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/srcu.h>
16 #include <linux/rcupdate.h>
17 #include <linux/sched.h>
18 #include <linux/sched/mm.h>
19 #include <linux/slab.h>
20
21 /* global SRCU for all MMs */
22 DEFINE_STATIC_SRCU(srcu);
23
24 #ifdef CONFIG_LOCKDEP
25 struct lockdep_map __mmu_notifier_invalidate_range_start_map = {
26 .name = "mmu_notifier_invalidate_range_start"
27 };
28 #endif
29
30 /*
31 * This function can't run concurrently against mmu_notifier_register
32 * because mm->mm_users > 0 during mmu_notifier_register and exit_mmap
33 * runs with mm_users == 0. Other tasks may still invoke mmu notifiers
34 * in parallel despite there being no task using this mm any more,
35 * through the vmas outside of the exit_mmap context, such as with
36 * vmtruncate. This serializes against mmu_notifier_unregister with
37 * the mmu_notifier_mm->lock in addition to SRCU and it serializes
38 * against the other mmu notifiers with SRCU. struct mmu_notifier_mm
39 * can't go away from under us as exit_mmap holds an mm_count pin
40 * itself.
41 */
42 void __mmu_notifier_release(struct mm_struct *mm)
43 {
44 struct mmu_notifier *mn;
45 int id;
46
47 /*
48 * SRCU here will block mmu_notifier_unregister until
49 * ->release returns.
50 */
51 id = srcu_read_lock(&srcu);
52 hlist_for_each_entry_rcu(mn, &mm->mmu_notifier_mm->list, hlist)
53 /*
54 * If ->release runs before mmu_notifier_unregister it must be
55 * handled, as it's the only way for the driver to flush all
56 * existing sptes and stop the driver from establishing any more
57 * sptes before all the pages in the mm are freed.
58 */
59 if (mn->ops->release)
60 mn->ops->release(mn, mm);
61
62 spin_lock(&mm->mmu_notifier_mm->lock);
63 while (unlikely(!hlist_empty(&mm->mmu_notifier_mm->list))) {
64 mn = hlist_entry(mm->mmu_notifier_mm->list.first,
65 struct mmu_notifier,
66 hlist);
67 /*
68 * We arrived before mmu_notifier_unregister so
69 * mmu_notifier_unregister will do nothing other than to wait
70 * for ->release to finish and for mmu_notifier_unregister to
71 * return.
72 */
73 hlist_del_init_rcu(&mn->hlist);
74 }
75 spin_unlock(&mm->mmu_notifier_mm->lock);
76 srcu_read_unlock(&srcu, id);
77
78 /*
79 * synchronize_srcu here prevents mmu_notifier_release from returning to
80 * exit_mmap (which would proceed with freeing all pages in the mm)
81 * until the ->release method returns, if it was invoked by
82 * mmu_notifier_unregister.
83 *
84 * The mmu_notifier_mm can't go away from under us because one mm_count
85 * is held by exit_mmap.
86 */
87 synchronize_srcu(&srcu);
88 }
89
90 /*
91 * If no young bitflag is supported by the hardware, ->clear_flush_young can
92 * unmap the address and return 1 or 0 depending if the mapping previously
93 * existed or not.
