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 <errno.h>
  35
  36 #include <urcu/arch.h>
  37 #include <urcu/assert.h>
  38 #include <urcu/tls-compat.h>
  39 #include <urcu/uatomic.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/wfstack.h>
  49
  50 /*
  51  * External synchronization used.
  52  */
  53 enum test_sync {
  54         TEST_SYNC_NONE = 0,
  55         TEST_SYNC_MUTEX,
  56 };
  57
  58 static enum test_sync test_sync;
  59
  60 static int test_force_sync;
  61
  62 static volatile int test_go, test_stop_enqueue, test_stop_dequeue;
  63
  64 static unsigned long rduration;
  65
  66 static unsigned long duration;
  67
  68 /* read-side C.S. duration, in loops */
  69 static unsigned long wdelay;
  70
  71 static inline void loop_sleep(unsigned long loops)
  72 {
  73         while (loops-- != 0)
  74                 caa_cpu_relax();
  75 }
  76
  77 static int verbose_mode;
  78
  79 static int test_pop, test_pop_all, test_wait_empty;
  80 static unsigned int test_enqueue_stopped;
  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 #ifdef 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 #ifdef 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         sched_setaffinity(0, sizeof(mask), &mask);
 120 #endif /* HAVE_SCHED_SETAFFINITY */
 121 }
 122
 123 /*
 124  * returns 0 if test should end.
 125  */
 126 static int test_duration_dequeue(void)
 127 {
 128         return !test_stop_dequeue;
 129 }
 130
 131 static int test_duration_enqueue(void)
 132 {
 133         return !test_stop_enqueue;
 134 }
 135
 136 static DEFINE_URCU_TLS(unsigned long long, nr_dequeues);
 137 static DEFINE_URCU_TLS(unsigned long long, nr_enqueues);
 138
 139 static DEFINE_URCU_TLS(unsigned long long, nr_successful_dequeues);
 140 static DEFINE_URCU_TLS(unsigned long long, nr_successful_enqueues);
 141 static DEFINE_URCU_TLS(unsigned long long, nr_empty_dest_enqueues);
 142 static DEFINE_URCU_TLS(unsigned long long, nr_pop_all);
 143 static DEFINE_URCU_TLS(unsigned long long, nr_pop_last);
 144
 145 static unsigned int nr_enqueuers;
 146 static unsigned int nr_dequeuers;
 147
 148 static struct cds_wfs_stack s;
 149
 150 static void *thr_enqueuer(void *_count)
 151 {
 152         unsigned long long *count = _count;
 153         bool was_nonempty;
 154
 155         printf_verbose("thread_begin %s, tid %lu\n",
 156                         "enqueuer", urcu_get_thread_id());
 157
 158         set_affinity();
 159
 160         while (!test_go)
 161         {
 162         }
 163         cmm_smp_mb();
 164
 165         for (;;) {
 166                 struct cds_wfs_node *node = malloc(sizeof(*node));
 167                 if (!node)
 168                         goto fail;
 169                 cds_wfs_node_init(node);
 170                 was_nonempty = cds_wfs_push(&s, node);
 171                 URCU_TLS(nr_successful_enqueues)++;
 172                 if (!was_nonempty)
 173                         URCU_TLS(nr_empty_dest_enqueues)++;
 174
 175                 if (caa_unlikely(wdelay))
 176                         loop_sleep(wdelay);
 177 fail:
 178                 URCU_TLS(nr_enqueues)++;
 179                 if (caa_unlikely(!test_duration_enqueue()))
 180                         break;
 181         }
 182
 183         uatomic_inc(&test_enqueue_stopped);
 184         count[0] = URCU_TLS(nr_enqueues);
 185         count[1] = URCU_TLS(nr_successful_enqueues);
 186         count[2] = URCU_TLS(nr_empty_dest_enqueues);
 187         printf_verbose("enqueuer thread_end, tid %lu, "
 188                         "enqueues %llu successful_enqueues %llu, "
 189                         "empty_dest_enqueues %llu\n",
 190                         urcu_get_thread_id(),
 191                         URCU_TLS(nr_enqueues),
 192                         URCU_TLS(nr_successful_enqueues),
 193                         URCU_TLS(nr_empty_dest_enqueues));
 194         return ((void*)1);
 195
 196 }
 197
 198 static void do_test_pop(enum test_sync sync)
 199 {
 200         struct cds_wfs_node *node;
 201         int state;
 202
 203         if (sync == TEST_SYNC_MUTEX)
 204                 cds_wfs_pop_lock(&s);
 205         node = __cds_wfs_pop_with_state_blocking(&s, &state);
 206         if (sync == TEST_SYNC_MUTEX)
 207                 cds_wfs_pop_unlock(&s);
 208
 209         if (node) {
 210                 if (state & CDS_WFS_STATE_LAST)
 211                         URCU_TLS(nr_pop_last)++;
 212                 free(node);
 213                 URCU_TLS(nr_successful_dequeues)++;
 214         }
 215         URCU_TLS(nr_dequeues)++;
 216 }
 217
 218 static void do_test_pop_all(enum test_sync sync)
 219 {
 220         struct cds_wfs_head *head;
 221         struct cds_wfs_node *node, *n;
 222
