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