tests/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
  44 #ifdef __linux__
  45 #include <syscall.h>
  46 #endif
  47
  48 /* hardcoded number of CPUs */
  49 #define NR_CPUS 16384
  50
  51 #if defined(_syscall0)
  52 _syscall0(pid_t, gettid)
  53 #elif defined(__NR_gettid)
  54 static inline pid_t gettid(void)
  55 {
  56         return syscall(__NR_gettid);
  57 }
  58 #else
  59 #warning "use pid as tid"
  60 static inline pid_t gettid(void)
  61 {
  62         return getpid();
  63 }
  64 #endif
  65
  66 #ifndef DYNAMIC_LINK_TEST
  67 #define _LGPL_SOURCE
  68 #endif
  69 #include <urcu/wfcqueue.h>
  70
  71 enum test_sync {
  72         TEST_SYNC_NONE = 0,
  73         TEST_SYNC_MUTEX,
  74 };
  75
  76 static enum test_sync test_sync;
  77
  78 static int test_force_sync;
  79
  80 static volatile int test_go, test_stop_enqueue, test_stop_dequeue;
  81
  82 static unsigned long rduration;
  83
  84 static unsigned long duration;
  85
  86 /* read-side C.S. duration, in loops */
  87 static unsigned long wdelay;
  88
  89 static inline void loop_sleep(unsigned long loops)
  90 {
  91         while (loops-- != 0)
  92                 caa_cpu_relax();
  93 }
  94
  95 static int verbose_mode;
  96
  97 static int test_dequeue, test_splice, test_wait_empty;
  98 static int test_enqueue_stopped;
  99
 100 #define printf_verbose(fmt, args...)            \
 101         do {                                    \
 102                 if (verbose_mode)               \
 103                         printf(fmt, ## args);   \
 104         } while (0)
 105
 106 static unsigned int cpu_affinities[NR_CPUS];
 107 static unsigned int next_aff = 0;
 108 static int use_affinity = 0;
 109
 110 pthread_mutex_t affinity_mutex = PTHREAD_MUTEX_INITIALIZER;
 111
 112 static void set_affinity(void)
 113 {
 114 #if HAVE_SCHED_SETAFFINITY
 115         cpu_set_t mask;
 116         int cpu, ret;
 117 #endif /* HAVE_SCHED_SETAFFINITY */
 118
 119         if (!use_affinity)
 120                 return;
 121
 122 #if HAVE_SCHED_SETAFFINITY
 123         ret = pthread_mutex_lock(&affinity_mutex);
 124         if (ret) {
 125                 perror("Error in pthread mutex lock");
 126                 exit(-1);
 127         }
 128         cpu = cpu_affinities[next_aff++];
 129         ret = pthread_mutex_unlock(&affinity_mutex);
 130         if (ret) {
 131                 perror("Error in pthread mutex unlock");
 132                 exit(-1);
 133         }
 134
 135         CPU_ZERO(&mask);
 136         CPU_SET(cpu, &mask);
 137 #if SCHED_SETAFFINITY_ARGS == 2
 138         sched_setaffinity(0, &mask);
 139 #else
 140         sched_setaffinity(0, sizeof(mask), &mask);
 141 #endif
 142 #endif /* HAVE_SCHED_SETAFFINITY */
 143 }
 144
 145 /*
 146  * returns 0 if test should end.
