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 <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/wfcqueue.h>
  49
  50 enum test_sync {
  51         TEST_SYNC_NONE = 0,
  52         TEST_SYNC_MUTEX,
  53 };
  54
  55 static enum test_sync test_sync;
  56
  57 static int test_force_sync;
  58
  59 static volatile int test_go, test_stop_enqueue, test_stop_dequeue;
  60
  61 static unsigned long rduration;
  62
  63 static unsigned long duration;
  64
  65 /* read-side C.S. duration, in loops */
  66 static unsigned long wdelay;
  67
  68 static inline void loop_sleep(unsigned long loops)
  69 {
  70         while (loops-- != 0)
  71                 caa_cpu_relax();
  72 }
  73
  74 static int verbose_mode;
  75
  76 static int test_dequeue, test_splice, test_wait_empty;
  77 static unsigned int test_enqueue_stopped;
  78
  79 #define printf_verbose(fmt, args...)            \
  80         do {                                    \
  81                 if (verbose_mode)               \
  82                         printf(fmt, ## args);   \
  83         } while (0)
  84
  85 static unsigned int cpu_affinities[NR_CPUS];
  86 static unsigned int next_aff = 0;
  87 static int use_affinity = 0;
  88
  89 pthread_mutex_t affinity_mutex = PTHREAD_MUTEX_INITIALIZER;
  90
  91 static void set_affinity(void)
  92 {
  93 #ifdef HAVE_SCHED_SETAFFINITY
  94         cpu_set_t mask;
  95         int cpu, ret;
  96 #endif /* HAVE_SCHED_SETAFFINITY */
  97
  98         if (!use_affinity)
  99                 return;
 100
 101 #ifdef HAVE_SCHED_SETAFFINITY
 102         ret = pthread_mutex_lock(&affinity_mutex);
 103         if (ret) {
 104                 perror("Error in pthread mutex lock");
 105                 exit(-1);
 106         }
 107         cpu = cpu_affinities[next_aff++];
 108         ret = pthread_mutex_unlock(&affinity_mutex);
 109         if (ret) {
 110                 perror("Error in pthread mutex unlock");
 111                 exit(-1);
 112         }
 113
 114         CPU_ZERO(&mask);
 115         CPU_SET(cpu, &mask);
 116         sched_setaffinity(0, sizeof(mask), &mask);
 117 #endif /* HAVE_SCHED_SETAFFINITY */
 118 }
 119
 120 /*
 121  * returns 0 if test should end.
 122  */
 123 static int test_duration_dequeue(void)
 124 {
 125         return !test_stop_dequeue;
 126 }
 127
 128 static int test_duration_enqueue(void)
 129 {
 130         return !test_stop_enqueue;
 131 }
 132
 133 static DEFINE_URCU_TLS(unsigned long long, nr_dequeues);
 134 static DEFINE_URCU_TLS(unsigned long long, nr_enqueues);
 135
 136 static DEFINE_URCU_TLS(unsigned long long, nr_successful_dequeues);
 137 static DEFINE_URCU_TLS(unsigned long long, nr_successful_enqueues);
 138 static DEFINE_URCU_TLS(unsigned long long, nr_empty_dest_enqueues);
 139 static DEFINE_URCU_TLS(unsigned long long, nr_splice);
 140 static DEFINE_URCU_TLS(unsigned long long, nr_dequeue_last);
 141
 142 static unsigned int nr_enqueuers;
 143 static unsigned int nr_dequeuers;
 144
 145 static struct cds_wfcq_head __attribute__((aligned(CAA_CACHE_LINE_SIZE))) head;
 146 static struct cds_wfcq_tail __attribute__((aligned(CAA_CACHE_LINE_SIZE))) tail;
 147
 148 static void *thr_enqueuer(void *_count)
 149 {
 150         unsigned long long *count = _count;
 151         bool was_nonempty;
 152
 153         printf_verbose("thread_begin %s, tid %lu\n",
 154                         "enqueuer", urcu_get_thread_id());
 155
 156         set_affinity();
 157
 158         while (!test_go)
 159         {
 160         }
 161         cmm_smp_mb();
 162
 163         for (;;) {
 164                 struct cds_wfcq_node *node = malloc(sizeof(*node));
 165                 if (!node)
 166                         goto fail;
 167                 cds_wfcq_node_init(node);
 168                 was_nonempty = cds_wfcq_enqueue(&head, &tail, node);
 169                 URCU_TLS(nr_successful_enqueues)++;
 170                 if (!was_nonempty)
 171                         URCU_TLS(nr_empty_dest_enqueues)++;
 172
 173                 if (caa_unlikely(wdelay))
 174                         loop_sleep(wdelay);
 175 fail:
 176                 URCU_TLS(nr_enqueues)++;
