blob: 479a1e751a73fb4781e1b856db60b2381909c9e6 [file] [log] [blame]
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001/*
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>
Paul Gortmakerb95f1b312011-10-16 02:01:52 -040014#include <linux/export.h>
Andrea Arcangelicddb8a52008-07-28 15:46:29 -070015#include <linux/mm.h>
16#include <linux/err.h>
Sagi Grimberg21a92732012-10-08 16:29:24 -070017#include <linux/srcu.h>
Andrea Arcangelicddb8a52008-07-28 15:46:29 -070018#include <linux/rcupdate.h>
19#include <linux/sched.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090020#include <linux/slab.h>
Andrea Arcangelicddb8a52008-07-28 15:46:29 -070021
Sagi Grimberg21a92732012-10-08 16:29:24 -070022/* global SRCU for all MMs */
Andrea Arcangeli70400302012-10-08 16:31:52 -070023static struct srcu_struct srcu;
Sagi Grimberg21a92732012-10-08 16:29:24 -070024
Andrea Arcangelicddb8a52008-07-28 15:46:29 -070025/*
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
Sagi Grimberg21a92732012-10-08 16:29:24 -070032 * the mmu_notifier_mm->lock in addition to SRCU and it serializes
33 * against the other mmu notifiers with SRCU. struct mmu_notifier_mm
Andrea Arcangelicddb8a52008-07-28 15:46:29 -070034 * can't go away from under us as exit_mmap holds an mm_count pin
35 * itself.
36 */
37void __mmu_notifier_release(struct mm_struct *mm)
38{
39 struct mmu_notifier *mn;
Xiao Guangrong3ad3d902012-07-31 16:45:52 -070040 struct hlist_node *n;
Sagi Grimberg21a92732012-10-08 16:29:24 -070041 int id;
Xiao Guangrong3ad3d902012-07-31 16:45:52 -070042
43 /*
Andrea Arcangeli70400302012-10-08 16:31:52 -070044 * SRCU here will block mmu_notifier_unregister until
Xiao Guangrong3ad3d902012-07-31 16:45:52 -070045 * ->release returns.
46 */
Sagi Grimberg21a92732012-10-08 16:29:24 -070047 id = srcu_read_lock(&srcu);
Xiao Guangrong3ad3d902012-07-31 16:45:52 -070048 hlist_for_each_entry_rcu(mn, n, &mm->mmu_notifier_mm->list, hlist)
49 /*
50 * if ->release runs before mmu_notifier_unregister it
51 * must be handled as it's the only way for the driver
52 * to flush all existing sptes and stop the driver
53 * from establishing any more sptes before all the
54 * pages in the mm are freed.
55 */
56 if (mn->ops->release)
57 mn->ops->release(mn, mm);
Sagi Grimberg21a92732012-10-08 16:29:24 -070058 srcu_read_unlock(&srcu, id);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -070059
60 spin_lock(&mm->mmu_notifier_mm->lock);
61 while (unlikely(!hlist_empty(&mm->mmu_notifier_mm->list))) {
62 mn = hlist_entry(mm->mmu_notifier_mm->list.first,
63 struct mmu_notifier,
64 hlist);
65 /*
66 * We arrived before mmu_notifier_unregister so
67 * mmu_notifier_unregister will do nothing other than
68 * to wait ->release to finish and
69 * mmu_notifier_unregister to return.
70 */
71 hlist_del_init_rcu(&mn->hlist);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -070072 }
73 spin_unlock(&mm->mmu_notifier_mm->lock);
74
75 /*
Sagi Grimberg21a92732012-10-08 16:29:24 -070076 * synchronize_srcu here prevents mmu_notifier_release to
Andrea Arcangelicddb8a52008-07-28 15:46:29 -070077 * return to exit_mmap (which would proceed freeing all pages
78 * in the mm) until the ->release method returns, if it was
79 * invoked by mmu_notifier_unregister.
80 *
81 * The mmu_notifier_mm can't go away from under us because one
82 * mm_count is hold by exit_mmap.
83 */
Sagi Grimberg21a92732012-10-08 16:29:24 -070084 synchronize_srcu(&srcu);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -070085}
86
87/*
88 * If no young bitflag is supported by the hardware, ->clear_flush_young can
89 * unmap the address and return 1 or 0 depending if the mapping previously
90 * existed or not.
