Linux Audio

Check our new training course

Loading...
v3.1
  1/*
  2 *  linux/mm/mmu_notifier.c
  3 *
  4 *  Copyright (C) 2008  Qumranet, Inc.
  5 *  Copyright (C) 2008  SGI
  6 *             Christoph Lameter <clameter@sgi.com>
  7 *
  8 *  This work is licensed under the terms of the GNU GPL, version 2. See
  9 *  the COPYING file in the top-level directory.
 10 */
 11
 12#include <linux/rculist.h>
 13#include <linux/mmu_notifier.h>
 14#include <linux/module.h>
 15#include <linux/mm.h>
 16#include <linux/err.h>
 
 17#include <linux/rcupdate.h>
 18#include <linux/sched.h>
 
 19#include <linux/slab.h>
 20
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 21/*
 22 * This function can't run concurrently against mmu_notifier_register
 23 * because mm->mm_users > 0 during mmu_notifier_register and exit_mmap
 24 * runs with mm_users == 0. Other tasks may still invoke mmu notifiers
 25 * in parallel despite there being no task using this mm any more,
 26 * through the vmas outside of the exit_mmap context, such as with
 27 * vmtruncate. This serializes against mmu_notifier_unregister with
 28 * the mmu_notifier_mm->lock in addition to RCU and it serializes
 29 * against the other mmu notifiers with RCU. struct mmu_notifier_mm
 30 * can't go away from under us as exit_mmap holds an mm_count pin
 31 * itself.
 32 */
 33void __mmu_notifier_release(struct mm_struct *mm)
 34{
 35	struct mmu_notifier *mn;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 36
 37	spin_lock(&mm->mmu_notifier_mm->lock);
 38	while (unlikely(!hlist_empty(&mm->mmu_notifier_mm->list))) {
 39		mn = hlist_entry(mm->mmu_notifier_mm->list.first,
 40				 struct mmu_notifier,
 41				 hlist);
 42		/*
 43		 * We arrived before mmu_notifier_unregister so
 44		 * mmu_notifier_unregister will do nothing other than
 45		 * to wait ->release to finish and
 46		 * mmu_notifier_unregister to return.
 47		 */
 48		hlist_del_init_rcu(&mn->hlist);
 49		/*
 50		 * RCU here will block mmu_notifier_unregister until
 51		 * ->release returns.
 52		 */
 53		rcu_read_lock();
 54		spin_unlock(&mm->mmu_notifier_mm->lock);
 55		/*
 56		 * if ->release runs before mmu_notifier_unregister it
 57		 * must be handled as it's the only way for the driver
 58		 * to flush all existing sptes and stop the driver
 59		 * from establishing any more sptes before all the
 60		 * pages in the mm are freed.
 61		 */
 62		if (mn->ops->release)
 63			mn->ops->release(mn, mm);
 64		rcu_read_unlock();
 65		spin_lock(&mm->mmu_notifier_mm->lock);
 66	}
 67	spin_unlock(&mm->mmu_notifier_mm->lock);
 
 68
 69	/*
 70	 * synchronize_rcu here prevents mmu_notifier_release to
 71	 * return to exit_mmap (which would proceed freeing all pages
 72	 * in the mm) until the ->release method returns, if it was
 73	 * invoked by mmu_notifier_unregister.
 74	 *
 75	 * The mmu_notifier_mm can't go away from under us because one
 76	 * mm_count is hold by exit_mmap.
 77	 */
 78	synchronize_rcu();
 79}
 80
 81/*
 82 * If no young bitflag is supported by the hardware, ->clear_flush_young can
 83 * unmap the address and return 1 or 0 depending if the mapping previously
 84 * existed or not.
 85 */
 86int __mmu_notifier_clear_flush_young(struct mm_struct *mm,
 87					unsigned long address)
 
