xref: /openbmc/linux/mm/migrate.c (revision abfbd895)
1 /*
2  * Memory Migration functionality - linux/mm/migrate.c
3  *
4  * Copyright (C) 2006 Silicon Graphics, Inc., Christoph Lameter
5  *
6  * Page migration was first developed in the context of the memory hotplug
7  * project. The main authors of the migration code are:
8  *
9  * IWAMOTO Toshihiro <iwamoto@valinux.co.jp>
10  * Hirokazu Takahashi <taka@valinux.co.jp>
11  * Dave Hansen <haveblue@us.ibm.com>
12  * Christoph Lameter
13  */
14 
15 #include <linux/migrate.h>
16 #include <linux/export.h>
17 #include <linux/swap.h>
18 #include <linux/swapops.h>
19 #include <linux/pagemap.h>
20 #include <linux/buffer_head.h>
21 #include <linux/mm_inline.h>
22 #include <linux/nsproxy.h>
23 #include <linux/pagevec.h>
24 #include <linux/ksm.h>
25 #include <linux/rmap.h>
26 #include <linux/topology.h>
27 #include <linux/cpu.h>
28 #include <linux/cpuset.h>
29 #include <linux/writeback.h>
30 #include <linux/mempolicy.h>
31 #include <linux/vmalloc.h>
32 #include <linux/security.h>
33 #include <linux/backing-dev.h>
34 #include <linux/syscalls.h>
35 #include <linux/hugetlb.h>
36 #include <linux/hugetlb_cgroup.h>
37 #include <linux/gfp.h>
38 #include <linux/balloon_compaction.h>
39 #include <linux/mmu_notifier.h>
40 #include <linux/page_idle.h>
41 
42 #include <asm/tlbflush.h>
43 
44 #define CREATE_TRACE_POINTS
45 #include <trace/events/migrate.h>
46 
47 #include "internal.h"
48 
49 /*
50  * migrate_prep() needs to be called before we start compiling a list of pages
51  * to be migrated using isolate_lru_page(). If scheduling work on other CPUs is
52  * undesirable, use migrate_prep_local()
53  */
54 int migrate_prep(void)
55 {
56 	/*
57 	 * Clear the LRU lists so pages can be isolated.
58 	 * Note that pages may be moved off the LRU after we have
59 	 * drained them. Those pages will fail to migrate like other
60 	 * pages that may be busy.
61 	 */
62 	lru_add_drain_all();
63 
64 	return 0;
65 }
66 
67 /* Do the necessary work of migrate_prep but not if it involves other CPUs */
68 int migrate_prep_local(void)
69 {
70 	lru_add_drain();
71 
72 	return 0;
73 }
74 
75 /*
76  * Put previously isolated pages back onto the appropriate lists
77  * from where they were once taken off for compaction/migration.
78  *
79  * This function shall be used whenever the isolated pageset has been
80  * built from lru, balloon, hugetlbfs page. See isolate_migratepages_range()
81  * and isolate_huge_page().
82  */
83 void putback_movable_pages(struct list_head *l)
84 {
85 	struct page *page;
86 	struct page *page2;
87 
88 	list_for_each_entry_safe(page, page2, l, lru) {
89 		if (unlikely(PageHuge(page))) {
90 			putback_active_hugepage(page);
91 			continue;
92 		}
93 		list_del(&page->lru);
94 		dec_zone_page_state(page, NR_ISOLATED_ANON +
95 				page_is_file_cache(page));
96 		if (unlikely(isolated_balloon_page(page)))
97 			balloon_page_putback(page);
98 		else
99 			putback_lru_page(page);
100 	}
101 }
102 
103 /*
104  * Restore a potential migration pte to a working pte entry
105  */
106 static int remove_migration_pte(struct page *new, struct vm_area_struct *vma,
107 				 unsigned long addr, void *old)
108 {
109 	struct mm_struct *mm = vma->vm_mm;
110 	swp_entry_t entry;
111  	pmd_t *pmd;
112 	pte_t *ptep, pte;
113  	spinlock_t *ptl;
114 
115 	if (unlikely(PageHuge(new))) {
116 		ptep = huge_pte_offset(mm, addr);
117 		if (!ptep)
118 			goto out;
119 		ptl = huge_pte_lockptr(hstate_vma(vma), mm, ptep);
120 	} else {
121 		pmd = mm_find_pmd(mm, addr);
122 		if (!pmd)
123 			goto out;
124 
125 		ptep = pte_offset_map(pmd, addr);
126 
127 		/*
128 		 * Peek to check is_swap_pte() before taking ptlock?  No, we
129 		 * can race mremap's move_ptes(), which skips anon_vma lock.
130 		 */
131 
132 		ptl = pte_lockptr(mm, pmd);
133 	}
134 
135  	spin_lock(ptl);
136 	pte = *ptep;
137 	if (!is_swap_pte(pte))
138 		goto unlock;
139 
140 	entry = pte_to_swp_entry(pte);
141 
142 	if (!is_migration_entry(entry) ||
143 	    migration_entry_to_page(entry) != old)
144 		goto unlock;
145 
146 	get_page(new);
147 	pte = pte_mkold(mk_pte(new, vma->vm_page_prot));
148 	if (pte_swp_soft_dirty(*ptep))
149 		pte = pte_mksoft_dirty(pte);
150 
151 	/* Recheck VMA as permissions can change since migration started  */
152 	if (is_write_migration_entry(entry))
153 		pte = maybe_mkwrite(pte, vma);
154 
155 #ifdef CONFIG_HUGETLB_PAGE
156 	if (PageHuge(new)) {
157 		pte = pte_mkhuge(pte);
158 		pte = arch_make_huge_pte(pte, vma, new, 0);
159 	}
160 #endif
161 	flush_dcache_page(new);
162 	set_pte_at(mm, addr, ptep, pte);
163 
164 	if (PageHuge(new)) {
165 		if (PageAnon(new))
166 			hugepage_add_anon_rmap(new, vma, addr);
167 		else
168 			page_dup_rmap(new);
169 	} else if (PageAnon(new))
170 		page_add_anon_rmap(new, vma, addr);
171 	else
172 		page_add_file_rmap(new);
173 
174 	if (vma->vm_flags & VM_LOCKED)
175 		mlock_vma_page(new);
176 
177 	/* No need to invalidate - it was non-present before */
178 	update_mmu_cache(vma, addr, ptep);
179 unlock:
180 	pte_unmap_unlock(ptep, ptl);
181 out:
182 	return SWAP_AGAIN;
183 }
184 
185 /*
186  * Get rid of all migration entries and replace them by
187  * references to the indicated page.
188  */
189 static void remove_migration_ptes(struct page *old, struct page *new)
190 {
191 	struct rmap_walk_control rwc = {
192 		.rmap_one = remove_migration_pte,
193 		.arg = old,
194 	};
195 
196 	rmap_walk(new, &rwc);
197 }
198 
199 /*
200  * Something used the pte of a page under migration. We need to
201  * get to the page and wait until migration is finished.
202  * When we return from this function the fault will be retried.
203  */
204 void __migration_entry_wait(struct mm_struct *mm, pte_t *ptep,
205 				spinlock_t *ptl)
206 {
207 	pte_t pte;
208 	swp_entry_t entry;
209 	struct page *page;
210 
211 	spin_lock(ptl);
212 	pte = *ptep;
213 	if (!is_swap_pte(pte))
214 		goto out;
215 
216 	entry = pte_to_swp_entry(pte);
217 	if (!is_migration_entry(entry))
218 		goto out;
219 
220 	page = migration_entry_to_page(entry);
221 
222 	/*
223 	 * Once radix-tree replacement of page migration started, page_count
224 	 * *must* be zero. And, we don't want to call wait_on_page_locked()
225 	 * against a page without get_page().
226 	 * So, we use get_page_unless_zero(), here. Even failed, page fault
227 	 * will occur again.
