src/urcu.c

   1 // SPDX-FileCopyrightText: 2009 Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
   2 // SPDX-FileCopyrightText: 2009 Paul E. McKenney, IBM Corporation.
   3 //
   4 // SPDX-License-Identifier: LGPL-2.1-or-later
   5
   6 /*
   7  * Userspace RCU library
   8  *
   9  * IBM's contributions to this file may be relicensed under LGPLv2 or later.
  10  */
  11
  12 #define URCU_NO_COMPAT_IDENTIFIERS
  13 #define _BSD_SOURCE
  14 #define _LGPL_SOURCE
  15 #define _DEFAULT_SOURCE
  16 #include <stdio.h>
  17 #include <pthread.h>
  18 #include <signal.h>
  19 #include <stdlib.h>
  20 #include <stdint.h>
  21 #include <string.h>
  22 #include <errno.h>
  23 #include <stdbool.h>
  24 #include <poll.h>
  25
  26 #include <urcu/config.h>
  27 #include <urcu/assert.h>
  28 #include <urcu/arch.h>
  29 #include <urcu/wfcqueue.h>
  30 #include <urcu/map/urcu.h>
  31 #include <urcu/static/urcu.h>
  32 #include <urcu/pointer.h>
  33 #include <urcu/tls-compat.h>
  34
  35 #include "urcu-die.h"
  36 #include "urcu-wait.h"
  37 #include "urcu-utils.h"
  38
  39 #define URCU_API_MAP
  40 /* Do not #define _LGPL_SOURCE to ensure we can emit the wrapper symbols */
  41 #undef _LGPL_SOURCE
  42 #include <urcu/urcu.h>
  43 #define _LGPL_SOURCE
  44
  45 /*
  46  * If a reader is really non-cooperative and refuses to commit its
  47  * rcu_active_readers count to memory (there is no barrier in the reader
  48  * per-se), kick it after 10 loops waiting for it.
  49  */
  50 #define KICK_READER_LOOPS       10
  51
  52 /*
  53  * Active attempts to check for reader Q.S. before calling futex().
  54  */
  55 #define RCU_QS_ACTIVE_ATTEMPTS 100
  56
  57 /* If the headers do not support membarrier system call, fall back on RCU_MB */
  58 #ifdef __NR_membarrier
  59 # define membarrier(...)                syscall(__NR_membarrier, __VA_ARGS__)
  60 #else
  61 # define membarrier(...)                -ENOSYS
  62 #endif
  63
  64 enum membarrier_cmd {
  65         MEMBARRIER_CMD_QUERY                            = 0,
  66         MEMBARRIER_CMD_SHARED                           = (1 << 0),
  67         /* reserved for MEMBARRIER_CMD_SHARED_EXPEDITED (1 << 1) */
  68         /* reserved for MEMBARRIER_CMD_PRIVATE (1 << 2) */
  69         MEMBARRIER_CMD_PRIVATE_EXPEDITED                = (1 << 3),
  70         MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED       = (1 << 4),
  71 };
  72
  73 #ifdef RCU_MEMBARRIER
  74 static int init_done;
  75 static int urcu_memb_has_sys_membarrier_private_expedited;
  76
  77 #ifndef CONFIG_RCU_FORCE_SYS_MEMBARRIER
  78 /*
  79  * Explicitly initialize to zero because we can't alias a non-static
  80  * uninitialized variable.
  81  */
  82 int urcu_memb_has_sys_membarrier = 0;
  83 #endif
  84
  85 void __attribute__((constructor)) rcu_init(void);
  86 #endif
  87
  88 #ifdef RCU_MB
  89 void rcu_init(void)
  90 {
  91 }
  92 #endif
  93
  94 #ifdef RCU_SIGNAL
  95 static int init_done;
  96
  97 void __attribute__((constructor)) rcu_init(void);
  98
  99 static DEFINE_URCU_TLS(int, rcu_signal_was_blocked);
 100 #endif
 101
 102 void __attribute__((destructor)) rcu_exit(void);
 103 static void urcu_call_rcu_exit(void);
 104
 105 /*
 106  * rcu_gp_lock ensures mutual exclusion between threads calling
 107  * synchronize_rcu().
 108  */
 109 static pthread_mutex_t rcu_gp_lock = PTHREAD_MUTEX_INITIALIZER;
 110 /*
 111  * rcu_registry_lock ensures mutual exclusion between threads
 112  * registering and unregistering themselves to/from the registry, and
 113  * with threads reading that registry from synchronize_rcu(). However,
 114  * this lock is not held all the way through the completion of awaiting
 115  * for the grace period. It is sporadically released between iterations
 116  * on the registry.
