tests/benchmark/test_urcu_lfq.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 #include <urcu.h>
  34 #include <urcu/cds.h>
  35
  36 static unsigned long rduration;
  37
  38 static unsigned long duration;
  39
  40 /* read-side C.S. duration, in loops */
  41 static unsigned long wdelay;
  42
  43 static inline void loop_sleep(unsigned long loops)
  44 {
  45         while (loops-- != 0)
  46                 caa_cpu_relax();
  47 }
  48
  49 static int verbose_mode;
  50
  51 #define printf_verbose(fmt, args...)            \
  52         do {                                    \
  53                 if (verbose_mode)               \
  54                         printf(fmt, args);      \
  55         } while (0)
  56
  57 static unsigned int cpu_affinities[NR_CPUS];
  58 static unsigned int next_aff = 0;
  59 static int use_affinity = 0;
  60
  61 pthread_mutex_t affinity_mutex = PTHREAD_MUTEX_INITIALIZER;
  62
  63 static void set_affinity(void)
  64 {
  65 #ifdef HAVE_SCHED_SETAFFINITY
  66         cpu_set_t mask;
  67         int cpu, ret;
  68 #endif /* HAVE_SCHED_SETAFFINITY */
  69
  70         if (!use_affinity)
  71                 return;
  72
  73 #ifdef HAVE_SCHED_SETAFFINITY
  74         ret = pthread_mutex_lock(&affinity_mutex);
  75         if (ret) {
  76                 perror("Error in pthread mutex lock");
  77                 exit(-1);
  78         }
  79         cpu = cpu_affinities[next_aff++];
  80         ret = pthread_mutex_unlock(&affinity_mutex);
  81         if (ret) {
  82                 perror("Error in pthread mutex unlock");
  83                 exit(-1);
  84         }
  85
  86         CPU_ZERO(&mask);
  87         CPU_SET(cpu, &mask);
  88         sched_setaffinity(0, sizeof(mask), &mask);
  89 #endif /* HAVE_SCHED_SETAFFINITY */
  90 }
  91
  92 /*
  93  * returns 0 if test should end.
  94  */
  95 static int test_duration_dequeue(void)
  96 {
  97         return test_duration_read();
  98 }
  99
 100 static int test_duration_enqueue(void)
 101 {
 102         return test_duration_write();
 103 }
 104
 105 static DEFINE_URCU_TLS(unsigned long long, nr_dequeues);
 106 static DEFINE_URCU_TLS(unsigned long long, nr_enqueues);
 107
 108 static DEFINE_URCU_TLS(unsigned long long, nr_successful_dequeues);
 109 static DEFINE_URCU_TLS(unsigned long long, nr_successful_enqueues);
 110
 111 static unsigned int nr_enqueuers;
 112 static unsigned int nr_dequeuers;
 113
 114 struct test {
 115         struct cds_lfq_node_rcu list;
 116         struct rcu_head rcu;
 117 };
 118
 119 static struct cds_lfq_queue_rcu q;
 120
 121 static
 122 void *thr_enqueuer(void *_count)
 123 {
 124         unsigned long long *count = _count;
 125
 126         printf_verbose("thread_begin %s, tid %lu\n",
 127                         "enqueuer", urcu_get_thread_id());
 128
 129         set_affinity();
 130
 131         rcu_register_thread();
 132
 133         wait_until_go();
 134
 135         for (;;) {
 136                 struct test *node = malloc(sizeof(*node));
 137                 if (!node)
 138                         goto fail;
 139                 cds_lfq_node_init_rcu(&node->list);
 140                 rcu_read_lock();
 141                 cds_lfq_enqueue_rcu(&q, &node->list);
 142                 rcu_read_unlock();
 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         rcu_unregister_thread();
 154
 155         count[0] = URCU_TLS(nr_enqueues);
 156         count[1] = URCU_TLS(nr_successful_enqueues);
 157         printf_verbose("enqueuer thread_end, tid %lu, "
 158                         "enqueues %llu successful_enqueues %llu\n",
 159                         urcu_get_thread_id(),
 160                         URCU_TLS(nr_enqueues),
 161                         URCU_TLS(nr_successful_enqueues));
 162         return ((void*)1);
 163
 164 }
 165
 166 static
 167 void free_node_cb(struct rcu_head *head)
 168 {
 169         struct test *node =
 170                 caa_container_of(head, struct test, rcu);
 171         free(node);
 172 }
 173
 174 static
 175 void *thr_dequeuer(void *_count)
 176 {
 177         unsigned long long *count = _count;
 178
 179         printf_verbose("thread_begin %s, tid %lu\n",
 180                         "dequeuer", urcu_get_thread_id());
 181
 182         set_affinity();
 183
 184         rcu_register_thread();
 185
 186         wait_until_go();
 187
 188         for (;;) {
 189                 struct cds_lfq_node_rcu *qnode;
 190
 191                 rcu_read_lock();
 192                 qnode = cds_lfq_dequeue_rcu(&q);
 193                 rcu_read_unlock();
