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/pfn_t.h>
42 #include <linux/memremap.h>
43 #include <linux/userfaultfd_k.h>
44 #include <linux/balloon_compaction.h>
45 #include <linux/mmu_notifier.h>
46 #include <linux/page_idle.h>
47 #include <linux/page_owner.h>
48 #include <linux/sched/mm.h>
49 #include <linux/ptrace.h>
51 #include <asm/tlbflush.h>
53 #define CREATE_TRACE_POINTS
54 #include <trace/events/migrate.h>
59 * migrate_prep() needs to be called before we start compiling a list of pages
60 * to be migrated using isolate_lru_page(). If scheduling work on other CPUs is
61 * undesirable, use migrate_prep_local()
63 int migrate_prep(void)
66 * Clear the LRU lists so pages can be isolated.
67 * Note that pages may be moved off the LRU after we have
68 * drained them. Those pages will fail to migrate like other
69 * pages that may be busy.
76 /* Do the necessary work of migrate_prep but not if it involves other CPUs */
77 int migrate_prep_local(void)
84 int isolate_movable_page(struct page
*page
, isolate_mode_t mode
)
86 struct address_space
*mapping
;
89 * Avoid burning cycles with pages that are yet under __free_pages(),
90 * or just got freed under us.
92 * In case we 'win' a race for a movable page being freed under us and
93 * raise its refcount preventing __free_pages() from doing its job
94 * the put_page() at the end of this block will take care of
95 * release this page, thus avoiding a nasty leakage.
97 if (unlikely(!get_page_unless_zero(page
)))
101 * Check PageMovable before holding a PG_lock because page's owner
102 * assumes anybody doesn't touch PG_lock of newly allocated page
103 * so unconditionally grabbing the lock ruins page's owner side.
105 if (unlikely(!__PageMovable(page
)))
108 * As movable pages are not isolated from LRU lists, concurrent
109 * compaction threads can race against page migration functions
110 * as well as race against the releasing a page.
112 * In order to avoid having an already isolated movable page
113 * being (wrongly) re-isolated while it is under migration,
114 * or to avoid attempting to isolate pages being released,
115 * lets be sure we have the page lock
116 * before proceeding with the movable page isolation steps.
118 if (unlikely(!trylock_page(page
)))
121 if (!PageMovable(page
) || PageIsolated(page
))
122 goto out_no_isolated
;
124 mapping
= page_mapping(page
);
125 VM_BUG_ON_PAGE(!mapping
, page
);
127 if (!mapping
->a_ops
->isolate_page(page
, mode
))
128 goto out_no_isolated
;
130 /* Driver shouldn't use PG_isolated bit of page->flags */
131 WARN_ON_ONCE(PageIsolated(page
));
132 __SetPageIsolated(page
);
145 /* It should be called on page which is PG_movable */
146 void putback_movable_page(struct page
*page
)
148 struct address_space
*mapping
;
150 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
151 VM_BUG_ON_PAGE(!PageMovable(page
), page
);
152 VM_BUG_ON_PAGE(!PageIsolated(page
), page
);
154 mapping
= page_mapping(page
);
155 mapping
->a_ops
->putback_page(page
);
156 __ClearPageIsolated(page
);
160 * Put previously isolated pages back onto the appropriate lists
161 * from where they were once taken off for compaction/migration.
163 * This function shall be used whenever the isolated pageset has been
164 * built from lru, balloon, hugetlbfs page. See isolate_migratepages_range()
165 * and isolate_huge_page().
167 void putback_movable_pages(struct list_head
*l
)
172 list_for_each_entry_safe(page
, page2
, l
, lru
) {
173 if (unlikely(PageHuge(page
))) {
174 putback_active_hugepage(page
);
177 list_del(&page
->lru
);
179 * We isolated non-lru movable page so here we can use
180 * __PageMovable because LRU page's mapping cannot have
181 * PAGE_MAPPING_MOVABLE.
183 if (unlikely(__PageMovable(page
))) {
184 VM_BUG_ON_PAGE(!PageIsolated(page
), page
);
186 if (PageMovable(page
))
187 putback_movable_page(page
);
189 __ClearPageIsolated(page
);
193 mod_node_page_state(page_pgdat(page
), NR_ISOLATED_ANON
+
194 page_is_file_cache(page
), -hpage_nr_pages(page
));
195 putback_lru_page(page
);
201 * Restore a potential migration pte to a working pte entry
203 static bool remove_migration_pte(struct page
*page
, struct vm_area_struct
*vma
,
204 unsigned long addr
, void *old
)
206 struct page_vma_mapped_walk pvmw
= {
210 .flags
= PVMW_SYNC
| PVMW_MIGRATION
,
216 VM_BUG_ON_PAGE(PageTail(page
), page
);
217 while (page_vma_mapped_walk(&pvmw
)) {
221 new = page
- pvmw
.page
->index
+
222 linear_page_index(vma
, pvmw
.address
);
224 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
225 /* PMD-mapped THP migration entry */
227 VM_BUG_ON_PAGE(PageHuge(page
) || !PageTransCompound(page
), page
);
228 remove_migration_pmd(&pvmw
, new);
234 pte
= pte_mkold(mk_pte(new, READ_ONCE(vma
->vm_page_prot
)));
235 if (pte_swp_soft_dirty(*pvmw
.pte
))
236 pte
= pte_mksoft_dirty(pte
);
239 * Recheck VMA as permissions can change since migration started
241 entry
= pte_to_swp_entry(*pvmw
.pte
);
242 if (is_write_migration_entry(entry
))
243 pte
= maybe_mkwrite(pte
, vma
);
245 if (unlikely(is_zone_device_page(new))) {
246 if (is_device_private_page(new)) {
247 entry
= make_device_private_entry(new, pte_write(pte
));
248 pte
= swp_entry_to_pte(entry
);
252 #ifdef CONFIG_HUGETLB_PAGE
254 pte
= pte_mkhuge(pte
);
255 pte
= arch_make_huge_pte(pte
, vma
, new, 0);
256 set_huge_pte_at(vma
->vm_mm
, pvmw
.address
, pvmw
.pte
, pte
);
258 hugepage_add_anon_rmap(new, vma
, pvmw
.address
);
260 page_dup_rmap(new, true);
264 set_pte_at(vma
->vm_mm
, pvmw
.address
, pvmw
.pte
, pte
);
267 page_add_anon_rmap(new, vma
, pvmw
.address
, false);
269 page_add_file_rmap(new, false);
271 if (vma
->vm_flags
& VM_LOCKED
&& !PageTransCompound(new))
274 if (PageTransHuge(page
) && PageMlocked(page
))
275 clear_page_mlock(page
);
277 /* No need to invalidate - it was non-present before */
278 update_mmu_cache(vma
, pvmw
.address
, pvmw
.pte
);
285 * Get rid of all migration entries and replace them by
286 * references to the indicated page.
288 void remove_migration_ptes(struct page
*old
, struct page
*new, bool locked
)
290 struct rmap_walk_control rwc
= {
291 .rmap_one
= remove_migration_pte
,
296 rmap_walk_locked(new, &rwc
);
298 rmap_walk(new, &rwc
);
302 * Something used the pte of a page under migration. We need to
303 * get to the page and wait until migration is finished.
304 * When we return from this function the fault will be retried.
306 void __migration_entry_wait(struct mm_struct
*mm
, pte_t
*ptep
,
315 if (!is_swap_pte(pte
))
318 entry
= pte_to_swp_entry(pte
);
319 if (!is_migration_entry(entry
))
322 page
= migration_entry_to_page(entry
);
325 * Once page cache replacement of page migration started, page_count
326 * is zero; but we must not call put_and_wait_on_page_locked() without
327 * a ref. Use get_page_unless_zero(), and just fault again if it fails.
329 if (!get_page_unless_zero(page
))
331 pte_unmap_unlock(ptep
, ptl
);
332 put_and_wait_on_page_locked(page
);
335 pte_unmap_unlock(ptep
, ptl
);
338 void migration_entry_wait(struct mm_struct
*mm
, pmd_t
*pmd
,
339 unsigned long address
)
341 spinlock_t
*ptl
= pte_lockptr(mm
, pmd
);
342 pte_t
*ptep
= pte_offset_map(pmd
, address
);
343 __migration_entry_wait(mm
, ptep
, ptl
);
346 void migration_entry_wait_huge(struct vm_area_struct
*vma
,
347 struct mm_struct
*mm
, pte_t
*pte
)
349 spinlock_t
*ptl
= huge_pte_lockptr(hstate_vma(vma
), mm
, pte
);
350 __migration_entry_wait(mm
, pte
, ptl
);
353 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
354 void pmd_migration_entry_wait(struct mm_struct
*mm
, pmd_t
*pmd
)
359 ptl
= pmd_lock(mm
, pmd
);
360 if (!is_pmd_migration_entry(*pmd
))
362 page
= migration_entry_to_page(pmd_to_swp_entry(*pmd
));
363 if (!get_page_unless_zero(page
))
366 put_and_wait_on_page_locked(page
);
373 static int expected_page_refs(struct address_space
*mapping
, struct page
*page
)
375 int expected_count
= 1;
378 * Device public or private pages have an extra refcount as they are
381 expected_count
+= is_device_private_page(page
);
383 expected_count
+= hpage_nr_pages(page
) + page_has_private(page
);
385 return expected_count
;
389 * Replace the page in the mapping.
391 * The number of remaining references must be:
392 * 1 for anonymous pages without a mapping
393 * 2 for pages with a mapping
394 * 3 for pages with a mapping and PagePrivate/PagePrivate2 set.
396 int migrate_page_move_mapping(struct address_space
*mapping
,
397 struct page
*newpage
, struct page
*page
, int extra_count
)
399 XA_STATE(xas
, &mapping
->i_pages
, page_index(page
));
400 struct zone
*oldzone
, *newzone
;
402 int expected_count
= expected_page_refs(mapping
, page
) + extra_count
;
405 /* Anonymous page without mapping */
406 if (page_count(page
) != expected_count
)
409 /* No turning back from here */
410 newpage
->index
= page
->index
;
411 newpage
->mapping
= page
->mapping
;
412 if (PageSwapBacked(page
))
413 __SetPageSwapBacked(newpage
);
415 return MIGRATEPAGE_SUCCESS
;
418 oldzone
= page_zone(page
);
419 newzone
= page_zone(newpage
);
422 if (page_count(page
) != expected_count
|| xas_load(&xas
) != page
) {
423 xas_unlock_irq(&xas
);
427 if (!page_ref_freeze(page
, expected_count
)) {
428 xas_unlock_irq(&xas
);
433 * Now we know that no one else is looking at the page:
434 * no turning back from here.
