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 "bpf_util.h"
14#include "testing_helpers.h"
15
16struct 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
25static 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
33void setup_libbpf(void)
34{
35 libbpf_set_strict_mode(LIBBPF_STRICT_ALL);
36 libbpf_set_print(libbpf_print_fn);
37}
38
39void 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
50void 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
68void 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
84void
85grace_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
103void
104grace_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
121void 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
159void 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
171void 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
192void 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
207void 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
239const char *argp_program_version = "benchmark";
240const char *argp_program_bug_address = "<bpf@vger.kernel.org>";
241const 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
254enum {
255 ARG_PROD_AFFINITY_SET = 1000,
256 ARG_CONS_AFFINITY_SET = 1001,
257};
258
259static 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
275extern struct argp bench_ringbufs_argp;
276extern struct argp bench_bloom_map_argp;
277extern struct argp bench_bpf_loop_argp;
278extern struct argp bench_local_storage_argp;
279extern struct argp bench_local_storage_rcu_tasks_trace_argp;
280extern struct argp bench_strncmp_argp;
281extern struct argp bench_hashmap_lookup_argp;
282extern struct argp bench_local_storage_create_argp;
283extern struct argp bench_htab_mem_argp;
284extern struct argp bench_trigger_batch_argp;
285extern struct argp bench_crypto_argp;
286
287static 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 */
304static int pos_args;
305
306static 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
379static 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
391static 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
411static void collect_measurements(long delta_ns);
412
413static __u64 last_time_ns;
414static 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 */
425static 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
448static 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
463static 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
482static struct bench_state {
483 int res_cnt;
484 struct bench_res *results;
485 pthread_t *consumers;
486 pthread_t *producers;
487} state;
488
489const struct bench *bench = NULL;
490
491extern const struct bench bench_count_global;
492extern const struct bench bench_count_local;
493extern const struct bench bench_rename_base;
494extern const struct bench bench_rename_kprobe;
495extern const struct bench bench_rename_kretprobe;
496extern const struct bench bench_rename_rawtp;
497extern const struct bench bench_rename_fentry;
498extern const struct bench bench_rename_fexit;
499
500/* pure counting benchmarks to establish theoretical lmits */
501extern const struct bench bench_trig_usermode_count;
502extern const struct bench bench_trig_syscall_count;
503extern const struct bench bench_trig_kernel_count;
504
505/* batched, staying mostly in-kernel benchmarks */
506extern const struct bench bench_trig_kprobe;
507extern const struct bench bench_trig_kretprobe;
508extern const struct bench bench_trig_kprobe_multi;
509extern const struct bench bench_trig_kretprobe_multi;
510extern const struct bench bench_trig_fentry;
511extern const struct bench bench_trig_fexit;
512extern const struct bench bench_trig_fmodret;
513extern const struct bench bench_trig_tp;
514extern const struct bench bench_trig_rawtp;
515
516/* uprobe/uretprobe benchmarks */
517extern const struct bench bench_trig_uprobe_nop;
518extern const struct bench bench_trig_uretprobe_nop;
519extern const struct bench bench_trig_uprobe_push;
520extern const struct bench bench_trig_uretprobe_push;
521extern const struct bench bench_trig_uprobe_ret;
522extern const struct bench bench_trig_uretprobe_ret;
523extern const struct bench bench_trig_uprobe_multi_nop;
524extern const struct bench bench_trig_uretprobe_multi_nop;