 194
 195                 if (qnode) {
 196                         struct test *node;
 197
 198                         node = caa_container_of(qnode, struct test, list);
 199                         call_rcu(&node->rcu, free_node_cb);
 200                         URCU_TLS(nr_successful_dequeues)++;
 201                 }
 202
 203                 URCU_TLS(nr_dequeues)++;
 204                 if (caa_unlikely(!test_duration_dequeue()))
 205                         break;
 206                 if (caa_unlikely(rduration))
 207                         loop_sleep(rduration);
 208         }
 209
 210         rcu_unregister_thread();
 211         printf_verbose("dequeuer thread_end, tid %lu, "
 212                         "dequeues %llu, successful_dequeues %llu\n",
 213                         urcu_get_thread_id(),
 214                         URCU_TLS(nr_dequeues),
 215                         URCU_TLS(nr_successful_dequeues));
 216         count[0] = URCU_TLS(nr_dequeues);
 217         count[1] = URCU_TLS(nr_successful_dequeues);
 218         return ((void*)2);
 219 }
 220
 221 static
 222 void test_end(unsigned long long *nr_dequeues_l)
 223 {
 224         struct cds_lfq_node_rcu *snode;
 225
 226         do {
 227                 snode = cds_lfq_dequeue_rcu(&q);
 228                 if (snode) {
 229                         struct test *node;
 230
 231                         node = caa_container_of(snode, struct test, list);
 232                         free(node);     /* no more concurrent access */
 233                         (*nr_dequeues_l)++;
 234                 }
 235         } while (snode);
 236 }
 237
 238 static
 239 void show_usage(char **argv)
 240 {
 241         printf("Usage : %s nr_dequeuers nr_enqueuers duration (s) <OPTIONS>\n",
 242                 argv[0]);
 243         printf("OPTIONS:\n");
 244         printf("        [-d delay] (enqueuer period (in loops))\n");
 245         printf("        [-c duration] (dequeuer period (in loops))\n");
 246         printf("        [-v] (verbose output)\n");
 247         printf("        [-a cpu#] [-a cpu#]... (affinity)\n");
 248         printf("\n");
 249 }
 250
 251 int main(int argc, char **argv)
 252 {
 253         int err;
 254         pthread_t *tid_enqueuer, *tid_dequeuer;
 255         void *tret;
 256         unsigned long long *count_enqueuer, *count_dequeuer;
 257         unsigned long long tot_enqueues = 0, tot_dequeues = 0;
 258         unsigned long long tot_successful_enqueues = 0,
 259                            tot_successful_dequeues = 0;
 260         unsigned long long end_dequeues = 0;
 261         int i, a;
 262         unsigned int i_thr;
 263
 264         if (argc < 4) {
 265                 show_usage(argv);
 266                 return -1;
 267         }
 268
 269         err = sscanf(argv[1], "%u", &nr_dequeuers);
 270         if (err != 1) {
 271                 show_usage(argv);
 272                 return -1;
 273         }
 274
 275         err = sscanf(argv[2], "%u", &nr_enqueuers);
 276         if (err != 1) {
 277                 show_usage(argv);
 278                 return -1;
 279         }
 280
 281         err = sscanf(argv[3], "%lu", &duration);
 282         if (err != 1) {
 283                 show_usage(argv);
 284                 return -1;
 285         }
 286
 287         for (i = 4; i < argc; i++) {
 288                 if (argv[i][0] != '-')
 289                         continue;
 290                 switch (argv[i][1]) {
 291                 case 'a':
 292                         if (argc < i + 2) {
 293                                 show_usage(argv);
 294                                 return -1;
 295                         }
 296                         a = atoi(argv[++i]);
 297                         cpu_affinities[next_aff++] = a;
 298                         use_affinity = 1;
 299                         printf_verbose("Adding CPU %d affinity\n", a);
 300                         break;
 301                 case 'c':
 302                         if (argc < i + 2) {
 303                                 show_usage(argv);
 304                                 return -1;
 305                         }
 306                         rduration = atol(argv[++i]);
 307                         break;
 308                 case 'd':
 309                         if (argc < i + 2) {
 310                                 show_usage(argv);
 311                                 return -1;
 312                         }
 313                         wdelay = atol(argv[++i]);
 314                         break;
