Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1 | #ifndef __LINUX_SEQLOCK_H |
| 2 | #define __LINUX_SEQLOCK_H |
| 3 | /* |
| 4 | * Reader/writer consistent mechanism without starving writers. This type of |
Robert P. J. Day | d08df60 | 2007-02-17 19:07:33 +0100 | [diff] [blame] | 5 | * lock for data where the reader wants a consistent set of information |
Waiman Long | 1370e97 | 2013-09-12 10:55:34 -0400 | [diff] [blame] | 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. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 15 | * |
Waiman Long | 1370e97 | 2013-09-12 10:55:34 -0400 | [diff] [blame] | 16 | * This is not as cache friendly as brlock. Also, this may not work well |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 17 | * for data that contains pointers, because any writer could |
| 18 | * invalidate a pointer that a reader was following. |
| 19 | * |
Waiman Long | 1370e97 | 2013-09-12 10:55:34 -0400 | [diff] [blame] | 20 | * Expected non-blocking reader usage: |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 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 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 35 | #include <linux/spinlock.h> |
| 36 | #include <linux/preempt.h> |
John Stultz | 1ca7d67 | 2013-10-07 15:51:59 -0700 | [diff] [blame] | 37 | #include <linux/lockdep.h> |
David Howells | 56a2105 | 2011-06-11 12:29:58 +0100 | [diff] [blame] | 38 | #include <asm/processor.h> |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 39 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 40 | /* |
| 41 | * Version using sequence counter only. |
| 42 | * This can be used when code has its own mutex protecting the |
| 43 | * updating starting before the write_seqcountbeqin() and ending |
| 44 | * after the write_seqcount_end(). |
| 45 | */ |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 46 | typedef struct seqcount { |
| 47 | unsigned sequence; |
John Stultz | 1ca7d67 | 2013-10-07 15:51:59 -0700 | [diff] [blame] | 48 | #ifdef CONFIG_DEBUG_LOCK_ALLOC |
| 49 | struct lockdep_map dep_map; |
| 50 | #endif |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 51 | } seqcount_t; |
| 52 | |
John Stultz | 1ca7d67 | 2013-10-07 15:51:59 -0700 | [diff] [blame] | 53 | static inline void __seqcount_init(seqcount_t *s, const char *name, |
| 54 | struct lock_class_key *key) |
| 55 | { |
| 56 | /* |
| 57 | * Make sure we are not reinitializing a held lock: |
| 58 | */ |
| 59 | lockdep_init_map(&s->dep_map, name, key, 0); |
| 60 | s->sequence = 0; |
| 61 | } |
| 62 | |
| 63 | #ifdef CONFIG_DEBUG_LOCK_ALLOC |
| 64 | # define SEQCOUNT_DEP_MAP_INIT(lockname) \ |
| 65 | .dep_map = { .name = #lockname } \ |
| 66 | |
| 67 | # define seqcount_init(s) \ |
| 68 | do { \ |
| 69 | static struct lock_class_key __key; \ |
| 70 | __seqcount_init((s), #s, &__key); \ |
| 71 | } while (0) |
| 72 | |
| 73 | static inline void seqcount_lockdep_reader_access(const seqcount_t *s) |
| 74 | { |
| 75 | seqcount_t *l = (seqcount_t *)s; |
| 76 | unsigned long flags; |
| 77 | |
| 78 | local_irq_save(flags); |
| 79 | seqcount_acquire_read(&l->dep_map, 0, 0, _RET_IP_); |
| 80 | seqcount_release(&l->dep_map, 1, _RET_IP_); |
| 81 | local_irq_restore(flags); |
| 82 | } |
| 83 | |
| 84 | #else |
| 85 | # define SEQCOUNT_DEP_MAP_INIT(lockname) |
| 86 | # define seqcount_init(s) __seqcount_init(s, NULL, NULL) |
| 87 | # define seqcount_lockdep_reader_access(x) |
| 88 | #endif |
| 89 | |
| 90 | #define SEQCNT_ZERO(lockname) { .sequence = 0, SEQCOUNT_DEP_MAP_INIT(lockname)} |
