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 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 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 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 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 } 102 103 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 } 116 117 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 } 126 127 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 } 134 135 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 ctx.skel->rodata->stacktrace = env.stacktrace; 150 } 151 152 static void load_ctx(void) 153 { 154 int err; 155 156 err = trigger_bench__load(ctx.skel); 157 if (err) { 158 fprintf(stderr, "failed to open skeleton\n"); 159 exit(1); 160 } 161 } 162 163 static void attach_bpf(struct bpf_program *prog) 164 { 165 struct bpf_link *link; 166 167 link = bpf_program__attach(prog); 168 if (!link) { 169 fprintf(stderr, "failed to attach program!\n"); 170 exit(1); 171 } 172 } 173 174 static void trigger_syscall_count_setup(void) 175 { 176 ctx.usermode_counters = true; 177 } 178 179 /* Batched, staying mostly in-kernel triggering setups */ 180 static void trigger_kernel_count_setup(void) 181 { 182 setup_ctx(); 183 bpf_program__set_autoload(ctx.skel->progs.trigger_driver, false); 184 bpf_program__set_autoload(ctx.skel->progs.trigger_kernel_count, true); 185 load_ctx(); 186 /* override driver program */ 187 ctx.driver_prog_fd = bpf_program__fd(ctx.skel->progs.trigger_kernel_count); 188 } 189 190 static void trigger_kprobe_setup(void) 191 { 192 setup_ctx(); 193 bpf_program__set_autoload(ctx.skel->progs.bench_trigger_kprobe, true); 194 load_ctx(); 195 attach_bpf(ctx.skel->progs.bench_trigger_kprobe); 196 } 197 198 static void trigger_kretprobe_setup(void) 199 { 200 setup_ctx(); 201 bpf_program__set_autoload(ctx.skel->progs.bench_trigger_kretprobe, true); 202 load_ctx(); 203 attach_bpf(ctx.skel->progs.bench_trigger_kretprobe); 204 } 205 206 static void trigger_kprobe_multi_setup(void) 207 { 208 setup_ctx(); 209 bpf_program__set_autoload(ctx.skel->progs.bench_trigger_kprobe_multi, true); 210 load_ctx(); 211 attach_bpf(ctx.skel->progs.bench_trigger_kprobe_multi); 212 } 213 214 static void trigger_kretprobe_multi_setup(void) 215 { 216 setup_ctx(); 217 bpf_program__set_autoload(ctx.skel->progs.bench_trigger_kretprobe_multi, true); 218 load_ctx(); 219 attach_bpf(ctx.skel->progs.bench_trigger_kretprobe_multi); 220 } 221 222 static void trigger_fentry_setup(void) 223 { 224 setup_ctx(); 225 bpf_program__set_autoload(ctx.skel->progs.bench_trigger_fentry, true); 226 load_ctx(); 227 attach_bpf(ctx.skel->progs.bench_trigger_fentry); 228 } 229 230 static void attach_ksyms_all(struct bpf_program *empty, bool kretprobe) 231 { 232 LIBBPF_OPTS(bpf_kprobe_multi_opts, opts); 233 char **syms = NULL; 234 size_t cnt = 0; 235 236 /* Some recursive functions will be skipped in 237 * bpf_get_ksyms -> skip_entry, as they can introduce sufficient 238 * overhead. However, it's difficut to skip all the recursive 239 * functions for a debug kernel. 240 * 241 * So, don't run the kprobe-multi-all and kretprobe-multi-all on 242 * a debug kernel. 