1 #include <linux/types.h> 2 #include <stdlib.h> 3 #include <unistd.h> 4 #include <stdio.h> 5 #include <ctype.h> 6 #include <string.h> 7 8 #include "parse-events.h" 9 #include "evlist.h" 10 #include "evsel.h" 11 #include "thread_map.h" 12 #include "cpumap.h" 13 #include "machine.h" 14 #include "event.h" 15 #include "thread.h" 16 17 #include "tests.h" 18 19 #define BUFSZ 1024 20 #define READLEN 128 21 22 struct state { 23 u64 done[1024]; 24 size_t done_cnt; 25 }; 26 27 static unsigned int hex(char c) 28 { 29 if (c >= '0' && c <= '9') 30 return c - '0'; 31 if (c >= 'a' && c <= 'f') 32 return c - 'a' + 10; 33 return c - 'A' + 10; 34 } 35 36 static void read_objdump_line(const char *line, size_t line_len, void **buf, 37 size_t *len) 38 { 39 const char *p; 40 size_t i; 41 42 /* Skip to a colon */ 43 p = strchr(line, ':'); 44 if (!p) 45 return; 46 i = p + 1 - line; 47 48 /* Read bytes */ 49 while (*len) { 50 char c1, c2; 51 52 /* Skip spaces */ 53 for (; i < line_len; i++) { 54 if (!isspace(line[i])) 55 break; 56 } 57 /* Get 2 hex digits */ 58 if (i >= line_len || !isxdigit(line[i])) 59 break; 60 c1 = line[i++]; 61 if (i >= line_len || !isxdigit(line[i])) 62 break; 63 c2 = line[i++]; 64 /* Followed by a space */ 65 if (i < line_len && line[i] && !isspace(line[i])) 66 break; 67 /* Store byte */ 68 *(unsigned char *)*buf = (hex(c1) << 4) | hex(c2); 69 *buf += 1; 70 *len -= 1; 71 } 72 } 73 74 static int read_objdump_output(FILE *f, void **buf, size_t *len) 75 { 76 char *line = NULL; 77 size_t line_len; 78 ssize_t ret; 79 int err = 0; 80 81 while (1) { 82 ret = getline(&line, &line_len, f); 83 if (feof(f)) 84 break; 85 if (ret < 0) { 86 pr_debug("getline failed\n"); 87 err = -1; 88 break; 89 } 90 read_objdump_line(line, ret, buf, len); 91 } 92 93 free(line); 94 95 return err; 96 } 97 98 static int read_via_objdump(const char *filename, u64 addr, void *buf, 99 size_t len) 100 { 101 char cmd[PATH_MAX * 2]; 102 const char *fmt; 103 FILE *f; 104 int ret; 105 106 fmt = "%s -d --start-address=0x%"PRIx64" --stop-address=0x%"PRIx64" %s"; 107 ret = snprintf(cmd, sizeof(cmd), fmt, "objdump", addr, addr + len, 108 filename); 109 if (ret <= 0 || (size_t)ret >= sizeof(cmd)) 110 return -1; 111 112 pr_debug("Objdump command is: %s\n", cmd); 113 114 /* Ignore objdump errors */ 115 strcat(cmd, " 2>/dev/null"); 116 117 f = popen(cmd, "r"); 118 if (!f) { 119 pr_debug("popen failed\n"); 120 return -1; 121 } 122 123 ret = read_objdump_output(f, &buf, &len); 124 if (len) { 125 pr_debug("objdump read too few bytes\n"); 126 if (!ret) 127 ret = len; 128 } 129 130 pclose(f); 131 132 return ret; 133 } 134 135 static int read_object_code(u64 addr, size_t len, u8 cpumode, 136 struct thread *thread, struct state *state) 137 { 138 struct addr_location al; 139 unsigned char buf1[BUFSZ]; 140 unsigned char buf2[BUFSZ]; 141 size_t ret_len; 142 u64 objdump_addr; 143 int ret; 144 145 pr_debug("Reading object code for memory address: %#"PRIx64"\n", addr); 146 147 thread__find_addr_map(thread, cpumode, MAP__FUNCTION, addr, &al); 148 if (!al.map || !al.map->dso) { 149 pr_debug("thread__find_addr_map failed\n"); 150 return -1; 151 } 152 153 pr_debug("File is: %s\n", al.map->dso->long_name); 154 155 if (al.map->dso->symtab_type == DSO_BINARY_TYPE__KALLSYMS && 156 !dso__is_kcore(al.map->dso)) { 157 pr_debug("Unexpected kernel address - skipping\n"); 158 