xref: /openbmc/linux/mm/migrate.c (revision 84764a41)
1 /*
2  * Memory Migration functionality - linux/mm/migration.c
3  *
4  * Copyright (C) 2006 Silicon Graphics, Inc., Christoph Lameter
5  *
6  * Page migration was first developed in the context of the memory hotplug
7  * project. The main authors of the migration code are:
8  *
9  * IWAMOTO Toshihiro <iwamoto@valinux.co.jp>
10  * Hirokazu Takahashi <taka@valinux.co.jp>
11  * Dave Hansen <haveblue@us.ibm.com>
12  * Christoph Lameter
13  */
14 
15 #include <linux/migrate.h>
16 #include <linux/export.h>
17 #include <linux/swap.h>
18 #include <linux/swapops.h>
19 #include <linux/pagemap.h>
20 #include <linux/buffer_head.h>
21 #include <linux/mm_inline.h>
22 #include <linux/nsproxy.h>
23 #include <linux/pagevec.h>
24 #include <linux/ksm.h>
25 #include <linux/rmap.h>
26 #include <linux/topology.h>
27 #include <linux/cpu.h>
28 #include <linux/cpuset.h>
29 #include <linux/writeback.h>
30 #include <linux/mempolicy.h>
31 #include <linux/vmalloc.h>
32 #include <linux/security.h>
33 #include <linux/memcontrol.h>
34 #include <linux/syscalls.h>
35 #include <linux/hugetlb.h>
36 #include <linux/hugetlb_cgroup.h>
37 #include <linux/gfp.h>
38 
39 #include <asm/tlbflush.h>
40 
41 #include "internal.h"
42 
43 /*
44  * migrate_prep() needs to be called before we start compiling a list of pages
45  * to be migrated using isolate_lru_page(). If scheduling work on other CPUs is
46  * undesirable, use migrate_prep_local()
47  */
48 int migrate_prep(void)
49 {
50 	/*
51 	 * Clear the LRU lists so pages can be isolated.
52 	 * Note that pages may be moved off the LRU after we have
53 	 * drained them. Those pages will fail to migrate like other
54 	 * pages that may be busy.
55 	 */
56 	lru_add_drain_all();
57 
58 	return 0;
59 }
60 
61 /* Do the necessary work of migrate_prep but not if it involves other CPUs */
62 int migrate_prep_local(void)
63 {
64 	lru_add_drain();
65 
66 	return 0;
67 }
68 
69 /*
70  * Add isolated pages on the list back to the LRU under page lock
71  * to avoid leaking evictable pages back onto unevictable list.
72  */
73 void putback_lru_pages(struct list_head *l)
74 {
75 	struct page *page;
76 	struct page *page2;
77 
78 	list_for_each_entry_safe(page, page2, l, lru) {
79 		list_del(&page->lru);
80 		dec_zone_page_state(page, NR_ISOLATED_ANON +
81 				page_is_file_cache(page));
82 		putback_lru_page(page);
83 	}
84 }
85 
86 /*
87  * Restore a potential migration pte to a working pte entry
88  */
89 static int remove_migration_pte(struct page *new, struct vm_area_struct *vma,
90 				 unsigned long addr, void *old)
91 {
92 	struct mm_struct *mm = vma->vm_mm;
93 	swp_entry_t entry;
94  	pgd_t *pgd;
95  	pud_t *pud;
96  	pmd_t *pmd;
97 	pte_t *ptep, pte;
98  	spinlock_t *ptl;
99 
100 	if (unlikely(PageHuge(new))) {
101 		ptep = huge_pte_offset(mm, addr);
102 		if (!ptep)
103 			goto out;
104 		ptl = &mm->page_table_lock;
105 	} else {
106 		pgd = pgd_offset(mm, addr);
107 		if (!pgd_present(*pgd))
108 			goto out;
109 
110 		pud = pud_offset(pgd, addr);
111 		if (!pud_present(*pud))
112 			goto out;
113 
114 		pmd = pmd_offset(pud, addr);
115 		if (pmd_trans_huge(*pmd))
116 			goto out;
117 		if (!pmd_present(*pmd))
118 			goto out;
119 
120 		ptep = pte_offset_map(pmd, addr);
121 
122 		/*
123 		 * Peek to check is_swap_pte() before taking ptlock?  No, we
124 		 * can race mremap's move_ptes(), which skips anon_vma lock.
125 		 */
126 
127 		ptl = pte_lockptr(mm, pmd);
128 	}
129 
130  	spin_lock(ptl);
131 	pte = *ptep;
132 	if (!is_swap_pte(pte))
133 		goto unlock;
134 
135 	entry = pte_to_swp_entry(pte);
136 
137 	if (!is_migration_entry(entry) ||
138 	    migration_entry_to_page(entry) != old)
139 		goto unlock;
140 
141 	get_page(new);
142 	pte = pte_mkold(mk_pte(new, vma->vm_page_prot));
143 	if (is_write_migration_entry(entry))
144 		pte = pte_mkwrite(pte);
145 #ifdef CONFIG_HUGETLB_PAGE
146 	if (PageHuge(new))
147 		pte = pte_mkhuge(pte);
148 #endif
149 	flush_cache_page(vma, addr, pte_pfn(pte));
150 	set_pte_at(mm, addr, ptep, pte);
151 
152 	if (PageHuge(new)) {
153 		if (PageAnon(new))
154 			hugepage_add_anon_rmap(new, vma, addr);
155 		else
156 			page_dup_rmap(new);
157 	} else if (PageAnon(new))
158 		page_add_anon_rmap(new, vma, addr);
159 	else
160 		page_add_file_rmap(new);
161 
162 	/* No need to invalidate - it was non-present before */
163 	update_mmu_cache(vma, addr, ptep);
164 unlock:
165 	pte_unmap_unlock(ptep, ptl);
166 out:
167 	return SWAP_AGAIN;
168 }
169 
170 /*
171  * Get rid of all migration entries and replace them by
172  * references to the indicated page.
