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/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/slab.h>
21
22/* global SRCU for all MMs */
23static struct srcu_struct srcu;
24
25/*
26 * This function allows mmu_notifier::release callback to delay a call to
27 * a function that will free appropriate resources. The function must be
28 * quick and must not block.
29 */
30void mmu_notifier_call_srcu(struct rcu_head *rcu,
31			    void (*func)(struct rcu_head *rcu))
32{
33	call_srcu(&srcu, rcu, func);
34}
35EXPORT_SYMBOL_GPL(mmu_notifier_call_srcu);
36
37void mmu_notifier_synchronize(void)
38{
39	/* Wait for any running method to finish. */
40	srcu_barrier(&srcu);
41}
42EXPORT_SYMBOL_GPL(mmu_notifier_synchronize);
43
44/*
45 * This function can't run concurrently against mmu_notifier_register
46 * because mm->mm_users > 0 during mmu_notifier_register and exit_mmap
47 * runs with mm_users == 0. Other tasks may still invoke mmu notifiers
48 * in parallel despite there being no task using this mm any more,
49 * through the vmas outside of the exit_mmap context, such as with
50 * vmtruncate. This serializes against mmu_notifier_unregister with
51 * the mmu_notifier_mm->lock in addition to SRCU and it serializes
52 * against the other mmu notifiers with SRCU. struct mmu_notifier_mm
53 * can't go away from under us as exit_mmap holds an mm_count pin
54 * itself.
55 */
56void __mmu_notifier_release(struct mm_struct *mm)
57{
58	struct mmu_notifier *mn;
59	int id;
60
61	/*
62	 * SRCU here will block mmu_notifier_unregister until
63	 * ->release returns.
64	 */
65	id = srcu_read_lock(&srcu);
66	hlist_for_each_entry_rcu(mn, &mm->mmu_notifier_mm->list, hlist)
67		/*
68		 * If ->release runs before mmu_notifier_unregister it must be
69		 * handled, as it's the only way for the driver to flush all
70		 * existing sptes and stop the driver from establishing any more
71		 * sptes before all the pages in the mm are freed.
72		 */
73		if (mn->ops->release)
74			mn->ops->release(mn, mm);
75
76	spin_lock(&mm->mmu_notifier_mm->lock);
77	while (unlikely(!hlist_empty(&mm->mmu_notifier_mm->list))) {
78		mn = hlist_entry(mm->mmu_notifier_mm->list.first,
79				 struct mmu_notifier,
80				 hlist);
81		/*
82		 * We arrived before mmu_notifier_unregister so
83		 * mmu_notifier_unregister will do nothing other than to wait
84		 * for ->release to finish and for mmu_notifier_unregister to
85		 * return.
86		 */
87		hlist_del_init_rcu(&mn->hlist);
88	}
89	spin_unlock(&mm->mmu_notifier_mm->lock);
90	srcu_read_unlock(&srcu, id);
91
92	/*
93	 * synchronize_srcu here prevents mmu_notifier_release from returning to
94	 * exit_mmap (which would proceed with freeing all pages in the mm)
95	 * until the ->release method returns, if it was invoked by
96	 * mmu_notifier_unregister.
97	 *
98	 * The mmu_notifier_mm can't go away from under us because one mm_count
99	 * is held by exit_mmap.
100	 */
101	synchronize_srcu(&srcu);
102}
103
104/*
105 * If no young bitflag is supported by the hardware, ->clear_flush_young can
106 * unmap the address and return 1 or 0 depending if the mapping previously
107 * existed or not.
