xref: /openbmc/linux/mm/rmap.c (revision 5d0e4d78)
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  *           anon_vma->rwsem
29  *             mm->page_table_lock or pte_lock
30  *               zone_lru_lock (in mark_page_accessed, isolate_lru_page)
31  *               swap_lock (in swap_duplicate, swap_info_get)
32  *                 mmlist_lock (in mmput, drain_mmlist and others)
33  *                 mapping->private_lock (in __set_page_dirty_buffers)
34  *                   mem_cgroup_{begin,end}_page_stat (memcg->move_lock)
35  *                     mapping->tree_lock (widely used)
36  *                 inode->i_lock (in set_page_dirty's __mark_inode_dirty)
37  *                 bdi.wb->list_lock (in set_page_dirty's __mark_inode_dirty)
38  *                   sb_lock (within inode_lock in fs/fs-writeback.c)
39  *                   mapping->tree_lock (widely used, in set_page_dirty,
40  *                             in arch-dependent flush_dcache_mmap_lock,
41  *                             within bdi.wb->list_lock in __sync_single_inode)
42  *
43  * anon_vma->rwsem,mapping->i_mutex      (memory_failure, collect_procs_anon)
44  *   ->tasklist_lock
45  *     pte map lock
46  */
47 
48 #include <linux/mm.h>
49 #include <linux/sched/mm.h>
50 #include <linux/sched/task.h>
51 #include <linux/pagemap.h>
52 #include <linux/swap.h>
53 #include <linux/swapops.h>
54 #include <linux/slab.h>
55 #include <linux/init.h>
56 #include <linux/ksm.h>
57 #include <linux/rmap.h>
58 #include <linux/rcupdate.h>
59 #include <linux/export.h>
60 #include <linux/memcontrol.h>
61 #include <linux/mmu_notifier.h>
62 #include <linux/migrate.h>
63 #include <linux/hugetlb.h>
64 #include <linux/backing-dev.h>
65 #include <linux/page_idle.h>
66 
67 #include <asm/tlbflush.h>
68 
69 #include <trace/events/tlb.h>
70 
71 #include "internal.h"
72 
73 static struct kmem_cache *anon_vma_cachep;
74 static struct kmem_cache *anon_vma_chain_cachep;
75 
76 static inline struct anon_vma *anon_vma_alloc(void)
77 {
78 	struct anon_vma *anon_vma;
79 
80 	anon_vma = kmem_cache_alloc(anon_vma_cachep, GFP_KERNEL);
81 	if (anon_vma) {
82 		atomic_set(&anon_vma->refcount, 1);
83 		anon_vma->degree = 1;	/* Reference for first vma */
84 		anon_vma->parent = anon_vma;
85 		/*
86 		 * Initialise the anon_vma root to point to itself. If called
87 		 * from fork, the root will be reset to the parents anon_vma.
88 		 */
89 		anon_vma->root = anon_vma;
90 	}
91 
92 	return anon_vma;
93 }
94 
95 static inline void anon_vma_free(struct anon_vma *anon_vma)
96 {
97 	VM_BUG_ON(atomic_read(&anon_vma->refcount));
98 
99 	/*
100 	 * Synchronize against page_lock_anon_vma_read() such that
101 	 * we can safely hold the lock without the anon_vma getting
102 	 * freed.
103 	 *
104 	 * Relies on the full mb implied by the atomic_dec_and_test() from
105 	 * put_anon_vma() against the acquire barrier implied by
106 	 * down_read_trylock() from page_lock_anon_vma_read(). This orders:
107 	 *
108 	 * page_lock_anon_vma_read()	VS	put_anon_vma()
109 	 *   down_read_trylock()		  atomic_dec_and_test()
110 	 *   LOCK				  MB
111 	 *   atomic_read()			  rwsem_is_locked()
112 	 *
113 	 * LOCK should suffice since the actual taking of the lock must
114 	 * happen _before_ what follows.
115 	 */
116 	might_sleep();
117 	if (rwsem_is_locked(&anon_vma->root->rwsem)) {
118 		anon_vma_lock_write(anon_vma);
119 		anon_vma_unlock_write(anon_vma);
120 	}
121 
122 	kmem_cache_free(anon_vma_cachep, anon_vma);
123 }
124 
125 static inline struct anon_vma_chain *anon_vma_chain_alloc(gfp_t gfp)
126 {
127 	return kmem_cache_alloc(anon_vma_chain_cachep, gfp);
128 }
129 
130 static void anon_vma_chain_free(struct anon_vma_chain *anon_vma_chain)
131 {
132 	kmem_cache_free(anon_vma_chain_cachep, anon_vma_chain);
133 }
134 
135 static void anon_vma_chain_link(struct vm_area_struct *vma,
136 				struct anon_vma_chain *avc,
137 				struct anon_vma *anon_vma)
138 {
139 	avc->vma = vma;
140 	avc->anon_vma = anon_vma;
141 	list_add(&avc->same_vma, &vma->anon_vma_chain);
142 	anon_vma_interval_tree_insert(avc, &anon_vma->rb_root);
143 }
144 
145 /**
146  * __anon_vma_prepare - attach an anon_vma to a memory region
147  * @vma: the memory region in question
148  *
149  * This makes sure the memory mapping described by 'vma' has
150  * an 'anon_vma' attached to it, so that we can associate the
151  * anonymous pages mapped into it with that anon_vma.
152  *
153  * The common case will be that we already have one, which
154  * is handled inline by anon_vma_prepare(). But if
155  * not we either need to find an adjacent mapping that we
156  * can re-use the anon_vma from (very common when the only
157  * reason for splitting a vma has been mprotect()), or we
158  * allocate a new one.
159  *
160  * Anon-vma allocations are very subtle, because we may have
161  * optimistically looked up an anon_vma in page_lock_anon_vma_read()
162  * and that may actually touch the spinlock even in the newly
163  * allocated vma (it depends on RCU to make sure that the
164  * anon_vma isn't actually destroyed).
165  *
166  * As a result, we need to do proper anon_vma locking even
167  * for the new allocation. At the same time, we do not want
168  * to do any locking for the common case of already having
169  * an anon_vma.
170  *
171  * This must be called with the mmap_sem held for reading.
172  */
173 int __anon_vma_prepare(struct vm_area_struct *vma)
174 {
175 	struct mm_struct *mm = vma->vm_mm;
176 	struct anon_vma *anon_vma, *allocated;
177 	struct anon_vma_chain *avc;
178 
179 	might_sleep();
180 
181 	avc = anon_vma_chain_alloc(GFP_KERNEL);
182 	if (!avc)
183 		goto out_enomem;
184 
185 	anon_vma = find_mergeable_anon_vma(vma);
186 	allocated = NULL;
187 	if (!anon_vma) {
188 		anon_vma = anon_vma_alloc();
189 		if (unlikely(!anon_vma))
190 			goto out_enomem_free_avc;
191 		allocated = anon_vma;
192 	}
193 
194 	anon_vma_lock_write(anon_vma);
195 	/* page_table_lock to protect against threads */
196 	spin_lock(&mm->page_table_lock);
197 	if (likely(!vma->anon_vma)) {
198 		vma->anon_vma = anon_vma;
199 		anon_vma_chain_link(vma, avc, anon_vma);
200 		/* vma reference or self-parent link for new root */
201 		anon_vma->degree++;
202 		allocated = NULL;
203 		avc = NULL;
204 	}
205 	spin_unlock(&mm->page_table_lock);
206 	anon_vma_unlock_write(anon_vma);
207 
208 	if (unlikely(allocated))
209 		put_anon_vma(allocated);
210 	if (unlikely(avc))
211 		anon_vma_chain_free(avc);
212 
213 	return 0;
214 
215  out_enomem_free_avc:
216 	anon_vma_chain_free(avc);
217  out_enomem:
218 	return -ENOMEM;
219 }
220 
221 /*
222  * This is a useful helper function for locking the anon_vma root as
223  * we traverse the vma->anon_vma_chain, looping over anon_vma's that
224  * have the same vma.
225  *
226  * Such anon_vma's should have the same root, so you'd expect to see
227  * just a single mutex_lock for the whole traversal.
