xref: /openbmc/linux/mm/rmap.c (revision 6774def6)
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  *   mm->mmap_sem
25  *     page->flags PG_locked (lock_page)
26  *       mapping->i_mmap_mutex
27  *         anon_vma->rwsem
28  *           mm->page_table_lock or pte_lock
29  *             zone->lru_lock (in mark_page_accessed, isolate_lru_page)
30  *             swap_lock (in swap_duplicate, swap_info_get)
31  *               mmlist_lock (in mmput, drain_mmlist and others)
32  *               mapping->private_lock (in __set_page_dirty_buffers)
33  *               inode->i_lock (in set_page_dirty's __mark_inode_dirty)
34  *               bdi.wb->list_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 bdi.wb->list_lock in __sync_single_inode)
39  *
40  * anon_vma->rwsem,mapping->i_mutex      (memory_failure, collect_procs_anon)
41  *   ->tasklist_lock
42  *     pte map lock
43  */
44 
45 #include <linux/mm.h>
46 #include <linux/pagemap.h>
47 #include <linux/swap.h>
48 #include <linux/swapops.h>
49 #include <linux/slab.h>
50 #include <linux/init.h>
51 #include <linux/ksm.h>
52 #include <linux/rmap.h>
53 #include <linux/rcupdate.h>
54 #include <linux/export.h>
55 #include <linux/memcontrol.h>
56 #include <linux/mmu_notifier.h>
57 #include <linux/migrate.h>
58 #include <linux/hugetlb.h>
59 #include <linux/backing-dev.h>
60 
61 #include <asm/tlbflush.h>
62 
63 #include "internal.h"
64 
65 static struct kmem_cache *anon_vma_cachep;
66 static struct kmem_cache *anon_vma_chain_cachep;
67 
68 static inline struct anon_vma *anon_vma_alloc(void)
69 {
70 	struct anon_vma *anon_vma;
71 
72 	anon_vma = kmem_cache_alloc(anon_vma_cachep, GFP_KERNEL);
73 	if (anon_vma) {
74 		atomic_set(&anon_vma->refcount, 1);
75 		/*
76 		 * Initialise the anon_vma root to point to itself. If called
77 		 * from fork, the root will be reset to the parents anon_vma.
78 		 */
79 		anon_vma->root = anon_vma;
80 	}
81 
82 	return anon_vma;
83 }
84 
85 static inline void anon_vma_free(struct anon_vma *anon_vma)
86 {
87 	VM_BUG_ON(atomic_read(&anon_vma->refcount));
88 
89 	/*
90 	 * Synchronize against page_lock_anon_vma_read() such that
91 	 * we can safely hold the lock without the anon_vma getting
92 	 * freed.
93 	 *
94 	 * Relies on the full mb implied by the atomic_dec_and_test() from
95 	 * put_anon_vma() against the acquire barrier implied by
96 	 * down_read_trylock() from page_lock_anon_vma_read(). This orders:
97 	 *
98 	 * page_lock_anon_vma_read()	VS	put_anon_vma()
99 	 *   down_read_trylock()		  atomic_dec_and_test()
100 	 *   LOCK				  MB
101 	 *   atomic_read()			  rwsem_is_locked()
102 	 *
103 	 * LOCK should suffice since the actual taking of the lock must
104 	 * happen _before_ what follows.
105 	 */
106 	might_sleep();
107 	if (rwsem_is_locked(&anon_vma->root->rwsem)) {
108 		anon_vma_lock_write(anon_vma);
109 		anon_vma_unlock_write(anon_vma);
110 	}
111 
112 	kmem_cache_free(anon_vma_cachep, anon_vma);
113 }
114 
115 static inline struct anon_vma_chain *anon_vma_chain_alloc(gfp_t gfp)
116 {
117 	return kmem_cache_alloc(anon_vma_chain_cachep, gfp);
118 }
119 
120 static void anon_vma_chain_free(struct anon_vma_chain *anon_vma_chain)
121 {
122 	kmem_cache_free(anon_vma_chain_cachep, anon_vma_chain);
123 }
124 
125 static void anon_vma_chain_link(struct vm_area_struct *vma,
126 				struct anon_vma_chain *avc,
127 				struct anon_vma *anon_vma)
128 {
129 	avc->vma = vma;
130 	avc->anon_vma = anon_vma;
131 	list_add(&avc->same_vma, &vma->anon_vma_chain);
132 	anon_vma_interval_tree_insert(avc, &anon_vma->rb_root);
133 }
134 
135 /**
136  * anon_vma_prepare - attach an anon_vma to a memory region
137  * @vma: the memory region in question
138  *
139  * This makes sure the memory mapping described by 'vma' has
140  * an 'anon_vma' attached to it, so that we can associate the
141  * anonymous pages mapped into it with that anon_vma.
142  *
143  * The common case will be that we already have one, but if
144  * not we either need to find an adjacent mapping that we
145  * can re-use the anon_vma from (very common when the only
146  * reason for splitting a vma has been mprotect()), or we
147  * allocate a new one.
148  *
149  * Anon-vma allocations are very subtle, because we may have
150  * optimistically looked up an anon_vma in page_lock_anon_vma_read()
151  * and that may actually touch the spinlock even in the newly
152  * allocated vma (it depends on RCU to make sure that the
153  * anon_vma isn't actually destroyed).
154  *
155  * As a result, we need to do proper anon_vma locking even
156  * for the new allocation. At the same time, we do not want
157  * to do any locking for the common case of already having
158  * an anon_vma.
159  *
160  * This must be called with the mmap_sem held for reading.
161  */
162 int anon_vma_prepare(struct vm_area_struct *vma)
163 {
164 	struct anon_vma *anon_vma = vma->anon_vma;
165 	struct anon_vma_chain *avc;
166 
167 	might_sleep();
168 	if (unlikely(!anon_vma)) {
169 		struct mm_struct *mm = vma->vm_mm;
170 		struct anon_vma *allocated;
171 
172 		avc = anon_vma_chain_alloc(GFP_KERNEL);
173 		if (!avc)
174 			goto out_enomem;
175 
176 		anon_vma = find_mergeable_anon_vma(vma);
177 		allocated = NULL;
178 		if (!anon_vma) {
179 			anon_vma = anon_vma_alloc();
180 			if (unlikely(!anon_vma))
181 				goto out_enomem_free_avc;
182 			allocated = anon_vma;
183 		}
184 
185 		anon_vma_lock_write(anon_vma);
186 		/* page_table_lock to protect against threads */
187 		spin_lock(&mm->page_table_lock);
188 		if (likely(!vma->anon_vma)) {
189 			vma->anon_vma = anon_vma;
190 			anon_vma_chain_link(vma, avc, anon_vma);
191 			allocated = NULL;
192 			avc = NULL;
193 		}
194 		spin_unlock(&mm->page_table_lock);
195 		anon_vma_unlock_write(anon_vma);
196 
197 		if (unlikely(allocated))
198 			put_anon_vma(allocated);
199 		if (unlikely(avc))
200 			anon_vma_chain_free(avc);
201 	}
202 	return 0;
203 
204  out_enomem_free_avc:
205 	anon_vma_chain_free(avc);
206  out_enomem:
207 	return -ENOMEM;
208 }
209 
210 /*
211  * This is a useful helper function for locking the anon_vma root as
212  * we traverse the vma->anon_vma_chain, looping over anon_vma's that
213  * have the same vma.
214  *
215  * Such anon_vma's should have the same root, so you'd expect to see
216  * just a single mutex_lock for the whole traversal.
217  */
218 static inline struct anon_vma *lock_anon_vma_root(struct anon_vma *root, struct anon_vma *anon_vma)
219 {
220 	struct anon_vma *new_root = anon_vma->root;
221 	if (new_root != root) {
222 		if (WARN_ON_ONCE(root))
223 			up_write(&root->rwsem);
224 		root = new_root;
225 		down_write(&root->rwsem);
226 	}
227 	return root;
228 }
229 
230 static inline void unlock_anon_vma_root(struct anon_vma *root)
231 {
232 	if (root)
233 		up_write(&root->rwsem);
234 }
235 
236 /*
237  * Attach the anon_vmas from src to dst.
238  * Returns 0 on success, -ENOMEM on failure.
