urcu.c

   1 /*
   2  * urcu.c
   3  *
   4  * Userspace RCU library
   5  *
   6  * Copyright February 2009 - Mathieu Desnoyers <mathieu.desnoyers@polymtl.ca>
   7  *
   8  * Distributed under GPLv2
   9  */
  10
  11 #include <stdio.h>
  12 #include <pthread.h>
  13 #include <signal.h>
  14 #include <assert.h>
  15 #include <stdlib.h>
  16 #include <string.h>
  17
  18 #include "urcu.h"
  19
  20 pthread_mutex_t urcu_mutex = PTHREAD_MUTEX_INITIALIZER;
  21
  22 /*
  23  * Global grace period counter.
  24  * Contains the current RCU_GP_CTR_BIT.
  25  * Also has a RCU_GP_CTR_BIT of 1, to accelerate the reader fast path.
  26  */
  27 long urcu_gp_ctr = RCU_GP_COUNT;
  28
  29 long __thread urcu_active_readers;
  30
  31 /* Thread IDs of registered readers */
  32 #define INIT_NUM_THREADS 4
  33
  34 struct reader_data {
  35         pthread_t tid;
  36         long *urcu_active_readers;
  37 };
  38
  39 #ifdef DEBUG_YIELD
  40 unsigned int yield_active;
  41 unsigned int __thread rand_yield;
  42 #endif
  43
  44 static struct reader_data *reader_data;
  45 static int num_readers, alloc_readers;
  46 #ifndef DEBUG_FULL_MB
  47 static int sig_done;
  48 #endif
  49
  50 void internal_urcu_lock(void)
  51 {
  52         int ret;
  53         ret = pthread_mutex_lock(&urcu_mutex);
  54         if (ret) {
  55                 perror("Error in pthread mutex lock");
  56                 exit(-1);
  57         }
  58 }
  59
  60 void internal_urcu_unlock(void)
  61 {
  62         int ret;
  63
  64         ret = pthread_mutex_unlock(&urcu_mutex);
  65         if (ret) {
  66                 perror("Error in pthread mutex unlock");
  67                 exit(-1);
  68         }
  69 }
  70
  71 /*
  72  * called with urcu_mutex held.
  73  */
  74 static void switch_next_urcu_qparity(void)
  75 {
  76         urcu_gp_ctr ^= RCU_GP_CTR_BIT;
  77 }
  78
  79 #ifdef DEBUG_FULL_MB
  80 static void force_mb_single_thread(pthread_t tid)
  81 {
  82         smp_mb();
  83 }
  84
  85 static void force_mb_all_threads(void)
  86 {
  87         smp_mb();
  88 }
  89 #else
  90
  91 static void force_mb_single_thread(pthread_t tid)
  92 {
  93         assert(reader_data);
  94         sig_done = 0;
  95         /*
  96          * pthread_kill has a smp_mb(). But beware, we assume it performs
  97          * a cache flush on architectures with non-coherent cache. Let's play
  98          * safe and don't assume anything : we use smp_mc() to make sure the
  99          * cache flush is enforced.
 100          * smp_mb();    write sig_done before sending the signals
 101          */
 102         smp_mc();       /* write sig_done before sending the signals */
 103         pthread_kill(tid, SIGURCU);
 104         /*
 105          * Wait for sighandler (and thus mb()) to execute on every thread.
 106          * BUSY-LOOP.
 107          */
 108         while (LOAD_SHARED(sig_done) < 1)
 109                 cpu_relax();
 110         smp_mb();       /* read sig_done before ending the barrier */
 111 }
 112
 113 static void force_mb_all_threads(void)
 114 {
 115         struct reader_data *index;
 116         /*
 117          * Ask for each threads to execute a smp_mb() so we can consider the
 118          * compiler barriers around rcu read lock as real memory barriers.
 119          */
 120         if (!reader_data)
 121                 return;
 122         sig_done = 0;
 123         /*
 124          * pthread_kill has a smp_mb(). But beware, we assume it performs
 125          * a cache flush on architectures with non-coherent cache. Let's play
 126          * safe and don't assume anything : we use smp_mc() to make sure the
 127          * cache flush is enforced.
 128          * smp_mb();    write sig_done before sending the signals
 129          */
 130         smp_mc();       /* write sig_done before sending the signals */
 131         for (index = reader_data; index < reader_data + num_readers; index++)
 132                 pthread_kill(index->tid, SIGURCU);
 133         /*
 134          * Wait for sighandler (and thus mb()) to execute on every thread.
 135          * BUSY-LOOP.
 136          */
 137         while (LOAD_SHARED(sig_done) < num_readers)
 138                 cpu_relax();
 139         smp_mb();       /* read sig_done before ending the barrier */
 140 }
 141 #endif
 142
 143 void wait_for_quiescent_state(void)
 144 {
 145         struct reader_data *index;
 146
 147         if (!reader_data)
 148                 return;
 149         /*
 150          * Wait for each thread urcu_active_readers count to become 0.
 151          */
 152         for (index = reader_data; index < reader_data + num_readers; index++) {
 153                 int wait_loops = 0;
 154                 /*
 155                  * BUSY-LOOP. Force the reader thread to commit its
 156                  * urcu_active_readers update to memory if we wait for too long.
