Commit | Line | Data |
---|---|---|
20c8ccb1 | 1 | // SPDX-License-Identifier: GPL-2.0-only |
cddb8a5c AA |
2 | /* |
3 | * linux/mm/mmu_notifier.c | |
4 | * | |
5 | * Copyright (C) 2008 Qumranet, Inc. | |
6 | * Copyright (C) 2008 SGI | |
93e205a7 | 7 | * Christoph Lameter <cl@linux.com> |
cddb8a5c AA |
8 | */ |
9 | ||
10 | #include <linux/rculist.h> | |
11 | #include <linux/mmu_notifier.h> | |
b95f1b31 | 12 | #include <linux/export.h> |
cddb8a5c AA |
13 | #include <linux/mm.h> |
14 | #include <linux/err.h> | |
99cb252f | 15 | #include <linux/interval_tree.h> |
21a92735 | 16 | #include <linux/srcu.h> |
cddb8a5c AA |
17 | #include <linux/rcupdate.h> |
18 | #include <linux/sched.h> | |
6e84f315 | 19 | #include <linux/sched/mm.h> |
5a0e3ad6 | 20 | #include <linux/slab.h> |
cddb8a5c | 21 | |
21a92735 | 22 | /* global SRCU for all MMs */ |
dde8da6c | 23 | DEFINE_STATIC_SRCU(srcu); |
21a92735 | 24 | |
23b68395 DV |
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 | ||
56f434f4 | 31 | /* |
984cfe4e JG |
32 | * The mmu_notifier_subscriptions structure is allocated and installed in |
33 | * mm->notifier_subscriptions inside the mm_take_all_locks() protected | |
56f434f4 JG |
34 | * critical section and it's released only when mm_count reaches zero |
35 | * in mmdrop(). | |
36 | */ | |
984cfe4e | 37 | struct mmu_notifier_subscriptions { |
56f434f4 JG |
38 | /* all mmu notifiers registered in this mm are queued in this list */ |
39 | struct hlist_head list; | |
99cb252f | 40 | bool has_itree; |
56f434f4 JG |
41 | /* to serialize the list modifications and hlist_unhashed */ |
42 | spinlock_t lock; | |
99cb252f JG |
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; | |
56f434f4 JG |
48 | }; |
49 | ||
99cb252f JG |
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 | |
984cfe4e | 68 | * - subscriptions->invalidate_seq & 1 == True (odd) |
99cb252f JG |
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 | |
984cfe4e | 74 | * - subscriptions->invalidate_seq & 1 == False (even) |
99cb252f JG |
75 | * - some range on the mm_struct is being invalidated |
76 | * - the itree is allowed to change | |
77 | * | |
984cfe4e | 78 | * Operations on notifier_subscriptions->invalidate_seq (under spinlock): |
99cb252f JG |
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 | |
5292e24a | 84 | * no mmu_interval_notifier monitoring the VA. |
99cb252f | 85 | */ |
984cfe4e JG |
86 | static bool |
87 | mn_itree_is_invalidating(struct mmu_notifier_subscriptions *subscriptions) | |
99cb252f | 88 | { |
984cfe4e JG |
89 | lockdep_assert_held(&subscriptions->lock); |
90 | return subscriptions->invalidate_seq & 1; | |
99cb252f JG |
91 | } |
92 | ||
93 | static struct mmu_interval_notifier * | |
984cfe4e | 94 | mn_itree_inv_start_range(struct mmu_notifier_subscriptions *subscriptions, |
99cb252f JG |
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 | ||
984cfe4e JG |
101 | spin_lock(&subscriptions->lock); |
102 | subscriptions->active_invalidate_ranges++; | |
103 | node = interval_tree_iter_first(&subscriptions->itree, range->start, | |
99cb252f JG |
104 | range->end - 1); |
105 | if (node) { | |
984cfe4e | 106 | subscriptions->invalidate_seq |= 1; |
99cb252f JG |
107 | res = container_of(node, struct mmu_interval_notifier, |
108 | interval_tree); | |
109 | } | |
110 | ||
984cfe4e JG |
111 | *seq = subscriptions->invalidate_seq; |
112 | spin_unlock(&subscriptions->lock); | |
99cb252f JG |
113 | return res; |
114 | } | |
115 | ||
116 | static struct mmu_interval_notifier * | |
5292e24a | 117 | mn_itree_inv_next(struct mmu_interval_notifier *interval_sub, |
99cb252f JG |
118 | const struct mmu_notifier_range *range) |
119 | { | |
120 | struct interval_tree_node *node; | |
121 | ||
5292e24a JG |
122 | node = interval_tree_iter_next(&interval_sub->interval_tree, |
123 | range->start, range->end - 1); | |
99cb252f JG |
124 | if (!node) |
125 | return NULL; | |
126 | return container_of(node, struct mmu_interval_notifier, interval_tree); | |
127 | } | |
128 | ||
984cfe4e | 129 | static void mn_itree_inv_end(struct mmu_notifier_subscriptions *subscriptions) |
99cb252f | 130 | { |
5292e24a | 131 | struct mmu_interval_notifier *interval_sub; |
99cb252f JG |
132 | struct hlist_node *next; |
133 | ||
984cfe4e JG |
134 | spin_lock(&subscriptions->lock); |
135 | if (--subscriptions->active_invalidate_ranges || | |
136 | !mn_itree_is_invalidating(subscriptions)) { | |
137 | spin_unlock(&subscriptions->lock); | |
99cb252f JG |
138 | return; |
139 | } | |
140 | ||
141 | /* Make invalidate_seq even */ | |
984cfe4e | 142 | subscriptions->invalidate_seq++; |
99cb252f JG |
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 | */ | |
5292e24a JG |
150 | hlist_for_each_entry_safe(interval_sub, next, |
151 | &subscriptions->deferred_list, | |
99cb252f | 152 | deferred_item) { |
5292e24a JG |
153 | if (RB_EMPTY_NODE(&interval_sub->interval_tree.rb)) |
154 | interval_tree_insert(&interval_sub->interval_tree, | |
984cfe4e | 155 | &subscriptions->itree); |
99cb252f | 156 | else |
5292e24a | 157 | interval_tree_remove(&interval_sub->interval_tree, |
984cfe4e | 158 | &subscriptions->itree); |
5292e24a | 159 | hlist_del(&interval_sub->deferred_item); |
99cb252f | 160 | } |
984cfe4e | 161 | spin_unlock(&subscriptions->lock); |
99cb252f | 162 | |
984cfe4e | 163 | wake_up_all(&subscriptions->wq); |
99cb252f JG |
164 | } |
165 | ||
166 | /** | |
167 | * mmu_interval_read_begin - Begin a read side critical section against a VA | |
