tests/benchmark/test_urcu_wfq.c

   1 // SPDX-FileCopyrightText: 2010 Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
   2 // SPDX-FileCopyrightText: 2010 Paolo Bonzini <pbonzini@redhat.com>
   3 //
   4 // SPDX-License-Identifier: GPL-2.0-or-later
   5
   6 /*
   7  * Userspace RCU library - example RCU-based lock-free queue
   8  */
   9
  10 #include <stdio.h>
  11 #include <pthread.h>
  12 #include <stdlib.h>
  13 #include <stdint.h>
  14 #include <stdbool.h>
  15 #include <string.h>
  16 #include <sys/types.h>
  17 #include <sys/wait.h>
  18 #include <unistd.h>
  19 #include <stdio.h>
  20 #include <errno.h>
  21
  22 #include <urcu/arch.h>
  23 #include <urcu/assert.h>
  24 #include <urcu/tls-compat.h>
  25 #include "thread-id.h"
  26
  27 /* hardcoded number of CPUs */
  28 #define NR_CPUS 16384
  29
  30 #ifndef DYNAMIC_LINK_TEST
  31 #define _LGPL_SOURCE
  32 #endif
  33
  34 /* Remove deprecation warnings from test build. */
  35 #define CDS_WFQ_DEPRECATED
  36
  37 #include <urcu.h>
  38 #include <urcu/wfqueue.h>
  39
  40 static volatile int test_go, test_stop;
  41
  42 static unsigned long rduration;
  43
  44 static unsigned long duration;
  45
  46 /* read-side C.S. duration, in loops */
  47 static unsigned long wdelay;
  48
  49 static inline void loop_sleep(unsigned long loops)
  50 {
  51         while (loops-- != 0)
  52                 caa_cpu_relax();
  53 }
  54
  55 static int verbose_mode;
  56
  57 #define printf_verbose(fmt, args...)            \
  58         do {                                    \
  59                 if (verbose_mode)               \
  60                         printf(fmt, args);      \
  61         } while (0)
  62
  63 static unsigned int cpu_affinities[NR_CPUS];
  64 static unsigned int next_aff = 0;
  65 static int use_affinity = 0;
  66
  67 pthread_mutex_t affinity_mutex = PTHREAD_MUTEX_INITIALIZER;
  68
  69 static void set_affinity(void)
  70 {
  71 #ifdef HAVE_SCHED_SETAFFINITY
  72         cpu_set_t mask;
  73         int cpu, ret;
  74 #endif /* HAVE_SCHED_SETAFFINITY */
  75
  76         if (!use_affinity)
  77                 return;
  78
  79 #ifdef HAVE_SCHED_SETAFFINITY
  80         ret = pthread_mutex_lock(&affinity_mutex);
  81         if (ret) {
  82                 perror("Error in pthread mutex lock");
  83                 exit(-1);
  84         }
  85         cpu = cpu_affinities[next_aff++];
  86         ret = pthread_mutex_unlock(&affinity_mutex);
  87         if (ret) {
  88                 perror("Error in pthread mutex unlock");
  89                 exit(-1);
  90         }
  91
  92         CPU_ZERO(&mask);
  93         CPU_SET(cpu, &mask);
  94         sched_setaffinity(0, sizeof(mask), &mask);
  95 #endif /* HAVE_SCHED_SETAFFINITY */
  96 }
  97
  98 /*
  99  * returns 0 if test should end.
 100  */
 101 static int test_duration_dequeue(void)
 102 {
 103         return !test_stop;
 104 }
 105
 106 static int test_duration_enqueue(void)
 107 {
 108         return !test_stop;
 109 }
 110
 111 static DEFINE_URCU_TLS(unsigned long long, nr_dequeues);
 112 static DEFINE_URCU_TLS(unsigned long long, nr_enqueues);
 113
 114 static DEFINE_URCU_TLS(unsigned long long, nr_successful_dequeues);
 115 static DEFINE_URCU_TLS(unsigned long long, nr_successful_enqueues);
 116
 117 static unsigned int nr_enqueuers;
 118 static unsigned int nr_dequeuers;
 119
 120 static struct cds_wfq_queue q;
 121
 122 static
 123 void *thr_enqueuer(void *_count)
 124 {
 125         unsigned long long *count = _count;
 126
 127         printf_verbose("thread_begin %s, tid %lu\n",
 128                         "enqueuer", urcu_get_thread_id());
 129
 130         set_affinity();
 131
 132         while (!test_go)
 133         {
 134         }
 135         cmm_smp_mb();
 136
 137         for (;;) {
 138                 struct cds_wfq_node *node = malloc(sizeof(*node));
 139                 if (!node)
 140                         goto fail;
 141                 cds_wfq_node_init(node);
 142                 cds_wfq_enqueue(&q, node);
 143                 URCU_TLS(nr_successful_enqueues)++;
 144
 145                 if (caa_unlikely(wdelay))
 146                         loop_sleep(wdelay);
 147 fail:
 148                 URCU_TLS(nr_enqueues)++;
 149                 if (caa_unlikely(!test_duration_enqueue()))
 150                         break;
 151         }
 152
 153         count[0] = URCU_TLS(nr_enqueues);
 154         count[1] = URCU_TLS(nr_successful_enqueues);
 155         printf_verbose("enqueuer thread_end, tid %lu, "
 156                         "enqueues %llu successful_enqueues %llu\n",
