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/compaction.h>
35 #include <linux/syscalls.h>
36 #include <linux/hugetlb.h>
37 #include <linux/hugetlb_cgroup.h>
38 #include <linux/gfp.h>
39 #include <linux/balloon_compaction.h>
40 #include <linux/mmu_notifier.h>
41 #include <linux/page_idle.h>
42 #include <linux/page_owner.h>
43 #include <linux/sched/mm.h>
44 #include <linux/ptrace.h>
46 #include <asm/tlbflush.h>
48 #define CREATE_TRACE_POINTS
49 #include <trace/events/migrate.h>
54 * migrate_prep() needs to be called before we start compiling a list of pages
55 * to be migrated using isolate_lru_page(). If scheduling work on other CPUs is
56 * undesirable, use migrate_prep_local()
58 int migrate_prep(void)
61 * Clear the LRU lists so pages can be isolated.
62 * Note that pages may be moved off the LRU after we have
63 * drained them. Those pages will fail to migrate like other
64 * pages that may be busy.
71 /* Do the necessary work of migrate_prep but not if it involves other CPUs */
72 int migrate_prep_local(void)
79 int isolate_movable_page(struct page
*page
, isolate_mode_t mode
)
81 struct address_space
*mapping
;
84 * Avoid burning cycles with pages that are yet under __free_pages(),
85 * or just got freed under us.
87 * In case we 'win' a race for a movable page being freed under us and
88 * raise its refcount preventing __free_pages() from doing its job
89 * the put_page() at the end of this block will take care of
90 * release this page, thus avoiding a nasty leakage.
92 if (unlikely(!get_page_unless_zero(page
)))
96 * Check PageMovable before holding a PG_lock because page's owner
97 * assumes anybody doesn't touch PG_lock of newly allocated page
98 * so unconditionally grapping the lock ruins page's owner side.
100 if (unlikely(!__PageMovable(page
)))
103 * As movable pages are not isolated from LRU lists, concurrent
104 * compaction threads can race against page migration functions
105 * as well as race against the releasing a page.
107 * In order to avoid having an already isolated movable page
108 * being (wrongly) re-isolated while it is under migration,
109 * or to avoid attempting to isolate pages being released,
110 * lets be sure we have the page lock
111 * before proceeding with the movable page isolation steps.
113 if (unlikely(!trylock_page(page
)))
116 if (!PageMovable(page
) || PageIsolated(page
))
117 goto out_no_isolated
;
119 mapping
= page_mapping(page
);
120 VM_BUG_ON_PAGE(!mapping
, page
);
122 if (!mapping
->a_ops
->isolate_page(page
, mode
))
123 goto out_no_isolated
;
125 /* Driver shouldn't use PG_isolated bit of page->flags */
126 WARN_ON_ONCE(PageIsolated(page
));
127 __SetPageIsolated(page
);
140 /* It should be called on page which is PG_movable */
141 void putback_movable_page(struct page
*page
)
143 struct address_space
*mapping
;
145 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
146 VM_BUG_ON_PAGE(!PageMovable(page
), page
);
147 VM_BUG_ON_PAGE(!PageIsolated(page
), page
);
149 mapping
= page_mapping(page
);
150 mapping
->a_ops
->putback_page(page
);
151 __ClearPageIsolated(page
);
155 * Put previously isolated pages back onto the appropriate lists
156 * from where they were once taken off for compaction/migration.
158 * This function shall be used whenever the isolated pageset has been
159 * built from lru, balloon, hugetlbfs page. See isolate_migratepages_range()
160 * and isolate_huge_page().
162 void putback_movable_pages(struct list_head
*l
)
167 list_for_each_entry_safe(page
, page2
, l
, lru
) {
168 if (unlikely(PageHuge(page
))) {
169 putback_active_hugepage(page
);
172 list_del(&page
->lru
);
174 * We isolated non-lru movable page so here we can use
175 * __PageMovable because LRU page's mapping cannot have
176 * PAGE_MAPPING_MOVABLE.
178 if (unlikely(__PageMovable(page
))) {
179 VM_BUG_ON_PAGE(!PageIsolated(page
), page
);
181 if (PageMovable(page
))
182 putback_movable_page(page
);
184 __ClearPageIsolated(page
);
188 dec_node_page_state(page
, NR_ISOLATED_ANON
+
189 page_is_file_cache(page
));
190 putback_lru_page(page
);
196 * Restore a potential migration pte to a working pte entry
198 static bool remove_migration_pte(struct page
*page
, struct vm_area_struct
*vma
,
199 unsigned long addr
, void *old
)
201 struct page_vma_mapped_walk pvmw
= {
205 .flags
= PVMW_SYNC
| PVMW_MIGRATION
,
211 VM_BUG_ON_PAGE(PageTail(page
), page
);
212 while (page_vma_mapped_walk(&pvmw
)) {
216 new = page
- pvmw
.page
->index
+
217 linear_page_index(vma
, pvmw
.address
);
220 pte
= pte_mkold(mk_pte(new, READ_ONCE(vma
->vm_page_prot
)));
221 if (pte_swp_soft_dirty(*pvmw
.pte
))
222 pte
= pte_mksoft_dirty(pte
);
225 * Recheck VMA as permissions can change since migration started
227 entry
= pte_to_swp_entry(*pvmw
.pte
);
228 if (is_write_migration_entry(entry
))
229 pte
= maybe_mkwrite(pte
, vma
);
231 flush_dcache_page(new);
232 #ifdef CONFIG_HUGETLB_PAGE
234 pte
= pte_mkhuge(pte
);
235 pte
= arch_make_huge_pte(pte
, vma
, new, 0);
236 set_huge_pte_at(vma
->vm_mm
, pvmw
.address
, pvmw
.pte
, pte
);
238 hugepage_add_anon_rmap(new, vma
, pvmw
.address
);
240 page_dup_rmap(new, true);
244 set_pte_at(vma
->vm_mm
, pvmw
.address
, pvmw
.pte
, pte
);
247 page_add_anon_rmap(new, vma
, pvmw
.address
, false);
249 page_add_file_rmap(new, false);
251 if (vma
->vm_flags
& VM_LOCKED
&& !PageTransCompound(new))
254 /* No need to invalidate - it was non-present before */
255 update_mmu_cache(vma
, pvmw
.address
, pvmw
.pte
);
262 * Get rid of all migration entries and replace them by
263 * references to the indicated page.
265 void remove_migration_ptes(struct page
*old
, struct page
*new, bool locked
)
267 struct rmap_walk_control rwc
= {
268 .rmap_one
= remove_migration_pte
,
273 rmap_walk_locked(new, &rwc
);
275 rmap_walk(new, &rwc
);
279 * Something used the pte of a page under migration. We need to
280 * get to the page and wait until migration is finished.
281 * When we return from this function the fault will be retried.
283 void __migration_entry_wait(struct mm_struct
*mm
, pte_t
*ptep
,
292 if (!is_swap_pte(pte
))
295 entry
= pte_to_swp_entry(pte
);
296 if (!is_migration_entry(entry
))
299 page
= migration_entry_to_page(entry
);
302 * Once radix-tree replacement of page migration started, page_count
303 * *must* be zero. And, we don't want to call wait_on_page_locked()
304 * against a page without get_page().
