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 #define _GNU_SOURCE
  25 #include "config.h"
  26 #include <stdio.h>
  27 #include <pthread.h>
  28 #include <stdlib.h>
  29 #include <stdint.h>
  30 #include <stdbool.h>
  31 #include <string.h>
  32 #include <sys/types.h>
  33 #include <sys/wait.h>
  34 #include <unistd.h>
  35 #include <stdio.h>
  36 #include <assert.h>
  37 #include <errno.h>
  38
  39 #include <urcu/arch.h>
  40 #include <urcu/tls-compat.h>
  41 #include <urcu/uatomic.h>
  42 #include "cpuset.h"
  43 #include "thread-id.h"
  44
  45 /* hardcoded number of CPUs */
  46 #define NR_CPUS 16384
  47
  48 #ifndef DYNAMIC_LINK_TEST
  49 #define _LGPL_SOURCE
  50 #endif
  51 #include <urcu/wfcqueue.h>
  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_dequeue, test_splice, test_wait_empty;
  80 static 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 #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_dequeue;
 133 }
 134
 135 static int test_duration_enqueue(void)
 136 {
 137         return !test_stop_enqueue;
 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 static DEFINE_URCU_TLS(unsigned long long, nr_empty_dest_enqueues);
 146 static DEFINE_URCU_TLS(unsigned long long, nr_splice);
 147 static DEFINE_URCU_TLS(unsigned long long, nr_dequeue_last);
 148
 149 static unsigned int nr_enqueuers;
 150 static unsigned int nr_dequeuers;
 151
 152 static struct cds_wfcq_head __attribute__((aligned(CAA_CACHE_LINE_SIZE))) head;
 153 static struct cds_wfcq_tail __attribute__((aligned(CAA_CACHE_LINE_SIZE))) tail;
 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_wfcq_node *node = malloc(sizeof(*node));
 172                 if (!node)
 173                         goto fail;
 174                 cds_wfcq_node_init(node);
 175                 was_nonempty = cds_wfcq_enqueue(&head, &tail, 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_dequeue(enum test_sync sync)
 204 {
 205         struct cds_wfcq_node *node;
 206         int state;
 207
 208         if (sync == TEST_SYNC_MUTEX)
 209                 node = cds_wfcq_dequeue_with_state_blocking(&head, &tail,
 210                                 &state);
 211         else
 212                 node = __cds_wfcq_dequeue_with_state_blocking(&head, &tail,
 213                                 &state);
 214
 215         if (state & CDS_WFCQ_STATE_LAST)
 216                 URCU_TLS(nr_dequeue_last)++;
 217
 218         if (node) {
 219                 free(node);
 220                 URCU_TLS(nr_successful_dequeues)++;
 221         }
 222         URCU_TLS(nr_dequeues)++;
 223 }
 224
 225 static void do_test_splice(enum test_sync sync)
 226 {
 227         struct cds_wfcq_head tmp_head;
 228         struct cds_wfcq_tail tmp_tail;
 229         struct cds_wfcq_node *node, *n;
 230         enum cds_wfcq_ret ret;
 231
 232         cds_wfcq_init(&tmp_head, &tmp_tail);
 233
 234         if (sync == TEST_SYNC_MUTEX)
 235                 ret = cds_wfcq_splice_blocking(&tmp_head, &tmp_tail,
 236                         &head, &tail);
 237         else
 238                 ret = __cds_wfcq_splice_blocking(&tmp_head, &tmp_tail,
 239                         &head, &tail);
 240
 241         switch (ret) {
 242         case CDS_WFCQ_RET_WOULDBLOCK:
 243                 assert(0);      /* blocking call */
 244                 break;
 245         case CDS_WFCQ_RET_DEST_EMPTY:
 246                 URCU_TLS(nr_splice)++;
 247                 URCU_TLS(nr_dequeue_last)++;
 248                 /* ok */
 249                 break;
 250         case CDS_WFCQ_RET_DEST_NON_EMPTY:
 251                 assert(0);      /* entirely unexpected */
 252                 break;
 253         case CDS_WFCQ_RET_SRC_EMPTY:
 254                 /* ok, we could even skip iteration on dest if we wanted */
 255                 break;
 256         }
 257
 258         __cds_wfcq_for_each_blocking_safe(&tmp_head, &tmp_tail, node, n) {
 259                 free(node);
 260                 URCU_TLS(nr_successful_dequeues)++;
 261                 URCU_TLS(nr_dequeues)++;
 262         }
 263 }
 264
 265 static void *thr_dequeuer(void *_count)
 266 {
 267         unsigned long long *count = _count;
 268         unsigned int counter = 0;
 269
 270         printf_verbose("thread_begin %s, tid %lu\n",
 271                         "dequeuer", urcu_get_thread_id());
 272
 273         set_affinity();
 274
 275         while (!test_go)
