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 <stdlib.h>
  32 #include <stdint.h>
  33 #include <string.h>
  34 #include <errno.h>
  35 #include <poll.h>
  36
  37 #include <urcu/assert.h>
  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
  75 /*
  76  * Active attempts to check for reader Q.S. before calling futex().
  77  */
  78 #define RCU_QS_ACTIVE_ATTEMPTS 100
  79
  80 /*
  81  * Written to only by each individual reader. Read by both the reader and the
  82  * writers.
  83  */
  84 DEFINE_URCU_TLS(struct urcu_qsbr_reader, urcu_qsbr_reader);
  85
  86 static CDS_LIST_HEAD(registry);
  87
  88 /*
  89  * Queue keeping threads awaiting to wait for a grace period. Contains
  90  * struct gp_waiters_thread objects.
  91  */
  92 static DEFINE_URCU_WAIT_QUEUE(gp_waiters);
  93
  94 static void mutex_lock(pthread_mutex_t *mutex)
  95 {
  96         int ret;
  97
  98 #ifndef DISTRUST_SIGNALS_EXTREME
  99         ret = pthread_mutex_lock(mutex);
 100         if (ret)
 101                 urcu_die(ret);
 102 #else /* #ifndef DISTRUST_SIGNALS_EXTREME */
 103         while ((ret = pthread_mutex_trylock(mutex)) != 0) {
 104                 if (ret != EBUSY && ret != EINTR)
 105                         urcu_die(ret);
 106                 poll(NULL,0,10);
 107         }
 108 #endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */
 109 }
 110
 111 static void mutex_unlock(pthread_mutex_t *mutex)
 112 {
 113         int ret;
 114
 115         ret = pthread_mutex_unlock(mutex);
 116         if (ret)
 117                 urcu_die(ret);
 118 }
 119
 120 /*
 121  * synchronize_rcu() waiting. Single thread.
 122  */
 123 static void wait_gp(void)
 124 {
 125         /* Read reader_gp before read futex */
 126         cmm_smp_rmb();
 127         while (uatomic_read(&urcu_qsbr_gp.futex) == -1) {
 128                 if (!futex_noasync(&urcu_qsbr_gp.futex, FUTEX_WAIT, -1, NULL, NULL, 0)) {
 129                         /*
 130                          * Prior queued wakeups queued by unrelated code
 131                          * using the same address can cause futex wait to
 132                          * return 0 even through the futex value is still
 133                          * -1 (spurious wakeups). Check the value again
 134                          * in user-space to validate whether it really
 135                          * differs from -1.
 136                          */
 137                         continue;
 138                 }
 139                 switch (errno) {
 140                 case EAGAIN:
 141                         /* Value already changed. */
 142                         return;
 143                 case EINTR:
 144                         /* Retry if interrupted by signal. */
 145                         break;  /* Get out of switch. Check again. */
 146                 default:
 147                         /* Unexpected error. */
 148                         urcu_die(errno);
 149                 }
 150         }
 151 }
 152
 153 /*
 154  * Always called with rcu_registry lock held. Releases this lock between
 155  * iterations and grabs it again. Holds the lock when it returns.
 156  */
 157 static void wait_for_readers(struct cds_list_head *input_readers,
 158                         struct cds_list_head *cur_snap_readers,
 159                         struct cds_list_head *qsreaders)
 160 {
 161         unsigned int wait_loops = 0;
 162         struct urcu_qsbr_reader *index, *tmp;
 163
 164         /*
 165          * Wait for each thread URCU_TLS(urcu_qsbr_reader).ctr to either
 166          * indicate quiescence (offline), or for them to observe the
 167          * current urcu_qsbr_gp.ctr value.
 168          */
 169         for (;;) {
 170                 if (wait_loops < RCU_QS_ACTIVE_ATTEMPTS)
 171                         wait_loops++;
 172                 if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS) {
 173                         uatomic_set(&urcu_qsbr_gp.futex, -1);
 174                         /*
 175                          * Write futex before write waiting (the other side
 176                          * reads them in the opposite order).
