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 extern struct argp bench_ringbufs_argp;
134 
135 static const struct argp_child bench_parsers[] = {
136 	{ &bench_ringbufs_argp, 0, "Ring buffers benchmark", 0 },
137 	{},
138 };
139 
140 static error_t parse_arg(int key, char *arg, struct argp_state *state)
141 {
142 	static int pos_args;
143 
144 	switch (key) {
145 	case 'v':
146 		env.verbose = true;
147 		break;
148 	case 'l':
149 		env.list = true;
150 		break;
151 	case 'd':
152 		env.duration_sec = strtol(arg, NULL, 10);
153 		if (env.duration_sec <= 0) {
154 			fprintf(stderr, "Invalid duration: %s\n", arg);
155 			argp_usage(state);
156 		}
157 		break;
158 	case 'w':
159 		env.warmup_sec = strtol(arg, NULL, 10);
160 		if (env.warmup_sec <= 0) {
161 			fprintf(stderr, "Invalid warm-up duration: %s\n", arg);
162 			argp_usage(state);
163 		}
164 		break;
165 	case 'p':
166 		env.producer_cnt = strtol(arg, NULL, 10);
167 		if (env.producer_cnt <= 0) {
168 			fprintf(stderr, "Invalid producer count: %s\n", arg);
169 			argp_usage(state);
170 		}
171 		break;
172 	case 'c':
173 		env.consumer_cnt = strtol(arg, NULL, 10);
174 		if (env.consumer_cnt <= 0) {
175 			fprintf(stderr, "Invalid consumer count: %s\n", arg);
176 			argp_usage(state);
177 		}
178 		break;
179 	case 'a':
180 		env.affinity = true;
181 		break;
182 	case ARG_PROD_AFFINITY_SET:
183 		env.affinity = true;
184 		if (parse_num_list(arg, &env.prod_cpus.cpus,
185 				   &env.prod_cpus.cpus_len)) {
186 			fprintf(stderr, "Invalid format of CPU set for producers.");
187 			argp_usage(state);
188 		}
189 		break;
190 	case ARG_CONS_AFFINITY_SET:
191 		env.affinity = true;
192 		if (parse_num_list(arg, &env.cons_cpus.cpus,
193 				   &env.cons_cpus.cpus_len)) {
194 			fprintf(stderr, "Invalid format of CPU set for consumers.");
195 			argp_usage(state);
196 		}
197 		break;
198 	case ARGP_KEY_ARG:
199 		if (pos_args++) {
200 			fprintf(stderr,
201 				"Unrecognized positional argument: %s\n", arg);
202 			argp_usage(state);
203 		}
204 		env.bench_name = strdup(arg);
205 		break;
206 	default:
207 		return ARGP_ERR_UNKNOWN;
208 	}
209 	return 0;
210 }
211 
212 static void parse_cmdline_args(int argc, char **argv)
213 {
214 	static const struct argp argp = {
215 		.options = opts,
216 		.parser = parse_arg,
217 		.doc = argp_program_doc,
218 		.children = bench_parsers,
219 	};
220 	if (argp_parse(&argp, argc, argv, 0, NULL, NULL))
221 		exit(1);
222 	if (!env.list && !env.bench_name) {
223 		argp_help(&argp, stderr, ARGP_HELP_DOC, "bench");
224 		exit(1);
225 	}
226 }
227 
228 static void collect_measurements(long delta_ns);
229 
230 static __u64 last_time_ns;
231 static void sigalarm_handler(int signo)
232 {
233 	long new_time_ns = get_time_ns();
234 	long delta_ns = new_time_ns - last_time_ns;
235 
236 	collect_measurements(delta_ns);
237 
238 	last_time_ns = new_time_ns;
239 }
240 
241 /* set up periodic 1-second timer */
242 static void setup_timer()
243 {
244 	static struct sigaction sigalarm_action = {
245 		.sa_handler = sigalarm_handler,
246 	};
247 	struct itimerval timer_settings = {};
248 	int err;
