]>
Commit | Line | Data |
---|---|---|
cddb8a5c AA |
1 | /* |
2 | * linux/mm/mmu_notifier.c | |
3 | * | |
4 | * Copyright (C) 2008 Qumranet, Inc. | |
5 | * Copyright (C) 2008 SGI | |
93e205a7 | 6 | * Christoph Lameter <cl@linux.com> |
cddb8a5c AA |
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> | |
b95f1b31 | 14 | #include <linux/export.h> |
cddb8a5c AA |
15 | #include <linux/mm.h> |
16 | #include <linux/err.h> | |
21a92735 | 17 | #include <linux/srcu.h> |
cddb8a5c AA |
18 | #include <linux/rcupdate.h> |
19 | #include <linux/sched.h> | |
6e84f315 | 20 | #include <linux/sched/mm.h> |
5a0e3ad6 | 21 | #include <linux/slab.h> |
cddb8a5c | 22 | |
21a92735 | 23 | /* global SRCU for all MMs */ |
dde8da6c | 24 | DEFINE_STATIC_SRCU(srcu); |
21a92735 | 25 | |
b972216e PZ |
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 | */ | |
31 | void mmu_notifier_call_srcu(struct rcu_head *rcu, | |
32 | void (*func)(struct rcu_head *rcu)) | |
33 | { | |
34 | call_srcu(&srcu, rcu, func); | |
35 | } | |
36 | EXPORT_SYMBOL_GPL(mmu_notifier_call_srcu); | |
37 | ||
38 | void mmu_notifier_synchronize(void) | |
39 | { | |
40 | /* Wait for any running method to finish. */ | |
41 | srcu_barrier(&srcu); | |
42 | } | |
43 | EXPORT_SYMBOL_GPL(mmu_notifier_synchronize); | |
44 | ||
cddb8a5c AA |
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 | |
21a92735 SG |
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 | |
cddb8a5c AA |
54 | * can't go away from under us as exit_mmap holds an mm_count pin |
55 | * itself. | |
56 | */ | |
57 | void __mmu_notifier_release(struct mm_struct *mm) | |
58 | { | |
59 | struct mmu_notifier *mn; | |
21a92735 | 60 | int id; |
3ad3d901 XG |
61 | |
62 | /* | |
d34883d4 XG |
63 | * SRCU here will block mmu_notifier_unregister until |
64 | * ->release returns. | |
3ad3d901 | 65 | */ |
21a92735 | 66 | id = srcu_read_lock(&srcu); |
d34883d4 XG |
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); | |
d34883d4 | 76 | |
cddb8a5c AA |
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 | /* | |
d34883d4 XG |
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. | |
cddb8a5c AA |
87 | */ |
88 | hlist_del_init_rcu(&mn->hlist); | |
cddb8a5c AA |
89 | } |
90 | spin_unlock(&mm->mmu_notifier_mm->lock); | |
b972216e | 91 | srcu_read_unlock(&srcu, id); |
cddb8a5c AA |
92 | |
93 | /* | |
d34883d4 XG |
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. | |
cddb8a5c | 101 | */ |
21a92735 | 102 | synchronize_srcu(&srcu); |
cddb8a5c AA |
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 | */ | |
110 | int __mmu_notifier_clear_flush_young(struct mm_struct *mm, | |
57128468 ALC |
111 | unsigned long start, |
112 | unsigned long end) | |
cddb8a5c AA |
113 | { |
114 | struct mmu_notifier *mn; | |
21a92735 | 115 | int young = 0, id; |
cddb8a5c | 116 | |
21a92735 | 117 | id = srcu_read_lock(&srcu); |
b67bfe0d | 118 | hlist_for_each_entry_rcu(mn, &mm->mmu_notifier_mm->list, hlist) { |
cddb8a5c | 119 | if (mn->ops->clear_flush_young) |
57128468 | 120 | young |= mn->ops->clear_flush_young(mn, mm, start, end); |
cddb8a5c | 121 | } |
21a92735 | 122 | srcu_read_unlock(&srcu, id); |
cddb8a5c AA |
123 | |
124 | return young; | |
125 | } | |
126 | ||
1d7715c6 VD |
127 | int __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 | ||
8ee53820 AA |
144 | int __mmu_notifier_test_young(struct mm_struct *mm, |
