Loading...
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 */
22DEFINE_STATIC_SRCU(srcu);
23
24#ifdef CONFIG_LOCKDEP
25struct 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 */
42void __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 */
95int __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
112int __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
129int __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
148void __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
162int __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
193void __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
231void __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 */
249int __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
308out_clean:
309 kfree(mmu_notifier_mm);
310 return ret;
311}
312EXPORT_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 */
333int 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}
342EXPORT_SYMBOL_GPL(mmu_notifier_register);
343
344static struct mmu_notifier *
345find_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 */
382struct 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;
404out_free:
405 mn->ops->free_notifier(mn);
406 return ERR_PTR(ret);
407}
408EXPORT_SYMBOL_GPL(mmu_notifier_get_locked);
409
410/* this is called after the last mmu_notifier_unregister() returned */
411void __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 */
428void 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}
467EXPORT_SYMBOL_GPL(mmu_notifier_unregister);
468
469static 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 */
501void 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
514out_unlock:
515 spin_unlock(&mm->mmu_notifier_mm->lock);
516}
517EXPORT_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 */
532void mmu_notifier_synchronize(void)
533{
534 synchronize_srcu(&srcu);
535}
536EXPORT_SYMBOL_GPL(mmu_notifier_synchronize);
537
538bool
539mmu_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}
546EXPORT_SYMBOL_GPL(mmu_notifier_range_update_to_read_only);
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 */
23DEFINE_STATIC_SRCU(srcu);
24
25#ifdef CONFIG_LOCKDEP
26struct 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 */
37struct 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 */
86static bool
87mn_itree_is_invalidating(struct mmu_notifier_subscriptions *subscriptions)
88{
89 lockdep_assert_held(&subscriptions->lock);
90 return subscriptions->invalidate_seq & 1;
91}
92
93static struct mmu_interval_notifier *
94mn_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
116static struct mmu_interval_notifier *
117mn_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
129static 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 */
185unsigned long
186mmu_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_tree_invalidate_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_range():
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}
259EXPORT_SYMBOL_GPL(mmu_interval_read_begin);
260
261static 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 */
299static 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
348void __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 */
365int __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
385int __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
405int __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
427void __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
444static 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
466out_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
475static 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
536int __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
552static void
553mn_hlist_invalidate_end(struct mmu_notifier_subscriptions *subscriptions,
554 struct mmu_notifier_range *range, bool only_end)
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 /*
563 * Call invalidate_range here too to avoid the need for the
564 * subsystem of having to register an invalidate_range_end
565 * call-back when there is invalidate_range already. Usually a
566 * subsystem registers either invalidate_range_start()/end() or
567 * invalidate_range(), so this will be no additional overhead
568 * (besides the pointer check).
569 *
570 * We skip call to invalidate_range() if we know it is safe ie
571 * call site use mmu_notifier_invalidate_range_only_end() which
572 * is safe to do when we know that a call to invalidate_range()
573 * already happen under page table lock.
574 */
575 if (!only_end && subscription->ops->invalidate_range)
576 subscription->ops->invalidate_range(subscription,
577 range->mm,
578 range->start,
579 range->end);
580 if (subscription->ops->invalidate_range_end) {
581 if (!mmu_notifier_range_blockable(range))
582 non_block_start();
583 subscription->ops->invalidate_range_end(subscription,
584 range);
585 if (!mmu_notifier_range_blockable(range))
586 non_block_end();
587 }
588 }
589 srcu_read_unlock(&srcu, id);
590}
591
592void __mmu_notifier_invalidate_range_end(struct mmu_notifier_range *range,
593 bool only_end)
594{
595 struct mmu_notifier_subscriptions *subscriptions =
596 range->mm->notifier_subscriptions;
597
598 lock_map_acquire(&__mmu_notifier_invalidate_range_start_map);
599 if (subscriptions->has_itree)
600 mn_itree_inv_end(subscriptions);
601
602 if (!hlist_empty(&subscriptions->list))
603 mn_hlist_invalidate_end(subscriptions, range, only_end);
604 lock_map_release(&__mmu_notifier_invalidate_range_start_map);
605}
606
607void __mmu_notifier_invalidate_range(struct mm_struct *mm,
608 unsigned long start, unsigned long end)
609{
610 struct mmu_notifier *subscription;
611 int id;
612
613 id = srcu_read_lock(&srcu);
614 hlist_for_each_entry_rcu(subscription,
615 &mm->notifier_subscriptions->list, hlist,
616 srcu_read_lock_held(&srcu)) {
617 if (subscription->ops->invalidate_range)
618 subscription->ops->invalidate_range(subscription, mm,
619 start, end);
620 }
621 srcu_read_unlock(&srcu, id);
622}
623
624/*
625 * Same as mmu_notifier_register but here the caller must hold the mmap_lock in
626 * write mode. A NULL mn signals the notifier is being registered for itree
627 * mode.