 147  */
 148 static int test_duration_dequeue(void)
 149 {
 150         return !test_stop_dequeue;
 151 }
 152
 153 static int test_duration_enqueue(void)
 154 {
 155         return !test_stop_enqueue;
 156 }
 157
 158 static DEFINE_URCU_TLS(unsigned long long, nr_dequeues);
 159 static DEFINE_URCU_TLS(unsigned long long, nr_enqueues);
 160
 161 static DEFINE_URCU_TLS(unsigned long long, nr_successful_dequeues);
 162 static DEFINE_URCU_TLS(unsigned long long, nr_successful_enqueues);
 163 static DEFINE_URCU_TLS(unsigned long long, nr_empty_dest_enqueues);
 164 static DEFINE_URCU_TLS(unsigned long long, nr_splice);
 165 static DEFINE_URCU_TLS(unsigned long long, nr_dequeue_last);
 166
 167 static unsigned int nr_enqueuers;
 168 static unsigned int nr_dequeuers;
 169
 170 static struct cds_wfcq_head __attribute__((aligned(CAA_CACHE_LINE_SIZE))) head;
 171 static struct cds_wfcq_tail __attribute__((aligned(CAA_CACHE_LINE_SIZE))) tail;
 172
 173 static void *thr_enqueuer(void *_count)
 174 {
 175         unsigned long long *count = _count;
 176         bool was_nonempty;
 177
 178         printf_verbose("thread_begin %s, thread id : %lx, tid %lu\n",
 179                         "enqueuer", (unsigned long) pthread_self(),
 180                         (unsigned long) gettid());
 181
 182         set_affinity();
 183
 184         while (!test_go)
 185         {
 186         }
 187         cmm_smp_mb();
 188
 189         for (;;) {
 190                 struct cds_wfcq_node *node = malloc(sizeof(*node));
 191                 if (!node)
 192                         goto fail;
 193                 cds_wfcq_node_init(node);
 194                 was_nonempty = cds_wfcq_enqueue(&head, &tail, node);
 195                 URCU_TLS(nr_successful_enqueues)++;
 196                 if (!was_nonempty)
 197                         URCU_TLS(nr_empty_dest_enqueues)++;
 198
 199                 if (caa_unlikely(wdelay))
 200                         loop_sleep(wdelay);
 201 fail:
 202                 URCU_TLS(nr_enqueues)++;
 203                 if (caa_unlikely(!test_duration_enqueue()))
 204                         break;
 205         }
 206
 207         uatomic_inc(&test_enqueue_stopped);
 208         count[0] = URCU_TLS(nr_enqueues);
 209         count[1] = URCU_TLS(nr_successful_enqueues);
 210         count[2] = URCU_TLS(nr_empty_dest_enqueues);
 211         printf_verbose("enqueuer thread_end, thread id : %lx, tid %lu, "
 212                        "enqueues %llu successful_enqueues %llu, "
 213                        "empty_dest_enqueues %llu\n",
 214                        pthread_self(),
 215                         (unsigned long) gettid(),
 216                        URCU_TLS(nr_enqueues),
 217                        URCU_TLS(nr_successful_enqueues),
 218                        URCU_TLS(nr_empty_dest_enqueues));
 219         return ((void*)1);
 220
 221 }
 222
 223 static void do_test_dequeue(enum test_sync sync)
 224 {
 225         struct cds_wfcq_node *node;
 226         int state;
 227
 228         if (sync == TEST_SYNC_MUTEX)
 229                 node = cds_wfcq_dequeue_with_state_blocking(&head, &tail,
 230                                 &state);
 231         else
 232                 node = __cds_wfcq_dequeue_with_state_blocking(&head, &tail,
 233                                 &state);
 234
 235         if (state & CDS_WFCQ_STATE_LAST)
 236                 URCU_TLS(nr_dequeue_last)++;
 237
 238         if (node) {
 239                 free(node);
 240                 URCU_TLS(nr_successful_dequeues)++;
 241         }
 242         URCU_TLS(nr_dequeues)++;
 243 }
 244
 245 static void do_test_splice(enum test_sync sync)
 246 {
 247         struct cds_wfcq_head tmp_head;
 248         struct cds_wfcq_tail tmp_tail;
 249         struct cds_wfcq_node *node, *n;
 250         enum cds_wfcq_ret ret;
 251
 252         cds_wfcq_init(&tmp_head, &tmp_tail);
 253
 254         if (sync == TEST_SYNC_MUTEX)
 255                 ret = cds_wfcq_splice_blocking(&tmp_head, &tmp_tail,
 256                         &head, &tail);
 257         else
 258                 ret = __cds_wfcq_splice_blocking(&tmp_head, &tmp_tail,
 259                         &head, &tail);
 260
 261         switch (ret) {
 262         case CDS_WFCQ_RET_WOULDBLOCK:
 263                 assert(0);      /* blocking call */
 264                 break;
 265         case CDS_WFCQ_RET_DEST_EMPTY:
 266                 URCU_TLS(nr_splice)++;
 267                 URCU_TLS(nr_dequeue_last)++;
 268                 /* ok */
 269                 break;
 270         case CDS_WFCQ_RET_DEST_NON_EMPTY:
 271                 assert(0);      /* entirely unexpected */
 272                 break;
