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>
42 #include <asm/tlbflush.h>
44 #define CREATE_TRACE_POINTS
45 #include <trace/events/migrate.h>
50 * migrate_prep() needs to be called before we start compiling a list of pages
51 * to be migrated using isolate_lru_page(). If scheduling work on other CPUs is
52 * undesirable, use migrate_prep_local()
54 int migrate_prep(void)
57 * Clear the LRU lists so pages can be isolated.
58 * Note that pages may be moved off the LRU after we have
59 * drained them. Those pages will fail to migrate like other
60 * pages that may be busy.
67 /* Do the necessary work of migrate_prep but not if it involves other CPUs */
68 int migrate_prep_local(void)
76 * Put previously isolated pages back onto the appropriate lists
77 * from where they were once taken off for compaction/migration.
79 * This function shall be used whenever the isolated pageset has been
80 * built from lru, balloon, hugetlbfs page. See isolate_migratepages_range()
81 * and isolate_huge_page().
83 void putback_movable_pages(struct list_head
*l
)
88 list_for_each_entry_safe(page
, page2
, l
, lru
) {
89 if (unlikely(PageHuge(page
))) {
90 putback_active_hugepage(page
);
94 dec_zone_page_state(page
, NR_ISOLATED_ANON
+
95 page_is_file_cache(page
));
96 if (unlikely(isolated_balloon_page(page
)))
97 balloon_page_putback(page
);
99 putback_lru_page(page
);
104 * Restore a potential migration pte to a working pte entry
106 static int remove_migration_pte(struct page
*new, struct vm_area_struct
*vma
,
107 unsigned long addr
, void *old
)
109 struct mm_struct
*mm
= vma
->vm_mm
;
115 if (unlikely(PageHuge(new))) {
116 ptep
= huge_pte_offset(mm
, addr
);
119 ptl
= huge_pte_lockptr(hstate_vma(vma
), mm
, ptep
);
121 pmd
= mm_find_pmd(mm
, addr
);
125 ptep
= pte_offset_map(pmd
, addr
);
128 * Peek to check is_swap_pte() before taking ptlock? No, we
129 * can race mremap's move_ptes(), which skips anon_vma lock.
132 ptl
= pte_lockptr(mm
, pmd
);
137 if (!is_swap_pte(pte
))
140 entry
= pte_to_swp_entry(pte
);
142 if (!is_migration_entry(entry
) ||
143 migration_entry_to_page(entry
) != old
)
147 pte
= pte_mkold(mk_pte(new, vma
->vm_page_prot
));
148 if (pte_swp_soft_dirty(*ptep
))
149 pte
= pte_mksoft_dirty(pte
);
151 /* Recheck VMA as permissions can change since migration started */
152 if (is_write_migration_entry(entry
))
153 pte
= maybe_mkwrite(pte
, vma
);
155 #ifdef CONFIG_HUGETLB_PAGE
157 pte
= pte_mkhuge(pte
);
158 pte
= arch_make_huge_pte(pte
, vma
, new, 0);
161 flush_dcache_page(new);
162 set_pte_at(mm
, addr
, ptep
, pte
);
166 hugepage_add_anon_rmap(new, vma
, addr
);
169 } else if (PageAnon(new))
170 page_add_anon_rmap(new, vma
, addr
);
172 page_add_file_rmap(new);
174 if (vma
->vm_flags
& VM_LOCKED
)
177 /* No need to invalidate - it was non-present before */
178 update_mmu_cache(vma
, addr
, ptep
);
180 pte_unmap_unlock(ptep
, ptl
);
186 * Get rid of all migration entries and replace them by
187 * references to the indicated page.
189 static void remove_migration_ptes(struct page
*old
, struct page
*new)
191 struct rmap_walk_control rwc
= {
192 .rmap_one
= remove_migration_pte
,
196 rmap_walk(new, &rwc
);
200 * Something used the pte of a page under migration. We need to
201 * get to the page and wait until migration is finished.
202 * When we return from this function the fault will be retried.
204 void __migration_entry_wait(struct mm_struct
*mm
, pte_t
*ptep
,
213 if (!is_swap_pte(pte
))
216 entry
= pte_to_swp_entry(pte
);
217 if (!is_migration_entry(entry
))
220 page
= migration_entry_to_page(entry
);
223 * Once radix-tree replacement of page migration started, page_count
224 * *must* be zero. And, we don't want to call wait_on_page_locked()
225 * against a page without get_page().
226 * So, we use get_page_unless_zero(), here. Even failed, page fault
229 if (!get_page_unless_zero(page
))
231 pte_unmap_unlock(ptep
, ptl
);
232 wait_on_page_locked(page
);
236 pte_unmap_unlock(ptep
, ptl
);
239 void migration_entry_wait(struct mm_struct
*mm
, pmd_t
*pmd
,
240 unsigned long address
)
242 spinlock_t
*ptl
= pte_lockptr(mm
, pmd
);
243 pte_t
*ptep
= pte_offset_map(pmd
, address
);
244 __migration_entry_wait(mm
, ptep
, ptl
);
247 void migration_entry_wait_huge(struct vm_area_struct
*vma
,
248 struct mm_struct
*mm
, pte_t
*pte
)
250 spinlock_t
*ptl
= huge_pte_lockptr(hstate_vma(vma
), mm
, pte
);
251 __migration_entry_wait(mm
, pte
, ptl
);
255 /* Returns true if all buffers are successfully locked */
256 static bool buffer_migrate_lock_buffers(struct buffer_head
*head
,
257 enum migrate_mode mode
)
259 struct buffer_head
*bh
= head
;
261 /* Simple case, sync compaction */
262 if (mode
!= MIGRATE_ASYNC
) {
266 bh
= bh
->b_this_page
;
268 } while (bh
!= head
);
273 /* async case, we cannot block on lock_buffer so use trylock_buffer */
276 if (!trylock_buffer(bh
)) {
278 * We failed to lock the buffer and cannot stall in
279 * async migration. Release the taken locks
281 struct buffer_head
*failed_bh
= bh
;
284 while (bh
!= failed_bh
) {
287 bh
= bh
->b_this_page
;
292 bh
= bh
->b_this_page
;
293 } while (bh
!= head
);
297 static inline bool buffer_migrate_lock_buffers(struct buffer_head
*head
,
298 enum migrate_mode mode
)
302 #endif /* CONFIG_BLOCK */
305 * Replace the page in the mapping.
