src/urcu-qsbr.c

   1 /*
   2  * urcu-qsbr.c
   3  *
   4  * Userspace RCU QSBR 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 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 <stdint.h>
  34 #include <string.h>
  35 #include <errno.h>
  36 #include <poll.h>
  37
  38 #include <urcu/wfcqueue.h>
  39 #include <urcu/map/urcu-qsbr.h>
  40 #define BUILD_QSBR_LIB
  41 #include <urcu/static/urcu-qsbr.h>
  42 #include <urcu/pointer.h>
  43 #include <urcu/tls-compat.h>
  44
  45 #include "urcu-die.h"
  46 #include "urcu-wait.h"
  47 #include "urcu-utils.h"
  48
  49 #define URCU_API_MAP
  50 /* Do not #define _LGPL_SOURCE to ensure we can emit the wrapper symbols */
  51 #undef _LGPL_SOURCE
  52 #include <urcu/urcu-qsbr.h>
  53 #define _LGPL_SOURCE
  54
  55 void __attribute__((destructor)) urcu_qsbr_exit(void);
  56 static void urcu_call_rcu_exit(void);
  57
  58 /*
  59  * rcu_gp_lock ensures mutual exclusion between threads calling
  60  * synchronize_rcu().
  61  */
  62 static pthread_mutex_t rcu_gp_lock = PTHREAD_MUTEX_INITIALIZER;
  63 /*
  64  * rcu_registry_lock ensures mutual exclusion between threads
  65  * registering and unregistering themselves to/from the registry, and
  66  * with threads reading that registry from synchronize_rcu(). However,
  67  * this lock is not held all the way through the completion of awaiting
  68  * for the grace period. It is sporadically released between iterations
  69  * on the registry.
  70  * rcu_registry_lock may nest inside rcu_gp_lock.
  71  */
  72 static pthread_mutex_t rcu_registry_lock = PTHREAD_MUTEX_INITIALIZER;
  73 struct urcu_gp urcu_qsbr_gp = { .ctr = URCU_QSBR_GP_ONLINE };
  74 URCU_ATTR_ALIAS("urcu_qsbr_gp") extern struct urcu_gp rcu_gp_qsbr;
  75
  76 /*
  77  * Active attempts to check for reader Q.S. before calling futex().
  78  */
  79 #define RCU_QS_ACTIVE_ATTEMPTS 100
  80
  81 /*
  82  * Written to only by each individual reader. Read by both the reader and the
  83  * writers.
  84  */
  85 DEFINE_URCU_TLS(struct urcu_qsbr_reader, urcu_qsbr_reader);
  86 DEFINE_URCU_TLS_ALIAS(struct urcu_qsbr_reader, urcu_qsbr_reader, rcu_reader_qsbr);
  87
  88 static CDS_LIST_HEAD(registry);
  89
  90 /*
  91  * Queue keeping threads awaiting to wait for a grace period. Contains
  92  * struct gp_waiters_thread objects.
  93  */
  94 static DEFINE_URCU_WAIT_QUEUE(gp_waiters);
  95
  96 static void mutex_lock(pthread_mutex_t *mutex)
  97 {
  98         int ret;
  99
 100 #ifndef DISTRUST_SIGNALS_EXTREME
 101         ret = pthread_mutex_lock(mutex);
 102         if (ret)
 103                 urcu_die(ret);
 104 #else /* #ifndef DISTRUST_SIGNALS_EXTREME */
 105         while ((ret = pthread_mutex_trylock(mutex)) != 0) {
 106                 if (ret != EBUSY && ret != EINTR)
 107                         urcu_die(ret);
 108                 poll(NULL,0,10);
 109         }
 110 #endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */
 111 }
 112
 113 static void mutex_unlock(pthread_mutex_t *mutex)
 114 {
 115         int ret;
 116
 117         ret = pthread_mutex_unlock(mutex);
 118         if (ret)
 119                 urcu_die(ret);
 120 }
 121
 122 /*
 123  * synchronize_rcu() waiting. Single thread.
