src/urcu-bp.c

   1 // SPDX-FileCopyrightText: 2009 Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
   2 // SPDX-FileCopyrightText: 2009 Paul E. McKenney, IBM Corporation.
   3 //
   4 // SPDX-License-Identifier: LGPL-2.1-or-later
   5
   6 /*
   7  * Userspace RCU library, "bulletproof" version.
   8  *
   9  * IBM's contributions to this file may be relicensed under LGPLv2 or later.
  10  */
  11
  12 #define URCU_NO_COMPAT_IDENTIFIERS
  13 #define _LGPL_SOURCE
  14 #include <stdio.h>
  15 #include <pthread.h>
  16 #include <signal.h>
  17 #include <stdlib.h>
  18 #include <string.h>
  19 #include <errno.h>
  20 #include <poll.h>
  21 #include <unistd.h>
  22 #include <stdbool.h>
  23 #include <sys/mman.h>
  24
  25 #include <urcu/assert.h>
  26 #include <urcu/config.h>
  27 #include <urcu/arch.h>
  28 #include <urcu/wfcqueue.h>
  29 #include <urcu/map/urcu-bp.h>
  30 #include <urcu/static/urcu-bp.h>
  31 #include <urcu/pointer.h>
  32 #include <urcu/tls-compat.h>
  33
  34 #include "urcu-die.h"
  35 #include "urcu-utils.h"
  36
  37 #define URCU_API_MAP
  38 /* Do not #define _LGPL_SOURCE to ensure we can emit the wrapper symbols */
  39 #undef _LGPL_SOURCE
  40 #include <urcu/urcu-bp.h>
  41 #define _LGPL_SOURCE
  42
  43 #ifndef MAP_ANONYMOUS
  44 #define MAP_ANONYMOUS MAP_ANON
  45 #endif
  46
  47 #ifdef __linux__
  48 static
  49 void *mremap_wrapper(void *old_address, size_t old_size,
  50                 size_t new_size, int flags)
  51 {
  52         return mremap(old_address, old_size, new_size, flags);
  53 }
  54 #else
  55
  56 #define MREMAP_MAYMOVE  1
  57 #define MREMAP_FIXED    2
  58
  59 /*
  60  * mremap wrapper for non-Linux systems not allowing MAYMOVE.
  61  * This is not generic.
  62 */
  63 static
  64 void *mremap_wrapper(void *old_address __attribute__((unused)),
  65                 size_t old_size __attribute__((unused)),
  66                 size_t new_size __attribute__((unused)),
  67                 int flags)
  68 {
  69         urcu_posix_assert(!(flags & MREMAP_MAYMOVE));
  70
  71         return MAP_FAILED;
  72 }
  73 #endif
  74
  75 /* Sleep delay in ms */
  76 #define RCU_SLEEP_DELAY_MS      10
  77 #define INIT_NR_THREADS         8
  78 #define ARENA_INIT_ALLOC                \
  79         sizeof(struct registry_chunk)   \
  80         + INIT_NR_THREADS * sizeof(struct urcu_bp_reader)
  81
  82 /*
  83  * Active attempts to check for reader Q.S. before calling sleep().
  84  */
  85 #define RCU_QS_ACTIVE_ATTEMPTS 100
  86
  87 static
  88 int urcu_bp_refcount;
  89
  90 /* If the headers do not support membarrier system call, fall back smp_mb. */
  91 #ifdef __NR_membarrier
  92 # define membarrier(...)                syscall(__NR_membarrier, __VA_ARGS__)
  93 #else
  94 # define membarrier(...)                -ENOSYS
  95 #endif
  96
  97 enum membarrier_cmd {
  98         MEMBARRIER_CMD_QUERY                            = 0,
  99         MEMBARRIER_CMD_SHARED                           = (1 << 0),
 100         /* reserved for MEMBARRIER_CMD_SHARED_EXPEDITED (1 << 1) */
 101         /* reserved for MEMBARRIER_CMD_PRIVATE (1 << 2) */
 102         MEMBARRIER_CMD_PRIVATE_EXPEDITED                = (1 << 3),
 103         MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED       = (1 << 4),
 104 };
 105
 106 static
 107 void __attribute__((constructor)) _urcu_bp_init(void);
 108 static
 109 void urcu_bp_exit(void);
 110 static
 111 void __attribute__((destructor)) urcu_bp_exit_destructor(void);
 112 static void urcu_call_rcu_exit(void);
 113
 114 #ifndef CONFIG_RCU_FORCE_SYS_MEMBARRIER
 115 int urcu_bp_has_sys_membarrier;
 116 #endif
 117
 118 /*
 119  * rcu_gp_lock ensures mutual exclusion between threads calling
 120  * synchronize_rcu().
