1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  *  linux/mm/mmu_notifier.c
4  *
5  *  Copyright (C) 2008  Qumranet, Inc.
6  *  Copyright (C) 2008  SGI
7  *             Christoph Lameter <cl@linux.com>
8  */
9 
10 #include <linux/rculist.h>
11 #include <linux/mmu_notifier.h>
12 #include <linux/export.h>
13 #include <linux/mm.h>
14 #include <linux/err.h>
15 #include <linux/interval_tree.h>
16 #include <linux/srcu.h>
17 #include <linux/rcupdate.h>
18 #include <linux/sched.h>
19 #include <linux/sched/mm.h>
20 #include <linux/slab.h>
21 
22 /* global SRCU for all MMs */
23 DEFINE_STATIC_SRCU(srcu);
24 
25 #ifdef CONFIG_LOCKDEP
26 struct lockdep_map __mmu_notifier_invalidate_range_start_map = {
27 	.name = "mmu_notifier_invalidate_range_start"
28 };
29 #endif
30 
31 /*
32  * The mmu_notifier_subscriptions structure is allocated and installed in
33  * mm->notifier_subscriptions inside the mm_take_all_locks() protected
34  * critical section and it's released only when mm_count reaches zero
35  * in mmdrop().
36  */
37 struct mmu_notifier_subscriptions {
38 	/* all mmu notifiers registered in this mm are queued in this list */
39 	struct hlist_head list;
40 	bool has_itree;
41 	/* to serialize the list modifications and hlist_unhashed */
42 	spinlock_t lock;
43 	unsigned long invalidate_seq;
44 	unsigned long active_invalidate_ranges;
45 	struct rb_root_cached itree;
46 	wait_queue_head_t wq;
47 	struct hlist_head deferred_list;
48 };
49 
50 /*
51  * This is a collision-retry read-side/write-side 'lock', a lot like a
52  * seqcount, however this allows multiple write-sides to hold it at
53  * once. Conceptually the write side is protecting the values of the PTEs in
54  * this mm, such that PTES cannot be read into SPTEs (shadow PTEs) while any
55  * writer exists.
56  *
57  * Note that the core mm creates nested invalidate_range_start()/end() regions
58  * within the same thread, and runs invalidate_range_start()/end() in parallel
59  * on multiple CPUs. This is designed to not reduce concurrency or block
60  * progress on the mm side.
61  *
62  * As a secondary function, holding the full write side also serves to prevent
63  * writers for the itree, this is an optimization to avoid extra locking
64  * during invalidate_range_start/end notifiers.
65  *
66  * The write side has two states, fully excluded:
67  *  - mm->active_invalidate_ranges != 0
68  *  - subscriptions->invalidate_seq & 1 == True (odd)
69  *  - some range on the mm_struct is being invalidated
70  *  - the itree is not allowed to change
71  *
72  * And partially excluded:
73  *  - mm->active_invalidate_ranges != 0
74  *  - subscriptions->invalidate_seq & 1 == False (even)
75  *  - some range on the mm_struct is being invalidated
76  *  - the itree is allowed to change
77  *
78  * Operations on notifier_subscriptions->invalidate_seq (under spinlock):
79  *    seq |= 1  # Begin writing
80  *    seq++     # Release the writing state
81  *    seq & 1   # True if a writer exists
82  *
83  * The later state avoids some expensive work on inv_end in the common case of
84  * no mmu_interval_notifier monitoring the VA.
85  */
86 static bool
mn_itree_is_invalidating(struct mmu_notifier_subscriptions * subscriptions)87 mn_itree_is_invalidating(struct mmu_notifier_subscriptions *subscriptions)
88 {
89 	lockdep_assert_held(&subscriptions->lock);
90 	return subscriptions->invalidate_seq & 1;
91 }
92 
93 static struct mmu_interval_notifier *
mn_itree_inv_start_range(struct mmu_notifier_subscriptions * subscriptions,const struct mmu_notifier_range * range,unsigned long * seq)94 mn_itree_inv_start_range(struct mmu_notifier_subscriptions *subscriptions,
95 			 const struct mmu_notifier_range *range,
96 			 unsigned long *seq)
97 {
98 	struct interval_tree_node *node;
99 	struct mmu_interval_notifier *res = NULL;
100 
101 	spin_lock(&subscriptions->lock);
102 	subscriptions->active_invalidate_ranges++;
103 	node = interval_tree_iter_first(&subscriptions->itree, range->start,
104 					range->end - 1);
105 	if (node) {
106 		subscriptions->invalidate_seq |= 1;
107 		res = container_of(node, struct mmu_interval_notifier,
108 				   interval_tree);
109 	}
110 
111 	*seq = subscriptions->invalidate_seq;
112 	spin_unlock(&subscriptions->lock);
113 	return res;
114 }
115 
116 static struct mmu_interval_notifier *
mn_itree_inv_next(struct mmu_interval_notifier * interval_sub,const struct mmu_notifier_range * range)117 mn_itree_inv_next(struct mmu_interval_notifier *interval_sub,
118 		  const struct mmu_notifier_range *range)
119 {
120 	struct interval_tree_node *node;
121 
122 	node = interval_tree_iter_next(&interval_sub->interval_tree,
123 				       range->start, range->end - 1);
124 	if (!node)
125 		return NULL;
126 	return container_of(node, struct mmu_interval_notifier, interval_tree);
127 }
128 
mn_itree_inv_end(struct mmu_notifier_subscriptions * subscriptions)129 static void mn_itree_inv_end(struct mmu_notifier_subscriptions *subscriptions)
130 {
131 	struct mmu_interval_notifier *interval_sub;
132 	struct hlist_node *next;
133 
134 	spin_lock(&subscriptions->lock);
135 	if (--subscriptions->active_invalidate_ranges ||
136 	    !mn_itree_is_invalidating(subscriptions)) {
137 		spin_unlock(&subscriptions->lock);
138 		return;
139 	}
140 
141 	/* Make invalidate_seq even */
142 	subscriptions->invalidate_seq++;
143 
144 	/*
145 	 * The inv_end incorporates a deferred mechanism like rtnl_unlock().
