2 * Memory Migration functionality - linux/mm/migration.c
4 * Copyright (C) 2006 Silicon Graphics, Inc., Christoph Lameter
6 * Page migration was first developed in the context of the memory hotplug
7 * project. The main authors of the migration code are:
9 * IWAMOTO Toshihiro <iwamoto@valinux.co.jp>
10 * Hirokazu Takahashi <taka@valinux.co.jp>
11 * Dave Hansen <haveblue@us.ibm.com>
15 #include <linux/migrate.h>
16 #include <linux/export.h>
17 #include <linux/swap.h>
18 #include <linux/swapops.h>
19 #include <linux/pagemap.h>
20 #include <linux/buffer_head.h>
21 #include <linux/mm_inline.h>
22 #include <linux/nsproxy.h>
23 #include <linux/pagevec.h>
24 #include <linux/ksm.h>
25 #include <linux/rmap.h>
26 #include <linux/topology.h>
27 #include <linux/cpu.h>
28 #include <linux/cpuset.h>
29 #include <linux/writeback.h>
30 #include <linux/mempolicy.h>
31 #include <linux/vmalloc.h>
32 #include <linux/security.h>
33 #include <linux/memcontrol.h>
34 #include <linux/syscalls.h>
35 #include <linux/hugetlb.h>
36 #include <linux/hugetlb_cgroup.h>
37 #include <linux/gfp.h>
38 #include <linux/balloon_compaction.h>
40 #include <asm/tlbflush.h>
42 #define CREATE_TRACE_POINTS
43 #include <trace/events/migrate.h>
48 * migrate_prep() needs to be called before we start compiling a list of pages
49 * to be migrated using isolate_lru_page(). If scheduling work on other CPUs is
50 * undesirable, use migrate_prep_local()
52 int migrate_prep(void)
55 * Clear the LRU lists so pages can be isolated.
56 * Note that pages may be moved off the LRU after we have
57 * drained them. Those pages will fail to migrate like other
58 * pages that may be busy.
65 /* Do the necessary work of migrate_prep but not if it involves other CPUs */
66 int migrate_prep_local(void)
74 * Add isolated pages on the list back to the LRU under page lock
75 * to avoid leaking evictable pages back onto unevictable list.
77 void putback_lru_pages(struct list_head
*l
)
82 list_for_each_entry_safe(page
, page2
, l
, lru
) {
84 dec_zone_page_state(page
, NR_ISOLATED_ANON
+
85 page_is_file_cache(page
));
86 putback_lru_page(page
);
91 * Put previously isolated pages back onto the appropriate lists
92 * from where they were once taken off for compaction/migration.
94 * This function shall be used instead of putback_lru_pages(),
95 * whenever the isolated pageset has been built by isolate_migratepages_range()
97 void putback_movable_pages(struct list_head
*l
)
102 list_for_each_entry_safe(page
, page2
, l
, lru
) {
103 list_del(&page
->lru
);
104 dec_zone_page_state(page
, NR_ISOLATED_ANON
+
105 page_is_file_cache(page
));
106 if (unlikely(balloon_page_movable(page
)))
107 balloon_page_putback(page
);
109 putback_lru_page(page
);
114 * Restore a potential migration pte to a working pte entry
116 static int remove_migration_pte(struct page
*new, struct vm_area_struct
*vma
,
117 unsigned long addr
, void *old
)
119 struct mm_struct
*mm
= vma
->vm_mm
;
125 if (unlikely(PageHuge(new))) {
126 ptep
= huge_pte_offset(mm
, addr
);
129 ptl
= &mm
->page_table_lock
;
131 pmd
= mm_find_pmd(mm
, addr
);
134 if (pmd_trans_huge(*pmd
))
137 ptep
= pte_offset_map(pmd
, addr
);
140 * Peek to check is_swap_pte() before taking ptlock? No, we
141 * can race mremap's move_ptes(), which skips anon_vma lock.
144 ptl
= pte_lockptr(mm
, pmd
);
149 if (!is_swap_pte(pte
))
152 entry
= pte_to_swp_entry(pte
);
154 if (!is_migration_entry(entry
) ||
155 migration_entry_to_page(entry
) != old
)
159 pte
= pte_mkold(mk_pte(new, vma
->vm_page_prot
));
160 if (is_write_migration_entry(entry
))
161 pte
= pte_mkwrite(pte
);
162 #ifdef CONFIG_HUGETLB_PAGE
164 pte
= pte_mkhuge(pte
);
165 pte
= arch_make_huge_pte(pte
, vma
, new, 0);
168 flush_cache_page(vma
, addr
, pte_pfn(pte
));
169 set_pte_at(mm
, addr
, ptep
, pte
);
173 hugepage_add_anon_rmap(new, vma
, addr
);
176 } else if (PageAnon(new))
177 page_add_anon_rmap(new, vma
, addr
);
179 page_add_file_rmap(new);
181 /* No need to invalidate - it was non-present before */
182 update_mmu_cache(vma
, addr
, ptep
);
184 pte_unmap_unlock(ptep
, ptl
);
190 * Get rid of all migration entries and replace them by
191 * references to the indicated page.
193 static void remove_migration_ptes(struct page
*old
, struct page
*new)
195 rmap_walk(new, remove_migration_pte
, old
);
199 * Something used the pte of a page under migration. We need to
200 * get to the page and wait until migration is finished.
201 * When we return from this function the fault will be retried.
203 void migration_entry_wait(struct mm_struct
*mm
, pmd_t
*pmd
,
204 unsigned long address
)
211 ptep
= pte_offset_map_lock(mm
, pmd
, address
, &ptl
);
213 if (!is_swap_pte(pte
))
216 entry
= pte_to_swp_entry(pte
);
217 if (!is_migration_entry(entry
))
220 page
= migration_entry_to_page(entry
);
223 * Once radix-tree replacement of page migration started, page_count
224 * *must* be zero. And, we don't want to call wait_on_page_locked()
225 * against a page without get_page().
