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 = READ_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 = READ_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 = READ_ONCE(s->sequence);
 183         smp_rmb();
 184         return ret & ~1;
 185 }
 186
 187 /**
 188  * __read_seqcount_retry - end a seq-read critical section (without barrier)
 189  * @s: pointer to seqcount_t
 190  * @start: count, from read_seqcount_begin
 191  * Returns: 1 if retry is required, else 0
 192  *
 193  * __read_seqcount_retry is like read_seqcount_retry, but has no smp_rmb()
 194  * barrier. Callers should ensure that smp_rmb() or equivalent ordering is
 195  * provided before actually loading any of the variables that are to be
 196  * protected in this critical section.
 197  *
 198  * Use carefully, only in critical code, and comment how the barrier is
 199  * provided.
 200  */
 201 static inline int __read_seqcount_retry(const seqcount_t *s, unsigned start)
 202 {
 203         return unlikely(s->sequence != start);
 204 }
 205
 206 /**
 207  * read_seqcount_retry - end a seq-read critical section
 208  * @s: pointer to seqcount_t
 209  * @start: count, from read_seqcount_begin
 210  * Returns: 1 if retry is required, else 0
 211  *
 212  * read_seqcount_retry closes a read critical section of the given seqcount.
 213  * If the critical section was invalid, it must be ignored (and typically
 214  * retried).
 215  */
 216 static inline int read_seqcount_retry(const seqcount_t *s, unsigned start)
 217 {
 218         smp_rmb();
 219         return __read_seqcount_retry(s, start);
 220 }
 221
 222
 223
 224 static inline void raw_write_seqcount_begin(seqcount_t *s)
 225 {
 226         s->sequence++;
 227         smp_wmb();
 228 }
 229
 230 static inline void raw_write_seqcount_end(seqcount_t *s)
 231 {
 232         smp_wmb();
 233         s->sequence++;
 234 }
 235
 236 /**
 237  * raw_write_seqcount_barrier - do a seq write barrier
 238  * @s: pointer to seqcount_t
 239  *
 240  * This can be used to provide an ordering guarantee instead of the
 241  * usual consistency guarantee. It is one wmb cheaper, because we can
 242  * collapse the two back-to-back wmb()s.
 243  *
 244  *      seqcount_t seq;
 245  *      bool X = true, Y = false;
 246  *
 247  *      void read(void)
 248  *      {
 249  *              bool x, y;
 250  *
 251  *              do {
 252  *                      int s = read_seqcount_begin(&seq);
 253  *
 254  *                      x = X; y = Y;
 255  *
 256  *              } while (read_seqcount_retry(&seq, s));
 257  *
 258  *              BUG_ON(!x && !y);
 259  *      }
 260  *
 261  *      void write(void)
 262  *      {
 263  *              Y = true;
 264  *
 265  *              raw_write_seqcount_barrier(seq);
 266  *
 267  *              X = false;
 268  *      }
 269  */
 270 static inline void raw_write_seqcount_barrier(seqcount_t *s)
 271 {
 272         s->sequence++;
 273         smp_wmb();
 274         s->sequence++;
 275 }
 276
 277 /*
 278  * raw_write_seqcount_latch - redirect readers to even/odd copy
 279  * @s: pointer to seqcount_t
 280  */
 281 static inline void raw_write_seqcount_latch(seqcount_t *s)
 282 {
 283        smp_wmb();      /* prior stores before incrementing "sequence" */
 284        s->sequence++;
 285        smp_wmb();      /* increment "sequence" before following stores */
 286 }
 287
 288 /*
 289  * Sequence counter only version assumes that callers are using their
 290  * own mutexing.
