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_write_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))) {
219 entry
= make_device_private_entry(new, pte_write(pte
));
220 pte
= swp_entry_to_pte(entry
);
221 if (pte_swp_soft_dirty(*pvmw
.pte
))
222 pte
= pte_swp_mksoft_dirty(pte
);
223 if (pte_swp_uffd_wp(*pvmw
.pte
))
224 pte
= pte_swp_mkuffd_wp(pte
);
227 #ifdef CONFIG_HUGETLB_PAGE
229 pte
= pte_mkhuge(pte
);
230 pte
= arch_make_huge_pte(pte
, vma
, new, 0);
231 set_huge_pte_at(vma
->vm_mm
, pvmw
.address
, pvmw
.pte
, pte
);
233 hugepage_add_anon_rmap(new, vma
, pvmw
.address
);
235 page_dup_rmap(new, true);
239 set_pte_at(vma
->vm_mm
, pvmw
.address
, pvmw
.pte
, pte
);
242 page_add_anon_rmap(new, vma
, pvmw
.address
, false);
244 page_add_file_rmap(new, false);
246 if (vma
->vm_flags
& VM_LOCKED
&& !PageTransCompound(new))
249 if (PageTransHuge(page
) && PageMlocked(page
))
250 clear_page_mlock(page
);
252 /* No need to invalidate - it was non-present before */
253 update_mmu_cache(vma
, pvmw
.address
, pvmw
.pte
);
260 * Get rid of all migration entries and replace them by
261 * references to the indicated page.
263 void remove_migration_ptes(struct page
*old
, struct page
*new, bool locked
)
265 struct rmap_walk_control rwc
= {
266 .rmap_one
= remove_migration_pte
,
271 rmap_walk_locked(new, &rwc
);
273 rmap_walk(new, &rwc
);
277 * Something used the pte of a page under migration. We need to
278 * get to the page and wait until migration is finished.
279 * When we return from this function the fault will be retried.
281 void __migration_entry_wait(struct mm_struct
*mm
, pte_t
*ptep
,
290 if (!is_swap_pte(pte
))
293 entry
= pte_to_swp_entry(pte
);
294 if (!is_migration_entry(entry
))
297 page
= migration_entry_to_page(entry
);
300 * Once page cache replacement of page migration started, page_count
301 * is zero; but we must not call put_and_wait_on_page_locked() without
302 * a ref. Use get_page_unless_zero(), and just fault again if it fails.
304 if (!get_page_unless_zero(page
))
306 pte_unmap_unlock(ptep
, ptl
);
307 put_and_wait_on_page_locked(page
, TASK_UNINTERRUPTIBLE
);
310 pte_unmap_unlock(ptep
, ptl
);
313 void migration_entry_wait(struct mm_struct
*mm
, pmd_t
*pmd
,
314 unsigned long address
)
316 spinlock_t
*ptl
= pte_lockptr(mm
, pmd
);
317 pte_t
*ptep
= pte_offset_map(pmd
, address
);
318 __migration_entry_wait(mm
, ptep
, ptl
);
321 void migration_entry_wait_huge(struct vm_area_struct
*vma
,
322 struct mm_struct
*mm
, pte_t
*pte
)
324 spinlock_t
*ptl
= huge_pte_lockptr(hstate_vma(vma
), mm
, pte
);
325 __migration_entry_wait(mm
, pte
, ptl
);
328 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
329 void pmd_migration_entry_wait(struct mm_struct
*mm
, pmd_t
*pmd
)
334 ptl
= pmd_lock(mm
, pmd
);
335 if (!is_pmd_migration_entry(*pmd
))
337 page
= migration_entry_to_page(pmd_to_swp_entry(*pmd
));
338 if (!get_page_unless_zero(page
))
341 put_and_wait_on_page_locked(page
, TASK_UNINTERRUPTIBLE
);
348 static int expected_page_refs(struct address_space
*mapping
, struct page
*page
)
350 int expected_count
= 1;
353 * Device private pages have an extra refcount as they are
356 expected_count
+= is_device_private_page(page
);
358 expected_count
+= thp_nr_pages(page
) + page_has_private(page
);
360 return expected_count
;
364 * Replace the page in the mapping.
366 * The number of remaining references must be:
367 * 1 for anonymous pages without a mapping
368 * 2 for pages with a mapping
369 * 3 for pages with a mapping and PagePrivate/PagePrivate2 set.
371 int migrate_page_move_mapping(struct address_space
*mapping
,
372 struct page
*newpage
, struct page
*page
, int extra_count
)
374 XA_STATE(xas
, &mapping
->i_pages
, page_index(page
));
375 struct zone
*oldzone
, *newzone
;
377 int expected_count
= expected_page_refs(mapping
, page
) + extra_count
;
378 int nr
= thp_nr_pages(page
);
381 /* Anonymous page without mapping */
382 if (page_count(page
) != expected_count
)
385 /* No turning back from here */
386 newpage
->index
= page
->index
;
387 newpage
->mapping
= page
->mapping
;
388 if (PageSwapBacked(page
))
389 __SetPageSwapBacked(newpage
);
391 return MIGRATEPAGE_SUCCESS
;
394 oldzone
= page_zone(page
);
395 newzone
= page_zone(newpage
);
398 if (page_count(page
) != expected_count
|| xas_load(&xas
) != page
) {
399 xas_unlock_irq(&xas
);
403 if (!page_ref_freeze(page
, expected_count
)) {
404 xas_unlock_irq(&xas
);
409 * Now we know that no one else is looking at the page:
410 * no turning back from here.
412 newpage
->index
= page
->index
;
413 newpage
->mapping
= page
->mapping
;
414 page_ref_add(newpage
, nr
); /* add cache reference */
415 if (PageSwapBacked(page
)) {
416 __SetPageSwapBacked(newpage
);
417 if (PageSwapCache(page
)) {
418 SetPageSwapCache(newpage
);
419 set_page_private(newpage
, page_private(page
));
422 VM_BUG_ON_PAGE(PageSwapCache(page
), page
);
425 /* Move dirty while page refs frozen and newpage not yet exposed */
426 dirty
= PageDirty(page
);
428 ClearPageDirty(page
);
429 SetPageDirty(newpage
);
432 xas_store(&xas
, newpage
);
433 if (PageTransHuge(page
)) {
436 for (i
= 1; i
< nr
; i
++) {
438 xas_store(&xas
, newpage
);
443 * Drop cache reference from old page by unfreezing
444 * to one less reference.
445 * We know this isn't the last reference.
447 page_ref_unfreeze(page
, expected_count
- nr
);
450 /* Leave irq disabled to prevent preemption while updating stats */
453 * If moved to a different zone then also account
454 * the page for that zone. Other VM counters will be
455 * taken care of when we establish references to the
456 * new page and drop references to the old page.
458 * Note that anonymous pages are accounted for
459 * via NR_FILE_PAGES and NR_ANON_MAPPED if they
460 * are mapped to swap space.
462 if (newzone
!= oldzone
) {
463 struct lruvec
*old_lruvec
, *new_lruvec
;
464 struct mem_cgroup
*memcg
;
466 memcg
= page_memcg(page
);
467 old_lruvec
= mem_cgroup_lruvec(memcg
, oldzone
->zone_pgdat
);
468 new_lruvec
= mem_cgroup_lruvec(memcg
, newzone
->zone_pgdat
);
470 __mod_lruvec_state(old_lruvec
, NR_FILE_PAGES
, -nr
);
471 __mod_lruvec_state(new_lruvec
, NR_FILE_PAGES
, nr
);
472 if (PageSwapBacked(page
) && !PageSwapCache(page
)) {
473 __mod_lruvec_state(old_lruvec
, NR_SHMEM
, -nr
);
474 __mod_lruvec_state(new_lruvec
, NR_SHMEM
, nr
);
477 if (PageSwapCache(page
)) {
478 __mod_lruvec_state(old_lruvec
, NR_SWAPCACHE
, -nr
);
479 __mod_lruvec_state(new_lruvec
, NR_SWAPCACHE
, nr
);
482 if (dirty
&& mapping_can_writeback(mapping
)) {
483 __mod_lruvec_state(old_lruvec
, NR_FILE_DIRTY
, -nr
);
484 __mod_zone_page_state(oldzone
, NR_ZONE_WRITE_PENDING
, -nr
);
485 __mod_lruvec_state(new_lruvec
, NR_FILE_DIRTY
, nr
);
486 __mod_zone_page_state(newzone
, NR_ZONE_WRITE_PENDING
, nr
);
491 return MIGRATEPAGE_SUCCESS
;
493 EXPORT_SYMBOL(migrate_page_move_mapping
);
496 * The expected number of remaining references is the same as that
497 * of migrate_page_move_mapping().
499 int migrate_huge_page_move_mapping(struct address_space
*mapping
,
500 struct page
*newpage
, struct page
*page
)
502 XA_STATE(xas
, &mapping
->i_pages
, page_index(page
));
506 expected_count
= 2 + page_has_private(page
);
507 if (page_count(page
) != expected_count
|| xas_load(&xas
) != page
) {
508 xas_unlock_irq(&xas
);
512 if (!page_ref_freeze(page
, expected_count
)) {
513 xas_unlock_irq(&xas
);
517 newpage
->index
= page
->index
;
518 newpage
->mapping
= page
->mapping
;
522 xas_store(&xas
, newpage
);
524 page_ref_unfreeze(page
, expected_count
- 1);
526 xas_unlock_irq(&xas
);
528 return MIGRATEPAGE_SUCCESS
;
532 * Gigantic pages are so large that we do not guarantee that page++ pointer
533 * arithmetic will work across the entire page. We need something more
536 static void __copy_gigantic_page(struct page
*dst
, struct page
*src
,
540 struct page
*dst_base
= dst
;
541 struct page
*src_base
= src
;
543 for (i
= 0; i
< nr_pages
; ) {
545 copy_highpage(dst
, src
);
548 dst
= mem_map_next(dst
, dst_base
, i
);
549 src
= mem_map_next(src
, src_base
, i
);
553 static void copy_huge_page(struct page
*dst
, struct page
*src
)
560 struct hstate
*h
= page_hstate(src
);
561 nr_pages
= pages_per_huge_page(h
);
563 if (unlikely(nr_pages
> MAX_ORDER_NR_PAGES
)) {
564 __copy_gigantic_page(dst
, src
, nr_pages
);
569 BUG_ON(!PageTransHuge(src
));
570 nr_pages
= thp_nr_pages(src
);
573 for (i
= 0; i
< nr_pages
; i
++) {
575 copy_highpage(dst
+ i
, src
+ i
);
580 * Copy the page to its new location
582 void migrate_page_states(struct page
*newpage
, struct page
*page
)
587 SetPageError(newpage
);
588 if (PageReferenced(page
))
589 SetPageReferenced(newpage
);
590 if (PageUptodate(page
))
591 SetPageUptodate(newpage
);
592 if (TestClearPageActive(page
)) {
593 VM_BUG_ON_PAGE(PageUnevictable(page
), page
);
594 SetPageActive(newpage
);
595 } else if (TestClearPageUnevictable(page
))
596 SetPageUnevictable(newpage
);
597 if (PageWorkingset(page
))
598 SetPageWorkingset(newpage
);
599 if (PageChecked(page
))
600 SetPageChecked(newpage
);
601 if (PageMappedToDisk(page
))
602 SetPageMappedToDisk(newpage
);
604 /* Move dirty on pages not done by migrate_page_move_mapping() */
606 SetPageDirty(newpage
);
608 if (page_is_young(page
))
609 set_page_young(newpage
);
610 if (page_is_idle(page
))
611 set_page_idle(newpage
);
614 * Copy NUMA information to the new page, to prevent over-eager
615 * future migrations of this same page.
