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 <sched.h>
  38 #include <errno.h>
  39
  40 #include <urcu/arch.h>
  41 #include <urcu/tls-compat.h>
  42
  43 #ifdef __linux__
  44 #include <syscall.h>
  45 #endif
  46
  47 /* hardcoded number of CPUs */
  48 #define NR_CPUS 16384
  49
  50 #if defined(_syscall0)
  51 _syscall0(pid_t, gettid)
  52 #elif defined(__NR_gettid)
  53 static inline pid_t gettid(void)
  54 {
  55         return syscall(__NR_gettid);
  56 }
  57 #else
  58 #warning "use pid as tid"
  59 static inline pid_t gettid(void)
  60 {
  61         return getpid();
  62 }
  63 #endif
  64
  65 #ifndef DYNAMIC_LINK_TEST
  66 #define _LGPL_SOURCE
  67 #endif
  68 #include <urcu.h>
  69 #include <urcu/wfcqueue.h>
  70
  71 enum test_sync {
  72         TEST_SYNC_MUTEX = 0,
  73         TEST_SYNC_NONE,
  74 };
  75
  76 static enum test_sync test_sync;
  77
  78 static volatile int test_go, test_stop_enqueue, test_stop_dequeue;
  79
  80 static unsigned long rduration;
  81
  82 static unsigned long duration;
  83
  84 /* read-side C.S. duration, in loops */
  85 static unsigned long wdelay;
  86
  87 static inline void loop_sleep(unsigned long loops)
  88 {
  89         while (loops-- != 0)
  90                 caa_cpu_relax();
  91 }
  92
  93 static int verbose_mode;
  94
  95 static int test_dequeue, test_splice, test_wait_empty;
  96 static int test_enqueue_stopped;
  97
  98 #define printf_verbose(fmt, args...)            \
  99         do {                                    \
 100                 if (verbose_mode)               \
 101                         printf(fmt, ## args);   \
 102         } while (0)
 103
 104 static unsigned int cpu_affinities[NR_CPUS];
 105 static unsigned int next_aff = 0;
 106 static int use_affinity = 0;
 107
 108 pthread_mutex_t affinity_mutex = PTHREAD_MUTEX_INITIALIZER;
 109
 110 #ifndef HAVE_CPU_SET_T
 111 typedef unsigned long cpu_set_t;
 112 # define CPU_ZERO(cpuset) do { *(cpuset) = 0; } while(0)
 113 # define CPU_SET(cpu, cpuset) do { *(cpuset) |= (1UL << (cpu)); } while(0)
 114 #endif
 115
 116 static void set_affinity(void)
 117 {
 118 #if HAVE_SCHED_SETAFFINITY
 119         cpu_set_t mask;
 120         int cpu, ret;
 121 #endif /* HAVE_SCHED_SETAFFINITY */
 122
 123         if (!use_affinity)
 124                 return;
 125
 126 #if HAVE_SCHED_SETAFFINITY
 127         ret = pthread_mutex_lock(&affinity_mutex);
 128         if (ret) {
 129                 perror("Error in pthread mutex lock");
 130                 exit(-1);
 131         }
 132         cpu = cpu_affinities[next_aff++];
 133         ret = pthread_mutex_unlock(&affinity_mutex);
 134         if (ret) {
 135                 perror("Error in pthread mutex unlock");
 136                 exit(-1);
 137         }
 138
 139         CPU_ZERO(&mask);
 140         CPU_SET(cpu, &mask);
 141 #if SCHED_SETAFFINITY_ARGS == 2
 142         sched_setaffinity(0, &mask);
 143 #else
 144         sched_setaffinity(0, sizeof(mask), &mask);
 145 #endif
 146 #endif /* HAVE_SCHED_SETAFFINITY */
 147 }
 148
 149 /*
 150  * returns 0 if test should end.
