| 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 stack |
| 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 <urcu/uatomic.h> |
| 26 | #include "thread-id.h" |
| 27 | |
| 28 | /* hardcoded number of CPUs */ |
| 29 | #define NR_CPUS 16384 |
| 30 | |
| 31 | #ifndef DYNAMIC_LINK_TEST |
| 32 | #define _LGPL_SOURCE |
| 33 | #endif |
| 34 | #include <urcu/wfstack.h> |
| 35 | |
| 36 | /* |
| 37 | * External synchronization used. |
| 38 | */ |
| 39 | enum test_sync { |
| 40 | TEST_SYNC_NONE = 0, |
| 41 | TEST_SYNC_MUTEX, |
| 42 | }; |
| 43 | |
| 44 | static enum test_sync test_sync; |
| 45 | |
| 46 | static int test_force_sync; |
| 47 | |
| 48 | static volatile int test_go, test_stop_enqueue, test_stop_dequeue; |
| 49 | |
| 50 | static unsigned long rduration; |
| 51 | |
| 52 | static unsigned long duration; |
| 53 | |
| 54 | /* read-side C.S. duration, in loops */ |
| 55 | static unsigned long wdelay; |
| 56 | |
| 57 | static inline void loop_sleep(unsigned long loops) |
| 58 | { |
| 59 | while (loops-- != 0) |
| 60 | caa_cpu_relax(); |
| 61 | } |
| 62 | |
| 63 | static int verbose_mode; |
| 64 | |
| 65 | static int test_pop, test_pop_all, test_wait_empty; |
| 66 | static unsigned int test_enqueue_stopped; |
| 67 | |
| 68 | #define printf_verbose(fmt, args...) \ |
| 69 | do { \ |
| 70 | if (verbose_mode) \ |
| 71 | printf(fmt, ## args); \ |
| 72 | } while (0) |
| 73 | |
| 74 | static unsigned int cpu_affinities[NR_CPUS]; |
| 75 | static unsigned int next_aff = 0; |
| 76 | static int use_affinity = 0; |
| 77 | |
| 78 | pthread_mutex_t affinity_mutex = PTHREAD_MUTEX_INITIALIZER; |
| 79 | |
| 80 | static void set_affinity(void) |
| 81 | { |
| 82 | #ifdef HAVE_SCHED_SETAFFINITY |
| 83 | cpu_set_t mask; |
| 84 | int cpu, ret; |
| 85 | #endif /* HAVE_SCHED_SETAFFINITY */ |
| 86 | |
| 87 | if (!use_affinity) |
| 88 | return; |
| 89 | |
| 90 | #ifdef HAVE_SCHED_SETAFFINITY |
| 91 | ret = pthread_mutex_lock(&affinity_mutex); |
| 92 | if (ret) { |
| 93 | perror("Error in pthread mutex lock"); |
| 94 | exit(-1); |
| 95 | } |
| 96 | cpu = cpu_affinities[next_aff++]; |
| 97 | ret = pthread_mutex_unlock(&affinity_mutex); |
| 98 | if (ret) { |
| 99 | perror("Error in pthread mutex unlock"); |
| 100 | exit(-1); |
| 101 | } |
| 102 | |
| 103 | CPU_ZERO(&mask); |
| 104 | CPU_SET(cpu, &mask); |
| 105 | sched_setaffinity(0, sizeof(mask), &mask); |
| 106 | #endif /* HAVE_SCHED_SETAFFINITY */ |
| 107 | } |
| 108 | |
| 109 | /* |
| 110 | * returns 0 if test should end. |
| 111 | */ |
| 112 | static int test_duration_dequeue(void) |
| 113 | { |
| 114 | return !test_stop_dequeue; |
| 115 | } |
| 116 | |
| 117 | static int test_duration_enqueue(void) |
| 118 | { |
| 119 | return !test_stop_enqueue; |
| 120 | } |
| 121 | |