617 cpupid
= page_cpupid_xchg_last(page
, -1);
618 page_cpupid_xchg_last(newpage
, cpupid
);
620 ksm_migrate_page(newpage
, page
);
622 * Please do not reorder this without considering how mm/ksm.c's
623 * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache().
625 if (PageSwapCache(page
))
626 ClearPageSwapCache(page
);
627 ClearPagePrivate(page
);
628 set_page_private(page
, 0);
631 * If any waiters have accumulated on the new page then
634 if (PageWriteback(newpage
))
635 end_page_writeback(newpage
);
638 * PG_readahead shares the same bit with PG_reclaim. The above
639 * end_page_writeback() may clear PG_readahead mistakenly, so set the
642 if (PageReadahead(page
))
643 SetPageReadahead(newpage
);
645 copy_page_owner(page
, newpage
);
648 mem_cgroup_migrate(page
, newpage
);
650 EXPORT_SYMBOL(migrate_page_states
);
652 void migrate_page_copy(struct page
*newpage
, struct page
*page
)
654 if (PageHuge(page
) || PageTransHuge(page
))
655 copy_huge_page(newpage
, page
);
657 copy_highpage(newpage
, page
);
659 migrate_page_states(newpage
, page
);
661 EXPORT_SYMBOL(migrate_page_copy
);
663 /************************************************************
664 * Migration functions
665 ***********************************************************/
668 * Common logic to directly migrate a single LRU page suitable for
669 * pages that do not use PagePrivate/PagePrivate2.
671 * Pages are locked upon entry and exit.
673 int migrate_page(struct address_space
*mapping
,
674 struct page
*newpage
, struct page
*page
,
675 enum migrate_mode mode
)
679 BUG_ON(PageWriteback(page
)); /* Writeback must be complete */
681 rc
= migrate_page_move_mapping(mapping
, newpage
, page
, 0);
683 if (rc
!= MIGRATEPAGE_SUCCESS
)
686 if (mode
!= MIGRATE_SYNC_NO_COPY
)
687 migrate_page_copy(newpage
, page
);
689 migrate_page_states(newpage
, page
);
690 return MIGRATEPAGE_SUCCESS
;
692 EXPORT_SYMBOL(migrate_page
);
695 /* Returns true if all buffers are successfully locked */
696 static bool buffer_migrate_lock_buffers(struct buffer_head
*head
,
697 enum migrate_mode mode
)
699 struct buffer_head
*bh
= head
;
701 /* Simple case, sync compaction */
702 if (mode
!= MIGRATE_ASYNC
) {
705 bh
= bh
->b_this_page
;
707 } while (bh
!= head
);
712 /* async case, we cannot block on lock_buffer so use trylock_buffer */
714 if (!trylock_buffer(bh
)) {
716 * We failed to lock the buffer and cannot stall in
717 * async migration. Release the taken locks
719 struct buffer_head
*failed_bh
= bh
;
721 while (bh
!= failed_bh
) {
723 bh
= bh
->b_this_page
;
728 bh
= bh
->b_this_page
;
729 } while (bh
!= head
);
733 static int __buffer_migrate_page(struct address_space
*mapping
,
734 struct page
*newpage
, struct page
*page
, enum migrate_mode mode
,
737 struct buffer_head
*bh
, *head
;
741 if (!page_has_buffers(page
))
742 return migrate_page(mapping
, newpage
, page
, mode
);
744 /* Check whether page does not have extra refs before we do more work */
745 expected_count
= expected_page_refs(mapping
, page
);
746 if (page_count(page
) != expected_count
)
749 head
= page_buffers(page
);
750 if (!buffer_migrate_lock_buffers(head
, mode
))
755 bool invalidated
= false;
759 spin_lock(&mapping
->private_lock
);
762 if (atomic_read(&bh
->b_count
)) {
766 bh
= bh
->b_this_page
;
767 } while (bh
!= head
);
773 spin_unlock(&mapping
->private_lock
);
774 invalidate_bh_lrus();
776 goto recheck_buffers
;
780 rc
= migrate_page_move_mapping(mapping
, newpage
, page
, 0);
781 if (rc
!= MIGRATEPAGE_SUCCESS
)
784 attach_page_private(newpage
, detach_page_private(page
));
788 set_bh_page(bh
, newpage
, bh_offset(bh
));
789 bh
= bh
->b_this_page
;
791 } while (bh
!= head
);
793 if (mode
!= MIGRATE_SYNC_NO_COPY
)
794 migrate_page_copy(newpage
, page
);
796 migrate_page_states(newpage
, page
);
798 rc
= MIGRATEPAGE_SUCCESS
;
801 spin_unlock(&mapping
->private_lock
);
805 bh
= bh
->b_this_page
;
807 } while (bh
!= head
);
813 * Migration function for pages with buffers. This function can only be used
814 * if the underlying filesystem guarantees that no other references to "page"
815 * exist. For example attached buffer heads are accessed only under page lock.
817 int buffer_migrate_page(struct address_space
*mapping
,
818 struct page
*newpage
, struct page
*page
, enum migrate_mode mode
)
820 return __buffer_migrate_page(mapping
, newpage
, page
, mode
, false);
822 EXPORT_SYMBOL(buffer_migrate_page
);
825 * Same as above except that this variant is more careful and checks that there
826 * are also no buffer head references. This function is the right one for
827 * mappings where buffer heads are directly looked up and referenced (such as
828 * block device mappings).
830 int buffer_migrate_page_norefs(struct address_space
*mapping
,
831 struct page
*newpage
, struct page
*page
, enum migrate_mode mode
)
833 return __buffer_migrate_page(mapping
, newpage
, page
, mode
, true);
838 * Writeback a page to clean the dirty state
840 static int writeout(struct address_space
*mapping
, struct page
*page
)
842 struct writeback_control wbc
= {
843 .sync_mode
= WB_SYNC_NONE
,
846 .range_end
= LLONG_MAX
,
851 if (!mapping
->a_ops
->writepage
)
852 /* No write method for the address space */
855 if (!clear_page_dirty_for_io(page
))
856 /* Someone else already triggered a write */
860 * A dirty page may imply that the underlying filesystem has
861 * the page on some queue. So the page must be clean for
862 * migration. Writeout may mean we loose the lock and the
863 * page state is no longer what we checked for earlier.
864 * At this point we know that the migration attempt cannot
867 remove_migration_ptes(page
, page
, false);
869 rc
= mapping
->a_ops
->writepage(page
, &wbc
);
871 if (rc
!= AOP_WRITEPAGE_ACTIVATE
)
872 /* unlocked. Relock */
875 return (rc
< 0) ? -EIO
: -EAGAIN
;
879 * Default handling if a filesystem does not provide a migration function.
881 static int fallback_migrate_page(struct address_space
*mapping
,
882 struct page
*newpage
, struct page
*page
, enum migrate_mode mode
)
884 if (PageDirty(page
)) {
885 /* Only writeback pages in full synchronous migration */
888 case MIGRATE_SYNC_NO_COPY
:
893 return writeout(mapping
, page
);
897 * Buffers may be managed in a filesystem specific way.
898 * We must have no buffers or drop them.
900 if (page_has_private(page
) &&
901 !try_to_release_page(page
, GFP_KERNEL
))
902 return mode
== MIGRATE_SYNC
? -EAGAIN
: -EBUSY
;
904 return migrate_page(mapping
, newpage
, page
, mode
);
908 * Move a page to a newly allocated page
909 * The page is locked and all ptes have been successfully removed.
911 * The new page will have replaced the old page if this function
916 * MIGRATEPAGE_SUCCESS - success
918 static int move_to_new_page(struct page
*newpage
, struct page
*page
,
919 enum migrate_mode mode
)
921 struct address_space
*mapping
;
923 bool is_lru
= !__PageMovable(page
);
925 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
926 VM_BUG_ON_PAGE(!PageLocked(newpage
), newpage
);
928 mapping
= page_mapping(page
);
930 if (likely(is_lru
)) {
932 rc
= migrate_page(mapping
, newpage
, page
, mode
);
933 else if (mapping
->a_ops
->migratepage
)
935 * Most pages have a mapping and most filesystems
936 * provide a migratepage callback. Anonymous pages
937 * are part of swap space which also has its own
938 * migratepage callback. This is the most common path
939 * for page migration.
941 rc
= mapping
->a_ops
->migratepage(mapping
, newpage
,
944 rc
= fallback_migrate_page(mapping
, newpage
,
948 * In case of non-lru page, it could be released after
949 * isolation step. In that case, we shouldn't try migration.
951 VM_BUG_ON_PAGE(!PageIsolated(page
), page
);
952 if (!PageMovable(page
)) {
953 rc
= MIGRATEPAGE_SUCCESS
;
954 __ClearPageIsolated(page
);
958 rc
= mapping
->a_ops
->migratepage(mapping
, newpage
,
960 WARN_ON_ONCE(rc
== MIGRATEPAGE_SUCCESS
&&
961 !PageIsolated(page
));
965 * When successful, old pagecache page->mapping must be cleared before
966 * page is freed; but stats require that PageAnon be left as PageAnon.
968 if (rc
== MIGRATEPAGE_SUCCESS
) {
969 if (__PageMovable(page
)) {
970 VM_BUG_ON_PAGE(!PageIsolated(page
), page
);
973 * We clear PG_movable under page_lock so any compactor
974 * cannot try to migrate this page.
