tests/benchmark/test_urcu_lfs_rcu.c

   1 /*
   2  * test_urcu_lfs_rcu.c
   3  *
   4  * Userspace RCU library - example RCU-based lock-free stack
   5  *
   6  * Copyright February 2010 - Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
   7  * Copyright February 2010 - Paolo Bonzini <pbonzini@redhat.com>
   8  *
   9  * This program is free software; you can redistribute it and/or modify
  10  * it under the terms of the GNU General Public License as published by
  11  * the Free Software Foundation; either version 2 of the License, or
  12  * (at your option) any later version.
  13  *
  14  * This program 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
  17  * GNU General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU General Public License along
  20  * with this program; if not, write to the Free Software Foundation, Inc.,
  21  * 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
  22  */
  23
  24 #include <stdio.h>
  25 #include <pthread.h>
  26 #include <stdlib.h>
  27 #include <stdint.h>
  28 #include <stdbool.h>
  29 #include <string.h>
  30 #include <sys/types.h>
  31 #include <sys/wait.h>
  32 #include <unistd.h>
  33 #include <stdio.h>
  34 #include <assert.h>
  35 #include <errno.h>
  36
  37 #include <urcu/arch.h>
  38 #include <urcu/tls-compat.h>
  39 #include "cpuset.h"
  40 #include "thread-id.h"
  41
  42 /* hardcoded number of CPUs */
  43 #define NR_CPUS 16384
  44
  45 #ifndef DYNAMIC_LINK_TEST
  46 #define _LGPL_SOURCE
  47 #endif
  48 #include <urcu.h>
  49
  50 /* Remove deprecation warnings from test build. */
  51 #define CDS_LFS_RCU_DEPRECATED
  52
  53 #include <urcu/cds.h>
  54
  55 static volatile int test_go, test_stop;
  56
  57 static unsigned long rduration;
  58
  59 static unsigned long duration;
  60
  61 /* read-side C.S. duration, in loops */
  62 static unsigned long wdelay;
  63
  64 static inline void loop_sleep(unsigned long loops)
  65 {
  66         while (loops-- != 0)
  67                 caa_cpu_relax();
  68 }
  69
  70 static int verbose_mode;
  71
  72 #define printf_verbose(fmt, args...)            \
  73         do {                                    \
  74                 if (verbose_mode)               \
  75                         printf(fmt, args);      \
  76         } while (0)
  77
  78 static unsigned int cpu_affinities[NR_CPUS];
  79 static unsigned int next_aff = 0;
  80 static int use_affinity = 0;
  81
  82 pthread_mutex_t affinity_mutex = PTHREAD_MUTEX_INITIALIZER;
  83
  84 static void set_affinity(void)
  85 {
  86 #ifdef HAVE_SCHED_SETAFFINITY
  87         cpu_set_t mask;
  88         int cpu, ret;
  89 #endif /* HAVE_SCHED_SETAFFINITY */
  90
  91         if (!use_affinity)
  92                 return;
  93
  94 #ifdef HAVE_SCHED_SETAFFINITY
  95         ret = pthread_mutex_lock(&affinity_mutex);
  96         if (ret) {
  97                 perror("Error in pthread mutex lock");
  98                 exit(-1);
  99         }
 100         cpu = cpu_affinities[next_aff++];
 101         ret = pthread_mutex_unlock(&affinity_mutex);
 102         if (ret) {
 103                 perror("Error in pthread mutex unlock");
 104                 exit(-1);
 105         }
 106
 107         CPU_ZERO(&mask);
 108         CPU_SET(cpu, &mask);
 109 #if SCHED_SETAFFINITY_ARGS == 2
 110         sched_setaffinity(0, &mask);
 111 #else
 112         sched_setaffinity(0, sizeof(mask), &mask);
 113 #endif
 114 #endif /* HAVE_SCHED_SETAFFINITY */
 115 }
 116
 117 /*
 118  * returns 0 if test should end.
