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
) ||
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.
510 WARN_ON(ops
->invalidate_range_end
);
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.
522 hlist_for_each_entry_rcu(subscription
, &subscriptions
->list
,
523 hlist
, srcu_read_lock_held(&srcu
)) {
524 if (!subscription
->ops
->invalidate_range_end
)
527 subscription
->ops
->invalidate_range_end(subscription
,
531 srcu_read_unlock(&srcu
, id
);
536 int __mmu_notifier_invalidate_range_start(struct mmu_notifier_range
*range
)
538 struct mmu_notifier_subscriptions
*subscriptions
=
539 range
->mm
->notifier_subscriptions
;
542 if (subscriptions
->has_itree
) {
543 ret
= mn_itree_invalidate(subscriptions
, range
);
547 if (!hlist_empty(&subscriptions
->list
))
548 return mn_hlist_invalidate_range_start(subscriptions
, range
);
553 mn_hlist_invalidate_end(struct mmu_notifier_subscriptions
*subscriptions
,
554 struct mmu_notifier_range
*range
, bool only_end
)
556 struct mmu_notifier
*subscription
;
559 id
= srcu_read_lock(&srcu
);
560 hlist_for_each_entry_rcu(subscription
, &subscriptions
->list
, hlist
,
561 srcu_read_lock_held(&srcu
)) {
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).
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.
575 if (!only_end
&& subscription
->ops
->invalidate_range
)
576 subscription
->ops
->invalidate_range(subscription
,
580 if (subscription
->ops
->invalidate_range_end
) {
581 if (!mmu_notifier_range_blockable(range
))
583 subscription
->ops
->invalidate_range_end(subscription
,
585 if (!mmu_notifier_range_blockable(range
))
589 srcu_read_unlock(&srcu
, id
);
592 void __mmu_notifier_invalidate_range_end(struct mmu_notifier_range
*range
,
595 struct mmu_notifier_subscriptions
*subscriptions
=
596 range
->mm
->notifier_subscriptions
;
598 lock_map_acquire(&__mmu_notifier_invalidate_range_start_map
);
599 if (subscriptions
->has_itree
)
600 mn_itree_inv_end(subscriptions
);
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
);
607 void __mmu_notifier_invalidate_range(struct mm_struct
*mm
,
608 unsigned long start
, unsigned long end
)
610 struct mmu_notifier
*subscription
;
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
,
621 srcu_read_unlock(&srcu
, id
);
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
629 int __mmu_notifier_register(struct mmu_notifier
*subscription
,
630 struct mm_struct
*mm
)
632 struct mmu_notifier_subscriptions
*subscriptions
= NULL
;
635 mmap_assert_write_locked(mm
);
636 BUG_ON(atomic_read(&mm
->mm_users
) <= 0);
638 if (!mm
->notifier_subscriptions
) {
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.
644 subscriptions
= kzalloc(
645 sizeof(struct mmu_notifier_subscriptions
), GFP_KERNEL
);
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
);
657 ret
= mm_take_all_locks(mm
);
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().
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.
678 smp_store_release(&mm
->notifier_subscriptions
, subscriptions
);
681 /* Pairs with the mmdrop in mmu_notifier_unregister_* */
683 subscription
->mm
= mm
;
684 subscription
->users
= 1;
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
);
691 mm
->notifier_subscriptions
->has_itree
= true;
693 mm_drop_all_locks(mm
);
694 BUG_ON(atomic_read(&mm
->mm_users
) <= 0);
698 kfree(subscriptions
);
701 EXPORT_SYMBOL_GPL(__mmu_notifier_register
);
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
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
716 * mmu_notifier_unregister() or mmu_notifier_put() must be always called to
717 * unregister the notifier.
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().
722 int mmu_notifier_register(struct mmu_notifier
*subscription
,
723 struct mm_struct
*mm
)
728 ret
= __mmu_notifier_register(subscription
, mm
);
729 mmap_write_unlock(mm
);
732 EXPORT_SYMBOL_GPL(mmu_notifier_register
);
734 static struct mmu_notifier
*
735 find_get_mmu_notifier(struct mm_struct
*mm
, const struct mmu_notifier_ops
*ops
)
737 struct mmu_notifier
*subscription
;
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
)
746 if (likely(subscription
->users
!= UINT_MAX
))
747 subscription
->users
++;
749 subscription
= ERR_PTR(-EOVERFLOW
);
750 spin_unlock(&mm
->notifier_subscriptions
->lock
);
753 spin_unlock(&mm
->notifier_subscriptions
->lock
);
758 * mmu_notifier_get_locked - Return the single struct mmu_notifier for
760 * @ops: The operations struct being subscribe with
761 * @mm : The mm to attach notifiers too
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
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.