628 */
629int __mmu_notifier_register(struct mmu_notifier *subscription,
630 struct mm_struct *mm)
631{
632 struct mmu_notifier_subscriptions *subscriptions = NULL;
633 int ret;
634
635 mmap_assert_write_locked(mm);
636 BUG_ON(atomic_read(&mm->mm_users) <= 0);
637
638 if (!mm->notifier_subscriptions) {
639 /*
640 * kmalloc cannot be called under mm_take_all_locks(), but we
641 * know that mm->notifier_subscriptions can't change while we
642 * hold the write side of the mmap_lock.
643 */
644 subscriptions = kzalloc(
645 sizeof(struct mmu_notifier_subscriptions), GFP_KERNEL);
646 if (!subscriptions)
647 return -ENOMEM;
648
649 INIT_HLIST_HEAD(&subscriptions->list);
650 spin_lock_init(&subscriptions->lock);
651 subscriptions->invalidate_seq = 2;
652 subscriptions->itree = RB_ROOT_CACHED;
653 init_waitqueue_head(&subscriptions->wq);
654 INIT_HLIST_HEAD(&subscriptions->deferred_list);
655 }
656
657 ret = mm_take_all_locks(mm);
658 if (unlikely(ret))
659 goto out_clean;
660
661 /*
662 * Serialize the update against mmu_notifier_unregister. A
663 * side note: mmu_notifier_release can't run concurrently with
664 * us because we hold the mm_users pin (either implicitly as
665 * current->mm or explicitly with get_task_mm() or similar).
666 * We can't race against any other mmu notifier method either
667 * thanks to mm_take_all_locks().
668 *
669 * release semantics on the initialization of the
670 * mmu_notifier_subscriptions's contents are provided for unlocked
671 * readers. acquire can only be used while holding the mmgrab or
672 * mmget, and is safe because once created the
673 * mmu_notifier_subscriptions is not freed until the mm is destroyed.
674 * As above, users holding the mmap_lock or one of the
675 * mm_take_all_locks() do not need to use acquire semantics.
676 */
677 if (subscriptions)
678 smp_store_release(&mm->notifier_subscriptions, subscriptions);
679
680 if (subscription) {
681 /* Pairs with the mmdrop in mmu_notifier_unregister_* */
682 mmgrab(mm);
683 subscription->mm = mm;
684 subscription->users = 1;
685
686 spin_lock(&mm->notifier_subscriptions->lock);
687 hlist_add_head_rcu(&subscription->hlist,
688 &mm->notifier_subscriptions->list);
689 spin_unlock(&mm->notifier_subscriptions->lock);
690 } else
691 mm->notifier_subscriptions->has_itree = true;
692
693 mm_drop_all_locks(mm);
694 BUG_ON(atomic_read(&mm->mm_users) <= 0);
695 return 0;
696
697out_clean:
698 kfree(subscriptions);
699 return ret;
700}
701EXPORT_SYMBOL_GPL(__mmu_notifier_register);
702
703/**
704 * mmu_notifier_register - Register a notifier on a mm
705 * @subscription: The notifier to attach
706 * @mm: The mm to attach the notifier to
707 *
708 * Must not hold mmap_lock nor any other VM related lock when calling
709 * this registration function. Must also ensure mm_users can't go down
710 * to zero while this runs to avoid races with mmu_notifier_release,
711 * so mm has to be current->mm or the mm should be pinned safely such
712 * as with get_task_mm(). If the mm is not current->mm, the mm_users
713 * pin should be released by calling mmput after mmu_notifier_register
714 * returns.