| 122 | static DEFINE_URCU_TLS(unsigned long long, nr_dequeues); |
| 123 | static DEFINE_URCU_TLS(unsigned long long, nr_enqueues); |
| 124 | |
| 125 | static DEFINE_URCU_TLS(unsigned long long, nr_successful_dequeues); |
| 126 | static DEFINE_URCU_TLS(unsigned long long, nr_successful_enqueues); |
| 127 | static DEFINE_URCU_TLS(unsigned long long, nr_empty_dest_enqueues); |
| 128 | static DEFINE_URCU_TLS(unsigned long long, nr_pop_all); |
| 129 | static DEFINE_URCU_TLS(unsigned long long, nr_pop_last); |
| 130 | |
| 131 | static unsigned int nr_enqueuers; |
| 132 | static unsigned int nr_dequeuers; |
| 133 | |
| 134 | static struct cds_wfs_stack s; |
| 135 | |
| 136 | static void *thr_enqueuer(void *_count) |
| 137 | { |
| 138 | unsigned long long *count = _count; |
| 139 | bool was_nonempty; |
| 140 | |
| 141 | printf_verbose("thread_begin %s, tid %lu\n", |
| 142 | "enqueuer", urcu_get_thread_id()); |
| 143 | |
| 144 | set_affinity(); |
| 145 | |
| 146 | while (!test_go) |
| 147 | { |
| 148 | } |
| 149 | cmm_smp_mb(); |
| 150 | |
| 151 | for (;;) { |
| 152 | struct cds_wfs_node *node = malloc(sizeof(*node)); |
| 153 | if (!node) |
| 154 | goto fail; |
| 155 | cds_wfs_node_init(node); |
| 156 | was_nonempty = cds_wfs_push(&s, node); |
| 157 | URCU_TLS(nr_successful_enqueues)++; |
| 158 | if (!was_nonempty) |
| 159 | URCU_TLS(nr_empty_dest_enqueues)++; |
| 160 | |
| 161 | if (caa_unlikely(wdelay)) |
| 162 | loop_sleep(wdelay); |
| 163 | fail: |
| 164 | URCU_TLS(nr_enqueues)++; |
| 165 | if (caa_unlikely(!test_duration_enqueue())) |
| 166 | break; |
| 167 | } |
| 168 | |
| 169 | uatomic_inc(&test_enqueue_stopped); |
| 170 | count[0] = URCU_TLS(nr_enqueues); |
| 171 | count[1] = URCU_TLS(nr_successful_enqueues); |
| 172 | count[2] = URCU_TLS(nr_empty_dest_enqueues); |
| 173 | printf_verbose("enqueuer thread_end, tid %lu, " |
| 174 | "enqueues %llu successful_enqueues %llu, " |
| 175 | "empty_dest_enqueues %llu\n", |
| 176 | urcu_get_thread_id(), |
| 177 | URCU_TLS(nr_enqueues), |
| 178 | URCU_TLS(nr_successful_enqueues), |
| 179 | URCU_TLS(nr_empty_dest_enqueues)); |
| 180 | return ((void*)1); |
| 181 | |
| 182 | } |
| 183 | |
| 184 | static void do_test_pop(enum test_sync sync) |
| 185 | { |
| 186 | struct cds_wfs_node *node; |
| 187 | int state; |
| 188 | |
| 189 | if (sync == TEST_SYNC_MUTEX) |
| 190 | cds_wfs_pop_lock(&s); |
| 191 | node = __cds_wfs_pop_with_state_blocking(&s, &state); |
| 192 | if (sync == TEST_SYNC_MUTEX) |
| 193 | cds_wfs_pop_unlock(&s); |
| 194 | |
| 195 | if (node) { |
| 196 | if (state & CDS_WFS_STATE_LAST) |
| 197 | URCU_TLS(nr_pop_last)++; |
| 198 | free(node); |
| 199 | URCU_TLS(nr_successful_dequeues)++; |
| 200 | } |
| 201 | URCU_TLS(nr_dequeues)++; |
| 202 | } |
| 203 | |
| 204 | static void do_test_pop_all(enum test_sync sync) |
