xref: /openbmc/linux/mm/migrate.c (revision ba6e8564)
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/pagevec.h>
23 #include <linux/rmap.h>
24 #include <linux/topology.h>
25 #include <linux/cpu.h>
26 #include <linux/cpuset.h>
27 #include <linux/writeback.h>
28 #include <linux/mempolicy.h>
29 #include <linux/vmalloc.h>
30 #include <linux/security.h>
31 
32 #include "internal.h"
33 
34 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
35 
36 /*
37  * Isolate one page from the LRU lists. If successful put it onto
38  * the indicated list with elevated page count.
39  *
40  * Result:
41  *  -EBUSY: page not on LRU list
42  *  0: page removed from LRU list and added to the specified list.
43  */
44 int isolate_lru_page(struct page *page, struct list_head *pagelist)
45 {
46 	int ret = -EBUSY;
47 
48 	if (PageLRU(page)) {
49 		struct zone *zone = page_zone(page);
50 
51 		spin_lock_irq(&zone->lru_lock);
52 		if (PageLRU(page)) {
53 			ret = 0;
54 			get_page(page);
55 			ClearPageLRU(page);
56 			if (PageActive(page))
57 				del_page_from_active_list(zone, page);
58 			else
59 				del_page_from_inactive_list(zone, page);
60 			list_add_tail(&page->lru, pagelist);
61 		}
62 		spin_unlock_irq(&zone->lru_lock);
63 	}
64 	return ret;
65 }
66 
67 /*
68  * migrate_prep() needs to be called before we start compiling a list of pages
69  * to be migrated using isolate_lru_page().
70  */
71 int migrate_prep(void)
72 {
73 	/*
74 	 * Clear the LRU lists so pages can be isolated.
75 	 * Note that pages may be moved off the LRU after we have
76 	 * drained them. Those pages will fail to migrate like other
77 	 * pages that may be busy.
78 	 */
79 	lru_add_drain_all();
80 
81 	return 0;
82 }
83 
84 static inline void move_to_lru(struct page *page)
85 {
86 	if (PageActive(page)) {
87 		/*
88 		 * lru_cache_add_active checks that
89 		 * the PG_active bit is off.
90 		 */
91 		ClearPageActive(page);
92 		lru_cache_add_active(page);
93 	} else {
94 		lru_cache_add(page);
95 	}
96 	put_page(page);
97 }
98 
99 /*
100  * Add isolated pages on the list back to the LRU.
101  *
102  * returns the number of pages put back.
103  */
104 int putback_lru_pages(struct list_head *l)
105 {
106 	struct page *page;
107 	struct page *page2;
108 	int count = 0;
109 
110 	list_for_each_entry_safe(page, page2, l, lru) {
111 		list_del(&page->lru);
112 		move_to_lru(page);
113 		count++;
114 	}
115 	return count;
116 }
117 
118 static inline int is_swap_pte(pte_t pte)
119 {
120 	return !pte_none(pte) && !pte_present(pte) && !pte_file(pte);
121 }
122 
123 /*
124  * Restore a potential migration pte to a working pte entry
125  */
126 static void remove_migration_pte(struct vm_area_struct *vma,
127 		struct page *old, struct page *new)
128 {
129 	struct mm_struct *mm = vma->vm_mm;
130 	swp_entry_t entry;
131  	pgd_t *pgd;
132  	pud_t *pud;
133  	pmd_t *pmd;
134 	pte_t *ptep, pte;
135  	spinlock_t *ptl;
136 	unsigned long addr = page_address_in_vma(new, vma);
137 
138 	if (addr == -EFAULT)
139 		return;
140 
141  	pgd = pgd_offset(mm, addr);
142 	if (!pgd_present(*pgd))
143                 return;
144 
145 	pud = pud_offset(pgd, addr);
146 	if (!pud_present(*pud))
147                 return;
148 
149 	pmd = pmd_offset(pud, addr);
150 	if (!pmd_present(*pmd))
151 		return;
152 
153 	ptep = pte_offset_map(pmd, addr);
154 
155 	if (!is_swap_pte(*ptep)) {
156 		pte_unmap(ptep);
157  		return;
158  	}
159 
160  	ptl = pte_lockptr(mm, pmd);
161  	spin_lock(ptl);
162 	pte = *ptep;
163 	if (!is_swap_pte(pte))
164 		goto out;
165 
166 	entry = pte_to_swp_entry(pte);
167 
168 	if (!is_migration_entry(entry) || migration_entry_to_page(entry) != old)
169 		goto out;
170 
171 	get_page(new);
172 	pte = pte_mkold(mk_pte(new, vma->vm_page_prot));
173 	if (is_write_migration_entry(entry))
174 		pte = pte_mkwrite(pte);
175 	set_pte_at(mm, addr, ptep, pte);
176 
177 	if (PageAnon(new))
178 		page_add_anon_rmap(new, vma, addr);
179 	else
180 		page_add_file_rmap(new);
181 
182 	/* No need to invalidate - it was non-present before */
183 	update_mmu_cache(vma, addr, pte);
184 	lazy_mmu_prot_update(pte);
185 
186 out:
187 	pte_unmap_unlock(ptep, ptl);
188 }
189 
190 /*
191  * Note that remove_file_migration_ptes will only work on regular mappings,
192  * Nonlinear mappings do not use migration entries.