228  */
229 static inline struct anon_vma *lock_anon_vma_root(struct anon_vma *root, struct anon_vma *anon_vma)
230 {
231 	struct anon_vma *new_root = anon_vma->root;
232 	if (new_root != root) {
233 		if (WARN_ON_ONCE(root))
234 			up_write(&root->rwsem);
235 		root = new_root;
236 		down_write(&root->rwsem);
237 	}
238 	return root;
239 }
240 
241 static inline void unlock_anon_vma_root(struct anon_vma *root)
242 {
243 	if (root)
244 		up_write(&root->rwsem);
245 }
246 
247 /*
248  * Attach the anon_vmas from src to dst.
249  * Returns 0 on success, -ENOMEM on failure.
250  *
251  * If dst->anon_vma is NULL this function tries to find and reuse existing
252  * anon_vma which has no vmas and only one child anon_vma. This prevents
253  * degradation of anon_vma hierarchy to endless linear chain in case of
254  * constantly forking task. On the other hand, an anon_vma with more than one
255  * child isn't reused even if there was no alive vma, thus rmap walker has a
256  * good chance of avoiding scanning the whole hierarchy when it searches where
257  * page is mapped.
258  */
259 int anon_vma_clone(struct vm_area_struct *dst, struct vm_area_struct *src)
260 {
261 	struct anon_vma_chain *avc, *pavc;
262 	struct anon_vma *root = NULL;
263 
264 	list_for_each_entry_reverse(pavc, &src->anon_vma_chain, same_vma) {
265 		struct anon_vma *anon_vma;
266 
267 		avc = anon_vma_chain_alloc(GFP_NOWAIT | __GFP_NOWARN);
268 		if (unlikely(!avc)) {
269 			unlock_anon_vma_root(root);
270 			root = NULL;
271 			avc = anon_vma_chain_alloc(GFP_KERNEL);
272 			if (!avc)
273 				goto enomem_failure;
274 		}
275 		anon_vma = pavc->anon_vma;
276 		root = lock_anon_vma_root(root, anon_vma);
277 		anon_vma_chain_link(dst, avc, anon_vma);
278 
279 		/*
280 		 * Reuse existing anon_vma if its degree lower than two,
281 		 * that means it has no vma and only one anon_vma child.
282 		 *
283 		 * Do not chose parent anon_vma, otherwise first child
284 		 * will always reuse it. Root anon_vma is never reused:
285 		 * it has self-parent reference and at least one child.
286 		 */
287 		if (!dst->anon_vma && anon_vma != src->anon_vma &&
288 				anon_vma->degree < 2)
289 			dst->anon_vma = anon_vma;
290 	}
291 	if (dst->anon_vma)
292 		dst->anon_vma->degree++;
293 	unlock_anon_vma_root(root);
294 	return 0;
295 
296  enomem_failure:
297 	/*
298 	 * dst->anon_vma is dropped here otherwise its degree can be incorrectly
299 	 * decremented in unlink_anon_vmas().
300 	 * We can safely do this because callers of anon_vma_clone() don't care
301 	 * about dst->anon_vma if anon_vma_clone() failed.
302 	 */
303 	dst->anon_vma = NULL;
304 	unlink_anon_vmas(dst);
305 	return -ENOMEM;
306 }
307 
308 /*
309  * Attach vma to its own anon_vma, as well as to the anon_vmas that
310  * the corresponding VMA in the parent process is attached to.
311  * Returns 0 on success, non-zero on failure.
312  */
313 int anon_vma_fork(struct vm_area_struct *vma, struct vm_area_struct *pvma)
314 {
315 	struct anon_vma_chain *avc;
316 	struct anon_vma *anon_vma;
317 	int error;
318 
319 	/* Don't bother if the parent process has no anon_vma here. */
320 	if (!pvma->anon_vma)
321 		return 0;
322 
323 	/* Drop inherited anon_vma, we'll reuse existing or allocate new. */
324 	vma->anon_vma = NULL;
325 
326 	/*
327 	 * First, attach the new VMA to the parent VMA's anon_vmas,
328 	 * so rmap can find non-COWed pages in child processes.
329 	 */
330 	error = anon_vma_clone(vma, pvma);
331 	if (error)
332 		return error;
333 
334 	/* An existing anon_vma has been reused, all done then. */
335 	if (vma->anon_vma)
336 		return 0;
337 
338 	/* Then add our own anon_vma. */
339 	anon_vma = anon_vma_alloc();
340 	if (!anon_vma)
341 		goto out_error;
342 	avc = anon_vma_chain_alloc(GFP_KERNEL);
343 	if (!avc)
344 		goto out_error_free_anon_vma;
345 
346 	/*
347 	 * The root anon_vma's spinlock is the lock actually used when we
348 	 * lock any of the anon_vmas in this anon_vma tree.
349 	 */
350 	anon_vma->root = pvma->anon_vma->root;
351 	anon_vma->parent = pvma->anon_vma;
352 	/*
353 	 * With refcounts, an anon_vma can stay around longer than the
354 	 * process it belongs to. The root anon_vma needs to be pinned until
355 	 * this anon_vma is freed, because the lock lives in the root.
356 	 */
357 	get_anon_vma(anon_vma->root);
358 	/* Mark this anon_vma as the one where our new (COWed) pages go. */
359 	vma->anon_vma = anon_vma;
360 	anon_vma_lock_write(anon_vma);
361 	anon_vma_chain_link(vma, avc, anon_vma);
362 	anon_vma->parent->degree++;
363 	anon_vma_unlock_write(anon_vma);
364 
365 	return 0;
366 
367  out_error_free_anon_vma:
368 	put_anon_vma(anon_vma);
369  out_error:
370 	unlink_anon_vmas(vma);
371 	return -ENOMEM;
372 }
373 
374 void unlink_anon_vmas(struct vm_area_struct *vma)
375 {
376 	struct anon_vma_chain *avc, *next;
377 	struct anon_vma *root = NULL;
378 
379 	/*
380 	 * Unlink each anon_vma chained to the VMA.  This list is ordered
381 	 * from newest to oldest, ensuring the root anon_vma gets freed last.
382 	 */
383 	list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
384 		struct anon_vma *anon_vma = avc->anon_vma;
385 
386 		root = lock_anon_vma_root(root, anon_vma);
387 		anon_vma_interval_tree_remove(avc, &anon_vma->rb_root);
388 
389 		/*
390 		 * Leave empty anon_vmas on the list - we'll need
391 		 * to free them outside the lock.
392 		 */
393 		if (RB_EMPTY_ROOT(&anon_vma->rb_root)) {
394 			anon_vma->parent->degree--;
395 			continue;
396 		}
397 
398 		list_del(&avc->same_vma);
399 		anon_vma_chain_free(avc);
400 	}
401 	if (vma->anon_vma)
402 		vma->anon_vma->degree--;
403 	unlock_anon_vma_root(root);
404 
405 	/*
406 	 * Iterate the list once more, it now only contains empty and unlinked
407 	 * anon_vmas, destroy them. Could not do before due to __put_anon_vma()
408 	 * needing to write-acquire the anon_vma->root->rwsem.
409 	 */
410 	list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
411 		struct anon_vma *anon_vma = avc->anon_vma;
412 
413 		VM_WARN_ON(anon_vma->degree);
414 		put_anon_vma(anon_vma);
415 
416 		list_del(&avc->same_vma);
417 		anon_vma_chain_free(avc);
418 	}
419 }
420 
421 static void anon_vma_ctor(void *data)
422 {
423 	struct anon_vma *anon_vma = data;
424 
425 	init_rwsem(&anon_vma->rwsem);
426 	atomic_set(&anon_vma->refcount, 0);
427 	anon_vma->rb_root = RB_ROOT;
428 }
429 
430 void __init anon_vma_init(void)
431 {
432 	anon_vma_cachep = kmem_cache_create("anon_vma", sizeof(struct anon_vma),
433 			0, SLAB_TYPESAFE_BY_RCU|SLAB_PANIC|SLAB_ACCOUNT,
434 			anon_vma_ctor);
435 	anon_vma_chain_cachep = KMEM_CACHE(anon_vma_chain,
436 			SLAB_PANIC|SLAB_ACCOUNT);
437 }
438 
439 /*
440  * Getting a lock on a stable anon_vma from a page off the LRU is tricky!
441  *
442  * Since there is no serialization what so ever against page_remove_rmap()
443  * the best this function can do is return a locked anon_vma that might
444  * have been relevant to this page.
