xref: /linux/tools/testing/selftests/bpf/bench.c (revision 55d0969c451159cff86949b38c39171cab962069)
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 	{ "prod-affinity", ARG_PROD_AFFINITY_SET, "CPUSET", 0,
269 	  "Set of CPUs for producer threads; implies --affinity"},
270 	{ "cons-affinity", ARG_CONS_AFFINITY_SET, "CPUSET", 0,
271 	  "Set of CPUs for consumer threads; implies --affinity"},
272 	{},
273 };
274 
275 extern struct argp bench_ringbufs_argp;
276 extern struct argp bench_bloom_map_argp;
277 extern struct argp bench_bpf_loop_argp;
278 extern struct argp bench_local_storage_argp;
279 extern struct argp bench_local_storage_rcu_tasks_trace_argp;
280 extern struct argp bench_strncmp_argp;
281 extern struct argp bench_hashmap_lookup_argp;
282 extern struct argp bench_local_storage_create_argp;
283 extern struct argp bench_htab_mem_argp;
284 extern struct argp bench_trigger_batch_argp;
285 extern struct argp bench_crypto_argp;
286 
287 static const struct argp_child bench_parsers[] = {
288 	{ &bench_ringbufs_argp, 0, "Ring buffers benchmark", 0 },
289 	{ &bench_bloom_map_argp, 0, "Bloom filter map benchmark", 0 },
290 	{ &bench_bpf_loop_argp, 0, "bpf_loop helper benchmark", 0 },
291 	{ &bench_local_storage_argp, 0, "local_storage benchmark", 0 },
292 	{ &bench_strncmp_argp, 0, "bpf_strncmp helper benchmark", 0 },
293 	{ &bench_local_storage_rcu_tasks_trace_argp, 0,
294 		"local_storage RCU Tasks Trace slowdown benchmark", 0 },
295 	{ &bench_hashmap_lookup_argp, 0, "Hashmap lookup benchmark", 0 },
296 	{ &bench_local_storage_create_argp, 0, "local-storage-create benchmark", 0 },
297 	{ &bench_htab_mem_argp, 0, "hash map memory benchmark", 0 },
298 	{ &bench_trigger_batch_argp, 0, "BPF triggering benchmark", 0 },
299 	{ &bench_crypto_argp, 0, "bpf crypto benchmark", 0 },
300 	{},
301 };
302 
303 /* Make pos_args global, so that we can run argp_parse twice, if necessary */
304 static int pos_args;
305 
306 static error_t parse_arg(int key, char *arg, struct argp_state *state)
307 {
308 	switch (key) {
309 	case 'v':
310 		env.verbose = true;
311 		break;
312 	case 'l':
313 		env.list = true;
314 		break;
315 	case 'd':
316 		env.duration_sec = strtol(arg, NULL, 10);
317 		if (env.duration_sec <= 0) {
318 			fprintf(stderr, "Invalid duration: %s\n", arg);
319 			argp_usage(state);
320 		}
321 		break;
322 	case 'w':
323 		env.warmup_sec = strtol(arg, NULL, 10);
324 		if (env.warmup_sec <= 0) {
325 			fprintf(stderr, "Invalid warm-up duration: %s\n", arg);
326 			argp_usage(state);
327 		}
328 		break;
329 	case 'p':
330 		env.producer_cnt = strtol(arg, NULL, 10);
331 		if (env.producer_cnt < 0) {
332 			fprintf(stderr, "Invalid producer count: %s\n", arg);
333 			argp_usage(state);
334 		}
335 		break;
336 	case 'c':
337 		env.consumer_cnt = strtol(arg, NULL, 10);
338 		if (env.consumer_cnt < 0) {
339 			fprintf(stderr, "Invalid consumer count: %s\n", arg);
340 			argp_usage(state);
341 		}
342 		break;
343 	case 'a':
344 		env.affinity = true;
345 		break;
346 	case 'q':
347 		env.quiet = true;
348 		break;
349 	case ARG_PROD_AFFINITY_SET:
350 		env.affinity = true;