 124  */
 125 static void wait_gp(void)
 126 {
 127         /* Read reader_gp before read futex */
 128         cmm_smp_rmb();
 129         while (uatomic_read(&urcu_qsbr_gp.futex) == -1) {
 130                 if (!futex_noasync(&urcu_qsbr_gp.futex, FUTEX_WAIT, -1, NULL, NULL, 0)) {
 131                         /*
 132                          * Prior queued wakeups queued by unrelated code
 133                          * using the same address can cause futex wait to
 134                          * return 0 even through the futex value is still
 135                          * -1 (spurious wakeups). Check the value again
 136                          * in user-space to validate whether it really
 137                          * differs from -1.
 138                          */
 139                         continue;
 140                 }
 141                 switch (errno) {
 142                 case EAGAIN:
 143                         /* Value already changed. */
 144                         return;
 145                 case EINTR:
 146                         /* Retry if interrupted by signal. */
 147                         break;  /* Get out of switch. Check again. */
 148                 default:
 149                         /* Unexpected error. */
 150                         urcu_die(errno);
 151                 }
 152         }
 153 }
 154
 155 /*
 156  * Always called with rcu_registry lock held. Releases this lock between
 157  * iterations and grabs it again. Holds the lock when it returns.
 158  */
 159 static void wait_for_readers(struct cds_list_head *input_readers,
 160                         struct cds_list_head *cur_snap_readers,
 161                         struct cds_list_head *qsreaders)
 162 {
 163         unsigned int wait_loops = 0;
 164         struct urcu_qsbr_reader *index, *tmp;
 165
 166         /*
 167          * Wait for each thread URCU_TLS(urcu_qsbr_reader).ctr to either
 168          * indicate quiescence (offline), or for them to observe the
 169          * current urcu_qsbr_gp.ctr value.
 170          */
 171         for (;;) {
 172                 if (wait_loops < RCU_QS_ACTIVE_ATTEMPTS)
 173                         wait_loops++;
 174                 if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS) {
 175                         uatomic_set(&urcu_qsbr_gp.futex, -1);
 176                         /*
 177                          * Write futex before write waiting (the other side
 178                          * reads them in the opposite order).
 179                          */
 180                         cmm_smp_wmb();
 181                         cds_list_for_each_entry(index, input_readers, node) {
 182                                 _CMM_STORE_SHARED(index->waiting, 1);
 183                         }
 184                         /* Write futex before read reader_gp */
 185                         cmm_smp_mb();
 186                 }
 187                 cds_list_for_each_entry_safe(index, tmp, input_readers, node) {
 188                         switch (urcu_qsbr_reader_state(&index->ctr)) {
 189                         case URCU_READER_ACTIVE_CURRENT:
 190                                 if (cur_snap_readers) {
 191                                         cds_list_move(&index->node,
 192                                                 cur_snap_readers);
 193                                         break;
 194                                 }
 195                                 /* Fall-through */
 196                         case URCU_READER_INACTIVE:
 197                                 cds_list_move(&index->node, qsreaders);
 198                                 break;
 199                         case URCU_READER_ACTIVE_OLD:
 200                                 /*
 201                                  * Old snapshot. Leaving node in
 202                                  * input_readers will make us busy-loop
 203                                  * until the snapshot becomes current or
 204                                  * the reader becomes inactive.
 205                                  */
 206                                 break;
 207                         }
 208                 }
 209
 210                 if (cds_list_empty(input_readers)) {
 211                         if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS) {
 212                                 /* Read reader_gp before write futex */
 213                                 cmm_smp_mb();
 214                                 uatomic_set(&urcu_qsbr_gp.futex, 0);
 215                         }
 216                         break;
 217                 } else {
 218                         /* Temporarily unlock the registry lock. */
 219                         mutex_unlock(&rcu_registry_lock);
 220                         if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS) {
 221                                 wait_gp();
 222                         } else {
 223 #ifndef HAS_INCOHERENT_CACHES
 224                                 caa_cpu_relax();
 225 #else /* #ifndef HAS_INCOHERENT_CACHES */
 226                                 cmm_smp_mb();
 227 #endif /* #else #ifndef HAS_INCOHERENT_CACHES */
 228                         }
 229                         /* Re-lock the registry lock before the next loop. */
 230                         mutex_lock(&rcu_registry_lock);
 231                 }
 232         }
 233 }
 234
 235 /*
 236  * Using a two-subphases algorithm for architectures with smaller than 64-bit
 237  * long-size to ensure we do not encounter an overflow bug.
