2 * linux/mm/mmu_notifier.c
4 * Copyright (C) 2008 Qumranet, Inc.
5 * Copyright (C) 2008 SGI
6 * Christoph Lameter <cl@linux.com>
8 * This work is licensed under the terms of the GNU GPL, version 2. See
9 * the COPYING file in the top-level directory.
12 #include <linux/rculist.h>
13 #include <linux/mmu_notifier.h>
14 #include <linux/export.h>
16 #include <linux/err.h>
17 #include <linux/srcu.h>
18 #include <linux/rcupdate.h>
19 #include <linux/sched.h>
20 #include <linux/sched/mm.h>
21 #include <linux/slab.h>
23 /* global SRCU for all MMs */
24 DEFINE_STATIC_SRCU(srcu);
27 * This function allows mmu_notifier::release callback to delay a call to
28 * a function that will free appropriate resources. The function must be
29 * quick and must not block.
31 void mmu_notifier_call_srcu(struct rcu_head *rcu,
32 void (*func)(struct rcu_head *rcu))
34 call_srcu(&srcu, rcu, func);
36 EXPORT_SYMBOL_GPL(mmu_notifier_call_srcu);
38 void mmu_notifier_synchronize(void)
40 /* Wait for any running method to finish. */
43 EXPORT_SYMBOL_GPL(mmu_notifier_synchronize);
46 * This function can't run concurrently against mmu_notifier_register
47 * because mm->mm_users > 0 during mmu_notifier_register and exit_mmap
48 * runs with mm_users == 0. Other tasks may still invoke mmu notifiers
49 * in parallel despite there being no task using this mm any more,
50 * through the vmas outside of the exit_mmap context, such as with
51 * vmtruncate. This serializes against mmu_notifier_unregister with
52 * the mmu_notifier_mm->lock in addition to SRCU and it serializes
53 * against the other mmu notifiers with SRCU. struct mmu_notifier_mm
54 * can't go away from under us as exit_mmap holds an mm_count pin
57 void __mmu_notifier_release(struct mm_struct *mm)
59 struct mmu_notifier *mn;
63 * SRCU here will block mmu_notifier_unregister until
66 id = srcu_read_lock(&srcu);
67 hlist_for_each_entry_rcu(mn, &mm->mmu_notifier_mm->list, hlist)
69 * If ->release runs before mmu_notifier_unregister it must be
70 * handled, as it's the only way for the driver to flush all
71 * existing sptes and stop the driver from establishing any more
72 * sptes before all the pages in the mm are freed.
75 mn->ops->release(mn, mm);
77 spin_lock(&mm->mmu_notifier_mm->lock);
78 while (unlikely(!hlist_empty(&mm->mmu_notifier_mm->list))) {
79 mn = hlist_entry(mm->mmu_notifier_mm->list.first,
83 * We arrived before mmu_notifier_unregister so
84 * mmu_notifier_unregister will do nothing other than to wait
85 * for ->release to finish and for mmu_notifier_unregister to
88 hlist_del_init_rcu(&mn->hlist);
90 spin_unlock(&mm->mmu_notifier_mm->lock);
91 srcu_read_unlock(&srcu, id);
94 * synchronize_srcu here prevents mmu_notifier_release from returning to
95 * exit_mmap (which would proceed with freeing all pages in the mm)
96 * until the ->release method returns, if it was invoked by
97 * mmu_notifier_unregister.
99 * The mmu_notifier_mm can't go away from under us because one mm_count
100 * is held by exit_mmap.
