Commit | Line | Data |
---|---|---|
cddb8a5c AA |
1 | #ifndef _LINUX_MMU_NOTIFIER_H |
2 | #define _LINUX_MMU_NOTIFIER_H | |
3 | ||
4 | #include <linux/list.h> | |
5 | #include <linux/spinlock.h> | |
6 | #include <linux/mm_types.h> | |
21a92735 | 7 | #include <linux/srcu.h> |
cddb8a5c AA |
8 | |
9 | struct mmu_notifier; | |
10 | struct mmu_notifier_ops; | |
11 | ||
12 | #ifdef CONFIG_MMU_NOTIFIER | |
13 | ||
14 | /* | |
15 | * The mmu notifier_mm structure is allocated and installed in | |
16 | * mm->mmu_notifier_mm inside the mm_take_all_locks() protected | |
17 | * critical section and it's released only when mm_count reaches zero | |
18 | * in mmdrop(). | |
19 | */ | |
20 | struct mmu_notifier_mm { | |
21 | /* all mmu notifiers registerd in this mm are queued in this list */ | |
22 | struct hlist_head list; | |
23 | /* to serialize the list modifications and hlist_unhashed */ | |
24 | spinlock_t lock; | |
25 | }; | |
26 | ||
27 | struct mmu_notifier_ops { | |
28 | /* | |
29 | * Called either by mmu_notifier_unregister or when the mm is | |
30 | * being destroyed by exit_mmap, always before all pages are | |
31 | * freed. This can run concurrently with other mmu notifier | |
32 | * methods (the ones invoked outside the mm context) and it | |
33 | * should tear down all secondary mmu mappings and freeze the | |
34 | * secondary mmu. If this method isn't implemented you've to | |
35 | * be sure that nothing could possibly write to the pages | |
36 | * through the secondary mmu by the time the last thread with | |
37 | * tsk->mm == mm exits. | |
38 | * | |
39 | * As side note: the pages freed after ->release returns could | |
40 | * be immediately reallocated by the gart at an alias physical | |
41 | * address with a different cache model, so if ->release isn't | |
42 | * implemented because all _software_ driven memory accesses | |
43 | * through the secondary mmu are terminated by the time the | |
44 | * last thread of this mm quits, you've also to be sure that | |
45 | * speculative _hardware_ operations can't allocate dirty | |
46 | * cachelines in the cpu that could not be snooped and made | |
47 | * coherent with the other read and write operations happening | |
48 | * through the gart alias address, so leading to memory | |
49 | * corruption. | |
50 | */ | |
51 | void (*release)(struct mmu_notifier *mn, | |
52 | struct mm_struct *mm); | |
53 | ||
54 | /* | |
55 | * clear_flush_young is called after the VM is | |
56 | * test-and-clearing the young/accessed bitflag in the | |
57 | * pte. This way the VM will provide proper aging to the | |
58 | * accesses to the page through the secondary MMUs and not | |
59 | * only to the ones through the Linux pte. | |
57128468 ALC |
60 | * Start-end is necessary in case the secondary MMU is mapping the page |
61 | * at a smaller granularity than the primary MMU. | |
cddb8a5c AA |
62 | */ |
63 | int (*clear_flush_young)(struct mmu_notifier *mn, | |
64 | struct mm_struct *mm, | |
57128468 ALC |
65 | unsigned long start, |
66 | unsigned long end); | |
cddb8a5c | 67 | |
8ee53820 AA |
68 | /* |
69 | * test_young is called to check the young/accessed bitflag in | |
70 | * the secondary pte. This is used to know if the page is | |
71 | * frequently used without actually clearing the flag or tearing | |
72 | * down the secondary mapping on the page. | |
73 | */ | |
74 | int (*test_young)(struct mmu_notifier *mn, | |
75 | struct mm_struct *mm, | |
76 | unsigned long address); | |
77 | ||
828502d3 IE |
78 | /* |
79 | * change_pte is called in cases that pte mapping to page is changed: | |
80 | * for example, when ksm remaps pte to point to a new shared page. | |
81 | */ | |
82 | void (*change_pte)(struct mmu_notifier *mn, | |
83 | struct mm_struct *mm, | |