307 * The number of remaining references must be:
308 * 1 for anonymous pages without a mapping
309 * 2 for pages with a mapping
310 * 3 for pages with a mapping and PagePrivate/PagePrivate2 set.
312 int migrate_page_move_mapping(struct address_space
*mapping
,
313 struct page
*newpage
, struct page
*page
,
314 struct buffer_head
*head
, enum migrate_mode mode
,
317 struct zone
*oldzone
, *newzone
;
319 int expected_count
= 1 + extra_count
;
323 /* Anonymous page without mapping */
324 if (page_count(page
) != expected_count
)
327 /* No turning back from here */
328 set_page_memcg(newpage
, page_memcg(page
));
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 set_page_memcg(newpage
, page_memcg(page
));
376 newpage
->index
= page
->index
;
377 newpage
->mapping
= page
->mapping
;
378 if (PageSwapBacked(page
))
379 SetPageSwapBacked(newpage
);
381 get_page(newpage
); /* add cache reference */
382 if (PageSwapCache(page
)) {
383 SetPageSwapCache(newpage
);
384 set_page_private(newpage
, page_private(page
));
387 /* Move dirty while page refs frozen and newpage not yet exposed */
388 dirty
= PageDirty(page
);
390 ClearPageDirty(page
);
391 SetPageDirty(newpage
);
394 radix_tree_replace_slot(pslot
, newpage
);
397 * Drop cache reference from old page by unfreezing
398 * to one less reference.
399 * We know this isn't the last reference.
401 page_unfreeze_refs(page
, expected_count
- 1);
403 spin_unlock(&mapping
->tree_lock
);
404 /* Leave irq disabled to prevent preemption while updating stats */
407 * If moved to a different zone then also account
408 * the page for that zone. Other VM counters will be
409 * taken care of when we establish references to the
410 * new page and drop references to the old page.
412 * Note that anonymous pages are accounted for
413 * via NR_FILE_PAGES and NR_ANON_PAGES if they
414 * are mapped to swap space.
416 if (newzone
!= oldzone
) {
417 __dec_zone_state(oldzone
, NR_FILE_PAGES
);
418 __inc_zone_state(newzone
, NR_FILE_PAGES
);
419 if (PageSwapBacked(page
) && !PageSwapCache(page
)) {
420 __dec_zone_state(oldzone
, NR_SHMEM
);
421 __inc_zone_state(newzone
, NR_SHMEM
);
423 if (dirty
&& mapping_cap_account_dirty(mapping
)) {
424 __dec_zone_state(oldzone
, NR_FILE_DIRTY
);
425 __inc_zone_state(newzone
, NR_FILE_DIRTY
);
430 return MIGRATEPAGE_SUCCESS
;
432 EXPORT_SYMBOL(migrate_page_move_mapping
);
435 * The expected number of remaining references is the same as that
436 * of migrate_page_move_mapping().
438 int migrate_huge_page_move_mapping(struct address_space
*mapping
,
439 struct page
*newpage
, struct page
*page
)
444 spin_lock_irq(&mapping
->tree_lock
);
446 pslot
= radix_tree_lookup_slot(&mapping
->page_tree
,
449 expected_count
= 2 + page_has_private(page
);
450 if (page_count(page
) != expected_count
||
451 radix_tree_deref_slot_protected(pslot
, &mapping
->tree_lock
) != page
) {
452 spin_unlock_irq(&mapping
->tree_lock
);
456 if (!page_freeze_refs(page
, expected_count
)) {
457 spin_unlock_irq(&mapping
->tree_lock
);
461 set_page_memcg(newpage
, page_memcg(page
));
462 newpage
->index
= page
->index
;
463 newpage
->mapping
= page
->mapping
;
466 radix_tree_replace_slot(pslot
, newpage
);
468 page_unfreeze_refs(page
, expected_count
- 1);
470 spin_unlock_irq(&mapping
->tree_lock
);
471 return MIGRATEPAGE_SUCCESS
;
475 * Gigantic pages are so large that we do not guarantee that page++ pointer
476 * arithmetic will work across the entire page. We need something more
479 static void __copy_gigantic_page(struct page
*dst
, struct page
*src
,
483 struct page
*dst_base
= dst
;
484 struct page
*src_base
= src
;
486 for (i
= 0; i
< nr_pages
; ) {
488 copy_highpage(dst
, src
);
491 dst
= mem_map_next(dst
, dst_base
, i
);
492 src
= mem_map_next(src
, src_base
, i
);
496 static void copy_huge_page(struct page
*dst
, struct page
*src
)
503 struct hstate
*h
= page_hstate(src
);
504 nr_pages
= pages_per_huge_page(h
);
506 if (unlikely(nr_pages
> MAX_ORDER_NR_PAGES
)) {
507 __copy_gigantic_page(dst
, src
, nr_pages
);
512 BUG_ON(!PageTransHuge(src
));
513 nr_pages
= hpage_nr_pages(src
);
516 for (i
= 0; i
< nr_pages
; i
++) {
518 copy_highpage(dst
+ i
, src
+ i
);
523 * Copy the page to its new location
525 void migrate_page_copy(struct page
*newpage
, struct page
*page
)
529 if (PageHuge(page
) || PageTransHuge(page
))
530 copy_huge_page(newpage
, page
);
532 copy_highpage(newpage
, page
);
535 SetPageError(newpage
);
536 if (PageReferenced(page
))
537 SetPageReferenced(newpage
);
538 if (PageUptodate(page
))
539 SetPageUptodate(newpage
);
540 if (TestClearPageActive(page
)) {
541 VM_BUG_ON_PAGE(PageUnevictable(page
), page
);
542 SetPageActive(newpage
);
543 } else if (TestClearPageUnevictable(page
))
544 SetPageUnevictable(newpage
);
545 if (PageChecked(page
))
546 SetPageChecked(newpage
);
547 if (PageMappedToDisk(page
))
548 SetPageMappedToDisk(newpage
);
550 /* Move dirty on pages not done by migrate_page_move_mapping() */
552 SetPageDirty(newpage
);
554 if (page_is_young(page
))
555 set_page_young(newpage
);
556 if (page_is_idle(page
))
557 set_page_idle(newpage
);
560 * Copy NUMA information to the new page, to prevent over-eager
561 * future migrations of this same page.