436 newpage
->index
= page
->index
;
437 newpage
->mapping
= page
->mapping
;
438 page_ref_add(newpage
, hpage_nr_pages(page
)); /* add cache reference */
439 if (PageSwapBacked(page
)) {
440 __SetPageSwapBacked(newpage
);
441 if (PageSwapCache(page
)) {
442 SetPageSwapCache(newpage
);
443 set_page_private(newpage
, page_private(page
));
446 VM_BUG_ON_PAGE(PageSwapCache(page
), page
);
449 /* Move dirty while page refs frozen and newpage not yet exposed */
450 dirty
= PageDirty(page
);
452 ClearPageDirty(page
);
453 SetPageDirty(newpage
);
456 xas_store(&xas
, newpage
);
457 if (PageTransHuge(page
)) {
460 for (i
= 1; i
< HPAGE_PMD_NR
; i
++) {
462 xas_store(&xas
, newpage
+ i
);
467 * Drop cache reference from old page by unfreezing
468 * to one less reference.
469 * We know this isn't the last reference.
471 page_ref_unfreeze(page
, expected_count
- hpage_nr_pages(page
));
474 /* Leave irq disabled to prevent preemption while updating stats */
477 * If moved to a different zone then also account
478 * the page for that zone. Other VM counters will be
479 * taken care of when we establish references to the
480 * new page and drop references to the old page.
482 * Note that anonymous pages are accounted for
483 * via NR_FILE_PAGES and NR_ANON_MAPPED if they
484 * are mapped to swap space.
486 if (newzone
!= oldzone
) {
487 __dec_node_state(oldzone
->zone_pgdat
, NR_FILE_PAGES
);
488 __inc_node_state(newzone
->zone_pgdat
, NR_FILE_PAGES
);
489 if (PageSwapBacked(page
) && !PageSwapCache(page
)) {
490 __dec_node_state(oldzone
->zone_pgdat
, NR_SHMEM
);
491 __inc_node_state(newzone
->zone_pgdat
, NR_SHMEM
);
493 if (dirty
&& mapping_cap_account_dirty(mapping
)) {
494 __dec_node_state(oldzone
->zone_pgdat
, NR_FILE_DIRTY
);
495 __dec_zone_state(oldzone
, NR_ZONE_WRITE_PENDING
);
496 __inc_node_state(newzone
->zone_pgdat
, NR_FILE_DIRTY
);
497 __inc_zone_state(newzone
, NR_ZONE_WRITE_PENDING
);
502 return MIGRATEPAGE_SUCCESS
;
504 EXPORT_SYMBOL(migrate_page_move_mapping
);
507 * The expected number of remaining references is the same as that
508 * of migrate_page_move_mapping().
510 int migrate_huge_page_move_mapping(struct address_space
*mapping
,
511 struct page
*newpage
, struct page
*page
)
513 XA_STATE(xas
, &mapping
->i_pages
, page_index(page
));
517 expected_count
= 2 + page_has_private(page
);
518 if (page_count(page
) != expected_count
|| xas_load(&xas
) != page
) {
519 xas_unlock_irq(&xas
);
523 if (!page_ref_freeze(page
, expected_count
)) {
524 xas_unlock_irq(&xas
);
528 newpage
->index
= page
->index
;
529 newpage
->mapping
= page
->mapping
;
533 xas_store(&xas
, newpage
);
535 page_ref_unfreeze(page
, expected_count
- 1);
537 xas_unlock_irq(&xas
);
539 return MIGRATEPAGE_SUCCESS
;
543 * Gigantic pages are so large that we do not guarantee that page++ pointer
544 * arithmetic will work across the entire page. We need something more
547 static void __copy_gigantic_page(struct page
*dst
, struct page
*src
,
551 struct page
*dst_base
= dst
;
552 struct page
*src_base
= src
;
554 for (i
= 0; i
< nr_pages
; ) {
556 copy_highpage(dst
, src
);
559 dst
= mem_map_next(dst
, dst_base
, i
);
560 src
= mem_map_next(src
, src_base
, i
);
564 static void copy_huge_page(struct page
*dst
, struct page
*src
)
571 struct hstate
*h
= page_hstate(src
);
572 nr_pages
= pages_per_huge_page(h
);
574 if (unlikely(nr_pages
> MAX_ORDER_NR_PAGES
)) {
575 __copy_gigantic_page(dst
, src
, nr_pages
);
580 BUG_ON(!PageTransHuge(src
));
581 nr_pages
= hpage_nr_pages(src
);
584 for (i
= 0; i
< nr_pages
; i
++) {
586 copy_highpage(dst
+ i
, src
+ i
);
591 * Copy the page to its new location
593 void migrate_page_states(struct page
*newpage
, struct page
*page
)
598 SetPageError(newpage
);
599 if (PageReferenced(page
))
600 SetPageReferenced(newpage
);
601 if (PageUptodate(page
))
602 SetPageUptodate(newpage
);
603 if (TestClearPageActive(page
)) {
604 VM_BUG_ON_PAGE(PageUnevictable(page
), page
);
605 SetPageActive(newpage
);
606 } else if (TestClearPageUnevictable(page
))
607 SetPageUnevictable(newpage
);
608 if (PageWorkingset(page
))
609 SetPageWorkingset(newpage
);
610 if (PageChecked(page
))
611 SetPageChecked(newpage
);
612 if (PageMappedToDisk(page
))
613 SetPageMappedToDisk(newpage
);
615 /* Move dirty on pages not done by migrate_page_move_mapping() */
617 SetPageDirty(newpage
);
619 if (page_is_young(page
))
620 set_page_young(newpage
);
621 if (page_is_idle(page
))
622 set_page_idle(newpage
);
625 * Copy NUMA information to the new page, to prevent over-eager
626 * future migrations of this same page.
628 cpupid
= page_cpupid_xchg_last(page
, -1);
629 page_cpupid_xchg_last(newpage
, cpupid
);
631 ksm_migrate_page(newpage
, page
);
633 * Please do not reorder this without considering how mm/ksm.c's
634 * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache().
636 if (PageSwapCache(page
))
637 ClearPageSwapCache(page
);
638 ClearPagePrivate(page
);
639 set_page_private(page
, 0);
642 * If any waiters have accumulated on the new page then
645 if (PageWriteback(newpage
))
646 end_page_writeback(newpage
);
648 copy_page_owner(page
, newpage
);
650 mem_cgroup_migrate(page
, newpage
);
652 EXPORT_SYMBOL(migrate_page_states
);
654 void migrate_page_copy(struct page
*newpage
, struct page
*page
)
656 if (PageHuge(page
) || PageTransHuge(page
))
657 copy_huge_page(newpage
, page
);
659 copy_highpage(newpage
, page
);
661 migrate_page_states(newpage
, page
);
663 EXPORT_SYMBOL(migrate_page_copy
);
665 /************************************************************
666 * Migration functions
667 ***********************************************************/
670 * Common logic to directly migrate a single LRU page suitable for
671 * pages that do not use PagePrivate/PagePrivate2.
673 * Pages are locked upon entry and exit.
675 int migrate_page(struct address_space
*mapping
,
676 struct page
*newpage
, struct page
*page
,
677 enum migrate_mode mode
)
681 BUG_ON(PageWriteback(page
)); /* Writeback must be complete */
683 rc
= migrate_page_move_mapping(mapping
, newpage
, page
, 0);
685 if (rc
!= MIGRATEPAGE_SUCCESS
)
688 if (mode
!= MIGRATE_SYNC_NO_COPY
)
689 migrate_page_copy(newpage
, page
);
691 migrate_page_states(newpage
, page
);
692 return MIGRATEPAGE_SUCCESS
;
694 EXPORT_SYMBOL(migrate_page
);
697 /* Returns true if all buffers are successfully locked */
698 static bool buffer_migrate_lock_buffers(struct buffer_head
*head
,
699 enum migrate_mode mode
)
701 struct buffer_head
*bh
= head
;
703 /* Simple case, sync compaction */
704 if (mode
!= MIGRATE_ASYNC
) {
707 bh
= bh
->b_this_page
;
709 } while (bh
!= head
);
714 /* async case, we cannot block on lock_buffer so use trylock_buffer */
716 if (!trylock_buffer(bh
)) {
718 * We failed to lock the buffer and cannot stall in
719 * async migration. Release the taken locks
721 struct buffer_head
*failed_bh
= bh
;
723 while (bh
!= failed_bh
) {
725 bh
= bh
->b_this_page
;
730 bh
= bh
->b_this_page
;
731 } while (bh
!= head
);
735 static int __buffer_migrate_page(struct address_space
*mapping
,
736 struct page
*newpage
, struct page
*page
, enum migrate_mode mode
,
739 struct buffer_head
*bh
, *head
;
743 if (!page_has_buffers(page
))
744 return migrate_page(mapping
, newpage
, page
, mode
);
746 /* Check whether page does not have extra refs before we do more work */
747 expected_count
= expected_page_refs(mapping
, page
);
748 if (page_count(page
) != expected_count
)
751 head
= page_buffers(page
);
752 if (!buffer_migrate_lock_buffers(head
, mode
))
757 bool invalidated
= false;
761 spin_lock(&mapping
->private_lock
);
764 if (atomic_read(&bh
->b_count
)) {
768 bh
= bh
->b_this_page
;
769 } while (bh
!= head
);
775 spin_unlock(&mapping
->private_lock
);
776 invalidate_bh_lrus();
778 goto recheck_buffers
;
782 rc
= migrate_page_move_mapping(mapping
, newpage
, page
, 0);
783 if (rc
!= MIGRATEPAGE_SUCCESS
)
786 ClearPagePrivate(page
);
787 set_page_private(newpage
, page_private(page
));
788 set_page_private(page
, 0);
794 set_bh_page(bh
, newpage
, bh_offset(bh
));
795 bh
= bh
->b_this_page
;
797 } while (bh
!= head
);
799 SetPagePrivate(newpage
);
801 if (mode
!= MIGRATE_SYNC_NO_COPY
)
802 migrate_page_copy(newpage
, page
);
804 migrate_page_states(newpage
, page
);
806 rc
= MIGRATEPAGE_SUCCESS
;
809 spin_unlock(&mapping
->private_lock
);
813 bh
= bh
->b_this_page
;
815 } while (bh
!= head
);
821 * Migration function for pages with buffers. This function can only be used
822 * if the underlying filesystem guarantees that no other references to "page"
823 * exist. For example attached buffer heads are accessed only under page lock.
