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_begin - 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  * raw_read_seqcount_begin 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 raw_read_seqcount_begin(const seqcount_t *s)
 129 {
 130         unsigned ret = __read_seqcount_begin(s);
 131         smp_rmb();
 132         return ret;
 133 }
 134
 135 /**
 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 {
 146         seqcount_lockdep_reader_access(s);
 147         return raw_read_seqcount_begin(s);
 148 }
 149
 150 /**
 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);
 167         smp_rmb();
 168         return ret & ~1;
 169 }
 170
 171 /**
 172  * __read_seqcount_retry - end a seq-read critical section (without barrier)
 173  * @s: pointer to seqcount_t
 174  * @start: count, from read_seqcount_begin
 175  * Returns: 1 if retry is required, else 0
 176  *
 177  * __read_seqcount_retry is like read_seqcount_retry, but has no smp_rmb()
 178  * barrier. Callers should ensure that smp_rmb() or equivalent ordering is
 179  * provided before actually loading any of the variables that are to be
 180  * protected in this critical section.
 181  *
 182  * Use carefully, only in critical code, and comment how the barrier is
 183  * provided.
 184  */
 185 static inline int __read_seqcount_retry(const seqcount_t *s, unsigned start)
 186 {
 187         return unlikely(s->sequence != start);
 188 }
 189
 190 /**
 191  * read_seqcount_retry - end a seq-read critical section
 192  * @s: pointer to seqcount_t
 193  * @start: count, from read_seqcount_begin
 194  * Returns: 1 if retry is required, else 0
 195  *
 196  * read_seqcount_retry closes a read critical section of the given seqcount.
 197  * If the critical section was invalid, it must be ignored (and typically
 198  * retried).
 199  */
 200 static inline int read_seqcount_retry(const seqcount_t *s, unsigned start)
 201 {
 202         smp_rmb();
 203         return __read_seqcount_retry(s, start);
 204 }
 205
 206
 207
 208 static inline void raw_write_seqcount_begin(seqcount_t *s)
 209 {
 210         s->sequence++;
 211         smp_wmb();
 212 }
 213
 214 static inline void raw_write_seqcount_end(seqcount_t *s)
 215 {
 216         smp_wmb();
 217         s->sequence++;
 218 }
 219
 220 /*
 221  * Sequence counter only version assumes that callers are using their
 222  * own mutexing.
 223  */
 224 static inline void write_seqcount_begin_nested(seqcount_t *s, int subclass)
 225 {
 226         raw_write_seqcount_begin(s);
 227         seqcount_acquire(&s->dep_map, subclass, 0, _RET_IP_);
 228 }
 229
 230 static inline void write_seqcount_begin(seqcount_t *s)
 231 {
 232         write_seqcount_begin_nested(s, 0);
 233 }
 234
 235 static inline void write_seqcount_end(seqcount_t *s)
 236 {
 237         seqcount_release(&s->dep_map, 1, _RET_IP_);
 238         raw_write_seqcount_end(s);
 239 }
 240
 241 /**
 242  * write_seqcount_barrier - invalidate in-progress read-side seq operations
 243  * @s: pointer to seqcount_t
 244  *
 245  * After write_seqcount_barrier, no read-side seq operations will complete
 246  * successfully and see data older than this.
 247  */
 248 static inline void write_seqcount_barrier(seqcount_t *s)
 249 {
 250         smp_wmb();
 251         s->sequence+=2;
 252 }
 253
 254 typedef struct {
 255         struct seqcount seqcount;
 256         spinlock_t lock;
 257 } seqlock_t;
 258
 259 /*
 260  * These macros triggered gcc-3.x compile-time problems.  We think these are
 261  * OK now.  Be cautious.
 262  */
 263 #define __SEQLOCK_UNLOCKED(lockname)                    \
 264         {                                               \
 265                 .seqcount = SEQCNT_ZERO(lockname),      \
 266                 .lock = __SPIN_LOCK_UNLOCKED(lockname)  \
 267         }
 268
 269 #define seqlock_init(x)                                 \
 270         do {                                            \
 271                 seqcount_init(&(x)->seqcount);          \
 272                 spin_lock_init(&(x)->lock);             \
 273         } while (0)
 274
 275 #define DEFINE_SEQLOCK(x) \
 276                 seqlock_t x = __SEQLOCK_UNLOCKED(x)
 277
 278 /*
 279  * Read side functions for starting and finalizing a read side section.
 280  */
 281 static inline unsigned read_seqbegin(const seqlock_t *sl)
 282 {
 283         return read_seqcount_begin(&sl->seqcount);
 284 }
 285
 286 static inline unsigned read_seqretry(const seqlock_t *sl, unsigned start)
 287 {
 288         return read_seqcount_retry(&sl->seqcount, start);
 289 }
 290
 291 /*
 292  * Lock out other writers and update the count.
 293  * Acts like a normal spin_lock/unlock.
 294  * Don't need preempt_disable() because that is in the spin_lock already.
