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_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 URCU_ATTR_ALIAS("urcu_bp_gp") extern struct urcu_bp_gp rcu_gp_bp;
 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 DEFINE_URCU_TLS_ALIAS(struct urcu_bp_reader *, urcu_bp_reader, rcu_reader_bp);
 159
 160 static CDS_LIST_HEAD(registry);
 161
 162 struct registry_chunk {
 163         size_t data_len;                /* data length */
 164         size_t used;                    /* amount of data used */
 165         struct cds_list_head node;      /* chunk_list node */
 166         char data[];
 167 };
 168
 169 struct registry_arena {
 170         struct cds_list_head chunk_list;
 171 };
 172
 173 static struct registry_arena registry_arena = {
 174         .chunk_list = CDS_LIST_HEAD_INIT(registry_arena.chunk_list),
 175 };
 176
 177 /* Saved fork signal mask, protected by rcu_gp_lock */
 178 static sigset_t saved_fork_signal_mask;
 179
 180 static void mutex_lock(pthread_mutex_t *mutex)
 181 {
 182         int ret;
 183
 184 #ifndef DISTRUST_SIGNALS_EXTREME
 185         ret = pthread_mutex_lock(mutex);
 186         if (ret)
 187                 urcu_die(ret);
 188 #else /* #ifndef DISTRUST_SIGNALS_EXTREME */
 189         while ((ret = pthread_mutex_trylock(mutex)) != 0) {
 190                 if (ret != EBUSY && ret != EINTR)
 191                         urcu_die(ret);
 192                 poll(NULL,0,10);
 193         }
 194 #endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */
 195 }
 196
 197 static void mutex_unlock(pthread_mutex_t *mutex)
 198 {
 199         int ret;
 200
 201         ret = pthread_mutex_unlock(mutex);
 202         if (ret)
 203                 urcu_die(ret);
 204 }
 205
 206 static void smp_mb_master(void)
 207 {
 208         if (caa_likely(urcu_bp_has_sys_membarrier)) {
 209                 if (membarrier(MEMBARRIER_CMD_PRIVATE_EXPEDITED, 0))
 210                         urcu_die(errno);
 211         } else {
 212                 cmm_smp_mb();
 213         }
 214 }
 215
 216 /*
 217  * Always called with rcu_registry lock held. Releases this lock between
 218  * iterations and grabs it again. Holds the lock when it returns.
 219  */
 220 static void wait_for_readers(struct cds_list_head *input_readers,
 221                         struct cds_list_head *cur_snap_readers,
 222                         struct cds_list_head *qsreaders)
 223 {
 224         unsigned int wait_loops = 0;
 225         struct urcu_bp_reader *index, *tmp;
 226
 227         /*
 228          * Wait for each thread URCU_TLS(urcu_bp_reader).ctr to either
 229          * indicate quiescence (not nested), or observe the current
 230          * rcu_gp.ctr value.
 231          */
 232         for (;;) {
 233                 if (wait_loops < RCU_QS_ACTIVE_ATTEMPTS)
 234                         wait_loops++;
 235
 236                 cds_list_for_each_entry_safe(index, tmp, input_readers, node) {
 237                         switch (urcu_bp_reader_state(&index->ctr)) {
 238                         case URCU_BP_READER_ACTIVE_CURRENT:
 239                                 if (cur_snap_readers) {
 240                                         cds_list_move(&index->node,
 241                                                 cur_snap_readers);
 242                                         break;
 243                                 }
 244                                 /* Fall-through */
 245                         case URCU_BP_READER_INACTIVE:
 246                                 cds_list_move(&index->node, qsreaders);
 247                                 break;
 248                         case URCU_BP_READER_ACTIVE_OLD:
 249                                 /*
 250                                  * Old snapshot. Leaving node in
 251                                  * input_readers will make us busy-loop
 252                                  * until the snapshot becomes current or
 253                                  * the reader becomes inactive.
