tests/benchmark/test_urcu_wfs.c

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