| 91 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 92 | |
Nick Piggin | 3c22cd5 | 2011-01-07 17:49:51 +1100 | [diff] [blame] | 93 | /** |
| 94 | * __read_seqcount_begin - begin a seq-read critical section (without barrier) |
| 95 | * @s: pointer to seqcount_t |
| 96 | * Returns: count to be passed to read_seqcount_retry |
| 97 | * |
| 98 | * __read_seqcount_begin is like read_seqcount_begin, but has no smp_rmb() |
| 99 | * barrier. Callers should ensure that smp_rmb() or equivalent ordering is |
| 100 | * provided before actually loading any of the variables that are to be |
| 101 | * protected in this critical section. |
| 102 | * |
| 103 | * Use carefully, only in critical code, and comment how the barrier is |
| 104 | * provided. |
| 105 | */ |
| 106 | static inline unsigned __read_seqcount_begin(const seqcount_t *s) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 107 | { |
Ingo Molnar | 88a411c | 2008-04-03 09:06:13 +0200 | [diff] [blame] | 108 | unsigned ret; |
| 109 | |
| 110 | repeat: |
Linus Torvalds | 2f62427 | 2012-05-04 14:46:02 -0700 | [diff] [blame] | 111 | ret = ACCESS_ONCE(s->sequence); |
Ingo Molnar | 88a411c | 2008-04-03 09:06:13 +0200 | [diff] [blame] | 112 | if (unlikely(ret & 1)) { |
| 113 | cpu_relax(); |
| 114 | goto repeat; |
| 115 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 116 | return ret; |
| 117 | } |
| 118 | |
Nick Piggin | 3c22cd5 | 2011-01-07 17:49:51 +1100 | [diff] [blame] | 119 | /** |
John Stultz | 1ca7d67 | 2013-10-07 15:51:59 -0700 | [diff] [blame] | 120 | * read_seqcount_begin_no_lockdep - start seq-read critical section w/o lockdep |
| 121 | * @s: pointer to seqcount_t |
| 122 | * Returns: count to be passed to read_seqcount_retry |
| 123 | * |
| 124 | * read_seqcount_begin_no_lockdep opens a read critical section of the given |
| 125 | * seqcount, but without any lockdep checking. Validity of the critical |
| 126 | * section is tested by checking read_seqcount_retry function. |
| 127 | */ |
| 128 | static inline unsigned read_seqcount_begin_no_lockdep(const seqcount_t *s) |
| 129 | { |
| 130 | unsigned ret = __read_seqcount_begin(s); |
| 131 | smp_rmb(); |
| 132 | return ret; |
| 133 | } |
| 134 | |
| 135 | /** |
Nick Piggin | 3c22cd5 | 2011-01-07 17:49:51 +1100 | [diff] [blame] | 136 | * read_seqcount_begin - begin a seq-read critical section |
| 137 | * @s: pointer to seqcount_t |
| 138 | * Returns: count to be passed to read_seqcount_retry |
| 139 | * |
| 140 | * read_seqcount_begin opens a read critical section of the given seqcount. |
| 141 | * Validity of the critical section is tested by checking read_seqcount_retry |
| 142 | * function. |
| 143 | */ |
| 144 | static inline unsigned read_seqcount_begin(const seqcount_t *s) |
| 145 | { |
John Stultz | 1ca7d67 | 2013-10-07 15:51:59 -0700 | [diff] [blame] | 146 | seqcount_lockdep_reader_access(s); |
| 147 | return read_seqcount_begin_no_lockdep(s); |
Nick Piggin | 3c22cd5 | 2011-01-07 17:49:51 +1100 | [diff] [blame] | 148 | } |
| 149 | |
| 150 | /** |
Linus Torvalds | 4f988f1 | 2012-05-04 15:13:54 -0700 | [diff] [blame] | 151 | * raw_seqcount_begin - begin a seq-read critical section |
| 152 | * @s: pointer to seqcount_t |
| 153 | * Returns: count to be passed to read_seqcount_retry |
| 154 | * |
| 155 | * raw_seqcount_begin opens a read critical section of the given seqcount. |
| 156 | * Validity of the critical section is tested by checking read_seqcount_retry |
| 157 | * function. |
| 158 | * |