563 cpupid
= page_cpupid_xchg_last(page
, -1);
564 page_cpupid_xchg_last(newpage
, cpupid
);
566 ksm_migrate_page(newpage
, page
);
568 * Please do not reorder this without considering how mm/ksm.c's
569 * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache().
571 if (PageSwapCache(page
))
572 ClearPageSwapCache(page
);
573 ClearPagePrivate(page
);
574 set_page_private(page
, 0);
577 * If any waiters have accumulated on the new page then
580 if (PageWriteback(newpage
))
581 end_page_writeback(newpage
);
583 EXPORT_SYMBOL(migrate_page_copy
);
585 /************************************************************
586 * Migration functions
587 ***********************************************************/
590 * Common logic to directly migrate a single page suitable for
591 * pages that do not use PagePrivate/PagePrivate2.
593 * Pages are locked upon entry and exit.
595 int migrate_page(struct address_space
*mapping
,
596 struct page
*newpage
, struct page
*page
,
597 enum migrate_mode mode
)
601 BUG_ON(PageWriteback(page
)); /* Writeback must be complete */
603 rc
= migrate_page_move_mapping(mapping
, newpage
, page
, NULL
, mode
, 0);
605 if (rc
!= MIGRATEPAGE_SUCCESS
)
608 migrate_page_copy(newpage
, page
);
609 return MIGRATEPAGE_SUCCESS
;
611 EXPORT_SYMBOL(migrate_page
);
615 * Migration function for pages with buffers. This function can only be used
616 * if the underlying filesystem guarantees that no other references to "page"
619 int buffer_migrate_page(struct address_space
*mapping
,
620 struct page
*newpage
, struct page
*page
, enum migrate_mode mode
)
622 struct buffer_head
*bh
, *head
;
625 if (!page_has_buffers(page
))
626 return migrate_page(mapping
, newpage
, page
, mode
);
628 head
= page_buffers(page
);
630 rc
= migrate_page_move_mapping(mapping
, newpage
, page
, head
, mode
, 0);
632 if (rc
!= MIGRATEPAGE_SUCCESS
)
636 * In the async case, migrate_page_move_mapping locked the buffers
637 * with an IRQ-safe spinlock held. In the sync case, the buffers
638 * need to be locked now
640 if (mode
!= MIGRATE_ASYNC
)
641 BUG_ON(!buffer_migrate_lock_buffers(head
, mode
));
643 ClearPagePrivate(page
);
644 set_page_private(newpage
, page_private(page
));
645 set_page_private(page
, 0);
651 set_bh_page(bh
, newpage
, bh_offset(bh
));
652 bh
= bh
->b_this_page
;
654 } while (bh
!= head
);
656 SetPagePrivate(newpage
);
658 migrate_page_copy(newpage
, page
);
664 bh
= bh
->b_this_page
;
666 } while (bh
!= head
);
668 return MIGRATEPAGE_SUCCESS
;
670 EXPORT_SYMBOL(buffer_migrate_page
);
674 * Writeback a page to clean the dirty state
676 static int writeout(struct address_space
*mapping
, struct page
*page
)
678 struct writeback_control wbc
= {
679 .sync_mode
= WB_SYNC_NONE
,
682 .range_end
= LLONG_MAX
,
687 if (!mapping
->a_ops
->writepage
)
688 /* No write method for the address space */
691 if (!clear_page_dirty_for_io(page
))
692 /* Someone else already triggered a write */
696 * A dirty page may imply that the underlying filesystem has
697 * the page on some queue. So the page must be clean for
698 * migration. Writeout may mean we loose the lock and the
699 * page state is no longer what we checked for earlier.
700 * At this point we know that the migration attempt cannot
703 remove_migration_ptes(page
, page
);
705 rc
= mapping
->a_ops
->writepage(page
, &wbc
);
707 if (rc
!= AOP_WRITEPAGE_ACTIVATE
)
708 /* unlocked. Relock */
711 return (rc
< 0) ? -EIO
: -EAGAIN
;
715 * Default handling if a filesystem does not provide a migration function.
717 static int fallback_migrate_page(struct address_space
*mapping
,
718 struct page
*newpage
, struct page
*page
, enum migrate_mode mode
)
720 if (PageDirty(page
)) {
721 /* Only writeback pages in full synchronous migration */
722 if (mode
!= MIGRATE_SYNC
)
724 return writeout(mapping
, page
);
728 * Buffers may be managed in a filesystem specific way.
729 * We must have no buffers or drop them.
731 if (page_has_private(page
) &&
732 !try_to_release_page(page
, GFP_KERNEL
))
735 return migrate_page(mapping
, newpage
, page
, mode
);
739 * Move a page to a newly allocated page
740 * The page is locked and all ptes have been successfully removed.
742 * The new page will have replaced the old page if this function
747 * MIGRATEPAGE_SUCCESS - success
749 static int move_to_new_page(struct page
*newpage
, struct page
*page
,
750 enum migrate_mode mode
)
752 struct address_space
*mapping
;
755 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
756 VM_BUG_ON_PAGE(!PageLocked(newpage
), newpage
);
758 mapping
= page_mapping(page
);
760 rc
= migrate_page(mapping
, newpage
, page
, mode
);
761 else if (mapping
->a_ops
->migratepage
)
763 * Most pages have a mapping and most filesystems provide a
764 * migratepage callback. Anonymous pages are part of swap
765 * space which also has its own migratepage callback. This
766 * is the most common path for page migration.
768 rc
= mapping
->a_ops
->migratepage(mapping
, newpage
, page
, mode
);
770 rc
= fallback_migrate_page(mapping
, newpage
, page
, mode
);
773 * When successful, old pagecache page->mapping must be cleared before
774 * page is freed; but stats require that PageAnon be left as PageAnon.
776 if (rc
== MIGRATEPAGE_SUCCESS
) {
777 set_page_memcg(page
, NULL
);
779 page
->mapping
= NULL
;
784 static int __unmap_and_move(struct page
*page
, struct page
*newpage
,
785 int force
, enum migrate_mode mode
)
788 int page_was_mapped
= 0;
789 struct anon_vma
*anon_vma
= NULL
;
791 if (!trylock_page(page
)) {
792 if (!force
|| mode
== MIGRATE_ASYNC
)
796 * It's not safe for direct compaction to call lock_page.
