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 #include <errno.h>
  18 #include <poll.h>
  19
  20 #include "urcu.h"
  21
  22 pthread_mutex_t urcu_mutex = PTHREAD_MUTEX_INITIALIZER;
  23
  24 /*
  25  * Global grace period counter.
  26  * Contains the current RCU_GP_CTR_BIT.
  27  * Also has a RCU_GP_CTR_BIT of 1, to accelerate the reader fast path.
  28  * Written to only by writer with mutex taken. Read by both writer and readers.
  29  */
  30 long urcu_gp_ctr = RCU_GP_COUNT;
  31
  32 /*
  33  * Written to only by each individual reader. Read by both the reader and the
  34  * writers.
  35  */
  36 long __thread urcu_active_readers;
  37
  38 /* Thread IDs of registered readers */
  39 #define INIT_NUM_THREADS 4
  40
  41 struct reader_registry {
  42         pthread_t tid;
  43         long *urcu_active_readers;
  44         char *need_mb;
  45 };
  46
  47 #ifdef DEBUG_YIELD
  48 unsigned int yield_active;
  49 unsigned int __thread rand_yield;
  50 #endif
  51
  52 static struct reader_registry *registry;
  53 static char __thread need_mb;
  54 static int num_readers, alloc_readers;
  55
  56 void internal_urcu_lock(void)
  57 {
  58         int ret;
  59
  60 #ifndef DISTRUST_SIGNALS_EXTREME
  61         ret = pthread_mutex_lock(&urcu_mutex);
  62         if (ret) {
  63                 perror("Error in pthread mutex lock");
  64                 exit(-1);
  65         }
  66 #else /* #ifndef DISTRUST_SIGNALS_EXTREME */
  67         while ((ret = pthread_mutex_trylock(&urcu_mutex)) != 0) {
  68                 if (ret != EBUSY && ret != EINTR) {
  69                         printf("ret = %d, errno = %d\n", ret, errno);
  70                         perror("Error in pthread mutex lock");
  71                         exit(-1);
  72                 }
  73                 if (need_mb) {
  74                         smp_mb();
  75                         need_mb = 0;
  76                         smp_mb();
  77                 }
  78                 poll(NULL,0,10);
  79         }
  80 #endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */
  81 }
  82
  83 void internal_urcu_unlock(void)
  84 {
  85         int ret;
  86
  87         ret = pthread_mutex_unlock(&urcu_mutex);
  88         if (ret) {
  89                 perror("Error in pthread mutex unlock");
  90                 exit(-1);
  91         }
  92 }
  93
  94 /*
  95  * called with urcu_mutex held.
  96  */
  97 static void switch_next_urcu_qparity(void)
  98 {
  99         STORE_SHARED(urcu_gp_ctr, urcu_gp_ctr ^ RCU_GP_CTR_BIT);
 100 }
 101
 102 #ifdef DEBUG_FULL_MB
 103 #ifdef HAS_INCOHERENT_CACHES
 104 static void force_mb_single_thread(struct reader_registry *index)
 105 {
 106         smp_mb();
 107 }
 108 #endif /* #ifdef HAS_INCOHERENT_CACHES */
 109
 110 static void force_mb_all_threads(void)
 111 {
 112         smp_mb();
 113 }
 114 #else /* #ifdef DEBUG_FULL_MB */
 115 #ifdef HAS_INCOHERENT_CACHES
 116 static void force_mb_single_thread(struct reader_registry *index)
 117 {
 118         assert(registry);
 119         /*
 120          * pthread_kill has a smp_mb(). But beware, we assume it performs
 121          * a cache flush on architectures with non-coherent cache. Let's play
 122          * safe and don't assume anything : we use smp_mc() to make sure the
 123          * cache flush is enforced.
 124          */
 125         *index->need_mb = 1;
 126         smp_mc();       /* write ->need_mb before sending the signals */
 127         pthread_kill(index->tid, SIGURCU);
 128         smp_mb();
 129         /*
 130          * Wait for sighandler (and thus mb()) to execute on every thread.
 131          * BUSY-LOOP.