return 0; 159 } 160 161 pr_debug("On file address is: %#"PRIx64"\n", al.addr); 162 163 if (len > BUFSZ) 164 len = BUFSZ; 165 166 /* Do not go off the map */ 167 if (addr + len > al.map->end) 168 len = al.map->end - addr; 169 170 /* Read the object code using perf */ 171 ret_len = dso__data_read_offset(al.map->dso, thread->mg->machine, 172 al.addr, buf1, len); 173 if (ret_len != len) { 174 pr_debug("dso__data_read_offset failed\n"); 175 return -1; 176 } 177 178 /* 179 * Converting addresses for use by objdump requires more information. 180 * map__load() does that. See map__rip_2objdump() for details. 181 */ 182 if (map__load(al.map, NULL)) 183 return -1; 184 185 /* objdump struggles with kcore - try each map only once */ 186 if (dso__is_kcore(al.map->dso)) { 187 size_t d; 188 189 for (d = 0; d < state->done_cnt; d++) { 190 if (state->done[d] == al.map->start) { 191 pr_debug("kcore map tested already"); 192 pr_debug(" - skipping\n"); 193 return 0; 194 } 195 } 196 if (state->done_cnt >= ARRAY_SIZE(state->done)) { 197 pr_debug("Too many kcore maps - skipping\n"); 198 return 0; 199 } 200 state->done[state->done_cnt++] = al.map->start; 201 } 202 203 /* Read the object code using objdump */ 204 objdump_addr = map__rip_2objdump(al.map, al.addr); 205 ret = read_via_objdump(al.map->dso->long_name, objdump_addr, buf2, len); 206 if (ret > 0) { 207 /* 208 * The kernel maps are inaccurate - assume objdump is right in 209 * that case. 210 */ 211 if (cpumode == PERF_RECORD_MISC_KERNEL || 212 cpumode == PERF_RECORD_MISC_GUEST_KERNEL) { 213 len -= ret; 214 if (len) { 215 pr_debug("Reducing len to %zu\n", len); 216 } else if (dso__is_kcore(al.map->dso)) { 217 /* 218 * objdump cannot handle very large segments 219 * that may be found in kcore. 220 */ 221 pr_debug("objdump failed for kcore"); 222 pr_debug(" - skipping\n"); 223 return 0; 224 } else { 225 return -1; 226 } 227 } 228 } 229 if (ret < 0) { 230 pr_debug("read_via_objdump failed\n"); 231 return -1; 232 } 233 234 /* The results should be identical */ 235 if (memcmp(buf1, buf2, len)) { 236 pr_debug("Bytes read differ from those read by objdump\n"); 237 return -1; 238 } 239 pr_debug("Bytes read match those read by objdump\n"); 240 241 return 0; 242 } 243 244 static int process_sample_event(struct machine *machine, 245 struct perf_evlist *evlist, 246 union perf_event *event, struct state *state) 247 { 248 struct perf_sample sample; 249 struct thread *thread; 250 u8 cpumode; 251 252 if (perf_evlist__parse_sample(evlist, event, &sample)) { 253 pr_debug("perf_evlist__parse_sample failed\n"); 254 return -1; 255 } 256 257 thread = machine__findnew_thread(machine, sample.pid, sample.tid); 258 if (!thread) { 259 pr_debug("machine__findnew_thread failed\n"); 260 return -1; 261 } 262 263 cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK; 264 265 return read_object_code(sample.ip, READLEN, cpumode, thread, state); 266 } 267 268 static int process_event(struct machine *machine, struct perf_evlist *evlist, 269 union perf_event *event, struct state *state) 270 { 271 if (event->header.type == PERF_RECORD_SAMPLE) 272 return process_sample_event(machine, evlist, event, state); 273 274 if (event->header.type == PERF_RECORD_THROTTLE || 275 event->header.type == PERF_RECORD_UNTHROTTLE) 276 return 0; 277 278 if (event->header.type < PERF_RECORD_MAX) { 279 int ret; 280 281 ret = machine__process_event(machine, event, NULL); 282 if (ret < 0) 283 pr_debug("machine__process_event failed, event type %u\n", 284 event->header.type); 285 return ret; 286 } 287 288 return 0; 289 } 290 291 static int process_events(struct machine *machine, struct perf_evlist *evlist, 292 struct state *state) 293 { 294 union perf_event *event; 295 int i, ret; 296 297 for (i = 0; i < evlist->nr_mmaps; i++) { 298 while ((event = perf_evlist__mmap_read(evlist, i)) != NULL) { 299 ret = process_event(machine, evlist, event, state); 300 perf_evlist__mmap_consume(evlist, i); 301 if (ret < 0) 302 return ret; 303 } 304 } 305 return 0; 306 } 307 308 static int comp(const void *a, const void *b) 309 { 310 return *(int *)a - *(int *)b; 311 } 312 313 static void do_sort_something(void) 314 { 315 int buf[40960], i; 316 317 for (i = 0; i < (int)ARRAY_SIZE(buf); i++) 318 buf[i] = ARRAY_SIZE(buf) - i - 1; 319 320 qsort(buf, ARRAY_SIZE(buf), sizeof(int), comp); 321 322 for (i = 0; i < (int)ARRAY_SIZE(buf); i++) { 323 if (buf[i] != i) { 324 pr_debug("qsort failed\n"); 325 break; 326 } 327 } 328 } 329 330 static void sort_something(void) 331 { 332 int i; 333 334 for (i = 0; i < 10; i++) 335 do_sort_something(); 336 } 337 338 static void syscall_something(void) 339 { 340 int pipefd[2]; 341 int i; 342 343 for (i = 0; i < 1000; i++) { 344 if (pipe(pipefd) < 0) { 345 pr_debug("pipe failed\n"); 346 break; 347 } 348 close(pipefd[1]); 349 close(pipefd[0]); 350 } 351 } 352 353 static void fs_something(void) 354 { 355 const char *test_file_name = "temp-perf-code-reading-test-file--"; 356 FILE *f; 357 int i; 358 359 for (i = 0; i < 1000; i++) { 360 f = fopen(test_file_name, "w+"); 361 if (f) { 362 fclose(f); 363 unlink(test_file_name); 364 } 365 } 366 } 367 368 static void do_something(void) 369 { 370 fs_something(); 371 372 sort_something(); 373 374 syscall_something(); 375 } 376 377 enum { 378 TEST_CODE_READING_OK, 379 TEST_CODE_READING_NO_VMLINUX, 380 TEST_CODE_READING_NO_KCORE, 381 TEST_CODE_READING_NO_ACCESS, 382 TEST_CODE_READING_NO_KERNEL_OBJ, 383 }; 384 385 static int do_test_code_reading(bool try_kcore) 386 { 387 struct machines machines; 388 struct machine *machine; 389 struct thread *thread; 390 struct record_opts opts = { 391 .mmap_pages = UINT_MAX, 392 .user_freq = UINT_MAX, 393 .user_interval = ULLONG_MAX, 394 .freq = 4000, 395 .target = { 396 .uses_mmap = true, 397 }, 398 }; 399 struct state state = { 400 .done_cnt = 0, 401 }; 402 struct thread_map *threads = NULL; 403 struct cpu_map *cpus = NULL; 404 struct perf_evlist *evlist = NULL; 405 struct perf_evsel *evsel = NULL; 406 int err = -1, ret; 407 pid_t pid; 408 struct map *map; 409 bool have_vmlinux, have_kcore, excl_kernel = false; 410 411 pid = getpid(); 412 413 machines__init(&machines); 414 machine = &machines.host; 415 416 ret = machine__create_kernel_maps(machine); 417 if (ret < 0) { 418 pr_debug("machine__create_kernel_maps failed\n"); 419 goto out_err; 420 } 421 422 /* Force the use of kallsyms instead of vmlinux to try kcore */ 423 if (try_kcore) 424 symbol_conf.kallsyms_name = "/proc/kallsyms"; 425 426 /* Load kernel map */ 427 map = machine->vmlinux_maps[MAP__FUNCTION]; 428 ret = map__load(map, NULL); 429 if (ret < 0) { 430 pr_debug("map__load failed\n"); 431 goto out_err; 432 } 433 have_vmlinux = dso__is_vmlinux(map->dso); 434 have_kcore = dso__is_kcore(map->dso); 