src/urcu.c

   1 /*
   2  * urcu.c
   3  *
   4  * Userspace RCU library
   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 _BSD_SOURCE
  27 #define _LGPL_SOURCE
  28 #define _DEFAULT_SOURCE
  29 #include <stdio.h>
  30 #include <pthread.h>
  31 #include <signal.h>
  32 #include <assert.h>
  33 #include <stdlib.h>
  34 #include <stdint.h>
  35 #include <string.h>
  36 #include <errno.h>
  37 #include <stdbool.h>
  38 #include <poll.h>
  39
  40 #include "urcu/arch.h"
  41 #include "urcu/wfcqueue.h"
  42 #include "urcu/map/urcu.h"
  43 #include "urcu/static/urcu.h"
  44 #include "urcu-pointer.h"
  45 #include "urcu/tls-compat.h"
  46
  47 #include "urcu-die.h"
  48 #include "urcu-wait.h"
  49
  50 /* Do not #define _LGPL_SOURCE to ensure we can emit the wrapper symbols */
  51 #undef _LGPL_SOURCE
  52 #include "urcu.h"
  53 #define _LGPL_SOURCE
  54
  55 /*
  56  * If a reader is really non-cooperative and refuses to commit its
  57  * rcu_active_readers count to memory (there is no barrier in the reader
  58  * per-se), kick it after 10 loops waiting for it.
  59  */
  60 #define KICK_READER_LOOPS       10
  61
  62 /*
  63  * Active attempts to check for reader Q.S. before calling futex().
  64  */
  65 #define RCU_QS_ACTIVE_ATTEMPTS 100
  66
  67 /* If the headers do not support membarrier system call, fall back on RCU_MB */
  68 #ifdef __NR_membarrier
  69 # define membarrier(...)                syscall(__NR_membarrier, __VA_ARGS__)
  70 #else
  71 # define membarrier(...)                -ENOSYS
  72 #endif
  73
  74 enum membarrier_cmd {
  75         MEMBARRIER_CMD_QUERY                            = 0,
  76         MEMBARRIER_CMD_SHARED                           = (1 << 0),
  77         /* reserved for MEMBARRIER_CMD_SHARED_EXPEDITED (1 << 1) */
  78         /* reserved for MEMBARRIER_CMD_PRIVATE (1 << 2) */
  79         MEMBARRIER_CMD_PRIVATE_EXPEDITED                = (1 << 3),
  80         MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED       = (1 << 4),
  81 };
  82
  83 #ifdef RCU_MEMBARRIER
  84 static int init_done;
  85 static int has_sys_membarrier_private_expedited;
  86
  87 #ifndef CONFIG_RCU_FORCE_SYS_MEMBARRIER
  88 int rcu_has_sys_membarrier_memb;
  89 #endif
  90
  91 void __attribute__((constructor)) rcu_init(void);
  92 #endif
  93
  94 #ifdef RCU_MB
  95 void rcu_init(void)
  96 {
  97 }
  98 #endif
  99
 100 #ifdef RCU_SIGNAL
 101 static int init_done;
 102
 103 void __attribute__((constructor)) rcu_init(void);
 104 void __attribute__((destructor)) rcu_exit(void);
 105 #endif
 106
 107 /*
 108  * rcu_gp_lock ensures mutual exclusion between threads calling
 109  * synchronize_rcu().
 110  */
 111 static pthread_mutex_t rcu_gp_lock = PTHREAD_MUTEX_INITIALIZER;
 112 /*
 113  * rcu_registry_lock ensures mutual exclusion between threads
 114  * registering and unregistering themselves to/from the registry, and
 115  * with threads reading that registry from synchronize_rcu(). However,
 116  * this lock is not held all the way through the completion of awaiting
 117  * for the grace period. It is sporadically released between iterations
 118  * on the registry.
 119  * rcu_registry_lock may nest inside rcu_gp_lock.