239  */
240 int anon_vma_clone(struct vm_area_struct *dst, struct vm_area_struct *src)
241 {
242 	struct anon_vma_chain *avc, *pavc;
243 	struct anon_vma *root = NULL;
244 
245 	list_for_each_entry_reverse(pavc, &src->anon_vma_chain, same_vma) {
246 		struct anon_vma *anon_vma;
247 
248 		avc = anon_vma_chain_alloc(GFP_NOWAIT | __GFP_NOWARN);
249 		if (unlikely(!avc)) {
250 			unlock_anon_vma_root(root);
251 			root = NULL;
252 			avc = anon_vma_chain_alloc(GFP_KERNEL);
253 			if (!avc)
254 				goto enomem_failure;
255 		}
256 		anon_vma = pavc->anon_vma;
257 		root = lock_anon_vma_root(root, anon_vma);
258 		anon_vma_chain_link(dst, avc, anon_vma);
259 	}
260 	unlock_anon_vma_root(root);
261 	return 0;
262 
263  enomem_failure:
264 	unlink_anon_vmas(dst);
265 	return -ENOMEM;
266 }
267 
268 /*
269  * Attach vma to its own anon_vma, as well as to the anon_vmas that
270  * the corresponding VMA in the parent process is attached to.
271  * Returns 0 on success, non-zero on failure.
272  */
273 int anon_vma_fork(struct vm_area_struct *vma, struct vm_area_struct *pvma)
274 {
275 	struct anon_vma_chain *avc;
276 	struct anon_vma *anon_vma;
277 
278 	/* Don't bother if the parent process has no anon_vma here. */
279 	if (!pvma->anon_vma)
280 		return 0;
281 
282 	/*
283 	 * First, attach the new VMA to the parent VMA's anon_vmas,
284 	 * so rmap can find non-COWed pages in child processes.
285 	 */
286 	if (anon_vma_clone(vma, pvma))
287 		return -ENOMEM;
288 
289 	/* Then add our own anon_vma. */
290 	anon_vma = anon_vma_alloc();
291 	if (!anon_vma)
292 		goto out_error;
293 	avc = anon_vma_chain_alloc(GFP_KERNEL);
294 	if (!avc)
295 		goto out_error_free_anon_vma;
296 
297 	/*
298 	 * The root anon_vma's spinlock is the lock actually used when we
299 	 * lock any of the anon_vmas in this anon_vma tree.
300 	 */
301 	anon_vma->root = pvma->anon_vma->root;
302 	/*
303 	 * With refcounts, an anon_vma can stay around longer than the
304 	 * process it belongs to. The root anon_vma needs to be pinned until
305 	 * this anon_vma is freed, because the lock lives in the root.
306 	 */
307 	get_anon_vma(anon_vma->root);
308 	/* Mark this anon_vma as the one where our new (COWed) pages go. */
309 	vma->anon_vma = anon_vma;
310 	anon_vma_lock_write(anon_vma);
311 	anon_vma_chain_link(vma, avc, anon_vma);
312 	anon_vma_unlock_write(anon_vma);
313 
314 	return 0;
315 
316  out_error_free_anon_vma:
317 	put_anon_vma(anon_vma);
318  out_error:
319 	unlink_anon_vmas(vma);
320 	return -ENOMEM;
321 }
322 
323 void unlink_anon_vmas(struct vm_area_struct *vma)
324 {
325 	struct anon_vma_chain *avc, *next;
326 	struct anon_vma *root = NULL;
327 
328 	/*
329 	 * Unlink each anon_vma chained to the VMA.  This list is ordered
330 	 * from newest to oldest, ensuring the root anon_vma gets freed last.
331 	 */
332 	list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
333 		struct anon_vma *anon_vma = avc->anon_vma;
334 
335 		root = lock_anon_vma_root(root, anon_vma);
336 		anon_vma_interval_tree_remove(avc, &anon_vma->rb_root);
337 
338 		/*
339 		 * Leave empty anon_vmas on the list - we'll need
340 		 * to free them outside the lock.
341 		 */
342 		if (RB_EMPTY_ROOT(&anon_vma->rb_root))
343 			continue;
344 
345 		list_del(&avc->same_vma);
346 		anon_vma_chain_free(avc);
347 	}
348 	unlock_anon_vma_root(root);
349 
350 	/*
351 	 * Iterate the list once more, it now only contains empty and unlinked
352 	 * anon_vmas, destroy them. Could not do before due to __put_anon_vma()
353 	 * needing to write-acquire the anon_vma->root->rwsem.
354 	 */
355 	list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
356 		struct anon_vma *anon_vma = avc->anon_vma;
357 
358 		put_anon_vma(anon_vma);
359 
360 		list_del(&avc->same_vma);
361 		anon_vma_chain_free(avc);
362 	}
363 }
364 
365 static void anon_vma_ctor(void *data)
366 {
367 	struct anon_vma *anon_vma = data;
368 
369 	init_rwsem(&anon_vma->rwsem);
370 	atomic_set(&anon_vma->refcount, 0);
371 	anon_vma->rb_root = RB_ROOT;
372 }
373 
374 void __init anon_vma_init(void)
375 {
376 	anon_vma_cachep = kmem_cache_create("anon_vma", sizeof(struct anon_vma),
377 			0, SLAB_DESTROY_BY_RCU|SLAB_PANIC, anon_vma_ctor);
378 	anon_vma_chain_cachep = KMEM_CACHE(anon_vma_chain, SLAB_PANIC);
379 }
380 
381 /*
382  * Getting a lock on a stable anon_vma from a page off the LRU is tricky!
383  *
384  * Since there is no serialization what so ever against page_remove_rmap()
385  * the best this function can do is return a locked anon_vma that might
386  * have been relevant to this page.
387  *
388  * The page might have been remapped to a different anon_vma or the anon_vma
389  * returned may already be freed (and even reused).
390  *
391  * In case it was remapped to a different anon_vma, the new anon_vma will be a
392  * child of the old anon_vma, and the anon_vma lifetime rules will therefore
393  * ensure that any anon_vma obtained from the page will still be valid for as
394  * long as we observe page_mapped() [ hence all those page_mapped() tests ].
395  *
396  * All users of this function must be very careful when walking the anon_vma
397  * chain and verify that the page in question is indeed mapped in it
398  * [ something equivalent to page_mapped_in_vma() ].
399  *
400  * Since anon_vma's slab is DESTROY_BY_RCU and we know from page_remove_rmap()
401  * that the anon_vma pointer from page->mapping is valid if there is a
402  * mapcount, we can dereference the anon_vma after observing those.
403  */
404 struct anon_vma *page_get_anon_vma(struct page *page)
405 {
406 	struct anon_vma *anon_vma = NULL;
407 	unsigned long anon_mapping;
408 
409 	rcu_read_lock();
410 	anon_mapping = (unsigned long) ACCESS_ONCE(page->mapping);
411 	if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
412 		goto out;
413 	if (!page_mapped(page))
414 		goto out;
415 
416 	anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
417 	if (!atomic_inc_not_zero(&anon_vma->refcount)) {
418 		anon_vma = NULL;
419 		goto out;
420 	}
421 
422 	/*
423 	 * If this page is still mapped, then its anon_vma cannot have been
424 	 * freed.  But if it has been unmapped, we have no security against the
425 	 * anon_vma structure being freed and reused (for another anon_vma:
426 	 * SLAB_DESTROY_BY_RCU guarantees that - so the atomic_inc_not_zero()
427 	 * above cannot corrupt).
428 	 */
429 	if (!page_mapped(page)) {
430 		rcu_read_unlock();
431 		put_anon_vma(anon_vma);
432 		return NULL;
433 	}
434 out:
435 	rcu_read_unlock();
436 
437 	return anon_vma;
438 }
439 
440 /*
441  * Similar to page_get_anon_vma() except it locks the anon_vma.
442  *
443  * Its a little more complex as it tries to keep the fast path to a single
444  * atomic op -- the trylock. If we fail the trylock, we fall back to getting a
445  * reference like with page_get_anon_vma() and then block on the mutex.
446  */
447 struct anon_vma *page_lock_anon_vma_read(struct page *page)
448 {
449 	struct anon_vma *anon_vma = NULL;
450 	struct anon_vma *root_anon_vma;
451 	unsigned long anon_mapping;
452 
453 	rcu_read_lock();
454 	anon_mapping = (unsigned long) ACCESS_ONCE(page->mapping);
455 	if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
456 		goto out;
457 	if (!page_mapped(page))
458 		goto out;
459 
460 	anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
461 	root_anon_vma = ACCESS_ONCE(anon_vma->root);
462 	if (down_read_trylock(&root_anon_vma->rwsem)) {
463 		/*
464 		 * If the page is still mapped, then this anon_vma is still
465 		 * its anon_vma, and holding the mutex ensures that it will
466 		 * not go away, see anon_vma_free().
