• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * mm/rmap.c - physical to virtual reverse mappings
3  *
4  * Copyright 2001, Rik van Riel <riel@conectiva.com.br>
5  * Released under the General Public License (GPL).
6  *
7  * Simple, low overhead reverse mapping scheme.
8  * Please try to keep this thing as modular as possible.
9  *
10  * Provides methods for unmapping each kind of mapped page:
11  * the anon methods track anonymous pages, and
12  * the file methods track pages belonging to an inode.
13  *
14  * Original design by Rik van Riel <riel@conectiva.com.br> 2001
15  * File methods by Dave McCracken <dmccr@us.ibm.com> 2003, 2004
16  * Anonymous methods by Andrea Arcangeli <andrea@suse.de> 2004
17  * Contributions by Hugh Dickins 2003, 2004
18  */
19 
20 /*
21  * Lock ordering in mm:
22  *
23  * inode->i_mutex	(while writing or truncating, not reading or faulting)
24  *   mm->mmap_sem
25  *     page->flags PG_locked (lock_page)
26  *       mapping->i_mmap_rwsem
27  *         anon_vma->rwsem
28  *           mm->page_table_lock or pte_lock
29  *             zone->lru_lock (in mark_page_accessed, isolate_lru_page)
30  *             swap_lock (in swap_duplicate, swap_info_get)
31  *               mmlist_lock (in mmput, drain_mmlist and others)
32  *               mapping->private_lock (in __set_page_dirty_buffers)
33  *                 mem_cgroup_{begin,end}_page_stat (memcg->move_lock)
34  *                   mapping->tree_lock (widely used)
35  *               inode->i_lock (in set_page_dirty's __mark_inode_dirty)
36  *               bdi.wb->list_lock (in set_page_dirty's __mark_inode_dirty)
37  *                 sb_lock (within inode_lock in fs/fs-writeback.c)
38  *                 mapping->tree_lock (widely used, in set_page_dirty,
39  *                           in arch-dependent flush_dcache_mmap_lock,
40  *                           within bdi.wb->list_lock in __sync_single_inode)
41  *
42  * anon_vma->rwsem,mapping->i_mutex      (memory_failure, collect_procs_anon)
43  *   ->tasklist_lock
44  *     pte map lock
45  */
46 
47 #include <linux/mm.h>
48 #include <linux/pagemap.h>
49 #include <linux/swap.h>
50 #include <linux/swapops.h>
51 #include <linux/slab.h>
52 #include <linux/init.h>
53 #include <linux/ksm.h>
54 #include <linux/rmap.h>
55 #include <linux/rcupdate.h>
56 #include <linux/export.h>
57 #include <linux/memcontrol.h>
58 #include <linux/mmu_notifier.h>
59 #include <linux/migrate.h>
60 #include <linux/hugetlb.h>
61 #include <linux/backing-dev.h>
62 #include <linux/page_idle.h>
63 
64 #include <asm/tlbflush.h>
65 
66 #include <trace/events/tlb.h>
67 
68 #include "internal.h"
69 
70 static struct kmem_cache *anon_vma_cachep;
71 static struct kmem_cache *anon_vma_chain_cachep;
72 
anon_vma_alloc(void)73 static inline struct anon_vma *anon_vma_alloc(void)
74 {
75 	struct anon_vma *anon_vma;
76 
77 	anon_vma = kmem_cache_alloc(anon_vma_cachep, GFP_KERNEL);
78 	if (anon_vma) {
79 		atomic_set(&anon_vma->refcount, 1);
80 		anon_vma->degree = 1;	/* Reference for first vma */
81 		anon_vma->parent = anon_vma;
82 		/*
83 		 * Initialise the anon_vma root to point to itself. If called
84 		 * from fork, the root will be reset to the parents anon_vma.
85 		 */
86 		anon_vma->root = anon_vma;
87 	}
88 
89 	return anon_vma;
90 }
91 
anon_vma_free(struct anon_vma * anon_vma)92 static inline void anon_vma_free(struct anon_vma *anon_vma)
93 {
94 	VM_BUG_ON(atomic_read(&anon_vma->refcount));
95 
96 	/*
97 	 * Synchronize against page_lock_anon_vma_read() such that
98 	 * we can safely hold the lock without the anon_vma getting
99 	 * freed.
100 	 *
101 	 * Relies on the full mb implied by the atomic_dec_and_test() from
102 	 * put_anon_vma() against the acquire barrier implied by
103 	 * down_read_trylock() from page_lock_anon_vma_read(). This orders:
104 	 *
105 	 * page_lock_anon_vma_read()	VS	put_anon_vma()
106 	 *   down_read_trylock()		  atomic_dec_and_test()
107 	 *   LOCK				  MB
108 	 *   atomic_read()			  rwsem_is_locked()
109 	 *
110 	 * LOCK should suffice since the actual taking of the lock must
111 	 * happen _before_ what follows.
112 	 */
113 	might_sleep();
114 	if (rwsem_is_locked(&anon_vma->root->rwsem)) {
115 		anon_vma_lock_write(anon_vma);
116 		anon_vma_unlock_write(anon_vma);
117 	}
118 
119 	kmem_cache_free(anon_vma_cachep, anon_vma);
120 }
121 
anon_vma_chain_alloc(gfp_t gfp)122 static inline struct anon_vma_chain *anon_vma_chain_alloc(gfp_t gfp)
123 {
124 	return kmem_cache_alloc(anon_vma_chain_cachep, gfp);
125 }
126 
anon_vma_chain_free(struct anon_vma_chain * anon_vma_chain)127 static void anon_vma_chain_free(struct anon_vma_chain *anon_vma_chain)
128 {
129 	kmem_cache_free(anon_vma_chain_cachep, anon_vma_chain);
130 }
131 
anon_vma_chain_link(struct vm_area_struct * vma,struct anon_vma_chain * avc,struct anon_vma * anon_vma)132 static void anon_vma_chain_link(struct vm_area_struct *vma,
133 				struct anon_vma_chain *avc,
134 				struct anon_vma *anon_vma)
135 {
136 	avc->vma = vma;
137 	avc->anon_vma = anon_vma;
138 	list_add(&avc->same_vma, &vma->anon_vma_chain);
139 	anon_vma_interval_tree_insert(avc, &anon_vma->rb_root);
140 }
141 
142 /**
143  * anon_vma_prepare - attach an anon_vma to a memory region
144  * @vma: the memory region in question
145  *
146  * This makes sure the memory mapping described by 'vma' has
147  * an 'anon_vma' attached to it, so that we can associate the
148  * anonymous pages mapped into it with that anon_vma.
149  *
150  * The common case will be that we already have one, but if
151  * not we either need to find an adjacent mapping that we
152  * can re-use the anon_vma from (very common when the only
153  * reason for splitting a vma has been mprotect()), or we
154  * allocate a new one.
155  *
156  * Anon-vma allocations are very subtle, because we may have
157  * optimistically looked up an anon_vma in page_lock_anon_vma_read()
158  * and that may actually touch the spinlock even in the newly
159  * allocated vma (it depends on RCU to make sure that the
160  * anon_vma isn't actually destroyed).
161  *
162  * As a result, we need to do proper anon_vma locking even
163  * for the new allocation. At the same time, we do not want
164  * to do any locking for the common case of already having
165  * an anon_vma.
166  *
167  * This must be called with the mmap_sem held for reading.
168  */
anon_vma_prepare(struct vm_area_struct * vma)169 int anon_vma_prepare(struct vm_area_struct *vma)
170 {
171 	struct anon_vma *anon_vma = vma->anon_vma;
172 	struct anon_vma_chain *avc;
173 
174 	might_sleep();
175 	if (unlikely(!anon_vma)) {
176 		struct mm_struct *mm = vma->vm_mm;
177 		struct anon_vma *allocated;
178 
179 		avc = anon_vma_chain_alloc(GFP_KERNEL);
180 		if (!avc)
181 			goto out_enomem;
182 
183 		anon_vma = find_mergeable_anon_vma(vma);
184 		allocated = NULL;
185 		if (!anon_vma) {
186 			anon_vma = anon_vma_alloc();
187 			if (unlikely(!anon_vma))
188 				goto out_enomem_free_avc;
189 			allocated = anon_vma;
190 		}
191 
192 		anon_vma_lock_write(anon_vma);
193 		/* page_table_lock to protect against threads */
194 		spin_lock(&mm->page_table_lock);
195 		if (likely(!vma->anon_vma)) {
196 			vma->anon_vma = anon_vma;
197 			anon_vma_chain_link(vma, avc, anon_vma);
198 			/* vma reference or self-parent link for new root */
199 			anon_vma->degree++;
200 			allocated = NULL;
201 			avc = NULL;
202 		}
203 		spin_unlock(&mm->page_table_lock);
204 		anon_vma_unlock_write(anon_vma);
205 
206 		if (unlikely(allocated))
207 			put_anon_vma(allocated);
208 		if (unlikely(avc))
209 			anon_vma_chain_free(avc);
210 	}
211 	return 0;
212 
213  out_enomem_free_avc:
214 	anon_vma_chain_free(avc);
215  out_enomem:
216 	return -ENOMEM;
217 }
218 
219 /*
220  * This is a useful helper function for locking the anon_vma root as
221  * we traverse the vma->anon_vma_chain, looping over anon_vma's that
222  * have the same vma.
