1 // SPDX-License-Identifier: GPL-2.0
3 * Memory Migration functionality - linux/mm/migrate.c
5 * Copyright (C) 2006 Silicon Graphics, Inc., Christoph Lameter
7 * Page migration was first developed in the context of the memory hotplug
8 * project. The main authors of the migration code are:
10 * IWAMOTO Toshihiro <iwamoto@valinux.co.jp>
11 * Hirokazu Takahashi <taka@valinux.co.jp>
12 * Dave Hansen <haveblue@us.ibm.com>
16 #include <linux/migrate.h>
17 #include <linux/export.h>
18 #include <linux/swap.h>
19 #include <linux/swapops.h>
20 #include <linux/pagemap.h>
21 #include <linux/buffer_head.h>
22 #include <linux/mm_inline.h>
23 #include <linux/nsproxy.h>
24 #include <linux/pagevec.h>
25 #include <linux/ksm.h>
26 #include <linux/rmap.h>
27 #include <linux/topology.h>
28 #include <linux/cpu.h>
29 #include <linux/cpuset.h>
30 #include <linux/writeback.h>
31 #include <linux/mempolicy.h>
32 #include <linux/vmalloc.h>
33 #include <linux/security.h>
34 #include <linux/backing-dev.h>
35 #include <linux/compaction.h>
36 #include <linux/syscalls.h>
37 #include <linux/compat.h>
38 #include <linux/hugetlb.h>
39 #include <linux/hugetlb_cgroup.h>
40 #include <linux/gfp.h>
41 #include <linux/pagewalk.h>
42 #include <linux/pfn_t.h>
43 #include <linux/memremap.h>
44 #include <linux/userfaultfd_k.h>
45 #include <linux/balloon_compaction.h>
46 #include <linux/mmu_notifier.h>
47 #include <linux/page_idle.h>
48 #include <linux/page_owner.h>
49 #include <linux/sched/mm.h>
50 #include <linux/ptrace.h>
51 #include <linux/oom.h>
53 #include <asm/tlbflush.h>
55 #define CREATE_TRACE_POINTS
56 #include <trace/events/migrate.h>
61 * migrate_prep() needs to be called before we start compiling a list of pages
62 * to be migrated using isolate_lru_page(). If scheduling work on other CPUs is
63 * undesirable, use migrate_prep_local()
65 void migrate_prep(void)
68 * Clear the LRU lists so pages can be isolated.
69 * Note that pages may be moved off the LRU after we have
70 * drained them. Those pages will fail to migrate like other
71 * pages that may be busy.
76 /* Do the necessary work of migrate_prep but not if it involves other CPUs */
77 void migrate_prep_local(void)
82 int isolate_movable_page(struct page
*page
, isolate_mode_t mode
)
84 struct address_space
*mapping
;
87 * Avoid burning cycles with pages that are yet under __free_pages(),
88 * or just got freed under us.
90 * In case we 'win' a race for a movable page being freed under us and
91 * raise its refcount preventing __free_pages() from doing its job
92 * the put_page() at the end of this block will take care of
93 * release this page, thus avoiding a nasty leakage.
95 if (unlikely(!get_page_unless_zero(page
)))
99 * Check PageMovable before holding a PG_lock because page's owner
100 * assumes anybody doesn't touch PG_lock of newly allocated page
101 * so unconditionally grabbing the lock ruins page's owner side.
103 if (unlikely(!__PageMovable(page
)))
106 * As movable pages are not isolated from LRU lists, concurrent
107 * compaction threads can race against page migration functions
108 * as well as race against the releasing a page.
110 * In order to avoid having an already isolated movable page
111 * being (wrongly) re-isolated while it is under migration,
112 * or to avoid attempting to isolate pages being released,
113 * lets be sure we have the page lock
114 * before proceeding with the movable page isolation steps.
116 if (unlikely(!trylock_page(page
)))
119 if (!PageMovable(page
) || PageIsolated(page
))
120 goto out_no_isolated
;
122 mapping
= page_mapping(page
);
123 VM_BUG_ON_PAGE(!mapping
, page
);
125 if (!mapping
->a_ops
->isolate_page(page
, mode
))
126 goto out_no_isolated
;
128 /* Driver shouldn't use PG_isolated bit of page->flags */
129 WARN_ON_ONCE(PageIsolated(page
));
130 __SetPageIsolated(page
);
143 /* It should be called on page which is PG_movable */
144 void putback_movable_page(struct page
*page
)
146 struct address_space
*mapping
;
148 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
149 VM_BUG_ON_PAGE(!PageMovable(page
), page
);
150 VM_BUG_ON_PAGE(!PageIsolated(page
), page
);
152 mapping
= page_mapping(page
);
153 mapping
->a_ops
->putback_page(page
);
154 __ClearPageIsolated(page
);
158 * Put previously isolated pages back onto the appropriate lists
159 * from where they were once taken off for compaction/migration.
161 * This function shall be used whenever the isolated pageset has been
162 * built from lru, balloon, hugetlbfs page. See isolate_migratepages_range()
163 * and isolate_huge_page().
165 void putback_movable_pages(struct list_head
*l
)
170 list_for_each_entry_safe(page
, page2
, l
, lru
) {
171 if (unlikely(PageHuge(page
))) {
172 putback_active_hugepage(page
);
175 list_del(&page
->lru
);
177 * We isolated non-lru movable page so here we can use
178 * __PageMovable because LRU page's mapping cannot have
179 * PAGE_MAPPING_MOVABLE.
181 if (unlikely(__PageMovable(page
))) {
182 VM_BUG_ON_PAGE(!PageIsolated(page
), page
);
184 if (PageMovable(page
))
185 putback_movable_page(page
);
187 __ClearPageIsolated(page
);
191 mod_node_page_state(page_pgdat(page
), NR_ISOLATED_ANON
+
192 page_is_file_lru(page
), -thp_nr_pages(page
));
193 putback_lru_page(page
);
199 * Restore a potential migration pte to a working pte entry
201 static bool remove_migration_pte(struct page
*page
, struct vm_area_struct
*vma
,
202 unsigned long addr
, void *old
)
204 struct page_vma_mapped_walk pvmw
= {
208 .flags
= PVMW_SYNC
| PVMW_MIGRATION
,
214 VM_BUG_ON_PAGE(PageTail(page
), page
);
215 while (page_vma_mapped_walk(&pvmw
)) {
219 new = page
- pvmw
.page
->index
+
220 linear_page_index(vma
, pvmw
.address
);
222 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
223 /* PMD-mapped THP migration entry */
225 VM_BUG_ON_PAGE(PageHuge(page
) || !PageTransCompound(page
), page
);
226 remove_migration_pmd(&pvmw
, new);
232 pte
= pte_mkold(mk_pte(new, READ_ONCE(vma
->vm_page_prot
)));
233 if (pte_swp_soft_dirty(*pvmw
.pte
))
234 pte
= pte_mksoft_dirty(pte
);
237 * Recheck VMA as permissions can change since migration started
239 entry
= pte_to_swp_entry(*pvmw
.pte
);
240 if (is_write_migration_entry(entry
))
241 pte
= maybe_mkwrite(pte
, vma
);
242 else if (pte_swp_uffd_wp(*pvmw
.pte
))
243 pte
= pte_mkuffd_wp(pte
);
245 if (unlikely(is_device_private_page(new))) {
246 entry
= make_device_private_entry(new, pte_write(pte
));
247 pte
= swp_entry_to_pte(entry
);
248 if (pte_swp_soft_dirty(*pvmw
.pte
))
249 pte
= pte_swp_mksoft_dirty(pte
);
250 if (pte_swp_uffd_wp(*pvmw
.pte
))
251 pte
= pte_swp_mkuffd_wp(pte
);
254 #ifdef CONFIG_HUGETLB_PAGE
256 pte
= pte_mkhuge(pte
);
257 pte
= arch_make_huge_pte(pte
, vma
, new, 0);
258 set_huge_pte_at(vma
->vm_mm
, pvmw
.address
, pvmw
.pte
, pte
);
260 hugepage_add_anon_rmap(new, vma
, pvmw
.address
);
262 page_dup_rmap(new, true);
266 set_pte_at(vma
->vm_mm
, pvmw
.address
, pvmw
.pte
, pte
);
269 page_add_anon_rmap(new, vma
, pvmw
.address
, false);
271 page_add_file_rmap(new, false);
273 if (vma
->vm_flags
& VM_LOCKED
&& !PageTransCompound(new))
276 if (PageTransHuge(page
) && PageMlocked(page
))
277 clear_page_mlock(page
);
279 /* No need to invalidate - it was non-present before */
280 update_mmu_cache(vma
, pvmw
.address
, pvmw
.pte
);
287 * Get rid of all migration entries and replace them by
288 * references to the indicated page.
290 void remove_migration_ptes(struct page
*old
, struct page
*new, bool locked
)
292 struct rmap_walk_control rwc
= {
293 .rmap_one
= remove_migration_pte
,
298 rmap_walk_locked(new, &rwc
);
300 rmap_walk(new, &rwc
);
304 * Something used the pte of a page under migration. We need to
305 * get to the page and wait until migration is finished.
306 * When we return from this function the fault will be retried.
308 void __migration_entry_wait(struct mm_struct
*mm
, pte_t
*ptep
,
317 if (!is_swap_pte(pte
))
320 entry
= pte_to_swp_entry(pte
);
321 if (!is_migration_entry(entry
))
324 page
= migration_entry_to_page(entry
);
327 * Once page cache replacement of page migration started, page_count
328 * is zero; but we must not call put_and_wait_on_page_locked() without
329 * a ref. Use get_page_unless_zero(), and just fault again if it fails.
331 if (!get_page_unless_zero(page
))
333 pte_unmap_unlock(ptep
, ptl
);
334 put_and_wait_on_page_locked(page
);
337 pte_unmap_unlock(ptep
, ptl
);
340 void migration_entry_wait(struct mm_struct
*mm
, pmd_t
*pmd
,
341 unsigned long address
)
343 spinlock_t
*ptl
= pte_lockptr(mm
, pmd
);
344 pte_t
*ptep
= pte_offset_map(pmd
, address
);
345 __migration_entry_wait(mm
, ptep
, ptl
);
348 void migration_entry_wait_huge(struct vm_area_struct
*vma
,
349 struct mm_struct
*mm
, pte_t
*pte
)
351 spinlock_t
*ptl
= huge_pte_lockptr(hstate_vma(vma
), mm
, pte
);
352 __migration_entry_wait(mm
, pte
, ptl
);
355 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
356 void pmd_migration_entry_wait(struct mm_struct
*mm
, pmd_t
*pmd
)
361 ptl
= pmd_lock(mm
, pmd
);
362 if (!is_pmd_migration_entry(*pmd
))
364 page
= migration_entry_to_page(pmd_to_swp_entry(*pmd
));
365 if (!get_page_unless_zero(page
))
368 put_and_wait_on_page_locked(page
);
375 static int expected_page_refs(struct address_space
*mapping
, struct page
*page
)
377 int expected_count
= 1;
380 * Device private pages have an extra refcount as they are
383 expected_count
+= is_device_private_page(page
);
385 expected_count
+= thp_nr_pages(page
) + page_has_private(page
);
387 return expected_count
;
391 * Replace the page in the mapping.
393 * The number of remaining references must be:
394 * 1 for anonymous pages without a mapping
395 * 2 for pages with a mapping
396 * 3 for pages with a mapping and PagePrivate/PagePrivate2 set.
