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/mmap_lock.h>
9#include <linux/srcu.h>
10#include <linux/interval_tree.h>
11
12struct mmu_notifier_subscriptions;
13struct mmu_notifier;
14struct mmu_notifier_range;
15struct mmu_interval_notifier;
16
17/**
18 * enum mmu_notifier_event - reason for the mmu notifier callback
19 * @MMU_NOTIFY_UNMAP: either munmap() that unmap the range or a mremap() that
20 * move the range
21 *
22 * @MMU_NOTIFY_CLEAR: clear page table entry (many reasons for this like
23 * madvise() or replacing a page by another one, ...).
24 *
25 * @MMU_NOTIFY_PROTECTION_VMA: update is due to protection change for the range
26 * ie using the vma access permission (vm_page_prot) to update the whole range
27 * is enough no need to inspect changes to the CPU page table (mprotect()
28 * syscall)
29 *
30 * @MMU_NOTIFY_PROTECTION_PAGE: update is due to change in read/write flag for
31 * pages in the range so to mirror those changes the user must inspect the CPU
32 * page table (from the end callback).
33 *
34 * @MMU_NOTIFY_SOFT_DIRTY: soft dirty accounting (still same page and same
35 * access flags). User should soft dirty the page in the end callback to make
36 * sure that anyone relying on soft dirtiness catch pages that might be written
37 * through non CPU mappings.
38 *
39 * @MMU_NOTIFY_RELEASE: used during mmu_interval_notifier invalidate to signal
40 * that the mm refcount is zero and the range is no longer accessible.
41 *
42 * @MMU_NOTIFY_MIGRATE: used during migrate_vma_collect() invalidate to signal
43 * a device driver to possibly ignore the invalidation if the
44 * owner field matches the driver's device private pgmap owner.
45 *
46 * @MMU_NOTIFY_EXCLUSIVE: conversion of a page table entry to device-exclusive.
47 * The owner is initialized to the value provided by the caller of
48 * make_device_exclusive(), such that this caller can filter out these
49 * events.
50 */
51enum mmu_notifier_event {
52 MMU_NOTIFY_UNMAP = 0,
53 MMU_NOTIFY_CLEAR,
54 MMU_NOTIFY_PROTECTION_VMA,
55 MMU_NOTIFY_PROTECTION_PAGE,
56 MMU_NOTIFY_SOFT_DIRTY,
57 MMU_NOTIFY_RELEASE,
58 MMU_NOTIFY_MIGRATE,
59 MMU_NOTIFY_EXCLUSIVE,
60};
61
62#define MMU_NOTIFIER_RANGE_BLOCKABLE (1 << 0)
63
64struct mmu_notifier_ops {
65 /*
66 * Called either by mmu_notifier_unregister or when the mm is
67 * being destroyed by exit_mmap, always before all pages are
68 * freed. This can run concurrently with other mmu notifier
69 * methods (the ones invoked outside the mm context) and it
70 * should tear down all secondary mmu mappings and freeze the
71 * secondary mmu. If this method isn't implemented you've to
72 * be sure that nothing could possibly write to the pages
73 * through the secondary mmu by the time the last thread with
74 * tsk->mm == mm exits.
75 *
76 * As side note: the pages freed after ->release returns could
77 * be immediately reallocated by the gart at an alias physical
78 * address with a different cache model, so if ->release isn't
79 * implemented because all _software_ driven memory accesses
80 * through the secondary mmu are terminated by the time the
81 * last thread of this mm quits, you've also to be sure that
82 * speculative _hardware_ operations can't allocate dirty
83 * cachelines in the cpu that could not be snooped and made
84 * coherent with the other read and write operations happening
85 * through the gart alias address, so leading to memory
86 * corruption.
87 */
88 void (*release)(struct mmu_notifier *subscription,
89 struct mm_struct *mm);
90
91 /*
92 * clear_flush_young is called after the VM is
93 * test-and-clearing the young/accessed bitflag in the
94 * pte. This way the VM will provide proper aging to the
95 * accesses to the page through the secondary MMUs and not
96 * only to the ones through the Linux pte.
97 * Start-end is necessary in case the secondary MMU is mapping the page
98 * at a smaller granularity than the primary MMU.