226 * So, we use get_page_unless_zero(), here. Even failed, page fault
229 if (!get_page_unless_zero(page
))
231 pte_unmap_unlock(ptep
, ptl
);
232 wait_on_page_locked(page
);
236 pte_unmap_unlock(ptep
, ptl
);
240 /* Returns true if all buffers are successfully locked */
241 static bool buffer_migrate_lock_buffers(struct buffer_head
*head
,
242 enum migrate_mode mode
)
244 struct buffer_head
*bh
= head
;
246 /* Simple case, sync compaction */
247 if (mode
!= MIGRATE_ASYNC
) {
251 bh
= bh
->b_this_page
;
253 } while (bh
!= head
);
258 /* async case, we cannot block on lock_buffer so use trylock_buffer */
261 if (!trylock_buffer(bh
)) {
263 * We failed to lock the buffer and cannot stall in
264 * async migration. Release the taken locks
266 struct buffer_head
*failed_bh
= bh
;
269 while (bh
!= failed_bh
) {
272 bh
= bh
->b_this_page
;
277 bh
= bh
->b_this_page
;
278 } while (bh
!= head
);
282 static inline bool buffer_migrate_lock_buffers(struct buffer_head
*head
,
283 enum migrate_mode mode
)
287 #endif /* CONFIG_BLOCK */
290 * Replace the page in the mapping.
292 * The number of remaining references must be:
293 * 1 for anonymous pages without a mapping
294 * 2 for pages with a mapping
295 * 3 for pages with a mapping and PagePrivate/PagePrivate2 set.
297 static int migrate_page_move_mapping(struct address_space
*mapping
,
298 struct page
*newpage
, struct page
*page
,
299 struct buffer_head
*head
, enum migrate_mode mode
)
301 int expected_count
= 0;
305 /* Anonymous page without mapping */
306 if (page_count(page
) != 1)
308 return MIGRATEPAGE_SUCCESS
;
311 spin_lock_irq(&mapping
->tree_lock
);
313 pslot
= radix_tree_lookup_slot(&mapping
->page_tree
,
316 expected_count
= 2 + page_has_private(page
);
317 if (page_count(page
) != expected_count
||
318 radix_tree_deref_slot_protected(pslot
, &mapping
->tree_lock
) != page
) {
319 spin_unlock_irq(&mapping
->tree_lock
);
323 if (!page_freeze_refs(page
, expected_count
)) {
324 spin_unlock_irq(&mapping
->tree_lock
);
329 * In the async migration case of moving a page with buffers, lock the
330 * buffers using trylock before the mapping is moved. If the mapping
331 * was moved, we later failed to lock the buffers and could not move
332 * the mapping back due to an elevated page count, we would have to
333 * block waiting on other references to be dropped.
335 if (mode
== MIGRATE_ASYNC
&& head
&&
336 !buffer_migrate_lock_buffers(head
, mode
)) {
337 page_unfreeze_refs(page
, expected_count
);
338 spin_unlock_irq(&mapping
->tree_lock
);
343 * Now we know that no one else is looking at the page.
345 get_page(newpage
); /* add cache reference */
346 if (PageSwapCache(page
)) {
347 SetPageSwapCache(newpage
);
348 set_page_private(newpage
, page_private(page
));
351 radix_tree_replace_slot(pslot
, newpage
);
354 * Drop cache reference from old page by unfreezing
355 * to one less reference.
356 * We know this isn't the last reference.
358 page_unfreeze_refs(page
, expected_count
- 1);
361 * If moved to a different zone then also account
362 * the page for that zone. Other VM counters will be
363 * taken care of when we establish references to the
364 * new page and drop references to the old page.
366 * Note that anonymous pages are accounted for
367 * via NR_FILE_PAGES and NR_ANON_PAGES if they
368 * are mapped to swap space.
370 __dec_zone_page_state(page
, NR_FILE_PAGES
);
371 __inc_zone_page_state(newpage
, NR_FILE_PAGES
);
372 if (!PageSwapCache(page
) && PageSwapBacked(page
)) {
373 __dec_zone_page_state(page
, NR_SHMEM
);
374 __inc_zone_page_state(newpage
, NR_SHMEM
);
376 spin_unlock_irq(&mapping
->tree_lock
);
378 return MIGRATEPAGE_SUCCESS
;
382 * The expected number of remaining references is the same as that
383 * of migrate_page_move_mapping().
385 int migrate_huge_page_move_mapping(struct address_space
*mapping
,
386 struct page
*newpage
, struct page
*page
)
392 if (page_count(page
) != 1)
394 return MIGRATEPAGE_SUCCESS
;
397 spin_lock_irq(&mapping
->tree_lock
);
399 pslot
= radix_tree_lookup_slot(&mapping
->page_tree
,
402 expected_count
= 2 + page_has_private(page
);
403 if (page_count(page
) != expected_count
||
404 radix_tree_deref_slot_protected(pslot
, &mapping
->tree_lock
) != page
) {
405 spin_unlock_irq(&mapping
->tree_lock
);
409 if (!page_freeze_refs(page
, expected_count
)) {
410 spin_unlock_irq(&mapping
->tree_lock
);
416 radix_tree_replace_slot(pslot
, newpage
);
418 page_unfreeze_refs(page
, expected_count
- 1);
420 spin_unlock_irq(&mapping
->tree_lock
);
421 return MIGRATEPAGE_SUCCESS
;
425 * Copy the page to its new location
427 void migrate_page_copy(struct page
*newpage
, struct page
*page
)
429 if (PageHuge(page
) || PageTransHuge(page
))
430 copy_huge_page(newpage
, page
);
432 copy_highpage(newpage
, page
);
435 SetPageError(newpage
);
436 if (PageReferenced(page
))
437 SetPageReferenced(newpage
);
438 if (PageUptodate(page
))
439 SetPageUptodate(newpage
);
440 if (TestClearPageActive(page
)) {
441 VM_BUG_ON(PageUnevictable(page
));
442 SetPageActive(newpage
);
443 } else if (TestClearPageUnevictable(page
))
444 SetPageUnevictable(newpage
);
445 if (PageChecked(page
))
446 SetPageChecked(newpage
);
447 if (PageMappedToDisk(page
))
448 SetPageMappedToDisk(newpage
);
450 if (PageDirty(page
)) {
451 clear_page_dirty_for_io(page
);
453 * Want to mark the page and the radix tree as dirty, and
454 * redo the accounting that clear_page_dirty_for_io undid,
455 * but we can't use set_page_dirty because that function
456 * is actually a signal that all of the page has become dirty.
