xref: /linux-6.15/include/linux/mmu_notifier.h (revision bcefe12e)
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>
7 
8 struct mmu_notifier;
9 struct mmu_notifier_ops;
10 
11 #ifdef CONFIG_MMU_NOTIFIER
12 
13 /*
14  * The mmu notifier_mm structure is allocated and installed in
15  * mm->mmu_notifier_mm inside the mm_take_all_locks() protected
16  * critical section and it's released only when mm_count reaches zero
17  * in mmdrop().
18  */
19 struct mmu_notifier_mm {
20 	/* all mmu notifiers registerd in this mm are queued in this list */
21 	struct hlist_head list;
22 	/* to serialize the list modifications and hlist_unhashed */
23 	spinlock_t lock;
24 };
25 
26 struct mmu_notifier_ops {
27 	/*
28 	 * Called either by mmu_notifier_unregister or when the mm is
29 	 * being destroyed by exit_mmap, always before all pages are
30 	 * freed. This can run concurrently with other mmu notifier
31 	 * methods (the ones invoked outside the mm context) and it
32 	 * should tear down all secondary mmu mappings and freeze the
33 	 * secondary mmu. If this method isn't implemented you've to
34 	 * be sure that nothing could possibly write to the pages
35 	 * through the secondary mmu by the time the last thread with
36 	 * tsk->mm == mm exits.
37 	 *
38 	 * As side note: the pages freed after ->release returns could
39 	 * be immediately reallocated by the gart at an alias physical
40 	 * address with a different cache model, so if ->release isn't
41 	 * implemented because all _software_ driven memory accesses
42 	 * through the secondary mmu are terminated by the time the
43 	 * last thread of this mm quits, you've also to be sure that
44 	 * speculative _hardware_ operations can't allocate dirty
45 	 * cachelines in the cpu that could not be snooped and made
46 	 * coherent with the other read and write operations happening
47 	 * through the gart alias address, so leading to memory
48 	 * corruption.
49 	 */
50 	void (*release)(struct mmu_notifier *mn,
51 			struct mm_struct *mm);
52 
53 	/*
54 	 * clear_flush_young is called after the VM is
55 	 * test-and-clearing the young/accessed bitflag in the
56 	 * pte. This way the VM will provide proper aging to the
57 	 * accesses to the page through the secondary MMUs and not
58 	 * only to the ones through the Linux pte.
59 	 */
60 	int (*clear_flush_young)(struct mmu_notifier *mn,
61 				 struct mm_struct *mm,
62 				 unsigned long address);
63 
64 	/*
65 	 * change_pte is called in cases that pte mapping to page is changed:
66 	 * for example, when ksm remaps pte to point to a new shared page.
67 	 */
68 	void (*change_pte)(struct mmu_notifier *mn,
69 			   struct mm_struct *mm,
70 			   unsigned long address,
71 			   pte_t pte);
72 
73 	/*
74 	 * Before this is invoked any secondary MMU is still ok to
75 	 * read/write to the page previously pointed to by the Linux
76 	 * pte because the page hasn't been freed yet and it won't be
77 	 * freed until this returns. If required set_page_dirty has to
78 	 * be called internally to this method.
79 	 */
80 	void (*invalidate_page)(struct mmu_notifier *mn,
81 				struct mm_struct *mm,
82 				unsigned long address);
83 
84 	/*
85 	 * invalidate_range_start() and invalidate_range_end() must be
86 	 * paired and are called only when the mmap_sem and/or the
87 	 * locks protecting the reverse maps are held. The subsystem
88 	 * must guarantee that no additional references are taken to
89 	 * the pages in the range established between the call to
90 	 * invalidate_range_start() and the matching call to
91 	 * invalidate_range_end().
92 	 *
93 	 * Invalidation of multiple concurrent ranges may be
94 	 * optionally permitted by the driver. Either way the
95 	 * establishment of sptes is forbidden in the range passed to
96 	 * invalidate_range_begin/end for the whole duration of the
97 	 * invalidate_range_begin/end critical section.
98 	 *
99 	 * invalidate_range_start() is called when all pages in the
100 	 * range are still mapped and have at least a refcount of one.
101 	 *
102 	 * invalidate_range_end() is called when all pages in the
103 	 * range have been unmapped and the pages have been freed by
104 	 * the VM.
