1 #include <errno.h> 2 #include <linux/kernel.h> 3 #include <linux/types.h> 4 #include <inttypes.h> 5 #include <stdlib.h> 6 #include <unistd.h> 7 #include <stdio.h> 8 #include <string.h> 9 #include <sys/param.h> 10 11 #include "parse-events.h" 12 #include "evlist.h" 13 #include "evsel.h" 14 #include "thread_map.h" 15 #include "cpumap.h" 16 #include "machine.h" 17 #include "event.h" 18 #include "thread.h" 19 20 #include "tests.h" 21 22 #include "sane_ctype.h" 23 24 #define BUFSZ 1024 25 #define READLEN 128 26 27 struct state { 28 u64 done[1024]; 29 size_t done_cnt; 30 }; 31 32 static unsigned int hex(char c) 33 { 34 if (c >= '0' && c <= '9') 35 return c - '0'; 36 if (c >= 'a' && c <= 'f') 37 return c - 'a' + 10; 38 return c - 'A' + 10; 39 } 40 41 static size_t read_objdump_chunk(const char **line, unsigned char **buf, 42 size_t *buf_len) 43 { 44 size_t bytes_read = 0; 45 unsigned char *chunk_start = *buf; 46 47 /* Read bytes */ 48 while (*buf_len > 0) { 49 char c1, c2; 50 51 /* Get 2 hex digits */ 52 c1 = *(*line)++; 53 if (!isxdigit(c1)) 54 break; 55 c2 = *(*line)++; 56 if (!isxdigit(c2)) 57 break; 58 59 /* Store byte and advance buf */ 60 **buf = (hex(c1) << 4) | hex(c2); 61 (*buf)++; 62 (*buf_len)--; 63 bytes_read++; 64 65 /* End of chunk? */ 66 if (isspace(**line)) 67 break; 68 } 69 70 /* 71 * objdump will display raw insn as LE if code endian 72 * is LE and bytes_per_chunk > 1. In that case reverse 73 * the chunk we just read. 74 * 75 * see disassemble_bytes() at binutils/objdump.c for details 76 * how objdump chooses display endian) 77 */ 78 if (bytes_read > 1 && !bigendian()) { 79 unsigned char *chunk_end = chunk_start + bytes_read - 1; 80 unsigned char tmp; 81 82 while (chunk_start < chunk_end) { 83 tmp = *chunk_start; 84 *chunk_start = *chunk_end; 85 *chunk_end = tmp; 86 chunk_start++; 87 chunk_end--; 88 } 89 } 90 91 return bytes_read; 92 } 93 94 static size_t read_objdump_line(const char *line, unsigned char *buf, 95 size_t buf_len) 96 { 97 const char *p; 98 size_t ret, bytes_read = 0; 99 100 /* Skip to a colon */ 101 p = strchr(line, ':'); 102 if (!p) 103 return 0; 104 p++; 105 106 /* Skip initial spaces */ 107 while (*p) { 108 if (!isspace(*p)) 109 break; 110 p++; 111 } 112 113 do { 114 ret = read_objdump_chunk(&p, &buf, &buf_len); 115 bytes_read += ret; 116 p++; 117 } while (ret > 0); 118 119 /* return number of successfully read bytes */ 120 return bytes_read; 121 } 122 123 static int read_objdump_output(FILE *f, void *buf, size_t *len, u64 start_addr) 124 { 125 char *line = NULL; 126 size_t line_len, off_last = 0; 127 ssize_t ret; 128 int err = 0; 129 u64 addr, last_addr = start_addr; 130 131 while (off_last < *len) { 132 size_t off, read_bytes, written_bytes; 133 unsigned char tmp[BUFSZ]; 134 135 ret = getline(&line, &line_len, f); 136 if (feof(f)) 137 break; 138 if (ret < 0) { 139 pr_debug("getline failed\n"); 140 err = -1; 141 break; 142 } 143 144 /* read objdump data into temporary buffer */ 145 read_bytes = read_objdump_line(line, tmp, sizeof(tmp)); 146 if (!read_bytes) 147 continue; 148 149 if (sscanf(line, "%"PRIx64, &addr) != 1) 150 continue; 151 if (addr < last_addr) { 152 pr_debug("addr going backwards, read beyond section?