 151  */
 152 static int test_duration_dequeue(void)
 153 {
 154         return !test_stop_dequeue;
 155 }
 156
 157 static int test_duration_enqueue(void)
 158 {
 159         return !test_stop_enqueue;
 160 }
 161
 162 static DEFINE_URCU_TLS(unsigned long long, nr_dequeues);
 163 static DEFINE_URCU_TLS(unsigned long long, nr_enqueues);
 164
 165 static DEFINE_URCU_TLS(unsigned long long, nr_successful_dequeues);
 166 static DEFINE_URCU_TLS(unsigned long long, nr_successful_enqueues);
 167 static DEFINE_URCU_TLS(unsigned long long, nr_empty_dest_enqueues);
 168 static DEFINE_URCU_TLS(unsigned long long, nr_splice);
 169
 170 static unsigned int nr_enqueuers;
 171 static unsigned int nr_dequeuers;
 172
 173 static struct cds_wfcq_head __attribute__((aligned(CAA_CACHE_LINE_SIZE))) head;
 174 static struct cds_wfcq_tail __attribute__((aligned(CAA_CACHE_LINE_SIZE))) tail;
 175
 176 static void *thr_enqueuer(void *_count)
 177 {
 178         unsigned long long *count = _count;
 179         bool was_nonempty;
 180
 181         printf_verbose("thread_begin %s, thread id : %lx, tid %lu\n",
 182                         "enqueuer", (unsigned long) pthread_self(),
 183                         (unsigned long) gettid());
 184
 185         set_affinity();
 186
 187         while (!test_go)
 188         {
 189         }
 190         cmm_smp_mb();
 191
 192         for (;;) {
 193                 struct cds_wfcq_node *node = malloc(sizeof(*node));
 194                 if (!node)
 195                         goto fail;
 196                 cds_wfcq_node_init(node);
 197                 was_nonempty = cds_wfcq_enqueue(&head, &tail, node);
 198                 URCU_TLS(nr_successful_enqueues)++;
 199                 if (!was_nonempty)
 200                         URCU_TLS(nr_empty_dest_enqueues)++;
 201
 202                 if (caa_unlikely(wdelay))
 203                         loop_sleep(wdelay);
 204 fail:
 205                 URCU_TLS(nr_enqueues)++;
 206                 if (caa_unlikely(!test_duration_enqueue()))
 207                         break;
 208         }
 209
 210         uatomic_inc(&test_enqueue_stopped);
 211         count[0] = URCU_TLS(nr_enqueues);
 212         count[1] = URCU_TLS(nr_successful_enqueues);
 213         count[2] = URCU_TLS(nr_empty_dest_enqueues);
 214         printf_verbose("enqueuer thread_end, thread id : %lx, tid %lu, "
 215                        "enqueues %llu successful_enqueues %llu, "
 216                        "empty_dest_enqueues %llu\n",
 217                        pthread_self(),
 218                         (unsigned long) gettid(),
 219                        URCU_TLS(nr_enqueues),
 220                        URCU_TLS(nr_successful_enqueues),
 221                        URCU_TLS(nr_empty_dest_enqueues));
 222         return ((void*)1);
 223
 224 }
 225
 226 static void do_test_dequeue(enum test_sync sync)
 227 {
 228         struct cds_wfcq_node *node;
 229
 230         if (sync == TEST_SYNC_MUTEX)
 231                 node = cds_wfcq_dequeue_blocking(&head, &tail);
 232         else
 233                 node = __cds_wfcq_dequeue_blocking(&head, &tail);
 234
 235         if (node) {
 236                 free(node);
 237                 URCU_TLS(nr_successful_dequeues)++;
 238         }
 239         URCU_TLS(nr_dequeues)++;
 240 }
 241
 242 static void do_test_splice(enum test_sync sync)
 243 {
 244         struct cds_wfcq_head tmp_head;
 245         struct cds_wfcq_tail tmp_tail;
 246         struct cds_wfcq_node *node, *n;
 247         enum cds_wfcq_ret ret;
 248
 249         cds_wfcq_init(&tmp_head, &tmp_tail);
 250
 251         if (sync == TEST_SYNC_MUTEX)
 252                 ret = cds_wfcq_splice_blocking(&tmp_head, &tmp_tail,
 253                         &head, &tail);
 254         else
 255                 ret = __cds_wfcq_splice_blocking(&tmp_head, &tmp_tail,
 256                         &head, &tail);
 257
 258         switch (ret) {
 259         case CDS_WFCQ_RET_WOULDBLOCK:
 260                 assert(0);      /* blocking call */
 261                 break;
 262         case CDS_WFCQ_RET_DEST_EMPTY:
 263                 URCU_TLS(nr_splice)++;
 264                 /* ok */
 265                 break;
 266         case CDS_WFCQ_RET_DEST_NON_EMPTY:
 267                 assert(0);      /* entirely unexpected */
 268                 break;
 269         case CDS_WFCQ_RET_SRC_EMPTY:
 270                 /* ok, we could even skip iteration on dest if we wanted */
 271                 break;
 272         }
 273
 274         __cds_wfcq_for_each_blocking_safe(&tmp_head, &tmp_tail, node, n) {
 275                 free(node);
 276                 URCU_TLS(nr_successful_dequeues)++;
