xref: /linux/mm/rmap.c (revision bd628c1bed7902ec1f24ba0fe70758949146abbe)
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  *       hugetlbfs_i_mmap_rwsem_key (in huge_pmd_share)
27  *         mapping->i_mmap_rwsem
28  *           hugetlb_fault_mutex (hugetlbfs specific page fault mutex)
29  *           anon_vma->rwsem
30  *             mm->page_table_lock or pte_lock
31  *               zone_lru_lock (in mark_page_accessed, isolate_lru_page)
32  *               swap_lock (in swap_duplicate, swap_info_get)
33  *                 mmlist_lock (in mmput, drain_mmlist and others)
34  *                 mapping->private_lock (in __set_page_dirty_buffers)
35  *                   mem_cgroup_{begin,end}_page_stat (memcg->move_lock)
36  *                     i_pages lock (widely used)
37  *                 inode->i_lock (in set_page_dirty's __mark_inode_dirty)
38  *                 bdi.wb->list_lock (in set_page_dirty's __mark_inode_dirty)
39  *                   sb_lock (within inode_lock in fs/fs-writeback.c)
40  *                   i_pages lock (widely used, in set_page_dirty,
41  *                             in arch-dependent flush_dcache_mmap_lock,
42  *                             within bdi.wb->list_lock in __sync_single_inode)
43  *
44  * anon_vma->rwsem,mapping->i_mutex      (memory_failure, collect_procs_anon)
45  *   ->tasklist_lock
46  *     pte map lock
47  */
48 
49 #include <linux/mm.h>
50 #include <linux/sched/mm.h>
51 #include <linux/sched/task.h>
52 #include <linux/pagemap.h>
53 #include <linux/swap.h>
54 #include <linux/swapops.h>
55 #include <linux/slab.h>
56 #include <linux/init.h>
57 #include <linux/ksm.h>
58 #include <linux/rmap.h>
59 #include <linux/rcupdate.h>
60 #include <linux/export.h>
61 #include <linux/memcontrol.h>
62 #include <linux/mmu_notifier.h>
63 #include <linux/migrate.h>
64 #include <linux/hugetlb.h>
65 #include <linux/backing-dev.h>
66 #include <linux/page_idle.h>
67 #include <linux/memremap.h>
68 #include <linux/userfaultfd_k.h>
69 
70 #include <asm/tlbflush.h>
71 
72 #include <trace/events/tlb.h>
73 
74 #include "internal.h"
75 
76 static struct kmem_cache *anon_vma_cachep;
77 static struct kmem_cache *anon_vma_chain_cachep;
78 
79 static inline struct anon_vma *anon_vma_alloc(void)
80 {
81 	struct anon_vma *anon_vma;
82 
83 	anon_vma = kmem_cache_alloc(anon_vma_cachep, GFP_KERNEL);
84 	if (anon_vma) {
85 		atomic_set(&anon_vma->refcount, 1);
86 		anon_vma->degree = 1;	/* Reference for first vma */
87 		anon_vma->parent = anon_vma;
88 		/*
89 		 * Initialise the anon_vma root to point to itself. If called
90 		 * from fork, the root will be reset to the parents anon_vma.
91 		 */
92 		anon_vma->root = anon_vma;
93 	}
94 
95 	return anon_vma;
96 }
97 
98 static inline void anon_vma_free(struct anon_vma *anon_vma)
99 {
100 	VM_BUG_ON(atomic_read(&anon_vma->refcount));
101 
102 	/*
103 	 * Synchronize against page_lock_anon_vma_read() such that
104 	 * we can safely hold the lock without the anon_vma getting
105 	 * freed.
106 	 *
107 	 * Relies on the full mb implied by the atomic_dec_and_test() from
108 	 * put_anon_vma() against the acquire barrier implied by
109 	 * down_read_trylock() from page_lock_anon_vma_read(). This orders:
110 	 *
111 	 * page_lock_anon_vma_read()	VS	put_anon_vma()
112 	 *   down_read_trylock()		  atomic_dec_and_test()
113 	 *   LOCK				  MB
114 	 *   atomic_read()			  rwsem_is_locked()
115 	 *
116 	 * LOCK should suffice since the actual taking of the lock must
117 	 * happen _before_ what follows.
118 	 */
119 	might_sleep();
120 	if (rwsem_is_locked(&anon_vma->root->rwsem)) {
121 		anon_vma_lock_write(anon_vma);
122 		anon_vma_unlock_write(anon_vma);
123 	}
124 
125 	kmem_cache_free(anon_vma_cachep, anon_vma);
126 }
127 
128 static inline struct anon_vma_chain *anon_vma_chain_alloc(gfp_t gfp)
129 {
130 	return kmem_cache_alloc(anon_vma_chain_cachep, gfp);
131 }
132 
133 static void anon_vma_chain_free(struct anon_vma_chain *anon_vma_chain)
134 {
135 	kmem_cache_free(anon_vma_chain_cachep, anon_vma_chain);
136 }
137 
138 static void anon_vma_chain_link(struct vm_area_struct *vma,
139 				struct anon_vma_chain *avc,
140 				struct anon_vma *anon_vma)
141 {
142 	avc->vma = vma;
143 	avc->anon_vma = anon_vma;
144 	list_add(&avc->same_vma, &vma->anon_vma_chain);
145 	anon_vma_interval_tree_insert(avc, &anon_vma->rb_root);
146 }
147 
148 /**
149  * __anon_vma_prepare - attach an anon_vma to a memory region
150  * @vma: the memory region in question
151  *
152  * This makes sure the memory mapping described by 'vma' has
153  * an 'anon_vma' attached to it, so that we can associate the
154  * anonymous pages mapped into it with that anon_vma.
155  *
156  * The common case will be that we already have one, which
157  * is handled inline by anon_vma_prepare(). But if
158  * not we either need to find an adjacent mapping that we
159  * can re-use the anon_vma from (very common when the only
160  * reason for splitting a vma has been mprotect()), or we
161  * allocate a new one.
162  *
163  * Anon-vma allocations are very subtle, because we may have
164  * optimistically looked up an anon_vma in page_lock_anon_vma_read()
165  * and that may actually touch the spinlock even in the newly
166  * allocated vma (it depends on RCU to make sure that the
167  * anon_vma isn't actually destroyed).
168  *
169  * As a result, we need to do proper anon_vma locking even
170  * for the new allocation. At the same time, we do not want
171  * to do any locking for the common case of already having
172  * an anon_vma.
173  *
174  * This must be called with the mmap_sem held for reading.
175  */
176 int __anon_vma_prepare(struct vm_area_struct *vma)
177 {
178 	struct mm_struct *mm = vma->vm_mm;
179 	struct anon_vma *anon_vma, *allocated;
180 	struct anon_vma_chain *avc;
181 
182 	might_sleep();
183 
184 	avc = anon_vma_chain_alloc(GFP_KERNEL);
185 	if (!avc)
186 		goto out_enomem;
187 
188 	anon_vma = find_mergeable_anon_vma(vma);
189 	allocated = NULL;
190 	if (!anon_vma) {
191 		anon_vma = anon_vma_alloc();
192 		if (unlikely(!anon_vma))
193 			goto out_enomem_free_avc;
194 		allocated = anon_vma;
195 	}
196 
197 	anon_vma_lock_write(anon_vma);
198 	/* page_table_lock to protect against threads */
199 	spin_lock(&mm->page_table_lock);
200 	if (likely(!vma->anon_vma)) {
201 		vma->anon_vma = anon_vma;
202 		anon_vma_chain_link(vma, avc, anon_vma);
203 		/* vma reference or self-parent link for new root */
204 		anon_vma->degree++;
205 		allocated = NULL;
206 		avc = NULL;
207 	}
208 	spin_unlock(&mm->page_table_lock);
209 	anon_vma_unlock_write(anon_vma);
210 
211 	if (unlikely(allocated))
212 		put_anon_vma(allocated);
213 	if (unlikely(avc))
214 		anon_vma_chain_free(avc);
215 
216 	return 0;
217 
218  out_enomem_free_avc:
219 	anon_vma_chain_free(avc);
220  out_enomem:
221 	return -ENOMEM;
222 }
223 
224 /*
225  * This is a useful helper function for locking the anon_vma root as
226  * we traverse the vma->anon_vma_chain, looping over anon_vma's that
227  * have the same vma.
228  *
229  * Such anon_vma's should have the same root, so you'd expect to see
230  * just a single mutex_lock for the whole traversal.
231  */
232 static inline struct anon_vma *lock_anon_vma_root(struct anon_vma *root, struct anon_vma *anon_vma)
233 {
234 	struct anon_vma *new_root = anon_vma->root;
235 	if (new_root != root) {
236 		if (WARN_ON_ONCE(root))
237 			up_write(&root->rwsem);
238 		root = new_root;
239 		down_write(&root->rwsem);
240 	}
241 	return root;
242 }
243 
244 static inline void unlock_anon_vma_root(struct anon_vma *root)
245 {
246 	if (root)
247 		up_write(&root->rwsem);
248 }
249 
250 /*
251  * Attach the anon_vmas from src to dst.
252  * Returns 0 on success, -ENOMEM on failure.