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().
774 struct mmu_notifier
*mmu_notifier_get_locked(const struct mmu_notifier_ops
*ops
,
775 struct mm_struct
*mm
)
777 struct mmu_notifier
*subscription
;
780 mmap_assert_write_locked(mm
);
782 if (mm
->notifier_subscriptions
) {
783 subscription
= find_get_mmu_notifier(mm
, ops
);
788 subscription
= ops
->alloc_notifier(mm
);
789 if (IS_ERR(subscription
))
791 subscription
->ops
= ops
;
792 ret
= __mmu_notifier_register(subscription
, mm
);
797 subscription
->ops
->free_notifier(subscription
);
800 EXPORT_SYMBOL_GPL(mmu_notifier_get_locked
);
802 /* this is called after the last mmu_notifier_unregister() returned */
803 void __mmu_notifier_subscriptions_destroy(struct mm_struct
*mm
)
805 BUG_ON(!hlist_empty(&mm
->notifier_subscriptions
->list
));
806 kfree(mm
->notifier_subscriptions
);
807 mm
->notifier_subscriptions
= LIST_POISON1
; /* debug */
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.
820 void mmu_notifier_unregister(struct mmu_notifier
*subscription
,
821 struct mm_struct
*mm
)
823 BUG_ON(atomic_read(&mm
->mm_count
) <= 0);
825 if (!hlist_unhashed(&subscription
->hlist
)) {
827 * SRCU here will force exit_mmap to wait for ->release to
828 * finish before freeing the pages.
832 id
= srcu_read_lock(&srcu
);
834 * exit_mmap will block in mmu_notifier_release to guarantee
835 * that ->release is called before freeing the pages.
837 if (subscription
->ops
->release
)
838 subscription
->ops
->release(subscription
, mm
);
839 srcu_read_unlock(&srcu
, id
);
841 spin_lock(&mm
->notifier_subscriptions
->lock
);
843 * Can not use list_del_rcu() since __mmu_notifier_release
844 * can delete it before we hold the lock.
846 hlist_del_init_rcu(&subscription
->hlist
);
847 spin_unlock(&mm
->notifier_subscriptions
->lock
);
851 * Wait for any running method to finish, of course including
852 * ->release if it was run by mmu_notifier_release instead of us.
854 synchronize_srcu(&srcu
);
856 BUG_ON(atomic_read(&mm
->mm_count
) <= 0);
860 EXPORT_SYMBOL_GPL(mmu_notifier_unregister
);
862 static void mmu_notifier_free_rcu(struct rcu_head
*rcu
)
864 struct mmu_notifier
*subscription
=
865 container_of(rcu
, struct mmu_notifier
, rcu
);
866 struct mm_struct
*mm
= subscription
->mm
;
868 subscription
->ops
->free_notifier(subscription
);
869 /* Pairs with the get in __mmu_notifier_register() */
874 * mmu_notifier_put - Release the reference on the notifier
875 * @subscription: The notifier to act on
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.
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.
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.
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().
892 * Modules calling this function must call mmu_notifier_synchronize() in
893 * their __exit functions to ensure the async work is completed.
895 void mmu_notifier_put(struct mmu_notifier
*subscription
)
897 struct mm_struct
*mm
= subscription
->mm
;
899 spin_lock(&mm
->notifier_subscriptions
->lock
);
900 if (WARN_ON(!subscription
->users
) || --subscription
->users
)
902 hlist_del_init_rcu(&subscription
->hlist
);
903 spin_unlock(&mm
->notifier_subscriptions
->lock
);
905 call_srcu(&srcu
, &subscription
->rcu
, mmu_notifier_free_rcu
);
909 spin_unlock(&mm
->notifier_subscriptions
->lock
);
911 EXPORT_SYMBOL_GPL(mmu_notifier_put
);
913 static 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
)
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
;
923 * Note that the representation of the intervals in the interval tree
924 * considers the ending point as contained in the interval.
927 check_add_overflow(start
, length
- 1,
928 &interval_sub
->interval_tree
.last
))
931 /* Must call with a mmget() held */
932 if (WARN_ON(atomic_read(&mm
->mm_users
) <= 0))
935 /* pairs with mmdrop in mmu_interval_notifier_remove() */
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.
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.
948 * In all cases the value for the interval_sub->invalidate_seq should be
949 * odd, see mmu_interval_read_begin()
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
);
957 subscriptions
->invalidate_seq
|= 1;
958 interval_tree_insert(&interval_sub
->interval_tree
,
959 &subscriptions
->itree
);
961 interval_sub
->invalidate_seq
= subscriptions
->invalidate_seq
;
963 WARN_ON(mn_itree_is_invalidating(subscriptions
));
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
970 interval_sub
->invalidate_seq
=
971 subscriptions
->invalidate_seq
- 1;
972 interval_tree_insert(&interval_sub
->interval_tree
,
973 &subscriptions
->itree
);
975 spin_unlock(&subscriptions
->lock
);
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
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.