228 	 */
229 	if (!get_page_unless_zero(page))
230 		goto out;
231 	pte_unmap_unlock(ptep, ptl);
232 	wait_on_page_locked(page);
233 	put_page(page);
234 	return;
235 out:
236 	pte_unmap_unlock(ptep, ptl);
237 }
238 
239 void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
240 				unsigned long address)
241 {
242 	spinlock_t *ptl = pte_lockptr(mm, pmd);
243 	pte_t *ptep = pte_offset_map(pmd, address);
244 	__migration_entry_wait(mm, ptep, ptl);
245 }
246 
247 void migration_entry_wait_huge(struct vm_area_struct *vma,
248 		struct mm_struct *mm, pte_t *pte)
249 {
250 	spinlock_t *ptl = huge_pte_lockptr(hstate_vma(vma), mm, pte);
251 	__migration_entry_wait(mm, pte, ptl);
252 }
253 
254 #ifdef CONFIG_BLOCK
255 /* Returns true if all buffers are successfully locked */
256 static bool buffer_migrate_lock_buffers(struct buffer_head *head,
257 							enum migrate_mode mode)
258 {
259 	struct buffer_head *bh = head;
260 
261 	/* Simple case, sync compaction */
262 	if (mode != MIGRATE_ASYNC) {
263 		do {
264 			get_bh(bh);
265 			lock_buffer(bh);
266 			bh = bh->b_this_page;
267 
268 		} while (bh != head);
269 
270 		return true;
271 	}
272 
273 	/* async case, we cannot block on lock_buffer so use trylock_buffer */
274 	do {
275 		get_bh(bh);
276 		if (!trylock_buffer(bh)) {
277 			/*
278 			 * We failed to lock the buffer and cannot stall in
279 			 * async migration. Release the taken locks
280 			 */
281 			struct buffer_head *failed_bh = bh;
282 			put_bh(failed_bh);
283 			bh = head;
284 			while (bh != failed_bh) {
285 				unlock_buffer(bh);
286 				put_bh(bh);
287 				bh = bh->b_this_page;
288 			}
289 			return false;
290 		}
291 
292 		bh = bh->b_this_page;
293 	} while (bh != head);
294 	return true;
295 }
296 #else
297 static inline bool buffer_migrate_lock_buffers(struct buffer_head *head,
298 							enum migrate_mode mode)
299 {
300 	return true;
301 }
302 #endif /* CONFIG_BLOCK */
303 
304 /*
305  * Replace the page in the mapping.
306  *
307  * The number of remaining references must be:
308  * 1 for anonymous pages without a mapping
309  * 2 for pages with a mapping
310  * 3 for pages with a mapping and PagePrivate/PagePrivate2 set.
311  */
312 int migrate_page_move_mapping(struct address_space *mapping,
313 		struct page *newpage, struct page *page,
314 		struct buffer_head *head, enum migrate_mode mode,
315 		int extra_count)
316 {
317 	struct zone *oldzone, *newzone;
318 	int dirty;
319 	int expected_count = 1 + extra_count;
320 	void **pslot;
321 
322 	if (!mapping) {
323 		/* Anonymous page without mapping */
324 		if (page_count(page) != expected_count)
325 			return -EAGAIN;
326 
327 		/* No turning back from here */
328 		set_page_memcg(newpage, page_memcg(page));
329 		newpage->index = page->index;
330 		newpage->mapping = page->mapping;
331 		if (PageSwapBacked(page))
332 			SetPageSwapBacked(newpage);
333 
334 		return MIGRATEPAGE_SUCCESS;
335 	}
336 
337 	oldzone = page_zone(page);
338 	newzone = page_zone(newpage);
339 
340 	spin_lock_irq(&mapping->tree_lock);
341 
342 	pslot = radix_tree_lookup_slot(&mapping->page_tree,
343  					page_index(page));
344 
345 	expected_count += 1 + page_has_private(page);
346 	if (page_count(page) != expected_count ||
347 		radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) {
348 		spin_unlock_irq(&mapping->tree_lock);
349 		return -EAGAIN;
350 	}
351 
352 	if (!page_freeze_refs(page, expected_count)) {
353 		spin_unlock_irq(&mapping->tree_lock);
354 		return -EAGAIN;
355 	}
356 
357 	/*
358 	 * In the async migration case of moving a page with buffers, lock the
359 	 * buffers using trylock before the mapping is moved. If the mapping
360 	 * was moved, we later failed to lock the buffers and could not move
361 	 * the mapping back due to an elevated page count, we would have to
362 	 * block waiting on other references to be dropped.
363 	 */
364 	if (mode == MIGRATE_ASYNC && head &&
365 			!buffer_migrate_lock_buffers(head, mode)) {
366 		page_unfreeze_refs(page, expected_count);
367 		spin_unlock_irq(&mapping->tree_lock);
368 		return -EAGAIN;
369 	}
370 
371 	/*
372 	 * Now we know that no one else is looking at the page:
373 	 * no turning back from here.
374 	 */
375 	set_page_memcg(newpage, page_memcg(page));
376 	newpage->index = page->index;
377 	newpage->mapping = page->mapping;
378 	if (PageSwapBacked(page))
379 		SetPageSwapBacked(newpage);
380 
381 	get_page(newpage);	/* add cache reference */
382 	if (PageSwapCache(page)) {
383 		SetPageSwapCache(newpage);
384 		set_page_private(newpage, page_private(page));
385 	}
386 
387 	/* Move dirty while page refs frozen and newpage not yet exposed */
388 	dirty = PageDirty(page);
389 	if (dirty) {
390 		ClearPageDirty(page);
391 		SetPageDirty(newpage);
392 	}
393 
394 	radix_tree_replace_slot(pslot, newpage);
395 
396 	/*
397 	 * Drop cache reference from old page by unfreezing
398 	 * to one less reference.
399 	 * We know this isn't the last reference.
400 	 */
401 	page_unfreeze_refs(page, expected_count - 1);
402 
403 	spin_unlock(&mapping->tree_lock);
404 	/* Leave irq disabled to prevent preemption while updating stats */
405 
406 	/*
407 	 * If moved to a different zone then also account
408 	 * the page for that zone. Other VM counters will be
409 	 * taken care of when we establish references to the
410 	 * new page and drop references to the old page.
411 	 *
412 	 * Note that anonymous pages are accounted for
413 	 * via NR_FILE_PAGES and NR_ANON_PAGES if they
414 	 * are mapped to swap space.
415 	 */
416 	if (newzone != oldzone) {
417 		__dec_zone_state(oldzone, NR_FILE_PAGES);
418 		__inc_zone_state(newzone, NR_FILE_PAGES);
419 		if (PageSwapBacked(page) && !PageSwapCache(page)) {
420 			__dec_zone_state(oldzone, NR_SHMEM);
421 			__inc_zone_state(newzone, NR_SHMEM);
422 		}
423 		if (dirty && mapping_cap_account_dirty(mapping)) {
424 			__dec_zone_state(oldzone, NR_FILE_DIRTY);
425 			__inc_zone_state(newzone, NR_FILE_DIRTY);
426 		}
427 	}
428 	local_irq_enable();
429 
430 	return MIGRATEPAGE_SUCCESS;
431 }
432 
433 /*
434  * The expected number of remaining references is the same as that
435  * of migrate_page_move_mapping().
436  */
437 int migrate_huge_page_move_mapping(struct address_space *mapping,
438 				   struct page *newpage, struct page *page)
439 {
440 	int expected_count;
441 	void **pslot;
442 
443 	spin_lock_irq(&mapping->tree_lock);
444 
445 	pslot = radix_tree_lookup_slot(&mapping->page_tree,
446 					page_index(page));
447 
448 	expected_count = 2 + page_has_private(page);
449 	if (page_count(page) != expected_count ||
450 		radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) {
451 		spin_unlock_irq(&mapping->tree_lock);
452 		return -EAGAIN;
453 	}
454 
455 	if (!page_freeze_refs(page, expected_count)) {
456 		spin_unlock_irq(&mapping->tree_lock);
457 		return -EAGAIN;
458 	}
459 
460 	set_page_memcg(newpage, page_memcg(page));
461 	newpage->index = page->index;
462 	newpage->mapping = page->mapping;
463 	get_page(newpage);
464 
465 	radix_tree_replace_slot(pslot, newpage);
466 
467 	page_unfreeze_refs(page, expected_count - 1);
468 
469 	spin_unlock_irq(&mapping->tree_lock);
470 	return MIGRATEPAGE_SUCCESS;
471 }
472 
473 /*
474  * Gigantic pages are so large that we do not guarantee that page++ pointer
475  * arithmetic will work across the entire page.  We need something more
476  * specialized.
