space */
1357 for (j
= 0; j
< chunk_nr_pages
; j
++) {
1358 const void __user
*p
;
1362 if (get_user(p
, pages
+ j
+ chunk_start
))
1364 pm
[j
].addr
= (unsigned long) p
;
1366 if (get_user(node
, nodes
+ j
+ chunk_start
))
1370 if (node
< 0 || node
>= MAX_NUMNODES
)
1373 if (!node_state(node
, N_MEMORY
))
1377 if (!node_isset(node
, task_nodes
))
1383 /* End marker for this chunk */
1384 pm
[chunk_nr_pages
].node
= MAX_NUMNODES
;
1386 /* Migrate this chunk */
1387 err
= do_move_page_to_node_array(mm
, pm
,
1388 flags
& MPOL_MF_MOVE_ALL
);
1392 /* Return status information */
1393 for (j
= 0; j
< chunk_nr_pages
; j
++)
1394 if (put_user(pm
[j
].status
, status
+ j
+ chunk_start
)) {
1402 free_page((unsigned long)pm
);
1408 * Determine the nodes of an array of pages and store it in an array of status.
1410 static void do_pages_stat_array(struct mm_struct
*mm
, unsigned long nr_pages
,
1411 const void __user
**pages
, int *status
)
1415 down_read(&mm
->mmap_sem
);
1417 for (i
= 0; i
< nr_pages
; i
++) {
1418 unsigned long addr
= (unsigned long)(*pages
);
1419 struct vm_area_struct
*vma
;
1423 vma
= find_vma(mm
, addr
);
1424 if (!vma
|| addr
< vma
->vm_start
)
1427 /* FOLL_DUMP to ignore special (like zero) pages */
1428 page
= follow_page(vma
, addr
, FOLL_DUMP
);
1430 err
= PTR_ERR(page
);
1434 err
= page
? page_to_nid(page
) : -ENOENT
;
1442 up_read(&mm
->mmap_sem
);
1446 * Determine the nodes of a user array of pages and store it in
1447 * a user array of status.
1449 static int do_pages_stat(struct mm_struct
*mm
, unsigned long nr_pages
,
1450 const void __user
* __user
*pages
,
1453 #define DO_PAGES_STAT_CHUNK_NR 16
1454 const void __user
*chunk_pages
[DO_PAGES_STAT_CHUNK_NR
];
1455 int chunk_status
[DO_PAGES_STAT_CHUNK_NR
];
1458 unsigned long chunk_nr
;
1460 chunk_nr
= nr_pages
;
1461 if (chunk_nr
> DO_PAGES_STAT_CHUNK_NR
)
1462 chunk_nr
= DO_PAGES_STAT_CHUNK_NR
;
1464 if (copy_from_user(chunk_pages
, pages
, chunk_nr
* sizeof(*chunk_pages
)))
1467 do_pages_stat_array(mm
, chunk_nr
, chunk_pages
, chunk_status
);
1469 if (copy_to_user(status
, chunk_status
, chunk_nr
* sizeof(*status
)))
1474 nr_pages
-= chunk_nr
;
1476 return nr_pages
? -EFAULT
: 0;
1480 * Move a list of pages in the address space of the currently executing
1483 SYSCALL_DEFINE6(move_pages
, pid_t
, pid
, unsigned long, nr_pages
,
1484 const void __user
* __user
*, pages
,
1485 const int __user
*, nodes
,
1486 int __user
*, status
, int, flags
)
1488 const struct cred
*cred
= current_cred(), *tcred
;
1489 struct task_struct
*task
;
1490 struct mm_struct
*mm
;
1492 nodemask_t task_nodes
;
1495 if (flags
& ~(MPOL_MF_MOVE
|MPOL_MF_MOVE_ALL
))
1498 if ((flags
& MPOL_MF_MOVE_ALL
) && !capable(CAP_SYS_NICE
))
1501 /* Find the mm_struct */
1503 task
= pid
? find_task_by_vpid(pid
) : current
;
1508 get_task_struct(task
);
1511 * Check if this process has the right to modify the specified
1512 * process. The right exists if the process has administrative
1513 * capabilities, superuser privileges or the same
1514 * userid as the target process.
