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 * This function can't run concurrently against mmu_notifier_register
26 * because mm->mm_users > 0 during mmu_notifier_register and exit_mmap
27 * runs with mm_users == 0. Other tasks may still invoke mmu notifiers
28 * in parallel despite there being no task using this mm any more,
29 * through the vmas outside of the exit_mmap context, such as with
30 * vmtruncate. This serializes against mmu_notifier_unregister with
31 * the mmu_notifier_mm->lock in addition to SRCU and it serializes
32 * against the other mmu notifiers with SRCU. struct mmu_notifier_mm
33 * can't go away from under us as exit_mmap holds an mm_count pin
36 void __mmu_notifier_release(struct mm_struct *mm)
38 struct mmu_notifier *mn;
42 * SRCU here will block mmu_notifier_unregister until
45 id = srcu_read_lock(&srcu);
46 hlist_for_each_entry_rcu(mn, &mm->mmu_notifier_mm->list, hlist)
48 * If ->release runs before mmu_notifier_unregister it must be
49 * handled, as it's the only way for the driver to flush all
50 * existing sptes and stop the driver from establishing any more
51 * sptes before all the pages in the mm are freed.
54 mn->ops->release(mn, mm);
56 spin_lock(&mm->mmu_notifier_mm->lock);
57 while (unlikely(!hlist_empty(&mm->mmu_notifier_mm->list))) {
58 mn = hlist_entry(mm->mmu_notifier_mm->list.first,
62 * We arrived before mmu_notifier_unregister so
63 * mmu_notifier_unregister will do nothing other than to wait
64 * for ->release to finish and for mmu_notifier_unregister to
67 hlist_del_init_rcu(&mn->hlist);
69 spin_unlock(&mm->mmu_notifier_mm->lock);
70 srcu_read_unlock(&srcu, id);
73 * synchronize_srcu here prevents mmu_notifier_release from returning to
74 * exit_mmap (which would proceed with freeing all pages in the mm)
75 * until the ->release method returns, if it was invoked by
76 * mmu_notifier_unregister.
78 * The mmu_notifier_mm can't go away from under us because one mm_count
79 * is held by exit_mmap.
81 synchronize_srcu(&srcu);
85 * If no young bitflag is supported by the hardware, ->clear_flush_young can
86 * unmap the address and return 1 or 0 depending if the mapping previously
89 int __mmu_notifier_clear_flush_young(struct mm_struct *mm,
93 struct mmu_notifier *mn;
96 id = srcu_read_lock(&srcu);
97 hlist_for_each_entry_rcu(mn, &mm->mmu_notifier_mm->list, hlist) {
98 if (mn->ops->clear_flush_young)
99 young |= mn->ops->clear_flush_young(mn, mm, start, end);
101 srcu_read_unlock(&srcu, id);
106 int __mmu_notifier_clear_young(struct mm_struct *mm,
110 struct mmu_notifier *mn;
113 id = srcu_read_lock(&srcu);
114 hlist_for_each_entry_rcu(mn, &mm->mmu_notifier_mm->list, hlist) {
115 if (mn->ops->clear_young)
116 young |= mn->ops->clear_young(mn, mm, start, end);
118 srcu_read_unlock(&srcu, id);
123 int __mmu_notifier_test_young(struct mm_struct *mm,
124 unsigned long address)
126 struct mmu_notifier *mn;
129 id = srcu_read_lock(&srcu);
130 hlist_for_each_entry_rcu(mn, &mm->mmu_notifier_mm->list, hlist) {
131 if (mn->ops->test_young) {
132 young = mn->ops->test_young(mn, mm, address);
137 srcu_read_unlock(&srcu, id);
142 void __mmu_notifier_change_pte(struct mm_struct *mm, unsigned long address,
145 struct mmu_notifier *mn;
148 id = srcu_read_lock(&srcu);
