1 /*
2 * Memory Migration functionality - linux/mm/migration.c
3 *
4 * Copyright (C) 2006 Silicon Graphics, Inc., Christoph Lameter
5 *
6 * Page migration was first developed in the context of the memory hotplug
7 * project. The main authors of the migration code are:
8 *
9 * IWAMOTO Toshihiro <iwamoto@valinux.co.jp>
10 * Hirokazu Takahashi <taka@valinux.co.jp>
11 * Dave Hansen <haveblue@us.ibm.com>
12 * Christoph Lameter <clameter@sgi.com>
13 */
14
15 #include <linux/migrate.h>
16 #include <linux/module.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/rmap.h>
25 #include <linux/topology.h>
26 #include <linux/cpu.h>
27 #include <linux/cpuset.h>
28 #include <linux/writeback.h>
29 #include <linux/mempolicy.h>
30 #include <linux/vmalloc.h>
31 #include <linux/security.h>
32 #include <linux/memcontrol.h>
33
34 #include "internal.h"
35
36 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
37
38 /*
39 * Isolate one page from the LRU lists. If successful put it onto
40 * the indicated list with elevated page count.
41 *
42 * Result:
43 * -EBUSY: page not on LRU list
44 * 0: page removed from LRU list and added to the specified list.
45 */
46 int isolate_lru_page(struct page *page, struct list_head *pagelist)
47 {
48 int ret = -EBUSY;
49
50 if (PageLRU(page)) {
51 struct zone *zone = page_zone(page);
52
53 spin_lock_irq(&zone->lru_lock);
54 if (PageLRU(page) && get_page_unless_zero(page)) {
55 ret = 0;
56 ClearPageLRU(page);
57 if (PageActive(page))
58 del_page_from_active_list(zone, page);
59 else
60 del_page_from_inactive_list(zone, page);
61 list_add_tail(&page->lru, pagelist);
62 }
63 spin_unlock_irq(&zone->lru_lock);
64 }
65 return ret;
66 }
67
68 /*
69 * migrate_prep() needs to be called before we start compiling a list of pages
70 * to be migrated using isolate_lru_page().
71 */
72 int migrate_prep(void)
73 {
74 /*
75 * Clear the LRU lists so pages can be isolated.
76 * Note that pages may be moved off the LRU after we have
77 * drained them. Those pages will fail to migrate like other
78 * pages that may be busy.
79 */
80 lru_add_drain_all();
81
82 return 0;
83 }
84
85 static inline void move_to_lru(struct page *page)
86 {
87 if (PageActive(page)) {
88 /*
89 * lru_cache_add_active checks that
90 * the PG_active bit is off.
91 */
92 ClearPageActive(page);
93 lru_cache_add_active(page);
94 } else {
95 lru_cache_add(page);
96 }
97 put_page(page);
98 }
99
100 /*
101 * Add isolated pages on the list back to the LRU.
102 *
103 * returns the number of pages put back.
104 */
105 int putback_lru_pages(struct list_head *l)
106 {
107 struct page *page;
108 struct page *page2;
109 int count = 0;
110
111 list_for_each_entry_safe(page, page2, l, lru) {
112 list_del(&page->lru);
113 move_to_lru(page);
114 count++;
115 }
116 return count;
117 }
118
119 /*
120 * Restore a potential migration pte to a working pte entry
121 */
122 static void remove_migration_pte(struct vm_area_struct *vma,
123 struct page *old, struct page *new)
124 {
125 struct mm_struct *mm = vma->vm_mm;
126 swp_entry_t entry;
127 pgd_t *pgd;
128 pud_t *pud;
129 pmd_t *pmd;
130 pte_t *ptep, pte;
131 spinlock_t *ptl;
132 unsigned long addr = page_address_in_vma(new, vma);
133
134 if (addr == -EFAULT)
135 return;
136
137 pgd = pgd_offset(mm, addr);
138 if (!pgd_present(*pgd))
139 return;
140
141 pud = pud_offset(pgd, addr);
142 if (!pud_present(*pud))
143 return;
144
145 pmd = pmd_offset(pud, addr);
146 if (!pmd_present(*pmd))
147 return;
148
149 ptep = pte_offset_map(pmd, addr);
150
151 if (!is_swap_pte(*ptep)) {
152 pte_unmap(ptep);
153 return;
154 }
155
156 ptl = pte_lockptr(mm, pmd);
157 spin_lock(ptl);
158 pte = *ptep;
159 if (!is_swap_pte(pte))
160 goto out;
161
162 entry = pte_to_swp_entry(pte);
163
164 if (!is_migration_entry(entry) || migration_entry_to_page(entry) != old)
165 goto out;
166
167 /*
168 * Yes, ignore the return value from a GFP_ATOMIC mem_cgroup_charge.
