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 _GNU_SOURCE
  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
  48 /* Do not #define _LGPL_SOURCE to ensure we can emit the wrapper symbols */
  49 #undef _LGPL_SOURCE
  50 #include "urcu-qsbr.h"
  51 #define _LGPL_SOURCE
  52
  53 void __attribute__((destructor)) rcu_exit(void);
  54
  55 /*
  56  * rcu_gp_lock ensures mutual exclusion between threads calling
  57  * synchronize_rcu().
  58  */
  59 static pthread_mutex_t rcu_gp_lock = PTHREAD_MUTEX_INITIALIZER;
  60 /*
  61  * rcu_registry_lock ensures mutual exclusion between threads
  62  * registering and unregistering themselves to/from the registry, and
  63  * with threads reading that registry from synchronize_rcu(). However,
  64  * this lock is not held all the way through the completion of awaiting
  65  * for the grace period. It is sporadically released between iterations
  66  * on the registry.
  67  * rcu_registry_lock may nest inside rcu_gp_lock.
  68  */
  69 static pthread_mutex_t rcu_registry_lock = PTHREAD_MUTEX_INITIALIZER;
  70 struct rcu_gp rcu_gp = { .ctr = RCU_GP_ONLINE };
  71
  72 /*
  73  * Active attempts to check for reader Q.S. before calling futex().
  74  */
  75 #define RCU_QS_ACTIVE_ATTEMPTS 100
  76
  77 /*
  78  * Written to only by each individual reader. Read by both the reader and the
  79  * writers.
  80  */
  81 DEFINE_URCU_TLS(struct rcu_reader, rcu_reader);
  82
  83 static CDS_LIST_HEAD(registry);
  84
  85 /*
  86  * Queue keeping threads awaiting to wait for a grace period. Contains
  87  * struct gp_waiters_thread objects.
  88  */
  89 static DEFINE_URCU_WAIT_QUEUE(gp_waiters);
  90
  91 static void mutex_lock(pthread_mutex_t *mutex)
  92 {
  93         int ret;
  94
  95 #ifndef DISTRUST_SIGNALS_EXTREME
  96         ret = pthread_mutex_lock(mutex);
  97         if (ret)
  98                 urcu_die(ret);
  99 #else /* #ifndef DISTRUST_SIGNALS_EXTREME */
 100         while ((ret = pthread_mutex_trylock(mutex)) != 0) {
 101                 if (ret != EBUSY && ret != EINTR)
 102                         urcu_die(ret);
 103                 poll(NULL,0,10);
 104         }
 105 #endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */
 106 }
 107
 108 static void mutex_unlock(pthread_mutex_t *mutex)
 109 {
 110         int ret;
 111
 112         ret = pthread_mutex_unlock(mutex);
 113         if (ret)
 114                 urcu_die(ret);
 115 }
 116
 117 /*
 118  * synchronize_rcu() waiting. Single thread.
 119  */
 120 static void wait_gp(void)
 121 {
 122         /* Read reader_gp before read futex */
 123         cmm_smp_rmb();
 124         if (uatomic_read(&rcu_gp.futex) == -1)
 125                 futex_noasync(&rcu_gp.futex, FUTEX_WAIT, -1,
 126                       NULL, NULL, 0);
 127 }
 128
 129 /*
 130  * Always called with rcu_registry lock held. Releases this lock between
 131  * iterations and grabs it again. Holds the lock when it returns.
 132  */
 133 static void wait_for_readers(struct cds_list_head *input_readers,
 134                         struct cds_list_head *cur_snap_readers,
 135                         struct cds_list_head *qsreaders)
 136 {
 137         unsigned int wait_loops = 0;
 138         struct rcu_reader *index, *tmp;
 139
 140         /*
 141          * Wait for each thread URCU_TLS(rcu_reader).ctr to either
 142          * indicate quiescence (offline), or for them to observe the
 143          * current rcu_gp.ctr value.