 177                          */
 178                         cmm_smp_wmb();
 179                         cds_list_for_each_entry(index, input_readers, node) {
 180                                 _CMM_STORE_SHARED(index->waiting, 1);
 181                         }
 182                         /* Write futex before read reader_gp */
 183                         cmm_smp_mb();
 184                 }
 185                 cds_list_for_each_entry_safe(index, tmp, input_readers, node) {
 186                         switch (urcu_qsbr_reader_state(&index->ctr)) {
 187                         case URCU_READER_ACTIVE_CURRENT:
 188                                 if (cur_snap_readers) {
 189                                         cds_list_move(&index->node,
 190                                                 cur_snap_readers);
 191                                         break;
 192                                 }
 193                                 /* Fall-through */
 194                         case URCU_READER_INACTIVE:
 195                                 cds_list_move(&index->node, qsreaders);
 196                                 break;
 197                         case URCU_READER_ACTIVE_OLD:
 198                                 /*
 199                                  * Old snapshot. Leaving node in
 200                                  * input_readers will make us busy-loop
 201                                  * until the snapshot becomes current or
 202                                  * the reader becomes inactive.
 203                                  */
 204                                 break;
 205                         }
 206                 }
 207
 208                 if (cds_list_empty(input_readers)) {
 209                         if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS) {
 210                                 /* Read reader_gp before write futex */
 211                                 cmm_smp_mb();
 212                                 uatomic_set(&urcu_qsbr_gp.futex, 0);
 213                         }
 214                         break;
 215                 } else {
 216                         /* Temporarily unlock the registry lock. */
 217                         mutex_unlock(&rcu_registry_lock);
 218                         if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS) {
 219                                 wait_gp();
 220                         } else {
 221 #ifndef HAS_INCOHERENT_CACHES
 222                                 caa_cpu_relax();
 223 #else /* #ifndef HAS_INCOHERENT_CACHES */
 224                                 cmm_smp_mb();
 225 #endif /* #else #ifndef HAS_INCOHERENT_CACHES */
 226                         }
 227                         /* Re-lock the registry lock before the next loop. */
 228                         mutex_lock(&rcu_registry_lock);
 229                 }
 230         }
 231 }
 232
 233 /*
 234  * Using a two-subphases algorithm for architectures with smaller than 64-bit
 235  * long-size to ensure we do not encounter an overflow bug.
 236  */
 237
 238 #if (CAA_BITS_PER_LONG < 64)
 239 void urcu_qsbr_synchronize_rcu(void)
 240 {
 241         CDS_LIST_HEAD(cur_snap_readers);
 242         CDS_LIST_HEAD(qsreaders);
 243         unsigned long was_online;
 244         DEFINE_URCU_WAIT_NODE(wait, URCU_WAIT_WAITING);
 245         struct urcu_waiters waiters;
 246
 247         was_online = urcu_qsbr_read_ongoing();
 248
 249         /* All threads should read qparity before accessing data structure
 250          * where new ptr points to.  In the "then" case, rcu_thread_offline
 251          * includes a memory barrier.
 252          *
 253          * Mark the writer thread offline to make sure we don't wait for
 254          * our own quiescent state. This allows using synchronize_rcu()
 255          * in threads registered as readers.
 256          */
 257         if (was_online)
 258                 urcu_qsbr_thread_offline();
 259         else
 260                 cmm_smp_mb();
 261
 262         /*
 263          * Add ourself to gp_waiters queue of threads awaiting to wait
 264          * for a grace period. Proceed to perform the grace period only
 265          * if we are the first thread added into the queue.
 266          */
 267         if (urcu_wait_add(&gp_waiters, &wait) != 0) {
 268                 /* Not first in queue: will be awakened by another thread. */
 269                 urcu_adaptative_busy_wait(&wait);
 270                 goto gp_end;
 271         }
 272         /* We won't need to wake ourself up */
 273         urcu_wait_set_state(&wait, URCU_WAIT_RUNNING);
 274
 275         mutex_lock(&rcu_gp_lock);
 276
 277         /*
 278          * Move all waiters into our local queue.
 279          */
 280         urcu_move_waiters(&waiters, &gp_waiters);
 281
 282         mutex_lock(&rcu_registry_lock);
 283
 284         if (cds_list_empty(&registry))
 285                 goto out;
 286
 287         /*
 288          * Wait for readers to observe original parity or be quiescent.