146 	 * Adds and removes are queued until the final inv_end happens then
147 	 * they are progressed. This arrangement for tree updates is used to
148 	 * avoid using a blocking lock during invalidate_range_start.
149 	 */
150 	hlist_for_each_entry_safe(interval_sub, next,
151 				  &subscriptions->deferred_list,
152 				  deferred_item) {
153 		if (RB_EMPTY_NODE(&interval_sub->interval_tree.rb))
154 			interval_tree_insert(&interval_sub->interval_tree,
155 					     &subscriptions->itree);
156 		else
157 			interval_tree_remove(&interval_sub->interval_tree,
158 					     &subscriptions->itree);
159 		hlist_del(&interval_sub->deferred_item);
160 	}
161 	spin_unlock(&subscriptions->lock);
162 
163 	wake_up_all(&subscriptions->wq);
164 }
165 
166 /**
167  * mmu_interval_read_begin - Begin a read side critical section against a VA
168  *                           range
169  * @interval_sub: The interval subscription
170  *
171  * mmu_iterval_read_begin()/mmu_iterval_read_retry() implement a
172  * collision-retry scheme similar to seqcount for the VA range under
173  * subscription. If the mm invokes invalidation during the critical section
174  * then mmu_interval_read_retry() will return true.
175  *
176  * This is useful to obtain shadow PTEs where teardown or setup of the SPTEs
177  * require a blocking context.  The critical region formed by this can sleep,
178  * and the required 'user_lock' can also be a sleeping lock.
179  *
180  * The caller is required to provide a 'user_lock' to serialize both teardown
181  * and setup.
182  *
183  * The return value should be passed to mmu_interval_read_retry().
184  */
185 unsigned long
mmu_interval_read_begin(struct mmu_interval_notifier * interval_sub)186 mmu_interval_read_begin(struct mmu_interval_notifier *interval_sub)
187 {
188 	struct mmu_notifier_subscriptions *subscriptions =
189 		interval_sub->mm->notifier_subscriptions;
190 	unsigned long seq;
191 	bool is_invalidating;
192 
193 	/*
194 	 * If the subscription has a different seq value under the user_lock
195 	 * than we started with then it has collided.
196 	 *
197 	 * If the subscription currently has the same seq value as the
198 	 * subscriptions seq, then it is currently between
199 	 * invalidate_start/end and is colliding.
200 	 *
201 	 * The locking looks broadly like this:
202 	 *   mn_itree_inv_start():                 mmu_interval_read_begin():
203 	 *                                         spin_lock
204 	 *                                          seq = READ_ONCE(interval_sub->invalidate_seq);
205 	 *                                          seq == subs->invalidate_seq
206 	 *                                         spin_unlock
207 	 *    spin_lock
208 	 *     seq = ++subscriptions->invalidate_seq
209 	 *    spin_unlock
210 	 *     op->invalidate():
211 	 *       user_lock
212 	 *        mmu_interval_set_seq()
213 	 *         interval_sub->invalidate_seq = seq
214 	 *       user_unlock
215 	 *
216 	 *                          [Required: mmu_interval_read_retry() == true]
217 	 *
218 	 *   mn_itree_inv_end():
219 	 *    spin_lock
220 	 *     seq = ++subscriptions->invalidate_seq
221 	 *    spin_unlock
222 	 *
223 	 *                                        user_lock
224 	 *                                         mmu_interval_read_retry():
225 	 *                                          interval_sub->invalidate_seq != seq
226 	 *                                        user_unlock
227 	 *
228 	 * Barriers are not needed here as any races here are closed by an
229 	 * eventual mmu_interval_read_retry(), which provides a barrier via the
230 	 * user_lock.
231 	 */
232 	spin_lock(&subscriptions->lock);
233 	/* Pairs with the WRITE_ONCE in mmu_interval_set_seq() */
234 	seq = READ_ONCE(interval_sub->invalidate_seq);
235 	is_invalidating = seq == subscriptions->invalidate_seq;
236 	spin_unlock(&subscriptions->lock);
237 
238 	/*
239 	 * interval_sub->invalidate_seq must always be set to an odd value via
240 	 * mmu_interval_set_seq() using the provided cur_seq from
241 	 * mn_itree_inv_start_range(). This ensures that if seq does wrap we
242 	 * will always clear the below sleep in some reasonable time as
243 	 * subscriptions->invalidate_seq is even in the idle state.
244 	 */
245 	lock_map_acquire(&__mmu_notifier_invalidate_range_start_map);
246 	lock_map_release(&__mmu_notifier_invalidate_range_start_map);
247 	if (is_invalidating)
248 		wait_event(subscriptions->wq,
249 			   READ_ONCE(subscriptions->invalidate_seq) != seq);
250 
251 	/*
252 	 * Notice that mmu_interval_read_retry() can already be true at this
253 	 * point, avoiding loops here allows the caller to provide a global
254 	 * time bound.
255 	 */
256 
257 	return seq;
258 }
259 EXPORT_SYMBOL_GPL(mmu_interval_read_begin);
260 
mn_itree_release(struct mmu_notifier_subscriptions * subscriptions,struct mm_struct * mm)261 static void mn_itree_release(struct mmu_notifier_subscriptions *subscriptions,
262 			     struct mm_struct *mm)
263 {
264 	struct mmu_notifier_range range = {
265 		.flags = MMU_NOTIFIER_RANGE_BLOCKABLE,
266 		.event = MMU_NOTIFY_RELEASE,
267 		.mm = mm,
268 		.start = 0,
269 		.end = ULONG_MAX,
270 	};
271 	struct mmu_interval_notifier *interval_sub;
272 	unsigned long cur_seq;
273 	bool ret;
274 
275 	for (interval_sub =
276 		     mn_itree_inv_start_range(subscriptions, &range, &cur_seq);
277 	     interval_sub;
278 	     interval_sub = mn_itree_inv_next(interval_sub, &range)) {
279 		ret = interval_sub->ops->invalidate(interval_sub, &range,
280 						    cur_seq);
281 		WARN_ON(!ret);
282 	}
283 
284 	mn_itree_inv_end(subscriptions);
285 }
286 
287 /*
288  * This function can't run concurrently against mmu_notifier_register
289  * because mm->mm_users > 0 during mmu_notifier_register and exit_mmap
290  * runs with mm_users == 0. Other tasks may still invoke mmu notifiers
291  * in parallel despite there being no task using this mm any more,
292  * through the vmas outside of the exit_mmap context, such as with
293  * vmtruncate. This serializes against mmu_notifier_unregister with
294  * the notifier_subscriptions->lock in addition to SRCU and it serializes
295  * against the other mmu notifiers with SRCU. struct mmu_notifier_subscriptions
296  * can't go away from under us as exit_mmap holds an mm_count pin
297  * itself.