 277                 URCU_TLS(nr_dequeues)++;
 278         }
 279 }
 280
 281 static void *thr_dequeuer(void *_count)
 282 {
 283         unsigned long long *count = _count;
 284         unsigned int counter;
 285
 286         printf_verbose("thread_begin %s, thread id : %lx, tid %lu\n",
 287                         "dequeuer", (unsigned long) pthread_self(),
 288                         (unsigned long) gettid());
 289
 290         set_affinity();
 291
 292         while (!test_go)
 293         {
 294         }
 295         cmm_smp_mb();
 296
 297         for (;;) {
 298                 if (test_dequeue && test_splice) {
 299                         if (counter & 1)
 300                                 do_test_dequeue(test_sync);
 301                         else
 302                                 do_test_splice(test_sync);
 303                         counter++;
 304                 } else {
 305                         if (test_dequeue)
 306                                 do_test_dequeue(test_sync);
 307                         else
 308                                 do_test_splice(test_sync);
 309                 }
 310                 if (caa_unlikely(!test_duration_dequeue()))
 311                         break;
 312                 if (caa_unlikely(rduration))
 313                         loop_sleep(rduration);
 314         }
 315
 316         printf_verbose("dequeuer thread_end, thread id : %lx, tid %lu, "
 317                        "dequeues %llu, successful_dequeues %llu, "
 318                        "nr_splice %llu\n",
 319                        pthread_self(),
 320                         (unsigned long) gettid(),
 321                        URCU_TLS(nr_dequeues), URCU_TLS(nr_successful_dequeues),
 322                        URCU_TLS(nr_splice));
 323         count[0] = URCU_TLS(nr_dequeues);
 324         count[1] = URCU_TLS(nr_successful_dequeues);
 325         count[2] = URCU_TLS(nr_splice);
 326         return ((void*)2);
 327 }
 328
 329 static void test_end(unsigned long long *nr_dequeues)
 330 {
 331         struct cds_wfcq_node *node;
 332
 333         do {
 334                 node = cds_wfcq_dequeue_blocking(&head, &tail);
 335                 if (node) {
 336                         free(node);
 337                         (*nr_dequeues)++;
 338                 }
 339         } while (node);
 340 }
 341
 342 static void show_usage(int argc, char **argv)
 343 {
 344         printf("Usage : %s nr_dequeuers nr_enqueuers duration (s)", argv[0]);
 345         printf(" [-d delay] (enqueuer period (in loops))");
 346         printf(" [-c duration] (dequeuer period (in loops))");
 347         printf(" [-v] (verbose output)");
 348         printf(" [-a cpu#] [-a cpu#]... (affinity)");
 349         printf(" [-q] (test dequeue)");
 350         printf(" [-s] (test splice, enabled by default)");
 351         printf(" [-M] (use mutex external synchronization)");
 352         printf(" [-0] (use no external synchronization)");
 353         printf("      Note: default: mutex external synchronization used.");
 354         printf(" [-w] Wait for dequeuer to empty queue");
 355         printf("\n");
 356 }
 357
 358 int main(int argc, char **argv)
 359 {
 360         int err;
 361         pthread_t *tid_enqueuer, *tid_dequeuer;
 362         void *tret;
 363         unsigned long long *count_enqueuer, *count_dequeuer;
 364         unsigned long long tot_enqueues = 0, tot_dequeues = 0;
 365         unsigned long long tot_successful_enqueues = 0,
 366                            tot_successful_dequeues = 0,
 367                            tot_empty_dest_enqueues = 0,
 368                            tot_splice = 0;
 369         unsigned long long end_dequeues = 0;
 370         int i, a, retval = 0;
 371
 372         if (argc < 4) {
 373                 show_usage(argc, argv);
 374                 return -1;
 375         }
 376
 377         err = sscanf(argv[1], "%u", &nr_dequeuers);
 378         if (err != 1) {
 379                 show_usage(argc, argv);
 380                 return -1;
 381         }
 382
 383         err = sscanf(argv[2], "%u", &nr_enqueuers);
 384         if (err != 1) {
 385                 show_usage(argc, argv);
 386                 return -1;
 387         }
 388
 389         err = sscanf(argv[3], "%lu", &duration);
 390         if (err != 1) {
 391                 show_usage(argc, argv);
 392                 return -1;
 393         }
 394
 395         for (i = 4; i < argc; i++) {
 396                 if (argv[i][0] != '-')
 397                         continue;
 398                 switch (argv[i][1]) {
 399                 case 'a':
 400                         if (argc < i + 2) {