169 * Failure is not an option here: we're now expected to remove every
170 * migration pte, and will cause crashes otherwise. Normally this
171 * is not an issue: mem_cgroup_prepare_migration bumped up the old
172 * page_cgroup count for safety, that's now attached to the new page,
173 * so this charge should just be another incrementation of the count,
174 * to keep in balance with rmap.c's mem_cgroup_uncharging. But if
175 * there's been a force_empty, those reference counts may no longer
176 * be reliable, and this charge can actually fail: oh well, we don't
177 * make the situation any worse by proceeding as if it had succeeded.
178 */
179 mem_cgroup_charge(new, mm, GFP_ATOMIC);
180
181 get_page(new);
182 pte = pte_mkold(mk_pte(new, vma->vm_page_prot));
183 if (is_write_migration_entry(entry))
184 pte = pte_mkwrite(pte);
185 flush_cache_page(vma, addr, pte_pfn(pte));
186 set_pte_at(mm, addr, ptep, pte);
187
188 if (PageAnon(new))
189 page_add_anon_rmap(new, vma, addr);
190 else
191 page_add_file_rmap(new);
192
193 /* No need to invalidate - it was non-present before */
194 update_mmu_cache(vma, addr, pte);
195
196 out:
197 pte_unmap_unlock(ptep, ptl);
198 }
199
200 /*
201 * Note that remove_file_migration_ptes will only work on regular mappings,
202 * Nonlinear mappings do not use migration entries.
203 */
204 static void remove_file_migration_ptes(struct page *old, struct page *new)
205 {
206 struct vm_area_struct *vma;
207 struct address_space *mapping = page_mapping(new);
208 struct prio_tree_iter iter;
209 pgoff_t pgoff = new->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
210
211 if (!mapping)
212 return;
213
214 spin_lock(&mapping->i_mmap_lock);
215
216 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff)
217 remove_migration_pte(vma, old, new);
218
219 spin_unlock(&mapping->i_mmap_lock);
220 }
221
222 /*
223 * Must hold mmap_sem lock on at least one of the vmas containing
224 * the page so that the anon_vma cannot vanish.
225 */
226 static void remove_anon_migration_ptes(struct page *old, struct page *new)
227 {
228 struct anon_vma *anon_vma;
229 struct vm_area_struct *vma;
230 unsigned long mapping;
231
232 mapping = (unsigned long)new->mapping;
233
234 if (!mapping || (mapping & PAGE_MAPPING_ANON) == 0)
235 return;
236
237 /*
238 * We hold the mmap_sem lock. So no need to call page_lock_anon_vma.
239 */
240 anon_vma = (struct anon_vma *) (mapping - PAGE_MAPPING_ANON);
241 spin_lock(&anon_vma->lock);
242
243 list_for_each_entry(vma, &anon_vma->head, anon_vma_node)
244 remove_migration_pte(vma, old, new);
245
246 spin_unlock(&anon_vma->lock);
247 }
248
249 /*
250 * Get rid of all migration entries and replace them by
251 * references to the indicated page.