173  */
174 static void remove_migration_ptes(struct page *old, struct page *new)
175 {
176 	rmap_walk(new, remove_migration_pte, old);
177 }
178 
179 /*
180  * Something used the pte of a page under migration. We need to
181  * get to the page and wait until migration is finished.
182  * When we return from this function the fault will be retried.
183  */
184 void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
185 				unsigned long address)
186 {
187 	pte_t *ptep, pte;
188 	spinlock_t *ptl;
189 	swp_entry_t entry;
190 	struct page *page;
191 
192 	ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
193 	pte = *ptep;
194 	if (!is_swap_pte(pte))
195 		goto out;
196 
197 	entry = pte_to_swp_entry(pte);
198 	if (!is_migration_entry(entry))
199 		goto out;
200 
201 	page = migration_entry_to_page(entry);
202 
203 	/*
204 	 * Once radix-tree replacement of page migration started, page_count
205 	 * *must* be zero. And, we don't want to call wait_on_page_locked()
206 	 * against a page without get_page().
207 	 * So, we use get_page_unless_zero(), here. Even failed, page fault
208 	 * will occur again.
209 	 */
210 	if (!get_page_unless_zero(page))
211 		goto out;
212 	pte_unmap_unlock(ptep, ptl);
213 	wait_on_page_locked(page);
214 	put_page(page);
215 	return;
216 out:
217 	pte_unmap_unlock(ptep, ptl);
218 }
219 
220 #ifdef CONFIG_BLOCK
221 /* Returns true if all buffers are successfully locked */
222 static bool buffer_migrate_lock_buffers(struct buffer_head *head,
223 							enum migrate_mode mode)
224 {
225 	struct buffer_head *bh = head;
226 
227 	/* Simple case, sync compaction */
228 	if (mode != MIGRATE_ASYNC) {
229 		do {
230 			get_bh(bh);
231 			lock_buffer(bh);
232 			bh = bh->b_this_page;
233 
234 		} while (bh != head);
235 
236 		return true;
237 	}
238 
239 	/* async case, we cannot block on lock_buffer so use trylock_buffer */
240 	do {
241 		get_bh(bh);
242 		if (!trylock_buffer(bh)) {
243 			/*
244 			 * We failed to lock the buffer and cannot stall in
245 			 * async migration. Release the taken locks
246 			 */
247 			struct buffer_head *failed_bh = bh;
248 			put_bh(failed_bh);
249 			bh = head;
250 			while (bh != failed_bh) {
251 				unlock_buffer(bh);
252 				put_bh(bh);
253 				bh = bh->b_this_page;
254 			}
255 			return false;
256 		}
257 
258 		bh = bh->b_this_page;
259 	} while (bh != head);
260 	return true;
261 }
262 #else
263 static inline bool buffer_migrate_lock_buffers(struct buffer_head *head,
264 							enum migrate_mode mode)
265 {
266 	return true;
267 }
268 #endif /* CONFIG_BLOCK */
269 
270 /*
271  * Replace the page in the mapping.
272  *
273  * The number of remaining references must be:
274  * 1 for anonymous pages without a mapping
275  * 2 for pages with a mapping
276  * 3 for pages with a mapping and PagePrivate/PagePrivate2 set.
277  */
278 static int migrate_page_move_mapping(struct address_space *mapping,
279 		struct page *newpage, struct page *page,
280 		struct buffer_head *head, enum migrate_mode mode)
281 {
282 	int expected_count;
283 	void **pslot;
284 
285 	if (!mapping) {
286 		/* Anonymous page without mapping */
287 		if (page_count(page) != 1)
288 			return -EAGAIN;
289 		return 0;
290 	}
291 
292 	spin_lock_irq(&mapping->tree_lock);
293 
294 	pslot = radix_tree_lookup_slot(&mapping->page_tree,
295  					page_index(page));
296 
297 	expected_count = 2 + page_has_private(page);
298 	if (page_count(page) != expected_count ||
299 		radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) {
300 		spin_unlock_irq(&mapping->tree_lock);
301 		return -EAGAIN;
302 	}
303 
304 	if (!page_freeze_refs(page, expected_count)) {
305 		spin_unlock_irq(&mapping->tree_lock);
306 		return -EAGAIN;
307 	}
308 
309 	/*
310 	 * In the async migration case of moving a page with buffers, lock the
311 	 * buffers using trylock before the mapping is moved. If the mapping
312 	 * was moved, we later failed to lock the buffers and could not move
313 	 * the mapping back due to an elevated page count, we would have to
314 	 * block waiting on other references to be dropped.
315 	 */
316 	if (mode == MIGRATE_ASYNC && head &&
317 			!buffer_migrate_lock_buffers(head, mode)) {
318 		page_unfreeze_refs(page, expected_count);
319 		spin_unlock_irq(&mapping->tree_lock);
320 		return -EAGAIN;
321 	}
322 
323 	/*
324 	 * Now we know that no one else is looking at the page.
325 	 */
326 	get_page(newpage);	/* add cache reference */
327 	if (PageSwapCache(page)) {
328 		SetPageSwapCache(newpage);
329 		set_page_private(newpage, page_private(page));
330 	}
331 
332 	radix_tree_replace_slot(pslot, newpage);
333 
334 	/*
335 	 * Drop cache reference from old page by unfreezing
336 	 * to one less reference.
337 	 * We know this isn't the last reference.
338 	 */
339 	page_unfreeze_refs(page, expected_count - 1);
340 
341 	/*
342 	 * If moved to a different zone then also account
343 	 * the page for that zone. Other VM counters will be
344 	 * taken care of when we establish references to the
345 	 * new page and drop references to the old page.
346 	 *
347 	 * Note that anonymous pages are accounted for
348 	 * via NR_FILE_PAGES and NR_ANON_PAGES if they
349 	 * are mapped to swap space.