 254                                  */
 255                                 break;
 256                         }
 257                 }
 258
 259                 if (cds_list_empty(input_readers)) {
 260                         break;
 261                 } else {
 262                         /* Temporarily unlock the registry lock. */
 263                         mutex_unlock(&rcu_registry_lock);
 264                         if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS)
 265                                 (void) poll(NULL, 0, RCU_SLEEP_DELAY_MS);
 266                         else
 267                                 caa_cpu_relax();
 268                         /* Re-lock the registry lock before the next loop. */
 269                         mutex_lock(&rcu_registry_lock);
 270                 }
 271         }
 272 }
 273
 274 void urcu_bp_synchronize_rcu(void)
 275 {
 276         CDS_LIST_HEAD(cur_snap_readers);
 277         CDS_LIST_HEAD(qsreaders);
 278         sigset_t newmask, oldmask;
 279         int ret;
 280
 281         ret = sigfillset(&newmask);
 282         assert(!ret);
 283         ret = pthread_sigmask(SIG_BLOCK, &newmask, &oldmask);
 284         assert(!ret);
 285
 286         mutex_lock(&rcu_gp_lock);
 287
 288         mutex_lock(&rcu_registry_lock);
 289
 290         if (cds_list_empty(&registry))
 291                 goto out;
 292
 293         /* All threads should read qparity before accessing data structure
 294          * where new ptr points to. */
 295         /* Write new ptr before changing the qparity */
 296         smp_mb_master();
 297
 298         /*
 299          * Wait for readers to observe original parity or be quiescent.
 300          * wait_for_readers() can release and grab again rcu_registry_lock
 301          * interally.
 302          */
 303         wait_for_readers(&registry, &cur_snap_readers, &qsreaders);
 304
 305         /*
 306          * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
 307          * model easier to understand. It does not have a big performance impact
 308          * anyway, given this is the write-side.
 309          */
 310         cmm_smp_mb();
 311
 312         /* Switch parity: 0 -> 1, 1 -> 0 */
 313         CMM_STORE_SHARED(rcu_gp.ctr, rcu_gp.ctr ^ URCU_BP_GP_CTR_PHASE);
 314
 315         /*
 316          * Must commit qparity update to memory before waiting for other parity
 317          * quiescent state. Failure to do so could result in the writer waiting
 318          * forever while new readers are always accessing data (no progress).
 319          * Ensured by CMM_STORE_SHARED and CMM_LOAD_SHARED.
 320          */
 321
 322         /*
 323          * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
 324          * model easier to understand. It does not have a big performance impact
 325          * anyway, given this is the write-side.
 326          */
 327         cmm_smp_mb();
 328
 329         /*
 330          * Wait for readers to observe new parity or be quiescent.
 331          * wait_for_readers() can release and grab again rcu_registry_lock
 332          * interally.
 333          */
 334         wait_for_readers(&cur_snap_readers, NULL, &qsreaders);
 335
 336         /*
 337          * Put quiescent reader list back into registry.
 338          */
 339         cds_list_splice(&qsreaders, &registry);
 340
 341         /*
 342          * Finish waiting for reader threads before letting the old ptr being
 343          * freed.
 344          */
 345         smp_mb_master();
 346 out:
 347         mutex_unlock(&rcu_registry_lock);
 348         mutex_unlock(&rcu_gp_lock);
 349         ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL);
 350         assert(!ret);
 351 }
 352 URCU_ATTR_ALIAS("urcu_bp_synchronize_rcu") void synchronize_rcu_bp();
 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 URCU_ATTR_ALIAS("urcu_bp_read_lock") void rcu_read_lock_bp();
 363
 364 void urcu_bp_read_unlock(void)
 365 {
 366         _urcu_bp_read_unlock();
 367 }
 368 URCU_ATTR_ALIAS("urcu_bp_read_unlock") void rcu_read_unlock_bp();
 369
 370 int urcu_bp_read_ongoing(void)
 371 {
 372         return _urcu_bp_read_ongoing();
 373 }
 374 URCU_ATTR_ALIAS("urcu_bp_read_ongoing") int rcu_read_ongoing_bp();
 375
 376 /*
 377  * Only grow for now. If empty, allocate a ARENA_INIT_ALLOC sized chunk.
 378  * Else, try expanding the last chunk. If this fails, allocate a new
 379  * chunk twice as big as the last chunk.
 380  * Memory used by chunks _never_ moves. A chunk could theoretically be
 381  * freed when all "used" slots are released, but we don't do it at this
 382  * point.