477  */
478 static void __copy_gigantic_page(struct page *dst, struct page *src,
479 				int nr_pages)
480 {
481 	int i;
482 	struct page *dst_base = dst;
483 	struct page *src_base = src;
484 
485 	for (i = 0; i < nr_pages; ) {
486 		cond_resched();
487 		copy_highpage(dst, src);
488 
489 		i++;
490 		dst = mem_map_next(dst, dst_base, i);
491 		src = mem_map_next(src, src_base, i);
492 	}
493 }
494 
495 static void copy_huge_page(struct page *dst, struct page *src)
496 {
497 	int i;
498 	int nr_pages;
499 
500 	if (PageHuge(src)) {
501 		/* hugetlbfs page */
502 		struct hstate *h = page_hstate(src);
503 		nr_pages = pages_per_huge_page(h);
504 
505 		if (unlikely(nr_pages > MAX_ORDER_NR_PAGES)) {
506 			__copy_gigantic_page(dst, src, nr_pages);
507 			return;
508 		}
509 	} else {
510 		/* thp page */
511 		BUG_ON(!PageTransHuge(src));
512 		nr_pages = hpage_nr_pages(src);
513 	}
514 
515 	for (i = 0; i < nr_pages; i++) {
516 		cond_resched();
517 		copy_highpage(dst + i, src + i);
518 	}
519 }
520 
521 /*
522  * Copy the page to its new location
523  */
524 void migrate_page_copy(struct page *newpage, struct page *page)
525 {
526 	int cpupid;
527 
528 	if (PageHuge(page) || PageTransHuge(page))
529 		copy_huge_page(newpage, page);
530 	else
531 		copy_highpage(newpage, page);
532 
533 	if (PageError(page))
534 		SetPageError(newpage);
535 	if (PageReferenced(page))
536 		SetPageReferenced(newpage);
537 	if (PageUptodate(page))
538 		SetPageUptodate(newpage);
539 	if (TestClearPageActive(page)) {
540 		VM_BUG_ON_PAGE(PageUnevictable(page), page);
541 		SetPageActive(newpage);
542 	} else if (TestClearPageUnevictable(page))
543 		SetPageUnevictable(newpage);
544 	if (PageChecked(page))
545 		SetPageChecked(newpage);
546 	if (PageMappedToDisk(page))
547 		SetPageMappedToDisk(newpage);
548 
549 	/* Move dirty on pages not done by migrate_page_move_mapping() */
550 	if (PageDirty(page))
551 		SetPageDirty(newpage);
552 
553 	if (page_is_young(page))
554 		set_page_young(newpage);
555 	if (page_is_idle(page))
556 		set_page_idle(newpage);
557 
558 	/*
559 	 * Copy NUMA information to the new page, to prevent over-eager
560 	 * future migrations of this same page.
561 	 */
562 	cpupid = page_cpupid_xchg_last(page, -1);
563 	page_cpupid_xchg_last(newpage, cpupid);
564 
565 	ksm_migrate_page(newpage, page);
566 	/*
567 	 * Please do not reorder this without considering how mm/ksm.c's
568 	 * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache().
569 	 */
570 	if (PageSwapCache(page))
571 		ClearPageSwapCache(page);
572 	ClearPagePrivate(page);
573 	set_page_private(page, 0);
574 
575 	/*
576 	 * If any waiters have accumulated on the new page then
577 	 * wake them up.
578 	 */
579 	if (PageWriteback(newpage))
580 		end_page_writeback(newpage);
581 }
582 
583 /************************************************************
584  *                    Migration functions
585  ***********************************************************/
586 
587 /*
588  * Common logic to directly migrate a single page suitable for
589  * pages that do not use PagePrivate/PagePrivate2.
590  *
591  * Pages are locked upon entry and exit.
592  */
593 int migrate_page(struct address_space *mapping,
594 		struct page *newpage, struct page *page,
595 		enum migrate_mode mode)
596 {
597 	int rc;
598 
599 	BUG_ON(PageWriteback(page));	/* Writeback must be complete */
600 
601 	rc = migrate_page_move_mapping(mapping, newpage, page, NULL, mode, 0);
602 
603 	if (rc != MIGRATEPAGE_SUCCESS)
604 		return rc;
605 
606 	migrate_page_copy(newpage, page);
607 	return MIGRATEPAGE_SUCCESS;
608 }
609 EXPORT_SYMBOL(migrate_page);
610 
611 #ifdef CONFIG_BLOCK
612 /*
613  * Migration function for pages with buffers. This function can only be used
614  * if the underlying filesystem guarantees that no other references to "page"
615  * exist.
616  */
617 int buffer_migrate_page(struct address_space *mapping,
618 		struct page *newpage, struct page *page, enum migrate_mode mode)
619 {
620 	struct buffer_head *bh, *head;
621 	int rc;
622 
623 	if (!page_has_buffers(page))
624 		return migrate_page(mapping, newpage, page, mode);
625 
626 	head = page_buffers(page);
627 
628 	rc = migrate_page_move_mapping(mapping, newpage, page, head, mode, 0);
629 
630 	if (rc != MIGRATEPAGE_SUCCESS)
631 		return rc;
632 
633 	/*
634 	 * In the async case, migrate_page_move_mapping locked the buffers
635 	 * with an IRQ-safe spinlock held. In the sync case, the buffers
636 	 * need to be locked now
637 	 */
638 	if (mode != MIGRATE_ASYNC)
639 		BUG_ON(!buffer_migrate_lock_buffers(head, mode));
640 
641 	ClearPagePrivate(page);
642 	set_page_private(newpage, page_private(page));
643 	set_page_private(page, 0);
644 	put_page(page);
645 	get_page(newpage);
646 
647 	bh = head;
648 	do {
649 		set_bh_page(bh, newpage, bh_offset(bh));
650 		bh = bh->b_this_page;
651 
652 	} while (bh != head);
653 
654 	SetPagePrivate(newpage);
655 
656 	migrate_page_copy(newpage, page);
657 
658 	bh = head;
659 	do {
660 		unlock_buffer(bh);
661  		put_bh(bh);
662 		bh = bh->b_this_page;
663 
664 	} while (bh != head);
665 
666 	return MIGRATEPAGE_SUCCESS;
667 }
668 EXPORT_SYMBOL(buffer_migrate_page);
669 #endif
670 
671 /*
672  * Writeback a page to clean the dirty state
673  */
674 static int writeout(struct address_space *mapping, struct page *page)
675 {
676 	struct writeback_control wbc = {
677 		.sync_mode = WB_SYNC_NONE,
678 		.nr_to_write = 1,
679 		.range_start = 0,
680 		.range_end = LLONG_MAX,
681 		.for_reclaim = 1
682 	};
683 	int rc;
684 
685 	if (!mapping->a_ops->writepage)
686 		/* No write method for the address space */
687 		return -EINVAL;
688 
689 	if (!clear_page_dirty_for_io(page))
690 		/* Someone else already triggered a write */
691 		return -EAGAIN;
692 
693 	/*
694 	 * A dirty page may imply that the underlying filesystem has
695 	 * the page on some queue. So the page must be clean for
696 	 * migration. Writeout may mean we loose the lock and the
697 	 * page state is no longer what we checked for earlier.
698 	 * At this point we know that the migration attempt cannot
699 	 * be successful.
700 	 */
701 	remove_migration_ptes(page, page);
702 
703 	rc = mapping->a_ops->writepage(page, &wbc);
704 
705 	if (rc != AOP_WRITEPAGE_ACTIVATE)
706 		/* unlocked. Relock */
707 		lock_page(page);
708 
709 	return (rc < 0) ? -EIO : -EAGAIN;
710 }
711 
712 /*
713  * Default handling if a filesystem does not provide a migration function.
714  */
715 static int fallback_migrate_page(struct address_space *mapping,
716 	struct page *newpage, struct page *page, enum migrate_mode mode)
717 {
718 	if (PageDirty(page)) {
719 		/* Only writeback pages in full synchronous migration */
720 		if (mode != MIGRATE_SYNC)
721 			return -EBUSY;
722 		return writeout(mapping, page);
723 	}
724 
725 	/*
726 	 * Buffers may be managed in a filesystem specific way.
727 	 * We must have no buffers or drop them.
728 	 */
729 	if (page_has_private(page) &&
730 	    !try_to_release_page(page, GFP_KERNEL))
731 		return -EAGAIN;
732 
733 	return migrate_page(mapping, newpage, page, mode);
734 }
735 
736 /*
737  * Move a page to a newly allocated page
738  * The page is locked and all ptes have been successfully removed.
739  *
740  * The new page will have replaced the old page if this function
741  * is successful.
