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 void setup_libbpf(void)
33 {
34 	libbpf_set_strict_mode(LIBBPF_STRICT_ALL);
35 	libbpf_set_print(libbpf_print_fn);
36 }
37 
38 void false_hits_report_progress(int iter, struct bench_res *res, long delta_ns)
39 {
40 	long total = res->false_hits  + res->hits + res->drops;
41 
42 	printf("Iter %3d (%7.3lfus): ",
43 	       iter, (delta_ns - 1000000000) / 1000.0);
44 
45 	printf("%ld false hits of %ld total operations. Percentage = %2.2f %%\n",
46 	       res->false_hits, total, ((float)res->false_hits / total) * 100);
47 }
48 
49 void false_hits_report_final(struct bench_res res[], int res_cnt)
50 {
51 	long total_hits = 0, total_drops = 0, total_false_hits = 0, total_ops = 0;
52 	int i;
53 
54 	for (i = 0; i < res_cnt; i++) {
55 		total_hits += res[i].hits;
56 		total_false_hits += res[i].false_hits;
57 		total_drops += res[i].drops;
58 	}
59 	total_ops = total_hits + total_false_hits + total_drops;
60 
61 	printf("Summary: %ld false hits of %ld total operations. ",
62 	       total_false_hits, total_ops);
63 	printf("Percentage =  %2.2f %%\n",
64 	       ((float)total_false_hits / total_ops) * 100);
65 }
66 
67 void hits_drops_report_progress(int iter, struct bench_res *res, long delta_ns)
68 {
69 	double hits_per_sec, drops_per_sec;
70 	double hits_per_prod;
71 
72 	hits_per_sec = res->hits / 1000000.0 / (delta_ns / 1000000000.0);
73 	hits_per_prod = hits_per_sec / env.producer_cnt;
74 	drops_per_sec = res->drops / 1000000.0 / (delta_ns / 1000000000.0);
75 
76 	printf("Iter %3d (%7.3lfus): ",
77 	       iter, (delta_ns - 1000000000) / 1000.0);
78 
79 	printf("hits %8.3lfM/s (%7.3lfM/prod), drops %8.3lfM/s, total operations %8.3lfM/s\n",
80 	       hits_per_sec, hits_per_prod, drops_per_sec, hits_per_sec + drops_per_sec);
81 }
82 
83 void hits_drops_report_final(struct bench_res res[], int res_cnt)
84 {
85 	int i;
86 	double hits_mean = 0.0, drops_mean = 0.0, total_ops_mean = 0.0;
87 	double hits_stddev = 0.0, drops_stddev = 0.0, total_ops_stddev = 0.0;
88 	double total_ops;
89 
90 	for (i = 0; i < res_cnt; i++) {
91 		hits_mean += res[i].hits / 1000000.0 / (0.0 + res_cnt);
92 		drops_mean += res[i].drops / 1000000.0 / (0.0 + res_cnt);
93 	}
94 	total_ops_mean = hits_mean + drops_mean;
95 
96 	if (res_cnt > 1)  {
97 		for (i = 0; i < res_cnt; i++) {
98 			hits_stddev += (hits_mean - res[i].hits / 1000000.0) *
99 				       (hits_mean - res[i].hits / 1000000.0) /
100 				       (res_cnt - 1.0);
101 			drops_stddev += (drops_mean - res[i].drops / 1000000.0) *
102 					(drops_mean - res[i].drops / 1000000.0) /
103 					(res_cnt - 1.0);
104 			total_ops = res[i].hits + res[i].drops;
105 			total_ops_stddev += (total_ops_mean - total_ops / 1000000.0) *
106 					(total_ops_mean - total_ops / 1000000.0) /
107 					(res_cnt - 1.0);
108 		}
109 		hits_stddev = sqrt(hits_stddev);
110 		drops_stddev = sqrt(drops_stddev);
111 		total_ops_stddev = sqrt(total_ops_stddev);
112 	}
113 	printf("Summary: hits %8.3lf \u00B1 %5.3lfM/s (%7.3lfM/prod), ",
114 	       hits_mean, hits_stddev, hits_mean / env.producer_cnt);
115 	printf("drops %8.3lf \u00B1 %5.3lfM/s, ",
116 	       drops_mean, drops_stddev);
117 	printf("total operations %8.3lf \u00B1 %5.3lfM/s\n",
118 	       total_ops_mean, total_ops_stddev);
119 }
120 
