1 // SPDX-License-Identifier: GPL-2.0 2 /* Copyright (c) 2020 Facebook */ 3 #define _GNU_SOURCE 4 #include <argp.h> 5 #include <linux/compiler.h> 6 #include <sys/time.h> 7 #include <sched.h> 8 #include <fcntl.h> 9 #include <pthread.h> 10 #include <sys/sysinfo.h> 11 #include <signal.h> 12 #include "bench.h" 13 #include "bpf_util.h" 14 #include "testing_helpers.h" 15 16 struct env env = { 17 .warmup_sec = 1, 18 .duration_sec = 5, 19 .affinity = false, 20 .quiet = false, 21 .consumer_cnt = 0, 22 .producer_cnt = 1, 23 }; 24 25 static int libbpf_print_fn(enum libbpf_print_level level, 26 const char *format, va_list args) 27 { 28 if (level == LIBBPF_DEBUG && !env.verbose) 29 return 0; 30 return vfprintf(stderr, format, args); 31 } 32 33 void setup_libbpf(void) 34 { 35 libbpf_set_strict_mode(LIBBPF_STRICT_ALL); 36 libbpf_set_print(libbpf_print_fn); 37 } 38 39 void false_hits_report_progress(int iter, struct bench_res *res, long delta_ns) 40 { 41 long total = res->false_hits + res->hits + res->drops; 42 43 printf("Iter %3d (%7.3lfus): ", 44 iter, (delta_ns - 1000000000) / 1000.0); 45 46 printf("%ld false hits of %ld total operations. Percentage = %2.2f %%\n", 47 res->false_hits, total, ((float)res->false_hits / total) * 100); 48 } 49 50 void false_hits_report_final(struct bench_res res[], int res_cnt) 51 { 52 long total_hits = 0, total_drops = 0, total_false_hits = 0, total_ops = 0; 53 int i; 54 55 for (i = 0; i < res_cnt; i++) { 56 total_hits += res[i].hits; 57 total_false_hits += res[i].false_hits; 58 total_drops += res[i].drops; 59 } 60 total_ops = total_hits + total_false_hits + total_drops; 61 62 printf("Summary: %ld false hits of %ld total operations. ", 63 total_false_hits, total_ops); 64 printf("Percentage = %2.2f %%\n", 65 ((float)total_false_hits / total_ops) * 100); 66 } 67 68 void hits_drops_report_progress(int iter, struct bench_res *res, long delta_ns) 69 { 70 double hits_per_sec, drops_per_sec; 71 double hits_per_prod; 72 73 hits_per_sec = res->hits / 1000000.0 / (delta_ns / 1000000000.0); 74 hits_per_prod = hits_per_sec / env.producer_cnt; 75 drops_per_sec = res->drops / 1000000.0 / (delta_ns / 1000000000.0); 76 77 printf("Iter %3d (%7.3lfus): ", 78 iter, (delta_ns - 1000000000) / 1000.0); 79 80 printf("hits %8.3lfM/s (%7.3lfM/prod), drops %8.3lfM/s, total operations %8.3lfM/s\n", 81 hits_per_sec, hits_per_prod, drops_per_sec, hits_per_sec + drops_per_sec); 82 } 83 84 void 85 grace_period_latency_basic_stats(struct bench_res res[], int res_cnt, struct basic_stats *gp_stat) 86 { 87 int i; 88 89 memset(gp_stat, 0, sizeof(struct basic_stats)); 90 91 for (i = 0; i < res_cnt; i++) 92 gp_stat->mean += res[i].gp_ns / 1000.0 / (double)res[i].gp_ct / (0.0 + res_cnt); 93 94 #define IT_MEAN_DIFF (res[i].gp_ns / 1000.0 / (double)res[i].gp_ct - gp_stat->mean) 95 if (res_cnt > 1) { 96 for (i = 0; i < res_cnt; i++) 97 gp_stat->stddev += (IT_MEAN_DIFF * IT_MEAN_DIFF) / (res_cnt - 1.0); 98 } 99 gp_stat->stddev = sqrt(gp_stat->stddev); 100 #undef IT_MEAN_DIFF 101 } 102 103 void 104 grace_period_ticks_basic_stats(struct bench_res res[], int res_cnt, struct basic_stats *gp_stat) 105 { 106 int i; 107 108 memset(gp_stat, 0, sizeof(struct basic_stats)); 109 for (i = 0; i < res_cnt; i++) 110 gp_stat->mean += res[i].stime / (double)res[i].gp_ct / (0.0 + res_cnt); 111 112 #define IT_MEAN_DIFF (res[i].stime / (double)res[i].gp_ct - gp_stat->mean) 113 if (res_cnt > 1) { 114 for (i = 0; i < res_cnt; i++) 115 gp_stat->stddev += (IT_MEAN_DIFF * IT_MEAN_DIFF) / (res_cnt - 1.0); 116 } 117 gp_stat->stddev = sqrt(gp_stat->stddev); 118 #undef IT_MEAN_DIFF 119 } 120 121 void hits_drops_report_final(struct bench_res res[], int res_cnt) 122 { 123 int i; 124 double hits_mean = 0.0, drops_mean = 0.0, total_ops_mean = 0.0; 125 double hits_stddev = 0.0, drops_stddev = 0.0, total_ops_stddev = 0.0; 126 double total_ops; 127 128 for (i = 0; i < res_cnt; i++) { 129 hits_mean += res[i].hits / 1000000.0 / (0.0 + res_cnt); 130 drops_mean += res[i].drops / 1000000.0 / (0.0 + res_cnt); 131 } 132 total_ops_mean = hits_mean + drops_mean; 133 134 if (res_cnt > 1) { 135 for (i = 0; i < res_cnt; i++) { 136 hits_stddev += (hits_mean - res[i].hits / 1000000.0) * 137 (hits_mean - res[i].hits / 1000000.0) / 138 (res_cnt - 1.0); 139 drops_stddev += (drops_mean - res[i].drops / 1000000.0) * 140 (drops_mean - res[i].drops / 1000000.0) / 141 (res_cnt - 1.0); 142 total_ops = res[i].hits + res[i].drops; 143 total_ops_stddev += (total_ops_mean - total_ops / 1000000.0) * 144 (total_ops_mean - total_ops / 1000000.0) / 145 (res_cnt - 1.0); 146 } 147 hits_stddev = sqrt(hits_stddev); 148 drops_stddev = sqrt(drops_stddev); 149 total_ops_stddev = sqrt(total_ops_stddev); 150 } 151 printf("Summary: hits %8.3lf \u00B1 %5.3lfM/s (%7.3lfM/prod), ", 152 hits_mean, hits_stddev, hits_mean / env.producer_cnt); 153 printf("drops %8.3lf \u00B1 %5.3lfM/s, ", 154 drops_mean, drops_stddev); 155 printf("total operations %8.3lf \u00B1 %5.3lfM/s\n", 156 total_ops_mean, total_ops_stddev); 157 } 158 159 void ops_report_progress(int iter, struct bench_res *res, long delta_ns) 160 { 161 double hits_per_sec, hits_per_prod; 162 163 hits_per_sec = res->hits / 1000000.0 / (delta_ns / 1000000000.0); 164 hits_per_prod = hits_per_sec / env.producer_cnt; 165 166 printf("Iter %3d (%7.3lfus): ", iter, (delta_ns - 1000000000) / 1000.0); 167 168 printf("hits %8.3lfM/s (%7.3lfM/prod)\n", hits_per_sec, hits_per_prod); 169 } 170 171 void ops_report_final(struct bench_res res[], int res_cnt) 172 { 173 double hits_mean = 0.0, hits_stddev = 0.0; 174 int i; 175 176 for (i = 0; i < res_cnt; i++) 177 hits_mean += res[i].hits / 1000000.0 / (0.0 + res_cnt); 178 179 if (res_cnt > 1) { 180 for (i = 0; i < res_cnt; i++) 181 hits_stddev += (hits_mean - res[i].hits / 1000000.0) * 182 (hits_mean - res[i].hits / 1000000.0) / 183 (res_cnt - 1.0); 184 185 hits_stddev = sqrt(hits_stddev); 186 } 187 printf("Summary: throughput %8.3lf \u00B1 %5.3lf M ops/s (%7.3lfM ops/prod), ", 188 hits_mean, hits_stddev, hits_mean / env.producer_cnt); 189 printf("latency %8.3lf ns/op\n", 1000.0 / hits_mean * env.producer_cnt); 190 } 191 192 void local_storage_report_progress(int iter, struct bench_res *res, 193 long delta_ns) 194 { 195 double important_hits_per_sec, hits_per_sec; 196 double delta_sec = delta_ns / 1000000000.0; 197 198 hits_per_sec = res->hits / 1000000.0 / delta_sec; 199 important_hits_per_sec = res->important_hits / 1000000.0 / delta_sec; 200 201 printf("Iter %3d (%7.3lfus): ", iter, (delta_ns - 1000000000) / 1000.0); 202 203 