94 */
95 int __mmu_notifier_clear_flush_young(struct mm_struct *mm,
96 unsigned long start,
97 unsigned long end)
98 {
99 struct mmu_notifier *mn;
100 int young = 0, id;
101
102 id = srcu_read_lock(&srcu);
103 hlist_for_each_entry_rcu(mn, &mm->mmu_notifier_mm->list, hlist) {
104 if (mn->ops->clear_flush_young)
105 young |= mn->ops->clear_flush_young(mn, mm, start, end);
106 }
107 srcu_read_unlock(&srcu, id);
108
109 return young;
110 }
111
112 int __mmu_notifier_clear_young(struct mm_struct *mm,
113 unsigned long start,
114 unsigned long end)
115 {
116 struct mmu_notifier *mn;
117 int young = 0, id;
118
119 id = srcu_read_lock(&srcu);
120 hlist_for_each_entry_rcu(mn, &mm->mmu_notifier_mm->list, hlist) {
121 if (mn->ops->clear_young)
122 young |= mn->ops->clear_young(mn, mm, start, end);
123 }
124 srcu_read_unlock(&srcu, id);
125
126 return young;
127 }
128
129 int __mmu_notifier_test_young(struct mm_struct *mm,
130 unsigned long address)
131 {
132 struct mmu_notifier *mn;
133 int young = 0, id;
134
135 id = srcu_read_lock(&srcu);
136 hlist_for_each_entry_rcu(mn, &mm->mmu_notifier_mm->list, hlist) {
137 if (mn->ops->test_young) {
138 young = mn->ops->test_young(mn, mm, address);
139 if (young)
140 break;
141 }
142 }
143 srcu_read_unlock(&srcu, id);
144
145 return young;
146 }
147
148 void __mmu_notifier_change_pte(struct mm_struct *mm, unsigned long address,
149 pte_t pte)
150 {
151 struct mmu_notifier *mn;
152 int id;
153
154 id = srcu_read_lock(&srcu);
155 hlist_for_each_entry_rcu(mn, &mm->mmu_notifier_mm->list, hlist) {
156 if (mn->ops->change_pte)
157 mn->ops->change_pte(mn, mm, address, pte);
158 }
159 srcu_read_unlock(&srcu, id);
160 }
161
162 int __mmu_notifier_invalidate_range_start(struct mmu_notifier_range *range)
163 {
164 struct mmu_notifier *mn;
165 int ret = 0;
166 int id;
167
168 id = srcu_read_lock(&srcu);
169 hlist_for_each_entry_rcu(mn, &range->mm->mmu_notifier_mm->list, hlist) {
170 if (mn->ops->invalidate_range_start) {
171 int _ret;
172
173 if (!mmu_notifier_range_blockable(range))
174 non_block_start();
175 _ret = mn->ops->invalidate_range_start(mn, range);
176 if (!mmu_notifier_range_blockable(range))
177 non_block_end();
178 if (_ret) {
179 pr_info("%pS callback failed with %d in %sblockable context.\n",
180 mn->ops->invalidate_range_start, _ret,
181 !mmu_notifier_range_blockable(range) ? "non-" : "");
182 WARN_ON(mmu_notifier_range_blockable(range) ||
183 _ret != -EAGAIN);
184 ret = _ret;
185 }
186 }
187 }
188 srcu_read_unlock(&srcu, id);
189
190 return ret;
191 }
192
193 void __mmu_notifier_invalidate_range_end(struct mmu_notifier_range *range,
194 bool only_end)
195 {
196 struct mmu_notifier *mn;
197 int id;
198
199 lock_map_acquire(&__mmu_notifier_invalidate_range_start_map);
200 id = srcu_read_lock(&srcu);
201 hlist_for_each_entry_rcu(mn, &range->mm->mmu_notifier_mm->list, hlist) {
202 /*
203 * Call invalidate_range here too to avoid the need for the
204 * subsystem of having to register an invalidate_range_end
205 * call-back when there is invalidate_range already. Usually a
206 * subsystem registers either invalidate_range_start()/end() or
207 * invalidate_range(), so this will be no additional overhead
208 * (besides the pointer check).
209 *
210 * We skip call to invalidate_range() if we know it is safe ie
211 * call site use mmu_notifier_invalidate_range_only_end() which
212 * is safe to do when we know that a call to invalidate_range()
213 * already happen under page table lock.
214 */
215 if (!only_end && mn->ops->invalidate_range)
216 mn->ops->invalidate_range(mn, range->mm,
217 range->start,
218 range->end);
219 if (mn->ops->invalidate_range_end) {
220 if (!mmu_notifier_range_blockable(range))
221 non_block_start();
222 mn->ops->invalidate_range_end(mn, range);
223 if (!mmu_notifier_range_blockable(range))
224 non_block_end();
225 }
226 }
227 srcu_read_unlock(&srcu, id);
228 lock_map_release(&__mmu_notifier_invalidate_range_start_map);
229 }
230
231 void __mmu_notifier_invalidate_range(struct mm_struct *mm,
232 unsigned long start, unsigned long end)
233 {
234 struct mmu_notifier *mn;
235 int id;
236
237 id = srcu_read_lock(&srcu);
238 hlist_for_each_entry_rcu(mn, &mm->mmu_notifier_mm->list, hlist) {
239 if (mn->ops->invalidate_range)
240 mn->ops->invalidate_range(mn, mm, start, end);
241 }
242 srcu_read_unlock(&srcu, id);
243 }
244
245 /*
246 * Same as mmu_notifier_register but here the caller must hold the
247 * mmap_sem in write mode.
