]> git.proxmox.com Git - mirror_ubuntu-artful-kernel.git/blob - mm/migrate.c
Merge branch 'for-linus-4.5' of git://git.kernel.org/pub/scm/linux/kernel/git/mason...
[mirror_ubuntu-artful-kernel.git] / mm / migrate.c
1 /*
2 * Memory Migration functionality - linux/mm/migrate.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/backing-dev.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>
39 #include <linux/mmu_notifier.h>
40 #include <linux/page_idle.h>
41
42 #include <asm/tlbflush.h>
43
44 #define CREATE_TRACE_POINTS
45 #include <trace/events/migrate.h>
46
47 #include "internal.h"
48
49 /*
50 * migrate_prep() needs to be called before we start compiling a list of pages
51 * to be migrated using isolate_lru_page(). If scheduling work on other CPUs is
52 * undesirable, use migrate_prep_local()
53 */
54 int migrate_prep(void)
55 {
56 /*
57 * Clear the LRU lists so pages can be isolated.
58 * Note that pages may be moved off the LRU after we have
59 * drained them. Those pages will fail to migrate like other
60 * pages that may be busy.
61 */
62 lru_add_drain_all();
63
64 return 0;
65 }
66
67 /* Do the necessary work of migrate_prep but not if it involves other CPUs */
68 int migrate_prep_local(void)
69 {
70 lru_add_drain();
71
72 return 0;
73 }
74
75 /*
76 * Put previously isolated pages back onto the appropriate lists
77 * from where they were once taken off for compaction/migration.
78 *
79 * This function shall be used whenever the isolated pageset has been
80 * built from lru, balloon, hugetlbfs page. See isolate_migratepages_range()
81 * and isolate_huge_page().
82 */
83 void putback_movable_pages(struct list_head *l)
84 {
85 struct page *page;
86 struct page *page2;
87
88 list_for_each_entry_safe(page, page2, l, lru) {
89 if (unlikely(PageHuge(page))) {
90 putback_active_hugepage(page);
91 continue;
92 }
93 list_del(&page->lru);
94 dec_zone_page_state(page, NR_ISOLATED_ANON +
95 page_is_file_cache(page));
96 if (unlikely(isolated_balloon_page(page)))
97 balloon_page_putback(page);
98 else
99 putback_lru_page(page);
100 }
101 }
102
103 /*
104 * Restore a potential migration pte to a working pte entry
105 */
106 static int remove_migration_pte(struct page *new, struct vm_area_struct *vma,
107 unsigned long addr, void *old)
108 {
109 struct mm_struct *mm = vma->vm_mm;
110 swp_entry_t entry;
111 pmd_t *pmd;
112 pte_t *ptep, pte;
113 spinlock_t *ptl;
114
115 if (unlikely(PageHuge(new))) {
116 ptep = huge_pte_offset(mm, addr);
117 if (!ptep)
118 goto out;
119 ptl = huge_pte_lockptr(hstate_vma(vma), mm, ptep);
120 } else {
121 pmd = mm_find_pmd(mm, addr);
122 if (!pmd)
123 goto out;
124
125 ptep = pte_offset_map(pmd, addr);
126
127 /*
128 * Peek to check is_swap_pte() before taking ptlock? No, we
129 * can race mremap's move_ptes(), which skips anon_vma lock.
130 */
131
132 ptl = pte_lockptr(mm, pmd);
133 }
134
135 spin_lock(ptl);
136 pte = *ptep;
137 if (!is_swap_pte(pte))
138 goto unlock;
139
140 entry = pte_to_swp_entry(pte);
141
142 if (!is_migration_entry(entry) ||
143 migration_entry_to_page(entry) != old)
144 goto unlock;
145
146 get_page(new);
147 pte = pte_mkold(mk_pte(new, vma->vm_page_prot));
148 if (pte_swp_soft_dirty(*ptep))
149 pte = pte_mksoft_dirty(pte);
150
151 /* Recheck VMA as permissions can change since migration started */
152 if (is_write_migration_entry(entry))
153 pte = maybe_mkwrite(pte, vma);
154
155 #ifdef CONFIG_HUGETLB_PAGE
156 if (PageHuge(new)) {
157 pte = pte_mkhuge(pte);
158 pte = arch_make_huge_pte(pte, vma, new, 0);
159 }
160 #endif
161 flush_dcache_page(new);
162 set_pte_at(mm, addr, ptep, pte);
163
164 if (PageHuge(new)) {
165 if (PageAnon(new))
166 hugepage_add_anon_rmap(new, vma, addr);
167 else
168 page_dup_rmap(new, true);
169 } else if (PageAnon(new))
170 page_add_anon_rmap(new, vma, addr, false);
171 else
172 page_add_file_rmap(new);
173
174 if (vma->vm_flags & VM_LOCKED)
175 mlock_vma_page(new);
176
177 /* No need to invalidate - it was non-present before */
178 update_mmu_cache(vma, addr, ptep);
179 unlock:
180 pte_unmap_unlock(ptep, ptl);
181 out:
182 return SWAP_AGAIN;
183 }
184
185 /*
186 * Get rid of all migration entries and replace them by
187 * references to the indicated page.
188 */
189 static void remove_migration_ptes(struct page *old, struct page *new)
190 {
191 struct rmap_walk_control rwc = {
192 .rmap_one = remove_migration_pte,
193 .arg = old,
194 };
195
196 rmap_walk(new, &rwc);
197 }
198
199 /*
200 * Something used the pte of a page under migration. We need to
201 * get to the page and wait until migration is finished.
202 * When we return from this function the fault will be retried.
203 */
204 void __migration_entry_wait(struct mm_struct *mm, pte_t *ptep,
205 spinlock_t *ptl)
206 {
207 pte_t pte;
208 swp_entry_t entry;
209 struct page *page;
210
211 spin_lock(ptl);
212 pte = *ptep;
213 if (!is_swap_pte(pte))
214 goto out;
215
216 entry = pte_to_swp_entry(pte);
217 if (!is_migration_entry(entry))
218 goto out;
219
220 page = migration_entry_to_page(entry);
221
222 /*
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
227 * will occur again.
228 */
229 if (!get_page_unless_zero(page))
230 goto out;
231 pte_unmap_unlock(ptep, ptl);
232 wait_on_page_locked(page);
233 put_page(page);
234 return;
235 out:
236 pte_unmap_unlock(ptep, ptl);
237 }
238
239 void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
240 unsigned long address)
241 {
242 spinlock_t *ptl = pte_lockptr(mm, pmd);
243 pte_t *ptep = pte_offset_map(pmd, address);
244 __migration_entry_wait(mm, ptep, ptl);
245 }
246
247 void migration_entry_wait_huge(struct vm_area_struct *vma,
248 struct mm_struct *mm, pte_t *pte)
249 {
250 spinlock_t *ptl = huge_pte_lockptr(hstate_vma(vma), mm, pte);
251 __migration_entry_wait(mm, pte, ptl);
252 }
253
254 #ifdef CONFIG_BLOCK
255 /* Returns true if all buffers are successfully locked */
256 static bool buffer_migrate_lock_buffers(struct buffer_head *head,
257 enum migrate_mode mode)
258 {
259 struct buffer_head *bh = head;
260
261 /* Simple case, sync compaction */
262 if (mode != MIGRATE_ASYNC) {
263 do {
264 get_bh(bh);
265 lock_buffer(bh);
266 bh = bh->b_this_page;
267
268 } while (bh != head);
269
270 return true;
271 }
272
273 /* async case, we cannot block on lock_buffer so use trylock_buffer */
274 do {
275 get_bh(bh);
276 if (!trylock_buffer(bh)) {
277 /*
278 * We failed to lock the buffer and cannot stall in
279 * async migration. Release the taken locks
280 */
281 struct buffer_head *failed_bh = bh;
282 put_bh(failed_bh);
283 bh = head;
284 while (bh != failed_bh) {
285 unlock_buffer(bh);
286 put_bh(bh);
287 bh = bh->b_this_page;
288 }
289 return false;
290 }
291
292 bh = bh->b_this_page;
293 } while (bh != head);
294 return true;
295 }
296 #else
297 static inline bool buffer_migrate_lock_buffers(struct buffer_head *head,
298 enum migrate_mode mode)
299 {
300 return true;
301 }
302 #endif /* CONFIG_BLOCK */
303
304 /*
305 * Replace the page in the mapping.