351 		if (parse_num_list(arg, &env.prod_cpus.cpus,
352 				   &env.prod_cpus.cpus_len)) {
353 			fprintf(stderr, "Invalid format of CPU set for producers.");
354 			argp_usage(state);
355 		}
356 		break;
357 	case ARG_CONS_AFFINITY_SET:
358 		env.affinity = true;
359 		if (parse_num_list(arg, &env.cons_cpus.cpus,
360 				   &env.cons_cpus.cpus_len)) {
361 			fprintf(stderr, "Invalid format of CPU set for consumers.");
362 			argp_usage(state);
363 		}
364 		break;
365 	case ARGP_KEY_ARG:
366 		if (pos_args++) {
367 			fprintf(stderr,
368 				"Unrecognized positional argument: %s\n", arg);
369 			argp_usage(state);
370 		}
371 		env.bench_name = strdup(arg);
372 		break;
373 	default:
374 		return ARGP_ERR_UNKNOWN;
375 	}
376 	return 0;
377 }
378 
379 static void parse_cmdline_args_init(int argc, char **argv)
380 {
381 	static const struct argp argp = {
382 		.options = opts,
383 		.parser = parse_arg,
384 		.doc = argp_program_doc,
385 		.children = bench_parsers,
386 	};
387 	if (argp_parse(&argp, argc, argv, 0, NULL, NULL))
388 		exit(1);
389 }
390 
391 static void parse_cmdline_args_final(int argc, char **argv)
392 {
393 	struct argp_child bench_parsers[2] = {};
394 	const struct argp argp = {
395 		.options = opts,
396 		.parser = parse_arg,
397 		.doc = argp_program_doc,
398 		.children = bench_parsers,
399 	};
400 
401 	/* Parse arguments the second time with the correct set of parsers */
402 	if (bench->argp) {
403 		bench_parsers[0].argp = bench->argp;
404 		bench_parsers[0].header = bench->name;
405 		pos_args = 0;
406 		if (argp_parse(&argp, argc, argv, 0, NULL, NULL))
407 			exit(1);
408 	}
409 }
410 
411 static void collect_measurements(long delta_ns);
412 
413 static __u64 last_time_ns;
414 static void sigalarm_handler(int signo)
415 {
416 	long new_time_ns = get_time_ns();
417 	long delta_ns = new_time_ns - last_time_ns;
418 
419 	collect_measurements(delta_ns);
420 
421 	last_time_ns = new_time_ns;
422 }
423 
424 /* set up periodic 1-second timer */
425 static void setup_timer()
426 {
427 	static struct sigaction sigalarm_action = {
428 		.sa_handler = sigalarm_handler,
429 	};
430 	struct itimerval timer_settings = {};
431 	int err;
432 
433 	last_time_ns = get_time_ns();
434 	err = sigaction(SIGALRM, &sigalarm_action, NULL);
435 	if (err < 0) {
436 		fprintf(stderr, "failed to install SIGALRM handler: %d\n", -errno);
437 		exit(1);
438 	}
439 	timer_settings.it_interval.tv_sec = 1;
440 	timer_settings.it_value.tv_sec = 1;
441 	err = setitimer(ITIMER_REAL, &timer_settings, NULL);
442 	if (err < 0) {
443 		fprintf(stderr, "failed to arm interval timer: %d\n", -errno);
444 		exit(1);
445 	}
446 }
447 
448 static void set_thread_affinity(pthread_t thread, int cpu)
449 {
450 	cpu_set_t cpuset;
451 	int err;
452 
453 	CPU_ZERO(&cpuset);
454 	CPU_SET(cpu, &cpuset);
455 	err = pthread_setaffinity_np(thread, sizeof(cpuset), &cpuset);
456 	if (err) {
457 		fprintf(stderr, "setting affinity to CPU #%d failed: %d\n",
458 			cpu, -err);
459 		exit(1);
460 	}
461 }
462 
463 static int next_cpu(struct cpu_set *cpu_set)
464 {
465 	if (cpu_set->cpus) {
