xref: /openbmc/linux/mm/rmap.c (revision e868d61272caa648214046a096e5a6bfc068dc8c)
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 <hugh@veritas.com> 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/kallsyms.h>
51 
52 #include <asm/tlbflush.h>
53 
54 struct kmem_cache *anon_vma_cachep;
55 
56 static inline void validate_anon_vma(struct vm_area_struct *find_vma)
57 {
58 #ifdef CONFIG_DEBUG_VM
59 	struct anon_vma *anon_vma = find_vma->anon_vma;
60 	struct vm_area_struct *vma;
61 	unsigned int mapcount = 0;
62 	int found = 0;
63 
64 	list_for_each_entry(vma, &anon_vma->head, anon_vma_node) {
65 		mapcount++;
66 		BUG_ON(mapcount > 100000);
67 		if (vma == find_vma)
68 			found = 1;
69 	}
70 	BUG_ON(!found);
71 #endif
72 }
73 
74 /* This must be called under the mmap_sem. */
75 int anon_vma_prepare(struct vm_area_struct *vma)
76 {
77 	struct anon_vma *anon_vma = vma->anon_vma;
78 
79 	might_sleep();
80 	if (unlikely(!anon_vma)) {
81 		struct mm_struct *mm = vma->vm_mm;
82 		struct anon_vma *allocated, *locked;
83 
84 		anon_vma = find_mergeable_anon_vma(vma);
85 		if (anon_vma) {
86 			allocated = NULL;
87 			locked = anon_vma;
88 			spin_lock(&locked->lock);
89 		} else {
90 			anon_vma = anon_vma_alloc();
91 			if (unlikely(!anon_vma))
92 				return -ENOMEM;
93 			allocated = anon_vma;
94 			locked = NULL;
95 		}
96 
97 		/* page_table_lock to protect against threads */
98 		spin_lock(&mm->page_table_lock);
99 		if (likely(!vma->anon_vma)) {
100 			vma->anon_vma = anon_vma;
101 			list_add_tail(&vma->anon_vma_node, &anon_vma->head);
102 			allocated = NULL;
103 		}
104 		spin_unlock(&mm->page_table_lock);
105 
106 		if (locked)
107 			spin_unlock(&locked->lock);
108 		if (unlikely(allocated))
109 			anon_vma_free(allocated);
110 	}
111 	return 0;
112 }
113 
114 void __anon_vma_merge(struct vm_area_struct *vma, struct vm_area_struct *next)
115 {
116 	BUG_ON(vma->anon_vma != next->anon_vma);
117 	list_del(&next->anon_vma_node);
118 }
119 
120 void __anon_vma_link(struct vm_area_struct *vma)
121 {
122 	struct anon_vma *anon_vma = vma->anon_vma;
123 
124 	if (anon_vma) {
125 		list_add_tail(&vma->anon_vma_node, &anon_vma->head);
126 		validate_anon_vma(vma);
127 	}
128 }
129 
130 void anon_vma_link(struct vm_area_struct *vma)
131 {
132 	struct anon_vma *anon_vma = vma->anon_vma;
133 
134 	if (anon_vma) {
135 		spin_lock(&anon_vma->lock);
136 		list_add_tail(&vma->anon_vma_node, &anon_vma->head);
137 		validate_anon_vma(vma);
138 		spin_unlock(&anon_vma->lock);
139 	}
140 }
141 
142 void anon_vma_unlink(struct vm_area_struct *vma)
143 {
144 	struct anon_vma *anon_vma = vma->anon_vma;
145 	int empty;
146 
147 	if (!anon_vma)
148 		return;
149 
150 	spin_lock(&anon_vma->lock);
151 	validate_anon_vma(vma);
152 	list_del(&vma->anon_vma_node);
153 
154 	/* We must garbage collect the anon_vma if it's empty */
155 	empty = list_empty(&anon_vma->head);
156 	spin_unlock(&anon_vma->lock);
157 
158 	if (empty)
159 		anon_vma_free(anon_vma);
160 }
161 
162 static void anon_vma_ctor(void *data, struct kmem_cache *cachep,
163 			  unsigned long flags)
164 {
165 	if (flags & SLAB_CTOR_CONSTRUCTOR) {
166 		struct anon_vma *anon_vma = data;
167 
168 		spin_lock_init(&anon_vma->lock);
169 		INIT_LIST_HEAD(&anon_vma->head);
170 	}
171 }
172 
173 void __init anon_vma_init(void)
174 {
175 	anon_vma_cachep = kmem_cache_create("anon_vma", sizeof(struct anon_vma),
176 			0, SLAB_DESTROY_BY_RCU|SLAB_PANIC, anon_vma_ctor, NULL);
177 }
178 
179 /*
180  * Getting a lock on a stable anon_vma from a page off the LRU is
181  * tricky: page_lock_anon_vma rely on RCU to guard against the races.
