src/urcu-bp.c

   1 /*
   2  * urcu-bp.c
   3  *
   4  * Userspace RCU library, "bulletproof" version.
   5  *
   6  * Copyright (c) 2009 Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
   7  * Copyright (c) 2009 Paul E. McKenney, IBM Corporation.
   8  *
   9  * This library is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU Lesser General Public
  11  * License as published by the Free Software Foundation; either
  12  * version 2.1 of the License, or (at your option) any later version.
  13  *
  14  * This library is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * Lesser General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU Lesser General Public
  20  * License along with this library; if not, write to the Free Software
  21  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  22  *
  23  * IBM's contributions to this file may be relicensed under LGPLv2 or later.
  24  */
  25
  26 #define URCU_NO_COMPAT_IDENTIFIERS
  27 #define _LGPL_SOURCE
  28 #include <stdio.h>
  29 #include <pthread.h>
  30 #include <signal.h>
  31 #include <assert.h>
  32 #include <stdlib.h>
  33 #include <string.h>
  34 #include <errno.h>
  35 #include <poll.h>
  36 #include <unistd.h>
  37 #include <stdbool.h>
  38 #include <sys/mman.h>
  39
  40 #include <urcu/config.h>
  41 #include <urcu/arch.h>
  42 #include <urcu/wfcqueue.h>
  43 #include <urcu/map/urcu-bp.h>
  44 #include <urcu/static/urcu-bp.h>
  45 #include <urcu/pointer.h>
  46 #include <urcu/tls-compat.h>
  47
  48 #include "urcu-die.h"
  49 #include "urcu-utils.h"
  50
  51 #define URCU_API_MAP
  52 /* Do not #define _LGPL_SOURCE to ensure we can emit the wrapper symbols */
  53 #undef _LGPL_SOURCE
  54 #include <urcu/urcu-bp.h>
  55 #define _LGPL_SOURCE
  56
  57 #ifndef MAP_ANONYMOUS
  58 #define MAP_ANONYMOUS MAP_ANON
  59 #endif
  60
  61 #ifdef __linux__
  62 static
  63 void *mremap_wrapper(void *old_address, size_t old_size,
  64                 size_t new_size, int flags)
  65 {
  66         return mremap(old_address, old_size, new_size, flags);
  67 }
  68 #else
  69
  70 #define MREMAP_MAYMOVE  1
  71 #define MREMAP_FIXED    2
  72
  73 /*
  74  * mremap wrapper for non-Linux systems not allowing MAYMOVE.
  75  * This is not generic.
  76 */
  77 static
  78 void *mremap_wrapper(void *old_address, size_t old_size,
  79                 size_t new_size, int flags)
  80 {
  81         assert(!(flags & MREMAP_MAYMOVE));
  82
  83         return MAP_FAILED;
  84 }
  85 #endif
  86
  87 /* Sleep delay in ms */
  88 #define RCU_SLEEP_DELAY_MS      10
  89 #define INIT_NR_THREADS         8
  90 #define ARENA_INIT_ALLOC                \
  91         sizeof(struct registry_chunk)   \
  92         + INIT_NR_THREADS * sizeof(struct urcu_bp_reader)
  93
  94 /*
  95  * Active attempts to check for reader Q.S. before calling sleep().
  96  */
  97 #define RCU_QS_ACTIVE_ATTEMPTS 100
  98
  99 static
 100 int urcu_bp_refcount;
 101
 102 /* If the headers do not support membarrier system call, fall back smp_mb. */
 103 #ifdef __NR_membarrier
 104 # define membarrier(...)                syscall(__NR_membarrier, __VA_ARGS__)
 105 #else
 106 # define membarrier(...)                -ENOSYS
 107 #endif
 108
 109 enum membarrier_cmd {
 110         MEMBARRIER_CMD_QUERY                            = 0,
 111         MEMBARRIER_CMD_SHARED                           = (1 << 0),
 112         /* reserved for MEMBARRIER_CMD_SHARED_EXPEDITED (1 << 1) */
 113         /* reserved for MEMBARRIER_CMD_PRIVATE (1 << 2) */
 114         MEMBARRIER_CMD_PRIVATE_EXPEDITED                = (1 << 3),
 115         MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED       = (1 << 4),
 116 };
 117
 118 static
 119 void __attribute__((constructor)) _urcu_bp_init(void);
 120 static
 121 void __attribute__((destructor)) urcu_bp_exit(void);
 122
 123 #ifndef CONFIG_RCU_FORCE_SYS_MEMBARRIER
 124 int urcu_bp_has_sys_membarrier;
 125 #endif
 126
 127 /*
 128  * rcu_gp_lock ensures mutual exclusion between threads calling
 129  * synchronize_rcu().