168 | * range | |
d49653f3 | 169 | * @interval_sub: The interval subscription |
99cb252f JG |
170 | * |
171 | * mmu_iterval_read_begin()/mmu_iterval_read_retry() implement a | |
5292e24a JG |
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. | |
99cb252f JG |
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 | */ | |
5292e24a JG |
185 | unsigned long |
186 | mmu_interval_read_begin(struct mmu_interval_notifier *interval_sub) | |
99cb252f | 187 | { |
984cfe4e | 188 | struct mmu_notifier_subscriptions *subscriptions = |
5292e24a | 189 | interval_sub->mm->notifier_subscriptions; |
99cb252f JG |
190 | unsigned long seq; |
191 | bool is_invalidating; | |
192 | ||
193 | /* | |
5292e24a JG |
194 | * If the subscription has a different seq value under the user_lock |
195 | * than we started with then it has collided. | |
99cb252f | 196 | * |
5292e24a JG |
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. | |
99cb252f JG |
200 | * |
201 | * The locking looks broadly like this: | |
202 | * mn_tree_invalidate_start(): mmu_interval_read_begin(): | |
203 | * spin_lock | |
5292e24a | 204 | * seq = READ_ONCE(interval_sub->invalidate_seq); |
984cfe4e | 205 | * seq == subs->invalidate_seq |
99cb252f JG |
206 | * spin_unlock |
207 | * spin_lock | |
984cfe4e | 208 | * seq = ++subscriptions->invalidate_seq |
99cb252f JG |
209 | * spin_unlock |
210 | * op->invalidate_range(): | |
211 | * user_lock | |
212 | * mmu_interval_set_seq() | |
5292e24a | 213 | * interval_sub->invalidate_seq = seq |
99cb252f JG |
214 | * user_unlock |
215 | * | |
216 | * [Required: mmu_interval_read_retry() == true] | |
217 | * | |
218 | * mn_itree_inv_end(): | |
219 | * spin_lock | |
984cfe4e | 220 | * seq = ++subscriptions->invalidate_seq |
99cb252f JG |
221 | * spin_unlock |
222 | * | |
223 | * user_lock | |
224 | * mmu_interval_read_retry(): | |
5292e24a | 225 | * interval_sub->invalidate_seq != seq |
99cb252f JG |
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 | */ | |
984cfe4e | 232 | spin_lock(&subscriptions->lock); |
99cb252f | 233 | /* Pairs with the WRITE_ONCE in mmu_interval_set_seq() */ |
5292e24a | 234 | seq = READ_ONCE(interval_sub->invalidate_seq); |
984cfe4e JG |
235 | is_invalidating = seq == subscriptions->invalidate_seq; |
236 | spin_unlock(&subscriptions->lock); | |
99cb252f JG |
237 | |
238 | /* | |
5292e24a | 239 | * interval_sub->invalidate_seq must always be set to an odd value via |
99cb252f JG |
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 | |
984cfe4e | 243 | * subscriptions->invalidate_seq is even in the idle state. |
99cb252f JG |
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) | |
984cfe4e JG |
248 | wait_event(subscriptions->wq, |
249 | READ_ONCE(subscriptions->invalidate_seq) != seq); | |
99cb252f JG |
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 | ||
984cfe4e | 261 | static void mn_itree_release(struct mmu_notifier_subscriptions *subscriptions, |
99cb252f JG |
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 | }; | |
5292e24a | 271 | struct mmu_interval_notifier *interval_sub; |
99cb252f JG |
272 | unsigned long cur_seq; |
273 | bool ret; | |
274 | ||
5292e24a JG |
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); | |
99cb252f JG |
281 | WARN_ON(!ret); |
282 | } | |
283 | ||
984cfe4e | 284 | mn_itree_inv_end(subscriptions); |
99cb252f JG |
285 | } |
286 | ||
cddb8a5c AA |
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 | |
984cfe4e JG |
294 | * the notifier_subscriptions->lock in addition to SRCU and it serializes |
295 | * against the other mmu notifiers with SRCU. struct mmu_notifier_subscriptions | |
cddb8a5c AA |
296 | * can't go away from under us as exit_mmap holds an mm_count pin |
297 | * itself. | |
298 | */ | |
984cfe4e | 299 | static void mn_hlist_release(struct mmu_notifier_subscriptions *subscriptions, |
99cb252f | 300 | struct mm_struct *mm) |
cddb8a5c | 301 | { |
1991722a | 302 | struct mmu_notifier *subscription; |
21a92735 | 303 | int id; |
3ad3d901 XG |
304 | |
305 | /* | |
d34883d4 XG |
306 | * SRCU here will block mmu_notifier_unregister until |
307 | * ->release returns. | |
3ad3d901 | 308 | */ |
21a92735 | 309 | id = srcu_read_lock(&srcu); |
63886bad QC |
310 | hlist_for_each_entry_rcu(subscription, &subscriptions->list, hlist, |
311 | srcu_read_lock_held(&srcu)) | |
d34883d4 XG |
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 | */ | |
1991722a JG |
318 | if (subscription->ops->release) |
319 | subscription->ops->release(subscription, mm); | |
d34883d4 | 320 | |
984cfe4e JG |
321 | spin_lock(&subscriptions->lock); |
322 | while (unlikely(!hlist_empty(&subscriptions->list))) { | |
1991722a JG |
323 | subscription = hlist_entry(subscriptions->list.first, |
324 | struct mmu_notifier, hlist); | |
cddb8a5c | 325 | /* |
d34883d4 XG |
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. | |
cddb8a5c | 330 | */ |
1991722a | 331 | hlist_del_init_rcu(&subscription->hlist); |
cddb8a5c | 332 | } |
984cfe4e | 333 | spin_unlock(&subscriptions->lock); |
b972216e | 334 | srcu_read_unlock(&srcu, id); |
cddb8a5c AA |
335 | |
336 | /* | |
d34883d4 XG |
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 | * | |
984cfe4e JG |
342 | * The notifier_subscriptions can't go away from under us because |
343 | * one mm_count is held by exit_mmap. | |
cddb8a5c | 344 | */ |
21a92735 | 345 | synchronize_srcu(&srcu); |
cddb8a5c AA |
346 | } |
347 | ||
99cb252f JG |