182  */
183 static struct anon_vma *page_lock_anon_vma(struct page *page)
184 {
185 	struct anon_vma *anon_vma;
186 	unsigned long anon_mapping;
187 
188 	rcu_read_lock();
189 	anon_mapping = (unsigned long) page->mapping;
190 	if (!(anon_mapping & PAGE_MAPPING_ANON))
191 		goto out;
192 	if (!page_mapped(page))
193 		goto out;
194 
195 	anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
196 	spin_lock(&anon_vma->lock);
197 	return anon_vma;
198 out:
199 	rcu_read_unlock();
200 	return NULL;
201 }
202 
203 static void page_unlock_anon_vma(struct anon_vma *anon_vma)
204 {
205 	spin_unlock(&anon_vma->lock);
206 	rcu_read_unlock();
207 }
208 
209 /*
210  * At what user virtual address is page expected in vma?
211  */
212 static inline unsigned long
213 vma_address(struct page *page, struct vm_area_struct *vma)
214 {
215 	pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
216 	unsigned long address;
217 
218 	address = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
219 	if (unlikely(address < vma->vm_start || address >= vma->vm_end)) {
220 		/* page should be within any vma from prio_tree_next */
221 		BUG_ON(!PageAnon(page));
222 		return -EFAULT;
223 	}
224 	return address;
225 }
226 
227 /*
228  * At what user virtual address is page expected in vma? checking that the
229  * page matches the vma: currently only used on anon pages, by unuse_vma;
230  */
231 unsigned long page_address_in_vma(struct page *page, struct vm_area_struct *vma)
232 {
233 	if (PageAnon(page)) {
234 		if ((void *)vma->anon_vma !=
235 		    (void *)page->mapping - PAGE_MAPPING_ANON)
236 			return -EFAULT;
237 	} else if (page->mapping && !(vma->vm_flags & VM_NONLINEAR)) {
238 		if (!vma->vm_file ||
239 		    vma->vm_file->f_mapping != page->mapping)
240 			return -EFAULT;
241 	} else
242 		return -EFAULT;
243 	return vma_address(page, vma);
244 }
245 
246 /*
247  * Check that @page is mapped at @address into @mm.
248  *
249  * On success returns with pte mapped and locked.
250  */
251 pte_t *page_check_address(struct page *page, struct mm_struct *mm,
252 			  unsigned long address, spinlock_t **ptlp)
253 {
254 	pgd_t *pgd;
255 	pud_t *pud;
256 	pmd_t *pmd;
257 	pte_t *pte;
258 	spinlock_t *ptl;
259 
260 	pgd = pgd_offset(mm, address);
261 	if (!pgd_present(*pgd))
262 		return NULL;
263 
264 	pud = pud_offset(pgd, address);
265 	if (!pud_present(*pud))
266 		return NULL;
267 
268 	pmd = pmd_offset(pud, address);
269 	if (!pmd_present(*pmd))
270 		return NULL;
271 
272 	pte = pte_offset_map(pmd, address);
273 	/* Make a quick check before getting the lock */
274 	if (!pte_present(*pte)) {
275 		pte_unmap(pte);
276 		return NULL;
277 	}
278 
279 	ptl = pte_lockptr(mm, pmd);
280 	spin_lock(ptl);
281 	if (pte_present(*pte) && page_to_pfn(page) == pte_pfn(*pte)) {
282 		*ptlp = ptl;
283 		return pte;
284 	}
285 	pte_unmap_unlock(pte, ptl);
286 	return NULL;
287 }
288 
289 /*
290  * Subfunctions of page_referenced: page_referenced_one called
291  * repeatedly from either page_referenced_anon or page_referenced_file.
