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 /* Thread IDs of registered readers */
  69 #define INIT_NUM_THREADS 4
  70
  71 #ifdef DEBUG_YIELD
  72 unsigned int yield_active;
  73 unsigned int __thread rand_yield;
  74 #endif
  75
  76 static LIST_HEAD(registry);
  77
  78 static void internal_urcu_lock(void)
  79 {
  80         int ret;
  81
  82 #ifndef DISTRUST_SIGNALS_EXTREME
  83         ret = pthread_mutex_lock(&urcu_mutex);
  84         if (ret) {
  85                 perror("Error in pthread mutex lock");
  86                 exit(-1);
  87         }
  88 #else /* #ifndef DISTRUST_SIGNALS_EXTREME */
  89         while ((ret = pthread_mutex_trylock(&urcu_mutex)) != 0) {
  90                 if (ret != EBUSY && ret != EINTR) {
  91                         printf("ret = %d, errno = %d\n", ret, errno);
  92                         perror("Error in pthread mutex lock");
  93                         exit(-1);
  94                 }
  95                 if (urcu_reader.need_mb) {
  96                         smp_mb();
  97                         urcu_reader.need_mb = 0;
  98                         smp_mb();
  99                 }
 100                 poll(NULL,0,10);
 101         }
 102 #endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */
 103 }
 104
 105 static void internal_urcu_unlock(void)
 106 {
 107         int ret;
 108
 109         ret = pthread_mutex_unlock(&urcu_mutex);
 110         if (ret) {
 111                 perror("Error in pthread mutex unlock");
 112                 exit(-1);
 113         }
 114 }
 115
 116 /*
 117  * called with urcu_mutex held.
 118  */
 119 static void switch_next_urcu_qparity(void)
 120 {
 121         STORE_SHARED(urcu_gp_ctr, urcu_gp_ctr ^ RCU_GP_CTR_BIT);
 122 }
 123
 124 #ifdef URCU_MB
 125 static void force_mb_single_thread(struct urcu_reader *index)
 126 {
 127         smp_mb();
 128 }
 129
 130 static void force_mb_all_threads(void)
 131 {
 132         smp_mb();
 133 }
 134 #else /* #ifdef URCU_MB */
 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
 158 static void force_mb_all_threads(void)
 159 {
 160         struct urcu_reader *index;
 161
 162         /*
 163          * Ask for each threads to execute a smp_mb() so we can consider the
 164          * compiler barriers around rcu read lock as real memory barriers.
 165          */
 166         if (list_empty(&registry))
 167                 return;
 168         /*
 169          * pthread_kill has a smp_mb(). But beware, we assume it performs
 170          * a cache flush on architectures with non-coherent cache. Let's play
 171          * safe and don't assume anything : we use smp_mc() to make sure the
 172          * cache flush is enforced.
 173          */
 174         list_for_each_entry(index, &registry, head) {
 175                 index->need_mb = 1;
 176                 smp_mc();       /* write need_mb before sending the signal */
 177                 pthread_kill(index->tid, SIGURCU);
 178         }
 179         /*
 180          * Wait for sighandler (and thus mb()) to execute on every thread.
 181          *
 182          * Note that the pthread_kill() will never be executed on systems
 183          * that correctly deliver signals in a timely manner.  However, it
 184          * is not uncommon for kernels to have bugs that can result in
 185          * lost or unduly delayed signals.
 186          *
 187          * If you are seeing the below pthread_kill() executing much at
 188          * all, we suggest testing the underlying kernel and filing the
 189          * relevant bug report.  For Linux kernels, we recommend getting
 190          * the Linux Test Project (LTP).
 191          */
 192         list_for_each_entry(index, &registry, head) {
 193                 while (index->need_mb) {
 194                         pthread_kill(index->tid, SIGURCU);
 195                         poll(NULL, 0, 1);
 196                 }
 197         }
 198         smp_mb();       /* read ->need_mb before ending the barrier */
 199 }
 200 #endif /* #else #ifdef URCU_MB */
 201
 202 /*
 203  * synchronize_rcu() waiting. Single thread.
