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>
60 int isolate_movable_page(struct page
*page
, isolate_mode_t mode
)
62 struct address_space
*mapping
;
65 * Avoid burning cycles with pages that are yet under __free_pages(),
66 * or just got freed under us.
68 * In case we 'win' a race for a movable page being freed under us and
69 * raise its refcount preventing __free_pages() from doing its job
70 * the put_page() at the end of this block will take care of
71 * release this page, thus avoiding a nasty leakage.
73 if (unlikely(!get_page_unless_zero(page
)))
77 * Check PageMovable before holding a PG_lock because page's owner
78 * assumes anybody doesn't touch PG_lock of newly allocated page
79 * so unconditionally grabbing the lock ruins page's owner side.
81 if (unlikely(!__PageMovable(page
)))
84 * As movable pages are not isolated from LRU lists, concurrent
85 * compaction threads can race against page migration functions
86 * as well as race against the releasing a page.
88 * In order to avoid having an already isolated movable page
89 * being (wrongly) re-isolated while it is under migration,
90 * or to avoid attempting to isolate pages being released,
91 * lets be sure we have the page lock
92 * before proceeding with the movable page isolation steps.
94 if (unlikely(!trylock_page(page
)))
97 if (!PageMovable(page
) || PageIsolated(page
))
100 mapping
= page_mapping(page
);
101 VM_BUG_ON_PAGE(!mapping
, page
);
103 if (!mapping
->a_ops
->isolate_page(page
, mode
))
104 goto out_no_isolated
;
106 /* Driver shouldn't use PG_isolated bit of page->flags */
107 WARN_ON_ONCE(PageIsolated(page
));
108 __SetPageIsolated(page
);
121 static void putback_movable_page(struct page
*page
)
123 struct address_space
*mapping
;
125 mapping
= page_mapping(page
);
126 mapping
->a_ops
->putback_page(page
);
127 __ClearPageIsolated(page
);
131 * Put previously isolated pages back onto the appropriate lists
132 * from where they were once taken off for compaction/migration.
134 * This function shall be used whenever the isolated pageset has been
135 * built from lru, balloon, hugetlbfs page. See isolate_migratepages_range()
136 * and isolate_huge_page().
138 void putback_movable_pages(struct list_head
*l
)
143 list_for_each_entry_safe(page
, page2
, l
, lru
) {
144 if (unlikely(PageHuge(page
))) {
145 putback_active_hugepage(page
);
148 list_del(&page
->lru
);
150 * We isolated non-lru movable page so here we can use
151 * __PageMovable because LRU page's mapping cannot have
152 * PAGE_MAPPING_MOVABLE.
154 if (unlikely(__PageMovable(page
))) {
155 VM_BUG_ON_PAGE(!PageIsolated(page
), page
);
157 if (PageMovable(page
))
158 putback_movable_page(page
);
160 __ClearPageIsolated(page
);
164 mod_node_page_state(page_pgdat(page
), NR_ISOLATED_ANON
+
165 page_is_file_lru(page
), -thp_nr_pages(page
));
166 putback_lru_page(page
);
172 * Restore a potential migration pte to a working pte entry
174 static bool remove_migration_pte(struct page
*page
, struct vm_area_struct
*vma
,
175 unsigned long addr
, void *old
)
177 struct page_vma_mapped_walk pvmw
= {
181 .flags
= PVMW_SYNC
| PVMW_MIGRATION
,
187 VM_BUG_ON_PAGE(PageTail(page
), page
);
188 while (page_vma_mapped_walk(&pvmw
)) {
192 new = page
- pvmw
.page
->index
+
193 linear_page_index(vma
, pvmw
.address
);
195 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
196 /* PMD-mapped THP migration entry */
198 VM_BUG_ON_PAGE(PageHuge(page
) || !PageTransCompound(page
), page
);
199 remove_migration_pmd(&pvmw
, new);
205 pte
= pte_mkold(mk_pte(new, READ_ONCE(vma
->vm_page_prot
)));
206 if (pte_swp_soft_dirty(*pvmw
.pte
))
207 pte
= pte_mksoft_dirty(pte
);
210 * Recheck VMA as permissions can change since migration started
212 entry
= pte_to_swp_entry(*pvmw
.pte
);
213 if (is_writable_migration_entry(entry
))
214 pte
= maybe_mkwrite(pte
, vma
);
215 else if (pte_swp_uffd_wp(*pvmw
.pte
))
216 pte
= pte_mkuffd_wp(pte
);
218 if (unlikely(is_device_private_page(new))) {
220 entry
= make_writable_device_private_entry(
223 entry
= make_readable_device_private_entry(
225 pte
= swp_entry_to_pte(entry
);
226 if (pte_swp_soft_dirty(*pvmw
.pte
))
227 pte
= pte_swp_mksoft_dirty(pte
);
228 if (pte_swp_uffd_wp(*pvmw
.pte
))
229 pte
= pte_swp_mkuffd_wp(pte
);
232 #ifdef CONFIG_HUGETLB_PAGE
234 unsigned int shift
= huge_page_shift(hstate_vma(vma
));
236 pte
= pte_mkhuge(pte
);
237 pte
= arch_make_huge_pte(pte
, shift
, vma
->vm_flags
);
238 set_huge_pte_at(vma
->vm_mm
, pvmw
.address
, pvmw
.pte
, pte
);
240 hugepage_add_anon_rmap(new, vma
, pvmw
.address
);
242 page_dup_rmap(new, true);
246 set_pte_at(vma
->vm_mm
, pvmw
.address
, pvmw
.pte
, pte
);
249 page_add_anon_rmap(new, vma
, pvmw
.address
, false);
251 page_add_file_rmap(new, false);
253 if (vma
->vm_flags
& VM_LOCKED
&& !PageTransCompound(new))
256 if (PageTransHuge(page
) && PageMlocked(page
))
257 clear_page_mlock(page
);
259 /* No need to invalidate - it was non-present before */
260 update_mmu_cache(vma
, pvmw
.address
, pvmw
.pte
);
267 * Get rid of all migration entries and replace them by
268 * references to the indicated page.
270 void remove_migration_ptes(struct page
*old
, struct page
*new, bool locked
)
272 struct rmap_walk_control rwc
= {
273 .rmap_one
= remove_migration_pte
,
278 rmap_walk_locked(new, &rwc
);
280 rmap_walk(new, &rwc
);
284 * Something used the pte of a page under migration. We need to
285 * get to the page and wait until migration is finished.
286 * When we return from this function the fault will be retried.
288 void __migration_entry_wait(struct mm_struct
*mm
, pte_t
*ptep
,
297 if (!is_swap_pte(pte
))
300 entry
= pte_to_swp_entry(pte
);
301 if (!is_migration_entry(entry
))
304 page
= pfn_swap_entry_to_page(entry
);
305 page
= compound_head(page
);
308 * Once page cache replacement of page migration started, page_count
309 * is zero; but we must not call put_and_wait_on_page_locked() without
310 * a ref. Use get_page_unless_zero(), and just fault again if it fails.
312 if (!get_page_unless_zero(page
))
314 pte_unmap_unlock(ptep
, ptl
);
315 put_and_wait_on_page_locked(page
, TASK_UNINTERRUPTIBLE
);
318 pte_unmap_unlock(ptep
, ptl
);
321 void migration_entry_wait(struct mm_struct
*mm
, pmd_t
*pmd
,
322 unsigned long address
)
324 spinlock_t
*ptl
= pte_lockptr(mm
, pmd
);
325 pte_t
*ptep
= pte_offset_map(pmd
, address
);
326 __migration_entry_wait(mm
, ptep
, ptl
);
329 void migration_entry_wait_huge(struct vm_area_struct
*vma
,
330 struct mm_struct
*mm
, pte_t
*pte
)
332 spinlock_t
*ptl
= huge_pte_lockptr(hstate_vma(vma
), mm
, pte
);
333 __migration_entry_wait(mm
, pte
, ptl
);
336 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
337 void pmd_migration_entry_wait(struct mm_struct
*mm
, pmd_t
*pmd
)
342 ptl
= pmd_lock(mm
, pmd
);
343 if (!is_pmd_migration_entry(*pmd
))
345 page
= pfn_swap_entry_to_page(pmd_to_swp_entry(*pmd
));
346 if (!get_page_unless_zero(page
))
349 put_and_wait_on_page_locked(page
, TASK_UNINTERRUPTIBLE
);
356 static int expected_page_refs(struct address_space
*mapping
, struct page
*page
)
358 int expected_count
= 1;
361 * Device private pages have an extra refcount as they are
364 expected_count
+= is_device_private_page(page
);
366 expected_count
+= thp_nr_pages(page
) + page_has_private(page
);
368 return expected_count
;
372 * Replace the page in the mapping.
374 * The number of remaining references must be:
375 * 1 for anonymous pages without a mapping
376 * 2 for pages with a mapping
377 * 3 for pages with a mapping and PagePrivate/PagePrivate2 set.
379 int migrate_page_move_mapping(struct address_space
*mapping
,
380 struct page
*newpage
, struct page
*page
, int extra_count
)
382 XA_STATE(xas
, &mapping
->i_pages
, page_index(page
));
383 struct zone
*oldzone
, *newzone
;
385 int expected_count
= expected_page_refs(mapping
, page
) + extra_count
;
386 int nr
= thp_nr_pages(page
);
389 /* Anonymous page without mapping */
390 if (page_count(page
) != expected_count
)
393 /* No turning back from here */
394 newpage
->index
= page
->index
;
395 newpage
->mapping
= page
->mapping
;
396 if (PageSwapBacked(page
))
397 __SetPageSwapBacked(newpage
);
399 return MIGRATEPAGE_SUCCESS
;
402 oldzone
= page_zone(page
);
403 newzone
= page_zone(newpage
);
406 if (page_count(page
) != expected_count
|| xas_load(&xas
) != page
) {
407 xas_unlock_irq(&xas
);
411 if (!page_ref_freeze(page
, expected_count
)) {
412 xas_unlock_irq(&xas
);
417 * Now we know that no one else is looking at the page:
418 * no turning back from here.