306 *
307 * The number of remaining references must be:
308 * 1 for anonymous pages without a mapping
309 * 2 for pages with a mapping
310 * 3 for pages with a mapping and PagePrivate/PagePrivate2 set.
311 */
312 int migrate_page_move_mapping(struct address_space *mapping,
313 struct page *newpage, struct page *page,
314 struct buffer_head *head, enum migrate_mode mode,
315 int extra_count)
316 {
317 struct zone *oldzone, *newzone;
318 int dirty;
319 int expected_count = 1 + extra_count;
320 void **pslot;
321
322 if (!mapping) {
323 /* Anonymous page without mapping */
324 if (page_count(page) != expected_count)
325 return -EAGAIN;
326
327 /* No turning back from here */
328 set_page_memcg(newpage, page_memcg(page));
329 newpage->index = page->index;
330 newpage->mapping = page->mapping;
331 if (PageSwapBacked(page))
332 SetPageSwapBacked(newpage);
333
334 return MIGRATEPAGE_SUCCESS;
335 }
336
337 oldzone = page_zone(page);
338 newzone = page_zone(newpage);
339
340 spin_lock_irq(&mapping->tree_lock);
341
342 pslot = radix_tree_lookup_slot(&mapping->page_tree,
343 page_index(page));
344
345 expected_count += 1 + page_has_private(page);
346 if (page_count(page) != expected_count ||
347 radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) {
348 spin_unlock_irq(&mapping->tree_lock);
349 return -EAGAIN;
350 }
351
352 if (!page_freeze_refs(page, expected_count)) {
353 spin_unlock_irq(&mapping->tree_lock);
354 return -EAGAIN;
355 }
356
357 /*
358 * In the async migration case of moving a page with buffers, lock the
359 * buffers using trylock before the mapping is moved. If the mapping
360 * was moved, we later failed to lock the buffers and could not move
361 * the mapping back due to an elevated page count, we would have to
362 * block waiting on other references to be dropped.
363 */
364 if (mode == MIGRATE_ASYNC && head &&
365 !buffer_migrate_lock_buffers(head, mode)) {
366 page_unfreeze_refs(page, expected_count);
367 spin_unlock_irq(&mapping->tree_lock);
368 return -EAGAIN;
369 }
370
371 /*
372 * Now we know that no one else is looking at the page:
373 * no turning back from here.
374 */
375 set_page_memcg(newpage, page_memcg(page));
376 newpage->index = page->index;
377 newpage->mapping = page->mapping;
378 if (PageSwapBacked(page))
379 SetPageSwapBacked(newpage);
380
381 get_page(newpage); /* add cache reference */
382 if (PageSwapCache(page)) {
383 SetPageSwapCache(newpage);
384 set_page_private(newpage, page_private(page));
385 }
386
387 /* Move dirty while page refs frozen and newpage not yet exposed */
388 dirty = PageDirty(page);
389 if (dirty) {
390 ClearPageDirty(page);
391 SetPageDirty(newpage);
392 }
393
394 radix_tree_replace_slot(pslot, newpage);
395
396 /*
397 * Drop cache reference from old page by unfreezing
398 * to one less reference.
399 * We know this isn't the last reference.
400 */
401 page_unfreeze_refs(page, expected_count - 1);
402
403 spin_unlock(&mapping->tree_lock);
404 /* Leave irq disabled to prevent preemption while updating stats */
405
406 /*
407 * If moved to a different zone then also account
408 * the page for that zone. Other VM counters will be
409 * taken care of when we establish references to the
410 * new page and drop references to the old page.
411 *
412 * Note that anonymous pages are accounted for
413 * via NR_FILE_PAGES and NR_ANON_PAGES if they
414 * are mapped to swap space.
415 */
416 if (newzone != oldzone) {
417 __dec_zone_state(oldzone, NR_FILE_PAGES);
418 __inc_zone_state(newzone, NR_FILE_PAGES);
419 if (PageSwapBacked(page) && !PageSwapCache(page)) {
420 __dec_zone_state(oldzone, NR_SHMEM);
421 __inc_zone_state(newzone, NR_SHMEM);
422 }
423 if (dirty && mapping_cap_account_dirty(mapping)) {
424 __dec_zone_state(oldzone, NR_FILE_DIRTY);
425 __inc_zone_state(newzone, NR_FILE_DIRTY);
426 }
427 }
428 local_irq_enable();
429
430 return MIGRATEPAGE_SUCCESS;
431 }
432
433 /*
434 * The expected number of remaining references is the same as that
435 * of migrate_page_move_mapping().
436 */
437 int migrate_huge_page_move_mapping(struct address_space *mapping,
438 struct page *newpage, struct page *page)
439 {
440 int expected_count;
441 void **pslot;
442
443 spin_lock_irq(&mapping->tree_lock);
444
445 pslot = radix_tree_lookup_slot(&mapping->page_tree,
446 page_index(page));
447
448 expected_count = 2 + page_has_private(page);
449 if (page_count(page) != expected_count ||
450 radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) {
451 spin_unlock_irq(&mapping->tree_lock);
452 return -EAGAIN;
453 }
454
455 if (!page_freeze_refs(page, expected_count)) {
456 spin_unlock_irq(&mapping->tree_lock);
457 return -EAGAIN;
458 }
459
460 set_page_memcg(newpage, page_memcg(page));
461 newpage->index = page->index;
462 newpage->mapping = page->mapping;
463 get_page(newpage);
464
465 radix_tree_replace_slot(pslot, newpage);
466
467 page_unfreeze_refs(page, expected_count - 1);
468
469 spin_unlock_irq(&mapping->tree_lock);
470 return MIGRATEPAGE_SUCCESS;
471 }
472
473 /*
474 * Gigantic pages are so large that we do not guarantee that page++ pointer
475 * arithmetic will work across the entire page. We need something more
476 * specialized.
477 */
478 static void __copy_gigantic_page(struct page *dst, struct page *src,
479 int nr_pages)
480 {
481 int i;
482 struct page *dst_base = dst;
483 struct page *src_base = src;
484
485 for (i = 0; i < nr_pages; ) {
486 cond_resched();
487 copy_highpage(dst, src);
488
489 i++;
490 dst = mem_map_next(dst, dst_base, i);
491 src = mem_map_next(src, src_base, i);
492 }
493 }
494
495 static void copy_huge_page(struct page *dst, struct page *src)
496 {
497 int i;
498 int nr_pages;
499
500 if (PageHuge(src)) {
501 /* hugetlbfs page */
502 struct hstate *h = page_hstate(src);
503 nr_pages = pages_per_huge_page(h);
504
505 if (unlikely(nr_pages > MAX_ORDER_NR_PAGES)) {
506 __copy_gigantic_page(dst, src, nr_pages);
507 return;
508 }
509 } else {
510 /* thp page */
511 BUG_ON(!PageTransHuge(src));
512 nr_pages = hpage_nr_pages(src);
513 }
514
515 for (i = 0; i < nr_pages; i++) {
516 cond_resched();
517 copy_highpage(dst + i, src + i);
518 }
519 }
520
521 /*
522 * Copy the page to its new location
523 */
524 void migrate_page_copy(struct page *newpage, struct page *page)
525 {
526 int cpupid;
527
528 if (PageHuge(page) || PageTransHuge(page))
529 copy_huge_page(newpage, page);
530 else
531 copy_highpage(newpage, page);
532
533 if (PageError(page))
534 SetPageError(newpage);
535 if (PageReferenced(page))
536 SetPageReferenced(newpage);
537 if (PageUptodate(page))
538 SetPageUptodate(newpage);
539 if (TestClearPageActive(page)) {
540 VM_BUG_ON_PAGE(PageUnevictable(page), page);
541 SetPageActive(newpage);
542 } else if (TestClearPageUnevictable(page))
543 SetPageUnevictable(newpage);
544 if (PageChecked(page))
545 SetPageChecked(newpage);
546 if (PageMappedToDisk(page))
547 SetPageMappedToDisk(newpage);
548
549 /* Move dirty on pages not done by migrate_page_move_mapping() */
550 if (PageDirty(page))
551 SetPageDirty(newpage);
552
553 if (page_is_young(page))
554 set_page_young(newpage);
555 if (page_is_idle(page))
556 set_page_idle(newpage);
557
558 /*
559 * Copy NUMA information to the new page, to prevent over-eager
560 * future migrations of this same page.
