Loading...
1/*
2 * Memory Migration functionality - linux/mm/migration.c
3 *
4 * Copyright (C) 2006 Silicon Graphics, Inc., Christoph Lameter
5 *
6 * Page migration was first developed in the context of the memory hotplug
7 * project. The main authors of the migration code are:
8 *
9 * IWAMOTO Toshihiro <iwamoto@valinux.co.jp>
10 * Hirokazu Takahashi <taka@valinux.co.jp>
11 * Dave Hansen <haveblue@us.ibm.com>
12 * Christoph Lameter
13 */
14
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/gfp.h>
37
38#include <asm/tlbflush.h>
39
40#include "internal.h"
41
42/*
43 * migrate_prep() needs to be called before we start compiling a list of pages
44 * to be migrated using isolate_lru_page(). If scheduling work on other CPUs is
45 * undesirable, use migrate_prep_local()
46 */
47int migrate_prep(void)
48{
49 /*
50 * Clear the LRU lists so pages can be isolated.
51 * Note that pages may be moved off the LRU after we have
52 * drained them. Those pages will fail to migrate like other
53 * pages that may be busy.
54 */
55 lru_add_drain_all();
56
57 return 0;
58}
59
60/* Do the necessary work of migrate_prep but not if it involves other CPUs */
61int migrate_prep_local(void)
62{
63 lru_add_drain();
64
65 return 0;
66}
67
68/*
69 * Add isolated pages on the list back to the LRU under page lock
70 * to avoid leaking evictable pages back onto unevictable list.
71 */
72void putback_lru_pages(struct list_head *l)
73{
74 struct page *page;
75 struct page *page2;
76
77 list_for_each_entry_safe(page, page2, l, lru) {
78 list_del(&page->lru);
79 dec_zone_page_state(page, NR_ISOLATED_ANON +
80 page_is_file_cache(page));
81 putback_lru_page(page);
82 }
83}
84
85/*
86 * Restore a potential migration pte to a working pte entry
87 */
88static int remove_migration_pte(struct page *new, struct vm_area_struct *vma,
89 unsigned long addr, void *old)
90{
91 struct mm_struct *mm = vma->vm_mm;
92 swp_entry_t entry;
93 pgd_t *pgd;
94 pud_t *pud;
95 pmd_t *pmd;
96 pte_t *ptep, pte;
97 spinlock_t *ptl;
98
99 if (unlikely(PageHuge(new))) {
100 ptep = huge_pte_offset(mm, addr);
101 if (!ptep)
102 goto out;
103 ptl = &mm->page_table_lock;
104 } else {
105 pgd = pgd_offset(mm, addr);
106 if (!pgd_present(*pgd))
107 goto out;
108
109 pud = pud_offset(pgd, addr);
110 if (!pud_present(*pud))
111 goto out;
112
113 pmd = pmd_offset(pud, addr);
114 if (pmd_trans_huge(*pmd))
115 goto out;
116 if (!pmd_present(*pmd))
117 goto out;
118
119 ptep = pte_offset_map(pmd, addr);
120
121 /*
122 * Peek to check is_swap_pte() before taking ptlock? No, we
123 * can race mremap's move_ptes(), which skips anon_vma lock.
124 */
125
126 ptl = pte_lockptr(mm, pmd);
127 }
128
129 spin_lock(ptl);
130 pte = *ptep;
131 if (!is_swap_pte(pte))
132 goto unlock;
133
134 entry = pte_to_swp_entry(pte);
135
136 if (!is_migration_entry(entry) ||
137 migration_entry_to_page(entry) != old)
138 goto unlock;
139
140 get_page(new);
141 pte = pte_mkold(mk_pte(new, vma->vm_page_prot));
142 if (is_write_migration_entry(entry))
143 pte = pte_mkwrite(pte);
144#ifdef CONFIG_HUGETLB_PAGE
145 if (PageHuge(new))
146 pte = pte_mkhuge(pte);
147#endif
148 flush_cache_page(vma, addr, pte_pfn(pte));
149 set_pte_at(mm, addr, ptep, pte);
150
151 if (PageHuge(new)) {
152 if (PageAnon(new))
153 hugepage_add_anon_rmap(new, vma, addr);
154 else
155 page_dup_rmap(new);
156 } else if (PageAnon(new))
157 page_add_anon_rmap(new, vma, addr);
158 else
159 page_add_file_rmap(new);
160
161 /* No need to invalidate - it was non-present before */
162 update_mmu_cache(vma, addr, ptep);
163unlock:
164 pte_unmap_unlock(ptep, ptl);
165out:
166 return SWAP_AGAIN;
167}
168
169/*
170 * Get rid of all migration entries and replace them by
171 * references to the indicated page.
172 */
173static void remove_migration_ptes(struct page *old, struct page *new)
174{
175 rmap_walk(new, remove_migration_pte, old);
176}
177
178/*
179 * Something used the pte of a page under migration. We need to
180 * get to the page and wait until migration is finished.
181 * When we return from this function the fault will be retried.
182 */
183void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
184 unsigned long address)
185{
186 pte_t *ptep, pte;
187 spinlock_t *ptl;
188 swp_entry_t entry;
189 struct page *page;
190
191 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
192 pte = *ptep;
193 if (!is_swap_pte(pte))
194 goto out;
195
196 entry = pte_to_swp_entry(pte);
197 if (!is_migration_entry(entry))
198 goto out;
199
200 page = migration_entry_to_page(entry);
201
202 /*
203 * Once radix-tree replacement of page migration started, page_count
204 * *must* be zero. And, we don't want to call wait_on_page_locked()
205 * against a page without get_page().
