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/pagewalk.h>
42 #include <linux/pfn_t.h>
43 #include <linux/memremap.h>
44 #include <linux/userfaultfd_k.h>
45 #include <linux/balloon_compaction.h>
46 #include <linux/mmu_notifier.h>
47 #include <linux/page_idle.h>
48 #include <linux/page_owner.h>
49 #include <linux/sched/mm.h>
50 #include <linux/ptrace.h>
52 #include <asm/tlbflush.h>
54 #define CREATE_TRACE_POINTS
55 #include <trace/events/migrate.h>
60 * migrate_prep() needs to be called before we start compiling a list of pages
61 * to be migrated using isolate_lru_page(). If scheduling work on other CPUs is
62 * undesirable, use migrate_prep_local()
64 int migrate_prep(void)
67 * Clear the LRU lists so pages can be isolated.
68 * Note that pages may be moved off the LRU after we have
69 * drained them. Those pages will fail to migrate like other
70 * pages that may be busy.
77 /* Do the necessary work of migrate_prep but not if it involves other CPUs */
78 int migrate_prep_local(void)
85 int isolate_movable_page(struct page
*page
, isolate_mode_t mode
)
87 struct address_space
*mapping
;
90 * Avoid burning cycles with pages that are yet under __free_pages(),
91 * or just got freed under us.
93 * In case we 'win' a race for a movable page being freed under us and
94 * raise its refcount preventing __free_pages() from doing its job
95 * the put_page() at the end of this block will take care of
96 * release this page, thus avoiding a nasty leakage.
98 if (unlikely(!get_page_unless_zero(page
)))
102 * Check PageMovable before holding a PG_lock because page's owner
103 * assumes anybody doesn't touch PG_lock of newly allocated page
104 * so unconditionally grabbing the lock ruins page's owner side.
106 if (unlikely(!__PageMovable(page
)))
109 * As movable pages are not isolated from LRU lists, concurrent
110 * compaction threads can race against page migration functions
111 * as well as race against the releasing a page.
113 * In order to avoid having an already isolated movable page
114 * being (wrongly) re-isolated while it is under migration,
115 * or to avoid attempting to isolate pages being released,
116 * lets be sure we have the page lock
117 * before proceeding with the movable page isolation steps.
119 if (unlikely(!trylock_page(page
)))
122 if (!PageMovable(page
) || PageIsolated(page
))
123 goto out_no_isolated
;
125 mapping
= page_mapping(page
);
126 VM_BUG_ON_PAGE(!mapping
, page
);
128 if (!mapping
->a_ops
->isolate_page(page
, mode
))
129 goto out_no_isolated
;
131 /* Driver shouldn't use PG_isolated bit of page->flags */
132 WARN_ON_ONCE(PageIsolated(page
));
133 __SetPageIsolated(page
);
146 /* It should be called on page which is PG_movable */
147 void putback_movable_page(struct page
*page
)
149 struct address_space
*mapping
;
151 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
152 VM_BUG_ON_PAGE(!PageMovable(page
), page
);
153 VM_BUG_ON_PAGE(!PageIsolated(page
), page
);
155 mapping
= page_mapping(page
);
156 mapping
->a_ops
->putback_page(page
);
157 __ClearPageIsolated(page
);
161 * Put previously isolated pages back onto the appropriate lists
162 * from where they were once taken off for compaction/migration.
164 * This function shall be used whenever the isolated pageset has been
165 * built from lru, balloon, hugetlbfs page. See isolate_migratepages_range()
166 * and isolate_huge_page().
168 void putback_movable_pages(struct list_head
*l
)
173 list_for_each_entry_safe(page
, page2
, l
, lru
) {
174 if (unlikely(PageHuge(page
))) {
175 putback_active_hugepage(page
);
178 list_del(&page
->lru
);
180 * We isolated non-lru movable page so here we can use
181 * __PageMovable because LRU page's mapping cannot have
182 * PAGE_MAPPING_MOVABLE.
184 if (unlikely(__PageMovable(page
))) {
185 VM_BUG_ON_PAGE(!PageIsolated(page
), page
);
187 if (PageMovable(page
))
188 putback_movable_page(page
);
190 __ClearPageIsolated(page
);
194 mod_node_page_state(page_pgdat(page
), NR_ISOLATED_ANON
+
195 page_is_file_cache(page
), -hpage_nr_pages(page
));
196 putback_lru_page(page
);
202 * Restore a potential migration pte to a working pte entry
204 static bool remove_migration_pte(struct page
*page
, struct vm_area_struct
*vma
,
205 unsigned long addr
, void *old
)
207 struct page_vma_mapped_walk pvmw
= {
211 .flags
= PVMW_SYNC
| PVMW_MIGRATION
,
217 VM_BUG_ON_PAGE(PageTail(page
), page
);
218 while (page_vma_mapped_walk(&pvmw
)) {
222 new = page
- pvmw
.page
->index
+
223 linear_page_index(vma
, pvmw
.address
);
225 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
226 /* PMD-mapped THP migration entry */
228 VM_BUG_ON_PAGE(PageHuge(page
) || !PageTransCompound(page
), page
);
229 remove_migration_pmd(&pvmw
, new);
235 pte
= pte_mkold(mk_pte(new, READ_ONCE(vma
->vm_page_prot
)));
236 if (pte_swp_soft_dirty(*pvmw
.pte
))
237 pte
= pte_mksoft_dirty(pte
);
240 * Recheck VMA as permissions can change since migration started
242 entry
= pte_to_swp_entry(*pvmw
.pte
);
243 if (is_write_migration_entry(entry
))
244 pte
= maybe_mkwrite(pte
, vma
);
246 if (unlikely(is_zone_device_page(new))) {
247 if (is_device_private_page(new)) {
248 entry
= make_device_private_entry(new, pte_write(pte
));
249 pte
= swp_entry_to_pte(entry
);
253 #ifdef CONFIG_HUGETLB_PAGE
255 pte
= pte_mkhuge(pte
);
256 pte
= arch_make_huge_pte(pte
, vma
, new, 0);
257 set_huge_pte_at(vma
->vm_mm
, pvmw
.address
, pvmw
.pte
, pte
);
259 hugepage_add_anon_rmap(new, vma
, pvmw
.address
);
261 page_dup_rmap(new, true);
265 set_pte_at(vma
->vm_mm
, pvmw
.address
, pvmw
.pte
, pte
);
268 page_add_anon_rmap(new, vma
, pvmw
.address
, false);
270 page_add_file_rmap(new, false);
272 if (vma
->vm_flags
& VM_LOCKED
&& !PageTransCompound(new))
275 if (PageTransHuge(page
) && PageMlocked(page
))
276 clear_page_mlock(page
);
278 /* No need to invalidate - it was non-present before */
279 update_mmu_cache(vma
, pvmw
.address
, pvmw
.pte
);
286 * Get rid of all migration entries and replace them by
287 * references to the indicated page.
289 void remove_migration_ptes(struct page
*old
, struct page
*new, bool locked
)
291 struct rmap_walk_control rwc
= {
292 .rmap_one
= remove_migration_pte
,
297 rmap_walk_locked(new, &rwc
);
299 rmap_walk(new, &rwc
);
303 * Something used the pte of a page under migration. We need to
304 * get to the page and wait until migration is finished.
305 * When we return from this function the fault will be retried.
307 void __migration_entry_wait(struct mm_struct
*mm
, pte_t
*ptep
,
316 if (!is_swap_pte(pte
))
319 entry
= pte_to_swp_entry(pte
);
320 if (!is_migration_entry(entry
))
323 page
= migration_entry_to_page(entry
);
326 * Once page cache replacement of page migration started, page_count
327 * is zero; but we must not call put_and_wait_on_page_locked() without
328 * a ref. Use get_page_unless_zero(), and just fault again if it fails.
330 if (!get_page_unless_zero(page
))
332 pte_unmap_unlock(ptep
, ptl
);
333 put_and_wait_on_page_locked(page
);
336 pte_unmap_unlock(ptep
, ptl
);
339 void migration_entry_wait(struct mm_struct
*mm
, pmd_t
*pmd
,
340 unsigned long address
)
342 spinlock_t
*ptl
= pte_lockptr(mm
, pmd
);
343 pte_t
*ptep
= pte_offset_map(pmd
, address
);
344 __migration_entry_wait(mm
, ptep
, ptl
);
347 void migration_entry_wait_huge(struct vm_area_struct
*vma
,
348 struct mm_struct
*mm
, pte_t
*pte
)
350 spinlock_t
*ptl
= huge_pte_lockptr(hstate_vma(vma
), mm
, pte
);
351 __migration_entry_wait(mm
, pte
, ptl
);
354 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
355 void pmd_migration_entry_wait(struct mm_struct
*mm
, pmd_t
*pmd
)
360 ptl
= pmd_lock(mm
, pmd
);
361 if (!is_pmd_migration_entry(*pmd
))
363 page
= migration_entry_to_page(pmd_to_swp_entry(*pmd
));
364 if (!get_page_unless_zero(page
))
367 put_and_wait_on_page_locked(page
);
374 static int expected_page_refs(struct address_space
*mapping
, struct page
*page
)
376 int expected_count
= 1;
379 * Device public or private pages have an extra refcount as they are
382 expected_count
+= is_device_private_page(page
);
384 expected_count
+= hpage_nr_pages(page
) + page_has_private(page
);
386 return expected_count
;
390 * Replace the page in the mapping.
392 * The number of remaining references must be:
393 * 1 for anonymous pages without a mapping
394 * 2 for pages with a mapping
395 * 3 for pages with a mapping and PagePrivate/PagePrivate2 set.
397 int migrate_page_move_mapping(struct address_space
*mapping
,
398 struct page
*newpage
, struct page
*page
, int extra_count
)
400 XA_STATE(xas
, &mapping
->i_pages
, page_index(page
));
401 struct zone
*oldzone
, *newzone
;
403 int expected_count
= expected_page_refs(mapping
, page
) + extra_count
;
406 /* Anonymous page without mapping */
407 if (page_count(page
) != expected_count
)
410 /* No turning back from here */
411 newpage
->index
= page
->index
;
412 newpage
->mapping
= page
->mapping
;
413 if (PageSwapBacked(page
))
414 __SetPageSwapBacked(newpage
);
416 return MIGRATEPAGE_SUCCESS
;
419 oldzone
= page_zone(page
);
420 newzone
= page_zone(newpage
);
423 if (page_count(page
) != expected_count
|| xas_load(&xas
) != page
) {
424 xas_unlock_irq(&xas
);
428 if (!page_ref_freeze(page
, expected_count
)) {
429 xas_unlock_irq(&xas
);
434 * Now we know that no one else is looking at the page:
435 * no turning back from here.