printf("hits %8.3lfM/s ", hits_per_sec); 204 printf("important_hits %8.3lfM/s\n", important_hits_per_sec); 205 } 206 207 void local_storage_report_final(struct bench_res res[], int res_cnt) 208 { 209 double important_hits_mean = 0.0, important_hits_stddev = 0.0; 210 double hits_mean = 0.0, hits_stddev = 0.0; 211 int i; 212 213 for (i = 0; i < res_cnt; i++) { 214 hits_mean += res[i].hits / 1000000.0 / (0.0 + res_cnt); 215 important_hits_mean += res[i].important_hits / 1000000.0 / (0.0 + res_cnt); 216 } 217 218 if (res_cnt > 1) { 219 for (i = 0; i < res_cnt; i++) { 220 hits_stddev += (hits_mean - res[i].hits / 1000000.0) * 221 (hits_mean - res[i].hits / 1000000.0) / 222 (res_cnt - 1.0); 223 important_hits_stddev += 224 (important_hits_mean - res[i].important_hits / 1000000.0) * 225 (important_hits_mean - res[i].important_hits / 1000000.0) / 226 (res_cnt - 1.0); 227 } 228 229 hits_stddev = sqrt(hits_stddev); 230 important_hits_stddev = sqrt(important_hits_stddev); 231 } 232 printf("Summary: hits throughput %8.3lf \u00B1 %5.3lf M ops/s, ", 233 hits_mean, hits_stddev); 234 printf("hits latency %8.3lf ns/op, ", 1000.0 / hits_mean); 235 printf("important_hits throughput %8.3lf \u00B1 %5.3lf M ops/s\n", 236 important_hits_mean, important_hits_stddev); 237 } 238 239 const char *argp_program_version = "benchmark"; 240 const char *argp_program_bug_address = "<bpf@vger.kernel.org>"; 241 const char argp_program_doc[] = 242 "benchmark Generic benchmarking framework.\n" 243 "\n" 244 "This tool runs benchmarks.\n" 245 "\n" 246 "USAGE: benchmark <bench-name>\n" 247 "\n" 248 "EXAMPLES:\n" 249 " # run 'count-local' benchmark with 1 producer and 1 consumer\n" 250 " benchmark count-local\n" 251 " # run 'count-local' with 16 producer and 8 consumer thread, pinned to CPUs\n" 252 " benchmark -p16 -c8 -a count-local\n"; 253 254 enum { 255 ARG_PROD_AFFINITY_SET = 1000, 256 ARG_CONS_AFFINITY_SET = 1001, 257 }; 258 259 static const struct argp_option opts[] = { 260 { "list", 'l', NULL, 0, "List available benchmarks"}, 261 { "duration", 'd', "SEC", 0, "Duration of benchmark, seconds"}, 262 { "warmup", 'w', "SEC", 0, "Warm-up period, seconds"}, 263 { "producers", 'p', "NUM", 0, "Number of producer threads"}, 264 { "consumers", 'c', "NUM", 0, "Number of consumer threads"}, 265 { "verbose", 'v', NULL, 0, "Verbose debug output"}, 266 { "affinity", 'a', NULL, 0, "Set consumer/producer thread affinity"}, 267 { "quiet", 'q', NULL, 0, "Be more quiet"}, 268 { "stacktrace", 's', NULL, 0, "Get stack trace"}, 269 { "prod-affinity", ARG_PROD_AFFINITY_SET, "CPUSET", 0, 270 "Set of CPUs for producer threads; implies --affinity"}, 271 { "cons-affinity", ARG_CONS_AFFINITY_SET, "CPUSET", 0, 272 "Set of CPUs for consumer threads; implies --affinity"}, 273 {}, 274 }; 275 276 extern struct argp bench_ringbufs_argp; 277 extern struct argp bench_bloom_map_argp; 278 extern struct argp bench_bpf_loop_argp; 279 extern struct argp bench_local_storage_argp; 280 extern struct argp bench_local_storage_rcu_tasks_trace_argp; 281 extern struct argp bench_strncmp_argp; 282 extern struct argp bench_hashmap_lookup_argp; 283 extern struct argp bench_local_storage_create_argp; 284 extern struct argp bench_htab_mem_argp; 285 extern struct argp bench_trigger_batch_argp; 286 