xref: /linux-6.15/include/linux/mmu_notifier.h (revision e00a844a)
1 /* SPDX-License-Identifier: GPL-2.0 */
2 #ifndef _LINUX_MMU_NOTIFIER_H
3 #define _LINUX_MMU_NOTIFIER_H
4 
5 #include <linux/list.h>
6 #include <linux/spinlock.h>
7 #include <linux/mm_types.h>
8 #include <linux/srcu.h>
9 
10 struct mmu_notifier;
11 struct mmu_notifier_ops;
12 
13 #ifdef CONFIG_MMU_NOTIFIER
14 
15 /*
16  * The mmu notifier_mm structure is allocated and installed in
17  * mm->mmu_notifier_mm inside the mm_take_all_locks() protected
18  * critical section and it's released only when mm_count reaches zero
19  * in mmdrop().
20  */
21 struct mmu_notifier_mm {
22 	/* all mmu notifiers registerd in this mm are queued in this list */
23 	struct hlist_head list;
24 	/* to serialize the list modifications and hlist_unhashed */
25 	spinlock_t lock;
26 };
27 
28 struct mmu_notifier_ops {
29 	/*
30 	 * Called either by mmu_notifier_unregister or when the mm is
31 	 * being destroyed by exit_mmap, always before all pages are
32 	 * freed. This can run concurrently with other mmu notifier
33 	 * methods (the ones invoked outside the mm context) and it
34 	 * should tear down all secondary mmu mappings and freeze the
35 	 * secondary mmu. If this method isn't implemented you've to
36 	 * be sure that nothing could possibly write to the pages
37 	 * through the secondary mmu by the time the last thread with
38 	 * tsk->mm == mm exits.
39 	 *
40 	 * As side note: the pages freed after ->release returns could
41 	 * be immediately reallocated by the gart at an alias physical
42 	 * address with a different cache model, so if ->release isn't
43 	 * implemented because all _software_ driven memory accesses
44 	 * through the secondary mmu are terminated by the time the
45 	 * last thread of this mm quits, you've also to be sure that
46 	 * speculative _hardware_ operations can't allocate dirty
47 	 * cachelines in the cpu that could not be snooped and made
48 	 * coherent with the other read and write operations happening
49 	 * through the gart alias address, so leading to memory
50 	 * corruption.
51 	 */
52 	void (*release)(struct mmu_notifier *mn,
53 			struct mm_struct *mm);
54 
55 	/*
56 	 * clear_flush_young is called after the VM is
57 	 * test-and-clearing the young/accessed bitflag in the
58 	 * pte. This way the VM will provide proper aging to the
59 	 * accesses to the page through the secondary MMUs and not
60 	 * only to the ones through the Linux pte.
61 	 * Start-end is necessary in case the secondary MMU is mapping the page
62 	 * at a smaller granularity than the primary MMU.
63 	 */
64 	int (*clear_flush_young)(struct mmu_notifier *mn,
65 				 struct mm_struct *mm,
66 				 unsigned long start,
67 				 unsigned long end);
68 
69 	/*
70 	 * clear_young is a lightweight version of clear_flush_young. Like the
71 	 * latter, it is supposed to test-and-clear the young/accessed bitflag
72 	 * in the secondary pte, but it may omit flushing the secondary tlb.
73 	 */
74 	int (*clear_young)(struct mmu_notifier *mn,
75 			   struct mm_struct *mm,
76 			   unsigned long start,
77 			   unsigned long end);
78 
79 	/*
80 	 * test_young is called to check the young/accessed bitflag in
81 	 * the secondary pte. This is used to know if the page is
82 	 * frequently used without actually clearing the flag or tearing
83 	 * down the secondary mapping on the page.
84 	 */
85 	int (*test_young)(struct mmu_notifier *mn,
86 			  struct mm_struct *mm,
87 			  unsigned long address);
88 
89 	/*
90 	 * change_pte is called in cases that pte mapping to page is changed:
91 	 * for example, when ksm remaps pte to point to a new shared page.
92 	 */
93 	void (*change_pte)(struct mmu_notifier *mn,
94 			   struct mm_struct *mm,
95 			   unsigned long address,
96 			   pte_t pte);
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
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().
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);
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. For more in depth
160 	 * discussion on this see Documentation/vm/mmu_notifier.txt
161 	 *
162 	 * The invalidate_range() function is called under the ptl
163 	 * spin-lock and not allowed to sleep.