 120  */
 121 static pthread_mutex_t rcu_registry_lock = PTHREAD_MUTEX_INITIALIZER;
 122 struct rcu_gp rcu_gp = { .ctr = RCU_GP_COUNT };
 123
 124 /*
 125  * Written to only by each individual reader. Read by both the reader and the
 126  * writers.
 127  */
 128 DEFINE_URCU_TLS(struct rcu_reader, rcu_reader);
 129
 130 static CDS_LIST_HEAD(registry);
 131
 132 /*
 133  * Queue keeping threads awaiting to wait for a grace period. Contains
 134  * struct gp_waiters_thread objects.
 135  */
 136 static DEFINE_URCU_WAIT_QUEUE(gp_waiters);
 137
 138 static void mutex_lock(pthread_mutex_t *mutex)
 139 {
 140         int ret;
 141
 142 #ifndef DISTRUST_SIGNALS_EXTREME
 143         ret = pthread_mutex_lock(mutex);
 144         if (ret)
 145                 urcu_die(ret);
 146 #else /* #ifndef DISTRUST_SIGNALS_EXTREME */
 147         while ((ret = pthread_mutex_trylock(mutex)) != 0) {
 148                 if (ret != EBUSY && ret != EINTR)
 149                         urcu_die(ret);
 150                 if (CMM_LOAD_SHARED(URCU_TLS(rcu_reader).need_mb)) {
 151                         cmm_smp_mb();
 152                         _CMM_STORE_SHARED(URCU_TLS(rcu_reader).need_mb, 0);
 153                         cmm_smp_mb();
 154                 }
 155                 (void) poll(NULL, 0, 10);
 156         }
 157 #endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */
 158 }
 159
 160 static void mutex_unlock(pthread_mutex_t *mutex)
 161 {
 162         int ret;
 163
 164         ret = pthread_mutex_unlock(mutex);
 165         if (ret)
 166                 urcu_die(ret);
 167 }
 168
 169 #ifdef RCU_MEMBARRIER
 170 static void smp_mb_master(void)
 171 {
 172         if (caa_likely(rcu_has_sys_membarrier_memb)) {
 173                 if (membarrier(has_sys_membarrier_private_expedited ?
 174                                 MEMBARRIER_CMD_PRIVATE_EXPEDITED :
 175                                 MEMBARRIER_CMD_SHARED, 0))
 176                         urcu_die(errno);
 177         } else {
 178                 cmm_smp_mb();
 179         }
 180 }
 181 #endif
 182
 183 #ifdef RCU_MB
 184 static void smp_mb_master(void)
 185 {
 186         cmm_smp_mb();
 187 }
 188 #endif
 189
 190 #ifdef RCU_SIGNAL
 191 static void force_mb_all_readers(void)
 192 {
 193         struct rcu_reader *index;
 194
 195         /*
 196          * Ask for each threads to execute a cmm_smp_mb() so we can consider the
 197          * compiler barriers around rcu read lock as real memory barriers.
 198          */
 199         if (cds_list_empty(&registry))
 200                 return;
 201         /*
 202          * pthread_kill has a cmm_smp_mb(). But beware, we assume it performs
 203          * a cache flush on architectures with non-coherent cache. Let's play
 204          * safe and don't assume anything : we use cmm_smp_mc() to make sure the
 205          * cache flush is enforced.
 206          */
 207         cds_list_for_each_entry(index, &registry, node) {
 208                 CMM_STORE_SHARED(index->need_mb, 1);
 209                 pthread_kill(index->tid, SIGRCU);
 210         }
 211         /*
 212          * Wait for sighandler (and thus mb()) to execute on every thread.
 213          *
 214          * Note that the pthread_kill() will never be executed on systems
 215          * that correctly deliver signals in a timely manner.  However, it
 216          * is not uncommon for kernels to have bugs that can result in
 217          * lost or unduly delayed signals.
 218          *
 219          * If you are seeing the below pthread_kill() executing much at
 220          * all, we suggest testing the underlying kernel and filing the
 221          * relevant bug report.  For Linux kernels, we recommend getting
 222          * the Linux Test Project (LTP).
