blob: b77486d152cd307228dae87523b884b80f300111 [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>
7
8struct mmu_notifier;
9struct mmu_notifier_ops;
10
11#ifdef CONFIG_MMU_NOTIFIER
12
13/*
14 * The mmu notifier_mm structure is allocated and installed in
15 * mm->mmu_notifier_mm inside the mm_take_all_locks() protected
16 * critical section and it's released only when mm_count reaches zero
17 * in mmdrop().
18 */
19struct mmu_notifier_mm {
20 /* all mmu notifiers registerd in this mm are queued in this list */
21 struct hlist_head list;
22 /* to serialize the list modifications and hlist_unhashed */
23 spinlock_t lock;
24};
25
26struct mmu_notifier_ops {
27 /*
28 * Called either by mmu_notifier_unregister or when the mm is
29 * being destroyed by exit_mmap, always before all pages are
30 * freed. This can run concurrently with other mmu notifier
31 * methods (the ones invoked outside the mm context) and it
32 * should tear down all secondary mmu mappings and freeze the
33 * secondary mmu. If this method isn't implemented you've to
34 * be sure that nothing could possibly write to the pages
35 * through the secondary mmu by the time the last thread with
36 * tsk->mm == mm exits.
37 *
38 * As side note: the pages freed after ->release returns could
39 * be immediately reallocated by the gart at an alias physical
40 * address with a different cache model, so if ->release isn't
41 * implemented because all _software_ driven memory accesses
42 * through the secondary mmu are terminated by the time the
43 * last thread of this mm quits, you've also to be sure that
44 * speculative _hardware_ operations can't allocate dirty
45 * cachelines in the cpu that could not be snooped and made
46 * coherent with the other read and write operations happening
47 * through the gart alias address, so leading to memory
48 * corruption.
49 */
50 void (*release)(struct mmu_notifier *mn,
51 struct mm_struct *mm);
52
53 /*
54 * clear_flush_young is called after the VM is
55 * test-and-clearing the young/accessed bitflag in the
56 * pte. This way the VM will provide proper aging to the
57 * accesses to the page through the secondary MMUs and not
58 * only to the ones through the Linux pte.
59 */
60 int (*clear_flush_young)(struct mmu_notifier *mn,
61 struct mm_struct *mm,
62 unsigned long address);
63
64 /*
65 * Before this is invoked any secondary MMU is still ok to
66 * read/write to the page previously pointed to by the Linux
67 * pte because the page hasn't been freed yet and it won't be
68 * freed until this returns. If required set_page_dirty has to
69 * be called internally to this method.
70 */
71 void (*invalidate_page)(struct mmu_notifier *mn,
72 struct mm_struct *mm,
73 unsigned long address);
74
75 /*
76 * invalidate_range_start() and invalidate_range_end() must be
77 * paired and are called only when the mmap_sem and/or the
78 * locks protecting the reverse maps are held. The subsystem
79 * must guarantee that no additional references are taken to
80 * the pages in the range established between the call to
81 * invalidate_range_start() and the matching call to
82 * invalidate_range_end().
83 *
84 * Invalidation of multiple concurrent ranges may be
85 * optionally permitted by the driver. Either way the
86 * establishment of sptes is forbidden in the range passed to
87 * invalidate_range_begin/end for the whole duration of the
88 * invalidate_range_begin/end critical section.
89 *
90 * invalidate_range_start() is called when all pages in the
91 * range are still mapped and have at least a refcount of one.
92 *
93 * invalidate_range_end() is called when all pages in the
94 * range have been unmapped and the pages have been freed by
95 * the VM.
96 *
97 * The VM will remove the page table entries and potentially
98 * the page between invalidate_range_start() and
99 * invalidate_range_end(). If the page must not be freed
100 * because of pending I/O or other circumstances then the
101 * invalidate_range_start() callback (or the initial mapping
102 * by the driver) must make sure that the refcount is kept
103 * elevated.
104 *
105 * If the driver increases the refcount when the pages are
106 * initially mapped into an address space then either
107 * invalidate_range_start() or invalidate_range_end() may
108 * decrease the refcount. If the refcount is decreased on
109 * invalidate_range_start() then the VM can free pages as page
110 * table entries are removed. If the refcount is only
111 * droppped on invalidate_range_end() then the driver itself
112 * will drop the last refcount but it must take care to flush
113 * any secondary tlb before doing the final free on the
114 * page. Pages will no longer be referenced by the linux
115 * address space but may still be referenced by sptes until
116 * the last refcount is dropped.
117 */
118 void (*invalidate_range_start)(struct mmu_notifier *mn,
119 struct mm_struct *mm,
120 unsigned long start, unsigned long end);
121 void (*invalidate_range_end)(struct mmu_notifier *mn,
122 struct mm_struct *mm,
123 unsigned long start, unsigned long end);
124};
125
126/*
127 * The notifier chains are protected by mmap_sem and/or the reverse map
128 * semaphores. Notifier chains are only changed when all reverse maps and
129 * the mmap_sem locks are taken.
130 *
131 * Therefore notifier chains can only be traversed when either
132 *
133 * 1. mmap_sem is held.
134 * 2. One of the reverse map locks is held (i_mmap_lock or anon_vma->lock).
