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/srcu.h>
16 #include <linux/rcupdate.h>
17 #include <linux/sched.h>
18 #include <linux/sched/mm.h>
19 #include <linux/slab.h>
21 /* global SRCU for all MMs */
22 DEFINE_STATIC_SRCU(srcu);
25 struct lockdep_map __mmu_notifier_invalidate_range_start_map = {
26 .name = "mmu_notifier_invalidate_range_start"
31 * This function can't run concurrently against mmu_notifier_register
32 * because mm->mm_users > 0 during mmu_notifier_register and exit_mmap
33 * runs with mm_users == 0. Other tasks may still invoke mmu notifiers
34 * in parallel despite there being no task using this mm any more,
35 * through the vmas outside of the exit_mmap context, such as with
36 * vmtruncate. This serializes against mmu_notifier_unregister with
37 * the mmu_notifier_mm->lock in addition to SRCU and it serializes
38 * against the other mmu notifiers with SRCU. struct mmu_notifier_mm
39 * can't go away from under us as exit_mmap holds an mm_count pin
42 void __mmu_notifier_release(struct mm_struct *mm)
44 struct mmu_notifier *mn;
48 * SRCU here will block mmu_notifier_unregister until
51 id = srcu_read_lock(&srcu);
52 hlist_for_each_entry_rcu(mn, &mm->mmu_notifier_mm->list, hlist)
54 * If ->release runs before mmu_notifier_unregister it must be
55 * handled, as it's the only way for the driver to flush all
56 * existing sptes and stop the driver from establishing any more
57 * sptes before all the pages in the mm are freed.
60 mn->ops->release(mn, mm);
62 spin_lock(&mm->mmu_notifier_mm->lock);
63 while (unlikely(!hlist_empty(&mm->mmu_notifier_mm->list))) {
64 mn = hlist_entry(mm->mmu_notifier_mm->list.first,
68 * We arrived before mmu_notifier_unregister so
69 * mmu_notifier_unregister will do nothing other than to wait
70 * for ->release to finish and for mmu_notifier_unregister to
73 hlist_del_init_rcu(&mn->hlist);
75 spin_unlock(&mm->mmu_notifier_mm->lock);
76 srcu_read_unlock(&srcu, id);
79 * synchronize_srcu here prevents mmu_notifier_release from returning to
80 * exit_mmap (which would proceed with freeing all pages in the mm)
81 * until the ->release method returns, if it was invoked by
82 * mmu_notifier_unregister.
84 * The mmu_notifier_mm can't go away from under us because one mm_count
85 * is held by exit_mmap.
87 synchronize_srcu(&srcu);
91 * If no young bitflag is supported by the hardware, ->clear_flush_young can
92 * unmap the address and return 1 or 0 depending if the mapping previously
95 int __mmu_notifier_clear_flush_young(struct mm_struct *mm,
99 struct mmu_notifier *mn;
102 id = srcu_read_lock(&srcu);
103 hlist_for_each_entry_rcu(mn, &mm->mmu_notifier_mm->list, hlist) {
104 if (mn->ops->clear_flush_young)
105 young |= mn->ops->clear_flush_young(mn, mm, start, end);
107 srcu_read_unlock(&srcu, id);
112 int __mmu_notifier_clear_young(struct mm_struct *mm,
116 struct mmu_notifier *mn;
119 id = srcu_read_lock(&srcu);
120 hlist_for_each_entry_rcu(mn, &mm->mmu_notifier_mm->list, hlist) {
121 if (mn->ops->clear_young)
122 young |= mn->ops->clear_young(mn, mm, start, end);
124 srcu_read_unlock(&srcu, id);
129 int __mmu_notifier_test_young(struct mm_struct *mm,
130 unsigned long address)
132 struct mmu_notifier *mn;
135 id = srcu_read_lock(&srcu);
136 hlist_for_each_entry_rcu(mn, &mm->mmu_notifier_mm->list, hlist) {
137 if (mn->ops->test_young) {
138 young = mn->ops->test_young(mn, mm, address);
143 srcu_read_unlock(&srcu, id);
148 void __mmu_notifier_change_pte(struct mm_struct *mm, unsigned long address,
151 struct mmu_notifier *mn;
154 id = srcu_read_lock(&srcu);