extern struct argp bench_crypto_argp; 287 extern struct argp bench_sockmap_argp; 288 extern struct argp bench_lpm_trie_map_argp; 289 extern struct argp bench_xdp_lb_argp; 290 291 static const struct argp_child bench_parsers[] = { 292 { &bench_ringbufs_argp, 0, "Ring buffers benchmark", 0 }, 293 { &bench_bloom_map_argp, 0, "Bloom filter map benchmark", 0 }, 294 { &bench_bpf_loop_argp, 0, "bpf_loop helper benchmark", 0 }, 295 { &bench_local_storage_argp, 0, "local_storage benchmark", 0 }, 296 { &bench_strncmp_argp, 0, "bpf_strncmp helper benchmark", 0 }, 297 { &bench_local_storage_rcu_tasks_trace_argp, 0, 298 "local_storage RCU Tasks Trace slowdown benchmark", 0 }, 299 { &bench_hashmap_lookup_argp, 0, "Hashmap lookup benchmark", 0 }, 300 { &bench_local_storage_create_argp, 0, "local-storage-create benchmark", 0 }, 301 { &bench_htab_mem_argp, 0, "hash map memory benchmark", 0 }, 302 { &bench_trigger_batch_argp, 0, "BPF triggering benchmark", 0 }, 303 { &bench_crypto_argp, 0, "bpf crypto benchmark", 0 }, 304 { &bench_sockmap_argp, 0, "bpf sockmap benchmark", 0 }, 305 { &bench_lpm_trie_map_argp, 0, "LPM trie map benchmark", 0 }, 306 { &bench_xdp_lb_argp, 0, "XDP load-balancer benchmark", 0 }, 307 {}, 308 }; 309 310 /* Make pos_args global, so that we can run argp_parse twice, if necessary */ 311 static int pos_args; 312 313 static error_t parse_arg(int key, char *arg, struct argp_state *state) 314 { 315 switch (key) { 316 case 'v': 317 env.verbose = true; 318 break; 319 case 'l': 320 env.list = true; 321 break; 322 case 'd': 323 env.duration_sec = strtol(arg, NULL, 10); 324 if (env.duration_sec <= 0) { 325 fprintf(stderr, "Invalid duration: %s\n", arg); 326 argp_usage(state); 327 } 328 break; 329 case 'w': 330 env.warmup_sec = strtol(arg, NULL, 10); 331 if (env.warmup_sec <= 0) { 332 fprintf(stderr, "Invalid warm-up duration: %s\n", arg); 333 argp_usage(state); 334 } 335 break; 336 case 'p': 337 env.producer_cnt = strtol(arg, NULL, 10); 338 if (env.producer_cnt < 0) { 339 fprintf(stderr, "Invalid producer count: %s\n", arg); 340 argp_usage(state); 341 } 342 break; 343 case 'c': 344 env.consumer_cnt = strtol(arg, NULL, 10); 345 if (env.consumer_cnt < 0) { 346 fprintf(stderr, "Invalid consumer count: %s\n", arg); 347 argp_usage(state); 348 } 349 break; 350 case 'a': 351 env.affinity = true; 352 break; 353 case 'q': 354 env.quiet = true; 355 break; 356 case 's': 357 env.stacktrace = true; 358 break; 359 case ARG_PROD_AFFINITY_SET: 360 env.affinity = true; 361 if (parse_num_list(arg, &env.prod_cpus.cpus, 362 &env.prod_cpus.cpus_len)) { 363 fprintf(stderr, "Invalid format of CPU set for producers."); 364 argp_usage(state); 365 } 366 break; 367 case ARG_CONS_AFFINITY_SET: 368 env.affinity = true; 369 if (parse_num_list(arg, &env.cons_cpus.cpus, 370 &env.cons_cpus.cpus_len)) { 371 fprintf(stderr, "Invalid format of CPU set for consumers."); 372 argp_usage(state); 373 } 374 break; 375 case ARGP_KEY_ARG: 376 if (pos_args++) { 377 fprintf(stderr, 378 "Unrecognized positional argument: %s\n", arg); 379 argp_usage(state); 380 } 381 env.bench_name = strdup(arg); 382 break; 383 default: 384 return ARGP_ERR_UNKNOWN; 385 } 386 return 0; 387 } 388 389 static void parse_cmdline_args_init(int argc, char **argv) 390 { 391 static const