561 */
562 cpupid = page_cpupid_xchg_last(page, -1);
563 page_cpupid_xchg_last(newpage, cpupid);
564
565 ksm_migrate_page(newpage, page);
566 /*
567 * Please do not reorder this without considering how mm/ksm.c's
568 * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache().
569 */
570 if (PageSwapCache(page))
571 ClearPageSwapCache(page);
572 ClearPagePrivate(page);
573 set_page_private(page, 0);
574
575 /*
576 * If any waiters have accumulated on the new page then
577 * wake them up.
578 */
579 if (PageWriteback(newpage))
580 end_page_writeback(newpage);
581 }
582
583 /************************************************************
584 * Migration functions
585 ***********************************************************/
586
587 /*
588 * Common logic to directly migrate a single page suitable for
589 * pages that do not use PagePrivate/PagePrivate2.
590 *
591 * Pages are locked upon entry and exit.
592 */
593 int migrate_page(struct address_space *mapping,
594 struct page *newpage, struct page *page,
595 enum migrate_mode mode)
596 {
597 int rc;
598
599 BUG_ON(PageWriteback(page)); /* Writeback must be complete */
600
601 rc = migrate_page_move_mapping(mapping, newpage, page, NULL, mode, 0);
602
603 if (rc != MIGRATEPAGE_SUCCESS)
604 return rc;
605
606 migrate_page_copy(newpage, page);
607 return MIGRATEPAGE_SUCCESS;
608 }
609 EXPORT_SYMBOL(migrate_page);
610
611 #ifdef CONFIG_BLOCK
612 /*
613 * Migration function for pages with buffers. This function can only be used
614 * if the underlying filesystem guarantees that no other references to "page"
615 * exist.
616 */
617 int buffer_migrate_page(struct address_space *mapping,
618 struct page *newpage, struct page *page, enum migrate_mode mode)
619 {
620 struct buffer_head *bh, *head;
621 int rc;
622
623 if (!page_has_buffers(page))
624 return migrate_page(mapping, newpage, page, mode);
625
626 head = page_buffers(page);
627
628 rc = migrate_page_move_mapping(mapping, newpage, page, head, mode, 0);
629
630 if (rc != MIGRATEPAGE_SUCCESS)
631 return rc;
632
633 /*
634 * In the async case, migrate_page_move_mapping locked the buffers
635 * with an IRQ-safe spinlock held. In the sync case, the buffers
636 * need to be locked now
637 */
638 if (mode != MIGRATE_ASYNC)
639 BUG_ON(!buffer_migrate_lock_buffers(head, mode));
640
641 ClearPagePrivate(page);
642 set_page_private(newpage, page_private(page));
643 set_page_private(page, 0);
644 put_page(page);
645 get_page(newpage);
646
647 bh = head;
648 do {
649 set_bh_page(bh, newpage, bh_offset(bh));
650 bh = bh->b_this_page;
651
652 } while (bh != head);
653
654 SetPagePrivate(newpage);
655
656 migrate_page_copy(newpage, page);
657
658 bh = head;
659 do {
660 unlock_buffer(bh);
661 put_bh(bh);
662 bh = bh->b_this_page;
663
664 } while (bh != head);
665
666 return MIGRATEPAGE_SUCCESS;
667 }
668 EXPORT_SYMBOL(buffer_migrate_page);
669 #endif
670
671 /*
672 * Writeback a page to clean the dirty state
673 */
674 static int writeout(struct address_space *mapping, struct page *page)
675 {
676 struct writeback_control wbc = {
677 .sync_mode = WB_SYNC_NONE,
678 .nr_to_write = 1,
679 .range_start = 0,
680 .range_end = LLONG_MAX,
681 .for_reclaim = 1
682 };
683 int rc;
684
685 if (!mapping->a_ops->writepage)
686 /* No write method for the address space */
687 return -EINVAL;
688
689 if (!clear_page_dirty_for_io(page))
690 /* Someone else already triggered a write */
691 return -EAGAIN;
692
693 /*
694 * A dirty page may imply that the underlying filesystem has
695 * the page on some queue. So the page must be clean for
696 * migration. Writeout may mean we loose the lock and the
697 * page state is no longer what we checked for earlier.
698 * At this point we know that the migration attempt cannot
699 * be successful.
700 */
701 remove_migration_ptes(page, page);
702
703 rc = mapping->a_ops->writepage(page, &wbc);
704
705 if (rc != AOP_WRITEPAGE_ACTIVATE)
706 /* unlocked. Relock */
707 lock_page(page);
708
709 return (rc < 0) ? -EIO : -EAGAIN;
710 }
711
712 /*
713 * Default handling if a filesystem does not provide a migration function.
714 */
715 static int fallback_migrate_page(struct address_space *mapping,
716 struct page *newpage, struct page *page, enum migrate_mode mode)
717 {
718 if (PageDirty(page)) {
719 /* Only writeback pages in full synchronous migration */
720 if (mode != MIGRATE_SYNC)
721 return -EBUSY;
722 return writeout(mapping, page);
723 }
724
725 /*
726 * Buffers may be managed in a filesystem specific way.
727 * We must have no buffers or drop them.
728 */
729 if (page_has_private(page) &&
730 !try_to_release_page(page, GFP_KERNEL))
731 return -EAGAIN;
732
733 return migrate_page(mapping, newpage, page, mode);
734 }
735
736 /*
737 * Move a page to a newly allocated page
738 * The page is locked and all ptes have been successfully removed.
739 *
740 * The new page will have replaced the old page if this function
741 * is successful.
742 *
743 * Return value:
744 * < 0 - error code
745 * MIGRATEPAGE_SUCCESS - success
746 */
747 static int move_to_new_page(struct page *newpage, struct page *page,
748 enum migrate_mode mode)
749 {
750 struct address_space *mapping;
751 int rc;
752
753 VM_BUG_ON_PAGE(!PageLocked(page), page);
754 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
755
756 mapping = page_mapping(page);
757 if (!mapping)
758 rc = migrate_page(mapping, newpage, page, mode);
759 else if (mapping->a_ops->migratepage)
760 /*
761 * Most pages have a mapping and most filesystems provide a
762 * migratepage callback. Anonymous pages are part of swap
763 * space which also has its own migratepage callback. This
764 * is the most common path for page migration.
765 */
766 rc = mapping->a_ops->migratepage(mapping, newpage, page, mode);
767 else
768 rc = fallback_migrate_page(mapping, newpage, page, mode);
769
770 /*
771 * When successful, old pagecache page->mapping must be cleared before
772 * page is freed; but stats require that PageAnon be left as PageAnon.