348 | void __mmu_notifier_release(struct mm_struct *mm) |
349 | { | |
984cfe4e JG |
350 | struct mmu_notifier_subscriptions *subscriptions = |
351 | mm->notifier_subscriptions; | |
99cb252f | 352 | |
984cfe4e JG |
353 | if (subscriptions->has_itree) |
354 | mn_itree_release(subscriptions, mm); | |
99cb252f | 355 | |
984cfe4e JG |
356 | if (!hlist_empty(&subscriptions->list)) |
357 | mn_hlist_release(subscriptions, mm); | |
99cb252f JG |
358 | } |
359 | ||
cddb8a5c AA |
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 | */ | |
365 | int __mmu_notifier_clear_flush_young(struct mm_struct *mm, | |
57128468 ALC |
366 | unsigned long start, |
367 | unsigned long end) | |
cddb8a5c | 368 | { |
1991722a | 369 | struct mmu_notifier *subscription; |
21a92735 | 370 | int young = 0, id; |
cddb8a5c | 371 | |
21a92735 | 372 | id = srcu_read_lock(&srcu); |
1991722a | 373 | hlist_for_each_entry_rcu(subscription, |
63886bad QC |
374 | &mm->notifier_subscriptions->list, hlist, |
375 | srcu_read_lock_held(&srcu)) { | |
1991722a JG |
376 | if (subscription->ops->clear_flush_young) |
377 | young |= subscription->ops->clear_flush_young( | |
378 | subscription, mm, start, end); | |
cddb8a5c | 379 | } |
21a92735 | 380 | srcu_read_unlock(&srcu, id); |
cddb8a5c AA |
381 | |
382 | return young; | |
383 | } | |
384 | ||
1d7715c6 VD |
385 | int __mmu_notifier_clear_young(struct mm_struct *mm, |
386 | unsigned long start, | |
387 | unsigned long end) | |
388 | { | |
1991722a | 389 | struct mmu_notifier *subscription; |
1d7715c6 VD |
390 | int young = 0, id; |
391 | ||
392 | id = srcu_read_lock(&srcu); | |
1991722a | 393 | hlist_for_each_entry_rcu(subscription, |
63886bad QC |
394 | &mm->notifier_subscriptions->list, hlist, |
395 | srcu_read_lock_held(&srcu)) { | |
1991722a JG |
396 | if (subscription->ops->clear_young) |
397 | young |= subscription->ops->clear_young(subscription, | |
398 | mm, start, end); | |
1d7715c6 VD |
399 | } |
400 | srcu_read_unlock(&srcu, id); | |
401 | ||
402 | return young; | |
403 | } | |
404 | ||
8ee53820 AA |
405 | int __mmu_notifier_test_young(struct mm_struct *mm, |
406 | unsigned long address) | |
407 | { | |
1991722a | 408 | struct mmu_notifier *subscription; |
21a92735 | 409 | int young = 0, id; |
8ee53820 | 410 | |
21a92735 | 411 | id = srcu_read_lock(&srcu); |
1991722a | 412 | hlist_for_each_entry_rcu(subscription, |
63886bad QC |
413 | &mm->notifier_subscriptions->list, hlist, |
414 | srcu_read_lock_held(&srcu)) { | |
1991722a JG |
415 | if (subscription->ops->test_young) { |
416 | young = subscription->ops->test_young(subscription, mm, | |
417 | address); | |
8ee53820 AA |
418 | if (young) |
419 | break; | |
420 | } | |
421 | } | |
21a92735 | 422 | srcu_read_unlock(&srcu, id); |
8ee53820 AA |
423 | |
424 | return young; | |
425 | } | |
426 | ||
828502d3 IE |
427 | void __mmu_notifier_change_pte(struct mm_struct *mm, unsigned long address, |
428 | pte_t pte) | |
429 | { | |
1991722a | 430 | struct mmu_notifier *subscription; |
21a92735 | 431 | int id; |
828502d3 | 432 | |
21a92735 | 433 | id = srcu_read_lock(&srcu); |
1991722a | 434 | hlist_for_each_entry_rcu(subscription, |
63886bad QC |
435 | &mm->notifier_subscriptions->list, hlist, |
436 | srcu_read_lock_held(&srcu)) { | |
1991722a JG |
437 | if (subscription->ops->change_pte) |
438 | subscription->ops->change_pte(subscription, mm, address, | |
439 | pte); | |
828502d3 | 440 | } |
21a92735 | 441 | srcu_read_unlock(&srcu, id); |
828502d3 IE |
442 | } |
443 | ||
984cfe4e | 444 | static int mn_itree_invalidate(struct mmu_notifier_subscriptions *subscriptions, |
99cb252f JG |
445 | const struct mmu_notifier_range *range) |
446 | { | |
5292e24a | 447 | struct mmu_interval_notifier *interval_sub; |
99cb252f JG |
448 | unsigned long cur_seq; |
449 | ||
5292e24a JG |
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)) { | |
99cb252f JG |
454 | bool ret; |
455 | ||
5292e24a JG |
456 | ret = interval_sub->ops->invalidate(interval_sub, range, |
457 | cur_seq); | |
99cb252f JG |
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 | */ | |
984cfe4e | 471 | mn_itree_inv_end(subscriptions); |
99cb252f JG |
472 | return -EAGAIN; |
473 | } | |
474 | ||
984cfe4e JG |
475 | static int mn_hlist_invalidate_range_start( |
476 | struct mmu_notifier_subscriptions *subscriptions, | |
477 | struct mmu_notifier_range *range) | |
cddb8a5c | 478 | { |
1991722a | 479 | struct mmu_notifier *subscription; |
93065ac7 | 480 | int ret = 0; |
21a92735 | 481 | int id; |
cddb8a5c | 482 | |
21a92735 | 483 | id = srcu_read_lock(&srcu); |
63886bad QC |
484 | hlist_for_each_entry_rcu(subscription, &subscriptions->list, hlist, |
485 | srcu_read_lock_held(&srcu)) { | |
1991722a JG |
486 | const struct mmu_notifier_ops *ops = subscription->ops; |
487 | ||
488 | if (ops->invalidate_range_start) { | |
ba170f76 DV |
489 | int _ret; |
490 | ||
491 | if (!mmu_notifier_range_blockable(range)) | |
492 | non_block_start(); | |
1991722a | 493 | _ret = ops->invalidate_range_start(subscription, range); |
ba170f76 DV |
494 | if (!mmu_notifier_range_blockable(range)) |
495 | non_block_end(); | |
93065ac7 MH |
496 | if (_ret) { |
497 | pr_info("%pS callback failed with %d in %sblockable context.\n", | |
1991722a JG |
498 | ops->invalidate_range_start, _ret, |
499 | !mmu_notifier_range_blockable(range) ? | |
500 | "non-" : | |
501 | ""); | |
8402ce61 | 502 | WARN_ON(mmu_notifier_range_blockable(range) || |
df2ec764 | 503 | _ret != -EAGAIN); |
