1 // SPDX-License-Identifier: GPL-2.0-only
3 * linux/mm/mmu_notifier.c
5 * Copyright (C) 2008 Qumranet, Inc.
6 * Copyright (C) 2008 SGI
7 * Christoph Lameter <cl@linux.com>
10 #include <linux/rculist.h>
11 #include <linux/mmu_notifier.h>
12 #include <linux/export.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>
22 /* global SRCU for all MMs */
23 DEFINE_STATIC_SRCU(srcu);
26 struct lockdep_map __mmu_notifier_invalidate_range_start_map = {
27 .name = "mmu_notifier_invalidate_range_start"
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
37 struct mmu_notifier_subscriptions {
38 /* all mmu notifiers registered in this mm are queued in this list */
39 struct hlist_head list;
41 /* to serialize the list modifications and hlist_unhashed */
43 unsigned long invalidate_seq;
44 unsigned long active_invalidate_ranges;
45 struct rb_root_cached itree;
47 struct hlist_head deferred_list;
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
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.
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.
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
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
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
83 * The later state avoids some expensive work on inv_end in the common case of
84 * no mmu_interval_notifier monitoring the VA.
87 mn_itree_is_invalidating(struct mmu_notifier_subscriptions *subscriptions)
89 lockdep_assert_held(&subscriptions->lock);
90 return subscriptions->invalidate_seq & 1;
93 static struct mmu_interval_notifier *
94 mn_itree_inv_start_range(struct mmu_notifier_subscriptions *subscriptions,
95 const struct mmu_notifier_range *range,
98 struct interval_tree_node *node;
99 struct mmu_interval_notifier *res = NULL;
101 spin_lock(&subscriptions->lock);
102 subscriptions->active_invalidate_ranges++;
103 node = interval_tree_iter_first(&subscriptions->itree, range->start,
106 subscriptions->invalidate_seq |= 1;
107 res = container_of(node, struct mmu_interval_notifier,
111 *seq = subscriptions->invalidate_seq;
112 spin_unlock(&subscriptions->lock);
116 static struct mmu_interval_notifier *
117 mn_itree_inv_next(struct mmu_interval_notifier *interval_sub,
118 const struct mmu_notifier_range *range)
120 struct interval_tree_node *node;
122 node = interval_tree_iter_next(&interval_sub->interval_tree,
123 range->start, range->end - 1);
126 return container_of(node, struct mmu_interval_notifier, interval_tree);
129 static void mn_itree_inv_end(struct mmu_notifier_subscriptions *subscriptions)
131 struct mmu_interval_notifier *interval_sub;
132 struct hlist_node *next;
134 spin_lock(&subscriptions->lock);
135 if (--subscriptions->active_invalidate_ranges ||
136 !mn_itree_is_invalidating(subscriptions)) {
137 spin_unlock(&subscriptions->lock);
141 /* Make invalidate_seq even */
142 subscriptions->invalidate_seq++;
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.
150 hlist_for_each_entry_safe(interval_sub, next,
151 &subscriptions->deferred_list,
153 if (RB_EMPTY_NODE(&interval_sub->interval_tree.rb))
154 interval_tree_insert(&interval_sub->interval_tree,
155 &subscriptions->itree);
157 interval_tree_remove(&interval_sub->interval_tree,
158 &subscriptions->itree);
159 hlist_del(&interval_sub->deferred_item);
161 spin_unlock(&subscriptions->lock);
163 wake_up_all(&subscriptions->wq);
167 * mmu_interval_read_begin - Begin a read side critical section against a VA
169 * @interval_sub: The interval subscription
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.
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.
180 * The caller is required to provide a 'user_lock' to serialize both teardown
183 * The return value should be passed to mmu_interval_read_retry().
186 mmu_interval_read_begin(struct mmu_interval_notifier *interval_sub)
188 struct mmu_notifier_subscriptions *subscriptions =
189 interval_sub->mm->notifier_subscriptions;
191 bool is_invalidating;
194 * If the subscription has a different seq value under the user_lock
195 * than we started with then it has collided.
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.
