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_itree_inv_start(): mmu_interval_read_begin():
204 * seq = READ_ONCE(interval_sub->invalidate_seq);
205 * seq == subs->invalidate_seq
208 * seq = ++subscriptions->invalidate_seq
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 static int mn_itree_invalidate(struct mmu_notifier_subscriptions
*subscriptions
,
428 const struct mmu_notifier_range
*range
)
430 struct mmu_interval_notifier
*interval_sub
;
431 unsigned long cur_seq
;
434 mn_itree_inv_start_range(subscriptions
, range
, &cur_seq
);
436 interval_sub
= mn_itree_inv_next(interval_sub
, range
)) {
439 ret
= interval_sub
->ops
->invalidate(interval_sub
, range
,
442 if (WARN_ON(mmu_notifier_range_blockable(range
)))
444 goto out_would_block
;
451 * On -EAGAIN the non-blocking caller is not allowed to call
452 * invalidate_range_end()
454 mn_itree_inv_end(subscriptions
);
458 static int mn_hlist_invalidate_range_start(
459 struct mmu_notifier_subscriptions
*subscriptions
,
460 struct mmu_notifier_range
*range
)
462 struct mmu_notifier
*subscription
;
466 id
= srcu_read_lock(&srcu
);
467 hlist_for_each_entry_rcu(subscription
, &subscriptions
->list
, hlist
,
468 srcu_read_lock_held(&srcu
)) {
469 const struct mmu_notifier_ops
*ops
= subscription
->ops
;
471 if (ops
->invalidate_range_start
) {
474 if (!mmu_notifier_range_blockable(range
))
476 _ret
= ops
->invalidate_range_start(subscription
, range
);
477 if (!mmu_notifier_range_blockable(range
))
480 pr_info("%pS callback failed with %d in %sblockable context.\n",
481 ops
->invalidate_range_start
, _ret
,
482 !mmu_notifier_range_blockable(range
) ?
485 WARN_ON(mmu_notifier_range_blockable(range
) ||
488 * We call all the notifiers on any EAGAIN,
489 * there is no way for a notifier to know if
490 * its start method failed, thus a start that
491 * does EAGAIN can't also do end.
493 WARN_ON(ops
->invalidate_range_end
);
501 * Must be non-blocking to get here. If there are multiple
502 * notifiers and one or more failed start, any that succeeded
503 * start are expecting their end to be called. Do so now.
505 hlist_for_each_entry_rcu(subscription
, &subscriptions
->list
,
506 hlist
, srcu_read_lock_held(&srcu
)) {
507 if (!subscription
->ops
->invalidate_range_end
)
510 subscription
->ops
->invalidate_range_end(subscription
,
514 srcu_read_unlock(&srcu
, id
);
519 int __mmu_notifier_invalidate_range_start(struct mmu_notifier_range
*range
)
521 struct mmu_notifier_subscriptions
*subscriptions
=
522 range
->mm
->notifier_subscriptions
;
525 if (subscriptions
->has_itree
) {
526 ret
= mn_itree_invalidate(subscriptions
, range
);
530 if (!hlist_empty(&subscriptions
->list
))
531 return mn_hlist_invalidate_range_start(subscriptions
, range
);
536 mn_hlist_invalidate_end(struct mmu_notifier_subscriptions
*subscriptions
,
537 struct mmu_notifier_range
*range
)
539 struct mmu_notifier
*subscription
;
542 id
= srcu_read_lock(&srcu
);
543 hlist_for_each_entry_rcu(subscription
, &subscriptions
->list
, hlist
,
544 srcu_read_lock_held(&srcu
)) {
545 if (subscription
->ops
->invalidate_range_end
) {
546 if (!mmu_notifier_range_blockable(range
))
548 subscription
->ops
->invalidate_range_end(subscription