435 436 /* 2nd time through we just try kcore */ 437 if (try_kcore && !have_kcore) 438 return TEST_CODE_READING_NO_KCORE; 439 440 /* No point getting kernel events if there is no kernel object */ 441 if (!have_vmlinux && !have_kcore) 442 excl_kernel = true; 443 444 threads = thread_map__new_by_tid(pid); 445 if (!threads) { 446 pr_debug("thread_map__new_by_tid failed\n"); 447 goto out_err; 448 } 449 450 ret = perf_event__synthesize_thread_map(NULL, threads, 451 perf_event__process, machine, false); 452 if (ret < 0) { 453 pr_debug("perf_event__synthesize_thread_map failed\n"); 454 goto out_err; 455 } 456 457 thread = machine__findnew_thread(machine, pid, pid); 458 if (!thread) { 459 pr_debug("machine__findnew_thread failed\n"); 460 goto out_err; 461 } 462 463 cpus = cpu_map__new(NULL); 464 if (!cpus) { 465 pr_debug("cpu_map__new failed\n"); 466 goto out_err; 467 } 468 469 while (1) { 470 const char *str; 471 472 evlist = perf_evlist__new(); 473 if (!evlist) { 474 pr_debug("perf_evlist__new failed\n"); 475 goto out_err; 476 } 477 478 perf_evlist__set_maps(evlist, cpus, threads); 479 480 if (excl_kernel) 481 str = "cycles:u"; 482 else 483 str = "cycles"; 484 pr_debug("Parsing event '%s'\n", str); 485 ret = parse_events(evlist, str, NULL); 486 if (ret < 0) { 487 pr_debug("parse_events failed\n"); 488 goto out_err; 489 } 490 491 perf_evlist__config(evlist, &opts); 492 493 evsel = perf_evlist__first(evlist); 494 495 evsel->attr.comm = 1; 496 evsel->attr.disabled = 1; 497 evsel->attr.enable_on_exec = 0; 498 499 ret = perf_evlist__open(evlist); 500 if (ret < 0) { 501 if (!excl_kernel) { 502 excl_kernel = true; 503 perf_evlist__set_maps(evlist, NULL, NULL); 504 perf_evlist__delete(evlist); 505 evlist = NULL; 506 continue; 507 } 508 pr_debug("perf_evlist__open failed\n"); 509 goto out_err; 510 } 511 break; 512 } 513 514 ret = perf_evlist__mmap(evlist, UINT_MAX, false); 515 if (ret < 0) { 516 pr_debug("perf_evlist__mmap failed\n"); 517 goto out_err; 518 } 519 520 perf_evlist__enable(evlist); 521 522 do_something(); 523 524 perf_evlist__disable(evlist); 525 526 ret = process_events(machine, evlist, &state); 527 if (ret < 0) 528 goto out_err; 529 530 if (!have_vmlinux && !have_kcore && !try_kcore) 531 err = TEST_CODE_READING_NO_KERNEL_OBJ; 532 else if (!have_vmlinux && !try_kcore) 533 err = TEST_CODE_READING_NO_VMLINUX; 534 else if (excl_kernel) 535 err = TEST_CODE_READING_NO_ACCESS; 536 else 537 err = TEST_CODE_READING_OK; 538 out_err: 539 if (evlist) { 540 perf_evlist__delete(evlist); 541 } else { 542 cpu_map__delete(cpus); 543 thread_map__delete(threads); 544 } 545 machines__destroy_kernel_maps(&machines); 546 machine__delete_threads(machine); 547 machines__exit(&machines); 548 549 return err; 550 } 551 552 int test__code_reading(void) 553 { 554 int ret; 555 556 ret = do_test_code_reading(false); 557 if (!ret) 558 ret = do_test_code_reading(true); 559 560 switch (ret) { 561 case TEST_CODE_READING_OK: 562 return 0; 563 case TEST_CODE_READING_NO_VMLINUX: 564 fprintf(stderr, " (no vmlinux)"); 565 return 0; 566 case TEST_CODE_READING_NO_KCORE: 567 fprintf(stderr, " (no kcore)"); 568 return 0; 569 case TEST_CODE_READING_NO_ACCESS: 570 fprintf(stderr, " (no access)"); 571 return 0; 572 case TEST_CODE_READING_NO_KERNEL_OBJ: 573 fprintf(stderr, " (no kernel obj)"); 574 return 0; 575 default: 576 return -1; 577 }; 578 } 579