xref: /openbmc/linux/mm/rmap.c (revision fd589a8f)
1 /*
2  * mm/rmap.c - physical to virtual reverse mappings
3  *
4  * Copyright 2001, Rik van Riel <riel@conectiva.com.br>
5  * Released under the General Public License (GPL).
6  *
7  * Simple, low overhead reverse mapping scheme.
8  * Please try to keep this thing as modular as possible.
9  *
10  * Provides methods for unmapping each kind of mapped page:
11  * the anon methods track anonymous pages, and
12  * the file methods track pages belonging to an inode.
13  *
14  * Original design by Rik van Riel <riel@conectiva.com.br> 2001
15  * File methods by Dave McCracken <dmccr@us.ibm.com> 2003, 2004
16  * Anonymous methods by Andrea Arcangeli <andrea@suse.de> 2004
17  * Contributions by Hugh Dickins 2003, 2004
18  */
19 
20 /*
21  * Lock ordering in mm:
22  *
23  * inode->i_mutex	(while writing or truncating, not reading or faulting)
24  *   inode->i_alloc_sem (vmtruncate_range)
25  *   mm->mmap_sem
26  *     page->flags PG_locked (lock_page)
27  *       mapping->i_mmap_lock
28  *         anon_vma->lock
29  *           mm->page_table_lock or pte_lock
30  *             zone->lru_lock (in mark_page_accessed, isolate_lru_page)
31  *             swap_lock (in swap_duplicate, swap_info_get)
32  *               mmlist_lock (in mmput, drain_mmlist and others)
33  *               mapping->private_lock (in __set_page_dirty_buffers)
34  *               inode_lock (in set_page_dirty's __mark_inode_dirty)
35  *                 sb_lock (within inode_lock in fs/fs-writeback.c)
36  *                 mapping->tree_lock (widely used, in set_page_dirty,
37  *                           in arch-dependent flush_dcache_mmap_lock,
38  *                           within inode_lock in __sync_single_inode)
39  */
40 
41 #include <linux/mm.h>
42 #include <linux/pagemap.h>
43 #include <linux/swap.h>
44 #include <linux/swapops.h>
45 #include <linux/slab.h>
46 #include <linux/init.h>
47 #include <linux/rmap.h>
48 #include <linux/rcupdate.h>
49 #include <linux/module.h>
50 #include <linux/memcontrol.h>
51 #include <linux/mmu_notifier.h>
52 #include <linux/migrate.h>
53 
54 #include <asm/tlbflush.h>
55 
56 #include "internal.h"
57 
58 static struct kmem_cache *anon_vma_cachep;
59 
60 static inline struct anon_vma *anon_vma_alloc(void)
61 {
62 	return kmem_cache_alloc(anon_vma_cachep, GFP_KERNEL);
63 }
64 
65 static inline void anon_vma_free(struct anon_vma *anon_vma)
66 {
67 	kmem_cache_free(anon_vma_cachep, anon_vma);
68 }
69 
70 /**
71  * anon_vma_prepare - attach an anon_vma to a memory region
72  * @vma: the memory region in question
73  *
74  * This makes sure the memory mapping described by 'vma' has
75  * an 'anon_vma' attached to it, so that we can associate the
76  * anonymous pages mapped into it with that anon_vma.
77  *
78  * The common case will be that we already have one, but if
79  * if not we either need to find an adjacent mapping that we
80  * can re-use the anon_vma from (very common when the only
81  * reason for splitting a vma has been mprotect()), or we
82  * allocate a new one.
83  *
84  * Anon-vma allocations are very subtle, because we may have
85  * optimistically looked up an anon_vma in page_lock_anon_vma()
86  * and that may actually touch the spinlock even in the newly
87  * allocated vma (it depends on RCU to make sure that the
88  * anon_vma isn't actually destroyed).
89  *
90  * As a result, we need to do proper anon_vma locking even
91  * for the new allocation. At the same time, we do not want
92  * to do any locking for the common case of already having
93  * an anon_vma.
94  *
95  * This must be called with the mmap_sem held for reading.
96  */
97 int anon_vma_prepare(struct vm_area_struct *vma)
98 {
99 	struct anon_vma *anon_vma = vma->anon_vma;
100 
101 	might_sleep();
102 	if (unlikely(!anon_vma)) {
103 		struct mm_struct *mm = vma->vm_mm;
104 		struct anon_vma *allocated;
105 
106 		anon_vma = find_mergeable_anon_vma(vma);
107 		allocated = NULL;
108 		if (!anon_vma) {
109 			anon_vma = anon_vma_alloc();
110 			if (unlikely(!anon_vma))
111 				return -ENOMEM;
112 			allocated = anon_vma;
113 		}
114 		spin_lock(&anon_vma->lock);
115 
116 		/* page_table_lock to protect against threads */
117 		spin_lock(&mm->page_table_lock);
118 		if (likely(!vma->anon_vma)) {
119 			vma->anon_vma = anon_vma;
120 			list_add_tail(&vma->anon_vma_node, &anon_vma->head);
121 			allocated = NULL;
122 		}
123 		spin_unlock(&mm->page_table_lock);
124 
125 		spin_unlock(&anon_vma->lock);
126 		if (unlikely(allocated))
127 			anon_vma_free(allocated);
128 	}
129 	return 0;
130 }
131 
132 void __anon_vma_merge(struct vm_area_struct *vma, struct vm_area_struct *next)
133 {
134 	BUG_ON(vma->anon_vma != next->anon_vma);
135 	list_del(&next->anon_vma_node);
136 }
137 
138 void __anon_vma_link(struct vm_area_struct *vma)
139 {
140 	struct anon_vma *anon_vma = vma->anon_vma;
141 
142 	if (anon_vma)
143 		list_add_tail(&vma->anon_vma_node, &anon_vma->head);
144 }
145 
146 void anon_vma_link(struct vm_area_struct *vma)
147 {
148 	struct anon_vma *anon_vma = vma->anon_vma;
149 
150 	if (anon_vma) {
151 		spin_lock(&anon_vma->lock);
152 		list_add_tail(&vma->anon_vma_node, &anon_vma->head);
153 		spin_unlock(&anon_vma->lock);
154 	}
155 }
156 
157 void anon_vma_unlink(struct vm_area_struct *vma)
158 {
159 	struct anon_vma *anon_vma = vma->anon_vma;
160 	int empty;
161 
162 	if (!anon_vma)
163 		return;
164 
165 	spin_lock(&anon_vma->lock);
166 	list_del(&vma->anon_vma_node);
167 
168 	/* We must garbage collect the anon_vma if it's empty */
169 	empty = list_empty(&anon_vma->head);
170 	spin_unlock(&anon_vma->lock);
171 
172 	if (empty)
173 		anon_vma_free(anon_vma);
174 }
175 
176 static void anon_vma_ctor(void *data)
177 {
178 	struct anon_vma *anon_vma = data;
179 
180 	spin_lock_init(&anon_vma->lock);
181 	INIT_LIST_HEAD(&anon_vma->head);
182 }
183 
184 void __init anon_vma_init(void)
185 {
186 	anon_vma_cachep = kmem_cache_create("anon_vma", sizeof(struct anon_vma),
187 			0, SLAB_DESTROY_BY_RCU|SLAB_PANIC, anon_vma_ctor);
188 }
189 
190 /*
191  * Getting a lock on a stable anon_vma from a page off the LRU is
192  * tricky: page_lock_anon_vma rely on RCU to guard against the races.