825 int buffer_migrate_page(struct address_space
*mapping
,
826 struct page
*newpage
, struct page
*page
, enum migrate_mode mode
)
828 return __buffer_migrate_page(mapping
, newpage
, page
, mode
, false);
830 EXPORT_SYMBOL(buffer_migrate_page
);
833 * Same as above except that this variant is more careful and checks that there
834 * are also no buffer head references. This function is the right one for
835 * mappings where buffer heads are directly looked up and referenced (such as
836 * block device mappings).
838 int buffer_migrate_page_norefs(struct address_space
*mapping
,
839 struct page
*newpage
, struct page
*page
, enum migrate_mode mode
)
841 return __buffer_migrate_page(mapping
, newpage
, page
, mode
, true);
846 * Writeback a page to clean the dirty state
848 static int writeout(struct address_space
*mapping
, struct page
*page
)
850 struct writeback_control wbc
= {
851 .sync_mode
= WB_SYNC_NONE
,
854 .range_end
= LLONG_MAX
,
859 if (!mapping
->a_ops
->writepage
)
860 /* No write method for the address space */
863 if (!clear_page_dirty_for_io(page
))
864 /* Someone else already triggered a write */
868 * A dirty page may imply that the underlying filesystem has
869 * the page on some queue. So the page must be clean for
870 * migration. Writeout may mean we loose the lock and the
871 * page state is no longer what we checked for earlier.
872 * At this point we know that the migration attempt cannot
875 remove_migration_ptes(page
, page
, false);
877 rc
= mapping
->a_ops
->writepage(page
, &wbc
);
879 if (rc
!= AOP_WRITEPAGE_ACTIVATE
)
880 /* unlocked. Relock */
883 return (rc
< 0) ? -EIO
: -EAGAIN
;
887 * Default handling if a filesystem does not provide a migration function.
889 static int fallback_migrate_page(struct address_space
*mapping
,
890 struct page
*newpage
, struct page
*page
, enum migrate_mode mode
)
892 if (PageDirty(page
)) {
893 /* Only writeback pages in full synchronous migration */
896 case MIGRATE_SYNC_NO_COPY
:
901 return writeout(mapping
, page
);
905 * Buffers may be managed in a filesystem specific way.
906 * We must have no buffers or drop them.
908 if (page_has_private(page
) &&
909 !try_to_release_page(page
, GFP_KERNEL
))
910 return mode
== MIGRATE_SYNC
? -EAGAIN
: -EBUSY
;
912 return migrate_page(mapping
, newpage
, page
, mode
);
916 * Move a page to a newly allocated page
917 * The page is locked and all ptes have been successfully removed.
919 * The new page will have replaced the old page if this function
924 * MIGRATEPAGE_SUCCESS - success
926 static int move_to_new_page(struct page
*newpage
, struct page
*page
,
927 enum migrate_mode mode
)
929 struct address_space
*mapping
;
931 bool is_lru
= !__PageMovable(page
);
933 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
934 VM_BUG_ON_PAGE(!PageLocked(newpage
), newpage
);
936 mapping
= page_mapping(page
);
938 if (likely(is_lru
)) {
940 rc
= migrate_page(mapping
, newpage
, page
, mode
);
941 else if (mapping
->a_ops
->migratepage
)
943 * Most pages have a mapping and most filesystems
944 * provide a migratepage callback. Anonymous pages
945 * are part of swap space which also has its own
946 * migratepage callback. This is the most common path
947 * for page migration.
949 rc
= mapping
->a_ops
->migratepage(mapping
, newpage
,
952 rc
= fallback_migrate_page(mapping
, newpage
,
956 * In case of non-lru page, it could be released after
957 * isolation step. In that case, we shouldn't try migration.
959 VM_BUG_ON_PAGE(!PageIsolated(page
), page
);
960 if (!PageMovable(page
)) {
961 rc
= MIGRATEPAGE_SUCCESS
;
962 __ClearPageIsolated(page
);
966 rc
= mapping
->a_ops
->migratepage(mapping
, newpage
,
968 WARN_ON_ONCE(rc
== MIGRATEPAGE_SUCCESS
&&
969 !PageIsolated(page
));
973 * When successful, old pagecache page->mapping must be cleared before
974 * page is freed; but stats require that PageAnon be left as PageAnon.
976 if (rc
== MIGRATEPAGE_SUCCESS
) {
977 if (__PageMovable(page
)) {
978 VM_BUG_ON_PAGE(!PageIsolated(page
), page
);
981 * We clear PG_movable under page_lock so any compactor
982 * cannot try to migrate this page.
984 __ClearPageIsolated(page
);
988 * Anonymous and movable page->mapping will be cleard by
989 * free_pages_prepare so don't reset it here for keeping
990 * the type to work PageAnon, for example.
992 if (!PageMappingFlags(page
))
993 page
->mapping
= NULL
;
995 if (likely(!is_zone_device_page(newpage
)))
996 flush_dcache_page(newpage
);
1003 static int __unmap_and_move(struct page
*page
, struct page
*newpage
,
1004 int force
, enum migrate_mode mode
)
1007 int page_was_mapped
= 0;
1008 struct anon_vma
*anon_vma
= NULL
;
1009 bool is_lru
= !__PageMovable(page
);
1011 if (!trylock_page(page
)) {
1012 if (!force
|| mode
== MIGRATE_ASYNC
)
1016 * It's not safe for direct compaction to call lock_page.
1017 * For example, during page readahead pages are added locked
1018 * to the LRU. Later, when the IO completes the pages are
1019 * marked uptodate and unlocked. However, the queueing
1020 * could be merging multiple pages for one bio (e.g.
1021 * mpage_readpages). If an allocation happens for the
1022 * second or third page, the process can end up locking
1023 * the same page twice and deadlocking. Rather than
1024 * trying to be clever about what pages can be locked,
1025 * avoid the use of lock_page for direct compaction
1028 if (current
->flags
& PF_MEMALLOC
)
1034 if (PageWriteback(page
)) {
1036 * Only in the case of a full synchronous migration is it
1037 * necessary to wait for PageWriteback. In the async case,
1038 * the retry loop is too short and in the sync-light case,
1039 * the overhead of stalling is too much
1043 case MIGRATE_SYNC_NO_COPY
:
1051 wait_on_page_writeback(page
);
1055 * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
1056 * we cannot notice that anon_vma is freed while we migrates a page.
1057 * This get_anon_vma() delays freeing anon_vma pointer until the end
1058 * of migration. File cache pages are no problem because of page_lock()
1059 * File Caches may use write_page() or lock_page() in migration, then,
1060 * just care Anon page here.
1062 * Only page_get_anon_vma() understands the subtleties of
1063 * getting a hold on an anon_vma from outside one of its mms.
1064 * But if we cannot get anon_vma, then we won't need it anyway,
1065 * because that implies that the anon page is no longer mapped
1066 * (and cannot be remapped so long as we hold the page lock).
1068 if (PageAnon(page
) && !PageKsm(page
))
1069 anon_vma
= page_get_anon_vma(page
);
1072 * Block others from accessing the new page when we get around to
1073 * establishing additional references. We are usually the only one
1074 * holding a reference to newpage at this point. We used to have a BUG
1075 * here if trylock_page(newpage) fails, but would like to allow for
1076 * cases where there might be a race with the previous use of newpage.
1077 * This is much like races on refcount of oldpage: just don't BUG().
1079 if (unlikely(!trylock_page(newpage
)))
1082 if (unlikely(!is_lru
)) {
1083 rc
= move_to_new_page(newpage
, page
, mode
);
1084 goto out_unlock_both
;
1088 * Corner case handling:
1089 * 1. When a new swap-cache page is read into, it is added to the LRU
1090 * and treated as swapcache but it has no rmap yet.
1091 * Calling try_to_unmap() against a page->mapping==NULL page will
1092 * trigger a BUG. So handle it here.
1093 * 2. An orphaned page (see truncate_complete_page) might have
1094 * fs-private metadata. The page can be picked up due to memory
1095 * offlining. Everywhere else except page reclaim, the page is
1096 * invisible to the vm, so the page can not be migrated. So try to
1097 * free the metadata, so the page can be freed.
1099 if (!page
->mapping
) {
1100 VM_BUG_ON_PAGE(PageAnon(page
), page
);
1101 if (page_has_private(page
)) {
1102 try_to_free_buffers(page
);
1103 goto out_unlock_both
;
1105 } else if (page_mapped(page
)) {
1106 /* Establish migration ptes */
1107 VM_BUG_ON_PAGE(PageAnon(page
) && !PageKsm(page
) && !anon_vma
,
1110 TTU_MIGRATION
|TTU_IGNORE_MLOCK
|TTU_IGNORE_ACCESS
);
1111 page_was_mapped
= 1;
1114 if (!page_mapped(page
))
1115 rc
= move_to_new_page(newpage
, page
, mode
);
1117 if (page_was_mapped
)
1118 remove_migration_ptes(page
,
1119 rc
== MIGRATEPAGE_SUCCESS
? newpage
: page
, false);
1122 unlock_page(newpage
);
1124 /* Drop an anon_vma reference if we took one */
1126 put_anon_vma(anon_vma
);
1130 * If migration is successful, decrease refcount of the newpage
1131 * which will not free the page because new page owner increased
1132 * refcounter. As well, if it is LRU page, add the page to LRU
1133 * list in here. Use the old state of the isolated source page to
1134 * determine if we migrated a LRU page. newpage was already unlocked
1135 * and possibly modified by its owner - don't rely on the page
1138 if (rc
== MIGRATEPAGE_SUCCESS
) {
1139 if (unlikely(!is_lru
))
1142 putback_lru_page(newpage
);
1149 * gcc 4.7 and 4.8 on arm get an ICEs when inlining unmap_and_move(). Work
1152 #if defined(CONFIG_ARM) && \
1153 defined(GCC_VERSION) && GCC_VERSION < 40900 && GCC_VERSION >= 40700
1154 #define ICE_noinline noinline
1156 #define ICE_noinline
1160 * Obtain the lock on page, remove all ptes and migrate the page
1161 * to the newly allocated page in newpage.
