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;
  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;
  96
  97 #define printf_verbose(fmt, args...)            \
  98         do {                                    \
  99                 if (verbose_mode)               \
 100                         printf(fmt, ## args);   \
 101         } while (0)
 102
 103 static unsigned int cpu_affinities[NR_CPUS];
 104 static unsigned int next_aff = 0;
 105 static int use_affinity = 0;
 106
 107 pthread_mutex_t affinity_mutex = PTHREAD_MUTEX_INITIALIZER;
 108
 109 #ifndef HAVE_CPU_SET_T
 110 typedef unsigned long cpu_set_t;
 111 # define CPU_ZERO(cpuset) do { *(cpuset) = 0; } while(0)
 112 # define CPU_SET(cpu, cpuset) do { *(cpuset) |= (1UL << (cpu)); } while(0)
 113 #endif
 114
 115 static void set_affinity(void)
 116 {
 117 #if HAVE_SCHED_SETAFFINITY
 118         cpu_set_t mask;
 119         int cpu, ret;
 120 #endif /* HAVE_SCHED_SETAFFINITY */
 121
 122         if (!use_affinity)
 123                 return;
 124
 125 #if HAVE_SCHED_SETAFFINITY
 126         ret = pthread_mutex_lock(&affinity_mutex);
 127         if (ret) {
 128                 perror("Error in pthread mutex lock");
 129                 exit(-1);
 130         }
 131         cpu = cpu_affinities[next_aff++];
 132         ret = pthread_mutex_unlock(&affinity_mutex);
 133         if (ret) {
 134                 perror("Error in pthread mutex unlock");
 135                 exit(-1);
 136         }
 137
 138         CPU_ZERO(&mask);
 139         CPU_SET(cpu, &mask);
 140 #if SCHED_SETAFFINITY_ARGS == 2
 141         sched_setaffinity(0, &mask);
 142 #else
 143         sched_setaffinity(0, sizeof(mask), &mask);
 144 #endif
 145 #endif /* HAVE_SCHED_SETAFFINITY */
 146 }
 147
 148 /*
 149  * returns 0 if test should end.
 150  */
 151 static int test_duration_dequeue(void)
 152 {
 153         return !test_stop;
 154 }
 155
 156 static int test_duration_enqueue(void)
 157 {
 158         return !test_stop;
 159 }
 160
 161 static DEFINE_URCU_TLS(unsigned long long, nr_dequeues);
 162 static DEFINE_URCU_TLS(unsigned long long, nr_enqueues);
 163
 164 static DEFINE_URCU_TLS(unsigned long long, nr_successful_dequeues);
 165 static DEFINE_URCU_TLS(unsigned long long, nr_successful_enqueues);
 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
 177         printf_verbose("thread_begin %s, thread id : %lx, tid %lu\n",
 178                         "enqueuer", pthread_self(), (unsigned long)gettid());
 179
 180         set_affinity();
 181
 182         while (!test_go)
 183         {
 184         }
 185         cmm_smp_mb();
 186
 187         for (;;) {
 188                 struct cds_wfcq_node *node = malloc(sizeof(*node));
 189                 if (!node)
 190                         goto fail;
 191                 cds_wfcq_node_init(node);
 192                 cds_wfcq_enqueue(&head, &tail, node);
 193                 URCU_TLS(nr_successful_enqueues)++;
 194
 195                 if (caa_unlikely(wdelay))
 196                         loop_sleep(wdelay);
 197 fail:
 198                 URCU_TLS(nr_enqueues)++;
 199                 if (caa_unlikely(!test_duration_enqueue()))
 200                         break;
 201         }
 202
 203         count[0] = URCU_TLS(nr_enqueues);
 204         count[1] = URCU_TLS(nr_successful_enqueues);
 205         printf_verbose("enqueuer thread_end, thread id : %lx, tid %lu, "
 206                        "enqueues %llu successful_enqueues %llu\n",
 207                        pthread_self(), (unsigned long)gettid(),
 208                        URCU_TLS(nr_enqueues), URCU_TLS(nr_successful_enqueues));
 209         return ((void*)1);
 210
 211 }
 212
 213 static void do_test_dequeue(enum test_sync sync)
 214 {
 215         struct cds_wfcq_node *node;
 216
 217         if (sync == TEST_SYNC_MUTEX)