193  */
194 static struct anon_vma *page_lock_anon_vma(struct page *page)
195 {
196 	struct anon_vma *anon_vma;
197 	unsigned long anon_mapping;
198 
199 	rcu_read_lock();
200 	anon_mapping = (unsigned long) page->mapping;
201 	if (!(anon_mapping & PAGE_MAPPING_ANON))
202 		goto out;
203 	if (!page_mapped(page))
204 		goto out;
205 
206 	anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
207 	spin_lock(&anon_vma->lock);
208 	return anon_vma;
209 out:
210 	rcu_read_unlock();
211 	return NULL;
212 }
213 
214 static void page_unlock_anon_vma(struct anon_vma *anon_vma)
215 {
216 	spin_unlock(&anon_vma->lock);
217 	rcu_read_unlock();
218 }
219 
220 /*
221  * At what user virtual address is page expected in @vma?
222  * Returns virtual address or -EFAULT if page's index/offset is not
223  * within the range mapped the @vma.
224  */
225 static inline unsigned long
226 vma_address(struct page *page, struct vm_area_struct *vma)
227 {
228 	pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
229 	unsigned long address;
230 
231 	address = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
232 	if (unlikely(address < vma->vm_start || address >= vma->vm_end)) {
233 		/* page should be within @vma mapping range */
234 		return -EFAULT;
235 	}
236 	return address;
237 }
238 
239 /*
240  * At what user virtual address is page expected in vma? checking that the
241  * page matches the vma: currently only used on anon pages, by unuse_vma;
242  */
243 unsigned long page_address_in_vma(struct page *page, struct vm_area_struct *vma)
244 {
245 	if (PageAnon(page)) {
246 		if ((void *)vma->anon_vma !=
247 		    (void *)page->mapping - PAGE_MAPPING_ANON)
248 			return -EFAULT;
249 	} else if (page->mapping && !(vma->vm_flags & VM_NONLINEAR)) {
250 		if (!vma->vm_file ||
251 		    vma->vm_file->f_mapping != page->mapping)
252 			return -EFAULT;
253 	} else
254 		return -EFAULT;
255 	return vma_address(page, vma);
256 }
257 
258 /*
259  * Check that @page is mapped at @address into @mm.
260  *
261  * If @sync is false, page_check_address may perform a racy check to avoid
262  * the page table lock when the pte is not present (helpful when reclaiming
263  * highly shared pages).
264  *
265  * On success returns with pte mapped and locked.
266  */
267 pte_t *page_check_address(struct page *page, struct mm_struct *mm,
268 			  unsigned long address, spinlock_t **ptlp, int sync)
269 {
270 	pgd_t *pgd;
271 	pud_t *pud;
272 	pmd_t *pmd;
273 	pte_t *pte;
274 	spinlock_t *ptl;
275 
276 	pgd = pgd_offset(mm, address);
277 	if (!pgd_present(*pgd))
278 		return NULL;
279 
280 	pud = pud_offset(pgd, address);
281 	if (!pud_present(*pud))
282 		return NULL;
283 
284 	pmd = pmd_offset(pud, address);
285 	if (!pmd_present(*pmd))
286 		return NULL;
287 
288 	pte = pte_offset_map(pmd, address);
289 	/* Make a quick check before getting the lock */
290 	if (!sync && !pte_present(*pte)) {
291 		pte_unmap(pte);
292 		return NULL;
293 	}
294 
295 	ptl = pte_lockptr(mm, pmd);
296 	spin_lock(ptl);
297 	if (pte_present(*pte) && page_to_pfn(page) == pte_pfn(*pte)) {
298 		*ptlp = ptl;
299 		return pte;
300 	}
301 	pte_unmap_unlock(pte, ptl);
302 	return NULL;
303 }
304 
305 /**
306  * page_mapped_in_vma - check whether a page is really mapped in a VMA
307  * @page: the page to test
308  * @vma: the VMA to test
309  *
310  * Returns 1 if the page is mapped into the page tables of the VMA, 0
311  * if the page is not mapped into the page tables of this VMA.  Only
312  * valid for normal file or anonymous VMAs.