 273         case CDS_WFCQ_RET_SRC_EMPTY:
 274                 /* ok, we could even skip iteration on dest if we wanted */
 275                 break;
 276         }
 277
 278         __cds_wfcq_for_each_blocking_safe(&tmp_head, &tmp_tail, node, n) {
 279                 free(node);
 280                 URCU_TLS(nr_successful_dequeues)++;
 281                 URCU_TLS(nr_dequeues)++;
 282         }
 283 }
 284
 285 static void *thr_dequeuer(void *_count)
 286 {
 287         unsigned long long *count = _count;
 288         unsigned int counter;
 289
 290         printf_verbose("thread_begin %s, thread id : %lx, tid %lu\n",
 291                         "dequeuer", (unsigned long) pthread_self(),
 292                         (unsigned long) gettid());
 293
 294         set_affinity();
 295
 296         while (!test_go)
 297         {
 298         }
 299         cmm_smp_mb();
 300
 301         for (;;) {
 302                 if (test_dequeue && test_splice) {
 303                         if (counter & 1)
 304                                 do_test_dequeue(test_sync);
 305                         else
 306                                 do_test_splice(test_sync);
 307                         counter++;
 308                 } else {
 309                         if (test_dequeue)
 310                                 do_test_dequeue(test_sync);
 311                         else
 312                                 do_test_splice(test_sync);
 313                 }
 314                 if (caa_unlikely(!test_duration_dequeue()))
 315                         break;
 316                 if (caa_unlikely(rduration))
 317                         loop_sleep(rduration);
 318         }
 319
 320         printf_verbose("dequeuer thread_end, thread id : %lx, tid %lu, "
 321                        "dequeues %llu, successful_dequeues %llu, "
 322                        "nr_splice %llu\n",
 323                        pthread_self(),
 324                         (unsigned long) gettid(),
 325                        URCU_TLS(nr_dequeues), URCU_TLS(nr_successful_dequeues),
 326                        URCU_TLS(nr_splice));
 327         count[0] = URCU_TLS(nr_dequeues);
 328         count[1] = URCU_TLS(nr_successful_dequeues);
 329         count[2] = URCU_TLS(nr_splice);
 330         count[3] = URCU_TLS(nr_dequeue_last);
 331         return ((void*)2);
 332 }
 333
 334 static void test_end(unsigned long long *nr_dequeues,
 335                 unsigned long long *nr_dequeue_last)
 336 {
 337         struct cds_wfcq_node *node;
 338         int state;
 339
 340         do {
 341                 node = cds_wfcq_dequeue_with_state_blocking(&head, &tail,
 342                                 &state);
 343                 if (node) {
 344                         if (state & CDS_WFCQ_STATE_LAST)
 345                                 (*nr_dequeue_last)++;
 346                         free(node);
 347                         (*nr_dequeues)++;
 348                 }
 349         } while (node);
 350 }
 351
 352 static void show_usage(int argc, char **argv)
 353 {
 354         printf("Usage : %s nr_dequeuers nr_enqueuers duration (s)", argv[0]);
 355         printf(" [-d delay] (enqueuer period (in loops))");
 356         printf(" [-c duration] (dequeuer period (in loops))");
 357         printf(" [-v] (verbose output)");
 358         printf(" [-a cpu#] [-a cpu#]... (affinity)");
 359         printf(" [-q] (test dequeue)");
 360         printf(" [-s] (test splice, enabled by default)");
 361         printf(" [-M] (use mutex external synchronization)");
 362         printf("      Note: default: no external synchronization used.");
 363         printf(" [-f] (force user-provided synchronization)");
 364         printf(" [-w] Wait for dequeuer to empty queue");
 365         printf("\n");
 366 }
 367
 368 int main(int argc, char **argv)
 369 {
 370         int err;
 371         pthread_t *tid_enqueuer, *tid_dequeuer;
 372         void *tret;
 373         unsigned long long *count_enqueuer, *count_dequeuer;
 374         unsigned long long tot_enqueues = 0, tot_dequeues = 0;
 375         unsigned long long tot_successful_enqueues = 0,
 376                            tot_successful_dequeues = 0,
 377                            tot_empty_dest_enqueues = 0,
 378                            tot_splice = 0, tot_dequeue_last = 0;
 379         unsigned long long end_dequeues = 0;
 380         int i, a, retval = 0;
 381
 382         if (argc < 4) {
 383                 show_usage(argc, argv);
 384                 return -1;
 385         }
 386
 387         err = sscanf(argv[1], "%u", &nr_dequeuers);
 388         if (err != 1) {
