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                 return;
 126         while (futex_noasync(&rcu_gp.futex, FUTEX_WAIT, -1,
 127                         NULL, NULL, 0)) {
 128                 switch (errno) {
 129                 case EWOULDBLOCK:
 130                         /* Value already changed. */
 131                         return;
 132                 case EINTR:
 133                         /* Retry if interrupted by signal. */
 134                         break;  /* Get out of switch. */
 135                 default:
 136                         /* Unexpected error. */
 137                         urcu_die(errno);
 138                 }
 139         }
 140 }
 141
 142 /*
 143  * Always called with rcu_registry lock held. Releases this lock between
 144  * iterations and grabs it again. Holds the lock when it returns.
 145  */
 146 static void wait_for_readers(struct cds_list_head *input_readers,
 147                         struct cds_list_head *cur_snap_readers,
 148                         struct cds_list_head *qsreaders)
 149 {
 150         unsigned int wait_loops = 0;
 151         struct rcu_reader *index, *tmp;
 152
 153         /*
 154          * Wait for each thread URCU_TLS(rcu_reader).ctr to either
 155          * indicate quiescence (offline), or for them to observe the
 156          * current rcu_gp.ctr value.
 157          */
 158         for (;;) {
 159                 if (wait_loops < RCU_QS_ACTIVE_ATTEMPTS)
 160                         wait_loops++;
 161                 if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS) {
 162                         uatomic_set(&rcu_gp.futex, -1);
 163                         /*
 164                          * Write futex before write waiting (the other side
 165                          * reads them in the opposite order).
 166                          */
 167                         cmm_smp_wmb();
 168                         cds_list_for_each_entry(index, input_readers, node) {
 169                                 _CMM_STORE_SHARED(index->waiting, 1);
 170                         }
 171                         /* Write futex before read reader_gp */
 172                         cmm_smp_mb();
 173                 }
 174                 cds_list_for_each_entry_safe(index, tmp, input_readers, node) {
 175                         switch (rcu_reader_state(&index->ctr)) {
 176                         case RCU_READER_ACTIVE_CURRENT:
 177                                 if (cur_snap_readers) {
 178                                         cds_list_move(&index->node,
 179                                                 cur_snap_readers);
 180                                         break;
 181                                 }
 182                                 /* Fall-through */
 183                         case RCU_READER_INACTIVE:
 184                                 cds_list_move(&index->node, qsreaders);
 185                                 break;
 186                         case RCU_READER_ACTIVE_OLD:
 187                                 /*
 188                                  * Old snapshot. Leaving node in
 189                                  * input_readers will make us busy-loop
 190                                  * until the snapshot becomes current or
 191                                  * the reader becomes inactive.
 192                                  */
 193                                 break;
 194                         }
 195                 }
 196
 197                 if (cds_list_empty(input_readers)) {
 198                         if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS) {
 199                                 /* Read reader_gp before write futex */
 200                                 cmm_smp_mb();
 201                                 uatomic_set(&rcu_gp.futex, 0);
 202                         }
 203                         break;
 204                 } else {
 205                         /* Temporarily unlock the registry lock. */
 206                         mutex_unlock(&rcu_registry_lock);
 207                         if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS) {
 208                                 wait_gp();
 209                         } else {
 210 #ifndef HAS_INCOHERENT_CACHES
 211                                 caa_cpu_relax();
 212 #else /* #ifndef HAS_INCOHERENT_CACHES */
 213                                 cmm_smp_mb();
 214 #endif /* #else #ifndef HAS_INCOHERENT_CACHES */
 215                         }
 216                         /* Re-lock the registry lock before the next loop. */
 217                         mutex_lock(&rcu_registry_lock);
 218                 }
 219         }
 220 }
 221
 222 /*
 223  * Using a two-subphases algorithm for architectures with smaller than 64-bit
 224  * long-size to ensure we do not encounter an overflow bug.
 225  */
 226
 227 #if (CAA_BITS_PER_LONG < 64)
 228 void synchronize_rcu(void)
 229 {
 230         CDS_LIST_HEAD(cur_snap_readers);
 231         CDS_LIST_HEAD(qsreaders);
 232         unsigned long was_online;
 233         DEFINE_URCU_WAIT_NODE(wait, URCU_WAIT_WAITING);
 234         struct urcu_waiters waiters;
 235
 236         was_online = rcu_read_ongoing();
 237
 238         /* All threads should read qparity before accessing data structure
 239          * where new ptr points to.  In the "then" case, rcu_thread_offline
 240          * includes a memory barrier.
