mm: migrate: support non-lru movable page migration
[deliverable/linux.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/compaction.h>
35 #include <linux/syscalls.h>
36 #include <linux/hugetlb.h>
37 #include <linux/hugetlb_cgroup.h>
38 #include <linux/gfp.h>
39 #include <linux/balloon_compaction.h>
40 #include <linux/mmu_notifier.h>
41 #include <linux/page_idle.h>
42 #include <linux/page_owner.h>
43
44 #include <asm/tlbflush.h>
45
46 #define CREATE_TRACE_POINTS
47 #include <trace/events/migrate.h>
48
49 #include "internal.h"
50
51 /*
52 * migrate_prep() needs to be called before we start compiling a list of pages
53 * to be migrated using isolate_lru_page(). If scheduling work on other CPUs is
54 * undesirable, use migrate_prep_local()
55 */
56 int migrate_prep(void)
57 {
58 /*
59 * Clear the LRU lists so pages can be isolated.
60 * Note that pages may be moved off the LRU after we have
61 * drained them. Those pages will fail to migrate like other
62 * pages that may be busy.
63 */
64 lru_add_drain_all();
65
66 return 0;
67 }
68
69 /* Do the necessary work of migrate_prep but not if it involves other CPUs */
70 int migrate_prep_local(void)
71 {
72 lru_add_drain();
73
74 return 0;
75 }
76
77 bool isolate_movable_page(struct page *page, isolate_mode_t mode)
78 {
79 struct address_space *mapping;
80
81 /*
82 * Avoid burning cycles with pages that are yet under __free_pages(),
83 * or just got freed under us.
84 *
85 * In case we 'win' a race for a movable page being freed under us and
86 * raise its refcount preventing __free_pages() from doing its job
87 * the put_page() at the end of this block will take care of
88 * release this page, thus avoiding a nasty leakage.
89 */
90 if (unlikely(!get_page_unless_zero(page)))
91 goto out;
92
93 /*
94 * Check PageMovable before holding a PG_lock because page's owner
95 * assumes anybody doesn't touch PG_lock of newly allocated page
96 * so unconditionally grapping the lock ruins page's owner side.
97 */
98 if (unlikely(!__PageMovable(page)))
99 goto out_putpage;
100 /*
101 * As movable pages are not isolated from LRU lists, concurrent
102 * compaction threads can race against page migration functions
103 * as well as race against the releasing a page.
104 *
105 * In order to avoid having an already isolated movable page
106 * being (wrongly) re-isolated while it is under migration,
107 * or to avoid attempting to isolate pages being released,
108 * lets be sure we have the page lock
109 * before proceeding with the movable page isolation steps.
110 */
111 if (unlikely(!trylock_page(page)))
112 goto out_putpage;
113
114 if (!PageMovable(page) || PageIsolated(page))
115 goto out_no_isolated;
116
117 mapping = page_mapping(page);
118 VM_BUG_ON_PAGE(!mapping, page);
119
120 if (!mapping->a_ops->isolate_page(page, mode))
121 goto out_no_isolated;
122
123 /* Driver shouldn't use PG_isolated bit of page->flags */
124 WARN_ON_ONCE(PageIsolated(page));
125 __SetPageIsolated(page);
126 unlock_page(page);
127
128 return true;
129
130 out_no_isolated:
131 unlock_page(page);
132 out_putpage:
133 put_page(page);
134 out:
135 return false;
136 }
137
138 /* It should be called on page which is PG_movable */
139 void putback_movable_page(struct page *page)
140 {
141 struct address_space *mapping;
142
143 VM_BUG_ON_PAGE(!PageLocked(page), page);
144 VM_BUG_ON_PAGE(!PageMovable(page), page);
145 VM_BUG_ON_PAGE(!PageIsolated(page), page);
146
147 mapping = page_mapping(page);
148 mapping->a_ops->putback_page(page);
149 __ClearPageIsolated(page);
150 }
151
152 /*
153 * Put previously isolated pages back onto the appropriate lists
154 * from where they were once taken off for compaction/migration.
155 *
156 * This function shall be used whenever the isolated pageset has been
157 * built from lru, balloon, hugetlbfs page. See isolate_migratepages_range()
158 * and isolate_huge_page().
159 */
160 void putback_movable_pages(struct list_head *l)
161 {
162 struct page *page;
163 struct page *page2;
164
165 list_for_each_entry_safe(page, page2, l, lru) {
166 if (unlikely(PageHuge(page))) {
167 putback_active_hugepage(page);
168 continue;
169 }
170 list_del(&page->lru);
171 dec_zone_page_state(page, NR_ISOLATED_ANON +
172 page_is_file_cache(page));
173 if (unlikely(isolated_balloon_page(page))) {
174 balloon_page_putback(page);
175 /*
176 * We isolated non-lru movable page so here we can use
177 * __PageMovable because LRU page's mapping cannot have
178 * PAGE_MAPPING_MOVABLE.
179 */
180 } else if (unlikely(__PageMovable(page))) {
181 VM_BUG_ON_PAGE(!PageIsolated(page), page);
182 lock_page(page);
183 if (PageMovable(page))
184 putback_movable_page(page);
185 else
186 __ClearPageIsolated(page);
187 unlock_page(page);
188 put_page(page);
189 } else {
190 putback_lru_page(page);
191 }
192 }
193 }
194
195 /*
196 * Restore a potential migration pte to a working pte entry
197 */
198 static int remove_migration_pte(struct page *new, struct vm_area_struct *vma,
199 unsigned long addr, void *old)
200 {
201 struct mm_struct *mm = vma->vm_mm;
202 swp_entry_t entry;
203 pmd_t *pmd;
204 pte_t *ptep, pte;
205 spinlock_t *ptl;
206
207 if (unlikely(PageHuge(new))) {
208 ptep = huge_pte_offset(mm, addr);
209 if (!ptep)
210 goto out;
211 ptl = huge_pte_lockptr(hstate_vma(vma), mm, ptep);
212 } else {
213 pmd = mm_find_pmd(mm, addr);
214 if (!pmd)
215 goto out;
216
217 ptep = pte_offset_map(pmd, addr);
218
219 /*
220 * Peek to check is_swap_pte() before taking ptlock? No, we
221 * can race mremap's move_ptes(), which skips anon_vma lock.
222 */
223
224 ptl = pte_lockptr(mm, pmd);
225 }
226
227 spin_lock(ptl);
228 pte = *ptep;
229 if (!is_swap_pte(pte))
230 goto unlock;
231
232 entry = pte_to_swp_entry(pte);
233
234 if (!is_migration_entry(entry) ||
235 migration_entry_to_page(entry) != old)
236 goto unlock;
237
238 get_page(new);
239 pte = pte_mkold(mk_pte(new, vma->vm_page_prot));
240 if (pte_swp_soft_dirty(*ptep))
241 pte = pte_mksoft_dirty(pte);
242
243 /* Recheck VMA as permissions can change since migration started */
244 if (is_write_migration_entry(entry))
245 pte = maybe_mkwrite(pte, vma);
246
247 #ifdef CONFIG_HUGETLB_PAGE
248 if (PageHuge(new)) {
249 pte = pte_mkhuge(pte);
250 pte = arch_make_huge_pte(pte, vma, new, 0);
251 }
252 #endif
253 flush_dcache_page(new);
254 set_pte_at(mm, addr, ptep, pte);
255
256 if (PageHuge(new)) {
257 if (PageAnon(new))
258 hugepage_add_anon_rmap(new, vma, addr);
259 else
260 page_dup_rmap(new, true);
261 } else if (PageAnon(new))
262 page_add_anon_rmap(new, vma, addr, false);
263 else
264 page_add_file_rmap(new);
265
266 if (vma->vm_flags & VM_LOCKED && !PageTransCompound(new))
267 mlock_vma_page(new);
268
269 /* No need to invalidate - it was non-present before */
270 update_mmu_cache(vma, addr, ptep);
271 unlock:
272 pte_unmap_unlock(ptep, ptl);
273 out:
274 return SWAP_AGAIN;
275 }
276
277 /*
278 * Get rid of all migration entries and replace them by
279 * references to the indicated page.
280 */
281 void remove_migration_ptes(struct page *old, struct page *new, bool locked)
282 {
283 struct rmap_walk_control rwc = {
284 .rmap_one = remove_migration_pte,
285 .arg = old,
286 };
287
288 if (locked)
289 rmap_walk_locked(new, &rwc);
290 else
291 rmap_walk(new, &rwc);
292 }
293
294 /*
295 * Something used the pte of a page under migration. We need to
296 * get to the page and wait until migration is finished.