742  *
743  * Return value:
744  *   < 0 - error code
745  *  MIGRATEPAGE_SUCCESS - success
746  */
747 static int move_to_new_page(struct page *newpage, struct page *page,
748 				enum migrate_mode mode)
749 {
750 	struct address_space *mapping;
751 	int rc;
752 
753 	VM_BUG_ON_PAGE(!PageLocked(page), page);
754 	VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
755 
756 	mapping = page_mapping(page);
757 	if (!mapping)
758 		rc = migrate_page(mapping, newpage, page, mode);
759 	else if (mapping->a_ops->migratepage)
760 		/*
761 		 * Most pages have a mapping and most filesystems provide a
762 		 * migratepage callback. Anonymous pages are part of swap
763 		 * space which also has its own migratepage callback. This
764 		 * is the most common path for page migration.
765 		 */
766 		rc = mapping->a_ops->migratepage(mapping, newpage, page, mode);
767 	else
768 		rc = fallback_migrate_page(mapping, newpage, page, mode);
769 
770 	/*
771 	 * When successful, old pagecache page->mapping must be cleared before
772 	 * page is freed; but stats require that PageAnon be left as PageAnon.
773 	 */
774 	if (rc == MIGRATEPAGE_SUCCESS) {
775 		set_page_memcg(page, NULL);
776 		if (!PageAnon(page))
777 			page->mapping = NULL;
778 	}
779 	return rc;
780 }
781 
782 static int __unmap_and_move(struct page *page, struct page *newpage,
783 				int force, enum migrate_mode mode)
784 {
785 	int rc = -EAGAIN;
786 	int page_was_mapped = 0;
787 	struct anon_vma *anon_vma = NULL;
788 
789 	if (!trylock_page(page)) {
790 		if (!force || mode == MIGRATE_ASYNC)
791 			goto out;
792 
793 		/*
794 		 * It's not safe for direct compaction to call lock_page.
795 		 * For example, during page readahead pages are added locked
796 		 * to the LRU. Later, when the IO completes the pages are
797 		 * marked uptodate and unlocked. However, the queueing
798 		 * could be merging multiple pages for one bio (e.g.
799 		 * mpage_readpages). If an allocation happens for the
800 		 * second or third page, the process can end up locking
801 		 * the same page twice and deadlocking. Rather than
802 		 * trying to be clever about what pages can be locked,
803 		 * avoid the use of lock_page for direct compaction
804 		 * altogether.
805 		 */
806 		if (current->flags & PF_MEMALLOC)
807 			goto out;
808 
809 		lock_page(page);
810 	}
811 
812 	if (PageWriteback(page)) {
813 		/*
814 		 * Only in the case of a full synchronous migration is it
815 		 * necessary to wait for PageWriteback. In the async case,
816 		 * the retry loop is too short and in the sync-light case,
817 		 * the overhead of stalling is too much
818 		 */
819 		if (mode != MIGRATE_SYNC) {
820 			rc = -EBUSY;
821 			goto out_unlock;
822 		}
823 		if (!force)
824 			goto out_unlock;
825 		wait_on_page_writeback(page);
826 	}
827 
828 	/*
829 	 * By try_to_unmap(), page->mapcount goes down to 0 here. In this case,
830 	 * we cannot notice that anon_vma is freed while we migrates a page.
831 	 * This get_anon_vma() delays freeing anon_vma pointer until the end
832 	 * of migration. File cache pages are no problem because of page_lock()
833 	 * File Caches may use write_page() or lock_page() in migration, then,
834 	 * just care Anon page here.
835 	 *
836 	 * Only page_get_anon_vma() understands the subtleties of
837 	 * getting a hold on an anon_vma from outside one of its mms.
838 	 * But if we cannot get anon_vma, then we won't need it anyway,
839 	 * because that implies that the anon page is no longer mapped
840 	 * (and cannot be remapped so long as we hold the page lock).
841 	 */
842 	if (PageAnon(page) && !PageKsm(page))
843 		anon_vma = page_get_anon_vma(page);
844 
845 	/*
846 	 * Block others from accessing the new page when we get around to
847 	 * establishing additional references. We are usually the only one
848 	 * holding a reference to newpage at this point. We used to have a BUG
849 	 * here if trylock_page(newpage) fails, but would like to allow for
850 	 * cases where there might be a race with the previous use of newpage.
851 	 * This is much like races on refcount of oldpage: just don't BUG().
852 	 */
853 	if (unlikely(!trylock_page(newpage)))
854 		goto out_unlock;
855 
856 	if (unlikely(isolated_balloon_page(page))) {
857 		/*
858 		 * A ballooned page does not need any special attention from
859 		 * physical to virtual reverse mapping procedures.
860 		 * Skip any attempt to unmap PTEs or to remap swap cache,
861 		 * in order to avoid burning cycles at rmap level, and perform
862 		 * the page migration right away (proteced by page lock).
863 		 */
864 		rc = balloon_page_migrate(newpage, page, mode);
865 		goto out_unlock_both;
866 	}
867 
868 	/*
869 	 * Corner case handling:
870 	 * 1. When a new swap-cache page is read into, it is added to the LRU
871 	 * and treated as swapcache but it has no rmap yet.
872 	 * Calling try_to_unmap() against a page->mapping==NULL page will
873 	 * trigger a BUG.  So handle it here.
874 	 * 2. An orphaned page (see truncate_complete_page) might have
875 	 * fs-private metadata. The page can be picked up due to memory
876 	 * offlining.  Everywhere else except page reclaim, the page is
877 	 * invisible to the vm, so the page can not be migrated.  So try to
878 	 * free the metadata, so the page can be freed.
879 	 */
880 	if (!page->mapping) {
881 		VM_BUG_ON_PAGE(PageAnon(page), page);
882 		if (page_has_private(page)) {
883 			try_to_free_buffers(page);
884 			goto out_unlock_both;
885 		}
886 	} else if (page_mapped(page)) {
887 		/* Establish migration ptes */
888 		VM_BUG_ON_PAGE(PageAnon(page) && !PageKsm(page) && !anon_vma,
889 				page);
890 		try_to_unmap(page,
891 			TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
892 		page_was_mapped = 1;
893 	}
894 
895 	if (!page_mapped(page))
896 		rc = move_to_new_page(newpage, page, mode);
897 
898 	if (page_was_mapped)
899 		remove_migration_ptes(page,
900 			rc == MIGRATEPAGE_SUCCESS ? newpage : page);
901 
902 out_unlock_both:
903 	unlock_page(newpage);
904 out_unlock:
905 	/* Drop an anon_vma reference if we took one */
906 	if (anon_vma)
907 		put_anon_vma(anon_vma);
908 	unlock_page(page);
909 out:
910 	return rc;
911 }
912 
913 /*
914  * gcc 4.7 and 4.8 on arm get an ICEs when inlining unmap_and_move().  Work
915  * around it.
916  */
917 #if (GCC_VERSION >= 40700 && GCC_VERSION < 40900) && defined(CONFIG_ARM)
918 #define ICE_noinline noinline
919 #else
920 #define ICE_noinline
921 #endif
922 
923 /*
924  * Obtain the lock on page, remove all ptes and migrate the page
925  * to the newly allocated page in newpage.
926  */
927 static ICE_noinline int unmap_and_move(new_page_t get_new_page,
928 				   free_page_t put_new_page,
929 				   unsigned long private, struct page *page,
930 				   int force, enum migrate_mode mode,
931 				   enum migrate_reason reason)
932 {
933 	int rc = MIGRATEPAGE_SUCCESS;
934 	int *result = NULL;
935 	struct page *newpage;
936 
937 	newpage = get_new_page(page, private, &result);
938 	if (!newpage)
939 		return -ENOMEM;
940 
941 	if (page_count(page) == 1) {
942 		/* page was freed from under us. So we are done. */
943 		goto out;
944 	}
945 
946 	if (unlikely(PageTransHuge(page)))
947 		if (unlikely(split_huge_page(page)))
948 			goto out;
949 
950 	rc = __unmap_and_move(page, newpage, force, mode);
951 	if (rc == MIGRATEPAGE_SUCCESS)
952 		put_new_page = NULL;
953 
954 out:
955 	if (rc != -EAGAIN) {
956 		/*
957 		 * A page that has been migrated has all references
958 		 * removed and will be freed. A page that has not been
959 		 * migrated will have kepts its references and be
960 		 * restored.
961 		 */
962 		list_del(&page->lru);
963 		dec_zone_page_state(page, NR_ISOLATED_ANON +
964 				page_is_file_cache(page));
965 		/* Soft-offlined page shouldn't go through lru cache list */
966 		if (reason == MR_MEMORY_FAILURE) {
967 			put_page(page);
968 			if (!test_set_page_hwpoison(page))
969 				num_poisoned_pages_inc();
970 		} else
971 			putback_lru_page(page);
972 	}
973 
974 	/*
975 	 * If migration was not successful and there's a freeing callback, use
976 	 * it.  Otherwise, putback_lru_page() will drop the reference grabbed
977 	 * during isolation.
