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>
51 #include <linux/oom.h>
53 #include <asm/tlbflush.h>
55 #define CREATE_TRACE_POINTS
56 #include <trace/events/migrate.h>
61 * migrate_prep() needs to be called before we start compiling a list of pages
62 * to be migrated using isolate_lru_page(). If scheduling work on other CPUs is
63 * undesirable, use migrate_prep_local()
65 int migrate_prep(void)
68 * Clear the LRU lists so pages can be isolated.
69 * Note that pages may be moved off the LRU after we have
70 * drained them. Those pages will fail to migrate like other
71 * pages that may be busy.
78 /* Do the necessary work of migrate_prep but not if it involves other CPUs */
79 int migrate_prep_local(void)
86 int isolate_movable_page(struct page
*page
, isolate_mode_t mode
)
88 struct address_space
*mapping
;
91 * Avoid burning cycles with pages that are yet under __free_pages(),
92 * or just got freed under us.
94 * In case we 'win' a race for a movable page being freed under us and
95 * raise its refcount preventing __free_pages() from doing its job
96 * the put_page() at the end of this block will take care of
97 * release this page, thus avoiding a nasty leakage.
99 if (unlikely(!get_page_unless_zero(page
)))
103 * Check PageMovable before holding a PG_lock because page's owner
104 * assumes anybody doesn't touch PG_lock of newly allocated page
105 * so unconditionally grabbing the lock ruins page's owner side.
107 if (unlikely(!__PageMovable(page
)))
110 * As movable pages are not isolated from LRU lists, concurrent
111 * compaction threads can race against page migration functions
112 * as well as race against the releasing a page.
114 * In order to avoid having an already isolated movable page
115 * being (wrongly) re-isolated while it is under migration,
116 * or to avoid attempting to isolate pages being released,
117 * lets be sure we have the page lock
118 * before proceeding with the movable page isolation steps.
120 if (unlikely(!trylock_page(page
)))
123 if (!PageMovable(page
) || PageIsolated(page
))
124 goto out_no_isolated
;
126 mapping
= page_mapping(page
);
127 VM_BUG_ON_PAGE(!mapping
, page
);
129 if (!mapping
->a_ops
->isolate_page(page
, mode
))
130 goto out_no_isolated
;
132 /* Driver shouldn't use PG_isolated bit of page->flags */
133 WARN_ON_ONCE(PageIsolated(page
));
134 __SetPageIsolated(page
);
147 /* It should be called on page which is PG_movable */
148 void putback_movable_page(struct page
*page
)
150 struct address_space
*mapping
;
152 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
153 VM_BUG_ON_PAGE(!PageMovable(page
), page
);
154 VM_BUG_ON_PAGE(!PageIsolated(page
), page
);
156 mapping
= page_mapping(page
);
157 mapping
->a_ops
->putback_page(page
);
158 __ClearPageIsolated(page
);
162 * Put previously isolated pages back onto the appropriate lists
163 * from where they were once taken off for compaction/migration.
165 * This function shall be used whenever the isolated pageset has been
166 * built from lru, balloon, hugetlbfs page. See isolate_migratepages_range()
167 * and isolate_huge_page().
169 void putback_movable_pages(struct list_head
*l
)
174 list_for_each_entry_safe(page
, page2
, l
, lru
) {
175 if (unlikely(PageHuge(page
))) {
176 putback_active_hugepage(page
);
179 list_del(&page
->lru
);
181 * We isolated non-lru movable page so here we can use
182 * __PageMovable because LRU page's mapping cannot have
183 * PAGE_MAPPING_MOVABLE.
185 if (unlikely(__PageMovable(page
))) {
186 VM_BUG_ON_PAGE(!PageIsolated(page
), page
);
188 if (PageMovable(page
))
189 putback_movable_page(page
);
191 __ClearPageIsolated(page
);
195 mod_node_page_state(page_pgdat(page
), NR_ISOLATED_ANON
+
196 page_is_file_cache(page
), -hpage_nr_pages(page
));
197 putback_lru_page(page
);
203 * Restore a potential migration pte to a working pte entry
205 static bool remove_migration_pte(struct page
*page
, struct vm_area_struct
*vma
,
206 unsigned long addr
, void *old
)
208 struct page_vma_mapped_walk pvmw
= {
212 .flags
= PVMW_SYNC
| PVMW_MIGRATION
,
218 VM_BUG_ON_PAGE(PageTail(page
), page
);
219 while (page_vma_mapped_walk(&pvmw
)) {
223 new = page
- pvmw
.page
->index
+
224 linear_page_index(vma
, pvmw
.address
);
226 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
227 /* PMD-mapped THP migration entry */
229 VM_BUG_ON_PAGE(PageHuge(page
) || !PageTransCompound(page
), page
);
230 remove_migration_pmd(&pvmw
, new);
236 pte
= pte_mkold(mk_pte(new, READ_ONCE(vma
->vm_page_prot
)));
237 if (pte_swp_soft_dirty(*pvmw
.pte
))
238 pte
= pte_mksoft_dirty(pte
);
241 * Recheck VMA as permissions can change since migration started
243 entry
= pte_to_swp_entry(*pvmw
.pte
);
244 if (is_write_migration_entry(entry
))
245 pte
= maybe_mkwrite(pte
, vma
);
247 if (unlikely(is_zone_device_page(new))) {
248 if (is_device_private_page(new)) {
249 entry
= make_device_private_entry(new, pte_write(pte
));
250 pte
= swp_entry_to_pte(entry
);
254 #ifdef CONFIG_HUGETLB_PAGE
256 pte
= pte_mkhuge(pte
);
257 pte
= arch_make_huge_pte(pte
, vma
, new, 0);
258 set_huge_pte_at(vma
->vm_mm
, pvmw
.address
, pvmw
.pte
, pte
);
260 hugepage_add_anon_rmap(new, vma
, pvmw
.address
);
262 page_dup_rmap(new, true);
266 set_pte_at(vma
->vm_mm
, pvmw
.address
, pvmw
.pte
, pte
);
269 page_add_anon_rmap(new, vma
, pvmw
.address
, false);
271 page_add_file_rmap(new, false);
273 if (vma
->vm_flags
& VM_LOCKED
&& !PageTransCompound(new))
276 if (PageTransHuge(page
) && PageMlocked(page
))
277 clear_page_mlock(page
);
279 /* No need to invalidate - it was non-present before */
280 update_mmu_cache(vma
, pvmw
.address
, pvmw
.pte
);
287 * Get rid of all migration entries and replace them by
288 * references to the indicated page.
290 void remove_migration_ptes(struct page
*old
, struct page
*new, bool locked
)
292 struct rmap_walk_control rwc
= {
293 .rmap_one
= remove_migration_pte
,
298 rmap_walk_locked(new, &rwc
);
300 rmap_walk(new, &rwc
);
304 * Something used the pte of a page under migration. We need to
305 * get to the page and wait until migration is finished.
306 * When we return from this function the fault will be retried.
308 void __migration_entry_wait(struct mm_struct
*mm
, pte_t
*ptep
,
317 if (!is_swap_pte(pte
))
320 entry
= pte_to_swp_entry(pte
);
321 if (!is_migration_entry(entry
))
324 page
= migration_entry_to_page(entry
);
327 * Once page cache replacement of page migration started, page_count
328 * is zero; but we must not call put_and_wait_on_page_locked() without
329 * a ref. Use get_page_unless_zero(), and just fault again if it fails.
331 if (!get_page_unless_zero(page
))
333 pte_unmap_unlock(ptep
, ptl
);
334 put_and_wait_on_page_locked(page
);
337 pte_unmap_unlock(ptep
, ptl
);
340 void migration_entry_wait(struct mm_struct
*mm
, pmd_t
*pmd
,
341 unsigned long address
)
343 spinlock_t
*ptl
= pte_lockptr(mm
, pmd
);
344 pte_t
*ptep
= pte_offset_map(pmd
, address
);
345 __migration_entry_wait(mm
, ptep
, ptl
);
348 void migration_entry_wait_huge(struct vm_area_struct
*vma
,
349 struct mm_struct
*mm
, pte_t
*pte
)
351 spinlock_t
*ptl
= huge_pte_lockptr(hstate_vma(vma
), mm
, pte
);
352 __migration_entry_wait(mm
, pte
, ptl
);
355 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
356 void pmd_migration_entry_wait(struct mm_struct
*mm
, pmd_t
*pmd
)
361 ptl
= pmd_lock(mm
, pmd
);
362 if (!is_pmd_migration_entry(*pmd
))
364 page
= migration_entry_to_page(pmd_to_swp_entry(*pmd
));
365 if (!get_page_unless_zero(page
))
368 put_and_wait_on_page_locked(page
);
375 static int expected_page_refs(struct address_space
*mapping
, struct page
*page
)
377 int expected_count
= 1;
380 * Device public or private pages have an extra refcount as they are
383 expected_count
+= is_device_private_page(page
);
385 expected_count
+= hpage_nr_pages(page
) + page_has_private(page
);
387 return expected_count
;
391 * Replace the page in the mapping.
393 * The number of remaining references must be:
394 * 1 for anonymous pages without a mapping
395 * 2 for pages with a mapping
396 * 3 for pages with a mapping and PagePrivate/PagePrivate2 set.
398 int migrate_page_move_mapping(struct address_space
*mapping
,
399 struct page
*newpage
, struct page
*page
, int extra_count
)
401 XA_STATE(xas
, &mapping
->i_pages
, page_index(page
));
402 struct zone
*oldzone
, *newzone
;
404 int expected_count
= expected_page_refs(mapping
, page
) + extra_count
;
407 /* Anonymous page without mapping */
408 if (page_count(page
) != expected_count
)
411 /* No turning back from here */
412 newpage
->index
= page
->index
;
413 newpage
->mapping
= page
->mapping
;
414 if (PageSwapBacked(page
))
415 __SetPageSwapBacked(newpage
);
417 return MIGRATEPAGE_SUCCESS
;
420 oldzone
= page_zone(page
);
421 newzone
= page_zone(newpage
);
424 if (page_count(page
) != expected_count
|| xas_load(&xas
) != page
) {
425 xas_unlock_irq(&xas
);
429 if (!page_ref_freeze(page
, expected_count
)) {
430 xas_unlock_irq(&xas
);
435 * Now we know that no one else is looking at the page:
436 * no turning back from here.
