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
 125 #ifndef CONFIG_RCU_FORCE_SYS_MEMBARRIER
 126 int urcu_bp_has_sys_membarrier;
 127 #endif
 128
 129 /*
 130  * rcu_gp_lock ensures mutual exclusion between threads calling
 131  * synchronize_rcu().
 132  */
 133 static pthread_mutex_t rcu_gp_lock = PTHREAD_MUTEX_INITIALIZER;
 134 /*
 135  * rcu_registry_lock ensures mutual exclusion between threads
 136  * registering and unregistering themselves to/from the registry, and
 137  * with threads reading that registry from synchronize_rcu(). However,
 138  * this lock is not held all the way through the completion of awaiting
 139  * for the grace period. It is sporadically released between iterations
 140  * on the registry.
 141  * rcu_registry_lock may nest inside rcu_gp_lock.
 142  */
 143 static pthread_mutex_t rcu_registry_lock = PTHREAD_MUTEX_INITIALIZER;
 144
 145 static pthread_mutex_t init_lock = PTHREAD_MUTEX_INITIALIZER;
 146 static int initialized;
 147
 148 static pthread_key_t urcu_bp_key;
 149
 150 struct urcu_bp_gp urcu_bp_gp = { .ctr = URCU_BP_GP_COUNT };
 151
 152 /*
 153  * Pointer to registry elements. Written to only by each individual reader. Read
 154  * by both the reader and the writers.
 155  */
 156 DEFINE_URCU_TLS(struct urcu_bp_reader *, urcu_bp_reader);
 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         urcu_posix_assert(!ret);
 281         ret = pthread_sigmask(SIG_BLOCK, &newmask, &oldmask);
 282         urcu_posix_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          * internally.
 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          * internally.
 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         urcu_posix_assert(!ret);
 349 }
 350
 351 /*
 352  * library wrappers to be used by non-LGPL compatible source code.
 353  */
 354
 355 void urcu_bp_read_lock(void)
 356 {
 357         _urcu_bp_read_lock();
 358 }
 359
 360 void urcu_bp_read_unlock(void)
 361 {
 362         _urcu_bp_read_unlock();
 363 }
 364
 365 int urcu_bp_read_ongoing(void)
 366 {
 367         return _urcu_bp_read_ongoing();
 368 }
 369
 370 /*
 371  * Only grow for now. If empty, allocate a ARENA_INIT_ALLOC sized chunk.
 372  * Else, try expanding the last chunk. If this fails, allocate a new
 373  * chunk twice as big as the last chunk.
 374  * Memory used by chunks _never_ moves. A chunk could theoretically be
 375  * freed when all "used" slots are released, but we don't do it at this
 376  * point.
 377  */
 378 static
 379 void expand_arena(struct registry_arena *arena)
 380 {
 381         struct registry_chunk *new_chunk, *last_chunk;
 382         size_t old_chunk_len, new_chunk_len;
 383
 384         /* No chunk. */
 385         if (cds_list_empty(&arena->chunk_list)) {
 386                 urcu_posix_assert(ARENA_INIT_ALLOC >=
 387                         sizeof(struct registry_chunk)
 388                         + sizeof(struct rcu_reader));
 389                 new_chunk_len = ARENA_INIT_ALLOC;
 390                 new_chunk = (struct registry_chunk *) mmap(NULL,
 391                         new_chunk_len,
 392                         PROT_READ | PROT_WRITE,
 393                         MAP_ANONYMOUS | MAP_PRIVATE,
 394                         -1, 0);
 395                 if (new_chunk == MAP_FAILED)
 396                         abort();
 397                 memset(new_chunk, 0, new_chunk_len);
 398                 new_chunk->data_len =
 399                         new_chunk_len - sizeof(struct registry_chunk);
 400                 cds_list_add_tail(&new_chunk->node, &arena->chunk_list);
 401                 return;         /* We're done. */
 402         }
 403
 404         /* Try expanding last chunk. */
 405         last_chunk = cds_list_entry(arena->chunk_list.prev,
 406                 struct registry_chunk, node);
 407         old_chunk_len =
 408                 last_chunk->data_len + sizeof(struct registry_chunk);
 409         new_chunk_len = old_chunk_len << 1;
 410
 411         /* Don't allow memory mapping to move, just expand. */
 412         new_chunk = mremap_wrapper(last_chunk, old_chunk_len,
 413                 new_chunk_len, 0);
 414         if (new_chunk != MAP_FAILED) {
 415                 /* Should not have moved. */
 416                 urcu_posix_assert(new_chunk == last_chunk);
 417                 memset((char *) last_chunk + old_chunk_len, 0,
 418                         new_chunk_len - old_chunk_len);
 419                 last_chunk->data_len =
 420                         new_chunk_len - sizeof(struct registry_chunk);
