urcu.c

   1 /*
   2  * urcu.c
   3  *
   4  * Userspace RCU library
   5  *
   6  * Copyright (c) 2009 Mathieu Desnoyers <mathieu.desnoyers@polymtl.ca>
   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 #include <stdio.h>
  27 #include <pthread.h>
  28 #include <signal.h>
  29 #include <assert.h>
  30 #include <stdlib.h>
  31 #include <string.h>
  32 #include <errno.h>
  33 #include <poll.h>
  34
  35 #include "urcu-static.h"
  36 /* Do not #define _LGPL_SOURCE to ensure we can emit the wrapper symbols */
  37 #include "urcu.h"
  38
  39 #ifndef URCU_MB
  40 static int init_done;
  41
  42 void __attribute__((constructor)) urcu_init(void);
  43 void __attribute__((destructor)) urcu_exit(void);
  44 #else
  45 void urcu_init(void)
  46 {
  47 }
  48 #endif
  49
  50 static pthread_mutex_t urcu_mutex = PTHREAD_MUTEX_INITIALIZER;
  51
  52 int gp_futex;
  53
  54 /*
  55  * Global grace period counter.
  56  * Contains the current RCU_GP_CTR_BIT.
  57  * Also has a RCU_GP_COUNT of 1, to accelerate the reader fast path.
  58  * Written to only by writer with mutex taken. Read by both writer and readers.
  59  */
  60 long urcu_gp_ctr = RCU_GP_COUNT;
  61
  62 /*
  63  * Written to only by each individual reader. Read by both the reader and the
  64  * writers.
  65  */
  66 struct urcu_reader __thread urcu_reader;
  67
  68 #ifdef DEBUG_YIELD
  69 unsigned int yield_active;
  70 unsigned int __thread rand_yield;
  71 #endif
  72
  73 static LIST_HEAD(registry);
  74
  75 static void internal_urcu_lock(void)
  76 {
  77         int ret;
  78
  79 #ifndef DISTRUST_SIGNALS_EXTREME
  80         ret = pthread_mutex_lock(&urcu_mutex);
  81         if (ret) {
  82                 perror("Error in pthread mutex lock");
  83                 exit(-1);
  84         }
  85 #else /* #ifndef DISTRUST_SIGNALS_EXTREME */
  86         while ((ret = pthread_mutex_trylock(&urcu_mutex)) != 0) {
  87                 if (ret != EBUSY && ret != EINTR) {
  88                         printf("ret = %d, errno = %d\n", ret, errno);
  89                         perror("Error in pthread mutex lock");
  90                         exit(-1);
  91                 }
  92                 if (urcu_reader.need_mb) {
  93                         smp_mb();
  94                         urcu_reader.need_mb = 0;
  95                         smp_mb();
  96                 }
  97                 poll(NULL,0,10);
  98         }
  99 #endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */
 100 }
 101
 102 static void internal_urcu_unlock(void)
 103 {
 104         int ret;
 105
 106         ret = pthread_mutex_unlock(&urcu_mutex);
 107         if (ret) {
 108                 perror("Error in pthread mutex unlock");
 109                 exit(-1);
 110         }
 111 }
 112
 113 /*
 114  * called with urcu_mutex held.
 115  */
 116 static void switch_next_urcu_qparity(void)
 117 {
 118         STORE_SHARED(urcu_gp_ctr, urcu_gp_ctr ^ RCU_GP_CTR_BIT);
 119 }
 120
 121 #ifdef URCU_MB
 122 #if 0 /* unused */
 123 static void force_mb_single_thread(struct urcu_reader *index)
 124 {
 125         smp_mb();
 126 }
 127 #endif //0
 128
 129 static void force_mb_all_threads(void)
 130 {
 131         smp_mb();
 132 }
 133 #else /* #ifdef URCU_MB */
 134 #if 0 /* unused */
 135 static void force_mb_single_thread(struct urcu_reader *index)
 136 {
 137         assert(!list_empty(&registry));
 138         /*
 139          * pthread_kill has a smp_mb(). But beware, we assume it performs
 140          * a cache flush on architectures with non-coherent cache. Let's play
 141          * safe and don't assume anything : we use smp_mc() to make sure the
 142          * cache flush is enforced.
 143          */
 144         index->need_mb = 1;
 145         smp_mc();       /* write ->need_mb before sending the signals */
 146         pthread_kill(index->tid, SIGURCU);
 147         smp_mb();
 148         /*
 149          * Wait for sighandler (and thus mb()) to execute on every thread.
 150          * BUSY-LOOP.
 151          */
 152         while (index->need_mb) {
 153                 poll(NULL, 0, 1);
 154         }
 155         smp_mb();       /* read ->need_mb before ending the barrier */
 156 }
 157 #endif //0
 158
 159 static void force_mb_all_threads(void)
 160 {
 161         struct urcu_reader *index;
 162
 163         /*
 164          * Ask for each threads to execute a smp_mb() so we can consider the
 165          * compiler barriers around rcu read lock as real memory barriers.