108 */
109int __mmu_notifier_clear_flush_young(struct mm_struct *mm,
110					unsigned long start,
111					unsigned long end)
112{
113	struct mmu_notifier *mn;
114	int young = 0, id;
115
116	id = srcu_read_lock(&srcu);
117	hlist_for_each_entry_rcu(mn, &mm->mmu_notifier_mm->list, hlist) {
118		if (mn->ops->clear_flush_young)
119			young |= mn->ops->clear_flush_young(mn, mm, start, end);
120	}
121	srcu_read_unlock(&srcu, id);
122
123	return young;
124}
125
126int __mmu_notifier_test_young(struct mm_struct *mm,
127			      unsigned long address)
128{
129	struct mmu_notifier *mn;
130	int young = 0, id;
131
132	id = srcu_read_lock(&srcu);
133	hlist_for_each_entry_rcu(mn, &mm->mmu_notifier_mm->list, hlist) {
134		if (mn->ops->test_young) {
135			young = mn->ops->test_young(mn, mm, address);
136			if (young)
137				break;
138		}
139	}
140	srcu_read_unlock(&srcu, id);
141
142	return young;
143}
144
145void __mmu_notifier_change_pte(struct mm_struct *mm, unsigned long address,
146			       pte_t pte)
147{
148	struct mmu_notifier *mn;
149	int id;
150
151	id = srcu_read_lock(&srcu);
152	hlist_for_each_entry_rcu(mn, &mm->mmu_notifier_mm->list, hlist) {
153		if (mn->ops->change_pte)
154			mn->ops->change_pte(mn, mm, address, pte);
155	}
156	srcu_read_unlock(&srcu, id);
157}
158
159void __mmu_notifier_invalidate_page(struct mm_struct *mm,
160					  unsigned long address)
161{
162	struct mmu_notifier *mn;
163	int id;
164
165	id = srcu_read_lock(&srcu);
166	hlist_for_each_entry_rcu(mn, &mm->mmu_notifier_mm->list, hlist) {
167		if (mn->ops->invalidate_page)
168			mn->ops->invalidate_page(mn, mm, address);
169	}
170	srcu_read_unlock(&srcu, id);
171}
172
173void __mmu_notifier_invalidate_range_start(struct mm_struct *mm,
174				  unsigned long start, unsigned long end)
175{
176	struct mmu_notifier *mn;
177	int id;
178
179	id = srcu_read_lock(&srcu);
180	hlist_for_each_entry_rcu(mn, &mm->mmu_notifier_mm->list, hlist) {
181		if (mn->ops->invalidate_range_start)
182			mn->ops->invalidate_range_start(mn, mm, start, end);
183	}
184	srcu_read_unlock(&srcu, id);
185}
186EXPORT_SYMBOL_GPL(__mmu_notifier_invalidate_range_start);
187
188void __mmu_notifier_invalidate_range_end(struct mm_struct *mm,
189				  unsigned long start, unsigned long end)
190{
191	struct mmu_notifier *mn;
192	int id;
193
194	id = srcu_read_lock(&srcu);
195	hlist_for_each_entry_rcu(mn, &mm->mmu_notifier_mm->list, hlist) {
196		/*
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).
203		 */
204		if (mn->ops->invalidate_range)
205			mn->ops->invalidate_range(mn, mm, start, end);
206		if (mn->ops->invalidate_range_end)
207			mn->ops->invalidate_range_end(mn, mm, start, end);
208	}
209	srcu_read_unlock(&srcu, id);
210}
211EXPORT_SYMBOL_GPL(__mmu_notifier_invalidate_range_end);
212
213void __mmu_notifier_invalidate_range(struct mm_struct *mm,
214				  unsigned long start, unsigned long end)
215{
216	struct mmu_notifier *mn;
217	int id;
218
219	id = srcu_read_lock(&srcu);
220	hlist_for_each_entry_rcu(mn, &mm->mmu_notifier_mm->list, hlist) {
221		if (mn->ops->invalidate_range)
222			mn->ops->invalidate_range(mn, mm, start, end);
223	}
224	srcu_read_unlock(&srcu, id);
225}
226EXPORT_SYMBOL_GPL(__mmu_notifier_invalidate_range);
227
228static int do_mmu_notifier_register(struct mmu_notifier *mn,
229				    struct mm_struct *mm,
230				    int take_mmap_sem)
231{
232	struct mmu_notifier_mm *mmu_notifier_mm;
233	int ret;
234
235	BUG_ON(atomic_read(&mm->mm_users) <= 0);
236
237	/*
238	 * Verify that mmu_notifier_init() already run and the global srcu is
239	 * initialized.
240	 */
241	BUG_ON(!srcu.per_cpu_ref);
242
243	ret = -ENOMEM;
244	mmu_notifier_mm = kmalloc(sizeof(struct mmu_notifier_mm), GFP_KERNEL);
245	if (unlikely(!mmu_notifier_mm))
246		goto out;
247
248	if (take_mmap_sem)
249		down_write(&mm->mmap_sem);
250	ret = mm_take_all_locks(mm);
251	if (unlikely(ret))
252		goto out_clean;
253
254	if (!mm_has_notifiers(mm)) {
255		INIT_HLIST_HEAD(&mmu_notifier_mm->list);
256		spin_lock_init(&mmu_notifier_mm->lock);
257
258		mm->mmu_notifier_mm = mmu_notifier_mm;
259		mmu_notifier_mm = NULL;
260	}
261	atomic_inc(&mm->mm_count);
262
263	/*
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().