467 		 */
468 		if (!page_mapped(page)) {
469 			up_read(&root_anon_vma->rwsem);
470 			anon_vma = NULL;
471 		}
472 		goto out;
473 	}
474 
475 	/* trylock failed, we got to sleep */
476 	if (!atomic_inc_not_zero(&anon_vma->refcount)) {
477 		anon_vma = NULL;
478 		goto out;
479 	}
480 
481 	if (!page_mapped(page)) {
482 		rcu_read_unlock();
483 		put_anon_vma(anon_vma);
484 		return NULL;
485 	}
486 
487 	/* we pinned the anon_vma, its safe to sleep */
488 	rcu_read_unlock();
489 	anon_vma_lock_read(anon_vma);
490 
491 	if (atomic_dec_and_test(&anon_vma->refcount)) {
492 		/*
493 		 * Oops, we held the last refcount, release the lock
494 		 * and bail -- can't simply use put_anon_vma() because
495 		 * we'll deadlock on the anon_vma_lock_write() recursion.
496 		 */
497 		anon_vma_unlock_read(anon_vma);
498 		__put_anon_vma(anon_vma);
499 		anon_vma = NULL;
500 	}
501 
502 	return anon_vma;
503 
504 out:
505 	rcu_read_unlock();
506 	return anon_vma;
507 }
508 
509 void page_unlock_anon_vma_read(struct anon_vma *anon_vma)
510 {
511 	anon_vma_unlock_read(anon_vma);
512 }
513 
514 /*
515  * At what user virtual address is page expected in @vma?
516  */
517 static inline unsigned long
518 __vma_address(struct page *page, struct vm_area_struct *vma)
519 {
520 	pgoff_t pgoff = page_to_pgoff(page);
521 	return vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
522 }
523 
524 inline unsigned long
525 vma_address(struct page *page, struct vm_area_struct *vma)
526 {
527 	unsigned long address = __vma_address(page, vma);
528 
529 	/* page should be within @vma mapping range */
530 	VM_BUG_ON_VMA(address < vma->vm_start || address >= vma->vm_end, vma);
531 
532 	return address;
533 }
534 
535 /*
536  * At what user virtual address is page expected in vma?
537  * Caller should check the page is actually part of the vma.
538  */
539 unsigned long page_address_in_vma(struct page *page, struct vm_area_struct *vma)
540 {
541 	unsigned long address;
542 	if (PageAnon(page)) {
543 		struct anon_vma *page__anon_vma = page_anon_vma(page);
544 		/*
545 		 * Note: swapoff's unuse_vma() is more efficient with this
546 		 * check, and needs it to match anon_vma when KSM is active.
547 		 */
548 		if (!vma->anon_vma || !page__anon_vma ||
549 		    vma->anon_vma->root != page__anon_vma->root)
550 			return -EFAULT;
551 	} else if (page->mapping && !(vma->vm_flags & VM_NONLINEAR)) {
552 		if (!vma->vm_file ||
553 		    vma->vm_file->f_mapping != page->mapping)
554 			return -EFAULT;
555 	} else
556 		return -EFAULT;
557 	address = __vma_address(page, vma);
558 	if (unlikely(address < vma->vm_start || address >= vma->vm_end))
559 		return -EFAULT;
560 	return address;
561 }
562 
563 pmd_t *mm_find_pmd(struct mm_struct *mm, unsigned long address)
564 {
565 	pgd_t *pgd;
566 	pud_t *pud;
567 	pmd_t *pmd = NULL;
568 	pmd_t pmde;
569 
570 	pgd = pgd_offset(mm, address);
571 	if (!pgd_present(*pgd))
572 		goto out;
573 
574 	pud = pud_offset(pgd, address);
575 	if (!pud_present(*pud))
576 		goto out;
577 
578 	pmd = pmd_offset(pud, address);
579 	/*
580 	 * Some THP functions use the sequence pmdp_clear_flush(), set_pmd_at()
581 	 * without holding anon_vma lock for write.  So when looking for a
582 	 * genuine pmde (in which to find pte), test present and !THP together.
583 	 */
584 	pmde = ACCESS_ONCE(*pmd);
585 	if (!pmd_present(pmde) || pmd_trans_huge(pmde))
586 		pmd = NULL;
587 out:
588 	return pmd;
589 }
590 
591 /*
592  * Check that @page is mapped at @address into @mm.
593  *
594  * If @sync is false, page_check_address may perform a racy check to avoid
595  * the page table lock when the pte is not present (helpful when reclaiming
596  * highly shared pages).
597  *
598  * On success returns with pte mapped and locked.
599  */
600 pte_t *__page_check_address(struct page *page, struct mm_struct *mm,
601 			  unsigned long address, spinlock_t **ptlp, int sync)
602 {
603 	pmd_t *pmd;
604 	pte_t *pte;
605 	spinlock_t *ptl;
606 
607 	if (unlikely(PageHuge(page))) {
608 		/* when pud is not present, pte will be NULL */
609 		pte = huge_pte_offset(mm, address);
610 		if (!pte)
611 			return NULL;
612 
613 		ptl = huge_pte_lockptr(page_hstate(page), mm, pte);
614 		goto check;
615 	}
616 
617 	pmd = mm_find_pmd(mm, address);
618 	if (!pmd)
619 		return NULL;
620 
621 	pte = pte_offset_map(pmd, address);
622 	/* Make a quick check before getting the lock */
623 	if (!sync && !pte_present(*pte)) {
624 		pte_unmap(pte);
625 		return NULL;
626 	}
627 
628 	ptl = pte_lockptr(mm, pmd);
629 check:
630 	spin_lock(ptl);
631 	if (pte_present(*pte) && page_to_pfn(page) == pte_pfn(*pte)) {
632 		*ptlp = ptl;
633 		return pte;
634 	}
635 	pte_unmap_unlock(pte, ptl);
636 	return NULL;
637 }
638 
639 /**
640  * page_mapped_in_vma - check whether a page is really mapped in a VMA
641  * @page: the page to test
642  * @vma: the VMA to test
643  *
644  * Returns 1 if the page is mapped into the page tables of the VMA, 0
645  * if the page is not mapped into the page tables of this VMA.  Only
646  * valid for normal file or anonymous VMAs.
647  */
648 int page_mapped_in_vma(struct page *page, struct vm_area_struct *vma)
649 {
650 	unsigned long address;
651 	pte_t *pte;
652 	spinlock_t *ptl;
653 
654 	address = __vma_address(page, vma);
655 	if (unlikely(address < vma->vm_start || address >= vma->vm_end))
656 		return 0;
657 	pte = page_check_address(page, vma->vm_mm, address, &ptl, 1);
658 	if (!pte)			/* the page is not in this mm */
659 		return 0;
660 	pte_unmap_unlock(pte, ptl);
661 
662 	return 1;
663 }
664 
665 struct page_referenced_arg {
666 	int mapcount;
667 	int referenced;
668 	unsigned long vm_flags;
669 	struct mem_cgroup *memcg;
670 };
671 /*
672  * arg: page_referenced_arg will be passed
673  */
674 static int page_referenced_one(struct page *page, struct vm_area_struct *vma,
675 			unsigned long address, void *arg)
676 {
677 	struct mm_struct *mm = vma->vm_mm;
678 	spinlock_t *ptl;
679 	int referenced = 0;
680 	struct page_referenced_arg *pra = arg;
681 
682 	if (unlikely(PageTransHuge(page))) {
683 		pmd_t *pmd;
684 
685 		/*
686 		 * rmap might return false positives; we must filter
687 		 * these out using page_check_address_pmd().
688 		 */
689 		pmd = page_check_address_pmd(page, mm, address,
690 					     PAGE_CHECK_ADDRESS_PMD_FLAG, &ptl);
691 		if (!pmd)
692 			return SWAP_AGAIN;
693 
694 		if (vma->vm_flags & VM_LOCKED) {
695 			spin_unlock(ptl);
696 			pra->vm_flags |= VM_LOCKED;
697 			return SWAP_FAIL; /* To break the loop */
698 		}
699 
700 		/* go ahead even if the pmd is pmd_trans_splitting() */
701 		if (pmdp_clear_flush_young_notify(vma, address, pmd))
702 			referenced++;
703 		spin_unlock(ptl);
704 	} else {
705 		pte_t *pte;
706 
707 		/*
708 		 * rmap might return false positives; we must filter
709 		 * these out using page_check_address().
