blob: a1a210d59961a855964b03ee9a7eae7c74d7c86f [file] [log] [blame]
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001#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>
Sagi Grimberg21a92732012-10-08 16:29:24 -07007#include <linux/srcu.h>
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07008
9struct mmu_notifier;
10struct mmu_notifier_ops;
11
12#ifdef CONFIG_MMU_NOTIFIER
13
14/*
15 * The mmu notifier_mm structure is allocated and installed in
16 * mm->mmu_notifier_mm inside the mm_take_all_locks() protected
17 * critical section and it's released only when mm_count reaches zero
18 * in mmdrop().
19 */
20struct mmu_notifier_mm {
21 /* all mmu notifiers registerd in this mm are queued in this list */
22 struct hlist_head list;
23 /* to serialize the list modifications and hlist_unhashed */
24 spinlock_t lock;
25};
26
27struct mmu_notifier_ops {
28 /*
29 * Called either by mmu_notifier_unregister or when the mm is
30 * being destroyed by exit_mmap, always before all pages are
31 * freed. This can run concurrently with other mmu notifier
32 * methods (the ones invoked outside the mm context) and it
33 * should tear down all secondary mmu mappings and freeze the
34 * secondary mmu. If this method isn't implemented you've to
35 * be sure that nothing could possibly write to the pages
36 * through the secondary mmu by the time the last thread with
37 * tsk->mm == mm exits.
38 *
39 * As side note: the pages freed after ->release returns could
40 * be immediately reallocated by the gart at an alias physical
41 * address with a different cache model, so if ->release isn't
42 * implemented because all _software_ driven memory accesses
43 * through the secondary mmu are terminated by the time the
44 * last thread of this mm quits, you've also to be sure that
45 * speculative _hardware_ operations can't allocate dirty
46 * cachelines in the cpu that could not be snooped and made
47 * coherent with the other read and write operations happening
48 * through the gart alias address, so leading to memory
49 * corruption.
50 */
51 void (*release)(struct mmu_notifier *mn,
52 struct mm_struct *mm);
53
54 /*
55 * clear_flush_young is called after the VM is
56 * test-and-clearing the young/accessed bitflag in the
57 * pte. This way the VM will provide proper aging to the
58 * accesses to the page through the secondary MMUs and not
59 * only to the ones through the Linux pte.
Andres Lagar-Cavilla57128462014-09-22 14:54:42 -070060 * Start-end is necessary in case the secondary MMU is mapping the page
61 * at a smaller granularity than the primary MMU.
Andrea Arcangelicddb8a52008-07-28 15:46:29 -070062 */
63 int (*clear_flush_young)(struct mmu_notifier *mn,
64 struct mm_struct *mm,
Andres Lagar-Cavilla57128462014-09-22 14:54:42 -070065 unsigned long start,
66 unsigned long end);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -070067
68 /*
Vladimir Davydov1d7715c2015-09-09 15:35:41 -070069 * clear_young is a lightweight version of clear_flush_young. Like the
70 * latter, it is supposed to test-and-clear the young/accessed bitflag
71 * in the secondary pte, but it may omit flushing the secondary tlb.
72 */
73 int (*clear_young)(struct mmu_notifier *mn,
74 struct mm_struct *mm,
75 unsigned long start,
76 unsigned long end);
77
78 /*
Andrea Arcangeli8ee53822011-01-13 15:47:10 -080079 * test_young is called to check the young/accessed bitflag in
80 * the secondary pte. This is used to know if the page is
81 * frequently used without actually clearing the flag or tearing
82 * down the secondary mapping on the page.
83 */
84 int (*test_young)(struct mmu_notifier *mn,
85 struct mm_struct *mm,
86 unsigned long address);
87
88 /*
Izik Eidus828502d2009-09-21 17:01:51 -070089 * change_pte is called in cases that pte mapping to page is changed:
90 * for example, when ksm remaps pte to point to a new shared page.