 289          * wait_for_readers() can release and grab again rcu_registry_lock
 290          * internally.
 291          */
 292         wait_for_readers(&registry, &cur_snap_readers, &qsreaders);
 293
 294         /*
 295          * Must finish waiting for quiescent state for original parity
 296          * before committing next urcu_qsbr_gp.ctr update to memory. Failure
 297          * to do so could result in the writer waiting forever while new
 298          * readers are always accessing data (no progress).  Enforce
 299          * compiler-order of load URCU_TLS(urcu_qsbr_reader).ctr before store
 300          * to urcu_qsbr_gp.ctr.
 301          */
 302         cmm_barrier();
 303
 304         /*
 305          * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
 306          * model easier to understand. It does not have a big performance impact
 307          * anyway, given this is the write-side.
 308          */
 309         cmm_smp_mb();
 310
 311         /* Switch parity: 0 -> 1, 1 -> 0 */
 312         CMM_STORE_SHARED(urcu_qsbr_gp.ctr, urcu_qsbr_gp.ctr ^ URCU_QSBR_GP_CTR);
 313
 314         /*
 315          * Must commit urcu_qsbr_gp.ctr update to memory before waiting for
 316          * quiescent state. Failure to do so could result in the writer
 317          * waiting forever while new readers are always accessing data
 318          * (no progress). Enforce compiler-order of store to urcu_qsbr_gp.ctr
 319          * before load URCU_TLS(urcu_qsbr_reader).ctr.
 320          */
 321         cmm_barrier();
 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          * internally.
 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 out:
 342         mutex_unlock(&rcu_registry_lock);
 343         mutex_unlock(&rcu_gp_lock);
 344         urcu_wake_all_waiters(&waiters);
 345 gp_end:
 346         /*
 347          * Finish waiting for reader threads before letting the old ptr being
 348          * freed.
 349          */
 350         if (was_online)
 351                 urcu_qsbr_thread_online();
 352         else
 353                 cmm_smp_mb();
 354 }
 355 #else /* !(CAA_BITS_PER_LONG < 64) */
 356 void urcu_qsbr_synchronize_rcu(void)
 357 {
 358         CDS_LIST_HEAD(qsreaders);
 359         unsigned long was_online;
 360         DEFINE_URCU_WAIT_NODE(wait, URCU_WAIT_WAITING);
 361         struct urcu_waiters waiters;
 362
 363         was_online = urcu_qsbr_read_ongoing();
 364
 365         /*
 366          * Mark the writer thread offline to make sure we don't wait for
 367          * our own quiescent state. This allows using synchronize_rcu()
 368          * in threads registered as readers.
 369          */
 370         if (was_online)
 371                 urcu_qsbr_thread_offline();
 372         else
 373                 cmm_smp_mb();
 374
 375         /*
 376          * Add ourself to gp_waiters queue of threads awaiting to wait
 377          * for a grace period. Proceed to perform the grace period only
 378          * if we are the first thread added into the queue.
 379          */
 380         if (urcu_wait_add(&gp_waiters, &wait) != 0) {
 381                 /* Not first in queue: will be awakened by another thread. */
 382                 urcu_adaptative_busy_wait(&wait);
 383                 goto gp_end;
 384         }
 385         /* We won't need to wake ourself up */
 386         urcu_wait_set_state(&wait, URCU_WAIT_RUNNING);
 387
 388         mutex_lock(&rcu_gp_lock);
 389
 390         /*
 391          * Move all waiters into our local queue.
 392          */
 393         urcu_move_waiters(&waiters, &gp_waiters);
 394
 395         mutex_lock(&rcu_registry_lock);
 396
 397         if (cds_list_empty(&registry))
 398                 goto out;
 399
 400         /* Increment current G.P. */
 401         CMM_STORE_SHARED(urcu_qsbr_gp.ctr, urcu_qsbr_gp.ctr + URCU_QSBR_GP_CTR);
 402
 403         /*
 404          * Must commit urcu_qsbr_gp.ctr update to memory before waiting for
 405          * quiescent state. Failure to do so could result in the writer
 406          * waiting forever while new readers are always accessing data
 407          * (no progress). Enforce compiler-order of store to urcu_qsbr_gp.ctr
 408          * before load URCU_TLS(urcu_qsbr_reader).ctr.