297 * When we return from this function the fault will be retried.
298 */
299 void __migration_entry_wait(struct mm_struct *mm, pte_t *ptep,
300 spinlock_t *ptl)
301 {
302 pte_t pte;
303 swp_entry_t entry;
304 struct page *page;
305
306 spin_lock(ptl);
307 pte = *ptep;
308 if (!is_swap_pte(pte))
309 goto out;
310
311 entry = pte_to_swp_entry(pte);
312 if (!is_migration_entry(entry))
313 goto out;
314
315 page = migration_entry_to_page(entry);
316
317 /*
318 * Once radix-tree replacement of page migration started, page_count
319 * *must* be zero. And, we don't want to call wait_on_page_locked()
320 * against a page without get_page().
321 * So, we use get_page_unless_zero(), here. Even failed, page fault
322 * will occur again.
323 */
324 if (!get_page_unless_zero(page))
325 goto out;
326 pte_unmap_unlock(ptep, ptl);
327 wait_on_page_locked(page);
328 put_page(page);
329 return;
330 out:
331 pte_unmap_unlock(ptep, ptl);
332 }
333
334 void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
335 unsigned long address)
336 {
337 spinlock_t *ptl = pte_lockptr(mm, pmd);
338 pte_t *ptep = pte_offset_map(pmd, address);
339 __migration_entry_wait(mm, ptep, ptl);
340 }
341
342 void migration_entry_wait_huge(struct vm_area_struct *vma,
343 struct mm_struct *mm, pte_t *pte)
344 {
345 spinlock_t *ptl = huge_pte_lockptr(hstate_vma(vma), mm, pte);
346 __migration_entry_wait(mm, pte, ptl);
347 }
348
349 #ifdef CONFIG_BLOCK
350 /* Returns true if all buffers are successfully locked */
351 static bool buffer_migrate_lock_buffers(struct buffer_head *head,
352 enum migrate_mode mode)
353 {
354 struct buffer_head *bh = head;
355
356 /* Simple case, sync compaction */
357 if (mode != MIGRATE_ASYNC) {
358 do {
359 get_bh(bh);
360 lock_buffer(bh);
361 bh = bh->b_this_page;
362
363 } while (bh != head);
364
365 return true;
366 }
367
368 /* async case, we cannot block on lock_buffer so use trylock_buffer */
369 do {
370 get_bh(bh);
371 if (!trylock_buffer(bh)) {
372 /*
373 * We failed to lock the buffer and cannot stall in
374 * async migration. Release the taken locks
375 */
376 struct buffer_head *failed_bh = bh;
377 put_bh(failed_bh);
378 bh = head;
379 while (bh != failed_bh) {
380 unlock_buffer(bh);
381 put_bh(bh);
382 bh = bh->b_this_page;
383 }
384 return false;
385 }
386
387 bh = bh->b_this_page;
388 } while (bh != head);
389 return true;
390 }
391 #else
392 static inline bool buffer_migrate_lock_buffers(struct buffer_head *head,
393 enum migrate_mode mode)
394 {
395 return true;
396 }
397 #endif /* CONFIG_BLOCK */
398
399 /*
400 * Replace the page in the mapping.
401 *
402 * The number of remaining references must be:
403 * 1 for anonymous pages without a mapping
404 * 2 for pages with a mapping
405 * 3 for pages with a mapping and PagePrivate/PagePrivate2 set.
406 */
407 int migrate_page_move_mapping(struct address_space *mapping,
408 struct page *newpage, struct page *page,
409 struct buffer_head *head, enum migrate_mode mode,
410 int extra_count)
411 {
412 struct zone *oldzone, *newzone;
413 int dirty;
414 int expected_count = 1 + extra_count;
415 void **pslot;
416
417 if (!mapping) {
418 /* Anonymous page without mapping */
419 if (page_count(page) != expected_count)
420 return -EAGAIN;
421
422 /* No turning back from here */
423 newpage->index = page->index;
424 newpage->mapping = page->mapping;
425 if (PageSwapBacked(page))
426 __SetPageSwapBacked(newpage);
427
428 return MIGRATEPAGE_SUCCESS;
429 }
430
431 oldzone = page_zone(page);
432 newzone = page_zone(newpage);
433
434 spin_lock_irq(&mapping->tree_lock);
435
436 pslot = radix_tree_lookup_slot(&mapping->page_tree,
437 page_index(page));
438
439 expected_count += 1 + page_has_private(page);
440 if (page_count(page) != expected_count ||
441 radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) {
442 spin_unlock_irq(&mapping->tree_lock);
443 return -EAGAIN;
444 }
445
446 if (!page_ref_freeze(page, expected_count)) {
447 spin_unlock_irq(&mapping->tree_lock);
448 return -EAGAIN;
449 }
450
451 /*
452 * In the async migration case of moving a page with buffers, lock the
453 * buffers using trylock before the mapping is moved. If the mapping
454 * was moved, we later failed to lock the buffers and could not move
455 * the mapping back due to an elevated page count, we would have to
456 * block waiting on other references to be dropped.
457 */
458 if (mode == MIGRATE_ASYNC && head &&
459 !buffer_migrate_lock_buffers(head, mode)) {
460 page_ref_unfreeze(page, expected_count);
461 spin_unlock_irq(&mapping->tree_lock);
462 return -EAGAIN;
463 }
464
465 /*
466 * Now we know that no one else is looking at the page:
467 * no turning back from here.
468 */
469 newpage->index = page->index;
470 newpage->mapping = page->mapping;
471 if (PageSwapBacked(page))
472 __SetPageSwapBacked(newpage);
473
474 get_page(newpage); /* add cache reference */
475 if (PageSwapCache(page)) {
476 SetPageSwapCache(newpage);
477 set_page_private(newpage, page_private(page));
478 }
479
480 /* Move dirty while page refs frozen and newpage not yet exposed */
481 dirty = PageDirty(page);
482 if (dirty) {
483 ClearPageDirty(page);
484 SetPageDirty(newpage);
485 }
486
487 radix_tree_replace_slot(pslot, newpage);
488
489 /*
490 * Drop cache reference from old page by unfreezing
491 * to one less reference.
492 * We know this isn't the last reference.
493 */
494 page_ref_unfreeze(page, expected_count - 1);
495
496 spin_unlock(&mapping->tree_lock);
497 /* Leave irq disabled to prevent preemption while updating stats */
498
499 /*
500 * If moved to a different zone then also account
501 * the page for that zone. Other VM counters will be
502 * taken care of when we establish references to the
503 * new page and drop references to the old page.
504 *
505 * Note that anonymous pages are accounted for
506 * via NR_FILE_PAGES and NR_ANON_PAGES if they
507 * are mapped to swap space.
508 */
509 if (newzone != oldzone) {
510 __dec_zone_state(oldzone, NR_FILE_PAGES);
511 __inc_zone_state(newzone, NR_FILE_PAGES);
512 if (PageSwapBacked(page) && !PageSwapCache(page)) {
513 __dec_zone_state(oldzone, NR_SHMEM);
514 __inc_zone_state(newzone, NR_SHMEM);
515 }
516 if (dirty && mapping_cap_account_dirty(mapping)) {
517 __dec_zone_state(oldzone, NR_FILE_DIRTY);
518 __inc_zone_state(newzone, NR_FILE_DIRTY);
519 }
520 }
521 local_irq_enable();
522
523 return MIGRATEPAGE_SUCCESS;
524 }
525 EXPORT_SYMBOL(migrate_page_move_mapping);
526
527 /*
528 * The expected number of remaining references is the same as that
529 * of migrate_page_move_mapping().
530 */
531 int migrate_huge_page_move_mapping(struct address_space *mapping,
532 struct page *newpage, struct page *page)
533 {
534 int expected_count;
535 void **pslot;
536
537 spin_lock_irq(&mapping->tree_lock);
538
539 pslot = radix_tree_lookup_slot(&mapping->page_tree,
540 page_index(page));
541
542 expected_count = 2 + page_has_private(page);
543 if (page_count(page) != expected_count ||
544 radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) {
545 spin_unlock_irq(&mapping->tree_lock);
546 return -EAGAIN;
547 }
548
549 if (!page_ref_freeze(page, expected_count)) {
550 spin_unlock_irq(&mapping->tree_lock);
551 return -EAGAIN;
552 }
553
554 newpage->index = page->index;
555 newpage->mapping = page->mapping;
556
557 get_page(newpage);
558
559 radix_tree_replace_slot(pslot, newpage);
560
561 page_ref_unfreeze(page, expected_count - 1);
562
563 spin_unlock_irq(&mapping->tree_lock);
564
565 return MIGRATEPAGE_SUCCESS;
566 }
567
568 /*
569 * Gigantic pages are so large that we do not guarantee that page++ pointer
570 * arithmetic will work across the entire page. We need something more
571 * specialized.