248 */
249 int __mmu_notifier_register(struct mmu_notifier *mn, struct mm_struct *mm)
250 {
251 struct mmu_notifier_mm *mmu_notifier_mm = NULL;
252 int ret;
253
254 lockdep_assert_held_write(&mm->mmap_sem);
255 BUG_ON(atomic_read(&mm->mm_users) <= 0);
256
257 if (IS_ENABLED(CONFIG_LOCKDEP)) {
258 fs_reclaim_acquire(GFP_KERNEL);
259 lock_map_acquire(&__mmu_notifier_invalidate_range_start_map);
260 lock_map_release(&__mmu_notifier_invalidate_range_start_map);
261 fs_reclaim_release(GFP_KERNEL);
262 }
263
264 mn->mm = mm;
265 mn->users = 1;
266
267 if (!mm->mmu_notifier_mm) {
268 /*
269 * kmalloc cannot be called under mm_take_all_locks(), but we
270 * know that mm->mmu_notifier_mm can't change while we hold
271 * the write side of the mmap_sem.
272 */
273 mmu_notifier_mm =
274 kmalloc(sizeof(struct mmu_notifier_mm), GFP_KERNEL);
275 if (!mmu_notifier_mm)
276 return -ENOMEM;
277
278 INIT_HLIST_HEAD(&mmu_notifier_mm->list);
279 spin_lock_init(&mmu_notifier_mm->lock);
280 }
281
282 ret = mm_take_all_locks(mm);
283 if (unlikely(ret))
284 goto out_clean;
285
286 /* Pairs with the mmdrop in mmu_notifier_unregister_* */
287 mmgrab(mm);
288
289 /*
290 * Serialize the update against mmu_notifier_unregister. A
291 * side note: mmu_notifier_release can't run concurrently with
292 * us because we hold the mm_users pin (either implicitly as
293 * current->mm or explicitly with get_task_mm() or similar).
294 * We can't race against any other mmu notifier method either
295 * thanks to mm_take_all_locks().
296 */
297 if (mmu_notifier_mm)
298 mm->mmu_notifier_mm = mmu_notifier_mm;
299
300 spin_lock(&mm->mmu_notifier_mm->lock);
301 hlist_add_head_rcu(&mn->hlist, &mm->mmu_notifier_mm->list);
302 spin_unlock(&mm->mmu_notifier_mm->lock);
303
304 mm_drop_all_locks(mm);
305 BUG_ON(atomic_read(&mm->mm_users) <= 0);
306 return 0;
307
308 out_clean:
309 kfree(mmu_notifier_mm);
310 return ret;
311 }
312 EXPORT_SYMBOL_GPL(__mmu_notifier_register);
313
314 /**
315 * mmu_notifier_register - Register a notifier on a mm
316 * @mn: The notifier to attach
317 * @mm: The mm to attach the notifier to
318 *
319 * Must not hold mmap_sem nor any other VM related lock when calling
320 * this registration function. Must also ensure mm_users can't go down
321 * to zero while this runs to avoid races with mmu_notifier_release,
322 * so mm has to be current->mm or the mm should be pinned safely such
323 * as with get_task_mm(). If the mm is not current->mm, the mm_users
324 * pin should be released by calling mmput after mmu_notifier_register
325 * returns.
326 *
327 * mmu_notifier_unregister() or mmu_notifier_put() must be always called to
328 * unregister the notifier.
329 *
330 * While the caller has a mmu_notifier get the mn->mm pointer will remain
331 * valid, and can be converted to an active mm pointer via mmget_not_zero().
