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 urcu_bp_exit(void);
 124 static
 125 void __attribute__((destructor)) urcu_bp_exit_destructor(void);
 126 static void urcu_call_rcu_exit(void);
 127
 128 #ifndef CONFIG_RCU_FORCE_SYS_MEMBARRIER
 129 int urcu_bp_has_sys_membarrier;
 130 #endif
 131
 132 /*
 133  * rcu_gp_lock ensures mutual exclusion between threads calling
 134  * synchronize_rcu().
 135  */
 136 static pthread_mutex_t rcu_gp_lock = PTHREAD_MUTEX_INITIALIZER;
 137 /*
 138  * rcu_registry_lock ensures mutual exclusion between threads
 139  * registering and unregistering themselves to/from the registry, and
 140  * with threads reading that registry from synchronize_rcu(). However,
 141  * this lock is not held all the way through the completion of awaiting
 142  * for the grace period. It is sporadically released between iterations
 143  * on the registry.
 144  * rcu_registry_lock may nest inside rcu_gp_lock.
 145  */
 146 static pthread_mutex_t rcu_registry_lock = PTHREAD_MUTEX_INITIALIZER;
 147
 148 static pthread_mutex_t init_lock = PTHREAD_MUTEX_INITIALIZER;
 149 static int initialized;
 150
 151 static pthread_key_t urcu_bp_key;
 152
 153 struct urcu_bp_gp urcu_bp_gp = { .ctr = URCU_BP_GP_COUNT };
 154
 155 /*
 156  * Pointer to registry elements. Written to only by each individual reader. Read
 157  * by both the reader and the writers.
 158  */
 159 DEFINE_URCU_TLS(struct urcu_bp_reader *, urcu_bp_reader);
 160
 161 static CDS_LIST_HEAD(registry);
 162
 163 struct registry_chunk {
 164         size_t data_len;                /* data length */
 165         size_t used;                    /* amount of data used */
 166         struct cds_list_head node;      /* chunk_list node */
 167         char data[];
 168 };
 169
 170 struct registry_arena {
 171         struct cds_list_head chunk_list;
 172 };
 173
 174 static struct registry_arena registry_arena = {
 175         .chunk_list = CDS_LIST_HEAD_INIT(registry_arena.chunk_list),
 176 };
 177
 178 /* Saved fork signal mask, protected by rcu_gp_lock */
 179 static sigset_t saved_fork_signal_mask;
 180
 181 static void mutex_lock(pthread_mutex_t *mutex)
 182 {
 183         int ret;
 184
 185 #ifndef DISTRUST_SIGNALS_EXTREME
 186         ret = pthread_mutex_lock(mutex);
 187         if (ret)
 188                 urcu_die(ret);
 189 #else /* #ifndef DISTRUST_SIGNALS_EXTREME */
 190         while ((ret = pthread_mutex_trylock(mutex)) != 0) {
 191                 if (ret != EBUSY && ret != EINTR)
 192                         urcu_die(ret);
 193                 poll(NULL,0,10);
 194         }
 195 #endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */
 196 }
 197
 198 static void mutex_unlock(pthread_mutex_t *mutex)
 199 {
 200         int ret;
 201
 202         ret = pthread_mutex_unlock(mutex);
 203         if (ret)
 204                 urcu_die(ret);
 205 }
 206
 207 static void smp_mb_master(void)
 208 {
 209         if (caa_likely(urcu_bp_has_sys_membarrier)) {
 210                 if (membarrier(MEMBARRIER_CMD_PRIVATE_EXPEDITED, 0))
 211                         urcu_die(errno);
 212         } else {
 213                 cmm_smp_mb();
 214         }
 215 }
 216
 217 /*
 218  * Always called with rcu_registry lock held. Releases this lock between
 219  * iterations and grabs it again. Holds the lock when it returns.
