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 /*
  53  * Global grace period counter.
  54  * Contains the current RCU_GP_CTR_BIT.
  55  * Also has a RCU_GP_COUNT of 1, to accelerate the reader fast path.
  56  * Written to only by writer with mutex taken. Read by both writer and readers.
  57  */
  58 long urcu_gp_ctr = RCU_GP_COUNT;
  59
  60 /*
  61  * Written to only by each individual reader. Read by both the reader and the
  62  * writers.
  63  */
  64 struct urcu_reader_status __thread urcu_reader_status;
  65
  66 /* Thread IDs of registered readers */
  67 #define INIT_NUM_THREADS 4
  68
  69 struct reader_registry {
  70         pthread_t tid;
  71         struct urcu_reader_status *urcu_reader_status;
  72         char *need_mb;
  73 };
  74
  75 #ifdef DEBUG_YIELD
  76 unsigned int yield_active;
  77 unsigned int __thread rand_yield;
  78 #endif
  79
  80 static struct reader_registry *registry;
  81 static char __thread need_mb;
  82 static int num_readers, alloc_readers;
  83
  84 static void internal_urcu_lock(void)
  85 {
  86         int ret;
  87
  88 #ifndef DISTRUST_SIGNALS_EXTREME
  89         ret = pthread_mutex_lock(&urcu_mutex);
  90         if (ret) {
  91                 perror("Error in pthread mutex lock");
  92                 exit(-1);
  93         }
  94 #else /* #ifndef DISTRUST_SIGNALS_EXTREME */
  95         while ((ret = pthread_mutex_trylock(&urcu_mutex)) != 0) {
  96                 if (ret != EBUSY && ret != EINTR) {
  97                         printf("ret = %d, errno = %d\n", ret, errno);
  98                         perror("Error in pthread mutex lock");
  99                         exit(-1);
 100                 }
 101                 if (need_mb) {
 102                         smp_mb();
 103                         need_mb = 0;
 104                         smp_mb();
 105                 }
 106                 poll(NULL,0,10);
 107         }
 108 #endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */
 109 }
 110
 111 static void internal_urcu_unlock(void)
 112 {
 113         int ret;
 114
 115         ret = pthread_mutex_unlock(&urcu_mutex);
 116         if (ret) {
 117                 perror("Error in pthread mutex unlock");
 118                 exit(-1);
 119         }
 120 }
 121
 122 /*
 123  * called with urcu_mutex held.
 124  */
 125 static void switch_next_urcu_qparity(void)
 126 {
 127         STORE_SHARED(urcu_gp_ctr, urcu_gp_ctr ^ RCU_GP_CTR_BIT);
 128 }
 129
 130 #ifdef URCU_MB
 131 #ifdef HAS_INCOHERENT_CACHES
 132 static void force_mb_single_thread(struct reader_registry *index)
 133 {
 134         smp_mb();
 135 }
 136 #endif /* #ifdef HAS_INCOHERENT_CACHES */
 137
 138 static void force_mb_all_threads(void)
 139 {
 140         smp_mb();
 141 }
 142 #else /* #ifdef URCU_MB */
 143 #ifdef HAS_INCOHERENT_CACHES
 144 static void force_mb_single_thread(struct reader_registry *index)
 145 {
 146         assert(registry);
 147         /*
 148          * pthread_kill has a smp_mb(). But beware, we assume it performs
 149          * a cache flush on architectures with non-coherent cache. Let's play
 150          * safe and don't assume anything : we use smp_mc() to make sure the
 151          * cache flush is enforced.
 152          */
 153         *index->need_mb = 1;
 154         smp_mc();       /* write ->need_mb before sending the signals */
 155         pthread_kill(index->tid, SIGURCU);
 156         smp_mb();
 157         /*
 158          * Wait for sighandler (and thus mb()) to execute on every thread.
 159          * BUSY-LOOP.