773 */
774 if (rc == MIGRATEPAGE_SUCCESS) {
775 set_page_memcg(page, NULL);
776 if (!PageAnon(page))
777 page->mapping = NULL;
778 }
779 return rc;
780 }
781
782 static int __unmap_and_move(struct page *page, struct page *newpage,
783 int force, enum migrate_mode mode)
784 {
785 int rc = -EAGAIN;
786 int page_was_mapped = 0;
787 struct anon_vma *anon_vma = NULL;
788
789 if (!trylock_page(page)) {
790 if (!force || mode == MIGRATE_ASYNC)
791 goto out;
792
793 /*
794 * It's not safe for direct compaction to call lock_page.
795 * For example, during page readahead pages are added locked
796 * to the LRU. Later, when the IO completes the pages are
797 * marked uptodate and unlocked. However, the queueing
798 * could be merging multiple pages for one bio (e.g.
799 * mpage_readpages). If an allocation happens for the
800 * second or third page, the process can end up locking
801 * the same page twice and deadlocking. Rather than
802 * trying to be clever about what pages can be locked,
803 * avoid the use of lock_page for direct compaction
804 * altogether.
805 */
806 if (current->flags & PF_MEMALLOC)
807 goto out;
808
809 lock_page(page);
810 }
811
812 if (PageWriteback(page)) {
813 /*
814 * Only in the case of a full synchronous migration is it
815 * necessary to wait for PageWriteback. In the async case,
816 * the retry loop is too short and in the sync-light case,
817 * the overhead of stalling is too much
818 */
819 if (mode != MIGRATE_SYNC) {
820 rc = -EBUSY;
821 goto out_unlock;
822 }
823 if (!force)
824 goto out_unlock;
825 wait_on_page_writeback(page);
826 }
827
828 /*
829 * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
830 * we cannot notice that anon_vma is freed while we migrates a page.
831 * This get_anon_vma() delays freeing anon_vma pointer until the end
832 * of migration. File cache pages are no problem because of page_lock()
833 * File Caches may use write_page() or lock_page() in migration, then,
834 * just care Anon page here.
835 *
836 * Only page_get_anon_vma() understands the subtleties of
837 * getting a hold on an anon_vma from outside one of its mms.
838 * But if we cannot get anon_vma, then we won't need it anyway,
839 * because that implies that the anon page is no longer mapped
840 * (and cannot be remapped so long as we hold the page lock).
841 */
842 if (PageAnon(page) && !PageKsm(page))
843 anon_vma = page_get_anon_vma(page);
844
845 /*
846 * Block others from accessing the new page when we get around to
847 * establishing additional references. We are usually the only one
848 * holding a reference to newpage at this point. We used to have a BUG
849 * here if trylock_page(newpage) fails, but would like to allow for
850 * cases where there might be a race with the previous use of newpage.
851 * This is much like races on refcount of oldpage: just don't BUG().
852 */
853 if (unlikely(!trylock_page(newpage)))
854 goto out_unlock;
855
856 if (unlikely(isolated_balloon_page(page))) {
857 /*
858 * A ballooned page does not need any special attention from
859 * physical to virtual reverse mapping procedures.
860 * Skip any attempt to unmap PTEs or to remap swap cache,
861 * in order to avoid burning cycles at rmap level, and perform
862 * the page migration right away (proteced by page lock).
863 */
864 rc = balloon_page_migrate(newpage, page, mode);
865 goto out_unlock_both;
866 }
867
868 /*
869 * Corner case handling:
870 * 1. When a new swap-cache page is read into, it is added to the LRU
871 * and treated as swapcache but it has no rmap yet.
872 * Calling try_to_unmap() against a page->mapping==NULL page will
873 * trigger a BUG. So handle it here.
874 * 2. An orphaned page (see truncate_complete_page) might have
875 * fs-private metadata. The page can be picked up due to memory
876 * offlining. Everywhere else except page reclaim, the page is
877 * invisible to the vm, so the page can not be migrated. So try to
878 * free the metadata, so the page can be freed.
879 */
880 if (!page->mapping) {
881 VM_BUG_ON_PAGE(PageAnon(page), page);
882 if (page_has_private(page)) {
883 try_to_free_buffers(page);
884 goto out_unlock_both;
885 }
886 } else if (page_mapped(page)) {
887 /* Establish migration ptes */
888 VM_BUG_ON_PAGE(PageAnon(page) && !PageKsm(page) && !anon_vma,
889 page);
890 try_to_unmap(page,
891 TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
892 page_was_mapped = 1;
893 }
894
895 if (!page_mapped(page))
896 rc = move_to_new_page(newpage, page, mode);
897
898 if (page_was_mapped)
899 remove_migration_ptes(page,
900 rc == MIGRATEPAGE_SUCCESS ? newpage : page);
901
902 out_unlock_both:
903 unlock_page(newpage);
904 out_unlock:
905 /* Drop an anon_vma reference if we took one */
906 if (anon_vma)
907 put_anon_vma(anon_vma);
908 unlock_page(page);
909 out:
910 return rc;
911 }
912
913 /*
914 * gcc 4.7 and 4.8 on arm get an ICEs when inlining unmap_and_move(). Work
915 * around it.
916 */
917 #if (GCC_VERSION >= 40700 && GCC_VERSION < 40900) && defined(CONFIG_ARM)
918 #define ICE_noinline noinline
919 #else
920 #define ICE_noinline
921 #endif
922
923 /*
924 * Obtain the lock on page, remove all ptes and migrate the page
925 * to the newly allocated page in newpage.
926 */
927 static ICE_noinline int unmap_and_move(new_page_t get_new_page,
928 free_page_t put_new_page,
929 unsigned long private, struct page *page,
930 int force, enum migrate_mode mode,
931 enum migrate_reason reason)
932 {
933 int rc = MIGRATEPAGE_SUCCESS;
934 int *result = NULL;
935 struct page *newpage;
936
937 newpage = get_new_page(page, private, &result);
938 if (!newpage)
939 return -ENOMEM;
940
941 if (page_count(page) == 1) {
942 /* page was freed from under us. So we are done. */
943 goto out;
944 }
945
946 if (unlikely(PageTransHuge(page))) {
947 lock_page(page);
948 rc = split_huge_page(page);
949 unlock_page(page);
950 if (rc)
951 goto out;
952 }
953
954 rc = __unmap_and_move(page, newpage, force, mode);
955 if (rc == MIGRATEPAGE_SUCCESS)
956 put_new_page = NULL;
957
958 out:
959 if (rc != -EAGAIN) {
960 /*
961 * A page that has been migrated has all references
962 * removed and will be freed. A page that has not been
963 * migrated will have kepts its references and be
964 * restored.
965 */
966 list_del(&page->lru);
967 dec_zone_page_state(page, NR_ISOLATED_ANON +
968 page_is_file_cache(page));
969 /* Soft-offlined page shouldn't go through lru cache list */
970 if (reason == MR_MEMORY_FAILURE) {
971 put_page(page);
972 if (!test_set_page_hwpoison(page))
973 num_poisoned_pages_inc();
974 } else
975 putback_lru_page(page);
976 }
977
978 /*
979 * If migration was not successful and there's a freeing callback, use
980 * it. Otherwise, putback_lru_page() will drop the reference grabbed
981 * during isolation.
982 */
983 if (put_new_page)
984 put_new_page(newpage, private);
985 else if (unlikely(__is_movable_balloon_page(newpage))) {
986 /* drop our reference, page already in the balloon */
987 put_page(newpage);
988 } else
989 putback_lru_page(newpage);
990
991 if (result) {
992 if (rc)
993 *result = rc;
994 else
995 *result = page_to_nid(newpage);
996 }
997 return rc;
998 }
999
1000 /*
1001 * Counterpart of unmap_and_move_page() for hugepage migration.