| 205 | { |
| 206 | struct cds_wfs_head *head; |
| 207 | struct cds_wfs_node *node, *n; |
| 208 | |
| 209 | if (sync == TEST_SYNC_MUTEX) |
| 210 | cds_wfs_pop_lock(&s); |
| 211 | head = __cds_wfs_pop_all(&s); |
| 212 | if (sync == TEST_SYNC_MUTEX) |
| 213 | cds_wfs_pop_unlock(&s); |
| 214 | |
| 215 | /* Check if empty */ |
| 216 | if (cds_wfs_first(head) == NULL) |
| 217 | return; |
| 218 | |
| 219 | URCU_TLS(nr_pop_all)++; |
| 220 | URCU_TLS(nr_pop_last)++; |
| 221 | |
| 222 | cds_wfs_for_each_blocking_safe(head, node, n) { |
| 223 | free(node); |
| 224 | URCU_TLS(nr_successful_dequeues)++; |
| 225 | URCU_TLS(nr_dequeues)++; |
| 226 | } |
| 227 | } |
| 228 | |
| 229 | static void *thr_dequeuer(void *_count) |
| 230 | { |
| 231 | unsigned long long *count = _count; |
| 232 | unsigned int counter = 0; |
| 233 | |
| 234 | printf_verbose("thread_begin %s, tid %lu\n", |
| 235 | "dequeuer", urcu_get_thread_id()); |
| 236 | |
| 237 | set_affinity(); |
| 238 | |
| 239 | while (!test_go) |
| 240 | { |
| 241 | } |
| 242 | cmm_smp_mb(); |
| 243 | |
| 244 | urcu_posix_assert(test_pop || test_pop_all); |
| 245 | |
| 246 | for (;;) { |
| 247 | if (test_pop && test_pop_all) { |
| 248 | if (counter & 1) |
| 249 | do_test_pop(test_sync); |
| 250 | else |
| 251 | do_test_pop_all(test_sync); |
| 252 | counter++; |
| 253 | } else { |
| 254 | if (test_pop) |
| 255 | do_test_pop(test_sync); |
| 256 | else |
| 257 | do_test_pop_all(test_sync); |
| 258 | } |
| 259 | |
| 260 | if (caa_unlikely(!test_duration_dequeue())) |
| 261 | break; |
| 262 | if (caa_unlikely(rduration)) |
| 263 | loop_sleep(rduration); |
| 264 | } |
| 265 | |
| 266 | printf_verbose("dequeuer thread_end, tid %lu, " |
| 267 | "dequeues %llu, successful_dequeues %llu " |
| 268 | "pop_all %llu pop_last %llu\n", |
| 269 | urcu_get_thread_id(), |
| 270 | URCU_TLS(nr_dequeues), URCU_TLS(nr_successful_dequeues), |
| 271 | URCU_TLS(nr_pop_all), |
| 272 | URCU_TLS(nr_pop_last)); |
| 273 | count[0] = URCU_TLS(nr_dequeues); |
| 274 | count[1] = URCU_TLS(nr_successful_dequeues); |
| 275 | count[2] = URCU_TLS(nr_pop_all); |
| 276 | count[3] = URCU_TLS(nr_pop_last); |
| 277 | return ((void*)2); |
| 278 | } |
| 279 | |
| 280 | static void test_end(unsigned long long *nr_dequeues_l, |
| 281 | unsigned long long *nr_pop_last_l) |
| 282 | { |
| 283 | struct cds_wfs_node *node; |
| 284 | int state; |
| 285 | |
| 286 | do { |
| 287 | node = cds_wfs_pop_with_state_blocking(&s, &state); |
| 288 | if (node) { |
| 289 | if (state & CDS_WFS_STATE_LAST) |
| 290 | (*nr_pop_last_l)++; |
| 291 | free(node); |
| 292 | (*nr_dequeues_l)++; |
| 293 | } |
| 294 | } while (node); |
| 295 | } |
| 296 | |
| 297 | static void show_usage(char **argv) |
| 298 | { |
| 299 | printf("Usage : %s nr_dequeuers nr_enqueuers duration (s) <OPTIONS>\n", |