525extern const struct bench bench_trig_uprobe_multi_push;
526extern const struct bench bench_trig_uretprobe_multi_push;
527extern const struct bench bench_trig_uprobe_multi_ret;
528extern const struct bench bench_trig_uretprobe_multi_ret;
529
530extern const struct bench bench_rb_libbpf;
531extern const struct bench bench_rb_custom;
532extern const struct bench bench_pb_libbpf;
533extern const struct bench bench_pb_custom;
534extern const struct bench bench_bloom_lookup;
535extern const struct bench bench_bloom_update;
536extern const struct bench bench_bloom_false_positive;
537extern const struct bench bench_hashmap_without_bloom;
538extern const struct bench bench_hashmap_with_bloom;
539extern const struct bench bench_bpf_loop;
540extern const struct bench bench_strncmp_no_helper;
541extern const struct bench bench_strncmp_helper;
542extern const struct bench bench_bpf_hashmap_full_update;
543extern const struct bench bench_local_storage_cache_seq_get;
544extern const struct bench bench_local_storage_cache_interleaved_get;
545extern const struct bench bench_local_storage_cache_hashmap_control;
546extern const struct bench bench_local_storage_tasks_trace;
547extern const struct bench bench_bpf_hashmap_lookup;
548extern const struct bench bench_local_storage_create;
549extern const struct bench bench_htab_mem;
550extern const struct bench bench_crypto_encrypt;
551extern const struct bench bench_crypto_decrypt;
552
553static 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
614static 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
634static 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
697static pthread_mutex_t bench_done_mtx = PTHREAD_MUTEX_INITIALIZER;
698static pthread_cond_t bench_done = PTHREAD_COND_INITIALIZER;
699
700static 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
716int 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}
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_kretprobe;
499extern const struct bench bench_trig_kprobe_multi;
500extern const struct bench bench_trig_kretprobe_multi;
501extern const struct bench bench_trig_fentry;
502extern const struct bench bench_trig_fexit;
503extern const struct bench bench_trig_fentry_sleep;
504extern const struct bench bench_trig_fmodret;
505extern const struct bench bench_trig_uprobe_base;
506extern const struct bench bench_trig_uprobe_nop;
507extern const struct bench bench_trig_uretprobe_nop;
508extern const struct bench bench_trig_uprobe_push;
509extern const struct bench bench_trig_uretprobe_push;
510extern const struct bench bench_trig_uprobe_ret;
511extern const struct bench bench_trig_uretprobe_ret;
512extern const struct bench bench_rb_libbpf;
513extern const struct bench bench_rb_custom;
514extern const struct bench bench_pb_libbpf;
515extern const struct bench bench_pb_custom;
516extern const struct bench bench_bloom_lookup;
517extern const struct bench bench_bloom_update;
518extern const struct bench bench_bloom_false_positive;
519extern const struct bench bench_hashmap_without_bloom;
520extern const struct bench bench_hashmap_with_bloom;
521extern const struct bench bench_bpf_loop;
522extern const struct bench bench_strncmp_no_helper;
523extern const struct bench bench_strncmp_helper;
524extern const struct bench bench_bpf_hashmap_full_update;
525extern const struct bench bench_local_storage_cache_seq_get;
526extern const struct bench bench_local_storage_cache_interleaved_get;
527extern const struct bench bench_local_storage_cache_hashmap_control;
528extern const struct bench bench_local_storage_tasks_trace;
529extern const struct bench bench_bpf_hashmap_lookup;
530extern const struct bench bench_local_storage_create;
531extern const struct bench bench_htab_mem;
532
533static const struct bench *benchs[] = {
534 &bench_count_global,
535 &bench_count_local,
536 &bench_rename_base,
537 &bench_rename_kprobe,
538 &bench_rename_kretprobe,
539 &bench_rename_rawtp,
540 &bench_rename_fentry,
541 &bench_rename_fexit,
542 &bench_trig_base,
543 &bench_trig_tp,
544 &bench_trig_rawtp,
545 &bench_trig_kprobe,
546 &bench_trig_kretprobe,
547 &bench_trig_kprobe_multi,
548 &bench_trig_kretprobe_multi,
549 &bench_trig_fentry,
550 &bench_trig_fexit,
551 &bench_trig_fentry_sleep,
552 &bench_trig_fmodret,
553 &bench_trig_uprobe_base,
554 &bench_trig_uprobe_nop,