102 synchronize_srcu(&srcu);
106 * If no young bitflag is supported by the hardware, ->clear_flush_young can
107 * unmap the address and return 1 or 0 depending if the mapping previously
110 int __mmu_notifier_clear_flush_young(struct mm_struct *mm,
114 struct mmu_notifier *mn;
117 id = srcu_read_lock(&srcu);
118 hlist_for_each_entry_rcu(mn, &mm->mmu_notifier_mm->list, hlist) {
119 if (mn->ops->clear_flush_young)
120 young |= mn->ops->clear_flush_young(mn, mm, start, end);
122 srcu_read_unlock(&srcu, id);
127 int __mmu_notifier_clear_young(struct mm_struct *mm,
131 struct mmu_notifier *mn;
134 id = srcu_read_lock(&srcu);
135 hlist_for_each_entry_rcu(mn, &mm->mmu_notifier_mm->list, hlist) {
136 if (mn->ops->clear_young)
137 young |= mn->ops->clear_young(mn, mm, start, end);
139 srcu_read_unlock(&srcu, id);
144 int __mmu_notifier_test_young(struct mm_struct *mm,
145 unsigned long address)
147 struct mmu_notifier *mn;
150 id = srcu_read_lock(&srcu);
151 hlist_for_each_entry_rcu(mn, &mm->mmu_notifier_mm->list, hlist) {
152 if (mn->ops->test_young) {
153 young = mn->ops->test_young(mn, mm, address);
158 srcu_read_unlock(&srcu, id);
163 void __mmu_notifier_change_pte(struct mm_struct *mm, unsigned long address,
166 struct mmu_notifier *mn;
169 id = srcu_read_lock(&srcu);
170 hlist_for_each_entry_rcu(mn, &mm->mmu_notifier_mm->list, hlist) {
171 if (mn->ops->change_pte)
172 mn->ops->change_pte(mn, mm, address, pte);
174 srcu_read_unlock(&srcu, id);
177 int __mmu_notifier_invalidate_range_start(struct mm_struct *mm,
178 unsigned long start, unsigned long end,
181 struct mmu_notifier *mn;
185 id = srcu_read_lock(&srcu);
186 hlist_for_each_entry_rcu(mn, &mm->mmu_notifier_mm->list, hlist) {
187 if (mn->ops->invalidate_range_start) {
188 int _ret = mn->ops->invalidate_range_start(mn, mm, start, end, blockable);
190 pr_info("%pS callback failed with %d in %sblockable context.\n",
191 mn->ops->invalidate_range_start, _ret,
192 !blockable ? "non-" : "");
197 srcu_read_unlock(&srcu, id);
201 EXPORT_SYMBOL_GPL(__mmu_notifier_invalidate_range_start);
203 void __mmu_notifier_invalidate_range_end(struct mm_struct *mm,
208 struct mmu_notifier *mn;
211 id = srcu_read_lock(&srcu);
212 hlist_for_each_entry_rcu(mn, &mm->mmu_notifier_mm->list, hlist) {
214 * Call invalidate_range here too to avoid the need for the
215 * subsystem of having to register an invalidate_range_end
216 * call-back when there is invalidate_range already. Usually a
217 * subsystem registers either invalidate_range_start()/end() or
218 * invalidate_range(), so this will be no additional overhead
219 * (besides the pointer check).
221 * We skip call to invalidate_range() if we know it is safe ie
222 * call site use mmu_notifier_invalidate_range_only_end() which
223 * is safe to do when we know that a call to invalidate_range()
224 * already happen under page table lock.
226 if (!only_end && mn->ops->invalidate_range)
227 mn->ops->invalidate_range(mn, mm, start, end);
228 if (mn->ops->invalidate_range_end)
229 mn->ops->invalidate_range_end(mn, mm, start, end);
231 srcu_read_unlock(&srcu, id);
233 EXPORT_SYMBOL_GPL(__mmu_notifier_invalidate_range_end);
235 void __mmu_notifier_invalidate_range(struct mm_struct *mm,
236 unsigned long start, unsigned long end)
238 struct mmu_notifier *mn;
241 id = srcu_read_lock(&srcu);
242 hlist_for_each_entry_rcu(mn, &mm->mmu_notifier_mm->list, hlist) {
243 if (mn->ops->invalidate_range)
244 mn->ops->invalidate_range(mn, mm, start, end);
246 srcu_read_unlock(&srcu, id);
248 EXPORT_SYMBOL_GPL(__mmu_notifier_invalidate_range);
251 * Must be called while holding mm->mmap_sem for either read or write.
252 * The result is guaranteed to be valid until mm->mmap_sem is dropped.
254 bool mm_has_blockable_invalidate_notifiers(struct mm_struct *mm)
256 struct mmu_notifier *mn;
260 WARN_ON_ONCE(!rwsem_is_locked(&mm->mmap_sem));
262 if (!mm_has_notifiers(mm))
265 id = srcu_read_lock(&srcu);
266 hlist_for_each_entry_rcu(mn, &mm->mmu_notifier_mm->list, hlist) {
267 if (!mn->ops->invalidate_range &&
268 !mn->ops->invalidate_range_start &&
269 !mn->ops->invalidate_range_end)
272 if (!(mn->ops->flags & MMU_INVALIDATE_DOES_NOT_BLOCK)) {
277 srcu_read_unlock(&srcu, id);
281 static int do_mmu_notifier_register(struct mmu_notifier *mn,
282 struct mm_struct *mm,
285 struct mmu_notifier_mm *mmu_notifier_mm;
288 BUG_ON(atomic_read(&mm->mm_users) <= 0);
291 mmu_notifier_mm = kmalloc(sizeof(struct mmu_notifier_mm), GFP_KERNEL);
292 if (unlikely(!mmu_notifier_mm))
296 down_write(&mm->mmap_sem);
297 ret = mm_take_all_locks(mm);
301 if (!mm_has_notifiers(mm)) {
302 INIT_HLIST_HEAD(&mmu_notifier_mm->list);
303 spin_lock_init(&mmu_notifier_mm->lock);
305 mm->mmu_notifier_mm = mmu_notifier_mm;
306 mmu_notifier_mm = NULL;
311 * Serialize the update against mmu_notifier_unregister. A
312 * side note: mmu_notifier_release can't run concurrently with
313 * us because we hold the mm_users pin (either implicitly as
314 * current->mm or explicitly with get_task_mm() or similar).