105 	 *
106 	 * The VM will remove the page table entries and potentially
107 	 * the page between invalidate_range_start() and
108 	 * invalidate_range_end(). If the page must not be freed
109 	 * because of pending I/O or other circumstances then the
110 	 * invalidate_range_start() callback (or the initial mapping
111 	 * by the driver) must make sure that the refcount is kept
112 	 * elevated.
113 	 *
114 	 * If the driver increases the refcount when the pages are
115 	 * initially mapped into an address space then either
116 	 * invalidate_range_start() or invalidate_range_end() may
117 	 * decrease the refcount. If the refcount is decreased on
118 	 * invalidate_range_start() then the VM can free pages as page
119 	 * table entries are removed.  If the refcount is only
120 	 * droppped on invalidate_range_end() then the driver itself
121 	 * will drop the last refcount but it must take care to flush
122 	 * any secondary tlb before doing the final free on the
123 	 * page. Pages will no longer be referenced by the linux
124 	 * address space but may still be referenced by sptes until
125 	 * the last refcount is dropped.
126 	 */
127 	void (*invalidate_range_start)(struct mmu_notifier *mn,
128 				       struct mm_struct *mm,
129 				       unsigned long start, unsigned long end);
130 	void (*invalidate_range_end)(struct mmu_notifier *mn,
131 				     struct mm_struct *mm,
132 				     unsigned long start, unsigned long end);
133 };
134 
135 /*
136  * The notifier chains are protected by mmap_sem and/or the reverse map
137  * semaphores. Notifier chains are only changed when all reverse maps and
138  * the mmap_sem locks are taken.
139  *
140  * Therefore notifier chains can only be traversed when either
141  *
142  * 1. mmap_sem is held.
143  * 2. One of the reverse map locks is held (i_mmap_lock or anon_vma->lock).
144  * 3. No other concurrent thread can access the list (release)
145  */
146 struct mmu_notifier {
147 	struct hlist_node hlist;
148 	const struct mmu_notifier_ops *ops;
149 };
150 
151 static inline int mm_has_notifiers(struct mm_struct *mm)
152 {
153 	return unlikely(mm->mmu_notifier_mm);
154 }
155 
156 extern int mmu_notifier_register(struct mmu_notifier *mn,
157 				 struct mm_struct *mm);
158 extern int __mmu_notifier_register(struct mmu_notifier *mn,
159 				   struct mm_struct *mm);
160 extern void mmu_notifier_unregister(struct mmu_notifier *mn,
161 				    struct mm_struct *mm);
162 extern void __mmu_notifier_mm_destroy(struct mm_struct *mm);
163 extern void __mmu_notifier_release(struct mm_struct *mm);
164 extern int __mmu_notifier_clear_flush_young(struct mm_struct *mm,
165 					  unsigned long address);
166 extern void __mmu_notifier_change_pte(struct mm_struct *mm,
167 				      unsigned long address, pte_t pte);
168 extern void __mmu_notifier_invalidate_page(struct mm_struct *mm,
169 					  unsigned long address);
170 extern void __mmu_notifier_invalidate_range_start(struct mm_struct *mm,
171 				  unsigned long start, unsigned long end);
172 extern void __mmu_notifier_invalidate_range_end(struct mm_struct *mm,
173 				  unsigned long start, unsigned long end);
174 
175 static inline void mmu_notifier_release(struct mm_struct *mm)
176 {
177 	if (mm_has_notifiers(mm))
178 		__mmu_notifier_release(mm);
179 }
180 
181 static inline int mmu_notifier_clear_flush_young(struct mm_struct *mm,
182 					  unsigned long address)
183 {
184 	if (mm_has_notifiers(mm))
185 		return __mmu_notifier_clear_flush_young(mm, address);
186 	return 0;
187 }
188 
189 static inline void mmu_notifier_change_pte(struct mm_struct *mm,
190 					   unsigned long address, pte_t pte)
191 {
192 	if (mm_has_notifiers(mm))
193 		__mmu_notifier_change_pte(mm, address, pte);
194 }
195 
196 static inline void mmu_notifier_invalidate_page(struct mm_struct *mm,
197 					  unsigned long address)
198 {
199 	if (mm_has_notifiers(mm))
200 		__mmu_notifier_invalidate_page(mm, address);
201 }
202 
203 static inline void mmu_notifier_invalidate_range_start(struct mm_struct *mm,
204 				  unsigned long start, unsigned long end)
205 {
206 	if (mm_has_notifiers(mm))
207 		__mmu_notifier_invalidate_range_start(mm, start, end);
208 }
209 
210 static inline void mmu_notifier_invalidate_range_end(struct mm_struct *mm,
211 				  unsigned long start, unsigned long end)
212 {
213 	if (mm_has_notifiers(mm))
214 		__mmu_notifier_invalidate_range_end(mm, start, end);
215 }
216 
217 static inline void mmu_notifier_mm_init(struct mm_struct *mm)
218 {
219 	mm->mmu_notifier_mm = NULL;
220 }
221 
222 static inline void mmu_notifier_mm_destroy(struct mm_struct *mm)
223 {
224 	if (mm_has_notifiers(mm))
225 		__mmu_notifier_mm_destroy(mm);
226 }
227 
228 /*
229  * These two macros will sometime replace ptep_clear_flush.