 389                 show_usage(argc, argv);
 390                 return -1;
 391         }
 392
 393         err = sscanf(argv[2], "%u", &nr_enqueuers);
 394         if (err != 1) {
 395                 show_usage(argc, argv);
 396                 return -1;
 397         }
 398
 399         err = sscanf(argv[3], "%lu", &duration);
 400         if (err != 1) {
 401                 show_usage(argc, argv);
 402                 return -1;
 403         }
 404
 405         for (i = 4; i < argc; i++) {
 406                 if (argv[i][0] != '-')
 407                         continue;
 408                 switch (argv[i][1]) {
 409                 case 'a':
 410                         if (argc < i + 2) {
 411                                 show_usage(argc, argv);
 412                                 return -1;
 413                         }
 414                         a = atoi(argv[++i]);
 415                         cpu_affinities[next_aff++] = a;
 416                         use_affinity = 1;
 417                         printf_verbose("Adding CPU %d affinity\n", a);
 418                         break;
 419                 case 'c':
 420                         if (argc < i + 2) {
 421                                 show_usage(argc, argv);
 422                                 return -1;
 423                         }
 424                         rduration = atol(argv[++i]);
 425                         break;
 426                 case 'd':
 427                         if (argc < i + 2) {
 428                                 show_usage(argc, argv);
 429                                 return -1;
 430                         }
 431                         wdelay = atol(argv[++i]);
 432                         break;
 433                 case 'v':
 434                         verbose_mode = 1;
 435                         break;
 436                 case 'q':
 437                         test_dequeue = 1;
 438                         break;
 439                 case 's':
 440                         test_splice = 1;
 441                         break;
 442                 case 'M':
 443                         test_sync = TEST_SYNC_MUTEX;
 444                         break;
 445                 case 'w':
 446                         test_wait_empty = 1;
 447                         break;
 448                 case 'f':
 449                         test_force_sync = 1;
 450                         break;
 451                 }
 452         }
 453
 454         /* activate splice test by default */
 455         if (!test_dequeue && !test_splice)
 456                 test_splice = 1;
 457
 458         if (test_sync == TEST_SYNC_NONE && nr_dequeuers > 1 && test_dequeue) {
 459                 if (test_force_sync) {
 460                         fprintf(stderr, "[WARNING] Using dequeue concurrently "
 461                                 "with other dequeue or splice without external "
 462                                 "synchronization. Expect run-time failure.\n");
 463                 } else {
 464                         printf("Enforcing mutex synchronization\n");
 465                         test_sync = TEST_SYNC_MUTEX;
 466                 }
 467         }
 468
 469         printf_verbose("running test for %lu seconds, %u enqueuers, "
 470                        "%u dequeuers.\n",
 471                        duration, nr_enqueuers, nr_dequeuers);
 472         if (test_dequeue)
 473                 printf_verbose("dequeue test activated.\n");
 474         else
 475                 printf_verbose("splice test activated.\n");
 476         if (test_sync == TEST_SYNC_MUTEX)
 477                 printf_verbose("External sync: mutex.\n");
 478         else
 479                 printf_verbose("External sync: none.\n");
 480         if (test_wait_empty)
 481                 printf_verbose("Wait for dequeuers to empty queue.\n");
 482         printf_verbose("Writer delay : %lu loops.\n", rduration);
 483         printf_verbose("Reader duration : %lu loops.\n", wdelay);
 484         printf_verbose("thread %-6s, thread id : %lx, tid %lu\n",
 485                         "main", (unsigned long) pthread_self(),
 486                         (unsigned long) gettid());
 487
 488         tid_enqueuer = malloc(sizeof(*tid_enqueuer) * nr_enqueuers);
 489         tid_dequeuer = malloc(sizeof(*tid_dequeuer) * nr_dequeuers);
 490         count_enqueuer = malloc(3 * sizeof(*count_enqueuer) * nr_enqueuers);
 491         count_dequeuer = malloc(4 * sizeof(*count_dequeuer) * nr_dequeuers);
 492         cds_wfcq_init(&head, &tail);
 493
 494         next_aff = 0;
 495
 496         for (i = 0; i < nr_enqueuers; i++) {
 497                 err = pthread_create(&tid_enqueuer[i], NULL, thr_enqueuer,