313  */
314 static int page_mapped_in_vma(struct page *page, struct vm_area_struct *vma)
315 {
316 	unsigned long address;
317 	pte_t *pte;
318 	spinlock_t *ptl;
319 
320 	address = vma_address(page, vma);
321 	if (address == -EFAULT)		/* out of vma range */
322 		return 0;
323 	pte = page_check_address(page, vma->vm_mm, address, &ptl, 1);
324 	if (!pte)			/* the page is not in this mm */
325 		return 0;
326 	pte_unmap_unlock(pte, ptl);
327 
328 	return 1;
329 }
330 
331 /*
332  * Subfunctions of page_referenced: page_referenced_one called
333  * repeatedly from either page_referenced_anon or page_referenced_file.
334  */
335 static int page_referenced_one(struct page *page,
336 			       struct vm_area_struct *vma,
337 			       unsigned int *mapcount,
338 			       unsigned long *vm_flags)
339 {
340 	struct mm_struct *mm = vma->vm_mm;
341 	unsigned long address;
342 	pte_t *pte;
343 	spinlock_t *ptl;
344 	int referenced = 0;
345 
346 	address = vma_address(page, vma);
347 	if (address == -EFAULT)
348 		goto out;
349 
350 	pte = page_check_address(page, mm, address, &ptl, 0);
351 	if (!pte)
352 		goto out;
353 
354 	/*
355 	 * Don't want to elevate referenced for mlocked page that gets this far,
356 	 * in order that it progresses to try_to_unmap and is moved to the
357 	 * unevictable list.
358 	 */
359 	if (vma->vm_flags & VM_LOCKED) {
360 		*mapcount = 1;	/* break early from loop */
361 		*vm_flags |= VM_LOCKED;
362 		goto out_unmap;
363 	}
364 
365 	if (ptep_clear_flush_young_notify(vma, address, pte)) {
366 		/*
367 		 * Don't treat a reference through a sequentially read
368 		 * mapping as such.  If the page has been used in
369 		 * another mapping, we will catch it; if this other
370 		 * mapping is already gone, the unmap path will have
371 		 * set PG_referenced or activated the page.
372 		 */
373 		if (likely(!VM_SequentialReadHint(vma)))
374 			referenced++;
375 	}
376 
377 	/* Pretend the page is referenced if the task has the
378 	   swap token and is in the middle of a page fault. */
379 	if (mm != current->mm && has_swap_token(mm) &&
380 			rwsem_is_locked(&mm->mmap_sem))
381 		referenced++;
382 
383 out_unmap:
384 	(*mapcount)--;
385 	pte_unmap_unlock(pte, ptl);
386 out:
387 	if (referenced)
388 		*vm_flags |= vma->vm_flags;
389 	return referenced;
390 }
391 
392 static int page_referenced_anon(struct page *page,
393 				struct mem_cgroup *mem_cont,
394 				unsigned long *vm_flags)
395 {
396 	unsigned int mapcount;
397 	struct anon_vma *anon_vma;
398 	struct vm_area_struct *vma;
399 	int referenced = 0;
400 
401 	anon_vma = page_lock_anon_vma(page);
402 	if (!anon_vma)
403 		return referenced;
404 
405 	mapcount = page_mapcount(page);
406 	list_for_each_entry(vma, &anon_vma->head, anon_vma_node) {
407 		/*
408 		 * If we are reclaiming on behalf of a cgroup, skip
409 		 * counting on behalf of references from different
410 		 * cgroups
411 		 */
412 		if (mem_cont && !mm_match_cgroup(vma->vm_mm, mem_cont))
413 			continue;
414 		referenced += page_referenced_one(page, vma,
415 						  &mapcount, vm_flags);
416 		if (!mapcount)
417 			break;
418 	}
419 
420 	page_unlock_anon_vma(anon_vma);
421 	return referenced;
422 }
423 
424 /**
425  * page_referenced_file - referenced check for object-based rmap
426  * @page: the page we're checking references on.
427  * @mem_cont: target memory controller
428  * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
429  *
430  * For an object-based mapped page, find all the places it is mapped and
431  * check/clear the referenced flag.  This is done by following the page->mapping
432  * pointer, then walking the chain of vmas it holds.  It returns the number
433  * of references it found.
434  *
435  * This function is only called from page_referenced for object-based pages.
436  */
437 static int page_referenced_file(struct page *page,
438 				struct mem_cgroup *mem_cont,
439 				unsigned long *vm_flags)
440 {
441 	unsigned int mapcount;
442 	struct address_space *mapping = page->mapping;
443 	pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
444 	struct vm_area_struct *vma;
445 	struct prio_tree_iter iter;
446 	int referenced = 0;
447 
448 	/*
449 	 * The caller's checks on page->mapping and !PageAnon have made
450 	 * sure that this is a file page: the check for page->mapping
451 	 * excludes the case just before it gets set on an anon page.
452 	 */
453 	BUG_ON(PageAnon(page));
454 
455 	/*
456 	 * The page lock not only makes sure that page->mapping cannot
457 	 * suddenly be NULLified by truncation, it makes sure that the
458 	 * structure at mapping cannot be freed and reused yet,
459 	 * so we can safely take mapping->i_mmap_lock.
460 	 */
461 	BUG_ON(!PageLocked(page));
462 
463 	spin_lock(&mapping->i_mmap_lock);
464 
465 	/*
466 	 * i_mmap_lock does not stabilize mapcount at all, but mapcount
467 	 * is more likely to be accurate if we note it after spinning.
468 	 */
469 	mapcount = page_mapcount(page);
470 
471 	vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
472 		/*
473 		 * If we are reclaiming on behalf of a cgroup, skip
474 		 * counting on behalf of references from different
475 		 * cgroups
476 		 */
477 		if (mem_cont && !mm_match_cgroup(vma->vm_mm, mem_cont))
478 			continue;
479 		referenced += page_referenced_one(page, vma,
480 						  &mapcount, vm_flags);
481 		if (!mapcount)
482 			break;
483 	}
484 
485 	spin_unlock(&mapping->i_mmap_lock);
486 	return referenced;
487 }
488 
489 /**
490  * page_referenced - test if the page was referenced
491  * @page: the page to test
492  * @is_locked: caller holds lock on the page
493  * @mem_cont: target memory controller
494  * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
495  *
496  * Quick test_and_clear_referenced for all mappings to a page,
497  * returns the number of ptes which referenced the page.