164 	 *
165 	 * Note that this function might be called with just a sub-range
166 	 * of what was passed to invalidate_range_start()/end(), if
167 	 * called between those functions.
168 	 */
169 	void (*invalidate_range)(struct mmu_notifier *mn, struct mm_struct *mm,
170 				 unsigned long start, unsigned long end);
171 };
172 
173 /*
174  * The notifier chains are protected by mmap_sem and/or the reverse map
175  * semaphores. Notifier chains are only changed when all reverse maps and
176  * the mmap_sem locks are taken.
177  *
178  * Therefore notifier chains can only be traversed when either
179  *
180  * 1. mmap_sem is held.
181  * 2. One of the reverse map locks is held (i_mmap_rwsem or anon_vma->rwsem).
182  * 3. No other concurrent thread can access the list (release)
183  */
184 struct mmu_notifier {
185 	struct hlist_node hlist;
186 	const struct mmu_notifier_ops *ops;
187 };
188 
189 static inline int mm_has_notifiers(struct mm_struct *mm)
190 {
191 	return unlikely(mm->mmu_notifier_mm);
192 }
193 
194 extern int mmu_notifier_register(struct mmu_notifier *mn,
195 				 struct mm_struct *mm);
196 extern int __mmu_notifier_register(struct mmu_notifier *mn,
197 				   struct mm_struct *mm);
198 extern void mmu_notifier_unregister(struct mmu_notifier *mn,
199 				    struct mm_struct *mm);
200 extern void mmu_notifier_unregister_no_release(struct mmu_notifier *mn,
201 					       struct mm_struct *mm);
202 extern void __mmu_notifier_mm_destroy(struct mm_struct *mm);
203 extern void __mmu_notifier_release(struct mm_struct *mm);
204 extern int __mmu_notifier_clear_flush_young(struct mm_struct *mm,
205 					  unsigned long start,
206 					  unsigned long end);
207 extern int __mmu_notifier_clear_young(struct mm_struct *mm,
208 				      unsigned long start,
209 				      unsigned long end);
210 extern int __mmu_notifier_test_young(struct mm_struct *mm,
211 				     unsigned long address);
212 extern void __mmu_notifier_change_pte(struct mm_struct *mm,
213 				      unsigned long address, pte_t pte);
214 extern void __mmu_notifier_invalidate_range_start(struct mm_struct *mm,
215 				  unsigned long start, unsigned long end);
216 extern void __mmu_notifier_invalidate_range_end(struct mm_struct *mm,
217 				  unsigned long start, unsigned long end,
218 				  bool only_end);
219 extern void __mmu_notifier_invalidate_range(struct mm_struct *mm,
220 				  unsigned long start, unsigned long end);
221 
222 static inline void mmu_notifier_release(struct mm_struct *mm)
223 {
224 	if (mm_has_notifiers(mm))
225 		__mmu_notifier_release(mm);
226 }
227 
228 static inline int mmu_notifier_clear_flush_young(struct mm_struct *mm,
229 					  unsigned long start,
230 					  unsigned long end)
231 {
232 	if (mm_has_notifiers(mm))
233 		return __mmu_notifier_clear_flush_young(mm, start, end);
234 	return 0;
235 }
236 
237 static inline int mmu_notifier_clear_young(struct mm_struct *mm,
238 					   unsigned long start,
239 					   unsigned long end)
240 {
241 	if (mm_has_notifiers(mm))
242 		return __mmu_notifier_clear_young(mm, start, end);
243 	return 0;
244 }
245 
246 static inline int mmu_notifier_test_young(struct mm_struct *mm,
247 					  unsigned long address)
248 {
249 	if (mm_has_notifiers(mm))
250 		return __mmu_notifier_test_young(mm, address);
251 	return 0;
252 }
253 
254 static inline void mmu_notifier_change_pte(struct mm_struct *mm,
255 					   unsigned long address, pte_t pte)
256 {
257 	if (mm_has_notifiers(mm))
258 		__mmu_notifier_change_pte(mm, address, pte);
259 }
260 
261 static inline void mmu_notifier_invalidate_range_start(struct mm_struct *mm,
262 				  unsigned long start, unsigned long end)
263 {
264 	if (mm_has_notifiers(mm))
265 		__mmu_notifier_invalidate_range_start(mm, start, end);
266 }
267 
268 static inline void mmu_notifier_invalidate_range_end(struct mm_struct *mm,
269 				  unsigned long start, unsigned long end)
270 {
271 	if (mm_has_notifiers(mm))