1002 *
1003 * This function doesn't wait the completion of hugepage I/O
1004 * because there is no race between I/O and migration for hugepage.
1005 * Note that currently hugepage I/O occurs only in direct I/O
1006 * where no lock is held and PG_writeback is irrelevant,
1007 * and writeback status of all subpages are counted in the reference
1008 * count of the head page (i.e. if all subpages of a 2MB hugepage are
1009 * under direct I/O, the reference of the head page is 512 and a bit more.)
1010 * This means that when we try to migrate hugepage whose subpages are
1011 * doing direct I/O, some references remain after try_to_unmap() and
1012 * hugepage migration fails without data corruption.
1013 *
1014 * There is also no race when direct I/O is issued on the page under migration,
1015 * because then pte is replaced with migration swap entry and direct I/O code
1016 * will wait in the page fault for migration to complete.
1017 */
1018 static int unmap_and_move_huge_page(new_page_t get_new_page,
1019 free_page_t put_new_page, unsigned long private,
1020 struct page *hpage, int force,
1021 enum migrate_mode mode)
1022 {
1023 int rc = -EAGAIN;
1024 int *result = NULL;
1025 int page_was_mapped = 0;
1026 struct page *new_hpage;
1027 struct anon_vma *anon_vma = NULL;
1028
1029 /*
1030 * Movability of hugepages depends on architectures and hugepage size.
1031 * This check is necessary because some callers of hugepage migration
1032 * like soft offline and memory hotremove don't walk through page
1033 * tables or check whether the hugepage is pmd-based or not before
1034 * kicking migration.
1035 */
1036 if (!hugepage_migration_supported(page_hstate(hpage))) {
1037 putback_active_hugepage(hpage);
1038 return -ENOSYS;
1039 }
1040
1041 new_hpage = get_new_page(hpage, private, &result);
1042 if (!new_hpage)
1043 return -ENOMEM;
1044
1045 if (!trylock_page(hpage)) {
1046 if (!force || mode != MIGRATE_SYNC)
1047 goto out;
1048 lock_page(hpage);
1049 }
1050
1051 if (PageAnon(hpage))
1052 anon_vma = page_get_anon_vma(hpage);
1053
1054 if (unlikely(!trylock_page(new_hpage)))
1055 goto put_anon;
1056
1057 if (page_mapped(hpage)) {
1058 try_to_unmap(hpage,
1059 TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
1060 page_was_mapped = 1;
1061 }
1062
1063 if (!page_mapped(hpage))
1064 rc = move_to_new_page(new_hpage, hpage, mode);
1065
1066 if (page_was_mapped)
1067 remove_migration_ptes(hpage,
1068 rc == MIGRATEPAGE_SUCCESS ? new_hpage : hpage);
1069
1070 unlock_page(new_hpage);
1071
1072 put_anon:
1073 if (anon_vma)
1074 put_anon_vma(anon_vma);
1075
1076 if (rc == MIGRATEPAGE_SUCCESS) {
1077 hugetlb_cgroup_migrate(hpage, new_hpage);
1078 put_new_page = NULL;
1079 }
1080
1081 unlock_page(hpage);
1082 out:
1083 if (rc != -EAGAIN)
1084 putback_active_hugepage(hpage);
1085
1086 /*
1087 * If migration was not successful and there's a freeing callback, use
1088 * it. Otherwise, put_page() will drop the reference grabbed during
1089 * isolation.
1090 */
1091 if (put_new_page)
1092 put_new_page(new_hpage, private);
1093 else
1094 putback_active_hugepage(new_hpage);
1095
1096 if (result) {
1097 if (rc)
1098 *result = rc;
1099 else
1100 *result = page_to_nid(new_hpage);
1101 }
1102 return rc;
1103 }
1104
1105 /*
1106 * migrate_pages - migrate the pages specified in a list, to the free pages
1107 * supplied as the target for the page migration
1108 *
1109 * @from: The list of pages to be migrated.
1110 * @get_new_page: The function used to allocate free pages to be used
1111 * as the target of the page migration.
1112 * @put_new_page: The function used to free target pages if migration
1113 * fails, or NULL if no special handling is necessary.
1114 * @private: Private data to be passed on to get_new_page()
1115 * @mode: The migration mode that specifies the constraints for
1116 * page migration, if any.
1117 * @reason: The reason for page migration.
1118 *
1119 * The function returns after 10 attempts or if no pages are movable any more
1120 * because the list has become empty or no retryable pages exist any more.
1121 * The caller should call putback_movable_pages() to return pages to the LRU
1122 * or free list only if ret != 0.
1123 *
1124 * Returns the number of pages that were not migrated, or an error code.
1125 */
1126 int migrate_pages(struct list_head *from, new_page_t get_new_page,
1127 free_page_t put_new_page, unsigned long private,
1128 enum migrate_mode mode, int reason)
1129 {
1130 int retry = 1;
1131 int nr_failed = 0;
1132 int nr_succeeded = 0;
1133 int pass = 0;
1134 struct page *page;
1135 struct page *page2;
1136 int swapwrite = current->flags & PF_SWAPWRITE;
1137 int rc;
1138
1139 if (!swapwrite)
1140 current->flags |= PF_SWAPWRITE;
1141
1142 for(pass = 0; pass < 10 && retry; pass++) {
1143 retry = 0;
1144
1145 list_for_each_entry_safe(page, page2, from, lru) {
1146 cond_resched();
1147
1148 if (PageHuge(page))
1149 rc = unmap_and_move_huge_page(get_new_page,
1150 put_new_page, private, page,
1151 pass > 2, mode);
1152 else
1153 rc = unmap_and_move(get_new_page, put_new_page,
1154 private, page, pass > 2, mode,
1155 reason);
1156
1157 switch(rc) {
1158 case -ENOMEM:
1159 goto out;
1160 case -EAGAIN:
1161 retry++;
1162 break;
1163 case MIGRATEPAGE_SUCCESS:
1164 nr_succeeded++;
1165 break;
1166 default:
1167 /*
1168 * Permanent failure (-EBUSY, -ENOSYS, etc.):
1169 * unlike -EAGAIN case, the failed page is
1170 * removed from migration page list and not
1171 * retried in the next outer loop.
1172 */
1173 nr_failed++;
1174 break;
1175 }
1176 }
1177 }
1178 nr_failed += retry;
1179 rc = nr_failed;
1180 out:
1181 if (nr_succeeded)
1182 count_vm_events(PGMIGRATE_SUCCESS, nr_succeeded);
1183 if (nr_failed)
1184 count_vm_events(PGMIGRATE_FAIL, nr_failed);
1185 trace_mm_migrate_pages(nr_succeeded, nr_failed, mode, reason);
1186
1187 if (!swapwrite)
1188 current->flags &= ~PF_SWAPWRITE;
1189
1190 return rc;
1191 }
1192
1193 #ifdef CONFIG_NUMA
1194 /*
1195 * Move a list of individual pages
1196 */
1197 struct page_to_node {
1198 unsigned long addr;
1199 struct page *page;
1200 int node;
1201 int status;
1202 };
1203
1204 static struct page *new_page_node(struct page *p, unsigned long private,
1205 int **result)
1206 {
1207 struct page_to_node *pm = (struct page_to_node *)private;
1208
1209 while (pm->node != MAX_NUMNODES && pm->page != p)
1210 pm++;
1211
1212 if (pm->node == MAX_NUMNODES)
1213 return NULL;
1214
1215 *result = &pm->status;
1216
1217 if (PageHuge(p))
1218 return alloc_huge_page_node(page_hstate(compound_head(p)),
1219 pm->node);
1220 else
1221 return __alloc_pages_node(pm->node,
1222 GFP_HIGHUSER_MOVABLE | __GFP_THISNODE, 0);
1223 }
1224
1225 /*
1226 * Move a set of pages as indicated in the pm array. The addr
1227 * field must be set to the virtual address of the page to be moved
1228 * and the node number must contain a valid target node.