498  */
499 int page_referenced(struct page *page,
500 		    int is_locked,
501 		    struct mem_cgroup *mem_cont,
502 		    unsigned long *vm_flags)
503 {
504 	int referenced = 0;
505 
506 	if (TestClearPageReferenced(page))
507 		referenced++;
508 
509 	*vm_flags = 0;
510 	if (page_mapped(page) && page->mapping) {
511 		if (PageAnon(page))
512 			referenced += page_referenced_anon(page, mem_cont,
513 								vm_flags);
514 		else if (is_locked)
515 			referenced += page_referenced_file(page, mem_cont,
516 								vm_flags);
517 		else if (!trylock_page(page))
518 			referenced++;
519 		else {
520 			if (page->mapping)
521 				referenced += page_referenced_file(page,
522 							mem_cont, vm_flags);
523 			unlock_page(page);
524 		}
525 	}
526 
527 	if (page_test_and_clear_young(page))
528 		referenced++;
529 
530 	return referenced;
531 }
532 
533 static int page_mkclean_one(struct page *page, struct vm_area_struct *vma)
534 {
535 	struct mm_struct *mm = vma->vm_mm;
536 	unsigned long address;
537 	pte_t *pte;
538 	spinlock_t *ptl;
539 	int ret = 0;
540 
541 	address = vma_address(page, vma);
542 	if (address == -EFAULT)
543 		goto out;
544 
545 	pte = page_check_address(page, mm, address, &ptl, 1);
546 	if (!pte)
547 		goto out;
548 
549 	if (pte_dirty(*pte) || pte_write(*pte)) {
550 		pte_t entry;
551 
552 		flush_cache_page(vma, address, pte_pfn(*pte));
553 		entry = ptep_clear_flush_notify(vma, address, pte);
554 		entry = pte_wrprotect(entry);
555 		entry = pte_mkclean(entry);
556 		set_pte_at(mm, address, pte, entry);
557 		ret = 1;
558 	}
559 
560 	pte_unmap_unlock(pte, ptl);
561 out:
562 	return ret;
563 }
564 
565 static int page_mkclean_file(struct address_space *mapping, struct page *page)
566 {
567 	pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
568 	struct vm_area_struct *vma;
569 	struct prio_tree_iter iter;
570 	int ret = 0;
571 
572 	BUG_ON(PageAnon(page));
573 
574 	spin_lock(&mapping->i_mmap_lock);
575 	vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
576 		if (vma->vm_flags & VM_SHARED)
577 			ret += page_mkclean_one(page, vma);
578 	}
579 	spin_unlock(&mapping->i_mmap_lock);
580 	return ret;
581 }
582 
583 int page_mkclean(struct page *page)
584 {
585 	int ret = 0;
586 
587 	BUG_ON(!PageLocked(page));
588 
589 	if (page_mapped(page)) {
590 		struct address_space *mapping = page_mapping(page);
591 		if (mapping) {
592 			ret = page_mkclean_file(mapping, page);
593 			if (page_test_dirty(page)) {
594 				page_clear_dirty(page);
595 				ret = 1;
596 			}
597 		}
598 	}
599 
600 	return ret;
601 }
602 EXPORT_SYMBOL_GPL(page_mkclean);
603 
604 /**
605  * __page_set_anon_rmap - setup new anonymous rmap
606  * @page:	the page to add the mapping to
607  * @vma:	the vm area in which the mapping is added
608  * @address:	the user virtual address mapped
609  */
610 static void __page_set_anon_rmap(struct page *page,
611 	struct vm_area_struct *vma, unsigned long address)
612 {
613 	struct anon_vma *anon_vma = vma->anon_vma;
614 
615 	BUG_ON(!anon_vma);
616 	anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
617 	page->mapping = (struct address_space *) anon_vma;
618 
619 	page->index = linear_page_index(vma, address);
620 
621 	/*
622 	 * nr_mapped state can be updated without turning off
623 	 * interrupts because it is not modified via interrupt.
624 	 */
625 	__inc_zone_page_state(page, NR_ANON_PAGES);
626 }
627 
628 /**
629  * __page_check_anon_rmap - sanity check anonymous rmap addition
630  * @page:	the page to add the mapping to
631  * @vma:	the vm area in which the mapping is added
632  * @address:	the user virtual address mapped
633  */
634 static void __page_check_anon_rmap(struct page *page,
635 	struct vm_area_struct *vma, unsigned long address)
636 {
637 #ifdef CONFIG_DEBUG_VM
638 	/*
639 	 * The page's anon-rmap details (mapping and index) are guaranteed to
640 	 * be set up correctly at this point.
641 	 *
642 	 * We have exclusion against page_add_anon_rmap because the caller
643 	 * always holds the page locked, except if called from page_dup_rmap,
644 	 * in which case the page is already known to be setup.
645 	 *
646 	 * We have exclusion against page_add_new_anon_rmap because those pages
647 	 * are initially only visible via the pagetables, and the pte is locked
648 	 * over the call to page_add_new_anon_rmap.
649 	 */
650 	struct anon_vma *anon_vma = vma->anon_vma;
651 	anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
652 	BUG_ON(page->mapping != (struct address_space *)anon_vma);
653 	BUG_ON(page->index != linear_page_index(vma, address));
654 #endif
655 }
656 
657 /**
658  * page_add_anon_rmap - add pte mapping to an anonymous page
659  * @page:	the page to add the mapping to
660  * @vma:	the vm area in which the mapping is added
661  * @address:	the user virtual address mapped
662  *
663  * The caller needs to hold the pte lock and the page must be locked.