252 */
253 static void remove_migration_ptes(struct page *old, struct page *new)
254 {
255 if (PageAnon(new))
256 remove_anon_migration_ptes(old, new);
257 else
258 remove_file_migration_ptes(old, new);
259 }
260
261 /*
262 * Something used the pte of a page under migration. We need to
263 * get to the page and wait until migration is finished.
264 * When we return from this function the fault will be retried.
265 *
266 * This function is called from do_swap_page().
267 */
268 void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
269 unsigned long address)
270 {
271 pte_t *ptep, pte;
272 spinlock_t *ptl;
273 swp_entry_t entry;
274 struct page *page;
275
276 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
277 pte = *ptep;
278 if (!is_swap_pte(pte))
279 goto out;
280
281 entry = pte_to_swp_entry(pte);
282 if (!is_migration_entry(entry))
283 goto out;
284
285 page = migration_entry_to_page(entry);
286
287 get_page(page);
288 pte_unmap_unlock(ptep, ptl);
289 wait_on_page_locked(page);
290 put_page(page);
291 return;
292 out:
293 pte_unmap_unlock(ptep, ptl);
294 }
295
296 /*
297 * Replace the page in the mapping.
298 *
299 * The number of remaining references must be:
300 * 1 for anonymous pages without a mapping
301 * 2 for pages with a mapping
302 * 3 for pages with a mapping and PagePrivate set.
303 */
304 static int migrate_page_move_mapping(struct address_space *mapping,
305 struct page *newpage, struct page *page)
306 {
307 void **pslot;
308 struct radix_tree_context ctx;
309
310 if (!mapping) {
311 /* Anonymous page without mapping */
312 if (page_count(page) != 1)
313 return -EAGAIN;
314 return 0;
315 }
316
317 init_radix_tree_context(&ctx, &mapping->page_tree);
318 lock_page_ref_irq(page);
319 radix_tree_lock(&ctx);
320 pslot = radix_tree_lookup_slot(ctx.tree, page_index(page));
321
322 if (page_count(page) != 2 + !!PagePrivate(page) ||
323 (struct page *)radix_tree_deref_slot(pslot) != page) {
324 radix_tree_unlock(&ctx);
325 unlock_page_ref_irq(page);
326 return -EAGAIN;
327 }
328
329 /*
330 * Now we know that no one else is looking at the page.
331 */
332 get_page(newpage); /* add cache reference */
333 #ifdef CONFIG_SWAP
334 if (PageSwapCache(page)) {
335 SetPageSwapCache(newpage);
336 set_page_private(newpage, page_private(page));
337 }
338 #endif
339
340 radix_tree_replace_slot(pslot, newpage);
341 page->mapping = NULL;
342 radix_tree_unlock(&ctx);
343
344 /*
345 * If moved to a different zone then also account
346 * the page for that zone. Other VM counters will be
347 * taken care of when we establish references to the
348 * new page and drop references to the old page.
349 *
350 * Note that anonymous pages are accounted for
351 * via NR_FILE_PAGES and NR_ANON_PAGES if they
352 * are mapped to swap space.
353 */
354 __dec_zone_page_state(page, NR_FILE_PAGES);
355 __inc_zone_page_state(newpage, NR_FILE_PAGES);
356
357 unlock_page_ref_irq(page);
358
359 /*
360 * Drop cache reference from old page.
361 * We know this isn't the last reference.
362 */
363 __put_page(page);
364
365 return 0;
366 }
367
368 /*
369 * Copy the page to its new location
370 */
371 static void migrate_page_copy(struct page *newpage, struct page *page)
372 {
373 copy_highpage(newpage, page);
374
375 if (PageError(page))
376 SetPageError(newpage);
377 if (PageReferenced(page))
378 SetPageReferenced(newpage);
379 if (PageUptodate(page))
380 SetPageUptodate(newpage);
381 if (PageActive(page))
382 SetPageActive(newpage);
383 if (PageChecked(page))
384 SetPageChecked(newpage);
385 if (PageMappedToDisk(page))
386 SetPageMappedToDisk(newpage);
387
388 if (PageDirty(page)) {
389 clear_page_dirty_for_io(page);
390 set_page_dirty(newpage);
391 }
392
393 #ifdef CONFIG_SWAP
394 ClearPageSwapCache(page);
395 #endif
396 ClearPageActive(page);
397 ClearPagePrivate(page);
398 set_page_private(page, 0);
399 page->mapping = NULL;
400
401 /*
402 * If any waiters have accumulated on the new page then
403 * wake them up.