206 * So, we use get_page_unless_zero(), here. Even failed, page fault
207 * will occur again.
208 */
209 if (!get_page_unless_zero(page))
210 goto out;
211 pte_unmap_unlock(ptep, ptl);
212 wait_on_page_locked(page);
213 put_page(page);
214 return;
215out:
216 pte_unmap_unlock(ptep, ptl);
217}
218
219#ifdef CONFIG_BLOCK
220/* Returns true if all buffers are successfully locked */
221static bool buffer_migrate_lock_buffers(struct buffer_head *head,
222 enum migrate_mode mode)
223{
224 struct buffer_head *bh = head;
225
226 /* Simple case, sync compaction */
227 if (mode != MIGRATE_ASYNC) {
228 do {
229 get_bh(bh);
230 lock_buffer(bh);
231 bh = bh->b_this_page;
232
233 } while (bh != head);
234
235 return true;
236 }
237
238 /* async case, we cannot block on lock_buffer so use trylock_buffer */
239 do {
240 get_bh(bh);
241 if (!trylock_buffer(bh)) {
242 /*
243 * We failed to lock the buffer and cannot stall in
244 * async migration. Release the taken locks
245 */
246 struct buffer_head *failed_bh = bh;
247 put_bh(failed_bh);
248 bh = head;
249 while (bh != failed_bh) {
250 unlock_buffer(bh);
251 put_bh(bh);
252 bh = bh->b_this_page;
253 }
254 return false;
255 }
256
257 bh = bh->b_this_page;
258 } while (bh != head);
259 return true;
260}
261#else
262static inline bool buffer_migrate_lock_buffers(struct buffer_head *head,
263 enum migrate_mode mode)
264{
265 return true;
266}
267#endif /* CONFIG_BLOCK */
268
269/*
270 * Replace the page in the mapping.
271 *
272 * The number of remaining references must be:
273 * 1 for anonymous pages without a mapping
274 * 2 for pages with a mapping
275 * 3 for pages with a mapping and PagePrivate/PagePrivate2 set.
276 */
277static int migrate_page_move_mapping(struct address_space *mapping,
278 struct page *newpage, struct page *page,
279 struct buffer_head *head, enum migrate_mode mode)
280{
281 int expected_count;
282 void **pslot;
283
284 if (!mapping) {
285 /* Anonymous page without mapping */
286 if (page_count(page) != 1)
287 return -EAGAIN;
288 return 0;
289 }
290
291 spin_lock_irq(&mapping->tree_lock);
292
293 pslot = radix_tree_lookup_slot(&mapping->page_tree,
294 page_index(page));
295
296 expected_count = 2 + page_has_private(page);
297 if (page_count(page) != expected_count ||
298 radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) {
299 spin_unlock_irq(&mapping->tree_lock);
300 return -EAGAIN;
301 }
302
303 if (!page_freeze_refs(page, expected_count)) {
304 spin_unlock_irq(&mapping->tree_lock);
305 return -EAGAIN;
306 }
307
308 /*
309 * In the async migration case of moving a page with buffers, lock the
310 * buffers using trylock before the mapping is moved. If the mapping
311 * was moved, we later failed to lock the buffers and could not move
312 * the mapping back due to an elevated page count, we would have to
313 * block waiting on other references to be dropped.
314 */
315 if (mode == MIGRATE_ASYNC && head &&
316 !buffer_migrate_lock_buffers(head, mode)) {
317 page_unfreeze_refs(page, expected_count);
318 spin_unlock_irq(&mapping->tree_lock);
319 return -EAGAIN;
320 }
321
322 /*
323 * Now we know that no one else is looking at the page.
324 */
325 get_page(newpage); /* add cache reference */
326 if (PageSwapCache(page)) {
327 SetPageSwapCache(newpage);
328 set_page_private(newpage, page_private(page));
329 }
330
331 radix_tree_replace_slot(pslot, newpage);
332
333 /*
334 * Drop cache reference from old page by unfreezing
335 * to one less reference.
336 * We know this isn't the last reference.
337 */
338 page_unfreeze_refs(page, expected_count - 1);
339
340 /*
341 * If moved to a different zone then also account
342 * the page for that zone. Other VM counters will be
343 * taken care of when we establish references to the
344 * new page and drop references to the old page.
345 *
346 * Note that anonymous pages are accounted for
347 * via NR_FILE_PAGES and NR_ANON_PAGES if they
348 * are mapped to swap space.
349 */
350 __dec_zone_page_state(page, NR_FILE_PAGES);
351 __inc_zone_page_state(newpage, NR_FILE_PAGES);
352 if (!PageSwapCache(page) && PageSwapBacked(page)) {
353 __dec_zone_page_state(page, NR_SHMEM);
354 __inc_zone_page_state(newpage, NR_SHMEM);
355 }
356 spin_unlock_irq(&mapping->tree_lock);
357
358 return 0;
359}
360
361/*
362 * The expected number of remaining references is the same as that
363 * of migrate_page_move_mapping().
