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 #include <stdio.h>
  30 #include <pthread.h>
  31 #include <signal.h>
  32 #include <assert.h>
  33 #include <stdlib.h>
  34 #include <stdint.h>
  35 #include <string.h>
  36 #include <errno.h>
  37 #include <poll.h>
  38
  39 #include "urcu/wfqueue.h"
  40 #include "urcu/map/urcu.h"
  41 #include "urcu/static/urcu.h"
  42
  43 /* Do not #define _LGPL_SOURCE to ensure we can emit the wrapper symbols */
  44 #undef _LGPL_SOURCE
  45 #include "urcu.h"
  46 #define _LGPL_SOURCE
  47
  48 /*
  49  * If a reader is really non-cooperative and refuses to commit its
  50  * rcu_active_readers count to memory (there is no barrier in the reader
  51  * per-se), kick it after a few loops waiting for it.
  52  */
  53 #define KICK_READER_LOOPS 10000
  54
  55 /*
  56  * Active attempts to check for reader Q.S. before calling futex().
  57  */
  58 #define RCU_QS_ACTIVE_ATTEMPTS 100
  59
  60 #ifdef RCU_MEMBARRIER
  61 static int init_done;
  62 int has_sys_membarrier;
  63
  64 void __attribute__((constructor)) rcu_init(void);
  65 #endif
  66
  67 #ifdef RCU_MB
  68 void rcu_init(void)
  69 {
  70 }
  71 #endif
  72
  73 #ifdef RCU_SIGNAL
  74 static int init_done;
  75
  76 void __attribute__((constructor)) rcu_init(void);
  77 void __attribute__((destructor)) rcu_exit(void);
  78 #endif
  79
  80 static pthread_mutex_t rcu_gp_lock = PTHREAD_MUTEX_INITIALIZER;
  81
  82 int32_t gp_futex;
  83
  84 /*
  85  * Global grace period counter.
  86  * Contains the current RCU_GP_CTR_PHASE.
  87  * Also has a RCU_GP_COUNT of 1, to accelerate the reader fast path.
  88  * Written to only by writer with mutex taken. Read by both writer and readers.
  89  */
  90 unsigned long rcu_gp_ctr = RCU_GP_COUNT;
  91
  92 /*
  93  * Written to only by each individual reader. Read by both the reader and the
  94  * writers.
  95  */
  96 struct rcu_reader __thread rcu_reader;
  97
  98 #ifdef DEBUG_YIELD
  99 unsigned int yield_active;
 100 unsigned int __thread rand_yield;
 101 #endif
 102
 103 static CDS_LIST_HEAD(registry);
 104
 105 static void mutex_lock(pthread_mutex_t *mutex)
 106 {
 107         int ret;
 108
 109 #ifndef DISTRUST_SIGNALS_EXTREME
 110         ret = pthread_mutex_lock(mutex);
 111         if (ret) {
 112                 perror("Error in pthread mutex lock");
 113                 exit(-1);
 114         }
 115 #else /* #ifndef DISTRUST_SIGNALS_EXTREME */
 116         while ((ret = pthread_mutex_trylock(mutex)) != 0) {
 117                 if (ret != EBUSY && ret != EINTR) {
 118                         printf("ret = %d, errno = %d\n", ret, errno);
 119                         perror("Error in pthread mutex lock");
 120                         exit(-1);
 121                 }
 122                 if (CMM_LOAD_SHARED(rcu_reader.need_mb)) {
 123                         cmm_smp_mb();
 124                         _CMM_STORE_SHARED(rcu_reader.need_mb, 0);
 125                         cmm_smp_mb();
 126                 }
 127                 poll(NULL,0,10);
 128         }
 129 #endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */
 130 }
 131
 132 static void mutex_unlock(pthread_mutex_t *mutex)
 133 {
 134         int ret;
 135
 136         ret = pthread_mutex_unlock(mutex);
 137         if (ret) {
 138                 perror("Error in pthread mutex unlock");
 139                 exit(-1);
 140         }
 141 }
 142
 143 #ifdef RCU_MEMBARRIER
 144 static void smp_mb_master(int group)
 145 {
 146         if (likely(has_sys_membarrier))
 147                 membarrier(MEMBARRIER_EXPEDITED);
 148         else
 149                 cmm_smp_mb();
 150 }
 151 #endif
 152
 153 #ifdef RCU_MB
 154 static void smp_mb_master(int group)
 155 {
 156         cmm_smp_mb();
 157 }
 158 #endif
 159
 160 #ifdef RCU_SIGNAL
 161 static void force_mb_all_readers(void)
 162 {
 163         struct rcu_reader *index;
 164
 165         /*
 166          * Ask for each threads to execute a cmm_smp_mb() so we can consider the
 167          * compiler barriers around rcu read lock as real memory barriers.
