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_registry {
  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_registry *registry;
  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 #if 0
  58         int ret;
  59         /* Mutex sleeping does not play well with busy-waiting loop. */
  60         ret = pthread_mutex_lock(&urcu_mutex);
  61         if (ret) {
  62                 perror("Error in pthread mutex lock");
  63                 exit(-1);
  64         }
  65 #endif
  66         while (pthread_mutex_trylock(&urcu_mutex) != 0)
  67                 cpu_relax();
  68 }
  69
  70 void internal_urcu_unlock(void)
  71 {
  72         int ret;
  73
  74         ret = pthread_mutex_unlock(&urcu_mutex);
  75         if (ret) {
  76                 perror("Error in pthread mutex unlock");
  77                 exit(-1);
  78         }
  79 }
  80
  81 /*
  82  * called with urcu_mutex held.
  83  */
  84 static void switch_next_urcu_qparity(void)
  85 {
  86         STORE_SHARED(urcu_gp_ctr, urcu_gp_ctr ^ RCU_GP_CTR_BIT);
  87 }
  88
  89 #ifdef DEBUG_FULL_MB
  90 static void force_mb_single_thread(pthread_t tid)
  91 {
  92         smp_mb();
  93 }
  94
  95 static void force_mb_all_threads(void)
  96 {
  97         smp_mb();
  98 }
  99 #else
 100
 101 static void force_mb_single_thread(pthread_t tid)
 102 {
 103         assert(registry);
 104         sig_done = 0;
 105         /*
 106          * pthread_kill has a smp_mb(). But beware, we assume it performs
 107          * a cache flush on architectures with non-coherent cache. Let's play
 108          * safe and don't assume anything : we use smp_mc() to make sure the
 109          * cache flush is enforced.
 110          * smp_mb();    write sig_done before sending the signals
 111          */
 112         smp_mc();       /* write sig_done before sending the signals */
 113         pthread_kill(tid, SIGURCU);
 114         /*
 115          * Wait for sighandler (and thus mb()) to execute on every thread.
 116          * BUSY-LOOP.
 117          */
 118         while (LOAD_SHARED(sig_done) < 1)
 119                 cpu_relax();
 120         smp_mb();       /* read sig_done before ending the barrier */
 121 }
 122
 123 static void force_mb_all_threads(void)
 124 {
 125         struct reader_registry *index;
 126         /*
 127          * Ask for each threads to execute a smp_mb() so we can consider the
 128          * compiler barriers around rcu read lock as real memory barriers.
 129          */
 130         if (!registry)
 131                 return;
 132         sig_done = 0;
 133         /*
 134          * pthread_kill has a smp_mb(). But beware, we assume it performs
 135          * a cache flush on architectures with non-coherent cache. Let's play
 136          * safe and don't assume anything : we use smp_mc() to make sure the
 137          * cache flush is enforced.
 138          * smp_mb();    write sig_done before sending the signals
 139          */
 140         smp_mc();       /* write sig_done before sending the signals */
 141         for (index = registry; index < registry + num_readers; index++)
 142                 pthread_kill(index->tid, SIGURCU);
 143         /*
 144          * Wait for sighandler (and thus mb()) to execute on every thread.
 145          * BUSY-LOOP.
 146          */
 147         while (LOAD_SHARED(sig_done) < num_readers)
 148                 cpu_relax();
 149         smp_mb();       /* read sig_done before ending the barrier */
 150 }
 151 #endif
 152
 153 void wait_for_quiescent_state(void)
 154 {
 155         struct reader_registry *index;
 156
 157         if (!registry)
 158                 return;
 159         /*
 160          * Wait for each thread urcu_active_readers count to become 0.
 161          */
 162         for (index = registry; index < registry + num_readers; index++) {
 163                 int wait_loops = 0;
 164                 /*
 165                  * BUSY-LOOP. Force the reader thread to commit its
 166                  * urcu_active_readers update to memory if we wait for too long.
 167                  */
 168                 while (rcu_old_gp_ongoing(index->urcu_active_readers)) {
 169                         if (wait_loops++ == KICK_READER_LOOPS) {
 170                                 force_mb_single_thread(index->tid);
 171                                 wait_loops = 0;
 172                         } else {
 173                                 cpu_relax();
 174                         }
 175                 }
 176         }
 177 }
 178
 179 void synchronize_rcu(void)
 180 {
 181         internal_urcu_lock();
 182
 183         /* All threads should read qparity before accessing data structure
 184          * where new ptr points to. Must be done within internal_urcu_lock
 185          * because it iterates on reader threads.*/
 186         /* Write new ptr before changing the qparity */
 187         force_mb_all_threads();
 188
 189         switch_next_urcu_qparity();     /* 0 -> 1 */
 190
 191         /*
 192          * Must commit qparity update to memory before waiting for parity
 193          * 0 quiescent state. Failure to do so could result in the writer
 194          * waiting forever while new readers are always accessing data (no
 195          * progress).
