tests/benchmark/test_urcu_wfs.c

   1 /*
   2  * test_urcu_wfs.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 <urcu/uatomic.h>
  40 #include "cpuset.h"
  41 #include "thread-id.h"
  42
  43 /* hardcoded number of CPUs */
  44 #define NR_CPUS 16384
  45
  46 #ifndef DYNAMIC_LINK_TEST
  47 #define _LGPL_SOURCE
  48 #endif
  49 #include <urcu/wfstack.h>
  50
  51 /*
  52  * External synchronization used.
  53  */
  54 enum test_sync {
  55         TEST_SYNC_NONE = 0,
  56         TEST_SYNC_MUTEX,
  57 };
  58
  59 static enum test_sync test_sync;
  60
  61 static int test_force_sync;
  62
  63 static volatile int test_go, test_stop_enqueue, test_stop_dequeue;
  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, test_wait_empty;
  81 static int test_enqueue_stopped;
  82
  83 #define printf_verbose(fmt, args...)            \
  84         do {                                    \
  85                 if (verbose_mode)               \
  86                         printf(fmt, ## args);   \
  87         } while (0)
  88
  89 static unsigned int cpu_affinities[NR_CPUS];
  90 static unsigned int next_aff = 0;
  91 static int use_affinity = 0;
  92
  93 pthread_mutex_t affinity_mutex = PTHREAD_MUTEX_INITIALIZER;
  94
  95 static void set_affinity(void)
  96 {
  97 #if HAVE_SCHED_SETAFFINITY
  98         cpu_set_t mask;
  99         int cpu, ret;
 100 #endif /* HAVE_SCHED_SETAFFINITY */
 101
 102         if (!use_affinity)
 103                 return;
 104
 105 #if HAVE_SCHED_SETAFFINITY
 106         ret = pthread_mutex_lock(&affinity_mutex);
 107         if (ret) {
 108                 perror("Error in pthread mutex lock");
 109                 exit(-1);
 110         }
 111         cpu = cpu_affinities[next_aff++];
 112         ret = pthread_mutex_unlock(&affinity_mutex);
 113         if (ret) {
 114                 perror("Error in pthread mutex unlock");
 115                 exit(-1);
 116         }
 117
 118         CPU_ZERO(&mask);
 119         CPU_SET(cpu, &mask);
 120 #if SCHED_SETAFFINITY_ARGS == 2
 121         sched_setaffinity(0, &mask);
 122 #else
 123         sched_setaffinity(0, sizeof(mask), &mask);
 124 #endif
 125 #endif /* HAVE_SCHED_SETAFFINITY */
 126 }
 127
 128 /*
 129  * returns 0 if test should end.
 130  */
 131 static int test_duration_dequeue(void)
 132 {
 133         return !test_stop_dequeue;
 134 }
 135
 136 static int test_duration_enqueue(void)
 137 {
 138         return !test_stop_enqueue;
 139 }
 140
 141 static DEFINE_URCU_TLS(unsigned long long, nr_dequeues);
 142 static DEFINE_URCU_TLS(unsigned long long, nr_enqueues);
 143
 144 static DEFINE_URCU_TLS(unsigned long long, nr_successful_dequeues);
 145 static DEFINE_URCU_TLS(unsigned long long, nr_successful_enqueues);
 146 static DEFINE_URCU_TLS(unsigned long long, nr_empty_dest_enqueues);
 147 static DEFINE_URCU_TLS(unsigned long long, nr_pop_all);
 148 static DEFINE_URCU_TLS(unsigned long long, nr_pop_last);
 149
 150 static unsigned int nr_enqueuers;
 151 static unsigned int nr_dequeuers;
 152
 153 static struct cds_wfs_stack s;
 154
 155 static void *thr_enqueuer(void *_count)
 156 {
 157         unsigned long long *count = _count;
 158         bool was_nonempty;
 159
 160         printf_verbose("thread_begin %s, tid %lu\n",
 161                         "enqueuer", urcu_get_thread_id());
 162
 163         set_affinity();
 164
 165         while (!test_go)
 166         {
 167         }
 168         cmm_smp_mb();
 169
 170         for (;;) {
 171                 struct cds_wfs_node *node = malloc(sizeof(*node));
 172                 if (!node)
 173                         goto fail;
 174                 cds_wfs_node_init(node);
 175                 was_nonempty = cds_wfs_push(&s, node);
 176                 URCU_TLS(nr_successful_enqueues)++;
 177                 if (!was_nonempty)