292  */
293 static int page_referenced_one(struct page *page,
294 	struct vm_area_struct *vma, unsigned int *mapcount)
295 {
296 	struct mm_struct *mm = vma->vm_mm;
297 	unsigned long address;
298 	pte_t *pte;
299 	spinlock_t *ptl;
300 	int referenced = 0;
301 
302 	address = vma_address(page, vma);
303 	if (address == -EFAULT)
304 		goto out;
305 
306 	pte = page_check_address(page, mm, address, &ptl);
307 	if (!pte)
308 		goto out;
309 
310 	if (ptep_clear_flush_young(vma, address, pte))
311 		referenced++;
312 
313 	/* Pretend the page is referenced if the task has the
314 	   swap token and is in the middle of a page fault. */
315 	if (mm != current->mm && has_swap_token(mm) &&
316 			rwsem_is_locked(&mm->mmap_sem))
317 		referenced++;
318 
319 	(*mapcount)--;
320 	pte_unmap_unlock(pte, ptl);
321 out:
322 	return referenced;
323 }
324 
325 static int page_referenced_anon(struct page *page)
326 {
327 	unsigned int mapcount;
328 	struct anon_vma *anon_vma;
329 	struct vm_area_struct *vma;
330 	int referenced = 0;
331 
332 	anon_vma = page_lock_anon_vma(page);
333 	if (!anon_vma)
334 		return referenced;
335 
336 	mapcount = page_mapcount(page);
337 	list_for_each_entry(vma, &anon_vma->head, anon_vma_node) {
338 		referenced += page_referenced_one(page, vma, &mapcount);
339 		if (!mapcount)
340 			break;
341 	}
342 
343 	page_unlock_anon_vma(anon_vma);
344 	return referenced;
345 }
346 
347 /**
348  * page_referenced_file - referenced check for object-based rmap
349  * @page: the page we're checking references on.
350  *
351  * For an object-based mapped page, find all the places it is mapped and
352  * check/clear the referenced flag.  This is done by following the page->mapping
353  * pointer, then walking the chain of vmas it holds.  It returns the number
354  * of references it found.
355  *
356  * This function is only called from page_referenced for object-based pages.
357  */
358 static int page_referenced_file(struct page *page)
359 {
360 	unsigned int mapcount;
361 	struct address_space *mapping = page->mapping;
362 	pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
363 	struct vm_area_struct *vma;
364 	struct prio_tree_iter iter;
365 	int referenced = 0;
366 
367 	/*
368 	 * The caller's checks on page->mapping and !PageAnon have made
369 	 * sure that this is a file page: the check for page->mapping
370 	 * excludes the case just before it gets set on an anon page.
371 	 */
372 	BUG_ON(PageAnon(page));
373 
374 	/*
375 	 * The page lock not only makes sure that page->mapping cannot
376 	 * suddenly be NULLified by truncation, it makes sure that the
377 	 * structure at mapping cannot be freed and reused yet,
378 	 * so we can safely take mapping->i_mmap_lock.
379 	 */
380 	BUG_ON(!PageLocked(page));
381 
382 	spin_lock(&mapping->i_mmap_lock);
383 
384 	/*
385 	 * i_mmap_lock does not stabilize mapcount at all, but mapcount
386 	 * is more likely to be accurate if we note it after spinning.
387 	 */
388 	mapcount = page_mapcount(page);
389 
390 	vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
391 		if ((vma->vm_flags & (VM_LOCKED|VM_MAYSHARE))
392 				  == (VM_LOCKED|VM_MAYSHARE)) {
393 			referenced++;
394 			break;
395 		}
396 		referenced += page_referenced_one(page, vma, &mapcount);
397 		if (!mapcount)
398 			break;
399 	}
400 
401 	spin_unlock(&mapping->i_mmap_lock);
402 	return referenced;
403 }
404 
405 /**
406  * page_referenced - test if the page was referenced
407  * @page: the page to test
408  * @is_locked: caller holds lock on the page
409  *
410  * Quick test_and_clear_referenced for all mappings to a page,
411  * returns the number of ptes which referenced the page.