978 	 */
979 	if (put_new_page)
980 		put_new_page(newpage, private);
981 	else if (unlikely(__is_movable_balloon_page(newpage))) {
982 		/* drop our reference, page already in the balloon */
983 		put_page(newpage);
984 	} else
985 		putback_lru_page(newpage);
986 
987 	if (result) {
988 		if (rc)
989 			*result = rc;
990 		else
991 			*result = page_to_nid(newpage);
992 	}
993 	return rc;
994 }
995 
996 /*
997  * Counterpart of unmap_and_move_page() for hugepage migration.
998  *
999  * This function doesn't wait the completion of hugepage I/O
1000  * because there is no race between I/O and migration for hugepage.
1001  * Note that currently hugepage I/O occurs only in direct I/O
1002  * where no lock is held and PG_writeback is irrelevant,
1003  * and writeback status of all subpages are counted in the reference
1004  * count of the head page (i.e. if all subpages of a 2MB hugepage are
1005  * under direct I/O, the reference of the head page is 512 and a bit more.)
1006  * This means that when we try to migrate hugepage whose subpages are
1007  * doing direct I/O, some references remain after try_to_unmap() and
1008  * hugepage migration fails without data corruption.
1009  *
1010  * There is also no race when direct I/O is issued on the page under migration,
1011  * because then pte is replaced with migration swap entry and direct I/O code
1012  * will wait in the page fault for migration to complete.
1013  */
1014 static int unmap_and_move_huge_page(new_page_t get_new_page,
1015 				free_page_t put_new_page, unsigned long private,
1016 				struct page *hpage, int force,
1017 				enum migrate_mode mode)
1018 {
1019 	int rc = -EAGAIN;
1020 	int *result = NULL;
1021 	int page_was_mapped = 0;
1022 	struct page *new_hpage;
1023 	struct anon_vma *anon_vma = NULL;
1024 
1025 	/*
1026 	 * Movability of hugepages depends on architectures and hugepage size.
1027 	 * This check is necessary because some callers of hugepage migration
1028 	 * like soft offline and memory hotremove don't walk through page
1029 	 * tables or check whether the hugepage is pmd-based or not before
1030 	 * kicking migration.
1031 	 */
1032 	if (!hugepage_migration_supported(page_hstate(hpage))) {
1033 		putback_active_hugepage(hpage);
1034 		return -ENOSYS;
1035 	}
1036 
1037 	new_hpage = get_new_page(hpage, private, &result);
1038 	if (!new_hpage)
1039 		return -ENOMEM;
1040 
1041 	if (!trylock_page(hpage)) {
1042 		if (!force || mode != MIGRATE_SYNC)
1043 			goto out;
1044 		lock_page(hpage);
1045 	}
1046 
1047 	if (PageAnon(hpage))
1048 		anon_vma = page_get_anon_vma(hpage);
1049 
1050 	if (unlikely(!trylock_page(new_hpage)))
1051 		goto put_anon;
1052 
1053 	if (page_mapped(hpage)) {
1054 		try_to_unmap(hpage,
1055 			TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
1056 		page_was_mapped = 1;
1057 	}
1058 
1059 	if (!page_mapped(hpage))
1060 		rc = move_to_new_page(new_hpage, hpage, mode);
1061 
1062 	if (page_was_mapped)
1063 		remove_migration_ptes(hpage,
1064 			rc == MIGRATEPAGE_SUCCESS ? new_hpage : hpage);
1065 
1066 	unlock_page(new_hpage);
1067 
1068 put_anon:
1069 	if (anon_vma)
1070 		put_anon_vma(anon_vma);
1071 
1072 	if (rc == MIGRATEPAGE_SUCCESS) {
1073 		hugetlb_cgroup_migrate(hpage, new_hpage);
1074 		put_new_page = NULL;
1075 	}
1076 
1077 	unlock_page(hpage);
1078 out:
1079 	if (rc != -EAGAIN)
1080 		putback_active_hugepage(hpage);
1081 
1082 	/*
1083 	 * If migration was not successful and there's a freeing callback, use
1084 	 * it.  Otherwise, put_page() will drop the reference grabbed during
1085 	 * isolation.
1086 	 */
1087 	if (put_new_page)
1088 		put_new_page(new_hpage, private);
1089 	else
1090 		putback_active_hugepage(new_hpage);
1091 
1092 	if (result) {
1093 		if (rc)
1094 			*result = rc;
1095 		else
1096 			*result = page_to_nid(new_hpage);
1097 	}
1098 	return rc;
1099 }
1100 
1101 /*
1102  * migrate_pages - migrate the pages specified in a list, to the free pages
1103  *		   supplied as the target for the page migration
1104  *
1105  * @from:		The list of pages to be migrated.
1106  * @get_new_page:	The function used to allocate free pages to be used
1107  *			as the target of the page migration.
1108  * @put_new_page:	The function used to free target pages if migration
1109  *			fails, or NULL if no special handling is necessary.
1110  * @private:		Private data to be passed on to get_new_page()
1111  * @mode:		The migration mode that specifies the constraints for
1112  *			page migration, if any.
1113  * @reason:		The reason for page migration.
1114  *
1115  * The function returns after 10 attempts or if no pages are movable any more
1116  * because the list has become empty or no retryable pages exist any more.
1117  * The caller should call putback_movable_pages() to return pages to the LRU
1118  * or free list only if ret != 0.
1119  *
1120  * Returns the number of pages that were not migrated, or an error code.
1121  */
1122 int migrate_pages(struct list_head *from, new_page_t get_new_page,
1123 		free_page_t put_new_page, unsigned long private,
1124 		enum migrate_mode mode, int reason)
1125 {
1126 	int retry = 1;
1127 	int nr_failed = 0;
1128 	int nr_succeeded = 0;
1129 	int pass = 0;
1130 	struct page *page;
1131 	struct page *page2;
1132 	int swapwrite = current->flags & PF_SWAPWRITE;
1133 	int rc;
1134 
1135 	if (!swapwrite)
1136 		current->flags |= PF_SWAPWRITE;
1137 
1138 	for(pass = 0; pass < 10 && retry; pass++) {
1139 		retry = 0;
1140 
1141 		list_for_each_entry_safe(page, page2, from, lru) {
1142 			cond_resched();
1143 
1144 			if (PageHuge(page))
1145 				rc = unmap_and_move_huge_page(get_new_page,
1146 						put_new_page, private, page,
1147 						pass > 2, mode);
1148 			else
1149 				rc = unmap_and_move(get_new_page, put_new_page,
1150 						private, page, pass > 2, mode,
1151 						reason);
1152 
1153 			switch(rc) {
1154 			case -ENOMEM:
1155 				goto out;
1156 			case -EAGAIN:
1157 				retry++;
1158 				break;
1159 			case MIGRATEPAGE_SUCCESS:
1160 				nr_succeeded++;
1161 				break;
1162 			default:
1163 				/*
1164 				 * Permanent failure (-EBUSY, -ENOSYS, etc.):
1165 				 * unlike -EAGAIN case, the failed page is
1166 				 * removed from migration page list and not
1167 				 * retried in the next outer loop.
1168 				 */
1169 				nr_failed++;
1170 				break;
1171 			}
1172 		}
1173 	}
1174 	nr_failed += retry;
1175 	rc = nr_failed;
1176 out:
1177 	if (nr_succeeded)
1178 		count_vm_events(PGMIGRATE_SUCCESS, nr_succeeded);
1179 	if (nr_failed)
1180 		count_vm_events(PGMIGRATE_FAIL, nr_failed);
1181 	trace_mm_migrate_pages(nr_succeeded, nr_failed, mode, reason);
1182 
1183 	if (!swapwrite)
1184 		current->flags &= ~PF_SWAPWRITE;
1185 
1186 	return rc;
1187 }
1188 
1189 #ifdef CONFIG_NUMA
1190 /*
1191  * Move a list of individual pages
1192  */
1193 struct page_to_node {
1194 	unsigned long addr;
1195 	struct page *page;
1196 	int node;
1197 	int status;
1198 };
1199 
1200 static struct page *new_page_node(struct page *p, unsigned long private,
1201 		int **result)
1202 {
1203 	struct page_to_node *pm = (struct page_to_node *)private;
1204 
1205 	while (pm->node != MAX_NUMNODES && pm->page != p)
1206 		pm++;
1207 
1208 	if (pm->node == MAX_NUMNODES)
1209 		return NULL;
1210 
1211 	*result = &pm->status;
1212 
1213 	if (PageHuge(p))
1214 		return alloc_huge_page_node(page_hstate(compound_head(p)),
1215 					pm->node);
1216 	else
1217 		return __alloc_pages_node(pm->node,
1218 				GFP_HIGHUSER_MOVABLE | __GFP_THISNODE, 0);
1219 }
1220 
1221 /*
1222  * Move a set of pages as indicated in the pm array. The addr
1223  * field must be set to the virtual address of the page to be moved
1224  * and the node number must contain a valid target node.