710 		 */
711 		pte = page_check_address(page, mm, address, &ptl, 0);
712 		if (!pte)
713 			return SWAP_AGAIN;
714 
715 		if (vma->vm_flags & VM_LOCKED) {
716 			pte_unmap_unlock(pte, ptl);
717 			pra->vm_flags |= VM_LOCKED;
718 			return SWAP_FAIL; /* To break the loop */
719 		}
720 
721 		if (ptep_clear_flush_young_notify(vma, address, pte)) {
722 			/*
723 			 * Don't treat a reference through a sequentially read
724 			 * mapping as such.  If the page has been used in
725 			 * another mapping, we will catch it; if this other
726 			 * mapping is already gone, the unmap path will have
727 			 * set PG_referenced or activated the page.
728 			 */
729 			if (likely(!(vma->vm_flags & VM_SEQ_READ)))
730 				referenced++;
731 		}
732 		pte_unmap_unlock(pte, ptl);
733 	}
734 
735 	if (referenced) {
736 		pra->referenced++;
737 		pra->vm_flags |= vma->vm_flags;
738 	}
739 
740 	pra->mapcount--;
741 	if (!pra->mapcount)
742 		return SWAP_SUCCESS; /* To break the loop */
743 
744 	return SWAP_AGAIN;
745 }
746 
747 static bool invalid_page_referenced_vma(struct vm_area_struct *vma, void *arg)
748 {
749 	struct page_referenced_arg *pra = arg;
750 	struct mem_cgroup *memcg = pra->memcg;
751 
752 	if (!mm_match_cgroup(vma->vm_mm, memcg))
753 		return true;
754 
755 	return false;
756 }
757 
758 /**
759  * page_referenced - test if the page was referenced
760  * @page: the page to test
761  * @is_locked: caller holds lock on the page
762  * @memcg: target memory cgroup
763  * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
764  *
765  * Quick test_and_clear_referenced for all mappings to a page,
766  * returns the number of ptes which referenced the page.
767  */
768 int page_referenced(struct page *page,
769 		    int is_locked,
770 		    struct mem_cgroup *memcg,
771 		    unsigned long *vm_flags)
772 {
773 	int ret;
774 	int we_locked = 0;
775 	struct page_referenced_arg pra = {
776 		.mapcount = page_mapcount(page),
777 		.memcg = memcg,
778 	};
779 	struct rmap_walk_control rwc = {
780 		.rmap_one = page_referenced_one,
781 		.arg = (void *)&pra,
782 		.anon_lock = page_lock_anon_vma_read,
783 	};
784 
785 	*vm_flags = 0;
786 	if (!page_mapped(page))
787 		return 0;
788 
789 	if (!page_rmapping(page))
790 		return 0;
791 
792 	if (!is_locked && (!PageAnon(page) || PageKsm(page))) {
793 		we_locked = trylock_page(page);
794 		if (!we_locked)
795 			return 1;
796 	}
797 
798 	/*
799 	 * If we are reclaiming on behalf of a cgroup, skip
800 	 * counting on behalf of references from different
801 	 * cgroups
802 	 */
803 	if (memcg) {
804 		rwc.invalid_vma = invalid_page_referenced_vma;
805 	}
806 
807 	ret = rmap_walk(page, &rwc);
808 	*vm_flags = pra.vm_flags;
809 
810 	if (we_locked)
811 		unlock_page(page);
812 
813 	return pra.referenced;
814 }
815 
816 static int page_mkclean_one(struct page *page, struct vm_area_struct *vma,
817 			    unsigned long address, void *arg)
818 {
819 	struct mm_struct *mm = vma->vm_mm;
820 	pte_t *pte;
821 	spinlock_t *ptl;
822 	int ret = 0;
823 	int *cleaned = arg;
824 
825 	pte = page_check_address(page, mm, address, &ptl, 1);
826 	if (!pte)
827 		goto out;
828 
829 	if (pte_dirty(*pte) || pte_write(*pte)) {
830 		pte_t entry;
831 
832 		flush_cache_page(vma, address, pte_pfn(*pte));
833 		entry = ptep_clear_flush(vma, address, pte);
834 		entry = pte_wrprotect(entry);
835 		entry = pte_mkclean(entry);
836 		set_pte_at(mm, address, pte, entry);
837 		ret = 1;
838 	}
839 
840 	pte_unmap_unlock(pte, ptl);
841 
842 	if (ret) {
843 		mmu_notifier_invalidate_page(mm, address);
844 		(*cleaned)++;
845 	}
846 out:
847 	return SWAP_AGAIN;
848 }
849 
850 static bool invalid_mkclean_vma(struct vm_area_struct *vma, void *arg)
851 {
852 	if (vma->vm_flags & VM_SHARED)
853 		return false;
854 
855 	return true;
856 }
857 
858 int page_mkclean(struct page *page)
859 {
860 	int cleaned = 0;
861 	struct address_space *mapping;
862 	struct rmap_walk_control rwc = {
863 		.arg = (void *)&cleaned,
864 		.rmap_one = page_mkclean_one,
865 		.invalid_vma = invalid_mkclean_vma,
866 	};
867 
868 	BUG_ON(!PageLocked(page));
869 
870 	if (!page_mapped(page))
871 		return 0;
872 
873 	mapping = page_mapping(page);
874 	if (!mapping)
875 		return 0;
876 
877 	rmap_walk(page, &rwc);
878 
879 	return cleaned;
880 }
881 EXPORT_SYMBOL_GPL(page_mkclean);
882 
883 /**
884  * page_move_anon_rmap - move a page to our anon_vma
885  * @page:	the page to move to our anon_vma
886  * @vma:	the vma the page belongs to
887  * @address:	the user virtual address mapped
888  *
889  * When a page belongs exclusively to one process after a COW event,
890  * that page can be moved into the anon_vma that belongs to just that
891  * process, so the rmap code will not search the parent or sibling
892  * processes.
893  */
894 void page_move_anon_rmap(struct page *page,
895 	struct vm_area_struct *vma, unsigned long address)
896 {
897 	struct anon_vma *anon_vma = vma->anon_vma;
898 
899 	VM_BUG_ON_PAGE(!PageLocked(page), page);
900 	VM_BUG_ON_VMA(!anon_vma, vma);
901 	VM_BUG_ON_PAGE(page->index != linear_page_index(vma, address), page);
902 
903 	anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
904 	page->mapping = (struct address_space *) anon_vma;
905 }
906 
907 /**
908  * __page_set_anon_rmap - set up new anonymous rmap
909  * @page:	Page to add to rmap
910  * @vma:	VM area to add page to.
911  * @address:	User virtual address of the mapping
912  * @exclusive:	the page is exclusively owned by the current process
913  */
914 static void __page_set_anon_rmap(struct page *page,
915 	struct vm_area_struct *vma, unsigned long address, int exclusive)
916 {
917 	struct anon_vma *anon_vma = vma->anon_vma;
918 
919 	BUG_ON(!anon_vma);
920 
921 	if (PageAnon(page))
922 		return;
923 
924 	/*
925 	 * If the page isn't exclusively mapped into this vma,
926 	 * we must use the _oldest_ possible anon_vma for the
927 	 * page mapping!
928 	 */
929 	if (!exclusive)
930 		anon_vma = anon_vma->root;
931 
932 	anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
933 	page->mapping = (struct address_space *) anon_vma;
934 	page->index = linear_page_index(vma, address);
935 }
936 
937 /**
938  * __page_check_anon_rmap - sanity check anonymous rmap addition
939  * @page:	the page to add the mapping to
940  * @vma:	the vm area in which the mapping is added
941  * @address:	the user virtual address mapped
942  */
943 static void __page_check_anon_rmap(struct page *page,
944 	struct vm_area_struct *vma, unsigned long address)
945 {
946 #ifdef CONFIG_DEBUG_VM
947 	/*
948 	 * The page's anon-rmap details (mapping and index) are guaranteed to
949 	 * be set up correctly at this point.
950 	 *
951 	 * We have exclusion against page_add_anon_rmap because the caller
952 	 * always holds the page locked, except if called from page_dup_rmap,
953 	 * in which case the page is already known to be setup.
954 	 *
955 	 * We have exclusion against page_add_new_anon_rmap because those pages
956 	 * are initially only visible via the pagetables, and the pte is locked
957 	 * over the call to page_add_new_anon_rmap.