572 */
573 static void __copy_gigantic_page(struct page *dst, struct page *src,
574 int nr_pages)
575 {
576 int i;
577 struct page *dst_base = dst;
578 struct page *src_base = src;
579
580 for (i = 0; i < nr_pages; ) {
581 cond_resched();
582 copy_highpage(dst, src);
583
584 i++;
585 dst = mem_map_next(dst, dst_base, i);
586 src = mem_map_next(src, src_base, i);
587 }
588 }
589
590 static void copy_huge_page(struct page *dst, struct page *src)
591 {
592 int i;
593 int nr_pages;
594
595 if (PageHuge(src)) {
596 /* hugetlbfs page */
597 struct hstate *h = page_hstate(src);
598 nr_pages = pages_per_huge_page(h);
599
600 if (unlikely(nr_pages > MAX_ORDER_NR_PAGES)) {
601 __copy_gigantic_page(dst, src, nr_pages);
602 return;
603 }
604 } else {
605 /* thp page */
606 BUG_ON(!PageTransHuge(src));
607 nr_pages = hpage_nr_pages(src);
608 }
609
610 for (i = 0; i < nr_pages; i++) {
611 cond_resched();
612 copy_highpage(dst + i, src + i);
613 }
614 }
615
616 /*
617 * Copy the page to its new location
618 */
619 void migrate_page_copy(struct page *newpage, struct page *page)
620 {
621 int cpupid;
622
623 if (PageHuge(page) || PageTransHuge(page))
624 copy_huge_page(newpage, page);
625 else
626 copy_highpage(newpage, page);
627
628 if (PageError(page))
629 SetPageError(newpage);
630 if (PageReferenced(page))
631 SetPageReferenced(newpage);
632 if (PageUptodate(page))
633 SetPageUptodate(newpage);
634 if (TestClearPageActive(page)) {
635 VM_BUG_ON_PAGE(PageUnevictable(page), page);
636 SetPageActive(newpage);
637 } else if (TestClearPageUnevictable(page))
638 SetPageUnevictable(newpage);
639 if (PageChecked(page))
640 SetPageChecked(newpage);
641 if (PageMappedToDisk(page))
642 SetPageMappedToDisk(newpage);
643
644 /* Move dirty on pages not done by migrate_page_move_mapping() */
645 if (PageDirty(page))
646 SetPageDirty(newpage);
647
648 if (page_is_young(page))
649 set_page_young(newpage);
650 if (page_is_idle(page))
651 set_page_idle(newpage);
652
653 /*
654 * Copy NUMA information to the new page, to prevent over-eager
655 * future migrations of this same page.
656 */
657 cpupid = page_cpupid_xchg_last(page, -1);
658 page_cpupid_xchg_last(newpage, cpupid);
659
660 ksm_migrate_page(newpage, page);
661 /*
662 * Please do not reorder this without considering how mm/ksm.c's
663 * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache().
664 */
665 if (PageSwapCache(page))
666 ClearPageSwapCache(page);
667 ClearPagePrivate(page);
668 set_page_private(page, 0);
669
670 /*
671 * If any waiters have accumulated on the new page then
672 * wake them up.
673 */
674 if (PageWriteback(newpage))
675 end_page_writeback(newpage);
676
677 copy_page_owner(page, newpage);
678
679 mem_cgroup_migrate(page, newpage);
680 }
681 EXPORT_SYMBOL(migrate_page_copy);
682
683 /************************************************************
684 * Migration functions
685 ***********************************************************/
686
687 /*
688 * Common logic to directly migrate a single LRU page suitable for
689 * pages that do not use PagePrivate/PagePrivate2.
690 *
691 * Pages are locked upon entry and exit.
692 */
693 int migrate_page(struct address_space *mapping,
694 struct page *newpage, struct page *page,
695 enum migrate_mode mode)
696 {
697 int rc;
698
699 BUG_ON(PageWriteback(page)); /* Writeback must be complete */
700
701 rc = migrate_page_move_mapping(mapping, newpage, page, NULL, mode, 0);
702
703 if (rc != MIGRATEPAGE_SUCCESS)
704 return rc;
705
706 migrate_page_copy(newpage, page);
707 return MIGRATEPAGE_SUCCESS;
708 }
709 EXPORT_SYMBOL(migrate_page);
710
711 #ifdef CONFIG_BLOCK
712 /*
713 * Migration function for pages with buffers. This function can only be used
714 * if the underlying filesystem guarantees that no other references to "page"
715 * exist.
716 */
717 int buffer_migrate_page(struct address_space *mapping,
718 struct page *newpage, struct page *page, enum migrate_mode mode)
719 {
720 struct buffer_head *bh, *head;
721 int rc;
722
723 if (!page_has_buffers(page))
724 return migrate_page(mapping, newpage, page, mode);
725
726 head = page_buffers(page);
727
728 rc = migrate_page_move_mapping(mapping, newpage, page, head, mode, 0);
729
730 if (rc != MIGRATEPAGE_SUCCESS)
731 return rc;
732
733 /*
734 * In the async case, migrate_page_move_mapping locked the buffers
735 * with an IRQ-safe spinlock held. In the sync case, the buffers
736 * need to be locked now
737 */
738 if (mode != MIGRATE_ASYNC)
739 BUG_ON(!buffer_migrate_lock_buffers(head, mode));
740
741 ClearPagePrivate(page);
742 set_page_private(newpage, page_private(page));
743 set_page_private(page, 0);
744 put_page(page);
745 get_page(newpage);
746
747 bh = head;
748 do {
749 set_bh_page(bh, newpage, bh_offset(bh));
750 bh = bh->b_this_page;
751
752 } while (bh != head);
753
754 SetPagePrivate(newpage);
755
756 migrate_page_copy(newpage, page);
757
758 bh = head;
759 do {
760 unlock_buffer(bh);
761 put_bh(bh);
762 bh = bh->b_this_page;
763
764 } while (bh != head);
765
766 return MIGRATEPAGE_SUCCESS;
767 }
768 EXPORT_SYMBOL(buffer_migrate_page);
769 #endif
770
771 /*
772 * Writeback a page to clean the dirty state
773 */
774 static int writeout(struct address_space *mapping, struct page *page)
775 {
776 struct writeback_control wbc = {
777 .sync_mode = WB_SYNC_NONE,
778 .nr_to_write = 1,
779 .range_start = 0,
780 .range_end = LLONG_MAX,
781 .for_reclaim = 1
782 };
783 int rc;
784
785 if (!mapping->a_ops->writepage)
786 /* No write method for the address space */
787 return -EINVAL;
788
789 if (!clear_page_dirty_for_io(page))
790 /* Someone else already triggered a write */
791 return -EAGAIN;
792
793 /*
794 * A dirty page may imply that the underlying filesystem has
795 * the page on some queue. So the page must be clean for
796 * migration. Writeout may mean we loose the lock and the
797 * page state is no longer what we checked for earlier.
798 * At this point we know that the migration attempt cannot
799 * be successful.
800 */
801 remove_migration_ptes(page, page, false);
802
803 rc = mapping->a_ops->writepage(page, &wbc);
804
805 if (rc != AOP_WRITEPAGE_ACTIVATE)
806 /* unlocked. Relock */
807 lock_page(page);
808
809 return (rc < 0) ? -EIO : -EAGAIN;
810 }
811
812 /*
813 * Default handling if a filesystem does not provide a migration function.
814 */
815 static int fallback_migrate_page(struct address_space *mapping,
816 struct page *newpage, struct page *page, enum migrate_mode mode)
817 {
818 if (PageDirty(page)) {
819 /* Only writeback pages in full synchronous migration */
820 if (mode != MIGRATE_SYNC)
821 return -EBUSY;
822 return writeout(mapping, page);
823 }
824
825 /*
826 * Buffers may be managed in a filesystem specific way.
827 * We must have no buffers or drop them.
828 */
829 if (page_has_private(page) &&
830 !try_to_release_page(page, GFP_KERNEL))
831 return -EAGAIN;
832
833 return migrate_page(mapping, newpage, page, mode);
834 }
835
836 /*
837 * Move a page to a newly allocated page
838 * The page is locked and all ptes have been successfully removed.
839 *
840 * The new page will have replaced the old page if this function
841 * is successful.