 276         {
 277         }
 278         cmm_smp_mb();
 279
 280         for (;;) {
 281                 if (test_dequeue && test_splice) {
 282                         if (counter & 1)
 283                                 do_test_dequeue(test_sync);
 284                         else
 285                                 do_test_splice(test_sync);
 286                         counter++;
 287                 } else {
 288                         if (test_dequeue)
 289                                 do_test_dequeue(test_sync);
 290                         else
 291                                 do_test_splice(test_sync);
 292                 }
 293                 if (caa_unlikely(!test_duration_dequeue()))
 294                         break;
 295                 if (caa_unlikely(rduration))
 296                         loop_sleep(rduration);
 297         }
 298
 299         printf_verbose("dequeuer thread_end, tid %lu, "
 300                         "dequeues %llu, successful_dequeues %llu, "
 301                         "nr_splice %llu\n",
 302                         urcu_get_thread_id(),
 303                         URCU_TLS(nr_dequeues), URCU_TLS(nr_successful_dequeues),
 304                         URCU_TLS(nr_splice));
 305         count[0] = URCU_TLS(nr_dequeues);
 306         count[1] = URCU_TLS(nr_successful_dequeues);
 307         count[2] = URCU_TLS(nr_splice);
 308         count[3] = URCU_TLS(nr_dequeue_last);
 309         return ((void*)2);
 310 }
 311
 312 static void test_end(unsigned long long *nr_dequeues,
 313                 unsigned long long *nr_dequeue_last)
 314 {
 315         struct cds_wfcq_node *node;
 316         int state;
 317
 318         do {
 319                 node = cds_wfcq_dequeue_with_state_blocking(&head, &tail,
 320                                 &state);
 321                 if (node) {
 322                         if (state & CDS_WFCQ_STATE_LAST)
 323                                 (*nr_dequeue_last)++;
 324                         free(node);
 325                         (*nr_dequeues)++;
 326                 }
 327         } while (node);
 328 }
 329
 330 static void show_usage(int argc, char **argv)
 331 {
 332         printf("Usage : %s nr_dequeuers nr_enqueuers duration (s) <OPTIONS>\n",
 333                 argv[0]);
 334         printf("OPTIONS:\n");
 335         printf("        [-d delay] (enqueuer period (in loops))\n");
 336         printf("        [-c duration] (dequeuer period (in loops))\n");
 337         printf("        [-v] (verbose output)\n");
 338         printf("        [-a cpu#] [-a cpu#]... (affinity)\n");
 339         printf("        [-q] (test dequeue)\n");
 340         printf("        [-s] (test splice, enabled by default)\n");
 341         printf("        [-M] (use mutex external synchronization)\n");
 342         printf("                Note: default: no external synchronization used.\n");
 343         printf("        [-f] (force user-provided synchronization)\n");
 344         printf("        [-w] Wait for dequeuer to empty queue\n");
 345         printf("\n");
 346 }
 347
 348 int main(int argc, char **argv)
 349 {
 350         int err;
 351         pthread_t *tid_enqueuer, *tid_dequeuer;
 352         void *tret;
 353         unsigned long long *count_enqueuer, *count_dequeuer;
 354         unsigned long long tot_enqueues = 0, tot_dequeues = 0;
 355         unsigned long long tot_successful_enqueues = 0,
 356                            tot_successful_dequeues = 0,
 357                            tot_empty_dest_enqueues = 0,
 358                            tot_splice = 0, tot_dequeue_last = 0;
 359         unsigned long long end_dequeues = 0;
 360         int i, a, retval = 0;
 361
 362         if (argc < 4) {
 363                 show_usage(argc, argv);
 364                 return -1;
 365         }
 366
 367         err = sscanf(argv[1], "%u", &nr_dequeuers);
 368         if (err != 1) {
 369                 show_usage(argc, argv);
 370                 return -1;
 371         }
 372
 373         err = sscanf(argv[2], "%u", &nr_enqueuers);
 374         if (err != 1) {
 375                 show_usage(argc, argv);
 376                 return -1;
 377         }
 378
 379         err = sscanf(argv[3], "%lu", &duration);
 380         if (err != 1) {
 381                 show_usage(argc, argv);
 382                 return -1;
 383         }
 384
