1 // SPDX-License-Identifier: GPL-2.0
2 /* Copyright (c) 2020 Facebook */
3 #define _GNU_SOURCE
4 #include <argp.h>
5 #include <unistd.h>
6 #include <stdint.h>
7 #include "bpf_util.h"
8 #include "bench.h"
9 #include "trigger_bench.skel.h"
10 #include "trace_helpers.h"
11
12 #define MAX_TRIG_BATCH_ITERS 1000
13
14 static struct {
15 __u32 batch_iters;
16 } args = {
17 .batch_iters = 100,
18 };
19
20 enum {
21 ARG_TRIG_BATCH_ITERS = 7000,
22 };
23
24 static const struct argp_option opts[] = {
25 { "trig-batch-iters", ARG_TRIG_BATCH_ITERS, "BATCH_ITER_CNT", 0,
26 "Number of in-kernel iterations per one driver test run"},
27 {},
28 };
29
parse_arg(int key,char * arg,struct argp_state * state)30 static error_t parse_arg(int key, char *arg, struct argp_state *state)
31 {
32 long ret;
33
34 switch (key) {
35 case ARG_TRIG_BATCH_ITERS:
36 ret = strtol(arg, NULL, 10);
37 if (ret < 1 || ret > MAX_TRIG_BATCH_ITERS) {
38 fprintf(stderr, "invalid --trig-batch-iters value (should be between %d and %d)\n",
39 1, MAX_TRIG_BATCH_ITERS);
40 argp_usage(state);
41 }
42 args.batch_iters = ret;
43 break;
44 default:
45 return ARGP_ERR_UNKNOWN;
46 }
47
48 return 0;
49 }
50
51 const struct argp bench_trigger_batch_argp = {
52 .options = opts,
53 .parser = parse_arg,
54 };
55
56 /* adjust slot shift in inc_hits() if changing */
57 #define MAX_BUCKETS 256
58
59 #pragma GCC diagnostic ignored "-Wattributes"
60
61 /* BPF triggering benchmarks */
62 static struct trigger_ctx {
63 struct trigger_bench *skel;
64 bool usermode_counters;
65 int driver_prog_fd;
66 } ctx;
67
68 static struct counter base_hits[MAX_BUCKETS];
69
inc_counter(struct counter * counters)70 static __always_inline void inc_counter(struct counter *counters)
71 {
72 static __thread int tid = 0;
73 unsigned slot;
74
75 if (unlikely(tid == 0))
76 tid = sys_gettid();
77
78 /* multiplicative hashing, it's fast */
79 slot = 2654435769U * tid;
80 slot >>= 24;
81
82 atomic_inc(&base_hits[slot].value); /* use highest byte as an index */
83 }
84
sum_and_reset_counters(struct counter * counters)85 static long sum_and_reset_counters(struct counter *counters)
86 {
87 int i;
88 long sum = 0;
89
90 for (i = 0; i < MAX_BUCKETS; i++)
91 sum += atomic_swap(&counters[i].value, 0);
92 return sum;
93 }
94
trigger_validate(void)95 static void trigger_validate(void)
96 {
97 if (env.consumer_cnt != 0) {
98 fprintf(stderr, "benchmark doesn't support consumer!\n");
99 exit(1);
100 }
101 }
102
trigger_producer(void * input)103 static void *trigger_producer(void *input)
104 {
105 if (ctx.usermode_counters) {
106 while (true) {
107 (void)syscall(__NR_getpgid);
108 inc_counter(base_hits);
109 }
110 } else {
111 while (true)
112 (void)syscall(__NR_getpgid);
113 }
114 return NULL;
115 }
116
trigger_producer_batch(void * input)117 static void *trigger_producer_batch(void *input)
118 {
119 int fd = ctx.driver_prog_fd ?: bpf_program__fd(ctx.skel->progs.trigger_driver);
120
121 while (true)
122 bpf_prog_test_run_opts(fd, NULL);
123
124 return NULL;
125 }
126
trigger_measure(struct bench_res * res)127 static void trigger_measure(struct bench_res *res)
