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_GLOBAL(struct rcu_reader, rcu_reader);
  82
  83 #ifdef DEBUG_YIELD
  84 unsigned int rcu_yield_active;
  85 __DEFINE_URCU_TLS_GLOBAL(unsigned int, rcu_rand_yield);
  86 #endif
  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         if (uatomic_read(&rcu_gp.futex) != -1)
 130                 return;
 131         while (futex_noasync(&rcu_gp.futex, FUTEX_WAIT, -1,
 132                         NULL, NULL, 0)) {
 133                 switch (errno) {
 134                 case EWOULDBLOCK:
 135                         /* Value already changed. */
 136                         return;
 137                 case EINTR:
 138                         /* Retry if interrupted by signal. */
 139                         break;  /* Get out of switch. */
 140                 default:
 141                         /* Unexpected error. */
 142                         urcu_die(errno);
 143                 }
 144         }
 145 }
 146
 147 /*
 148  * Always called with rcu_registry lock held. Releases this lock between
 149  * iterations and grabs it again. Holds the lock when it returns.
 150  */
 151 static void wait_for_readers(struct cds_list_head *input_readers,
 152                         struct cds_list_head *cur_snap_readers,
 153                         struct cds_list_head *qsreaders)
 154 {
 155         unsigned int wait_loops = 0;
 156         struct rcu_reader *index, *tmp;
 157
 158         /*
 159          * Wait for each thread URCU_TLS(rcu_reader).ctr to either
 160          * indicate quiescence (offline), or for them to observe the
 161          * current rcu_gp.ctr value.
 162          */
 163         for (;;) {
 164                 if (wait_loops < RCU_QS_ACTIVE_ATTEMPTS)
 165                         wait_loops++;
 166                 if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS) {
 167                         uatomic_set(&rcu_gp.futex, -1);
 168                         /*
 169                          * Write futex before write waiting (the other side
 170                          * reads them in the opposite order).
 171                          */
 172                         cmm_smp_wmb();
 173                         cds_list_for_each_entry(index, input_readers, node) {
 174                                 _CMM_STORE_SHARED(index->waiting, 1);
 175                         }
 176                         /* Write futex before read reader_gp */
 177                         cmm_smp_mb();
 178                 }
 179                 cds_list_for_each_entry_safe(index, tmp, input_readers, node) {
 180                         switch (rcu_reader_state(&index->ctr)) {
 181                         case RCU_READER_ACTIVE_CURRENT:
 182                                 if (cur_snap_readers) {
 183                                         cds_list_move(&index->node,
 184                                                 cur_snap_readers);
 185                                         break;
 186                                 }
 187                                 /* Fall-through */
 188                         case RCU_READER_INACTIVE:
 189                                 cds_list_move(&index->node, qsreaders);
 190                                 break;
 191                         case RCU_READER_ACTIVE_OLD:
 192                                 /*
 193                                  * Old snapshot. Leaving node in
 194                                  * input_readers will make us busy-loop
 195                                  * until the snapshot becomes current or
 196                                  * the reader becomes inactive.
 197                                  */
 198                                 break;
 199                         }
 200                 }
 201
 202                 if (cds_list_empty(input_readers)) {
 203                         if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS) {
 204                                 /* Read reader_gp before write futex */
 205                                 cmm_smp_mb();
 206                                 uatomic_set(&rcu_gp.futex, 0);
 207                         }
 208                         break;
 209                 } else {
 210                         /* Temporarily unlock the registry lock. */
 211                         mutex_unlock(&rcu_registry_lock);
 212                         if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS) {
 213                                 wait_gp();
 214                         } else {
 215 #ifndef HAS_INCOHERENT_CACHES
 216                                 caa_cpu_relax();
 217 #else /* #ifndef HAS_INCOHERENT_CACHES */
 218                                 cmm_smp_mb();
 219 #endif /* #else #ifndef HAS_INCOHERENT_CACHES */
 220                         }
 221                         /* Re-lock the registry lock before the next loop. */
 222                         mutex_lock(&rcu_registry_lock);
 223                 }
 224         }
 225 }
 226
 227 /*
 228  * Using a two-subphases algorithm for architectures with smaller than 64-bit
 229  * long-size to ensure we do not encounter an overflow bug.
 230  */
 231
 232 #if (CAA_BITS_PER_LONG < 64)
 233 void synchronize_rcu(void)
 234 {
 235         CDS_LIST_HEAD(cur_snap_readers);
 236         CDS_LIST_HEAD(qsreaders);
 237         unsigned long was_online;
 238         DEFINE_URCU_WAIT_NODE(wait, URCU_WAIT_WAITING);
 239         struct urcu_waiters waiters;
 240
 241         was_online = rcu_read_ongoing();
 242
 243         /* All threads should read qparity before accessing data structure
 244          * where new ptr points to.  In the "then" case, rcu_thread_offline
 245          * includes a memory barrier.
