tools/testing/selftests/bpf/bench.c

   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 <sys/resource.h>
  12 #include <signal.h>
  13 #include "bench.h"
  14 #include "testing_helpers.h"
  15
  16 struct env env = {
  17         .warmup_sec = 1,
  18         .duration_sec = 5,
  19         .affinity = false,
  20         .consumer_cnt = 1,
  21         .producer_cnt = 1,
  22 };
  23
  24 static int libbpf_print_fn(enum libbpf_print_level level,
  25                     const char *format, va_list args)
  26 {
  27         if (level == LIBBPF_DEBUG && !env.verbose)
  28                 return 0;
  29         return vfprintf(stderr, format, args);
  30 }
  31
  32 static int bump_memlock_rlimit(void)
  33 {
  34         struct rlimit rlim_new = {
  35                 .rlim_cur       = RLIM_INFINITY,
  36                 .rlim_max       = RLIM_INFINITY,
  37         };
  38
  39         return setrlimit(RLIMIT_MEMLOCK, &rlim_new);
  40 }
  41
  42 void setup_libbpf()
  43 {
  44         int err;
  45
  46         libbpf_set_print(libbpf_print_fn);
  47
  48         err = bump_memlock_rlimit();
  49         if (err)
  50                 fprintf(stderr, "failed to increase RLIMIT_MEMLOCK: %d", err);
  51 }
  52
  53 void hits_drops_report_progress(int iter, struct bench_res *res, long delta_ns)
  54 {
  55         double hits_per_sec, drops_per_sec;
  56         double hits_per_prod;
  57
  58         hits_per_sec = res->hits / 1000000.0 / (delta_ns / 1000000000.0);
  59         hits_per_prod = hits_per_sec / env.producer_cnt;
  60         drops_per_sec = res->drops / 1000000.0 / (delta_ns / 1000000000.0);
  61
  62         printf("Iter %3d (%7.3lfus): ",
  63                iter, (delta_ns - 1000000000) / 1000.0);
  64
  65         printf("hits %8.3lfM/s (%7.3lfM/prod), drops %8.3lfM/s\n",
  66                hits_per_sec, hits_per_prod, drops_per_sec);
  67 }
  68
  69 void hits_drops_report_final(struct bench_res res[], int res_cnt)
  70 {
  71         int i;
  72         double hits_mean = 0.0, drops_mean = 0.0;
  73         double hits_stddev = 0.0, drops_stddev = 0.0;
  74
  75         for (i = 0; i < res_cnt; i++) {
  76                 hits_mean += res[i].hits / 1000000.0 / (0.0 + res_cnt);
  77                 drops_mean += res[i].drops / 1000000.0 / (0.0 + res_cnt);
  78         }
  79
  80         if (res_cnt > 1)  {
  81                 for (i = 0; i < res_cnt; i++) {
  82                         hits_stddev += (hits_mean - res[i].hits / 1000000.0) *
  83                                        (hits_mean - res[i].hits / 1000000.0) /
  84                                        (res_cnt - 1.0);
  85                         drops_stddev += (drops_mean - res[i].drops / 1000000.0) *
  86                                         (drops_mean - res[i].drops / 1000000.0) /
  87                                         (res_cnt - 1.0);
  88                 }
  89                 hits_stddev = sqrt(hits_stddev);
  90                 drops_stddev = sqrt(drops_stddev);
  91         }
  92         printf("Summary: hits %8.3lf \u00B1 %5.3lfM/s (%7.3lfM/prod), ",
  93                hits_mean, hits_stddev, hits_mean / env.producer_cnt);
  94         printf("drops %8.3lf \u00B1 %5.3lfM/s\n",
  95                drops_mean, drops_stddev);
  96 }
  97
  98 const char *argp_program_version = "benchmark";
  99 const char *argp_program_bug_address = "<bpf@vger.kernel.org>";
 100 const char argp_program_doc[] =
 101 "benchmark    Generic benchmarking framework.\n"
 102 "\n"
 103 "This tool runs benchmarks.\n"
 104 "\n"
 105 "USAGE: benchmark <bench-name>\n"
 106 "\n"
 107 "EXAMPLES:\n"
 108 "    # run 'count-local' benchmark with 1 producer and 1 consumer\n"
 109 "    benchmark count-local\n"
 110 "    # run 'count-local' with 16 producer and 8 consumer thread, pinned to CPUs\n"
 111 "    benchmark -p16 -c8 -a count-local\n";
 112
 113 enum {
 114         ARG_PROD_AFFINITY_SET = 1000,
 115         ARG_CONS_AFFINITY_SET = 1001,
 116 };
 117
 118 static const struct argp_option opts[] = {
 119         { "list", 'l', NULL, 0, "List available benchmarks"},
 120         { "duration", 'd', "SEC", 0, "Duration of benchmark, seconds"},