272 		__mmu_notifier_invalidate_range_end(mm, start, end, false);
273 }
274 
275 static inline void mmu_notifier_invalidate_range_only_end(struct mm_struct *mm,
276 				  unsigned long start, unsigned long end)
277 {
278 	if (mm_has_notifiers(mm))
279 		__mmu_notifier_invalidate_range_end(mm, start, end, true);
280 }
281 
282 static inline void mmu_notifier_invalidate_range(struct mm_struct *mm,
283 				  unsigned long start, unsigned long end)
284 {
285 	if (mm_has_notifiers(mm))
286 		__mmu_notifier_invalidate_range(mm, start, end);
287 }
288 
289 static inline void mmu_notifier_mm_init(struct mm_struct *mm)
290 {
291 	mm->mmu_notifier_mm = NULL;
292 }
293 
294 static inline void mmu_notifier_mm_destroy(struct mm_struct *mm)
295 {
296 	if (mm_has_notifiers(mm))
297 		__mmu_notifier_mm_destroy(mm);
298 }
299 
300 #define ptep_clear_flush_young_notify(__vma, __address, __ptep)		\
301 ({									\
302 	int __young;							\
303 	struct vm_area_struct *___vma = __vma;				\
304 	unsigned long ___address = __address;				\
305 	__young = ptep_clear_flush_young(___vma, ___address, __ptep);	\
306 	__young |= mmu_notifier_clear_flush_young(___vma->vm_mm,	\
307 						  ___address,		\
308 						  ___address +		\
309 							PAGE_SIZE);	\
310 	__young;							\
311 })
312 
313 #define pmdp_clear_flush_young_notify(__vma, __address, __pmdp)		\
314 ({									\
315 	int __young;							\
316 	struct vm_area_struct *___vma = __vma;				\
317 	unsigned long ___address = __address;				\
318 	__young = pmdp_clear_flush_young(___vma, ___address, __pmdp);	\
319 	__young |= mmu_notifier_clear_flush_young(___vma->vm_mm,	\
320 						  ___address,		\
321 						  ___address +		\
322 							PMD_SIZE);	\
323 	__young;							\
324 })
325 
326 #define ptep_clear_young_notify(__vma, __address, __ptep)		\
327 ({									\
328 	int __young;							\
329 	struct vm_area_struct *___vma = __vma;				\
330 	unsigned long ___address = __address;				\
331 	__young = ptep_test_and_clear_young(___vma, ___address, __ptep);\
332 	__young |= mmu_notifier_clear_young(___vma->vm_mm, ___address,	\
333 					    ___address + PAGE_SIZE);	\
334 	__young;							\
335 })
336 
337 #define pmdp_clear_young_notify(__vma, __address, __pmdp)		\
338 ({									\
339 	int __young;							\
340 	struct vm_area_struct *___vma = __vma;				\
341 	unsigned long ___address = __address;				\
342 	__young = pmdp_test_and_clear_young(___vma, ___address, __pmdp);\
343 	__young |= mmu_notifier_clear_young(___vma->vm_mm, ___address,	\
344 					    ___address + PMD_SIZE);	\
345 	__young;							\
346 })
347 
348 #define	ptep_clear_flush_notify(__vma, __address, __ptep)		\
349 ({									\
350 	unsigned long ___addr = __address & PAGE_MASK;			\
351 	struct mm_struct *___mm = (__vma)->vm_mm;			\
352 	pte_t ___pte;							\
353 									\
354 	___pte = ptep_clear_flush(__vma, __address, __ptep);		\
355 	mmu_notifier_invalidate_range(___mm, ___addr,			\
356 					___addr + PAGE_SIZE);		\
357 									\
358 	___pte;								\
359 })
360 
361 #define pmdp_huge_clear_flush_notify(__vma, __haddr, __pmd)		\
362 ({									\
363 	unsigned long ___haddr = __haddr & HPAGE_PMD_MASK;		\
364 	struct mm_struct *___mm = (__vma)->vm_mm;			\
365 	pmd_t ___pmd;							\
366 									\
367 	___pmd = pmdp_huge_clear_flush(__vma, __haddr, __pmd);		\
368 	mmu_notifier_invalidate_range(___mm, ___haddr,			\
369 				      ___haddr + HPAGE_PMD_SIZE);	\
370 									\
371 	___pmd;								\
372 })
373 
374 #define pudp_huge_clear_flush_notify(__vma, __haddr, __pud)		\
375 ({									\
376 	unsigned long ___haddr = __haddr & HPAGE_PUD_MASK;		\
377 	struct mm_struct *___mm = (__vma)->vm_mm;			\
378 	pud_t ___pud;							\
379 									\
380 	___pud = pudp_huge_clear_flush(__vma, __haddr, __pud);		\
381 	mmu_notifier_invalidate_range(___mm, ___haddr,			\
382 				      ___haddr + HPAGE_PUD_SIZE);	\
383 									\
384 	___pud;								\
385 })
386 
387 /*
388  * set_pte_at_notify() sets the pte _after_ running the notifier.