350 	 */
351 	__dec_zone_page_state(page, NR_FILE_PAGES);
352 	__inc_zone_page_state(newpage, NR_FILE_PAGES);
353 	if (!PageSwapCache(page) && PageSwapBacked(page)) {
354 		__dec_zone_page_state(page, NR_SHMEM);
355 		__inc_zone_page_state(newpage, NR_SHMEM);
356 	}
357 	spin_unlock_irq(&mapping->tree_lock);
358 
359 	return 0;
360 }
361 
362 /*
363  * The expected number of remaining references is the same as that
364  * of migrate_page_move_mapping().
365  */
366 int migrate_huge_page_move_mapping(struct address_space *mapping,
367 				   struct page *newpage, struct page *page)
368 {
369 	int expected_count;
370 	void **pslot;
371 
372 	if (!mapping) {
373 		if (page_count(page) != 1)
374 			return -EAGAIN;
375 		return 0;
376 	}
377 
378 	spin_lock_irq(&mapping->tree_lock);
379 
380 	pslot = radix_tree_lookup_slot(&mapping->page_tree,
381 					page_index(page));
382 
383 	expected_count = 2 + page_has_private(page);
384 	if (page_count(page) != expected_count ||
385 		radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) {
386 		spin_unlock_irq(&mapping->tree_lock);
387 		return -EAGAIN;
388 	}
389 
390 	if (!page_freeze_refs(page, expected_count)) {
391 		spin_unlock_irq(&mapping->tree_lock);
392 		return -EAGAIN;
393 	}
394 
395 	get_page(newpage);
396 
397 	radix_tree_replace_slot(pslot, newpage);
398 
399 	page_unfreeze_refs(page, expected_count - 1);
400 
401 	spin_unlock_irq(&mapping->tree_lock);
402 	return 0;
403 }
404 
405 /*
406  * Copy the page to its new location
407  */
408 void migrate_page_copy(struct page *newpage, struct page *page)
409 {
410 	if (PageHuge(page))
411 		copy_huge_page(newpage, page);
412 	else
413 		copy_highpage(newpage, page);
414 
415 	if (PageError(page))
416 		SetPageError(newpage);
417 	if (PageReferenced(page))
418 		SetPageReferenced(newpage);
419 	if (PageUptodate(page))
420 		SetPageUptodate(newpage);
421 	if (TestClearPageActive(page)) {
422 		VM_BUG_ON(PageUnevictable(page));
423 		SetPageActive(newpage);
424 	} else if (TestClearPageUnevictable(page))
425 		SetPageUnevictable(newpage);
426 	if (PageChecked(page))
427 		SetPageChecked(newpage);
428 	if (PageMappedToDisk(page))
429 		SetPageMappedToDisk(newpage);
430 
431 	if (PageDirty(page)) {
432 		clear_page_dirty_for_io(page);
433 		/*
434 		 * Want to mark the page and the radix tree as dirty, and
435 		 * redo the accounting that clear_page_dirty_for_io undid,
436 		 * but we can't use set_page_dirty because that function
437 		 * is actually a signal that all of the page has become dirty.
438 		 * Whereas only part of our page may be dirty.
439 		 */
440 		if (PageSwapBacked(page))
441 			SetPageDirty(newpage);
442 		else
443 			__set_page_dirty_nobuffers(newpage);
444  	}
445 
446 	mlock_migrate_page(newpage, page);
447 	ksm_migrate_page(newpage, page);
448 
449 	ClearPageSwapCache(page);
450 	ClearPagePrivate(page);
451 	set_page_private(page, 0);
452 
453 	/*
454 	 * If any waiters have accumulated on the new page then
455 	 * wake them up.
456 	 */
457 	if (PageWriteback(newpage))
458 		end_page_writeback(newpage);
459 }
460 
461 /************************************************************
462  *                    Migration functions
463  ***********************************************************/
464 
465 /* Always fail migration. Used for mappings that are not movable */
466 int fail_migrate_page(struct address_space *mapping,
467 			struct page *newpage, struct page *page)
468 {
469 	return -EIO;
470 }
471 EXPORT_SYMBOL(fail_migrate_page);
472 
473 /*
474  * Common logic to directly migrate a single page suitable for
475  * pages that do not use PagePrivate/PagePrivate2.
476  *
477  * Pages are locked upon entry and exit.
478  */
479 int migrate_page(struct address_space *mapping,
480 		struct page *newpage, struct page *page,
481 		enum migrate_mode mode)
482 {
483 	int rc;
484 
485 	BUG_ON(PageWriteback(page));	/* Writeback must be complete */
486 
487 	rc = migrate_page_move_mapping(mapping, newpage, page, NULL, mode);
488 
489 	if (rc)
490 		return rc;
491 
492 	migrate_page_copy(newpage, page);
493 	return 0;
494 }
495 EXPORT_SYMBOL(migrate_page);
496 
497 #ifdef CONFIG_BLOCK
498 /*
499  * Migration function for pages with buffers. This function can only be used
500  * if the underlying filesystem guarantees that no other references to "page"
501  * exist.