128 {
129 if (ctx.usermode_counters)
130 res->hits = sum_and_reset_counters(base_hits);
131 else
132 res->hits = sum_and_reset_counters(ctx.skel->bss->hits);
133 }
134
setup_ctx(void)135 static void setup_ctx(void)
136 {
137 setup_libbpf();
138
139 ctx.skel = trigger_bench__open();
140 if (!ctx.skel) {
141 fprintf(stderr, "failed to open skeleton\n");
142 exit(1);
143 }
144
145 /* default "driver" BPF program */
146 bpf_program__set_autoload(ctx.skel->progs.trigger_driver, true);
147
148 ctx.skel->rodata->batch_iters = args.batch_iters;
149 }
150
load_ctx(void)151 static void load_ctx(void)
152 {
153 int err;
154
155 err = trigger_bench__load(ctx.skel);
156 if (err) {
157 fprintf(stderr, "failed to open skeleton\n");
158 exit(1);
159 }
160 }
161
attach_bpf(struct bpf_program * prog)162 static void attach_bpf(struct bpf_program *prog)
163 {
164 struct bpf_link *link;
165
166 link = bpf_program__attach(prog);
167 if (!link) {
168 fprintf(stderr, "failed to attach program!\n");
169 exit(1);
170 }
171 }
172
trigger_syscall_count_setup(void)173 static void trigger_syscall_count_setup(void)
174 {
175 ctx.usermode_counters = true;
176 }
177
178 /* Batched, staying mostly in-kernel triggering setups */
trigger_kernel_count_setup(void)179 static void trigger_kernel_count_setup(void)
180 {
181 setup_ctx();
182 bpf_program__set_autoload(ctx.skel->progs.trigger_driver, false);
183 bpf_program__set_autoload(ctx.skel->progs.trigger_count, true);
184 load_ctx();
185 /* override driver program */
186 ctx.driver_prog_fd = bpf_program__fd(ctx.skel->progs.trigger_count);
187 }
188
trigger_kprobe_setup(void)189 static void trigger_kprobe_setup(void)
190 {
191 setup_ctx();
192 bpf_program__set_autoload(ctx.skel->progs.bench_trigger_kprobe, true);
193 load_ctx();
194 attach_bpf(ctx.skel->progs.bench_trigger_kprobe);
195 }
196
trigger_kretprobe_setup(void)197 static void trigger_kretprobe_setup(void)
198 {
199 setup_ctx();
200 bpf_program__set_autoload(ctx.skel->progs.bench_trigger_kretprobe, true);
201 load_ctx();
202 attach_bpf(ctx.skel->progs.bench_trigger_kretprobe);
203 }
204
trigger_kprobe_multi_setup(void)205 static void trigger_kprobe_multi_setup(void)
206 {
207 setup_ctx();
208 bpf_program__set_autoload(ctx.skel->progs.bench_trigger_kprobe_multi, true);
209 load_ctx();
210 attach_bpf(ctx.skel->progs.bench_trigger_kprobe_multi);
211 }
212
trigger_kretprobe_multi_setup(void)213 static void trigger_kretprobe_multi_setup(void)
214 {
215 setup_ctx();
216 bpf_program__set_autoload(ctx.skel->progs.bench_trigger_kretprobe_multi, true);
217 load_ctx();
218 attach_bpf(ctx.skel->progs.bench_trigger_kretprobe_multi);
219 }
220
trigger_fentry_setup(void)221 static void trigger_fentry_setup(void)
222 {
223 setup_ctx();
224 bpf_program__set_autoload(ctx.skel->progs.bench_trigger_fentry, true);
225 load_ctx();
226 attach_bpf(ctx.skel->progs.bench_trigger_fentry);
227 }
228
trigger_fexit_setup(void)229 static void trigger_fexit_setup(void)
230 {
231 setup_ctx();
232 bpf_program__set_autoload(ctx.skel->progs.bench_trigger_fexit, true);