398 int migrate_page_move_mapping(struct address_space
*mapping
,
399 struct page
*newpage
, struct page
*page
, int extra_count
)
401 XA_STATE(xas
, &mapping
->i_pages
, page_index(page
));
402 struct zone
*oldzone
, *newzone
;
404 int expected_count
= expected_page_refs(mapping
, page
) + extra_count
;
405 int nr
= thp_nr_pages(page
);
408 /* Anonymous page without mapping */
409 if (page_count(page
) != expected_count
)
412 /* No turning back from here */
413 newpage
->index
= page
->index
;
414 newpage
->mapping
= page
->mapping
;
415 if (PageSwapBacked(page
))
416 __SetPageSwapBacked(newpage
);
418 return MIGRATEPAGE_SUCCESS
;
421 oldzone
= page_zone(page
);
422 newzone
= page_zone(newpage
);
425 if (page_count(page
) != expected_count
|| xas_load(&xas
) != page
) {
426 xas_unlock_irq(&xas
);
430 if (!page_ref_freeze(page
, expected_count
)) {
431 xas_unlock_irq(&xas
);
436 * Now we know that no one else is looking at the page:
437 * no turning back from here.
439 newpage
->index
= page
->index
;
440 newpage
->mapping
= page
->mapping
;
441 page_ref_add(newpage
, nr
); /* add cache reference */
442 if (PageSwapBacked(page
)) {
443 __SetPageSwapBacked(newpage
);
444 if (PageSwapCache(page
)) {
445 SetPageSwapCache(newpage
);
446 set_page_private(newpage
, page_private(page
));
449 VM_BUG_ON_PAGE(PageSwapCache(page
), page
);
452 /* Move dirty while page refs frozen and newpage not yet exposed */
453 dirty
= PageDirty(page
);
455 ClearPageDirty(page
);
456 SetPageDirty(newpage
);
459 xas_store(&xas
, newpage
);
460 if (PageTransHuge(page
)) {
463 for (i
= 1; i
< nr
; i
++) {
465 xas_store(&xas
, newpage
);
470 * Drop cache reference from old page by unfreezing
471 * to one less reference.
472 * We know this isn't the last reference.
474 page_ref_unfreeze(page
, expected_count
- nr
);
477 /* Leave irq disabled to prevent preemption while updating stats */
480 * If moved to a different zone then also account
481 * the page for that zone. Other VM counters will be
482 * taken care of when we establish references to the
483 * new page and drop references to the old page.
485 * Note that anonymous pages are accounted for
486 * via NR_FILE_PAGES and NR_ANON_MAPPED if they
487 * are mapped to swap space.
489 if (newzone
!= oldzone
) {
490 struct lruvec
*old_lruvec
, *new_lruvec
;
491 struct mem_cgroup
*memcg
;
493 memcg
= page_memcg(page
);
494 old_lruvec
= mem_cgroup_lruvec(memcg
, oldzone
->zone_pgdat
);
495 new_lruvec
= mem_cgroup_lruvec(memcg
, newzone
->zone_pgdat
);
497 __mod_lruvec_state(old_lruvec
, NR_FILE_PAGES
, -nr
);
498 __mod_lruvec_state(new_lruvec
, NR_FILE_PAGES
, nr
);
499 if (PageSwapBacked(page
) && !PageSwapCache(page
)) {
500 __mod_lruvec_state(old_lruvec
, NR_SHMEM
, -nr
);
501 __mod_lruvec_state(new_lruvec
, NR_SHMEM
, nr
);
503 if (dirty
&& mapping_can_writeback(mapping
)) {
504 __mod_lruvec_state(old_lruvec
, NR_FILE_DIRTY
, -nr
);
505 __mod_zone_page_state(oldzone
, NR_ZONE_WRITE_PENDING
, -nr
);
506 __mod_lruvec_state(new_lruvec
, NR_FILE_DIRTY
, nr
);
507 __mod_zone_page_state(newzone
, NR_ZONE_WRITE_PENDING
, nr
);
512 return MIGRATEPAGE_SUCCESS
;
514 EXPORT_SYMBOL(migrate_page_move_mapping
);
517 * The expected number of remaining references is the same as that
518 * of migrate_page_move_mapping().
520 int migrate_huge_page_move_mapping(struct address_space
*mapping
,
521 struct page
*newpage
, struct page
*page
)
523 XA_STATE(xas
, &mapping
->i_pages
, page_index(page
));
527 expected_count
= 2 + page_has_private(page
);
528 if (page_count(page
) != expected_count
|| xas_load(&xas
) != page
) {
529 xas_unlock_irq(&xas
);
533 if (!page_ref_freeze(page
, expected_count
)) {
534 xas_unlock_irq(&xas
);
538 newpage
->index
= page
->index
;
539 newpage
->mapping
= page
->mapping
;
543 xas_store(&xas
, newpage
);
545 page_ref_unfreeze(page
, expected_count
- 1);
547 xas_unlock_irq(&xas
);
549 return MIGRATEPAGE_SUCCESS
;
553 * Gigantic pages are so large that we do not guarantee that page++ pointer
554 * arithmetic will work across the entire page. We need something more
557 static void __copy_gigantic_page(struct page
*dst
, struct page
*src
,
561 struct page
*dst_base
= dst
;
562 struct page
*src_base
= src
;
564 for (i
= 0; i
< nr_pages
; ) {
566 copy_highpage(dst
, src
);
569 dst
= mem_map_next(dst
, dst_base
, i
);
570 src
= mem_map_next(src
, src_base
, i
);
574 static void copy_huge_page(struct page
*dst
, struct page
*src
)
581 struct hstate
*h
= page_hstate(src
);
582 nr_pages
= pages_per_huge_page(h
);
584 if (unlikely(nr_pages
> MAX_ORDER_NR_PAGES
)) {
585 __copy_gigantic_page(dst
, src
, nr_pages
);
590 BUG_ON(!PageTransHuge(src
));
591 nr_pages
= thp_nr_pages(src
);
594 for (i
= 0; i
< nr_pages
; i
++) {
596 copy_highpage(dst
+ i
, src
+ i
);
601 * Copy the page to its new location
603 void migrate_page_states(struct page
*newpage
, struct page
*page
)
608 SetPageError(newpage
);
609 if (PageReferenced(page
))
610 SetPageReferenced(newpage
);
611 if (PageUptodate(page
))
612 SetPageUptodate(newpage
);
613 if (TestClearPageActive(page
)) {
614 VM_BUG_ON_PAGE(PageUnevictable(page
), page
);
615 SetPageActive(newpage
);
616 } else if (TestClearPageUnevictable(page
))
617 SetPageUnevictable(newpage
);
618 if (PageWorkingset(page
))
619 SetPageWorkingset(newpage
);
620 if (PageChecked(page
))
621 SetPageChecked(newpage
);
622 if (PageMappedToDisk(page
))
623 SetPageMappedToDisk(newpage
);
625 /* Move dirty on pages not done by migrate_page_move_mapping() */
627 SetPageDirty(newpage
);
629 if (page_is_young(page
))
630 set_page_young(newpage
);
631 if (page_is_idle(page
))
632 set_page_idle(newpage
);
635 * Copy NUMA information to the new page, to prevent over-eager
636 * future migrations of this same page.
638 cpupid
= page_cpupid_xchg_last(page
, -1);
639 page_cpupid_xchg_last(newpage
, cpupid
);
641 ksm_migrate_page(newpage
, page
);
643 * Please do not reorder this without considering how mm/ksm.c's
644 * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache().
646 if (PageSwapCache(page
))
647 ClearPageSwapCache(page
);
648 ClearPagePrivate(page
);
649 set_page_private(page
, 0);
652 * If any waiters have accumulated on the new page then
655 if (PageWriteback(newpage
))
656 end_page_writeback(newpage
);
659 * PG_readahead shares the same bit with PG_reclaim. The above
660 * end_page_writeback() may clear PG_readahead mistakenly, so set the
663 if (PageReadahead(page
))
664 SetPageReadahead(newpage
);
666 copy_page_owner(page
, newpage
);
669 mem_cgroup_migrate(page
, newpage
);
671 EXPORT_SYMBOL(migrate_page_states
);
673 void migrate_page_copy(struct page
*newpage
, struct page
*page
)
675 if (PageHuge(page
) || PageTransHuge(page
))
676 copy_huge_page(newpage
, page
);
678 copy_highpage(newpage
, page
);
680 migrate_page_states(newpage
, page
);
682 EXPORT_SYMBOL(migrate_page_copy
);
684 /************************************************************
685 * Migration functions
686 ***********************************************************/
689 * Common logic to directly migrate a single LRU page suitable for
690 * pages that do not use PagePrivate/PagePrivate2.
692 * Pages are locked upon entry and exit.
694 int migrate_page(struct address_space
*mapping
,
695 struct page
*newpage
, struct page
*page
,
696 enum migrate_mode mode
)
700 BUG_ON(PageWriteback(page
)); /* Writeback must be complete */
702 rc
= migrate_page_move_mapping(mapping
, newpage
, page
, 0);
704 if (rc
!= MIGRATEPAGE_SUCCESS
)
707 if (mode
!= MIGRATE_SYNC_NO_COPY
)
708 migrate_page_copy(newpage
, page
);
710 migrate_page_states(newpage
, page
);
711 return MIGRATEPAGE_SUCCESS
;
713 EXPORT_SYMBOL(migrate_page
);
716 /* Returns true if all buffers are successfully locked */
717 static bool buffer_migrate_lock_buffers(struct buffer_head
*head
,
718 enum migrate_mode mode
)
720 struct buffer_head
*bh
= head
;
722 /* Simple case, sync compaction */
723 if (mode
!= MIGRATE_ASYNC
) {
726 bh
= bh
->b_this_page
;
728 } while (bh
!= head
);
733 /* async case, we cannot block on lock_buffer so use trylock_buffer */
735 if (!trylock_buffer(bh
)) {
737 * We failed to lock the buffer and cannot stall in
738 * async migration. Release the taken locks
740 struct buffer_head
*failed_bh
= bh
;
742 while (bh
!= failed_bh
) {
744 bh
= bh
->b_this_page
;
749 bh
= bh
->b_this_page
;
750 } while (bh
!= head
);
754 static int __buffer_migrate_page(struct address_space
*mapping
,
755 struct page
*newpage
, struct page
*page
, enum migrate_mode mode
,
758 struct buffer_head
*bh
, *head
;
762 if (!page_has_buffers(page
))
763 return migrate_page(mapping
, newpage
, page
, mode
);
765 /* Check whether page does not have extra refs before we do more work */
766 expected_count
= expected_page_refs(mapping
, page
);
767 if (page_count(page
) != expected_count
)
770 head
= page_buffers(page
);
771 if (!buffer_migrate_lock_buffers(head
, mode
))
776 bool invalidated
= false;
780 spin_lock(&mapping
->private_lock
);
783 if (atomic_read(&bh
->b_count
)) {
787 bh
= bh
->b_this_page
;
788 } while (bh
!= head
);
794 spin_unlock(&mapping
->private_lock
);
795 invalidate_bh_lrus();
797 goto recheck_buffers
;
801 rc
= migrate_page_move_mapping(mapping
, newpage
, page
, 0);
802 if (rc
!= MIGRATEPAGE_SUCCESS
)
805 attach_page_private(newpage
, detach_page_private(page
));
809 set_bh_page(bh
, newpage
, bh_offset(bh
));
810 bh
= bh
->b_this_page
;
812 } while (bh
!= head
);
814 if (mode
!= MIGRATE_SYNC_NO_COPY
)
815 migrate_page_copy(newpage
, page
);
817 migrate_page_states(newpage
, page
);
819 rc
= MIGRATEPAGE_SUCCESS
;
822 spin_unlock(&mapping
->private_lock
);
826 bh
= bh
->b_this_page
;
828 } while (bh
!= head
);
834 * Migration function for pages with buffers. This function can only be used
835 * if the underlying filesystem guarantees that no other references to "page"
836 * exist. For example attached buffer heads are accessed only under page lock.