253  *
254  * If dst->anon_vma is NULL this function tries to find and reuse existing
255  * anon_vma which has no vmas and only one child anon_vma. This prevents
256  * degradation of anon_vma hierarchy to endless linear chain in case of
257  * constantly forking task. On the other hand, an anon_vma with more than one
258  * child isn't reused even if there was no alive vma, thus rmap walker has a
259  * good chance of avoiding scanning the whole hierarchy when it searches where
260  * page is mapped.
261  */
262 int anon_vma_clone(struct vm_area_struct *dst, struct vm_area_struct *src)
263 {
264 	struct anon_vma_chain *avc, *pavc;
265 	struct anon_vma *root = NULL;
266 
267 	list_for_each_entry_reverse(pavc, &src->anon_vma_chain, same_vma) {
268 		struct anon_vma *anon_vma;
269 
270 		avc = anon_vma_chain_alloc(GFP_NOWAIT | __GFP_NOWARN);
271 		if (unlikely(!avc)) {
272 			unlock_anon_vma_root(root);
273 			root = NULL;
274 			avc = anon_vma_chain_alloc(GFP_KERNEL);
275 			if (!avc)
276 				goto enomem_failure;
277 		}
278 		anon_vma = pavc->anon_vma;
279 		root = lock_anon_vma_root(root, anon_vma);
280 		anon_vma_chain_link(dst, avc, anon_vma);
281 
282 		/*
283 		 * Reuse existing anon_vma if its degree lower than two,
284 		 * that means it has no vma and only one anon_vma child.
285 		 *
286 		 * Do not chose parent anon_vma, otherwise first child
287 		 * will always reuse it. Root anon_vma is never reused:
288 		 * it has self-parent reference and at least one child.
289 		 */
290 		if (!dst->anon_vma && anon_vma != src->anon_vma &&
291 				anon_vma->degree < 2)
292 			dst->anon_vma = anon_vma;
293 	}
294 	if (dst->anon_vma)
295 		dst->anon_vma->degree++;
296 	unlock_anon_vma_root(root);
297 	return 0;
298 
299  enomem_failure:
300 	/*
301 	 * dst->anon_vma is dropped here otherwise its degree can be incorrectly
302 	 * decremented in unlink_anon_vmas().
303 	 * We can safely do this because callers of anon_vma_clone() don't care
304 	 * about dst->anon_vma if anon_vma_clone() failed.
305 	 */
306 	dst->anon_vma = NULL;
307 	unlink_anon_vmas(dst);
308 	return -ENOMEM;
309 }
310 
311 /*
312  * Attach vma to its own anon_vma, as well as to the anon_vmas that
313  * the corresponding VMA in the parent process is attached to.
314  * Returns 0 on success, non-zero on failure.
315  */
316 int anon_vma_fork(struct vm_area_struct *vma, struct vm_area_struct *pvma)
317 {
318 	struct anon_vma_chain *avc;
319 	struct anon_vma *anon_vma;
320 	int error;
321 
322 	/* Don't bother if the parent process has no anon_vma here. */
323 	if (!pvma->anon_vma)
324 		return 0;
325 
326 	/* Drop inherited anon_vma, we'll reuse existing or allocate new. */
327 	vma->anon_vma = NULL;
328 
329 	/*
330 	 * First, attach the new VMA to the parent VMA's anon_vmas,
331 	 * so rmap can find non-COWed pages in child processes.
332 	 */
333 	error = anon_vma_clone(vma, pvma);
334 	if (error)
335 		return error;
336 
337 	/* An existing anon_vma has been reused, all done then. */
338 	if (vma->anon_vma)
339 		return 0;
340 
341 	/* Then add our own anon_vma. */
342 	anon_vma = anon_vma_alloc();
343 	if (!anon_vma)
344 		goto out_error;
345 	avc = anon_vma_chain_alloc(GFP_KERNEL);
346 	if (!avc)
347 		goto out_error_free_anon_vma;
348 
349 	/*
350 	 * The root anon_vma's spinlock is the lock actually used when we
351 	 * lock any of the anon_vmas in this anon_vma tree.
352 	 */
353 	anon_vma->root = pvma->anon_vma->root;
354 	anon_vma->parent = pvma->anon_vma;
355 	/*
356 	 * With refcounts, an anon_vma can stay around longer than the
357 	 * process it belongs to. The root anon_vma needs to be pinned until
358 	 * this anon_vma is freed, because the lock lives in the root.
359 	 */
360 	get_anon_vma(anon_vma->root);
361 	/* Mark this anon_vma as the one where our new (COWed) pages go. */
362 	vma->anon_vma = anon_vma;
363 	anon_vma_lock_write(anon_vma);
364 	anon_vma_chain_link(vma, avc, anon_vma);
365 	anon_vma->parent->degree++;
366 	anon_vma_unlock_write(anon_vma);
367 
368 	return 0;
369 
370  out_error_free_anon_vma:
371 	put_anon_vma(anon_vma);
372  out_error:
373 	unlink_anon_vmas(vma);
374 	return -ENOMEM;
375 }
376 
377 void unlink_anon_vmas(struct vm_area_struct *vma)
378 {
379 	struct anon_vma_chain *avc, *next;
380 	struct anon_vma *root = NULL;
381 
382 	/*
383 	 * Unlink each anon_vma chained to the VMA.  This list is ordered
384 	 * from newest to oldest, ensuring the root anon_vma gets freed last.
385 	 */
386 	list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
387 		struct anon_vma *anon_vma = avc->anon_vma;
388 
389 		root = lock_anon_vma_root(root, anon_vma);
390 		anon_vma_interval_tree_remove(avc, &anon_vma->rb_root);
391 
392 		/*
393 		 * Leave empty anon_vmas on the list - we'll need
394 		 * to free them outside the lock.
395 		 */
396 		if (RB_EMPTY_ROOT(&anon_vma->rb_root.rb_root)) {
397 			anon_vma->parent->degree--;
398 			continue;
399 		}
400 
401 		list_del(&avc->same_vma);
402 		anon_vma_chain_free(avc);
403 	}
404 	if (vma->anon_vma)
405 		vma->anon_vma->degree--;
406 	unlock_anon_vma_root(root);
407 
408 	/*
409 	 * Iterate the list once more, it now only contains empty and unlinked
410 	 * anon_vmas, destroy them. Could not do before due to __put_anon_vma()
411 	 * needing to write-acquire the anon_vma->root->rwsem.
412 	 */
413 	list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
414 		struct anon_vma *anon_vma = avc->anon_vma;
415 
416 		VM_WARN_ON(anon_vma->degree);
417 		put_anon_vma(anon_vma);
418 
419 		list_del(&avc->same_vma);
420 		anon_vma_chain_free(avc);
421 	}
422 }
423 
424 static void anon_vma_ctor(void *data)
425 {
426 	struct anon_vma *anon_vma = data;
427 
428 	init_rwsem(&anon_vma->rwsem);
429 	atomic_set(&anon_vma->refcount, 0);
430 	anon_vma->rb_root = RB_ROOT_CACHED;
431 }
432 
433 void __init anon_vma_init(void)
434 {
435 	anon_vma_cachep = kmem_cache_create("anon_vma", sizeof(struct anon_vma),
436 			0, SLAB_TYPESAFE_BY_RCU|SLAB_PANIC|SLAB_ACCOUNT,
437 			anon_vma_ctor);
438 	anon_vma_chain_cachep = KMEM_CACHE(anon_vma_chain,
439 			SLAB_PANIC|SLAB_ACCOUNT);
440 }
441 
442 /*
443  * Getting a lock on a stable anon_vma from a page off the LRU is tricky!
444  *
445  * Since there is no serialization what so ever against page_remove_rmap()
446  * the best this function can do is return a locked anon_vma that might
447  * have been relevant to this page.
448  *
449  * The page might have been remapped to a different anon_vma or the anon_vma
450  * returned may already be freed (and even reused).
451  *
452  * In case it was remapped to a different anon_vma, the new anon_vma will be a
453  * child of the old anon_vma, and the anon_vma lifetime rules will therefore
454  * ensure that any anon_vma obtained from the page will still be valid for as
455  * long as we observe page_mapped() [ hence all those page_mapped() tests ].
456  *
457  * All users of this function must be very careful when walking the anon_vma
458  * chain and verify that the page in question is indeed mapped in it
459  * [ something equivalent to page_mapped_in_vma() ].
460  *
461  * Since anon_vma's slab is DESTROY_BY_RCU and we know from page_remove_rmap()
462  * that the anon_vma pointer from page->mapping is valid if there is a
463  * mapcount, we can dereference the anon_vma after observing those.
464  */
465 struct anon_vma *page_get_anon_vma(struct page *page)
466 {
467 	struct anon_vma *anon_vma = NULL;
468 	unsigned long anon_mapping;
469 
470 	rcu_read_lock();
471 	anon_mapping = (unsigned long)READ_ONCE(page->mapping);
472 	if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
473 		goto out;
474 	if (!page_mapped(page))
475 		goto out;
476 
477 	anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
478 	if (!atomic_inc_not_zero(&anon_vma->refcount)) {
479 		anon_vma = NULL;
480 		goto out;
481 	}
482 
483 	/*
484 	 * If this page is still mapped, then its anon_vma cannot have been
485 	 * freed.  But if it has been unmapped, we have no security against the
486 	 * anon_vma structure being freed and reused (for another anon_vma:
487 	 * SLAB_TYPESAFE_BY_RCU guarantees that - so the atomic_inc_not_zero()
488 	 * above cannot corrupt).
