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 static void urcu_call_rcu_exit(void);
 125
 126 #ifndef CONFIG_RCU_FORCE_SYS_MEMBARRIER
 127 int urcu_bp_has_sys_membarrier;
 128 #endif
 129
 130 /*
 131  * rcu_gp_lock ensures mutual exclusion between threads calling
 132  * synchronize_rcu().
 133  */
 134 static pthread_mutex_t rcu_gp_lock = PTHREAD_MUTEX_INITIALIZER;
 135 /*
 136  * rcu_registry_lock ensures mutual exclusion between threads
 137  * registering and unregistering themselves to/from the registry, and
 138  * with threads reading that registry from synchronize_rcu(). However,
 139  * this lock is not held all the way through the completion of awaiting
 140  * for the grace period. It is sporadically released between iterations
 141  * on the registry.
 142  * rcu_registry_lock may nest inside rcu_gp_lock.
 143  */
 144 static pthread_mutex_t rcu_registry_lock = PTHREAD_MUTEX_INITIALIZER;
 145
 146 static pthread_mutex_t init_lock = PTHREAD_MUTEX_INITIALIZER;
 147 static int initialized;
 148
 149 static pthread_key_t urcu_bp_key;
 150
 151 struct urcu_bp_gp urcu_bp_gp = { .ctr = URCU_BP_GP_COUNT };
 152 URCU_ATTR_ALIAS("urcu_bp_gp") extern struct urcu_bp_gp rcu_gp_bp;
 153
 154 /*
 155  * Pointer to registry elements. Written to only by each individual reader. Read
 156  * by both the reader and the writers.
 157  */
 158 DEFINE_URCU_TLS(struct urcu_bp_reader *, urcu_bp_reader);
 159 DEFINE_URCU_TLS_ALIAS(struct urcu_bp_reader *, urcu_bp_reader, rcu_reader_bp);
 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         assert(!ret);
 284         ret = pthread_sigmask(SIG_BLOCK, &newmask, &oldmask);
 285         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          * interally.
 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          * interally.
 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         assert(!ret);
 352 }
 353 URCU_ATTR_ALIAS("urcu_bp_synchronize_rcu") void synchronize_rcu_bp();
 354
 355 /*
 356  * library wrappers to be used by non-LGPL compatible source code.
 357  */
 358
 359 void urcu_bp_read_lock(void)
 360 {
 361         _urcu_bp_read_lock();
 362 }
 363 URCU_ATTR_ALIAS("urcu_bp_read_lock") void rcu_read_lock_bp();
 364
 365 void urcu_bp_read_unlock(void)
 366 {
 367         _urcu_bp_read_unlock();
 368 }
 369 URCU_ATTR_ALIAS("urcu_bp_read_unlock") void rcu_read_unlock_bp();
 370
 371 int urcu_bp_read_ongoing(void)
 372 {
 373         return _urcu_bp_read_ongoing();
 374 }
 375 URCU_ATTR_ALIAS("urcu_bp_read_ongoing") int rcu_read_ongoing_bp();
 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_len, new_chunk_len;
 390
 391         /* No chunk. */
 392         if (cds_list_empty(&arena->chunk_list)) {
 393                 assert(ARENA_INIT_ALLOC >=
 394                         sizeof(struct registry_chunk)
 395                         + sizeof(struct rcu_reader));
 396                 new_chunk_len = ARENA_INIT_ALLOC;
 397                 new_chunk = (struct registry_chunk *) mmap(NULL,
 398                         new_chunk_len,
 399                         PROT_READ | PROT_WRITE,
 400                         MAP_ANONYMOUS | MAP_PRIVATE,
 401                         -1, 0);
 402                 if (new_chunk == MAP_FAILED)
 403                         abort();
 404                 memset(new_chunk, 0, new_chunk_len);
 405                 new_chunk->data_len =
 406                         new_chunk_len - sizeof(struct registry_chunk);
 407                 cds_list_add_tail(&new_chunk->node, &arena->chunk_list);
 408                 return;         /* We're done. */
 409         }
 410
 411         /* Try expanding last chunk. */
 412         last_chunk = cds_list_entry(arena->chunk_list.prev,
 413                 struct registry_chunk, node);
 414         old_chunk_len =
 415                 last_chunk->data_len + sizeof(struct registry_chunk);
 416         new_chunk_len = old_chunk_len << 1;
 417
 418         /* Don't allow memory mapping to move, just expand. */
 419         new_chunk = mremap_wrapper(last_chunk, old_chunk_len,
 420                 new_chunk_len, 0);
 421         if (new_chunk != MAP_FAILED) {
 422                 /* Should not have moved. */
 423                 assert(new_chunk == last_chunk);
 424                 memset((char *) last_chunk + old_chunk_len, 0,
 425                         new_chunk_len - old_chunk_len);
 426                 last_chunk->data_len =
 427                         new_chunk_len - sizeof(struct registry_chunk);
