2 * linux/mm/mmu_notifier.c
4 * Copyright (C) 2008 Qumranet, Inc.
5 * Copyright (C) 2008 SGI
6 * Christoph Lameter <clameter@sgi.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/slab.h>
22 /* global SRCU for all MMs */
23 static struct srcu_struct srcu;
26 * This function can't run concurrently against mmu_notifier_register
27 * because mm->mm_users > 0 during mmu_notifier_register and exit_mmap
28 * runs with mm_users == 0. Other tasks may still invoke mmu notifiers
29 * in parallel despite there being no task using this mm any more,
30 * through the vmas outside of the exit_mmap context, such as with
31 * vmtruncate. This serializes against mmu_notifier_unregister with
32 * the mmu_notifier_mm->lock in addition to SRCU and it serializes
33 * against the other mmu notifiers with SRCU. struct mmu_notifier_mm
34 * can't go away from under us as exit_mmap holds an mm_count pin
37 void __mmu_notifier_release(struct mm_struct *mm)
39 struct mmu_notifier *mn;
43 * SRCU here will block mmu_notifier_unregister until
46 id = srcu_read_lock(&srcu);
47 hlist_for_each_entry_rcu(mn, &mm->mmu_notifier_mm->list, hlist)
49 * If ->release runs before mmu_notifier_unregister it must be
50 * handled, as it's the only way for the driver to flush all
51 * existing sptes and stop the driver from establishing any more
52 * sptes before all the pages in the mm are freed.
55 mn->ops->release(mn, mm);
56 srcu_read_unlock(&srcu, id);
58 spin_lock(&mm->mmu_notifier_mm->lock);
59 while (unlikely(!hlist_empty(&mm->mmu_notifier_mm->list))) {
60 mn = hlist_entry(mm->mmu_notifier_mm->list.first,
64 * We arrived before mmu_notifier_unregister so
65 * mmu_notifier_unregister will do nothing other than to wait
66 * for ->release to finish and for mmu_notifier_unregister to
69 hlist_del_init_rcu(&mn->hlist);
71 spin_unlock(&mm->mmu_notifier_mm->lock);
74 * synchronize_srcu here prevents mmu_notifier_release from returning to
75 * exit_mmap (which would proceed with freeing all pages in the mm)
76 * until the ->release method returns, if it was invoked by
77 * mmu_notifier_unregister.
79 * The mmu_notifier_mm can't go away from under us because one mm_count
80 * is held by exit_mmap.
82 synchronize_srcu(&srcu);
86 * If no young bitflag is supported by the hardware, ->clear_flush_young can
87 * unmap the address and return 1 or 0 depending if the mapping previously
90 int __mmu_notifier_clear_flush_young(struct mm_struct *mm,
91 unsigned long address)
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, address);
101 srcu_read_unlock(&srcu, id);
106 int __mmu_notifier_test_young(struct mm_struct *mm,
107 unsigned long address)
109 struct mmu_notifier *mn;
112 id = srcu_read_lock(&srcu);
113 hlist_for_each_entry_rcu(mn, &mm->mmu_notifier_mm->list, hlist) {
114 if (mn->ops->test_young) {
115 young = mn->ops->test_young(mn, mm, address);
120 srcu_read_unlock(&srcu, id);
125 void __mmu_notifier_change_pte(struct mm_struct *mm, unsigned long address,
128 struct mmu_notifier *mn;
131 id = srcu_read_lock(&srcu);
132 hlist_for_each_entry_rcu(mn, &mm->mmu_notifier_mm->list, hlist) {
133 if (mn->ops->change_pte)
134 mn->ops->change_pte(mn, mm, address, pte);
136 srcu_read_unlock(&srcu, id);
139 void __mmu_notifier_invalidate_page(struct mm_struct *mm,
140 unsigned long address)
142 struct mmu_notifier *mn;
145 id = srcu_read_lock(&srcu);
146 hlist_for_each_entry_rcu(mn, &mm->mmu_notifier_mm->list, hlist) {
147 if (mn->ops->invalidate_page)
148 mn->ops->invalidate_page(mn, mm, address);
150 srcu_read_unlock(&srcu, id);
153 void __mmu_notifier_invalidate_range_start(struct mm_struct *mm,
154 unsigned long start, unsigned long end)
156 struct mmu_notifier *mn;
159 id = srcu_read_lock(&srcu);
160 hlist_for_each_entry_rcu(mn, &mm->mmu_notifier_mm->list, hlist) {
161 if (mn->ops->invalidate_range_start)
162 mn->ops->invalidate_range_start(mn, mm, start, end);
164 srcu_read_unlock(&srcu, id);
166 EXPORT_SYMBOL_GPL(__mmu_notifier_invalidate_range_start);
168 void __mmu_notifier_invalidate_range_end(struct mm_struct *mm,
169 unsigned long start, unsigned long end)
171 struct mmu_notifier *mn;
174 id = srcu_read_lock(&srcu);
175 hlist_for_each_entry_rcu(mn, &mm->mmu_notifier_mm->list, hlist) {
176 if (mn->ops->invalidate_range_end)
177 mn->ops->invalidate_range_end(mn, mm, start, end);
179 srcu_read_unlock(&srcu, id);
181 EXPORT_SYMBOL_GPL(__mmu_notifier_invalidate_range_end);
183 static int do_mmu_notifier_register(struct mmu_notifier *mn,
184 struct mm_struct *mm,
187 struct mmu_notifier_mm *mmu_notifier_mm;
190 BUG_ON(atomic_read(&mm->mm_users) <= 0);
193 * Verify that mmu_notifier_init() already run and the global srcu is
196 BUG_ON(!srcu.per_cpu_ref);
199 mmu_notifier_mm = kmalloc(sizeof(struct mmu_notifier_mm), GFP_KERNEL);
200 if (unlikely(!mmu_notifier_mm))
204 down_write(&mm->mmap_sem);
205 ret = mm_take_all_locks(mm);
209 if (!mm_has_notifiers(mm)) {
210 INIT_HLIST_HEAD(&mmu_notifier_mm->list);
211 spin_lock_init(&mmu_notifier_mm->lock);
213 mm->mmu_notifier_mm = mmu_notifier_mm;
214 mmu_notifier_mm = NULL;
216 atomic_inc(&mm->mm_count);
219 * Serialize the update against mmu_notifier_unregister. A
220 * side note: mmu_notifier_release can't run concurrently with
221 * us because we hold the mm_users pin (either implicitly as
222 * current->mm or explicitly with get_task_mm() or similar).