struct argp argp = { 392 .options = opts, 393 .parser = parse_arg, 394 .doc = argp_program_doc, 395 .children = bench_parsers, 396 }; 397 if (argp_parse(&argp, argc, argv, 0, NULL, NULL)) 398 exit(1); 399 } 400 401 static void parse_cmdline_args_final(int argc, char **argv) 402 { 403 struct argp_child bench_parsers[2] = {}; 404 const struct argp argp = { 405 .options = opts, 406 .parser = parse_arg, 407 .doc = argp_program_doc, 408 .children = bench_parsers, 409 }; 410 411 /* Parse arguments the second time with the correct set of parsers */ 412 if (bench->argp) { 413 bench_parsers[0].argp = bench->argp; 414 bench_parsers[0].header = bench->name; 415 pos_args = 0; 416 if (argp_parse(&argp, argc, argv, 0, NULL, NULL)) 417 exit(1); 418 } 419 } 420 421 static void collect_measurements(long delta_ns); 422 423 static __u64 last_time_ns; 424 static void sigalarm_handler(int signo) 425 { 426 long new_time_ns = get_time_ns(); 427 long delta_ns = new_time_ns - last_time_ns; 428 429 collect_measurements(delta_ns); 430 431 last_time_ns = new_time_ns; 432 } 433 434 /* set up periodic 1-second timer */ 435 static void setup_timer() 436 { 437 static struct sigaction sigalarm_action = { 438 .sa_handler = sigalarm_handler, 439 }; 440 struct itimerval timer_settings = {}; 441 int err; 442 443 last_time_ns = get_time_ns(); 444 err = sigaction(SIGALRM, &sigalarm_action, NULL); 445 if (err < 0) { 446 fprintf(stderr, "failed to install SIGALRM handler: %d\n", -errno); 447 exit(1); 448 } 449 timer_settings.it_interval.tv_sec = 1; 450 timer_settings.it_value.tv_sec = 1; 451 err = setitimer(ITIMER_REAL, &timer_settings, NULL); 452 if (err < 0) { 453 fprintf(stderr, "failed to arm interval timer: %d\n", -errno); 454 exit(1); 455 } 456 } 457 458 static void set_thread_affinity(pthread_t thread, int cpu) 459 { 460 cpu_set_t cpuset; 461 int err; 462 463 CPU_ZERO(&cpuset); 464 CPU_SET(cpu, &cpuset); 465 err = pthread_setaffinity_np(thread, sizeof(cpuset), &cpuset); 466 if (err) { 467 fprintf(stderr, "setting affinity to CPU #%d failed: %d\n", 468 cpu, -err); 469 exit(1); 470 } 471 } 472 473 static int next_cpu(struct cpu_set *cpu_set) 474 { 475 if (cpu_set->cpus) { 476 int i; 477 478 /* find next available CPU */ 479 for (i = cpu_set->next_cpu; i < cpu_set->cpus_len; i++) { 480 if (cpu_set->cpus[i]) { 481 cpu_set->next_cpu = i + 1; 482 return i; 483 } 484 } 485 fprintf(stderr, "Not enough CPUs specified, need CPU #%d or higher.\n", i); 486 exit(1); 487 } 488 489 return cpu_set->next_cpu++ % env.nr_cpus; 490 } 491 492 static struct bench_state { 493 int res_cnt; 494 struct bench_res *results; 495 pthread_t *consumers; 496 pthread_t *producers; 497 } state; 498 499 const struct bench *bench = NULL; 500 501 extern const struct bench bench_count_global; 502 extern const struct bench bench_count_local; 503 extern const struct bench bench_rename_base; 504 extern const struct bench bench_rename_kprobe; 505 extern const struct bench bench_rename_kretprobe; 506 extern const struct bench bench_rename_rawtp; 507 extern const struct bench bench_rename_fentry; 508 extern const struct bench bench_rename_fexit; 509 510 /* pure counting benchmarks to establish theoretical limits */ 511 extern const struct bench bench_trig_usermode_count; 512 extern