tests/benchmark/test_urcu_lfs.c

   1 /*
   2  * test_urcu_lfs.c
   3  *
   4  * Userspace RCU library - example lock-free stack
   5  *
   6  * Copyright 2010-2012 - 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 #include <stdio.h>
  25 #include <pthread.h>
  26 #include <stdlib.h>
  27 #include <stdint.h>
  28 #include <stdbool.h>
  29 #include <string.h>
  30 #include <sys/types.h>
  31 #include <sys/wait.h>
  32 #include <unistd.h>
  33 #include <stdio.h>
  34 #include <assert.h>
  35 #include <errno.h>
  36
  37 #include <urcu/arch.h>
  38 #include <urcu/tls-compat.h>
  39 #include "thread-id.h"
  40
  41 /* hardcoded number of CPUs */
  42 #define NR_CPUS 16384
  43
  44 #ifndef DYNAMIC_LINK_TEST
  45 #define _LGPL_SOURCE
  46 #endif
  47 #include <urcu.h>
  48 #include <urcu/cds.h>
  49
  50 #define POISON_PTR      ((void *) 0x42UL)
  51
  52 /*
  53  * External synchronization used.
  54  */
  55 enum test_sync {
  56         TEST_SYNC_NONE = 0,
  57         TEST_SYNC_RCU,
  58 };
  59
  60 static enum test_sync test_sync;
  61
  62 static volatile int test_go, test_stop;
  63
  64 static unsigned long rduration;
  65
  66 static unsigned long duration;
  67
  68 /* read-side C.S. duration, in loops */
  69 static unsigned long wdelay;
  70
  71 static inline void loop_sleep(unsigned long loops)
  72 {
  73         while (loops-- != 0)
  74                 caa_cpu_relax();
  75 }
  76
  77 static int verbose_mode;
  78
  79 static int test_pop, test_pop_all;
  80
  81 #define printf_verbose(fmt, args...)            \
  82         do {                                    \
  83                 if (verbose_mode)               \
  84                         printf(fmt, ## args);   \
  85         } while (0)
  86
  87 static unsigned int cpu_affinities[NR_CPUS];
  88 static unsigned int next_aff = 0;
  89 static int use_affinity = 0;
  90
  91 pthread_mutex_t affinity_mutex = PTHREAD_MUTEX_INITIALIZER;
  92
  93 static void set_affinity(void)
  94 {
  95 #if HAVE_SCHED_SETAFFINITY
  96         cpu_set_t mask;
  97         int cpu, ret;
  98 #endif /* HAVE_SCHED_SETAFFINITY */
  99
 100         if (!use_affinity)
 101                 return;
 102
 103 #if HAVE_SCHED_SETAFFINITY
 104         ret = pthread_mutex_lock(&affinity_mutex);
 105         if (ret) {
 106                 perror("Error in pthread mutex lock");
 107                 exit(-1);
 108         }
 109         cpu = cpu_affinities[next_aff++];
 110         ret = pthread_mutex_unlock(&affinity_mutex);
 111         if (ret) {
 112                 perror("Error in pthread mutex unlock");
 113                 exit(-1);
 114         }
 115
 116         CPU_ZERO(&mask);
 117         CPU_SET(cpu, &mask);
 118         sched_setaffinity(0, sizeof(mask), &mask);
 119 #endif /* HAVE_SCHED_SETAFFINITY */
 120 }
 121
 122 /*
 123  * returns 0 if test should end.