155 hlist_for_each_entry_rcu(mn, &mm->mmu_notifier_mm->list, hlist) {
156 if (mn->ops->change_pte)
157 mn->ops->change_pte(mn, mm, address, pte);
159 srcu_read_unlock(&srcu, id);
162 int __mmu_notifier_invalidate_range_start(struct mmu_notifier_range *range)
164 struct mmu_notifier *mn;
168 id = srcu_read_lock(&srcu);
169 hlist_for_each_entry_rcu(mn, &range->mm->mmu_notifier_mm->list, hlist) {
170 if (mn->ops->invalidate_range_start) {
171 int _ret = mn->ops->invalidate_range_start(mn, range);
173 pr_info("%pS callback failed with %d in %sblockable context.\n",
174 mn->ops->invalidate_range_start, _ret,
175 !mmu_notifier_range_blockable(range) ? "non-" : "");
176 WARN_ON(mmu_notifier_range_blockable(range) ||
182 srcu_read_unlock(&srcu, id);
187 void __mmu_notifier_invalidate_range_end(struct mmu_notifier_range *range,
190 struct mmu_notifier *mn;
193 lock_map_acquire(&__mmu_notifier_invalidate_range_start_map);
194 id = srcu_read_lock(&srcu);
195 hlist_for_each_entry_rcu(mn, &range->mm->mmu_notifier_mm->list, hlist) {
197 * Call invalidate_range here too to avoid the need for the
198 * subsystem of having to register an invalidate_range_end
199 * call-back when there is invalidate_range already. Usually a
200 * subsystem registers either invalidate_range_start()/end() or
201 * invalidate_range(), so this will be no additional overhead
202 * (besides the pointer check).
204 * We skip call to invalidate_range() if we know it is safe ie
205 * call site use mmu_notifier_invalidate_range_only_end() which
206 * is safe to do when we know that a call to invalidate_range()
207 * already happen under page table lock.
209 if (!only_end && mn->ops->invalidate_range)
210 mn->ops->invalidate_range(mn, range->mm,
213 if (mn->ops->invalidate_range_end)
214 mn->ops->invalidate_range_end(mn, range);
216 srcu_read_unlock(&srcu, id);
217 lock_map_release(&__mmu_notifier_invalidate_range_start_map);
220 void __mmu_notifier_invalidate_range(struct mm_struct *mm,
221 unsigned long start, unsigned long end)
223 struct mmu_notifier *mn;
226 id = srcu_read_lock(&srcu);
227 hlist_for_each_entry_rcu(mn, &mm->mmu_notifier_mm->list, hlist) {
228 if (mn->ops->invalidate_range)
229 mn->ops->invalidate_range(mn, mm, start, end);
231 srcu_read_unlock(&srcu, id);
235 * Same as mmu_notifier_register but here the caller must hold the
236 * mmap_sem in write mode.
238 int __mmu_notifier_register(struct mmu_notifier *mn, struct mm_struct *mm)
240 struct mmu_notifier_mm *mmu_notifier_mm = NULL;
243 lockdep_assert_held_write(&mm->mmap_sem);
244 BUG_ON(atomic_read(&mm->mm_users) <= 0);
246 if (IS_ENABLED(CONFIG_LOCKDEP)) {
247 fs_reclaim_acquire(GFP_KERNEL);
248 lock_map_acquire(&__mmu_notifier_invalidate_range_start_map);
249 lock_map_release(&__mmu_notifier_invalidate_range_start_map);
250 fs_reclaim_release(GFP_KERNEL);
256 if (!mm->mmu_notifier_mm) {
258 * kmalloc cannot be called under mm_take_all_locks(), but we
259 * know that mm->mmu_notifier_mm can't change while we hold
260 * the write side of the mmap_sem.
263 kmalloc(sizeof(struct mmu_notifier_mm), GFP_KERNEL);
264 if (!mmu_notifier_mm)
267 INIT_HLIST_HEAD(&mmu_notifier_mm->list);
268 spin_lock_init(&mmu_notifier_mm->lock);
271 ret = mm_take_all_locks(mm);
275 /* Pairs with the mmdrop in mmu_notifier_unregister_* */
279 * Serialize the update against mmu_notifier_unregister. A
280 * side note: mmu_notifier_release can't run concurrently with
281 * us because we hold the mm_users pin (either implicitly as
282 * current->mm or explicitly with get_task_mm() or similar).