976 __ClearPageIsolated(page
);
980 * Anonymous and movable page->mapping will be cleared by
981 * free_pages_prepare so don't reset it here for keeping
982 * the type to work PageAnon, for example.
984 if (!PageMappingFlags(page
))
985 page
->mapping
= NULL
;
987 if (likely(!is_zone_device_page(newpage
)))
988 flush_dcache_page(newpage
);
995 static int __unmap_and_move(struct page
*page
, struct page
*newpage
,
996 int force
, enum migrate_mode mode
)
999 int page_was_mapped
= 0;
1000 struct anon_vma
*anon_vma
= NULL
;
1001 bool is_lru
= !__PageMovable(page
);
1003 if (!trylock_page(page
)) {
1004 if (!force
|| mode
== MIGRATE_ASYNC
)
1008 * It's not safe for direct compaction to call lock_page.
1009 * For example, during page readahead pages are added locked
1010 * to the LRU. Later, when the IO completes the pages are
1011 * marked uptodate and unlocked. However, the queueing
1012 * could be merging multiple pages for one bio (e.g.
1013 * mpage_readahead). If an allocation happens for the
1014 * second or third page, the process can end up locking
1015 * the same page twice and deadlocking. Rather than
1016 * trying to be clever about what pages can be locked,
1017 * avoid the use of lock_page for direct compaction
1020 if (current
->flags
& PF_MEMALLOC
)
1026 if (PageWriteback(page
)) {
1028 * Only in the case of a full synchronous migration is it
1029 * necessary to wait for PageWriteback. In the async case,
1030 * the retry loop is too short and in the sync-light case,
1031 * the overhead of stalling is too much
1035 case MIGRATE_SYNC_NO_COPY
:
1043 wait_on_page_writeback(page
);
1047 * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
1048 * we cannot notice that anon_vma is freed while we migrates a page.
1049 * This get_anon_vma() delays freeing anon_vma pointer until the end
1050 * of migration. File cache pages are no problem because of page_lock()
1051 * File Caches may use write_page() or lock_page() in migration, then,
1052 * just care Anon page here.
1054 * Only page_get_anon_vma() understands the subtleties of
1055 * getting a hold on an anon_vma from outside one of its mms.
1056 * But if we cannot get anon_vma, then we won't need it anyway,
1057 * because that implies that the anon page is no longer mapped
1058 * (and cannot be remapped so long as we hold the page lock).
1060 if (PageAnon(page
) && !PageKsm(page
))
1061 anon_vma
= page_get_anon_vma(page
);
1064 * Block others from accessing the new page when we get around to
1065 * establishing additional references. We are usually the only one
1066 * holding a reference to newpage at this point. We used to have a BUG
1067 * here if trylock_page(newpage) fails, but would like to allow for
1068 * cases where there might be a race with the previous use of newpage.
1069 * This is much like races on refcount of oldpage: just don't BUG().
1071 if (unlikely(!trylock_page(newpage
)))
1074 if (unlikely(!is_lru
)) {
1075 rc
= move_to_new_page(newpage
, page
, mode
);
1076 goto out_unlock_both
;
1080 * Corner case handling:
1081 * 1. When a new swap-cache page is read into, it is added to the LRU
1082 * and treated as swapcache but it has no rmap yet.
1083 * Calling try_to_unmap() against a page->mapping==NULL page will
1084 * trigger a BUG. So handle it here.
1085 * 2. An orphaned page (see truncate_cleanup_page) might have
1086 * fs-private metadata. The page can be picked up due to memory
1087 * offlining. Everywhere else except page reclaim, the page is
1088 * invisible to the vm, so the page can not be migrated. So try to
1089 * free the metadata, so the page can be freed.
1091 if (!page
->mapping
) {
1092 VM_BUG_ON_PAGE(PageAnon(page
), page
);
1093 if (page_has_private(page
)) {
1094 try_to_free_buffers(page
);
1095 goto out_unlock_both
;
1097 } else if (page_mapped(page
)) {
1098 /* Establish migration ptes */
1099 VM_BUG_ON_PAGE(PageAnon(page
) && !PageKsm(page
) && !anon_vma
,
1101 try_to_unmap(page
, TTU_MIGRATION
|TTU_IGNORE_MLOCK
);
1102 page_was_mapped
= 1;
1105 if (!page_mapped(page
))
1106 rc
= move_to_new_page(newpage
, page
, mode
);
1108 if (page_was_mapped
)
1109 remove_migration_ptes(page
,
1110 rc
== MIGRATEPAGE_SUCCESS
? newpage
: page
, false);
1113 unlock_page(newpage
);
1115 /* Drop an anon_vma reference if we took one */
1117 put_anon_vma(anon_vma
);
1121 * If migration is successful, decrease refcount of the newpage
1122 * which will not free the page because new page owner increased
1123 * refcounter. As well, if it is LRU page, add the page to LRU
1124 * list in here. Use the old state of the isolated source page to
1125 * determine if we migrated a LRU page. newpage was already unlocked
1126 * and possibly modified by its owner - don't rely on the page
1129 if (rc
== MIGRATEPAGE_SUCCESS
) {
1130 if (unlikely(!is_lru
))
1133 putback_lru_page(newpage
);
1140 * Obtain the lock on page, remove all ptes and migrate the page
1141 * to the newly allocated page in newpage.
1143 static int unmap_and_move(new_page_t get_new_page
,
1144 free_page_t put_new_page
,
1145 unsigned long private, struct page
*page
,
1146 int force
, enum migrate_mode mode
,
1147 enum migrate_reason reason
,
1148 struct list_head
*ret
)
1150 int rc
= MIGRATEPAGE_SUCCESS
;
1151 struct page
*newpage
= NULL
;
1153 if (!thp_migration_supported() && PageTransHuge(page
))
1156 if (page_count(page
) == 1) {
1157 /* page was freed from under us. So we are done. */
1158 ClearPageActive(page
);
1159 ClearPageUnevictable(page
);
1160 if (unlikely(__PageMovable(page
))) {
1162 if (!PageMovable(page
))
1163 __ClearPageIsolated(page
);
1169 newpage
= get_new_page(page
, private);
1173 rc
= __unmap_and_move(page
, newpage
, force
, mode
);
1174 if (rc
== MIGRATEPAGE_SUCCESS
)
1175 set_page_owner_migrate_reason(newpage
, reason
);
1178 if (rc
!= -EAGAIN
) {
1180 * A page that has been migrated has all references
1181 * removed and will be freed. A page that has not been
1182 * migrated will have kept its references and be restored.
1184 list_del(&page
->lru
);
1188 * If migration is successful, releases reference grabbed during
1189 * isolation. Otherwise, restore the page to right list unless
1192 if (rc
== MIGRATEPAGE_SUCCESS
) {
1194 * Compaction can migrate also non-LRU pages which are
1195 * not accounted to NR_ISOLATED_*. They can be recognized
1198 if (likely(!__PageMovable(page
)))
1199 mod_node_page_state(page_pgdat(page
), NR_ISOLATED_ANON
+
1200 page_is_file_lru(page
), -thp_nr_pages(page
));
1202 if (reason
!= MR_MEMORY_FAILURE
)
1204 * We release the page in page_handle_poison.
1209 list_add_tail(&page
->lru
, ret
);
1212 put_new_page(newpage
, private);
1221 * Counterpart of unmap_and_move_page() for hugepage migration.
1223 * This function doesn't wait the completion of hugepage I/O
1224 * because there is no race between I/O and migration for hugepage.
1225 * Note that currently hugepage I/O occurs only in direct I/O
1226 * where no lock is held and PG_writeback is irrelevant,
1227 * and writeback status of all subpages are counted in the reference
1228 * count of the head page (i.e. if all subpages of a 2MB hugepage are
1229 * under direct I/O, the reference of the head page is 512 and a bit more.)
1230 * This means that when we try to migrate hugepage whose subpages are
1231 * doing direct I/O, some references remain after try_to_unmap() and
1232 * hugepage migration fails without data corruption.
1234 * There is also no race when direct I/O is issued on the page under migration,
1235 * because then pte is replaced with migration swap entry and direct I/O code
1236 * will wait in the page fault for migration to complete.
1238 static int unmap_and_move_huge_page(new_page_t get_new_page
,
1239 free_page_t put_new_page
, unsigned long private,
1240 struct page
*hpage
, int force
,
1241 enum migrate_mode mode
, int reason
,
1242 struct list_head
*ret
)
1245 int page_was_mapped
= 0;
1246 struct page
*new_hpage
;
1247 struct anon_vma
*anon_vma
= NULL
;
1248 struct address_space
*mapping
= NULL
;
1251 * Migratability of hugepages depends on architectures and their size.
1252 * This check is necessary because some callers of hugepage migration
1253 * like soft offline and memory hotremove don't walk through page
1254 * tables or check whether the hugepage is pmd-based or not before
1255 * kicking migration.
1257 if (!hugepage_migration_supported(page_hstate(hpage
))) {
1258 list_move_tail(&hpage
->lru
, ret
);
1262 if (page_count(hpage
) == 1) {
1263 /* page was freed from under us. So we are done. */
1264 putback_active_hugepage(hpage
);
1265 return MIGRATEPAGE_SUCCESS
;
1268 new_hpage
= get_new_page(hpage
, private);
1272 if (!trylock_page(hpage
)) {
1277 case MIGRATE_SYNC_NO_COPY
:
1286 * Check for pages which are in the process of being freed. Without
1287 * page_mapping() set, hugetlbfs specific move page routine will not
1288 * be called and we could leak usage counts for subpools.
1290 if (page_private(hpage
) && !page_mapping(hpage
)) {
1295 if (PageAnon(hpage
))
1296 anon_vma
= page_get_anon_vma(hpage
);
1298 if (unlikely(!trylock_page(new_hpage
)))
1301 if (page_mapped(hpage
)) {
1302 bool mapping_locked
= false;
1303 enum ttu_flags ttu
= TTU_MIGRATION
|TTU_IGNORE_MLOCK
;
1305 if (!PageAnon(hpage
)) {
1307 * In shared mappings, try_to_unmap could potentially
1308 * call huge_pmd_unshare. Because of this, take
1309 * semaphore in write mode here and set TTU_RMAP_LOCKED
1310 * to let lower levels know we have taken the lock.