93065ac7 MH |
504 | ret = _ret; |
505 | } | |
506 | } | |
cddb8a5c | 507 | } |
21a92735 | 508 | srcu_read_unlock(&srcu, id); |
93065ac7 MH |
509 | |
510 | return ret; | |
cddb8a5c AA |
511 | } |
512 | ||
99cb252f JG |
513 | int __mmu_notifier_invalidate_range_start(struct mmu_notifier_range *range) |
514 | { | |
984cfe4e JG |
515 | struct mmu_notifier_subscriptions *subscriptions = |
516 | range->mm->notifier_subscriptions; | |
99cb252f JG |
517 | int ret; |
518 | ||
984cfe4e JG |
519 | if (subscriptions->has_itree) { |
520 | ret = mn_itree_invalidate(subscriptions, range); | |
99cb252f JG |
521 | if (ret) |
522 | return ret; | |
523 | } | |
984cfe4e JG |
524 | if (!hlist_empty(&subscriptions->list)) |
525 | return mn_hlist_invalidate_range_start(subscriptions, range); | |
99cb252f JG |
526 | return 0; |
527 | } | |
528 | ||
984cfe4e JG |
529 | static void |
530 | mn_hlist_invalidate_end(struct mmu_notifier_subscriptions *subscriptions, | |
531 | struct mmu_notifier_range *range, bool only_end) | |
cddb8a5c | 532 | { |
1991722a | 533 | struct mmu_notifier *subscription; |
21a92735 | 534 | int id; |
cddb8a5c | 535 | |
21a92735 | 536 | id = srcu_read_lock(&srcu); |
63886bad QC |
537 | hlist_for_each_entry_rcu(subscription, &subscriptions->list, hlist, |
538 | srcu_read_lock_held(&srcu)) { | |
0f0a327f JR |
539 | /* |
540 | * Call invalidate_range here too to avoid the need for the | |
541 | * subsystem of having to register an invalidate_range_end | |
542 | * call-back when there is invalidate_range already. Usually a | |
543 | * subsystem registers either invalidate_range_start()/end() or | |
544 | * invalidate_range(), so this will be no additional overhead | |
545 | * (besides the pointer check). | |
4645b9fe JG |
546 | * |
547 | * We skip call to invalidate_range() if we know it is safe ie | |
548 | * call site use mmu_notifier_invalidate_range_only_end() which | |
549 | * is safe to do when we know that a call to invalidate_range() | |
550 | * already happen under page table lock. | |
0f0a327f | 551 | */ |
1991722a JG |
552 | if (!only_end && subscription->ops->invalidate_range) |
553 | subscription->ops->invalidate_range(subscription, | |
554 | range->mm, | |
555 | range->start, | |
556 | range->end); | |
557 | if (subscription->ops->invalidate_range_end) { | |
ba170f76 DV |
558 | if (!mmu_notifier_range_blockable(range)) |
559 | non_block_start(); | |
1991722a JG |
560 | subscription->ops->invalidate_range_end(subscription, |
561 | range); | |
ba170f76 DV |
562 | if (!mmu_notifier_range_blockable(range)) |
563 | non_block_end(); | |
564 | } | |
cddb8a5c | 565 | } |
21a92735 | 566 | srcu_read_unlock(&srcu, id); |
99cb252f JG |
567 | } |
568 | ||
569 | void __mmu_notifier_invalidate_range_end(struct mmu_notifier_range *range, | |
570 | bool only_end) | |
571 | { | |
984cfe4e JG |
572 | struct mmu_notifier_subscriptions *subscriptions = |
573 | range->mm->notifier_subscriptions; | |
99cb252f JG |
574 | |
575 | lock_map_acquire(&__mmu_notifier_invalidate_range_start_map); | |
984cfe4e JG |
576 | if (subscriptions->has_itree) |
577 | mn_itree_inv_end(subscriptions); | |
99cb252f | 578 | |
984cfe4e JG |
579 | if (!hlist_empty(&subscriptions->list)) |
580 | mn_hlist_invalidate_end(subscriptions, range, only_end); | |
23b68395 | 581 | lock_map_release(&__mmu_notifier_invalidate_range_start_map); |
cddb8a5c AA |
582 | } |
583 | ||
0f0a327f JR |
584 | void __mmu_notifier_invalidate_range(struct mm_struct *mm, |
585 | unsigned long start, unsigned long end) | |
586 | { | |
1991722a | 587 | struct mmu_notifier *subscription; |
0f0a327f JR |
588 | int id; |
589 | ||
590 | id = srcu_read_lock(&srcu); | |
1991722a | 591 | hlist_for_each_entry_rcu(subscription, |
63886bad QC |
592 | &mm->notifier_subscriptions->list, hlist, |
593 | srcu_read_lock_held(&srcu)) { | |
1991722a JG |
594 | if (subscription->ops->invalidate_range) |
595 | subscription->ops->invalidate_range(subscription, mm, | |
596 | start, end); | |
0f0a327f JR |
597 | } |
598 | srcu_read_unlock(&srcu, id); | |
599 | } | |
0f0a327f | 600 | |
56c57103 | 601 | /* |
c1e8d7c6 | 602 | * Same as mmu_notifier_register but here the caller must hold the mmap_lock in |
99cb252f JG |
603 | * write mode. A NULL mn signals the notifier is being registered for itree |
604 | * mode. | |
56c57103 | 605 | */ |
1991722a JG |
606 | int __mmu_notifier_register(struct mmu_notifier *subscription, |
607 | struct mm_struct *mm) | |
cddb8a5c | 608 | { |
984cfe4e | 609 | struct mmu_notifier_subscriptions *subscriptions = NULL; |
cddb8a5c AA |
610 | int ret; |
611 | ||
42fc5414 | 612 | mmap_assert_write_locked(mm); |
cddb8a5c AA |
613 | BUG_ON(atomic_read(&mm->mm_users) <= 0); |
614 | ||
984cfe4e | 615 | if (!mm->notifier_subscriptions) { |
70df291b JG |
616 | /* |
617 | * kmalloc cannot be called under mm_take_all_locks(), but we | |
984cfe4e | 618 | * know that mm->notifier_subscriptions can't change while we |
c1e8d7c6 | 619 | * hold the write side of the mmap_lock. |
70df291b | 620 | */ |
984cfe4e JG |
621 | subscriptions = kzalloc( |
622 | sizeof(struct mmu_notifier_subscriptions), GFP_KERNEL); | |
623 | if (!subscriptions) | |
70df291b JG |
624 | return -ENOMEM; |
625 | ||
984cfe4e JG |
626 | INIT_HLIST_HEAD(&subscriptions->list); |
627 | spin_lock_init(&subscriptions->lock); | |
628 | subscriptions->invalidate_seq = 2; | |
629 | subscriptions->itree = RB_ROOT_CACHED; | |
630 | init_waitqueue_head(&subscriptions->wq); | |