283 * We can't race against any other mmu notifier method either
284 * thanks to mm_take_all_locks().
287 mm->mmu_notifier_mm = mmu_notifier_mm;
289 spin_lock(&mm->mmu_notifier_mm->lock);
290 hlist_add_head_rcu(&mn->hlist, &mm->mmu_notifier_mm->list);
291 spin_unlock(&mm->mmu_notifier_mm->lock);
293 mm_drop_all_locks(mm);
294 BUG_ON(atomic_read(&mm->mm_users) <= 0);
298 kfree(mmu_notifier_mm);
301 EXPORT_SYMBOL_GPL(__mmu_notifier_register);
304 * mmu_notifier_register - Register a notifier on a mm
305 * @mn: The notifier to attach
306 * @mm: The mm to attach the notifier to
308 * Must not hold mmap_sem nor any other VM related lock when calling
309 * this registration function. Must also ensure mm_users can't go down
310 * to zero while this runs to avoid races with mmu_notifier_release,
311 * so mm has to be current->mm or the mm should be pinned safely such
312 * as with get_task_mm(). If the mm is not current->mm, the mm_users
313 * pin should be released by calling mmput after mmu_notifier_register
316 * mmu_notifier_unregister() or mmu_notifier_put() must be always called to
317 * unregister the notifier.
319 * While the caller has a mmu_notifier get the mn->mm pointer will remain
320 * valid, and can be converted to an active mm pointer via mmget_not_zero().
322 int mmu_notifier_register(struct mmu_notifier *mn, struct mm_struct *mm)
326 down_write(&mm->mmap_sem);
327 ret = __mmu_notifier_register(mn, mm);
328 up_write(&mm->mmap_sem);
331 EXPORT_SYMBOL_GPL(mmu_notifier_register);
333 static struct mmu_notifier *
334 find_get_mmu_notifier(struct mm_struct *mm, const struct mmu_notifier_ops *ops)
336 struct mmu_notifier *mn;
338 spin_lock(&mm->mmu_notifier_mm->lock);
339 hlist_for_each_entry_rcu (mn, &mm->mmu_notifier_mm->list, hlist) {
343 if (likely(mn->users != UINT_MAX))
346 mn = ERR_PTR(-EOVERFLOW);
347 spin_unlock(&mm->mmu_notifier_mm->lock);
350 spin_unlock(&mm->mmu_notifier_mm->lock);
355 * mmu_notifier_get_locked - Return the single struct mmu_notifier for
357 * @ops: The operations struct being subscribe with
358 * @mm : The mm to attach notifiers too
360 * This function either allocates a new mmu_notifier via
361 * ops->alloc_notifier(), or returns an already existing notifier on the
362 * list. The value of the ops pointer is used to determine when two notifiers
365 * Each call to mmu_notifier_get() must be paired with a call to
366 * mmu_notifier_put(). The caller must hold the write side of mm->mmap_sem.
368 * While the caller has a mmu_notifier get the mm pointer will remain valid,
369 * and can be converted to an active mm pointer via mmget_not_zero().
371 struct mmu_notifier *mmu_notifier_get_locked(const struct mmu_notifier_ops *ops,
372 struct mm_struct *mm)
374 struct mmu_notifier *mn;
377 lockdep_assert_held_write(&mm->mmap_sem);
379 if (mm->mmu_notifier_mm) {
380 mn = find_get_mmu_notifier(mm, ops);
385 mn = ops->alloc_notifier(mm);
389 ret = __mmu_notifier_register(mn, mm);
394 mn->ops->free_notifier(mn);
397 EXPORT_SYMBOL_GPL(mmu_notifier_get_locked);
399 /* this is called after the last mmu_notifier_unregister() returned */
400 void __mmu_notifier_mm_destroy(struct mm_struct *mm)
402 BUG_ON(!hlist_empty(&mm->mmu_notifier_mm->list));
403 kfree(mm->mmu_notifier_mm);
404 mm->mmu_notifier_mm = LIST_POISON1; /* debug */
408 * This releases the mm_count pin automatically and frees the mm
409 * structure if it was the last user of it. It serializes against
410 * running mmu notifiers with SRCU and against mmu_notifier_unregister
411 * with the unregister lock + SRCU. All sptes must be dropped before
412 * calling mmu_notifier_unregister. ->release or any other notifier
413 * method may be invoked concurrently with mmu_notifier_unregister,
414 * and only after mmu_notifier_unregister returned we're guaranteed
415 * that ->release or any other method can't run anymore.
