1 // SPDX-License-Identifier: GPL-2.0
3 * Memory Migration functionality - linux/mm/migrate.c
5 * Copyright (C) 2006 Silicon Graphics, Inc., Christoph Lameter
7 * Page migration was first developed in the context of the memory hotplug
8 * project. The main authors of the migration code are:
10 * IWAMOTO Toshihiro <iwamoto@valinux.co.jp>
11 * Hirokazu Takahashi <taka@valinux.co.jp>
12 * Dave Hansen <haveblue@us.ibm.com>
16 #include <linux/migrate.h>
17 #include <linux/export.h>
18 #include <linux/swap.h>
19 #include <linux/swapops.h>
20 #include <linux/pagemap.h>
21 #include <linux/buffer_head.h>
22 #include <linux/mm_inline.h>
23 #include <linux/nsproxy.h>
24 #include <linux/pagevec.h>
25 #include <linux/ksm.h>
26 #include <linux/rmap.h>
27 #include <linux/topology.h>
28 #include <linux/cpu.h>
29 #include <linux/cpuset.h>
30 #include <linux/writeback.h>
31 #include <linux/mempolicy.h>
32 #include <linux/vmalloc.h>
33 #include <linux/security.h>
34 #include <linux/backing-dev.h>
35 #include <linux/compaction.h>
36 #include <linux/syscalls.h>
37 #include <linux/compat.h>
38 #include <linux/hugetlb.h>
39 #include <linux/hugetlb_cgroup.h>
40 #include <linux/gfp.h>
41 #include <linux/pfn_t.h>
42 #include <linux/memremap.h>
43 #include <linux/userfaultfd_k.h>
44 #include <linux/balloon_compaction.h>
45 #include <linux/mmu_notifier.h>
46 #include <linux/page_idle.h>
47 #include <linux/page_owner.h>
48 #include <linux/sched/mm.h>
49 #include <linux/ptrace.h>
51 #include <asm/tlbflush.h>
53 #define CREATE_TRACE_POINTS
54 #include <trace/events/migrate.h>
59 * migrate_prep() needs to be called before we start compiling a list of pages
60 * to be migrated using isolate_lru_page(). If scheduling work on other CPUs is
61 * undesirable, use migrate_prep_local()
63 int migrate_prep(void)
66 * Clear the LRU lists so pages can be isolated.
67 * Note that pages may be moved off the LRU after we have
68 * drained them. Those pages will fail to migrate like other
69 * pages that may be busy.
76 /* Do the necessary work of migrate_prep but not if it involves other CPUs */
77 int migrate_prep_local(void)
84 int isolate_movable_page(struct page
*page
, isolate_mode_t mode
)
86 struct address_space
*mapping
;
89 * Avoid burning cycles with pages that are yet under __free_pages(),
90 * or just got freed under us.
92 * In case we 'win' a race for a movable page being freed under us and
93 * raise its refcount preventing __free_pages() from doing its job
94 * the put_page() at the end of this block will take care of
95 * release this page, thus avoiding a nasty leakage.
97 if (unlikely(!get_page_unless_zero(page
)))
101 * Check PageMovable before holding a PG_lock because page's owner
102 * assumes anybody doesn't touch PG_lock of newly allocated page
103 * so unconditionally grapping the lock ruins page's owner side.
105 if (unlikely(!__PageMovable(page
)))
108 * As movable pages are not isolated from LRU lists, concurrent
109 * compaction threads can race against page migration functions
110 * as well as race against the releasing a page.
112 * In order to avoid having an already isolated movable page
113 * being (wrongly) re-isolated while it is under migration,
114 * or to avoid attempting to isolate pages being released,
115 * lets be sure we have the page lock
116 * before proceeding with the movable page isolation steps.
118 if (unlikely(!trylock_page(page
)))
121 if (!PageMovable(page
) || PageIsolated(page
))
122 goto out_no_isolated
;
124 mapping
= page_mapping(page
);
125 VM_BUG_ON_PAGE(!mapping
, page
);
127 if (!mapping
->a_ops
->isolate_page(page
, mode
))
128 goto out_no_isolated
;
130 /* Driver shouldn't use PG_isolated bit of page->flags */
131 WARN_ON_ONCE(PageIsolated(page
));
132 __SetPageIsolated(page
);
145 /* It should be called on page which is PG_movable */
146 void putback_movable_page(struct page
*page
)
148 struct address_space
*mapping
;
150 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
151 VM_BUG_ON_PAGE(!PageMovable(page
), page
);
152 VM_BUG_ON_PAGE(!PageIsolated(page
), page
);
154 mapping
= page_mapping(page
);
155 mapping
->a_ops
->putback_page(page
);
156 __ClearPageIsolated(page
);
160 * Put previously isolated pages back onto the appropriate lists
161 * from where they were once taken off for compaction/migration.
163 * This function shall be used whenever the isolated pageset has been
164 * built from lru, balloon, hugetlbfs page. See isolate_migratepages_range()
165 * and isolate_huge_page().
167 void putback_movable_pages(struct list_head
*l
)
172 list_for_each_entry_safe(page
, page2
, l
, lru
) {
173 if (unlikely(PageHuge(page
))) {
174 putback_active_hugepage(page
);
177 list_del(&page
->lru
);
179 * We isolated non-lru movable page so here we can use
180 * __PageMovable because LRU page's mapping cannot have
181 * PAGE_MAPPING_MOVABLE.
183 if (unlikely(__PageMovable(page
))) {
184 VM_BUG_ON_PAGE(!PageIsolated(page
), page
);
186 if (PageMovable(page
))
187 putback_movable_page(page
);
189 __ClearPageIsolated(page
);
193 mod_node_page_state(page_pgdat(page
), NR_ISOLATED_ANON
+
194 page_is_file_cache(page
), -hpage_nr_pages(page
));
195 putback_lru_page(page
);
201 * Restore a potential migration pte to a working pte entry
203 static bool remove_migration_pte(struct page
*page
, struct vm_area_struct
*vma
,
204 unsigned long addr
, void *old
)
206 struct page_vma_mapped_walk pvmw
= {
210 .flags
= PVMW_SYNC
| PVMW_MIGRATION
,
216 VM_BUG_ON_PAGE(PageTail(page
), page
);
217 while (page_vma_mapped_walk(&pvmw
)) {
221 new = page
- pvmw
.page
->index
+
222 linear_page_index(vma
, pvmw
.address
);
224 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
225 /* PMD-mapped THP migration entry */
227 VM_BUG_ON_PAGE(PageHuge(page
) || !PageTransCompound(page
), page
);
228 remove_migration_pmd(&pvmw
, new);
234 pte
= pte_mkold(mk_pte(new, READ_ONCE(vma
->vm_page_prot
)));
235 if (pte_swp_soft_dirty(*pvmw
.pte
))
236 pte
= pte_mksoft_dirty(pte
);
239 * Recheck VMA as permissions can change since migration started
241 entry
= pte_to_swp_entry(*pvmw
.pte
);
242 if (is_write_migration_entry(entry
))
243 pte
= maybe_mkwrite(pte
, vma
);
245 if (unlikely(is_zone_device_page(new))) {
246 if (is_device_private_page(new)) {
247 entry
= make_device_private_entry(new, pte_write(pte
));
248 pte
= swp_entry_to_pte(entry
);
249 } else if (is_device_public_page(new)) {
250 pte
= pte_mkdevmap(pte
);
251 flush_dcache_page(new);
254 flush_dcache_page(new);
256 #ifdef CONFIG_HUGETLB_PAGE
258 pte
= pte_mkhuge(pte
);
259 pte
= arch_make_huge_pte(pte
, vma
, new, 0);
260 set_huge_pte_at(vma
->vm_mm
, pvmw
.address
, pvmw
.pte
, pte
);
262 hugepage_add_anon_rmap(new, vma
, pvmw
.address
);
264 page_dup_rmap(new, true);
268 set_pte_at(vma
->vm_mm
, pvmw
.address
, pvmw
.pte
, pte
);
271 page_add_anon_rmap(new, vma
, pvmw
.address
, false);
273 page_add_file_rmap(new, false);
275 if (vma
->vm_flags
& VM_LOCKED
&& !PageTransCompound(new))
278 if (PageTransHuge(page
) && PageMlocked(page
))
279 clear_page_mlock(page
);
281 /* No need to invalidate - it was non-present before */
282 update_mmu_cache(vma
, pvmw
.address
, pvmw
.pte
);
289 * Get rid of all migration entries and replace them by
290 * references to the indicated page.
292 void remove_migration_ptes(struct page
*old
, struct page
*new, bool locked
)
294 struct rmap_walk_control rwc
= {
295 .rmap_one
= remove_migration_pte
,
300 rmap_walk_locked(new, &rwc
);
302 rmap_walk(new, &rwc
);
306 * Something used the pte of a page under migration. We need to
307 * get to the page and wait until migration is finished.
308 * When we return from this function the fault will be retried.
310 void __migration_entry_wait(struct mm_struct
*mm
, pte_t
*ptep
,
319 if (!is_swap_pte(pte
))
322 entry
= pte_to_swp_entry(pte
);
323 if (!is_migration_entry(entry
))
326 page
= migration_entry_to_page(entry
);
329 * Once radix-tree replacement of page migration started, page_count
330 * *must* be zero. And, we don't want to call wait_on_page_locked()
331 * against a page without get_page().
332 * So, we use get_page_unless_zero(), here. Even failed, page fault
335 if (!get_page_unless_zero(page
))
337 pte_unmap_unlock(ptep
, ptl
);
338 wait_on_page_locked(page
);
342 pte_unmap_unlock(ptep
, ptl
);
345 void migration_entry_wait(struct mm_struct
*mm
, pmd_t
*pmd
,
346 unsigned long address
)
348 spinlock_t
*ptl
= pte_lockptr(mm
, pmd
);
349 pte_t
*ptep
= pte_offset_map(pmd
, address
);
350 __migration_entry_wait(mm
, ptep
, ptl
);
353 void migration_entry_wait_huge(struct vm_area_struct
*vma
,
354 struct mm_struct
*mm
, pte_t
*pte
)
356 spinlock_t
*ptl
= huge_pte_lockptr(hstate_vma(vma
), mm
, pte
);
357 __migration_entry_wait(mm
, pte
, ptl
);
360 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
361 void pmd_migration_entry_wait(struct mm_struct
*mm
, pmd_t
*pmd
)
366 ptl
= pmd_lock(mm
, pmd
);
367 if (!is_pmd_migration_entry(*pmd
))
369 page
= migration_entry_to_page(pmd_to_swp_entry(*pmd
));
370 if (!get_page_unless_zero(page
))
373 wait_on_page_locked(page
);
382 /* Returns true if all buffers are successfully locked */
383 static bool buffer_migrate_lock_buffers(struct buffer_head
*head
,
384 enum migrate_mode mode
)
386 struct buffer_head
*bh
= head
;
388 /* Simple case, sync compaction */
389 if (mode
!= MIGRATE_ASYNC
) {
393 bh
= bh
->b_this_page
;
395 } while (bh
!= head
);
400 /* async case, we cannot block on lock_buffer so use trylock_buffer */
403 if (!trylock_buffer(bh
)) {
405 * We failed to lock the buffer and cannot stall in
406 * async migration. Release the taken locks
408 struct buffer_head
*failed_bh
= bh
;
411 while (bh
!= failed_bh
) {
414 bh
= bh
->b_this_page
;
419 bh
= bh
->b_this_page
;
420 } while (bh
!= head
);
424 static inline bool buffer_migrate_lock_buffers(struct buffer_head
*head
,
425 enum migrate_mode mode
)
429 #endif /* CONFIG_BLOCK */
432 * Replace the page in the mapping.