958 	 */
959 	BUG_ON(page_anon_vma(page)->root != vma->anon_vma->root);
960 	BUG_ON(page->index != linear_page_index(vma, address));
961 #endif
962 }
963 
964 /**
965  * page_add_anon_rmap - add pte mapping to an anonymous page
966  * @page:	the page to add the mapping to
967  * @vma:	the vm area in which the mapping is added
968  * @address:	the user virtual address mapped
969  *
970  * The caller needs to hold the pte lock, and the page must be locked in
971  * the anon_vma case: to serialize mapping,index checking after setting,
972  * and to ensure that PageAnon is not being upgraded racily to PageKsm
973  * (but PageKsm is never downgraded to PageAnon).
974  */
975 void page_add_anon_rmap(struct page *page,
976 	struct vm_area_struct *vma, unsigned long address)
977 {
978 	do_page_add_anon_rmap(page, vma, address, 0);
979 }
980 
981 /*
982  * Special version of the above for do_swap_page, which often runs
983  * into pages that are exclusively owned by the current process.
984  * Everybody else should continue to use page_add_anon_rmap above.
985  */
986 void do_page_add_anon_rmap(struct page *page,
987 	struct vm_area_struct *vma, unsigned long address, int exclusive)
988 {
989 	int first = atomic_inc_and_test(&page->_mapcount);
990 	if (first) {
991 		/*
992 		 * We use the irq-unsafe __{inc|mod}_zone_page_stat because
993 		 * these counters are not modified in interrupt context, and
994 		 * pte lock(a spinlock) is held, which implies preemption
995 		 * disabled.
996 		 */
997 		if (PageTransHuge(page))
998 			__inc_zone_page_state(page,
999 					      NR_ANON_TRANSPARENT_HUGEPAGES);
1000 		__mod_zone_page_state(page_zone(page), NR_ANON_PAGES,
1001 				hpage_nr_pages(page));
1002 	}
1003 	if (unlikely(PageKsm(page)))
1004 		return;
1005 
1006 	VM_BUG_ON_PAGE(!PageLocked(page), page);
1007 	/* address might be in next vma when migration races vma_adjust */
1008 	if (first)
1009 		__page_set_anon_rmap(page, vma, address, exclusive);
1010 	else
1011 		__page_check_anon_rmap(page, vma, address);
1012 }
1013 
1014 /**
1015  * page_add_new_anon_rmap - add pte mapping to a new anonymous page
1016  * @page:	the page to add the mapping to
1017  * @vma:	the vm area in which the mapping is added
1018  * @address:	the user virtual address mapped
1019  *
1020  * Same as page_add_anon_rmap but must only be called on *new* pages.
1021  * This means the inc-and-test can be bypassed.
1022  * Page does not have to be locked.
1023  */
1024 void page_add_new_anon_rmap(struct page *page,
1025 	struct vm_area_struct *vma, unsigned long address)
1026 {
1027 	VM_BUG_ON_VMA(address < vma->vm_start || address >= vma->vm_end, vma);
1028 	SetPageSwapBacked(page);
1029 	atomic_set(&page->_mapcount, 0); /* increment count (starts at -1) */
1030 	if (PageTransHuge(page))
1031 		__inc_zone_page_state(page, NR_ANON_TRANSPARENT_HUGEPAGES);
1032 	__mod_zone_page_state(page_zone(page), NR_ANON_PAGES,
1033 			hpage_nr_pages(page));
1034 	__page_set_anon_rmap(page, vma, address, 1);
1035 }
1036 
1037 /**
1038  * page_add_file_rmap - add pte mapping to a file page
1039  * @page: the page to add the mapping to
1040  *
1041  * The caller needs to hold the pte lock.
1042  */
1043 void page_add_file_rmap(struct page *page)
1044 {
1045 	struct mem_cgroup *memcg;
1046 	unsigned long flags;
1047 	bool locked;
1048 
1049 	memcg = mem_cgroup_begin_page_stat(page, &locked, &flags);
1050 	if (atomic_inc_and_test(&page->_mapcount)) {
1051 		__inc_zone_page_state(page, NR_FILE_MAPPED);
1052 		mem_cgroup_inc_page_stat(memcg, MEM_CGROUP_STAT_FILE_MAPPED);
1053 	}
1054 	mem_cgroup_end_page_stat(memcg, locked, flags);
1055 }
1056 
1057 static void page_remove_file_rmap(struct page *page)
1058 {
1059 	struct mem_cgroup *memcg;
1060 	unsigned long flags;
1061 	bool locked;
1062 
1063 	memcg = mem_cgroup_begin_page_stat(page, &locked, &flags);
1064 
1065 	/* page still mapped by someone else? */
1066 	if (!atomic_add_negative(-1, &page->_mapcount))
1067 		goto out;
1068 
1069 	/* Hugepages are not counted in NR_FILE_MAPPED for now. */
1070 	if (unlikely(PageHuge(page)))
1071 		goto out;
1072 
1073 	/*
1074 	 * We use the irq-unsafe __{inc|mod}_zone_page_stat because
1075 	 * these counters are not modified in interrupt context, and
1076 	 * pte lock(a spinlock) is held, which implies preemption disabled.
1077 	 */
1078 	__dec_zone_page_state(page, NR_FILE_MAPPED);
1079 	mem_cgroup_dec_page_stat(memcg, MEM_CGROUP_STAT_FILE_MAPPED);
1080 
1081 	if (unlikely(PageMlocked(page)))
1082 		clear_page_mlock(page);
1083 out:
1084 	mem_cgroup_end_page_stat(memcg, locked, flags);
1085 }
1086 
1087 /**
1088  * page_remove_rmap - take down pte mapping from a page
1089  * @page: page to remove mapping from
1090  *
1091  * The caller needs to hold the pte lock.
1092  */
1093 void page_remove_rmap(struct page *page)
1094 {
1095 	if (!PageAnon(page)) {
1096 		page_remove_file_rmap(page);
1097 		return;
1098 	}
1099 
1100 	/* page still mapped by someone else? */
1101 	if (!atomic_add_negative(-1, &page->_mapcount))
1102 		return;
1103 
1104 	/* Hugepages are not counted in NR_ANON_PAGES for now. */
1105 	if (unlikely(PageHuge(page)))
1106 		return;
1107 
1108 	/*
1109 	 * We use the irq-unsafe __{inc|mod}_zone_page_stat because
1110 	 * these counters are not modified in interrupt context, and
1111 	 * pte lock(a spinlock) is held, which implies preemption disabled.
1112 	 */
1113 	if (PageTransHuge(page))
1114 		__dec_zone_page_state(page, NR_ANON_TRANSPARENT_HUGEPAGES);
1115 
1116 	__mod_zone_page_state(page_zone(page), NR_ANON_PAGES,
1117 			      -hpage_nr_pages(page));
1118 
1119 	if (unlikely(PageMlocked(page)))
1120 		clear_page_mlock(page);
1121 
1122 	/*
1123 	 * It would be tidy to reset the PageAnon mapping here,
1124 	 * but that might overwrite a racing page_add_anon_rmap
1125 	 * which increments mapcount after us but sets mapping
1126 	 * before us: so leave the reset to free_hot_cold_page,
1127 	 * and remember that it's only reliable while mapped.
1128 	 * Leaving it set also helps swapoff to reinstate ptes
1129 	 * faster for those pages still in swapcache.
1130 	 */
1131 }
1132 
1133 /*
1134  * @arg: enum ttu_flags will be passed to this argument
1135  */
1136 static int try_to_unmap_one(struct page *page, struct vm_area_struct *vma,
1137 		     unsigned long address, void *arg)
1138 {
1139 	struct mm_struct *mm = vma->vm_mm;
1140 	pte_t *pte;
1141 	pte_t pteval;
1142 	spinlock_t *ptl;
1143 	int ret = SWAP_AGAIN;
1144 	enum ttu_flags flags = (enum ttu_flags)arg;
1145 
1146 	pte = page_check_address(page, mm, address, &ptl, 0);
1147 	if (!pte)
1148 		goto out;
1149 
1150 	/*
1151 	 * If the page is mlock()d, we cannot swap it out.
1152 	 * If it's recently referenced (perhaps page_referenced
1153 	 * skipped over this mm) then we should reactivate it.