91 */
92 void (*change_pte)(struct mmu_notifier *mn,
93 struct mm_struct *mm,
94 unsigned long address,
95 pte_t pte);
96
97 /*
Andrea Arcangelicddb8a52008-07-28 15:46:29 -070098 * Before this is invoked any secondary MMU is still ok to
99 * read/write to the page previously pointed to by the Linux
100 * pte because the page hasn't been freed yet and it won't be
101 * freed until this returns. If required set_page_dirty has to
102 * be called internally to this method.
103 */
104 void (*invalidate_page)(struct mmu_notifier *mn,
105 struct mm_struct *mm,
106 unsigned long address);
107
108 /*
109 * invalidate_range_start() and invalidate_range_end() must be
110 * paired and are called only when the mmap_sem and/or the
Joerg Roedel0f0a3272014-11-13 13:46:09 +1100111 * locks protecting the reverse maps are held. If the subsystem
112 * can't guarantee that no additional references are taken to
113 * the pages in the range, it has to implement the
114 * invalidate_range() notifier to remove any references taken
115 * after invalidate_range_start().
Andrea Arcangelicddb8a52008-07-28 15:46:29 -0700116 *
117 * Invalidation of multiple concurrent ranges may be
118 * optionally permitted by the driver. Either way the
119 * establishment of sptes is forbidden in the range passed to
120 * invalidate_range_begin/end for the whole duration of the
121 * invalidate_range_begin/end critical section.
122 *
123 * invalidate_range_start() is called when all pages in the
124 * range are still mapped and have at least a refcount of one.
125 *
126 * invalidate_range_end() is called when all pages in the
127 * range have been unmapped and the pages have been freed by
128 * the VM.
129 *
130 * The VM will remove the page table entries and potentially
131 * the page between invalidate_range_start() and
132 * invalidate_range_end(). If the page must not be freed
133 * because of pending I/O or other circumstances then the
134 * invalidate_range_start() callback (or the initial mapping
135 * by the driver) must make sure that the refcount is kept
136 * elevated.
137 *
138 * If the driver increases the refcount when the pages are
139 * initially mapped into an address space then either
140 * invalidate_range_start() or invalidate_range_end() may
141 * decrease the refcount. If the refcount is decreased on
142 * invalidate_range_start() then the VM can free pages as page
143 * table entries are removed. If the refcount is only
144 * droppped on invalidate_range_end() then the driver itself
145 * will drop the last refcount but it must take care to flush
146 * any secondary tlb before doing the final free on the
147 * page. Pages will no longer be referenced by the linux
148 * address space but may still be referenced by sptes until
149 * the last refcount is dropped.
150 */
151 void (*invalidate_range_start)(struct mmu_notifier *mn,
152 struct mm_struct *mm,
153 unsigned long start, unsigned long end);
154 void (*invalidate_range_end)(struct mmu_notifier *mn,
155 struct mm_struct *mm,
156 unsigned long start, unsigned long end);
Joerg Roedel0f0a3272014-11-13 13:46:09 +1100157
158 /*
159 * invalidate_range() is either called between
160 * invalidate_range_start() and invalidate_range_end() when the
161 * VM has to free pages that where unmapped, but before the
162 * pages are actually freed, or outside of _start()/_end() when
163 * a (remote) TLB is necessary.
164 *
165 * If invalidate_range() is used to manage a non-CPU TLB with
166 * shared page-tables, it not necessary to implement the
167 * invalidate_range_start()/end() notifiers, as
168 * invalidate_range() alread catches the points in time when an
169 * external TLB range needs to be flushed.
170 *
171 * The invalidate_range() function is called under the ptl
172 * spin-lock and not allowed to sleep.
173 *
174 * Note that this function might be called with just a sub-range
175 * of what was passed to invalidate_range_start()/end(), if
176 * called between those functions.