420 newpage
->index
= page
->index
;
421 newpage
->mapping
= page
->mapping
;
422 page_ref_add(newpage
, nr
); /* add cache reference */
423 if (PageSwapBacked(page
)) {
424 __SetPageSwapBacked(newpage
);
425 if (PageSwapCache(page
)) {
426 SetPageSwapCache(newpage
);
427 set_page_private(newpage
, page_private(page
));
430 VM_BUG_ON_PAGE(PageSwapCache(page
), page
);
433 /* Move dirty while page refs frozen and newpage not yet exposed */
434 dirty
= PageDirty(page
);
436 ClearPageDirty(page
);
437 SetPageDirty(newpage
);
440 xas_store(&xas
, newpage
);
441 if (PageTransHuge(page
)) {
444 for (i
= 1; i
< nr
; i
++) {
446 xas_store(&xas
, newpage
);
451 * Drop cache reference from old page by unfreezing
452 * to one less reference.
453 * We know this isn't the last reference.
455 page_ref_unfreeze(page
, expected_count
- nr
);
458 /* Leave irq disabled to prevent preemption while updating stats */
461 * If moved to a different zone then also account
462 * the page for that zone. Other VM counters will be
463 * taken care of when we establish references to the
464 * new page and drop references to the old page.
466 * Note that anonymous pages are accounted for
467 * via NR_FILE_PAGES and NR_ANON_MAPPED if they
468 * are mapped to swap space.
470 if (newzone
!= oldzone
) {
471 struct lruvec
*old_lruvec
, *new_lruvec
;
472 struct mem_cgroup
*memcg
;
474 memcg
= page_memcg(page
);
475 old_lruvec
= mem_cgroup_lruvec(memcg
, oldzone
->zone_pgdat
);
476 new_lruvec
= mem_cgroup_lruvec(memcg
, newzone
->zone_pgdat
);
478 __mod_lruvec_state(old_lruvec
, NR_FILE_PAGES
, -nr
);
479 __mod_lruvec_state(new_lruvec
, NR_FILE_PAGES
, nr
);
480 if (PageSwapBacked(page
) && !PageSwapCache(page
)) {
481 __mod_lruvec_state(old_lruvec
, NR_SHMEM
, -nr
);
482 __mod_lruvec_state(new_lruvec
, NR_SHMEM
, nr
);
485 if (PageSwapCache(page
)) {
486 __mod_lruvec_state(old_lruvec
, NR_SWAPCACHE
, -nr
);
487 __mod_lruvec_state(new_lruvec
, NR_SWAPCACHE
, nr
);
490 if (dirty
&& mapping_can_writeback(mapping
)) {
491 __mod_lruvec_state(old_lruvec
, NR_FILE_DIRTY
, -nr
);
492 __mod_zone_page_state(oldzone
, NR_ZONE_WRITE_PENDING
, -nr
);
493 __mod_lruvec_state(new_lruvec
, NR_FILE_DIRTY
, nr
);
494 __mod_zone_page_state(newzone
, NR_ZONE_WRITE_PENDING
, nr
);
499 return MIGRATEPAGE_SUCCESS
;
501 EXPORT_SYMBOL(migrate_page_move_mapping
);
504 * The expected number of remaining references is the same as that
505 * of migrate_page_move_mapping().
507 int migrate_huge_page_move_mapping(struct address_space
*mapping
,
508 struct page
*newpage
, struct page
*page
)
510 XA_STATE(xas
, &mapping
->i_pages
, page_index(page
));
514 expected_count
= 2 + page_has_private(page
);
515 if (page_count(page
) != expected_count
|| xas_load(&xas
) != page
) {
516 xas_unlock_irq(&xas
);
520 if (!page_ref_freeze(page
, expected_count
)) {
521 xas_unlock_irq(&xas
);
525 newpage
->index
= page
->index
;
526 newpage
->mapping
= page
->mapping
;
530 xas_store(&xas
, newpage
);
532 page_ref_unfreeze(page
, expected_count
- 1);
534 xas_unlock_irq(&xas
);
536 return MIGRATEPAGE_SUCCESS
;
540 * Copy the page to its new location
542 void migrate_page_states(struct page
*newpage
, struct page
*page
)
547 SetPageError(newpage
);
548 if (PageReferenced(page
))
549 SetPageReferenced(newpage
);
550 if (PageUptodate(page
))
551 SetPageUptodate(newpage
);
552 if (TestClearPageActive(page
)) {
553 VM_BUG_ON_PAGE(PageUnevictable(page
), page
);
554 SetPageActive(newpage
);
555 } else if (TestClearPageUnevictable(page
))
556 SetPageUnevictable(newpage
);
557 if (PageWorkingset(page
))
558 SetPageWorkingset(newpage
);
559 if (PageChecked(page
))
560 SetPageChecked(newpage
);
561 if (PageMappedToDisk(page
))
562 SetPageMappedToDisk(newpage
);
564 /* Move dirty on pages not done by migrate_page_move_mapping() */
566 SetPageDirty(newpage
);
568 if (page_is_young(page
))
569 set_page_young(newpage
);
570 if (page_is_idle(page
))
571 set_page_idle(newpage
);
574 * Copy NUMA information to the new page, to prevent over-eager
575 * future migrations of this same page.
577 cpupid
= page_cpupid_xchg_last(page
, -1);
578 page_cpupid_xchg_last(newpage
, cpupid
);
580 ksm_migrate_page(newpage
, page
);
582 * Please do not reorder this without considering how mm/ksm.c's
583 * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache().
585 if (PageSwapCache(page
))
586 ClearPageSwapCache(page
);
587 ClearPagePrivate(page
);
589 /* page->private contains hugetlb specific flags */
591 set_page_private(page
, 0);
594 * If any waiters have accumulated on the new page then
597 if (PageWriteback(newpage
))
598 end_page_writeback(newpage
);
601 * PG_readahead shares the same bit with PG_reclaim. The above
602 * end_page_writeback() may clear PG_readahead mistakenly, so set the
605 if (PageReadahead(page
))
606 SetPageReadahead(newpage
);
608 copy_page_owner(page
, newpage
);
611 mem_cgroup_migrate(page
, newpage
);
613 EXPORT_SYMBOL(migrate_page_states
);
615 void migrate_page_copy(struct page
*newpage
, struct page
*page
)
617 if (PageHuge(page
) || PageTransHuge(page
))
618 copy_huge_page(newpage
, page
);
620 copy_highpage(newpage
, page
);
622 migrate_page_states(newpage
, page
);
624 EXPORT_SYMBOL(migrate_page_copy
);
626 /************************************************************
627 * Migration functions
628 ***********************************************************/
631 * Common logic to directly migrate a single LRU page suitable for
632 * pages that do not use PagePrivate/PagePrivate2.
634 * Pages are locked upon entry and exit.
636 int migrate_page(struct address_space
*mapping
,
637 struct page
*newpage
, struct page
*page
,
638 enum migrate_mode mode
)
642 BUG_ON(PageWriteback(page
)); /* Writeback must be complete */
644 rc
= migrate_page_move_mapping(mapping
, newpage
, page
, 0);
646 if (rc
!= MIGRATEPAGE_SUCCESS
)
649 if (mode
!= MIGRATE_SYNC_NO_COPY
)
650 migrate_page_copy(newpage
, page
);
652 migrate_page_states(newpage
, page
);
653 return MIGRATEPAGE_SUCCESS
;
655 EXPORT_SYMBOL(migrate_page
);
658 /* Returns true if all buffers are successfully locked */
659 static bool buffer_migrate_lock_buffers(struct buffer_head
*head
,
660 enum migrate_mode mode
)
662 struct buffer_head
*bh
= head
;
664 /* Simple case, sync compaction */
665 if (mode
!= MIGRATE_ASYNC
) {
668 bh
= bh
->b_this_page
;
670 } while (bh
!= head
);
675 /* async case, we cannot block on lock_buffer so use trylock_buffer */
677 if (!trylock_buffer(bh
)) {
679 * We failed to lock the buffer and cannot stall in
680 * async migration. Release the taken locks
682 struct buffer_head
*failed_bh
= bh
;
684 while (bh
!= failed_bh
) {
686 bh
= bh
->b_this_page
;
691 bh
= bh
->b_this_page
;
692 } while (bh
!= head
);
696 static int __buffer_migrate_page(struct address_space
*mapping
,
697 struct page
*newpage
, struct page
*page
, enum migrate_mode mode
,
700 struct buffer_head
*bh
, *head
;
704 if (!page_has_buffers(page
))
705 return migrate_page(mapping
, newpage
, page
, mode
);
707 /* Check whether page does not have extra refs before we do more work */
708 expected_count
= expected_page_refs(mapping
, page
);
709 if (page_count(page
) != expected_count
)
712 head
= page_buffers(page
);
713 if (!buffer_migrate_lock_buffers(head
, mode
))
718 bool invalidated
= false;
722 spin_lock(&mapping
->private_lock
);
725 if (atomic_read(&bh
->b_count
)) {
729 bh
= bh
->b_this_page
;
730 } while (bh
!= head
);
736 spin_unlock(&mapping
->private_lock
);
737 invalidate_bh_lrus();
739 goto recheck_buffers
;
743 rc
= migrate_page_move_mapping(mapping
, newpage
, page
, 0);
744 if (rc
!= MIGRATEPAGE_SUCCESS
)
747 attach_page_private(newpage
, detach_page_private(page
));
751 set_bh_page(bh
, newpage
, bh_offset(bh
));
752 bh
= bh
->b_this_page
;
754 } while (bh
!= head
);
756 if (mode
!= MIGRATE_SYNC_NO_COPY
)
757 migrate_page_copy(newpage
, page
);
759 migrate_page_states(newpage
, page
);
761 rc
= MIGRATEPAGE_SUCCESS
;
764 spin_unlock(&mapping
->private_lock
);
768 bh
= bh
->b_this_page
;
770 } while (bh
!= head
);
776 * Migration function for pages with buffers. This function can only be used
777 * if the underlying filesystem guarantees that no other references to "page"
778 * exist. For example attached buffer heads are accessed only under page lock.
780 int buffer_migrate_page(struct address_space
*mapping
,
781 struct page
*newpage
, struct page
*page
, enum migrate_mode mode
)
783 return __buffer_migrate_page(mapping
, newpage
, page
, mode
, false);
785 EXPORT_SYMBOL(buffer_migrate_page
);
788 * Same as above except that this variant is more careful and checks that there
789 * are also no buffer head references. This function is the right one for
790 * mappings where buffer heads are directly looked up and referenced (such as
791 * block device mappings).