 117  * rcu_registry_lock may nest inside rcu_gp_lock.
 118  */
 119 static pthread_mutex_t rcu_registry_lock = PTHREAD_MUTEX_INITIALIZER;
 120 struct urcu_gp rcu_gp = { .ctr = URCU_GP_COUNT };
 121
 122 /*
 123  * Written to only by each individual reader. Read by both the reader and the
 124  * writers.
 125  */
 126 DEFINE_URCU_TLS(struct urcu_reader, rcu_reader);
 127
 128 static CDS_LIST_HEAD(registry);
 129
 130 /*
 131  * Queue keeping threads awaiting to wait for a grace period. Contains
 132  * struct gp_waiters_thread objects.
 133  */
 134 static DEFINE_URCU_WAIT_QUEUE(gp_waiters);
 135
 136 static void mutex_lock(pthread_mutex_t *mutex)
 137 {
 138         int ret;
 139
 140 #ifndef DISTRUST_SIGNALS_EXTREME
 141         ret = pthread_mutex_lock(mutex);
 142         if (ret)
 143                 urcu_die(ret);
 144 #else /* #ifndef DISTRUST_SIGNALS_EXTREME */
 145         while ((ret = pthread_mutex_trylock(mutex)) != 0) {
 146                 if (ret != EBUSY && ret != EINTR)
 147                         urcu_die(ret);
 148                 if (CMM_LOAD_SHARED(URCU_TLS(rcu_reader).need_mb)) {
 149                         cmm_smp_mb();
 150                         _CMM_STORE_SHARED(URCU_TLS(rcu_reader).need_mb, 0);
 151                         cmm_smp_mb();
 152                 }
 153                 (void) poll(NULL, 0, 10);
 154         }
 155 #endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */
 156 }
 157
 158 static void mutex_unlock(pthread_mutex_t *mutex)
 159 {
 160         int ret;
 161
 162         ret = pthread_mutex_unlock(mutex);
 163         if (ret)
 164                 urcu_die(ret);
 165 }
 166
 167 #ifdef RCU_MEMBARRIER
 168 static void smp_mb_master(void)
 169 {
 170         if (caa_likely(urcu_memb_has_sys_membarrier)) {
 171                 if (membarrier(urcu_memb_has_sys_membarrier_private_expedited ?
 172                                 MEMBARRIER_CMD_PRIVATE_EXPEDITED :
 173                                 MEMBARRIER_CMD_SHARED, 0))
 174                         urcu_die(errno);
 175         } else {
 176                 cmm_smp_mb();
 177         }
 178 }
 179 #endif
 180
 181 #ifdef RCU_MB
 182 static void smp_mb_master(void)
 183 {
 184         cmm_smp_mb();
 185 }
 186 #endif
 187
 188 #ifdef RCU_SIGNAL
 189 static void force_mb_all_readers(void)
 190 {
 191         struct urcu_reader *index;
 192
 193         /*
 194          * Ask for each threads to execute a cmm_smp_mb() so we can consider the
 195          * compiler barriers around rcu read lock as real memory barriers.
 196          */
 197         if (cds_list_empty(&registry))
 198                 return;
 199         /*
 200          * pthread_kill has a cmm_smp_mb(). But beware, we assume it performs
 201          * a cache flush on architectures with non-coherent cache. Let's play
 202          * safe and don't assume anything : we use cmm_smp_mc() to make sure the
 203          * cache flush is enforced.
 204          */
 205         cds_list_for_each_entry(index, &registry, node) {
 206                 CMM_STORE_SHARED(index->need_mb, 1);
 207                 pthread_kill(index->tid, SIGRCU);
 208         }
 209         /*
 210          * Wait for sighandler (and thus mb()) to execute on every thread.
 211          *
 212          * Note that the pthread_kill() will never be executed on systems
 213          * that correctly deliver signals in a timely manner.  However, it
 214          * is not uncommon for kernels to have bugs that can result in
 215          * lost or unduly delayed signals.
 216          *
 217          * If you are seeing the below pthread_kill() executing much at
 218          * all, we suggest testing the underlying kernel and filing the
 219          * relevant bug report.  For Linux kernels, we recommend getting
 220          * the Linux Test Project (LTP).
