blob: f4259e496f83a6d0465bba0a95e918980395030f [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
Christoph Lameter93e205a2016-03-17 14:21:15 -07006 * Christoph Lameter <cl@linux.com>
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07007 *
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/*
Peter Zijlstrab9722162014-08-06 16:08:20 -070026 * 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/*
Andrea Arcangelicddb8a52008-07-28 15:46:29 -070045 * 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
Sagi Grimberg21a92732012-10-08 16:29:24 -070051 * the mmu_notifier_mm->lock in addition to SRCU and it serializes
52 * against the other mmu notifiers with SRCU. struct mmu_notifier_mm
Andrea Arcangelicddb8a52008-07-28 15:46:29 -070053 * 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;
Sagi Grimberg21a92732012-10-08 16:29:24 -070059 int id;
Xiao Guangrong3ad3d902012-07-31 16:45:52 -070060
61 /*
Xiao Guangrongd34883d2013-05-24 15:55:11 -070062 * SRCU here will block mmu_notifier_unregister until
63 * ->release returns.
Xiao Guangrong3ad3d902012-07-31 16:45:52 -070064 */
Sagi Grimberg21a92732012-10-08 16:29:24 -070065 id = srcu_read_lock(&srcu);
Xiao Guangrongd34883d2013-05-24 15:55:11 -070066 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);
Xiao Guangrongd34883d2013-05-24 15:55:11 -070075
Andrea Arcangelicddb8a52008-07-28 15:46:29 -070076 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 /*
Xiao Guangrongd34883d2013-05-24 15:55:11 -070082 * 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.
Andrea Arcangelicddb8a52008-07-28 15:46:29 -070086 */
87 hlist_del_init_rcu(&mn->hlist);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -070088 }
89 spin_unlock(&mm->mmu_notifier_mm->lock);
Peter Zijlstrab9722162014-08-06 16:08:20 -070090 srcu_read_unlock(&srcu, id);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -070091
92 /*
Xiao Guangrongd34883d2013-05-24 15:55:11 -070093 * 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.
Andrea Arcangelicddb8a52008-07-28 15:46:29 -0700100 */
Sagi Grimberg21a92732012-10-08 16:29:24 -0700101 synchronize_srcu(&srcu);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -0700102}
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,
Andres Lagar-Cavilla57128462014-09-22 14:54:42 -0700110 unsigned long start,
111 unsigned long end)
Andrea Arcangelicddb8a52008-07-28 15:46:29 -0700112{
113 struct mmu_notifier *mn;
Sagi Grimberg21a92732012-10-08 16:29:24 -0700114 int young = 0, id;
Andrea Arcangelicddb8a52008-07-28 15:46:29 -0700115
Sagi Grimberg21a92732012-10-08 16:29:24 -0700116 id = srcu_read_lock(&srcu);
Sasha Levinb67bfe02013-02-27 17:06:00 -0800117 hlist_for_each_entry_rcu(mn, &mm->mmu_notifier_mm->list, hlist) {
Andrea Arcangelicddb8a52008-07-28 15:46:29 -0700118 if (mn->ops->clear_flush_young)
Andres Lagar-Cavilla57128462014-09-22 14:54:42 -0700119 young |= mn->ops->clear_flush_young(mn, mm, start, end);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -0700120 }