715 *
716 * mmu_notifier_unregister() or mmu_notifier_put() must be always called to
717 * unregister the notifier.
718 *
719 * While the caller has a mmu_notifier get the subscription->mm pointer will remain
720 * valid, and can be converted to an active mm pointer via mmget_not_zero().
721 */
722int mmu_notifier_register(struct mmu_notifier *subscription,
723 struct mm_struct *mm)
724{
725 int ret;
726
727 mmap_write_lock(mm);
728 ret = __mmu_notifier_register(subscription, mm);
729 mmap_write_unlock(mm);
730 return ret;
731}
732EXPORT_SYMBOL_GPL(mmu_notifier_register);
733
734static struct mmu_notifier *
735find_get_mmu_notifier(struct mm_struct *mm, const struct mmu_notifier_ops *ops)
736{
737 struct mmu_notifier *subscription;
738
739 spin_lock(&mm->notifier_subscriptions->lock);
740 hlist_for_each_entry_rcu(subscription,
741 &mm->notifier_subscriptions->list, hlist,
742 lockdep_is_held(&mm->notifier_subscriptions->lock)) {
743 if (subscription->ops != ops)
744 continue;
745
746 if (likely(subscription->users != UINT_MAX))
747 subscription->users++;
748 else
749 subscription = ERR_PTR(-EOVERFLOW);
750 spin_unlock(&mm->notifier_subscriptions->lock);
751 return subscription;
752 }
753 spin_unlock(&mm->notifier_subscriptions->lock);
754 return NULL;
755}
756
757/**
758 * mmu_notifier_get_locked - Return the single struct mmu_notifier for
759 * the mm & ops
760 * @ops: The operations struct being subscribe with
761 * @mm : The mm to attach notifiers too
762 *
763 * This function either allocates a new mmu_notifier via
764 * ops->alloc_notifier(), or returns an already existing notifier on the
765 * list. The value of the ops pointer is used to determine when two notifiers
766 * are the same.
767 *
768 * Each call to mmu_notifier_get() must be paired with a call to
769 * mmu_notifier_put(). The caller must hold the write side of mm->mmap_lock.
770 *
771 * While the caller has a mmu_notifier get the mm pointer will remain valid,
772 * and can be converted to an active mm pointer via mmget_not_zero().
773 */
774struct mmu_notifier *mmu_notifier_get_locked(const struct mmu_notifier_ops *ops,
775 struct mm_struct *mm)
776{
777 struct mmu_notifier *subscription;
778 int ret;
779
780 mmap_assert_write_locked(mm);
781
782 if (mm->notifier_subscriptions) {
783 subscription = find_get_mmu_notifier(mm, ops);
784 if (subscription)
785 return subscription;
786 }
787
788 subscription = ops->alloc_notifier(mm);
789 if (IS_ERR(subscription))
790 return subscription;
791 subscription->ops = ops;
792 ret = __mmu_notifier_register(subscription, mm);
793 if (ret)
794 goto out_free;
795 return subscription;
796out_free:
797 subscription->ops->free_notifier(subscription);
798 return ERR_PTR(ret);
799}
800EXPORT_SYMBOL_GPL(mmu_notifier_get_locked);
801
802/* this is called after the last mmu_notifier_unregister() returned */
803void __mmu_notifier_subscriptions_destroy(struct mm_struct *mm)
804{
805 BUG_ON(!hlist_empty(&mm->notifier_subscriptions->list));
806 kfree(mm->notifier_subscriptions);
807 mm->notifier_subscriptions = LIST_POISON1; /* debug */
808}
809
810/*
811 * This releases the mm_count pin automatically and frees the mm
812 * structure if it was the last user of it. It serializes against
813 * running mmu notifiers with SRCU and against mmu_notifier_unregister
814 * with the unregister lock + SRCU. All sptes must be dropped before
815 * calling mmu_notifier_unregister. ->release or any other notifier
816 * method may be invoked concurrently with mmu_notifier_unregister,
817 * and only after mmu_notifier_unregister returned we're guaranteed
818 * that ->release or any other method can't run anymore.