 160          */
 161         while (*index->need_mb) {
 162                 poll(NULL, 0, 1);
 163         }
 164         smp_mb();       /* read ->need_mb before ending the barrier */
 165 }
 166 #endif /* #ifdef HAS_INCOHERENT_CACHES */
 167
 168 static void force_mb_all_threads(void)
 169 {
 170         struct reader_registry *index;
 171         /*
 172          * Ask for each threads to execute a smp_mb() so we can consider the
 173          * compiler barriers around rcu read lock as real memory barriers.
 174          */
 175         if (!registry)
 176                 return;
 177         /*
 178          * pthread_kill has a smp_mb(). But beware, we assume it performs
 179          * a cache flush on architectures with non-coherent cache. Let's play
 180          * safe and don't assume anything : we use smp_mc() to make sure the
 181          * cache flush is enforced.
 182          */
 183         for (index = registry; index < registry + num_readers; index++) {
 184                 *index->need_mb = 1;
 185                 smp_mc();       /* write need_mb before sending the signal */
 186                 pthread_kill(index->tid, SIGURCU);
 187         }
 188         /*
 189          * Wait for sighandler (and thus mb()) to execute on every thread.
 190          *
 191          * Note that the pthread_kill() will never be executed on systems
 192          * that correctly deliver signals in a timely manner.  However, it
 193          * is not uncommon for kernels to have bugs that can result in
 194          * lost or unduly delayed signals.
 195          *
 196          * If you are seeing the below pthread_kill() executing much at
 197          * all, we suggest testing the underlying kernel and filing the
 198          * relevant bug report.  For Linux kernels, we recommend getting
 199          * the Linux Test Project (LTP).
 200          */
 201         for (index = registry; index < registry + num_readers; index++) {
 202                 while (*index->need_mb) {
 203                         pthread_kill(index->tid, SIGURCU);
 204                         poll(NULL, 0, 1);
 205                 }
 206         }
 207         smp_mb();       /* read ->need_mb before ending the barrier */
 208 }
 209 #endif /* #else #ifdef URCU_MB */
 210
 211 void wait_for_quiescent_state(void)
 212 {
 213         struct reader_registry *index;
 214
 215         if (!registry)
 216                 return;
 217         /*
 218          * Wait for each thread active_readers count to become 0.
 219          */
 220         for (index = registry; index < registry + num_readers; index++) {
 221                 int wait_loops = 0;
 222
 223                 if (likely(!rcu_old_gp_ongoing(
 224                                 &index->urcu_reader_status->active_readers)))
 225                         continue;
 226
 227                 index->urcu_reader_status->gp_waiting = 1;
 228 #ifndef HAS_INCOHERENT_CACHES
 229                 while (rcu_old_gp_ongoing(
 230                                 &index->urcu_reader_status->active_readers)) {
 231                         if (wait_loops++ == RCU_QS_ACTIVE_ATTEMPTS) {
 232                                 sched_yield();  /* ideally sched_yield_to() */
 233                         } else {
 234                                 cpu_relax();
 235                         }
 236                 }
 237 #else /* #ifndef HAS_INCOHERENT_CACHES */
 238                 /*
 239                  * BUSY-LOOP. Force the reader thread to commit its
 240                  * active_readers update to memory if we wait for too long.
 241                  */
 242                 while (rcu_old_gp_ongoing(
 243                                 &index->urcu_reader_status->active_readers)) {
 244                         switch (wait_loops++) {
 245                         case RCU_QS_ACTIVE_ATTEMPTS:
 246                                 sched_yield();  /* ideally sched_yield_to() */
 247                                 break;
 248                         case KICK_READER_LOOPS:
 249                                 force_mb_single_thread(index);
 250                                 wait_loops = 0;
 251                                 break;
 252                         default:
 253                                 cpu_relax();
 254                         }
 255                 }
 256 #endif /* #else #ifndef HAS_INCOHERENT_CACHES */
 257                 index->urcu_reader_status->gp_waiting = 0;
 258         }
 259 }
 260
 261 void synchronize_rcu(void)
 262 {
 263         internal_urcu_lock();
 264
 265         /* All threads should read qparity before accessing data structure
 266          * where new ptr points to. Must be done within internal_urcu_lock
 267          * because it iterates on reader threads.*/
 268         /* Write new ptr before changing the qparity */
 269         force_mb_all_threads();
 270
 271         STORE_SHARED(urcu_gp_ctr, urcu_gp_ctr ^ RCU_GP_ONGOING);
 272
 273         switch_next_urcu_qparity();     /* 0 -> 1 */
 274
 275         /*
 276          * Must commit qparity update to memory before waiting for parity
 277          * 0 quiescent state. Failure to do so could result in the writer
 278          * waiting forever while new readers are always accessing data (no
 279          * progress).