445  *
446  * The page might have been remapped to a different anon_vma or the anon_vma
447  * returned may already be freed (and even reused).
448  *
449  * In case it was remapped to a different anon_vma, the new anon_vma will be a
450  * child of the old anon_vma, and the anon_vma lifetime rules will therefore
451  * ensure that any anon_vma obtained from the page will still be valid for as
452  * long as we observe page_mapped() [ hence all those page_mapped() tests ].
453  *
454  * All users of this function must be very careful when walking the anon_vma
455  * chain and verify that the page in question is indeed mapped in it
456  * [ something equivalent to page_mapped_in_vma() ].
457  *
458  * Since anon_vma's slab is DESTROY_BY_RCU and we know from page_remove_rmap()
459  * that the anon_vma pointer from page->mapping is valid if there is a
460  * mapcount, we can dereference the anon_vma after observing those.
461  */
462 struct anon_vma *page_get_anon_vma(struct page *page)
463 {
464 	struct anon_vma *anon_vma = NULL;
465 	unsigned long anon_mapping;
466 
467 	rcu_read_lock();
468 	anon_mapping = (unsigned long)READ_ONCE(page->mapping);
469 	if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
470 		goto out;
471 	if (!page_mapped(page))
472 		goto out;
473 
474 	anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
475 	if (!atomic_inc_not_zero(&anon_vma->refcount)) {
476 		anon_vma = NULL;
477 		goto out;
478 	}
479 
480 	/*
481 	 * If this page is still mapped, then its anon_vma cannot have been
482 	 * freed.  But if it has been unmapped, we have no security against the
483 	 * anon_vma structure being freed and reused (for another anon_vma:
484 	 * SLAB_TYPESAFE_BY_RCU guarantees that - so the atomic_inc_not_zero()
485 	 * above cannot corrupt).
486 	 */
487 	if (!page_mapped(page)) {
488 		rcu_read_unlock();
489 		put_anon_vma(anon_vma);
490 		return NULL;
491 	}
492 out:
493 	rcu_read_unlock();
494 
495 	return anon_vma;
496 }
497 
498 /*
499  * Similar to page_get_anon_vma() except it locks the anon_vma.
500  *
501  * Its a little more complex as it tries to keep the fast path to a single
502  * atomic op -- the trylock. If we fail the trylock, we fall back to getting a
503  * reference like with page_get_anon_vma() and then block on the mutex.
504  */
505 struct anon_vma *page_lock_anon_vma_read(struct page *page)
506 {
507 	struct anon_vma *anon_vma = NULL;
508 	struct anon_vma *root_anon_vma;
509 	unsigned long anon_mapping;
510 
511 	rcu_read_lock();
512 	anon_mapping = (unsigned long)READ_ONCE(page->mapping);
513 	if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
514 		goto out;
515 	if (!page_mapped(page))
516 		goto out;
517 
518 	anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
519 	root_anon_vma = READ_ONCE(anon_vma->root);
520 	if (down_read_trylock(&root_anon_vma->rwsem)) {
521 		/*
522 		 * If the page is still mapped, then this anon_vma is still
523 		 * its anon_vma, and holding the mutex ensures that it will
524 		 * not go away, see anon_vma_free().
525 		 */
526 		if (!page_mapped(page)) {
527 			up_read(&root_anon_vma->rwsem);
528 			anon_vma = NULL;
529 		}
530 		goto out;
531 	}
532 
533 	/* trylock failed, we got to sleep */
534 	if (!atomic_inc_not_zero(&anon_vma->refcount)) {
535 		anon_vma = NULL;
536 		goto out;
537 	}
538 
539 	if (!page_mapped(page)) {
540 		rcu_read_unlock();
541 		put_anon_vma(anon_vma);
542 		return NULL;
543 	}
544 
545 	/* we pinned the anon_vma, its safe to sleep */
546 	rcu_read_unlock();
547 	anon_vma_lock_read(anon_vma);
548 
549 	if (atomic_dec_and_test(&anon_vma->refcount)) {
550 		/*
551 		 * Oops, we held the last refcount, release the lock
552 		 * and bail -- can't simply use put_anon_vma() because
553 		 * we'll deadlock on the anon_vma_lock_write() recursion.
554 		 */
555 		anon_vma_unlock_read(anon_vma);
556 		__put_anon_vma(anon_vma);
557 		anon_vma = NULL;
558 	}
559 
560 	return anon_vma;
561 
562 out:
563 	rcu_read_unlock();
564 	return anon_vma;
565 }
566 
567 void page_unlock_anon_vma_read(struct anon_vma *anon_vma)
568 {
569 	anon_vma_unlock_read(anon_vma);
570 }
571 
572 #ifdef CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH
573 /*
574  * Flush TLB entries for recently unmapped pages from remote CPUs. It is
575  * important if a PTE was dirty when it was unmapped that it's flushed
576  * before any IO is initiated on the page to prevent lost writes. Similarly,
577  * it must be flushed before freeing to prevent data leakage.
578  */
579 void try_to_unmap_flush(void)
580 {
581 	struct tlbflush_unmap_batch *tlb_ubc = &current->tlb_ubc;
582 
583 	if (!tlb_ubc->flush_required)
584 		return;
585 
586 	arch_tlbbatch_flush(&tlb_ubc->arch);
587 	tlb_ubc->flush_required = false;
588 	tlb_ubc->writable = false;
589 }
590 
591 /* Flush iff there are potentially writable TLB entries that can race with IO */
592 void try_to_unmap_flush_dirty(void)
593 {
594 	struct tlbflush_unmap_batch *tlb_ubc = &current->tlb_ubc;
595 
596 	if (tlb_ubc->writable)
597 		try_to_unmap_flush();
598 }
599 
600 static void set_tlb_ubc_flush_pending(struct mm_struct *mm, bool writable)
601 {
602 	struct tlbflush_unmap_batch *tlb_ubc = &current->tlb_ubc;
603 
604 	arch_tlbbatch_add_mm(&tlb_ubc->arch, mm);
605 	tlb_ubc->flush_required = true;
606 
607 	/*
608 	 * If the PTE was dirty then it's best to assume it's writable. The
609 	 * caller must use try_to_unmap_flush_dirty() or try_to_unmap_flush()
610 	 * before the page is queued for IO.
611 	 */
612 	if (writable)
613 		tlb_ubc->writable = true;
614 }
615 
616 /*
617  * Returns true if the TLB flush should be deferred to the end of a batch of
618  * unmap operations to reduce IPIs.
619  */
620 static bool should_defer_flush(struct mm_struct *mm, enum ttu_flags flags)
621 {
622 	bool should_defer = false;
623 
624 	if (!(flags & TTU_BATCH_FLUSH))
625 		return false;
626 
627 	/* If remote CPUs need to be flushed then defer batch the flush */
628 	if (cpumask_any_but(mm_cpumask(mm), get_cpu()) < nr_cpu_ids)
629 		should_defer = true;
630 	put_cpu();
631 
632 	return should_defer;
633 }
634 #else
635 static void set_tlb_ubc_flush_pending(struct mm_struct *mm, bool writable)
636 {
637 }
638 
639 static bool should_defer_flush(struct mm_struct *mm, enum ttu_flags flags)
640 {
641 	return false;
642 }
643 #endif /* CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH */
644 
645 /*
646  * At what user virtual address is page expected in vma?
647  * Caller should check the page is actually part of the vma.
648  */
649 unsigned long page_address_in_vma(struct page *page, struct vm_area_struct *vma)
650 {
651 	unsigned long address;
652 	if (PageAnon(page)) {
653 		struct anon_vma *page__anon_vma = page_anon_vma(page);
654 		/*
655 		 * Note: swapoff's unuse_vma() is more efficient with this
656 		 * check, and needs it to match anon_vma when KSM is active.
