tests/benchmark/test_urcu_lfs.c

   1 /*
   2  * test_urcu_lfs.c
   3  *
   4  * Userspace RCU library - example lock-free stack
   5  *
   6  * Copyright 2010-2012 - 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 #define _GNU_SOURCE
  25 #include "config.h"
  26 #include <stdio.h>
  27 #include <pthread.h>
  28 #include <stdlib.h>
  29 #include <stdint.h>
  30 #include <stdbool.h>
  31 #include <string.h>
  32 #include <sys/types.h>
  33 #include <sys/wait.h>
  34 #include <unistd.h>
  35 #include <stdio.h>
  36 #include <assert.h>
  37 #include <errno.h>
  38
  39 #include <urcu/arch.h>
  40 #include <urcu/tls-compat.h>
  41 #include "cpuset.h"
  42 #include "thread-id.h"
  43
  44 /* hardcoded number of CPUs */
  45 #define NR_CPUS 16384
  46
  47 #ifndef DYNAMIC_LINK_TEST
  48 #define _LGPL_SOURCE
  49 #endif
  50 #include <urcu.h>
  51 #include <urcu/cds.h>
  52
  53 #define POISON_PTR      ((void *) 0x42UL)
  54
  55 /*
  56  * External synchronization used.
  57  */
  58 enum test_sync {
  59         TEST_SYNC_NONE = 0,
  60         TEST_SYNC_RCU,
  61 };
  62
  63 static enum test_sync test_sync;
  64
  65 static volatile int test_go, test_stop;
  66
  67 static unsigned long rduration;
  68
  69 static unsigned long duration;
  70
  71 /* read-side C.S. duration, in loops */
  72 static unsigned long wdelay;
  73
  74 static inline void loop_sleep(unsigned long loops)
  75 {
  76         while (loops-- != 0)
  77                 caa_cpu_relax();
  78 }
  79
  80 static int verbose_mode;
  81
  82 static int test_pop, test_pop_all;
  83
  84 #define printf_verbose(fmt, args...)            \
  85         do {                                    \
  86                 if (verbose_mode)               \
  87                         printf(fmt, ## args);   \
  88         } while (0)
  89
  90 static unsigned int cpu_affinities[NR_CPUS];
  91 static unsigned int next_aff = 0;
  92 static int use_affinity = 0;
  93
  94 pthread_mutex_t affinity_mutex = PTHREAD_MUTEX_INITIALIZER;
  95
  96 static void set_affinity(void)
  97 {
  98 #if HAVE_SCHED_SETAFFINITY
  99         cpu_set_t mask;
 100         int cpu, ret;
 101 #endif /* HAVE_SCHED_SETAFFINITY */
 102
 103         if (!use_affinity)
 104                 return;
 105
 106 #if HAVE_SCHED_SETAFFINITY
 107         ret = pthread_mutex_lock(&affinity_mutex);
 108         if (ret) {
 109                 perror("Error in pthread mutex lock");
 110                 exit(-1);
 111         }
 112         cpu = cpu_affinities[next_aff++];
 113         ret = pthread_mutex_unlock(&affinity_mutex);
 114         if (ret) {
 115                 perror("Error in pthread mutex unlock");
 116                 exit(-1);
 117         }
 118
 119         CPU_ZERO(&mask);
 120         CPU_SET(cpu, &mask);
 121 #if SCHED_SETAFFINITY_ARGS == 2
 122         sched_setaffinity(0, &mask);
 123 #else
 124         sched_setaffinity(0, sizeof(mask), &mask);
 125 #endif
 126 #endif /* HAVE_SCHED_SETAFFINITY */
 127 }
 128
 129 /*
 130  * returns 0 if test should end.