177 */
178 void (*invalidate_range)(struct mmu_notifier *mn, struct mm_struct *mm,
179 unsigned long start, unsigned long end);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -0700180};
181
182/*
183 * The notifier chains are protected by mmap_sem and/or the reverse map
184 * semaphores. Notifier chains are only changed when all reverse maps and
185 * the mmap_sem locks are taken.
186 *
187 * Therefore notifier chains can only be traversed when either
188 *
189 * 1. mmap_sem is held.
Davidlohr Buesoc8c06ef2014-12-12 16:54:24 -0800190 * 2. One of the reverse map locks is held (i_mmap_rwsem or anon_vma->rwsem).
Andrea Arcangelicddb8a52008-07-28 15:46:29 -0700191 * 3. No other concurrent thread can access the list (release)
192 */
193struct mmu_notifier {
194 struct hlist_node hlist;
195 const struct mmu_notifier_ops *ops;
196};
197
198static inline int mm_has_notifiers(struct mm_struct *mm)
199{
200 return unlikely(mm->mmu_notifier_mm);
201}
202
203extern int mmu_notifier_register(struct mmu_notifier *mn,
204 struct mm_struct *mm);
205extern int __mmu_notifier_register(struct mmu_notifier *mn,
206 struct mm_struct *mm);
207extern void mmu_notifier_unregister(struct mmu_notifier *mn,
208 struct mm_struct *mm);
Peter Zijlstrab9722162014-08-06 16:08:20 -0700209extern void mmu_notifier_unregister_no_release(struct mmu_notifier *mn,
210 struct mm_struct *mm);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -0700211extern void __mmu_notifier_mm_destroy(struct mm_struct *mm);
212extern void __mmu_notifier_release(struct mm_struct *mm);
213extern int __mmu_notifier_clear_flush_young(struct mm_struct *mm,
Andres Lagar-Cavilla57128462014-09-22 14:54:42 -0700214 unsigned long start,
215 unsigned long end);
Vladimir Davydov1d7715c2015-09-09 15:35:41 -0700216extern int __mmu_notifier_clear_young(struct mm_struct *mm,
217 unsigned long start,
218 unsigned long end);
Andrea Arcangeli8ee53822011-01-13 15:47:10 -0800219extern int __mmu_notifier_test_young(struct mm_struct *mm,
220 unsigned long address);
Izik Eidus828502d2009-09-21 17:01:51 -0700221extern void __mmu_notifier_change_pte(struct mm_struct *mm,
222 unsigned long address, pte_t pte);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -0700223extern void __mmu_notifier_invalidate_page(struct mm_struct *mm,
224 unsigned long address);
225extern void __mmu_notifier_invalidate_range_start(struct mm_struct *mm,
226 unsigned long start, unsigned long end);
227extern void __mmu_notifier_invalidate_range_end(struct mm_struct *mm,
228 unsigned long start, unsigned long end);
Joerg Roedel0f0a3272014-11-13 13:46:09 +1100229extern void __mmu_notifier_invalidate_range(struct mm_struct *mm,
230 unsigned long start, unsigned long end);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -0700231
232static inline void mmu_notifier_release(struct mm_struct *mm)
233{
234 if (mm_has_notifiers(mm))
235 __mmu_notifier_release(mm);
236}
237
238static inline int mmu_notifier_clear_flush_young(struct mm_struct *mm,
Andres Lagar-Cavilla57128462014-09-22 14:54:42 -0700239 unsigned long start,
240 unsigned long end)
Andrea Arcangelicddb8a52008-07-28 15:46:29 -0700241{
242 if (mm_has_notifiers(mm))