364 */
365int migrate_huge_page_move_mapping(struct address_space *mapping,
366 struct page *newpage, struct page *page)
367{
368 int expected_count;
369 void **pslot;
370
371 if (!mapping) {
372 if (page_count(page) != 1)
373 return -EAGAIN;
374 return 0;
375 }
376
377 spin_lock_irq(&mapping->tree_lock);
378
379 pslot = radix_tree_lookup_slot(&mapping->page_tree,
380 page_index(page));
381
382 expected_count = 2 + page_has_private(page);
383 if (page_count(page) != expected_count ||
384 radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) {
385 spin_unlock_irq(&mapping->tree_lock);
386 return -EAGAIN;
387 }
388
389 if (!page_freeze_refs(page, expected_count)) {
390 spin_unlock_irq(&mapping->tree_lock);
391 return -EAGAIN;
392 }
393
394 get_page(newpage);
395
396 radix_tree_replace_slot(pslot, newpage);
397
398 page_unfreeze_refs(page, expected_count - 1);
399
400 spin_unlock_irq(&mapping->tree_lock);
401 return 0;
402}
403
404/*
405 * Copy the page to its new location
406 */
407void migrate_page_copy(struct page *newpage, struct page *page)
408{
409 if (PageHuge(page))
410 copy_huge_page(newpage, page);
411 else
412 copy_highpage(newpage, page);
413
414 if (PageError(page))
415 SetPageError(newpage);
416 if (PageReferenced(page))
417 SetPageReferenced(newpage);
418 if (PageUptodate(page))
419 SetPageUptodate(newpage);
420 if (TestClearPageActive(page)) {
421 VM_BUG_ON(PageUnevictable(page));
422 SetPageActive(newpage);
423 } else if (TestClearPageUnevictable(page))
424 SetPageUnevictable(newpage);
425 if (PageChecked(page))
426 SetPageChecked(newpage);
427 if (PageMappedToDisk(page))
428 SetPageMappedToDisk(newpage);
429
430 if (PageDirty(page)) {
431 clear_page_dirty_for_io(page);
432 /*
433 * Want to mark the page and the radix tree as dirty, and
434 * redo the accounting that clear_page_dirty_for_io undid,
435 * but we can't use set_page_dirty because that function
436 * is actually a signal that all of the page has become dirty.
437 * Whereas only part of our page may be dirty.
438 */
439 if (PageSwapBacked(page))
440 SetPageDirty(newpage);
441 else
442 __set_page_dirty_nobuffers(newpage);
443 }
444
445 mlock_migrate_page(newpage, page);
446 ksm_migrate_page(newpage, page);
447
448 ClearPageSwapCache(page);
449 ClearPagePrivate(page);
450 set_page_private(page, 0);
451
452 /*
453 * If any waiters have accumulated on the new page then
454 * wake them up.
455 */
456 if (PageWriteback(newpage))
457 end_page_writeback(newpage);
458}
459
460/************************************************************
461 * Migration functions
462 ***********************************************************/
463
464/* Always fail migration. Used for mappings that are not movable */
465int fail_migrate_page(struct address_space *mapping,
466 struct page *newpage, struct page *page)
467{
468 return -EIO;
469}
470EXPORT_SYMBOL(fail_migrate_page);
471
472/*
473 * Common logic to directly migrate a single page suitable for
474 * pages that do not use PagePrivate/PagePrivate2.
475 *
476 * Pages are locked upon entry and exit.
477 */
478int migrate_page(struct address_space *mapping,
479 struct page *newpage, struct page *page,
480 enum migrate_mode mode)
481{
482 int rc;
483
484 BUG_ON(PageWriteback(page)); /* Writeback must be complete */
485
486 rc = migrate_page_move_mapping(mapping, newpage, page, NULL, mode);
487
488 if (rc)
489 return rc;
490
491 migrate_page_copy(newpage, page);
492 return 0;
493}
494EXPORT_SYMBOL(migrate_page);
495
496#ifdef CONFIG_BLOCK
497/*
498 * Migration function for pages with buffers. This function can only be used
499 * if the underlying filesystem guarantees that no other references to "page"
500 * exist.
501 */
502int buffer_migrate_page(struct address_space *mapping,
503 struct page *newpage, struct page *page, enum migrate_mode mode)
504{
505 struct buffer_head *bh, *head;
506 int rc;
507
508 if (!page_has_buffers(page))
509 return migrate_page(mapping, newpage, page, mode);
510
511 head = page_buffers(page);
512
513 rc = migrate_page_move_mapping(mapping, newpage, page, head, mode);
514
515 if (rc)
516 return rc;
517
518 /*
519 * In the async case, migrate_page_move_mapping locked the buffers
520 * with an IRQ-safe spinlock held. In the sync case, the buffers
521 * need to be locked now
522 */
523 if (mode != MIGRATE_ASYNC)
524 BUG_ON(!buffer_migrate_lock_buffers(head, mode));
525
526 ClearPagePrivate(page);
527 set_page_private(newpage, page_private(page));
528 set_page_private(page, 0);
529 put_page(page);
530 get_page(newpage);
531
532 bh = head;
533 do {
534 set_bh_page(bh, newpage, bh_offset(bh));
535 bh = bh->b_this_page;
536
537 } while (bh != head);
538
539 SetPagePrivate(newpage);
540
541 migrate_page_copy(newpage, page);
542
543 bh = head;
544 do {
545 unlock_buffer(bh);
546 put_bh(bh);
547 bh = bh->b_this_page;
548
549 } while (bh != head);
550
551 return 0;
552}
553EXPORT_SYMBOL(buffer_migrate_page);
554#endif
555
556/*
557 * Writeback a page to clean the dirty state
558 */
559static int writeout(struct address_space *mapping, struct page *page)
560{
561 struct writeback_control wbc = {
562 .sync_mode = WB_SYNC_NONE,
563 .nr_to_write = 1,
564 .range_start = 0,
565 .range_end = LLONG_MAX,
566 .for_reclaim = 1
567 };
568 int rc;
569
570 if (!mapping->a_ops->writepage)
571 /* No write method for the address space */
572 return -EINVAL;
573
574 if (!clear_page_dirty_for_io(page))
575 /* Someone else already triggered a write */
576 return -EAGAIN;
577
578 /*
579 * A dirty page may imply that the underlying filesystem has
580 * the page on some queue. So the page must be clean for
581 * migration. Writeout may mean we loose the lock and the
582 * page state is no longer what we checked for earlier.