502  */
503 int buffer_migrate_page(struct address_space *mapping,
504 		struct page *newpage, struct page *page, enum migrate_mode mode)
505 {
506 	struct buffer_head *bh, *head;
507 	int rc;
508 
509 	if (!page_has_buffers(page))
510 		return migrate_page(mapping, newpage, page, mode);
511 
512 	head = page_buffers(page);
513 
514 	rc = migrate_page_move_mapping(mapping, newpage, page, head, mode);
515 
516 	if (rc)
517 		return rc;
518 
519 	/*
520 	 * In the async case, migrate_page_move_mapping locked the buffers
521 	 * with an IRQ-safe spinlock held. In the sync case, the buffers
522 	 * need to be locked now
523 	 */
524 	if (mode != MIGRATE_ASYNC)
525 		BUG_ON(!buffer_migrate_lock_buffers(head, mode));
526 
527 	ClearPagePrivate(page);
528 	set_page_private(newpage, page_private(page));
529 	set_page_private(page, 0);
530 	put_page(page);
531 	get_page(newpage);
532 
533 	bh = head;
534 	do {
535 		set_bh_page(bh, newpage, bh_offset(bh));
536 		bh = bh->b_this_page;
537 
538 	} while (bh != head);
539 
540 	SetPagePrivate(newpage);
541 
542 	migrate_page_copy(newpage, page);
543 
544 	bh = head;
545 	do {
546 		unlock_buffer(bh);
547  		put_bh(bh);
548 		bh = bh->b_this_page;
549 
550 	} while (bh != head);
551 
552 	return 0;
553 }
554 EXPORT_SYMBOL(buffer_migrate_page);
555 #endif
556 
557 /*
558  * Writeback a page to clean the dirty state
559  */
560 static int writeout(struct address_space *mapping, struct page *page)
561 {
562 	struct writeback_control wbc = {
563 		.sync_mode = WB_SYNC_NONE,
564 		.nr_to_write = 1,
565 		.range_start = 0,
566 		.range_end = LLONG_MAX,
567 		.for_reclaim = 1
568 	};
569 	int rc;
570 
571 	if (!mapping->a_ops->writepage)
572 		/* No write method for the address space */
573 		return -EINVAL;
574 
575 	if (!clear_page_dirty_for_io(page))
576 		/* Someone else already triggered a write */
577 		return -EAGAIN;
578 
579 	/*
580 	 * A dirty page may imply that the underlying filesystem has
581 	 * the page on some queue. So the page must be clean for
582 	 * migration. Writeout may mean we loose the lock and the
583 	 * page state is no longer what we checked for earlier.
584 	 * At this point we know that the migration attempt cannot
585 	 * be successful.
586 	 */
587 	remove_migration_ptes(page, page);
588 
589 	rc = mapping->a_ops->writepage(page, &wbc);
590 
591 	if (rc != AOP_WRITEPAGE_ACTIVATE)
592 		/* unlocked. Relock */
593 		lock_page(page);
594 
595 	return (rc < 0) ? -EIO : -EAGAIN;
596 }
597 
598 /*
599  * Default handling if a filesystem does not provide a migration function.
600  */
601 static int fallback_migrate_page(struct address_space *mapping,
602 	struct page *newpage, struct page *page, enum migrate_mode mode)
603 {
604 	if (PageDirty(page)) {
605 		/* Only writeback pages in full synchronous migration */
606 		if (mode != MIGRATE_SYNC)
607 			return -EBUSY;
608 		return writeout(mapping, page);
609 	}
610 
611 	/*
612 	 * Buffers may be managed in a filesystem specific way.
613 	 * We must have no buffers or drop them.
614 	 */
615 	if (page_has_private(page) &&
616 	    !try_to_release_page(page, GFP_KERNEL))
617 		return -EAGAIN;
618 
619 	return migrate_page(mapping, newpage, page, mode);
620 }
621 
622 /*
623  * Move a page to a newly allocated page
624  * The page is locked and all ptes have been successfully removed.
625  *
626  * The new page will have replaced the old page if this function
627  * is successful.
628  *
629  * Return value:
630  *   < 0 - error code
631  *  == 0 - success
632  */
633 static int move_to_new_page(struct page *newpage, struct page *page,
634 				int remap_swapcache, enum migrate_mode mode)
635 {
636 	struct address_space *mapping;
637 	int rc;
638 
639 	/*
640 	 * Block others from accessing the page when we get around to
641 	 * establishing additional references. We are the only one
642 	 * holding a reference to the new page at this point.
643 	 */
644 	if (!trylock_page(newpage))
645 		BUG();
646 
647 	/* Prepare mapping for the new page.*/
648 	newpage->index = page->index;
649 	newpage->mapping = page->mapping;
650 	if (PageSwapBacked(page))
651 		SetPageSwapBacked(newpage);
652 
653 	mapping = page_mapping(page);
654 	if (!mapping)
655 		rc = migrate_page(mapping, newpage, page, mode);
656 	else if (mapping->a_ops->migratepage)
657 		/*
658 		 * Most pages have a mapping and most filesystems provide a
659 		 * migratepage callback. Anonymous pages are part of swap
660 		 * space which also has its own migratepage callback. This
661 		 * is the most common path for page migration.
662 		 */
663 		rc = mapping->a_ops->migratepage(mapping,
664 						newpage, page, mode);
665 	else
666 		rc = fallback_migrate_page(mapping, newpage, page, mode);
667 
668 	if (rc) {
669 		newpage->mapping = NULL;
670 	} else {
671 		if (remap_swapcache)
672 			remove_migration_ptes(page, newpage);
673 		page->mapping = NULL;
674 	}
675 
676 	unlock_page(newpage);
677 
678 	return rc;
679 }
680 
681 static int __unmap_and_move(struct page *page, struct page *newpage,
682 			int force, bool offlining, enum migrate_mode mode)
683 {
684 	int rc = -EAGAIN;
685 	int remap_swapcache = 1;
686 	struct mem_cgroup *mem;
687 	struct anon_vma *anon_vma = NULL;
688 
689 	if (!trylock_page(page)) {
690 		if (!force || mode == MIGRATE_ASYNC)
691 			goto out;
692 
693 		/*
694 		 * It's not safe for direct compaction to call lock_page.
695 		 * For example, during page readahead pages are added locked
696 		 * to the LRU. Later, when the IO completes the pages are
697 		 * marked uptodate and unlocked. However, the queueing
698 		 * could be merging multiple pages for one bio (e.g.
699 		 * mpage_readpages). If an allocation happens for the
700 		 * second or third page, the process can end up locking
701 		 * the same page twice and deadlocking. Rather than
702 		 * trying to be clever about what pages can be locked,
703 		 * avoid the use of lock_page for direct compaction
704 		 * altogether.