99 */
100 int (*clear_flush_young)(struct mmu_notifier *subscription,
101 struct mm_struct *mm,
102 unsigned long start,
103 unsigned long end);
104
105 /*
106 * clear_young is a lightweight version of clear_flush_young. Like the
107 * latter, it is supposed to test-and-clear the young/accessed bitflag
108 * in the secondary pte, but it may omit flushing the secondary tlb.
109 */
110 int (*clear_young)(struct mmu_notifier *subscription,
111 struct mm_struct *mm,
112 unsigned long start,
113 unsigned long end);
114
115 /*
116 * test_young is called to check the young/accessed bitflag in
117 * the secondary pte. This is used to know if the page is
118 * frequently used without actually clearing the flag or tearing
119 * down the secondary mapping on the page.
120 */
121 int (*test_young)(struct mmu_notifier *subscription,
122 struct mm_struct *mm,
123 unsigned long address);
124
125 /*
126 * invalidate_range_start() and invalidate_range_end() must be
127 * paired and are called only when the mmap_lock and/or the
128 * locks protecting the reverse maps are held. If the subsystem
129 * can't guarantee that no additional references are taken to
130 * the pages in the range, it has to implement the
131 * invalidate_range() notifier to remove any references taken
132 * after invalidate_range_start().
133 *
134 * Invalidation of multiple concurrent ranges may be
135 * optionally permitted by the driver. Either way the
136 * establishment of sptes is forbidden in the range passed to
137 * invalidate_range_begin/end for the whole duration of the
138 * invalidate_range_begin/end critical section.
139 *
140 * invalidate_range_start() is called when all pages in the
141 * range are still mapped and have at least a refcount of one.
142 *
143 * invalidate_range_end() is called when all pages in the
144 * range have been unmapped and the pages have been freed by
145 * the VM.
146 *
147 * The VM will remove the page table entries and potentially
148 * the page between invalidate_range_start() and
149 * invalidate_range_end(). If the page must not be freed
150 * because of pending I/O or other circumstances then the
151 * invalidate_range_start() callback (or the initial mapping
152 * by the driver) must make sure that the refcount is kept
153 * elevated.
154 *
155 * If the driver increases the refcount when the pages are
156 * initially mapped into an address space then either
157 * invalidate_range_start() or invalidate_range_end() may
158 * decrease the refcount. If the refcount is decreased on
159 * invalidate_range_start() then the VM can free pages as page
160 * table entries are removed. If the refcount is only
161 * dropped on invalidate_range_end() then the driver itself
162 * will drop the last refcount but it must take care to flush
163 * any secondary tlb before doing the final free on the
164 * page. Pages will no longer be referenced by the linux
165 * address space but may still be referenced by sptes until
166 * the last refcount is dropped.
167 *
168 * If blockable argument is set to false then the callback cannot
169 * sleep and has to return with -EAGAIN if sleeping would be required.
170 * 0 should be returned otherwise. Please note that notifiers that can
171 * fail invalidate_range_start are not allowed to implement
172 * invalidate_range_end, as there is no mechanism for informing the
173 * notifier that its start failed.
174 */
175 int (*invalidate_range_start)(struct mmu_notifier *subscription,
176 const struct mmu_notifier_range *range);
177 void (*invalidate_range_end)(struct mmu_notifier *subscription,
178 const struct mmu_notifier_range *range);
179
180 /*
181 * arch_invalidate_secondary_tlbs() is used to manage a non-CPU TLB
182 * which shares page-tables with the CPU. The
183 * invalidate_range_start()/end() callbacks should not be implemented as
184 * invalidate_secondary_tlbs() already catches the points in time when
185 * an external TLB needs to be flushed.
186 *
187 * This requires arch_invalidate_secondary_tlbs() to be called while
188 * holding the ptl spin-lock and therefore this callback is not allowed
189 * to sleep.
190 *
191 * This is called by architecture code whenever invalidating a TLB
192 * entry. It is assumed that any secondary TLB has the same rules for
193 * when invalidations are required. If this is not the case architecture
194 * code will need to call this explicitly when required for secondary
195 * TLB invalidation.