223  *
224  * Such anon_vma's should have the same root, so you'd expect to see
225  * just a single mutex_lock for the whole traversal.
226  */
lock_anon_vma_root(struct anon_vma * root,struct anon_vma * anon_vma)227 static inline struct anon_vma *lock_anon_vma_root(struct anon_vma *root, struct anon_vma *anon_vma)
228 {
229 	struct anon_vma *new_root = anon_vma->root;
230 	if (new_root != root) {
231 		if (WARN_ON_ONCE(root))
232 			up_write(&root->rwsem);
233 		root = new_root;
234 		down_write(&root->rwsem);
235 	}
236 	return root;
237 }
238 
unlock_anon_vma_root(struct anon_vma * root)239 static inline void unlock_anon_vma_root(struct anon_vma *root)
240 {
241 	if (root)
242 		up_write(&root->rwsem);
243 }
244 
245 /*
246  * Attach the anon_vmas from src to dst.
247  * Returns 0 on success, -ENOMEM on failure.
248  *
249  * If dst->anon_vma is NULL this function tries to find and reuse existing
250  * anon_vma which has no vmas and only one child anon_vma. This prevents
251  * degradation of anon_vma hierarchy to endless linear chain in case of
252  * constantly forking task. On the other hand, an anon_vma with more than one
253  * child isn't reused even if there was no alive vma, thus rmap walker has a
254  * good chance of avoiding scanning the whole hierarchy when it searches where
255  * page is mapped.
256  */
anon_vma_clone(struct vm_area_struct * dst,struct vm_area_struct * src)257 int anon_vma_clone(struct vm_area_struct *dst, struct vm_area_struct *src)
258 {
259 	struct anon_vma_chain *avc, *pavc;
260 	struct anon_vma *root = NULL;
261 
262 	list_for_each_entry_reverse(pavc, &src->anon_vma_chain, same_vma) {
263 		struct anon_vma *anon_vma;
264 
265 		avc = anon_vma_chain_alloc(GFP_NOWAIT | __GFP_NOWARN);
266 		if (unlikely(!avc)) {
267 			unlock_anon_vma_root(root);
268 			root = NULL;
269 			avc = anon_vma_chain_alloc(GFP_KERNEL);
270 			if (!avc)
271 				goto enomem_failure;
272 		}
273 		anon_vma = pavc->anon_vma;
274 		root = lock_anon_vma_root(root, anon_vma);
275 		anon_vma_chain_link(dst, avc, anon_vma);
276 
277 		/*
278 		 * Reuse existing anon_vma if its degree lower than two,
279 		 * that means it has no vma and only one anon_vma child.
280 		 *
281 		 * Do not chose parent anon_vma, otherwise first child
282 		 * will always reuse it. Root anon_vma is never reused:
283 		 * it has self-parent reference and at least one child.
284 		 */
285 		if (!dst->anon_vma && anon_vma != src->anon_vma &&
286 				anon_vma->degree < 2)
287 			dst->anon_vma = anon_vma;
288 	}
289 	if (dst->anon_vma)
290 		dst->anon_vma->degree++;
291 	unlock_anon_vma_root(root);
292 	return 0;
293 
294  enomem_failure:
295 	/*
296 	 * dst->anon_vma is dropped here otherwise its degree can be incorrectly
297 	 * decremented in unlink_anon_vmas().
298 	 * We can safely do this because callers of anon_vma_clone() don't care
299 	 * about dst->anon_vma if anon_vma_clone() failed.
300 	 */
301 	dst->anon_vma = NULL;
302 	unlink_anon_vmas(dst);
303 	return -ENOMEM;
304 }
305 
306 /*
307  * Attach vma to its own anon_vma, as well as to the anon_vmas that
308  * the corresponding VMA in the parent process is attached to.
309  * Returns 0 on success, non-zero on failure.
310  */
anon_vma_fork(struct vm_area_struct * vma,struct vm_area_struct * pvma)311 int anon_vma_fork(struct vm_area_struct *vma, struct vm_area_struct *pvma)
312 {
313 	struct anon_vma_chain *avc;
314 	struct anon_vma *anon_vma;
315 	int error;
316 
317 	/* Don't bother if the parent process has no anon_vma here. */
318 	if (!pvma->anon_vma)
319 		return 0;
320 
321 	/* Drop inherited anon_vma, we'll reuse existing or allocate new. */
322 	vma->anon_vma = NULL;
323 
324 	/*
325 	 * First, attach the new VMA to the parent VMA's anon_vmas,
326 	 * so rmap can find non-COWed pages in child processes.
327 	 */
328 	error = anon_vma_clone(vma, pvma);
329 	if (error)
330 		return error;
331 
332 	/* An existing anon_vma has been reused, all done then. */
333 	if (vma->anon_vma)
334 		return 0;
335 
336 	/* Then add our own anon_vma. */
337 	anon_vma = anon_vma_alloc();
338 	if (!anon_vma)
339 		goto out_error;
340 	avc = anon_vma_chain_alloc(GFP_KERNEL);
341 	if (!avc)
342 		goto out_error_free_anon_vma;
343 
344 	/*
345 	 * The root anon_vma's spinlock is the lock actually used when we
346 	 * lock any of the anon_vmas in this anon_vma tree.
347 	 */
348 	anon_vma->root = pvma->anon_vma->root;
349 	anon_vma->parent = pvma->anon_vma;
350 	/*
351 	 * With refcounts, an anon_vma can stay around longer than the
352 	 * process it belongs to. The root anon_vma needs to be pinned until
353 	 * this anon_vma is freed, because the lock lives in the root.
354 	 */
355 	get_anon_vma(anon_vma->root);
356 	/* Mark this anon_vma as the one where our new (COWed) pages go. */
357 	vma->anon_vma = anon_vma;
358 	anon_vma_lock_write(anon_vma);
359 	anon_vma_chain_link(vma, avc, anon_vma);
360 	anon_vma->parent->degree++;
361 	anon_vma_unlock_write(anon_vma);
362 
363 	return 0;
364 
365  out_error_free_anon_vma:
366 	put_anon_vma(anon_vma);
367  out_error:
368 	unlink_anon_vmas(vma);
369 	return -ENOMEM;
370 }
371 
unlink_anon_vmas(struct vm_area_struct * vma)372 void unlink_anon_vmas(struct vm_area_struct *vma)
373 {
374 	struct anon_vma_chain *avc, *next;
375 	struct anon_vma *root = NULL;
376 
377 	/*
378 	 * Unlink each anon_vma chained to the VMA.  This list is ordered
379 	 * from newest to oldest, ensuring the root anon_vma gets freed last.
380 	 */
381 	list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
382 		struct anon_vma *anon_vma = avc->anon_vma;
383 
384 		root = lock_anon_vma_root(root, anon_vma);
385 		anon_vma_interval_tree_remove(avc, &anon_vma->rb_root);
386 
387 		/*
388 		 * Leave empty anon_vmas on the list - we'll need
389 		 * to free them outside the lock.
390 		 */
391 		if (RB_EMPTY_ROOT(&anon_vma->rb_root)) {
392 			anon_vma->parent->degree--;
393 			continue;
394 		}
395 
396 		list_del(&avc->same_vma);
397 		anon_vma_chain_free(avc);
398 	}
399 	if (vma->anon_vma)
400 		vma->anon_vma->degree--;
401 	unlock_anon_vma_root(root);
402 
403 	/*
404 	 * Iterate the list once more, it now only contains empty and unlinked
405 	 * anon_vmas, destroy them. Could not do before due to __put_anon_vma()
406 	 * needing to write-acquire the anon_vma->root->rwsem.
407 	 */
408 	list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) {
409 		struct anon_vma *anon_vma = avc->anon_vma;
410 
411 		VM_WARN_ON(anon_vma->degree);
412 		put_anon_vma(anon_vma);
413 
414 		list_del(&avc->same_vma);
415 		anon_vma_chain_free(avc);
416 	}
417 }
418 
anon_vma_ctor(void * data)419 static void anon_vma_ctor(void *data)
420 {
421 	struct anon_vma *anon_vma = data;
422 
423 	init_rwsem(&anon_vma->rwsem);
424 	atomic_set(&anon_vma->refcount, 0);
425 	anon_vma->rb_root = RB_ROOT;
426 }
427 
anon_vma_init(void)428 void __init anon_vma_init(void)
429 {
430 	anon_vma_cachep = kmem_cache_create("anon_vma", sizeof(struct anon_vma),
431 			0, SLAB_DESTROY_BY_RCU|SLAB_PANIC, anon_vma_ctor);
432 	anon_vma_chain_cachep = KMEM_CACHE(anon_vma_chain, SLAB_PANIC);
433 }
434 
435 /*
436  * Getting a lock on a stable anon_vma from a page off the LRU is tricky!
437  *
438  * Since there is no serialization what so ever against page_remove_rmap()
439  * the best this function can do is return a locked anon_vma that might
440  * have been relevant to this page.
441  *
442  * The page might have been remapped to a different anon_vma or the anon_vma
443  * returned may already be freed (and even reused).
444  *
445  * In case it was remapped to a different anon_vma, the new anon_vma will be a
446  * child of the old anon_vma, and the anon_vma lifetime rules will therefore
447  * ensure that any anon_vma obtained from the page will still be valid for as
448  * long as we observe page_mapped() [ hence all those page_mapped() tests ].