 220  */
 221 static void wait_for_readers(struct cds_list_head *input_readers,
 222                         struct cds_list_head *cur_snap_readers,
 223                         struct cds_list_head *qsreaders)
 224 {
 225         unsigned int wait_loops = 0;
 226         struct urcu_bp_reader *index, *tmp;
 227
 228         /*
 229          * Wait for each thread URCU_TLS(urcu_bp_reader).ctr to either
 230          * indicate quiescence (not nested), or observe the current
 231          * rcu_gp.ctr value.
 232          */
 233         for (;;) {
 234                 if (wait_loops < RCU_QS_ACTIVE_ATTEMPTS)
 235                         wait_loops++;
 236
 237                 cds_list_for_each_entry_safe(index, tmp, input_readers, node) {
 238                         switch (urcu_bp_reader_state(&index->ctr)) {
 239                         case URCU_BP_READER_ACTIVE_CURRENT:
 240                                 if (cur_snap_readers) {
 241                                         cds_list_move(&index->node,
 242                                                 cur_snap_readers);
 243                                         break;
 244                                 }
 245                                 /* Fall-through */
 246                         case URCU_BP_READER_INACTIVE:
 247                                 cds_list_move(&index->node, qsreaders);
 248                                 break;
 249                         case URCU_BP_READER_ACTIVE_OLD:
 250                                 /*
 251                                  * Old snapshot. Leaving node in
 252                                  * input_readers will make us busy-loop
 253                                  * until the snapshot becomes current or
 254                                  * the reader becomes inactive.
 255                                  */
 256                                 break;
 257                         }
 258                 }
 259
 260                 if (cds_list_empty(input_readers)) {
 261                         break;
 262                 } else {
 263                         /* Temporarily unlock the registry lock. */
 264                         mutex_unlock(&rcu_registry_lock);
 265                         if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS)
 266                                 (void) poll(NULL, 0, RCU_SLEEP_DELAY_MS);
 267                         else
 268                                 caa_cpu_relax();
 269                         /* Re-lock the registry lock before the next loop. */
 270                         mutex_lock(&rcu_registry_lock);
 271                 }
 272         }
 273 }
 274
 275 void urcu_bp_synchronize_rcu(void)
 276 {
 277         CDS_LIST_HEAD(cur_snap_readers);
 278         CDS_LIST_HEAD(qsreaders);
 279         sigset_t newmask, oldmask;
 280         int ret;
 281
 282         ret = sigfillset(&newmask);
 283         urcu_posix_assert(!ret);
 284         ret = pthread_sigmask(SIG_BLOCK, &newmask, &oldmask);
 285         urcu_posix_assert(!ret);
 286
 287         mutex_lock(&rcu_gp_lock);
 288
 289         mutex_lock(&rcu_registry_lock);
 290
 291         if (cds_list_empty(&registry))
 292                 goto out;
 293
 294         /* All threads should read qparity before accessing data structure
 295          * where new ptr points to. */
 296         /* Write new ptr before changing the qparity */
 297         smp_mb_master();
 298
 299         /*
 300          * Wait for readers to observe original parity or be quiescent.
 301          * wait_for_readers() can release and grab again rcu_registry_lock
 302          * internally.
 303          */
 304         wait_for_readers(&registry, &cur_snap_readers, &qsreaders);
 305
 306         /*
 307          * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
 308          * model easier to understand. It does not have a big performance impact
 309          * anyway, given this is the write-side.
 310          */
 311         cmm_smp_mb();
 312
 313         /* Switch parity: 0 -> 1, 1 -> 0 */
 314         CMM_STORE_SHARED(rcu_gp.ctr, rcu_gp.ctr ^ URCU_BP_GP_CTR_PHASE);
 315
 316         /*
 317          * Must commit qparity update to memory before waiting for other parity
 318          * quiescent state. Failure to do so could result in the writer waiting
 319          * forever while new readers are always accessing data (no progress).
 320          * Ensured by CMM_STORE_SHARED and CMM_LOAD_SHARED.
 321          */
 322
 323         /*
 324          * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
 325          * model easier to understand. It does not have a big performance impact
 326          * anyway, given this is the write-side.