664  */
665 void page_add_anon_rmap(struct page *page,
666 	struct vm_area_struct *vma, unsigned long address)
667 {
668 	VM_BUG_ON(!PageLocked(page));
669 	VM_BUG_ON(address < vma->vm_start || address >= vma->vm_end);
670 	if (atomic_inc_and_test(&page->_mapcount))
671 		__page_set_anon_rmap(page, vma, address);
672 	else
673 		__page_check_anon_rmap(page, vma, address);
674 }
675 
676 /**
677  * page_add_new_anon_rmap - add pte mapping to a new anonymous page
678  * @page:	the page to add the mapping to
679  * @vma:	the vm area in which the mapping is added
680  * @address:	the user virtual address mapped
681  *
682  * Same as page_add_anon_rmap but must only be called on *new* pages.
683  * This means the inc-and-test can be bypassed.
684  * Page does not have to be locked.
685  */
686 void page_add_new_anon_rmap(struct page *page,
687 	struct vm_area_struct *vma, unsigned long address)
688 {
689 	VM_BUG_ON(address < vma->vm_start || address >= vma->vm_end);
690 	SetPageSwapBacked(page);
691 	atomic_set(&page->_mapcount, 0); /* increment count (starts at -1) */
692 	__page_set_anon_rmap(page, vma, address);
693 	if (page_evictable(page, vma))
694 		lru_cache_add_lru(page, LRU_ACTIVE_ANON);
695 	else
696 		add_page_to_unevictable_list(page);
697 }
698 
699 /**
700  * page_add_file_rmap - add pte mapping to a file page
701  * @page: the page to add the mapping to
702  *
703  * The caller needs to hold the pte lock.
704  */
705 void page_add_file_rmap(struct page *page)
706 {
707 	if (atomic_inc_and_test(&page->_mapcount)) {
708 		__inc_zone_page_state(page, NR_FILE_MAPPED);
709 		mem_cgroup_update_mapped_file_stat(page, 1);
710 	}
711 }
712 
713 #ifdef CONFIG_DEBUG_VM
714 /**
715  * page_dup_rmap - duplicate pte mapping to a page
716  * @page:	the page to add the mapping to
717  * @vma:	the vm area being duplicated
718  * @address:	the user virtual address mapped
719  *
720  * For copy_page_range only: minimal extract from page_add_file_rmap /
721  * page_add_anon_rmap, avoiding unnecessary tests (already checked) so it's
722  * quicker.
723  *
724  * The caller needs to hold the pte lock.
725  */
726 void page_dup_rmap(struct page *page, struct vm_area_struct *vma, unsigned long address)
727 {
728 	if (PageAnon(page))
729 		__page_check_anon_rmap(page, vma, address);
730 	atomic_inc(&page->_mapcount);
731 }
732 #endif
733 
734 /**
735  * page_remove_rmap - take down pte mapping from a page
736  * @page: page to remove mapping from
737  *
738  * The caller needs to hold the pte lock.
739  */
740 void page_remove_rmap(struct page *page)
741 {
742 	if (atomic_add_negative(-1, &page->_mapcount)) {
743 		/*
744 		 * Now that the last pte has gone, s390 must transfer dirty
745 		 * flag from storage key to struct page.  We can usually skip
746 		 * this if the page is anon, so about to be freed; but perhaps
747 		 * not if it's in swapcache - there might be another pte slot
748 		 * containing the swap entry, but page not yet written to swap.
749 		 */
750 		if ((!PageAnon(page) || PageSwapCache(page)) &&
751 		    page_test_dirty(page)) {
752 			page_clear_dirty(page);
753 			set_page_dirty(page);
754 		}
755 		if (PageAnon(page))
756 			mem_cgroup_uncharge_page(page);
757 		__dec_zone_page_state(page,
758 			PageAnon(page) ? NR_ANON_PAGES : NR_FILE_MAPPED);
759 		mem_cgroup_update_mapped_file_stat(page, -1);
760 		/*
761 		 * It would be tidy to reset the PageAnon mapping here,
762 		 * but that might overwrite a racing page_add_anon_rmap
763 		 * which increments mapcount after us but sets mapping
764 		 * before us: so leave the reset to free_hot_cold_page,
765 		 * and remember that it's only reliable while mapped.
766 		 * Leaving it set also helps swapoff to reinstate ptes
767 		 * faster for those pages still in swapcache.
768 		 */
769 	}
770 }
771 
772 /*
773  * Subfunctions of try_to_unmap: try_to_unmap_one called
774  * repeatedly from either try_to_unmap_anon or try_to_unmap_file.
775  */
776 static int try_to_unmap_one(struct page *page, struct vm_area_struct *vma,
777 				int migration)
778 {
779 	struct mm_struct *mm = vma->vm_mm;
780 	unsigned long address;
781 	pte_t *pte;
782 	pte_t pteval;
783 	spinlock_t *ptl;
784 	int ret = SWAP_AGAIN;
785 
786 	address = vma_address(page, vma);
787 	if (address == -EFAULT)
788 		goto out;
789 
790 	pte = page_check_address(page, mm, address, &ptl, 0);
791 	if (!pte)
792 		goto out;
793 
794 	/*
795 	 * If the page is mlock()d, we cannot swap it out.
796 	 * If it's recently referenced (perhaps page_referenced
797 	 * skipped over this mm) then we should reactivate it.
798 	 */
799 	if (!migration) {
800 		if (vma->vm_flags & VM_LOCKED) {
801 			ret = SWAP_MLOCK;
802 			goto out_unmap;
803 		}
804 		if (ptep_clear_flush_young_notify(vma, address, pte)) {
805 			ret = SWAP_FAIL;
806 			goto out_unmap;
807 		}
808   	}
809 
810 	/* Nuke the page table entry. */
811 	flush_cache_page(vma, address, page_to_pfn(page));
812 	pteval = ptep_clear_flush_notify(vma, address, pte);
813 
814 	/* Move the dirty bit to the physical page now the pte is gone. */
815 	if (pte_dirty(pteval))
816 		set_page_dirty(page);
817 
818 	/* Update high watermark before we lower rss */
819 	update_hiwater_rss(mm);
820 
821 	if (PageAnon(page)) {
822 		swp_entry_t entry = { .val = page_private(page) };
823 
824 		if (PageSwapCache(page)) {
825 			/*
826 			 * Store the swap location in the pte.