145 | unsigned long address) | |
146 | { | |
147 | struct mmu_notifier *mn; | |
21a92735 | 148 | int young = 0, id; |
8ee53820 | 149 | |
21a92735 | 150 | id = srcu_read_lock(&srcu); |
b67bfe0d | 151 | hlist_for_each_entry_rcu(mn, &mm->mmu_notifier_mm->list, hlist) { |
8ee53820 AA |
152 | if (mn->ops->test_young) { |
153 | young = mn->ops->test_young(mn, mm, address); | |
154 | if (young) | |
155 | break; | |
156 | } | |
157 | } | |
21a92735 | 158 | srcu_read_unlock(&srcu, id); |
8ee53820 AA |
159 | |
160 | return young; | |
161 | } | |
162 | ||
828502d3 IE |
163 | void __mmu_notifier_change_pte(struct mm_struct *mm, unsigned long address, |
164 | pte_t pte) | |
165 | { | |
166 | struct mmu_notifier *mn; | |
21a92735 | 167 | int id; |
828502d3 | 168 | |
21a92735 | 169 | id = srcu_read_lock(&srcu); |
b67bfe0d | 170 | hlist_for_each_entry_rcu(mn, &mm->mmu_notifier_mm->list, hlist) { |
828502d3 IE |
171 | if (mn->ops->change_pte) |
172 | mn->ops->change_pte(mn, mm, address, pte); | |
828502d3 | 173 | } |
21a92735 | 174 | srcu_read_unlock(&srcu, id); |
828502d3 IE |
175 | } |
176 | ||
cddb8a5c AA |
177 | void __mmu_notifier_invalidate_range_start(struct mm_struct *mm, |
178 | unsigned long start, unsigned long end) | |
179 | { | |
180 | struct mmu_notifier *mn; | |
21a92735 | 181 | int id; |
cddb8a5c | 182 | |
21a92735 | 183 | id = srcu_read_lock(&srcu); |
b67bfe0d | 184 | hlist_for_each_entry_rcu(mn, &mm->mmu_notifier_mm->list, hlist) { |
cddb8a5c AA |
185 | if (mn->ops->invalidate_range_start) |
186 | mn->ops->invalidate_range_start(mn, mm, start, end); | |
187 | } | |
21a92735 | 188 | srcu_read_unlock(&srcu, id); |
cddb8a5c | 189 | } |
fa794199 | 190 | EXPORT_SYMBOL_GPL(__mmu_notifier_invalidate_range_start); |
cddb8a5c AA |
191 | |
192 | void __mmu_notifier_invalidate_range_end(struct mm_struct *mm, | |
193 | unsigned long start, unsigned long end) | |
194 | { | |
195 | struct mmu_notifier *mn; | |
21a92735 | 196 | int id; |
cddb8a5c | 197 | |
21a92735 | 198 | id = srcu_read_lock(&srcu); |
b67bfe0d | 199 | hlist_for_each_entry_rcu(mn, &mm->mmu_notifier_mm->list, hlist) { |
0f0a327f JR |
200 | /* |
201 | * Call invalidate_range here too to avoid the need for the | |
202 | * subsystem of having to register an invalidate_range_end | |
203 | * call-back when there is invalidate_range already. Usually a | |
204 | * subsystem registers either invalidate_range_start()/end() or | |
205 | * invalidate_range(), so this will be no additional overhead | |
206 | * (besides the pointer check). | |
207 | */ | |
208 | if (mn->ops->invalidate_range) | |
209 | mn->ops->invalidate_range(mn, mm, start, end); | |
cddb8a5c AA |
210 | if (mn->ops->invalidate_range_end) |
211 | mn->ops->invalidate_range_end(mn, mm, start, end); | |
212 | } | |
21a92735 | 213 | srcu_read_unlock(&srcu, id); |
cddb8a5c | 214 | } |
fa794199 | 215 | EXPORT_SYMBOL_GPL(__mmu_notifier_invalidate_range_end); |
cddb8a5c | 216 | |
0f0a327f JR |
217 | void __mmu_notifier_invalidate_range(struct mm_struct *mm, |
218 | unsigned long start, unsigned long end) | |
219 | { | |
220 | struct mmu_notifier *mn; | |
221 | int id; | |
222 | ||
223 | id = srcu_read_lock(&srcu); | |
224 | hlist_for_each_entry_rcu(mn, &mm->mmu_notifier_mm->list, hlist) { | |
225 | if (mn->ops->invalidate_range) | |
226 | mn->ops->invalidate_range(mn, mm, start, end); | |
227 | } | |