 221          */
 222         cds_list_for_each_entry(index, &registry, node) {
 223                 while (CMM_LOAD_SHARED(index->need_mb)) {
 224                         pthread_kill(index->tid, SIGRCU);
 225                         (void) poll(NULL, 0, 1);
 226                 }
 227         }
 228         cmm_smp_mb();   /* read ->need_mb before ending the barrier */
 229 }
 230
 231 static void smp_mb_master(void)
 232 {
 233         force_mb_all_readers();
 234 }
 235 #endif /* #ifdef RCU_SIGNAL */
 236
 237 /*
 238  * synchronize_rcu() waiting. Single thread.
 239  * Always called with rcu_registry lock held. Releases this lock and
 240  * grabs it again. Holds the lock when it returns.
 241  */
 242 static void wait_gp(void)
 243 {
 244         /*
 245          * Read reader_gp before read futex. smp_mb_master() needs to
 246          * be called with the rcu registry lock held in RCU_SIGNAL
 247          * flavor.
 248          */
 249         smp_mb_master();
 250         /* Temporarily unlock the registry lock. */
 251         mutex_unlock(&rcu_registry_lock);
 252         while (uatomic_read(&rcu_gp.futex) == -1) {
 253                 if (!futex_async(&rcu_gp.futex, FUTEX_WAIT, -1, NULL, NULL, 0)) {
 254                         /*
 255                          * Prior queued wakeups queued by unrelated code
 256                          * using the same address can cause futex wait to
 257                          * return 0 even through the futex value is still
 258                          * -1 (spurious wakeups). Check the value again
 259                          * in user-space to validate whether it really
 260                          * differs from -1.
 261                          */
 262                         continue;
 263                 }
 264                 switch (errno) {
 265                 case EAGAIN:
 266                         /* Value already changed. */
 267                         goto end;
 268                 case EINTR:
 269                         /* Retry if interrupted by signal. */
 270                         break;  /* Get out of switch. Check again. */
 271                 default:
 272                         /* Unexpected error. */
 273                         urcu_die(errno);
 274                 }
 275         }
 276 end:
 277         /*
 278          * Re-lock the registry lock before the next loop.
 279          */
 280         mutex_lock(&rcu_registry_lock);
 281 }
 282
 283 /*
 284  * Always called with rcu_registry lock held. Releases this lock between
 285  * iterations and grabs it again. Holds the lock when it returns.
 286  */
 287 static void wait_for_readers(struct cds_list_head *input_readers,
 288                         struct cds_list_head *cur_snap_readers,
 289                         struct cds_list_head *qsreaders)
 290 {
 291         unsigned int wait_loops = 0;
 292         struct urcu_reader *index, *tmp;
 293 #ifdef HAS_INCOHERENT_CACHES
 294         unsigned int wait_gp_loops = 0;
 295 #endif /* HAS_INCOHERENT_CACHES */
 296
 297         /*
 298          * Wait for each thread URCU_TLS(rcu_reader).ctr to either
 299          * indicate quiescence (not nested), or observe the current
 300          * rcu_gp.ctr value.
 301          */
 302         for (;;) {
 303                 if (wait_loops < RCU_QS_ACTIVE_ATTEMPTS)
 304                         wait_loops++;
 305                 if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS) {
 306                         uatomic_dec(&rcu_gp.futex);
 307                         /* Write futex before read reader_gp */
 308                         smp_mb_master();
 309                 }
 310
 311                 cds_list_for_each_entry_safe(index, tmp, input_readers, node) {
 312                         switch (urcu_common_reader_state(&rcu_gp, &index->ctr)) {
 313                         case URCU_READER_ACTIVE_CURRENT:
 314                                 if (cur_snap_readers) {
 315                                         cds_list_move(&index->node,
 316                                                 cur_snap_readers);
 317                                         break;
 318                                 }
 319                                 /* Fall-through */
 320                         case URCU_READER_INACTIVE:
 321                                 cds_list_move(&index->node, qsreaders);
 322                                 break;
 323                         case URCU_READER_ACTIVE_OLD:
 324                                 /*
 325                                  * Old snapshot. Leaving node in
 326                                  * input_readers will make us busy-loop
 327                                  * until the snapshot becomes current or
 328                                  * the reader becomes inactive.