819 */
820void mmu_notifier_unregister(struct mmu_notifier *subscription,
821 struct mm_struct *mm)
822{
823 BUG_ON(atomic_read(&mm->mm_count) <= 0);
824
825 if (!hlist_unhashed(&subscription->hlist)) {
826 /*
827 * SRCU here will force exit_mmap to wait for ->release to
828 * finish before freeing the pages.
829 */
830 int id;
831
832 id = srcu_read_lock(&srcu);
833 /*
834 * exit_mmap will block in mmu_notifier_release to guarantee
835 * that ->release is called before freeing the pages.
836 */
837 if (subscription->ops->release)
838 subscription->ops->release(subscription, mm);
839 srcu_read_unlock(&srcu, id);
840
841 spin_lock(&mm->notifier_subscriptions->lock);
842 /*
843 * Can not use list_del_rcu() since __mmu_notifier_release
844 * can delete it before we hold the lock.
845 */
846 hlist_del_init_rcu(&subscription->hlist);
847 spin_unlock(&mm->notifier_subscriptions->lock);
848 }
849
850 /*
851 * Wait for any running method to finish, of course including
852 * ->release if it was run by mmu_notifier_release instead of us.
853 */
854 synchronize_srcu(&srcu);
855
856 BUG_ON(atomic_read(&mm->mm_count) <= 0);
857
858 mmdrop(mm);
859}
860EXPORT_SYMBOL_GPL(mmu_notifier_unregister);
861
862static void mmu_notifier_free_rcu(struct rcu_head *rcu)
863{
864 struct mmu_notifier *subscription =
865 container_of(rcu, struct mmu_notifier, rcu);
866 struct mm_struct *mm = subscription->mm;
867
868 subscription->ops->free_notifier(subscription);
869 /* Pairs with the get in __mmu_notifier_register() */
870 mmdrop(mm);
871}
872
873/**
874 * mmu_notifier_put - Release the reference on the notifier
875 * @subscription: The notifier to act on
876 *
877 * This function must be paired with each mmu_notifier_get(), it releases the
878 * reference obtained by the get. If this is the last reference then process
879 * to free the notifier will be run asynchronously.
880 *
881 * Unlike mmu_notifier_unregister() the get/put flow only calls ops->release
882 * when the mm_struct is destroyed. Instead free_notifier is always called to
883 * release any resources held by the user.
884 *
885 * As ops->release is not guaranteed to be called, the user must ensure that
886 * all sptes are dropped, and no new sptes can be established before
887 * mmu_notifier_put() is called.
888 *
889 * This function can be called from the ops->release callback, however the
890 * caller must still ensure it is called pairwise with mmu_notifier_get().
891 *
892 * Modules calling this function must call mmu_notifier_synchronize() in
893 * their __exit functions to ensure the async work is completed.
894 */
895void mmu_notifier_put(struct mmu_notifier *subscription)
896{
897 struct mm_struct *mm = subscription->mm;
898
899 spin_lock(&mm->notifier_subscriptions->lock);
900 if (WARN_ON(!subscription->users) || --subscription->users)
901 goto out_unlock;
902 hlist_del_init_rcu(&subscription->hlist);
903 spin_unlock(&mm->notifier_subscriptions->lock);
904
905 call_srcu(&srcu, &subscription->rcu, mmu_notifier_free_rcu);
906 return;
907
908out_unlock:
909 spin_unlock(&mm->notifier_subscriptions->lock);
910}
911EXPORT_SYMBOL_GPL(mmu_notifier_put);
912
913static int __mmu_interval_notifier_insert(
914 struct mmu_interval_notifier *interval_sub, struct mm_struct *mm,
915 struct mmu_notifier_subscriptions *subscriptions, unsigned long start,
916 unsigned long length, const struct mmu_interval_notifier_ops *ops)
917{
918 interval_sub->mm = mm;
919 interval_sub->ops = ops;
920 RB_CLEAR_NODE(&interval_sub->interval_tree.rb);
921 interval_sub->interval_tree.start = start;
922 /*
923 * Note that the representation of the intervals in the interval tree
924 * considers the ending point as contained in the interval.