,
550 if (!mmu_notifier_range_blockable(range
))
554 srcu_read_unlock(&srcu
, id
);
557 void __mmu_notifier_invalidate_range_end(struct mmu_notifier_range
*range
)
559 struct mmu_notifier_subscriptions
*subscriptions
=
560 range
->mm
->notifier_subscriptions
;
562 lock_map_acquire(&__mmu_notifier_invalidate_range_start_map
);
563 if (subscriptions
->has_itree
)
564 mn_itree_inv_end(subscriptions
);
566 if (!hlist_empty(&subscriptions
->list
))
567 mn_hlist_invalidate_end(subscriptions
, range
);
568 lock_map_release(&__mmu_notifier_invalidate_range_start_map
);
571 void __mmu_notifier_arch_invalidate_secondary_tlbs(struct mm_struct
*mm
,
572 unsigned long start
, unsigned long end
)
574 struct mmu_notifier
*subscription
;
577 id
= srcu_read_lock(&srcu
);
578 hlist_for_each_entry_rcu(subscription
,
579 &mm
->notifier_subscriptions
->list
, hlist
,
580 srcu_read_lock_held(&srcu
)) {
581 if (subscription
->ops
->arch_invalidate_secondary_tlbs
)
582 subscription
->ops
->arch_invalidate_secondary_tlbs(
586 srcu_read_unlock(&srcu
, id
);
590 * Same as mmu_notifier_register but here the caller must hold the mmap_lock in
591 * write mode. A NULL mn signals the notifier is being registered for itree
594 int __mmu_notifier_register(struct mmu_notifier
*subscription
,
595 struct mm_struct
*mm
)
597 struct mmu_notifier_subscriptions
*subscriptions
= NULL
;
600 mmap_assert_write_locked(mm
);
601 BUG_ON(atomic_read(&mm
->mm_users
) <= 0);
604 * Subsystems should only register for invalidate_secondary_tlbs() or
605 * invalidate_range_start()/end() callbacks, not both.
607 if (WARN_ON_ONCE(subscription
&&
608 (subscription
->ops
->arch_invalidate_secondary_tlbs
&&
609 (subscription
->ops
->invalidate_range_start
||
610 subscription
->ops
->invalidate_range_end
))))
613 if (!mm
->notifier_subscriptions
) {
615 * kmalloc cannot be called under mm_take_all_locks(), but we
616 * know that mm->notifier_subscriptions can't change while we
617 * hold the write side of the mmap_lock.
619 subscriptions
= kzalloc(
620 sizeof(struct mmu_notifier_subscriptions
), GFP_KERNEL
);
624 INIT_HLIST_HEAD(&subscriptions
->list
);
625 spin_lock_init(&subscriptions
->lock
);
626 subscriptions
->invalidate_seq
= 2;
627 subscriptions
->itree
= RB_ROOT_CACHED
;
628 init_waitqueue_head(&subscriptions
->wq
);
629 INIT_HLIST_HEAD(&subscriptions
->deferred_list
);
632 ret
= mm_take_all_locks(mm
);
637 * Serialize the update against mmu_notifier_unregister. A
638 * side note: mmu_notifier_release can't run concurrently with
639 * us because we hold the mm_users pin (either implicitly as
640 * current->mm or explicitly with get_task_mm() or similar).
641 * We can't race against any other mmu notifier method either
642 * thanks to mm_take_all_locks().
644 * release semantics on the initialization of the
645 * mmu_notifier_subscriptions's contents are provided for unlocked
646 * readers. acquire can only be used while holding the mmgrab or
647 * mmget, and is safe because once created the
648 * mmu_notifier_subscriptions is not freed until the mm is destroyed.
649 * As above, users holding the mmap_lock or one of the
650 * mm_take_all_locks() do not need to use acquire semantics.