489 	 */
490 	if (!page_mapped(page)) {
491 		rcu_read_unlock();
492 		put_anon_vma(anon_vma);
493 		return NULL;
494 	}
495 out:
496 	rcu_read_unlock();
497 
498 	return anon_vma;
499 }
500 
501 /*
502  * Similar to page_get_anon_vma() except it locks the anon_vma.
503  *
504  * Its a little more complex as it tries to keep the fast path to a single
505  * atomic op -- the trylock. If we fail the trylock, we fall back to getting a
506  * reference like with page_get_anon_vma() and then block on the mutex.
507  */
508 struct anon_vma *page_lock_anon_vma_read(struct page *page)
509 {
510 	struct anon_vma *anon_vma = NULL;
511 	struct anon_vma *root_anon_vma;
512 	unsigned long anon_mapping;
513 
514 	rcu_read_lock();
515 	anon_mapping = (unsigned long)READ_ONCE(page->mapping);
516 	if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
517 		goto out;
518 	if (!page_mapped(page))
519 		goto out;
520 
521 	anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
522 	root_anon_vma = READ_ONCE(anon_vma->root);
523 	if (down_read_trylock(&root_anon_vma->rwsem)) {
524 		/*
525 		 * If the page is still mapped, then this anon_vma is still
526 		 * its anon_vma, and holding the mutex ensures that it will
527 		 * not go away, see anon_vma_free().
528 		 */
529 		if (!page_mapped(page)) {
530 			up_read(&root_anon_vma->rwsem);
531 			anon_vma = NULL;
532 		}
533 		goto out;
534 	}
535 
536 	/* trylock failed, we got to sleep */
537 	if (!atomic_inc_not_zero(&anon_vma->refcount)) {
538 		anon_vma = NULL;
539 		goto out;
540 	}
541 
542 	if (!page_mapped(page)) {
543 		rcu_read_unlock();
544 		put_anon_vma(anon_vma);
545 		return NULL;
546 	}
547 
548 	/* we pinned the anon_vma, its safe to sleep */
549 	rcu_read_unlock();
550 	anon_vma_lock_read(anon_vma);
551 
552 	if (atomic_dec_and_test(&anon_vma->refcount)) {
553 		/*
554 		 * Oops, we held the last refcount, release the lock
555 		 * and bail -- can't simply use put_anon_vma() because
556 		 * we'll deadlock on the anon_vma_lock_write() recursion.
557 		 */
558 		anon_vma_unlock_read(anon_vma);
559 		__put_anon_vma(anon_vma);
560 		anon_vma = NULL;
561 	}
562 
563 	return anon_vma;
564 
565 out:
566 	rcu_read_unlock();
567 	return anon_vma;
568 }
569 
570 void page_unlock_anon_vma_read(struct anon_vma *anon_vma)
571 {
572 	anon_vma_unlock_read(anon_vma);
573 }
574 
575 #ifdef CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH
576 /*
577  * Flush TLB entries for recently unmapped pages from remote CPUs. It is
578  * important if a PTE was dirty when it was unmapped that it's flushed
579  * before any IO is initiated on the page to prevent lost writes. Similarly,
580  * it must be flushed before freeing to prevent data leakage.
581  */
582 void try_to_unmap_flush(void)
583 {
584 	struct tlbflush_unmap_batch *tlb_ubc = &current->tlb_ubc;
585 
586 	if (!tlb_ubc->flush_required)
587 		return;
588 
589 	arch_tlbbatch_flush(&tlb_ubc->arch);
590 	tlb_ubc->flush_required = false;
591 	tlb_ubc->writable = false;
592 }
593 
594 /* Flush iff there are potentially writable TLB entries that can race with IO */
595 void try_to_unmap_flush_dirty(void)
596 {
597 	struct tlbflush_unmap_batch *tlb_ubc = &current->tlb_ubc;
598 
599 	if (tlb_ubc->writable)
600 		try_to_unmap_flush();
601 }
602 
603 static void set_tlb_ubc_flush_pending(struct mm_struct *mm, bool writable)
604 {
605 	struct tlbflush_unmap_batch *tlb_ubc = &current->tlb_ubc;
606 
607 	arch_tlbbatch_add_mm(&tlb_ubc->arch, mm);
608 	tlb_ubc->flush_required = true;
609 
610 	/*
611 	 * Ensure compiler does not re-order the setting of tlb_flush_batched
612 	 * before the PTE is cleared.
613 	 */
614 	barrier();
615 	mm->tlb_flush_batched = true;
616 
617 	/*
618 	 * If the PTE was dirty then it's best to assume it's writable. The
619 	 * caller must use try_to_unmap_flush_dirty() or try_to_unmap_flush()
620 	 * before the page is queued for IO.
621 	 */
622 	if (writable)
623 		tlb_ubc->writable = true;
624 }
625 
626 /*
627  * Returns true if the TLB flush should be deferred to the end of a batch of
628  * unmap operations to reduce IPIs.
629  */
630 static bool should_defer_flush(struct mm_struct *mm, enum ttu_flags flags)
631 {
632 	bool should_defer = false;
633 
634 	if (!(flags & TTU_BATCH_FLUSH))
635 		return false;
636 
637 	/* If remote CPUs need to be flushed then defer batch the flush */
638 	if (cpumask_any_but(mm_cpumask(mm), get_cpu()) < nr_cpu_ids)
639 		should_defer = true;
640 	put_cpu();
641 
642 	return should_defer;
643 }
644 
645 /*
646  * Reclaim unmaps pages under the PTL but do not flush the TLB prior to
647  * releasing the PTL if TLB flushes are batched. It's possible for a parallel
648  * operation such as mprotect or munmap to race between reclaim unmapping
649  * the page and flushing the page. If this race occurs, it potentially allows
650  * access to data via a stale TLB entry. Tracking all mm's that have TLB
651  * batching in flight would be expensive during reclaim so instead track
652  * whether TLB batching occurred in the past and if so then do a flush here
653  * if required. This will cost one additional flush per reclaim cycle paid
654  * by the first operation at risk such as mprotect and mumap.
655  *
656  * This must be called under the PTL so that an access to tlb_flush_batched
657  * that is potentially a "reclaim vs mprotect/munmap/etc" race will synchronise
658  * via the PTL.
659  */
660 void flush_tlb_batched_pending(struct mm_struct *mm)
661 {
662 	if (mm->tlb_flush_batched) {
663 		flush_tlb_mm(mm);
664 
665 		/*
666 		 * Do not allow the compiler to re-order the clearing of
667 		 * tlb_flush_batched before the tlb is flushed.
668 		 */
669 		barrier();
670 		mm->tlb_flush_batched = false;
671 	}
672 }
673 #else
674 static void set_tlb_ubc_flush_pending(struct mm_struct *mm, bool writable)
675 {
676 }
677 
678 static bool should_defer_flush(struct mm_struct *mm, enum ttu_flags flags)
679 {
680 	return false;
681 }
682 #endif /* CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH */
683 
684 /*
685  * At what user virtual address is page expected in vma?
686  * Caller should check the page is actually part of the vma.
687  */
688 unsigned long page_address_in_vma(struct page *page, struct vm_area_struct *vma)
689 {
690 	unsigned long address;
691 	if (PageAnon(page)) {
692 		struct anon_vma *page__anon_vma = page_anon_vma(page);
693 		/*
694 		 * Note: swapoff's unuse_vma() is more efficient with this
695 		 * check, and needs it to match anon_vma when KSM is active.
696 		 */
697 		if (!vma->anon_vma || !page__anon_vma ||
698 		    vma->anon_vma->root != page__anon_vma->root)
699 			return -EFAULT;
700 	} else if (page->mapping) {
701 		if (!vma->vm_file || vma->vm_file->f_mapping != page->mapping)
702 			return -EFAULT;
703 	} else
704 		return -EFAULT;
705 	address = __vma_address(page, vma);
706 	if (unlikely(address < vma->vm_start || address >= vma->vm_end))
707 		return -EFAULT;
708 	return address;
709 }
710 
711 pmd_t *mm_find_pmd(struct mm_struct *mm, unsigned long address)
712 {
713 	pgd_t *pgd;
714 	p4d_t *p4d;
715 	pud_t *pud;
716 	pmd_t *pmd = NULL;
717 	pmd_t pmde;
718 
719 	pgd = pgd_offset(mm, address);
720 	if (!pgd_present(*pgd))
721 		goto out;
722 
723 	p4d = p4d_offset(pgd, address);
724 	if (!p4d_present(*p4d))
725 		goto out;
726 
727 	pud = pud_offset(p4d, address);
728 	if (!pud_present(*pud))
729 		goto out;
730 
731 	pmd = pmd_offset(pud, address);
732 	/*
733 	 * Some THP functions use the sequence pmdp_huge_clear_flush(), set_pmd_at()
734 	 * without holding anon_vma lock for write.  So when looking for a
735 	 * genuine pmde (in which to find pte), test present and !THP together.