 121         { "warmup", 'w', "SEC", 0, "Warm-up period, seconds"},
 122         { "producers", 'p', "NUM", 0, "Number of producer threads"},
 123         { "consumers", 'c', "NUM", 0, "Number of consumer threads"},
 124         { "verbose", 'v', NULL, 0, "Verbose debug output"},
 125         { "affinity", 'a', NULL, 0, "Set consumer/producer thread affinity"},
 126         { "prod-affinity", ARG_PROD_AFFINITY_SET, "CPUSET", 0,
 127           "Set of CPUs for producer threads; implies --affinity"},
 128         { "cons-affinity", ARG_CONS_AFFINITY_SET, "CPUSET", 0,
 129           "Set of CPUs for consumer threads; implies --affinity"},
 130         {},
 131 };
 132
 133 static error_t parse_arg(int key, char *arg, struct argp_state *state)
 134 {
 135         static int pos_args;
 136
 137         switch (key) {
 138         case 'v':
 139                 env.verbose = true;
 140                 break;
 141         case 'l':
 142                 env.list = true;
 143                 break;
 144         case 'd':
 145                 env.duration_sec = strtol(arg, NULL, 10);
 146                 if (env.duration_sec <= 0) {
 147                         fprintf(stderr, "Invalid duration: %s\n", arg);
 148                         argp_usage(state);
 149                 }
 150                 break;
 151         case 'w':
 152                 env.warmup_sec = strtol(arg, NULL, 10);
 153                 if (env.warmup_sec <= 0) {
 154                         fprintf(stderr, "Invalid warm-up duration: %s\n", arg);
 155                         argp_usage(state);
 156                 }
 157                 break;
 158         case 'p':
 159                 env.producer_cnt = strtol(arg, NULL, 10);
 160                 if (env.producer_cnt <= 0) {
 161                         fprintf(stderr, "Invalid producer count: %s\n", arg);
 162                         argp_usage(state);
 163                 }
 164                 break;
 165         case 'c':
 166                 env.consumer_cnt = strtol(arg, NULL, 10);
 167                 if (env.consumer_cnt <= 0) {
 168                         fprintf(stderr, "Invalid consumer count: %s\n", arg);
 169                         argp_usage(state);
 170                 }
 171                 break;
 172         case 'a':
 173                 env.affinity = true;
 174                 break;
 175         case ARG_PROD_AFFINITY_SET:
 176                 env.affinity = true;
 177                 if (parse_num_list(arg, &env.prod_cpus.cpus,
 178                                    &env.prod_cpus.cpus_len)) {
 179                         fprintf(stderr, "Invalid format of CPU set for producers.");
 180                         argp_usage(state);
 181                 }
 182                 break;
 183         case ARG_CONS_AFFINITY_SET:
 184                 env.affinity = true;
 185                 if (parse_num_list(arg, &env.cons_cpus.cpus,
 186                                    &env.cons_cpus.cpus_len)) {
 187                         fprintf(stderr, "Invalid format of CPU set for consumers.");
 188                         argp_usage(state);
 189                 }
 190                 break;
 191         case ARGP_KEY_ARG:
 192                 if (pos_args++) {
 193                         fprintf(stderr,
 194                                 "Unrecognized positional argument: %s\n", arg);
 195                         argp_usage(state);
 196                 }
 197                 env.bench_name = strdup(arg);
 198                 break;
 199         default:
 200                 return ARGP_ERR_UNKNOWN;
 201         }
 202         return 0;
 203 }
 204
 205 static void parse_cmdline_args(int argc, char **argv)
 206 {
 207         static const struct argp argp = {
 208                 .options = opts,
 209                 .parser = parse_arg,
 210                 .doc = argp_program_doc,
 211         };
 212         if (argp_parse(&argp, argc, argv, 0, NULL, NULL))
 213                 exit(1);
 214         if (!env.list && !env.bench_name) {
 215                 argp_help(&argp, stderr, ARGP_HELP_DOC, "bench");
 216                 exit(1);
 217         }
 218 }
 219
 220 static void collect_measurements(long delta_ns);