1516 tcred
= __task_cred(task
);
1517 if (!uid_eq(cred
->euid
, tcred
->suid
) && !uid_eq(cred
->euid
, tcred
->uid
) &&
1518 !uid_eq(cred
->uid
, tcred
->suid
) && !uid_eq(cred
->uid
, tcred
->uid
) &&
1519 !capable(CAP_SYS_NICE
)) {
1526 err
= security_task_movememory(task
);
1530 task_nodes
= cpuset_mems_allowed(task
);
1531 mm
= get_task_mm(task
);
1532 put_task_struct(task
);
1538 err
= do_pages_move(mm
, task_nodes
, nr_pages
, pages
,
1539 nodes
, status
, flags
);
1541 err
= do_pages_stat(mm
, nr_pages
, pages
, status
);
1547 put_task_struct(task
);
1551 #ifdef CONFIG_NUMA_BALANCING
1553 * Returns true if this is a safe migration target node for misplaced NUMA
1554 * pages. Currently it only checks the watermarks which crude
1556 static bool migrate_balanced_pgdat(struct pglist_data
*pgdat
,
1557 unsigned long nr_migrate_pages
)
1560 for (z
= pgdat
->nr_zones
- 1; z
>= 0; z
--) {
1561 struct zone
*zone
= pgdat
->node_zones
+ z
;
1563 if (!populated_zone(zone
))
1566 if (!zone_reclaimable(zone
))
1569 /* Avoid waking kswapd by allocating pages_to_migrate pages. */
1570 if (!zone_watermark_ok(zone
, 0,
1571 high_wmark_pages(zone
) +
1580 static struct page
*alloc_misplaced_dst_page(struct page
*page
,
1584 int nid
= (int) data
;
1585 struct page
*newpage
;
1587 newpage
= __alloc_pages_node(nid
,
1588 (GFP_HIGHUSER_MOVABLE
|
1589 __GFP_THISNODE
| __GFP_NOMEMALLOC
|
1590 __GFP_NORETRY
| __GFP_NOWARN
) &
1597 * page migration rate limiting control.
1598 * Do not migrate more than @pages_to_migrate in a @migrate_interval_millisecs
1599 * window of time. Default here says do not migrate more than 1280M per second.
1601 static unsigned int migrate_interval_millisecs __read_mostly
= 100;
1602 static unsigned int ratelimit_pages __read_mostly
= 128 << (20 - PAGE_SHIFT
);
1604 /* Returns true if the node is migrate rate-limited after the update */
1605 static bool numamigrate_update_ratelimit(pg_data_t
*pgdat
,
1606 unsigned long nr_pages
)
1609 * Rate-limit the amount of data that is being migrated to a node.
1610 * Optimal placement is no good if the memory bus is saturated and
1611 * all the time is being spent migrating!
1613 if (time_after(jiffies
, pgdat
->numabalancing_migrate_next_window
)) {
1614 spin_lock(&pgdat
->numabalancing_migrate_lock
);
1615 pgdat
->numabalancing_migrate_nr_pages
= 0;
1616 pgdat
->numabalancing_migrate_next_window
= jiffies
+
1617 msecs_to_jiffies(migrate_interval_millisecs
);
1618 spin_unlock(&pgdat
->numabalancing_migrate_lock
);
1620 if (pgdat
->numabalancing_migrate_nr_pages
> ratelimit_pages
) {
1621 trace_mm_numa_migrate_ratelimit(current
, pgdat
->node_id
,
1627 * This is an unlocked non-atomic update so errors are possible.
1628 * The consequences are failing to migrate when we potentiall should
1629 * have which is not severe enough to warrant locking. If it is ever
1630 * a problem, it can be converted to a per-cpu counter.
1632 pgdat
->numabalancing_migrate_nr_pages
+= nr_pages
;
1636 static int numamigrate_isolate_page(pg_data_t
*pgdat
, struct page
*page
)
1640 VM_BUG_ON_PAGE(compound_order(page
) && !PageTransHuge(page
), page
);
1642 /* Avoid migrating to a node that is nearly full */
1643 if (!migrate_balanced_pgdat(pgdat
, 1UL << compound_order(page
)))
1646 if (isolate_lru_page(page
))
1650 * migrate_misplaced_transhuge_page() skips page migration's usual
1651 * check on page_count(), so we must do it here, now that the page
1652 * has been isolated: a GUP pin, or any other pin, prevents migration.