437 newpage
->index
= page
->index
;
438 newpage
->mapping
= page
->mapping
;
439 page_ref_add(newpage
, hpage_nr_pages(page
)); /* add cache reference */
440 if (PageSwapBacked(page
)) {
441 __SetPageSwapBacked(newpage
);
442 if (PageSwapCache(page
)) {
443 SetPageSwapCache(newpage
);
444 set_page_private(newpage
, page_private(page
));
447 VM_BUG_ON_PAGE(PageSwapCache(page
), page
);
450 /* Move dirty while page refs frozen and newpage not yet exposed */
451 dirty
= PageDirty(page
);
453 ClearPageDirty(page
);
454 SetPageDirty(newpage
);
457 xas_store(&xas
, newpage
);
458 if (PageTransHuge(page
)) {
461 for (i
= 1; i
< HPAGE_PMD_NR
; i
++) {
463 xas_store(&xas
, newpage
);
468 * Drop cache reference from old page by unfreezing
469 * to one less reference.
470 * We know this isn't the last reference.
472 page_ref_unfreeze(page
, expected_count
- hpage_nr_pages(page
));
475 /* Leave irq disabled to prevent preemption while updating stats */
478 * If moved to a different zone then also account
479 * the page for that zone. Other VM counters will be
480 * taken care of when we establish references to the
481 * new page and drop references to the old page.
483 * Note that anonymous pages are accounted for
484 * via NR_FILE_PAGES and NR_ANON_MAPPED if they
485 * are mapped to swap space.
487 if (newzone
!= oldzone
) {
488 __dec_node_state(oldzone
->zone_pgdat
, NR_FILE_PAGES
);
489 __inc_node_state(newzone
->zone_pgdat
, NR_FILE_PAGES
);
490 if (PageSwapBacked(page
) && !PageSwapCache(page
)) {
491 __dec_node_state(oldzone
->zone_pgdat
, NR_SHMEM
);
492 __inc_node_state(newzone
->zone_pgdat
, NR_SHMEM
);
494 if (dirty
&& mapping_cap_account_dirty(mapping
)) {
495 __dec_node_state(oldzone
->zone_pgdat
, NR_FILE_DIRTY
);
496 __dec_zone_state(oldzone
, NR_ZONE_WRITE_PENDING
);
497 __inc_node_state(newzone
->zone_pgdat
, NR_FILE_DIRTY
);
498 __inc_zone_state(newzone
, NR_ZONE_WRITE_PENDING
);
503 return MIGRATEPAGE_SUCCESS
;
505 EXPORT_SYMBOL(migrate_page_move_mapping
);
508 * The expected number of remaining references is the same as that
509 * of migrate_page_move_mapping().
511 int migrate_huge_page_move_mapping(struct address_space
*mapping
,
512 struct page
*newpage
, struct page
*page
)
514 XA_STATE(xas
, &mapping
->i_pages
, page_index(page
));
518 expected_count
= 2 + page_has_private(page
);
519 if (page_count(page
) != expected_count
|| xas_load(&xas
) != page
) {
520 xas_unlock_irq(&xas
);
524 if (!page_ref_freeze(page
, expected_count
)) {
525 xas_unlock_irq(&xas
);
529 newpage
->index
= page
->index
;
530 newpage
->mapping
= page
->mapping
;
534 xas_store(&xas
, newpage
);
536 page_ref_unfreeze(page
, expected_count
- 1);
538 xas_unlock_irq(&xas
);
540 return MIGRATEPAGE_SUCCESS
;
544 * Gigantic pages are so large that we do not guarantee that page++ pointer
545 * arithmetic will work across the entire page. We need something more
548 static void __copy_gigantic_page(struct page
*dst
, struct page
*src
,
552 struct page
*dst_base
= dst
;
553 struct page
*src_base
= src
;
555 for (i
= 0; i
< nr_pages
; ) {
557 copy_highpage(dst
, src
);
560 dst
= mem_map_next(dst
, dst_base
, i
);
561 src
= mem_map_next(src
, src_base
, i
);
565 static void copy_huge_page(struct page
*dst
, struct page
*src
)
572 struct hstate
*h
= page_hstate(src
);
573 nr_pages
= pages_per_huge_page(h
);
575 if (unlikely(nr_pages
> MAX_ORDER_NR_PAGES
)) {
576 __copy_gigantic_page(dst
, src
, nr_pages
);
581 BUG_ON(!PageTransHuge(src
));
582 nr_pages
= hpage_nr_pages(src
);
585 for (i
= 0; i
< nr_pages
; i
++) {
587 copy_highpage(dst
+ i
, src
+ i
);
592 * Copy the page to its new location
594 void migrate_page_states(struct page
*newpage
, struct page
*page
)
599 SetPageError(newpage
);
600 if (PageReferenced(page
))
601 SetPageReferenced(newpage
);
602 if (PageUptodate(page
))
603 SetPageUptodate(newpage
);
604 if (TestClearPageActive(page
)) {
605 VM_BUG_ON_PAGE(PageUnevictable(page
), page
);
606 SetPageActive(newpage
);
607 } else if (TestClearPageUnevictable(page
))
608 SetPageUnevictable(newpage
);
609 if (PageWorkingset(page
))
610 SetPageWorkingset(newpage
);
611 if (PageChecked(page
))
612 SetPageChecked(newpage
);
613 if (PageMappedToDisk(page
))
614 SetPageMappedToDisk(newpage
);
616 /* Move dirty on pages not done by migrate_page_move_mapping() */
618 SetPageDirty(newpage
);
620 if (page_is_young(page
))
621 set_page_young(newpage
);
622 if (page_is_idle(page
))
623 set_page_idle(newpage
);
626 * Copy NUMA information to the new page, to prevent over-eager
627 * future migrations of this same page.
629 cpupid
= page_cpupid_xchg_last(page
, -1);
630 page_cpupid_xchg_last(newpage
, cpupid
);
632 ksm_migrate_page(newpage
, page
);
634 * Please do not reorder this without considering how mm/ksm.c's
635 * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache().
637 if (PageSwapCache(page
))
638 ClearPageSwapCache(page
);
639 ClearPagePrivate(page
);
640 set_page_private(page
, 0);
643 * If any waiters have accumulated on the new page then
646 if (PageWriteback(newpage
))
647 end_page_writeback(newpage
);
649 copy_page_owner(page
, newpage
);
651 mem_cgroup_migrate(page
, newpage
);
653 EXPORT_SYMBOL(migrate_page_states
);
655 void migrate_page_copy(struct page
*newpage
, struct page
*page
)
657 if (PageHuge(page
) || PageTransHuge(page
))
658 copy_huge_page(newpage
, page
);
660 copy_highpage(newpage
, page
);
662 migrate_page_states(newpage
, page
);
664 EXPORT_SYMBOL(migrate_page_copy
);
666 /************************************************************
667 * Migration functions
668 ***********************************************************/
671 * Common logic to directly migrate a single LRU page suitable for
672 * pages that do not use PagePrivate/PagePrivate2.
674 * Pages are locked upon entry and exit.
676 int migrate_page(struct address_space
*mapping
,
677 struct page
*newpage
, struct page
*page
,
678 enum migrate_mode mode
)
682 BUG_ON(PageWriteback(page
)); /* Writeback must be complete */
684 rc
= migrate_page_move_mapping(mapping
, newpage
, page
, 0);
686 if (rc
!= MIGRATEPAGE_SUCCESS
)
689 if (mode
!= MIGRATE_SYNC_NO_COPY
)
690 migrate_page_copy(newpage
, page
);
692 migrate_page_states(newpage
, page
);
693 return MIGRATEPAGE_SUCCESS
;
695 EXPORT_SYMBOL(migrate_page
);
698 /* Returns true if all buffers are successfully locked */
699 static bool buffer_migrate_lock_buffers(struct buffer_head
*head
,
700 enum migrate_mode mode
)
702 struct buffer_head
*bh
= head
;
704 /* Simple case, sync compaction */
705 if (mode
!= MIGRATE_ASYNC
) {
708 bh
= bh
->b_this_page
;
710 } while (bh
!= head
);
715 /* async case, we cannot block on lock_buffer so use trylock_buffer */
717 if (!trylock_buffer(bh
)) {
719 * We failed to lock the buffer and cannot stall in
720 * async migration. Release the taken locks
722 struct buffer_head
*failed_bh
= bh
;
724 while (bh
!= failed_bh
) {
726 bh
= bh
->b_this_page
;
731 bh
= bh
->b_this_page
;
732 } while (bh
!= head
);
736 static int __buffer_migrate_page(struct address_space
*mapping
,
737 struct page
*newpage
, struct page
*page
, enum migrate_mode mode
,
740 struct buffer_head
*bh
, *head
;
744 if (!page_has_buffers(page
))
745 return migrate_page(mapping
, newpage
, page
, mode
);
747 /* Check whether page does not have extra refs before we do more work */
748 expected_count
= expected_page_refs(mapping
, page
);
749 if (page_count(page
) != expected_count
)
752 head
= page_buffers(page
);
753 if (!buffer_migrate_lock_buffers(head
, mode
))
758 bool invalidated
= false;
762 spin_lock(&mapping
->private_lock
);
765 if (atomic_read(&bh
->b_count
)) {
769 bh
= bh
->b_this_page
;
770 } while (bh
!= head
);
776 spin_unlock(&mapping
->private_lock
);
777 invalidate_bh_lrus();
779 goto recheck_buffers
;
783 rc
= migrate_page_move_mapping(mapping
, newpage
, page
, 0);
784 if (rc
!= MIGRATEPAGE_SUCCESS
)
787 ClearPagePrivate(page
);
788 set_page_private(newpage
, page_private(page
));
789 set_page_private(page
, 0);
795 set_bh_page(bh
, newpage
, bh_offset(bh
));
796 bh
= bh
->b_this_page
;
798 } while (bh
!= head
);
800 SetPagePrivate(newpage
);
802 if (mode
!= MIGRATE_SYNC_NO_COPY
)
803 migrate_page_copy(newpage
, page
);
805 migrate_page_states(newpage
, page
);
807 rc
= MIGRATEPAGE_SUCCESS
;
810 spin_unlock(&mapping
->private_lock
);
814 bh
= bh
->b_this_page
;
816 } while (bh
!= head
);
822 * Migration function for pages with buffers. This function can only be used
823 * if the underlying filesystem guarantees that no other references to "page"
824 * exist. For example attached buffer heads are accessed only under page lock.