736 	 */
737 	pmde = *pmd;
738 	barrier();
739 	if (!pmd_present(pmde) || pmd_trans_huge(pmde))
740 		pmd = NULL;
741 out:
742 	return pmd;
743 }
744 
745 struct page_referenced_arg {
746 	int mapcount;
747 	int referenced;
748 	unsigned long vm_flags;
749 	struct mem_cgroup *memcg;
750 };
751 /*
752  * arg: page_referenced_arg will be passed
753  */
754 static bool page_referenced_one(struct page *page, struct vm_area_struct *vma,
755 			unsigned long address, void *arg)
756 {
757 	struct page_referenced_arg *pra = arg;
758 	struct page_vma_mapped_walk pvmw = {
759 		.page = page,
760 		.vma = vma,
761 		.address = address,
762 	};
763 	int referenced = 0;
764 
765 	while (page_vma_mapped_walk(&pvmw)) {
766 		address = pvmw.address;
767 
768 		if (vma->vm_flags & VM_LOCKED) {
769 			page_vma_mapped_walk_done(&pvmw);
770 			pra->vm_flags |= VM_LOCKED;
771 			return false; /* To break the loop */
772 		}
773 
774 		if (pvmw.pte) {
775 			if (ptep_clear_flush_young_notify(vma, address,
776 						pvmw.pte)) {
777 				/*
778 				 * Don't treat a reference through
779 				 * a sequentially read mapping as such.
780 				 * If the page has been used in another mapping,
781 				 * we will catch it; if this other mapping is
782 				 * already gone, the unmap path will have set
783 				 * PG_referenced or activated the page.
784 				 */
785 				if (likely(!(vma->vm_flags & VM_SEQ_READ)))
786 					referenced++;
787 			}
788 		} else if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE)) {
789 			if (pmdp_clear_flush_young_notify(vma, address,
790 						pvmw.pmd))
791 				referenced++;
792 		} else {
793 			/* unexpected pmd-mapped page? */
794 			WARN_ON_ONCE(1);
795 		}
796 
797 		pra->mapcount--;
798 	}
799 
800 	if (referenced)
801 		clear_page_idle(page);
802 	if (test_and_clear_page_young(page))
803 		referenced++;
804 
805 	if (referenced) {
806 		pra->referenced++;
807 		pra->vm_flags |= vma->vm_flags;
808 	}
809 
810 	if (!pra->mapcount)
811 		return false; /* To break the loop */
812 
813 	return true;
814 }
815 
816 static bool invalid_page_referenced_vma(struct vm_area_struct *vma, void *arg)
817 {
818 	struct page_referenced_arg *pra = arg;
819 	struct mem_cgroup *memcg = pra->memcg;
820 
821 	if (!mm_match_cgroup(vma->vm_mm, memcg))
822 		return true;
823 
824 	return false;
825 }
826 
827 /**
828  * page_referenced - test if the page was referenced
829  * @page: the page to test
830  * @is_locked: caller holds lock on the page
831  * @memcg: target memory cgroup
832  * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
833  *
834  * Quick test_and_clear_referenced for all mappings to a page,
835  * returns the number of ptes which referenced the page.
836  */
837 int page_referenced(struct page *page,
838 		    int is_locked,
839 		    struct mem_cgroup *memcg,
840 		    unsigned long *vm_flags)
841 {
842 	int we_locked = 0;
843 	struct page_referenced_arg pra = {
844 		.mapcount = total_mapcount(page),
845 		.memcg = memcg,
846 	};
847 	struct rmap_walk_control rwc = {
848 		.rmap_one = page_referenced_one,
849 		.arg = (void *)&pra,
850 		.anon_lock = page_lock_anon_vma_read,
851 	};
852 
853 	*vm_flags = 0;
854 	if (!page_mapped(page))
855 		return 0;
856 
857 	if (!page_rmapping(page))
858 		return 0;
859 
860 	if (!is_locked && (!PageAnon(page) || PageKsm(page))) {
861 		we_locked = trylock_page(page);
862 		if (!we_locked)
863 			return 1;
864 	}
865 
866 	/*
867 	 * If we are reclaiming on behalf of a cgroup, skip
868 	 * counting on behalf of references from different
869 	 * cgroups
870 	 */
871 	if (memcg) {
872 		rwc.invalid_vma = invalid_page_referenced_vma;
873 	}
874 
875 	rmap_walk(page, &rwc);
876 	*vm_flags = pra.vm_flags;
877 
878 	if (we_locked)
879 		unlock_page(page);
880 
881 	return pra.referenced;
882 }
883 
884 static bool page_mkclean_one(struct page *page, struct vm_area_struct *vma,
885 			    unsigned long address, void *arg)
886 {
887 	struct page_vma_mapped_walk pvmw = {
888 		.page = page,
889 		.vma = vma,
890 		.address = address,
891 		.flags = PVMW_SYNC,
892 	};
893 	struct mmu_notifier_range range;
894 	int *cleaned = arg;
895 
896 	/*
897 	 * We have to assume the worse case ie pmd for invalidation. Note that
898 	 * the page can not be free from this function.
899 	 */
900 	mmu_notifier_range_init(&range, vma->vm_mm, address,
901 				min(vma->vm_end, address +
902 				    (PAGE_SIZE << compound_order(page))));
903 	mmu_notifier_invalidate_range_start(&range);
904 
905 	while (page_vma_mapped_walk(&pvmw)) {
906 		unsigned long cstart;
907 		int ret = 0;
908 
909 		cstart = address = pvmw.address;
910 		if (pvmw.pte) {
911 			pte_t entry;
912 			pte_t *pte = pvmw.pte;
913 
914 			if (!pte_dirty(*pte) && !pte_write(*pte))
915 				continue;
916 
917 			flush_cache_page(vma, address, pte_pfn(*pte));
918 			entry = ptep_clear_flush(vma, address, pte);
919 			entry = pte_wrprotect(entry);
920 			entry = pte_mkclean(entry);
921 			set_pte_at(vma->vm_mm, address, pte, entry);
922 			ret = 1;
923 		} else {
924 #ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
925 			pmd_t *pmd = pvmw.pmd;
926 			pmd_t entry;
927 
928 			if (!pmd_dirty(*pmd) && !pmd_write(*pmd))
929 				continue;
930 
931 			flush_cache_page(vma, address, page_to_pfn(page));
932 			entry = pmdp_huge_clear_flush(vma, address, pmd);
933 			entry = pmd_wrprotect(entry);
934 			entry = pmd_mkclean(entry);
935 			set_pmd_at(vma->vm_mm, address, pmd, entry);
936 			cstart &= PMD_MASK;
937 			ret = 1;
938 #else
939 			/* unexpected pmd-mapped page? */
940 			WARN_ON_ONCE(1);
941 #endif
942 		}
943 
944 		/*
945 		 * No need to call mmu_notifier_invalidate_range() as we are
946 		 * downgrading page table protection not changing it to point
947 		 * to a new page.
948 		 *
949 		 * See Documentation/vm/mmu_notifier.rst
950 		 */
951 		if (ret)
952 			(*cleaned)++;
953 	}
954 
955 	mmu_notifier_invalidate_range_end(&range);
956 
957 	return true;
958 }
959 
960 static bool invalid_mkclean_vma(struct vm_area_struct *vma, void *arg)
961 {
962 	if (vma->vm_flags & VM_SHARED)
963 		return false;
964 
965 	return true;
966 }
967 
968 int page_mkclean(struct page *page)
969 {
970 	int cleaned = 0;
971 	struct address_space *mapping;
972 	struct rmap_walk_control rwc = {
973 		.arg = (void *)&cleaned,
974 		.rmap_one = page_mkclean_one,
975 		.invalid_vma = invalid_mkclean_vma,
976 	};
977 
978 	BUG_ON(!PageLocked(page));
979 
980 	if (!page_mapped(page))
981 		return 0;
982 
983 	mapping = page_mapping(page);
984 	if (!mapping)
985 		return 0;
986 
987 	rmap_walk(page, &rwc);
988 
989 	return cleaned;
990 }
991 EXPORT_SYMBOL_GPL(page_mkclean);
992 
993 /**
994  * page_move_anon_rmap - move a page to our anon_vma
995  * @page:	the page to move to our anon_vma
996  * @vma:	the vma the page belongs to
997  *
998  * When a page belongs exclusively to one process after a COW event,
999  * that page can be moved into the anon_vma that belongs to just that
1000  * process, so the rmap code will not search the parent or sibling
1001  * processes.
1002  */
1003 void page_move_anon_rmap(struct page *page, struct vm_area_struct *vma)
1004 {
1005 	struct anon_vma *anon_vma = vma->anon_vma;
1006 
1007 	page = compound_head(page);
1008 
1009 	VM_BUG_ON_PAGE(!PageLocked(page), page);
1010 	VM_BUG_ON_VMA(!anon_vma, vma);
1011 
1012 	anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
1013 	/*
1014 	 * Ensure that anon_vma and the PAGE_MAPPING_ANON bit are written
1015 	 * simultaneously, so a concurrent reader (eg page_referenced()'s
1016 	 * PageAnon()) will not see one without the other.
1017 	 */
1018 	WRITE_ONCE(page->mapping, (struct address_space *) anon_vma);
1019 }
1020 
1021 /**
1022  * __page_set_anon_rmap - set up new anonymous rmap
1023  * @page:	Page or Hugepage to add to rmap
1024  * @vma:	VM area to add page to.
1025  * @address:	User virtual address of the mapping
1026  * @exclusive:	the page is exclusively owned by the current process
1027  */
1028 static void __page_set_anon_rmap(struct page *page,
1029 	struct vm_area_struct *vma, unsigned long address, int exclusive)
1030 {
1031 	struct anon_vma *anon_vma = vma->anon_vma;
1032 
1033 	BUG_ON(!anon_vma);
1034 
1035 	if (PageAnon(page))
1036 		return;
1037 
1038 	/*
1039 	 * If the page isn't exclusively mapped into this vma,
1040 	 * we must use the _oldest_ possible anon_vma for the
1041 	 * page mapping!