404 */
405 if (PageWriteback(newpage))
406 end_page_writeback(newpage);
407 }
408
409 /************************************************************
410 * Migration functions
411 ***********************************************************/
412
413 /* Always fail migration. Used for mappings that are not movable */
414 int fail_migrate_page(struct address_space *mapping,
415 struct page *newpage, struct page *page)
416 {
417 return -EIO;
418 }
419 EXPORT_SYMBOL(fail_migrate_page);
420
421 /*
422 * Common logic to directly migrate a single page suitable for
423 * pages that do not use PagePrivate.
424 *
425 * Pages are locked upon entry and exit.
426 */
427 int migrate_page(struct address_space *mapping,
428 struct page *newpage, struct page *page)
429 {
430 int rc;
431
432 BUG_ON(PageWriteback(page)); /* Writeback must be complete */
433
434 rc = migrate_page_move_mapping(mapping, newpage, page);
435
436 if (rc)
437 return rc;
438
439 migrate_page_copy(newpage, page);
440 return 0;
441 }
442 EXPORT_SYMBOL(migrate_page);
443
444 #ifdef CONFIG_BLOCK
445 /*
446 * Migration function for pages with buffers. This function can only be used
447 * if the underlying filesystem guarantees that no other references to "page"
448 * exist.
449 */
450 int buffer_migrate_page(struct address_space *mapping,
451 struct page *newpage, struct page *page)
452 {
453 struct buffer_head *bh, *head;
454 int rc;
455
456 if (!page_has_buffers(page))
457 return migrate_page(mapping, newpage, page);
458
459 head = page_buffers(page);
460
461 rc = migrate_page_move_mapping(mapping, newpage, page);
462
463 if (rc)
464 return rc;
465
466 bh = head;
467 do {
468 get_bh(bh);
469 lock_buffer(bh);
470 bh = bh->b_this_page;
471
472 } while (bh != head);
473
474 ClearPagePrivate(page);
475 set_page_private(newpage, page_private(page));
476 set_page_private(page, 0);
477 put_page(page);
478 get_page(newpage);
479
480 bh = head;
481 do {
482 set_bh_page(bh, newpage, bh_offset(bh));
483 bh = bh->b_this_page;
484
485 } while (bh != head);
486
487 SetPagePrivate(newpage);
488
489 migrate_page_copy(newpage, page);
490
491 bh = head;
492 do {
493 unlock_buffer(bh);
494 put_bh(bh);
495 bh = bh->b_this_page;
496
497 } while (bh != head);
498
499 return 0;
500 }
501 EXPORT_SYMBOL(buffer_migrate_page);
502 #endif
503
504 /*
505 * Writeback a page to clean the dirty state
506 */
507 static int writeout(struct address_space *mapping, struct page *page)
508 {
509 struct writeback_control wbc = {
510 .sync_mode = WB_SYNC_NONE,
511 .nr_to_write = 1,
512 .range_start = 0,
513 .range_end = LLONG_MAX,
514 .nonblocking = 1,
515 .for_reclaim = 1
516 };
517 int rc;
518
519 if (!mapping->a_ops->writepage)
520 /* No write method for the address space */
521 return -EINVAL;
522
523 if (!clear_page_dirty_for_io(page))
524 /* Someone else already triggered a write */
525 return -EAGAIN;
526
527 /*
528 * A dirty page may imply that the underlying filesystem has
529 * the page on some queue. So the page must be clean for
530 * migration. Writeout may mean we loose the lock and the
531 * page state is no longer what we checked for earlier.
532 * At this point we know that the migration attempt cannot
533 * be successful.
