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 /**
14 * enum mmu_notifier_event - reason for the mmu notifier callback
15 * @MMU_NOTIFY_UNMAP: either munmap() that unmap the range or a mremap() that
16 * move the range
17 *
18 * @MMU_NOTIFY_CLEAR: clear page table entry (many reasons for this like
19 * madvise() or replacing a page by another one, ...).
20 *
21 * @MMU_NOTIFY_PROTECTION_VMA: update is due to protection change for the range
22 * ie using the vma access permission (vm_page_prot) to update the whole range
23 * is enough no need to inspect changes to the CPU page table (mprotect()
24 * syscall)
25 *
26 * @MMU_NOTIFY_PROTECTION_PAGE: update is due to change in read/write flag for
27 * pages in the range so to mirror those changes the user must inspect the CPU
28 * page table (from the end callback).
29 *
30 * @MMU_NOTIFY_SOFT_DIRTY: soft dirty accounting (still same page and same
31 * access flags). User should soft dirty the page in the end callback to make
32 * sure that anyone relying on soft dirtyness catch pages that might be written
33 * through non CPU mappings.
34 */
35 enum mmu_notifier_event {
36 MMU_NOTIFY_UNMAP = 0,
37 MMU_NOTIFY_CLEAR,
38 MMU_NOTIFY_PROTECTION_VMA,
39 MMU_NOTIFY_PROTECTION_PAGE,
40 MMU_NOTIFY_SOFT_DIRTY,
41 };
42
43 #ifdef CONFIG_MMU_NOTIFIER
44
45 #ifdef CONFIG_LOCKDEP
46 extern struct lockdep_map __mmu_notifier_invalidate_range_start_map;
47 #endif
48
49 /*
50 * The mmu notifier_mm structure is allocated and installed in
51 * mm->mmu_notifier_mm inside the mm_take_all_locks() protected
52 * critical section and it's released only when mm_count reaches zero
53 * in mmdrop().
54 */
55 struct mmu_notifier_mm {
56 /* all mmu notifiers registerd in this mm are queued in this list */
57 struct hlist_head list;
58 /* to serialize the list modifications and hlist_unhashed */
59 spinlock_t lock;
60 };
61
62 #define MMU_NOTIFIER_RANGE_BLOCKABLE (1 << 0)
63
64 struct mmu_notifier_range {
65 struct vm_area_struct *vma;
66 struct mm_struct *mm;
67 unsigned long start;
68 unsigned long end;
69 unsigned flags;
70 enum mmu_notifier_event event;
71 };
72
73 struct mmu_notifier_ops {
74 /*
75 * Called either by mmu_notifier_unregister or when the mm is
76 * being destroyed by exit_mmap, always before all pages are
77 * freed. This can run concurrently with other mmu notifier
78 * methods (the ones invoked outside the mm context) and it
79 * should tear down all secondary mmu mappings and freeze the
80 * secondary mmu. If this method isn't implemented you've to
81 * be sure that nothing could possibly write to the pages
82 * through the secondary mmu by the time the last thread with
83 * tsk->mm == mm exits.
84 *
85 * As side note: the pages freed after ->release returns could
86 * be immediately reallocated by the gart at an alias physical
87 * address with a different cache model, so if ->release isn't
88 * implemented because all _software_ driven memory accesses
89 * through the secondary mmu are terminated by the time the
90 * last thread of this mm quits, you've also to be sure that
91 * speculative _hardware_ operations can't allocate dirty
92 * cachelines in the cpu that could not be snooped and made
93 * coherent with the other read and write operations happening
94 * through the gart alias address, so leading to memory
95 * corruption.
96 */
97 void (*release)(struct mmu_notifier *mn,
98 struct mm_struct *mm);
99
100 /*
101 * clear_flush_young is called after the VM is
102 * test-and-clearing the young/accessed bitflag in the
103 * pte. This way the VM will provide proper aging to the
104 * accesses to the page through the secondary MMUs and not
105 * only to the ones through the Linux pte.