305 * So, we use get_page_unless_zero(), here. Even failed, page fault
308 if (!get_page_unless_zero(page
))
310 pte_unmap_unlock(ptep
, ptl
);
311 wait_on_page_locked(page
);
315 pte_unmap_unlock(ptep
, ptl
);
318 void migration_entry_wait(struct mm_struct
*mm
, pmd_t
*pmd
,
319 unsigned long address
)
321 spinlock_t
*ptl
= pte_lockptr(mm
, pmd
);
322 pte_t
*ptep
= pte_offset_map(pmd
, address
);
323 __migration_entry_wait(mm
, ptep
, ptl
);
326 void migration_entry_wait_huge(struct vm_area_struct
*vma
,
327 struct mm_struct
*mm
, pte_t
*pte
)
329 spinlock_t
*ptl
= huge_pte_lockptr(hstate_vma(vma
), mm
, pte
);
330 __migration_entry_wait(mm
, pte
, ptl
);
334 /* Returns true if all buffers are successfully locked */
335 static bool buffer_migrate_lock_buffers(struct buffer_head
*head
,
336 enum migrate_mode mode
)
338 struct buffer_head
*bh
= head
;
340 /* Simple case, sync compaction */
341 if (mode
!= MIGRATE_ASYNC
) {
345 bh
= bh
->b_this_page
;
347 } while (bh
!= head
);
352 /* async case, we cannot block on lock_buffer so use trylock_buffer */
355 if (!trylock_buffer(bh
)) {
357 * We failed to lock the buffer and cannot stall in
358 * async migration. Release the taken locks
360 struct buffer_head
*failed_bh
= bh
;
363 while (bh
!= failed_bh
) {
366 bh
= bh
->b_this_page
;
371 bh
= bh
->b_this_page
;
372 } while (bh
!= head
);
376 static inline bool buffer_migrate_lock_buffers(struct buffer_head
*head
,
377 enum migrate_mode mode
)
381 #endif /* CONFIG_BLOCK */
384 * Replace the page in the mapping.
386 * The number of remaining references must be:
387 * 1 for anonymous pages without a mapping
388 * 2 for pages with a mapping
389 * 3 for pages with a mapping and PagePrivate/PagePrivate2 set.
391 int migrate_page_move_mapping(struct address_space
*mapping
,
392 struct page
*newpage
, struct page
*page
,
393 struct buffer_head
*head
, enum migrate_mode mode
,
396 struct zone
*oldzone
, *newzone
;
398 int expected_count
= 1 + extra_count
;
402 /* Anonymous page without mapping */
403 if (page_count(page
) != expected_count
)
406 /* No turning back from here */
407 newpage
->index
= page
->index
;
408 newpage
->mapping
= page
->mapping
;
409 if (PageSwapBacked(page
))
410 __SetPageSwapBacked(newpage
);
412 return MIGRATEPAGE_SUCCESS
;
415 oldzone
= page_zone(page
);
416 newzone
= page_zone(newpage
);
418 spin_lock_irq(&mapping
->tree_lock
);
420 pslot
= radix_tree_lookup_slot(&mapping
->page_tree
,
423 expected_count
+= 1 + page_has_private(page
);
424 if (page_count(page
) != expected_count
||
425 radix_tree_deref_slot_protected(pslot
, &mapping
->tree_lock
) != page
) {
426 spin_unlock_irq(&mapping
->tree_lock
);
430 if (!page_ref_freeze(page
, expected_count
)) {
431 spin_unlock_irq(&mapping
->tree_lock
);
436 * In the async migration case of moving a page with buffers, lock the
437 * buffers using trylock before the mapping is moved. If the mapping
438 * was moved, we later failed to lock the buffers and could not move
439 * the mapping back due to an elevated page count, we would have to
440 * block waiting on other references to be dropped.
442 if (mode
== MIGRATE_ASYNC
&& head
&&
443 !buffer_migrate_lock_buffers(head
, mode
)) {
444 page_ref_unfreeze(page
, expected_count
);
445 spin_unlock_irq(&mapping
->tree_lock
);
450 * Now we know that no one else is looking at the page:
451 * no turning back from here.
453 newpage
->index
= page
->index
;
454 newpage
->mapping
= page
->mapping
;
455 get_page(newpage
); /* add cache reference */
456 if (PageSwapBacked(page
)) {
457 __SetPageSwapBacked(newpage
);
458 if (PageSwapCache(page
)) {
459 SetPageSwapCache(newpage
);
460 set_page_private(newpage
, page_private(page
));
463 VM_BUG_ON_PAGE(PageSwapCache(page
), page
);
466 /* Move dirty while page refs frozen and newpage not yet exposed */
467 dirty
= PageDirty(page
);
469 ClearPageDirty(page
);
470 SetPageDirty(newpage
);
473 radix_tree_replace_slot(&mapping
->page_tree
, pslot
, newpage
);
476 * Drop cache reference from old page by unfreezing
477 * to one less reference.
478 * We know this isn't the last reference.
480 page_ref_unfreeze(page
, expected_count
- 1);
482 spin_unlock(&mapping
->tree_lock
);
483 /* Leave irq disabled to prevent preemption while updating stats */
486 * If moved to a different zone then also account
487 * the page for that zone. Other VM counters will be
488 * taken care of when we establish references to the
489 * new page and drop references to the old page.
491 * Note that anonymous pages are accounted for
492 * via NR_FILE_PAGES and NR_ANON_MAPPED if they
493 * are mapped to swap space.
495 if (newzone
!= oldzone
) {
496 __dec_node_state(oldzone
->zone_pgdat
, NR_FILE_PAGES
);
497 __inc_node_state(newzone
->zone_pgdat
, NR_FILE_PAGES
);
498 if (PageSwapBacked(page
) && !PageSwapCache(page
)) {
499 __dec_node_state(oldzone
->zone_pgdat
, NR_SHMEM
);
500 __inc_node_state(newzone
->zone_pgdat
, NR_SHMEM
);
502 if (dirty
&& mapping_cap_account_dirty(mapping
)) {
503 __dec_node_state(oldzone
->zone_pgdat
, NR_FILE_DIRTY
);
504 __dec_zone_state(oldzone
, NR_ZONE_WRITE_PENDING
);
505 __inc_node_state(newzone
->zone_pgdat
, NR_FILE_DIRTY
);
506 __inc_zone_state(newzone
, NR_ZONE_WRITE_PENDING
);
511 return MIGRATEPAGE_SUCCESS
;
513 EXPORT_SYMBOL(migrate_page_move_mapping
);
516 * The expected number of remaining references is the same as that
517 * of migrate_page_move_mapping().