842 *
843 * Return value:
844 * < 0 - error code
845 * MIGRATEPAGE_SUCCESS - success
846 */
847 static int move_to_new_page(struct page *newpage, struct page *page,
848 enum migrate_mode mode)
849 {
850 struct address_space *mapping;
851 int rc = -EAGAIN;
852 bool is_lru = !__PageMovable(page);
853
854 VM_BUG_ON_PAGE(!PageLocked(page), page);
855 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
856
857 mapping = page_mapping(page);
858
859 if (likely(is_lru)) {
860 if (!mapping)
861 rc = migrate_page(mapping, newpage, page, mode);
862 else if (mapping->a_ops->migratepage)
863 /*
864 * Most pages have a mapping and most filesystems
865 * provide a migratepage callback. Anonymous pages
866 * are part of swap space which also has its own
867 * migratepage callback. This is the most common path
868 * for page migration.
869 */
870 rc = mapping->a_ops->migratepage(mapping, newpage,
871 page, mode);
872 else
873 rc = fallback_migrate_page(mapping, newpage,
874 page, mode);
875 } else {
876 /*
877 * In case of non-lru page, it could be released after
878 * isolation step. In that case, we shouldn't try migration.
879 */
880 VM_BUG_ON_PAGE(!PageIsolated(page), page);
881 if (!PageMovable(page)) {
882 rc = MIGRATEPAGE_SUCCESS;
883 __ClearPageIsolated(page);
884 goto out;
885 }
886
887 rc = mapping->a_ops->migratepage(mapping, newpage,
888 page, mode);
889 WARN_ON_ONCE(rc == MIGRATEPAGE_SUCCESS &&
890 !PageIsolated(page));
891 }
892
893 /*
894 * When successful, old pagecache page->mapping must be cleared before
895 * page is freed; but stats require that PageAnon be left as PageAnon.
896 */
897 if (rc == MIGRATEPAGE_SUCCESS) {
898 if (__PageMovable(page)) {
899 VM_BUG_ON_PAGE(!PageIsolated(page), page);
900
901 /*
902 * We clear PG_movable under page_lock so any compactor
903 * cannot try to migrate this page.
904 */
905 __ClearPageIsolated(page);
906 }
907
908 /*
909 * Anonymous and movable page->mapping will be cleard by
910 * free_pages_prepare so don't reset it here for keeping
911 * the type to work PageAnon, for example.
912 */
913 if (!PageMappingFlags(page))
914 page->mapping = NULL;
915 }
916 out:
917 return rc;
918 }
919
920 static int __unmap_and_move(struct page *page, struct page *newpage,
921 int force, enum migrate_mode mode)
922 {
923 int rc = -EAGAIN;
924 int page_was_mapped = 0;
925 struct anon_vma *anon_vma = NULL;
926 bool is_lru = !__PageMovable(page);
927
928 if (!trylock_page(page)) {
929 if (!force || mode == MIGRATE_ASYNC)
930 goto out;
931
932 /*
933 * It's not safe for direct compaction to call lock_page.
934 * For example, during page readahead pages are added locked
935 * to the LRU. Later, when the IO completes the pages are
936 * marked uptodate and unlocked. However, the queueing
937 * could be merging multiple pages for one bio (e.g.
938 * mpage_readpages). If an allocation happens for the
939 * second or third page, the process can end up locking
940 * the same page twice and deadlocking. Rather than
941 * trying to be clever about what pages can be locked,
942 * avoid the use of lock_page for direct compaction
943 * altogether.
944 */
945 if (current->flags & PF_MEMALLOC)
946 goto out;
947
948 lock_page(page);
949 }
950
951 if (PageWriteback(page)) {
952 /*
953 * Only in the case of a full synchronous migration is it
954 * necessary to wait for PageWriteback. In the async case,
955 * the retry loop is too short and in the sync-light case,
956 * the overhead of stalling is too much
957 */
958 if (mode != MIGRATE_SYNC) {
959 rc = -EBUSY;
960 goto out_unlock;
961 }
962 if (!force)
963 goto out_unlock;
964 wait_on_page_writeback(page);
965 }
966
967 /*
968 * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
969 * we cannot notice that anon_vma is freed while we migrates a page.
970 * This get_anon_vma() delays freeing anon_vma pointer until the end
971 * of migration. File cache pages are no problem because of page_lock()
972 * File Caches may use write_page() or lock_page() in migration, then,
973 * just care Anon page here.
974 *
975 * Only page_get_anon_vma() understands the subtleties of
976 * getting a hold on an anon_vma from outside one of its mms.
977 * But if we cannot get anon_vma, then we won't need it anyway,
978 * because that implies that the anon page is no longer mapped
979 * (and cannot be remapped so long as we hold the page lock).
980 */
981 if (PageAnon(page) && !PageKsm(page))
982 anon_vma = page_get_anon_vma(page);
983
984 /*
985 * Block others from accessing the new page when we get around to
986 * establishing additional references. We are usually the only one
987 * holding a reference to newpage at this point. We used to have a BUG
988 * here if trylock_page(newpage) fails, but would like to allow for
989 * cases where there might be a race with the previous use of newpage.
990 * This is much like races on refcount of oldpage: just don't BUG().
991 */
992 if (unlikely(!trylock_page(newpage)))
993 goto out_unlock;
994
995 if (unlikely(isolated_balloon_page(page))) {
996 /*
997 * A ballooned page does not need any special attention from
998 * physical to virtual reverse mapping procedures.
999 * Skip any attempt to unmap PTEs or to remap swap cache,
1000 * in order to avoid burning cycles at rmap level, and perform
1001 * the page migration right away (proteced by page lock).
1002 */
1003 rc = balloon_page_migrate(newpage, page, mode);
1004 goto out_unlock_both;
1005 }
1006
1007 if (unlikely(!is_lru)) {
1008 rc = move_to_new_page(newpage, page, mode);
1009 goto out_unlock_both;
1010 }
1011
1012 /*
1013 * Corner case handling:
1014 * 1. When a new swap-cache page is read into, it is added to the LRU
1015 * and treated as swapcache but it has no rmap yet.
1016 * Calling try_to_unmap() against a page->mapping==NULL page will
1017 * trigger a BUG. So handle it here.
1018 * 2. An orphaned page (see truncate_complete_page) might have
1019 * fs-private metadata. The page can be picked up due to memory
1020 * offlining. Everywhere else except page reclaim, the page is
1021 * invisible to the vm, so the page can not be migrated. So try to
1022 * free the metadata, so the page can be freed.
1023 */
1024 if (!page->mapping) {
1025 VM_BUG_ON_PAGE(PageAnon(page), page);
1026 if (page_has_private(page)) {
1027 try_to_free_buffers(page);
1028 goto out_unlock_both;
1029 }
1030 } else if (page_mapped(page)) {
1031 /* Establish migration ptes */
1032 VM_BUG_ON_PAGE(PageAnon(page) && !PageKsm(page) && !anon_vma,
1033 page);
1034 try_to_unmap(page,
1035 TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
1036 page_was_mapped = 1;
1037 }
1038
1039 if (!page_mapped(page))
1040 rc = move_to_new_page(newpage, page, mode);
1041
1042 if (page_was_mapped)
1043 remove_migration_ptes(page,
1044 rc == MIGRATEPAGE_SUCCESS ? newpage : page, false);
1045
1046 out_unlock_both:
1047 unlock_page(newpage);
1048 out_unlock:
1049 /* Drop an anon_vma reference if we took one */
1050 if (anon_vma)
1051 put_anon_vma(anon_vma);
1052 unlock_page(page);
1053 out:
1054 /*
1055 * If migration is successful, decrease refcount of the newpage
1056 * which will not free the page because new page owner increased
1057 * refcounter. As well, if it is LRU page, add the page to LRU
1058 * list in here.
1059 */
1060 if (rc == MIGRATEPAGE_SUCCESS) {
1061 if (unlikely(__is_movable_balloon_page(newpage) ||
1062 __PageMovable(newpage)))
1063 put_page(newpage);
1064 else
1065 putback_lru_page(newpage);
1066 }
1067
1068 return rc;
1069 }
1070
1071 /*
1072 * gcc 4.7 and 4.8 on arm get an ICEs when inlining unmap_and_move(). Work
1073 * around it.
1074 */
1075 #if (GCC_VERSION >= 40700 && GCC_VERSION < 40900) && defined(CONFIG_ARM)
1076 #define ICE_noinline noinline
1077 #else
1078 #define ICE_noinline
1079 #endif
1080
1081 /*
1082 * Obtain the lock on page, remove all ptes and migrate the page
1083 * to the newly allocated page in newpage.