\n"); 153 break; 154 } 155 last_addr = addr; 156 157 /* copy it from temporary buffer to 'buf' according 158 * to address on current objdump line */ 159 off = addr - start_addr; 160 if (off >= *len) 161 break; 162 written_bytes = MIN(read_bytes, *len - off); 163 memcpy(buf + off, tmp, written_bytes); 164 off_last = off + written_bytes; 165 } 166 167 /* len returns number of bytes that could not be read */ 168 *len -= off_last; 169 170 free(line); 171 172 return err; 173 } 174 175 static int read_via_objdump(const char *filename, u64 addr, void *buf, 176 size_t len) 177 { 178 char cmd[PATH_MAX * 2]; 179 const char *fmt; 180 FILE *f; 181 int ret; 182 183 fmt = "%s -z -d --start-address=0x%"PRIx64" --stop-address=0x%"PRIx64" %s"; 184 ret = snprintf(cmd, sizeof(cmd), fmt, "objdump", addr, addr + len, 185 filename); 186 if (ret <= 0 || (size_t)ret >= sizeof(cmd)) 187 return -1; 188 189 pr_debug("Objdump command is: %s\n", cmd); 190 191 /* Ignore objdump errors */ 192 strcat(cmd, " 2>/dev/null"); 193 194 f = popen(cmd, "r"); 195 if (!f) { 196 pr_debug("popen failed\n"); 197 return -1; 198 } 199 200 ret = read_objdump_output(f, buf, &len, addr); 201 if (len) { 202 pr_debug("objdump read too few bytes: %zd\n", len); 203 if (!ret) 204 ret = len; 205 } 206 207 pclose(f); 208 209 return ret; 210 } 211 212 static void dump_buf(unsigned char *buf, size_t len) 213 { 214 size_t i; 215 216 for (i = 0; i < len; i++) { 217 pr_debug("0x%02x ", buf[i]); 218 if (i % 16 == 15) 219 pr_debug("\n"); 220 } 221 pr_debug("\n"); 222 } 223 224 static int read_object_code(u64 addr, size_t len, u8 cpumode, 225 struct thread *thread, struct state *state) 226 { 227 struct addr_location al; 228 unsigned char buf1[BUFSZ]; 229 unsigned char buf2[BUFSZ]; 230 size_t ret_len; 231 u64 objdump_addr; 232 const char *objdump_name; 233 char decomp_name[KMOD_DECOMP_LEN]; 234 int ret; 235 236 pr_debug("Reading object code for memory address: %#"PRIx64"\n", addr); 237 238 thread__find_addr_map(thread, cpumode, MAP__FUNCTION, addr, &al); 239 if (!al.map || !al.map->dso) { 240 pr_debug("thread__find_addr_map failed\n"); 241 return -1; 242 } 243 244 pr_debug("File is: %s\n", al.map->dso->long_name); 245 246 if (al.map->dso->symtab_type == DSO_BINARY_TYPE__KALLSYMS && 247 !dso__is_kcore(al.map->dso)) { 248 pr_debug("Unexpected kernel address - skipping\n"); 249 return 0; 250 } 251 252 pr_debug("On file address is: %#"PRIx64"\n", al.addr); 253 254 if (len > BUFSZ) 255 len = BUFSZ; 256 257 /* Do not go off the map */ 258 if (addr + len > al.map->end) 259 len = al.map->end - addr; 260 261 /* Read the object code using perf */ 262 ret_len = dso__data_read_offset(al.map->dso, thread->mg->machine, 263 al.addr, buf1, len); 264 if (ret_len != len) { 265 pr_debug("dso__data_read_offset failed\n"); 266 return -1; 267 } 268 269 /* 270 * Converting addresses for use by objdump requires more information. 271 * map__load() does that. See map__rip_2objdump() for details. 272 */ 273 if (map__load(al.map)) 274 return -1; 275 276 /* objdump struggles with kcore - try each map only once */ 277 if (dso__is_kcore(al.map->dso)) { 278 size_t d; 279 280 for (d = 0; d < state->done_cnt; d++) { 281 if (state->done[d] == al.map->start) { 282 pr_debug("kcore map tested already"); 283 pr_debug(" - skipping\n"); 284 return 0; 285 } 286 } 287 if (state->done_cnt >= ARRAY_SIZE(state->done)) { 288 pr_debug("Too many kcore maps - skipping\n"); 289 return 0; 290 } 291 state->done[state->done_cnt++] = al.map->start; 292 } 293 294 objdump_name = al.map->dso->long_name; 295 if (dso__needs_decompress(al.map->dso)) { 296 if (dso__decompress_kmodule_path(al.map->dso, objdump_name, 297 decomp_name, 298 sizeof(decomp_name)) < 0) { 299 pr_debug("decompression failed\n"); 300 return -1; 301 } 302 303 objdump_name = decomp_name; 304 } 305 306 /* Read the object code using objdump */ 307 objdump_addr = map__rip_2objdump(al.map, al.addr); 308 ret = read_via_objdump(objdump_name, objdump_addr, buf2, len); 309 310 if (dso__needs_decompress(al.map->dso)) 311 unlink(objdump_name); 312 313 if (ret > 0) { 314 /* 315 * The kernel maps are inaccurate - assume objdump is right in 316 * that case. 317 */ 318 if (cpumode == PERF_RECORD_MISC_KERNEL || 319 cpumode == PERF_RECORD_MISC_GUEST_KERNEL) { 320 len -= ret; 321 if (len) { 322 pr_debug("Reducing len to %zu\n", len); 323 } else if (dso__is_kcore(al.map->dso)) { 324 /* 325 * objdump cannot handle very large segments 326 * that may be found in kcore. 