 130  */
 131 static pthread_mutex_t rcu_gp_lock = PTHREAD_MUTEX_INITIALIZER;
 132 /*
 133  * rcu_registry_lock ensures mutual exclusion between threads
 134  * registering and unregistering themselves to/from the registry, and
 135  * with threads reading that registry from synchronize_rcu(). However,
 136  * this lock is not held all the way through the completion of awaiting
 137  * for the grace period. It is sporadically released between iterations
 138  * on the registry.
 139  * rcu_registry_lock may nest inside rcu_gp_lock.
 140  */
 141 static pthread_mutex_t rcu_registry_lock = PTHREAD_MUTEX_INITIALIZER;
 142
 143 static pthread_mutex_t init_lock = PTHREAD_MUTEX_INITIALIZER;
 144 static int initialized;
 145
 146 static pthread_key_t urcu_bp_key;
 147
 148 struct urcu_bp_gp urcu_bp_gp = { .ctr = URCU_BP_GP_COUNT };
 149 URCU_ATTR_ALIAS("urcu_bp_gp") extern struct urcu_bp_gp rcu_gp_bp;
 150
 151 /*
 152  * Pointer to registry elements. Written to only by each individual reader. Read
 153  * by both the reader and the writers.
 154  */
 155 DEFINE_URCU_TLS(struct urcu_bp_reader *, urcu_bp_reader);
 156 DEFINE_URCU_TLS_ALIAS(struct urcu_bp_reader *, urcu_bp_reader, rcu_reader_bp);
 157
 158 static CDS_LIST_HEAD(registry);
 159
 160 struct registry_chunk {
 161         size_t data_len;                /* data length */
 162         size_t used;                    /* amount of data used */
 163         struct cds_list_head node;      /* chunk_list node */
 164         char data[];
 165 };
 166
 167 struct registry_arena {
 168         struct cds_list_head chunk_list;
 169 };
 170
 171 static struct registry_arena registry_arena = {
 172         .chunk_list = CDS_LIST_HEAD_INIT(registry_arena.chunk_list),
 173 };
 174
 175 /* Saved fork signal mask, protected by rcu_gp_lock */
 176 static sigset_t saved_fork_signal_mask;
 177
 178 static void mutex_lock(pthread_mutex_t *mutex)
 179 {
 180         int ret;
 181
 182 #ifndef DISTRUST_SIGNALS_EXTREME
 183         ret = pthread_mutex_lock(mutex);
 184         if (ret)
 185                 urcu_die(ret);
 186 #else /* #ifndef DISTRUST_SIGNALS_EXTREME */
 187         while ((ret = pthread_mutex_trylock(mutex)) != 0) {
 188                 if (ret != EBUSY && ret != EINTR)
 189                         urcu_die(ret);
 190                 poll(NULL,0,10);
 191         }
 192 #endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */
 193 }
 194
 195 static void mutex_unlock(pthread_mutex_t *mutex)
 196 {
 197         int ret;
 198
 199         ret = pthread_mutex_unlock(mutex);
 200         if (ret)
 201                 urcu_die(ret);
 202 }
 203
 204 static void smp_mb_master(void)
 205 {
 206         if (caa_likely(urcu_bp_has_sys_membarrier)) {
 207                 if (membarrier(MEMBARRIER_CMD_PRIVATE_EXPEDITED, 0))
 208                         urcu_die(errno);
 209         } else {
 210                 cmm_smp_mb();
 211         }
 212 }
 213
 214 /*
 215  * Always called with rcu_registry lock held. Releases this lock between
 216  * iterations and grabs it again. Holds the lock when it returns.
 217  */
 218 static void wait_for_readers(struct cds_list_head *input_readers,
 219                         struct cds_list_head *cur_snap_readers,
 220                         struct cds_list_head *qsreaders)
 221 {
 222         unsigned int wait_loops = 0;
 223         struct urcu_bp_reader *index, *tmp;
 224
 225         /*
 226          * Wait for each thread URCU_TLS(urcu_bp_reader).ctr to either
 227          * indicate quiescence (not nested), or observe the current
 228          * rcu_gp.ctr value.