 166          */
 167         if (list_empty(&registry))
 168                 return;
 169         /*
 170          * pthread_kill has a smp_mb(). But beware, we assume it performs
 171          * a cache flush on architectures with non-coherent cache. Let's play
 172          * safe and don't assume anything : we use smp_mc() to make sure the
 173          * cache flush is enforced.
 174          */
 175         list_for_each_entry(index, &registry, head) {
 176                 index->need_mb = 1;
 177                 smp_mc();       /* write need_mb before sending the signal */
 178                 pthread_kill(index->tid, SIGURCU);
 179         }
 180         /*
 181          * Wait for sighandler (and thus mb()) to execute on every thread.
 182          *
 183          * Note that the pthread_kill() will never be executed on systems
 184          * that correctly deliver signals in a timely manner.  However, it
 185          * is not uncommon for kernels to have bugs that can result in
 186          * lost or unduly delayed signals.
 187          *
 188          * If you are seeing the below pthread_kill() executing much at
 189          * all, we suggest testing the underlying kernel and filing the
 190          * relevant bug report.  For Linux kernels, we recommend getting
 191          * the Linux Test Project (LTP).
 192          */
 193         list_for_each_entry(index, &registry, head) {
 194                 while (index->need_mb) {
 195                         pthread_kill(index->tid, SIGURCU);
 196                         poll(NULL, 0, 1);
 197                 }
 198         }
 199         smp_mb();       /* read ->need_mb before ending the barrier */
 200 }
 201 #endif /* #else #ifdef URCU_MB */
 202
 203 /*
 204  * synchronize_rcu() waiting. Single thread.
 205  */
 206 static void wait_gp(void)
 207 {
 208         /* Read reader_gp before read futex */
 209         force_mb_all_threads();
 210         if (uatomic_read(&gp_futex) == -1)
 211                 futex(&gp_futex, FUTEX_WAIT, -1,
 212                       NULL, NULL, 0);
 213 }
 214
 215 void wait_for_quiescent_state(void)
 216 {
 217         LIST_HEAD(qsreaders);
 218         int wait_loops = 0;
 219         struct urcu_reader *index, *tmp;
 220
 221         if (list_empty(&registry))
 222                 return;
 223         /*
 224          * Wait for each thread urcu_reader.ctr count to become 0.
 225          */
 226         for (;;) {
 227                 wait_loops++;
 228                 if (wait_loops == RCU_QS_ACTIVE_ATTEMPTS) {
 229                         uatomic_dec(&gp_futex);
 230                         /* Write futex before read reader_gp */
 231                         force_mb_all_threads();
 232                 }
 233
 234                 list_for_each_entry_safe(index, tmp, &registry, head) {
 235                         if (!rcu_old_gp_ongoing(&index->ctr))
 236                                 list_move(&index->head, &qsreaders);
 237                 }
 238
 239 #ifndef HAS_INCOHERENT_CACHES
 240                 if (list_empty(&registry)) {
 241                         if (wait_loops == RCU_QS_ACTIVE_ATTEMPTS) {
 242                                 /* Read reader_gp before write futex */
 243                                 force_mb_all_threads();
 244                                 uatomic_set(&gp_futex, 0);
 245                         }
 246                         break;
 247                 } else {
 248                         if (wait_loops == RCU_QS_ACTIVE_ATTEMPTS)
 249                                 wait_gp();
 250                         else
 251                                 cpu_relax();
 252                 }
 253 #else /* #ifndef HAS_INCOHERENT_CACHES */
 254                 /*
 255                  * BUSY-LOOP. Force the reader thread to commit its
 256                  * urcu_reader.ctr update to memory if we wait for too long.
 257                  */
 258                 if (list_empty(&registry)) {
 259                         if (wait_loops == RCU_QS_ACTIVE_ATTEMPTS) {
 260                                 /* Read reader_gp before write futex */
 261                                 force_mb_all_threads();
 262                                 uatomic_set(&gp_futex, 0);
 263                         }
 264                         break;
 265                 } else {
 266                         switch (wait_loops) {
 267                         case RCU_QS_ACTIVE_ATTEMPTS:
 268                                 wait_gp();
 269                                 break; /* only escape switch */
 270                         case KICK_READER_LOOPS:
 271                                 force_mb_all_threads();
 272                                 wait_loops = 0;
 273                                 break; /* only escape switch */
 274                         default:
 275                                 cpu_relax();
 276                         }
 277                 }
 278 #endif /* #else #ifndef HAS_INCOHERENT_CACHES */
 279         }
 280         /* put back the reader list in the registry */
 281         list_splice(&qsreaders, &registry);
 282 }
 283
 284 void synchronize_rcu(void)
 285 {
 286         internal_urcu_lock();
 287
 288         /* All threads should read qparity before accessing data structure
 289          * where new ptr points to. Must be done within internal_urcu_lock
 290          * because it iterates on reader threads.*/
 291         /* Write new ptr before changing the qparity */
 292         force_mb_all_threads();
 293
 294         switch_next_urcu_qparity();     /* 0 -> 1 */
 295
 296         /*
 297          * Must commit qparity update to memory before waiting for parity
 298          * 0 quiescent state. Failure to do so could result in the writer
 299          * waiting forever while new readers are always accessing data (no
 300          * progress).