228 | srcu_read_unlock(&srcu, id); | |
229 | } | |
230 | EXPORT_SYMBOL_GPL(__mmu_notifier_invalidate_range); | |
231 | ||
cddb8a5c AA |
232 | static int do_mmu_notifier_register(struct mmu_notifier *mn, |
233 | struct mm_struct *mm, | |
234 | int take_mmap_sem) | |
235 | { | |
236 | struct mmu_notifier_mm *mmu_notifier_mm; | |
237 | int ret; | |
238 | ||
239 | BUG_ON(atomic_read(&mm->mm_users) <= 0); | |
240 | ||
35cfa2b0 GS |
241 | ret = -ENOMEM; |
242 | mmu_notifier_mm = kmalloc(sizeof(struct mmu_notifier_mm), GFP_KERNEL); | |
243 | if (unlikely(!mmu_notifier_mm)) | |
244 | goto out; | |
245 | ||
cddb8a5c AA |
246 | if (take_mmap_sem) |
247 | down_write(&mm->mmap_sem); | |
248 | ret = mm_take_all_locks(mm); | |
249 | if (unlikely(ret)) | |
35cfa2b0 | 250 | goto out_clean; |
cddb8a5c AA |
251 | |
252 | if (!mm_has_notifiers(mm)) { | |
253 | INIT_HLIST_HEAD(&mmu_notifier_mm->list); | |
254 | spin_lock_init(&mmu_notifier_mm->lock); | |
e0f3c3f7 | 255 | |
cddb8a5c | 256 | mm->mmu_notifier_mm = mmu_notifier_mm; |
35cfa2b0 | 257 | mmu_notifier_mm = NULL; |
cddb8a5c | 258 | } |
f1f10076 | 259 | mmgrab(mm); |
cddb8a5c AA |
260 | |
261 | /* | |
262 | * Serialize the update against mmu_notifier_unregister. A | |
263 | * side note: mmu_notifier_release can't run concurrently with | |
264 | * us because we hold the mm_users pin (either implicitly as | |
265 | * current->mm or explicitly with get_task_mm() or similar). | |
266 | * We can't race against any other mmu notifier method either | |
267 | * thanks to mm_take_all_locks(). | |
268 | */ | |
269 | spin_lock(&mm->mmu_notifier_mm->lock); | |
270 | hlist_add_head(&mn->hlist, &mm->mmu_notifier_mm->list); | |
271 | spin_unlock(&mm->mmu_notifier_mm->lock); | |
272 | ||
273 | mm_drop_all_locks(mm); | |
35cfa2b0 | 274 | out_clean: |
cddb8a5c AA |
275 | if (take_mmap_sem) |
276 | up_write(&mm->mmap_sem); | |
35cfa2b0 GS |
277 | kfree(mmu_notifier_mm); |
278 | out: | |
cddb8a5c AA |
279 | BUG_ON(atomic_read(&mm->mm_users) <= 0); |
280 | return ret; | |
281 | } | |
282 | ||
283 | /* | |
284 | * Must not hold mmap_sem nor any other VM related lock when calling | |
285 | * this registration function. Must also ensure mm_users can't go down | |
286 | * to zero while this runs to avoid races with mmu_notifier_release, | |
287 | * so mm has to be current->mm or the mm should be pinned safely such | |
288 | * as with get_task_mm(). If the mm is not current->mm, the mm_users | |
289 | * pin should be released by calling mmput after mmu_notifier_register | |
290 | * returns. mmu_notifier_unregister must be always called to | |
291 | * unregister the notifier. mm_count is automatically pinned to allow | |
292 | * mmu_notifier_unregister to safely run at any time later, before or | |
293 | * after exit_mmap. ->release will always be called before exit_mmap | |
294 | * frees the pages. | |
295 | */ | |
296 | int mmu_notifier_register(struct mmu_notifier *mn, struct mm_struct *mm) | |
297 | { | |
298 | return do_mmu_notifier_register(mn, mm, 1); | |
299 | } | |
300 | EXPORT_SYMBOL_GPL(mmu_notifier_register); | |
301 | ||
302 | /* | |
303 | * Same as mmu_notifier_register but here the caller must hold the | |
304 | * mmap_sem in write mode. | |
305 | */ | |
306 | int __mmu_notifier_register(struct mmu_notifier *mn, struct mm_struct *mm) | |
307 | { | |
308 | return do_mmu_notifier_register(mn, mm, 0); | |