1229 * The pm array ends with node = MAX_NUMNODES.
1230 */
1231 static int do_move_page_to_node_array(struct mm_struct *mm,
1232 struct page_to_node *pm,
1233 int migrate_all)
1234 {
1235 int err;
1236 struct page_to_node *pp;
1237 LIST_HEAD(pagelist);
1238
1239 down_read(&mm->mmap_sem);
1240
1241 /*
1242 * Build a list of pages to migrate
1243 */
1244 for (pp = pm; pp->node != MAX_NUMNODES; pp++) {
1245 struct vm_area_struct *vma;
1246 struct page *page;
1247
1248 err = -EFAULT;
1249 vma = find_vma(mm, pp->addr);
1250 if (!vma || pp->addr < vma->vm_start || !vma_migratable(vma))
1251 goto set_status;
1252
1253 /* FOLL_DUMP to ignore special (like zero) pages */
1254 page = follow_page(vma, pp->addr,
1255 FOLL_GET | FOLL_SPLIT | FOLL_DUMP);
1256
1257 err = PTR_ERR(page);
1258 if (IS_ERR(page))
1259 goto set_status;
1260
1261 err = -ENOENT;
1262 if (!page)
1263 goto set_status;
1264
1265 pp->page = page;
1266 err = page_to_nid(page);
1267
1268 if (err == pp->node)
1269 /*
1270 * Node already in the right place
1271 */
1272 goto put_and_set;
1273
1274 err = -EACCES;
1275 if (page_mapcount(page) > 1 &&
1276 !migrate_all)
1277 goto put_and_set;
1278
1279 if (PageHuge(page)) {
1280 if (PageHead(page))
1281 isolate_huge_page(page, &pagelist);
1282 goto put_and_set;
1283 }
1284
1285 err = isolate_lru_page(page);
1286 if (!err) {
1287 list_add_tail(&page->lru, &pagelist);
1288 inc_zone_page_state(page, NR_ISOLATED_ANON +
1289 page_is_file_cache(page));
1290 }
1291 put_and_set:
1292 /*
1293 * Either remove the duplicate refcount from
1294 * isolate_lru_page() or drop the page ref if it was
1295 * not isolated.
1296 */
1297 put_page(page);
1298 set_status:
1299 pp->status = err;
1300 }
1301
1302 err = 0;
1303 if (!list_empty(&pagelist)) {
1304 err = migrate_pages(&pagelist, new_page_node, NULL,
1305 (unsigned long)pm, MIGRATE_SYNC, MR_SYSCALL);
1306 if (err)
1307 putback_movable_pages(&pagelist);
1308 }
1309
1310 up_read(&mm->mmap_sem);
1311 return err;
1312 }
1313
1314 /*
1315 * Migrate an array of page address onto an array of nodes and fill
1316 * the corresponding array of status.
1317 */
1318 static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes,
1319 unsigned long nr_pages,
1320 const void __user * __user *pages,
1321 const int __user *nodes,
1322 int __user *status, int flags)
1323 {
1324 struct page_to_node *pm;
1325 unsigned long chunk_nr_pages;
1326 unsigned long chunk_start;
1327 int err;
1328
1329 err = -ENOMEM;
1330 pm = (struct page_to_node *)__get_free_page(GFP_KERNEL);
1331 if (!pm)
1332 goto out;
1333
1334 migrate_prep();
1335
1336 /*
1337 * Store a chunk of page_to_node array in a page,
1338 * but keep the last one as a marker
1339 */
1340 chunk_nr_pages = (PAGE_SIZE / sizeof(struct page_to_node)) - 1;
1341
1342 for (chunk_start = 0;
1343 chunk_start < nr_pages;
1344 chunk_start += chunk_nr_pages) {
1345 int j;
1346
1347 if (chunk_start + chunk_nr_pages > nr_pages)
1348 chunk_nr_pages = nr_pages - chunk_start;
1349
1350 /* fill the chunk pm with addrs and nodes from user-space */
1351 for (j = 0; j < chunk_nr_pages; j++) {
1352 const void __user *p;
1353 int node;
1354
1355 err = -EFAULT;
1356 if (get_user(p, pages + j + chunk_start))
1357 goto out_pm;
1358 pm[j].addr = (unsigned long) p;
1359
1360 if (get_user(node, nodes + j + chunk_start))
1361 goto out_pm;
1362
1363 err = -ENODEV;
1364 if (node < 0 || node >= MAX_NUMNODES)
1365 goto out_pm;
1366
1367 if (!node_state(node, N_MEMORY))
1368 goto out_pm;
1369
1370 err = -EACCES;
1371 if (!node_isset(node, task_nodes))
1372 goto out_pm;
1373
1374 pm[j].node = node;
1375 }
1376
1377 /* End marker for this chunk */
1378 pm[chunk_nr_pages].node = MAX_NUMNODES;
1379
1380 /* Migrate this chunk */
1381 err = do_move_page_to_node_array(mm, pm,
1382 flags & MPOL_MF_MOVE_ALL);
1383 if (err < 0)
1384 goto out_pm;
1385
1386 /* Return status information */
1387 for (j = 0; j < chunk_nr_pages; j++)
1388 if (put_user(pm[j].status, status + j + chunk_start)) {
1389 err = -EFAULT;
1390 goto out_pm;
1391 }
1392 }
1393 err = 0;
1394
1395 out_pm:
1396 free_page((unsigned long)pm);
1397 out:
1398 return err;
1399 }
1400
1401 /*
1402 * Determine the nodes of an array of pages and store it in an array of status.
1403 */
1404 static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages,
1405 const void __user **pages, int *status)
1406 {
1407 unsigned long i;
1408
1409 down_read(&mm->mmap_sem);
1410
1411 for (i = 0; i < nr_pages; i++) {
1412 unsigned long addr = (unsigned long)(*pages);
1413 struct vm_area_struct *vma;
1414 struct page *page;
1415 int err = -EFAULT;
1416
1417 vma = find_vma(mm, addr);
1418 if (!vma || addr < vma->vm_start)
1419 goto set_status;
1420
1421 /* FOLL_DUMP to ignore special (like zero) pages */
1422 page = follow_page(vma, addr, FOLL_DUMP);
1423
1424 err = PTR_ERR(page);
1425 if (IS_ERR(page))
1426 goto set_status;
1427
1428 err = page ? page_to_nid(page) : -ENOENT;
1429 set_status:
1430 *status = err;
1431
1432 pages++;
1433 status++;
1434 }
1435
1436 up_read(&mm->mmap_sem);
1437 }
1438
1439 /*
1440 * Determine the nodes of a user array of pages and store it in
1441 * a user array of status.
1442 */
1443 static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages,
1444 const void __user * __user *pages,
1445 int __user *status)
1446 {
1447 #define DO_PAGES_STAT_CHUNK_NR 16
1448 const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR];
1449 int chunk_status[DO_PAGES_STAT_CHUNK_NR];
1450
1451 while (nr_pages) {
1452 unsigned long chunk_nr;
1453
1454 chunk_nr = nr_pages;
1455 if (chunk_nr > DO_PAGES_STAT_CHUNK_NR)
1456 chunk_nr = DO_PAGES_STAT_CHUNK_NR;
1457
1458 if (copy_from_user(chunk_pages, pages, chunk_nr * sizeof(*chunk_pages)))
1459 break;
1460
1461 do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status);
1462
1463 if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status)))
1464 break;
1465
1466 pages += chunk_nr;
1467 status += chunk_nr;
1468 nr_pages -= chunk_nr;
1469 }
1470 return nr_pages ? -EFAULT : 0;
1471 }
1472
1473 /*
1474 * Move a list of pages in the address space of the currently executing
1475 * process.