1042 	 */
1043 	if (!exclusive)
1044 		anon_vma = anon_vma->root;
1045 
1046 	anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
1047 	page->mapping = (struct address_space *) anon_vma;
1048 	page->index = linear_page_index(vma, address);
1049 }
1050 
1051 /**
1052  * __page_check_anon_rmap - sanity check anonymous rmap addition
1053  * @page:	the page to add the mapping to
1054  * @vma:	the vm area in which the mapping is added
1055  * @address:	the user virtual address mapped
1056  */
1057 static void __page_check_anon_rmap(struct page *page,
1058 	struct vm_area_struct *vma, unsigned long address)
1059 {
1060 #ifdef CONFIG_DEBUG_VM
1061 	/*
1062 	 * The page's anon-rmap details (mapping and index) are guaranteed to
1063 	 * be set up correctly at this point.
1064 	 *
1065 	 * We have exclusion against page_add_anon_rmap because the caller
1066 	 * always holds the page locked, except if called from page_dup_rmap,
1067 	 * in which case the page is already known to be setup.
1068 	 *
1069 	 * We have exclusion against page_add_new_anon_rmap because those pages
1070 	 * are initially only visible via the pagetables, and the pte is locked
1071 	 * over the call to page_add_new_anon_rmap.
1072 	 */
1073 	BUG_ON(page_anon_vma(page)->root != vma->anon_vma->root);
1074 	BUG_ON(page_to_pgoff(page) != linear_page_index(vma, address));
1075 #endif
1076 }
1077 
1078 /**
1079  * page_add_anon_rmap - add pte mapping to an anonymous page
1080  * @page:	the page to add the mapping to
1081  * @vma:	the vm area in which the mapping is added
1082  * @address:	the user virtual address mapped
1083  * @compound:	charge the page as compound or small page
1084  *
1085  * The caller needs to hold the pte lock, and the page must be locked in
1086  * the anon_vma case: to serialize mapping,index checking after setting,
1087  * and to ensure that PageAnon is not being upgraded racily to PageKsm
1088  * (but PageKsm is never downgraded to PageAnon).
1089  */
1090 void page_add_anon_rmap(struct page *page,
1091 	struct vm_area_struct *vma, unsigned long address, bool compound)
1092 {
1093 	do_page_add_anon_rmap(page, vma, address, compound ? RMAP_COMPOUND : 0);
1094 }
1095 
1096 /*
1097  * Special version of the above for do_swap_page, which often runs
1098  * into pages that are exclusively owned by the current process.
1099  * Everybody else should continue to use page_add_anon_rmap above.
1100  */
1101 void do_page_add_anon_rmap(struct page *page,
1102 	struct vm_area_struct *vma, unsigned long address, int flags)
1103 {
1104 	bool compound = flags & RMAP_COMPOUND;
1105 	bool first;
1106 
1107 	if (compound) {
1108 		atomic_t *mapcount;
1109 		VM_BUG_ON_PAGE(!PageLocked(page), page);
1110 		VM_BUG_ON_PAGE(!PageTransHuge(page), page);
1111 		mapcount = compound_mapcount_ptr(page);
1112 		first = atomic_inc_and_test(mapcount);
1113 	} else {
1114 		first = atomic_inc_and_test(&page->_mapcount);
1115 	}
1116 
1117 	if (first) {
1118 		int nr = compound ? hpage_nr_pages(page) : 1;
1119 		/*
1120 		 * We use the irq-unsafe __{inc|mod}_zone_page_stat because
1121 		 * these counters are not modified in interrupt context, and
1122 		 * pte lock(a spinlock) is held, which implies preemption
1123 		 * disabled.
1124 		 */
1125 		if (compound)
1126 			__inc_node_page_state(page, NR_ANON_THPS);
1127 		__mod_node_page_state(page_pgdat(page), NR_ANON_MAPPED, nr);
1128 	}
1129 	if (unlikely(PageKsm(page)))
1130 		return;
1131 
1132 	VM_BUG_ON_PAGE(!PageLocked(page), page);
1133 
1134 	/* address might be in next vma when migration races vma_adjust */
1135 	if (first)
1136 		__page_set_anon_rmap(page, vma, address,
1137 				flags & RMAP_EXCLUSIVE);
1138 	else
1139 		__page_check_anon_rmap(page, vma, address);
1140 }
1141 
1142 /**
1143  * page_add_new_anon_rmap - add pte mapping to a new anonymous page
1144  * @page:	the page to add the mapping to
1145  * @vma:	the vm area in which the mapping is added
1146  * @address:	the user virtual address mapped
1147  * @compound:	charge the page as compound or small page
1148  *
1149  * Same as page_add_anon_rmap but must only be called on *new* pages.
1150  * This means the inc-and-test can be bypassed.
1151  * Page does not have to be locked.
1152  */
1153 void page_add_new_anon_rmap(struct page *page,
1154 	struct vm_area_struct *vma, unsigned long address, bool compound)
1155 {
1156 	int nr = compound ? hpage_nr_pages(page) : 1;
1157 
1158 	VM_BUG_ON_VMA(address < vma->vm_start || address >= vma->vm_end, vma);
1159 	__SetPageSwapBacked(page);
1160 	if (compound) {
1161 		VM_BUG_ON_PAGE(!PageTransHuge(page), page);
1162 		/* increment count (starts at -1) */
1163 		atomic_set(compound_mapcount_ptr(page), 0);
1164 		__inc_node_page_state(page, NR_ANON_THPS);
1165 	} else {
1166 		/* Anon THP always mapped first with PMD */
1167 		VM_BUG_ON_PAGE(PageTransCompound(page), page);
1168 		/* increment count (starts at -1) */
1169 		atomic_set(&page->_mapcount, 0);
1170 	}
1171 	__mod_node_page_state(page_pgdat(page), NR_ANON_MAPPED, nr);
1172 	__page_set_anon_rmap(page, vma, address, 1);
1173 }
1174 
1175 /**
1176  * page_add_file_rmap - add pte mapping to a file page
1177  * @page: the page to add the mapping to
1178  * @compound: charge the page as compound or small page
1179  *
1180  * The caller needs to hold the pte lock.
1181  */
1182 void page_add_file_rmap(struct page *page, bool compound)
1183 {
1184 	int i, nr = 1;
1185 
1186 	VM_BUG_ON_PAGE(compound && !PageTransHuge(page), page);
1187 	lock_page_memcg(page);
1188 	if (compound && PageTransHuge(page)) {
1189 		for (i = 0, nr = 0; i < HPAGE_PMD_NR; i++) {
1190 			if (atomic_inc_and_test(&page[i]._mapcount))
1191 				nr++;
1192 		}
1193 		if (!atomic_inc_and_test(compound_mapcount_ptr(page)))
1194 			goto out;
1195 		VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
1196 		__inc_node_page_state(page, NR_SHMEM_PMDMAPPED);
1197 	} else {
1198 		if (PageTransCompound(page) && page_mapping(page)) {
1199 			VM_WARN_ON_ONCE(!PageLocked(page));
1200 
1201 			SetPageDoubleMap(compound_head(page));
1202 			if (PageMlocked(page))
1203 				clear_page_mlock(compound_head(page));
1204 		}
1205 		if (!atomic_inc_and_test(&page->_mapcount))
1206 			goto out;
1207 	}
1208 	__mod_lruvec_page_state(page, NR_FILE_MAPPED, nr);
1209 out:
1210 	unlock_page_memcg(page);
1211 }
1212 
1213 static void page_remove_file_rmap(struct page *page, bool compound)
1214 {
1215 	int i, nr = 1;
1216 
1217 	VM_BUG_ON_PAGE(compound && !PageHead(page), page);
1218 	lock_page_memcg(page);
1219 
1220 	/* Hugepages are not counted in NR_FILE_MAPPED for now. */
1221 	if (unlikely(PageHuge(page))) {
1222 		/* hugetlb pages are always mapped with pmds */
1223 		atomic_dec(compound_mapcount_ptr(page));
1224 		goto out;
1225 	}
1226 
1227 	/* page still mapped by someone else? */
1228 	if (compound && PageTransHuge(page)) {
1229 		for (i = 0, nr = 0; i < HPAGE_PMD_NR; i++) {
1230 			if (atomic_add_negative(-1, &page[i]._mapcount))
1231 				nr++;
1232 		}
1233 		if (!atomic_add_negative(-1, compound_mapcount_ptr(page)))
1234 			goto out;
1235 		VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
1236 		__dec_node_page_state(page, NR_SHMEM_PMDMAPPED);
1237 	} else {
1238 		if (!atomic_add_negative(-1, &page->_mapcount))
1239 			goto out;
1240 	}
1241 
1242 	/*
1243 	 * We use the irq-unsafe __{inc|mod}_lruvec_page_state because
1244 	 * these counters are not modified in interrupt context, and
1245 	 * pte lock(a spinlock) is held, which implies preemption disabled.