 280          * Ensured by STORE_SHARED and LOAD_SHARED.
 281          */
 282
 283         /*
 284          * Adding a 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         smp_mb();
 289
 290         /*
 291          * Wait for previous parity to be empty of readers.
 292          */
 293         wait_for_quiescent_state();     /* Wait readers in parity 0 */
 294
 295         /*
 296          * Must finish waiting for quiescent state for parity 0 before
 297          * committing qparity update to memory. Failure to do so could result in
 298          * the writer waiting forever while new readers are always accessing
 299          * data (no progress).
 300          * Ensured by STORE_SHARED and LOAD_SHARED.
 301          */
 302
 303         /*
 304          * Adding a smp_mb() which is _not_ formally required, but makes the
 305          * model easier to understand. It does not have a big performance impact
 306          * anyway, given this is the write-side.
 307          */
 308         smp_mb();
 309
 310         switch_next_urcu_qparity();     /* 1 -> 0 */
 311
 312         /*
 313          * Must commit qparity update to memory before waiting for parity
 314          * 1 quiescent state. Failure to do so could result in the writer
 315          * waiting forever while new readers are always accessing data (no
 316          * progress).
 317          * Ensured by STORE_SHARED and LOAD_SHARED.
 318          */
 319
 320         /*
 321          * Adding a smp_mb() which is _not_ formally required, but makes the
 322          * model easier to understand. It does not have a big performance impact
 323          * anyway, given this is the write-side.
 324          */
 325         smp_mb();
 326
 327         /*
 328          * Wait for previous parity to be empty of readers.
 329          */
 330         wait_for_quiescent_state();     /* Wait readers in parity 1 */
 331
 332         STORE_SHARED(urcu_gp_ctr, urcu_gp_ctr ^ RCU_GP_ONGOING);
 333
 334         /* Finish waiting for reader threads before letting the old ptr being
 335          * freed. Must be done within internal_urcu_lock because it iterates on
 336          * reader threads. */
 337         force_mb_all_threads();
 338
 339         internal_urcu_unlock();
 340 }
 341
 342 /*
 343  * library wrappers to be used by non-LGPL compatible source code.
 344  */
 345
 346 void rcu_read_lock(void)
 347 {
 348         _rcu_read_lock();
 349 }
 350
 351 void rcu_read_unlock(void)
 352 {
 353         _rcu_read_unlock();
 354 }
 355
 356 void *rcu_dereference(void *p)
 357 {
 358         return _rcu_dereference(p);
 359 }
 360
 361 void *rcu_assign_pointer_sym(void **p, void *v)
 362 {
 363         wmb();
 364         return STORE_SHARED(p, v);
 365 }
 366
 367 void *rcu_xchg_pointer_sym(void **p, void *v)
 368 {
 369         wmb();
 370         return xchg(p, v);
 371 }
 372
 373 void *rcu_cmpxchg_pointer_sym(void **p, void *old, void *_new)
 374 {
 375         wmb();
 376         return cmpxchg(p, old, _new);
 377 }
 378
 379 void *rcu_publish_content_sym(void **p, void *v)
 380 {
 381         void *oldptr;
 382
 383         oldptr = _rcu_xchg_pointer(p, v);
 384         synchronize_rcu();
 385         return oldptr;
 386 }
 387
 388 static void rcu_add_reader(pthread_t id)
 389 {
 390         struct reader_registry *oldarray;
 391
 392         if (!registry) {
 393                 alloc_readers = INIT_NUM_THREADS;
 394                 num_readers = 0;
 395                 registry =
 396                         malloc(sizeof(struct reader_registry) * alloc_readers);
 397         }
 398         if (alloc_readers < num_readers + 1) {
 399                 oldarray = registry;
 400                 registry = malloc(sizeof(struct reader_registry)
 401                                 * (alloc_readers << 1));
 402                 memcpy(registry, oldarray,
 403                         sizeof(struct reader_registry) * alloc_readers);
 404                 alloc_readers <<= 1;
 405                 free(oldarray);
 406         }
 407         registry[num_readers].tid = id;
 408         /* reference to the TLS of _this_ reader thread. */
 409         registry[num_readers].urcu_reader_status = &urcu_reader_status;
 410         registry[num_readers].need_mb = &need_mb;
 411         num_readers++;
 412 }
 413
 414 /*
 415  * Never shrink (implementation limitation).