583 * At this point we know that the migration attempt cannot
584 * be successful.
585 */
586 remove_migration_ptes(page, page);
587
588 rc = mapping->a_ops->writepage(page, &wbc);
589
590 if (rc != AOP_WRITEPAGE_ACTIVATE)
591 /* unlocked. Relock */
592 lock_page(page);
593
594 return (rc < 0) ? -EIO : -EAGAIN;
595}
596
597/*
598 * Default handling if a filesystem does not provide a migration function.
599 */
600static int fallback_migrate_page(struct address_space *mapping,
601 struct page *newpage, struct page *page, enum migrate_mode mode)
602{
603 if (PageDirty(page)) {
604 /* Only writeback pages in full synchronous migration */
605 if (mode != MIGRATE_SYNC)
606 return -EBUSY;
607 return writeout(mapping, page);
608 }
609
610 /*
611 * Buffers may be managed in a filesystem specific way.
612 * We must have no buffers or drop them.
613 */
614 if (page_has_private(page) &&
615 !try_to_release_page(page, GFP_KERNEL))
616 return -EAGAIN;
617
618 return migrate_page(mapping, newpage, page, mode);
619}
620
621/*
622 * Move a page to a newly allocated page
623 * The page is locked and all ptes have been successfully removed.
624 *
625 * The new page will have replaced the old page if this function
626 * is successful.
627 *
628 * Return value:
629 * < 0 - error code
630 * == 0 - success
631 */
632static int move_to_new_page(struct page *newpage, struct page *page,
633 int remap_swapcache, enum migrate_mode mode)
634{
635 struct address_space *mapping;
636 int rc;
637
638 /*
639 * Block others from accessing the page when we get around to
640 * establishing additional references. We are the only one
641 * holding a reference to the new page at this point.
642 */
643 if (!trylock_page(newpage))
644 BUG();
645
646 /* Prepare mapping for the new page.*/
647 newpage->index = page->index;
648 newpage->mapping = page->mapping;
649 if (PageSwapBacked(page))
650 SetPageSwapBacked(newpage);
651
652 mapping = page_mapping(page);
653 if (!mapping)
654 rc = migrate_page(mapping, newpage, page, mode);
655 else if (mapping->a_ops->migratepage)
656 /*
657 * Most pages have a mapping and most filesystems provide a
658 * migratepage callback. Anonymous pages are part of swap
659 * space which also has its own migratepage callback. This
660 * is the most common path for page migration.
661 */
662 rc = mapping->a_ops->migratepage(mapping,
663 newpage, page, mode);
664 else
665 rc = fallback_migrate_page(mapping, newpage, page, mode);
666
667 if (rc) {
668 newpage->mapping = NULL;
669 } else {
670 if (remap_swapcache)
671 remove_migration_ptes(page, newpage);
672 page->mapping = NULL;
673 }
674
675 unlock_page(newpage);
676
677 return rc;
678}
679
680static int __unmap_and_move(struct page *page, struct page *newpage,
681 int force, bool offlining, enum migrate_mode mode)
682{
683 int rc = -EAGAIN;
684 int remap_swapcache = 1;
685 int charge = 0;
686 struct mem_cgroup *mem;
687 struct anon_vma *anon_vma = NULL;
688
689 if (!trylock_page(page)) {
690 if (!force || mode == MIGRATE_ASYNC)
691 goto out;
692
693 /*
694 * It's not safe for direct compaction to call lock_page.
695 * For example, during page readahead pages are added locked
696 * to the LRU. Later, when the IO completes the pages are
697 * marked uptodate and unlocked. However, the queueing
698 * could be merging multiple pages for one bio (e.g.
699 * mpage_readpages). If an allocation happens for the
700 * second or third page, the process can end up locking
701 * the same page twice and deadlocking. Rather than
702 * trying to be clever about what pages can be locked,
703 * avoid the use of lock_page for direct compaction
704 * altogether.
705 */
706 if (current->flags & PF_MEMALLOC)
707 goto out;
708
709 lock_page(page);
710 }
711
712 /*
713 * Only memory hotplug's offline_pages() caller has locked out KSM,
714 * and can safely migrate a KSM page. The other cases have skipped
715 * PageKsm along with PageReserved - but it is only now when we have
716 * the page lock that we can be certain it will not go KSM beneath us
717 * (KSM will not upgrade a page from PageAnon to PageKsm when it sees
718 * its pagecount raised, but only here do we take the page lock which
719 * serializes that).