298  */
mn_hlist_release(struct mmu_notifier_subscriptions * subscriptions,struct mm_struct * mm)299 static void mn_hlist_release(struct mmu_notifier_subscriptions *subscriptions,
300 			     struct mm_struct *mm)
301 {
302 	struct mmu_notifier *subscription;
303 	int id;
304 
305 	/*
306 	 * SRCU here will block mmu_notifier_unregister until
307 	 * ->release returns.
308 	 */
309 	id = srcu_read_lock(&srcu);
310 	hlist_for_each_entry_rcu(subscription, &subscriptions->list, hlist,
311 				 srcu_read_lock_held(&srcu))
312 		/*
313 		 * If ->release runs before mmu_notifier_unregister it must be
314 		 * handled, as it's the only way for the driver to flush all
315 		 * existing sptes and stop the driver from establishing any more
316 		 * sptes before all the pages in the mm are freed.
317 		 */
318 		if (subscription->ops->release)
319 			subscription->ops->release(subscription, mm);
320 
321 	spin_lock(&subscriptions->lock);
322 	while (unlikely(!hlist_empty(&subscriptions->list))) {
323 		subscription = hlist_entry(subscriptions->list.first,
324 					   struct mmu_notifier, hlist);
325 		/*
326 		 * We arrived before mmu_notifier_unregister so
327 		 * mmu_notifier_unregister will do nothing other than to wait
328 		 * for ->release to finish and for mmu_notifier_unregister to
329 		 * return.
330 		 */
331 		hlist_del_init_rcu(&subscription->hlist);
332 	}
333 	spin_unlock(&subscriptions->lock);
334 	srcu_read_unlock(&srcu, id);
335 
336 	/*
337 	 * synchronize_srcu here prevents mmu_notifier_release from returning to
338 	 * exit_mmap (which would proceed with freeing all pages in the mm)
339 	 * until the ->release method returns, if it was invoked by
340 	 * mmu_notifier_unregister.
341 	 *
342 	 * The notifier_subscriptions can't go away from under us because
343 	 * one mm_count is held by exit_mmap.
344 	 */
345 	synchronize_srcu(&srcu);
346 }
347 
__mmu_notifier_release(struct mm_struct * mm)348 void __mmu_notifier_release(struct mm_struct *mm)
349 {
350 	struct mmu_notifier_subscriptions *subscriptions =
351 		mm->notifier_subscriptions;
352 
353 	if (subscriptions->has_itree)
354 		mn_itree_release(subscriptions, mm);
355 
356 	if (!hlist_empty(&subscriptions->list))
357 		mn_hlist_release(subscriptions, mm);
358 }
359 
360 /*
361  * If no young bitflag is supported by the hardware, ->clear_flush_young can
362  * unmap the address and return 1 or 0 depending if the mapping previously
363  * existed or not.
364  */
__mmu_notifier_clear_flush_young(struct mm_struct * mm,unsigned long start,unsigned long end)365 int __mmu_notifier_clear_flush_young(struct mm_struct *mm,
366 					unsigned long start,
367 					unsigned long end)
368 {
369 	struct mmu_notifier *subscription;
370 	int young = 0, id;
371 
372 	id = srcu_read_lock(&srcu);
373 	hlist_for_each_entry_rcu(subscription,
374 				 &mm->notifier_subscriptions->list, hlist,
375 				 srcu_read_lock_held(&srcu)) {
376 		if (subscription->ops->clear_flush_young)
377 			young |= subscription->ops->clear_flush_young(
378 				subscription, mm, start, end);
379 	}
380 	srcu_read_unlock(&srcu, id);
381 
382 	return young;
383 }
384 
__mmu_notifier_clear_young(struct mm_struct * mm,unsigned long start,unsigned long end)385 int __mmu_notifier_clear_young(struct mm_struct *mm,
386 			       unsigned long start,
387 			       unsigned long end)
388 {
389 	struct mmu_notifier *subscription;
390 	int young = 0, id;
391 
392 	id = srcu_read_lock(&srcu);
393 	hlist_for_each_entry_rcu(subscription,
394 				 &mm->notifier_subscriptions->list, hlist,
395 				 srcu_read_lock_held(&srcu)) {
396 		if (subscription->ops->clear_young)
397 			young |= subscription->ops->clear_young(subscription,
398 								mm, start, end);
399 	}
400 	srcu_read_unlock(&srcu, id);
401 
402 	return young;
403 }
404 
__mmu_notifier_test_young(struct mm_struct * mm,unsigned long address)405 int __mmu_notifier_test_young(struct mm_struct *mm,
406 			      unsigned long address)
407 {
408 	struct mmu_notifier *subscription;
409 	int young = 0, id;
410 
411 	id = srcu_read_lock(&srcu);
412 	hlist_for_each_entry_rcu(subscription,
413 				 &mm->notifier_subscriptions->list, hlist,
414 				 srcu_read_lock_held(&srcu)) {
415 		if (subscription->ops->test_young) {
416 			young = subscription->ops->test_young(subscription, mm,
417 							      address);
418 			if (young)
419 				break;
420 		}
421 	}
422 	srcu_read_unlock(&srcu, id);
423 
424 	return young;
425 }
426 
__mmu_notifier_change_pte(struct mm_struct * mm,unsigned long address,pte_t pte)427 void __mmu_notifier_change_pte(struct mm_struct *mm, unsigned long address,
428 			       pte_t pte)
429 {
430 	struct mmu_notifier *subscription;
431 	int id;
432 
433 	id = srcu_read_lock(&srcu);
434 	hlist_for_each_entry_rcu(subscription,
435 				 &mm->notifier_subscriptions->list, hlist,
436 				 srcu_read_lock_held(&srcu)) {
437 		if (subscription->ops->change_pte)
438 			subscription->ops->change_pte(subscription, mm, address,
439 						      pte);
440 	}
441 	srcu_read_unlock(&srcu, id);
442 }
443 
mn_itree_invalidate(struct mmu_notifier_subscriptions * subscriptions,const struct mmu_notifier_range * range)444 static int mn_itree_invalidate(struct mmu_notifier_subscriptions *subscriptions,
445 			       const struct mmu_notifier_range *range)
446 {
447 	struct mmu_interval_notifier *interval_sub;
448 	unsigned long cur_seq;
449 
450 	for (interval_sub =