332 */
333 int mmu_notifier_register(struct mmu_notifier *mn, struct mm_struct *mm)
334 {
335 int ret;
336
337 down_write(&mm->mmap_sem);
338 ret = __mmu_notifier_register(mn, mm);
339 up_write(&mm->mmap_sem);
340 return ret;
341 }
342 EXPORT_SYMBOL_GPL(mmu_notifier_register);
343
344 static struct mmu_notifier *
345 find_get_mmu_notifier(struct mm_struct *mm, const struct mmu_notifier_ops *ops)
346 {
347 struct mmu_notifier *mn;
348
349 spin_lock(&mm->mmu_notifier_mm->lock);
350 hlist_for_each_entry_rcu (mn, &mm->mmu_notifier_mm->list, hlist) {
351 if (mn->ops != ops)
352 continue;
353
354 if (likely(mn->users != UINT_MAX))
355 mn->users++;
356 else
357 mn = ERR_PTR(-EOVERFLOW);
358 spin_unlock(&mm->mmu_notifier_mm->lock);
359 return mn;
360 }
361 spin_unlock(&mm->mmu_notifier_mm->lock);
362 return NULL;
363 }
364
365 /**
366 * mmu_notifier_get_locked - Return the single struct mmu_notifier for
367 * the mm & ops
368 * @ops: The operations struct being subscribe with
369 * @mm : The mm to attach notifiers too
370 *
371 * This function either allocates a new mmu_notifier via
372 * ops->alloc_notifier(), or returns an already existing notifier on the
373 * list. The value of the ops pointer is used to determine when two notifiers
374 * are the same.
375 *
376 * Each call to mmu_notifier_get() must be paired with a call to
377 * mmu_notifier_put(). The caller must hold the write side of mm->mmap_sem.
378 *
379 * While the caller has a mmu_notifier get the mm pointer will remain valid,
380 * and can be converted to an active mm pointer via mmget_not_zero().
381 */
382 struct mmu_notifier *mmu_notifier_get_locked(const struct mmu_notifier_ops *ops,
383 struct mm_struct *mm)
384 {
385 struct mmu_notifier *mn;
386 int ret;
387
388 lockdep_assert_held_write(&mm->mmap_sem);
389
390 if (mm->mmu_notifier_mm) {
391 mn = find_get_mmu_notifier(mm, ops);
392 if (mn)
393 return mn;
394 }
395
396 mn = ops->alloc_notifier(mm);
397 if (IS_ERR(mn))
398 return mn;
399 mn->ops = ops;
400 ret = __mmu_notifier_register(mn, mm);
401 if (ret)
402 goto out_free;
403 return mn;
404 out_free:
405 mn->ops->free_notifier(mn);
406 return ERR_PTR(ret);
407 }
408 EXPORT_SYMBOL_GPL(mmu_notifier_get_locked);
409
410 /* this is called after the last mmu_notifier_unregister() returned */
411 void __mmu_notifier_mm_destroy(struct mm_struct *mm)
412 {
413 BUG_ON(!hlist_empty(&mm->mmu_notifier_mm->list));
414 kfree(mm->mmu_notifier_mm);
415 mm->mmu_notifier_mm = LIST_POISON1; /* debug */
416 }
417
418 /*
419 * This releases the mm_count pin automatically and frees the mm
420 * structure if it was the last user of it. It serializes against
421 * running mmu notifiers with SRCU and against mmu_notifier_unregister
422 * with the unregister lock + SRCU. All sptes must be dropped before
423 * calling mmu_notifier_unregister. ->release or any other notifier
424 * method may be invoked concurrently with mmu_notifier_unregister,
425 * and only after mmu_notifier_unregister returned we're guaranteed
426 * that ->release or any other method can't run anymore.
427 */
428 void mmu_notifier_unregister(struct mmu_notifier *mn, struct mm_struct *mm)
429 {
430 BUG_ON(atomic_read(&mm->mm_count) <= 0);
431
432 if (!hlist_unhashed(&mn->hlist)) {
433 /*
434 * SRCU here will force exit_mmap to wait for ->release to
435 * finish before freeing the pages.
436 */
437 int id;
438
439 id = srcu_read_lock(&srcu);
440 /*
441 * exit_mmap will block in mmu_notifier_release to guarantee
442 * that ->release is called before freeing the pages.