438 newpage
->index
= page
->index
;
439 newpage
->mapping
= page
->mapping
;
440 page_ref_add(newpage
, hpage_nr_pages(page
)); /* add cache reference */
441 if (PageSwapBacked(page
)) {
442 __SetPageSwapBacked(newpage
);
443 if (PageSwapCache(page
)) {
444 SetPageSwapCache(newpage
);
445 set_page_private(newpage
, page_private(page
));
448 VM_BUG_ON_PAGE(PageSwapCache(page
), page
);
451 /* Move dirty while page refs frozen and newpage not yet exposed */
452 dirty
= PageDirty(page
);
454 ClearPageDirty(page
);
455 SetPageDirty(newpage
);
458 xas_store(&xas
, newpage
);
459 if (PageTransHuge(page
)) {
462 for (i
= 1; i
< HPAGE_PMD_NR
; i
++) {
464 xas_store(&xas
, newpage
);
469 * Drop cache reference from old page by unfreezing
470 * to one less reference.
471 * We know this isn't the last reference.
473 page_ref_unfreeze(page
, expected_count
- hpage_nr_pages(page
));
476 /* Leave irq disabled to prevent preemption while updating stats */
479 * If moved to a different zone then also account
480 * the page for that zone. Other VM counters will be
481 * taken care of when we establish references to the
482 * new page and drop references to the old page.
484 * Note that anonymous pages are accounted for
485 * via NR_FILE_PAGES and NR_ANON_MAPPED if they
486 * are mapped to swap space.
488 if (newzone
!= oldzone
) {
489 __dec_node_state(oldzone
->zone_pgdat
, NR_FILE_PAGES
);
490 __inc_node_state(newzone
->zone_pgdat
, NR_FILE_PAGES
);
491 if (PageSwapBacked(page
) && !PageSwapCache(page
)) {
492 __dec_node_state(oldzone
->zone_pgdat
, NR_SHMEM
);
493 __inc_node_state(newzone
->zone_pgdat
, NR_SHMEM
);
495 if (dirty
&& mapping_cap_account_dirty(mapping
)) {
496 __dec_node_state(oldzone
->zone_pgdat
, NR_FILE_DIRTY
);
497 __dec_zone_state(oldzone
, NR_ZONE_WRITE_PENDING
);
498 __inc_node_state(newzone
->zone_pgdat
, NR_FILE_DIRTY
);
499 __inc_zone_state(newzone
, NR_ZONE_WRITE_PENDING
);
504 return MIGRATEPAGE_SUCCESS
;
506 EXPORT_SYMBOL(migrate_page_move_mapping
);
509 * The expected number of remaining references is the same as that
510 * of migrate_page_move_mapping().
512 int migrate_huge_page_move_mapping(struct address_space
*mapping
,
513 struct page
*newpage
, struct page
*page
)
515 XA_STATE(xas
, &mapping
->i_pages
, page_index(page
));
519 expected_count
= 2 + page_has_private(page
);
520 if (page_count(page
) != expected_count
|| xas_load(&xas
) != page
) {
521 xas_unlock_irq(&xas
);
525 if (!page_ref_freeze(page
, expected_count
)) {
526 xas_unlock_irq(&xas
);
530 newpage
->index
= page
->index
;
531 newpage
->mapping
= page
->mapping
;
535 xas_store(&xas
, newpage
);
537 page_ref_unfreeze(page
, expected_count
- 1);
539 xas_unlock_irq(&xas
);
541 return MIGRATEPAGE_SUCCESS
;
545 * Gigantic pages are so large that we do not guarantee that page++ pointer
546 * arithmetic will work across the entire page. We need something more
549 static void __copy_gigantic_page(struct page
*dst
, struct page
*src
,
553 struct page
*dst_base
= dst
;
554 struct page
*src_base
= src
;
556 for (i
= 0; i
< nr_pages
; ) {
558 copy_highpage(dst
, src
);
561 dst
= mem_map_next(dst
, dst_base
, i
);
562 src
= mem_map_next(src
, src_base
, i
);
566 static void copy_huge_page(struct page
*dst
, struct page
*src
)
573 struct hstate
*h
= page_hstate(src
);
574 nr_pages
= pages_per_huge_page(h
);
576 if (unlikely(nr_pages
> MAX_ORDER_NR_PAGES
)) {
577 __copy_gigantic_page(dst
, src
, nr_pages
);
582 BUG_ON(!PageTransHuge(src
));
583 nr_pages
= hpage_nr_pages(src
);
586 for (i
= 0; i
< nr_pages
; i
++) {
588 copy_highpage(dst
+ i
, src
+ i
);
593 * Copy the page to its new location
595 void migrate_page_states(struct page
*newpage
, struct page
*page
)
600 SetPageError(newpage
);
601 if (PageReferenced(page
))
602 SetPageReferenced(newpage
);
603 if (PageUptodate(page
))
604 SetPageUptodate(newpage
);
605 if (TestClearPageActive(page
)) {
606 VM_BUG_ON_PAGE(PageUnevictable(page
), page
);
607 SetPageActive(newpage
);
608 } else if (TestClearPageUnevictable(page
))
609 SetPageUnevictable(newpage
);
610 if (PageWorkingset(page
))
611 SetPageWorkingset(newpage
);
612 if (PageChecked(page
))
613 SetPageChecked(newpage
);
614 if (PageMappedToDisk(page
))
615 SetPageMappedToDisk(newpage
);
617 /* Move dirty on pages not done by migrate_page_move_mapping() */
619 SetPageDirty(newpage
);
621 if (page_is_young(page
))
622 set_page_young(newpage
);
623 if (page_is_idle(page
))
624 set_page_idle(newpage
);
627 * Copy NUMA information to the new page, to prevent over-eager
628 * future migrations of this same page.
630 cpupid
= page_cpupid_xchg_last(page
, -1);
631 page_cpupid_xchg_last(newpage
, cpupid
);
633 ksm_migrate_page(newpage
, page
);
635 * Please do not reorder this without considering how mm/ksm.c's
636 * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache().
638 if (PageSwapCache(page
))
639 ClearPageSwapCache(page
);
640 ClearPagePrivate(page
);
641 set_page_private(page
, 0);
644 * If any waiters have accumulated on the new page then
647 if (PageWriteback(newpage
))
648 end_page_writeback(newpage
);
650 copy_page_owner(page
, newpage
);
652 mem_cgroup_migrate(page
, newpage
);
654 EXPORT_SYMBOL(migrate_page_states
);
656 void migrate_page_copy(struct page
*newpage
, struct page
*page
)
658 if (PageHuge(page
) || PageTransHuge(page
))
659 copy_huge_page(newpage
, page
);
661 copy_highpage(newpage
, page
);
663 migrate_page_states(newpage
, page
);
665 EXPORT_SYMBOL(migrate_page_copy
);
667 /************************************************************
668 * Migration functions
669 ***********************************************************/
672 * Common logic to directly migrate a single LRU page suitable for
673 * pages that do not use PagePrivate/PagePrivate2.
675 * Pages are locked upon entry and exit.
677 int migrate_page(struct address_space
*mapping
,
678 struct page
*newpage
, struct page
*page
,
679 enum migrate_mode mode
)
683 BUG_ON(PageWriteback(page
)); /* Writeback must be complete */
685 rc
= migrate_page_move_mapping(mapping
, newpage
, page
, 0);
687 if (rc
!= MIGRATEPAGE_SUCCESS
)
690 if (mode
!= MIGRATE_SYNC_NO_COPY
)
691 migrate_page_copy(newpage
, page
);
693 migrate_page_states(newpage
, page
);
694 return MIGRATEPAGE_SUCCESS
;
696 EXPORT_SYMBOL(migrate_page
);
699 /* Returns true if all buffers are successfully locked */
700 static bool buffer_migrate_lock_buffers(struct buffer_head
*head
,
701 enum migrate_mode mode
)
703 struct buffer_head
*bh
= head
;
705 /* Simple case, sync compaction */
706 if (mode
!= MIGRATE_ASYNC
) {
709 bh
= bh
->b_this_page
;
711 } while (bh
!= head
);
716 /* async case, we cannot block on lock_buffer so use trylock_buffer */
718 if (!trylock_buffer(bh
)) {
720 * We failed to lock the buffer and cannot stall in
721 * async migration. Release the taken locks
723 struct buffer_head
*failed_bh
= bh
;
725 while (bh
!= failed_bh
) {
727 bh
= bh
->b_this_page
;
732 bh
= bh
->b_this_page
;
733 } while (bh
!= head
);
737 static int __buffer_migrate_page(struct address_space
*mapping
,
738 struct page
*newpage
, struct page
*page
, enum migrate_mode mode
,
741 struct buffer_head
*bh
, *head
;
745 if (!page_has_buffers(page
))
746 return migrate_page(mapping
, newpage
, page
, mode
);
748 /* Check whether page does not have extra refs before we do more work */
749 expected_count
= expected_page_refs(mapping
, page
);
750 if (page_count(page
) != expected_count
)
753 head
= page_buffers(page
);
754 if (!buffer_migrate_lock_buffers(head
, mode
))
759 bool invalidated
= false;
763 spin_lock(&mapping
->private_lock
);
766 if (atomic_read(&bh
->b_count
)) {
770 bh
= bh
->b_this_page
;
771 } while (bh
!= head
);
777 spin_unlock(&mapping
->private_lock
);
778 invalidate_bh_lrus();
780 goto recheck_buffers
;
784 rc
= migrate_page_move_mapping(mapping
, newpage
, page
, 0);
785 if (rc
!= MIGRATEPAGE_SUCCESS
)
788 ClearPagePrivate(page
);
789 set_page_private(newpage
, page_private(page
));
790 set_page_private(page
, 0);
796 set_bh_page(bh
, newpage
, bh_offset(bh
));
797 bh
= bh
->b_this_page
;
799 } while (bh
!= head
);
801 SetPagePrivate(newpage
);
803 if (mode
!= MIGRATE_SYNC_NO_COPY
)
804 migrate_page_copy(newpage
, page
);
806 migrate_page_states(newpage
, page
);
808 rc
= MIGRATEPAGE_SUCCESS
;
811 spin_unlock(&mapping
->private_lock
);
815 bh
= bh
->b_this_page
;
817 } while (bh
!= head
);
823 * Migration function for pages with buffers. This function can only be used
824 * if the underlying filesystem guarantees that no other references to "page"
825 * exist. For example attached buffer heads are accessed only under page lock.