270	 */
271	spin_lock(&mm->mmu_notifier_mm->lock);
272	hlist_add_head(&mn->hlist, &mm->mmu_notifier_mm->list);
273	spin_unlock(&mm->mmu_notifier_mm->lock);
274
275	mm_drop_all_locks(mm);
276out_clean:
277	if (take_mmap_sem)
278		up_write(&mm->mmap_sem);
279	kfree(mmu_notifier_mm);
280out:
281	BUG_ON(atomic_read(&mm->mm_users) <= 0);
282	return ret;
283}
284
285/*
286 * Must not hold mmap_sem nor any other VM related lock when calling
287 * this registration function. Must also ensure mm_users can't go down
288 * to zero while this runs to avoid races with mmu_notifier_release,
289 * so mm has to be current->mm or the mm should be pinned safely such
290 * as with get_task_mm(). If the mm is not current->mm, the mm_users
291 * pin should be released by calling mmput after mmu_notifier_register
292 * returns. mmu_notifier_unregister must be always called to
293 * unregister the notifier. mm_count is automatically pinned to allow
294 * mmu_notifier_unregister to safely run at any time later, before or
295 * after exit_mmap. ->release will always be called before exit_mmap
296 * frees the pages.
297 */
298int mmu_notifier_register(struct mmu_notifier *mn, struct mm_struct *mm)
299{
300	return do_mmu_notifier_register(mn, mm, 1);
301}
302EXPORT_SYMBOL_GPL(mmu_notifier_register);
303
304/*
305 * Same as mmu_notifier_register but here the caller must hold the
306 * mmap_sem in write mode.
307 */
308int __mmu_notifier_register(struct mmu_notifier *mn, struct mm_struct *mm)
309{
310	return do_mmu_notifier_register(mn, mm, 0);
311}
312EXPORT_SYMBOL_GPL(__mmu_notifier_register);
313
314/* this is called after the last mmu_notifier_unregister() returned */
315void __mmu_notifier_mm_destroy(struct mm_struct *mm)
316{
317	BUG_ON(!hlist_empty(&mm->mmu_notifier_mm->list));
318	kfree(mm->mmu_notifier_mm);
319	mm->mmu_notifier_mm = LIST_POISON1; /* debug */
320}
321
322/*
323 * This releases the mm_count pin automatically and frees the mm
324 * structure if it was the last user of it. It serializes against
325 * running mmu notifiers with SRCU and against mmu_notifier_unregister
326 * with the unregister lock + SRCU. All sptes must be dropped before
327 * calling mmu_notifier_unregister. ->release or any other notifier
328 * method may be invoked concurrently with mmu_notifier_unregister,
329 * and only after mmu_notifier_unregister returned we're guaranteed
330 * that ->release or any other method can't run anymore.
331 */
332void mmu_notifier_unregister(struct mmu_notifier *mn, struct mm_struct *mm)
333{
334	BUG_ON(atomic_read(&mm->mm_count) <= 0);
335
336	if (!hlist_unhashed(&mn->hlist)) {
337		/*
338		 * SRCU here will force exit_mmap to wait for ->release to
339		 * finish before freeing the pages.
340		 */
341		int id;
342
343		id = srcu_read_lock(&srcu);
344		/*
345		 * exit_mmap will block in mmu_notifier_release to guarantee
346		 * that ->release is called before freeing the pages.
347		 */
348		if (mn->ops->release)
349			mn->ops->release(mn, mm);
350		srcu_read_unlock(&srcu, id);
351
352		spin_lock(&mm->mmu_notifier_mm->lock);
353		/*
354		 * Can not use list_del_rcu() since __mmu_notifier_release
355		 * can delete it before we hold the lock.
356		 */
357		hlist_del_init_rcu(&mn->hlist);
358		spin_unlock(&mm->mmu_notifier_mm->lock);
359	}
360
361	/*
362	 * Wait for any running method to finish, of course including
363	 * ->release if it was run by mmu_notifier_release instead of us.
364	 */
365	synchronize_srcu(&srcu);
366
367	BUG_ON(atomic_read(&mm->mm_count) <= 0);
368
369	mmdrop(mm);
370}
371EXPORT_SYMBOL_GPL(mmu_notifier_unregister);
372
373/*
374 * Same as mmu_notifier_unregister but no callback and no srcu synchronization.
375 */
376void mmu_notifier_unregister_no_release(struct mmu_notifier *mn,
377					struct mm_struct *mm)
378{
379	spin_lock(&mm->mmu_notifier_mm->lock);
380	/*
381	 * Can not use list_del_rcu() since __mmu_notifier_release
382	 * can delete it before we hold the lock.
383	 */
384	hlist_del_init_rcu(&mn->hlist);
385	spin_unlock(&mm->mmu_notifier_mm->lock);
386
387	BUG_ON(atomic_read(&mm->mm_count) <= 0);
388	mmdrop(mm);
389}
390EXPORT_SYMBOL_GPL(mmu_notifier_unregister_no_release);
391
392static int __init mmu_notifier_init(void)
393{
394	return init_srcu_struct(&srcu);
395}
396subsys_initcall(mmu_notifier_init);
397