720 */
721 if (PageKsm(page) && !offlining) {
722 rc = -EBUSY;
723 goto unlock;
724 }
725
726 /* charge against new page */
727 charge = mem_cgroup_prepare_migration(page, newpage, &mem, GFP_KERNEL);
728 if (charge == -ENOMEM) {
729 rc = -ENOMEM;
730 goto unlock;
731 }
732 BUG_ON(charge);
733
734 if (PageWriteback(page)) {
735 /*
736 * Only in the case of a full syncronous migration is it
737 * necessary to wait for PageWriteback. In the async case,
738 * the retry loop is too short and in the sync-light case,
739 * the overhead of stalling is too much
740 */
741 if (mode != MIGRATE_SYNC) {
742 rc = -EBUSY;
743 goto uncharge;
744 }
745 if (!force)
746 goto uncharge;
747 wait_on_page_writeback(page);
748 }
749 /*
750 * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
751 * we cannot notice that anon_vma is freed while we migrates a page.
752 * This get_anon_vma() delays freeing anon_vma pointer until the end
753 * of migration. File cache pages are no problem because of page_lock()
754 * File Caches may use write_page() or lock_page() in migration, then,
755 * just care Anon page here.
756 */
757 if (PageAnon(page)) {
758 /*
759 * Only page_lock_anon_vma() understands the subtleties of
760 * getting a hold on an anon_vma from outside one of its mms.
761 */
762 anon_vma = page_get_anon_vma(page);
763 if (anon_vma) {
764 /*
765 * Anon page
766 */
767 } else if (PageSwapCache(page)) {
768 /*
769 * We cannot be sure that the anon_vma of an unmapped
770 * swapcache page is safe to use because we don't
771 * know in advance if the VMA that this page belonged
772 * to still exists. If the VMA and others sharing the
773 * data have been freed, then the anon_vma could
774 * already be invalid.
775 *
776 * To avoid this possibility, swapcache pages get
777 * migrated but are not remapped when migration
778 * completes
779 */
780 remap_swapcache = 0;
781 } else {
782 goto uncharge;
783 }
784 }
785
786 /*
787 * Corner case handling:
788 * 1. When a new swap-cache page is read into, it is added to the LRU
789 * and treated as swapcache but it has no rmap yet.
790 * Calling try_to_unmap() against a page->mapping==NULL page will
791 * trigger a BUG. So handle it here.
792 * 2. An orphaned page (see truncate_complete_page) might have
793 * fs-private metadata. The page can be picked up due to memory
794 * offlining. Everywhere else except page reclaim, the page is
795 * invisible to the vm, so the page can not be migrated. So try to
796 * free the metadata, so the page can be freed.
797 */
798 if (!page->mapping) {
799 VM_BUG_ON(PageAnon(page));
800 if (page_has_private(page)) {
801 try_to_free_buffers(page);
802 goto uncharge;
803 }
804 goto skip_unmap;
805 }
806
807 /* Establish migration ptes or remove ptes */
808 try_to_unmap(page, TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
809
810skip_unmap:
811 if (!page_mapped(page))
812 rc = move_to_new_page(newpage, page, remap_swapcache, mode);
813
814 if (rc && remap_swapcache)
815 remove_migration_ptes(page, page);
816
817 /* Drop an anon_vma reference if we took one */
818 if (anon_vma)
819 put_anon_vma(anon_vma);
820
821uncharge:
822 if (!charge)
823 mem_cgroup_end_migration(mem, page, newpage, rc == 0);
824unlock:
825 unlock_page(page);
826out:
827 return rc;
828}
829
830/*
831 * Obtain the lock on page, remove all ptes and migrate the page
832 * to the newly allocated page in newpage.
833 */
834static int unmap_and_move(new_page_t get_new_page, unsigned long private,
835 struct page *page, int force, bool offlining,
836 enum migrate_mode mode)
837{
838 int rc = 0;
839 int *result = NULL;
840 struct page *newpage = get_new_page(page, private, &result);
841
842 if (!newpage)
843 return -ENOMEM;
844
845 if (page_count(page) == 1) {
846 /* page was freed from under us. So we are done. */
847 goto out;
848 }
849
850 if (unlikely(PageTransHuge(page)))
851 if (unlikely(split_huge_page(page)))
852 goto out;
853
854 rc = __unmap_and_move(page, newpage, force, offlining, mode);
855out:
856 if (rc != -EAGAIN) {
857 /*
858 * A page that has been migrated has all references
859 * removed and will be freed. A page that has not been
860 * migrated will have kepts its references and be
861 * restored.
862 */
863 list_del(&page->lru);
864 dec_zone_page_state(page, NR_ISOLATED_ANON +
865 page_is_file_cache(page));
866 putback_lru_page(page);
867 }
868 /*
869 * Move the new page to the LRU. If migration was not successful
870 * then this will free the page.
871 */
872 putback_lru_page(newpage);
873 if (result) {
874 if (rc)
875 *result = rc;
876 else
877 *result = page_to_nid(newpage);
878 }
879 return rc;
880}
881
882/*
883 * Counterpart of unmap_and_move_page() for hugepage migration.
884 *
885 * This function doesn't wait the completion of hugepage I/O
886 * because there is no race between I/O and migration for hugepage.
887 * Note that currently hugepage I/O occurs only in direct I/O
888 * where no lock is held and PG_writeback is irrelevant,
889 * and writeback status of all subpages are counted in the reference
890 * count of the head page (i.e. if all subpages of a 2MB hugepage are
891 * under direct I/O, the reference of the head page is 512 and a bit more.)