 315                 case 'v':
 316                         verbose_mode = 1;
 317                         break;
 318                 }
 319         }
 320
 321         printf_verbose("running test for %lu seconds, %u enqueuers, "
 322                        "%u dequeuers.\n",
 323                        duration, nr_enqueuers, nr_dequeuers);
 324         printf_verbose("Writer delay : %lu loops.\n", rduration);
 325         printf_verbose("Reader duration : %lu loops.\n", wdelay);
 326         printf_verbose("thread %-6s, tid %lu\n",
 327                         "main", urcu_get_thread_id());
 328
 329         tid_enqueuer = calloc(nr_enqueuers, sizeof(*tid_enqueuer));
 330         tid_dequeuer = calloc(nr_dequeuers, sizeof(*tid_dequeuer));
 331         count_enqueuer = calloc(nr_enqueuers, 2 * sizeof(*count_enqueuer));
 332         count_dequeuer = calloc(nr_dequeuers, 2 * sizeof(*count_dequeuer));
 333         cds_lfq_init_rcu(&q, call_rcu);
 334         err = create_all_cpu_call_rcu_data(0);
 335         if (err) {
 336                 printf("Per-CPU call_rcu() worker threads unavailable. Using default global worker thread.\n");
 337         }
 338
 339         next_aff = 0;
 340
 341         for (i_thr = 0; i_thr < nr_enqueuers; i_thr++) {
 342                 err = pthread_create(&tid_enqueuer[i_thr], NULL, thr_enqueuer,
 343                                      &count_enqueuer[2 * i_thr]);
 344                 if (err != 0)
 345                         exit(1);
 346         }
 347         for (i_thr = 0; i_thr < nr_dequeuers; i_thr++) {
 348                 err = pthread_create(&tid_dequeuer[i_thr], NULL, thr_dequeuer,
 349                                      &count_dequeuer[2 * i_thr]);
 350                 if (err != 0)
 351                         exit(1);
 352         }
 353
 354         cmm_smp_mb();
 355
 356         begin_test();
 357
 358         for (i_thr = 0; i_thr < duration; i_thr++) {
 359                 sleep(1);
 360                 if (verbose_mode) {
 361                         fwrite(".", sizeof(char), 1, stdout);
 362                         fflush(stdout);
 363                 }
 364         }
 365
 366         end_test();
 367
 368         for (i_thr = 0; i_thr < nr_enqueuers; i_thr++) {
 369                 err = pthread_join(tid_enqueuer[i_thr], &tret);
 370                 if (err != 0)
 371                         exit(1);
 372                 tot_enqueues += count_enqueuer[2 * i_thr];
 373                 tot_successful_enqueues += count_enqueuer[2 * i_thr + 1];
 374         }
 375         for (i_thr = 0; i_thr < nr_dequeuers; i_thr++) {
 376                 err = pthread_join(tid_dequeuer[i_thr], &tret);
 377                 if (err != 0)
 378                         exit(1);
 379                 tot_dequeues += count_dequeuer[2 * i_thr];
 380                 tot_successful_dequeues += count_dequeuer[2 * i_thr + 1];
 381         }
 382
 383         test_end(&end_dequeues);
 384         err = cds_lfq_destroy_rcu(&q);
 385         urcu_posix_assert(!err);
 386
 387         printf_verbose("total number of enqueues : %llu, dequeues %llu\n",
 388                        tot_enqueues, tot_dequeues);
 389         printf_verbose("total number of successful enqueues : %llu, "
 390                        "successful dequeues %llu\n",
 391                        tot_successful_enqueues, tot_successful_dequeues);
 392         printf("SUMMARY %-25s testdur %4lu nr_enqueuers %3u wdelay %6lu "
 393                 "nr_dequeuers %3u "
 394                 "rdur %6lu nr_enqueues %12llu nr_dequeues %12llu "
 395                 "successful enqueues %12llu successful dequeues %12llu "
 396                 "end_dequeues %llu nr_ops %12llu\n",
 397                 argv[0], duration, nr_enqueuers, wdelay,
 398                 nr_dequeuers, rduration, tot_enqueues, tot_dequeues,
 399                 tot_successful_enqueues,
 400                 tot_successful_dequeues, end_dequeues,
 401                 tot_enqueues + tot_dequeues);
 402         if (tot_successful_enqueues != tot_successful_dequeues + end_dequeues)
 403                 printf("WARNING! Discrepancy between nr succ. enqueues %llu vs "
 404                        "succ. dequeues + end dequeues %llu.\n",
 405                        tot_successful_enqueues,
 406                        tot_successful_dequeues + end_dequeues);
 407
 408         free_all_cpu_call_rcu_data();
 409         free(count_enqueuer);
 410         free(count_dequeuer);
 411         free(tid_enqueuer);
 412         free(tid_dequeuer);
 413
 414         return 0;
 415 }