 157                  */
 158                 while (rcu_old_gp_ongoing(index->urcu_active_readers)) {
 159                         if (wait_loops++ == KICK_READER_LOOPS) {
 160                                 force_mb_single_thread(index->tid);
 161                                 wait_loops = 0;
 162                         } else {
 163                                 cpu_relax();
 164                         }
 165                 }
 166         }
 167 }
 168
 169 void synchronize_rcu(void)
 170 {
 171         internal_urcu_lock();
 172
 173         /* All threads should read qparity before accessing data structure
 174          * where new ptr points to. Must be done within internal_urcu_lock
 175          * because it iterates on reader threads.*/
 176         /* Write new ptr before changing the qparity */
 177         force_mb_all_threads();
 178
 179         switch_next_urcu_qparity();     /* 0 -> 1 */
 180
 181         /*
 182          * Must commit qparity update to memory before waiting for parity
 183          * 0 quiescent state. Failure to do so could result in the writer
 184          * waiting forever while new readers are always accessing data (no
 185          * progress).
 186          */
 187         smp_mc();
 188
 189         /*
 190          * Wait for previous parity to be empty of readers.
 191          */
 192         wait_for_quiescent_state();     /* Wait readers in parity 0 */
 193
 194         /*
 195          * Must finish waiting for quiescent state for parity 0 before
 196          * committing qparity update to memory. Failure to do so could result in
 197          * the writer waiting forever while new readers are always accessing
 198          * data (no progress).
 199          */
 200         smp_mc();
 201
 202         switch_next_urcu_qparity();     /* 1 -> 0 */
 203
 204         /*
 205          * Must commit qparity update to memory before waiting for parity
 206          * 1 quiescent state. Failure to do so could result in the writer
 207          * waiting forever while new readers are always accessing data (no
 208          * progress).
 209          */
 210         smp_mc();
 211
 212         /*
 213          * Wait for previous parity to be empty of readers.
 214          */
 215         wait_for_quiescent_state();     /* Wait readers in parity 1 */
 216
 217         /* Finish waiting for reader threads before letting the old ptr being
 218          * freed. Must be done within internal_urcu_lock because it iterates on
 219          * reader threads. */
 220         force_mb_all_threads();
 221
 222         internal_urcu_unlock();
 223 }
 224
 225 void urcu_add_reader(pthread_t id)
 226 {
 227         struct reader_data *oldarray;
 228
 229         if (!reader_data) {
 230                 alloc_readers = INIT_NUM_THREADS;
 231                 num_readers = 0;
 232                 reader_data =
 233                         malloc(sizeof(struct reader_data) * alloc_readers);
 234         }
 235         if (alloc_readers < num_readers + 1) {
 236                 oldarray = reader_data;
 237                 reader_data = malloc(sizeof(struct reader_data)
 238                                 * (alloc_readers << 1));
 239                 memcpy(reader_data, oldarray,
 240                         sizeof(struct reader_data) * alloc_readers);
 241                 alloc_readers <<= 1;
 242                 free(oldarray);
 243         }
 244         reader_data[num_readers].tid = id;
 245         /* reference to the TLS of _this_ reader thread. */
 246         reader_data[num_readers].urcu_active_readers = &urcu_active_readers;
 247         num_readers++;
 248 }
 249
 250 /*
 251  * Never shrink (implementation limitation).
 252  * This is O(nb threads). Eventually use a hash table.
 253  */
 254 void urcu_remove_reader(pthread_t id)
 255 {
 256         struct reader_data *index;
 257
 258         assert(reader_data != NULL);
 259         for (index = reader_data; index < reader_data + num_readers; index++) {
 260                 if (pthread_equal(index->tid, id)) {
 261                         memcpy(index, &reader_data[num_readers - 1],
 262                                 sizeof(struct reader_data));
 263                         reader_data[num_readers - 1].tid = 0;
 264                         reader_data[num_readers - 1].urcu_active_readers = NULL;
 265                         num_readers--;
 266                         return;
 267                 }
 268         }
 269         /* Hrm not found, forgot to register ? */
 270         assert(0);
 271 }
 272
 273 void urcu_register_thread(void)
 274 {
 275         internal_urcu_lock();
 276         urcu_add_reader(pthread_self());
 277         internal_urcu_unlock();
 278 }
 279
 280 void urcu_unregister_thread(void)
 281 {
 282         internal_urcu_lock();
 283         urcu_remove_reader(pthread_self());
 284         internal_urcu_unlock();
 285 }
 286
 287 #ifndef DEBUG_FULL_MB
 288 void sigurcu_handler(int signo, siginfo_t *siginfo, void *context)
 289 {
 290         /*
 291          * Executing this smp_mb() is the only purpose of this signal handler.
 292          * It punctually promotes barrier() into smp_mb() on every thread it is
 293          * executed on.
 294          */
 295         smp_mb();
 296         atomic_inc(&sig_done);
 297 }
 298
 299 void __attribute__((constructor)) urcu_init(void)
 300 {
 301         struct sigaction act;
 302         int ret;
 303
 304         act.sa_sigaction = sigurcu_handler;
 305         ret = sigaction(SIGURCU, &act, NULL);
 306         if (ret) {
 307                 perror("Error in sigaction");
 308                 exit(-1);
 309         }
 310 }
 311
 312 void __attribute__((destructor)) urcu_exit(void)
 313 {
 314         struct sigaction act;
 315         int ret;
 316
 317         ret = sigaction(SIGURCU, NULL, &act);
 318         if (ret) {
 319                 perror("Error in sigaction");
 320                 exit(-1);
 321         }
 322         assert(act.sa_sigaction == sigurcu_handler);
 323         free(reader_data);
 324 }
 325 #endif