434 * The number of remaining references must be:
435 * 1 for anonymous pages without a mapping
436 * 2 for pages with a mapping
437 * 3 for pages with a mapping and PagePrivate/PagePrivate2 set.
439 int migrate_page_move_mapping(struct address_space
*mapping
,
440 struct page
*newpage
, struct page
*page
,
441 struct buffer_head
*head
, enum migrate_mode mode
,
444 struct zone
*oldzone
, *newzone
;
446 int expected_count
= 1 + extra_count
;
450 * Device public or private pages have an extra refcount as they are
453 expected_count
+= is_device_private_page(page
);
454 expected_count
+= is_device_public_page(page
);
457 /* Anonymous page without mapping */
458 if (page_count(page
) != expected_count
)
461 /* No turning back from here */
462 newpage
->index
= page
->index
;
463 newpage
->mapping
= page
->mapping
;
464 if (PageSwapBacked(page
))
465 __SetPageSwapBacked(newpage
);
467 return MIGRATEPAGE_SUCCESS
;
470 oldzone
= page_zone(page
);
471 newzone
= page_zone(newpage
);
473 xa_lock_irq(&mapping
->i_pages
);
475 pslot
= radix_tree_lookup_slot(&mapping
->i_pages
,
478 expected_count
+= hpage_nr_pages(page
) + page_has_private(page
);
479 if (page_count(page
) != expected_count
||
480 radix_tree_deref_slot_protected(pslot
,
481 &mapping
->i_pages
.xa_lock
) != page
) {
482 xa_unlock_irq(&mapping
->i_pages
);
486 if (!page_ref_freeze(page
, expected_count
)) {
487 xa_unlock_irq(&mapping
->i_pages
);
492 * In the async migration case of moving a page with buffers, lock the
493 * buffers using trylock before the mapping is moved. If the mapping
494 * was moved, we later failed to lock the buffers and could not move
495 * the mapping back due to an elevated page count, we would have to
496 * block waiting on other references to be dropped.
498 if (mode
== MIGRATE_ASYNC
&& head
&&
499 !buffer_migrate_lock_buffers(head
, mode
)) {
500 page_ref_unfreeze(page
, expected_count
);
501 xa_unlock_irq(&mapping
->i_pages
);
506 * Now we know that no one else is looking at the page:
507 * no turning back from here.
509 newpage
->index
= page
->index
;
510 newpage
->mapping
= page
->mapping
;
511 page_ref_add(newpage
, hpage_nr_pages(page
)); /* add cache reference */
512 if (PageSwapBacked(page
)) {
513 __SetPageSwapBacked(newpage
);
514 if (PageSwapCache(page
)) {
515 SetPageSwapCache(newpage
);
516 set_page_private(newpage
, page_private(page
));
519 VM_BUG_ON_PAGE(PageSwapCache(page
), page
);
522 /* Move dirty while page refs frozen and newpage not yet exposed */
523 dirty
= PageDirty(page
);
525 ClearPageDirty(page
);
526 SetPageDirty(newpage
);
529 radix_tree_replace_slot(&mapping
->i_pages
, pslot
, newpage
);
530 if (PageTransHuge(page
)) {
532 int index
= page_index(page
);
534 for (i
= 1; i
< HPAGE_PMD_NR
; i
++) {
535 pslot
= radix_tree_lookup_slot(&mapping
->i_pages
,
537 radix_tree_replace_slot(&mapping
->i_pages
, pslot
,
543 * Drop cache reference from old page by unfreezing
544 * to one less reference.
545 * We know this isn't the last reference.
547 page_ref_unfreeze(page
, expected_count
- hpage_nr_pages(page
));
549 xa_unlock(&mapping
->i_pages
);
550 /* Leave irq disabled to prevent preemption while updating stats */
553 * If moved to a different zone then also account
554 * the page for that zone. Other VM counters will be
555 * taken care of when we establish references to the
556 * new page and drop references to the old page.
558 * Note that anonymous pages are accounted for
559 * via NR_FILE_PAGES and NR_ANON_MAPPED if they
560 * are mapped to swap space.
562 if (newzone
!= oldzone
) {
563 __dec_node_state(oldzone
->zone_pgdat
, NR_FILE_PAGES
);
564 __inc_node_state(newzone
->zone_pgdat
, NR_FILE_PAGES
);
565 if (PageSwapBacked(page
) && !PageSwapCache(page
)) {
566 __dec_node_state(oldzone
->zone_pgdat
, NR_SHMEM
);
567 __inc_node_state(newzone
->zone_pgdat
, NR_SHMEM
);
569 if (dirty
&& mapping_cap_account_dirty(mapping
)) {
570 __dec_node_state(oldzone
->zone_pgdat
, NR_FILE_DIRTY
);
571 __dec_zone_state(oldzone
, NR_ZONE_WRITE_PENDING
);
572 __inc_node_state(newzone
->zone_pgdat
, NR_FILE_DIRTY
);
573 __inc_zone_state(newzone
, NR_ZONE_WRITE_PENDING
);
578 return MIGRATEPAGE_SUCCESS
;
580 EXPORT_SYMBOL(migrate_page_move_mapping
);
583 * The expected number of remaining references is the same as that
584 * of migrate_page_move_mapping().
586 int migrate_huge_page_move_mapping(struct address_space
*mapping
,
587 struct page
*newpage
, struct page
*page
)
592 xa_lock_irq(&mapping
->i_pages
);
594 pslot
= radix_tree_lookup_slot(&mapping
->i_pages
, page_index(page
));
596 expected_count
= 2 + page_has_private(page
);
597 if (page_count(page
) != expected_count
||
598 radix_tree_deref_slot_protected(pslot
, &mapping
->i_pages
.xa_lock
) != page
) {
599 xa_unlock_irq(&mapping
->i_pages
);
603 if (!page_ref_freeze(page
, expected_count
)) {
604 xa_unlock_irq(&mapping
->i_pages
);
608 newpage
->index
= page
->index
;
609 newpage
->mapping
= page
->mapping
;
613 radix_tree_replace_slot(&mapping
->i_pages
, pslot
, newpage
);
615 page_ref_unfreeze(page
, expected_count
- 1);
617 xa_unlock_irq(&mapping
->i_pages
);
619 return MIGRATEPAGE_SUCCESS
;
623 * Gigantic pages are so large that we do not guarantee that page++ pointer
624 * arithmetic will work across the entire page. We need something more
627 static void __copy_gigantic_page(struct page
*dst
, struct page
*src
,
631 struct page
*dst_base
= dst
;
632 struct page
*src_base
= src
;
634 for (i
= 0; i
< nr_pages
; ) {
636 copy_highpage(dst
, src
);
639 dst
= mem_map_next(dst
, dst_base
, i
);
640 src
= mem_map_next(src
, src_base
, i
);
644 static void copy_huge_page(struct page
*dst
, struct page
*src
)
651 struct hstate
*h
= page_hstate(src
);
652 nr_pages
= pages_per_huge_page(h
);
654 if (unlikely(nr_pages
> MAX_ORDER_NR_PAGES
)) {
655 __copy_gigantic_page(dst
, src
, nr_pages
);
660 BUG_ON(!PageTransHuge(src
));
661 nr_pages
= hpage_nr_pages(src
);
664 for (i
= 0; i
< nr_pages
; i
++) {
666 copy_highpage(dst
+ i
, src
+ i
);
671 * Copy the page to its new location
673 void migrate_page_states(struct page
*newpage
, struct page
*page
)
678 SetPageError(newpage
);
679 if (PageReferenced(page
))
680 SetPageReferenced(newpage
);
681 if (PageUptodate(page
))
682 SetPageUptodate(newpage
);
683 if (TestClearPageActive(page
)) {
684 VM_BUG_ON_PAGE(PageUnevictable(page
), page
);
685 SetPageActive(newpage
);
686 } else if (TestClearPageUnevictable(page
))
687 SetPageUnevictable(newpage
);
688 if (PageChecked(page
))
689 SetPageChecked(newpage
);
690 if (PageMappedToDisk(page
))
691 SetPageMappedToDisk(newpage
);
693 /* Move dirty on pages not done by migrate_page_move_mapping() */
695 SetPageDirty(newpage
);
697 if (page_is_young(page
))
698 set_page_young(newpage
);
699 if (page_is_idle(page
))
700 set_page_idle(newpage
);
703 * Copy NUMA information to the new page, to prevent over-eager
704 * future migrations of this same page.
706 cpupid
= page_cpupid_xchg_last(page
, -1);
707 page_cpupid_xchg_last(newpage
, cpupid
);
709 ksm_migrate_page(newpage
, page
);
711 * Please do not reorder this without considering how mm/ksm.c's
712 * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache().
714 if (PageSwapCache(page
))
715 ClearPageSwapCache(page
);
716 ClearPagePrivate(page
);
717 set_page_private(page
, 0);
720 * If any waiters have accumulated on the new page then
723 if (PageWriteback(newpage
))
724 end_page_writeback(newpage
);
726 copy_page_owner(page
, newpage
);
728 mem_cgroup_migrate(page
, newpage
);
730 EXPORT_SYMBOL(migrate_page_states
);
732 void migrate_page_copy(struct page
*newpage
, struct page
*page
)
734 if (PageHuge(page
) || PageTransHuge(page
))
735 copy_huge_page(newpage
, page
);
737 copy_highpage(newpage
, page
);
739 migrate_page_states(newpage
, page
);
741 EXPORT_SYMBOL(migrate_page_copy
);
743 /************************************************************
744 * Migration functions
745 ***********************************************************/
748 * Common logic to directly migrate a single LRU page suitable for
749 * pages that do not use PagePrivate/PagePrivate2.
751 * Pages are locked upon entry and exit.