149 hlist_for_each_entry_rcu(mn, &mm->mmu_notifier_mm->list, hlist) {
150 if (mn->ops->change_pte)
151 mn->ops->change_pte(mn, mm, address, pte);
153 srcu_read_unlock(&srcu, id);
156 int __mmu_notifier_invalidate_range_start(struct mmu_notifier_range *range)
158 struct mmu_notifier *mn;
162 id = srcu_read_lock(&srcu);
163 hlist_for_each_entry_rcu(mn, &range->mm->mmu_notifier_mm->list, hlist) {
164 if (mn->ops->invalidate_range_start) {
165 int _ret = mn->ops->invalidate_range_start(mn, range);
167 pr_info("%pS callback failed with %d in %sblockable context.\n",
168 mn->ops->invalidate_range_start, _ret,
169 !mmu_notifier_range_blockable(range) ? "non-" : "");
170 WARN_ON(mmu_notifier_range_blockable(range) ||
176 srcu_read_unlock(&srcu, id);
181 void __mmu_notifier_invalidate_range_end(struct mmu_notifier_range *range,
184 struct mmu_notifier *mn;
187 id = srcu_read_lock(&srcu);
188 hlist_for_each_entry_rcu(mn, &range->mm->mmu_notifier_mm->list, hlist) {
190 * Call invalidate_range here too to avoid the need for the
191 * subsystem of having to register an invalidate_range_end
192 * call-back when there is invalidate_range already. Usually a
193 * subsystem registers either invalidate_range_start()/end() or
194 * invalidate_range(), so this will be no additional overhead
195 * (besides the pointer check).
197 * We skip call to invalidate_range() if we know it is safe ie
198 * call site use mmu_notifier_invalidate_range_only_end() which
199 * is safe to do when we know that a call to invalidate_range()
200 * already happen under page table lock.
202 if (!only_end && mn->ops->invalidate_range)
203 mn->ops->invalidate_range(mn, range->mm,
206 if (mn->ops->invalidate_range_end)
207 mn->ops->invalidate_range_end(mn, range);
209 srcu_read_unlock(&srcu, id);
212 void __mmu_notifier_invalidate_range(struct mm_struct *mm,
213 unsigned long start, unsigned long end)
215 struct mmu_notifier *mn;
218 id = srcu_read_lock(&srcu);
219 hlist_for_each_entry_rcu(mn, &mm->mmu_notifier_mm->list, hlist) {
220 if (mn->ops->invalidate_range)
221 mn->ops->invalidate_range(mn, mm, start, end);
223 srcu_read_unlock(&srcu, id);
227 * Same as mmu_notifier_register but here the caller must hold the
228 * mmap_sem in write mode.
230 int __mmu_notifier_register(struct mmu_notifier *mn, struct mm_struct *mm)
232 struct mmu_notifier_mm *mmu_notifier_mm = NULL;
235 lockdep_assert_held_write(&mm->mmap_sem);
236 BUG_ON(atomic_read(&mm->mm_users) <= 0);
241 if (!mm->mmu_notifier_mm) {
243 * kmalloc cannot be called under mm_take_all_locks(), but we
244 * know that mm->mmu_notifier_mm can't change while we hold
245 * the write side of the mmap_sem.
248 kmalloc(sizeof(struct mmu_notifier_mm), GFP_KERNEL);
249 if (!mmu_notifier_mm)
252 INIT_HLIST_HEAD(&mmu_notifier_mm->list);
253 spin_lock_init(&mmu_notifier_mm->lock);
256 ret = mm_take_all_locks(mm);
260 /* Pairs with the mmdrop in mmu_notifier_unregister_* */
264 * Serialize the update against mmu_notifier_unregister. A
265 * side note: mmu_notifier_release can't run concurrently with
266 * us because we hold the mm_users pin (either implicitly as
267 * current->mm or explicitly with get_task_mm() or similar).
268 * We can't race against any other mmu notifier method either
269 * thanks to mm_take_all_locks().