555 &bench_trig_uretprobe_nop,
556 &bench_trig_uprobe_push,
557 &bench_trig_uretprobe_push,
558 &bench_trig_uprobe_ret,
559 &bench_trig_uretprobe_ret,
560 &bench_rb_libbpf,
561 &bench_rb_custom,
562 &bench_pb_libbpf,
563 &bench_pb_custom,
564 &bench_bloom_lookup,
565 &bench_bloom_update,
566 &bench_bloom_false_positive,
567 &bench_hashmap_without_bloom,
568 &bench_hashmap_with_bloom,
569 &bench_bpf_loop,
570 &bench_strncmp_no_helper,
571 &bench_strncmp_helper,
572 &bench_bpf_hashmap_full_update,
573 &bench_local_storage_cache_seq_get,
574 &bench_local_storage_cache_interleaved_get,
575 &bench_local_storage_cache_hashmap_control,
576 &bench_local_storage_tasks_trace,
577 &bench_bpf_hashmap_lookup,
578 &bench_local_storage_create,
579 &bench_htab_mem,
580};
581
582static void find_benchmark(void)
583{
584 int i;
585
586 if (!env.bench_name) {
587 fprintf(stderr, "benchmark name is not specified\n");
588 exit(1);
589 }
590 for (i = 0; i < ARRAY_SIZE(benchs); i++) {
591 if (strcmp(benchs[i]->name, env.bench_name) == 0) {
592 bench = benchs[i];
593 break;
594 }
595 }
596 if (!bench) {
597 fprintf(stderr, "benchmark '%s' not found\n", env.bench_name);
598 exit(1);
599 }
600}
601
602static void setup_benchmark(void)
603{
604 int i, err;
605
606 if (!env.quiet)
607 printf("Setting up benchmark '%s'...\n", bench->name);
608
609 state.producers = calloc(env.producer_cnt, sizeof(*state.producers));
610 state.consumers = calloc(env.consumer_cnt, sizeof(*state.consumers));
611 state.results = calloc(env.duration_sec + env.warmup_sec + 2,
612 sizeof(*state.results));
613 if (!state.producers || !state.consumers || !state.results)
614 exit(1);
615
616 if (bench->validate)
617 bench->validate();
618 if (bench->setup)
619 bench->setup();
620
621 for (i = 0; i < env.consumer_cnt; i++) {
622 if (!bench->consumer_thread) {
623 fprintf(stderr, "benchmark doesn't support consumers!\n");
624 exit(1);
625 }
626 err = pthread_create(&state.consumers[i], NULL,
627 bench->consumer_thread, (void *)(long)i);
628 if (err) {
629 fprintf(stderr, "failed to create consumer thread #%d: %d\n",
630 i, -err);
631 exit(1);
632 }
633 if (env.affinity)
634 set_thread_affinity(state.consumers[i],
635 next_cpu(&env.cons_cpus));
636 }
637
638 /* unless explicit producer CPU list is specified, continue after
639 * last consumer CPU
640 */
641 if (!env.prod_cpus.cpus)
642 env.prod_cpus.next_cpu = env.cons_cpus.next_cpu;
643
644 for (i = 0; i < env.producer_cnt; i++) {
645 if (!bench->producer_thread) {
646 fprintf(stderr, "benchmark doesn't support producers!\n");
647 exit(1);
648 }
649 err = pthread_create(&state.producers[i], NULL,
650 bench->producer_thread, (void *)(long)i);
651 if (err) {
652 fprintf(stderr, "failed to create producer thread #%d: %d\n",
653 i, -err);
654 exit(1);
655 }
656 if (env.affinity)
657 set_thread_affinity(state.producers[i],
658 next_cpu(&env.prod_cpus));
659 }
660
661 if (!env.quiet)
662 printf("Benchmark '%s' started.\n", bench->name);
663}
664
665static pthread_mutex_t bench_done_mtx = PTHREAD_MUTEX_INITIALIZER;
666static pthread_cond_t bench_done = PTHREAD_COND_INITIALIZER;
667
668static void collect_measurements(long delta_ns) {
669 int iter = state.res_cnt++;
670 struct bench_res *res = &state.results[iter];
671
672 bench->measure(res);
673
674 if (bench->report_progress)
675 bench->report_progress(iter, res, delta_ns);
676
677 if (iter == env.duration_sec + env.warmup_sec) {
678 pthread_mutex_lock(&bench_done_mtx);
679 pthread_cond_signal(&bench_done);
680 pthread_mutex_unlock(&bench_done_mtx);
681 }
682}
683
684int main(int argc, char **argv)
685{
686 env.nr_cpus = get_nprocs();
687 parse_cmdline_args_init(argc, argv);
688
689 if (env.list) {
690 int i;
691
692 printf("Available benchmarks:\n");
693 for (i = 0; i < ARRAY_SIZE(benchs); i++) {
694 printf("- %s\n", benchs[i]->name);
695 }
696 return 0;
697 }
698
699 find_benchmark();
700 parse_cmdline_args_final(argc, argv);
701
702 setup_benchmark();
703
704 setup_timer();
705
706 pthread_mutex_lock(&bench_done_mtx);
707 pthread_cond_wait(&bench_done, &bench_done_mtx);
708 pthread_mutex_unlock(&bench_done_mtx);
709
710 if (bench->report_final)
711 /* skip first sample */
712 bench->report_final(state.results + env.warmup_sec,
713 state.res_cnt - env.warmup_sec);
714
715 return 0;
716}