1653 * The expected page count is 3: 1 for page's mapcount and 1 for the
1654 * caller's pin and 1 for the reference taken by isolate_lru_page().
1656 if (PageTransHuge(page
) && page_count(page
) != 3) {
1657 putback_lru_page(page
);
1661 page_lru
= page_is_file_cache(page
);
1662 mod_zone_page_state(page_zone(page
), NR_ISOLATED_ANON
+ page_lru
,
1663 hpage_nr_pages(page
));
1666 * Isolating the page has taken another reference, so the
1667 * caller's reference can be safely dropped without the page
1668 * disappearing underneath us during migration.
1674 bool pmd_trans_migrating(pmd_t pmd
)
1676 struct page
*page
= pmd_page(pmd
);
1677 return PageLocked(page
);
1681 * Attempt to migrate a misplaced page to the specified destination
1682 * node. Caller is expected to have an elevated reference count on
1683 * the page that will be dropped by this function before returning.
1685 int migrate_misplaced_page(struct page
*page
, struct vm_area_struct
*vma
,
1688 pg_data_t
*pgdat
= NODE_DATA(node
);
1691 LIST_HEAD(migratepages
);
1694 * Don't migrate file pages that are mapped in multiple processes
1695 * with execute permissions as they are probably shared libraries.
1697 if (page_mapcount(page
) != 1 && page_is_file_cache(page
) &&
1698 (vma
->vm_flags
& VM_EXEC
))
1702 * Rate-limit the amount of data that is being migrated to a node.
1703 * Optimal placement is no good if the memory bus is saturated and
1704 * all the time is being spent migrating!
1706 if (numamigrate_update_ratelimit(pgdat
, 1))
1709 isolated
= numamigrate_isolate_page(pgdat
, page
);
1713 list_add(&page
->lru
, &migratepages
);
1714 nr_remaining
= migrate_pages(&migratepages
, alloc_misplaced_dst_page
,
1715 NULL
, node
, MIGRATE_ASYNC
,
1718 if (!list_empty(&migratepages
)) {
1719 list_del(&page
->lru
);
1720 dec_zone_page_state(page
, NR_ISOLATED_ANON
+
1721 page_is_file_cache(page
));
1722 putback_lru_page(page
);
1726 count_vm_numa_event(NUMA_PAGE_MIGRATE
);
1727 BUG_ON(!list_empty(&migratepages
));
1734 #endif /* CONFIG_NUMA_BALANCING */
1736 #if defined(CONFIG_NUMA_BALANCING) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
1738 * Migrates a THP to a given target node. page must be locked and is unlocked
1741 int migrate_misplaced_transhuge_page(struct mm_struct
*mm
,
1742 struct vm_area_struct
*vma
,
1743 pmd_t
*pmd
, pmd_t entry
,
1744 unsigned long address
,
1745 struct page
*page
, int node
)
1748 pg_data_t
*pgdat
= NODE_DATA(node
);
1750 struct page
*new_page
= NULL
;
1751 int page_lru
= page_is_file_cache(page
);
1752 unsigned long mmun_start
= address
& HPAGE_PMD_MASK
;
1753 unsigned long mmun_end
= mmun_start
+ HPAGE_PMD_SIZE
;
1757 * Rate-limit the amount of data that is being migrated to a node.
1758 * Optimal placement is no good if the memory bus is saturated and
1759 * all the time is being spent migrating!
