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