892 * This means that when we try to migrate hugepage whose subpages are
893 * doing direct I/O, some references remain after try_to_unmap() and
894 * hugepage migration fails without data corruption.
895 *
896 * There is also no race when direct I/O is issued on the page under migration,
897 * because then pte is replaced with migration swap entry and direct I/O code
898 * will wait in the page fault for migration to complete.
899 */
900static int unmap_and_move_huge_page(new_page_t get_new_page,
901 unsigned long private, struct page *hpage,
902 int force, bool offlining,
903 enum migrate_mode mode)
904{
905 int rc = 0;
906 int *result = NULL;
907 struct page *new_hpage = get_new_page(hpage, private, &result);
908 struct anon_vma *anon_vma = NULL;
909
910 if (!new_hpage)
911 return -ENOMEM;
912
913 rc = -EAGAIN;
914
915 if (!trylock_page(hpage)) {
916 if (!force || mode != MIGRATE_SYNC)
917 goto out;
918 lock_page(hpage);
919 }
920
921 if (PageAnon(hpage))
922 anon_vma = page_get_anon_vma(hpage);
923
924 try_to_unmap(hpage, TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
925
926 if (!page_mapped(hpage))
927 rc = move_to_new_page(new_hpage, hpage, 1, mode);
928
929 if (rc)
930 remove_migration_ptes(hpage, hpage);
931
932 if (anon_vma)
933 put_anon_vma(anon_vma);
934 unlock_page(hpage);
935
936out:
937 if (rc != -EAGAIN) {
938 list_del(&hpage->lru);
939 put_page(hpage);
940 }
941
942 put_page(new_hpage);
943
944 if (result) {
945 if (rc)
946 *result = rc;
947 else
948 *result = page_to_nid(new_hpage);
949 }
950 return rc;
951}
952
953/*
954 * migrate_pages
955 *
956 * The function takes one list of pages to migrate and a function
957 * that determines from the page to be migrated and the private data
958 * the target of the move and allocates the page.
959 *
960 * The function returns after 10 attempts or if no pages
961 * are movable anymore because to has become empty
962 * or no retryable pages exist anymore.
963 * Caller should call putback_lru_pages to return pages to the LRU
964 * or free list only if ret != 0.
965 *
966 * Return: Number of pages not migrated or error code.
967 */
968int migrate_pages(struct list_head *from,
969 new_page_t get_new_page, unsigned long private, bool offlining,
970 enum migrate_mode mode)
971{
972 int retry = 1;
973 int nr_failed = 0;
974 int pass = 0;
975 struct page *page;
976 struct page *page2;
977 int swapwrite = current->flags & PF_SWAPWRITE;
978 int rc;
979
980 if (!swapwrite)
981 current->flags |= PF_SWAPWRITE;
982
983 for(pass = 0; pass < 10 && retry; pass++) {
984 retry = 0;
985
986 list_for_each_entry_safe(page, page2, from, lru) {
987 cond_resched();
988
989 rc = unmap_and_move(get_new_page, private,
990 page, pass > 2, offlining,
991 mode);
992
993 switch(rc) {
994 case -ENOMEM:
995 goto out;
996 case -EAGAIN:
997 retry++;
998 break;
999 case 0:
1000 break;
1001 default:
1002 /* Permanent failure */
1003 nr_failed++;
1004 break;
1005 }
1006 }
1007 }
1008 rc = 0;
1009out:
1010 if (!swapwrite)
1011 current->flags &= ~PF_SWAPWRITE;
1012
1013 if (rc)
1014 return rc;
1015
1016 return nr_failed + retry;
1017}
1018
1019int migrate_huge_pages(struct list_head *from,
1020 new_page_t get_new_page, unsigned long private, bool offlining,
1021 enum migrate_mode mode)
1022{
1023 int retry = 1;
1024 int nr_failed = 0;
1025 int pass = 0;
1026 struct page *page;
1027 struct page *page2;
1028 int rc;
1029
1030 for (pass = 0; pass < 10 && retry; pass++) {
1031 retry = 0;
1032
1033 list_for_each_entry_safe(page, page2, from, lru) {
1034 cond_resched();
1035
1036 rc = unmap_and_move_huge_page(get_new_page,
1037 private, page, pass > 2, offlining,
1038 mode);
1039
1040 switch(rc) {
1041 case -ENOMEM:
1042 goto out;
1043 case -EAGAIN:
1044 retry++;
1045 break;
1046 case 0:
1047 break;
1048 default:
1049 /* Permanent failure */
1050 nr_failed++;
1051 break;
1052 }
1053 }
1054 }
1055 rc = 0;
1056out:
1057 if (rc)
1058 return rc;
1059
1060 return nr_failed + retry;
1061}
1062
1063#ifdef CONFIG_NUMA
1064/*
1065 * Move a list of individual pages
1066 */
1067struct page_to_node {
1068 unsigned long addr;
1069 struct page *page;
1070 int node;
1071 int status;
1072};
1073
1074static struct page *new_page_node(struct page *p, unsigned long private,
1075 int **result)
1076{
1077 struct page_to_node *pm = (struct page_to_node *)private;
1078
1079 while (pm->node != MAX_NUMNODES && pm->page != p)
1080 pm++;
1081
1082 if (pm->node == MAX_NUMNODES)
1083 return NULL;
1084
1085 *result = &pm->status;
1086
1087 return alloc_pages_exact_node(pm->node,
1088 GFP_HIGHUSER_MOVABLE | GFP_THISNODE, 0);
1089}
1090
1091/*
1092 * Move a set of pages as indicated in the pm array. The addr
1093 * field must be set to the virtual address of the page to be moved
1094 * and the node number must contain a valid target node.