519 int migrate_huge_page_move_mapping(struct address_space
*mapping
,
520 struct page
*newpage
, struct page
*page
)
525 spin_lock_irq(&mapping
->tree_lock
);
527 pslot
= radix_tree_lookup_slot(&mapping
->page_tree
,
530 expected_count
= 2 + page_has_private(page
);
531 if (page_count(page
) != expected_count
||
532 radix_tree_deref_slot_protected(pslot
, &mapping
->tree_lock
) != page
) {
533 spin_unlock_irq(&mapping
->tree_lock
);
537 if (!page_ref_freeze(page
, expected_count
)) {
538 spin_unlock_irq(&mapping
->tree_lock
);
542 newpage
->index
= page
->index
;
543 newpage
->mapping
= page
->mapping
;
547 radix_tree_replace_slot(&mapping
->page_tree
, pslot
, newpage
);
549 page_ref_unfreeze(page
, expected_count
- 1);
551 spin_unlock_irq(&mapping
->tree_lock
);
553 return MIGRATEPAGE_SUCCESS
;
557 * Gigantic pages are so large that we do not guarantee that page++ pointer
558 * arithmetic will work across the entire page. We need something more
561 static void __copy_gigantic_page(struct page
*dst
, struct page
*src
,
565 struct page
*dst_base
= dst
;
566 struct page
*src_base
= src
;
568 for (i
= 0; i
< nr_pages
; ) {
570 copy_highpage(dst
, src
);
573 dst
= mem_map_next(dst
, dst_base
, i
);
574 src
= mem_map_next(src
, src_base
, i
);
578 static void copy_huge_page(struct page
*dst
, struct page
*src
)
585 struct hstate
*h
= page_hstate(src
);
586 nr_pages
= pages_per_huge_page(h
);
588 if (unlikely(nr_pages
> MAX_ORDER_NR_PAGES
)) {
589 __copy_gigantic_page(dst
, src
, nr_pages
);
594 BUG_ON(!PageTransHuge(src
));
595 nr_pages
= hpage_nr_pages(src
);
598 for (i
= 0; i
< nr_pages
; i
++) {
600 copy_highpage(dst
+ i
, src
+ i
);
605 * Copy the page to its new location
607 void migrate_page_copy(struct page
*newpage
, struct page
*page
)
611 if (PageHuge(page
) || PageTransHuge(page
))
612 copy_huge_page(newpage
, page
);
614 copy_highpage(newpage
, page
);
617 SetPageError(newpage
);
618 if (PageReferenced(page
))
619 SetPageReferenced(newpage
);
620 if (PageUptodate(page
))
621 SetPageUptodate(newpage
);
622 if (TestClearPageActive(page
)) {
623 VM_BUG_ON_PAGE(PageUnevictable(page
), page
);
624 SetPageActive(newpage
);
625 } else if (TestClearPageUnevictable(page
))
626 SetPageUnevictable(newpage
);
627 if (PageChecked(page
))
628 SetPageChecked(newpage
);
629 if (PageMappedToDisk(page
))
630 SetPageMappedToDisk(newpage
);
632 /* Move dirty on pages not done by migrate_page_move_mapping() */
634 SetPageDirty(newpage
);
636 if (page_is_young(page
))
637 set_page_young(newpage
);
638 if (page_is_idle(page
))
639 set_page_idle(newpage
);
642 * Copy NUMA information to the new page, to prevent over-eager
643 * future migrations of this same page.
645 cpupid
= page_cpupid_xchg_last(page
, -1);
646 page_cpupid_xchg_last(newpage
, cpupid
);
648 ksm_migrate_page(newpage
, page
);
650 * Please do not reorder this without considering how mm/ksm.c's
651 * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache().
653 if (PageSwapCache(page
))
654 ClearPageSwapCache(page
);
655 ClearPagePrivate(page
);
656 set_page_private(page
, 0);
659 * If any waiters have accumulated on the new page then
662 if (PageWriteback(newpage
))
663 end_page_writeback(newpage
);
665 copy_page_owner(page
, newpage
);
667 mem_cgroup_migrate(page
, newpage
);
669 EXPORT_SYMBOL(migrate_page_copy
);
671 /************************************************************
672 * Migration functions
673 ***********************************************************/
676 * Common logic to directly migrate a single LRU page suitable for
677 * pages that do not use PagePrivate/PagePrivate2.
679 * Pages are locked upon entry and exit.
681 int migrate_page(struct address_space
*mapping
,
682 struct page
*newpage
, struct page
*page
,
683 enum migrate_mode mode
)
687 BUG_ON(PageWriteback(page
)); /* Writeback must be complete */
689 rc
= migrate_page_move_mapping(mapping
, newpage
, page
, NULL
, mode
, 0);
691 if (rc
!= MIGRATEPAGE_SUCCESS
)
694 migrate_page_copy(newpage
, page
);
695 return MIGRATEPAGE_SUCCESS
;
697 EXPORT_SYMBOL(migrate_page
);
701 * Migration function for pages with buffers. This function can only be used
702 * if the underlying filesystem guarantees that no other references to "page"
705 int buffer_migrate_page(struct address_space
*mapping
,
706 struct page
*newpage
, struct page
*page
, enum migrate_mode mode
)
708 struct buffer_head
*bh
, *head
;
711 if (!page_has_buffers(page
))
712 return migrate_page(mapping
, newpage
, page
, mode
);
714 head
= page_buffers(page
);
716 rc
= migrate_page_move_mapping(mapping
, newpage
, page
, head
, mode
, 0);
718 if (rc
!= MIGRATEPAGE_SUCCESS
)
722 * In the async case, migrate_page_move_mapping locked the buffers
723 * with an IRQ-safe spinlock held. In the sync case, the buffers
724 * need to be locked now
726 if (mode
!= MIGRATE_ASYNC
)
727 BUG_ON(!buffer_migrate_lock_buffers(head
, mode
));
729 ClearPagePrivate(page
);
730 set_page_private(newpage
, page_private(page
));
731 set_page_private(page
, 0);
737 set_bh_page(bh
, newpage
, bh_offset(bh
));
738 bh
= bh
->b_this_page
;
740 } while (bh
!= head
);
742 SetPagePrivate(newpage
);
744 migrate_page_copy(newpage
, page
);
750 bh
= bh
->b_this_page
;
752 } while (bh
!= head
);
754 return MIGRATEPAGE_SUCCESS
;
756 EXPORT_SYMBOL(buffer_migrate_page
);
760 * Writeback a page to clean the dirty state
762 static int writeout(struct address_space
*mapping
, struct page
*page
)
764 struct writeback_control wbc
= {
765 .sync_mode
= WB_SYNC_NONE
,
768 .range_end
= LLONG_MAX
,
773 if (!mapping
->a_ops
->writepage
)
774 /* No write method for the address space */
777 if (!clear_page_dirty_for_io(page
))
778 /* Someone else already triggered a write */
782 * A dirty page may imply that the underlying filesystem has
783 * the page on some queue. So the page must be clean for
784 * migration. Writeout may mean we loose the lock and the
785 * page state is no longer what we checked for earlier.
786 * At this point we know that the migration attempt cannot
789 remove_migration_ptes(page
, page
, false);
791 rc
= mapping
->a_ops
->writepage(page
, &wbc
);
793 if (rc
!= AOP_WRITEPAGE_ACTIVATE
)
794 /* unlocked. Relock */
797 return (rc
< 0) ? -EIO
: -EAGAIN
;
801 * Default handling if a filesystem does not provide a migration function.
