urcu.c

   1 /*
   2  * urcu.c
   3  *
   4  * Userspace RCU 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 _BSD_SOURCE
  27 #define _GNU_SOURCE
  28 #define _LGPL_SOURCE
  29 #define _DEFAULT_SOURCE
  30 #include <stdio.h>
  31 #include <pthread.h>
  32 #include <signal.h>
  33 #include <assert.h>
  34 #include <stdlib.h>
  35 #include <stdint.h>
  36 #include <string.h>
  37 #include <errno.h>
  38 #include <poll.h>
  39
  40 #include "urcu/wfqueue.h"
  41 #include "urcu/map/urcu.h"
  42 #include "urcu/static/urcu.h"
  43 #include "urcu-pointer.h"
  44 #include "urcu/tls-compat.h"
  45
  46 #include "urcu-die.h"
  47
  48 /* Do not #define _LGPL_SOURCE to ensure we can emit the wrapper symbols */
  49 #undef _LGPL_SOURCE
  50 #include "urcu.h"
  51 #define _LGPL_SOURCE
  52
  53 /*
  54  * If a reader is really non-cooperative and refuses to commit its
  55  * rcu_active_readers count to memory (there is no barrier in the reader
  56  * per-se), kick it after 10 loops waiting for it.
  57  */
  58 #define KICK_READER_LOOPS       10
  59
  60 /*
  61  * Active attempts to check for reader Q.S. before calling futex().
  62  */
  63 #define RCU_QS_ACTIVE_ATTEMPTS 100
  64
  65 #ifdef RCU_MEMBARRIER
  66 static int init_done;
  67 int has_sys_membarrier;
  68
  69 void __attribute__((constructor)) rcu_init(void);
  70 #endif
  71
  72 #ifdef RCU_MB
  73 void rcu_init(void)
  74 {
  75 }
  76 #endif
  77
  78 #ifdef RCU_SIGNAL
  79 static int init_done;
  80
  81 void __attribute__((constructor)) rcu_init(void);
  82 void __attribute__((destructor)) rcu_exit(void);
  83 #endif
  84
  85 /*
  86  * rcu_gp_lock ensures mutual exclusion between threads calling
  87  * synchronize_rcu().
  88  */
  89 static pthread_mutex_t rcu_gp_lock = PTHREAD_MUTEX_INITIALIZER;
  90 /*
  91  * rcu_registry_lock ensures mutual exclusion between threads
  92  * registering and unregistering themselves to/from the registry, and
  93  * with threads reading that registry from synchronize_rcu(). However,
  94  * this lock is not held all the way through the completion of awaiting
  95  * for the grace period. It is sporadically released between iterations
  96  * on the registry.
  97  * rcu_registry_lock may nest inside rcu_gp_lock.
  98  */
  99 static pthread_mutex_t rcu_registry_lock = PTHREAD_MUTEX_INITIALIZER;
 100
 101 int32_t gp_futex;
 102
 103 /*
 104  * Global grace period counter.
 105  * Contains the current RCU_GP_CTR_PHASE.
 106  * Also has a RCU_GP_COUNT of 1, to accelerate the reader fast path.
 107  * Written to only by writer with mutex taken. Read by both writer and readers.
 108  */
 109 unsigned long rcu_gp_ctr = RCU_GP_COUNT;
 110 /*
 111  * Written to only by each individual reader. Read by both the reader and the
 112  * writers.
 113  */
 114 __DEFINE_URCU_TLS_GLOBAL(struct rcu_reader, rcu_reader);
 115
 116 #ifdef DEBUG_YIELD
 117 unsigned int yield_active;
 118 __DEFINE_URCU_TLS_GLOBAL(unsigned int, rand_yield);
 119 #endif
 120
 121 static CDS_LIST_HEAD(registry);
 122
 123 static void mutex_lock(pthread_mutex_t *mutex)
 124 {
 125         int ret;
 126
 127 #ifndef DISTRUST_SIGNALS_EXTREME
 128         ret = pthread_mutex_lock(mutex);
 129         if (ret)
 130                 urcu_die(ret);
 131 #else /* #ifndef DISTRUST_SIGNALS_EXTREME */
 132         while ((ret = pthread_mutex_trylock(mutex)) != 0) {
 133                 if (ret != EBUSY && ret != EINTR)
 134                         urcu_die(ret);
 135                 if (CMM_LOAD_SHARED(URCU_TLS(rcu_reader).need_mb)) {
 136                         cmm_smp_mb();
 137                         _CMM_STORE_SHARED(URCU_TLS(rcu_reader).need_mb, 0);
 138                         cmm_smp_mb();
 139                 }
 140                 poll(NULL,0,10);
 141         }
 142 #endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */
 143 }
 144
 145 static void mutex_unlock(pthread_mutex_t *mutex)
 146 {
 147         int ret;
 148
 149         ret = pthread_mutex_unlock(mutex);
 150         if (ret)
 151                 urcu_die(ret);
 152 }
 153
 154 #ifdef RCU_MEMBARRIER
 155 static void smp_mb_master(int group)
 156 {
 157         if (caa_likely(has_sys_membarrier))
 158                 membarrier(MEMBARRIER_EXPEDITED);
 159         else
 160                 cmm_smp_mb();
 161 }
 162 #endif
 163
 164 #ifdef RCU_MB
 165 static void smp_mb_master(int group)
 166 {
 167         cmm_smp_mb();
 168 }
 169 #endif
 170
 171 #ifdef RCU_SIGNAL
 172 static void force_mb_all_readers(void)
 173 {
 174         struct rcu_reader *index;
 175
 176         /*
 177          * Ask for each threads to execute a cmm_smp_mb() so we can consider the
 178          * compiler barriers around rcu read lock as real memory barriers.