838 int buffer_migrate_page(struct address_space
*mapping
,
839 struct page
*newpage
, struct page
*page
, enum migrate_mode mode
)
841 return __buffer_migrate_page(mapping
, newpage
, page
, mode
, false);
843 EXPORT_SYMBOL(buffer_migrate_page
);
846 * Same as above except that this variant is more careful and checks that there
847 * are also no buffer head references. This function is the right one for
848 * mappings where buffer heads are directly looked up and referenced (such as
849 * block device mappings).
851 int buffer_migrate_page_norefs(struct address_space
*mapping
,
852 struct page
*newpage
, struct page
*page
, enum migrate_mode mode
)
854 return __buffer_migrate_page(mapping
, newpage
, page
, mode
, true);
859 * Writeback a page to clean the dirty state
861 static int writeout(struct address_space
*mapping
, struct page
*page
)
863 struct writeback_control wbc
= {
864 .sync_mode
= WB_SYNC_NONE
,
867 .range_end
= LLONG_MAX
,
872 if (!mapping
->a_ops
->writepage
)
873 /* No write method for the address space */
876 if (!clear_page_dirty_for_io(page
))
877 /* Someone else already triggered a write */
881 * A dirty page may imply that the underlying filesystem has
882 * the page on some queue. So the page must be clean for
883 * migration. Writeout may mean we loose the lock and the
884 * page state is no longer what we checked for earlier.
885 * At this point we know that the migration attempt cannot
888 remove_migration_ptes(page
, page
, false);
890 rc
= mapping
->a_ops
->writepage(page
, &wbc
);
892 if (rc
!= AOP_WRITEPAGE_ACTIVATE
)
893 /* unlocked. Relock */
896 return (rc
< 0) ? -EIO
: -EAGAIN
;
900 * Default handling if a filesystem does not provide a migration function.
902 static int fallback_migrate_page(struct address_space
*mapping
,
903 struct page
*newpage
, struct page
*page
, enum migrate_mode mode
)
905 if (PageDirty(page
)) {
906 /* Only writeback pages in full synchronous migration */
909 case MIGRATE_SYNC_NO_COPY
:
914 return writeout(mapping
, page
);
918 * Buffers may be managed in a filesystem specific way.
919 * We must have no buffers or drop them.
921 if (page_has_private(page
) &&
922 !try_to_release_page(page
, GFP_KERNEL
))
923 return mode
== MIGRATE_SYNC
? -EAGAIN
: -EBUSY
;
925 return migrate_page(mapping
, newpage
, page
, mode
);
929 * Move a page to a newly allocated page
930 * The page is locked and all ptes have been successfully removed.
932 * The new page will have replaced the old page if this function
937 * MIGRATEPAGE_SUCCESS - success
939 static int move_to_new_page(struct page
*newpage
, struct page
*page
,
940 enum migrate_mode mode
)
942 struct address_space
*mapping
;
944 bool is_lru
= !__PageMovable(page
);
946 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
947 VM_BUG_ON_PAGE(!PageLocked(newpage
), newpage
);
949 mapping
= page_mapping(page
);
951 if (likely(is_lru
)) {
953 rc
= migrate_page(mapping
, newpage
, page
, mode
);
954 else if (mapping
->a_ops
->migratepage
)
956 * Most pages have a mapping and most filesystems
957 * provide a migratepage callback. Anonymous pages
958 * are part of swap space which also has its own
959 * migratepage callback. This is the most common path
960 * for page migration.
962 rc
= mapping
->a_ops
->migratepage(mapping
, newpage
,
965 rc
= fallback_migrate_page(mapping
, newpage
,
969 * In case of non-lru page, it could be released after
970 * isolation step. In that case, we shouldn't try migration.
972 VM_BUG_ON_PAGE(!PageIsolated(page
), page
);
973 if (!PageMovable(page
)) {
974 rc
= MIGRATEPAGE_SUCCESS
;
975 __ClearPageIsolated(page
);
979 rc
= mapping
->a_ops
->migratepage(mapping
, newpage
,
981 WARN_ON_ONCE(rc
== MIGRATEPAGE_SUCCESS
&&
982 !PageIsolated(page
));
986 * When successful, old pagecache page->mapping must be cleared before
987 * page is freed; but stats require that PageAnon be left as PageAnon.
989 if (rc
== MIGRATEPAGE_SUCCESS
) {
990 if (__PageMovable(page
)) {
991 VM_BUG_ON_PAGE(!PageIsolated(page
), page
);
994 * We clear PG_movable under page_lock so any compactor
995 * cannot try to migrate this page.
997 __ClearPageIsolated(page
);
1001 * Anonymous and movable page->mapping will be cleared by
1002 * free_pages_prepare so don't reset it here for keeping
1003 * the type to work PageAnon, for example.
1005 if (!PageMappingFlags(page
))
1006 page
->mapping
= NULL
;
1008 if (likely(!is_zone_device_page(newpage
)))
1009 flush_dcache_page(newpage
);
1016 static int __unmap_and_move(struct page
*page
, struct page
*newpage
,
1017 int force
, enum migrate_mode mode
)
1020 int page_was_mapped
= 0;
1021 struct anon_vma
*anon_vma
= NULL
;
1022 bool is_lru
= !__PageMovable(page
);
1024 if (!trylock_page(page
)) {
1025 if (!force
|| mode
== MIGRATE_ASYNC
)
1029 * It's not safe for direct compaction to call lock_page.
1030 * For example, during page readahead pages are added locked
1031 * to the LRU. Later, when the IO completes the pages are
1032 * marked uptodate and unlocked. However, the queueing
1033 * could be merging multiple pages for one bio (e.g.
1034 * mpage_readahead). If an allocation happens for the
1035 * second or third page, the process can end up locking
1036 * the same page twice and deadlocking. Rather than
1037 * trying to be clever about what pages can be locked,
1038 * avoid the use of lock_page for direct compaction
1041 if (current
->flags
& PF_MEMALLOC
)
1047 if (PageWriteback(page
)) {
1049 * Only in the case of a full synchronous migration is it
1050 * necessary to wait for PageWriteback. In the async case,
1051 * the retry loop is too short and in the sync-light case,
1052 * the overhead of stalling is too much
1056 case MIGRATE_SYNC_NO_COPY
:
1064 wait_on_page_writeback(page
);
1068 * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
1069 * we cannot notice that anon_vma is freed while we migrates a page.
1070 * This get_anon_vma() delays freeing anon_vma pointer until the end
1071 * of migration. File cache pages are no problem because of page_lock()
1072 * File Caches may use write_page() or lock_page() in migration, then,
1073 * just care Anon page here.
1075 * Only page_get_anon_vma() understands the subtleties of
1076 * getting a hold on an anon_vma from outside one of its mms.
1077 * But if we cannot get anon_vma, then we won't need it anyway,
1078 * because that implies that the anon page is no longer mapped
1079 * (and cannot be remapped so long as we hold the page lock).
1081 if (PageAnon(page
) && !PageKsm(page
))
1082 anon_vma
= page_get_anon_vma(page
);
1085 * Block others from accessing the new page when we get around to
1086 * establishing additional references. We are usually the only one
1087 * holding a reference to newpage at this point. We used to have a BUG
1088 * here if trylock_page(newpage) fails, but would like to allow for
1089 * cases where there might be a race with the previous use of newpage.
1090 * This is much like races on refcount of oldpage: just don't BUG().
1092 if (unlikely(!trylock_page(newpage
)))
1095 if (unlikely(!is_lru
)) {
1096 rc
= move_to_new_page(newpage
, page
, mode
);
1097 goto out_unlock_both
;
1101 * Corner case handling:
1102 * 1. When a new swap-cache page is read into, it is added to the LRU
1103 * and treated as swapcache but it has no rmap yet.
1104 * Calling try_to_unmap() against a page->mapping==NULL page will
1105 * trigger a BUG. So handle it here.
1106 * 2. An orphaned page (see truncate_cleanup_page) might have
1107 * fs-private metadata. The page can be picked up due to memory
1108 * offlining. Everywhere else except page reclaim, the page is
1109 * invisible to the vm, so the page can not be migrated. So try to
1110 * free the metadata, so the page can be freed.
1112 if (!page
->mapping
) {
1113 VM_BUG_ON_PAGE(PageAnon(page
), page
);
1114 if (page_has_private(page
)) {
1115 try_to_free_buffers(page
);
1116 goto out_unlock_both
;
1118 } else if (page_mapped(page
)) {
1119 /* Establish migration ptes */
1120 VM_BUG_ON_PAGE(PageAnon(page
) && !PageKsm(page
) && !anon_vma
,
1122 try_to_unmap(page
, TTU_MIGRATION
|TTU_IGNORE_MLOCK
);
1123 page_was_mapped
= 1;
1126 if (!page_mapped(page
))
1127 rc
= move_to_new_page(newpage
, page
, mode
);
1129 if (page_was_mapped
)
1130 remove_migration_ptes(page
,
1131 rc
== MIGRATEPAGE_SUCCESS
? newpage
: page
, false);
1134 unlock_page(newpage
);
1136 /* Drop an anon_vma reference if we took one */
1138 put_anon_vma(anon_vma
);
1142 * If migration is successful, decrease refcount of the newpage
1143 * which will not free the page because new page owner increased
1144 * refcounter. As well, if it is LRU page, add the page to LRU
1145 * list in here. Use the old state of the isolated source page to
1146 * determine if we migrated a LRU page. newpage was already unlocked
1147 * and possibly modified by its owner - don't rely on the page
1150 if (rc
== MIGRATEPAGE_SUCCESS
) {
1151 if (unlikely(!is_lru
))
1154 putback_lru_page(newpage
);
1161 * Obtain the lock on page, remove all ptes and migrate the page
1162 * to the newly allocated page in newpage.
1164 static int unmap_and_move(new_page_t get_new_page
,
1165 free_page_t put_new_page
,
1166 unsigned long private, struct page
*page
,
1167 int force
, enum migrate_mode mode
,
1168 enum migrate_reason reason
,
1169 struct list_head
*ret
)
1171 int rc
= MIGRATEPAGE_SUCCESS
;
1172 struct page
*newpage
= NULL
;
1174 if (!thp_migration_supported() && PageTransHuge(page
))
1177 if (page_count(page
) == 1) {
1178 /* page was freed from under us. So we are done. */
1179 ClearPageActive(page
);
1180 ClearPageUnevictable(page
);
1181 if (unlikely(__PageMovable(page
))) {
1183 if (!PageMovable(page
))
1184 __ClearPageIsolated(page
);
1190 newpage
= get_new_page(page
, private);
1194 rc
= __unmap_and_move(page
, newpage
, force
, mode
);
1195 if (rc
== MIGRATEPAGE_SUCCESS
)
1196 set_page_owner_migrate_reason(newpage
, reason
);
1199 if (rc
!= -EAGAIN
) {
1201 * A page that has been migrated has all references
1202 * removed and will be freed. A page that has not been
1203 * migrated will have kept its references and be restored.