| 159 | * Unlike read_seqcount_begin(), this function will not wait for the count |
| 160 | * to stabilize. If a writer is active when we begin, we will fail the |
| 161 | * read_seqcount_retry() instead of stabilizing at the beginning of the |
| 162 | * critical section. |
| 163 | */ |
| 164 | static inline unsigned raw_seqcount_begin(const seqcount_t *s) |
| 165 | { |
| 166 | unsigned ret = ACCESS_ONCE(s->sequence); |
John Stultz | 1ca7d67 | 2013-10-07 15:51:59 -0700 | [diff] [blame] | 167 | |
| 168 | seqcount_lockdep_reader_access(s); |
Linus Torvalds | 4f988f1 | 2012-05-04 15:13:54 -0700 | [diff] [blame] | 169 | smp_rmb(); |
| 170 | return ret & ~1; |
| 171 | } |
| 172 | |
| 173 | /** |
Nick Piggin | 3c22cd5 | 2011-01-07 17:49:51 +1100 | [diff] [blame] | 174 | * __read_seqcount_retry - end a seq-read critical section (without barrier) |
| 175 | * @s: pointer to seqcount_t |
| 176 | * @start: count, from read_seqcount_begin |
| 177 | * Returns: 1 if retry is required, else 0 |
| 178 | * |
| 179 | * __read_seqcount_retry is like read_seqcount_retry, but has no smp_rmb() |
| 180 | * barrier. Callers should ensure that smp_rmb() or equivalent ordering is |
| 181 | * provided before actually loading any of the variables that are to be |
| 182 | * protected in this critical section. |
| 183 | * |
| 184 | * Use carefully, only in critical code, and comment how the barrier is |
| 185 | * provided. |
| 186 | */ |
| 187 | static inline int __read_seqcount_retry(const seqcount_t *s, unsigned start) |
| 188 | { |
| 189 | return unlikely(s->sequence != start); |
| 190 | } |
| 191 | |
| 192 | /** |
| 193 | * read_seqcount_retry - end a seq-read critical section |
| 194 | * @s: pointer to seqcount_t |
| 195 | * @start: count, from read_seqcount_begin |
| 196 | * Returns: 1 if retry is required, else 0 |
| 197 | * |
| 198 | * read_seqcount_retry closes a read critical section of the given seqcount. |
| 199 | * If the critical section was invalid, it must be ignored (and typically |
| 200 | * retried). |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 201 | */ |
Ingo Molnar | 88a411c | 2008-04-03 09:06:13 +0200 | [diff] [blame] | 202 | static inline int read_seqcount_retry(const seqcount_t *s, unsigned start) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 203 | { |
| 204 | smp_rmb(); |
Nick Piggin | 3c22cd5 | 2011-01-07 17:49:51 +1100 | [diff] [blame] | 205 | return __read_seqcount_retry(s, start); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 206 | } |
| 207 | |
| 208 | |
| 209 | /* |
| 210 | * Sequence counter only version assumes that callers are using their |
| 211 | * own mutexing. |
| 212 | */ |
John Stultz | 1ca7d67 | 2013-10-07 15:51:59 -0700 | [diff] [blame] | 213 | static inline void write_seqcount_begin_nested(seqcount_t *s, int subclass) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 214 | { |
| 215 | s->sequence++; |
| 216 | smp_wmb(); |
John Stultz | 1ca7d67 | 2013-10-07 15:51:59 -0700 | [diff] [blame] | 217 | seqcount_acquire(&s->dep_map, subclass, 0, _RET_IP_); |
| 218 | } |
| 219 | |
| 220 | static inline void write_seqcount_begin(seqcount_t *s) |
| 221 | { |
| 222 | write_seqcount_begin_nested(s, 0); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 223 | } |
| 224 | |
| 225 | static inline void write_seqcount_end(seqcount_t *s) |
| 226 | { |
John Stultz | 1ca7d67 | 2013-10-07 15:51:59 -0700 | [diff] [blame] | 227 | seqcount_release(&s->dep_map, 1, _RET_IP_); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 228 | smp_wmb(); |