753 int migrate_page(struct address_space
*mapping
,
754 struct page
*newpage
, struct page
*page
,
755 enum migrate_mode mode
)
759 BUG_ON(PageWriteback(page
)); /* Writeback must be complete */
761 rc
= migrate_page_move_mapping(mapping
, newpage
, page
, NULL
, mode
, 0);
763 if (rc
!= MIGRATEPAGE_SUCCESS
)
766 if (mode
!= MIGRATE_SYNC_NO_COPY
)
767 migrate_page_copy(newpage
, page
);
769 migrate_page_states(newpage
, page
);
770 return MIGRATEPAGE_SUCCESS
;
772 EXPORT_SYMBOL(migrate_page
);
776 * Migration function for pages with buffers. This function can only be used
777 * if the underlying filesystem guarantees that no other references to "page"
780 int buffer_migrate_page(struct address_space
*mapping
,
781 struct page
*newpage
, struct page
*page
, enum migrate_mode mode
)
783 struct buffer_head
*bh
, *head
;
786 if (!page_has_buffers(page
))
787 return migrate_page(mapping
, newpage
, page
, mode
);
789 head
= page_buffers(page
);
791 rc
= migrate_page_move_mapping(mapping
, newpage
, page
, head
, mode
, 0);
793 if (rc
!= MIGRATEPAGE_SUCCESS
)
797 * In the async case, migrate_page_move_mapping locked the buffers
798 * with an IRQ-safe spinlock held. In the sync case, the buffers
799 * need to be locked now
801 if (mode
!= MIGRATE_ASYNC
)
802 BUG_ON(!buffer_migrate_lock_buffers(head
, mode
));
804 ClearPagePrivate(page
);
805 set_page_private(newpage
, page_private(page
));
806 set_page_private(page
, 0);
812 set_bh_page(bh
, newpage
, bh_offset(bh
));
813 bh
= bh
->b_this_page
;
815 } while (bh
!= head
);
817 SetPagePrivate(newpage
);
819 if (mode
!= MIGRATE_SYNC_NO_COPY
)
820 migrate_page_copy(newpage
, page
);
822 migrate_page_states(newpage
, page
);
828 bh
= bh
->b_this_page
;
830 } while (bh
!= head
);
832 return MIGRATEPAGE_SUCCESS
;
834 EXPORT_SYMBOL(buffer_migrate_page
);
838 * Writeback a page to clean the dirty state
840 static int writeout(struct address_space
*mapping
, struct page
*page
)
842 struct writeback_control wbc
= {
843 .sync_mode
= WB_SYNC_NONE
,
846 .range_end
= LLONG_MAX
,
851 if (!mapping
->a_ops
->writepage
)
852 /* No write method for the address space */
855 if (!clear_page_dirty_for_io(page
))
856 /* Someone else already triggered a write */
860 * A dirty page may imply that the underlying filesystem has
861 * the page on some queue. So the page must be clean for
862 * migration. Writeout may mean we loose the lock and the
863 * page state is no longer what we checked for earlier.
864 * At this point we know that the migration attempt cannot
867 remove_migration_ptes(page
, page
, false);
869 rc
= mapping
->a_ops
->writepage(page
, &wbc
);
871 if (rc
!= AOP_WRITEPAGE_ACTIVATE
)
872 /* unlocked. Relock */
875 return (rc
< 0) ? -EIO
: -EAGAIN
;
879 * Default handling if a filesystem does not provide a migration function.
881 static int fallback_migrate_page(struct address_space
*mapping
,
882 struct page
*newpage
, struct page
*page
, enum migrate_mode mode
)
884 if (PageDirty(page
)) {
885 /* Only writeback pages in full synchronous migration */
888 case MIGRATE_SYNC_NO_COPY
:
893 return writeout(mapping
, page
);
897 * Buffers may be managed in a filesystem specific way.
898 * We must have no buffers or drop them.
900 if (page_has_private(page
) &&
901 !try_to_release_page(page
, GFP_KERNEL
))
904 return migrate_page(mapping
, newpage
, page
, mode
);
908 * Move a page to a newly allocated page
909 * The page is locked and all ptes have been successfully removed.
911 * The new page will have replaced the old page if this function
916 * MIGRATEPAGE_SUCCESS - success
918 static int move_to_new_page(struct page
*newpage
, struct page
*page
,
919 enum migrate_mode mode
)
921 struct address_space
*mapping
;
923 bool is_lru
= !__PageMovable(page
);
925 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
926 VM_BUG_ON_PAGE(!PageLocked(newpage
), newpage
);
928 mapping
= page_mapping(page
);
930 if (likely(is_lru
)) {
932 rc
= migrate_page(mapping
, newpage
, page
, mode
);
933 else if (mapping
->a_ops
->migratepage
)
935 * Most pages have a mapping and most filesystems
936 * provide a migratepage callback. Anonymous pages
937 * are part of swap space which also has its own
938 * migratepage callback. This is the most common path
939 * for page migration.
941 rc
= mapping
->a_ops
->migratepage(mapping
, newpage
,
944 rc
= fallback_migrate_page(mapping
, newpage
,
948 * In case of non-lru page, it could be released after
949 * isolation step. In that case, we shouldn't try migration.
951 VM_BUG_ON_PAGE(!PageIsolated(page
), page
);
952 if (!PageMovable(page
)) {
953 rc
= MIGRATEPAGE_SUCCESS
;
954 __ClearPageIsolated(page
);
958 rc
= mapping
->a_ops
->migratepage(mapping
, newpage
,
960 WARN_ON_ONCE(rc
== MIGRATEPAGE_SUCCESS
&&
961 !PageIsolated(page
));
965 * When successful, old pagecache page->mapping must be cleared before
966 * page is freed; but stats require that PageAnon be left as PageAnon.
968 if (rc
== MIGRATEPAGE_SUCCESS
) {
969 if (__PageMovable(page
)) {
970 VM_BUG_ON_PAGE(!PageIsolated(page
), page
);
973 * We clear PG_movable under page_lock so any compactor
974 * cannot try to migrate this page.
976 __ClearPageIsolated(page
);
980 * Anonymous and movable page->mapping will be cleard by
981 * free_pages_prepare so don't reset it here for keeping
982 * the type to work PageAnon, for example.
984 if (!PageMappingFlags(page
))
985 page
->mapping
= NULL
;
991 static int __unmap_and_move(struct page
*page
, struct page
*newpage
,
992 int force
, enum migrate_mode mode
)
995 int page_was_mapped
= 0;
996 struct anon_vma
*anon_vma
= NULL
;
997 bool is_lru
= !__PageMovable(page
);
999 if (!trylock_page(page
)) {
1000 if (!force
|| mode
== MIGRATE_ASYNC
)
1004 * It's not safe for direct compaction to call lock_page.
1005 * For example, during page readahead pages are added locked
1006 * to the LRU. Later, when the IO completes the pages are
1007 * marked uptodate and unlocked. However, the queueing
1008 * could be merging multiple pages for one bio (e.g.
1009 * mpage_readpages). If an allocation happens for the
1010 * second or third page, the process can end up locking
1011 * the same page twice and deadlocking. Rather than
1012 * trying to be clever about what pages can be locked,
1013 * avoid the use of lock_page for direct compaction
1016 if (current
->flags
& PF_MEMALLOC
)
1022 if (PageWriteback(page
)) {
1024 * Only in the case of a full synchronous migration is it
1025 * necessary to wait for PageWriteback. In the async case,
1026 * the retry loop is too short and in the sync-light case,
1027 * the overhead of stalling is too much
1031 case MIGRATE_SYNC_NO_COPY
:
1039 wait_on_page_writeback(page
);
1043 * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
1044 * we cannot notice that anon_vma is freed while we migrates a page.
1045 * This get_anon_vma() delays freeing anon_vma pointer until the end
1046 * of migration. File cache pages are no problem because of page_lock()
1047 * File Caches may use write_page() or lock_page() in migration, then,
1048 * just care Anon page here.
1050 * Only page_get_anon_vma() understands the subtleties of
1051 * getting a hold on an anon_vma from outside one of its mms.
1052 * But if we cannot get anon_vma, then we won't need it anyway,
1053 * because that implies that the anon page is no longer mapped
1054 * (and cannot be remapped so long as we hold the page lock).
1056 if (PageAnon(page
) && !PageKsm(page
))
1057 anon_vma
= page_get_anon_vma(page
);
1060 * Block others from accessing the new page when we get around to
1061 * establishing additional references. We are usually the only one
1062 * holding a reference to newpage at this point. We used to have a BUG
1063 * here if trylock_page(newpage) fails, but would like to allow for
1064 * cases where there might be a race with the previous use of newpage.
1065 * This is much like races on refcount of oldpage: just don't BUG().
1067 if (unlikely(!trylock_page(newpage
)))
1070 if (unlikely(!is_lru
)) {
1071 rc
= move_to_new_page(newpage
, page
, mode
);
1072 goto out_unlock_both
;
1076 * Corner case handling:
1077 * 1. When a new swap-cache page is read into, it is added to the LRU
1078 * and treated as swapcache but it has no rmap yet.
1079 * Calling try_to_unmap() against a page->mapping==NULL page will
1080 * trigger a BUG. So handle it here.
1081 * 2. An orphaned page (see truncate_complete_page) might have
1082 * fs-private metadata. The page can be picked up due to memory
1083 * offlining. Everywhere else except page reclaim, the page is
1084 * invisible to the vm, so the page can not be migrated. So try to
1085 * free the metadata, so the page can be freed.
1087 if (!page
->mapping
) {
1088 VM_BUG_ON_PAGE(PageAnon(page
), page
);
1089 if (page_has_private(page
)) {
1090 try_to_free_buffers(page
);
1091 goto out_unlock_both
;
1093 } else if (page_mapped(page
)) {
1094 /* Establish migration ptes */
1095 VM_BUG_ON_PAGE(PageAnon(page
) && !PageKsm(page
) && !anon_vma
,
1098 TTU_MIGRATION
|TTU_IGNORE_MLOCK
|TTU_IGNORE_ACCESS
);
1099 page_was_mapped
= 1;
1102 if (!page_mapped(page
))
1103 rc
= move_to_new_page(newpage
, page
, mode
);
1105 if (page_was_mapped
)
1106 remove_migration_ptes(page
,
1107 rc
== MIGRATEPAGE_SUCCESS
? newpage
: page
, false);
1110 unlock_page(newpage
);
1112 /* Drop an anon_vma reference if we took one */
1114 put_anon_vma(anon_vma
);
1118 * If migration is successful, decrease refcount of the newpage
1119 * which will not free the page because new page owner increased
1120 * refcounter. As well, if it is LRU page, add the page to LRU
1123 if (rc
== MIGRATEPAGE_SUCCESS
) {
1124 if (unlikely(__PageMovable(newpage
)))
1127 putback_lru_page(newpage
);
1134 * gcc 4.7 and 4.8 on arm get an ICEs when inlining unmap_and_move(). Work
1137 #if defined(CONFIG_ARM) && \
1138 defined(GCC_VERSION) && GCC_VERSION < 40900 && GCC_VERSION >= 40700
1139 #define ICE_noinline noinline
1141 #define ICE_noinline
1145 * Obtain the lock on page, remove all ptes and migrate the page
1146 * to the newly allocated page in newpage.