457 * Whereas only part of our page may be dirty.
459 if (PageSwapBacked(page
))
460 SetPageDirty(newpage
);
462 __set_page_dirty_nobuffers(newpage
);
465 mlock_migrate_page(newpage
, page
);
466 ksm_migrate_page(newpage
, page
);
468 * Please do not reorder this without considering how mm/ksm.c's
469 * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache().
471 ClearPageSwapCache(page
);
472 ClearPagePrivate(page
);
473 set_page_private(page
, 0);
476 * If any waiters have accumulated on the new page then
479 if (PageWriteback(newpage
))
480 end_page_writeback(newpage
);
483 /************************************************************
484 * Migration functions
485 ***********************************************************/
487 /* Always fail migration. Used for mappings that are not movable */
488 int fail_migrate_page(struct address_space
*mapping
,
489 struct page
*newpage
, struct page
*page
)
493 EXPORT_SYMBOL(fail_migrate_page
);
496 * Common logic to directly migrate a single page suitable for
497 * pages that do not use PagePrivate/PagePrivate2.
499 * Pages are locked upon entry and exit.
501 int migrate_page(struct address_space
*mapping
,
502 struct page
*newpage
, struct page
*page
,
503 enum migrate_mode mode
)
507 BUG_ON(PageWriteback(page
)); /* Writeback must be complete */
509 rc
= migrate_page_move_mapping(mapping
, newpage
, page
, NULL
, mode
);
511 if (rc
!= MIGRATEPAGE_SUCCESS
)
514 migrate_page_copy(newpage
, page
);
515 return MIGRATEPAGE_SUCCESS
;
517 EXPORT_SYMBOL(migrate_page
);
521 * Migration function for pages with buffers. This function can only be used
522 * if the underlying filesystem guarantees that no other references to "page"
525 int buffer_migrate_page(struct address_space
*mapping
,
526 struct page
*newpage
, struct page
*page
, enum migrate_mode mode
)
528 struct buffer_head
*bh
, *head
;
531 if (!page_has_buffers(page
))
532 return migrate_page(mapping
, newpage
, page
, mode
);
534 head
= page_buffers(page
);
536 rc
= migrate_page_move_mapping(mapping
, newpage
, page
, head
, mode
);
538 if (rc
!= MIGRATEPAGE_SUCCESS
)
542 * In the async case, migrate_page_move_mapping locked the buffers
543 * with an IRQ-safe spinlock held. In the sync case, the buffers
544 * need to be locked now
546 if (mode
!= MIGRATE_ASYNC
)
547 BUG_ON(!buffer_migrate_lock_buffers(head
, mode
));
549 ClearPagePrivate(page
);
550 set_page_private(newpage
, page_private(page
));
551 set_page_private(page
, 0);
557 set_bh_page(bh
, newpage
, bh_offset(bh
));
558 bh
= bh
->b_this_page
;
560 } while (bh
!= head
);
562 SetPagePrivate(newpage
);
564 migrate_page_copy(newpage
, page
);
570 bh
= bh
->b_this_page
;
572 } while (bh
!= head
);
574 return MIGRATEPAGE_SUCCESS
;
576 EXPORT_SYMBOL(buffer_migrate_page
);
580 * Writeback a page to clean the dirty state
582 static int writeout(struct address_space
*mapping
, struct page
*page
)
584 struct writeback_control wbc
= {
585 .sync_mode
= WB_SYNC_NONE
,
588 .range_end
= LLONG_MAX
,
593 if (!mapping
->a_ops
->writepage
)
594 /* No write method for the address space */
597 if (!clear_page_dirty_for_io(page
))
598 /* Someone else already triggered a write */
602 * A dirty page may imply that the underlying filesystem has
603 * the page on some queue. So the page must be clean for
604 * migration. Writeout may mean we loose the lock and the
605 * page state is no longer what we checked for earlier.
606 * At this point we know that the migration attempt cannot
609 remove_migration_ptes(page
, page
);
611 rc
= mapping
->a_ops
->writepage(page
, &wbc
);
613 if (rc
!= AOP_WRITEPAGE_ACTIVATE
)
614 /* unlocked. Relock */
617 return (rc
< 0) ? -EIO
: -EAGAIN
;
621 * Default handling if a filesystem does not provide a migration function.
623 static int fallback_migrate_page(struct address_space
*mapping
,
624 struct page
*newpage
, struct page
*page
, enum migrate_mode mode
)
626 if (PageDirty(page
)) {
627 /* Only writeback pages in full synchronous migration */
628 if (mode
!= MIGRATE_SYNC
)
630 return writeout(mapping
, page
);
634 * Buffers may be managed in a filesystem specific way.
635 * We must have no buffers or drop them.
637 if (page_has_private(page
) &&
638 !try_to_release_page(page
, GFP_KERNEL
))
641 return migrate_page(mapping
, newpage
, page
, mode
);
645 * Move a page to a newly allocated page
646 * The page is locked and all ptes have been successfully removed.
648 * The new page will have replaced the old page if this function
653 * MIGRATEPAGE_SUCCESS - success
655 static int move_to_new_page(struct page
*newpage
, struct page
*page
,
656 int remap_swapcache
, enum migrate_mode mode
)
658 struct address_space
*mapping
;
662 * Block others from accessing the page when we get around to
663 * establishing additional references. We are the only one
664 * holding a reference to the new page at this point.
666 if (!trylock_page(newpage
))
669 /* Prepare mapping for the new page.*/
670 newpage
->index
= page
->index
;
671 newpage
->mapping
= page
->mapping
;
672 if (PageSwapBacked(page
))
673 SetPageSwapBacked(newpage
);
675 mapping
= page_mapping(page
);
677 rc
= migrate_page(mapping
, newpage
, page
, mode
);
678 else if (mapping
->a_ops
->migratepage
)
680 * Most pages have a mapping and most filesystems provide a
681 * migratepage callback. Anonymous pages are part of swap
682 * space which also has its own migratepage callback. This
683 * is the most common path for page migration.