84 | unsigned long address, | |
85 | pte_t pte); | |
86 | ||
cddb8a5c AA |
87 | /* |
88 | * Before this is invoked any secondary MMU is still ok to | |
89 | * read/write to the page previously pointed to by the Linux | |
90 | * pte because the page hasn't been freed yet and it won't be | |
91 | * freed until this returns. If required set_page_dirty has to | |
92 | * be called internally to this method. | |
93 | */ | |
94 | void (*invalidate_page)(struct mmu_notifier *mn, | |
95 | struct mm_struct *mm, | |
96 | unsigned long address); | |
97 | ||
98 | /* | |
99 | * invalidate_range_start() and invalidate_range_end() must be | |
100 | * paired and are called only when the mmap_sem and/or the | |
0f0a327f JR |
101 | * locks protecting the reverse maps are held. If the subsystem |
102 | * can't guarantee that no additional references are taken to | |
103 | * the pages in the range, it has to implement the | |
104 | * invalidate_range() notifier to remove any references taken | |
105 | * after invalidate_range_start(). | |
cddb8a5c AA |
106 | * |
107 | * Invalidation of multiple concurrent ranges may be | |
108 | * optionally permitted by the driver. Either way the | |
109 | * establishment of sptes is forbidden in the range passed to | |
110 | * invalidate_range_begin/end for the whole duration of the | |
111 | * invalidate_range_begin/end critical section. | |
112 | * | |
113 | * invalidate_range_start() is called when all pages in the | |
114 | * range are still mapped and have at least a refcount of one. | |
115 | * | |
116 | * invalidate_range_end() is called when all pages in the | |
117 | * range have been unmapped and the pages have been freed by | |
118 | * the VM. | |
119 | * | |
120 | * The VM will remove the page table entries and potentially | |
121 | * the page between invalidate_range_start() and | |
122 | * invalidate_range_end(). If the page must not be freed | |
123 | * because of pending I/O or other circumstances then the | |
124 | * invalidate_range_start() callback (or the initial mapping | |
125 | * by the driver) must make sure that the refcount is kept | |
126 | * elevated. | |
127 | * | |
128 | * If the driver increases the refcount when the pages are | |
129 | * initially mapped into an address space then either | |
130 | * invalidate_range_start() or invalidate_range_end() may | |
131 | * decrease the refcount. If the refcount is decreased on | |
132 | * invalidate_range_start() then the VM can free pages as page | |
133 | * table entries are removed. If the refcount is only | |
134 | * droppped on invalidate_range_end() then the driver itself | |
135 | * will drop the last refcount but it must take care to flush | |
136 | * any secondary tlb before doing the final free on the | |
137 | * page. Pages will no longer be referenced by the linux | |
138 | * address space but may still be referenced by sptes until | |
139 | * the last refcount is dropped. | |
140 | */ | |
141 | void (*invalidate_range_start)(struct mmu_notifier *mn, | |
142 | struct mm_struct *mm, | |
143 | unsigned long start, unsigned long end); | |
144 | void (*invalidate_range_end)(struct mmu_notifier *mn, | |
145 | struct mm_struct *mm, | |
146 | unsigned long start, unsigned long end); | |
0f0a327f JR |
147 | |
148 | /* | |
149 | * invalidate_range() is either called between | |
150 | * invalidate_range_start() and invalidate_range_end() when the | |
151 | * VM has to free pages that where unmapped, but before the | |
152 | * pages are actually freed, or outside of _start()/_end() when | |
153 | * a (remote) TLB is necessary. | |
154 | * | |
155 | * If invalidate_range() is used to manage a non-CPU TLB with | |
156 | * shared page-tables, it not necessary to implement the | |