827 			 * See handle_pte_fault() ...
828 			 */
829 			swap_duplicate(entry);
830 			if (list_empty(&mm->mmlist)) {
831 				spin_lock(&mmlist_lock);
832 				if (list_empty(&mm->mmlist))
833 					list_add(&mm->mmlist, &init_mm.mmlist);
834 				spin_unlock(&mmlist_lock);
835 			}
836 			dec_mm_counter(mm, anon_rss);
837 		} else if (PAGE_MIGRATION) {
838 			/*
839 			 * Store the pfn of the page in a special migration
840 			 * pte. do_swap_page() will wait until the migration
841 			 * pte is removed and then restart fault handling.
842 			 */
843 			BUG_ON(!migration);
844 			entry = make_migration_entry(page, pte_write(pteval));
845 		}
846 		set_pte_at(mm, address, pte, swp_entry_to_pte(entry));
847 		BUG_ON(pte_file(*pte));
848 	} else if (PAGE_MIGRATION && migration) {
849 		/* Establish migration entry for a file page */
850 		swp_entry_t entry;
851 		entry = make_migration_entry(page, pte_write(pteval));
852 		set_pte_at(mm, address, pte, swp_entry_to_pte(entry));
853 	} else
854 		dec_mm_counter(mm, file_rss);
855 
856 
857 	page_remove_rmap(page);
858 	page_cache_release(page);
859 
860 out_unmap:
861 	pte_unmap_unlock(pte, ptl);
862 out:
863 	return ret;
864 }
865 
866 /*
867  * objrmap doesn't work for nonlinear VMAs because the assumption that
868  * offset-into-file correlates with offset-into-virtual-addresses does not hold.
869  * Consequently, given a particular page and its ->index, we cannot locate the
870  * ptes which are mapping that page without an exhaustive linear search.
871  *
872  * So what this code does is a mini "virtual scan" of each nonlinear VMA which
873  * maps the file to which the target page belongs.  The ->vm_private_data field
874  * holds the current cursor into that scan.  Successive searches will circulate
875  * around the vma's virtual address space.
876  *
877  * So as more replacement pressure is applied to the pages in a nonlinear VMA,
878  * more scanning pressure is placed against them as well.   Eventually pages
879  * will become fully unmapped and are eligible for eviction.
880  *
881  * For very sparsely populated VMAs this is a little inefficient - chances are
882  * there there won't be many ptes located within the scan cluster.  In this case
883  * maybe we could scan further - to the end of the pte page, perhaps.
884  *
885  * Mlocked pages:  check VM_LOCKED under mmap_sem held for read, if we can
886  * acquire it without blocking.  If vma locked, mlock the pages in the cluster,
887  * rather than unmapping them.  If we encounter the "check_page" that vmscan is
888  * trying to unmap, return SWAP_MLOCK, else default SWAP_AGAIN.
889  */
890 #define CLUSTER_SIZE	min(32*PAGE_SIZE, PMD_SIZE)
891 #define CLUSTER_MASK	(~(CLUSTER_SIZE - 1))
892 
893 static int try_to_unmap_cluster(unsigned long cursor, unsigned int *mapcount,
894 		struct vm_area_struct *vma, struct page *check_page)
895 {
896 	struct mm_struct *mm = vma->vm_mm;
897 	pgd_t *pgd;
898 	pud_t *pud;
899 	pmd_t *pmd;
900 	pte_t *pte;
901 	pte_t pteval;
902 	spinlock_t *ptl;
903 	struct page *page;
904 	unsigned long address;
905 	unsigned long end;
906 	int ret = SWAP_AGAIN;
907 	int locked_vma = 0;
908 
909 	address = (vma->vm_start + cursor) & CLUSTER_MASK;
910 	end = address + CLUSTER_SIZE;
911 	if (address < vma->vm_start)
912 		address = vma->vm_start;
913 	if (end > vma->vm_end)
914 		end = vma->vm_end;
915 
916 	pgd = pgd_offset(mm, address);
917 	if (!pgd_present(*pgd))
918 		return ret;
919 
920 	pud = pud_offset(pgd, address);
921 	if (!pud_present(*pud))
922 		return ret;
923 
924 	pmd = pmd_offset(pud, address);
925 	if (!pmd_present(*pmd))
926 		return ret;
927 
928 	/*
929 	 * MLOCK_PAGES => feature is configured.
930 	 * if we can acquire the mmap_sem for read, and vma is VM_LOCKED,
931 	 * keep the sem while scanning the cluster for mlocking pages.