827 int buffer_migrate_page(struct address_space
*mapping
,
828 struct page
*newpage
, struct page
*page
, enum migrate_mode mode
)
830 return __buffer_migrate_page(mapping
, newpage
, page
, mode
, false);
832 EXPORT_SYMBOL(buffer_migrate_page
);
835 * Same as above except that this variant is more careful and checks that there
836 * are also no buffer head references. This function is the right one for
837 * mappings where buffer heads are directly looked up and referenced (such as
838 * block device mappings).
840 int buffer_migrate_page_norefs(struct address_space
*mapping
,
841 struct page
*newpage
, struct page
*page
, enum migrate_mode mode
)
843 return __buffer_migrate_page(mapping
, newpage
, page
, mode
, true);
848 * Writeback a page to clean the dirty state
850 static int writeout(struct address_space
*mapping
, struct page
*page
)
852 struct writeback_control wbc
= {
853 .sync_mode
= WB_SYNC_NONE
,
856 .range_end
= LLONG_MAX
,
861 if (!mapping
->a_ops
->writepage
)
862 /* No write method for the address space */
865 if (!clear_page_dirty_for_io(page
))
866 /* Someone else already triggered a write */
870 * A dirty page may imply that the underlying filesystem has
871 * the page on some queue. So the page must be clean for
872 * migration. Writeout may mean we loose the lock and the
873 * page state is no longer what we checked for earlier.
874 * At this point we know that the migration attempt cannot
877 remove_migration_ptes(page
, page
, false);
879 rc
= mapping
->a_ops
->writepage(page
, &wbc
);
881 if (rc
!= AOP_WRITEPAGE_ACTIVATE
)
882 /* unlocked. Relock */
885 return (rc
< 0) ? -EIO
: -EAGAIN
;
889 * Default handling if a filesystem does not provide a migration function.
891 static int fallback_migrate_page(struct address_space
*mapping
,
892 struct page
*newpage
, struct page
*page
, enum migrate_mode mode
)
894 if (PageDirty(page
)) {
895 /* Only writeback pages in full synchronous migration */
898 case MIGRATE_SYNC_NO_COPY
:
903 return writeout(mapping
, page
);
907 * Buffers may be managed in a filesystem specific way.
908 * We must have no buffers or drop them.
910 if (page_has_private(page
) &&
911 !try_to_release_page(page
, GFP_KERNEL
))
912 return mode
== MIGRATE_SYNC
? -EAGAIN
: -EBUSY
;
914 return migrate_page(mapping
, newpage
, page
, mode
);
918 * Move a page to a newly allocated page
919 * The page is locked and all ptes have been successfully removed.
921 * The new page will have replaced the old page if this function
926 * MIGRATEPAGE_SUCCESS - success
928 static int move_to_new_page(struct page
*newpage
, struct page
*page
,
929 enum migrate_mode mode
)
931 struct address_space
*mapping
;
933 bool is_lru
= !__PageMovable(page
);
935 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
936 VM_BUG_ON_PAGE(!PageLocked(newpage
), newpage
);
938 mapping
= page_mapping(page
);
940 if (likely(is_lru
)) {
942 rc
= migrate_page(mapping
, newpage
, page
, mode
);
943 else if (mapping
->a_ops
->migratepage
)
945 * Most pages have a mapping and most filesystems
946 * provide a migratepage callback. Anonymous pages
947 * are part of swap space which also has its own
948 * migratepage callback. This is the most common path
949 * for page migration.
951 rc
= mapping
->a_ops
->migratepage(mapping
, newpage
,
954 rc
= fallback_migrate_page(mapping
, newpage
,
958 * In case of non-lru page, it could be released after
959 * isolation step. In that case, we shouldn't try migration.
961 VM_BUG_ON_PAGE(!PageIsolated(page
), page
);
962 if (!PageMovable(page
)) {
963 rc
= MIGRATEPAGE_SUCCESS
;
964 __ClearPageIsolated(page
);
968 rc
= mapping
->a_ops
->migratepage(mapping
, newpage
,
970 WARN_ON_ONCE(rc
== MIGRATEPAGE_SUCCESS
&&
971 !PageIsolated(page
));
975 * When successful, old pagecache page->mapping must be cleared before
976 * page is freed; but stats require that PageAnon be left as PageAnon.
978 if (rc
== MIGRATEPAGE_SUCCESS
) {
979 if (__PageMovable(page
)) {
980 VM_BUG_ON_PAGE(!PageIsolated(page
), page
);
983 * We clear PG_movable under page_lock so any compactor
984 * cannot try to migrate this page.
986 __ClearPageIsolated(page
);
990 * Anonymous and movable page->mapping will be cleared by
991 * free_pages_prepare so don't reset it here for keeping
992 * the type to work PageAnon, for example.
994 if (!PageMappingFlags(page
))
995 page
->mapping
= NULL
;
997 if (likely(!is_zone_device_page(newpage
)))
998 flush_dcache_page(newpage
);
1005 static int __unmap_and_move(struct page
*page
, struct page
*newpage
,
1006 int force
, enum migrate_mode mode
)
1009 int page_was_mapped
= 0;
1010 struct anon_vma
*anon_vma
= NULL
;
1011 bool is_lru
= !__PageMovable(page
);
1013 if (!trylock_page(page
)) {
1014 if (!force
|| mode
== MIGRATE_ASYNC
)
1018 * It's not safe for direct compaction to call lock_page.
1019 * For example, during page readahead pages are added locked
1020 * to the LRU. Later, when the IO completes the pages are
1021 * marked uptodate and unlocked. However, the queueing
1022 * could be merging multiple pages for one bio (e.g.
1023 * mpage_readpages). If an allocation happens for the
1024 * second or third page, the process can end up locking
1025 * the same page twice and deadlocking. Rather than
1026 * trying to be clever about what pages can be locked,
1027 * avoid the use of lock_page for direct compaction
1030 if (current
->flags
& PF_MEMALLOC
)
1036 if (PageWriteback(page
)) {
1038 * Only in the case of a full synchronous migration is it
1039 * necessary to wait for PageWriteback. In the async case,
1040 * the retry loop is too short and in the sync-light case,
1041 * the overhead of stalling is too much
1045 case MIGRATE_SYNC_NO_COPY
:
1053 wait_on_page_writeback(page
);
1057 * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
1058 * we cannot notice that anon_vma is freed while we migrates a page.
1059 * This get_anon_vma() delays freeing anon_vma pointer until the end
1060 * of migration. File cache pages are no problem because of page_lock()
1061 * File Caches may use write_page() or lock_page() in migration, then,
1062 * just care Anon page here.
1064 * Only page_get_anon_vma() understands the subtleties of
1065 * getting a hold on an anon_vma from outside one of its mms.
1066 * But if we cannot get anon_vma, then we won't need it anyway,
1067 * because that implies that the anon page is no longer mapped
1068 * (and cannot be remapped so long as we hold the page lock).
1070 if (PageAnon(page
) && !PageKsm(page
))
1071 anon_vma
= page_get_anon_vma(page
);
1074 * Block others from accessing the new page when we get around to
1075 * establishing additional references. We are usually the only one
1076 * holding a reference to newpage at this point. We used to have a BUG
1077 * here if trylock_page(newpage) fails, but would like to allow for
1078 * cases where there might be a race with the previous use of newpage.
1079 * This is much like races on refcount of oldpage: just don't BUG().
1081 if (unlikely(!trylock_page(newpage
)))
1084 if (unlikely(!is_lru
)) {
1085 rc
= move_to_new_page(newpage
, page
, mode
);
1086 goto out_unlock_both
;
1090 * Corner case handling:
1091 * 1. When a new swap-cache page is read into, it is added to the LRU
1092 * and treated as swapcache but it has no rmap yet.
1093 * Calling try_to_unmap() against a page->mapping==NULL page will
1094 * trigger a BUG. So handle it here.
1095 * 2. An orphaned page (see truncate_complete_page) might have
1096 * fs-private metadata. The page can be picked up due to memory
1097 * offlining. Everywhere else except page reclaim, the page is
1098 * invisible to the vm, so the page can not be migrated. So try to
1099 * free the metadata, so the page can be freed.
1101 if (!page
->mapping
) {
1102 VM_BUG_ON_PAGE(PageAnon(page
), page
);
1103 if (page_has_private(page
)) {
1104 try_to_free_buffers(page
);
1105 goto out_unlock_both
;
1107 } else if (page_mapped(page
)) {
1108 /* Establish migration ptes */
1109 VM_BUG_ON_PAGE(PageAnon(page
) && !PageKsm(page
) && !anon_vma
,
1112 TTU_MIGRATION
|TTU_IGNORE_MLOCK
|TTU_IGNORE_ACCESS
);
1113 page_was_mapped
= 1;
1116 if (!page_mapped(page
))
1117 rc
= move_to_new_page(newpage
, page
, mode
);
1119 if (page_was_mapped
)
1120 remove_migration_ptes(page
,
1121 rc
== MIGRATEPAGE_SUCCESS
? newpage
: page
, false);
1124 unlock_page(newpage
);
1126 /* Drop an anon_vma reference if we took one */
1128 put_anon_vma(anon_vma
);
1132 * If migration is successful, decrease refcount of the newpage
1133 * which will not free the page because new page owner increased
1134 * refcounter. As well, if it is LRU page, add the page to LRU
1135 * list in here. Use the old state of the isolated source page to
1136 * determine if we migrated a LRU page. newpage was already unlocked
1137 * and possibly modified by its owner - don't rely on the page
1140 if (rc
== MIGRATEPAGE_SUCCESS
) {
1141 if (unlikely(!is_lru
))
1144 putback_lru_page(newpage
);
1151 * gcc 4.7 and 4.8 on arm get an ICEs when inlining unmap_and_move(). Work
1154 #if defined(CONFIG_ARM) && \
1155 defined(GCC_VERSION) && GCC_VERSION < 40900 && GCC_VERSION >= 40700
1156 #define ICE_noinline noinline
1158 #define ICE_noinline
1162 * Obtain the lock on page, remove all ptes and migrate the page
1163 * to the newly allocated page in newpage.