 168          */
 169         if (cds_list_empty(&registry))
 170                 return;
 171         /*
 172          * pthread_kill has a cmm_smp_mb(). But beware, we assume it performs
 173          * a cache flush on architectures with non-coherent cache. Let's play
 174          * safe and don't assume anything : we use cmm_smp_mc() to make sure the
 175          * cache flush is enforced.
 176          */
 177         cds_list_for_each_entry(index, &registry, node) {
 178                 CMM_STORE_SHARED(index->need_mb, 1);
 179                 pthread_kill(index->tid, SIGRCU);
 180         }
 181         /*
 182          * Wait for sighandler (and thus mb()) to execute on every thread.
 183          *
 184          * Note that the pthread_kill() will never be executed on systems
 185          * that correctly deliver signals in a timely manner.  However, it
 186          * is not uncommon for kernels to have bugs that can result in
 187          * lost or unduly delayed signals.
 188          *
 189          * If you are seeing the below pthread_kill() executing much at
 190          * all, we suggest testing the underlying kernel and filing the
 191          * relevant bug report.  For Linux kernels, we recommend getting
 192          * the Linux Test Project (LTP).
 193          */
 194         cds_list_for_each_entry(index, &registry, node) {
 195                 while (CMM_LOAD_SHARED(index->need_mb)) {
 196                         pthread_kill(index->tid, SIGRCU);
 197                         poll(NULL, 0, 1);
 198                 }
 199         }
 200         cmm_smp_mb();   /* read ->need_mb before ending the barrier */
 201 }
 202
 203 static void smp_mb_master(int group)
 204 {
 205         force_mb_all_readers();
 206 }
 207 #endif /* #ifdef RCU_SIGNAL */
 208
 209 /*
 210  * synchronize_rcu() waiting. Single thread.
 211  */
 212 static void wait_gp(void)
 213 {
 214         /* Read reader_gp before read futex */
 215         smp_mb_master(RCU_MB_GROUP);
 216         if (uatomic_read(&gp_futex) == -1)
 217                 futex_async(&gp_futex, FUTEX_WAIT, -1,
 218                       NULL, NULL, 0);
 219 }
 220
 221 void update_counter_and_wait(void)
 222 {
 223         CDS_LIST_HEAD(qsreaders);
 224         int wait_loops = 0;
 225         struct rcu_reader *index, *tmp;
 226
 227         /* Switch parity: 0 -> 1, 1 -> 0 */
 228         CMM_STORE_SHARED(rcu_gp_ctr, rcu_gp_ctr ^ RCU_GP_CTR_PHASE);
 229
 230         /*
 231          * Must commit rcu_gp_ctr update to memory before waiting for quiescent
 232          * state. Failure to do so could result in the writer waiting forever
 233          * while new readers are always accessing data (no progress). Enforce
 234          * compiler-order of store to rcu_gp_ctr before load rcu_reader ctr.
 235          */
 236         cmm_barrier();
 237
 238         /*
 239          *
 240          * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
 241          * model easier to understand. It does not have a big performance impact
 242          * anyway, given this is the write-side.