1312 mapping
= hugetlb_page_mapping_lock_write(hpage
);
1313 if (unlikely(!mapping
))
1314 goto unlock_put_anon
;
1316 mapping_locked
= true;
1317 ttu
|= TTU_RMAP_LOCKED
;
1320 try_to_unmap(hpage
, ttu
);
1321 page_was_mapped
= 1;
1324 i_mmap_unlock_write(mapping
);
1327 if (!page_mapped(hpage
))
1328 rc
= move_to_new_page(new_hpage
, hpage
, mode
);
1330 if (page_was_mapped
)
1331 remove_migration_ptes(hpage
,
1332 rc
== MIGRATEPAGE_SUCCESS
? new_hpage
: hpage
, false);
1335 unlock_page(new_hpage
);
1339 put_anon_vma(anon_vma
);
1341 if (rc
== MIGRATEPAGE_SUCCESS
) {
1342 move_hugetlb_state(hpage
, new_hpage
, reason
);
1343 put_new_page
= NULL
;
1349 if (rc
== MIGRATEPAGE_SUCCESS
)
1350 putback_active_hugepage(hpage
);
1351 else if (rc
!= -EAGAIN
)
1352 list_move_tail(&hpage
->lru
, ret
);
1355 * If migration was not successful and there's a freeing callback, use
1356 * it. Otherwise, put_page() will drop the reference grabbed during
1360 put_new_page(new_hpage
, private);
1362 putback_active_hugepage(new_hpage
);
1367 static inline int try_split_thp(struct page
*page
, struct page
**page2
,
1368 struct list_head
*from
)
1373 rc
= split_huge_page_to_list(page
, from
);
1376 list_safe_reset_next(page
, *page2
, lru
);
1382 * migrate_pages - migrate the pages specified in a list, to the free pages
1383 * supplied as the target for the page migration
1385 * @from: The list of pages to be migrated.
1386 * @get_new_page: The function used to allocate free pages to be used
1387 * as the target of the page migration.
1388 * @put_new_page: The function used to free target pages if migration
1389 * fails, or NULL if no special handling is necessary.
1390 * @private: Private data to be passed on to get_new_page()
1391 * @mode: The migration mode that specifies the constraints for
1392 * page migration, if any.
1393 * @reason: The reason for page migration.
1395 * The function returns after 10 attempts or if no pages are movable any more
1396 * because the list has become empty or no retryable pages exist any more.
1397 * It is caller's responsibility to call putback_movable_pages() to return pages
1398 * to the LRU or free list only if ret != 0.
1400 * Returns the number of pages that were not migrated, or an error code.
1402 int migrate_pages(struct list_head
*from
, new_page_t get_new_page
,
1403 free_page_t put_new_page
, unsigned long private,
1404 enum migrate_mode mode
, int reason
)
1409 int nr_succeeded
= 0;
1410 int nr_thp_succeeded
= 0;
1411 int nr_thp_failed
= 0;
1412 int nr_thp_split
= 0;
1414 bool is_thp
= false;
1417 int swapwrite
= current
->flags
& PF_SWAPWRITE
;
1418 int rc
, nr_subpages
;
1419 LIST_HEAD(ret_pages
);
1421 trace_mm_migrate_pages_start(mode
, reason
);
1424 current
->flags
|= PF_SWAPWRITE
;
1426 for (pass
= 0; pass
< 10 && (retry
|| thp_retry
); pass
++) {
1430 list_for_each_entry_safe(page
, page2
, from
, lru
) {
1433 * THP statistics is based on the source huge page.
1434 * Capture required information that might get lost
1437 is_thp
= PageTransHuge(page
) && !PageHuge(page
);
1438 nr_subpages
= thp_nr_pages(page
);
1442 rc
= unmap_and_move_huge_page(get_new_page
,
1443 put_new_page
, private, page
,
1444 pass
> 2, mode
, reason
,
1447 rc
= unmap_and_move(get_new_page
, put_new_page
,
1448 private, page
, pass
> 2, mode
,
1449 reason
, &ret_pages
);
1452 * Success: non hugetlb page will be freed, hugetlb
1453 * page will be put back
1454 * -EAGAIN: stay on the from list
1455 * -ENOMEM: stay on the from list
1456 * Other errno: put on ret_pages list then splice to
1461 * THP migration might be unsupported or the
1462 * allocation could've failed so we should
1463 * retry on the same page with the THP split
1466 * Head page is retried immediately and tail
1467 * pages are added to the tail of the list so
1468 * we encounter them after the rest of the list
1472 /* THP migration is unsupported */
1474 if (!try_split_thp(page
, &page2
, from
)) {
1480 nr_failed
+= nr_subpages
;
1484 /* Hugetlb migration is unsupported */
1489 * When memory is low, don't bother to try to migrate
1490 * other pages, just exit.
1493 if (!try_split_thp(page
, &page2
, from
)) {
1499 nr_failed
+= nr_subpages
;
1511 case MIGRATEPAGE_SUCCESS
:
1514 nr_succeeded
+= nr_subpages
;
1521 * Permanent failure (-EBUSY, etc.):
1522 * unlike -EAGAIN case, the failed page is
1523 * removed from migration page list and not
1524 * retried in the next outer loop.
1528 nr_failed
+= nr_subpages
;
1536 nr_failed
+= retry
+ thp_retry
;
1537 nr_thp_failed
+= thp_retry
;
1541 * Put the permanent failure page back to migration list, they
1542 * will be put back to the right list by the caller.
1544 list_splice(&ret_pages
, from
);
1546 count_vm_events(PGMIGRATE_SUCCESS
, nr_succeeded
);
1547 count_vm_events(PGMIGRATE_FAIL
, nr_failed
);
1548 count_vm_events(THP_MIGRATION_SUCCESS
, nr_thp_succeeded
);
1549 count_vm_events(THP_MIGRATION_FAIL
, nr_thp_failed
);
1550 count_vm_events(THP_MIGRATION_SPLIT
, nr_thp_split
);
1551 trace_mm_migrate_pages(nr_succeeded
, nr_failed
, nr_thp_succeeded
,
1552 nr_thp_failed
, nr_thp_split
, mode
, reason
);
1555 current
->flags
&= ~PF_SWAPWRITE
;
1560 struct page
*alloc_migration_target(struct page
*page
, unsigned long private)
1562 struct migration_target_control
*mtc
;
1564 unsigned int order
= 0;
1565 struct page
*new_page
= NULL
;
1569 mtc
= (struct migration_target_control
*)private;
1570 gfp_mask
= mtc
->gfp_mask
;
1572 if (nid
== NUMA_NO_NODE
)
1573 nid
= page_to_nid(page
);
1575 if (PageHuge(page
)) {
1576 struct hstate
*h
= page_hstate(compound_head(page
));
1578 gfp_mask
= htlb_modify_alloc_mask(h
, gfp_mask
);
1579 return alloc_huge_page_nodemask(h
, nid
, mtc
->nmask
, gfp_mask
);
1582 if (PageTransHuge(page
)) {
1584 * clear __GFP_RECLAIM to make the migration callback
1585 * consistent with regular THP allocations.
1587 gfp_mask
&= ~__GFP_RECLAIM
;
1588 gfp_mask
|= GFP_TRANSHUGE
;
1589 order
= HPAGE_PMD_ORDER
;
1591 zidx
= zone_idx(page_zone(page
));
1592 if (is_highmem_idx(zidx
) || zidx
== ZONE_MOVABLE
)
1593 gfp_mask
|= __GFP_HIGHMEM
;
1595 new_page
= __alloc_pages(gfp_mask
, order
, nid
, mtc
->nmask
);
1597 if (new_page
&& PageTransHuge(new_page
))
1598 prep_transhuge_page(new_page
);
1605 static int store_status(int __user
*status
, int start
, int value
, int nr
)
1608 if (put_user(value
, status
+ start
))
1616 static int do_move_pages_to_node(struct mm_struct
*mm
,
1617 struct list_head
*pagelist
, int node
)
1620 struct migration_target_control mtc
= {
1622 .gfp_mask
= GFP_HIGHUSER_MOVABLE
| __GFP_THISNODE
,
1625 err
= migrate_pages(pagelist
, alloc_migration_target
, NULL
,
1626 (unsigned long)&mtc
, MIGRATE_SYNC
, MR_SYSCALL
);
1628 putback_movable_pages(pagelist
);
1633 * Resolves the given address to a struct page, isolates it from the LRU and
1634 * puts it to the given pagelist.
1636 * errno - if the page cannot be found/isolated
1637 * 0 - when it doesn't have to be migrated because it is already on the
1639 * 1 - when it has been queued
1641 static int add_page_for_migration(struct mm_struct
*mm
, unsigned long addr
,
1642 int node
, struct list_head
*pagelist
, bool migrate_all
)
1644 struct vm_area_struct
*vma
;
1646 unsigned int follflags
;
1651 vma
= find_vma(mm
, addr
);
1652 if (!vma
|| addr
< vma
->vm_start
|| !vma_migratable(vma
))
1655 /* FOLL_DUMP to ignore special (like zero) pages */
1656 follflags
= FOLL_GET
| FOLL_DUMP
;
1657 page
= follow_page(vma
, addr
, follflags
);
1659 err
= PTR_ERR(page
);
1668 if (page_to_nid(page
) == node
)
1672 if (page_mapcount(page
) > 1 && !migrate_all
)
1675 if (PageHuge(page
)) {
1676 if (PageHead(page
)) {
1677 isolate_huge_page(page
, pagelist
);
1683 head
= compound_head(page
);
1684 err
= isolate_lru_page(head
);
1689 list_add_tail(&head
->lru
, pagelist
);
1690 mod_node_page_state(page_pgdat(head
),
1691 NR_ISOLATED_ANON
+ page_is_file_lru(head
),
1692 thp_nr_pages(head
));
1696 * Either remove the duplicate refcount from
1697 * isolate_lru_page() or drop the page ref if it was
1702 mmap_read_unlock(mm
);
1706 static int move_pages_and_store_status(struct mm_struct
*mm
, int node
,
1707 struct list_head
*pagelist
, int __user
*status
,
1708 int start
, int i
, unsigned long nr_pages
)
1712 if (list_empty(pagelist
))
1715 err
= do_move_pages_to_node(mm
, pagelist
, node
);
1718 * Positive err means the number of failed
1719 * pages to migrate. Since we are going to
1720 * abort and return the number of non-migrated
1721 * pages, so need to include the rest of the
1722 * nr_pages that have not been attempted as
1726 err
+= nr_pages
- i
- 1;
1729 return store_status(status
, start
, node
, i
- start
);
1733 * Migrate an array of page address onto an array of nodes and fill
1734 * the corresponding array of status.