 131  */
 132 static int test_duration_dequeue(void)
 133 {
 134         return !test_stop;
 135 }
 136
 137 static int test_duration_enqueue(void)
 138 {
 139         return !test_stop;
 140 }
 141
 142 static DEFINE_URCU_TLS(unsigned long long, nr_dequeues);
 143 static DEFINE_URCU_TLS(unsigned long long, nr_enqueues);
 144
 145 static DEFINE_URCU_TLS(unsigned long long, nr_successful_dequeues);
 146 static DEFINE_URCU_TLS(unsigned long long, nr_successful_enqueues);
 147
 148 static unsigned int nr_enqueuers;
 149 static unsigned int nr_dequeuers;
 150
 151 struct test {
 152         struct cds_lfs_node list;
 153         struct rcu_head rcu;
 154 };
 155
 156 static struct cds_lfs_stack s;
 157
 158 static void *thr_enqueuer(void *_count)
 159 {
 160         unsigned long long *count = _count;
 161
 162         printf_verbose("thread_begin %s, tid %lu\n",
 163                         "enqueuer", urcu_get_thread_id());
 164
 165         set_affinity();
 166
 167         rcu_register_thread();
 168
 169         while (!test_go)
 170         {
 171         }
 172         cmm_smp_mb();
 173
 174         for (;;) {
 175                 struct test *node = malloc(sizeof(*node));
 176                 if (!node)
 177                         goto fail;
 178                 cds_lfs_node_init(&node->list);
 179                 cds_lfs_push(&s, &node->list);
 180                 URCU_TLS(nr_successful_enqueues)++;
 181
 182                 if (caa_unlikely(wdelay))
 183                         loop_sleep(wdelay);
 184 fail:
 185                 URCU_TLS(nr_enqueues)++;
 186                 if (caa_unlikely(!test_duration_enqueue()))
 187                         break;
 188         }
 189
 190         rcu_unregister_thread();
 191
 192         count[0] = URCU_TLS(nr_enqueues);
 193         count[1] = URCU_TLS(nr_successful_enqueues);
 194         printf_verbose("enqueuer thread_end, tid %lu, "
 195                         "enqueues %llu successful_enqueues %llu\n",
 196                         urcu_get_thread_id(),
 197                         URCU_TLS(nr_enqueues),
 198                         URCU_TLS(nr_successful_enqueues));
 199         return ((void*)1);
 200
 201 }
 202
 203 static
 204 void free_node_cb(struct rcu_head *head)
 205 {
 206         struct test *node =
 207                 caa_container_of(head, struct test, rcu);
 208         free(node);
 209 }
 210
 211 static
 212 void do_test_pop(enum test_sync sync)
 213 {
 214         struct cds_lfs_node *snode;
 215
 216         if (sync == TEST_SYNC_RCU)
 217                 rcu_read_lock();
 218         snode = __cds_lfs_pop(&s);
 219         if (sync == TEST_SYNC_RCU)
 220                 rcu_read_unlock();
 221         if (snode) {
 222                 struct test *node;
 223
 224                 snode->next = POISON_PTR;
 225                 node = caa_container_of(snode,
 226                         struct test, list);
 227                 if (sync == TEST_SYNC_RCU)
 228                         call_rcu(&node->rcu, free_node_cb);
 229                 else
 230                         free(node);
 231                 URCU_TLS(nr_successful_dequeues)++;
 232         }
 233         URCU_TLS(nr_dequeues)++;
 234 }
 235
 236 static
 237 void do_test_pop_all(enum test_sync sync)
 238 {
 239         struct cds_lfs_node *snode;
 240         struct cds_lfs_head *head;
 241         struct cds_lfs_node *n;
 242
 243         head = __cds_lfs_pop_all(&s);
 244         cds_lfs_for_each_safe(head, snode, n) {
 245                 struct test *node;
 246
 247                 snode->next = POISON_PTR;
 248                 node = caa_container_of(snode, struct test, list);
 249                 if (sync == TEST_SYNC_RCU)
 250                         call_rcu(&node->rcu, free_node_cb);
 251                 else
 252                         free(node);
 253                 URCU_TLS(nr_successful_dequeues)++;
 254                 URCU_TLS(nr_dequeues)++;
 255         }
 256
 257 }
 258
 259 static void *thr_dequeuer(void *_count)
 260 {
 261         unsigned long long *count = _count;
 262         unsigned int counter = 0;
 263
 264         printf_verbose("thread_begin %s, tid %lu\n",
 265                         "dequeuer", urcu_get_thread_id());
 266
 267         set_affinity();
 268
 269         rcu_register_thread();
 270
 271         while (!test_go)
 272         {
 273         }
 274         cmm_smp_mb();
 275
 276         assert(test_pop || test_pop_all);
 277
 278         for (;;) {
 279                 if (test_pop && test_pop_all) {
 280                         /* both pop and pop all */
 281                         if (counter & 1)
 282                                 do_test_pop(test_sync);
 283                         else
 284                                 do_test_pop_all(test_sync);
 285                         counter++;
 286                 } else {
 287                         if (test_pop)
 288                                 do_test_pop(test_sync);
