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 <linux/lockdep.h>
  38 #include <asm/processor.h>
  39
  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  */
  46 typedef struct seqcount {
  47         unsigned sequence;
  48 #ifdef CONFIG_DEBUG_LOCK_ALLOC
  49         struct lockdep_map dep_map;
  50 #endif
  51 } seqcount_t;
  52
  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
  92
  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)
 107 {
 108         unsigned ret;
 109
 110 repeat:
 111         ret = ACCESS_ONCE(s->sequence);
 112         if (unlikely(ret & 1)) {
 113                 cpu_relax();
 114                 goto repeat;
 115         }
 116         return ret;
 117 }
 118
 119 /**
 120  * raw_read_seqcount - Read the raw seqcount
 121  * @s: pointer to seqcount_t
 122  * Returns: count to be passed to read_seqcount_retry
 123  *
 124  * raw_read_seqcount opens a read critical section of the given
 125  * seqcount without any lockdep checking and without checking or
 126  * masking the LSB. Calling code is responsible for handling that.
 127  */
 128 static inline unsigned raw_read_seqcount(const seqcount_t *s)
 129 {
 130         unsigned ret = ACCESS_ONCE(s->sequence);
 131         smp_rmb();
 132         return ret;
 133 }
 134
 135 /**
 136  * raw_read_seqcount_begin - start seq-read critical section w/o lockdep
 137  * @s: pointer to seqcount_t
 138  * Returns: count to be passed to read_seqcount_retry
 139  *
 140  * raw_read_seqcount_begin opens a read critical section of the given
 141  * seqcount, but without any lockdep checking. Validity of the critical
 142  * section is tested by checking read_seqcount_retry function.
 143  */
 144 static inline unsigned raw_read_seqcount_begin(const seqcount_t *s)
 145 {
 146         unsigned ret = __read_seqcount_begin(s);
 147         smp_rmb();
 148         return ret;
 149 }
 150
 151 /**
 152  * read_seqcount_begin - begin a seq-read critical section
 153  * @s: pointer to seqcount_t
 154  * Returns: count to be passed to read_seqcount_retry
 155  *
 156  * read_seqcount_begin opens a read critical section of the given seqcount.
 157  * Validity of the critical section is tested by checking read_seqcount_retry
 158  * function.
 159  */
 160 static inline unsigned read_seqcount_begin(const seqcount_t *s)
 161 {
 162         seqcount_lockdep_reader_access(s);
 163         return raw_read_seqcount_begin(s);
 164 }
 165
 166 /**
 167  * raw_seqcount_begin - begin a seq-read critical section
 168  * @s: pointer to seqcount_t
 169  * Returns: count to be passed to read_seqcount_retry
 170  *
 171  * raw_seqcount_begin opens a read critical section of the given seqcount.
 172  * Validity of the critical section is tested by checking read_seqcount_retry
 173  * function.
 174  *
 175  * Unlike read_seqcount_begin(), this function will not wait for the count
 176  * to stabilize. If a writer is active when we begin, we will fail the
 177  * read_seqcount_retry() instead of stabilizing at the beginning of the
 178  * critical section.
 179  */
 180 static inline unsigned raw_seqcount_begin(const seqcount_t *s)
 181 {
 182         unsigned ret = ACCESS_ONCE(s->sequence);
 183
 184         seqcount_lockdep_reader_access(s);
 185         smp_rmb();
 186         return ret & ~1;
 187 }
 188
 189 /**
 190  * __read_seqcount_retry - end a seq-read critical section (without barrier)
 191  * @s: pointer to seqcount_t
 192  * @start: count, from read_seqcount_begin
 193  * Returns: 1 if retry is required, else 0
 194  *
 195  * __read_seqcount_retry is like read_seqcount_retry, but has no smp_rmb()
 196  * barrier. Callers should ensure that smp_rmb() or equivalent ordering is
 197  * provided before actually loading any of the variables that are to be
 198  * protected in this critical section.
 199  *
 200  * Use carefully, only in critical code, and comment how the barrier is
 201  * provided.
 202  */
 203 static inline int __read_seqcount_retry(const seqcount_t *s, unsigned start)
 204 {
 205         return unlikely(s->sequence != start);
 206 }
 207
 208 /**
 209  * read_seqcount_retry - end a seq-read critical section
 210  * @s: pointer to seqcount_t
 211  * @start: count, from read_seqcount_begin
 212  * Returns: 1 if retry is required, else 0
 213  *
 214  * read_seqcount_retry closes a read critical section of the given seqcount.