 178                         URCU_TLS(nr_empty_dest_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         uatomic_inc(&test_enqueue_stopped);
 189         count[0] = URCU_TLS(nr_enqueues);
 190         count[1] = URCU_TLS(nr_successful_enqueues);
 191         count[2] = URCU_TLS(nr_empty_dest_enqueues);
 192         printf_verbose("enqueuer thread_end, tid %lu, "
 193                         "enqueues %llu successful_enqueues %llu, "
 194                         "empty_dest_enqueues %llu\n",
 195                         urcu_get_thread_id(),
 196                         URCU_TLS(nr_enqueues),
 197                         URCU_TLS(nr_successful_enqueues),
 198                         URCU_TLS(nr_empty_dest_enqueues));
 199         return ((void*)1);
 200
 201 }
 202
 203 static void do_test_pop(enum test_sync sync)
 204 {
 205         struct cds_wfs_node *node;
 206         int state;
 207
 208         if (sync == TEST_SYNC_MUTEX)
 209                 cds_wfs_pop_lock(&s);
 210         node = __cds_wfs_pop_with_state_blocking(&s, &state);
 211         if (sync == TEST_SYNC_MUTEX)
 212                 cds_wfs_pop_unlock(&s);
 213
 214         if (node) {
 215                 if (state & CDS_WFS_STATE_LAST)
 216                         URCU_TLS(nr_pop_last)++;
 217                 free(node);
 218                 URCU_TLS(nr_successful_dequeues)++;
 219         }
 220         URCU_TLS(nr_dequeues)++;
 221 }
 222
 223 static void do_test_pop_all(enum test_sync sync)
 224 {
 225         struct cds_wfs_head *head;
 226         struct cds_wfs_node *node, *n;
 227
 228         if (sync == TEST_SYNC_MUTEX)
 229                 cds_wfs_pop_lock(&s);
 230         head = __cds_wfs_pop_all(&s);
 231         if (sync == TEST_SYNC_MUTEX)
 232                 cds_wfs_pop_unlock(&s);
 233
 234         /* Check if empty */
 235         if (cds_wfs_first(head) == NULL)
 236                 return;
 237
 238         URCU_TLS(nr_pop_all)++;
 239         URCU_TLS(nr_pop_last)++;
 240
 241         cds_wfs_for_each_blocking_safe(head, node, n) {
 242                 free(node);
 243                 URCU_TLS(nr_successful_dequeues)++;
 244                 URCU_TLS(nr_dequeues)++;
 245         }
 246 }
 247
 248 static void *thr_dequeuer(void *_count)
 249 {
 250         unsigned long long *count = _count;
 251         unsigned int counter = 0;
 252
 253         printf_verbose("thread_begin %s, tid %lu\n",
 254                         "dequeuer", urcu_get_thread_id());
 255
 256         set_affinity();
 257
 258         while (!test_go)
 259         {
 260         }
 261         cmm_smp_mb();
 262
 263         assert(test_pop || test_pop_all);
 264
 265         for (;;) {
 266                 if (test_pop && test_pop_all) {
 267                         if (counter & 1)
 268                                 do_test_pop(test_sync);
 269                         else
 270                                 do_test_pop_all(test_sync);
 271                         counter++;
 272                 } else {
 273                         if (test_pop)
 274                                 do_test_pop(test_sync);
 275                         else
 276                                 do_test_pop_all(test_sync);
 277                 }
 278
 279                 if (caa_unlikely(!test_duration_dequeue()))
 280                         break;
 281                 if (caa_unlikely(rduration))
 282                         loop_sleep(rduration);
 283         }
 284
 285         printf_verbose("dequeuer thread_end, tid %lu, "
 286                         "dequeues %llu, successful_dequeues %llu "
 287                         "pop_all %llu pop_last %llu\n",
 288                         urcu_get_thread_id(),
 289                         URCU_TLS(nr_dequeues), URCU_TLS(nr_successful_dequeues),
 290                         URCU_TLS(nr_pop_all),
 291                         URCU_TLS(nr_pop_last));
 292         count[0] = URCU_TLS(nr_dequeues);