 218                 node = cds_wfcq_dequeue_blocking(&head, &tail);
 219         else
 220                 node = __cds_wfcq_dequeue_blocking(&head, &tail);
 221
 222         if (node) {
 223                 free(node);
 224                 URCU_TLS(nr_successful_dequeues)++;
 225         }
 226         URCU_TLS(nr_dequeues)++;
 227 }
 228
 229 static void do_test_splice(enum test_sync sync)
 230 {
 231         struct cds_wfcq_head tmp_head;
 232         struct cds_wfcq_tail tmp_tail;
 233         struct cds_wfcq_node *node, *n;
 234
 235         cds_wfcq_init(&tmp_head, &tmp_tail);
 236
 237         if (sync == TEST_SYNC_MUTEX)
 238                 cds_wfcq_splice_blocking(&tmp_head, &tmp_tail,
 239                         &head, &tail);
 240         else
 241                 __cds_wfcq_splice_blocking(&tmp_head, &tmp_tail,
 242                         &head, &tail);
 243
 244         __cds_wfcq_for_each_blocking_safe(&tmp_head, &tmp_tail, node, n) {
 245                 free(node);
 246                 URCU_TLS(nr_successful_dequeues)++;
 247                 URCU_TLS(nr_dequeues)++;
 248         }
 249 }
 250
 251 static void *thr_dequeuer(void *_count)
 252 {
 253         unsigned long long *count = _count;
 254         unsigned int counter;
 255
 256         printf_verbose("thread_begin %s, thread id : %lx, tid %lu\n",
 257                         "dequeuer", pthread_self(), (unsigned long)gettid());
 258
 259         set_affinity();
 260
 261         while (!test_go)
 262         {
 263         }
 264         cmm_smp_mb();
 265
 266         for (;;) {
 267                 if (test_dequeue && test_splice) {
 268                         if (counter & 1)
 269                                 do_test_dequeue(test_sync);
 270                         else
 271                                 do_test_splice(test_sync);
 272                         counter++;
 273                 } else {
 274                         if (test_dequeue)
 275                                 do_test_dequeue(test_sync);
 276                         else
 277                                 do_test_splice(test_sync);
 278                 }
 279                 if (caa_unlikely(!test_duration_dequeue()))
 280                         break;
 281                 if (caa_unlikely(rduration))
 282                         loop_sleep(rduration);
 283         }
 284
 285         printf_verbose("dequeuer thread_end, thread id : %lx, tid %lu, "
 286                        "dequeues %llu, successful_dequeues %llu\n",
 287                        pthread_self(), (unsigned long)gettid(),
 288                        URCU_TLS(nr_dequeues), URCU_TLS(nr_successful_dequeues));
 289         count[0] = URCU_TLS(nr_dequeues);
 290         count[1] = URCU_TLS(nr_successful_dequeues);
 291         return ((void*)2);
 292 }
 293
 294 static void test_end(unsigned long long *nr_dequeues)
 295 {
 296         struct cds_wfcq_node *node;
 297
 298         do {
 299                 node = cds_wfcq_dequeue_blocking(&head, &tail);
 300                 if (node) {
 301                         free(node);
 302                         (*nr_dequeues)++;
 303                 }
 304         } while (node);
 305 }
 306
 307 static void show_usage(int argc, char **argv)
 308 {
 309         printf("Usage : %s nr_dequeuers nr_enqueuers duration (s)", argv[0]);
 310         printf(" [-d delay] (enqueuer period (in loops))");
 311         printf(" [-c duration] (dequeuer period (in loops))");
 312         printf(" [-v] (verbose output)");
 313         printf(" [-a cpu#] [-a cpu#]... (affinity)");
 314         printf(" [-q] (test dequeue)");
 315         printf(" [-s] (test splice, enabled by default)");
 316         printf(" [-M] (use mutex external synchronization)");
 317         printf(" [-0] (use no external synchronization)");