 119  */
 120 static int test_duration_dequeue(void)
 121 {
 122         return !test_stop;
 123 }
 124
 125 static int test_duration_enqueue(void)
 126 {
 127         return !test_stop;
 128 }
 129
 130 static DEFINE_URCU_TLS(unsigned long long, nr_dequeues);
 131 static DEFINE_URCU_TLS(unsigned long long, nr_enqueues);
 132
 133 static DEFINE_URCU_TLS(unsigned long long, nr_successful_dequeues);
 134 static DEFINE_URCU_TLS(unsigned long long, nr_successful_enqueues);
 135
 136 static unsigned int nr_enqueuers;
 137 static unsigned int nr_dequeuers;
 138
 139 struct test {
 140         struct cds_lfs_node_rcu list;
 141         struct rcu_head rcu;
 142 };
 143
 144 static struct cds_lfs_stack_rcu s;
 145
 146 static
 147 void *thr_enqueuer(void *_count)
 148 {
 149         unsigned long long *count = _count;
 150
 151         printf_verbose("thread_begin %s, tid %lu\n",
 152                         "enqueuer", urcu_get_thread_id());
 153
 154         set_affinity();
 155
 156         rcu_register_thread();
 157
 158         while (!test_go)
 159         {
 160         }
 161         cmm_smp_mb();
 162
 163         for (;;) {
 164                 struct test *node = malloc(sizeof(*node));
 165                 if (!node)
 166                         goto fail;
 167                 cds_lfs_node_init_rcu(&node->list);
 168                 /* No rcu read-side is needed for push */
 169                 cds_lfs_push_rcu(&s, &node->list);
 170                 URCU_TLS(nr_successful_enqueues)++;
 171
 172                 if (caa_unlikely(wdelay))
 173                         loop_sleep(wdelay);
 174 fail:
 175                 URCU_TLS(nr_enqueues)++;
 176                 if (caa_unlikely(!test_duration_enqueue()))
 177                         break;
 178         }
 179
 180         rcu_unregister_thread();
 181
 182         count[0] = URCU_TLS(nr_enqueues);
 183         count[1] = URCU_TLS(nr_successful_enqueues);
 184         printf_verbose("enqueuer thread_end, tid %lu, "
 185                         "enqueues %llu successful_enqueues %llu\n",
 186                         urcu_get_thread_id(),
 187                         URCU_TLS(nr_enqueues),
 188                         URCU_TLS(nr_successful_enqueues));
 189         return ((void*)1);
 190
 191 }
 192
 193 static
 194 void free_node_cb(struct rcu_head *head)
 195 {
 196         struct test *node =
 197                 caa_container_of(head, struct test, rcu);
 198         free(node);
 199 }
 200
 201 static
 202 void *thr_dequeuer(void *_count)
 203 {
 204         unsigned long long *count = _count;
 205
 206         printf_verbose("thread_begin %s, tid %lu\n",
 207                         "dequeuer", urcu_get_thread_id());
 208
 209         set_affinity();
 210
 211         rcu_register_thread();
 212
 213         while (!test_go)
 214         {
 215         }
 216         cmm_smp_mb();
 217
 218         for (;;) {
 219                 struct cds_lfs_node_rcu *snode;
 220
 221                 rcu_read_lock();
 222                 snode = cds_lfs_pop_rcu(&s);
 223                 rcu_read_unlock();
 224                 if (snode) {
 225                         struct test *node;
 226
 227                         node = caa_container_of(snode, struct test, list);
 228                         call_rcu(&node->rcu, free_node_cb);
 229                         URCU_TLS(nr_successful_dequeues)++;
 230                 }
 231                 URCU_TLS(nr_dequeues)++;
 232                 if (caa_unlikely(!test_duration_dequeue()))
 233                         break;
 234                 if (caa_unlikely(rduration))
 235                         loop_sleep(rduration);
 236         }
 237
 238         rcu_unregister_thread();
 239
 240         printf_verbose("dequeuer thread_end, tid %lu, "
 241                         "dequeues %llu, successful_dequeues %llu\n",
 242                         urcu_get_thread_id(),
 243                         URCU_TLS(nr_dequeues),
 244                         URCU_TLS(nr_successful_dequeues));