272 mm->mmu_notifier_mm = mmu_notifier_mm;
274 spin_lock(&mm->mmu_notifier_mm->lock);
275 hlist_add_head_rcu(&mn->hlist, &mm->mmu_notifier_mm->list);
276 spin_unlock(&mm->mmu_notifier_mm->lock);
278 mm_drop_all_locks(mm);
279 BUG_ON(atomic_read(&mm->mm_users) <= 0);
283 kfree(mmu_notifier_mm);
286 EXPORT_SYMBOL_GPL(__mmu_notifier_register);
289 * mmu_notifier_register - Register a notifier on a mm
290 * @mn: The notifier to attach
291 * @mm: The mm to attach the notifier to
293 * Must not hold mmap_sem nor any other VM related lock when calling
294 * this registration function. Must also ensure mm_users can't go down
295 * to zero while this runs to avoid races with mmu_notifier_release,
296 * so mm has to be current->mm or the mm should be pinned safely such
297 * as with get_task_mm(). If the mm is not current->mm, the mm_users
298 * pin should be released by calling mmput after mmu_notifier_register
301 * mmu_notifier_unregister() or mmu_notifier_put() must be always called to
302 * unregister the notifier.
304 * While the caller has a mmu_notifier get the mn->mm pointer will remain
305 * valid, and can be converted to an active mm pointer via mmget_not_zero().
307 int mmu_notifier_register(struct mmu_notifier *mn, struct mm_struct *mm)
311 down_write(&mm->mmap_sem);
312 ret = __mmu_notifier_register(mn, mm);
313 up_write(&mm->mmap_sem);
316 EXPORT_SYMBOL_GPL(mmu_notifier_register);
318 static struct mmu_notifier *
319 find_get_mmu_notifier(struct mm_struct *mm, const struct mmu_notifier_ops *ops)
321 struct mmu_notifier *mn;
323 spin_lock(&mm->mmu_notifier_mm->lock);
324 hlist_for_each_entry_rcu (mn, &mm->mmu_notifier_mm->list, hlist) {
328 if (likely(mn->users != UINT_MAX))
331 mn = ERR_PTR(-EOVERFLOW);
332 spin_unlock(&mm->mmu_notifier_mm->lock);
335 spin_unlock(&mm->mmu_notifier_mm->lock);
340 * mmu_notifier_get_locked - Return the single struct mmu_notifier for
342 * @ops: The operations struct being subscribe with
343 * @mm : The mm to attach notifiers too
345 * This function either allocates a new mmu_notifier via
346 * ops->alloc_notifier(), or returns an already existing notifier on the
347 * list. The value of the ops pointer is used to determine when two notifiers
350 * Each call to mmu_notifier_get() must be paired with a call to
351 * mmu_notifier_put(). The caller must hold the write side of mm->mmap_sem.
353 * While the caller has a mmu_notifier get the mm pointer will remain valid,
354 * and can be converted to an active mm pointer via mmget_not_zero().
356 struct mmu_notifier *mmu_notifier_get_locked(const struct mmu_notifier_ops *ops,
357 struct mm_struct *mm)
359 struct mmu_notifier *mn;
362 lockdep_assert_held_write(&mm->mmap_sem);
364 if (mm->mmu_notifier_mm) {
365 mn = find_get_mmu_notifier(mm, ops);
370 mn = ops->alloc_notifier(mm);
374 ret = __mmu_notifier_register(mn, mm);
379 mn->ops->free_notifier(mn);
382 EXPORT_SYMBOL_GPL(mmu_notifier_get_locked);
384 /* this is called after the last mmu_notifier_unregister() returned */
385 void __mmu_notifier_mm_destroy(struct mm_struct *mm)
387 BUG_ON(!hlist_empty(&mm->mmu_notifier_mm->list));
388 kfree(mm->mmu_notifier_mm);
389 mm->mmu_notifier_mm = LIST_POISON1; /* debug */
393 * This releases the mm_count pin automatically and frees the mm
394 * structure if it was the last user of it. It serializes against
395 * running mmu notifiers with SRCU and against mmu_notifier_unregister
396 * with the unregister lock + SRCU. All sptes must be dropped before
397 * calling mmu_notifier_unregister. ->release or any other notifier
398 * method may be invoked concurrently with mmu_notifier_unregister,
399 * and only after mmu_notifier_unregister returned we're guaranteed
400 * that ->release or any other method can't run anymore.
