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 __attribute__((unused)),
  79                 size_t old_size __attribute__((unused)),
  80                 size_t new_size __attribute__((unused)),
  81                 int flags)
  82 {
  83         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_READER_COUNT       8
  92
  93 /*
  94  * Active attempts to check for reader Q.S. before calling sleep().
  95  */
  96 #define RCU_QS_ACTIVE_ATTEMPTS 100
  97
  98 static
  99 int urcu_bp_refcount;
 100
 101 /* If the headers do not support membarrier system call, fall back smp_mb. */
 102 #ifdef __NR_membarrier
 103 # define membarrier(...)                syscall(__NR_membarrier, __VA_ARGS__)
 104 #else
 105 # define membarrier(...)                -ENOSYS
 106 #endif
 107
 108 enum membarrier_cmd {
 109         MEMBARRIER_CMD_QUERY                            = 0,
 110         MEMBARRIER_CMD_SHARED                           = (1 << 0),
 111         /* reserved for MEMBARRIER_CMD_SHARED_EXPEDITED (1 << 1) */
 112         /* reserved for MEMBARRIER_CMD_PRIVATE (1 << 2) */
 113         MEMBARRIER_CMD_PRIVATE_EXPEDITED                = (1 << 3),
 114         MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED       = (1 << 4),
 115 };
 116
 117 static
 118 void __attribute__((constructor)) _urcu_bp_init(void);
 119 static
 120 void urcu_bp_exit(void);
 121 static
 122 void __attribute__((destructor)) urcu_bp_exit_destructor(void);
 123 static void urcu_call_rcu_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 capacity;                /* capacity of this chunk (in elements) */
 162         size_t used;                    /* count of elements used */
 163         struct cds_list_head node;      /* chunk_list node */
 164         struct urcu_bp_reader readers[];
 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 /* Get the size of a chunk's allocation from its capacity (an element count). */
 215 static size_t chunk_allocation_size(size_t capacity)
 216 {
 217         return (capacity * sizeof(struct urcu_bp_reader)) +
 218                 sizeof(struct registry_chunk);
 219 }
 220
 221 /*
 222  * Always called with rcu_registry lock held. Releases this lock between
 223  * iterations and grabs it again. Holds the lock when it returns.
 224  */
 225 static void wait_for_readers(struct cds_list_head *input_readers,
 226                         struct cds_list_head *cur_snap_readers,
 227                         struct cds_list_head *qsreaders)
 228 {
 229         unsigned int wait_loops = 0;
 230         struct urcu_bp_reader *index, *tmp;
 231
 232         /*
 233          * Wait for each thread URCU_TLS(urcu_bp_reader).ctr to either
 234          * indicate quiescence (not nested), or observe the current
 235          * rcu_gp.ctr value.
 236          */
 237         for (;;) {
 238                 if (wait_loops < RCU_QS_ACTIVE_ATTEMPTS)
 239                         wait_loops++;
 240
 241                 cds_list_for_each_entry_safe(index, tmp, input_readers, node) {
 242                         switch (urcu_bp_reader_state(&index->ctr)) {
 243                         case URCU_BP_READER_ACTIVE_CURRENT:
 244                                 if (cur_snap_readers) {
 245                                         cds_list_move(&index->node,
 246                                                 cur_snap_readers);
 247                                         break;
 248                                 }
 249                                 /* Fall-through */
 250                         case URCU_BP_READER_INACTIVE:
 251                                 cds_list_move(&index->node, qsreaders);
 252                                 break;
 253                         case URCU_BP_READER_ACTIVE_OLD:
 254                                 /*
 255                                  * Old snapshot. Leaving node in
 256                                  * input_readers will make us busy-loop
 257                                  * until the snapshot becomes current or
 258                                  * the reader becomes inactive.