const struct bench bench_trig_syscall_count; 513 extern const struct bench bench_trig_kernel_count; 514 515 /* batched, staying mostly in-kernel benchmarks */ 516 extern const struct bench bench_trig_kprobe; 517 extern const struct bench bench_trig_kretprobe; 518 extern const struct bench bench_trig_kprobe_multi; 519 extern const struct bench bench_trig_kretprobe_multi; 520 extern const struct bench bench_trig_fentry; 521 extern const struct bench bench_trig_kprobe_multi_all; 522 extern const struct bench bench_trig_kretprobe_multi_all; 523 extern const struct bench bench_trig_fexit; 524 extern const struct bench bench_trig_fmodret; 525 extern const struct bench bench_trig_tp; 526 extern const struct bench bench_trig_rawtp; 527 528 /* uprobe/uretprobe benchmarks */ 529 extern const struct bench bench_trig_uprobe_nop; 530 extern const struct bench bench_trig_uretprobe_nop; 531 extern const struct bench bench_trig_uprobe_push; 532 extern const struct bench bench_trig_uretprobe_push; 533 extern const struct bench bench_trig_uprobe_ret; 534 extern const struct bench bench_trig_uretprobe_ret; 535 extern const struct bench bench_trig_uprobe_multi_nop; 536 extern const struct bench bench_trig_uretprobe_multi_nop; 537 extern const struct bench bench_trig_uprobe_multi_push; 538 extern const struct bench bench_trig_uretprobe_multi_push; 539 extern const struct bench bench_trig_uprobe_multi_ret; 540 extern const struct bench bench_trig_uretprobe_multi_ret; 541 #ifdef __x86_64__ 542 extern const struct bench bench_trig_uprobe_nop5; 543 extern const struct bench bench_trig_uretprobe_nop5; 544 extern const struct bench bench_trig_uprobe_multi_nop5; 545 extern const struct bench bench_trig_uretprobe_multi_nop5; 546 extern const struct bench bench_trig_usdt_nop; 547 extern const struct bench bench_trig_usdt_nop5; 548 #endif 549 550 extern const struct bench bench_rb_libbpf; 551 extern const struct bench bench_rb_custom; 552 extern const struct bench bench_pb_libbpf; 553 extern const struct bench bench_pb_custom; 554 extern const struct bench bench_bloom_lookup; 555 extern const struct bench bench_bloom_update; 556 extern const struct bench bench_bloom_false_positive; 557 extern const struct bench bench_hashmap_without_bloom; 558 extern const struct bench bench_hashmap_with_bloom; 559 extern const struct bench bench_bpf_loop; 560 extern const struct bench bench_strncmp_no_helper; 561 extern const struct bench bench_strncmp_helper; 562 extern const struct bench bench_bpf_hashmap_full_update; 563 extern const struct bench bench_bpf_rhashmap_full_update; 564 extern const struct bench bench_local_storage_cache_seq_get; 565 extern const struct bench bench_local_storage_cache_interleaved_get; 566 extern const struct bench bench_local_storage_cache_hashmap_control; 567 extern const struct bench bench_local_storage_tasks_trace; 568 extern const struct bench bench_bpf_hashmap_lookup; 569 extern const struct bench bench_bpf_rhashmap_lookup; 570 extern const struct bench bench_local_storage_create; 571 extern const struct bench bench_htab_mem; 572 extern const struct bench bench_rhtab_mem; 573 extern const struct bench bench_crypto_encrypt; 574 extern const struct bench bench_crypto_decrypt; 575 extern const struct bench bench_sockmap; 