 179          */
 180         if (cds_list_empty(&registry))
 181                 return;
 182         /*
 183          * pthread_kill has a cmm_smp_mb(). But beware, we assume it performs
 184          * a cache flush on architectures with non-coherent cache. Let's play
 185          * safe and don't assume anything : we use cmm_smp_mc() to make sure the
 186          * cache flush is enforced.
 187          */
 188         cds_list_for_each_entry(index, &registry, node) {
 189                 CMM_STORE_SHARED(index->need_mb, 1);
 190                 pthread_kill(index->tid, SIGRCU);
 191         }
 192         /*
 193          * Wait for sighandler (and thus mb()) to execute on every thread.
 194          *
 195          * Note that the pthread_kill() will never be executed on systems
 196          * that correctly deliver signals in a timely manner.  However, it
 197          * is not uncommon for kernels to have bugs that can result in
 198          * lost or unduly delayed signals.
 199          *
 200          * If you are seeing the below pthread_kill() executing much at
 201          * all, we suggest testing the underlying kernel and filing the
 202          * relevant bug report.  For Linux kernels, we recommend getting
 203          * the Linux Test Project (LTP).
 204          */
 205         cds_list_for_each_entry(index, &registry, node) {
 206                 while (CMM_LOAD_SHARED(index->need_mb)) {
 207                         pthread_kill(index->tid, SIGRCU);
 208                         poll(NULL, 0, 1);
 209                 }
 210         }
 211         cmm_smp_mb();   /* read ->need_mb before ending the barrier */
 212 }
 213
 214 static void smp_mb_master(int group)
 215 {
 216         force_mb_all_readers();
 217 }
 218 #endif /* #ifdef RCU_SIGNAL */
 219
 220 /*
 221  * synchronize_rcu() waiting. Single thread.
 222  * Always called with rcu_registry lock held. Releases this lock and
 223  * grabs it again. Holds the lock when it returns.
 224  */
 225 static void wait_gp(void)
 226 {
 227         /*
 228          * Read reader_gp before read futex. smp_mb_master() needs to
 229          * be called with the rcu registry lock held in RCU_SIGNAL
 230          * flavor.
 231          */
 232         smp_mb_master(RCU_MB_GROUP);
 233         /* Temporarily unlock the registry lock. */
 234         mutex_unlock(&rcu_registry_lock);
 235         if (uatomic_read(&gp_futex) != -1)
 236                 goto end;
 237         while (futex_async(&gp_futex, FUTEX_WAIT, -1,
 238                         NULL, NULL, 0)) {
 239                 switch (errno) {
 240                 case EWOULDBLOCK:
 241                         /* Value already changed. */
 242                         goto end;
 243                 case EINTR:
 244                         /* Retry if interrupted by signal. */
 245                         break;  /* Get out of switch. */
 246                 default:
 247                         /* Unexpected error. */
 248                         urcu_die(errno);
 249                 }
 250         }
 251 end:
 252         /*
 253          * Re-lock the registry lock before the next loop.
 254          */
 255         mutex_lock(&rcu_registry_lock);
 256 }
 257
 258 /*
 259  * Always called with rcu_registry lock held. Releases this lock between
 260  * iterations and grabs it again. Holds the lock when it returns.