417 void mmu_notifier_unregister(struct mmu_notifier *mn, struct mm_struct *mm)
419 BUG_ON(atomic_read(&mm->mm_count) <= 0);
421 if (!hlist_unhashed(&mn->hlist)) {
423 * SRCU here will force exit_mmap to wait for ->release to
424 * finish before freeing the pages.
428 id = srcu_read_lock(&srcu);
430 * exit_mmap will block in mmu_notifier_release to guarantee
431 * that ->release is called before freeing the pages.
433 if (mn->ops->release)
434 mn->ops->release(mn, mm);
435 srcu_read_unlock(&srcu, id);
437 spin_lock(&mm->mmu_notifier_mm->lock);
439 * Can not use list_del_rcu() since __mmu_notifier_release
440 * can delete it before we hold the lock.
442 hlist_del_init_rcu(&mn->hlist);
443 spin_unlock(&mm->mmu_notifier_mm->lock);
447 * Wait for any running method to finish, of course including
448 * ->release if it was run by mmu_notifier_release instead of us.
450 synchronize_srcu(&srcu);
452 BUG_ON(atomic_read(&mm->mm_count) <= 0);
456 EXPORT_SYMBOL_GPL(mmu_notifier_unregister);
458 static void mmu_notifier_free_rcu(struct rcu_head *rcu)
460 struct mmu_notifier *mn = container_of(rcu, struct mmu_notifier, rcu);
461 struct mm_struct *mm = mn->mm;
463 mn->ops->free_notifier(mn);
464 /* Pairs with the get in __mmu_notifier_register() */
469 * mmu_notifier_put - Release the reference on the notifier
470 * @mn: The notifier to act on
472 * This function must be paired with each mmu_notifier_get(), it releases the
473 * reference obtained by the get. If this is the last reference then process
474 * to free the notifier will be run asynchronously.
476 * Unlike mmu_notifier_unregister() the get/put flow only calls ops->release
477 * when the mm_struct is destroyed. Instead free_notifier is always called to
478 * release any resources held by the user.
480 * As ops->release is not guaranteed to be called, the user must ensure that
481 * all sptes are dropped, and no new sptes can be established before
482 * mmu_notifier_put() is called.
484 * This function can be called from the ops->release callback, however the
485 * caller must still ensure it is called pairwise with mmu_notifier_get().
487 * Modules calling this function must call mmu_notifier_synchronize() in
488 * their __exit functions to ensure the async work is completed.
490 void mmu_notifier_put(struct mmu_notifier *mn)
492 struct mm_struct *mm = mn->mm;
494 spin_lock(&mm->mmu_notifier_mm->lock);
495 if (WARN_ON(!mn->users) || --mn->users)
497 hlist_del_init_rcu(&mn->hlist);
498 spin_unlock(&mm->mmu_notifier_mm->lock);
500 call_srcu(&srcu, &mn->rcu, mmu_notifier_free_rcu);
504 spin_unlock(&mm->mmu_notifier_mm->lock);
506 EXPORT_SYMBOL_GPL(mmu_notifier_put);
509 * mmu_notifier_synchronize - Ensure all mmu_notifiers are freed
511 * This function ensures that all outstanding async SRU work from
512 * mmu_notifier_put() is completed. After it returns any mmu_notifier_ops
513 * associated with an unused mmu_notifier will no longer be called.
515 * Before using the caller must ensure that all of its mmu_notifiers have been
516 * fully released via mmu_notifier_put().
518 * Modules using the mmu_notifier_put() API should call this in their __exit
519 * function to avoid module unloading races.
521 void mmu_notifier_synchronize(void)
523 synchronize_srcu(&srcu);
525 EXPORT_SYMBOL_GPL(mmu_notifier_synchronize);
528 mmu_notifier_range_update_to_read_only(const struct mmu_notifier_range *range)
530 if (!range->vma || range->event != MMU_NOTIFY_PROTECTION_VMA)
532 /* Return true if the vma still have the read flag set. */
533 return range->vma->vm_flags & VM_READ;
535 EXPORT_SYMBOL_GPL(mmu_notifier_range_update_to_read_only);