 177                 if (caa_unlikely(!test_duration_enqueue()))
 178                         break;
 179         }
 180
 181         uatomic_inc(&test_enqueue_stopped);
 182         count[0] = URCU_TLS(nr_enqueues);
 183         count[1] = URCU_TLS(nr_successful_enqueues);
 184         count[2] = URCU_TLS(nr_empty_dest_enqueues);
 185         printf_verbose("enqueuer thread_end, tid %lu, "
 186                         "enqueues %llu successful_enqueues %llu, "
 187                         "empty_dest_enqueues %llu\n",
 188                         urcu_get_thread_id(),
 189                         URCU_TLS(nr_enqueues),
 190                         URCU_TLS(nr_successful_enqueues),
 191                         URCU_TLS(nr_empty_dest_enqueues));
 192         return ((void*)1);
 193
 194 }
 195
 196 static void do_test_dequeue(enum test_sync sync)
 197 {
 198         struct cds_wfcq_node *node;
 199         int state;
 200
 201         if (sync == TEST_SYNC_MUTEX)
 202                 node = cds_wfcq_dequeue_with_state_blocking(&head, &tail,
 203                                 &state);
 204         else
 205                 node = __cds_wfcq_dequeue_with_state_blocking(&head, &tail,
 206                                 &state);
 207
 208         if (state & CDS_WFCQ_STATE_LAST)
 209                 URCU_TLS(nr_dequeue_last)++;
 210
 211         if (node) {
 212                 free(node);
 213                 URCU_TLS(nr_successful_dequeues)++;
 214         }
 215         URCU_TLS(nr_dequeues)++;
 216 }
 217
 218 static void do_test_splice(enum test_sync sync)
 219 {
 220         struct cds_wfcq_head tmp_head;
 221         struct cds_wfcq_tail tmp_tail;
 222         struct cds_wfcq_node *node, *n;
 223         enum cds_wfcq_ret ret;
 224
 225         cds_wfcq_init(&tmp_head, &tmp_tail);
 226
 227         if (sync == TEST_SYNC_MUTEX)
 228                 ret = cds_wfcq_splice_blocking(&tmp_head, &tmp_tail,
 229                         &head, &tail);
 230         else
 231                 ret = __cds_wfcq_splice_blocking(&tmp_head, &tmp_tail,
 232                         &head, &tail);
 233
 234         switch (ret) {
 235         case CDS_WFCQ_RET_WOULDBLOCK:
 236                 urcu_posix_assert(0);   /* blocking call */
 237                 break;
 238         case CDS_WFCQ_RET_DEST_EMPTY:
 239                 URCU_TLS(nr_splice)++;
 240                 URCU_TLS(nr_dequeue_last)++;
 241                 /* ok */
 242                 break;
 243         case CDS_WFCQ_RET_DEST_NON_EMPTY:
 244                 urcu_posix_assert(0);   /* entirely unexpected */
 245                 break;
 246         case CDS_WFCQ_RET_SRC_EMPTY:
 247                 /* ok, we could even skip iteration on dest if we wanted */
 248                 break;
 249         }
 250
 251         __cds_wfcq_for_each_blocking_safe(&tmp_head, &tmp_tail, node, n) {
 252                 free(node);
 253                 URCU_TLS(nr_successful_dequeues)++;
 254                 URCU_TLS(nr_dequeues)++;
 255         }
 256         cds_wfcq_destroy(&tmp_head, &tmp_tail);
 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         while (!test_go)
 270         {
 271         }
 272         cmm_smp_mb();
 273
 274         for (;;) {
 275                 if (test_dequeue && test_splice) {
 276                         if (counter & 1)
 277                                 do_test_dequeue(test_sync);
 278                         else
 279                                 do_test_splice(test_sync);
 280                         counter++;
 281                 } else {
 282                         if (test_dequeue)
 283                                 do_test_dequeue(test_sync);
 284                         else
 285                                 do_test_splice(test_sync);
 286                 }
 287                 if (caa_unlikely(!test_duration_dequeue()))
 288                         break;
 289                 if (caa_unlikely(rduration))
 290                         loop_sleep(rduration);
 291         }
 292
 293         printf_verbose("dequeuer thread_end, tid %lu, "
 294                         "dequeues %llu, successful_dequeues %llu, "