230  * ptep_clear_flush is impleemnted as macro itself, so this also is
231  * implemented as a macro until ptep_clear_flush will converted to an
232  * inline function, to diminish the risk of compilation failure. The
233  * invalidate_page method over time can be moved outside the PT lock
234  * and these two macros can be later removed.
235  */
236 #define ptep_clear_flush_notify(__vma, __address, __ptep)		\
237 ({									\
238 	pte_t __pte;							\
239 	struct vm_area_struct *___vma = __vma;				\
240 	unsigned long ___address = __address;				\
241 	__pte = ptep_clear_flush(___vma, ___address, __ptep);		\
242 	mmu_notifier_invalidate_page(___vma->vm_mm, ___address);	\
243 	__pte;								\
244 })
245 
246 #define ptep_clear_flush_young_notify(__vma, __address, __ptep)		\
247 ({									\
248 	int __young;							\
249 	struct vm_area_struct *___vma = __vma;				\
250 	unsigned long ___address = __address;				\
251 	__young = ptep_clear_flush_young(___vma, ___address, __ptep);	\
252 	__young |= mmu_notifier_clear_flush_young(___vma->vm_mm,	\
253 						  ___address);		\
254 	__young;							\
255 })
256 
257 #define set_pte_at_notify(__mm, __address, __ptep, __pte)		\
258 ({									\
259 	struct mm_struct *___mm = __mm;					\
260 	unsigned long ___address = __address;				\
261 	pte_t ___pte = __pte;						\
262 									\
263 	set_pte_at(___mm, ___address, __ptep, ___pte);			\
264 	mmu_notifier_change_pte(___mm, ___address, ___pte);		\
265 })
266 
267 #else /* CONFIG_MMU_NOTIFIER */
268 
269 static inline void mmu_notifier_release(struct mm_struct *mm)
270 {
271 }
272 
273 static inline int mmu_notifier_clear_flush_young(struct mm_struct *mm,
274 					  unsigned long address)
275 {
276 	return 0;
277 }
278 
279 static inline void mmu_notifier_change_pte(struct mm_struct *mm,
280 					   unsigned long address, pte_t pte)
281 {
282 }
283 
284 static inline void mmu_notifier_invalidate_page(struct mm_struct *mm,
285 					  unsigned long address)
286 {
287 }
288 
289 static inline void mmu_notifier_invalidate_range_start(struct mm_struct *mm,
290 				  unsigned long start, unsigned long end)
291 {
292 }
293 
294 static inline void mmu_notifier_invalidate_range_end(struct mm_struct *mm,
295 				  unsigned long start, unsigned long end)
296 {
297 }
298 
299 static inline void mmu_notifier_mm_init(struct mm_struct *mm)
300 {
301 }
302 
303 static inline void mmu_notifier_mm_destroy(struct mm_struct *mm)
304 {
305 }
306 
307 #define ptep_clear_flush_young_notify ptep_clear_flush_young
308 #define ptep_clear_flush_notify ptep_clear_flush
309 #define set_pte_at_notify set_pte_at
310 
311 #endif /* CONFIG_MMU_NOTIFIER */
312 
313 #endif /* _LINUX_MMU_NOTIFIER_H */
314