 291  */
 292 static inline void write_seqcount_begin_nested(seqcount_t *s, int subclass)
 293 {
 294         raw_write_seqcount_begin(s);
 295         seqcount_acquire(&s->dep_map, subclass, 0, _RET_IP_);
 296 }
 297
 298 static inline void write_seqcount_begin(seqcount_t *s)
 299 {
 300         write_seqcount_begin_nested(s, 0);
 301 }
 302
 303 static inline void write_seqcount_end(seqcount_t *s)
 304 {
 305         seqcount_release(&s->dep_map, 1, _RET_IP_);
 306         raw_write_seqcount_end(s);
 307 }
 308
 309 /**
 310  * write_seqcount_invalidate - invalidate in-progress read-side seq operations
 311  * @s: pointer to seqcount_t
 312  *
 313  * After write_seqcount_invalidate, no read-side seq operations will complete
 314  * successfully and see data older than this.
 315  */
 316 static inline void write_seqcount_invalidate(seqcount_t *s)
 317 {
 318         smp_wmb();
 319         s->sequence+=2;
 320 }
 321
 322 typedef struct {
 323         struct seqcount seqcount;
 324         spinlock_t lock;
 325 } seqlock_t;
 326
 327 /*
 328  * These macros triggered gcc-3.x compile-time problems.  We think these are
 329  * OK now.  Be cautious.
 330  */
 331 #define __SEQLOCK_UNLOCKED(lockname)                    \
 332         {                                               \
 333                 .seqcount = SEQCNT_ZERO(lockname),      \
 334                 .lock = __SPIN_LOCK_UNLOCKED(lockname)  \
 335         }
 336
 337 #define seqlock_init(x)                                 \
 338         do {                                            \
 339                 seqcount_init(&(x)->seqcount);          \
 340                 spin_lock_init(&(x)->lock);             \
 341         } while (0)
 342
 343 #define DEFINE_SEQLOCK(x) \
 344                 seqlock_t x = __SEQLOCK_UNLOCKED(x)
 345
 346 /*
 347  * Read side functions for starting and finalizing a read side section.
 348  */
 349 static inline unsigned read_seqbegin(const seqlock_t *sl)
 350 {
 351         return read_seqcount_begin(&sl->seqcount);
 352 }
 353
 354 static inline unsigned read_seqretry(const seqlock_t *sl, unsigned start)
 355 {
 356         return read_seqcount_retry(&sl->seqcount, start);
 357 }
 358
 359 /*
 360  * Lock out other writers and update the count.
 361  * Acts like a normal spin_lock/unlock.
 362  * Don't need preempt_disable() because that is in the spin_lock already.
 363  */
 364 static inline void write_seqlock(seqlock_t *sl)
 365 {
 366         spin_lock(&sl->lock);
 367         write_seqcount_begin(&sl->seqcount);
 368 }
 369
 370 static inline void write_sequnlock(seqlock_t *sl)
 371 {
 372         write_seqcount_end(&sl->seqcount);
 373         spin_unlock(&sl->lock);
 374 }
 375
 376 static inline void write_seqlock_bh(seqlock_t *sl)
 377 {
 378         spin_lock_bh(&sl->lock);
 379         write_seqcount_begin(&sl->seqcount);
 380 }
 381
 382 static inline void write_sequnlock_bh(seqlock_t *sl)
 383 {
 384         write_seqcount_end(&sl->seqcount);
 385         spin_unlock_bh(&sl->lock);
 386 }
 387
 388 static inline void write_seqlock_irq(seqlock_t *sl)
 389 {
 390         spin_lock_irq(&sl->lock);
 391         write_seqcount_begin(&sl->seqcount);
 392 }
 393
 394 static inline void write_sequnlock_irq(seqlock_t *sl)
 395 {
 396         write_seqcount_end(&sl->seqcount);
 397         spin_unlock_irq(&sl->lock);
 398 }
 399
 400 static inline unsigned long __write_seqlock_irqsave(seqlock_t *sl)
 401 {
 402         unsigned long flags;
 403
 404         spin_lock_irqsave(&sl->lock, flags);
 405         write_seqcount_begin(&sl->seqcount);
 406         return flags;
 407 }
 408
 409 #define write_seqlock_irqsave(lock, flags)                              \
 410         do { flags = __write_seqlock_irqsave(lock); } while (0)
 411
 412 static inline void
 413 write_sequnlock_irqrestore(seqlock_t *sl, unsigned long flags)
 414 {
 415         write_seqcount_end(&sl->seqcount);
 416         spin_unlock_irqrestore(&sl->lock, flags);
 417 }
 418
 419 /*
 420  * A locking reader exclusively locks out other writers and locking readers,
 421  * but doesn't update the sequence number. Acts like a normal spin_lock/unlock.