 318         printf("      Note: default: mutex external synchronization used.");
 319         printf("\n");
 320 }
 321
 322 int main(int argc, char **argv)
 323 {
 324         int err;
 325         pthread_t *tid_enqueuer, *tid_dequeuer;
 326         void *tret;
 327         unsigned long long *count_enqueuer, *count_dequeuer;
 328         unsigned long long tot_enqueues = 0, tot_dequeues = 0;
 329         unsigned long long tot_successful_enqueues = 0,
 330                            tot_successful_dequeues = 0;
 331         unsigned long long end_dequeues = 0;
 332         int i, a;
 333
 334         if (argc < 4) {
 335                 show_usage(argc, argv);
 336                 return -1;
 337         }
 338
 339         err = sscanf(argv[1], "%u", &nr_dequeuers);
 340         if (err != 1) {
 341                 show_usage(argc, argv);
 342                 return -1;
 343         }
 344
 345         err = sscanf(argv[2], "%u", &nr_enqueuers);
 346         if (err != 1) {
 347                 show_usage(argc, argv);
 348                 return -1;
 349         }
 350
 351         err = sscanf(argv[3], "%lu", &duration);
 352         if (err != 1) {
 353                 show_usage(argc, argv);
 354                 return -1;
 355         }
 356
 357         for (i = 4; i < argc; i++) {
 358                 if (argv[i][0] != '-')
 359                         continue;
 360                 switch (argv[i][1]) {
 361                 case 'a':
 362                         if (argc < i + 2) {
 363                                 show_usage(argc, argv);
 364                                 return -1;
 365                         }
 366                         a = atoi(argv[++i]);
 367                         cpu_affinities[next_aff++] = a;
 368                         use_affinity = 1;
 369                         printf_verbose("Adding CPU %d affinity\n", a);
 370                         break;
 371                 case 'c':
 372                         if (argc < i + 2) {
 373                                 show_usage(argc, argv);
 374                                 return -1;
 375                         }
 376                         rduration = atol(argv[++i]);
 377                         break;
 378                 case 'd':
 379                         if (argc < i + 2) {
 380                                 show_usage(argc, argv);
 381                                 return -1;
 382                         }
 383                         wdelay = atol(argv[++i]);
 384                         break;
 385                 case 'v':
 386                         verbose_mode = 1;
 387                         break;
 388                 case 'q':
 389                         test_dequeue = 1;
 390                         break;
 391                 case 's':
 392                         test_splice = 1;
 393                         break;
 394                 case 'M':
 395                         test_sync = TEST_SYNC_MUTEX;
 396                         break;
 397                 case '0':
 398                         test_sync = TEST_SYNC_NONE;
 399                         break;
 400                 }
 401         }
 402
 403         /* activate splice test by default */
 404         if (!test_dequeue && !test_splice)
 405                 test_splice = 1;
 406
 407         printf_verbose("running test for %lu seconds, %u enqueuers, "
 408                        "%u dequeuers.\n",
 409                        duration, nr_enqueuers, nr_dequeuers);
 410         if (test_dequeue)
 411                 printf_verbose("dequeue test activated.\n");
 412         else
 413                 printf_verbose("splice test activated.\n");
 414         if (test_sync == TEST_SYNC_MUTEX)
 415                 printf_verbose("External sync: mutex.\n");
 416         else