 243          */
 244         cmm_smp_mb();
 245
 246         /*
 247          * Wait for each thread rcu_reader.ctr count to become 0.
 248          */
 249         for (;;) {
 250                 wait_loops++;
 251                 if (wait_loops == RCU_QS_ACTIVE_ATTEMPTS) {
 252                         uatomic_dec(&gp_futex);
 253                         /* Write futex before read reader_gp */
 254                         smp_mb_master(RCU_MB_GROUP);
 255                 }
 256
 257                 cds_list_for_each_entry_safe(index, tmp, &registry, node) {
 258                         if (!rcu_gp_ongoing(&index->ctr))
 259                                 cds_list_move(&index->node, &qsreaders);
 260                 }
 261
 262 #ifndef HAS_INCOHERENT_CACHES
 263                 if (cds_list_empty(&registry)) {
 264                         if (wait_loops == RCU_QS_ACTIVE_ATTEMPTS) {
 265                                 /* Read reader_gp before write futex */
 266                                 smp_mb_master(RCU_MB_GROUP);
 267                                 uatomic_set(&gp_futex, 0);
 268                         }
 269                         break;
 270                 } else {
 271                         if (wait_loops == RCU_QS_ACTIVE_ATTEMPTS)
 272                                 wait_gp();
 273                         else
 274                                 caa_cpu_relax();
 275                 }
 276 #else /* #ifndef HAS_INCOHERENT_CACHES */
 277                 /*
 278                  * BUSY-LOOP. Force the reader thread to commit its
 279                  * rcu_reader.ctr update to memory if we wait for too long.
 280                  */
 281                 if (cds_list_empty(&registry)) {
 282                         if (wait_loops == RCU_QS_ACTIVE_ATTEMPTS) {
 283                                 /* Read reader_gp before write futex */
 284                                 smp_mb_master(RCU_MB_GROUP);
 285                                 uatomic_set(&gp_futex, 0);
 286                         }
 287                         break;
 288                 } else {
 289                         switch (wait_loops) {
 290                         case RCU_QS_ACTIVE_ATTEMPTS:
 291                                 wait_gp();
 292                                 break; /* only escape switch */
 293                         case KICK_READER_LOOPS:
 294                                 smp_mb_master(RCU_MB_GROUP);
 295                                 wait_loops = 0;
 296                                 break; /* only escape switch */
 297                         default:
 298                                 caa_cpu_relax();
 299                         }
 300                 }
 301 #endif /* #else #ifndef HAS_INCOHERENT_CACHES */
 302         }
 303         /* put back the reader list in the registry */
 304         cds_list_splice(&qsreaders, &registry);
 305 }
 306
 307 void synchronize_rcu(void)
 308 {
 309         mutex_lock(&rcu_gp_lock);
 310
 311         if (cds_list_empty(&registry))
 312                 goto out;
 313
 314         /* All threads should read qparity before accessing data structure
 315          * where new ptr points to. Must be done within rcu_gp_lock because it
 316          * iterates on reader threads.*/
 317         /* Write new ptr before changing the qparity */
 318         smp_mb_master(RCU_MB_GROUP);
 319
 320         /*
 321          * Wait for previous parity to be empty of readers.
 322          */
 323         update_counter_and_wait();      /* 0 -> 1, wait readers in parity 0 */
 324
 325         /*
 326          * Must finish waiting for quiescent state for parity 0 before
 327          * committing next rcu_gp_ctr update to memory. Failure to do so could
 328          * result in the writer waiting forever while new readers are always
 329          * accessing data (no progress).  Enforce compiler-order of load
 330          * rcu_reader ctr before store to rcu_gp_ctr.
 331          */
 332         cmm_barrier();
 333
 334         /*
 335          * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
 336          * model easier to understand. It does not have a big performance impact
 337          * anyway, given this is the write-side.
 338          */
 339         cmm_smp_mb();
 340
 341         /*
 342          * Wait for previous parity to be empty of readers.