451 		     mn_itree_inv_start_range(subscriptions, range, &cur_seq);
452 	     interval_sub;
453 	     interval_sub = mn_itree_inv_next(interval_sub, range)) {
454 		bool ret;
455 
456 		ret = interval_sub->ops->invalidate(interval_sub, range,
457 						    cur_seq);
458 		if (!ret) {
459 			if (WARN_ON(mmu_notifier_range_blockable(range)))
460 				continue;
461 			goto out_would_block;
462 		}
463 	}
464 	return 0;
465 
466 out_would_block:
467 	/*
468 	 * On -EAGAIN the non-blocking caller is not allowed to call
469 	 * invalidate_range_end()
470 	 */
471 	mn_itree_inv_end(subscriptions);
472 	return -EAGAIN;
473 }
474 
mn_hlist_invalidate_range_start(struct mmu_notifier_subscriptions * subscriptions,struct mmu_notifier_range * range)475 static int mn_hlist_invalidate_range_start(
476 	struct mmu_notifier_subscriptions *subscriptions,
477 	struct mmu_notifier_range *range)
478 {
479 	struct mmu_notifier *subscription;
480 	int ret = 0;
481 	int id;
482 
483 	id = srcu_read_lock(&srcu);
484 	hlist_for_each_entry_rcu(subscription, &subscriptions->list, hlist,
485 				 srcu_read_lock_held(&srcu)) {
486 		const struct mmu_notifier_ops *ops = subscription->ops;
487 
488 		if (ops->invalidate_range_start) {
489 			int _ret;
490 
491 			if (!mmu_notifier_range_blockable(range))
492 				non_block_start();
493 			_ret = ops->invalidate_range_start(subscription, range);
494 			if (!mmu_notifier_range_blockable(range))
495 				non_block_end();
496 			if (_ret) {
497 				pr_info("%pS callback failed with %d in %sblockable context.\n",
498 					ops->invalidate_range_start, _ret,
499 					!mmu_notifier_range_blockable(range) ?
500 						"non-" :
501 						"");
502 				WARN_ON(mmu_notifier_range_blockable(range) ||
503 					_ret != -EAGAIN);
504 				/*
505 				 * We call all the notifiers on any EAGAIN,
506 				 * there is no way for a notifier to know if
507 				 * its start method failed, thus a start that
508 				 * does EAGAIN can't also do end.
509 				 */
510 				WARN_ON(ops->invalidate_range_end);
511 				ret = _ret;
512 			}
513 		}
514 	}
515 
516 	if (ret) {
517 		/*
518 		 * Must be non-blocking to get here.  If there are multiple
519 		 * notifiers and one or more failed start, any that succeeded
520 		 * start are expecting their end to be called.  Do so now.
521 		 */
522 		hlist_for_each_entry_rcu(subscription, &subscriptions->list,
523 					 hlist, srcu_read_lock_held(&srcu)) {
524 			if (!subscription->ops->invalidate_range_end)
525 				continue;
526 
527 			subscription->ops->invalidate_range_end(subscription,
528 								range);
529 		}
530 	}
531 	srcu_read_unlock(&srcu, id);
532 
533 	return ret;
534 }
535 
__mmu_notifier_invalidate_range_start(struct mmu_notifier_range * range)536 int __mmu_notifier_invalidate_range_start(struct mmu_notifier_range *range)
537 {
538 	struct mmu_notifier_subscriptions *subscriptions =
539 		range->mm->notifier_subscriptions;
540 	int ret;
541 
542 	if (subscriptions->has_itree) {
543 		ret = mn_itree_invalidate(subscriptions, range);
544 		if (ret)
545 			return ret;
546 	}
547 	if (!hlist_empty(&subscriptions->list))
548 		return mn_hlist_invalidate_range_start(subscriptions, range);
549 	return 0;
550 }
551 
552 static void
mn_hlist_invalidate_end(struct mmu_notifier_subscriptions * subscriptions,struct mmu_notifier_range * range)553 mn_hlist_invalidate_end(struct mmu_notifier_subscriptions *subscriptions,
554 			struct mmu_notifier_range *range)
555 {
556 	struct mmu_notifier *subscription;
557 	int id;
558 
559 	id = srcu_read_lock(&srcu);
560 	hlist_for_each_entry_rcu(subscription, &subscriptions->list, hlist,
561 				 srcu_read_lock_held(&srcu)) {
562 		if (subscription->ops->invalidate_range_end) {
563 			if (!mmu_notifier_range_blockable(range))
564 				non_block_start();
565 			subscription->ops->invalidate_range_end(subscription,
566 								range);
567 			if (!mmu_notifier_range_blockable(range))
568 				non_block_end();
569 		}
570 	}
571 	srcu_read_unlock(&srcu, id);
572 }
573 
__mmu_notifier_invalidate_range_end(struct mmu_notifier_range * range)574 void __mmu_notifier_invalidate_range_end(struct mmu_notifier_range *range)
575 {
576 	struct mmu_notifier_subscriptions *subscriptions =
577 		range->mm->notifier_subscriptions;
578 
579 	lock_map_acquire(&__mmu_notifier_invalidate_range_start_map);
580 	if (subscriptions->has_itree)
581 		mn_itree_inv_end(subscriptions);
582 
583 	if (!hlist_empty(&subscriptions->list))
584 		mn_hlist_invalidate_end(subscriptions, range);
585 	lock_map_release(&__mmu_notifier_invalidate_range_start_map);
586 }
587 
__mmu_notifier_arch_invalidate_secondary_tlbs(struct mm_struct * mm,unsigned long start,unsigned long end)588 void __mmu_notifier_arch_invalidate_secondary_tlbs(struct mm_struct *mm,
589 					unsigned long start, unsigned long end)
590 {
591 	struct mmu_notifier *subscription;
592 	int id;
593 
594 	id = srcu_read_lock(&srcu);
595 	hlist_for_each_entry_rcu(subscription,
596 				 &mm->notifier_subscriptions->list, hlist,
597 				 srcu_read_lock_held(&srcu)) {
598 		if (subscription->ops->arch_invalidate_secondary_tlbs)
599 			subscription->ops->arch_invalidate_secondary_tlbs(
600 				subscription, mm,
601 				start, end);
602 	}
603 	srcu_read_unlock(&srcu, id);
604 }
605 
606 /*
607  * Same as mmu_notifier_register but here the caller must hold the mmap_lock in
608  * write mode. A NULL mn signals the notifier is being registered for itree
609  * mode.
