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
 166 static unsigned int nr_enqueuers;
 167 static unsigned int nr_dequeuers;
 168
 169 static struct cds_wfcq_head __attribute__((aligned(CAA_CACHE_LINE_SIZE))) head;
 170 static struct cds_wfcq_tail __attribute__((aligned(CAA_CACHE_LINE_SIZE))) tail;
 171
 172 static void *thr_enqueuer(void *_count)
 173 {
 174         unsigned long long *count = _count;
 175         bool was_nonempty;
 176
 177         printf_verbose("thread_begin %s, thread id : %lx, tid %lu\n",
 178                         "enqueuer", (unsigned long) pthread_self(),
 179                         (unsigned long) gettid());
 180
 181         set_affinity();
 182
 183         while (!test_go)
 184         {
 185         }
 186         cmm_smp_mb();
 187
 188         for (;;) {
 189                 struct cds_wfcq_node *node = malloc(sizeof(*node));
 190                 if (!node)
 191                         goto fail;
 192                 cds_wfcq_node_init(node);
 193                 was_nonempty = cds_wfcq_enqueue(&head, &tail, node);
 194                 URCU_TLS(nr_successful_enqueues)++;
 195                 if (!was_nonempty)
 196                         URCU_TLS(nr_empty_dest_enqueues)++;
 197
 198                 if (caa_unlikely(wdelay))
 199                         loop_sleep(wdelay);
 200 fail:
 201                 URCU_TLS(nr_enqueues)++;
 202                 if (caa_unlikely(!test_duration_enqueue()))
 203                         break;
 204         }
 205
 206         uatomic_inc(&test_enqueue_stopped);
 207         count[0] = URCU_TLS(nr_enqueues);
 208         count[1] = URCU_TLS(nr_successful_enqueues);
 209         count[2] = URCU_TLS(nr_empty_dest_enqueues);
 210         printf_verbose("enqueuer thread_end, thread id : %lx, tid %lu, "
 211                        "enqueues %llu successful_enqueues %llu, "
 212                        "empty_dest_enqueues %llu\n",
 213                        pthread_self(),
 214                         (unsigned long) gettid(),
 215                        URCU_TLS(nr_enqueues),
 216                        URCU_TLS(nr_successful_enqueues),
 217                        URCU_TLS(nr_empty_dest_enqueues));
 218         return ((void*)1);
 219
 220 }
 221
 222 static void do_test_dequeue(enum test_sync sync)
 223 {
 224         struct cds_wfcq_node *node;
 225
 226         if (sync == TEST_SYNC_MUTEX)
 227                 node = cds_wfcq_dequeue_blocking(&head, &tail);
 228         else
 229                 node = __cds_wfcq_dequeue_blocking(&head, &tail);
 230
 231         if (node) {
 232                 free(node);
 233                 URCU_TLS(nr_successful_dequeues)++;
 234         }
 235         URCU_TLS(nr_dequeues)++;
 236 }
 237
 238 static void do_test_splice(enum test_sync sync)
 239 {
 240         struct cds_wfcq_head tmp_head;
 241         struct cds_wfcq_tail tmp_tail;
 242         struct cds_wfcq_node *node, *n;
 243         enum cds_wfcq_ret ret;
 244
 245         cds_wfcq_init(&tmp_head, &tmp_tail);
 246
 247         if (sync == TEST_SYNC_MUTEX)
 248                 ret = cds_wfcq_splice_blocking(&tmp_head, &tmp_tail,
 249                         &head, &tail);
 250         else
 251                 ret = __cds_wfcq_splice_blocking(&tmp_head, &tmp_tail,
 252                         &head, &tail);
 253
 254         switch (ret) {
 255         case CDS_WFCQ_RET_WOULDBLOCK:
 256                 assert(0);      /* blocking call */
 257                 break;
 258         case CDS_WFCQ_RET_DEST_EMPTY:
 259                 URCU_TLS(nr_splice)++;
 260                 /* ok */
 261                 break;
 262         case CDS_WFCQ_RET_DEST_NON_EMPTY:
 263                 assert(0);      /* entirely unexpected */
 264                 break;
 265         case CDS_WFCQ_RET_SRC_EMPTY:
 266                 /* ok, we could even skip iteration on dest if we wanted */
 267                 break;
 268         }
 269
 270         __cds_wfcq_for_each_blocking_safe(&tmp_head, &tmp_tail, node, n) {
 271                 free(node);
 272                 URCU_TLS(nr_successful_dequeues)++;
 273                 URCU_TLS(nr_dequeues)++;
 274         }
 275 }
 276
 277 static void *thr_dequeuer(void *_count)
 278 {
 279         unsigned long long *count = _count;
 280         unsigned int counter;
 281