196 */
197 void (*arch_invalidate_secondary_tlbs)(
198 struct mmu_notifier *subscription,
199 struct mm_struct *mm,
200 unsigned long start,
201 unsigned long end);
202
203 /*
204 * These callbacks are used with the get/put interface to manage the
205 * lifetime of the mmu_notifier memory. alloc_notifier() returns a new
206 * notifier for use with the mm.
207 *
208 * free_notifier() is only called after the mmu_notifier has been
209 * fully put, calls to any ops callback are prevented and no ops
210 * callbacks are currently running. It is called from a SRCU callback
211 * and cannot sleep.
212 */
213 struct mmu_notifier *(*alloc_notifier)(struct mm_struct *mm);
214 void (*free_notifier)(struct mmu_notifier *subscription);
215};
216
217/*
218 * The notifier chains are protected by mmap_lock and/or the reverse map
219 * semaphores. Notifier chains are only changed when all reverse maps and
220 * the mmap_lock locks are taken.
221 *
222 * Therefore notifier chains can only be traversed when either
223 *
224 * 1. mmap_lock is held.
225 * 2. One of the reverse map locks is held (i_mmap_rwsem or anon_vma->rwsem).
226 * 3. No other concurrent thread can access the list (release)
227 */
228struct mmu_notifier {
229 struct hlist_node hlist;
230 const struct mmu_notifier_ops *ops;
231 struct mm_struct *mm;
232 struct rcu_head rcu;
233 unsigned int users;
234};
235
236/**
237 * struct mmu_interval_notifier_ops
238 * @invalidate: Upon return the caller must stop using any SPTEs within this
239 * range. This function can sleep. Return false only if sleeping
240 * was required but mmu_notifier_range_blockable(range) is false.
241 */
242struct mmu_interval_notifier_ops {
243 bool (*invalidate)(struct mmu_interval_notifier *interval_sub,
244 const struct mmu_notifier_range *range,
245 unsigned long cur_seq);
246};
247
248struct mmu_interval_notifier {
249 struct interval_tree_node interval_tree;
250 const struct mmu_interval_notifier_ops *ops;
251 struct mm_struct *mm;
252 struct hlist_node deferred_item;
253 unsigned long invalidate_seq;
254};
255
256#ifdef CONFIG_MMU_NOTIFIER
257
258#ifdef CONFIG_LOCKDEP
259extern struct lockdep_map __mmu_notifier_invalidate_range_start_map;
260#endif
261
262struct mmu_notifier_range {
263 struct mm_struct *mm;
264 unsigned long start;
265 unsigned long end;
266 unsigned flags;
267 enum mmu_notifier_event event;
268 void *owner;
269};
270
271static inline int mm_has_notifiers(struct mm_struct *mm)
272{
273 return unlikely(mm->notifier_subscriptions);
274}
275
276struct mmu_notifier *mmu_notifier_get_locked(const struct mmu_notifier_ops *ops,
277 struct mm_struct *mm);
278static inline struct mmu_notifier *
279mmu_notifier_get(const struct mmu_notifier_ops *ops, struct mm_struct *mm)
280{
281 struct mmu_notifier *ret;
282
283 mmap_write_lock(mm);
284 ret = mmu_notifier_get_locked(ops, mm);
285 mmap_write_unlock(mm);
286 return ret;
287}
288void mmu_notifier_put(struct mmu_notifier *subscription);
289void mmu_notifier_synchronize(void);
290
291extern int mmu_notifier_register(struct mmu_notifier *subscription,
292 struct mm_struct *mm);
293extern int __mmu_notifier_register(struct mmu_notifier *subscription,
294 struct mm_struct *mm);
295extern void mmu_notifier_unregister(struct mmu_notifier *subscription,
296 struct mm_struct *mm);
297
298unsigned long
299mmu_interval_read_begin(struct mmu_interval_notifier *interval_sub);
300int mmu_interval_notifier_insert(struct mmu_interval_notifier *interval_sub,
301 struct mm_struct *mm, unsigned long start,
302 unsigned long length,
303 const struct mmu_interval_notifier_ops *ops);
304int mmu_interval_notifier_insert_locked(
305 struct mmu_interval_notifier *interval_sub, struct mm_struct *mm,
306 unsigned long start, unsigned long length,
307 const struct mmu_interval_notifier_ops *ops);
308void mmu_interval_notifier_remove(struct mmu_interval_notifier *interval_sub);
309
310/**
311 * mmu_interval_set_seq - Save the invalidation sequence
312 * @interval_sub - The subscription passed to invalidate
313 * @cur_seq - The cur_seq passed to the invalidate() callback
314 *
315 * This must be called unconditionally from the invalidate callback of a
316 * struct mmu_interval_notifier_ops under the same lock that is used to call
317 * mmu_interval_read_retry(). It updates the sequence number for later use by
318 * mmu_interval_read_retry(). The provided cur_seq will always be odd.