1165 static ICE_noinline
int unmap_and_move(new_page_t get_new_page
,
1166 free_page_t put_new_page
,
1167 unsigned long private, struct page
*page
,
1168 int force
, enum migrate_mode mode
,
1169 enum migrate_reason reason
)
1171 int rc
= MIGRATEPAGE_SUCCESS
;
1172 struct page
*newpage
= NULL
;
1174 if (!thp_migration_supported() && PageTransHuge(page
))
1177 if (page_count(page
) == 1) {
1178 /* page was freed from under us. So we are done. */
1179 ClearPageActive(page
);
1180 ClearPageUnevictable(page
);
1181 if (unlikely(__PageMovable(page
))) {
1183 if (!PageMovable(page
))
1184 __ClearPageIsolated(page
);
1190 newpage
= get_new_page(page
, private);
1194 rc
= __unmap_and_move(page
, newpage
, force
, mode
);
1195 if (rc
== MIGRATEPAGE_SUCCESS
)
1196 set_page_owner_migrate_reason(newpage
, reason
);
1199 if (rc
!= -EAGAIN
) {
1201 * A page that has been migrated has all references
1202 * removed and will be freed. A page that has not been
1203 * migrated will have kept its references and be restored.
1205 list_del(&page
->lru
);
1208 * Compaction can migrate also non-LRU pages which are
1209 * not accounted to NR_ISOLATED_*. They can be recognized
1212 if (likely(!__PageMovable(page
)))
1213 mod_node_page_state(page_pgdat(page
), NR_ISOLATED_ANON
+
1214 page_is_file_cache(page
), -hpage_nr_pages(page
));
1218 * If migration is successful, releases reference grabbed during
1219 * isolation. Otherwise, restore the page to right list unless
1222 if (rc
== MIGRATEPAGE_SUCCESS
) {
1224 if (reason
== MR_MEMORY_FAILURE
) {
1226 * Set PG_HWPoison on just freed page
1227 * intentionally. Although it's rather weird,
1228 * it's how HWPoison flag works at the moment.
1230 if (set_hwpoison_free_buddy_page(page
))
1231 num_poisoned_pages_inc();
1234 if (rc
!= -EAGAIN
) {
1235 if (likely(!__PageMovable(page
))) {
1236 putback_lru_page(page
);
1241 if (PageMovable(page
))
1242 putback_movable_page(page
);
1244 __ClearPageIsolated(page
);
1250 put_new_page(newpage
, private);
1259 * Counterpart of unmap_and_move_page() for hugepage migration.
1261 * This function doesn't wait the completion of hugepage I/O
1262 * because there is no race between I/O and migration for hugepage.
1263 * Note that currently hugepage I/O occurs only in direct I/O
1264 * where no lock is held and PG_writeback is irrelevant,
1265 * and writeback status of all subpages are counted in the reference
1266 * count of the head page (i.e. if all subpages of a 2MB hugepage are
1267 * under direct I/O, the reference of the head page is 512 and a bit more.)
1268 * This means that when we try to migrate hugepage whose subpages are
1269 * doing direct I/O, some references remain after try_to_unmap() and
1270 * hugepage migration fails without data corruption.
1272 * There is also no race when direct I/O is issued on the page under migration,
1273 * because then pte is replaced with migration swap entry and direct I/O code
1274 * will wait in the page fault for migration to complete.
1276 static int unmap_and_move_huge_page(new_page_t get_new_page
,
1277 free_page_t put_new_page
, unsigned long private,
1278 struct page
*hpage
, int force
,
1279 enum migrate_mode mode
, int reason
)
1282 int page_was_mapped
= 0;
1283 struct page
*new_hpage
;
1284 struct anon_vma
*anon_vma
= NULL
;
1285 struct address_space
*mapping
= NULL
;
1288 * Migratability of hugepages depends on architectures and their size.
1289 * This check is necessary because some callers of hugepage migration
1290 * like soft offline and memory hotremove don't walk through page
1291 * tables or check whether the hugepage is pmd-based or not before
1292 * kicking migration.
1294 if (!hugepage_migration_supported(page_hstate(hpage
))) {
1295 putback_active_hugepage(hpage
);
1299 new_hpage
= get_new_page(hpage
, private);
1303 if (!trylock_page(hpage
)) {
1308 case MIGRATE_SYNC_NO_COPY
:
1317 * Check for pages which are in the process of being freed. Without
1318 * page_mapping() set, hugetlbfs specific move page routine will not
1319 * be called and we could leak usage counts for subpools.
1321 if (page_private(hpage
) && !page_mapping(hpage
)) {
1326 if (PageAnon(hpage
))
1327 anon_vma
= page_get_anon_vma(hpage
);
1329 if (unlikely(!trylock_page(new_hpage
)))
1332 if (page_mapped(hpage
)) {
1334 * try_to_unmap could potentially call huge_pmd_unshare.
1335 * Because of this, take semaphore in write mode here and
1336 * set TTU_RMAP_LOCKED to let lower levels know we have
1339 mapping
= hugetlb_page_mapping_lock_write(hpage
);
1340 if (unlikely(!mapping
))
1341 goto unlock_put_anon
;
1344 TTU_MIGRATION
|TTU_IGNORE_MLOCK
|TTU_IGNORE_ACCESS
|
1346 page_was_mapped
= 1;
1348 * Leave mapping locked until after subsequent call to
1349 * remove_migration_ptes()
1353 if (!page_mapped(hpage
))
1354 rc
= move_to_new_page(new_hpage
, hpage
, mode
);
1356 if (page_was_mapped
) {
1357 remove_migration_ptes(hpage
,
1358 rc
== MIGRATEPAGE_SUCCESS
? new_hpage
: hpage
, true);
1359 i_mmap_unlock_write(mapping
);
1363 unlock_page(new_hpage
);
1367 put_anon_vma(anon_vma
);
1369 if (rc
== MIGRATEPAGE_SUCCESS
) {
1370 move_hugetlb_state(hpage
, new_hpage
, reason
);
1371 put_new_page
= NULL
;
1378 putback_active_hugepage(hpage
);
1381 * If migration was not successful and there's a freeing callback, use
1382 * it. Otherwise, put_page() will drop the reference grabbed during
1386 put_new_page(new_hpage
, private);
1388 putback_active_hugepage(new_hpage
);
1394 * migrate_pages - migrate the pages specified in a list, to the free pages
1395 * supplied as the target for the page migration
1397 * @from: The list of pages to be migrated.
1398 * @get_new_page: The function used to allocate free pages to be used
1399 * as the target of the page migration.
1400 * @put_new_page: The function used to free target pages if migration
1401 * fails, or NULL if no special handling is necessary.
1402 * @private: Private data to be passed on to get_new_page()
1403 * @mode: The migration mode that specifies the constraints for
1404 * page migration, if any.
1405 * @reason: The reason for page migration.
1407 * The function returns after 10 attempts or if no pages are movable any more
1408 * because the list has become empty or no retryable pages exist any more.
1409 * The caller should call putback_movable_pages() to return pages to the LRU
1410 * or free list only if ret != 0.
1412 * Returns the number of pages that were not migrated, or an error code.
1414 int migrate_pages(struct list_head
*from
, new_page_t get_new_page
,
1415 free_page_t put_new_page
, unsigned long private,
1416 enum migrate_mode mode
, int reason
)
1420 int nr_succeeded
= 0;
1424 int swapwrite
= current
->flags
& PF_SWAPWRITE
;
1428 current
->flags
|= PF_SWAPWRITE
;
1430 for(pass
= 0; pass
< 10 && retry
; pass
++) {
1433 list_for_each_entry_safe(page
, page2
, from
, lru
) {
1438 rc
= unmap_and_move_huge_page(get_new_page
,
1439 put_new_page
, private, page
,
1440 pass
> 2, mode
, reason
);
1442 rc
= unmap_and_move(get_new_page
, put_new_page
,
1443 private, page
, pass
> 2, mode
,
1449 * THP migration might be unsupported or the
1450 * allocation could've failed so we should
1451 * retry on the same page with the THP split
1454 * Head page is retried immediately and tail
1455 * pages are added to the tail of the list so
1456 * we encounter them after the rest of the list
1459 if (PageTransHuge(page
) && !PageHuge(page
)) {
1461 rc
= split_huge_page_to_list(page
, from
);
1464 list_safe_reset_next(page
, page2
, lru
);
1473 case MIGRATEPAGE_SUCCESS
:
1478 * Permanent failure (-EBUSY, -ENOSYS, etc.):
1479 * unlike -EAGAIN case, the failed page is
1480 * removed from migration page list and not
1481 * retried in the next outer loop.
1492 count_vm_events(PGMIGRATE_SUCCESS
, nr_succeeded
);
1494 count_vm_events(PGMIGRATE_FAIL
, nr_failed
);
1495 trace_mm_migrate_pages(nr_succeeded
, nr_failed
, mode
, reason
);
1498 current
->flags
&= ~PF_SWAPWRITE
;
1505 static int store_status(int __user
*status
, int start
, int value
, int nr
)
1508 if (put_user(value
, status
+ start
))
1516 static int do_move_pages_to_node(struct mm_struct
*mm
,
1517 struct list_head
*pagelist
, int node
)
1521 if (list_empty(pagelist
))
1524 err
= migrate_pages(pagelist
, alloc_new_node_page
, NULL
, node
,
1525 MIGRATE_SYNC
, MR_SYSCALL
);
1527 putback_movable_pages(pagelist
);
1532 * Resolves the given address to a struct page, isolates it from the LRU and
1533 * puts it to the given pagelist.
1535 * errno - if the page cannot be found/isolated
1536 * 0 - when it doesn't have to be migrated because it is already on the
1538 * 1 - when it has been queued
1540 static int add_page_for_migration(struct mm_struct
*mm
, unsigned long addr
,
1541 int node
, struct list_head
*pagelist
, bool migrate_all
)
1543 struct vm_area_struct
*vma
;
1545 unsigned int follflags
;
1548 down_read(&mm
->mmap_sem
);
1550 vma
= find_vma(mm
, addr
);
1551 if (!vma
|| addr
< vma
->vm_start
|| !vma_migratable(vma
))
1554 /* FOLL_DUMP to ignore special (like zero) pages */
1555 follflags
= FOLL_GET
| FOLL_DUMP
;
1556 page
= follow_page(vma
, addr
, follflags
);
1558 err
= PTR_ERR(page
);
1567 if (page_to_nid(page
) == node
)
1571 if (page_mapcount(page
) > 1 && !migrate_all
)
1574 if (PageHuge(page
)) {
1575 if (PageHead(page
)) {
1576 isolate_huge_page(page
, pagelist
);
1582 head
= compound_head(page
);
1583 err
= isolate_lru_page(head
);
1588 list_add_tail(&head
->lru
, pagelist
);
1589 mod_node_page_state(page_pgdat(head
),
1590 NR_ISOLATED_ANON
+ page_is_file_cache(head
),
1591 hpage_nr_pages(head
));
1595 * Either remove the duplicate refcount from
1596 * isolate_lru_page() or drop the page ref if it was
1601 up_read(&mm
->mmap_sem
);
1606 * Migrate an array of page address onto an array of nodes and fill
1607 * the corresponding array of status.