1205 list_del(&page
->lru
);
1209 * If migration is successful, releases reference grabbed during
1210 * isolation. Otherwise, restore the page to right list unless
1213 if (rc
== MIGRATEPAGE_SUCCESS
) {
1215 * Compaction can migrate also non-LRU pages which are
1216 * not accounted to NR_ISOLATED_*. They can be recognized
1219 if (likely(!__PageMovable(page
)))
1220 mod_node_page_state(page_pgdat(page
), NR_ISOLATED_ANON
+
1221 page_is_file_lru(page
), -thp_nr_pages(page
));
1223 if (reason
!= MR_MEMORY_FAILURE
)
1225 * We release the page in page_handle_poison.
1230 list_add_tail(&page
->lru
, ret
);
1233 put_new_page(newpage
, private);
1242 * Counterpart of unmap_and_move_page() for hugepage migration.
1244 * This function doesn't wait the completion of hugepage I/O
1245 * because there is no race between I/O and migration for hugepage.
1246 * Note that currently hugepage I/O occurs only in direct I/O
1247 * where no lock is held and PG_writeback is irrelevant,
1248 * and writeback status of all subpages are counted in the reference
1249 * count of the head page (i.e. if all subpages of a 2MB hugepage are
1250 * under direct I/O, the reference of the head page is 512 and a bit more.)
1251 * This means that when we try to migrate hugepage whose subpages are
1252 * doing direct I/O, some references remain after try_to_unmap() and
1253 * hugepage migration fails without data corruption.
1255 * There is also no race when direct I/O is issued on the page under migration,
1256 * because then pte is replaced with migration swap entry and direct I/O code
1257 * will wait in the page fault for migration to complete.
1259 static int unmap_and_move_huge_page(new_page_t get_new_page
,
1260 free_page_t put_new_page
, unsigned long private,
1261 struct page
*hpage
, int force
,
1262 enum migrate_mode mode
, int reason
,
1263 struct list_head
*ret
)
1266 int page_was_mapped
= 0;
1267 struct page
*new_hpage
;
1268 struct anon_vma
*anon_vma
= NULL
;
1269 struct address_space
*mapping
= NULL
;
1272 * Migratability of hugepages depends on architectures and their size.
1273 * This check is necessary because some callers of hugepage migration
1274 * like soft offline and memory hotremove don't walk through page
1275 * tables or check whether the hugepage is pmd-based or not before
1276 * kicking migration.
1278 if (!hugepage_migration_supported(page_hstate(hpage
))) {
1279 list_move_tail(&hpage
->lru
, ret
);
1283 if (page_count(hpage
) == 1) {
1284 /* page was freed from under us. So we are done. */
1285 putback_active_hugepage(hpage
);
1286 return MIGRATEPAGE_SUCCESS
;
1289 new_hpage
= get_new_page(hpage
, private);
1293 if (!trylock_page(hpage
)) {
1298 case MIGRATE_SYNC_NO_COPY
:
1307 * Check for pages which are in the process of being freed. Without
1308 * page_mapping() set, hugetlbfs specific move page routine will not
1309 * be called and we could leak usage counts for subpools.
1311 if (page_private(hpage
) && !page_mapping(hpage
)) {
1316 if (PageAnon(hpage
))
1317 anon_vma
= page_get_anon_vma(hpage
);
1319 if (unlikely(!trylock_page(new_hpage
)))
1322 if (page_mapped(hpage
)) {
1323 bool mapping_locked
= false;
1324 enum ttu_flags ttu
= TTU_MIGRATION
|TTU_IGNORE_MLOCK
;
1326 if (!PageAnon(hpage
)) {
1328 * In shared mappings, try_to_unmap could potentially
1329 * call huge_pmd_unshare. Because of this, take
1330 * semaphore in write mode here and set TTU_RMAP_LOCKED
1331 * to let lower levels know we have taken the lock.
1333 mapping
= hugetlb_page_mapping_lock_write(hpage
);
1334 if (unlikely(!mapping
))
1335 goto unlock_put_anon
;
1337 mapping_locked
= true;
1338 ttu
|= TTU_RMAP_LOCKED
;
1341 try_to_unmap(hpage
, ttu
);
1342 page_was_mapped
= 1;
1345 i_mmap_unlock_write(mapping
);
1348 if (!page_mapped(hpage
))
1349 rc
= move_to_new_page(new_hpage
, hpage
, mode
);
1351 if (page_was_mapped
)
1352 remove_migration_ptes(hpage
,
1353 rc
== MIGRATEPAGE_SUCCESS
? new_hpage
: hpage
, false);
1356 unlock_page(new_hpage
);
1360 put_anon_vma(anon_vma
);
1362 if (rc
== MIGRATEPAGE_SUCCESS
) {
1363 move_hugetlb_state(hpage
, new_hpage
, reason
);
1364 put_new_page
= NULL
;
1370 if (rc
== MIGRATEPAGE_SUCCESS
)
1371 putback_active_hugepage(hpage
);
1372 else if (rc
!= -EAGAIN
&& rc
!= MIGRATEPAGE_SUCCESS
)
1373 list_move_tail(&hpage
->lru
, ret
);
1376 * If migration was not successful and there's a freeing callback, use
1377 * it. Otherwise, put_page() will drop the reference grabbed during
1381 put_new_page(new_hpage
, private);
1383 putback_active_hugepage(new_hpage
);
1388 static inline int try_split_thp(struct page
*page
, struct page
**page2
,
1389 struct list_head
*from
)
1394 rc
= split_huge_page_to_list(page
, from
);
1397 list_safe_reset_next(page
, *page2
, lru
);
1403 * migrate_pages - migrate the pages specified in a list, to the free pages
1404 * supplied as the target for the page migration
1406 * @from: The list of pages to be migrated.
1407 * @get_new_page: The function used to allocate free pages to be used
1408 * as the target of the page migration.
1409 * @put_new_page: The function used to free target pages if migration
1410 * fails, or NULL if no special handling is necessary.
1411 * @private: Private data to be passed on to get_new_page()
1412 * @mode: The migration mode that specifies the constraints for
1413 * page migration, if any.
1414 * @reason: The reason for page migration.
1416 * The function returns after 10 attempts or if no pages are movable any more
1417 * because the list has become empty or no retryable pages exist any more.
1418 * It is caller's responsibility to call putback_movable_pages() to return pages
1419 * to the LRU or free list only if ret != 0.
1421 * Returns the number of pages that were not migrated, or an error code.
1423 int migrate_pages(struct list_head
*from
, new_page_t get_new_page
,
1424 free_page_t put_new_page
, unsigned long private,
1425 enum migrate_mode mode
, int reason
)
1430 int nr_succeeded
= 0;
1431 int nr_thp_succeeded
= 0;
1432 int nr_thp_failed
= 0;
1433 int nr_thp_split
= 0;
1435 bool is_thp
= false;
1438 int swapwrite
= current
->flags
& PF_SWAPWRITE
;
1439 int rc
, nr_subpages
;
1440 LIST_HEAD(ret_pages
);
1443 current
->flags
|= PF_SWAPWRITE
;
1445 for (pass
= 0; pass
< 10 && (retry
|| thp_retry
); pass
++) {
1449 list_for_each_entry_safe(page
, page2
, from
, lru
) {
1452 * THP statistics is based on the source huge page.
1453 * Capture required information that might get lost
1456 is_thp
= PageTransHuge(page
) && !PageHuge(page
);
1457 nr_subpages
= thp_nr_pages(page
);
1461 rc
= unmap_and_move_huge_page(get_new_page
,
1462 put_new_page
, private, page
,
1463 pass
> 2, mode
, reason
,
1466 rc
= unmap_and_move(get_new_page
, put_new_page
,
1467 private, page
, pass
> 2, mode
,
1468 reason
, &ret_pages
);
1471 * Success: non hugetlb page will be freed, hugetlb
1472 * page will be put back
1473 * -EAGAIN: stay on the from list
1474 * -ENOMEM: stay on the from list
1475 * Other errno: put on ret_pages list then splice to
1480 * THP migration might be unsupported or the
1481 * allocation could've failed so we should
1482 * retry on the same page with the THP split
1485 * Head page is retried immediately and tail
1486 * pages are added to the tail of the list so
1487 * we encounter them after the rest of the list
1491 /* THP migration is unsupported */
1493 if (!try_split_thp(page
, &page2
, from
)) {
1499 nr_failed
+= nr_subpages
;
1503 /* Hugetlb migration is unsupported */
1508 * When memory is low, don't bother to try to migrate
1509 * other pages, just exit.
1512 if (!try_split_thp(page
, &page2
, from
)) {
1518 nr_failed
+= nr_subpages
;
1530 case MIGRATEPAGE_SUCCESS
:
1533 nr_succeeded
+= nr_subpages
;
1540 * Permanent failure (-EBUSY, etc.):
1541 * unlike -EAGAIN case, the failed page is
1542 * removed from migration page list and not
1543 * retried in the next outer loop.
1547 nr_failed
+= nr_subpages
;
1555 nr_failed
+= retry
+ thp_retry
;
1556 nr_thp_failed
+= thp_retry
;
1560 * Put the permanent failure page back to migration list, they
1561 * will be put back to the right list by the caller.
1563 list_splice(&ret_pages
, from
);
1565 count_vm_events(PGMIGRATE_SUCCESS
, nr_succeeded
);
1566 count_vm_events(PGMIGRATE_FAIL
, nr_failed
);
1567 count_vm_events(THP_MIGRATION_SUCCESS
, nr_thp_succeeded
);
1568 count_vm_events(THP_MIGRATION_FAIL
, nr_thp_failed
);
1569 count_vm_events(THP_MIGRATION_SPLIT
, nr_thp_split
);
1570 trace_mm_migrate_pages(nr_succeeded
, nr_failed
, nr_thp_succeeded
,
1571 nr_thp_failed
, nr_thp_split
, mode
, reason
);
1574 current
->flags
&= ~PF_SWAPWRITE
;
1579 struct page
*alloc_migration_target(struct page
*page
, unsigned long private)
1581 struct migration_target_control
*mtc
;
1583 unsigned int order
= 0;
1584 struct page
*new_page
= NULL
;
1588 mtc
= (struct migration_target_control
*)private;
1589 gfp_mask
= mtc
->gfp_mask
;
1591 if (nid
== NUMA_NO_NODE
)
1592 nid
= page_to_nid(page
);
1594 if (PageHuge(page
)) {
1595 struct hstate
*h
= page_hstate(compound_head(page
));
1597 gfp_mask
= htlb_modify_alloc_mask(h
, gfp_mask
);
1598 return alloc_huge_page_nodemask(h
, nid
, mtc
->nmask
, gfp_mask
);
1601 if (PageTransHuge(page
)) {
1603 * clear __GFP_RECLAIM to make the migration callback
1604 * consistent with regular THP allocations.