685 rc
= mapping
->a_ops
->migratepage(mapping
,
686 newpage
, page
, mode
);
688 rc
= fallback_migrate_page(mapping
, newpage
, page
, mode
);
690 if (rc
!= MIGRATEPAGE_SUCCESS
) {
691 newpage
->mapping
= NULL
;
694 remove_migration_ptes(page
, newpage
);
695 page
->mapping
= NULL
;
698 unlock_page(newpage
);
703 static int __unmap_and_move(struct page
*page
, struct page
*newpage
,
704 int force
, enum migrate_mode mode
)
707 int remap_swapcache
= 1;
708 struct mem_cgroup
*mem
;
709 struct anon_vma
*anon_vma
= NULL
;
711 if (!trylock_page(page
)) {
712 if (!force
|| mode
== MIGRATE_ASYNC
)
716 * It's not safe for direct compaction to call lock_page.
717 * For example, during page readahead pages are added locked
718 * to the LRU. Later, when the IO completes the pages are
719 * marked uptodate and unlocked. However, the queueing
720 * could be merging multiple pages for one bio (e.g.
721 * mpage_readpages). If an allocation happens for the
722 * second or third page, the process can end up locking
723 * the same page twice and deadlocking. Rather than
724 * trying to be clever about what pages can be locked,
725 * avoid the use of lock_page for direct compaction
728 if (current
->flags
& PF_MEMALLOC
)
734 /* charge against new page */
735 mem_cgroup_prepare_migration(page
, newpage
, &mem
);
737 if (PageWriteback(page
)) {
739 * Only in the case of a full synchronous migration is it
740 * necessary to wait for PageWriteback. In the async case,
741 * the retry loop is too short and in the sync-light case,
742 * the overhead of stalling is too much
744 if (mode
!= MIGRATE_SYNC
) {
750 wait_on_page_writeback(page
);
753 * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
754 * we cannot notice that anon_vma is freed while we migrates a page.
755 * This get_anon_vma() delays freeing anon_vma pointer until the end
756 * of migration. File cache pages are no problem because of page_lock()
757 * File Caches may use write_page() or lock_page() in migration, then,
758 * just care Anon page here.
760 if (PageAnon(page
) && !PageKsm(page
)) {
762 * Only page_lock_anon_vma_read() understands the subtleties of
763 * getting a hold on an anon_vma from outside one of its mms.
765 anon_vma
= page_get_anon_vma(page
);
770 } else if (PageSwapCache(page
)) {
772 * We cannot be sure that the anon_vma of an unmapped
773 * swapcache page is safe to use because we don't
774 * know in advance if the VMA that this page belonged
775 * to still exists. If the VMA and others sharing the
776 * data have been freed, then the anon_vma could
777 * already be invalid.
779 * To avoid this possibility, swapcache pages get
780 * migrated but are not remapped when migration
789 if (unlikely(balloon_page_movable(page
))) {
791 * A ballooned page does not need any special attention from
792 * physical to virtual reverse mapping procedures.
793 * Skip any attempt to unmap PTEs or to remap swap cache,
794 * in order to avoid burning cycles at rmap level, and perform
795 * the page migration right away (proteced by page lock).
797 rc
= balloon_page_migrate(newpage
, page
, mode
);
802 * Corner case handling:
803 * 1. When a new swap-cache page is read into, it is added to the LRU
804 * and treated as swapcache but it has no rmap yet.
805 * Calling try_to_unmap() against a page->mapping==NULL page will
806 * trigger a BUG. So handle it here.
807 * 2. An orphaned page (see truncate_complete_page) might have
808 * fs-private metadata. The page can be picked up due to memory
809 * offlining. Everywhere else except page reclaim, the page is
810 * invisible to the vm, so the page can not be migrated. So try to
811 * free the metadata, so the page can be freed.
813 if (!page
->mapping
) {
814 VM_BUG_ON(PageAnon(page
));
815 if (page_has_private(page
)) {
816 try_to_free_buffers(page
);
822 /* Establish migration ptes or remove ptes */
823 try_to_unmap(page
, TTU_MIGRATION
|TTU_IGNORE_MLOCK
|TTU_IGNORE_ACCESS
);
826 if (!page_mapped(page
))
827 rc
= move_to_new_page(newpage
, page
, remap_swapcache
, mode
);
829 if (rc
&& remap_swapcache
)
830 remove_migration_ptes(page
, page
);
832 /* Drop an anon_vma reference if we took one */
834 put_anon_vma(anon_vma
);
837 mem_cgroup_end_migration(mem
, page
, newpage
,
838 (rc
== MIGRATEPAGE_SUCCESS
||
839 rc
== MIGRATEPAGE_BALLOON_SUCCESS
));
846 * Obtain the lock on page, remove all ptes and migrate the page
847 * to the newly allocated page in newpage.
849 static int unmap_and_move(new_page_t get_new_page
, unsigned long private,
850 struct page
*page
, int force
, enum migrate_mode mode
)
854 struct page
*newpage
= get_new_page(page
, private, &result
);
859 if (page_count(page
) == 1) {
860 /* page was freed from under us. So we are done. */
864 if (unlikely(PageTransHuge(page
)))
865 if (unlikely(split_huge_page(page
)))
868 rc
= __unmap_and_move(page
, newpage
, force
, mode
);
870 if (unlikely(rc
== MIGRATEPAGE_BALLOON_SUCCESS
)) {
872 * A ballooned page has been migrated already.
873 * Now, it's the time to wrap-up counters,
874 * handle the page back to Buddy and return.
876 dec_zone_page_state(page
, NR_ISOLATED_ANON
+
877 page_is_file_cache(page
));
878 balloon_page_free(page
);
879 return MIGRATEPAGE_SUCCESS
;
884 * A page that has been migrated has all references
885 * removed and will be freed. A page that has not been
886 * migrated will have kepts its references and be
889 list_del(&page
->lru
);
890 dec_zone_page_state(page
, NR_ISOLATED_ANON
+
891 page_is_file_cache(page
));
892 putback_lru_page(page
);
895 * Move the new page to the LRU. If migration was not successful
896 * then this will free the page.