1609 static int do_pages_move(struct mm_struct
*mm
, nodemask_t task_nodes
,
1610 unsigned long nr_pages
,
1611 const void __user
* __user
*pages
,
1612 const int __user
*nodes
,
1613 int __user
*status
, int flags
)
1615 int current_node
= NUMA_NO_NODE
;
1616 LIST_HEAD(pagelist
);
1622 for (i
= start
= 0; i
< nr_pages
; i
++) {
1623 const void __user
*p
;
1628 if (get_user(p
, pages
+ i
))
1630 if (get_user(node
, nodes
+ i
))
1632 addr
= (unsigned long)untagged_addr(p
);
1635 if (node
< 0 || node
>= MAX_NUMNODES
)
1637 if (!node_state(node
, N_MEMORY
))
1641 if (!node_isset(node
, task_nodes
))
1644 if (current_node
== NUMA_NO_NODE
) {
1645 current_node
= node
;
1647 } else if (node
!= current_node
) {
1648 err
= do_move_pages_to_node(mm
, &pagelist
, current_node
);
1651 * Positive err means the number of failed
1652 * pages to migrate. Since we are going to
1653 * abort and return the number of non-migrated
1654 * pages, so need to incude the rest of the
1655 * nr_pages that have not been attempted as
1659 err
+= nr_pages
- i
- 1;
1662 err
= store_status(status
, start
, current_node
, i
- start
);
1666 current_node
= node
;
1670 * Errors in the page lookup or isolation are not fatal and we simply
1671 * report them via status
1673 err
= add_page_for_migration(mm
, addr
, current_node
,
1674 &pagelist
, flags
& MPOL_MF_MOVE_ALL
);
1677 /* The page is already on the target node */
1678 err
= store_status(status
, i
, current_node
, 1);
1682 } else if (err
> 0) {
1683 /* The page is successfully queued for migration */
1687 err
= store_status(status
, i
, err
, 1);
1691 err
= do_move_pages_to_node(mm
, &pagelist
, current_node
);
1694 err
+= nr_pages
- i
- 1;
1698 err
= store_status(status
, start
, current_node
, i
- start
);
1702 current_node
= NUMA_NO_NODE
;
1705 if (list_empty(&pagelist
))
1708 /* Make sure we do not overwrite the existing error */
1709 err1
= do_move_pages_to_node(mm
, &pagelist
, current_node
);
1711 * Don't have to report non-attempted pages here since:
1712 * - If the above loop is done gracefully all pages have been
1714 * - If the above loop is aborted it means a fatal error
1715 * happened, should return ret.
1718 err1
= store_status(status
, start
, current_node
, i
- start
);
1726 * Determine the nodes of an array of pages and store it in an array of status.
1728 static void do_pages_stat_array(struct mm_struct
*mm
, unsigned long nr_pages
,
1729 const void __user
**pages
, int *status
)
1733 down_read(&mm
->mmap_sem
);
1735 for (i
= 0; i
< nr_pages
; i
++) {
1736 unsigned long addr
= (unsigned long)(*pages
);
1737 struct vm_area_struct
*vma
;
1741 vma
= find_vma(mm
, addr
);
1742 if (!vma
|| addr
< vma
->vm_start
)
1745 /* FOLL_DUMP to ignore special (like zero) pages */
1746 page
= follow_page(vma
, addr
, FOLL_DUMP
);
1748 err
= PTR_ERR(page
);
1752 err
= page
? page_to_nid(page
) : -ENOENT
;
1760 up_read(&mm
->mmap_sem
);
1764 * Determine the nodes of a user array of pages and store it in
1765 * a user array of status.
1767 static int do_pages_stat(struct mm_struct
*mm
, unsigned long nr_pages
,
1768 const void __user
* __user
*pages
,
1771 #define DO_PAGES_STAT_CHUNK_NR 16
1772 const void __user
*chunk_pages
[DO_PAGES_STAT_CHUNK_NR
];
1773 int chunk_status
[DO_PAGES_STAT_CHUNK_NR
];
1776 unsigned long chunk_nr
;
1778 chunk_nr
= nr_pages
;
1779 if (chunk_nr
> DO_PAGES_STAT_CHUNK_NR
)
1780 chunk_nr
= DO_PAGES_STAT_CHUNK_NR
;
1782 if (copy_from_user(chunk_pages
, pages
, chunk_nr
* sizeof(*chunk_pages
)))
1785 do_pages_stat_array(mm
, chunk_nr
, chunk_pages
, chunk_status
);
1787 if (copy_to_user(status
, chunk_status
, chunk_nr
* sizeof(*status
)))
1792 nr_pages
-= chunk_nr
;
1794 return nr_pages
? -EFAULT
: 0;
1798 * Move a list of pages in the address space of the currently executing
1801 static int kernel_move_pages(pid_t pid
, unsigned long nr_pages
,
1802 const void __user
* __user
*pages
,
1803 const int __user
*nodes
,
1804 int __user
*status
, int flags
)
1806 struct task_struct
*task
;
1807 struct mm_struct
*mm
;
1809 nodemask_t task_nodes
;
1812 if (flags
& ~(MPOL_MF_MOVE
|MPOL_MF_MOVE_ALL
))
1815 if ((flags
& MPOL_MF_MOVE_ALL
) && !capable(CAP_SYS_NICE
))
1818 /* Find the mm_struct */
1820 task
= pid
? find_task_by_vpid(pid
) : current
;
1825 get_task_struct(task
);
1828 * Check if this process has the right to modify the specified
1829 * process. Use the regular "ptrace_may_access()" checks.
1831 if (!ptrace_may_access(task
, PTRACE_MODE_READ_REALCREDS
)) {
1838 err
= security_task_movememory(task
);
1842 task_nodes
= cpuset_mems_allowed(task
);
1843 mm
= get_task_mm(task
);
1844 put_task_struct(task
);
1850 err
= do_pages_move(mm
, task_nodes
, nr_pages
, pages
,
1851 nodes
, status
, flags
);
1853 err
= do_pages_stat(mm
, nr_pages
, pages
, status
);
1859 put_task_struct(task
);
1863 SYSCALL_DEFINE6(move_pages
, pid_t
, pid
, unsigned long, nr_pages
,
1864 const void __user
* __user
*, pages
,
1865 const int __user
*, nodes
,
1866 int __user
*, status
, int, flags
)
1868 return kernel_move_pages(pid
, nr_pages
, pages
, nodes
, status
, flags
);
1871 #ifdef CONFIG_COMPAT
1872 COMPAT_SYSCALL_DEFINE6(move_pages
, pid_t
, pid
, compat_ulong_t
, nr_pages
,
1873 compat_uptr_t __user
*, pages32
,
1874 const int __user
*, nodes
,
1875 int __user
*, status
,
1878 const void __user
* __user
*pages
;
1881 pages
= compat_alloc_user_space(nr_pages
* sizeof(void *));
1882 for (i
= 0; i
< nr_pages
; i
++) {
1885 if (get_user(p
, pages32
+ i
) ||
1886 put_user(compat_ptr(p
), pages
+ i
))
1889 return kernel_move_pages(pid
, nr_pages
, pages
, nodes
, status
, flags
);
1891 #endif /* CONFIG_COMPAT */
1893 #ifdef CONFIG_NUMA_BALANCING
1895 * Returns true if this is a safe migration target node for misplaced NUMA
1896 * pages. Currently it only checks the watermarks which crude
1898 static bool migrate_balanced_pgdat(struct pglist_data
*pgdat
,
1899 unsigned long nr_migrate_pages
)
1903 for (z
= pgdat
->nr_zones
- 1; z
>= 0; z
--) {
1904 struct zone
*zone
= pgdat
->node_zones
+ z
;
1906 if (!populated_zone(zone
))
1909 /* Avoid waking kswapd by allocating pages_to_migrate pages. */
1910 if (!zone_watermark_ok(zone
, 0,
1911 high_wmark_pages(zone
) +
1920 static struct page
*alloc_misplaced_dst_page(struct page
*page
,
1923 int nid
= (int) data
;
1924 struct page
*newpage
;
1926 newpage
= __alloc_pages_node(nid
,
1927 (GFP_HIGHUSER_MOVABLE
|
1928 __GFP_THISNODE
| __GFP_NOMEMALLOC
|
1929 __GFP_NORETRY
| __GFP_NOWARN
) &
1935 static int numamigrate_isolate_page(pg_data_t
*pgdat
, struct page
*page
)
1939 VM_BUG_ON_PAGE(compound_order(page
) && !PageTransHuge(page
), page
);
1941 /* Avoid migrating to a node that is nearly full */
1942 if (!migrate_balanced_pgdat(pgdat
, compound_nr(page
)))
1945 if (isolate_lru_page(page
))
1949 * migrate_misplaced_transhuge_page() skips page migration's usual
1950 * check on page_count(), so we must do it here, now that the page
1951 * has been isolated: a GUP pin, or any other pin, prevents migration.
1952 * The expected page count is 3: 1 for page's mapcount and 1 for the
1953 * caller's pin and 1 for the reference taken by isolate_lru_page().
1955 if (PageTransHuge(page
) && page_count(page
) != 3) {
1956 putback_lru_page(page
);
1960 page_lru
= page_is_file_cache(page
);
1961 mod_node_page_state(page_pgdat(page
), NR_ISOLATED_ANON
+ page_lru
,
1962 hpage_nr_pages(page
));
1965 * Isolating the page has taken another reference, so the
1966 * caller's reference can be safely dropped without the page
1967 * disappearing underneath us during migration.
1973 bool pmd_trans_migrating(pmd_t pmd
)
1975 struct page
*page
= pmd_page(pmd
);
1976 return PageLocked(page
);
1980 * Attempt to migrate a misplaced page to the specified destination
1981 * node. Caller is expected to have an elevated reference count on
1982 * the page that will be dropped by this function before returning.
1984 int migrate_misplaced_page(struct page
*page
, struct vm_area_struct
*vma
,
1987 pg_data_t
*pgdat
= NODE_DATA(node
);
1990 LIST_HEAD(migratepages
);
1993 * Don't migrate file pages that are mapped in multiple processes
1994 * with execute permissions as they are probably shared libraries.