 282         printf_verbose("thread_begin %s, thread id : %lx, tid %lu\n",
 283                         "dequeuer", (unsigned long) pthread_self(),
 284                         (unsigned long) gettid());
 285
 286         set_affinity();
 287
 288         while (!test_go)
 289         {
 290         }
 291         cmm_smp_mb();
 292
 293         for (;;) {
 294                 if (test_dequeue && test_splice) {
 295                         if (counter & 1)
 296                                 do_test_dequeue(test_sync);
 297                         else
 298                                 do_test_splice(test_sync);
 299                         counter++;
 300                 } else {
 301                         if (test_dequeue)
 302                                 do_test_dequeue(test_sync);
 303                         else
 304                                 do_test_splice(test_sync);
 305                 }
 306                 if (caa_unlikely(!test_duration_dequeue()))
 307                         break;
 308                 if (caa_unlikely(rduration))
 309                         loop_sleep(rduration);
 310         }
 311
 312         printf_verbose("dequeuer thread_end, thread id : %lx, tid %lu, "
 313                        "dequeues %llu, successful_dequeues %llu, "
 314                        "nr_splice %llu\n",
 315                        pthread_self(),
 316                         (unsigned long) gettid(),
 317                        URCU_TLS(nr_dequeues), URCU_TLS(nr_successful_dequeues),
 318                        URCU_TLS(nr_splice));
 319         count[0] = URCU_TLS(nr_dequeues);
 320         count[1] = URCU_TLS(nr_successful_dequeues);
 321         count[2] = URCU_TLS(nr_splice);
 322         return ((void*)2);
 323 }
 324
 325 static void test_end(unsigned long long *nr_dequeues)
 326 {
 327         struct cds_wfcq_node *node;
 328
 329         do {
 330                 node = cds_wfcq_dequeue_blocking(&head, &tail);
 331                 if (node) {
 332                         free(node);
 333                         (*nr_dequeues)++;
 334                 }
 335         } while (node);
 336 }
 337
 338 static void show_usage(int argc, char **argv)
 339 {
 340         printf("Usage : %s nr_dequeuers nr_enqueuers duration (s)", argv[0]);
 341         printf(" [-d delay] (enqueuer period (in loops))");
 342         printf(" [-c duration] (dequeuer period (in loops))");
 343         printf(" [-v] (verbose output)");
 344         printf(" [-a cpu#] [-a cpu#]... (affinity)");
 345         printf(" [-q] (test dequeue)");
 346         printf(" [-s] (test splice, enabled by default)");
 347         printf(" [-M] (use mutex external synchronization)");
 348         printf("      Note: default: no external synchronization used.");
 349         printf(" [-f] (force user-provided synchronization)");
 350         printf(" [-w] Wait for dequeuer to empty queue");
 351         printf("\n");
 352 }
 353
 354 int main(int argc, char **argv)
 355 {
 356         int err;
 357         pthread_t *tid_enqueuer, *tid_dequeuer;
 358         void *tret;
 359         unsigned long long *count_enqueuer, *count_dequeuer;
 360         unsigned long long tot_enqueues = 0, tot_dequeues = 0;
 361         unsigned long long tot_successful_enqueues = 0,
 362                            tot_successful_dequeues = 0,
 363                            tot_empty_dest_enqueues = 0,
 364                            tot_splice = 0;
 365         unsigned long long end_dequeues = 0;
 366         int i, a, retval = 0;
 367
 368         if (argc < 4) {
 369                 show_usage(argc, argv);
 370                 return -1;
 371         }
 372
 373         err = sscanf(argv[1], "%u", &nr_dequeuers);
 374         if (err != 1) {
 375                 show_usage(argc, argv);
 376                 return -1;
 377         }
 378
 379         err = sscanf(argv[2], "%u", &nr_enqueuers);
 380         if (err != 1) {
 381                 show_usage(argc, argv);
 382                 return -1;
 383         }
 384
 385         err = sscanf(argv[3], "%lu", &duration);
 386         if (err != 1) {
 387                 show_usage(argc, argv);
 388                 return -1;
 389         }
 390
 391         for (i = 4; i < argc; i++) {
 392                 if (argv[i][0] != '-')
 393                         continue;