| 300 | argv[0]); |
| 301 | printf("OPTIONS:\n"); |
| 302 | printf(" [-d delay] (enqueuer period (in loops))\n"); |
| 303 | printf(" [-c duration] (dequeuer period (in loops))\n"); |
| 304 | printf(" [-v] (verbose output)\n"); |
| 305 | printf(" [-a cpu#] [-a cpu#]... (affinity)\n"); |
| 306 | printf(" [-p] (test pop)\n"); |
| 307 | printf(" [-P] (test pop_all, enabled by default)\n"); |
| 308 | printf(" [-M] (use mutex external synchronization)\n"); |
| 309 | printf(" Note: default: no external synchronization used.\n"); |
| 310 | printf(" [-f] (force user-provided synchronization)\n"); |
| 311 | printf(" [-w] Wait for dequeuer to empty stack\n"); |
| 312 | printf("\n"); |
| 313 | } |
| 314 | |
| 315 | int main(int argc, char **argv) |
| 316 | { |
| 317 | int err; |
| 318 | pthread_t *tid_enqueuer, *tid_dequeuer; |
| 319 | void *tret; |
| 320 | unsigned long long *count_enqueuer, *count_dequeuer; |
| 321 | unsigned long long tot_enqueues = 0, tot_dequeues = 0; |
| 322 | unsigned long long tot_successful_enqueues = 0, |
| 323 | tot_successful_dequeues = 0, |
| 324 | tot_empty_dest_enqueues = 0, |
| 325 | tot_pop_all = 0, tot_pop_last = 0; |
| 326 | unsigned long long end_dequeues = 0; |
| 327 | int i, a, retval = 0; |
| 328 | unsigned int i_thr; |
| 329 | |
| 330 | if (argc < 4) { |
| 331 | show_usage(argv); |
| 332 | return -1; |
| 333 | } |
| 334 | |
| 335 | err = sscanf(argv[1], "%u", &nr_dequeuers); |
| 336 | if (err != 1) { |
| 337 | show_usage(argv); |
| 338 | return -1; |
| 339 | } |
| 340 | |
| 341 | err = sscanf(argv[2], "%u", &nr_enqueuers); |
| 342 | if (err != 1) { |
| 343 | show_usage(argv); |
| 344 | return -1; |
| 345 | } |
| 346 | |
| 347 | err = sscanf(argv[3], "%lu", &duration); |
| 348 | if (err != 1) { |
| 349 | show_usage(argv); |
| 350 | return -1; |
| 351 | } |
| 352 | |
| 353 | for (i = 4; i < argc; i++) { |
| 354 | if (argv[i][0] != '-') |
| 355 | continue; |
| 356 | switch (argv[i][1]) { |
| 357 | case 'a': |
| 358 | if (argc < i + 2) { |
| 359 | show_usage(argv); |
| 360 | return -1; |
| 361 | } |
| 362 | a = atoi(argv[++i]); |
| 363 | cpu_affinities[next_aff++] = a; |
| 364 | use_affinity = 1; |
| 365 | printf_verbose("Adding CPU %d affinity\n", a); |
| 366 | break; |
| 367 | case 'c': |
| 368 | if (argc < i + 2) { |
| 369 | show_usage(argv); |
| 370 | return -1; |
| 371 | } |
| 372 | rduration = atol(argv[++i]); |
| 373 | break; |
| 374 | case 'd': |
| 375 | if (argc < i + 2) { |
| 376 | show_usage(argv); |
| 377 | return -1; |
| 378 | } |
| 379 | wdelay = atol(argv[++i]); |
| 380 | break; |
| 381 | case 'v': |
| 382 | verbose_mode = 1; |
| 383 | break; |
| 384 | case 'p': |
| 385 | test_pop = 1; |
| 386 | break; |
| 387 | case 'P': |
| 388 | test_pop_all = 1; |
| 389 | break; |
| 390 | case 'M': |