797 * For example, during page readahead pages are added locked
798 * to the LRU. Later, when the IO completes the pages are
799 * marked uptodate and unlocked. However, the queueing
800 * could be merging multiple pages for one bio (e.g.
801 * mpage_readpages). If an allocation happens for the
802 * second or third page, the process can end up locking
803 * the same page twice and deadlocking. Rather than
804 * trying to be clever about what pages can be locked,
805 * avoid the use of lock_page for direct compaction
808 if (current
->flags
& PF_MEMALLOC
)
814 if (PageWriteback(page
)) {
816 * Only in the case of a full synchronous migration is it
817 * necessary to wait for PageWriteback. In the async case,
818 * the retry loop is too short and in the sync-light case,
819 * the overhead of stalling is too much
821 if (mode
!= MIGRATE_SYNC
) {
827 wait_on_page_writeback(page
);
831 * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
832 * we cannot notice that anon_vma is freed while we migrates a page.
833 * This get_anon_vma() delays freeing anon_vma pointer until the end
834 * of migration. File cache pages are no problem because of page_lock()
835 * File Caches may use write_page() or lock_page() in migration, then,
836 * just care Anon page here.
838 * Only page_get_anon_vma() understands the subtleties of
839 * getting a hold on an anon_vma from outside one of its mms.
840 * But if we cannot get anon_vma, then we won't need it anyway,
841 * because that implies that the anon page is no longer mapped
842 * (and cannot be remapped so long as we hold the page lock).
844 if (PageAnon(page
) && !PageKsm(page
))
845 anon_vma
= page_get_anon_vma(page
);
848 * Block others from accessing the new page when we get around to
849 * establishing additional references. We are usually the only one
850 * holding a reference to newpage at this point. We used to have a BUG
851 * here if trylock_page(newpage) fails, but would like to allow for
852 * cases where there might be a race with the previous use of newpage.
853 * This is much like races on refcount of oldpage: just don't BUG().
855 if (unlikely(!trylock_page(newpage
)))
858 if (unlikely(isolated_balloon_page(page
))) {
860 * A ballooned page does not need any special attention from
861 * physical to virtual reverse mapping procedures.
862 * Skip any attempt to unmap PTEs or to remap swap cache,
863 * in order to avoid burning cycles at rmap level, and perform
864 * the page migration right away (proteced by page lock).
866 rc
= balloon_page_migrate(newpage
, page
, mode
);
867 goto out_unlock_both
;
871 * Corner case handling:
872 * 1. When a new swap-cache page is read into, it is added to the LRU
873 * and treated as swapcache but it has no rmap yet.
874 * Calling try_to_unmap() against a page->mapping==NULL page will
875 * trigger a BUG. So handle it here.
876 * 2. An orphaned page (see truncate_complete_page) might have
877 * fs-private metadata. The page can be picked up due to memory
878 * offlining. Everywhere else except page reclaim, the page is
879 * invisible to the vm, so the page can not be migrated. So try to
880 * free the metadata, so the page can be freed.
882 if (!page
->mapping
) {
883 VM_BUG_ON_PAGE(PageAnon(page
), page
);
884 if (page_has_private(page
)) {
885 try_to_free_buffers(page
);
886 goto out_unlock_both
;
888 } else if (page_mapped(page
)) {
889 /* Establish migration ptes */
890 VM_BUG_ON_PAGE(PageAnon(page
) && !PageKsm(page
) && !anon_vma
,
893 TTU_MIGRATION
|TTU_IGNORE_MLOCK
|TTU_IGNORE_ACCESS
);
897 if (!page_mapped(page
))
898 rc
= move_to_new_page(newpage
, page
, mode
);
901 remove_migration_ptes(page
,
902 rc
== MIGRATEPAGE_SUCCESS
? newpage
: page
);
905 unlock_page(newpage
);
907 /* Drop an anon_vma reference if we took one */
909 put_anon_vma(anon_vma
);
916 * gcc 4.7 and 4.8 on arm get an ICEs when inlining unmap_and_move(). Work
919 #if (GCC_VERSION >= 40700 && GCC_VERSION < 40900) && defined(CONFIG_ARM)
920 #define ICE_noinline noinline
926 * Obtain the lock on page, remove all ptes and migrate the page
927 * to the newly allocated page in newpage.
929 static ICE_noinline
int unmap_and_move(new_page_t get_new_page
,
930 free_page_t put_new_page
,
931 unsigned long private, struct page
*page
,
932 int force
, enum migrate_mode mode
,
933 enum migrate_reason reason
)
935 int rc
= MIGRATEPAGE_SUCCESS
;
937 struct page
*newpage
;
939 newpage
= get_new_page(page
, private, &result
);
943 if (page_count(page
) == 1) {
944 /* page was freed from under us. So we are done. */
948 if (unlikely(PageTransHuge(page
)))
949 if (unlikely(split_huge_page(page
)))
952 rc
= __unmap_and_move(page
, newpage
, force
, mode
);
953 if (rc
== MIGRATEPAGE_SUCCESS
)
959 * A page that has been migrated has all references
960 * removed and will be freed. A page that has not been
961 * migrated will have kepts its references and be
964 list_del(&page
->lru
);
965 dec_zone_page_state(page
, NR_ISOLATED_ANON
+
966 page_is_file_cache(page
));
967 /* Soft-offlined page shouldn't go through lru cache list */
968 if (reason
== MR_MEMORY_FAILURE
&& rc
== MIGRATEPAGE_SUCCESS
) {
970 * With this release, we free successfully migrated
971 * page and set PG_HWPoison on just freed page
972 * intentionally. Although it's rather weird, it's how
973 * HWPoison flag works at the moment.
976 if (!test_set_page_hwpoison(page
))
977 num_poisoned_pages_inc();
979 putback_lru_page(page
);
983 * If migration was not successful and there's a freeing callback, use
984 * it. Otherwise, putback_lru_page() will drop the reference grabbed
988 put_new_page(newpage
, private);
989 else if (unlikely(__is_movable_balloon_page(newpage
))) {
990 /* drop our reference, page already in the balloon */
993 putback_lru_page(newpage
);
999 *result
= page_to_nid(newpage
);
1005 * Counterpart of unmap_and_move_page() for hugepage migration.
1007 * This function doesn't wait the completion of hugepage I/O
1008 * because there is no race between I/O and migration for hugepage.
