1 /* SPDX-License-Identifier: GPL-2.0 */
2 #ifndef _LINUX_MMU_NOTIFIER_H
3 #define _LINUX_MMU_NOTIFIER_H
5 #include <linux/types.h>
6 #include <linux/list.h>
7 #include <linux/spinlock.h>
8 #include <linux/mm_types.h>
9 #include <linux/srcu.h>
12 struct mmu_notifier_ops;
14 /* mmu_notifier_ops flags */
15 #define MMU_INVALIDATE_DOES_NOT_BLOCK (0x01)
17 #ifdef CONFIG_MMU_NOTIFIER
20 * The mmu notifier_mm structure is allocated and installed in
21 * mm->mmu_notifier_mm inside the mm_take_all_locks() protected
22 * critical section and it's released only when mm_count reaches zero
25 struct mmu_notifier_mm {
26 /* all mmu notifiers registerd in this mm are queued in this list */
27 struct hlist_head list;
28 /* to serialize the list modifications and hlist_unhashed */
32 struct mmu_notifier_ops {
34 * Flags to specify behavior of callbacks for this MMU notifier.
35 * Used to determine which context an operation may be called.
37 * MMU_INVALIDATE_DOES_NOT_BLOCK: invalidate_range_* callbacks do not
43 * Called either by mmu_notifier_unregister or when the mm is
44 * being destroyed by exit_mmap, always before all pages are
45 * freed. This can run concurrently with other mmu notifier
46 * methods (the ones invoked outside the mm context) and it
47 * should tear down all secondary mmu mappings and freeze the
48 * secondary mmu. If this method isn't implemented you've to
49 * be sure that nothing could possibly write to the pages
50 * through the secondary mmu by the time the last thread with
51 * tsk->mm == mm exits.
53 * As side note: the pages freed after ->release returns could
54 * be immediately reallocated by the gart at an alias physical
55 * address with a different cache model, so if ->release isn't
56 * implemented because all _software_ driven memory accesses
57 * through the secondary mmu are terminated by the time the
58 * last thread of this mm quits, you've also to be sure that
59 * speculative _hardware_ operations can't allocate dirty
60 * cachelines in the cpu that could not be snooped and made
61 * coherent with the other read and write operations happening
62 * through the gart alias address, so leading to memory
65 void (*release)(struct mmu_notifier *mn,
66 struct mm_struct *mm);
69 * clear_flush_young is called after the VM is
70 * test-and-clearing the young/accessed bitflag in the
71 * pte. This way the VM will provide proper aging to the
72 * accesses to the page through the secondary MMUs and not
73 * only to the ones through the Linux pte.
74 * Start-end is necessary in case the secondary MMU is mapping the page
75 * at a smaller granularity than the primary MMU.
77 int (*clear_flush_young)(struct mmu_notifier *mn,
83 * clear_young is a lightweight version of clear_flush_young. Like the
84 * latter, it is supposed to test-and-clear the young/accessed bitflag
85 * in the secondary pte, but it may omit flushing the secondary tlb.
87 int (*clear_young)(struct mmu_notifier *mn,
93 * test_young is called to check the young/accessed bitflag in
94 * the secondary pte. This is used to know if the page is
95 * frequently used without actually clearing the flag or tearing
96 * down the secondary mapping on the page.
98 int (*test_young)(struct mmu_notifier *mn,
100 unsigned long address);
103 * change_pte is called in cases that pte mapping to page is changed:
104 * for example, when ksm remaps pte to point to a new shared page.
106 void (*change_pte)(struct mmu_notifier *mn,
107 struct mm_struct *mm,
108 unsigned long address,
112 * invalidate_range_start() and invalidate_range_end() must be
113 * paired and are called only when the mmap_sem and/or the
114 * locks protecting the reverse maps are held. If the subsystem
115 * can't guarantee that no additional references are taken to
116 * the pages in the range, it has to implement the
117 * invalidate_range() notifier to remove any references taken
118 * after invalidate_range_start().
120 * Invalidation of multiple concurrent ranges may be
121 * optionally permitted by the driver. Either way the
122 * establishment of sptes is forbidden in the range passed to
123 * invalidate_range_begin/end for the whole duration of the
124 * invalidate_range_begin/end critical section.
126 * invalidate_range_start() is called when all pages in the
127 * range are still mapped and have at least a refcount of one.