1736 static int do_pages_move(struct mm_struct
*mm
, nodemask_t task_nodes
,
1737 unsigned long nr_pages
,
1738 const void __user
* __user
*pages
,
1739 const int __user
*nodes
,
1740 int __user
*status
, int flags
)
1742 int current_node
= NUMA_NO_NODE
;
1743 LIST_HEAD(pagelist
);
1747 lru_cache_disable();
1749 for (i
= start
= 0; i
< nr_pages
; i
++) {
1750 const void __user
*p
;
1755 if (get_user(p
, pages
+ i
))
1757 if (get_user(node
, nodes
+ i
))
1759 addr
= (unsigned long)untagged_addr(p
);
1762 if (node
< 0 || node
>= MAX_NUMNODES
)
1764 if (!node_state(node
, N_MEMORY
))
1768 if (!node_isset(node
, task_nodes
))
1771 if (current_node
== NUMA_NO_NODE
) {
1772 current_node
= node
;
1774 } else if (node
!= current_node
) {
1775 err
= move_pages_and_store_status(mm
, current_node
,
1776 &pagelist
, status
, start
, i
, nr_pages
);
1780 current_node
= node
;
1784 * Errors in the page lookup or isolation are not fatal and we simply
1785 * report them via status
1787 err
= add_page_for_migration(mm
, addr
, current_node
,
1788 &pagelist
, flags
& MPOL_MF_MOVE_ALL
);
1791 /* The page is successfully queued for migration */
1796 * If the page is already on the target node (!err), store the
1797 * node, otherwise, store the err.
1799 err
= store_status(status
, i
, err
? : current_node
, 1);
1803 err
= move_pages_and_store_status(mm
, current_node
, &pagelist
,
1804 status
, start
, i
, nr_pages
);
1807 current_node
= NUMA_NO_NODE
;
1810 /* Make sure we do not overwrite the existing error */
1811 err1
= move_pages_and_store_status(mm
, current_node
, &pagelist
,
1812 status
, start
, i
, nr_pages
);
1821 * Determine the nodes of an array of pages and store it in an array of status.
1823 static void do_pages_stat_array(struct mm_struct
*mm
, unsigned long nr_pages
,
1824 const void __user
**pages
, int *status
)
1830 for (i
= 0; i
< nr_pages
; i
++) {
1831 unsigned long addr
= (unsigned long)(*pages
);
1832 struct vm_area_struct
*vma
;
1836 vma
= find_vma(mm
, addr
);
1837 if (!vma
|| addr
< vma
->vm_start
)
1840 /* FOLL_DUMP to ignore special (like zero) pages */
1841 page
= follow_page(vma
, addr
, FOLL_DUMP
);
1843 err
= PTR_ERR(page
);
1847 err
= page
? page_to_nid(page
) : -ENOENT
;
1855 mmap_read_unlock(mm
);
1859 * Determine the nodes of a user array of pages and store it in
1860 * a user array of status.
1862 static int do_pages_stat(struct mm_struct
*mm
, unsigned long nr_pages
,
1863 const void __user
* __user
*pages
,
1866 #define DO_PAGES_STAT_CHUNK_NR 16
1867 const void __user
*chunk_pages
[DO_PAGES_STAT_CHUNK_NR
];
1868 int chunk_status
[DO_PAGES_STAT_CHUNK_NR
];
1871 unsigned long chunk_nr
;
1873 chunk_nr
= nr_pages
;
1874 if (chunk_nr
> DO_PAGES_STAT_CHUNK_NR
)
1875 chunk_nr
= DO_PAGES_STAT_CHUNK_NR
;
1877 if (copy_from_user(chunk_pages
, pages
, chunk_nr
* sizeof(*chunk_pages
)))
1880 do_pages_stat_array(mm
, chunk_nr
, chunk_pages
, chunk_status
);
1882 if (copy_to_user(status
, chunk_status
, chunk_nr
* sizeof(*status
)))
1887 nr_pages
-= chunk_nr
;
1889 return nr_pages
? -EFAULT
: 0;
1892 static struct mm_struct
*find_mm_struct(pid_t pid
, nodemask_t
*mem_nodes
)
1894 struct task_struct
*task
;
1895 struct mm_struct
*mm
;
1898 * There is no need to check if current process has the right to modify
1899 * the specified process when they are same.
1903 *mem_nodes
= cpuset_mems_allowed(current
);
1907 /* Find the mm_struct */
1909 task
= find_task_by_vpid(pid
);
1912 return ERR_PTR(-ESRCH
);
1914 get_task_struct(task
);
1917 * Check if this process has the right to modify the specified
1918 * process. Use the regular "ptrace_may_access()" checks.
1920 if (!ptrace_may_access(task
, PTRACE_MODE_READ_REALCREDS
)) {
1922 mm
= ERR_PTR(-EPERM
);
1927 mm
= ERR_PTR(security_task_movememory(task
));
1930 *mem_nodes
= cpuset_mems_allowed(task
);
1931 mm
= get_task_mm(task
);
1933 put_task_struct(task
);
1935 mm
= ERR_PTR(-EINVAL
);
1940 * Move a list of pages in the address space of the currently executing
1943 static int kernel_move_pages(pid_t pid
, unsigned long nr_pages
,
1944 const void __user
* __user
*pages
,
1945 const int __user
*nodes
,
1946 int __user
*status
, int flags
)
1948 struct mm_struct
*mm
;
1950 nodemask_t task_nodes
;
1953 if (flags
& ~(MPOL_MF_MOVE
|MPOL_MF_MOVE_ALL
))
1956 if ((flags
& MPOL_MF_MOVE_ALL
) && !capable(CAP_SYS_NICE
))
1959 mm
= find_mm_struct(pid
, &task_nodes
);
1964 err
= do_pages_move(mm
, task_nodes
, nr_pages
, pages
,
1965 nodes
, status
, flags
);
1967 err
= do_pages_stat(mm
, nr_pages
, pages
, status
);
1973 SYSCALL_DEFINE6(move_pages
, pid_t
, pid
, unsigned long, nr_pages
,
1974 const void __user
* __user
*, pages
,
1975 const int __user
*, nodes
,
1976 int __user
*, status
, int, flags
)
1978 return kernel_move_pages(pid
, nr_pages
, pages
, nodes
, status
, flags
);
1981 #ifdef CONFIG_COMPAT
1982 COMPAT_SYSCALL_DEFINE6(move_pages
, pid_t
, pid
, compat_ulong_t
, nr_pages
,
1983 compat_uptr_t __user
*, pages32
,
1984 const int __user
*, nodes
,
1985 int __user
*, status
,
1988 const void __user
* __user
*pages
;
1991 pages
= compat_alloc_user_space(nr_pages
* sizeof(void *));
1992 for (i
= 0; i
< nr_pages
; i
++) {
1995 if (get_user(p
, pages32
+ i
) ||
1996 put_user(compat_ptr(p
), pages
+ i
))
1999 return kernel_move_pages(pid
, nr_pages
, pages
, nodes
, status
, flags
);
2001 #endif /* CONFIG_COMPAT */
2003 #ifdef CONFIG_NUMA_BALANCING
2005 * Returns true if this is a safe migration target node for misplaced NUMA
2006 * pages. Currently it only checks the watermarks which crude
2008 static bool migrate_balanced_pgdat(struct pglist_data
*pgdat
,
2009 unsigned long nr_migrate_pages
)
2013 for (z
= pgdat
->nr_zones
- 1; z
>= 0; z
--) {
2014 struct zone
*zone
= pgdat
->node_zones
+ z
;
2016 if (!populated_zone(zone
))
2019 /* Avoid waking kswapd by allocating pages_to_migrate pages. */
2020 if (!zone_watermark_ok(zone
, 0,
2021 high_wmark_pages(zone
) +
2030 static struct page
*alloc_misplaced_dst_page(struct page
*page
,
2033 int nid
= (int) data
;
2034 struct page
*newpage
;
2036 newpage
= __alloc_pages_node(nid
,
2037 (GFP_HIGHUSER_MOVABLE
|
2038 __GFP_THISNODE
| __GFP_NOMEMALLOC
|
2039 __GFP_NORETRY
| __GFP_NOWARN
) &
2045 static int numamigrate_isolate_page(pg_data_t
*pgdat
, struct page
*page
)
2049 VM_BUG_ON_PAGE(compound_order(page
) && !PageTransHuge(page
), page
);
2051 /* Avoid migrating to a node that is nearly full */
2052 if (!migrate_balanced_pgdat(pgdat
, compound_nr(page
)))
2055 if (isolate_lru_page(page
))
2059 * migrate_misplaced_transhuge_page() skips page migration's usual
2060 * check on page_count(), so we must do it here, now that the page
2061 * has been isolated: a GUP pin, or any other pin, prevents migration.
2062 * The expected page count is 3: 1 for page's mapcount and 1 for the
2063 * caller's pin and 1 for the reference taken by isolate_lru_page().
2065 if (PageTransHuge(page
) && page_count(page
) != 3) {
2066 putback_lru_page(page
);
2070 page_lru
= page_is_file_lru(page
);
2071 mod_node_page_state(page_pgdat(page
), NR_ISOLATED_ANON
+ page_lru
,
2072 thp_nr_pages(page
));
2075 * Isolating the page has taken another reference, so the
2076 * caller's reference can be safely dropped without the page
2077 * disappearing underneath us during migration.
2083 bool pmd_trans_migrating(pmd_t pmd
)
2085 struct page
*page
= pmd_page(pmd
);
2086 return PageLocked(page
);
2090 * Attempt to migrate a misplaced page to the specified destination
2091 * node. Caller is expected to have an elevated reference count on
2092 * the page that will be dropped by this function before returning.
2094 int migrate_misplaced_page(struct page
*page
, struct vm_area_struct
*vma
,
2097 pg_data_t
*pgdat
= NODE_DATA(node
);
2100 LIST_HEAD(migratepages
);
2103 * Don't migrate file pages that are mapped in multiple processes
2104 * with execute permissions as they are probably shared libraries.
2106 if (page_mapcount(page
) != 1 && page_is_file_lru(page
) &&
2107 (vma
->vm_flags
& VM_EXEC
))
2111 * Also do not migrate dirty pages as not all filesystems can move
2112 * dirty pages in MIGRATE_ASYNC mode which is a waste of cycles.
