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 void __attribute__((destructor)) rcu_exit(void);
  41
  42 static pthread_mutex_t rcu_gp_lock = PTHREAD_MUTEX_INITIALIZER;
  43
  44 int gp_futex;
  45
  46 /*
  47  * Global grace period counter.
  48  */
  49 unsigned long rcu_gp_ctr = RCU_GP_ONLINE;
  50
  51 /*
  52  * Written to only by each individual reader. Read by both the reader and the
  53  * writers.
  54  */
  55 struct rcu_reader __thread rcu_reader;
  56
  57 #ifdef DEBUG_YIELD
  58 unsigned int yield_active;
  59 unsigned int __thread rand_yield;
  60 #endif
  61
  62 static LIST_HEAD(registry);
  63
  64 static void mutex_lock(pthread_mutex_t *mutex)
  65 {
  66         int ret;
  67
  68 #ifndef DISTRUST_SIGNALS_EXTREME
  69         ret = pthread_mutex_lock(mutex);
  70         if (ret) {
  71                 perror("Error in pthread mutex lock");
  72                 exit(-1);
  73         }
  74 #else /* #ifndef DISTRUST_SIGNALS_EXTREME */
  75         while ((ret = pthread_mutex_trylock(mutex)) != 0) {
  76                 if (ret != EBUSY && ret != EINTR) {
  77                         printf("ret = %d, errno = %d\n", ret, errno);
  78                         perror("Error in pthread mutex lock");
  79                         exit(-1);
  80                 }
  81                 poll(NULL,0,10);
  82         }
  83 #endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */
  84 }
  85
  86 static void mutex_unlock(pthread_mutex_t *mutex)
  87 {
  88         int ret;
  89
  90         ret = pthread_mutex_unlock(mutex);
  91         if (ret) {
  92                 perror("Error in pthread mutex unlock");
  93                 exit(-1);
  94         }
  95 }
  96
  97 /*
  98  * synchronize_rcu() waiting. Single thread.
  99  */
 100 static void wait_gp(void)
 101 {
 102         /* Read reader_gp before read futex */
 103         smp_rmb();
 104         if (uatomic_read(&gp_futex) == -1)
 105                 futex_noasync(&gp_futex, FUTEX_WAIT, -1,
 106                       NULL, NULL, 0);
 107 }
 108
 109 static void update_counter_and_wait(void)
 110 {
 111         LIST_HEAD(qsreaders);
 112         int wait_loops = 0;
 113         struct rcu_reader *index, *tmp;
 114
 115 #if (BITS_PER_LONG < 64)
 116         /* Switch parity: 0 -> 1, 1 -> 0 */
 117         STORE_SHARED(rcu_gp_ctr, rcu_gp_ctr ^ RCU_GP_CTR);
 118 #else   /* !(BITS_PER_LONG < 64) */
 119         /* Increment current G.P. */
 120         STORE_SHARED(rcu_gp_ctr, rcu_gp_ctr + RCU_GP_CTR);
 121 #endif  /* !(BITS_PER_LONG < 64) */
 122
 123         /*
 124          * Wait for each thread rcu_reader_qs_gp count to become 0.
 125          */
 126         for (;;) {
 127                 wait_loops++;
 128                 if (wait_loops == RCU_QS_ACTIVE_ATTEMPTS) {
 129                         uatomic_dec(&gp_futex);
 130                         /* Write futex before read reader_gp */
 131                         smp_mb();
 132                 }
 133
 134                 list_for_each_entry_safe(index, tmp, &registry, head) {
 135                         if (!rcu_gp_ongoing(&index->ctr))
 136                                 list_move(&index->head, &qsreaders);
 137                 }
 138
 139                 if (list_empty(&registry)) {
 140                         if (wait_loops == RCU_QS_ACTIVE_ATTEMPTS) {
 141                                 /* Read reader_gp before write futex */
 142                                 smp_mb();
 143                                 uatomic_set(&gp_futex, 0);
 144                         }
 145                         break;
 146                 } else {
 147                         if (wait_loops == RCU_QS_ACTIVE_ATTEMPTS) {
 148                                 wait_gp();
 149                         } else {
 150 #ifndef HAS_INCOHERENT_CACHES
 151                                 cpu_relax();
 152 #else /* #ifndef HAS_INCOHERENT_CACHES */
 153                                 smp_mb();
 154 #endif /* #else #ifndef HAS_INCOHERENT_CACHES */
 155                         }
 156                 }
 157         }
 158         /* put back the reader list in the registry */
 159         list_splice(&qsreaders, &registry);
 160 }
 161
 162 /*
 163  * Using a two-subphases algorithm for architectures with smaller than 64-bit
 164  * long-size to ensure we do not encounter an overflow bug.