243 */ 244 if (bpf_get_ksyms(&syms, &cnt, true)) { 245 fprintf(stderr, "failed to get ksyms\n"); 246 exit(1); 247 } 248 249 opts.syms = (const char **) syms; 250 opts.cnt = cnt; 251 opts.retprobe = kretprobe; 252 /* attach empty to all the kernel functions except bpf_get_numa_node_id. */ 253 if (!bpf_program__attach_kprobe_multi_opts(empty, NULL, &opts)) { 254 fprintf(stderr, "failed to attach bpf_program__attach_kprobe_multi_opts to all\n"); 255 exit(1); 256 } 257 } 258 259 static void trigger_kprobe_multi_all_setup(void) 260 { 261 struct bpf_program *prog, *empty; 262 263 setup_ctx(); 264 empty = ctx.skel->progs.bench_kprobe_multi_empty; 265 prog = ctx.skel->progs.bench_trigger_kprobe_multi; 266 bpf_program__set_autoload(empty, true); 267 bpf_program__set_autoload(prog, true); 268 load_ctx(); 269 270 attach_ksyms_all(empty, false); 271 attach_bpf(prog); 272 } 273 274 static void trigger_kretprobe_multi_all_setup(void) 275 { 276 struct bpf_program *prog, *empty; 277 278 setup_ctx(); 279 empty = ctx.skel->progs.bench_kretprobe_multi_empty; 280 prog = ctx.skel->progs.bench_trigger_kretprobe_multi; 281 bpf_program__set_autoload(empty, true); 282 bpf_program__set_autoload(prog, true); 283 load_ctx(); 284 285 attach_ksyms_all(empty, true); 286 attach_bpf(prog); 287 } 288 289 static void trigger_fexit_setup(void) 290 { 291 setup_ctx(); 292 bpf_program__set_autoload(ctx.skel->progs.bench_trigger_fexit, true); 293 load_ctx(); 294 attach_bpf(ctx.skel->progs.bench_trigger_fexit); 295 } 296 297 static void trigger_fmodret_setup(void) 298 { 299 setup_ctx(); 300 bpf_program__set_autoload(ctx.skel->progs.trigger_driver, false); 301 bpf_program__set_autoload(ctx.skel->progs.trigger_driver_kfunc, true); 302 bpf_program__set_autoload(ctx.skel->progs.bench_trigger_fmodret, true); 303 load_ctx(); 304 /* override driver program */ 305 ctx.driver_prog_fd = bpf_program__fd(ctx.skel->progs.trigger_driver_kfunc); 306 attach_bpf(ctx.skel->progs.bench_trigger_fmodret); 307 } 308 309 static void trigger_tp_setup(void) 310 { 311 setup_ctx(); 312 bpf_program__set_autoload(ctx.skel->progs.trigger_driver, false); 313 bpf_program__set_autoload(ctx.skel->progs.trigger_driver_kfunc, true); 314 bpf_program__set_autoload(ctx.skel->progs.bench_trigger_tp, true); 315 load_ctx(); 316 /* override driver program */ 317 ctx.driver_prog_fd = bpf_program__fd(ctx.skel->progs.trigger_driver_kfunc); 318 attach_bpf(ctx.skel->progs.bench_trigger_tp); 319 } 320 321 static void trigger_rawtp_setup(void) 322 { 323 setup_ctx(); 324 bpf_program__set_autoload(ctx.skel->progs.trigger_driver, false); 325 bpf_program__set_autoload(ctx.skel->progs.trigger_driver_kfunc, true); 326 bpf_program__set_autoload(ctx.skel->progs.bench_trigger_rawtp, true); 327 load_ctx(); 328 /* override driver program */ 329 ctx.driver_prog_fd = bpf_program__fd(ctx.skel->progs.trigger_driver_kfunc); 330 attach_bpf(ctx.skel->progs.bench_trigger_rawtp); 331 } 332 333 /* make sure call is not inlined and not avoided by compiler, so __weak and 334 * inline asm volatile in the body of the function 335 * 336 * There is a performance difference between uprobing at nop location vs other 337 * instructions. So use two different targets, one of which starts with nop 338 * and another doesn't. 339 * 340 * GCC doesn't generate stack setup preamble for these functions due to them 341 * having no input arguments and doing nothing in the body. 342 */ 343 __nocf_check __weak void uprobe_target_nop(void) 344 { 345 asm volatile ("nop"); 346 } 347 348 __weak void opaque_noop_func(void) 349 { 350 } 351 352 __nocf_check __weak int uprobe_target_push(void) 353 { 354 /* overhead of function call is negligible compared to uprobe 355 * triggering, so this shouldn't affect benchmark results much 356 */ 357 opaque_noop_func(); 358 return 1; 359 } 360 361 __nocf_check __weak void uprobe_target_ret(void) 362 { 363 asm volatile (""); 364 } 365 366 static void *uprobe_producer_count(void *input) 367 { 368 while (true) { 369 uprobe_target_nop(); 