389  * This is safe to start by updating the secondary MMUs, because the primary MMU
390  * pte invalidate must have already happened with a ptep_clear_flush() before
391  * set_pte_at_notify() has been invoked.  Updating the secondary MMUs first is
392  * required when we change both the protection of the mapping from read-only to
393  * read-write and the pfn (like during copy on write page faults). Otherwise the
394  * old page would remain mapped readonly in the secondary MMUs after the new
395  * page is already writable by some CPU through the primary MMU.
396  */
397 #define set_pte_at_notify(__mm, __address, __ptep, __pte)		\
398 ({									\
399 	struct mm_struct *___mm = __mm;					\
400 	unsigned long ___address = __address;				\
401 	pte_t ___pte = __pte;						\
402 									\
403 	mmu_notifier_change_pte(___mm, ___address, ___pte);		\
404 	set_pte_at(___mm, ___address, __ptep, ___pte);			\
405 })
406 
407 extern void mmu_notifier_call_srcu(struct rcu_head *rcu,
408 				   void (*func)(struct rcu_head *rcu));
409 extern void mmu_notifier_synchronize(void);
410 
411 #else /* CONFIG_MMU_NOTIFIER */
412 
413 static inline int mm_has_notifiers(struct mm_struct *mm)
414 {
415 	return 0;
416 }
417 
418 static inline void mmu_notifier_release(struct mm_struct *mm)
419 {
420 }
421 
422 static inline int mmu_notifier_clear_flush_young(struct mm_struct *mm,
423 					  unsigned long start,
424 					  unsigned long end)
425 {
426 	return 0;
427 }
428 
429 static inline int mmu_notifier_test_young(struct mm_struct *mm,
430 					  unsigned long address)
431 {
432 	return 0;
433 }
434 
435 static inline void mmu_notifier_change_pte(struct mm_struct *mm,
436 					   unsigned long address, pte_t pte)
437 {
438 }
439 
440 static inline void mmu_notifier_invalidate_range_start(struct mm_struct *mm,
441 				  unsigned long start, unsigned long end)
442 {
443 }
444 
445 static inline void mmu_notifier_invalidate_range_end(struct mm_struct *mm,
446 				  unsigned long start, unsigned long end)
447 {
448 }
449 
450 static inline void mmu_notifier_invalidate_range_only_end(struct mm_struct *mm,
451 				  unsigned long start, unsigned long end)
452 {
453 }
454 
455 static inline void mmu_notifier_invalidate_range(struct mm_struct *mm,
456 				  unsigned long start, unsigned long end)
457 {
458 }
459 
460 static inline void mmu_notifier_mm_init(struct mm_struct *mm)
461 {
462 }
463 
464 static inline void mmu_notifier_mm_destroy(struct mm_struct *mm)
465 {
466 }
467 
468 #define ptep_clear_flush_young_notify ptep_clear_flush_young
469 #define pmdp_clear_flush_young_notify pmdp_clear_flush_young
470 #define ptep_clear_young_notify ptep_test_and_clear_young
471 #define pmdp_clear_young_notify pmdp_test_and_clear_young
472 #define	ptep_clear_flush_notify ptep_clear_flush
473 #define pmdp_huge_clear_flush_notify pmdp_huge_clear_flush
474 #define pudp_huge_clear_flush_notify pudp_huge_clear_flush
475 #define set_pte_at_notify set_pte_at
476 
477 #endif /* CONFIG_MMU_NOTIFIER */
478 
479 #endif /* _LINUX_MMU_NOTIFIER_H */
480