 88{
 89	struct mmu_notifier *mn;
 90	struct hlist_node *n;
 91	int young = 0;
 92
 93	rcu_read_lock();
 94	hlist_for_each_entry_rcu(mn, n, &mm->mmu_notifier_mm->list, hlist) {
 95		if (mn->ops->clear_flush_young)
 96			young |= mn->ops->clear_flush_young(mn, mm, address);
 97	}
 98	rcu_read_unlock();
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 99
100	return young;
101}
102
103int __mmu_notifier_test_young(struct mm_struct *mm,
104			      unsigned long address)
105{
106	struct mmu_notifier *mn;
107	struct hlist_node *n;
108	int young = 0;
109
110	rcu_read_lock();
111	hlist_for_each_entry_rcu(mn, n, &mm->mmu_notifier_mm->list, hlist) {
112		if (mn->ops->test_young) {
113			young = mn->ops->test_young(mn, mm, address);
114			if (young)
115				break;
116		}
117	}
118	rcu_read_unlock();
119
120	return young;
121}
122
123void __mmu_notifier_change_pte(struct mm_struct *mm, unsigned long address,
124			       pte_t pte)
125{
126	struct mmu_notifier *mn;
127	struct hlist_node *n;
128
129	rcu_read_lock();
130	hlist_for_each_entry_rcu(mn, n, &mm->mmu_notifier_mm->list, hlist) {
131		if (mn->ops->change_pte)
132			mn->ops->change_pte(mn, mm, address, pte);
133		/*
134		 * Some drivers don't have change_pte,
135		 * so we must call invalidate_page in that case.
136		 */
137		else if (mn->ops->invalidate_page)
138			mn->ops->invalidate_page(mn, mm, address);
139	}
140	rcu_read_unlock();
141}
142
143void __mmu_notifier_invalidate_page(struct mm_struct *mm,
144					  unsigned long address)
145{
146	struct mmu_notifier *mn;
147	struct hlist_node *n;
148
149	rcu_read_lock();
150	hlist_for_each_entry_rcu(mn, n, &mm->mmu_notifier_mm->list, hlist) {
151		if (mn->ops->invalidate_page)
152			mn->ops->invalidate_page(mn, mm, address);
153	}
154	rcu_read_unlock();
155}
156
157void __mmu_notifier_invalidate_range_start(struct mm_struct *mm,
158				  unsigned long start, unsigned long end)
159{
160	struct mmu_notifier *mn;
161	struct hlist_node *n;
162
163	rcu_read_lock();
164	hlist_for_each_entry_rcu(mn, n, &mm->mmu_notifier_mm->list, hlist) {
165		if (mn->ops->invalidate_range_start)
166			mn->ops->invalidate_range_start(mn, mm, start, end);
167	}
168	rcu_read_unlock();
169}
 
170
171void __mmu_notifier_invalidate_range_end(struct mm_struct *mm,
172				  unsigned long start, unsigned long end)
 
 
173{
174	struct mmu_notifier *mn;
175	struct hlist_node *n;
176
177	rcu_read_lock();
178	hlist_for_each_entry_rcu(mn, n, &mm->mmu_notifier_mm->list, hlist) {
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
179		if (mn->ops->invalidate_range_end)
180			mn->ops->invalidate_range_end(mn, mm, start, end);
181	}
182	rcu_read_unlock();
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
183}
184
185static int do_mmu_notifier_register(struct mmu_notifier *mn,
186				    struct mm_struct *mm,
187				    int take_mmap_sem)
188{
189	struct mmu_notifier_mm *mmu_notifier_mm;
190	int ret;
191
192	BUG_ON(atomic_read(&mm->mm_users) <= 0);
193
194	ret = -ENOMEM;
195	mmu_notifier_mm = kmalloc(sizeof(struct mmu_notifier_mm), GFP_KERNEL);
196	if (unlikely(!mmu_notifier_mm))
197		goto out;
198
199	if (take_mmap_sem)
200		down_write(&mm->mmap_sem);
201	ret = mm_take_all_locks(mm);
202	if (unlikely(ret))
203		goto out_cleanup;
204
205	if (!mm_has_notifiers(mm)) {
206		INIT_HLIST_HEAD(&mmu_notifier_mm->list);
207		spin_lock_init(&mmu_notifier_mm->lock);
 