121 void ops_report_progress(int iter, struct bench_res *res, long delta_ns)
122 {
123 	double hits_per_sec, hits_per_prod;
124 
125 	hits_per_sec = res->hits / 1000000.0 / (delta_ns / 1000000000.0);
126 	hits_per_prod = hits_per_sec / env.producer_cnt;
127 
128 	printf("Iter %3d (%7.3lfus): ", iter, (delta_ns - 1000000000) / 1000.0);
129 
130 	printf("hits %8.3lfM/s (%7.3lfM/prod)\n", hits_per_sec, hits_per_prod);
131 }
132 
133 void ops_report_final(struct bench_res res[], int res_cnt)
134 {
135 	double hits_mean = 0.0, hits_stddev = 0.0;
136 	int i;
137 
138 	for (i = 0; i < res_cnt; i++)
139 		hits_mean += res[i].hits / 1000000.0 / (0.0 + res_cnt);
140 
141 	if (res_cnt > 1)  {
142 		for (i = 0; i < res_cnt; i++)
143 			hits_stddev += (hits_mean - res[i].hits / 1000000.0) *
144 				       (hits_mean - res[i].hits / 1000000.0) /
145 				       (res_cnt - 1.0);
146 
147 		hits_stddev = sqrt(hits_stddev);
148 	}
149 	printf("Summary: throughput %8.3lf \u00B1 %5.3lf M ops/s (%7.3lfM ops/prod), ",
150 	       hits_mean, hits_stddev, hits_mean / env.producer_cnt);
151 	printf("latency %8.3lf ns/op\n", 1000.0 / hits_mean * env.producer_cnt);
152 }
153 
154 const char *argp_program_version = "benchmark";
155 const char *argp_program_bug_address = "<bpf@vger.kernel.org>";
156 const char argp_program_doc[] =
157 "benchmark    Generic benchmarking framework.\n"
158 "\n"
159 "This tool runs benchmarks.\n"
160 "\n"
161 "USAGE: benchmark <bench-name>\n"
162 "\n"
163 "EXAMPLES:\n"
164 "    # run 'count-local' benchmark with 1 producer and 1 consumer\n"
165 "    benchmark count-local\n"
166 "    # run 'count-local' with 16 producer and 8 consumer thread, pinned to CPUs\n"
167 "    benchmark -p16 -c8 -a count-local\n";
168 
169 enum {
170 	ARG_PROD_AFFINITY_SET = 1000,
171 	ARG_CONS_AFFINITY_SET = 1001,
172 };
173 
174 static const struct argp_option opts[] = {
175 	{ "list", 'l', NULL, 0, "List available benchmarks"},
176 	{ "duration", 'd', "SEC", 0, "Duration of benchmark, seconds"},
177 	{ "warmup", 'w', "SEC", 0, "Warm-up period, seconds"},
178 	{ "producers", 'p', "NUM", 0, "Number of producer threads"},
179 	{ "consumers", 'c', "NUM", 0, "Number of consumer threads"},
180 	{ "verbose", 'v', NULL, 0, "Verbose debug output"},
181 	{ "affinity", 'a', NULL, 0, "Set consumer/producer thread affinity"},
182 	{ "prod-affinity", ARG_PROD_AFFINITY_SET, "CPUSET", 0,
183 	  "Set of CPUs for producer threads; implies --affinity"},
184 	{ "cons-affinity", ARG_CONS_AFFINITY_SET, "CPUSET", 0,
185 	  "Set of CPUs for consumer threads; implies --affinity"},
186 	{},
187 };
188 
189 extern struct argp bench_ringbufs_argp;
190 extern struct argp bench_bloom_map_argp;
191 extern struct argp bench_bpf_loop_argp;
192 extern struct argp bench_strncmp_argp;
193 
194 static const struct argp_child bench_parsers[] = {
195 	{ &bench_ringbufs_argp, 0, "Ring buffers benchmark", 0 },
196 	{ &bench_bloom_map_argp, 0, "Bloom filter map benchmark", 0 },
197 	{ &bench_bpf_loop_argp, 0, "bpf_loop helper benchmark", 0 },
198 	{ &bench_strncmp_argp, 0, "bpf_strncmp helper benchmark", 0 },
199 	{},
200 };
201 
202 static error_t parse_arg(int key, char *arg, struct argp_state *state)
203 {
204 	static int pos_args;
205 
206 	switch (key) {
207 	case 'v':
208 		env.verbose = true;
209 		break;
210 	case 'l':
211 		env.list = true;
212 		break;
213 	case 'd':