449  *
450  * All users of this function must be very careful when walking the anon_vma
451  * chain and verify that the page in question is indeed mapped in it
452  * [ something equivalent to page_mapped_in_vma() ].
453  *
454  * Since anon_vma's slab is DESTROY_BY_RCU and we know from page_remove_rmap()
455  * that the anon_vma pointer from page->mapping is valid if there is a
456  * mapcount, we can dereference the anon_vma after observing those.
457  */
page_get_anon_vma(struct page * page)458 struct anon_vma *page_get_anon_vma(struct page *page)
459 {
460 	struct anon_vma *anon_vma = NULL;
461 	unsigned long anon_mapping;
462 
463 	rcu_read_lock();
464 	anon_mapping = (unsigned long)READ_ONCE(page->mapping);
465 	if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
466 		goto out;
467 	if (!page_mapped(page))
468 		goto out;
469 
470 	anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
471 	if (!atomic_inc_not_zero(&anon_vma->refcount)) {
472 		anon_vma = NULL;
473 		goto out;
474 	}
475 
476 	/*
477 	 * If this page is still mapped, then its anon_vma cannot have been
478 	 * freed.  But if it has been unmapped, we have no security against the
479 	 * anon_vma structure being freed and reused (for another anon_vma:
480 	 * SLAB_DESTROY_BY_RCU guarantees that - so the atomic_inc_not_zero()
481 	 * above cannot corrupt).
482 	 */
483 	if (!page_mapped(page)) {
484 		rcu_read_unlock();
485 		put_anon_vma(anon_vma);
486 		return NULL;
487 	}
488 out:
489 	rcu_read_unlock();
490 
491 	return anon_vma;
492 }
493 
494 /*
495  * Similar to page_get_anon_vma() except it locks the anon_vma.
496  *
497  * Its a little more complex as it tries to keep the fast path to a single
498  * atomic op -- the trylock. If we fail the trylock, we fall back to getting a
499  * reference like with page_get_anon_vma() and then block on the mutex.
500  */
page_lock_anon_vma_read(struct page * page)501 struct anon_vma *page_lock_anon_vma_read(struct page *page)
502 {
503 	struct anon_vma *anon_vma = NULL;
504 	struct anon_vma *root_anon_vma;
505 	unsigned long anon_mapping;
506 
507 	rcu_read_lock();
508 	anon_mapping = (unsigned long)READ_ONCE(page->mapping);
509 	if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)
510 		goto out;
511 	if (!page_mapped(page))
512 		goto out;
513 
514 	anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
515 	root_anon_vma = READ_ONCE(anon_vma->root);
516 	if (down_read_trylock(&root_anon_vma->rwsem)) {
517 		/*
518 		 * If the page is still mapped, then this anon_vma is still
519 		 * its anon_vma, and holding the mutex ensures that it will
520 		 * not go away, see anon_vma_free().
521 		 */
522 		if (!page_mapped(page)) {
523 			up_read(&root_anon_vma->rwsem);
524 			anon_vma = NULL;
525 		}
526 		goto out;
527 	}
528 
529 	/* trylock failed, we got to sleep */
530 	if (!atomic_inc_not_zero(&anon_vma->refcount)) {
531 		anon_vma = NULL;
532 		goto out;
533 	}
534 
535 	if (!page_mapped(page)) {
536 		rcu_read_unlock();
537 		put_anon_vma(anon_vma);
538 		return NULL;
539 	}
540 
541 	/* we pinned the anon_vma, its safe to sleep */
542 	rcu_read_unlock();
543 	anon_vma_lock_read(anon_vma);
544 
545 	if (atomic_dec_and_test(&anon_vma->refcount)) {
546 		/*
547 		 * Oops, we held the last refcount, release the lock
548 		 * and bail -- can't simply use put_anon_vma() because
549 		 * we'll deadlock on the anon_vma_lock_write() recursion.
550 		 */
551 		anon_vma_unlock_read(anon_vma);
552 		__put_anon_vma(anon_vma);
553 		anon_vma = NULL;
554 	}
555 
556 	return anon_vma;
557 
558 out:
559 	rcu_read_unlock();
560 	return anon_vma;
561 }
562 
page_unlock_anon_vma_read(struct anon_vma * anon_vma)563 void page_unlock_anon_vma_read(struct anon_vma *anon_vma)
564 {
565 	anon_vma_unlock_read(anon_vma);
566 }
567 
568 /*
569  * At what user virtual address is page expected in @vma?
570  */
571 static inline unsigned long
__vma_address(struct page * page,struct vm_area_struct * vma)572 __vma_address(struct page *page, struct vm_area_struct *vma)
573 {
574 	pgoff_t pgoff = page_to_pgoff(page);
575 	return vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
576 }
577 
578 inline unsigned long
vma_address(struct page * page,struct vm_area_struct * vma)579 vma_address(struct page *page, struct vm_area_struct *vma)
580 {
581 	unsigned long address = __vma_address(page, vma);
582 
583 	/* page should be within @vma mapping range */
584 	VM_BUG_ON_VMA(address < vma->vm_start || address >= vma->vm_end, vma);
585 
586 	return address;
587 }
588 
589 #ifdef CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH
590 /*
591  * Flush TLB entries for recently unmapped pages from remote CPUs. It is
592  * important if a PTE was dirty when it was unmapped that it's flushed
593  * before any IO is initiated on the page to prevent lost writes. Similarly,
594  * it must be flushed before freeing to prevent data leakage.
595  */
try_to_unmap_flush(void)596 void try_to_unmap_flush(void)
597 {
598 	struct tlbflush_unmap_batch *tlb_ubc = &current->tlb_ubc;
599 	int cpu;
600 
601 	if (!tlb_ubc->flush_required)
602 		return;
603 
604 	cpu = get_cpu();
605 
606 	if (cpumask_test_cpu(cpu, &tlb_ubc->cpumask)) {
607 		count_vm_tlb_event(NR_TLB_LOCAL_FLUSH_ALL);
608 		local_flush_tlb();
609 		trace_tlb_flush(TLB_LOCAL_SHOOTDOWN, TLB_FLUSH_ALL);
610 	}
611 
612 	if (cpumask_any_but(&tlb_ubc->cpumask, cpu) < nr_cpu_ids)
613 		flush_tlb_others(&tlb_ubc->cpumask, NULL, 0, TLB_FLUSH_ALL);
614 	cpumask_clear(&tlb_ubc->cpumask);
615 	tlb_ubc->flush_required = false;
616 	tlb_ubc->writable = false;
617 	put_cpu();
618 }
619 
620 /* Flush iff there are potentially writable TLB entries that can race with IO */
try_to_unmap_flush_dirty(void)621 void try_to_unmap_flush_dirty(void)
622 {
623 	struct tlbflush_unmap_batch *tlb_ubc = &current->tlb_ubc;
624 
625 	if (tlb_ubc->writable)
626 		try_to_unmap_flush();
627 }
628 
set_tlb_ubc_flush_pending(struct mm_struct * mm,struct page * page,bool writable)629 static void set_tlb_ubc_flush_pending(struct mm_struct *mm,
630 		struct page *page, bool writable)
631 {
632 	struct tlbflush_unmap_batch *tlb_ubc = &current->tlb_ubc;
633 
634 	cpumask_or(&tlb_ubc->cpumask, &tlb_ubc->cpumask, mm_cpumask(mm));
635 	tlb_ubc->flush_required = true;
636 
637 	/*
638 	 * Ensure compiler does not re-order the setting of tlb_flush_batched
639 	 * before the PTE is cleared.
640 	 */
641 	barrier();
642 	mm->tlb_flush_batched = true;
643 
644 	/*
645 	 * If the PTE was dirty then it's best to assume it's writable. The
646 	 * caller must use try_to_unmap_flush_dirty() or try_to_unmap_flush()
647 	 * before the page is queued for IO.
648 	 */
649 	if (writable)
650 		tlb_ubc->writable = true;
651 }
652 
653 /*
654  * Returns true if the TLB flush should be deferred to the end of a batch of
655  * unmap operations to reduce IPIs.
656  */
should_defer_flush(struct mm_struct * mm,enum ttu_flags flags)657 static bool should_defer_flush(struct mm_struct *mm, enum ttu_flags flags)
658 {
659 	bool should_defer = false;
660 
661 	if (!(flags & TTU_BATCH_FLUSH))
662 		return false;
663 
664 	/* If remote CPUs need to be flushed then defer batch the flush */
665 	if (cpumask_any_but(mm_cpumask(mm), get_cpu()) < nr_cpu_ids)
666 		should_defer = true;
667 	put_cpu();
668 
669 	return should_defer;
670 }
671 
672 /*
673  * Reclaim unmaps pages under the PTL but do not flush the TLB prior to
674  * releasing the PTL if TLB flushes are batched. It's possible for a parallel
675  * operation such as mprotect or munmap to race between reclaim unmapping
676  * the page and flushing the page. If this race occurs, it potentially allows
677  * access to data via a stale TLB entry. Tracking all mm's that have TLB
678  * batching in flight would be expensive during reclaim so instead track
679  * whether TLB batching occurred in the past and if so then do a flush here
680  * if required. This will cost one additional flush per reclaim cycle paid
681  * by the first operation at risk such as mprotect and mumap.