1246 	 */
1247 	__mod_lruvec_page_state(page, NR_FILE_MAPPED, -nr);
1248 
1249 	if (unlikely(PageMlocked(page)))
1250 		clear_page_mlock(page);
1251 out:
1252 	unlock_page_memcg(page);
1253 }
1254 
1255 static void page_remove_anon_compound_rmap(struct page *page)
1256 {
1257 	int i, nr;
1258 
1259 	if (!atomic_add_negative(-1, compound_mapcount_ptr(page)))
1260 		return;
1261 
1262 	/* Hugepages are not counted in NR_ANON_PAGES for now. */
1263 	if (unlikely(PageHuge(page)))
1264 		return;
1265 
1266 	if (!IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE))
1267 		return;
1268 
1269 	__dec_node_page_state(page, NR_ANON_THPS);
1270 
1271 	if (TestClearPageDoubleMap(page)) {
1272 		/*
1273 		 * Subpages can be mapped with PTEs too. Check how many of
1274 		 * themi are still mapped.
1275 		 */
1276 		for (i = 0, nr = 0; i < HPAGE_PMD_NR; i++) {
1277 			if (atomic_add_negative(-1, &page[i]._mapcount))
1278 				nr++;
1279 		}
1280 	} else {
1281 		nr = HPAGE_PMD_NR;
1282 	}
1283 
1284 	if (unlikely(PageMlocked(page)))
1285 		clear_page_mlock(page);
1286 
1287 	if (nr) {
1288 		__mod_node_page_state(page_pgdat(page), NR_ANON_MAPPED, -nr);
1289 		deferred_split_huge_page(page);
1290 	}
1291 }
1292 
1293 /**
1294  * page_remove_rmap - take down pte mapping from a page
1295  * @page:	page to remove mapping from
1296  * @compound:	uncharge the page as compound or small page
1297  *
1298  * The caller needs to hold the pte lock.
1299  */
1300 void page_remove_rmap(struct page *page, bool compound)
1301 {
1302 	if (!PageAnon(page))
1303 		return page_remove_file_rmap(page, compound);
1304 
1305 	if (compound)
1306 		return page_remove_anon_compound_rmap(page);
1307 
1308 	/* page still mapped by someone else? */
1309 	if (!atomic_add_negative(-1, &page->_mapcount))
1310 		return;
1311 
1312 	/*
1313 	 * We use the irq-unsafe __{inc|mod}_zone_page_stat because
1314 	 * these counters are not modified in interrupt context, and
1315 	 * pte lock(a spinlock) is held, which implies preemption disabled.
1316 	 */
1317 	__dec_node_page_state(page, NR_ANON_MAPPED);
1318 
1319 	if (unlikely(PageMlocked(page)))
1320 		clear_page_mlock(page);
1321 
1322 	if (PageTransCompound(page))
1323 		deferred_split_huge_page(compound_head(page));
1324 
1325 	/*
1326 	 * It would be tidy to reset the PageAnon mapping here,
1327 	 * but that might overwrite a racing page_add_anon_rmap
1328 	 * which increments mapcount after us but sets mapping
1329 	 * before us: so leave the reset to free_unref_page,
1330 	 * and remember that it's only reliable while mapped.
1331 	 * Leaving it set also helps swapoff to reinstate ptes
1332 	 * faster for those pages still in swapcache.
1333 	 */
1334 }
1335 
1336 /*
1337  * @arg: enum ttu_flags will be passed to this argument
1338  */
1339 static bool try_to_unmap_one(struct page *page, struct vm_area_struct *vma,
1340 		     unsigned long address, void *arg)
1341 {
1342 	struct mm_struct *mm = vma->vm_mm;
1343 	struct page_vma_mapped_walk pvmw = {
1344 		.page = page,
1345 		.vma = vma,
1346 		.address = address,
1347 	};
1348 	pte_t pteval;
1349 	struct page *subpage;
1350 	bool ret = true;
1351 	struct mmu_notifier_range range;
1352 	enum ttu_flags flags = (enum ttu_flags)arg;
1353 
1354 	/* munlock has nothing to gain from examining un-locked vmas */
1355 	if ((flags & TTU_MUNLOCK) && !(vma->vm_flags & VM_LOCKED))
1356 		return true;
1357 
1358 	if (IS_ENABLED(CONFIG_MIGRATION) && (flags & TTU_MIGRATION) &&
1359 	    is_zone_device_page(page) && !is_device_private_page(page))
1360 		return true;
1361 
1362 	if (flags & TTU_SPLIT_HUGE_PMD) {
1363 		split_huge_pmd_address(vma, address,
1364 				flags & TTU_SPLIT_FREEZE, page);
1365 	}
1366 
1367 	/*
1368 	 * For THP, we have to assume the worse case ie pmd for invalidation.
1369 	 * For hugetlb, it could be much worse if we need to do pud
1370 	 * invalidation in the case of pmd sharing.
1371 	 *
1372 	 * Note that the page can not be free in this function as call of
1373 	 * try_to_unmap() must hold a reference on the page.
1374 	 */
1375 	mmu_notifier_range_init(&range, vma->vm_mm, vma->vm_start,
1376 				min(vma->vm_end, vma->vm_start +
1377 				    (PAGE_SIZE << compound_order(page))));
1378 	if (PageHuge(page)) {
1379 		/*
1380 		 * If sharing is possible, start and end will be adjusted
1381 		 * accordingly.
1382 		 *
1383 		 * If called for a huge page, caller must hold i_mmap_rwsem
1384 		 * in write mode as it is possible to call huge_pmd_unshare.
1385 		 */
1386 		adjust_range_if_pmd_sharing_possible(vma, &range.start,
1387 						     &range.end);
1388 	}
1389 	mmu_notifier_invalidate_range_start(&range);
1390 
1391 	while (page_vma_mapped_walk(&pvmw)) {
1392 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1393 		/* PMD-mapped THP migration entry */
1394 		if (!pvmw.pte && (flags & TTU_MIGRATION)) {
1395 			VM_BUG_ON_PAGE(PageHuge(page) || !PageTransCompound(page), page);
1396 
1397 			set_pmd_migration_entry(&pvmw, page);
1398 			continue;
1399 		}
1400 #endif
1401 
1402 		/*
1403 		 * If the page is mlock()d, we cannot swap it out.
1404 		 * If it's recently referenced (perhaps page_referenced
1405 		 * skipped over this mm) then we should reactivate it.
1406 		 */
1407 		if (!(flags & TTU_IGNORE_MLOCK)) {
1408 			if (vma->vm_flags & VM_LOCKED) {
1409 				/* PTE-mapped THP are never mlocked */
1410 				if (!PageTransCompound(page)) {
1411 					/*
1412 					 * Holding pte lock, we do *not* need
1413 					 * mmap_sem here
1414 					 */
1415 					mlock_vma_page(page);
1416 				}
1417 				ret = false;
1418 				page_vma_mapped_walk_done(&pvmw);
1419 				break;
1420 			}
1421 			if (flags & TTU_MUNLOCK)
1422 				continue;
1423 		}
1424 
1425 		/* Unexpected PMD-mapped THP? */
1426 		VM_BUG_ON_PAGE(!pvmw.pte, page);
1427 
1428 		subpage = page - page_to_pfn(page) + pte_pfn(*pvmw.pte);
1429 		address = pvmw.address;
1430 
1431 		if (PageHuge(page)) {
1432 			if (huge_pmd_unshare(mm, &address, pvmw.pte)) {
1433 				/*
1434 				 * huge_pmd_unshare unmapped an entire PMD
1435 				 * page.  There is no way of knowing exactly
1436 				 * which PMDs may be cached for this mm, so
1437 				 * we must flush them all.  start/end were
1438 				 * already adjusted above to cover this range.
1439 				 */
1440 				flush_cache_range(vma, range.start, range.end);
1441 				flush_tlb_range(vma, range.start, range.end);
1442 				mmu_notifier_invalidate_range(mm, range.start,
1443 							      range.end);
1444 
1445 				/*
1446 				 * The ref count of the PMD page was dropped
1447 				 * which is part of the way map counting
1448 				 * is done for shared PMDs.  Return 'true'
1449 				 * here.  When there is no other sharing,
1450 				 * huge_pmd_unshare returns false and we will
1451 				 * unmap the actual page and drop map count
1452 				 * to zero.
1453 				 */
1454 				page_vma_mapped_walk_done(&pvmw);
1455 				break;
1456 			}
1457 		}
1458 
1459 		if (IS_ENABLED(CONFIG_MIGRATION) &&
1460 		    (flags & TTU_MIGRATION) &&
1461 		    is_zone_device_page(page)) {
1462 			swp_entry_t entry;
1463 			pte_t swp_pte;
1464 
1465 			pteval = ptep_get_and_clear(mm, pvmw.address, pvmw.pte);
1466 
1467 			/*
1468 			 * Store the pfn of the page in a special migration
1469 			 * pte. do_swap_page() will wait until the migration
1470 			 * pte is removed and then restart fault handling.
1471 			 */
1472 			entry = make_migration_entry(page, 0);
1473 			swp_pte = swp_entry_to_pte(entry);
1474 			if (pte_soft_dirty(pteval))
1475 				swp_pte = pte_swp_mksoft_dirty(swp_pte);
1476 			set_pte_at(mm, pvmw.address, pvmw.pte, swp_pte);
1477 			/*
1478 			 * No need to invalidate here it will synchronize on
1479 			 * against the special swap migration pte.