214 		env.duration_sec = strtol(arg, NULL, 10);
215 		if (env.duration_sec <= 0) {
216 			fprintf(stderr, "Invalid duration: %s\n", arg);
217 			argp_usage(state);
218 		}
219 		break;
220 	case 'w':
221 		env.warmup_sec = strtol(arg, NULL, 10);
222 		if (env.warmup_sec <= 0) {
223 			fprintf(stderr, "Invalid warm-up duration: %s\n", arg);
224 			argp_usage(state);
225 		}
226 		break;
227 	case 'p':
228 		env.producer_cnt = strtol(arg, NULL, 10);
229 		if (env.producer_cnt <= 0) {
230 			fprintf(stderr, "Invalid producer count: %s\n", arg);
231 			argp_usage(state);
232 		}
233 		break;
234 	case 'c':
235 		env.consumer_cnt = strtol(arg, NULL, 10);
236 		if (env.consumer_cnt <= 0) {
237 			fprintf(stderr, "Invalid consumer count: %s\n", arg);
238 			argp_usage(state);
239 		}
240 		break;
241 	case 'a':
242 		env.affinity = true;
243 		break;
244 	case ARG_PROD_AFFINITY_SET:
245 		env.affinity = true;
246 		if (parse_num_list(arg, &env.prod_cpus.cpus,
247 				   &env.prod_cpus.cpus_len)) {
248 			fprintf(stderr, "Invalid format of CPU set for producers.");
249 			argp_usage(state);
250 		}
251 		break;
252 	case ARG_CONS_AFFINITY_SET:
253 		env.affinity = true;
254 		if (parse_num_list(arg, &env.cons_cpus.cpus,
255 				   &env.cons_cpus.cpus_len)) {
256 			fprintf(stderr, "Invalid format of CPU set for consumers.");
257 			argp_usage(state);
258 		}
259 		break;
260 	case ARGP_KEY_ARG:
261 		if (pos_args++) {
262 			fprintf(stderr,
263 				"Unrecognized positional argument: %s\n", arg);
264 			argp_usage(state);
265 		}
266 		env.bench_name = strdup(arg);
267 		break;
268 	default:
269 		return ARGP_ERR_UNKNOWN;
270 	}
271 	return 0;
272 }
273 
274 static void parse_cmdline_args(int argc, char **argv)
275 {
276 	static const struct argp argp = {
277 		.options = opts,
278 		.parser = parse_arg,
279 		.doc = argp_program_doc,
280 		.children = bench_parsers,
281 	};
282 	if (argp_parse(&argp, argc, argv, 0, NULL, NULL))
283 		exit(1);
284 	if (!env.list && !env.bench_name) {
285 		argp_help(&argp, stderr, ARGP_HELP_DOC, "bench");
286 		exit(1);
287 	}
288 }
289 
290 static void collect_measurements(long delta_ns);
291 
292 static __u64 last_time_ns;
293 static void sigalarm_handler(int signo)
294 {
295 	long new_time_ns = get_time_ns();
296 	long delta_ns = new_time_ns - last_time_ns;
297 
298 	collect_measurements(delta_ns);
299 
300 	last_time_ns = new_time_ns;
301 }
302 
303 /* set up periodic 1-second timer */
304 static void setup_timer()
305 {
306 	static struct sigaction sigalarm_action = {
307 		.sa_handler = sigalarm_handler,
308 	};
309 	struct itimerval timer_settings = {};
310 	int err;
311 
312 	last_time_ns = get_time_ns();
313 	err = sigaction(SIGALRM, &sigalarm_action, NULL);
314 	if (err < 0) {
315 		fprintf(stderr, "failed to install SIGALRM handler: %d\n", -errno);
316 		exit(1);
317 	}
318 	timer_settings.it_interval.tv_sec = 1;
319 	timer_settings.it_value.tv_sec = 1;
320 	err = setitimer(ITIMER_REAL, &timer_settings, NULL);
321 	if (err < 0) {
322 		fprintf(stderr, "failed to arm interval timer: %d\n", -errno);
323 		exit(1);
324 	}
325 }
326 
327 static void set_thread_affinity(pthread_t thread, int cpu)
328 {
329 	cpu_set_t cpuset;
330 
331 	CPU_ZERO(&cpuset);
332 	CPU_SET(cpu, &cpuset);
333 	if (pthread_setaffinity_np(thread, sizeof(cpuset), &cpuset)) {