 215  * If the critical section was invalid, it must be ignored (and typically
 216  * retried).
 217  */
 218 static inline int read_seqcount_retry(const seqcount_t *s, unsigned start)
 219 {
 220         smp_rmb();
 221         return __read_seqcount_retry(s, start);
 222 }
 223
 224
 225
 226 static inline void raw_write_seqcount_begin(seqcount_t *s)
 227 {
 228         s->sequence++;
 229         smp_wmb();
 230 }
 231
 232 static inline void raw_write_seqcount_end(seqcount_t *s)
 233 {
 234         smp_wmb();
 235         s->sequence++;
 236 }
 237
 238 /*
 239  * raw_write_seqcount_latch - redirect readers to even/odd copy
 240  * @s: pointer to seqcount_t
 241  */
 242 static inline void raw_write_seqcount_latch(seqcount_t *s)
 243 {
 244        smp_wmb();      /* prior stores before incrementing "sequence" */
 245        s->sequence++;
 246        smp_wmb();      /* increment "sequence" before following stores */
 247 }
 248
 249 /*
 250  * Sequence counter only version assumes that callers are using their
 251  * own mutexing.
 252  */
 253 static inline void write_seqcount_begin_nested(seqcount_t *s, int subclass)
 254 {
 255         raw_write_seqcount_begin(s);
 256         seqcount_acquire(&s->dep_map, subclass, 0, _RET_IP_);
 257 }
 258
 259 static inline void write_seqcount_begin(seqcount_t *s)
 260 {
 261         write_seqcount_begin_nested(s, 0);
 262 }
 263
 264 static inline void write_seqcount_end(seqcount_t *s)
 265 {
 266         seqcount_release(&s->dep_map, 1, _RET_IP_);
 267         raw_write_seqcount_end(s);
 268 }
 269
 270 /**
 271  * write_seqcount_barrier - invalidate in-progress read-side seq operations
 272  * @s: pointer to seqcount_t
 273  *
 274  * After write_seqcount_barrier, no read-side seq operations will complete
 275  * successfully and see data older than this.
 276  */
 277 static inline void write_seqcount_barrier(seqcount_t *s)
 278 {
 279         smp_wmb();
 280         s->sequence+=2;
 281 }
 282
 283 typedef struct {
 284         struct seqcount seqcount;
 285         spinlock_t lock;
 286 } seqlock_t;
 287
 288 /*
 289  * These macros triggered gcc-3.x compile-time problems.  We think these are
 290  * OK now.  Be cautious.
 291  */
 292 #define __SEQLOCK_UNLOCKED(lockname)                    \
 293         {                                               \
 294                 .seqcount = SEQCNT_ZERO(lockname),      \
 295                 .lock = __SPIN_LOCK_UNLOCKED(lockname)  \
 296         }
 297
 298 #define seqlock_init(x)                                 \
 299         do {                                            \
 300                 seqcount_init(&(x)->seqcount);          \
 301                 spin_lock_init(&(x)->lock);             \
 302         } while (0)
 303
 304 #define DEFINE_SEQLOCK(x) \
 305                 seqlock_t x = __SEQLOCK_UNLOCKED(x)
 306
 307 /*
 308  * Read side functions for starting and finalizing a read side section.
 309  */
 310 static inline unsigned read_seqbegin(const seqlock_t *sl)
 311 {
 312         return read_seqcount_begin(&sl->seqcount);
 313 }
 314
 315 static inline unsigned read_seqretry(const seqlock_t *sl, unsigned start)
 316 {
 317         return read_seqcount_retry(&sl->seqcount, start);
 318 }
 319
 320 /*
 321  * Lock out other writers and update the count.
 322  * Acts like a normal spin_lock/unlock.
 323  * Don't need preempt_disable() because that is in the spin_lock already.