 223          */
 224         cds_list_for_each_entry(index, &registry, node) {
 225                 while (CMM_LOAD_SHARED(index->need_mb)) {
 226                         pthread_kill(index->tid, SIGRCU);
 227                         (void) poll(NULL, 0, 1);
 228                 }
 229         }
 230         cmm_smp_mb();   /* read ->need_mb before ending the barrier */
 231 }
 232
 233 static void smp_mb_master(void)
 234 {
 235         force_mb_all_readers();
 236 }
 237 #endif /* #ifdef RCU_SIGNAL */
 238
 239 /*
 240  * synchronize_rcu() waiting. Single thread.
 241  * Always called with rcu_registry lock held. Releases this lock and
 242  * grabs it again. Holds the lock when it returns.
 243  */
 244 static void wait_gp(void)
 245 {
 246         /*
 247          * Read reader_gp before read futex. smp_mb_master() needs to
 248          * be called with the rcu registry lock held in RCU_SIGNAL
 249          * flavor.
 250          */
 251         smp_mb_master();
 252         /* Temporarily unlock the registry lock. */
 253         mutex_unlock(&rcu_registry_lock);
 254         if (uatomic_read(&rcu_gp.futex) != -1)
 255                 goto end;
 256         while (futex_async(&rcu_gp.futex, FUTEX_WAIT, -1,
 257                         NULL, NULL, 0)) {
 258                 switch (errno) {
 259                 case EWOULDBLOCK:
 260                         /* Value already changed. */
 261                         goto end;
 262                 case EINTR:
 263                         /* Retry if interrupted by signal. */
 264                         break;  /* Get out of switch. */
 265                 default:
 266                         /* Unexpected error. */
 267                         urcu_die(errno);
 268                 }
 269         }
 270 end:
 271         /*
 272          * Re-lock the registry lock before the next loop.
 273          */
 274         mutex_lock(&rcu_registry_lock);
 275 }
 276
 277 /*
 278  * Always called with rcu_registry lock held. Releases this lock between
 279  * iterations and grabs it again. Holds the lock when it returns.
 280  */
 281 static void wait_for_readers(struct cds_list_head *input_readers,
 282                         struct cds_list_head *cur_snap_readers,
 283                         struct cds_list_head *qsreaders)
 284 {
 285         unsigned int wait_loops = 0;
 286         struct rcu_reader *index, *tmp;
 287 #ifdef HAS_INCOHERENT_CACHES
 288         unsigned int wait_gp_loops = 0;
 289 #endif /* HAS_INCOHERENT_CACHES */
 290
 291         /*
 292          * Wait for each thread URCU_TLS(rcu_reader).ctr to either
 293          * indicate quiescence (not nested), or observe the current
 294          * rcu_gp.ctr value.
 295          */
 296         for (;;) {
 297                 if (wait_loops < RCU_QS_ACTIVE_ATTEMPTS)
 298                         wait_loops++;
 299                 if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS) {
 300                         uatomic_dec(&rcu_gp.futex);
 301                         /* Write futex before read reader_gp */
 302                         smp_mb_master();
 303                 }
 304
 305                 cds_list_for_each_entry_safe(index, tmp, input_readers, node) {
 306                         switch (rcu_reader_state(&index->ctr)) {
 307                         case RCU_READER_ACTIVE_CURRENT:
 308                                 if (cur_snap_readers) {
 309                                         cds_list_move(&index->node,
 310                                                 cur_snap_readers);
 311                                         break;
 312                                 }
 313                                 /* Fall-through */
 314                         case RCU_READER_INACTIVE:
 315                                 cds_list_move(&index->node, qsreaders);
 316                                 break;
 317                         case RCU_READER_ACTIVE_OLD:
 318                                 /*
 319                                  * Old snapshot. Leaving node in
 320                                  * input_readers will make us busy-loop
 321                                  * until the snapshot becomes current or
 322                                  * the reader becomes inactive.