898 putback_lru_page(newpage
);
903 *result
= page_to_nid(newpage
);
909 * Counterpart of unmap_and_move_page() for hugepage migration.
911 * This function doesn't wait the completion of hugepage I/O
912 * because there is no race between I/O and migration for hugepage.
913 * Note that currently hugepage I/O occurs only in direct I/O
914 * where no lock is held and PG_writeback is irrelevant,
915 * and writeback status of all subpages are counted in the reference
916 * count of the head page (i.e. if all subpages of a 2MB hugepage are
917 * under direct I/O, the reference of the head page is 512 and a bit more.)
918 * This means that when we try to migrate hugepage whose subpages are
919 * doing direct I/O, some references remain after try_to_unmap() and
920 * hugepage migration fails without data corruption.
922 * There is also no race when direct I/O is issued on the page under migration,
923 * because then pte is replaced with migration swap entry and direct I/O code
924 * will wait in the page fault for migration to complete.
926 static int unmap_and_move_huge_page(new_page_t get_new_page
,
927 unsigned long private, struct page
*hpage
,
928 int force
, enum migrate_mode mode
)
932 struct page
*new_hpage
= get_new_page(hpage
, private, &result
);
933 struct anon_vma
*anon_vma
= NULL
;
940 if (!trylock_page(hpage
)) {
941 if (!force
|| mode
!= MIGRATE_SYNC
)
947 anon_vma
= page_get_anon_vma(hpage
);
949 try_to_unmap(hpage
, TTU_MIGRATION
|TTU_IGNORE_MLOCK
|TTU_IGNORE_ACCESS
);
951 if (!page_mapped(hpage
))
952 rc
= move_to_new_page(new_hpage
, hpage
, 1, mode
);
955 remove_migration_ptes(hpage
, hpage
);
958 put_anon_vma(anon_vma
);
961 hugetlb_cgroup_migrate(hpage
, new_hpage
);
970 *result
= page_to_nid(new_hpage
);
976 * migrate_pages - migrate the pages specified in a list, to the free pages
977 * supplied as the target for the page migration
979 * @from: The list of pages to be migrated.
980 * @get_new_page: The function used to allocate free pages to be used
981 * as the target of the page migration.
982 * @private: Private data to be passed on to get_new_page()
983 * @mode: The migration mode that specifies the constraints for
984 * page migration, if any.
985 * @reason: The reason for page migration.
987 * The function returns after 10 attempts or if no pages are movable any more
988 * because the list has become empty or no retryable pages exist any more.
989 * The caller should call putback_lru_pages() to return pages to the LRU
990 * or free list only if ret != 0.
992 * Returns the number of pages that were not migrated, or an error code.
994 int migrate_pages(struct list_head
*from
, new_page_t get_new_page
,
995 unsigned long private, enum migrate_mode mode
, int reason
)
999 int nr_succeeded
= 0;
1003 int swapwrite
= current
->flags
& PF_SWAPWRITE
;
1007 current
->flags
|= PF_SWAPWRITE
;
1009 for(pass
= 0; pass
< 10 && retry
; pass
++) {
1012 list_for_each_entry_safe(page
, page2
, from
, lru
) {
1015 rc
= unmap_and_move(get_new_page
, private,
1016 page
, pass
> 2, mode
);
1024 case MIGRATEPAGE_SUCCESS
:
1028 /* Permanent failure */
1034 rc
= nr_failed
+ retry
;
1037 count_vm_events(PGMIGRATE_SUCCESS
, nr_succeeded
);
1039 count_vm_events(PGMIGRATE_FAIL
, nr_failed
);
1040 trace_mm_migrate_pages(nr_succeeded
, nr_failed
, mode
, reason
);
1043 current
->flags
&= ~PF_SWAPWRITE
;
1048 int migrate_huge_page(struct page
*hpage
, new_page_t get_new_page
,
1049 unsigned long private, enum migrate_mode mode
)
1053 for (pass
= 0; pass
< 10; pass
++) {
1054 rc
= unmap_and_move_huge_page(get_new_page
, private,
1055 hpage
, pass
> 2, mode
);
1063 case MIGRATEPAGE_SUCCESS
:
1076 * Move a list of individual pages
1078 struct page_to_node
{
1085 static struct page
*new_page_node(struct page
*p
, unsigned long private,
1088 struct page_to_node
*pm
= (struct page_to_node
*)private;
1090 while (pm
->node
!= MAX_NUMNODES
&& pm
->page
!= p
)
1093 if (pm
->node
== MAX_NUMNODES
)
1096 *result
= &pm
->status
;
1098 return alloc_pages_exact_node(pm
->node
,
1099 GFP_HIGHUSER_MOVABLE
| GFP_THISNODE
, 0);
1103 * Move a set of pages as indicated in the pm array. The addr
1104 * field must be set to the virtual address of the page to be moved
1105 * and the node number must contain a valid target node.
1106 * The pm array ends with node = MAX_NUMNODES.