129 * invalidate_range_end() is called when all pages in the
130 * range have been unmapped and the pages have been freed by
133 * The VM will remove the page table entries and potentially
134 * the page between invalidate_range_start() and
135 * invalidate_range_end(). If the page must not be freed
136 * because of pending I/O or other circumstances then the
137 * invalidate_range_start() callback (or the initial mapping
138 * by the driver) must make sure that the refcount is kept
141 * If the driver increases the refcount when the pages are
142 * initially mapped into an address space then either
143 * invalidate_range_start() or invalidate_range_end() may
144 * decrease the refcount. If the refcount is decreased on
145 * invalidate_range_start() then the VM can free pages as page
146 * table entries are removed. If the refcount is only
147 * droppped on invalidate_range_end() then the driver itself
148 * will drop the last refcount but it must take care to flush
149 * any secondary tlb before doing the final free on the
150 * page. Pages will no longer be referenced by the linux
151 * address space but may still be referenced by sptes until
152 * the last refcount is dropped.
154 * If blockable argument is set to false then the callback cannot
155 * sleep and has to return with -EAGAIN. 0 should be returned
156 * otherwise. Please note that if invalidate_range_start approves
157 * a non-blocking behavior then the same applies to
158 * invalidate_range_end.
161 int (*invalidate_range_start)(struct mmu_notifier *mn,
162 struct mm_struct *mm,
163 unsigned long start, unsigned long end,
165 void (*invalidate_range_end)(struct mmu_notifier *mn,
166 struct mm_struct *mm,
167 unsigned long start, unsigned long end);
170 * invalidate_range() is either called between
171 * invalidate_range_start() and invalidate_range_end() when the
172 * VM has to free pages that where unmapped, but before the
173 * pages are actually freed, or outside of _start()/_end() when
174 * a (remote) TLB is necessary.
176 * If invalidate_range() is used to manage a non-CPU TLB with
177 * shared page-tables, it not necessary to implement the
178 * invalidate_range_start()/end() notifiers, as
179 * invalidate_range() alread catches the points in time when an
180 * external TLB range needs to be flushed. For more in depth
181 * discussion on this see Documentation/vm/mmu_notifier.rst
183 * Note that this function might be called with just a sub-range
184 * of what was passed to invalidate_range_start()/end(), if
185 * called between those functions.
187 * If this callback cannot block, and invalidate_range_{start,end}
188 * cannot block, mmu_notifier_ops.flags should have
189 * MMU_INVALIDATE_DOES_NOT_BLOCK set.
191 void (*invalidate_range)(struct mmu_notifier *mn, struct mm_struct *mm,
192 unsigned long start, unsigned long end);
196 * The notifier chains are protected by mmap_sem and/or the reverse map
197 * semaphores. Notifier chains are only changed when all reverse maps and
198 * the mmap_sem locks are taken.
200 * Therefore notifier chains can only be traversed when either
202 * 1. mmap_sem is held.
203 * 2. One of the reverse map locks is held (i_mmap_rwsem or anon_vma->rwsem).
204 * 3. No other concurrent thread can access the list (release)
206 struct mmu_notifier {
207 struct hlist_node hlist;
208 const struct mmu_notifier_ops *ops;
211 static inline int mm_has_notifiers(struct mm_struct *mm)
213 return unlikely(mm->mmu_notifier_mm);
216 extern int mmu_notifier_register(struct mmu_notifier *mn,
217 struct mm_struct *mm);
218 extern int __mmu_notifier_register(struct mmu_notifier *mn,
219 struct mm_struct *mm);
220 extern void mmu_notifier_unregister(struct mmu_notifier *mn,
221 struct mm_struct *mm);
222 extern void mmu_notifier_unregister_no_release(struct mmu_notifier *mn,
223 struct mm_struct *mm);
224 extern void __mmu_notifier_mm_destroy(struct mm_struct *mm);
225 extern void __mmu_notifier_release(struct mm_struct *mm);
226 extern int __mmu_notifier_clear_flush_young(struct mm_struct *mm,
229 extern int __mmu_notifier_clear_young(struct mm_struct *mm,
232 extern int __mmu_notifier_test_young(struct mm_struct *mm,
233 unsigned long address);
234 extern void __mmu_notifier_change_pte(struct mm_struct *mm,
235 unsigned long address, pte_t pte);
236 extern int __mmu_notifier_invalidate_range_start(struct mm_struct *mm,
237 unsigned long start, unsigned long end,
239 extern void __mmu_notifier_invalidate_range_end(struct mm_struct *mm,
240 unsigned long start, unsigned long end,