793 int buffer_migrate_page_norefs(struct address_space
*mapping
,
794 struct page
*newpage
, struct page
*page
, enum migrate_mode mode
)
796 return __buffer_migrate_page(mapping
, newpage
, page
, mode
, true);
801 * Writeback a page to clean the dirty state
803 static int writeout(struct address_space
*mapping
, struct page
*page
)
805 struct writeback_control wbc
= {
806 .sync_mode
= WB_SYNC_NONE
,
809 .range_end
= LLONG_MAX
,
814 if (!mapping
->a_ops
->writepage
)
815 /* No write method for the address space */
818 if (!clear_page_dirty_for_io(page
))
819 /* Someone else already triggered a write */
823 * A dirty page may imply that the underlying filesystem has
824 * the page on some queue. So the page must be clean for
825 * migration. Writeout may mean we loose the lock and the
826 * page state is no longer what we checked for earlier.
827 * At this point we know that the migration attempt cannot
830 remove_migration_ptes(page
, page
, false);
832 rc
= mapping
->a_ops
->writepage(page
, &wbc
);
834 if (rc
!= AOP_WRITEPAGE_ACTIVATE
)
835 /* unlocked. Relock */
838 return (rc
< 0) ? -EIO
: -EAGAIN
;
842 * Default handling if a filesystem does not provide a migration function.
844 static int fallback_migrate_page(struct address_space
*mapping
,
845 struct page
*newpage
, struct page
*page
, enum migrate_mode mode
)
847 if (PageDirty(page
)) {
848 /* Only writeback pages in full synchronous migration */
851 case MIGRATE_SYNC_NO_COPY
:
856 return writeout(mapping
, page
);
860 * Buffers may be managed in a filesystem specific way.
861 * We must have no buffers or drop them.
863 if (page_has_private(page
) &&
864 !try_to_release_page(page
, GFP_KERNEL
))
865 return mode
== MIGRATE_SYNC
? -EAGAIN
: -EBUSY
;
867 return migrate_page(mapping
, newpage
, page
, mode
);
871 * Move a page to a newly allocated page
872 * The page is locked and all ptes have been successfully removed.
874 * The new page will have replaced the old page if this function
879 * MIGRATEPAGE_SUCCESS - success
881 static int move_to_new_page(struct page
*newpage
, struct page
*page
,
882 enum migrate_mode mode
)
884 struct address_space
*mapping
;
886 bool is_lru
= !__PageMovable(page
);
888 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
889 VM_BUG_ON_PAGE(!PageLocked(newpage
), newpage
);
891 mapping
= page_mapping(page
);
893 if (likely(is_lru
)) {
895 rc
= migrate_page(mapping
, newpage
, page
, mode
);
896 else if (mapping
->a_ops
->migratepage
)
898 * Most pages have a mapping and most filesystems
899 * provide a migratepage callback. Anonymous pages
900 * are part of swap space which also has its own
901 * migratepage callback. This is the most common path
902 * for page migration.
904 rc
= mapping
->a_ops
->migratepage(mapping
, newpage
,
907 rc
= fallback_migrate_page(mapping
, newpage
,
911 * In case of non-lru page, it could be released after
912 * isolation step. In that case, we shouldn't try migration.
914 VM_BUG_ON_PAGE(!PageIsolated(page
), page
);
915 if (!PageMovable(page
)) {
916 rc
= MIGRATEPAGE_SUCCESS
;
917 __ClearPageIsolated(page
);
921 rc
= mapping
->a_ops
->migratepage(mapping
, newpage
,
923 WARN_ON_ONCE(rc
== MIGRATEPAGE_SUCCESS
&&
924 !PageIsolated(page
));
928 * When successful, old pagecache page->mapping must be cleared before
929 * page is freed; but stats require that PageAnon be left as PageAnon.
931 if (rc
== MIGRATEPAGE_SUCCESS
) {
932 if (__PageMovable(page
)) {
933 VM_BUG_ON_PAGE(!PageIsolated(page
), page
);
936 * We clear PG_movable under page_lock so any compactor
937 * cannot try to migrate this page.
939 __ClearPageIsolated(page
);
943 * Anonymous and movable page->mapping will be cleared by
944 * free_pages_prepare so don't reset it here for keeping
945 * the type to work PageAnon, for example.
947 if (!PageMappingFlags(page
))
948 page
->mapping
= NULL
;
950 if (likely(!is_zone_device_page(newpage
)))
951 flush_dcache_page(newpage
);
958 static int __unmap_and_move(struct page
*page
, struct page
*newpage
,
959 int force
, enum migrate_mode mode
)
962 int page_was_mapped
= 0;
963 struct anon_vma
*anon_vma
= NULL
;
964 bool is_lru
= !__PageMovable(page
);
966 if (!trylock_page(page
)) {
967 if (!force
|| mode
== MIGRATE_ASYNC
)
971 * It's not safe for direct compaction to call lock_page.
972 * For example, during page readahead pages are added locked
973 * to the LRU. Later, when the IO completes the pages are
974 * marked uptodate and unlocked. However, the queueing
975 * could be merging multiple pages for one bio (e.g.
976 * mpage_readahead). If an allocation happens for the
977 * second or third page, the process can end up locking
978 * the same page twice and deadlocking. Rather than
979 * trying to be clever about what pages can be locked,
980 * avoid the use of lock_page for direct compaction
983 if (current
->flags
& PF_MEMALLOC
)
989 if (PageWriteback(page
)) {
991 * Only in the case of a full synchronous migration is it
992 * necessary to wait for PageWriteback. In the async case,
993 * the retry loop is too short and in the sync-light case,
994 * the overhead of stalling is too much
998 case MIGRATE_SYNC_NO_COPY
:
1006 wait_on_page_writeback(page
);
1010 * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
1011 * we cannot notice that anon_vma is freed while we migrates a page.
1012 * This get_anon_vma() delays freeing anon_vma pointer until the end
1013 * of migration. File cache pages are no problem because of page_lock()
1014 * File Caches may use write_page() or lock_page() in migration, then,
1015 * just care Anon page here.
1017 * Only page_get_anon_vma() understands the subtleties of
1018 * getting a hold on an anon_vma from outside one of its mms.
1019 * But if we cannot get anon_vma, then we won't need it anyway,
1020 * because that implies that the anon page is no longer mapped
1021 * (and cannot be remapped so long as we hold the page lock).
1023 if (PageAnon(page
) && !PageKsm(page
))
1024 anon_vma
= page_get_anon_vma(page
);
1027 * Block others from accessing the new page when we get around to
1028 * establishing additional references. We are usually the only one
1029 * holding a reference to newpage at this point. We used to have a BUG
1030 * here if trylock_page(newpage) fails, but would like to allow for
1031 * cases where there might be a race with the previous use of newpage.
1032 * This is much like races on refcount of oldpage: just don't BUG().
1034 if (unlikely(!trylock_page(newpage
)))
1037 if (unlikely(!is_lru
)) {
1038 rc
= move_to_new_page(newpage
, page
, mode
);
1039 goto out_unlock_both
;
1043 * Corner case handling:
1044 * 1. When a new swap-cache page is read into, it is added to the LRU
1045 * and treated as swapcache but it has no rmap yet.
1046 * Calling try_to_unmap() against a page->mapping==NULL page will
1047 * trigger a BUG. So handle it here.
1048 * 2. An orphaned page (see truncate_cleanup_page) might have
1049 * fs-private metadata. The page can be picked up due to memory
1050 * offlining. Everywhere else except page reclaim, the page is
1051 * invisible to the vm, so the page can not be migrated. So try to
1052 * free the metadata, so the page can be freed.
1054 if (!page
->mapping
) {
1055 VM_BUG_ON_PAGE(PageAnon(page
), page
);
1056 if (page_has_private(page
)) {
1057 try_to_free_buffers(page
);
1058 goto out_unlock_both
;
1060 } else if (page_mapped(page
)) {
1061 /* Establish migration ptes */
1062 VM_BUG_ON_PAGE(PageAnon(page
) && !PageKsm(page
) && !anon_vma
,
1064 try_to_migrate(page
, 0);
1065 page_was_mapped
= 1;
1068 if (!page_mapped(page
))
1069 rc
= move_to_new_page(newpage
, page
, mode
);
1071 if (page_was_mapped
)
1072 remove_migration_ptes(page
,
1073 rc
== MIGRATEPAGE_SUCCESS
? newpage
: page
, false);
1076 unlock_page(newpage
);
1078 /* Drop an anon_vma reference if we took one */
1080 put_anon_vma(anon_vma
);
1084 * If migration is successful, decrease refcount of the newpage
1085 * which will not free the page because new page owner increased
1086 * refcounter. As well, if it is LRU page, add the page to LRU
1087 * list in here. Use the old state of the isolated source page to
1088 * determine if we migrated a LRU page. newpage was already unlocked
1089 * and possibly modified by its owner - don't rely on the page
1092 if (rc
== MIGRATEPAGE_SUCCESS
) {
1093 if (unlikely(!is_lru
))
1096 putback_lru_page(newpage
);
1103 * Obtain the lock on page, remove all ptes and migrate the page
1104 * to the newly allocated page in newpage.
1106 static int unmap_and_move(new_page_t get_new_page
,
1107 free_page_t put_new_page
,
1108 unsigned long private, struct page
*page
,
1109 int force
, enum migrate_mode mode
,
1110 enum migrate_reason reason
,
1111 struct list_head
*ret
)
1113 int rc
= MIGRATEPAGE_SUCCESS
;
1114 struct page
*newpage
= NULL
;
1116 if (!thp_migration_supported() && PageTransHuge(page
))
1119 if (page_count(page
) == 1) {
1120 /* page was freed from under us. So we are done. */
1121 ClearPageActive(page
);
1122 ClearPageUnevictable(page
);
1123 if (unlikely(__PageMovable(page
))) {
1125 if (!PageMovable(page
))
1126 __ClearPageIsolated(page
);
1132 newpage
= get_new_page(page
, private);
1136 rc
= __unmap_and_move(page
, newpage
, force
, mode
);
1137 if (rc
== MIGRATEPAGE_SUCCESS
)
1138 set_page_owner_migrate_reason(newpage
, reason
);
1141 if (rc
!= -EAGAIN
) {
1143 * A page that has been migrated has all references
1144 * removed and will be freed. A page that has not been
1145 * migrated will have kept its references and be restored.
