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  * Written to only by writer with mutex taken. Read by both writer and readers.
  27  */
  28 long urcu_gp_ctr = RCU_GP_COUNT;
  29
  30 /*
  31  * Written to only by each individual reader. Read by both the reader and the
  32  * writers.
  33  */
  34 long __thread urcu_active_readers;
  35
  36 /* Thread IDs of registered readers */
  37 #define INIT_NUM_THREADS 4
  38
  39 struct reader_data {
  40         pthread_t tid;
  41         long *urcu_active_readers;
  42 };
  43
  44 #ifdef DEBUG_YIELD
  45 unsigned int yield_active;
  46 unsigned int __thread rand_yield;
  47 #endif
  48
  49 static struct reader_data *reader_data;
  50 static int num_readers, alloc_readers;
  51 #ifndef DEBUG_FULL_MB
  52 static int sig_done;
  53 #endif
  54
  55 void internal_urcu_lock(void)
  56 {
  57         int ret;
  58         ret = pthread_mutex_lock(&urcu_mutex);
  59         if (ret) {
  60                 perror("Error in pthread mutex lock");
  61                 exit(-1);
  62         }
  63 }
  64
  65 void internal_urcu_unlock(void)
  66 {
  67         int ret;
  68
  69         ret = pthread_mutex_unlock(&urcu_mutex);
  70         if (ret) {
  71                 perror("Error in pthread mutex unlock");
  72                 exit(-1);
  73         }
  74 }
  75
  76 /*
  77  * called with urcu_mutex held.
  78  */
  79 static void switch_next_urcu_qparity(void)
  80 {
  81         STORE_SHARED(urcu_gp_ctr, urcu_gp_ctr ^ RCU_GP_CTR_BIT);
  82 }
  83
  84 #ifdef DEBUG_FULL_MB
  85 static void force_mb_single_thread(pthread_t tid)
  86 {
  87         smp_mb();
  88 }
  89
  90 static void force_mb_all_threads(void)
  91 {
  92         smp_mb();
  93 }
  94 #else
  95
  96 static void force_mb_single_thread(pthread_t tid)
  97 {
  98         assert(reader_data);
  99         sig_done = 0;
 100         /*
 101          * pthread_kill has a smp_mb(). But beware, we assume it performs
 102          * a cache flush on architectures with non-coherent cache. Let's play
 103          * safe and don't assume anything : we use smp_mc() to make sure the
 104          * cache flush is enforced.
 105          * smp_mb();    write sig_done before sending the signals
 106          */
 107         smp_mc();       /* write sig_done before sending the signals */
 108         pthread_kill(tid, SIGURCU);
 109         /*
 110          * Wait for sighandler (and thus mb()) to execute on every thread.
 111          * BUSY-LOOP.
 112          */
 113         while (LOAD_SHARED(sig_done) < 1)
 114                 cpu_relax();
 115         smp_mb();       /* read sig_done before ending the barrier */
 116 }
 117
 118 static void force_mb_all_threads(void)
 119 {
 120         struct reader_data *index;
 121         /*
 122          * Ask for each threads to execute a smp_mb() so we can consider the
 123          * compiler barriers around rcu read lock as real memory barriers.
 124          */
 125         if (!reader_data)
 126                 return;
 127         sig_done = 0;
 128         /*
 129          * pthread_kill has a smp_mb(). But beware, we assume it performs
 130          * a cache flush on architectures with non-coherent cache. Let's play
 131          * safe and don't assume anything : we use smp_mc() to make sure the
 132          * cache flush is enforced.
 133          * smp_mb();    write sig_done before sending the signals
 134          */
 135         smp_mc();       /* write sig_done before sending the signals */
 136         for (index = reader_data; index < reader_data + num_readers; index++)
 137                 pthread_kill(index->tid, SIGURCU);
 138         /*
 139          * Wait for sighandler (and thus mb()) to execute on every thread.
 140          * BUSY-LOOP.
 141          */
 142         while (LOAD_SHARED(sig_done) < num_readers)
 143                 cpu_relax();
 144         smp_mb();       /* read sig_done before ending the barrier */
 145 }
 146 #endif
 147
 148 void wait_for_quiescent_state(void)
 149 {
 150         struct reader_data *index;
 151
 152         if (!reader_data)
 153                 return;
 154         /*
 155          * Wait for each thread urcu_active_readers count to become 0.
 156          */
 157         for (index = reader_data; index < reader_data + num_readers; index++) {
 158                 int wait_loops = 0;
 159                 /*
 160                  * BUSY-LOOP. Force the reader thread to commit its
 161                  * urcu_active_readers update to memory if we wait for too long.
 162                  */
 163                 while (rcu_old_gp_ongoing(index->urcu_active_readers)) {
 164                         if (wait_loops++ == KICK_READER_LOOPS) {
 165                                 force_mb_single_thread(index->tid);
 166                                 wait_loops = 0;
 167                         } else {
 168                                 cpu_relax();
 169                         }
 170                 }
 171         }
 172 }
 173
 174 void synchronize_rcu(void)
 175 {
 176         internal_urcu_lock();
 177
 178         /* All threads should read qparity before accessing data structure
 179          * where new ptr points to. Must be done within internal_urcu_lock
 180          * because it iterates on reader threads.*/
 181         /* Write new ptr before changing the qparity */
 182         force_mb_all_threads();
 183
 184         switch_next_urcu_qparity();     /* 0 -> 1 */
 185
 186         /*
 187          * Must commit qparity update to memory before waiting for parity
 188          * 0 quiescent state. Failure to do so could result in the writer
 189          * waiting forever while new readers are always accessing data (no
 190          * progress).
