tests/benchmark/test_urcu_wfcq.c

   1 /*
   2  * test_urcu_wfcq.c
   3  *
   4  * Userspace RCU library - example RCU-based lock-free concurrent queue
   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/wfcqueue.h>
  50
  51 enum test_sync {
  52         TEST_SYNC_NONE = 0,
  53         TEST_SYNC_MUTEX,
  54 };
  55
  56 static enum test_sync test_sync;
  57
  58 static int test_force_sync;
  59
  60 static volatile int test_go, test_stop_enqueue, test_stop_dequeue;
  61
  62 static unsigned long rduration;
  63
  64 static unsigned long duration;
  65
  66 /* read-side C.S. duration, in loops */
  67 static unsigned long wdelay;
  68
  69 static inline void loop_sleep(unsigned long loops)
  70 {
  71         while (loops-- != 0)
  72                 caa_cpu_relax();
  73 }
  74
  75 static int verbose_mode;
  76
  77 static int test_dequeue, test_splice, test_wait_empty;
  78 static int test_enqueue_stopped;
  79
  80 #define printf_verbose(fmt, args...)            \
  81         do {                                    \
  82                 if (verbose_mode)               \
  83                         printf(fmt, ## args);   \
  84         } while (0)
  85
  86 static unsigned int cpu_affinities[NR_CPUS];
  87 static unsigned int next_aff = 0;
  88 static int use_affinity = 0;
  89
  90 pthread_mutex_t affinity_mutex = PTHREAD_MUTEX_INITIALIZER;
  91
  92 static void set_affinity(void)
  93 {
  94 #if HAVE_SCHED_SETAFFINITY
  95         cpu_set_t mask;
  96         int cpu, ret;
  97 #endif /* HAVE_SCHED_SETAFFINITY */
  98
  99         if (!use_affinity)
 100                 return;
 101
 102 #if HAVE_SCHED_SETAFFINITY
 103         ret = pthread_mutex_lock(&affinity_mutex);
 104         if (ret) {
 105                 perror("Error in pthread mutex lock");
 106                 exit(-1);
 107         }
 108         cpu = cpu_affinities[next_aff++];
 109         ret = pthread_mutex_unlock(&affinity_mutex);
 110         if (ret) {
 111                 perror("Error in pthread mutex unlock");
 112                 exit(-1);
 113         }
 114
 115         CPU_ZERO(&mask);
 116         CPU_SET(cpu, &mask);
 117 #if SCHED_SETAFFINITY_ARGS == 2
 118         sched_setaffinity(0, &mask);
 119 #else
 120         sched_setaffinity(0, sizeof(mask), &mask);
 121 #endif
 122 #endif /* HAVE_SCHED_SETAFFINITY */
 123 }
 124
 125 /*
 126  * returns 0 if test should end.
 127  */
 128 static int test_duration_dequeue(void)
 129 {
 130         return !test_stop_dequeue;
 131 }
 132
 133 static int test_duration_enqueue(void)
 134 {
 135         return !test_stop_enqueue;
 136 }
 137
 138 static DEFINE_URCU_TLS(unsigned long long, nr_dequeues);
 139 static DEFINE_URCU_TLS(unsigned long long, nr_enqueues);
 140
 141 static DEFINE_URCU_TLS(unsigned long long, nr_successful_dequeues);
 142 static DEFINE_URCU_TLS(unsigned long long, nr_successful_enqueues);
 143 static DEFINE_URCU_TLS(unsigned long long, nr_empty_dest_enqueues);
 144 static DEFINE_URCU_TLS(unsigned long long, nr_splice);
 145 static DEFINE_URCU_TLS(unsigned long long, nr_dequeue_last);
 146
 147 static unsigned int nr_enqueuers;
 148 static unsigned int nr_dequeuers;
 149
 150 static struct cds_wfcq_head __attribute__((aligned(CAA_CACHE_LINE_SIZE))) head;
 151 static struct cds_wfcq_tail __attribute__((aligned(CAA_CACHE_LINE_SIZE))) tail;
 152
 153 static void *thr_enqueuer(void *_count)
 154 {
 155         unsigned long long *count = _count;
 156         bool was_nonempty;
 157
 158         printf_verbose("thread_begin %s, tid %lu\n",
 159                         "enqueuer", urcu_get_thread_id());
 160
 161         set_affinity();
 162
 163         while (!test_go)
 164         {
 165         }
 166         cmm_smp_mb();
 167
 168         for (;;) {
 169                 struct cds_wfcq_node *node = malloc(sizeof(*node));
 170                 if (!node)
 171                         goto fail;
 172                 cds_wfcq_node_init(node);
 173                 was_nonempty = cds_wfcq_enqueue(&head, &tail, node);
 174                 URCU_TLS(nr_successful_enqueues)++;
 175                 if (!was_nonempty)
 176                         URCU_TLS(nr_empty_dest_enqueues)++;
 177
 178                 if (caa_unlikely(wdelay))