1095 * The pm array ends with node = MAX_NUMNODES.
1096 */
1097static int do_move_page_to_node_array(struct mm_struct *mm,
1098 struct page_to_node *pm,
1099 int migrate_all)
1100{
1101 int err;
1102 struct page_to_node *pp;
1103 LIST_HEAD(pagelist);
1104
1105 down_read(&mm->mmap_sem);
1106
1107 /*
1108 * Build a list of pages to migrate
1109 */
1110 for (pp = pm; pp->node != MAX_NUMNODES; pp++) {
1111 struct vm_area_struct *vma;
1112 struct page *page;
1113
1114 err = -EFAULT;
1115 vma = find_vma(mm, pp->addr);
1116 if (!vma || pp->addr < vma->vm_start || !vma_migratable(vma))
1117 goto set_status;
1118
1119 page = follow_page(vma, pp->addr, FOLL_GET|FOLL_SPLIT);
1120
1121 err = PTR_ERR(page);
1122 if (IS_ERR(page))
1123 goto set_status;
1124
1125 err = -ENOENT;
1126 if (!page)
1127 goto set_status;
1128
1129 /* Use PageReserved to check for zero page */
1130 if (PageReserved(page) || PageKsm(page))
1131 goto put_and_set;
1132
1133 pp->page = page;
1134 err = page_to_nid(page);
1135
1136 if (err == pp->node)
1137 /*
1138 * Node already in the right place
1139 */
1140 goto put_and_set;
1141
1142 err = -EACCES;
1143 if (page_mapcount(page) > 1 &&
1144 !migrate_all)
1145 goto put_and_set;
1146
1147 err = isolate_lru_page(page);
1148 if (!err) {
1149 list_add_tail(&page->lru, &pagelist);
1150 inc_zone_page_state(page, NR_ISOLATED_ANON +
1151 page_is_file_cache(page));
1152 }
1153put_and_set:
1154 /*
1155 * Either remove the duplicate refcount from
1156 * isolate_lru_page() or drop the page ref if it was
1157 * not isolated.
1158 */
1159 put_page(page);
1160set_status:
1161 pp->status = err;
1162 }
1163
1164 err = 0;
1165 if (!list_empty(&pagelist)) {
1166 err = migrate_pages(&pagelist, new_page_node,
1167 (unsigned long)pm, 0, MIGRATE_SYNC);
1168 if (err)
1169 putback_lru_pages(&pagelist);
1170 }
1171
1172 up_read(&mm->mmap_sem);
1173 return err;
1174}
1175
1176/*
1177 * Migrate an array of page address onto an array of nodes and fill
1178 * the corresponding array of status.
1179 */
1180static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes,
1181 unsigned long nr_pages,
1182 const void __user * __user *pages,
1183 const int __user *nodes,
1184 int __user *status, int flags)
1185{
1186 struct page_to_node *pm;
1187 unsigned long chunk_nr_pages;
1188 unsigned long chunk_start;
1189 int err;
1190
1191 err = -ENOMEM;
1192 pm = (struct page_to_node *)__get_free_page(GFP_KERNEL);
1193 if (!pm)
1194 goto out;
1195
1196 migrate_prep();
1197
1198 /*
1199 * Store a chunk of page_to_node array in a page,
1200 * but keep the last one as a marker
1201 */
1202 chunk_nr_pages = (PAGE_SIZE / sizeof(struct page_to_node)) - 1;
1203
1204 for (chunk_start = 0;
1205 chunk_start < nr_pages;
1206 chunk_start += chunk_nr_pages) {
1207 int j;
1208
1209 if (chunk_start + chunk_nr_pages > nr_pages)
1210 chunk_nr_pages = nr_pages - chunk_start;
1211
1212 /* fill the chunk pm with addrs and nodes from user-space */
1213 for (j = 0; j < chunk_nr_pages; j++) {
1214 const void __user *p;
1215 int node;
1216
1217 err = -EFAULT;
1218 if (get_user(p, pages + j + chunk_start))
1219 goto out_pm;
1220 pm[j].addr = (unsigned long) p;
1221
1222 if (get_user(node, nodes + j + chunk_start))
1223 goto out_pm;
1224
1225 err = -ENODEV;
1226 if (node < 0 || node >= MAX_NUMNODES)
1227 goto out_pm;
1228
1229 if (!node_state(node, N_HIGH_MEMORY))
1230 goto out_pm;
1231
1232 err = -EACCES;
1233 if (!node_isset(node, task_nodes))
1234 goto out_pm;
1235
1236 pm[j].node = node;
1237 }
1238
1239 /* End marker for this chunk */
1240 pm[chunk_nr_pages].node = MAX_NUMNODES;
1241
1242 /* Migrate this chunk */
1243 err = do_move_page_to_node_array(mm, pm,
1244 flags & MPOL_MF_MOVE_ALL);
1245 if (err < 0)
1246 goto out_pm;
1247
1248 /* Return status information */
1249 for (j = 0; j < chunk_nr_pages; j++)
1250 if (put_user(pm[j].status, status + j + chunk_start)) {
1251 err = -EFAULT;
1252 goto out_pm;
1253 }
1254 }
1255 err = 0;
1256
1257out_pm:
1258 free_page((unsigned long)pm);
1259out:
1260 return err;
1261}
1262
1263/*
1264 * Determine the nodes of an array of pages and store it in an array of status.