 144          */
 145         for (;;) {
 146                 if (wait_loops < RCU_QS_ACTIVE_ATTEMPTS)
 147                         wait_loops++;
 148                 if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS) {
 149                         uatomic_set(&rcu_gp.futex, -1);
 150                         /*
 151                          * Write futex before write waiting (the other side
 152                          * reads them in the opposite order).
 153                          */
 154                         cmm_smp_wmb();
 155                         cds_list_for_each_entry(index, input_readers, node) {
 156                                 _CMM_STORE_SHARED(index->waiting, 1);
 157                         }
 158                         /* Write futex before read reader_gp */
 159                         cmm_smp_mb();
 160                 }
 161                 cds_list_for_each_entry_safe(index, tmp, input_readers, node) {
 162                         switch (rcu_reader_state(&index->ctr)) {
 163                         case RCU_READER_ACTIVE_CURRENT:
 164                                 if (cur_snap_readers) {
 165                                         cds_list_move(&index->node,
 166                                                 cur_snap_readers);
 167                                         break;
 168                                 }
 169                                 /* Fall-through */
 170                         case RCU_READER_INACTIVE:
 171                                 cds_list_move(&index->node, qsreaders);
 172                                 break;
 173                         case RCU_READER_ACTIVE_OLD:
 174                                 /*
 175                                  * Old snapshot. Leaving node in
 176                                  * input_readers will make us busy-loop
 177                                  * until the snapshot becomes current or
 178                                  * the reader becomes inactive.
 179                                  */
 180                                 break;
 181                         }
 182                 }
 183
 184                 if (cds_list_empty(input_readers)) {
 185                         if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS) {
 186                                 /* Read reader_gp before write futex */
 187                                 cmm_smp_mb();
 188                                 uatomic_set(&rcu_gp.futex, 0);
 189                         }
 190                         break;
 191                 } else {
 192                         /* Temporarily unlock the registry lock. */
 193                         mutex_unlock(&rcu_registry_lock);
 194                         if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS) {
 195                                 wait_gp();
 196                         } else {
 197 #ifndef HAS_INCOHERENT_CACHES
 198                                 caa_cpu_relax();
 199 #else /* #ifndef HAS_INCOHERENT_CACHES */
 200                                 cmm_smp_mb();
 201 #endif /* #else #ifndef HAS_INCOHERENT_CACHES */
 202                         }
 203                         /* Re-lock the registry lock before the next loop. */
 204                         mutex_lock(&rcu_registry_lock);
 205                 }
 206         }
 207 }
 208
 209 /*
 210  * Using a two-subphases algorithm for architectures with smaller than 64-bit
 211  * long-size to ensure we do not encounter an overflow bug.
 212  */
 213
 214 #if (CAA_BITS_PER_LONG < 64)
 215 void synchronize_rcu(void)
 216 {
 217         CDS_LIST_HEAD(cur_snap_readers);
 218         CDS_LIST_HEAD(qsreaders);
 219         unsigned long was_online;
 220         DEFINE_URCU_WAIT_NODE(wait, URCU_WAIT_WAITING);
 221         struct urcu_waiters waiters;
 222
 223         was_online = rcu_read_ongoing();
 224
 225         /* All threads should read qparity before accessing data structure
 226          * where new ptr points to.  In the "then" case, rcu_thread_offline
 227          * includes a memory barrier.
 228          *
 229          * Mark the writer thread offline to make sure we don't wait for
 230          * our own quiescent state. This allows using synchronize_rcu()
 231          * in threads registered as readers.
 232          */
 233         if (was_online)
 234                 rcu_thread_offline();
 235         else
 236                 cmm_smp_mb();
 237
 238         /*
 239          * Add ourself to gp_waiters queue of threads awaiting to wait
 240          * for a grace period. Proceed to perform the grace period only
 241          * if we are the first thread added into the queue.
 242          */
 243         if (urcu_wait_add(&gp_waiters, &wait) != 0) {
 244                 /* Not first in queue: will be awakened by another thread. */
 245                 urcu_adaptative_busy_wait(&wait);
 246                 goto gp_end;
 247         }
 248         /* We won't need to wake ourself up */
 249         urcu_wait_set_state(&wait, URCU_WAIT_RUNNING);
 250
 251         mutex_lock(&rcu_gp_lock);
 252
 253         /*
 254          * Move all waiters into our local queue.