242 extern void __mmu_notifier_invalidate_range(struct mm_struct *mm,
243 unsigned long start, unsigned long end);
244 extern bool mm_has_blockable_invalidate_notifiers(struct mm_struct *mm);
246 static inline void mmu_notifier_release(struct mm_struct *mm)
248 if (mm_has_notifiers(mm))
249 __mmu_notifier_release(mm);
252 static inline int mmu_notifier_clear_flush_young(struct mm_struct *mm,
256 if (mm_has_notifiers(mm))
257 return __mmu_notifier_clear_flush_young(mm, start, end);
261 static inline int mmu_notifier_clear_young(struct mm_struct *mm,
265 if (mm_has_notifiers(mm))
266 return __mmu_notifier_clear_young(mm, start, end);
270 static inline int mmu_notifier_test_young(struct mm_struct *mm,
271 unsigned long address)
273 if (mm_has_notifiers(mm))
274 return __mmu_notifier_test_young(mm, address);
278 static inline void mmu_notifier_change_pte(struct mm_struct *mm,
279 unsigned long address, pte_t pte)
281 if (mm_has_notifiers(mm))
282 __mmu_notifier_change_pte(mm, address, pte);
285 static inline void mmu_notifier_invalidate_range_start(struct mm_struct *mm,
286 unsigned long start, unsigned long end)
288 if (mm_has_notifiers(mm))
289 __mmu_notifier_invalidate_range_start(mm, start, end, true);
292 static inline int mmu_notifier_invalidate_range_start_nonblock(struct mm_struct *mm,
293 unsigned long start, unsigned long end)
295 if (mm_has_notifiers(mm))
296 return __mmu_notifier_invalidate_range_start(mm, start, end, false);
300 static inline void mmu_notifier_invalidate_range_end(struct mm_struct *mm,
301 unsigned long start, unsigned long end)
303 if (mm_has_notifiers(mm))
304 __mmu_notifier_invalidate_range_end(mm, start, end, false);
307 static inline void mmu_notifier_invalidate_range_only_end(struct mm_struct *mm,
308 unsigned long start, unsigned long end)
310 if (mm_has_notifiers(mm))
311 __mmu_notifier_invalidate_range_end(mm, start, end, true);
314 static inline void mmu_notifier_invalidate_range(struct mm_struct *mm,
315 unsigned long start, unsigned long end)
317 if (mm_has_notifiers(mm))
318 __mmu_notifier_invalidate_range(mm, start, end);
321 static inline void mmu_notifier_mm_init(struct mm_struct *mm)
323 mm->mmu_notifier_mm = NULL;
326 static inline void mmu_notifier_mm_destroy(struct mm_struct *mm)
328 if (mm_has_notifiers(mm))
329 __mmu_notifier_mm_destroy(mm);
332 #define ptep_clear_flush_young_notify(__vma, __address, __ptep) \
335 struct vm_area_struct *___vma = __vma; \
336 unsigned long ___address = __address; \
337 __young = ptep_clear_flush_young(___vma, ___address, __ptep); \
338 __young |= mmu_notifier_clear_flush_young(___vma->vm_mm, \
345 #define pmdp_clear_flush_young_notify(__vma, __address, __pmdp) \
348 struct vm_area_struct *___vma = __vma; \
349 unsigned long ___address = __address; \
350 __young = pmdp_clear_flush_young(___vma, ___address, __pmdp); \
351 __young |= mmu_notifier_clear_flush_young(___vma->vm_mm, \
358 #define ptep_clear_young_notify(__vma, __address, __ptep) \
361 struct vm_area_struct *___vma = __vma; \
362 unsigned long ___address = __address; \
363 __young = ptep_test_and_clear_young(___vma, ___address, __ptep);\
364 __young |= mmu_notifier_clear_young(___vma->vm_mm, ___address, \
365 ___address + PAGE_SIZE); \
369 #define pmdp_clear_young_notify(__vma, __address, __pmdp) \
372 struct vm_area_struct *___vma = __vma; \
373 unsigned long ___address = __address; \
374 __young = pmdp_test_and_clear_young(___vma, ___address, __pmdp);\
375 __young |= mmu_notifier_clear_young(___vma->vm_mm, ___address, \
376 ___address + PMD_SIZE); \
380 #define ptep_clear_flush_notify(__vma, __address, __ptep) \
382 unsigned long ___addr = __address & PAGE_MASK; \
383 struct mm_struct *___mm = (__vma)->vm_mm; \
386 ___pte = ptep_clear_flush(__vma, __address, __ptep); \
387 mmu_notifier_invalidate_range(___mm, ___addr, \
388 ___addr + PAGE_SIZE); \
393 #define pmdp_huge_clear_flush_notify(__vma, __haddr, __pmd) \
395 unsigned long ___haddr = __haddr & HPAGE_PMD_MASK; \
396 struct mm_struct *___mm = (__vma)->vm_mm; \
399 ___pmd = pmdp_huge_clear_flush(__vma, __haddr, __pmd); \
400 mmu_notifier_invalidate_range(___mm, ___haddr, \
401 ___haddr + HPAGE_PMD_SIZE); \
406 #define pudp_huge_clear_flush_notify(__vma, __haddr, __pud) \
408 unsigned long ___haddr = __haddr & HPAGE_PUD_MASK; \
409 struct mm_struct *___mm = (__vma)->vm_mm; \
412 ___pud = pudp_huge_clear_flush(__vma, __haddr, __pud); \
413 mmu_notifier_invalidate_range(___mm, ___haddr, \
414 ___haddr + HPAGE_PUD_SIZE); \
420 * set_pte_at_notify() sets the pte _after_ running the notifier.