Sagi Grimberg21a92732012-10-08 16:29:24 -0700121 srcu_read_unlock(&srcu, id);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -0700122
123 return young;
124}
125
Vladimir Davydov1d7715c2015-09-09 15:35:41 -0700126int __mmu_notifier_clear_young(struct mm_struct *mm,
127 unsigned long start,
128 unsigned long end)
129{
130 struct mmu_notifier *mn;
131 int young = 0, id;
132
133 id = srcu_read_lock(&srcu);
134 hlist_for_each_entry_rcu(mn, &mm->mmu_notifier_mm->list, hlist) {
135 if (mn->ops->clear_young)
136 young |= mn->ops->clear_young(mn, mm, start, end);
137 }
138 srcu_read_unlock(&srcu, id);
139
140 return young;
141}
142
Andrea Arcangeli8ee53822011-01-13 15:47:10 -0800143int __mmu_notifier_test_young(struct mm_struct *mm,
144 unsigned long address)
145{
146 struct mmu_notifier *mn;
Sagi Grimberg21a92732012-10-08 16:29:24 -0700147 int young = 0, id;
Andrea Arcangeli8ee53822011-01-13 15:47:10 -0800148
Sagi Grimberg21a92732012-10-08 16:29:24 -0700149 id = srcu_read_lock(&srcu);
Sasha Levinb67bfe02013-02-27 17:06:00 -0800150 hlist_for_each_entry_rcu(mn, &mm->mmu_notifier_mm->list, hlist) {
Andrea Arcangeli8ee53822011-01-13 15:47:10 -0800151 if (mn->ops->test_young) {
152 young = mn->ops->test_young(mn, mm, address);
153 if (young)
154 break;
155 }
156 }
Sagi Grimberg21a92732012-10-08 16:29:24 -0700157 srcu_read_unlock(&srcu, id);
Andrea Arcangeli8ee53822011-01-13 15:47:10 -0800158
159 return young;
160}
161
Izik Eidus828502d2009-09-21 17:01:51 -0700162void __mmu_notifier_change_pte(struct mm_struct *mm, unsigned long address,
163 pte_t pte)
164{
165 struct mmu_notifier *mn;
Sagi Grimberg21a92732012-10-08 16:29:24 -0700166 int id;
Izik Eidus828502d2009-09-21 17:01:51 -0700167
Sagi Grimberg21a92732012-10-08 16:29:24 -0700168 id = srcu_read_lock(&srcu);
Sasha Levinb67bfe02013-02-27 17:06:00 -0800169 hlist_for_each_entry_rcu(mn, &mm->mmu_notifier_mm->list, hlist) {
Izik Eidus828502d2009-09-21 17:01:51 -0700170 if (mn->ops->change_pte)
171 mn->ops->change_pte(mn, mm, address, pte);
Izik Eidus828502d2009-09-21 17:01:51 -0700172 }
Sagi Grimberg21a92732012-10-08 16:29:24 -0700173 srcu_read_unlock(&srcu, id);
Izik Eidus828502d2009-09-21 17:01:51 -0700174}
175
Andrea Arcangelicddb8a52008-07-28 15:46:29 -0700176void __mmu_notifier_invalidate_page(struct mm_struct *mm,
177 unsigned long address)
178{
179 struct mmu_notifier *mn;
Sagi Grimberg21a92732012-10-08 16:29:24 -0700180 int id;
Andrea Arcangelicddb8a52008-07-28 15:46:29 -0700181
Sagi Grimberg21a92732012-10-08 16:29:24 -0700182 id = srcu_read_lock(&srcu);
Sasha Levinb67bfe02013-02-27 17:06:00 -0800183 hlist_for_each_entry_rcu(mn, &mm->mmu_notifier_mm->list, hlist) {
Andrea Arcangelicddb8a52008-07-28 15:46:29 -0700184 if (mn->ops->invalidate_page)
185 mn->ops->invalidate_page(mn, mm, address);
186 }
Sagi Grimberg21a92732012-10-08 16:29:24 -0700187 srcu_read_unlock(&srcu, id);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -0700188}
189