576 extern const struct bench bench_lpm_trie_noop; 577 extern const struct bench bench_lpm_trie_baseline; 578 extern const struct bench bench_lpm_trie_lookup; 579 extern const struct bench bench_lpm_trie_insert; 580 extern const struct bench bench_lpm_trie_update; 581 extern const struct bench bench_lpm_trie_delete; 582 extern const struct bench bench_lpm_trie_free; 583 extern const struct bench bench_bpf_nop; 584 extern const struct bench bench_xdp_lb; 585 586 static const struct bench *benchs[] = { 587 &bench_count_global, 588 &bench_count_local, 589 &bench_rename_base, 590 &bench_rename_kprobe, 591 &bench_rename_kretprobe, 592 &bench_rename_rawtp, 593 &bench_rename_fentry, 594 &bench_rename_fexit, 595 /* pure counting benchmarks for establishing theoretical limits */ 596 &bench_trig_usermode_count, 597 &bench_trig_kernel_count, 598 &bench_trig_syscall_count, 599 /* batched, staying mostly in-kernel triggers */ 600 &bench_trig_kprobe, 601 &bench_trig_kretprobe, 602 &bench_trig_kprobe_multi, 603 &bench_trig_kretprobe_multi, 604 &bench_trig_fentry, 605 &bench_trig_kprobe_multi_all, 606 &bench_trig_kretprobe_multi_all, 607 &bench_trig_fexit, 608 &bench_trig_fmodret, 609 &bench_trig_tp, 610 &bench_trig_rawtp, 611 /* uprobes */ 612 &bench_trig_uprobe_nop, 613 &bench_trig_uretprobe_nop, 614 &bench_trig_uprobe_push, 615 &bench_trig_uretprobe_push, 616 &bench_trig_uprobe_ret, 617 &bench_trig_uretprobe_ret, 618 &bench_trig_uprobe_multi_nop, 619 &bench_trig_uretprobe_multi_nop, 620 &bench_trig_uprobe_multi_push, 621 &bench_trig_uretprobe_multi_push, 622 &bench_trig_uprobe_multi_ret, 623 &bench_trig_uretprobe_multi_ret, 624 #ifdef __x86_64__ 625 &bench_trig_uprobe_nop5, 626 &bench_trig_uretprobe_nop5, 627 &bench_trig_uprobe_multi_nop5, 628 &bench_trig_uretprobe_multi_nop5, 629 &bench_trig_usdt_nop, 630 &bench_trig_usdt_nop5, 631 #endif 632 /* ringbuf/perfbuf benchmarks */ 633 &bench_rb_libbpf, 634 &bench_rb_custom, 635 &bench_pb_libbpf, 636 &bench_pb_custom, 637 &bench_bloom_lookup, 638 &bench_bloom_update, 639 &bench_bloom_false_positive, 640 &bench_hashmap_without_bloom, 641 &bench_hashmap_with_bloom, 642 &bench_bpf_loop, 643 &bench_strncmp_no_helper, 644 &bench_strncmp_helper, 645 &bench_bpf_hashmap_full_update, 646 &bench_bpf_rhashmap_full_update, 647 &bench_local_storage_cache_seq_get, 648 &bench_local_storage_cache_interleaved_get, 649 &bench_local_storage_cache_hashmap_control, 650 &bench_local_storage_tasks_trace, 651 &bench_bpf_hashmap_lookup, 652 &bench_bpf_rhashmap_lookup, 653 &bench_local_storage_create, 654 &bench_htab_mem, 655 &bench_rhtab_mem, 656 &bench_crypto_encrypt, 657 &bench_crypto_decrypt, 658 &bench_sockmap, 659 &bench_lpm_trie_noop, 660 &bench_lpm_trie_baseline, 661 &bench_lpm_trie_lookup, 662 &bench_lpm_trie_insert, 663 &bench_lpm_trie_update, 664 &bench_lpm_trie_delete, 665 &bench_lpm_trie_free, 666 &bench_bpf_nop, 667 &bench_xdp_lb, 668 }; 669 670 static void find_benchmark(void) 671 { 672 int i; 673 674 if (!env.bench_name) { 675 fprintf(stderr, "benchmark name is not specified\n"); 676 exit(1); 677 } 678 for (i = 0; i < ARRAY_SIZE(benchs); i++) { 679 if (strcmp(benchs[i]->name, env.bench_name) == 0) { 680 bench = benchs[i]; 681 break; 682 } 683 } 684 if (!bench) { 685 fprintf(stderr, "benchmark '%s' not found\n", env.bench_name); 686 exit(1); 687 } 688 } 689 690 static void setup_benchmark(void) 691 { 692 int i, err; 693 694 if (!env.quiet) 695 printf("Setting up benchmark '%s'...