 157                         urcu_get_thread_id(),
 158                         URCU_TLS(nr_enqueues),
 159                         URCU_TLS(nr_successful_enqueues));
 160         return ((void*)1);
 161
 162 }
 163
 164 static
 165 void *thr_dequeuer(void *_count)
 166 {
 167         unsigned long long *count = _count;
 168
 169         printf_verbose("thread_begin %s, tid %lu\n",
 170                         "dequeuer", urcu_get_thread_id());
 171
 172         set_affinity();
 173
 174         while (!test_go)
 175         {
 176         }
 177         cmm_smp_mb();
 178
 179         for (;;) {
 180                 struct cds_wfq_node *node = cds_wfq_dequeue_blocking(&q);
 181
 182                 if (node) {
 183                         free(node);
 184                         URCU_TLS(nr_successful_dequeues)++;
 185                 }
 186
 187                 URCU_TLS(nr_dequeues)++;
 188                 if (caa_unlikely(!test_duration_dequeue()))
 189                         break;
 190                 if (caa_unlikely(rduration))
 191                         loop_sleep(rduration);
 192         }
 193
 194         printf_verbose("dequeuer thread_end, tid %lu, "
 195                         "dequeues %llu, successful_dequeues %llu\n",
 196                         urcu_get_thread_id(),
 197                         URCU_TLS(nr_dequeues),
 198                         URCU_TLS(nr_successful_dequeues));
 199         count[0] = URCU_TLS(nr_dequeues);
 200         count[1] = URCU_TLS(nr_successful_dequeues);
 201         return ((void*)2);
 202 }
 203
 204 static
 205 void test_end(unsigned long long *nr_dequeues_l)
 206 {
 207         struct cds_wfq_node *node;
 208
 209         do {
 210                 node = cds_wfq_dequeue_blocking(&q);
 211                 if (node) {
 212                         free(node);
 213                         (*nr_dequeues_l)++;
 214                 }
 215         } while (node);
 216 }
 217
 218 static
 219 void show_usage(char **argv)
 220 {
 221         printf("Usage : %s nr_dequeuers nr_enqueuers duration (s) <OPTIONS>\n",
 222                 argv[0]);
 223         printf("OPTIONS:\n");
 224         printf("        [-d delay] (enqueuer period (in loops))\n");
 225         printf("        [-c duration] (dequeuer period (in loops))\n");
 226         printf("        [-v] (verbose output)\n");
 227         printf("        [-a cpu#] [-a cpu#]... (affinity)\n");
 228         printf("\n");
 229 }
 230
 231 int main(int argc, char **argv)
 232 {
 233         int err;
 234         pthread_t *tid_enqueuer, *tid_dequeuer;
 235         void *tret;
 236         unsigned long long *count_enqueuer, *count_dequeuer;
 237         unsigned long long tot_enqueues = 0, tot_dequeues = 0;
 238         unsigned long long tot_successful_enqueues = 0,
 239                            tot_successful_dequeues = 0;
 240         unsigned long long end_dequeues = 0;
 241         int i, a;
 242         unsigned int i_thr;
 243
 244         if (argc < 4) {
 245                 show_usage(argv);
 246                 return -1;
 247         }
 248
 249         err = sscanf(argv[1], "%u", &nr_dequeuers);
 250         if (err != 1) {
 251                 show_usage(argv);
 252                 return -1;
 253         }
 254
 255         err = sscanf(argv[2], "%u", &nr_enqueuers);
 256         if (err != 1) {
 257                 show_usage(argv);
 258                 return -1;
 259         }
 260
 261         err = sscanf(argv[3], "%lu", &duration);
 262         if (err != 1) {
 263                 show_usage(argv);
 264                 return -1;
 265         }
 266
 267         for (i = 4; i < argc; i++) {
 268                 if (argv[i][0] != '-')
 269                         continue;
 270                 switch (argv[i][1]) {
 271                 case 'a':
 272                         if (argc < i + 2) {
 273                                 show_usage(argv);
 274                                 return -1;
 275                         }
 276                         a = atoi(argv[++i]);
 277                         cpu_affinities[next_aff++] = a;
 278                         use_affinity = 1;
 279                         printf_verbose("Adding CPU %d affinity\n", a);
 280                         break;
 281                 case 'c':
 282                         if (argc < i + 2) {
 283                                 show_usage(argv);
 284                                 return -1;
 285                         }
 286                         rduration = atol(argv[++i]);
 287                         break;
