]> git.karo-electronics.de Git - karo-tx-linux.git/blob - include/linux/seqlock.h
sched/numa: Skip some page migrations after a shared fault
[karo-tx-linux.git] / include / linux / seqlock.h
1 #ifndef __LINUX_SEQLOCK_H
2 #define __LINUX_SEQLOCK_H
3 /*
4  * Reader/writer consistent mechanism without starving writers. This type of
5  * lock for data where the reader wants a consistent set of information
6  * and is willing to retry if the information changes. There are two types
7  * of readers:
8  * 1. Sequence readers which never block a writer but they may have to retry
9  *    if a writer is in progress by detecting change in sequence number.
10  *    Writers do not wait for a sequence reader.
11  * 2. Locking readers which will wait if a writer or another locking reader
12  *    is in progress. A locking reader in progress will also block a writer
13  *    from going forward. Unlike the regular rwlock, the read lock here is
14  *    exclusive so that only one locking reader can get it.
15  *
16  * This is not as cache friendly as brlock. Also, this may not work well
17  * for data that contains pointers, because any writer could
18  * invalidate a pointer that a reader was following.
19  *
20  * Expected non-blocking reader usage:
21  *      do {
22  *          seq = read_seqbegin(&foo);
23  *      ...
24  *      } while (read_seqretry(&foo, seq));
25  *
26  *
27  * On non-SMP the spin locks disappear but the writer still needs
28  * to increment the sequence variables because an interrupt routine could
29  * change the state of the data.
30  *
31  * Based on x86_64 vsyscall gettimeofday 
32  * by Keith Owens and Andrea Arcangeli
33  */
34
35 #include <linux/spinlock.h>
36 #include <linux/preempt.h>
37 #include <asm/processor.h>
38
39 /*
40  * Version using sequence counter only.
41  * This can be used when code has its own mutex protecting the
42  * updating starting before the write_seqcountbeqin() and ending
43  * after the write_seqcount_end().
44  */
45 typedef struct seqcount {
46         unsigned sequence;
47 } seqcount_t;
48
49 #define SEQCNT_ZERO { 0 }
50 #define seqcount_init(x)        do { *(x) = (seqcount_t) SEQCNT_ZERO; } while (0)
51
52 /**
53  * __read_seqcount_begin - begin a seq-read critical section (without barrier)
54  * @s: pointer to seqcount_t
55  * Returns: count to be passed to read_seqcount_retry
56  *
57  * __read_seqcount_begin is like read_seqcount_begin, but has no smp_rmb()
58  * barrier. Callers should ensure that smp_rmb() or equivalent ordering is
59  * provided before actually loading any of the variables that are to be
60  * protected in this critical section.
61  *
62  * Use carefully, only in critical code, and comment how the barrier is
63  * provided.
64  */
65 static inline unsigned __read_seqcount_begin(const seqcount_t *s)
66 {
67         unsigned ret;
68
69 repeat:
70         ret = ACCESS_ONCE(s->sequence);
71         if (unlikely(ret & 1)) {
72                 cpu_relax();
73                 goto repeat;
74         }
75         return ret;
76 }
77
78 /**
79  * read_seqcount_begin - begin a seq-read critical section
80  * @s: pointer to seqcount_t
81  * Returns: count to be passed to read_seqcount_retry
82  *
83  * read_seqcount_begin opens a read critical section of the given seqcount.
84  * Validity of the critical section is tested by checking read_seqcount_retry
85  * function.
86  */
87 static inline unsigned read_seqcount_begin(const seqcount_t *s)
88 {
89         unsigned ret = __read_seqcount_begin(s);
90         smp_rmb();
91         return ret;
92 }
93
94 /**
95  * raw_seqcount_begin - begin a seq-read critical section
96  * @s: pointer to seqcount_t
97  * Returns: count to be passed to read_seqcount_retry
98  *
99  * raw_seqcount_begin opens a read critical section of the given seqcount.
100  * Validity of the critical section is tested by checking read_seqcount_retry
101  * function.
102  *
103  * Unlike read_seqcount_begin(), this function will not wait for the count
104  * to stabilize. If a writer is active when we begin, we will fail the
105  * read_seqcount_retry() instead of stabilizing at the beginning of the
106  * critical section.
107  */
108 static inline unsigned raw_seqcount_begin(const seqcount_t *s)
109 {
110         unsigned ret = ACCESS_ONCE(s->sequence);
111         smp_rmb();
112         return ret & ~1;
113 }
114
115 /**
116  * __read_seqcount_retry - end a seq-read critical section (without barrier)
117  * @s: pointer to seqcount_t
118  * @start: count, from read_seqcount_begin
119  * Returns: 1 if retry is required, else 0
120  *
121  * __read_seqcount_retry is like read_seqcount_retry, but has no smp_rmb()
122  * barrier. Callers should ensure that smp_rmb() or equivalent ordering is
123  * provided before actually loading any of the variables that are to be
124  * protected in this critical section.
125  *
126  * Use carefully, only in critical code, and comment how the barrier is
127  * provided.
128  */
129 static inline int __read_seqcount_retry(const seqcount_t *s, unsigned start)
130 {
131         return unlikely(s->sequence != start);
132 }
133
134 /**
135  * read_seqcount_retry - end a seq-read critical section
136  * @s: pointer to seqcount_t
137  * @start: count, from read_seqcount_begin
138  * Returns: 1 if retry is required, else 0
139  *
140  * read_seqcount_retry closes a read critical section of the given seqcount.
141  * If the critical section was invalid, it must be ignored (and typically
142  * retried).
143  */
144 static inline int read_seqcount_retry(const seqcount_t *s, unsigned start)
145 {
146         smp_rmb();
147         return __read_seqcount_retry(s, start);
148 }
149
150
151 /*
152  * Sequence counter only version assumes that callers are using their
153  * own mutexing.