190void __mmu_notifier_invalidate_range_start(struct mm_struct *mm,
191 unsigned long start, unsigned long end)
192{
193 struct mmu_notifier *mn;
Sagi Grimberg21a92732012-10-08 16:29:24 -0700194 int id;
Andrea Arcangelicddb8a52008-07-28 15:46:29 -0700195
Sagi Grimberg21a92732012-10-08 16:29:24 -0700196 id = srcu_read_lock(&srcu);
Sasha Levinb67bfe02013-02-27 17:06:00 -0800197 hlist_for_each_entry_rcu(mn, &mm->mmu_notifier_mm->list, hlist) {
Andrea Arcangelicddb8a52008-07-28 15:46:29 -0700198 if (mn->ops->invalidate_range_start)
199 mn->ops->invalidate_range_start(mn, mm, start, end);
200 }
Sagi Grimberg21a92732012-10-08 16:29:24 -0700201 srcu_read_unlock(&srcu, id);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -0700202}
Cliff Wickmanfa794192013-02-22 16:35:58 -0800203EXPORT_SYMBOL_GPL(__mmu_notifier_invalidate_range_start);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -0700204
205void __mmu_notifier_invalidate_range_end(struct mm_struct *mm,
206 unsigned long start, unsigned long end)
207{
208 struct mmu_notifier *mn;
Sagi Grimberg21a92732012-10-08 16:29:24 -0700209 int id;
Andrea Arcangelicddb8a52008-07-28 15:46:29 -0700210
Sagi Grimberg21a92732012-10-08 16:29:24 -0700211 id = srcu_read_lock(&srcu);
Sasha Levinb67bfe02013-02-27 17:06:00 -0800212 hlist_for_each_entry_rcu(mn, &mm->mmu_notifier_mm->list, hlist) {
Joerg Roedel0f0a3272014-11-13 13:46:09 +1100213 /*
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).
220 */
221 if (mn->ops->invalidate_range)
222 mn->ops->invalidate_range(mn, mm, start, end);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -0700223 if (mn->ops->invalidate_range_end)
224 mn->ops->invalidate_range_end(mn, mm, start, end);
225 }
Sagi Grimberg21a92732012-10-08 16:29:24 -0700226 srcu_read_unlock(&srcu, id);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -0700227}
Cliff Wickmanfa794192013-02-22 16:35:58 -0800228EXPORT_SYMBOL_GPL(__mmu_notifier_invalidate_range_end);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -0700229
Joerg Roedel0f0a3272014-11-13 13:46:09 +1100230void __mmu_notifier_invalidate_range(struct mm_struct *mm,
231 unsigned long start, unsigned long end)
232{
233 struct mmu_notifier *mn;
234 int id;
235
236 id = srcu_read_lock(&srcu);
237 hlist_for_each_entry_rcu(mn, &mm->mmu_notifier_mm->list, hlist) {
238 if (mn->ops->invalidate_range)
239 mn->ops->invalidate_range(mn, mm, start, end);
240 }
241 srcu_read_unlock(&srcu, id);
242}
243EXPORT_SYMBOL_GPL(__mmu_notifier_invalidate_range);
244
Andrea Arcangelicddb8a52008-07-28 15:46:29 -0700245static int do_mmu_notifier_register(struct mmu_notifier *mn,
246 struct mm_struct *mm,
247 int take_mmap_sem)
248{
249 struct mmu_notifier_mm *mmu_notifier_mm;
250 int ret;
251
252 BUG_ON(atomic_read(&mm->mm_users) <= 0);
253
Sagi Grimberg21a92732012-10-08 16:29:24 -0700254 /*
Gavin Shan35cfa2b2012-10-25 13:38:01 -0700255 * Verify that mmu_notifier_init() already run and the global srcu is
256 * initialized.