 229          */
 230         for (;;) {
 231                 if (wait_loops < RCU_QS_ACTIVE_ATTEMPTS)
 232                         wait_loops++;
 233
 234                 cds_list_for_each_entry_safe(index, tmp, input_readers, node) {
 235                         switch (urcu_bp_reader_state(&index->ctr)) {
 236                         case URCU_BP_READER_ACTIVE_CURRENT:
 237                                 if (cur_snap_readers) {
 238                                         cds_list_move(&index->node,
 239                                                 cur_snap_readers);
 240                                         break;
 241                                 }
 242                                 /* Fall-through */
 243                         case URCU_BP_READER_INACTIVE:
 244                                 cds_list_move(&index->node, qsreaders);
 245                                 break;
 246                         case URCU_BP_READER_ACTIVE_OLD:
 247                                 /*
 248                                  * Old snapshot. Leaving node in
 249                                  * input_readers will make us busy-loop
 250                                  * until the snapshot becomes current or
 251                                  * the reader becomes inactive.
 252                                  */
 253                                 break;
 254                         }
 255                 }
 256
 257                 if (cds_list_empty(input_readers)) {
 258                         break;
 259                 } else {
 260                         /* Temporarily unlock the registry lock. */
 261                         mutex_unlock(&rcu_registry_lock);
 262                         if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS)
 263                                 (void) poll(NULL, 0, RCU_SLEEP_DELAY_MS);
 264                         else
 265                                 caa_cpu_relax();
 266                         /* Re-lock the registry lock before the next loop. */
 267                         mutex_lock(&rcu_registry_lock);
 268                 }
 269         }
 270 }
 271
 272 void urcu_bp_synchronize_rcu(void)
 273 {
 274         CDS_LIST_HEAD(cur_snap_readers);
 275         CDS_LIST_HEAD(qsreaders);
 276         sigset_t newmask, oldmask;
 277         int ret;
 278
 279         ret = sigfillset(&newmask);
 280         assert(!ret);
 281         ret = pthread_sigmask(SIG_BLOCK, &newmask, &oldmask);
 282         assert(!ret);
 283
 284         mutex_lock(&rcu_gp_lock);
 285
 286         mutex_lock(&rcu_registry_lock);
 287
 288         if (cds_list_empty(&registry))
 289                 goto out;
 290
 291         /* All threads should read qparity before accessing data structure
 292          * where new ptr points to. */
 293         /* Write new ptr before changing the qparity */
 294         smp_mb_master();
 295
 296         /*
 297          * Wait for readers to observe original parity or be quiescent.
 298          * wait_for_readers() can release and grab again rcu_registry_lock
 299          * interally.
 300          */
 301         wait_for_readers(&registry, &cur_snap_readers, &qsreaders);
 302
 303         /*
 304          * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
 305          * model easier to understand. It does not have a big performance impact
 306          * anyway, given this is the write-side.
 307          */
 308         cmm_smp_mb();
 309
 310         /* Switch parity: 0 -> 1, 1 -> 0 */
 311         CMM_STORE_SHARED(rcu_gp.ctr, rcu_gp.ctr ^ URCU_BP_GP_CTR_PHASE);
 312
 313         /*
 314          * Must commit qparity update to memory before waiting for other parity
 315          * quiescent state. Failure to do so could result in the writer waiting
 316          * forever while new readers are always accessing data (no progress).
 317          * Ensured by CMM_STORE_SHARED and CMM_LOAD_SHARED.
 318          */
 319
 320         /*
 321          * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
 322          * model easier to understand. It does not have a big performance impact
 323          * anyway, given this is the write-side.
 324          */
 325         cmm_smp_mb();
 326
 327         /*
 328          * Wait for readers to observe new parity or be quiescent.
 329          * wait_for_readers() can release and grab again rcu_registry_lock
 330          * interally.
 331          */
 332         wait_for_readers(&cur_snap_readers, NULL, &qsreaders);
 333
 334         /*
 335          * Put quiescent reader list back into registry.
 336          */
 337         cds_list_splice(&qsreaders, &registry);
 338
 339         /*
 340          * Finish waiting for reader threads before letting the old ptr being
 341          * freed.