193  */
194 static void remove_file_migration_ptes(struct page *old, struct page *new)
195 {
196 	struct vm_area_struct *vma;
197 	struct address_space *mapping = page_mapping(new);
198 	struct prio_tree_iter iter;
199 	pgoff_t pgoff = new->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
200 
201 	if (!mapping)
202 		return;
203 
204 	spin_lock(&mapping->i_mmap_lock);
205 
206 	vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff)
207 		remove_migration_pte(vma, old, new);
208 
209 	spin_unlock(&mapping->i_mmap_lock);
210 }
211 
212 /*
213  * Must hold mmap_sem lock on at least one of the vmas containing
214  * the page so that the anon_vma cannot vanish.
215  */
216 static void remove_anon_migration_ptes(struct page *old, struct page *new)
217 {
218 	struct anon_vma *anon_vma;
219 	struct vm_area_struct *vma;
220 	unsigned long mapping;
221 
222 	mapping = (unsigned long)new->mapping;
223 
224 	if (!mapping || (mapping & PAGE_MAPPING_ANON) == 0)
225 		return;
226 
227 	/*
228 	 * We hold the mmap_sem lock. So no need to call page_lock_anon_vma.
229 	 */
230 	anon_vma = (struct anon_vma *) (mapping - PAGE_MAPPING_ANON);
231 	spin_lock(&anon_vma->lock);
232 
233 	list_for_each_entry(vma, &anon_vma->head, anon_vma_node)
234 		remove_migration_pte(vma, old, new);
235 
236 	spin_unlock(&anon_vma->lock);
237 }
238 
239 /*
240  * Get rid of all migration entries and replace them by
241  * references to the indicated page.
242  */
243 static void remove_migration_ptes(struct page *old, struct page *new)
244 {
245 	if (PageAnon(new))
246 		remove_anon_migration_ptes(old, new);
247 	else
248 		remove_file_migration_ptes(old, new);
249 }
250 
251 /*
252  * Something used the pte of a page under migration. We need to
253  * get to the page and wait until migration is finished.
254  * When we return from this function the fault will be retried.
255  *
256  * This function is called from do_swap_page().
257  */
258 void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
259 				unsigned long address)
260 {
261 	pte_t *ptep, pte;
262 	spinlock_t *ptl;
263 	swp_entry_t entry;
264 	struct page *page;
265 
266 	ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
267 	pte = *ptep;
268 	if (!is_swap_pte(pte))
269 		goto out;
270 
271 	entry = pte_to_swp_entry(pte);
272 	if (!is_migration_entry(entry))
273 		goto out;
274 
275 	page = migration_entry_to_page(entry);
276 
277 	get_page(page);
278 	pte_unmap_unlock(ptep, ptl);
279 	wait_on_page_locked(page);
280 	put_page(page);
281 	return;
282 out:
283 	pte_unmap_unlock(ptep, ptl);
284 }
285 
286 /*
287  * Replace the page in the mapping.
288  *
289  * The number of remaining references must be:
290  * 1 for anonymous pages without a mapping
291  * 2 for pages with a mapping
292  * 3 for pages with a mapping and PagePrivate set.