 289                         else
 290                                 do_test_pop_all(test_sync);
 291                 }
 292
 293                 if (caa_unlikely(!test_duration_dequeue()))
 294                         break;
 295                 if (caa_unlikely(rduration))
 296                         loop_sleep(rduration);
 297         }
 298
 299         rcu_unregister_thread();
 300
 301         printf_verbose("dequeuer thread_end, tid %lu, "
 302                         "dequeues %llu, successful_dequeues %llu\n",
 303                         urcu_get_thread_id(),
 304                         URCU_TLS(nr_dequeues),
 305                         URCU_TLS(nr_successful_dequeues));
 306         count[0] = URCU_TLS(nr_dequeues);
 307         count[1] = URCU_TLS(nr_successful_dequeues);
 308         return ((void*)2);
 309 }
 310
 311 static void test_end(struct cds_lfs_stack *s, unsigned long long *nr_dequeues)
 312 {
 313         struct cds_lfs_node *snode;
 314
 315         do {
 316                 snode = __cds_lfs_pop(s);
 317                 if (snode) {
 318                         struct test *node;
 319
 320                         node = caa_container_of(snode, struct test, list);
 321                         free(node);
 322                         (*nr_dequeues)++;
 323                 }
 324         } while (snode);
 325 }
 326
 327 static void show_usage(int argc, char **argv)
 328 {
 329         printf("Usage : %s nr_dequeuers nr_enqueuers duration (s) <OPTIONS>\n",
 330                 argv[0]);
 331         printf("OPTIONS:\n");
 332         printf("        [-d delay] (enqueuer period (in loops))\n");
 333         printf("        [-c duration] (dequeuer period (in loops))\n");
 334         printf("        [-v] (verbose output)\n");
 335         printf("        [-a cpu#] [-a cpu#]... (affinity)\n");
 336         printf("        [-p] (test pop)\n");
 337         printf("        [-P] (test pop_all, enabled by default)\n");
 338         printf("        [-R] (use RCU external synchronization)\n");
 339         printf("                Note: default: no external synchronization used.\n");
 340         printf("\n");
 341 }
 342
 343 int main(int argc, char **argv)
 344 {
 345         int err;
 346         pthread_t *tid_enqueuer, *tid_dequeuer;
 347         void *tret;
 348         unsigned long long *count_enqueuer, *count_dequeuer;
 349         unsigned long long tot_enqueues = 0, tot_dequeues = 0;
 350         unsigned long long tot_successful_enqueues = 0,
 351                            tot_successful_dequeues = 0;
 352         unsigned long long end_dequeues = 0;
 353         int i, a;
 354
 355         if (argc < 4) {
 356                 show_usage(argc, argv);
 357                 return -1;
 358         }
 359
 360         err = sscanf(argv[1], "%u", &nr_dequeuers);
 361         if (err != 1) {
 362                 show_usage(argc, argv);
 363                 return -1;
 364         }
 365
 366         err = sscanf(argv[2], "%u", &nr_enqueuers);
 367         if (err != 1) {
 368                 show_usage(argc, argv);
 369                 return -1;
 370         }
 371
 372         err = sscanf(argv[3], "%lu", &duration);
 373         if (err != 1) {
 374                 show_usage(argc, argv);
 375                 return -1;
 376         }
 377
 378         for (i = 4; i < argc; i++) {
 379                 if (argv[i][0] != '-')
 380                         continue;
 381                 switch (argv[i][1]) {
 382                 case 'a':
 383                         if (argc < i + 2) {
 384                                 show_usage(argc, argv);
 385                                 return -1;
 386                         }
 387                         a = atoi(argv[++i]);
 388                         cpu_affinities[next_aff++] = a;
 389                         use_affinity = 1;
 390                         printf_verbose("Adding CPU %d affinity\n", a);
 391                         break;
 392                 case 'c':
 393                         if (argc < i + 2) {
 394                                 show_usage(argc, argv);
 395                                 return -1;
 396                         }
 397                         rduration = atol(argv[++i]);
 398                         break;
 399                 case 'd':
 400                         if (argc < i + 2) {
 401                                 show_usage(argc, argv);
 402                                 return -1;
 403                         }
 404                         wdelay = atol(argv[++i]);
 405                         break;
 406                 case 'v':
 407                         verbose_mode = 1;
 408                         break;
 409                 case 'p':
 410                         test_pop = 1;
 411                         break;
 412                 case 'P':
 413                         test_pop_all = 1;
 414                         break;