 245         count[0] = URCU_TLS(nr_dequeues);
 246         count[1] = URCU_TLS(nr_successful_dequeues);
 247         return ((void*)2);
 248 }
 249
 250 static
 251 void test_end(unsigned long long *nr_dequeues_l)
 252 {
 253         struct cds_lfs_node_rcu *snode;
 254
 255         do {
 256                 snode = cds_lfs_pop_rcu(&s);
 257                 if (snode) {
 258                         struct test *node;
 259
 260                         node = caa_container_of(snode, struct test, list);
 261                         free(node);
 262                         (*nr_dequeues_l)++;
 263                 }
 264         } while (snode);
 265 }
 266
 267 static
 268 void show_usage(char **argv)
 269 {
 270         printf("Usage : %s nr_dequeuers nr_enqueuers duration (s) <OPTIONS>\n",
 271                 argv[0]);
 272         printf("OPTIONS:\n");
 273         printf("        [-d delay] (enqueuer period (in loops))\n");
 274         printf("        [-c duration] (dequeuer period (in loops))\n");
 275         printf("        [-v] (verbose output)\n");
 276         printf("        [-a cpu#] [-a cpu#]... (affinity)\n");
 277         printf("\n");
 278 }
 279
 280 int main(int argc, char **argv)
 281 {
 282         int err;
 283         pthread_t *tid_enqueuer, *tid_dequeuer;
 284         void *tret;
 285         unsigned long long *count_enqueuer, *count_dequeuer;
 286         unsigned long long tot_enqueues = 0, tot_dequeues = 0;
 287         unsigned long long tot_successful_enqueues = 0,
 288                            tot_successful_dequeues = 0;
 289         unsigned long long end_dequeues = 0;
 290         int i, a;
 291         unsigned int i_thr;
 292
 293         if (argc < 4) {
 294                 show_usage(argv);
 295                 return -1;
 296         }
 297
 298         err = sscanf(argv[1], "%u", &nr_dequeuers);
 299         if (err != 1) {
 300                 show_usage(argv);
 301                 return -1;
 302         }
 303
 304         err = sscanf(argv[2], "%u", &nr_enqueuers);
 305         if (err != 1) {
 306                 show_usage(argv);
 307                 return -1;
 308         }
 309
 310         err = sscanf(argv[3], "%lu", &duration);
 311         if (err != 1) {
 312                 show_usage(argv);
 313                 return -1;
 314         }
 315
 316         for (i = 4; i < argc; i++) {
 317                 if (argv[i][0] != '-')
 318                         continue;
 319                 switch (argv[i][1]) {
 320                 case 'a':
 321                         if (argc < i + 2) {
 322                                 show_usage(argv);
 323                                 return -1;
 324                         }
 325                         a = atoi(argv[++i]);
 326                         cpu_affinities[next_aff++] = a;
 327                         use_affinity = 1;
 328                         printf_verbose("Adding CPU %d affinity\n", a);
 329                         break;
 330                 case 'c':
 331                         if (argc < i + 2) {
 332                                 show_usage(argv);
 333                                 return -1;
 334                         }
 335                         rduration = atol(argv[++i]);
 336                         break;
 337                 case 'd':
 338                         if (argc < i + 2) {
 339                                 show_usage(argv);
 340                                 return -1;
 341                         }
 342                         wdelay = atol(argv[++i]);
 343                         break;
 344                 case 'v':
 345                         verbose_mode = 1;
 346                         break;
 347                 }
 348         }
 349
 350         printf_verbose("running test for %lu seconds, %u enqueuers, "
 351                        "%u dequeuers.\n",
 352                        duration, nr_enqueuers, nr_dequeuers);
 353         printf_verbose("Writer delay : %lu loops.\n", rduration);
 354         printf_verbose("Reader duration : %lu loops.\n", wdelay);