1996 if (page_mapcount(page
) != 1 && page_is_file_cache(page
) &&
1997 (vma
->vm_flags
& VM_EXEC
))
2001 * Also do not migrate dirty pages as not all filesystems can move
2002 * dirty pages in MIGRATE_ASYNC mode which is a waste of cycles.
2004 if (page_is_file_cache(page
) && PageDirty(page
))
2007 isolated
= numamigrate_isolate_page(pgdat
, page
);
2011 list_add(&page
->lru
, &migratepages
);
2012 nr_remaining
= migrate_pages(&migratepages
, alloc_misplaced_dst_page
,
2013 NULL
, node
, MIGRATE_ASYNC
,
2016 if (!list_empty(&migratepages
)) {
2017 list_del(&page
->lru
);
2018 dec_node_page_state(page
, NR_ISOLATED_ANON
+
2019 page_is_file_cache(page
));
2020 putback_lru_page(page
);
2024 count_vm_numa_event(NUMA_PAGE_MIGRATE
);
2025 BUG_ON(!list_empty(&migratepages
));
2032 #endif /* CONFIG_NUMA_BALANCING */
2034 #if defined(CONFIG_NUMA_BALANCING) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
2036 * Migrates a THP to a given target node. page must be locked and is unlocked
2039 int migrate_misplaced_transhuge_page(struct mm_struct
*mm
,
2040 struct vm_area_struct
*vma
,
2041 pmd_t
*pmd
, pmd_t entry
,
2042 unsigned long address
,
2043 struct page
*page
, int node
)
2046 pg_data_t
*pgdat
= NODE_DATA(node
);
2048 struct page
*new_page
= NULL
;
2049 int page_lru
= page_is_file_cache(page
);
2050 unsigned long start
= address
& HPAGE_PMD_MASK
;
2052 new_page
= alloc_pages_node(node
,
2053 (GFP_TRANSHUGE_LIGHT
| __GFP_THISNODE
),
2057 prep_transhuge_page(new_page
);
2059 isolated
= numamigrate_isolate_page(pgdat
, page
);
2065 /* Prepare a page as a migration target */
2066 __SetPageLocked(new_page
);
2067 if (PageSwapBacked(page
))
2068 __SetPageSwapBacked(new_page
);
2070 /* anon mapping, we can simply copy page->mapping to the new page: */
2071 new_page
->mapping
= page
->mapping
;
2072 new_page
->index
= page
->index
;
2073 /* flush the cache before copying using the kernel virtual address */
2074 flush_cache_range(vma
, start
, start
+ HPAGE_PMD_SIZE
);
2075 migrate_page_copy(new_page
, page
);
2076 WARN_ON(PageLRU(new_page
));
2078 /* Recheck the target PMD */
2079 ptl
= pmd_lock(mm
, pmd
);
2080 if (unlikely(!pmd_same(*pmd
, entry
) || !page_ref_freeze(page
, 2))) {
2083 /* Reverse changes made by migrate_page_copy() */
2084 if (TestClearPageActive(new_page
))
2085 SetPageActive(page
);
2086 if (TestClearPageUnevictable(new_page
))
2087 SetPageUnevictable(page
);
2089 unlock_page(new_page
);
2090 put_page(new_page
); /* Free it */
2092 /* Retake the callers reference and putback on LRU */
2094 putback_lru_page(page
);
2095 mod_node_page_state(page_pgdat(page
),
2096 NR_ISOLATED_ANON
+ page_lru
, -HPAGE_PMD_NR
);
2101 entry
= mk_huge_pmd(new_page
, vma
->vm_page_prot
);
2102 entry
= maybe_pmd_mkwrite(pmd_mkdirty(entry
), vma
);
2105 * Overwrite the old entry under pagetable lock and establish
2106 * the new PTE. Any parallel GUP will either observe the old
2107 * page blocking on the page lock, block on the page table
2108 * lock or observe the new page. The SetPageUptodate on the
2109 * new page and page_add_new_anon_rmap guarantee the copy is
2110 * visible before the pagetable update.
2112 page_add_anon_rmap(new_page
, vma
, start
, true);
2114 * At this point the pmd is numa/protnone (i.e. non present) and the TLB
2115 * has already been flushed globally. So no TLB can be currently
2116 * caching this non present pmd mapping. There's no need to clear the
2117 * pmd before doing set_pmd_at(), nor to flush the TLB after
2118 * set_pmd_at(). Clearing the pmd here would introduce a race
2119 * condition against MADV_DONTNEED, because MADV_DONTNEED only holds the
2120 * mmap_sem for reading. If the pmd is set to NULL at any given time,
2121 * MADV_DONTNEED won't wait on the pmd lock and it'll skip clearing this
2124 set_pmd_at(mm
, start
, pmd
, entry
);
2125 update_mmu_cache_pmd(vma
, address
, &entry
);
2127 page_ref_unfreeze(page
, 2);
2128 mlock_migrate_page(new_page
, page
);
2129 page_remove_rmap(page
, true);
2130 set_page_owner_migrate_reason(new_page
, MR_NUMA_MISPLACED
);
2134 /* Take an "isolate" reference and put new page on the LRU. */
2136 putback_lru_page(new_page
);
2138 unlock_page(new_page
);
2140 put_page(page
); /* Drop the rmap reference */
2141 put_page(page
); /* Drop the LRU isolation reference */
2143 count_vm_events(PGMIGRATE_SUCCESS
, HPAGE_PMD_NR
);
2144 count_vm_numa_events(NUMA_PAGE_MIGRATE
, HPAGE_PMD_NR
);
2146 mod_node_page_state(page_pgdat(page
),
2147 NR_ISOLATED_ANON
+ page_lru
,
2152 count_vm_events(PGMIGRATE_FAIL
, HPAGE_PMD_NR
);
2153 ptl
= pmd_lock(mm
, pmd
);
2154 if (pmd_same(*pmd
, entry
)) {
2155 entry
= pmd_modify(entry
, vma
->vm_page_prot
);
2156 set_pmd_at(mm
, start
, pmd
, entry
);
2157 update_mmu_cache_pmd(vma
, address
, &entry
);
2166 #endif /* CONFIG_NUMA_BALANCING */
2168 #endif /* CONFIG_NUMA */
2170 #ifdef CONFIG_DEVICE_PRIVATE
2171 static int migrate_vma_collect_hole(unsigned long start
,
2173 __always_unused
int depth
,
2174 struct mm_walk
*walk
)
2176 struct migrate_vma
*migrate
= walk
->private;
2179 for (addr
= start
; addr
< end
; addr
+= PAGE_SIZE
) {
2180 migrate
->src
[migrate
->npages
] = MIGRATE_PFN_MIGRATE
;
2181 migrate
->dst
[migrate
->npages
] = 0;
2189 static int migrate_vma_collect_skip(unsigned long start
,
2191 struct mm_walk
*walk
)
2193 struct migrate_vma
*migrate
= walk
->private;
2196 for (addr
= start
; addr
< end
; addr
+= PAGE_SIZE
) {
2197 migrate
->dst
[migrate
->npages
] = 0;
2198 migrate
->src
[migrate
->npages
++] = 0;
2204 static int migrate_vma_collect_pmd(pmd_t
*pmdp
,
2205 unsigned long start
,
2207 struct mm_walk
*walk
)
2209 struct migrate_vma
*migrate
= walk
->private;
2210 struct vm_area_struct
*vma
= walk
->vma
;
2211 struct mm_struct
*mm
= vma
->vm_mm
;
2212 unsigned long addr
= start
, unmapped
= 0;
2217 if (pmd_none(*pmdp
))
2218 return migrate_vma_collect_hole(start
, end
, -1, walk
);
2220 if (pmd_trans_huge(*pmdp
)) {
2223 ptl
= pmd_lock(mm
, pmdp
);
2224 if (unlikely(!pmd_trans_huge(*pmdp
))) {
2229 page
= pmd_page(*pmdp
);
2230 if (is_huge_zero_page(page
)) {
2232 split_huge_pmd(vma
, pmdp
, addr
);
2233 if (pmd_trans_unstable(pmdp
))
2234 return migrate_vma_collect_skip(start
, end
,
2241 if (unlikely(!trylock_page(page
)))
2242 return migrate_vma_collect_skip(start
, end
,
2244 ret
= split_huge_page(page
);
2248 return migrate_vma_collect_skip(start
, end
,
2250 if (pmd_none(*pmdp
))
2251 return migrate_vma_collect_hole(start
, end
, -1,
2256 if (unlikely(pmd_bad(*pmdp
)))
2257 return migrate_vma_collect_skip(start
, end
, walk
);
2259 ptep
= pte_offset_map_lock(mm
, pmdp
, addr
, &ptl
);
2260 arch_enter_lazy_mmu_mode();
2262 for (; addr
< end
; addr
+= PAGE_SIZE
, ptep
++) {
2263 unsigned long mpfn
= 0, pfn
;
2270 if (pte_none(pte
)) {
2271 mpfn
= MIGRATE_PFN_MIGRATE
;
2276 if (!pte_present(pte
)) {
2278 * Only care about unaddressable device page special
2279 * page table entry. Other special swap entries are not
2280 * migratable, and we ignore regular swapped page.
2282 entry
= pte_to_swp_entry(pte
);
2283 if (!is_device_private_entry(entry
))
2286 page
= device_private_entry_to_page(entry
);
2287 if (page
->pgmap
->owner
!= migrate
->src_owner
)
2290 mpfn
= migrate_pfn(page_to_pfn(page
)) |
2291 MIGRATE_PFN_MIGRATE
;
2292 if (is_write_device_private_entry(entry
))
2293 mpfn
|= MIGRATE_PFN_WRITE
;
2295 if (migrate
->src_owner
)
2298 if (is_zero_pfn(pfn
)) {
2299 mpfn
= MIGRATE_PFN_MIGRATE
;
2303 page
= vm_normal_page(migrate
->vma
, addr
, pte
);
2304 mpfn
= migrate_pfn(pfn
) | MIGRATE_PFN_MIGRATE
;
2305 mpfn
|= pte_write(pte
) ? MIGRATE_PFN_WRITE
: 0;
2308 /* FIXME support THP */
2309 if (!page
|| !page
->mapping
|| PageTransCompound(page
)) {
2315 * By getting a reference on the page we pin it and that blocks
2316 * any kind of migration. Side effect is that it "freezes" the
2319 * We drop this reference after isolating the page from the lru
2320 * for non device page (device page are not on the lru and thus
2321 * can't be dropped from it).