657 		 */
658 		if (!vma->anon_vma || !page__anon_vma ||
659 		    vma->anon_vma->root != page__anon_vma->root)
660 			return -EFAULT;
661 	} else if (page->mapping) {
662 		if (!vma->vm_file || vma->vm_file->f_mapping != page->mapping)
663 			return -EFAULT;
664 	} else
665 		return -EFAULT;
666 	address = __vma_address(page, vma);
667 	if (unlikely(address < vma->vm_start || address >= vma->vm_end))
668 		return -EFAULT;
669 	return address;
670 }
671 
672 pmd_t *mm_find_pmd(struct mm_struct *mm, unsigned long address)
673 {
674 	pgd_t *pgd;
675 	p4d_t *p4d;
676 	pud_t *pud;
677 	pmd_t *pmd = NULL;
678 	pmd_t pmde;
679 
680 	pgd = pgd_offset(mm, address);
681 	if (!pgd_present(*pgd))
682 		goto out;
683 
684 	p4d = p4d_offset(pgd, address);
685 	if (!p4d_present(*p4d))
686 		goto out;
687 
688 	pud = pud_offset(p4d, address);
689 	if (!pud_present(*pud))
690 		goto out;
691 
692 	pmd = pmd_offset(pud, address);
693 	/*
694 	 * Some THP functions use the sequence pmdp_huge_clear_flush(), set_pmd_at()
695 	 * without holding anon_vma lock for write.  So when looking for a
696 	 * genuine pmde (in which to find pte), test present and !THP together.
697 	 */
698 	pmde = *pmd;
699 	barrier();
700 	if (!pmd_present(pmde) || pmd_trans_huge(pmde))
701 		pmd = NULL;
702 out:
703 	return pmd;
704 }
705 
706 struct page_referenced_arg {
707 	int mapcount;
708 	int referenced;
709 	unsigned long vm_flags;
710 	struct mem_cgroup *memcg;
711 };
712 /*
713  * arg: page_referenced_arg will be passed
714  */
715 static bool page_referenced_one(struct page *page, struct vm_area_struct *vma,
716 			unsigned long address, void *arg)
717 {
718 	struct page_referenced_arg *pra = arg;
719 	struct page_vma_mapped_walk pvmw = {
720 		.page = page,
721 		.vma = vma,
722 		.address = address,
723 	};
724 	int referenced = 0;
725 
726 	while (page_vma_mapped_walk(&pvmw)) {
727 		address = pvmw.address;
728 
729 		if (vma->vm_flags & VM_LOCKED) {
730 			page_vma_mapped_walk_done(&pvmw);
731 			pra->vm_flags |= VM_LOCKED;
732 			return false; /* To break the loop */
733 		}
734 
735 		if (pvmw.pte) {
736 			if (ptep_clear_flush_young_notify(vma, address,
737 						pvmw.pte)) {
738 				/*
739 				 * Don't treat a reference through
740 				 * a sequentially read mapping as such.
741 				 * If the page has been used in another mapping,
742 				 * we will catch it; if this other mapping is
743 				 * already gone, the unmap path will have set
744 				 * PG_referenced or activated the page.
745 				 */
746 				if (likely(!(vma->vm_flags & VM_SEQ_READ)))
747 					referenced++;
748 			}
749 		} else if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE)) {
750 			if (pmdp_clear_flush_young_notify(vma, address,
751 						pvmw.pmd))
752 				referenced++;
753 		} else {
754 			/* unexpected pmd-mapped page? */
755 			WARN_ON_ONCE(1);
756 		}
757 
758 		pra->mapcount--;
759 	}
760 
761 	if (referenced)
762 		clear_page_idle(page);
763 	if (test_and_clear_page_young(page))
764 		referenced++;
765 
766 	if (referenced) {
767 		pra->referenced++;
768 		pra->vm_flags |= vma->vm_flags;
769 	}
770 
771 	if (!pra->mapcount)
772 		return false; /* To break the loop */
773 
774 	return true;
775 }
776 
777 static bool invalid_page_referenced_vma(struct vm_area_struct *vma, void *arg)
778 {
779 	struct page_referenced_arg *pra = arg;
780 	struct mem_cgroup *memcg = pra->memcg;
781 
782 	if (!mm_match_cgroup(vma->vm_mm, memcg))
783 		return true;
784 
785 	return false;
786 }
787 
788 /**
789  * page_referenced - test if the page was referenced
790  * @page: the page to test
791  * @is_locked: caller holds lock on the page
792  * @memcg: target memory cgroup
793  * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
794  *
795  * Quick test_and_clear_referenced for all mappings to a page,
796  * returns the number of ptes which referenced the page.
797  */
798 int page_referenced(struct page *page,
799 		    int is_locked,
800 		    struct mem_cgroup *memcg,
801 		    unsigned long *vm_flags)
802 {
803 	int we_locked = 0;
804 	struct page_referenced_arg pra = {
805 		.mapcount = total_mapcount(page),
806 		.memcg = memcg,
807 	};
808 	struct rmap_walk_control rwc = {
809 		.rmap_one = page_referenced_one,
810 		.arg = (void *)&pra,
811 		.anon_lock = page_lock_anon_vma_read,
812 	};
813 
814 	*vm_flags = 0;
815 	if (!page_mapped(page))
816 		return 0;
817 
818 	if (!page_rmapping(page))
819 		return 0;
820 
821 	if (!is_locked && (!PageAnon(page) || PageKsm(page))) {
822 		we_locked = trylock_page(page);
823 		if (!we_locked)
824 			return 1;
825 	}
826 
827 	/*
828 	 * If we are reclaiming on behalf of a cgroup, skip
829 	 * counting on behalf of references from different
830 	 * cgroups
831 	 */
832 	if (memcg) {
833 		rwc.invalid_vma = invalid_page_referenced_vma;
834 	}
835 
836 	rmap_walk(page, &rwc);
837 	*vm_flags = pra.vm_flags;
838 
839 	if (we_locked)
840 		unlock_page(page);
841 
842 	return pra.referenced;
843 }
844 
845 static bool page_mkclean_one(struct page *page, struct vm_area_struct *vma,
846 			    unsigned long address, void *arg)
847 {
848 	struct page_vma_mapped_walk pvmw = {
849 		.page = page,
850 		.vma = vma,
851 		.address = address,
852 		.flags = PVMW_SYNC,
853 	};
854 	int *cleaned = arg;
855 
856 	while (page_vma_mapped_walk(&pvmw)) {
857 		int ret = 0;
858 		address = pvmw.address;
859 		if (pvmw.pte) {
860 			pte_t entry;
861 			pte_t *pte = pvmw.pte;
862 
863 			if (!pte_dirty(*pte) && !pte_write(*pte))
864 				continue;
865 
866 			flush_cache_page(vma, address, pte_pfn(*pte));
867 			entry = ptep_clear_flush(vma, address, pte);
868 			entry = pte_wrprotect(entry);
869 			entry = pte_mkclean(entry);
870 			set_pte_at(vma->vm_mm, address, pte, entry);
871 			ret = 1;
872 		} else {
873 #ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
874 			pmd_t *pmd = pvmw.pmd;
875 			pmd_t entry;
876 
877 			if (!pmd_dirty(*pmd) && !pmd_write(*pmd))
878 				continue;
879 
880 			flush_cache_page(vma, address, page_to_pfn(page));
881 			entry = pmdp_huge_clear_flush(vma, address, pmd);
882 			entry = pmd_wrprotect(entry);
883 			entry = pmd_mkclean(entry);
884 			set_pmd_at(vma->vm_mm, address, pmd, entry);
885 			ret = 1;
886 #else
887 			/* unexpected pmd-mapped page? */
888 			WARN_ON_ONCE(1);
889 #endif
890 		}
891 
892 		if (ret) {
893 			mmu_notifier_invalidate_page(vma->vm_mm, address);
894 			(*cleaned)++;
895 		}
896 	}
897 
898 	return true;
899 }
900 
901 static bool invalid_mkclean_vma(struct vm_area_struct *vma, void *arg)
902 {
903 	if (vma->vm_flags & VM_SHARED)
904 		return false;
905 
906 	return true;
907 }
908 
909 int page_mkclean(struct page *page)
910 {
911 	int cleaned = 0;
912 	struct address_space *mapping;
913 	struct rmap_walk_control rwc = {
914 		.arg = (void *)&cleaned,
915 		.rmap_one = page_mkclean_one,
916 		.invalid_vma = invalid_mkclean_vma,
917 	};
918 
919 	BUG_ON(!PageLocked(page));
920 
921 	if (!page_mapped(page))
922 		return 0;
923 
924 	mapping = page_mapping(page);
925 	if (!mapping)
926 		return 0;
927 
928 	rmap_walk(page, &rwc);
929 
930 	return cleaned;
931 }
932 EXPORT_SYMBOL_GPL(page_mkclean);
933 
934 /**
935  * page_move_anon_rmap - move a page to our anon_vma
936  * @page:	the page to move to our anon_vma
937  * @vma:	the vma the page belongs to
938  *
939  * When a page belongs exclusively to one process after a COW event,
940  * that page can be moved into the anon_vma that belongs to just that
941  * process, so the rmap code will not search the parent or sibling
942  * processes.