631 | INIT_HLIST_HEAD(&subscriptions->deferred_list); | |
70df291b | 632 | } |
35cfa2b0 | 633 | |
cddb8a5c AA |
634 | ret = mm_take_all_locks(mm); |
635 | if (unlikely(ret)) | |
35cfa2b0 | 636 | goto out_clean; |
cddb8a5c | 637 | |
cddb8a5c AA |
638 | /* |
639 | * Serialize the update against mmu_notifier_unregister. A | |
640 | * side note: mmu_notifier_release can't run concurrently with | |
641 | * us because we hold the mm_users pin (either implicitly as | |
642 | * current->mm or explicitly with get_task_mm() or similar). | |
643 | * We can't race against any other mmu notifier method either | |
644 | * thanks to mm_take_all_locks(). | |
99cb252f | 645 | * |
984cfe4e JG |
646 | * release semantics on the initialization of the |
647 | * mmu_notifier_subscriptions's contents are provided for unlocked | |
648 | * readers. acquire can only be used while holding the mmgrab or | |
649 | * mmget, and is safe because once created the | |
650 | * mmu_notifier_subscriptions is not freed until the mm is destroyed. | |
c1e8d7c6 | 651 | * As above, users holding the mmap_lock or one of the |
99cb252f | 652 | * mm_take_all_locks() do not need to use acquire semantics. |
cddb8a5c | 653 | */ |
984cfe4e JG |
654 | if (subscriptions) |
655 | smp_store_release(&mm->notifier_subscriptions, subscriptions); | |
70df291b | 656 | |
1991722a | 657 | if (subscription) { |
99cb252f JG |
658 | /* Pairs with the mmdrop in mmu_notifier_unregister_* */ |
659 | mmgrab(mm); | |
1991722a JG |
660 | subscription->mm = mm; |
661 | subscription->users = 1; | |
99cb252f | 662 | |
984cfe4e | 663 | spin_lock(&mm->notifier_subscriptions->lock); |
1991722a | 664 | hlist_add_head_rcu(&subscription->hlist, |
984cfe4e JG |
665 | &mm->notifier_subscriptions->list); |
666 | spin_unlock(&mm->notifier_subscriptions->lock); | |
99cb252f | 667 | } else |
984cfe4e | 668 | mm->notifier_subscriptions->has_itree = true; |
cddb8a5c AA |
669 | |
670 | mm_drop_all_locks(mm); | |
70df291b JG |
671 | BUG_ON(atomic_read(&mm->mm_users) <= 0); |
672 | return 0; | |
673 | ||
35cfa2b0 | 674 | out_clean: |
984cfe4e | 675 | kfree(subscriptions); |
cddb8a5c AA |
676 | return ret; |
677 | } | |
56c57103 | 678 | EXPORT_SYMBOL_GPL(__mmu_notifier_register); |
cddb8a5c | 679 | |
2c7933f5 JG |
680 | /** |
681 | * mmu_notifier_register - Register a notifier on a mm | |
d49653f3 | 682 | * @subscription: The notifier to attach |
2c7933f5 JG |
683 | * @mm: The mm to attach the notifier to |
684 | * | |
c1e8d7c6 | 685 | * Must not hold mmap_lock nor any other VM related lock when calling |
cddb8a5c AA |
686 | * this registration function. Must also ensure mm_users can't go down |
687 | * to zero while this runs to avoid races with mmu_notifier_release, | |
688 | * so mm has to be current->mm or the mm should be pinned safely such | |
689 | * as with get_task_mm(). If the mm is not current->mm, the mm_users | |
690 | * pin should be released by calling mmput after mmu_notifier_register | |
2c7933f5 JG |
691 | * returns. |
692 | * | |
693 | * mmu_notifier_unregister() or mmu_notifier_put() must be always called to | |
694 | * unregister the notifier. | |
695 | * | |
1991722a | 696 | * While the caller has a mmu_notifier get the subscription->mm pointer will remain |
2c7933f5 | 697 | * valid, and can be converted to an active mm pointer via mmget_not_zero(). |
cddb8a5c | 698 | */ |
1991722a JG |
699 | int mmu_notifier_register(struct mmu_notifier *subscription, |
700 | struct mm_struct *mm) | |
cddb8a5c | 701 | { |
56c57103 | 702 | int ret; |
cddb8a5c | 703 | |
d8ed45c5 | 704 | mmap_write_lock(mm); |
1991722a | 705 | ret = __mmu_notifier_register(subscription, mm); |
d8ed45c5 | 706 | mmap_write_unlock(mm); |
56c57103 | 707 | return ret; |
cddb8a5c | 708 | } |
56c57103 | 709 | EXPORT_SYMBOL_GPL(mmu_notifier_register); |
cddb8a5c | 710 | |
2c7933f5 JG |
711 | static struct mmu_notifier * |
712 | find_get_mmu_notifier(struct mm_struct *mm, const struct mmu_notifier_ops *ops) | |
713 | { | |
1991722a | 714 | struct mmu_notifier *subscription; |
2c7933f5 | 715 | |
984cfe4e | 716 | spin_lock(&mm->notifier_subscriptions->lock); |
1991722a | 717 | hlist_for_each_entry_rcu(subscription, |
63886bad QC |
718 | &mm->notifier_subscriptions->list, hlist, |
719 | lockdep_is_held(&mm->notifier_subscriptions->lock)) { | |
1991722a | 720 | if (subscription->ops != ops) |
2c7933f5 JG |
721 | continue; |
722 | ||
1991722a JG |
723 | if (likely(subscription->users != UINT_MAX)) |
724 | subscription->users++; | |
2c7933f5 | 725 | else |
1991722a | 726 | subscription = ERR_PTR(-EOVERFLOW); |
984cfe4e | 727 | spin_unlock(&mm->notifier_subscriptions->lock); |
1991722a | 728 | return subscription; |
2c7933f5 | 729 | } |
984cfe4e | 730 | spin_unlock(&mm->notifier_subscriptions->lock); |
2c7933f5 JG |
731 | return NULL; |
732 | } | |
733 | ||
734 | /** | |
735 | * mmu_notifier_get_locked - Return the single struct mmu_notifier for | |
736 | * the mm & ops | |
737 | * @ops: The operations struct being subscribe with | |
738 | * @mm : The mm to attach notifiers too | |
739 | * | |
740 | * This function either allocates a new mmu_notifier via | |
741 | * ops->alloc_notifier(), or returns an already existing notifier on the | |
742 | * list. The value of the ops pointer is used to determine when two notifiers | |
743 | * are the same. | |
744 | * | |
745 | * Each call to mmu_notifier_get() must be paired with a call to | |