 428                 return;         /* We're done. */
 429         }
 430
 431         /* Remap did not succeed, we need to add a new chunk. */
 432         new_chunk = (struct registry_chunk *) mmap(NULL,
 433                 new_chunk_len,
 434                 PROT_READ | PROT_WRITE,
 435                 MAP_ANONYMOUS | MAP_PRIVATE,
 436                 -1, 0);
 437         if (new_chunk == MAP_FAILED)
 438                 abort();
 439         memset(new_chunk, 0, new_chunk_len);
 440         new_chunk->data_len =
 441                 new_chunk_len - sizeof(struct registry_chunk);
 442         cds_list_add_tail(&new_chunk->node, &arena->chunk_list);
 443 }
 444
 445 static
 446 struct rcu_reader *arena_alloc(struct registry_arena *arena)
 447 {
 448         struct registry_chunk *chunk;
 449         struct rcu_reader *rcu_reader_reg;
 450         int expand_done = 0;    /* Only allow to expand once per alloc */
 451         size_t len = sizeof(struct rcu_reader);
 452
 453 retry:
 454         cds_list_for_each_entry(chunk, &arena->chunk_list, node) {
 455                 if (chunk->data_len - chunk->used < len)
 456                         continue;
 457                 /* Find spot */
 458                 for (rcu_reader_reg = (struct rcu_reader *) &chunk->data[0];
 459                                 rcu_reader_reg < (struct rcu_reader *) &chunk->data[chunk->data_len];
 460                                 rcu_reader_reg++) {
 461                         if (!rcu_reader_reg->alloc) {
 462                                 rcu_reader_reg->alloc = 1;
 463                                 chunk->used += len;
 464                                 return rcu_reader_reg;
 465                         }
 466                 }
 467         }
 468
 469         if (!expand_done) {
 470                 expand_arena(arena);
 471                 expand_done = 1;
 472                 goto retry;
 473         }
 474
 475         return NULL;
 476 }
 477
 478 /* Called with signals off and mutex locked */
 479 static
 480 void add_thread(void)
 481 {
 482         struct rcu_reader *rcu_reader_reg;
 483         int ret;
 484
 485         rcu_reader_reg = arena_alloc(&registry_arena);
 486         if (!rcu_reader_reg)
 487                 abort();
 488         ret = pthread_setspecific(urcu_bp_key, rcu_reader_reg);
 489         if (ret)
 490                 abort();
 491
 492         /* Add to registry */
 493         rcu_reader_reg->tid = pthread_self();
 494         assert(rcu_reader_reg->ctr == 0);
 495         cds_list_add(&rcu_reader_reg->node, &registry);
 496         /*
 497          * Reader threads are pointing to the reader registry. This is
 498          * why its memory should never be relocated.
 499          */
 500         URCU_TLS(urcu_bp_reader) = rcu_reader_reg;
 501 }
 502
 503 /* Called with mutex locked */
 504 static
 505 void cleanup_thread(struct registry_chunk *chunk,
 506                 struct rcu_reader *rcu_reader_reg)
 507 {
 508         rcu_reader_reg->ctr = 0;
 509         cds_list_del(&rcu_reader_reg->node);
 510         rcu_reader_reg->tid = 0;
 511         rcu_reader_reg->alloc = 0;
 512         chunk->used -= sizeof(struct rcu_reader);
 513 }
 514
 515 static
 516 struct registry_chunk *find_chunk(struct rcu_reader *rcu_reader_reg)
 517 {
 518         struct registry_chunk *chunk;
 519
 520         cds_list_for_each_entry(chunk, &registry_arena.chunk_list, node) {
 521                 if (rcu_reader_reg < (struct rcu_reader *) &chunk->data[0])
 522                         continue;
 523                 if (rcu_reader_reg >= (struct rcu_reader *) &chunk->data[chunk->data_len])
 524                         continue;
 525                 return chunk;
 526         }
 527         return NULL;
 528 }
 529
 530 /* Called with signals off and mutex locked */
 531 static
 532 void remove_thread(struct rcu_reader *rcu_reader_reg)
 533 {
 534         cleanup_thread(find_chunk(rcu_reader_reg), rcu_reader_reg);
 535         URCU_TLS(urcu_bp_reader) = NULL;
 536 }
 537
 538 /* Disable signals, take mutex, add to registry */
 539 void urcu_bp_register(void)
 540 {
 541         sigset_t newmask, oldmask;
 542         int ret;
 543
 544         ret = sigfillset(&newmask);
 545         if (ret)
 546                 abort();
 547         ret = pthread_sigmask(SIG_BLOCK, &newmask, &oldmask);
 548         if (ret)
 549                 abort();
 550
 551         /*
 552          * Check if a signal concurrently registered our thread since
 553          * the check in rcu_read_lock().