1225  * The pm array ends with node = MAX_NUMNODES.
1226  */
1227 static int do_move_page_to_node_array(struct mm_struct *mm,
1228 				      struct page_to_node *pm,
1229 				      int migrate_all)
1230 {
1231 	int err;
1232 	struct page_to_node *pp;
1233 	LIST_HEAD(pagelist);
1234 
1235 	down_read(&mm->mmap_sem);
1236 
1237 	/*
1238 	 * Build a list of pages to migrate
1239 	 */
1240 	for (pp = pm; pp->node != MAX_NUMNODES; pp++) {
1241 		struct vm_area_struct *vma;
1242 		struct page *page;
1243 
1244 		err = -EFAULT;
1245 		vma = find_vma(mm, pp->addr);
1246 		if (!vma || pp->addr < vma->vm_start || !vma_migratable(vma))
1247 			goto set_status;
1248 
1249 		/* FOLL_DUMP to ignore special (like zero) pages */
1250 		page = follow_page(vma, pp->addr,
1251 				FOLL_GET | FOLL_SPLIT | FOLL_DUMP);
1252 
1253 		err = PTR_ERR(page);
1254 		if (IS_ERR(page))
1255 			goto set_status;
1256 
1257 		err = -ENOENT;
1258 		if (!page)
1259 			goto set_status;
1260 
1261 		pp->page = page;
1262 		err = page_to_nid(page);
1263 
1264 		if (err == pp->node)
1265 			/*
1266 			 * Node already in the right place
1267 			 */
1268 			goto put_and_set;
1269 
1270 		err = -EACCES;
1271 		if (page_mapcount(page) > 1 &&
1272 				!migrate_all)
1273 			goto put_and_set;
1274 
1275 		if (PageHuge(page)) {
1276 			if (PageHead(page))
1277 				isolate_huge_page(page, &pagelist);
1278 			goto put_and_set;
1279 		}
1280 
1281 		err = isolate_lru_page(page);
1282 		if (!err) {
1283 			list_add_tail(&page->lru, &pagelist);
1284 			inc_zone_page_state(page, NR_ISOLATED_ANON +
1285 					    page_is_file_cache(page));
1286 		}
1287 put_and_set:
1288 		/*
1289 		 * Either remove the duplicate refcount from
1290 		 * isolate_lru_page() or drop the page ref if it was
1291 		 * not isolated.
1292 		 */
1293 		put_page(page);
1294 set_status:
1295 		pp->status = err;
1296 	}
1297 
1298 	err = 0;
1299 	if (!list_empty(&pagelist)) {
1300 		err = migrate_pages(&pagelist, new_page_node, NULL,
1301 				(unsigned long)pm, MIGRATE_SYNC, MR_SYSCALL);
1302 		if (err)
1303 			putback_movable_pages(&pagelist);
1304 	}
1305 
1306 	up_read(&mm->mmap_sem);
1307 	return err;
1308 }
1309 
1310 /*
1311  * Migrate an array of page address onto an array of nodes and fill
1312  * the corresponding array of status.
1313  */
1314 static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes,
1315 			 unsigned long nr_pages,
1316 			 const void __user * __user *pages,
1317 			 const int __user *nodes,
1318 			 int __user *status, int flags)
1319 {
1320 	struct page_to_node *pm;
1321 	unsigned long chunk_nr_pages;
1322 	unsigned long chunk_start;
1323 	int err;
1324 
1325 	err = -ENOMEM;
1326 	pm = (struct page_to_node *)__get_free_page(GFP_KERNEL);
1327 	if (!pm)
1328 		goto out;
1329 
1330 	migrate_prep();
1331 
1332 	/*
1333 	 * Store a chunk of page_to_node array in a page,
1334 	 * but keep the last one as a marker
1335 	 */
1336 	chunk_nr_pages = (PAGE_SIZE / sizeof(struct page_to_node)) - 1;
1337 
1338 	for (chunk_start = 0;
1339 	     chunk_start < nr_pages;
1340 	     chunk_start += chunk_nr_pages) {
1341 		int j;
1342 
1343 		if (chunk_start + chunk_nr_pages > nr_pages)
1344 			chunk_nr_pages = nr_pages - chunk_start;
1345 
1346 		/* fill the chunk pm with addrs and nodes from user-space */
1347 		for (j = 0; j < chunk_nr_pages; j++) {
1348 			const void __user *p;
1349 			int node;
1350 
1351 			err = -EFAULT;
1352 			if (get_user(p, pages + j + chunk_start))
1353 				goto out_pm;
1354 			pm[j].addr = (unsigned long) p;
1355 
1356 			if (get_user(node, nodes + j + chunk_start))
1357 				goto out_pm;
1358 
1359 			err = -ENODEV;
1360 			if (node < 0 || node >= MAX_NUMNODES)
1361 				goto out_pm;
1362 
1363 			if (!node_state(node, N_MEMORY))
1364 				goto out_pm;
1365 
1366 			err = -EACCES;
1367 			if (!node_isset(node, task_nodes))
1368 				goto out_pm;
1369 
1370 			pm[j].node = node;
1371 		}
1372 
1373 		/* End marker for this chunk */
1374 		pm[chunk_nr_pages].node = MAX_NUMNODES;
1375 
1376 		/* Migrate this chunk */
1377 		err = do_move_page_to_node_array(mm, pm,
1378 						 flags & MPOL_MF_MOVE_ALL);
1379 		if (err < 0)
1380 			goto out_pm;
1381 
1382 		/* Return status information */
1383 		for (j = 0; j < chunk_nr_pages; j++)
1384 			if (put_user(pm[j].status, status + j + chunk_start)) {
1385 				err = -EFAULT;
1386 				goto out_pm;
1387 			}
1388 	}
1389 	err = 0;
1390 
1391 out_pm:
1392 	free_page((unsigned long)pm);
1393 out:
1394 	return err;
1395 }
1396 
1397 /*
1398  * Determine the nodes of an array of pages and store it in an array of status.
1399  */
1400 static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages,
1401 				const void __user **pages, int *status)
1402 {
1403 	unsigned long i;
1404 
1405 	down_read(&mm->mmap_sem);
1406 
1407 	for (i = 0; i < nr_pages; i++) {
1408 		unsigned long addr = (unsigned long)(*pages);
1409 		struct vm_area_struct *vma;
1410 		struct page *page;
1411 		int err = -EFAULT;
1412 
1413 		vma = find_vma(mm, addr);
1414 		if (!vma || addr < vma->vm_start)
1415 			goto set_status;
1416 
1417 		/* FOLL_DUMP to ignore special (like zero) pages */
1418 		page = follow_page(vma, addr, FOLL_DUMP);
1419 
1420 		err = PTR_ERR(page);
1421 		if (IS_ERR(page))
1422 			goto set_status;
1423 
1424 		err = page ? page_to_nid(page) : -ENOENT;
1425 set_status:
1426 		*status = err;
1427 
1428 		pages++;
1429 		status++;
1430 	}
1431 
1432 	up_read(&mm->mmap_sem);
1433 }
1434 
1435 /*
1436  * Determine the nodes of a user array of pages and store it in
1437  * a user array of status.
1438  */
1439 static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages,
1440 			 const void __user * __user *pages,
1441 			 int __user *status)
1442 {
1443 #define DO_PAGES_STAT_CHUNK_NR 16
1444 	const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR];
1445 	int chunk_status[DO_PAGES_STAT_CHUNK_NR];
1446 
1447 	while (nr_pages) {
1448 		unsigned long chunk_nr;
1449 
1450 		chunk_nr = nr_pages;
1451 		if (chunk_nr > DO_PAGES_STAT_CHUNK_NR)
1452 			chunk_nr = DO_PAGES_STAT_CHUNK_NR;
1453 
1454 		if (copy_from_user(chunk_pages, pages, chunk_nr * sizeof(*chunk_pages)))
1455 			break;
1456 
1457 		do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status);
1458 
1459 		if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status)))
1460 			break;
1461 
1462 		pages += chunk_nr;
1463 		status += chunk_nr;
1464 		nr_pages -= chunk_nr;
1465 	}
1466 	return nr_pages ? -EFAULT : 0;
1467 }
1468 
1469 /*
1470  * Move a list of pages in the address space of the currently executing
1471  * process.