319 *
320 * If the caller does not call mmu_interval_read_begin() or
321 * mmu_interval_read_retry() then this call is not required.
322 */
323static inline void
324mmu_interval_set_seq(struct mmu_interval_notifier *interval_sub,
325 unsigned long cur_seq)
326{
327 WRITE_ONCE(interval_sub->invalidate_seq, cur_seq);
328}
329
330/**
331 * mmu_interval_read_retry - End a read side critical section against a VA range
332 * interval_sub: The subscription
333 * seq: The return of the paired mmu_interval_read_begin()
334 *
335 * This MUST be called under a user provided lock that is also held
336 * unconditionally by op->invalidate() when it calls mmu_interval_set_seq().
337 *
338 * Each call should be paired with a single mmu_interval_read_begin() and
339 * should be used to conclude the read side.
340 *
341 * Returns true if an invalidation collided with this critical section, and
342 * the caller should retry.
343 */
344static inline bool
345mmu_interval_read_retry(struct mmu_interval_notifier *interval_sub,
346 unsigned long seq)
347{
348 return interval_sub->invalidate_seq != seq;
349}
350
351/**
352 * mmu_interval_check_retry - Test if a collision has occurred
353 * interval_sub: The subscription
354 * seq: The return of the matching mmu_interval_read_begin()
355 *
356 * This can be used in the critical section between mmu_interval_read_begin()
357 * and mmu_interval_read_retry(). A return of true indicates an invalidation
358 * has collided with this critical region and a future
359 * mmu_interval_read_retry() will return true.
360 *
361 * False is not reliable and only suggests a collision may not have
362 * occurred. It can be called many times and does not have to hold the user
363 * provided lock.
364 *
365 * This call can be used as part of loops and other expensive operations to
366 * expedite a retry.
367 */
368static inline bool
369mmu_interval_check_retry(struct mmu_interval_notifier *interval_sub,
370 unsigned long seq)
371{
372 /* Pairs with the WRITE_ONCE in mmu_interval_set_seq() */
373 return READ_ONCE(interval_sub->invalidate_seq) != seq;
374}
375
376extern void __mmu_notifier_subscriptions_destroy(struct mm_struct *mm);
377extern void __mmu_notifier_release(struct mm_struct *mm);
378extern int __mmu_notifier_clear_flush_young(struct mm_struct *mm,
379 unsigned long start,
380 unsigned long end);
381extern int __mmu_notifier_clear_young(struct mm_struct *mm,
382 unsigned long start,
383 unsigned long end);
384extern int __mmu_notifier_test_young(struct mm_struct *mm,
385 unsigned long address);
386extern int __mmu_notifier_invalidate_range_start(struct mmu_notifier_range *r);
387extern void __mmu_notifier_invalidate_range_end(struct mmu_notifier_range *r);
388extern void __mmu_notifier_arch_invalidate_secondary_tlbs(struct mm_struct *mm,
389 unsigned long start, unsigned long end);
390extern bool
391mmu_notifier_range_update_to_read_only(const struct mmu_notifier_range *range);
392
393static inline bool
394mmu_notifier_range_blockable(const struct mmu_notifier_range *range)
395{
396 return (range->flags & MMU_NOTIFIER_RANGE_BLOCKABLE);
397}
398
399static inline void mmu_notifier_release(struct mm_struct *mm)
400{
401 if (mm_has_notifiers(mm))
402 __mmu_notifier_release(mm);
403}
404
405static inline int mmu_notifier_clear_flush_young(struct mm_struct *mm,
406 unsigned long start,
407 unsigned long end)
408{
409 if (mm_has_notifiers(mm))
410 return __mmu_notifier_clear_flush_young(mm, start, end);
411 return 0;
412}
413