 238  */
 239
 240 #if (CAA_BITS_PER_LONG < 64)
 241 void urcu_qsbr_synchronize_rcu(void)
 242 {
 243         CDS_LIST_HEAD(cur_snap_readers);
 244         CDS_LIST_HEAD(qsreaders);
 245         unsigned long was_online;
 246         DEFINE_URCU_WAIT_NODE(wait, URCU_WAIT_WAITING);
 247         struct urcu_waiters waiters;
 248
 249         was_online = urcu_qsbr_read_ongoing();
 250
 251         /* All threads should read qparity before accessing data structure
 252          * where new ptr points to.  In the "then" case, rcu_thread_offline
 253          * includes a memory barrier.
 254          *
 255          * Mark the writer thread offline to make sure we don't wait for
 256          * our own quiescent state. This allows using synchronize_rcu()
 257          * in threads registered as readers.
 258          */
 259         if (was_online)
 260                 urcu_qsbr_thread_offline();
 261         else
 262                 cmm_smp_mb();
 263
 264         /*
 265          * Add ourself to gp_waiters queue of threads awaiting to wait
 266          * for a grace period. Proceed to perform the grace period only
 267          * if we are the first thread added into the queue.
 268          */
 269         if (urcu_wait_add(&gp_waiters, &wait) != 0) {
 270                 /* Not first in queue: will be awakened by another thread. */
 271                 urcu_adaptative_busy_wait(&wait);
 272                 goto gp_end;
 273         }
 274         /* We won't need to wake ourself up */
 275         urcu_wait_set_state(&wait, URCU_WAIT_RUNNING);
 276
 277         mutex_lock(&rcu_gp_lock);
 278
 279         /*
 280          * Move all waiters into our local queue.
 281          */
 282         urcu_move_waiters(&waiters, &gp_waiters);
 283
 284         mutex_lock(&rcu_registry_lock);
 285
 286         if (cds_list_empty(&registry))
 287                 goto out;
 288
 289         /*
 290          * Wait for readers to observe original parity or be quiescent.
 291          * wait_for_readers() can release and grab again rcu_registry_lock
 292          * interally.
 293          */
 294         wait_for_readers(&registry, &cur_snap_readers, &qsreaders);
 295
 296         /*
 297          * Must finish waiting for quiescent state for original parity
 298          * before committing next urcu_qsbr_gp.ctr update to memory. Failure
 299          * to do so could result in the writer waiting forever while new
 300          * readers are always accessing data (no progress).  Enforce
 301          * compiler-order of load URCU_TLS(urcu_qsbr_reader).ctr before store
 302          * to urcu_qsbr_gp.ctr.
 303          */
 304         cmm_barrier();
 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(urcu_qsbr_gp.ctr, urcu_qsbr_gp.ctr ^ URCU_QSBR_GP_CTR);
 315
 316         /*
 317          * Must commit urcu_qsbr_gp.ctr update to memory before waiting for
 318          * quiescent state. Failure to do so could result in the writer
 319          * waiting forever while new readers are always accessing data
 320          * (no progress). Enforce compiler-order of store to urcu_qsbr_gp.ctr
 321          * before load URCU_TLS(urcu_qsbr_reader).ctr.
 322          */
 323         cmm_barrier();
 324
 325         /*
 326          * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
 327          * model easier to understand. It does not have a big performance impact
 328          * anyway, given this is the write-side.
 329          */
 330         cmm_smp_mb();
 331
 332         /*
 333          * Wait for readers to observe new parity or be quiescent.
 334          * wait_for_readers() can release and grab again rcu_registry_lock
 335          * interally.