826 int buffer_migrate_page(struct address_space
*mapping
,
827 struct page
*newpage
, struct page
*page
, enum migrate_mode mode
)
829 return __buffer_migrate_page(mapping
, newpage
, page
, mode
, false);
831 EXPORT_SYMBOL(buffer_migrate_page
);
834 * Same as above except that this variant is more careful and checks that there
835 * are also no buffer head references. This function is the right one for
836 * mappings where buffer heads are directly looked up and referenced (such as
837 * block device mappings).
839 int buffer_migrate_page_norefs(struct address_space
*mapping
,
840 struct page
*newpage
, struct page
*page
, enum migrate_mode mode
)
842 return __buffer_migrate_page(mapping
, newpage
, page
, mode
, true);
847 * Writeback a page to clean the dirty state
849 static int writeout(struct address_space
*mapping
, struct page
*page
)
851 struct writeback_control wbc
= {
852 .sync_mode
= WB_SYNC_NONE
,
855 .range_end
= LLONG_MAX
,
860 if (!mapping
->a_ops
->writepage
)
861 /* No write method for the address space */
864 if (!clear_page_dirty_for_io(page
))
865 /* Someone else already triggered a write */
869 * A dirty page may imply that the underlying filesystem has
870 * the page on some queue. So the page must be clean for
871 * migration. Writeout may mean we loose the lock and the
872 * page state is no longer what we checked for earlier.
873 * At this point we know that the migration attempt cannot
876 remove_migration_ptes(page
, page
, false);
878 rc
= mapping
->a_ops
->writepage(page
, &wbc
);
880 if (rc
!= AOP_WRITEPAGE_ACTIVATE
)
881 /* unlocked. Relock */
884 return (rc
< 0) ? -EIO
: -EAGAIN
;
888 * Default handling if a filesystem does not provide a migration function.
890 static int fallback_migrate_page(struct address_space
*mapping
,
891 struct page
*newpage
, struct page
*page
, enum migrate_mode mode
)
893 if (PageDirty(page
)) {
894 /* Only writeback pages in full synchronous migration */
897 case MIGRATE_SYNC_NO_COPY
:
902 return writeout(mapping
, page
);
906 * Buffers may be managed in a filesystem specific way.
907 * We must have no buffers or drop them.
909 if (page_has_private(page
) &&
910 !try_to_release_page(page
, GFP_KERNEL
))
911 return mode
== MIGRATE_SYNC
? -EAGAIN
: -EBUSY
;
913 return migrate_page(mapping
, newpage
, page
, mode
);
917 * Move a page to a newly allocated page
918 * The page is locked and all ptes have been successfully removed.
920 * The new page will have replaced the old page if this function
925 * MIGRATEPAGE_SUCCESS - success
927 static int move_to_new_page(struct page
*newpage
, struct page
*page
,
928 enum migrate_mode mode
)
930 struct address_space
*mapping
;
932 bool is_lru
= !__PageMovable(page
);
934 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
935 VM_BUG_ON_PAGE(!PageLocked(newpage
), newpage
);
937 mapping
= page_mapping(page
);
939 if (likely(is_lru
)) {
941 rc
= migrate_page(mapping
, newpage
, page
, mode
);
942 else if (mapping
->a_ops
->migratepage
)
944 * Most pages have a mapping and most filesystems
945 * provide a migratepage callback. Anonymous pages
946 * are part of swap space which also has its own
947 * migratepage callback. This is the most common path
948 * for page migration.
950 rc
= mapping
->a_ops
->migratepage(mapping
, newpage
,
953 rc
= fallback_migrate_page(mapping
, newpage
,
957 * In case of non-lru page, it could be released after
958 * isolation step. In that case, we shouldn't try migration.
960 VM_BUG_ON_PAGE(!PageIsolated(page
), page
);
961 if (!PageMovable(page
)) {
962 rc
= MIGRATEPAGE_SUCCESS
;
963 __ClearPageIsolated(page
);
967 rc
= mapping
->a_ops
->migratepage(mapping
, newpage
,
969 WARN_ON_ONCE(rc
== MIGRATEPAGE_SUCCESS
&&
970 !PageIsolated(page
));
974 * When successful, old pagecache page->mapping must be cleared before
975 * page is freed; but stats require that PageAnon be left as PageAnon.
977 if (rc
== MIGRATEPAGE_SUCCESS
) {
978 if (__PageMovable(page
)) {
979 VM_BUG_ON_PAGE(!PageIsolated(page
), page
);
982 * We clear PG_movable under page_lock so any compactor
983 * cannot try to migrate this page.
985 __ClearPageIsolated(page
);
989 * Anonymous and movable page->mapping will be cleard by
990 * free_pages_prepare so don't reset it here for keeping
991 * the type to work PageAnon, for example.
993 if (!PageMappingFlags(page
))
994 page
->mapping
= NULL
;
996 if (likely(!is_zone_device_page(newpage
)))
997 flush_dcache_page(newpage
);
1004 static int __unmap_and_move(struct page
*page
, struct page
*newpage
,
1005 int force
, enum migrate_mode mode
)
1008 int page_was_mapped
= 0;
1009 struct anon_vma
*anon_vma
= NULL
;
1010 bool is_lru
= !__PageMovable(page
);
1012 if (!trylock_page(page
)) {
1013 if (!force
|| mode
== MIGRATE_ASYNC
)
1017 * It's not safe for direct compaction to call lock_page.
1018 * For example, during page readahead pages are added locked
1019 * to the LRU. Later, when the IO completes the pages are
1020 * marked uptodate and unlocked. However, the queueing
1021 * could be merging multiple pages for one bio (e.g.
1022 * mpage_readpages). If an allocation happens for the
1023 * second or third page, the process can end up locking
1024 * the same page twice and deadlocking. Rather than
1025 * trying to be clever about what pages can be locked,
1026 * avoid the use of lock_page for direct compaction
1029 if (current
->flags
& PF_MEMALLOC
)
1035 if (PageWriteback(page
)) {
1037 * Only in the case of a full synchronous migration is it
1038 * necessary to wait for PageWriteback. In the async case,
1039 * the retry loop is too short and in the sync-light case,
1040 * the overhead of stalling is too much
1044 case MIGRATE_SYNC_NO_COPY
:
1052 wait_on_page_writeback(page
);
1056 * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
1057 * we cannot notice that anon_vma is freed while we migrates a page.
1058 * This get_anon_vma() delays freeing anon_vma pointer until the end
1059 * of migration. File cache pages are no problem because of page_lock()
1060 * File Caches may use write_page() or lock_page() in migration, then,
1061 * just care Anon page here.
1063 * Only page_get_anon_vma() understands the subtleties of
1064 * getting a hold on an anon_vma from outside one of its mms.
1065 * But if we cannot get anon_vma, then we won't need it anyway,
1066 * because that implies that the anon page is no longer mapped
1067 * (and cannot be remapped so long as we hold the page lock).
1069 if (PageAnon(page
) && !PageKsm(page
))
1070 anon_vma
= page_get_anon_vma(page
);
1073 * Block others from accessing the new page when we get around to
1074 * establishing additional references. We are usually the only one
1075 * holding a reference to newpage at this point. We used to have a BUG
1076 * here if trylock_page(newpage) fails, but would like to allow for
1077 * cases where there might be a race with the previous use of newpage.
1078 * This is much like races on refcount of oldpage: just don't BUG().
1080 if (unlikely(!trylock_page(newpage
)))
1083 if (unlikely(!is_lru
)) {
1084 rc
= move_to_new_page(newpage
, page
, mode
);
1085 goto out_unlock_both
;
1089 * Corner case handling:
1090 * 1. When a new swap-cache page is read into, it is added to the LRU
1091 * and treated as swapcache but it has no rmap yet.
1092 * Calling try_to_unmap() against a page->mapping==NULL page will
1093 * trigger a BUG. So handle it here.
1094 * 2. An orphaned page (see truncate_complete_page) might have
1095 * fs-private metadata. The page can be picked up due to memory
1096 * offlining. Everywhere else except page reclaim, the page is
1097 * invisible to the vm, so the page can not be migrated. So try to
1098 * free the metadata, so the page can be freed.
1100 if (!page
->mapping
) {
1101 VM_BUG_ON_PAGE(PageAnon(page
), page
);
1102 if (page_has_private(page
)) {
1103 try_to_free_buffers(page
);
1104 goto out_unlock_both
;
1106 } else if (page_mapped(page
)) {
1107 /* Establish migration ptes */
1108 VM_BUG_ON_PAGE(PageAnon(page
) && !PageKsm(page
) && !anon_vma
,
1111 TTU_MIGRATION
|TTU_IGNORE_MLOCK
|TTU_IGNORE_ACCESS
);
1112 page_was_mapped
= 1;
1115 if (!page_mapped(page
))
1116 rc
= move_to_new_page(newpage
, page
, mode
);
1118 if (page_was_mapped
)
1119 remove_migration_ptes(page
,
1120 rc
== MIGRATEPAGE_SUCCESS
? newpage
: page
, false);
1123 unlock_page(newpage
);
1125 /* Drop an anon_vma reference if we took one */
1127 put_anon_vma(anon_vma
);
1131 * If migration is successful, decrease refcount of the newpage
1132 * which will not free the page because new page owner increased
1133 * refcounter. As well, if it is LRU page, add the page to LRU
1134 * list in here. Use the old state of the isolated source page to
1135 * determine if we migrated a LRU page. newpage was already unlocked
1136 * and possibly modified by its owner - don't rely on the page
1139 if (rc
== MIGRATEPAGE_SUCCESS
) {
1140 if (unlikely(!is_lru
))
1143 putback_lru_page(newpage
);
1150 * gcc 4.7 and 4.8 on arm get an ICEs when inlining unmap_and_move(). Work
1153 #if defined(CONFIG_ARM) && \
1154 defined(GCC_VERSION) && GCC_VERSION < 40900 && GCC_VERSION >= 40700
1155 #define ICE_noinline noinline
1157 #define ICE_noinline
1161 * Obtain the lock on page, remove all ptes and migrate the page
1162 * to the newly allocated page in newpage.