315 * We can't race against any other mmu notifier method either
316 * thanks to mm_take_all_locks().
318 spin_lock(&mm->mmu_notifier_mm->lock);
319 hlist_add_head(&mn->hlist, &mm->mmu_notifier_mm->list);
320 spin_unlock(&mm->mmu_notifier_mm->lock);
322 mm_drop_all_locks(mm);
325 up_write(&mm->mmap_sem);
326 kfree(mmu_notifier_mm);
328 BUG_ON(atomic_read(&mm->mm_users) <= 0);
333 * Must not hold mmap_sem nor any other VM related lock when calling
334 * this registration function. Must also ensure mm_users can't go down
335 * to zero while this runs to avoid races with mmu_notifier_release,
336 * so mm has to be current->mm or the mm should be pinned safely such
337 * as with get_task_mm(). If the mm is not current->mm, the mm_users
338 * pin should be released by calling mmput after mmu_notifier_register
339 * returns. mmu_notifier_unregister must be always called to
340 * unregister the notifier. mm_count is automatically pinned to allow
341 * mmu_notifier_unregister to safely run at any time later, before or
342 * after exit_mmap. ->release will always be called before exit_mmap
345 int mmu_notifier_register(struct mmu_notifier *mn, struct mm_struct *mm)
347 return do_mmu_notifier_register(mn, mm, 1);
349 EXPORT_SYMBOL_GPL(mmu_notifier_register);
352 * Same as mmu_notifier_register but here the caller must hold the
353 * mmap_sem in write mode.
355 int __mmu_notifier_register(struct mmu_notifier *mn, struct mm_struct *mm)
357 return do_mmu_notifier_register(mn, mm, 0);
359 EXPORT_SYMBOL_GPL(__mmu_notifier_register);
361 /* this is called after the last mmu_notifier_unregister() returned */
362 void __mmu_notifier_mm_destroy(struct mm_struct *mm)
364 BUG_ON(!hlist_empty(&mm->mmu_notifier_mm->list));
365 kfree(mm->mmu_notifier_mm);
366 mm->mmu_notifier_mm = LIST_POISON1; /* debug */
370 * This releases the mm_count pin automatically and frees the mm
371 * structure if it was the last user of it. It serializes against
372 * running mmu notifiers with SRCU and against mmu_notifier_unregister
373 * with the unregister lock + SRCU. All sptes must be dropped before
374 * calling mmu_notifier_unregister. ->release or any other notifier
375 * method may be invoked concurrently with mmu_notifier_unregister,
376 * and only after mmu_notifier_unregister returned we're guaranteed
377 * that ->release or any other method can't run anymore.
379 void mmu_notifier_unregister(struct mmu_notifier *mn, struct mm_struct *mm)
381 BUG_ON(atomic_read(&mm->mm_count) <= 0);
383 if (!hlist_unhashed(&mn->hlist)) {
385 * SRCU here will force exit_mmap to wait for ->release to
386 * finish before freeing the pages.
390 id = srcu_read_lock(&srcu);
392 * exit_mmap will block in mmu_notifier_release to guarantee
393 * that ->release is called before freeing the pages.
395 if (mn->ops->release)
396 mn->ops->release(mn, mm);
397 srcu_read_unlock(&srcu, id);
399 spin_lock(&mm->mmu_notifier_mm->lock);
401 * Can not use list_del_rcu() since __mmu_notifier_release
402 * can delete it before we hold the lock.
404 hlist_del_init_rcu(&mn->hlist);
405 spin_unlock(&mm->mmu_notifier_mm->lock);
409 * Wait for any running method to finish, of course including
410 * ->release if it was run by mmu_notifier_release instead of us.
412 synchronize_srcu(&srcu);
414 BUG_ON(atomic_read(&mm->mm_count) <= 0);
418 EXPORT_SYMBOL_GPL(mmu_notifier_unregister);
421 * Same as mmu_notifier_unregister but no callback and no srcu synchronization.
423 void mmu_notifier_unregister_no_release(struct mmu_notifier *mn,
424 struct mm_struct *mm)
426 spin_lock(&mm->mmu_notifier_mm->lock);
428 * Can not use list_del_rcu() since __mmu_notifier_release
429 * can delete it before we hold the lock.
431 hlist_del_init_rcu(&mn->hlist);
432 spin_unlock(&mm->mmu_notifier_mm->lock);
434 BUG_ON(atomic_read(&mm->mm_count) <= 0);
437 EXPORT_SYMBOL_GPL(mmu_notifier_unregister_no_release);