 336          */
 337         wait_for_readers(&cur_snap_readers, NULL, &qsreaders);
 338
 339         /*
 340          * Put quiescent reader list back into registry.
 341          */
 342         cds_list_splice(&qsreaders, &registry);
 343 out:
 344         mutex_unlock(&rcu_registry_lock);
 345         mutex_unlock(&rcu_gp_lock);
 346         urcu_wake_all_waiters(&waiters);
 347 gp_end:
 348         /*
 349          * Finish waiting for reader threads before letting the old ptr being
 350          * freed.
 351          */
 352         if (was_online)
 353                 urcu_qsbr_thread_online();
 354         else
 355                 cmm_smp_mb();
 356 }
 357 #else /* !(CAA_BITS_PER_LONG < 64) */
 358 void urcu_qsbr_synchronize_rcu(void)
 359 {
 360         CDS_LIST_HEAD(qsreaders);
 361         unsigned long was_online;
 362         DEFINE_URCU_WAIT_NODE(wait, URCU_WAIT_WAITING);
 363         struct urcu_waiters waiters;
 364
 365         was_online = urcu_qsbr_read_ongoing();
 366
 367         /*
 368          * Mark the writer thread offline to make sure we don't wait for
 369          * our own quiescent state. This allows using synchronize_rcu()
 370          * in threads registered as readers.
 371          */
 372         if (was_online)
 373                 urcu_qsbr_thread_offline();
 374         else
 375                 cmm_smp_mb();
 376
 377         /*
 378          * Add ourself to gp_waiters queue of threads awaiting to wait
 379          * for a grace period. Proceed to perform the grace period only
 380          * if we are the first thread added into the queue.
 381          */
 382         if (urcu_wait_add(&gp_waiters, &wait) != 0) {
 383                 /* Not first in queue: will be awakened by another thread. */
 384                 urcu_adaptative_busy_wait(&wait);
 385                 goto gp_end;
 386         }
 387         /* We won't need to wake ourself up */
 388         urcu_wait_set_state(&wait, URCU_WAIT_RUNNING);
 389
 390         mutex_lock(&rcu_gp_lock);
 391
 392         /*
 393          * Move all waiters into our local queue.
 394          */
 395         urcu_move_waiters(&waiters, &gp_waiters);
 396
 397         mutex_lock(&rcu_registry_lock);
 398
 399         if (cds_list_empty(&registry))
 400                 goto out;
 401
 402         /* Increment current G.P. */
 403         CMM_STORE_SHARED(urcu_qsbr_gp.ctr, urcu_qsbr_gp.ctr + URCU_QSBR_GP_CTR);
 404
 405         /*
 406          * Must commit urcu_qsbr_gp.ctr update to memory before waiting for
 407          * quiescent state. Failure to do so could result in the writer
 408          * waiting forever while new readers are always accessing data
 409          * (no progress). Enforce compiler-order of store to urcu_qsbr_gp.ctr
 410          * before load URCU_TLS(urcu_qsbr_reader).ctr.
 411          */
 412         cmm_barrier();
 413
 414         /*
 415          * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
 416          * model easier to understand. It does not have a big performance impact
 417          * anyway, given this is the write-side.
 418          */
 419         cmm_smp_mb();
 420
 421         /*
 422          * Wait for readers to observe new count of be quiescent.
 423          * wait_for_readers() can release and grab again rcu_registry_lock
 424          * interally.
 425          */
 426         wait_for_readers(&registry, NULL, &qsreaders);
 427
 428         /*
 429          * Put quiescent reader list back into registry.
 430          */
 431         cds_list_splice(&qsreaders, &registry);
 432 out:
 433         mutex_unlock(&rcu_registry_lock);
 434         mutex_unlock(&rcu_gp_lock);
 435         urcu_wake_all_waiters(&waiters);
 436 gp_end:
 437         if (was_online)
 438                 urcu_qsbr_thread_online();
 439         else
 440                 cmm_smp_mb();
 441 }
 442 #endif  /* !(CAA_BITS_PER_LONG < 64) */
 443 URCU_ATTR_ALIAS("urcu_qsbr_synchronize_rcu")
 444 void synchronize_rcu_qsbr();
 445
 446 /*
 447  * library wrappers to be used by non-LGPL compatible source code.