 409          */
 410         cmm_barrier();
 411
 412         /*
 413          * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
 414          * model easier to understand. It does not have a big performance impact
 415          * anyway, given this is the write-side.
 416          */
 417         cmm_smp_mb();
 418
 419         /*
 420          * Wait for readers to observe new count of be quiescent.
 421          * wait_for_readers() can release and grab again rcu_registry_lock
 422          * internally.
 423          */
 424         wait_for_readers(&registry, NULL, &qsreaders);
 425
 426         /*
 427          * Put quiescent reader list back into registry.
 428          */
 429         cds_list_splice(&qsreaders, &registry);
 430 out:
 431         mutex_unlock(&rcu_registry_lock);
 432         mutex_unlock(&rcu_gp_lock);
 433         urcu_wake_all_waiters(&waiters);
 434 gp_end:
 435         if (was_online)
 436                 urcu_qsbr_thread_online();
 437         else
 438                 cmm_smp_mb();
 439 }
 440 #endif  /* !(CAA_BITS_PER_LONG < 64) */
 441
 442 /*
 443  * library wrappers to be used by non-LGPL compatible source code.
 444  */
 445
 446 void urcu_qsbr_read_lock(void)
 447 {
 448         _urcu_qsbr_read_lock();
 449 }
 450
 451 void urcu_qsbr_read_unlock(void)
 452 {
 453         _urcu_qsbr_read_unlock();
 454 }
 455
 456 int urcu_qsbr_read_ongoing(void)
 457 {
 458         return _urcu_qsbr_read_ongoing();
 459 }
 460 void rcu_read_ongoing_qsbr();
 461
 462 void urcu_qsbr_quiescent_state(void)
 463 {
 464         _urcu_qsbr_quiescent_state();
 465 }
 466 void rcu_quiescent_state_qsbr();
 467
 468 void urcu_qsbr_thread_offline(void)
 469 {
 470         _urcu_qsbr_thread_offline();
 471 }
 472 void rcu_thread_offline_qsbr();
 473
 474 void urcu_qsbr_thread_online(void)
 475 {
 476         _urcu_qsbr_thread_online();
 477 }
 478
 479 void urcu_qsbr_register_thread(void)
 480 {
 481         URCU_TLS(urcu_qsbr_reader).tid = pthread_self();
 482         urcu_posix_assert(URCU_TLS(urcu_qsbr_reader).ctr == 0);
 483
 484         mutex_lock(&rcu_registry_lock);
 485         urcu_posix_assert(!URCU_TLS(urcu_qsbr_reader).registered);
 486         URCU_TLS(urcu_qsbr_reader).registered = 1;
 487         cds_list_add(&URCU_TLS(urcu_qsbr_reader).node, &registry);
 488         mutex_unlock(&rcu_registry_lock);
 489         _urcu_qsbr_thread_online();
 490 }
 491
 492 void urcu_qsbr_unregister_thread(void)
 493 {
 494         /*
 495          * We have to make the thread offline otherwise we end up dealocking
 496          * with a waiting writer.
 497          */
 498         _urcu_qsbr_thread_offline();
 499         urcu_posix_assert(URCU_TLS(urcu_qsbr_reader).registered);
 500         URCU_TLS(urcu_qsbr_reader).registered = 0;
 501         mutex_lock(&rcu_registry_lock);
 502         cds_list_del(&URCU_TLS(urcu_qsbr_reader).node);
 503         mutex_unlock(&rcu_registry_lock);
 504 }
 505
 506 void urcu_qsbr_exit(void)
 507 {
 508         /*
 509          * Assertion disabled because call_rcu threads are now rcu
 510          * readers, and left running at exit.
 511          * urcu_posix_assert(cds_list_empty(&registry));
 512          */
 513         urcu_call_rcu_exit();
 514 }
 515
 516 DEFINE_RCU_FLAVOR(rcu_flavor);
 517
 518 #include "urcu-call-rcu-impl.h"
 519 #include "urcu-defer-impl.h"
 520 #include "urcu-poll-impl.h"