 385         for (i = 4; i < argc; i++) {
 386                 if (argv[i][0] != '-')
 387                         continue;
 388                 switch (argv[i][1]) {
 389                 case 'a':
 390                         if (argc < i + 2) {
 391                                 show_usage(argc, argv);
 392                                 return -1;
 393                         }
 394                         a = atoi(argv[++i]);
 395                         cpu_affinities[next_aff++] = a;
 396                         use_affinity = 1;
 397                         printf_verbose("Adding CPU %d affinity\n", a);
 398                         break;
 399                 case 'c':
 400                         if (argc < i + 2) {
 401                                 show_usage(argc, argv);
 402                                 return -1;
 403                         }
 404                         rduration = atol(argv[++i]);
 405                         break;
 406                 case 'd':
 407                         if (argc < i + 2) {
 408                                 show_usage(argc, argv);
 409                                 return -1;
 410                         }
 411                         wdelay = atol(argv[++i]);
 412                         break;
 413                 case 'v':
 414                         verbose_mode = 1;
 415                         break;
 416                 case 'q':
 417                         test_dequeue = 1;
 418                         break;
 419                 case 's':
 420                         test_splice = 1;
 421                         break;
 422                 case 'M':
 423                         test_sync = TEST_SYNC_MUTEX;
 424                         break;
 425                 case 'w':
 426                         test_wait_empty = 1;
 427                         break;
 428                 case 'f':
 429                         test_force_sync = 1;
 430                         break;
 431                 }
 432         }
 433
 434         /* activate splice test by default */
 435         if (!test_dequeue && !test_splice)
 436                 test_splice = 1;
 437
 438         if (test_sync == TEST_SYNC_NONE && nr_dequeuers > 1 && test_dequeue) {
 439                 if (test_force_sync) {
 440                         fprintf(stderr, "[WARNING] Using dequeue concurrently "
 441                                 "with other dequeue or splice without external "
 442                                 "synchronization. Expect run-time failure.\n");
 443                 } else {
 444                         printf("Enforcing mutex synchronization\n");
 445                         test_sync = TEST_SYNC_MUTEX;
 446                 }
 447         }
 448
 449         printf_verbose("running test for %lu seconds, %u enqueuers, "
 450                        "%u dequeuers.\n",
 451                        duration, nr_enqueuers, nr_dequeuers);
 452         if (test_dequeue)
 453                 printf_verbose("dequeue test activated.\n");
 454         else
 455                 printf_verbose("splice test activated.\n");
 456         if (test_sync == TEST_SYNC_MUTEX)
 457                 printf_verbose("External sync: mutex.\n");
 458         else
 459                 printf_verbose("External sync: none.\n");
 460         if (test_wait_empty)
 461                 printf_verbose("Wait for dequeuers to empty queue.\n");
 462         printf_verbose("Writer delay : %lu loops.\n", rduration);
 463         printf_verbose("Reader duration : %lu loops.\n", wdelay);
 464         printf_verbose("thread %-6s, tid %lu\n",
 465                         "main", urcu_get_thread_id());
 466
 467         tid_enqueuer = calloc(nr_enqueuers, sizeof(*tid_enqueuer));
 468         tid_dequeuer = calloc(nr_dequeuers, sizeof(*tid_dequeuer));
 469         count_enqueuer = calloc(nr_enqueuers, 3 * sizeof(*count_enqueuer));
 470         count_dequeuer = calloc(nr_dequeuers, 4 * sizeof(*count_dequeuer));
 471         cds_wfcq_init(&head, &tail);
 472
 473         next_aff = 0;
 474
 475         for (i = 0; i < nr_enqueuers; i++) {
 476                 err = pthread_create(&tid_enqueuer[i], NULL, thr_enqueuer,
 477                                      &count_enqueuer[3 * i]);
 478                 if (err != 0)
 479                         exit(1);
 480         }
 481         for (i = 0; i < nr_dequeuers; i++) {