249 
250 	last_time_ns = get_time_ns();
251 	err = sigaction(SIGALRM, &sigalarm_action, NULL);
252 	if (err < 0) {
253 		fprintf(stderr, "failed to install SIGALRM handler: %d\n", -errno);
254 		exit(1);
255 	}
256 	timer_settings.it_interval.tv_sec = 1;
257 	timer_settings.it_value.tv_sec = 1;
258 	err = setitimer(ITIMER_REAL, &timer_settings, NULL);
259 	if (err < 0) {
260 		fprintf(stderr, "failed to arm interval timer: %d\n", -errno);
261 		exit(1);
262 	}
263 }
264 
265 static void set_thread_affinity(pthread_t thread, int cpu)
266 {
267 	cpu_set_t cpuset;
268 
269 	CPU_ZERO(&cpuset);
270 	CPU_SET(cpu, &cpuset);
271 	if (pthread_setaffinity_np(thread, sizeof(cpuset), &cpuset)) {
272 		fprintf(stderr, "setting affinity to CPU #%d failed: %d\n",
273 			cpu, errno);
274 		exit(1);
275 	}
276 }
277 
278 static int next_cpu(struct cpu_set *cpu_set)
279 {
280 	if (cpu_set->cpus) {
281 		int i;
282 
283 		/* find next available CPU */
284 		for (i = cpu_set->next_cpu; i < cpu_set->cpus_len; i++) {
285 			if (cpu_set->cpus[i]) {
286 				cpu_set->next_cpu = i + 1;
287 				return i;
288 			}
289 		}
290 		fprintf(stderr, "Not enough CPUs specified, need CPU #%d or higher.\n", i);
291 		exit(1);
292 	}
293 
294 	return cpu_set->next_cpu++;
295 }
296 
297 static struct bench_state {
298 	int res_cnt;
299 	struct bench_res *results;
300 	pthread_t *consumers;
301 	pthread_t *producers;
302 } state;
303 
304 const struct bench *bench = NULL;
305 
306 extern const struct bench bench_count_global;
307 extern const struct bench bench_count_local;
308 extern const struct bench bench_rename_base;
309 extern const struct bench bench_rename_kprobe;
310 extern const struct bench bench_rename_kretprobe;
311 extern const struct bench bench_rename_rawtp;
312 extern const struct bench bench_rename_fentry;
313 extern const struct bench bench_rename_fexit;
314 extern const struct bench bench_rename_fmodret;
315 extern const struct bench bench_trig_base;
316 extern const struct bench bench_trig_tp;
317 extern const struct bench bench_trig_rawtp;
318 extern const struct bench bench_trig_kprobe;
319 extern const struct bench bench_trig_fentry;
320 extern const struct bench bench_trig_fmodret;
321 extern const struct bench bench_rb_libbpf;
322 extern const struct bench bench_rb_custom;
323 extern const struct bench bench_pb_libbpf;
324 extern const struct bench bench_pb_custom;
325 
326 static const struct bench *benchs[] = {
327 	&bench_count_global,
328 	&bench_count_local,
329 	&bench_rename_base,
330 	&bench_rename_kprobe,
331 	&bench_rename_kretprobe,
332 	&bench_rename_rawtp,
333 	&bench_rename_fentry,
334 	&bench_rename_fexit,
335 	&bench_rename_fmodret,
336 	&bench_trig_base,
337 	&bench_trig_tp,
338 	&bench_trig_rawtp,
339 	&bench_trig_kprobe,
340 	&bench_trig_fentry,
341 	&bench_trig_fmodret,
342 	&bench_rb_libbpf,
343 	&bench_rb_custom,
344 	&bench_pb_libbpf,
345 	&bench_pb_custom,
346 };
347 
348 static void setup_benchmark()
349 {
350 	int i, err;
351 
352 	if (!env.bench_name) {
353 		fprintf(stderr, "benchmark name is not specified\n");
354 		exit(1);
355 	}
356 