 342          */
 343         smp_mb_master();
 344 out:
 345         mutex_unlock(&rcu_registry_lock);
 346         mutex_unlock(&rcu_gp_lock);
 347         ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL);
 348         assert(!ret);
 349 }
 350 URCU_ATTR_ALIAS("urcu_bp_synchronize_rcu") void synchronize_rcu_bp();
 351
 352 /*
 353  * library wrappers to be used by non-LGPL compatible source code.
 354  */
 355
 356 void urcu_bp_read_lock(void)
 357 {
 358         _urcu_bp_read_lock();
 359 }
 360 URCU_ATTR_ALIAS("urcu_bp_read_lock") void rcu_read_lock_bp();
 361
 362 void urcu_bp_read_unlock(void)
 363 {
 364         _urcu_bp_read_unlock();
 365 }
 366 URCU_ATTR_ALIAS("urcu_bp_read_unlock") void rcu_read_unlock_bp();
 367
 368 int urcu_bp_read_ongoing(void)
 369 {
 370         return _urcu_bp_read_ongoing();
 371 }
 372 URCU_ATTR_ALIAS("urcu_bp_read_ongoing") int rcu_read_ongoing_bp();
 373
 374 /*
 375  * Only grow for now. If empty, allocate a ARENA_INIT_ALLOC sized chunk.
 376  * Else, try expanding the last chunk. If this fails, allocate a new
 377  * chunk twice as big as the last chunk.
 378  * Memory used by chunks _never_ moves. A chunk could theoretically be
 379  * freed when all "used" slots are released, but we don't do it at this
 380  * point.
 381  */
 382 static
 383 void expand_arena(struct registry_arena *arena)
 384 {
 385         struct registry_chunk *new_chunk, *last_chunk;
 386         size_t old_chunk_len, new_chunk_len;
 387
 388         /* No chunk. */
 389         if (cds_list_empty(&arena->chunk_list)) {
 390                 assert(ARENA_INIT_ALLOC >=
 391                         sizeof(struct registry_chunk)
 392                         + sizeof(struct rcu_reader));
 393                 new_chunk_len = ARENA_INIT_ALLOC;
 394                 new_chunk = (struct registry_chunk *) mmap(NULL,
 395                         new_chunk_len,
 396                         PROT_READ | PROT_WRITE,
 397                         MAP_ANONYMOUS | MAP_PRIVATE,
 398                         -1, 0);
 399                 if (new_chunk == MAP_FAILED)
 400                         abort();
 401                 memset(new_chunk, 0, new_chunk_len);
 402                 new_chunk->data_len =
 403                         new_chunk_len - sizeof(struct registry_chunk);
 404                 cds_list_add_tail(&new_chunk->node, &arena->chunk_list);
 405                 return;         /* We're done. */
 406         }
 407
 408         /* Try expanding last chunk. */
 409         last_chunk = cds_list_entry(arena->chunk_list.prev,
 410                 struct registry_chunk, node);
 411         old_chunk_len =
 412                 last_chunk->data_len + sizeof(struct registry_chunk);
 413         new_chunk_len = old_chunk_len << 1;
 414
 415         /* Don't allow memory mapping to move, just expand. */
 416         new_chunk = mremap_wrapper(last_chunk, old_chunk_len,
 417                 new_chunk_len, 0);
 418         if (new_chunk != MAP_FAILED) {
 419                 /* Should not have moved. */
 420                 assert(new_chunk == last_chunk);
 421                 memset((char *) last_chunk + old_chunk_len, 0,
 422                         new_chunk_len - old_chunk_len);
 423                 last_chunk->data_len =
 424                         new_chunk_len - sizeof(struct registry_chunk);
 425                 return;         /* We're done. */
 426         }
 427
 428         /* Remap did not succeed, we need to add a new chunk. */
 429         new_chunk = (struct registry_chunk *) mmap(NULL,
 430                 new_chunk_len,
 431                 PROT_READ | PROT_WRITE,
 432                 MAP_ANONYMOUS | MAP_PRIVATE,
 433                 -1, 0);
 434         if (new_chunk == MAP_FAILED)
 435                 abort();
 436         memset(new_chunk, 0, new_chunk_len);
 437         new_chunk->data_len =
 438                 new_chunk_len - sizeof(struct registry_chunk);
 439         cds_list_add_tail(&new_chunk->node, &arena->chunk_list);
 440 }
 441
 442 static