1009 * Note that currently hugepage I/O occurs only in direct I/O
1010 * where no lock is held and PG_writeback is irrelevant,
1011 * and writeback status of all subpages are counted in the reference
1012 * count of the head page (i.e. if all subpages of a 2MB hugepage are
1013 * under direct I/O, the reference of the head page is 512 and a bit more.)
1014 * This means that when we try to migrate hugepage whose subpages are
1015 * doing direct I/O, some references remain after try_to_unmap() and
1016 * hugepage migration fails without data corruption.
1018 * There is also no race when direct I/O is issued on the page under migration,
1019 * because then pte is replaced with migration swap entry and direct I/O code
1020 * will wait in the page fault for migration to complete.
1022 static int unmap_and_move_huge_page(new_page_t get_new_page
,
1023 free_page_t put_new_page
, unsigned long private,
1024 struct page
*hpage
, int force
,
1025 enum migrate_mode mode
)
1029 int page_was_mapped
= 0;
1030 struct page
*new_hpage
;
1031 struct anon_vma
*anon_vma
= NULL
;
1034 * Movability of hugepages depends on architectures and hugepage size.
1035 * This check is necessary because some callers of hugepage migration
1036 * like soft offline and memory hotremove don't walk through page
1037 * tables or check whether the hugepage is pmd-based or not before
1038 * kicking migration.
1040 if (!hugepage_migration_supported(page_hstate(hpage
))) {
1041 putback_active_hugepage(hpage
);
1045 new_hpage
= get_new_page(hpage
, private, &result
);
1049 if (!trylock_page(hpage
)) {
1050 if (!force
|| mode
!= MIGRATE_SYNC
)
1055 if (PageAnon(hpage
))
1056 anon_vma
= page_get_anon_vma(hpage
);
1058 if (unlikely(!trylock_page(new_hpage
)))
1061 if (page_mapped(hpage
)) {
1063 TTU_MIGRATION
|TTU_IGNORE_MLOCK
|TTU_IGNORE_ACCESS
);
1064 page_was_mapped
= 1;
1067 if (!page_mapped(hpage
))
1068 rc
= move_to_new_page(new_hpage
, hpage
, mode
);
1070 if (page_was_mapped
)
1071 remove_migration_ptes(hpage
,
1072 rc
== MIGRATEPAGE_SUCCESS
? new_hpage
: hpage
);
1074 unlock_page(new_hpage
);
1078 put_anon_vma(anon_vma
);
1080 if (rc
== MIGRATEPAGE_SUCCESS
) {
1081 hugetlb_cgroup_migrate(hpage
, new_hpage
);
1082 put_new_page
= NULL
;
1088 putback_active_hugepage(hpage
);
1091 * If migration was not successful and there's a freeing callback, use
1092 * it. Otherwise, put_page() will drop the reference grabbed during
1096 put_new_page(new_hpage
, private);
1098 putback_active_hugepage(new_hpage
);
1104 *result
= page_to_nid(new_hpage
);
1110 * migrate_pages - migrate the pages specified in a list, to the free pages
1111 * supplied as the target for the page migration
1113 * @from: The list of pages to be migrated.
1114 * @get_new_page: The function used to allocate free pages to be used
1115 * as the target of the page migration.
1116 * @put_new_page: The function used to free target pages if migration
1117 * fails, or NULL if no special handling is necessary.
1118 * @private: Private data to be passed on to get_new_page()
1119 * @mode: The migration mode that specifies the constraints for
1120 * page migration, if any.
1121 * @reason: The reason for page migration.
1123 * The function returns after 10 attempts or if no pages are movable any more
1124 * because the list has become empty or no retryable pages exist any more.
1125 * The caller should call putback_movable_pages() to return pages to the LRU
1126 * or free list only if ret != 0.
1128 * Returns the number of pages that were not migrated, or an error code.
1130 int migrate_pages(struct list_head
*from
, new_page_t get_new_page
,
1131 free_page_t put_new_page
, unsigned long private,
1132 enum migrate_mode mode
, int reason
)
1136 int nr_succeeded
= 0;
1140 int swapwrite
= current
->flags
& PF_SWAPWRITE
;
1144 current
->flags
|= PF_SWAPWRITE
;
1146 for(pass
= 0; pass
< 10 && retry
; pass
++) {
1149 list_for_each_entry_safe(page
, page2
, from
, lru
) {
1153 rc
= unmap_and_move_huge_page(get_new_page
,
1154 put_new_page
, private, page
,
1157 rc
= unmap_and_move(get_new_page
, put_new_page
,
1158 private, page
, pass
> 2, mode
,
1167 case MIGRATEPAGE_SUCCESS
:
1172 * Permanent failure (-EBUSY, -ENOSYS, etc.):
1173 * unlike -EAGAIN case, the failed page is
1174 * removed from migration page list and not
1175 * retried in the next outer loop.
1186 count_vm_events(PGMIGRATE_SUCCESS
, nr_succeeded
);
1188 count_vm_events(PGMIGRATE_FAIL
, nr_failed
);
1189 trace_mm_migrate_pages(nr_succeeded
, nr_failed
, mode
, reason
);
1192 current
->flags
&= ~PF_SWAPWRITE
;
1199 * Move a list of individual pages
1201 struct page_to_node
{
1208 static struct page
*new_page_node(struct page
*p
, unsigned long private,
1211 struct page_to_node
*pm
= (struct page_to_node
*)private;
1213 while (pm
->node
!= MAX_NUMNODES
&& pm
->page
!= p
)
1216 if (pm
->node
== MAX_NUMNODES
)
1219 *result
= &pm
->status
;
1222 return alloc_huge_page_node(page_hstate(compound_head(p
)),
1225 return __alloc_pages_node(pm
->node
,
1226 GFP_HIGHUSER_MOVABLE
| __GFP_THISNODE
, 0);
1230 * Move a set of pages as indicated in the pm array. The addr
1231 * field must be set to the virtual address of the page to be moved
1232 * and the node number must contain a valid target node.
1233 * The pm array ends with node = MAX_NUMNODES.