233 load_ctx();
234 attach_bpf(ctx.skel->progs.bench_trigger_fexit);
235 }
236
trigger_fmodret_setup(void)237 static void trigger_fmodret_setup(void)
238 {
239 setup_ctx();
240 bpf_program__set_autoload(ctx.skel->progs.trigger_driver, false);
241 bpf_program__set_autoload(ctx.skel->progs.trigger_driver_kfunc, true);
242 bpf_program__set_autoload(ctx.skel->progs.bench_trigger_fmodret, true);
243 load_ctx();
244 /* override driver program */
245 ctx.driver_prog_fd = bpf_program__fd(ctx.skel->progs.trigger_driver_kfunc);
246 attach_bpf(ctx.skel->progs.bench_trigger_fmodret);
247 }
248
trigger_tp_setup(void)249 static void trigger_tp_setup(void)
250 {
251 setup_ctx();
252 bpf_program__set_autoload(ctx.skel->progs.trigger_driver, false);
253 bpf_program__set_autoload(ctx.skel->progs.trigger_driver_kfunc, true);
254 bpf_program__set_autoload(ctx.skel->progs.bench_trigger_tp, true);
255 load_ctx();
256 /* override driver program */
257 ctx.driver_prog_fd = bpf_program__fd(ctx.skel->progs.trigger_driver_kfunc);
258 attach_bpf(ctx.skel->progs.bench_trigger_tp);
259 }
260
trigger_rawtp_setup(void)261 static void trigger_rawtp_setup(void)
262 {
263 setup_ctx();
264 bpf_program__set_autoload(ctx.skel->progs.trigger_driver, false);
265 bpf_program__set_autoload(ctx.skel->progs.trigger_driver_kfunc, true);
266 bpf_program__set_autoload(ctx.skel->progs.bench_trigger_rawtp, true);
267 load_ctx();
268 /* override driver program */
269 ctx.driver_prog_fd = bpf_program__fd(ctx.skel->progs.trigger_driver_kfunc);
270 attach_bpf(ctx.skel->progs.bench_trigger_rawtp);
271 }
272
273 /* make sure call is not inlined and not avoided by compiler, so __weak and
274 * inline asm volatile in the body of the function
275 *
276 * There is a performance difference between uprobing at nop location vs other
277 * instructions. So use two different targets, one of which starts with nop
278 * and another doesn't.
279 *
280 * GCC doesn't generate stack setup preamble for these functions due to them
281 * having no input arguments and doing nothing in the body.
282 */
uprobe_target_nop(void)283 __nocf_check __weak void uprobe_target_nop(void)
284 {
285 asm volatile ("nop");
286 }
287
opaque_noop_func(void)288 __weak void opaque_noop_func(void)
289 {
290 }
291
uprobe_target_push(void)292 __nocf_check __weak int uprobe_target_push(void)
293 {
294 /* overhead of function call is negligible compared to uprobe
295 * triggering, so this shouldn't affect benchmark results much
296 */
297 opaque_noop_func();
298 return 1;
299 }
300
uprobe_target_ret(void)301 __nocf_check __weak void uprobe_target_ret(void)
302 {
303 asm volatile ("");
304 }
305
uprobe_producer_count(void * input)306 static void *uprobe_producer_count(void *input)
307 {
308 while (true) {
309 uprobe_target_nop();
310 inc_counter(base_hits);
311 }
312 return NULL;
313 }
314
uprobe_producer_nop(void * input)315 static void *uprobe_producer_nop(void *input)
316 {
317 while (true)
318 uprobe_target_nop();
319 return NULL;
320 }
321
uprobe_producer_push(void * input)322 static void *uprobe_producer_push(void *input)