682  *
683  * This must be called under the PTL so that an access to tlb_flush_batched
684  * that is potentially a "reclaim vs mprotect/munmap/etc" race will synchronise
685  * via the PTL.
686  */
flush_tlb_batched_pending(struct mm_struct * mm)687 void flush_tlb_batched_pending(struct mm_struct *mm)
688 {
689 	if (mm->tlb_flush_batched) {
690 		flush_tlb_mm(mm);
691 
692 		/*
693 		 * Do not allow the compiler to re-order the clearing of
694 		 * tlb_flush_batched before the tlb is flushed.
695 		 */
696 		barrier();
697 		mm->tlb_flush_batched = false;
698 	}
699 }
700 #else
set_tlb_ubc_flush_pending(struct mm_struct * mm,struct page * page,bool writable)701 static void set_tlb_ubc_flush_pending(struct mm_struct *mm,
702 		struct page *page, bool writable)
703 {
704 }
705 
should_defer_flush(struct mm_struct * mm,enum ttu_flags flags)706 static bool should_defer_flush(struct mm_struct *mm, enum ttu_flags flags)
707 {
708 	return false;
709 }
710 #endif /* CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH */
711 
712 /*
713  * At what user virtual address is page expected in vma?
714  * Caller should check the page is actually part of the vma.
715  */
page_address_in_vma(struct page * page,struct vm_area_struct * vma)716 unsigned long page_address_in_vma(struct page *page, struct vm_area_struct *vma)
717 {
718 	unsigned long address;
719 	if (PageAnon(page)) {
720 		struct anon_vma *page__anon_vma = page_anon_vma(page);
721 		/*
722 		 * Note: swapoff's unuse_vma() is more efficient with this
723 		 * check, and needs it to match anon_vma when KSM is active.
724 		 */
725 		if (!vma->anon_vma || !page__anon_vma ||
726 		    vma->anon_vma->root != page__anon_vma->root)
727 			return -EFAULT;
728 	} else if (page->mapping) {
729 		if (!vma->vm_file || vma->vm_file->f_mapping != page->mapping)
730 			return -EFAULT;
731 	} else
732 		return -EFAULT;
733 	address = __vma_address(page, vma);
734 	if (unlikely(address < vma->vm_start || address >= vma->vm_end))
735 		return -EFAULT;
736 	return address;
737 }
738 
mm_find_pmd(struct mm_struct * mm,unsigned long address)739 pmd_t *mm_find_pmd(struct mm_struct *mm, unsigned long address)
740 {
741 	pgd_t *pgd;
742 	pud_t *pud;
743 	pmd_t *pmd = NULL;
744 	pmd_t pmde;
745 
746 	pgd = pgd_offset(mm, address);
747 	if (!pgd_present(*pgd))
748 		goto out;
749 
750 	pud = pud_offset(pgd, address);
751 	if (!pud_present(*pud))
752 		goto out;
753 
754 	pmd = pmd_offset(pud, address);
755 	/*
756 	 * Some THP functions use the sequence pmdp_huge_clear_flush(), set_pmd_at()
757 	 * without holding anon_vma lock for write.  So when looking for a
758 	 * genuine pmde (in which to find pte), test present and !THP together.
759 	 */
760 	pmde = *pmd;
761 	barrier();
762 	if (!pmd_present(pmde) || pmd_trans_huge(pmde))
763 		pmd = NULL;
764 out:
765 	return pmd;
766 }
767 
768 /*
769  * Check that @page is mapped at @address into @mm.
770  *
771  * If @sync is false, page_check_address may perform a racy check to avoid
772  * the page table lock when the pte is not present (helpful when reclaiming
773  * highly shared pages).
774  *
775  * On success returns with pte mapped and locked.
776  */
__page_check_address(struct page * page,struct mm_struct * mm,unsigned long address,spinlock_t ** ptlp,int sync)777 pte_t *__page_check_address(struct page *page, struct mm_struct *mm,
778 			  unsigned long address, spinlock_t **ptlp, int sync)
779 {
780 	pmd_t *pmd;
781 	pte_t *pte;
782 	spinlock_t *ptl;
783 
784 	if (unlikely(PageHuge(page))) {
785 		/* when pud is not present, pte will be NULL */
786 		pte = huge_pte_offset(mm, address);
787 		if (!pte)
788 			return NULL;
789 
790 		ptl = huge_pte_lockptr(page_hstate(page), mm, pte);
791 		goto check;
792 	}
793 
794 	pmd = mm_find_pmd(mm, address);
795 	if (!pmd)
796 		return NULL;
797 
798 	pte = pte_offset_map(pmd, address);
799 	/* Make a quick check before getting the lock */
800 	if (!sync && !pte_present(*pte)) {
801 		pte_unmap(pte);
802 		return NULL;
803 	}
804 
805 	ptl = pte_lockptr(mm, pmd);
806 check:
807 	spin_lock(ptl);
808 	if (pte_present(*pte) && page_to_pfn(page) == pte_pfn(*pte)) {
809 		*ptlp = ptl;
810 		return pte;
811 	}
812 	pte_unmap_unlock(pte, ptl);
813 	return NULL;
814 }
815 
816 /**
817  * page_mapped_in_vma - check whether a page is really mapped in a VMA
818  * @page: the page to test
819  * @vma: the VMA to test
820  *
821  * Returns 1 if the page is mapped into the page tables of the VMA, 0
822  * if the page is not mapped into the page tables of this VMA.  Only
823  * valid for normal file or anonymous VMAs.
824  */
page_mapped_in_vma(struct page * page,struct vm_area_struct * vma)825 int page_mapped_in_vma(struct page *page, struct vm_area_struct *vma)
826 {
827 	unsigned long address;
828 	pte_t *pte;
829 	spinlock_t *ptl;
830 
831 	address = __vma_address(page, vma);
832 	if (unlikely(address < vma->vm_start || address >= vma->vm_end))
833 		return 0;
834 	pte = page_check_address(page, vma->vm_mm, address, &ptl, 1);
835 	if (!pte)			/* the page is not in this mm */
836 		return 0;
837 	pte_unmap_unlock(pte, ptl);
838 
839 	return 1;
840 }
841 
842 struct page_referenced_arg {
843 	int mapcount;
844 	int referenced;
845 	unsigned long vm_flags;
846 	struct mem_cgroup *memcg;
847 };
848 /*
849  * arg: page_referenced_arg will be passed
850  */
page_referenced_one(struct page * page,struct vm_area_struct * vma,unsigned long address,void * arg)851 static int page_referenced_one(struct page *page, struct vm_area_struct *vma,
852 			unsigned long address, void *arg)
853 {
854 	struct mm_struct *mm = vma->vm_mm;
855 	spinlock_t *ptl;
856 	int referenced = 0;
857 	struct page_referenced_arg *pra = arg;
858 
859 	if (unlikely(PageTransHuge(page))) {
860 		pmd_t *pmd;
861 
862 		/*
863 		 * rmap might return false positives; we must filter
864 		 * these out using page_check_address_pmd().
865 		 */
866 		pmd = page_check_address_pmd(page, mm, address,
867 					     PAGE_CHECK_ADDRESS_PMD_FLAG, &ptl);
868 		if (!pmd)
869 			return SWAP_AGAIN;
870 
871 		if (vma->vm_flags & VM_LOCKED) {
872 			spin_unlock(ptl);
873 			pra->vm_flags |= VM_LOCKED;
874 			return SWAP_FAIL; /* To break the loop */
875 		}
876 
877 		/* go ahead even if the pmd is pmd_trans_splitting() */
878 		if (pmdp_clear_flush_young_notify(vma, address, pmd))
879 			referenced++;
880 		spin_unlock(ptl);
881 	} else {
882 		pte_t *pte;
883 
884 		/*
885 		 * rmap might return false positives; we must filter
886 		 * these out using page_check_address().
887 		 */
888 		pte = page_check_address(page, mm, address, &ptl, 0);
889 		if (!pte)
890 			return SWAP_AGAIN;
891 
892 		if (vma->vm_flags & VM_LOCKED) {
893 			pte_unmap_unlock(pte, ptl);
894 			pra->vm_flags |= VM_LOCKED;
895 			return SWAP_FAIL; /* To break the loop */
896 		}
897 
898 		if (ptep_clear_flush_young_notify(vma, address, pte)) {
899 			/*
900 			 * Don't treat a reference through a sequentially read
901 			 * mapping as such.  If the page has been used in
902 			 * another mapping, we will catch it; if this other
903 			 * mapping is already gone, the unmap path will have
904 			 * set PG_referenced or activated the page.