 554          */
 555         if (URCU_TLS(urcu_bp_reader))
 556                 goto end;
 557
 558         /*
 559          * Take care of early registration before urcu_bp constructor.
 560          */
 561         _urcu_bp_init();
 562
 563         mutex_lock(&rcu_registry_lock);
 564         add_thread();
 565         mutex_unlock(&rcu_registry_lock);
 566 end:
 567         ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL);
 568         if (ret)
 569                 abort();
 570 }
 571 URCU_ATTR_ALIAS("urcu_bp_register") void rcu_bp_register();
 572
 573 void urcu_bp_register_thread(void)
 574 {
 575         if (caa_unlikely(!URCU_TLS(urcu_bp_reader)))
 576                 urcu_bp_register(); /* If not yet registered. */
 577 }
 578
 579 /* Disable signals, take mutex, remove from registry */
 580 static
 581 void urcu_bp_unregister(struct rcu_reader *rcu_reader_reg)
 582 {
 583         sigset_t newmask, oldmask;
 584         int ret;
 585
 586         ret = sigfillset(&newmask);
 587         if (ret)
 588                 abort();
 589         ret = pthread_sigmask(SIG_BLOCK, &newmask, &oldmask);
 590         if (ret)
 591                 abort();
 592
 593         mutex_lock(&rcu_registry_lock);
 594         remove_thread(rcu_reader_reg);
 595         mutex_unlock(&rcu_registry_lock);
 596         ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL);
 597         if (ret)
 598                 abort();
 599         urcu_bp_exit();
 600 }
 601
 602 /*
 603  * Remove thread from the registry when it exits, and flag it as
 604  * destroyed so garbage collection can take care of it.
 605  */
 606 static
 607 void urcu_bp_thread_exit_notifier(void *rcu_key)
 608 {
 609         urcu_bp_unregister(rcu_key);
 610 }
 611
 612 #ifdef CONFIG_RCU_FORCE_SYS_MEMBARRIER
 613 static
 614 void urcu_bp_sys_membarrier_status(bool available)
 615 {
 616         if (!available)
 617                 abort();
 618 }
 619 #else
 620 static
 621 void urcu_bp_sys_membarrier_status(bool available)
 622 {
 623         if (!available)
 624                 return;
 625         urcu_bp_has_sys_membarrier = 1;
 626 }
 627 #endif
 628
 629 static
 630 void urcu_bp_sys_membarrier_init(void)
 631 {
 632         bool available = false;
 633         int mask;
 634
 635         mask = membarrier(MEMBARRIER_CMD_QUERY, 0);
 636         if (mask >= 0) {
 637                 if (mask & MEMBARRIER_CMD_PRIVATE_EXPEDITED) {
 638                         if (membarrier(MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED, 0))
 639                                 urcu_die(errno);
 640                         available = true;
 641                 }
 642         }
 643         urcu_bp_sys_membarrier_status(available);
 644 }
 645
 646 static
 647 void _urcu_bp_init(void)
 648 {
 649         mutex_lock(&init_lock);
 650         if (!urcu_bp_refcount++) {
 651                 int ret;
 652
 653                 ret = pthread_key_create(&urcu_bp_key,
 654                                 urcu_bp_thread_exit_notifier);
 655                 if (ret)
 656                         abort();
 657                 urcu_bp_sys_membarrier_init();
 658                 initialized = 1;
 659         }
 660         mutex_unlock(&init_lock);
 661 }
 662
 663 static
 664 void urcu_bp_exit(void)
 665 {
 666         urcu_call_rcu_exit();
 667
 668         mutex_lock(&init_lock);
 669         if (!--urcu_bp_refcount) {
 670                 struct registry_chunk *chunk, *tmp;
 671                 int ret;
 672
 673                 cds_list_for_each_entry_safe(chunk, tmp,
 674                                 &registry_arena.chunk_list, node) {
 675                         munmap((void *) chunk, chunk->data_len
