urcu-qsbr.c

   1 /*
   2  * urcu-qsbr.c
   3  *
   4  * Userspace RCU QSBR 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 #define BUILD_QSBR_LIB
  36 #include "urcu-qsbr-static.h"
  37 /* Do not #define _LGPL_SOURCE to ensure we can emit the wrapper symbols */
  38 #include "urcu-qsbr.h"
  39
  40 static pthread_mutex_t urcu_mutex = PTHREAD_MUTEX_INITIALIZER;
  41
  42 /*
  43  * Global grace period counter.
  44  */
  45 unsigned long urcu_gp_ctr = RCU_GP_ONLINE;
  46
  47 /*
  48  * Written to only by each individual reader. Read by both the reader and the
  49  * writers.
  50  */
  51 struct urcu_reader_status __thread urcu_reader_status;
  52
  53 /* Thread IDs of registered readers */
  54 #define INIT_NUM_THREADS 4
  55
  56 struct reader_registry {
  57         pthread_t tid;
  58         struct urcu_reader_status *urcu_reader_status;
  59 };
  60
  61 #ifdef DEBUG_YIELD
  62 unsigned int yield_active;
  63 unsigned int __thread rand_yield;
  64 #endif
  65
  66 static struct reader_registry *registry;
  67 static int num_readers, alloc_readers;
  68
  69 static void internal_urcu_lock(void)
  70 {
  71         int ret;
  72
  73 #ifndef DISTRUST_SIGNALS_EXTREME
  74         ret = pthread_mutex_lock(&urcu_mutex);
  75         if (ret) {
  76                 perror("Error in pthread mutex lock");
  77                 exit(-1);
  78         }
  79 #else /* #ifndef DISTRUST_SIGNALS_EXTREME */
  80         while ((ret = pthread_mutex_trylock(&urcu_mutex)) != 0) {
  81                 if (ret != EBUSY && ret != EINTR) {
  82                         printf("ret = %d, errno = %d\n", ret, errno);
  83                         perror("Error in pthread mutex lock");
  84                         exit(-1);
  85                 }
  86                 poll(NULL,0,10);
  87         }
  88 #endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */
  89 }
  90
  91 static void internal_urcu_unlock(void)
  92 {
  93         int ret;
  94
  95         ret = pthread_mutex_unlock(&urcu_mutex);
  96         if (ret) {
  97                 perror("Error in pthread mutex unlock");
  98                 exit(-1);
  99         }
 100 }
 101
 102 /*
 103  * synchronize_rcu() waiting. Single thread.
 104  */
 105 static void wait_for_quiescent_state(void)
 106 {
 107         struct reader_registry *index;
 108
 109         if (!registry)
 110                 return;
 111         /*
 112          * Wait for each thread rcu_reader qs_gp count to become 0.
 113          */
 114         for (index = registry; index < registry + num_readers; index++) {
 115                 int wait_loops = 0;
 116
 117                 index->urcu_reader_status->gp_waiting = 1;
 118                 while (rcu_gp_ongoing(&index->urcu_reader_status->qs_gp)) {
 119                         if (wait_loops++ == RCU_QS_ACTIVE_ATTEMPTS) {
 120                                 sched_yield();  /* ideally sched_yield_to() */
 121                                 wait_loops = 0;
 122                         } else {
 123 #ifndef HAS_INCOHERENT_CACHES
 124                                 cpu_relax();
 125 #else /* #ifndef HAS_INCOHERENT_CACHES */
 126                                 smp_mb();
 127 #endif /* #else #ifndef HAS_INCOHERENT_CACHES */
 128                         }
 129                 }
 130                 index->urcu_reader_status->gp_waiting = 0;
 131         }
 132 }
 133
 134 /*
 135  * Using a two-subphases algorithm for architectures with smaller than 64-bit
 136  * long-size to ensure we do not encounter an overflow bug.
 137  */
 138
 139 #if (BITS_PER_LONG < 64)
 140 /*
 141  * called with urcu_mutex held.