 394                 switch (argv[i][1]) {
 395                 case 'a':
 396                         if (argc < i + 2) {
 397                                 show_usage(argc, argv);
 398                                 return -1;
 399                         }
 400                         a = atoi(argv[++i]);
 401                         cpu_affinities[next_aff++] = a;
 402                         use_affinity = 1;
 403                         printf_verbose("Adding CPU %d affinity\n", a);
 404                         break;
 405                 case 'c':
 406                         if (argc < i + 2) {
 407                                 show_usage(argc, argv);
 408                                 return -1;
 409                         }
 410                         rduration = atol(argv[++i]);
 411                         break;
 412                 case 'd':
 413                         if (argc < i + 2) {
 414                                 show_usage(argc, argv);
 415                                 return -1;
 416                         }
 417                         wdelay = atol(argv[++i]);
 418                         break;
 419                 case 'v':
 420                         verbose_mode = 1;
 421                         break;
 422                 case 'q':
 423                         test_dequeue = 1;
 424                         break;
 425                 case 's':
 426                         test_splice = 1;
 427                         break;
 428                 case 'M':
 429                         test_sync = TEST_SYNC_MUTEX;
 430                         break;
 431                 case 'w':
 432                         test_wait_empty = 1;
 433                         break;
 434                 case 'f':
 435                         test_force_sync = 1;
 436                         break;
 437                 }
 438         }
 439
 440         /* activate splice test by default */
 441         if (!test_dequeue && !test_splice)
 442                 test_splice = 1;
 443
 444         if (test_sync == TEST_SYNC_NONE && nr_dequeuers > 1 && test_dequeue) {
 445                 if (test_force_sync) {
 446                         fprintf(stderr, "[WARNING] Using dequeue concurrently "
 447                                 "with other dequeue or splice without external "
 448                                 "synchronization. Expect run-time failure.\n");
 449                 } else {
 450                         printf("Enforcing mutex synchronization\n");
 451                         test_sync = TEST_SYNC_MUTEX;
 452                 }
 453         }
 454
 455         printf_verbose("running test for %lu seconds, %u enqueuers, "
 456                        "%u dequeuers.\n",
 457                        duration, nr_enqueuers, nr_dequeuers);
 458         if (test_dequeue)
 459                 printf_verbose("dequeue test activated.\n");
 460         else
 461                 printf_verbose("splice test activated.\n");
 462         if (test_sync == TEST_SYNC_MUTEX)
 463                 printf_verbose("External sync: mutex.\n");
 464         else
 465                 printf_verbose("External sync: none.\n");
 466         if (test_wait_empty)
 467                 printf_verbose("Wait for dequeuers to empty queue.\n");
 468         printf_verbose("Writer delay : %lu loops.\n", rduration);
 469         printf_verbose("Reader duration : %lu loops.\n", wdelay);
 470         printf_verbose("thread %-6s, thread id : %lx, tid %lu\n",
 471                         "main", (unsigned long) pthread_self(),
 472                         (unsigned long) gettid());
 473
 474         tid_enqueuer = calloc(nr_enqueuers, sizeof(*tid_enqueuer));
 475         tid_dequeuer = calloc(nr_dequeuers, sizeof(*tid_dequeuer));
 476         count_enqueuer = calloc(nr_enqueuers, 3 * sizeof(*count_enqueuer));
 477         count_dequeuer = calloc(nr_dequeuers, 3 * sizeof(*count_dequeuer));
 478
 479         cds_wfcq_init(&head, &tail);
 480
 481         next_aff = 0;
 482
 483         for (i = 0; i < nr_enqueuers; i++) {
 484                 err = pthread_create(&tid_enqueuer[i], NULL, thr_enqueuer,
 485                                      &count_enqueuer[3 * i]);
 486                 if (err != 0)
 487                         exit(1);
 488         }