1163 static ICE_noinline
int unmap_and_move(new_page_t get_new_page
,
1164 free_page_t put_new_page
,
1165 unsigned long private, struct page
*page
,
1166 int force
, enum migrate_mode mode
,
1167 enum migrate_reason reason
)
1169 int rc
= MIGRATEPAGE_SUCCESS
;
1170 struct page
*newpage
;
1172 if (!thp_migration_supported() && PageTransHuge(page
))
1175 newpage
= get_new_page(page
, private);
1179 if (page_count(page
) == 1) {
1180 /* page was freed from under us. So we are done. */
1181 ClearPageActive(page
);
1182 ClearPageUnevictable(page
);
1183 if (unlikely(__PageMovable(page
))) {
1185 if (!PageMovable(page
))
1186 __ClearPageIsolated(page
);
1190 put_new_page(newpage
, private);
1196 rc
= __unmap_and_move(page
, newpage
, force
, mode
);
1197 if (rc
== MIGRATEPAGE_SUCCESS
)
1198 set_page_owner_migrate_reason(newpage
, reason
);
1201 if (rc
!= -EAGAIN
) {
1203 * A page that has been migrated has all references
1204 * removed and will be freed. A page that has not been
1205 * migrated will have kepts its references and be
1208 list_del(&page
->lru
);
1211 * Compaction can migrate also non-LRU pages which are
1212 * not accounted to NR_ISOLATED_*. They can be recognized
1215 if (likely(!__PageMovable(page
)))
1216 mod_node_page_state(page_pgdat(page
), NR_ISOLATED_ANON
+
1217 page_is_file_cache(page
), -hpage_nr_pages(page
));
1221 * If migration is successful, releases reference grabbed during
1222 * isolation. Otherwise, restore the page to right list unless
1225 if (rc
== MIGRATEPAGE_SUCCESS
) {
1227 if (reason
== MR_MEMORY_FAILURE
) {
1229 * Set PG_HWPoison on just freed page
1230 * intentionally. Although it's rather weird,
1231 * it's how HWPoison flag works at the moment.
1233 if (set_hwpoison_free_buddy_page(page
))
1234 num_poisoned_pages_inc();
1237 if (rc
!= -EAGAIN
) {
1238 if (likely(!__PageMovable(page
))) {
1239 putback_lru_page(page
);
1244 if (PageMovable(page
))
1245 putback_movable_page(page
);
1247 __ClearPageIsolated(page
);
1253 put_new_page(newpage
, private);
1262 * Counterpart of unmap_and_move_page() for hugepage migration.
1264 * This function doesn't wait the completion of hugepage I/O
1265 * because there is no race between I/O and migration for hugepage.
1266 * Note that currently hugepage I/O occurs only in direct I/O
1267 * where no lock is held and PG_writeback is irrelevant,
1268 * and writeback status of all subpages are counted in the reference
1269 * count of the head page (i.e. if all subpages of a 2MB hugepage are
1270 * under direct I/O, the reference of the head page is 512 and a bit more.)
1271 * This means that when we try to migrate hugepage whose subpages are
1272 * doing direct I/O, some references remain after try_to_unmap() and
1273 * hugepage migration fails without data corruption.
1275 * There is also no race when direct I/O is issued on the page under migration,
1276 * because then pte is replaced with migration swap entry and direct I/O code
1277 * will wait in the page fault for migration to complete.
1279 static int unmap_and_move_huge_page(new_page_t get_new_page
,
1280 free_page_t put_new_page
, unsigned long private,
1281 struct page
*hpage
, int force
,
1282 enum migrate_mode mode
, int reason
)
1285 int page_was_mapped
= 0;
1286 struct page
*new_hpage
;
1287 struct anon_vma
*anon_vma
= NULL
;
1290 * Migratability of hugepages depends on architectures and their size.
1291 * This check is necessary because some callers of hugepage migration
1292 * like soft offline and memory hotremove don't walk through page
1293 * tables or check whether the hugepage is pmd-based or not before
1294 * kicking migration.
1296 if (!hugepage_migration_supported(page_hstate(hpage
))) {
1297 putback_active_hugepage(hpage
);
1301 new_hpage
= get_new_page(hpage
, private);
1305 if (!trylock_page(hpage
)) {
1310 case MIGRATE_SYNC_NO_COPY
:
1319 * Check for pages which are in the process of being freed. Without
1320 * page_mapping() set, hugetlbfs specific move page routine will not
1321 * be called and we could leak usage counts for subpools.
1323 if (page_private(hpage
) && !page_mapping(hpage
)) {
1328 if (PageAnon(hpage
))
1329 anon_vma
= page_get_anon_vma(hpage
);
1331 if (unlikely(!trylock_page(new_hpage
)))
1334 if (page_mapped(hpage
)) {
1336 TTU_MIGRATION
|TTU_IGNORE_MLOCK
|TTU_IGNORE_ACCESS
);
1337 page_was_mapped
= 1;
1340 if (!page_mapped(hpage
))
1341 rc
= move_to_new_page(new_hpage
, hpage
, mode
);
1343 if (page_was_mapped
)
1344 remove_migration_ptes(hpage
,
1345 rc
== MIGRATEPAGE_SUCCESS
? new_hpage
: hpage
, false);
1347 unlock_page(new_hpage
);
1351 put_anon_vma(anon_vma
);
1353 if (rc
== MIGRATEPAGE_SUCCESS
) {
1354 move_hugetlb_state(hpage
, new_hpage
, reason
);
1355 put_new_page
= NULL
;
1362 putback_active_hugepage(hpage
);
1365 * If migration was not successful and there's a freeing callback, use
1366 * it. Otherwise, put_page() will drop the reference grabbed during
1370 put_new_page(new_hpage
, private);
1372 putback_active_hugepage(new_hpage
);
1378 * migrate_pages - migrate the pages specified in a list, to the free pages
1379 * supplied as the target for the page migration
1381 * @from: The list of pages to be migrated.
1382 * @get_new_page: The function used to allocate free pages to be used
1383 * as the target of the page migration.
1384 * @put_new_page: The function used to free target pages if migration
1385 * fails, or NULL if no special handling is necessary.
1386 * @private: Private data to be passed on to get_new_page()
1387 * @mode: The migration mode that specifies the constraints for
1388 * page migration, if any.
1389 * @reason: The reason for page migration.
1391 * The function returns after 10 attempts or if no pages are movable any more
1392 * because the list has become empty or no retryable pages exist any more.
1393 * The caller should call putback_movable_pages() to return pages to the LRU
1394 * or free list only if ret != 0.
1396 * Returns the number of pages that were not migrated, or an error code.
1398 int migrate_pages(struct list_head
*from
, new_page_t get_new_page
,
1399 free_page_t put_new_page
, unsigned long private,
1400 enum migrate_mode mode
, int reason
)
1404 int nr_succeeded
= 0;
1408 int swapwrite
= current
->flags
& PF_SWAPWRITE
;
1412 current
->flags
|= PF_SWAPWRITE
;
1414 for(pass
= 0; pass
< 10 && retry
; pass
++) {
1417 list_for_each_entry_safe(page
, page2
, from
, lru
) {
1422 rc
= unmap_and_move_huge_page(get_new_page
,
1423 put_new_page
, private, page
,
1424 pass
> 2, mode
, reason
);
1426 rc
= unmap_and_move(get_new_page
, put_new_page
,
1427 private, page
, pass
> 2, mode
,
1433 * THP migration might be unsupported or the
1434 * allocation could've failed so we should
1435 * retry on the same page with the THP split
1438 * Head page is retried immediately and tail
1439 * pages are added to the tail of the list so
1440 * we encounter them after the rest of the list
1443 if (PageTransHuge(page
) && !PageHuge(page
)) {
1445 rc
= split_huge_page_to_list(page
, from
);
1448 list_safe_reset_next(page
, page2
, lru
);
1457 case MIGRATEPAGE_SUCCESS
:
1462 * Permanent failure (-EBUSY, -ENOSYS, etc.):
1463 * unlike -EAGAIN case, the failed page is
1464 * removed from migration page list and not
1465 * retried in the next outer loop.
1476 count_vm_events(PGMIGRATE_SUCCESS
, nr_succeeded
);
1478 count_vm_events(PGMIGRATE_FAIL
, nr_failed
);
1479 trace_mm_migrate_pages(nr_succeeded
, nr_failed
, mode
, reason
);
1482 current
->flags
&= ~PF_SWAPWRITE
;
1489 static int store_status(int __user
*status
, int start
, int value
, int nr
)
1492 if (put_user(value
, status
+ start
))
1500 static int do_move_pages_to_node(struct mm_struct
*mm
,
1501 struct list_head
*pagelist
, int node
)
1505 if (list_empty(pagelist
))
1508 err
= migrate_pages(pagelist
, alloc_new_node_page
, NULL
, node
,
1509 MIGRATE_SYNC
, MR_SYSCALL
);
1511 putback_movable_pages(pagelist
);
1516 * Resolves the given address to a struct page, isolates it from the LRU and
1517 * puts it to the given pagelist.
1519 * errno - if the page cannot be found/isolated
1520 * 0 - when it doesn't have to be migrated because it is already on the
1522 * 1 - when it has been queued
1524 static int add_page_for_migration(struct mm_struct
*mm
, unsigned long addr
,
1525 int node
, struct list_head
*pagelist
, bool migrate_all
)
1527 struct vm_area_struct
*vma
;
1529 unsigned int follflags
;
1532 down_read(&mm
->mmap_sem
);
1534 vma
= find_vma(mm
, addr
);
1535 if (!vma
|| addr
< vma
->vm_start
|| !vma_migratable(vma
))
1538 /* FOLL_DUMP to ignore special (like zero) pages */
1539 follflags
= FOLL_GET
| FOLL_DUMP
;
1540 page
= follow_page(vma
, addr
, follflags
);
1542 err
= PTR_ERR(page
);
1551 if (page_to_nid(page
) == node
)
1555 if (page_mapcount(page
) > 1 && !migrate_all
)
1558 if (PageHuge(page
)) {
1559 if (PageHead(page
)) {
1560 isolate_huge_page(page
, pagelist
);
1566 head
= compound_head(page
);
1567 err
= isolate_lru_page(head
);
1572 list_add_tail(&head
->lru
, pagelist
);
1573 mod_node_page_state(page_pgdat(head
),
1574 NR_ISOLATED_ANON
+ page_is_file_cache(head
),
1575 hpage_nr_pages(head
));
1579 * Either remove the duplicate refcount from
1580 * isolate_lru_page() or drop the page ref if it was
1585 up_read(&mm
->mmap_sem
);
1590 * Migrate an array of page address onto an array of nodes and fill
1591 * the corresponding array of status.