414static inline int mmu_notifier_clear_young(struct mm_struct *mm,
415 unsigned long start,
416 unsigned long end)
417{
418 if (mm_has_notifiers(mm))
419 return __mmu_notifier_clear_young(mm, start, end);
420 return 0;
421}
422
423static inline int mmu_notifier_test_young(struct mm_struct *mm,
424 unsigned long address)
425{
426 if (mm_has_notifiers(mm))
427 return __mmu_notifier_test_young(mm, address);
428 return 0;
429}
430
431static inline void
432mmu_notifier_invalidate_range_start(struct mmu_notifier_range *range)
433{
434 might_sleep();
435
436 lock_map_acquire(&__mmu_notifier_invalidate_range_start_map);
437 if (mm_has_notifiers(mm: range->mm)) {
438 range->flags |= MMU_NOTIFIER_RANGE_BLOCKABLE;
439 __mmu_notifier_invalidate_range_start(r: range);
440 }
441 lock_map_release(&__mmu_notifier_invalidate_range_start_map);
442}
443
444/*
445 * This version of mmu_notifier_invalidate_range_start() avoids blocking, but it
446 * can return an error if a notifier can't proceed without blocking, in which
447 * case you're not allowed to modify PTEs in the specified range.
448 *
449 * This is mainly intended for OOM handling.
450 */
451static inline int __must_check
452mmu_notifier_invalidate_range_start_nonblock(struct mmu_notifier_range *range)
453{
454 int ret = 0;
455
456 lock_map_acquire(&__mmu_notifier_invalidate_range_start_map);
457 if (mm_has_notifiers(mm: range->mm)) {
458 range->flags &= ~MMU_NOTIFIER_RANGE_BLOCKABLE;
459 ret = __mmu_notifier_invalidate_range_start(r: range);
460 }
461 lock_map_release(&__mmu_notifier_invalidate_range_start_map);
462 return ret;
463}
464
465static inline void
466mmu_notifier_invalidate_range_end(struct mmu_notifier_range *range)
467{
468 if (mmu_notifier_range_blockable(range))
469 might_sleep();
470
471 if (mm_has_notifiers(mm: range->mm))
472 __mmu_notifier_invalidate_range_end(r: range);
473}
474
475static inline void mmu_notifier_arch_invalidate_secondary_tlbs(struct mm_struct *mm,
476 unsigned long start, unsigned long end)
477{
478 if (mm_has_notifiers(mm))
479 __mmu_notifier_arch_invalidate_secondary_tlbs(mm, start, end);
480}
481
482static inline void mmu_notifier_subscriptions_init(struct mm_struct *mm)
483{
484 mm->notifier_subscriptions = NULL;
485}
486
487static inline void mmu_notifier_subscriptions_destroy(struct mm_struct *mm)
488{
489 if (mm_has_notifiers(mm))
490 __mmu_notifier_subscriptions_destroy(mm);
491}
492
493
494static inline void mmu_notifier_range_init(struct mmu_notifier_range *range,
495 enum mmu_notifier_event event,
496 unsigned flags,
497 struct mm_struct *mm,
498 unsigned long start,
499 unsigned long end)
500{
501 range->event = event;
502 range->mm = mm;
503 range->start = start;
504 range->end = end;
505 range->flags = flags;
506}
507
508static inline void mmu_notifier_range_init_owner(
509 struct mmu_notifier_range *range,
510 enum mmu_notifier_event event, unsigned int flags,
511 struct mm_struct *mm, unsigned long start,
512 unsigned long end, void *owner)
513{
514 mmu_notifier_range_init(range, event, flags, mm, start, end);
515 range->owner = owner;
516}
517
518#define ptep_clear_flush_young_notify(__vma, __address, __ptep) \
519({ \
520 int __young; \
521 struct vm_area_struct *___vma = __vma; \
522 unsigned long ___address = __address; \
523 __young = ptep_clear_flush_young(___vma, ___address, __ptep); \
524 __young |= mmu_notifier_clear_flush_young(___vma->vm_mm, \
525 ___address, \
526 ___address + \
527 PAGE_SIZE); \
528 __young; \
529})
530
531#define pmdp_clear_flush_young_notify(__vma, __address, __pmdp) \