 329                                  */
 330                                 break;
 331                         }
 332                 }
 333
 334 #ifndef HAS_INCOHERENT_CACHES
 335                 if (cds_list_empty(input_readers)) {
 336                         if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS) {
 337                                 /* Read reader_gp before write futex */
 338                                 smp_mb_master();
 339                                 uatomic_set(&rcu_gp.futex, 0);
 340                         }
 341                         break;
 342                 } else {
 343                         if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS) {
 344                                 /* wait_gp unlocks/locks registry lock. */
 345                                 wait_gp();
 346                         } else {
 347                                 /* Temporarily unlock the registry lock. */
 348                                 mutex_unlock(&rcu_registry_lock);
 349                                 caa_cpu_relax();
 350                                 /*
 351                                  * Re-lock the registry lock before the
 352                                  * next loop.
 353                                  */
 354                                 mutex_lock(&rcu_registry_lock);
 355                         }
 356                 }
 357 #else /* #ifndef HAS_INCOHERENT_CACHES */
 358                 /*
 359                  * BUSY-LOOP. Force the reader thread to commit its
 360                  * URCU_TLS(rcu_reader).ctr update to memory if we wait
 361                  * for too long.
 362                  */
 363                 if (cds_list_empty(input_readers)) {
 364                         if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS) {
 365                                 /* Read reader_gp before write futex */
 366                                 smp_mb_master();
 367                                 uatomic_set(&rcu_gp.futex, 0);
 368                         }
 369                         break;
 370                 } else {
 371                         if (wait_gp_loops == KICK_READER_LOOPS) {
 372                                 smp_mb_master();
 373                                 wait_gp_loops = 0;
 374                         }
 375                         if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS) {
 376                                 /* wait_gp unlocks/locks registry lock. */
 377                                 wait_gp();
 378                                 wait_gp_loops++;
 379                         } else {
 380                                 /* Temporarily unlock the registry lock. */
 381                                 mutex_unlock(&rcu_registry_lock);
 382                                 caa_cpu_relax();
 383                                 /*
 384                                  * Re-lock the registry lock before the
 385                                  * next loop.
 386                                  */
 387                                 mutex_lock(&rcu_registry_lock);
 388                         }
 389                 }
 390 #endif /* #else #ifndef HAS_INCOHERENT_CACHES */
 391         }
 392 }
 393
 394 void synchronize_rcu(void)
 395 {
 396         CDS_LIST_HEAD(cur_snap_readers);
 397         CDS_LIST_HEAD(qsreaders);
 398         DEFINE_URCU_WAIT_NODE(wait, URCU_WAIT_WAITING);
 399         struct urcu_waiters waiters;
 400
 401         /*
 402          * Add ourself to gp_waiters queue of threads awaiting to wait
 403          * for a grace period. Proceed to perform the grace period only
 404          * if we are the first thread added into the queue.
 405          * The implicit memory barrier before urcu_wait_add()
 406          * orders prior memory accesses of threads put into the wait
 407          * queue before their insertion into the wait queue.
 408          */
 409         if (urcu_wait_add(&gp_waiters, &wait) != 0) {
 410                 /* Not first in queue: will be awakened by another thread. */
 411                 urcu_adaptative_busy_wait(&wait);
 412                 /* Order following memory accesses after grace period. */
 413                 cmm_smp_mb();
 414                 return;
 415         }
 416         /* We won't need to wake ourself up */
 417         urcu_wait_set_state(&wait, URCU_WAIT_RUNNING);
 418
 419         mutex_lock(&rcu_gp_lock);
 420
 421         /*
 422          * Move all waiters into our local queue.
 423          */
 424         urcu_move_waiters(&waiters, &gp_waiters);
 425
 426         mutex_lock(&rcu_registry_lock);
 427
 428         if (cds_list_empty(&registry))
 429                 goto out;
 430
 431         /*
 432          * All threads should read qparity before accessing data structure
 433          * where new ptr points to. Must be done within rcu_registry_lock
 434          * because it iterates on reader threads.
 435          */
 436         /* Write new ptr before changing the qparity */
 437         smp_mb_master();
 438
 439         /*
 440          * Wait for readers to observe original parity or be quiescent.
 441          * wait_for_readers() can release and grab again rcu_registry_lock
 442          * internally.
 443          */
 444         wait_for_readers(&registry, &cur_snap_readers, &qsreaders);
 445
 446         /*
 447          * Must finish waiting for quiescent state for original parity before
 448          * committing next rcu_gp.ctr update to memory. Failure to do so could
 449          * result in the writer waiting forever while new readers are always
 450          * accessing data (no progress).  Enforce compiler-order of load
 451          * URCU_TLS(rcu_reader).ctr before store to rcu_gp.ctr.