334 		fprintf(stderr, "setting affinity to CPU #%d failed: %d\n",
335 			cpu, errno);
336 		exit(1);
337 	}
338 }
339 
340 static int next_cpu(struct cpu_set *cpu_set)
341 {
342 	if (cpu_set->cpus) {
343 		int i;
344 
345 		/* find next available CPU */
346 		for (i = cpu_set->next_cpu; i < cpu_set->cpus_len; i++) {
347 			if (cpu_set->cpus[i]) {
348 				cpu_set->next_cpu = i + 1;
349 				return i;
350 			}
351 		}
352 		fprintf(stderr, "Not enough CPUs specified, need CPU #%d or higher.\n", i);
353 		exit(1);
354 	}
355 
356 	return cpu_set->next_cpu++;
357 }
358 
359 static struct bench_state {
360 	int res_cnt;
361 	struct bench_res *results;
362 	pthread_t *consumers;
363 	pthread_t *producers;
364 } state;
365 
366 const struct bench *bench = NULL;
367 
368 extern const struct bench bench_count_global;
369 extern const struct bench bench_count_local;
370 extern const struct bench bench_rename_base;
371 extern const struct bench bench_rename_kprobe;
372 extern const struct bench bench_rename_kretprobe;
373 extern const struct bench bench_rename_rawtp;
374 extern const struct bench bench_rename_fentry;
375 extern const struct bench bench_rename_fexit;
376 extern const struct bench bench_trig_base;
377 extern const struct bench bench_trig_tp;
378 extern const struct bench bench_trig_rawtp;
379 extern const struct bench bench_trig_kprobe;
380 extern const struct bench bench_trig_fentry;
381 extern const struct bench bench_trig_fentry_sleep;
382 extern const struct bench bench_trig_fmodret;
383 extern const struct bench bench_trig_uprobe_base;
384 extern const struct bench bench_trig_uprobe_with_nop;
385 extern const struct bench bench_trig_uretprobe_with_nop;
386 extern const struct bench bench_trig_uprobe_without_nop;
387 extern const struct bench bench_trig_uretprobe_without_nop;
388 extern const struct bench bench_rb_libbpf;
389 extern const struct bench bench_rb_custom;
390 extern const struct bench bench_pb_libbpf;
391 extern const struct bench bench_pb_custom;
392 extern const struct bench bench_bloom_lookup;
393 extern const struct bench bench_bloom_update;
394 extern const struct bench bench_bloom_false_positive;
395 extern const struct bench bench_hashmap_without_bloom;
396 extern const struct bench bench_hashmap_with_bloom;
397 extern const struct bench bench_bpf_loop;
398 extern const struct bench bench_strncmp_no_helper;
399 extern const struct bench bench_strncmp_helper;
400 
401 static const struct bench *benchs[] = {
402 	&bench_count_global,
403 	&bench_count_local,
404 	&bench_rename_base,
405 	&bench_rename_kprobe,
406 	&bench_rename_kretprobe,
407 	&bench_rename_rawtp,
408 	&bench_rename_fentry,
409 	&bench_rename_fexit,
410 	&bench_trig_base,
411 	&bench_trig_tp,
412 	&bench_trig_rawtp,
413 	&bench_trig_kprobe,
414 	&bench_trig_fentry,
415 	&bench_trig_fentry_sleep,
416 	&bench_trig_fmodret,
417 	&bench_trig_uprobe_base,
418 	&bench_trig_uprobe_with_nop,
419 	&bench_trig_uretprobe_with_nop,
420 	&bench_trig_uprobe_without_nop,
421 	&bench_trig_uretprobe_without_nop,
422 	&bench_rb_libbpf,
423 	&bench_rb_custom,
424 	&bench_pb_libbpf,
425 	&bench_pb_custom,
426 	&bench_bloom_lookup,
427 	&bench_bloom_update,
428 	&bench_bloom_false_positive,
429 	&bench_hashmap_without_bloom,
430 	&bench_hashmap_with_bloom,
431 	&bench_bpf_loop,
432 	&bench_strncmp_no_helper,
433 	&bench_strncmp_helper,