1593 static int do_pages_move(struct mm_struct
*mm
, nodemask_t task_nodes
,
1594 unsigned long nr_pages
,
1595 const void __user
* __user
*pages
,
1596 const int __user
*nodes
,
1597 int __user
*status
, int flags
)
1599 int current_node
= NUMA_NO_NODE
;
1600 LIST_HEAD(pagelist
);
1606 for (i
= start
= 0; i
< nr_pages
; i
++) {
1607 const void __user
*p
;
1612 if (get_user(p
, pages
+ i
))
1614 if (get_user(node
, nodes
+ i
))
1616 addr
= (unsigned long)p
;
1619 if (node
< 0 || node
>= MAX_NUMNODES
)
1621 if (!node_state(node
, N_MEMORY
))
1625 if (!node_isset(node
, task_nodes
))
1628 if (current_node
== NUMA_NO_NODE
) {
1629 current_node
= node
;
1631 } else if (node
!= current_node
) {
1632 err
= do_move_pages_to_node(mm
, &pagelist
, current_node
);
1635 err
= store_status(status
, start
, current_node
, i
- start
);
1639 current_node
= node
;
1643 * Errors in the page lookup or isolation are not fatal and we simply
1644 * report them via status
1646 err
= add_page_for_migration(mm
, addr
, current_node
,
1647 &pagelist
, flags
& MPOL_MF_MOVE_ALL
);
1650 /* The page is already on the target node */
1651 err
= store_status(status
, i
, current_node
, 1);
1655 } else if (err
> 0) {
1656 /* The page is successfully queued for migration */
1660 err
= store_status(status
, i
, err
, 1);
1664 err
= do_move_pages_to_node(mm
, &pagelist
, current_node
);
1668 err
= store_status(status
, start
, current_node
, i
- start
);
1672 current_node
= NUMA_NO_NODE
;
1675 if (list_empty(&pagelist
))
1678 /* Make sure we do not overwrite the existing error */
1679 err1
= do_move_pages_to_node(mm
, &pagelist
, current_node
);
1681 err1
= store_status(status
, start
, current_node
, i
- start
);
1689 * Determine the nodes of an array of pages and store it in an array of status.
1691 static void do_pages_stat_array(struct mm_struct
*mm
, unsigned long nr_pages
,
1692 const void __user
**pages
, int *status
)
1696 down_read(&mm
->mmap_sem
);
1698 for (i
= 0; i
< nr_pages
; i
++) {
1699 unsigned long addr
= (unsigned long)(*pages
);
1700 struct vm_area_struct
*vma
;
1704 vma
= find_vma(mm
, addr
);
1705 if (!vma
|| addr
< vma
->vm_start
)
1708 /* FOLL_DUMP to ignore special (like zero) pages */
1709 page
= follow_page(vma
, addr
, FOLL_DUMP
);
1711 err
= PTR_ERR(page
);
1715 err
= page
? page_to_nid(page
) : -ENOENT
;
1723 up_read(&mm
->mmap_sem
);
1727 * Determine the nodes of a user array of pages and store it in
1728 * a user array of status.
1730 static int do_pages_stat(struct mm_struct
*mm
, unsigned long nr_pages
,
1731 const void __user
* __user
*pages
,
1734 #define DO_PAGES_STAT_CHUNK_NR 16
1735 const void __user
*chunk_pages
[DO_PAGES_STAT_CHUNK_NR
];
1736 int chunk_status
[DO_PAGES_STAT_CHUNK_NR
];
1739 unsigned long chunk_nr
;
1741 chunk_nr
= nr_pages
;
1742 if (chunk_nr
> DO_PAGES_STAT_CHUNK_NR
)
1743 chunk_nr
= DO_PAGES_STAT_CHUNK_NR
;
1745 if (copy_from_user(chunk_pages
, pages
, chunk_nr
* sizeof(*chunk_pages
)))
1748 do_pages_stat_array(mm
, chunk_nr
, chunk_pages
, chunk_status
);
1750 if (copy_to_user(status
, chunk_status
, chunk_nr
* sizeof(*status
)))
1755 nr_pages
-= chunk_nr
;
1757 return nr_pages
? -EFAULT
: 0;
1761 * Move a list of pages in the address space of the currently executing
1764 static int kernel_move_pages(pid_t pid
, unsigned long nr_pages
,
1765 const void __user
* __user
*pages
,
1766 const int __user
*nodes
,
1767 int __user
*status
, int flags
)
1769 struct task_struct
*task
;
1770 struct mm_struct
*mm
;
1772 nodemask_t task_nodes
;
1775 if (flags
& ~(MPOL_MF_MOVE
|MPOL_MF_MOVE_ALL
))
1778 if ((flags
& MPOL_MF_MOVE_ALL
) && !capable(CAP_SYS_NICE
))
1781 /* Find the mm_struct */
1783 task
= pid
? find_task_by_vpid(pid
) : current
;
1788 get_task_struct(task
);
1791 * Check if this process has the right to modify the specified
1792 * process. Use the regular "ptrace_may_access()" checks.
1794 if (!ptrace_may_access(task
, PTRACE_MODE_READ_REALCREDS
)) {
1801 err
= security_task_movememory(task
);
1805 task_nodes
= cpuset_mems_allowed(task
);
1806 mm
= get_task_mm(task
);
1807 put_task_struct(task
);
1813 err
= do_pages_move(mm
, task_nodes
, nr_pages
, pages
,
1814 nodes
, status
, flags
);
1816 err
= do_pages_stat(mm
, nr_pages
, pages
, status
);
1822 put_task_struct(task
);
1826 SYSCALL_DEFINE6(move_pages
, pid_t
, pid
, unsigned long, nr_pages
,
1827 const void __user
* __user
*, pages
,
1828 const int __user
*, nodes
,
1829 int __user
*, status
, int, flags
)
1831 return kernel_move_pages(pid
, nr_pages
, pages
, nodes
, status
, flags
);
1834 #ifdef CONFIG_COMPAT
1835 COMPAT_SYSCALL_DEFINE6(move_pages
, pid_t
, pid
, compat_ulong_t
, nr_pages
,
1836 compat_uptr_t __user
*, pages32
,
1837 const int __user
*, nodes
,
1838 int __user
*, status
,
1841 const void __user
* __user
*pages
;
1844 pages
= compat_alloc_user_space(nr_pages
* sizeof(void *));
1845 for (i
= 0; i
< nr_pages
; i
++) {
1848 if (get_user(p
, pages32
+ i
) ||
1849 put_user(compat_ptr(p
), pages
+ i
))
1852 return kernel_move_pages(pid
, nr_pages
, pages
, nodes
, status
, flags
);
1854 #endif /* CONFIG_COMPAT */
1856 #ifdef CONFIG_NUMA_BALANCING
1858 * Returns true if this is a safe migration target node for misplaced NUMA
1859 * pages. Currently it only checks the watermarks which crude
1861 static bool migrate_balanced_pgdat(struct pglist_data
*pgdat
,
1862 unsigned long nr_migrate_pages
)
1866 for (z
= pgdat
->nr_zones
- 1; z
>= 0; z
--) {
1867 struct zone
*zone
= pgdat
->node_zones
+ z
;
1869 if (!populated_zone(zone
))
1872 /* Avoid waking kswapd by allocating pages_to_migrate pages. */
1873 if (!zone_watermark_ok(zone
, 0,
1874 high_wmark_pages(zone
) +
1883 static struct page
*alloc_misplaced_dst_page(struct page
*page
,
1886 int nid
= (int) data
;
1887 struct page
*newpage
;
1889 newpage
= __alloc_pages_node(nid
,
1890 (GFP_HIGHUSER_MOVABLE
|
1891 __GFP_THISNODE
| __GFP_NOMEMALLOC
|
1892 __GFP_NORETRY
| __GFP_NOWARN
) &
1898 static int numamigrate_isolate_page(pg_data_t
*pgdat
, struct page
*page
)
1902 VM_BUG_ON_PAGE(compound_order(page
) && !PageTransHuge(page
), page
);
1904 /* Avoid migrating to a node that is nearly full */
1905 if (!migrate_balanced_pgdat(pgdat
, 1UL << compound_order(page
)))
1908 if (isolate_lru_page(page
))
1912 * migrate_misplaced_transhuge_page() skips page migration's usual
1913 * check on page_count(), so we must do it here, now that the page
1914 * has been isolated: a GUP pin, or any other pin, prevents migration.
1915 * The expected page count is 3: 1 for page's mapcount and 1 for the
1916 * caller's pin and 1 for the reference taken by isolate_lru_page().
1918 if (PageTransHuge(page
) && page_count(page
) != 3) {
1919 putback_lru_page(page
);
1923 page_lru
= page_is_file_cache(page
);
1924 mod_node_page_state(page_pgdat(page
), NR_ISOLATED_ANON
+ page_lru
,
1925 hpage_nr_pages(page
));
1928 * Isolating the page has taken another reference, so the
1929 * caller's reference can be safely dropped without the page
1930 * disappearing underneath us during migration.
1936 bool pmd_trans_migrating(pmd_t pmd
)
1938 struct page
*page
= pmd_page(pmd
);
1939 return PageLocked(page
);
1943 * Attempt to migrate a misplaced page to the specified destination
1944 * node. Caller is expected to have an elevated reference count on
1945 * the page that will be dropped by this function before returning.
1947 int migrate_misplaced_page(struct page
*page
, struct vm_area_struct
*vma
,
1950 pg_data_t
*pgdat
= NODE_DATA(node
);
1953 LIST_HEAD(migratepages
);
1956 * Don't migrate file pages that are mapped in multiple processes
1957 * with execute permissions as they are probably shared libraries.
1959 if (page_mapcount(page
) != 1 && page_is_file_cache(page
) &&
1960 (vma
->vm_flags
& VM_EXEC
))
1964 * Also do not migrate dirty pages as not all filesystems can move
1965 * dirty pages in MIGRATE_ASYNC mode which is a waste of cycles.