1084 */
1085 static ICE_noinline int unmap_and_move(new_page_t get_new_page,
1086 free_page_t put_new_page,
1087 unsigned long private, struct page *page,
1088 int force, enum migrate_mode mode,
1089 enum migrate_reason reason)
1090 {
1091 int rc = MIGRATEPAGE_SUCCESS;
1092 int *result = NULL;
1093 struct page *newpage;
1094
1095 newpage = get_new_page(page, private, &result);
1096 if (!newpage)
1097 return -ENOMEM;
1098
1099 if (page_count(page) == 1) {
1100 /* page was freed from under us. So we are done. */
1101 ClearPageActive(page);
1102 ClearPageUnevictable(page);
1103 if (unlikely(__PageMovable(page))) {
1104 lock_page(page);
1105 if (!PageMovable(page))
1106 __ClearPageIsolated(page);
1107 unlock_page(page);
1108 }
1109 if (put_new_page)
1110 put_new_page(newpage, private);
1111 else
1112 put_page(newpage);
1113 goto out;
1114 }
1115
1116 if (unlikely(PageTransHuge(page))) {
1117 lock_page(page);
1118 rc = split_huge_page(page);
1119 unlock_page(page);
1120 if (rc)
1121 goto out;
1122 }
1123
1124 rc = __unmap_and_move(page, newpage, force, mode);
1125 if (rc == MIGRATEPAGE_SUCCESS)
1126 set_page_owner_migrate_reason(newpage, reason);
1127
1128 out:
1129 if (rc != -EAGAIN) {
1130 /*
1131 * A page that has been migrated has all references
1132 * removed and will be freed. A page that has not been
1133 * migrated will have kepts its references and be
1134 * restored.
1135 */
1136 list_del(&page->lru);
1137 dec_zone_page_state(page, NR_ISOLATED_ANON +
1138 page_is_file_cache(page));
1139 }
1140
1141 /*
1142 * If migration is successful, releases reference grabbed during
1143 * isolation. Otherwise, restore the page to right list unless
1144 * we want to retry.
1145 */
1146 if (rc == MIGRATEPAGE_SUCCESS) {
1147 put_page(page);
1148 if (reason == MR_MEMORY_FAILURE) {
1149 /*
1150 * Set PG_HWPoison on just freed page
1151 * intentionally. Although it's rather weird,
1152 * it's how HWPoison flag works at the moment.
1153 */
1154 if (!test_set_page_hwpoison(page))
1155 num_poisoned_pages_inc();
1156 }
1157 } else {
1158 if (rc != -EAGAIN) {
1159 if (likely(!__PageMovable(page))) {
1160 putback_lru_page(page);
1161 goto put_new;
1162 }
1163
1164 lock_page(page);
1165 if (PageMovable(page))
1166 putback_movable_page(page);
1167 else
1168 __ClearPageIsolated(page);
1169 unlock_page(page);
1170 put_page(page);
1171 }
1172 put_new:
1173 if (put_new_page)
1174 put_new_page(newpage, private);
1175 else
1176 put_page(newpage);
1177 }
1178
1179 if (result) {
1180 if (rc)
1181 *result = rc;
1182 else
1183 *result = page_to_nid(newpage);
1184 }
1185 return rc;
1186 }
1187
1188 /*
1189 * Counterpart of unmap_and_move_page() for hugepage migration.
1190 *
1191 * This function doesn't wait the completion of hugepage I/O
1192 * because there is no race between I/O and migration for hugepage.
1193 * Note that currently hugepage I/O occurs only in direct I/O
1194 * where no lock is held and PG_writeback is irrelevant,
1195 * and writeback status of all subpages are counted in the reference
1196 * count of the head page (i.e. if all subpages of a 2MB hugepage are
1197 * under direct I/O, the reference of the head page is 512 and a bit more.)
1198 * This means that when we try to migrate hugepage whose subpages are
1199 * doing direct I/O, some references remain after try_to_unmap() and
1200 * hugepage migration fails without data corruption.
1201 *
1202 * There is also no race when direct I/O is issued on the page under migration,
1203 * because then pte is replaced with migration swap entry and direct I/O code
1204 * will wait in the page fault for migration to complete.
1205 */
1206 static int unmap_and_move_huge_page(new_page_t get_new_page,
1207 free_page_t put_new_page, unsigned long private,
1208 struct page *hpage, int force,
1209 enum migrate_mode mode, int reason)
1210 {
1211 int rc = -EAGAIN;
1212 int *result = NULL;
1213 int page_was_mapped = 0;
1214 struct page *new_hpage;
1215 struct anon_vma *anon_vma = NULL;
1216
1217 /*
1218 * Movability of hugepages depends on architectures and hugepage size.
1219 * This check is necessary because some callers of hugepage migration
1220 * like soft offline and memory hotremove don't walk through page
1221 * tables or check whether the hugepage is pmd-based or not before
1222 * kicking migration.
1223 */
1224 if (!hugepage_migration_supported(page_hstate(hpage))) {
1225 putback_active_hugepage(hpage);
1226 return -ENOSYS;
1227 }
1228
1229 new_hpage = get_new_page(hpage, private, &result);
1230 if (!new_hpage)
1231 return -ENOMEM;
1232
1233 if (!trylock_page(hpage)) {
1234 if (!force || mode != MIGRATE_SYNC)
1235 goto out;
1236 lock_page(hpage);
1237 }
1238
1239 if (PageAnon(hpage))
1240 anon_vma = page_get_anon_vma(hpage);
1241
1242 if (unlikely(!trylock_page(new_hpage)))
1243 goto put_anon;
1244
1245 if (page_mapped(hpage)) {
1246 try_to_unmap(hpage,
1247 TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
1248 page_was_mapped = 1;
1249 }
1250
1251 if (!page_mapped(hpage))
1252 rc = move_to_new_page(new_hpage, hpage, mode);
1253
1254 if (page_was_mapped)
1255 remove_migration_ptes(hpage,
1256 rc == MIGRATEPAGE_SUCCESS ? new_hpage : hpage, false);
1257
1258 unlock_page(new_hpage);
1259
1260 put_anon:
1261 if (anon_vma)
1262 put_anon_vma(anon_vma);
1263
1264 if (rc == MIGRATEPAGE_SUCCESS) {
1265 hugetlb_cgroup_migrate(hpage, new_hpage);
1266 put_new_page = NULL;
1267 set_page_owner_migrate_reason(new_hpage, reason);
1268 }
1269
1270 unlock_page(hpage);
1271 out:
1272 if (rc != -EAGAIN)
1273 putback_active_hugepage(hpage);
1274
1275 /*
1276 * If migration was not successful and there's a freeing callback, use
1277 * it. Otherwise, put_page() will drop the reference grabbed during
1278 * isolation.
1279 */
1280 if (put_new_page)
1281 put_new_page(new_hpage, private);
1282 else
1283 putback_active_hugepage(new_hpage);
1284
1285 if (result) {
1286 if (rc)
1287 *result = rc;
1288 else
1289 *result = page_to_nid(new_hpage);
1290 }
1291 return rc;
1292 }
1293
1294 /*
1295 * migrate_pages - migrate the pages specified in a list, to the free pages
1296 * supplied as the target for the page migration
1297 *
1298 * @from: The list of pages to be migrated.
1299 * @get_new_page: The function used to allocate free pages to be used
1300 * as the target of the page migration.
1301 * @put_new_page: The function used to free target pages if migration
1302 * fails, or NULL if no special handling is necessary.
1303 * @private: Private data to be passed on to get_new_page()
1304 * @mode: The migration mode that specifies the constraints for
1305 * page migration, if any.
1306 * @reason: The reason for page migration.
1307 *
1308 * The function returns after 10 attempts or if no pages are movable any more
1309 * because the list has become empty or no retryable pages exist any more.
1310 * The caller should call putback_movable_pages() to return pages to the LRU
1311 * or free list only if ret != 0.
1312 *
1313 * Returns the number of pages that were not migrated, or an error code.