 121  */
 122 static pthread_mutex_t rcu_gp_lock = PTHREAD_MUTEX_INITIALIZER;
 123 /*
 124  * rcu_registry_lock ensures mutual exclusion between threads
 125  * registering and unregistering themselves to/from the registry, and
 126  * with threads reading that registry from synchronize_rcu(). However,
 127  * this lock is not held all the way through the completion of awaiting
 128  * for the grace period. It is sporadically released between iterations
 129  * on the registry.
 130  * rcu_registry_lock may nest inside rcu_gp_lock.
 131  */
 132 static pthread_mutex_t rcu_registry_lock = PTHREAD_MUTEX_INITIALIZER;
 133
 134 static pthread_mutex_t init_lock = PTHREAD_MUTEX_INITIALIZER;
 135 static int initialized;
 136
 137 static pthread_key_t urcu_bp_key;
 138
 139 struct urcu_bp_gp urcu_bp_gp = { .ctr = URCU_BP_GP_COUNT };
 140
 141 /*
 142  * Pointer to registry elements. Written to only by each individual reader. Read
 143  * by both the reader and the writers.
 144  */
 145 DEFINE_URCU_TLS(struct urcu_bp_reader *, urcu_bp_reader);
 146
 147 static CDS_LIST_HEAD(registry);
 148
 149 struct registry_chunk {
 150         size_t data_len;                /* data length */
 151         size_t used;                    /* amount of data used */
 152         struct cds_list_head node;      /* chunk_list node */
 153         char data[];
 154 };
 155
 156 struct registry_arena {
 157         struct cds_list_head chunk_list;
 158 };
 159
 160 static struct registry_arena registry_arena = {
 161         .chunk_list = CDS_LIST_HEAD_INIT(registry_arena.chunk_list),
 162 };
 163
 164 /* Saved fork signal mask, protected by rcu_gp_lock */
 165 static sigset_t saved_fork_signal_mask;
 166
 167 static void mutex_lock(pthread_mutex_t *mutex)
 168 {
 169         int ret;
 170
 171 #ifndef DISTRUST_SIGNALS_EXTREME
 172         ret = pthread_mutex_lock(mutex);
 173         if (ret)
 174                 urcu_die(ret);
 175 #else /* #ifndef DISTRUST_SIGNALS_EXTREME */
 176         while ((ret = pthread_mutex_trylock(mutex)) != 0) {
 177                 if (ret != EBUSY && ret != EINTR)
 178                         urcu_die(ret);
 179                 poll(NULL,0,10);
 180         }
 181 #endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */
 182 }
 183
 184 static void mutex_unlock(pthread_mutex_t *mutex)
 185 {
 186         int ret;
 187
 188         ret = pthread_mutex_unlock(mutex);
 189         if (ret)
 190                 urcu_die(ret);
 191 }
 192
 193 static void smp_mb_master(void)
 194 {
 195         if (caa_likely(urcu_bp_has_sys_membarrier)) {
 196                 if (membarrier(MEMBARRIER_CMD_PRIVATE_EXPEDITED, 0))
 197                         urcu_die(errno);
 198         } else {
 199                 cmm_smp_mb();
 200         }
 201 }
 202
 203 /*
 204  * Always called with rcu_registry lock held. Releases this lock between
 205  * iterations and grabs it again. Holds the lock when it returns.