1148 static ICE_noinline
int unmap_and_move(new_page_t get_new_page
,
1149 free_page_t put_new_page
,
1150 unsigned long private, struct page
*page
,
1151 int force
, enum migrate_mode mode
,
1152 enum migrate_reason reason
)
1154 int rc
= MIGRATEPAGE_SUCCESS
;
1155 struct page
*newpage
;
1157 if (!thp_migration_supported() && PageTransHuge(page
))
1160 newpage
= get_new_page(page
, private);
1164 if (page_count(page
) == 1) {
1165 /* page was freed from under us. So we are done. */
1166 ClearPageActive(page
);
1167 ClearPageUnevictable(page
);
1168 if (unlikely(__PageMovable(page
))) {
1170 if (!PageMovable(page
))
1171 __ClearPageIsolated(page
);
1175 put_new_page(newpage
, private);
1181 rc
= __unmap_and_move(page
, newpage
, force
, mode
);
1182 if (rc
== MIGRATEPAGE_SUCCESS
)
1183 set_page_owner_migrate_reason(newpage
, reason
);
1186 if (rc
!= -EAGAIN
) {
1188 * A page that has been migrated has all references
1189 * removed and will be freed. A page that has not been
1190 * migrated will have kepts its references and be
1193 list_del(&page
->lru
);
1196 * Compaction can migrate also non-LRU pages which are
1197 * not accounted to NR_ISOLATED_*. They can be recognized
1200 if (likely(!__PageMovable(page
)))
1201 mod_node_page_state(page_pgdat(page
), NR_ISOLATED_ANON
+
1202 page_is_file_cache(page
), -hpage_nr_pages(page
));
1206 * If migration is successful, releases reference grabbed during
1207 * isolation. Otherwise, restore the page to right list unless
1210 if (rc
== MIGRATEPAGE_SUCCESS
) {
1212 if (reason
== MR_MEMORY_FAILURE
) {
1214 * Set PG_HWPoison on just freed page
1215 * intentionally. Although it's rather weird,
1216 * it's how HWPoison flag works at the moment.
1218 if (set_hwpoison_free_buddy_page(page
))
1219 num_poisoned_pages_inc();
1222 if (rc
!= -EAGAIN
) {
1223 if (likely(!__PageMovable(page
))) {
1224 putback_lru_page(page
);
1229 if (PageMovable(page
))
1230 putback_movable_page(page
);
1232 __ClearPageIsolated(page
);
1238 put_new_page(newpage
, private);
1247 * Counterpart of unmap_and_move_page() for hugepage migration.
1249 * This function doesn't wait the completion of hugepage I/O
1250 * because there is no race between I/O and migration for hugepage.
1251 * Note that currently hugepage I/O occurs only in direct I/O
1252 * where no lock is held and PG_writeback is irrelevant,
1253 * and writeback status of all subpages are counted in the reference
1254 * count of the head page (i.e. if all subpages of a 2MB hugepage are
1255 * under direct I/O, the reference of the head page is 512 and a bit more.)
1256 * This means that when we try to migrate hugepage whose subpages are
1257 * doing direct I/O, some references remain after try_to_unmap() and
1258 * hugepage migration fails without data corruption.
1260 * There is also no race when direct I/O is issued on the page under migration,
1261 * because then pte is replaced with migration swap entry and direct I/O code
1262 * will wait in the page fault for migration to complete.
1264 static int unmap_and_move_huge_page(new_page_t get_new_page
,
1265 free_page_t put_new_page
, unsigned long private,
1266 struct page
*hpage
, int force
,
1267 enum migrate_mode mode
, int reason
)
1270 int page_was_mapped
= 0;
1271 struct page
*new_hpage
;
1272 struct anon_vma
*anon_vma
= NULL
;
1275 * Movability of hugepages depends on architectures and hugepage size.
1276 * This check is necessary because some callers of hugepage migration
1277 * like soft offline and memory hotremove don't walk through page
1278 * tables or check whether the hugepage is pmd-based or not before
1279 * kicking migration.
1281 if (!hugepage_migration_supported(page_hstate(hpage
))) {
1282 putback_active_hugepage(hpage
);
1286 new_hpage
= get_new_page(hpage
, private);
1290 if (!trylock_page(hpage
)) {
1295 case MIGRATE_SYNC_NO_COPY
:
1303 if (PageAnon(hpage
))
1304 anon_vma
= page_get_anon_vma(hpage
);
1306 if (unlikely(!trylock_page(new_hpage
)))
1309 if (page_mapped(hpage
)) {
1311 TTU_MIGRATION
|TTU_IGNORE_MLOCK
|TTU_IGNORE_ACCESS
);
1312 page_was_mapped
= 1;
1315 if (!page_mapped(hpage
))
1316 rc
= move_to_new_page(new_hpage
, hpage
, mode
);
1318 if (page_was_mapped
)
1319 remove_migration_ptes(hpage
,
1320 rc
== MIGRATEPAGE_SUCCESS
? new_hpage
: hpage
, false);
1322 unlock_page(new_hpage
);
1326 put_anon_vma(anon_vma
);
1328 if (rc
== MIGRATEPAGE_SUCCESS
) {
1329 move_hugetlb_state(hpage
, new_hpage
, reason
);
1330 put_new_page
= NULL
;
1336 putback_active_hugepage(hpage
);
1339 * If migration was not successful and there's a freeing callback, use
1340 * it. Otherwise, put_page() will drop the reference grabbed during
1344 put_new_page(new_hpage
, private);
1346 putback_active_hugepage(new_hpage
);
1352 * migrate_pages - migrate the pages specified in a list, to the free pages
1353 * supplied as the target for the page migration
1355 * @from: The list of pages to be migrated.
1356 * @get_new_page: The function used to allocate free pages to be used
1357 * as the target of the page migration.
1358 * @put_new_page: The function used to free target pages if migration
1359 * fails, or NULL if no special handling is necessary.
1360 * @private: Private data to be passed on to get_new_page()
1361 * @mode: The migration mode that specifies the constraints for
1362 * page migration, if any.
1363 * @reason: The reason for page migration.
1365 * The function returns after 10 attempts or if no pages are movable any more
1366 * because the list has become empty or no retryable pages exist any more.
1367 * The caller should call putback_movable_pages() to return pages to the LRU
1368 * or free list only if ret != 0.
1370 * Returns the number of pages that were not migrated, or an error code.
1372 int migrate_pages(struct list_head
*from
, new_page_t get_new_page
,
1373 free_page_t put_new_page
, unsigned long private,
1374 enum migrate_mode mode
, int reason
)
1378 int nr_succeeded
= 0;
1382 int swapwrite
= current
->flags
& PF_SWAPWRITE
;
1386 current
->flags
|= PF_SWAPWRITE
;
1388 for(pass
= 0; pass
< 10 && retry
; pass
++) {
1391 list_for_each_entry_safe(page
, page2
, from
, lru
) {
1396 rc
= unmap_and_move_huge_page(get_new_page
,
1397 put_new_page
, private, page
,
1398 pass
> 2, mode
, reason
);
1400 rc
= unmap_and_move(get_new_page
, put_new_page
,
1401 private, page
, pass
> 2, mode
,
1407 * THP migration might be unsupported or the
1408 * allocation could've failed so we should
1409 * retry on the same page with the THP split
1412 * Head page is retried immediately and tail
1413 * pages are added to the tail of the list so
1414 * we encounter them after the rest of the list
1417 if (PageTransHuge(page
)) {
1419 rc
= split_huge_page_to_list(page
, from
);
1422 list_safe_reset_next(page
, page2
, lru
);
1431 case MIGRATEPAGE_SUCCESS
:
1436 * Permanent failure (-EBUSY, -ENOSYS, etc.):
1437 * unlike -EAGAIN case, the failed page is
1438 * removed from migration page list and not
1439 * retried in the next outer loop.
1450 count_vm_events(PGMIGRATE_SUCCESS
, nr_succeeded
);
1452 count_vm_events(PGMIGRATE_FAIL
, nr_failed
);
1453 trace_mm_migrate_pages(nr_succeeded
, nr_failed
, mode
, reason
);
1456 current
->flags
&= ~PF_SWAPWRITE
;
1463 static int store_status(int __user
*status
, int start
, int value
, int nr
)
1466 if (put_user(value
, status
+ start
))
1474 static int do_move_pages_to_node(struct mm_struct
*mm
,
1475 struct list_head
*pagelist
, int node
)
1479 if (list_empty(pagelist
))
1482 err
= migrate_pages(pagelist
, alloc_new_node_page
, NULL
, node
,
1483 MIGRATE_SYNC
, MR_SYSCALL
);
1485 putback_movable_pages(pagelist
);
1490 * Resolves the given address to a struct page, isolates it from the LRU and
1491 * puts it to the given pagelist.
1492 * Returns -errno if the page cannot be found/isolated or 0 when it has been
1493 * queued or the page doesn't need to be migrated because it is already on
1496 static int add_page_for_migration(struct mm_struct
*mm
, unsigned long addr
,
1497 int node
, struct list_head
*pagelist
, bool migrate_all
)
1499 struct vm_area_struct
*vma
;
1501 unsigned int follflags
;
1504 down_read(&mm
->mmap_sem
);
1506 vma
= find_vma(mm
, addr
);
1507 if (!vma
|| addr
< vma
->vm_start
|| !vma_migratable(vma
))
1510 /* FOLL_DUMP to ignore special (like zero) pages */
1511 follflags
= FOLL_GET
| FOLL_DUMP
;
1512 page
= follow_page(vma
, addr
, follflags
);
1514 err
= PTR_ERR(page
);
1523 if (page_to_nid(page
) == node
)
1527 if (page_mapcount(page
) > 1 && !migrate_all
)
1530 if (PageHuge(page
)) {
1531 if (PageHead(page
)) {
1532 isolate_huge_page(page
, pagelist
);
1538 head
= compound_head(page
);
1539 err
= isolate_lru_page(head
);
1544 list_add_tail(&head
->lru
, pagelist
);
1545 mod_node_page_state(page_pgdat(head
),
1546 NR_ISOLATED_ANON
+ page_is_file_cache(head
),
1547 hpage_nr_pages(head
));
1551 * Either remove the duplicate refcount from
1552 * isolate_lru_page() or drop the page ref if it was
1557 up_read(&mm
->mmap_sem
);
1562 * Migrate an array of page address onto an array of nodes and fill
1563 * the corresponding array of status.