 223         if (sync == TEST_SYNC_MUTEX)
 224                 cds_wfs_pop_lock(&s);
 225         head = __cds_wfs_pop_all(&s);
 226         if (sync == TEST_SYNC_MUTEX)
 227                 cds_wfs_pop_unlock(&s);
 228
 229         /* Check if empty */
 230         if (cds_wfs_first(head) == NULL)
 231                 return;
 232
 233         URCU_TLS(nr_pop_all)++;
 234         URCU_TLS(nr_pop_last)++;
 235
 236         cds_wfs_for_each_blocking_safe(head, node, n) {
 237                 free(node);
 238                 URCU_TLS(nr_successful_dequeues)++;
 239                 URCU_TLS(nr_dequeues)++;
 240         }
 241 }
 242
 243 static void *thr_dequeuer(void *_count)
 244 {
 245         unsigned long long *count = _count;
 246         unsigned int counter = 0;
 247
 248         printf_verbose("thread_begin %s, tid %lu\n",
 249                         "dequeuer", urcu_get_thread_id());
 250
 251         set_affinity();
 252
 253         while (!test_go)
 254         {
 255         }
 256         cmm_smp_mb();
 257
 258         urcu_posix_assert(test_pop || test_pop_all);
 259
 260         for (;;) {
 261                 if (test_pop && test_pop_all) {
 262                         if (counter & 1)
 263                                 do_test_pop(test_sync);
 264                         else
 265                                 do_test_pop_all(test_sync);
 266                         counter++;
 267                 } else {
 268                         if (test_pop)
 269                                 do_test_pop(test_sync);
 270                         else
 271                                 do_test_pop_all(test_sync);
 272                 }
 273
 274                 if (caa_unlikely(!test_duration_dequeue()))
 275                         break;
 276                 if (caa_unlikely(rduration))
 277                         loop_sleep(rduration);
 278         }
 279
 280         printf_verbose("dequeuer thread_end, tid %lu, "
 281                         "dequeues %llu, successful_dequeues %llu "
 282                         "pop_all %llu pop_last %llu\n",
 283                         urcu_get_thread_id(),
 284                         URCU_TLS(nr_dequeues), URCU_TLS(nr_successful_dequeues),
 285                         URCU_TLS(nr_pop_all),
 286                         URCU_TLS(nr_pop_last));
 287         count[0] = URCU_TLS(nr_dequeues);
 288         count[1] = URCU_TLS(nr_successful_dequeues);
 289         count[2] = URCU_TLS(nr_pop_all);
 290         count[3] = URCU_TLS(nr_pop_last);
 291         return ((void*)2);
 292 }
 293
 294 static void test_end(unsigned long long *nr_dequeues_l,
 295                 unsigned long long *nr_pop_last_l)
 296 {
 297         struct cds_wfs_node *node;
 298         int state;
 299
 300         do {
 301                 node = cds_wfs_pop_with_state_blocking(&s, &state);
 302                 if (node) {
 303                         if (state & CDS_WFS_STATE_LAST)
 304                                 (*nr_pop_last_l)++;
 305                         free(node);
 306                         (*nr_dequeues_l)++;
 307                 }
 308         } while (node);
 309 }
 310
 311 static void show_usage(char **argv)
 312 {
 313         printf("Usage : %s nr_dequeuers nr_enqueuers duration (s) <OPTIONS>\n",
 314                 argv[0]);
 315         printf("OPTIONS:\n");
 316         printf("        [-d delay] (enqueuer period (in loops))\n");
 317         printf("        [-c duration] (dequeuer period (in loops))\n");
 318         printf("        [-v] (verbose output)\n");
 319         printf("        [-a cpu#] [-a cpu#]... (affinity)\n");
 320         printf("        [-p] (test pop)\n");
 321         printf("        [-P] (test pop_all, enabled by default)\n");
 322         printf("        [-M] (use mutex external synchronization)\n");
 323         printf("                Note: default: no external synchronization used.\n");
 324         printf("        [-f] (force user-provided synchronization)\n");
 325         printf("        [-w] Wait for dequeuer to empty stack\n");
 326         printf("\n");
 327 }
 328
 329 int main(int argc, char **argv)
 330 {
 331         int err;
 332         pthread_t *tid_enqueuer, *tid_dequeuer;
 333         void *tret;
 334         unsigned long long *count_enqueuer, *count_dequeuer;
 335         unsigned long long tot_enqueues = 0, tot_dequeues = 0;
 336         unsigned long long tot_successful_enqueues = 0,
 337                            tot_successful_dequeues = 0,
 338                            tot_empty_dest_enqueues = 0,
 339                            tot_pop_all = 0, tot_pop_last = 0;
 340         unsigned long long end_dequeues = 0;
 341         int i, a, retval = 0;
 342         unsigned int i_thr;
 343
 344         if (argc < 4) {