293  */
294 static int migrate_page_move_mapping(struct address_space *mapping,
295 		struct page *newpage, struct page *page)
296 {
297 	void **pslot;
298 
299 	if (!mapping) {
300 		/* Anonymous page */
301 		if (page_count(page) != 1)
302 			return -EAGAIN;
303 		return 0;
304 	}
305 
306 	write_lock_irq(&mapping->tree_lock);
307 
308 	pslot = radix_tree_lookup_slot(&mapping->page_tree,
309  					page_index(page));
310 
311 	if (page_count(page) != 2 + !!PagePrivate(page) ||
312 			(struct page *)radix_tree_deref_slot(pslot) != page) {
313 		write_unlock_irq(&mapping->tree_lock);
314 		return -EAGAIN;
315 	}
316 
317 	/*
318 	 * Now we know that no one else is looking at the page.
319 	 */
320 	get_page(newpage);	/* add cache reference */
321 #ifdef CONFIG_SWAP
322 	if (PageSwapCache(page)) {
323 		SetPageSwapCache(newpage);
324 		set_page_private(newpage, page_private(page));
325 	}
326 #endif
327 
328 	radix_tree_replace_slot(pslot, newpage);
329 
330 	/*
331 	 * Drop cache reference from old page.
332 	 * We know this isn't the last reference.
333 	 */
334 	__put_page(page);
335 
336 	write_unlock_irq(&mapping->tree_lock);
337 
338 	return 0;
339 }
340 
341 /*
342  * Copy the page to its new location
343  */
344 static void migrate_page_copy(struct page *newpage, struct page *page)
345 {
346 	copy_highpage(newpage, page);
347 
348 	if (PageError(page))
349 		SetPageError(newpage);
350 	if (PageReferenced(page))
351 		SetPageReferenced(newpage);
352 	if (PageUptodate(page))
353 		SetPageUptodate(newpage);
354 	if (PageActive(page))
355 		SetPageActive(newpage);
356 	if (PageChecked(page))
357 		SetPageChecked(newpage);
358 	if (PageMappedToDisk(page))
359 		SetPageMappedToDisk(newpage);
360 
361 	if (PageDirty(page)) {
362 		clear_page_dirty_for_io(page);
363 		set_page_dirty(newpage);
364  	}
365 
366 #ifdef CONFIG_SWAP
367 	ClearPageSwapCache(page);
368 #endif
369 	ClearPageActive(page);
370 	ClearPagePrivate(page);
371 	set_page_private(page, 0);
372 	page->mapping = NULL;
373 
374 	/*
375 	 * If any waiters have accumulated on the new page then
376 	 * wake them up.
377 	 */
378 	if (PageWriteback(newpage))
379 		end_page_writeback(newpage);
380 }
381 
382 /************************************************************
383  *                    Migration functions
384  ***********************************************************/
385 
386 /* Always fail migration. Used for mappings that are not movable */
387 int fail_migrate_page(struct address_space *mapping,
388 			struct page *newpage, struct page *page)
389 {
390 	return -EIO;
391 }
392 EXPORT_SYMBOL(fail_migrate_page);
393 
394 /*
395  * Common logic to directly migrate a single page suitable for
396  * pages that do not use PagePrivate.
397  *
398  * Pages are locked upon entry and exit.
399  */
400 int migrate_page(struct address_space *mapping,
401 		struct page *newpage, struct page *page)
402 {
403 	int rc;
404 
405 	BUG_ON(PageWriteback(page));	/* Writeback must be complete */
406 
407 	rc = migrate_page_move_mapping(mapping, newpage, page);
408 
409 	if (rc)
410 		return rc;
411 
412 	migrate_page_copy(newpage, page);
413 	return 0;
414 }
415 EXPORT_SYMBOL(migrate_page);
416 
417 #ifdef CONFIG_BLOCK
418 /*
419  * Migration function for pages with buffers. This function can only be used
420  * if the underlying filesystem guarantees that no other references to "page"
421  * exist.