803 static int fallback_migrate_page(struct address_space
*mapping
,
804 struct page
*newpage
, struct page
*page
, enum migrate_mode mode
)
806 if (PageDirty(page
)) {
807 /* Only writeback pages in full synchronous migration */
808 if (mode
!= MIGRATE_SYNC
)
810 return writeout(mapping
, page
);
814 * Buffers may be managed in a filesystem specific way.
815 * We must have no buffers or drop them.
817 if (page_has_private(page
) &&
818 !try_to_release_page(page
, GFP_KERNEL
))
821 return migrate_page(mapping
, newpage
, page
, mode
);
825 * Move a page to a newly allocated page
826 * The page is locked and all ptes have been successfully removed.
828 * The new page will have replaced the old page if this function
833 * MIGRATEPAGE_SUCCESS - success
835 static int move_to_new_page(struct page
*newpage
, struct page
*page
,
836 enum migrate_mode mode
)
838 struct address_space
*mapping
;
840 bool is_lru
= !__PageMovable(page
);
842 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
843 VM_BUG_ON_PAGE(!PageLocked(newpage
), newpage
);
845 mapping
= page_mapping(page
);
847 if (likely(is_lru
)) {
849 rc
= migrate_page(mapping
, newpage
, page
, mode
);
850 else if (mapping
->a_ops
->migratepage
)
852 * Most pages have a mapping and most filesystems
853 * provide a migratepage callback. Anonymous pages
854 * are part of swap space which also has its own
855 * migratepage callback. This is the most common path
856 * for page migration.
858 rc
= mapping
->a_ops
->migratepage(mapping
, newpage
,
861 rc
= fallback_migrate_page(mapping
, newpage
,
865 * In case of non-lru page, it could be released after
866 * isolation step. In that case, we shouldn't try migration.
868 VM_BUG_ON_PAGE(!PageIsolated(page
), page
);
869 if (!PageMovable(page
)) {
870 rc
= MIGRATEPAGE_SUCCESS
;
871 __ClearPageIsolated(page
);
875 rc
= mapping
->a_ops
->migratepage(mapping
, newpage
,
877 WARN_ON_ONCE(rc
== MIGRATEPAGE_SUCCESS
&&
878 !PageIsolated(page
));
882 * When successful, old pagecache page->mapping must be cleared before
883 * page is freed; but stats require that PageAnon be left as PageAnon.
885 if (rc
== MIGRATEPAGE_SUCCESS
) {
886 if (__PageMovable(page
)) {
887 VM_BUG_ON_PAGE(!PageIsolated(page
), page
);
890 * We clear PG_movable under page_lock so any compactor
891 * cannot try to migrate this page.
893 __ClearPageIsolated(page
);
897 * Anonymous and movable page->mapping will be cleard by
898 * free_pages_prepare so don't reset it here for keeping
899 * the type to work PageAnon, for example.
901 if (!PageMappingFlags(page
))
902 page
->mapping
= NULL
;
908 static int __unmap_and_move(struct page
*page
, struct page
*newpage
,
909 int force
, enum migrate_mode mode
)
912 int page_was_mapped
= 0;
913 struct anon_vma
*anon_vma
= NULL
;
914 bool is_lru
= !__PageMovable(page
);
916 if (!trylock_page(page
)) {
917 if (!force
|| mode
== MIGRATE_ASYNC
)
921 * It's not safe for direct compaction to call lock_page.
922 * For example, during page readahead pages are added locked
923 * to the LRU. Later, when the IO completes the pages are
924 * marked uptodate and unlocked. However, the queueing
925 * could be merging multiple pages for one bio (e.g.
926 * mpage_readpages). If an allocation happens for the
927 * second or third page, the process can end up locking
928 * the same page twice and deadlocking. Rather than
929 * trying to be clever about what pages can be locked,
930 * avoid the use of lock_page for direct compaction
933 if (current
->flags
& PF_MEMALLOC
)
939 if (PageWriteback(page
)) {
941 * Only in the case of a full synchronous migration is it
942 * necessary to wait for PageWriteback. In the async case,
943 * the retry loop is too short and in the sync-light case,
944 * the overhead of stalling is too much
946 if (mode
!= MIGRATE_SYNC
) {
952 wait_on_page_writeback(page
);
956 * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
957 * we cannot notice that anon_vma is freed while we migrates a page.
958 * This get_anon_vma() delays freeing anon_vma pointer until the end
959 * of migration. File cache pages are no problem because of page_lock()
960 * File Caches may use write_page() or lock_page() in migration, then,
961 * just care Anon page here.
963 * Only page_get_anon_vma() understands the subtleties of
964 * getting a hold on an anon_vma from outside one of its mms.
965 * But if we cannot get anon_vma, then we won't need it anyway,
966 * because that implies that the anon page is no longer mapped
967 * (and cannot be remapped so long as we hold the page lock).
969 if (PageAnon(page
) && !PageKsm(page
))
970 anon_vma
= page_get_anon_vma(page
);
973 * Block others from accessing the new page when we get around to
974 * establishing additional references. We are usually the only one
975 * holding a reference to newpage at this point. We used to have a BUG
976 * here if trylock_page(newpage) fails, but would like to allow for
977 * cases where there might be a race with the previous use of newpage.
978 * This is much like races on refcount of oldpage: just don't BUG().
980 if (unlikely(!trylock_page(newpage
)))
983 if (unlikely(!is_lru
)) {
984 rc
= move_to_new_page(newpage
, page
, mode
);
985 goto out_unlock_both
;
989 * Corner case handling:
990 * 1. When a new swap-cache page is read into, it is added to the LRU
991 * and treated as swapcache but it has no rmap yet.
992 * Calling try_to_unmap() against a page->mapping==NULL page will
993 * trigger a BUG. So handle it here.
994 * 2. An orphaned page (see truncate_complete_page) might have
995 * fs-private metadata. The page can be picked up due to memory
996 * offlining. Everywhere else except page reclaim, the page is
997 * invisible to the vm, so the page can not be migrated. So try to
998 * free the metadata, so the page can be freed.