91 */
92int __mmu_notifier_clear_flush_young(struct mm_struct *mm,
93 unsigned long address)
94{
95 struct mmu_notifier *mn;
96 struct hlist_node *n;
Sagi Grimberg21a92732012-10-08 16:29:24 -070097 int young = 0, id;
Andrea Arcangelicddb8a52008-07-28 15:46:29 -070098
Sagi Grimberg21a92732012-10-08 16:29:24 -070099 id = srcu_read_lock(&srcu);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -0700100 hlist_for_each_entry_rcu(mn, n, &mm->mmu_notifier_mm->list, hlist) {
101 if (mn->ops->clear_flush_young)
102 young |= mn->ops->clear_flush_young(mn, mm, address);
103 }
Sagi Grimberg21a92732012-10-08 16:29:24 -0700104 srcu_read_unlock(&srcu, id);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -0700105
106 return young;
107}
108
Andrea Arcangeli8ee53822011-01-13 15:47:10 -0800109int __mmu_notifier_test_young(struct mm_struct *mm,
110 unsigned long address)
111{
112 struct mmu_notifier *mn;
113 struct hlist_node *n;
Sagi Grimberg21a92732012-10-08 16:29:24 -0700114 int young = 0, id;
Andrea Arcangeli8ee53822011-01-13 15:47:10 -0800115
Sagi Grimberg21a92732012-10-08 16:29:24 -0700116 id = srcu_read_lock(&srcu);
Andrea Arcangeli8ee53822011-01-13 15:47:10 -0800117 hlist_for_each_entry_rcu(mn, n, &mm->mmu_notifier_mm->list, hlist) {
118 if (mn->ops->test_young) {
119 young = mn->ops->test_young(mn, mm, address);
120 if (young)
121 break;
122 }
123 }
Sagi Grimberg21a92732012-10-08 16:29:24 -0700124 srcu_read_unlock(&srcu, id);
Andrea Arcangeli8ee53822011-01-13 15:47:10 -0800125
126 return young;
127}
128
Izik Eidus828502d2009-09-21 17:01:51 -0700129void __mmu_notifier_change_pte(struct mm_struct *mm, unsigned long address,
130 pte_t pte)
131{
132 struct mmu_notifier *mn;
133 struct hlist_node *n;
Sagi Grimberg21a92732012-10-08 16:29:24 -0700134 int id;
Izik Eidus828502d2009-09-21 17:01:51 -0700135
Sagi Grimberg21a92732012-10-08 16:29:24 -0700136 id = srcu_read_lock(&srcu);
Izik Eidus828502d2009-09-21 17:01:51 -0700137 hlist_for_each_entry_rcu(mn, n, &mm->mmu_notifier_mm->list, hlist) {
138 if (mn->ops->change_pte)
139 mn->ops->change_pte(mn, mm, address, pte);
Izik Eidus828502d2009-09-21 17:01:51 -0700140 }
Sagi Grimberg21a92732012-10-08 16:29:24 -0700141 srcu_read_unlock(&srcu, id);
Izik Eidus828502d2009-09-21 17:01:51 -0700142}
143
Andrea Arcangelicddb8a52008-07-28 15:46:29 -0700144void __mmu_notifier_invalidate_page(struct mm_struct *mm,
145 unsigned long address)
146{
147 struct mmu_notifier *mn;
148 struct hlist_node *n;
Sagi Grimberg21a92732012-10-08 16:29:24 -0700149 int id;
Andrea Arcangelicddb8a52008-07-28 15:46:29 -0700150
Sagi Grimberg21a92732012-10-08 16:29:24 -0700151 id = srcu_read_lock(&srcu);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -0700152 hlist_for_each_entry_rcu(mn, n, &mm->mmu_notifier_mm->list, hlist) {
153 if (mn->ops->invalidate_page)
154 mn->ops->invalidate_page(mn, mm, address);
155 }
Sagi Grimberg21a92732012-10-08 16:29:24 -0700156 srcu_read_unlock(&srcu, id);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -0700157}
158
159void __mmu_notifier_invalidate_range_start(struct mm_struct *mm,
160 unsigned long start, unsigned long end)
161{
162 struct mmu_notifier *mn;
163 struct hlist_node *n;
Sagi Grimberg21a92732012-10-08 16:29:24 -0700164 int id;
Andrea Arcangelicddb8a52008-07-28 15:46:29 -0700165
Sagi Grimberg21a92732012-10-08 16:29:24 -0700166 id = srcu_read_lock(&srcu);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -0700167 hlist_for_each_entry_rcu(mn, n, &mm->mmu_notifier_mm->list, hlist) {
168 if (mn->ops->invalidate_range_start)
169 mn->ops->invalidate_range_start(mn, mm, start, end);
170 }
Sagi Grimberg21a92732012-10-08 16:29:24 -0700171 srcu_read_unlock(&srcu, id);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -0700172}