925 */
926 if (length == 0 ||
927 check_add_overflow(start, length - 1,
928 &interval_sub->interval_tree.last))
929 return -EOVERFLOW;
930
931 /* Must call with a mmget() held */
932 if (WARN_ON(atomic_read(&mm->mm_users) <= 0))
933 return -EINVAL;
934
935 /* pairs with mmdrop in mmu_interval_notifier_remove() */
936 mmgrab(mm);
937
938 /*
939 * If some invalidate_range_start/end region is going on in parallel
940 * we don't know what VA ranges are affected, so we must assume this
941 * new range is included.
942 *
943 * If the itree is invalidating then we are not allowed to change
944 * it. Retrying until invalidation is done is tricky due to the
945 * possibility for live lock, instead defer the add to
946 * mn_itree_inv_end() so this algorithm is deterministic.
947 *
948 * In all cases the value for the interval_sub->invalidate_seq should be
949 * odd, see mmu_interval_read_begin()
950 */
951 spin_lock(&subscriptions->lock);
952 if (subscriptions->active_invalidate_ranges) {
953 if (mn_itree_is_invalidating(subscriptions))
954 hlist_add_head(&interval_sub->deferred_item,
955 &subscriptions->deferred_list);
956 else {
957 subscriptions->invalidate_seq |= 1;
958 interval_tree_insert(&interval_sub->interval_tree,
959 &subscriptions->itree);
960 }
961 interval_sub->invalidate_seq = subscriptions->invalidate_seq;
962 } else {
963 WARN_ON(mn_itree_is_invalidating(subscriptions));
964 /*
965 * The starting seq for a subscription not under invalidation
966 * should be odd, not equal to the current invalidate_seq and
967 * invalidate_seq should not 'wrap' to the new seq any time
968 * soon.
969 */
970 interval_sub->invalidate_seq =
971 subscriptions->invalidate_seq - 1;
972 interval_tree_insert(&interval_sub->interval_tree,
973 &subscriptions->itree);
974 }
975 spin_unlock(&subscriptions->lock);
976 return 0;
977}
978
979/**
980 * mmu_interval_notifier_insert - Insert an interval notifier
981 * @interval_sub: Interval subscription to register
982 * @start: Starting virtual address to monitor
983 * @length: Length of the range to monitor
984 * @mm: mm_struct to attach to
985 * @ops: Interval notifier operations to be called on matching events
986 *
987 * This function subscribes the interval notifier for notifications from the
988 * mm. Upon return the ops related to mmu_interval_notifier will be called
989 * whenever an event that intersects with the given range occurs.
990 *
991 * Upon return the range_notifier may not be present in the interval tree yet.
992 * The caller must use the normal interval notifier read flow via
993 * mmu_interval_read_begin() to establish SPTEs for this range.
994 */
995int mmu_interval_notifier_insert(struct mmu_interval_notifier *interval_sub,
996 struct mm_struct *mm, unsigned long start,
997 unsigned long length,
998 const struct mmu_interval_notifier_ops *ops)
999{
1000 struct mmu_notifier_subscriptions *subscriptions;
1001 int ret;
1002
1003 might_lock(&mm->mmap_lock);
1004
1005 subscriptions = smp_load_acquire(&mm->notifier_subscriptions);
1006 if (!subscriptions || !subscriptions->has_itree) {
1007 ret = mmu_notifier_register(NULL, mm);
1008 if (ret)
1009 return ret;
1010 subscriptions = mm->notifier_subscriptions;
1011 }
1012 return __mmu_interval_notifier_insert(interval_sub, mm, subscriptions,
1013 start, length, ops);
1014}
1015EXPORT_SYMBOL_GPL(mmu_interval_notifier_insert);
1016
1017int mmu_interval_notifier_insert_locked(
1018 struct mmu_interval_notifier *interval_sub, struct mm_struct *mm,
1019 unsigned long start, unsigned long length,
1020 const struct mmu_interval_notifier_ops *ops)
1021{
1022 struct mmu_notifier_subscriptions *subscriptions =
1023 mm->notifier_subscriptions;
1024 int ret;
1025
1026 mmap_assert_write_locked(mm);
1027
1028 if (!subscriptions || !subscriptions->has_itree) {
1029 ret = __mmu_notifier_register(NULL, mm);
1030 if (ret)
1031 return ret;
1032 subscriptions = mm->notifier_subscriptions;
1033 }
1034 return __mmu_interval_notifier_insert(interval_sub, mm, subscriptions,
1035 start, length, ops);
1036}
1037EXPORT_SYMBOL_GPL(mmu_interval_notifier_insert_locked);
1038
1039/**
1040 * mmu_interval_notifier_remove - Remove a interval notifier
1041 * @interval_sub: Interval subscription to unregister
1042 *
1043 * This function must be paired with mmu_interval_notifier_insert(). It cannot
1044 * be called from any ops callback.