1147 list_del(&page
->lru
);
1151 * If migration is successful, releases reference grabbed during
1152 * isolation. Otherwise, restore the page to right list unless
1155 if (rc
== MIGRATEPAGE_SUCCESS
) {
1157 * Compaction can migrate also non-LRU pages which are
1158 * not accounted to NR_ISOLATED_*. They can be recognized
1161 if (likely(!__PageMovable(page
)))
1162 mod_node_page_state(page_pgdat(page
), NR_ISOLATED_ANON
+
1163 page_is_file_lru(page
), -thp_nr_pages(page
));
1165 if (reason
!= MR_MEMORY_FAILURE
)
1167 * We release the page in page_handle_poison.
1172 list_add_tail(&page
->lru
, ret
);
1175 put_new_page(newpage
, private);
1184 * Counterpart of unmap_and_move_page() for hugepage migration.
1186 * This function doesn't wait the completion of hugepage I/O
1187 * because there is no race between I/O and migration for hugepage.
1188 * Note that currently hugepage I/O occurs only in direct I/O
1189 * where no lock is held and PG_writeback is irrelevant,
1190 * and writeback status of all subpages are counted in the reference
1191 * count of the head page (i.e. if all subpages of a 2MB hugepage are
1192 * under direct I/O, the reference of the head page is 512 and a bit more.)
1193 * This means that when we try to migrate hugepage whose subpages are
1194 * doing direct I/O, some references remain after try_to_unmap() and
1195 * hugepage migration fails without data corruption.
1197 * There is also no race when direct I/O is issued on the page under migration,
1198 * because then pte is replaced with migration swap entry and direct I/O code
1199 * will wait in the page fault for migration to complete.
1201 static int unmap_and_move_huge_page(new_page_t get_new_page
,
1202 free_page_t put_new_page
, unsigned long private,
1203 struct page
*hpage
, int force
,
1204 enum migrate_mode mode
, int reason
,
1205 struct list_head
*ret
)
1208 int page_was_mapped
= 0;
1209 struct page
*new_hpage
;
1210 struct anon_vma
*anon_vma
= NULL
;
1211 struct address_space
*mapping
= NULL
;
1214 * Migratability of hugepages depends on architectures and their size.
1215 * This check is necessary because some callers of hugepage migration
1216 * like soft offline and memory hotremove don't walk through page
1217 * tables or check whether the hugepage is pmd-based or not before
1218 * kicking migration.
1220 if (!hugepage_migration_supported(page_hstate(hpage
))) {
1221 list_move_tail(&hpage
->lru
, ret
);
1225 if (page_count(hpage
) == 1) {
1226 /* page was freed from under us. So we are done. */
1227 putback_active_hugepage(hpage
);
1228 return MIGRATEPAGE_SUCCESS
;
1231 new_hpage
= get_new_page(hpage
, private);
1235 if (!trylock_page(hpage
)) {
1240 case MIGRATE_SYNC_NO_COPY
:
1249 * Check for pages which are in the process of being freed. Without
1250 * page_mapping() set, hugetlbfs specific move page routine will not
1251 * be called and we could leak usage counts for subpools.
1253 if (hugetlb_page_subpool(hpage
) && !page_mapping(hpage
)) {
1258 if (PageAnon(hpage
))
1259 anon_vma
= page_get_anon_vma(hpage
);
1261 if (unlikely(!trylock_page(new_hpage
)))
1264 if (page_mapped(hpage
)) {
1265 bool mapping_locked
= false;
1266 enum ttu_flags ttu
= 0;
1268 if (!PageAnon(hpage
)) {
1270 * In shared mappings, try_to_unmap could potentially
1271 * call huge_pmd_unshare. Because of this, take
1272 * semaphore in write mode here and set TTU_RMAP_LOCKED
1273 * to let lower levels know we have taken the lock.
1275 mapping
= hugetlb_page_mapping_lock_write(hpage
);
1276 if (unlikely(!mapping
))
1277 goto unlock_put_anon
;
1279 mapping_locked
= true;
1280 ttu
|= TTU_RMAP_LOCKED
;
1283 try_to_migrate(hpage
, ttu
);
1284 page_was_mapped
= 1;
1287 i_mmap_unlock_write(mapping
);
1290 if (!page_mapped(hpage
))
1291 rc
= move_to_new_page(new_hpage
, hpage
, mode
);
1293 if (page_was_mapped
)
1294 remove_migration_ptes(hpage
,
1295 rc
== MIGRATEPAGE_SUCCESS
? new_hpage
: hpage
, false);
1298 unlock_page(new_hpage
);
1302 put_anon_vma(anon_vma
);
1304 if (rc
== MIGRATEPAGE_SUCCESS
) {
1305 move_hugetlb_state(hpage
, new_hpage
, reason
);
1306 put_new_page
= NULL
;
1312 if (rc
== MIGRATEPAGE_SUCCESS
)
1313 putback_active_hugepage(hpage
);
1314 else if (rc
!= -EAGAIN
)
1315 list_move_tail(&hpage
->lru
, ret
);
1318 * If migration was not successful and there's a freeing callback, use
1319 * it. Otherwise, put_page() will drop the reference grabbed during
1323 put_new_page(new_hpage
, private);
1325 putback_active_hugepage(new_hpage
);
1330 static inline int try_split_thp(struct page
*page
, struct page
**page2
,
1331 struct list_head
*from
)
1336 rc
= split_huge_page_to_list(page
, from
);
1339 list_safe_reset_next(page
, *page2
, lru
);
1345 * migrate_pages - migrate the pages specified in a list, to the free pages
1346 * supplied as the target for the page migration
1348 * @from: The list of pages to be migrated.
1349 * @get_new_page: The function used to allocate free pages to be used
1350 * as the target of the page migration.
1351 * @put_new_page: The function used to free target pages if migration
1352 * fails, or NULL if no special handling is necessary.
1353 * @private: Private data to be passed on to get_new_page()
1354 * @mode: The migration mode that specifies the constraints for
1355 * page migration, if any.
1356 * @reason: The reason for page migration.
1358 * The function returns after 10 attempts or if no pages are movable any more
1359 * because the list has become empty or no retryable pages exist any more.
1360 * It is caller's responsibility to call putback_movable_pages() to return pages
1361 * to the LRU or free list only if ret != 0.
1363 * Returns the number of pages that were not migrated, or an error code.
1365 int migrate_pages(struct list_head
*from
, new_page_t get_new_page
,
1366 free_page_t put_new_page
, unsigned long private,
1367 enum migrate_mode mode
, int reason
)
1372 int nr_succeeded
= 0;
1373 int nr_thp_succeeded
= 0;
1374 int nr_thp_failed
= 0;
1375 int nr_thp_split
= 0;
1377 bool is_thp
= false;
1380 int swapwrite
= current
->flags
& PF_SWAPWRITE
;
1381 int rc
, nr_subpages
;
1382 LIST_HEAD(ret_pages
);
1383 bool nosplit
= (reason
== MR_NUMA_MISPLACED
);
1385 trace_mm_migrate_pages_start(mode
, reason
);
1388 current
->flags
|= PF_SWAPWRITE
;
1390 for (pass
= 0; pass
< 10 && (retry
|| thp_retry
); pass
++) {
1394 list_for_each_entry_safe(page
, page2
, from
, lru
) {
1397 * THP statistics is based on the source huge page.
1398 * Capture required information that might get lost
1401 is_thp
= PageTransHuge(page
) && !PageHuge(page
);
1402 nr_subpages
= thp_nr_pages(page
);
1406 rc
= unmap_and_move_huge_page(get_new_page
,
1407 put_new_page
, private, page
,
1408 pass
> 2, mode
, reason
,
1411 rc
= unmap_and_move(get_new_page
, put_new_page
,
1412 private, page
, pass
> 2, mode
,
1413 reason
, &ret_pages
);
1416 * Success: non hugetlb page will be freed, hugetlb
1417 * page will be put back
1418 * -EAGAIN: stay on the from list
1419 * -ENOMEM: stay on the from list
1420 * Other errno: put on ret_pages list then splice to
1425 * THP migration might be unsupported or the
1426 * allocation could've failed so we should
1427 * retry on the same page with the THP split
1430 * Head page is retried immediately and tail
1431 * pages are added to the tail of the list so
1432 * we encounter them after the rest of the list
1436 /* THP migration is unsupported */
1438 if (!try_split_thp(page
, &page2
, from
)) {
1444 nr_failed
+= nr_subpages
;
1448 /* Hugetlb migration is unsupported */
1453 * When memory is low, don't bother to try to migrate
1454 * other pages, just exit.
1455 * THP NUMA faulting doesn't split THP to retry.
1457 if (is_thp
&& !nosplit
) {
1458 if (!try_split_thp(page
, &page2
, from
)) {
1464 nr_failed
+= nr_subpages
;
1476 case MIGRATEPAGE_SUCCESS
:
1479 nr_succeeded
+= nr_subpages
;
1486 * Permanent failure (-EBUSY, etc.):
1487 * unlike -EAGAIN case, the failed page is
1488 * removed from migration page list and not
1489 * retried in the next outer loop.
1493 nr_failed
+= nr_subpages
;
1501 nr_failed
+= retry
+ thp_retry
;
1502 nr_thp_failed
+= thp_retry
;
1506 * Put the permanent failure page back to migration list, they
1507 * will be put back to the right list by the caller.
1509 list_splice(&ret_pages
, from
);
1511 count_vm_events(PGMIGRATE_SUCCESS
, nr_succeeded
);
1512 count_vm_events(PGMIGRATE_FAIL
, nr_failed
);
1513 count_vm_events(THP_MIGRATION_SUCCESS
, nr_thp_succeeded
);
1514 count_vm_events(THP_MIGRATION_FAIL
, nr_thp_failed
);
1515 count_vm_events(THP_MIGRATION_SPLIT
, nr_thp_split
);
1516 trace_mm_migrate_pages(nr_succeeded
, nr_failed
, nr_thp_succeeded
,
1517 nr_thp_failed
, nr_thp_split
, mode
, reason
);
1520 current
->flags
&= ~PF_SWAPWRITE
;
1525 struct page
*alloc_migration_target(struct page
*page
, unsigned long private)
1527 struct migration_target_control
*mtc
;
1529 unsigned int order
= 0;
1530 struct page
*new_page
= NULL
;
1534 mtc
= (struct migration_target_control
*)private;
1535 gfp_mask
= mtc
->gfp_mask
;
1537 if (nid
== NUMA_NO_NODE
)
1538 nid
= page_to_nid(page
);
1540 if (PageHuge(page
)) {
1541 struct hstate
*h
= page_hstate(compound_head(page
));
1543 gfp_mask
= htlb_modify_alloc_mask(h
, gfp_mask
);
1544 return alloc_huge_page_nodemask(h
, nid
, mtc
->nmask
, gfp_mask
);
1547 if (PageTransHuge(page
)) {
1549 * clear __GFP_RECLAIM to make the migration callback
1550 * consistent with regular THP allocations.