157 | * invalidate_range_start()/end() notifiers, as | |
158 | * invalidate_range() alread catches the points in time when an | |
159 | * external TLB range needs to be flushed. | |
160 | * | |
161 | * The invalidate_range() function is called under the ptl | |
162 | * spin-lock and not allowed to sleep. | |
163 | * | |
164 | * Note that this function might be called with just a sub-range | |
165 | * of what was passed to invalidate_range_start()/end(), if | |
166 | * called between those functions. | |
167 | */ | |
168 | void (*invalidate_range)(struct mmu_notifier *mn, struct mm_struct *mm, | |
169 | unsigned long start, unsigned long end); | |
cddb8a5c AA |
170 | }; |
171 | ||
172 | /* | |
173 | * The notifier chains are protected by mmap_sem and/or the reverse map | |
174 | * semaphores. Notifier chains are only changed when all reverse maps and | |
175 | * the mmap_sem locks are taken. | |
176 | * | |
177 | * Therefore notifier chains can only be traversed when either | |
178 | * | |
179 | * 1. mmap_sem is held. | |
c8c06efa | 180 | * 2. One of the reverse map locks is held (i_mmap_rwsem or anon_vma->rwsem). |
cddb8a5c AA |
181 | * 3. No other concurrent thread can access the list (release) |
182 | */ | |
183 | struct mmu_notifier { | |
184 | struct hlist_node hlist; | |
185 | const struct mmu_notifier_ops *ops; | |
186 | }; | |
187 | ||
188 | static inline int mm_has_notifiers(struct mm_struct *mm) | |
189 | { | |
190 | return unlikely(mm->mmu_notifier_mm); | |
191 | } | |
192 | ||
193 | extern int mmu_notifier_register(struct mmu_notifier *mn, | |
194 | struct mm_struct *mm); | |
195 | extern int __mmu_notifier_register(struct mmu_notifier *mn, | |
196 | struct mm_struct *mm); | |
197 | extern void mmu_notifier_unregister(struct mmu_notifier *mn, | |
198 | struct mm_struct *mm); | |
b972216e PZ |
199 | extern void mmu_notifier_unregister_no_release(struct mmu_notifier *mn, |
200 | struct mm_struct *mm); | |
cddb8a5c AA |
201 | extern void __mmu_notifier_mm_destroy(struct mm_struct *mm); |
202 | extern void __mmu_notifier_release(struct mm_struct *mm); | |
203 | extern int __mmu_notifier_clear_flush_young(struct mm_struct *mm, | |
57128468 ALC |
204 | unsigned long start, |
205 | unsigned long end); | |
8ee53820 AA |
206 | extern int __mmu_notifier_test_young(struct mm_struct *mm, |
207 | unsigned long address); | |
828502d3 IE |
208 | extern void __mmu_notifier_change_pte(struct mm_struct *mm, |
209 | unsigned long address, pte_t pte); | |
cddb8a5c AA |
210 | extern void __mmu_notifier_invalidate_page(struct mm_struct *mm, |
211 | unsigned long address); | |
212 | extern void __mmu_notifier_invalidate_range_start(struct mm_struct *mm, | |
213 | unsigned long start, unsigned long end); | |
214 | extern void __mmu_notifier_invalidate_range_end(struct mm_struct *mm, | |
215 | unsigned long start, unsigned long end); | |
0f0a327f JR |
216 | extern void __mmu_notifier_invalidate_range(struct mm_struct *mm, |
217 | unsigned long start, unsigned long end); | |
cddb8a5c AA |
218 | |
219 | static inline void mmu_notifier_release(struct mm_struct *mm) | |
220 | { | |
221 | if (mm_has_notifiers(mm)) | |
222 | __mmu_notifier_release(mm); | |
223 | } | |
224 | ||
225 | static inline int mmu_notifier_clear_flush_young(struct mm_struct *mm, | |
57128468 ALC |
226 | unsigned long start, |
227 | unsigned long end) | |
cddb8a5c AA |
228 | { |
229 | if (mm_has_notifiers(mm)) | |
57128468 | 230 | return __mmu_notifier_clear_flush_young(mm, start, end); |
cddb8a5c AA |
231 | return 0; |
232 | } | |
233 | ||
8ee53820 AA |
234 | static inline int mmu_notifier_test_young(struct mm_struct *mm, |
235 | unsigned long address) | |
236 | { | |
237 | if (mm_has_notifiers(mm)) | |
238 | return __mmu_notifier_test_young(mm, address); | |