 221
 222 static __u64 last_time_ns;
 223 static void sigalarm_handler(int signo)
 224 {
 225         long new_time_ns = get_time_ns();
 226         long delta_ns = new_time_ns - last_time_ns;
 227
 228         collect_measurements(delta_ns);
 229
 230         last_time_ns = new_time_ns;
 231 }
 232
 233 /* set up periodic 1-second timer */
 234 static void setup_timer()
 235 {
 236         static struct sigaction sigalarm_action = {
 237                 .sa_handler = sigalarm_handler,
 238         };
 239         struct itimerval timer_settings = {};
 240         int err;
 241
 242         last_time_ns = get_time_ns();
 243         err = sigaction(SIGALRM, &sigalarm_action, NULL);
 244         if (err < 0) {
 245                 fprintf(stderr, "failed to install SIGALRM handler: %d\n", -errno);
 246                 exit(1);
 247         }
 248         timer_settings.it_interval.tv_sec = 1;
 249         timer_settings.it_value.tv_sec = 1;
 250         err = setitimer(ITIMER_REAL, &timer_settings, NULL);
 251         if (err < 0) {
 252                 fprintf(stderr, "failed to arm interval timer: %d\n", -errno);
 253                 exit(1);
 254         }
 255 }
 256
 257 static void set_thread_affinity(pthread_t thread, int cpu)
 258 {
 259         cpu_set_t cpuset;
 260
 261         CPU_ZERO(&cpuset);
 262         CPU_SET(cpu, &cpuset);
 263         if (pthread_setaffinity_np(thread, sizeof(cpuset), &cpuset)) {
 264                 fprintf(stderr, "setting affinity to CPU #%d failed: %d\n",
 265                         cpu, errno);
 266                 exit(1);
 267         }
 268 }
 269
 270 static int next_cpu(struct cpu_set *cpu_set)
 271 {
 272         if (cpu_set->cpus) {
 273                 int i;
 274
 275                 /* find next available CPU */
 276                 for (i = cpu_set->next_cpu; i < cpu_set->cpus_len; i++) {
 277                         if (cpu_set->cpus[i]) {
 278                                 cpu_set->next_cpu = i + 1;
 279                                 return i;
 280                         }
 281                 }
 282                 fprintf(stderr, "Not enough CPUs specified, need CPU #%d or higher.\n", i);
 283                 exit(1);
 284         }
 285
 286         return cpu_set->next_cpu++;
 287 }
 288
 289 static struct bench_state {
 290         int res_cnt;
 291         struct bench_res *results;
 292         pthread_t *consumers;
 293         pthread_t *producers;
 294 } state;
 295
 296 const struct bench *bench = NULL;
 297
 298 extern const struct bench bench_count_global;
 299 extern const struct bench bench_count_local;
 300 extern const struct bench bench_rename_base;
 301 extern const struct bench bench_rename_kprobe;
 302 extern const struct bench bench_rename_kretprobe;
 303 extern const struct bench bench_rename_rawtp;
 304 extern const struct bench bench_rename_fentry;
 305 extern const struct bench bench_rename_fexit;
 306 extern const struct bench bench_rename_fmodret;
 307 extern const struct bench bench_trig_base;
 308 extern const struct bench bench_trig_tp;
 309 extern const struct bench bench_trig_rawtp;
 310 extern const struct bench bench_trig_kprobe;
 311 extern const struct bench bench_trig_fentry;
 312 extern const struct bench bench_trig_fmodret;
 313
 314 static const struct bench *benchs[] = {
 315         &bench_count_global,
 316         &bench_count_local,
 317         &bench_rename_base,
 318         &bench_rename_kprobe,
 319         &bench_rename_kretprobe,
 320         &bench_rename_rawtp,
 321         &bench_rename_fentry,
 322         &bench_rename_fexit,
 323         &bench_rename_fmodret,
 324         &bench_trig_base,
 325         &bench_trig_tp,
 326         &bench_trig_rawtp,
 327         &bench_trig_kprobe,
 328         &bench_trig_fentry,
 329         &bench_trig_fmodret,
 330 };
 331
 332 static void setup_benchmark()
 333 {
 334         int i, err;
 335
 336         if (!env.bench_name) {
 337                 fprintf(stderr, "benchmark name is not specified\n");
 338                 exit(1);
 339         }
 340
 341         for (i = 0; i < ARRAY_SIZE(benchs); i++) {