610  */
__mmu_notifier_register(struct mmu_notifier * subscription,struct mm_struct * mm)611 int __mmu_notifier_register(struct mmu_notifier *subscription,
612 			    struct mm_struct *mm)
613 {
614 	struct mmu_notifier_subscriptions *subscriptions = NULL;
615 	int ret;
616 
617 	mmap_assert_write_locked(mm);
618 	BUG_ON(atomic_read(&mm->mm_users) <= 0);
619 
620 	/*
621 	 * Subsystems should only register for invalidate_secondary_tlbs() or
622 	 * invalidate_range_start()/end() callbacks, not both.
623 	 */
624 	if (WARN_ON_ONCE(subscription &&
625 			 (subscription->ops->arch_invalidate_secondary_tlbs &&
626 			 (subscription->ops->invalidate_range_start ||
627 			  subscription->ops->invalidate_range_end))))
628 		return -EINVAL;
629 
630 	if (!mm->notifier_subscriptions) {
631 		/*
632 		 * kmalloc cannot be called under mm_take_all_locks(), but we
633 		 * know that mm->notifier_subscriptions can't change while we
634 		 * hold the write side of the mmap_lock.
635 		 */
636 		subscriptions = kzalloc(
637 			sizeof(struct mmu_notifier_subscriptions), GFP_KERNEL);
638 		if (!subscriptions)
639 			return -ENOMEM;
640 
641 		INIT_HLIST_HEAD(&subscriptions->list);
642 		spin_lock_init(&subscriptions->lock);
643 		subscriptions->invalidate_seq = 2;
644 		subscriptions->itree = RB_ROOT_CACHED;
645 		init_waitqueue_head(&subscriptions->wq);
646 		INIT_HLIST_HEAD(&subscriptions->deferred_list);
647 	}
648 
649 	ret = mm_take_all_locks(mm);
650 	if (unlikely(ret))
651 		goto out_clean;
652 
653 	/*
654 	 * Serialize the update against mmu_notifier_unregister. A
655 	 * side note: mmu_notifier_release can't run concurrently with
656 	 * us because we hold the mm_users pin (either implicitly as
657 	 * current->mm or explicitly with get_task_mm() or similar).
658 	 * We can't race against any other mmu notifier method either
659 	 * thanks to mm_take_all_locks().
660 	 *
661 	 * release semantics on the initialization of the
662 	 * mmu_notifier_subscriptions's contents are provided for unlocked
663 	 * readers.  acquire can only be used while holding the mmgrab or
664 	 * mmget, and is safe because once created the
665 	 * mmu_notifier_subscriptions is not freed until the mm is destroyed.
666 	 * As above, users holding the mmap_lock or one of the
667 	 * mm_take_all_locks() do not need to use acquire semantics.
668 	 */
669 	if (subscriptions)
670 		smp_store_release(&mm->notifier_subscriptions, subscriptions);
671 
672 	if (subscription) {
673 		/* Pairs with the mmdrop in mmu_notifier_unregister_* */
674 		mmgrab(mm);
675 		subscription->mm = mm;
676 		subscription->users = 1;
677 
678 		spin_lock(&mm->notifier_subscriptions->lock);
679 		hlist_add_head_rcu(&subscription->hlist,
680 				   &mm->notifier_subscriptions->list);
681 		spin_unlock(&mm->notifier_subscriptions->lock);
682 	} else
683 		mm->notifier_subscriptions->has_itree = true;
684 
685 	mm_drop_all_locks(mm);
686 	BUG_ON(atomic_read(&mm->mm_users) <= 0);
687 	return 0;
688 
689 out_clean:
690 	kfree(subscriptions);
691 	return ret;
692 }
693 EXPORT_SYMBOL_GPL(__mmu_notifier_register);
694 
695 /**
696  * mmu_notifier_register - Register a notifier on a mm
697  * @subscription: The notifier to attach
698  * @mm: The mm to attach the notifier to
699  *
700  * Must not hold mmap_lock nor any other VM related lock when calling
701  * this registration function. Must also ensure mm_users can't go down
702  * to zero while this runs to avoid races with mmu_notifier_release,
703  * so mm has to be current->mm or the mm should be pinned safely such
704  * as with get_task_mm(). If the mm is not current->mm, the mm_users
705  * pin should be released by calling mmput after mmu_notifier_register
706  * returns.
707  *
708  * mmu_notifier_unregister() or mmu_notifier_put() must be always called to
709  * unregister the notifier.
710  *
711  * While the caller has a mmu_notifier get the subscription->mm pointer will remain
712  * valid, and can be converted to an active mm pointer via mmget_not_zero().