1314 */
1315 int migrate_pages(struct list_head *from, new_page_t get_new_page,
1316 free_page_t put_new_page, unsigned long private,
1317 enum migrate_mode mode, int reason)
1318 {
1319 int retry = 1;
1320 int nr_failed = 0;
1321 int nr_succeeded = 0;
1322 int pass = 0;
1323 struct page *page;
1324 struct page *page2;
1325 int swapwrite = current->flags & PF_SWAPWRITE;
1326 int rc;
1327
1328 if (!swapwrite)
1329 current->flags |= PF_SWAPWRITE;
1330
1331 for(pass = 0; pass < 10 && retry; pass++) {
1332 retry = 0;
1333
1334 list_for_each_entry_safe(page, page2, from, lru) {
1335 cond_resched();
1336
1337 if (PageHuge(page))
1338 rc = unmap_and_move_huge_page(get_new_page,
1339 put_new_page, private, page,
1340 pass > 2, mode, reason);
1341 else
1342 rc = unmap_and_move(get_new_page, put_new_page,
1343 private, page, pass > 2, mode,
1344 reason);
1345
1346 switch(rc) {
1347 case -ENOMEM:
1348 nr_failed++;
1349 goto out;
1350 case -EAGAIN:
1351 retry++;
1352 break;
1353 case MIGRATEPAGE_SUCCESS:
1354 nr_succeeded++;
1355 break;
1356 default:
1357 /*
1358 * Permanent failure (-EBUSY, -ENOSYS, etc.):
1359 * unlike -EAGAIN case, the failed page is
1360 * removed from migration page list and not
1361 * retried in the next outer loop.
1362 */
1363 nr_failed++;
1364 break;
1365 }
1366 }
1367 }
1368 nr_failed += retry;
1369 rc = nr_failed;
1370 out:
1371 if (nr_succeeded)
1372 count_vm_events(PGMIGRATE_SUCCESS, nr_succeeded);
1373 if (nr_failed)
1374 count_vm_events(PGMIGRATE_FAIL, nr_failed);
1375 trace_mm_migrate_pages(nr_succeeded, nr_failed, mode, reason);
1376
1377 if (!swapwrite)
1378 current->flags &= ~PF_SWAPWRITE;
1379
1380 return rc;
1381 }
1382
1383 #ifdef CONFIG_NUMA
1384 /*
1385 * Move a list of individual pages
1386 */
1387 struct page_to_node {
1388 unsigned long addr;
1389 struct page *page;
1390 int node;
1391 int status;
1392 };
1393
1394 static struct page *new_page_node(struct page *p, unsigned long private,
1395 int **result)
1396 {
1397 struct page_to_node *pm = (struct page_to_node *)private;
1398
1399 while (pm->node != MAX_NUMNODES && pm->page != p)
1400 pm++;
1401
1402 if (pm->node == MAX_NUMNODES)
1403 return NULL;
1404
1405 *result = &pm->status;
1406
1407 if (PageHuge(p))
1408 return alloc_huge_page_node(page_hstate(compound_head(p)),
1409 pm->node);
1410 else
1411 return __alloc_pages_node(pm->node,
1412 GFP_HIGHUSER_MOVABLE | __GFP_THISNODE, 0);
1413 }
1414
1415 /*
1416 * Move a set of pages as indicated in the pm array. The addr
1417 * field must be set to the virtual address of the page to be moved
1418 * and the node number must contain a valid target node.
1419 * The pm array ends with node = MAX_NUMNODES.
1420 */
1421 static int do_move_page_to_node_array(struct mm_struct *mm,
1422 struct page_to_node *pm,
1423 int migrate_all)
1424 {
1425 int err;
1426 struct page_to_node *pp;
1427 LIST_HEAD(pagelist);
1428
1429 down_read(&mm->mmap_sem);
1430
1431 /*
1432 * Build a list of pages to migrate
1433 */
1434 for (pp = pm; pp->node != MAX_NUMNODES; pp++) {
1435 struct vm_area_struct *vma;
1436 struct page *page;
1437
1438 err = -EFAULT;
1439 vma = find_vma(mm, pp->addr);
1440 if (!vma || pp->addr < vma->vm_start || !vma_migratable(vma))
1441 goto set_status;
1442
1443 /* FOLL_DUMP to ignore special (like zero) pages */
1444 page = follow_page(vma, pp->addr,
1445 FOLL_GET | FOLL_SPLIT | FOLL_DUMP);
1446
1447 err = PTR_ERR(page);
1448 if (IS_ERR(page))
1449 goto set_status;
1450
1451 err = -ENOENT;
1452 if (!page)
1453 goto set_status;
1454
1455 pp->page = page;
1456 err = page_to_nid(page);
1457
1458 if (err == pp->node)
1459 /*
1460 * Node already in the right place
1461 */
1462 goto put_and_set;
1463
1464 err = -EACCES;
1465 if (page_mapcount(page) > 1 &&
1466 !migrate_all)
1467 goto put_and_set;
1468
1469 if (PageHuge(page)) {
1470 if (PageHead(page))
1471 isolate_huge_page(page, &pagelist);
1472 goto put_and_set;
1473 }
1474
1475 err = isolate_lru_page(page);
1476 if (!err) {
1477 list_add_tail(&page->lru, &pagelist);
1478 inc_zone_page_state(page, NR_ISOLATED_ANON +
1479 page_is_file_cache(page));
1480 }
1481 put_and_set:
1482 /*
1483 * Either remove the duplicate refcount from
1484 * isolate_lru_page() or drop the page ref if it was
1485 * not isolated.
1486 */
1487 put_page(page);
1488 set_status:
1489 pp->status = err;
1490 }
1491
1492 err = 0;
1493 if (!list_empty(&pagelist)) {
1494 err = migrate_pages(&pagelist, new_page_node, NULL,
1495 (unsigned long)pm, MIGRATE_SYNC, MR_SYSCALL);
1496 if (err)
1497 putback_movable_pages(&pagelist);
1498 }
1499
1500 up_read(&mm->mmap_sem);
1501 return err;
1502 }
1503
1504 /*
1505 * Migrate an array of page address onto an array of nodes and fill
1506 * the corresponding array of status.
1507 */
1508 static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes,
1509 unsigned long nr_pages,
1510 const void __user * __user *pages,
1511 const int __user *nodes,
1512 int __user *status, int flags)
1513 {
1514 struct page_to_node *pm;
1515 unsigned long chunk_nr_pages;
1516 unsigned long chunk_start;
1517 int err;
1518
1519 err = -ENOMEM;
1520 pm = (struct page_to_node *)__get_free_page(GFP_KERNEL);
1521 if (!pm)
1522 goto out;
1523
1524 migrate_prep();
1525
1526 /*
1527 * Store a chunk of page_to_node array in a page,
1528 * but keep the last one as a marker
1529 */
1530 chunk_nr_pages = (PAGE_SIZE / sizeof(struct page_to_node)) - 1;
1531
1532 for (chunk_start = 0;
1533 chunk_start < nr_pages;
1534 chunk_start += chunk_nr_pages) {
1535 int j;
1536
1537 if (chunk_start + chunk_nr_pages > nr_pages)
1538 chunk_nr_pages = nr_pages - chunk_start;
1539
1540 /* fill the chunk pm with addrs and nodes from user-space */
1541 for (j = 0; j < chunk_nr_pages; j++) {
1542 const void __user *p;
1543 int node;
1544
1545 err = -EFAULT;
1546 if (get_user(p, pages + j + chunk_start))
1547 goto out_pm;
1548 pm[j].addr = (unsigned long) p;
1549
1550 if (get_user(node, nodes + j + chunk_start))
1551 goto out_pm;
1552
1553 err = -ENODEV;
1554 if (node < 0 || node >= MAX_NUMNODES)
1555 goto out_pm;
1556
1557 if (!node_state(node, N_MEMORY))
1558 goto out_pm;
1559
1560 err = -EACCES;
1561 if (!node_isset(node, task_nodes))
1562 goto out_pm;
1563
1564 pm[j].node = node;
1565 }
1566
1567 /* End marker for this chunk */
1568 pm[chunk_nr_pages].node = MAX_NUMNODES;
1569
1570 /* Migrate this chunk */
1571 err = do_move_page_to_node_array(mm, pm,
1572 flags & MPOL_MF_MOVE_ALL);
1573 if (err < 0)
1574 goto out_pm;
1575
1576 /* Return status information */
1577 for (j = 0; j < chunk_nr_pages; j++)
1578 if (put_user(pm[j].status, status + j + chunk_start)) {
1579 err = -EFAULT;
1580 goto out_pm;
1581 }
1582 }
1583 err = 0;
1584
1585 out_pm:
1586 free_page((unsigned long)pm);
1587 out:
1588 return err;
1589 }
1590
1591 /*
1592 * Determine the nodes of an array of pages and store it in an array of status.