1606 gfp_mask
&= ~__GFP_RECLAIM
;
1607 gfp_mask
|= GFP_TRANSHUGE
;
1608 order
= HPAGE_PMD_ORDER
;
1610 zidx
= zone_idx(page_zone(page
));
1611 if (is_highmem_idx(zidx
) || zidx
== ZONE_MOVABLE
)
1612 gfp_mask
|= __GFP_HIGHMEM
;
1614 new_page
= __alloc_pages_nodemask(gfp_mask
, order
, nid
, mtc
->nmask
);
1616 if (new_page
&& PageTransHuge(new_page
))
1617 prep_transhuge_page(new_page
);
1624 static int store_status(int __user
*status
, int start
, int value
, int nr
)
1627 if (put_user(value
, status
+ start
))
1635 static int do_move_pages_to_node(struct mm_struct
*mm
,
1636 struct list_head
*pagelist
, int node
)
1639 struct migration_target_control mtc
= {
1641 .gfp_mask
= GFP_HIGHUSER_MOVABLE
| __GFP_THISNODE
,
1644 err
= migrate_pages(pagelist
, alloc_migration_target
, NULL
,
1645 (unsigned long)&mtc
, MIGRATE_SYNC
, MR_SYSCALL
);
1647 putback_movable_pages(pagelist
);
1652 * Resolves the given address to a struct page, isolates it from the LRU and
1653 * puts it to the given pagelist.
1655 * errno - if the page cannot be found/isolated
1656 * 0 - when it doesn't have to be migrated because it is already on the
1658 * 1 - when it has been queued
1660 static int add_page_for_migration(struct mm_struct
*mm
, unsigned long addr
,
1661 int node
, struct list_head
*pagelist
, bool migrate_all
)
1663 struct vm_area_struct
*vma
;
1665 unsigned int follflags
;
1670 vma
= find_vma(mm
, addr
);
1671 if (!vma
|| addr
< vma
->vm_start
|| !vma_migratable(vma
))
1674 /* FOLL_DUMP to ignore special (like zero) pages */
1675 follflags
= FOLL_GET
| FOLL_DUMP
;
1676 page
= follow_page(vma
, addr
, follflags
);
1678 err
= PTR_ERR(page
);
1687 if (page_to_nid(page
) == node
)
1691 if (page_mapcount(page
) > 1 && !migrate_all
)
1694 if (PageHuge(page
)) {
1695 if (PageHead(page
)) {
1696 isolate_huge_page(page
, pagelist
);
1702 head
= compound_head(page
);
1703 err
= isolate_lru_page(head
);
1708 list_add_tail(&head
->lru
, pagelist
);
1709 mod_node_page_state(page_pgdat(head
),
1710 NR_ISOLATED_ANON
+ page_is_file_lru(head
),
1711 thp_nr_pages(head
));
1715 * Either remove the duplicate refcount from
1716 * isolate_lru_page() or drop the page ref if it was
1721 mmap_read_unlock(mm
);
1725 static int move_pages_and_store_status(struct mm_struct
*mm
, int node
,
1726 struct list_head
*pagelist
, int __user
*status
,
1727 int start
, int i
, unsigned long nr_pages
)
1731 if (list_empty(pagelist
))
1734 err
= do_move_pages_to_node(mm
, pagelist
, node
);
1737 * Positive err means the number of failed
1738 * pages to migrate. Since we are going to
1739 * abort and return the number of non-migrated
1740 * pages, so need to include the rest of the
1741 * nr_pages that have not been attempted as
1745 err
+= nr_pages
- i
- 1;
1748 return store_status(status
, start
, node
, i
- start
);
1752 * Migrate an array of page address onto an array of nodes and fill
1753 * the corresponding array of status.
1755 static int do_pages_move(struct mm_struct
*mm
, nodemask_t task_nodes
,
1756 unsigned long nr_pages
,
1757 const void __user
* __user
*pages
,
1758 const int __user
*nodes
,
1759 int __user
*status
, int flags
)
1761 int current_node
= NUMA_NO_NODE
;
1762 LIST_HEAD(pagelist
);
1768 for (i
= start
= 0; i
< nr_pages
; i
++) {
1769 const void __user
*p
;
1774 if (get_user(p
, pages
+ i
))
1776 if (get_user(node
, nodes
+ i
))
1778 addr
= (unsigned long)untagged_addr(p
);
1781 if (node
< 0 || node
>= MAX_NUMNODES
)
1783 if (!node_state(node
, N_MEMORY
))
1787 if (!node_isset(node
, task_nodes
))
1790 if (current_node
== NUMA_NO_NODE
) {
1791 current_node
= node
;
1793 } else if (node
!= current_node
) {
1794 err
= move_pages_and_store_status(mm
, current_node
,
1795 &pagelist
, status
, start
, i
, nr_pages
);
1799 current_node
= node
;
1803 * Errors in the page lookup or isolation are not fatal and we simply
1804 * report them via status
1806 err
= add_page_for_migration(mm
, addr
, current_node
,
1807 &pagelist
, flags
& MPOL_MF_MOVE_ALL
);
1810 /* The page is successfully queued for migration */
1815 * If the page is already on the target node (!err), store the
1816 * node, otherwise, store the err.
1818 err
= store_status(status
, i
, err
? : current_node
, 1);
1822 err
= move_pages_and_store_status(mm
, current_node
, &pagelist
,
1823 status
, start
, i
, nr_pages
);
1826 current_node
= NUMA_NO_NODE
;
1829 /* Make sure we do not overwrite the existing error */
1830 err1
= move_pages_and_store_status(mm
, current_node
, &pagelist
,
1831 status
, start
, i
, nr_pages
);
1839 * Determine the nodes of an array of pages and store it in an array of status.
1841 static void do_pages_stat_array(struct mm_struct
*mm
, unsigned long nr_pages
,
1842 const void __user
**pages
, int *status
)
1848 for (i
= 0; i
< nr_pages
; i
++) {
1849 unsigned long addr
= (unsigned long)(*pages
);
1850 struct vm_area_struct
*vma
;
1854 vma
= find_vma(mm
, addr
);
1855 if (!vma
|| addr
< vma
->vm_start
)
1858 /* FOLL_DUMP to ignore special (like zero) pages */
1859 page
= follow_page(vma
, addr
, FOLL_DUMP
);
1861 err
= PTR_ERR(page
);
1865 err
= page
? page_to_nid(page
) : -ENOENT
;
1873 mmap_read_unlock(mm
);
1877 * Determine the nodes of a user array of pages and store it in
1878 * a user array of status.
1880 static int do_pages_stat(struct mm_struct
*mm
, unsigned long nr_pages
,
1881 const void __user
* __user
*pages
,
1884 #define DO_PAGES_STAT_CHUNK_NR 16
1885 const void __user
*chunk_pages
[DO_PAGES_STAT_CHUNK_NR
];
1886 int chunk_status
[DO_PAGES_STAT_CHUNK_NR
];
1889 unsigned long chunk_nr
;
1891 chunk_nr
= nr_pages
;
1892 if (chunk_nr
> DO_PAGES_STAT_CHUNK_NR
)
1893 chunk_nr
= DO_PAGES_STAT_CHUNK_NR
;
1895 if (copy_from_user(chunk_pages
, pages
, chunk_nr
* sizeof(*chunk_pages
)))
1898 do_pages_stat_array(mm
, chunk_nr
, chunk_pages
, chunk_status
);
1900 if (copy_to_user(status
, chunk_status
, chunk_nr
* sizeof(*status
)))
1905 nr_pages
-= chunk_nr
;
1907 return nr_pages
? -EFAULT
: 0;
1910 static struct mm_struct
*find_mm_struct(pid_t pid
, nodemask_t
*mem_nodes
)
1912 struct task_struct
*task
;
1913 struct mm_struct
*mm
;
1916 * There is no need to check if current process has the right to modify
1917 * the specified process when they are same.
1921 *mem_nodes
= cpuset_mems_allowed(current
);
1925 /* Find the mm_struct */
1927 task
= find_task_by_vpid(pid
);
1930 return ERR_PTR(-ESRCH
);
1932 get_task_struct(task
);
1935 * Check if this process has the right to modify the specified
1936 * process. Use the regular "ptrace_may_access()" checks.
1938 if (!ptrace_may_access(task
, PTRACE_MODE_READ_REALCREDS
)) {
1940 mm
= ERR_PTR(-EPERM
);
1945 mm
= ERR_PTR(security_task_movememory(task
));
1948 *mem_nodes
= cpuset_mems_allowed(task
);
1949 mm
= get_task_mm(task
);
1951 put_task_struct(task
);
1953 mm
= ERR_PTR(-EINVAL
);
1958 * Move a list of pages in the address space of the currently executing
1961 static int kernel_move_pages(pid_t pid
, unsigned long nr_pages
,
1962 const void __user
* __user
*pages
,
1963 const int __user
*nodes
,
1964 int __user
*status
, int flags
)
1966 struct mm_struct
*mm
;
1968 nodemask_t task_nodes
;
1971 if (flags
& ~(MPOL_MF_MOVE
|MPOL_MF_MOVE_ALL
))
1974 if ((flags
& MPOL_MF_MOVE_ALL
) && !capable(CAP_SYS_NICE
))
1977 mm
= find_mm_struct(pid
, &task_nodes
);
1982 err
= do_pages_move(mm
, task_nodes
, nr_pages
, pages
,
1983 nodes
, status
, flags
);
1985 err
= do_pages_stat(mm
, nr_pages
, pages
, status
);
1991 SYSCALL_DEFINE6(move_pages
, pid_t
, pid
, unsigned long, nr_pages
,
1992 const void __user
* __user
*, pages
,
1993 const int __user
*, nodes
,
1994 int __user
*, status
, int, flags
)
1996 return kernel_move_pages(pid
, nr_pages
, pages
, nodes
, status
, flags
);
1999 #ifdef CONFIG_COMPAT
2000 COMPAT_SYSCALL_DEFINE6(move_pages
, pid_t
, pid
, compat_ulong_t
, nr_pages
,
2001 compat_uptr_t __user
*, pages32
,
2002 const int __user
*, nodes
,
2003 int __user
*, status
,
2006 const void __user
* __user
*pages
;
2009 pages
= compat_alloc_user_space(nr_pages
* sizeof(void *));
2010 for (i
= 0; i
< nr_pages
; i
++) {
2013 if (get_user(p
, pages32
+ i
) ||
2014 put_user(compat_ptr(p
), pages
+ i
))
2017 return kernel_move_pages(pid
, nr_pages
, pages
, nodes
, status
, flags
);
2019 #endif /* CONFIG_COMPAT */
2021 #ifdef CONFIG_NUMA_BALANCING
2023 * Returns true if this is a safe migration target node for misplaced NUMA
2024 * pages. Currently it only checks the watermarks which crude
2026 static bool migrate_balanced_pgdat(struct pglist_data
*pgdat
,
2027 unsigned long nr_migrate_pages
)
2031 for (z
= pgdat
->nr_zones
- 1; z
>= 0; z
--) {
2032 struct zone
*zone
= pgdat
->node_zones
+ z
;
2034 if (!populated_zone(zone
))
2037 /* Avoid waking kswapd by allocating pages_to_migrate pages. */
2038 if (!zone_watermark_ok(zone
, 0,
2039 high_wmark_pages(zone
) +
2048 static struct page
*alloc_misplaced_dst_page(struct page
*page
,
2051 int nid
= (int) data
;
2052 struct page
*newpage
;
2054 newpage
= __alloc_pages_node(nid
,
2055 (GFP_HIGHUSER_MOVABLE
|
2056 __GFP_THISNODE
| __GFP_NOMEMALLOC
|
2057 __GFP_NORETRY
| __GFP_NOWARN
) &
2063 static int numamigrate_isolate_page(pg_data_t
*pgdat
, struct page
*page
)
2067 VM_BUG_ON_PAGE(compound_order(page
) && !PageTransHuge(page
), page
);
2069 /* Avoid migrating to a node that is nearly full */
2070 if (!migrate_balanced_pgdat(pgdat
, compound_nr(page
)))
2073 if (isolate_lru_page(page
))
2077 * migrate_misplaced_transhuge_page() skips page migration's usual
2078 * check on page_count(), so we must do it here, now that the page
2079 * has been isolated: a GUP pin, or any other pin, prevents migration.