135 * 3. No other concurrent thread can access the list (release)
136 */
137struct mmu_notifier {
138 struct hlist_node hlist;
139 const struct mmu_notifier_ops *ops;
140};
141
142static inline int mm_has_notifiers(struct mm_struct *mm)
143{
144 return unlikely(mm->mmu_notifier_mm);
145}
146
147extern int mmu_notifier_register(struct mmu_notifier *mn,
148 struct mm_struct *mm);
149extern int __mmu_notifier_register(struct mmu_notifier *mn,
150 struct mm_struct *mm);
151extern void mmu_notifier_unregister(struct mmu_notifier *mn,
152 struct mm_struct *mm);
153extern void __mmu_notifier_mm_destroy(struct mm_struct *mm);
154extern void __mmu_notifier_release(struct mm_struct *mm);
155extern int __mmu_notifier_clear_flush_young(struct mm_struct *mm,
156 unsigned long address);
157extern void __mmu_notifier_invalidate_page(struct mm_struct *mm,
158 unsigned long address);
159extern void __mmu_notifier_invalidate_range_start(struct mm_struct *mm,
160 unsigned long start, unsigned long end);
161extern void __mmu_notifier_invalidate_range_end(struct mm_struct *mm,
162 unsigned long start, unsigned long end);
163
164static inline void mmu_notifier_release(struct mm_struct *mm)
165{
166 if (mm_has_notifiers(mm))
167 __mmu_notifier_release(mm);
168}
169
170static inline int mmu_notifier_clear_flush_young(struct mm_struct *mm,
171 unsigned long address)
172{
173 if (mm_has_notifiers(mm))
174 return __mmu_notifier_clear_flush_young(mm, address);
175 return 0;
176}
177
178static inline void mmu_notifier_invalidate_page(struct mm_struct *mm,
179 unsigned long address)
180{
181 if (mm_has_notifiers(mm))
182 __mmu_notifier_invalidate_page(mm, address);
183}
184
185static inline void mmu_notifier_invalidate_range_start(struct mm_struct *mm,
186 unsigned long start, unsigned long end)
187{
188 if (mm_has_notifiers(mm))
189 __mmu_notifier_invalidate_range_start(mm, start, end);
190}
191
192static inline void mmu_notifier_invalidate_range_end(struct mm_struct *mm,
193 unsigned long start, unsigned long end)
194{
195 if (mm_has_notifiers(mm))
196 __mmu_notifier_invalidate_range_end(mm, start, end);
197}
198
199static inline void mmu_notifier_mm_init(struct mm_struct *mm)
200{
201 mm->mmu_notifier_mm = NULL;
202}
203
204static inline void mmu_notifier_mm_destroy(struct mm_struct *mm)
205{
206 if (mm_has_notifiers(mm))
207 __mmu_notifier_mm_destroy(mm);
208}
209
210/*
211 * These two macros will sometime replace ptep_clear_flush.
212 * ptep_clear_flush is impleemnted as macro itself, so this also is
213 * implemented as a macro until ptep_clear_flush will converted to an
214 * inline function, to diminish the risk of compilation failure. The
215 * invalidate_page method over time can be moved outside the PT lock
216 * and these two macros can be later removed.
217 */
218#define ptep_clear_flush_notify(__vma, __address, __ptep) \
219({ \
220 pte_t __pte; \
221 struct vm_area_struct *___vma = __vma; \
222 unsigned long ___address = __address; \
223 __pte = ptep_clear_flush(___vma, ___address, __ptep); \
224 mmu_notifier_invalidate_page(___vma->vm_mm, ___address); \
225 __pte; \
226})
227
228#define ptep_clear_flush_young_notify(__vma, __address, __ptep) \
229({ \
230 int __young; \
231 struct vm_area_struct *___vma = __vma; \
232 unsigned long ___address = __address; \
233 __young = ptep_clear_flush_young(___vma, ___address, __ptep); \
234 __young |= mmu_notifier_clear_flush_young(___vma->vm_mm, \
235 ___address); \
236 __young; \
237})
238
239#else /* CONFIG_MMU_NOTIFIER */
240
241static inline void mmu_notifier_release(struct mm_struct *mm)
242{
243}
244
245static inline int mmu_notifier_clear_flush_young(struct mm_struct *mm,
246 unsigned long address)
247{
248 return 0;
249}
250
251static inline void mmu_notifier_invalidate_page(struct mm_struct *mm,
252 unsigned long address)
253{
254}
255
256static inline void mmu_notifier_invalidate_range_start(struct mm_struct *mm,
257 unsigned long start, unsigned long end)
258{
259}
260
261static inline void mmu_notifier_invalidate_range_end(struct mm_struct *mm,
262 unsigned long start, unsigned long end)
263{
264}
265
266static inline void mmu_notifier_mm_init(struct mm_struct *mm)
267{
268}
269
270static inline void mmu_notifier_mm_destroy(struct mm_struct *mm)
271{
272}
273
274#define ptep_clear_flush_young_notify ptep_clear_flush_young
275#define ptep_clear_flush_notify ptep_clear_flush
276
277#endif /* CONFIG_MMU_NOTIFIER */
278
279#endif /* _LINUX_MMU_NOTIFIER_H */