 124  */
 125 static int test_duration_dequeue(void)
 126 {
 127         return !test_stop;
 128 }
 129
 130 static int test_duration_enqueue(void)
 131 {
 132         return !test_stop;
 133 }
 134
 135 static DEFINE_URCU_TLS(unsigned long long, nr_dequeues);
 136 static DEFINE_URCU_TLS(unsigned long long, nr_enqueues);
 137
 138 static DEFINE_URCU_TLS(unsigned long long, nr_successful_dequeues);
 139 static DEFINE_URCU_TLS(unsigned long long, nr_successful_enqueues);
 140
 141 static unsigned int nr_enqueuers;
 142 static unsigned int nr_dequeuers;
 143
 144 struct test {
 145         struct cds_lfs_node list;
 146         struct rcu_head rcu;
 147 };
 148
 149 static struct cds_lfs_stack s;
 150
 151 static void *thr_enqueuer(void *_count)
 152 {
 153         unsigned long long *count = _count;
 154
 155         printf_verbose("thread_begin %s, tid %lu\n",
 156                         "enqueuer", urcu_get_thread_id());
 157
 158         set_affinity();
 159
 160         rcu_register_thread();
 161
 162         while (!test_go)
 163         {
 164         }
 165         cmm_smp_mb();
 166
 167         for (;;) {
 168                 struct test *node = malloc(sizeof(*node));
 169                 if (!node)
 170                         goto fail;
 171                 cds_lfs_node_init(&node->list);
 172                 cds_lfs_push(&s, &node->list);
 173                 URCU_TLS(nr_successful_enqueues)++;
 174
 175                 if (caa_unlikely(wdelay))
 176                         loop_sleep(wdelay);
 177 fail:
 178                 URCU_TLS(nr_enqueues)++;
 179                 if (caa_unlikely(!test_duration_enqueue()))
 180                         break;
 181         }
 182
 183         rcu_unregister_thread();
 184
 185         count[0] = URCU_TLS(nr_enqueues);
 186         count[1] = URCU_TLS(nr_successful_enqueues);
 187         printf_verbose("enqueuer thread_end, tid %lu, "
 188                         "enqueues %llu successful_enqueues %llu\n",
 189                         urcu_get_thread_id(),
 190                         URCU_TLS(nr_enqueues),
 191                         URCU_TLS(nr_successful_enqueues));
 192         return ((void*)1);
 193
 194 }
 195
 196 static
 197 void free_node_cb(struct rcu_head *head)
 198 {
 199         struct test *node =
 200                 caa_container_of(head, struct test, rcu);
 201         free(node);
 202 }
 203
 204 static
 205 void do_test_pop(enum test_sync sync)
 206 {
 207         struct cds_lfs_node *snode;
 208
 209         if (sync == TEST_SYNC_RCU)
 210                 rcu_read_lock();
 211         snode = __cds_lfs_pop(&s);
 212         if (sync == TEST_SYNC_RCU)
 213                 rcu_read_unlock();
 214         if (snode) {
 215                 struct test *node;
 216
 217                 snode->next = POISON_PTR;
 218                 node = caa_container_of(snode,
 219                         struct test, list);
 220                 if (sync == TEST_SYNC_RCU)
 221                         call_rcu(&node->rcu, free_node_cb);
 222                 else
 223                         free(node);
 224                 URCU_TLS(nr_successful_dequeues)++;
 225         }
 226         URCU_TLS(nr_dequeues)++;
 227 }
 228
 229 static
 230 void do_test_pop_all(enum test_sync sync)
 231 {
 232         struct cds_lfs_node *snode;
 233         struct cds_lfs_head *head;
 234         struct cds_lfs_node *n;
 235
 236         head = __cds_lfs_pop_all(&s);
 237         cds_lfs_for_each_safe(head, snode, n) {
 238                 struct test *node;
 239
 240                 snode->next = POISON_PTR;
 241                 node = caa_container_of(snode, struct test, list);
 242                 if (sync == TEST_SYNC_RCU)
 243                         call_rcu(&node->rcu, free_node_cb);
 244                 else
 245                         free(node);
 246                 URCU_TLS(nr_successful_dequeues)++;
 247                 URCU_TLS(nr_dequeues)++;
 248         }
 249
 250 }
 251
 252 static void *thr_dequeuer(void *_count)
 253 {
 254         unsigned long long *count = _count;
 255         unsigned int counter = 0;
 256
 257         printf_verbose("thread_begin %s, tid %lu\n",
 258                         "dequeuer", urcu_get_thread_id());
 259
 260         set_affinity();
 261
 262         rcu_register_thread();
 263
 264         while (!test_go)
 265         {
 266         }
 267         cmm_smp_mb();
 268
 269         assert(test_pop || test_pop_all);
 270
 271         for (;;) {
 272                 if (test_pop && test_pop_all) {
 273                         /* both pop and pop all */
 274                         if (counter & 1)
 275                                 do_test_pop(test_sync);
 276                         else
 277                                 do_test_pop_all(test_sync);
 278                         counter++;
 279                 } else {