1108 static int do_move_page_to_node_array(struct mm_struct
*mm
,
1109 struct page_to_node
*pm
,
1113 struct page_to_node
*pp
;
1114 LIST_HEAD(pagelist
);
1116 down_read(&mm
->mmap_sem
);
1119 * Build a list of pages to migrate
1121 for (pp
= pm
; pp
->node
!= MAX_NUMNODES
; pp
++) {
1122 struct vm_area_struct
*vma
;
1126 vma
= find_vma(mm
, pp
->addr
);
1127 if (!vma
|| pp
->addr
< vma
->vm_start
|| !vma_migratable(vma
))
1130 page
= follow_page(vma
, pp
->addr
, FOLL_GET
|FOLL_SPLIT
);
1132 err
= PTR_ERR(page
);
1140 /* Use PageReserved to check for zero page */
1141 if (PageReserved(page
))
1145 err
= page_to_nid(page
);
1147 if (err
== pp
->node
)
1149 * Node already in the right place
1154 if (page_mapcount(page
) > 1 &&
1158 err
= isolate_lru_page(page
);
1160 list_add_tail(&page
->lru
, &pagelist
);
1161 inc_zone_page_state(page
, NR_ISOLATED_ANON
+
1162 page_is_file_cache(page
));
1166 * Either remove the duplicate refcount from
1167 * isolate_lru_page() or drop the page ref if it was
1176 if (!list_empty(&pagelist
)) {
1177 err
= migrate_pages(&pagelist
, new_page_node
,
1178 (unsigned long)pm
, MIGRATE_SYNC
, MR_SYSCALL
);
1180 putback_lru_pages(&pagelist
);
1183 up_read(&mm
->mmap_sem
);
1188 * Migrate an array of page address onto an array of nodes and fill
1189 * the corresponding array of status.
1191 static int do_pages_move(struct mm_struct
*mm
, nodemask_t task_nodes
,
1192 unsigned long nr_pages
,
1193 const void __user
* __user
*pages
,
1194 const int __user
*nodes
,
1195 int __user
*status
, int flags
)
1197 struct page_to_node
*pm
;
1198 unsigned long chunk_nr_pages
;
1199 unsigned long chunk_start
;
1203 pm
= (struct page_to_node
*)__get_free_page(GFP_KERNEL
);
1210 * Store a chunk of page_to_node array in a page,
1211 * but keep the last one as a marker
1213 chunk_nr_pages
= (PAGE_SIZE
/ sizeof(struct page_to_node
)) - 1;
1215 for (chunk_start
= 0;
1216 chunk_start
< nr_pages
;
1217 chunk_start
+= chunk_nr_pages
) {
1220 if (chunk_start
+ chunk_nr_pages
> nr_pages
)
1221 chunk_nr_pages
= nr_pages
- chunk_start
;
1223 /* fill the chunk pm with addrs and nodes from user-space */
1224 for (j
= 0; j
< chunk_nr_pages
; j
++) {
1225 const void __user
*p
;
1229 if (get_user(p
, pages
+ j
+ chunk_start
))
1231 pm
[j
].addr
= (unsigned long) p
;
1233 if (get_user(node
, nodes
+ j
+ chunk_start
))
1237 if (node
< 0 || node
>= MAX_NUMNODES
)
1240 if (!node_state(node
, N_MEMORY
))
1244 if (!node_isset(node
, task_nodes
))
1250 /* End marker for this chunk */
1251 pm
[chunk_nr_pages
].node
= MAX_NUMNODES
;
1253 /* Migrate this chunk */
1254 err
= do_move_page_to_node_array(mm
, pm
,
1255 flags
& MPOL_MF_MOVE_ALL
);
1259 /* Return status information */
1260 for (j
= 0; j
< chunk_nr_pages
; j
++)
1261 if (put_user(pm
[j
].status
, status
+ j
+ chunk_start
)) {
1269 free_page((unsigned long)pm
);
1275 * Determine the nodes of an array of pages and store it in an array of status.
1277 static void do_pages_stat_array(struct mm_struct
*mm
, unsigned long nr_pages
,
1278 const void __user
**pages
, int *status
)
1282 down_read(&mm
->mmap_sem
);
1284 for (i
= 0; i
< nr_pages
; i
++) {
1285 unsigned long addr
= (unsigned long)(*pages
);
1286 struct vm_area_struct
*vma
;
1290 vma
= find_vma(mm
, addr
);
1291 if (!vma
|| addr
< vma
->vm_start
)
1294 page
= follow_page(vma
, addr
, 0);
1296 err
= PTR_ERR(page
);
1301 /* Use PageReserved to check for zero page */
1302 if (!page
|| PageReserved(page
))
1305 err
= page_to_nid(page
);
1313 up_read(&mm
->mmap_sem
);
1317 * Determine the nodes of a user array of pages and store it in
1318 * a user array of status.
1320 static int do_pages_stat(struct mm_struct
*mm
, unsigned long nr_pages
,
1321 const void __user
* __user
*pages
,
1324 #define DO_PAGES_STAT_CHUNK_NR 16
1325 const void __user
*chunk_pages
[DO_PAGES_STAT_CHUNK_NR
];
1326 int chunk_status
[DO_PAGES_STAT_CHUNK_NR
];
1329 unsigned long chunk_nr
;
1331 chunk_nr
= nr_pages
;
1332 if (chunk_nr
> DO_PAGES_STAT_CHUNK_NR
)
1333 chunk_nr
= DO_PAGES_STAT_CHUNK_NR
;
1335 if (copy_from_user(chunk_pages
, pages
, chunk_nr
* sizeof(*chunk_pages
)))
1338 do_pages_stat_array(mm
, chunk_nr
, chunk_pages
, chunk_status
);
1340 if (copy_to_user(status
, chunk_status
, chunk_nr
* sizeof(*status
)))
1345 nr_pages
-= chunk_nr
;
1347 return nr_pages
? -EFAULT
: 0;
1351 * Move a list of pages in the address space of the currently executing
1354 SYSCALL_DEFINE6(move_pages
, pid_t
, pid
, unsigned long, nr_pages
,
1355 const void __user
* __user
*, pages
,
1356 const int __user
*, nodes
,
1357 int __user
*, status
, int, flags
)
1359 const struct cred
*cred
= current_cred(), *tcred
;
1360 struct task_struct
*task
;
1361 struct mm_struct
*mm
;
1363 nodemask_t task_nodes
;
1366 if (flags
& ~(MPOL_MF_MOVE
|MPOL_MF_MOVE_ALL
))
1369 if ((flags
& MPOL_MF_MOVE_ALL
) && !capable(CAP_SYS_NICE
))
1372 /* Find the mm_struct */
1374 task
= pid
? find_task_by_vpid(pid
) : current
;
1379 get_task_struct(task
);
1382 * Check if this process has the right to modify the specified
1383 * process. The right exists if the process has administrative
1384 * capabilities, superuser privileges or the same
1385 * userid as the target process.