1480 			 */
1481 			goto discard;
1482 		}
1483 
1484 		if (!(flags & TTU_IGNORE_ACCESS)) {
1485 			if (ptep_clear_flush_young_notify(vma, address,
1486 						pvmw.pte)) {
1487 				ret = false;
1488 				page_vma_mapped_walk_done(&pvmw);
1489 				break;
1490 			}
1491 		}
1492 
1493 		/* Nuke the page table entry. */
1494 		flush_cache_page(vma, address, pte_pfn(*pvmw.pte));
1495 		if (should_defer_flush(mm, flags)) {
1496 			/*
1497 			 * We clear the PTE but do not flush so potentially
1498 			 * a remote CPU could still be writing to the page.
1499 			 * If the entry was previously clean then the
1500 			 * architecture must guarantee that a clear->dirty
1501 			 * transition on a cached TLB entry is written through
1502 			 * and traps if the PTE is unmapped.
1503 			 */
1504 			pteval = ptep_get_and_clear(mm, address, pvmw.pte);
1505 
1506 			set_tlb_ubc_flush_pending(mm, pte_dirty(pteval));
1507 		} else {
1508 			pteval = ptep_clear_flush(vma, address, pvmw.pte);
1509 		}
1510 
1511 		/* Move the dirty bit to the page. Now the pte is gone. */
1512 		if (pte_dirty(pteval))
1513 			set_page_dirty(page);
1514 
1515 		/* Update high watermark before we lower rss */
1516 		update_hiwater_rss(mm);
1517 
1518 		if (PageHWPoison(page) && !(flags & TTU_IGNORE_HWPOISON)) {
1519 			pteval = swp_entry_to_pte(make_hwpoison_entry(subpage));
1520 			if (PageHuge(page)) {
1521 				int nr = 1 << compound_order(page);
1522 				hugetlb_count_sub(nr, mm);
1523 				set_huge_swap_pte_at(mm, address,
1524 						     pvmw.pte, pteval,
1525 						     vma_mmu_pagesize(vma));
1526 			} else {
1527 				dec_mm_counter(mm, mm_counter(page));
1528 				set_pte_at(mm, address, pvmw.pte, pteval);
1529 			}
1530 
1531 		} else if (pte_unused(pteval) && !userfaultfd_armed(vma)) {
1532 			/*
1533 			 * The guest indicated that the page content is of no
1534 			 * interest anymore. Simply discard the pte, vmscan
1535 			 * will take care of the rest.
1536 			 * A future reference will then fault in a new zero
1537 			 * page. When userfaultfd is active, we must not drop
1538 			 * this page though, as its main user (postcopy
1539 			 * migration) will not expect userfaults on already
1540 			 * copied pages.
1541 			 */
1542 			dec_mm_counter(mm, mm_counter(page));
1543 			/* We have to invalidate as we cleared the pte */
1544 			mmu_notifier_invalidate_range(mm, address,
1545 						      address + PAGE_SIZE);
1546 		} else if (IS_ENABLED(CONFIG_MIGRATION) &&
1547 				(flags & (TTU_MIGRATION|TTU_SPLIT_FREEZE))) {
1548 			swp_entry_t entry;
1549 			pte_t swp_pte;
1550 
1551 			if (arch_unmap_one(mm, vma, address, pteval) < 0) {
1552 				set_pte_at(mm, address, pvmw.pte, pteval);
1553 				ret = false;
1554 				page_vma_mapped_walk_done(&pvmw);
1555 				break;
1556 			}
1557 
1558 			/*
1559 			 * Store the pfn of the page in a special migration
1560 			 * pte. do_swap_page() will wait until the migration
1561 			 * pte is removed and then restart fault handling.
1562 			 */
1563 			entry = make_migration_entry(subpage,
1564 					pte_write(pteval));
1565 			swp_pte = swp_entry_to_pte(entry);
1566 			if (pte_soft_dirty(pteval))
1567 				swp_pte = pte_swp_mksoft_dirty(swp_pte);
1568 			set_pte_at(mm, address, pvmw.pte, swp_pte);
1569 			/*
1570 			 * No need to invalidate here it will synchronize on
1571 			 * against the special swap migration pte.
1572 			 */
1573 		} else if (PageAnon(page)) {
1574 			swp_entry_t entry = { .val = page_private(subpage) };
1575 			pte_t swp_pte;
1576 			/*
1577 			 * Store the swap location in the pte.
1578 			 * See handle_pte_fault() ...
1579 			 */
1580 			if (unlikely(PageSwapBacked(page) != PageSwapCache(page))) {
1581 				WARN_ON_ONCE(1);
1582 				ret = false;
1583 				/* We have to invalidate as we cleared the pte */
1584 				mmu_notifier_invalidate_range(mm, address,
1585 							address + PAGE_SIZE);
1586 				page_vma_mapped_walk_done(&pvmw);
1587 				break;
1588 			}
1589 
1590 			/* MADV_FREE page check */
1591 			if (!PageSwapBacked(page)) {
1592 				if (!PageDirty(page)) {
1593 					/* Invalidate as we cleared the pte */
1594 					mmu_notifier_invalidate_range(mm,
1595 						address, address + PAGE_SIZE);
1596 					dec_mm_counter(mm, MM_ANONPAGES);
1597 					goto discard;
1598 				}
1599 
1600 				/*
1601 				 * If the page was redirtied, it cannot be
1602 				 * discarded. Remap the page to page table.
1603 				 */
1604 				set_pte_at(mm, address, pvmw.pte, pteval);
1605 				SetPageSwapBacked(page);
1606 				ret = false;
1607 				page_vma_mapped_walk_done(&pvmw);
1608 				break;
1609 			}
1610 
1611 			if (swap_duplicate(entry) < 0) {
1612 				set_pte_at(mm, address, pvmw.pte, pteval);
1613 				ret = false;
1614 				page_vma_mapped_walk_done(&pvmw);
1615 				break;
1616 			}
1617 			if (arch_unmap_one(mm, vma, address, pteval) < 0) {
1618 				set_pte_at(mm, address, pvmw.pte, pteval);
1619 				ret = false;
1620 				page_vma_mapped_walk_done(&pvmw);
1621 				break;
1622 			}
1623 			if (list_empty(&mm->mmlist)) {
1624 				spin_lock(&mmlist_lock);
1625 				if (list_empty(&mm->mmlist))
1626 					list_add(&mm->mmlist, &init_mm.mmlist);
1627 				spin_unlock(&mmlist_lock);
1628 			}
1629 			dec_mm_counter(mm, MM_ANONPAGES);
1630 			inc_mm_counter(mm, MM_SWAPENTS);
1631 			swp_pte = swp_entry_to_pte(entry);
1632 			if (pte_soft_dirty(pteval))
1633 				swp_pte = pte_swp_mksoft_dirty(swp_pte);
1634 			set_pte_at(mm, address, pvmw.pte, swp_pte);
1635 			/* Invalidate as we cleared the pte */
1636 			mmu_notifier_invalidate_range(mm, address,
1637 						      address + PAGE_SIZE);
1638 		} else {
1639 			/*
1640 			 * This is a locked file-backed page, thus it cannot
1641 			 * be removed from the page cache and replaced by a new
1642 			 * page before mmu_notifier_invalidate_range_end, so no
1643 			 * concurrent thread might update its page table to
1644 			 * point at new page while a device still is using this
1645 			 * page.
1646 			 *
1647 			 * See Documentation/vm/mmu_notifier.rst
1648 			 */
1649 			dec_mm_counter(mm, mm_counter_file(page));
1650 		}
1651 discard:
1652 		/*
1653 		 * No need to call mmu_notifier_invalidate_range() it has be
1654 		 * done above for all cases requiring it to happen under page
1655 		 * table lock before mmu_notifier_invalidate_range_end()
1656 		 *
1657 		 * See Documentation/vm/mmu_notifier.rst
1658 		 */
1659 		page_remove_rmap(subpage, PageHuge(page));
1660 		put_page(page);
1661 	}
1662 
1663 	mmu_notifier_invalidate_range_end(&range);
1664 
1665 	return ret;
1666 }
1667 
1668 bool is_vma_temporary_stack(struct vm_area_struct *vma)
1669 {
1670 	int maybe_stack = vma->vm_flags & (VM_GROWSDOWN | VM_GROWSUP);
1671 
1672 	if (!maybe_stack)
1673 		return false;
1674 
1675 	if ((vma->vm_flags & VM_STACK_INCOMPLETE_SETUP) ==
1676 						VM_STACK_INCOMPLETE_SETUP)
1677 		return true;
1678 
1679 	return false;
1680 }
1681 
1682 static bool invalid_migration_vma(struct vm_area_struct *vma, void *arg)
1683 {
1684 	return is_vma_temporary_stack(vma);
1685 }
1686 
1687 static int page_mapcount_is_zero(struct page *page)
1688 {
1689 	return !total_mapcount(page);
1690 }
1691 
1692 /**
1693  * try_to_unmap - try to remove all page table mappings to a page
1694  * @page: the page to get unmapped
1695  * @flags: action and flags
1696  *
1697  * Tries to remove all the page table entries which are mapping this
1698  * page, used in the pageout path.  Caller must hold the page lock.
1699  *
1700  * If unmap is successful, return true. Otherwise, false.
1701  */
1702 bool try_to_unmap(struct page *page, enum ttu_flags flags)
1703 {
1704 	struct rmap_walk_control rwc = {
1705 		.rmap_one = try_to_unmap_one,
1706 		.arg = (void *)flags,
1707 		.done = page_mapcount_is_zero,
1708 		.anon_lock = page_lock_anon_vma_read,
1709 	};
1710 
1711 	/*
1712 	 * During exec, a temporary VMA is setup and later moved.