1154 	 */
1155 	if (!(flags & TTU_IGNORE_MLOCK)) {
1156 		if (vma->vm_flags & VM_LOCKED)
1157 			goto out_mlock;
1158 
1159 		if (flags & TTU_MUNLOCK)
1160 			goto out_unmap;
1161 	}
1162 	if (!(flags & TTU_IGNORE_ACCESS)) {
1163 		if (ptep_clear_flush_young_notify(vma, address, pte)) {
1164 			ret = SWAP_FAIL;
1165 			goto out_unmap;
1166 		}
1167   	}
1168 
1169 	/* Nuke the page table entry. */
1170 	flush_cache_page(vma, address, page_to_pfn(page));
1171 	pteval = ptep_clear_flush(vma, address, pte);
1172 
1173 	/* Move the dirty bit to the physical page now the pte is gone. */
1174 	if (pte_dirty(pteval))
1175 		set_page_dirty(page);
1176 
1177 	/* Update high watermark before we lower rss */
1178 	update_hiwater_rss(mm);
1179 
1180 	if (PageHWPoison(page) && !(flags & TTU_IGNORE_HWPOISON)) {
1181 		if (!PageHuge(page)) {
1182 			if (PageAnon(page))
1183 				dec_mm_counter(mm, MM_ANONPAGES);
1184 			else
1185 				dec_mm_counter(mm, MM_FILEPAGES);
1186 		}
1187 		set_pte_at(mm, address, pte,
1188 			   swp_entry_to_pte(make_hwpoison_entry(page)));
1189 	} else if (pte_unused(pteval)) {
1190 		/*
1191 		 * The guest indicated that the page content is of no
1192 		 * interest anymore. Simply discard the pte, vmscan
1193 		 * will take care of the rest.
1194 		 */
1195 		if (PageAnon(page))
1196 			dec_mm_counter(mm, MM_ANONPAGES);
1197 		else
1198 			dec_mm_counter(mm, MM_FILEPAGES);
1199 	} else if (PageAnon(page)) {
1200 		swp_entry_t entry = { .val = page_private(page) };
1201 		pte_t swp_pte;
1202 
1203 		if (PageSwapCache(page)) {
1204 			/*
1205 			 * Store the swap location in the pte.
1206 			 * See handle_pte_fault() ...
1207 			 */
1208 			if (swap_duplicate(entry) < 0) {
1209 				set_pte_at(mm, address, pte, pteval);
1210 				ret = SWAP_FAIL;
1211 				goto out_unmap;
1212 			}
1213 			if (list_empty(&mm->mmlist)) {
1214 				spin_lock(&mmlist_lock);
1215 				if (list_empty(&mm->mmlist))
1216 					list_add(&mm->mmlist, &init_mm.mmlist);
1217 				spin_unlock(&mmlist_lock);
1218 			}
1219 			dec_mm_counter(mm, MM_ANONPAGES);
1220 			inc_mm_counter(mm, MM_SWAPENTS);
1221 		} else if (IS_ENABLED(CONFIG_MIGRATION)) {
1222 			/*
1223 			 * Store the pfn of the page in a special migration
1224 			 * pte. do_swap_page() will wait until the migration
1225 			 * pte is removed and then restart fault handling.
1226 			 */
1227 			BUG_ON(!(flags & TTU_MIGRATION));
1228 			entry = make_migration_entry(page, pte_write(pteval));
1229 		}
1230 		swp_pte = swp_entry_to_pte(entry);
1231 		if (pte_soft_dirty(pteval))
1232 			swp_pte = pte_swp_mksoft_dirty(swp_pte);
1233 		set_pte_at(mm, address, pte, swp_pte);
1234 		BUG_ON(pte_file(*pte));
1235 	} else if (IS_ENABLED(CONFIG_MIGRATION) &&
1236 		   (flags & TTU_MIGRATION)) {
1237 		/* Establish migration entry for a file page */
1238 		swp_entry_t entry;
1239 		entry = make_migration_entry(page, pte_write(pteval));
1240 		set_pte_at(mm, address, pte, swp_entry_to_pte(entry));
1241 	} else
1242 		dec_mm_counter(mm, MM_FILEPAGES);
1243 
1244 	page_remove_rmap(page);
1245 	page_cache_release(page);
1246 
1247 out_unmap:
1248 	pte_unmap_unlock(pte, ptl);
1249 	if (ret != SWAP_FAIL && !(flags & TTU_MUNLOCK))
1250 		mmu_notifier_invalidate_page(mm, address);
1251 out:
1252 	return ret;
1253 
1254 out_mlock:
1255 	pte_unmap_unlock(pte, ptl);
1256 
1257 
1258 	/*
1259 	 * We need mmap_sem locking, Otherwise VM_LOCKED check makes
1260 	 * unstable result and race. Plus, We can't wait here because
1261 	 * we now hold anon_vma->rwsem or mapping->i_mmap_mutex.
1262 	 * if trylock failed, the page remain in evictable lru and later
1263 	 * vmscan could retry to move the page to unevictable lru if the
1264 	 * page is actually mlocked.
1265 	 */
1266 	if (down_read_trylock(&vma->vm_mm->mmap_sem)) {
1267 		if (vma->vm_flags & VM_LOCKED) {
1268 			mlock_vma_page(page);
1269 			ret = SWAP_MLOCK;
1270 		}
1271 		up_read(&vma->vm_mm->mmap_sem);
1272 	}
1273 	return ret;
1274 }
1275 
1276 /*
1277  * objrmap doesn't work for nonlinear VMAs because the assumption that
1278  * offset-into-file correlates with offset-into-virtual-addresses does not hold.
1279  * Consequently, given a particular page and its ->index, we cannot locate the
1280  * ptes which are mapping that page without an exhaustive linear search.
1281  *
1282  * So what this code does is a mini "virtual scan" of each nonlinear VMA which
1283  * maps the file to which the target page belongs.  The ->vm_private_data field
1284  * holds the current cursor into that scan.  Successive searches will circulate
1285  * around the vma's virtual address space.
1286  *
1287  * So as more replacement pressure is applied to the pages in a nonlinear VMA,
1288  * more scanning pressure is placed against them as well.   Eventually pages
1289  * will become fully unmapped and are eligible for eviction.
1290  *
1291  * For very sparsely populated VMAs this is a little inefficient - chances are
1292  * there there won't be many ptes located within the scan cluster.  In this case
1293  * maybe we could scan further - to the end of the pte page, perhaps.
1294  *
1295  * Mlocked pages:  check VM_LOCKED under mmap_sem held for read, if we can
1296  * acquire it without blocking.  If vma locked, mlock the pages in the cluster,
1297  * rather than unmapping them.  If we encounter the "check_page" that vmscan is
1298  * trying to unmap, return SWAP_MLOCK, else default SWAP_AGAIN.
1299  */
1300 #define CLUSTER_SIZE	min(32*PAGE_SIZE, PMD_SIZE)
1301 #define CLUSTER_MASK	(~(CLUSTER_SIZE - 1))
1302 
1303 static int try_to_unmap_cluster(unsigned long cursor, unsigned int *mapcount,
1304 		struct vm_area_struct *vma, struct page *check_page)
1305 {
1306 	struct mm_struct *mm = vma->vm_mm;
1307 	pmd_t *pmd;
1308 	pte_t *pte;
1309 	pte_t pteval;
1310 	spinlock_t *ptl;
1311 	struct page *page;
1312 	unsigned long address;
1313 	unsigned long mmun_start;	/* For mmu_notifiers */
1314 	unsigned long mmun_end;		/* For mmu_notifiers */
1315 	unsigned long end;
1316 	int ret = SWAP_AGAIN;
1317 	int locked_vma = 0;
1318 
1319 	address = (vma->vm_start + cursor) & CLUSTER_MASK;
1320 	end = address + CLUSTER_SIZE;
1321 	if (address < vma->vm_start)
1322 		address = vma->vm_start;
1323 	if (end > vma->vm_end)
1324 		end = vma->vm_end;
1325 
1326 	pmd = mm_find_pmd(mm, address);
1327 	if (!pmd)
1328 		return ret;
1329 
1330 	mmun_start = address;
1331 	mmun_end   = end;
1332 	mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1333 
1334 	/*
1335 	 * If we can acquire the mmap_sem for read, and vma is VM_LOCKED,
1336 	 * keep the sem while scanning the cluster for mlocking pages.