1472  */
1473 SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages,
1474 		const void __user * __user *, pages,
1475 		const int __user *, nodes,
1476 		int __user *, status, int, flags)
1477 {
1478 	const struct cred *cred = current_cred(), *tcred;
1479 	struct task_struct *task;
1480 	struct mm_struct *mm;
1481 	int err;
1482 	nodemask_t task_nodes;
1483 
1484 	/* Check flags */
1485 	if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
1486 		return -EINVAL;
1487 
1488 	if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1489 		return -EPERM;
1490 
1491 	/* Find the mm_struct */
1492 	rcu_read_lock();
1493 	task = pid ? find_task_by_vpid(pid) : current;
1494 	if (!task) {
1495 		rcu_read_unlock();
1496 		return -ESRCH;
1497 	}
1498 	get_task_struct(task);
1499 
1500 	/*
1501 	 * Check if this process has the right to modify the specified
1502 	 * process. The right exists if the process has administrative
1503 	 * capabilities, superuser privileges or the same
1504 	 * userid as the target process.
1505 	 */
1506 	tcred = __task_cred(task);
1507 	if (!uid_eq(cred->euid, tcred->suid) && !uid_eq(cred->euid, tcred->uid) &&
1508 	    !uid_eq(cred->uid,  tcred->suid) && !uid_eq(cred->uid,  tcred->uid) &&
1509 	    !capable(CAP_SYS_NICE)) {
1510 		rcu_read_unlock();
1511 		err = -EPERM;
1512 		goto out;
1513 	}
1514 	rcu_read_unlock();
1515 
1516  	err = security_task_movememory(task);
1517  	if (err)
1518 		goto out;
1519 
1520 	task_nodes = cpuset_mems_allowed(task);
1521 	mm = get_task_mm(task);
1522 	put_task_struct(task);
1523 
1524 	if (!mm)
1525 		return -EINVAL;
1526 
1527 	if (nodes)
1528 		err = do_pages_move(mm, task_nodes, nr_pages, pages,
1529 				    nodes, status, flags);
1530 	else
1531 		err = do_pages_stat(mm, nr_pages, pages, status);
1532 
1533 	mmput(mm);
1534 	return err;
1535 
1536 out:
1537 	put_task_struct(task);
1538 	return err;
1539 }
1540 
1541 #ifdef CONFIG_NUMA_BALANCING
1542 /*
1543  * Returns true if this is a safe migration target node for misplaced NUMA
1544  * pages. Currently it only checks the watermarks which crude
1545  */
1546 static bool migrate_balanced_pgdat(struct pglist_data *pgdat,
1547 				   unsigned long nr_migrate_pages)
1548 {
1549 	int z;
1550 	for (z = pgdat->nr_zones - 1; z >= 0; z--) {
1551 		struct zone *zone = pgdat->node_zones + z;
1552 
1553 		if (!populated_zone(zone))
1554 			continue;
1555 
1556 		if (!zone_reclaimable(zone))
1557 			continue;
1558 
1559 		/* Avoid waking kswapd by allocating pages_to_migrate pages. */
1560 		if (!zone_watermark_ok(zone, 0,
1561 				       high_wmark_pages(zone) +
1562 				       nr_migrate_pages,
1563 				       0, 0))
1564 			continue;
1565 		return true;
1566 	}
1567 	return false;
1568 }
1569 
1570 static struct page *alloc_misplaced_dst_page(struct page *page,
1571 					   unsigned long data,
1572 					   int **result)
1573 {
1574 	int nid = (int) data;
1575 	struct page *newpage;
1576 
1577 	newpage = __alloc_pages_node(nid,
1578 					 (GFP_HIGHUSER_MOVABLE |
1579 					  __GFP_THISNODE | __GFP_NOMEMALLOC |
1580 					  __GFP_NORETRY | __GFP_NOWARN) &
1581 					 ~(__GFP_IO | __GFP_FS), 0);
1582 
1583 	return newpage;
1584 }
1585 
1586 /*
1587  * page migration rate limiting control.
1588  * Do not migrate more than @pages_to_migrate in a @migrate_interval_millisecs
1589  * window of time. Default here says do not migrate more than 1280M per second.
1590  */
1591 static unsigned int migrate_interval_millisecs __read_mostly = 100;
1592 static unsigned int ratelimit_pages __read_mostly = 128 << (20 - PAGE_SHIFT);
1593 
1594 /* Returns true if the node is migrate rate-limited after the update */
1595 static bool numamigrate_update_ratelimit(pg_data_t *pgdat,
1596 					unsigned long nr_pages)
1597 {
1598 	/*
1599 	 * Rate-limit the amount of data that is being migrated to a node.
1600 	 * Optimal placement is no good if the memory bus is saturated and
1601 	 * all the time is being spent migrating!
1602 	 */
1603 	if (time_after(jiffies, pgdat->numabalancing_migrate_next_window)) {
1604 		spin_lock(&pgdat->numabalancing_migrate_lock);
1605 		pgdat->numabalancing_migrate_nr_pages = 0;
1606 		pgdat->numabalancing_migrate_next_window = jiffies +
1607 			msecs_to_jiffies(migrate_interval_millisecs);
1608 		spin_unlock(&pgdat->numabalancing_migrate_lock);
1609 	}
1610 	if (pgdat->numabalancing_migrate_nr_pages > ratelimit_pages) {
1611 		trace_mm_numa_migrate_ratelimit(current, pgdat->node_id,
1612 								nr_pages);
1613 		return true;
1614 	}
1615 
1616 	/*
1617 	 * This is an unlocked non-atomic update so errors are possible.
1618 	 * The consequences are failing to migrate when we potentiall should
1619 	 * have which is not severe enough to warrant locking. If it is ever
1620 	 * a problem, it can be converted to a per-cpu counter.
1621 	 */
1622 	pgdat->numabalancing_migrate_nr_pages += nr_pages;
1623 	return false;
1624 }
1625 
1626 static int numamigrate_isolate_page(pg_data_t *pgdat, struct page *page)
1627 {
1628 	int page_lru;
1629 
1630 	VM_BUG_ON_PAGE(compound_order(page) && !PageTransHuge(page), page);
1631 
1632 	/* Avoid migrating to a node that is nearly full */
1633 	if (!migrate_balanced_pgdat(pgdat, 1UL << compound_order(page)))
1634 		return 0;
1635 
1636 	if (isolate_lru_page(page))
1637 		return 0;
1638 
1639 	/*
1640 	 * migrate_misplaced_transhuge_page() skips page migration's usual
1641 	 * check on page_count(), so we must do it here, now that the page
1642 	 * has been isolated: a GUP pin, or any other pin, prevents migration.
1643 	 * The expected page count is 3: 1 for page's mapcount and 1 for the
1644 	 * caller's pin and 1 for the reference taken by isolate_lru_page().
1645 	 */
1646 	if (PageTransHuge(page) && page_count(page) != 3) {
1647 		putback_lru_page(page);
1648 		return 0;
1649 	}
1650 
1651 	page_lru = page_is_file_cache(page);
1652 	mod_zone_page_state(page_zone(page), NR_ISOLATED_ANON + page_lru,
1653 				hpage_nr_pages(page));
1654 
1655 	/*
1656 	 * Isolating the page has taken another reference, so the
1657 	 * caller's reference can be safely dropped without the page
1658 	 * disappearing underneath us during migration.
1659 	 */
1660 	put_page(page);
1661 	return 1;
1662 }
1663 
1664 bool pmd_trans_migrating(pmd_t pmd)
1665 {
1666 	struct page *page = pmd_page(pmd);
1667 	return PageLocked(page);
1668 }
1669 
1670 /*
1671  * Attempt to migrate a misplaced page to the specified destination
1672  * node. Caller is expected to have an elevated reference count on
1673  * the page that will be dropped by this function before returning.