705 		 */
706 		if (current->flags & PF_MEMALLOC)
707 			goto out;
708 
709 		lock_page(page);
710 	}
711 
712 	/*
713 	 * Only memory hotplug's offline_pages() caller has locked out KSM,
714 	 * and can safely migrate a KSM page.  The other cases have skipped
715 	 * PageKsm along with PageReserved - but it is only now when we have
716 	 * the page lock that we can be certain it will not go KSM beneath us
717 	 * (KSM will not upgrade a page from PageAnon to PageKsm when it sees
718 	 * its pagecount raised, but only here do we take the page lock which
719 	 * serializes that).
720 	 */
721 	if (PageKsm(page) && !offlining) {
722 		rc = -EBUSY;
723 		goto unlock;
724 	}
725 
726 	/* charge against new page */
727 	mem_cgroup_prepare_migration(page, newpage, &mem);
728 
729 	if (PageWriteback(page)) {
730 		/*
731 		 * Only in the case of a full syncronous migration is it
732 		 * necessary to wait for PageWriteback. In the async case,
733 		 * the retry loop is too short and in the sync-light case,
734 		 * the overhead of stalling is too much
735 		 */
736 		if (mode != MIGRATE_SYNC) {
737 			rc = -EBUSY;
738 			goto uncharge;
739 		}
740 		if (!force)
741 			goto uncharge;
742 		wait_on_page_writeback(page);
743 	}
744 	/*
745 	 * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
746 	 * we cannot notice that anon_vma is freed while we migrates a page.
747 	 * This get_anon_vma() delays freeing anon_vma pointer until the end
748 	 * of migration. File cache pages are no problem because of page_lock()
749 	 * File Caches may use write_page() or lock_page() in migration, then,
750 	 * just care Anon page here.
751 	 */
752 	if (PageAnon(page)) {
753 		/*
754 		 * Only page_lock_anon_vma() understands the subtleties of
755 		 * getting a hold on an anon_vma from outside one of its mms.
756 		 */
757 		anon_vma = page_get_anon_vma(page);
758 		if (anon_vma) {
759 			/*
760 			 * Anon page
761 			 */
762 		} else if (PageSwapCache(page)) {
763 			/*
764 			 * We cannot be sure that the anon_vma of an unmapped
765 			 * swapcache page is safe to use because we don't
766 			 * know in advance if the VMA that this page belonged
767 			 * to still exists. If the VMA and others sharing the
768 			 * data have been freed, then the anon_vma could
769 			 * already be invalid.
770 			 *
771 			 * To avoid this possibility, swapcache pages get
772 			 * migrated but are not remapped when migration
773 			 * completes
774 			 */
775 			remap_swapcache = 0;
776 		} else {
777 			goto uncharge;
778 		}
779 	}
780 
781 	/*
782 	 * Corner case handling:
783 	 * 1. When a new swap-cache page is read into, it is added to the LRU
784 	 * and treated as swapcache but it has no rmap yet.
785 	 * Calling try_to_unmap() against a page->mapping==NULL page will
786 	 * trigger a BUG.  So handle it here.
787 	 * 2. An orphaned page (see truncate_complete_page) might have
788 	 * fs-private metadata. The page can be picked up due to memory
789 	 * offlining.  Everywhere else except page reclaim, the page is
790 	 * invisible to the vm, so the page can not be migrated.  So try to
791 	 * free the metadata, so the page can be freed.
792 	 */
793 	if (!page->mapping) {
794 		VM_BUG_ON(PageAnon(page));
795 		if (page_has_private(page)) {
796 			try_to_free_buffers(page);
797 			goto uncharge;
798 		}
799 		goto skip_unmap;
800 	}
801 
802 	/* Establish migration ptes or remove ptes */
803 	try_to_unmap(page, TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
804 
805 skip_unmap:
806 	if (!page_mapped(page))
807 		rc = move_to_new_page(newpage, page, remap_swapcache, mode);
808 
809 	if (rc && remap_swapcache)
810 		remove_migration_ptes(page, page);
811 
812 	/* Drop an anon_vma reference if we took one */
813 	if (anon_vma)
814 		put_anon_vma(anon_vma);
815 
816 uncharge:
817 	mem_cgroup_end_migration(mem, page, newpage, rc == 0);
818 unlock:
819 	unlock_page(page);
820 out:
821 	return rc;
822 }
823 
824 /*
825  * Obtain the lock on page, remove all ptes and migrate the page
826  * to the newly allocated page in newpage.
827  */
828 static int unmap_and_move(new_page_t get_new_page, unsigned long private,
829 			struct page *page, int force, bool offlining,
830 			enum migrate_mode mode)
831 {
832 	int rc = 0;
833 	int *result = NULL;
834 	struct page *newpage = get_new_page(page, private, &result);
835 
836 	if (!newpage)
837 		return -ENOMEM;
838 
839 	if (page_count(page) == 1) {
840 		/* page was freed from under us. So we are done. */
841 		goto out;
842 	}
843 
844 	if (unlikely(PageTransHuge(page)))
845 		if (unlikely(split_huge_page(page)))
846 			goto out;
847 
848 	rc = __unmap_and_move(page, newpage, force, offlining, mode);
849 out:
850 	if (rc != -EAGAIN) {
851 		/*
852 		 * A page that has been migrated has all references
853 		 * removed and will be freed. A page that has not been
854 		 * migrated will have kepts its references and be
855 		 * restored.
856 		 */
857 		list_del(&page->lru);
858 		dec_zone_page_state(page, NR_ISOLATED_ANON +
859 				page_is_file_cache(page));
860 		putback_lru_page(page);
861 	}
862 	/*
863 	 * Move the new page to the LRU. If migration was not successful
864 	 * then this will free the page.