1164 static ICE_noinline
int unmap_and_move(new_page_t get_new_page
,
1165 free_page_t put_new_page
,
1166 unsigned long private, struct page
*page
,
1167 int force
, enum migrate_mode mode
,
1168 enum migrate_reason reason
)
1170 int rc
= MIGRATEPAGE_SUCCESS
;
1171 struct page
*newpage
;
1173 if (!thp_migration_supported() && PageTransHuge(page
))
1176 newpage
= get_new_page(page
, private);
1180 if (page_count(page
) == 1) {
1181 /* page was freed from under us. So we are done. */
1182 ClearPageActive(page
);
1183 ClearPageUnevictable(page
);
1184 if (unlikely(__PageMovable(page
))) {
1186 if (!PageMovable(page
))
1187 __ClearPageIsolated(page
);
1191 put_new_page(newpage
, private);
1197 rc
= __unmap_and_move(page
, newpage
, force
, mode
);
1198 if (rc
== MIGRATEPAGE_SUCCESS
)
1199 set_page_owner_migrate_reason(newpage
, reason
);
1202 if (rc
!= -EAGAIN
) {
1204 * A page that has been migrated has all references
1205 * removed and will be freed. A page that has not been
1206 * migrated will have kepts its references and be
1209 list_del(&page
->lru
);
1212 * Compaction can migrate also non-LRU pages which are
1213 * not accounted to NR_ISOLATED_*. They can be recognized
1216 if (likely(!__PageMovable(page
)))
1217 mod_node_page_state(page_pgdat(page
), NR_ISOLATED_ANON
+
1218 page_is_file_cache(page
), -hpage_nr_pages(page
));
1222 * If migration is successful, releases reference grabbed during
1223 * isolation. Otherwise, restore the page to right list unless
1226 if (rc
== MIGRATEPAGE_SUCCESS
) {
1228 if (reason
== MR_MEMORY_FAILURE
) {
1230 * Set PG_HWPoison on just freed page
1231 * intentionally. Although it's rather weird,
1232 * it's how HWPoison flag works at the moment.
1234 if (set_hwpoison_free_buddy_page(page
))
1235 num_poisoned_pages_inc();
1238 if (rc
!= -EAGAIN
) {
1239 if (likely(!__PageMovable(page
))) {
1240 putback_lru_page(page
);
1245 if (PageMovable(page
))
1246 putback_movable_page(page
);
1248 __ClearPageIsolated(page
);
1254 put_new_page(newpage
, private);
1263 * Counterpart of unmap_and_move_page() for hugepage migration.
1265 * This function doesn't wait the completion of hugepage I/O
1266 * because there is no race between I/O and migration for hugepage.
1267 * Note that currently hugepage I/O occurs only in direct I/O
1268 * where no lock is held and PG_writeback is irrelevant,
1269 * and writeback status of all subpages are counted in the reference
1270 * count of the head page (i.e. if all subpages of a 2MB hugepage are
1271 * under direct I/O, the reference of the head page is 512 and a bit more.)
1272 * This means that when we try to migrate hugepage whose subpages are
1273 * doing direct I/O, some references remain after try_to_unmap() and
1274 * hugepage migration fails without data corruption.
1276 * There is also no race when direct I/O is issued on the page under migration,
1277 * because then pte is replaced with migration swap entry and direct I/O code
1278 * will wait in the page fault for migration to complete.
1280 static int unmap_and_move_huge_page(new_page_t get_new_page
,
1281 free_page_t put_new_page
, unsigned long private,
1282 struct page
*hpage
, int force
,
1283 enum migrate_mode mode
, int reason
)
1286 int page_was_mapped
= 0;
1287 struct page
*new_hpage
;
1288 struct anon_vma
*anon_vma
= NULL
;
1291 * Migratability of hugepages depends on architectures and their size.
1292 * This check is necessary because some callers of hugepage migration
1293 * like soft offline and memory hotremove don't walk through page
1294 * tables or check whether the hugepage is pmd-based or not before
1295 * kicking migration.
1297 if (!hugepage_migration_supported(page_hstate(hpage
))) {
1298 putback_active_hugepage(hpage
);
1302 new_hpage
= get_new_page(hpage
, private);
1306 if (!trylock_page(hpage
)) {
1311 case MIGRATE_SYNC_NO_COPY
:
1320 * Check for pages which are in the process of being freed. Without
1321 * page_mapping() set, hugetlbfs specific move page routine will not
1322 * be called and we could leak usage counts for subpools.
1324 if (page_private(hpage
) && !page_mapping(hpage
)) {
1329 if (PageAnon(hpage
))
1330 anon_vma
= page_get_anon_vma(hpage
);
1332 if (unlikely(!trylock_page(new_hpage
)))
1335 if (page_mapped(hpage
)) {
1337 TTU_MIGRATION
|TTU_IGNORE_MLOCK
|TTU_IGNORE_ACCESS
);
1338 page_was_mapped
= 1;
1341 if (!page_mapped(hpage
))
1342 rc
= move_to_new_page(new_hpage
, hpage
, mode
);
1344 if (page_was_mapped
)
1345 remove_migration_ptes(hpage
,
1346 rc
== MIGRATEPAGE_SUCCESS
? new_hpage
: hpage
, false);
1348 unlock_page(new_hpage
);
1352 put_anon_vma(anon_vma
);
1354 if (rc
== MIGRATEPAGE_SUCCESS
) {
1355 move_hugetlb_state(hpage
, new_hpage
, reason
);
1356 put_new_page
= NULL
;
1363 putback_active_hugepage(hpage
);
1366 * If migration was not successful and there's a freeing callback, use
1367 * it. Otherwise, put_page() will drop the reference grabbed during
1371 put_new_page(new_hpage
, private);
1373 putback_active_hugepage(new_hpage
);
1379 * migrate_pages - migrate the pages specified in a list, to the free pages
1380 * supplied as the target for the page migration
1382 * @from: The list of pages to be migrated.
1383 * @get_new_page: The function used to allocate free pages to be used
1384 * as the target of the page migration.
1385 * @put_new_page: The function used to free target pages if migration
1386 * fails, or NULL if no special handling is necessary.
1387 * @private: Private data to be passed on to get_new_page()
1388 * @mode: The migration mode that specifies the constraints for
1389 * page migration, if any.
1390 * @reason: The reason for page migration.
1392 * The function returns after 10 attempts or if no pages are movable any more
1393 * because the list has become empty or no retryable pages exist any more.
1394 * The caller should call putback_movable_pages() to return pages to the LRU
1395 * or free list only if ret != 0.
1397 * Returns the number of pages that were not migrated, or an error code.
1399 int migrate_pages(struct list_head
*from
, new_page_t get_new_page
,
1400 free_page_t put_new_page
, unsigned long private,
1401 enum migrate_mode mode
, int reason
)
1405 int nr_succeeded
= 0;
1409 int swapwrite
= current
->flags
& PF_SWAPWRITE
;
1413 current
->flags
|= PF_SWAPWRITE
;
1415 for(pass
= 0; pass
< 10 && retry
; pass
++) {
1418 list_for_each_entry_safe(page
, page2
, from
, lru
) {
1423 rc
= unmap_and_move_huge_page(get_new_page
,
1424 put_new_page
, private, page
,
1425 pass
> 2, mode
, reason
);
1427 rc
= unmap_and_move(get_new_page
, put_new_page
,
1428 private, page
, pass
> 2, mode
,
1434 * THP migration might be unsupported or the
1435 * allocation could've failed so we should
1436 * retry on the same page with the THP split
1439 * Head page is retried immediately and tail
1440 * pages are added to the tail of the list so
1441 * we encounter them after the rest of the list
1444 if (PageTransHuge(page
) && !PageHuge(page
)) {
1446 rc
= split_huge_page_to_list(page
, from
);
1449 list_safe_reset_next(page
, page2
, lru
);
1458 case MIGRATEPAGE_SUCCESS
:
1463 * Permanent failure (-EBUSY, -ENOSYS, etc.):
1464 * unlike -EAGAIN case, the failed page is
1465 * removed from migration page list and not
1466 * retried in the next outer loop.
1477 count_vm_events(PGMIGRATE_SUCCESS
, nr_succeeded
);
1479 count_vm_events(PGMIGRATE_FAIL
, nr_failed
);
1480 trace_mm_migrate_pages(nr_succeeded
, nr_failed
, mode
, reason
);
1483 current
->flags
&= ~PF_SWAPWRITE
;
1490 static int store_status(int __user
*status
, int start
, int value
, int nr
)
1493 if (put_user(value
, status
+ start
))
1501 static int do_move_pages_to_node(struct mm_struct
*mm
,
1502 struct list_head
*pagelist
, int node
)
1506 if (list_empty(pagelist
))
1509 err
= migrate_pages(pagelist
, alloc_new_node_page
, NULL
, node
,
1510 MIGRATE_SYNC
, MR_SYSCALL
);
1512 putback_movable_pages(pagelist
);
1517 * Resolves the given address to a struct page, isolates it from the LRU and
1518 * puts it to the given pagelist.
1519 * Returns -errno if the page cannot be found/isolated or 0 when it has been
1520 * queued or the page doesn't need to be migrated because it is already on
1523 static int add_page_for_migration(struct mm_struct
*mm
, unsigned long addr
,
1524 int node
, struct list_head
*pagelist
, bool migrate_all
)
1526 struct vm_area_struct
*vma
;
1528 unsigned int follflags
;
1531 down_read(&mm
->mmap_sem
);
1533 vma
= find_vma(mm
, addr
);
1534 if (!vma
|| addr
< vma
->vm_start
|| !vma_migratable(vma
))
1537 /* FOLL_DUMP to ignore special (like zero) pages */
1538 follflags
= FOLL_GET
| FOLL_DUMP
;
1539 page
= follow_page(vma
, addr
, follflags
);
1541 err
= PTR_ERR(page
);
1550 if (page_to_nid(page
) == node
)
1554 if (page_mapcount(page
) > 1 && !migrate_all
)
1557 if (PageHuge(page
)) {
1558 if (PageHead(page
)) {
1559 isolate_huge_page(page
, pagelist
);
1565 head
= compound_head(page
);
1566 err
= isolate_lru_page(head
);
1571 list_add_tail(&head
->lru
, pagelist
);
1572 mod_node_page_state(page_pgdat(head
),
1573 NR_ISOLATED_ANON
+ page_is_file_cache(head
),
1574 hpage_nr_pages(head
));
1578 * Either remove the duplicate refcount from
1579 * isolate_lru_page() or drop the page ref if it was
1584 up_read(&mm
->mmap_sem
);
1589 * Migrate an array of page address onto an array of nodes and fill
1590 * the corresponding array of status.
