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