932 	 */
933 	if (MLOCK_PAGES && down_read_trylock(&vma->vm_mm->mmap_sem)) {
934 		locked_vma = (vma->vm_flags & VM_LOCKED);
935 		if (!locked_vma)
936 			up_read(&vma->vm_mm->mmap_sem); /* don't need it */
937 	}
938 
939 	pte = pte_offset_map_lock(mm, pmd, address, &ptl);
940 
941 	/* Update high watermark before we lower rss */
942 	update_hiwater_rss(mm);
943 
944 	for (; address < end; pte++, address += PAGE_SIZE) {
945 		if (!pte_present(*pte))
946 			continue;
947 		page = vm_normal_page(vma, address, *pte);
948 		BUG_ON(!page || PageAnon(page));
949 
950 		if (locked_vma) {
951 			mlock_vma_page(page);   /* no-op if already mlocked */
952 			if (page == check_page)
953 				ret = SWAP_MLOCK;
954 			continue;	/* don't unmap */
955 		}
956 
957 		if (ptep_clear_flush_young_notify(vma, address, pte))
958 			continue;
959 
960 		/* Nuke the page table entry. */
961 		flush_cache_page(vma, address, pte_pfn(*pte));
962 		pteval = ptep_clear_flush_notify(vma, address, pte);
963 
964 		/* If nonlinear, store the file page offset in the pte. */
965 		if (page->index != linear_page_index(vma, address))
966 			set_pte_at(mm, address, pte, pgoff_to_pte(page->index));
967 
968 		/* Move the dirty bit to the physical page now the pte is gone. */
969 		if (pte_dirty(pteval))
970 			set_page_dirty(page);
971 
972 		page_remove_rmap(page);
973 		page_cache_release(page);
974 		dec_mm_counter(mm, file_rss);
975 		(*mapcount)--;
976 	}
977 	pte_unmap_unlock(pte - 1, ptl);
978 	if (locked_vma)
979 		up_read(&vma->vm_mm->mmap_sem);
980 	return ret;
981 }
982 
983 /*
984  * common handling for pages mapped in VM_LOCKED vmas
985  */
986 static int try_to_mlock_page(struct page *page, struct vm_area_struct *vma)
987 {
988 	int mlocked = 0;
989 
990 	if (down_read_trylock(&vma->vm_mm->mmap_sem)) {
991 		if (vma->vm_flags & VM_LOCKED) {
992 			mlock_vma_page(page);
993 			mlocked++;	/* really mlocked the page */
994 		}
995 		up_read(&vma->vm_mm->mmap_sem);
996 	}
997 	return mlocked;
998 }
999 
1000 /**
1001  * try_to_unmap_anon - unmap or unlock anonymous page using the object-based
1002  * rmap method
1003  * @page: the page to unmap/unlock
1004  * @unlock:  request for unlock rather than unmap [unlikely]
1005  * @migration:  unmapping for migration - ignored if @unlock
1006  *
1007  * Find all the mappings of a page using the mapping pointer and the vma chains
1008  * contained in the anon_vma struct it points to.
1009  *
1010  * This function is only called from try_to_unmap/try_to_munlock for
1011  * anonymous pages.
1012  * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1013  * where the page was found will be held for write.  So, we won't recheck
1014  * vm_flags for that VMA.  That should be OK, because that vma shouldn't be
1015  * 'LOCKED.
1016  */
1017 static int try_to_unmap_anon(struct page *page, int unlock, int migration)
1018 {
1019 	struct anon_vma *anon_vma;
1020 	struct vm_area_struct *vma;
1021 	unsigned int mlocked = 0;
1022 	int ret = SWAP_AGAIN;
1023 
1024 	if (MLOCK_PAGES && unlikely(unlock))
1025 		ret = SWAP_SUCCESS;	/* default for try_to_munlock() */
1026 
1027 	anon_vma = page_lock_anon_vma(page);
1028 	if (!anon_vma)
1029 		return ret;
1030 
1031 	list_for_each_entry(vma, &anon_vma->head, anon_vma_node) {
1032 		if (MLOCK_PAGES && unlikely(unlock)) {
1033 			if (!((vma->vm_flags & VM_LOCKED) &&
1034 			      page_mapped_in_vma(page, vma)))
1035 				continue;  /* must visit all unlocked vmas */
1036 			ret = SWAP_MLOCK;  /* saw at least one mlocked vma */
1037 		} else {
1038 			ret = try_to_unmap_one(page, vma, migration);
1039 			if (ret == SWAP_FAIL || !page_mapped(page))
1040 				break;
1041 		}
1042 		if (ret == SWAP_MLOCK) {
1043 			mlocked = try_to_mlock_page(page, vma);
1044 			if (mlocked)
1045 				break;	/* stop if actually mlocked page */
1046 		}
1047 	}
1048 
1049 	page_unlock_anon_vma(anon_vma);
1050 
1051 	if (mlocked)
1052 		ret = SWAP_MLOCK;	/* actually mlocked the page */
1053 	else if (ret == SWAP_MLOCK)
1054 		ret = SWAP_AGAIN;	/* saw VM_LOCKED vma */
1055 
1056 	return ret;
1057 }
1058 
1059 /**
1060  * try_to_unmap_file - unmap/unlock file page using the object-based rmap method
1061  * @page: the page to unmap/unlock
1062  * @unlock:  request for unlock rather than unmap [unlikely]
1063  * @migration:  unmapping for migration - ignored if @unlock
1064  *
1065  * Find all the mappings of a page using the mapping pointer and the vma chains
1066  * contained in the address_space struct it points to.
1067  *
1068  * This function is only called from try_to_unmap/try_to_munlock for
1069  * object-based pages.
1070  * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1071  * where the page was found will be held for write.  So, we won't recheck
1072  * vm_flags for that VMA.  That should be OK, because that vma shouldn't be
1073  * 'LOCKED.