534 */
535 remove_migration_ptes(page, page);
536
537 rc = mapping->a_ops->writepage(page, &wbc);
538 if (rc < 0)
539 /* I/O Error writing */
540 return -EIO;
541
542 if (rc != AOP_WRITEPAGE_ACTIVATE)
543 /* unlocked. Relock */
544 lock_page(page);
545
546 return -EAGAIN;
547 }
548
549 /*
550 * Default handling if a filesystem does not provide a migration function.
551 */
552 static int fallback_migrate_page(struct address_space *mapping,
553 struct page *newpage, struct page *page)
554 {
555 if (PageDirty(page))
556 return writeout(mapping, page);
557
558 /*
559 * Buffers may be managed in a filesystem specific way.
560 * We must have no buffers or drop them.
561 */
562 if (PagePrivate(page) &&
563 !try_to_release_page(page, GFP_KERNEL))
564 return -EAGAIN;
565
566 return migrate_page(mapping, newpage, page);
567 }
568
569 /*
570 * Move a page to a newly allocated page
571 * The page is locked and all ptes have been successfully removed.
572 *
573 * The new page will have replaced the old page if this function
574 * is successful.
575 */
576 static int move_to_new_page(struct page *newpage, struct page *page)
577 {
578 struct address_space *mapping;
579 int rc;
580
581 /*
582 * Block others from accessing the page when we get around to
583 * establishing additional references. We are the only one
584 * holding a reference to the new page at this point.
585 */
586 if (TestSetPageLocked(newpage))
587 BUG();
588
589 /* Prepare mapping for the new page.*/
590 newpage->index = page->index;
591 newpage->mapping = page->mapping;
592
593 mapping = page_mapping(page);
594 if (!mapping)
595 rc = migrate_page(mapping, newpage, page);
596 else if (mapping->a_ops->migratepage)
597 /*
598 * Most pages have a mapping and most filesystems
599 * should provide a migration function. Anonymous
600 * pages are part of swap space which also has its
601 * own migration function. This is the most common
602 * path for page migration.
603 */
604 rc = mapping->a_ops->migratepage(mapping,
605 newpage, page);
606 else
607 rc = fallback_migrate_page(mapping, newpage, page);
608
609 if (!rc) {
610 mem_cgroup_page_migration(page, newpage);
611 remove_migration_ptes(page, newpage);
612 } else
613 newpage->mapping = NULL;
614
615 unlock_page(newpage);
616
617 return rc;
618 }
619
620 /*
621 * Obtain the lock on page, remove all ptes and migrate the page
622 * to the newly allocated page in newpage.
623 */
624 static int unmap_and_move(new_page_t get_new_page, unsigned long private,
625 struct page *page, int force)
626 {
627 int rc = 0;
628 int *result = NULL;
629 struct page *newpage = get_new_page(page, private, &result);
630 int rcu_locked = 0;
631 int charge = 0;
632
633 if (!newpage)
634 return -ENOMEM;
635
636 if (page_count(page) == 1)
637 /* page was freed from under us. So we are done. */
638 goto move_newpage;
639
640 rc = -EAGAIN;
641 if (TestSetPageLocked(page)) {
642 if (!force)
643 goto move_newpage;
644 lock_page(page);
645 }
646
647 if (PageWriteback(page)) {
648 if (!force)
649 goto unlock;
650 wait_on_page_writeback(page);
651 }
652 /*
653 * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
654 * we cannot notice that anon_vma is freed while we migrates a page.
655 * This rcu_read_lock() delays freeing anon_vma pointer until the end
656 * of migration. File cache pages are no problem because of page_lock()
657 * File Caches may use write_page() or lock_page() in migration, then,
658 * just care Anon page here.
659 */
660 if (PageAnon(page)) {
661 rcu_read_lock();
662 rcu_locked = 1;
663 }
664
665 /*
666 * Corner case handling:
667 * 1. When a new swap-cache page is read into, it is added to the LRU
668 * and treated as swapcache but it has no rmap yet.