713  */
mmu_notifier_register(struct mmu_notifier * subscription,struct mm_struct * mm)714 int mmu_notifier_register(struct mmu_notifier *subscription,
715 			  struct mm_struct *mm)
716 {
717 	int ret;
718 
719 	mmap_write_lock(mm);
720 	ret = __mmu_notifier_register(subscription, mm);
721 	mmap_write_unlock(mm);
722 	return ret;
723 }
724 EXPORT_SYMBOL_GPL(mmu_notifier_register);
725 
726 static struct mmu_notifier *
find_get_mmu_notifier(struct mm_struct * mm,const struct mmu_notifier_ops * ops)727 find_get_mmu_notifier(struct mm_struct *mm, const struct mmu_notifier_ops *ops)
728 {
729 	struct mmu_notifier *subscription;
730 
731 	spin_lock(&mm->notifier_subscriptions->lock);
732 	hlist_for_each_entry_rcu(subscription,
733 				 &mm->notifier_subscriptions->list, hlist,
734 				 lockdep_is_held(&mm->notifier_subscriptions->lock)) {
735 		if (subscription->ops != ops)
736 			continue;
737 
738 		if (likely(subscription->users != UINT_MAX))
739 			subscription->users++;
740 		else
741 			subscription = ERR_PTR(-EOVERFLOW);
742 		spin_unlock(&mm->notifier_subscriptions->lock);
743 		return subscription;
744 	}
745 	spin_unlock(&mm->notifier_subscriptions->lock);
746 	return NULL;
747 }
748 
749 /**
750  * mmu_notifier_get_locked - Return the single struct mmu_notifier for
751  *                           the mm & ops
752  * @ops: The operations struct being subscribe with
753  * @mm : The mm to attach notifiers too
754  *
755  * This function either allocates a new mmu_notifier via
756  * ops->alloc_notifier(), or returns an already existing notifier on the
757  * list. The value of the ops pointer is used to determine when two notifiers
758  * are the same.
759  *
760  * Each call to mmu_notifier_get() must be paired with a call to
761  * mmu_notifier_put(). The caller must hold the write side of mm->mmap_lock.
762  *
763  * While the caller has a mmu_notifier get the mm pointer will remain valid,
764  * and can be converted to an active mm pointer via mmget_not_zero().
765  */
mmu_notifier_get_locked(const struct mmu_notifier_ops * ops,struct mm_struct * mm)766 struct mmu_notifier *mmu_notifier_get_locked(const struct mmu_notifier_ops *ops,
767 					     struct mm_struct *mm)
768 {
769 	struct mmu_notifier *subscription;
770 	int ret;
771 
772 	mmap_assert_write_locked(mm);
773 
774 	if (mm->notifier_subscriptions) {
775 		subscription = find_get_mmu_notifier(mm, ops);
776 		if (subscription)
777 			return subscription;
778 	}
779 
780 	subscription = ops->alloc_notifier(mm);
781 	if (IS_ERR(subscription))
782 		return subscription;
783 	subscription->ops = ops;
784 	ret = __mmu_notifier_register(subscription, mm);
785 	if (ret)
786 		goto out_free;
787 	return subscription;
788 out_free:
789 	subscription->ops->free_notifier(subscription);
790 	return ERR_PTR(ret);
791 }
792 EXPORT_SYMBOL_GPL(mmu_notifier_get_locked);
793 
794 /* this is called after the last mmu_notifier_unregister() returned */
__mmu_notifier_subscriptions_destroy(struct mm_struct * mm)795 void __mmu_notifier_subscriptions_destroy(struct mm_struct *mm)
796 {
797 	BUG_ON(!hlist_empty(&mm->notifier_subscriptions->list));
798 	kfree(mm->notifier_subscriptions);
799 	mm->notifier_subscriptions = LIST_POISON1; /* debug */
800 }
801 
802 /*
803  * This releases the mm_count pin automatically and frees the mm
804  * structure if it was the last user of it. It serializes against
805  * running mmu notifiers with SRCU and against mmu_notifier_unregister
806  * with the unregister lock + SRCU. All sptes must be dropped before
807  * calling mmu_notifier_unregister. ->release or any other notifier
808  * method may be invoked concurrently with mmu_notifier_unregister,
809  * and only after mmu_notifier_unregister returned we're guaranteed
810  * that ->release or any other method can't run anymore.
811  */
mmu_notifier_unregister(struct mmu_notifier * subscription,struct mm_struct * mm)812 void mmu_notifier_unregister(struct mmu_notifier *subscription,
813 			     struct mm_struct *mm)
814 {
815 	BUG_ON(atomic_read(&mm->mm_count) <= 0);
816 
817 	if (!hlist_unhashed(&subscription->hlist)) {
818 		/*
819 		 * SRCU here will force exit_mmap to wait for ->release to
820 		 * finish before freeing the pages.
821 		 */
822 		int id;
823 
824 		id = srcu_read_lock(&srcu);
825 		/*
826 		 * exit_mmap will block in mmu_notifier_release to guarantee
827 		 * that ->release is called before freeing the pages.
828 		 */
829 		if (subscription->ops->release)
830 			subscription->ops->release(subscription, mm);
831 		srcu_read_unlock(&srcu, id);
832 
833 		spin_lock(&mm->notifier_subscriptions->lock);
834 		/*
835 		 * Can not use list_del_rcu() since __mmu_notifier_release
836 		 * can delete it before we hold the lock.
837 		 */
838 		hlist_del_init_rcu(&subscription->hlist);
839 		spin_unlock(&mm->notifier_subscriptions->lock);
840 	}
841 
842 	/*
843 	 * Wait for any running method to finish, of course including
844 	 * ->release if it was run by mmu_notifier_release instead of us.
