kernel/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 static inline void __raw_write_seqcount_begin(seqcount_t *s)
 223 {
 224         s->sequence++;
 225         smp_wmb();
 226 }
 227
 228 static inline void raw_write_seqcount_begin(seqcount_t *s)
 229 {
 230         preempt_disable_rt();
 231         __raw_write_seqcount_begin(s);
 232 }
 233
 234 static inline void __raw_write_seqcount_end(seqcount_t *s)
 235 {
 236         smp_wmb();
 237         s->sequence++;
 238 }
 239
 240 static inline void raw_write_seqcount_end(seqcount_t *s)
 241 {
 242         __raw_write_seqcount_end(s);
 243         preempt_enable_rt();
 244 }
 245
 246 /*
 247  * raw_write_seqcount_latch - redirect readers to even/odd copy
 248  * @s: pointer to seqcount_t
 249  */
 250 static inline void raw_write_seqcount_latch(seqcount_t *s)
 251 {
 252        smp_wmb();      /* prior stores before incrementing "sequence" */
 253        s->sequence++;
 254        smp_wmb();      /* increment "sequence" before following stores */
 255 }
 256
 257 /*
 258  * Sequence counter only version assumes that callers are using their
 259  * own mutexing.
 260  */
 261 static inline void write_seqcount_begin_nested(seqcount_t *s, int subclass)
 262 {
 263         raw_write_seqcount_begin(s);
 264         seqcount_acquire(&s->dep_map, subclass, 0, _RET_IP_);
 265 }
 266
 267 static inline void write_seqcount_begin(seqcount_t *s)
 268 {
 269         write_seqcount_begin_nested(s, 0);
 270 }
 271
 272 static inline void write_seqcount_end(seqcount_t *s)
 273 {
 274         seqcount_release(&s->dep_map, 1, _RET_IP_);
 275         raw_write_seqcount_end(s);
 276 }
 277
 278 /**
 279  * write_seqcount_barrier - invalidate in-progress read-side seq operations
 280  * @s: pointer to seqcount_t
 281  *
 282  * After write_seqcount_barrier, no read-side seq operations will complete
 283  * successfully and see data older than this.
 284  */
 285 static inline void write_seqcount_barrier(seqcount_t *s)
 286 {
 287         smp_wmb();
 288         s->sequence+=2;
 289 }
 290
 291 typedef struct {
 292         struct seqcount seqcount;
 293         spinlock_t lock;
 294 } seqlock_t;
 295
 296 /*
 297  * These macros triggered gcc-3.x compile-time problems.  We think these are
 298  * OK now.  Be cautious.
 299  */
 300 #define __SEQLOCK_UNLOCKED(lockname)                    \
 301         {                                               \
 302                 .seqcount = SEQCNT_ZERO(lockname),      \
 303                 .lock = __SPIN_LOCK_UNLOCKED(lockname)  \
 304         }
 305
 306 #define seqlock_init(x)                                 \
 307         do {                                            \
 308                 seqcount_init(&(x)->seqcount);          \
 309                 spin_lock_init(&(x)->lock);             \
 310         } while (0)
 311
 312 #define DEFINE_SEQLOCK(x) \
 313                 seqlock_t x = __SEQLOCK_UNLOCKED(x)
 314
 315 /*
 316  * Read side functions for starting and finalizing a read side section.
 317  */
 318 #ifndef CONFIG_PREEMPT_RT_FULL
 319 static inline unsigned read_seqbegin(const seqlock_t *sl)
 320 {
 321         return read_seqcount_begin(&sl->seqcount);
 322 }
 323 #else
 324 /*
 325  * Starvation safe read side for RT
 326  */
 327 static inline unsigned read_seqbegin(seqlock_t *sl)
 328 {
 329         unsigned ret;
 330
 331 repeat:
 332         ret = ACCESS_ONCE(sl->seqcount.sequence);
 333         if (unlikely(ret & 1)) {
 334                 /*
 335                  * Take the lock and let the writer proceed (i.e. evtl
 336                  * boost it), otherwise we could loop here forever.
 337                  */
 338                 spin_unlock_wait(&sl->lock);
 339                 goto repeat;
 340         }
 341         return ret;
 342 }
 343 #endif
 344
 345 static inline unsigned read_seqretry(const seqlock_t *sl, unsigned start)
 346 {
 347         return read_seqcount_retry(&sl->seqcount, start);
 348 }
 349
 350 /*
 351  * Lock out other writers and update the count.
 352  * Acts like a normal spin_lock/unlock.
 353  * Don't need preempt_disable() because that is in the spin_lock already.