 196          * Ensured by STORE_SHARED and LOAD_SHARED.
 197          */
 198
 199         /*
 200          * Wait for previous parity to be empty of readers.
 201          */
 202         wait_for_quiescent_state();     /* Wait readers in parity 0 */
 203
 204         /*
 205          * Must finish waiting for quiescent state for parity 0 before
 206          * committing qparity update to memory. Failure to do so could result in
 207          * the writer waiting forever while new readers are always accessing
 208          * data (no progress).
 209          * Ensured by STORE_SHARED and LOAD_SHARED.
 210          */
 211
 212         switch_next_urcu_qparity();     /* 1 -> 0 */
 213
 214         /*
 215          * Must commit qparity update to memory before waiting for parity
 216          * 1 quiescent state. Failure to do so could result in the writer
 217          * waiting forever while new readers are always accessing data (no
 218          * progress).
 219          * Ensured by STORE_SHARED and LOAD_SHARED.
 220          */
 221
 222         /*
 223          * Wait for previous parity to be empty of readers.
 224          */
 225         wait_for_quiescent_state();     /* Wait readers in parity 1 */
 226
 227         /* Finish waiting for reader threads before letting the old ptr being
 228          * freed. Must be done within internal_urcu_lock because it iterates on
 229          * reader threads. */
 230         force_mb_all_threads();
 231
 232         internal_urcu_unlock();
 233 }
 234
 235 void urcu_add_reader(pthread_t id)
 236 {
 237         struct reader_registry *oldarray;
 238
 239         if (!registry) {
 240                 alloc_readers = INIT_NUM_THREADS;
 241                 num_readers = 0;
 242                 registry =
 243                         malloc(sizeof(struct reader_registry) * alloc_readers);
 244         }
 245         if (alloc_readers < num_readers + 1) {
 246                 oldarray = registry;
 247                 registry = malloc(sizeof(struct reader_registry)
 248                                 * (alloc_readers << 1));
 249                 memcpy(registry, oldarray,
 250                         sizeof(struct reader_registry) * alloc_readers);
 251                 alloc_readers <<= 1;
 252                 free(oldarray);
 253         }
 254         registry[num_readers].tid = id;
 255         /* reference to the TLS of _this_ reader thread. */
 256         registry[num_readers].urcu_active_readers = &urcu_active_readers;
 257         num_readers++;
 258 }
 259
 260 /*
 261  * Never shrink (implementation limitation).
 262  * This is O(nb threads). Eventually use a hash table.
 263  */
 264 void urcu_remove_reader(pthread_t id)
 265 {
 266         struct reader_registry *index;
 267
 268         assert(registry != NULL);
 269         for (index = registry; index < registry + num_readers; index++) {
 270                 if (pthread_equal(index->tid, id)) {
 271                         memcpy(index, &registry[num_readers - 1],
 272                                 sizeof(struct reader_registry));
 273                         registry[num_readers - 1].tid = 0;
 274                         registry[num_readers - 1].urcu_active_readers = NULL;
 275                         num_readers--;
 276                         return;
 277                 }
 278         }
 279         /* Hrm not found, forgot to register ? */
 280         assert(0);
 281 }
 282
 283 void urcu_register_thread(void)
 284 {
 285         internal_urcu_lock();
 286         urcu_add_reader(pthread_self());
 287         internal_urcu_unlock();
 288 }
 289
 290 void urcu_unregister_thread(void)
 291 {
 292         internal_urcu_lock();
 293         urcu_remove_reader(pthread_self());
 294         internal_urcu_unlock();
 295 }
 296
 297 #ifndef DEBUG_FULL_MB
 298 void sigurcu_handler(int signo, siginfo_t *siginfo, void *context)
 299 {
 300         /*
 301          * Executing this smp_mb() is the only purpose of this signal handler.
 302          * It punctually promotes barrier() into smp_mb() on every thread it is
 303          * executed on.
 304          */
 305         smp_mb();
 306         atomic_inc(&sig_done);
 307 }
 308
 309 void __attribute__((constructor)) urcu_init(void)
 310 {
 311         struct sigaction act;
 312         int ret;
 313
 314         act.sa_sigaction = sigurcu_handler;
 315         ret = sigaction(SIGURCU, &act, NULL);
 316         if (ret) {
 317                 perror("Error in sigaction");
 318                 exit(-1);
 319         }
 320 }
 321
 322 void __attribute__((destructor)) urcu_exit(void)
 323 {
 324         struct sigaction act;
 325         int ret;
 326
 327         ret = sigaction(SIGURCU, NULL, &act);
 328         if (ret) {
 329                 perror("Error in sigaction");
 330                 exit(-1);
 331         }
 332         assert(act.sa_sigaction == sigurcu_handler);
 333         free(registry);
 334 }
 335 #endif