 179                         loop_sleep(wdelay);
 180 fail:
 181                 URCU_TLS(nr_enqueues)++;
 182                 if (caa_unlikely(!test_duration_enqueue()))
 183                         break;
 184         }
 185
 186         uatomic_inc(&test_enqueue_stopped);
 187         count[0] = URCU_TLS(nr_enqueues);
 188         count[1] = URCU_TLS(nr_successful_enqueues);
 189         count[2] = URCU_TLS(nr_empty_dest_enqueues);
 190         printf_verbose("enqueuer thread_end, tid %lu, "
 191                         "enqueues %llu successful_enqueues %llu, "
 192                         "empty_dest_enqueues %llu\n",
 193                         urcu_get_thread_id(),
 194                         URCU_TLS(nr_enqueues),
 195                         URCU_TLS(nr_successful_enqueues),
 196                         URCU_TLS(nr_empty_dest_enqueues));
 197         return ((void*)1);
 198
 199 }
 200
 201 static void do_test_dequeue(enum test_sync sync)
 202 {
 203         struct cds_wfcq_node *node;
 204         int state;
 205
 206         if (sync == TEST_SYNC_MUTEX)
 207                 node = cds_wfcq_dequeue_with_state_blocking(&head, &tail,
 208                                 &state);
 209         else
 210                 node = __cds_wfcq_dequeue_with_state_blocking(&head, &tail,
 211                                 &state);
 212
 213         if (state & CDS_WFCQ_STATE_LAST)
 214                 URCU_TLS(nr_dequeue_last)++;
 215
 216         if (node) {
 217                 free(node);
 218                 URCU_TLS(nr_successful_dequeues)++;
 219         }
 220         URCU_TLS(nr_dequeues)++;
 221 }
 222
 223 static void do_test_splice(enum test_sync sync)
 224 {
 225         struct cds_wfcq_head tmp_head;
 226         struct cds_wfcq_tail tmp_tail;
 227         struct cds_wfcq_node *node, *n;
 228         enum cds_wfcq_ret ret;
 229
 230         cds_wfcq_init(&tmp_head, &tmp_tail);
 231
 232         if (sync == TEST_SYNC_MUTEX)
 233                 ret = cds_wfcq_splice_blocking(&tmp_head, &tmp_tail,
 234                         &head, &tail);
 235         else
 236                 ret = __cds_wfcq_splice_blocking(&tmp_head, &tmp_tail,
 237                         &head, &tail);
 238
 239         switch (ret) {
 240         case CDS_WFCQ_RET_WOULDBLOCK:
 241                 assert(0);      /* blocking call */
 242                 break;
 243         case CDS_WFCQ_RET_DEST_EMPTY:
 244                 URCU_TLS(nr_splice)++;
 245                 URCU_TLS(nr_dequeue_last)++;
 246                 /* ok */
 247                 break;
 248         case CDS_WFCQ_RET_DEST_NON_EMPTY:
 249                 assert(0);      /* entirely unexpected */
 250                 break;
 251         case CDS_WFCQ_RET_SRC_EMPTY:
 252                 /* ok, we could even skip iteration on dest if we wanted */
 253                 break;
 254         }
 255
 256         __cds_wfcq_for_each_blocking_safe(&tmp_head, &tmp_tail, node, n) {
 257                 free(node);
 258                 URCU_TLS(nr_successful_dequeues)++;
 259                 URCU_TLS(nr_dequeues)++;
 260         }
 261         cds_wfcq_destroy(&tmp_head, &tmp_tail);
 262 }
 263
 264 static void *thr_dequeuer(void *_count)
 265 {
 266         unsigned long long *count = _count;
 267         unsigned int counter = 0;
 268
 269         printf_verbose("thread_begin %s, tid %lu\n",
 270                         "dequeuer", urcu_get_thread_id());
 271
 272         set_affinity();
 273
 274         while (!test_go)
 275         {
 276         }
 277         cmm_smp_mb();
 278
 279         for (;;) {
 280                 if (test_dequeue && test_splice) {
 281                         if (counter & 1)
 282                                 do_test_dequeue(test_sync);
 283                         else
 284                                 do_test_splice(test_sync);
 285                         counter++;
 286                 } else {
 287                         if (test_dequeue)
 288                                 do_test_dequeue(test_sync);
 289                         else
 290                                 do_test_splice(test_sync);
 291                 }
 292                 if (caa_unlikely(!test_duration_dequeue()))