1476 */
1477 SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages,
1478 const void __user * __user *, pages,
1479 const int __user *, nodes,
1480 int __user *, status, int, flags)
1481 {
1482 const struct cred *cred = current_cred(), *tcred;
1483 struct task_struct *task;
1484 struct mm_struct *mm;
1485 int err;
1486 nodemask_t task_nodes;
1487
1488 /* Check flags */
1489 if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
1490 return -EINVAL;
1491
1492 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1493 return -EPERM;
1494
1495 /* Find the mm_struct */
1496 rcu_read_lock();
1497 task = pid ? find_task_by_vpid(pid) : current;
1498 if (!task) {
1499 rcu_read_unlock();
1500 return -ESRCH;
1501 }
1502 get_task_struct(task);
1503
1504 /*
1505 * Check if this process has the right to modify the specified
1506 * process. The right exists if the process has administrative
1507 * capabilities, superuser privileges or the same
1508 * userid as the target process.
1509 */
1510 tcred = __task_cred(task);
1511 if (!uid_eq(cred->euid, tcred->suid) && !uid_eq(cred->euid, tcred->uid) &&
1512 !uid_eq(cred->uid, tcred->suid) && !uid_eq(cred->uid, tcred->uid) &&
1513 !capable(CAP_SYS_NICE)) {
1514 rcu_read_unlock();
1515 err = -EPERM;
1516 goto out;
1517 }
1518 rcu_read_unlock();
1519
1520 err = security_task_movememory(task);
1521 if (err)
1522 goto out;
1523
1524 task_nodes = cpuset_mems_allowed(task);
1525 mm = get_task_mm(task);
1526 put_task_struct(task);
1527
1528 if (!mm)
1529 return -EINVAL;
1530
1531 if (nodes)
1532 err = do_pages_move(mm, task_nodes, nr_pages, pages,
1533 nodes, status, flags);
1534 else
1535 err = do_pages_stat(mm, nr_pages, pages, status);
1536
1537 mmput(mm);
1538 return err;
1539
1540 out:
1541 put_task_struct(task);
1542 return err;
1543 }
1544
1545 #ifdef CONFIG_NUMA_BALANCING
1546 /*
1547 * Returns true if this is a safe migration target node for misplaced NUMA
1548 * pages. Currently it only checks the watermarks which crude
1549 */
1550 static bool migrate_balanced_pgdat(struct pglist_data *pgdat,
1551 unsigned long nr_migrate_pages)
1552 {
1553 int z;
1554 for (z = pgdat->nr_zones - 1; z >= 0; z--) {
1555 struct zone *zone = pgdat->node_zones + z;
1556
1557 if (!populated_zone(zone))
1558 continue;
1559
1560 if (!zone_reclaimable(zone))
1561 continue;
1562
1563 /* Avoid waking kswapd by allocating pages_to_migrate pages. */
1564 if (!zone_watermark_ok(zone, 0,
1565 high_wmark_pages(zone) +
1566 nr_migrate_pages,
1567 0, 0))
1568 continue;
1569 return true;
1570 }
1571 return false;
1572 }
1573
1574 static struct page *alloc_misplaced_dst_page(struct page *page,
1575 unsigned long data,
1576 int **result)
1577 {
1578 int nid = (int) data;
1579 struct page *newpage;
1580
1581 newpage = __alloc_pages_node(nid,
1582 (GFP_HIGHUSER_MOVABLE |
1583 __GFP_THISNODE | __GFP_NOMEMALLOC |
1584 __GFP_NORETRY | __GFP_NOWARN) &
1585 ~(__GFP_IO | __GFP_FS), 0);
1586
1587 return newpage;
1588 }
1589
1590 /*
1591 * page migration rate limiting control.
1592 * Do not migrate more than @pages_to_migrate in a @migrate_interval_millisecs
1593 * window of time. Default here says do not migrate more than 1280M per second.
1594 */
1595 static unsigned int migrate_interval_millisecs __read_mostly = 100;
1596 static unsigned int ratelimit_pages __read_mostly = 128 << (20 - PAGE_SHIFT);
1597
1598 /* Returns true if the node is migrate rate-limited after the update */
1599 static bool numamigrate_update_ratelimit(pg_data_t *pgdat,
1600 unsigned long nr_pages)
1601 {
1602 /*
1603 * Rate-limit the amount of data that is being migrated to a node.
1604 * Optimal placement is no good if the memory bus is saturated and
1605 * all the time is being spent migrating!
1606 */
1607 if (time_after(jiffies, pgdat->numabalancing_migrate_next_window)) {
1608 spin_lock(&pgdat->numabalancing_migrate_lock);
1609 pgdat->numabalancing_migrate_nr_pages = 0;
1610 pgdat->numabalancing_migrate_next_window = jiffies +
1611 msecs_to_jiffies(migrate_interval_millisecs);
1612 spin_unlock(&pgdat->numabalancing_migrate_lock);
1613 }
1614 if (pgdat->numabalancing_migrate_nr_pages > ratelimit_pages) {
1615 trace_mm_numa_migrate_ratelimit(current, pgdat->node_id,
1616 nr_pages);
1617 return true;
1618 }
1619
1620 /*
1621 * This is an unlocked non-atomic update so errors are possible.
1622 * The consequences are failing to migrate when we potentiall should
1623 * have which is not severe enough to warrant locking. If it is ever
1624 * a problem, it can be converted to a per-cpu counter.
1625 */
1626 pgdat->numabalancing_migrate_nr_pages += nr_pages;
1627 return false;
1628 }
1629
1630 static int numamigrate_isolate_page(pg_data_t *pgdat, struct page *page)
1631 {
1632 int page_lru;
1633
1634 VM_BUG_ON_PAGE(compound_order(page) && !PageTransHuge(page), page);
1635
1636 /* Avoid migrating to a node that is nearly full */
1637 if (!migrate_balanced_pgdat(pgdat, 1UL << compound_order(page)))
1638 return 0;
1639
1640 if (isolate_lru_page(page))
1641 return 0;
1642
1643 /*
1644 * migrate_misplaced_transhuge_page() skips page migration's usual
1645 * check on page_count(), so we must do it here, now that the page
1646 * has been isolated: a GUP pin, or any other pin, prevents migration.
1647 * The expected page count is 3: 1 for page's mapcount and 1 for the
1648 * caller's pin and 1 for the reference taken by isolate_lru_page().
1649 */
1650 if (PageTransHuge(page) && page_count(page) != 3) {
1651 putback_lru_page(page);
1652 return 0;
1653 }
1654
1655 page_lru = page_is_file_cache(page);
1656 mod_zone_page_state(page_zone(page), NR_ISOLATED_ANON + page_lru,
1657 hpage_nr_pages(page));
1658
1659 /*
1660 * Isolating the page has taken another reference, so the
1661 * caller's reference can be safely dropped without the page
1662 * disappearing underneath us during migration.
1663 */
1664 put_page(page);
1665 return 1;
1666 }
1667
1668 bool pmd_trans_migrating(pmd_t pmd)
1669 {
1670 struct page *page = pmd_page(pmd);
1671 return PageLocked(page);
1672 }
1673
1674 /*
1675 * Attempt to migrate a misplaced page to the specified destination
1676 * node. Caller is expected to have an elevated reference count on
1677 * the page that will be dropped by this function before returning.
1678 */
1679 int migrate_misplaced_page(struct page *page, struct vm_area_struct *vma,
1680 int node)
1681 {
1682 pg_data_t *pgdat = NODE_DATA(node);
1683 int isolated;
1684 int nr_remaining;
1685 LIST_HEAD(migratepages);
1686
1687 /*
1688 * Don't migrate file pages that are mapped in multiple processes
1689 * with execute permissions as they are probably shared libraries.