 448  */
 449
 450 void urcu_qsbr_read_lock(void)
 451 {
 452         _urcu_qsbr_read_lock();
 453 }
 454 URCU_ATTR_ALIAS("urcu_qsbr_read_lock") void rcu_read_lock_qsbr();
 455
 456 void urcu_qsbr_read_unlock(void)
 457 {
 458         _urcu_qsbr_read_unlock();
 459 }
 460 URCU_ATTR_ALIAS("urcu_qsbr_read_unlock") void rcu_read_unlock_qsbr();
 461
 462 int urcu_qsbr_read_ongoing(void)
 463 {
 464         return _urcu_qsbr_read_ongoing();
 465 }
 466 URCU_ATTR_ALIAS("urcu_qsbr_read_ongoing")
 467 void rcu_read_ongoing_qsbr();
 468
 469 void urcu_qsbr_quiescent_state(void)
 470 {
 471         _urcu_qsbr_quiescent_state();
 472 }
 473 URCU_ATTR_ALIAS("urcu_qsbr_quiescent_state")
 474 void rcu_quiescent_state_qsbr();
 475
 476 void urcu_qsbr_thread_offline(void)
 477 {
 478         _urcu_qsbr_thread_offline();
 479 }
 480 URCU_ATTR_ALIAS("urcu_qsbr_thread_offline")
 481 void rcu_thread_offline_qsbr();
 482
 483 void urcu_qsbr_thread_online(void)
 484 {
 485         _urcu_qsbr_thread_online();
 486 }
 487 URCU_ATTR_ALIAS("urcu_qsbr_thread_online")
 488 void rcu_thread_online_qsbr();
 489
 490 void urcu_qsbr_register_thread(void)
 491 {
 492         URCU_TLS(urcu_qsbr_reader).tid = pthread_self();
 493         assert(URCU_TLS(urcu_qsbr_reader).ctr == 0);
 494
 495         mutex_lock(&rcu_registry_lock);
 496         assert(!URCU_TLS(urcu_qsbr_reader).registered);
 497         URCU_TLS(urcu_qsbr_reader).registered = 1;
 498         cds_list_add(&URCU_TLS(urcu_qsbr_reader).node, &registry);
 499         mutex_unlock(&rcu_registry_lock);
 500         _urcu_qsbr_thread_online();
 501 }
 502 URCU_ATTR_ALIAS("urcu_qsbr_register_thread")
 503 void rcu_register_thread_qsbr();
 504
 505 void urcu_qsbr_unregister_thread(void)
 506 {
 507         /*
 508          * We have to make the thread offline otherwise we end up dealocking
 509          * with a waiting writer.
 510          */
 511         _urcu_qsbr_thread_offline();
 512         assert(URCU_TLS(urcu_qsbr_reader).registered);
 513         URCU_TLS(urcu_qsbr_reader).registered = 0;
 514         mutex_lock(&rcu_registry_lock);
 515         cds_list_del(&URCU_TLS(urcu_qsbr_reader).node);
 516         mutex_unlock(&rcu_registry_lock);
 517 }
 518 URCU_ATTR_ALIAS("urcu_qsbr_unregister_thread")
 519 void rcu_unregister_thread_qsbr();
 520
 521 void urcu_qsbr_exit(void)
 522 {
 523         /*
 524          * Assertion disabled because call_rcu threads are now rcu
 525          * readers, and left running at exit.
 526          * assert(cds_list_empty(&registry));
 527          */
 528         urcu_call_rcu_exit();
 529 }
 530 URCU_ATTR_ALIAS("urcu_qsbr_exit") void rcu_exit_qsbr();
 531
 532 DEFINE_RCU_FLAVOR(rcu_flavor);
 533 DEFINE_RCU_FLAVOR_ALIAS(rcu_flavor, alias_rcu_flavor);
 534
 535 #include "urcu-call-rcu-impl.h"
 536 #include "urcu-defer-impl.h"