173
174void __mmu_notifier_invalidate_range_end(struct mm_struct *mm,
175 unsigned long start, unsigned long end)
176{
177 struct mmu_notifier *mn;
178 struct hlist_node *n;
Sagi Grimberg21a92732012-10-08 16:29:24 -0700179 int id;
Andrea Arcangelicddb8a52008-07-28 15:46:29 -0700180
Sagi Grimberg21a92732012-10-08 16:29:24 -0700181 id = srcu_read_lock(&srcu);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -0700182 hlist_for_each_entry_rcu(mn, n, &mm->mmu_notifier_mm->list, hlist) {
183 if (mn->ops->invalidate_range_end)
184 mn->ops->invalidate_range_end(mn, mm, start, end);
185 }
Sagi Grimberg21a92732012-10-08 16:29:24 -0700186 srcu_read_unlock(&srcu, id);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -0700187}
188
189static int do_mmu_notifier_register(struct mmu_notifier *mn,
190 struct mm_struct *mm,
191 int take_mmap_sem)
192{
193 struct mmu_notifier_mm *mmu_notifier_mm;
194 int ret;
195
196 BUG_ON(atomic_read(&mm->mm_users) <= 0);
197
Sagi Grimberg21a92732012-10-08 16:29:24 -0700198 /*
199 * Verify that mmu_notifier_init() already run and the global srcu is
200 * initialized.
201 */
202 BUG_ON(!srcu.per_cpu_ref);
203
Andrea Arcangelicddb8a52008-07-28 15:46:29 -0700204 if (take_mmap_sem)
205 down_write(&mm->mmap_sem);
206 ret = mm_take_all_locks(mm);
207 if (unlikely(ret))
Gavin Shane0f3c3f2012-10-08 16:29:26 -0700208 goto out;
Andrea Arcangelicddb8a52008-07-28 15:46:29 -0700209
210 if (!mm_has_notifiers(mm)) {
Gavin Shane0f3c3f2012-10-08 16:29:26 -0700211 mmu_notifier_mm = kmalloc(sizeof(struct mmu_notifier_mm),
212 GFP_KERNEL);
213 if (unlikely(!mmu_notifier_mm)) {
214 ret = -ENOMEM;
215 goto out_of_mem;
216 }
Andrea Arcangelicddb8a52008-07-28 15:46:29 -0700217 INIT_HLIST_HEAD(&mmu_notifier_mm->list);
218 spin_lock_init(&mmu_notifier_mm->lock);
Gavin Shane0f3c3f2012-10-08 16:29:26 -0700219
Andrea Arcangelicddb8a52008-07-28 15:46:29 -0700220 mm->mmu_notifier_mm = mmu_notifier_mm;
Andrea Arcangelicddb8a52008-07-28 15:46:29 -0700221 }
222 atomic_inc(&mm->mm_count);
223
224 /*
225 * Serialize the update against mmu_notifier_unregister. A
226 * side note: mmu_notifier_release can't run concurrently with
227 * us because we hold the mm_users pin (either implicitly as
228 * current->mm or explicitly with get_task_mm() or similar).
229 * We can't race against any other mmu notifier method either
230 * thanks to mm_take_all_locks().
231 */
232 spin_lock(&mm->mmu_notifier_mm->lock);
233 hlist_add_head(&mn->hlist, &mm->mmu_notifier_mm->list);
234 spin_unlock(&mm->mmu_notifier_mm->lock);
235
Gavin Shane0f3c3f2012-10-08 16:29:26 -0700236out_of_mem:
Andrea Arcangelicddb8a52008-07-28 15:46:29 -0700237 mm_drop_all_locks(mm);
Gavin Shane0f3c3f2012-10-08 16:29:26 -0700238out:
Andrea Arcangelicddb8a52008-07-28 15:46:29 -0700239 if (take_mmap_sem)
240 up_write(&mm->mmap_sem);
Gavin Shane0f3c3f2012-10-08 16:29:26 -0700241
Andrea Arcangelicddb8a52008-07-28 15:46:29 -0700242 BUG_ON(atomic_read(&mm->mm_users) <= 0);
243 return ret;
244}
245
246/*
247 * Must not hold mmap_sem nor any other VM related lock when calling
248 * this registration function. Must also ensure mm_users can't go down
249 * to zero while this runs to avoid races with mmu_notifier_release,
250 * so mm has to be current->mm or the mm should be pinned safely such
251 * as with get_task_mm(). If the mm is not current->mm, the mm_users
252 * pin should be released by calling mmput after mmu_notifier_register
253 * returns. mmu_notifier_unregister must be always called to
254 * unregister the notifier. mm_count is automatically pinned to allow
255 * mmu_notifier_unregister to safely run at any time later, before or
256 * after exit_mmap. ->release will always be called before exit_mmap
257 * frees the pages.