1265 */
1266static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages,
1267 const void __user **pages, int *status)
1268{
1269 unsigned long i;
1270
1271 down_read(&mm->mmap_sem);
1272
1273 for (i = 0; i < nr_pages; i++) {
1274 unsigned long addr = (unsigned long)(*pages);
1275 struct vm_area_struct *vma;
1276 struct page *page;
1277 int err = -EFAULT;
1278
1279 vma = find_vma(mm, addr);
1280 if (!vma || addr < vma->vm_start)
1281 goto set_status;
1282
1283 page = follow_page(vma, addr, 0);
1284
1285 err = PTR_ERR(page);
1286 if (IS_ERR(page))
1287 goto set_status;
1288
1289 err = -ENOENT;
1290 /* Use PageReserved to check for zero page */
1291 if (!page || PageReserved(page) || PageKsm(page))
1292 goto set_status;
1293
1294 err = page_to_nid(page);
1295set_status:
1296 *status = err;
1297
1298 pages++;
1299 status++;
1300 }
1301
1302 up_read(&mm->mmap_sem);
1303}
1304
1305/*
1306 * Determine the nodes of a user array of pages and store it in
1307 * a user array of status.
1308 */
1309static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages,
1310 const void __user * __user *pages,
1311 int __user *status)
1312{
1313#define DO_PAGES_STAT_CHUNK_NR 16
1314 const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR];
1315 int chunk_status[DO_PAGES_STAT_CHUNK_NR];
1316
1317 while (nr_pages) {
1318 unsigned long chunk_nr;
1319
1320 chunk_nr = nr_pages;
1321 if (chunk_nr > DO_PAGES_STAT_CHUNK_NR)
1322 chunk_nr = DO_PAGES_STAT_CHUNK_NR;
1323
1324 if (copy_from_user(chunk_pages, pages, chunk_nr * sizeof(*chunk_pages)))
1325 break;
1326
1327 do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status);
1328
1329 if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status)))
1330 break;
1331
1332 pages += chunk_nr;
1333 status += chunk_nr;
1334 nr_pages -= chunk_nr;
1335 }
1336 return nr_pages ? -EFAULT : 0;
1337}
1338
1339/*
1340 * Move a list of pages in the address space of the currently executing
1341 * process.
1342 */
1343SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages,
1344 const void __user * __user *, pages,
1345 const int __user *, nodes,
1346 int __user *, status, int, flags)
1347{
1348 const struct cred *cred = current_cred(), *tcred;
1349 struct task_struct *task;
1350 struct mm_struct *mm;
1351 int err;
1352 nodemask_t task_nodes;
1353
1354 /* Check flags */
1355 if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
1356 return -EINVAL;
1357
1358 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1359 return -EPERM;
1360
1361 /* Find the mm_struct */
1362 rcu_read_lock();
1363 task = pid ? find_task_by_vpid(pid) : current;
1364 if (!task) {
1365 rcu_read_unlock();
1366 return -ESRCH;
1367 }
1368 get_task_struct(task);
1369
1370 /*
1371 * Check if this process has the right to modify the specified
1372 * process. The right exists if the process has administrative
1373 * capabilities, superuser privileges or the same
1374 * userid as the target process.
1375 */
1376 tcred = __task_cred(task);
1377 if (!uid_eq(cred->euid, tcred->suid) && !uid_eq(cred->euid, tcred->uid) &&
1378 !uid_eq(cred->uid, tcred->suid) && !uid_eq(cred->uid, tcred->uid) &&
1379 !capable(CAP_SYS_NICE)) {
1380 rcu_read_unlock();
1381 err = -EPERM;
1382 goto out;
1383 }
1384 rcu_read_unlock();
1385
1386 err = security_task_movememory(task);
1387 if (err)
1388 goto out;
1389
1390 task_nodes = cpuset_mems_allowed(task);
1391 mm = get_task_mm(task);
1392 put_task_struct(task);
1393
1394 if (!mm)
1395 return -EINVAL;
1396
1397 if (nodes)
1398 err = do_pages_move(mm, task_nodes, nr_pages, pages,
1399 nodes, status, flags);
1400 else
1401 err = do_pages_stat(mm, nr_pages, pages, status);
1402
1403 mmput(mm);
1404 return err;
1405
1406out:
1407 put_task_struct(task);
1408 return err;
1409}
1410
1411/*
1412 * Call migration functions in the vma_ops that may prepare
1413 * memory in a vm for migration. migration functions may perform
1414 * the migration for vmas that do not have an underlying page struct.
1415 */
1416int migrate_vmas(struct mm_struct *mm, const nodemask_t *to,
1417 const nodemask_t *from, unsigned long flags)
1418{
1419 struct vm_area_struct *vma;
1420 int err = 0;
1421
1422 for (vma = mm->mmap; vma && !err; vma = vma->vm_next) {
1423 if (vma->vm_ops && vma->vm_ops->migrate) {
1424 err = vma->vm_ops->migrate(vma, to, from, flags);
1425 if (err)
1426 break;
1427 }
1428 }
1429 return err;
1430}
1431#endif