c1e8d7c6 | 746 | * mmu_notifier_put(). The caller must hold the write side of mm->mmap_lock. |
2c7933f5 JG |
747 | * |
748 | * While the caller has a mmu_notifier get the mm pointer will remain valid, | |
749 | * and can be converted to an active mm pointer via mmget_not_zero(). | |
750 | */ | |
751 | struct mmu_notifier *mmu_notifier_get_locked(const struct mmu_notifier_ops *ops, | |
752 | struct mm_struct *mm) | |
753 | { | |
1991722a | 754 | struct mmu_notifier *subscription; |
2c7933f5 JG |
755 | int ret; |
756 | ||
42fc5414 | 757 | mmap_assert_write_locked(mm); |
2c7933f5 | 758 | |
984cfe4e | 759 | if (mm->notifier_subscriptions) { |
1991722a JG |
760 | subscription = find_get_mmu_notifier(mm, ops); |
761 | if (subscription) | |
762 | return subscription; | |
2c7933f5 JG |
763 | } |
764 | ||
1991722a JG |
765 | subscription = ops->alloc_notifier(mm); |
766 | if (IS_ERR(subscription)) | |
767 | return subscription; | |
768 | subscription->ops = ops; | |
769 | ret = __mmu_notifier_register(subscription, mm); | |
2c7933f5 JG |
770 | if (ret) |
771 | goto out_free; | |
1991722a | 772 | return subscription; |
2c7933f5 | 773 | out_free: |
1991722a | 774 | subscription->ops->free_notifier(subscription); |
2c7933f5 JG |
775 | return ERR_PTR(ret); |
776 | } | |
777 | EXPORT_SYMBOL_GPL(mmu_notifier_get_locked); | |
778 | ||
cddb8a5c | 779 | /* this is called after the last mmu_notifier_unregister() returned */ |
984cfe4e | 780 | void __mmu_notifier_subscriptions_destroy(struct mm_struct *mm) |
cddb8a5c | 781 | { |
984cfe4e JG |
782 | BUG_ON(!hlist_empty(&mm->notifier_subscriptions->list)); |
783 | kfree(mm->notifier_subscriptions); | |
784 | mm->notifier_subscriptions = LIST_POISON1; /* debug */ | |
cddb8a5c AA |
785 | } |
786 | ||
787 | /* | |
788 | * This releases the mm_count pin automatically and frees the mm | |
789 | * structure if it was the last user of it. It serializes against | |
21a92735 SG |
790 | * running mmu notifiers with SRCU and against mmu_notifier_unregister |
791 | * with the unregister lock + SRCU. All sptes must be dropped before | |
cddb8a5c AA |
792 | * calling mmu_notifier_unregister. ->release or any other notifier |
793 | * method may be invoked concurrently with mmu_notifier_unregister, | |
794 | * and only after mmu_notifier_unregister returned we're guaranteed | |
795 | * that ->release or any other method can't run anymore. | |
796 | */ | |
1991722a JG |
797 | void mmu_notifier_unregister(struct mmu_notifier *subscription, |
798 | struct mm_struct *mm) | |
cddb8a5c AA |
799 | { |
800 | BUG_ON(atomic_read(&mm->mm_count) <= 0); | |
801 | ||
1991722a | 802 | if (!hlist_unhashed(&subscription->hlist)) { |
d34883d4 XG |
803 | /* |
804 | * SRCU here will force exit_mmap to wait for ->release to | |
805 | * finish before freeing the pages. | |
806 | */ | |
21a92735 | 807 | int id; |
3ad3d901 | 808 | |
d34883d4 | 809 | id = srcu_read_lock(&srcu); |
cddb8a5c | 810 | /* |
d34883d4 XG |
811 | * exit_mmap will block in mmu_notifier_release to guarantee |
812 | * that ->release is called before freeing the pages. | |
cddb8a5c | 813 | */ |
1991722a JG |
814 | if (subscription->ops->release) |
815 | subscription->ops->release(subscription, mm); | |
d34883d4 | 816 | srcu_read_unlock(&srcu, id); |
3ad3d901 | 817 | |
984cfe4e | 818 | spin_lock(&mm->notifier_subscriptions->lock); |
751efd86 | 819 | /* |
d34883d4 XG |
820 | * Can not use list_del_rcu() since __mmu_notifier_release |
821 | * can delete it before we hold the lock. | |
751efd86 | 822 | */ |
1991722a | 823 | hlist_del_init_rcu(&subscription->hlist); |
984cfe4e | 824 | spin_unlock(&mm->notifier_subscriptions->lock); |
d34883d4 | 825 | } |
cddb8a5c AA |
826 | |
827 | /* | |
d34883d4 | 828 | * Wait for any running method to finish, of course including |
83a35e36 | 829 | * ->release if it was run by mmu_notifier_release instead of us. |
cddb8a5c | 830 | */ |
21a92735 | 831 | synchronize_srcu(&srcu); |
cddb8a5c AA |
832 | |
833 | BUG_ON(atomic_read(&mm->mm_count) <= 0); | |
834 | ||
835 | mmdrop(mm); | |
836 | } | |
837 | EXPORT_SYMBOL_GPL(mmu_notifier_unregister); | |
21a92735 | 838 | |
2c7933f5 JG |
839 | static void mmu_notifier_free_rcu(struct rcu_head *rcu) |
840 | { | |
1991722a JG |
841 | struct mmu_notifier *subscription = |
842 | container_of(rcu, struct mmu_notifier, rcu); | |
843 | struct mm_struct *mm = subscription->mm; | |
2c7933f5 | 844 | |
1991722a | 845 | subscription->ops->free_notifier(subscription); |
2c7933f5 JG |
846 | /* Pairs with the get in __mmu_notifier_register() */ |
847 | mmdrop(mm); | |
848 | } | |
849 | ||
850 | /** | |
851 | * mmu_notifier_put - Release the reference on the notifier | |
d49653f3 | 852 | * @subscription: The notifier to act on |
2c7933f5 JG |
853 | * |
854 | * This function must be paired with each mmu_notifier_get(), it releases the | |
855 | * reference obtained by the get. If this is the last reference then process | |
856 | * to free the notifier will be run asynchronously. | |
857 | * | |
858 | * Unlike mmu_notifier_unregister() the get/put flow only calls ops->release | |
859 | * when the mm_struct is destroyed. Instead free_notifier is always called to | |
860 | * release any resources held by the user. | |
861 | * | |
862 | * As ops->release is not guaranteed to be called, the user must ensure that | |
863 | * all sptes are dropped, and no new sptes can be established before | |
864 | * mmu_notifier_put() is called. | |
865 | * | |
