]>
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 | ||
cddb8a5c AA |
38 | /* |
39 | * This function can't run concurrently against mmu_notifier_register | |
40 | * because mm->mm_users > 0 during mmu_notifier_register and exit_mmap | |
41 | * runs with mm_users == 0. Other tasks may still invoke mmu notifiers | |
42 | * in parallel despite there being no task using this mm any more, | |
43 | * through the vmas outside of the exit_mmap context, such as with | |
44 | * vmtruncate. This serializes against mmu_notifier_unregister with | |
21a92735 SG |
45 | * the mmu_notifier_mm->lock in addition to SRCU and it serializes |
46 | * against the other mmu notifiers with SRCU. struct mmu_notifier_mm | |
cddb8a5c AA |
47 | * can't go away from under us as exit_mmap holds an mm_count pin |
48 | * itself. | |
49 | */ | |
50 | void __mmu_notifier_release(struct mm_struct *mm) | |
51 | { | |
52 | struct mmu_notifier *mn; | |
21a92735 | 53 | int id; |
3ad3d901 XG |
54 | |
55 | /* | |
d34883d4 XG |
56 | * SRCU here will block mmu_notifier_unregister until |
57 | * ->release returns. | |
3ad3d901 | 58 | */ |
21a92735 | 59 | id = srcu_read_lock(&srcu); |
d34883d4 XG |
60 | hlist_for_each_entry_rcu(mn, &mm->mmu_notifier_mm->list, hlist) |
61 | /* | |
62 | * If ->release runs before mmu_notifier_unregister it must be | |
63 | * handled, as it's the only way for the driver to flush all | |
64 | * existing sptes and stop the driver from establishing any more | |
65 | * sptes before all the pages in the mm are freed. | |
66 | */ | |
67 | if (mn->ops->release) | |
68 | mn->ops->release(mn, mm); | |
d34883d4 | 69 | |
cddb8a5c AA |
70 | spin_lock(&mm->mmu_notifier_mm->lock); |
71 | while (unlikely(!hlist_empty(&mm->mmu_notifier_mm->list))) { | |
72 | mn = hlist_entry(mm->mmu_notifier_mm->list.first, | |
73 | struct mmu_notifier, | |
74 | hlist); | |
75 | /* | |
d34883d4 XG |
76 | * We arrived before mmu_notifier_unregister so |
77 | * mmu_notifier_unregister will do nothing other than to wait | |
78 | * for ->release to finish and for mmu_notifier_unregister to | |
79 | * return. | |
cddb8a5c AA |
80 | */ |
81 | hlist_del_init_rcu(&mn->hlist); | |
cddb8a5c AA |
82 | } |
83 | spin_unlock(&mm->mmu_notifier_mm->lock); | |
b972216e | 84 | srcu_read_unlock(&srcu, id); |
cddb8a5c AA |
85 | |
86 | /* | |
d34883d4 XG |
87 | * synchronize_srcu here prevents mmu_notifier_release from returning to |
88 | * exit_mmap (which would proceed with freeing all pages in the mm) | |
89 | * until the ->release method returns, if it was invoked by | |
90 | * mmu_notifier_unregister. | |
91 | * | |
92 | * The mmu_notifier_mm can't go away from under us because one mm_count | |
93 | * is held by exit_mmap. | |
cddb8a5c | 94 | */ |
21a92735 | 95 | synchronize_srcu(&srcu); |
cddb8a5c AA |
96 | } |
97 | ||
98 | /* | |
99 | * If no young bitflag is supported by the hardware, ->clear_flush_young can | |
100 | * unmap the address and return 1 or 0 depending if the mapping previously | |
101 | * existed or not. | |
102 | */ | |
103 | int __mmu_notifier_clear_flush_young(struct mm_struct *mm, | |
57128468 ALC |
104 | unsigned long start, |
105 | unsigned long end) | |