2114 if (page_is_file_lru(page
) && PageDirty(page
))
2117 isolated
= numamigrate_isolate_page(pgdat
, page
);
2121 list_add(&page
->lru
, &migratepages
);
2122 nr_remaining
= migrate_pages(&migratepages
, alloc_misplaced_dst_page
,
2123 NULL
, node
, MIGRATE_ASYNC
,
2126 if (!list_empty(&migratepages
)) {
2127 list_del(&page
->lru
);
2128 dec_node_page_state(page
, NR_ISOLATED_ANON
+
2129 page_is_file_lru(page
));
2130 putback_lru_page(page
);
2134 count_vm_numa_event(NUMA_PAGE_MIGRATE
);
2135 BUG_ON(!list_empty(&migratepages
));
2142 #endif /* CONFIG_NUMA_BALANCING */
2144 #if defined(CONFIG_NUMA_BALANCING) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
2146 * Migrates a THP to a given target node. page must be locked and is unlocked
2149 int migrate_misplaced_transhuge_page(struct mm_struct
*mm
,
2150 struct vm_area_struct
*vma
,
2151 pmd_t
*pmd
, pmd_t entry
,
2152 unsigned long address
,
2153 struct page
*page
, int node
)
2156 pg_data_t
*pgdat
= NODE_DATA(node
);
2158 struct page
*new_page
= NULL
;
2159 int page_lru
= page_is_file_lru(page
);
2160 unsigned long start
= address
& HPAGE_PMD_MASK
;
2162 new_page
= alloc_pages_node(node
,
2163 (GFP_TRANSHUGE_LIGHT
| __GFP_THISNODE
),
2167 prep_transhuge_page(new_page
);
2169 isolated
= numamigrate_isolate_page(pgdat
, page
);
2175 /* Prepare a page as a migration target */
2176 __SetPageLocked(new_page
);
2177 if (PageSwapBacked(page
))
2178 __SetPageSwapBacked(new_page
);
2180 /* anon mapping, we can simply copy page->mapping to the new page: */
2181 new_page
->mapping
= page
->mapping
;
2182 new_page
->index
= page
->index
;
2183 /* flush the cache before copying using the kernel virtual address */
2184 flush_cache_range(vma
, start
, start
+ HPAGE_PMD_SIZE
);
2185 migrate_page_copy(new_page
, page
);
2186 WARN_ON(PageLRU(new_page
));
2188 /* Recheck the target PMD */
2189 ptl
= pmd_lock(mm
, pmd
);
2190 if (unlikely(!pmd_same(*pmd
, entry
) || !page_ref_freeze(page
, 2))) {
2193 /* Reverse changes made by migrate_page_copy() */
2194 if (TestClearPageActive(new_page
))
2195 SetPageActive(page
);
2196 if (TestClearPageUnevictable(new_page
))
2197 SetPageUnevictable(page
);
2199 unlock_page(new_page
);
2200 put_page(new_page
); /* Free it */
2202 /* Retake the callers reference and putback on LRU */
2204 putback_lru_page(page
);
2205 mod_node_page_state(page_pgdat(page
),
2206 NR_ISOLATED_ANON
+ page_lru
, -HPAGE_PMD_NR
);
2211 entry
= mk_huge_pmd(new_page
, vma
->vm_page_prot
);
2212 entry
= maybe_pmd_mkwrite(pmd_mkdirty(entry
), vma
);
2215 * Overwrite the old entry under pagetable lock and establish
2216 * the new PTE. Any parallel GUP will either observe the old
2217 * page blocking on the page lock, block on the page table
2218 * lock or observe the new page. The SetPageUptodate on the
2219 * new page and page_add_new_anon_rmap guarantee the copy is
2220 * visible before the pagetable update.
2222 page_add_anon_rmap(new_page
, vma
, start
, true);
2224 * At this point the pmd is numa/protnone (i.e. non present) and the TLB
2225 * has already been flushed globally. So no TLB can be currently
2226 * caching this non present pmd mapping. There's no need to clear the
2227 * pmd before doing set_pmd_at(), nor to flush the TLB after
2228 * set_pmd_at(). Clearing the pmd here would introduce a race
2229 * condition against MADV_DONTNEED, because MADV_DONTNEED only holds the
2230 * mmap_lock for reading. If the pmd is set to NULL at any given time,
2231 * MADV_DONTNEED won't wait on the pmd lock and it'll skip clearing this
2234 set_pmd_at(mm
, start
, pmd
, entry
);
2235 update_mmu_cache_pmd(vma
, address
, &entry
);
2237 page_ref_unfreeze(page
, 2);
2238 mlock_migrate_page(new_page
, page
);
2239 page_remove_rmap(page
, true);
2240 set_page_owner_migrate_reason(new_page
, MR_NUMA_MISPLACED
);
2244 /* Take an "isolate" reference and put new page on the LRU. */
2246 putback_lru_page(new_page
);
2248 unlock_page(new_page
);
2250 put_page(page
); /* Drop the rmap reference */
2251 put_page(page
); /* Drop the LRU isolation reference */
2253 count_vm_events(PGMIGRATE_SUCCESS
, HPAGE_PMD_NR
);
2254 count_vm_numa_events(NUMA_PAGE_MIGRATE
, HPAGE_PMD_NR
);
2256 mod_node_page_state(page_pgdat(page
),
2257 NR_ISOLATED_ANON
+ page_lru
,
2262 count_vm_events(PGMIGRATE_FAIL
, HPAGE_PMD_NR
);
2263 ptl
= pmd_lock(mm
, pmd
);
2264 if (pmd_same(*pmd
, entry
)) {
2265 entry
= pmd_modify(entry
, vma
->vm_page_prot
);
2266 set_pmd_at(mm
, start
, pmd
, entry
);
2267 update_mmu_cache_pmd(vma
, address
, &entry
);
2276 #endif /* CONFIG_NUMA_BALANCING */
2278 #endif /* CONFIG_NUMA */
2280 #ifdef CONFIG_DEVICE_PRIVATE
2281 static int migrate_vma_collect_skip(unsigned long start
,
2283 struct mm_walk
*walk
)
2285 struct migrate_vma
*migrate
= walk
->private;
2288 for (addr
= start
; addr
< end
; addr
+= PAGE_SIZE
) {
2289 migrate
->dst
[migrate
->npages
] = 0;
2290 migrate
->src
[migrate
->npages
++] = 0;
2296 static int migrate_vma_collect_hole(unsigned long start
,
2298 __always_unused
int depth
,
2299 struct mm_walk
*walk
)
2301 struct migrate_vma
*migrate
= walk
->private;
2304 /* Only allow populating anonymous memory. */
2305 if (!vma_is_anonymous(walk
->vma
))
2306 return migrate_vma_collect_skip(start
, end
, walk
);
2308 for (addr
= start
; addr
< end
; addr
+= PAGE_SIZE
) {
2309 migrate
->src
[migrate
->npages
] = MIGRATE_PFN_MIGRATE
;
2310 migrate
->dst
[migrate
->npages
] = 0;
2318 static int migrate_vma_collect_pmd(pmd_t
*pmdp
,
2319 unsigned long start
,
2321 struct mm_walk
*walk
)
2323 struct migrate_vma
*migrate
= walk
->private;
2324 struct vm_area_struct
*vma
= walk
->vma
;
2325 struct mm_struct
*mm
= vma
->vm_mm
;
2326 unsigned long addr
= start
, unmapped
= 0;
2331 if (pmd_none(*pmdp
))
2332 return migrate_vma_collect_hole(start
, end
, -1, walk
);
2334 if (pmd_trans_huge(*pmdp
)) {
2337 ptl
= pmd_lock(mm
, pmdp
);
2338 if (unlikely(!pmd_trans_huge(*pmdp
))) {
2343 page
= pmd_page(*pmdp
);
2344 if (is_huge_zero_page(page
)) {
2346 split_huge_pmd(vma
, pmdp
, addr
);
2347 if (pmd_trans_unstable(pmdp
))
2348 return migrate_vma_collect_skip(start
, end
,
2355 if (unlikely(!trylock_page(page
)))
2356 return migrate_vma_collect_skip(start
, end
,
2358 ret
= split_huge_page(page
);
2362 return migrate_vma_collect_skip(start
, end
,
2364 if (pmd_none(*pmdp
))
2365 return migrate_vma_collect_hole(start
, end
, -1,
2370 if (unlikely(pmd_bad(*pmdp
)))
2371 return migrate_vma_collect_skip(start
, end
, walk
);
2373 ptep
= pte_offset_map_lock(mm
, pmdp
, addr
, &ptl
);
2374 arch_enter_lazy_mmu_mode();
2376 for (; addr
< end
; addr
+= PAGE_SIZE
, ptep
++) {
2377 unsigned long mpfn
= 0, pfn
;
2384 if (pte_none(pte
)) {
2385 if (vma_is_anonymous(vma
)) {
2386 mpfn
= MIGRATE_PFN_MIGRATE
;
2392 if (!pte_present(pte
)) {
2394 * Only care about unaddressable device page special
2395 * page table entry. Other special swap entries are not
2396 * migratable, and we ignore regular swapped page.
2398 entry
= pte_to_swp_entry(pte
);
2399 if (!is_device_private_entry(entry
))
2402 page
= device_private_entry_to_page(entry
);
2403 if (!(migrate
->flags
&
2404 MIGRATE_VMA_SELECT_DEVICE_PRIVATE
) ||
2405 page
->pgmap
->owner
!= migrate
->pgmap_owner
)
2408 mpfn
= migrate_pfn(page_to_pfn(page
)) |
2409 MIGRATE_PFN_MIGRATE
;
2410 if (is_write_device_private_entry(entry
))
2411 mpfn
|= MIGRATE_PFN_WRITE
;
2413 if (!(migrate
->flags
& MIGRATE_VMA_SELECT_SYSTEM
))
2416 if (is_zero_pfn(pfn
)) {
2417 mpfn
= MIGRATE_PFN_MIGRATE
;
2421 page
= vm_normal_page(migrate
->vma
, addr
, pte
);
2422 mpfn
= migrate_pfn(pfn
) | MIGRATE_PFN_MIGRATE
;
2423 mpfn
|= pte_write(pte
) ? MIGRATE_PFN_WRITE
: 0;
2426 /* FIXME support THP */
2427 if (!page
|| !page
->mapping
|| PageTransCompound(page
)) {
2433 * By getting a reference on the page we pin it and that blocks
2434 * any kind of migration. Side effect is that it "freezes" the
2437 * We drop this reference after isolating the page from the lru
2438 * for non device page (device page are not on the lru and thus
2439 * can't be dropped from it).
2445 * Optimize for the common case where page is only mapped once
2446 * in one process. If we can lock the page, then we can safely
2447 * set up a special migration page table entry now.