905 			 */
906 			if (likely(!(vma->vm_flags & VM_SEQ_READ)))
907 				referenced++;
908 		}
909 		pte_unmap_unlock(pte, ptl);
910 	}
911 
912 	if (referenced)
913 		clear_page_idle(page);
914 	if (test_and_clear_page_young(page))
915 		referenced++;
916 
917 	if (referenced) {
918 		pra->referenced++;
919 		pra->vm_flags |= vma->vm_flags;
920 	}
921 
922 	pra->mapcount--;
923 	if (!pra->mapcount)
924 		return SWAP_SUCCESS; /* To break the loop */
925 
926 	return SWAP_AGAIN;
927 }
928 
invalid_page_referenced_vma(struct vm_area_struct * vma,void * arg)929 static bool invalid_page_referenced_vma(struct vm_area_struct *vma, void *arg)
930 {
931 	struct page_referenced_arg *pra = arg;
932 	struct mem_cgroup *memcg = pra->memcg;
933 
934 	if (!mm_match_cgroup(vma->vm_mm, memcg))
935 		return true;
936 
937 	return false;
938 }
939 
940 /**
941  * page_referenced - test if the page was referenced
942  * @page: the page to test
943  * @is_locked: caller holds lock on the page
944  * @memcg: target memory cgroup
945  * @vm_flags: collect encountered vma->vm_flags who actually referenced the page
946  *
947  * Quick test_and_clear_referenced for all mappings to a page,
948  * returns the number of ptes which referenced the page.
949  */
page_referenced(struct page * page,int is_locked,struct mem_cgroup * memcg,unsigned long * vm_flags)950 int page_referenced(struct page *page,
951 		    int is_locked,
952 		    struct mem_cgroup *memcg,
953 		    unsigned long *vm_flags)
954 {
955 	int ret;
956 	int we_locked = 0;
957 	struct page_referenced_arg pra = {
958 		.mapcount = page_mapcount(page),
959 		.memcg = memcg,
960 	};
961 	struct rmap_walk_control rwc = {
962 		.rmap_one = page_referenced_one,
963 		.arg = (void *)&pra,
964 		.anon_lock = page_lock_anon_vma_read,
965 	};
966 
967 	*vm_flags = 0;
968 	if (!page_mapped(page))
969 		return 0;
970 
971 	if (!page_rmapping(page))
972 		return 0;
973 
974 	if (!is_locked && (!PageAnon(page) || PageKsm(page))) {
975 		we_locked = trylock_page(page);
976 		if (!we_locked)
977 			return 1;
978 	}
979 
980 	/*
981 	 * If we are reclaiming on behalf of a cgroup, skip
982 	 * counting on behalf of references from different
983 	 * cgroups
984 	 */
985 	if (memcg) {
986 		rwc.invalid_vma = invalid_page_referenced_vma;
987 	}
988 
989 	ret = rmap_walk(page, &rwc);
990 	*vm_flags = pra.vm_flags;
991 
992 	if (we_locked)
993 		unlock_page(page);
994 
995 	return pra.referenced;
996 }
997 
page_mkclean_one(struct page * page,struct vm_area_struct * vma,unsigned long address,void * arg)998 static int page_mkclean_one(struct page *page, struct vm_area_struct *vma,
999 			    unsigned long address, void *arg)
1000 {
1001 	struct mm_struct *mm = vma->vm_mm;
1002 	pte_t *pte;
1003 	spinlock_t *ptl;
1004 	int ret = 0;
1005 	int *cleaned = arg;
1006 
1007 	pte = page_check_address(page, mm, address, &ptl, 1);
1008 	if (!pte)
1009 		goto out;
1010 
1011 	if (pte_dirty(*pte) || pte_write(*pte)) {
1012 		pte_t entry;
1013 
1014 		flush_cache_page(vma, address, pte_pfn(*pte));
1015 		entry = ptep_clear_flush(vma, address, pte);
1016 		entry = pte_wrprotect(entry);
1017 		entry = pte_mkclean(entry);
1018 		set_pte_at(mm, address, pte, entry);
1019 		ret = 1;
1020 	}
1021 
1022 	pte_unmap_unlock(pte, ptl);
1023 
1024 	if (ret) {
1025 		mmu_notifier_invalidate_page(mm, address);
1026 		(*cleaned)++;
1027 	}
1028 out:
1029 	return SWAP_AGAIN;
1030 }
1031 
invalid_mkclean_vma(struct vm_area_struct * vma,void * arg)1032 static bool invalid_mkclean_vma(struct vm_area_struct *vma, void *arg)
1033 {
1034 	if (vma->vm_flags & VM_SHARED)
1035 		return false;
1036 
1037 	return true;
1038 }
1039 
page_mkclean(struct page * page)1040 int page_mkclean(struct page *page)
1041 {
1042 	int cleaned = 0;
1043 	struct address_space *mapping;
1044 	struct rmap_walk_control rwc = {
1045 		.arg = (void *)&cleaned,
1046 		.rmap_one = page_mkclean_one,
1047 		.invalid_vma = invalid_mkclean_vma,
1048 	};
1049 
1050 	BUG_ON(!PageLocked(page));
1051 
1052 	if (!page_mapped(page))
1053 		return 0;
1054 
1055 	mapping = page_mapping(page);
1056 	if (!mapping)
1057 		return 0;
1058 
1059 	rmap_walk(page, &rwc);
1060 
1061 	return cleaned;
1062 }
1063 EXPORT_SYMBOL_GPL(page_mkclean);
1064 
1065 /**
1066  * page_move_anon_rmap - move a page to our anon_vma
1067  * @page:	the page to move to our anon_vma
1068  * @vma:	the vma the page belongs to
1069  * @address:	the user virtual address mapped
1070  *
1071  * When a page belongs exclusively to one process after a COW event,
1072  * that page can be moved into the anon_vma that belongs to just that
1073  * process, so the rmap code will not search the parent or sibling
1074  * processes.
1075  */
page_move_anon_rmap(struct page * page,struct vm_area_struct * vma,unsigned long address)1076 void page_move_anon_rmap(struct page *page,
1077 	struct vm_area_struct *vma, unsigned long address)
1078 {
1079 	struct anon_vma *anon_vma = vma->anon_vma;
1080 
1081 	VM_BUG_ON_PAGE(!PageLocked(page), page);
1082 	VM_BUG_ON_VMA(!anon_vma, vma);
1083 	VM_BUG_ON_PAGE(page->index != linear_page_index(vma, address), page);
1084 
1085 	anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
1086 	/*
1087 	 * Ensure that anon_vma and the PAGE_MAPPING_ANON bit are written
1088 	 * simultaneously, so a concurrent reader (eg page_referenced()'s
1089 	 * PageAnon()) will not see one without the other.
1090 	 */
1091 	WRITE_ONCE(page->mapping, (struct address_space *) anon_vma);
1092 }
1093 
1094 /**
1095  * __page_set_anon_rmap - set up new anonymous rmap
1096  * @page:	Page to add to rmap
1097  * @vma:	VM area to add page to.
1098  * @address:	User virtual address of the mapping
1099  * @exclusive:	the page is exclusively owned by the current process
1100  */
__page_set_anon_rmap(struct page * page,struct vm_area_struct * vma,unsigned long address,int exclusive)1101 static void __page_set_anon_rmap(struct page *page,
1102 	struct vm_area_struct *vma, unsigned long address, int exclusive)
1103 {
1104 	struct anon_vma *anon_vma = vma->anon_vma;
1105 
1106 	BUG_ON(!anon_vma);
1107 
1108 	if (PageAnon(page))
1109 		return;
1110 
1111 	/*
1112 	 * If the page isn't exclusively mapped into this vma,
1113 	 * we must use the _oldest_ possible anon_vma for the
1114 	 * page mapping!
1115 	 */
1116 	if (!exclusive)
1117 		anon_vma = anon_vma->root;
1118 
1119 	anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
1120 	page->mapping = (struct address_space *) anon_vma;
1121 	page->index = linear_page_index(vma, address);
1122 }
1123 
1124 /**
1125  * __page_check_anon_rmap - sanity check anonymous rmap addition
1126  * @page:	the page to add the mapping to
1127  * @vma:	the vm area in which the mapping is added
1128  * @address:	the user virtual address mapped
1129  */
__page_check_anon_rmap(struct page * page,struct vm_area_struct * vma,unsigned long address)1130 static void __page_check_anon_rmap(struct page *page,
1131 	struct vm_area_struct *vma, unsigned long address)
1132 {
1133 #ifdef CONFIG_DEBUG_VM
1134 	/*
1135 	 * The page's anon-rmap details (mapping and index) are guaranteed to
1136 	 * be set up correctly at this point.
1137 	 *
1138 	 * We have exclusion against page_add_anon_rmap because the caller
1139 	 * always holds the page locked, except if called from page_dup_rmap,
1140 	 * in which case the page is already known to be setup.
1141 	 *
1142 	 * We have exclusion against page_add_new_anon_rmap because those pages
1143 	 * are initially only visible via the pagetables, and the pte is locked
1144 	 * over the call to page_add_new_anon_rmap.
1145 	 */
1146 	BUG_ON(page_anon_vma(page)->root != vma->anon_vma->root);
1147 	BUG_ON(page->index != linear_page_index(vma, address));
1148 #endif
1149 }
1150 
1151 /**
1152  * page_add_anon_rmap - add pte mapping to an anonymous page
1153  * @page:	the page to add the mapping to
1154  * @vma:	the vm area in which the mapping is added
1155  * @address:	the user virtual address mapped
1156  *
1157  * The caller needs to hold the pte lock, and the page must be locked in
1158  * the anon_vma case: to serialize mapping,index checking after setting,
1159  * and to ensure that PageAnon is not being upgraded racily to PageKsm
1160  * (but PageKsm is never downgraded to PageAnon).