 354  */
 355 static inline void write_seqlock(seqlock_t *sl)
 356 {
 357         spin_lock(&sl->lock);
 358         __raw_write_seqcount_begin(&sl->seqcount);
 359 }
 360
 361 static inline void write_sequnlock(seqlock_t *sl)
 362 {
 363         __raw_write_seqcount_end(&sl->seqcount);
 364         spin_unlock(&sl->lock);
 365 }
 366
 367 static inline void write_seqlock_bh(seqlock_t *sl)
 368 {
 369         spin_lock_bh(&sl->lock);
 370         __raw_write_seqcount_begin(&sl->seqcount);
 371 }
 372
 373 static inline void write_sequnlock_bh(seqlock_t *sl)
 374 {
 375         __raw_write_seqcount_end(&sl->seqcount);
 376         spin_unlock_bh(&sl->lock);
 377 }
 378
 379 static inline void write_seqlock_irq(seqlock_t *sl)
 380 {
 381         spin_lock_irq(&sl->lock);
 382         __raw_write_seqcount_begin(&sl->seqcount);
 383 }
 384
 385 static inline void write_sequnlock_irq(seqlock_t *sl)
 386 {
 387         __raw_write_seqcount_end(&sl->seqcount);
 388         spin_unlock_irq(&sl->lock);
 389 }
 390
 391 static inline unsigned long __write_seqlock_irqsave(seqlock_t *sl)
 392 {
 393         unsigned long flags;
 394
 395         spin_lock_irqsave(&sl->lock, flags);
 396         __raw_write_seqcount_begin(&sl->seqcount);
 397         return flags;
 398 }
 399
 400 #define write_seqlock_irqsave(lock, flags)                              \
 401         do { flags = __write_seqlock_irqsave(lock); } while (0)
 402
 403 static inline void
 404 write_sequnlock_irqrestore(seqlock_t *sl, unsigned long flags)
 405 {
 406         __raw_write_seqcount_end(&sl->seqcount);
 407         spin_unlock_irqrestore(&sl->lock, flags);
 408 }
 409
 410 /*
 411  * A locking reader exclusively locks out other writers and locking readers,
 412  * but doesn't update the sequence number. Acts like a normal spin_lock/unlock.
 413  * Don't need preempt_disable() because that is in the spin_lock already.
 414  */
 415 static inline void read_seqlock_excl(seqlock_t *sl)
 416 {
 417         spin_lock(&sl->lock);
 418 }
 419
 420 static inline void read_sequnlock_excl(seqlock_t *sl)
 421 {
 422         spin_unlock(&sl->lock);
 423 }
 424
 425 /**
 426  * read_seqbegin_or_lock - begin a sequence number check or locking block
 427  * @lock: sequence lock
 428  * @seq : sequence number to be checked
 429  *
 430  * First try it once optimistically without taking the lock. If that fails,
 431  * take the lock. The sequence number is also used as a marker for deciding
 432  * whether to be a reader (even) or writer (odd).
 433  * N.B. seq must be initialized to an even number to begin with.
 434  */
 435 static inline void read_seqbegin_or_lock(seqlock_t *lock, int *seq)
 436 {
 437         if (!(*seq & 1))        /* Even */
 438                 *seq = read_seqbegin(lock);
 439         else                    /* Odd */
 440                 read_seqlock_excl(lock);
 441 }
 442
 443 static inline int need_seqretry(seqlock_t *lock, int seq)
 444 {
 445         return !(seq & 1) && read_seqretry(lock, seq);
 446 }
 447
 448 static inline void done_seqretry(seqlock_t *lock, int seq)
 449 {
 450         if (seq & 1)
 451                 read_sequnlock_excl(lock);
 452 }
 453
 454 static inline void read_seqlock_excl_bh(seqlock_t *sl)
 455 {
 456         spin_lock_bh(&sl->lock);
 457 }
 458
 459 static inline void read_sequnlock_excl_bh(seqlock_t *sl)
 460 {
 461         spin_unlock_bh(&sl->lock);
 462 }
 463
 464 static inline void read_seqlock_excl_irq(seqlock_t *sl)
 465 {
 466         spin_lock_irq(&sl->lock);
 467 }
 468
 469 static inline void read_sequnlock_excl_irq(seqlock_t *sl)
 470 {
 471         spin_unlock_irq(&sl->lock);
 472 }
 473
 474 static inline unsigned long __read_seqlock_excl_irqsave(seqlock_t *sl)
 475 {
 476         unsigned long flags;
 477
 478         spin_lock_irqsave(&sl->lock, flags);
 479         return flags;
 480 }
 481
 482 #define read_seqlock_excl_irqsave(lock, flags)                          \
 483         do { flags = __read_seqlock_excl_irqsave(lock); } while (0)
 484
 485 static inline void
 486 read_sequnlock_excl_irqrestore(seqlock_t *sl, unsigned long flags)
 487 {
 488         spin_unlock_irqrestore(&sl->lock, flags);
 489 }
 490
 491 static inline unsigned long
 492 read_seqbegin_or_lock_irqsave(seqlock_t *lock, int *seq)
 493 {
 494         unsigned long flags = 0;
 495
 496         if (!(*seq & 1))        /* Even */
 497                 *seq = read_seqbegin(lock);
 498         else                    /* Odd */
 499                 read_seqlock_excl_irqsave(lock, flags);
 500
 501         return flags;
 502 }
 503
 504 static inline void
 505 done_seqretry_irqrestore(seqlock_t *lock, int seq, unsigned long flags)
 506 {
 507         if (seq & 1)
 508                 read_sequnlock_excl_irqrestore(lock, flags);
 509 }
 510 #endif /* __LINUX_SEQLOCK_H */