 421                 return;         /* We're done. */
 422         }
 423
 424         /* Remap did not succeed, we need to add a new chunk. */
 425         new_chunk = (struct registry_chunk *) mmap(NULL,
 426                 new_chunk_len,
 427                 PROT_READ | PROT_WRITE,
 428                 MAP_ANONYMOUS | MAP_PRIVATE,
 429                 -1, 0);
 430         if (new_chunk == MAP_FAILED)
 431                 abort();
 432         memset(new_chunk, 0, new_chunk_len);
 433         new_chunk->data_len =
 434                 new_chunk_len - sizeof(struct registry_chunk);
 435         cds_list_add_tail(&new_chunk->node, &arena->chunk_list);
 436 }
 437
 438 static
 439 struct rcu_reader *arena_alloc(struct registry_arena *arena)
 440 {
 441         struct registry_chunk *chunk;
 442         struct rcu_reader *rcu_reader_reg;
 443         int expand_done = 0;    /* Only allow to expand once per alloc */
 444         size_t len = sizeof(struct rcu_reader);
 445
 446 retry:
 447         cds_list_for_each_entry(chunk, &arena->chunk_list, node) {
 448                 if (chunk->data_len - chunk->used < len)
 449                         continue;
 450                 /* Find spot */
 451                 for (rcu_reader_reg = (struct rcu_reader *) &chunk->data[0];
 452                                 rcu_reader_reg < (struct rcu_reader *) &chunk->data[chunk->data_len];
 453                                 rcu_reader_reg++) {
 454                         if (!rcu_reader_reg->alloc) {
 455                                 rcu_reader_reg->alloc = 1;
 456                                 chunk->used += len;
 457                                 return rcu_reader_reg;
 458                         }
 459                 }
 460         }
 461
 462         if (!expand_done) {
 463                 expand_arena(arena);
 464                 expand_done = 1;
 465                 goto retry;
 466         }
 467
 468         return NULL;
 469 }
 470
 471 /* Called with signals off and mutex locked */
 472 static
 473 void add_thread(void)
 474 {
 475         struct rcu_reader *rcu_reader_reg;
 476         int ret;
 477
 478         rcu_reader_reg = arena_alloc(&registry_arena);
 479         if (!rcu_reader_reg)
 480                 abort();
 481         ret = pthread_setspecific(urcu_bp_key, rcu_reader_reg);
 482         if (ret)
 483                 abort();
 484
 485         /* Add to registry */
 486         rcu_reader_reg->tid = pthread_self();
 487         urcu_posix_assert(rcu_reader_reg->ctr == 0);
 488         cds_list_add(&rcu_reader_reg->node, &registry);
 489         /*
 490          * Reader threads are pointing to the reader registry. This is
 491          * why its memory should never be relocated.
 492          */
 493         URCU_TLS(urcu_bp_reader) = rcu_reader_reg;
 494 }
 495
 496 /* Called with mutex locked */
 497 static
 498 void cleanup_thread(struct registry_chunk *chunk,
 499                 struct rcu_reader *rcu_reader_reg)
 500 {
 501         rcu_reader_reg->ctr = 0;
 502         cds_list_del(&rcu_reader_reg->node);
 503         rcu_reader_reg->tid = 0;
 504         rcu_reader_reg->alloc = 0;
 505         chunk->used -= sizeof(struct rcu_reader);
 506 }
 507
 508 static
 509 struct registry_chunk *find_chunk(struct rcu_reader *rcu_reader_reg)
 510 {
 511         struct registry_chunk *chunk;
 512
 513         cds_list_for_each_entry(chunk, &registry_arena.chunk_list, node) {
 514                 if (rcu_reader_reg < (struct rcu_reader *) &chunk->data[0])
 515                         continue;
 516                 if (rcu_reader_reg >= (struct rcu_reader *) &chunk->data[chunk->data_len])
 517                         continue;
 518                 return chunk;
 519         }
 520         return NULL;
 521 }
 522
 523 /* Called with signals off and mutex locked */
 524 static
 525 void remove_thread(struct rcu_reader *rcu_reader_reg)
 526 {
 527         cleanup_thread(find_chunk(rcu_reader_reg), rcu_reader_reg);
 528         URCU_TLS(urcu_bp_reader) = NULL;
 529 }
 530
 531 /* Disable signals, take mutex, add to registry */
 532 void urcu_bp_register(void)
 533 {
 534         sigset_t newmask, oldmask;
 535         int ret;
 536
 537         ret = sigfillset(&newmask);
 538         if (ret)
 539                 abort();
 540         ret = pthread_sigmask(SIG_BLOCK, &newmask, &oldmask);
 541         if (ret)
 542                 abort();
 543
 544         /*
 545          * Check if a signal concurrently registered our thread since
 546          * the check in rcu_read_lock().