669 * Calling try_to_unmap() against a page->mapping==NULL page will
670 * trigger a BUG. So handle it here.
671 * 2. An orphaned page (see truncate_complete_page) might have
672 * fs-private metadata. The page can be picked up due to memory
673 * offlining. Everywhere else except page reclaim, the page is
674 * invisible to the vm, so the page can not be migrated. So try to
675 * free the metadata, so the page can be freed.
676 */
677 if (!page->mapping) {
678 if (!PageAnon(page) && PagePrivate(page)) {
679 /*
680 * Go direct to try_to_free_buffers() here because
681 * a) that's what try_to_release_page() would do anyway
682 * b) we may be under rcu_read_lock() here, so we can't
683 * use GFP_KERNEL which is what try_to_release_page()
684 * needs to be effective.
685 */
686 try_to_free_buffers(page);
687 }
688 goto rcu_unlock;
689 }
690
691 charge = mem_cgroup_prepare_migration(page);
692 /* Establish migration ptes or remove ptes */
693 try_to_unmap(page, 1);
694
695 if (!page_mapped(page))
696 rc = move_to_new_page(newpage, page);
697
698 if (rc) {
699 remove_migration_ptes(page, page);
700 if (charge)
701 mem_cgroup_end_migration(page);
702 } else if (charge)
703 mem_cgroup_end_migration(newpage);
704 rcu_unlock:
705 if (rcu_locked)
706 rcu_read_unlock();
707
708 unlock:
709
710 unlock_page(page);
711
712 if (rc != -EAGAIN) {
713 /*
714 * A page that has been migrated has all references
715 * removed and will be freed. A page that has not been
716 * migrated will have kepts its references and be
717 * restored.
718 */
719 list_del(&page->lru);
720 move_to_lru(page);
721 }
722
723 move_newpage:
724 /*
725 * Move the new page to the LRU. If migration was not successful
726 * then this will free the page.
727 */
728 move_to_lru(newpage);
729 if (result) {
730 if (rc)
731 *result = rc;
732 else
733 *result = page_to_nid(newpage);
734 }
735 return rc;
736 }
737
738 /*
739 * migrate_pages
740 *
741 * The function takes one list of pages to migrate and a function
742 * that determines from the page to be migrated and the private data
743 * the target of the move and allocates the page.
744 *
745 * The function returns after 10 attempts or if no pages
746 * are movable anymore because to has become empty
747 * or no retryable pages exist anymore. All pages will be
748 * returned to the LRU or freed.
749 *
750 * Return: Number of pages not migrated or error code.
751 */
752 int migrate_pages(struct list_head *from,
753 new_page_t get_new_page, unsigned long private)
754 {
755 int retry = 1;
756 int nr_failed = 0;
757 int pass = 0;
758 struct page *page;
759 struct page *page2;
760 int swapwrite = current->flags & PF_SWAPWRITE;
761 int rc;
762
763 if (!swapwrite)
764 current->flags |= PF_SWAPWRITE;
765
766 for(pass = 0; pass < 10 && retry; pass++) {
767 retry = 0;
768
769 list_for_each_entry_safe(page, page2, from, lru) {
770 cond_resched();
771
772 rc = unmap_and_move(get_new_page, private,
773 page, pass > 2);
774
775 switch(rc) {
776 case -ENOMEM:
777 goto out;
778 case -EAGAIN:
779 retry++;
780 break;
781 case 0:
782 break;
783 default:
784 /* Permanent failure */
785 nr_failed++;
786 break;
787 }
788 }
789 }
790 rc = 0;
791 out:
792 if (!swapwrite)
793 current->flags &= ~PF_SWAPWRITE;
794
795 putback_lru_pages(from);
796
797 if (rc)
798 return rc;
799
800 return nr_failed + retry;
801 }
802
803 #ifdef CONFIG_NUMA
804 /*
805 * Move a list of individual pages
806 */
807 struct page_to_node {
808 unsigned long addr;
809 struct page *page;
810 int node;
811 int status;
812 };
813
814 static struct page *new_page_node(struct page *p, unsigned long private,
815 int **result)
816 {
817 struct page_to_node *pm = (struct page_to_node *)private;
818
819 while (pm->node != MAX_NUMNODES && pm->page != p)
820 pm++;
821
822 if (pm->node == MAX_NUMNODES)
823 return NULL;
824
825 *result = &pm->status;
826
827 return alloc_pages_node(pm->node,
828 GFP_HIGHUSER_MOVABLE | GFP_THISNODE, 0);
829 }
830
831 /*
832 * Move a set of pages as indicated in the pm array. The addr
833 * field must be set to the virtual address of the page to be moved
834 * and the node number must contain a valid target node.