1967 if (page_is_file_cache(page
) && PageDirty(page
))
1970 isolated
= numamigrate_isolate_page(pgdat
, page
);
1974 list_add(&page
->lru
, &migratepages
);
1975 nr_remaining
= migrate_pages(&migratepages
, alloc_misplaced_dst_page
,
1976 NULL
, node
, MIGRATE_ASYNC
,
1979 if (!list_empty(&migratepages
)) {
1980 list_del(&page
->lru
);
1981 dec_node_page_state(page
, NR_ISOLATED_ANON
+
1982 page_is_file_cache(page
));
1983 putback_lru_page(page
);
1987 count_vm_numa_event(NUMA_PAGE_MIGRATE
);
1988 BUG_ON(!list_empty(&migratepages
));
1995 #endif /* CONFIG_NUMA_BALANCING */
1997 #if defined(CONFIG_NUMA_BALANCING) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
1999 * Migrates a THP to a given target node. page must be locked and is unlocked
2002 int migrate_misplaced_transhuge_page(struct mm_struct
*mm
,
2003 struct vm_area_struct
*vma
,
2004 pmd_t
*pmd
, pmd_t entry
,
2005 unsigned long address
,
2006 struct page
*page
, int node
)
2009 pg_data_t
*pgdat
= NODE_DATA(node
);
2011 struct page
*new_page
= NULL
;
2012 int page_lru
= page_is_file_cache(page
);
2013 unsigned long start
= address
& HPAGE_PMD_MASK
;
2015 new_page
= alloc_pages_node(node
,
2016 (GFP_TRANSHUGE_LIGHT
| __GFP_THISNODE
),
2020 prep_transhuge_page(new_page
);
2022 isolated
= numamigrate_isolate_page(pgdat
, page
);
2028 /* Prepare a page as a migration target */
2029 __SetPageLocked(new_page
);
2030 if (PageSwapBacked(page
))
2031 __SetPageSwapBacked(new_page
);
2033 /* anon mapping, we can simply copy page->mapping to the new page: */
2034 new_page
->mapping
= page
->mapping
;
2035 new_page
->index
= page
->index
;
2036 /* flush the cache before copying using the kernel virtual address */
2037 flush_cache_range(vma
, start
, start
+ HPAGE_PMD_SIZE
);
2038 migrate_page_copy(new_page
, page
);
2039 WARN_ON(PageLRU(new_page
));
2041 /* Recheck the target PMD */
2042 ptl
= pmd_lock(mm
, pmd
);
2043 if (unlikely(!pmd_same(*pmd
, entry
) || !page_ref_freeze(page
, 2))) {
2046 /* Reverse changes made by migrate_page_copy() */
2047 if (TestClearPageActive(new_page
))
2048 SetPageActive(page
);
2049 if (TestClearPageUnevictable(new_page
))
2050 SetPageUnevictable(page
);
2052 unlock_page(new_page
);
2053 put_page(new_page
); /* Free it */
2055 /* Retake the callers reference and putback on LRU */
2057 putback_lru_page(page
);
2058 mod_node_page_state(page_pgdat(page
),
2059 NR_ISOLATED_ANON
+ page_lru
, -HPAGE_PMD_NR
);
2064 entry
= mk_huge_pmd(new_page
, vma
->vm_page_prot
);
2065 entry
= maybe_pmd_mkwrite(pmd_mkdirty(entry
), vma
);
2068 * Overwrite the old entry under pagetable lock and establish
2069 * the new PTE. Any parallel GUP will either observe the old
2070 * page blocking on the page lock, block on the page table
2071 * lock or observe the new page. The SetPageUptodate on the
2072 * new page and page_add_new_anon_rmap guarantee the copy is
2073 * visible before the pagetable update.
2075 page_add_anon_rmap(new_page
, vma
, start
, true);
2077 * At this point the pmd is numa/protnone (i.e. non present) and the TLB
2078 * has already been flushed globally. So no TLB can be currently
2079 * caching this non present pmd mapping. There's no need to clear the
2080 * pmd before doing set_pmd_at(), nor to flush the TLB after
2081 * set_pmd_at(). Clearing the pmd here would introduce a race
2082 * condition against MADV_DONTNEED, because MADV_DONTNEED only holds the
2083 * mmap_sem for reading. If the pmd is set to NULL at any given time,
2084 * MADV_DONTNEED won't wait on the pmd lock and it'll skip clearing this
2087 set_pmd_at(mm
, start
, pmd
, entry
);
2088 update_mmu_cache_pmd(vma
, address
, &entry
);
2090 page_ref_unfreeze(page
, 2);
2091 mlock_migrate_page(new_page
, page
);
2092 page_remove_rmap(page
, true);
2093 set_page_owner_migrate_reason(new_page
, MR_NUMA_MISPLACED
);
2097 /* Take an "isolate" reference and put new page on the LRU. */
2099 putback_lru_page(new_page
);
2101 unlock_page(new_page
);
2103 put_page(page
); /* Drop the rmap reference */
2104 put_page(page
); /* Drop the LRU isolation reference */
2106 count_vm_events(PGMIGRATE_SUCCESS
, HPAGE_PMD_NR
);
2107 count_vm_numa_events(NUMA_PAGE_MIGRATE
, HPAGE_PMD_NR
);
2109 mod_node_page_state(page_pgdat(page
),
2110 NR_ISOLATED_ANON
+ page_lru
,
2115 count_vm_events(PGMIGRATE_FAIL
, HPAGE_PMD_NR
);
2116 ptl
= pmd_lock(mm
, pmd
);
2117 if (pmd_same(*pmd
, entry
)) {
2118 entry
= pmd_modify(entry
, vma
->vm_page_prot
);
2119 set_pmd_at(mm
, start
, pmd
, entry
);
2120 update_mmu_cache_pmd(vma
, address
, &entry
);
2129 #endif /* CONFIG_NUMA_BALANCING */
2131 #endif /* CONFIG_NUMA */
2133 #if defined(CONFIG_MIGRATE_VMA_HELPER)
2134 struct migrate_vma
{
2135 struct vm_area_struct
*vma
;
2138 unsigned long cpages
;
2139 unsigned long npages
;
2140 unsigned long start
;
2144 static int migrate_vma_collect_hole(unsigned long start
,
2146 struct mm_walk
*walk
)
2148 struct migrate_vma
*migrate
= walk
->private;
2151 for (addr
= start
& PAGE_MASK
; addr
< end
; addr
+= PAGE_SIZE
) {
2152 migrate
->src
[migrate
->npages
] = MIGRATE_PFN_MIGRATE
;
2153 migrate
->dst
[migrate
->npages
] = 0;
2161 static int migrate_vma_collect_skip(unsigned long start
,
2163 struct mm_walk
*walk
)
2165 struct migrate_vma
*migrate
= walk
->private;
2168 for (addr
= start
& PAGE_MASK
; addr
< end
; addr
+= PAGE_SIZE
) {
2169 migrate
->dst
[migrate
->npages
] = 0;
2170 migrate
->src
[migrate
->npages
++] = 0;
2176 static int migrate_vma_collect_pmd(pmd_t
*pmdp
,
2177 unsigned long start
,
2179 struct mm_walk
*walk
)
2181 struct migrate_vma
*migrate
= walk
->private;
2182 struct vm_area_struct
*vma
= walk
->vma
;
2183 struct mm_struct
*mm
= vma
->vm_mm
;
2184 unsigned long addr
= start
, unmapped
= 0;
2189 if (pmd_none(*pmdp
))
2190 return migrate_vma_collect_hole(start
, end
, walk
);
2192 if (pmd_trans_huge(*pmdp
)) {
2195 ptl
= pmd_lock(mm
, pmdp
);
2196 if (unlikely(!pmd_trans_huge(*pmdp
))) {
2201 page
= pmd_page(*pmdp
);
2202 if (is_huge_zero_page(page
)) {
2204 split_huge_pmd(vma
, pmdp
, addr
);
2205 if (pmd_trans_unstable(pmdp
))
2206 return migrate_vma_collect_skip(start
, end
,
2213 if (unlikely(!trylock_page(page
)))
2214 return migrate_vma_collect_skip(start
, end
,
2216 ret
= split_huge_page(page
);
2220 return migrate_vma_collect_skip(start
, end
,
2222 if (pmd_none(*pmdp
))
2223 return migrate_vma_collect_hole(start
, end
,
2228 if (unlikely(pmd_bad(*pmdp
)))
2229 return migrate_vma_collect_skip(start
, end
, walk
);
2231 ptep
= pte_offset_map_lock(mm
, pmdp
, addr
, &ptl
);
2232 arch_enter_lazy_mmu_mode();
2234 for (; addr
< end
; addr
+= PAGE_SIZE
, ptep
++) {
2235 unsigned long mpfn
, pfn
;
2243 if (pte_none(pte
)) {
2244 mpfn
= MIGRATE_PFN_MIGRATE
;
2250 if (!pte_present(pte
)) {
2254 * Only care about unaddressable device page special
2255 * page table entry. Other special swap entries are not
2256 * migratable, and we ignore regular swapped page.
2258 entry
= pte_to_swp_entry(pte
);
2259 if (!is_device_private_entry(entry
))
2262 page
= device_private_entry_to_page(entry
);
2263 mpfn
= migrate_pfn(page_to_pfn(page
))|
2264 MIGRATE_PFN_DEVICE
| MIGRATE_PFN_MIGRATE
;
2265 if (is_write_device_private_entry(entry
))
2266 mpfn
|= MIGRATE_PFN_WRITE
;
2268 if (is_zero_pfn(pfn
)) {
2269 mpfn
= MIGRATE_PFN_MIGRATE
;
2274 page
= vm_normal_page(migrate
->vma
, addr
, pte
);
2275 mpfn
= migrate_pfn(pfn
) | MIGRATE_PFN_MIGRATE
;
2276 mpfn
|= pte_write(pte
) ? MIGRATE_PFN_WRITE
: 0;
2279 /* FIXME support THP */
2280 if (!page
|| !page
->mapping
|| PageTransCompound(page
)) {
2284 pfn
= page_to_pfn(page
);
2287 * By getting a reference on the page we pin it and that blocks
2288 * any kind of migration. Side effect is that it "freezes" the
2291 * We drop this reference after isolating the page from the lru
2292 * for non device page (device page are not on the lru and thus
2293 * can't be dropped from it).