 259                                  */
 260                                 break;
 261                         }
 262                 }
 263
 264                 if (cds_list_empty(input_readers)) {
 265                         break;
 266                 } else {
 267                         /* Temporarily unlock the registry lock. */
 268                         mutex_unlock(&rcu_registry_lock);
 269                         if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS)
 270                                 (void) poll(NULL, 0, RCU_SLEEP_DELAY_MS);
 271                         else
 272                                 caa_cpu_relax();
 273                         /* Re-lock the registry lock before the next loop. */
 274                         mutex_lock(&rcu_registry_lock);
 275                 }
 276         }
 277 }
 278
 279 void urcu_bp_synchronize_rcu(void)
 280 {
 281         CDS_LIST_HEAD(cur_snap_readers);
 282         CDS_LIST_HEAD(qsreaders);
 283         sigset_t newmask, oldmask;
 284         int ret;
 285
 286         ret = sigfillset(&newmask);
 287         assert(!ret);
 288         ret = pthread_sigmask(SIG_BLOCK, &newmask, &oldmask);
 289         assert(!ret);
 290
 291         mutex_lock(&rcu_gp_lock);
 292
 293         mutex_lock(&rcu_registry_lock);
 294
 295         if (cds_list_empty(&registry))
 296                 goto out;
 297
 298         /* All threads should read qparity before accessing data structure
 299          * where new ptr points to. */
 300         /* Write new ptr before changing the qparity */
 301         smp_mb_master();
 302
 303         /*
 304          * Wait for readers to observe original parity or be quiescent.
 305          * wait_for_readers() can release and grab again rcu_registry_lock
 306          * interally.
 307          */
 308         wait_for_readers(&registry, &cur_snap_readers, &qsreaders);
 309
 310         /*
 311          * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
 312          * model easier to understand. It does not have a big performance impact
 313          * anyway, given this is the write-side.
 314          */
 315         cmm_smp_mb();
 316
 317         /* Switch parity: 0 -> 1, 1 -> 0 */
 318         CMM_STORE_SHARED(rcu_gp.ctr, rcu_gp.ctr ^ URCU_BP_GP_CTR_PHASE);
 319
 320         /*
 321          * Must commit qparity update to memory before waiting for other parity
 322          * quiescent state. Failure to do so could result in the writer waiting
 323          * forever while new readers are always accessing data (no progress).
 324          * Ensured by CMM_STORE_SHARED and CMM_LOAD_SHARED.
 325          */
 326
 327         /*
 328          * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
 329          * model easier to understand. It does not have a big performance impact
 330          * anyway, given this is the write-side.
 331          */
 332         cmm_smp_mb();
 333
 334         /*
 335          * Wait for readers to observe new parity or be quiescent.
 336          * wait_for_readers() can release and grab again rcu_registry_lock
 337          * interally.
 338          */
 339         wait_for_readers(&cur_snap_readers, NULL, &qsreaders);
 340
 341         /*
 342          * Put quiescent reader list back into registry.
 343          */
 344         cds_list_splice(&qsreaders, &registry);
 345
 346         /*
 347          * Finish waiting for reader threads before letting the old ptr being
 348          * freed.
 349          */
 350         smp_mb_master();
 351 out:
 352         mutex_unlock(&rcu_registry_lock);
 353         mutex_unlock(&rcu_gp_lock);
 354         ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL);
 355         assert(!ret);
 356 }
 357
 358 /*
 359  * library wrappers to be used by non-LGPL compatible source code.
 360  */
 361
 362 void urcu_bp_read_lock(void)
 363 {
 364         _urcu_bp_read_lock();
 365 }
 366
 367 void urcu_bp_read_unlock(void)
 368 {
 369         _urcu_bp_read_unlock();
 370 }
 371
 372 int urcu_bp_read_ongoing(void)
 373 {
 374         return _urcu_bp_read_ongoing();
 375 }
 376
 377 /*
 378  * Only grow for now. If empty, allocate a ARENA_INIT_ALLOC sized chunk.
 379  * Else, try expanding the last chunk. If this fails, allocate a new
 380  * chunk twice as big as the last chunk.
 381  * Memory used by chunks _never_ moves. A chunk could theoretically be
 382  * freed when all "used" slots are released, but we don't do it at this
 383  * point.