 355         printf_verbose("thread %-6s, tid %lu\n",
 356                         "main", urcu_get_thread_id());
 357
 358         tid_enqueuer = calloc(nr_enqueuers, sizeof(*tid_enqueuer));
 359         tid_dequeuer = calloc(nr_dequeuers, sizeof(*tid_dequeuer));
 360         count_enqueuer = calloc(nr_enqueuers, 2 * sizeof(*count_enqueuer));
 361         count_dequeuer = calloc(nr_dequeuers, 2 * sizeof(*count_dequeuer));
 362         cds_lfs_init_rcu(&s);
 363         err = create_all_cpu_call_rcu_data(0);
 364         if (err) {
 365                 printf("Per-CPU call_rcu() worker threads unavailable. Using default global worker thread.\n");
 366         }
 367
 368         next_aff = 0;
 369
 370         for (i_thr = 0; i_thr < nr_enqueuers; i_thr++) {
 371                 err = pthread_create(&tid_enqueuer[i_thr], NULL, thr_enqueuer,
 372                                      &count_enqueuer[2 * i_thr]);
 373                 if (err != 0)
 374                         exit(1);
 375         }
 376         for (i_thr = 0; i_thr < nr_dequeuers; i_thr++) {
 377                 err = pthread_create(&tid_dequeuer[i_thr], NULL, thr_dequeuer,
 378                                      &count_dequeuer[2 * i_thr]);
 379                 if (err != 0)
 380                         exit(1);
 381         }
 382
 383         cmm_smp_mb();
 384
 385         test_go = 1;
 386
 387         for (i_thr = 0; i_thr < duration; i_thr++) {
 388                 sleep(1);
 389                 if (verbose_mode) {
 390                         fwrite(".", sizeof(char), 1, stdout);
 391                         fflush(stdout);
 392                 }
 393         }
 394
 395         test_stop = 1;
 396
 397         for (i_thr = 0; i_thr < nr_enqueuers; i_thr++) {
 398                 err = pthread_join(tid_enqueuer[i_thr], &tret);
 399                 if (err != 0)
 400                         exit(1);
 401                 tot_enqueues += count_enqueuer[2 * i_thr];
 402                 tot_successful_enqueues += count_enqueuer[2 * i_thr + 1];
 403         }
 404         for (i_thr = 0; i_thr < nr_dequeuers; i_thr++) {
 405                 err = pthread_join(tid_dequeuer[i_thr], &tret);
 406                 if (err != 0)
 407                         exit(1);
 408                 tot_dequeues += count_dequeuer[2 * i_thr];
 409                 tot_successful_dequeues += count_dequeuer[2 * i_thr + 1];
 410         }
 411
 412         test_end(&end_dequeues);
 413
 414         printf_verbose("total number of enqueues : %llu, dequeues %llu\n",
 415                        tot_enqueues, tot_dequeues);
 416         printf_verbose("total number of successful enqueues : %llu, "
 417                        "successful dequeues %llu\n",
 418                        tot_successful_enqueues, tot_successful_dequeues);
 419         printf("SUMMARY %-25s testdur %4lu nr_enqueuers %3u wdelay %6lu "
 420                 "nr_dequeuers %3u "
 421                 "rdur %6lu nr_enqueues %12llu nr_dequeues %12llu "
 422                 "successful enqueues %12llu successful dequeues %12llu "
 423                 "end_dequeues %llu nr_ops %12llu\n",
 424                 argv[0], duration, nr_enqueuers, wdelay,
 425                 nr_dequeuers, rduration, tot_enqueues, tot_dequeues,
 426                 tot_successful_enqueues,
 427                 tot_successful_dequeues, end_dequeues,
 428                 tot_enqueues + tot_dequeues);
 429         if (tot_successful_enqueues != tot_successful_dequeues + end_dequeues)
 430                 printf("WARNING! Discrepancy between nr succ. enqueues %llu vs "
 431                        "succ. dequeues + end dequeues %llu.\n",
 432                        tot_successful_enqueues,
 433                        tot_successful_dequeues + end_dequeues);
 434
 435         free_all_cpu_call_rcu_data();
 436         free(count_enqueuer);
 437         free(count_dequeuer);
 438         free(tid_enqueuer);
 439         free(tid_dequeuer);
 440         return 0;
 441 }