2299 * Optimize for the common case where page is only mapped once
2300 * in one process. If we can lock the page, then we can safely
2301 * set up a special migration page table entry now.
2303 if (trylock_page(page
)) {
2306 mpfn
|= MIGRATE_PFN_LOCKED
;
2307 ptep_get_and_clear(mm
, addr
, ptep
);
2309 /* Setup special migration page table entry */
2310 entry
= make_migration_entry(page
, mpfn
&
2312 swp_pte
= swp_entry_to_pte(entry
);
2313 if (pte_soft_dirty(pte
))
2314 swp_pte
= pte_swp_mksoft_dirty(swp_pte
);
2315 set_pte_at(mm
, addr
, ptep
, swp_pte
);
2318 * This is like regular unmap: we remove the rmap and
2319 * drop page refcount. Page won't be freed, as we took
2320 * a reference just above.
2322 page_remove_rmap(page
, false);
2325 if (pte_present(pte
))
2330 migrate
->dst
[migrate
->npages
] = 0;
2331 migrate
->src
[migrate
->npages
++] = mpfn
;
2333 arch_leave_lazy_mmu_mode();
2334 pte_unmap_unlock(ptep
- 1, ptl
);
2336 /* Only flush the TLB if we actually modified any entries */
2338 flush_tlb_range(walk
->vma
, start
, end
);
2344 * migrate_vma_collect() - collect pages over a range of virtual addresses
2345 * @migrate: migrate struct containing all migration information
2347 * This will walk the CPU page table. For each virtual address backed by a
2348 * valid page, it updates the src array and takes a reference on the page, in
2349 * order to pin the page until we lock it and unmap it.
2351 static void migrate_vma_collect(struct migrate_vma
*migrate
)
2353 struct mmu_notifier_range range
;
2354 struct mm_walk mm_walk
= {
2355 .pmd_entry
= migrate_vma_collect_pmd
,
2356 .pte_hole
= migrate_vma_collect_hole
,
2357 .vma
= migrate
->vma
,
2358 .mm
= migrate
->vma
->vm_mm
,
2362 mmu_notifier_range_init(&range
, MMU_NOTIFY_CLEAR
, 0, NULL
, mm_walk
.mm
,
2365 mmu_notifier_invalidate_range_start(&range
);
2366 walk_page_range(migrate
->start
, migrate
->end
, &mm_walk
);
2367 mmu_notifier_invalidate_range_end(&range
);
2369 migrate
->end
= migrate
->start
+ (migrate
->npages
<< PAGE_SHIFT
);
2373 * migrate_vma_check_page() - check if page is pinned or not
2374 * @page: struct page to check
2376 * Pinned pages cannot be migrated. This is the same test as in
2377 * migrate_page_move_mapping(), except that here we allow migration of a
2380 static bool migrate_vma_check_page(struct page
*page
)
2383 * One extra ref because caller holds an extra reference, either from
2384 * isolate_lru_page() for a regular page, or migrate_vma_collect() for
2390 * FIXME support THP (transparent huge page), it is bit more complex to
2391 * check them than regular pages, because they can be mapped with a pmd
2392 * or with a pte (split pte mapping).
2394 if (PageCompound(page
))
2397 /* Page from ZONE_DEVICE have one extra reference */
2398 if (is_zone_device_page(page
)) {
2400 * Private page can never be pin as they have no valid pte and
2401 * GUP will fail for those. Yet if there is a pending migration
2402 * a thread might try to wait on the pte migration entry and
2403 * will bump the page reference count. Sadly there is no way to
2404 * differentiate a regular pin from migration wait. Hence to
2405 * avoid 2 racing thread trying to migrate back to CPU to enter
2406 * infinite loop (one stoping migration because the other is
2407 * waiting on pte migration entry). We always return true here.
2409 * FIXME proper solution is to rework migration_entry_wait() so
2410 * it does not need to take a reference on page.
2412 return is_device_private_page(page
);
2415 /* For file back page */
2416 if (page_mapping(page
))
2417 extra
+= 1 + page_has_private(page
);
2419 if ((page_count(page
) - extra
) > page_mapcount(page
))
2426 * migrate_vma_prepare() - lock pages and isolate them from the lru
2427 * @migrate: migrate struct containing all migration information
2429 * This locks pages that have been collected by migrate_vma_collect(). Once each
2430 * page is locked it is isolated from the lru (for non-device pages). Finally,
2431 * the ref taken by migrate_vma_collect() is dropped, as locked pages cannot be
2432 * migrated by concurrent kernel threads.
2434 static void migrate_vma_prepare(struct migrate_vma
*migrate
)
2436 const unsigned long npages
= migrate
->npages
;
2437 const unsigned long start
= migrate
->start
;
2438 unsigned long addr
, i
, restore
= 0;
2439 bool allow_drain
= true;
2443 for (i
= 0; (i
< npages
) && migrate
->cpages
; i
++) {
2444 struct page
*page
= migrate_pfn_to_page(migrate
->src
[i
]);
2450 if (!(migrate
->src
[i
] & MIGRATE_PFN_LOCKED
)) {
2452 * Because we are migrating several pages there can be
2453 * a deadlock between 2 concurrent migration where each
2454 * are waiting on each other page lock.
2456 * Make migrate_vma() a best effort thing and backoff
2457 * for any page we can not lock right away.
2459 if (!trylock_page(page
)) {
2460 migrate
->src
[i
] = 0;
2466 migrate
->src
[i
] |= MIGRATE_PFN_LOCKED
;
2469 /* ZONE_DEVICE pages are not on LRU */
2470 if (!is_zone_device_page(page
)) {
2471 if (!PageLRU(page
) && allow_drain
) {
2472 /* Drain CPU's pagevec */
2473 lru_add_drain_all();
2474 allow_drain
= false;
2477 if (isolate_lru_page(page
)) {
2479 migrate
->src
[i
] &= ~MIGRATE_PFN_MIGRATE
;
2483 migrate
->src
[i
] = 0;
2491 /* Drop the reference we took in collect */
2495 if (!migrate_vma_check_page(page
)) {
2497 migrate
->src
[i
] &= ~MIGRATE_PFN_MIGRATE
;
2501 if (!is_zone_device_page(page
)) {
2503 putback_lru_page(page
);
2506 migrate
->src
[i
] = 0;
2510 if (!is_zone_device_page(page
))
2511 putback_lru_page(page
);
2518 for (i
= 0, addr
= start
; i
< npages
&& restore
; i
++, addr
+= PAGE_SIZE
) {
2519 struct page
*page
= migrate_pfn_to_page(migrate
->src
[i
]);
2521 if (!page
|| (migrate
->src
[i
] & MIGRATE_PFN_MIGRATE
))
2524 remove_migration_pte(page
, migrate
->vma
, addr
, page
);
2526 migrate
->src
[i
] = 0;
2534 * migrate_vma_unmap() - replace page mapping with special migration pte entry
2535 * @migrate: migrate struct containing all migration information
2537 * Replace page mapping (CPU page table pte) with a special migration pte entry
2538 * and check again if it has been pinned. Pinned pages are restored because we
2539 * cannot migrate them.
2541 * This is the last step before we call the device driver callback to allocate
2542 * destination memory and copy contents of original page over to new page.
2544 static void migrate_vma_unmap(struct migrate_vma
*migrate
)
2546 int flags
= TTU_MIGRATION
| TTU_IGNORE_MLOCK
| TTU_IGNORE_ACCESS
;
2547 const unsigned long npages
= migrate
->npages
;
2548 const unsigned long start
= migrate
->start
;
2549 unsigned long addr
, i
, restore
= 0;
2551 for (i
= 0; i
< npages
; i
++) {
2552 struct page
*page
= migrate_pfn_to_page(migrate
->src
[i
]);
2554 if (!page
|| !(migrate
->src
[i
] & MIGRATE_PFN_MIGRATE
))
2557 if (page_mapped(page
)) {
2558 try_to_unmap(page
, flags
);
2559 if (page_mapped(page
))
2563 if (migrate_vma_check_page(page
))
2567 migrate
->src
[i
] &= ~MIGRATE_PFN_MIGRATE
;
2572 for (addr
= start
, i
= 0; i
< npages
&& restore
; addr
+= PAGE_SIZE
, i
++) {
2573 struct page
*page
= migrate_pfn_to_page(migrate
->src
[i
]);
2575 if (!page
|| (migrate
->src
[i
] & MIGRATE_PFN_MIGRATE
))
2578 remove_migration_ptes(page
, page
, false);
2580 migrate
->src
[i
] = 0;
2584 if (is_zone_device_page(page
))
2587 putback_lru_page(page
);
2591 static void migrate_vma_insert_page(struct migrate_vma
*migrate
,
2597 struct vm_area_struct
*vma
= migrate
->vma
;
2598 struct mm_struct
*mm
= vma
->vm_mm
;
2599 struct mem_cgroup
*memcg
;
2609 /* Only allow populating anonymous memory */
2610 if (!vma_is_anonymous(vma
))
2613 pgdp
= pgd_offset(mm
, addr
);
2614 p4dp
= p4d_alloc(mm
, pgdp
, addr
);
2617 pudp
= pud_alloc(mm
, p4dp
, addr
);
2620 pmdp
= pmd_alloc(mm
, pudp
, addr
);
2624 if (pmd_trans_huge(*pmdp
) || pmd_devmap(*pmdp
))
2628 * Use pte_alloc() instead of pte_alloc_map(). We can't run
2629 * pte_offset_map() on pmds where a huge pmd might be created
2630 * from a different thread.
2632 * pte_alloc_map() is safe to use under down_write(mmap_sem) or when
2633 * parallel threads are excluded by other means.
2635 * Here we only have down_read(mmap_sem).
2637 if (pte_alloc(mm
, pmdp
))
2640 /* See the comment in pte_alloc_one_map() */
2641 if (unlikely(pmd_trans_unstable(pmdp
)))
2644 if (unlikely(anon_vma_prepare(vma
)))
2646 if (mem_cgroup_try_charge(page
, vma
->vm_mm
, GFP_KERNEL
, &memcg
, false))
2650 * The memory barrier inside __SetPageUptodate makes sure that
2651 * preceding stores to the page contents become visible before
2652 * the set_pte_at() write.