 343          */
 344         update_counter_and_wait();      /* 1 -> 0, wait readers in parity 1 */
 345
 346         /* Finish waiting for reader threads before letting the old ptr being
 347          * freed. Must be done within rcu_gp_lock because it iterates on reader
 348          * threads. */
 349         smp_mb_master(RCU_MB_GROUP);
 350 out:
 351         mutex_unlock(&rcu_gp_lock);
 352 }
 353
 354 /*
 355  * library wrappers to be used by non-LGPL compatible source code.
 356  */
 357
 358 void rcu_read_lock(void)
 359 {
 360         _rcu_read_lock();
 361 }
 362
 363 void rcu_read_unlock(void)
 364 {
 365         _rcu_read_unlock();
 366 }
 367
 368 void rcu_register_thread(void)
 369 {
 370         rcu_reader.tid = pthread_self();
 371         assert(rcu_reader.need_mb == 0);
 372         assert(!(rcu_reader.ctr & RCU_GP_CTR_NEST_MASK));
 373
 374         mutex_lock(&rcu_gp_lock);
 375         rcu_init();     /* In case gcc does not support constructor attribute */
 376         cds_list_add(&rcu_reader.node, &registry);
 377         mutex_unlock(&rcu_gp_lock);
 378 }
 379
 380 void rcu_unregister_thread(void)
 381 {
 382         mutex_lock(&rcu_gp_lock);
 383         cds_list_del(&rcu_reader.node);
 384         mutex_unlock(&rcu_gp_lock);
 385 }
 386
 387 #ifdef RCU_MEMBARRIER
 388 void rcu_init(void)
 389 {
 390         if (init_done)
 391                 return;
 392         init_done = 1;
 393         if (!membarrier(MEMBARRIER_EXPEDITED | MEMBARRIER_QUERY))
 394                 has_sys_membarrier = 1;
 395 }
 396 #endif
 397
 398 #ifdef RCU_SIGNAL
 399 static void sigrcu_handler(int signo, siginfo_t *siginfo, void *context)
 400 {
 401         /*
 402          * Executing this cmm_smp_mb() is the only purpose of this signal handler.
 403          * It punctually promotes cmm_barrier() into cmm_smp_mb() on every thread it is
 404          * executed on.
 405          */
 406         cmm_smp_mb();
 407         _CMM_STORE_SHARED(rcu_reader.need_mb, 0);
 408         cmm_smp_mb();
 409 }
 410
 411 /*
 412  * rcu_init constructor. Called when the library is linked, but also when
 413  * reader threads are calling rcu_register_thread().
 414  * Should only be called by a single thread at a given time. This is ensured by
 415  * holing the rcu_gp_lock from rcu_register_thread() or by running at library
 416  * load time, which should not be executed by multiple threads nor concurrently
 417  * with rcu_register_thread() anyway.
 418  */
 419 void rcu_init(void)
 420 {
 421         struct sigaction act;
 422         int ret;
 423
 424         if (init_done)
 425                 return;
 426         init_done = 1;
 427
 428         act.sa_sigaction = sigrcu_handler;
 429         act.sa_flags = SA_SIGINFO | SA_RESTART;
 430         sigemptyset(&act.sa_mask);
 431         ret = sigaction(SIGRCU, &act, NULL);
 432         if (ret) {
 433                 perror("Error in sigaction");
 434                 exit(-1);
 435         }
 436 }
 437
 438 void rcu_exit(void)
 439 {
 440         struct sigaction act;
 441         int ret;
 442
 443         ret = sigaction(SIGRCU, NULL, &act);
 444         if (ret) {
 445                 perror("Error in sigaction");
 446                 exit(-1);
 447         }
 448         assert(act.sa_sigaction == sigrcu_handler);
 449         assert(cds_list_empty(&registry));
 450 }
 451
 452 #endif /* #ifdef RCU_SIGNAL */
 453
 454 #include "urcu-call-rcu-impl.h"
 455 #include "urcu-defer-impl.h"