 293                         break;
 294                 if (caa_unlikely(rduration))
 295                         loop_sleep(rduration);
 296         }
 297
 298         printf_verbose("dequeuer thread_end, tid %lu, "
 299                         "dequeues %llu, successful_dequeues %llu, "
 300                         "nr_splice %llu\n",
 301                         urcu_get_thread_id(),
 302                         URCU_TLS(nr_dequeues), URCU_TLS(nr_successful_dequeues),
 303                         URCU_TLS(nr_splice));
 304         count[0] = URCU_TLS(nr_dequeues);
 305         count[1] = URCU_TLS(nr_successful_dequeues);
 306         count[2] = URCU_TLS(nr_splice);
 307         count[3] = URCU_TLS(nr_dequeue_last);
 308         return ((void*)2);
 309 }
 310
 311 static void test_end(unsigned long long *nr_dequeues,
 312                 unsigned long long *nr_dequeue_last)
 313 {
 314         struct cds_wfcq_node *node;
 315         int state;
 316
 317         do {
 318                 node = cds_wfcq_dequeue_with_state_blocking(&head, &tail,
 319                                 &state);
 320                 if (node) {
 321                         if (state & CDS_WFCQ_STATE_LAST)
 322                                 (*nr_dequeue_last)++;
 323                         free(node);
 324                         (*nr_dequeues)++;
 325                 }
 326         } while (node);
 327 }
 328
 329 static void show_usage(int argc, char **argv)
 330 {
 331         printf("Usage : %s nr_dequeuers nr_enqueuers duration (s) <OPTIONS>\n",
 332                 argv[0]);
 333         printf("OPTIONS:\n");
 334         printf("        [-d delay] (enqueuer period (in loops))\n");
 335         printf("        [-c duration] (dequeuer period (in loops))\n");
 336         printf("        [-v] (verbose output)\n");
 337         printf("        [-a cpu#] [-a cpu#]... (affinity)\n");
 338         printf("        [-q] (test dequeue)\n");
 339         printf("        [-s] (test splice, enabled by default)\n");
 340         printf("        [-M] (use mutex external synchronization)\n");
 341         printf("                Note: default: no external synchronization used.\n");
 342         printf("        [-f] (force user-provided synchronization)\n");
 343         printf("        [-w] Wait for dequeuer to empty queue\n");
 344         printf("\n");
 345 }
 346
 347 int main(int argc, char **argv)
 348 {
 349         int err;
 350         pthread_t *tid_enqueuer, *tid_dequeuer;
 351         void *tret;
 352         unsigned long long *count_enqueuer, *count_dequeuer;
 353         unsigned long long tot_enqueues = 0, tot_dequeues = 0;
 354         unsigned long long tot_successful_enqueues = 0,
 355                            tot_successful_dequeues = 0,
 356                            tot_empty_dest_enqueues = 0,
 357                            tot_splice = 0, tot_dequeue_last = 0;
 358         unsigned long long end_dequeues = 0;
 359         int i, a, retval = 0;
 360
 361         if (argc < 4) {
 362                 show_usage(argc, argv);
 363                 return -1;
 364         }
 365
 366         err = sscanf(argv[1], "%u", &nr_dequeuers);
 367         if (err != 1) {
 368                 show_usage(argc, argv);
 369                 return -1;
 370         }
 371
 372         err = sscanf(argv[2], "%u", &nr_enqueuers);
 373         if (err != 1) {
 374                 show_usage(argc, argv);
 375                 return -1;
 376         }
 377
 378         err = sscanf(argv[3], "%lu", &duration);
 379         if (err != 1) {
 380                 show_usage(argc, argv);
 381                 return -1;
 382         }
 383
 384         for (i = 4; i < argc; i++) {
 385                 if (argv[i][0] != '-')
 386                         continue;
 387                 switch (argv[i][1]) {
 388                 case 'a':
 389                         if (argc < i + 2) {
 390                                 show_usage(argc, argv);
 391                                 return -1;
 392                         }
 393                         a = atoi(argv[++i]);
 394                         cpu_affinities[next_aff++] = a;
 395                         use_affinity = 1;