 384  */
 385 static
 386 void expand_arena(struct registry_arena *arena)
 387 {
 388         struct registry_chunk *new_chunk, *last_chunk;
 389         size_t old_chunk_size_bytes, new_chunk_size_bytes, new_capacity;
 390
 391         /* No chunk. */
 392         if (cds_list_empty(&arena->chunk_list)) {
 393                 new_chunk_size_bytes = chunk_allocation_size(INIT_READER_COUNT);
 394                 new_chunk = (struct registry_chunk *) mmap(NULL,
 395                         new_chunk_size_bytes,
 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_size_bytes);
 402                 new_chunk->capacity = INIT_READER_COUNT;
 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_size_bytes = chunk_allocation_size(last_chunk->capacity);
 411         new_capacity = last_chunk->capacity << 1;
 412         new_chunk_size_bytes = chunk_allocation_size(new_capacity);
 413
 414         /* Don't allow memory mapping to move, just expand. */
 415         new_chunk = mremap_wrapper(last_chunk, old_chunk_size_bytes,
 416                 new_chunk_size_bytes, 0);
 417         if (new_chunk != MAP_FAILED) {
 418                 /* Should not have moved. */
 419                 assert(new_chunk == last_chunk);
 420                 memset((char *) last_chunk + old_chunk_size_bytes, 0,
 421                         new_chunk_size_bytes - old_chunk_size_bytes);
 422                 last_chunk->capacity = new_capacity;
 423                 return;         /* We're done. */
 424         }
 425
 426         /* Remap did not succeed, we need to add a new chunk. */
 427         new_chunk = (struct registry_chunk *) mmap(NULL,
 428                 new_chunk_size_bytes,
 429                 PROT_READ | PROT_WRITE,
 430                 MAP_ANONYMOUS | MAP_PRIVATE,
 431                 -1, 0);
 432         if (new_chunk == MAP_FAILED)
 433                 abort();
 434         memset(new_chunk, 0, new_chunk_size_bytes);
 435         new_chunk->capacity = new_capacity;
 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         int expand_done = 0;    /* Only allow to expand once per alloc */
 444
 445 retry:
 446         cds_list_for_each_entry(chunk, &arena->chunk_list, node) {
 447                 size_t spot_idx;
 448
 449                 /* Skip fully used chunks. */
 450                 if (chunk->used == chunk->capacity) {
 451                         continue;
 452                 }
 453
 454                 /* Find a spot. */
 455                 for (spot_idx = 0; spot_idx < chunk->capacity; spot_idx++) {
 456                         if (!chunk->readers[spot_idx].alloc) {
 457                                 chunk->readers[spot_idx].alloc = 1;
 458                                 chunk->used++;
 459                                 return &chunk->readers[spot_idx];
 460                         }
 461                 }
 462         }
 463
 464         if (!expand_done) {
 465                 expand_arena(arena);
 466                 expand_done = 1;
 467                 goto retry;
 468         }
 469
 470         return NULL;
 471 }
 472
 473 /* Called with signals off and mutex locked */
 474 static
 475 void add_thread(void)
 476 {
 477         struct rcu_reader *rcu_reader_reg;
 478         int ret;
 479
 480         rcu_reader_reg = arena_alloc(&registry_arena);
 481         if (!rcu_reader_reg)
 482                 abort();
 483         ret = pthread_setspecific(urcu_bp_key, rcu_reader_reg);
 484         if (ret)
 485                 abort();
 486
 487         /* Add to registry */
 488         rcu_reader_reg->tid = pthread_self();
 489         assert(rcu_reader_reg->ctr == 0);
 490         cds_list_add(&rcu_reader_reg->node, &registry);
 491         /*
 492          * Reader threads are pointing to the reader registry. This is
 493          * why its memory should never be relocated.