 401                                 show_usage(argc, argv);
 402                                 return -1;
 403                         }
 404                         a = atoi(argv[++i]);
 405                         cpu_affinities[next_aff++] = a;
 406                         use_affinity = 1;
 407                         printf_verbose("Adding CPU %d affinity\n", a);
 408                         break;
 409                 case 'c':
 410                         if (argc < i + 2) {
 411                                 show_usage(argc, argv);
 412                                 return -1;
 413                         }
 414                         rduration = atol(argv[++i]);
 415                         break;
 416                 case 'd':
 417                         if (argc < i + 2) {
 418                                 show_usage(argc, argv);
 419                                 return -1;
 420                         }
 421                         wdelay = atol(argv[++i]);
 422                         break;
 423                 case 'v':
 424                         verbose_mode = 1;
 425                         break;
 426                 case 'q':
 427                         test_dequeue = 1;
 428                         break;
 429                 case 's':
 430                         test_splice = 1;
 431                         break;
 432                 case 'M':
 433                         test_sync = TEST_SYNC_MUTEX;
 434                         break;
 435                 case '0':
 436                         test_sync = TEST_SYNC_NONE;
 437                         break;
 438                 case 'w':
 439                         test_wait_empty = 1;
 440                         break;
 441                 }
 442         }
 443
 444         /* activate splice test by default */
 445         if (!test_dequeue && !test_splice)
 446                 test_splice = 1;
 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, thread id : %lx, tid %lu\n",
 464                         "main", (unsigned long) pthread_self(),
 465                         (unsigned long) gettid());
 466
 467         tid_enqueuer = malloc(sizeof(*tid_enqueuer) * nr_enqueuers);
 468         tid_dequeuer = malloc(sizeof(*tid_dequeuer) * nr_dequeuers);
 469         count_enqueuer = malloc(3 * sizeof(*count_enqueuer) * nr_enqueuers);
 470         count_dequeuer = malloc(3 * sizeof(*count_dequeuer) * nr_dequeuers);
 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[3 * 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[3 * i];
 524                 tot_successful_dequeues += count_dequeuer[3 * i + 1];
 525                 tot_splice += count_dequeuer[3 * i + 2];
 526         }
 527
 528         test_end(&end_dequeues);
 529
 530         printf_verbose("total number of enqueues : %llu, dequeues %llu\n",
 531                        tot_enqueues, tot_dequeues);
 532         printf_verbose("total number of successful enqueues : %llu, "
 533                        "enqueues to empty dest : %llu, "
 534                        "successful dequeues %llu, ",
 535                        "splice : %llu\n",
 536                        tot_successful_enqueues,
 537                        tot_empty_dest_enqueues,
 538                        tot_successful_dequeues,
 539                        tot_splice);
 540         printf("SUMMARY %-25s testdur %4lu nr_enqueuers %3u wdelay %6lu "
 541                 "nr_dequeuers %3u "
 542                 "rdur %6lu nr_enqueues %12llu nr_dequeues %12llu "
 543                 "successful enqueues %12llu enqueues to empty dest %12llu "
 544                 "successful dequeues %12llu splice %12llu "
 545                 "end_dequeues %llu nr_ops %12llu\n",
 546                 argv[0], duration, nr_enqueuers, wdelay,
 547                 nr_dequeuers, rduration, tot_enqueues, tot_dequeues,
 548                 tot_successful_enqueues,
 549                 tot_empty_dest_enqueues,
 550                 tot_successful_dequeues, tot_splice, 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 (test_wait_empty && test_splice && !test_dequeue
 567                         && tot_empty_dest_enqueues != tot_splice) {
 568                 printf("WARNING! Discrepancy between empty enqueue (%llu) and "
 569                         "number of non-empty splice (%llu)\n",
 570                         tot_empty_dest_enqueues,
 571                         tot_splice);
 572                 retval = 1;
 573         }
 574         free(count_enqueuer);
 575         free(count_dequeuer);
 576         free(tid_enqueuer);
 577         free(tid_dequeuer);
 578         return retval;
 579 }