 482                 err = pthread_create(&tid_dequeuer[i], NULL, thr_dequeuer,
 483                                      &count_dequeuer[4 * i]);
 484                 if (err != 0)
 485                         exit(1);
 486         }
 487
 488         cmm_smp_mb();
 489
 490         test_go = 1;
 491
 492         for (i = 0; i < duration; i++) {
 493                 sleep(1);
 494                 if (verbose_mode)
 495                         write (1, ".", 1);
 496         }
 497
 498         test_stop_enqueue = 1;
 499
 500         if (test_wait_empty) {
 501                 while (nr_enqueuers != uatomic_read(&test_enqueue_stopped)) {
 502                         sleep(1);
 503                 }
 504                 while (!cds_wfcq_empty(&head, &tail)) {
 505                         sleep(1);
 506                 }
 507         }
 508
 509         test_stop_dequeue = 1;
 510
 511         for (i = 0; i < nr_enqueuers; i++) {
 512                 err = pthread_join(tid_enqueuer[i], &tret);
 513                 if (err != 0)
 514                         exit(1);
 515                 tot_enqueues += count_enqueuer[3 * i];
 516                 tot_successful_enqueues += count_enqueuer[3 * i + 1];
 517                 tot_empty_dest_enqueues += count_enqueuer[3 * i + 2];
 518         }
 519         for (i = 0; i < nr_dequeuers; i++) {
 520                 err = pthread_join(tid_dequeuer[i], &tret);
 521                 if (err != 0)
 522                         exit(1);
 523                 tot_dequeues += count_dequeuer[4 * i];
 524                 tot_successful_dequeues += count_dequeuer[4 * i + 1];
 525                 tot_splice += count_dequeuer[4 * i + 2];
 526                 tot_dequeue_last += count_dequeuer[4 * i + 3];
 527         }
 528
 529         test_end(&end_dequeues, &tot_dequeue_last);
 530
 531         printf_verbose("total number of enqueues : %llu, dequeues %llu\n",
 532                        tot_enqueues, tot_dequeues);
 533         printf_verbose("total number of successful enqueues : %llu, "
 534                        "enqueues to empty dest : %llu, "
 535                        "successful dequeues %llu, "
 536                        "splice : %llu, dequeue_last : %llu\n",
 537                        tot_successful_enqueues,
 538                        tot_empty_dest_enqueues,
 539                        tot_successful_dequeues,
 540                        tot_splice, tot_dequeue_last);
 541         printf("SUMMARY %-25s testdur %4lu nr_enqueuers %3u wdelay %6lu "
 542                 "nr_dequeuers %3u "
 543                 "rdur %6lu nr_enqueues %12llu nr_dequeues %12llu "
 544                 "successful enqueues %12llu enqueues to empty dest %12llu "
 545                 "successful dequeues %12llu splice %12llu "
 546                 "dequeue_last %llu "
 547                 "end_dequeues %llu nr_ops %12llu\n",
 548                 argv[0], duration, nr_enqueuers, wdelay,
 549                 nr_dequeuers, rduration, tot_enqueues, tot_dequeues,
 550                 tot_successful_enqueues,
 551                 tot_empty_dest_enqueues,
 552                 tot_successful_dequeues, tot_splice, tot_dequeue_last,
 553                 end_dequeues,
 554                 tot_enqueues + tot_dequeues);
 555
 556         if (tot_successful_enqueues != tot_successful_dequeues + end_dequeues) {
 557                 printf("WARNING! Discrepancy between nr succ. enqueues %llu vs "
 558                        "succ. dequeues + end dequeues %llu.\n",
 559                        tot_successful_enqueues,
 560                        tot_successful_dequeues + end_dequeues);
 561                 retval = 1;
 562         }
 563
 564         /*
 565          * If only using splice to dequeue, the enqueuer should see
 566          * exactly as many empty queues than the number of non-empty
 567          * src splice.
 568          */
 569         if (tot_empty_dest_enqueues != tot_dequeue_last) {
 570                 printf("WARNING! Discrepancy between empty enqueue (%llu) and "
 571                         "number of dequeue of last element (%llu)\n",
 572                         tot_empty_dest_enqueues,
 573                         tot_dequeue_last);
 574                 retval = 1;
 575         }
 576         free(count_enqueuer);
 577         free(count_dequeuer);
 578         free(tid_enqueuer);
 579         free(tid_dequeuer);
 580         return retval;
 581 }