 241          *
 242          * Mark the writer thread offline to make sure we don't wait for
 243          * our own quiescent state. This allows using synchronize_rcu()
 244          * in threads registered as readers.
 245          */
 246         if (was_online)
 247                 rcu_thread_offline();
 248         else
 249                 cmm_smp_mb();
 250
 251         /*
 252          * Add ourself to gp_waiters queue of threads awaiting to wait
 253          * for a grace period. Proceed to perform the grace period only
 254          * if we are the first thread added into the queue.
 255          */
 256         if (urcu_wait_add(&gp_waiters, &wait) != 0) {
 257                 /* Not first in queue: will be awakened by another thread. */
 258                 urcu_adaptative_busy_wait(&wait);
 259                 goto gp_end;
 260         }
 261         /* We won't need to wake ourself up */
 262         urcu_wait_set_state(&wait, URCU_WAIT_RUNNING);
 263
 264         mutex_lock(&rcu_gp_lock);
 265
 266         /*
 267          * Move all waiters into our local queue.
 268          */
 269         urcu_move_waiters(&waiters, &gp_waiters);
 270
 271         mutex_lock(&rcu_registry_lock);
 272
 273         if (cds_list_empty(&registry))
 274                 goto out;
 275
 276         /*
 277          * Wait for readers to observe original parity or be quiescent.
 278          * wait_for_readers() can release and grab again rcu_registry_lock
 279          * interally.
 280          */
 281         wait_for_readers(&registry, &cur_snap_readers, &qsreaders);
 282
 283         /*
 284          * Must finish waiting for quiescent state for original parity
 285          * before committing next rcu_gp.ctr update to memory. Failure
 286          * to do so could result in the writer waiting forever while new
 287          * readers are always accessing data (no progress).  Enforce
 288          * compiler-order of load URCU_TLS(rcu_reader).ctr before store
 289          * to rcu_gp.ctr.
 290          */
 291         cmm_barrier();
 292
 293         /*
 294          * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
 295          * model easier to understand. It does not have a big performance impact
 296          * anyway, given this is the write-side.
 297          */
 298         cmm_smp_mb();
 299
 300         /* Switch parity: 0 -> 1, 1 -> 0 */
 301         CMM_STORE_SHARED(rcu_gp.ctr, rcu_gp.ctr ^ RCU_GP_CTR);
 302
 303         /*
 304          * Must commit rcu_gp.ctr update to memory before waiting for
 305          * quiescent state. Failure to do so could result in the writer
 306          * waiting forever while new readers are always accessing data
 307          * (no progress). Enforce compiler-order of store to rcu_gp.ctr
 308          * before load URCU_TLS(rcu_reader).ctr.
 309          */
 310         cmm_barrier();
 311
 312         /*
 313          * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
 314          * model easier to understand. It does not have a big performance impact
 315          * anyway, given this is the write-side.
 316          */
 317         cmm_smp_mb();
 318
 319         /*
 320          * Wait for readers to observe new parity or be quiescent.
 321          * wait_for_readers() can release and grab again rcu_registry_lock
 322          * interally.
 323          */
 324         wait_for_readers(&cur_snap_readers, NULL, &qsreaders);
 325
 326         /*
 327          * Put quiescent reader list back into registry.
 328          */
 329         cds_list_splice(&qsreaders, &registry);
 330 out:
 331         mutex_unlock(&rcu_registry_lock);
 332         mutex_unlock(&rcu_gp_lock);
 333         urcu_wake_all_waiters(&waiters);
 334 gp_end:
 335         /*
 336          * Finish waiting for reader threads before letting the old ptr being
 337          * freed.
 338          */
 339         if (was_online)
 340                 rcu_thread_online();
 341         else
 342                 cmm_smp_mb();
 343 }
 344 #else /* !(CAA_BITS_PER_LONG < 64) */
 345 void synchronize_rcu(void)
 346 {
 347         CDS_LIST_HEAD(qsreaders);
 348         unsigned long was_online;
 349         DEFINE_URCU_WAIT_NODE(wait, URCU_WAIT_WAITING);
 350         struct urcu_waiters waiters;
 351
 352         was_online = rcu_read_ongoing();
 353
 354         /*
 355          * Mark the writer thread offline to make sure we don't wait for
 356          * our own quiescent state. This allows using synchronize_rcu()
 357          * in threads registered as readers.
 358          */
 359         if (was_online)
 360                 rcu_thread_offline();
 361         else
 362                 cmm_smp_mb();
 363
 364         /*
 365          * Add ourself to gp_waiters queue of threads awaiting to wait
 366          * for a grace period. Proceed to perform the grace period only
 367          * if we are the first thread added into the queue.