208		mm->mmu_notifier_mm = mmu_notifier_mm;
209		mmu_notifier_mm = NULL;
210	}
211	atomic_inc(&mm->mm_count);
212
213	/*
214	 * Serialize the update against mmu_notifier_unregister. A
215	 * side note: mmu_notifier_release can't run concurrently with
216	 * us because we hold the mm_users pin (either implicitly as
217	 * current->mm or explicitly with get_task_mm() or similar).
218	 * We can't race against any other mmu notifier method either
219	 * thanks to mm_take_all_locks().
220	 */
221	spin_lock(&mm->mmu_notifier_mm->lock);
222	hlist_add_head(&mn->hlist, &mm->mmu_notifier_mm->list);
223	spin_unlock(&mm->mmu_notifier_mm->lock);
224
225	mm_drop_all_locks(mm);
226out_cleanup:
227	if (take_mmap_sem)
228		up_write(&mm->mmap_sem);
229	/* kfree() does nothing if mmu_notifier_mm is NULL */
230	kfree(mmu_notifier_mm);
231out:
232	BUG_ON(atomic_read(&mm->mm_users) <= 0);
233	return ret;
234}
235
236/*
237 * Must not hold mmap_sem nor any other VM related lock when calling
238 * this registration function. Must also ensure mm_users can't go down
239 * to zero while this runs to avoid races with mmu_notifier_release,
240 * so mm has to be current->mm or the mm should be pinned safely such
241 * as with get_task_mm(). If the mm is not current->mm, the mm_users
242 * pin should be released by calling mmput after mmu_notifier_register
243 * returns. mmu_notifier_unregister must be always called to
244 * unregister the notifier. mm_count is automatically pinned to allow
245 * mmu_notifier_unregister to safely run at any time later, before or
246 * after exit_mmap. ->release will always be called before exit_mmap
247 * frees the pages.
248 */
249int mmu_notifier_register(struct mmu_notifier *mn, struct mm_struct *mm)
250{
251	return do_mmu_notifier_register(mn, mm, 1);
252}
253EXPORT_SYMBOL_GPL(mmu_notifier_register);
254
255/*
256 * Same as mmu_notifier_register but here the caller must hold the
257 * mmap_sem in write mode.
258 */
259int __mmu_notifier_register(struct mmu_notifier *mn, struct mm_struct *mm)
260{
261	return do_mmu_notifier_register(mn, mm, 0);
262}
263EXPORT_SYMBOL_GPL(__mmu_notifier_register);
264
265/* this is called after the last mmu_notifier_unregister() returned */
266void __mmu_notifier_mm_destroy(struct mm_struct *mm)
267{
268	BUG_ON(!hlist_empty(&mm->mmu_notifier_mm->list));
269	kfree(mm->mmu_notifier_mm);
270	mm->mmu_notifier_mm = LIST_POISON1; /* debug */
271}
272
273/*
274 * This releases the mm_count pin automatically and frees the mm
275 * structure if it was the last user of it. It serializes against
276 * running mmu notifiers with RCU and against mmu_notifier_unregister
277 * with the unregister lock + RCU. All sptes must be dropped before
278 * calling mmu_notifier_unregister. ->release or any other notifier
279 * method may be invoked concurrently with mmu_notifier_unregister,
280 * and only after mmu_notifier_unregister returned we're guaranteed
281 * that ->release or any other method can't run anymore.
282 */
283void mmu_notifier_unregister(struct mmu_notifier *mn, struct mm_struct *mm)
284{
285	BUG_ON(atomic_read(&mm->mm_count) <= 0);
286
287	spin_lock(&mm->mmu_notifier_mm->lock);
288	if (!hlist_unhashed(&mn->hlist)) {
289		hlist_del_rcu(&mn->hlist);
290
291		/*
292		 * RCU here will force exit_mmap to wait ->release to finish
293		 * before freeing the pages.
294		 */
295		rcu_read_lock();
296		spin_unlock(&mm->mmu_notifier_mm->lock);
 
297		/*
298		 * exit_mmap will block in mmu_notifier_release to
299		 * guarantee ->release is called before freeing the
300		 * pages.
301		 */
302		if (mn->ops->release)
303			mn->ops->release(mn, mm);
304		rcu_read_unlock();
305	} else
 
 
 
 
 
 
306		spin_unlock(&mm->mmu_notifier_mm->lock);
 