653 smp_store_release(&mm
->notifier_subscriptions
, subscriptions
);
656 /* Pairs with the mmdrop in mmu_notifier_unregister_* */
658 subscription
->mm
= mm
;
659 subscription
->users
= 1;
661 spin_lock(&mm
->notifier_subscriptions
->lock
);
662 hlist_add_head_rcu(&subscription
->hlist
,
663 &mm
->notifier_subscriptions
->list
);
664 spin_unlock(&mm
->notifier_subscriptions
->lock
);
666 mm
->notifier_subscriptions
->has_itree
= true;
668 mm_drop_all_locks(mm
);
669 BUG_ON(atomic_read(&mm
->mm_users
) <= 0);
673 kfree(subscriptions
);
676 EXPORT_SYMBOL_GPL(__mmu_notifier_register
);
679 * mmu_notifier_register - Register a notifier on a mm
680 * @subscription: The notifier to attach
681 * @mm: The mm to attach the notifier to
683 * Must not hold mmap_lock nor any other VM related lock when calling
684 * this registration function. Must also ensure mm_users can't go down
685 * to zero while this runs to avoid races with mmu_notifier_release,
686 * so mm has to be current->mm or the mm should be pinned safely such
687 * as with get_task_mm(). If the mm is not current->mm, the mm_users
688 * pin should be released by calling mmput after mmu_notifier_register
691 * mmu_notifier_unregister() or mmu_notifier_put() must be always called to
692 * unregister the notifier.
694 * While the caller has a mmu_notifier get the subscription->mm pointer will remain
695 * valid, and can be converted to an active mm pointer via mmget_not_zero().
697 int mmu_notifier_register(struct mmu_notifier
*subscription
,
698 struct mm_struct
*mm
)
703 ret
= __mmu_notifier_register(subscription
, mm
);
704 mmap_write_unlock(mm
);
707 EXPORT_SYMBOL_GPL(mmu_notifier_register
);
709 static struct mmu_notifier
*
710 find_get_mmu_notifier(struct mm_struct
*mm
, const struct mmu_notifier_ops
*ops
)
712 struct mmu_notifier
*subscription
;
714 spin_lock(&mm
->notifier_subscriptions
->lock
);
715 hlist_for_each_entry_rcu(subscription
,
716 &mm
->notifier_subscriptions
->list
, hlist
,
717 lockdep_is_held(&mm
->notifier_subscriptions
->lock
)) {
718 if (subscription
->ops
!= ops
)
721 if (likely(subscription
->users
!= UINT_MAX
))
722 subscription
->users
++;
724 subscription
= ERR_PTR(-EOVERFLOW
);
725 spin_unlock(&mm
->notifier_subscriptions
->lock
);
728 spin_unlock(&mm
->notifier_subscriptions
->lock
);
733 * mmu_notifier_get_locked - Return the single struct mmu_notifier for
735 * @ops: The operations struct being subscribe with
736 * @mm : The mm to attach notifiers too
738 * This function either allocates a new mmu_notifier via
739 * ops->alloc_notifier(), or returns an already existing notifier on the
740 * list. The value of the ops pointer is used to determine when two notifiers
743 * Each call to mmu_notifier_get() must be paired with a call to
744 * mmu_notifier_put(). The caller must hold the write side of mm->mmap_lock.
746 * While the caller has a mmu_notifier get the mm pointer will remain valid,
747 * and can be converted to an active mm pointer via mmget_not_zero().
749 struct mmu_notifier
*mmu_notifier_get_locked(const struct mmu_notifier_ops
*ops
,
750 struct mm_struct
*mm
)
752 struct mmu_notifier
*subscription
;
755 mmap_assert_write_locked(mm
);
757 if (mm
->notifier_subscriptions
) {
758 subscription
= find_get_mmu_notifier(mm
, ops
);
763 subscription
= ops
->alloc_notifier(mm
);
764 if (IS_ERR(subscription
))
766 subscription
->ops
= ops
;
767 ret
= __mmu_notifier_register(subscription
, mm
);
772 subscription
->ops
->free_notifier(subscription
);
775 EXPORT_SYMBOL_GPL(mmu_notifier_get_locked
);
777 /* this is called after the last mmu_notifier_unregister() returned */
778 void __mmu_notifier_subscriptions_destroy(struct mm_struct
*mm
)
780 BUG_ON(!hlist_empty(&mm
->notifier_subscriptions
->list
));
781 kfree(mm
->notifier_subscriptions
);
782 mm
->notifier_subscriptions
= LIST_POISON1
; /* debug */
786 * This releases the mm_count pin automatically and frees the mm
787 * structure if it was the last user of it. It serializes against
788 * running mmu notifiers with SRCU and against mmu_notifier_unregister
789 * with the unregister lock + SRCU. All sptes must be dropped before
790 * calling mmu_notifier_unregister. ->release or any other notifier
791 * method may be invoked concurrently with mmu_notifier_unregister,
792 * and only after mmu_notifier_unregister returned we're guaranteed
793 * that ->release or any other method can't run anymore.