309 | } | |
310 | EXPORT_SYMBOL_GPL(__mmu_notifier_register); | |
311 | ||
312 | /* this is called after the last mmu_notifier_unregister() returned */ | |
313 | void __mmu_notifier_mm_destroy(struct mm_struct *mm) | |
314 | { | |
315 | BUG_ON(!hlist_empty(&mm->mmu_notifier_mm->list)); | |
316 | kfree(mm->mmu_notifier_mm); | |
317 | mm->mmu_notifier_mm = LIST_POISON1; /* debug */ | |
318 | } | |
319 | ||
320 | /* | |
321 | * This releases the mm_count pin automatically and frees the mm | |
322 | * structure if it was the last user of it. It serializes against | |
21a92735 SG |
323 | * running mmu notifiers with SRCU and against mmu_notifier_unregister |
324 | * with the unregister lock + SRCU. All sptes must be dropped before | |
cddb8a5c AA |
325 | * calling mmu_notifier_unregister. ->release or any other notifier |
326 | * method may be invoked concurrently with mmu_notifier_unregister, | |
327 | * and only after mmu_notifier_unregister returned we're guaranteed | |
328 | * that ->release or any other method can't run anymore. | |
329 | */ | |
330 | void mmu_notifier_unregister(struct mmu_notifier *mn, struct mm_struct *mm) | |
331 | { | |
332 | BUG_ON(atomic_read(&mm->mm_count) <= 0); | |
333 | ||
cddb8a5c | 334 | if (!hlist_unhashed(&mn->hlist)) { |
d34883d4 XG |
335 | /* |
336 | * SRCU here will force exit_mmap to wait for ->release to | |
337 | * finish before freeing the pages. | |
338 | */ | |
21a92735 | 339 | int id; |
3ad3d901 | 340 | |
d34883d4 | 341 | id = srcu_read_lock(&srcu); |
cddb8a5c | 342 | /* |
d34883d4 XG |
343 | * exit_mmap will block in mmu_notifier_release to guarantee |
344 | * that ->release is called before freeing the pages. | |
cddb8a5c AA |
345 | */ |
346 | if (mn->ops->release) | |
347 | mn->ops->release(mn, mm); | |
d34883d4 | 348 | srcu_read_unlock(&srcu, id); |
3ad3d901 | 349 | |
d34883d4 | 350 | spin_lock(&mm->mmu_notifier_mm->lock); |
751efd86 | 351 | /* |
d34883d4 XG |
352 | * Can not use list_del_rcu() since __mmu_notifier_release |
353 | * can delete it before we hold the lock. | |
751efd86 | 354 | */ |
d34883d4 | 355 | hlist_del_init_rcu(&mn->hlist); |
cddb8a5c | 356 | spin_unlock(&mm->mmu_notifier_mm->lock); |
d34883d4 | 357 | } |
cddb8a5c AA |
358 | |
359 | /* | |
d34883d4 | 360 | * Wait for any running method to finish, of course including |
83a35e36 | 361 | * ->release if it was run by mmu_notifier_release instead of us. |
cddb8a5c | 362 | */ |
21a92735 | 363 | synchronize_srcu(&srcu); |
cddb8a5c AA |
364 | |
365 | BUG_ON(atomic_read(&mm->mm_count) <= 0); | |
366 | ||
367 | mmdrop(mm); | |
368 | } | |
369 | EXPORT_SYMBOL_GPL(mmu_notifier_unregister); | |
21a92735 | 370 | |
b972216e PZ |
371 | /* |
372 | * Same as mmu_notifier_unregister but no callback and no srcu synchronization. | |
373 | */ | |
374 | void mmu_notifier_unregister_no_release(struct mmu_notifier *mn, | |
375 | struct mm_struct *mm) | |
376 | { | |
377 | spin_lock(&mm->mmu_notifier_mm->lock); | |
378 | /* | |
379 | * Can not use list_del_rcu() since __mmu_notifier_release | |
380 | * can delete it before we hold the lock. | |
381 | */ | |
382 | hlist_del_init_rcu(&mn->hlist); | |
383 | spin_unlock(&mm->mmu_notifier_mm->lock); | |
384 | ||
385 | BUG_ON(atomic_read(&mm->mm_count) <= 0); | |
386 | mmdrop(mm); | |
387 | } | |
388 | EXPORT_SYMBOL_GPL(mmu_notifier_unregister_no_release); |