1690 */
1691 if (page_mapcount(page) != 1 && page_is_file_cache(page) &&
1692 (vma->vm_flags & VM_EXEC))
1693 goto out;
1694
1695 /*
1696 * Rate-limit the amount of data that is being migrated to a node.
1697 * Optimal placement is no good if the memory bus is saturated and
1698 * all the time is being spent migrating!
1699 */
1700 if (numamigrate_update_ratelimit(pgdat, 1))
1701 goto out;
1702
1703 isolated = numamigrate_isolate_page(pgdat, page);
1704 if (!isolated)
1705 goto out;
1706
1707 list_add(&page->lru, &migratepages);
1708 nr_remaining = migrate_pages(&migratepages, alloc_misplaced_dst_page,
1709 NULL, node, MIGRATE_ASYNC,
1710 MR_NUMA_MISPLACED);
1711 if (nr_remaining) {
1712 if (!list_empty(&migratepages)) {
1713 list_del(&page->lru);
1714 dec_zone_page_state(page, NR_ISOLATED_ANON +
1715 page_is_file_cache(page));
1716 putback_lru_page(page);
1717 }
1718 isolated = 0;
1719 } else
1720 count_vm_numa_event(NUMA_PAGE_MIGRATE);
1721 BUG_ON(!list_empty(&migratepages));
1722 return isolated;
1723
1724 out:
1725 put_page(page);
1726 return 0;
1727 }
1728 #endif /* CONFIG_NUMA_BALANCING */
1729
1730 #if defined(CONFIG_NUMA_BALANCING) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
1731 /*
1732 * Migrates a THP to a given target node. page must be locked and is unlocked
1733 * before returning.
1734 */
1735 int migrate_misplaced_transhuge_page(struct mm_struct *mm,
1736 struct vm_area_struct *vma,
1737 pmd_t *pmd, pmd_t entry,
1738 unsigned long address,
1739 struct page *page, int node)
1740 {
1741 spinlock_t *ptl;
1742 pg_data_t *pgdat = NODE_DATA(node);
1743 int isolated = 0;
1744 struct page *new_page = NULL;
1745 int page_lru = page_is_file_cache(page);
1746 unsigned long mmun_start = address & HPAGE_PMD_MASK;
1747 unsigned long mmun_end = mmun_start + HPAGE_PMD_SIZE;
1748 pmd_t orig_entry;
1749
1750 /*
1751 * Rate-limit the amount of data that is being migrated to a node.
1752 * Optimal placement is no good if the memory bus is saturated and
1753 * all the time is being spent migrating!
1754 */
1755 if (numamigrate_update_ratelimit(pgdat, HPAGE_PMD_NR))
1756 goto out_dropref;
1757
1758 new_page = alloc_pages_node(node,
1759 (GFP_TRANSHUGE | __GFP_THISNODE) & ~__GFP_RECLAIM,
1760 HPAGE_PMD_ORDER);
1761 if (!new_page)
1762 goto out_fail;
1763 prep_transhuge_page(new_page);
1764
1765 isolated = numamigrate_isolate_page(pgdat, page);
1766 if (!isolated) {
1767 put_page(new_page);
1768 goto out_fail;
1769 }
1770
1771 if (mm_tlb_flush_pending(mm))
1772 flush_tlb_range(vma, mmun_start, mmun_end);
1773
1774 /* Prepare a page as a migration target */
1775 __SetPageLocked(new_page);
1776 SetPageSwapBacked(new_page);
1777
1778 /* anon mapping, we can simply copy page->mapping to the new page: */
1779 new_page->mapping = page->mapping;
1780 new_page->index = page->index;
1781 migrate_page_copy(new_page, page);
1782 WARN_ON(PageLRU(new_page));
1783
1784 /* Recheck the target PMD */
1785 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1786 ptl = pmd_lock(mm, pmd);
1787 if (unlikely(!pmd_same(*pmd, entry) || page_count(page) != 2)) {
1788 fail_putback:
1789 spin_unlock(ptl);
1790 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1791
1792 /* Reverse changes made by migrate_page_copy() */
1793 if (TestClearPageActive(new_page))
1794 SetPageActive(page);
1795 if (TestClearPageUnevictable(new_page))
1796 SetPageUnevictable(page);
1797
1798 unlock_page(new_page);
1799 put_page(new_page); /* Free it */
1800
1801 /* Retake the callers reference and putback on LRU */
1802 get_page(page);
1803 putback_lru_page(page);
1804 mod_zone_page_state(page_zone(page),
1805 NR_ISOLATED_ANON + page_lru, -HPAGE_PMD_NR);
1806
1807 goto out_unlock;
1808 }
1809
1810 orig_entry = *pmd;
1811 entry = mk_pmd(new_page, vma->vm_page_prot);
1812 entry = pmd_mkhuge(entry);
1813 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1814
1815 /*
1816 * Clear the old entry under pagetable lock and establish the new PTE.
1817 * Any parallel GUP will either observe the old page blocking on the
1818 * page lock, block on the page table lock or observe the new page.
1819 * The SetPageUptodate on the new page and page_add_new_anon_rmap
1820 * guarantee the copy is visible before the pagetable update.
1821 */
1822 flush_cache_range(vma, mmun_start, mmun_end);
1823 page_add_anon_rmap(new_page, vma, mmun_start, true);
1824 pmdp_huge_clear_flush_notify(vma, mmun_start, pmd);
1825 set_pmd_at(mm, mmun_start, pmd, entry);
1826 flush_tlb_range(vma, mmun_start, mmun_end);
1827 update_mmu_cache_pmd(vma, address, &entry);
1828
1829 if (page_count(page) != 2) {
1830 set_pmd_at(mm, mmun_start, pmd, orig_entry);
1831 flush_tlb_range(vma, mmun_start, mmun_end);
1832 mmu_notifier_invalidate_range(mm, mmun_start, mmun_end);
1833 update_mmu_cache_pmd(vma, address, &entry);
1834 page_remove_rmap(new_page, true);
1835 goto fail_putback;
1836 }
1837
1838 mlock_migrate_page(new_page, page);
1839 set_page_memcg(new_page, page_memcg(page));
1840 set_page_memcg(page, NULL);
1841 page_remove_rmap(page, true);
1842
1843 spin_unlock(ptl);
1844 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1845
1846 /* Take an "isolate" reference and put new page on the LRU. */
1847 get_page(new_page);
1848 putback_lru_page(new_page);
1849
1850 unlock_page(new_page);
1851 unlock_page(page);
1852 put_page(page); /* Drop the rmap reference */
1853 put_page(page); /* Drop the LRU isolation reference */
1854
1855 count_vm_events(PGMIGRATE_SUCCESS, HPAGE_PMD_NR);
1856 count_vm_numa_events(NUMA_PAGE_MIGRATE, HPAGE_PMD_NR);
1857
1858 mod_zone_page_state(page_zone(page),
1859 NR_ISOLATED_ANON + page_lru,
1860 -HPAGE_PMD_NR);
1861 return isolated;
1862
1863 out_fail:
1864 count_vm_events(PGMIGRATE_FAIL, HPAGE_PMD_NR);
1865 out_dropref:
1866 ptl = pmd_lock(mm, pmd);
1867 if (pmd_same(*pmd, entry)) {
1868 entry = pmd_modify(entry, vma->vm_page_prot);
1869 set_pmd_at(mm, mmun_start, pmd, entry);
1870 update_mmu_cache_pmd(vma, address, &entry);
1871 }
1872 spin_unlock(ptl);
1873
1874 out_unlock:
1875 unlock_page(page);
1876 put_page(page);
1877 return 0;
1878 }
1879 #endif /* CONFIG_NUMA_BALANCING */
1880
1881 #endif /* CONFIG_NUMA */