1235 static int do_move_page_to_node_array(struct mm_struct
*mm
,
1236 struct page_to_node
*pm
,
1240 struct page_to_node
*pp
;
1241 LIST_HEAD(pagelist
);
1243 down_read(&mm
->mmap_sem
);
1246 * Build a list of pages to migrate
1248 for (pp
= pm
; pp
->node
!= MAX_NUMNODES
; pp
++) {
1249 struct vm_area_struct
*vma
;
1253 vma
= find_vma(mm
, pp
->addr
);
1254 if (!vma
|| pp
->addr
< vma
->vm_start
|| !vma_migratable(vma
))
1257 /* FOLL_DUMP to ignore special (like zero) pages */
1258 page
= follow_page(vma
, pp
->addr
,
1259 FOLL_GET
| FOLL_SPLIT
| FOLL_DUMP
);
1261 err
= PTR_ERR(page
);
1270 err
= page_to_nid(page
);
1272 if (err
== pp
->node
)
1274 * Node already in the right place
1279 if (page_mapcount(page
) > 1 &&
1283 if (PageHuge(page
)) {
1285 isolate_huge_page(page
, &pagelist
);
1289 err
= isolate_lru_page(page
);
1291 list_add_tail(&page
->lru
, &pagelist
);
1292 inc_zone_page_state(page
, NR_ISOLATED_ANON
+
1293 page_is_file_cache(page
));
1297 * Either remove the duplicate refcount from
1298 * isolate_lru_page() or drop the page ref if it was
1307 if (!list_empty(&pagelist
)) {
1308 err
= migrate_pages(&pagelist
, new_page_node
, NULL
,
1309 (unsigned long)pm
, MIGRATE_SYNC
, MR_SYSCALL
);
1311 putback_movable_pages(&pagelist
);
1314 up_read(&mm
->mmap_sem
);
1319 * Migrate an array of page address onto an array of nodes and fill
1320 * the corresponding array of status.
1322 static int do_pages_move(struct mm_struct
*mm
, nodemask_t task_nodes
,
1323 unsigned long nr_pages
,
1324 const void __user
* __user
*pages
,
1325 const int __user
*nodes
,
1326 int __user
*status
, int flags
)
1328 struct page_to_node
*pm
;
1329 unsigned long chunk_nr_pages
;
1330 unsigned long chunk_start
;
1334 pm
= (struct page_to_node
*)__get_free_page(GFP_KERNEL
);
1341 * Store a chunk of page_to_node array in a page,
1342 * but keep the last one as a marker
1344 chunk_nr_pages
= (PAGE_SIZE
/ sizeof(struct page_to_node
)) - 1;
1346 for (chunk_start
= 0;
1347 chunk_start
< nr_pages
;
1348 chunk_start
+= chunk_nr_pages
) {
1351 if (chunk_start
+ chunk_nr_pages
> nr_pages
)
1352 chunk_nr_pages
= nr_pages
- chunk_start
;
1354 /* fill the chunk pm with addrs and nodes from user-space */
1355 for (j
= 0; j
< chunk_nr_pages
; j
++) {
1356 const void __user
*p
;
1360 if (get_user(p
, pages
+ j
+ chunk_start
))
1362 pm
[j
].addr
= (unsigned long) p
;
1364 if (get_user(node
, nodes
+ j
+ chunk_start
))
1368 if (node
< 0 || node
>= MAX_NUMNODES
)
1371 if (!node_state(node
, N_MEMORY
))
1375 if (!node_isset(node
, task_nodes
))
1381 /* End marker for this chunk */
1382 pm
[chunk_nr_pages
].node
= MAX_NUMNODES
;
1384 /* Migrate this chunk */
1385 err
= do_move_page_to_node_array(mm
, pm
,
1386 flags
& MPOL_MF_MOVE_ALL
);
1390 /* Return status information */
1391 for (j
= 0; j
< chunk_nr_pages
; j
++)
1392 if (put_user(pm
[j
].status
, status
+ j
+ chunk_start
)) {
1400 free_page((unsigned long)pm
);
1406 * Determine the nodes of an array of pages and store it in an array of status.
1408 static void do_pages_stat_array(struct mm_struct
*mm
, unsigned long nr_pages
,
1409 const void __user
**pages
, int *status
)
1413 down_read(&mm
->mmap_sem
);
1415 for (i
= 0; i
< nr_pages
; i
++) {
1416 unsigned long addr
= (unsigned long)(*pages
);
1417 struct vm_area_struct
*vma
;
1421 vma
= find_vma(mm
, addr
);
1422 if (!vma
|| addr
< vma
->vm_start
)
1425 /* FOLL_DUMP to ignore special (like zero) pages */
1426 page
= follow_page(vma
, addr
, FOLL_DUMP
);
1428 err
= PTR_ERR(page
);
1432 err
= page
? page_to_nid(page
) : -ENOENT
;
1440 up_read(&mm
->mmap_sem
);
1444 * Determine the nodes of a user array of pages and store it in
1445 * a user array of status.
1447 static int do_pages_stat(struct mm_struct
*mm
, unsigned long nr_pages
,
1448 const void __user
* __user
*pages
,
1451 #define DO_PAGES_STAT_CHUNK_NR 16
1452 const void __user
*chunk_pages
[DO_PAGES_STAT_CHUNK_NR
];
1453 int chunk_status
[DO_PAGES_STAT_CHUNK_NR
];
1456 unsigned long chunk_nr
;
1458 chunk_nr
= nr_pages
;
1459 if (chunk_nr
> DO_PAGES_STAT_CHUNK_NR
)
1460 chunk_nr
= DO_PAGES_STAT_CHUNK_NR
;
1462 if (copy_from_user(chunk_pages
, pages
, chunk_nr
* sizeof(*chunk_pages
)))
1465 do_pages_stat_array(mm
, chunk_nr
, chunk_pages
, chunk_status
);
1467 if (copy_to_user(status
, chunk_status
, chunk_nr
* sizeof(*status
)))
1472 nr_pages
-= chunk_nr
;
1474 return nr_pages
? -EFAULT
: 0;
1478 * Move a list of pages in the address space of the currently executing
1481 SYSCALL_DEFINE6(move_pages
, pid_t
, pid
, unsigned long, nr_pages
,
1482 const void __user
* __user
*, pages
,
1483 const int __user
*, nodes
,
1484 int __user
*, status
, int, flags
)
1486 const struct cred
*cred
= current_cred(), *tcred
;
1487 struct task_struct
*task
;
1488 struct mm_struct
*mm
;
1490 nodemask_t task_nodes
;
1493 if (flags
& ~(MPOL_MF_MOVE
|MPOL_MF_MOVE_ALL
))
1496 if ((flags
& MPOL_MF_MOVE_ALL
) && !capable(CAP_SYS_NICE
))
1499 /* Find the mm_struct */
1501 task
= pid
? find_task_by_vpid(pid
) : current
;
1506 get_task_struct(task
);
1509 * Check if this process has the right to modify the specified
1510 * process. The right exists if the process has administrative
1511 * capabilities, superuser privileges or the same
1512 * userid as the target process.