 547          */
 548         if (URCU_TLS(urcu_bp_reader))
 549                 goto end;
 550
 551         /*
 552          * Take care of early registration before urcu_bp constructor.
 553          */
 554         _urcu_bp_init();
 555
 556         mutex_lock(&rcu_registry_lock);
 557         add_thread();
 558         mutex_unlock(&rcu_registry_lock);
 559 end:
 560         ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL);
 561         if (ret)
 562                 abort();
 563 }
 564
 565 void urcu_bp_register_thread(void)
 566 {
 567         if (caa_unlikely(!URCU_TLS(urcu_bp_reader)))
 568                 urcu_bp_register(); /* If not yet registered. */
 569 }
 570
 571 /* Disable signals, take mutex, remove from registry */
 572 static
 573 void urcu_bp_unregister(struct rcu_reader *rcu_reader_reg)
 574 {
 575         sigset_t newmask, oldmask;
 576         int ret;
 577
 578         ret = sigfillset(&newmask);
 579         if (ret)
 580                 abort();
 581         ret = pthread_sigmask(SIG_BLOCK, &newmask, &oldmask);
 582         if (ret)
 583                 abort();
 584
 585         mutex_lock(&rcu_registry_lock);
 586         remove_thread(rcu_reader_reg);
 587         mutex_unlock(&rcu_registry_lock);
 588         ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL);
 589         if (ret)
 590                 abort();
 591         urcu_bp_exit();
 592 }
 593
 594 /*
 595  * Remove thread from the registry when it exits, and flag it as
 596  * destroyed so garbage collection can take care of it.
 597  */
 598 static
 599 void urcu_bp_thread_exit_notifier(void *rcu_key)
 600 {
 601         urcu_bp_unregister(rcu_key);
 602 }
 603
 604 #ifdef CONFIG_RCU_FORCE_SYS_MEMBARRIER
 605 static
 606 void urcu_bp_sys_membarrier_status(bool available)
 607 {
 608         if (!available)
 609                 abort();
 610 }
 611 #else
 612 static
 613 void urcu_bp_sys_membarrier_status(bool available)
 614 {
 615         if (!available)
 616                 return;
 617         urcu_bp_has_sys_membarrier = 1;
 618 }
 619 #endif
 620
 621 static
 622 void urcu_bp_sys_membarrier_init(void)
 623 {
 624         bool available = false;
 625         int mask;
 626
 627         mask = membarrier(MEMBARRIER_CMD_QUERY, 0);
 628         if (mask >= 0) {
 629                 if (mask & MEMBARRIER_CMD_PRIVATE_EXPEDITED) {
 630                         if (membarrier(MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED, 0))
 631                                 urcu_die(errno);
 632                         available = true;
 633                 }
 634         }
 635         urcu_bp_sys_membarrier_status(available);
 636 }
 637
 638 static
 639 void _urcu_bp_init(void)
 640 {
 641         mutex_lock(&init_lock);
 642         if (!urcu_bp_refcount++) {
 643                 int ret;
 644
 645                 ret = pthread_key_create(&urcu_bp_key,
 646                                 urcu_bp_thread_exit_notifier);
 647                 if (ret)
 648                         abort();
 649                 urcu_bp_sys_membarrier_init();
 650                 initialized = 1;
 651         }
 652         mutex_unlock(&init_lock);
 653 }
 654
 655 static
 656 void urcu_bp_exit(void)
 657 {
 658         mutex_lock(&init_lock);
 659         if (!--urcu_bp_refcount) {
 660                 struct registry_chunk *chunk, *tmp;
 661                 int ret;
 662
 663                 cds_list_for_each_entry_safe(chunk, tmp,
 664                                 &registry_arena.chunk_list, node) {
 665                         munmap((void *) chunk, chunk->data_len
 666                                         + sizeof(struct registry_chunk));
 667                 }
 668                 CDS_INIT_LIST_HEAD(&registry_arena.chunk_list);
 669                 ret = pthread_key_delete(urcu_bp_key);
 670                 if (ret)
 671                         abort();
 672         }
 673         mutex_unlock(&init_lock);
 674 }
 675
 676 /*
 677  * Holding the rcu_gp_lock and rcu_registry_lock across fork will make
 678  * sure we fork() don't race with a concurrent thread executing with
 679  * any of those locks held. This ensures that the registry and data
 680  * protected by rcu_gp_lock are in a coherent state in the child.