1565 static int do_pages_move(struct mm_struct
*mm
, nodemask_t task_nodes
,
1566 unsigned long nr_pages
,
1567 const void __user
* __user
*pages
,
1568 const int __user
*nodes
,
1569 int __user
*status
, int flags
)
1571 int current_node
= NUMA_NO_NODE
;
1572 LIST_HEAD(pagelist
);
1578 for (i
= start
= 0; i
< nr_pages
; i
++) {
1579 const void __user
*p
;
1584 if (get_user(p
, pages
+ i
))
1586 if (get_user(node
, nodes
+ i
))
1588 addr
= (unsigned long)p
;
1591 if (node
< 0 || node
>= MAX_NUMNODES
)
1593 if (!node_state(node
, N_MEMORY
))
1597 if (!node_isset(node
, task_nodes
))
1600 if (current_node
== NUMA_NO_NODE
) {
1601 current_node
= node
;
1603 } else if (node
!= current_node
) {
1604 err
= do_move_pages_to_node(mm
, &pagelist
, current_node
);
1607 err
= store_status(status
, start
, current_node
, i
- start
);
1611 current_node
= node
;
1615 * Errors in the page lookup or isolation are not fatal and we simply
1616 * report them via status
1618 err
= add_page_for_migration(mm
, addr
, current_node
,
1619 &pagelist
, flags
& MPOL_MF_MOVE_ALL
);
1623 err
= store_status(status
, i
, err
, 1);
1627 err
= do_move_pages_to_node(mm
, &pagelist
, current_node
);
1631 err
= store_status(status
, start
, current_node
, i
- start
);
1635 current_node
= NUMA_NO_NODE
;
1638 if (list_empty(&pagelist
))
1641 /* Make sure we do not overwrite the existing error */
1642 err1
= do_move_pages_to_node(mm
, &pagelist
, current_node
);
1644 err1
= store_status(status
, start
, current_node
, i
- start
);
1652 * Determine the nodes of an array of pages and store it in an array of status.
1654 static void do_pages_stat_array(struct mm_struct
*mm
, unsigned long nr_pages
,
1655 const void __user
**pages
, int *status
)
1659 down_read(&mm
->mmap_sem
);
1661 for (i
= 0; i
< nr_pages
; i
++) {
1662 unsigned long addr
= (unsigned long)(*pages
);
1663 struct vm_area_struct
*vma
;
1667 vma
= find_vma(mm
, addr
);
1668 if (!vma
|| addr
< vma
->vm_start
)
1671 /* FOLL_DUMP to ignore special (like zero) pages */
1672 page
= follow_page(vma
, addr
, FOLL_DUMP
);
1674 err
= PTR_ERR(page
);
1678 err
= page
? page_to_nid(page
) : -ENOENT
;
1686 up_read(&mm
->mmap_sem
);
1690 * Determine the nodes of a user array of pages and store it in
1691 * a user array of status.
1693 static int do_pages_stat(struct mm_struct
*mm
, unsigned long nr_pages
,
1694 const void __user
* __user
*pages
,
1697 #define DO_PAGES_STAT_CHUNK_NR 16
1698 const void __user
*chunk_pages
[DO_PAGES_STAT_CHUNK_NR
];
1699 int chunk_status
[DO_PAGES_STAT_CHUNK_NR
];
1702 unsigned long chunk_nr
;
1704 chunk_nr
= nr_pages
;
1705 if (chunk_nr
> DO_PAGES_STAT_CHUNK_NR
)
1706 chunk_nr
= DO_PAGES_STAT_CHUNK_NR
;
1708 if (copy_from_user(chunk_pages
, pages
, chunk_nr
* sizeof(*chunk_pages
)))
1711 do_pages_stat_array(mm
, chunk_nr
, chunk_pages
, chunk_status
);
1713 if (copy_to_user(status
, chunk_status
, chunk_nr
* sizeof(*status
)))
1718 nr_pages
-= chunk_nr
;
1720 return nr_pages
? -EFAULT
: 0;
1724 * Move a list of pages in the address space of the currently executing
1727 static int kernel_move_pages(pid_t pid
, unsigned long nr_pages
,
1728 const void __user
* __user
*pages
,
1729 const int __user
*nodes
,
1730 int __user
*status
, int flags
)
1732 struct task_struct
*task
;
1733 struct mm_struct
*mm
;
1735 nodemask_t task_nodes
;
1738 if (flags
& ~(MPOL_MF_MOVE
|MPOL_MF_MOVE_ALL
))
1741 if ((flags
& MPOL_MF_MOVE_ALL
) && !capable(CAP_SYS_NICE
))
1744 /* Find the mm_struct */
1746 task
= pid
? find_task_by_vpid(pid
) : current
;
1751 get_task_struct(task
);
1754 * Check if this process has the right to modify the specified
1755 * process. Use the regular "ptrace_may_access()" checks.
1757 if (!ptrace_may_access(task
, PTRACE_MODE_READ_REALCREDS
)) {
1764 err
= security_task_movememory(task
);
1768 task_nodes
= cpuset_mems_allowed(task
);
1769 mm
= get_task_mm(task
);
1770 put_task_struct(task
);
1776 err
= do_pages_move(mm
, task_nodes
, nr_pages
, pages
,
1777 nodes
, status
, flags
);
1779 err
= do_pages_stat(mm
, nr_pages
, pages
, status
);
1785 put_task_struct(task
);
1789 SYSCALL_DEFINE6(move_pages
, pid_t
, pid
, unsigned long, nr_pages
,
1790 const void __user
* __user
*, pages
,
1791 const int __user
*, nodes
,
1792 int __user
*, status
, int, flags
)
1794 return kernel_move_pages(pid
, nr_pages
, pages
, nodes
, status
, flags
);
1797 #ifdef CONFIG_COMPAT
1798 COMPAT_SYSCALL_DEFINE6(move_pages
, pid_t
, pid
, compat_ulong_t
, nr_pages
,
1799 compat_uptr_t __user
*, pages32
,
1800 const int __user
*, nodes
,
1801 int __user
*, status
,
1804 const void __user
* __user
*pages
;
1807 pages
= compat_alloc_user_space(nr_pages
* sizeof(void *));
1808 for (i
= 0; i
< nr_pages
; i
++) {
1811 if (get_user(p
, pages32
+ i
) ||
1812 put_user(compat_ptr(p
), pages
+ i
))
1815 return kernel_move_pages(pid
, nr_pages
, pages
, nodes
, status
, flags
);
1817 #endif /* CONFIG_COMPAT */
1819 #ifdef CONFIG_NUMA_BALANCING
1821 * Returns true if this is a safe migration target node for misplaced NUMA
1822 * pages. Currently it only checks the watermarks which crude
1824 static bool migrate_balanced_pgdat(struct pglist_data
*pgdat
,
1825 unsigned long nr_migrate_pages
)
1829 for (z
= pgdat
->nr_zones
- 1; z
>= 0; z
--) {
1830 struct zone
*zone
= pgdat
->node_zones
+ z
;
1832 if (!populated_zone(zone
))
1835 /* Avoid waking kswapd by allocating pages_to_migrate pages. */
1836 if (!zone_watermark_ok(zone
, 0,
1837 high_wmark_pages(zone
) +
1846 static struct page
*alloc_misplaced_dst_page(struct page
*page
,
1849 int nid
= (int) data
;
1850 struct page
*newpage
;
1852 newpage
= __alloc_pages_node(nid
,
1853 (GFP_HIGHUSER_MOVABLE
|
1854 __GFP_THISNODE
| __GFP_NOMEMALLOC
|
1855 __GFP_NORETRY
| __GFP_NOWARN
) &
1862 * page migration rate limiting control.
1863 * Do not migrate more than @pages_to_migrate in a @migrate_interval_millisecs
1864 * window of time. Default here says do not migrate more than 1280M per second.
1866 static unsigned int migrate_interval_millisecs __read_mostly
= 100;
1867 static unsigned int ratelimit_pages __read_mostly
= 128 << (20 - PAGE_SHIFT
);
1869 /* Returns true if the node is migrate rate-limited after the update */
1870 static bool numamigrate_update_ratelimit(pg_data_t
*pgdat
,
1871 unsigned long nr_pages
)
1874 * Rate-limit the amount of data that is being migrated to a node.
1875 * Optimal placement is no good if the memory bus is saturated and
1876 * all the time is being spent migrating!
1878 if (time_after(jiffies
, pgdat
->numabalancing_migrate_next_window
)) {
1879 spin_lock(&pgdat
->numabalancing_migrate_lock
);
1880 pgdat
->numabalancing_migrate_nr_pages
= 0;
1881 pgdat
->numabalancing_migrate_next_window
= jiffies
+
1882 msecs_to_jiffies(migrate_interval_millisecs
);
1883 spin_unlock(&pgdat
->numabalancing_migrate_lock
);
1885 if (pgdat
->numabalancing_migrate_nr_pages
> ratelimit_pages
) {
1886 trace_mm_numa_migrate_ratelimit(current
, pgdat
->node_id
,
1892 * This is an unlocked non-atomic update so errors are possible.
1893 * The consequences are failing to migrate when we potentiall should
1894 * have which is not severe enough to warrant locking. If it is ever
1895 * a problem, it can be converted to a per-cpu counter.
1897 pgdat
->numabalancing_migrate_nr_pages
+= nr_pages
;
1901 static int numamigrate_isolate_page(pg_data_t
*pgdat
, struct page
*page
)
1905 VM_BUG_ON_PAGE(compound_order(page
) && !PageTransHuge(page
), page
);
1907 /* Avoid migrating to a node that is nearly full */
1908 if (!migrate_balanced_pgdat(pgdat
, 1UL << compound_order(page
)))
1911 if (isolate_lru_page(page
))
1915 * migrate_misplaced_transhuge_page() skips page migration's usual
1916 * check on page_count(), so we must do it here, now that the page
1917 * has been isolated: a GUP pin, or any other pin, prevents migration.
1918 * The expected page count is 3: 1 for page's mapcount and 1 for the
1919 * caller's pin and 1 for the reference taken by isolate_lru_page().
1921 if (PageTransHuge(page
) && page_count(page
) != 3) {
1922 putback_lru_page(page
);
1926 page_lru
= page_is_file_cache(page
);
1927 mod_node_page_state(page_pgdat(page
), NR_ISOLATED_ANON
+ page_lru
,
1928 hpage_nr_pages(page
));
1931 * Isolating the page has taken another reference, so the
1932 * caller's reference can be safely dropped without the page
1933 * disappearing underneath us during migration.
1939 bool pmd_trans_migrating(pmd_t pmd
)
1941 struct page
*page
= pmd_page(pmd
);
1942 return PageLocked(page
);
1946 * Attempt to migrate a misplaced page to the specified destination
1947 * node. Caller is expected to have an elevated reference count on
1948 * the page that will be dropped by this function before returning.