466 		int i;
467 
468 		/* find next available CPU */
469 		for (i = cpu_set->next_cpu; i < cpu_set->cpus_len; i++) {
470 			if (cpu_set->cpus[i]) {
471 				cpu_set->next_cpu = i + 1;
472 				return i;
473 			}
474 		}
475 		fprintf(stderr, "Not enough CPUs specified, need CPU #%d or higher.\n", i);
476 		exit(1);
477 	}
478 
479 	return cpu_set->next_cpu++ % env.nr_cpus;
480 }
481 
482 static struct bench_state {
483 	int res_cnt;
484 	struct bench_res *results;
485 	pthread_t *consumers;
486 	pthread_t *producers;
487 } state;
488 
489 const struct bench *bench = NULL;
490 
491 extern const struct bench bench_count_global;
492 extern const struct bench bench_count_local;
493 extern const struct bench bench_rename_base;
494 extern const struct bench bench_rename_kprobe;
495 extern const struct bench bench_rename_kretprobe;
496 extern const struct bench bench_rename_rawtp;
497 extern const struct bench bench_rename_fentry;
498 extern const struct bench bench_rename_fexit;
499 
500 /* pure counting benchmarks to establish theoretical lmits */
501 extern const struct bench bench_trig_usermode_count;
502 extern const struct bench bench_trig_syscall_count;
503 extern const struct bench bench_trig_kernel_count;
504 
505 /* batched, staying mostly in-kernel benchmarks */
506 extern const struct bench bench_trig_kprobe;
507 extern const struct bench bench_trig_kretprobe;
508 extern const struct bench bench_trig_kprobe_multi;
509 extern const struct bench bench_trig_kretprobe_multi;
510 extern const struct bench bench_trig_fentry;
511 extern const struct bench bench_trig_fexit;
512 extern const struct bench bench_trig_fmodret;
513 extern const struct bench bench_trig_tp;
514 extern const struct bench bench_trig_rawtp;
515 
516 /* uprobe/uretprobe benchmarks */
517 extern const struct bench bench_trig_uprobe_nop;
518 extern const struct bench bench_trig_uretprobe_nop;
519 extern const struct bench bench_trig_uprobe_push;
520 extern const struct bench bench_trig_uretprobe_push;
521 extern const struct bench bench_trig_uprobe_ret;
522 extern const struct bench bench_trig_uretprobe_ret;
523 extern const struct bench bench_trig_uprobe_multi_nop;
524 extern const struct bench bench_trig_uretprobe_multi_nop;
525 extern const struct bench bench_trig_uprobe_multi_push;
526 extern const struct bench bench_trig_uretprobe_multi_push;
527 extern const struct bench bench_trig_uprobe_multi_ret;
528 extern const struct bench bench_trig_uretprobe_multi_ret;
529 
530 extern const struct bench bench_rb_libbpf;
531 extern const struct bench bench_rb_custom;
532 extern const struct bench bench_pb_libbpf;
533 extern const struct bench bench_pb_custom;
534 extern const struct bench bench_bloom_lookup;
535 extern const struct bench bench_bloom_update;
536 extern const struct bench bench_bloom_false_positive;
537 extern const struct bench bench_hashmap_without_bloom;
538 extern const struct bench bench_hashmap_with_bloom;
539 extern const struct bench bench_bpf_loop;
540 extern const struct bench bench_strncmp_no_helper;
541 extern const struct bench bench_strncmp_helper;