2327 * Optimize for the common case where page is only mapped once
2328 * in one process. If we can lock the page, then we can safely
2329 * set up a special migration page table entry now.
2331 if (trylock_page(page
)) {
2334 mpfn
|= MIGRATE_PFN_LOCKED
;
2335 ptep_get_and_clear(mm
, addr
, ptep
);
2337 /* Setup special migration page table entry */
2338 entry
= make_migration_entry(page
, mpfn
&
2340 swp_pte
= swp_entry_to_pte(entry
);
2341 if (pte_soft_dirty(pte
))
2342 swp_pte
= pte_swp_mksoft_dirty(swp_pte
);
2343 set_pte_at(mm
, addr
, ptep
, swp_pte
);
2346 * This is like regular unmap: we remove the rmap and
2347 * drop page refcount. Page won't be freed, as we took
2348 * a reference just above.
2350 page_remove_rmap(page
, false);
2353 if (pte_present(pte
))
2358 migrate
->dst
[migrate
->npages
] = 0;
2359 migrate
->src
[migrate
->npages
++] = mpfn
;
2361 arch_leave_lazy_mmu_mode();
2362 pte_unmap_unlock(ptep
- 1, ptl
);
2364 /* Only flush the TLB if we actually modified any entries */
2366 flush_tlb_range(walk
->vma
, start
, end
);
2371 static const struct mm_walk_ops migrate_vma_walk_ops
= {
2372 .pmd_entry
= migrate_vma_collect_pmd
,
2373 .pte_hole
= migrate_vma_collect_hole
,
2377 * migrate_vma_collect() - collect pages over a range of virtual addresses
2378 * @migrate: migrate struct containing all migration information
2380 * This will walk the CPU page table. For each virtual address backed by a
2381 * valid page, it updates the src array and takes a reference on the page, in
2382 * order to pin the page until we lock it and unmap it.
2384 static void migrate_vma_collect(struct migrate_vma
*migrate
)
2386 struct mmu_notifier_range range
;
2388 mmu_notifier_range_init(&range
, MMU_NOTIFY_CLEAR
, 0, NULL
,
2389 migrate
->vma
->vm_mm
, migrate
->start
, migrate
->end
);
2390 mmu_notifier_invalidate_range_start(&range
);
2392 walk_page_range(migrate
->vma
->vm_mm
, migrate
->start
, migrate
->end
,
2393 &migrate_vma_walk_ops
, migrate
);
2395 mmu_notifier_invalidate_range_end(&range
);
2396 migrate
->end
= migrate
->start
+ (migrate
->npages
<< PAGE_SHIFT
);
2400 * migrate_vma_check_page() - check if page is pinned or not
2401 * @page: struct page to check
2403 * Pinned pages cannot be migrated. This is the same test as in
2404 * migrate_page_move_mapping(), except that here we allow migration of a
2407 static bool migrate_vma_check_page(struct page
*page
)
2410 * One extra ref because caller holds an extra reference, either from
2411 * isolate_lru_page() for a regular page, or migrate_vma_collect() for
2417 * FIXME support THP (transparent huge page), it is bit more complex to
2418 * check them than regular pages, because they can be mapped with a pmd
2419 * or with a pte (split pte mapping).
2421 if (PageCompound(page
))
2424 /* Page from ZONE_DEVICE have one extra reference */
2425 if (is_zone_device_page(page
)) {
2427 * Private page can never be pin as they have no valid pte and
2428 * GUP will fail for those. Yet if there is a pending migration
2429 * a thread might try to wait on the pte migration entry and
2430 * will bump the page reference count. Sadly there is no way to
2431 * differentiate a regular pin from migration wait. Hence to
2432 * avoid 2 racing thread trying to migrate back to CPU to enter
2433 * infinite loop (one stoping migration because the other is
2434 * waiting on pte migration entry). We always return true here.
2436 * FIXME proper solution is to rework migration_entry_wait() so
2437 * it does not need to take a reference on page.
2439 return is_device_private_page(page
);
2442 /* For file back page */
2443 if (page_mapping(page
))
2444 extra
+= 1 + page_has_private(page
);
2446 if ((page_count(page
) - extra
) > page_mapcount(page
))
2453 * migrate_vma_prepare() - lock pages and isolate them from the lru
2454 * @migrate: migrate struct containing all migration information
2456 * This locks pages that have been collected by migrate_vma_collect(). Once each
2457 * page is locked it is isolated from the lru (for non-device pages). Finally,
2458 * the ref taken by migrate_vma_collect() is dropped, as locked pages cannot be
2459 * migrated by concurrent kernel threads.
2461 static void migrate_vma_prepare(struct migrate_vma
*migrate
)
2463 const unsigned long npages
= migrate
->npages
;
2464 const unsigned long start
= migrate
->start
;
2465 unsigned long addr
, i
, restore
= 0;
2466 bool allow_drain
= true;
2470 for (i
= 0; (i
< npages
) && migrate
->cpages
; i
++) {
2471 struct page
*page
= migrate_pfn_to_page(migrate
->src
[i
]);
2477 if (!(migrate
->src
[i
] & MIGRATE_PFN_LOCKED
)) {
2479 * Because we are migrating several pages there can be
2480 * a deadlock between 2 concurrent migration where each
2481 * are waiting on each other page lock.
2483 * Make migrate_vma() a best effort thing and backoff
2484 * for any page we can not lock right away.
2486 if (!trylock_page(page
)) {
2487 migrate
->src
[i
] = 0;
2493 migrate
->src
[i
] |= MIGRATE_PFN_LOCKED
;
2496 /* ZONE_DEVICE pages are not on LRU */
2497 if (!is_zone_device_page(page
)) {
2498 if (!PageLRU(page
) && allow_drain
) {
2499 /* Drain CPU's pagevec */
2500 lru_add_drain_all();
2501 allow_drain
= false;
2504 if (isolate_lru_page(page
)) {
2506 migrate
->src
[i
] &= ~MIGRATE_PFN_MIGRATE
;
2510 migrate
->src
[i
] = 0;
2518 /* Drop the reference we took in collect */
2522 if (!migrate_vma_check_page(page
)) {
2524 migrate
->src
[i
] &= ~MIGRATE_PFN_MIGRATE
;
2528 if (!is_zone_device_page(page
)) {
2530 putback_lru_page(page
);
2533 migrate
->src
[i
] = 0;
2537 if (!is_zone_device_page(page
))
2538 putback_lru_page(page
);
2545 for (i
= 0, addr
= start
; i
< npages
&& restore
; i
++, addr
+= PAGE_SIZE
) {
2546 struct page
*page
= migrate_pfn_to_page(migrate
->src
[i
]);
2548 if (!page
|| (migrate
->src
[i
] & MIGRATE_PFN_MIGRATE
))
2551 remove_migration_pte(page
, migrate
->vma
, addr
, page
);
2553 migrate
->src
[i
] = 0;
2561 * migrate_vma_unmap() - replace page mapping with special migration pte entry
2562 * @migrate: migrate struct containing all migration information
2564 * Replace page mapping (CPU page table pte) with a special migration pte entry
2565 * and check again if it has been pinned. Pinned pages are restored because we
2566 * cannot migrate them.
2568 * This is the last step before we call the device driver callback to allocate
2569 * destination memory and copy contents of original page over to new page.
2571 static void migrate_vma_unmap(struct migrate_vma
*migrate
)
2573 int flags
= TTU_MIGRATION
| TTU_IGNORE_MLOCK
| TTU_IGNORE_ACCESS
;
2574 const unsigned long npages
= migrate
->npages
;
2575 const unsigned long start
= migrate
->start
;
2576 unsigned long addr
, i
, restore
= 0;
2578 for (i
= 0; i
< npages
; i
++) {
2579 struct page
*page
= migrate_pfn_to_page(migrate
->src
[i
]);
2581 if (!page
|| !(migrate
->src
[i
] & MIGRATE_PFN_MIGRATE
))
2584 if (page_mapped(page
)) {
2585 try_to_unmap(page
, flags
);
2586 if (page_mapped(page
))
2590 if (migrate_vma_check_page(page
))
2594 migrate
->src
[i
] &= ~MIGRATE_PFN_MIGRATE
;
2599 for (addr
= start
, i
= 0; i
< npages
&& restore
; addr
+= PAGE_SIZE
, i
++) {
2600 struct page
*page
= migrate_pfn_to_page(migrate
->src
[i
]);
2602 if (!page
|| (migrate
->src
[i
] & MIGRATE_PFN_MIGRATE
))
2605 remove_migration_ptes(page
, page
, false);
2607 migrate
->src
[i
] = 0;
2611 if (is_zone_device_page(page
))
2614 putback_lru_page(page
);
2619 * migrate_vma_setup() - prepare to migrate a range of memory
2620 * @args: contains the vma, start, and and pfns arrays for the migration
2622 * Returns: negative errno on failures, 0 when 0 or more pages were migrated
2625 * Prepare to migrate a range of memory virtual address range by collecting all
2626 * the pages backing each virtual address in the range, saving them inside the
2627 * src array. Then lock those pages and unmap them. Once the pages are locked
2628 * and unmapped, check whether each page is pinned or not. Pages that aren't
2629 * pinned have the MIGRATE_PFN_MIGRATE flag set (by this function) in the
2630 * corresponding src array entry. Then restores any pages that are pinned, by
2631 * remapping and unlocking those pages.
2633 * The caller should then allocate destination memory and copy source memory to
2634 * it for all those entries (ie with MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE
2635 * flag set). Once these are allocated and copied, the caller must update each
2636 * corresponding entry in the dst array with the pfn value of the destination
2637 * page and with the MIGRATE_PFN_VALID and MIGRATE_PFN_LOCKED flags set
2638 * (destination pages must have their struct pages locked, via lock_page()).
2640 * Note that the caller does not have to migrate all the pages that are marked
2641 * with MIGRATE_PFN_MIGRATE flag in src array unless this is a migration from
2642 * device memory to system memory. If the caller cannot migrate a device page
2643 * back to system memory, then it must return VM_FAULT_SIGBUS, which has severe
2644 * consequences for the userspace process, so it must be avoided if at all
2647 * For empty entries inside CPU page table (pte_none() or pmd_none() is true) we
2648 * do set MIGRATE_PFN_MIGRATE flag inside the corresponding source array thus
2649 * allowing the caller to allocate device memory for those unback virtual
2650 * address. For this the caller simply has to allocate device memory and
2651 * properly set the destination entry like for regular migration. Note that
2652 * this can still fails and thus inside the device driver must check if the
2653 * migration was successful for those entries after calling migrate_vma_pages()
2654 * just like for regular migration.