412  */
413 int page_referenced(struct page *page, int is_locked)
414 {
415 	int referenced = 0;
416 
417 	if (page_test_and_clear_young(page))
418 		referenced++;
419 
420 	if (TestClearPageReferenced(page))
421 		referenced++;
422 
423 	if (page_mapped(page) && page->mapping) {
424 		if (PageAnon(page))
425 			referenced += page_referenced_anon(page);
426 		else if (is_locked)
427 			referenced += page_referenced_file(page);
428 		else if (TestSetPageLocked(page))
429 			referenced++;
430 		else {
431 			if (page->mapping)
432 				referenced += page_referenced_file(page);
433 			unlock_page(page);
434 		}
435 	}
436 	return referenced;
437 }
438 
439 static int page_mkclean_one(struct page *page, struct vm_area_struct *vma)
440 {
441 	struct mm_struct *mm = vma->vm_mm;
442 	unsigned long address;
443 	pte_t *pte;
444 	spinlock_t *ptl;
445 	int ret = 0;
446 
447 	address = vma_address(page, vma);
448 	if (address == -EFAULT)
449 		goto out;
450 
451 	pte = page_check_address(page, mm, address, &ptl);
452 	if (!pte)
453 		goto out;
454 
455 	if (pte_dirty(*pte) || pte_write(*pte)) {
456 		pte_t entry;
457 
458 		flush_cache_page(vma, address, pte_pfn(*pte));
459 		entry = ptep_clear_flush(vma, address, pte);
460 		entry = pte_wrprotect(entry);
461 		entry = pte_mkclean(entry);
462 		set_pte_at(mm, address, pte, entry);
463 		lazy_mmu_prot_update(entry);
464 		ret = 1;
465 	}
466 
467 	pte_unmap_unlock(pte, ptl);
468 out:
469 	return ret;
470 }
471 
472 static int page_mkclean_file(struct address_space *mapping, struct page *page)
473 {
474 	pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
475 	struct vm_area_struct *vma;
476 	struct prio_tree_iter iter;
477 	int ret = 0;
478 
479 	BUG_ON(PageAnon(page));
480 
481 	spin_lock(&mapping->i_mmap_lock);
482 	vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
483 		if (vma->vm_flags & VM_SHARED)
484 			ret += page_mkclean_one(page, vma);
485 	}
486 	spin_unlock(&mapping->i_mmap_lock);
487 	return ret;
488 }
489 
490 int page_mkclean(struct page *page)
491 {
492 	int ret = 0;
493 
494 	BUG_ON(!PageLocked(page));
495 
496 	if (page_mapped(page)) {
497 		struct address_space *mapping = page_mapping(page);
498 		if (mapping)
499 			ret = page_mkclean_file(mapping, page);
500 		if (page_test_dirty(page)) {
501 			page_clear_dirty(page);
502 			ret = 1;
503 		}
504 	}
505 
506 	return ret;
507 }
508 EXPORT_SYMBOL_GPL(page_mkclean);
509 
510 /**
511  * page_set_anon_rmap - setup new anonymous rmap
512  * @page:	the page to add the mapping to
513  * @vma:	the vm area in which the mapping is added
514  * @address:	the user virtual address mapped
515  */
516 static void __page_set_anon_rmap(struct page *page,
517 	struct vm_area_struct *vma, unsigned long address)
518 {
519 	struct anon_vma *anon_vma = vma->anon_vma;
520 
521 	BUG_ON(!anon_vma);
522 	anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
523 	page->mapping = (struct address_space *) anon_vma;
524 
525 	page->index = linear_page_index(vma, address);
526 
527 	/*
528 	 * nr_mapped state can be updated without turning off
529 	 * interrupts because it is not modified via interrupt.