542 extern const struct bench bench_bpf_hashmap_full_update;
543 extern const struct bench bench_local_storage_cache_seq_get;
544 extern const struct bench bench_local_storage_cache_interleaved_get;
545 extern const struct bench bench_local_storage_cache_hashmap_control;
546 extern const struct bench bench_local_storage_tasks_trace;
547 extern const struct bench bench_bpf_hashmap_lookup;
548 extern const struct bench bench_local_storage_create;
549 extern const struct bench bench_htab_mem;
550 extern const struct bench bench_crypto_encrypt;
551 extern const struct bench bench_crypto_decrypt;
552 
553 static const struct bench *benchs[] = {
554 	&bench_count_global,
555 	&bench_count_local,
556 	&bench_rename_base,
557 	&bench_rename_kprobe,
558 	&bench_rename_kretprobe,
559 	&bench_rename_rawtp,
560 	&bench_rename_fentry,
561 	&bench_rename_fexit,
562 	/* pure counting benchmarks for establishing theoretical limits */
563 	&bench_trig_usermode_count,
564 	&bench_trig_kernel_count,
565 	&bench_trig_syscall_count,
566 	/* batched, staying mostly in-kernel triggers */
567 	&bench_trig_kprobe,
568 	&bench_trig_kretprobe,
569 	&bench_trig_kprobe_multi,
570 	&bench_trig_kretprobe_multi,
571 	&bench_trig_fentry,
572 	&bench_trig_fexit,
573 	&bench_trig_fmodret,
574 	&bench_trig_tp,
575 	&bench_trig_rawtp,
576 	/* uprobes */
577 	&bench_trig_uprobe_nop,
578 	&bench_trig_uretprobe_nop,
579 	&bench_trig_uprobe_push,
580 	&bench_trig_uretprobe_push,
581 	&bench_trig_uprobe_ret,
582 	&bench_trig_uretprobe_ret,
583 	&bench_trig_uprobe_multi_nop,
584 	&bench_trig_uretprobe_multi_nop,
585 	&bench_trig_uprobe_multi_push,
586 	&bench_trig_uretprobe_multi_push,
587 	&bench_trig_uprobe_multi_ret,
588 	&bench_trig_uretprobe_multi_ret,
589 	/* ringbuf/perfbuf benchmarks */
590 	&bench_rb_libbpf,
591 	&bench_rb_custom,
592 	&bench_pb_libbpf,
593 	&bench_pb_custom,
594 	&bench_bloom_lookup,
595 	&bench_bloom_update,
596 	&bench_bloom_false_positive,
597 	&bench_hashmap_without_bloom,
598 	&bench_hashmap_with_bloom,
599 	&bench_bpf_loop,
600 	&bench_strncmp_no_helper,
601 	&bench_strncmp_helper,
602 	&bench_bpf_hashmap_full_update,
603 	&bench_local_storage_cache_seq_get,
604 	&bench_local_storage_cache_interleaved_get,
605 	&bench_local_storage_cache_hashmap_control,
606 	&bench_local_storage_tasks_trace,
607 	&bench_bpf_hashmap_lookup,
608 	&bench_local_storage_create,
609 	&bench_htab_mem,
610 	&bench_crypto_encrypt,
611 	&bench_crypto_decrypt,
612 };
613 
614 static void find_benchmark(void)
615 {
616 	int i;
617 
618 	if (!env.bench_name) {
619 		fprintf(stderr, "benchmark name is not specified\n");
620 		exit(1);
621 	}
622 	for (i = 0; i < ARRAY_SIZE(benchs); i++) {
623 		if (strcmp(benchs[i]->name, env.bench_name) == 0) {
624 			bench = benchs[i];
625 			break;
626 		}
627 	}
628 	if (!bench) {
629 		fprintf(stderr, "benchmark '%s' not found\n", env.bench_name);
630 		exit(1);
631 	}
632 }
633 
634 static void setup_benchmark(void)
635 {
636 	int i, err;
637 
638 	if (!env.quiet)
639 		printf("Setting up benchmark '%s'...\n", bench->name);
640 
641 	state.producers = calloc(env.producer_cnt, sizeof(*state.producers));
