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