154  */
155 static inline void write_seqcount_begin(seqcount_t *s)
156 {
157         s->sequence++;
158         smp_wmb();
159 }
160
161 static inline void write_seqcount_end(seqcount_t *s)
162 {
163         smp_wmb();
164         s->sequence++;
165 }
166
167 /**
168  * write_seqcount_barrier - invalidate in-progress read-side seq operations
169  * @s: pointer to seqcount_t
170  *
171  * After write_seqcount_barrier, no read-side seq operations will complete
172  * successfully and see data older than this.
173  */
174 static inline void write_seqcount_barrier(seqcount_t *s)
175 {
176         smp_wmb();
177         s->sequence+=2;
178 }
179
180 typedef struct {
181         struct seqcount seqcount;
182         spinlock_t lock;
183 } seqlock_t;
184
185 /*
186  * These macros triggered gcc-3.x compile-time problems.  We think these are
187  * OK now.  Be cautious.
188  */
189 #define __SEQLOCK_UNLOCKED(lockname)                    \
190         {                                               \
191                 .seqcount = SEQCNT_ZERO,                \
192                 .lock = __SPIN_LOCK_UNLOCKED(lockname)  \
193         }
194
195 #define seqlock_init(x)                                 \
196         do {                                            \
197                 seqcount_init(&(x)->seqcount);          \
198                 spin_lock_init(&(x)->lock);             \
199         } while (0)
200
201 #define DEFINE_SEQLOCK(x) \
202                 seqlock_t x = __SEQLOCK_UNLOCKED(x)
203
204 /*
205  * Read side functions for starting and finalizing a read side section.
206  */
207 static inline unsigned read_seqbegin(const seqlock_t *sl)
208 {
209         return read_seqcount_begin(&sl->seqcount);
210 }
211
212 static inline unsigned read_seqretry(const seqlock_t *sl, unsigned start)
213 {
214         return read_seqcount_retry(&sl->seqcount, start);
215 }
216
217 /*
218  * Lock out other writers and update the count.
219  * Acts like a normal spin_lock/unlock.
220  * Don't need preempt_disable() because that is in the spin_lock already.
221  */
222 static inline void write_seqlock(seqlock_t *sl)
223 {
224         spin_lock(&sl->lock);
225         write_seqcount_begin(&sl->seqcount);
226 }
227
228 static inline void write_sequnlock(seqlock_t *sl)
229 {
230         write_seqcount_end(&sl->seqcount);
231         spin_unlock(&sl->lock);
232 }
233
234 static inline void write_seqlock_bh(seqlock_t *sl)
235 {
236         spin_lock_bh(&sl->lock);
237         write_seqcount_begin(&sl->seqcount);
238 }
239
240 static inline void write_sequnlock_bh(seqlock_t *sl)
241 {
242         write_seqcount_end(&sl->seqcount);
243         spin_unlock_bh(&sl->lock);
244 }
245
246 static inline void write_seqlock_irq(seqlock_t *sl)
247 {
248         spin_lock_irq(&sl->lock);
249         write_seqcount_begin(&sl->seqcount);
250 }
251
252 static inline void write_sequnlock_irq(seqlock_t *sl)
253 {
254         write_seqcount_end(&sl->seqcount);
255         spin_unlock_irq(&sl->lock);
256 }
257
258 static inline unsigned long __write_seqlock_irqsave(seqlock_t *sl)
259 {
260         unsigned long flags;
261
262         spin_lock_irqsave(&sl->lock, flags);
263         write_seqcount_begin(&sl->seqcount);
264         return flags;
265 }
266
267 #define write_seqlock_irqsave(lock, flags)                              \
268         do { flags = __write_seqlock_irqsave(lock); } while (0)
269
270 static inline void
271 write_sequnlock_irqrestore(seqlock_t *sl, unsigned long flags)
272 {
273         write_seqcount_end(&sl->seqcount);
274         spin_unlock_irqrestore(&sl->lock, flags);
275 }
276
277 /*
278  * A locking reader exclusively locks out other writers and locking readers,
279  * but doesn't update the sequence number. Acts like a normal spin_lock/unlock.
280  * Don't need preempt_disable() because that is in the spin_lock already.
281  */
282 static inline void read_seqlock_excl(seqlock_t *sl)
283 {
284         spin_lock(&sl->lock);
285 }
286
287 static inline void read_sequnlock_excl(seqlock_t *sl)
288 {
289         spin_unlock(&sl->lock);
290 }
291
292 static inline void read_seqlock_excl_bh(seqlock_t *sl)
293 {
294         spin_lock_bh(&sl->lock);
295 }
296
297 static inline void read_sequnlock_excl_bh(seqlock_t *sl)
298 {
299         spin_unlock_bh(&sl->lock);
300 }
301
302 static inline void read_seqlock_excl_irq(seqlock_t *sl)
303 {
304         spin_lock_irq(&sl->lock);
305 }
306
307 static inline void read_sequnlock_excl_irq(seqlock_t *sl)
308 {
309         spin_unlock_irq(&sl->lock);
310 }
311
312 static inline unsigned long __read_seqlock_excl_irqsave(seqlock_t *sl)
313 {
314         unsigned long flags;
315
316         spin_lock_irqsave(&sl->lock, flags);
317         return flags;
318 }
319
320 #define read_seqlock_excl_irqsave(lock, flags)                          \
321         do { flags = __read_seqlock_excl_irqsave(lock); } while (0)
322
323 static inline void
324 read_sequnlock_excl_irqrestore(seqlock_t *sl, unsigned long flags)
325 {
326         spin_unlock_irqrestore(&sl->lock, flags);
327 }
328
329 #endif /* __LINUX_SEQLOCK_H */