cddb8a5c AA |
106 | { |
107 | struct mmu_notifier *mn; | |
21a92735 | 108 | int young = 0, id; |
cddb8a5c | 109 | |
21a92735 | 110 | id = srcu_read_lock(&srcu); |
b67bfe0d | 111 | hlist_for_each_entry_rcu(mn, &mm->mmu_notifier_mm->list, hlist) { |
cddb8a5c | 112 | if (mn->ops->clear_flush_young) |
57128468 | 113 | young |= mn->ops->clear_flush_young(mn, mm, start, end); |
cddb8a5c | 114 | } |
21a92735 | 115 | srcu_read_unlock(&srcu, id); |
cddb8a5c AA |
116 | |
117 | return young; | |
118 | } | |
119 | ||
1d7715c6 VD |
120 | int __mmu_notifier_clear_young(struct mm_struct *mm, |
121 | unsigned long start, | |
122 | unsigned long end) | |
123 | { | |
124 | struct mmu_notifier *mn; | |
125 | int young = 0, id; | |
126 | ||
127 | id = srcu_read_lock(&srcu); | |
128 | hlist_for_each_entry_rcu(mn, &mm->mmu_notifier_mm->list, hlist) { | |
129 | if (mn->ops->clear_young) | |
130 | young |= mn->ops->clear_young(mn, mm, start, end); | |
131 | } | |
132 | srcu_read_unlock(&srcu, id); | |
133 | ||
134 | return young; | |
135 | } | |
136 | ||
8ee53820 AA |
137 | int __mmu_notifier_test_young(struct mm_struct *mm, |
138 | unsigned long address) | |
139 | { | |
140 | struct mmu_notifier *mn; | |
21a92735 | 141 | int young = 0, id; |
8ee53820 | 142 | |
21a92735 | 143 | id = srcu_read_lock(&srcu); |
b67bfe0d | 144 | hlist_for_each_entry_rcu(mn, &mm->mmu_notifier_mm->list, hlist) { |
8ee53820 AA |
145 | if (mn->ops->test_young) { |
146 | young = mn->ops->test_young(mn, mm, address); | |
147 | if (young) | |
148 | break; | |
149 | } | |
150 | } | |
21a92735 | 151 | srcu_read_unlock(&srcu, id); |
8ee53820 AA |
152 | |
153 | return young; | |
154 | } | |
155 | ||
828502d3 IE |
156 | void __mmu_notifier_change_pte(struct mm_struct *mm, unsigned long address, |
157 | pte_t pte) | |
158 | { | |
159 | struct mmu_notifier *mn; | |
21a92735 | 160 | int id; |
828502d3 | 161 | |
21a92735 | 162 | id = srcu_read_lock(&srcu); |
b67bfe0d | 163 | hlist_for_each_entry_rcu(mn, &mm->mmu_notifier_mm->list, hlist) { |
828502d3 IE |
164 | if (mn->ops->change_pte) |
165 | mn->ops->change_pte(mn, mm, address, pte); | |
828502d3 | 166 | } |
21a92735 | 167 | srcu_read_unlock(&srcu, id); |
828502d3 IE |
168 | } |
169 | ||
93065ac7 MH |
170 | int __mmu_notifier_invalidate_range_start(struct mm_struct *mm, |
171 | unsigned long start, unsigned long end, | |
172 | bool blockable) | |
cddb8a5c | 173 | { |
5d6527a7 | 174 | struct mmu_notifier_range _range, *range = &_range; |
cddb8a5c | 175 | struct mmu_notifier *mn; |
93065ac7 | 176 | int ret = 0; |
21a92735 | 177 | int id; |
cddb8a5c | 178 | |
5d6527a7 JG |
179 | range->blockable = blockable; |
180 | range->start = start; | |
181 | range->end = end; | |
182 | range->mm = mm; | |
183 | ||
21a92735 | 184 | id = srcu_read_lock(&srcu); |
b67bfe0d | 185 | hlist_for_each_entry_rcu(mn, &mm->mmu_notifier_mm->list, hlist) { |
93065ac7 | 186 | if (mn->ops->invalidate_range_start) { |
5d6527a7 | 187 | int _ret = mn->ops->invalidate_range_start(mn, range); |
93065ac7 MH |
188 | if (_ret) { |
189 | pr_info("%pS callback failed with %d in %sblockable context.\n", | |
190 | mn->ops->invalidate_range_start, _ret, | |