835 */
836 static int do_move_pages(struct mm_struct *mm, struct page_to_node *pm,
837 int migrate_all)
838 {
839 int err;
840 struct page_to_node *pp;
841 LIST_HEAD(pagelist);
842
843 down_read(&mm->mmap_sem);
844
845 /*
846 * Build a list of pages to migrate
847 */
848 migrate_prep();
849 for (pp = pm; pp->node != MAX_NUMNODES; pp++) {
850 struct vm_area_struct *vma;
851 struct page *page;
852
853 /*
854 * A valid page pointer that will not match any of the
855 * pages that will be moved.
856 */
857 pp->page = ZERO_PAGE(0);
858
859 err = -EFAULT;
860 vma = find_vma(mm, pp->addr);
861 if (!vma || !vma_migratable(vma))
862 goto set_status;
863
864 page = follow_page(vma, pp->addr, FOLL_GET);
865 err = -ENOENT;
866 if (!page)
867 goto set_status;
868
869 if (PageReserved(page)) /* Check for zero page */
870 goto put_and_set;
871
872 pp->page = page;
873 err = page_to_nid(page);
874
875 if (err == pp->node)
876 /*
877 * Node already in the right place
878 */
879 goto put_and_set;
880
881 err = -EACCES;
882 if (page_mapcount(page) > 1 &&
883 !migrate_all)
884 goto put_and_set;
885
886 err = isolate_lru_page(page, &pagelist);
887 put_and_set:
888 /*
889 * Either remove the duplicate refcount from
890 * isolate_lru_page() or drop the page ref if it was
891 * not isolated.
892 */
893 put_page(page);
894 set_status:
895 pp->status = err;
896 }
897
898 if (!list_empty(&pagelist))
899 err = migrate_pages(&pagelist, new_page_node,
900 (unsigned long)pm);
901 else
902 err = -ENOENT;
903
904 up_read(&mm->mmap_sem);
905 return err;
906 }
907
908 /*
909 * Determine the nodes of a list of pages. The addr in the pm array
910 * must have been set to the virtual address of which we want to determine
911 * the node number.
912 */
913 static int do_pages_stat(struct mm_struct *mm, struct page_to_node *pm)
914 {
915 down_read(&mm->mmap_sem);
916
917 for ( ; pm->node != MAX_NUMNODES; pm++) {
918 struct vm_area_struct *vma;
919 struct page *page;
920 int err;
921
922 err = -EFAULT;
923 vma = find_vma(mm, pm->addr);
924 if (!vma)
925 goto set_status;
926
927 page = follow_page(vma, pm->addr, 0);
928 err = -ENOENT;
929 /* Use PageReserved to check for zero page */
930 if (!page || PageReserved(page))
931 goto set_status;
932
933 err = page_to_nid(page);
934 set_status:
935 pm->status = err;
936 }
937
938 up_read(&mm->mmap_sem);
939 return 0;
940 }
941
942 /*
943 * Move a list of pages in the address space of the currently executing
944 * process.