865 	 */
866 	putback_lru_page(newpage);
867 	if (result) {
868 		if (rc)
869 			*result = rc;
870 		else
871 			*result = page_to_nid(newpage);
872 	}
873 	return rc;
874 }
875 
876 /*
877  * Counterpart of unmap_and_move_page() for hugepage migration.
878  *
879  * This function doesn't wait the completion of hugepage I/O
880  * because there is no race between I/O and migration for hugepage.
881  * Note that currently hugepage I/O occurs only in direct I/O
882  * where no lock is held and PG_writeback is irrelevant,
883  * and writeback status of all subpages are counted in the reference
884  * count of the head page (i.e. if all subpages of a 2MB hugepage are
885  * under direct I/O, the reference of the head page is 512 and a bit more.)
886  * This means that when we try to migrate hugepage whose subpages are
887  * doing direct I/O, some references remain after try_to_unmap() and
888  * hugepage migration fails without data corruption.
889  *
890  * There is also no race when direct I/O is issued on the page under migration,
891  * because then pte is replaced with migration swap entry and direct I/O code
892  * will wait in the page fault for migration to complete.
893  */
894 static int unmap_and_move_huge_page(new_page_t get_new_page,
895 				unsigned long private, struct page *hpage,
896 				int force, bool offlining,
897 				enum migrate_mode mode)
898 {
899 	int rc = 0;
900 	int *result = NULL;
901 	struct page *new_hpage = get_new_page(hpage, private, &result);
902 	struct anon_vma *anon_vma = NULL;
903 
904 	if (!new_hpage)
905 		return -ENOMEM;
906 
907 	rc = -EAGAIN;
908 
909 	if (!trylock_page(hpage)) {
910 		if (!force || mode != MIGRATE_SYNC)
911 			goto out;
912 		lock_page(hpage);
913 	}
914 
915 	if (PageAnon(hpage))
916 		anon_vma = page_get_anon_vma(hpage);
917 
918 	try_to_unmap(hpage, TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
919 
920 	if (!page_mapped(hpage))
921 		rc = move_to_new_page(new_hpage, hpage, 1, mode);
922 
923 	if (rc)
924 		remove_migration_ptes(hpage, hpage);
925 
926 	if (anon_vma)
927 		put_anon_vma(anon_vma);
928 
929 	if (!rc)
930 		hugetlb_cgroup_migrate(hpage, new_hpage);
931 
932 	unlock_page(hpage);
933 out:
934 	put_page(new_hpage);
935 	if (result) {
936 		if (rc)
937 			*result = rc;
938 		else
939 			*result = page_to_nid(new_hpage);
940 	}
941 	return rc;
942 }
943 
944 /*
945  * migrate_pages
946  *
947  * The function takes one list of pages to migrate and a function
948  * that determines from the page to be migrated and the private data
949  * the target of the move and allocates the page.
950  *
951  * The function returns after 10 attempts or if no pages
952  * are movable anymore because to has become empty
953  * or no retryable pages exist anymore.
954  * Caller should call putback_lru_pages to return pages to the LRU
955  * or free list only if ret != 0.
956  *
957  * Return: Number of pages not migrated or error code.
958  */
959 int migrate_pages(struct list_head *from,
960 		new_page_t get_new_page, unsigned long private, bool offlining,
961 		enum migrate_mode mode)
962 {
963 	int retry = 1;
964 	int nr_failed = 0;
965 	int pass = 0;
966 	struct page *page;
967 	struct page *page2;
968 	int swapwrite = current->flags & PF_SWAPWRITE;
969 	int rc;
970 
971 	if (!swapwrite)
972 		current->flags |= PF_SWAPWRITE;
973 
974 	for(pass = 0; pass < 10 && retry; pass++) {
975 		retry = 0;
976 
977 		list_for_each_entry_safe(page, page2, from, lru) {
978 			cond_resched();
979 
980 			rc = unmap_and_move(get_new_page, private,
981 						page, pass > 2, offlining,
982 						mode);
983 
984 			switch(rc) {
985 			case -ENOMEM:
986 				goto out;
987 			case -EAGAIN:
988 				retry++;
989 				break;
990 			case 0:
991 				break;
992 			default:
993 				/* Permanent failure */
994 				nr_failed++;
995 				break;
996 			}
997 		}
998 	}
999 	rc = 0;
1000 out:
1001 	if (!swapwrite)
1002 		current->flags &= ~PF_SWAPWRITE;
1003 
1004 	if (rc)
1005 		return rc;
1006 
1007 	return nr_failed + retry;
1008 }
1009 
1010 int migrate_huge_page(struct page *hpage, new_page_t get_new_page,
1011 		      unsigned long private, bool offlining,
1012 		      enum migrate_mode mode)
1013 {
1014 	int pass, rc;
1015 
1016 	for (pass = 0; pass < 10; pass++) {
1017 		rc = unmap_and_move_huge_page(get_new_page,
1018 					      private, hpage, pass > 2, offlining,
1019 					      mode);
1020 		switch (rc) {
1021 		case -ENOMEM:
1022 			goto out;
1023 		case -EAGAIN:
1024 			/* try again */
1025 			cond_resched();
1026 			break;
1027 		case 0:
1028 			goto out;
1029 		default:
1030 			rc = -EIO;
1031 			goto out;
1032 		}
1033 	}
1034 out:
1035 	return rc;
1036 }
1037 
1038 #ifdef CONFIG_NUMA
1039 /*
1040  * Move a list of individual pages
1041  */
1042 struct page_to_node {
1043 	unsigned long addr;
1044 	struct page *page;
1045 	int node;
1046 	int status;
1047 };
1048 
1049 static struct page *new_page_node(struct page *p, unsigned long private,
1050 		int **result)
1051 {
1052 	struct page_to_node *pm = (struct page_to_node *)private;
1053 
1054 	while (pm->node != MAX_NUMNODES && pm->page != p)
1055 		pm++;
1056 
1057 	if (pm->node == MAX_NUMNODES)
1058 		return NULL;
1059 
1060 	*result = &pm->status;
1061 
1062 	return alloc_pages_exact_node(pm->node,
1063 				GFP_HIGHUSER_MOVABLE | GFP_THISNODE, 0);
1064 }
1065 
1066 /*
1067  * Move a set of pages as indicated in the pm array. The addr
1068  * field must be set to the virtual address of the page to be moved
1069  * and the node number must contain a valid target node.