1337 	 */
1338 	if (down_read_trylock(&vma->vm_mm->mmap_sem)) {
1339 		locked_vma = (vma->vm_flags & VM_LOCKED);
1340 		if (!locked_vma)
1341 			up_read(&vma->vm_mm->mmap_sem); /* don't need it */
1342 	}
1343 
1344 	pte = pte_offset_map_lock(mm, pmd, address, &ptl);
1345 
1346 	/* Update high watermark before we lower rss */
1347 	update_hiwater_rss(mm);
1348 
1349 	for (; address < end; pte++, address += PAGE_SIZE) {
1350 		if (!pte_present(*pte))
1351 			continue;
1352 		page = vm_normal_page(vma, address, *pte);
1353 		BUG_ON(!page || PageAnon(page));
1354 
1355 		if (locked_vma) {
1356 			if (page == check_page) {
1357 				/* we know we have check_page locked */
1358 				mlock_vma_page(page);
1359 				ret = SWAP_MLOCK;
1360 			} else if (trylock_page(page)) {
1361 				/*
1362 				 * If we can lock the page, perform mlock.
1363 				 * Otherwise leave the page alone, it will be
1364 				 * eventually encountered again later.
1365 				 */
1366 				mlock_vma_page(page);
1367 				unlock_page(page);
1368 			}
1369 			continue;	/* don't unmap */
1370 		}
1371 
1372 		/*
1373 		 * No need for _notify because we're within an
1374 		 * mmu_notifier_invalidate_range_ {start|end} scope.
1375 		 */
1376 		if (ptep_clear_flush_young(vma, address, pte))
1377 			continue;
1378 
1379 		/* Nuke the page table entry. */
1380 		flush_cache_page(vma, address, pte_pfn(*pte));
1381 		pteval = ptep_clear_flush(vma, address, pte);
1382 
1383 		/* If nonlinear, store the file page offset in the pte. */
1384 		if (page->index != linear_page_index(vma, address)) {
1385 			pte_t ptfile = pgoff_to_pte(page->index);
1386 			if (pte_soft_dirty(pteval))
1387 				ptfile = pte_file_mksoft_dirty(ptfile);
1388 			set_pte_at(mm, address, pte, ptfile);
1389 		}
1390 
1391 		/* Move the dirty bit to the physical page now the pte is gone. */
1392 		if (pte_dirty(pteval))
1393 			set_page_dirty(page);
1394 
1395 		page_remove_rmap(page);
1396 		page_cache_release(page);
1397 		dec_mm_counter(mm, MM_FILEPAGES);
1398 		(*mapcount)--;
1399 	}
1400 	pte_unmap_unlock(pte - 1, ptl);
1401 	mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1402 	if (locked_vma)
1403 		up_read(&vma->vm_mm->mmap_sem);
1404 	return ret;
1405 }
1406 
1407 static int try_to_unmap_nonlinear(struct page *page,
1408 		struct address_space *mapping, void *arg)
1409 {
1410 	struct vm_area_struct *vma;
1411 	int ret = SWAP_AGAIN;
1412 	unsigned long cursor;
1413 	unsigned long max_nl_cursor = 0;
1414 	unsigned long max_nl_size = 0;
1415 	unsigned int mapcount;
1416 
1417 	list_for_each_entry(vma,
1418 		&mapping->i_mmap_nonlinear, shared.nonlinear) {
1419 
1420 		cursor = (unsigned long) vma->vm_private_data;
1421 		if (cursor > max_nl_cursor)
1422 			max_nl_cursor = cursor;
1423 		cursor = vma->vm_end - vma->vm_start;
1424 		if (cursor > max_nl_size)
1425 			max_nl_size = cursor;
1426 	}
1427 
1428 	if (max_nl_size == 0) {	/* all nonlinears locked or reserved ? */
1429 		return SWAP_FAIL;
1430 	}
1431 
1432 	/*
1433 	 * We don't try to search for this page in the nonlinear vmas,
1434 	 * and page_referenced wouldn't have found it anyway.  Instead
1435 	 * just walk the nonlinear vmas trying to age and unmap some.
1436 	 * The mapcount of the page we came in with is irrelevant,
1437 	 * but even so use it as a guide to how hard we should try?
1438 	 */
1439 	mapcount = page_mapcount(page);
1440 	if (!mapcount)
1441 		return ret;
1442 
1443 	cond_resched();
1444 
1445 	max_nl_size = (max_nl_size + CLUSTER_SIZE - 1) & CLUSTER_MASK;
1446 	if (max_nl_cursor == 0)
1447 		max_nl_cursor = CLUSTER_SIZE;
1448 
1449 	do {
1450 		list_for_each_entry(vma,
1451 			&mapping->i_mmap_nonlinear, shared.nonlinear) {
1452 
1453 			cursor = (unsigned long) vma->vm_private_data;
1454 			while (cursor < max_nl_cursor &&
1455 				cursor < vma->vm_end - vma->vm_start) {
1456 				if (try_to_unmap_cluster(cursor, &mapcount,
1457 						vma, page) == SWAP_MLOCK)
1458 					ret = SWAP_MLOCK;
1459 				cursor += CLUSTER_SIZE;
1460 				vma->vm_private_data = (void *) cursor;
1461 				if ((int)mapcount <= 0)
1462 					return ret;
1463 			}
1464 			vma->vm_private_data = (void *) max_nl_cursor;
1465 		}
1466 		cond_resched();
1467 		max_nl_cursor += CLUSTER_SIZE;
1468 	} while (max_nl_cursor <= max_nl_size);
1469 
1470 	/*
1471 	 * Don't loop forever (perhaps all the remaining pages are
1472 	 * in locked vmas).  Reset cursor on all unreserved nonlinear
1473 	 * vmas, now forgetting on which ones it had fallen behind.
1474 	 */
1475 	list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.nonlinear)
1476 		vma->vm_private_data = NULL;
1477 
1478 	return ret;
1479 }
1480 
1481 bool is_vma_temporary_stack(struct vm_area_struct *vma)
1482 {
1483 	int maybe_stack = vma->vm_flags & (VM_GROWSDOWN | VM_GROWSUP);
1484 
1485 	if (!maybe_stack)
1486 		return false;
1487 
1488 	if ((vma->vm_flags & VM_STACK_INCOMPLETE_SETUP) ==
1489 						VM_STACK_INCOMPLETE_SETUP)
1490 		return true;
1491 
1492 	return false;
1493 }
1494 
1495 static bool invalid_migration_vma(struct vm_area_struct *vma, void *arg)
1496 {
1497 	return is_vma_temporary_stack(vma);
1498 }
1499 
1500 static int page_not_mapped(struct page *page)
1501 {
1502 	return !page_mapped(page);
1503 };
1504 
1505 /**
1506  * try_to_unmap - try to remove all page table mappings to a page
1507  * @page: the page to get unmapped
1508  * @flags: action and flags
1509  *
1510  * Tries to remove all the page table entries which are mapping this
1511  * page, used in the pageout path.  Caller must hold the page lock.
1512  * Return values are:
1513  *
1514  * SWAP_SUCCESS	- we succeeded in removing all mappings
1515  * SWAP_AGAIN	- we missed a mapping, try again later
1516  * SWAP_FAIL	- the page is unswappable
1517  * SWAP_MLOCK	- page is mlocked.
1518  */
1519 int try_to_unmap(struct page *page, enum ttu_flags flags)
1520 {
1521 	int ret;
1522 	struct rmap_walk_control rwc = {
1523 		.rmap_one = try_to_unmap_one,
1524 		.arg = (void *)flags,
1525 		.done = page_not_mapped,
1526 		.file_nonlinear = try_to_unmap_nonlinear,
1527 		.anon_lock = page_lock_anon_vma_read,
1528 	};
1529 
1530 	VM_BUG_ON_PAGE(!PageHuge(page) && PageTransHuge(page), page);
1531 
1532 	/*
1533 	 * During exec, a temporary VMA is setup and later moved.
1534 	 * The VMA is moved under the anon_vma lock but not the
1535 	 * page tables leading to a race where migration cannot
1536 	 * find the migration ptes. Rather than increasing the
1537 	 * locking requirements of exec(), migration skips
1538 	 * temporary VMAs until after exec() completes.
1539 	 */
1540 	if ((flags & TTU_MIGRATION) && !PageKsm(page) && PageAnon(page))
1541 		rwc.invalid_vma = invalid_migration_vma;
1542 
1543 	ret = rmap_walk(page, &rwc);
1544 
1545 	if (ret != SWAP_MLOCK && !page_mapped(page))
1546 		ret = SWAP_SUCCESS;
1547 	return ret;
1548 }
1549 
1550 /**
1551  * try_to_munlock - try to munlock a page
1552  * @page: the page to be munlocked
1553  *
1554  * Called from munlock code.  Checks all of the VMAs mapping the page
1555  * to make sure nobody else has this page mlocked. The page will be
1556  * returned with PG_mlocked cleared if no other vmas have it mlocked.