| 391 | test_sync = TEST_SYNC_MUTEX; |
| 392 | break; |
| 393 | case 'w': |
| 394 | test_wait_empty = 1; |
| 395 | break; |
| 396 | case 'f': |
| 397 | test_force_sync = 1; |
| 398 | break; |
| 399 | } |
| 400 | } |
| 401 | |
| 402 | /* activate pop_all test by default */ |
| 403 | if (!test_pop && !test_pop_all) |
| 404 | test_pop_all = 1; |
| 405 | |
| 406 | if (test_sync == TEST_SYNC_NONE && nr_dequeuers > 1 && test_pop) { |
| 407 | if (test_force_sync) { |
| 408 | fprintf(stderr, "[WARNING] Using pop concurrently " |
| 409 | "with other pop or pop_all without external " |
| 410 | "synchronization. Expect run-time failure.\n"); |
| 411 | } else { |
| 412 | printf("Enforcing mutex synchronization\n"); |
| 413 | test_sync = TEST_SYNC_MUTEX; |
| 414 | } |
| 415 | } |
| 416 | |
| 417 | printf_verbose("running test for %lu seconds, %u enqueuers, " |
| 418 | "%u dequeuers.\n", |
| 419 | duration, nr_enqueuers, nr_dequeuers); |
| 420 | if (test_pop) |
| 421 | printf_verbose("pop test activated.\n"); |
| 422 | if (test_pop_all) |
| 423 | printf_verbose("pop_all test activated.\n"); |
| 424 | if (test_sync == TEST_SYNC_MUTEX) |
| 425 | printf_verbose("External sync: mutex.\n"); |
| 426 | else |
| 427 | printf_verbose("External sync: none.\n"); |
| 428 | if (test_wait_empty) |
| 429 | printf_verbose("Wait for dequeuers to empty stack.\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, 3 * sizeof(*count_enqueuer)); |
| 438 | count_dequeuer = calloc(nr_dequeuers, 4 * sizeof(*count_dequeuer)); |
| 439 | cds_wfs_init(&s); |
| 440 | |
| 441 | next_aff = 0; |
| 442 | |
| 443 | for (i_thr = 0; i_thr < nr_enqueuers; i_thr++) { |
| 444 | err = pthread_create(&tid_enqueuer[i], NULL, thr_enqueuer, |
| 445 | &count_enqueuer[3 * i_thr]); |
| 446 | if (err != 0) |
| 447 | exit(1); |
| 448 | } |
| 449 | for (i_thr = 0; i_thr < nr_dequeuers; i_thr++) { |
| 450 | err = pthread_create(&tid_dequeuer[i_thr], NULL, thr_dequeuer, |
| 451 | &count_dequeuer[4 * i_thr]); |
| 452 | if (err != 0) |
| 453 | exit(1); |
| 454 | } |
| 455 | |
| 456 | cmm_smp_mb(); |
| 457 | |
| 458 | test_go = 1; |
| 459 | |
| 460 | for (i_thr = 0; i_thr < duration; i_thr++) { |
| 461 | sleep(1); |
| 462 | if (verbose_mode) { |
| 463 | fwrite(".", sizeof(char), 1, stdout); |
| 464 | fflush(stdout); |
| 465 | } |
| 466 | } |
| 467 | |
| 468 | test_stop_enqueue = 1; |
| 469 | |
| 470 | if (test_wait_empty) { |
| 471 | while (nr_enqueuers != uatomic_read(&test_enqueue_stopped)) { |
| 472 | sleep(1); |
| 473 | } |
| 474 | while (!cds_wfs_empty(&s)) { |
| 475 | sleep(1); |
| 476 | } |
| 477 | } |
| 478 | |
| 479 | test_stop_dequeue = 1; |
| 480 | |
| 481 | for (i_thr = 0; i_thr < nr_enqueuers; i_thr++) { |
| 482 | err = pthread_join(tid_enqueuer[i_thr], &tret); |
| 483 | if (err != 0) |