1950 int migrate_misplaced_page(struct page
*page
, struct vm_area_struct
*vma
,
1953 pg_data_t
*pgdat
= NODE_DATA(node
);
1956 LIST_HEAD(migratepages
);
1959 * Don't migrate file pages that are mapped in multiple processes
1960 * with execute permissions as they are probably shared libraries.
1962 if (page_mapcount(page
) != 1 && page_is_file_cache(page
) &&
1963 (vma
->vm_flags
& VM_EXEC
))
1967 * Also do not migrate dirty pages as not all filesystems can move
1968 * dirty pages in MIGRATE_ASYNC mode which is a waste of cycles.
1970 if (page_is_file_cache(page
) && PageDirty(page
))
1974 * Rate-limit the amount of data that is being migrated to a node.
1975 * Optimal placement is no good if the memory bus is saturated and
1976 * all the time is being spent migrating!
1978 if (numamigrate_update_ratelimit(pgdat
, 1))
1981 isolated
= numamigrate_isolate_page(pgdat
, page
);
1985 list_add(&page
->lru
, &migratepages
);
1986 nr_remaining
= migrate_pages(&migratepages
, alloc_misplaced_dst_page
,
1987 NULL
, node
, MIGRATE_ASYNC
,
1990 if (!list_empty(&migratepages
)) {
1991 list_del(&page
->lru
);
1992 dec_node_page_state(page
, NR_ISOLATED_ANON
+
1993 page_is_file_cache(page
));
1994 putback_lru_page(page
);
1998 count_vm_numa_event(NUMA_PAGE_MIGRATE
);
1999 BUG_ON(!list_empty(&migratepages
));
2006 #endif /* CONFIG_NUMA_BALANCING */
2008 #if defined(CONFIG_NUMA_BALANCING) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
2010 * Migrates a THP to a given target node. page must be locked and is unlocked
2013 int migrate_misplaced_transhuge_page(struct mm_struct
*mm
,
2014 struct vm_area_struct
*vma
,
2015 pmd_t
*pmd
, pmd_t entry
,
2016 unsigned long address
,
2017 struct page
*page
, int node
)
2020 pg_data_t
*pgdat
= NODE_DATA(node
);
2022 struct page
*new_page
= NULL
;
2023 int page_lru
= page_is_file_cache(page
);
2024 unsigned long mmun_start
= address
& HPAGE_PMD_MASK
;
2025 unsigned long mmun_end
= mmun_start
+ HPAGE_PMD_SIZE
;
2028 * Rate-limit the amount of data that is being migrated to a node.
2029 * Optimal placement is no good if the memory bus is saturated and
2030 * all the time is being spent migrating!
2032 if (numamigrate_update_ratelimit(pgdat
, HPAGE_PMD_NR
))
2035 new_page
= alloc_pages_node(node
,
2036 (GFP_TRANSHUGE_LIGHT
| __GFP_THISNODE
),
2040 prep_transhuge_page(new_page
);
2042 isolated
= numamigrate_isolate_page(pgdat
, page
);
2048 /* Prepare a page as a migration target */
2049 __SetPageLocked(new_page
);
2050 if (PageSwapBacked(page
))
2051 __SetPageSwapBacked(new_page
);
2053 /* anon mapping, we can simply copy page->mapping to the new page: */
2054 new_page
->mapping
= page
->mapping
;
2055 new_page
->index
= page
->index
;
2056 migrate_page_copy(new_page
, page
);
2057 WARN_ON(PageLRU(new_page
));
2059 /* Recheck the target PMD */
2060 mmu_notifier_invalidate_range_start(mm
, mmun_start
, mmun_end
);
2061 ptl
= pmd_lock(mm
, pmd
);
2062 if (unlikely(!pmd_same(*pmd
, entry
) || !page_ref_freeze(page
, 2))) {
2064 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
2066 /* Reverse changes made by migrate_page_copy() */
2067 if (TestClearPageActive(new_page
))
2068 SetPageActive(page
);
2069 if (TestClearPageUnevictable(new_page
))
2070 SetPageUnevictable(page
);
2072 unlock_page(new_page
);
2073 put_page(new_page
); /* Free it */
2075 /* Retake the callers reference and putback on LRU */
2077 putback_lru_page(page
);
2078 mod_node_page_state(page_pgdat(page
),
2079 NR_ISOLATED_ANON
+ page_lru
, -HPAGE_PMD_NR
);
2084 entry
= mk_huge_pmd(new_page
, vma
->vm_page_prot
);
2085 entry
= maybe_pmd_mkwrite(pmd_mkdirty(entry
), vma
);
2088 * Clear the old entry under pagetable lock and establish the new PTE.
2089 * Any parallel GUP will either observe the old page blocking on the
2090 * page lock, block on the page table lock or observe the new page.
2091 * The SetPageUptodate on the new page and page_add_new_anon_rmap
2092 * guarantee the copy is visible before the pagetable update.
2094 flush_cache_range(vma
, mmun_start
, mmun_end
);
2095 page_add_anon_rmap(new_page
, vma
, mmun_start
, true);
2096 pmdp_huge_clear_flush_notify(vma
, mmun_start
, pmd
);
2097 set_pmd_at(mm
, mmun_start
, pmd
, entry
);
2098 update_mmu_cache_pmd(vma
, address
, &entry
);
2100 page_ref_unfreeze(page
, 2);
2101 mlock_migrate_page(new_page
, page
);
2102 page_remove_rmap(page
, true);
2103 set_page_owner_migrate_reason(new_page
, MR_NUMA_MISPLACED
);
2107 * No need to double call mmu_notifier->invalidate_range() callback as
2108 * the above pmdp_huge_clear_flush_notify() did already call it.
2110 mmu_notifier_invalidate_range_only_end(mm
, mmun_start
, mmun_end
);
2112 /* Take an "isolate" reference and put new page on the LRU. */
2114 putback_lru_page(new_page
);
2116 unlock_page(new_page
);
2118 put_page(page
); /* Drop the rmap reference */
2119 put_page(page
); /* Drop the LRU isolation reference */
2121 count_vm_events(PGMIGRATE_SUCCESS
, HPAGE_PMD_NR
);
2122 count_vm_numa_events(NUMA_PAGE_MIGRATE
, HPAGE_PMD_NR
);
2124 mod_node_page_state(page_pgdat(page
),
2125 NR_ISOLATED_ANON
+ page_lru
,
2130 count_vm_events(PGMIGRATE_FAIL
, HPAGE_PMD_NR
);
2132 ptl
= pmd_lock(mm
, pmd
);
2133 if (pmd_same(*pmd
, entry
)) {
2134 entry
= pmd_modify(entry
, vma
->vm_page_prot
);
2135 set_pmd_at(mm
, mmun_start
, pmd
, entry
);
2136 update_mmu_cache_pmd(vma
, address
, &entry
);
2145 #endif /* CONFIG_NUMA_BALANCING */
2147 #endif /* CONFIG_NUMA */
2149 #if defined(CONFIG_MIGRATE_VMA_HELPER)
2150 struct migrate_vma
{
2151 struct vm_area_struct
*vma
;
2154 unsigned long cpages
;
2155 unsigned long npages
;
2156 unsigned long start
;
2160 static int migrate_vma_collect_hole(unsigned long start
,
2162 struct mm_walk
*walk
)
2164 struct migrate_vma
*migrate
= walk
->private;
2167 for (addr
= start
& PAGE_MASK
; addr
< end
; addr
+= PAGE_SIZE
) {
2168 migrate
->src
[migrate
->npages
] = MIGRATE_PFN_MIGRATE
;
2169 migrate
->dst
[migrate
->npages
] = 0;
2177 static int migrate_vma_collect_skip(unsigned long start
,
2179 struct mm_walk
*walk
)
2181 struct migrate_vma
*migrate
= walk
->private;
2184 for (addr
= start
& PAGE_MASK
; addr
< end
; addr
+= PAGE_SIZE
) {
2185 migrate
->dst
[migrate
->npages
] = 0;
2186 migrate
->src
[migrate
->npages
++] = 0;
2192 static int migrate_vma_collect_pmd(pmd_t
*pmdp
,
2193 unsigned long start
,
2195 struct mm_walk
*walk
)
2197 struct migrate_vma
*migrate
= walk
->private;
2198 struct vm_area_struct
*vma
= walk
->vma
;
2199 struct mm_struct
*mm
= vma
->vm_mm
;
2200 unsigned long addr
= start
, unmapped
= 0;
2205 if (pmd_none(*pmdp
))
2206 return migrate_vma_collect_hole(start
, end
, walk
);
2208 if (pmd_trans_huge(*pmdp
)) {
2211 ptl
= pmd_lock(mm
, pmdp
);
2212 if (unlikely(!pmd_trans_huge(*pmdp
))) {
2217 page
= pmd_page(*pmdp
);
2218 if (is_huge_zero_page(page
)) {
2220 split_huge_pmd(vma
, pmdp
, addr
);
2221 if (pmd_trans_unstable(pmdp
))
2222 return migrate_vma_collect_skip(start
, end
,
2229 if (unlikely(!trylock_page(page
)))
2230 return migrate_vma_collect_skip(start
, end
,
2232 ret
= split_huge_page(page
);
2236 return migrate_vma_collect_skip(start
, end
,
2238 if (pmd_none(*pmdp
))
2239 return migrate_vma_collect_hole(start
, end
,
2244 if (unlikely(pmd_bad(*pmdp
)))
2245 return migrate_vma_collect_skip(start
, end
, walk
);
2247 ptep
= pte_offset_map_lock(mm
, pmdp
, addr
, &ptl
);
2248 arch_enter_lazy_mmu_mode();
2250 for (; addr
< end
; addr
+= PAGE_SIZE
, ptep
++) {
2251 unsigned long mpfn
, pfn
;
2259 if (pte_none(pte
)) {
2260 mpfn
= MIGRATE_PFN_MIGRATE
;
2266 if (!pte_present(pte
)) {
2270 * Only care about unaddressable device page special
2271 * page table entry. Other special swap entries are not
2272 * migratable, and we ignore regular swapped page.
2274 entry
= pte_to_swp_entry(pte
);
2275 if (!is_device_private_entry(entry
))
2278 page
= device_private_entry_to_page(entry
);
2279 mpfn
= migrate_pfn(page_to_pfn(page
))|
2280 MIGRATE_PFN_DEVICE
| MIGRATE_PFN_MIGRATE
;
2281 if (is_write_device_private_entry(entry
))
2282 mpfn
|= MIGRATE_PFN_WRITE
;
2284 if (is_zero_pfn(pfn
)) {
2285 mpfn
= MIGRATE_PFN_MIGRATE
;
2290 page
= _vm_normal_page(migrate
->vma
, addr
, pte
, true);
2291 mpfn
= migrate_pfn(pfn
) | MIGRATE_PFN_MIGRATE
;
2292 mpfn
|= pte_write(pte
) ? MIGRATE_PFN_WRITE
: 0;
2295 /* FIXME support THP */
2296 if (!page
|| !page
->mapping
|| PageTransCompound(page
)) {
2300 pfn
= page_to_pfn(page
);
2303 * By getting a reference on the page we pin it and that blocks
2304 * any kind of migration. Side effect is that it "freezes" the
2307 * We drop this reference after isolating the page from the lru
2308 * for non device page (device page are not on the lru and thus
2309 * can't be dropped from it).