1161  */
page_add_anon_rmap(struct page * page,struct vm_area_struct * vma,unsigned long address)1162 void page_add_anon_rmap(struct page *page,
1163 	struct vm_area_struct *vma, unsigned long address)
1164 {
1165 	do_page_add_anon_rmap(page, vma, address, 0);
1166 }
1167 
1168 /*
1169  * Special version of the above for do_swap_page, which often runs
1170  * into pages that are exclusively owned by the current process.
1171  * Everybody else should continue to use page_add_anon_rmap above.
1172  */
do_page_add_anon_rmap(struct page * page,struct vm_area_struct * vma,unsigned long address,int exclusive)1173 void do_page_add_anon_rmap(struct page *page,
1174 	struct vm_area_struct *vma, unsigned long address, int exclusive)
1175 {
1176 	int first = atomic_inc_and_test(&page->_mapcount);
1177 	if (first) {
1178 		/*
1179 		 * We use the irq-unsafe __{inc|mod}_zone_page_stat because
1180 		 * these counters are not modified in interrupt context, and
1181 		 * pte lock(a spinlock) is held, which implies preemption
1182 		 * disabled.
1183 		 */
1184 		if (PageTransHuge(page))
1185 			__inc_zone_page_state(page,
1186 					      NR_ANON_TRANSPARENT_HUGEPAGES);
1187 		__mod_zone_page_state(page_zone(page), NR_ANON_PAGES,
1188 				hpage_nr_pages(page));
1189 	}
1190 	if (unlikely(PageKsm(page)))
1191 		return;
1192 
1193 	VM_BUG_ON_PAGE(!PageLocked(page), page);
1194 	/* address might be in next vma when migration races vma_adjust */
1195 	if (first)
1196 		__page_set_anon_rmap(page, vma, address, exclusive);
1197 	else
1198 		__page_check_anon_rmap(page, vma, address);
1199 }
1200 
1201 /**
1202  * page_add_new_anon_rmap - add pte mapping to a new anonymous page
1203  * @page:	the page to add the mapping to
1204  * @vma:	the vm area in which the mapping is added
1205  * @address:	the user virtual address mapped
1206  *
1207  * Same as page_add_anon_rmap but must only be called on *new* pages.
1208  * This means the inc-and-test can be bypassed.
1209  * Page does not have to be locked.
1210  */
page_add_new_anon_rmap(struct page * page,struct vm_area_struct * vma,unsigned long address)1211 void page_add_new_anon_rmap(struct page *page,
1212 	struct vm_area_struct *vma, unsigned long address)
1213 {
1214 	VM_BUG_ON_VMA(address < vma->vm_start || address >= vma->vm_end, vma);
1215 	SetPageSwapBacked(page);
1216 	atomic_set(&page->_mapcount, 0); /* increment count (starts at -1) */
1217 	if (PageTransHuge(page))
1218 		__inc_zone_page_state(page, NR_ANON_TRANSPARENT_HUGEPAGES);
1219 	__mod_zone_page_state(page_zone(page), NR_ANON_PAGES,
1220 			hpage_nr_pages(page));
1221 	__page_set_anon_rmap(page, vma, address, 1);
1222 }
1223 
1224 /**
1225  * page_add_file_rmap - add pte mapping to a file page
1226  * @page: the page to add the mapping to
1227  *
1228  * The caller needs to hold the pte lock.
1229  */
page_add_file_rmap(struct page * page)1230 void page_add_file_rmap(struct page *page)
1231 {
1232 	struct mem_cgroup *memcg;
1233 
1234 	memcg = mem_cgroup_begin_page_stat(page);
1235 	if (atomic_inc_and_test(&page->_mapcount)) {
1236 		__inc_zone_page_state(page, NR_FILE_MAPPED);
1237 		mem_cgroup_inc_page_stat(memcg, MEM_CGROUP_STAT_FILE_MAPPED);
1238 	}
1239 	mem_cgroup_end_page_stat(memcg);
1240 }
1241 
page_remove_file_rmap(struct page * page)1242 static void page_remove_file_rmap(struct page *page)
1243 {
1244 	struct mem_cgroup *memcg;
1245 
1246 	memcg = mem_cgroup_begin_page_stat(page);
1247 
1248 	/* page still mapped by someone else? */
1249 	if (!atomic_add_negative(-1, &page->_mapcount))
1250 		goto out;
1251 
1252 	/* Hugepages are not counted in NR_FILE_MAPPED for now. */
1253 	if (unlikely(PageHuge(page)))
1254 		goto out;
1255 
1256 	/*
1257 	 * We use the irq-unsafe __{inc|mod}_zone_page_stat because
1258 	 * these counters are not modified in interrupt context, and
1259 	 * pte lock(a spinlock) is held, which implies preemption disabled.
1260 	 */
1261 	__dec_zone_page_state(page, NR_FILE_MAPPED);
1262 	mem_cgroup_dec_page_stat(memcg, MEM_CGROUP_STAT_FILE_MAPPED);
1263 
1264 	if (unlikely(PageMlocked(page)))
1265 		clear_page_mlock(page);
1266 out:
1267 	mem_cgroup_end_page_stat(memcg);
1268 }
1269 
1270 /**
1271  * page_remove_rmap - take down pte mapping from a page
1272  * @page: page to remove mapping from
1273  *
1274  * The caller needs to hold the pte lock.
1275  */
page_remove_rmap(struct page * page)1276 void page_remove_rmap(struct page *page)
1277 {
1278 	if (!PageAnon(page)) {
1279 		page_remove_file_rmap(page);
1280 		return;
1281 	}
1282 
1283 	/* page still mapped by someone else? */
1284 	if (!atomic_add_negative(-1, &page->_mapcount))
1285 		return;
1286 
1287 	/* Hugepages are not counted in NR_ANON_PAGES for now. */
1288 	if (unlikely(PageHuge(page)))
1289 		return;
1290 
1291 	/*
1292 	 * We use the irq-unsafe __{inc|mod}_zone_page_stat because
1293 	 * these counters are not modified in interrupt context, and
1294 	 * pte lock(a spinlock) is held, which implies preemption disabled.
1295 	 */
1296 	if (PageTransHuge(page))
1297 		__dec_zone_page_state(page, NR_ANON_TRANSPARENT_HUGEPAGES);
1298 
1299 	__mod_zone_page_state(page_zone(page), NR_ANON_PAGES,
1300 			      -hpage_nr_pages(page));
1301 
1302 	if (unlikely(PageMlocked(page)))
1303 		clear_page_mlock(page);
1304 
1305 	/*
1306 	 * It would be tidy to reset the PageAnon mapping here,
1307 	 * but that might overwrite a racing page_add_anon_rmap
1308 	 * which increments mapcount after us but sets mapping
1309 	 * before us: so leave the reset to free_hot_cold_page,
1310 	 * and remember that it's only reliable while mapped.
1311 	 * Leaving it set also helps swapoff to reinstate ptes
1312 	 * faster for those pages still in swapcache.
1313 	 */
1314 }
1315 
1316 /*
1317  * @arg: enum ttu_flags will be passed to this argument
1318  */
try_to_unmap_one(struct page * page,struct vm_area_struct * vma,unsigned long address,void * arg)1319 static int try_to_unmap_one(struct page *page, struct vm_area_struct *vma,
1320 		     unsigned long address, void *arg)
1321 {
1322 	struct mm_struct *mm = vma->vm_mm;
1323 	pte_t *pte;
1324 	pte_t pteval;
1325 	spinlock_t *ptl;
1326 	int ret = SWAP_AGAIN;
1327 	unsigned long sh_address;
1328 	bool pmd_sharing_possible = false;
1329 	unsigned long spmd_start, spmd_end;
1330 	enum ttu_flags flags = (enum ttu_flags)arg;
1331 
1332 	/* munlock has nothing to gain from examining un-locked vmas */
1333 	if ((flags & TTU_MUNLOCK) && !(vma->vm_flags & VM_LOCKED))
1334 		goto out;
1335 
1336 	/*
1337 	 * Only use the range_start/end mmu notifiers if huge pmd sharing
1338 	 * is possible.  In the normal case, mmu_notifier_invalidate_page
1339 	 * is sufficient as we only unmap a page.  However, if we unshare
1340 	 * a pmd, we will unmap a PUD_SIZE range.
1341 	 */
1342 	if (PageHuge(page)) {
1343 		spmd_start = address;
1344 		spmd_end = spmd_start + vma_mmu_pagesize(vma);
1345 
1346 		/*
1347 		 * Check if pmd sharing is possible.  If possible, we could
1348 		 * unmap a PUD_SIZE range.  spmd_start/spmd_end will be
1349 		 * modified if sharing is possible.
1350 		 */
1351 		adjust_range_if_pmd_sharing_possible(vma, &spmd_start,
1352 								&spmd_end);
1353 		if (spmd_end - spmd_start != vma_mmu_pagesize(vma)) {
1354 			sh_address = address;
1355 
1356 			pmd_sharing_possible = true;
1357 			mmu_notifier_invalidate_range_start(vma->vm_mm,
1358 							spmd_start, spmd_end);
1359 		}
1360 	}
1361 
1362 	pte = page_check_address(page, mm, address, &ptl, 0);
1363 	if (!pte)
1364 		goto out;
1365 
1366 	/*
1367 	 * If the page is mlock()d, we cannot swap it out.