1074  */
1075 static int try_to_unmap_file(struct page *page, int unlock, int migration)
1076 {
1077 	struct address_space *mapping = page->mapping;
1078 	pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1079 	struct vm_area_struct *vma;
1080 	struct prio_tree_iter iter;
1081 	int ret = SWAP_AGAIN;
1082 	unsigned long cursor;
1083 	unsigned long max_nl_cursor = 0;
1084 	unsigned long max_nl_size = 0;
1085 	unsigned int mapcount;
1086 	unsigned int mlocked = 0;
1087 
1088 	if (MLOCK_PAGES && unlikely(unlock))
1089 		ret = SWAP_SUCCESS;	/* default for try_to_munlock() */
1090 
1091 	spin_lock(&mapping->i_mmap_lock);
1092 	vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
1093 		if (MLOCK_PAGES && unlikely(unlock)) {
1094 			if (!((vma->vm_flags & VM_LOCKED) &&
1095 						page_mapped_in_vma(page, vma)))
1096 				continue;	/* must visit all vmas */
1097 			ret = SWAP_MLOCK;
1098 		} else {
1099 			ret = try_to_unmap_one(page, vma, migration);
1100 			if (ret == SWAP_FAIL || !page_mapped(page))
1101 				goto out;
1102 		}
1103 		if (ret == SWAP_MLOCK) {
1104 			mlocked = try_to_mlock_page(page, vma);
1105 			if (mlocked)
1106 				break;  /* stop if actually mlocked page */
1107 		}
1108 	}
1109 
1110 	if (mlocked)
1111 		goto out;
1112 
1113 	if (list_empty(&mapping->i_mmap_nonlinear))
1114 		goto out;
1115 
1116 	list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
1117 						shared.vm_set.list) {
1118 		if (MLOCK_PAGES && unlikely(unlock)) {
1119 			if (!(vma->vm_flags & VM_LOCKED))
1120 				continue;	/* must visit all vmas */
1121 			ret = SWAP_MLOCK;	/* leave mlocked == 0 */
1122 			goto out;		/* no need to look further */
1123 		}
1124 		if (!MLOCK_PAGES && !migration && (vma->vm_flags & VM_LOCKED))
1125 			continue;
1126 		cursor = (unsigned long) vma->vm_private_data;
1127 		if (cursor > max_nl_cursor)
1128 			max_nl_cursor = cursor;
1129 		cursor = vma->vm_end - vma->vm_start;
1130 		if (cursor > max_nl_size)
1131 			max_nl_size = cursor;
1132 	}
1133 
1134 	if (max_nl_size == 0) {	/* all nonlinears locked or reserved ? */
1135 		ret = SWAP_FAIL;
1136 		goto out;
1137 	}
1138 
1139 	/*
1140 	 * We don't try to search for this page in the nonlinear vmas,
1141 	 * and page_referenced wouldn't have found it anyway.  Instead
1142 	 * just walk the nonlinear vmas trying to age and unmap some.
1143 	 * The mapcount of the page we came in with is irrelevant,
1144 	 * but even so use it as a guide to how hard we should try?
1145 	 */
1146 	mapcount = page_mapcount(page);
1147 	if (!mapcount)
1148 		goto out;
1149 	cond_resched_lock(&mapping->i_mmap_lock);
1150 
1151 	max_nl_size = (max_nl_size + CLUSTER_SIZE - 1) & CLUSTER_MASK;
1152 	if (max_nl_cursor == 0)
1153 		max_nl_cursor = CLUSTER_SIZE;
1154 
1155 	do {
1156 		list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
1157 						shared.vm_set.list) {
1158 			if (!MLOCK_PAGES && !migration &&
1159 			    (vma->vm_flags & VM_LOCKED))
1160 				continue;
1161 			cursor = (unsigned long) vma->vm_private_data;
1162 			while ( cursor < max_nl_cursor &&
1163 				cursor < vma->vm_end - vma->vm_start) {
1164 				ret = try_to_unmap_cluster(cursor, &mapcount,
1165 								vma, page);
1166 				if (ret == SWAP_MLOCK)
1167 					mlocked = 2;	/* to return below */
1168 				cursor += CLUSTER_SIZE;
1169 				vma->vm_private_data = (void *) cursor;
1170 				if ((int)mapcount <= 0)
1171 					goto out;
1172 			}
1173 			vma->vm_private_data = (void *) max_nl_cursor;
1174 		}
1175 		cond_resched_lock(&mapping->i_mmap_lock);
1176 		max_nl_cursor += CLUSTER_SIZE;
1177 	} while (max_nl_cursor <= max_nl_size);
1178 
1179 	/*
1180 	 * Don't loop forever (perhaps all the remaining pages are
1181 	 * in locked vmas).  Reset cursor on all unreserved nonlinear
1182 	 * vmas, now forgetting on which ones it had fallen behind.
1183 	 */
1184 	list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list)
1185 		vma->vm_private_data = NULL;
1186 out:
1187 	spin_unlock(&mapping->i_mmap_lock);
1188 	if (mlocked)
1189 		ret = SWAP_MLOCK;	/* actually mlocked the page */
1190 	else if (ret == SWAP_MLOCK)
1191 		ret = SWAP_AGAIN;	/* saw VM_LOCKED vma */
1192 	return ret;
1193 }
1194 
1195 /**
1196  * try_to_unmap - try to remove all page table mappings to a page
1197  * @page: the page to get unmapped
1198  * @migration: migration flag
1199  *
1200  * Tries to remove all the page table entries which are mapping this
1201  * page, used in the pageout path.  Caller must hold the page lock.
1202  * Return values are:
1203  *
1204  * SWAP_SUCCESS	- we succeeded in removing all mappings
1205  * SWAP_AGAIN	- we missed a mapping, try again later
1206  * SWAP_FAIL	- the page is unswappable
1207  * SWAP_MLOCK	- page is mlocked.
1208  */
1209 int try_to_unmap(struct page *page, int migration)
1210 {
1211 	int ret;
1212 
1213 	BUG_ON(!PageLocked(page));
1214 
1215 	if (PageAnon(page))
1216 		ret = try_to_unmap_anon(page, 0, migration);
1217 	else
1218 		ret = try_to_unmap_file(page, 0, migration);
1219 	if (ret != SWAP_MLOCK && !page_mapped(page))
1220 		ret = SWAP_SUCCESS;
1221 	return ret;
1222 }
1223 
1224 /**
1225  * try_to_munlock - try to munlock a page
1226  * @page: the page to be munlocked
1227  *
1228  * Called from munlock code.  Checks all of the VMAs mapping the page
1229  * to make sure nobody else has this page mlocked. The page will be
1230  * returned with PG_mlocked cleared if no other vmas have it mlocked.
1231  *
1232  * Return values are:
1233  *
1234  * SWAP_SUCCESS	- no vma's holding page mlocked.
1235  * SWAP_AGAIN	- page mapped in mlocked vma -- couldn't acquire mmap sem
1236  * SWAP_MLOCK	- page is now mlocked.
1237  */
1238 int try_to_munlock(struct page *page)
1239 {
1240 	VM_BUG_ON(!PageLocked(page) || PageLRU(page));
1241 
1242 	if (PageAnon(page))
1243 		return try_to_unmap_anon(page, 1, 0);
1244 	else
1245 		return try_to_unmap_file(page, 1, 0);
1246 }
1247 
1248