 383  */
 384 static
 385 void expand_arena(struct registry_arena *arena)
 386 {
 387         struct registry_chunk *new_chunk, *last_chunk;
 388         size_t old_chunk_len, new_chunk_len;
 389
 390         /* No chunk. */
 391         if (cds_list_empty(&arena->chunk_list)) {
 392                 assert(ARENA_INIT_ALLOC >=
 393                         sizeof(struct registry_chunk)
 394                         + sizeof(struct rcu_reader));
 395                 new_chunk_len = ARENA_INIT_ALLOC;
 396                 new_chunk = (struct registry_chunk *) mmap(NULL,
 397                         new_chunk_len,
 398                         PROT_READ | PROT_WRITE,
 399                         MAP_ANONYMOUS | MAP_PRIVATE,
 400                         -1, 0);
 401                 if (new_chunk == MAP_FAILED)
 402                         abort();
 403                 memset(new_chunk, 0, new_chunk_len);
 404                 new_chunk->data_len =
 405                         new_chunk_len - sizeof(struct registry_chunk);
 406                 cds_list_add_tail(&new_chunk->node, &arena->chunk_list);
 407                 return;         /* We're done. */
 408         }
 409
 410         /* Try expanding last chunk. */
 411         last_chunk = cds_list_entry(arena->chunk_list.prev,
 412                 struct registry_chunk, node);
 413         old_chunk_len =
 414                 last_chunk->data_len + sizeof(struct registry_chunk);
 415         new_chunk_len = old_chunk_len << 1;
 416
 417         /* Don't allow memory mapping to move, just expand. */
 418         new_chunk = mremap_wrapper(last_chunk, old_chunk_len,
 419                 new_chunk_len, 0);
 420         if (new_chunk != MAP_FAILED) {
 421                 /* Should not have moved. */
 422                 assert(new_chunk == last_chunk);
 423                 memset((char *) last_chunk + old_chunk_len, 0,
 424                         new_chunk_len - old_chunk_len);
 425                 last_chunk->data_len =
 426                         new_chunk_len - sizeof(struct registry_chunk);
 427                 return;         /* We're done. */
 428         }
 429
 430         /* Remap did not succeed, we need to add a new chunk. */
 431         new_chunk = (struct registry_chunk *) mmap(NULL,
 432                 new_chunk_len,
 433                 PROT_READ | PROT_WRITE,
 434                 MAP_ANONYMOUS | MAP_PRIVATE,
 435                 -1, 0);
 436         if (new_chunk == MAP_FAILED)
 437                 abort();
 438         memset(new_chunk, 0, new_chunk_len);
 439         new_chunk->data_len =
 440                 new_chunk_len - sizeof(struct registry_chunk);
 441         cds_list_add_tail(&new_chunk->node, &arena->chunk_list);
 442 }
 443
 444 static
 445 struct rcu_reader *arena_alloc(struct registry_arena *arena)
 446 {
 447         struct registry_chunk *chunk;
 448         struct rcu_reader *rcu_reader_reg;
 449         int expand_done = 0;    /* Only allow to expand once per alloc */
 450         size_t len = sizeof(struct rcu_reader);
 451
 452 retry:
 453         cds_list_for_each_entry(chunk, &arena->chunk_list, node) {
 454                 if (chunk->data_len - chunk->used < len)
 455                         continue;
 456                 /* Find spot */
 457                 for (rcu_reader_reg = (struct rcu_reader *) &chunk->data[0];
 458                                 rcu_reader_reg < (struct rcu_reader *) &chunk->data[chunk->data_len];
 459                                 rcu_reader_reg++) {
 460                         if (!rcu_reader_reg->alloc) {
 461                                 rcu_reader_reg->alloc = 1;
 462                                 chunk->used += len;
 463                                 return rcu_reader_reg;
 464                         }
 465                 }
 466         }
 467
 468         if (!expand_done) {
 469                 expand_arena(arena);
 470                 expand_done = 1;
 471                 goto retry;
 472         }
 473
 474         return NULL;
 475 }
 476
 477 /* Called with signals off and mutex locked */
 478 static
 479 void add_thread(void)
 480 {
 481         struct rcu_reader *rcu_reader_reg;
 482         int ret;
 483
 484         rcu_reader_reg = arena_alloc(&registry_arena);
 485         if (!rcu_reader_reg)
 486                 abort();
 487         ret = pthread_setspecific(urcu_bp_key, rcu_reader_reg);
 488         if (ret)
 489                 abort();
 490
 491         /* Add to registry */
 492         rcu_reader_reg->tid = pthread_self();
 493         assert(rcu_reader_reg->ctr == 0);
 494         cds_list_add(&rcu_reader_reg->node, &registry);
 495         /*
 496          * Reader threads are pointing to the reader registry. This is
 497          * why its memory should never be relocated.