 295  */
 296 static inline void write_seqlock(seqlock_t *sl)
 297 {
 298         spin_lock(&sl->lock);
 299         write_seqcount_begin(&sl->seqcount);
 300 }
 301
 302 static inline void write_sequnlock(seqlock_t *sl)
 303 {
 304         write_seqcount_end(&sl->seqcount);
 305         spin_unlock(&sl->lock);
 306 }
 307
 308 static inline void write_seqlock_bh(seqlock_t *sl)
 309 {
 310         spin_lock_bh(&sl->lock);
 311         write_seqcount_begin(&sl->seqcount);
 312 }
 313
 314 static inline void write_sequnlock_bh(seqlock_t *sl)
 315 {
 316         write_seqcount_end(&sl->seqcount);
 317         spin_unlock_bh(&sl->lock);
 318 }
 319
 320 static inline void write_seqlock_irq(seqlock_t *sl)
 321 {
 322         spin_lock_irq(&sl->lock);
 323         write_seqcount_begin(&sl->seqcount);
 324 }
 325
 326 static inline void write_sequnlock_irq(seqlock_t *sl)
 327 {
 328         write_seqcount_end(&sl->seqcount);
 329         spin_unlock_irq(&sl->lock);
 330 }
 331
 332 static inline unsigned long __write_seqlock_irqsave(seqlock_t *sl)
 333 {
 334         unsigned long flags;
 335
 336         spin_lock_irqsave(&sl->lock, flags);
 337         write_seqcount_begin(&sl->seqcount);
 338         return flags;
 339 }
 340
 341 #define write_seqlock_irqsave(lock, flags)                              \
 342         do { flags = __write_seqlock_irqsave(lock); } while (0)
 343
 344 static inline void
 345 write_sequnlock_irqrestore(seqlock_t *sl, unsigned long flags)
 346 {
 347         write_seqcount_end(&sl->seqcount);
 348         spin_unlock_irqrestore(&sl->lock, flags);
 349 }
 350
 351 /*
 352  * A locking reader exclusively locks out other writers and locking readers,
 353  * but doesn't update the sequence number. Acts like a normal spin_lock/unlock.
 354  * Don't need preempt_disable() because that is in the spin_lock already.
 355  */
 356 static inline void read_seqlock_excl(seqlock_t *sl)
 357 {
 358         spin_lock(&sl->lock);
 359 }
 360
 361 static inline void read_sequnlock_excl(seqlock_t *sl)
 362 {
 363         spin_unlock(&sl->lock);
 364 }
 365
 366 /**
 367  * read_seqbegin_or_lock - begin a sequence number check or locking block
 368  * @lock: sequence lock
 369  * @seq : sequence number to be checked
 370  *
 371  * First try it once optimistically without taking the lock. If that fails,
 372  * take the lock. The sequence number is also used as a marker for deciding
 373  * whether to be a reader (even) or writer (odd).
 374  * N.B. seq must be initialized to an even number to begin with.
 375  */
 376 static inline void read_seqbegin_or_lock(seqlock_t *lock, int *seq)
 377 {
 378         if (!(*seq & 1))        /* Even */
 379                 *seq = read_seqbegin(lock);
 380         else                    /* Odd */
 381                 read_seqlock_excl(lock);
 382 }
 383
 384 static inline int need_seqretry(seqlock_t *lock, int seq)
 385 {
 386         return !(seq & 1) && read_seqretry(lock, seq);
 387 }
 388
 389 static inline void done_seqretry(seqlock_t *lock, int seq)
 390 {
 391         if (seq & 1)
 392                 read_sequnlock_excl(lock);
 393 }
 394
 395 static inline void read_seqlock_excl_bh(seqlock_t *sl)
 396 {
 397         spin_lock_bh(&sl->lock);
 398 }
 399
 400 static inline void read_sequnlock_excl_bh(seqlock_t *sl)
 401 {
 402         spin_unlock_bh(&sl->lock);
 403 }
 404
 405 static inline void read_seqlock_excl_irq(seqlock_t *sl)
 406 {
 407         spin_lock_irq(&sl->lock);
 408 }
 409
 410 static inline void read_sequnlock_excl_irq(seqlock_t *sl)
 411 {
 412         spin_unlock_irq(&sl->lock);
 413 }
 414
 415 static inline unsigned long __read_seqlock_excl_irqsave(seqlock_t *sl)
 416 {
 417         unsigned long flags;
 418
 419         spin_lock_irqsave(&sl->lock, flags);
 420         return flags;
 421 }
 422
 423 #define read_seqlock_excl_irqsave(lock, flags)                          \
 424         do { flags = __read_seqlock_excl_irqsave(lock); } while (0)
 425
 426 static inline void
 427 read_sequnlock_excl_irqrestore(seqlock_t *sl, unsigned long flags)
 428 {
 429         spin_unlock_irqrestore(&sl->lock, flags);
 430 }
 431
 432 #endif /* __LINUX_SEQLOCK_H */