 261  */
 262 void update_counter_and_wait(void)
 263 {
 264         CDS_LIST_HEAD(qsreaders);
 265         unsigned int wait_loops = 0;
 266         struct rcu_reader *index, *tmp;
 267 #ifdef HAS_INCOHERENT_CACHES
 268         unsigned int wait_gp_loops = 0;
 269 #endif /* HAS_INCOHERENT_CACHES */
 270
 271         /* Switch parity: 0 -> 1, 1 -> 0 */
 272         CMM_STORE_SHARED(rcu_gp_ctr, rcu_gp_ctr ^ RCU_GP_CTR_PHASE);
 273
 274         /*
 275          * Must commit rcu_gp_ctr update to memory before waiting for quiescent
 276          * state. Failure to do so could result in the writer waiting forever
 277          * while new readers are always accessing data (no progress). Enforce
 278          * compiler-order of store to rcu_gp_ctr before load rcu_reader ctr.
 279          */
 280         cmm_barrier();
 281
 282         /*
 283          *
 284          * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
 285          * model easier to understand. It does not have a big performance impact
 286          * anyway, given this is the write-side.
 287          */
 288         cmm_smp_mb();
 289
 290         /*
 291          * Wait for each thread URCU_TLS(rcu_reader).ctr count to become 0.
 292          */
 293         for (;;) {
 294                 if (wait_loops < RCU_QS_ACTIVE_ATTEMPTS)
 295                         wait_loops++;
 296                 if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS) {
 297                         uatomic_dec(&gp_futex);
 298                         /* Write futex before read reader_gp */
 299                         smp_mb_master(RCU_MB_GROUP);
 300                 }
 301
 302                 cds_list_for_each_entry_safe(index, tmp, &registry, node) {
 303                         if (!rcu_gp_ongoing(&index->ctr))
 304                                 cds_list_move(&index->node, &qsreaders);
 305                 }
 306
 307 #ifndef HAS_INCOHERENT_CACHES
 308                 if (cds_list_empty(&registry)) {
 309                         if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS) {
 310                                 /* Read reader_gp before write futex */
 311                                 smp_mb_master(RCU_MB_GROUP);
 312                                 uatomic_set(&gp_futex, 0);
 313                         }
 314                         break;
 315                 } else {
 316                         if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS) {
 317                                 /* wait_gp unlocks/locks registry lock. */
 318                                 wait_gp();
 319                         } else {
 320                                 /* Temporarily unlock the registry lock. */
 321                                 mutex_unlock(&rcu_registry_lock);
 322                                 caa_cpu_relax();
 323                                 /*
 324                                  * Re-lock the registry lock before the
 325                                  * next loop.
 326                                  */
 327                                 mutex_lock(&rcu_registry_lock);
 328                         }
 329                 }
 330 #else /* #ifndef HAS_INCOHERENT_CACHES */
 331                 /*
 332                  * BUSY-LOOP. Force the reader thread to commit its
 333                  * URCU_TLS(rcu_reader).ctr update to memory if we wait
 334                  * for too long.
 335                  */
 336                 if (cds_list_empty(&registry)) {
 337                         if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS) {
 338                                 /* Read reader_gp before write futex */
 339                                 smp_mb_master(RCU_MB_GROUP);
 340                                 uatomic_set(&gp_futex, 0);
 341                         }
 342                         break;
 343                 } else {
 344                         if (wait_gp_loops == KICK_READER_LOOPS) {
 345                                 smp_mb_master(RCU_MB_GROUP);
 346                                 wait_gp_loops = 0;
 347                         }
 348                         if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS) {
 349                                 /* wait_gp unlocks/locks registry lock. */
 350                                 wait_gp();
 351                                 wait_gp_loops++;
 352                         } else {
 353                                 /* Temporarily unlock the registry lock. */
 354                                 mutex_unlock(&rcu_registry_lock);
 355                                 caa_cpu_relax();
 356                                 /*
 357                                  * Re-lock the registry lock before the
 358                                  * next loop.
 359                                  */
 360                                 mutex_lock(&rcu_registry_lock);
 361                         }
 362                 }
 363 #endif /* #else #ifndef HAS_INCOHERENT_CACHES */
 364         }
 365         /* put back the reader list in the registry */
 366         cds_list_splice(&qsreaders, &registry);
 367 }
 368
 369 void synchronize_rcu(void)
 370 {
 371         mutex_lock(&rcu_gp_lock);
 372         mutex_lock(&rcu_registry_lock);
 373
 374         if (cds_list_empty(&registry))
 375                 goto out;
 376
 377         /*
 378          * All threads should read qparity before accessing data structure
 379          * where new ptr points to. Must be done within rcu_registry_lock
 380          * because it iterates on reader threads.
 381          */
 382         /* Write new ptr before changing the qparity */
 383         smp_mb_master(RCU_MB_GROUP);
 384
 385         /*
 386          * Wait for previous parity to be empty of readers.
 387          * update_counter_and_wait() can release and grab again
 388          * rcu_registry_lock interally.