 415                 case 'R':
 416                         test_sync = TEST_SYNC_RCU;
 417                         break;
 418                 }
 419         }
 420
 421         /* activate pop_all test by default */
 422         if (!test_pop && !test_pop_all)
 423                 test_pop_all = 1;
 424
 425         printf_verbose("running test for %lu seconds, %u enqueuers, "
 426                        "%u dequeuers.\n",
 427                        duration, nr_enqueuers, nr_dequeuers);
 428         if (test_pop)
 429                 printf_verbose("pop test activated.\n");
 430         if (test_pop_all)
 431                 printf_verbose("pop_all test activated.\n");
 432         if (test_sync == TEST_SYNC_RCU)
 433                 printf_verbose("External sync: RCU.\n");
 434         else
 435                 printf_verbose("External sync: none.\n");
 436         printf_verbose("Writer delay : %lu loops.\n", rduration);
 437         printf_verbose("Reader duration : %lu loops.\n", wdelay);
 438         printf_verbose("thread %-6s, tid %lu\n",
 439                         "main", urcu_get_thread_id());
 440
 441         tid_enqueuer = calloc(nr_enqueuers, sizeof(*tid_enqueuer));
 442         tid_dequeuer = calloc(nr_dequeuers, sizeof(*tid_dequeuer));
 443         count_enqueuer = calloc(nr_enqueuers, 2 * sizeof(*count_enqueuer));
 444         count_dequeuer = calloc(nr_dequeuers, 2 * sizeof(*count_dequeuer));
 445         cds_lfs_init(&s);
 446         err = create_all_cpu_call_rcu_data(0);
 447         if (err) {
 448                 printf("Per-CPU call_rcu() worker threads unavailable. Using default global worker thread.\n");
 449         }
 450
 451         next_aff = 0;
 452
 453         for (i = 0; i < nr_enqueuers; i++) {
 454                 err = pthread_create(&tid_enqueuer[i], NULL, thr_enqueuer,
 455                                      &count_enqueuer[2 * i]);
 456                 if (err != 0)
 457                         exit(1);
 458         }
 459         for (i = 0; i < nr_dequeuers; i++) {
 460                 err = pthread_create(&tid_dequeuer[i], NULL, thr_dequeuer,
 461                                      &count_dequeuer[2 * i]);
 462                 if (err != 0)
 463                         exit(1);
 464         }
 465
 466         cmm_smp_mb();
 467
 468         test_go = 1;
 469
 470         for (i = 0; i < duration; i++) {
 471                 sleep(1);
 472                 if (verbose_mode) {
 473                         fwrite(".", sizeof(char), 1, stdout);
 474                         fflush(stdout);
 475                 }
 476         }
 477
 478         test_stop = 1;
 479
 480         for (i = 0; i < nr_enqueuers; i++) {
 481                 err = pthread_join(tid_enqueuer[i], &tret);
 482                 if (err != 0)
 483                         exit(1);
 484                 tot_enqueues += count_enqueuer[2 * i];
 485                 tot_successful_enqueues += count_enqueuer[2 * i + 1];
 486         }
 487         for (i = 0; i < nr_dequeuers; i++) {
 488                 err = pthread_join(tid_dequeuer[i], &tret);
 489                 if (err != 0)
 490                         exit(1);
 491                 tot_dequeues += count_dequeuer[2 * i];
 492                 tot_successful_dequeues += count_dequeuer[2 * i + 1];
 493         }
 494
 495         test_end(&s, &end_dequeues);
 496
 497         printf_verbose("total number of enqueues : %llu, dequeues %llu\n",
 498                        tot_enqueues, tot_dequeues);
 499         printf_verbose("total number of successful enqueues : %llu, "
 500                        "successful dequeues %llu\n",
 501                        tot_successful_enqueues, tot_successful_dequeues);
 502         printf("SUMMARY %-25s testdur %4lu nr_enqueuers %3u wdelay %6lu "
 503                 "nr_dequeuers %3u "
 504                 "rdur %6lu nr_enqueues %12llu nr_dequeues %12llu "
 505                 "successful enqueues %12llu successful dequeues %12llu "
 506                 "end_dequeues %llu nr_ops %12llu\n",
 507                 argv[0], duration, nr_enqueuers, wdelay,
 508                 nr_dequeuers, rduration, tot_enqueues, tot_dequeues,
 509                 tot_successful_enqueues,
 510                 tot_successful_dequeues, end_dequeues,
 511                 tot_enqueues + tot_dequeues);
 512         if (tot_successful_enqueues != tot_successful_dequeues + end_dequeues)
 513                 printf("WARNING! Discrepancy between nr succ. enqueues %llu vs "
 514                        "succ. dequeues + end dequeues %llu.\n",
 515                        tot_successful_enqueues,
 516                        tot_successful_dequeues + end_dequeues);
 517
 518         free_all_cpu_call_rcu_data();
 519         cds_lfs_destroy(&s);
 520         free(count_enqueuer);
 521         free(count_dequeuer);
 522         free(tid_enqueuer);
 523         free(tid_dequeuer);
 524         return 0;
 525 }