 301          * Ensured by STORE_SHARED and LOAD_SHARED.
 302          */
 303
 304         /*
 305          * Adding a 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         smp_mb();
 310
 311         /*
 312          * Wait for previous parity to be empty of readers.
 313          */
 314         wait_for_quiescent_state();     /* Wait readers in parity 0 */
 315
 316         /*
 317          * Must finish waiting for quiescent state for parity 0 before
 318          * committing qparity update to memory. Failure to do so could result in
 319          * the writer waiting forever while new readers are always accessing
 320          * data (no progress).
 321          * Ensured by STORE_SHARED and LOAD_SHARED.
 322          */
 323
 324         /*
 325          * Adding a smp_mb() which is _not_ formally required, but makes the
 326          * model easier to understand. It does not have a big performance impact
 327          * anyway, given this is the write-side.
 328          */
 329         smp_mb();
 330
 331         switch_next_urcu_qparity();     /* 1 -> 0 */
 332
 333         /*
 334          * Must commit qparity update to memory before waiting for parity
 335          * 1 quiescent state. Failure to do so could result in the writer
 336          * waiting forever while new readers are always accessing data (no
 337          * progress).
 338          * Ensured by STORE_SHARED and LOAD_SHARED.
 339          */
 340
 341         /*
 342          * Adding a smp_mb() which is _not_ formally required, but makes the
 343          * model easier to understand. It does not have a big performance impact
 344          * anyway, given this is the write-side.
 345          */
 346         smp_mb();
 347
 348         /*
 349          * Wait for previous parity to be empty of readers.
 350          */
 351         wait_for_quiescent_state();     /* Wait readers in parity 1 */
 352
 353         /* Finish waiting for reader threads before letting the old ptr being
 354          * freed. Must be done within internal_urcu_lock because it iterates on
 355          * reader threads. */
 356         force_mb_all_threads();
 357
 358         internal_urcu_unlock();
 359 }
 360
 361 /*
 362  * library wrappers to be used by non-LGPL compatible source code.
 363  */
 364
 365 void rcu_read_lock(void)
 366 {
 367         _rcu_read_lock();
 368 }
 369
 370 void rcu_read_unlock(void)
 371 {
 372         _rcu_read_unlock();
 373 }
 374
 375 void rcu_register_thread(void)
 376 {
 377         urcu_reader.tid = pthread_self();
 378         assert(urcu_reader.need_mb == 0);
 379         assert(urcu_reader.ctr == 0);
 380
 381         internal_urcu_lock();
 382         urcu_init();    /* In case gcc does not support constructor attribute */
 383         list_add(&urcu_reader.head, &registry);
 384         internal_urcu_unlock();
 385 }
 386
 387 void rcu_unregister_thread(void)
 388 {
 389         internal_urcu_lock();
 390         list_del(&urcu_reader.head);
 391         internal_urcu_unlock();
 392 }
 393
 394 #ifndef URCU_MB
 395 static void sigurcu_handler(int signo, siginfo_t *siginfo, void *context)
 396 {
 397         /*
 398          * Executing this smp_mb() is the only purpose of this signal handler.
 399          * It punctually promotes barrier() into smp_mb() on every thread it is
 400          * executed on.
 401          */
 402         smp_mb();
 403         urcu_reader.need_mb = 0;
 404         smp_mb();
 405 }
 406
 407 /*
 408  * urcu_init constructor. Called when the library is linked, but also when
 409  * reader threads are calling rcu_register_thread().
 410  * Should only be called by a single thread at a given time. This is ensured by
 411  * holing the internal_urcu_lock() from rcu_register_thread() or by running at
 412  * library load time, which should not be executed by multiple threads nor
 413  * concurrently with rcu_register_thread() anyway.
 414  */
 415 void urcu_init(void)
 416 {
 417         struct sigaction act;
 418         int ret;
 419
 420         if (init_done)
 421                 return;
 422         init_done = 1;
 423
 424         act.sa_sigaction = sigurcu_handler;
 425         act.sa_flags = SA_SIGINFO | SA_RESTART;
 426         sigemptyset(&act.sa_mask);
 427         ret = sigaction(SIGURCU, &act, NULL);
 428         if (ret) {
 429                 perror("Error in sigaction");
 430                 exit(-1);
 431         }
 432 }
 433
 434 void urcu_exit(void)
 435 {
 436         struct sigaction act;
 437         int ret;
 438
 439         ret = sigaction(SIGURCU, NULL, &act);
 440         if (ret) {
 441                 perror("Error in sigaction");
 442                 exit(-1);
 443         }
 444         assert(act.sa_sigaction == sigurcu_handler);
 445         assert(list_empty(&registry));
 446 }
 447 #endif /* #ifndef URCU_MB */