 327          */
 328         cmm_smp_mb();
 329
 330         /*
 331          * Wait for readers to observe new parity or be quiescent.
 332          * wait_for_readers() can release and grab again rcu_registry_lock
 333          * internally.
 334          */
 335         wait_for_readers(&cur_snap_readers, NULL, &qsreaders);
 336
 337         /*
 338          * Put quiescent reader list back into registry.
 339          */
 340         cds_list_splice(&qsreaders, &registry);
 341
 342         /*
 343          * Finish waiting for reader threads before letting the old ptr being
 344          * freed.
 345          */
 346         smp_mb_master();
 347 out:
 348         mutex_unlock(&rcu_registry_lock);
 349         mutex_unlock(&rcu_gp_lock);
 350         ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL);
 351         urcu_posix_assert(!ret);
 352 }
 353
 354 /*
 355  * library wrappers to be used by non-LGPL compatible source code.
 356  */
 357
 358 void urcu_bp_read_lock(void)
 359 {
 360         _urcu_bp_read_lock();
 361 }
 362
 363 void urcu_bp_read_unlock(void)
 364 {
 365         _urcu_bp_read_unlock();
 366 }
 367
 368 int urcu_bp_read_ongoing(void)
 369 {
 370         return _urcu_bp_read_ongoing();
 371 }
 372
 373 /*
 374  * Only grow for now. If empty, allocate a ARENA_INIT_ALLOC sized chunk.
 375  * Else, try expanding the last chunk. If this fails, allocate a new
 376  * chunk twice as big as the last chunk.
 377  * Memory used by chunks _never_ moves. A chunk could theoretically be
 378  * freed when all "used" slots are released, but we don't do it at this
 379  * point.
 380  */
 381 static
 382 void expand_arena(struct registry_arena *arena)
 383 {
 384         struct registry_chunk *new_chunk, *last_chunk;
 385         size_t old_chunk_len, new_chunk_len;
 386
 387         /* No chunk. */
 388         if (cds_list_empty(&arena->chunk_list)) {
 389                 urcu_posix_assert(ARENA_INIT_ALLOC >=
 390                         sizeof(struct registry_chunk)
 391                         + sizeof(struct rcu_reader));
 392                 new_chunk_len = ARENA_INIT_ALLOC;
 393                 new_chunk = (struct registry_chunk *) mmap(NULL,
 394                         new_chunk_len,
 395                         PROT_READ | PROT_WRITE,
 396                         MAP_ANONYMOUS | MAP_PRIVATE,
 397                         -1, 0);
 398                 if (new_chunk == MAP_FAILED)
 399                         abort();
 400                 memset(new_chunk, 0, new_chunk_len);
 401                 new_chunk->data_len =
 402                         new_chunk_len - sizeof(struct registry_chunk);
 403                 cds_list_add_tail(&new_chunk->node, &arena->chunk_list);
 404                 return;         /* We're done. */
 405         }
 406
 407         /* Try expanding last chunk. */
 408         last_chunk = cds_list_entry(arena->chunk_list.prev,
 409                 struct registry_chunk, node);
 410         old_chunk_len =
 411                 last_chunk->data_len + sizeof(struct registry_chunk);
 412         new_chunk_len = old_chunk_len << 1;
 413
 414         /* Don't allow memory mapping to move, just expand. */
 415         new_chunk = mremap_wrapper(last_chunk, old_chunk_len,
 416                 new_chunk_len, 0);
 417         if (new_chunk != MAP_FAILED) {
 418                 /* Should not have moved. */
 419                 urcu_posix_assert(new_chunk == last_chunk);
 420                 memset((char *) last_chunk + old_chunk_len, 0,
 421                         new_chunk_len - old_chunk_len);
 422                 last_chunk->data_len =
 423                         new_chunk_len - sizeof(struct registry_chunk);
 424                 return;         /* We're done. */
 425         }
 426
 427         /* Remap did not succeed, we need to add a new chunk. */
 428         new_chunk = (struct registry_chunk *) mmap(NULL,
 429                 new_chunk_len,
 430                 PROT_READ | PROT_WRITE,
 431                 MAP_ANONYMOUS | MAP_PRIVATE,
 432                 -1, 0);
 433         if (new_chunk == MAP_FAILED)
 434                 abort();