1552 gfp_mask
&= ~__GFP_RECLAIM
;
1553 gfp_mask
|= GFP_TRANSHUGE
;
1554 order
= HPAGE_PMD_ORDER
;
1556 zidx
= zone_idx(page_zone(page
));
1557 if (is_highmem_idx(zidx
) || zidx
== ZONE_MOVABLE
)
1558 gfp_mask
|= __GFP_HIGHMEM
;
1560 new_page
= __alloc_pages(gfp_mask
, order
, nid
, mtc
->nmask
);
1562 if (new_page
&& PageTransHuge(new_page
))
1563 prep_transhuge_page(new_page
);
1570 static int store_status(int __user
*status
, int start
, int value
, int nr
)
1573 if (put_user(value
, status
+ start
))
1581 static int do_move_pages_to_node(struct mm_struct
*mm
,
1582 struct list_head
*pagelist
, int node
)
1585 struct migration_target_control mtc
= {
1587 .gfp_mask
= GFP_HIGHUSER_MOVABLE
| __GFP_THISNODE
,
1590 err
= migrate_pages(pagelist
, alloc_migration_target
, NULL
,
1591 (unsigned long)&mtc
, MIGRATE_SYNC
, MR_SYSCALL
);
1593 putback_movable_pages(pagelist
);
1598 * Resolves the given address to a struct page, isolates it from the LRU and
1599 * puts it to the given pagelist.
1601 * errno - if the page cannot be found/isolated
1602 * 0 - when it doesn't have to be migrated because it is already on the
1604 * 1 - when it has been queued
1606 static int add_page_for_migration(struct mm_struct
*mm
, unsigned long addr
,
1607 int node
, struct list_head
*pagelist
, bool migrate_all
)
1609 struct vm_area_struct
*vma
;
1611 unsigned int follflags
;
1616 vma
= find_vma(mm
, addr
);
1617 if (!vma
|| addr
< vma
->vm_start
|| !vma_migratable(vma
))
1620 /* FOLL_DUMP to ignore special (like zero) pages */
1621 follflags
= FOLL_GET
| FOLL_DUMP
;
1622 page
= follow_page(vma
, addr
, follflags
);
1624 err
= PTR_ERR(page
);
1633 if (page_to_nid(page
) == node
)
1637 if (page_mapcount(page
) > 1 && !migrate_all
)
1640 if (PageHuge(page
)) {
1641 if (PageHead(page
)) {
1642 isolate_huge_page(page
, pagelist
);
1648 head
= compound_head(page
);
1649 err
= isolate_lru_page(head
);
1654 list_add_tail(&head
->lru
, pagelist
);
1655 mod_node_page_state(page_pgdat(head
),
1656 NR_ISOLATED_ANON
+ page_is_file_lru(head
),
1657 thp_nr_pages(head
));
1661 * Either remove the duplicate refcount from
1662 * isolate_lru_page() or drop the page ref if it was
1667 mmap_read_unlock(mm
);
1671 static int move_pages_and_store_status(struct mm_struct
*mm
, int node
,
1672 struct list_head
*pagelist
, int __user
*status
,
1673 int start
, int i
, unsigned long nr_pages
)
1677 if (list_empty(pagelist
))
1680 err
= do_move_pages_to_node(mm
, pagelist
, node
);
1683 * Positive err means the number of failed
1684 * pages to migrate. Since we are going to
1685 * abort and return the number of non-migrated
1686 * pages, so need to include the rest of the
1687 * nr_pages that have not been attempted as
1691 err
+= nr_pages
- i
- 1;
1694 return store_status(status
, start
, node
, i
- start
);
1698 * Migrate an array of page address onto an array of nodes and fill
1699 * the corresponding array of status.
1701 static int do_pages_move(struct mm_struct
*mm
, nodemask_t task_nodes
,
1702 unsigned long nr_pages
,
1703 const void __user
* __user
*pages
,
1704 const int __user
*nodes
,
1705 int __user
*status
, int flags
)
1707 int current_node
= NUMA_NO_NODE
;
1708 LIST_HEAD(pagelist
);
1712 lru_cache_disable();
1714 for (i
= start
= 0; i
< nr_pages
; i
++) {
1715 const void __user
*p
;
1720 if (get_user(p
, pages
+ i
))
1722 if (get_user(node
, nodes
+ i
))
1724 addr
= (unsigned long)untagged_addr(p
);
1727 if (node
< 0 || node
>= MAX_NUMNODES
)
1729 if (!node_state(node
, N_MEMORY
))
1733 if (!node_isset(node
, task_nodes
))
1736 if (current_node
== NUMA_NO_NODE
) {
1737 current_node
= node
;
1739 } else if (node
!= current_node
) {
1740 err
= move_pages_and_store_status(mm
, current_node
,
1741 &pagelist
, status
, start
, i
, nr_pages
);
1745 current_node
= node
;
1749 * Errors in the page lookup or isolation are not fatal and we simply
1750 * report them via status
1752 err
= add_page_for_migration(mm
, addr
, current_node
,
1753 &pagelist
, flags
& MPOL_MF_MOVE_ALL
);
1756 /* The page is successfully queued for migration */
1761 * If the page is already on the target node (!err), store the
1762 * node, otherwise, store the err.
1764 err
= store_status(status
, i
, err
? : current_node
, 1);
1768 err
= move_pages_and_store_status(mm
, current_node
, &pagelist
,
1769 status
, start
, i
, nr_pages
);
1772 current_node
= NUMA_NO_NODE
;
1775 /* Make sure we do not overwrite the existing error */
1776 err1
= move_pages_and_store_status(mm
, current_node
, &pagelist
,
1777 status
, start
, i
, nr_pages
);
1786 * Determine the nodes of an array of pages and store it in an array of status.
1788 static void do_pages_stat_array(struct mm_struct
*mm
, unsigned long nr_pages
,
1789 const void __user
**pages
, int *status
)
1795 for (i
= 0; i
< nr_pages
; i
++) {
1796 unsigned long addr
= (unsigned long)(*pages
);
1797 struct vm_area_struct
*vma
;
1801 vma
= vma_lookup(mm
, addr
);
1805 /* FOLL_DUMP to ignore special (like zero) pages */
1806 page
= follow_page(vma
, addr
, FOLL_DUMP
);
1808 err
= PTR_ERR(page
);
1812 err
= page
? page_to_nid(page
) : -ENOENT
;
1820 mmap_read_unlock(mm
);
1824 * Determine the nodes of a user array of pages and store it in
1825 * a user array of status.
1827 static int do_pages_stat(struct mm_struct
*mm
, unsigned long nr_pages
,
1828 const void __user
* __user
*pages
,
1831 #define DO_PAGES_STAT_CHUNK_NR 16
1832 const void __user
*chunk_pages
[DO_PAGES_STAT_CHUNK_NR
];
1833 int chunk_status
[DO_PAGES_STAT_CHUNK_NR
];
1836 unsigned long chunk_nr
;
1838 chunk_nr
= nr_pages
;
1839 if (chunk_nr
> DO_PAGES_STAT_CHUNK_NR
)
1840 chunk_nr
= DO_PAGES_STAT_CHUNK_NR
;
1842 if (copy_from_user(chunk_pages
, pages
, chunk_nr
* sizeof(*chunk_pages
)))
1845 do_pages_stat_array(mm
, chunk_nr
, chunk_pages
, chunk_status
);
1847 if (copy_to_user(status
, chunk_status
, chunk_nr
* sizeof(*status
)))
1852 nr_pages
-= chunk_nr
;
1854 return nr_pages
? -EFAULT
: 0;
1857 static struct mm_struct
*find_mm_struct(pid_t pid
, nodemask_t
*mem_nodes
)
1859 struct task_struct
*task
;
1860 struct mm_struct
*mm
;
1863 * There is no need to check if current process has the right to modify
1864 * the specified process when they are same.
1868 *mem_nodes
= cpuset_mems_allowed(current
);
1872 /* Find the mm_struct */
1874 task
= find_task_by_vpid(pid
);
1877 return ERR_PTR(-ESRCH
);
1879 get_task_struct(task
);
1882 * Check if this process has the right to modify the specified
1883 * process. Use the regular "ptrace_may_access()" checks.