 323                                  */
 324                                 break;
 325                         }
 326                 }
 327
 328 #ifndef HAS_INCOHERENT_CACHES
 329                 if (cds_list_empty(input_readers)) {
 330                         if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS) {
 331                                 /* Read reader_gp before write futex */
 332                                 smp_mb_master();
 333                                 uatomic_set(&rcu_gp.futex, 0);
 334                         }
 335                         break;
 336                 } else {
 337                         if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS) {
 338                                 /* wait_gp unlocks/locks registry lock. */
 339                                 wait_gp();
 340                         } else {
 341                                 /* Temporarily unlock the registry lock. */
 342                                 mutex_unlock(&rcu_registry_lock);
 343                                 caa_cpu_relax();
 344                                 /*
 345                                  * Re-lock the registry lock before the
 346                                  * next loop.
 347                                  */
 348                                 mutex_lock(&rcu_registry_lock);
 349                         }
 350                 }
 351 #else /* #ifndef HAS_INCOHERENT_CACHES */
 352                 /*
 353                  * BUSY-LOOP. Force the reader thread to commit its
 354                  * URCU_TLS(rcu_reader).ctr update to memory if we wait
 355                  * for too long.
 356                  */
 357                 if (cds_list_empty(input_readers)) {
 358                         if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS) {
 359                                 /* Read reader_gp before write futex */
 360                                 smp_mb_master();
 361                                 uatomic_set(&rcu_gp.futex, 0);
 362                         }
 363                         break;
 364                 } else {
 365                         if (wait_gp_loops == KICK_READER_LOOPS) {
 366                                 smp_mb_master();
 367                                 wait_gp_loops = 0;
 368                         }
 369                         if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS) {
 370                                 /* wait_gp unlocks/locks registry lock. */
 371                                 wait_gp();
 372                                 wait_gp_loops++;
 373                         } else {
 374                                 /* Temporarily unlock the registry lock. */
 375                                 mutex_unlock(&rcu_registry_lock);
 376                                 caa_cpu_relax();
 377                                 /*
 378                                  * Re-lock the registry lock before the
 379                                  * next loop.
 380                                  */
 381                                 mutex_lock(&rcu_registry_lock);
 382                         }
 383                 }
 384 #endif /* #else #ifndef HAS_INCOHERENT_CACHES */
 385         }
 386 }
 387
 388 void synchronize_rcu(void)
 389 {
 390         CDS_LIST_HEAD(cur_snap_readers);
 391         CDS_LIST_HEAD(qsreaders);
 392         DEFINE_URCU_WAIT_NODE(wait, URCU_WAIT_WAITING);
 393         struct urcu_waiters waiters;
 394
 395         /*
 396          * Add ourself to gp_waiters queue of threads awaiting to wait
 397          * for a grace period. Proceed to perform the grace period only
 398          * if we are the first thread added into the queue.
 399          * The implicit memory barrier before urcu_wait_add()
 400          * orders prior memory accesses of threads put into the wait
 401          * queue before their insertion into the wait queue.
 402          */
 403         if (urcu_wait_add(&gp_waiters, &wait) != 0) {
 404                 /* Not first in queue: will be awakened by another thread. */
 405                 urcu_adaptative_busy_wait(&wait);
 406                 /* Order following memory accesses after grace period. */
 407                 cmm_smp_mb();
 408                 return;
 409         }
 410         /* We won't need to wake ourself up */
 411         urcu_wait_set_state(&wait, URCU_WAIT_RUNNING);
 412
 413         mutex_lock(&rcu_gp_lock);
 414
 415         /*
 416          * Move all waiters into our local queue.
 417          */
 418         urcu_move_waiters(&waiters, &gp_waiters);
 419
 420         mutex_lock(&rcu_registry_lock);
 421
 422         if (cds_list_empty(&registry))
 423                 goto out;
 424
 425         /*
 426          * All threads should read qparity before accessing data structure
 427          * where new ptr points to. Must be done within rcu_registry_lock
 428          * because it iterates on reader threads.