 489         for (i = 0; i < nr_dequeuers; i++) {
 490                 err = pthread_create(&tid_dequeuer[i], NULL, thr_dequeuer,
 491                                      &count_dequeuer[3 * i]);
 492                 if (err != 0)
 493                         exit(1);
 494         }
 495
 496         cmm_smp_mb();
 497
 498         test_go = 1;
 499
 500         for (i = 0; i < duration; i++) {
 501                 sleep(1);
 502                 if (verbose_mode)
 503                         write (1, ".", 1);
 504         }
 505
 506         test_stop_enqueue = 1;
 507
 508         if (test_wait_empty) {
 509                 while (nr_enqueuers != uatomic_read(&test_enqueue_stopped)) {
 510                         sleep(1);
 511                 }
 512                 while (!cds_wfcq_empty(&head, &tail)) {
 513                         sleep(1);
 514                 }
 515         }
 516
 517         test_stop_dequeue = 1;
 518
 519         for (i = 0; i < nr_enqueuers; i++) {
 520                 err = pthread_join(tid_enqueuer[i], &tret);
 521                 if (err != 0)
 522                         exit(1);
 523                 tot_enqueues += count_enqueuer[3 * i];
 524                 tot_successful_enqueues += count_enqueuer[3 * i + 1];
 525                 tot_empty_dest_enqueues += count_enqueuer[3 * i + 2];
 526         }
 527         for (i = 0; i < nr_dequeuers; i++) {
 528                 err = pthread_join(tid_dequeuer[i], &tret);
 529                 if (err != 0)
 530                         exit(1);
 531                 tot_dequeues += count_dequeuer[3 * i];
 532                 tot_successful_dequeues += count_dequeuer[3 * i + 1];
 533                 tot_splice += count_dequeuer[3 * i + 2];
 534         }
 535
 536         test_end(&end_dequeues);
 537
 538         printf_verbose("total number of enqueues : %llu, dequeues %llu\n",
 539                        tot_enqueues, tot_dequeues);
 540         printf_verbose("total number of successful enqueues : %llu, "
 541                        "enqueues to empty dest : %llu, "
 542                        "successful dequeues %llu, "
 543                        "splice : %llu\n",
 544                        tot_successful_enqueues,
 545                        tot_empty_dest_enqueues,
 546                        tot_successful_dequeues,
 547                        tot_splice);
 548         printf("SUMMARY %-25s testdur %4lu nr_enqueuers %3u wdelay %6lu "
 549                 "nr_dequeuers %3u "
 550                 "rdur %6lu nr_enqueues %12llu nr_dequeues %12llu "
 551                 "successful enqueues %12llu enqueues to empty dest %12llu "
 552                 "successful dequeues %12llu splice %12llu "
 553                 "end_dequeues %llu nr_ops %12llu\n",
 554                 argv[0], duration, nr_enqueuers, wdelay,
 555                 nr_dequeuers, rduration, tot_enqueues, tot_dequeues,
 556                 tot_successful_enqueues,
 557                 tot_empty_dest_enqueues,
 558                 tot_successful_dequeues, tot_splice, end_dequeues,
 559                 tot_enqueues + tot_dequeues);
 560
 561         if (tot_successful_enqueues != tot_successful_dequeues + end_dequeues) {
 562                 printf("WARNING! Discrepancy between nr succ. enqueues %llu vs "
 563                        "succ. dequeues + end dequeues %llu.\n",
 564                        tot_successful_enqueues,
 565                        tot_successful_dequeues + end_dequeues);
 566                 retval = 1;
 567         }
 568
 569         /*
 570          * If only using splice to dequeue, the enqueuer should see
 571          * exactly as many empty queues than the number of non-empty
 572          * src splice.
 573          */
 574         if (test_wait_empty && test_splice && !test_dequeue
 575                         && tot_empty_dest_enqueues != tot_splice) {
 576                 printf("WARNING! Discrepancy between empty enqueue (%llu) and "
 577                         "number of non-empty splice (%llu)\n",
 578                         tot_empty_dest_enqueues,
 579                         tot_splice);
 580                 retval = 1;
 581         }
 582         free(count_enqueuer);
 583         free(count_dequeuer);
 584         free(tid_enqueuer);
 585         free(tid_dequeuer);
 586         return retval;
 587 }