 206  */
 207 static void wait_for_readers(struct cds_list_head *input_readers,
 208                         struct cds_list_head *cur_snap_readers,
 209                         struct cds_list_head *qsreaders)
 210 {
 211         unsigned int wait_loops = 0;
 212         struct urcu_bp_reader *index, *tmp;
 213
 214         /*
 215          * Wait for each thread URCU_TLS(urcu_bp_reader).ctr to either
 216          * indicate quiescence (not nested), or observe the current
 217          * rcu_gp.ctr value.
 218          */
 219         for (;;) {
 220                 if (wait_loops < RCU_QS_ACTIVE_ATTEMPTS)
 221                         wait_loops++;
 222
 223                 cds_list_for_each_entry_safe(index, tmp, input_readers, node) {
 224                         switch (urcu_bp_reader_state(&index->ctr)) {
 225                         case URCU_BP_READER_ACTIVE_CURRENT:
 226                                 if (cur_snap_readers) {
 227                                         cds_list_move(&index->node,
 228                                                 cur_snap_readers);
 229                                         break;
 230                                 }
 231                                 /* Fall-through */
 232                         case URCU_BP_READER_INACTIVE:
 233                                 cds_list_move(&index->node, qsreaders);
 234                                 break;
 235                         case URCU_BP_READER_ACTIVE_OLD:
 236                                 /*
 237                                  * Old snapshot. Leaving node in
 238                                  * input_readers will make us busy-loop
 239                                  * until the snapshot becomes current or
 240                                  * the reader becomes inactive.
 241                                  */
 242                                 break;
 243                         }
 244                 }
 245
 246                 if (cds_list_empty(input_readers)) {
 247                         break;
 248                 } else {
 249                         /* Temporarily unlock the registry lock. */
 250                         mutex_unlock(&rcu_registry_lock);
 251                         if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS)
 252                                 (void) poll(NULL, 0, RCU_SLEEP_DELAY_MS);
 253                         else
 254                                 caa_cpu_relax();
 255                         /* Re-lock the registry lock before the next loop. */
 256                         mutex_lock(&rcu_registry_lock);
 257                 }
 258         }
 259 }
 260
 261 void urcu_bp_synchronize_rcu(void)
 262 {
 263         CDS_LIST_HEAD(cur_snap_readers);
 264         CDS_LIST_HEAD(qsreaders);
 265         sigset_t newmask, oldmask;
 266         int ret;
 267
 268         ret = sigfillset(&newmask);
 269         urcu_posix_assert(!ret);
 270         ret = pthread_sigmask(SIG_BLOCK, &newmask, &oldmask);
 271         urcu_posix_assert(!ret);
 272
 273         mutex_lock(&rcu_gp_lock);
 274
 275         mutex_lock(&rcu_registry_lock);
 276
 277         if (cds_list_empty(&registry))
 278                 goto out;
 279
 280         /* All threads should read qparity before accessing data structure
 281          * where new ptr points to. */
 282         /* Write new ptr before changing the qparity */
 283         smp_mb_master();
 284
 285         /*
 286          * Wait for readers to observe original parity or be quiescent.
 287          * wait_for_readers() can release and grab again rcu_registry_lock
 288          * internally.
 289          */
 290         wait_for_readers(&registry, &cur_snap_readers, &qsreaders);
 291
 292         /*
 293          * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
 294          * model easier to understand. It does not have a big performance impact
 295          * anyway, given this is the write-side.
 296          */
 297         cmm_smp_mb();
 298
 299         /* Switch parity: 0 -> 1, 1 -> 0 */
 300         CMM_STORE_SHARED(rcu_gp.ctr, rcu_gp.ctr ^ URCU_BP_GP_CTR_PHASE);
 301
 302         /*
 303          * Must commit qparity update to memory before waiting for other parity
 304          * quiescent state. Failure to do so could result in the writer waiting
 305          * forever while new readers are always accessing data (no progress).
 306          * Ensured by CMM_STORE_SHARED and CMM_LOAD_SHARED.
 307          */
 308
 309         /*
 310          * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
 311          * model easier to understand. It does not have a big performance impact
 312          * anyway, given this is the write-side.
 313          */
 314         cmm_smp_mb();
 315
 316         /*
 317          * Wait for readers to observe new parity or be quiescent.