 368          */
 369         if (urcu_wait_add(&gp_waiters, &wait) != 0) {
 370                 /* Not first in queue: will be awakened by another thread. */
 371                 urcu_adaptative_busy_wait(&wait);
 372                 goto gp_end;
 373         }
 374         /* We won't need to wake ourself up */
 375         urcu_wait_set_state(&wait, URCU_WAIT_RUNNING);
 376
 377         mutex_lock(&rcu_gp_lock);
 378
 379         /*
 380          * Move all waiters into our local queue.
 381          */
 382         urcu_move_waiters(&waiters, &gp_waiters);
 383
 384         mutex_lock(&rcu_registry_lock);
 385
 386         if (cds_list_empty(&registry))
 387                 goto out;
 388
 389         /* Increment current G.P. */
 390         CMM_STORE_SHARED(rcu_gp.ctr, rcu_gp.ctr + RCU_GP_CTR);
 391
 392         /*
 393          * Must commit rcu_gp.ctr update to memory before waiting for
 394          * quiescent state. Failure to do so could result in the writer
 395          * waiting forever while new readers are always accessing data
 396          * (no progress). Enforce compiler-order of store to rcu_gp.ctr
 397          * before load URCU_TLS(rcu_reader).ctr.
 398          */
 399         cmm_barrier();
 400
 401         /*
 402          * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
 403          * model easier to understand. It does not have a big performance impact
 404          * anyway, given this is the write-side.
 405          */
 406         cmm_smp_mb();
 407
 408         /*
 409          * Wait for readers to observe new count of be quiescent.
 410          * wait_for_readers() can release and grab again rcu_registry_lock
 411          * interally.
 412          */
 413         wait_for_readers(&registry, NULL, &qsreaders);
 414
 415         /*
 416          * Put quiescent reader list back into registry.
 417          */
 418         cds_list_splice(&qsreaders, &registry);
 419 out:
 420         mutex_unlock(&rcu_registry_lock);
 421         mutex_unlock(&rcu_gp_lock);
 422         urcu_wake_all_waiters(&waiters);
 423 gp_end:
 424         if (was_online)
 425                 rcu_thread_online();
 426         else
 427                 cmm_smp_mb();
 428 }
 429 #endif  /* !(CAA_BITS_PER_LONG < 64) */
 430
 431 /*
 432  * library wrappers to be used by non-LGPL compatible source code.
 433  */
 434
 435 void rcu_read_lock(void)
 436 {
 437         _rcu_read_lock();
 438 }
 439
 440 void rcu_read_unlock(void)
 441 {
 442         _rcu_read_unlock();
 443 }
 444
 445 int rcu_read_ongoing(void)
 446 {
 447         return _rcu_read_ongoing();
 448 }
 449
 450 void rcu_quiescent_state(void)
 451 {
 452         _rcu_quiescent_state();
 453 }
 454
 455 void rcu_thread_offline(void)
 456 {
 457         _rcu_thread_offline();
 458 }
 459
 460 void rcu_thread_online(void)
 461 {
 462         _rcu_thread_online();
 463 }
 464
 465 void rcu_register_thread(void)
 466 {
 467         URCU_TLS(rcu_reader).tid = pthread_self();
 468         assert(URCU_TLS(rcu_reader).ctr == 0);
 469
 470         mutex_lock(&rcu_registry_lock);
 471         assert(!URCU_TLS(rcu_reader).registered);
 472         URCU_TLS(rcu_reader).registered = 1;
 473         cds_list_add(&URCU_TLS(rcu_reader).node, &registry);
 474         mutex_unlock(&rcu_registry_lock);
 475         _rcu_thread_online();
 476 }
 477
 478 void rcu_unregister_thread(void)
 479 {
 480         /*
 481          * We have to make the thread offline otherwise we end up dealocking
 482          * with a waiting writer.
 483          */
 484         _rcu_thread_offline();
 485         assert(URCU_TLS(rcu_reader).registered);
 486         URCU_TLS(rcu_reader).registered = 0;
 487         mutex_lock(&rcu_registry_lock);
 488         cds_list_del(&URCU_TLS(rcu_reader).node);
 489         mutex_unlock(&rcu_registry_lock);
 490 }
 491
 492 void rcu_exit(void)
 493 {
 494         /*
 495          * Assertion disabled because call_rcu threads are now rcu
 496          * readers, and left running at exit.
 497          * assert(cds_list_empty(&registry));
 498          */
 499 }
 500
 501 DEFINE_RCU_FLAVOR(rcu_flavor);
 502
 503 #include "urcu-call-rcu-impl.h"
 504 #include "urcu-defer-impl.h"