 345                 show_usage(argv);
 346                 return -1;
 347         }
 348
 349         err = sscanf(argv[1], "%u", &nr_dequeuers);
 350         if (err != 1) {
 351                 show_usage(argv);
 352                 return -1;
 353         }
 354
 355         err = sscanf(argv[2], "%u", &nr_enqueuers);
 356         if (err != 1) {
 357                 show_usage(argv);
 358                 return -1;
 359         }
 360
 361         err = sscanf(argv[3], "%lu", &duration);
 362         if (err != 1) {
 363                 show_usage(argv);
 364                 return -1;
 365         }
 366
 367         for (i = 4; i < argc; i++) {
 368                 if (argv[i][0] != '-')
 369                         continue;
 370                 switch (argv[i][1]) {
 371                 case 'a':
 372                         if (argc < i + 2) {
 373                                 show_usage(argv);
 374                                 return -1;
 375                         }
 376                         a = atoi(argv[++i]);
 377                         cpu_affinities[next_aff++] = a;
 378                         use_affinity = 1;
 379                         printf_verbose("Adding CPU %d affinity\n", a);
 380                         break;
 381                 case 'c':
 382                         if (argc < i + 2) {
 383                                 show_usage(argv);
 384                                 return -1;
 385                         }
 386                         rduration = atol(argv[++i]);
 387                         break;
 388                 case 'd':
 389                         if (argc < i + 2) {
 390                                 show_usage(argv);
 391                                 return -1;
 392                         }
 393                         wdelay = atol(argv[++i]);
 394                         break;
 395                 case 'v':
 396                         verbose_mode = 1;
 397                         break;
 398                 case 'p':
 399                         test_pop = 1;
 400                         break;
 401                 case 'P':
 402                         test_pop_all = 1;
 403                         break;
 404                 case 'M':
 405                         test_sync = TEST_SYNC_MUTEX;
 406                         break;
 407                 case 'w':
 408                         test_wait_empty = 1;
 409                         break;
 410                 case 'f':
 411                         test_force_sync = 1;
 412                         break;
 413                 }
 414         }
 415
 416         /* activate pop_all test by default */
 417         if (!test_pop && !test_pop_all)
 418                 test_pop_all = 1;
 419
 420         if (test_sync == TEST_SYNC_NONE && nr_dequeuers > 1 && test_pop) {
 421                 if (test_force_sync) {
 422                         fprintf(stderr, "[WARNING] Using pop concurrently "
 423                                 "with other pop or pop_all without external "
 424                                 "synchronization. Expect run-time failure.\n");
 425                 } else {
 426                         printf("Enforcing mutex synchronization\n");
 427                         test_sync = TEST_SYNC_MUTEX;
 428                 }
 429         }
 430
 431         printf_verbose("running test for %lu seconds, %u enqueuers, "
 432                        "%u dequeuers.\n",
 433                        duration, nr_enqueuers, nr_dequeuers);
 434         if (test_pop)
 435                 printf_verbose("pop test activated.\n");
 436         if (test_pop_all)
 437                 printf_verbose("pop_all test activated.\n");
 438         if (test_sync == TEST_SYNC_MUTEX)
 439                 printf_verbose("External sync: mutex.\n");
 440         else
 441                 printf_verbose("External sync: none.\n");
 442         if (test_wait_empty)
 443                 printf_verbose("Wait for dequeuers to empty stack.\n");
 444         printf_verbose("Writer delay : %lu loops.\n", rduration);
 445         printf_verbose("Reader duration : %lu loops.\n", wdelay);
 446         printf_verbose("thread %-6s, tid %lu\n",
 447                         "main", urcu_get_thread_id());
 448
 449         tid_enqueuer = calloc(nr_enqueuers, sizeof(*tid_enqueuer));
 450         tid_dequeuer = calloc(nr_dequeuers, sizeof(*tid_dequeuer));
 451         count_enqueuer = calloc(nr_enqueuers, 3 * sizeof(*count_enqueuer));
 452         count_dequeuer = calloc(nr_dequeuers, 4 * sizeof(*count_dequeuer));
 453         cds_wfs_init(&s);
 454
 455         next_aff = 0;
 456
 457         for (i_thr = 0; i_thr < nr_enqueuers; i_thr++) {
 458                 err = pthread_create(&tid_enqueuer[i], NULL, thr_enqueuer,