1885 if (!ptrace_may_access(task
, PTRACE_MODE_READ_REALCREDS
)) {
1887 mm
= ERR_PTR(-EPERM
);
1892 mm
= ERR_PTR(security_task_movememory(task
));
1895 *mem_nodes
= cpuset_mems_allowed(task
);
1896 mm
= get_task_mm(task
);
1898 put_task_struct(task
);
1900 mm
= ERR_PTR(-EINVAL
);
1905 * Move a list of pages in the address space of the currently executing
1908 static int kernel_move_pages(pid_t pid
, unsigned long nr_pages
,
1909 const void __user
* __user
*pages
,
1910 const int __user
*nodes
,
1911 int __user
*status
, int flags
)
1913 struct mm_struct
*mm
;
1915 nodemask_t task_nodes
;
1918 if (flags
& ~(MPOL_MF_MOVE
|MPOL_MF_MOVE_ALL
))
1921 if ((flags
& MPOL_MF_MOVE_ALL
) && !capable(CAP_SYS_NICE
))
1924 mm
= find_mm_struct(pid
, &task_nodes
);
1929 err
= do_pages_move(mm
, task_nodes
, nr_pages
, pages
,
1930 nodes
, status
, flags
);
1932 err
= do_pages_stat(mm
, nr_pages
, pages
, status
);
1938 SYSCALL_DEFINE6(move_pages
, pid_t
, pid
, unsigned long, nr_pages
,
1939 const void __user
* __user
*, pages
,
1940 const int __user
*, nodes
,
1941 int __user
*, status
, int, flags
)
1943 return kernel_move_pages(pid
, nr_pages
, pages
, nodes
, status
, flags
);
1946 #ifdef CONFIG_COMPAT
1947 COMPAT_SYSCALL_DEFINE6(move_pages
, pid_t
, pid
, compat_ulong_t
, nr_pages
,
1948 compat_uptr_t __user
*, pages32
,
1949 const int __user
*, nodes
,
1950 int __user
*, status
,
1953 const void __user
* __user
*pages
;
1956 pages
= compat_alloc_user_space(nr_pages
* sizeof(void *));
1957 for (i
= 0; i
< nr_pages
; i
++) {
1960 if (get_user(p
, pages32
+ i
) ||
1961 put_user(compat_ptr(p
), pages
+ i
))
1964 return kernel_move_pages(pid
, nr_pages
, pages
, nodes
, status
, flags
);
1966 #endif /* CONFIG_COMPAT */
1968 #ifdef CONFIG_NUMA_BALANCING
1970 * Returns true if this is a safe migration target node for misplaced NUMA
1971 * pages. Currently it only checks the watermarks which crude
1973 static bool migrate_balanced_pgdat(struct pglist_data
*pgdat
,
1974 unsigned long nr_migrate_pages
)
1978 for (z
= pgdat
->nr_zones
- 1; z
>= 0; z
--) {
1979 struct zone
*zone
= pgdat
->node_zones
+ z
;
1981 if (!populated_zone(zone
))
1984 /* Avoid waking kswapd by allocating pages_to_migrate pages. */
1985 if (!zone_watermark_ok(zone
, 0,
1986 high_wmark_pages(zone
) +
1995 static struct page
*alloc_misplaced_dst_page(struct page
*page
,
1998 int nid
= (int) data
;
1999 struct page
*newpage
;
2001 newpage
= __alloc_pages_node(nid
,
2002 (GFP_HIGHUSER_MOVABLE
|
2003 __GFP_THISNODE
| __GFP_NOMEMALLOC
|
2004 __GFP_NORETRY
| __GFP_NOWARN
) &
2010 static struct page
*alloc_misplaced_dst_page_thp(struct page
*page
,
2013 int nid
= (int) data
;
2014 struct page
*newpage
;
2016 newpage
= alloc_pages_node(nid
, (GFP_TRANSHUGE_LIGHT
| __GFP_THISNODE
),
2021 prep_transhuge_page(newpage
);
2027 static int numamigrate_isolate_page(pg_data_t
*pgdat
, struct page
*page
)
2031 VM_BUG_ON_PAGE(compound_order(page
) && !PageTransHuge(page
), page
);
2033 /* Do not migrate THP mapped by multiple processes */
2034 if (PageTransHuge(page
) && total_mapcount(page
) > 1)
2037 /* Avoid migrating to a node that is nearly full */
2038 if (!migrate_balanced_pgdat(pgdat
, compound_nr(page
)))
2041 if (isolate_lru_page(page
))
2044 page_lru
= page_is_file_lru(page
);
2045 mod_node_page_state(page_pgdat(page
), NR_ISOLATED_ANON
+ page_lru
,
2046 thp_nr_pages(page
));
2049 * Isolating the page has taken another reference, so the
2050 * caller's reference can be safely dropped without the page
2051 * disappearing underneath us during migration.
2058 * Attempt to migrate a misplaced page to the specified destination
2059 * node. Caller is expected to have an elevated reference count on
2060 * the page that will be dropped by this function before returning.
2062 int migrate_misplaced_page(struct page
*page
, struct vm_area_struct
*vma
,
2065 pg_data_t
*pgdat
= NODE_DATA(node
);
2068 LIST_HEAD(migratepages
);
2071 int nr_pages
= thp_nr_pages(page
);
2074 * PTE mapped THP or HugeTLB page can't reach here so the page could
2075 * be either base page or THP. And it must be head page if it is
2078 compound
= PageTransHuge(page
);
2081 new = alloc_misplaced_dst_page_thp
;
2083 new = alloc_misplaced_dst_page
;
2086 * Don't migrate file pages that are mapped in multiple processes
2087 * with execute permissions as they are probably shared libraries.
2089 if (page_mapcount(page
) != 1 && page_is_file_lru(page
) &&
2090 (vma
->vm_flags
& VM_EXEC
))
2094 * Also do not migrate dirty pages as not all filesystems can move
2095 * dirty pages in MIGRATE_ASYNC mode which is a waste of cycles.
2097 if (page_is_file_lru(page
) && PageDirty(page
))
2100 isolated
= numamigrate_isolate_page(pgdat
, page
);
2104 list_add(&page
->lru
, &migratepages
);
2105 nr_remaining
= migrate_pages(&migratepages
, *new, NULL
, node
,
2106 MIGRATE_ASYNC
, MR_NUMA_MISPLACED
);
2108 if (!list_empty(&migratepages
)) {
2109 list_del(&page
->lru
);
2110 mod_node_page_state(page_pgdat(page
), NR_ISOLATED_ANON
+
2111 page_is_file_lru(page
), -nr_pages
);
2112 putback_lru_page(page
);
2116 count_vm_numa_events(NUMA_PAGE_MIGRATE
, nr_pages
);
2117 BUG_ON(!list_empty(&migratepages
));
2124 #endif /* CONFIG_NUMA_BALANCING */
2125 #endif /* CONFIG_NUMA */
2127 #ifdef CONFIG_DEVICE_PRIVATE
2128 static int migrate_vma_collect_skip(unsigned long start
,
2130 struct mm_walk
*walk
)
2132 struct migrate_vma
*migrate
= walk
->private;
2135 for (addr
= start
; addr
< end
; addr
+= PAGE_SIZE
) {
2136 migrate
->dst
[migrate
->npages
] = 0;
2137 migrate
->src
[migrate
->npages
++] = 0;
2143 static int migrate_vma_collect_hole(unsigned long start
,
2145 __always_unused
int depth
,
2146 struct mm_walk
*walk
)
2148 struct migrate_vma
*migrate
= walk
->private;
2151 /* Only allow populating anonymous memory. */
2152 if (!vma_is_anonymous(walk
->vma
))
2153 return migrate_vma_collect_skip(start
, end
, walk
);
2155 for (addr
= start
; addr
< end
; addr
+= PAGE_SIZE
) {
2156 migrate
->src
[migrate
->npages
] = MIGRATE_PFN_MIGRATE
;
2157 migrate
->dst
[migrate
->npages
] = 0;
2165 static int migrate_vma_collect_pmd(pmd_t
*pmdp
,
2166 unsigned long start
,
2168 struct mm_walk
*walk
)
2170 struct migrate_vma
*migrate
= walk
->private;
2171 struct vm_area_struct
*vma
= walk
->vma
;
2172 struct mm_struct
*mm
= vma
->vm_mm
;
2173 unsigned long addr
= start
, unmapped
= 0;
2178 if (pmd_none(*pmdp
))
2179 return migrate_vma_collect_hole(start
, end
, -1, walk
);
2181 if (pmd_trans_huge(*pmdp
)) {
2184 ptl
= pmd_lock(mm
, pmdp
);
2185 if (unlikely(!pmd_trans_huge(*pmdp
))) {
2190 page
= pmd_page(*pmdp
);
2191 if (is_huge_zero_page(page
)) {
2193 split_huge_pmd(vma
, pmdp
, addr
);
2194 if (pmd_trans_unstable(pmdp
))
2195 return migrate_vma_collect_skip(start
, end
,
2202 if (unlikely(!trylock_page(page
)))
2203 return migrate_vma_collect_skip(start
, end
,
2205 ret
= split_huge_page(page
);
2209 return migrate_vma_collect_skip(start
, end
,
2211 if (pmd_none(*pmdp
))
2212 return migrate_vma_collect_hole(start
, end
, -1,
2217 if (unlikely(pmd_bad(*pmdp
)))
2218 return migrate_vma_collect_skip(start
, end
, walk
);
2220 ptep
= pte_offset_map_lock(mm
, pmdp
, addr
, &ptl
);
2221 arch_enter_lazy_mmu_mode();
2223 for (; addr
< end
; addr
+= PAGE_SIZE
, ptep
++) {
2224 unsigned long mpfn
= 0, pfn
;
2231 if (pte_none(pte
)) {
2232 if (vma_is_anonymous(vma
)) {
2233 mpfn
= MIGRATE_PFN_MIGRATE
;
2239 if (!pte_present(pte
)) {
2241 * Only care about unaddressable device page special
2242 * page table entry. Other special swap entries are not
2243 * migratable, and we ignore regular swapped page.
2245 entry
= pte_to_swp_entry(pte
);
2246 if (!is_device_private_entry(entry
))
2249 page
= pfn_swap_entry_to_page(entry
);
2250 if (!(migrate
->flags
&
2251 MIGRATE_VMA_SELECT_DEVICE_PRIVATE
) ||
2252 page
->pgmap
->owner
!= migrate
->pgmap_owner
)
2255 mpfn
= migrate_pfn(page_to_pfn(page
)) |
2256 MIGRATE_PFN_MIGRATE
;
2257 if (is_writable_device_private_entry(entry
))
2258 mpfn
|= MIGRATE_PFN_WRITE
;
2260 if (!(migrate
->flags
& MIGRATE_VMA_SELECT_SYSTEM
))
2263 if (is_zero_pfn(pfn
)) {
2264 mpfn
= MIGRATE_PFN_MIGRATE
;
2268 page
= vm_normal_page(migrate
->vma
, addr
, pte
);
2269 mpfn
= migrate_pfn(pfn
) | MIGRATE_PFN_MIGRATE
;
2270 mpfn
|= pte_write(pte
) ? MIGRATE_PFN_WRITE
: 0;
2273 /* FIXME support THP */
2274 if (!page
|| !page
->mapping
|| PageTransCompound(page
)) {
2280 * By getting a reference on the page we pin it and that blocks
2281 * any kind of migration. Side effect is that it "freezes" the
2284 * We drop this reference after isolating the page from the lru
2285 * for non device page (device page are not on the lru and thus
2286 * can't be dropped from it).
2292 * Optimize for the common case where page is only mapped once
2293 * in one process. If we can lock the page, then we can safely
2294 * set up a special migration page table entry now.
2296 if (trylock_page(page
)) {
2299 mpfn
|= MIGRATE_PFN_LOCKED
;
2300 ptep_get_and_clear(mm
, addr
, ptep
);
2302 /* Setup special migration page table entry */
2303 if (mpfn
& MIGRATE_PFN_WRITE
)
2304 entry
= make_writable_migration_entry(
2307 entry
= make_readable_migration_entry(
2309 swp_pte
= swp_entry_to_pte(entry
);
2310 if (pte_present(pte
)) {
2311 if (pte_soft_dirty(pte
))
2312 swp_pte
= pte_swp_mksoft_dirty(swp_pte
);
2313 if (pte_uffd_wp(pte
))
2314 swp_pte
= pte_swp_mkuffd_wp(swp_pte
);
2316 if (pte_swp_soft_dirty(pte
))
2317 swp_pte
= pte_swp_mksoft_dirty(swp_pte
);
2318 if (pte_swp_uffd_wp(pte
))
2319 swp_pte
= pte_swp_mkuffd_wp(swp_pte
);
2321 set_pte_at(mm
, addr
, ptep
, swp_pte
);
2324 * This is like regular unmap: we remove the rmap and
2325 * drop page refcount. Page won't be freed, as we took
2326 * a reference just above.