1593 */
1594 static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages,
1595 const void __user **pages, int *status)
1596 {
1597 unsigned long i;
1598
1599 down_read(&mm->mmap_sem);
1600
1601 for (i = 0; i < nr_pages; i++) {
1602 unsigned long addr = (unsigned long)(*pages);
1603 struct vm_area_struct *vma;
1604 struct page *page;
1605 int err = -EFAULT;
1606
1607 vma = find_vma(mm, addr);
1608 if (!vma || addr < vma->vm_start)
1609 goto set_status;
1610
1611 /* FOLL_DUMP to ignore special (like zero) pages */
1612 page = follow_page(vma, addr, FOLL_DUMP);
1613
1614 err = PTR_ERR(page);
1615 if (IS_ERR(page))
1616 goto set_status;
1617
1618 err = page ? page_to_nid(page) : -ENOENT;
1619 set_status:
1620 *status = err;
1621
1622 pages++;
1623 status++;
1624 }
1625
1626 up_read(&mm->mmap_sem);
1627 }
1628
1629 /*
1630 * Determine the nodes of a user array of pages and store it in
1631 * a user array of status.
1632 */
1633 static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages,
1634 const void __user * __user *pages,
1635 int __user *status)
1636 {
1637 #define DO_PAGES_STAT_CHUNK_NR 16
1638 const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR];
1639 int chunk_status[DO_PAGES_STAT_CHUNK_NR];
1640
1641 while (nr_pages) {
1642 unsigned long chunk_nr;
1643
1644 chunk_nr = nr_pages;
1645 if (chunk_nr > DO_PAGES_STAT_CHUNK_NR)
1646 chunk_nr = DO_PAGES_STAT_CHUNK_NR;
1647
1648 if (copy_from_user(chunk_pages, pages, chunk_nr * sizeof(*chunk_pages)))
1649 break;
1650
1651 do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status);
1652
1653 if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status)))
1654 break;
1655
1656 pages += chunk_nr;
1657 status += chunk_nr;
1658 nr_pages -= chunk_nr;
1659 }
1660 return nr_pages ? -EFAULT : 0;
1661 }
1662
1663 /*
1664 * Move a list of pages in the address space of the currently executing
1665 * process.
1666 */
1667 SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages,
1668 const void __user * __user *, pages,
1669 const int __user *, nodes,
1670 int __user *, status, int, flags)
1671 {
1672 const struct cred *cred = current_cred(), *tcred;
1673 struct task_struct *task;
1674 struct mm_struct *mm;
1675 int err;
1676 nodemask_t task_nodes;
1677
1678 /* Check flags */
1679 if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
1680 return -EINVAL;
1681
1682 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1683 return -EPERM;
1684
1685 /* Find the mm_struct */
1686 rcu_read_lock();
1687 task = pid ? find_task_by_vpid(pid) : current;
1688 if (!task) {
1689 rcu_read_unlock();
1690 return -ESRCH;
1691 }
1692 get_task_struct(task);
1693
1694 /*
1695 * Check if this process has the right to modify the specified
1696 * process. The right exists if the process has administrative
1697 * capabilities, superuser privileges or the same
1698 * userid as the target process.
1699 */
1700 tcred = __task_cred(task);
1701 if (!uid_eq(cred->euid, tcred->suid) && !uid_eq(cred->euid, tcred->uid) &&
1702 !uid_eq(cred->uid, tcred->suid) && !uid_eq(cred->uid, tcred->uid) &&
1703 !capable(CAP_SYS_NICE)) {
1704 rcu_read_unlock();
1705 err = -EPERM;
1706 goto out;
1707 }
1708 rcu_read_unlock();
1709
1710 err = security_task_movememory(task);
1711 if (err)
1712 goto out;
1713
1714 task_nodes = cpuset_mems_allowed(task);
1715 mm = get_task_mm(task);
1716 put_task_struct(task);
1717
1718 if (!mm)
1719 return -EINVAL;
1720
1721 if (nodes)
1722 err = do_pages_move(mm, task_nodes, nr_pages, pages,
1723 nodes, status, flags);
1724 else
1725 err = do_pages_stat(mm, nr_pages, pages, status);
1726
1727 mmput(mm);
1728 return err;
1729
1730 out:
1731 put_task_struct(task);
1732 return err;
1733 }
1734
1735 #ifdef CONFIG_NUMA_BALANCING
1736 /*
1737 * Returns true if this is a safe migration target node for misplaced NUMA
1738 * pages. Currently it only checks the watermarks which crude
1739 */
1740 static bool migrate_balanced_pgdat(struct pglist_data *pgdat,
1741 unsigned long nr_migrate_pages)
1742 {
1743 int z;
1744 for (z = pgdat->nr_zones - 1; z >= 0; z--) {
1745 struct zone *zone = pgdat->node_zones + z;
1746
1747 if (!populated_zone(zone))
1748 continue;
1749
1750 if (!zone_reclaimable(zone))
1751 continue;
1752
1753 /* Avoid waking kswapd by allocating pages_to_migrate pages. */
1754 if (!zone_watermark_ok(zone, 0,
1755 high_wmark_pages(zone) +
1756 nr_migrate_pages,
1757 0, 0))
1758 continue;
1759 return true;
1760 }
1761 return false;
1762 }
1763
1764 static struct page *alloc_misplaced_dst_page(struct page *page,
1765 unsigned long data,
1766 int **result)
1767 {
1768 int nid = (int) data;
1769 struct page *newpage;
1770
1771 newpage = __alloc_pages_node(nid,
1772 (GFP_HIGHUSER_MOVABLE |
1773 __GFP_THISNODE | __GFP_NOMEMALLOC |
1774 __GFP_NORETRY | __GFP_NOWARN) &
1775 ~__GFP_RECLAIM, 0);
1776
1777 return newpage;
1778 }
1779
1780 /*
1781 * page migration rate limiting control.
1782 * Do not migrate more than @pages_to_migrate in a @migrate_interval_millisecs
1783 * window of time. Default here says do not migrate more than 1280M per second.
1784 */
1785 static unsigned int migrate_interval_millisecs __read_mostly = 100;
1786 static unsigned int ratelimit_pages __read_mostly = 128 << (20 - PAGE_SHIFT);
1787
1788 /* Returns true if the node is migrate rate-limited after the update */
1789 static bool numamigrate_update_ratelimit(pg_data_t *pgdat,
1790 unsigned long nr_pages)
1791 {
1792 /*
1793 * Rate-limit the amount of data that is being migrated to a node.
1794 * Optimal placement is no good if the memory bus is saturated and
1795 * all the time is being spent migrating!
1796 */
1797 if (time_after(jiffies, pgdat->numabalancing_migrate_next_window)) {
1798 spin_lock(&pgdat->numabalancing_migrate_lock);
1799 pgdat->numabalancing_migrate_nr_pages = 0;
1800 pgdat->numabalancing_migrate_next_window = jiffies +
1801 msecs_to_jiffies(migrate_interval_millisecs);
1802 spin_unlock(&pgdat->numabalancing_migrate_lock);
1803 }
1804 if (pgdat->numabalancing_migrate_nr_pages > ratelimit_pages) {
1805 trace_mm_numa_migrate_ratelimit(current, pgdat->node_id,
1806 nr_pages);
1807 return true;
1808 }
1809
1810 /*
1811 * This is an unlocked non-atomic update so errors are possible.
1812 * The consequences are failing to migrate when we potentiall should
1813 * have which is not severe enough to warrant locking. If it is ever
1814 * a problem, it can be converted to a per-cpu counter.
1815 */
1816 pgdat->numabalancing_migrate_nr_pages += nr_pages;
1817 return false;
1818 }
1819
1820 static int numamigrate_isolate_page(pg_data_t *pgdat, struct page *page)
1821 {
1822 int page_lru;
1823
1824 VM_BUG_ON_PAGE(compound_order(page) && !PageTransHuge(page), page);
1825
1826 /* Avoid migrating to a node that is nearly full */
1827 if (!migrate_balanced_pgdat(pgdat, 1UL << compound_order(page)))
1828 return 0;
1829
1830 if (isolate_lru_page(page))
1831 return 0;
1832
1833 /*
1834 * migrate_misplaced_transhuge_page() skips page migration's usual
1835 * check on page_count(), so we must do it here, now that the page
1836 * has been isolated: a GUP pin, or any other pin, prevents migration.
1837 * The expected page count is 3: 1 for page's mapcount and 1 for the
1838 * caller's pin and 1 for the reference taken by isolate_lru_page().
1839 */
1840 if (PageTransHuge(page) && page_count(page) != 3) {
1841 putback_lru_page(page);
1842 return 0;
1843 }
1844
1845 page_lru = page_is_file_cache(page);
1846 mod_zone_page_state(page_zone(page), NR_ISOLATED_ANON + page_lru,
1847 hpage_nr_pages(page));
1848
1849 /*
1850 * Isolating the page has taken another reference, so the
1851 * caller's reference can be safely dropped without the page
1852 * disappearing underneath us during migration.