191 | !blockable ? "non-" : ""); | |
192 | ret = _ret; | |
193 | } | |
194 | } | |
cddb8a5c | 195 | } |
21a92735 | 196 | srcu_read_unlock(&srcu, id); |
93065ac7 MH |
197 | |
198 | return ret; | |
cddb8a5c | 199 | } |
fa794199 | 200 | EXPORT_SYMBOL_GPL(__mmu_notifier_invalidate_range_start); |
cddb8a5c AA |
201 | |
202 | void __mmu_notifier_invalidate_range_end(struct mm_struct *mm, | |
4645b9fe JG |
203 | unsigned long start, |
204 | unsigned long end, | |
205 | bool only_end) | |
cddb8a5c | 206 | { |
5d6527a7 | 207 | struct mmu_notifier_range _range, *range = &_range; |
cddb8a5c | 208 | struct mmu_notifier *mn; |
21a92735 | 209 | int id; |
cddb8a5c | 210 | |
5d6527a7 JG |
211 | /* |
212 | * The end call back will never be call if the start refused to go | |
213 | * through because of blockable was false so here assume that we | |
214 | * can block. | |
215 | */ | |
216 | range->blockable = true; | |
217 | range->start = start; | |
218 | range->end = end; | |
219 | range->mm = mm; | |
220 | ||
21a92735 | 221 | id = srcu_read_lock(&srcu); |
b67bfe0d | 222 | hlist_for_each_entry_rcu(mn, &mm->mmu_notifier_mm->list, hlist) { |
0f0a327f JR |
223 | /* |
224 | * Call invalidate_range here too to avoid the need for the | |
225 | * subsystem of having to register an invalidate_range_end | |
226 | * call-back when there is invalidate_range already. Usually a | |
227 | * subsystem registers either invalidate_range_start()/end() or | |
228 | * invalidate_range(), so this will be no additional overhead | |
229 | * (besides the pointer check). | |
4645b9fe JG |
230 | * |
231 | * We skip call to invalidate_range() if we know it is safe ie | |
232 | * call site use mmu_notifier_invalidate_range_only_end() which | |
233 | * is safe to do when we know that a call to invalidate_range() | |
234 | * already happen under page table lock. | |
0f0a327f | 235 | */ |
4645b9fe | 236 | if (!only_end && mn->ops->invalidate_range) |
0f0a327f | 237 | mn->ops->invalidate_range(mn, mm, start, end); |
cddb8a5c | 238 | if (mn->ops->invalidate_range_end) |
5d6527a7 | 239 | mn->ops->invalidate_range_end(mn, range); |
cddb8a5c | 240 | } |
21a92735 | 241 | srcu_read_unlock(&srcu, id); |
cddb8a5c | 242 | } |
fa794199 | 243 | EXPORT_SYMBOL_GPL(__mmu_notifier_invalidate_range_end); |
cddb8a5c | 244 | |
0f0a327f JR |
245 | void __mmu_notifier_invalidate_range(struct mm_struct *mm, |
246 | unsigned long start, unsigned long end) | |
247 | { | |
248 | struct mmu_notifier *mn; | |
249 | int id; | |
250 | ||
251 | id = srcu_read_lock(&srcu); | |
252 | hlist_for_each_entry_rcu(mn, &mm->mmu_notifier_mm->list, hlist) { | |
253 | if (mn->ops->invalidate_range) | |
254 | mn->ops->invalidate_range(mn, mm, start, end); | |
255 | } | |
256 | srcu_read_unlock(&srcu, id); | |
257 | } | |
258 | EXPORT_SYMBOL_GPL(__mmu_notifier_invalidate_range); | |
259 | ||
cddb8a5c AA |
260 | static int do_mmu_notifier_register(struct mmu_notifier *mn, |
261 | struct mm_struct *mm, | |
262 | int take_mmap_sem) | |
263 | { | |
264 | struct mmu_notifier_mm *mmu_notifier_mm; | |
265 | int ret; | |
266 | ||
267 | BUG_ON(atomic_read(&mm->mm_users) <= 0); | |
268 | ||