2328 page_remove_rmap(page
, false);
2331 if (pte_present(pte
))
2336 migrate
->dst
[migrate
->npages
] = 0;
2337 migrate
->src
[migrate
->npages
++] = mpfn
;
2339 arch_leave_lazy_mmu_mode();
2340 pte_unmap_unlock(ptep
- 1, ptl
);
2342 /* Only flush the TLB if we actually modified any entries */
2344 flush_tlb_range(walk
->vma
, start
, end
);
2349 static const struct mm_walk_ops migrate_vma_walk_ops
= {
2350 .pmd_entry
= migrate_vma_collect_pmd
,
2351 .pte_hole
= migrate_vma_collect_hole
,
2355 * migrate_vma_collect() - collect pages over a range of virtual addresses
2356 * @migrate: migrate struct containing all migration information
2358 * This will walk the CPU page table. For each virtual address backed by a
2359 * valid page, it updates the src array and takes a reference on the page, in
2360 * order to pin the page until we lock it and unmap it.
2362 static void migrate_vma_collect(struct migrate_vma
*migrate
)
2364 struct mmu_notifier_range range
;
2367 * Note that the pgmap_owner is passed to the mmu notifier callback so
2368 * that the registered device driver can skip invalidating device
2369 * private page mappings that won't be migrated.
2371 mmu_notifier_range_init_owner(&range
, MMU_NOTIFY_MIGRATE
, 0,
2372 migrate
->vma
, migrate
->vma
->vm_mm
, migrate
->start
, migrate
->end
,
2373 migrate
->pgmap_owner
);
2374 mmu_notifier_invalidate_range_start(&range
);
2376 walk_page_range(migrate
->vma
->vm_mm
, migrate
->start
, migrate
->end
,
2377 &migrate_vma_walk_ops
, migrate
);
2379 mmu_notifier_invalidate_range_end(&range
);
2380 migrate
->end
= migrate
->start
+ (migrate
->npages
<< PAGE_SHIFT
);
2384 * migrate_vma_check_page() - check if page is pinned or not
2385 * @page: struct page to check
2387 * Pinned pages cannot be migrated. This is the same test as in
2388 * migrate_page_move_mapping(), except that here we allow migration of a
2391 static bool migrate_vma_check_page(struct page
*page
)
2394 * One extra ref because caller holds an extra reference, either from
2395 * isolate_lru_page() for a regular page, or migrate_vma_collect() for
2401 * FIXME support THP (transparent huge page), it is bit more complex to
2402 * check them than regular pages, because they can be mapped with a pmd
2403 * or with a pte (split pte mapping).
2405 if (PageCompound(page
))
2408 /* Page from ZONE_DEVICE have one extra reference */
2409 if (is_zone_device_page(page
)) {
2411 * Private page can never be pin as they have no valid pte and
2412 * GUP will fail for those. Yet if there is a pending migration
2413 * a thread might try to wait on the pte migration entry and
2414 * will bump the page reference count. Sadly there is no way to
2415 * differentiate a regular pin from migration wait. Hence to
2416 * avoid 2 racing thread trying to migrate back to CPU to enter
2417 * infinite loop (one stopping migration because the other is
2418 * waiting on pte migration entry). We always return true here.
2420 * FIXME proper solution is to rework migration_entry_wait() so
2421 * it does not need to take a reference on page.
2423 return is_device_private_page(page
);
2426 /* For file back page */
2427 if (page_mapping(page
))
2428 extra
+= 1 + page_has_private(page
);
2430 if ((page_count(page
) - extra
) > page_mapcount(page
))
2437 * migrate_vma_prepare() - lock pages and isolate them from the lru
2438 * @migrate: migrate struct containing all migration information
2440 * This locks pages that have been collected by migrate_vma_collect(). Once each
2441 * page is locked it is isolated from the lru (for non-device pages). Finally,
2442 * the ref taken by migrate_vma_collect() is dropped, as locked pages cannot be
2443 * migrated by concurrent kernel threads.
2445 static void migrate_vma_prepare(struct migrate_vma
*migrate
)
2447 const unsigned long npages
= migrate
->npages
;
2448 const unsigned long start
= migrate
->start
;
2449 unsigned long addr
, i
, restore
= 0;
2450 bool allow_drain
= true;
2454 for (i
= 0; (i
< npages
) && migrate
->cpages
; i
++) {
2455 struct page
*page
= migrate_pfn_to_page(migrate
->src
[i
]);
2461 if (!(migrate
->src
[i
] & MIGRATE_PFN_LOCKED
)) {
2463 * Because we are migrating several pages there can be
2464 * a deadlock between 2 concurrent migration where each
2465 * are waiting on each other page lock.
2467 * Make migrate_vma() a best effort thing and backoff
2468 * for any page we can not lock right away.
2470 if (!trylock_page(page
)) {
2471 migrate
->src
[i
] = 0;
2477 migrate
->src
[i
] |= MIGRATE_PFN_LOCKED
;
2480 /* ZONE_DEVICE pages are not on LRU */
2481 if (!is_zone_device_page(page
)) {
2482 if (!PageLRU(page
) && allow_drain
) {
2483 /* Drain CPU's pagevec */
2484 lru_add_drain_all();
2485 allow_drain
= false;
2488 if (isolate_lru_page(page
)) {
2490 migrate
->src
[i
] &= ~MIGRATE_PFN_MIGRATE
;
2494 migrate
->src
[i
] = 0;
2502 /* Drop the reference we took in collect */
2506 if (!migrate_vma_check_page(page
)) {
2508 migrate
->src
[i
] &= ~MIGRATE_PFN_MIGRATE
;
2512 if (!is_zone_device_page(page
)) {
2514 putback_lru_page(page
);
2517 migrate
->src
[i
] = 0;
2521 if (!is_zone_device_page(page
))
2522 putback_lru_page(page
);
2529 for (i
= 0, addr
= start
; i
< npages
&& restore
; i
++, addr
+= PAGE_SIZE
) {
2530 struct page
*page
= migrate_pfn_to_page(migrate
->src
[i
]);
2532 if (!page
|| (migrate
->src
[i
] & MIGRATE_PFN_MIGRATE
))
2535 remove_migration_pte(page
, migrate
->vma
, addr
, page
);
2537 migrate
->src
[i
] = 0;
2545 * migrate_vma_unmap() - replace page mapping with special migration pte entry
2546 * @migrate: migrate struct containing all migration information
2548 * Replace page mapping (CPU page table pte) with a special migration pte entry
2549 * and check again if it has been pinned. Pinned pages are restored because we
2550 * cannot migrate them.
2552 * This is the last step before we call the device driver callback to allocate
2553 * destination memory and copy contents of original page over to new page.
2555 static void migrate_vma_unmap(struct migrate_vma
*migrate
)
2557 const unsigned long npages
= migrate
->npages
;
2558 const unsigned long start
= migrate
->start
;
2559 unsigned long addr
, i
, restore
= 0;
2561 for (i
= 0; i
< npages
; i
++) {
2562 struct page
*page
= migrate_pfn_to_page(migrate
->src
[i
]);
2564 if (!page
|| !(migrate
->src
[i
] & MIGRATE_PFN_MIGRATE
))
2567 if (page_mapped(page
)) {
2568 try_to_migrate(page
, 0);
2569 if (page_mapped(page
))
2573 if (migrate_vma_check_page(page
))
2577 migrate
->src
[i
] &= ~MIGRATE_PFN_MIGRATE
;
2582 for (addr
= start
, i
= 0; i
< npages
&& restore
; addr
+= PAGE_SIZE
, i
++) {
2583 struct page
*page
= migrate_pfn_to_page(migrate
->src
[i
]);
2585 if (!page
|| (migrate
->src
[i
] & MIGRATE_PFN_MIGRATE
))
2588 remove_migration_ptes(page
, page
, false);
2590 migrate
->src
[i
] = 0;
2594 if (is_zone_device_page(page
))
2597 putback_lru_page(page
);
2602 * migrate_vma_setup() - prepare to migrate a range of memory
2603 * @args: contains the vma, start, and pfns arrays for the migration
2605 * Returns: negative errno on failures, 0 when 0 or more pages were migrated
2608 * Prepare to migrate a range of memory virtual address range by collecting all
2609 * the pages backing each virtual address in the range, saving them inside the
2610 * src array. Then lock those pages and unmap them. Once the pages are locked
2611 * and unmapped, check whether each page is pinned or not. Pages that aren't
2612 * pinned have the MIGRATE_PFN_MIGRATE flag set (by this function) in the
2613 * corresponding src array entry. Then restores any pages that are pinned, by
2614 * remapping and unlocking those pages.
2616 * The caller should then allocate destination memory and copy source memory to
2617 * it for all those entries (ie with MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE
2618 * flag set). Once these are allocated and copied, the caller must update each
2619 * corresponding entry in the dst array with the pfn value of the destination
2620 * page and with the MIGRATE_PFN_VALID and MIGRATE_PFN_LOCKED flags set
2621 * (destination pages must have their struct pages locked, via lock_page()).
2623 * Note that the caller does not have to migrate all the pages that are marked
2624 * with MIGRATE_PFN_MIGRATE flag in src array unless this is a migration from
2625 * device memory to system memory. If the caller cannot migrate a device page
2626 * back to system memory, then it must return VM_FAULT_SIGBUS, which has severe
2627 * consequences for the userspace process, so it must be avoided if at all
2630 * For empty entries inside CPU page table (pte_none() or pmd_none() is true) we
2631 * do set MIGRATE_PFN_MIGRATE flag inside the corresponding source array thus
2632 * allowing the caller to allocate device memory for those unbacked virtual
2633 * addresses. For this the caller simply has to allocate device memory and
2634 * properly set the destination entry like for regular migration. Note that
2635 * this can still fail, and thus inside the device driver you must check if the
2636 * migration was successful for those entries after calling migrate_vma_pages(),
2637 * just like for regular migration.