795 void mmu_notifier_unregister(struct mmu_notifier
*subscription
,
796 struct mm_struct
*mm
)
798 BUG_ON(atomic_read(&mm
->mm_count
) <= 0);
800 if (!hlist_unhashed(&subscription
->hlist
)) {
802 * SRCU here will force exit_mmap to wait for ->release to
803 * finish before freeing the pages.
807 id
= srcu_read_lock(&srcu
);
809 * exit_mmap will block in mmu_notifier_release to guarantee
810 * that ->release is called before freeing the pages.
812 if (subscription
->ops
->release
)
813 subscription
->ops
->release(subscription
, mm
);
814 srcu_read_unlock(&srcu
, id
);
816 spin_lock(&mm
->notifier_subscriptions
->lock
);
818 * Can not use list_del_rcu() since __mmu_notifier_release
819 * can delete it before we hold the lock.
821 hlist_del_init_rcu(&subscription
->hlist
);
822 spin_unlock(&mm
->notifier_subscriptions
->lock
);
826 * Wait for any running method to finish, of course including
827 * ->release if it was run by mmu_notifier_release instead of us.
829 synchronize_srcu(&srcu
);
831 BUG_ON(atomic_read(&mm
->mm_count
) <= 0);
835 EXPORT_SYMBOL_GPL(mmu_notifier_unregister
);
837 static void mmu_notifier_free_rcu(struct rcu_head
*rcu
)
839 struct mmu_notifier
*subscription
=
840 container_of(rcu
, struct mmu_notifier
, rcu
);
841 struct mm_struct
*mm
= subscription
->mm
;
843 subscription
->ops
->free_notifier(subscription
);
844 /* Pairs with the get in __mmu_notifier_register() */
849 * mmu_notifier_put - Release the reference on the notifier
850 * @subscription: The notifier to act on
852 * This function must be paired with each mmu_notifier_get(), it releases the
853 * reference obtained by the get. If this is the last reference then process
854 * to free the notifier will be run asynchronously.
856 * Unlike mmu_notifier_unregister() the get/put flow only calls ops->release
857 * when the mm_struct is destroyed. Instead free_notifier is always called to
858 * release any resources held by the user.
860 * As ops->release is not guaranteed to be called, the user must ensure that
861 * all sptes are dropped, and no new sptes can be established before
862 * mmu_notifier_put() is called.
864 * This function can be called from the ops->release callback, however the
865 * caller must still ensure it is called pairwise with mmu_notifier_get().
867 * Modules calling this function must call mmu_notifier_synchronize() in
868 * their __exit functions to ensure the async work is completed.
870 void mmu_notifier_put(struct mmu_notifier
*subscription
)
872 struct mm_struct
*mm
= subscription
->mm
;
874 spin_lock(&mm
->notifier_subscriptions
->lock
);
875 if (WARN_ON(!subscription
->users
) || --subscription
->users
)
877 hlist_del_init_rcu(&subscription
->hlist
);
878 spin_unlock(&mm
->notifier_subscriptions
->lock
);
880 call_srcu(&srcu
, &subscription
->rcu
, mmu_notifier_free_rcu
);
884 spin_unlock(&mm
->notifier_subscriptions
->lock
);
886 EXPORT_SYMBOL_GPL(mmu_notifier_put
);
888 static int __mmu_interval_notifier_insert(
889 struct mmu_interval_notifier
*interval_sub
, struct mm_struct
*mm
,
890 struct mmu_notifier_subscriptions
*subscriptions
, unsigned long start
,
891 unsigned long length
, const struct mmu_interval_notifier_ops
*ops
)
893 interval_sub
->mm
= mm
;
894 interval_sub
->ops
= ops
;
895 RB_CLEAR_NODE(&interval_sub
->interval_tree
.rb
);
896 interval_sub
->interval_tree
.start
= start
;
898 * Note that the representation of the intervals in the interval tree
899 * considers the ending point as contained in the interval.