530 	 */
531 	__inc_zone_page_state(page, NR_ANON_PAGES);
532 }
533 
534 /**
535  * page_add_anon_rmap - add pte mapping to an anonymous page
536  * @page:	the page to add the mapping to
537  * @vma:	the vm area in which the mapping is added
538  * @address:	the user virtual address mapped
539  *
540  * The caller needs to hold the pte lock.
541  */
542 void page_add_anon_rmap(struct page *page,
543 	struct vm_area_struct *vma, unsigned long address)
544 {
545 	if (atomic_inc_and_test(&page->_mapcount))
546 		__page_set_anon_rmap(page, vma, address);
547 	/* else checking page index and mapping is racy */
548 }
549 
550 /*
551  * page_add_new_anon_rmap - add pte mapping to a new anonymous page
552  * @page:	the page to add the mapping to
553  * @vma:	the vm area in which the mapping is added
554  * @address:	the user virtual address mapped
555  *
556  * Same as page_add_anon_rmap but must only be called on *new* pages.
557  * This means the inc-and-test can be bypassed.
558  */
559 void page_add_new_anon_rmap(struct page *page,
560 	struct vm_area_struct *vma, unsigned long address)
561 {
562 	atomic_set(&page->_mapcount, 0); /* elevate count by 1 (starts at -1) */
563 	__page_set_anon_rmap(page, vma, address);
564 }
565 
566 /**
567  * page_add_file_rmap - add pte mapping to a file page
568  * @page: the page to add the mapping to
569  *
570  * The caller needs to hold the pte lock.
571  */
572 void page_add_file_rmap(struct page *page)
573 {
574 	if (atomic_inc_and_test(&page->_mapcount))
575 		__inc_zone_page_state(page, NR_FILE_MAPPED);
576 }
577 
578 /**
579  * page_remove_rmap - take down pte mapping from a page
580  * @page: page to remove mapping from
581  *
582  * The caller needs to hold the pte lock.
583  */
584 void page_remove_rmap(struct page *page, struct vm_area_struct *vma)
585 {
586 	if (atomic_add_negative(-1, &page->_mapcount)) {
587 		if (unlikely(page_mapcount(page) < 0)) {
588 			printk (KERN_EMERG "Eeek! page_mapcount(page) went negative! (%d)\n", page_mapcount(page));
589 			printk (KERN_EMERG "  page pfn = %lx\n", page_to_pfn(page));
590 			printk (KERN_EMERG "  page->flags = %lx\n", page->flags);
591 			printk (KERN_EMERG "  page->count = %x\n", page_count(page));
592 			printk (KERN_EMERG "  page->mapping = %p\n", page->mapping);
593 			print_symbol (KERN_EMERG "  vma->vm_ops = %s\n", (unsigned long)vma->vm_ops);
594 			if (vma->vm_ops)
595 				print_symbol (KERN_EMERG "  vma->vm_ops->nopage = %s\n", (unsigned long)vma->vm_ops->nopage);
596 			if (vma->vm_file && vma->vm_file->f_op)
597 				print_symbol (KERN_EMERG "  vma->vm_file->f_op->mmap = %s\n", (unsigned long)vma->vm_file->f_op->mmap);
598 			BUG();
599 		}
600 
601 		/*
602 		 * It would be tidy to reset the PageAnon mapping here,
603 		 * but that might overwrite a racing page_add_anon_rmap
604 		 * which increments mapcount after us but sets mapping
605 		 * before us: so leave the reset to free_hot_cold_page,
606 		 * and remember that it's only reliable while mapped.
607 		 * Leaving it set also helps swapoff to reinstate ptes
608 		 * faster for those pages still in swapcache.
609 		 */
610 		if (page_test_dirty(page)) {
611 			page_clear_dirty(page);
612 			set_page_dirty(page);
613 		}
614 		__dec_zone_page_state(page,
615 				PageAnon(page) ? NR_ANON_PAGES : NR_FILE_MAPPED);
616 	}
617 }
618 
619 /*
620  * Subfunctions of try_to_unmap: try_to_unmap_one called
621  * repeatedly from either try_to_unmap_anon or try_to_unmap_file.