 246          *
 247          * Mark the writer thread offline to make sure we don't wait for
 248          * our own quiescent state. This allows using synchronize_rcu()
 249          * in threads registered as readers.
 250          */
 251         if (was_online)
 252                 rcu_thread_offline();
 253         else
 254                 cmm_smp_mb();
 255
 256         /*
 257          * Add ourself to gp_waiters queue of threads awaiting to wait
 258          * for a grace period. Proceed to perform the grace period only
 259          * if we are the first thread added into the queue.
 260          */
 261         if (urcu_wait_add(&gp_waiters, &wait) != 0) {
 262                 /* Not first in queue: will be awakened by another thread. */
 263                 urcu_adaptative_busy_wait(&wait);
 264                 goto gp_end;
 265         }
 266         /* We won't need to wake ourself up */
 267         urcu_wait_set_state(&wait, URCU_WAIT_RUNNING);
 268
 269         mutex_lock(&rcu_gp_lock);
 270
 271         /*
 272          * Move all waiters into our local queue.
 273          */
 274         urcu_move_waiters(&waiters, &gp_waiters);
 275
 276         mutex_lock(&rcu_registry_lock);
 277
 278         if (cds_list_empty(&registry))
 279                 goto out;
 280
 281         /*
 282          * Wait for readers to observe original parity or be quiescent.
 283          * wait_for_readers() can release and grab again rcu_registry_lock
 284          * interally.
 285          */
 286         wait_for_readers(&registry, &cur_snap_readers, &qsreaders);
 287
 288         /*
 289          * Must finish waiting for quiescent state for original parity
 290          * before committing next rcu_gp.ctr update to memory. Failure
 291          * to do so could result in the writer waiting forever while new
 292          * readers are always accessing data (no progress).  Enforce
 293          * compiler-order of load URCU_TLS(rcu_reader).ctr before store
 294          * to rcu_gp.ctr.
 295          */
 296         cmm_barrier();
 297
 298         /*
 299          * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
 300          * model easier to understand. It does not have a big performance impact
 301          * anyway, given this is the write-side.
 302          */
 303         cmm_smp_mb();
 304
 305         /* Switch parity: 0 -> 1, 1 -> 0 */
 306         CMM_STORE_SHARED(rcu_gp.ctr, rcu_gp.ctr ^ RCU_GP_CTR);
 307
 308         /*
 309          * Must commit rcu_gp.ctr update to memory before waiting for
 310          * quiescent state. Failure to do so could result in the writer
 311          * waiting forever while new readers are always accessing data
 312          * (no progress). Enforce compiler-order of store to rcu_gp.ctr
 313          * before load URCU_TLS(rcu_reader).ctr.
 314          */
 315         cmm_barrier();
 316
 317         /*
 318          * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
 319          * model easier to understand. It does not have a big performance impact
 320          * anyway, given this is the write-side.
 321          */
 322         cmm_smp_mb();
 323
 324         /*
 325          * Wait for readers to observe new parity or be quiescent.
 326          * wait_for_readers() can release and grab again rcu_registry_lock
 327          * interally.
 328          */
 329         wait_for_readers(&cur_snap_readers, NULL, &qsreaders);
 330
 331         /*
 332          * Put quiescent reader list back into registry.
 333          */
 334         cds_list_splice(&qsreaders, &registry);
 335 out:
 336         mutex_unlock(&rcu_registry_lock);
 337         mutex_unlock(&rcu_gp_lock);
 338         urcu_wake_all_waiters(&waiters);
 339 gp_end:
 340         /*
 341          * Finish waiting for reader threads before letting the old ptr being
 342          * freed.
 343          */
 344         if (was_online)
 345                 rcu_thread_online();
 346         else
 347                 cmm_smp_mb();
 348 }
 349 #else /* !(CAA_BITS_PER_LONG < 64) */
 350 void synchronize_rcu(void)
 351 {
 352         CDS_LIST_HEAD(qsreaders);
 353         unsigned long was_online;
 354         DEFINE_URCU_WAIT_NODE(wait, URCU_WAIT_WAITING);
 355         struct urcu_waiters waiters;
 356
 357         was_online = rcu_read_ongoing();
 358
 359         /*
 360          * Mark the writer thread offline to make sure we don't wait for
 361          * our own quiescent state. This allows using synchronize_rcu()
 362          * in threads registered as readers.