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.
995 int 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
)
1000 struct mmu_notifier_subscriptions
*subscriptions
;
1003 might_lock(&mm
->mmap_lock
);
1005 subscriptions
= smp_load_acquire(&mm
->notifier_subscriptions
);
1006 if (!subscriptions
|| !subscriptions
->has_itree
) {
1007 ret
= mmu_notifier_register(NULL
, mm
);
1010 subscriptions
= mm
->notifier_subscriptions
;
1012 return __mmu_interval_notifier_insert(interval_sub
, mm
, subscriptions
,
1013 start
, length
, ops
);
1015 EXPORT_SYMBOL_GPL(mmu_interval_notifier_insert
);
1017 int 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
)
1022 struct mmu_notifier_subscriptions
*subscriptions
=
1023 mm
->notifier_subscriptions
;
1026 mmap_assert_write_locked(mm
);
1028 if (!subscriptions
|| !subscriptions
->has_itree
) {
1029 ret
= __mmu_notifier_register(NULL
, mm
);
1032 subscriptions
= mm
->notifier_subscriptions
;
1034 return __mmu_interval_notifier_insert(interval_sub
, mm
, subscriptions
,
1035 start
, length
, ops
);
1037 EXPORT_SYMBOL_GPL(mmu_interval_notifier_insert_locked
);
1040 * mmu_interval_notifier_remove - Remove a interval notifier
1041 * @interval_sub: Interval subscription to unregister
1043 * This function must be paired with mmu_interval_notifier_insert(). It cannot
1044 * be called from any ops callback.
1046 * Once this returns ops callbacks are no longer running on other CPUs and
1047 * will not be called in future.
1049 void mmu_interval_notifier_remove(struct mmu_interval_notifier
*interval_sub
)
1051 struct mm_struct
*mm
= interval_sub
->mm
;
1052 struct mmu_notifier_subscriptions
*subscriptions
=
1053 mm
->notifier_subscriptions
;
1054 unsigned long seq
= 0;
1058 spin_lock(&subscriptions
->lock
);
1059 if (mn_itree_is_invalidating(subscriptions
)) {
1061 * remove is being called after insert put this on the
1062 * deferred list, but before the deferred list was processed.
1064 if (RB_EMPTY_NODE(&interval_sub
->interval_tree
.rb
)) {
1065 hlist_del(&interval_sub
->deferred_item
);
1067 hlist_add_head(&interval_sub
->deferred_item
,
1068 &subscriptions
->deferred_list
);
1069 seq
= subscriptions
->invalidate_seq
;
1072 WARN_ON(RB_EMPTY_NODE(&interval_sub
->interval_tree
.rb
));
1073 interval_tree_remove(&interval_sub
->interval_tree
,
1074 &subscriptions
->itree
);
1076 spin_unlock(&subscriptions
->lock
);
1079 * The possible sleep on progress in the invalidation requires the
1080 * caller not hold any locks held by invalidation callbacks.
1082 lock_map_acquire(&__mmu_notifier_invalidate_range_start_map
);
1083 lock_map_release(&__mmu_notifier_invalidate_range_start_map
);
1085 wait_event(subscriptions
->wq
,
1086 READ_ONCE(subscriptions
->invalidate_seq
) != seq
);
1088 /* pairs with mmgrab in mmu_interval_notifier_insert() */
1091 EXPORT_SYMBOL_GPL(mmu_interval_notifier_remove
);
1094 * mmu_notifier_synchronize - Ensure all mmu_notifiers are freed
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.
1100 * Before using the caller must ensure that all of its mmu_notifiers have been
1101 * fully released via mmu_notifier_put().
1103 * Modules using the mmu_notifier_put() API should call this in their __exit
1104 * function to avoid module unloading races.
1106 void mmu_notifier_synchronize(void)
1108 synchronize_srcu(&srcu
);
1110 EXPORT_SYMBOL_GPL(mmu_notifier_synchronize
);
1113 mmu_notifier_range_update_to_read_only(const struct mmu_notifier_range
*range
)
1115 if (!range
->vma
|| range
->event
!= MMU_NOTIFY_PROTECTION_VMA
)
1117 /* Return true if the vma still have the read flag set. */
1118 return range
->vma
->vm_flags
& VM_READ
;
1120 EXPORT_SYMBOL_GPL(mmu_notifier_range_update_to_read_only
);