 295                         "nr_splice %llu\n",
 296                         urcu_get_thread_id(),
 297                         URCU_TLS(nr_dequeues), URCU_TLS(nr_successful_dequeues),
 298                         URCU_TLS(nr_splice));
 299         count[0] = URCU_TLS(nr_dequeues);
 300         count[1] = URCU_TLS(nr_successful_dequeues);
 301         count[2] = URCU_TLS(nr_splice);
 302         count[3] = URCU_TLS(nr_dequeue_last);
 303         return ((void*)2);
 304 }
 305
 306 static void test_end(unsigned long long *nr_dequeues_l,
 307                 unsigned long long *nr_dequeue_last_l)
 308 {
 309         struct cds_wfcq_node *node;
 310         int state;
 311
 312         do {
 313                 node = cds_wfcq_dequeue_with_state_blocking(&head, &tail,
 314                                 &state);
 315                 if (node) {
 316                         if (state & CDS_WFCQ_STATE_LAST)
 317                                 (*nr_dequeue_last_l)++;
 318                         free(node);
 319                         (*nr_dequeues_l)++;
 320                 }
 321         } while (node);
 322 }
 323
 324 static void show_usage(char **argv)
 325 {
 326         printf("Usage : %s nr_dequeuers nr_enqueuers duration (s) <OPTIONS>\n",
 327                 argv[0]);
 328         printf("OPTIONS:\n");
 329         printf("        [-d delay] (enqueuer period (in loops))\n");
 330         printf("        [-c duration] (dequeuer period (in loops))\n");
 331         printf("        [-v] (verbose output)\n");
 332         printf("        [-a cpu#] [-a cpu#]... (affinity)\n");
 333         printf("        [-q] (test dequeue)\n");
 334         printf("        [-s] (test splice, enabled by default)\n");
 335         printf("        [-M] (use mutex external synchronization)\n");
 336         printf("                Note: default: no external synchronization used.\n");
 337         printf("        [-f] (force user-provided synchronization)\n");
 338         printf("        [-w] Wait for dequeuer to empty queue\n");
 339         printf("\n");
 340 }
 341
 342 int main(int argc, char **argv)
 343 {
 344         int err;
 345         pthread_t *tid_enqueuer, *tid_dequeuer;
 346         void *tret;
 347         unsigned long long *count_enqueuer, *count_dequeuer;
 348         unsigned long long tot_enqueues = 0, tot_dequeues = 0;
 349         unsigned long long tot_successful_enqueues = 0,
 350                            tot_successful_dequeues = 0,
 351                            tot_empty_dest_enqueues = 0,
 352                            tot_splice = 0, tot_dequeue_last = 0;
 353         unsigned long long end_dequeues = 0;
 354         int i, a, retval = 0;
 355         unsigned int i_thr;
 356
 357         if (argc < 4) {
 358                 show_usage(argv);
 359                 return -1;
 360         }
 361
 362         err = sscanf(argv[1], "%u", &nr_dequeuers);
 363         if (err != 1) {
 364                 show_usage(argv);
 365                 return -1;
 366         }
 367
 368         err = sscanf(argv[2], "%u", &nr_enqueuers);
 369         if (err != 1) {
 370                 show_usage(argv);
 371                 return -1;
 372         }
 373
 374         err = sscanf(argv[3], "%lu", &duration);
 375         if (err != 1) {
 376                 show_usage(argv);
 377                 return -1;
 378         }
 379
 380         for (i = 4; i < argc; i++) {
 381                 if (argv[i][0] != '-')
 382                         continue;
 383                 switch (argv[i][1]) {
 384                 case 'a':
 385                         if (argc < i + 2) {
 386                                 show_usage(argv);
 387                                 return -1;
 388                         }
 389                         a = atoi(argv[++i]);
 390                         cpu_affinities[next_aff++] = a;
 391                         use_affinity = 1;
 392                         printf_verbose("Adding CPU %d affinity\n", a);
 393                         break;
 394                 case 'c':
 395                         if (argc < i + 2) {
 396                                 show_usage(argv);