1387 tcred
= __task_cred(task
);
1388 if (!uid_eq(cred
->euid
, tcred
->suid
) && !uid_eq(cred
->euid
, tcred
->uid
) &&
1389 !uid_eq(cred
->uid
, tcred
->suid
) && !uid_eq(cred
->uid
, tcred
->uid
) &&
1390 !capable(CAP_SYS_NICE
)) {
1397 err
= security_task_movememory(task
);
1401 task_nodes
= cpuset_mems_allowed(task
);
1402 mm
= get_task_mm(task
);
1403 put_task_struct(task
);
1409 err
= do_pages_move(mm
, task_nodes
, nr_pages
, pages
,
1410 nodes
, status
, flags
);
1412 err
= do_pages_stat(mm
, nr_pages
, pages
, status
);
1418 put_task_struct(task
);
1423 * Call migration functions in the vma_ops that may prepare
1424 * memory in a vm for migration. migration functions may perform
1425 * the migration for vmas that do not have an underlying page struct.
1427 int migrate_vmas(struct mm_struct
*mm
, const nodemask_t
*to
,
1428 const nodemask_t
*from
, unsigned long flags
)
1430 struct vm_area_struct
*vma
;
1433 for (vma
= mm
->mmap
; vma
&& !err
; vma
= vma
->vm_next
) {
1434 if (vma
->vm_ops
&& vma
->vm_ops
->migrate
) {
1435 err
= vma
->vm_ops
->migrate(vma
, to
, from
, flags
);
1443 #ifdef CONFIG_NUMA_BALANCING
1445 * Returns true if this is a safe migration target node for misplaced NUMA
1446 * pages. Currently it only checks the watermarks which crude
1448 static bool migrate_balanced_pgdat(struct pglist_data
*pgdat
,
1449 unsigned long nr_migrate_pages
)
1452 for (z
= pgdat
->nr_zones
- 1; z
>= 0; z
--) {
1453 struct zone
*zone
= pgdat
->node_zones
+ z
;
1455 if (!populated_zone(zone
))
1458 if (zone
->all_unreclaimable
)
1461 /* Avoid waking kswapd by allocating pages_to_migrate pages. */
1462 if (!zone_watermark_ok(zone
, 0,
1463 high_wmark_pages(zone
) +
1472 static struct page
*alloc_misplaced_dst_page(struct page
*page
,
1476 int nid
= (int) data
;
1477 struct page
*newpage
;
1479 newpage
= alloc_pages_exact_node(nid
,
1480 (GFP_HIGHUSER_MOVABLE
| GFP_THISNODE
|
1481 __GFP_NOMEMALLOC
| __GFP_NORETRY
|
1485 page_nid_xchg_last(newpage
, page_nid_last(page
));
1491 * page migration rate limiting control.
1492 * Do not migrate more than @pages_to_migrate in a @migrate_interval_millisecs
1493 * window of time. Default here says do not migrate more than 1280M per second.
1494 * If a node is rate-limited then PTE NUMA updates are also rate-limited. However
1495 * as it is faults that reset the window, pte updates will happen unconditionally
1496 * if there has not been a fault since @pteupdate_interval_millisecs after the
1497 * throttle window closed.
1499 static unsigned int migrate_interval_millisecs __read_mostly
= 100;
1500 static unsigned int pteupdate_interval_millisecs __read_mostly
= 1000;
1501 static unsigned int ratelimit_pages __read_mostly
= 128 << (20 - PAGE_SHIFT
);
1503 /* Returns true if NUMA migration is currently rate limited */
1504 bool migrate_ratelimited(int node
)
1506 pg_data_t
*pgdat
= NODE_DATA(node
);
1508 if (time_after(jiffies
, pgdat
->numabalancing_migrate_next_window
+
1509 msecs_to_jiffies(pteupdate_interval_millisecs
)))
1512 if (pgdat
->numabalancing_migrate_nr_pages
< ratelimit_pages
)
1518 /* Returns true if the node is migrate rate-limited after the update */
1519 bool numamigrate_update_ratelimit(pg_data_t
*pgdat
, unsigned long nr_pages
)
1521 bool rate_limited
= false;
1524 * Rate-limit the amount of data that is being migrated to a node.
1525 * Optimal placement is no good if the memory bus is saturated and
1526 * all the time is being spent migrating!
1528 spin_lock(&pgdat
->numabalancing_migrate_lock
);
1529 if (time_after(jiffies
, pgdat
->numabalancing_migrate_next_window
)) {
1530 pgdat
->numabalancing_migrate_nr_pages
= 0;
1531 pgdat
->numabalancing_migrate_next_window
= jiffies
+
1532 msecs_to_jiffies(migrate_interval_millisecs
);
1534 if (pgdat
->numabalancing_migrate_nr_pages
> ratelimit_pages
)
1535 rate_limited
= true;
1537 pgdat
->numabalancing_migrate_nr_pages
+= nr_pages
;
1538 spin_unlock(&pgdat
->numabalancing_migrate_lock
);
1540 return rate_limited
;
1543 int numamigrate_isolate_page(pg_data_t
*pgdat
, struct page
*page
)
1547 VM_BUG_ON(compound_order(page
) && !PageTransHuge(page
));
1549 /* Avoid migrating to a node that is nearly full */
1550 if (!migrate_balanced_pgdat(pgdat
, 1UL << compound_order(page
)))
1553 if (isolate_lru_page(page
))
1557 * migrate_misplaced_transhuge_page() skips page migration's usual
1558 * check on page_count(), so we must do it here, now that the page
1559 * has been isolated: a GUP pin, or any other pin, prevents migration.
1560 * The expected page count is 3: 1 for page's mapcount and 1 for the
1561 * caller's pin and 1 for the reference taken by isolate_lru_page().
1563 if (PageTransHuge(page
) && page_count(page
) != 3) {
1564 putback_lru_page(page
);
1568 page_lru
= page_is_file_cache(page
);
1569 mod_zone_page_state(page_zone(page
), NR_ISOLATED_ANON
+ page_lru
,
1570 hpage_nr_pages(page
));
1573 * Isolating the page has taken another reference, so the
1574 * caller's reference can be safely dropped without the page
1575 * disappearing underneath us during migration.