 293         count[1] = URCU_TLS(nr_successful_dequeues);
 294         count[2] = URCU_TLS(nr_pop_all);
 295         count[3] = URCU_TLS(nr_pop_last);
 296         return ((void*)2);
 297 }
 298
 299 static void test_end(struct cds_wfs_stack *s, unsigned long long *nr_dequeues,
 300                 unsigned long long *nr_pop_last)
 301 {
 302         struct cds_wfs_node *node;
 303         int state;
 304
 305         do {
 306                 node = cds_wfs_pop_with_state_blocking(s, &state);
 307                 if (node) {
 308                         if (state & CDS_WFS_STATE_LAST)
 309                                 (*nr_pop_last)++;
 310                         free(node);
 311                         (*nr_dequeues)++;
 312                 }
 313         } while (node);
 314 }
 315
 316 static void show_usage(int argc, char **argv)
 317 {
 318         printf("Usage : %s nr_dequeuers nr_enqueuers duration (s) <OPTIONS>\n",
 319                 argv[0]);
 320         printf("OPTIONS:\n");
 321         printf("        [-d delay] (enqueuer period (in loops))\n");
 322         printf("        [-c duration] (dequeuer period (in loops))\n");
 323         printf("        [-v] (verbose output)\n");
 324         printf("        [-a cpu#] [-a cpu#]... (affinity)\n");
 325         printf("        [-p] (test pop)\n");
 326         printf("        [-P] (test pop_all, enabled by default)\n");
 327         printf("        [-M] (use mutex external synchronization)\n");
 328         printf("                Note: default: no external synchronization used.\n");
 329         printf("        [-f] (force user-provided synchronization)\n");
 330         printf("        [-w] Wait for dequeuer to empty stack\n");
 331         printf("\n");
 332 }
 333
 334 int main(int argc, char **argv)
 335 {
 336         int err;
 337         pthread_t *tid_enqueuer, *tid_dequeuer;
 338         void *tret;
 339         unsigned long long *count_enqueuer, *count_dequeuer;
 340         unsigned long long tot_enqueues = 0, tot_dequeues = 0;
 341         unsigned long long tot_successful_enqueues = 0,
 342                            tot_successful_dequeues = 0,
 343                            tot_empty_dest_enqueues = 0,
 344                            tot_pop_all = 0, tot_pop_last = 0;
 345         unsigned long long end_dequeues = 0;
 346         int i, a, retval = 0;
 347
 348         if (argc < 4) {
 349                 show_usage(argc, argv);
 350                 return -1;
 351         }
 352
 353         err = sscanf(argv[1], "%u", &nr_dequeuers);
 354         if (err != 1) {
 355                 show_usage(argc, argv);
 356                 return -1;
 357         }
 358
 359         err = sscanf(argv[2], "%u", &nr_enqueuers);
 360         if (err != 1) {
 361                 show_usage(argc, argv);
 362                 return -1;
 363         }
 364
 365         err = sscanf(argv[3], "%lu", &duration);
 366         if (err != 1) {
 367                 show_usage(argc, argv);
 368                 return -1;
 369         }
 370
 371         for (i = 4; i < argc; i++) {
 372                 if (argv[i][0] != '-')
 373                         continue;
 374                 switch (argv[i][1]) {
 375                 case 'a':
 376                         if (argc < i + 2) {
 377                                 show_usage(argc, argv);
 378                                 return -1;
 379                         }
 380                         a = atoi(argv[++i]);
 381                         cpu_affinities[next_aff++] = a;
 382                         use_affinity = 1;
 383                         printf_verbose("Adding CPU %d affinity\n", a);
 384                         break;
 385                 case 'c':
 386                         if (argc < i + 2) {
 387                                 show_usage(argc, argv);
 388                                 return -1;
 389                         }