35cfa2b0 GS |
269 | ret = -ENOMEM; |
270 | mmu_notifier_mm = kmalloc(sizeof(struct mmu_notifier_mm), GFP_KERNEL); | |
271 | if (unlikely(!mmu_notifier_mm)) | |
272 | goto out; | |
273 | ||
cddb8a5c AA |
274 | if (take_mmap_sem) |
275 | down_write(&mm->mmap_sem); | |
276 | ret = mm_take_all_locks(mm); | |
277 | if (unlikely(ret)) | |
35cfa2b0 | 278 | goto out_clean; |
cddb8a5c AA |
279 | |
280 | if (!mm_has_notifiers(mm)) { | |
281 | INIT_HLIST_HEAD(&mmu_notifier_mm->list); | |
282 | spin_lock_init(&mmu_notifier_mm->lock); | |
e0f3c3f7 | 283 | |
cddb8a5c | 284 | mm->mmu_notifier_mm = mmu_notifier_mm; |
35cfa2b0 | 285 | mmu_notifier_mm = NULL; |
cddb8a5c | 286 | } |
f1f10076 | 287 | mmgrab(mm); |
cddb8a5c AA |
288 | |
289 | /* | |
290 | * Serialize the update against mmu_notifier_unregister. A | |
291 | * side note: mmu_notifier_release can't run concurrently with | |
292 | * us because we hold the mm_users pin (either implicitly as | |
293 | * current->mm or explicitly with get_task_mm() or similar). | |
294 | * We can't race against any other mmu notifier method either | |
295 | * thanks to mm_take_all_locks(). | |
296 | */ | |
297 | spin_lock(&mm->mmu_notifier_mm->lock); | |
298 | hlist_add_head(&mn->hlist, &mm->mmu_notifier_mm->list); | |
299 | spin_unlock(&mm->mmu_notifier_mm->lock); | |
300 | ||
301 | mm_drop_all_locks(mm); | |
35cfa2b0 | 302 | out_clean: |
cddb8a5c AA |
303 | if (take_mmap_sem) |
304 | up_write(&mm->mmap_sem); | |
35cfa2b0 GS |
305 | kfree(mmu_notifier_mm); |
306 | out: | |
cddb8a5c AA |
307 | BUG_ON(atomic_read(&mm->mm_users) <= 0); |
308 | return ret; | |
309 | } | |
310 | ||
311 | /* | |
312 | * Must not hold mmap_sem nor any other VM related lock when calling | |
313 | * this registration function. Must also ensure mm_users can't go down | |
314 | * to zero while this runs to avoid races with mmu_notifier_release, | |
315 | * so mm has to be current->mm or the mm should be pinned safely such | |
316 | * as with get_task_mm(). If the mm is not current->mm, the mm_users | |
317 | * pin should be released by calling mmput after mmu_notifier_register | |
318 | * returns. mmu_notifier_unregister must be always called to | |
319 | * unregister the notifier. mm_count is automatically pinned to allow | |
320 | * mmu_notifier_unregister to safely run at any time later, before or | |
321 | * after exit_mmap. ->release will always be called before exit_mmap | |
322 | * frees the pages. | |
323 | */ | |
324 | int mmu_notifier_register(struct mmu_notifier *mn, struct mm_struct *mm) | |
325 | { | |
326 | return do_mmu_notifier_register(mn, mm, 1); | |
327 | } | |
328 | EXPORT_SYMBOL_GPL(mmu_notifier_register); | |
329 | ||
330 | /* | |
331 | * Same as mmu_notifier_register but here the caller must hold the | |
332 | * mmap_sem in write mode. | |
333 | */ | |
334 | int __mmu_notifier_register(struct mmu_notifier *mn, struct mm_struct *mm) | |
335 | { | |
336 | return do_mmu_notifier_register(mn, mm, 0); | |
337 | } | |
338 | EXPORT_SYMBOL_GPL(__mmu_notifier_register); | |
339 | ||
340 | /* this is called after the last mmu_notifier_unregister() returned */ | |
341 | void __mmu_notifier_mm_destroy(struct mm_struct *mm) | |