 459                                      &count_enqueuer[3 * i_thr]);
 460                 if (err != 0)
 461                         exit(1);
 462         }
 463         for (i_thr = 0; i_thr < nr_dequeuers; i_thr++) {
 464                 err = pthread_create(&tid_dequeuer[i_thr], NULL, thr_dequeuer,
 465                                      &count_dequeuer[4 * i_thr]);
 466                 if (err != 0)
 467                         exit(1);
 468         }
 469
 470         cmm_smp_mb();
 471
 472         test_go = 1;
 473
 474         for (i_thr = 0; i_thr < duration; i_thr++) {
 475                 sleep(1);
 476                 if (verbose_mode) {
 477                         fwrite(".", sizeof(char), 1, stdout);
 478                         fflush(stdout);
 479                 }
 480         }
 481
 482         test_stop_enqueue = 1;
 483
 484         if (test_wait_empty) {
 485                 while (nr_enqueuers != uatomic_read(&test_enqueue_stopped)) {
 486                         sleep(1);
 487                 }
 488                 while (!cds_wfs_empty(&s)) {
 489                         sleep(1);
 490                 }
 491         }
 492
 493         test_stop_dequeue = 1;
 494
 495         for (i_thr = 0; i_thr < nr_enqueuers; i_thr++) {
 496                 err = pthread_join(tid_enqueuer[i_thr], &tret);
 497                 if (err != 0)
 498                         exit(1);
 499                 tot_enqueues += count_enqueuer[3 * i_thr];
 500                 tot_successful_enqueues += count_enqueuer[3 * i_thr + 1];
 501                 tot_empty_dest_enqueues += count_enqueuer[3 * i_thr + 2];
 502         }
 503         for (i_thr = 0; i_thr < nr_dequeuers; i_thr++) {
 504                 err = pthread_join(tid_dequeuer[i_thr], &tret);
 505                 if (err != 0)
 506                         exit(1);
 507                 tot_dequeues += count_dequeuer[4 * i_thr];
 508                 tot_successful_dequeues += count_dequeuer[4 * i_thr + 1];
 509                 tot_pop_all += count_dequeuer[4 * i_thr + 2];
 510                 tot_pop_last += count_dequeuer[4 * i_thr + 3];
 511         }
 512
 513         test_end(&end_dequeues, &tot_pop_last);
 514
 515         printf_verbose("total number of enqueues : %llu, dequeues %llu\n",
 516                        tot_enqueues, tot_dequeues);
 517         printf_verbose("total number of successful enqueues : %llu, "
 518                        "enqueues to empty dest : %llu, "
 519                        "successful dequeues %llu, "
 520                        "pop_all : %llu, pop_last : %llu\n",
 521                        tot_successful_enqueues,
 522                        tot_empty_dest_enqueues,
 523                        tot_successful_dequeues,
 524                        tot_pop_all, tot_pop_last);
 525         printf("SUMMARY %-25s testdur %4lu nr_enqueuers %3u wdelay %6lu "
 526                 "nr_dequeuers %3u "
 527                 "rdur %6lu nr_enqueues %12llu nr_dequeues %12llu "
 528                 "successful enqueues %12llu enqueues to empty dest %12llu "
 529                 "successful dequeues %12llu pop_all %12llu "
 530                 "pop_last %llu end_dequeues %llu nr_ops %12llu\n",
 531                 argv[0], duration, nr_enqueuers, wdelay,
 532                 nr_dequeuers, rduration, tot_enqueues, tot_dequeues,
 533                 tot_successful_enqueues,
 534                 tot_empty_dest_enqueues,
 535                 tot_successful_dequeues, tot_pop_all, tot_pop_last,
 536                 end_dequeues,
 537                 tot_enqueues + tot_dequeues);
 538         if (tot_successful_enqueues != tot_successful_dequeues + end_dequeues) {
 539                 printf("WARNING! Discrepancy between nr succ. enqueues %llu vs "
 540                        "succ. dequeues + end dequeues %llu.\n",
 541                        tot_successful_enqueues,
 542                        tot_successful_dequeues + end_dequeues);
 543                 retval = 1;
 544         }
 545         /*
 546          * The enqueuer should see exactly as many empty queues than the
 547          * number of non-empty stacks dequeued.
 548          */
 549         if (tot_empty_dest_enqueues != tot_pop_last) {
 550                 printf("WARNING! Discrepancy between empty enqueue (%llu) and "
 551                         "number of pop last (%llu)\n",
 552                         tot_empty_dest_enqueues,
 553                         tot_pop_last);
 554                 retval = 1;
 555         }
 556
 557         cds_wfs_destroy(&s);
 558         free(count_enqueuer);
 559         free(count_dequeuer);
 560         free(tid_enqueuer);
 561         free(tid_dequeuer);
 562
 563         return retval;
 564 }