 255          */
 256         urcu_move_waiters(&waiters, &gp_waiters);
 257
 258         mutex_lock(&rcu_registry_lock);
 259
 260         if (cds_list_empty(&registry))
 261                 goto out;
 262
 263         /*
 264          * Wait for readers to observe original parity or be quiescent.
 265          * wait_for_readers() can release and grab again rcu_registry_lock
 266          * interally.
 267          */
 268         wait_for_readers(&registry, &cur_snap_readers, &qsreaders);
 269
 270         /*
 271          * Must finish waiting for quiescent state for original parity
 272          * before committing next rcu_gp.ctr update to memory. Failure
 273          * to do so could result in the writer waiting forever while new
 274          * readers are always accessing data (no progress).  Enforce
 275          * compiler-order of load URCU_TLS(rcu_reader).ctr before store
 276          * to rcu_gp.ctr.
 277          */
 278         cmm_barrier();
 279
 280         /*
 281          * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
 282          * model easier to understand. It does not have a big performance impact
 283          * anyway, given this is the write-side.
 284          */
 285         cmm_smp_mb();
 286
 287         /* Switch parity: 0 -> 1, 1 -> 0 */
 288         CMM_STORE_SHARED(rcu_gp.ctr, rcu_gp.ctr ^ RCU_GP_CTR);
 289
 290         /*
 291          * Must commit rcu_gp.ctr update to memory before waiting for
 292          * quiescent state. Failure to do so could result in the writer
 293          * waiting forever while new readers are always accessing data
 294          * (no progress). Enforce compiler-order of store to rcu_gp.ctr
 295          * before load URCU_TLS(rcu_reader).ctr.
 296          */
 297         cmm_barrier();
 298
 299         /*
 300          * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
 301          * model easier to understand. It does not have a big performance impact
 302          * anyway, given this is the write-side.
 303          */
 304         cmm_smp_mb();
 305
 306         /*
 307          * Wait for readers to observe new parity or be quiescent.
 308          * wait_for_readers() can release and grab again rcu_registry_lock
 309          * interally.
 310          */
 311         wait_for_readers(&cur_snap_readers, NULL, &qsreaders);
 312
 313         /*
 314          * Put quiescent reader list back into registry.
 315          */
 316         cds_list_splice(&qsreaders, &registry);
 317 out:
 318         mutex_unlock(&rcu_registry_lock);
 319         mutex_unlock(&rcu_gp_lock);
 320         urcu_wake_all_waiters(&waiters);
 321 gp_end:
 322         /*
 323          * Finish waiting for reader threads before letting the old ptr being
 324          * freed.
 325          */
 326         if (was_online)
 327                 rcu_thread_online();
 328         else
 329                 cmm_smp_mb();
 330 }
 331 #else /* !(CAA_BITS_PER_LONG < 64) */
 332 void synchronize_rcu(void)
 333 {
 334         CDS_LIST_HEAD(qsreaders);
 335         unsigned long was_online;
 336         DEFINE_URCU_WAIT_NODE(wait, URCU_WAIT_WAITING);
 337         struct urcu_waiters waiters;
 338
 339         was_online = rcu_read_ongoing();
 340
 341         /*
 342          * Mark the writer thread offline to make sure we don't wait for
 343          * our own quiescent state. This allows using synchronize_rcu()
 344          * in threads registered as readers.
 345          */
 346         if (was_online)
 347                 rcu_thread_offline();
 348         else
 349                 cmm_smp_mb();
 350
 351         /*
 352          * Add ourself to gp_waiters queue of threads awaiting to wait
 353          * for a grace period. Proceed to perform the grace period only
 354          * if we are the first thread added into the queue.
 355          */
 356         if (urcu_wait_add(&gp_waiters, &wait) != 0) {
 357                 /* Not first in queue: will be awakened by another thread. */
 358                 urcu_adaptative_busy_wait(&wait);
 359                 goto gp_end;
 360         }
 361         /* We won't need to wake ourself up */
 362         urcu_wait_set_state(&wait, URCU_WAIT_RUNNING);
 363
 364         mutex_lock(&rcu_gp_lock);
 365
 366         /*
 367          * Move all waiters into our local queue.