943  */
944 void page_move_anon_rmap(struct page *page, struct vm_area_struct *vma)
945 {
946 	struct anon_vma *anon_vma = vma->anon_vma;
947 
948 	page = compound_head(page);
949 
950 	VM_BUG_ON_PAGE(!PageLocked(page), page);
951 	VM_BUG_ON_VMA(!anon_vma, vma);
952 
953 	anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
954 	/*
955 	 * Ensure that anon_vma and the PAGE_MAPPING_ANON bit are written
956 	 * simultaneously, so a concurrent reader (eg page_referenced()'s
957 	 * PageAnon()) will not see one without the other.
958 	 */
959 	WRITE_ONCE(page->mapping, (struct address_space *) anon_vma);
960 }
961 
962 /**
963  * __page_set_anon_rmap - set up new anonymous rmap
964  * @page:	Page to add to rmap
965  * @vma:	VM area to add page to.
966  * @address:	User virtual address of the mapping
967  * @exclusive:	the page is exclusively owned by the current process
968  */
969 static void __page_set_anon_rmap(struct page *page,
970 	struct vm_area_struct *vma, unsigned long address, int exclusive)
971 {
972 	struct anon_vma *anon_vma = vma->anon_vma;
973 
974 	BUG_ON(!anon_vma);
975 
976 	if (PageAnon(page))
977 		return;
978 
979 	/*
980 	 * If the page isn't exclusively mapped into this vma,
981 	 * we must use the _oldest_ possible anon_vma for the
982 	 * page mapping!
983 	 */
984 	if (!exclusive)
985 		anon_vma = anon_vma->root;
986 
987 	anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
988 	page->mapping = (struct address_space *) anon_vma;
989 	page->index = linear_page_index(vma, address);
990 }
991 
992 /**
993  * __page_check_anon_rmap - sanity check anonymous rmap addition
994  * @page:	the page to add the mapping to
995  * @vma:	the vm area in which the mapping is added
996  * @address:	the user virtual address mapped
997  */
998 static void __page_check_anon_rmap(struct page *page,
999 	struct vm_area_struct *vma, unsigned long address)
1000 {
1001 #ifdef CONFIG_DEBUG_VM
1002 	/*
1003 	 * The page's anon-rmap details (mapping and index) are guaranteed to
1004 	 * be set up correctly at this point.
1005 	 *
1006 	 * We have exclusion against page_add_anon_rmap because the caller
1007 	 * always holds the page locked, except if called from page_dup_rmap,
1008 	 * in which case the page is already known to be setup.
1009 	 *
1010 	 * We have exclusion against page_add_new_anon_rmap because those pages
1011 	 * are initially only visible via the pagetables, and the pte is locked
1012 	 * over the call to page_add_new_anon_rmap.
1013 	 */
1014 	BUG_ON(page_anon_vma(page)->root != vma->anon_vma->root);
1015 	BUG_ON(page_to_pgoff(page) != linear_page_index(vma, address));
1016 #endif
1017 }
1018 
1019 /**
1020  * page_add_anon_rmap - add pte mapping to an anonymous page
1021  * @page:	the page to add the mapping to
1022  * @vma:	the vm area in which the mapping is added
1023  * @address:	the user virtual address mapped
1024  * @compound:	charge the page as compound or small page
1025  *
1026  * The caller needs to hold the pte lock, and the page must be locked in
1027  * the anon_vma case: to serialize mapping,index checking after setting,
1028  * and to ensure that PageAnon is not being upgraded racily to PageKsm
1029  * (but PageKsm is never downgraded to PageAnon).
1030  */
1031 void page_add_anon_rmap(struct page *page,
1032 	struct vm_area_struct *vma, unsigned long address, bool compound)
1033 {
1034 	do_page_add_anon_rmap(page, vma, address, compound ? RMAP_COMPOUND : 0);
1035 }
1036 
1037 /*
1038  * Special version of the above for do_swap_page, which often runs
1039  * into pages that are exclusively owned by the current process.
1040  * Everybody else should continue to use page_add_anon_rmap above.
1041  */
1042 void do_page_add_anon_rmap(struct page *page,
1043 	struct vm_area_struct *vma, unsigned long address, int flags)
1044 {
1045 	bool compound = flags & RMAP_COMPOUND;
1046 	bool first;
1047 
1048 	if (compound) {
1049 		atomic_t *mapcount;
1050 		VM_BUG_ON_PAGE(!PageLocked(page), page);
1051 		VM_BUG_ON_PAGE(!PageTransHuge(page), page);
1052 		mapcount = compound_mapcount_ptr(page);
1053 		first = atomic_inc_and_test(mapcount);
1054 	} else {
1055 		first = atomic_inc_and_test(&page->_mapcount);
1056 	}
1057 
1058 	if (first) {
1059 		int nr = compound ? hpage_nr_pages(page) : 1;
1060 		/*
1061 		 * We use the irq-unsafe __{inc|mod}_zone_page_stat because
1062 		 * these counters are not modified in interrupt context, and
1063 		 * pte lock(a spinlock) is held, which implies preemption
1064 		 * disabled.
1065 		 */
1066 		if (compound)
1067 			__inc_node_page_state(page, NR_ANON_THPS);
1068 		__mod_node_page_state(page_pgdat(page), NR_ANON_MAPPED, nr);
1069 	}
1070 	if (unlikely(PageKsm(page)))
1071 		return;
1072 
1073 	VM_BUG_ON_PAGE(!PageLocked(page), page);
1074 
1075 	/* address might be in next vma when migration races vma_adjust */
1076 	if (first)
1077 		__page_set_anon_rmap(page, vma, address,
1078 				flags & RMAP_EXCLUSIVE);
1079 	else
1080 		__page_check_anon_rmap(page, vma, address);
1081 }
1082 
1083 /**
1084  * page_add_new_anon_rmap - add pte mapping to a new anonymous page
1085  * @page:	the page to add the mapping to
1086  * @vma:	the vm area in which the mapping is added
1087  * @address:	the user virtual address mapped
1088  * @compound:	charge the page as compound or small page
1089  *
1090  * Same as page_add_anon_rmap but must only be called on *new* pages.
1091  * This means the inc-and-test can be bypassed.
1092  * Page does not have to be locked.
1093  */
1094 void page_add_new_anon_rmap(struct page *page,
1095 	struct vm_area_struct *vma, unsigned long address, bool compound)
1096 {
1097 	int nr = compound ? hpage_nr_pages(page) : 1;
1098 
1099 	VM_BUG_ON_VMA(address < vma->vm_start || address >= vma->vm_end, vma);
1100 	__SetPageSwapBacked(page);
1101 	if (compound) {
1102 		VM_BUG_ON_PAGE(!PageTransHuge(page), page);
1103 		/* increment count (starts at -1) */
1104 		atomic_set(compound_mapcount_ptr(page), 0);
1105 		__inc_node_page_state(page, NR_ANON_THPS);
1106 	} else {
1107 		/* Anon THP always mapped first with PMD */
1108 		VM_BUG_ON_PAGE(PageTransCompound(page), page);
1109 		/* increment count (starts at -1) */
1110 		atomic_set(&page->_mapcount, 0);
1111 	}
1112 	__mod_node_page_state(page_pgdat(page), NR_ANON_MAPPED, nr);
1113 	__page_set_anon_rmap(page, vma, address, 1);
1114 }
1115 
1116 /**
1117  * page_add_file_rmap - add pte mapping to a file page
1118  * @page: the page to add the mapping to
1119  *
1120  * The caller needs to hold the pte lock.