 288                 case 'd':
 289                         if (argc < i + 2) {
 290                                 show_usage(argv);
 291                                 return -1;
 292                         }
 293                         wdelay = atol(argv[++i]);
 294                         break;
 295                 case 'v':
 296                         verbose_mode = 1;
 297                         break;
 298                 }
 299         }
 300
 301         printf_verbose("running test for %lu seconds, %u enqueuers, "
 302                        "%u dequeuers.\n",
 303                        duration, nr_enqueuers, nr_dequeuers);
 304         printf_verbose("Writer delay : %lu loops.\n", rduration);
 305         printf_verbose("Reader duration : %lu loops.\n", wdelay);
 306         printf_verbose("thread %-6s, tid %lu\n",
 307                         "main", urcu_get_thread_id());
 308
 309         tid_enqueuer = calloc(nr_enqueuers, sizeof(*tid_enqueuer));
 310         tid_dequeuer = calloc(nr_dequeuers, sizeof(*tid_dequeuer));
 311         count_enqueuer = calloc(nr_enqueuers, 2 * sizeof(*count_enqueuer));
 312         count_dequeuer = calloc(nr_dequeuers, 2 * sizeof(*count_dequeuer));
 313         cds_wfq_init(&q);
 314
 315         next_aff = 0;
 316
 317         for (i_thr = 0; i_thr < nr_enqueuers; i_thr++) {
 318                 err = pthread_create(&tid_enqueuer[i_thr], NULL, thr_enqueuer,
 319                                      &count_enqueuer[2 * i_thr]);
 320                 if (err != 0)
 321                         exit(1);
 322         }
 323         for (i_thr = 0; i_thr < nr_dequeuers; i_thr++) {
 324                 err = pthread_create(&tid_dequeuer[i_thr], NULL, thr_dequeuer,
 325                                      &count_dequeuer[2 * i_thr]);
 326                 if (err != 0)
 327                         exit(1);
 328         }
 329
 330         cmm_smp_mb();
 331
 332         test_go = 1;
 333
 334         for (i_thr = 0; i_thr < duration; i_thr++) {
 335                 sleep(1);
 336                 if (verbose_mode) {
 337                         fwrite(".", sizeof(char), 1, stdout);
 338                         fflush(stdout);
 339                 }
 340         }
 341
 342         test_stop = 1;
 343
 344         for (i_thr = 0; i_thr < nr_enqueuers; i_thr++) {
 345                 err = pthread_join(tid_enqueuer[i_thr], &tret);
 346                 if (err != 0)
 347                         exit(1);
 348                 tot_enqueues += count_enqueuer[2 * i_thr];
 349                 tot_successful_enqueues += count_enqueuer[2 * i_thr + 1];
 350         }
 351         for (i_thr = 0; i_thr < nr_dequeuers; i_thr++) {
 352                 err = pthread_join(tid_dequeuer[i_thr], &tret);
 353                 if (err != 0)
 354                         exit(1);
 355                 tot_dequeues += count_dequeuer[2 * i_thr];
 356                 tot_successful_dequeues += count_dequeuer[2 * i_thr + 1];
 357         }
 358
 359         test_end(&end_dequeues);
 360
 361         printf_verbose("total number of enqueues : %llu, dequeues %llu\n",
 362                        tot_enqueues, tot_dequeues);
 363         printf_verbose("total number of successful enqueues : %llu, "
 364                        "successful dequeues %llu\n",
 365                        tot_successful_enqueues, tot_successful_dequeues);
 366         printf("SUMMARY %-25s testdur %4lu nr_enqueuers %3u wdelay %6lu "
 367                 "nr_dequeuers %3u "
 368                 "rdur %6lu nr_enqueues %12llu nr_dequeues %12llu "
 369                 "successful enqueues %12llu successful dequeues %12llu "
 370                 "end_dequeues %llu nr_ops %12llu\n",
 371                 argv[0], duration, nr_enqueuers, wdelay,
 372                 nr_dequeuers, rduration, tot_enqueues, tot_dequeues,
 373                 tot_successful_enqueues,
 374                 tot_successful_dequeues, end_dequeues,
 375                 tot_enqueues + tot_dequeues);
 376         if (tot_successful_enqueues != tot_successful_dequeues + end_dequeues)
 377                 printf("WARNING! Discrepancy between nr succ. enqueues %llu vs "
 378                        "succ. dequeues + end dequeues %llu.\n",
 379                        tot_successful_enqueues,
 380                        tot_successful_dequeues + end_dequeues);
 381
 382         cds_wfq_destroy(&q);
 383         free(count_enqueuer);
 384         free(count_dequeuer);
 385         free(tid_enqueuer);
 386         free(tid_dequeuer);
 387
 388         return 0;
 389 }