1368 	 * If it's recently referenced (perhaps page_referenced
1369 	 * skipped over this mm) then we should reactivate it.
1370 	 */
1371 	if (!(flags & TTU_IGNORE_MLOCK)) {
1372 		if (vma->vm_flags & VM_LOCKED) {
1373 			/* Holding pte lock, we do *not* need mmap_sem here */
1374 			mlock_vma_page(page);
1375 			ret = SWAP_MLOCK;
1376 			goto out_unmap;
1377 		}
1378 		if (flags & TTU_MUNLOCK)
1379 			goto out_unmap;
1380 	}
1381 	if (!(flags & TTU_IGNORE_ACCESS)) {
1382 		if (ptep_clear_flush_young_notify(vma, address, pte)) {
1383 			ret = SWAP_FAIL;
1384 			goto out_unmap;
1385 		}
1386   	}
1387 
1388 	/*
1389 	 * Call huge_pmd_unshare to potentially unshare a huge pmd.  Pass
1390 	 * sh_address as it will be modified if unsharing is successful.
1391 	 */
1392 	if (PageHuge(page) && huge_pmd_unshare(mm, &sh_address, pte)) {
1393 		/*
1394 		 * huge_pmd_unshare unmapped an entire PMD page.  There is
1395 		 * no way of knowing exactly which PMDs may be cached for
1396 		 * this mm, so flush them all.  spmd_start/spmd_end cover
1397 		 * this PUD_SIZE range.
1398 		 */
1399 		flush_cache_range(vma, spmd_start, spmd_end);
1400 		flush_tlb_range(vma, spmd_start, spmd_end);
1401 
1402 		/*
1403 		 * The ref count of the PMD page was dropped which is part
1404 		 * of the way map counting is done for shared PMDs.  When
1405 		 * there is no other sharing, huge_pmd_unshare returns false
1406 		 * and we will unmap the actual page and drop map count
1407 		 * to zero.
1408 		 */
1409 		goto out_unmap;
1410 	}
1411 
1412 	/* Nuke the page table entry. */
1413 	flush_cache_page(vma, address, page_to_pfn(page));
1414 	if (should_defer_flush(mm, flags)) {
1415 		/*
1416 		 * We clear the PTE but do not flush so potentially a remote
1417 		 * CPU could still be writing to the page. If the entry was
1418 		 * previously clean then the architecture must guarantee that
1419 		 * a clear->dirty transition on a cached TLB entry is written
1420 		 * through and traps if the PTE is unmapped.
1421 		 */
1422 		pteval = ptep_get_and_clear(mm, address, pte);
1423 
1424 		set_tlb_ubc_flush_pending(mm, page, pte_dirty(pteval));
1425 	} else {
1426 		pteval = ptep_clear_flush(vma, address, pte);
1427 	}
1428 
1429 	/* Move the dirty bit to the physical page now the pte is gone. */
1430 	if (pte_dirty(pteval))
1431 		set_page_dirty(page);
1432 
1433 	/* Update high watermark before we lower rss */
1434 	update_hiwater_rss(mm);
1435 
1436 	if (PageHWPoison(page) && !(flags & TTU_IGNORE_HWPOISON)) {
1437 		if (PageHuge(page)) {
1438 			hugetlb_count_sub(1 << compound_order(page), mm);
1439 		} else {
1440 			if (PageAnon(page))
1441 				dec_mm_counter(mm, MM_ANONPAGES);
1442 			else
1443 				dec_mm_counter(mm, MM_FILEPAGES);
1444 		}
1445 		set_pte_at(mm, address, pte,
1446 			   swp_entry_to_pte(make_hwpoison_entry(page)));
1447 	} else if (pte_unused(pteval)) {
1448 		/*
1449 		 * The guest indicated that the page content is of no
1450 		 * interest anymore. Simply discard the pte, vmscan
1451 		 * will take care of the rest.
1452 		 */
1453 		if (PageAnon(page))
1454 			dec_mm_counter(mm, MM_ANONPAGES);
1455 		else
1456 			dec_mm_counter(mm, MM_FILEPAGES);
1457 	} else if (IS_ENABLED(CONFIG_MIGRATION) && (flags & TTU_MIGRATION)) {
1458 		swp_entry_t entry;
1459 		pte_t swp_pte;
1460 		/*
1461 		 * Store the pfn of the page in a special migration
1462 		 * pte. do_swap_page() will wait until the migration
1463 		 * pte is removed and then restart fault handling.
1464 		 */
1465 		entry = make_migration_entry(page, pte_write(pteval));
1466 		swp_pte = swp_entry_to_pte(entry);
1467 		if (pte_soft_dirty(pteval))
1468 			swp_pte = pte_swp_mksoft_dirty(swp_pte);
1469 		set_pte_at(mm, address, pte, swp_pte);
1470 	} else if (PageAnon(page)) {
1471 		swp_entry_t entry = { .val = page_private(page) };
1472 		pte_t swp_pte;
1473 		/*
1474 		 * Store the swap location in the pte.
1475 		 * See handle_pte_fault() ...
1476 		 */
1477 		VM_BUG_ON_PAGE(!PageSwapCache(page), page);
1478 		if (swap_duplicate(entry) < 0) {
1479 			set_pte_at(mm, address, pte, pteval);
1480 			ret = SWAP_FAIL;
1481 			goto out_unmap;
1482 		}
1483 		if (list_empty(&mm->mmlist)) {
1484 			spin_lock(&mmlist_lock);
1485 			if (list_empty(&mm->mmlist))
1486 				list_add(&mm->mmlist, &init_mm.mmlist);
1487 			spin_unlock(&mmlist_lock);
1488 		}
1489 		dec_mm_counter(mm, MM_ANONPAGES);
1490 		inc_mm_counter(mm, MM_SWAPENTS);
1491 		swp_pte = swp_entry_to_pte(entry);
1492 		if (pte_soft_dirty(pteval))
1493 			swp_pte = pte_swp_mksoft_dirty(swp_pte);
1494 		set_pte_at(mm, address, pte, swp_pte);
1495 	} else
1496 		dec_mm_counter(mm, MM_FILEPAGES);
1497 
1498 	page_remove_rmap(page);
1499 	page_cache_release(page);
1500 
1501 out_unmap:
1502 	pte_unmap_unlock(pte, ptl);
1503 	if (ret != SWAP_FAIL && ret != SWAP_MLOCK && !(flags & TTU_MUNLOCK))
1504 		mmu_notifier_invalidate_page(mm, address);
1505 out:
1506 	if (pmd_sharing_possible)
1507 		mmu_notifier_invalidate_range_end(vma->vm_mm,
1508 							spmd_start, spmd_end);
1509 	return ret;
1510 }
1511 
is_vma_temporary_stack(struct vm_area_struct * vma)1512 bool is_vma_temporary_stack(struct vm_area_struct *vma)
1513 {
1514 	int maybe_stack = vma->vm_flags & (VM_GROWSDOWN | VM_GROWSUP);
1515 
1516 	if (!maybe_stack)
1517 		return false;
1518 
1519 	if ((vma->vm_flags & VM_STACK_INCOMPLETE_SETUP) ==
1520 						VM_STACK_INCOMPLETE_SETUP)
1521 		return true;
1522 
1523 	return false;
1524 }
1525 
invalid_migration_vma(struct vm_area_struct * vma,void * arg)1526 static bool invalid_migration_vma(struct vm_area_struct *vma, void *arg)
1527 {
1528 	return is_vma_temporary_stack(vma);
1529 }
1530 
page_not_mapped(struct page * page)1531 static int page_not_mapped(struct page *page)
1532 {
1533 	return !page_mapped(page);
1534 };
1535 
1536 /**
1537  * try_to_unmap - try to remove all page table mappings to a page
1538  * @page: the page to get unmapped
1539  * @flags: action and flags
1540  *
1541  * Tries to remove all the page table entries which are mapping this
1542  * page, used in the pageout path.  Caller must hold the page lock.
1543  * Return values are:
1544  *
1545  * SWAP_SUCCESS	- we succeeded in removing all mappings
1546  * SWAP_AGAIN	- we missed a mapping, try again later
1547  * SWAP_FAIL	- the page is unswappable
1548  * SWAP_MLOCK	- page is mlocked.
1549  */
try_to_unmap(struct page * page,enum ttu_flags flags)1550 int try_to_unmap(struct page *page, enum ttu_flags flags)
1551 {
1552 	int ret;
1553 	struct rmap_walk_control rwc = {
1554 		.rmap_one = try_to_unmap_one,
1555 		.arg = (void *)flags,
1556 		.done = page_not_mapped,
1557 		.anon_lock = page_lock_anon_vma_read,
1558 	};
1559 
1560 	VM_BUG_ON_PAGE(!PageHuge(page) && PageTransHuge(page), page);
1561 
1562 	/*
1563 	 * During exec, a temporary VMA is setup and later moved.
1564 	 * The VMA is moved under the anon_vma lock but not the
1565 	 * page tables leading to a race where migration cannot
1566 	 * find the migration ptes. Rather than increasing the
1567 	 * locking requirements of exec(), migration skips
1568 	 * temporary VMAs until after exec() completes.