642 	state.consumers = calloc(env.consumer_cnt, sizeof(*state.consumers));
643 	state.results = calloc(env.duration_sec + env.warmup_sec + 2,
644 			       sizeof(*state.results));
645 	if (!state.producers || !state.consumers || !state.results)
646 		exit(1);
647 
648 	if (bench->validate)
649 		bench->validate();
650 	if (bench->setup)
651 		bench->setup();
652 
653 	for (i = 0; i < env.consumer_cnt; i++) {
654 		if (!bench->consumer_thread) {
655 			fprintf(stderr, "benchmark doesn't support consumers!\n");
656 			exit(1);
657 		}
658 		err = pthread_create(&state.consumers[i], NULL,
659 				     bench->consumer_thread, (void *)(long)i);
660 		if (err) {
661 			fprintf(stderr, "failed to create consumer thread #%d: %d\n",
662 				i, -err);
663 			exit(1);
664 		}
665 		if (env.affinity)
666 			set_thread_affinity(state.consumers[i],
667 					    next_cpu(&env.cons_cpus));
668 	}
669 
670 	/* unless explicit producer CPU list is specified, continue after
671 	 * last consumer CPU
672 	 */
673 	if (!env.prod_cpus.cpus)
674 		env.prod_cpus.next_cpu = env.cons_cpus.next_cpu;
675 
676 	for (i = 0; i < env.producer_cnt; i++) {
677 		if (!bench->producer_thread) {
678 			fprintf(stderr, "benchmark doesn't support producers!\n");
679 			exit(1);
680 		}
681 		err = pthread_create(&state.producers[i], NULL,
682 				     bench->producer_thread, (void *)(long)i);
683 		if (err) {
684 			fprintf(stderr, "failed to create producer thread #%d: %d\n",
685 				i, -err);
686 			exit(1);
687 		}
688 		if (env.affinity)
689 			set_thread_affinity(state.producers[i],
690 					    next_cpu(&env.prod_cpus));
691 	}
692 
693 	if (!env.quiet)
694 		printf("Benchmark '%s' started.\n", bench->name);
695 }
696 
697 static pthread_mutex_t bench_done_mtx = PTHREAD_MUTEX_INITIALIZER;
698 static pthread_cond_t bench_done = PTHREAD_COND_INITIALIZER;
699 
700 static void collect_measurements(long delta_ns) {
701 	int iter = state.res_cnt++;
702 	struct bench_res *res = &state.results[iter];
703 
704 	bench->measure(res);
705 
706 	if (bench->report_progress)
707 		bench->report_progress(iter, res, delta_ns);
708 
709 	if (iter == env.duration_sec + env.warmup_sec) {
710 		pthread_mutex_lock(&bench_done_mtx);
711 		pthread_cond_signal(&bench_done);
712 		pthread_mutex_unlock(&bench_done_mtx);
713 	}
714 }
715 
716 int main(int argc, char **argv)
717 {
718 	env.nr_cpus = get_nprocs();
719 	parse_cmdline_args_init(argc, argv);
720 
721 	if (env.list) {
722 		int i;
723 
724 		printf("Available benchmarks:\n");
725 		for (i = 0; i < ARRAY_SIZE(benchs); i++) {
726 			printf("- %s\n", benchs[i]->name);
727 		}
728 		return 0;
729 	}
730 
731 	find_benchmark();
732 	parse_cmdline_args_final(argc, argv);
733 
734 	setup_benchmark();
735 
736 	setup_timer();
737 
738 	pthread_mutex_lock(&bench_done_mtx);
739 	pthread_cond_wait(&bench_done, &bench_done_mtx);
740 	pthread_mutex_unlock(&bench_done_mtx);
741 
742 	if (bench->report_final)
743 		/* skip first sample */
744 		bench->report_final(state.results + env.warmup_sec,
745 				    state.res_cnt - env.warmup_sec);
746 
747 	return 0;
748 }
749