 396                         printf_verbose("Adding CPU %d affinity\n", a);
 397                         break;
 398                 case 'c':
 399                         if (argc < i + 2) {
 400                                 show_usage(argc, argv);
 401                                 return -1;
 402                         }
 403                         rduration = atol(argv[++i]);
 404                         break;
 405                 case 'd':
 406                         if (argc < i + 2) {
 407                                 show_usage(argc, argv);
 408                                 return -1;
 409                         }
 410                         wdelay = atol(argv[++i]);
 411                         break;
 412                 case 'v':
 413                         verbose_mode = 1;
 414                         break;
 415                 case 'q':
 416                         test_dequeue = 1;
 417                         break;
 418                 case 's':
 419                         test_splice = 1;
 420                         break;
 421                 case 'M':
 422                         test_sync = TEST_SYNC_MUTEX;
 423                         break;
 424                 case 'w':
 425                         test_wait_empty = 1;
 426                         break;
 427                 case 'f':
 428                         test_force_sync = 1;
 429                         break;
 430                 }
 431         }
 432
 433         /* activate splice test by default */
 434         if (!test_dequeue && !test_splice)
 435                 test_splice = 1;
 436
 437         if (test_sync == TEST_SYNC_NONE && nr_dequeuers > 1 && test_dequeue) {
 438                 if (test_force_sync) {
 439                         fprintf(stderr, "[WARNING] Using dequeue concurrently "
 440                                 "with other dequeue or splice without external "
 441                                 "synchronization. Expect run-time failure.\n");
 442                 } else {
 443                         printf("Enforcing mutex synchronization\n");
 444                         test_sync = TEST_SYNC_MUTEX;
 445                 }
 446         }
 447
 448         printf_verbose("running test for %lu seconds, %u enqueuers, "
 449                        "%u dequeuers.\n",
 450                        duration, nr_enqueuers, nr_dequeuers);
 451         if (test_dequeue)
 452                 printf_verbose("dequeue test activated.\n");
 453         else
 454                 printf_verbose("splice test activated.\n");
 455         if (test_sync == TEST_SYNC_MUTEX)
 456                 printf_verbose("External sync: mutex.\n");
 457         else
 458                 printf_verbose("External sync: none.\n");
 459         if (test_wait_empty)
 460                 printf_verbose("Wait for dequeuers to empty queue.\n");
 461         printf_verbose("Writer delay : %lu loops.\n", rduration);
 462         printf_verbose("Reader duration : %lu loops.\n", wdelay);
 463         printf_verbose("thread %-6s, tid %lu\n",
 464                         "main", urcu_get_thread_id());
 465
 466         tid_enqueuer = calloc(nr_enqueuers, sizeof(*tid_enqueuer));
 467         tid_dequeuer = calloc(nr_dequeuers, sizeof(*tid_dequeuer));
 468         count_enqueuer = calloc(nr_enqueuers, 3 * sizeof(*count_enqueuer));
 469         count_dequeuer = calloc(nr_dequeuers, 4 * sizeof(*count_dequeuer));
 470         cds_wfcq_init(&head, &tail);
 471
 472         next_aff = 0;
 473
 474         for (i = 0; i < nr_enqueuers; i++) {
 475                 err = pthread_create(&tid_enqueuer[i], NULL, thr_enqueuer,
 476                                      &count_enqueuer[3 * i]);
 477                 if (err != 0)
 478                         exit(1);
 479         }
 480         for (i = 0; i < nr_dequeuers; i++) {
 481                 err = pthread_create(&tid_dequeuer[i], NULL, thr_dequeuer,
 482                                      &count_dequeuer[4 * i]);
 483                 if (err != 0)
 484                         exit(1);
 485         }
 486
 487         cmm_smp_mb();
 488
 489         test_go = 1;
 490
 491         for (i = 0; i < duration; i++) {