106 * Start-end is necessary in case the secondary MMU is mapping the page
107 * at a smaller granularity than the primary MMU.
108 */
109 int (*clear_flush_young)(struct mmu_notifier *mn,
110 struct mm_struct *mm,
111 unsigned long start,
112 unsigned long end);
113
114 /*
115 * clear_young is a lightweight version of clear_flush_young. Like the
116 * latter, it is supposed to test-and-clear the young/accessed bitflag
117 * in the secondary pte, but it may omit flushing the secondary tlb.
118 */
119 int (*clear_young)(struct mmu_notifier *mn,
120 struct mm_struct *mm,
121 unsigned long start,
122 unsigned long end);
123
124 /*
125 * test_young is called to check the young/accessed bitflag in
126 * the secondary pte. This is used to know if the page is
127 * frequently used without actually clearing the flag or tearing
128 * down the secondary mapping on the page.
129 */
130 int (*test_young)(struct mmu_notifier *mn,
131 struct mm_struct *mm,
132 unsigned long address);
133
134 /*
135 * change_pte is called in cases that pte mapping to page is changed:
136 * for example, when ksm remaps pte to point to a new shared page.
137 */
138 void (*change_pte)(struct mmu_notifier *mn,
139 struct mm_struct *mm,
140 unsigned long address,
141 pte_t pte);
142
143 /*
144 * invalidate_range_start() and invalidate_range_end() must be
145 * paired and are called only when the mmap_sem and/or the
146 * locks protecting the reverse maps are held. If the subsystem
147 * can't guarantee that no additional references are taken to
148 * the pages in the range, it has to implement the
149 * invalidate_range() notifier to remove any references taken
150 * after invalidate_range_start().
151 *
152 * Invalidation of multiple concurrent ranges may be
153 * optionally permitted by the driver. Either way the
154 * establishment of sptes is forbidden in the range passed to
155 * invalidate_range_begin/end for the whole duration of the
156 * invalidate_range_begin/end critical section.
157 *
158 * invalidate_range_start() is called when all pages in the
159 * range are still mapped and have at least a refcount of one.
160 *
161 * invalidate_range_end() is called when all pages in the
162 * range have been unmapped and the pages have been freed by
163 * the VM.
164 *
165 * The VM will remove the page table entries and potentially
166 * the page between invalidate_range_start() and
167 * invalidate_range_end(). If the page must not be freed
168 * because of pending I/O or other circumstances then the
169 * invalidate_range_start() callback (or the initial mapping
170 * by the driver) must make sure that the refcount is kept
171 * elevated.
172 *
173 * If the driver increases the refcount when the pages are
174 * initially mapped into an address space then either
175 * invalidate_range_start() or invalidate_range_end() may
176 * decrease the refcount. If the refcount is decreased on
177 * invalidate_range_start() then the VM can free pages as page
178 * table entries are removed. If the refcount is only
179 * droppped on invalidate_range_end() then the driver itself
180 * will drop the last refcount but it must take care to flush
181 * any secondary tlb before doing the final free on the
182 * page. Pages will no longer be referenced by the linux
183 * address space but may still be referenced by sptes until
184 * the last refcount is dropped.
185 *
186 * If blockable argument is set to false then the callback cannot
187 * sleep and has to return with -EAGAIN. 0 should be returned
188 * otherwise. Please note that if invalidate_range_start approves
189 * a non-blocking behavior then the same applies to
190 * invalidate_range_end.
191 *
192 */
193 int (*invalidate_range_start)(struct mmu_notifier *mn,
194 const struct mmu_notifier_range *range);
195 void (*invalidate_range_end)(struct mmu_notifier *mn,
196 const struct mmu_notifier_range *range);
197
198 /*
199 * invalidate_range() is either called between
200 * invalidate_range_start() and invalidate_range_end() when the
201 * VM has to free pages that where unmapped, but before the
202 * pages are actually freed, or outside of _start()/_end() when
203 * a (remote) TLB is necessary.