 204  */
 205 static void wait_gp(struct urcu_reader *index)
 206 {
 207         uatomic_dec(&gp_futex);
 208         force_mb_single_thread(index); /* Write futex before read reader_gp */
 209         if (!rcu_old_gp_ongoing(&index->ctr)) {
 210                 /* Read reader_gp before write futex */
 211                 force_mb_single_thread(index);
 212                 /* Callbacks are queued, don't wait. */
 213                 uatomic_set(&gp_futex, 0);
 214         } else {
 215                 /* Read reader_gp before read futex */
 216                 force_mb_single_thread(index);
 217                 if (uatomic_read(&gp_futex) == -1)
 218                         futex(&gp_futex, FUTEX_WAIT, -1,
 219                               NULL, NULL, 0);
 220         }
 221 }
 222
 223 void wait_for_quiescent_state(void)
 224 {
 225         struct urcu_reader *index;
 226
 227         if (list_empty(&registry))
 228                 return;
 229         /*
 230          * Wait for each thread urcu_reader.ctr count to become 0.
 231          */
 232         list_for_each_entry(index, &registry, head) {
 233                 int wait_loops = 0;
 234 #ifndef HAS_INCOHERENT_CACHES
 235                 while (rcu_old_gp_ongoing(&index->ctr)) {
 236                         if (wait_loops++ == RCU_QS_ACTIVE_ATTEMPTS) {
 237                                 wait_gp(index);
 238                         } else {
 239                                 cpu_relax();
 240                         }
 241                 }
 242 #else /* #ifndef HAS_INCOHERENT_CACHES */
 243                 /*
 244                  * BUSY-LOOP. Force the reader thread to commit its
 245                  * urcu_reader.ctr update to memory if we wait for too long.
 246                  */
 247                 while (rcu_old_gp_ongoing(&index->ctr)) {
 248                         switch (wait_loops++) {
 249                         case RCU_QS_ACTIVE_ATTEMPTS:
 250                                 wait_gp(index);
 251                                 break;
 252                         case KICK_READER_LOOPS:
 253                                 force_mb_single_thread(index);
 254                                 wait_loops = 0;
 255                                 break;
 256                         default:
 257                                 cpu_relax();
 258                         }
 259                 }
 260 #endif /* #else #ifndef HAS_INCOHERENT_CACHES */
 261         }
 262 }
 263
 264 void synchronize_rcu(void)
 265 {
 266         internal_urcu_lock();
 267
 268         /* All threads should read qparity before accessing data structure
 269          * where new ptr points to. Must be done within internal_urcu_lock
 270          * because it iterates on reader threads.*/
 271         /* Write new ptr before changing the qparity */
 272         force_mb_all_threads();
 273
 274         switch_next_urcu_qparity();     /* 0 -> 1 */
 275
 276         /*
 277          * Must commit qparity update to memory before waiting for parity
 278          * 0 quiescent state. Failure to do so could result in the writer
 279          * waiting forever while new readers are always accessing data (no
 280          * progress).
 281          * Ensured by STORE_SHARED and LOAD_SHARED.
 282          */
 283
 284         /*
 285          * Adding a smp_mb() which is _not_ formally required, but makes the
 286          * model easier to understand. It does not have a big performance impact
 287          * anyway, given this is the write-side.
 288          */
 289         smp_mb();
 290
 291         /*
 292          * Wait for previous parity to be empty of readers.
 293          */
 294         wait_for_quiescent_state();     /* Wait readers in parity 0 */
 295
 296         /*
 297          * Must finish waiting for quiescent state for parity 0 before
 298          * committing qparity update to memory. Failure to do so could result in
 299          * the writer waiting forever while new readers are always accessing
 300          * data (no 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         switch_next_urcu_qparity();     /* 1 -> 0 */
 312
 313         /*
 314          * Must commit qparity update to memory before waiting for parity
 315          * 1 quiescent state. Failure to do so could result in the writer
 316          * waiting forever while new readers are always accessing data (no
 317          * progress).
 318          * Ensured by STORE_SHARED and LOAD_SHARED.
 319          */
 320
 321         /*
 322          * Adding a smp_mb() which is _not_ formally required, but makes the
 323          * model easier to understand. It does not have a big performance impact
 324          * anyway, given this is the write-side.