 443 struct rcu_reader *arena_alloc(struct registry_arena *arena)
 444 {
 445         struct registry_chunk *chunk;
 446         struct rcu_reader *rcu_reader_reg;
 447         int expand_done = 0;    /* Only allow to expand once per alloc */
 448         size_t len = sizeof(struct rcu_reader);
 449
 450 retry:
 451         cds_list_for_each_entry(chunk, &arena->chunk_list, node) {
 452                 if (chunk->data_len - chunk->used < len)
 453                         continue;
 454                 /* Find spot */
 455                 for (rcu_reader_reg = (struct rcu_reader *) &chunk->data[0];
 456                                 rcu_reader_reg < (struct rcu_reader *) &chunk->data[chunk->data_len];
 457                                 rcu_reader_reg++) {
 458                         if (!rcu_reader_reg->alloc) {
 459                                 rcu_reader_reg->alloc = 1;
 460                                 chunk->used += len;
 461                                 return rcu_reader_reg;
 462                         }
 463                 }
 464         }
 465
 466         if (!expand_done) {
 467                 expand_arena(arena);
 468                 expand_done = 1;
 469                 goto retry;
 470         }
 471
 472         return NULL;
 473 }
 474
 475 /* Called with signals off and mutex locked */
 476 static
 477 void add_thread(void)
 478 {
 479         struct rcu_reader *rcu_reader_reg;
 480         int ret;
 481
 482         rcu_reader_reg = arena_alloc(&registry_arena);
 483         if (!rcu_reader_reg)
 484                 abort();
 485         ret = pthread_setspecific(urcu_bp_key, rcu_reader_reg);
 486         if (ret)
 487                 abort();
 488
 489         /* Add to registry */
 490         rcu_reader_reg->tid = pthread_self();
 491         assert(rcu_reader_reg->ctr == 0);
 492         cds_list_add(&rcu_reader_reg->node, &registry);
 493         /*
 494          * Reader threads are pointing to the reader registry. This is
 495          * why its memory should never be relocated.
 496          */
 497         URCU_TLS(urcu_bp_reader) = rcu_reader_reg;
 498 }
 499
 500 /* Called with mutex locked */
 501 static
 502 void cleanup_thread(struct registry_chunk *chunk,
 503                 struct rcu_reader *rcu_reader_reg)
 504 {
 505         rcu_reader_reg->ctr = 0;
 506         cds_list_del(&rcu_reader_reg->node);
 507         rcu_reader_reg->tid = 0;
 508         rcu_reader_reg->alloc = 0;
 509         chunk->used -= sizeof(struct rcu_reader);
 510 }
 511
 512 static
 513 struct registry_chunk *find_chunk(struct rcu_reader *rcu_reader_reg)
 514 {
 515         struct registry_chunk *chunk;
 516
 517         cds_list_for_each_entry(chunk, &registry_arena.chunk_list, node) {
 518                 if (rcu_reader_reg < (struct rcu_reader *) &chunk->data[0])
 519                         continue;
 520                 if (rcu_reader_reg >= (struct rcu_reader *) &chunk->data[chunk->data_len])
 521                         continue;
 522                 return chunk;
 523         }
 524         return NULL;
 525 }
 526
 527 /* Called with signals off and mutex locked */
 528 static
 529 void remove_thread(struct rcu_reader *rcu_reader_reg)
 530 {
 531         cleanup_thread(find_chunk(rcu_reader_reg), rcu_reader_reg);
 532         URCU_TLS(urcu_bp_reader) = NULL;
 533 }
 534
 535 /* Disable signals, take mutex, add to registry */
 536 void urcu_bp_register(void)
 537 {
 538         sigset_t newmask, oldmask;
 539         int ret;
 540
 541         ret = sigfillset(&newmask);
 542         if (ret)
 543                 abort();
 544         ret = pthread_sigmask(SIG_BLOCK, &newmask, &oldmask);
 545         if (ret)
 546                 abort();
 547
 548         /*
 549          * Check if a signal concurrently registered our thread since
 550          * the check in rcu_read_lock().
 551          */
 552         if (URCU_TLS(urcu_bp_reader))
 553                 goto end;
 554
 555         /*
 556          * Take care of early registration before urcu_bp constructor.