 342                 if (strcmp(benchs[i]->name, env.bench_name) == 0) {
 343                         bench = benchs[i];
 344                         break;
 345                 }
 346         }
 347         if (!bench) {
 348                 fprintf(stderr, "benchmark '%s' not found\n", env.bench_name);
 349                 exit(1);
 350         }
 351
 352         printf("Setting up benchmark '%s'...\n", bench->name);
 353
 354         state.producers = calloc(env.producer_cnt, sizeof(*state.producers));
 355         state.consumers = calloc(env.consumer_cnt, sizeof(*state.consumers));
 356         state.results = calloc(env.duration_sec + env.warmup_sec + 2,
 357                                sizeof(*state.results));
 358         if (!state.producers || !state.consumers || !state.results)
 359                 exit(1);
 360
 361         if (bench->validate)
 362                 bench->validate();
 363         if (bench->setup)
 364                 bench->setup();
 365
 366         for (i = 0; i < env.consumer_cnt; i++) {
 367                 err = pthread_create(&state.consumers[i], NULL,
 368                                      bench->consumer_thread, (void *)(long)i);
 369                 if (err) {
 370                         fprintf(stderr, "failed to create consumer thread #%d: %d\n",
 371                                 i, -errno);
 372                         exit(1);
 373                 }
 374                 if (env.affinity)
 375                         set_thread_affinity(state.consumers[i],
 376                                             next_cpu(&env.cons_cpus));
 377         }
 378
 379         /* unless explicit producer CPU list is specified, continue after
 380          * last consumer CPU
 381          */
 382         if (!env.prod_cpus.cpus)
 383                 env.prod_cpus.next_cpu = env.cons_cpus.next_cpu;
 384
 385         for (i = 0; i < env.producer_cnt; i++) {
 386                 err = pthread_create(&state.producers[i], NULL,
 387                                      bench->producer_thread, (void *)(long)i);
 388                 if (err) {
 389                         fprintf(stderr, "failed to create producer thread #%d: %d\n",
 390                                 i, -errno);
 391                         exit(1);
 392                 }
 393                 if (env.affinity)
 394                         set_thread_affinity(state.producers[i],
 395                                             next_cpu(&env.prod_cpus));
 396         }
 397
 398         printf("Benchmark '%s' started.\n", bench->name);
 399 }
 400
 401 static pthread_mutex_t bench_done_mtx = PTHREAD_MUTEX_INITIALIZER;
 402 static pthread_cond_t bench_done = PTHREAD_COND_INITIALIZER;
 403
 404 static void collect_measurements(long delta_ns) {
 405         int iter = state.res_cnt++;
 406         struct bench_res *res = &state.results[iter];
 407
 408         bench->measure(res);
 409
 410         if (bench->report_progress)
 411                 bench->report_progress(iter, res, delta_ns);
 412
 413         if (iter == env.duration_sec + env.warmup_sec) {
 414                 pthread_mutex_lock(&bench_done_mtx);
 415                 pthread_cond_signal(&bench_done);
 416                 pthread_mutex_unlock(&bench_done_mtx);
 417         }
 418 }
 419
 420 int main(int argc, char **argv)
 421 {
 422         parse_cmdline_args(argc, argv);
 423
 424         if (env.list) {
 425                 int i;
 426
 427                 printf("Available benchmarks:\n");
 428                 for (i = 0; i < ARRAY_SIZE(benchs); i++) {
 429                         printf("- %s\n", benchs[i]->name);
 430                 }
 431                 return 0;
 432         }
 433
 434         setup_benchmark();
 435
 436         setup_timer();
 437
 438         pthread_mutex_lock(&bench_done_mtx);
 439         pthread_cond_wait(&bench_done, &bench_done_mtx);
 440         pthread_mutex_unlock(&bench_done_mtx);
 441
 442         if (bench->report_final)
 443                 /* skip first sample */
 444                 bench->report_final(state.results + env.warmup_sec,
 445                                     state.res_cnt - env.warmup_sec);
 446
 447         return 0;
 448 }
 449