421 * This is safe to start by updating the secondary MMUs, because the primary MMU
422 * pte invalidate must have already happened with a ptep_clear_flush() before
423 * set_pte_at_notify() has been invoked. Updating the secondary MMUs first is
424 * required when we change both the protection of the mapping from read-only to
425 * read-write and the pfn (like during copy on write page faults). Otherwise the
426 * old page would remain mapped readonly in the secondary MMUs after the new
427 * page is already writable by some CPU through the primary MMU.
429 #define set_pte_at_notify(__mm, __address, __ptep, __pte) \
431 struct mm_struct *___mm = __mm; \
432 unsigned long ___address = __address; \
433 pte_t ___pte = __pte; \
435 mmu_notifier_change_pte(___mm, ___address, ___pte); \
436 set_pte_at(___mm, ___address, __ptep, ___pte); \
439 extern void mmu_notifier_call_srcu(struct rcu_head *rcu,
440 void (*func)(struct rcu_head *rcu));
441 extern void mmu_notifier_synchronize(void);
443 #else /* CONFIG_MMU_NOTIFIER */
445 static inline int mm_has_notifiers(struct mm_struct *mm)
450 static inline void mmu_notifier_release(struct mm_struct *mm)
454 static inline int mmu_notifier_clear_flush_young(struct mm_struct *mm,
461 static inline int mmu_notifier_test_young(struct mm_struct *mm,
462 unsigned long address)
467 static inline void mmu_notifier_change_pte(struct mm_struct *mm,
468 unsigned long address, pte_t pte)
472 static inline void mmu_notifier_invalidate_range_start(struct mm_struct *mm,
473 unsigned long start, unsigned long end)
477 static inline int mmu_notifier_invalidate_range_start_nonblock(struct mm_struct *mm,
478 unsigned long start, unsigned long end)
483 static inline void mmu_notifier_invalidate_range_end(struct mm_struct *mm,
484 unsigned long start, unsigned long end)
488 static inline void mmu_notifier_invalidate_range_only_end(struct mm_struct *mm,
489 unsigned long start, unsigned long end)
493 static inline void mmu_notifier_invalidate_range(struct mm_struct *mm,
494 unsigned long start, unsigned long end)
498 static inline bool mm_has_blockable_invalidate_notifiers(struct mm_struct *mm)
503 static inline void mmu_notifier_mm_init(struct mm_struct *mm)
507 static inline void mmu_notifier_mm_destroy(struct mm_struct *mm)
511 #define ptep_clear_flush_young_notify ptep_clear_flush_young
512 #define pmdp_clear_flush_young_notify pmdp_clear_flush_young
513 #define ptep_clear_young_notify ptep_test_and_clear_young
514 #define pmdp_clear_young_notify pmdp_test_and_clear_young
515 #define ptep_clear_flush_notify ptep_clear_flush
516 #define pmdp_huge_clear_flush_notify pmdp_huge_clear_flush
517 #define pudp_huge_clear_flush_notify pudp_huge_clear_flush
518 #define set_pte_at_notify set_pte_at
520 #endif /* CONFIG_MMU_NOTIFIER */
522 #endif /* _LINUX_MMU_NOTIFIER_H */