 681  */
 682 void urcu_bp_before_fork(void)
 683 {
 684         sigset_t newmask, oldmask;
 685         int ret;
 686
 687         ret = sigfillset(&newmask);
 688         urcu_posix_assert(!ret);
 689         ret = pthread_sigmask(SIG_BLOCK, &newmask, &oldmask);
 690         urcu_posix_assert(!ret);
 691         mutex_lock(&rcu_gp_lock);
 692         mutex_lock(&rcu_registry_lock);
 693         saved_fork_signal_mask = oldmask;
 694 }
 695
 696 void urcu_bp_after_fork_parent(void)
 697 {
 698         sigset_t oldmask;
 699         int ret;
 700
 701         oldmask = saved_fork_signal_mask;
 702         mutex_unlock(&rcu_registry_lock);
 703         mutex_unlock(&rcu_gp_lock);
 704         ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL);
 705         urcu_posix_assert(!ret);
 706 }
 707
 708 /*
 709  * Prune all entries from registry except our own thread. Fits the Linux
 710  * fork behavior. Called with rcu_gp_lock and rcu_registry_lock held.
 711  */
 712 static
 713 void urcu_bp_prune_registry(void)
 714 {
 715         struct registry_chunk *chunk;
 716         struct urcu_bp_reader *rcu_reader_reg;
 717
 718         cds_list_for_each_entry(chunk, &registry_arena.chunk_list, node) {
 719                 for (rcu_reader_reg = (struct urcu_bp_reader *) &chunk->data[0];
 720                                 rcu_reader_reg < (struct urcu_bp_reader *) &chunk->data[chunk->data_len];
 721                                 rcu_reader_reg++) {
 722                         if (!rcu_reader_reg->alloc)
 723                                 continue;
 724                         if (rcu_reader_reg->tid == pthread_self())
 725                                 continue;
 726                         cleanup_thread(chunk, rcu_reader_reg);
 727                 }
 728         }
 729 }
 730
 731 void urcu_bp_after_fork_child(void)
 732 {
 733         sigset_t oldmask;
 734         int ret;
 735
 736         urcu_bp_prune_registry();
 737         oldmask = saved_fork_signal_mask;
 738         mutex_unlock(&rcu_registry_lock);
 739         mutex_unlock(&rcu_gp_lock);
 740         ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL);
 741         urcu_posix_assert(!ret);
 742 }
 743
 744 void *urcu_bp_dereference_sym(void *p)
 745 {
 746         return _rcu_dereference(p);
 747 }
 748
 749 void *urcu_bp_set_pointer_sym(void **p, void *v)
 750 {
 751         cmm_wmb();
 752         uatomic_set(p, v);
 753         return v;
 754 }
 755
 756 void *urcu_bp_xchg_pointer_sym(void **p, void *v)
 757 {
 758         cmm_wmb();
 759         return uatomic_xchg(p, v);
 760 }
 761
 762 void *urcu_bp_cmpxchg_pointer_sym(void **p, void *old, void *_new)
 763 {
 764         cmm_wmb();
 765         return uatomic_cmpxchg(p, old, _new);
 766 }
 767
 768 DEFINE_RCU_FLAVOR(rcu_flavor);
 769
 770 #include "urcu-call-rcu-impl.h"
 771 #include "urcu-defer-impl.h"