201 * The locking looks broadly like this:
202 * mn_tree_invalidate_start(): mmu_interval_read_begin():
204 * seq = READ_ONCE(interval_sub->invalidate_seq);
205 * seq == subs->invalidate_seq
208 * seq = ++subscriptions->invalidate_seq
210 * op->invalidate_range():
212 * mmu_interval_set_seq()
213 * interval_sub->invalidate_seq = seq
216 * [Required: mmu_interval_read_retry() == true]
218 * mn_itree_inv_end():
220 * seq = ++subscriptions->invalidate_seq
224 * mmu_interval_read_retry():
225 * interval_sub->invalidate_seq != seq
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
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);
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.
245 lock_map_acquire(&__mmu_notifier_invalidate_range_start_map);
246 lock_map_release(&__mmu_notifier_invalidate_range_start_map);
248 wait_event(subscriptions->wq,
249 READ_ONCE(subscriptions->invalidate_seq) != seq);
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
259 EXPORT_SYMBOL_GPL(mmu_interval_read_begin);
261 static void mn_itree_release(struct mmu_notifier_subscriptions *subscriptions,
262 struct mm_struct *mm)
264 struct mmu_notifier_range range = {
265 .flags = MMU_NOTIFIER_RANGE_BLOCKABLE,
266 .event = MMU_NOTIFY_RELEASE,
271 struct mmu_interval_notifier *interval_sub;
272 unsigned long cur_seq;
276 mn_itree_inv_start_range(subscriptions, &range, &cur_seq);
278 interval_sub = mn_itree_inv_next(interval_sub, &range)) {
279 ret = interval_sub->ops->invalidate(interval_sub, &range,
284 mn_itree_inv_end(subscriptions);
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
299 static void mn_hlist_release(struct mmu_notifier_subscriptions *subscriptions,
300 struct mm_struct *mm)
302 struct mmu_notifier *subscription;
306 * SRCU here will block mmu_notifier_unregister until
309 id = srcu_read_lock(&srcu);
310 hlist_for_each_entry_rcu(subscription, &subscriptions->list, hlist,
311 srcu_read_lock_held(&srcu))
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.
318 if (subscription->ops->release)
319 subscription->ops->release(subscription, mm);
321 spin_lock(&subscriptions->lock);
322 while (unlikely(!hlist_empty(&subscriptions->list))) {
323 subscription = hlist_entry(subscriptions->list.first,
324 struct mmu_notifier, hlist);
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
331 hlist_del_init_rcu(&subscription->hlist);
333 spin_unlock(&subscriptions->lock);
334 srcu_read_unlock(&srcu, id);
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.
342 * The notifier_subscriptions can't go away from under us because
343 * one mm_count is held by exit_mmap.
345 synchronize_srcu(&srcu);
348 void __mmu_notifier_release(struct mm_struct *mm)
350 struct mmu_notifier_subscriptions *subscriptions =
351 mm->notifier_subscriptions;
353 if (subscriptions->has_itree)
354 mn_itree_release(subscriptions, mm);
356 if (!hlist_empty(&subscriptions->list))
357 mn_hlist_release(subscriptions, mm);
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
365 int __mmu_notifier_clear_flush_young(struct mm_struct *mm,
369 struct mmu_notifier *subscription;
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);
380 srcu_read_unlock(&srcu, id);
385 int __mmu_notifier_clear_young(struct mm_struct *mm,
389 struct mmu_notifier *subscription;
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,
400 srcu_read_unlock(&srcu, id);
405 int __mmu_notifier_test_young(struct mm_struct *mm,
406 unsigned long address)
408 struct mmu_notifier *subscription;
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,
422 srcu_read_unlock(&srcu, id);
427 void __mmu_notifier_change_pte(struct mm_struct *mm, unsigned long address,
430 struct mmu_notifier *subscription;
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,
441 srcu_read_unlock(&srcu, id);
444 static int mn_itree_invalidate(struct mmu_notifier_subscriptions *subscriptions,
445 const struct mmu_notifier_range *range)
447 struct mmu_interval_notifier *interval_sub;
448 unsigned long cur_seq;
451 mn_itree_inv_start_range(subscriptions, range, &cur_seq);
453 interval_sub = mn_itree_inv_next(interval_sub, range)) {
456 ret = interval_sub->ops->invalidate(interval_sub, range,
459 if (WARN_ON(mmu_notifier_range_blockable(range)))
461 goto out_would_block;
468 * On -EAGAIN the non-blocking caller is not allowed to call
469 * invalidate_range_end()
471 mn_itree_inv_end(subscriptions);
475 static int mn_hlist_invalidate_range_start(
476 struct mmu_notifier_subscriptions *subscriptions,
477 struct mmu_notifier_range *range)
479 struct mmu_notifier *subscription;
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;
488 if (ops->invalidate_range_start) {
491 if (!mmu_notifier_range_blockable(range))
493 _ret = ops->invalidate_range_start(subscription, range);
494 if (!mmu_notifier_range_blockable(range))
497 pr_info("%pS callback failed with %d in %sblockable context.\n",
498 ops->invalidate_range_start, _ret,
499 !mmu_notifier_range_blockable(range) ?