 429          */
 430         /* Write new ptr before changing the qparity */
 431         smp_mb_master();
 432
 433         /*
 434          * Wait for readers to observe original parity or be quiescent.
 435          * wait_for_readers() can release and grab again rcu_registry_lock
 436          * interally.
 437          */
 438         wait_for_readers(&registry, &cur_snap_readers, &qsreaders);
 439
 440         /*
 441          * Must finish waiting for quiescent state for original parity before
 442          * committing next rcu_gp.ctr update to memory. Failure to do so could
 443          * result in the writer waiting forever while new readers are always
 444          * accessing data (no progress).  Enforce compiler-order of load
 445          * URCU_TLS(rcu_reader).ctr before store to rcu_gp.ctr.
 446          */
 447         cmm_barrier();
 448
 449         /*
 450          * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
 451          * model easier to understand. It does not have a big performance impact
 452          * anyway, given this is the write-side.
 453          */
 454         cmm_smp_mb();
 455
 456         /* Switch parity: 0 -> 1, 1 -> 0 */
 457         CMM_STORE_SHARED(rcu_gp.ctr, rcu_gp.ctr ^ RCU_GP_CTR_PHASE);
 458
 459         /*
 460          * Must commit rcu_gp.ctr update to memory before waiting for quiescent
 461          * state. Failure to do so could result in the writer waiting forever
 462          * while new readers are always accessing data (no progress). Enforce
 463          * compiler-order of store to rcu_gp.ctr before load rcu_reader ctr.
 464          */
 465         cmm_barrier();
 466
 467         /*
 468          *
 469          * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
 470          * model easier to understand. It does not have a big performance impact
 471          * anyway, given this is the write-side.
 472          */
 473         cmm_smp_mb();
 474
 475         /*
 476          * Wait for readers to observe new parity or be quiescent.
 477          * wait_for_readers() can release and grab again rcu_registry_lock
 478          * interally.
 479          */
 480         wait_for_readers(&cur_snap_readers, NULL, &qsreaders);
 481
 482         /*
 483          * Put quiescent reader list back into registry.
 484          */
 485         cds_list_splice(&qsreaders, &registry);
 486
 487         /*
 488          * Finish waiting for reader threads before letting the old ptr
 489          * being freed. Must be done within rcu_registry_lock because it
 490          * iterates on reader threads.
 491          */
 492         smp_mb_master();
 493 out:
 494         mutex_unlock(&rcu_registry_lock);
 495         mutex_unlock(&rcu_gp_lock);
 496
 497         /*
 498          * Wakeup waiters only after we have completed the grace period
 499          * and have ensured the memory barriers at the end of the grace
 500          * period have been issued.
 501          */
 502         urcu_wake_all_waiters(&waiters);
 503 }
 504
 505 /*
 506  * library wrappers to be used by non-LGPL compatible source code.