2654 __SetPageUptodate(page
);
2656 if (is_zone_device_page(page
)) {
2657 if (is_device_private_page(page
)) {
2658 swp_entry_t swp_entry
;
2660 swp_entry
= make_device_private_entry(page
, vma
->vm_flags
& VM_WRITE
);
2661 entry
= swp_entry_to_pte(swp_entry
);
2664 entry
= mk_pte(page
, vma
->vm_page_prot
);
2665 if (vma
->vm_flags
& VM_WRITE
)
2666 entry
= pte_mkwrite(pte_mkdirty(entry
));
2669 ptep
= pte_offset_map_lock(mm
, pmdp
, addr
, &ptl
);
2671 if (pte_present(*ptep
)) {
2672 unsigned long pfn
= pte_pfn(*ptep
);
2674 if (!is_zero_pfn(pfn
)) {
2675 pte_unmap_unlock(ptep
, ptl
);
2676 mem_cgroup_cancel_charge(page
, memcg
, false);
2680 } else if (!pte_none(*ptep
)) {
2681 pte_unmap_unlock(ptep
, ptl
);
2682 mem_cgroup_cancel_charge(page
, memcg
, false);
2687 * Check for usefaultfd but do not deliver the fault. Instead,
2690 if (userfaultfd_missing(vma
)) {
2691 pte_unmap_unlock(ptep
, ptl
);
2692 mem_cgroup_cancel_charge(page
, memcg
, false);
2696 inc_mm_counter(mm
, MM_ANONPAGES
);
2697 page_add_new_anon_rmap(page
, vma
, addr
, false);
2698 mem_cgroup_commit_charge(page
, memcg
, false, false);
2699 if (!is_zone_device_page(page
))
2700 lru_cache_add_active_or_unevictable(page
, vma
);
2704 flush_cache_page(vma
, addr
, pte_pfn(*ptep
));
2705 ptep_clear_flush_notify(vma
, addr
, ptep
);
2706 set_pte_at_notify(mm
, addr
, ptep
, entry
);
2707 update_mmu_cache(vma
, addr
, ptep
);
2709 /* No need to invalidate - it was non-present before */
2710 set_pte_at(mm
, addr
, ptep
, entry
);
2711 update_mmu_cache(vma
, addr
, ptep
);
2714 pte_unmap_unlock(ptep
, ptl
);
2715 *src
= MIGRATE_PFN_MIGRATE
;
2719 *src
&= ~MIGRATE_PFN_MIGRATE
;
2723 * migrate_vma_pages() - migrate meta-data from src page to dst page
2724 * @migrate: migrate struct containing all migration information
2726 * This migrates struct page meta-data from source struct page to destination
2727 * struct page. This effectively finishes the migration from source page to the
2730 static void migrate_vma_pages(struct migrate_vma
*migrate
)
2732 const unsigned long npages
= migrate
->npages
;
2733 const unsigned long start
= migrate
->start
;
2734 struct mmu_notifier_range range
;
2735 unsigned long addr
, i
;
2736 bool notified
= false;
2738 for (i
= 0, addr
= start
; i
< npages
; addr
+= PAGE_SIZE
, i
++) {
2739 struct page
*newpage
= migrate_pfn_to_page(migrate
->dst
[i
]);
2740 struct page
*page
= migrate_pfn_to_page(migrate
->src
[i
]);
2741 struct address_space
*mapping
;
2745 migrate
->src
[i
] &= ~MIGRATE_PFN_MIGRATE
;
2750 if (!(migrate
->src
[i
] & MIGRATE_PFN_MIGRATE
)) {
2756 mmu_notifier_range_init(&range
,
2757 MMU_NOTIFY_CLEAR
, 0,
2759 migrate
->vma
->vm_mm
,
2760 addr
, migrate
->end
);
2761 mmu_notifier_invalidate_range_start(&range
);
2763 migrate_vma_insert_page(migrate
, addr
, newpage
,
2769 mapping
= page_mapping(page
);
2771 if (is_zone_device_page(newpage
)) {
2772 if (is_device_private_page(newpage
)) {
2774 * For now only support private anonymous when
2775 * migrating to un-addressable device memory.
2778 migrate
->src
[i
] &= ~MIGRATE_PFN_MIGRATE
;
2783 * Other types of ZONE_DEVICE page are not
2786 migrate
->src
[i
] &= ~MIGRATE_PFN_MIGRATE
;
2791 r
= migrate_page(mapping
, newpage
, page
, MIGRATE_SYNC_NO_COPY
);
2792 if (r
!= MIGRATEPAGE_SUCCESS
)
2793 migrate
->src
[i
] &= ~MIGRATE_PFN_MIGRATE
;
2797 * No need to double call mmu_notifier->invalidate_range() callback as
2798 * the above ptep_clear_flush_notify() inside migrate_vma_insert_page()
2799 * did already call it.
2802 mmu_notifier_invalidate_range_only_end(&range
);
2806 * migrate_vma_finalize() - restore CPU page table entry
2807 * @migrate: migrate struct containing all migration information
2809 * This replaces the special migration pte entry with either a mapping to the
2810 * new page if migration was successful for that page, or to the original page
2813 * This also unlocks the pages and puts them back on the lru, or drops the extra
2814 * refcount, for device pages.
2816 static void migrate_vma_finalize(struct migrate_vma
*migrate
)
2818 const unsigned long npages
= migrate
->npages
;
2821 for (i
= 0; i
< npages
; i
++) {
2822 struct page
*newpage
= migrate_pfn_to_page(migrate
->dst
[i
]);
2823 struct page
*page
= migrate_pfn_to_page(migrate
->src
[i
]);
2827 unlock_page(newpage
);
2833 if (!(migrate
->src
[i
] & MIGRATE_PFN_MIGRATE
) || !newpage
) {
2835 unlock_page(newpage
);
2841 remove_migration_ptes(page
, newpage
, false);
2845 if (is_zone_device_page(page
))
2848 putback_lru_page(page
);
2850 if (newpage
!= page
) {
2851 unlock_page(newpage
);
2852 if (is_zone_device_page(newpage
))
2855 putback_lru_page(newpage
);
2861 * migrate_vma() - migrate a range of memory inside vma
2863 * @ops: migration callback for allocating destination memory and copying
2864 * @vma: virtual memory area containing the range to be migrated
2865 * @start: start address of the range to migrate (inclusive)
2866 * @end: end address of the range to migrate (exclusive)
2867 * @src: array of hmm_pfn_t containing source pfns
2868 * @dst: array of hmm_pfn_t containing destination pfns
2869 * @private: pointer passed back to each of the callback
2870 * Returns: 0 on success, error code otherwise
2872 * This function tries to migrate a range of memory virtual address range, using
2873 * callbacks to allocate and copy memory from source to destination. First it
2874 * collects all the pages backing each virtual address in the range, saving this
2875 * inside the src array. Then it locks those pages and unmaps them. Once the pages
2876 * are locked and unmapped, it checks whether each page is pinned or not. Pages
2877 * that aren't pinned have the MIGRATE_PFN_MIGRATE flag set (by this function)
2878 * in the corresponding src array entry. It then restores any pages that are
2879 * pinned, by remapping and unlocking those pages.
2881 * At this point it calls the alloc_and_copy() callback. For documentation on
2882 * what is expected from that callback, see struct migrate_vma_ops comments in
2883 * include/linux/migrate.h
2885 * After the alloc_and_copy() callback, this function goes over each entry in
2886 * the src array that has the MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE flag
2887 * set. If the corresponding entry in dst array has MIGRATE_PFN_VALID flag set,
2888 * then the function tries to migrate struct page information from the source
2889 * struct page to the destination struct page. If it fails to migrate the struct
2890 * page information, then it clears the MIGRATE_PFN_MIGRATE flag in the src
2893 * At this point all successfully migrated pages have an entry in the src
2894 * array with MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE flag set and the dst
2895 * array entry with MIGRATE_PFN_VALID flag set.
2897 * It then calls the finalize_and_map() callback. See comments for "struct
2898 * migrate_vma_ops", in include/linux/migrate.h for details about
2899 * finalize_and_map() behavior.
2901 * After the finalize_and_map() callback, for successfully migrated pages, this
2902 * function updates the CPU page table to point to new pages, otherwise it
2903 * restores the CPU page table to point to the original source pages.
2905 * Function returns 0 after the above steps, even if no pages were migrated
2906 * (The function only returns an error if any of the arguments are invalid.)
2908 * Both src and dst array must be big enough for (end - start) >> PAGE_SHIFT
2909 * unsigned long entries.
2911 int migrate_vma(const struct migrate_vma_ops
*ops
,
2912 struct vm_area_struct
*vma
,
2913 unsigned long start
,
2919 struct migrate_vma migrate
;
2921 /* Sanity check the arguments */
2924 if (!vma
|| is_vm_hugetlb_page(vma
) || (vma
->vm_flags
& VM_SPECIAL
) ||
2927 if (start
< vma
->vm_start
|| start
>= vma
->vm_end
)
2929 if (end
<= vma
->vm_start
|| end
> vma
->vm_end
)
2931 if (!ops
|| !src
|| !dst
|| start
>= end
)
2934 memset(src
, 0, sizeof(*src
) * ((end
- start
) >> PAGE_SHIFT
));
2937 migrate
.start
= start
;
2943 /* Collect, and try to unmap source pages */
2944 migrate_vma_collect(&migrate
);
2945 if (!migrate
.cpages
)
2948 /* Lock and isolate page */
2949 migrate_vma_prepare(&migrate
);
2950 if (!migrate
.cpages
)
2954 migrate_vma_unmap(&migrate
);
2955 if (!migrate
.cpages
)
2959 * At this point pages are locked and unmapped, and thus they have
2960 * stable content and can safely be copied to destination memory that
2961 * is allocated by the callback.
2963 * Note that migration can fail in migrate_vma_struct_page() for each
2966 ops
->alloc_and_copy(vma
, src
, dst
, start
, end
, private);
2968 /* This does the real migration of struct page */
2969 migrate_vma_pages(&migrate
);
2971 ops
->finalize_and_map(vma
, src
, dst
, start
, end
, private);
2973 /* Unlock and remap pages */
2974 migrate_vma_finalize(&migrate
);
2978 EXPORT_SYMBOL(migrate_vma
);
2979 #endif /* defined(MIGRATE_VMA_HELPER) */