257 */
Sagi Grimberg21a92732012-10-08 16:29:24 -0700258 BUG_ON(!srcu.per_cpu_ref);
259
Gavin Shan35cfa2b2012-10-25 13:38:01 -0700260 ret = -ENOMEM;
261 mmu_notifier_mm = kmalloc(sizeof(struct mmu_notifier_mm), GFP_KERNEL);
262 if (unlikely(!mmu_notifier_mm))
263 goto out;
264
Andrea Arcangelicddb8a52008-07-28 15:46:29 -0700265 if (take_mmap_sem)
266 down_write(&mm->mmap_sem);
267 ret = mm_take_all_locks(mm);
268 if (unlikely(ret))
Gavin Shan35cfa2b2012-10-25 13:38:01 -0700269 goto out_clean;
Andrea Arcangelicddb8a52008-07-28 15:46:29 -0700270
271 if (!mm_has_notifiers(mm)) {
272 INIT_HLIST_HEAD(&mmu_notifier_mm->list);
273 spin_lock_init(&mmu_notifier_mm->lock);
Gavin Shane0f3c3f2012-10-08 16:29:26 -0700274
Andrea Arcangelicddb8a52008-07-28 15:46:29 -0700275 mm->mmu_notifier_mm = mmu_notifier_mm;
Gavin Shan35cfa2b2012-10-25 13:38:01 -0700276 mmu_notifier_mm = NULL;
Andrea Arcangelicddb8a52008-07-28 15:46:29 -0700277 }
278 atomic_inc(&mm->mm_count);
279
280 /*
281 * Serialize the update against mmu_notifier_unregister. A
282 * side note: mmu_notifier_release can't run concurrently with
283 * us because we hold the mm_users pin (either implicitly as
284 * current->mm or explicitly with get_task_mm() or similar).
285 * We can't race against any other mmu notifier method either
286 * thanks to mm_take_all_locks().
287 */
288 spin_lock(&mm->mmu_notifier_mm->lock);
289 hlist_add_head(&mn->hlist, &mm->mmu_notifier_mm->list);
290 spin_unlock(&mm->mmu_notifier_mm->lock);
291
292 mm_drop_all_locks(mm);
Gavin Shan35cfa2b2012-10-25 13:38:01 -0700293out_clean:
Andrea Arcangelicddb8a52008-07-28 15:46:29 -0700294 if (take_mmap_sem)
295 up_write(&mm->mmap_sem);
Gavin Shan35cfa2b2012-10-25 13:38:01 -0700296 kfree(mmu_notifier_mm);
297out:
Andrea Arcangelicddb8a52008-07-28 15:46:29 -0700298 BUG_ON(atomic_read(&mm->mm_users) <= 0);
299 return ret;
300}
301
302/*
303 * Must not hold mmap_sem nor any other VM related lock when calling
304 * this registration function. Must also ensure mm_users can't go down
305 * to zero while this runs to avoid races with mmu_notifier_release,
306 * so mm has to be current->mm or the mm should be pinned safely such
307 * as with get_task_mm(). If the mm is not current->mm, the mm_users
308 * pin should be released by calling mmput after mmu_notifier_register
309 * returns. mmu_notifier_unregister must be always called to
310 * unregister the notifier. mm_count is automatically pinned to allow
311 * mmu_notifier_unregister to safely run at any time later, before or
312 * after exit_mmap. ->release will always be called before exit_mmap
313 * frees the pages.
314 */
315int mmu_notifier_register(struct mmu_notifier *mn, struct mm_struct *mm)
316{
317 return do_mmu_notifier_register(mn, mm, 1);
318}
319EXPORT_SYMBOL_GPL(mmu_notifier_register);
320
321/*
322 * Same as mmu_notifier_register but here the caller must hold the
323 * mmap_sem in write mode.
324 */
325int __mmu_notifier_register(struct mmu_notifier *mn, struct mm_struct *mm)
326{
327 return do_mmu_notifier_register(mn, mm, 0);
328}
329EXPORT_SYMBOL_GPL(__mmu_notifier_register);
330
331/* this is called after the last mmu_notifier_unregister() returned */
332void __mmu_notifier_mm_destroy(struct mm_struct *mm)
333{
334 BUG_ON(!hlist_empty(&mm->mmu_notifier_mm->list));
335 kfree(mm->mmu_notifier_mm);
336 mm->mmu_notifier_mm = LIST_POISON1; /* debug */
337}
338
339/*
340 * This releases the mm_count pin automatically and frees the mm
341 * structure if it was the last user of it. It serializes against
Sagi Grimberg21a92732012-10-08 16:29:24 -0700342 * running mmu notifiers with SRCU and against mmu_notifier_unregister
343 * with the unregister lock + SRCU. All sptes must be dropped before
Andrea Arcangelicddb8a52008-07-28 15:46:29 -0700344 * calling mmu_notifier_unregister. ->release or any other notifier
345 * method may be invoked concurrently with mmu_notifier_unregister,
346 * and only after mmu_notifier_unregister returned we're guaranteed
347 * that ->release or any other method can't run anymore.