1853 */
1854 put_page(page);
1855 return 1;
1856 }
1857
1858 bool pmd_trans_migrating(pmd_t pmd)
1859 {
1860 struct page *page = pmd_page(pmd);
1861 return PageLocked(page);
1862 }
1863
1864 /*
1865 * Attempt to migrate a misplaced page to the specified destination
1866 * node. Caller is expected to have an elevated reference count on
1867 * the page that will be dropped by this function before returning.
1868 */
1869 int migrate_misplaced_page(struct page *page, struct vm_area_struct *vma,
1870 int node)
1871 {
1872 pg_data_t *pgdat = NODE_DATA(node);
1873 int isolated;
1874 int nr_remaining;
1875 LIST_HEAD(migratepages);
1876
1877 /*
1878 * Don't migrate file pages that are mapped in multiple processes
1879 * with execute permissions as they are probably shared libraries.
1880 */
1881 if (page_mapcount(page) != 1 && page_is_file_cache(page) &&
1882 (vma->vm_flags & VM_EXEC))
1883 goto out;
1884
1885 /*
1886 * Rate-limit the amount of data that is being migrated to a node.
1887 * Optimal placement is no good if the memory bus is saturated and
1888 * all the time is being spent migrating!
1889 */
1890 if (numamigrate_update_ratelimit(pgdat, 1))
1891 goto out;
1892
1893 isolated = numamigrate_isolate_page(pgdat, page);
1894 if (!isolated)
1895 goto out;
1896
1897 list_add(&page->lru, &migratepages);
1898 nr_remaining = migrate_pages(&migratepages, alloc_misplaced_dst_page,
1899 NULL, node, MIGRATE_ASYNC,
1900 MR_NUMA_MISPLACED);
1901 if (nr_remaining) {
1902 if (!list_empty(&migratepages)) {
1903 list_del(&page->lru);
1904 dec_zone_page_state(page, NR_ISOLATED_ANON +
1905 page_is_file_cache(page));
1906 putback_lru_page(page);
1907 }
1908 isolated = 0;
1909 } else
1910 count_vm_numa_event(NUMA_PAGE_MIGRATE);
1911 BUG_ON(!list_empty(&migratepages));
1912 return isolated;
1913
1914 out:
1915 put_page(page);
1916 return 0;
1917 }
1918 #endif /* CONFIG_NUMA_BALANCING */
1919
1920 #if defined(CONFIG_NUMA_BALANCING) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
1921 /*
1922 * Migrates a THP to a given target node. page must be locked and is unlocked
1923 * before returning.
1924 */
1925 int migrate_misplaced_transhuge_page(struct mm_struct *mm,
1926 struct vm_area_struct *vma,
1927 pmd_t *pmd, pmd_t entry,
1928 unsigned long address,
1929 struct page *page, int node)
1930 {
1931 spinlock_t *ptl;
1932 pg_data_t *pgdat = NODE_DATA(node);
1933 int isolated = 0;
1934 struct page *new_page = NULL;
1935 int page_lru = page_is_file_cache(page);
1936 unsigned long mmun_start = address & HPAGE_PMD_MASK;
1937 unsigned long mmun_end = mmun_start + HPAGE_PMD_SIZE;
1938 pmd_t orig_entry;
1939
1940 /*
1941 * Rate-limit the amount of data that is being migrated to a node.
1942 * Optimal placement is no good if the memory bus is saturated and
1943 * all the time is being spent migrating!
1944 */
1945 if (numamigrate_update_ratelimit(pgdat, HPAGE_PMD_NR))
1946 goto out_dropref;
1947
1948 new_page = alloc_pages_node(node,
1949 (GFP_TRANSHUGE | __GFP_THISNODE) & ~__GFP_RECLAIM,
1950 HPAGE_PMD_ORDER);
1951 if (!new_page)
1952 goto out_fail;
1953 prep_transhuge_page(new_page);
1954
1955 isolated = numamigrate_isolate_page(pgdat, page);
1956 if (!isolated) {
1957 put_page(new_page);
1958 goto out_fail;
1959 }
1960 /*
1961 * We are not sure a pending tlb flush here is for a huge page
1962 * mapping or not. Hence use the tlb range variant
1963 */
1964 if (mm_tlb_flush_pending(mm))
1965 flush_tlb_range(vma, mmun_start, mmun_end);
1966
1967 /* Prepare a page as a migration target */
1968 __SetPageLocked(new_page);
1969 __SetPageSwapBacked(new_page);
1970
1971 /* anon mapping, we can simply copy page->mapping to the new page: */
1972 new_page->mapping = page->mapping;
1973 new_page->index = page->index;
1974 migrate_page_copy(new_page, page);
1975 WARN_ON(PageLRU(new_page));
1976
1977 /* Recheck the target PMD */
1978 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1979 ptl = pmd_lock(mm, pmd);
1980 if (unlikely(!pmd_same(*pmd, entry) || page_count(page) != 2)) {
1981 fail_putback:
1982 spin_unlock(ptl);
1983 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1984
1985 /* Reverse changes made by migrate_page_copy() */
1986 if (TestClearPageActive(new_page))
1987 SetPageActive(page);
1988 if (TestClearPageUnevictable(new_page))
1989 SetPageUnevictable(page);
1990
1991 unlock_page(new_page);
1992 put_page(new_page); /* Free it */
1993
1994 /* Retake the callers reference and putback on LRU */
1995 get_page(page);
1996 putback_lru_page(page);
1997 mod_zone_page_state(page_zone(page),
1998 NR_ISOLATED_ANON + page_lru, -HPAGE_PMD_NR);
1999
2000 goto out_unlock;
2001 }
2002
2003 orig_entry = *pmd;
2004 entry = mk_pmd(new_page, vma->vm_page_prot);
2005 entry = pmd_mkhuge(entry);
2006 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
2007
2008 /*
2009 * Clear the old entry under pagetable lock and establish the new PTE.
2010 * Any parallel GUP will either observe the old page blocking on the
2011 * page lock, block on the page table lock or observe the new page.
2012 * The SetPageUptodate on the new page and page_add_new_anon_rmap
2013 * guarantee the copy is visible before the pagetable update.
2014 */
2015 flush_cache_range(vma, mmun_start, mmun_end);
2016 page_add_anon_rmap(new_page, vma, mmun_start, true);
2017 pmdp_huge_clear_flush_notify(vma, mmun_start, pmd);
2018 set_pmd_at(mm, mmun_start, pmd, entry);
2019 update_mmu_cache_pmd(vma, address, &entry);
2020
2021 if (page_count(page) != 2) {
2022 set_pmd_at(mm, mmun_start, pmd, orig_entry);
2023 flush_pmd_tlb_range(vma, mmun_start, mmun_end);
2024 mmu_notifier_invalidate_range(mm, mmun_start, mmun_end);
2025 update_mmu_cache_pmd(vma, address, &entry);
2026 page_remove_rmap(new_page, true);
2027 goto fail_putback;
2028 }
2029
2030 mlock_migrate_page(new_page, page);
2031 page_remove_rmap(page, true);
2032 set_page_owner_migrate_reason(new_page, MR_NUMA_MISPLACED);
2033
2034 spin_unlock(ptl);
2035 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2036
2037 /* Take an "isolate" reference and put new page on the LRU. */
2038 get_page(new_page);
2039 putback_lru_page(new_page);
2040
2041 unlock_page(new_page);
2042 unlock_page(page);
2043 put_page(page); /* Drop the rmap reference */
2044 put_page(page); /* Drop the LRU isolation reference */
2045
2046 count_vm_events(PGMIGRATE_SUCCESS, HPAGE_PMD_NR);
2047 count_vm_numa_events(NUMA_PAGE_MIGRATE, HPAGE_PMD_NR);
2048
2049 mod_zone_page_state(page_zone(page),
2050 NR_ISOLATED_ANON + page_lru,
2051 -HPAGE_PMD_NR);
2052 return isolated;
2053
2054 out_fail:
2055 count_vm_events(PGMIGRATE_FAIL, HPAGE_PMD_NR);
2056 out_dropref:
2057 ptl = pmd_lock(mm, pmd);
2058 if (pmd_same(*pmd, entry)) {
2059 entry = pmd_modify(entry, vma->vm_page_prot);
2060 set_pmd_at(mm, mmun_start, pmd, entry);
2061 update_mmu_cache_pmd(vma, address, &entry);
2062 }
2063 spin_unlock(ptl);
2064
2065 out_unlock:
2066 unlock_page(page);
2067 put_page(page);
2068 return 0;
2069 }
2070 #endif /* CONFIG_NUMA_BALANCING */
2071
2072 #endif /* CONFIG_NUMA */
This page took 0.106254 seconds and 5 git commands to generate.