622  */
623 static int try_to_unmap_one(struct page *page, struct vm_area_struct *vma,
624 				int migration)
625 {
626 	struct mm_struct *mm = vma->vm_mm;
627 	unsigned long address;
628 	pte_t *pte;
629 	pte_t pteval;
630 	spinlock_t *ptl;
631 	int ret = SWAP_AGAIN;
632 
633 	address = vma_address(page, vma);
634 	if (address == -EFAULT)
635 		goto out;
636 
637 	pte = page_check_address(page, mm, address, &ptl);
638 	if (!pte)
639 		goto out;
640 
641 	/*
642 	 * If the page is mlock()d, we cannot swap it out.
643 	 * If it's recently referenced (perhaps page_referenced
644 	 * skipped over this mm) then we should reactivate it.
645 	 */
646 	if (!migration && ((vma->vm_flags & VM_LOCKED) ||
647 			(ptep_clear_flush_young(vma, address, pte)))) {
648 		ret = SWAP_FAIL;
649 		goto out_unmap;
650 	}
651 
652 	/* Nuke the page table entry. */
653 	flush_cache_page(vma, address, page_to_pfn(page));
654 	pteval = ptep_clear_flush(vma, address, pte);
655 
656 	/* Move the dirty bit to the physical page now the pte is gone. */
657 	if (pte_dirty(pteval))
658 		set_page_dirty(page);
659 
660 	/* Update high watermark before we lower rss */
661 	update_hiwater_rss(mm);
662 
663 	if (PageAnon(page)) {
664 		swp_entry_t entry = { .val = page_private(page) };
665 
666 		if (PageSwapCache(page)) {
667 			/*
668 			 * Store the swap location in the pte.
669 			 * See handle_pte_fault() ...
670 			 */
671 			swap_duplicate(entry);
672 			if (list_empty(&mm->mmlist)) {
673 				spin_lock(&mmlist_lock);
674 				if (list_empty(&mm->mmlist))
675 					list_add(&mm->mmlist, &init_mm.mmlist);
676 				spin_unlock(&mmlist_lock);
677 			}
678 			dec_mm_counter(mm, anon_rss);
679 #ifdef CONFIG_MIGRATION
680 		} else {
681 			/*
682 			 * Store the pfn of the page in a special migration
683 			 * pte. do_swap_page() will wait until the migration
684 			 * pte is removed and then restart fault handling.
685 			 */
686 			BUG_ON(!migration);
687 			entry = make_migration_entry(page, pte_write(pteval));
688 #endif
689 		}
690 		set_pte_at(mm, address, pte, swp_entry_to_pte(entry));
691 		BUG_ON(pte_file(*pte));
692 	} else
693 #ifdef CONFIG_MIGRATION
694 	if (migration) {
695 		/* Establish migration entry for a file page */
696 		swp_entry_t entry;
697 		entry = make_migration_entry(page, pte_write(pteval));
698 		set_pte_at(mm, address, pte, swp_entry_to_pte(entry));
699 	} else
700 #endif
701 		dec_mm_counter(mm, file_rss);
702 
703 
704 	page_remove_rmap(page, vma);
705 	page_cache_release(page);
706 
707 out_unmap:
708 	pte_unmap_unlock(pte, ptl);
709 out:
710 	return ret;
711 }
712 
713 /*
714  * objrmap doesn't work for nonlinear VMAs because the assumption that
715  * offset-into-file correlates with offset-into-virtual-addresses does not hold.
716  * Consequently, given a particular page and its ->index, we cannot locate the
717  * ptes which are mapping that page without an exhaustive linear search.
718  *
719  * So what this code does is a mini "virtual scan" of each nonlinear VMA which
720  * maps the file to which the target page belongs.  The ->vm_private_data field
721  * holds the current cursor into that scan.  Successive searches will circulate
722  * around the vma's virtual address space.
723  *
724  * So as more replacement pressure is applied to the pages in a nonlinear VMA,
725  * more scanning pressure is placed against them as well.   Eventually pages
726  * will become fully unmapped and are eligible for eviction.
727  *
728  * For very sparsely populated VMAs this is a little inefficient - chances are
729  * there there won't be many ptes located within the scan cluster.  In this case
730  * maybe we could scan further - to the end of the pte page, perhaps.