 191          * Ensured by STORE_SHARED and LOAD_SHARED.
 192          */
 193
 194         /*
 195          * Wait for previous parity to be empty of readers.
 196          */
 197         wait_for_quiescent_state();     /* Wait readers in parity 0 */
 198
 199         /*
 200          * Must finish waiting for quiescent state for parity 0 before
 201          * committing qparity update to memory. Failure to do so could result in
 202          * the writer waiting forever while new readers are always accessing
 203          * data (no progress).
 204          * Ensured by STORE_SHARED and LOAD_SHARED.
 205          */
 206
 207         switch_next_urcu_qparity();     /* 1 -> 0 */
 208
 209         /*
 210          * Must commit qparity update to memory before waiting for parity
 211          * 1 quiescent state. Failure to do so could result in the writer
 212          * waiting forever while new readers are always accessing data (no
 213          * progress).
 214          * Ensured by STORE_SHARED and LOAD_SHARED.
 215          */
 216
 217         /*
 218          * Wait for previous parity to be empty of readers.
 219          */
 220         wait_for_quiescent_state();     /* Wait readers in parity 1 */
 221
 222         /* Finish waiting for reader threads before letting the old ptr being
 223          * freed. Must be done within internal_urcu_lock because it iterates on
 224          * reader threads. */
 225         force_mb_all_threads();
 226
 227         internal_urcu_unlock();
 228 }
 229
 230 void urcu_add_reader(pthread_t id)
 231 {
 232         struct reader_data *oldarray;
 233
 234         if (!reader_data) {
 235                 alloc_readers = INIT_NUM_THREADS;
 236                 num_readers = 0;
 237                 reader_data =
 238                         malloc(sizeof(struct reader_data) * alloc_readers);
 239         }
 240         if (alloc_readers < num_readers + 1) {
 241                 oldarray = reader_data;
 242                 reader_data = malloc(sizeof(struct reader_data)
 243                                 * (alloc_readers << 1));
 244                 memcpy(reader_data, oldarray,
 245                         sizeof(struct reader_data) * alloc_readers);
 246                 alloc_readers <<= 1;
 247                 free(oldarray);
 248         }
 249         reader_data[num_readers].tid = id;
 250         /* reference to the TLS of _this_ reader thread. */
 251         reader_data[num_readers].urcu_active_readers = &urcu_active_readers;
 252         num_readers++;
 253 }
 254
 255 /*
 256  * Never shrink (implementation limitation).
 257  * This is O(nb threads). Eventually use a hash table.
 258  */
 259 void urcu_remove_reader(pthread_t id)
 260 {
 261         struct reader_data *index;
 262
 263         assert(reader_data != NULL);
 264         for (index = reader_data; index < reader_data + num_readers; index++) {
 265                 if (pthread_equal(index->tid, id)) {
 266                         memcpy(index, &reader_data[num_readers - 1],
 267                                 sizeof(struct reader_data));
 268                         reader_data[num_readers - 1].tid = 0;
 269                         reader_data[num_readers - 1].urcu_active_readers = NULL;
 270                         num_readers--;
 271                         return;
 272                 }
 273         }
 274         /* Hrm not found, forgot to register ? */
 275         assert(0);
 276 }
 277
 278 void urcu_register_thread(void)
 279 {
 280         internal_urcu_lock();
 281         urcu_add_reader(pthread_self());
 282         internal_urcu_unlock();
 283 }
 284
 285 void urcu_unregister_thread(void)
 286 {
 287         internal_urcu_lock();
 288         urcu_remove_reader(pthread_self());
 289         internal_urcu_unlock();
 290 }
 291
 292 #ifndef DEBUG_FULL_MB
 293 void sigurcu_handler(int signo, siginfo_t *siginfo, void *context)
 294 {
 295         /*
 296          * Executing this smp_mb() is the only purpose of this signal handler.
 297          * It punctually promotes barrier() into smp_mb() on every thread it is
 298          * executed on.
 299          */
 300         smp_mb();
 301         atomic_inc(&sig_done);
 302 }
 303
 304 void __attribute__((constructor)) urcu_init(void)
 305 {
 306         struct sigaction act;
 307         int ret;
 308
 309         act.sa_sigaction = sigurcu_handler;
 310         ret = sigaction(SIGURCU, &act, NULL);
 311         if (ret) {
 312                 perror("Error in sigaction");
 313                 exit(-1);
 314         }
 315 }
 316
 317 void __attribute__((destructor)) urcu_exit(void)
 318 {
 319         struct sigaction act;
 320         int ret;
 321
 322         ret = sigaction(SIGURCU, NULL, &act);
 323         if (ret) {
 324                 perror("Error in sigaction");
 325                 exit(-1);
 326         }
 327         assert(act.sa_sigaction == sigurcu_handler);
 328         free(reader_data);
 329 }
 330 #endif