 368          */
 369         urcu_move_waiters(&waiters, &gp_waiters);
 370
 371         mutex_lock(&rcu_registry_lock);
 372
 373         if (cds_list_empty(&registry))
 374                 goto out;
 375
 376         /* Increment current G.P. */
 377         CMM_STORE_SHARED(rcu_gp.ctr, rcu_gp.ctr + RCU_GP_CTR);
 378
 379         /*
 380          * Must commit rcu_gp.ctr update to memory before waiting for
 381          * quiescent state. Failure to do so could result in the writer
 382          * waiting forever while new readers are always accessing data
 383          * (no progress). Enforce compiler-order of store to rcu_gp.ctr
 384          * before load URCU_TLS(rcu_reader).ctr.
 385          */
 386         cmm_barrier();
 387
 388         /*
 389          * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
 390          * model easier to understand. It does not have a big performance impact
 391          * anyway, given this is the write-side.
 392          */
 393         cmm_smp_mb();
 394
 395         /*
 396          * Wait for readers to observe new count of be quiescent.
 397          * wait_for_readers() can release and grab again rcu_registry_lock
 398          * interally.
 399          */
 400         wait_for_readers(&registry, NULL, &qsreaders);
 401
 402         /*
 403          * Put quiescent reader list back into registry.
 404          */
 405         cds_list_splice(&qsreaders, &registry);
 406 out:
 407         mutex_unlock(&rcu_registry_lock);
 408         mutex_unlock(&rcu_gp_lock);
 409         urcu_wake_all_waiters(&waiters);
 410 gp_end:
 411         if (was_online)
 412                 rcu_thread_online();
 413         else
 414                 cmm_smp_mb();
 415 }
 416 #endif  /* !(CAA_BITS_PER_LONG < 64) */
 417
 418 /*
 419  * library wrappers to be used by non-LGPL compatible source code.
 420  */
 421
 422 void rcu_read_lock(void)
 423 {
 424         _rcu_read_lock();
 425 }
 426
 427 void rcu_read_unlock(void)
 428 {
 429         _rcu_read_unlock();
 430 }
 431
 432 int rcu_read_ongoing(void)
 433 {
 434         return _rcu_read_ongoing();
 435 }
 436
 437 void rcu_quiescent_state(void)
 438 {
 439         _rcu_quiescent_state();
 440 }
 441
 442 void rcu_thread_offline(void)
 443 {
 444         _rcu_thread_offline();
 445 }
 446
 447 void rcu_thread_online(void)
 448 {
 449         _rcu_thread_online();
 450 }
 451
 452 void rcu_register_thread(void)
 453 {
 454         URCU_TLS(rcu_reader).tid = pthread_self();
 455         assert(URCU_TLS(rcu_reader).ctr == 0);
 456
 457         mutex_lock(&rcu_registry_lock);
 458         cds_list_add(&URCU_TLS(rcu_reader).node, &registry);
 459         mutex_unlock(&rcu_registry_lock);
 460         _rcu_thread_online();
 461 }
 462
 463 void rcu_unregister_thread(void)
 464 {
 465         /*
 466          * We have to make the thread offline otherwise we end up dealocking
 467          * with a waiting writer.
 468          */
 469         _rcu_thread_offline();
 470         mutex_lock(&rcu_registry_lock);
 471         cds_list_del(&URCU_TLS(rcu_reader).node);
 472         mutex_unlock(&rcu_registry_lock);
 473 }
 474
 475 void rcu_exit(void)
 476 {
 477         /*
 478          * Assertion disabled because call_rcu threads are now rcu
 479          * readers, and left running at exit.
 480          * assert(cds_list_empty(&registry));
 481          */
 482 }
 483
 484 DEFINE_RCU_FLAVOR(rcu_flavor);
 485
 486 #include "urcu-call-rcu-impl.h"
 487 #include "urcu-defer-impl.h"