 318          * wait_for_readers() can release and grab again rcu_registry_lock
 319          * internally.
 320          */
 321         wait_for_readers(&cur_snap_readers, NULL, &qsreaders);
 322
 323         /*
 324          * Put quiescent reader list back into registry.
 325          */
 326         cds_list_splice(&qsreaders, &registry);
 327
 328         /*
 329          * Finish waiting for reader threads before letting the old ptr being
 330          * freed.
 331          */
 332         smp_mb_master();
 333 out:
 334         mutex_unlock(&rcu_registry_lock);
 335         mutex_unlock(&rcu_gp_lock);
 336         ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL);
 337         urcu_posix_assert(!ret);
 338 }
 339
 340 /*
 341  * library wrappers to be used by non-LGPL compatible source code.
 342  */
 343
 344 void urcu_bp_read_lock(void)
 345 {
 346         _urcu_bp_read_lock();
 347 }
 348
 349 void urcu_bp_read_unlock(void)
 350 {
 351         _urcu_bp_read_unlock();
 352 }
 353
 354 int urcu_bp_read_ongoing(void)
 355 {
 356         return _urcu_bp_read_ongoing();
 357 }
 358
 359 /*
 360  * Only grow for now. If empty, allocate a ARENA_INIT_ALLOC sized chunk.
 361  * Else, try expanding the last chunk. If this fails, allocate a new
 362  * chunk twice as big as the last chunk.
 363  * Memory used by chunks _never_ moves. A chunk could theoretically be
 364  * freed when all "used" slots are released, but we don't do it at this
 365  * point.
 366  */
 367 static
 368 void expand_arena(struct registry_arena *arena)
 369 {
 370         struct registry_chunk *new_chunk, *last_chunk;
 371         size_t old_chunk_len, new_chunk_len;
 372
 373         /* No chunk. */
 374         if (cds_list_empty(&arena->chunk_list)) {
 375                 urcu_posix_assert(ARENA_INIT_ALLOC >=
 376                         sizeof(struct registry_chunk)
 377                         + sizeof(struct rcu_reader));
 378                 new_chunk_len = ARENA_INIT_ALLOC;
 379                 new_chunk = (struct registry_chunk *) mmap(NULL,
 380                         new_chunk_len,
 381                         PROT_READ | PROT_WRITE,
 382                         MAP_ANONYMOUS | MAP_PRIVATE,
 383                         -1, 0);
 384                 if (new_chunk == MAP_FAILED)
 385                         abort();
 386                 memset(new_chunk, 0, new_chunk_len);
 387                 new_chunk->data_len =
 388                         new_chunk_len - sizeof(struct registry_chunk);
 389                 cds_list_add_tail(&new_chunk->node, &arena->chunk_list);
 390                 return;         /* We're done. */
 391         }
 392
 393         /* Try expanding last chunk. */
 394         last_chunk = cds_list_entry(arena->chunk_list.prev,
 395                 struct registry_chunk, node);
 396         old_chunk_len =
 397                 last_chunk->data_len + sizeof(struct registry_chunk);
 398         new_chunk_len = old_chunk_len << 1;
 399
 400         /* Don't allow memory mapping to move, just expand. */
 401         new_chunk = mremap_wrapper(last_chunk, old_chunk_len,
 402                 new_chunk_len, 0);
 403         if (new_chunk != MAP_FAILED) {
 404                 /* Should not have moved. */
 405                 urcu_posix_assert(new_chunk == last_chunk);
 406                 memset((char *) last_chunk + old_chunk_len, 0,
 407                         new_chunk_len - old_chunk_len);
 408                 last_chunk->data_len =
 409                         new_chunk_len - sizeof(struct registry_chunk);
 410                 return;         /* We're done. */
 411         }
 412
 413         /* Remap did not succeed, we need to add a new chunk. */
 414         new_chunk = (struct registry_chunk *) mmap(NULL,
 415                 new_chunk_len,
 416                 PROT_READ | PROT_WRITE,
 417                 MAP_ANONYMOUS | MAP_PRIVATE,
 418                 -1, 0);
 419         if (new_chunk == MAP_FAILED)
 420                 abort();
 421         memset(new_chunk, 0, new_chunk_len);