502 WARN_ON(mmu_notifier_range_blockable(range) ||
508 srcu_read_unlock(&srcu, id);
513 int __mmu_notifier_invalidate_range_start(struct mmu_notifier_range *range)
515 struct mmu_notifier_subscriptions *subscriptions =
516 range->mm->notifier_subscriptions;
519 if (subscriptions->has_itree) {
520 ret = mn_itree_invalidate(subscriptions, range);
524 if (!hlist_empty(&subscriptions->list))
525 return mn_hlist_invalidate_range_start(subscriptions, range);
530 mn_hlist_invalidate_end(struct mmu_notifier_subscriptions *subscriptions,
531 struct mmu_notifier_range *range, bool only_end)
533 struct mmu_notifier *subscription;
536 id = srcu_read_lock(&srcu);
537 hlist_for_each_entry_rcu(subscription, &subscriptions->list, hlist,
538 srcu_read_lock_held(&srcu)) {
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).
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.
552 if (!only_end && subscription->ops->invalidate_range)
553 subscription->ops->invalidate_range(subscription,
557 if (subscription->ops->invalidate_range_end) {
558 if (!mmu_notifier_range_blockable(range))
560 subscription->ops->invalidate_range_end(subscription,
562 if (!mmu_notifier_range_blockable(range))
566 srcu_read_unlock(&srcu, id);
569 void __mmu_notifier_invalidate_range_end(struct mmu_notifier_range *range,
572 struct mmu_notifier_subscriptions *subscriptions =
573 range->mm->notifier_subscriptions;
575 lock_map_acquire(&__mmu_notifier_invalidate_range_start_map);
576 if (subscriptions->has_itree)
577 mn_itree_inv_end(subscriptions);
579 if (!hlist_empty(&subscriptions->list))
580 mn_hlist_invalidate_end(subscriptions, range, only_end);
581 lock_map_release(&__mmu_notifier_invalidate_range_start_map);
584 void __mmu_notifier_invalidate_range(struct mm_struct *mm,
585 unsigned long start, unsigned long end)
587 struct mmu_notifier *subscription;
590 id = srcu_read_lock(&srcu);
591 hlist_for_each_entry_rcu(subscription,
592 &mm->notifier_subscriptions->list, hlist,
593 srcu_read_lock_held(&srcu)) {
594 if (subscription->ops->invalidate_range)
595 subscription->ops->invalidate_range(subscription, mm,
598 srcu_read_unlock(&srcu, id);
602 * Same as mmu_notifier_register but here the caller must hold the mmap_lock in
603 * write mode. A NULL mn signals the notifier is being registered for itree
606 int __mmu_notifier_register(struct mmu_notifier *subscription,
607 struct mm_struct *mm)
609 struct mmu_notifier_subscriptions *subscriptions = NULL;
612 mmap_assert_write_locked(mm);
613 BUG_ON(atomic_read(&mm->mm_users) <= 0);
615 if (IS_ENABLED(CONFIG_LOCKDEP)) {
616 fs_reclaim_acquire(GFP_KERNEL);
617 lock_map_acquire(&__mmu_notifier_invalidate_range_start_map);
618 lock_map_release(&__mmu_notifier_invalidate_range_start_map);
619 fs_reclaim_release(GFP_KERNEL);
622 if (!mm->notifier_subscriptions) {
624 * kmalloc cannot be called under mm_take_all_locks(), but we
625 * know that mm->notifier_subscriptions can't change while we
626 * hold the write side of the mmap_lock.