307
308	/*
309	 * Wait any running method to finish, of course including
310	 * ->release if it was run by mmu_notifier_relase instead of us.
311	 */
312	synchronize_rcu();
313
314	BUG_ON(atomic_read(&mm->mm_count) <= 0);
315
316	mmdrop(mm);
317}
318EXPORT_SYMBOL_GPL(mmu_notifier_unregister);
v4.17
  1/*
  2 *  linux/mm/mmu_notifier.c
  3 *
  4 *  Copyright (C) 2008  Qumranet, Inc.
  5 *  Copyright (C) 2008  SGI
  6 *             Christoph Lameter <cl@linux.com>
  7 *
  8 *  This work is licensed under the terms of the GNU GPL, version 2. See
  9 *  the COPYING file in the top-level directory.
 10 */
 11
 12#include <linux/rculist.h>
 13#include <linux/mmu_notifier.h>
 14#include <linux/export.h>
 15#include <linux/mm.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>
 22
 23/* global SRCU for all MMs */
 24DEFINE_STATIC_SRCU(srcu);
 25
 26/*
 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.
 30 */
 31void mmu_notifier_call_srcu(struct rcu_head *rcu,
 32			    void (*func)(struct rcu_head *rcu))
 33{
 34	call_srcu(&srcu, rcu, func);
 35}
 36EXPORT_SYMBOL_GPL(mmu_notifier_call_srcu);
 37
 38void mmu_notifier_synchronize(void)
 39{
 40	/* Wait for any running method to finish. */
 41	srcu_barrier(&srcu);
 42}
 43EXPORT_SYMBOL_GPL(mmu_notifier_synchronize);
 44
 45/*
 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
 55 * itself.
 56 */
 57void __mmu_notifier_release(struct mm_struct *mm)
 58{
 59	struct mmu_notifier *mn;
 60	int id;
 61
 62	/*
 63	 * SRCU here will block mmu_notifier_unregister until
 64	 * ->release returns.
 65	 */
 66	id = srcu_read_lock(&srcu);
 67	hlist_for_each_entry_rcu(mn, &mm->mmu_notifier_mm->list, hlist)
 68		/*
 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.
 73		 */
 74		if (mn->ops->release)
 75			mn->ops->release(mn, mm);
 76
 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,
 80				 struct mmu_notifier,
 81				 hlist);
 82		/*
 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
 86		 * return.
 87		 */
 88		hlist_del_init_rcu(&mn->hlist);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 89	}
 90	spin_unlock(&mm->mmu_notifier_mm->lock);
 91	srcu_read_unlock(&srcu, id);
 92
 93	/*
 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.
 98	 *
 99	 * The mmu_notifier_mm can't go away from under us because one mm_count
100	 * is held by exit_mmap.
101	 */
102	synchronize_srcu(&srcu);
103}
104
105/*
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
108 * existed or not.
109 */
110int __mmu_notifier_clear_flush_young(struct mm_struct *mm,
111					unsigned long start,
112					unsigned long end)
113{
114	struct mmu_notifier *mn;
115	int young = 0, id;
 
116
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);
121	}
122	srcu_read_unlock(&srcu, id);
123
124	return young;
125}
126
127int __mmu_notifier_clear_young(struct mm_struct *mm,
128			       unsigned long start,
129			       unsigned long end)
130{
131	struct mmu_notifier *mn;
132	int young = 0, id;
133
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);
138	}
139	srcu_read_unlock(&srcu, id);
140
141	return young;
142}
143
144int __mmu_notifier_test_young(struct mm_struct *mm,
145			      unsigned long address)
146{
147	struct mmu_notifier *mn;
148	int young = 0, id;
 