 422         new_chunk->data_len =
 423                 new_chunk_len - sizeof(struct registry_chunk);
 424         cds_list_add_tail(&new_chunk->node, &arena->chunk_list);
 425 }
 426
 427 static
 428 struct rcu_reader *arena_alloc(struct registry_arena *arena)
 429 {
 430         struct registry_chunk *chunk;
 431         struct rcu_reader *rcu_reader_reg;
 432         int expand_done = 0;    /* Only allow to expand once per alloc */
 433         size_t len = sizeof(struct rcu_reader);
 434
 435 retry:
 436         cds_list_for_each_entry(chunk, &arena->chunk_list, node) {
 437                 if (chunk->data_len - chunk->used < len)
 438                         continue;
 439                 /* Find spot */
 440                 for (rcu_reader_reg = (struct rcu_reader *) &chunk->data[0];
 441                                 rcu_reader_reg < (struct rcu_reader *) &chunk->data[chunk->data_len];
 442                                 rcu_reader_reg++) {
 443                         if (!rcu_reader_reg->alloc) {
 444                                 rcu_reader_reg->alloc = 1;
 445                                 chunk->used += len;
 446                                 return rcu_reader_reg;
 447                         }
 448                 }
 449         }
 450
 451         if (!expand_done) {
 452                 expand_arena(arena);
 453                 expand_done = 1;
 454                 goto retry;
 455         }
 456
 457         return NULL;
 458 }
 459
 460 /* Called with signals off and mutex locked */
 461 static
 462 void add_thread(void)
 463 {
 464         struct rcu_reader *rcu_reader_reg;
 465         int ret;
 466
 467         rcu_reader_reg = arena_alloc(&registry_arena);
 468         if (!rcu_reader_reg)
 469                 abort();
 470         ret = pthread_setspecific(urcu_bp_key, rcu_reader_reg);
 471         if (ret)
 472                 abort();
 473
 474         /* Add to registry */
 475         rcu_reader_reg->tid = pthread_self();
 476         urcu_posix_assert(rcu_reader_reg->ctr == 0);
 477         cds_list_add(&rcu_reader_reg->node, &registry);
 478         /*
 479          * Reader threads are pointing to the reader registry. This is
 480          * why its memory should never be relocated.
 481          */
 482         URCU_TLS(urcu_bp_reader) = rcu_reader_reg;
 483 }
 484
 485 /* Called with mutex locked */
 486 static
 487 void cleanup_thread(struct registry_chunk *chunk,
 488                 struct rcu_reader *rcu_reader_reg)
 489 {
 490         rcu_reader_reg->ctr = 0;
 491         cds_list_del(&rcu_reader_reg->node);
 492         rcu_reader_reg->tid = 0;
 493         rcu_reader_reg->alloc = 0;
 494         chunk->used -= sizeof(struct rcu_reader);
 495 }
 496
 497 static
 498 struct registry_chunk *find_chunk(struct rcu_reader *rcu_reader_reg)
 499 {
 500         struct registry_chunk *chunk;
 501
 502         cds_list_for_each_entry(chunk, &registry_arena.chunk_list, node) {
 503                 if (rcu_reader_reg < (struct rcu_reader *) &chunk->data[0])
 504                         continue;
 505                 if (rcu_reader_reg >= (struct rcu_reader *) &chunk->data[chunk->data_len])
 506                         continue;
 507                 return chunk;
 508         }
 509         return NULL;
 510 }
 511
 512 /* Called with signals off and mutex locked */
 513 static
 514 void remove_thread(struct rcu_reader *rcu_reader_reg)
 515 {
 516         cleanup_thread(find_chunk(rcu_reader_reg), rcu_reader_reg);
 517         URCU_TLS(urcu_bp_reader) = NULL;
 518 }
 519
 520 /* Disable signals, take mutex, add to registry */
 521 void urcu_bp_register(void)
 522 {
 523         sigset_t newmask, oldmask;
 524         int ret;
 525
 526         ret = sigfillset(&newmask);
 527         if (ret)
 528                 abort();
 529         ret = pthread_sigmask(SIG_BLOCK, &newmask, &oldmask);
 530         if (ret)
 531                 abort();
 532
 533         /*
 534          * Check if a signal concurrently registered our thread since
 535          * the check in rcu_read_lock().