 452          */
 453         cmm_barrier();
 454
 455         /*
 456          * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
 457          * model easier to understand. It does not have a big performance impact
 458          * anyway, given this is the write-side.
 459          */
 460         cmm_smp_mb();
 461
 462         /* Switch parity: 0 -> 1, 1 -> 0 */
 463         CMM_STORE_SHARED(rcu_gp.ctr, rcu_gp.ctr ^ URCU_GP_CTR_PHASE);
 464
 465         /*
 466          * Must commit rcu_gp.ctr update to memory before waiting for quiescent
 467          * state. Failure to do so could result in the writer waiting forever
 468          * while new readers are always accessing data (no progress). Enforce
 469          * compiler-order of store to rcu_gp.ctr before load rcu_reader ctr.
 470          */
 471         cmm_barrier();
 472
 473         /*
 474          *
 475          * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
 476          * model easier to understand. It does not have a big performance impact
 477          * anyway, given this is the write-side.
 478          */
 479         cmm_smp_mb();
 480
 481         /*
 482          * Wait for readers to observe new parity or be quiescent.
 483          * wait_for_readers() can release and grab again rcu_registry_lock
 484          * internally.
 485          */
 486         wait_for_readers(&cur_snap_readers, NULL, &qsreaders);
 487
 488         /*
 489          * Put quiescent reader list back into registry.
 490          */
 491         cds_list_splice(&qsreaders, &registry);
 492
 493         /*
 494          * Finish waiting for reader threads before letting the old ptr
 495          * being freed. Must be done within rcu_registry_lock because it
 496          * iterates on reader threads.
 497          */
 498         smp_mb_master();
 499 out:
 500         mutex_unlock(&rcu_registry_lock);
 501         mutex_unlock(&rcu_gp_lock);
 502
 503         /*
 504          * Wakeup waiters only after we have completed the grace period
 505          * and have ensured the memory barriers at the end of the grace
 506          * period have been issued.
 507          */
 508         urcu_wake_all_waiters(&waiters);
 509 }
 510
 511 /*
 512  * library wrappers to be used by non-LGPL compatible source code.
 513  */
 514
 515 void rcu_read_lock(void)
 516 {
 517         _rcu_read_lock();
 518 }
 519
 520 void rcu_read_unlock(void)
 521 {
 522         _rcu_read_unlock();
 523 }
 524
 525 int rcu_read_ongoing(void)
 526 {
 527         return _rcu_read_ongoing();
 528 }
 529
 530 #ifdef RCU_SIGNAL
 531 /*
 532  * Make sure the signal used by the urcu-signal flavor is unblocked
 533  * while the thread is registered.
 534  */
 535 static
 536 void urcu_signal_unblock(void)
 537 {
 538         sigset_t mask, oldmask;
 539         int ret;
 540
 541         ret = sigemptyset(&mask);
 542         urcu_posix_assert(!ret);
 543         ret = sigaddset(&mask, SIGRCU);
 544         urcu_posix_assert(!ret);
 545         ret = pthread_sigmask(SIG_UNBLOCK, &mask, &oldmask);
 546         urcu_posix_assert(!ret);
 547         URCU_TLS(rcu_signal_was_blocked) = sigismember(&oldmask, SIGRCU);
 548 }
 549
 550 static
 551 void urcu_signal_restore(void)
 552 {
 553         sigset_t mask;
 554         int ret;
 555
 556         if (!URCU_TLS(rcu_signal_was_blocked))
 557                 return;
 558         ret = sigemptyset(&mask);
 559         urcu_posix_assert(!ret);
 560         ret = sigaddset(&mask, SIGRCU);
 561         urcu_posix_assert(!ret);
 562         ret = pthread_sigmask(SIG_BLOCK, &mask, NULL);
 563         urcu_posix_assert(!ret);
 564 }
 565 #else
 566 static
 567 void urcu_signal_unblock(void) { }
 568 static
 569 void urcu_signal_restore(void) { }
 570 #endif
 571
 572 void rcu_register_thread(void)
 573 {
 574         urcu_signal_unblock();
 575
 576         URCU_TLS(rcu_reader).tid = pthread_self();
 577         urcu_posix_assert(URCU_TLS(rcu_reader).need_mb == 0);
 578         urcu_posix_assert(!(URCU_TLS(rcu_reader).ctr & URCU_GP_CTR_NEST_MASK));
 579
 580         mutex_lock(&rcu_registry_lock);