204 *
205 * If invalidate_range() is used to manage a non-CPU TLB with
206 * shared page-tables, it not necessary to implement the
207 * invalidate_range_start()/end() notifiers, as
208 * invalidate_range() alread catches the points in time when an
209 * external TLB range needs to be flushed. For more in depth
210 * discussion on this see Documentation/vm/mmu_notifier.rst
211 *
212 * Note that this function might be called with just a sub-range
213 * of what was passed to invalidate_range_start()/end(), if
214 * called between those functions.
215 */
216 void (*invalidate_range)(struct mmu_notifier *mn, struct mm_struct *mm,
217 unsigned long start, unsigned long end);
218
219 /*
220 * These callbacks are used with the get/put interface to manage the
221 * lifetime of the mmu_notifier memory. alloc_notifier() returns a new
222 * notifier for use with the mm.
223 *
224 * free_notifier() is only called after the mmu_notifier has been
225 * fully put, calls to any ops callback are prevented and no ops
226 * callbacks are currently running. It is called from a SRCU callback
227 * and cannot sleep.
228 */
229 struct mmu_notifier *(*alloc_notifier)(struct mm_struct *mm);
230 void (*free_notifier)(struct mmu_notifier *mn);
231 };
232
233 /*
234 * The notifier chains are protected by mmap_sem and/or the reverse map
235 * semaphores. Notifier chains are only changed when all reverse maps and
236 * the mmap_sem locks are taken.
237 *
238 * Therefore notifier chains can only be traversed when either
239 *
240 * 1. mmap_sem is held.
241 * 2. One of the reverse map locks is held (i_mmap_rwsem or anon_vma->rwsem).
242 * 3. No other concurrent thread can access the list (release)
243 */
244 struct mmu_notifier {
245 struct hlist_node hlist;
246 const struct mmu_notifier_ops *ops;
247 struct mm_struct *mm;
248 struct rcu_head rcu;
249 unsigned int users;
250 };
251
252 static inline int mm_has_notifiers(struct mm_struct *mm)
253 {
254 return unlikely(mm->mmu_notifier_mm);
255 }
256
257 struct mmu_notifier *mmu_notifier_get_locked(const struct mmu_notifier_ops *ops,
258 struct mm_struct *mm);
259 static inline struct mmu_notifier *
260 mmu_notifier_get(const struct mmu_notifier_ops *ops, struct mm_struct *mm)
261 {
262 struct mmu_notifier *ret;
263
264 down_write(&mm->mmap_sem);
265 ret = mmu_notifier_get_locked(ops, mm);
266 up_write(&mm->mmap_sem);
267 return ret;
268 }
269 void mmu_notifier_put(struct mmu_notifier *mn);
270 void mmu_notifier_synchronize(void);
271
272 extern int mmu_notifier_register(struct mmu_notifier *mn,
273 struct mm_struct *mm);
274 extern int __mmu_notifier_register(struct mmu_notifier *mn,
275 struct mm_struct *mm);
276 extern void mmu_notifier_unregister(struct mmu_notifier *mn,
277 struct mm_struct *mm);
278 extern void __mmu_notifier_mm_destroy(struct mm_struct *mm);
279 extern void __mmu_notifier_release(struct mm_struct *mm);
280 extern int __mmu_notifier_clear_flush_young(struct mm_struct *mm,
281 unsigned long start,
282 unsigned long end);
283 extern int __mmu_notifier_clear_young(struct mm_struct *mm,
284 unsigned long start,
285 unsigned long end);
286 extern int __mmu_notifier_test_young(struct mm_struct *mm,
287 unsigned long address);
288 extern void __mmu_notifier_change_pte(struct mm_struct *mm,
289 unsigned long address, pte_t pte);