1070  * The pm array ends with node = MAX_NUMNODES.
1071  */
1072 static int do_move_page_to_node_array(struct mm_struct *mm,
1073 				      struct page_to_node *pm,
1074 				      int migrate_all)
1075 {
1076 	int err;
1077 	struct page_to_node *pp;
1078 	LIST_HEAD(pagelist);
1079 
1080 	down_read(&mm->mmap_sem);
1081 
1082 	/*
1083 	 * Build a list of pages to migrate
1084 	 */
1085 	for (pp = pm; pp->node != MAX_NUMNODES; pp++) {
1086 		struct vm_area_struct *vma;
1087 		struct page *page;
1088 
1089 		err = -EFAULT;
1090 		vma = find_vma(mm, pp->addr);
1091 		if (!vma || pp->addr < vma->vm_start || !vma_migratable(vma))
1092 			goto set_status;
1093 
1094 		page = follow_page(vma, pp->addr, FOLL_GET|FOLL_SPLIT);
1095 
1096 		err = PTR_ERR(page);
1097 		if (IS_ERR(page))
1098 			goto set_status;
1099 
1100 		err = -ENOENT;
1101 		if (!page)
1102 			goto set_status;
1103 
1104 		/* Use PageReserved to check for zero page */
1105 		if (PageReserved(page) || PageKsm(page))
1106 			goto put_and_set;
1107 
1108 		pp->page = page;
1109 		err = page_to_nid(page);
1110 
1111 		if (err == pp->node)
1112 			/*
1113 			 * Node already in the right place
1114 			 */
1115 			goto put_and_set;
1116 
1117 		err = -EACCES;
1118 		if (page_mapcount(page) > 1 &&
1119 				!migrate_all)
1120 			goto put_and_set;
1121 
1122 		err = isolate_lru_page(page);
1123 		if (!err) {
1124 			list_add_tail(&page->lru, &pagelist);
1125 			inc_zone_page_state(page, NR_ISOLATED_ANON +
1126 					    page_is_file_cache(page));
1127 		}
1128 put_and_set:
1129 		/*
1130 		 * Either remove the duplicate refcount from
1131 		 * isolate_lru_page() or drop the page ref if it was
1132 		 * not isolated.
1133 		 */
1134 		put_page(page);
1135 set_status:
1136 		pp->status = err;
1137 	}
1138 
1139 	err = 0;
1140 	if (!list_empty(&pagelist)) {
1141 		err = migrate_pages(&pagelist, new_page_node,
1142 				(unsigned long)pm, 0, MIGRATE_SYNC);
1143 		if (err)
1144 			putback_lru_pages(&pagelist);
1145 	}
1146 
1147 	up_read(&mm->mmap_sem);
1148 	return err;
1149 }
1150 
1151 /*
1152  * Migrate an array of page address onto an array of nodes and fill
1153  * the corresponding array of status.
1154  */
1155 static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes,
1156 			 unsigned long nr_pages,
1157 			 const void __user * __user *pages,
1158 			 const int __user *nodes,
1159 			 int __user *status, int flags)
1160 {
1161 	struct page_to_node *pm;
1162 	unsigned long chunk_nr_pages;
1163 	unsigned long chunk_start;
1164 	int err;
1165 
1166 	err = -ENOMEM;
1167 	pm = (struct page_to_node *)__get_free_page(GFP_KERNEL);
1168 	if (!pm)
1169 		goto out;
1170 
1171 	migrate_prep();
1172 
1173 	/*
1174 	 * Store a chunk of page_to_node array in a page,
1175 	 * but keep the last one as a marker
1176 	 */
1177 	chunk_nr_pages = (PAGE_SIZE / sizeof(struct page_to_node)) - 1;
1178 
1179 	for (chunk_start = 0;
1180 	     chunk_start < nr_pages;
1181 	     chunk_start += chunk_nr_pages) {
1182 		int j;
1183 
1184 		if (chunk_start + chunk_nr_pages > nr_pages)
1185 			chunk_nr_pages = nr_pages - chunk_start;
1186 
1187 		/* fill the chunk pm with addrs and nodes from user-space */
1188 		for (j = 0; j < chunk_nr_pages; j++) {
1189 			const void __user *p;
1190 			int node;
1191 
1192 			err = -EFAULT;
1193 			if (get_user(p, pages + j + chunk_start))
1194 				goto out_pm;
1195 			pm[j].addr = (unsigned long) p;
1196 
1197 			if (get_user(node, nodes + j + chunk_start))
1198 				goto out_pm;
1199 
1200 			err = -ENODEV;
1201 			if (node < 0 || node >= MAX_NUMNODES)
1202 				goto out_pm;
1203 
1204 			if (!node_state(node, N_HIGH_MEMORY))
1205 				goto out_pm;
1206 
1207 			err = -EACCES;
1208 			if (!node_isset(node, task_nodes))
1209 				goto out_pm;
1210 
1211 			pm[j].node = node;
1212 		}
1213 
1214 		/* End marker for this chunk */
1215 		pm[chunk_nr_pages].node = MAX_NUMNODES;
1216 
1217 		/* Migrate this chunk */
1218 		err = do_move_page_to_node_array(mm, pm,
1219 						 flags & MPOL_MF_MOVE_ALL);
1220 		if (err < 0)
1221 			goto out_pm;
1222 
1223 		/* Return status information */
1224 		for (j = 0; j < chunk_nr_pages; j++)
1225 			if (put_user(pm[j].status, status + j + chunk_start)) {
1226 				err = -EFAULT;
1227 				goto out_pm;
1228 			}
1229 	}
1230 	err = 0;
1231 
1232 out_pm:
1233 	free_page((unsigned long)pm);
1234 out:
1235 	return err;
1236 }
1237 
1238 /*
1239  * Determine the nodes of an array of pages and store it in an array of status.