 390                         rduration = atol(argv[++i]);
 391                         break;
 392                 case 'd':
 393                         if (argc < i + 2) {
 394                                 show_usage(argc, argv);
 395                                 return -1;
 396                         }
 397                         wdelay = atol(argv[++i]);
 398                         break;
 399                 case 'v':
 400                         verbose_mode = 1;
 401                         break;
 402                 case 'p':
 403                         test_pop = 1;
 404                         break;
 405                 case 'P':
 406                         test_pop_all = 1;
 407                         break;
 408                 case 'M':
 409                         test_sync = TEST_SYNC_MUTEX;
 410                         break;
 411                 case 'w':
 412                         test_wait_empty = 1;
 413                         break;
 414                 case 'f':
 415                         test_force_sync = 1;
 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         if (test_sync == TEST_SYNC_NONE && nr_dequeuers > 1 && test_pop) {
 425                 if (test_force_sync) {
 426                         fprintf(stderr, "[WARNING] Using pop concurrently "
 427                                 "with other pop or pop_all without external "
 428                                 "synchronization. Expect run-time failure.\n");
 429                 } else {
 430                         printf("Enforcing mutex synchronization\n");
 431                         test_sync = TEST_SYNC_MUTEX;
 432                 }
 433         }
 434
 435         printf_verbose("running test for %lu seconds, %u enqueuers, "
 436                        "%u dequeuers.\n",
 437                        duration, nr_enqueuers, nr_dequeuers);
 438         if (test_pop)
 439                 printf_verbose("pop test activated.\n");
 440         if (test_pop_all)
 441                 printf_verbose("pop_all test activated.\n");
 442         if (test_sync == TEST_SYNC_MUTEX)
 443                 printf_verbose("External sync: mutex.\n");
 444         else
 445                 printf_verbose("External sync: none.\n");
 446         if (test_wait_empty)
 447                 printf_verbose("Wait for dequeuers to empty stack.\n");
 448         printf_verbose("Writer delay : %lu loops.\n", rduration);
 449         printf_verbose("Reader duration : %lu loops.\n", wdelay);
 450         printf_verbose("thread %-6s, tid %lu\n",
 451                         "main", urcu_get_thread_id());
 452
 453         tid_enqueuer = calloc(nr_enqueuers, sizeof(*tid_enqueuer));
 454         tid_dequeuer = calloc(nr_dequeuers, sizeof(*tid_dequeuer));
 455         count_enqueuer = calloc(nr_enqueuers, 3 * sizeof(*count_enqueuer));
 456         count_dequeuer = calloc(nr_dequeuers, 4 * sizeof(*count_dequeuer));
 457         cds_wfs_init(&s);
 458
 459         next_aff = 0;
 460
 461         for (i = 0; i < nr_enqueuers; i++) {
 462                 err = pthread_create(&tid_enqueuer[i], NULL, thr_enqueuer,
 463                                      &count_enqueuer[3 * i]);
 464                 if (err != 0)
 465                         exit(1);
 466         }
 467         for (i = 0; i < nr_dequeuers; i++) {
 468                 err = pthread_create(&tid_dequeuer[i], NULL, thr_dequeuer,
 469                                      &count_dequeuer[4 * i]);
 470                 if (err != 0)
 471                         exit(1);
 472         }
 473
 474         cmm_smp_mb();
 475
 476         test_go = 1;
 477
 478         for (i = 0; i < duration; i++) {
 479                 sleep(1);
 480                 if (verbose_mode) {
 481                         fwrite(".", sizeof(char), 1, stdout);
 482                         fflush(stdout);
 483                 }