290 extern int __mmu_notifier_invalidate_range_start(struct mmu_notifier_range *r);
291 extern void __mmu_notifier_invalidate_range_end(struct mmu_notifier_range *r,
292 bool only_end);
293 extern void __mmu_notifier_invalidate_range(struct mm_struct *mm,
294 unsigned long start, unsigned long end);
295 extern bool
296 mmu_notifier_range_update_to_read_only(const struct mmu_notifier_range *range);
297
298 static inline bool
299 mmu_notifier_range_blockable(const struct mmu_notifier_range *range)
300 {
301 return (range->flags & MMU_NOTIFIER_RANGE_BLOCKABLE);
302 }
303
304 static inline void mmu_notifier_release(struct mm_struct *mm)
305 {
306 if (mm_has_notifiers(mm))
307 __mmu_notifier_release(mm);
308 }
309
310 static inline int mmu_notifier_clear_flush_young(struct mm_struct *mm,
311 unsigned long start,
312 unsigned long end)
313 {
314 if (mm_has_notifiers(mm))
315 return __mmu_notifier_clear_flush_young(mm, start, end);
316 return 0;
317 }
318
319 static inline int mmu_notifier_clear_young(struct mm_struct *mm,
320 unsigned long start,
321 unsigned long end)
322 {
323 if (mm_has_notifiers(mm))
324 return __mmu_notifier_clear_young(mm, start, end);
325 return 0;
326 }
327
328 static inline int mmu_notifier_test_young(struct mm_struct *mm,
329 unsigned long address)
330 {
331 if (mm_has_notifiers(mm))
332 return __mmu_notifier_test_young(mm, address);
333 return 0;
334 }
335
336 static inline void mmu_notifier_change_pte(struct mm_struct *mm,
337 unsigned long address, pte_t pte)
338 {
339 if (mm_has_notifiers(mm))
340 __mmu_notifier_change_pte(mm, address, pte);
341 }
342
343 static inline void
344 mmu_notifier_invalidate_range_start(struct mmu_notifier_range *range)
345 {
346 might_sleep();
347
348 lock_map_acquire(&__mmu_notifier_invalidate_range_start_map);
349 if (mm_has_notifiers(range->mm)) {
350 range->flags |= MMU_NOTIFIER_RANGE_BLOCKABLE;
351 __mmu_notifier_invalidate_range_start(range);
352 }
353 lock_map_release(&__mmu_notifier_invalidate_range_start_map);
354 }
355
356 static inline int
357 mmu_notifier_invalidate_range_start_nonblock(struct mmu_notifier_range *range)
358 {
359 int ret = 0;
360
361 lock_map_acquire(&__mmu_notifier_invalidate_range_start_map);
362 if (mm_has_notifiers(range->mm)) {
363 range->flags &= ~MMU_NOTIFIER_RANGE_BLOCKABLE;
364 ret = __mmu_notifier_invalidate_range_start(range);
365 }
366 lock_map_release(&__mmu_notifier_invalidate_range_start_map);
367 return ret;
368 }
369
370 static inline void
371 mmu_notifier_invalidate_range_end(struct mmu_notifier_range *range)
372 {
373 if (mmu_notifier_range_blockable(range))
374 might_sleep();
375
376 if (mm_has_notifiers(range->mm))
377 __mmu_notifier_invalidate_range_end(range, false);
378 }
379
380 static inline void
381 mmu_notifier_invalidate_range_only_end(struct mmu_notifier_range *range)
382 {
383 if (mm_has_notifiers(range->mm))
384 __mmu_notifier_invalidate_range_end(range, true);
385 }
386
387 static inline void mmu_notifier_invalidate_range(struct mm_struct *mm,
388 unsigned long start, unsigned long end)
389 {
390 if (mm_has_notifiers(mm))
391 __mmu_notifier_invalidate_range(mm, start, end);
392 }
393
394 static inline void mmu_notifier_mm_init(struct mm_struct *mm)
395 {
396 mm->mmu_notifier_mm = NULL;
397 }
398