 557          */
 558         _urcu_bp_init();
 559
 560         mutex_lock(&rcu_registry_lock);
 561         add_thread();
 562         mutex_unlock(&rcu_registry_lock);
 563 end:
 564         ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL);
 565         if (ret)
 566                 abort();
 567 }
 568 URCU_ATTR_ALIAS("urcu_bp_register") void rcu_bp_register();
 569
 570 void urcu_bp_register_thread(void)
 571 {
 572         if (caa_unlikely(!URCU_TLS(urcu_bp_reader)))
 573                 urcu_bp_register(); /* If not yet registered. */
 574 }
 575
 576 /* Disable signals, take mutex, remove from registry */
 577 static
 578 void urcu_bp_unregister(struct rcu_reader *rcu_reader_reg)
 579 {
 580         sigset_t newmask, oldmask;
 581         int ret;
 582
 583         ret = sigfillset(&newmask);
 584         if (ret)
 585                 abort();
 586         ret = pthread_sigmask(SIG_BLOCK, &newmask, &oldmask);
 587         if (ret)
 588                 abort();
 589
 590         mutex_lock(&rcu_registry_lock);
 591         remove_thread(rcu_reader_reg);
 592         mutex_unlock(&rcu_registry_lock);
 593         ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL);
 594         if (ret)
 595                 abort();
 596         urcu_bp_exit();
 597 }
 598
 599 /*
 600  * Remove thread from the registry when it exits, and flag it as
 601  * destroyed so garbage collection can take care of it.
 602  */
 603 static
 604 void urcu_bp_thread_exit_notifier(void *rcu_key)
 605 {
 606         urcu_bp_unregister(rcu_key);
 607 }
 608
 609 #ifdef CONFIG_RCU_FORCE_SYS_MEMBARRIER
 610 static
 611 void urcu_bp_sys_membarrier_status(bool available)
 612 {
 613         if (!available)
 614                 abort();
 615 }
 616 #else
 617 static
 618 void urcu_bp_sys_membarrier_status(bool available)
 619 {
 620         if (!available)
 621                 return;
 622         urcu_bp_has_sys_membarrier = 1;
 623 }
 624 #endif
 625
 626 static
 627 void urcu_bp_sys_membarrier_init(void)
 628 {
 629         bool available = false;
 630         int mask;
 631
 632         mask = membarrier(MEMBARRIER_CMD_QUERY, 0);
 633         if (mask >= 0) {
 634                 if (mask & MEMBARRIER_CMD_PRIVATE_EXPEDITED) {
 635                         if (membarrier(MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED, 0))
 636                                 urcu_die(errno);
 637                         available = true;
 638                 }
 639         }
 640         urcu_bp_sys_membarrier_status(available);
 641 }
 642
 643 static
 644 void _urcu_bp_init(void)
 645 {
 646         mutex_lock(&init_lock);
 647         if (!urcu_bp_refcount++) {
 648                 int ret;
 649
 650                 ret = pthread_key_create(&urcu_bp_key,
 651                                 urcu_bp_thread_exit_notifier);
 652                 if (ret)
 653                         abort();
 654                 urcu_bp_sys_membarrier_init();
 655                 initialized = 1;
 656         }
 657         mutex_unlock(&init_lock);
 658 }
 659
 660 static
 661 void urcu_bp_exit(void)
 662 {
 663         mutex_lock(&init_lock);
 664         if (!--urcu_bp_refcount) {
 665                 struct registry_chunk *chunk, *tmp;
 666                 int ret;
 667
 668                 cds_list_for_each_entry_safe(chunk, tmp,
 669                                 &registry_arena.chunk_list, node) {
 670                         munmap((void *) chunk, chunk->data_len
 671                                         + sizeof(struct registry_chunk));
 672                 }
 673                 CDS_INIT_LIST_HEAD(&registry_arena.chunk_list);
 674                 ret = pthread_key_delete(urcu_bp_key);
 675                 if (ret)
 676                         abort();
 677         }
 678         mutex_unlock(&init_lock);
 679 }
 680
 681 /*
 682  * Holding the rcu_gp_lock and rcu_registry_lock across fork will make
 683  * sure we fork() don't race with a concurrent thread executing with
 684  * any of those locks held. This ensures that the registry and data
 685  * protected by rcu_gp_lock are in a coherent state in the child.