348 */
349void mmu_notifier_unregister(struct mmu_notifier *mn, struct mm_struct *mm)
350{
351 BUG_ON(atomic_read(&mm->mm_count) <= 0);
352
Andrea Arcangelicddb8a52008-07-28 15:46:29 -0700353 if (!hlist_unhashed(&mn->hlist)) {
Xiao Guangrongd34883d2013-05-24 15:55:11 -0700354 /*
355 * SRCU here will force exit_mmap to wait for ->release to
356 * finish before freeing the pages.
357 */
Sagi Grimberg21a92732012-10-08 16:29:24 -0700358 int id;
Xiao Guangrong3ad3d902012-07-31 16:45:52 -0700359
Robin Holt751efd82013-02-22 16:35:34 -0800360 id = srcu_read_lock(&srcu);
Xiao Guangrongd34883d2013-05-24 15:55:11 -0700361 /*
362 * exit_mmap will block in mmu_notifier_release to guarantee
363 * that ->release is called before freeing the pages.
364 */
Robin Holt751efd82013-02-22 16:35:34 -0800365 if (mn->ops->release)
366 mn->ops->release(mn, mm);
Robin Holt751efd82013-02-22 16:35:34 -0800367 srcu_read_unlock(&srcu, id);
Xiao Guangrongd34883d2013-05-24 15:55:11 -0700368
369 spin_lock(&mm->mmu_notifier_mm->lock);
370 /*
371 * Can not use list_del_rcu() since __mmu_notifier_release
372 * can delete it before we hold the lock.
373 */
374 hlist_del_init_rcu(&mn->hlist);
Robin Holt751efd82013-02-22 16:35:34 -0800375 spin_unlock(&mm->mmu_notifier_mm->lock);
Xiao Guangrongd34883d2013-05-24 15:55:11 -0700376 }
Andrea Arcangelicddb8a52008-07-28 15:46:29 -0700377
378 /*
Xiao Guangrongd34883d2013-05-24 15:55:11 -0700379 * Wait for any running method to finish, of course including
Geert Uytterhoeven83a35e32013-06-28 11:27:31 +0200380 * ->release if it was run by mmu_notifier_release instead of us.
Andrea Arcangelicddb8a52008-07-28 15:46:29 -0700381 */
Sagi Grimberg21a92732012-10-08 16:29:24 -0700382 synchronize_srcu(&srcu);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -0700383
384 BUG_ON(atomic_read(&mm->mm_count) <= 0);
385
386 mmdrop(mm);
387}
388EXPORT_SYMBOL_GPL(mmu_notifier_unregister);
Sagi Grimberg21a92732012-10-08 16:29:24 -0700389
Peter Zijlstrab9722162014-08-06 16:08:20 -0700390/*
391 * Same as mmu_notifier_unregister but no callback and no srcu synchronization.
392 */
393void mmu_notifier_unregister_no_release(struct mmu_notifier *mn,
394 struct mm_struct *mm)
395{
396 spin_lock(&mm->mmu_notifier_mm->lock);
397 /*
398 * Can not use list_del_rcu() since __mmu_notifier_release
399 * can delete it before we hold the lock.
400 */
401 hlist_del_init_rcu(&mn->hlist);
402 spin_unlock(&mm->mmu_notifier_mm->lock);
403
404 BUG_ON(atomic_read(&mm->mm_count) <= 0);
405 mmdrop(mm);
406}
407EXPORT_SYMBOL_GPL(mmu_notifier_unregister_no_release);
408
Sagi Grimberg21a92732012-10-08 16:29:24 -0700409static int __init mmu_notifier_init(void)
410{
411 return init_srcu_struct(&srcu);
412}
Paul Gortmakera64fb3c2014-01-23 15:53:30 -0800413subsys_initcall(mmu_notifier_init);