| 484 | exit(1); |
| 485 | tot_enqueues += count_enqueuer[3 * i_thr]; |
| 486 | tot_successful_enqueues += count_enqueuer[3 * i_thr + 1]; |
| 487 | tot_empty_dest_enqueues += count_enqueuer[3 * i_thr + 2]; |
| 488 | } |
| 489 | for (i_thr = 0; i_thr < nr_dequeuers; i_thr++) { |
| 490 | err = pthread_join(tid_dequeuer[i_thr], &tret); |
| 491 | if (err != 0) |
| 492 | exit(1); |
| 493 | tot_dequeues += count_dequeuer[4 * i_thr]; |
| 494 | tot_successful_dequeues += count_dequeuer[4 * i_thr + 1]; |
| 495 | tot_pop_all += count_dequeuer[4 * i_thr + 2]; |
| 496 | tot_pop_last += count_dequeuer[4 * i_thr + 3]; |
| 497 | } |
| 498 | |
| 499 | test_end(&end_dequeues, &tot_pop_last); |
| 500 | |
| 501 | printf_verbose("total number of enqueues : %llu, dequeues %llu\n", |
| 502 | tot_enqueues, tot_dequeues); |
| 503 | printf_verbose("total number of successful enqueues : %llu, " |
| 504 | "enqueues to empty dest : %llu, " |
| 505 | "successful dequeues %llu, " |
| 506 | "pop_all : %llu, pop_last : %llu\n", |
| 507 | tot_successful_enqueues, |
| 508 | tot_empty_dest_enqueues, |
| 509 | tot_successful_dequeues, |
| 510 | tot_pop_all, tot_pop_last); |
| 511 | printf("SUMMARY %-25s testdur %4lu nr_enqueuers %3u wdelay %6lu " |
| 512 | "nr_dequeuers %3u " |
| 513 | "rdur %6lu nr_enqueues %12llu nr_dequeues %12llu " |
| 514 | "successful enqueues %12llu enqueues to empty dest %12llu " |
| 515 | "successful dequeues %12llu pop_all %12llu " |
| 516 | "pop_last %llu end_dequeues %llu nr_ops %12llu\n", |
| 517 | argv[0], duration, nr_enqueuers, wdelay, |
| 518 | nr_dequeuers, rduration, tot_enqueues, tot_dequeues, |
| 519 | tot_successful_enqueues, |
| 520 | tot_empty_dest_enqueues, |
| 521 | tot_successful_dequeues, tot_pop_all, tot_pop_last, |
| 522 | end_dequeues, |
| 523 | tot_enqueues + tot_dequeues); |
| 524 | if (tot_successful_enqueues != tot_successful_dequeues + end_dequeues) { |
| 525 | printf("WARNING! Discrepancy between nr succ. enqueues %llu vs " |
| 526 | "succ. dequeues + end dequeues %llu.\n", |
| 527 | tot_successful_enqueues, |
| 528 | tot_successful_dequeues + end_dequeues); |
| 529 | retval = 1; |
| 530 | } |
| 531 | /* |
| 532 | * The enqueuer should see exactly as many empty queues than the |
| 533 | * number of non-empty stacks dequeued. |
| 534 | */ |
| 535 | if (tot_empty_dest_enqueues != tot_pop_last) { |
| 536 | printf("WARNING! Discrepancy between empty enqueue (%llu) and " |
| 537 | "number of pop last (%llu)\n", |
| 538 | tot_empty_dest_enqueues, |
| 539 | tot_pop_last); |
| 540 | retval = 1; |
| 541 | } |
| 542 | |
| 543 | cds_wfs_destroy(&s); |
| 544 | free(count_enqueuer); |
| 545 | free(count_dequeuer); |
| 546 | free(tid_enqueuer); |
| 547 | free(tid_dequeuer); |
| 548 | |
| 549 | return retval; |
| 550 | } |