845 	 */
846 	synchronize_srcu(&srcu);
847 
848 	BUG_ON(atomic_read(&mm->mm_count) <= 0);
849 
850 	mmdrop(mm);
851 }
852 EXPORT_SYMBOL_GPL(mmu_notifier_unregister);
853 
mmu_notifier_free_rcu(struct rcu_head * rcu)854 static void mmu_notifier_free_rcu(struct rcu_head *rcu)
855 {
856 	struct mmu_notifier *subscription =
857 		container_of(rcu, struct mmu_notifier, rcu);
858 	struct mm_struct *mm = subscription->mm;
859 
860 	subscription->ops->free_notifier(subscription);
861 	/* Pairs with the get in __mmu_notifier_register() */
862 	mmdrop(mm);
863 }
864 
865 /**
866  * mmu_notifier_put - Release the reference on the notifier
867  * @subscription: The notifier to act on
868  *
869  * This function must be paired with each mmu_notifier_get(), it releases the
870  * reference obtained by the get. If this is the last reference then process
871  * to free the notifier will be run asynchronously.
872  *
873  * Unlike mmu_notifier_unregister() the get/put flow only calls ops->release
874  * when the mm_struct is destroyed. Instead free_notifier is always called to
875  * release any resources held by the user.
876  *
877  * As ops->release is not guaranteed to be called, the user must ensure that
878  * all sptes are dropped, and no new sptes can be established before
879  * mmu_notifier_put() is called.
880  *
881  * This function can be called from the ops->release callback, however the
882  * caller must still ensure it is called pairwise with mmu_notifier_get().
883  *
884  * Modules calling this function must call mmu_notifier_synchronize() in
885  * their __exit functions to ensure the async work is completed.
886  */
mmu_notifier_put(struct mmu_notifier * subscription)887 void mmu_notifier_put(struct mmu_notifier *subscription)
888 {
889 	struct mm_struct *mm = subscription->mm;
890 
891 	spin_lock(&mm->notifier_subscriptions->lock);
892 	if (WARN_ON(!subscription->users) || --subscription->users)
893 		goto out_unlock;
894 	hlist_del_init_rcu(&subscription->hlist);
895 	spin_unlock(&mm->notifier_subscriptions->lock);
896 
897 	call_srcu(&srcu, &subscription->rcu, mmu_notifier_free_rcu);
898 	return;
899 
900 out_unlock:
901 	spin_unlock(&mm->notifier_subscriptions->lock);
902 }
903 EXPORT_SYMBOL_GPL(mmu_notifier_put);
904 
__mmu_interval_notifier_insert(struct mmu_interval_notifier * interval_sub,struct mm_struct * mm,struct mmu_notifier_subscriptions * subscriptions,unsigned long start,unsigned long length,const struct mmu_interval_notifier_ops * ops)905 static int __mmu_interval_notifier_insert(
906 	struct mmu_interval_notifier *interval_sub, struct mm_struct *mm,
907 	struct mmu_notifier_subscriptions *subscriptions, unsigned long start,
908 	unsigned long length, const struct mmu_interval_notifier_ops *ops)
909 {
910 	interval_sub->mm = mm;
911 	interval_sub->ops = ops;
912 	RB_CLEAR_NODE(&interval_sub->interval_tree.rb);
913 	interval_sub->interval_tree.start = start;
914 	/*
915 	 * Note that the representation of the intervals in the interval tree
916 	 * considers the ending point as contained in the interval.
917 	 */
918 	if (length == 0 ||
919 	    check_add_overflow(start, length - 1,
920 			       &interval_sub->interval_tree.last))
921 		return -EOVERFLOW;
922 
923 	/* Must call with a mmget() held */
924 	if (WARN_ON(atomic_read(&mm->mm_users) <= 0))
925 		return -EINVAL;
926 
927 	/* pairs with mmdrop in mmu_interval_notifier_remove() */
928 	mmgrab(mm);
929 
930 	/*
931 	 * If some invalidate_range_start/end region is going on in parallel
932 	 * we don't know what VA ranges are affected, so we must assume this
933 	 * new range is included.
934 	 *
935 	 * If the itree is invalidating then we are not allowed to change
936 	 * it. Retrying until invalidation is done is tricky due to the
937 	 * possibility for live lock, instead defer the add to
938 	 * mn_itree_inv_end() so this algorithm is deterministic.
939 	 *
940 	 * In all cases the value for the interval_sub->invalidate_seq should be
941 	 * odd, see mmu_interval_read_begin()
942 	 */
943 	spin_lock(&subscriptions->lock);
944 	if (subscriptions->active_invalidate_ranges) {
945 		if (mn_itree_is_invalidating(subscriptions))
946 			hlist_add_head(&interval_sub->deferred_item,
947 				       &subscriptions->deferred_list);
948 		else {
949 			subscriptions->invalidate_seq |= 1;
950 			interval_tree_insert(&interval_sub->interval_tree,
951 					     &subscriptions->itree);
952 		}
953 		interval_sub->invalidate_seq = subscriptions->invalidate_seq;
954 	} else {
955 		WARN_ON(mn_itree_is_invalidating(subscriptions));
956 		/*
957 		 * The starting seq for a subscription not under invalidation
958 		 * should be odd, not equal to the current invalidate_seq and
959 		 * invalidate_seq should not 'wrap' to the new seq any time
960 		 * soon.
961 		 */
962 		interval_sub->invalidate_seq =
963 			subscriptions->invalidate_seq - 1;
964 		interval_tree_insert(&interval_sub->interval_tree,
965 				     &subscriptions->itree);
966 	}
967 	spin_unlock(&subscriptions->lock);
968 	return 0;
969 }
970 
971 /**
972  * mmu_interval_notifier_insert - Insert an interval notifier
973  * @interval_sub: Interval subscription to register
974  * @start: Starting virtual address to monitor
975  * @length: Length of the range to monitor
976  * @mm: mm_struct to attach to
977  * @ops: Interval notifier operations to be called on matching events
978  *
979  * This function subscribes the interval notifier for notifications from the
980  * mm.  Upon return the ops related to mmu_interval_notifier will be called
981  * whenever an event that intersects with the given range occurs.
982  *
983  * Upon return the range_notifier may not be present in the interval tree yet.
984  * The caller must use the normal interval notifier read flow via
985  * mmu_interval_read_begin() to establish SPTEs for this range.