443 */
444 if (mn->ops->release)
445 mn->ops->release(mn, mm);
446 srcu_read_unlock(&srcu, id);
447
448 spin_lock(&mm->mmu_notifier_mm->lock);
449 /*
450 * Can not use list_del_rcu() since __mmu_notifier_release
451 * can delete it before we hold the lock.
452 */
453 hlist_del_init_rcu(&mn->hlist);
454 spin_unlock(&mm->mmu_notifier_mm->lock);
455 }
456
457 /*
458 * Wait for any running method to finish, of course including
459 * ->release if it was run by mmu_notifier_release instead of us.
460 */
461 synchronize_srcu(&srcu);
462
463 BUG_ON(atomic_read(&mm->mm_count) <= 0);
464
465 mmdrop(mm);
466 }
467 EXPORT_SYMBOL_GPL(mmu_notifier_unregister);
468
469 static void mmu_notifier_free_rcu(struct rcu_head *rcu)
470 {
471 struct mmu_notifier *mn = container_of(rcu, struct mmu_notifier, rcu);
472 struct mm_struct *mm = mn->mm;
473
474 mn->ops->free_notifier(mn);
475 /* Pairs with the get in __mmu_notifier_register() */
476 mmdrop(mm);
477 }
478
479 /**
480 * mmu_notifier_put - Release the reference on the notifier
481 * @mn: The notifier to act on
482 *
483 * This function must be paired with each mmu_notifier_get(), it releases the
484 * reference obtained by the get. If this is the last reference then process
485 * to free the notifier will be run asynchronously.
486 *
487 * Unlike mmu_notifier_unregister() the get/put flow only calls ops->release
488 * when the mm_struct is destroyed. Instead free_notifier is always called to
489 * release any resources held by the user.
490 *
491 * As ops->release is not guaranteed to be called, the user must ensure that
492 * all sptes are dropped, and no new sptes can be established before
493 * mmu_notifier_put() is called.
494 *
495 * This function can be called from the ops->release callback, however the
496 * caller must still ensure it is called pairwise with mmu_notifier_get().
497 *
498 * Modules calling this function must call mmu_notifier_synchronize() in
499 * their __exit functions to ensure the async work is completed.
500 */
501 void mmu_notifier_put(struct mmu_notifier *mn)
502 {
503 struct mm_struct *mm = mn->mm;
504
505 spin_lock(&mm->mmu_notifier_mm->lock);
506 if (WARN_ON(!mn->users) || --mn->users)
507 goto out_unlock;
508 hlist_del_init_rcu(&mn->hlist);
509 spin_unlock(&mm->mmu_notifier_mm->lock);
510
511 call_srcu(&srcu, &mn->rcu, mmu_notifier_free_rcu);
512 return;
513
514 out_unlock:
515 spin_unlock(&mm->mmu_notifier_mm->lock);
516 }
517 EXPORT_SYMBOL_GPL(mmu_notifier_put);
518
519 /**
520 * mmu_notifier_synchronize - Ensure all mmu_notifiers are freed
521 *
522 * This function ensures that all outstanding async SRU work from
523 * mmu_notifier_put() is completed. After it returns any mmu_notifier_ops
524 * associated with an unused mmu_notifier will no longer be called.
525 *
526 * Before using the caller must ensure that all of its mmu_notifiers have been
527 * fully released via mmu_notifier_put().
528 *
529 * Modules using the mmu_notifier_put() API should call this in their __exit
530 * function to avoid module unloading races.
531 */
532 void mmu_notifier_synchronize(void)
533 {
534 synchronize_srcu(&srcu);
535 }
536 EXPORT_SYMBOL_GPL(mmu_notifier_synchronize);
537
538 bool
539 mmu_notifier_range_update_to_read_only(const struct mmu_notifier_range *range)
540 {
541 if (!range->vma || range->event != MMU_NOTIFY_PROTECTION_VMA)
542 return false;
543 /* Return true if the vma still have the read flag set. */
544 return range->vma->vm_flags & VM_READ;
545 }
546 EXPORT_SYMBOL_GPL(mmu_notifier_range_update_to_read_only);