1592 static int do_pages_move(struct mm_struct
*mm
, nodemask_t task_nodes
,
1593 unsigned long nr_pages
,
1594 const void __user
* __user
*pages
,
1595 const int __user
*nodes
,
1596 int __user
*status
, int flags
)
1598 int current_node
= NUMA_NO_NODE
;
1599 LIST_HEAD(pagelist
);
1605 for (i
= start
= 0; i
< nr_pages
; i
++) {
1606 const void __user
*p
;
1611 if (get_user(p
, pages
+ i
))
1613 if (get_user(node
, nodes
+ i
))
1615 addr
= (unsigned long)untagged_addr(p
);
1618 if (node
< 0 || node
>= MAX_NUMNODES
)
1620 if (!node_state(node
, N_MEMORY
))
1624 if (!node_isset(node
, task_nodes
))
1627 if (current_node
== NUMA_NO_NODE
) {
1628 current_node
= node
;
1630 } else if (node
!= current_node
) {
1631 err
= do_move_pages_to_node(mm
, &pagelist
, current_node
);
1634 err
= store_status(status
, start
, current_node
, i
- start
);
1638 current_node
= node
;
1642 * Errors in the page lookup or isolation are not fatal and we simply
1643 * report them via status
1645 err
= add_page_for_migration(mm
, addr
, current_node
,
1646 &pagelist
, flags
& MPOL_MF_MOVE_ALL
);
1650 err
= store_status(status
, i
, err
, 1);
1654 err
= do_move_pages_to_node(mm
, &pagelist
, current_node
);
1658 err
= store_status(status
, start
, current_node
, i
- start
);
1662 current_node
= NUMA_NO_NODE
;
1665 if (list_empty(&pagelist
))
1668 /* Make sure we do not overwrite the existing error */
1669 err1
= do_move_pages_to_node(mm
, &pagelist
, current_node
);
1671 err1
= store_status(status
, start
, current_node
, i
- start
);
1679 * Determine the nodes of an array of pages and store it in an array of status.
1681 static void do_pages_stat_array(struct mm_struct
*mm
, unsigned long nr_pages
,
1682 const void __user
**pages
, int *status
)
1686 down_read(&mm
->mmap_sem
);
1688 for (i
= 0; i
< nr_pages
; i
++) {
1689 unsigned long addr
= (unsigned long)(*pages
);
1690 struct vm_area_struct
*vma
;
1694 vma
= find_vma(mm
, addr
);
1695 if (!vma
|| addr
< vma
->vm_start
)
1698 /* FOLL_DUMP to ignore special (like zero) pages */
1699 page
= follow_page(vma
, addr
, FOLL_DUMP
);
1701 err
= PTR_ERR(page
);
1705 err
= page
? page_to_nid(page
) : -ENOENT
;
1713 up_read(&mm
->mmap_sem
);
1717 * Determine the nodes of a user array of pages and store it in
1718 * a user array of status.
1720 static int do_pages_stat(struct mm_struct
*mm
, unsigned long nr_pages
,
1721 const void __user
* __user
*pages
,
1724 #define DO_PAGES_STAT_CHUNK_NR 16
1725 const void __user
*chunk_pages
[DO_PAGES_STAT_CHUNK_NR
];
1726 int chunk_status
[DO_PAGES_STAT_CHUNK_NR
];
1729 unsigned long chunk_nr
;
1731 chunk_nr
= nr_pages
;
1732 if (chunk_nr
> DO_PAGES_STAT_CHUNK_NR
)
1733 chunk_nr
= DO_PAGES_STAT_CHUNK_NR
;
1735 if (copy_from_user(chunk_pages
, pages
, chunk_nr
* sizeof(*chunk_pages
)))
1738 do_pages_stat_array(mm
, chunk_nr
, chunk_pages
, chunk_status
);
1740 if (copy_to_user(status
, chunk_status
, chunk_nr
* sizeof(*status
)))
1745 nr_pages
-= chunk_nr
;
1747 return nr_pages
? -EFAULT
: 0;
1751 * Move a list of pages in the address space of the currently executing
1754 static int kernel_move_pages(pid_t pid
, unsigned long nr_pages
,
1755 const void __user
* __user
*pages
,
1756 const int __user
*nodes
,
1757 int __user
*status
, int flags
)
1759 struct task_struct
*task
;
1760 struct mm_struct
*mm
;
1762 nodemask_t task_nodes
;
1765 if (flags
& ~(MPOL_MF_MOVE
|MPOL_MF_MOVE_ALL
))
1768 if ((flags
& MPOL_MF_MOVE_ALL
) && !capable(CAP_SYS_NICE
))
1771 /* Find the mm_struct */
1773 task
= pid
? find_task_by_vpid(pid
) : current
;
1778 get_task_struct(task
);
1781 * Check if this process has the right to modify the specified
1782 * process. Use the regular "ptrace_may_access()" checks.
1784 if (!ptrace_may_access(task
, PTRACE_MODE_READ_REALCREDS
)) {
1791 err
= security_task_movememory(task
);
1795 task_nodes
= cpuset_mems_allowed(task
);
1796 mm
= get_task_mm(task
);
1797 put_task_struct(task
);
1803 err
= do_pages_move(mm
, task_nodes
, nr_pages
, pages
,
1804 nodes
, status
, flags
);
1806 err
= do_pages_stat(mm
, nr_pages
, pages
, status
);
1812 put_task_struct(task
);
1816 SYSCALL_DEFINE6(move_pages
, pid_t
, pid
, unsigned long, nr_pages
,
1817 const void __user
* __user
*, pages
,
1818 const int __user
*, nodes
,
1819 int __user
*, status
, int, flags
)
1821 return kernel_move_pages(pid
, nr_pages
, pages
, nodes
, status
, flags
);
1824 #ifdef CONFIG_COMPAT
1825 COMPAT_SYSCALL_DEFINE6(move_pages
, pid_t
, pid
, compat_ulong_t
, nr_pages
,
1826 compat_uptr_t __user
*, pages32
,
1827 const int __user
*, nodes
,
1828 int __user
*, status
,
1831 const void __user
* __user
*pages
;
1834 pages
= compat_alloc_user_space(nr_pages
* sizeof(void *));
1835 for (i
= 0; i
< nr_pages
; i
++) {
1838 if (get_user(p
, pages32
+ i
) ||
1839 put_user(compat_ptr(p
), pages
+ i
))
1842 return kernel_move_pages(pid
, nr_pages
, pages
, nodes
, status
, flags
);
1844 #endif /* CONFIG_COMPAT */
1846 #ifdef CONFIG_NUMA_BALANCING
1848 * Returns true if this is a safe migration target node for misplaced NUMA
1849 * pages. Currently it only checks the watermarks which crude
1851 static bool migrate_balanced_pgdat(struct pglist_data
*pgdat
,
1852 unsigned long nr_migrate_pages
)
1856 for (z
= pgdat
->nr_zones
- 1; z
>= 0; z
--) {
1857 struct zone
*zone
= pgdat
->node_zones
+ z
;
1859 if (!populated_zone(zone
))
1862 /* Avoid waking kswapd by allocating pages_to_migrate pages. */
1863 if (!zone_watermark_ok(zone
, 0,
1864 high_wmark_pages(zone
) +
1873 static struct page
*alloc_misplaced_dst_page(struct page
*page
,
1876 int nid
= (int) data
;
1877 struct page
*newpage
;
1879 newpage
= __alloc_pages_node(nid
,
1880 (GFP_HIGHUSER_MOVABLE
|
1881 __GFP_THISNODE
| __GFP_NOMEMALLOC
|
1882 __GFP_NORETRY
| __GFP_NOWARN
) &
1888 static int numamigrate_isolate_page(pg_data_t
*pgdat
, struct page
*page
)
1892 VM_BUG_ON_PAGE(compound_order(page
) && !PageTransHuge(page
), page
);
1894 /* Avoid migrating to a node that is nearly full */
1895 if (!migrate_balanced_pgdat(pgdat
, compound_nr(page
)))
1898 if (isolate_lru_page(page
))
1902 * migrate_misplaced_transhuge_page() skips page migration's usual
1903 * check on page_count(), so we must do it here, now that the page
1904 * has been isolated: a GUP pin, or any other pin, prevents migration.
1905 * The expected page count is 3: 1 for page's mapcount and 1 for the
1906 * caller's pin and 1 for the reference taken by isolate_lru_page().
1908 if (PageTransHuge(page
) && page_count(page
) != 3) {
1909 putback_lru_page(page
);
1913 page_lru
= page_is_file_cache(page
);
1914 mod_node_page_state(page_pgdat(page
), NR_ISOLATED_ANON
+ page_lru
,
1915 hpage_nr_pages(page
));
1918 * Isolating the page has taken another reference, so the
1919 * caller's reference can be safely dropped without the page
1920 * disappearing underneath us during migration.
1926 bool pmd_trans_migrating(pmd_t pmd
)
1928 struct page
*page
= pmd_page(pmd
);
1929 return PageLocked(page
);
1933 * Attempt to migrate a misplaced page to the specified destination
1934 * node. Caller is expected to have an elevated reference count on
1935 * the page that will be dropped by this function before returning.
1937 int migrate_misplaced_page(struct page
*page
, struct vm_area_struct
*vma
,
1940 pg_data_t
*pgdat
= NODE_DATA(node
);
1943 LIST_HEAD(migratepages
);
1946 * Don't migrate file pages that are mapped in multiple processes
1947 * with execute permissions as they are probably shared libraries.
1949 if (page_mapcount(page
) != 1 && page_is_file_cache(page
) &&
1950 (vma
->vm_flags
& VM_EXEC
))
1954 * Also do not migrate dirty pages as not all filesystems can move
1955 * dirty pages in MIGRATE_ASYNC mode which is a waste of cycles.