2639 * After that, the callers must call migrate_vma_pages() to go over each entry
2640 * in the src array that has the MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE flag
2641 * set. If the corresponding entry in dst array has MIGRATE_PFN_VALID flag set,
2642 * then migrate_vma_pages() to migrate struct page information from the source
2643 * struct page to the destination struct page. If it fails to migrate the
2644 * struct page information, then it clears the MIGRATE_PFN_MIGRATE flag in the
2647 * At this point all successfully migrated pages have an entry in the src
2648 * array with MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE flag set and the dst
2649 * array entry with MIGRATE_PFN_VALID flag set.
2651 * Once migrate_vma_pages() returns the caller may inspect which pages were
2652 * successfully migrated, and which were not. Successfully migrated pages will
2653 * have the MIGRATE_PFN_MIGRATE flag set for their src array entry.
2655 * It is safe to update device page table after migrate_vma_pages() because
2656 * both destination and source page are still locked, and the mmap_lock is held
2657 * in read mode (hence no one can unmap the range being migrated).
2659 * Once the caller is done cleaning up things and updating its page table (if it
2660 * chose to do so, this is not an obligation) it finally calls
2661 * migrate_vma_finalize() to update the CPU page table to point to new pages
2662 * for successfully migrated pages or otherwise restore the CPU page table to
2663 * point to the original source pages.
2665 int migrate_vma_setup(struct migrate_vma
*args
)
2667 long nr_pages
= (args
->end
- args
->start
) >> PAGE_SHIFT
;
2669 args
->start
&= PAGE_MASK
;
2670 args
->end
&= PAGE_MASK
;
2671 if (!args
->vma
|| is_vm_hugetlb_page(args
->vma
) ||
2672 (args
->vma
->vm_flags
& VM_SPECIAL
) || vma_is_dax(args
->vma
))
2676 if (args
->start
< args
->vma
->vm_start
||
2677 args
->start
>= args
->vma
->vm_end
)
2679 if (args
->end
<= args
->vma
->vm_start
|| args
->end
> args
->vma
->vm_end
)
2681 if (!args
->src
|| !args
->dst
)
2684 memset(args
->src
, 0, sizeof(*args
->src
) * nr_pages
);
2688 migrate_vma_collect(args
);
2691 migrate_vma_prepare(args
);
2693 migrate_vma_unmap(args
);
2696 * At this point pages are locked and unmapped, and thus they have
2697 * stable content and can safely be copied to destination memory that
2698 * is allocated by the drivers.
2703 EXPORT_SYMBOL(migrate_vma_setup
);
2706 * This code closely matches the code in:
2707 * __handle_mm_fault()
2708 * handle_pte_fault()
2709 * do_anonymous_page()
2710 * to map in an anonymous zero page but the struct page will be a ZONE_DEVICE
2713 static void migrate_vma_insert_page(struct migrate_vma
*migrate
,
2718 struct vm_area_struct
*vma
= migrate
->vma
;
2719 struct mm_struct
*mm
= vma
->vm_mm
;
2729 /* Only allow populating anonymous memory */
2730 if (!vma_is_anonymous(vma
))
2733 pgdp
= pgd_offset(mm
, addr
);
2734 p4dp
= p4d_alloc(mm
, pgdp
, addr
);
2737 pudp
= pud_alloc(mm
, p4dp
, addr
);
2740 pmdp
= pmd_alloc(mm
, pudp
, addr
);
2744 if (pmd_trans_huge(*pmdp
) || pmd_devmap(*pmdp
))
2748 * Use pte_alloc() instead of pte_alloc_map(). We can't run
2749 * pte_offset_map() on pmds where a huge pmd might be created
2750 * from a different thread.
2752 * pte_alloc_map() is safe to use under mmap_write_lock(mm) or when
2753 * parallel threads are excluded by other means.
2755 * Here we only have mmap_read_lock(mm).
2757 if (pte_alloc(mm
, pmdp
))
2760 /* See the comment in pte_alloc_one_map() */
2761 if (unlikely(pmd_trans_unstable(pmdp
)))
2764 if (unlikely(anon_vma_prepare(vma
)))
2766 if (mem_cgroup_charge(page
, vma
->vm_mm
, GFP_KERNEL
))
2770 * The memory barrier inside __SetPageUptodate makes sure that
2771 * preceding stores to the page contents become visible before
2772 * the set_pte_at() write.
2774 __SetPageUptodate(page
);
2776 if (is_zone_device_page(page
)) {
2777 if (is_device_private_page(page
)) {
2778 swp_entry_t swp_entry
;
2780 if (vma
->vm_flags
& VM_WRITE
)
2781 swp_entry
= make_writable_device_private_entry(
2784 swp_entry
= make_readable_device_private_entry(
2786 entry
= swp_entry_to_pte(swp_entry
);
2789 * For now we only support migrating to un-addressable
2792 pr_warn_once("Unsupported ZONE_DEVICE page type.\n");
2796 entry
= mk_pte(page
, vma
->vm_page_prot
);
2797 if (vma
->vm_flags
& VM_WRITE
)
2798 entry
= pte_mkwrite(pte_mkdirty(entry
));
2801 ptep
= pte_offset_map_lock(mm
, pmdp
, addr
, &ptl
);
2803 if (check_stable_address_space(mm
))
2806 if (pte_present(*ptep
)) {
2807 unsigned long pfn
= pte_pfn(*ptep
);
2809 if (!is_zero_pfn(pfn
))
2812 } else if (!pte_none(*ptep
))
2816 * Check for userfaultfd but do not deliver the fault. Instead,
2819 if (userfaultfd_missing(vma
))
2822 inc_mm_counter(mm
, MM_ANONPAGES
);
2823 page_add_new_anon_rmap(page
, vma
, addr
, false);
2824 if (!is_zone_device_page(page
))
2825 lru_cache_add_inactive_or_unevictable(page
, vma
);
2829 flush_cache_page(vma
, addr
, pte_pfn(*ptep
));
2830 ptep_clear_flush_notify(vma
, addr
, ptep
);
2831 set_pte_at_notify(mm
, addr
, ptep
, entry
);
2832 update_mmu_cache(vma
, addr
, ptep
);
2834 /* No need to invalidate - it was non-present before */
2835 set_pte_at(mm
, addr
, ptep
, entry
);
2836 update_mmu_cache(vma
, addr
, ptep
);
2839 pte_unmap_unlock(ptep
, ptl
);
2840 *src
= MIGRATE_PFN_MIGRATE
;
2844 pte_unmap_unlock(ptep
, ptl
);
2846 *src
&= ~MIGRATE_PFN_MIGRATE
;
2850 * migrate_vma_pages() - migrate meta-data from src page to dst page
2851 * @migrate: migrate struct containing all migration information
2853 * This migrates struct page meta-data from source struct page to destination
2854 * struct page. This effectively finishes the migration from source page to the
2857 void migrate_vma_pages(struct migrate_vma
*migrate
)
2859 const unsigned long npages
= migrate
->npages
;
2860 const unsigned long start
= migrate
->start
;
2861 struct mmu_notifier_range range
;
2862 unsigned long addr
, i
;
2863 bool notified
= false;
2865 for (i
= 0, addr
= start
; i
< npages
; addr
+= PAGE_SIZE
, i
++) {
2866 struct page
*newpage
= migrate_pfn_to_page(migrate
->dst
[i
]);
2867 struct page
*page
= migrate_pfn_to_page(migrate
->src
[i
]);
2868 struct address_space
*mapping
;
2872 migrate
->src
[i
] &= ~MIGRATE_PFN_MIGRATE
;
2877 if (!(migrate
->src
[i
] & MIGRATE_PFN_MIGRATE
))
2882 mmu_notifier_range_init_owner(&range
,
2883 MMU_NOTIFY_MIGRATE
, 0, migrate
->vma
,
2884 migrate
->vma
->vm_mm
, addr
, migrate
->end
,
2885 migrate
->pgmap_owner
);
2886 mmu_notifier_invalidate_range_start(&range
);
2888 migrate_vma_insert_page(migrate
, addr
, newpage
,
2893 mapping
= page_mapping(page
);
2895 if (is_zone_device_page(newpage
)) {
2896 if (is_device_private_page(newpage
)) {
2898 * For now only support private anonymous when
2899 * migrating to un-addressable device memory.
2902 migrate
->src
[i
] &= ~MIGRATE_PFN_MIGRATE
;
2907 * Other types of ZONE_DEVICE page are not
2910 migrate
->src
[i
] &= ~MIGRATE_PFN_MIGRATE
;
2915 r
= migrate_page(mapping
, newpage
, page
, MIGRATE_SYNC_NO_COPY
);
2916 if (r
!= MIGRATEPAGE_SUCCESS
)
2917 migrate
->src
[i
] &= ~MIGRATE_PFN_MIGRATE
;
2921 * No need to double call mmu_notifier->invalidate_range() callback as
2922 * the above ptep_clear_flush_notify() inside migrate_vma_insert_page()
2923 * did already call it.
2926 mmu_notifier_invalidate_range_only_end(&range
);
2928 EXPORT_SYMBOL(migrate_vma_pages
);
2931 * migrate_vma_finalize() - restore CPU page table entry
2932 * @migrate: migrate struct containing all migration information
2934 * This replaces the special migration pte entry with either a mapping to the
2935 * new page if migration was successful for that page, or to the original page
2938 * This also unlocks the pages and puts them back on the lru, or drops the extra
2939 * refcount, for device pages.
2941 void migrate_vma_finalize(struct migrate_vma
*migrate
)
2943 const unsigned long npages
= migrate
->npages
;
2946 for (i
= 0; i
< npages
; i
++) {
2947 struct page
*newpage
= migrate_pfn_to_page(migrate
->dst
[i
]);
2948 struct page
*page
= migrate_pfn_to_page(migrate
->src
[i
]);
2952 unlock_page(newpage
);
2958 if (!(migrate
->src
[i
] & MIGRATE_PFN_MIGRATE
) || !newpage
) {
2960 unlock_page(newpage
);
2966 remove_migration_ptes(page
, newpage
, false);
2969 if (is_zone_device_page(page
))
2972 putback_lru_page(page
);
2974 if (newpage
!= page
) {
2975 unlock_page(newpage
);
2976 if (is_zone_device_page(newpage
))
2979 putback_lru_page(newpage
);
2983 EXPORT_SYMBOL(migrate_vma_finalize
);
2984 #endif /* CONFIG_DEVICE_PRIVATE */