1514 tcred
= __task_cred(task
);
1515 if (!uid_eq(cred
->euid
, tcred
->suid
) && !uid_eq(cred
->euid
, tcred
->uid
) &&
1516 !uid_eq(cred
->uid
, tcred
->suid
) && !uid_eq(cred
->uid
, tcred
->uid
) &&
1517 !capable(CAP_SYS_NICE
)) {
1524 err
= security_task_movememory(task
);
1528 task_nodes
= cpuset_mems_allowed(task
);
1529 mm
= get_task_mm(task
);
1530 put_task_struct(task
);
1536 err
= do_pages_move(mm
, task_nodes
, nr_pages
, pages
,
1537 nodes
, status
, flags
);
1539 err
= do_pages_stat(mm
, nr_pages
, pages
, status
);
1545 put_task_struct(task
);
1549 #ifdef CONFIG_NUMA_BALANCING
1551 * Returns true if this is a safe migration target node for misplaced NUMA
1552 * pages. Currently it only checks the watermarks which crude
1554 static bool migrate_balanced_pgdat(struct pglist_data
*pgdat
,
1555 unsigned long nr_migrate_pages
)
1558 for (z
= pgdat
->nr_zones
- 1; z
>= 0; z
--) {
1559 struct zone
*zone
= pgdat
->node_zones
+ z
;
1561 if (!populated_zone(zone
))
1564 if (!zone_reclaimable(zone
))
1567 /* Avoid waking kswapd by allocating pages_to_migrate pages. */
1568 if (!zone_watermark_ok(zone
, 0,
1569 high_wmark_pages(zone
) +
1578 static struct page
*alloc_misplaced_dst_page(struct page
*page
,
1582 int nid
= (int) data
;
1583 struct page
*newpage
;
1585 newpage
= __alloc_pages_node(nid
,
1586 (GFP_HIGHUSER_MOVABLE
|
1587 __GFP_THISNODE
| __GFP_NOMEMALLOC
|
1588 __GFP_NORETRY
| __GFP_NOWARN
) &
1595 * page migration rate limiting control.
1596 * Do not migrate more than @pages_to_migrate in a @migrate_interval_millisecs
1597 * window of time. Default here says do not migrate more than 1280M per second.
1599 static unsigned int migrate_interval_millisecs __read_mostly
= 100;
1600 static unsigned int ratelimit_pages __read_mostly
= 128 << (20 - PAGE_SHIFT
);
1602 /* Returns true if the node is migrate rate-limited after the update */
1603 static bool numamigrate_update_ratelimit(pg_data_t
*pgdat
,
1604 unsigned long nr_pages
)
1607 * Rate-limit the amount of data that is being migrated to a node.
1608 * Optimal placement is no good if the memory bus is saturated and
1609 * all the time is being spent migrating!
1611 if (time_after(jiffies
, pgdat
->numabalancing_migrate_next_window
)) {
1612 spin_lock(&pgdat
->numabalancing_migrate_lock
);
1613 pgdat
->numabalancing_migrate_nr_pages
= 0;
1614 pgdat
->numabalancing_migrate_next_window
= jiffies
+
1615 msecs_to_jiffies(migrate_interval_millisecs
);
1616 spin_unlock(&pgdat
->numabalancing_migrate_lock
);
1618 if (pgdat
->numabalancing_migrate_nr_pages
> ratelimit_pages
) {
1619 trace_mm_numa_migrate_ratelimit(current
, pgdat
->node_id
,
1625 * This is an unlocked non-atomic update so errors are possible.
1626 * The consequences are failing to migrate when we potentiall should
1627 * have which is not severe enough to warrant locking. If it is ever
1628 * a problem, it can be converted to a per-cpu counter.
1630 pgdat
->numabalancing_migrate_nr_pages
+= nr_pages
;
1634 static int numamigrate_isolate_page(pg_data_t
*pgdat
, struct page
*page
)
1638 VM_BUG_ON_PAGE(compound_order(page
) && !PageTransHuge(page
), page
);
1640 /* Avoid migrating to a node that is nearly full */
1641 if (!migrate_balanced_pgdat(pgdat
, 1UL << compound_order(page
)))
1644 if (isolate_lru_page(page
))
1648 * migrate_misplaced_transhuge_page() skips page migration's usual
1649 * check on page_count(), so we must do it here, now that the page
1650 * has been isolated: a GUP pin, or any other pin, prevents migration.
1651 * The expected page count is 3: 1 for page's mapcount and 1 for the
1652 * caller's pin and 1 for the reference taken by isolate_lru_page().
1654 if (PageTransHuge(page
) && page_count(page
) != 3) {
1655 putback_lru_page(page
);
1659 page_lru
= page_is_file_cache(page
);
1660 mod_zone_page_state(page_zone(page
), NR_ISOLATED_ANON
+ page_lru
,
1661 hpage_nr_pages(page
));
1664 * Isolating the page has taken another reference, so the
1665 * caller's reference can be safely dropped without the page
1666 * disappearing underneath us during migration.
1672 bool pmd_trans_migrating(pmd_t pmd
)
1674 struct page
*page
= pmd_page(pmd
);
1675 return PageLocked(page
);
1679 * Attempt to migrate a misplaced page to the specified destination
1680 * node. Caller is expected to have an elevated reference count on
1681 * the page that will be dropped by this function before returning.
1683 int migrate_misplaced_page(struct page
*page
, struct vm_area_struct
*vma
,
1686 pg_data_t
*pgdat
= NODE_DATA(node
);
1689 LIST_HEAD(migratepages
);
1692 * Don't migrate file pages that are mapped in multiple processes
1693 * with execute permissions as they are probably shared libraries.
1695 if (page_mapcount(page
) != 1 && page_is_file_cache(page
) &&
1696 (vma
->vm_flags
& VM_EXEC
))
1700 * Rate-limit the amount of data that is being migrated to a node.
1701 * Optimal placement is no good if the memory bus is saturated and
1702 * all the time is being spent migrating!