628 subscriptions = kzalloc(
629 sizeof(struct mmu_notifier_subscriptions), GFP_KERNEL);
633 INIT_HLIST_HEAD(&subscriptions->list);
634 spin_lock_init(&subscriptions->lock);
635 subscriptions->invalidate_seq = 2;
636 subscriptions->itree = RB_ROOT_CACHED;
637 init_waitqueue_head(&subscriptions->wq);
638 INIT_HLIST_HEAD(&subscriptions->deferred_list);
641 ret = mm_take_all_locks(mm);
646 * Serialize the update against mmu_notifier_unregister. A
647 * side note: mmu_notifier_release can't run concurrently with
648 * us because we hold the mm_users pin (either implicitly as
649 * current->mm or explicitly with get_task_mm() or similar).
650 * We can't race against any other mmu notifier method either
651 * thanks to mm_take_all_locks().
653 * release semantics on the initialization of the
654 * mmu_notifier_subscriptions's contents are provided for unlocked
655 * readers. acquire can only be used while holding the mmgrab or
656 * mmget, and is safe because once created the
657 * mmu_notifier_subscriptions is not freed until the mm is destroyed.
658 * As above, users holding the mmap_lock or one of the
659 * mm_take_all_locks() do not need to use acquire semantics.
662 smp_store_release(&mm->notifier_subscriptions, subscriptions);
665 /* Pairs with the mmdrop in mmu_notifier_unregister_* */
667 subscription->mm = mm;
668 subscription->users = 1;
670 spin_lock(&mm->notifier_subscriptions->lock);
671 hlist_add_head_rcu(&subscription->hlist,
672 &mm->notifier_subscriptions->list);
673 spin_unlock(&mm->notifier_subscriptions->lock);
675 mm->notifier_subscriptions->has_itree = true;
677 mm_drop_all_locks(mm);
678 BUG_ON(atomic_read(&mm->mm_users) <= 0);
682 kfree(subscriptions);
685 EXPORT_SYMBOL_GPL(__mmu_notifier_register);
688 * mmu_notifier_register - Register a notifier on a mm
689 * @subscription: The notifier to attach
690 * @mm: The mm to attach the notifier to
692 * Must not hold mmap_lock nor any other VM related lock when calling
693 * this registration function. Must also ensure mm_users can't go down
694 * to zero while this runs to avoid races with mmu_notifier_release,
695 * so mm has to be current->mm or the mm should be pinned safely such
696 * as with get_task_mm(). If the mm is not current->mm, the mm_users
697 * pin should be released by calling mmput after mmu_notifier_register
700 * mmu_notifier_unregister() or mmu_notifier_put() must be always called to
701 * unregister the notifier.
703 * While the caller has a mmu_notifier get the subscription->mm pointer will remain
704 * valid, and can be converted to an active mm pointer via mmget_not_zero().
706 int mmu_notifier_register(struct mmu_notifier *subscription,
707 struct mm_struct *mm)
712 ret = __mmu_notifier_register(subscription, mm);
713 mmap_write_unlock(mm);
716 EXPORT_SYMBOL_GPL(mmu_notifier_register);
718 static struct mmu_notifier *
719 find_get_mmu_notifier(struct mm_struct *mm, const struct mmu_notifier_ops *ops)
721 struct mmu_notifier *subscription;
723 spin_lock(&mm->notifier_subscriptions->lock);
724 hlist_for_each_entry_rcu(subscription,
725 &mm->notifier_subscriptions->list, hlist,
726 lockdep_is_held(&mm->notifier_subscriptions->lock)) {
727 if (subscription->ops != ops)
730 if (likely(subscription->users != UINT_MAX))
731 subscription->users++;
733 subscription = ERR_PTR(-EOVERFLOW);
734 spin_unlock(&mm->notifier_subscriptions->lock);
737 spin_unlock(&mm->notifier_subscriptions->lock);
742 * mmu_notifier_get_locked - Return the single struct mmu_notifier for
744 * @ops: The operations struct being subscribe with
745 * @mm : The mm to attach notifiers too
747 * This function either allocates a new mmu_notifier via
748 * ops->alloc_notifier(), or returns an already existing notifier on the
749 * list. The value of the ops pointer is used to determine when two notifiers
752 * Each call to mmu_notifier_get() must be paired with a call to
753 * mmu_notifier_put(). The caller must hold the write side of mm->mmap_lock.