1582 * Attempt to migrate a misplaced page to the specified destination
1583 * node. Caller is expected to have an elevated reference count on
1584 * the page that will be dropped by this function before returning.
1586 int migrate_misplaced_page(struct page
*page
, int node
)
1588 pg_data_t
*pgdat
= NODE_DATA(node
);
1591 LIST_HEAD(migratepages
);
1594 * Don't migrate pages that are mapped in multiple processes.
1595 * TODO: Handle false sharing detection instead of this hammer
1597 if (page_mapcount(page
) != 1)
1601 * Rate-limit the amount of data that is being migrated to a node.
1602 * Optimal placement is no good if the memory bus is saturated and
1603 * all the time is being spent migrating!
1605 if (numamigrate_update_ratelimit(pgdat
, 1))
1608 isolated
= numamigrate_isolate_page(pgdat
, page
);
1612 list_add(&page
->lru
, &migratepages
);
1613 nr_remaining
= migrate_pages(&migratepages
, alloc_misplaced_dst_page
,
1614 node
, MIGRATE_ASYNC
, MR_NUMA_MISPLACED
);
1616 putback_lru_pages(&migratepages
);
1619 count_vm_numa_event(NUMA_PAGE_MIGRATE
);
1620 BUG_ON(!list_empty(&migratepages
));
1627 #endif /* CONFIG_NUMA_BALANCING */
1629 #if defined(CONFIG_NUMA_BALANCING) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
1631 * Migrates a THP to a given target node. page must be locked and is unlocked
1634 int migrate_misplaced_transhuge_page(struct mm_struct
*mm
,
1635 struct vm_area_struct
*vma
,
1636 pmd_t
*pmd
, pmd_t entry
,
1637 unsigned long address
,
1638 struct page
*page
, int node
)
1640 unsigned long haddr
= address
& HPAGE_PMD_MASK
;
1641 pg_data_t
*pgdat
= NODE_DATA(node
);
1643 struct page
*new_page
= NULL
;
1644 struct mem_cgroup
*memcg
= NULL
;
1645 int page_lru
= page_is_file_cache(page
);
1648 * Don't migrate pages that are mapped in multiple processes.
1649 * TODO: Handle false sharing detection instead of this hammer
1651 if (page_mapcount(page
) != 1)
1655 * Rate-limit the amount of data that is being migrated to a node.
1656 * Optimal placement is no good if the memory bus is saturated and
1657 * all the time is being spent migrating!
1659 if (numamigrate_update_ratelimit(pgdat
, HPAGE_PMD_NR
))
1662 new_page
= alloc_pages_node(node
,
1663 (GFP_TRANSHUGE
| GFP_THISNODE
) & ~__GFP_WAIT
, HPAGE_PMD_ORDER
);
1667 page_nid_xchg_last(new_page
, page_nid_last(page
));
1669 isolated
= numamigrate_isolate_page(pgdat
, page
);
1675 /* Prepare a page as a migration target */
1676 __set_page_locked(new_page
);
1677 SetPageSwapBacked(new_page
);
1679 /* anon mapping, we can simply copy page->mapping to the new page: */
1680 new_page
->mapping
= page
->mapping
;
1681 new_page
->index
= page
->index
;
1682 migrate_page_copy(new_page
, page
);
1683 WARN_ON(PageLRU(new_page
));
1685 /* Recheck the target PMD */
1686 spin_lock(&mm
->page_table_lock
);
1687 if (unlikely(!pmd_same(*pmd
, entry
))) {
1688 spin_unlock(&mm
->page_table_lock
);
1690 /* Reverse changes made by migrate_page_copy() */
1691 if (TestClearPageActive(new_page
))
1692 SetPageActive(page
);
1693 if (TestClearPageUnevictable(new_page
))
1694 SetPageUnevictable(page
);
1695 mlock_migrate_page(page
, new_page
);
1697 unlock_page(new_page
);
1698 put_page(new_page
); /* Free it */
1701 putback_lru_page(page
);
1703 count_vm_events(PGMIGRATE_FAIL
, HPAGE_PMD_NR
);
1709 * Traditional migration needs to prepare the memcg charge
1710 * transaction early to prevent the old page from being
1711 * uncharged when installing migration entries. Here we can
1712 * save the potential rollback and start the charge transfer
1713 * only when migration is already known to end successfully.
1715 mem_cgroup_prepare_migration(page
, new_page
, &memcg
);
1717 entry
= mk_pmd(new_page
, vma
->vm_page_prot
);
1718 entry
= pmd_mknonnuma(entry
);
1719 entry
= maybe_pmd_mkwrite(pmd_mkdirty(entry
), vma
);
1720 entry
= pmd_mkhuge(entry
);
1722 page_add_new_anon_rmap(new_page
, vma
, haddr
);
1724 set_pmd_at(mm
, haddr
, pmd
, entry
);
1725 update_mmu_cache_pmd(vma
, address
, &entry
);
1726 page_remove_rmap(page
);
1728 * Finish the charge transaction under the page table lock to
1729 * prevent split_huge_page() from dividing up the charge
1730 * before it's fully transferred to the new page.
1732 mem_cgroup_end_migration(memcg
, page
, new_page
, true);
1733 spin_unlock(&mm
->page_table_lock
);
1735 unlock_page(new_page
);
1737 put_page(page
); /* Drop the rmap reference */
1738 put_page(page
); /* Drop the LRU isolation reference */
1740 count_vm_events(PGMIGRATE_SUCCESS
, HPAGE_PMD_NR
);
1741 count_vm_numa_events(NUMA_PAGE_MIGRATE
, HPAGE_PMD_NR
);
1744 mod_zone_page_state(page_zone(page
),
1745 NR_ISOLATED_ANON
+ page_lru
,
1750 count_vm_events(PGMIGRATE_FAIL
, HPAGE_PMD_NR
);
1756 #endif /* CONFIG_NUMA_BALANCING */
1758 #endif /* CONFIG_NUMA */