 397                                 return -1;
 398                         }
 399                         rduration = atol(argv[++i]);
 400                         break;
 401                 case 'd':
 402                         if (argc < i + 2) {
 403                                 show_usage(argv);
 404                                 return -1;
 405                         }
 406                         wdelay = atol(argv[++i]);
 407                         break;
 408                 case 'v':
 409                         verbose_mode = 1;
 410                         break;
 411                 case 'q':
 412                         test_dequeue = 1;
 413                         break;
 414                 case 's':
 415                         test_splice = 1;
 416                         break;
 417                 case 'M':
 418                         test_sync = TEST_SYNC_MUTEX;
 419                         break;
 420                 case 'w':
 421                         test_wait_empty = 1;
 422                         break;
 423                 case 'f':
 424                         test_force_sync = 1;
 425                         break;
 426                 }
 427         }
 428
 429         /* activate splice test by default */
 430         if (!test_dequeue && !test_splice)
 431                 test_splice = 1;
 432
 433         if (test_sync == TEST_SYNC_NONE && nr_dequeuers > 1 && test_dequeue) {
 434                 if (test_force_sync) {
 435                         fprintf(stderr, "[WARNING] Using dequeue concurrently "
 436                                 "with other dequeue or splice without external "
 437                                 "synchronization. Expect run-time failure.\n");
 438                 } else {
 439                         printf("Enforcing mutex synchronization\n");
 440                         test_sync = TEST_SYNC_MUTEX;
 441                 }
 442         }
 443
 444         printf_verbose("running test for %lu seconds, %u enqueuers, "
 445                        "%u dequeuers.\n",
 446                        duration, nr_enqueuers, nr_dequeuers);
 447         if (test_dequeue)
 448                 printf_verbose("dequeue test activated.\n");
 449         else
 450                 printf_verbose("splice test activated.\n");
 451         if (test_sync == TEST_SYNC_MUTEX)
 452                 printf_verbose("External sync: mutex.\n");
 453         else
 454                 printf_verbose("External sync: none.\n");
 455         if (test_wait_empty)
 456                 printf_verbose("Wait for dequeuers to empty queue.\n");
 457         printf_verbose("Writer delay : %lu loops.\n", rduration);
 458         printf_verbose("Reader duration : %lu loops.\n", wdelay);
 459         printf_verbose("thread %-6s, tid %lu\n",
 460                         "main", urcu_get_thread_id());
 461
 462         tid_enqueuer = calloc(nr_enqueuers, sizeof(*tid_enqueuer));
 463         tid_dequeuer = calloc(nr_dequeuers, sizeof(*tid_dequeuer));
 464         count_enqueuer = calloc(nr_enqueuers, 3 * sizeof(*count_enqueuer));
 465         count_dequeuer = calloc(nr_dequeuers, 4 * sizeof(*count_dequeuer));
 466         cds_wfcq_init(&head, &tail);
 467
 468         next_aff = 0;
 469
 470         for (i_thr = 0; i_thr < nr_enqueuers; i_thr++) {
 471                 err = pthread_create(&tid_enqueuer[i_thr], NULL, thr_enqueuer,
 472                                      &count_enqueuer[3 * i_thr]);
 473                 if (err != 0)
 474                         exit(1);
 475         }
 476         for (i_thr = 0; i_thr < nr_dequeuers; i_thr++) {
 477                 err = pthread_create(&tid_dequeuer[i_thr], NULL, thr_dequeuer,
 478                                      &count_dequeuer[4 * i_thr]);
 479                 if (err != 0)
 480                         exit(1);
 481         }
 482
 483         cmm_smp_mb();
 484
 485         test_go = 1;
 486
 487         for (i_thr = 0; i_thr < duration; i_thr++) {
 488                 sleep(1);
 489                 if (verbose_mode) {
 490                         fwrite(".", sizeof(char), 1, stdout);