755 * While the caller has a mmu_notifier get the mm pointer will remain valid,
756 * and can be converted to an active mm pointer via mmget_not_zero().
758 struct mmu_notifier *mmu_notifier_get_locked(const struct mmu_notifier_ops *ops,
759 struct mm_struct *mm)
761 struct mmu_notifier *subscription;
764 mmap_assert_write_locked(mm);
766 if (mm->notifier_subscriptions) {
767 subscription = find_get_mmu_notifier(mm, ops);
772 subscription = ops->alloc_notifier(mm);
773 if (IS_ERR(subscription))
775 subscription->ops = ops;
776 ret = __mmu_notifier_register(subscription, mm);
781 subscription->ops->free_notifier(subscription);
784 EXPORT_SYMBOL_GPL(mmu_notifier_get_locked);
786 /* this is called after the last mmu_notifier_unregister() returned */
787 void __mmu_notifier_subscriptions_destroy(struct mm_struct *mm)
789 BUG_ON(!hlist_empty(&mm->notifier_subscriptions->list));
790 kfree(mm->notifier_subscriptions);
791 mm->notifier_subscriptions = LIST_POISON1; /* debug */
795 * This releases the mm_count pin automatically and frees the mm
796 * structure if it was the last user of it. It serializes against
797 * running mmu notifiers with SRCU and against mmu_notifier_unregister
798 * with the unregister lock + SRCU. All sptes must be dropped before
799 * calling mmu_notifier_unregister. ->release or any other notifier
800 * method may be invoked concurrently with mmu_notifier_unregister,
801 * and only after mmu_notifier_unregister returned we're guaranteed
802 * that ->release or any other method can't run anymore.
804 void mmu_notifier_unregister(struct mmu_notifier *subscription,
805 struct mm_struct *mm)
807 BUG_ON(atomic_read(&mm->mm_count) <= 0);
809 if (!hlist_unhashed(&subscription->hlist)) {
811 * SRCU here will force exit_mmap to wait for ->release to
812 * finish before freeing the pages.
816 id = srcu_read_lock(&srcu);
818 * exit_mmap will block in mmu_notifier_release to guarantee
819 * that ->release is called before freeing the pages.
821 if (subscription->ops->release)
822 subscription->ops->release(subscription, mm);
823 srcu_read_unlock(&srcu, id);
825 spin_lock(&mm->notifier_subscriptions->lock);
827 * Can not use list_del_rcu() since __mmu_notifier_release
828 * can delete it before we hold the lock.
830 hlist_del_init_rcu(&subscription->hlist);
831 spin_unlock(&mm->notifier_subscriptions->lock);
835 * Wait for any running method to finish, of course including
836 * ->release if it was run by mmu_notifier_release instead of us.
838 synchronize_srcu(&srcu);
840 BUG_ON(atomic_read(&mm->mm_count) <= 0);
844 EXPORT_SYMBOL_GPL(mmu_notifier_unregister);
846 static void mmu_notifier_free_rcu(struct rcu_head *rcu)
848 struct mmu_notifier *subscription =
849 container_of(rcu, struct mmu_notifier, rcu);
850 struct mm_struct *mm = subscription->mm;
852 subscription->ops->free_notifier(subscription);
853 /* Pairs with the get in __mmu_notifier_register() */
858 * mmu_notifier_put - Release the reference on the notifier
859 * @subscription: The notifier to act on
861 * This function must be paired with each mmu_notifier_get(), it releases the
862 * reference obtained by the get. If this is the last reference then process
863 * to free the notifier will be run asynchronously.
865 * Unlike mmu_notifier_unregister() the get/put flow only calls ops->release
866 * when the mm_struct is destroyed. Instead free_notifier is always called to
867 * release any resources held by the user.
869 * As ops->release is not guaranteed to be called, the user must ensure that
870 * all sptes are dropped, and no new sptes can be established before
871 * mmu_notifier_put() is called.
873 * This function can be called from the ops->release callback, however the
874 * caller must still ensure it is called pairwise with mmu_notifier_get().