1000 if (!page
->mapping
) {
1001 VM_BUG_ON_PAGE(PageAnon(page
), page
);
1002 if (page_has_private(page
)) {
1003 try_to_free_buffers(page
);
1004 goto out_unlock_both
;
1006 } else if (page_mapped(page
)) {
1007 /* Establish migration ptes */
1008 VM_BUG_ON_PAGE(PageAnon(page
) && !PageKsm(page
) && !anon_vma
,
1011 TTU_MIGRATION
|TTU_IGNORE_MLOCK
|TTU_IGNORE_ACCESS
);
1012 page_was_mapped
= 1;
1015 if (!page_mapped(page
))
1016 rc
= move_to_new_page(newpage
, page
, mode
);
1018 if (page_was_mapped
)
1019 remove_migration_ptes(page
,
1020 rc
== MIGRATEPAGE_SUCCESS
? newpage
: page
, false);
1023 unlock_page(newpage
);
1025 /* Drop an anon_vma reference if we took one */
1027 put_anon_vma(anon_vma
);
1031 * If migration is successful, decrease refcount of the newpage
1032 * which will not free the page because new page owner increased
1033 * refcounter. As well, if it is LRU page, add the page to LRU
1036 if (rc
== MIGRATEPAGE_SUCCESS
) {
1037 if (unlikely(__PageMovable(newpage
)))
1040 putback_lru_page(newpage
);
1047 * gcc 4.7 and 4.8 on arm get an ICEs when inlining unmap_and_move(). Work
1050 #if (GCC_VERSION >= 40700 && GCC_VERSION < 40900) && defined(CONFIG_ARM)
1051 #define ICE_noinline noinline
1053 #define ICE_noinline
1057 * Obtain the lock on page, remove all ptes and migrate the page
1058 * to the newly allocated page in newpage.
1060 static ICE_noinline
int unmap_and_move(new_page_t get_new_page
,
1061 free_page_t put_new_page
,
1062 unsigned long private, struct page
*page
,
1063 int force
, enum migrate_mode mode
,
1064 enum migrate_reason reason
)
1066 int rc
= MIGRATEPAGE_SUCCESS
;
1068 struct page
*newpage
;
1070 newpage
= get_new_page(page
, private, &result
);
1074 if (page_count(page
) == 1) {
1075 /* page was freed from under us. So we are done. */
1076 ClearPageActive(page
);
1077 ClearPageUnevictable(page
);
1078 if (unlikely(__PageMovable(page
))) {
1080 if (!PageMovable(page
))
1081 __ClearPageIsolated(page
);
1085 put_new_page(newpage
, private);
1091 if (unlikely(PageTransHuge(page
))) {
1093 rc
= split_huge_page(page
);
1099 rc
= __unmap_and_move(page
, newpage
, force
, mode
);
1100 if (rc
== MIGRATEPAGE_SUCCESS
)
1101 set_page_owner_migrate_reason(newpage
, reason
);
1104 if (rc
!= -EAGAIN
) {
1106 * A page that has been migrated has all references
1107 * removed and will be freed. A page that has not been
1108 * migrated will have kepts its references and be
1111 list_del(&page
->lru
);
1114 * Compaction can migrate also non-LRU pages which are
1115 * not accounted to NR_ISOLATED_*. They can be recognized
1118 if (likely(!__PageMovable(page
)))
1119 dec_node_page_state(page
, NR_ISOLATED_ANON
+
1120 page_is_file_cache(page
));
1124 * If migration is successful, releases reference grabbed during
1125 * isolation. Otherwise, restore the page to right list unless
1128 if (rc
== MIGRATEPAGE_SUCCESS
) {
1130 if (reason
== MR_MEMORY_FAILURE
) {
1132 * Set PG_HWPoison on just freed page
1133 * intentionally. Although it's rather weird,
1134 * it's how HWPoison flag works at the moment.
1136 if (!test_set_page_hwpoison(page
))
1137 num_poisoned_pages_inc();
1140 if (rc
!= -EAGAIN
) {
1141 if (likely(!__PageMovable(page
))) {
1142 putback_lru_page(page
);
1147 if (PageMovable(page
))
1148 putback_movable_page(page
);
1150 __ClearPageIsolated(page
);
1156 put_new_page(newpage
, private);
1165 *result
= page_to_nid(newpage
);
1171 * Counterpart of unmap_and_move_page() for hugepage migration.
1173 * This function doesn't wait the completion of hugepage I/O
1174 * because there is no race between I/O and migration for hugepage.
1175 * Note that currently hugepage I/O occurs only in direct I/O
1176 * where no lock is held and PG_writeback is irrelevant,
1177 * and writeback status of all subpages are counted in the reference
1178 * count of the head page (i.e. if all subpages of a 2MB hugepage are
1179 * under direct I/O, the reference of the head page is 512 and a bit more.)
1180 * This means that when we try to migrate hugepage whose subpages are
1181 * doing direct I/O, some references remain after try_to_unmap() and
1182 * hugepage migration fails without data corruption.
1184 * There is also no race when direct I/O is issued on the page under migration,
1185 * because then pte is replaced with migration swap entry and direct I/O code
1186 * will wait in the page fault for migration to complete.
1188 static int unmap_and_move_huge_page(new_page_t get_new_page
,
1189 free_page_t put_new_page
, unsigned long private,
1190 struct page
*hpage
, int force
,
1191 enum migrate_mode mode
, int reason
)
1195 int page_was_mapped
= 0;
1196 struct page
*new_hpage
;
1197 struct anon_vma
*anon_vma
= NULL
;
1200 * Movability of hugepages depends on architectures and hugepage size.
1201 * This check is necessary because some callers of hugepage migration
1202 * like soft offline and memory hotremove don't walk through page
1203 * tables or check whether the hugepage is pmd-based or not before
1204 * kicking migration.
1206 if (!hugepage_migration_supported(page_hstate(hpage
))) {
1207 putback_active_hugepage(hpage
);
1211 new_hpage
= get_new_page(hpage
, private, &result
);
1215 if (!trylock_page(hpage
)) {
1216 if (!force
|| mode
!= MIGRATE_SYNC
)
1221 if (PageAnon(hpage
))
1222 anon_vma
= page_get_anon_vma(hpage
);
1224 if (unlikely(!trylock_page(new_hpage
)))
1227 if (page_mapped(hpage
)) {
1229 TTU_MIGRATION
|TTU_IGNORE_MLOCK
|TTU_IGNORE_ACCESS
);
1230 page_was_mapped
= 1;
1233 if (!page_mapped(hpage
))
1234 rc
= move_to_new_page(new_hpage
, hpage
, mode
);
1236 if (page_was_mapped
)
1237 remove_migration_ptes(hpage
,
1238 rc
== MIGRATEPAGE_SUCCESS
? new_hpage
: hpage
, false);
1240 unlock_page(new_hpage
);
1244 put_anon_vma(anon_vma
);
1246 if (rc
== MIGRATEPAGE_SUCCESS
) {
1247 hugetlb_cgroup_migrate(hpage
, new_hpage
);
1248 put_new_page
= NULL
;
1249 set_page_owner_migrate_reason(new_hpage
, reason
);
1255 putback_active_hugepage(hpage
);
1256 if (reason
== MR_MEMORY_FAILURE
&& !test_set_page_hwpoison(hpage
))
1257 num_poisoned_pages_inc();
1260 * If migration was not successful and there's a freeing callback, use
1261 * it. Otherwise, put_page() will drop the reference grabbed during
1265 put_new_page(new_hpage
, private);
1267 putback_active_hugepage(new_hpage
);
1273 *result
= page_to_nid(new_hpage
);
1279 * migrate_pages - migrate the pages specified in a list, to the free pages
1280 * supplied as the target for the page migration
1282 * @from: The list of pages to be migrated.
1283 * @get_new_page: The function used to allocate free pages to be used
1284 * as the target of the page migration.