1557  *
1558  * Return values are:
1559  *
1560  * SWAP_AGAIN	- no vma is holding page mlocked, or,
1561  * SWAP_AGAIN	- page mapped in mlocked vma -- couldn't acquire mmap sem
1562  * SWAP_FAIL	- page cannot be located at present
1563  * SWAP_MLOCK	- page is now mlocked.
1564  */
1565 int try_to_munlock(struct page *page)
1566 {
1567 	int ret;
1568 	struct rmap_walk_control rwc = {
1569 		.rmap_one = try_to_unmap_one,
1570 		.arg = (void *)TTU_MUNLOCK,
1571 		.done = page_not_mapped,
1572 		/*
1573 		 * We don't bother to try to find the munlocked page in
1574 		 * nonlinears. It's costly. Instead, later, page reclaim logic
1575 		 * may call try_to_unmap() and recover PG_mlocked lazily.
1576 		 */
1577 		.file_nonlinear = NULL,
1578 		.anon_lock = page_lock_anon_vma_read,
1579 
1580 	};
1581 
1582 	VM_BUG_ON_PAGE(!PageLocked(page) || PageLRU(page), page);
1583 
1584 	ret = rmap_walk(page, &rwc);
1585 	return ret;
1586 }
1587 
1588 void __put_anon_vma(struct anon_vma *anon_vma)
1589 {
1590 	struct anon_vma *root = anon_vma->root;
1591 
1592 	anon_vma_free(anon_vma);
1593 	if (root != anon_vma && atomic_dec_and_test(&root->refcount))
1594 		anon_vma_free(root);
1595 }
1596 
1597 static struct anon_vma *rmap_walk_anon_lock(struct page *page,
1598 					struct rmap_walk_control *rwc)
1599 {
1600 	struct anon_vma *anon_vma;
1601 
1602 	if (rwc->anon_lock)
1603 		return rwc->anon_lock(page);
1604 
1605 	/*
1606 	 * Note: remove_migration_ptes() cannot use page_lock_anon_vma_read()
1607 	 * because that depends on page_mapped(); but not all its usages
1608 	 * are holding mmap_sem. Users without mmap_sem are required to
1609 	 * take a reference count to prevent the anon_vma disappearing
1610 	 */
1611 	anon_vma = page_anon_vma(page);
1612 	if (!anon_vma)
1613 		return NULL;
1614 
1615 	anon_vma_lock_read(anon_vma);
1616 	return anon_vma;
1617 }
1618 
1619 /*
1620  * rmap_walk_anon - do something to anonymous page using the object-based
1621  * rmap method
1622  * @page: the page to be handled
1623  * @rwc: control variable according to each walk type
1624  *
1625  * Find all the mappings of a page using the mapping pointer and the vma chains
1626  * contained in the anon_vma struct it points to.
1627  *
1628  * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1629  * where the page was found will be held for write.  So, we won't recheck
1630  * vm_flags for that VMA.  That should be OK, because that vma shouldn't be
1631  * LOCKED.
1632  */
1633 static int rmap_walk_anon(struct page *page, struct rmap_walk_control *rwc)
1634 {
1635 	struct anon_vma *anon_vma;
1636 	pgoff_t pgoff = page_to_pgoff(page);
1637 	struct anon_vma_chain *avc;
1638 	int ret = SWAP_AGAIN;
1639 
1640 	anon_vma = rmap_walk_anon_lock(page, rwc);
1641 	if (!anon_vma)
1642 		return ret;
1643 
1644 	anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1645 		struct vm_area_struct *vma = avc->vma;
1646 		unsigned long address = vma_address(page, vma);
1647 
1648 		if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
1649 			continue;
1650 
1651 		ret = rwc->rmap_one(page, vma, address, rwc->arg);
1652 		if (ret != SWAP_AGAIN)
1653 			break;
1654 		if (rwc->done && rwc->done(page))
1655 			break;
1656 	}
1657 	anon_vma_unlock_read(anon_vma);
1658 	return ret;
1659 }
1660 
1661 /*
1662  * rmap_walk_file - do something to file page using the object-based rmap method
1663  * @page: the page to be handled
1664  * @rwc: control variable according to each walk type
1665  *
1666  * Find all the mappings of a page using the mapping pointer and the vma chains
1667  * contained in the address_space struct it points to.
1668  *
1669  * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1670  * where the page was found will be held for write.  So, we won't recheck
1671  * vm_flags for that VMA.  That should be OK, because that vma shouldn't be
1672  * LOCKED.
1673  */
1674 static int rmap_walk_file(struct page *page, struct rmap_walk_control *rwc)
1675 {
1676 	struct address_space *mapping = page->mapping;
1677 	pgoff_t pgoff = page_to_pgoff(page);
1678 	struct vm_area_struct *vma;
1679 	int ret = SWAP_AGAIN;
1680 
1681 	/*
1682 	 * The page lock not only makes sure that page->mapping cannot
1683 	 * suddenly be NULLified by truncation, it makes sure that the
1684 	 * structure at mapping cannot be freed and reused yet,
1685 	 * so we can safely take mapping->i_mmap_mutex.
1686 	 */
1687 	VM_BUG_ON_PAGE(!PageLocked(page), page);
1688 
1689 	if (!mapping)
1690 		return ret;
1691 	mutex_lock(&mapping->i_mmap_mutex);
1692 	vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) {
1693 		unsigned long address = vma_address(page, vma);
1694 
1695 		if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
1696 			continue;
1697 
1698 		ret = rwc->rmap_one(page, vma, address, rwc->arg);
1699 		if (ret != SWAP_AGAIN)
1700 			goto done;
1701 		if (rwc->done && rwc->done(page))
1702 			goto done;
1703 	}
1704 
1705 	if (!rwc->file_nonlinear)
1706 		goto done;
1707 
1708 	if (list_empty(&mapping->i_mmap_nonlinear))
1709 		goto done;
1710 
1711 	ret = rwc->file_nonlinear(page, mapping, rwc->arg);
1712 
1713 done:
1714 	mutex_unlock(&mapping->i_mmap_mutex);
1715 	return ret;
1716 }
1717 
1718 int rmap_walk(struct page *page, struct rmap_walk_control *rwc)
1719 {
1720 	if (unlikely(PageKsm(page)))
1721 		return rmap_walk_ksm(page, rwc);
1722 	else if (PageAnon(page))
1723 		return rmap_walk_anon(page, rwc);
1724 	else
1725 		return rmap_walk_file(page, rwc);
1726 }
1727 
1728 #ifdef CONFIG_HUGETLB_PAGE
1729 /*
1730  * The following three functions are for anonymous (private mapped) hugepages.
1731  * Unlike common anonymous pages, anonymous hugepages have no accounting code
1732  * and no lru code, because we handle hugepages differently from common pages.
1733  */
1734 static void __hugepage_set_anon_rmap(struct page *page,
1735 	struct vm_area_struct *vma, unsigned long address, int exclusive)
1736 {
1737 	struct anon_vma *anon_vma = vma->anon_vma;
1738 
1739 	BUG_ON(!anon_vma);
1740 
1741 	if (PageAnon(page))
1742 		return;
1743 	if (!exclusive)
1744 		anon_vma = anon_vma->root;
1745 
1746 	anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
1747 	page->mapping = (struct address_space *) anon_vma;
1748 	page->index = linear_page_index(vma, address);
1749 }
1750 
1751 void hugepage_add_anon_rmap(struct page *page,
1752 			    struct vm_area_struct *vma, unsigned long address)
1753 {
1754 	struct anon_vma *anon_vma = vma->anon_vma;
1755 	int first;
1756 
1757 	BUG_ON(!PageLocked(page));
1758 	BUG_ON(!anon_vma);
1759 	/* address might be in next vma when migration races vma_adjust */
1760 	first = atomic_inc_and_test(&page->_mapcount);
1761 	if (first)
1762 		__hugepage_set_anon_rmap(page, vma, address, 0);
1763 }
1764 
1765 void hugepage_add_new_anon_rmap(struct page *page,
1766 			struct vm_area_struct *vma, unsigned long address)
1767 {
1768 	BUG_ON(address < vma->vm_start || address >= vma->vm_end);
1769 	atomic_set(&page->_mapcount, 0);
1770 	__hugepage_set_anon_rmap(page, vma, address, 1);
1771 }
1772 #endif /* CONFIG_HUGETLB_PAGE */
1773