\n", bench->name); 696 697 state.producers = calloc(env.producer_cnt, sizeof(*state.producers)); 698 state.consumers = calloc(env.consumer_cnt, sizeof(*state.consumers)); 699 state.results = calloc(env.duration_sec + env.warmup_sec + 2, 700 sizeof(*state.results)); 701 if (!state.producers || !state.consumers || !state.results) 702 exit(1); 703 704 if (bench->validate) 705 bench->validate(); 706 if (bench->setup) 707 bench->setup(); 708 709 for (i = 0; i < env.consumer_cnt; i++) { 710 if (!bench->consumer_thread) { 711 fprintf(stderr, "benchmark doesn't support consumers!\n"); 712 exit(1); 713 } 714 err = pthread_create(&state.consumers[i], NULL, 715 bench->consumer_thread, (void *)(long)i); 716 if (err) { 717 fprintf(stderr, "failed to create consumer thread #%d: %d\n", 718 i, -err); 719 exit(1); 720 } 721 if (env.affinity) 722 set_thread_affinity(state.consumers[i], 723 next_cpu(&env.cons_cpus)); 724 } 725 726 /* unless explicit producer CPU list is specified, continue after 727 * last consumer CPU 728 */ 729 if (!env.prod_cpus.cpus) 730 env.prod_cpus.next_cpu = env.cons_cpus.next_cpu; 731 732 for (i = 0; i < env.producer_cnt; i++) { 733 if (!bench->producer_thread) { 734 fprintf(stderr, "benchmark doesn't support producers!\n"); 735 exit(1); 736 } 737 err = pthread_create(&state.producers[i], NULL, 738 bench->producer_thread, (void *)(long)i); 739 if (err) { 740 fprintf(stderr, "failed to create producer thread #%d: %d\n", 741 i, -err); 742 exit(1); 743 } 744 if (env.affinity) 745 set_thread_affinity(state.producers[i], 746 next_cpu(&env.prod_cpus)); 747 } 748 749 if (!env.quiet) 750 printf("Benchmark '%s' started.\n", bench->name); 751 } 752 753 static pthread_mutex_t bench_done_mtx = PTHREAD_MUTEX_INITIALIZER; 754 static pthread_cond_t bench_done = PTHREAD_COND_INITIALIZER; 755 756 void bench_force_done(void) 757 { 758 pthread_mutex_lock(&bench_done_mtx); 759 pthread_cond_signal(&bench_done); 760 pthread_mutex_unlock(&bench_done_mtx); 761 } 762 763 static void collect_measurements(long delta_ns) { 764 int iter = state.res_cnt++; 765 struct bench_res *res = &state.results[iter]; 766 767 bench->measure(res); 768 769 if (bench->report_progress) 770 bench->report_progress(iter, res, delta_ns); 771 772 if (iter == env.duration_sec + env.warmup_sec) 773 bench_force_done(); 774 } 775 776 int main(int argc, char **argv) 777 { 778 env.nr_cpus = get_nprocs(); 779 parse_cmdline_args_init(argc, argv); 780 781 if (env.list) { 782 int i; 783 784 printf("Available benchmarks:\n"); 785 for (i = 0; i < ARRAY_SIZE(benchs); i++) { 786 printf("- %s\n", benchs[i]->name); 787 } 788 return 0; 789 } 790 791 find_benchmark(); 792 parse_cmdline_args_final(argc, argv); 793 794 setup_benchmark(); 795 796 setup_timer(); 797 798 pthread_mutex_lock(&bench_done_mtx); 799 pthread_cond_wait(&bench_done, &bench_done_mtx); 800 pthread_mutex_unlock(&bench_done_mtx); 801 802 if (bench->report_final) 803 /* skip first sample */ 804 bench->report_final(state.results + env.warmup_sec, 805 state.res_cnt - env.warmup_sec); 806 807 return 0; 808 } 809