2449 if (trylock_page(page
)) {
2452 mpfn
|= MIGRATE_PFN_LOCKED
;
2453 ptep_get_and_clear(mm
, addr
, ptep
);
2455 /* Setup special migration page table entry */
2456 entry
= make_migration_entry(page
, mpfn
&
2458 swp_pte
= swp_entry_to_pte(entry
);
2459 if (pte_present(pte
)) {
2460 if (pte_soft_dirty(pte
))
2461 swp_pte
= pte_swp_mksoft_dirty(swp_pte
);
2462 if (pte_uffd_wp(pte
))
2463 swp_pte
= pte_swp_mkuffd_wp(swp_pte
);
2465 if (pte_swp_soft_dirty(pte
))
2466 swp_pte
= pte_swp_mksoft_dirty(swp_pte
);
2467 if (pte_swp_uffd_wp(pte
))
2468 swp_pte
= pte_swp_mkuffd_wp(swp_pte
);
2470 set_pte_at(mm
, addr
, ptep
, swp_pte
);
2473 * This is like regular unmap: we remove the rmap and
2474 * drop page refcount. Page won't be freed, as we took
2475 * a reference just above.
2477 page_remove_rmap(page
, false);
2480 if (pte_present(pte
))
2485 migrate
->dst
[migrate
->npages
] = 0;
2486 migrate
->src
[migrate
->npages
++] = mpfn
;
2488 arch_leave_lazy_mmu_mode();
2489 pte_unmap_unlock(ptep
- 1, ptl
);
2491 /* Only flush the TLB if we actually modified any entries */
2493 flush_tlb_range(walk
->vma
, start
, end
);
2498 static const struct mm_walk_ops migrate_vma_walk_ops
= {
2499 .pmd_entry
= migrate_vma_collect_pmd
,
2500 .pte_hole
= migrate_vma_collect_hole
,
2504 * migrate_vma_collect() - collect pages over a range of virtual addresses
2505 * @migrate: migrate struct containing all migration information
2507 * This will walk the CPU page table. For each virtual address backed by a
2508 * valid page, it updates the src array and takes a reference on the page, in
2509 * order to pin the page until we lock it and unmap it.
2511 static void migrate_vma_collect(struct migrate_vma
*migrate
)
2513 struct mmu_notifier_range range
;
2516 * Note that the pgmap_owner is passed to the mmu notifier callback so
2517 * that the registered device driver can skip invalidating device
2518 * private page mappings that won't be migrated.
2520 mmu_notifier_range_init_migrate(&range
, 0, migrate
->vma
,
2521 migrate
->vma
->vm_mm
, migrate
->start
, migrate
->end
,
2522 migrate
->pgmap_owner
);
2523 mmu_notifier_invalidate_range_start(&range
);
2525 walk_page_range(migrate
->vma
->vm_mm
, migrate
->start
, migrate
->end
,
2526 &migrate_vma_walk_ops
, migrate
);
2528 mmu_notifier_invalidate_range_end(&range
);
2529 migrate
->end
= migrate
->start
+ (migrate
->npages
<< PAGE_SHIFT
);
2533 * migrate_vma_check_page() - check if page is pinned or not
2534 * @page: struct page to check
2536 * Pinned pages cannot be migrated. This is the same test as in
2537 * migrate_page_move_mapping(), except that here we allow migration of a
2540 static bool migrate_vma_check_page(struct page
*page
)
2543 * One extra ref because caller holds an extra reference, either from
2544 * isolate_lru_page() for a regular page, or migrate_vma_collect() for
2550 * FIXME support THP (transparent huge page), it is bit more complex to
2551 * check them than regular pages, because they can be mapped with a pmd
2552 * or with a pte (split pte mapping).
2554 if (PageCompound(page
))
2557 /* Page from ZONE_DEVICE have one extra reference */
2558 if (is_zone_device_page(page
)) {
2560 * Private page can never be pin as they have no valid pte and
2561 * GUP will fail for those. Yet if there is a pending migration
2562 * a thread might try to wait on the pte migration entry and
2563 * will bump the page reference count. Sadly there is no way to
2564 * differentiate a regular pin from migration wait. Hence to
2565 * avoid 2 racing thread trying to migrate back to CPU to enter
2566 * infinite loop (one stopping migration because the other is
2567 * waiting on pte migration entry). We always return true here.
2569 * FIXME proper solution is to rework migration_entry_wait() so
2570 * it does not need to take a reference on page.
2572 return is_device_private_page(page
);
2575 /* For file back page */
2576 if (page_mapping(page
))
2577 extra
+= 1 + page_has_private(page
);
2579 if ((page_count(page
) - extra
) > page_mapcount(page
))
2586 * migrate_vma_prepare() - lock pages and isolate them from the lru
2587 * @migrate: migrate struct containing all migration information
2589 * This locks pages that have been collected by migrate_vma_collect(). Once each
2590 * page is locked it is isolated from the lru (for non-device pages). Finally,
2591 * the ref taken by migrate_vma_collect() is dropped, as locked pages cannot be
2592 * migrated by concurrent kernel threads.
2594 static void migrate_vma_prepare(struct migrate_vma
*migrate
)
2596 const unsigned long npages
= migrate
->npages
;
2597 const unsigned long start
= migrate
->start
;
2598 unsigned long addr
, i
, restore
= 0;
2599 bool allow_drain
= true;
2603 for (i
= 0; (i
< npages
) && migrate
->cpages
; i
++) {
2604 struct page
*page
= migrate_pfn_to_page(migrate
->src
[i
]);
2610 if (!(migrate
->src
[i
] & MIGRATE_PFN_LOCKED
)) {
2612 * Because we are migrating several pages there can be
2613 * a deadlock between 2 concurrent migration where each
2614 * are waiting on each other page lock.
2616 * Make migrate_vma() a best effort thing and backoff
2617 * for any page we can not lock right away.
2619 if (!trylock_page(page
)) {
2620 migrate
->src
[i
] = 0;
2626 migrate
->src
[i
] |= MIGRATE_PFN_LOCKED
;
2629 /* ZONE_DEVICE pages are not on LRU */
2630 if (!is_zone_device_page(page
)) {
2631 if (!PageLRU(page
) && allow_drain
) {
2632 /* Drain CPU's pagevec */
2633 lru_add_drain_all();
2634 allow_drain
= false;
2637 if (isolate_lru_page(page
)) {
2639 migrate
->src
[i
] &= ~MIGRATE_PFN_MIGRATE
;
2643 migrate
->src
[i
] = 0;
2651 /* Drop the reference we took in collect */
2655 if (!migrate_vma_check_page(page
)) {
2657 migrate
->src
[i
] &= ~MIGRATE_PFN_MIGRATE
;
2661 if (!is_zone_device_page(page
)) {
2663 putback_lru_page(page
);
2666 migrate
->src
[i
] = 0;
2670 if (!is_zone_device_page(page
))
2671 putback_lru_page(page
);
2678 for (i
= 0, addr
= start
; i
< npages
&& restore
; i
++, addr
+= PAGE_SIZE
) {
2679 struct page
*page
= migrate_pfn_to_page(migrate
->src
[i
]);
2681 if (!page
|| (migrate
->src
[i
] & MIGRATE_PFN_MIGRATE
))
2684 remove_migration_pte(page
, migrate
->vma
, addr
, page
);
2686 migrate
->src
[i
] = 0;
2694 * migrate_vma_unmap() - replace page mapping with special migration pte entry
2695 * @migrate: migrate struct containing all migration information
2697 * Replace page mapping (CPU page table pte) with a special migration pte entry
2698 * and check again if it has been pinned. Pinned pages are restored because we
2699 * cannot migrate them.
2701 * This is the last step before we call the device driver callback to allocate
2702 * destination memory and copy contents of original page over to new page.
2704 static void migrate_vma_unmap(struct migrate_vma
*migrate
)
2706 int flags
= TTU_MIGRATION
| TTU_IGNORE_MLOCK
;
2707 const unsigned long npages
= migrate
->npages
;
2708 const unsigned long start
= migrate
->start
;
2709 unsigned long addr
, i
, restore
= 0;
2711 for (i
= 0; i
< npages
; i
++) {
2712 struct page
*page
= migrate_pfn_to_page(migrate
->src
[i
]);
2714 if (!page
|| !(migrate
->src
[i
] & MIGRATE_PFN_MIGRATE
))
2717 if (page_mapped(page
)) {
2718 try_to_unmap(page
, flags
);
2719 if (page_mapped(page
))
2723 if (migrate_vma_check_page(page
))
2727 migrate
->src
[i
] &= ~MIGRATE_PFN_MIGRATE
;
2732 for (addr
= start
, i
= 0; i
< npages
&& restore
; addr
+= PAGE_SIZE
, i
++) {
2733 struct page
*page
= migrate_pfn_to_page(migrate
->src
[i
]);
2735 if (!page
|| (migrate
->src
[i
] & MIGRATE_PFN_MIGRATE
))
2738 remove_migration_ptes(page
, page
, false);
2740 migrate
->src
[i
] = 0;
2744 if (is_zone_device_page(page
))
2747 putback_lru_page(page
);
2752 * migrate_vma_setup() - prepare to migrate a range of memory
2753 * @args: contains the vma, start, and pfns arrays for the migration
2755 * Returns: negative errno on failures, 0 when 0 or more pages were migrated
2758 * Prepare to migrate a range of memory virtual address range by collecting all
2759 * the pages backing each virtual address in the range, saving them inside the
2760 * src array. Then lock those pages and unmap them. Once the pages are locked
2761 * and unmapped, check whether each page is pinned or not. Pages that aren't
2762 * pinned have the MIGRATE_PFN_MIGRATE flag set (by this function) in the
2763 * corresponding src array entry. Then restores any pages that are pinned, by
2764 * remapping and unlocking those pages.
2766 * The caller should then allocate destination memory and copy source memory to
2767 * it for all those entries (ie with MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE
2768 * flag set). Once these are allocated and copied, the caller must update each
2769 * corresponding entry in the dst array with the pfn value of the destination
2770 * page and with the MIGRATE_PFN_VALID and MIGRATE_PFN_LOCKED flags set
2771 * (destination pages must have their struct pages locked, via lock_page()).
2773 * Note that the caller does not have to migrate all the pages that are marked
2774 * with MIGRATE_PFN_MIGRATE flag in src array unless this is a migration from
2775 * device memory to system memory. If the caller cannot migrate a device page
2776 * back to system memory, then it must return VM_FAULT_SIGBUS, which has severe
2777 * consequences for the userspace process, so it must be avoided if at all
2780 * For empty entries inside CPU page table (pte_none() or pmd_none() is true) we
2781 * do set MIGRATE_PFN_MIGRATE flag inside the corresponding source array thus
2782 * allowing the caller to allocate device memory for those unbacked virtual
2783 * addresses. For this the caller simply has to allocate device memory and
2784 * properly set the destination entry like for regular migration. Note that
2785 * this can still fail, and thus inside the device driver you must check if the
2786 * migration was successful for those entries after calling migrate_vma_pages(),
2787 * just like for regular migration.