 165  */
 166
 167 #if (BITS_PER_LONG < 64)
 168 void synchronize_rcu(void)
 169 {
 170         unsigned long was_online;
 171
 172         was_online = rcu_reader.ctr;
 173
 174         /* All threads should read qparity before accessing data structure
 175          * where new ptr points to.
 176          */
 177         /* Write new ptr before changing the qparity */
 178         smp_mb();
 179
 180         /*
 181          * Mark the writer thread offline to make sure we don't wait for
 182          * our own quiescent state. This allows using synchronize_rcu() in
 183          * threads registered as readers.
 184          */
 185         if (was_online)
 186                 STORE_SHARED(rcu_reader.ctr, 0);
 187
 188         mutex_lock(&rcu_gp_lock);
 189
 190         if (list_empty(&registry))
 191                 goto out;
 192
 193         /*
 194          * Wait for previous parity to be empty of readers.
 195          */
 196         update_counter_and_wait();      /* 0 -> 1, wait readers in parity 0 */
 197
 198         /*
 199          * Must finish waiting for quiescent state for parity 0 before
 200          * committing qparity update to memory. Failure to do so could result in
 201          * the writer waiting forever while new readers are always accessing
 202          * data (no progress).
 203          * Ensured by STORE_SHARED and LOAD_SHARED.
 204          */
 205
 206         /*
 207          * Adding a smp_mb() which is _not_ formally required, but makes the
 208          * model easier to understand. It does not have a big performance impact
 209          * anyway, given this is the write-side.
 210          */
 211         smp_mb();
 212
 213         /*
 214          * Wait for previous parity to be empty of readers.
 215          */
 216         update_counter_and_wait();      /* 1 -> 0, wait readers in parity 1 */
 217 out:
 218         mutex_unlock(&rcu_gp_lock);
 219
 220         /*
 221          * Finish waiting for reader threads before letting the old ptr being
 222          * freed.
 223          */
 224         if (was_online)
 225                 _STORE_SHARED(rcu_reader.ctr, LOAD_SHARED(rcu_gp_ctr));
 226         smp_mb();
 227 }
 228 #else /* !(BITS_PER_LONG < 64) */
 229 void synchronize_rcu(void)
 230 {
 231         unsigned long was_online;
 232
 233         was_online = rcu_reader.ctr;
 234
 235         /*
 236          * Mark the writer thread offline to make sure we don't wait for
 237          * our own quiescent state. This allows using synchronize_rcu() in
 238          * threads registered as readers.
 239          */
 240         smp_mb();
 241         if (was_online)
 242                 STORE_SHARED(rcu_reader.ctr, 0);
 243
 244         mutex_lock(&rcu_gp_lock);
 245         if (list_empty(&registry))
 246                 goto out;
 247         update_counter_and_wait();
 248 out:
 249         mutex_unlock(&rcu_gp_lock);
 250
 251         if (was_online)
 252                 _STORE_SHARED(rcu_reader.ctr, LOAD_SHARED(rcu_gp_ctr));
 253         smp_mb();
 254 }
 255 #endif  /* !(BITS_PER_LONG < 64) */
 256
 257 /*
 258  * library wrappers to be used by non-LGPL compatible source code.
 259  */
 260
 261 void rcu_read_lock(void)
 262 {
 263         _rcu_read_lock();
 264 }
 265
 266 void rcu_read_unlock(void)
 267 {
 268         _rcu_read_unlock();
 269 }
 270
 271 void rcu_quiescent_state(void)
 272 {
 273         _rcu_quiescent_state();
 274 }
 275
 276 void rcu_thread_offline(void)
 277 {
 278         _rcu_thread_offline();
 279 }
 280
 281 void rcu_thread_online(void)
 282 {
 283         _rcu_thread_online();
 284 }
 285
 286 void rcu_register_thread(void)
 287 {
 288         rcu_reader.tid = pthread_self();
 289         assert(rcu_reader.ctr == 0);
 290
 291         mutex_lock(&rcu_gp_lock);
 292         list_add(&rcu_reader.head, &registry);
 293         mutex_unlock(&rcu_gp_lock);
 294         _rcu_thread_online();
 295 }
 296
 297 void rcu_unregister_thread(void)
 298 {
 299         /*
 300          * We have to make the thread offline otherwise we end up dealocking
 301          * with a waiting writer.
 302          */
 303         _rcu_thread_offline();
 304         mutex_lock(&rcu_gp_lock);
 305         list_del(&rcu_reader.head);
 306         mutex_unlock(&rcu_gp_lock);
 307 }
 308
 309 void rcu_exit(void)
 310 {
 311         assert(list_empty(&registry));
 312 }