866 | * This function can be called from the ops->release callback, however the | |
867 | * caller must still ensure it is called pairwise with mmu_notifier_get(). | |
868 | * | |
869 | * Modules calling this function must call mmu_notifier_synchronize() in | |
870 | * their __exit functions to ensure the async work is completed. | |
871 | */ | |
1991722a | 872 | void mmu_notifier_put(struct mmu_notifier *subscription) |
2c7933f5 | 873 | { |
1991722a | 874 | struct mm_struct *mm = subscription->mm; |
2c7933f5 | 875 | |
984cfe4e | 876 | spin_lock(&mm->notifier_subscriptions->lock); |
1991722a | 877 | if (WARN_ON(!subscription->users) || --subscription->users) |
2c7933f5 | 878 | goto out_unlock; |
1991722a | 879 | hlist_del_init_rcu(&subscription->hlist); |
984cfe4e | 880 | spin_unlock(&mm->notifier_subscriptions->lock); |
2c7933f5 | 881 | |
1991722a | 882 | call_srcu(&srcu, &subscription->rcu, mmu_notifier_free_rcu); |
2c7933f5 JG |
883 | return; |
884 | ||
885 | out_unlock: | |
984cfe4e | 886 | spin_unlock(&mm->notifier_subscriptions->lock); |
2c7933f5 JG |
887 | } |
888 | EXPORT_SYMBOL_GPL(mmu_notifier_put); | |
889 | ||
99cb252f | 890 | static int __mmu_interval_notifier_insert( |
5292e24a | 891 | struct mmu_interval_notifier *interval_sub, struct mm_struct *mm, |
984cfe4e | 892 | struct mmu_notifier_subscriptions *subscriptions, unsigned long start, |
99cb252f JG |
893 | unsigned long length, const struct mmu_interval_notifier_ops *ops) |
894 | { | |
5292e24a JG |
895 | interval_sub->mm = mm; |
896 | interval_sub->ops = ops; | |
897 | RB_CLEAR_NODE(&interval_sub->interval_tree.rb); | |
898 | interval_sub->interval_tree.start = start; | |
99cb252f JG |
899 | /* |
900 | * Note that the representation of the intervals in the interval tree | |
901 | * considers the ending point as contained in the interval. | |
902 | */ | |
903 | if (length == 0 || | |
5292e24a JG |
904 | check_add_overflow(start, length - 1, |
905 | &interval_sub->interval_tree.last)) | |
99cb252f JG |
906 | return -EOVERFLOW; |
907 | ||
908 | /* Must call with a mmget() held */ | |
c9682d10 | 909 | if (WARN_ON(atomic_read(&mm->mm_users) <= 0)) |
99cb252f JG |
910 | return -EINVAL; |
911 | ||
912 | /* pairs with mmdrop in mmu_interval_notifier_remove() */ | |
913 | mmgrab(mm); | |
914 | ||
915 | /* | |
916 | * If some invalidate_range_start/end region is going on in parallel | |
917 | * we don't know what VA ranges are affected, so we must assume this | |
918 | * new range is included. | |
919 | * | |
920 | * If the itree is invalidating then we are not allowed to change | |
921 | * it. Retrying until invalidation is done is tricky due to the | |
922 | * possibility for live lock, instead defer the add to | |
923 | * mn_itree_inv_end() so this algorithm is deterministic. | |
924 | * | |
5292e24a | 925 | * In all cases the value for the interval_sub->invalidate_seq should be |
99cb252f JG |
926 | * odd, see mmu_interval_read_begin() |
927 | */ | |
984cfe4e JG |
928 | spin_lock(&subscriptions->lock); |
929 | if (subscriptions->active_invalidate_ranges) { | |
930 | if (mn_itree_is_invalidating(subscriptions)) | |
5292e24a | 931 | hlist_add_head(&interval_sub->deferred_item, |
984cfe4e | 932 | &subscriptions->deferred_list); |
99cb252f | 933 | else { |
984cfe4e | 934 | subscriptions->invalidate_seq |= 1; |
5292e24a | 935 | interval_tree_insert(&interval_sub->interval_tree, |
984cfe4e | 936 | &subscriptions->itree); |
99cb252f | 937 | } |
5292e24a | 938 | interval_sub->invalidate_seq = subscriptions->invalidate_seq; |
99cb252f | 939 | } else { |
984cfe4e | 940 | WARN_ON(mn_itree_is_invalidating(subscriptions)); |
99cb252f | 941 | /* |
5292e24a JG |
942 | * The starting seq for a subscription not under invalidation |
943 | * should be odd, not equal to the current invalidate_seq and | |
99cb252f JG |
944 | * invalidate_seq should not 'wrap' to the new seq any time |
945 | * soon. | |
946 | */ | |
5292e24a JG |
947 | interval_sub->invalidate_seq = |
948 | subscriptions->invalidate_seq - 1; | |
949 | interval_tree_insert(&interval_sub->interval_tree, | |
984cfe4e | 950 | &subscriptions->itree); |
99cb252f | 951 | } |
984cfe4e | 952 | spin_unlock(&subscriptions->lock); |
99cb252f JG |
953 | return 0; |
954 | } | |
955 | ||
956 | /** | |
957 | * mmu_interval_notifier_insert - Insert an interval notifier | |
5292e24a | 958 | * @interval_sub: Interval subscription to register |
99cb252f JG |
959 | * @start: Starting virtual address to monitor |
960 | * @length: Length of the range to monitor | |
d49653f3 KK |
961 | * @mm: mm_struct to attach to |
962 | * @ops: Interval notifier operations to be called on matching events | |
99cb252f JG |
963 | * |
964 | * This function subscribes the interval notifier for notifications from the | |
965 | * mm. Upon return the ops related to mmu_interval_notifier will be called | |
966 | * whenever an event that intersects with the given range occurs. | |
967 | * | |
968 | * Upon return the range_notifier may not be present in the interval tree yet. | |
969 | * The caller must use the normal interval notifier read flow via | |
970 | * mmu_interval_read_begin() to establish SPTEs for this range. | |
971 | */ | |
5292e24a | 972 | int mmu_interval_notifier_insert(struct mmu_interval_notifier *interval_sub, |
99cb252f JG |
973 | struct mm_struct *mm, unsigned long start, |
974 | unsigned long length, | |
975 | const struct mmu_interval_notifier_ops *ops) | |
976 | { | |