 132          */
 133         while (*index->need_mb) {
 134                 poll(NULL, 0, 1);
 135         }
 136         smp_mb();       /* read ->need_mb before ending the barrier */
 137 }
 138 #endif /* #ifdef HAS_INCOHERENT_CACHES */
 139
 140 static void force_mb_all_threads(void)
 141 {
 142         struct reader_registry *index;
 143         /*
 144          * Ask for each threads to execute a smp_mb() so we can consider the
 145          * compiler barriers around rcu read lock as real memory barriers.
 146          */
 147         if (!registry)
 148                 return;
 149         /*
 150          * pthread_kill has a smp_mb(). But beware, we assume it performs
 151          * a cache flush on architectures with non-coherent cache. Let's play
 152          * safe and don't assume anything : we use smp_mc() to make sure the
 153          * cache flush is enforced.
 154          */
 155         for (index = registry; index < registry + num_readers; index++) {
 156                 *index->need_mb = 1;
 157                 smp_mc();       /* write need_mb before sending the signal */
 158                 pthread_kill(index->tid, SIGURCU);
 159         }
 160         /*
 161          * Wait for sighandler (and thus mb()) to execute on every thread.
 162          *
 163          * Note that the pthread_kill() will never be executed on systems
 164          * that correctly deliver signals in a timely manner.  However, it
 165          * is not uncommon for kernels to have bugs that can result in
 166          * lost or unduly delayed signals.
 167          *
 168          * If you are seeing the below pthread_kill() executing much at
 169          * all, we suggest testing the underlying kernel and filing the
 170          * relevant bug report.  For Linux kernels, we recommend getting
 171          * the Linux Test Project (LTP).
 172          */
 173         for (index = registry; index < registry + num_readers; index++) {
 174                 while (*index->need_mb) {
 175                         pthread_kill(index->tid, SIGURCU);
 176                         poll(NULL, 0, 1);
 177                 }
 178         }
 179         smp_mb();       /* read ->need_mb before ending the barrier */
 180 }
 181 #endif /* #else #ifdef DEBUG_FULL_MB */
 182
 183 void wait_for_quiescent_state(void)
 184 {
 185         struct reader_registry *index;
 186
 187         if (!registry)
 188                 return;
 189         /*
 190          * Wait for each thread urcu_active_readers count to become 0.
 191          */
 192         for (index = registry; index < registry + num_readers; index++) {
 193 #ifndef HAS_INCOHERENT_CACHES
 194                 while (rcu_old_gp_ongoing(index->urcu_active_readers))
 195                         cpu_relax();
 196 #else /* #ifndef HAS_INCOHERENT_CACHES */
 197                 int wait_loops = 0;
 198                 /*
 199                  * BUSY-LOOP. Force the reader thread to commit its
 200                  * urcu_active_readers update to memory if we wait for too long.
 201                  */
 202                 while (rcu_old_gp_ongoing(index->urcu_active_readers)) {
 203                         if (wait_loops++ == KICK_READER_LOOPS) {
 204                                 force_mb_single_thread(index);
 205                                 wait_loops = 0;
 206                         } else {
 207                                 cpu_relax();
 208                         }
 209                 }
 210 #endif /* #else #ifndef HAS_INCOHERENT_CACHES */
 211         }
 212 }
 213
 214 void synchronize_rcu(void)
 215 {
 216         internal_urcu_lock();
 217
 218         /* All threads should read qparity before accessing data structure
 219          * where new ptr points to. Must be done within internal_urcu_lock
 220          * because it iterates on reader threads.*/
 221         /* Write new ptr before changing the qparity */
 222         force_mb_all_threads();
 223
 224         switch_next_urcu_qparity();     /* 0 -> 1 */
 225
 226         /*
 227          * Must commit qparity update to memory before waiting for parity
 228          * 0 quiescent state. Failure to do so could result in the writer
 229          * waiting forever while new readers are always accessing data (no
 230          * progress).
 231          * Ensured by STORE_SHARED and LOAD_SHARED.
 232          */
 233
 234         /*
 235          * Wait for previous parity to be empty of readers.
 236          */
 237         wait_for_quiescent_state();     /* Wait readers in parity 0 */
 238
 239         /*
 240          * Must finish waiting for quiescent state for parity 0 before
 241          * committing qparity update to memory. Failure to do so could result in
 242          * the writer waiting forever while new readers are always accessing
 243          * data (no progress).
 244          * Ensured by STORE_SHARED and LOAD_SHARED.