370 inc_counter(base_hits); 371 } 372 return NULL; 373 } 374 375 static void *uprobe_producer_nop(void *input) 376 { 377 while (true) 378 uprobe_target_nop(); 379 return NULL; 380 } 381 382 static void *uprobe_producer_push(void *input) 383 { 384 while (true) 385 uprobe_target_push(); 386 return NULL; 387 } 388 389 static void *uprobe_producer_ret(void *input) 390 { 391 while (true) 392 uprobe_target_ret(); 393 return NULL; 394 } 395 396 #ifdef __x86_64__ 397 __nocf_check __weak void uprobe_target_nop5(void) 398 { 399 asm volatile (".byte 0x0f, 0x1f, 0x44, 0x00, 0x00"); 400 } 401 402 static void *uprobe_producer_nop5(void *input) 403 { 404 while (true) 405 uprobe_target_nop5(); 406 return NULL; 407 } 408 #endif 409 410 static void usetup(bool use_retprobe, bool use_multi, void *target_addr) 411 { 412 size_t uprobe_offset; 413 struct bpf_link *link; 414 int err; 415 416 setup_libbpf(); 417 418 ctx.skel = trigger_bench__open(); 419 if (!ctx.skel) { 420 fprintf(stderr, "failed to open skeleton\n"); 421 exit(1); 422 } 423 424 if (use_multi) 425 bpf_program__set_autoload(ctx.skel->progs.bench_trigger_uprobe_multi, true); 426 else 427 bpf_program__set_autoload(ctx.skel->progs.bench_trigger_uprobe, true); 428 429 err = trigger_bench__load(ctx.skel); 430 if (err) { 431 fprintf(stderr, "failed to load skeleton\n"); 432 exit(1); 433 } 434 435 uprobe_offset = get_uprobe_offset(target_addr); 436 if (use_multi) { 437 LIBBPF_OPTS(bpf_uprobe_multi_opts, opts, 438 .retprobe = use_retprobe, 439 .cnt = 1, 440 .offsets = &uprobe_offset, 441 ); 442 link = bpf_program__attach_uprobe_multi( 443 ctx.skel->progs.bench_trigger_uprobe_multi, 444 -1 /* all PIDs */, "/proc/self/exe", NULL, &opts); 445 ctx.skel->links.bench_trigger_uprobe_multi = link; 446 } else { 447 link = bpf_program__attach_uprobe(ctx.skel->progs.bench_trigger_uprobe, 448 use_retprobe, 449 -1 /* all PIDs */, 450 "/proc/self/exe", 451 uprobe_offset); 452 ctx.skel->links.bench_trigger_uprobe = link; 453 } 454 if (!link) { 455 fprintf(stderr, "failed to attach %s!\n", use_multi ? "multi-uprobe" : "uprobe"); 456 exit(1); 457 } 458 } 459 460 static void usermode_count_setup(void) 461 { 462 ctx.usermode_counters = true; 463 } 464 465 static void uprobe_nop_setup(void) 466 { 467 usetup(false, false /* !use_multi */, &uprobe_target_nop); 468 } 469 470 static void uretprobe_nop_setup(void) 471 { 472 usetup(true, false /* !use_multi */, &uprobe_target_nop); 473 } 474 475 static void uprobe_push_setup(void) 476 { 477 usetup(false, false /* !use_multi */, &uprobe_target_push); 478 } 479 480 static void uretprobe_push_setup(void) 481 { 482 usetup(true, false /* !use_multi */, &uprobe_target_push); 483 } 484 485 static void uprobe_ret_setup(void) 486 { 487 usetup(false, false /* !use_multi */, &uprobe_target_ret); 488 } 489 490 static void uretprobe_ret_setup(void) 491 { 492 usetup(true, false /* !use_multi */, &uprobe_target_ret); 493 } 494 495 static void uprobe_multi_nop_setup(void) 496 { 497 usetup(false, true /* use_multi */, &uprobe_target_nop); 498 } 499 500 static void uretprobe_multi_nop_setup(void) 501 { 502 usetup(true, true /* use_multi */, &uprobe_target_nop); 503 } 504 505 static void uprobe_multi_push_setup(void) 506 { 507 usetup(false, true /* use_multi */, &uprobe_target_push); 508 } 509 510 static void uretprobe_multi_push_setup(void) 511 { 512 usetup(true, true /* use_multi */, &uprobe_target_push); 513 } 514 515 static void uprobe_multi_ret_setup(void) 516 { 517 usetup(false, true /* use_multi */, &uprobe_target_ret); 518 } 519 520 static void uretprobe_multi_ret_setup(void) 