 422  * Don't need preempt_disable() because that is in the spin_lock already.
 423  */
 424 static inline void read_seqlock_excl(seqlock_t *sl)
 425 {
 426         spin_lock(&sl->lock);
 427 }
 428
 429 static inline void read_sequnlock_excl(seqlock_t *sl)
 430 {
 431         spin_unlock(&sl->lock);
 432 }
 433
 434 /**
 435  * read_seqbegin_or_lock - begin a sequence number check or locking block
 436  * @lock: sequence lock
 437  * @seq : sequence number to be checked
 438  *
 439  * First try it once optimistically without taking the lock. If that fails,
 440  * take the lock. The sequence number is also used as a marker for deciding
 441  * whether to be a reader (even) or writer (odd).
 442  * N.B. seq must be initialized to an even number to begin with.
 443  */
 444 static inline void read_seqbegin_or_lock(seqlock_t *lock, int *seq)
 445 {
 446         if (!(*seq & 1))        /* Even */
 447                 *seq = read_seqbegin(lock);
 448         else                    /* Odd */
 449                 read_seqlock_excl(lock);
 450 }
 451
 452 static inline int need_seqretry(seqlock_t *lock, int seq)
 453 {
 454         return !(seq & 1) && read_seqretry(lock, seq);
 455 }
 456
 457 static inline void done_seqretry(seqlock_t *lock, int seq)
 458 {
 459         if (seq & 1)
 460                 read_sequnlock_excl(lock);
 461 }
 462
 463 static inline void read_seqlock_excl_bh(seqlock_t *sl)
 464 {
 465         spin_lock_bh(&sl->lock);
 466 }
 467
 468 static inline void read_sequnlock_excl_bh(seqlock_t *sl)
 469 {
 470         spin_unlock_bh(&sl->lock);
 471 }
 472
 473 static inline void read_seqlock_excl_irq(seqlock_t *sl)
 474 {
 475         spin_lock_irq(&sl->lock);
 476 }
 477
 478 static inline void read_sequnlock_excl_irq(seqlock_t *sl)
 479 {
 480         spin_unlock_irq(&sl->lock);
 481 }
 482
 483 static inline unsigned long __read_seqlock_excl_irqsave(seqlock_t *sl)
 484 {
 485         unsigned long flags;
 486
 487         spin_lock_irqsave(&sl->lock, flags);
 488         return flags;
 489 }
 490
 491 #define read_seqlock_excl_irqsave(lock, flags)                          \
 492         do { flags = __read_seqlock_excl_irqsave(lock); } while (0)
 493
 494 static inline void
 495 read_sequnlock_excl_irqrestore(seqlock_t *sl, unsigned long flags)
 496 {
 497         spin_unlock_irqrestore(&sl->lock, flags);
 498 }
 499
 500 static inline unsigned long
 501 read_seqbegin_or_lock_irqsave(seqlock_t *lock, int *seq)
 502 {
 503         unsigned long flags = 0;
 504
 505         if (!(*seq & 1))        /* Even */
 506                 *seq = read_seqbegin(lock);
 507         else                    /* Odd */
 508                 read_seqlock_excl_irqsave(lock, flags);
 509
 510         return flags;
 511 }
 512
 513 static inline void
 514 done_seqretry_irqrestore(seqlock_t *lock, int seq, unsigned long flags)
 515 {
 516         if (seq & 1)
 517                 read_sequnlock_excl_irqrestore(lock, flags);
 518 }
 519 #endif /* __LINUX_SEQLOCK_H */