 498          */
 499         URCU_TLS(urcu_bp_reader) = rcu_reader_reg;
 500 }
 501
 502 /* Called with mutex locked */
 503 static
 504 void cleanup_thread(struct registry_chunk *chunk,
 505                 struct rcu_reader *rcu_reader_reg)
 506 {
 507         rcu_reader_reg->ctr = 0;
 508         cds_list_del(&rcu_reader_reg->node);
 509         rcu_reader_reg->tid = 0;
 510         rcu_reader_reg->alloc = 0;
 511         chunk->used -= sizeof(struct rcu_reader);
 512 }
 513
 514 static
 515 struct registry_chunk *find_chunk(struct rcu_reader *rcu_reader_reg)
 516 {
 517         struct registry_chunk *chunk;
 518
 519         cds_list_for_each_entry(chunk, &registry_arena.chunk_list, node) {
 520                 if (rcu_reader_reg < (struct rcu_reader *) &chunk->data[0])
 521                         continue;
 522                 if (rcu_reader_reg >= (struct rcu_reader *) &chunk->data[chunk->data_len])
 523                         continue;
 524                 return chunk;
 525         }
 526         return NULL;
 527 }
 528
 529 /* Called with signals off and mutex locked */
 530 static
 531 void remove_thread(struct rcu_reader *rcu_reader_reg)
 532 {
 533         cleanup_thread(find_chunk(rcu_reader_reg), rcu_reader_reg);
 534         URCU_TLS(urcu_bp_reader) = NULL;
 535 }
 536
 537 /* Disable signals, take mutex, add to registry */
 538 void urcu_bp_register(void)
 539 {
 540         sigset_t newmask, oldmask;
 541         int ret;
 542
 543         ret = sigfillset(&newmask);
 544         if (ret)
 545                 abort();
 546         ret = pthread_sigmask(SIG_BLOCK, &newmask, &oldmask);
 547         if (ret)
 548                 abort();
 549
 550         /*
 551          * Check if a signal concurrently registered our thread since
 552          * the check in rcu_read_lock().
 553          */
 554         if (URCU_TLS(urcu_bp_reader))
 555                 goto end;
 556
 557         /*
 558          * Take care of early registration before urcu_bp constructor.
 559          */
 560         _urcu_bp_init();
 561
 562         mutex_lock(&rcu_registry_lock);
 563         add_thread();
 564         mutex_unlock(&rcu_registry_lock);
 565 end:
 566         ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL);
 567         if (ret)
 568                 abort();
 569 }
 570 URCU_ATTR_ALIAS("urcu_bp_register") void rcu_bp_register();
 571
 572 void urcu_bp_register_thread(void)
 573 {
 574         if (caa_unlikely(!URCU_TLS(urcu_bp_reader)))
 575                 urcu_bp_register(); /* If not yet registered. */
 576 }
 577
 578 /* Disable signals, take mutex, remove from registry */
 579 static
 580 void urcu_bp_unregister(struct rcu_reader *rcu_reader_reg)
 581 {
 582         sigset_t newmask, oldmask;
 583         int ret;
 584
 585         ret = sigfillset(&newmask);
 586         if (ret)
 587                 abort();
 588         ret = pthread_sigmask(SIG_BLOCK, &newmask, &oldmask);
 589         if (ret)
 590                 abort();
 591
 592         mutex_lock(&rcu_registry_lock);
 593         remove_thread(rcu_reader_reg);
 594         mutex_unlock(&rcu_registry_lock);
 595         ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL);
 596         if (ret)
 597                 abort();
 598         urcu_bp_exit();
 599 }
 600
 601 /*
 602  * Remove thread from the registry when it exits, and flag it as
 603  * destroyed so garbage collection can take care of it.