731  */
732 #define CLUSTER_SIZE	min(32*PAGE_SIZE, PMD_SIZE)
733 #define CLUSTER_MASK	(~(CLUSTER_SIZE - 1))
734 
735 static void try_to_unmap_cluster(unsigned long cursor,
736 	unsigned int *mapcount, struct vm_area_struct *vma)
737 {
738 	struct mm_struct *mm = vma->vm_mm;
739 	pgd_t *pgd;
740 	pud_t *pud;
741 	pmd_t *pmd;
742 	pte_t *pte;
743 	pte_t pteval;
744 	spinlock_t *ptl;
745 	struct page *page;
746 	unsigned long address;
747 	unsigned long end;
748 
749 	address = (vma->vm_start + cursor) & CLUSTER_MASK;
750 	end = address + CLUSTER_SIZE;
751 	if (address < vma->vm_start)
752 		address = vma->vm_start;
753 	if (end > vma->vm_end)
754 		end = vma->vm_end;
755 
756 	pgd = pgd_offset(mm, address);
757 	if (!pgd_present(*pgd))
758 		return;
759 
760 	pud = pud_offset(pgd, address);
761 	if (!pud_present(*pud))
762 		return;
763 
764 	pmd = pmd_offset(pud, address);
765 	if (!pmd_present(*pmd))
766 		return;
767 
768 	pte = pte_offset_map_lock(mm, pmd, address, &ptl);
769 
770 	/* Update high watermark before we lower rss */
771 	update_hiwater_rss(mm);
772 
773 	for (; address < end; pte++, address += PAGE_SIZE) {
774 		if (!pte_present(*pte))
775 			continue;
776 		page = vm_normal_page(vma, address, *pte);
777 		BUG_ON(!page || PageAnon(page));
778 
779 		if (ptep_clear_flush_young(vma, address, pte))
780 			continue;
781 
782 		/* Nuke the page table entry. */
783 		flush_cache_page(vma, address, pte_pfn(*pte));
784 		pteval = ptep_clear_flush(vma, address, pte);
785 
786 		/* If nonlinear, store the file page offset in the pte. */
787 		if (page->index != linear_page_index(vma, address))
788 			set_pte_at(mm, address, pte, pgoff_to_pte(page->index));
789 
790 		/* Move the dirty bit to the physical page now the pte is gone. */
791 		if (pte_dirty(pteval))
792 			set_page_dirty(page);
793 
794 		page_remove_rmap(page, vma);
795 		page_cache_release(page);
796 		dec_mm_counter(mm, file_rss);
797 		(*mapcount)--;
798 	}
799 	pte_unmap_unlock(pte - 1, ptl);
800 }
801 
802 static int try_to_unmap_anon(struct page *page, int migration)
803 {
804 	struct anon_vma *anon_vma;
805 	struct vm_area_struct *vma;
806 	int ret = SWAP_AGAIN;
807 
808 	anon_vma = page_lock_anon_vma(page);
809 	if (!anon_vma)
810 		return ret;
811 
812 	list_for_each_entry(vma, &anon_vma->head, anon_vma_node) {
813 		ret = try_to_unmap_one(page, vma, migration);
814 		if (ret == SWAP_FAIL || !page_mapped(page))
815 			break;
816 	}
817 
818 	page_unlock_anon_vma(anon_vma);
819 	return ret;
820 }
821 
822 /**
823  * try_to_unmap_file - unmap file page using the object-based rmap method
824  * @page: the page to unmap
825  *
826  * Find all the mappings of a page using the mapping pointer and the vma chains
827  * contained in the address_space struct it points to.
828  *
829  * This function is only called from try_to_unmap for object-based pages.