223 * We can't race against any other mmu notifier method either
224 * thanks to mm_take_all_locks().
226 spin_lock(&mm->mmu_notifier_mm->lock);
227 hlist_add_head(&mn->hlist, &mm->mmu_notifier_mm->list);
228 spin_unlock(&mm->mmu_notifier_mm->lock);
230 mm_drop_all_locks(mm);
233 up_write(&mm->mmap_sem);
234 kfree(mmu_notifier_mm);
236 BUG_ON(atomic_read(&mm->mm_users) <= 0);
241 * Must not hold mmap_sem nor any other VM related lock when calling
242 * this registration function. Must also ensure mm_users can't go down
243 * to zero while this runs to avoid races with mmu_notifier_release,
244 * so mm has to be current->mm or the mm should be pinned safely such
245 * as with get_task_mm(). If the mm is not current->mm, the mm_users
246 * pin should be released by calling mmput after mmu_notifier_register
247 * returns. mmu_notifier_unregister must be always called to
248 * unregister the notifier. mm_count is automatically pinned to allow
249 * mmu_notifier_unregister to safely run at any time later, before or
250 * after exit_mmap. ->release will always be called before exit_mmap
253 int mmu_notifier_register(struct mmu_notifier *mn, struct mm_struct *mm)
255 return do_mmu_notifier_register(mn, mm, 1);
257 EXPORT_SYMBOL_GPL(mmu_notifier_register);
260 * Same as mmu_notifier_register but here the caller must hold the
261 * mmap_sem in write mode.
263 int __mmu_notifier_register(struct mmu_notifier *mn, struct mm_struct *mm)
265 return do_mmu_notifier_register(mn, mm, 0);
267 EXPORT_SYMBOL_GPL(__mmu_notifier_register);
269 /* this is called after the last mmu_notifier_unregister() returned */
270 void __mmu_notifier_mm_destroy(struct mm_struct *mm)
272 BUG_ON(!hlist_empty(&mm->mmu_notifier_mm->list));
273 kfree(mm->mmu_notifier_mm);
274 mm->mmu_notifier_mm = LIST_POISON1; /* debug */
278 * This releases the mm_count pin automatically and frees the mm
279 * structure if it was the last user of it. It serializes against
280 * running mmu notifiers with SRCU and against mmu_notifier_unregister
281 * with the unregister lock + SRCU. All sptes must be dropped before
282 * calling mmu_notifier_unregister. ->release or any other notifier
283 * method may be invoked concurrently with mmu_notifier_unregister,
284 * and only after mmu_notifier_unregister returned we're guaranteed
285 * that ->release or any other method can't run anymore.
287 void mmu_notifier_unregister(struct mmu_notifier *mn, struct mm_struct *mm)
289 BUG_ON(atomic_read(&mm->mm_count) <= 0);
291 if (!hlist_unhashed(&mn->hlist)) {
293 * SRCU here will force exit_mmap to wait for ->release to
294 * finish before freeing the pages.
298 id = srcu_read_lock(&srcu);
300 * exit_mmap will block in mmu_notifier_release to guarantee
301 * that ->release is called before freeing the pages.
303 if (mn->ops->release)
304 mn->ops->release(mn, mm);
305 srcu_read_unlock(&srcu, id);
307 spin_lock(&mm->mmu_notifier_mm->lock);
309 * Can not use list_del_rcu() since __mmu_notifier_release
310 * can delete it before we hold the lock.
312 hlist_del_init_rcu(&mn->hlist);
313 spin_unlock(&mm->mmu_notifier_mm->lock);
317 * Wait for any running method to finish, of course including
318 * ->release if it was run by mmu_notifier_relase instead of us.
320 synchronize_srcu(&srcu);
322 BUG_ON(atomic_read(&mm->mm_count) <= 0);
326 EXPORT_SYMBOL_GPL(mmu_notifier_unregister);
328 static int __init mmu_notifier_init(void)
330 return init_srcu_struct(&srcu);
333 module_init(mmu_notifier_init);