986  */
mmu_interval_notifier_insert(struct mmu_interval_notifier * interval_sub,struct mm_struct * mm,unsigned long start,unsigned long length,const struct mmu_interval_notifier_ops * ops)987 int mmu_interval_notifier_insert(struct mmu_interval_notifier *interval_sub,
988 				 struct mm_struct *mm, unsigned long start,
989 				 unsigned long length,
990 				 const struct mmu_interval_notifier_ops *ops)
991 {
992 	struct mmu_notifier_subscriptions *subscriptions;
993 	int ret;
994 
995 	might_lock(&mm->mmap_lock);
996 
997 	subscriptions = smp_load_acquire(&mm->notifier_subscriptions);
998 	if (!subscriptions || !subscriptions->has_itree) {
999 		ret = mmu_notifier_register(NULL, mm);
1000 		if (ret)
1001 			return ret;
1002 		subscriptions = mm->notifier_subscriptions;
1003 	}
1004 	return __mmu_interval_notifier_insert(interval_sub, mm, subscriptions,
1005 					      start, length, ops);
1006 }
1007 EXPORT_SYMBOL_GPL(mmu_interval_notifier_insert);
1008 
mmu_interval_notifier_insert_locked(struct mmu_interval_notifier * interval_sub,struct mm_struct * mm,unsigned long start,unsigned long length,const struct mmu_interval_notifier_ops * ops)1009 int mmu_interval_notifier_insert_locked(
1010 	struct mmu_interval_notifier *interval_sub, struct mm_struct *mm,
1011 	unsigned long start, unsigned long length,
1012 	const struct mmu_interval_notifier_ops *ops)
1013 {
1014 	struct mmu_notifier_subscriptions *subscriptions =
1015 		mm->notifier_subscriptions;
1016 	int ret;
1017 
1018 	mmap_assert_write_locked(mm);
1019 
1020 	if (!subscriptions || !subscriptions->has_itree) {
1021 		ret = __mmu_notifier_register(NULL, mm);
1022 		if (ret)
1023 			return ret;
1024 		subscriptions = mm->notifier_subscriptions;
1025 	}
1026 	return __mmu_interval_notifier_insert(interval_sub, mm, subscriptions,
1027 					      start, length, ops);
1028 }
1029 EXPORT_SYMBOL_GPL(mmu_interval_notifier_insert_locked);
1030 
1031 static bool
mmu_interval_seq_released(struct mmu_notifier_subscriptions * subscriptions,unsigned long seq)1032 mmu_interval_seq_released(struct mmu_notifier_subscriptions *subscriptions,
1033 			  unsigned long seq)
1034 {
1035 	bool ret;
1036 
1037 	spin_lock(&subscriptions->lock);
1038 	ret = subscriptions->invalidate_seq != seq;
1039 	spin_unlock(&subscriptions->lock);
1040 	return ret;
1041 }
1042 
1043 /**
1044  * mmu_interval_notifier_remove - Remove a interval notifier
1045  * @interval_sub: Interval subscription to unregister
1046  *
1047  * This function must be paired with mmu_interval_notifier_insert(). It cannot
1048  * be called from any ops callback.
1049  *
1050  * Once this returns ops callbacks are no longer running on other CPUs and
1051  * will not be called in future.
1052  */
mmu_interval_notifier_remove(struct mmu_interval_notifier * interval_sub)1053 void mmu_interval_notifier_remove(struct mmu_interval_notifier *interval_sub)
1054 {
1055 	struct mm_struct *mm = interval_sub->mm;
1056 	struct mmu_notifier_subscriptions *subscriptions =
1057 		mm->notifier_subscriptions;
1058 	unsigned long seq = 0;
1059 
1060 	might_sleep();
1061 
1062 	spin_lock(&subscriptions->lock);
1063 	if (mn_itree_is_invalidating(subscriptions)) {
1064 		/*
1065 		 * remove is being called after insert put this on the
1066 		 * deferred list, but before the deferred list was processed.
1067 		 */
1068 		if (RB_EMPTY_NODE(&interval_sub->interval_tree.rb)) {
1069 			hlist_del(&interval_sub->deferred_item);
1070 		} else {
1071 			hlist_add_head(&interval_sub->deferred_item,
1072 				       &subscriptions->deferred_list);
1073 			seq = subscriptions->invalidate_seq;
1074 		}
1075 	} else {
1076 		WARN_ON(RB_EMPTY_NODE(&interval_sub->interval_tree.rb));
1077 		interval_tree_remove(&interval_sub->interval_tree,
1078 				     &subscriptions->itree);
1079 	}
1080 	spin_unlock(&subscriptions->lock);
1081 
1082 	/*
1083 	 * The possible sleep on progress in the invalidation requires the
1084 	 * caller not hold any locks held by invalidation callbacks.
1085 	 */
1086 	lock_map_acquire(&__mmu_notifier_invalidate_range_start_map);
1087 	lock_map_release(&__mmu_notifier_invalidate_range_start_map);
1088 	if (seq)
1089 		wait_event(subscriptions->wq,
1090 			   mmu_interval_seq_released(subscriptions, seq));
1091 
1092 	/* pairs with mmgrab in mmu_interval_notifier_insert() */
1093 	mmdrop(mm);
1094 }
1095 EXPORT_SYMBOL_GPL(mmu_interval_notifier_remove);
1096 
1097 /**
1098  * mmu_notifier_synchronize - Ensure all mmu_notifiers are freed
1099  *
1100  * This function ensures that all outstanding async SRU work from
1101  * mmu_notifier_put() is completed. After it returns any mmu_notifier_ops
1102  * associated with an unused mmu_notifier will no longer be called.
1103  *
1104  * Before using the caller must ensure that all of its mmu_notifiers have been
1105  * fully released via mmu_notifier_put().
1106  *
1107  * Modules using the mmu_notifier_put() API should call this in their __exit
1108  * function to avoid module unloading races.
1109  */
mmu_notifier_synchronize(void)1110 void mmu_notifier_synchronize(void)
1111 {
1112 	synchronize_srcu(&srcu);
1113 }
1114 EXPORT_SYMBOL_GPL(mmu_notifier_synchronize);
1115