 280                         if (test_pop)
 281                                 do_test_pop(test_sync);
 282                         else
 283                                 do_test_pop_all(test_sync);
 284                 }
 285
 286                 if (caa_unlikely(!test_duration_dequeue()))
 287                         break;
 288                 if (caa_unlikely(rduration))
 289                         loop_sleep(rduration);
 290         }
 291
 292         rcu_unregister_thread();
 293
 294         printf_verbose("dequeuer thread_end, tid %lu, "
 295                         "dequeues %llu, successful_dequeues %llu\n",
 296                         urcu_get_thread_id(),
 297                         URCU_TLS(nr_dequeues),
 298                         URCU_TLS(nr_successful_dequeues));
 299         count[0] = URCU_TLS(nr_dequeues);
 300         count[1] = URCU_TLS(nr_successful_dequeues);
 301         return ((void*)2);
 302 }
 303
 304 static void test_end(struct cds_lfs_stack *s, unsigned long long *nr_dequeues)
 305 {
 306         struct cds_lfs_node *snode;
 307
 308         do {
 309                 snode = __cds_lfs_pop(s);
 310                 if (snode) {
 311                         struct test *node;
 312
 313                         node = caa_container_of(snode, struct test, list);
 314                         free(node);
 315                         (*nr_dequeues)++;
 316                 }
 317         } while (snode);
 318 }
 319
 320 static void show_usage(int argc, char **argv)
 321 {
 322         printf("Usage : %s nr_dequeuers nr_enqueuers duration (s) <OPTIONS>\n",
 323                 argv[0]);
 324         printf("OPTIONS:\n");
 325         printf("        [-d delay] (enqueuer period (in loops))\n");
 326         printf("        [-c duration] (dequeuer period (in loops))\n");
 327         printf("        [-v] (verbose output)\n");
 328         printf("        [-a cpu#] [-a cpu#]... (affinity)\n");
 329         printf("        [-p] (test pop)\n");
 330         printf("        [-P] (test pop_all, enabled by default)\n");
 331         printf("        [-R] (use RCU external synchronization)\n");
 332         printf("                Note: default: no external synchronization used.\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         unsigned long long end_dequeues = 0;
 346         int i, a;
 347         unsigned int i_thr;
 348
 349         if (argc < 4) {
 350                 show_usage(argc, argv);
 351                 return -1;
 352         }
 353
 354         err = sscanf(argv[1], "%u", &nr_dequeuers);
 355         if (err != 1) {
 356                 show_usage(argc, argv);
 357                 return -1;
 358         }
 359
 360         err = sscanf(argv[2], "%u", &nr_enqueuers);
 361         if (err != 1) {
 362                 show_usage(argc, argv);
 363                 return -1;
 364         }
 365
 366         err = sscanf(argv[3], "%lu", &duration);
 367         if (err != 1) {
 368                 show_usage(argc, argv);
 369                 return -1;
 370         }
 371
 372         for (i = 4; i < argc; i++) {
 373                 if (argv[i][0] != '-')
 374                         continue;
 375                 switch (argv[i][1]) {
 376                 case 'a':
 377                         if (argc < i + 2) {
 378                                 show_usage(argc, argv);
 379                                 return -1;
 380                         }
 381                         a = atoi(argv[++i]);
 382                         cpu_affinities[next_aff++] = a;
 383                         use_affinity = 1;
 384                         printf_verbose("Adding CPU %d affinity\n", a);
 385                         break;
 386                 case 'c':
 387                         if (argc < i + 2) {
 388                                 show_usage(argc, argv);
 389                                 return -1;
 390                         }
 391                         rduration = atol(argv[++i]);
 392                         break;
 393                 case 'd':
 394                         if (argc < i + 2) {
 395                                 show_usage(argc, argv);
 396                                 return -1;
 397                         }
 398                         wdelay = atol(argv[++i]);
 399                         break;
 400                 case 'v':
 401                         verbose_mode = 1;
 402                         break;
 403                 case 'p':
 404                         test_pop = 1;
 405                         break;
 406                 case 'P':
 407                         test_pop_all = 1;
 408                         break;
 409                 case 'R':
 410                         test_sync = TEST_SYNC_RCU;
 411                         break;
 412                 }
 413         }
 414