422  */
423 int buffer_migrate_page(struct address_space *mapping,
424 		struct page *newpage, struct page *page)
425 {
426 	struct buffer_head *bh, *head;
427 	int rc;
428 
429 	if (!page_has_buffers(page))
430 		return migrate_page(mapping, newpage, page);
431 
432 	head = page_buffers(page);
433 
434 	rc = migrate_page_move_mapping(mapping, newpage, page);
435 
436 	if (rc)
437 		return rc;
438 
439 	bh = head;
440 	do {
441 		get_bh(bh);
442 		lock_buffer(bh);
443 		bh = bh->b_this_page;
444 
445 	} while (bh != head);
446 
447 	ClearPagePrivate(page);
448 	set_page_private(newpage, page_private(page));
449 	set_page_private(page, 0);
450 	put_page(page);
451 	get_page(newpage);
452 
453 	bh = head;
454 	do {
455 		set_bh_page(bh, newpage, bh_offset(bh));
456 		bh = bh->b_this_page;
457 
458 	} while (bh != head);
459 
460 	SetPagePrivate(newpage);
461 
462 	migrate_page_copy(newpage, page);
463 
464 	bh = head;
465 	do {
466 		unlock_buffer(bh);
467  		put_bh(bh);
468 		bh = bh->b_this_page;
469 
470 	} while (bh != head);
471 
472 	return 0;
473 }
474 EXPORT_SYMBOL(buffer_migrate_page);
475 #endif
476 
477 /*
478  * Writeback a page to clean the dirty state
479  */
480 static int writeout(struct address_space *mapping, struct page *page)
481 {
482 	struct writeback_control wbc = {
483 		.sync_mode = WB_SYNC_NONE,
484 		.nr_to_write = 1,
485 		.range_start = 0,
486 		.range_end = LLONG_MAX,
487 		.nonblocking = 1,
488 		.for_reclaim = 1
489 	};
490 	int rc;
491 
492 	if (!mapping->a_ops->writepage)
493 		/* No write method for the address space */
494 		return -EINVAL;
495 
496 	if (!clear_page_dirty_for_io(page))
497 		/* Someone else already triggered a write */
498 		return -EAGAIN;
499 
500 	/*
501 	 * A dirty page may imply that the underlying filesystem has
502 	 * the page on some queue. So the page must be clean for
503 	 * migration. Writeout may mean we loose the lock and the
504 	 * page state is no longer what we checked for earlier.
505 	 * At this point we know that the migration attempt cannot
506 	 * be successful.
507 	 */
508 	remove_migration_ptes(page, page);
509 
510 	rc = mapping->a_ops->writepage(page, &wbc);
511 	if (rc < 0)
512 		/* I/O Error writing */
513 		return -EIO;
514 
515 	if (rc != AOP_WRITEPAGE_ACTIVATE)
516 		/* unlocked. Relock */
517 		lock_page(page);
518 
519 	return -EAGAIN;
520 }
521 
522 /*
523  * Default handling if a filesystem does not provide a migration function.
524  */
525 static int fallback_migrate_page(struct address_space *mapping,
526 	struct page *newpage, struct page *page)
527 {
528 	if (PageDirty(page))
529 		return writeout(mapping, page);
530 
531 	/*
532 	 * Buffers may be managed in a filesystem specific way.
533 	 * We must have no buffers or drop them.
534 	 */
535 	if (PagePrivate(page) &&
536 	    !try_to_release_page(page, GFP_KERNEL))
537 		return -EAGAIN;
538 
539 	return migrate_page(mapping, newpage, page);
540 }
541 
542 /*
543  * Move a page to a newly allocated page
544  * The page is locked and all ptes have been successfully removed.
545  *
546  * The new page will have replaced the old page if this function
547  * is successful.
548  */
549 static int move_to_new_page(struct page *newpage, struct page *page)
550 {
551 	struct address_space *mapping;
552 	int rc;
553 
554 	/*
555 	 * Block others from accessing the page when we get around to
556 	 * establishing additional references. We are the only one
557 	 * holding a reference to the new page at this point.
558 	 */
559 	if (TestSetPageLocked(newpage))
560 		BUG();
561 
562 	/* Prepare mapping for the new page.*/
563 	newpage->index = page->index;
564 	newpage->mapping = page->mapping;
565 
566 	mapping = page_mapping(page);
567 	if (!mapping)
568 		rc = migrate_page(mapping, newpage, page);
569 	else if (mapping->a_ops->migratepage)
570 		/*
571 		 * Most pages have a mapping and most filesystems
572 		 * should provide a migration function. Anonymous
573 		 * pages are part of swap space which also has its
574 		 * own migration function. This is the most common
575 		 * path for page migration.