 417                 printf_verbose("External sync: none.\n");
 418         printf_verbose("Writer delay : %lu loops.\n", rduration);
 419         printf_verbose("Reader duration : %lu loops.\n", wdelay);
 420         printf_verbose("thread %-6s, thread id : %lx, tid %lu\n",
 421                         "main", pthread_self(), (unsigned long)gettid());
 422
 423         tid_enqueuer = malloc(sizeof(*tid_enqueuer) * nr_enqueuers);
 424         tid_dequeuer = malloc(sizeof(*tid_dequeuer) * nr_dequeuers);
 425         count_enqueuer = malloc(2 * sizeof(*count_enqueuer) * nr_enqueuers);
 426         count_dequeuer = malloc(2 * sizeof(*count_dequeuer) * nr_dequeuers);
 427         cds_wfcq_init(&head, &tail);
 428
 429         next_aff = 0;
 430
 431         for (i = 0; i < nr_enqueuers; i++) {
 432                 err = pthread_create(&tid_enqueuer[i], NULL, thr_enqueuer,
 433                                      &count_enqueuer[2 * i]);
 434                 if (err != 0)
 435                         exit(1);
 436         }
 437         for (i = 0; i < nr_dequeuers; i++) {
 438                 err = pthread_create(&tid_dequeuer[i], NULL, thr_dequeuer,
 439                                      &count_dequeuer[2 * i]);
 440                 if (err != 0)
 441                         exit(1);
 442         }
 443
 444         cmm_smp_mb();
 445
 446         test_go = 1;
 447
 448         for (i = 0; i < duration; i++) {
 449                 sleep(1);
 450                 if (verbose_mode)
 451                         write (1, ".", 1);
 452         }
 453
 454         test_stop = 1;
 455
 456         for (i = 0; i < nr_enqueuers; i++) {
 457                 err = pthread_join(tid_enqueuer[i], &tret);
 458                 if (err != 0)
 459                         exit(1);
 460                 tot_enqueues += count_enqueuer[2 * i];
 461                 tot_successful_enqueues += count_enqueuer[2 * i + 1];
 462         }
 463         for (i = 0; i < nr_dequeuers; i++) {
 464                 err = pthread_join(tid_dequeuer[i], &tret);
 465                 if (err != 0)
 466                         exit(1);
 467                 tot_dequeues += count_dequeuer[2 * i];
 468                 tot_successful_dequeues += count_dequeuer[2 * i + 1];
 469         }
 470
 471         test_end(&end_dequeues);
 472
 473         printf_verbose("total number of enqueues : %llu, dequeues %llu\n",
 474                        tot_enqueues, tot_dequeues);
 475         printf_verbose("total number of successful enqueues : %llu, "
 476                        "successful dequeues %llu\n",
 477                        tot_successful_enqueues, tot_successful_dequeues);
 478         printf("SUMMARY %-25s testdur %4lu nr_enqueuers %3u wdelay %6lu "
 479                 "nr_dequeuers %3u "
 480                 "rdur %6lu nr_enqueues %12llu nr_dequeues %12llu "
 481                 "successful enqueues %12llu successful dequeues %12llu "
 482                 "end_dequeues %llu nr_ops %12llu\n",
 483                 argv[0], duration, nr_enqueuers, wdelay,
 484                 nr_dequeuers, rduration, tot_enqueues, tot_dequeues,
 485                 tot_successful_enqueues,
 486                 tot_successful_dequeues, end_dequeues,
 487                 tot_enqueues + tot_dequeues);
 488         if (tot_successful_enqueues != tot_successful_dequeues + end_dequeues)
 489                 printf("WARNING! Discrepancy between nr succ. enqueues %llu vs "
 490                        "succ. dequeues + end dequeues %llu.\n",
 491                        tot_successful_enqueues,
 492                        tot_successful_dequeues + end_dequeues);
 493
 494         free(count_enqueuer);
 495         free(count_dequeuer);
 496         free(tid_enqueuer);
 497         free(tid_dequeuer);
 498         return 0;
 499 }