1957 if (page_is_file_cache(page
) && PageDirty(page
))
1960 isolated
= numamigrate_isolate_page(pgdat
, page
);
1964 list_add(&page
->lru
, &migratepages
);
1965 nr_remaining
= migrate_pages(&migratepages
, alloc_misplaced_dst_page
,
1966 NULL
, node
, MIGRATE_ASYNC
,
1969 if (!list_empty(&migratepages
)) {
1970 list_del(&page
->lru
);
1971 dec_node_page_state(page
, NR_ISOLATED_ANON
+
1972 page_is_file_cache(page
));
1973 putback_lru_page(page
);
1977 count_vm_numa_event(NUMA_PAGE_MIGRATE
);
1978 BUG_ON(!list_empty(&migratepages
));
1985 #endif /* CONFIG_NUMA_BALANCING */
1987 #if defined(CONFIG_NUMA_BALANCING) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
1989 * Migrates a THP to a given target node. page must be locked and is unlocked
1992 int migrate_misplaced_transhuge_page(struct mm_struct
*mm
,
1993 struct vm_area_struct
*vma
,
1994 pmd_t
*pmd
, pmd_t entry
,
1995 unsigned long address
,
1996 struct page
*page
, int node
)
1999 pg_data_t
*pgdat
= NODE_DATA(node
);
2001 struct page
*new_page
= NULL
;
2002 int page_lru
= page_is_file_cache(page
);
2003 unsigned long start
= address
& HPAGE_PMD_MASK
;
2005 new_page
= alloc_pages_node(node
,
2006 (GFP_TRANSHUGE_LIGHT
| __GFP_THISNODE
),
2010 prep_transhuge_page(new_page
);
2012 isolated
= numamigrate_isolate_page(pgdat
, page
);
2018 /* Prepare a page as a migration target */
2019 __SetPageLocked(new_page
);
2020 if (PageSwapBacked(page
))
2021 __SetPageSwapBacked(new_page
);
2023 /* anon mapping, we can simply copy page->mapping to the new page: */
2024 new_page
->mapping
= page
->mapping
;
2025 new_page
->index
= page
->index
;
2026 /* flush the cache before copying using the kernel virtual address */
2027 flush_cache_range(vma
, start
, start
+ HPAGE_PMD_SIZE
);
2028 migrate_page_copy(new_page
, page
);
2029 WARN_ON(PageLRU(new_page
));
2031 /* Recheck the target PMD */
2032 ptl
= pmd_lock(mm
, pmd
);
2033 if (unlikely(!pmd_same(*pmd
, entry
) || !page_ref_freeze(page
, 2))) {
2036 /* Reverse changes made by migrate_page_copy() */
2037 if (TestClearPageActive(new_page
))
2038 SetPageActive(page
);
2039 if (TestClearPageUnevictable(new_page
))
2040 SetPageUnevictable(page
);
2042 unlock_page(new_page
);
2043 put_page(new_page
); /* Free it */
2045 /* Retake the callers reference and putback on LRU */
2047 putback_lru_page(page
);
2048 mod_node_page_state(page_pgdat(page
),
2049 NR_ISOLATED_ANON
+ page_lru
, -HPAGE_PMD_NR
);
2054 entry
= mk_huge_pmd(new_page
, vma
->vm_page_prot
);
2055 entry
= maybe_pmd_mkwrite(pmd_mkdirty(entry
), vma
);
2058 * Overwrite the old entry under pagetable lock and establish
2059 * the new PTE. Any parallel GUP will either observe the old
2060 * page blocking on the page lock, block on the page table
2061 * lock or observe the new page. The SetPageUptodate on the
2062 * new page and page_add_new_anon_rmap guarantee the copy is
2063 * visible before the pagetable update.
2065 page_add_anon_rmap(new_page
, vma
, start
, true);
2067 * At this point the pmd is numa/protnone (i.e. non present) and the TLB
2068 * has already been flushed globally. So no TLB can be currently
2069 * caching this non present pmd mapping. There's no need to clear the
2070 * pmd before doing set_pmd_at(), nor to flush the TLB after
2071 * set_pmd_at(). Clearing the pmd here would introduce a race
2072 * condition against MADV_DONTNEED, because MADV_DONTNEED only holds the
2073 * mmap_sem for reading. If the pmd is set to NULL at any given time,
2074 * MADV_DONTNEED won't wait on the pmd lock and it'll skip clearing this
2077 set_pmd_at(mm
, start
, pmd
, entry
);
2078 update_mmu_cache_pmd(vma
, address
, &entry
);
2080 page_ref_unfreeze(page
, 2);
2081 mlock_migrate_page(new_page
, page
);
2082 page_remove_rmap(page
, true);
2083 set_page_owner_migrate_reason(new_page
, MR_NUMA_MISPLACED
);
2087 /* Take an "isolate" reference and put new page on the LRU. */
2089 putback_lru_page(new_page
);
2091 unlock_page(new_page
);
2093 put_page(page
); /* Drop the rmap reference */
2094 put_page(page
); /* Drop the LRU isolation reference */
2096 count_vm_events(PGMIGRATE_SUCCESS
, HPAGE_PMD_NR
);
2097 count_vm_numa_events(NUMA_PAGE_MIGRATE
, HPAGE_PMD_NR
);
2099 mod_node_page_state(page_pgdat(page
),
2100 NR_ISOLATED_ANON
+ page_lru
,
2105 count_vm_events(PGMIGRATE_FAIL
, HPAGE_PMD_NR
);
2106 ptl
= pmd_lock(mm
, pmd
);
2107 if (pmd_same(*pmd
, entry
)) {
2108 entry
= pmd_modify(entry
, vma
->vm_page_prot
);
2109 set_pmd_at(mm
, start
, pmd
, entry
);
2110 update_mmu_cache_pmd(vma
, address
, &entry
);
2119 #endif /* CONFIG_NUMA_BALANCING */
2121 #endif /* CONFIG_NUMA */
2123 #ifdef CONFIG_DEVICE_PRIVATE
2124 static int migrate_vma_collect_hole(unsigned long start
,
2126 struct mm_walk
*walk
)
2128 struct migrate_vma
*migrate
= walk
->private;
2131 for (addr
= start
& PAGE_MASK
; addr
< end
; addr
+= PAGE_SIZE
) {
2132 migrate
->src
[migrate
->npages
] = MIGRATE_PFN_MIGRATE
;
2133 migrate
->dst
[migrate
->npages
] = 0;
2141 static int migrate_vma_collect_skip(unsigned long start
,
2143 struct mm_walk
*walk
)
2145 struct migrate_vma
*migrate
= walk
->private;
2148 for (addr
= start
& PAGE_MASK
; addr
< end
; addr
+= PAGE_SIZE
) {
2149 migrate
->dst
[migrate
->npages
] = 0;
2150 migrate
->src
[migrate
->npages
++] = 0;
2156 static int migrate_vma_collect_pmd(pmd_t
*pmdp
,
2157 unsigned long start
,
2159 struct mm_walk
*walk
)
2161 struct migrate_vma
*migrate
= walk
->private;
2162 struct vm_area_struct
*vma
= walk
->vma
;
2163 struct mm_struct
*mm
= vma
->vm_mm
;
2164 unsigned long addr
= start
, unmapped
= 0;
2169 if (pmd_none(*pmdp
))
2170 return migrate_vma_collect_hole(start
, end
, walk
);
2172 if (pmd_trans_huge(*pmdp
)) {
2175 ptl
= pmd_lock(mm
, pmdp
);
2176 if (unlikely(!pmd_trans_huge(*pmdp
))) {
2181 page
= pmd_page(*pmdp
);
2182 if (is_huge_zero_page(page
)) {
2184 split_huge_pmd(vma
, pmdp
, addr
);
2185 if (pmd_trans_unstable(pmdp
))
2186 return migrate_vma_collect_skip(start
, end
,
2193 if (unlikely(!trylock_page(page
)))
2194 return migrate_vma_collect_skip(start
, end
,
2196 ret
= split_huge_page(page
);
2200 return migrate_vma_collect_skip(start
, end
,
2202 if (pmd_none(*pmdp
))
2203 return migrate_vma_collect_hole(start
, end
,
2208 if (unlikely(pmd_bad(*pmdp
)))
2209 return migrate_vma_collect_skip(start
, end
, walk
);
2211 ptep
= pte_offset_map_lock(mm
, pmdp
, addr
, &ptl
);
2212 arch_enter_lazy_mmu_mode();
2214 for (; addr
< end
; addr
+= PAGE_SIZE
, ptep
++) {
2215 unsigned long mpfn
, pfn
;
2222 if (pte_none(pte
)) {
2223 mpfn
= MIGRATE_PFN_MIGRATE
;
2228 if (!pte_present(pte
)) {
2232 * Only care about unaddressable device page special
2233 * page table entry. Other special swap entries are not
2234 * migratable, and we ignore regular swapped page.
2236 entry
= pte_to_swp_entry(pte
);
2237 if (!is_device_private_entry(entry
))
2240 page
= device_private_entry_to_page(entry
);
2241 mpfn
= migrate_pfn(page_to_pfn(page
)) |
2242 MIGRATE_PFN_MIGRATE
;
2243 if (is_write_device_private_entry(entry
))
2244 mpfn
|= MIGRATE_PFN_WRITE
;
2247 if (is_zero_pfn(pfn
)) {
2248 mpfn
= MIGRATE_PFN_MIGRATE
;
2252 page
= vm_normal_page(migrate
->vma
, addr
, pte
);
2253 mpfn
= migrate_pfn(pfn
) | MIGRATE_PFN_MIGRATE
;
2254 mpfn
|= pte_write(pte
) ? MIGRATE_PFN_WRITE
: 0;
2257 /* FIXME support THP */
2258 if (!page
|| !page
->mapping
|| PageTransCompound(page
)) {
2264 * By getting a reference on the page we pin it and that blocks
2265 * any kind of migration. Side effect is that it "freezes" the
2268 * We drop this reference after isolating the page from the lru
2269 * for non device page (device page are not on the lru and thus
2270 * can't be dropped from it).
2276 * Optimize for the common case where page is only mapped once
2277 * in one process. If we can lock the page, then we can safely
2278 * set up a special migration page table entry now.