2315 * Optimize for the common case where page is only mapped once
2316 * in one process. If we can lock the page, then we can safely
2317 * set up a special migration page table entry now.
2319 if (trylock_page(page
)) {
2322 mpfn
|= MIGRATE_PFN_LOCKED
;
2323 ptep_get_and_clear(mm
, addr
, ptep
);
2325 /* Setup special migration page table entry */
2326 entry
= make_migration_entry(page
, mpfn
&
2328 swp_pte
= swp_entry_to_pte(entry
);
2329 if (pte_soft_dirty(pte
))
2330 swp_pte
= pte_swp_mksoft_dirty(swp_pte
);
2331 set_pte_at(mm
, addr
, ptep
, swp_pte
);
2334 * This is like regular unmap: we remove the rmap and
2335 * drop page refcount. Page won't be freed, as we took
2336 * a reference just above.
2338 page_remove_rmap(page
, false);
2341 if (pte_present(pte
))
2346 migrate
->dst
[migrate
->npages
] = 0;
2347 migrate
->src
[migrate
->npages
++] = mpfn
;
2349 arch_leave_lazy_mmu_mode();
2350 pte_unmap_unlock(ptep
- 1, ptl
);
2352 /* Only flush the TLB if we actually modified any entries */
2354 flush_tlb_range(walk
->vma
, start
, end
);
2360 * migrate_vma_collect() - collect pages over a range of virtual addresses
2361 * @migrate: migrate struct containing all migration information
2363 * This will walk the CPU page table. For each virtual address backed by a
2364 * valid page, it updates the src array and takes a reference on the page, in
2365 * order to pin the page until we lock it and unmap it.
2367 static void migrate_vma_collect(struct migrate_vma
*migrate
)
2369 struct mm_walk mm_walk
;
2371 mm_walk
.pmd_entry
= migrate_vma_collect_pmd
;
2372 mm_walk
.pte_entry
= NULL
;
2373 mm_walk
.pte_hole
= migrate_vma_collect_hole
;
2374 mm_walk
.hugetlb_entry
= NULL
;
2375 mm_walk
.test_walk
= NULL
;
2376 mm_walk
.vma
= migrate
->vma
;
2377 mm_walk
.mm
= migrate
->vma
->vm_mm
;
2378 mm_walk
.private = migrate
;
2380 mmu_notifier_invalidate_range_start(mm_walk
.mm
,
2383 walk_page_range(migrate
->start
, migrate
->end
, &mm_walk
);
2384 mmu_notifier_invalidate_range_end(mm_walk
.mm
,
2388 migrate
->end
= migrate
->start
+ (migrate
->npages
<< PAGE_SHIFT
);
2392 * migrate_vma_check_page() - check if page is pinned or not
2393 * @page: struct page to check
2395 * Pinned pages cannot be migrated. This is the same test as in
2396 * migrate_page_move_mapping(), except that here we allow migration of a
2399 static bool migrate_vma_check_page(struct page
*page
)
2402 * One extra ref because caller holds an extra reference, either from
2403 * isolate_lru_page() for a regular page, or migrate_vma_collect() for
2409 * FIXME support THP (transparent huge page), it is bit more complex to
2410 * check them than regular pages, because they can be mapped with a pmd
2411 * or with a pte (split pte mapping).
2413 if (PageCompound(page
))
2416 /* Page from ZONE_DEVICE have one extra reference */
2417 if (is_zone_device_page(page
)) {
2419 * Private page can never be pin as they have no valid pte and
2420 * GUP will fail for those. Yet if there is a pending migration
2421 * a thread might try to wait on the pte migration entry and
2422 * will bump the page reference count. Sadly there is no way to
2423 * differentiate a regular pin from migration wait. Hence to
2424 * avoid 2 racing thread trying to migrate back to CPU to enter
2425 * infinite loop (one stoping migration because the other is
2426 * waiting on pte migration entry). We always return true here.
2428 * FIXME proper solution is to rework migration_entry_wait() so
2429 * it does not need to take a reference on page.
2431 if (is_device_private_page(page
))
2435 * Only allow device public page to be migrated and account for
2436 * the extra reference count imply by ZONE_DEVICE pages.
2438 if (!is_device_public_page(page
))
2443 /* For file back page */
2444 if (page_mapping(page
))
2445 extra
+= 1 + page_has_private(page
);
2447 if ((page_count(page
) - extra
) > page_mapcount(page
))
2454 * migrate_vma_prepare() - lock pages and isolate them from the lru
2455 * @migrate: migrate struct containing all migration information
2457 * This locks pages that have been collected by migrate_vma_collect(). Once each
2458 * page is locked it is isolated from the lru (for non-device pages). Finally,
2459 * the ref taken by migrate_vma_collect() is dropped, as locked pages cannot be
2460 * migrated by concurrent kernel threads.
2462 static void migrate_vma_prepare(struct migrate_vma
*migrate
)
2464 const unsigned long npages
= migrate
->npages
;
2465 const unsigned long start
= migrate
->start
;
2466 unsigned long addr
, i
, restore
= 0;
2467 bool allow_drain
= true;
2471 for (i
= 0; (i
< npages
) && migrate
->cpages
; i
++) {
2472 struct page
*page
= migrate_pfn_to_page(migrate
->src
[i
]);
2478 if (!(migrate
->src
[i
] & MIGRATE_PFN_LOCKED
)) {
2480 * Because we are migrating several pages there can be
2481 * a deadlock between 2 concurrent migration where each
2482 * are waiting on each other page lock.
2484 * Make migrate_vma() a best effort thing and backoff
2485 * for any page we can not lock right away.
2487 if (!trylock_page(page
)) {
2488 migrate
->src
[i
] = 0;
2494 migrate
->src
[i
] |= MIGRATE_PFN_LOCKED
;
2497 /* ZONE_DEVICE pages are not on LRU */
2498 if (!is_zone_device_page(page
)) {
2499 if (!PageLRU(page
) && allow_drain
) {
2500 /* Drain CPU's pagevec */
2501 lru_add_drain_all();
2502 allow_drain
= false;
2505 if (isolate_lru_page(page
)) {
2507 migrate
->src
[i
] &= ~MIGRATE_PFN_MIGRATE
;
2511 migrate
->src
[i
] = 0;
2519 /* Drop the reference we took in collect */
2523 if (!migrate_vma_check_page(page
)) {
2525 migrate
->src
[i
] &= ~MIGRATE_PFN_MIGRATE
;
2529 if (!is_zone_device_page(page
)) {
2531 putback_lru_page(page
);
2534 migrate
->src
[i
] = 0;
2538 if (!is_zone_device_page(page
))
2539 putback_lru_page(page
);
2546 for (i
= 0, addr
= start
; i
< npages
&& restore
; i
++, addr
+= PAGE_SIZE
) {
2547 struct page
*page
= migrate_pfn_to_page(migrate
->src
[i
]);
2549 if (!page
|| (migrate
->src
[i
] & MIGRATE_PFN_MIGRATE
))
2552 remove_migration_pte(page
, migrate
->vma
, addr
, page
);
2554 migrate
->src
[i
] = 0;
2562 * migrate_vma_unmap() - replace page mapping with special migration pte entry
2563 * @migrate: migrate struct containing all migration information
2565 * Replace page mapping (CPU page table pte) with a special migration pte entry
2566 * and check again if it has been pinned. Pinned pages are restored because we
2567 * cannot migrate them.
2569 * This is the last step before we call the device driver callback to allocate
2570 * destination memory and copy contents of original page over to new page.
2572 static void migrate_vma_unmap(struct migrate_vma
*migrate
)
2574 int flags
= TTU_MIGRATION
| TTU_IGNORE_MLOCK
| TTU_IGNORE_ACCESS
;
2575 const unsigned long npages
= migrate
->npages
;
2576 const unsigned long start
= migrate
->start
;
2577 unsigned long addr
, i
, restore
= 0;
2579 for (i
= 0; i
< npages
; i
++) {
2580 struct page
*page
= migrate_pfn_to_page(migrate
->src
[i
]);
2582 if (!page
|| !(migrate
->src
[i
] & MIGRATE_PFN_MIGRATE
))
2585 if (page_mapped(page
)) {
2586 try_to_unmap(page
, flags
);
2587 if (page_mapped(page
))
2591 if (migrate_vma_check_page(page
))
2595 migrate
->src
[i
] &= ~MIGRATE_PFN_MIGRATE
;
2600 for (addr
= start
, i
= 0; i
< npages
&& restore
; addr
+= PAGE_SIZE
, i
++) {
2601 struct page
*page
= migrate_pfn_to_page(migrate
->src
[i
]);
2603 if (!page
|| (migrate
->src
[i
] & MIGRATE_PFN_MIGRATE
))
2606 remove_migration_ptes(page
, page
, false);
2608 migrate
->src
[i
] = 0;
2612 if (is_zone_device_page(page
))
2615 putback_lru_page(page
);
2619 static void migrate_vma_insert_page(struct migrate_vma
*migrate
,
2625 struct vm_area_struct
*vma
= migrate
->vma
;
2626 struct mm_struct
*mm
= vma
->vm_mm
;
2627 struct mem_cgroup
*memcg
;
2637 /* Only allow populating anonymous memory */
2638 if (!vma_is_anonymous(vma
))
2641 pgdp
= pgd_offset(mm
, addr
);
2642 p4dp
= p4d_alloc(mm
, pgdp
, addr
);
2645 pudp
= pud_alloc(mm
, p4dp
, addr
);
2648 pmdp
= pmd_alloc(mm
, pudp
, addr
);
2652 if (pmd_trans_huge(*pmdp
) || pmd_devmap(*pmdp
))
2656 * Use pte_alloc() instead of pte_alloc_map(). We can't run
2657 * pte_offset_map() on pmds where a huge pmd might be created
2658 * from a different thread.
2660 * pte_alloc_map() is safe to use under down_write(mmap_sem) or when
2661 * parallel threads are excluded by other means.
2663 * Here we only have down_read(mmap_sem).
2665 if (pte_alloc(mm
, pmdp
, addr
))
2668 /* See the comment in pte_alloc_one_map() */
2669 if (unlikely(pmd_trans_unstable(pmdp
)))
2672 if (unlikely(anon_vma_prepare(vma
)))
2674 if (mem_cgroup_try_charge(page
, vma
->vm_mm
, GFP_KERNEL
, &memcg
, false))
2678 * The memory barrier inside __SetPageUptodate makes sure that
2679 * preceding stores to the page contents become visible before
2680 * the set_pte_at() write.