 494          */
 495         URCU_TLS(urcu_bp_reader) = rcu_reader_reg;
 496 }
 497
 498 /* Called with mutex locked */
 499 static
 500 void cleanup_thread(struct registry_chunk *chunk,
 501                 struct rcu_reader *rcu_reader_reg)
 502 {
 503         rcu_reader_reg->ctr = 0;
 504         cds_list_del(&rcu_reader_reg->node);
 505         rcu_reader_reg->tid = 0;
 506         rcu_reader_reg->alloc = 0;
 507         chunk->used--;
 508 }
 509
 510 static
 511 struct registry_chunk *find_chunk(struct rcu_reader *rcu_reader_reg)
 512 {
 513         struct registry_chunk *chunk;
 514
 515         cds_list_for_each_entry(chunk, &registry_arena.chunk_list, node) {
 516                 if (rcu_reader_reg < (struct urcu_bp_reader *) &chunk->readers[0])
 517                         continue;
 518                 if (rcu_reader_reg >= (struct urcu_bp_reader *) &chunk->readers[chunk->capacity])
 519                         continue;
 520                 return chunk;
 521         }
 522         return NULL;
 523 }
 524
 525 /* Called with signals off and mutex locked */
 526 static
 527 void remove_thread(struct rcu_reader *rcu_reader_reg)
 528 {
 529         cleanup_thread(find_chunk(rcu_reader_reg), rcu_reader_reg);
 530         URCU_TLS(urcu_bp_reader) = NULL;
 531 }
 532
 533 /* Disable signals, take mutex, add to registry */
 534 void urcu_bp_register(void)
 535 {
 536         sigset_t newmask, oldmask;
 537         int ret;
 538
 539         ret = sigfillset(&newmask);
 540         if (ret)
 541                 abort();
 542         ret = pthread_sigmask(SIG_BLOCK, &newmask, &oldmask);
 543         if (ret)
 544                 abort();
 545
 546         /*
 547          * Check if a signal concurrently registered our thread since
 548          * the check in rcu_read_lock().
 549          */
 550         if (URCU_TLS(urcu_bp_reader))
 551                 goto end;
 552
 553         /*
 554          * Take care of early registration before urcu_bp constructor.
 555          */
 556         _urcu_bp_init();
 557
 558         mutex_lock(&rcu_registry_lock);
 559         add_thread();
 560         mutex_unlock(&rcu_registry_lock);
 561 end:
 562         ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL);
 563         if (ret)
 564                 abort();
 565 }
 566
 567 void urcu_bp_register_thread(void)
 568 {
 569         if (caa_unlikely(!URCU_TLS(urcu_bp_reader)))
 570                 urcu_bp_register(); /* If not yet registered. */
 571 }
 572
 573 /* Disable signals, take mutex, remove from registry */
 574 static
 575 void urcu_bp_unregister(struct rcu_reader *rcu_reader_reg)
 576 {
 577         sigset_t newmask, oldmask;
 578         int ret;
 579
 580         ret = sigfillset(&newmask);
 581         if (ret)
 582                 abort();
 583         ret = pthread_sigmask(SIG_BLOCK, &newmask, &oldmask);
 584         if (ret)
 585                 abort();
 586
 587         mutex_lock(&rcu_registry_lock);
 588         remove_thread(rcu_reader_reg);
 589         mutex_unlock(&rcu_registry_lock);
 590         ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL);
 591         if (ret)
 592                 abort();
 593         urcu_bp_exit();
 594 }
 595
 596 /*
 597  * Remove thread from the registry when it exits, and flag it as
 598  * destroyed so garbage collection can take care of it.
 599  */
 600 static
 601 void urcu_bp_thread_exit_notifier(void *rcu_key)
 602 {
 603         urcu_bp_unregister(rcu_key);
 604 }
 605
 606 #ifdef CONFIG_RCU_FORCE_SYS_MEMBARRIER
 607 static
 608 void urcu_bp_sys_membarrier_status(bool available)
 609 {
 610         if (!available)
 611                 abort();
 612 }
 613 #else
 614 static
 615 void urcu_bp_sys_membarrier_status(bool available)
 616 {
 617         if (!available)
 618                 return;
 619         urcu_bp_has_sys_membarrier = 1;
 620 }
 621 #endif
 622
 623 static
 624 void urcu_bp_sys_membarrier_init(void)
 625 {
 626         bool available = false;
 627         int mask;
 628
 629         mask = membarrier(MEMBARRIER_CMD_QUERY, 0);
 630         if (mask >= 0) {
 631                 if (mask & MEMBARRIER_CMD_PRIVATE_EXPEDITED) {
 632                         if (membarrier(MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED, 0))
 633                                 urcu_die(errno);