149
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);
154			if (young)
155				break;
156		}
157	}
158	srcu_read_unlock(&srcu, id);
159
160	return young;
161}
162
163void __mmu_notifier_change_pte(struct mm_struct *mm, unsigned long address,
164			       pte_t pte)
165{
166	struct mmu_notifier *mn;
167	int id;
168
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);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
173	}
174	srcu_read_unlock(&srcu, id);
175}
176
177void __mmu_notifier_invalidate_range_start(struct mm_struct *mm,
178				  unsigned long start, unsigned long end)
179{
180	struct mmu_notifier *mn;
181	int id;
182
183	id = srcu_read_lock(&srcu);
184	hlist_for_each_entry_rcu(mn, &mm->mmu_notifier_mm->list, hlist) {
185		if (mn->ops->invalidate_range_start)
186			mn->ops->invalidate_range_start(mn, mm, start, end);
187	}
188	srcu_read_unlock(&srcu, id);
189}
190EXPORT_SYMBOL_GPL(__mmu_notifier_invalidate_range_start);
191
192void __mmu_notifier_invalidate_range_end(struct mm_struct *mm,
193					 unsigned long start,
194					 unsigned long end,
195					 bool only_end)
196{
197	struct mmu_notifier *mn;
198	int id;
199
200	id = srcu_read_lock(&srcu);
201	hlist_for_each_entry_rcu(mn, &mm->mmu_notifier_mm->list, hlist) {
202		/*
203		 * Call invalidate_range here too to avoid the need for the
204		 * subsystem of having to register an invalidate_range_end
205		 * call-back when there is invalidate_range already. Usually a
206		 * subsystem registers either invalidate_range_start()/end() or
207		 * invalidate_range(), so this will be no additional overhead
208		 * (besides the pointer check).
209		 *
210		 * We skip call to invalidate_range() if we know it is safe ie
211		 * call site use mmu_notifier_invalidate_range_only_end() which
212		 * is safe to do when we know that a call to invalidate_range()
213		 * already happen under page table lock.
214		 */
215		if (!only_end && mn->ops->invalidate_range)
216			mn->ops->invalidate_range(mn, mm, start, end);
217		if (mn->ops->invalidate_range_end)
218			mn->ops->invalidate_range_end(mn, mm, start, end);
219	}
220	srcu_read_unlock(&srcu, id);
221}
222EXPORT_SYMBOL_GPL(__mmu_notifier_invalidate_range_end);
223
224void __mmu_notifier_invalidate_range(struct mm_struct *mm,
225				  unsigned long start, unsigned long end)
226{
227	struct mmu_notifier *mn;
228	int id;
229
230	id = srcu_read_lock(&srcu);
231	hlist_for_each_entry_rcu(mn, &mm->mmu_notifier_mm->list, hlist) {
232		if (mn->ops->invalidate_range)
233			mn->ops->invalidate_range(mn, mm, start, end);
234	}
235	srcu_read_unlock(&srcu, id);
236}
237EXPORT_SYMBOL_GPL(__mmu_notifier_invalidate_range);
238
239/*
240 * Must be called while holding mm->mmap_sem for either read or write.
241 * The result is guaranteed to be valid until mm->mmap_sem is dropped.
242 */
243bool mm_has_blockable_invalidate_notifiers(struct mm_struct *mm)
244{
245	struct mmu_notifier *mn;
246	int id;
247	bool ret = false;
248
249	WARN_ON_ONCE(!rwsem_is_locked(&mm->mmap_sem));
250
251	if (!mm_has_notifiers(mm))
252		return ret;
253
254	id = srcu_read_lock(&srcu);
255	hlist_for_each_entry_rcu(mn, &mm->mmu_notifier_mm->list, hlist) {
256		if (!mn->ops->invalidate_range &&
257		    !mn->ops->invalidate_range_start &&
258		    !mn->ops->invalidate_range_end)
259				continue;
260
261		if (!(mn->ops->flags & MMU_INVALIDATE_DOES_NOT_BLOCK)) {
262			ret = true;
263			break;
264		}
265	}
266	srcu_read_unlock(&srcu, id);
267	return ret;
268}
269
270static int do_mmu_notifier_register(struct mmu_notifier *mn,
271				    struct mm_struct *mm,
272				    int take_mmap_sem)
273{
274	struct mmu_notifier_mm *mmu_notifier_mm;
275	int ret;
276
277	BUG_ON(atomic_read(&mm->mm_users) <= 0);
278
279	ret = -ENOMEM;
280	mmu_notifier_mm = kmalloc(sizeof(struct mmu_notifier_mm), GFP_KERNEL);
281	if (unlikely(!mmu_notifier_mm))
282		goto out;
283
284	if (take_mmap_sem)
285		down_write(&mm->mmap_sem);
286	ret = mm_take_all_locks(mm);
287	if (unlikely(ret))
288		goto out_clean;
289
290	if (!mm_has_notifiers(mm)) {
291		INIT_HLIST_HEAD(&mmu_notifier_mm->list);
292		spin_lock_init(&mmu_notifier_mm->lock);
293
294		mm->mmu_notifier_mm = mmu_notifier_mm;
295		mmu_notifier_mm = NULL;
296	}
297	mmgrab(mm);
298
299	/*
300	 * Serialize the update against mmu_notifier_unregister. A
301	 * side note: mmu_notifier_release can't run concurrently with
302	 * us because we hold the mm_users pin (either implicitly as
303	 * current->mm or explicitly with get_task_mm() or similar).
304	 * We can't race against any other mmu notifier method either
305	 * thanks to mm_take_all_locks().
306	 */
307	spin_lock(&mm->mmu_notifier_mm->lock);
308	hlist_add_head(&mn->hlist, &mm->mmu_notifier_mm->list);
309	spin_unlock(&mm->mmu_notifier_mm->lock);
310
311	mm_drop_all_locks(mm);
312out_clean:
313	if (take_mmap_sem)
314		up_write(&mm->mmap_sem);
 