258 */
259int mmu_notifier_register(struct mmu_notifier *mn, struct mm_struct *mm)
260{
261 return do_mmu_notifier_register(mn, mm, 1);
262}
263EXPORT_SYMBOL_GPL(mmu_notifier_register);
264
265/*
266 * Same as mmu_notifier_register but here the caller must hold the
267 * mmap_sem in write mode.
268 */
269int __mmu_notifier_register(struct mmu_notifier *mn, struct mm_struct *mm)
270{
271 return do_mmu_notifier_register(mn, mm, 0);
272}
273EXPORT_SYMBOL_GPL(__mmu_notifier_register);
274
275/* this is called after the last mmu_notifier_unregister() returned */
276void __mmu_notifier_mm_destroy(struct mm_struct *mm)
277{
278 BUG_ON(!hlist_empty(&mm->mmu_notifier_mm->list));
279 kfree(mm->mmu_notifier_mm);
280 mm->mmu_notifier_mm = LIST_POISON1; /* debug */
281}
282
283/*
284 * This releases the mm_count pin automatically and frees the mm
285 * structure if it was the last user of it. It serializes against
Sagi Grimberg21a92732012-10-08 16:29:24 -0700286 * running mmu notifiers with SRCU and against mmu_notifier_unregister
287 * with the unregister lock + SRCU. All sptes must be dropped before
Andrea Arcangelicddb8a52008-07-28 15:46:29 -0700288 * calling mmu_notifier_unregister. ->release or any other notifier
289 * method may be invoked concurrently with mmu_notifier_unregister,
290 * and only after mmu_notifier_unregister returned we're guaranteed
291 * that ->release or any other method can't run anymore.
292 */
293void mmu_notifier_unregister(struct mmu_notifier *mn, struct mm_struct *mm)
294{
295 BUG_ON(atomic_read(&mm->mm_count) <= 0);
296
Andrea Arcangelicddb8a52008-07-28 15:46:29 -0700297 if (!hlist_unhashed(&mn->hlist)) {
Andrea Arcangelicddb8a52008-07-28 15:46:29 -0700298 /*
Andrea Arcangeli70400302012-10-08 16:31:52 -0700299 * SRCU here will force exit_mmap to wait ->release to finish
Andrea Arcangelicddb8a52008-07-28 15:46:29 -0700300 * before freeing the pages.
301 */
Sagi Grimberg21a92732012-10-08 16:29:24 -0700302 int id;
Xiao Guangrong3ad3d902012-07-31 16:45:52 -0700303
Sagi Grimberg21a92732012-10-08 16:29:24 -0700304 id = srcu_read_lock(&srcu);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -0700305 /*
306 * exit_mmap will block in mmu_notifier_release to
307 * guarantee ->release is called before freeing the
308 * pages.
309 */
310 if (mn->ops->release)
311 mn->ops->release(mn, mm);
Sagi Grimberg21a92732012-10-08 16:29:24 -0700312 srcu_read_unlock(&srcu, id);
Xiao Guangrong3ad3d902012-07-31 16:45:52 -0700313
314 spin_lock(&mm->mmu_notifier_mm->lock);
315 hlist_del_rcu(&mn->hlist);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -0700316 spin_unlock(&mm->mmu_notifier_mm->lock);
Xiao Guangrong3ad3d902012-07-31 16:45:52 -0700317 }
Andrea Arcangelicddb8a52008-07-28 15:46:29 -0700318
319 /*
320 * Wait any running method to finish, of course including
321 * ->release if it was run by mmu_notifier_relase instead of us.
322 */
Sagi Grimberg21a92732012-10-08 16:29:24 -0700323 synchronize_srcu(&srcu);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -0700324
325 BUG_ON(atomic_read(&mm->mm_count) <= 0);
326
327 mmdrop(mm);
328}
329EXPORT_SYMBOL_GPL(mmu_notifier_unregister);
Sagi Grimberg21a92732012-10-08 16:29:24 -0700330
331static int __init mmu_notifier_init(void)
332{
333 return init_srcu_struct(&srcu);
334}
335
336module_init(mmu_notifier_init);