1761 if (numamigrate_update_ratelimit(pgdat
, HPAGE_PMD_NR
))
1764 new_page
= alloc_pages_node(node
,
1765 (GFP_TRANSHUGE
| __GFP_THISNODE
) & ~__GFP_RECLAIM
,
1769 prep_transhuge_page(new_page
);
1771 isolated
= numamigrate_isolate_page(pgdat
, page
);
1777 if (mm_tlb_flush_pending(mm
))
1778 flush_tlb_range(vma
, mmun_start
, mmun_end
);
1780 /* Prepare a page as a migration target */
1781 __SetPageLocked(new_page
);
1782 SetPageSwapBacked(new_page
);
1784 /* anon mapping, we can simply copy page->mapping to the new page: */
1785 new_page
->mapping
= page
->mapping
;
1786 new_page
->index
= page
->index
;
1787 migrate_page_copy(new_page
, page
);
1788 WARN_ON(PageLRU(new_page
));
1790 /* Recheck the target PMD */
1791 mmu_notifier_invalidate_range_start(mm
, mmun_start
, mmun_end
);
1792 ptl
= pmd_lock(mm
, pmd
);
1793 if (unlikely(!pmd_same(*pmd
, entry
) || page_count(page
) != 2)) {
1796 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
1798 /* Reverse changes made by migrate_page_copy() */
1799 if (TestClearPageActive(new_page
))
1800 SetPageActive(page
);
1801 if (TestClearPageUnevictable(new_page
))
1802 SetPageUnevictable(page
);
1804 unlock_page(new_page
);
1805 put_page(new_page
); /* Free it */
1807 /* Retake the callers reference and putback on LRU */
1809 putback_lru_page(page
);
1810 mod_zone_page_state(page_zone(page
),
1811 NR_ISOLATED_ANON
+ page_lru
, -HPAGE_PMD_NR
);
1817 entry
= mk_pmd(new_page
, vma
->vm_page_prot
);
1818 entry
= pmd_mkhuge(entry
);
1819 entry
= maybe_pmd_mkwrite(pmd_mkdirty(entry
), vma
);
1822 * Clear the old entry under pagetable lock and establish the new PTE.
1823 * Any parallel GUP will either observe the old page blocking on the
1824 * page lock, block on the page table lock or observe the new page.
1825 * The SetPageUptodate on the new page and page_add_new_anon_rmap
1826 * guarantee the copy is visible before the pagetable update.
1828 flush_cache_range(vma
, mmun_start
, mmun_end
);
1829 page_add_anon_rmap(new_page
, vma
, mmun_start
, true);
1830 pmdp_huge_clear_flush_notify(vma
, mmun_start
, pmd
);
1831 set_pmd_at(mm
, mmun_start
, pmd
, entry
);
1832 flush_tlb_range(vma
, mmun_start
, mmun_end
);
1833 update_mmu_cache_pmd(vma
, address
, &entry
);
1835 if (page_count(page
) != 2) {
1836 set_pmd_at(mm
, mmun_start
, pmd
, orig_entry
);
1837 flush_tlb_range(vma
, mmun_start
, mmun_end
);
1838 mmu_notifier_invalidate_range(mm
, mmun_start
, mmun_end
);
1839 update_mmu_cache_pmd(vma
, address
, &entry
);
1840 page_remove_rmap(new_page
, true);
1844 mlock_migrate_page(new_page
, page
);
1845 page_remove_rmap(page
, true);
1846 set_page_owner_migrate_reason(new_page
, MR_NUMA_MISPLACED
);
1849 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
1851 /* Take an "isolate" reference and put new page on the LRU. */
1853 putback_lru_page(new_page
);
1855 unlock_page(new_page
);
1857 put_page(page
); /* Drop the rmap reference */
1858 put_page(page
); /* Drop the LRU isolation reference */
1860 count_vm_events(PGMIGRATE_SUCCESS
, HPAGE_PMD_NR
);
1861 count_vm_numa_events(NUMA_PAGE_MIGRATE
, HPAGE_PMD_NR
);
1863 mod_zone_page_state(page_zone(page
),
1864 NR_ISOLATED_ANON
+ page_lru
,
1869 count_vm_events(PGMIGRATE_FAIL
, HPAGE_PMD_NR
);
1871 ptl
= pmd_lock(mm
, pmd
);
1872 if (pmd_same(*pmd
, entry
)) {
1873 entry
= pmd_modify(entry
, vma
->vm_page_prot
);
1874 set_pmd_at(mm
, mmun_start
, pmd
, entry
);
1875 update_mmu_cache_pmd(vma
, address
, &entry
);
1884 #endif /* CONFIG_NUMA_BALANCING */
1886 #endif /* CONFIG_NUMA */