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