1121  */
1122 void page_add_file_rmap(struct page *page, bool compound)
1123 {
1124 	int i, nr = 1;
1125 
1126 	VM_BUG_ON_PAGE(compound && !PageTransHuge(page), page);
1127 	lock_page_memcg(page);
1128 	if (compound && PageTransHuge(page)) {
1129 		for (i = 0, nr = 0; i < HPAGE_PMD_NR; i++) {
1130 			if (atomic_inc_and_test(&page[i]._mapcount))
1131 				nr++;
1132 		}
1133 		if (!atomic_inc_and_test(compound_mapcount_ptr(page)))
1134 			goto out;
1135 		VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
1136 		__inc_node_page_state(page, NR_SHMEM_PMDMAPPED);
1137 	} else {
1138 		if (PageTransCompound(page) && page_mapping(page)) {
1139 			VM_WARN_ON_ONCE(!PageLocked(page));
1140 
1141 			SetPageDoubleMap(compound_head(page));
1142 			if (PageMlocked(page))
1143 				clear_page_mlock(compound_head(page));
1144 		}
1145 		if (!atomic_inc_and_test(&page->_mapcount))
1146 			goto out;
1147 	}
1148 	__mod_lruvec_page_state(page, NR_FILE_MAPPED, nr);
1149 out:
1150 	unlock_page_memcg(page);
1151 }
1152 
1153 static void page_remove_file_rmap(struct page *page, bool compound)
1154 {
1155 	int i, nr = 1;
1156 
1157 	VM_BUG_ON_PAGE(compound && !PageHead(page), page);
1158 	lock_page_memcg(page);
1159 
1160 	/* Hugepages are not counted in NR_FILE_MAPPED for now. */
1161 	if (unlikely(PageHuge(page))) {
1162 		/* hugetlb pages are always mapped with pmds */
1163 		atomic_dec(compound_mapcount_ptr(page));
1164 		goto out;
1165 	}
1166 
1167 	/* page still mapped by someone else? */
1168 	if (compound && PageTransHuge(page)) {
1169 		for (i = 0, nr = 0; i < HPAGE_PMD_NR; i++) {
1170 			if (atomic_add_negative(-1, &page[i]._mapcount))
1171 				nr++;
1172 		}
1173 		if (!atomic_add_negative(-1, compound_mapcount_ptr(page)))
1174 			goto out;
1175 		VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
1176 		__dec_node_page_state(page, NR_SHMEM_PMDMAPPED);
1177 	} else {
1178 		if (!atomic_add_negative(-1, &page->_mapcount))
1179 			goto out;
1180 	}
1181 
1182 	/*
1183 	 * We use the irq-unsafe __{inc|mod}_lruvec_page_state because
1184 	 * these counters are not modified in interrupt context, and
1185 	 * pte lock(a spinlock) is held, which implies preemption disabled.
1186 	 */
1187 	__mod_lruvec_page_state(page, NR_FILE_MAPPED, -nr);
1188 
1189 	if (unlikely(PageMlocked(page)))
1190 		clear_page_mlock(page);
1191 out:
1192 	unlock_page_memcg(page);
1193 }
1194 
1195 static void page_remove_anon_compound_rmap(struct page *page)
1196 {
1197 	int i, nr;
1198 
1199 	if (!atomic_add_negative(-1, compound_mapcount_ptr(page)))
1200 		return;
1201 
1202 	/* Hugepages are not counted in NR_ANON_PAGES for now. */
1203 	if (unlikely(PageHuge(page)))
1204 		return;
1205 
1206 	if (!IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE))
1207 		return;
1208 
1209 	__dec_node_page_state(page, NR_ANON_THPS);
1210 
1211 	if (TestClearPageDoubleMap(page)) {
1212 		/*
1213 		 * Subpages can be mapped with PTEs too. Check how many of
1214 		 * themi are still mapped.
1215 		 */
1216 		for (i = 0, nr = 0; i < HPAGE_PMD_NR; i++) {
1217 			if (atomic_add_negative(-1, &page[i]._mapcount))
1218 				nr++;
1219 		}
1220 	} else {
1221 		nr = HPAGE_PMD_NR;
1222 	}
1223 
1224 	if (unlikely(PageMlocked(page)))
1225 		clear_page_mlock(page);
1226 
1227 	if (nr) {
1228 		__mod_node_page_state(page_pgdat(page), NR_ANON_MAPPED, -nr);
1229 		deferred_split_huge_page(page);
1230 	}
1231 }
1232 
1233 /**
1234  * page_remove_rmap - take down pte mapping from a page
1235  * @page:	page to remove mapping from
1236  * @compound:	uncharge the page as compound or small page
1237  *
1238  * The caller needs to hold the pte lock.
1239  */
1240 void page_remove_rmap(struct page *page, bool compound)
1241 {
1242 	if (!PageAnon(page))
1243 		return page_remove_file_rmap(page, compound);
1244 
1245 	if (compound)
1246 		return page_remove_anon_compound_rmap(page);
1247 
1248 	/* page still mapped by someone else? */
1249 	if (!atomic_add_negative(-1, &page->_mapcount))
1250 		return;
1251 
1252 	/*
1253 	 * We use the irq-unsafe __{inc|mod}_zone_page_stat because
1254 	 * these counters are not modified in interrupt context, and
1255 	 * pte lock(a spinlock) is held, which implies preemption disabled.
1256 	 */
1257 	__dec_node_page_state(page, NR_ANON_MAPPED);
1258 
1259 	if (unlikely(PageMlocked(page)))
1260 		clear_page_mlock(page);
1261 
1262 	if (PageTransCompound(page))
1263 		deferred_split_huge_page(compound_head(page));
1264 
1265 	/*
1266 	 * It would be tidy to reset the PageAnon mapping here,
1267 	 * but that might overwrite a racing page_add_anon_rmap
1268 	 * which increments mapcount after us but sets mapping
1269 	 * before us: so leave the reset to free_hot_cold_page,
1270 	 * and remember that it's only reliable while mapped.
1271 	 * Leaving it set also helps swapoff to reinstate ptes
1272 	 * faster for those pages still in swapcache.
1273 	 */
1274 }
1275 
1276 /*
1277  * @arg: enum ttu_flags will be passed to this argument
1278  */
1279 static bool try_to_unmap_one(struct page *page, struct vm_area_struct *vma,
1280 		     unsigned long address, void *arg)
1281 {
1282 	struct mm_struct *mm = vma->vm_mm;
1283 	struct page_vma_mapped_walk pvmw = {
1284 		.page = page,
1285 		.vma = vma,
1286 		.address = address,
1287 	};
1288 	pte_t pteval;
1289 	struct page *subpage;
1290 	bool ret = true;
1291 	enum ttu_flags flags = (enum ttu_flags)arg;
1292 
1293 	/* munlock has nothing to gain from examining un-locked vmas */
1294 	if ((flags & TTU_MUNLOCK) && !(vma->vm_flags & VM_LOCKED))
1295 		return true;
1296 
1297 	if (flags & TTU_SPLIT_HUGE_PMD) {
1298 		split_huge_pmd_address(vma, address,
1299 				flags & TTU_MIGRATION, page);
1300 	}
1301 
1302 	while (page_vma_mapped_walk(&pvmw)) {
1303 		/*
1304 		 * If the page is mlock()d, we cannot swap it out.
1305 		 * If it's recently referenced (perhaps page_referenced
1306 		 * skipped over this mm) then we should reactivate it.
1307 		 */
1308 		if (!(flags & TTU_IGNORE_MLOCK)) {
1309 			if (vma->vm_flags & VM_LOCKED) {
1310 				/* PTE-mapped THP are never mlocked */
1311 				if (!PageTransCompound(page)) {
1312 					/*
1313 					 * Holding pte lock, we do *not* need
1314 					 * mmap_sem here
1315 					 */
1316 					mlock_vma_page(page);
1317 				}
1318 				ret = false;
1319 				page_vma_mapped_walk_done(&pvmw);
1320 				break;
1321 			}
1322 			if (flags & TTU_MUNLOCK)
1323 				continue;
1324 		}
1325 
1326 		/* Unexpected PMD-mapped THP? */
1327 		VM_BUG_ON_PAGE(!pvmw.pte, page);
1328 
1329 		subpage = page - page_to_pfn(page) + pte_pfn(*pvmw.pte);
1330 		address = pvmw.address;
1331 
1332 
1333 		if (!(flags & TTU_IGNORE_ACCESS)) {
1334 			if (ptep_clear_flush_young_notify(vma, address,
1335 						pvmw.pte)) {
1336 				ret = false;
1337 				page_vma_mapped_walk_done(&pvmw);
1338 				break;
1339 			}
1340 		}
1341 
1342 		/* Nuke the page table entry. */
1343 		flush_cache_page(vma, address, pte_pfn(*pvmw.pte));
1344 		if (should_defer_flush(mm, flags)) {
1345 			/*
1346 			 * We clear the PTE but do not flush so potentially
1347 			 * a remote CPU could still be writing to the page.