576 		 */
577 		rc = mapping->a_ops->migratepage(mapping,
578 						newpage, page);
579 	else
580 		rc = fallback_migrate_page(mapping, newpage, page);
581 
582 	if (!rc)
583 		remove_migration_ptes(page, newpage);
584 	else
585 		newpage->mapping = NULL;
586 
587 	unlock_page(newpage);
588 
589 	return rc;
590 }
591 
592 /*
593  * Obtain the lock on page, remove all ptes and migrate the page
594  * to the newly allocated page in newpage.
595  */
596 static int unmap_and_move(new_page_t get_new_page, unsigned long private,
597 			struct page *page, int force)
598 {
599 	int rc = 0;
600 	int *result = NULL;
601 	struct page *newpage = get_new_page(page, private, &result);
602 
603 	if (!newpage)
604 		return -ENOMEM;
605 
606 	if (page_count(page) == 1)
607 		/* page was freed from under us. So we are done. */
608 		goto move_newpage;
609 
610 	rc = -EAGAIN;
611 	if (TestSetPageLocked(page)) {
612 		if (!force)
613 			goto move_newpage;
614 		lock_page(page);
615 	}
616 
617 	if (PageWriteback(page)) {
618 		if (!force)
619 			goto unlock;
620 		wait_on_page_writeback(page);
621 	}
622 
623 	/*
624 	 * Establish migration ptes or remove ptes
625 	 */
626 	try_to_unmap(page, 1);
627 	if (!page_mapped(page))
628 		rc = move_to_new_page(newpage, page);
629 
630 	if (rc)
631 		remove_migration_ptes(page, page);
632 
633 unlock:
634 	unlock_page(page);
635 
636 	if (rc != -EAGAIN) {
637  		/*
638  		 * A page that has been migrated has all references
639  		 * removed and will be freed. A page that has not been
640  		 * migrated will have kepts its references and be
641  		 * restored.
642  		 */
643  		list_del(&page->lru);
644  		move_to_lru(page);
645 	}
646 
647 move_newpage:
648 	/*
649 	 * Move the new page to the LRU. If migration was not successful
650 	 * then this will free the page.
651 	 */
652 	move_to_lru(newpage);
653 	if (result) {
654 		if (rc)
655 			*result = rc;
656 		else
657 			*result = page_to_nid(newpage);
658 	}
659 	return rc;
660 }
661 
662 /*
663  * migrate_pages
664  *
665  * The function takes one list of pages to migrate and a function
666  * that determines from the page to be migrated and the private data
667  * the target of the move and allocates the page.
668  *
669  * The function returns after 10 attempts or if no pages
670  * are movable anymore because to has become empty
671  * or no retryable pages exist anymore. All pages will be
672  * retruned to the LRU or freed.
673  *
674  * Return: Number of pages not migrated or error code.
675  */
676 int migrate_pages(struct list_head *from,
677 		new_page_t get_new_page, unsigned long private)
678 {
679 	int retry = 1;
680 	int nr_failed = 0;
681 	int pass = 0;
682 	struct page *page;
683 	struct page *page2;
684 	int swapwrite = current->flags & PF_SWAPWRITE;
685 	int rc;
686 
687 	if (!swapwrite)
688 		current->flags |= PF_SWAPWRITE;
689 
690 	for(pass = 0; pass < 10 && retry; pass++) {
691 		retry = 0;
692 
693 		list_for_each_entry_safe(page, page2, from, lru) {
694 			cond_resched();
695 
696 			rc = unmap_and_move(get_new_page, private,
697 						page, pass > 2);
698 
699 			switch(rc) {
700 			case -ENOMEM:
701 				goto out;
702 			case -EAGAIN:
703 				retry++;
704 				break;
705 			case 0:
706 				break;
707 			default:
708 				/* Permanent failure */
709 				nr_failed++;
710 				break;
711 			}
712 		}
713 	}
714 	rc = 0;
715 out:
716 	if (!swapwrite)
717 		current->flags &= ~PF_SWAPWRITE;
718 
719 	putback_lru_pages(from);
720 
721 	if (rc)
722 		return rc;
723 
724 	return nr_failed + retry;
725 }
726 
727 #ifdef CONFIG_NUMA
728 /*
729  * Move a list of individual pages
730  */
731 struct page_to_node {
732 	unsigned long addr;
733 	struct page *page;
734 	int node;
735 	int status;
736 };
737 
738 static struct page *new_page_node(struct page *p, unsigned long private,
739 		int **result)
740 {
741 	struct page_to_node *pm = (struct page_to_node *)private;
742 
743 	while (pm->node != MAX_NUMNODES && pm->page != p)
744 		pm++;
745 
746 	if (pm->node == MAX_NUMNODES)
747 		return NULL;
748 
749 	*result = &pm->status;
750 
751 	return alloc_pages_node(pm->node, GFP_HIGHUSER | GFP_THISNODE, 0);
752 }
753 
754 /*
755  * Move a set of pages as indicated in the pm array. The addr
756  * field must be set to the virtual address of the page to be moved
757  * and the node number must contain a valid target node.