945 */
946 asmlinkage long sys_move_pages(pid_t pid, unsigned long nr_pages,
947 const void __user * __user *pages,
948 const int __user *nodes,
949 int __user *status, int flags)
950 {
951 int err = 0;
952 int i;
953 struct task_struct *task;
954 nodemask_t task_nodes;
955 struct mm_struct *mm;
956 struct page_to_node *pm = NULL;
957
958 /* Check flags */
959 if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
960 return -EINVAL;
961
962 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
963 return -EPERM;
964
965 /* Find the mm_struct */
966 read_lock(&tasklist_lock);
967 task = pid ? find_task_by_vpid(pid) : current;
968 if (!task) {
969 read_unlock(&tasklist_lock);
970 return -ESRCH;
971 }
972 mm = get_task_mm(task);
973 read_unlock(&tasklist_lock);
974
975 if (!mm)
976 return -EINVAL;
977
978 /*
979 * Check if this process has the right to modify the specified
980 * process. The right exists if the process has administrative
981 * capabilities, superuser privileges or the same
982 * userid as the target process.
983 */
984 if ((current->euid != task->suid) && (current->euid != task->uid) &&
985 (current->uid != task->suid) && (current->uid != task->uid) &&
986 !capable(CAP_SYS_NICE)) {
987 err = -EPERM;
988 goto out2;
989 }
990
991 err = security_task_movememory(task);
992 if (err)
993 goto out2;
994
995
996 task_nodes = cpuset_mems_allowed(task);
997
998 /* Limit nr_pages so that the multiplication may not overflow */
999 if (nr_pages >= ULONG_MAX / sizeof(struct page_to_node) - 1) {
1000 err = -E2BIG;
1001 goto out2;
1002 }
1003
1004 pm = vmalloc((nr_pages + 1) * sizeof(struct page_to_node));
1005 if (!pm) {
1006 err = -ENOMEM;
1007 goto out2;
1008 }
1009
1010 /*
1011 * Get parameters from user space and initialize the pm
1012 * array. Return various errors if the user did something wrong.
1013 */
1014 for (i = 0; i < nr_pages; i++) {
1015 const void __user *p;
1016
1017 err = -EFAULT;
1018 if (get_user(p, pages + i))
1019 goto out;
1020
1021 pm[i].addr = (unsigned long)p;
1022 if (nodes) {
1023 int node;
1024
1025 if (get_user(node, nodes + i))
1026 goto out;
1027
1028 err = -ENODEV;
1029 if (!node_state(node, N_HIGH_MEMORY))
1030 goto out;
1031
1032 err = -EACCES;
1033 if (!node_isset(node, task_nodes))
1034 goto out;
1035
1036 pm[i].node = node;
1037 } else
1038 pm[i].node = 0; /* anything to not match MAX_NUMNODES */
1039 }
1040 /* End marker */
1041 pm[nr_pages].node = MAX_NUMNODES;
1042
1043 if (nodes)
1044 err = do_move_pages(mm, pm, flags & MPOL_MF_MOVE_ALL);
1045 else
1046 err = do_pages_stat(mm, pm);
1047
1048 if (err >= 0)
1049 /* Return status information */
1050 for (i = 0; i < nr_pages; i++)
1051 if (put_user(pm[i].status, status + i))
1052 err = -EFAULT;
1053
1054 out:
1055 vfree(pm);
1056 out2:
1057 mmput(mm);
1058 return err;
1059 }
1060 #endif
1061
1062 /*
1063 * Call migration functions in the vma_ops that may prepare
1064 * memory in a vm for migration. migration functions may perform
1065 * the migration for vmas that do not have an underlying page struct.
1066 */
1067 int migrate_vmas(struct mm_struct *mm, const nodemask_t *to,
1068 const nodemask_t *from, unsigned long flags)
1069 {
1070 struct vm_area_struct *vma;
1071 int err = 0;
1072
1073 for(vma = mm->mmap; vma->vm_next && !err; vma = vma->vm_next) {
1074 if (vma->vm_ops && vma->vm_ops->migrate) {
1075 err = vma->vm_ops->migrate(vma, to, from, flags);
1076 if (err)
1077 break;
1078 }
1079 }
1080 return err;
1081 }
1082
|
This page was automatically generated by the
LXR engine.
|