984cfe4e | 977 | struct mmu_notifier_subscriptions *subscriptions; |
99cb252f JG |
978 | int ret; |
979 | ||
da1c55f1 | 980 | might_lock(&mm->mmap_lock); |
99cb252f | 981 | |
984cfe4e JG |
982 | subscriptions = smp_load_acquire(&mm->notifier_subscriptions); |
983 | if (!subscriptions || !subscriptions->has_itree) { | |
99cb252f JG |
984 | ret = mmu_notifier_register(NULL, mm); |
985 | if (ret) | |
986 | return ret; | |
984cfe4e | 987 | subscriptions = mm->notifier_subscriptions; |
99cb252f | 988 | } |
5292e24a JG |
989 | return __mmu_interval_notifier_insert(interval_sub, mm, subscriptions, |
990 | start, length, ops); | |
99cb252f JG |
991 | } |
992 | EXPORT_SYMBOL_GPL(mmu_interval_notifier_insert); | |
993 | ||
994 | int mmu_interval_notifier_insert_locked( | |
5292e24a | 995 | struct mmu_interval_notifier *interval_sub, struct mm_struct *mm, |
99cb252f JG |
996 | unsigned long start, unsigned long length, |
997 | const struct mmu_interval_notifier_ops *ops) | |
998 | { | |
984cfe4e JG |
999 | struct mmu_notifier_subscriptions *subscriptions = |
1000 | mm->notifier_subscriptions; | |
99cb252f JG |
1001 | int ret; |
1002 | ||
42fc5414 | 1003 | mmap_assert_write_locked(mm); |
99cb252f | 1004 | |
984cfe4e | 1005 | if (!subscriptions || !subscriptions->has_itree) { |
99cb252f JG |
1006 | ret = __mmu_notifier_register(NULL, mm); |
1007 | if (ret) | |
1008 | return ret; | |
984cfe4e | 1009 | subscriptions = mm->notifier_subscriptions; |
99cb252f | 1010 | } |
5292e24a JG |
1011 | return __mmu_interval_notifier_insert(interval_sub, mm, subscriptions, |
1012 | start, length, ops); | |
99cb252f JG |
1013 | } |
1014 | EXPORT_SYMBOL_GPL(mmu_interval_notifier_insert_locked); | |
1015 | ||
1016 | /** | |
1017 | * mmu_interval_notifier_remove - Remove a interval notifier | |
5292e24a | 1018 | * @interval_sub: Interval subscription to unregister |
99cb252f JG |
1019 | * |
1020 | * This function must be paired with mmu_interval_notifier_insert(). It cannot | |
1021 | * be called from any ops callback. | |
1022 | * | |
1023 | * Once this returns ops callbacks are no longer running on other CPUs and | |
1024 | * will not be called in future. | |
1025 | */ | |
5292e24a | 1026 | void mmu_interval_notifier_remove(struct mmu_interval_notifier *interval_sub) |
99cb252f | 1027 | { |
5292e24a | 1028 | struct mm_struct *mm = interval_sub->mm; |
984cfe4e JG |
1029 | struct mmu_notifier_subscriptions *subscriptions = |
1030 | mm->notifier_subscriptions; | |
99cb252f JG |
1031 | unsigned long seq = 0; |
1032 | ||
1033 | might_sleep(); | |
1034 | ||
984cfe4e JG |
1035 | spin_lock(&subscriptions->lock); |
1036 | if (mn_itree_is_invalidating(subscriptions)) { | |
99cb252f JG |
1037 | /* |
1038 | * remove is being called after insert put this on the | |
1039 | * deferred list, but before the deferred list was processed. | |
1040 | */ | |
5292e24a JG |
1041 | if (RB_EMPTY_NODE(&interval_sub->interval_tree.rb)) { |
1042 | hlist_del(&interval_sub->deferred_item); | |
99cb252f | 1043 | } else { |
5292e24a | 1044 | hlist_add_head(&interval_sub->deferred_item, |
984cfe4e JG |
1045 | &subscriptions->deferred_list); |
1046 | seq = subscriptions->invalidate_seq; | |
99cb252f JG |
1047 | } |
1048 | } else { | |
5292e24a JG |
1049 | WARN_ON(RB_EMPTY_NODE(&interval_sub->interval_tree.rb)); |
1050 | interval_tree_remove(&interval_sub->interval_tree, | |
984cfe4e | 1051 | &subscriptions->itree); |
99cb252f | 1052 | } |
984cfe4e | 1053 | spin_unlock(&subscriptions->lock); |
99cb252f JG |
1054 | |
1055 | /* | |
1056 | * The possible sleep on progress in the invalidation requires the | |
1057 | * caller not hold any locks held by invalidation callbacks. | |
1058 | */ | |
1059 | lock_map_acquire(&__mmu_notifier_invalidate_range_start_map); | |
1060 | lock_map_release(&__mmu_notifier_invalidate_range_start_map); | |
1061 | if (seq) | |
984cfe4e JG |
1062 | wait_event(subscriptions->wq, |
1063 | READ_ONCE(subscriptions->invalidate_seq) != seq); | |
99cb252f JG |
1064 | |
1065 | /* pairs with mmgrab in mmu_interval_notifier_insert() */ | |
1066 | mmdrop(mm); | |
1067 | } | |
1068 | EXPORT_SYMBOL_GPL(mmu_interval_notifier_remove); | |
1069 | ||
2c7933f5 JG |
1070 | /** |
1071 | * mmu_notifier_synchronize - Ensure all mmu_notifiers are freed | |
1072 | * | |
1073 | * This function ensures that all outstanding async SRU work from | |
1074 | * mmu_notifier_put() is completed. After it returns any mmu_notifier_ops | |
1075 | * associated with an unused mmu_notifier will no longer be called. | |
1076 | * | |
1077 | * Before using the caller must ensure that all of its mmu_notifiers have been | |
1078 | * fully released via mmu_notifier_put(). | |
1079 | * | |
1080 | * Modules using the mmu_notifier_put() API should call this in their __exit | |
1081 | * function to avoid module unloading races. | |
1082 | */ | |
1083 | void mmu_notifier_synchronize(void) | |
1084 | { | |
1085 | synchronize_srcu(&srcu); | |
1086 | } | |
1087 | EXPORT_SYMBOL_GPL(mmu_notifier_synchronize); | |
1088 | ||
c6d23413 JG |
1089 | bool |
1090 | mmu_notifier_range_update_to_read_only(const struct mmu_notifier_range *range) | |
1091 | { | |
1092 | if (!range->vma || range->event != MMU_NOTIFY_PROTECTION_VMA) | |
1093 | return false; | |
1094 | /* Return true if the vma still have the read flag set. */ | |
1095 | return range->vma->vm_flags & VM_READ; | |
1096 | } | |
1097 | EXPORT_SYMBOL_GPL(mmu_notifier_range_update_to_read_only); |