2656 * After that, the callers must call migrate_vma_pages() to go over each entry
2657 * in the src array that has the MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE flag
2658 * set. If the corresponding entry in dst array has MIGRATE_PFN_VALID flag set,
2659 * then migrate_vma_pages() to migrate struct page information from the source
2660 * struct page to the destination struct page. If it fails to migrate the
2661 * struct page information, then it clears the MIGRATE_PFN_MIGRATE flag in the
2664 * At this point all successfully migrated pages have an entry in the src
2665 * array with MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE flag set and the dst
2666 * array entry with MIGRATE_PFN_VALID flag set.
2668 * Once migrate_vma_pages() returns the caller may inspect which pages were
2669 * successfully migrated, and which were not. Successfully migrated pages will
2670 * have the MIGRATE_PFN_MIGRATE flag set for their src array entry.
2672 * It is safe to update device page table after migrate_vma_pages() because
2673 * both destination and source page are still locked, and the mmap_sem is held
2674 * in read mode (hence no one can unmap the range being migrated).
2676 * Once the caller is done cleaning up things and updating its page table (if it
2677 * chose to do so, this is not an obligation) it finally calls
2678 * migrate_vma_finalize() to update the CPU page table to point to new pages
2679 * for successfully migrated pages or otherwise restore the CPU page table to
2680 * point to the original source pages.
2682 int migrate_vma_setup(struct migrate_vma
*args
)
2684 long nr_pages
= (args
->end
- args
->start
) >> PAGE_SHIFT
;
2686 args
->start
&= PAGE_MASK
;
2687 args
->end
&= PAGE_MASK
;
2688 if (!args
->vma
|| is_vm_hugetlb_page(args
->vma
) ||
2689 (args
->vma
->vm_flags
& VM_SPECIAL
) || vma_is_dax(args
->vma
))
2693 if (args
->start
< args
->vma
->vm_start
||
2694 args
->start
>= args
->vma
->vm_end
)
2696 if (args
->end
<= args
->vma
->vm_start
|| args
->end
> args
->vma
->vm_end
)
2698 if (!args
->src
|| !args
->dst
)
2701 memset(args
->src
, 0, sizeof(*args
->src
) * nr_pages
);
2705 migrate_vma_collect(args
);
2708 migrate_vma_prepare(args
);
2710 migrate_vma_unmap(args
);
2713 * At this point pages are locked and unmapped, and thus they have
2714 * stable content and can safely be copied to destination memory that
2715 * is allocated by the drivers.
2720 EXPORT_SYMBOL(migrate_vma_setup
);
2723 * This code closely matches the code in:
2724 * __handle_mm_fault()
2725 * handle_pte_fault()
2726 * do_anonymous_page()
2727 * to map in an anonymous zero page but the struct page will be a ZONE_DEVICE
2730 static void migrate_vma_insert_page(struct migrate_vma
*migrate
,
2736 struct vm_area_struct
*vma
= migrate
->vma
;
2737 struct mm_struct
*mm
= vma
->vm_mm
;
2738 struct mem_cgroup
*memcg
;
2748 /* Only allow populating anonymous memory */
2749 if (!vma_is_anonymous(vma
))
2752 pgdp
= pgd_offset(mm
, addr
);
2753 p4dp
= p4d_alloc(mm
, pgdp
, addr
);
2756 pudp
= pud_alloc(mm
, p4dp
, addr
);
2759 pmdp
= pmd_alloc(mm
, pudp
, addr
);
2763 if (pmd_trans_huge(*pmdp
) || pmd_devmap(*pmdp
))
2767 * Use pte_alloc() instead of pte_alloc_map(). We can't run
2768 * pte_offset_map() on pmds where a huge pmd might be created
2769 * from a different thread.
2771 * pte_alloc_map() is safe to use under down_write(mmap_sem) or when
2772 * parallel threads are excluded by other means.
2774 * Here we only have down_read(mmap_sem).
2776 if (pte_alloc(mm
, pmdp
))
2779 /* See the comment in pte_alloc_one_map() */
2780 if (unlikely(pmd_trans_unstable(pmdp
)))
2783 if (unlikely(anon_vma_prepare(vma
)))
2785 if (mem_cgroup_try_charge(page
, vma
->vm_mm
, GFP_KERNEL
, &memcg
, false))
2789 * The memory barrier inside __SetPageUptodate makes sure that
2790 * preceding stores to the page contents become visible before
2791 * the set_pte_at() write.
2793 __SetPageUptodate(page
);
2795 if (is_zone_device_page(page
)) {
2796 if (is_device_private_page(page
)) {
2797 swp_entry_t swp_entry
;
2799 swp_entry
= make_device_private_entry(page
, vma
->vm_flags
& VM_WRITE
);
2800 entry
= swp_entry_to_pte(swp_entry
);
2803 entry
= mk_pte(page
, vma
->vm_page_prot
);
2804 if (vma
->vm_flags
& VM_WRITE
)
2805 entry
= pte_mkwrite(pte_mkdirty(entry
));
2808 ptep
= pte_offset_map_lock(mm
, pmdp
, addr
, &ptl
);
2810 if (check_stable_address_space(mm
))
2813 if (pte_present(*ptep
)) {
2814 unsigned long pfn
= pte_pfn(*ptep
);
2816 if (!is_zero_pfn(pfn
))
2819 } else if (!pte_none(*ptep
))
2823 * Check for userfaultfd but do not deliver the fault. Instead,
2826 if (userfaultfd_missing(vma
))
2829 inc_mm_counter(mm
, MM_ANONPAGES
);
2830 page_add_new_anon_rmap(page
, vma
, addr
, false);
2831 mem_cgroup_commit_charge(page
, memcg
, false, false);
2832 if (!is_zone_device_page(page
))
2833 lru_cache_add_active_or_unevictable(page
, vma
);
2837 flush_cache_page(vma
, addr
, pte_pfn(*ptep
));
2838 ptep_clear_flush_notify(vma
, addr
, ptep
);
2839 set_pte_at_notify(mm
, addr
, ptep
, entry
);
2840 update_mmu_cache(vma
, addr
, ptep
);
2842 /* No need to invalidate - it was non-present before */
2843 set_pte_at(mm
, addr
, ptep
, entry
);
2844 update_mmu_cache(vma
, addr
, ptep
);
2847 pte_unmap_unlock(ptep
, ptl
);
2848 *src
= MIGRATE_PFN_MIGRATE
;
2852 pte_unmap_unlock(ptep
, ptl
);
2853 mem_cgroup_cancel_charge(page
, memcg
, false);
2855 *src
&= ~MIGRATE_PFN_MIGRATE
;
2859 * migrate_vma_pages() - migrate meta-data from src page to dst page
2860 * @migrate: migrate struct containing all migration information
2862 * This migrates struct page meta-data from source struct page to destination
2863 * struct page. This effectively finishes the migration from source page to the
2866 void migrate_vma_pages(struct migrate_vma
*migrate
)
2868 const unsigned long npages
= migrate
->npages
;
2869 const unsigned long start
= migrate
->start
;
2870 struct mmu_notifier_range range
;
2871 unsigned long addr
, i
;
2872 bool notified
= false;
2874 for (i
= 0, addr
= start
; i
< npages
; addr
+= PAGE_SIZE
, i
++) {
2875 struct page
*newpage
= migrate_pfn_to_page(migrate
->dst
[i
]);
2876 struct page
*page
= migrate_pfn_to_page(migrate
->src
[i
]);
2877 struct address_space
*mapping
;
2881 migrate
->src
[i
] &= ~MIGRATE_PFN_MIGRATE
;
2886 if (!(migrate
->src
[i
] & MIGRATE_PFN_MIGRATE
))
2891 mmu_notifier_range_init(&range
,
2892 MMU_NOTIFY_CLEAR
, 0,
2894 migrate
->vma
->vm_mm
,
2895 addr
, migrate
->end
);
2896 mmu_notifier_invalidate_range_start(&range
);
2898 migrate_vma_insert_page(migrate
, addr
, newpage
,
2904 mapping
= page_mapping(page
);
2906 if (is_zone_device_page(newpage
)) {
2907 if (is_device_private_page(newpage
)) {
2909 * For now only support private anonymous when
2910 * migrating to un-addressable device memory.
2913 migrate
->src
[i
] &= ~MIGRATE_PFN_MIGRATE
;
2918 * Other types of ZONE_DEVICE page are not
2921 migrate
->src
[i
] &= ~MIGRATE_PFN_MIGRATE
;
2926 r
= migrate_page(mapping
, newpage
, page
, MIGRATE_SYNC_NO_COPY
);
2927 if (r
!= MIGRATEPAGE_SUCCESS
)
2928 migrate
->src
[i
] &= ~MIGRATE_PFN_MIGRATE
;
2932 * No need to double call mmu_notifier->invalidate_range() callback as
2933 * the above ptep_clear_flush_notify() inside migrate_vma_insert_page()
2934 * did already call it.
2937 mmu_notifier_invalidate_range_only_end(&range
);
2939 EXPORT_SYMBOL(migrate_vma_pages
);
2942 * migrate_vma_finalize() - restore CPU page table entry
2943 * @migrate: migrate struct containing all migration information
2945 * This replaces the special migration pte entry with either a mapping to the
2946 * new page if migration was successful for that page, or to the original page
2949 * This also unlocks the pages and puts them back on the lru, or drops the extra
2950 * refcount, for device pages.
2952 void migrate_vma_finalize(struct migrate_vma
*migrate
)
2954 const unsigned long npages
= migrate
->npages
;
2957 for (i
= 0; i
< npages
; i
++) {
2958 struct page
*newpage
= migrate_pfn_to_page(migrate
->dst
[i
]);
2959 struct page
*page
= migrate_pfn_to_page(migrate
->src
[i
]);
2963 unlock_page(newpage
);
2969 if (!(migrate
->src
[i
] & MIGRATE_PFN_MIGRATE
) || !newpage
) {
2971 unlock_page(newpage
);
2977 remove_migration_ptes(page
, newpage
, false);
2981 if (is_zone_device_page(page
))
2984 putback_lru_page(page
);
2986 if (newpage
!= page
) {
2987 unlock_page(newpage
);
2988 if (is_zone_device_page(newpage
))
2991 putback_lru_page(newpage
);
2995 EXPORT_SYMBOL(migrate_vma_finalize
);
2996 #endif /* CONFIG_DEVICE_PRIVATE */