902 check_add_overflow(start
, length
- 1,
903 &interval_sub
->interval_tree
.last
))
906 /* Must call with a mmget() held */
907 if (WARN_ON(atomic_read(&mm
->mm_users
) <= 0))
910 /* pairs with mmdrop in mmu_interval_notifier_remove() */
914 * If some invalidate_range_start/end region is going on in parallel
915 * we don't know what VA ranges are affected, so we must assume this
916 * new range is included.
918 * If the itree is invalidating then we are not allowed to change
919 * it. Retrying until invalidation is done is tricky due to the
920 * possibility for live lock, instead defer the add to
921 * mn_itree_inv_end() so this algorithm is deterministic.
923 * In all cases the value for the interval_sub->invalidate_seq should be
924 * odd, see mmu_interval_read_begin()
926 spin_lock(&subscriptions
->lock
);
927 if (subscriptions
->active_invalidate_ranges
) {
928 if (mn_itree_is_invalidating(subscriptions
))
929 hlist_add_head(&interval_sub
->deferred_item
,
930 &subscriptions
->deferred_list
);
932 subscriptions
->invalidate_seq
|= 1;
933 interval_tree_insert(&interval_sub
->interval_tree
,
934 &subscriptions
->itree
);
936 interval_sub
->invalidate_seq
= subscriptions
->invalidate_seq
;
938 WARN_ON(mn_itree_is_invalidating(subscriptions
));
940 * The starting seq for a subscription not under invalidation
941 * should be odd, not equal to the current invalidate_seq and
942 * invalidate_seq should not 'wrap' to the new seq any time
945 interval_sub
->invalidate_seq
=
946 subscriptions
->invalidate_seq
- 1;
947 interval_tree_insert(&interval_sub
->interval_tree
,
948 &subscriptions
->itree
);
950 spin_unlock(&subscriptions
->lock
);
955 * mmu_interval_notifier_insert - Insert an interval notifier
956 * @interval_sub: Interval subscription to register
957 * @start: Starting virtual address to monitor
958 * @length: Length of the range to monitor
959 * @mm: mm_struct to attach to
960 * @ops: Interval notifier operations to be called on matching events
962 * This function subscribes the interval notifier for notifications from the
963 * mm. Upon return the ops related to mmu_interval_notifier will be called
964 * whenever an event that intersects with the given range occurs.
966 * Upon return the range_notifier may not be present in the interval tree yet.
967 * The caller must use the normal interval notifier read flow via
968 * mmu_interval_read_begin() to establish SPTEs for this range.