 245          */
 246
 247         switch_next_urcu_qparity();     /* 1 -> 0 */
 248
 249         /*
 250          * Must commit qparity update to memory before waiting for parity
 251          * 1 quiescent state. Failure to do so could result in the writer
 252          * waiting forever while new readers are always accessing data (no
 253          * progress).
 254          * Ensured by STORE_SHARED and LOAD_SHARED.
 255          */
 256
 257         /*
 258          * Wait for previous parity to be empty of readers.
 259          */
 260         wait_for_quiescent_state();     /* Wait readers in parity 1 */
 261
 262         /* Finish waiting for reader threads before letting the old ptr being
 263          * freed. Must be done within internal_urcu_lock because it iterates on
 264          * reader threads. */
 265         force_mb_all_threads();
 266
 267         internal_urcu_unlock();
 268 }
 269
 270 void urcu_add_reader(pthread_t id)
 271 {
 272         struct reader_registry *oldarray;
 273
 274         if (!registry) {
 275                 alloc_readers = INIT_NUM_THREADS;
 276                 num_readers = 0;
 277                 registry =
 278                         malloc(sizeof(struct reader_registry) * alloc_readers);
 279         }
 280         if (alloc_readers < num_readers + 1) {
 281                 oldarray = registry;
 282                 registry = malloc(sizeof(struct reader_registry)
 283                                 * (alloc_readers << 1));
 284                 memcpy(registry, oldarray,
 285                         sizeof(struct reader_registry) * alloc_readers);
 286                 alloc_readers <<= 1;
 287                 free(oldarray);
 288         }
 289         registry[num_readers].tid = id;
 290         /* reference to the TLS of _this_ reader thread. */
 291         registry[num_readers].urcu_active_readers = &urcu_active_readers;
 292         registry[num_readers].need_mb = &need_mb;
 293         num_readers++;
 294 }
 295
 296 /*
 297  * Never shrink (implementation limitation).
 298  * This is O(nb threads). Eventually use a hash table.
 299  */
 300 void urcu_remove_reader(pthread_t id)
 301 {
 302         struct reader_registry *index;
 303
 304         assert(registry != NULL);
 305         for (index = registry; index < registry + num_readers; index++) {
 306                 if (pthread_equal(index->tid, id)) {
 307                         memcpy(index, &registry[num_readers - 1],
 308                                 sizeof(struct reader_registry));
 309                         registry[num_readers - 1].tid = 0;
 310                         registry[num_readers - 1].urcu_active_readers = NULL;
 311                         num_readers--;
 312                         return;
 313                 }
 314         }
 315         /* Hrm not found, forgot to register ? */
 316         assert(0);
 317 }
 318
 319 void urcu_register_thread(void)
 320 {
 321         internal_urcu_lock();
 322         urcu_add_reader(pthread_self());
 323         internal_urcu_unlock();
 324 }
 325
 326 void urcu_unregister_thread(void)
 327 {
 328         internal_urcu_lock();
 329         urcu_remove_reader(pthread_self());
 330         internal_urcu_unlock();
 331 }
 332
 333 #ifndef DEBUG_FULL_MB
 334 void sigurcu_handler(int signo, siginfo_t *siginfo, void *context)
 335 {
 336         /*
 337          * Executing this smp_mb() is the only purpose of this signal handler.
 338          * It punctually promotes barrier() into smp_mb() on every thread it is
 339          * executed on.
 340          */
 341         smp_mb();
 342         need_mb = 0;
 343         smp_mb();
 344 }
 345
 346 void __attribute__((constructor)) urcu_init(void)
 347 {
 348         struct sigaction act;
 349         int ret;
 350
 351         act.sa_sigaction = sigurcu_handler;
 352         ret = sigaction(SIGURCU, &act, NULL);
 353         if (ret) {
 354                 perror("Error in sigaction");
 355                 exit(-1);
 356         }
 357 }
 358
 359 void __attribute__((destructor)) urcu_exit(void)
 360 {
 361         struct sigaction act;
 362         int ret;
 363
 364         ret = sigaction(SIGURCU, NULL, &act);
 365         if (ret) {
 366                 perror("Error in sigaction");
 367                 exit(-1);
 368         }
 369         assert(act.sa_sigaction == sigurcu_handler);
 370         free(registry);
 371 }
 372 #endif /* #ifndef DEBUG_FULL_MB */