 324  */
 325 static inline void write_seqlock(seqlock_t *sl)
 326 {
 327         spin_lock(&sl->lock);
 328         write_seqcount_begin(&sl->seqcount);
 329 }
 330
 331 static inline void write_sequnlock(seqlock_t *sl)
 332 {
 333         write_seqcount_end(&sl->seqcount);
 334         spin_unlock(&sl->lock);
 335 }
 336
 337 static inline void write_seqlock_bh(seqlock_t *sl)
 338 {
 339         spin_lock_bh(&sl->lock);
 340         write_seqcount_begin(&sl->seqcount);
 341 }
 342
 343 static inline void write_sequnlock_bh(seqlock_t *sl)
 344 {
 345         write_seqcount_end(&sl->seqcount);
 346         spin_unlock_bh(&sl->lock);
 347 }
 348
 349 static inline void write_seqlock_irq(seqlock_t *sl)
 350 {
 351         spin_lock_irq(&sl->lock);
 352         write_seqcount_begin(&sl->seqcount);
 353 }
 354
 355 static inline void write_sequnlock_irq(seqlock_t *sl)
 356 {
 357         write_seqcount_end(&sl->seqcount);
 358         spin_unlock_irq(&sl->lock);
 359 }
 360
 361 static inline unsigned long __write_seqlock_irqsave(seqlock_t *sl)
 362 {
 363         unsigned long flags;
 364
 365         spin_lock_irqsave(&sl->lock, flags);
 366         write_seqcount_begin(&sl->seqcount);
 367         return flags;
 368 }
 369
 370 #define write_seqlock_irqsave(lock, flags)                              \
 371         do { flags = __write_seqlock_irqsave(lock); } while (0)
 372
 373 static inline void
 374 write_sequnlock_irqrestore(seqlock_t *sl, unsigned long flags)
 375 {
 376         write_seqcount_end(&sl->seqcount);
 377         spin_unlock_irqrestore(&sl->lock, flags);
 378 }
 379
 380 /*
 381  * A locking reader exclusively locks out other writers and locking readers,
 382  * but doesn't update the sequence number. Acts like a normal spin_lock/unlock.
 383  * Don't need preempt_disable() because that is in the spin_lock already.
 384  */
 385 static inline void read_seqlock_excl(seqlock_t *sl)
 386 {
 387         spin_lock(&sl->lock);
 388 }
 389
 390 static inline void read_sequnlock_excl(seqlock_t *sl)
 391 {
 392         spin_unlock(&sl->lock);
 393 }
 394
 395 /**
 396  * read_seqbegin_or_lock - begin a sequence number check or locking block
 397  * @lock: sequence lock
 398  * @seq : sequence number to be checked
 399  *
 400  * First try it once optimistically without taking the lock. If that fails,
 401  * take the lock. The sequence number is also used as a marker for deciding
 402  * whether to be a reader (even) or writer (odd).
 403  * N.B. seq must be initialized to an even number to begin with.
 404  */
 405 static inline void read_seqbegin_or_lock(seqlock_t *lock, int *seq)
 406 {
 407         if (!(*seq & 1))        /* Even */
 408                 *seq = read_seqbegin(lock);
 409         else                    /* Odd */
 410                 read_seqlock_excl(lock);
 411 }
 412
 413 static inline int need_seqretry(seqlock_t *lock, int seq)
 414 {
 415         return !(seq & 1) && read_seqretry(lock, seq);
 416 }
 417
 418 static inline void done_seqretry(seqlock_t *lock, int seq)
 419 {
 420         if (seq & 1)
 421                 read_sequnlock_excl(lock);
 422 }
 423
 424 static inline void read_seqlock_excl_bh(seqlock_t *sl)
 425 {
 426         spin_lock_bh(&sl->lock);
 427 }
 428
 429 static inline void read_sequnlock_excl_bh(seqlock_t *sl)
 430 {
 431         spin_unlock_bh(&sl->lock);
 432 }
 433
 434 static inline void read_seqlock_excl_irq(seqlock_t *sl)
 435 {
 436         spin_lock_irq(&sl->lock);
 437 }
 438
 439 static inline void read_sequnlock_excl_irq(seqlock_t *sl)
 440 {
 441         spin_unlock_irq(&sl->lock);
 442 }
 443
 444 static inline unsigned long __read_seqlock_excl_irqsave(seqlock_t *sl)
 445 {
 446         unsigned long flags;
 447
 448         spin_lock_irqsave(&sl->lock, flags);
 449         return flags;
 450 }
 451
 452 #define read_seqlock_excl_irqsave(lock, flags)                          \
 453         do { flags = __read_seqlock_excl_irqsave(lock); } while (0)
 454
 455 static inline void
 456 read_sequnlock_excl_irqrestore(seqlock_t *sl, unsigned long flags)
 457 {
 458         spin_unlock_irqrestore(&sl->lock, flags);
 459 }
 460
 461 #endif /* __LINUX_SEQLOCK_H */