 435         memset(new_chunk, 0, new_chunk_len);
 436         new_chunk->data_len =
 437                 new_chunk_len - sizeof(struct registry_chunk);
 438         cds_list_add_tail(&new_chunk->node, &arena->chunk_list);
 439 }
 440
 441 static
 442 struct rcu_reader *arena_alloc(struct registry_arena *arena)
 443 {
 444         struct registry_chunk *chunk;
 445         struct rcu_reader *rcu_reader_reg;
 446         int expand_done = 0;    /* Only allow to expand once per alloc */
 447         size_t len = sizeof(struct rcu_reader);
 448
 449 retry:
 450         cds_list_for_each_entry(chunk, &arena->chunk_list, node) {
 451                 if (chunk->data_len - chunk->used < len)
 452                         continue;
 453                 /* Find spot */
 454                 for (rcu_reader_reg = (struct rcu_reader *) &chunk->data[0];
 455                                 rcu_reader_reg < (struct rcu_reader *) &chunk->data[chunk->data_len];
 456                                 rcu_reader_reg++) {
 457                         if (!rcu_reader_reg->alloc) {
 458                                 rcu_reader_reg->alloc = 1;
 459                                 chunk->used += len;
 460                                 return rcu_reader_reg;
 461                         }
 462                 }
 463         }
 464
 465         if (!expand_done) {
 466                 expand_arena(arena);
 467                 expand_done = 1;
 468                 goto retry;
 469         }
 470
 471         return NULL;
 472 }
 473
 474 /* Called with signals off and mutex locked */
 475 static
 476 void add_thread(void)
 477 {
 478         struct rcu_reader *rcu_reader_reg;
 479         int ret;
 480
 481         rcu_reader_reg = arena_alloc(&registry_arena);
 482         if (!rcu_reader_reg)
 483                 abort();
 484         ret = pthread_setspecific(urcu_bp_key, rcu_reader_reg);
 485         if (ret)
 486                 abort();
 487
 488         /* Add to registry */
 489         rcu_reader_reg->tid = pthread_self();
 490         urcu_posix_assert(rcu_reader_reg->ctr == 0);
 491         cds_list_add(&rcu_reader_reg->node, &registry);
 492         /*
 493          * Reader threads are pointing to the reader registry. This is
 494          * why its memory should never be relocated.
 495          */
 496         URCU_TLS(urcu_bp_reader) = rcu_reader_reg;
 497 }
 498
 499 /* Called with mutex locked */
 500 static
 501 void cleanup_thread(struct registry_chunk *chunk,
 502                 struct rcu_reader *rcu_reader_reg)
 503 {
 504         rcu_reader_reg->ctr = 0;
 505         cds_list_del(&rcu_reader_reg->node);
 506         rcu_reader_reg->tid = 0;
 507         rcu_reader_reg->alloc = 0;
 508         chunk->used -= sizeof(struct rcu_reader);
 509 }
 510
 511 static
 512 struct registry_chunk *find_chunk(struct rcu_reader *rcu_reader_reg)
 513 {
 514         struct registry_chunk *chunk;
 515
 516         cds_list_for_each_entry(chunk, &registry_arena.chunk_list, node) {
 517                 if (rcu_reader_reg < (struct rcu_reader *) &chunk->data[0])
 518                         continue;
 519                 if (rcu_reader_reg >= (struct rcu_reader *) &chunk->data[chunk->data_len])
 520                         continue;
 521                 return chunk;
 522         }
 523         return NULL;
 524 }
 525
 526 /* Called with signals off and mutex locked */
 527 static
 528 void remove_thread(struct rcu_reader *rcu_reader_reg)
 529 {
 530         cleanup_thread(find_chunk(rcu_reader_reg), rcu_reader_reg);
 531         URCU_TLS(urcu_bp_reader) = NULL;
 532 }
 533
 534 /* Disable signals, take mutex, add to registry */
 535 void urcu_bp_register(void)
 536 {
 537         sigset_t newmask, oldmask;
 538         int ret;
 539
 540         ret = sigfillset(&newmask);
 541         if (ret)
 542                 abort();
 543         ret = pthread_sigmask(SIG_BLOCK, &newmask, &oldmask);
 544         if (ret)
 545                 abort();
 546
 547         /*
 548          * Check if a signal concurrently registered our thread since
 549          * the check in rcu_read_lock().