399 static inline void mmu_notifier_mm_destroy(struct mm_struct *mm)
400 {
401 if (mm_has_notifiers(mm))
402 __mmu_notifier_mm_destroy(mm);
403 }
404
405
406 static inline void mmu_notifier_range_init(struct mmu_notifier_range *range,
407 enum mmu_notifier_event event,
408 unsigned flags,
409 struct vm_area_struct *vma,
410 struct mm_struct *mm,
411 unsigned long start,
412 unsigned long end)
413 {
414 range->vma = vma;
415 range->event = event;
416 range->mm = mm;
417 range->start = start;
418 range->end = end;
419 range->flags = flags;
420 }
421
422 #define ptep_clear_flush_young_notify(__vma, __address, __ptep) \
423 ({ \
424 int __young; \
425 struct vm_area_struct *___vma = __vma; \
426 unsigned long ___address = __address; \
427 __young = ptep_clear_flush_young(___vma, ___address, __ptep); \
428 __young |= mmu_notifier_clear_flush_young(___vma->vm_mm, \
429 ___address, \
430 ___address + \
431 PAGE_SIZE); \
432 __young; \
433 })
434
435 #define pmdp_clear_flush_young_notify(__vma, __address, __pmdp) \
436 ({ \
437 int __young; \
438 struct vm_area_struct *___vma = __vma; \
439 unsigned long ___address = __address; \
440 __young = pmdp_clear_flush_young(___vma, ___address, __pmdp); \
441 __young |= mmu_notifier_clear_flush_young(___vma->vm_mm, \
442 ___address, \
443 ___address + \
444 PMD_SIZE); \
445 __young; \
446 })
447
448 #define ptep_clear_young_notify(__vma, __address, __ptep) \
449 ({ \
450 int __young; \
451 struct vm_area_struct *___vma = __vma; \
452 unsigned long ___address = __address; \
453 __young = ptep_test_and_clear_young(___vma, ___address, __ptep);\
454 __young |= mmu_notifier_clear_young(___vma->vm_mm, ___address, \
455 ___address + PAGE_SIZE); \
456 __young; \
457 })
458
459 #define pmdp_clear_young_notify(__vma, __address, __pmdp) \
460 ({ \
461 int __young; \
462 struct vm_area_struct *___vma = __vma; \
463 unsigned long ___address = __address; \
464 __young = pmdp_test_and_clear_young(___vma, ___address, __pmdp);\
465 __young |= mmu_notifier_clear_young(___vma->vm_mm, ___address, \
466 ___address + PMD_SIZE); \
467 __young; \
468 })
469
470 #define ptep_clear_flush_notify(__vma, __address, __ptep) \
471 ({ \
472 unsigned long ___addr = __address & PAGE_MASK; \
473 struct mm_struct *___mm = (__vma)->vm_mm; \
474 pte_t ___pte; \
475 \
476 ___pte = ptep_clear_flush(__vma, __address, __ptep); \
477 mmu_notifier_invalidate_range(___mm, ___addr, \
478 ___addr + PAGE_SIZE); \
479 \
480 ___pte; \
481 })
482
483 #define pmdp_huge_clear_flush_notify(__vma, __haddr, __pmd) \
484 ({ \
485 unsigned long ___haddr = __haddr & HPAGE_PMD_MASK; \
486 struct mm_struct *___mm = (__vma)->vm_mm; \
487 pmd_t ___pmd; \
488 \
489 ___pmd = pmdp_huge_clear_flush(__vma, __haddr, __pmd); \
490 mmu_notifier_invalidate_range(___mm, ___haddr, \
491 ___haddr + HPAGE_PMD_SIZE); \
492 \
493 ___pmd; \
494 })
495
496 #define pudp_huge_clear_flush_notify(__vma, __haddr, __pud) \
497 ({ \
498 unsigned long ___haddr = __haddr & HPAGE_PUD_MASK; \
499 struct mm_struct *___mm = (__vma)->vm_mm; \
500 pud_t ___pud; \
501 \
502 ___pud = pudp_huge_clear_flush(__vma, __haddr, __pud); \
503 mmu_notifier_invalidate_range(___mm, ___haddr, \
504 ___haddr + HPAGE_PUD_SIZE); \
505 \
506 ___pud; \
507 })
508
509 /*
510 * set_pte_at_notify() sets the pte _after_ running the notifier.