 363          */
 364         if (was_online)
 365                 rcu_thread_offline();
 366         else
 367                 cmm_smp_mb();
 368
 369         /*
 370          * Add ourself to gp_waiters queue of threads awaiting to wait
 371          * for a grace period. Proceed to perform the grace period only
 372          * if we are the first thread added into the queue.
 373          */
 374         if (urcu_wait_add(&gp_waiters, &wait) != 0) {
 375                 /* Not first in queue: will be awakened by another thread. */
 376                 urcu_adaptative_busy_wait(&wait);
 377                 goto gp_end;
 378         }
 379         /* We won't need to wake ourself up */
 380         urcu_wait_set_state(&wait, URCU_WAIT_RUNNING);
 381
 382         mutex_lock(&rcu_gp_lock);
 383
 384         /*
 385          * Move all waiters into our local queue.
 386          */
 387         urcu_move_waiters(&waiters, &gp_waiters);
 388
 389         mutex_lock(&rcu_registry_lock);
 390
 391         if (cds_list_empty(&registry))
 392                 goto out;
 393
 394         /* Increment current G.P. */
 395         CMM_STORE_SHARED(rcu_gp.ctr, rcu_gp.ctr + RCU_GP_CTR);
 396
 397         /*
 398          * Must commit rcu_gp.ctr update to memory before waiting for
 399          * quiescent state. Failure to do so could result in the writer
 400          * waiting forever while new readers are always accessing data
 401          * (no progress). Enforce compiler-order of store to rcu_gp.ctr
 402          * before load URCU_TLS(rcu_reader).ctr.
 403          */
 404         cmm_barrier();
 405
 406         /*
 407          * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
 408          * model easier to understand. It does not have a big performance impact
 409          * anyway, given this is the write-side.
 410          */
 411         cmm_smp_mb();
 412
 413         /*
 414          * Wait for readers to observe new count of be quiescent.
 415          * wait_for_readers() can release and grab again rcu_registry_lock
 416          * interally.
 417          */
 418         wait_for_readers(&registry, NULL, &qsreaders);
 419
 420         /*
 421          * Put quiescent reader list back into registry.
 422          */
 423         cds_list_splice(&qsreaders, &registry);
 424 out:
 425         mutex_unlock(&rcu_registry_lock);
 426         mutex_unlock(&rcu_gp_lock);
 427         urcu_wake_all_waiters(&waiters);
 428 gp_end:
 429         if (was_online)
 430                 rcu_thread_online();
 431         else
 432                 cmm_smp_mb();
 433 }
 434 #endif  /* !(CAA_BITS_PER_LONG < 64) */
 435
 436 /*
 437  * library wrappers to be used by non-LGPL compatible source code.
 438  */
 439
 440 void rcu_read_lock(void)
 441 {
 442         _rcu_read_lock();
 443 }
 444
 445 void rcu_read_unlock(void)
 446 {
 447         _rcu_read_unlock();
 448 }
 449
 450 int rcu_read_ongoing(void)
 451 {
 452         return _rcu_read_ongoing();
 453 }
 454
 455 void rcu_quiescent_state(void)
 456 {
 457         _rcu_quiescent_state();
 458 }
 459
 460 void rcu_thread_offline(void)
 461 {
 462         _rcu_thread_offline();
 463 }
 464
 465 void rcu_thread_online(void)
 466 {
 467         _rcu_thread_online();
 468 }
 469
 470 void rcu_register_thread(void)
 471 {
 472         URCU_TLS(rcu_reader).tid = pthread_self();
 473         assert(URCU_TLS(rcu_reader).ctr == 0);
 474
 475         mutex_lock(&rcu_registry_lock);
 476         cds_list_add(&URCU_TLS(rcu_reader).node, &registry);
 477         mutex_unlock(&rcu_registry_lock);
 478         _rcu_thread_online();
 479 }
 480
 481 void rcu_unregister_thread(void)
 482 {
 483         /*
 484          * We have to make the thread offline otherwise we end up dealocking
 485          * with a waiting writer.
 486          */
 487         _rcu_thread_offline();
 488         mutex_lock(&rcu_registry_lock);
 489         cds_list_del(&URCU_TLS(rcu_reader).node);
 490         mutex_unlock(&rcu_registry_lock);
 491 }
 492
 493 void rcu_exit(void)
 494 {
 495         /*
 496          * Assertion disabled because call_rcu threads are now rcu
 497          * readers, and left running at exit.
 498          * assert(cds_list_empty(&registry));
 499          */
 500 }
 501
 502 DEFINE_RCU_FLAVOR(rcu_flavor);
 503
 504 #include "urcu-call-rcu-impl.h"
 505 #include "urcu-defer-impl.h"