2280 if (trylock_page(page
)) {
2283 mpfn
|= MIGRATE_PFN_LOCKED
;
2284 ptep_get_and_clear(mm
, addr
, ptep
);
2286 /* Setup special migration page table entry */
2287 entry
= make_migration_entry(page
, mpfn
&
2289 swp_pte
= swp_entry_to_pte(entry
);
2290 if (pte_soft_dirty(pte
))
2291 swp_pte
= pte_swp_mksoft_dirty(swp_pte
);
2292 set_pte_at(mm
, addr
, ptep
, swp_pte
);
2295 * This is like regular unmap: we remove the rmap and
2296 * drop page refcount. Page won't be freed, as we took
2297 * a reference just above.
2299 page_remove_rmap(page
, false);
2302 if (pte_present(pte
))
2307 migrate
->dst
[migrate
->npages
] = 0;
2308 migrate
->src
[migrate
->npages
++] = mpfn
;
2310 arch_leave_lazy_mmu_mode();
2311 pte_unmap_unlock(ptep
- 1, ptl
);
2313 /* Only flush the TLB if we actually modified any entries */
2315 flush_tlb_range(walk
->vma
, start
, end
);
2320 static const struct mm_walk_ops migrate_vma_walk_ops
= {
2321 .pmd_entry
= migrate_vma_collect_pmd
,
2322 .pte_hole
= migrate_vma_collect_hole
,
2326 * migrate_vma_collect() - collect pages over a range of virtual addresses
2327 * @migrate: migrate struct containing all migration information
2329 * This will walk the CPU page table. For each virtual address backed by a
2330 * valid page, it updates the src array and takes a reference on the page, in
2331 * order to pin the page until we lock it and unmap it.
2333 static void migrate_vma_collect(struct migrate_vma
*migrate
)
2335 struct mmu_notifier_range range
;
2337 mmu_notifier_range_init(&range
, MMU_NOTIFY_CLEAR
, 0, NULL
,
2338 migrate
->vma
->vm_mm
, migrate
->start
, migrate
->end
);
2339 mmu_notifier_invalidate_range_start(&range
);
2341 walk_page_range(migrate
->vma
->vm_mm
, migrate
->start
, migrate
->end
,
2342 &migrate_vma_walk_ops
, migrate
);
2344 mmu_notifier_invalidate_range_end(&range
);
2345 migrate
->end
= migrate
->start
+ (migrate
->npages
<< PAGE_SHIFT
);
2349 * migrate_vma_check_page() - check if page is pinned or not
2350 * @page: struct page to check
2352 * Pinned pages cannot be migrated. This is the same test as in
2353 * migrate_page_move_mapping(), except that here we allow migration of a
2356 static bool migrate_vma_check_page(struct page
*page
)
2359 * One extra ref because caller holds an extra reference, either from
2360 * isolate_lru_page() for a regular page, or migrate_vma_collect() for
2366 * FIXME support THP (transparent huge page), it is bit more complex to
2367 * check them than regular pages, because they can be mapped with a pmd
2368 * or with a pte (split pte mapping).
2370 if (PageCompound(page
))
2373 /* Page from ZONE_DEVICE have one extra reference */
2374 if (is_zone_device_page(page
)) {
2376 * Private page can never be pin as they have no valid pte and
2377 * GUP will fail for those. Yet if there is a pending migration
2378 * a thread might try to wait on the pte migration entry and
2379 * will bump the page reference count. Sadly there is no way to
2380 * differentiate a regular pin from migration wait. Hence to
2381 * avoid 2 racing thread trying to migrate back to CPU to enter
2382 * infinite loop (one stoping migration because the other is
2383 * waiting on pte migration entry). We always return true here.
2385 * FIXME proper solution is to rework migration_entry_wait() so
2386 * it does not need to take a reference on page.
2388 return is_device_private_page(page
);
2391 /* For file back page */
2392 if (page_mapping(page
))
2393 extra
+= 1 + page_has_private(page
);
2395 if ((page_count(page
) - extra
) > page_mapcount(page
))
2402 * migrate_vma_prepare() - lock pages and isolate them from the lru
2403 * @migrate: migrate struct containing all migration information
2405 * This locks pages that have been collected by migrate_vma_collect(). Once each
2406 * page is locked it is isolated from the lru (for non-device pages). Finally,
2407 * the ref taken by migrate_vma_collect() is dropped, as locked pages cannot be
2408 * migrated by concurrent kernel threads.
2410 static void migrate_vma_prepare(struct migrate_vma
*migrate
)
2412 const unsigned long npages
= migrate
->npages
;
2413 const unsigned long start
= migrate
->start
;
2414 unsigned long addr
, i
, restore
= 0;
2415 bool allow_drain
= true;
2419 for (i
= 0; (i
< npages
) && migrate
->cpages
; i
++) {
2420 struct page
*page
= migrate_pfn_to_page(migrate
->src
[i
]);
2426 if (!(migrate
->src
[i
] & MIGRATE_PFN_LOCKED
)) {
2428 * Because we are migrating several pages there can be
2429 * a deadlock between 2 concurrent migration where each
2430 * are waiting on each other page lock.
2432 * Make migrate_vma() a best effort thing and backoff
2433 * for any page we can not lock right away.
2435 if (!trylock_page(page
)) {
2436 migrate
->src
[i
] = 0;
2442 migrate
->src
[i
] |= MIGRATE_PFN_LOCKED
;
2445 /* ZONE_DEVICE pages are not on LRU */
2446 if (!is_zone_device_page(page
)) {
2447 if (!PageLRU(page
) && allow_drain
) {
2448 /* Drain CPU's pagevec */
2449 lru_add_drain_all();
2450 allow_drain
= false;
2453 if (isolate_lru_page(page
)) {
2455 migrate
->src
[i
] &= ~MIGRATE_PFN_MIGRATE
;
2459 migrate
->src
[i
] = 0;
2467 /* Drop the reference we took in collect */
2471 if (!migrate_vma_check_page(page
)) {
2473 migrate
->src
[i
] &= ~MIGRATE_PFN_MIGRATE
;
2477 if (!is_zone_device_page(page
)) {
2479 putback_lru_page(page
);
2482 migrate
->src
[i
] = 0;
2486 if (!is_zone_device_page(page
))
2487 putback_lru_page(page
);
2494 for (i
= 0, addr
= start
; i
< npages
&& restore
; i
++, addr
+= PAGE_SIZE
) {
2495 struct page
*page
= migrate_pfn_to_page(migrate
->src
[i
]);
2497 if (!page
|| (migrate
->src
[i
] & MIGRATE_PFN_MIGRATE
))
2500 remove_migration_pte(page
, migrate
->vma
, addr
, page
);
2502 migrate
->src
[i
] = 0;
2510 * migrate_vma_unmap() - replace page mapping with special migration pte entry
2511 * @migrate: migrate struct containing all migration information
2513 * Replace page mapping (CPU page table pte) with a special migration pte entry
2514 * and check again if it has been pinned. Pinned pages are restored because we
2515 * cannot migrate them.
2517 * This is the last step before we call the device driver callback to allocate
2518 * destination memory and copy contents of original page over to new page.
2520 static void migrate_vma_unmap(struct migrate_vma
*migrate
)
2522 int flags
= TTU_MIGRATION
| TTU_IGNORE_MLOCK
| TTU_IGNORE_ACCESS
;
2523 const unsigned long npages
= migrate
->npages
;
2524 const unsigned long start
= migrate
->start
;
2525 unsigned long addr
, i
, restore
= 0;
2527 for (i
= 0; i
< npages
; i
++) {
2528 struct page
*page
= migrate_pfn_to_page(migrate
->src
[i
]);
2530 if (!page
|| !(migrate
->src
[i
] & MIGRATE_PFN_MIGRATE
))
2533 if (page_mapped(page
)) {
2534 try_to_unmap(page
, flags
);
2535 if (page_mapped(page
))
2539 if (migrate_vma_check_page(page
))
2543 migrate
->src
[i
] &= ~MIGRATE_PFN_MIGRATE
;
2548 for (addr
= start
, i
= 0; i
< npages
&& restore
; addr
+= PAGE_SIZE
, i
++) {
2549 struct page
*page
= migrate_pfn_to_page(migrate
->src
[i
]);
2551 if (!page
|| (migrate
->src
[i
] & MIGRATE_PFN_MIGRATE
))
2554 remove_migration_ptes(page
, page
, false);
2556 migrate
->src
[i
] = 0;
2560 if (is_zone_device_page(page
))
2563 putback_lru_page(page
);
2568 * migrate_vma_setup() - prepare to migrate a range of memory
2569 * @args: contains the vma, start, and and pfns arrays for the migration
2571 * Returns: negative errno on failures, 0 when 0 or more pages were migrated
2574 * Prepare to migrate a range of memory virtual address range by collecting all
2575 * the pages backing each virtual address in the range, saving them inside the
2576 * src array. Then lock those pages and unmap them. Once the pages are locked
2577 * and unmapped, check whether each page is pinned or not. Pages that aren't
2578 * pinned have the MIGRATE_PFN_MIGRATE flag set (by this function) in the
2579 * corresponding src array entry. Then restores any pages that are pinned, by
2580 * remapping and unlocking those pages.
2582 * The caller should then allocate destination memory and copy source memory to
2583 * it for all those entries (ie with MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE
2584 * flag set). Once these are allocated and copied, the caller must update each
2585 * corresponding entry in the dst array with the pfn value of the destination
2586 * page and with the MIGRATE_PFN_VALID and MIGRATE_PFN_LOCKED flags set
2587 * (destination pages must have their struct pages locked, via lock_page()).
2589 * Note that the caller does not have to migrate all the pages that are marked
2590 * with MIGRATE_PFN_MIGRATE flag in src array unless this is a migration from
2591 * device memory to system memory. If the caller cannot migrate a device page
2592 * back to system memory, then it must return VM_FAULT_SIGBUS, which has severe
2593 * consequences for the userspace process, so it must be avoided if at all
2596 * For empty entries inside CPU page table (pte_none() or pmd_none() is true) we
2597 * do set MIGRATE_PFN_MIGRATE flag inside the corresponding source array thus
2598 * allowing the caller to allocate device memory for those unback virtual
2599 * address. For this the caller simply has to allocate device memory and
2600 * properly set the destination entry like for regular migration. Note that
2601 * this can still fails and thus inside the device driver must check if the
2602 * migration was successful for those entries after calling migrate_vma_pages()
2603 * just like for regular migration.