1285 * @put_new_page: The function used to free target pages if migration
1286 * fails, or NULL if no special handling is necessary.
1287 * @private: Private data to be passed on to get_new_page()
1288 * @mode: The migration mode that specifies the constraints for
1289 * page migration, if any.
1290 * @reason: The reason for page migration.
1292 * The function returns after 10 attempts or if no pages are movable any more
1293 * because the list has become empty or no retryable pages exist any more.
1294 * The caller should call putback_movable_pages() to return pages to the LRU
1295 * or free list only if ret != 0.
1297 * Returns the number of pages that were not migrated, or an error code.
1299 int migrate_pages(struct list_head
*from
, new_page_t get_new_page
,
1300 free_page_t put_new_page
, unsigned long private,
1301 enum migrate_mode mode
, int reason
)
1305 int nr_succeeded
= 0;
1309 int swapwrite
= current
->flags
& PF_SWAPWRITE
;
1313 current
->flags
|= PF_SWAPWRITE
;
1315 for(pass
= 0; pass
< 10 && retry
; pass
++) {
1318 list_for_each_entry_safe(page
, page2
, from
, lru
) {
1322 rc
= unmap_and_move_huge_page(get_new_page
,
1323 put_new_page
, private, page
,
1324 pass
> 2, mode
, reason
);
1326 rc
= unmap_and_move(get_new_page
, put_new_page
,
1327 private, page
, pass
> 2, mode
,
1337 case MIGRATEPAGE_SUCCESS
:
1342 * Permanent failure (-EBUSY, -ENOSYS, etc.):
1343 * unlike -EAGAIN case, the failed page is
1344 * removed from migration page list and not
1345 * retried in the next outer loop.
1356 count_vm_events(PGMIGRATE_SUCCESS
, nr_succeeded
);
1358 count_vm_events(PGMIGRATE_FAIL
, nr_failed
);
1359 trace_mm_migrate_pages(nr_succeeded
, nr_failed
, mode
, reason
);
1362 current
->flags
&= ~PF_SWAPWRITE
;
1369 * Move a list of individual pages
1371 struct page_to_node
{
1378 static struct page
*new_page_node(struct page
*p
, unsigned long private,
1381 struct page_to_node
*pm
= (struct page_to_node
*)private;
1383 while (pm
->node
!= MAX_NUMNODES
&& pm
->page
!= p
)
1386 if (pm
->node
== MAX_NUMNODES
)
1389 *result
= &pm
->status
;
1392 return alloc_huge_page_node(page_hstate(compound_head(p
)),
1395 return __alloc_pages_node(pm
->node
,
1396 GFP_HIGHUSER_MOVABLE
| __GFP_THISNODE
, 0);
1400 * Move a set of pages as indicated in the pm array. The addr
1401 * field must be set to the virtual address of the page to be moved
1402 * and the node number must contain a valid target node.
1403 * The pm array ends with node = MAX_NUMNODES.
1405 static int do_move_page_to_node_array(struct mm_struct
*mm
,
1406 struct page_to_node
*pm
,
1410 struct page_to_node
*pp
;
1411 LIST_HEAD(pagelist
);
1413 down_read(&mm
->mmap_sem
);
1416 * Build a list of pages to migrate
1418 for (pp
= pm
; pp
->node
!= MAX_NUMNODES
; pp
++) {
1419 struct vm_area_struct
*vma
;
1423 vma
= find_vma(mm
, pp
->addr
);
1424 if (!vma
|| pp
->addr
< vma
->vm_start
|| !vma_migratable(vma
))
1427 /* FOLL_DUMP to ignore special (like zero) pages */
1428 page
= follow_page(vma
, pp
->addr
,
1429 FOLL_GET
| FOLL_SPLIT
| FOLL_DUMP
);
1431 err
= PTR_ERR(page
);
1440 err
= page_to_nid(page
);
1442 if (err
== pp
->node
)
1444 * Node already in the right place
1449 if (page_mapcount(page
) > 1 &&
1453 if (PageHuge(page
)) {
1455 isolate_huge_page(page
, &pagelist
);
1459 err
= isolate_lru_page(page
);
1461 list_add_tail(&page
->lru
, &pagelist
);
1462 inc_node_page_state(page
, NR_ISOLATED_ANON
+
1463 page_is_file_cache(page
));
1467 * Either remove the duplicate refcount from
1468 * isolate_lru_page() or drop the page ref if it was
1477 if (!list_empty(&pagelist
)) {
1478 err
= migrate_pages(&pagelist
, new_page_node
, NULL
,
1479 (unsigned long)pm
, MIGRATE_SYNC
, MR_SYSCALL
);
1481 putback_movable_pages(&pagelist
);
1484 up_read(&mm
->mmap_sem
);
1489 * Migrate an array of page address onto an array of nodes and fill
1490 * the corresponding array of status.
1492 static int do_pages_move(struct mm_struct
*mm
, nodemask_t task_nodes
,
1493 unsigned long nr_pages
,
1494 const void __user
* __user
*pages
,
1495 const int __user
*nodes
,
1496 int __user
*status
, int flags
)
1498 struct page_to_node
*pm
;
1499 unsigned long chunk_nr_pages
;
1500 unsigned long chunk_start
;
1504 pm
= (struct page_to_node
*)__get_free_page(GFP_KERNEL
);
1511 * Store a chunk of page_to_node array in a page,
1512 * but keep the last one as a marker
1514 chunk_nr_pages
= (PAGE_SIZE
/ sizeof(struct page_to_node
)) - 1;
1516 for (chunk_start
= 0;
1517 chunk_start
< nr_pages
;
1518 chunk_start
+= chunk_nr_pages
) {
1521 if (chunk_start
+ chunk_nr_pages
> nr_pages
)
1522 chunk_nr_pages
= nr_pages
- chunk_start
;
1524 /* fill the chunk pm with addrs and nodes from user-space */
1525 for (j
= 0; j
< chunk_nr_pages
; j
++) {
1526 const void __user
*p
;
1530 if (get_user(p
, pages
+ j
+ chunk_start
))
1532 pm
[j
].addr
= (unsigned long) p
;
1534 if (get_user(node
, nodes
+ j
+ chunk_start
))
1538 if (node
< 0 || node
>= MAX_NUMNODES
)
1541 if (!node_state(node
, N_MEMORY
))
1545 if (!node_isset(node
, task_nodes
))
1551 /* End marker for this chunk */
1552 pm
[chunk_nr_pages
].node
= MAX_NUMNODES
;
1554 /* Migrate this chunk */
1555 err
= do_move_page_to_node_array(mm
, pm
,
1556 flags
& MPOL_MF_MOVE_ALL
);
1560 /* Return status information */
1561 for (j
= 0; j
< chunk_nr_pages
; j
++)
1562 if (put_user(pm
[j
].status
, status
+ j
+ chunk_start
)) {
1570 free_page((unsigned long)pm
);
1576 * Determine the nodes of an array of pages and store it in an array of status.