2682 __SetPageUptodate(page
);
2684 if (is_zone_device_page(page
)) {
2685 if (is_device_private_page(page
)) {
2686 swp_entry_t swp_entry
;
2688 swp_entry
= make_device_private_entry(page
, vma
->vm_flags
& VM_WRITE
);
2689 entry
= swp_entry_to_pte(swp_entry
);
2690 } else if (is_device_public_page(page
)) {
2691 entry
= pte_mkold(mk_pte(page
, READ_ONCE(vma
->vm_page_prot
)));
2692 if (vma
->vm_flags
& VM_WRITE
)
2693 entry
= pte_mkwrite(pte_mkdirty(entry
));
2694 entry
= pte_mkdevmap(entry
);
2697 entry
= mk_pte(page
, vma
->vm_page_prot
);
2698 if (vma
->vm_flags
& VM_WRITE
)
2699 entry
= pte_mkwrite(pte_mkdirty(entry
));
2702 ptep
= pte_offset_map_lock(mm
, pmdp
, addr
, &ptl
);
2704 if (pte_present(*ptep
)) {
2705 unsigned long pfn
= pte_pfn(*ptep
);
2707 if (!is_zero_pfn(pfn
)) {
2708 pte_unmap_unlock(ptep
, ptl
);
2709 mem_cgroup_cancel_charge(page
, memcg
, false);
2713 } else if (!pte_none(*ptep
)) {
2714 pte_unmap_unlock(ptep
, ptl
);
2715 mem_cgroup_cancel_charge(page
, memcg
, false);
2720 * Check for usefaultfd but do not deliver the fault. Instead,
2723 if (userfaultfd_missing(vma
)) {
2724 pte_unmap_unlock(ptep
, ptl
);
2725 mem_cgroup_cancel_charge(page
, memcg
, false);
2729 inc_mm_counter(mm
, MM_ANONPAGES
);
2730 page_add_new_anon_rmap(page
, vma
, addr
, false);
2731 mem_cgroup_commit_charge(page
, memcg
, false, false);
2732 if (!is_zone_device_page(page
))
2733 lru_cache_add_active_or_unevictable(page
, vma
);
2737 flush_cache_page(vma
, addr
, pte_pfn(*ptep
));
2738 ptep_clear_flush_notify(vma
, addr
, ptep
);
2739 set_pte_at_notify(mm
, addr
, ptep
, entry
);
2740 update_mmu_cache(vma
, addr
, ptep
);
2742 /* No need to invalidate - it was non-present before */
2743 set_pte_at(mm
, addr
, ptep
, entry
);
2744 update_mmu_cache(vma
, addr
, ptep
);
2747 pte_unmap_unlock(ptep
, ptl
);
2748 *src
= MIGRATE_PFN_MIGRATE
;
2752 *src
&= ~MIGRATE_PFN_MIGRATE
;
2756 * migrate_vma_pages() - migrate meta-data from src page to dst page
2757 * @migrate: migrate struct containing all migration information
2759 * This migrates struct page meta-data from source struct page to destination
2760 * struct page. This effectively finishes the migration from source page to the
2763 static void migrate_vma_pages(struct migrate_vma
*migrate
)
2765 const unsigned long npages
= migrate
->npages
;
2766 const unsigned long start
= migrate
->start
;
2767 struct vm_area_struct
*vma
= migrate
->vma
;
2768 struct mm_struct
*mm
= vma
->vm_mm
;
2769 unsigned long addr
, i
, mmu_start
;
2770 bool notified
= false;
2772 for (i
= 0, addr
= start
; i
< npages
; addr
+= PAGE_SIZE
, i
++) {
2773 struct page
*newpage
= migrate_pfn_to_page(migrate
->dst
[i
]);
2774 struct page
*page
= migrate_pfn_to_page(migrate
->src
[i
]);
2775 struct address_space
*mapping
;
2779 migrate
->src
[i
] &= ~MIGRATE_PFN_MIGRATE
;
2784 if (!(migrate
->src
[i
] & MIGRATE_PFN_MIGRATE
)) {
2790 mmu_notifier_invalidate_range_start(mm
,
2794 migrate_vma_insert_page(migrate
, addr
, newpage
,
2800 mapping
= page_mapping(page
);
2802 if (is_zone_device_page(newpage
)) {
2803 if (is_device_private_page(newpage
)) {
2805 * For now only support private anonymous when
2806 * migrating to un-addressable device memory.
2809 migrate
->src
[i
] &= ~MIGRATE_PFN_MIGRATE
;
2812 } else if (!is_device_public_page(newpage
)) {
2814 * Other types of ZONE_DEVICE page are not
2817 migrate
->src
[i
] &= ~MIGRATE_PFN_MIGRATE
;
2822 r
= migrate_page(mapping
, newpage
, page
, MIGRATE_SYNC_NO_COPY
);
2823 if (r
!= MIGRATEPAGE_SUCCESS
)
2824 migrate
->src
[i
] &= ~MIGRATE_PFN_MIGRATE
;
2828 * No need to double call mmu_notifier->invalidate_range() callback as
2829 * the above ptep_clear_flush_notify() inside migrate_vma_insert_page()
2830 * did already call it.
2833 mmu_notifier_invalidate_range_only_end(mm
, mmu_start
,
2838 * migrate_vma_finalize() - restore CPU page table entry
2839 * @migrate: migrate struct containing all migration information
2841 * This replaces the special migration pte entry with either a mapping to the
2842 * new page if migration was successful for that page, or to the original page
2845 * This also unlocks the pages and puts them back on the lru, or drops the extra
2846 * refcount, for device pages.
2848 static void migrate_vma_finalize(struct migrate_vma
*migrate
)
2850 const unsigned long npages
= migrate
->npages
;
2853 for (i
= 0; i
< npages
; i
++) {
2854 struct page
*newpage
= migrate_pfn_to_page(migrate
->dst
[i
]);
2855 struct page
*page
= migrate_pfn_to_page(migrate
->src
[i
]);
2859 unlock_page(newpage
);
2865 if (!(migrate
->src
[i
] & MIGRATE_PFN_MIGRATE
) || !newpage
) {
2867 unlock_page(newpage
);
2873 remove_migration_ptes(page
, newpage
, false);
2877 if (is_zone_device_page(page
))
2880 putback_lru_page(page
);
2882 if (newpage
!= page
) {
2883 unlock_page(newpage
);
2884 if (is_zone_device_page(newpage
))
2887 putback_lru_page(newpage
);
2893 * migrate_vma() - migrate a range of memory inside vma
2895 * @ops: migration callback for allocating destination memory and copying
2896 * @vma: virtual memory area containing the range to be migrated
2897 * @start: start address of the range to migrate (inclusive)
2898 * @end: end address of the range to migrate (exclusive)
2899 * @src: array of hmm_pfn_t containing source pfns
2900 * @dst: array of hmm_pfn_t containing destination pfns
2901 * @private: pointer passed back to each of the callback
2902 * Returns: 0 on success, error code otherwise
2904 * This function tries to migrate a range of memory virtual address range, using
2905 * callbacks to allocate and copy memory from source to destination. First it
2906 * collects all the pages backing each virtual address in the range, saving this
2907 * inside the src array. Then it locks those pages and unmaps them. Once the pages
2908 * are locked and unmapped, it checks whether each page is pinned or not. Pages
2909 * that aren't pinned have the MIGRATE_PFN_MIGRATE flag set (by this function)
2910 * in the corresponding src array entry. It then restores any pages that are
2911 * pinned, by remapping and unlocking those pages.
2913 * At this point it calls the alloc_and_copy() callback. For documentation on
2914 * what is expected from that callback, see struct migrate_vma_ops comments in
2915 * include/linux/migrate.h
2917 * After the alloc_and_copy() callback, this function goes over each entry in
2918 * the src array that has the MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE flag
2919 * set. If the corresponding entry in dst array has MIGRATE_PFN_VALID flag set,
2920 * then the function tries to migrate struct page information from the source
2921 * struct page to the destination struct page. If it fails to migrate the struct
2922 * page information, then it clears the MIGRATE_PFN_MIGRATE flag in the src
2925 * At this point all successfully migrated pages have an entry in the src
2926 * array with MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE flag set and the dst
2927 * array entry with MIGRATE_PFN_VALID flag set.
2929 * It then calls the finalize_and_map() callback. See comments for "struct
2930 * migrate_vma_ops", in include/linux/migrate.h for details about
2931 * finalize_and_map() behavior.
2933 * After the finalize_and_map() callback, for successfully migrated pages, this
2934 * function updates the CPU page table to point to new pages, otherwise it
2935 * restores the CPU page table to point to the original source pages.
2937 * Function returns 0 after the above steps, even if no pages were migrated
2938 * (The function only returns an error if any of the arguments are invalid.)
2940 * Both src and dst array must be big enough for (end - start) >> PAGE_SHIFT
2941 * unsigned long entries.
2943 int migrate_vma(const struct migrate_vma_ops
*ops
,
2944 struct vm_area_struct
*vma
,
2945 unsigned long start
,
2951 struct migrate_vma migrate
;
2953 /* Sanity check the arguments */
2956 if (!vma
|| is_vm_hugetlb_page(vma
) || (vma
->vm_flags
& VM_SPECIAL
) ||
2959 if (start
< vma
->vm_start
|| start
>= vma
->vm_end
)
2961 if (end
<= vma
->vm_start
|| end
> vma
->vm_end
)
2963 if (!ops
|| !src
|| !dst
|| start
>= end
)
2966 memset(src
, 0, sizeof(*src
) * ((end
- start
) >> PAGE_SHIFT
));
2969 migrate
.start
= start
;
2975 /* Collect, and try to unmap source pages */
2976 migrate_vma_collect(&migrate
);
2977 if (!migrate
.cpages
)
2980 /* Lock and isolate page */
2981 migrate_vma_prepare(&migrate
);
2982 if (!migrate
.cpages
)
2986 migrate_vma_unmap(&migrate
);
2987 if (!migrate
.cpages
)
2991 * At this point pages are locked and unmapped, and thus they have
2992 * stable content and can safely be copied to destination memory that
2993 * is allocated by the callback.
2995 * Note that migration can fail in migrate_vma_struct_page() for each
2998 ops
->alloc_and_copy(vma
, src
, dst
, start
, end
, private);
3000 /* This does the real migration of struct page */
3001 migrate_vma_pages(&migrate
);
3003 ops
->finalize_and_map(vma
, src
, dst
, start
, end
, private);
3005 /* Unlock and remap pages */
3006 migrate_vma_finalize(&migrate
);
3010 EXPORT_SYMBOL(migrate_vma
);
3011 #endif /* defined(MIGRATE_VMA_HELPER) */