970 int mmu_interval_notifier_insert(struct mmu_interval_notifier
*interval_sub
,
971 struct mm_struct
*mm
, unsigned long start
,
972 unsigned long length
,
973 const struct mmu_interval_notifier_ops
*ops
)
975 struct mmu_notifier_subscriptions
*subscriptions
;
978 might_lock(&mm
->mmap_lock
);
980 subscriptions
= smp_load_acquire(&mm
->notifier_subscriptions
);
981 if (!subscriptions
|| !subscriptions
->has_itree
) {
982 ret
= mmu_notifier_register(NULL
, mm
);
985 subscriptions
= mm
->notifier_subscriptions
;
987 return __mmu_interval_notifier_insert(interval_sub
, mm
, subscriptions
,
990 EXPORT_SYMBOL_GPL(mmu_interval_notifier_insert
);
992 int mmu_interval_notifier_insert_locked(
993 struct mmu_interval_notifier
*interval_sub
, struct mm_struct
*mm
,
994 unsigned long start
, unsigned long length
,
995 const struct mmu_interval_notifier_ops
*ops
)
997 struct mmu_notifier_subscriptions
*subscriptions
=
998 mm
->notifier_subscriptions
;
1001 mmap_assert_write_locked(mm
);
1003 if (!subscriptions
|| !subscriptions
->has_itree
) {
1004 ret
= __mmu_notifier_register(NULL
, mm
);
1007 subscriptions
= mm
->notifier_subscriptions
;
1009 return __mmu_interval_notifier_insert(interval_sub
, mm
, subscriptions
,
1010 start
, length
, ops
);
1012 EXPORT_SYMBOL_GPL(mmu_interval_notifier_insert_locked
);
1015 mmu_interval_seq_released(struct mmu_notifier_subscriptions
*subscriptions
,
1020 spin_lock(&subscriptions
->lock
);
1021 ret
= subscriptions
->invalidate_seq
!= seq
;
1022 spin_unlock(&subscriptions
->lock
);
1027 * mmu_interval_notifier_remove - Remove a interval notifier
1028 * @interval_sub: Interval subscription to unregister
1030 * This function must be paired with mmu_interval_notifier_insert(). It cannot
1031 * be called from any ops callback.
1033 * Once this returns ops callbacks are no longer running on other CPUs and
1034 * will not be called in future.
1036 void mmu_interval_notifier_remove(struct mmu_interval_notifier
*interval_sub
)
1038 struct mm_struct
*mm
= interval_sub
->mm
;
1039 struct mmu_notifier_subscriptions
*subscriptions
=
1040 mm
->notifier_subscriptions
;
1041 unsigned long seq
= 0;
1045 spin_lock(&subscriptions
->lock
);
1046 if (mn_itree_is_invalidating(subscriptions
)) {
1048 * remove is being called after insert put this on the
1049 * deferred list, but before the deferred list was processed.
1051 if (RB_EMPTY_NODE(&interval_sub
->interval_tree
.rb
)) {
1052 hlist_del(&interval_sub
->deferred_item
);
1054 hlist_add_head(&interval_sub
->deferred_item
,
1055 &subscriptions
->deferred_list
);
1056 seq
= subscriptions
->invalidate_seq
;
1059 WARN_ON(RB_EMPTY_NODE(&interval_sub
->interval_tree
.rb
));
1060 interval_tree_remove(&interval_sub
->interval_tree
,
1061 &subscriptions
->itree
);
1063 spin_unlock(&subscriptions
->lock
);
1066 * The possible sleep on progress in the invalidation requires the
1067 * caller not hold any locks held by invalidation callbacks.
1069 lock_map_acquire(&__mmu_notifier_invalidate_range_start_map
);
1070 lock_map_release(&__mmu_notifier_invalidate_range_start_map
);
1072 wait_event(subscriptions
->wq
,
1073 mmu_interval_seq_released(subscriptions
, seq
));
1075 /* pairs with mmgrab in mmu_interval_notifier_insert() */
1078 EXPORT_SYMBOL_GPL(mmu_interval_notifier_remove
);
1081 * mmu_notifier_synchronize - Ensure all mmu_notifiers are freed
1083 * This function ensures that all outstanding async SRU work from
1084 * mmu_notifier_put() is completed. After it returns any mmu_notifier_ops
1085 * associated with an unused mmu_notifier will no longer be called.
1087 * Before using the caller must ensure that all of its mmu_notifiers have been
1088 * fully released via mmu_notifier_put().
1090 * Modules using the mmu_notifier_put() API should call this in their __exit
1091 * function to avoid module unloading races.
1093 void mmu_notifier_synchronize(void)
1095 synchronize_srcu(&srcu
);
1097 EXPORT_SYMBOL_GPL(mmu_notifier_synchronize
);