Andres Lagar-Cavilla57128462014-09-22 14:54:42 -0700243 return __mmu_notifier_clear_flush_young(mm, start, end);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -0700244 return 0;
245}
246
Vladimir Davydov1d7715c2015-09-09 15:35:41 -0700247static inline int mmu_notifier_clear_young(struct mm_struct *mm,
248 unsigned long start,
249 unsigned long end)
250{
251 if (mm_has_notifiers(mm))
252 return __mmu_notifier_clear_young(mm, start, end);
253 return 0;
254}
255
Andrea Arcangeli8ee53822011-01-13 15:47:10 -0800256static inline int mmu_notifier_test_young(struct mm_struct *mm,
257 unsigned long address)
258{
259 if (mm_has_notifiers(mm))
260 return __mmu_notifier_test_young(mm, address);
261 return 0;
262}
263
Izik Eidus828502d2009-09-21 17:01:51 -0700264static inline void mmu_notifier_change_pte(struct mm_struct *mm,
265 unsigned long address, pte_t pte)
266{
267 if (mm_has_notifiers(mm))
268 __mmu_notifier_change_pte(mm, address, pte);
269}
270
Andrea Arcangelicddb8a52008-07-28 15:46:29 -0700271static inline void mmu_notifier_invalidate_page(struct mm_struct *mm,
272 unsigned long address)
273{
274 if (mm_has_notifiers(mm))
275 __mmu_notifier_invalidate_page(mm, address);
276}
277
278static inline void mmu_notifier_invalidate_range_start(struct mm_struct *mm,
279 unsigned long start, unsigned long end)
280{
281 if (mm_has_notifiers(mm))
282 __mmu_notifier_invalidate_range_start(mm, start, end);
283}
284
285static inline void mmu_notifier_invalidate_range_end(struct mm_struct *mm,
286 unsigned long start, unsigned long end)
287{
288 if (mm_has_notifiers(mm))
289 __mmu_notifier_invalidate_range_end(mm, start, end);
290}
291
Joerg Roedel1897bdc2014-11-13 13:46:09 +1100292static inline void mmu_notifier_invalidate_range(struct mm_struct *mm,
293 unsigned long start, unsigned long end)
294{
Joerg Roedel0f0a3272014-11-13 13:46:09 +1100295 if (mm_has_notifiers(mm))
296 __mmu_notifier_invalidate_range(mm, start, end);
Joerg Roedel1897bdc2014-11-13 13:46:09 +1100297}
298
Andrea Arcangelicddb8a52008-07-28 15:46:29 -0700299static inline void mmu_notifier_mm_init(struct mm_struct *mm)
300{
301 mm->mmu_notifier_mm = NULL;
302}
303
304static inline void mmu_notifier_mm_destroy(struct mm_struct *mm)
305{
306 if (mm_has_notifiers(mm))
307 __mmu_notifier_mm_destroy(mm);
308}
309
Andrea Arcangelicddb8a52008-07-28 15:46:29 -0700310#define ptep_clear_flush_young_notify(__vma, __address, __ptep) \
311({ \
312 int __young; \
313 struct vm_area_struct *___vma = __vma; \
314 unsigned long ___address = __address; \
315 __young = ptep_clear_flush_young(___vma, ___address, __ptep); \
316 __young |= mmu_notifier_clear_flush_young(___vma->vm_mm, \
Andres Lagar-Cavilla57128462014-09-22 14:54:42 -0700317 ___address, \
318 ___address + \
319 PAGE_SIZE); \
Andrea Arcangelicddb8a52008-07-28 15:46:29 -0700320 __young; \
321})
322
Andrea Arcangeli91a4ee22011-01-13 15:46:44 -0800323#define pmdp_clear_flush_young_notify(__vma, __address, __pmdp) \
324({ \
325 int __young; \
326 struct vm_area_struct *___vma = __vma; \
327 unsigned long ___address = __address; \
328 __young = pmdp_clear_flush_young(___vma, ___address, __pmdp); \
329 __young |= mmu_notifier_clear_flush_young(___vma->vm_mm, \
