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 pthread_mutex_t urcu_mutex = PTHREAD_MUTEX_INITIALIZER;
  40
  41 /*
  42  * Global grace period counter.
  43  * Contains the current RCU_GP_CTR_BIT.
  44  * Also has a RCU_GP_CTR_BIT of 1, to accelerate the reader fast path.
  45  * Written to only by writer with mutex taken. Read by both writer and readers.
  46  */
  47 long urcu_gp_ctr = RCU_GP_COUNT;
  48
  49 /*
  50  * Written to only by each individual reader. Read by both the reader and the
  51  * writers.
  52  */
  53 long __thread urcu_active_readers;
  54
  55 /* Thread IDs of registered readers */
  56 #define INIT_NUM_THREADS 4
  57
  58 struct reader_registry {
  59         pthread_t tid;
  60         long *urcu_active_readers;
  61         char *need_mb;
  62 };
  63
  64 #ifdef DEBUG_YIELD
  65 unsigned int yield_active;
  66 unsigned int __thread rand_yield;
  67 #endif
  68
  69 static struct reader_registry *registry;
  70 static char __thread need_mb;
  71 static int num_readers, alloc_readers;
  72
  73 void internal_urcu_lock(void)
  74 {
  75         int ret;
  76
  77 #ifndef DISTRUST_SIGNALS_EXTREME
  78         ret = pthread_mutex_lock(&urcu_mutex);
  79         if (ret) {
  80                 perror("Error in pthread mutex lock");
  81                 exit(-1);
  82         }
  83 #else /* #ifndef DISTRUST_SIGNALS_EXTREME */
  84         while ((ret = pthread_mutex_trylock(&urcu_mutex)) != 0) {
  85                 if (ret != EBUSY && ret != EINTR) {
  86                         printf("ret = %d, errno = %d\n", ret, errno);
  87                         perror("Error in pthread mutex lock");
  88                         exit(-1);
  89                 }
  90                 if (need_mb) {
  91                         smp_mb();
  92                         need_mb = 0;
  93                         smp_mb();
  94                 }
  95                 poll(NULL,0,10);
  96         }
  97 #endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */
  98 }
  99
 100 void internal_urcu_unlock(void)
 101 {
 102         int ret;
 103
 104         ret = pthread_mutex_unlock(&urcu_mutex);
 105         if (ret) {
 106                 perror("Error in pthread mutex unlock");
 107                 exit(-1);
 108         }
 109 }
 110
 111 /*
 112  * called with urcu_mutex held.
 113  */
 114 static void switch_next_urcu_qparity(void)
 115 {
 116         STORE_SHARED(urcu_gp_ctr, urcu_gp_ctr ^ RCU_GP_CTR_BIT);
 117 }
 118
 119 #ifdef DEBUG_FULL_MB
 120 #ifdef HAS_INCOHERENT_CACHES
 121 static void force_mb_single_thread(struct reader_registry *index)
 122 {
 123         smp_mb();
 124 }
 125 #endif /* #ifdef HAS_INCOHERENT_CACHES */
 126
 127 static void force_mb_all_threads(void)
 128 {
 129         smp_mb();
 130 }
 131 #else /* #ifdef DEBUG_FULL_MB */
 132 #ifdef HAS_INCOHERENT_CACHES
 133 static void force_mb_single_thread(struct reader_registry *index)
 134 {
 135         assert(registry);
 136         /*
 137          * pthread_kill has a smp_mb(). But beware, we assume it performs
 138          * a cache flush on architectures with non-coherent cache. Let's play
 139          * safe and don't assume anything : we use smp_mc() to make sure the
 140          * cache flush is enforced.
 141          */
 142         *index->need_mb = 1;
 143         smp_mc();       /* write ->need_mb before sending the signals */
 144         pthread_kill(index->tid, SIGURCU);
 145         smp_mb();
 146         /*
 147          * Wait for sighandler (and thus mb()) to execute on every thread.
 148          * BUSY-LOOP.
 149          */
 150         while (*index->need_mb) {
 151                 poll(NULL, 0, 1);
 152         }
 153         smp_mb();       /* read ->need_mb before ending the barrier */
 154 }
 155 #endif /* #ifdef HAS_INCOHERENT_CACHES */
 156
 157 static void force_mb_all_threads(void)
 158 {
 159         struct reader_registry *index;
 160         /*
 161          * Ask for each threads to execute a smp_mb() so we can consider the
 162          * compiler barriers around rcu read lock as real memory barriers.
 163          */
 164         if (!registry)
 165                 return;
 166         /*
 167          * pthread_kill has a smp_mb(). But beware, we assume it performs
 168          * a cache flush on architectures with non-coherent cache. Let's play
 169          * safe and don't assume anything : we use smp_mc() to make sure the
 170          * cache flush is enforced.