 491                         fflush(stdout);
 492                 }
 493         }
 494
 495         test_stop_enqueue = 1;
 496
 497         if (test_wait_empty) {
 498                 while (nr_enqueuers != uatomic_read(&test_enqueue_stopped)) {
 499                         sleep(1);
 500                 }
 501                 while (!cds_wfcq_empty(&head, &tail)) {
 502                         sleep(1);
 503                 }
 504         }
 505
 506         test_stop_dequeue = 1;
 507
 508         for (i_thr = 0; i_thr < nr_enqueuers; i_thr++) {
 509                 err = pthread_join(tid_enqueuer[i_thr], &tret);
 510                 if (err != 0)
 511                         exit(1);
 512                 tot_enqueues += count_enqueuer[3 * i_thr];
 513                 tot_successful_enqueues += count_enqueuer[3 * i_thr + 1];
 514                 tot_empty_dest_enqueues += count_enqueuer[3 * i_thr + 2];
 515         }
 516         for (i_thr = 0; i_thr < nr_dequeuers; i_thr++) {
 517                 err = pthread_join(tid_dequeuer[i_thr], &tret);
 518                 if (err != 0)
 519                         exit(1);
 520                 tot_dequeues += count_dequeuer[4 * i_thr];
 521                 tot_successful_dequeues += count_dequeuer[4 * i_thr + 1];
 522                 tot_splice += count_dequeuer[4 * i_thr + 2];
 523                 tot_dequeue_last += count_dequeuer[4 * i_thr + 3];
 524         }
 525
 526         test_end(&end_dequeues, &tot_dequeue_last);
 527
 528         printf_verbose("total number of enqueues : %llu, dequeues %llu\n",
 529                        tot_enqueues, tot_dequeues);
 530         printf_verbose("total number of successful enqueues : %llu, "
 531                        "enqueues to empty dest : %llu, "
 532                        "successful dequeues %llu, "
 533                        "splice : %llu, dequeue_last : %llu\n",
 534                        tot_successful_enqueues,
 535                        tot_empty_dest_enqueues,
 536                        tot_successful_dequeues,
 537                        tot_splice, tot_dequeue_last);
 538         printf("SUMMARY %-25s testdur %4lu nr_enqueuers %3u wdelay %6lu "
 539                 "nr_dequeuers %3u "
 540                 "rdur %6lu nr_enqueues %12llu nr_dequeues %12llu "
 541                 "successful enqueues %12llu enqueues to empty dest %12llu "
 542                 "successful dequeues %12llu splice %12llu "
 543                 "dequeue_last %llu "
 544                 "end_dequeues %llu nr_ops %12llu\n",
 545                 argv[0], duration, nr_enqueuers, wdelay,
 546                 nr_dequeuers, rduration, tot_enqueues, tot_dequeues,
 547                 tot_successful_enqueues,
 548                 tot_empty_dest_enqueues,
 549                 tot_successful_dequeues, tot_splice, tot_dequeue_last,
 550                 end_dequeues,
 551                 tot_enqueues + tot_dequeues);
 552
 553         if (tot_successful_enqueues != tot_successful_dequeues + end_dequeues) {
 554                 printf("WARNING! Discrepancy between nr succ. enqueues %llu vs "
 555                        "succ. dequeues + end dequeues %llu.\n",
 556                        tot_successful_enqueues,
 557                        tot_successful_dequeues + end_dequeues);
 558                 retval = 1;
 559         }
 560
 561         /*
 562          * If only using splice to dequeue, the enqueuer should see
 563          * exactly as many empty queues than the number of non-empty
 564          * src splice.
 565          */
 566         if (tot_empty_dest_enqueues != tot_dequeue_last) {
 567                 printf("WARNING! Discrepancy between empty enqueue (%llu) and "
 568                         "number of dequeue of last element (%llu)\n",
 569                         tot_empty_dest_enqueues,
 570                         tot_dequeue_last);
 571                 retval = 1;
 572         }
 573         cds_wfcq_destroy(&head, &tail);
 574         free(count_enqueuer);
 575         free(count_dequeuer);
 576         free(tid_enqueuer);
 577         free(tid_dequeuer);
 578
 579         return retval;
 580 }