521 { 522 usetup(true, true /* use_multi */, &uprobe_target_ret); 523 } 524 525 #ifdef __x86_64__ 526 static void uprobe_nop5_setup(void) 527 { 528 usetup(false, false /* !use_multi */, &uprobe_target_nop5); 529 } 530 531 static void uretprobe_nop5_setup(void) 532 { 533 usetup(true, false /* !use_multi */, &uprobe_target_nop5); 534 } 535 536 static void uprobe_multi_nop5_setup(void) 537 { 538 usetup(false, true /* use_multi */, &uprobe_target_nop5); 539 } 540 541 static void uretprobe_multi_nop5_setup(void) 542 { 543 usetup(true, true /* use_multi */, &uprobe_target_nop5); 544 } 545 #endif 546 547 const struct bench bench_trig_syscall_count = { 548 .name = "trig-syscall-count", 549 .validate = trigger_validate, 550 .setup = trigger_syscall_count_setup, 551 .producer_thread = trigger_producer, 552 .measure = trigger_measure, 553 .report_progress = hits_drops_report_progress, 554 .report_final = hits_drops_report_final, 555 }; 556 557 /* batched (staying mostly in kernel) kprobe/fentry benchmarks */ 558 #define BENCH_TRIG_KERNEL(KIND, NAME) \ 559 const struct bench bench_trig_##KIND = { \ 560 .name = "trig-" NAME, \ 561 .setup = trigger_##KIND##_setup, \ 562 .producer_thread = trigger_producer_batch, \ 563 .measure = trigger_measure, \ 564 .report_progress = hits_drops_report_progress, \ 565 .report_final = hits_drops_report_final, \ 566 .argp = &bench_trigger_batch_argp, \ 567 } 568 569 BENCH_TRIG_KERNEL(kernel_count, "kernel-count"); 570 BENCH_TRIG_KERNEL(kprobe, "kprobe"); 571 BENCH_TRIG_KERNEL(kretprobe, "kretprobe"); 572 BENCH_TRIG_KERNEL(kprobe_multi, "kprobe-multi"); 573 BENCH_TRIG_KERNEL(kretprobe_multi, "kretprobe-multi"); 574 BENCH_TRIG_KERNEL(fentry, "fentry"); 575 BENCH_TRIG_KERNEL(kprobe_multi_all, "kprobe-multi-all"); 576 BENCH_TRIG_KERNEL(kretprobe_multi_all, "kretprobe-multi-all"); 577 BENCH_TRIG_KERNEL(fexit, "fexit"); 578 BENCH_TRIG_KERNEL(fmodret, "fmodret"); 579 BENCH_TRIG_KERNEL(tp, "tp"); 580 BENCH_TRIG_KERNEL(rawtp, "rawtp"); 581 582 /* uprobe benchmarks */ 583 #define BENCH_TRIG_USERMODE(KIND, PRODUCER, NAME) \ 584 const struct bench bench_trig_##KIND = { \ 585 .name = "trig-" NAME, \ 586 .validate = trigger_validate, \ 587 .setup = KIND##_setup, \ 588 .producer_thread = uprobe_producer_##PRODUCER, \ 589 .measure = trigger_measure, \ 590 .report_progress = hits_drops_report_progress, \ 591 .report_final = hits_drops_report_final, \ 592 } 593 594 BENCH_TRIG_USERMODE(usermode_count, count, "usermode-count"); 595 BENCH_TRIG_USERMODE(uprobe_nop, nop, "uprobe-nop"); 596 BENCH_TRIG_USERMODE(uprobe_push, push, "uprobe-push"); 597 BENCH_TRIG_USERMODE(uprobe_ret, ret, "uprobe-ret"); 598 BENCH_TRIG_USERMODE(uretprobe_nop, nop, "uretprobe-nop"); 599 BENCH_TRIG_USERMODE(uretprobe_push, push, "uretprobe-push"); 600 BENCH_TRIG_USERMODE(uretprobe_ret, ret, "uretprobe-ret"); 601 BENCH_TRIG_USERMODE(uprobe_multi_nop, nop, "uprobe-multi-nop"); 602 BENCH_TRIG_USERMODE(uprobe_multi_push, push, "uprobe-multi-push"); 603 BENCH_TRIG_USERMODE(uprobe_multi_ret, ret, "uprobe-multi-ret"); 604 BENCH_TRIG_USERMODE(uretprobe_multi_nop, nop, "uretprobe-multi-nop"); 605 BENCH_TRIG_USERMODE(uretprobe_multi_push, push, "uretprobe-multi-push"); 606 BENCH_TRIG_USERMODE(uretprobe_multi_ret, ret, "uretprobe-multi-ret"); 607 #ifdef __x86_64__ 608 BENCH_TRIG_USERMODE(uprobe_nop5, nop5, "uprobe-nop5"); 609 BENCH_TRIG_USERMODE(uretprobe_nop5, nop5, "uretprobe-nop5"); 610 BENCH_TRIG_USERMODE(uprobe_multi_nop5, nop5, "uprobe-multi-nop5"); 611 BENCH_TRIG_USERMODE(uretprobe_multi_nop5, nop5, "uretprobe-multi-nop5"); 612 #endif 613