 389          */
 390         update_counter_and_wait();      /* 0 -> 1, wait readers in parity 0 */
 391
 392         /*
 393          * Must finish waiting for quiescent state for parity 0 before
 394          * committing next rcu_gp_ctr update to memory. Failure to do so could
 395          * result in the writer waiting forever while new readers are always
 396          * accessing data (no progress).  Enforce compiler-order of load
 397          * URCU_TLS(rcu_reader).ctr before store to rcu_gp_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 previous parity to be empty of readers.
 410          * update_counter_and_wait() can release and grab again
 411          * rcu_registry_lock interally.
 412          */
 413         update_counter_and_wait();      /* 1 -> 0, wait readers in parity 1 */
 414
 415         /*
 416          * Finish waiting for reader threads before letting the old ptr
 417          * being freed. Must be done within rcu_registry_lock because it
 418          * iterates on reader threads.
 419          */
 420         smp_mb_master(RCU_MB_GROUP);
 421 out:
 422         mutex_unlock(&rcu_registry_lock);
 423         mutex_unlock(&rcu_gp_lock);
 424 }
 425
 426 /*
 427  * library wrappers to be used by non-LGPL compatible source code.
 428  */
 429
 430 void rcu_read_lock(void)
 431 {
 432         _rcu_read_lock();
 433 }
 434
 435 void rcu_read_unlock(void)
 436 {
 437         _rcu_read_unlock();
 438 }
 439
 440 void rcu_register_thread(void)
 441 {
 442         URCU_TLS(rcu_reader).tid = pthread_self();
 443         assert(URCU_TLS(rcu_reader).need_mb == 0);
 444         assert(!(URCU_TLS(rcu_reader).ctr & RCU_GP_CTR_NEST_MASK));
 445
 446         mutex_lock(&rcu_registry_lock);
 447         rcu_init();     /* In case gcc does not support constructor attribute */
 448         cds_list_add(&URCU_TLS(rcu_reader).node, &registry);
 449         mutex_unlock(&rcu_registry_lock);
 450 }
 451
 452 void rcu_unregister_thread(void)
 453 {
 454         mutex_lock(&rcu_registry_lock);
 455         cds_list_del(&URCU_TLS(rcu_reader).node);
 456         mutex_unlock(&rcu_registry_lock);
 457 }
 458
 459 #ifdef RCU_MEMBARRIER
 460 void rcu_init(void)
 461 {
 462         if (init_done)
 463                 return;
 464         init_done = 1;
 465         if (!membarrier(MEMBARRIER_EXPEDITED | MEMBARRIER_QUERY))
 466                 has_sys_membarrier = 1;
 467 }
 468 #endif
 469
 470 #ifdef RCU_SIGNAL
 471 static void sigrcu_handler(int signo, siginfo_t *siginfo, void *context)
 472 {
 473         /*
 474          * Executing this cmm_smp_mb() is the only purpose of this signal handler.
 475          * It punctually promotes cmm_barrier() into cmm_smp_mb() on every thread it is
 476          * executed on.
 477          */
 478         cmm_smp_mb();
 479         _CMM_STORE_SHARED(URCU_TLS(rcu_reader).need_mb, 0);
 480         cmm_smp_mb();
 481 }
 482
 483 /*
 484  * rcu_init constructor. Called when the library is linked, but also when
 485  * reader threads are calling rcu_register_thread().
 486  * Should only be called by a single thread at a given time. This is ensured by
 487  * holing the rcu_registry_lock from rcu_register_thread() or by running
 488  * at library load time, which should not be executed by multiple
 489  * threads nor concurrently with rcu_register_thread() anyway.
 490  */
 491 void rcu_init(void)
 492 {
 493         struct sigaction act;
 494         int ret;
 495
 496         if (init_done)
 497                 return;
 498         init_done = 1;
 499
 500         act.sa_sigaction = sigrcu_handler;
 501         act.sa_flags = SA_SIGINFO | SA_RESTART;
 502         sigemptyset(&act.sa_mask);
 503         ret = sigaction(SIGRCU, &act, NULL);
 504         if (ret)
 505                 urcu_die(errno);
 506 }
 507
 508 void rcu_exit(void)
 509 {
 510         struct sigaction act;
 511         int ret;
 512
 513         ret = sigaction(SIGRCU, NULL, &act);
 514         if (ret)
 515                 urcu_die(errno);
 516         assert(act.sa_sigaction == sigrcu_handler);
 517         assert(cds_list_empty(&registry));
 518 }
 519
 520 #endif /* #ifdef RCU_SIGNAL */
 521
 522 DEFINE_RCU_FLAVOR(rcu_flavor);
 523
 524 #include "urcu-call-rcu-impl.h"
 525 #include "urcu-defer-impl.h"