434 };
435 
436 static void setup_benchmark()
437 {
438 	int i, err;
439 
440 	if (!env.bench_name) {
441 		fprintf(stderr, "benchmark name is not specified\n");
442 		exit(1);
443 	}
444 
445 	for (i = 0; i < ARRAY_SIZE(benchs); i++) {
446 		if (strcmp(benchs[i]->name, env.bench_name) == 0) {
447 			bench = benchs[i];
448 			break;
449 		}
450 	}
451 	if (!bench) {
452 		fprintf(stderr, "benchmark '%s' not found\n", env.bench_name);
453 		exit(1);
454 	}
455 
456 	printf("Setting up benchmark '%s'...\n", bench->name);
457 
458 	state.producers = calloc(env.producer_cnt, sizeof(*state.producers));
459 	state.consumers = calloc(env.consumer_cnt, sizeof(*state.consumers));
460 	state.results = calloc(env.duration_sec + env.warmup_sec + 2,
461 			       sizeof(*state.results));
462 	if (!state.producers || !state.consumers || !state.results)
463 		exit(1);
464 
465 	if (bench->validate)
466 		bench->validate();
467 	if (bench->setup)
468 		bench->setup();
469 
470 	for (i = 0; i < env.consumer_cnt; i++) {
471 		err = pthread_create(&state.consumers[i], NULL,
472 				     bench->consumer_thread, (void *)(long)i);
473 		if (err) {
474 			fprintf(stderr, "failed to create consumer thread #%d: %d\n",
475 				i, -errno);
476 			exit(1);
477 		}
478 		if (env.affinity)
479 			set_thread_affinity(state.consumers[i],
480 					    next_cpu(&env.cons_cpus));
481 	}
482 
483 	/* unless explicit producer CPU list is specified, continue after
484 	 * last consumer CPU
485 	 */
486 	if (!env.prod_cpus.cpus)
487 		env.prod_cpus.next_cpu = env.cons_cpus.next_cpu;
488 
489 	for (i = 0; i < env.producer_cnt; i++) {
490 		err = pthread_create(&state.producers[i], NULL,
491 				     bench->producer_thread, (void *)(long)i);
492 		if (err) {
493 			fprintf(stderr, "failed to create producer thread #%d: %d\n",
494 				i, -errno);
495 			exit(1);
496 		}
497 		if (env.affinity)
498 			set_thread_affinity(state.producers[i],
499 					    next_cpu(&env.prod_cpus));
500 	}
501 
502 	printf("Benchmark '%s' started.\n", bench->name);
503 }
504 
505 static pthread_mutex_t bench_done_mtx = PTHREAD_MUTEX_INITIALIZER;
506 static pthread_cond_t bench_done = PTHREAD_COND_INITIALIZER;
507 
508 static void collect_measurements(long delta_ns) {
509 	int iter = state.res_cnt++;
510 	struct bench_res *res = &state.results[iter];
511 
512 	bench->measure(res);
513 
514 	if (bench->report_progress)
515 		bench->report_progress(iter, res, delta_ns);
516 
517 	if (iter == env.duration_sec + env.warmup_sec) {
518 		pthread_mutex_lock(&bench_done_mtx);
519 		pthread_cond_signal(&bench_done);
520 		pthread_mutex_unlock(&bench_done_mtx);
521 	}
522 }
523 
524 int main(int argc, char **argv)
525 {
526 	parse_cmdline_args(argc, argv);
527 
528 	if (env.list) {
529 		int i;
530 
531 		printf("Available benchmarks:\n");
532 		for (i = 0; i < ARRAY_SIZE(benchs); i++) {
533 			printf("- %s\n", benchs[i]->name);
534 		}
535 		return 0;
536 	}
537 
538 	setup_benchmark();
539 
540 	setup_timer();
541 
542 	pthread_mutex_lock(&bench_done_mtx);
543 	pthread_cond_wait(&bench_done, &bench_done_mtx);
544 	pthread_mutex_unlock(&bench_done_mtx);
545 
546 	if (bench->report_final)
547 		/* skip first sample */
548 		bench->report_final(state.results + env.warmup_sec,
549 				    state.res_cnt - env.warmup_sec);
550 
551 	return 0;
552 }
553