2789 * After that, the callers must call migrate_vma_pages() to go over each entry
2790 * in the src array that has the MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE flag
2791 * set. If the corresponding entry in dst array has MIGRATE_PFN_VALID flag set,
2792 * then migrate_vma_pages() to migrate struct page information from the source
2793 * struct page to the destination struct page. If it fails to migrate the
2794 * struct page information, then it clears the MIGRATE_PFN_MIGRATE flag in the
2797 * At this point all successfully migrated pages have an entry in the src
2798 * array with MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE flag set and the dst
2799 * array entry with MIGRATE_PFN_VALID flag set.
2801 * Once migrate_vma_pages() returns the caller may inspect which pages were
2802 * successfully migrated, and which were not. Successfully migrated pages will
2803 * have the MIGRATE_PFN_MIGRATE flag set for their src array entry.
2805 * It is safe to update device page table after migrate_vma_pages() because
2806 * both destination and source page are still locked, and the mmap_lock is held
2807 * in read mode (hence no one can unmap the range being migrated).
2809 * Once the caller is done cleaning up things and updating its page table (if it
2810 * chose to do so, this is not an obligation) it finally calls
2811 * migrate_vma_finalize() to update the CPU page table to point to new pages
2812 * for successfully migrated pages or otherwise restore the CPU page table to
2813 * point to the original source pages.
2815 int migrate_vma_setup(struct migrate_vma
*args
)
2817 long nr_pages
= (args
->end
- args
->start
) >> PAGE_SHIFT
;
2819 args
->start
&= PAGE_MASK
;
2820 args
->end
&= PAGE_MASK
;
2821 if (!args
->vma
|| is_vm_hugetlb_page(args
->vma
) ||
2822 (args
->vma
->vm_flags
& VM_SPECIAL
) || vma_is_dax(args
->vma
))
2826 if (args
->start
< args
->vma
->vm_start
||
2827 args
->start
>= args
->vma
->vm_end
)
2829 if (args
->end
<= args
->vma
->vm_start
|| args
->end
> args
->vma
->vm_end
)
2831 if (!args
->src
|| !args
->dst
)
2834 memset(args
->src
, 0, sizeof(*args
->src
) * nr_pages
);
2838 migrate_vma_collect(args
);
2841 migrate_vma_prepare(args
);
2843 migrate_vma_unmap(args
);
2846 * At this point pages are locked and unmapped, and thus they have
2847 * stable content and can safely be copied to destination memory that
2848 * is allocated by the drivers.
2853 EXPORT_SYMBOL(migrate_vma_setup
);
2856 * This code closely matches the code in:
2857 * __handle_mm_fault()
2858 * handle_pte_fault()
2859 * do_anonymous_page()
2860 * to map in an anonymous zero page but the struct page will be a ZONE_DEVICE
2863 static void migrate_vma_insert_page(struct migrate_vma
*migrate
,
2868 struct vm_area_struct
*vma
= migrate
->vma
;
2869 struct mm_struct
*mm
= vma
->vm_mm
;
2879 /* Only allow populating anonymous memory */
2880 if (!vma_is_anonymous(vma
))
2883 pgdp
= pgd_offset(mm
, addr
);
2884 p4dp
= p4d_alloc(mm
, pgdp
, addr
);
2887 pudp
= pud_alloc(mm
, p4dp
, addr
);
2890 pmdp
= pmd_alloc(mm
, pudp
, addr
);
2894 if (pmd_trans_huge(*pmdp
) || pmd_devmap(*pmdp
))
2898 * Use pte_alloc() instead of pte_alloc_map(). We can't run
2899 * pte_offset_map() on pmds where a huge pmd might be created
2900 * from a different thread.
2902 * pte_alloc_map() is safe to use under mmap_write_lock(mm) or when
2903 * parallel threads are excluded by other means.
2905 * Here we only have mmap_read_lock(mm).
2907 if (pte_alloc(mm
, pmdp
))
2910 /* See the comment in pte_alloc_one_map() */
2911 if (unlikely(pmd_trans_unstable(pmdp
)))
2914 if (unlikely(anon_vma_prepare(vma
)))
2916 if (mem_cgroup_charge(page
, vma
->vm_mm
, GFP_KERNEL
))
2920 * The memory barrier inside __SetPageUptodate makes sure that
2921 * preceding stores to the page contents become visible before
2922 * the set_pte_at() write.
2924 __SetPageUptodate(page
);
2926 if (is_zone_device_page(page
)) {
2927 if (is_device_private_page(page
)) {
2928 swp_entry_t swp_entry
;
2930 swp_entry
= make_device_private_entry(page
, vma
->vm_flags
& VM_WRITE
);
2931 entry
= swp_entry_to_pte(swp_entry
);
2934 * For now we only support migrating to un-addressable
2937 pr_warn_once("Unsupported ZONE_DEVICE page type.\n");
2941 entry
= mk_pte(page
, vma
->vm_page_prot
);
2942 if (vma
->vm_flags
& VM_WRITE
)
2943 entry
= pte_mkwrite(pte_mkdirty(entry
));
2946 ptep
= pte_offset_map_lock(mm
, pmdp
, addr
, &ptl
);
2948 if (check_stable_address_space(mm
))
2951 if (pte_present(*ptep
)) {
2952 unsigned long pfn
= pte_pfn(*ptep
);
2954 if (!is_zero_pfn(pfn
))
2957 } else if (!pte_none(*ptep
))
2961 * Check for userfaultfd but do not deliver the fault. Instead,
2964 if (userfaultfd_missing(vma
))
2967 inc_mm_counter(mm
, MM_ANONPAGES
);
2968 page_add_new_anon_rmap(page
, vma
, addr
, false);
2969 if (!is_zone_device_page(page
))
2970 lru_cache_add_inactive_or_unevictable(page
, vma
);
2974 flush_cache_page(vma
, addr
, pte_pfn(*ptep
));
2975 ptep_clear_flush_notify(vma
, addr
, ptep
);
2976 set_pte_at_notify(mm
, addr
, ptep
, entry
);
2977 update_mmu_cache(vma
, addr
, ptep
);
2979 /* No need to invalidate - it was non-present before */
2980 set_pte_at(mm
, addr
, ptep
, entry
);
2981 update_mmu_cache(vma
, addr
, ptep
);
2984 pte_unmap_unlock(ptep
, ptl
);
2985 *src
= MIGRATE_PFN_MIGRATE
;
2989 pte_unmap_unlock(ptep
, ptl
);
2991 *src
&= ~MIGRATE_PFN_MIGRATE
;
2995 * migrate_vma_pages() - migrate meta-data from src page to dst page
2996 * @migrate: migrate struct containing all migration information
2998 * This migrates struct page meta-data from source struct page to destination
2999 * struct page. This effectively finishes the migration from source page to the
3002 void migrate_vma_pages(struct migrate_vma
*migrate
)
3004 const unsigned long npages
= migrate
->npages
;
3005 const unsigned long start
= migrate
->start
;
3006 struct mmu_notifier_range range
;
3007 unsigned long addr
, i
;
3008 bool notified
= false;
3010 for (i
= 0, addr
= start
; i
< npages
; addr
+= PAGE_SIZE
, i
++) {
3011 struct page
*newpage
= migrate_pfn_to_page(migrate
->dst
[i
]);
3012 struct page
*page
= migrate_pfn_to_page(migrate
->src
[i
]);
3013 struct address_space
*mapping
;
3017 migrate
->src
[i
] &= ~MIGRATE_PFN_MIGRATE
;
3022 if (!(migrate
->src
[i
] & MIGRATE_PFN_MIGRATE
))
3027 mmu_notifier_range_init_migrate(&range
, 0,
3028 migrate
->vma
, migrate
->vma
->vm_mm
,
3030 migrate
->pgmap_owner
);
3031 mmu_notifier_invalidate_range_start(&range
);
3033 migrate_vma_insert_page(migrate
, addr
, newpage
,
3038 mapping
= page_mapping(page
);
3040 if (is_zone_device_page(newpage
)) {
3041 if (is_device_private_page(newpage
)) {
3043 * For now only support private anonymous when
3044 * migrating to un-addressable device memory.
3047 migrate
->src
[i
] &= ~MIGRATE_PFN_MIGRATE
;
3052 * Other types of ZONE_DEVICE page are not
3055 migrate
->src
[i
] &= ~MIGRATE_PFN_MIGRATE
;
3060 r
= migrate_page(mapping
, newpage
, page
, MIGRATE_SYNC_NO_COPY
);
3061 if (r
!= MIGRATEPAGE_SUCCESS
)
3062 migrate
->src
[i
] &= ~MIGRATE_PFN_MIGRATE
;
3066 * No need to double call mmu_notifier->invalidate_range() callback as
3067 * the above ptep_clear_flush_notify() inside migrate_vma_insert_page()
3068 * did already call it.
3071 mmu_notifier_invalidate_range_only_end(&range
);
3073 EXPORT_SYMBOL(migrate_vma_pages
);
3076 * migrate_vma_finalize() - restore CPU page table entry
3077 * @migrate: migrate struct containing all migration information
3079 * This replaces the special migration pte entry with either a mapping to the
3080 * new page if migration was successful for that page, or to the original page
3083 * This also unlocks the pages and puts them back on the lru, or drops the extra
3084 * refcount, for device pages.
3086 void migrate_vma_finalize(struct migrate_vma
*migrate
)
3088 const unsigned long npages
= migrate
->npages
;
3091 for (i
= 0; i
< npages
; i
++) {
3092 struct page
*newpage
= migrate_pfn_to_page(migrate
->dst
[i
]);
3093 struct page
*page
= migrate_pfn_to_page(migrate
->src
[i
]);
3097 unlock_page(newpage
);
3103 if (!(migrate
->src
[i
] & MIGRATE_PFN_MIGRATE
) || !newpage
) {
3105 unlock_page(newpage
);
3111 remove_migration_ptes(page
, newpage
, false);
3114 if (is_zone_device_page(page
))
3117 putback_lru_page(page
);
3119 if (newpage
!= page
) {
3120 unlock_page(newpage
);
3121 if (is_zone_device_page(newpage
))
3124 putback_lru_page(newpage
);
3128 EXPORT_SYMBOL(migrate_vma_finalize
);
3129 #endif /* CONFIG_DEVICE_PRIVATE */