 536          */
 537         if (URCU_TLS(urcu_bp_reader))
 538                 goto end;
 539
 540         /*
 541          * Take care of early registration before urcu_bp constructor.
 542          */
 543         _urcu_bp_init();
 544
 545         mutex_lock(&rcu_registry_lock);
 546         add_thread();
 547         mutex_unlock(&rcu_registry_lock);
 548 end:
 549         ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL);
 550         if (ret)
 551                 abort();
 552 }
 553
 554 void urcu_bp_register_thread(void)
 555 {
 556         if (caa_unlikely(!URCU_TLS(urcu_bp_reader)))
 557                 urcu_bp_register(); /* If not yet registered. */
 558 }
 559
 560 /* Disable signals, take mutex, remove from registry */
 561 static
 562 void urcu_bp_unregister(struct rcu_reader *rcu_reader_reg)
 563 {
 564         sigset_t newmask, oldmask;
 565         int ret;
 566
 567         ret = sigfillset(&newmask);
 568         if (ret)
 569                 abort();
 570         ret = pthread_sigmask(SIG_BLOCK, &newmask, &oldmask);
 571         if (ret)
 572                 abort();
 573
 574         mutex_lock(&rcu_registry_lock);
 575         remove_thread(rcu_reader_reg);
 576         mutex_unlock(&rcu_registry_lock);
 577         ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL);
 578         if (ret)
 579                 abort();
 580         urcu_bp_exit();
 581 }
 582
 583 /*
 584  * Remove thread from the registry when it exits, and flag it as
 585  * destroyed so garbage collection can take care of it.
 586  */
 587 static
 588 void urcu_bp_thread_exit_notifier(void *rcu_key)
 589 {
 590         urcu_bp_unregister(rcu_key);
 591 }
 592
 593 #ifdef CONFIG_RCU_FORCE_SYS_MEMBARRIER
 594 static
 595 void urcu_bp_sys_membarrier_status(bool available)
 596 {
 597         if (!available)
 598                 abort();
 599 }
 600 #else
 601 static
 602 void urcu_bp_sys_membarrier_status(bool available)
 603 {
 604         if (!available)
 605                 return;
 606         urcu_bp_has_sys_membarrier = 1;
 607 }
 608 #endif
 609
 610 static
 611 void urcu_bp_sys_membarrier_init(void)
 612 {
 613         bool available = false;
 614         int mask;
 615
 616         mask = membarrier(MEMBARRIER_CMD_QUERY, 0);
 617         if (mask >= 0) {
 618                 if (mask & MEMBARRIER_CMD_PRIVATE_EXPEDITED) {
 619                         if (membarrier(MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED, 0))
 620                                 urcu_die(errno);
 621                         available = true;
 622                 }
 623         }
 624         urcu_bp_sys_membarrier_status(available);
 625 }
 626
 627 static
 628 void _urcu_bp_init(void)
 629 {
 630         mutex_lock(&init_lock);
 631         if (!urcu_bp_refcount++) {
 632                 int ret;
 633
 634                 ret = pthread_key_create(&urcu_bp_key,
 635                                 urcu_bp_thread_exit_notifier);
 636                 if (ret)
 637                         abort();
 638                 urcu_bp_sys_membarrier_init();
 639                 initialized = 1;
 640         }
 641         mutex_unlock(&init_lock);
 642 }
 643
 644 static
 645 void urcu_bp_exit(void)
 646 {
 647         mutex_lock(&init_lock);
 648         if (!--urcu_bp_refcount) {
 649                 struct registry_chunk *chunk, *tmp;
 650                 int ret;
 651
 652                 cds_list_for_each_entry_safe(chunk, tmp,
 653                                 &registry_arena.chunk_list, node) {
 654                         munmap((void *) chunk, chunk->data_len
 655                                         + sizeof(struct registry_chunk));
 656                 }
 657                 CDS_INIT_LIST_HEAD(&registry_arena.chunk_list);
 658                 ret = pthread_key_delete(urcu_bp_key);
 659                 if (ret)
 660                         abort();
 661         }
 662         mutex_unlock(&init_lock);
 663 }
 664
 665 static
 666 void urcu_bp_exit_destructor(void)
 667 {
 668         urcu_call_rcu_exit();
 669         urcu_bp_exit();
 670 }
 671
 672 /*
 673  * Holding the rcu_gp_lock and rcu_registry_lock across fork will make
 674  * sure we fork() don't race with a concurrent thread executing with
 675  * any of those locks held. This ensures that the registry and data
 676  * protected by rcu_gp_lock are in a coherent state in the child.