1704 if (numamigrate_update_ratelimit(pgdat
, 1))
1707 isolated
= numamigrate_isolate_page(pgdat
, page
);
1711 list_add(&page
->lru
, &migratepages
);
1712 nr_remaining
= migrate_pages(&migratepages
, alloc_misplaced_dst_page
,
1713 NULL
, node
, MIGRATE_ASYNC
,
1716 if (!list_empty(&migratepages
)) {
1717 list_del(&page
->lru
);
1718 dec_zone_page_state(page
, NR_ISOLATED_ANON
+
1719 page_is_file_cache(page
));
1720 putback_lru_page(page
);
1724 count_vm_numa_event(NUMA_PAGE_MIGRATE
);
1725 BUG_ON(!list_empty(&migratepages
));
1732 #endif /* CONFIG_NUMA_BALANCING */
1734 #if defined(CONFIG_NUMA_BALANCING) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
1736 * Migrates a THP to a given target node. page must be locked and is unlocked
1739 int migrate_misplaced_transhuge_page(struct mm_struct
*mm
,
1740 struct vm_area_struct
*vma
,
1741 pmd_t
*pmd
, pmd_t entry
,
1742 unsigned long address
,
1743 struct page
*page
, int node
)
1746 pg_data_t
*pgdat
= NODE_DATA(node
);
1748 struct page
*new_page
= NULL
;
1749 int page_lru
= page_is_file_cache(page
);
1750 unsigned long mmun_start
= address
& HPAGE_PMD_MASK
;
1751 unsigned long mmun_end
= mmun_start
+ HPAGE_PMD_SIZE
;
1755 * Rate-limit the amount of data that is being migrated to a node.
1756 * Optimal placement is no good if the memory bus is saturated and
1757 * all the time is being spent migrating!
1759 if (numamigrate_update_ratelimit(pgdat
, HPAGE_PMD_NR
))
1762 new_page
= alloc_pages_node(node
,
1763 (GFP_TRANSHUGE
| __GFP_THISNODE
) & ~__GFP_RECLAIM
,
1768 isolated
= numamigrate_isolate_page(pgdat
, page
);
1774 if (mm_tlb_flush_pending(mm
))
1775 flush_tlb_range(vma
, mmun_start
, mmun_end
);
1777 /* Prepare a page as a migration target */
1778 __set_page_locked(new_page
);
1779 SetPageSwapBacked(new_page
);
1781 /* anon mapping, we can simply copy page->mapping to the new page: */
1782 new_page
->mapping
= page
->mapping
;
1783 new_page
->index
= page
->index
;
1784 migrate_page_copy(new_page
, page
);
1785 WARN_ON(PageLRU(new_page
));
1787 /* Recheck the target PMD */
1788 mmu_notifier_invalidate_range_start(mm
, mmun_start
, mmun_end
);
1789 ptl
= pmd_lock(mm
, pmd
);
1790 if (unlikely(!pmd_same(*pmd
, entry
) || page_count(page
) != 2)) {
1793 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
1795 /* Reverse changes made by migrate_page_copy() */
1796 if (TestClearPageActive(new_page
))
1797 SetPageActive(page
);
1798 if (TestClearPageUnevictable(new_page
))
1799 SetPageUnevictable(page
);
1801 unlock_page(new_page
);
1802 put_page(new_page
); /* Free it */
1804 /* Retake the callers reference and putback on LRU */
1806 putback_lru_page(page
);
1807 mod_zone_page_state(page_zone(page
),
1808 NR_ISOLATED_ANON
+ page_lru
, -HPAGE_PMD_NR
);
1814 entry
= mk_pmd(new_page
, vma
->vm_page_prot
);
1815 entry
= pmd_mkhuge(entry
);
1816 entry
= maybe_pmd_mkwrite(pmd_mkdirty(entry
), vma
);
1819 * Clear the old entry under pagetable lock and establish the new PTE.
1820 * Any parallel GUP will either observe the old page blocking on the
1821 * page lock, block on the page table lock or observe the new page.
1822 * The SetPageUptodate on the new page and page_add_new_anon_rmap
1823 * guarantee the copy is visible before the pagetable update.
1825 flush_cache_range(vma
, mmun_start
, mmun_end
);
1826 page_add_anon_rmap(new_page
, vma
, mmun_start
);
1827 pmdp_huge_clear_flush_notify(vma
, mmun_start
, pmd
);
1828 set_pmd_at(mm
, mmun_start
, pmd
, entry
);
1829 flush_tlb_range(vma
, mmun_start
, mmun_end
);
1830 update_mmu_cache_pmd(vma
, address
, &entry
);
1832 if (page_count(page
) != 2) {
1833 set_pmd_at(mm
, mmun_start
, pmd
, orig_entry
);
1834 flush_tlb_range(vma
, mmun_start
, mmun_end
);
1835 mmu_notifier_invalidate_range(mm
, mmun_start
, mmun_end
);
1836 update_mmu_cache_pmd(vma
, address
, &entry
);
1837 page_remove_rmap(new_page
);
1841 mlock_migrate_page(new_page
, page
);
1842 set_page_memcg(new_page
, page_memcg(page
));
1843 set_page_memcg(page
, NULL
);
1844 page_remove_rmap(page
);
1847 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
1849 /* Take an "isolate" reference and put new page on the LRU. */
1851 putback_lru_page(new_page
);
1853 unlock_page(new_page
);
1855 put_page(page
); /* Drop the rmap reference */
1856 put_page(page
); /* Drop the LRU isolation reference */
1858 count_vm_events(PGMIGRATE_SUCCESS
, HPAGE_PMD_NR
);
1859 count_vm_numa_events(NUMA_PAGE_MIGRATE
, HPAGE_PMD_NR
);
1861 mod_zone_page_state(page_zone(page
),
1862 NR_ISOLATED_ANON
+ page_lru
,
1867 count_vm_events(PGMIGRATE_FAIL
, HPAGE_PMD_NR
);
1869 ptl
= pmd_lock(mm
, pmd
);
1870 if (pmd_same(*pmd
, entry
)) {
1871 entry
= pmd_modify(entry
, vma
->vm_page_prot
);
1872 set_pmd_at(mm
, mmun_start
, pmd
, entry
);
1873 update_mmu_cache_pmd(vma
, address
, &entry
);
1882 #endif /* CONFIG_NUMA_BALANCING */
1884 #endif /* CONFIG_NUMA */