1578 static void do_pages_stat_array(struct mm_struct
*mm
, unsigned long nr_pages
,
1579 const void __user
**pages
, int *status
)
1583 down_read(&mm
->mmap_sem
);
1585 for (i
= 0; i
< nr_pages
; i
++) {
1586 unsigned long addr
= (unsigned long)(*pages
);
1587 struct vm_area_struct
*vma
;
1591 vma
= find_vma(mm
, addr
);
1592 if (!vma
|| addr
< vma
->vm_start
)
1595 /* FOLL_DUMP to ignore special (like zero) pages */
1596 page
= follow_page(vma
, addr
, FOLL_DUMP
);
1598 err
= PTR_ERR(page
);
1602 err
= page
? page_to_nid(page
) : -ENOENT
;
1610 up_read(&mm
->mmap_sem
);
1614 * Determine the nodes of a user array of pages and store it in
1615 * a user array of status.
1617 static int do_pages_stat(struct mm_struct
*mm
, unsigned long nr_pages
,
1618 const void __user
* __user
*pages
,
1621 #define DO_PAGES_STAT_CHUNK_NR 16
1622 const void __user
*chunk_pages
[DO_PAGES_STAT_CHUNK_NR
];
1623 int chunk_status
[DO_PAGES_STAT_CHUNK_NR
];
1626 unsigned long chunk_nr
;
1628 chunk_nr
= nr_pages
;
1629 if (chunk_nr
> DO_PAGES_STAT_CHUNK_NR
)
1630 chunk_nr
= DO_PAGES_STAT_CHUNK_NR
;
1632 if (copy_from_user(chunk_pages
, pages
, chunk_nr
* sizeof(*chunk_pages
)))
1635 do_pages_stat_array(mm
, chunk_nr
, chunk_pages
, chunk_status
);
1637 if (copy_to_user(status
, chunk_status
, chunk_nr
* sizeof(*status
)))
1642 nr_pages
-= chunk_nr
;
1644 return nr_pages
? -EFAULT
: 0;
1648 * Move a list of pages in the address space of the currently executing
1651 SYSCALL_DEFINE6(move_pages
, pid_t
, pid
, unsigned long, nr_pages
,
1652 const void __user
* __user
*, pages
,
1653 const int __user
*, nodes
,
1654 int __user
*, status
, int, flags
)
1656 struct task_struct
*task
;
1657 struct mm_struct
*mm
;
1659 nodemask_t task_nodes
;
1662 if (flags
& ~(MPOL_MF_MOVE
|MPOL_MF_MOVE_ALL
))
1665 if ((flags
& MPOL_MF_MOVE_ALL
) && !capable(CAP_SYS_NICE
))
1668 /* Find the mm_struct */
1670 task
= pid
? find_task_by_vpid(pid
) : current
;
1675 get_task_struct(task
);
1678 * Check if this process has the right to modify the specified
1679 * process. Use the regular "ptrace_may_access()" checks.
1681 if (!ptrace_may_access(task
, PTRACE_MODE_READ_REALCREDS
)) {
1688 err
= security_task_movememory(task
);
1692 task_nodes
= cpuset_mems_allowed(task
);
1693 mm
= get_task_mm(task
);
1694 put_task_struct(task
);
1700 err
= do_pages_move(mm
, task_nodes
, nr_pages
, pages
,
1701 nodes
, status
, flags
);
1703 err
= do_pages_stat(mm
, nr_pages
, pages
, status
);
1709 put_task_struct(task
);
1713 #ifdef CONFIG_NUMA_BALANCING
1715 * Returns true if this is a safe migration target node for misplaced NUMA
1716 * pages. Currently it only checks the watermarks which crude
1718 static bool migrate_balanced_pgdat(struct pglist_data
*pgdat
,
1719 unsigned long nr_migrate_pages
)
1723 for (z
= pgdat
->nr_zones
- 1; z
>= 0; z
--) {
1724 struct zone
*zone
= pgdat
->node_zones
+ z
;
1726 if (!populated_zone(zone
))
1729 /* Avoid waking kswapd by allocating pages_to_migrate pages. */
1730 if (!zone_watermark_ok(zone
, 0,
1731 high_wmark_pages(zone
) +
1740 static struct page
*alloc_misplaced_dst_page(struct page
*page
,
1744 int nid
= (int) data
;
1745 struct page
*newpage
;
1747 newpage
= __alloc_pages_node(nid
,
1748 (GFP_HIGHUSER_MOVABLE
|
1749 __GFP_THISNODE
| __GFP_NOMEMALLOC
|
1750 __GFP_NORETRY
| __GFP_NOWARN
) &
1757 * page migration rate limiting control.
1758 * Do not migrate more than @pages_to_migrate in a @migrate_interval_millisecs
1759 * window of time. Default here says do not migrate more than 1280M per second.
1761 static unsigned int migrate_interval_millisecs __read_mostly
= 100;
1762 static unsigned int ratelimit_pages __read_mostly
= 128 << (20 - PAGE_SHIFT
);
1764 /* Returns true if the node is migrate rate-limited after the update */
1765 static bool numamigrate_update_ratelimit(pg_data_t
*pgdat
,
1766 unsigned long nr_pages
)
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 (time_after(jiffies
, pgdat
->numabalancing_migrate_next_window
)) {
1774 spin_lock(&pgdat
->numabalancing_migrate_lock
);
1775 pgdat
->numabalancing_migrate_nr_pages
= 0;
1776 pgdat
->numabalancing_migrate_next_window
= jiffies
+
1777 msecs_to_jiffies(migrate_interval_millisecs
);
1778 spin_unlock(&pgdat
->numabalancing_migrate_lock
);
1780 if (pgdat
->numabalancing_migrate_nr_pages
> ratelimit_pages
) {
1781 trace_mm_numa_migrate_ratelimit(current
, pgdat
->node_id
,
1787 * This is an unlocked non-atomic update so errors are possible.
1788 * The consequences are failing to migrate when we potentiall should
1789 * have which is not severe enough to warrant locking. If it is ever
1790 * a problem, it can be converted to a per-cpu counter.
1792 pgdat
->numabalancing_migrate_nr_pages
+= nr_pages
;
1796 static int numamigrate_isolate_page(pg_data_t
*pgdat
, struct page
*page
)
1800 VM_BUG_ON_PAGE(compound_order(page
) && !PageTransHuge(page
), page
);
1802 /* Avoid migrating to a node that is nearly full */
1803 if (!migrate_balanced_pgdat(pgdat
, 1UL << compound_order(page
)))
1806 if (isolate_lru_page(page
))
1810 * migrate_misplaced_transhuge_page() skips page migration's usual
1811 * check on page_count(), so we must do it here, now that the page
1812 * has been isolated: a GUP pin, or any other pin, prevents migration.
1813 * The expected page count is 3: 1 for page's mapcount and 1 for the
1814 * caller's pin and 1 for the reference taken by isolate_lru_page().