 484         }
 485
 486         test_stop_enqueue = 1;
 487
 488         if (test_wait_empty) {
 489                 while (nr_enqueuers != uatomic_read(&test_enqueue_stopped)) {
 490                         sleep(1);
 491                 }
 492                 while (!cds_wfs_empty(&s)) {
 493                         sleep(1);
 494                 }
 495         }
 496
 497         test_stop_dequeue = 1;
 498
 499         for (i = 0; i < nr_enqueuers; i++) {
 500                 err = pthread_join(tid_enqueuer[i], &tret);
 501                 if (err != 0)
 502                         exit(1);
 503                 tot_enqueues += count_enqueuer[3 * i];
 504                 tot_successful_enqueues += count_enqueuer[3 * i + 1];
 505                 tot_empty_dest_enqueues += count_enqueuer[3 * i + 2];
 506         }
 507         for (i = 0; i < nr_dequeuers; i++) {
 508                 err = pthread_join(tid_dequeuer[i], &tret);
 509                 if (err != 0)
 510                         exit(1);
 511                 tot_dequeues += count_dequeuer[4 * i];
 512                 tot_successful_dequeues += count_dequeuer[4 * i + 1];
 513                 tot_pop_all += count_dequeuer[4 * i + 2];
 514                 tot_pop_last += count_dequeuer[4 * i + 3];
 515         }
 516
 517         test_end(&s, &end_dequeues, &tot_pop_last);
 518
 519         printf_verbose("total number of enqueues : %llu, dequeues %llu\n",
 520                        tot_enqueues, tot_dequeues);
 521         printf_verbose("total number of successful enqueues : %llu, "
 522                        "enqueues to empty dest : %llu, "
 523                        "successful dequeues %llu, "
 524                        "pop_all : %llu, pop_last : %llu\n",
 525                        tot_successful_enqueues,
 526                        tot_empty_dest_enqueues,
 527                        tot_successful_dequeues,
 528                        tot_pop_all, tot_pop_last);
 529         printf("SUMMARY %-25s testdur %4lu nr_enqueuers %3u wdelay %6lu "
 530                 "nr_dequeuers %3u "
 531                 "rdur %6lu nr_enqueues %12llu nr_dequeues %12llu "
 532                 "successful enqueues %12llu enqueues to empty dest %12llu "
 533                 "successful dequeues %12llu pop_all %12llu "
 534                 "pop_last %llu end_dequeues %llu nr_ops %12llu\n",
 535                 argv[0], duration, nr_enqueuers, wdelay,
 536                 nr_dequeuers, rduration, tot_enqueues, tot_dequeues,
 537                 tot_successful_enqueues,
 538                 tot_empty_dest_enqueues,
 539                 tot_successful_dequeues, tot_pop_all, tot_pop_last,
 540                 end_dequeues,
 541                 tot_enqueues + tot_dequeues);
 542         if (tot_successful_enqueues != tot_successful_dequeues + end_dequeues) {
 543                 printf("WARNING! Discrepancy between nr succ. enqueues %llu vs "
 544                        "succ. dequeues + end dequeues %llu.\n",
 545                        tot_successful_enqueues,
 546                        tot_successful_dequeues + end_dequeues);
 547                 retval = 1;
 548         }
 549         /*
 550          * The enqueuer should see exactly as many empty queues than the
 551          * number of non-empty stacks dequeued.
 552          */
 553         if (tot_empty_dest_enqueues != tot_pop_last) {
 554                 printf("WARNING! Discrepancy between empty enqueue (%llu) and "
 555                         "number of pop last (%llu)\n",
 556                         tot_empty_dest_enqueues,
 557                         tot_pop_last);
 558                 retval = 1;
 559         }
 560         cds_wfs_destroy(&s);
 561         free(count_enqueuer);
 562         free(count_dequeuer);
 563         free(tid_enqueuer);
 564         free(tid_dequeuer);
 565         return retval;
 566 }