1569 	 */
1570 	if ((flags & TTU_MIGRATION) && !PageKsm(page) && PageAnon(page))
1571 		rwc.invalid_vma = invalid_migration_vma;
1572 
1573 	ret = rmap_walk(page, &rwc);
1574 
1575 	if (ret != SWAP_MLOCK && !page_mapped(page))
1576 		ret = SWAP_SUCCESS;
1577 	return ret;
1578 }
1579 
1580 /**
1581  * try_to_munlock - try to munlock a page
1582  * @page: the page to be munlocked
1583  *
1584  * Called from munlock code.  Checks all of the VMAs mapping the page
1585  * to make sure nobody else has this page mlocked. The page will be
1586  * returned with PG_mlocked cleared if no other vmas have it mlocked.
1587  *
1588  * Return values are:
1589  *
1590  * SWAP_AGAIN	- no vma is holding page mlocked, or,
1591  * SWAP_AGAIN	- page mapped in mlocked vma -- couldn't acquire mmap sem
1592  * SWAP_FAIL	- page cannot be located at present
1593  * SWAP_MLOCK	- page is now mlocked.
1594  */
try_to_munlock(struct page * page)1595 int try_to_munlock(struct page *page)
1596 {
1597 	int ret;
1598 	struct rmap_walk_control rwc = {
1599 		.rmap_one = try_to_unmap_one,
1600 		.arg = (void *)TTU_MUNLOCK,
1601 		.done = page_not_mapped,
1602 		.anon_lock = page_lock_anon_vma_read,
1603 
1604 	};
1605 
1606 	VM_BUG_ON_PAGE(!PageLocked(page) || PageLRU(page), page);
1607 
1608 	ret = rmap_walk(page, &rwc);
1609 	return ret;
1610 }
1611 
__put_anon_vma(struct anon_vma * anon_vma)1612 void __put_anon_vma(struct anon_vma *anon_vma)
1613 {
1614 	struct anon_vma *root = anon_vma->root;
1615 
1616 	anon_vma_free(anon_vma);
1617 	if (root != anon_vma && atomic_dec_and_test(&root->refcount))
1618 		anon_vma_free(root);
1619 }
1620 
rmap_walk_anon_lock(struct page * page,struct rmap_walk_control * rwc)1621 static struct anon_vma *rmap_walk_anon_lock(struct page *page,
1622 					struct rmap_walk_control *rwc)
1623 {
1624 	struct anon_vma *anon_vma;
1625 
1626 	if (rwc->anon_lock)
1627 		return rwc->anon_lock(page);
1628 
1629 	/*
1630 	 * Note: remove_migration_ptes() cannot use page_lock_anon_vma_read()
1631 	 * because that depends on page_mapped(); but not all its usages
1632 	 * are holding mmap_sem. Users without mmap_sem are required to
1633 	 * take a reference count to prevent the anon_vma disappearing
1634 	 */
1635 	anon_vma = page_anon_vma(page);
1636 	if (!anon_vma)
1637 		return NULL;
1638 
1639 	anon_vma_lock_read(anon_vma);
1640 	return anon_vma;
1641 }
1642 
1643 /*
1644  * rmap_walk_anon - do something to anonymous page using the object-based
1645  * rmap method
1646  * @page: the page to be handled
1647  * @rwc: control variable according to each walk type
1648  *
1649  * Find all the mappings of a page using the mapping pointer and the vma chains
1650  * contained in the anon_vma struct it points to.
1651  *
1652  * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1653  * where the page was found will be held for write.  So, we won't recheck
1654  * vm_flags for that VMA.  That should be OK, because that vma shouldn't be
1655  * LOCKED.
1656  */
rmap_walk_anon(struct page * page,struct rmap_walk_control * rwc)1657 static int rmap_walk_anon(struct page *page, struct rmap_walk_control *rwc)
1658 {
1659 	struct anon_vma *anon_vma;
1660 	pgoff_t pgoff;
1661 	struct anon_vma_chain *avc;
1662 	int ret = SWAP_AGAIN;
1663 
1664 	anon_vma = rmap_walk_anon_lock(page, rwc);
1665 	if (!anon_vma)
1666 		return ret;
1667 
1668 	pgoff = page_to_pgoff(page);
1669 	anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1670 		struct vm_area_struct *vma = avc->vma;
1671 		unsigned long address = vma_address(page, vma);
1672 
1673 		cond_resched();
1674 
1675 		if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
1676 			continue;
1677 
1678 		ret = rwc->rmap_one(page, vma, address, rwc->arg);
1679 		if (ret != SWAP_AGAIN)
1680 			break;
1681 		if (rwc->done && rwc->done(page))
1682 			break;
1683 	}
1684 	anon_vma_unlock_read(anon_vma);
1685 	return ret;
1686 }
1687 
1688 /*
1689  * rmap_walk_file - do something to file page using the object-based rmap method
1690  * @page: the page to be handled
1691  * @rwc: control variable according to each walk type
1692  *
1693  * Find all the mappings of a page using the mapping pointer and the vma chains
1694  * contained in the address_space struct it points to.
1695  *
1696  * When called from try_to_munlock(), the mmap_sem of the mm containing the vma
1697  * where the page was found will be held for write.  So, we won't recheck
1698  * vm_flags for that VMA.  That should be OK, because that vma shouldn't be
1699  * LOCKED.
1700  */
rmap_walk_file(struct page * page,struct rmap_walk_control * rwc)1701 static int rmap_walk_file(struct page *page, struct rmap_walk_control *rwc)
1702 {
1703 	struct address_space *mapping = page->mapping;
1704 	pgoff_t pgoff;
1705 	struct vm_area_struct *vma;
1706 	int ret = SWAP_AGAIN;
1707 
1708 	/*
1709 	 * The page lock not only makes sure that page->mapping cannot
1710 	 * suddenly be NULLified by truncation, it makes sure that the
1711 	 * structure at mapping cannot be freed and reused yet,
1712 	 * so we can safely take mapping->i_mmap_rwsem.
1713 	 */
1714 	VM_BUG_ON_PAGE(!PageLocked(page), page);
1715 
1716 	if (!mapping)
1717 		return ret;
1718 
1719 	pgoff = page_to_pgoff(page);
1720 	i_mmap_lock_read(mapping);
1721 	vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) {
1722 		unsigned long address = vma_address(page, vma);
1723 
1724 		cond_resched();
1725 
1726 		if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg))
1727 			continue;
1728 
1729 		ret = rwc->rmap_one(page, vma, address, rwc->arg);
1730 		if (ret != SWAP_AGAIN)
1731 			goto done;
1732 		if (rwc->done && rwc->done(page))
1733 			goto done;
1734 	}
1735 
1736 done:
1737 	i_mmap_unlock_read(mapping);
1738 	return ret;
1739 }
1740 
rmap_walk(struct page * page,struct rmap_walk_control * rwc)1741 int rmap_walk(struct page *page, struct rmap_walk_control *rwc)
1742 {
1743 	if (unlikely(PageKsm(page)))
1744 		return rmap_walk_ksm(page, rwc);
1745 	else if (PageAnon(page))
1746 		return rmap_walk_anon(page, rwc);
1747 	else
1748 		return rmap_walk_file(page, rwc);
1749 }
1750 
1751 #ifdef CONFIG_HUGETLB_PAGE
1752 /*
1753  * The following three functions are for anonymous (private mapped) hugepages.
1754  * Unlike common anonymous pages, anonymous hugepages have no accounting code
1755  * and no lru code, because we handle hugepages differently from common pages.
1756  */
__hugepage_set_anon_rmap(struct page * page,struct vm_area_struct * vma,unsigned long address,int exclusive)1757 static void __hugepage_set_anon_rmap(struct page *page,
1758 	struct vm_area_struct *vma, unsigned long address, int exclusive)
1759 {
1760 	struct anon_vma *anon_vma = vma->anon_vma;
1761 
1762 	BUG_ON(!anon_vma);
1763 
1764 	if (PageAnon(page))
1765 		return;
1766 	if (!exclusive)
1767 		anon_vma = anon_vma->root;
1768 
1769 	anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON;
1770 	page->mapping = (struct address_space *) anon_vma;
1771 	page->index = linear_page_index(vma, address);
1772 }
1773 
hugepage_add_anon_rmap(struct page * page,struct vm_area_struct * vma,unsigned long address)1774 void hugepage_add_anon_rmap(struct page *page,
1775 			    struct vm_area_struct *vma, unsigned long address)
1776 {
1777 	struct anon_vma *anon_vma = vma->anon_vma;
1778 	int first;
1779 
1780 	BUG_ON(!PageLocked(page));
1781 	BUG_ON(!anon_vma);
1782 	/* address might be in next vma when migration races vma_adjust */
1783 	first = atomic_inc_and_test(&page->_mapcount);
1784 	if (first)
1785 		__hugepage_set_anon_rmap(page, vma, address, 0);
1786 }
1787 
hugepage_add_new_anon_rmap(struct page * page,struct vm_area_struct * vma,unsigned long address)1788 void hugepage_add_new_anon_rmap(struct page *page,
1789 			struct vm_area_struct *vma, unsigned long address)
1790 {
1791 	BUG_ON(address < vma->vm_start || address >= vma->vm_end);
1792 	atomic_set(&page->_mapcount, 0);
1793 	__hugepage_set_anon_rmap(page, vma, address, 1);
1794 }
1795 #endif /* CONFIG_HUGETLB_PAGE */
1796