 325          */
 326         smp_mb();
 327
 328         /*
 329          * Wait for previous parity to be empty of readers.
 330          */
 331         wait_for_quiescent_state();     /* Wait readers in parity 1 */
 332
 333         /* Finish waiting for reader threads before letting the old ptr being
 334          * freed. Must be done within internal_urcu_lock because it iterates on
 335          * reader threads. */
 336         force_mb_all_threads();
 337
 338         internal_urcu_unlock();
 339 }
 340
 341 /*
 342  * library wrappers to be used by non-LGPL compatible source code.
 343  */
 344
 345 void rcu_read_lock(void)
 346 {
 347         _rcu_read_lock();
 348 }
 349
 350 void rcu_read_unlock(void)
 351 {
 352         _rcu_read_unlock();
 353 }
 354
 355 void *rcu_dereference(void *p)
 356 {
 357         return _rcu_dereference(p);
 358 }
 359
 360 void *rcu_assign_pointer_sym(void **p, void *v)
 361 {
 362         wmb();
 363         return STORE_SHARED(p, v);
 364 }
 365
 366 void *rcu_xchg_pointer_sym(void **p, void *v)
 367 {
 368         wmb();
 369         return uatomic_xchg(p, v);
 370 }
 371
 372 void *rcu_cmpxchg_pointer_sym(void **p, void *old, void *_new)
 373 {
 374         wmb();
 375         return uatomic_cmpxchg(p, old, _new);
 376 }
 377
 378 void *rcu_publish_content_sym(void **p, void *v)
 379 {
 380         void *oldptr;
 381
 382         oldptr = _rcu_xchg_pointer(p, v);
 383         synchronize_rcu();
 384         return oldptr;
 385 }
 386
 387 void rcu_register_thread(void)
 388 {
 389         internal_urcu_lock();
 390         urcu_init();    /* In case gcc does not support constructor attribute */
 391         urcu_reader.tid = pthread_self();
 392         assert(urcu_reader.need_mb == 0);
 393         assert(urcu_reader.ctr == 0);
 394         list_add(&urcu_reader.head, &registry);
 395         internal_urcu_unlock();
 396 }
 397
 398 void rcu_unregister_thread(void)
 399 {
 400         internal_urcu_lock();
 401         list_del(&urcu_reader.head);
 402         internal_urcu_unlock();
 403 }
 404
 405 #ifndef URCU_MB
 406 static void sigurcu_handler(int signo, siginfo_t *siginfo, void *context)
 407 {
 408         /*
 409          * Executing this smp_mb() is the only purpose of this signal handler.
 410          * It punctually promotes barrier() into smp_mb() on every thread it is
 411          * executed on.
 412          */
 413         smp_mb();
 414         urcu_reader.need_mb = 0;
 415         smp_mb();
 416 }
 417
 418 /*
 419  * urcu_init constructor. Called when the library is linked, but also when
 420  * reader threads are calling rcu_register_thread().
 421  * Should only be called by a single thread at a given time. This is ensured by
 422  * holing the internal_urcu_lock() from rcu_register_thread() or by running at
 423  * library load time, which should not be executed by multiple threads nor
 424  * concurrently with rcu_register_thread() anyway.
 425  */
 426 void urcu_init(void)
 427 {
 428         struct sigaction act;
 429         int ret;
 430
 431         if (init_done)
 432                 return;
 433         init_done = 1;
 434
 435         act.sa_sigaction = sigurcu_handler;
 436         act.sa_flags = SA_SIGINFO | SA_RESTART;
 437         sigemptyset(&act.sa_mask);
 438         ret = sigaction(SIGURCU, &act, NULL);
 439         if (ret) {
 440                 perror("Error in sigaction");
 441                 exit(-1);
 442         }
 443 }
 444
 445 void urcu_exit(void)
 446 {
 447         struct sigaction act;
 448         int ret;
 449
 450         ret = sigaction(SIGURCU, NULL, &act);
 451         if (ret) {
 452                 perror("Error in sigaction");
 453                 exit(-1);
 454         }
 455         assert(act.sa_sigaction == sigurcu_handler);
 456         assert(list_empty(&registry));
 457 }
 458 #endif /* #ifndef URCU_MB */