511 * This is safe to start by updating the secondary MMUs, because the primary MMU
512 * pte invalidate must have already happened with a ptep_clear_flush() before
513 * set_pte_at_notify() has been invoked. Updating the secondary MMUs first is
514 * required when we change both the protection of the mapping from read-only to
515 * read-write and the pfn (like during copy on write page faults). Otherwise the
516 * old page would remain mapped readonly in the secondary MMUs after the new
517 * page is already writable by some CPU through the primary MMU.
518 */
519 #define set_pte_at_notify(__mm, __address, __ptep, __pte) \
520 ({ \
521 struct mm_struct *___mm = __mm; \
522 unsigned long ___address = __address; \
523 pte_t ___pte = __pte; \
524 \
525 mmu_notifier_change_pte(___mm, ___address, ___pte); \
526 set_pte_at(___mm, ___address, __ptep, ___pte); \
527 })
528
529 #else /* CONFIG_MMU_NOTIFIER */
530
531 struct mmu_notifier_range {
532 unsigned long start;
533 unsigned long end;
534 };
535
536 static inline void _mmu_notifier_range_init(struct mmu_notifier_range *range,
537 unsigned long start,
538 unsigned long end)
539 {
540 range->start = start;
541 range->end = end;
542 }
543
544 #define mmu_notifier_range_init(range,event,flags,vma,mm,start,end) \
545 _mmu_notifier_range_init(range, start, end)
546
547 static inline bool
548 mmu_notifier_range_blockable(const struct mmu_notifier_range *range)
549 {
550 return true;
551 }
552
553 static inline int mm_has_notifiers(struct mm_struct *mm)
554 {
555 return 0;
556 }
557
558 static inline void mmu_notifier_release(struct mm_struct *mm)
559 {
560 }
561
562 static inline int mmu_notifier_clear_flush_young(struct mm_struct *mm,
563 unsigned long start,
564 unsigned long end)
565 {
566 return 0;
567 }
568
569 static inline int mmu_notifier_test_young(struct mm_struct *mm,
570 unsigned long address)
571 {
572 return 0;
573 }
574
575 static inline void mmu_notifier_change_pte(struct mm_struct *mm,
576 unsigned long address, pte_t pte)
577 {
578 }
579
580 static inline void
581 mmu_notifier_invalidate_range_start(struct mmu_notifier_range *range)
582 {
583 }
584
585 static inline int
586 mmu_notifier_invalidate_range_start_nonblock(struct mmu_notifier_range *range)
587 {
588 return 0;
589 }
590
591 static inline
592 void mmu_notifier_invalidate_range_end(struct mmu_notifier_range *range)
593 {
594 }
595
596 static inline void
597 mmu_notifier_invalidate_range_only_end(struct mmu_notifier_range *range)
598 {
599 }
600
601 static inline void mmu_notifier_invalidate_range(struct mm_struct *mm,
602 unsigned long start, unsigned long end)
603 {
604 }
605
606 static inline void mmu_notifier_mm_init(struct mm_struct *mm)
607 {
608 }
609
610 static inline void mmu_notifier_mm_destroy(struct mm_struct *mm)
611 {
612 }
613
614 #define mmu_notifier_range_update_to_read_only(r) false
615
616 #define ptep_clear_flush_young_notify ptep_clear_flush_young
617 #define pmdp_clear_flush_young_notify pmdp_clear_flush_young
618 #define ptep_clear_young_notify ptep_test_and_clear_young
619 #define pmdp_clear_young_notify pmdp_test_and_clear_young
620 #define ptep_clear_flush_notify ptep_clear_flush
621 #define pmdp_huge_clear_flush_notify pmdp_huge_clear_flush
622 #define pudp_huge_clear_flush_notify pudp_huge_clear_flush
623 #define set_pte_at_notify set_pte_at
624
625 static inline void mmu_notifier_synchronize(void)
626 {
627 }
628
629 #endif /* CONFIG_MMU_NOTIFIER */
630
631 #endif /* _LINUX_MMU_NOTIFIER_H */