239 | return 0; | |
240 | } | |
241 | ||
828502d3 IE |
242 | static inline void mmu_notifier_change_pte(struct mm_struct *mm, |
243 | unsigned long address, pte_t pte) | |
244 | { | |
245 | if (mm_has_notifiers(mm)) | |
246 | __mmu_notifier_change_pte(mm, address, pte); | |
247 | } | |
248 | ||
cddb8a5c AA |
249 | static inline void mmu_notifier_invalidate_page(struct mm_struct *mm, |
250 | unsigned long address) | |
251 | { | |
252 | if (mm_has_notifiers(mm)) | |
253 | __mmu_notifier_invalidate_page(mm, address); | |
254 | } | |
255 | ||
256 | static inline void mmu_notifier_invalidate_range_start(struct mm_struct *mm, | |
257 | unsigned long start, unsigned long end) | |
258 | { | |
259 | if (mm_has_notifiers(mm)) | |
260 | __mmu_notifier_invalidate_range_start(mm, start, end); | |
261 | } | |
262 | ||
263 | static inline void mmu_notifier_invalidate_range_end(struct mm_struct *mm, | |
264 | unsigned long start, unsigned long end) | |
265 | { | |
266 | if (mm_has_notifiers(mm)) | |
267 | __mmu_notifier_invalidate_range_end(mm, start, end); | |
268 | } | |
269 | ||
1897bdc4 JR |
270 | static inline void mmu_notifier_invalidate_range(struct mm_struct *mm, |
271 | unsigned long start, unsigned long end) | |
272 | { | |
0f0a327f JR |
273 | if (mm_has_notifiers(mm)) |
274 | __mmu_notifier_invalidate_range(mm, start, end); | |
1897bdc4 JR |
275 | } |
276 | ||
cddb8a5c AA |
277 | static inline void mmu_notifier_mm_init(struct mm_struct *mm) |
278 | { | |
279 | mm->mmu_notifier_mm = NULL; | |
280 | } | |
281 | ||
282 | static inline void mmu_notifier_mm_destroy(struct mm_struct *mm) | |
283 | { | |
284 | if (mm_has_notifiers(mm)) | |
285 | __mmu_notifier_mm_destroy(mm); | |
286 | } | |
287 | ||
cddb8a5c AA |
288 | #define ptep_clear_flush_young_notify(__vma, __address, __ptep) \ |
289 | ({ \ | |
290 | int __young; \ | |
291 | struct vm_area_struct *___vma = __vma; \ | |
292 | unsigned long ___address = __address; \ | |
293 | __young = ptep_clear_flush_young(___vma, ___address, __ptep); \ | |
294 | __young |= mmu_notifier_clear_flush_young(___vma->vm_mm, \ | |
57128468 ALC |
295 | ___address, \ |
296 | ___address + \ | |
297 | PAGE_SIZE); \ | |
cddb8a5c AA |
298 | __young; \ |
299 | }) | |
300 | ||
91a4ee26 AA |
301 | #define pmdp_clear_flush_young_notify(__vma, __address, __pmdp) \ |
302 | ({ \ | |
303 | int __young; \ | |
304 | struct vm_area_struct *___vma = __vma; \ | |
305 | unsigned long ___address = __address; \ | |
306 | __young = pmdp_clear_flush_young(___vma, ___address, __pmdp); \ | |
307 | __young |= mmu_notifier_clear_flush_young(___vma->vm_mm, \ | |
57128468 ALC |
308 | ___address, \ |
309 | ___address + \ | |
310 | PMD_SIZE); \ | |
91a4ee26 AA |
311 | __young; \ |
312 | }) | |
313 | ||
34ee645e JR |
314 | #define ptep_clear_flush_notify(__vma, __address, __ptep) \ |
315 | ({ \ | |
316 | unsigned long ___addr = __address & PAGE_MASK; \ | |
317 | struct mm_struct *___mm = (__vma)->vm_mm; \ | |
318 | pte_t ___pte; \ | |
319 | \ | |
320 | ___pte = ptep_clear_flush(__vma, __address, __ptep); \ | |
321 | mmu_notifier_invalidate_range(___mm, ___addr, \ | |
322 | ___addr + PAGE_SIZE); \ | |
323 | \ | |
324 | ___pte; \ | |
325 | }) | |
326 | ||
327 | #define pmdp_clear_flush_notify(__vma, __haddr, __pmd) \ | |
328 | ({ \ | |
329 | unsigned long ___haddr = __haddr & HPAGE_PMD_MASK; \ | |
330 | struct mm_struct *___mm = (__vma)->vm_mm; \ | |
331 | pmd_t ___pmd; \ | |
332 | \ | |
333 | ___pmd = pmdp_clear_flush(__vma, __haddr, __pmd); \ | |
334 | mmu_notifier_invalidate_range(___mm, ___haddr, \ | |
335 | ___haddr + HPAGE_PMD_SIZE); \ | |
336 | \ | |
337 | ___pmd; \ | |
338 | }) | |
339 | ||
340 | #define pmdp_get_and_clear_notify(__mm, __haddr, __pmd) \ | |
341 | ({ \ | |