1674  */
1675 int migrate_misplaced_page(struct page *page, struct vm_area_struct *vma,
1676 			   int node)
1677 {
1678 	pg_data_t *pgdat = NODE_DATA(node);
1679 	int isolated;
1680 	int nr_remaining;
1681 	LIST_HEAD(migratepages);
1682 
1683 	/*
1684 	 * Don't migrate file pages that are mapped in multiple processes
1685 	 * with execute permissions as they are probably shared libraries.
1686 	 */
1687 	if (page_mapcount(page) != 1 && page_is_file_cache(page) &&
1688 	    (vma->vm_flags & VM_EXEC))
1689 		goto out;
1690 
1691 	/*
1692 	 * Rate-limit the amount of data that is being migrated to a node.
1693 	 * Optimal placement is no good if the memory bus is saturated and
1694 	 * all the time is being spent migrating!
1695 	 */
1696 	if (numamigrate_update_ratelimit(pgdat, 1))
1697 		goto out;
1698 
1699 	isolated = numamigrate_isolate_page(pgdat, page);
1700 	if (!isolated)
1701 		goto out;
1702 
1703 	list_add(&page->lru, &migratepages);
1704 	nr_remaining = migrate_pages(&migratepages, alloc_misplaced_dst_page,
1705 				     NULL, node, MIGRATE_ASYNC,
1706 				     MR_NUMA_MISPLACED);
1707 	if (nr_remaining) {
1708 		if (!list_empty(&migratepages)) {
1709 			list_del(&page->lru);
1710 			dec_zone_page_state(page, NR_ISOLATED_ANON +
1711 					page_is_file_cache(page));
1712 			putback_lru_page(page);
1713 		}
1714 		isolated = 0;
1715 	} else
1716 		count_vm_numa_event(NUMA_PAGE_MIGRATE);
1717 	BUG_ON(!list_empty(&migratepages));
1718 	return isolated;
1719 
1720 out:
1721 	put_page(page);
1722 	return 0;
1723 }
1724 #endif /* CONFIG_NUMA_BALANCING */
1725 
1726 #if defined(CONFIG_NUMA_BALANCING) && defined(CONFIG_TRANSPARENT_HUGEPAGE)
1727 /*
1728  * Migrates a THP to a given target node. page must be locked and is unlocked
1729  * before returning.
1730  */
1731 int migrate_misplaced_transhuge_page(struct mm_struct *mm,
1732 				struct vm_area_struct *vma,
1733 				pmd_t *pmd, pmd_t entry,
1734 				unsigned long address,
1735 				struct page *page, int node)
1736 {
1737 	spinlock_t *ptl;
1738 	pg_data_t *pgdat = NODE_DATA(node);
1739 	int isolated = 0;
1740 	struct page *new_page = NULL;
1741 	int page_lru = page_is_file_cache(page);
1742 	unsigned long mmun_start = address & HPAGE_PMD_MASK;
1743 	unsigned long mmun_end = mmun_start + HPAGE_PMD_SIZE;
1744 	pmd_t orig_entry;
1745 
1746 	/*
1747 	 * Rate-limit the amount of data that is being migrated to a node.
1748 	 * Optimal placement is no good if the memory bus is saturated and
1749 	 * all the time is being spent migrating!
1750 	 */
1751 	if (numamigrate_update_ratelimit(pgdat, HPAGE_PMD_NR))
1752 		goto out_dropref;
1753 
1754 	new_page = alloc_pages_node(node,
1755 		(GFP_TRANSHUGE | __GFP_THISNODE) & ~__GFP_RECLAIM,
1756 		HPAGE_PMD_ORDER);
1757 	if (!new_page)
1758 		goto out_fail;
1759 
1760 	isolated = numamigrate_isolate_page(pgdat, page);
1761 	if (!isolated) {
1762 		put_page(new_page);
1763 		goto out_fail;
1764 	}
1765 
1766 	if (mm_tlb_flush_pending(mm))
1767 		flush_tlb_range(vma, mmun_start, mmun_end);
1768 
1769 	/* Prepare a page as a migration target */
1770 	__set_page_locked(new_page);
1771 	SetPageSwapBacked(new_page);
1772 
1773 	/* anon mapping, we can simply copy page->mapping to the new page: */
1774 	new_page->mapping = page->mapping;
1775 	new_page->index = page->index;
1776 	migrate_page_copy(new_page, page);
1777 	WARN_ON(PageLRU(new_page));
1778 
1779 	/* Recheck the target PMD */
1780 	mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1781 	ptl = pmd_lock(mm, pmd);
1782 	if (unlikely(!pmd_same(*pmd, entry) || page_count(page) != 2)) {
1783 fail_putback:
1784 		spin_unlock(ptl);
1785 		mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1786 
1787 		/* Reverse changes made by migrate_page_copy() */
1788 		if (TestClearPageActive(new_page))
1789 			SetPageActive(page);
1790 		if (TestClearPageUnevictable(new_page))
1791 			SetPageUnevictable(page);
1792 
1793 		unlock_page(new_page);
1794 		put_page(new_page);		/* Free it */
1795 
1796 		/* Retake the callers reference and putback on LRU */
1797 		get_page(page);
1798 		putback_lru_page(page);
1799 		mod_zone_page_state(page_zone(page),
1800 			 NR_ISOLATED_ANON + page_lru, -HPAGE_PMD_NR);
1801 
1802 		goto out_unlock;
1803 	}
1804 
1805 	orig_entry = *pmd;
1806 	entry = mk_pmd(new_page, vma->vm_page_prot);
1807 	entry = pmd_mkhuge(entry);
1808 	entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1809 
1810 	/*
1811 	 * Clear the old entry under pagetable lock and establish the new PTE.
1812 	 * Any parallel GUP will either observe the old page blocking on the
1813 	 * page lock, block on the page table lock or observe the new page.
1814 	 * The SetPageUptodate on the new page and page_add_new_anon_rmap
1815 	 * guarantee the copy is visible before the pagetable update.
1816 	 */
1817 	flush_cache_range(vma, mmun_start, mmun_end);
1818 	page_add_anon_rmap(new_page, vma, mmun_start);
1819 	pmdp_huge_clear_flush_notify(vma, mmun_start, pmd);
1820 	set_pmd_at(mm, mmun_start, pmd, entry);
1821 	flush_tlb_range(vma, mmun_start, mmun_end);
1822 	update_mmu_cache_pmd(vma, address, &entry);
1823 
1824 	if (page_count(page) != 2) {
1825 		set_pmd_at(mm, mmun_start, pmd, orig_entry);
1826 		flush_tlb_range(vma, mmun_start, mmun_end);
1827 		mmu_notifier_invalidate_range(mm, mmun_start, mmun_end);
1828 		update_mmu_cache_pmd(vma, address, &entry);
1829 		page_remove_rmap(new_page);
1830 		goto fail_putback;
1831 	}
1832 
1833 	mlock_migrate_page(new_page, page);
1834 	set_page_memcg(new_page, page_memcg(page));
1835 	set_page_memcg(page, NULL);
1836 	page_remove_rmap(page);
1837 
1838 	spin_unlock(ptl);
1839 	mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1840 
1841 	/* Take an "isolate" reference and put new page on the LRU. */
1842 	get_page(new_page);
1843 	putback_lru_page(new_page);
1844 
1845 	unlock_page(new_page);
1846 	unlock_page(page);
1847 	put_page(page);			/* Drop the rmap reference */
1848 	put_page(page);			/* Drop the LRU isolation reference */
1849 
1850 	count_vm_events(PGMIGRATE_SUCCESS, HPAGE_PMD_NR);
1851 	count_vm_numa_events(NUMA_PAGE_MIGRATE, HPAGE_PMD_NR);
1852 
1853 	mod_zone_page_state(page_zone(page),
1854 			NR_ISOLATED_ANON + page_lru,
1855 			-HPAGE_PMD_NR);
1856 	return isolated;
1857 
1858 out_fail:
1859 	count_vm_events(PGMIGRATE_FAIL, HPAGE_PMD_NR);
1860 out_dropref:
1861 	ptl = pmd_lock(mm, pmd);
1862 	if (pmd_same(*pmd, entry)) {
1863 		entry = pmd_modify(entry, vma->vm_page_prot);
1864 		set_pmd_at(mm, mmun_start, pmd, entry);
1865 		update_mmu_cache_pmd(vma, address, &entry);
1866 	}
1867 	spin_unlock(ptl);
1868 
1869 out_unlock:
1870 	unlock_page(page);
1871 	put_page(page);
1872 	return 0;
1873 }
1874 #endif /* CONFIG_NUMA_BALANCING */
1875 
1876 #endif /* CONFIG_NUMA */
1877