1240  */
1241 static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages,
1242 				const void __user **pages, int *status)
1243 {
1244 	unsigned long i;
1245 
1246 	down_read(&mm->mmap_sem);
1247 
1248 	for (i = 0; i < nr_pages; i++) {
1249 		unsigned long addr = (unsigned long)(*pages);
1250 		struct vm_area_struct *vma;
1251 		struct page *page;
1252 		int err = -EFAULT;
1253 
1254 		vma = find_vma(mm, addr);
1255 		if (!vma || addr < vma->vm_start)
1256 			goto set_status;
1257 
1258 		page = follow_page(vma, addr, 0);
1259 
1260 		err = PTR_ERR(page);
1261 		if (IS_ERR(page))
1262 			goto set_status;
1263 
1264 		err = -ENOENT;
1265 		/* Use PageReserved to check for zero page */
1266 		if (!page || PageReserved(page) || PageKsm(page))
1267 			goto set_status;
1268 
1269 		err = page_to_nid(page);
1270 set_status:
1271 		*status = err;
1272 
1273 		pages++;
1274 		status++;
1275 	}
1276 
1277 	up_read(&mm->mmap_sem);
1278 }
1279 
1280 /*
1281  * Determine the nodes of a user array of pages and store it in
1282  * a user array of status.
1283  */
1284 static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages,
1285 			 const void __user * __user *pages,
1286 			 int __user *status)
1287 {
1288 #define DO_PAGES_STAT_CHUNK_NR 16
1289 	const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR];
1290 	int chunk_status[DO_PAGES_STAT_CHUNK_NR];
1291 
1292 	while (nr_pages) {
1293 		unsigned long chunk_nr;
1294 
1295 		chunk_nr = nr_pages;
1296 		if (chunk_nr > DO_PAGES_STAT_CHUNK_NR)
1297 			chunk_nr = DO_PAGES_STAT_CHUNK_NR;
1298 
1299 		if (copy_from_user(chunk_pages, pages, chunk_nr * sizeof(*chunk_pages)))
1300 			break;
1301 
1302 		do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status);
1303 
1304 		if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status)))
1305 			break;
1306 
1307 		pages += chunk_nr;
1308 		status += chunk_nr;
1309 		nr_pages -= chunk_nr;
1310 	}
1311 	return nr_pages ? -EFAULT : 0;
1312 }
1313 
1314 /*
1315  * Move a list of pages in the address space of the currently executing
1316  * process.
1317  */
1318 SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages,
1319 		const void __user * __user *, pages,
1320 		const int __user *, nodes,
1321 		int __user *, status, int, flags)
1322 {
1323 	const struct cred *cred = current_cred(), *tcred;
1324 	struct task_struct *task;
1325 	struct mm_struct *mm;
1326 	int err;
1327 	nodemask_t task_nodes;
1328 
1329 	/* Check flags */
1330 	if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
1331 		return -EINVAL;
1332 
1333 	if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1334 		return -EPERM;
1335 
1336 	/* Find the mm_struct */
1337 	rcu_read_lock();
1338 	task = pid ? find_task_by_vpid(pid) : current;
1339 	if (!task) {
1340 		rcu_read_unlock();
1341 		return -ESRCH;
1342 	}
1343 	get_task_struct(task);
1344 
1345 	/*
1346 	 * Check if this process has the right to modify the specified
1347 	 * process. The right exists if the process has administrative
1348 	 * capabilities, superuser privileges or the same
1349 	 * userid as the target process.
1350 	 */
1351 	tcred = __task_cred(task);
1352 	if (!uid_eq(cred->euid, tcred->suid) && !uid_eq(cred->euid, tcred->uid) &&
1353 	    !uid_eq(cred->uid,  tcred->suid) && !uid_eq(cred->uid,  tcred->uid) &&
1354 	    !capable(CAP_SYS_NICE)) {
1355 		rcu_read_unlock();
1356 		err = -EPERM;
1357 		goto out;
1358 	}
1359 	rcu_read_unlock();
1360 
1361  	err = security_task_movememory(task);
1362  	if (err)
1363 		goto out;
1364 
1365 	task_nodes = cpuset_mems_allowed(task);
1366 	mm = get_task_mm(task);
1367 	put_task_struct(task);
1368 
1369 	if (!mm)
1370 		return -EINVAL;
1371 
1372 	if (nodes)
1373 		err = do_pages_move(mm, task_nodes, nr_pages, pages,
1374 				    nodes, status, flags);
1375 	else
1376 		err = do_pages_stat(mm, nr_pages, pages, status);
1377 
1378 	mmput(mm);
1379 	return err;
1380 
1381 out:
1382 	put_task_struct(task);
1383 	return err;
1384 }
1385 
1386 /*
1387  * Call migration functions in the vma_ops that may prepare
1388  * memory in a vm for migration. migration functions may perform
1389  * the migration for vmas that do not have an underlying page struct.
1390  */
1391 int migrate_vmas(struct mm_struct *mm, const nodemask_t *to,
1392 	const nodemask_t *from, unsigned long flags)
1393 {
1394  	struct vm_area_struct *vma;
1395  	int err = 0;
1396 
1397 	for (vma = mm->mmap; vma && !err; vma = vma->vm_next) {
1398  		if (vma->vm_ops && vma->vm_ops->migrate) {
1399  			err = vma->vm_ops->migrate(vma, to, from, flags);
1400  			if (err)
1401  				break;
1402  		}
1403  	}
1404  	return err;
1405 }
1406 #endif
1407