876 * Modules calling this function must call mmu_notifier_synchronize() in
877 * their __exit functions to ensure the async work is completed.
879 void mmu_notifier_put(struct mmu_notifier *subscription)
881 struct mm_struct *mm = subscription->mm;
883 spin_lock(&mm->notifier_subscriptions->lock);
884 if (WARN_ON(!subscription->users) || --subscription->users)
886 hlist_del_init_rcu(&subscription->hlist);
887 spin_unlock(&mm->notifier_subscriptions->lock);
889 call_srcu(&srcu, &subscription->rcu, mmu_notifier_free_rcu);
893 spin_unlock(&mm->notifier_subscriptions->lock);
895 EXPORT_SYMBOL_GPL(mmu_notifier_put);
897 static int __mmu_interval_notifier_insert(
898 struct mmu_interval_notifier *interval_sub, struct mm_struct *mm,
899 struct mmu_notifier_subscriptions *subscriptions, unsigned long start,
900 unsigned long length, const struct mmu_interval_notifier_ops *ops)
902 interval_sub->mm = mm;
903 interval_sub->ops = ops;
904 RB_CLEAR_NODE(&interval_sub->interval_tree.rb);
905 interval_sub->interval_tree.start = start;
907 * Note that the representation of the intervals in the interval tree
908 * considers the ending point as contained in the interval.
911 check_add_overflow(start, length - 1,
912 &interval_sub->interval_tree.last))
915 /* Must call with a mmget() held */
916 if (WARN_ON(atomic_read(&mm->mm_users) <= 0))
919 /* pairs with mmdrop in mmu_interval_notifier_remove() */
923 * If some invalidate_range_start/end region is going on in parallel
924 * we don't know what VA ranges are affected, so we must assume this
925 * new range is included.
927 * If the itree is invalidating then we are not allowed to change
928 * it. Retrying until invalidation is done is tricky due to the
929 * possibility for live lock, instead defer the add to
930 * mn_itree_inv_end() so this algorithm is deterministic.
932 * In all cases the value for the interval_sub->invalidate_seq should be
933 * odd, see mmu_interval_read_begin()
935 spin_lock(&subscriptions->lock);
936 if (subscriptions->active_invalidate_ranges) {
937 if (mn_itree_is_invalidating(subscriptions))
938 hlist_add_head(&interval_sub->deferred_item,
939 &subscriptions->deferred_list);
941 subscriptions->invalidate_seq |= 1;
942 interval_tree_insert(&interval_sub->interval_tree,
943 &subscriptions->itree);
945 interval_sub->invalidate_seq = subscriptions->invalidate_seq;
947 WARN_ON(mn_itree_is_invalidating(subscriptions));
949 * The starting seq for a subscription not under invalidation
950 * should be odd, not equal to the current invalidate_seq and
951 * invalidate_seq should not 'wrap' to the new seq any time
954 interval_sub->invalidate_seq =
955 subscriptions->invalidate_seq - 1;
956 interval_tree_insert(&interval_sub->interval_tree,
957 &subscriptions->itree);
959 spin_unlock(&subscriptions->lock);
964 * mmu_interval_notifier_insert - Insert an interval notifier
965 * @interval_sub: Interval subscription to register
966 * @start: Starting virtual address to monitor
967 * @length: Length of the range to monitor
968 * @mm: mm_struct to attach to
969 * @ops: Interval notifier operations to be called on matching events
971 * This function subscribes the interval notifier for notifications from the
972 * mm. Upon return the ops related to mmu_interval_notifier will be called
973 * whenever an event that intersects with the given range occurs.
975 * Upon return the range_notifier may not be present in the interval tree yet.
976 * The caller must use the normal interval notifier read flow via
977 * mmu_interval_read_begin() to establish SPTEs for this range.