 686  */
 687 void urcu_bp_before_fork(void)
 688 {
 689         sigset_t newmask, oldmask;
 690         int ret;
 691
 692         ret = sigfillset(&newmask);
 693         assert(!ret);
 694         ret = pthread_sigmask(SIG_BLOCK, &newmask, &oldmask);
 695         assert(!ret);
 696         mutex_lock(&rcu_gp_lock);
 697         mutex_lock(&rcu_registry_lock);
 698         saved_fork_signal_mask = oldmask;
 699 }
 700 URCU_ATTR_ALIAS("urcu_bp_before_fork") void rcu_bp_before_fork();
 701
 702 void urcu_bp_after_fork_parent(void)
 703 {
 704         sigset_t oldmask;
 705         int ret;
 706
 707         oldmask = saved_fork_signal_mask;
 708         mutex_unlock(&rcu_registry_lock);
 709         mutex_unlock(&rcu_gp_lock);
 710         ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL);
 711         assert(!ret);
 712 }
 713 URCU_ATTR_ALIAS("urcu_bp_after_fork_parent")
 714 void rcu_bp_after_fork_parent(void);
 715
 716 /*
 717  * Prune all entries from registry except our own thread. Fits the Linux
 718  * fork behavior. Called with rcu_gp_lock and rcu_registry_lock held.
 719  */
 720 static
 721 void urcu_bp_prune_registry(void)
 722 {
 723         struct registry_chunk *chunk;
 724         struct urcu_bp_reader *rcu_reader_reg;
 725
 726         cds_list_for_each_entry(chunk, &registry_arena.chunk_list, node) {
 727                 for (rcu_reader_reg = (struct urcu_bp_reader *) &chunk->data[0];
 728                                 rcu_reader_reg < (struct urcu_bp_reader *) &chunk->data[chunk->data_len];
 729                                 rcu_reader_reg++) {
 730                         if (!rcu_reader_reg->alloc)
 731                                 continue;
 732                         if (rcu_reader_reg->tid == pthread_self())
 733                                 continue;
 734                         cleanup_thread(chunk, rcu_reader_reg);
 735                 }
 736         }
 737 }
 738
 739 void urcu_bp_after_fork_child(void)
 740 {
 741         sigset_t oldmask;
 742         int ret;
 743
 744         urcu_bp_prune_registry();
 745         oldmask = saved_fork_signal_mask;
 746         mutex_unlock(&rcu_registry_lock);
 747         mutex_unlock(&rcu_gp_lock);
 748         ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL);
 749         assert(!ret);
 750 }
 751 URCU_ATTR_ALIAS("urcu_bp_after_fork_child")
 752 void rcu_bp_after_fork_child(void);
 753
 754 void *urcu_bp_dereference_sym(void *p)
 755 {
 756         return _rcu_dereference(p);
 757 }
 758 URCU_ATTR_ALIAS("urcu_bp_dereference_sym")
 759 void *rcu_dereference_sym_bp();
 760
 761 void *urcu_bp_set_pointer_sym(void **p, void *v)
 762 {
 763         cmm_wmb();
 764         uatomic_set(p, v);
 765         return v;
 766 }
 767 URCU_ATTR_ALIAS("urcu_bp_set_pointer_sym")
 768 void *rcu_set_pointer_sym_bp();
 769
 770 void *urcu_bp_xchg_pointer_sym(void **p, void *v)
 771 {
 772         cmm_wmb();
 773         return uatomic_xchg(p, v);
 774 }
 775 URCU_ATTR_ALIAS("urcu_bp_xchg_pointer_sym")
 776 void *rcu_xchg_pointer_sym_bp();
 777
 778 void *urcu_bp_cmpxchg_pointer_sym(void **p, void *old, void *_new)
 779 {
 780         cmm_wmb();
 781         return uatomic_cmpxchg(p, old, _new);
 782 }
 783 URCU_ATTR_ALIAS("urcu_bp_cmpxchg_pointer_sym")
 784 void *rcu_cmpxchg_pointer_sym_bp();
 785
 786 DEFINE_RCU_FLAVOR(rcu_flavor);
 787 DEFINE_RCU_FLAVOR_ALIAS(rcu_flavor, alias_rcu_flavor);
 788
 789 #include "urcu-call-rcu-impl.h"
 790 #include "urcu-defer-impl.h"