2605 * After that, the callers must call migrate_vma_pages() to go over each entry
2606 * in the src array that has the MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE flag
2607 * set. If the corresponding entry in dst array has MIGRATE_PFN_VALID flag set,
2608 * then migrate_vma_pages() to migrate struct page information from the source
2609 * struct page to the destination struct page. If it fails to migrate the
2610 * struct page information, then it clears the MIGRATE_PFN_MIGRATE flag in the
2613 * At this point all successfully migrated pages have an entry in the src
2614 * array with MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE flag set and the dst
2615 * array entry with MIGRATE_PFN_VALID flag set.
2617 * Once migrate_vma_pages() returns the caller may inspect which pages were
2618 * successfully migrated, and which were not. Successfully migrated pages will
2619 * have the MIGRATE_PFN_MIGRATE flag set for their src array entry.
2621 * It is safe to update device page table after migrate_vma_pages() because
2622 * both destination and source page are still locked, and the mmap_sem is held
2623 * in read mode (hence no one can unmap the range being migrated).
2625 * Once the caller is done cleaning up things and updating its page table (if it
2626 * chose to do so, this is not an obligation) it finally calls
2627 * migrate_vma_finalize() to update the CPU page table to point to new pages
2628 * for successfully migrated pages or otherwise restore the CPU page table to
2629 * point to the original source pages.
2631 int migrate_vma_setup(struct migrate_vma
*args
)
2633 long nr_pages
= (args
->end
- args
->start
) >> PAGE_SHIFT
;
2635 args
->start
&= PAGE_MASK
;
2636 args
->end
&= PAGE_MASK
;
2637 if (!args
->vma
|| is_vm_hugetlb_page(args
->vma
) ||
2638 (args
->vma
->vm_flags
& VM_SPECIAL
) || vma_is_dax(args
->vma
))
2642 if (args
->start
< args
->vma
->vm_start
||
2643 args
->start
>= args
->vma
->vm_end
)
2645 if (args
->end
<= args
->vma
->vm_start
|| args
->end
> args
->vma
->vm_end
)
2647 if (!args
->src
|| !args
->dst
)
2650 memset(args
->src
, 0, sizeof(*args
->src
) * nr_pages
);
2654 migrate_vma_collect(args
);
2657 migrate_vma_prepare(args
);
2659 migrate_vma_unmap(args
);
2662 * At this point pages are locked and unmapped, and thus they have
2663 * stable content and can safely be copied to destination memory that
2664 * is allocated by the drivers.
2669 EXPORT_SYMBOL(migrate_vma_setup
);
2671 static void migrate_vma_insert_page(struct migrate_vma
*migrate
,
2677 struct vm_area_struct
*vma
= migrate
->vma
;
2678 struct mm_struct
*mm
= vma
->vm_mm
;
2679 struct mem_cgroup
*memcg
;
2689 /* Only allow populating anonymous memory */
2690 if (!vma_is_anonymous(vma
))
2693 pgdp
= pgd_offset(mm
, addr
);
2694 p4dp
= p4d_alloc(mm
, pgdp
, addr
);
2697 pudp
= pud_alloc(mm
, p4dp
, addr
);
2700 pmdp
= pmd_alloc(mm
, pudp
, addr
);
2704 if (pmd_trans_huge(*pmdp
) || pmd_devmap(*pmdp
))
2708 * Use pte_alloc() instead of pte_alloc_map(). We can't run
2709 * pte_offset_map() on pmds where a huge pmd might be created
2710 * from a different thread.
2712 * pte_alloc_map() is safe to use under down_write(mmap_sem) or when
2713 * parallel threads are excluded by other means.
2715 * Here we only have down_read(mmap_sem).
2717 if (pte_alloc(mm
, pmdp
))
2720 /* See the comment in pte_alloc_one_map() */
2721 if (unlikely(pmd_trans_unstable(pmdp
)))
2724 if (unlikely(anon_vma_prepare(vma
)))
2726 if (mem_cgroup_try_charge(page
, vma
->vm_mm
, GFP_KERNEL
, &memcg
, false))
2730 * The memory barrier inside __SetPageUptodate makes sure that
2731 * preceding stores to the page contents become visible before
2732 * the set_pte_at() write.
2734 __SetPageUptodate(page
);
2736 if (is_zone_device_page(page
)) {
2737 if (is_device_private_page(page
)) {
2738 swp_entry_t swp_entry
;
2740 swp_entry
= make_device_private_entry(page
, vma
->vm_flags
& VM_WRITE
);
2741 entry
= swp_entry_to_pte(swp_entry
);
2744 entry
= mk_pte(page
, vma
->vm_page_prot
);
2745 if (vma
->vm_flags
& VM_WRITE
)
2746 entry
= pte_mkwrite(pte_mkdirty(entry
));
2749 ptep
= pte_offset_map_lock(mm
, pmdp
, addr
, &ptl
);
2751 if (pte_present(*ptep
)) {
2752 unsigned long pfn
= pte_pfn(*ptep
);
2754 if (!is_zero_pfn(pfn
)) {
2755 pte_unmap_unlock(ptep
, ptl
);
2756 mem_cgroup_cancel_charge(page
, memcg
, false);
2760 } else if (!pte_none(*ptep
)) {
2761 pte_unmap_unlock(ptep
, ptl
);
2762 mem_cgroup_cancel_charge(page
, memcg
, false);
2767 * Check for usefaultfd but do not deliver the fault. Instead,
2770 if (userfaultfd_missing(vma
)) {
2771 pte_unmap_unlock(ptep
, ptl
);
2772 mem_cgroup_cancel_charge(page
, memcg
, false);
2776 inc_mm_counter(mm
, MM_ANONPAGES
);
2777 page_add_new_anon_rmap(page
, vma
, addr
, false);
2778 mem_cgroup_commit_charge(page
, memcg
, false, false);
2779 if (!is_zone_device_page(page
))
2780 lru_cache_add_active_or_unevictable(page
, vma
);
2784 flush_cache_page(vma
, addr
, pte_pfn(*ptep
));
2785 ptep_clear_flush_notify(vma
, addr
, ptep
);
2786 set_pte_at_notify(mm
, addr
, ptep
, entry
);
2787 update_mmu_cache(vma
, addr
, ptep
);
2789 /* No need to invalidate - it was non-present before */
2790 set_pte_at(mm
, addr
, ptep
, entry
);
2791 update_mmu_cache(vma
, addr
, ptep
);
2794 pte_unmap_unlock(ptep
, ptl
);
2795 *src
= MIGRATE_PFN_MIGRATE
;
2799 *src
&= ~MIGRATE_PFN_MIGRATE
;
2803 * migrate_vma_pages() - migrate meta-data from src page to dst page
2804 * @migrate: migrate struct containing all migration information
2806 * This migrates struct page meta-data from source struct page to destination
2807 * struct page. This effectively finishes the migration from source page to the
2810 void migrate_vma_pages(struct migrate_vma
*migrate
)
2812 const unsigned long npages
= migrate
->npages
;
2813 const unsigned long start
= migrate
->start
;
2814 struct mmu_notifier_range range
;
2815 unsigned long addr
, i
;
2816 bool notified
= false;
2818 for (i
= 0, addr
= start
; i
< npages
; addr
+= PAGE_SIZE
, i
++) {
2819 struct page
*newpage
= migrate_pfn_to_page(migrate
->dst
[i
]);
2820 struct page
*page
= migrate_pfn_to_page(migrate
->src
[i
]);
2821 struct address_space
*mapping
;
2825 migrate
->src
[i
] &= ~MIGRATE_PFN_MIGRATE
;
2830 if (!(migrate
->src
[i
] & MIGRATE_PFN_MIGRATE
)) {
2836 mmu_notifier_range_init(&range
,
2837 MMU_NOTIFY_CLEAR
, 0,
2839 migrate
->vma
->vm_mm
,
2840 addr
, migrate
->end
);
2841 mmu_notifier_invalidate_range_start(&range
);
2843 migrate_vma_insert_page(migrate
, addr
, newpage
,
2849 mapping
= page_mapping(page
);
2851 if (is_zone_device_page(newpage
)) {
2852 if (is_device_private_page(newpage
)) {
2854 * For now only support private anonymous when
2855 * migrating to un-addressable device memory.
2858 migrate
->src
[i
] &= ~MIGRATE_PFN_MIGRATE
;
2863 * Other types of ZONE_DEVICE page are not
2866 migrate
->src
[i
] &= ~MIGRATE_PFN_MIGRATE
;
2871 r
= migrate_page(mapping
, newpage
, page
, MIGRATE_SYNC_NO_COPY
);
2872 if (r
!= MIGRATEPAGE_SUCCESS
)
2873 migrate
->src
[i
] &= ~MIGRATE_PFN_MIGRATE
;
2877 * No need to double call mmu_notifier->invalidate_range() callback as
2878 * the above ptep_clear_flush_notify() inside migrate_vma_insert_page()
2879 * did already call it.
2882 mmu_notifier_invalidate_range_only_end(&range
);
2884 EXPORT_SYMBOL(migrate_vma_pages
);
2887 * migrate_vma_finalize() - restore CPU page table entry
2888 * @migrate: migrate struct containing all migration information
2890 * This replaces the special migration pte entry with either a mapping to the
2891 * new page if migration was successful for that page, or to the original page
2894 * This also unlocks the pages and puts them back on the lru, or drops the extra
2895 * refcount, for device pages.
2897 void migrate_vma_finalize(struct migrate_vma
*migrate
)
2899 const unsigned long npages
= migrate
->npages
;
2902 for (i
= 0; i
< npages
; i
++) {
2903 struct page
*newpage
= migrate_pfn_to_page(migrate
->dst
[i
]);
2904 struct page
*page
= migrate_pfn_to_page(migrate
->src
[i
]);
2908 unlock_page(newpage
);
2914 if (!(migrate
->src
[i
] & MIGRATE_PFN_MIGRATE
) || !newpage
) {
2916 unlock_page(newpage
);
2922 remove_migration_ptes(page
, newpage
, false);
2926 if (is_zone_device_page(page
))
2929 putback_lru_page(page
);
2931 if (newpage
!= page
) {
2932 unlock_page(newpage
);
2933 if (is_zone_device_page(newpage
))
2936 putback_lru_page(newpage
);
2940 EXPORT_SYMBOL(migrate_vma_finalize
);
2941 #endif /* CONFIG_DEVICE_PRIVATE */