| 229 | s->sequence++; |
| 230 | } |
| 231 | |
Nick Piggin | 3c22cd5 | 2011-01-07 17:49:51 +1100 | [diff] [blame] | 232 | /** |
| 233 | * write_seqcount_barrier - invalidate in-progress read-side seq operations |
| 234 | * @s: pointer to seqcount_t |
| 235 | * |
| 236 | * After write_seqcount_barrier, no read-side seq operations will complete |
| 237 | * successfully and see data older than this. |
| 238 | */ |
| 239 | static inline void write_seqcount_barrier(seqcount_t *s) |
| 240 | { |
| 241 | smp_wmb(); |
| 242 | s->sequence+=2; |
| 243 | } |
| 244 | |
Thomas Gleixner | 6617fec | 2011-07-16 18:40:26 +0200 | [diff] [blame] | 245 | typedef struct { |
| 246 | struct seqcount seqcount; |
| 247 | spinlock_t lock; |
| 248 | } seqlock_t; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 249 | |
Thomas Gleixner | 6617fec | 2011-07-16 18:40:26 +0200 | [diff] [blame] | 250 | /* |
| 251 | * These macros triggered gcc-3.x compile-time problems. We think these are |
| 252 | * OK now. Be cautious. |
| 253 | */ |
| 254 | #define __SEQLOCK_UNLOCKED(lockname) \ |
| 255 | { \ |
John Stultz | 1ca7d67 | 2013-10-07 15:51:59 -0700 | [diff] [blame] | 256 | .seqcount = SEQCNT_ZERO(lockname), \ |
Thomas Gleixner | 6617fec | 2011-07-16 18:40:26 +0200 | [diff] [blame] | 257 | .lock = __SPIN_LOCK_UNLOCKED(lockname) \ |
| 258 | } |
| 259 | |
| 260 | #define seqlock_init(x) \ |
| 261 | do { \ |
| 262 | seqcount_init(&(x)->seqcount); \ |
| 263 | spin_lock_init(&(x)->lock); \ |
| 264 | } while (0) |
| 265 | |
| 266 | #define DEFINE_SEQLOCK(x) \ |
| 267 | seqlock_t x = __SEQLOCK_UNLOCKED(x) |
| 268 | |
| 269 | /* |
| 270 | * Read side functions for starting and finalizing a read side section. |
| 271 | */ |
| 272 | static inline unsigned read_seqbegin(const seqlock_t *sl) |
| 273 | { |
| 274 | return read_seqcount_begin(&sl->seqcount); |
| 275 | } |
| 276 | |
| 277 | static inline unsigned read_seqretry(const seqlock_t *sl, unsigned start) |
| 278 | { |
| 279 | return read_seqcount_retry(&sl->seqcount, start); |
| 280 | } |
| 281 | |
| 282 | /* |
| 283 | * Lock out other writers and update the count. |
| 284 | * Acts like a normal spin_lock/unlock. |
| 285 | * Don't need preempt_disable() because that is in the spin_lock already. |
| 286 | */ |
| 287 | static inline void write_seqlock(seqlock_t *sl) |
| 288 | { |
| 289 | spin_lock(&sl->lock); |
| 290 | write_seqcount_begin(&sl->seqcount); |
| 291 | } |
| 292 | |
| 293 | static inline void write_sequnlock(seqlock_t *sl) |
| 294 | { |
| 295 | write_seqcount_end(&sl->seqcount); |
| 296 | spin_unlock(&sl->lock); |
| 297 | } |
| 298 | |
| 299 | static inline void write_seqlock_bh(seqlock_t *sl) |
| 300 | { |
| 301 | spin_lock_bh(&sl->lock); |
| 302 | write_seqcount_begin(&sl->seqcount); |
| 303 | } |
| 304 | |
| 305 | static inline void write_sequnlock_bh(seqlock_t *sl) |
| 306 | { |
| 307 | write_seqcount_end(&sl->seqcount); |
| 308 | spin_unlock_bh(&sl->lock); |
| 309 | } |
| 310 | |
| 311 | static inline void write_seqlock_irq(seqlock_t *sl) |
| 312 | { |
| 313 | spin_lock_irq(&sl->lock); |
| 314 | write_seqcount_begin(&sl->seqcount); |
| 315 | } |
| 316 | |
| 317 | static inline void write_sequnlock_irq(seqlock_t *sl) |
| 318 | { |
| 319 | write_seqcount_end(&sl->seqcount); |
| 320 | spin_unlock_irq(&sl->lock); |
| 321 | } |
| 322 | |
| 323 | static inline unsigned long __write_seqlock_irqsave(seqlock_t *sl) |
| 324 | { |
| 325 | unsigned long flags; |
| 326 | |
| 327 | spin_lock_irqsave(&sl->lock, flags); |