830  */
831 static int try_to_unmap_file(struct page *page, int migration)
832 {
833 	struct address_space *mapping = page->mapping;
834 	pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
835 	struct vm_area_struct *vma;
836 	struct prio_tree_iter iter;
837 	int ret = SWAP_AGAIN;
838 	unsigned long cursor;
839 	unsigned long max_nl_cursor = 0;
840 	unsigned long max_nl_size = 0;
841 	unsigned int mapcount;
842 
843 	spin_lock(&mapping->i_mmap_lock);
844 	vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
845 		ret = try_to_unmap_one(page, vma, migration);
846 		if (ret == SWAP_FAIL || !page_mapped(page))
847 			goto out;
848 	}
849 
850 	if (list_empty(&mapping->i_mmap_nonlinear))
851 		goto out;
852 
853 	list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
854 						shared.vm_set.list) {
855 		if ((vma->vm_flags & VM_LOCKED) && !migration)
856 			continue;
857 		cursor = (unsigned long) vma->vm_private_data;
858 		if (cursor > max_nl_cursor)
859 			max_nl_cursor = cursor;
860 		cursor = vma->vm_end - vma->vm_start;
861 		if (cursor > max_nl_size)
862 			max_nl_size = cursor;
863 	}
864 
865 	if (max_nl_size == 0) {	/* any nonlinears locked or reserved */
866 		ret = SWAP_FAIL;
867 		goto out;
868 	}
869 
870 	/*
871 	 * We don't try to search for this page in the nonlinear vmas,
872 	 * and page_referenced wouldn't have found it anyway.  Instead
873 	 * just walk the nonlinear vmas trying to age and unmap some.
874 	 * The mapcount of the page we came in with is irrelevant,
875 	 * but even so use it as a guide to how hard we should try?
876 	 */
877 	mapcount = page_mapcount(page);
878 	if (!mapcount)
879 		goto out;
880 	cond_resched_lock(&mapping->i_mmap_lock);
881 
882 	max_nl_size = (max_nl_size + CLUSTER_SIZE - 1) & CLUSTER_MASK;
883 	if (max_nl_cursor == 0)
884 		max_nl_cursor = CLUSTER_SIZE;
885 
886 	do {
887 		list_for_each_entry(vma, &mapping->i_mmap_nonlinear,
888 						shared.vm_set.list) {
889 			if ((vma->vm_flags & VM_LOCKED) && !migration)
890 				continue;
891 			cursor = (unsigned long) vma->vm_private_data;
892 			while ( cursor < max_nl_cursor &&
893 				cursor < vma->vm_end - vma->vm_start) {
894 				try_to_unmap_cluster(cursor, &mapcount, vma);
895 				cursor += CLUSTER_SIZE;
896 				vma->vm_private_data = (void *) cursor;
897 				if ((int)mapcount <= 0)
898 					goto out;
899 			}
900 			vma->vm_private_data = (void *) max_nl_cursor;
901 		}
902 		cond_resched_lock(&mapping->i_mmap_lock);
903 		max_nl_cursor += CLUSTER_SIZE;
904 	} while (max_nl_cursor <= max_nl_size);
905 
906 	/*
907 	 * Don't loop forever (perhaps all the remaining pages are
908 	 * in locked vmas).  Reset cursor on all unreserved nonlinear
909 	 * vmas, now forgetting on which ones it had fallen behind.
910 	 */
911 	list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list)
912 		vma->vm_private_data = NULL;
913 out:
914 	spin_unlock(&mapping->i_mmap_lock);
915 	return ret;
916 }
917 
918 /**
919  * try_to_unmap - try to remove all page table mappings to a page
920  * @page: the page to get unmapped
921  *
922  * Tries to remove all the page table entries which are mapping this
923  * page, used in the pageout path.  Caller must hold the page lock.
924  * Return values are:
925  *
926  * SWAP_SUCCESS	- we succeeded in removing all mappings
927  * SWAP_AGAIN	- we missed a mapping, try again later
928  * SWAP_FAIL	- the page is unswappable
929  */
930 int try_to_unmap(struct page *page, int migration)
931 {
932 	int ret;
933 
934 	BUG_ON(!PageLocked(page));
935 
936 	if (PageAnon(page))
937 		ret = try_to_unmap_anon(page, migration);
938 	else
939 		ret = try_to_unmap_file(page, migration);
940 
941 	if (!page_mapped(page))
942 		ret = SWAP_SUCCESS;
943 	return ret;
944 }
945 
946