 550          */
 551         if (URCU_TLS(urcu_bp_reader))
 552                 goto end;
 553
 554         /*
 555          * Take care of early registration before urcu_bp constructor.
 556          */
 557         _urcu_bp_init();
 558
 559         mutex_lock(&rcu_registry_lock);
 560         add_thread();
 561         mutex_unlock(&rcu_registry_lock);
 562 end:
 563         ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL);
 564         if (ret)
 565                 abort();
 566 }
 567
 568 void urcu_bp_register_thread(void)
 569 {
 570         if (caa_unlikely(!URCU_TLS(urcu_bp_reader)))
 571                 urcu_bp_register(); /* If not yet registered. */
 572 }
 573
 574 /* Disable signals, take mutex, remove from registry */
 575 static
 576 void urcu_bp_unregister(struct rcu_reader *rcu_reader_reg)
 577 {
 578         sigset_t newmask, oldmask;
 579         int ret;
 580
 581         ret = sigfillset(&newmask);
 582         if (ret)
 583                 abort();
 584         ret = pthread_sigmask(SIG_BLOCK, &newmask, &oldmask);
 585         if (ret)
 586                 abort();
 587
 588         mutex_lock(&rcu_registry_lock);
 589         remove_thread(rcu_reader_reg);
 590         mutex_unlock(&rcu_registry_lock);
 591         ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL);
 592         if (ret)
 593                 abort();
 594         urcu_bp_exit();
 595 }
 596
 597 /*
 598  * Remove thread from the registry when it exits, and flag it as
 599  * destroyed so garbage collection can take care of it.
 600  */
 601 static
 602 void urcu_bp_thread_exit_notifier(void *rcu_key)
 603 {
 604         urcu_bp_unregister(rcu_key);
 605 }
 606
 607 #ifdef CONFIG_RCU_FORCE_SYS_MEMBARRIER
 608 static
 609 void urcu_bp_sys_membarrier_status(bool available)
 610 {
 611         if (!available)
 612                 abort();
 613 }
 614 #else
 615 static
 616 void urcu_bp_sys_membarrier_status(bool available)
 617 {
 618         if (!available)
 619                 return;
 620         urcu_bp_has_sys_membarrier = 1;
 621 }
 622 #endif
 623
 624 static
 625 void urcu_bp_sys_membarrier_init(void)
 626 {
 627         bool available = false;
 628         int mask;
 629
 630         mask = membarrier(MEMBARRIER_CMD_QUERY, 0);
 631         if (mask >= 0) {
 632                 if (mask & MEMBARRIER_CMD_PRIVATE_EXPEDITED) {
 633                         if (membarrier(MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED, 0))
 634                                 urcu_die(errno);
 635                         available = true;
 636                 }
 637         }
 638         urcu_bp_sys_membarrier_status(available);
 639 }
 640
 641 static
 642 void _urcu_bp_init(void)
 643 {
 644         mutex_lock(&init_lock);
 645         if (!urcu_bp_refcount++) {
 646                 int ret;
 647
 648                 ret = pthread_key_create(&urcu_bp_key,
 649                                 urcu_bp_thread_exit_notifier);
 650                 if (ret)
 651                         abort();
 652                 urcu_bp_sys_membarrier_init();
 653                 initialized = 1;
 654         }
 655         mutex_unlock(&init_lock);
 656 }
 657
 658 static
 659 void urcu_bp_exit(void)
 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 static
 680 void urcu_bp_exit_destructor(void)
 681 {
 682         urcu_call_rcu_exit();
 683         urcu_bp_exit();
 684 }
 685
 686 /*
 687  * Holding the rcu_gp_lock and rcu_registry_lock across fork will make
 688  * sure we fork() don't race with a concurrent thread executing with
 689  * any of those locks held. This ensures that the registry and data
 690  * protected by rcu_gp_lock are in a coherent state in the child.