 676                                         + sizeof(struct registry_chunk));
 677                 }
 678                 CDS_INIT_LIST_HEAD(&registry_arena.chunk_list);
 679                 ret = pthread_key_delete(urcu_bp_key);
 680                 if (ret)
 681                         abort();
 682         }
 683         mutex_unlock(&init_lock);
 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         assert(!ret);
 699         ret = pthread_sigmask(SIG_BLOCK, &newmask, &oldmask);
 700         assert(!ret);
 701         mutex_lock(&rcu_gp_lock);
 702         mutex_lock(&rcu_registry_lock);
 703         saved_fork_signal_mask = oldmask;
 704 }
 705 URCU_ATTR_ALIAS("urcu_bp_before_fork") void rcu_bp_before_fork();
 706
 707 void urcu_bp_after_fork_parent(void)
 708 {
 709         sigset_t oldmask;
 710         int ret;
 711
 712         oldmask = saved_fork_signal_mask;
 713         mutex_unlock(&rcu_registry_lock);
 714         mutex_unlock(&rcu_gp_lock);
 715         ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL);
 716         assert(!ret);
 717 }
 718 URCU_ATTR_ALIAS("urcu_bp_after_fork_parent")
 719 void rcu_bp_after_fork_parent(void);
 720
 721 /*
 722  * Prune all entries from registry except our own thread. Fits the Linux
 723  * fork behavior. Called with rcu_gp_lock and rcu_registry_lock held.
 724  */
 725 static
 726 void urcu_bp_prune_registry(void)
 727 {
 728         struct registry_chunk *chunk;
 729         struct urcu_bp_reader *rcu_reader_reg;
 730
 731         cds_list_for_each_entry(chunk, &registry_arena.chunk_list, node) {
 732                 for (rcu_reader_reg = (struct urcu_bp_reader *) &chunk->data[0];
 733                                 rcu_reader_reg < (struct urcu_bp_reader *) &chunk->data[chunk->data_len];
 734                                 rcu_reader_reg++) {
 735                         if (!rcu_reader_reg->alloc)
 736                                 continue;
 737                         if (rcu_reader_reg->tid == pthread_self())
 738                                 continue;
 739                         cleanup_thread(chunk, rcu_reader_reg);
 740                 }
 741         }
 742 }
 743
 744 void urcu_bp_after_fork_child(void)
 745 {
 746         sigset_t oldmask;
 747         int ret;
 748
 749         urcu_bp_prune_registry();
 750         oldmask = saved_fork_signal_mask;
 751         mutex_unlock(&rcu_registry_lock);
 752         mutex_unlock(&rcu_gp_lock);
 753         ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL);
 754         assert(!ret);
 755 }
 756 URCU_ATTR_ALIAS("urcu_bp_after_fork_child")
 757 void rcu_bp_after_fork_child(void);
 758
 759 void *urcu_bp_dereference_sym(void *p)
 760 {
 761         return _rcu_dereference(p);
 762 }
 763 URCU_ATTR_ALIAS("urcu_bp_dereference_sym")
 764 void *rcu_dereference_sym_bp();
 765
 766 void *urcu_bp_set_pointer_sym(void **p, void *v)
 767 {
 768         cmm_wmb();
 769         uatomic_set(p, v);
 770         return v;
 771 }
 772 URCU_ATTR_ALIAS("urcu_bp_set_pointer_sym")
 773 void *rcu_set_pointer_sym_bp();
 774
 775 void *urcu_bp_xchg_pointer_sym(void **p, void *v)
 776 {
 777         cmm_wmb();
 778         return uatomic_xchg(p, v);
 779 }
 780 URCU_ATTR_ALIAS("urcu_bp_xchg_pointer_sym")
 781 void *rcu_xchg_pointer_sym_bp();
 782
 783 void *urcu_bp_cmpxchg_pointer_sym(void **p, void *old, void *_new)
 784 {
 785         cmm_wmb();
 786         return uatomic_cmpxchg(p, old, _new);
 787 }
 788 URCU_ATTR_ALIAS("urcu_bp_cmpxchg_pointer_sym")
 789 void *rcu_cmpxchg_pointer_sym_bp();
 790
 791 DEFINE_RCU_FLAVOR(rcu_flavor);
 792 DEFINE_RCU_FLAVOR_ALIAS(rcu_flavor, alias_rcu_flavor);
 793
 794 #include "urcu-call-rcu-impl.h"
 795 #include "urcu-defer-impl.h"