 498                                      &count_enqueuer[3 * i]);
 499                 if (err != 0)
 500                         exit(1);
 501         }
 502         for (i = 0; i < nr_dequeuers; i++) {
 503                 err = pthread_create(&tid_dequeuer[i], NULL, thr_dequeuer,
 504                                      &count_dequeuer[4 * i]);
 505                 if (err != 0)
 506                         exit(1);
 507         }
 508
 509         cmm_smp_mb();
 510
 511         test_go = 1;
 512
 513         for (i = 0; i < duration; i++) {
 514                 sleep(1);
 515                 if (verbose_mode)
 516                         write (1, ".", 1);
 517         }
 518
 519         test_stop_enqueue = 1;
 520
 521         if (test_wait_empty) {
 522                 while (nr_enqueuers != uatomic_read(&test_enqueue_stopped)) {
 523                         sleep(1);
 524                 }
 525                 while (!cds_wfcq_empty(&head, &tail)) {
 526                         sleep(1);
 527                 }
 528         }
 529
 530         test_stop_dequeue = 1;
 531
 532         for (i = 0; i < nr_enqueuers; i++) {
 533                 err = pthread_join(tid_enqueuer[i], &tret);
 534                 if (err != 0)
 535                         exit(1);
 536                 tot_enqueues += count_enqueuer[3 * i];
 537                 tot_successful_enqueues += count_enqueuer[3 * i + 1];
 538                 tot_empty_dest_enqueues += count_enqueuer[3 * i + 2];
 539         }
 540         for (i = 0; i < nr_dequeuers; i++) {
 541                 err = pthread_join(tid_dequeuer[i], &tret);
 542                 if (err != 0)
 543                         exit(1);
 544                 tot_dequeues += count_dequeuer[4 * i];
 545                 tot_successful_dequeues += count_dequeuer[4 * i + 1];
 546                 tot_splice += count_dequeuer[4 * i + 2];
 547                 tot_dequeue_last += count_dequeuer[4 * i + 3];
 548         }
 549
 550         test_end(&end_dequeues, &tot_dequeue_last);
 551
 552         printf_verbose("total number of enqueues : %llu, dequeues %llu\n",
 553                        tot_enqueues, tot_dequeues);
 554         printf_verbose("total number of successful enqueues : %llu, "
 555                        "enqueues to empty dest : %llu, "
 556                        "successful dequeues %llu, "
 557                        "splice : %llu, dequeue_last : %llu\n",
 558                        tot_successful_enqueues,
 559                        tot_empty_dest_enqueues,
 560                        tot_successful_dequeues,
 561                        tot_splice, tot_dequeue_last);
 562         printf("SUMMARY %-25s testdur %4lu nr_enqueuers %3u wdelay %6lu "
 563                 "nr_dequeuers %3u "
 564                 "rdur %6lu nr_enqueues %12llu nr_dequeues %12llu "
 565                 "successful enqueues %12llu enqueues to empty dest %12llu "
 566                 "successful dequeues %12llu splice %12llu "
 567                 "dequeue_last %llu "
 568                 "end_dequeues %llu nr_ops %12llu\n",
 569                 argv[0], duration, nr_enqueuers, wdelay,
 570                 nr_dequeuers, rduration, tot_enqueues, tot_dequeues,
 571                 tot_successful_enqueues,
 572                 tot_empty_dest_enqueues,
 573                 tot_successful_dequeues, tot_splice, tot_dequeue_last,
 574                 end_dequeues,
 575                 tot_enqueues + tot_dequeues);
 576
 577         if (tot_successful_enqueues != tot_successful_dequeues + end_dequeues) {
 578                 printf("WARNING! Discrepancy between nr succ. enqueues %llu vs "
 579                        "succ. dequeues + end dequeues %llu.\n",
 580                        tot_successful_enqueues,
 581                        tot_successful_dequeues + end_dequeues);
 582                 retval = 1;
 583         }
 584
 585         /*
 586          * If only using splice to dequeue, the enqueuer should see
 587          * exactly as many empty queues than the number of non-empty
 588          * src splice.
 589          */
 590         if (tot_empty_dest_enqueues != tot_dequeue_last) {
 591                 printf("WARNING! Discrepancy between empty enqueue (%llu) and "
 592                         "number of dequeue of last element (%llu)\n",
 593                         tot_empty_dest_enqueues,
 594                         tot_dequeue_last);
 595                 retval = 1;
 596         }
 597         free(count_enqueuer);
 598         free(count_dequeuer);
 599         free(tid_enqueuer);
 600         free(tid_dequeuer);
 601         return retval;
 602 }