323 {
324 while (true)
325 uprobe_target_push();
326 return NULL;
327 }
328
uprobe_producer_ret(void * input)329 static void *uprobe_producer_ret(void *input)
330 {
331 while (true)
332 uprobe_target_ret();
333 return NULL;
334 }
335
usetup(bool use_retprobe,bool use_multi,void * target_addr)336 static void usetup(bool use_retprobe, bool use_multi, void *target_addr)
337 {
338 size_t uprobe_offset;
339 struct bpf_link *link;
340 int err;
341
342 setup_libbpf();
343
344 ctx.skel = trigger_bench__open();
345 if (!ctx.skel) {
346 fprintf(stderr, "failed to open skeleton\n");
347 exit(1);
348 }
349
350 if (use_multi)
351 bpf_program__set_autoload(ctx.skel->progs.bench_trigger_uprobe_multi, true);
352 else
353 bpf_program__set_autoload(ctx.skel->progs.bench_trigger_uprobe, true);
354
355 err = trigger_bench__load(ctx.skel);
356 if (err) {
357 fprintf(stderr, "failed to load skeleton\n");
358 exit(1);
359 }
360
361 uprobe_offset = get_uprobe_offset(target_addr);
362 if (use_multi) {
363 LIBBPF_OPTS(bpf_uprobe_multi_opts, opts,
364 .retprobe = use_retprobe,
365 .cnt = 1,
366 .offsets = &uprobe_offset,
367 );
368 link = bpf_program__attach_uprobe_multi(
369 ctx.skel->progs.bench_trigger_uprobe_multi,
370 -1 /* all PIDs */, "/proc/self/exe", NULL, &opts);
371 ctx.skel->links.bench_trigger_uprobe_multi = link;
372 } else {
373 link = bpf_program__attach_uprobe(ctx.skel->progs.bench_trigger_uprobe,
374 use_retprobe,
375 -1 /* all PIDs */,
376 "/proc/self/exe",
377 uprobe_offset);
378 ctx.skel->links.bench_trigger_uprobe = link;
379 }
380 if (!link) {
381 fprintf(stderr, "failed to attach %s!\n", use_multi ? "multi-uprobe" : "uprobe");
382 exit(1);
383 }
384 }
385
usermode_count_setup(void)386 static void usermode_count_setup(void)
387 {
388 ctx.usermode_counters = true;
389 }
390
uprobe_nop_setup(void)391 static void uprobe_nop_setup(void)
392 {
393 usetup(false, false /* !use_multi */, &uprobe_target_nop);
394 }
395
uretprobe_nop_setup(void)396 static void uretprobe_nop_setup(void)
397 {
398 usetup(true, false /* !use_multi */, &uprobe_target_nop);
399 }
400
uprobe_push_setup(void)401 static void uprobe_push_setup(void)
402 {
403 usetup(false, false /* !use_multi */, &uprobe_target_push);
404 }
405
uretprobe_push_setup(void)406 static void uretprobe_push_setup(void)
407 {
408 usetup(true, false /* !use_multi */, &uprobe_target_push);
409 }
410
uprobe_ret_setup(void)411 static void uprobe_ret_setup(void)
412 {
413 usetup(false, false /* !use_multi */, &uprobe_target_ret);
414 }
415
uretprobe_ret_setup(void)416 static void uretprobe_ret_setup(void)
417 {
418 usetup(true, false /* !use_multi */, &uprobe_target_ret);
419 }
420
uprobe_multi_nop_setup(void)421 static void uprobe_multi_nop_setup(void)
422 {
423 usetup(false, true /* use_multi */, &uprobe_target_nop);
424 }
425
uretprobe_multi_nop_setup(void)426 static void uretprobe_multi_nop_setup(void)
427 {
428 usetup(true, true /* use_multi */, &uprobe_target_nop);
429 }
430
uprobe_multi_push_setup(void)431 static void uprobe_multi_push_setup(void)
432 {