342 | unsigned long ___haddr = __haddr & HPAGE_PMD_MASK; \ | |
343 | pmd_t ___pmd; \ | |
344 | \ | |
345 | ___pmd = pmdp_get_and_clear(__mm, __haddr, __pmd); \ | |
346 | mmu_notifier_invalidate_range(__mm, ___haddr, \ | |
347 | ___haddr + HPAGE_PMD_SIZE); \ | |
348 | \ | |
349 | ___pmd; \ | |
350 | }) | |
351 | ||
48af0d7c XG |
352 | /* |
353 | * set_pte_at_notify() sets the pte _after_ running the notifier. | |
354 | * This is safe to start by updating the secondary MMUs, because the primary MMU | |
355 | * pte invalidate must have already happened with a ptep_clear_flush() before | |
356 | * set_pte_at_notify() has been invoked. Updating the secondary MMUs first is | |
357 | * required when we change both the protection of the mapping from read-only to | |
358 | * read-write and the pfn (like during copy on write page faults). Otherwise the | |
359 | * old page would remain mapped readonly in the secondary MMUs after the new | |
360 | * page is already writable by some CPU through the primary MMU. | |
361 | */ | |
828502d3 IE |
362 | #define set_pte_at_notify(__mm, __address, __ptep, __pte) \ |
363 | ({ \ | |
364 | struct mm_struct *___mm = __mm; \ | |
365 | unsigned long ___address = __address; \ | |
366 | pte_t ___pte = __pte; \ | |
367 | \ | |
828502d3 | 368 | mmu_notifier_change_pte(___mm, ___address, ___pte); \ |
48af0d7c | 369 | set_pte_at(___mm, ___address, __ptep, ___pte); \ |
828502d3 IE |
370 | }) |
371 | ||
b972216e PZ |
372 | extern void mmu_notifier_call_srcu(struct rcu_head *rcu, |
373 | void (*func)(struct rcu_head *rcu)); | |
374 | extern void mmu_notifier_synchronize(void); | |
375 | ||
cddb8a5c AA |
376 | #else /* CONFIG_MMU_NOTIFIER */ |
377 | ||
378 | static inline void mmu_notifier_release(struct mm_struct *mm) | |
379 | { | |
380 | } | |
381 | ||
382 | static inline int mmu_notifier_clear_flush_young(struct mm_struct *mm, | |
57128468 ALC |
383 | unsigned long start, |
384 | unsigned long end) | |
8ee53820 AA |
385 | { |
386 | return 0; | |
387 | } | |
388 | ||
389 | static inline int mmu_notifier_test_young(struct mm_struct *mm, | |
390 | unsigned long address) | |
cddb8a5c AA |
391 | { |
392 | return 0; | |
393 | } | |
394 | ||
828502d3 IE |
395 | static inline void mmu_notifier_change_pte(struct mm_struct *mm, |
396 | unsigned long address, pte_t pte) | |
397 | { | |
398 | } | |
399 | ||
cddb8a5c AA |
400 | static inline void mmu_notifier_invalidate_page(struct mm_struct *mm, |
401 | unsigned long address) | |
402 | { | |
403 | } | |
404 | ||
405 | static inline void mmu_notifier_invalidate_range_start(struct mm_struct *mm, | |
406 | unsigned long start, unsigned long end) | |
407 | { | |
408 | } | |
409 | ||
410 | static inline void mmu_notifier_invalidate_range_end(struct mm_struct *mm, | |
411 | unsigned long start, unsigned long end) | |
412 | { | |
413 | } | |
414 | ||
1897bdc4 JR |
415 | static inline void mmu_notifier_invalidate_range(struct mm_struct *mm, |
416 | unsigned long start, unsigned long end) | |
417 | { | |
418 | } | |
419 | ||
cddb8a5c AA |
420 | static inline void mmu_notifier_mm_init(struct mm_struct *mm) |
421 | { | |
422 | } | |
423 | ||
424 | static inline void mmu_notifier_mm_destroy(struct mm_struct *mm) | |
425 | { | |
426 | } | |
427 | ||
428 | #define ptep_clear_flush_young_notify ptep_clear_flush_young | |
91a4ee26 | 429 | #define pmdp_clear_flush_young_notify pmdp_clear_flush_young |
34ee645e JR |
430 | #define ptep_clear_flush_notify ptep_clear_flush |
431 | #define pmdp_clear_flush_notify pmdp_clear_flush | |
432 | #define pmdp_get_and_clear_notify pmdp_get_and_clear | |
828502d3 | 433 | #define set_pte_at_notify set_pte_at |
cddb8a5c AA |
434 | |
435 | #endif /* CONFIG_MMU_NOTIFIER */ | |
436 | ||
437 | #endif /* _LINUX_MMU_NOTIFIER_H */ |