 492                 sleep(1);
 493                 if (verbose_mode) {
 494                         fwrite(".", sizeof(char), 1, stdout);
 495                         fflush(stdout);
 496                 }
 497         }
 498
 499         test_stop_enqueue = 1;
 500
 501         if (test_wait_empty) {
 502                 while (nr_enqueuers != uatomic_read(&test_enqueue_stopped)) {
 503                         sleep(1);
 504                 }
 505                 while (!cds_wfcq_empty(&head, &tail)) {
 506                         sleep(1);
 507                 }
 508         }
 509
 510         test_stop_dequeue = 1;
 511
 512         for (i = 0; i < nr_enqueuers; i++) {
 513                 err = pthread_join(tid_enqueuer[i], &tret);
 514                 if (err != 0)
 515                         exit(1);
 516                 tot_enqueues += count_enqueuer[3 * i];
 517                 tot_successful_enqueues += count_enqueuer[3 * i + 1];
 518                 tot_empty_dest_enqueues += count_enqueuer[3 * i + 2];
 519         }
 520         for (i = 0; i < nr_dequeuers; i++) {
 521                 err = pthread_join(tid_dequeuer[i], &tret);
 522                 if (err != 0)
 523                         exit(1);
 524                 tot_dequeues += count_dequeuer[4 * i];
 525                 tot_successful_dequeues += count_dequeuer[4 * i + 1];
 526                 tot_splice += count_dequeuer[4 * i + 2];
 527                 tot_dequeue_last += count_dequeuer[4 * i + 3];
 528         }
 529
 530         test_end(&end_dequeues, &tot_dequeue_last);
 531
 532         printf_verbose("total number of enqueues : %llu, dequeues %llu\n",
 533                        tot_enqueues, tot_dequeues);
 534         printf_verbose("total number of successful enqueues : %llu, "
 535                        "enqueues to empty dest : %llu, "
 536                        "successful dequeues %llu, "
 537                        "splice : %llu, dequeue_last : %llu\n",
 538                        tot_successful_enqueues,
 539                        tot_empty_dest_enqueues,
 540                        tot_successful_dequeues,
 541                        tot_splice, tot_dequeue_last);
 542         printf("SUMMARY %-25s testdur %4lu nr_enqueuers %3u wdelay %6lu "
 543                 "nr_dequeuers %3u "
 544                 "rdur %6lu nr_enqueues %12llu nr_dequeues %12llu "
 545                 "successful enqueues %12llu enqueues to empty dest %12llu "
 546                 "successful dequeues %12llu splice %12llu "
 547                 "dequeue_last %llu "
 548                 "end_dequeues %llu nr_ops %12llu\n",
 549                 argv[0], duration, nr_enqueuers, wdelay,
 550                 nr_dequeuers, rduration, tot_enqueues, tot_dequeues,
 551                 tot_successful_enqueues,
 552                 tot_empty_dest_enqueues,
 553                 tot_successful_dequeues, tot_splice, tot_dequeue_last,
 554                 end_dequeues,
 555                 tot_enqueues + tot_dequeues);
 556
 557         if (tot_successful_enqueues != tot_successful_dequeues + end_dequeues) {
 558                 printf("WARNING! Discrepancy between nr succ. enqueues %llu vs "
 559                        "succ. dequeues + end dequeues %llu.\n",
 560                        tot_successful_enqueues,
 561                        tot_successful_dequeues + end_dequeues);
 562                 retval = 1;
 563         }
 564
 565         /*
 566          * If only using splice to dequeue, the enqueuer should see
 567          * exactly as many empty queues than the number of non-empty
 568          * src splice.
 569          */
 570         if (tot_empty_dest_enqueues != tot_dequeue_last) {
 571                 printf("WARNING! Discrepancy between empty enqueue (%llu) and "
 572                         "number of dequeue of last element (%llu)\n",
 573                         tot_empty_dest_enqueues,
 574                         tot_dequeue_last);
 575                 retval = 1;
 576         }
 577         cds_wfcq_destroy(&head, &tail);
 578         free(count_enqueuer);
 579         free(count_dequeuer);
 580         free(tid_enqueuer);
 581         free(tid_dequeuer);
 582         return retval;
 583 }