 604  */
 605 static
 606 void urcu_bp_thread_exit_notifier(void *rcu_key)
 607 {
 608         urcu_bp_unregister(rcu_key);
 609 }
 610
 611 #ifdef CONFIG_RCU_FORCE_SYS_MEMBARRIER
 612 static
 613 void urcu_bp_sys_membarrier_status(bool available)
 614 {
 615         if (!available)
 616                 abort();
 617 }
 618 #else
 619 static
 620 void urcu_bp_sys_membarrier_status(bool available)
 621 {
 622         if (!available)
 623                 return;
 624         urcu_bp_has_sys_membarrier = 1;
 625 }
 626 #endif
 627
 628 static
 629 void urcu_bp_sys_membarrier_init(void)
 630 {
 631         bool available = false;
 632         int mask;
 633
 634         mask = membarrier(MEMBARRIER_CMD_QUERY, 0);
 635         if (mask >= 0) {
 636                 if (mask & MEMBARRIER_CMD_PRIVATE_EXPEDITED) {
 637                         if (membarrier(MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED, 0))
 638                                 urcu_die(errno);
 639                         available = true;
 640                 }
 641         }
 642         urcu_bp_sys_membarrier_status(available);
 643 }
 644
 645 static
 646 void _urcu_bp_init(void)
 647 {
 648         mutex_lock(&init_lock);
 649         if (!urcu_bp_refcount++) {
 650                 int ret;
 651
 652                 ret = pthread_key_create(&urcu_bp_key,
 653                                 urcu_bp_thread_exit_notifier);
 654                 if (ret)
 655                         abort();
 656                 urcu_bp_sys_membarrier_init();
 657                 initialized = 1;
 658         }
 659         mutex_unlock(&init_lock);
 660 }
 661
 662 static
 663 void urcu_bp_exit(void)
 664 {
 665         mutex_lock(&init_lock);
 666         if (!--urcu_bp_refcount) {
 667                 struct registry_chunk *chunk, *tmp;
 668                 int ret;
 669
 670                 cds_list_for_each_entry_safe(chunk, tmp,
 671                                 &registry_arena.chunk_list, node) {
 672                         munmap((void *) chunk, chunk->data_len
 673                                         + sizeof(struct registry_chunk));
 674                 }
 675                 CDS_INIT_LIST_HEAD(&registry_arena.chunk_list);
 676                 ret = pthread_key_delete(urcu_bp_key);
 677                 if (ret)
 678                         abort();
 679         }
 680         mutex_unlock(&init_lock);
 681 }
 682
 683 /*
 684  * Holding the rcu_gp_lock and rcu_registry_lock across fork will make
 685  * sure we fork() don't race with a concurrent thread executing with
 686  * any of those locks held. This ensures that the registry and data
 687  * protected by rcu_gp_lock are in a coherent state in the child.
 688  */
 689 void urcu_bp_before_fork(void)
 690 {
 691         sigset_t newmask, oldmask;
 692         int ret;
 693
 694         ret = sigfillset(&newmask);
 695         assert(!ret);
 696         ret = pthread_sigmask(SIG_BLOCK, &newmask, &oldmask);
 697         assert(!ret);
 698         mutex_lock(&rcu_gp_lock);
 699         mutex_lock(&rcu_registry_lock);
 700         saved_fork_signal_mask = oldmask;
 701 }
 702 URCU_ATTR_ALIAS("urcu_bp_before_fork") void rcu_bp_before_fork();
 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 URCU_ATTR_ALIAS("urcu_bp_after_fork_parent")
 716 void rcu_bp_after_fork_parent(void);
 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         assert(!ret);
 752 }
 753 URCU_ATTR_ALIAS("urcu_bp_after_fork_child")
 754 void rcu_bp_after_fork_child(void);
 755
 756 void *urcu_bp_dereference_sym(void *p)
 757 {
 758         return _rcu_dereference(p);
 759 }
 760 URCU_ATTR_ALIAS("urcu_bp_dereference_sym")
 761 void *rcu_dereference_sym_bp();
 762
 763 void *urcu_bp_set_pointer_sym(void **p, void *v)
 764 {
 765         cmm_wmb();
 766         uatomic_set(p, v);
 767         return v;
 768 }
 769 URCU_ATTR_ALIAS("urcu_bp_set_pointer_sym")
 770 void *rcu_set_pointer_sym_bp();
 771
 772 void *urcu_bp_xchg_pointer_sym(void **p, void *v)
 773 {
 774         cmm_wmb();
 775         return uatomic_xchg(p, v);
 776 }
 777 URCU_ATTR_ALIAS("urcu_bp_xchg_pointer_sym")
 778 void *rcu_xchg_pointer_sym_bp();
 779
 780 void *urcu_bp_cmpxchg_pointer_sym(void **p, void *old, void *_new)
 781 {
 782         cmm_wmb();
 783         return uatomic_cmpxchg(p, old, _new);
 784 }
 785 URCU_ATTR_ALIAS("urcu_bp_cmpxchg_pointer_sym")
 786 void *rcu_cmpxchg_pointer_sym_bp();
 787
 788 DEFINE_RCU_FLAVOR(rcu_flavor);
 789 DEFINE_RCU_FLAVOR_ALIAS(rcu_flavor, alias_rcu_flavor);
 790
 791 #include "urcu-call-rcu-impl.h"
 792 #include "urcu-defer-impl.h"