758  */
759 static int do_move_pages(struct mm_struct *mm, struct page_to_node *pm,
760 				int migrate_all)
761 {
762 	int err;
763 	struct page_to_node *pp;
764 	LIST_HEAD(pagelist);
765 
766 	down_read(&mm->mmap_sem);
767 
768 	/*
769 	 * Build a list of pages to migrate
770 	 */
771 	migrate_prep();
772 	for (pp = pm; pp->node != MAX_NUMNODES; pp++) {
773 		struct vm_area_struct *vma;
774 		struct page *page;
775 
776 		/*
777 		 * A valid page pointer that will not match any of the
778 		 * pages that will be moved.
779 		 */
780 		pp->page = ZERO_PAGE(0);
781 
782 		err = -EFAULT;
783 		vma = find_vma(mm, pp->addr);
784 		if (!vma || !vma_migratable(vma))
785 			goto set_status;
786 
787 		page = follow_page(vma, pp->addr, FOLL_GET);
788 		err = -ENOENT;
789 		if (!page)
790 			goto set_status;
791 
792 		if (PageReserved(page))		/* Check for zero page */
793 			goto put_and_set;
794 
795 		pp->page = page;
796 		err = page_to_nid(page);
797 
798 		if (err == pp->node)
799 			/*
800 			 * Node already in the right place
801 			 */
802 			goto put_and_set;
803 
804 		err = -EACCES;
805 		if (page_mapcount(page) > 1 &&
806 				!migrate_all)
807 			goto put_and_set;
808 
809 		err = isolate_lru_page(page, &pagelist);
810 put_and_set:
811 		/*
812 		 * Either remove the duplicate refcount from
813 		 * isolate_lru_page() or drop the page ref if it was
814 		 * not isolated.
815 		 */
816 		put_page(page);
817 set_status:
818 		pp->status = err;
819 	}
820 
821 	if (!list_empty(&pagelist))
822 		err = migrate_pages(&pagelist, new_page_node,
823 				(unsigned long)pm);
824 	else
825 		err = -ENOENT;
826 
827 	up_read(&mm->mmap_sem);
828 	return err;
829 }
830 
831 /*
832  * Determine the nodes of a list of pages. The addr in the pm array
833  * must have been set to the virtual address of which we want to determine
834  * the node number.
835  */
836 static int do_pages_stat(struct mm_struct *mm, struct page_to_node *pm)
837 {
838 	down_read(&mm->mmap_sem);
839 
840 	for ( ; pm->node != MAX_NUMNODES; pm++) {
841 		struct vm_area_struct *vma;
842 		struct page *page;
843 		int err;
844 
845 		err = -EFAULT;
846 		vma = find_vma(mm, pm->addr);
847 		if (!vma)
848 			goto set_status;
849 
850 		page = follow_page(vma, pm->addr, 0);
851 		err = -ENOENT;
852 		/* Use PageReserved to check for zero page */
853 		if (!page || PageReserved(page))
854 			goto set_status;
855 
856 		err = page_to_nid(page);
857 set_status:
858 		pm->status = err;
859 	}
860 
861 	up_read(&mm->mmap_sem);
862 	return 0;
863 }
864 
865 /*
866  * Move a list of pages in the address space of the currently executing
867  * process.