1348 			 * If the entry was previously clean then the
1349 			 * architecture must guarantee that a clear->dirty
1350 			 * transition on a cached TLB entry is written through
1351 			 * and traps if the PTE is unmapped.
1352 			 */
1353 			pteval = ptep_get_and_clear(mm, address, pvmw.pte);
1354 
1355 			set_tlb_ubc_flush_pending(mm, pte_dirty(pteval));
1356 		} else {
1357 			pteval = ptep_clear_flush(vma, address, pvmw.pte);
1358 		}
1359 
1360 		/* Move the dirty bit to the page. Now the pte is gone. */
1361 		if (pte_dirty(pteval))
1362 			set_page_dirty(page);
1363 
1364 		/* Update high watermark before we lower rss */
1365 		update_hiwater_rss(mm);
1366 
1367 		if (PageHWPoison(page) && !(flags & TTU_IGNORE_HWPOISON)) {
1368 			pteval = swp_entry_to_pte(make_hwpoison_entry(subpage));
1369 			if (PageHuge(page)) {
1370 				int nr = 1 << compound_order(page);
1371 				hugetlb_count_sub(nr, mm);
1372 				set_huge_swap_pte_at(mm, address,
1373 						     pvmw.pte, pteval,
1374 						     vma_mmu_pagesize(vma));
1375 			} else {
1376 				dec_mm_counter(mm, mm_counter(page));
1377 				set_pte_at(mm, address, pvmw.pte, pteval);
1378 			}
1379 
1380 		} else if (pte_unused(pteval)) {
1381 			/*
1382 			 * The guest indicated that the page content is of no
1383 			 * interest anymore. Simply discard the pte, vmscan
1384 			 * will take care of the rest.
1385 			 */
1386 			dec_mm_counter(mm, mm_counter(page));
1387 		} else if (IS_ENABLED(CONFIG_MIGRATION) &&
1388 				(flags & TTU_MIGRATION)) {
1389 			swp_entry_t entry;
1390 			pte_t swp_pte;
1391 			/*
1392 			 * Store the pfn of the page in a special migration
1393 			 * pte. do_swap_page() will wait until the migration
1394 			 * pte is removed and then restart fault handling.
1395 			 */
1396 			entry = make_migration_entry(subpage,
1397 					pte_write(pteval));
1398 			swp_pte = swp_entry_to_pte(entry);
1399 			if (pte_soft_dirty(pteval))
1400 				swp_pte = pte_swp_mksoft_dirty(swp_pte);
1401 			set_pte_at(mm, address, pvmw.pte, swp_pte);
1402 		} else if (PageAnon(page)) {
1403 			swp_entry_t entry = { .val = page_private(subpage) };
1404 			pte_t swp_pte;
1405 			/*
1406 			 * Store the swap location in the pte.
1407 			 * See handle_pte_fault() ...
1408 			 */
1409 			if (unlikely(PageSwapBacked(page) != PageSwapCache(page))) {
1410 				WARN_ON_ONCE(1);
1411 				ret = false;
1412 				page_vma_mapped_walk_done(&pvmw);
1413 				break;
1414 			}
1415 
1416 			/* MADV_FREE page check */
1417 			if (!PageSwapBacked(page)) {
1418 				if (!PageDirty(page)) {
1419 					dec_mm_counter(mm, MM_ANONPAGES);
1420 					goto discard;
1421 				}
1422 
1423 				/*
1424 				 * If the page was redirtied, it cannot be
1425 				 * discarded. Remap the page to page table.
1426 				 */
1427 				set_pte_at(mm, address, pvmw.pte, pteval);
1428 				SetPageSwapBacked(page);
1429 				ret = false;
1430 				page_vma_mapped_walk_done(&pvmw);
1431 				break;
1432 			}
1433 
1434 			if (swap_duplicate(entry) < 0) {
1435 				set_pte_at(mm, address, pvmw.pte, pteval);
1436 				ret = false;
1437 				page_vma_mapped_walk_done(&pvmw);
1438 				break;
1439 			}
1440 			if (list_empty(&mm->mmlist)) {
1441 				spin_lock(&mmlist_lock);
1442 				if (list_empty(&mm->mmlist))
1443 					list_add(&mm->mmlist, &init_mm.mmlist);
1444 				spin_unlock(&mmlist_lock);
1445 			}
1446 			dec_mm_counter(mm, MM_ANONPAGES);
1447 			inc_mm_counter(mm, MM_SWAPENTS);
1448 			swp_pte = swp_entry_to_pte(entry);
1449 			if (pte_soft_dirty(pteval))
1450 				swp_pte = pte_swp_mksoft_dirty(swp_pte);
1451 			set_pte_at(mm, address, pvmw.pte, swp_pte);
1452 		} else
1453 			dec_mm_counter(mm, mm_counter_file(page));
1454 discard:
1455 		page_remove_rmap(subpage, PageHuge(page));
1456 		put_page(page);
1457 		mmu_notifier_invalidate_page(mm, address);
1458 	}
1459 	return ret;
1460 }
1461 
1462 bool is_vma_temporary_stack(struct vm_area_struct *vma)
1463 {
1464 	int maybe_stack = vma->vm_flags & (VM_GROWSDOWN | VM_GROWSUP);
1465 
1466 	if (!maybe_stack)
1467 		return false;
1468 
1469 	if ((vma->vm_flags & VM_STACK_INCOMPLETE_SETUP) ==
1470 						VM_STACK_INCOMPLETE_SETUP)
1471 		return true;
1472 
1473 	return false;
1474 }
1475 
1476 static bool invalid_migration_vma(struct vm_area_struct *vma, void *arg)
1477 {
1478 	return is_vma_temporary_stack(vma);
1479 }
1480 
1481 static int page_mapcount_is_zero(struct page *page)
1482 {
1483 	return !total_mapcount(page);
1484 }
1485 
1486 /**
1487  * try_to_unmap - try to remove all page table mappings to a page
1488  * @page: the page to get unmapped
1489  * @flags: action and flags
1490  *
1491  * Tries to remove all the page table entries which are mapping this
1492  * page, used in the pageout path.  Caller must hold the page lock.
1493  *
1494  * If unmap is successful, return true. Otherwise, false.
1495  */
1496 bool try_to_unmap(struct page *page, enum ttu_flags flags)
1497 {
1498 	struct rmap_walk_control rwc = {
1499 		.rmap_one = try_to_unmap_one,
1500 		.arg = (void *)flags,
1501 		.done = page_mapcount_is_zero,
1502 		.anon_lock = page_lock_anon_vma_read,
1503 	};
1504 
1505 	/*
1506 	 * During exec, a temporary VMA is setup and later moved.
1507 	 * The VMA is moved under the anon_vma lock but not the
1508 	 * page tables leading to a race where migration cannot
1509 	 * find the migration ptes. Rather than increasing the
1510 	 * locking requirements of exec(), migration skips
1511 	 * temporary VMAs until after exec() completes.
1512 	 */
1513 	if ((flags & TTU_MIGRATION) && !PageKsm(page) && PageAnon(page))
1514 		rwc.invalid_vma = invalid_migration_vma;
1515 
1516 	if (flags & TTU_RMAP_LOCKED)
1517 		rmap_walk_locked(page, &rwc);
1518 	else
1519 		rmap_walk(page, &rwc);
1520 
1521 	return !page_mapcount(page) ? true : false;
1522 }
1523 
1524 static int page_not_mapped(struct page *page)
1525 {
1526 	return !page_mapped(page);
1527 };
1528 
1529 /**
1530  * try_to_munlock - try to munlock a page
1531  * @page: the page to be munlocked
1532  *
1533  * Called from munlock code.  Checks all of the VMAs mapping the page
1534  * to make sure nobody else has this page mlocked. The page will be
1535  * returned with PG_mlocked cleared if no other vmas have it mlocked.