2080 * The expected page count is 3: 1 for page's mapcount and 1 for the
2081 * caller's pin and 1 for the reference taken by isolate_lru_page().
2083 if (PageTransHuge(page
) && page_count(page
) != 3) {
2084 putback_lru_page(page
);
2088 page_lru
= page_is_file_lru(page
);
2089 mod_node_page_state(page_pgdat(page
), NR_ISOLATED_ANON
+ page_lru
,
2090 thp_nr_pages(page
));
2093 * Isolating the page has taken another reference, so the
2094 * caller's reference can be safely dropped without the page
2095 * disappearing underneath us during migration.
2101 bool pmd_trans_migrating(pmd_t pmd
)
2103 struct page
*page
= pmd_page(pmd
);
2104 return PageLocked(page
);
2107 static inline bool is_shared_exec_page(struct vm_area_struct
*vma
,
2110 if (page_mapcount(page
) != 1 &&
2111 (page_is_file_lru(page
) || vma_is_shmem(vma
)) &&
2112 (vma
->vm_flags
& VM_EXEC
))
2119 * Attempt to migrate a misplaced page to the specified destination
2120 * node. Caller is expected to have an elevated reference count on
2121 * the page that will be dropped by this function before returning.
2123 int migrate_misplaced_page(struct page
*page
, struct vm_area_struct
*vma
,
2126 pg_data_t
*pgdat
= NODE_DATA(node
);
2129 LIST_HEAD(migratepages
);
2132 * Don't migrate file pages that are mapped in multiple processes
2133 * with execute permissions as they are probably shared libraries.
2135 if (is_shared_exec_page(vma
, page
))
2139 * Also do not migrate dirty pages as not all filesystems can move
2140 * dirty pages in MIGRATE_ASYNC mode which is a waste of cycles.
2142 if (page_is_file_lru(page
) && PageDirty(page
))
2145 isolated
= numamigrate_isolate_page(pgdat
, page
);
2149 list_add(&page
->lru
, &migratepages
);
2150 nr_remaining
= migrate_pages(&migratepages
, alloc_misplaced_dst_page
,
2151 NULL
, node
, MIGRATE_ASYNC
,
2154 if (!list_empty(&migratepages
)) {
2155 list_del(&page
->lru
);
2156 dec_node_page_state(page
, NR_ISOLATED_ANON
+
2157 page_is_file_lru(page
));
2158 putback_lru_page(page
);
2162 count_vm_numa_event(NUMA_PAGE_MIGRATE
);
2163 BUG_ON(!list_empty(&migratepages
));
2170 #endif /* CONFIG_NUMA_BALANCING */
2172 #if defined(CONFIG_NUMA_BALANCING) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
2174 * Migrates a THP to a given target node. page must be locked and is unlocked
2177 int migrate_misplaced_transhuge_page(struct mm_struct
*mm
,
2178 struct vm_area_struct
*vma
,
2179 pmd_t
*pmd
, pmd_t entry
,
2180 unsigned long address
,
2181 struct page
*page
, int node
)
2184 pg_data_t
*pgdat
= NODE_DATA(node
);
2186 struct page
*new_page
= NULL
;
2187 int page_lru
= page_is_file_lru(page
);
2188 unsigned long start
= address
& HPAGE_PMD_MASK
;
2190 if (is_shared_exec_page(vma
, page
))
2193 new_page
= alloc_pages_node(node
,
2194 (GFP_TRANSHUGE_LIGHT
| __GFP_THISNODE
),
2198 prep_transhuge_page(new_page
);
2200 isolated
= numamigrate_isolate_page(pgdat
, page
);
2206 /* Prepare a page as a migration target */
2207 __SetPageLocked(new_page
);
2208 if (PageSwapBacked(page
))
2209 __SetPageSwapBacked(new_page
);
2211 /* anon mapping, we can simply copy page->mapping to the new page: */
2212 new_page
->mapping
= page
->mapping
;
2213 new_page
->index
= page
->index
;
2214 /* flush the cache before copying using the kernel virtual address */
2215 flush_cache_range(vma
, start
, start
+ HPAGE_PMD_SIZE
);
2216 migrate_page_copy(new_page
, page
);
2217 WARN_ON(PageLRU(new_page
));
2219 /* Recheck the target PMD */
2220 ptl
= pmd_lock(mm
, pmd
);
2221 if (unlikely(!pmd_same(*pmd
, entry
) || !page_ref_freeze(page
, 2))) {
2224 /* Reverse changes made by migrate_page_copy() */
2225 if (TestClearPageActive(new_page
))
2226 SetPageActive(page
);
2227 if (TestClearPageUnevictable(new_page
))
2228 SetPageUnevictable(page
);
2230 unlock_page(new_page
);
2231 put_page(new_page
); /* Free it */
2233 /* Retake the callers reference and putback on LRU */
2235 putback_lru_page(page
);
2236 mod_node_page_state(page_pgdat(page
),
2237 NR_ISOLATED_ANON
+ page_lru
, -HPAGE_PMD_NR
);
2242 entry
= mk_huge_pmd(new_page
, vma
->vm_page_prot
);
2243 entry
= maybe_pmd_mkwrite(pmd_mkdirty(entry
), vma
);
2246 * Overwrite the old entry under pagetable lock and establish
2247 * the new PTE. Any parallel GUP will either observe the old
2248 * page blocking on the page lock, block on the page table
2249 * lock or observe the new page. The SetPageUptodate on the
2250 * new page and page_add_new_anon_rmap guarantee the copy is
2251 * visible before the pagetable update.
2253 page_add_anon_rmap(new_page
, vma
, start
, true);
2255 * At this point the pmd is numa/protnone (i.e. non present) and the TLB
2256 * has already been flushed globally. So no TLB can be currently
2257 * caching this non present pmd mapping. There's no need to clear the
2258 * pmd before doing set_pmd_at(), nor to flush the TLB after
2259 * set_pmd_at(). Clearing the pmd here would introduce a race
2260 * condition against MADV_DONTNEED, because MADV_DONTNEED only holds the
2261 * mmap_lock for reading. If the pmd is set to NULL at any given time,
2262 * MADV_DONTNEED won't wait on the pmd lock and it'll skip clearing this
2265 set_pmd_at(mm
, start
, pmd
, entry
);
2266 update_mmu_cache_pmd(vma
, address
, &entry
);
2268 page_ref_unfreeze(page
, 2);
2269 mlock_migrate_page(new_page
, page
);
2270 page_remove_rmap(page
, true);
2271 set_page_owner_migrate_reason(new_page
, MR_NUMA_MISPLACED
);
2275 /* Take an "isolate" reference and put new page on the LRU. */
2277 putback_lru_page(new_page
);
2279 unlock_page(new_page
);
2281 put_page(page
); /* Drop the rmap reference */
2282 put_page(page
); /* Drop the LRU isolation reference */
2284 count_vm_events(PGMIGRATE_SUCCESS
, HPAGE_PMD_NR
);
2285 count_vm_numa_events(NUMA_PAGE_MIGRATE
, HPAGE_PMD_NR
);
2287 mod_node_page_state(page_pgdat(page
),
2288 NR_ISOLATED_ANON
+ page_lru
,
2293 count_vm_events(PGMIGRATE_FAIL
, HPAGE_PMD_NR
);
2294 ptl
= pmd_lock(mm
, pmd
);
2295 if (pmd_same(*pmd
, entry
)) {
2296 entry
= pmd_modify(entry
, vma
->vm_page_prot
);
2297 set_pmd_at(mm
, start
, pmd
, entry
);
2298 update_mmu_cache_pmd(vma
, address
, &entry
);
2308 #endif /* CONFIG_NUMA_BALANCING */
2310 #endif /* CONFIG_NUMA */
2312 #ifdef CONFIG_DEVICE_PRIVATE
2313 static int migrate_vma_collect_hole(unsigned long start
,
2315 __always_unused
int depth
,
2316 struct mm_walk
*walk
)
2318 struct migrate_vma
*migrate
= walk
->private;
2321 /* Only allow populating anonymous memory. */
2322 if (!vma_is_anonymous(walk
->vma
)) {
2323 for (addr
= start
; addr
< end
; addr
+= PAGE_SIZE
) {
2324 migrate
->src
[migrate
->npages
] = 0;
2325 migrate
->dst
[migrate
->npages
] = 0;
2331 for (addr
= start
; addr
< end
; addr
+= PAGE_SIZE
) {
2332 migrate
->src
[migrate
->npages
] = MIGRATE_PFN_MIGRATE
;
2333 migrate
->dst
[migrate
->npages
] = 0;
2341 static int migrate_vma_collect_skip(unsigned long start
,
2343 struct mm_walk
*walk
)
2345 struct migrate_vma
*migrate
= walk
->private;
2348 for (addr
= start
; addr
< end
; addr
+= PAGE_SIZE
) {
2349 migrate
->dst
[migrate
->npages
] = 0;
2350 migrate
->src
[migrate
->npages
++] = 0;
2356 static int migrate_vma_collect_pmd(pmd_t
*pmdp
,
2357 unsigned long start
,
2359 struct mm_walk
*walk
)
2361 struct migrate_vma
*migrate
= walk
->private;
2362 struct vm_area_struct
*vma
= walk
->vma
;
2363 struct mm_struct
*mm
= vma
->vm_mm
;
2364 unsigned long addr
= start
, unmapped
= 0;
2369 if (pmd_none(*pmdp
))
2370 return migrate_vma_collect_hole(start
, end
, -1, walk
);
2372 if (pmd_trans_huge(*pmdp
)) {
2375 ptl
= pmd_lock(mm
, pmdp
);
2376 if (unlikely(!pmd_trans_huge(*pmdp
))) {
2381 page
= pmd_page(*pmdp
);
2382 if (is_huge_zero_page(page
)) {
2384 split_huge_pmd(vma
, pmdp
, addr
);
2385 if (pmd_trans_unstable(pmdp
))
2386 return migrate_vma_collect_skip(start
, end
,
2393 if (unlikely(!trylock_page(page
)))
2394 return migrate_vma_collect_skip(start
, end
,
2396 ret
= split_huge_page(page
);
2400 return migrate_vma_collect_skip(start
, end
,
2402 if (pmd_none(*pmdp
))
2403 return migrate_vma_collect_hole(start
, end
, -1,
2408 if (unlikely(pmd_bad(*pmdp
)))
2409 return migrate_vma_collect_skip(start
, end
, walk
);
2411 ptep
= pte_offset_map_lock(mm
, pmdp
, addr
, &ptl
);
2412 arch_enter_lazy_mmu_mode();
2414 for (; addr
< end
; addr
+= PAGE_SIZE
, ptep
++) {
2415 unsigned long mpfn
= 0, pfn
;
2422 if (pte_none(pte
)) {
2423 if (vma_is_anonymous(vma
)) {
2424 mpfn
= MIGRATE_PFN_MIGRATE
;
2430 if (!pte_present(pte
)) {
2432 * Only care about unaddressable device page special
2433 * page table entry. Other special swap entries are not
2434 * migratable, and we ignore regular swapped page.