868  */
869 asmlinkage long sys_move_pages(pid_t pid, unsigned long nr_pages,
870 			const void __user * __user *pages,
871 			const int __user *nodes,
872 			int __user *status, int flags)
873 {
874 	int err = 0;
875 	int i;
876 	struct task_struct *task;
877 	nodemask_t task_nodes;
878 	struct mm_struct *mm;
879 	struct page_to_node *pm = NULL;
880 
881 	/* Check flags */
882 	if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
883 		return -EINVAL;
884 
885 	if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
886 		return -EPERM;
887 
888 	/* Find the mm_struct */
889 	read_lock(&tasklist_lock);
890 	task = pid ? find_task_by_pid(pid) : current;
891 	if (!task) {
892 		read_unlock(&tasklist_lock);
893 		return -ESRCH;
894 	}
895 	mm = get_task_mm(task);
896 	read_unlock(&tasklist_lock);
897 
898 	if (!mm)
899 		return -EINVAL;
900 
901 	/*
902 	 * Check if this process has the right to modify the specified
903 	 * process. The right exists if the process has administrative
904 	 * capabilities, superuser privileges or the same
905 	 * userid as the target process.
906 	 */
907 	if ((current->euid != task->suid) && (current->euid != task->uid) &&
908 	    (current->uid != task->suid) && (current->uid != task->uid) &&
909 	    !capable(CAP_SYS_NICE)) {
910 		err = -EPERM;
911 		goto out2;
912 	}
913 
914  	err = security_task_movememory(task);
915  	if (err)
916  		goto out2;
917 
918 
919 	task_nodes = cpuset_mems_allowed(task);
920 
921 	/* Limit nr_pages so that the multiplication may not overflow */
922 	if (nr_pages >= ULONG_MAX / sizeof(struct page_to_node) - 1) {
923 		err = -E2BIG;
924 		goto out2;
925 	}
926 
927 	pm = vmalloc((nr_pages + 1) * sizeof(struct page_to_node));
928 	if (!pm) {
929 		err = -ENOMEM;
930 		goto out2;
931 	}
932 
933 	/*
934 	 * Get parameters from user space and initialize the pm
935 	 * array. Return various errors if the user did something wrong.
936 	 */
937 	for (i = 0; i < nr_pages; i++) {
938 		const void *p;
939 
940 		err = -EFAULT;
941 		if (get_user(p, pages + i))
942 			goto out;
943 
944 		pm[i].addr = (unsigned long)p;
945 		if (nodes) {
946 			int node;
947 
948 			if (get_user(node, nodes + i))
949 				goto out;
950 
951 			err = -ENODEV;
952 			if (!node_online(node))
953 				goto out;
954 
955 			err = -EACCES;
956 			if (!node_isset(node, task_nodes))
957 				goto out;
958 
959 			pm[i].node = node;
960 		} else
961 			pm[i].node = 0;	/* anything to not match MAX_NUMNODES */
962 	}
963 	/* End marker */
964 	pm[nr_pages].node = MAX_NUMNODES;
965 
966 	if (nodes)
967 		err = do_move_pages(mm, pm, flags & MPOL_MF_MOVE_ALL);
968 	else
969 		err = do_pages_stat(mm, pm);
970 
971 	if (err >= 0)
972 		/* Return status information */
973 		for (i = 0; i < nr_pages; i++)
974 			if (put_user(pm[i].status, status + i))
975 				err = -EFAULT;
976 
977 out:
978 	vfree(pm);
979 out2:
980 	mmput(mm);
981 	return err;
982 }
983 #endif
984 
985 /*
986  * Call migration functions in the vma_ops that may prepare
987  * memory in a vm for migration. migration functions may perform
988  * the migration for vmas that do not have an underlying page struct.
989  */
990 int migrate_vmas(struct mm_struct *mm, const nodemask_t *to,
991 	const nodemask_t *from, unsigned long flags)
992 {
993  	struct vm_area_struct *vma;
994  	int err = 0;
995 
996  	for(vma = mm->mmap; vma->vm_next && !err; vma = vma->vm_next) {
997  		if (vma->vm_ops && vma->vm_ops->migrate) {
998  			err = vma->vm_ops->migrate(vma, to, from, flags);
999  			if (err)
1000  				break;
1001  		}
1002  	}
1003  	return err;
1004 }
1005