 677  */
 678 void urcu_bp_before_fork(void)
 679 {
 680         sigset_t newmask, oldmask;
 681         int ret;
 682
 683         ret = sigfillset(&newmask);
 684         urcu_posix_assert(!ret);
 685         ret = pthread_sigmask(SIG_BLOCK, &newmask, &oldmask);
 686         urcu_posix_assert(!ret);
 687         mutex_lock(&rcu_gp_lock);
 688         mutex_lock(&rcu_registry_lock);
 689         saved_fork_signal_mask = oldmask;
 690 }
 691
 692 void urcu_bp_after_fork_parent(void)
 693 {
 694         sigset_t oldmask;
 695         int ret;
 696
 697         oldmask = saved_fork_signal_mask;
 698         mutex_unlock(&rcu_registry_lock);
 699         mutex_unlock(&rcu_gp_lock);
 700         ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL);
 701         urcu_posix_assert(!ret);
 702 }
 703
 704 /*
 705  * Prune all entries from registry except our own thread. Fits the Linux
 706  * fork behavior. Called with rcu_gp_lock and rcu_registry_lock held.
 707  */
 708 static
 709 void urcu_bp_prune_registry(void)
 710 {
 711         struct registry_chunk *chunk;
 712         struct urcu_bp_reader *rcu_reader_reg;
 713
 714         cds_list_for_each_entry(chunk, &registry_arena.chunk_list, node) {
 715                 for (rcu_reader_reg = (struct urcu_bp_reader *) &chunk->data[0];
 716                                 rcu_reader_reg < (struct urcu_bp_reader *) &chunk->data[chunk->data_len];
 717                                 rcu_reader_reg++) {
 718                         if (!rcu_reader_reg->alloc)
 719                                 continue;
 720                         if (rcu_reader_reg->tid == pthread_self())
 721                                 continue;
 722                         cleanup_thread(chunk, rcu_reader_reg);
 723                 }
 724         }
 725 }
 726
 727 void urcu_bp_after_fork_child(void)
 728 {
 729         sigset_t oldmask;
 730         int ret;
 731
 732         urcu_bp_prune_registry();
 733         oldmask = saved_fork_signal_mask;
 734         mutex_unlock(&rcu_registry_lock);
 735         mutex_unlock(&rcu_gp_lock);
 736         ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL);
 737         urcu_posix_assert(!ret);
 738 }
 739
 740 void *urcu_bp_dereference_sym(void *p)
 741 {
 742         return _rcu_dereference(p);
 743 }
 744
 745 void *urcu_bp_set_pointer_sym(void **p, void *v)
 746 {
 747         cmm_wmb();
 748         uatomic_set(p, v);
 749         return v;
 750 }
 751
 752 void *urcu_bp_xchg_pointer_sym(void **p, void *v)
 753 {
 754         cmm_wmb();
 755         return uatomic_xchg(p, v);
 756 }
 757
 758 void *urcu_bp_cmpxchg_pointer_sym(void **p, void *old, void *_new)
 759 {
 760         cmm_wmb();
 761         return uatomic_cmpxchg(p, old, _new);
 762 }
 763
 764 DEFINE_RCU_FLAVOR(rcu_flavor);
 765
 766 #include "urcu-call-rcu-impl.h"
 767 #include "urcu-defer-impl.h"
 768 #include "urcu-poll-impl.h"