 634                         available = true;
 635                 }
 636         }
 637         urcu_bp_sys_membarrier_status(available);
 638 }
 639
 640 static
 641 void _urcu_bp_init(void)
 642 {
 643         mutex_lock(&init_lock);
 644         if (!urcu_bp_refcount++) {
 645                 int ret;
 646
 647                 ret = pthread_key_create(&urcu_bp_key,
 648                                 urcu_bp_thread_exit_notifier);
 649                 if (ret)
 650                         abort();
 651                 urcu_bp_sys_membarrier_init();
 652                 initialized = 1;
 653         }
 654         mutex_unlock(&init_lock);
 655 }
 656
 657 static
 658 void urcu_bp_exit(void)
 659 {
 660         mutex_lock(&init_lock);
 661         if (!--urcu_bp_refcount) {
 662                 struct registry_chunk *chunk, *tmp;
 663                 int ret;
 664
 665                 cds_list_for_each_entry_safe(chunk, tmp,
 666                                 &registry_arena.chunk_list, node) {
 667                         munmap((void *) chunk, chunk_allocation_size(chunk->capacity));
 668                 }
 669                 CDS_INIT_LIST_HEAD(&registry_arena.chunk_list);
 670                 ret = pthread_key_delete(urcu_bp_key);
 671                 if (ret)
 672                         abort();
 673         }
 674         mutex_unlock(&init_lock);
 675 }
 676
 677 static
 678 void urcu_bp_exit_destructor(void)
 679 {
 680         urcu_call_rcu_exit();
 681         urcu_bp_exit();
 682 }
 683
 684 /*
 685  * Holding the rcu_gp_lock and rcu_registry_lock across fork will make
 686  * sure we fork() don't race with a concurrent thread executing with
 687  * any of those locks held. This ensures that the registry and data
 688  * protected by rcu_gp_lock are in a coherent state in the child.
 689  */
 690 void urcu_bp_before_fork(void)
 691 {
 692         sigset_t newmask, oldmask;
 693         int ret;
 694
 695         ret = sigfillset(&newmask);
 696         assert(!ret);
 697         ret = pthread_sigmask(SIG_BLOCK, &newmask, &oldmask);
 698         assert(!ret);
 699         mutex_lock(&rcu_gp_lock);
 700         mutex_lock(&rcu_registry_lock);
 701         saved_fork_signal_mask = oldmask;
 702 }
 703
 704 void urcu_bp_after_fork_parent(void)
 705 {
 706         sigset_t oldmask;
 707         int ret;
 708
 709         oldmask = saved_fork_signal_mask;
 710         mutex_unlock(&rcu_registry_lock);
 711         mutex_unlock(&rcu_gp_lock);
 712         ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL);
 713         assert(!ret);
 714 }
 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
 725         cds_list_for_each_entry(chunk, &registry_arena.chunk_list, node) {
 726                 size_t spot_idx;
 727
 728                 for (spot_idx = 0; spot_idx < chunk->capacity; spot_idx++) {
 729                         struct urcu_bp_reader *reader = &chunk->readers[spot_idx];
 730
 731                         if (!reader->alloc)
 732                                 continue;
 733                         if (reader->tid == pthread_self())
 734                                 continue;
 735                         cleanup_thread(chunk, reader);
 736                 }
 737         }
 738 }
 739
 740 void urcu_bp_after_fork_child(void)
 741 {
 742         sigset_t oldmask;
 743         int ret;
 744
 745         urcu_bp_prune_registry();
 746         oldmask = saved_fork_signal_mask;
 747         mutex_unlock(&rcu_registry_lock);
 748         mutex_unlock(&rcu_gp_lock);
 749         ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL);
 750         assert(!ret);
 751 }
 752
 753 void *urcu_bp_dereference_sym(void *p)
 754 {
 755         return _rcu_dereference(p);
 756 }
 757
 758 void *urcu_bp_set_pointer_sym(void **p, void *v)
 759 {
 760         cmm_wmb();
 761         uatomic_set(p, v);
 762         return v;
 763 }
 764
 765 void *urcu_bp_xchg_pointer_sym(void **p, void *v)
 766 {
 767         cmm_wmb();
 768         return uatomic_xchg(p, v);
 769 }
 770
 771 void *urcu_bp_cmpxchg_pointer_sym(void **p, void *old, void *_new)
 772 {
 773         cmm_wmb();
 774         return uatomic_cmpxchg(p, old, _new);
 775 }
 776
 777 DEFINE_RCU_FLAVOR(rcu_flavor);
 778
 779 #include "urcu-call-rcu-impl.h"
 780 #include "urcu-defer-impl.h"