 415         /* activate pop_all test by default */
 416         if (!test_pop && !test_pop_all)
 417                 test_pop_all = 1;
 418
 419         printf_verbose("running test for %lu seconds, %u enqueuers, "
 420                        "%u dequeuers.\n",
 421                        duration, nr_enqueuers, nr_dequeuers);
 422         if (test_pop)
 423                 printf_verbose("pop test activated.\n");
 424         if (test_pop_all)
 425                 printf_verbose("pop_all test activated.\n");
 426         if (test_sync == TEST_SYNC_RCU)
 427                 printf_verbose("External sync: RCU.\n");
 428         else
 429                 printf_verbose("External sync: none.\n");
 430         printf_verbose("Writer delay : %lu loops.\n", rduration);
 431         printf_verbose("Reader duration : %lu loops.\n", wdelay);
 432         printf_verbose("thread %-6s, tid %lu\n",
 433                         "main", urcu_get_thread_id());
 434
 435         tid_enqueuer = calloc(nr_enqueuers, sizeof(*tid_enqueuer));
 436         tid_dequeuer = calloc(nr_dequeuers, sizeof(*tid_dequeuer));
 437         count_enqueuer = calloc(nr_enqueuers, 2 * sizeof(*count_enqueuer));
 438         count_dequeuer = calloc(nr_dequeuers, 2 * sizeof(*count_dequeuer));
 439         cds_lfs_init(&s);
 440         err = create_all_cpu_call_rcu_data(0);
 441         if (err) {
 442                 printf("Per-CPU call_rcu() worker threads unavailable. Using default global worker thread.\n");
 443         }
 444
 445         next_aff = 0;
 446
 447         for (i_thr = 0; i_thr < nr_enqueuers; i_thr++) {
 448                 err = pthread_create(&tid_enqueuer[i_thr], NULL, thr_enqueuer,
 449                                      &count_enqueuer[2 * i_thr]);
 450                 if (err != 0)
 451                         exit(1);
 452         }
 453         for (i_thr = 0; i_thr < nr_dequeuers; i_thr++) {
 454                 err = pthread_create(&tid_dequeuer[i_thr], NULL, thr_dequeuer,
 455                                      &count_dequeuer[2 * i_thr]);
 456                 if (err != 0)
 457                         exit(1);
 458         }
 459
 460         cmm_smp_mb();
 461
 462         test_go = 1;
 463
 464         for (i_thr = 0; i_thr < duration; i_thr++) {
 465                 sleep(1);
 466                 if (verbose_mode) {
 467                         fwrite(".", sizeof(char), 1, stdout);
 468                         fflush(stdout);
 469                 }
 470         }
 471
 472         test_stop = 1;
 473
 474         for (i_thr = 0; i_thr < nr_enqueuers; i_thr++) {
 475                 err = pthread_join(tid_enqueuer[i_thr], &tret);
 476                 if (err != 0)
 477                         exit(1);
 478                 tot_enqueues += count_enqueuer[2 * i_thr];
 479                 tot_successful_enqueues += count_enqueuer[2 * i_thr + 1];
 480         }
 481         for (i_thr = 0; i_thr < nr_dequeuers; i_thr++) {
 482                 err = pthread_join(tid_dequeuer[i_thr], &tret);
 483                 if (err != 0)
 484                         exit(1);
 485                 tot_dequeues += count_dequeuer[2 * i_thr];
 486                 tot_successful_dequeues += count_dequeuer[2 * i_thr + 1];
 487         }
 488
 489         test_end(&s, &end_dequeues);
 490
 491         printf_verbose("total number of enqueues : %llu, dequeues %llu\n",
 492                        tot_enqueues, tot_dequeues);
 493         printf_verbose("total number of successful enqueues : %llu, "
 494                        "successful dequeues %llu\n",
 495                        tot_successful_enqueues, tot_successful_dequeues);
 496         printf("SUMMARY %-25s testdur %4lu nr_enqueuers %3u wdelay %6lu "
 497                 "nr_dequeuers %3u "
 498                 "rdur %6lu nr_enqueues %12llu nr_dequeues %12llu "
 499                 "successful enqueues %12llu successful dequeues %12llu "
 500                 "end_dequeues %llu nr_ops %12llu\n",
 501                 argv[0], duration, nr_enqueuers, wdelay,
 502                 nr_dequeuers, rduration, tot_enqueues, tot_dequeues,
 503                 tot_successful_enqueues,
 504                 tot_successful_dequeues, end_dequeues,
 505                 tot_enqueues + tot_dequeues);
 506         if (tot_successful_enqueues != tot_successful_dequeues + end_dequeues)
 507                 printf("WARNING! Discrepancy between nr succ. enqueues %llu vs "
 508                        "succ. dequeues + end dequeues %llu.\n",
 509                        tot_successful_enqueues,
 510                        tot_successful_dequeues + end_dequeues);
 511
 512         free_all_cpu_call_rcu_data();
 513         cds_lfs_destroy(&s);
 514         free(count_enqueuer);
 515         free(count_dequeuer);
 516         free(tid_enqueuer);
 517         free(tid_dequeuer);
 518
 519         return 0;
 520 }