 691  */
 692 void urcu_bp_before_fork(void)
 693 {
 694         sigset_t newmask, oldmask;
 695         int ret;
 696
 697         ret = sigfillset(&newmask);
 698         urcu_posix_assert(!ret);
 699         ret = pthread_sigmask(SIG_BLOCK, &newmask, &oldmask);
 700         urcu_posix_assert(!ret);
 701         mutex_lock(&rcu_gp_lock);
 702         mutex_lock(&rcu_registry_lock);
 703         saved_fork_signal_mask = oldmask;
 704 }
 705
 706 void urcu_bp_after_fork_parent(void)
 707 {
 708         sigset_t oldmask;
 709         int ret;
 710
 711         oldmask = saved_fork_signal_mask;
 712         mutex_unlock(&rcu_registry_lock);
 713         mutex_unlock(&rcu_gp_lock);
 714         ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL);
 715         urcu_posix_assert(!ret);
 716 }
 717
 718 /*
 719  * Prune all entries from registry except our own thread. Fits the Linux
 720  * fork behavior. Called with rcu_gp_lock and rcu_registry_lock held.
 721  */
 722 static
 723 void urcu_bp_prune_registry(void)
 724 {
 725         struct registry_chunk *chunk;
 726         struct urcu_bp_reader *rcu_reader_reg;
 727
 728         cds_list_for_each_entry(chunk, &registry_arena.chunk_list, node) {
 729                 for (rcu_reader_reg = (struct urcu_bp_reader *) &chunk->data[0];
 730                                 rcu_reader_reg < (struct urcu_bp_reader *) &chunk->data[chunk->data_len];
 731                                 rcu_reader_reg++) {
 732                         if (!rcu_reader_reg->alloc)
 733                                 continue;
 734                         if (rcu_reader_reg->tid == pthread_self())
 735                                 continue;
 736                         cleanup_thread(chunk, rcu_reader_reg);
 737                 }
 738         }
 739 }
 740
 741 void urcu_bp_after_fork_child(void)
 742 {
 743         sigset_t oldmask;
 744         int ret;
 745
 746         urcu_bp_prune_registry();
 747         oldmask = saved_fork_signal_mask;
 748         mutex_unlock(&rcu_registry_lock);
 749         mutex_unlock(&rcu_gp_lock);
 750         ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL);
 751         urcu_posix_assert(!ret);
 752 }
 753
 754 void *urcu_bp_dereference_sym(void *p)
 755 {
 756         return _rcu_dereference(p);
 757 }
 758
 759 void *urcu_bp_set_pointer_sym(void **p, void *v)
 760 {
 761         cmm_wmb();
 762         uatomic_set(p, v);
 763         return v;
 764 }
 765
 766 void *urcu_bp_xchg_pointer_sym(void **p, void *v)
 767 {
 768         cmm_wmb();
 769         return uatomic_xchg(p, v);
 770 }
 771
 772 void *urcu_bp_cmpxchg_pointer_sym(void **p, void *old, void *_new)
 773 {
 774         cmm_wmb();
 775         return uatomic_cmpxchg(p, old, _new);
 776 }
 777
 778 DEFINE_RCU_FLAVOR(rcu_flavor);
 779
 780 #include "urcu-call-rcu-impl.h"
 781 #include "urcu-defer-impl.h"
 782 #include "urcu-poll-impl.h"