 507  */
 508
 509 void rcu_read_lock(void)
 510 {
 511         _rcu_read_lock();
 512 }
 513
 514 void rcu_read_unlock(void)
 515 {
 516         _rcu_read_unlock();
 517 }
 518
 519 int rcu_read_ongoing(void)
 520 {
 521         return _rcu_read_ongoing();
 522 }
 523
 524 void rcu_register_thread(void)
 525 {
 526         URCU_TLS(rcu_reader).tid = pthread_self();
 527         assert(URCU_TLS(rcu_reader).need_mb == 0);
 528         assert(!(URCU_TLS(rcu_reader).ctr & RCU_GP_CTR_NEST_MASK));
 529
 530         mutex_lock(&rcu_registry_lock);
 531         assert(!URCU_TLS(rcu_reader).registered);
 532         URCU_TLS(rcu_reader).registered = 1;
 533         rcu_init();     /* In case gcc does not support constructor attribute */
 534         cds_list_add(&URCU_TLS(rcu_reader).node, &registry);
 535         mutex_unlock(&rcu_registry_lock);
 536 }
 537
 538 void rcu_unregister_thread(void)
 539 {
 540         mutex_lock(&rcu_registry_lock);
 541         assert(URCU_TLS(rcu_reader).registered);
 542         URCU_TLS(rcu_reader).registered = 0;
 543         cds_list_del(&URCU_TLS(rcu_reader).node);
 544         mutex_unlock(&rcu_registry_lock);
 545 }
 546
 547 #ifdef RCU_MEMBARRIER
 548
 549 #ifdef CONFIG_RCU_FORCE_SYS_MEMBARRIER
 550 static
 551 void rcu_sys_membarrier_status(bool available)
 552 {
 553         if (!available)
 554                 abort();
 555 }
 556 #else
 557 static
 558 void rcu_sys_membarrier_status(bool available)
 559 {
 560         if (!available)
 561                 return;
 562         rcu_has_sys_membarrier_memb = 1;
 563 }
 564 #endif
 565
 566 static
 567 void rcu_sys_membarrier_init(void)
 568 {
 569         bool available = false;
 570         int mask;
 571
 572         mask = membarrier(MEMBARRIER_CMD_QUERY, 0);
 573         if (mask >= 0) {
 574                 if (mask & MEMBARRIER_CMD_PRIVATE_EXPEDITED) {
 575                         if (membarrier(MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED, 0))
 576                                 urcu_die(errno);
 577                         has_sys_membarrier_private_expedited = 1;
 578                         available = true;
 579                 } else if (mask & MEMBARRIER_CMD_SHARED) {
 580                         available = true;
 581                 }
 582         }
 583         rcu_sys_membarrier_status(available);
 584 }
 585
 586 void rcu_init(void)
 587 {
 588         if (init_done)
 589                 return;
 590         init_done = 1;
 591         rcu_sys_membarrier_init();
 592 }
 593 #endif
 594
 595 #ifdef RCU_SIGNAL
 596 static void sigrcu_handler(int signo, siginfo_t *siginfo, void *context)
 597 {
 598         /*
 599          * Executing this cmm_smp_mb() is the only purpose of this signal handler.
 600          * It punctually promotes cmm_barrier() into cmm_smp_mb() on every thread it is
 601          * executed on.
 602          */
 603         cmm_smp_mb();
 604         _CMM_STORE_SHARED(URCU_TLS(rcu_reader).need_mb, 0);
 605         cmm_smp_mb();
 606 }
 607
 608 /*
 609  * rcu_init constructor. Called when the library is linked, but also when
 610  * reader threads are calling rcu_register_thread().
 611  * Should only be called by a single thread at a given time. This is ensured by
 612  * holing the rcu_registry_lock from rcu_register_thread() or by running
 613  * at library load time, which should not be executed by multiple
 614  * threads nor concurrently with rcu_register_thread() anyway.
 615  */
 616 void rcu_init(void)
 617 {
 618         struct sigaction act;
 619         int ret;
 620
 621         if (init_done)
 622                 return;
 623         init_done = 1;
 624
 625         act.sa_sigaction = sigrcu_handler;
 626         act.sa_flags = SA_SIGINFO | SA_RESTART;
 627         sigemptyset(&act.sa_mask);
 628         ret = sigaction(SIGRCU, &act, NULL);
 629         if (ret)
 630                 urcu_die(errno);
 631 }
 632
 633 void rcu_exit(void)
 634 {
 635         /*
 636          * Don't unregister the SIGRCU signal handler anymore, because
 637          * call_rcu threads could still be using it shortly before the
 638          * application exits.
 639          * Assertion disabled because call_rcu threads are now rcu
 640          * readers, and left running at exit.
 641          * assert(cds_list_empty(&registry));
 642          */
 643 }
 644
 645 #endif /* #ifdef RCU_SIGNAL */
 646
 647 DEFINE_RCU_FLAVOR(rcu_flavor);
 648
 649 #include "urcu-call-rcu-impl.h"
 650 #include "urcu-defer-impl.h"