1536  */
1537 
1538 void try_to_munlock(struct page *page)
1539 {
1540 	struct rmap_walk_control rwc = {
1541 		.rmap_one = try_to_unmap_one,
1542 		.arg = (void *)TTU_MUNLOCK,
1543 		.done = page_not_mapped,
1544 		.anon_lock = page_lock_anon_vma_read,
1545 
1546 	};
1547 
1548 	VM_BUG_ON_PAGE(!PageLocked(page) || PageLRU(page), page);
1549 	VM_BUG_ON_PAGE(PageCompound(page) && PageDoubleMap(page), page);
1550 
1551 	rmap_walk(page, &rwc);
1552 }
1553 
1554 void __put_anon_vma(struct anon_vma *anon_vma)
1555 {
1556 	struct anon_vma *root = anon_vma->root;
1557 
1558 	anon_vma_free(anon_vma);
1559 	if (root != anon_vma && atomic_dec_and_test(&root->refcount))
1560 		anon_vma_free(root);
1561 }
1562 
1563 static struct anon_vma *rmap_walk_anon_lock(struct page *page,
1564 					struct rmap_walk_control *rwc)
1565 {
1566 	struct anon_vma *anon_vma;
1567 
1568 	if (rwc->anon_lock)
1569 		return rwc->anon_lock(page);
1570 
1571 	/*
1572 	 * Note: remove_migration_ptes() cannot use page_lock_anon_vma_read()
1573 	 * because that depends on page_mapped(); but not all its usages
1574 	 * are holding mmap_sem. Users without mmap_sem are required to
1575 	 * take a reference count to prevent the anon_vma disappearing
1576 	 */
1577 	anon_vma = page_anon_vma(page);
1578 	if (!anon_vma)
1579 		return NULL;
1580 
1581 	anon_vma_lock_read(anon_vma);
1582 	return anon_vma;
1583 }
1584 
1585 /*
1586  * rmap_walk_anon - do something to anonymous page using the object-based
1587  * rmap method
1588  * @page: the page to be handled
1589  * @rwc: control variable according to each walk type
1590  *
1591  * Find all the mappings of a page using the mapping pointer and the vma chains
1592  * contained in the anon_vma struct it points to.
1593  *
1594  * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1595  * where the page was found will be held for write.  So, we won't recheck
1596  * vm_flags for that VMA.  That should be OK, because that vma shouldn't be
1597  * LOCKED.
1598  */
1599 static void rmap_walk_anon(struct page *page, struct rmap_walk_control *rwc,
1600 		bool locked)
1601 {
1602 	struct anon_vma *anon_vma;
1603 	pgoff_t pgoff_start, pgoff_end;
1604 	struct anon_vma_chain *avc;
1605 
1606 	if (locked) {
1607 		anon_vma = page_anon_vma(page);
1608 		/* anon_vma disappear under us? */
1609 		VM_BUG_ON_PAGE(!anon_vma, page);
1610 	} else {
1611 		anon_vma = rmap_walk_anon_lock(page, rwc);
1612 	}
1613 	if (!anon_vma)
1614 		return;
1615 
1616 	pgoff_start = page_to_pgoff(page);
1617 	pgoff_end = pgoff_start + hpage_nr_pages(page) - 1;
1618 	anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root,
1619 			pgoff_start, pgoff_end) {
1620 		struct vm_area_struct *vma = avc->vma;
1621 		unsigned long address = vma_address(page, vma);
1622 
1623 		cond_resched();
1624 
1625 		if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
1626 			continue;
1627 
1628 		if (!rwc->rmap_one(page, vma, address, rwc->arg))
1629 			break;
1630 		if (rwc->done && rwc->done(page))
1631 			break;
1632 	}
1633 
1634 	if (!locked)
1635 		anon_vma_unlock_read(anon_vma);
1636 }
1637 
1638 /*
1639  * rmap_walk_file - do something to file page using the object-based rmap method
1640  * @page: the page to be handled
1641  * @rwc: control variable according to each walk type
1642  *
1643  * Find all the mappings of a page using the mapping pointer and the vma chains
1644  * contained in the address_space struct it points to.
1645  *
1646  * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1647  * where the page was found will be held for write.  So, we won't recheck
1648  * vm_flags for that VMA.  That should be OK, because that vma shouldn't be
1649  * LOCKED.
1650  */
1651 static void rmap_walk_file(struct page *page, struct rmap_walk_control *rwc,
1652 		bool locked)
1653 {
1654 	struct address_space *mapping = page_mapping(page);
1655 	pgoff_t pgoff_start, pgoff_end;
1656 	struct vm_area_struct *vma;
1657 
1658 	/*
1659 	 * The page lock not only makes sure that page->mapping cannot
1660 	 * suddenly be NULLified by truncation, it makes sure that the
1661 	 * structure at mapping cannot be freed and reused yet,
1662 	 * so we can safely take mapping->i_mmap_rwsem.
1663 	 */
1664 	VM_BUG_ON_PAGE(!PageLocked(page), page);
1665 
1666 	if (!mapping)
1667 		return;
1668 
1669 	pgoff_start = page_to_pgoff(page);
1670 	pgoff_end = pgoff_start + hpage_nr_pages(page) - 1;
1671 	if (!locked)
1672 		i_mmap_lock_read(mapping);
1673 	vma_interval_tree_foreach(vma, &mapping->i_mmap,
1674 			pgoff_start, pgoff_end) {
1675 		unsigned long address = vma_address(page, vma);
1676 
1677 		cond_resched();
1678 
1679 		if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
1680 			continue;
1681 
1682 		if (!rwc->rmap_one(page, vma, address, rwc->arg))
1683 			goto done;
1684 		if (rwc->done && rwc->done(page))
1685 			goto done;
1686 	}
1687 
1688 done:
1689 	if (!locked)
1690 		i_mmap_unlock_read(mapping);
1691 }
1692 
1693 void rmap_walk(struct page *page, struct rmap_walk_control *rwc)
1694 {
1695 	if (unlikely(PageKsm(page)))
1696 		rmap_walk_ksm(page, rwc);
1697 	else if (PageAnon(page))
1698 		rmap_walk_anon(page, rwc, false);
1699 	else
1700 		rmap_walk_file(page, rwc, false);
1701 }
1702 
1703 /* Like rmap_walk, but caller holds relevant rmap lock */
1704 void rmap_walk_locked(struct page *page, struct rmap_walk_control *rwc)
1705 {
1706 	/* no ksm support for now */
1707 	VM_BUG_ON_PAGE(PageKsm(page), page);
1708 	if (PageAnon(page))
1709 		rmap_walk_anon(page, rwc, true);
1710 	else
1711 		rmap_walk_file(page, rwc, true);
1712 }
1713 
1714 #ifdef CONFIG_HUGETLB_PAGE
1715 /*
1716  * The following three functions are for anonymous (private mapped) hugepages.
1717  * Unlike common anonymous pages, anonymous hugepages have no accounting code
1718  * and no lru code, because we handle hugepages differently from common pages.
1719  */
1720 static void __hugepage_set_anon_rmap(struct page *page,
1721 	struct vm_area_struct *vma, unsigned long address, int exclusive)
1722 {
1723 	struct anon_vma *anon_vma = vma->anon_vma;
1724 
1725 	BUG_ON(!anon_vma);
1726 
1727 	if (PageAnon(page))
1728 		return;
1729 	if (!exclusive)
1730 		anon_vma = anon_vma->root;
1731 
1732 	anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
1733 	page->mapping = (struct address_space *) anon_vma;
1734 	page->index = linear_page_index(vma, address);
1735 }
1736 
1737 void hugepage_add_anon_rmap(struct page *page,
1738 			    struct vm_area_struct *vma, unsigned long address)
1739 {
1740 	struct anon_vma *anon_vma = vma->anon_vma;
1741 	int first;
1742 
1743 	BUG_ON(!PageLocked(page));
1744 	BUG_ON(!anon_vma);
1745 	/* address might be in next vma when migration races vma_adjust */
1746 	first = atomic_inc_and_test(compound_mapcount_ptr(page));
1747 	if (first)
1748 		__hugepage_set_anon_rmap(page, vma, address, 0);
1749 }
1750 
1751 void hugepage_add_new_anon_rmap(struct page *page,
1752 			struct vm_area_struct *vma, unsigned long address)
1753 {
1754 	BUG_ON(address < vma->vm_start || address >= vma->vm_end);
1755 	atomic_set(compound_mapcount_ptr(page), 0);
1756 	__hugepage_set_anon_rmap(page, vma, address, 1);
1757 }
1758 #endif /* CONFIG_HUGETLB_PAGE */
1759