 581         urcu_posix_assert(!URCU_TLS(rcu_reader).registered);
 582         URCU_TLS(rcu_reader).registered = 1;
 583         rcu_init();     /* In case gcc does not support constructor attribute */
 584         cds_list_add(&URCU_TLS(rcu_reader).node, &registry);
 585         mutex_unlock(&rcu_registry_lock);
 586 }
 587
 588 void rcu_unregister_thread(void)
 589 {
 590         mutex_lock(&rcu_registry_lock);
 591         urcu_posix_assert(URCU_TLS(rcu_reader).registered);
 592         URCU_TLS(rcu_reader).registered = 0;
 593         cds_list_del(&URCU_TLS(rcu_reader).node);
 594         mutex_unlock(&rcu_registry_lock);
 595
 596         urcu_signal_restore();
 597 }
 598
 599 #ifdef RCU_MEMBARRIER
 600
 601 #ifdef CONFIG_RCU_FORCE_SYS_MEMBARRIER
 602 static
 603 void rcu_sys_membarrier_status(bool available)
 604 {
 605         if (!available)
 606                 abort();
 607 }
 608 #else
 609 static
 610 void rcu_sys_membarrier_status(bool available)
 611 {
 612         if (!available)
 613                 return;
 614         urcu_memb_has_sys_membarrier = 1;
 615 }
 616 #endif
 617
 618 static
 619 void rcu_sys_membarrier_init(void)
 620 {
 621         bool available = false;
 622         int mask;
 623
 624         mask = membarrier(MEMBARRIER_CMD_QUERY, 0);
 625         if (mask >= 0) {
 626                 if (mask & MEMBARRIER_CMD_PRIVATE_EXPEDITED) {
 627                         if (membarrier(MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED, 0))
 628                                 urcu_die(errno);
 629                         urcu_memb_has_sys_membarrier_private_expedited = 1;
 630                         available = true;
 631                 } else if (mask & MEMBARRIER_CMD_SHARED) {
 632                         available = true;
 633                 }
 634         }
 635         rcu_sys_membarrier_status(available);
 636 }
 637
 638 void rcu_init(void)
 639 {
 640         if (init_done)
 641                 return;
 642         init_done = 1;
 643         rcu_sys_membarrier_init();
 644 }
 645 #endif
 646
 647 #ifdef RCU_SIGNAL
 648 static void sigrcu_handler(int signo __attribute__((unused)),
 649                 siginfo_t *siginfo __attribute__((unused)),
 650                 void *context __attribute__((unused)))
 651 {
 652         /*
 653          * Executing this cmm_smp_mb() is the only purpose of this signal handler.
 654          * It punctually promotes cmm_barrier() into cmm_smp_mb() on every thread it is
 655          * executed on.
 656          */
 657         cmm_smp_mb();
 658         _CMM_STORE_SHARED(URCU_TLS(rcu_reader).need_mb, 0);
 659         cmm_smp_mb();
 660 }
 661
 662 /*
 663  * rcu_init constructor. Called when the library is linked, but also when
 664  * reader threads are calling rcu_register_thread().
 665  * Should only be called by a single thread at a given time. This is ensured by
 666  * holing the rcu_registry_lock from rcu_register_thread() or by running
 667  * at library load time, which should not be executed by multiple
 668  * threads nor concurrently with rcu_register_thread() anyway.
 669  */
 670 void rcu_init(void)
 671 {
 672         struct sigaction act;
 673         int ret;
 674
 675         if (init_done)
 676                 return;
 677         init_done = 1;
 678
 679         act.sa_sigaction = sigrcu_handler;
 680         act.sa_flags = SA_SIGINFO | SA_RESTART;
 681         sigemptyset(&act.sa_mask);
 682         ret = sigaction(SIGRCU, &act, NULL);
 683         if (ret)
 684                 urcu_die(errno);
 685 }
 686
 687 /*
 688  * Don't unregister the SIGRCU signal handler anymore, because
 689  * call_rcu threads could still be using it shortly before the
 690  * application exits.
 691  * Assertion disabled because call_rcu threads are now rcu
 692  * readers, and left running at exit.
 693  * urcu_posix_assert(cds_list_empty(&registry));
 694  */
 695
 696 #endif /* #ifdef RCU_SIGNAL */
 697
 698 void rcu_exit(void)
 699 {
 700         urcu_call_rcu_exit();
 701 }
 702
 703 DEFINE_RCU_FLAVOR(rcu_flavor);
 704
 705 #include "urcu-call-rcu-impl.h"
 706 #include "urcu-defer-impl.h"
 707 #include "urcu-poll-impl.h"