 142  */
 143 static void switch_next_urcu_qparity(void)
 144 {
 145         STORE_SHARED(urcu_gp_ctr, urcu_gp_ctr ^ RCU_GP_CTR);
 146 }
 147
 148 void synchronize_rcu(void)
 149 {
 150         unsigned long was_online;
 151
 152         was_online = urcu_reader_status.qs_gp;
 153
 154         /* All threads should read qparity before accessing data structure
 155          * where new ptr points to.
 156          */
 157         /* Write new ptr before changing the qparity */
 158         smp_mb();
 159
 160         /*
 161          * Mark the writer thread offline to make sure we don't wait for
 162          * our own quiescent state. This allows using synchronize_rcu() in
 163          * threads registered as readers.
 164          */
 165         if (was_online)
 166                 STORE_SHARED(urcu_reader_status.qs_gp, 0);
 167
 168         internal_urcu_lock();
 169
 170         STORE_SHARED(urcu_gp_ctr, urcu_gp_ctr ^ RCU_GP_ONGOING);
 171
 172         switch_next_urcu_qparity();     /* 0 -> 1 */
 173
 174         /*
 175          * Must commit qparity update to memory before waiting for parity
 176          * 0 quiescent state. Failure to do so could result in the writer
 177          * waiting forever while new readers are always accessing data (no
 178          * progress).
 179          * Ensured by STORE_SHARED and LOAD_SHARED.
 180          */
 181
 182         /*
 183          * Wait for previous parity to be empty of readers.
 184          */
 185         wait_for_quiescent_state();     /* Wait readers in parity 0 */
 186
 187         /*
 188          * Must finish waiting for quiescent state for parity 0 before
 189          * committing qparity update to memory. Failure to do so could result in
 190          * the writer waiting forever while new readers are always accessing
 191          * data (no progress).
 192          * Ensured by STORE_SHARED and LOAD_SHARED.
 193          */
 194
 195         switch_next_urcu_qparity();     /* 1 -> 0 */
 196
 197         /*
 198          * Must commit qparity update to memory before waiting for parity
 199          * 1 quiescent state. Failure to do so could result in the writer
 200          * waiting forever while new readers are always accessing data (no
 201          * progress).
 202          * Ensured by STORE_SHARED and LOAD_SHARED.
 203          */
 204
 205         /*
 206          * Wait for previous parity to be empty of readers.
 207          */
 208         wait_for_quiescent_state();     /* Wait readers in parity 1 */
 209
 210         STORE_SHARED(urcu_gp_ctr, urcu_gp_ctr ^ RCU_GP_ONGOING);
 211
 212         internal_urcu_unlock();
 213
 214         /*
 215          * Finish waiting for reader threads before letting the old ptr being
 216          * freed.
 217          */
 218         if (was_online)
 219                 _STORE_SHARED(urcu_reader_status.qs_gp,
 220                               LOAD_SHARED(urcu_gp_ctr));
 221         smp_mb();
 222 }
 223 #else /* !(BITS_PER_LONG < 64) */
 224 void synchronize_rcu(void)
 225 {
 226         unsigned long was_online;
 227
 228         was_online = urcu_reader_status.qs_gp;
 229
 230         /*
 231          * Mark the writer thread offline to make sure we don't wait for
 232          * our own quiescent state. This allows using synchronize_rcu() in
 233          * threads registered as readers.
 234          */
 235         smp_mb();
 236         if (was_online)
 237                 STORE_SHARED(urcu_reader_status.qs_gp, 0);
 238
 239         internal_urcu_lock();
 240         STORE_SHARED(urcu_gp_ctr, urcu_gp_ctr ^ RCU_GP_ONGOING);
 241         STORE_SHARED(urcu_gp_ctr, urcu_gp_ctr + RCU_GP_CTR);
 242         wait_for_quiescent_state();
 243         STORE_SHARED(urcu_gp_ctr, urcu_gp_ctr ^ RCU_GP_ONGOING);
 244         internal_urcu_unlock();
 245
 246         if (was_online)
 247                 _STORE_SHARED(urcu_reader_status.qs_gp,
 248                               LOAD_SHARED(urcu_gp_ctr));
 249         smp_mb();
 250 }
 251 #endif  /* !(BITS_PER_LONG < 64) */
 252
 253 /*
 254  * library wrappers to be used by non-LGPL compatible source code.