979 int mmu_interval_notifier_insert(struct mmu_interval_notifier *interval_sub,
980 struct mm_struct *mm, unsigned long start,
981 unsigned long length,
982 const struct mmu_interval_notifier_ops *ops)
984 struct mmu_notifier_subscriptions *subscriptions;
987 might_lock(&mm->mmap_lock);
989 subscriptions = smp_load_acquire(&mm->notifier_subscriptions);
990 if (!subscriptions || !subscriptions->has_itree) {
991 ret = mmu_notifier_register(NULL, mm);
994 subscriptions = mm->notifier_subscriptions;
996 return __mmu_interval_notifier_insert(interval_sub, mm, subscriptions,
999 EXPORT_SYMBOL_GPL(mmu_interval_notifier_insert);
1001 int mmu_interval_notifier_insert_locked(
1002 struct mmu_interval_notifier *interval_sub, struct mm_struct *mm,
1003 unsigned long start, unsigned long length,
1004 const struct mmu_interval_notifier_ops *ops)
1006 struct mmu_notifier_subscriptions *subscriptions =
1007 mm->notifier_subscriptions;
1010 mmap_assert_write_locked(mm);
1012 if (!subscriptions || !subscriptions->has_itree) {
1013 ret = __mmu_notifier_register(NULL, mm);
1016 subscriptions = mm->notifier_subscriptions;
1018 return __mmu_interval_notifier_insert(interval_sub, mm, subscriptions,
1019 start, length, ops);
1021 EXPORT_SYMBOL_GPL(mmu_interval_notifier_insert_locked);
1024 * mmu_interval_notifier_remove - Remove a interval notifier
1025 * @interval_sub: Interval subscription to unregister
1027 * This function must be paired with mmu_interval_notifier_insert(). It cannot
1028 * be called from any ops callback.
1030 * Once this returns ops callbacks are no longer running on other CPUs and
1031 * will not be called in future.
1033 void mmu_interval_notifier_remove(struct mmu_interval_notifier *interval_sub)
1035 struct mm_struct *mm = interval_sub->mm;
1036 struct mmu_notifier_subscriptions *subscriptions =
1037 mm->notifier_subscriptions;
1038 unsigned long seq = 0;
1042 spin_lock(&subscriptions->lock);
1043 if (mn_itree_is_invalidating(subscriptions)) {
1045 * remove is being called after insert put this on the
1046 * deferred list, but before the deferred list was processed.
1048 if (RB_EMPTY_NODE(&interval_sub->interval_tree.rb)) {
1049 hlist_del(&interval_sub->deferred_item);
1051 hlist_add_head(&interval_sub->deferred_item,
1052 &subscriptions->deferred_list);
1053 seq = subscriptions->invalidate_seq;
1056 WARN_ON(RB_EMPTY_NODE(&interval_sub->interval_tree.rb));
1057 interval_tree_remove(&interval_sub->interval_tree,
1058 &subscriptions->itree);
1060 spin_unlock(&subscriptions->lock);
1063 * The possible sleep on progress in the invalidation requires the
1064 * caller not hold any locks held by invalidation callbacks.
1066 lock_map_acquire(&__mmu_notifier_invalidate_range_start_map);
1067 lock_map_release(&__mmu_notifier_invalidate_range_start_map);
1069 wait_event(subscriptions->wq,
1070 READ_ONCE(subscriptions->invalidate_seq) != seq);
1072 /* pairs with mmgrab in mmu_interval_notifier_insert() */
1075 EXPORT_SYMBOL_GPL(mmu_interval_notifier_remove);
1078 * mmu_notifier_synchronize - Ensure all mmu_notifiers are freed
1080 * This function ensures that all outstanding async SRU work from
1081 * mmu_notifier_put() is completed. After it returns any mmu_notifier_ops
1082 * associated with an unused mmu_notifier will no longer be called.
1084 * Before using the caller must ensure that all of its mmu_notifiers have been
1085 * fully released via mmu_notifier_put().
1087 * Modules using the mmu_notifier_put() API should call this in their __exit
1088 * function to avoid module unloading races.
1090 void mmu_notifier_synchronize(void)
1092 synchronize_srcu(&srcu);
1094 EXPORT_SYMBOL_GPL(mmu_notifier_synchronize);
1097 mmu_notifier_range_update_to_read_only(const struct mmu_notifier_range *range)
1099 if (!range->vma || range->event != MMU_NOTIFY_PROTECTION_VMA)
1101 /* Return true if the vma still have the read flag set. */
1102 return range->vma->vm_flags & VM_READ;
1104 EXPORT_SYMBOL_GPL(mmu_notifier_range_update_to_read_only);