532({ \
533 int __young; \
534 struct vm_area_struct *___vma = __vma; \
535 unsigned long ___address = __address; \
536 __young = pmdp_clear_flush_young(___vma, ___address, __pmdp); \
537 __young |= mmu_notifier_clear_flush_young(___vma->vm_mm, \
538 ___address, \
539 ___address + \
540 PMD_SIZE); \
541 __young; \
542})
543
544#define ptep_clear_young_notify(__vma, __address, __ptep) \
545({ \
546 int __young; \
547 struct vm_area_struct *___vma = __vma; \
548 unsigned long ___address = __address; \
549 __young = ptep_test_and_clear_young(___vma, ___address, __ptep);\
550 __young |= mmu_notifier_clear_young(___vma->vm_mm, ___address, \
551 ___address + PAGE_SIZE); \
552 __young; \
553})
554
555#define pmdp_clear_young_notify(__vma, __address, __pmdp) \
556({ \
557 int __young; \
558 struct vm_area_struct *___vma = __vma; \
559 unsigned long ___address = __address; \
560 __young = pmdp_test_and_clear_young(___vma, ___address, __pmdp);\
561 __young |= mmu_notifier_clear_young(___vma->vm_mm, ___address, \
562 ___address + PMD_SIZE); \
563 __young; \
564})
565
566#else /* CONFIG_MMU_NOTIFIER */
567
568struct mmu_notifier_range {
569 unsigned long start;
570 unsigned long end;
571};
572
573static inline void _mmu_notifier_range_init(struct mmu_notifier_range *range,
574 unsigned long start,
575 unsigned long end)
576{
577 range->start = start;
578 range->end = end;
579}
580
581#define mmu_notifier_range_init(range,event,flags,mm,start,end) \
582 _mmu_notifier_range_init(range, start, end)
583#define mmu_notifier_range_init_owner(range, event, flags, mm, start, \
584 end, owner) \
585 _mmu_notifier_range_init(range, start, end)
586
587static inline bool
588mmu_notifier_range_blockable(const struct mmu_notifier_range *range)
589{
590 return true;
591}
592
593static inline int mm_has_notifiers(struct mm_struct *mm)
594{
595 return 0;
596}
597
598static inline void mmu_notifier_release(struct mm_struct *mm)
599{
600}
601
602static inline int mmu_notifier_clear_flush_young(struct mm_struct *mm,
603 unsigned long start,
604 unsigned long end)
605{
606 return 0;
607}
608
609static inline int mmu_notifier_clear_young(struct mm_struct *mm,
610 unsigned long start,
611 unsigned long end)
612{
613 return 0;
614}
615
616static inline int mmu_notifier_test_young(struct mm_struct *mm,
617 unsigned long address)
618{
619 return 0;
620}
621
622static inline void
623mmu_notifier_invalidate_range_start(struct mmu_notifier_range *range)
624{
625}
626
627static inline int
628mmu_notifier_invalidate_range_start_nonblock(struct mmu_notifier_range *range)
629{
630 return 0;
631}
632
633static inline
634void mmu_notifier_invalidate_range_end(struct mmu_notifier_range *range)
635{
636}
637
638static inline void mmu_notifier_arch_invalidate_secondary_tlbs(struct mm_struct *mm,
639 unsigned long start, unsigned long end)
640{
641}
642
643static inline void mmu_notifier_subscriptions_init(struct mm_struct *mm)
644{
645}
646
647static inline void mmu_notifier_subscriptions_destroy(struct mm_struct *mm)
648{
649}
650
651#define mmu_notifier_range_update_to_read_only(r) false
652
653#define ptep_clear_flush_young_notify ptep_clear_flush_young
654#define pmdp_clear_flush_young_notify pmdp_clear_flush_young
655#define ptep_clear_young_notify ptep_test_and_clear_young
656#define pmdp_clear_young_notify pmdp_test_and_clear_young
657
658static inline void mmu_notifier_synchronize(void)
659{
660}
661
662#endif /* CONFIG_MMU_NOTIFIER */
663
664#endif /* _LINUX_MMU_NOTIFIER_H */
665