 171          */
 172         for (index = registry; index < registry + num_readers; index++) {
 173                 *index->need_mb = 1;
 174                 smp_mc();       /* write need_mb before sending the signal */
 175                 pthread_kill(index->tid, SIGURCU);
 176         }
 177         /*
 178          * Wait for sighandler (and thus mb()) to execute on every thread.
 179          *
 180          * Note that the pthread_kill() will never be executed on systems
 181          * that correctly deliver signals in a timely manner.  However, it
 182          * is not uncommon for kernels to have bugs that can result in
 183          * lost or unduly delayed signals.
 184          *
 185          * If you are seeing the below pthread_kill() executing much at
 186          * all, we suggest testing the underlying kernel and filing the
 187          * relevant bug report.  For Linux kernels, we recommend getting
 188          * the Linux Test Project (LTP).
 189          */
 190         for (index = registry; index < registry + num_readers; index++) {
 191                 while (*index->need_mb) {
 192                         pthread_kill(index->tid, SIGURCU);
 193                         poll(NULL, 0, 1);
 194                 }
 195         }
 196         smp_mb();       /* read ->need_mb before ending the barrier */
 197 }
 198 #endif /* #else #ifdef DEBUG_FULL_MB */
 199
 200 void wait_for_quiescent_state(void)
 201 {
 202         struct reader_registry *index;
 203
 204         if (!registry)
 205                 return;
 206         /*
 207          * Wait for each thread urcu_active_readers count to become 0.
 208          */
 209         for (index = registry; index < registry + num_readers; index++) {
 210 #ifndef HAS_INCOHERENT_CACHES
 211                 while (rcu_old_gp_ongoing(index->urcu_active_readers))
 212                         cpu_relax();
 213 #else /* #ifndef HAS_INCOHERENT_CACHES */
 214                 int wait_loops = 0;
 215                 /*
 216                  * BUSY-LOOP. Force the reader thread to commit its
 217                  * urcu_active_readers update to memory if we wait for too long.
 218                  */
 219                 while (rcu_old_gp_ongoing(index->urcu_active_readers)) {
 220                         if (wait_loops++ == KICK_READER_LOOPS) {
 221                                 force_mb_single_thread(index);
 222                                 wait_loops = 0;
 223                         } else {
 224                                 cpu_relax();
 225                         }
 226                 }
 227 #endif /* #else #ifndef HAS_INCOHERENT_CACHES */
 228         }
 229 }
 230
 231 void synchronize_rcu(void)
 232 {
 233         internal_urcu_lock();
 234
 235         /* All threads should read qparity before accessing data structure
 236          * where new ptr points to. Must be done within internal_urcu_lock
 237          * because it iterates on reader threads.*/
 238         /* Write new ptr before changing the qparity */
 239         force_mb_all_threads();
 240
 241         switch_next_urcu_qparity();     /* 0 -> 1 */
 242
 243         /*
 244          * Must commit qparity update to memory before waiting for parity
 245          * 0 quiescent state. Failure to do so could result in the writer
 246          * waiting forever while new readers are always accessing data (no
 247          * progress).
 248          * Ensured by STORE_SHARED and LOAD_SHARED.
 249          */
 250
 251         /*
 252          * Current RCU formal verification model assumes sequential execution of
 253          * the write-side. Add core synchronization instructions. Can be removed
 254          * if the formal model is extended to prove that reordering is still
 255          * correct.
 256          */
 257         sync_core();    /* Formal model assumes serialized execution */
 258
 259         /*
 260          * Wait for previous parity to be empty of readers.
 261          */
 262         wait_for_quiescent_state();     /* Wait readers in parity 0 */
 263
 264         /*
 265          * Must finish waiting for quiescent state for parity 0 before
 266          * committing qparity update to memory. Failure to do so could result in
 267          * the writer waiting forever while new readers are always accessing
 268          * data (no progress).
 269          * Ensured by STORE_SHARED and LOAD_SHARED.
 270          */
 271
 272         sync_core();    /* Formal model assumes serialized execution */
 273
 274         switch_next_urcu_qparity();     /* 1 -> 0 */
 275
 276         /*
 277          * Must commit qparity update to memory before waiting for parity
 278          * 1 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         sync_core();    /* Formal model assumes serialized execution */
 285
 286         /*
 287          * Wait for previous parity to be empty of readers.
 288          */
 289         wait_for_quiescent_state();     /* Wait readers in parity 1 */
 290
 291         /* Finish waiting for reader threads before letting the old ptr being
 292          * freed. Must be done within internal_urcu_lock because it iterates on
 293          * reader threads. */
 294         force_mb_all_threads();
 295
 296         internal_urcu_unlock();
 297 }
 298
 299 /*
 300  * library wrappers to be used by non-LGPL compatible source code.