315	kfree(mmu_notifier_mm);
316out:
317	BUG_ON(atomic_read(&mm->mm_users) <= 0);
318	return ret;
319}
320
321/*
322 * Must not hold mmap_sem nor any other VM related lock when calling
323 * this registration function. Must also ensure mm_users can't go down
324 * to zero while this runs to avoid races with mmu_notifier_release,
325 * so mm has to be current->mm or the mm should be pinned safely such
326 * as with get_task_mm(). If the mm is not current->mm, the mm_users
327 * pin should be released by calling mmput after mmu_notifier_register
328 * returns. mmu_notifier_unregister must be always called to
329 * unregister the notifier. mm_count is automatically pinned to allow
330 * mmu_notifier_unregister to safely run at any time later, before or
331 * after exit_mmap. ->release will always be called before exit_mmap
332 * frees the pages.
333 */
334int mmu_notifier_register(struct mmu_notifier *mn, struct mm_struct *mm)
335{
336	return do_mmu_notifier_register(mn, mm, 1);
337}
338EXPORT_SYMBOL_GPL(mmu_notifier_register);
339
340/*
341 * Same as mmu_notifier_register but here the caller must hold the
342 * mmap_sem in write mode.
343 */
344int __mmu_notifier_register(struct mmu_notifier *mn, struct mm_struct *mm)
345{
346	return do_mmu_notifier_register(mn, mm, 0);
347}
348EXPORT_SYMBOL_GPL(__mmu_notifier_register);
349
350/* this is called after the last mmu_notifier_unregister() returned */
351void __mmu_notifier_mm_destroy(struct mm_struct *mm)
352{
353	BUG_ON(!hlist_empty(&mm->mmu_notifier_mm->list));
354	kfree(mm->mmu_notifier_mm);
355	mm->mmu_notifier_mm = LIST_POISON1; /* debug */
356}
357
358/*
359 * This releases the mm_count pin automatically and frees the mm
360 * structure if it was the last user of it. It serializes against
361 * running mmu notifiers with SRCU and against mmu_notifier_unregister
362 * with the unregister lock + SRCU. All sptes must be dropped before
363 * calling mmu_notifier_unregister. ->release or any other notifier
364 * method may be invoked concurrently with mmu_notifier_unregister,
365 * and only after mmu_notifier_unregister returned we're guaranteed
366 * that ->release or any other method can't run anymore.
367 */
368void mmu_notifier_unregister(struct mmu_notifier *mn, struct mm_struct *mm)
369{
370	BUG_ON(atomic_read(&mm->mm_count) <= 0);
371
 
372	if (!hlist_unhashed(&mn->hlist)) {
 
 
373		/*
374		 * SRCU here will force exit_mmap to wait for ->release to
375		 * finish before freeing the pages.
376		 */
377		int id;
378
379		id = srcu_read_lock(&srcu);
380		/*
381		 * exit_mmap will block in mmu_notifier_release to guarantee
382		 * that ->release is called before freeing the pages.
 
383		 */
384		if (mn->ops->release)
385			mn->ops->release(mn, mm);
386		srcu_read_unlock(&srcu, id);
387
388		spin_lock(&mm->mmu_notifier_mm->lock);
389		/*
390		 * Can not use list_del_rcu() since __mmu_notifier_release
391		 * can delete it before we hold the lock.
392		 */
393		hlist_del_init_rcu(&mn->hlist);
394		spin_unlock(&mm->mmu_notifier_mm->lock);
395	}
396
397	/*
398	 * Wait for any running method to finish, of course including
399	 * ->release if it was run by mmu_notifier_release instead of us.
400	 */
401	synchronize_srcu(&srcu);
402
403	BUG_ON(atomic_read(&mm->mm_count) <= 0);
404
405	mmdrop(mm);
406}
407EXPORT_SYMBOL_GPL(mmu_notifier_unregister);
408
409/*
410 * Same as mmu_notifier_unregister but no callback and no srcu synchronization.
411 */
412void mmu_notifier_unregister_no_release(struct mmu_notifier *mn,
413					struct mm_struct *mm)
414{
415	spin_lock(&mm->mmu_notifier_mm->lock);
416	/*
417	 * Can not use list_del_rcu() since __mmu_notifier_release
418	 * can delete it before we hold the lock.
419	 */
420	hlist_del_init_rcu(&mn->hlist);
421	spin_unlock(&mm->mmu_notifier_mm->lock);
422
423	BUG_ON(atomic_read(&mm->mm_count) <= 0);
424	mmdrop(mm);
425}
426EXPORT_SYMBOL_GPL(mmu_notifier_unregister_no_release);