Andres Lagar-Cavilla57128462014-09-22 14:54:42 -0700330 ___address, \
331 ___address + \
332 PMD_SIZE); \
Andrea Arcangeli91a4ee22011-01-13 15:46:44 -0800333 __young; \
334})
335
Vladimir Davydov1d7715c2015-09-09 15:35:41 -0700336#define ptep_clear_young_notify(__vma, __address, __ptep) \
337({ \
338 int __young; \
339 struct vm_area_struct *___vma = __vma; \
340 unsigned long ___address = __address; \
341 __young = ptep_test_and_clear_young(___vma, ___address, __ptep);\
342 __young |= mmu_notifier_clear_young(___vma->vm_mm, ___address, \
343 ___address + PAGE_SIZE); \
344 __young; \
345})
346
347#define pmdp_clear_young_notify(__vma, __address, __pmdp) \
348({ \
349 int __young; \
350 struct vm_area_struct *___vma = __vma; \
351 unsigned long ___address = __address; \
352 __young = pmdp_test_and_clear_young(___vma, ___address, __pmdp);\
353 __young |= mmu_notifier_clear_young(___vma->vm_mm, ___address, \
354 ___address + PMD_SIZE); \
355 __young; \
356})
357
Joerg Roedel34ee6452014-11-13 13:46:09 +1100358#define ptep_clear_flush_notify(__vma, __address, __ptep) \
359({ \
360 unsigned long ___addr = __address & PAGE_MASK; \
361 struct mm_struct *___mm = (__vma)->vm_mm; \
362 pte_t ___pte; \
363 \
364 ___pte = ptep_clear_flush(__vma, __address, __ptep); \
365 mmu_notifier_invalidate_range(___mm, ___addr, \
366 ___addr + PAGE_SIZE); \
367 \
368 ___pte; \
369})
370
Aneesh Kumar K.V8809aa22015-06-24 16:57:44 -0700371#define pmdp_huge_clear_flush_notify(__vma, __haddr, __pmd) \
Joerg Roedel34ee6452014-11-13 13:46:09 +1100372({ \
373 unsigned long ___haddr = __haddr & HPAGE_PMD_MASK; \
374 struct mm_struct *___mm = (__vma)->vm_mm; \
375 pmd_t ___pmd; \
376 \
Aneesh Kumar K.V8809aa22015-06-24 16:57:44 -0700377 ___pmd = pmdp_huge_clear_flush(__vma, __haddr, __pmd); \
Joerg Roedel34ee6452014-11-13 13:46:09 +1100378 mmu_notifier_invalidate_range(___mm, ___haddr, \
379 ___haddr + HPAGE_PMD_SIZE); \
380 \
381 ___pmd; \
382})
383
Aneesh Kumar K.V8809aa22015-06-24 16:57:44 -0700384#define pmdp_huge_get_and_clear_notify(__mm, __haddr, __pmd) \
Joerg Roedel34ee6452014-11-13 13:46:09 +1100385({ \
386 unsigned long ___haddr = __haddr & HPAGE_PMD_MASK; \
387 pmd_t ___pmd; \
388 \
Aneesh Kumar K.V8809aa22015-06-24 16:57:44 -0700389 ___pmd = pmdp_huge_get_and_clear(__mm, __haddr, __pmd); \
Joerg Roedel34ee6452014-11-13 13:46:09 +1100390 mmu_notifier_invalidate_range(__mm, ___haddr, \
391 ___haddr + HPAGE_PMD_SIZE); \
392 \
393 ___pmd; \
394})
395
Xiao Guangrong48af0d72012-10-08 16:29:23 -0700396/*
397 * set_pte_at_notify() sets the pte _after_ running the notifier.
398 * This is safe to start by updating the secondary MMUs, because the primary MMU
399 * pte invalidate must have already happened with a ptep_clear_flush() before
400 * set_pte_at_notify() has been invoked. Updating the secondary MMUs first is
401 * required when we change both the protection of the mapping from read-only to
402 * read-write and the pfn (like during copy on write page faults). Otherwise the
403 * old page would remain mapped readonly in the secondary MMUs after the new
404 * page is already writable by some CPU through the primary MMU.