 255  */
 256
 257 void rcu_read_lock(void)
 258 {
 259         _rcu_read_lock();
 260 }
 261
 262 void rcu_read_unlock(void)
 263 {
 264         _rcu_read_unlock();
 265 }
 266
 267 void *rcu_dereference(void *p)
 268 {
 269         return _rcu_dereference(p);
 270 }
 271
 272 void *rcu_assign_pointer_sym(void **p, void *v)
 273 {
 274         wmb();
 275         return STORE_SHARED(p, v);
 276 }
 277
 278 void *rcu_cmpxchg_pointer_sym(void **p, void *old, void *_new)
 279 {
 280         wmb();
 281         return cmpxchg(p, old, _new);
 282 }
 283
 284 void *rcu_xchg_pointer_sym(void **p, void *v)
 285 {
 286         wmb();
 287         return xchg(p, v);
 288 }
 289
 290 void *rcu_publish_content_sym(void **p, void *v)
 291 {
 292         void *oldptr;
 293
 294         oldptr = _rcu_xchg_pointer(p, v);
 295         synchronize_rcu();
 296         return oldptr;
 297 }
 298
 299 void rcu_quiescent_state(void)
 300 {
 301         _rcu_quiescent_state();
 302 }
 303
 304 void rcu_thread_offline(void)
 305 {
 306         _rcu_thread_offline();
 307 }
 308
 309 void rcu_thread_online(void)
 310 {
 311         _rcu_thread_online();
 312 }
 313
 314 static void rcu_add_reader(pthread_t id)
 315 {
 316         struct reader_registry *oldarray;
 317
 318         if (!registry) {
 319                 alloc_readers = INIT_NUM_THREADS;
 320                 num_readers = 0;
 321                 registry =
 322                         malloc(sizeof(struct reader_registry) * alloc_readers);
 323         }
 324         if (alloc_readers < num_readers + 1) {
 325                 oldarray = registry;
 326                 registry = malloc(sizeof(struct reader_registry)
 327                                 * (alloc_readers << 1));
 328                 memcpy(registry, oldarray,
 329                         sizeof(struct reader_registry) * alloc_readers);
 330                 alloc_readers <<= 1;
 331                 free(oldarray);
 332         }
 333         registry[num_readers].tid = id;
 334         /* reference to the TLS of _this_ reader thread. */
 335         registry[num_readers].urcu_reader_status = &urcu_reader_status;
 336         num_readers++;
 337 }
 338
 339 /*
 340  * Never shrink (implementation limitation).
 341  * This is O(nb threads). Eventually use a hash table.
 342  */
 343 static void rcu_remove_reader(pthread_t id)
 344 {
 345         struct reader_registry *index;
 346
 347         assert(registry != NULL);
 348         for (index = registry; index < registry + num_readers; index++) {
 349                 if (pthread_equal(index->tid, id)) {
 350                         memcpy(index, &registry[num_readers - 1],
 351                                 sizeof(struct reader_registry));
 352                         registry[num_readers - 1].tid = 0;
 353                         registry[num_readers - 1].urcu_reader_status = NULL;
 354                         num_readers--;
 355                         return;
 356                 }
 357         }
 358         /* Hrm not found, forgot to register ? */
 359         assert(0);
 360 }
 361
 362 void rcu_register_thread(void)
 363 {
 364         internal_urcu_lock();
 365         rcu_add_reader(pthread_self());
 366         internal_urcu_unlock();
 367         _rcu_thread_online();
 368 }
 369
 370 void rcu_unregister_thread(void)
 371 {
 372         /*
 373          * We have to make the thread offline otherwise we end up dealocking
 374          * with a waiting writer.
 375          */
 376         _rcu_thread_offline();
 377         internal_urcu_lock();
 378         rcu_remove_reader(pthread_self());
 379         internal_urcu_unlock();
 380 }