 416  * This is O(nb threads). Eventually use a hash table.
 417  */
 418 static void rcu_remove_reader(pthread_t id)
 419 {
 420         struct reader_registry *index;
 421
 422         assert(registry != NULL);
 423         for (index = registry; index < registry + num_readers; index++) {
 424                 if (pthread_equal(index->tid, id)) {
 425                         memcpy(index, &registry[num_readers - 1],
 426                                 sizeof(struct reader_registry));
 427                         registry[num_readers - 1].tid = 0;
 428                         registry[num_readers - 1].urcu_reader_status = NULL;
 429                         num_readers--;
 430                         return;
 431                 }
 432         }
 433         /* Hrm not found, forgot to register ? */
 434         assert(0);
 435 }
 436
 437 void rcu_register_thread(void)
 438 {
 439         internal_urcu_lock();
 440         urcu_init();    /* In case gcc does not support constructor attribute */
 441         rcu_add_reader(pthread_self());
 442         internal_urcu_unlock();
 443 }
 444
 445 void rcu_unregister_thread(void)
 446 {
 447         internal_urcu_lock();
 448         rcu_remove_reader(pthread_self());
 449         internal_urcu_unlock();
 450 }
 451
 452 #ifndef URCU_MB
 453 static void sigurcu_handler(int signo, siginfo_t *siginfo, void *context)
 454 {
 455         /*
 456          * Executing this smp_mb() is the only purpose of this signal handler.
 457          * It punctually promotes barrier() into smp_mb() on every thread it is
 458          * executed on.
 459          */
 460         smp_mb();
 461         need_mb = 0;
 462         smp_mb();
 463 }
 464
 465 /*
 466  * urcu_init constructor. Called when the library is linked, but also when
 467  * reader threads are calling rcu_register_thread().
 468  * Should only be called by a single thread at a given time. This is ensured by
 469  * holing the internal_urcu_lock() from rcu_register_thread() or by running at
 470  * library load time, which should not be executed by multiple threads nor
 471  * concurrently with rcu_register_thread() anyway.
 472  */
 473 void urcu_init(void)
 474 {
 475         struct sigaction act;
 476         int ret;
 477
 478         if (init_done)
 479                 return;
 480         init_done = 1;
 481
 482         act.sa_sigaction = sigurcu_handler;
 483         act.sa_flags = SA_SIGINFO | SA_RESTART;
 484         sigemptyset(&act.sa_mask);
 485         ret = sigaction(SIGURCU, &act, NULL);
 486         if (ret) {
 487                 perror("Error in sigaction");
 488                 exit(-1);
 489         }
 490 }
 491
 492 void urcu_exit(void)
 493 {
 494         struct sigaction act;
 495         int ret;
 496
 497         ret = sigaction(SIGURCU, NULL, &act);
 498         if (ret) {
 499                 perror("Error in sigaction");
 500                 exit(-1);
 501         }
 502         assert(act.sa_sigaction == sigurcu_handler);
 503         free(registry);
 504 }
 505 #endif /* #ifndef URCU_MB */