405 */
Izik Eidus828502d2009-09-21 17:01:51 -0700406#define set_pte_at_notify(__mm, __address, __ptep, __pte) \
407({ \
408 struct mm_struct *___mm = __mm; \
409 unsigned long ___address = __address; \
410 pte_t ___pte = __pte; \
411 \
Izik Eidus828502d2009-09-21 17:01:51 -0700412 mmu_notifier_change_pte(___mm, ___address, ___pte); \
Xiao Guangrong48af0d72012-10-08 16:29:23 -0700413 set_pte_at(___mm, ___address, __ptep, ___pte); \
Izik Eidus828502d2009-09-21 17:01:51 -0700414})
415
Peter Zijlstrab9722162014-08-06 16:08:20 -0700416extern void mmu_notifier_call_srcu(struct rcu_head *rcu,
417 void (*func)(struct rcu_head *rcu));
418extern void mmu_notifier_synchronize(void);
419
Andrea Arcangelicddb8a52008-07-28 15:46:29 -0700420#else /* CONFIG_MMU_NOTIFIER */
421
422static inline void mmu_notifier_release(struct mm_struct *mm)
423{
424}
425
426static inline int mmu_notifier_clear_flush_young(struct mm_struct *mm,
Andres Lagar-Cavilla57128462014-09-22 14:54:42 -0700427 unsigned long start,
428 unsigned long end)
Andrea Arcangelicddb8a52008-07-28 15:46:29 -0700429{
430 return 0;
431}
432
Andrea Arcangeli8ee53822011-01-13 15:47:10 -0800433static inline int mmu_notifier_test_young(struct mm_struct *mm,
434 unsigned long address)
435{
436 return 0;
437}
438
Izik Eidus828502d2009-09-21 17:01:51 -0700439static inline void mmu_notifier_change_pte(struct mm_struct *mm,
440 unsigned long address, pte_t pte)
441{
442}
443
Andrea Arcangelicddb8a52008-07-28 15:46:29 -0700444static inline void mmu_notifier_invalidate_page(struct mm_struct *mm,
445 unsigned long address)
446{
447}
448
449static inline void mmu_notifier_invalidate_range_start(struct mm_struct *mm,
450 unsigned long start, unsigned long end)
451{
452}
453
454static inline void mmu_notifier_invalidate_range_end(struct mm_struct *mm,
455 unsigned long start, unsigned long end)
456{
457}
458
Joerg Roedel1897bdc2014-11-13 13:46:09 +1100459static inline void mmu_notifier_invalidate_range(struct mm_struct *mm,
460 unsigned long start, unsigned long end)
461{
462}
463
Andrea Arcangelicddb8a52008-07-28 15:46:29 -0700464static inline void mmu_notifier_mm_init(struct mm_struct *mm)
465{
466}
467
468static inline void mmu_notifier_mm_destroy(struct mm_struct *mm)
469{
470}
471
472#define ptep_clear_flush_young_notify ptep_clear_flush_young
Andrea Arcangeli91a4ee22011-01-13 15:46:44 -0800473#define pmdp_clear_flush_young_notify pmdp_clear_flush_young
Vladimir Davydov33c3fc72015-09-09 15:35:45 -0700474#define ptep_clear_young_notify ptep_test_and_clear_young
475#define pmdp_clear_young_notify pmdp_test_and_clear_young
Joerg Roedel34ee6452014-11-13 13:46:09 +1100476#define ptep_clear_flush_notify ptep_clear_flush
Aneesh Kumar K.V8809aa22015-06-24 16:57:44 -0700477#define pmdp_huge_clear_flush_notify pmdp_huge_clear_flush
478#define pmdp_huge_get_and_clear_notify pmdp_huge_get_and_clear
Izik Eidus828502d2009-09-21 17:01:51 -0700479#define set_pte_at_notify set_pte_at
Andrea Arcangelicddb8a52008-07-28 15:46:29 -0700480
481#endif /* CONFIG_MMU_NOTIFIER */
482
483#endif /* _LINUX_MMU_NOTIFIER_H */