433 usetup(false, true /* use_multi */, &uprobe_target_push);
434 }
435
uretprobe_multi_push_setup(void)436 static void uretprobe_multi_push_setup(void)
437 {
438 usetup(true, true /* use_multi */, &uprobe_target_push);
439 }
440
uprobe_multi_ret_setup(void)441 static void uprobe_multi_ret_setup(void)
442 {
443 usetup(false, true /* use_multi */, &uprobe_target_ret);
444 }
445
uretprobe_multi_ret_setup(void)446 static void uretprobe_multi_ret_setup(void)
447 {
448 usetup(true, true /* use_multi */, &uprobe_target_ret);
449 }
450
451 const struct bench bench_trig_syscall_count = {
452 .name = "trig-syscall-count",
453 .validate = trigger_validate,
454 .setup = trigger_syscall_count_setup,
455 .producer_thread = trigger_producer,
456 .measure = trigger_measure,
457 .report_progress = hits_drops_report_progress,
458 .report_final = hits_drops_report_final,
459 };
460
461 /* batched (staying mostly in kernel) kprobe/fentry benchmarks */
462 #define BENCH_TRIG_KERNEL(KIND, NAME) \
463 const struct bench bench_trig_##KIND = { \
464 .name = "trig-" NAME, \
465 .setup = trigger_##KIND##_setup, \
466 .producer_thread = trigger_producer_batch, \
467 .measure = trigger_measure, \
468 .report_progress = hits_drops_report_progress, \
469 .report_final = hits_drops_report_final, \
470 .argp = &bench_trigger_batch_argp, \
471 }
472
473 BENCH_TRIG_KERNEL(kernel_count, "kernel-count");
474 BENCH_TRIG_KERNEL(kprobe, "kprobe");
475 BENCH_TRIG_KERNEL(kretprobe, "kretprobe");
476 BENCH_TRIG_KERNEL(kprobe_multi, "kprobe-multi");
477 BENCH_TRIG_KERNEL(kretprobe_multi, "kretprobe-multi");
478 BENCH_TRIG_KERNEL(fentry, "fentry");
479 BENCH_TRIG_KERNEL(fexit, "fexit");
480 BENCH_TRIG_KERNEL(fmodret, "fmodret");
481 BENCH_TRIG_KERNEL(tp, "tp");
482 BENCH_TRIG_KERNEL(rawtp, "rawtp");
483
484 /* uprobe benchmarks */
485 #define BENCH_TRIG_USERMODE(KIND, PRODUCER, NAME) \
486 const struct bench bench_trig_##KIND = { \
487 .name = "trig-" NAME, \
488 .validate = trigger_validate, \
489 .setup = KIND##_setup, \
490 .producer_thread = uprobe_producer_##PRODUCER, \
491 .measure = trigger_measure, \
492 .report_progress = hits_drops_report_progress, \
493 .report_final = hits_drops_report_final, \
494 }
495
496 BENCH_TRIG_USERMODE(usermode_count, count, "usermode-count");
497 BENCH_TRIG_USERMODE(uprobe_nop, nop, "uprobe-nop");
498 BENCH_TRIG_USERMODE(uprobe_push, push, "uprobe-push");
499 BENCH_TRIG_USERMODE(uprobe_ret, ret, "uprobe-ret");
500 BENCH_TRIG_USERMODE(uretprobe_nop, nop, "uretprobe-nop");
501 BENCH_TRIG_USERMODE(uretprobe_push, push, "uretprobe-push");
502 BENCH_TRIG_USERMODE(uretprobe_ret, ret, "uretprobe-ret");
503 BENCH_TRIG_USERMODE(uprobe_multi_nop, nop, "uprobe-multi-nop");
504 BENCH_TRIG_USERMODE(uprobe_multi_push, push, "uprobe-multi-push");
505 BENCH_TRIG_USERMODE(uprobe_multi_ret, ret, "uprobe-multi-ret");
506 BENCH_TRIG_USERMODE(uretprobe_multi_nop, nop, "uretprobe-multi-nop");
507 BENCH_TRIG_USERMODE(uretprobe_multi_push, push, "uretprobe-multi-push");
508 BENCH_TRIG_USERMODE(uretprobe_multi_ret, ret, "uretprobe-multi-ret");
509