2436 entry
= pte_to_swp_entry(pte
);
2437 if (!is_device_private_entry(entry
))
2440 page
= device_private_entry_to_page(entry
);
2441 if (!(migrate
->flags
&
2442 MIGRATE_VMA_SELECT_DEVICE_PRIVATE
) ||
2443 page
->pgmap
->owner
!= migrate
->pgmap_owner
)
2446 mpfn
= migrate_pfn(page_to_pfn(page
)) |
2447 MIGRATE_PFN_MIGRATE
;
2448 if (is_write_device_private_entry(entry
))
2449 mpfn
|= MIGRATE_PFN_WRITE
;
2451 if (!(migrate
->flags
& MIGRATE_VMA_SELECT_SYSTEM
))
2454 if (is_zero_pfn(pfn
)) {
2455 mpfn
= MIGRATE_PFN_MIGRATE
;
2459 page
= vm_normal_page(migrate
->vma
, addr
, pte
);
2460 mpfn
= migrate_pfn(pfn
) | MIGRATE_PFN_MIGRATE
;
2461 mpfn
|= pte_write(pte
) ? MIGRATE_PFN_WRITE
: 0;
2464 /* FIXME support THP */
2465 if (!page
|| !page
->mapping
|| PageTransCompound(page
)) {
2471 * By getting a reference on the page we pin it and that blocks
2472 * any kind of migration. Side effect is that it "freezes" the
2475 * We drop this reference after isolating the page from the lru
2476 * for non device page (device page are not on the lru and thus
2477 * can't be dropped from it).
2483 * Optimize for the common case where page is only mapped once
2484 * in one process. If we can lock the page, then we can safely
2485 * set up a special migration page table entry now.
2487 if (trylock_page(page
)) {
2490 mpfn
|= MIGRATE_PFN_LOCKED
;
2491 ptep_get_and_clear(mm
, addr
, ptep
);
2493 /* Setup special migration page table entry */
2494 entry
= make_migration_entry(page
, mpfn
&
2496 swp_pte
= swp_entry_to_pte(entry
);
2497 if (pte_present(pte
)) {
2498 if (pte_soft_dirty(pte
))
2499 swp_pte
= pte_swp_mksoft_dirty(swp_pte
);
2500 if (pte_uffd_wp(pte
))
2501 swp_pte
= pte_swp_mkuffd_wp(swp_pte
);
2503 if (pte_swp_soft_dirty(pte
))
2504 swp_pte
= pte_swp_mksoft_dirty(swp_pte
);
2505 if (pte_swp_uffd_wp(pte
))
2506 swp_pte
= pte_swp_mkuffd_wp(swp_pte
);
2508 set_pte_at(mm
, addr
, ptep
, swp_pte
);
2511 * This is like regular unmap: we remove the rmap and
2512 * drop page refcount. Page won't be freed, as we took
2513 * a reference just above.
2515 page_remove_rmap(page
, false);
2518 if (pte_present(pte
))
2523 migrate
->dst
[migrate
->npages
] = 0;
2524 migrate
->src
[migrate
->npages
++] = mpfn
;
2526 arch_leave_lazy_mmu_mode();
2527 pte_unmap_unlock(ptep
- 1, ptl
);
2529 /* Only flush the TLB if we actually modified any entries */
2531 flush_tlb_range(walk
->vma
, start
, end
);
2536 static const struct mm_walk_ops migrate_vma_walk_ops
= {
2537 .pmd_entry
= migrate_vma_collect_pmd
,
2538 .pte_hole
= migrate_vma_collect_hole
,
2542 * migrate_vma_collect() - collect pages over a range of virtual addresses
2543 * @migrate: migrate struct containing all migration information
2545 * This will walk the CPU page table. For each virtual address backed by a
2546 * valid page, it updates the src array and takes a reference on the page, in
2547 * order to pin the page until we lock it and unmap it.
2549 static void migrate_vma_collect(struct migrate_vma
*migrate
)
2551 struct mmu_notifier_range range
;
2554 * Note that the pgmap_owner is passed to the mmu notifier callback so
2555 * that the registered device driver can skip invalidating device
2556 * private page mappings that won't be migrated.
2558 mmu_notifier_range_init_migrate(&range
, 0, migrate
->vma
,
2559 migrate
->vma
->vm_mm
, migrate
->start
, migrate
->end
,
2560 migrate
->pgmap_owner
);
2561 mmu_notifier_invalidate_range_start(&range
);
2563 walk_page_range(migrate
->vma
->vm_mm
, migrate
->start
, migrate
->end
,
2564 &migrate_vma_walk_ops
, migrate
);
2566 mmu_notifier_invalidate_range_end(&range
);
2567 migrate
->end
= migrate
->start
+ (migrate
->npages
<< PAGE_SHIFT
);
2571 * migrate_vma_check_page() - check if page is pinned or not
2572 * @page: struct page to check
2574 * Pinned pages cannot be migrated. This is the same test as in
2575 * migrate_page_move_mapping(), except that here we allow migration of a
2578 static bool migrate_vma_check_page(struct page
*page
)
2581 * One extra ref because caller holds an extra reference, either from
2582 * isolate_lru_page() for a regular page, or migrate_vma_collect() for
2588 * FIXME support THP (transparent huge page), it is bit more complex to
2589 * check them than regular pages, because they can be mapped with a pmd
2590 * or with a pte (split pte mapping).
2592 if (PageCompound(page
))
2595 /* Page from ZONE_DEVICE have one extra reference */
2596 if (is_zone_device_page(page
)) {
2598 * Private page can never be pin as they have no valid pte and
2599 * GUP will fail for those. Yet if there is a pending migration
2600 * a thread might try to wait on the pte migration entry and
2601 * will bump the page reference count. Sadly there is no way to
2602 * differentiate a regular pin from migration wait. Hence to
2603 * avoid 2 racing thread trying to migrate back to CPU to enter
2604 * infinite loop (one stopping migration because the other is
2605 * waiting on pte migration entry). We always return true here.
2607 * FIXME proper solution is to rework migration_entry_wait() so
2608 * it does not need to take a reference on page.
2610 return is_device_private_page(page
);
2613 /* For file back page */
2614 if (page_mapping(page
))
2615 extra
+= 1 + page_has_private(page
);
2617 if ((page_count(page
) - extra
) > page_mapcount(page
))
2624 * migrate_vma_prepare() - lock pages and isolate them from the lru
2625 * @migrate: migrate struct containing all migration information
2627 * This locks pages that have been collected by migrate_vma_collect(). Once each
2628 * page is locked it is isolated from the lru (for non-device pages). Finally,
2629 * the ref taken by migrate_vma_collect() is dropped, as locked pages cannot be
2630 * migrated by concurrent kernel threads.
2632 static void migrate_vma_prepare(struct migrate_vma
*migrate
)
2634 const unsigned long npages
= migrate
->npages
;
2635 const unsigned long start
= migrate
->start
;
2636 unsigned long addr
, i
, restore
= 0;
2637 bool allow_drain
= true;
2641 for (i
= 0; (i
< npages
) && migrate
->cpages
; i
++) {
2642 struct page
*page
= migrate_pfn_to_page(migrate
->src
[i
]);
2648 if (!(migrate
->src
[i
] & MIGRATE_PFN_LOCKED
)) {
2650 * Because we are migrating several pages there can be
2651 * a deadlock between 2 concurrent migration where each
2652 * are waiting on each other page lock.
2654 * Make migrate_vma() a best effort thing and backoff
2655 * for any page we can not lock right away.
2657 if (!trylock_page(page
)) {
2658 migrate
->src
[i
] = 0;
2664 migrate
->src
[i
] |= MIGRATE_PFN_LOCKED
;
2667 /* ZONE_DEVICE pages are not on LRU */
2668 if (!is_zone_device_page(page
)) {
2669 if (!PageLRU(page
) && allow_drain
) {
2670 /* Drain CPU's pagevec */
2671 lru_add_drain_all();
2672 allow_drain
= false;
2675 if (isolate_lru_page(page
)) {
2677 migrate
->src
[i
] &= ~MIGRATE_PFN_MIGRATE
;
2681 migrate
->src
[i
] = 0;
2689 /* Drop the reference we took in collect */
2693 if (!migrate_vma_check_page(page
)) {
2695 migrate
->src
[i
] &= ~MIGRATE_PFN_MIGRATE
;
2699 if (!is_zone_device_page(page
)) {
2701 putback_lru_page(page
);
2704 migrate
->src
[i
] = 0;
2708 if (!is_zone_device_page(page
))
2709 putback_lru_page(page
);
2716 for (i
= 0, addr
= start
; i
< npages
&& restore
; i
++, addr
+= PAGE_SIZE
) {
2717 struct page
*page
= migrate_pfn_to_page(migrate
->src
[i
]);
2719 if (!page
|| (migrate
->src
[i
] & MIGRATE_PFN_MIGRATE
))
2722 remove_migration_pte(page
, migrate
->vma
, addr
, page
);
2724 migrate
->src
[i
] = 0;
2732 * migrate_vma_unmap() - replace page mapping with special migration pte entry
2733 * @migrate: migrate struct containing all migration information
2735 * Replace page mapping (CPU page table pte) with a special migration pte entry
2736 * and check again if it has been pinned. Pinned pages are restored because we
2737 * cannot migrate them.
2739 * This is the last step before we call the device driver callback to allocate
2740 * destination memory and copy contents of original page over to new page.
2742 static void migrate_vma_unmap(struct migrate_vma
*migrate
)
2744 int flags
= TTU_MIGRATION
| TTU_IGNORE_MLOCK
;
2745 const unsigned long npages
= migrate
->npages
;
2746 const unsigned long start
= migrate
->start
;
2747 unsigned long addr
, i
, restore
= 0;
2749 for (i
= 0; i
< npages
; i
++) {
2750 struct page
*page
= migrate_pfn_to_page(migrate
->src
[i
]);
2752 if (!page
|| !(migrate
->src
[i
] & MIGRATE_PFN_MIGRATE
))
2755 if (page_mapped(page
)) {
2756 try_to_unmap(page
, flags
);
2757 if (page_mapped(page
))
2761 if (migrate_vma_check_page(page
))
2765 migrate
->src
[i
] &= ~MIGRATE_PFN_MIGRATE
;
2770 for (addr
= start
, i
= 0; i
< npages
&& restore
; addr
+= PAGE_SIZE
, i
++) {
2771 struct page
*page
= migrate_pfn_to_page(migrate
->src
[i
]);
2773 if (!page
|| (migrate
->src
[i
] & MIGRATE_PFN_MIGRATE
))
2776 remove_migration_ptes(page
, page
, false);
2778 migrate
->src
[i
] = 0;
2782 if (is_zone_device_page(page
))
2785 putback_lru_page(page
);
2790 * migrate_vma_setup() - prepare to migrate a range of memory
2791 * @args: contains the vma, start, and pfns arrays for the migration
2793 * Returns: negative errno on failures, 0 when 0 or more pages were migrated
2796 * Prepare to migrate a range of memory virtual address range by collecting all
2797 * the pages backing each virtual address in the range, saving them inside the
2798 * src array. Then lock those pages and unmap them. Once the pages are locked
2799 * and unmapped, check whether each page is pinned or not. Pages that aren't
2800 * pinned have the MIGRATE_PFN_MIGRATE flag set (by this function) in the
2801 * corresponding src array entry. Then restores any pages that are pinned, by
2802 * remapping and unlocking those pages.