357 	for (i = 0; i < ARRAY_SIZE(benchs); i++) {
358 		if (strcmp(benchs[i]->name, env.bench_name) == 0) {
359 			bench = benchs[i];
360 			break;
361 		}
362 	}
363 	if (!bench) {
364 		fprintf(stderr, "benchmark '%s' not found\n", env.bench_name);
365 		exit(1);
366 	}
367 
368 	printf("Setting up benchmark '%s'...\n", bench->name);
369 
370 	state.producers = calloc(env.producer_cnt, sizeof(*state.producers));
371 	state.consumers = calloc(env.consumer_cnt, sizeof(*state.consumers));
372 	state.results = calloc(env.duration_sec + env.warmup_sec + 2,
373 			       sizeof(*state.results));
374 	if (!state.producers || !state.consumers || !state.results)
375 		exit(1);
376 
377 	if (bench->validate)
378 		bench->validate();
379 	if (bench->setup)
380 		bench->setup();
381 
382 	for (i = 0; i < env.consumer_cnt; i++) {
383 		err = pthread_create(&state.consumers[i], NULL,
384 				     bench->consumer_thread, (void *)(long)i);
385 		if (err) {
386 			fprintf(stderr, "failed to create consumer thread #%d: %d\n",
387 				i, -errno);
388 			exit(1);
389 		}
390 		if (env.affinity)
391 			set_thread_affinity(state.consumers[i],
392 					    next_cpu(&env.cons_cpus));
393 	}
394 
395 	/* unless explicit producer CPU list is specified, continue after
396 	 * last consumer CPU
397 	 */
398 	if (!env.prod_cpus.cpus)
399 		env.prod_cpus.next_cpu = env.cons_cpus.next_cpu;
400 
401 	for (i = 0; i < env.producer_cnt; i++) {
402 		err = pthread_create(&state.producers[i], NULL,
403 				     bench->producer_thread, (void *)(long)i);
404 		if (err) {
405 			fprintf(stderr, "failed to create producer thread #%d: %d\n",
406 				i, -errno);
407 			exit(1);
408 		}
409 		if (env.affinity)
410 			set_thread_affinity(state.producers[i],
411 					    next_cpu(&env.prod_cpus));
412 	}
413 
414 	printf("Benchmark '%s' started.\n", bench->name);
415 }
416 
417 static pthread_mutex_t bench_done_mtx = PTHREAD_MUTEX_INITIALIZER;
418 static pthread_cond_t bench_done = PTHREAD_COND_INITIALIZER;
419 
420 static void collect_measurements(long delta_ns) {
421 	int iter = state.res_cnt++;
422 	struct bench_res *res = &state.results[iter];
423 
424 	bench->measure(res);
425 
426 	if (bench->report_progress)
427 		bench->report_progress(iter, res, delta_ns);
428 
429 	if (iter == env.duration_sec + env.warmup_sec) {
430 		pthread_mutex_lock(&bench_done_mtx);
431 		pthread_cond_signal(&bench_done);
432 		pthread_mutex_unlock(&bench_done_mtx);
433 	}
434 }
435 
436 int main(int argc, char **argv)
437 {
438 	parse_cmdline_args(argc, argv);
439 
440 	if (env.list) {
441 		int i;
442 
443 		printf("Available benchmarks:\n");
444 		for (i = 0; i < ARRAY_SIZE(benchs); i++) {
445 			printf("- %s\n", benchs[i]->name);
446 		}
447 		return 0;
448 	}
449 
450 	setup_benchmark();
451 
452 	setup_timer();
453 
454 	pthread_mutex_lock(&bench_done_mtx);
455 	pthread_cond_wait(&bench_done, &bench_done_mtx);
456 	pthread_mutex_unlock(&bench_done_mtx);
457 
458 	if (bench->report_final)
459 		/* skip first sample */
460 		bench->report_final(state.results + env.warmup_sec,
461 				    state.res_cnt - env.warmup_sec);
462 
463 	return 0;
464 }
465 
466