 301  */
 302
 303 void rcu_read_lock(void)
 304 {
 305         _rcu_read_lock();
 306 }
 307
 308 void rcu_read_unlock(void)
 309 {
 310         _rcu_read_unlock();
 311 }
 312
 313 void *rcu_dereference(void *p)
 314 {
 315         return _rcu_dereference(p);
 316 }
 317
 318 void *rcu_assign_pointer_sym(void **p, void *v)
 319 {
 320         wmb();
 321         return STORE_SHARED(p, v);
 322 }
 323
 324 void *rcu_xchg_pointer_sym(void **p, void *v)
 325 {
 326         wmb();
 327         return xchg(p, v);
 328 }
 329
 330 void *rcu_publish_content_sym(void **p, void *v)
 331 {
 332         void *oldptr;
 333
 334         oldptr = _rcu_xchg_pointer(p, v);
 335         synchronize_rcu();
 336         return oldptr;
 337 }
 338
 339 static void rcu_add_reader(pthread_t id)
 340 {
 341         struct reader_registry *oldarray;
 342
 343         if (!registry) {
 344                 alloc_readers = INIT_NUM_THREADS;
 345                 num_readers = 0;
 346                 registry =
 347                         malloc(sizeof(struct reader_registry) * alloc_readers);
 348         }
 349         if (alloc_readers < num_readers + 1) {
 350                 oldarray = registry;
 351                 registry = malloc(sizeof(struct reader_registry)
 352                                 * (alloc_readers << 1));
 353                 memcpy(registry, oldarray,
 354                         sizeof(struct reader_registry) * alloc_readers);
 355                 alloc_readers <<= 1;
 356                 free(oldarray);
 357         }
 358         registry[num_readers].tid = id;
 359         /* reference to the TLS of _this_ reader thread. */
 360         registry[num_readers].urcu_active_readers = &urcu_active_readers;
 361         registry[num_readers].need_mb = &need_mb;
 362         num_readers++;
 363 }
 364
 365 /*
 366  * Never shrink (implementation limitation).
 367  * This is O(nb threads). Eventually use a hash table.
 368  */
 369 static void rcu_remove_reader(pthread_t id)
 370 {
 371         struct reader_registry *index;
 372
 373         assert(registry != NULL);
 374         for (index = registry; index < registry + num_readers; index++) {
 375                 if (pthread_equal(index->tid, id)) {
 376                         memcpy(index, &registry[num_readers - 1],
 377                                 sizeof(struct reader_registry));
 378                         registry[num_readers - 1].tid = 0;
 379                         registry[num_readers - 1].urcu_active_readers = NULL;
 380                         num_readers--;
 381                         return;
 382                 }
 383         }
 384         /* Hrm not found, forgot to register ? */
 385         assert(0);
 386 }
 387
 388 void rcu_register_thread(void)
 389 {
 390         internal_urcu_lock();
 391         rcu_add_reader(pthread_self());
 392         internal_urcu_unlock();
 393 }
 394
 395 void rcu_unregister_thread(void)
 396 {
 397         internal_urcu_lock();
 398         rcu_remove_reader(pthread_self());
 399         internal_urcu_unlock();
 400 }
 401
 402 #ifndef DEBUG_FULL_MB
 403 static void sigurcu_handler(int signo, siginfo_t *siginfo, void *context)
 404 {
 405         /*
 406          * Executing this smp_mb() is the only purpose of this signal handler.
 407          * It punctually promotes barrier() into smp_mb() on every thread it is
 408          * executed on.
 409          */
 410         smp_mb();
 411         need_mb = 0;
 412         smp_mb();
 413 }
 414
 415 void __attribute__((constructor)) urcu_init(void)
 416 {
 417         struct sigaction act;
 418         int ret;
 419
 420         act.sa_sigaction = sigurcu_handler;
 421         ret = sigaction(SIGURCU, &act, NULL);
 422         if (ret) {
 423                 perror("Error in sigaction");
 424                 exit(-1);
 425         }
 426 }
 427
 428 void __attribute__((destructor)) urcu_exit(void)
 429 {
 430         struct sigaction act;
 431         int ret;
 432
 433         ret = sigaction(SIGURCU, NULL, &act);
 434         if (ret) {
 435                 perror("Error in sigaction");
 436                 exit(-1);
 437         }
 438         assert(act.sa_sigaction == sigurcu_handler);
 439         free(registry);
 440 }
 441 #endif /* #ifndef DEBUG_FULL_MB */