1713 	 * The VMA is moved under the anon_vma lock but not the
1714 	 * page tables leading to a race where migration cannot
1715 	 * find the migration ptes. Rather than increasing the
1716 	 * locking requirements of exec(), migration skips
1717 	 * temporary VMAs until after exec() completes.
1718 	 */
1719 	if ((flags & (TTU_MIGRATION|TTU_SPLIT_FREEZE))
1720 	    && !PageKsm(page) && PageAnon(page))
1721 		rwc.invalid_vma = invalid_migration_vma;
1722 
1723 	if (flags & TTU_RMAP_LOCKED)
1724 		rmap_walk_locked(page, &rwc);
1725 	else
1726 		rmap_walk(page, &rwc);
1727 
1728 	return !page_mapcount(page) ? true : false;
1729 }
1730 
1731 static int page_not_mapped(struct page *page)
1732 {
1733 	return !page_mapped(page);
1734 };
1735 
1736 /**
1737  * try_to_munlock - try to munlock a page
1738  * @page: the page to be munlocked
1739  *
1740  * Called from munlock code.  Checks all of the VMAs mapping the page
1741  * to make sure nobody else has this page mlocked. The page will be
1742  * returned with PG_mlocked cleared if no other vmas have it mlocked.
1743  */
1744 
1745 void try_to_munlock(struct page *page)
1746 {
1747 	struct rmap_walk_control rwc = {
1748 		.rmap_one = try_to_unmap_one,
1749 		.arg = (void *)TTU_MUNLOCK,
1750 		.done = page_not_mapped,
1751 		.anon_lock = page_lock_anon_vma_read,
1752 
1753 	};
1754 
1755 	VM_BUG_ON_PAGE(!PageLocked(page) || PageLRU(page), page);
1756 	VM_BUG_ON_PAGE(PageCompound(page) && PageDoubleMap(page), page);
1757 
1758 	rmap_walk(page, &rwc);
1759 }
1760 
1761 void __put_anon_vma(struct anon_vma *anon_vma)
1762 {
1763 	struct anon_vma *root = anon_vma->root;
1764 
1765 	anon_vma_free(anon_vma);
1766 	if (root != anon_vma && atomic_dec_and_test(&root->refcount))
1767 		anon_vma_free(root);
1768 }
1769 
1770 static struct anon_vma *rmap_walk_anon_lock(struct page *page,
1771 					struct rmap_walk_control *rwc)
1772 {
1773 	struct anon_vma *anon_vma;
1774 
1775 	if (rwc->anon_lock)
1776 		return rwc->anon_lock(page);
1777 
1778 	/*
1779 	 * Note: remove_migration_ptes() cannot use page_lock_anon_vma_read()
1780 	 * because that depends on page_mapped(); but not all its usages
1781 	 * are holding mmap_sem. Users without mmap_sem are required to
1782 	 * take a reference count to prevent the anon_vma disappearing
1783 	 */
1784 	anon_vma = page_anon_vma(page);
1785 	if (!anon_vma)
1786 		return NULL;
1787 
1788 	anon_vma_lock_read(anon_vma);
1789 	return anon_vma;
1790 }
1791 
1792 /*
1793  * rmap_walk_anon - do something to anonymous page using the object-based
1794  * rmap method
1795  * @page: the page to be handled
1796  * @rwc: control variable according to each walk type
1797  *
1798  * Find all the mappings of a page using the mapping pointer and the vma chains
1799  * contained in the anon_vma struct it points to.
1800  *
1801  * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1802  * where the page was found will be held for write.  So, we won't recheck
1803  * vm_flags for that VMA.  That should be OK, because that vma shouldn't be
1804  * LOCKED.
1805  */
1806 static void rmap_walk_anon(struct page *page, struct rmap_walk_control *rwc,
1807 		bool locked)
1808 {
1809 	struct anon_vma *anon_vma;
1810 	pgoff_t pgoff_start, pgoff_end;
1811 	struct anon_vma_chain *avc;
1812 
1813 	if (locked) {
1814 		anon_vma = page_anon_vma(page);
1815 		/* anon_vma disappear under us? */
1816 		VM_BUG_ON_PAGE(!anon_vma, page);
1817 	} else {
1818 		anon_vma = rmap_walk_anon_lock(page, rwc);
1819 	}
1820 	if (!anon_vma)
1821 		return;
1822 
1823 	pgoff_start = page_to_pgoff(page);
1824 	pgoff_end = pgoff_start + hpage_nr_pages(page) - 1;
1825 	anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root,
1826 			pgoff_start, pgoff_end) {
1827 		struct vm_area_struct *vma = avc->vma;
1828 		unsigned long address = vma_address(page, vma);
1829 
1830 		cond_resched();
1831 
1832 		if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
1833 			continue;
1834 
1835 		if (!rwc->rmap_one(page, vma, address, rwc->arg))
1836 			break;
1837 		if (rwc->done && rwc->done(page))
1838 			break;
1839 	}
1840 
1841 	if (!locked)
1842 		anon_vma_unlock_read(anon_vma);
1843 }
1844 
1845 /*
1846  * rmap_walk_file - do something to file page using the object-based rmap method
1847  * @page: the page to be handled
1848  * @rwc: control variable according to each walk type
1849  *
1850  * Find all the mappings of a page using the mapping pointer and the vma chains
1851  * contained in the address_space struct it points to.
1852  *
1853  * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1854  * where the page was found will be held for write.  So, we won't recheck
1855  * vm_flags for that VMA.  That should be OK, because that vma shouldn't be
1856  * LOCKED.
1857  */
1858 static void rmap_walk_file(struct page *page, struct rmap_walk_control *rwc,
1859 		bool locked)
1860 {
1861 	struct address_space *mapping = page_mapping(page);
1862 	pgoff_t pgoff_start, pgoff_end;
1863 	struct vm_area_struct *vma;
1864 
1865 	/*
1866 	 * The page lock not only makes sure that page->mapping cannot
1867 	 * suddenly be NULLified by truncation, it makes sure that the
1868 	 * structure at mapping cannot be freed and reused yet,
1869 	 * so we can safely take mapping->i_mmap_rwsem.
1870 	 */
1871 	VM_BUG_ON_PAGE(!PageLocked(page), page);
1872 
1873 	if (!mapping)
1874 		return;
1875 
1876 	pgoff_start = page_to_pgoff(page);
1877 	pgoff_end = pgoff_start + hpage_nr_pages(page) - 1;
1878 	if (!locked)
1879 		i_mmap_lock_read(mapping);
1880 	vma_interval_tree_foreach(vma, &mapping->i_mmap,
1881 			pgoff_start, pgoff_end) {
1882 		unsigned long address = vma_address(page, vma);
1883 
1884 		cond_resched();
1885 
1886 		if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
1887 			continue;
1888 
1889 		if (!rwc->rmap_one(page, vma, address, rwc->arg))
1890 			goto done;
1891 		if (rwc->done && rwc->done(page))
1892 			goto done;
1893 	}
1894 
1895 done:
1896 	if (!locked)
1897 		i_mmap_unlock_read(mapping);
1898 }
1899 
1900 void rmap_walk(struct page *page, struct rmap_walk_control *rwc)
1901 {
1902 	if (unlikely(PageKsm(page)))
1903 		rmap_walk_ksm(page, rwc);
1904 	else if (PageAnon(page))
1905 		rmap_walk_anon(page, rwc, false);
1906 	else
1907 		rmap_walk_file(page, rwc, false);
1908 }
1909 
1910 /* Like rmap_walk, but caller holds relevant rmap lock */
1911 void rmap_walk_locked(struct page *page, struct rmap_walk_control *rwc)
1912 {
1913 	/* no ksm support for now */
1914 	VM_BUG_ON_PAGE(PageKsm(page), page);
1915 	if (PageAnon(page))
1916 		rmap_walk_anon(page, rwc, true);
1917 	else
1918 		rmap_walk_file(page, rwc, true);
1919 }
1920 
1921 #ifdef CONFIG_HUGETLB_PAGE
1922 /*
1923  * The following two functions are for anonymous (private mapped) hugepages.
1924  * Unlike common anonymous pages, anonymous hugepages have no accounting code
1925  * and no lru code, because we handle hugepages differently from common pages.
1926  */
1927 void hugepage_add_anon_rmap(struct page *page,
1928 			    struct vm_area_struct *vma, unsigned long address)
1929 {
1930 	struct anon_vma *anon_vma = vma->anon_vma;
1931 	int first;
1932 
1933 	BUG_ON(!PageLocked(page));
1934 	BUG_ON(!anon_vma);
1935 	/* address might be in next vma when migration races vma_adjust */
1936 	first = atomic_inc_and_test(compound_mapcount_ptr(page));
1937 	if (first)
1938 		__page_set_anon_rmap(page, vma, address, 0);
1939 }
1940 
1941 void hugepage_add_new_anon_rmap(struct page *page,
1942 			struct vm_area_struct *vma, unsigned long address)
1943 {
1944 	BUG_ON(address < vma->vm_start || address >= vma->vm_end);
1945 	atomic_set(compound_mapcount_ptr(page), 0);
1946 	__page_set_anon_rmap(page, vma, address, 1);
1947 }
1948 #endif /* CONFIG_HUGETLB_PAGE */
1949