1045 *
1046 * Once this returns ops callbacks are no longer running on other CPUs and
1047 * will not be called in future.
1048 */
1049void mmu_interval_notifier_remove(struct mmu_interval_notifier *interval_sub)
1050{
1051 struct mm_struct *mm = interval_sub->mm;
1052 struct mmu_notifier_subscriptions *subscriptions =
1053 mm->notifier_subscriptions;
1054 unsigned long seq = 0;
1055
1056 might_sleep();
1057
1058 spin_lock(&subscriptions->lock);
1059 if (mn_itree_is_invalidating(subscriptions)) {
1060 /*
1061 * remove is being called after insert put this on the
1062 * deferred list, but before the deferred list was processed.
1063 */
1064 if (RB_EMPTY_NODE(&interval_sub->interval_tree.rb)) {
1065 hlist_del(&interval_sub->deferred_item);
1066 } else {
1067 hlist_add_head(&interval_sub->deferred_item,
1068 &subscriptions->deferred_list);
1069 seq = subscriptions->invalidate_seq;
1070 }
1071 } else {
1072 WARN_ON(RB_EMPTY_NODE(&interval_sub->interval_tree.rb));
1073 interval_tree_remove(&interval_sub->interval_tree,
1074 &subscriptions->itree);
1075 }
1076 spin_unlock(&subscriptions->lock);
1077
1078 /*
1079 * The possible sleep on progress in the invalidation requires the
1080 * caller not hold any locks held by invalidation callbacks.
1081 */
1082 lock_map_acquire(&__mmu_notifier_invalidate_range_start_map);
1083 lock_map_release(&__mmu_notifier_invalidate_range_start_map);
1084 if (seq)
1085 wait_event(subscriptions->wq,
1086 READ_ONCE(subscriptions->invalidate_seq) != seq);
1087
1088 /* pairs with mmgrab in mmu_interval_notifier_insert() */
1089 mmdrop(mm);
1090}
1091EXPORT_SYMBOL_GPL(mmu_interval_notifier_remove);
1092
1093/**
1094 * mmu_notifier_synchronize - Ensure all mmu_notifiers are freed
1095 *
1096 * This function ensures that all outstanding async SRU work from
1097 * mmu_notifier_put() is completed. After it returns any mmu_notifier_ops
1098 * associated with an unused mmu_notifier will no longer be called.
1099 *
1100 * Before using the caller must ensure that all of its mmu_notifiers have been
1101 * fully released via mmu_notifier_put().
1102 *
1103 * Modules using the mmu_notifier_put() API should call this in their __exit
1104 * function to avoid module unloading races.
1105 */
1106void mmu_notifier_synchronize(void)
1107{
1108 synchronize_srcu(&srcu);
1109}
1110EXPORT_SYMBOL_GPL(mmu_notifier_synchronize);
1111
1112bool
1113mmu_notifier_range_update_to_read_only(const struct mmu_notifier_range *range)
1114{
1115 if (!range->vma || range->event != MMU_NOTIFY_PROTECTION_VMA)
1116 return false;
1117 /* Return true if the vma still have the read flag set. */
1118 return range->vma->vm_flags & VM_READ;
1119}
1120EXPORT_SYMBOL_GPL(mmu_notifier_range_update_to_read_only);