| 328 | write_seqcount_begin(&sl->seqcount); |
| 329 | return flags; |
| 330 | } |
| 331 | |
| 332 | #define write_seqlock_irqsave(lock, flags) \ |
| 333 | do { flags = __write_seqlock_irqsave(lock); } while (0) |
| 334 | |
| 335 | static inline void |
| 336 | write_sequnlock_irqrestore(seqlock_t *sl, unsigned long flags) |
| 337 | { |
| 338 | write_seqcount_end(&sl->seqcount); |
| 339 | spin_unlock_irqrestore(&sl->lock, flags); |
| 340 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 341 | |
Waiman Long | 1370e97 | 2013-09-12 10:55:34 -0400 | [diff] [blame] | 342 | /* |
| 343 | * A locking reader exclusively locks out other writers and locking readers, |
| 344 | * but doesn't update the sequence number. Acts like a normal spin_lock/unlock. |
| 345 | * Don't need preempt_disable() because that is in the spin_lock already. |
| 346 | */ |
| 347 | static inline void read_seqlock_excl(seqlock_t *sl) |
| 348 | { |
| 349 | spin_lock(&sl->lock); |
| 350 | } |
| 351 | |
| 352 | static inline void read_sequnlock_excl(seqlock_t *sl) |
| 353 | { |
| 354 | spin_unlock(&sl->lock); |
| 355 | } |
| 356 | |
Al Viro | 2bc74fe | 2013-10-25 16:39:14 -0400 | [diff] [blame] | 357 | /** |
| 358 | * read_seqbegin_or_lock - begin a sequence number check or locking block |
| 359 | * @lock: sequence lock |
| 360 | * @seq : sequence number to be checked |
| 361 | * |
| 362 | * First try it once optimistically without taking the lock. If that fails, |
| 363 | * take the lock. The sequence number is also used as a marker for deciding |
| 364 | * whether to be a reader (even) or writer (odd). |
| 365 | * N.B. seq must be initialized to an even number to begin with. |
| 366 | */ |
| 367 | static inline void read_seqbegin_or_lock(seqlock_t *lock, int *seq) |
| 368 | { |
| 369 | if (!(*seq & 1)) /* Even */ |
| 370 | *seq = read_seqbegin(lock); |
| 371 | else /* Odd */ |
| 372 | read_seqlock_excl(lock); |
| 373 | } |
| 374 | |
| 375 | static inline int need_seqretry(seqlock_t *lock, int seq) |
| 376 | { |
| 377 | return !(seq & 1) && read_seqretry(lock, seq); |
| 378 | } |
| 379 | |
| 380 | static inline void done_seqretry(seqlock_t *lock, int seq) |
| 381 | { |
| 382 | if (seq & 1) |
| 383 | read_sequnlock_excl(lock); |
| 384 | } |
| 385 | |
Waiman Long | 1370e97 | 2013-09-12 10:55:34 -0400 | [diff] [blame] | 386 | static inline void read_seqlock_excl_bh(seqlock_t *sl) |
| 387 | { |
| 388 | spin_lock_bh(&sl->lock); |
| 389 | } |
| 390 | |
| 391 | static inline void read_sequnlock_excl_bh(seqlock_t *sl) |
| 392 | { |
| 393 | spin_unlock_bh(&sl->lock); |
| 394 | } |
| 395 | |
| 396 | static inline void read_seqlock_excl_irq(seqlock_t *sl) |
| 397 | { |
| 398 | spin_lock_irq(&sl->lock); |
| 399 | } |
| 400 | |
| 401 | static inline void read_sequnlock_excl_irq(seqlock_t *sl) |
| 402 | { |
| 403 | spin_unlock_irq(&sl->lock); |
| 404 | } |
| 405 | |
| 406 | static inline unsigned long __read_seqlock_excl_irqsave(seqlock_t *sl) |
| 407 | { |
| 408 | unsigned long flags; |
| 409 | |
| 410 | spin_lock_irqsave(&sl->lock, flags); |
| 411 | return flags; |
| 412 | } |
| 413 | |
| 414 | #define read_seqlock_excl_irqsave(lock, flags) \ |
| 415 | do { flags = __read_seqlock_excl_irqsave(lock); } while (0) |
| 416 | |
| 417 | static inline void |
| 418 | read_sequnlock_excl_irqrestore(seqlock_t *sl, unsigned long flags) |
| 419 | { |
| 420 | spin_unlock_irqrestore(&sl->lock, flags); |
| 421 | } |
| 422 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 423 | #endif /* __LINUX_SEQLOCK_H */ |