402 void mmu_notifier_unregister(struct mmu_notifier *mn, struct mm_struct *mm)
404 BUG_ON(atomic_read(&mm->mm_count) <= 0);
406 if (!hlist_unhashed(&mn->hlist)) {
408 * SRCU here will force exit_mmap to wait for ->release to
409 * finish before freeing the pages.
413 id = srcu_read_lock(&srcu);
415 * exit_mmap will block in mmu_notifier_release to guarantee
416 * that ->release is called before freeing the pages.
418 if (mn->ops->release)
419 mn->ops->release(mn, mm);
420 srcu_read_unlock(&srcu, id);
422 spin_lock(&mm->mmu_notifier_mm->lock);
424 * Can not use list_del_rcu() since __mmu_notifier_release
425 * can delete it before we hold the lock.
427 hlist_del_init_rcu(&mn->hlist);
428 spin_unlock(&mm->mmu_notifier_mm->lock);
432 * Wait for any running method to finish, of course including
433 * ->release if it was run by mmu_notifier_release instead of us.
435 synchronize_srcu(&srcu);
437 BUG_ON(atomic_read(&mm->mm_count) <= 0);
441 EXPORT_SYMBOL_GPL(mmu_notifier_unregister);
443 static void mmu_notifier_free_rcu(struct rcu_head *rcu)
445 struct mmu_notifier *mn = container_of(rcu, struct mmu_notifier, rcu);
446 struct mm_struct *mm = mn->mm;
448 mn->ops->free_notifier(mn);
449 /* Pairs with the get in __mmu_notifier_register() */
454 * mmu_notifier_put - Release the reference on the notifier
455 * @mn: The notifier to act on
457 * This function must be paired with each mmu_notifier_get(), it releases the
458 * reference obtained by the get. If this is the last reference then process
459 * to free the notifier will be run asynchronously.
461 * Unlike mmu_notifier_unregister() the get/put flow only calls ops->release
462 * when the mm_struct is destroyed. Instead free_notifier is always called to
463 * release any resources held by the user.
465 * As ops->release is not guaranteed to be called, the user must ensure that
466 * all sptes are dropped, and no new sptes can be established before
467 * mmu_notifier_put() is called.
469 * This function can be called from the ops->release callback, however the
470 * caller must still ensure it is called pairwise with mmu_notifier_get().
472 * Modules calling this function must call mmu_notifier_synchronize() in
473 * their __exit functions to ensure the async work is completed.
475 void mmu_notifier_put(struct mmu_notifier *mn)
477 struct mm_struct *mm = mn->mm;
479 spin_lock(&mm->mmu_notifier_mm->lock);
480 if (WARN_ON(!mn->users) || --mn->users)
482 hlist_del_init_rcu(&mn->hlist);
483 spin_unlock(&mm->mmu_notifier_mm->lock);
485 call_srcu(&srcu, &mn->rcu, mmu_notifier_free_rcu);
489 spin_unlock(&mm->mmu_notifier_mm->lock);
491 EXPORT_SYMBOL_GPL(mmu_notifier_put);
494 * mmu_notifier_synchronize - Ensure all mmu_notifiers are freed
496 * This function ensures that all outstanding async SRU work from
497 * mmu_notifier_put() is completed. After it returns any mmu_notifier_ops
498 * associated with an unused mmu_notifier will no longer be called.
500 * Before using the caller must ensure that all of its mmu_notifiers have been
501 * fully released via mmu_notifier_put().
503 * Modules using the mmu_notifier_put() API should call this in their __exit
504 * function to avoid module unloading races.
506 void mmu_notifier_synchronize(void)
508 synchronize_srcu(&srcu);
510 EXPORT_SYMBOL_GPL(mmu_notifier_synchronize);
513 mmu_notifier_range_update_to_read_only(const struct mmu_notifier_range *range)
515 if (!range->vma || range->event != MMU_NOTIFY_PROTECTION_VMA)
517 /* Return true if the vma still have the read flag set. */
518 return range->vma->vm_flags & VM_READ;
520 EXPORT_SYMBOL_GPL(mmu_notifier_range_update_to_read_only);