342 | { | |
343 | BUG_ON(!hlist_empty(&mm->mmu_notifier_mm->list)); | |
344 | kfree(mm->mmu_notifier_mm); | |
345 | mm->mmu_notifier_mm = LIST_POISON1; /* debug */ | |
346 | } | |
347 | ||
348 | /* | |
349 | * This releases the mm_count pin automatically and frees the mm | |
350 | * structure if it was the last user of it. It serializes against | |
21a92735 SG |
351 | * running mmu notifiers with SRCU and against mmu_notifier_unregister |
352 | * with the unregister lock + SRCU. All sptes must be dropped before | |
cddb8a5c AA |
353 | * calling mmu_notifier_unregister. ->release or any other notifier |
354 | * method may be invoked concurrently with mmu_notifier_unregister, | |
355 | * and only after mmu_notifier_unregister returned we're guaranteed | |
356 | * that ->release or any other method can't run anymore. | |
357 | */ | |
358 | void mmu_notifier_unregister(struct mmu_notifier *mn, struct mm_struct *mm) | |
359 | { | |
360 | BUG_ON(atomic_read(&mm->mm_count) <= 0); | |
361 | ||
cddb8a5c | 362 | if (!hlist_unhashed(&mn->hlist)) { |
d34883d4 XG |
363 | /* |
364 | * SRCU here will force exit_mmap to wait for ->release to | |
365 | * finish before freeing the pages. | |
366 | */ | |
21a92735 | 367 | int id; |
3ad3d901 | 368 | |
d34883d4 | 369 | id = srcu_read_lock(&srcu); |
cddb8a5c | 370 | /* |
d34883d4 XG |
371 | * exit_mmap will block in mmu_notifier_release to guarantee |
372 | * that ->release is called before freeing the pages. | |
cddb8a5c AA |
373 | */ |
374 | if (mn->ops->release) | |
375 | mn->ops->release(mn, mm); | |
d34883d4 | 376 | srcu_read_unlock(&srcu, id); |
3ad3d901 | 377 | |
d34883d4 | 378 | spin_lock(&mm->mmu_notifier_mm->lock); |
751efd86 | 379 | /* |
d34883d4 XG |
380 | * Can not use list_del_rcu() since __mmu_notifier_release |
381 | * can delete it before we hold the lock. | |
751efd86 | 382 | */ |
d34883d4 | 383 | hlist_del_init_rcu(&mn->hlist); |
cddb8a5c | 384 | spin_unlock(&mm->mmu_notifier_mm->lock); |
d34883d4 | 385 | } |
cddb8a5c AA |
386 | |
387 | /* | |
d34883d4 | 388 | * Wait for any running method to finish, of course including |
83a35e36 | 389 | * ->release if it was run by mmu_notifier_release instead of us. |
cddb8a5c | 390 | */ |
21a92735 | 391 | synchronize_srcu(&srcu); |
cddb8a5c AA |
392 | |
393 | BUG_ON(atomic_read(&mm->mm_count) <= 0); | |
394 | ||
395 | mmdrop(mm); | |
396 | } | |
397 | EXPORT_SYMBOL_GPL(mmu_notifier_unregister); | |
21a92735 | 398 | |
b972216e PZ |
399 | /* |
400 | * Same as mmu_notifier_unregister but no callback and no srcu synchronization. | |
401 | */ | |
402 | void mmu_notifier_unregister_no_release(struct mmu_notifier *mn, | |
403 | struct mm_struct *mm) | |
404 | { | |
405 | spin_lock(&mm->mmu_notifier_mm->lock); | |
406 | /* | |
407 | * Can not use list_del_rcu() since __mmu_notifier_release | |
408 | * can delete it before we hold the lock. | |
409 | */ | |
410 | hlist_del_init_rcu(&mn->hlist); | |
411 | spin_unlock(&mm->mmu_notifier_mm->lock); | |
412 | ||
413 | BUG_ON(atomic_read(&mm->mm_count) <= 0); | |
414 | mmdrop(mm); | |
415 | } | |
416 | EXPORT_SYMBOL_GPL(mmu_notifier_unregister_no_release); |