1816 if (PageTransHuge(page
) && page_count(page
) != 3) {
1817 putback_lru_page(page
);
1821 page_lru
= page_is_file_cache(page
);
1822 mod_node_page_state(page_pgdat(page
), NR_ISOLATED_ANON
+ page_lru
,
1823 hpage_nr_pages(page
));
1826 * Isolating the page has taken another reference, so the
1827 * caller's reference can be safely dropped without the page
1828 * disappearing underneath us during migration.
1834 bool pmd_trans_migrating(pmd_t pmd
)
1836 struct page
*page
= pmd_page(pmd
);
1837 return PageLocked(page
);
1841 * Attempt to migrate a misplaced page to the specified destination
1842 * node. Caller is expected to have an elevated reference count on
1843 * the page that will be dropped by this function before returning.
1845 int migrate_misplaced_page(struct page
*page
, struct vm_area_struct
*vma
,
1848 pg_data_t
*pgdat
= NODE_DATA(node
);
1851 LIST_HEAD(migratepages
);
1854 * Don't migrate file pages that are mapped in multiple processes
1855 * with execute permissions as they are probably shared libraries.
1857 if (page_mapcount(page
) != 1 && page_is_file_cache(page
) &&
1858 (vma
->vm_flags
& VM_EXEC
))
1862 * Rate-limit the amount of data that is being migrated to a node.
1863 * Optimal placement is no good if the memory bus is saturated and
1864 * all the time is being spent migrating!
1866 if (numamigrate_update_ratelimit(pgdat
, 1))
1869 isolated
= numamigrate_isolate_page(pgdat
, page
);
1873 list_add(&page
->lru
, &migratepages
);
1874 nr_remaining
= migrate_pages(&migratepages
, alloc_misplaced_dst_page
,
1875 NULL
, node
, MIGRATE_ASYNC
,
1878 if (!list_empty(&migratepages
)) {
1879 list_del(&page
->lru
);
1880 dec_node_page_state(page
, NR_ISOLATED_ANON
+
1881 page_is_file_cache(page
));
1882 putback_lru_page(page
);
1886 count_vm_numa_event(NUMA_PAGE_MIGRATE
);
1887 BUG_ON(!list_empty(&migratepages
));
1894 #endif /* CONFIG_NUMA_BALANCING */
1896 #if defined(CONFIG_NUMA_BALANCING) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
1898 * Migrates a THP to a given target node. page must be locked and is unlocked
1901 int migrate_misplaced_transhuge_page(struct mm_struct
*mm
,
1902 struct vm_area_struct
*vma
,
1903 pmd_t
*pmd
, pmd_t entry
,
1904 unsigned long address
,
1905 struct page
*page
, int node
)
1908 pg_data_t
*pgdat
= NODE_DATA(node
);
1910 struct page
*new_page
= NULL
;
1911 int page_lru
= page_is_file_cache(page
);
1912 unsigned long mmun_start
= address
& HPAGE_PMD_MASK
;
1913 unsigned long mmun_end
= mmun_start
+ HPAGE_PMD_SIZE
;
1916 * Rate-limit the amount of data that is being migrated to a node.
1917 * Optimal placement is no good if the memory bus is saturated and
1918 * all the time is being spent migrating!
1920 if (numamigrate_update_ratelimit(pgdat
, HPAGE_PMD_NR
))
1923 new_page
= alloc_pages_node(node
,
1924 (GFP_TRANSHUGE_LIGHT
| __GFP_THISNODE
),
1928 prep_transhuge_page(new_page
);
1930 isolated
= numamigrate_isolate_page(pgdat
, page
);
1936 /* Prepare a page as a migration target */
1937 __SetPageLocked(new_page
);
1938 if (PageSwapBacked(page
))
1939 __SetPageSwapBacked(new_page
);
1941 /* anon mapping, we can simply copy page->mapping to the new page: */
1942 new_page
->mapping
= page
->mapping
;
1943 new_page
->index
= page
->index
;
1944 migrate_page_copy(new_page
, page
);
1945 WARN_ON(PageLRU(new_page
));
1947 /* Recheck the target PMD */
1948 mmu_notifier_invalidate_range_start(mm
, mmun_start
, mmun_end
);
1949 ptl
= pmd_lock(mm
, pmd
);
1950 if (unlikely(!pmd_same(*pmd
, entry
) || !page_ref_freeze(page
, 2))) {
1952 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
1954 /* Reverse changes made by migrate_page_copy() */
1955 if (TestClearPageActive(new_page
))
1956 SetPageActive(page
);
1957 if (TestClearPageUnevictable(new_page
))
1958 SetPageUnevictable(page
);
1960 unlock_page(new_page
);
1961 put_page(new_page
); /* Free it */
1963 /* Retake the callers reference and putback on LRU */
1965 putback_lru_page(page
);
1966 mod_node_page_state(page_pgdat(page
),
1967 NR_ISOLATED_ANON
+ page_lru
, -HPAGE_PMD_NR
);
1972 entry
= mk_huge_pmd(new_page
, vma
->vm_page_prot
);
1973 entry
= maybe_pmd_mkwrite(pmd_mkdirty(entry
), vma
);
1976 * Clear the old entry under pagetable lock and establish the new PTE.
1977 * Any parallel GUP will either observe the old page blocking on the
1978 * page lock, block on the page table lock or observe the new page.
1979 * The SetPageUptodate on the new page and page_add_new_anon_rmap
1980 * guarantee the copy is visible before the pagetable update.
1982 flush_cache_range(vma
, mmun_start
, mmun_end
);
1983 page_add_anon_rmap(new_page
, vma
, mmun_start
, true);
1984 pmdp_huge_clear_flush_notify(vma
, mmun_start
, pmd
);
1985 set_pmd_at(mm
, mmun_start
, pmd
, entry
);
1986 update_mmu_cache_pmd(vma
, address
, &entry
);
1988 page_ref_unfreeze(page
, 2);
1989 mlock_migrate_page(new_page
, page
);
1990 page_remove_rmap(page
, true);
1991 set_page_owner_migrate_reason(new_page
, MR_NUMA_MISPLACED
);
1994 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
1996 /* Take an "isolate" reference and put new page on the LRU. */
1998 putback_lru_page(new_page
);
2000 unlock_page(new_page
);
2002 put_page(page
); /* Drop the rmap reference */
2003 put_page(page
); /* Drop the LRU isolation reference */
2005 count_vm_events(PGMIGRATE_SUCCESS
, HPAGE_PMD_NR
);
2006 count_vm_numa_events(NUMA_PAGE_MIGRATE
, HPAGE_PMD_NR
);
2008 mod_node_page_state(page_pgdat(page
),
2009 NR_ISOLATED_ANON
+ page_lru
,
2014 count_vm_events(PGMIGRATE_FAIL
, HPAGE_PMD_NR
);
2016 ptl
= pmd_lock(mm
, pmd
);
2017 if (pmd_same(*pmd
, entry
)) {
2018 entry
= pmd_modify(entry
, vma
->vm_page_prot
);
2019 set_pmd_at(mm
, mmun_start
, pmd
, entry
);
2020 update_mmu_cache_pmd(vma
, address
, &entry
);
2029 #endif /* CONFIG_NUMA_BALANCING */
2031 #endif /* CONFIG_NUMA */