2804 * The caller should then allocate destination memory and copy source memory to
2805 * it for all those entries (ie with MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE
2806 * flag set). Once these are allocated and copied, the caller must update each
2807 * corresponding entry in the dst array with the pfn value of the destination
2808 * page and with the MIGRATE_PFN_VALID and MIGRATE_PFN_LOCKED flags set
2809 * (destination pages must have their struct pages locked, via lock_page()).
2811 * Note that the caller does not have to migrate all the pages that are marked
2812 * with MIGRATE_PFN_MIGRATE flag in src array unless this is a migration from
2813 * device memory to system memory. If the caller cannot migrate a device page
2814 * back to system memory, then it must return VM_FAULT_SIGBUS, which has severe
2815 * consequences for the userspace process, so it must be avoided if at all
2818 * For empty entries inside CPU page table (pte_none() or pmd_none() is true) we
2819 * do set MIGRATE_PFN_MIGRATE flag inside the corresponding source array thus
2820 * allowing the caller to allocate device memory for those unback virtual
2821 * address. For this the caller simply has to allocate device memory and
2822 * properly set the destination entry like for regular migration. Note that
2823 * this can still fails and thus inside the device driver must check if the
2824 * migration was successful for those entries after calling migrate_vma_pages()
2825 * just like for regular migration.
2827 * After that, the callers must call migrate_vma_pages() to go over each entry
2828 * in the src array that has the MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE flag
2829 * set. If the corresponding entry in dst array has MIGRATE_PFN_VALID flag set,
2830 * then migrate_vma_pages() to migrate struct page information from the source
2831 * struct page to the destination struct page. If it fails to migrate the
2832 * struct page information, then it clears the MIGRATE_PFN_MIGRATE flag in the
2835 * At this point all successfully migrated pages have an entry in the src
2836 * array with MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE flag set and the dst
2837 * array entry with MIGRATE_PFN_VALID flag set.
2839 * Once migrate_vma_pages() returns the caller may inspect which pages were
2840 * successfully migrated, and which were not. Successfully migrated pages will
2841 * have the MIGRATE_PFN_MIGRATE flag set for their src array entry.
2843 * It is safe to update device page table after migrate_vma_pages() because
2844 * both destination and source page are still locked, and the mmap_lock is held
2845 * in read mode (hence no one can unmap the range being migrated).
2847 * Once the caller is done cleaning up things and updating its page table (if it
2848 * chose to do so, this is not an obligation) it finally calls
2849 * migrate_vma_finalize() to update the CPU page table to point to new pages
2850 * for successfully migrated pages or otherwise restore the CPU page table to
2851 * point to the original source pages.
2853 int migrate_vma_setup(struct migrate_vma
*args
)
2855 long nr_pages
= (args
->end
- args
->start
) >> PAGE_SHIFT
;
2857 args
->start
&= PAGE_MASK
;
2858 args
->end
&= PAGE_MASK
;
2859 if (!args
->vma
|| is_vm_hugetlb_page(args
->vma
) ||
2860 (args
->vma
->vm_flags
& VM_SPECIAL
) || vma_is_dax(args
->vma
))
2864 if (args
->start
< args
->vma
->vm_start
||
2865 args
->start
>= args
->vma
->vm_end
)
2867 if (args
->end
<= args
->vma
->vm_start
|| args
->end
> args
->vma
->vm_end
)
2869 if (!args
->src
|| !args
->dst
)
2872 memset(args
->src
, 0, sizeof(*args
->src
) * nr_pages
);
2876 migrate_vma_collect(args
);
2879 migrate_vma_prepare(args
);
2881 migrate_vma_unmap(args
);
2884 * At this point pages are locked and unmapped, and thus they have
2885 * stable content and can safely be copied to destination memory that
2886 * is allocated by the drivers.
2891 EXPORT_SYMBOL(migrate_vma_setup
);
2894 * This code closely matches the code in:
2895 * __handle_mm_fault()
2896 * handle_pte_fault()
2897 * do_anonymous_page()
2898 * to map in an anonymous zero page but the struct page will be a ZONE_DEVICE
2901 static void migrate_vma_insert_page(struct migrate_vma
*migrate
,
2906 struct vm_area_struct
*vma
= migrate
->vma
;
2907 struct mm_struct
*mm
= vma
->vm_mm
;
2917 /* Only allow populating anonymous memory */
2918 if (!vma_is_anonymous(vma
))
2921 pgdp
= pgd_offset(mm
, addr
);
2922 p4dp
= p4d_alloc(mm
, pgdp
, addr
);
2925 pudp
= pud_alloc(mm
, p4dp
, addr
);
2928 pmdp
= pmd_alloc(mm
, pudp
, addr
);
2932 if (pmd_trans_huge(*pmdp
) || pmd_devmap(*pmdp
))
2936 * Use pte_alloc() instead of pte_alloc_map(). We can't run
2937 * pte_offset_map() on pmds where a huge pmd might be created
2938 * from a different thread.
2940 * pte_alloc_map() is safe to use under mmap_write_lock(mm) or when
2941 * parallel threads are excluded by other means.
2943 * Here we only have mmap_read_lock(mm).
2945 if (pte_alloc(mm
, pmdp
))
2948 /* See the comment in pte_alloc_one_map() */
2949 if (unlikely(pmd_trans_unstable(pmdp
)))
2952 if (unlikely(anon_vma_prepare(vma
)))
2954 if (mem_cgroup_charge(page
, vma
->vm_mm
, GFP_KERNEL
))
2958 * The memory barrier inside __SetPageUptodate makes sure that
2959 * preceding stores to the page contents become visible before
2960 * the set_pte_at() write.
2962 __SetPageUptodate(page
);
2964 if (is_zone_device_page(page
)) {
2965 if (is_device_private_page(page
)) {
2966 swp_entry_t swp_entry
;
2968 swp_entry
= make_device_private_entry(page
, vma
->vm_flags
& VM_WRITE
);
2969 entry
= swp_entry_to_pte(swp_entry
);
2972 entry
= mk_pte(page
, vma
->vm_page_prot
);
2973 if (vma
->vm_flags
& VM_WRITE
)
2974 entry
= pte_mkwrite(pte_mkdirty(entry
));
2977 ptep
= pte_offset_map_lock(mm
, pmdp
, addr
, &ptl
);
2979 if (check_stable_address_space(mm
))
2982 if (pte_present(*ptep
)) {
2983 unsigned long pfn
= pte_pfn(*ptep
);
2985 if (!is_zero_pfn(pfn
))
2988 } else if (!pte_none(*ptep
))
2992 * Check for userfaultfd but do not deliver the fault. Instead,
2995 if (userfaultfd_missing(vma
))
2998 inc_mm_counter(mm
, MM_ANONPAGES
);
2999 page_add_new_anon_rmap(page
, vma
, addr
, false);
3000 if (!is_zone_device_page(page
))
3001 lru_cache_add_inactive_or_unevictable(page
, vma
);
3005 flush_cache_page(vma
, addr
, pte_pfn(*ptep
));
3006 ptep_clear_flush_notify(vma
, addr
, ptep
);
3007 set_pte_at_notify(mm
, addr
, ptep
, entry
);
3008 update_mmu_cache(vma
, addr
, ptep
);
3010 /* No need to invalidate - it was non-present before */
3011 set_pte_at(mm
, addr
, ptep
, entry
);
3012 update_mmu_cache(vma
, addr
, ptep
);
3015 pte_unmap_unlock(ptep
, ptl
);
3016 *src
= MIGRATE_PFN_MIGRATE
;
3020 pte_unmap_unlock(ptep
, ptl
);
3022 *src
&= ~MIGRATE_PFN_MIGRATE
;
3026 * migrate_vma_pages() - migrate meta-data from src page to dst page
3027 * @migrate: migrate struct containing all migration information
3029 * This migrates struct page meta-data from source struct page to destination
3030 * struct page. This effectively finishes the migration from source page to the
3033 void migrate_vma_pages(struct migrate_vma
*migrate
)
3035 const unsigned long npages
= migrate
->npages
;
3036 const unsigned long start
= migrate
->start
;
3037 struct mmu_notifier_range range
;
3038 unsigned long addr
, i
;
3039 bool notified
= false;
3041 for (i
= 0, addr
= start
; i
< npages
; addr
+= PAGE_SIZE
, i
++) {
3042 struct page
*newpage
= migrate_pfn_to_page(migrate
->dst
[i
]);
3043 struct page
*page
= migrate_pfn_to_page(migrate
->src
[i
]);
3044 struct address_space
*mapping
;
3048 migrate
->src
[i
] &= ~MIGRATE_PFN_MIGRATE
;
3053 if (!(migrate
->src
[i
] & MIGRATE_PFN_MIGRATE
))
3058 mmu_notifier_range_init_migrate(&range
, 0,
3059 migrate
->vma
, migrate
->vma
->vm_mm
,
3061 migrate
->pgmap_owner
);
3062 mmu_notifier_invalidate_range_start(&range
);
3064 migrate_vma_insert_page(migrate
, addr
, newpage
,
3069 mapping
= page_mapping(page
);
3071 if (is_zone_device_page(newpage
)) {
3072 if (is_device_private_page(newpage
)) {
3074 * For now only support private anonymous when
3075 * migrating to un-addressable device memory.
3078 migrate
->src
[i
] &= ~MIGRATE_PFN_MIGRATE
;
3083 * Other types of ZONE_DEVICE page are not
3086 migrate
->src
[i
] &= ~MIGRATE_PFN_MIGRATE
;
3091 r
= migrate_page(mapping
, newpage
, page
, MIGRATE_SYNC_NO_COPY
);
3092 if (r
!= MIGRATEPAGE_SUCCESS
)
3093 migrate
->src
[i
] &= ~MIGRATE_PFN_MIGRATE
;
3097 * No need to double call mmu_notifier->invalidate_range() callback as
3098 * the above ptep_clear_flush_notify() inside migrate_vma_insert_page()
3099 * did already call it.
3102 mmu_notifier_invalidate_range_only_end(&range
);
3104 EXPORT_SYMBOL(migrate_vma_pages
);
3107 * migrate_vma_finalize() - restore CPU page table entry
3108 * @migrate: migrate struct containing all migration information
3110 * This replaces the special migration pte entry with either a mapping to the
3111 * new page if migration was successful for that page, or to the original page
3114 * This also unlocks the pages and puts them back on the lru, or drops the extra
3115 * refcount, for device pages.
3117 void migrate_vma_finalize(struct migrate_vma
*migrate
)
3119 const unsigned long npages
= migrate
->npages
;
3122 for (i
= 0; i
< npages
; i
++) {
3123 struct page
*newpage
= migrate_pfn_to_page(migrate
->dst
[i
]);
3124 struct page
*page
= migrate_pfn_to_page(migrate
->src
[i
]);
3128 unlock_page(newpage
);
3134 if (!(migrate
->src
[i
] & MIGRATE_PFN_MIGRATE
) || !newpage
) {
3136 unlock_page(newpage
);
3142 remove_migration_ptes(page
, newpage
, false);
3145 if (is_zone_device_page(page
))
3148 putback_lru_page(page
);
3150 if (newpage
!= page
) {
3151 unlock_page(newpage
);
3152 if (is_zone_device_page(newpage
))
3155 putback_lru_page(newpage
);
3159 EXPORT_SYMBOL(migrate_vma_finalize
);
3160 #endif /* CONFIG_DEVICE_PRIVATE */