327 */ 328 pr_debug("objdump failed for kcore"); 329 pr_debug(" - skipping\n"); 330 return 0; 331 } else { 332 return -1; 333 } 334 } 335 } 336 if (ret < 0) { 337 pr_debug("read_via_objdump failed\n"); 338 return -1; 339 } 340 341 /* The results should be identical */ 342 if (memcmp(buf1, buf2, len)) { 343 pr_debug("Bytes read differ from those read by objdump\n"); 344 pr_debug("buf1 (dso):\n"); 345 dump_buf(buf1, len); 346 pr_debug("buf2 (objdump):\n"); 347 dump_buf(buf2, len); 348 return -1; 349 } 350 pr_debug("Bytes read match those read by objdump\n"); 351 352 return 0; 353 } 354 355 static int process_sample_event(struct machine *machine, 356 struct perf_evlist *evlist, 357 union perf_event *event, struct state *state) 358 { 359 struct perf_sample sample; 360 struct thread *thread; 361 int ret; 362 363 if (perf_evlist__parse_sample(evlist, event, &sample)) { 364 pr_debug("perf_evlist__parse_sample failed\n"); 365 return -1; 366 } 367 368 thread = machine__findnew_thread(machine, sample.pid, sample.tid); 369 if (!thread) { 370 pr_debug("machine__findnew_thread failed\n"); 371 return -1; 372 } 373 374 ret = read_object_code(sample.ip, READLEN, sample.cpumode, thread, state); 375 thread__put(thread); 376 return ret; 377 } 378 379 static int process_event(struct machine *machine, struct perf_evlist *evlist, 380 union perf_event *event, struct state *state) 381 { 382 if (event->header.type == PERF_RECORD_SAMPLE) 383 return process_sample_event(machine, evlist, event, state); 384 385 if (event->header.type == PERF_RECORD_THROTTLE || 386 event->header.type == PERF_RECORD_UNTHROTTLE) 387 return 0; 388 389 if (event->header.type < PERF_RECORD_MAX) { 390 int ret; 391 392 ret = machine__process_event(machine, event, NULL); 393 if (ret < 0) 394 pr_debug("machine__process_event failed, event type %u\n", 395 event->header.type); 396 return ret; 397 } 398 399 return 0; 400 } 401 402 static int process_events(struct machine *machine, struct perf_evlist *evlist, 403 struct state *state) 404 { 405 union perf_event *event; 406 int i, ret; 407 408 for (i = 0; i < evlist->nr_mmaps; i++) { 409 while ((event = perf_evlist__mmap_read(evlist, i)) != NULL) { 410 ret = process_event(machine, evlist, event, state); 411 perf_evlist__mmap_consume(evlist, i); 412 if (ret < 0) 413 return ret; 414 } 415 } 416 return 0; 417 } 418 419 static int comp(const void *a, const void *b) 420 { 421 return *(int *)a - *(int *)b; 422 } 423 424 static void do_sort_something(void) 425 { 426 int buf[40960], i; 427 428 for (i = 0; i < (int)ARRAY_SIZE(buf); i++) 429 buf[i] = ARRAY_SIZE(buf) - i - 1; 430 431 qsort(buf, ARRAY_SIZE(buf), sizeof(int), comp); 432 433 for (i = 0; i < (int)ARRAY_SIZE(buf); i++) { 434 if (buf[i] != i) { 435 pr_debug("qsort failed\n"); 436 break; 437 } 438 } 439 } 440 441 static void sort_something(void) 442 { 443 int i; 444 445 for (i = 0; i < 10; i++) 446 do_sort_something(); 447 } 448 449 static void syscall_something(void) 450 { 451 int pipefd[2]; 452 int i; 453 454 for (i = 0; i < 1000; i++) { 455 if (pipe(pipefd) < 0) { 456 pr_debug("pipe failed\n"); 457 break; 458 } 459 close(pipefd[1]); 460 close(pipefd[0]); 461 } 462 } 463 464 static void fs_something(void) 465 { 466 const char *test_file_name = "temp-perf-code-reading-test-file--"; 467 FILE *f; 468 int i; 469 470 for (i = 0; i < 1000; i++) { 471 f = fopen(test_file_name, "w+"); 472 if (f) { 473 fclose(f); 474 unlink(test_file_name); 475 } 476 } 477 } 478 479 static void do_something(void) 480 { 481 fs_something(); 482 483 sort_something(); 484 485 syscall_something(); 486 } 487 488 enum { 489 TEST_CODE_READING_OK, 490 TEST_CODE_READING_NO_VMLINUX, 491 TEST_CODE_READING_NO_KCORE, 492 TEST_CODE_READING_NO_ACCESS, 493 TEST_CODE_READING_NO_KERNEL_OBJ, 494 }; 495 496 static int do_test_code_reading(bool try_kcore) 497 { 498 struct machine *machine; 499 struct thread *thread; 500 struct record_opts opts = { 501 .mmap_pages = UINT_MAX, 502 .user_freq = UINT_MAX, 503 .user_interval = ULLONG_MAX, 504 .freq = 500, 505 .target = { 506 .uses_mmap = true, 507 }, 508 }; 509 struct state state = { 510 .done_cnt = 0, 511 }; 512 struct thread_map *threads = NULL; 513 struct cpu_map *cpus = NULL; 514 struct perf_evlist *evlist = NULL; 515 struct perf_evsel *evsel = NULL; 516 int err = -1, ret; 517 pid_t pid; 518 struct map *map; 519 bool have_vmlinux, have_kcore, excl_kernel = false; 520 521 pid = getpid(); 522 523 machine = machine__new_host(); 524 525 ret = machine__create_kernel_maps(machine); 526 if (ret < 0) { 527 pr_debug("machine__create_kernel_maps failed\n"); 528 goto out_err; 529 } 530 531 /* Force the use of kallsyms instead of vmlinux to try kcore */ 532 if (try_kcore) 533 symbol_conf.kallsyms_name = "/proc/kallsyms"; 534 535 /* Load kernel map */ 536 map = machine__kernel_map(machine); 537 ret = map__load(map); 538 if (ret < 0) { 539 pr_debug("map__load failed\n"); 540 goto out_err; 541 } 542 have_vmlinux = dso__is_vmlinux(map->dso); 543 have_kcore = dso__is_kcore(map->dso); 544 545 /* 2nd time through we just try kcore */ 546 if (try_kcore && !have_kcore) 547 return TEST_CODE_READING_NO_KCORE; 548 549 /* No point getting kernel events if there is no kernel object */ 550 if (!have_vmlinux && !have_kcore) 551 excl_kernel = true; 552 553 threads = thread_map__new_by_tid(pid); 554 if (!threads) { 555 pr_debug("thread_map__new_by_tid failed\n"); 556 goto out_err; 557 } 558 559 ret = perf_event__synthesize_thread_map(NULL, threads, 560 perf_event__process, machine, false, 500); 561 if (ret < 0) { 562 pr_debug("perf_event__synthesize_thread_map failed\n"); 563 goto out_err; 564 } 565 566 thread = machine__findnew_thread(machine, pid, pid); 567 if (!thread) { 568 pr_debug("machine__findnew_thread failed\n"); 569 goto out_put; 570 } 571 572 cpus = cpu_map__new(NULL); 573 if (!cpus) { 574 pr_debug("cpu_map__new failed\n"); 575 goto out_put; 576 } 577 578 while (1) { 579 const char *str; 580 581 evlist = perf_evlist__new(); 582 if (!evlist) { 583 pr_debug("perf_evlist__new failed\n"); 584 goto out_put; 585 } 586 587 perf_evlist__set_maps(evlist, cpus, threads); 588 589 if (excl_kernel) 590 str = "cycles:u"; 591 else 592 str = "cycles"; 593 pr_debug("Parsing event '%s'\n", str); 594 ret = parse_events(evlist, str, NULL); 595 if (ret < 0) { 596 pr_debug("parse_events failed\n"); 597 goto out_put; 598 } 599 600 perf_evlist__config(evlist, &opts, NULL); 601 602 evsel = perf_evlist__first(evlist); 603 604 evsel->attr.comm = 1; 605 evsel->attr.disabled = 1; 606 evsel->attr.enable_on_exec = 0; 607 608 ret = perf_evlist__open(evlist); 609 if (ret < 0) { 610 if (!excl_kernel) { 611 excl_kernel = true; 612 /* 613 * Both cpus and threads are now owned by evlist 614 * and will be freed by following perf_evlist__set_maps 615 * call. Getting refference to keep them alive. 616 */ 617 cpu_map__get(cpus); 618 thread_map__get(threads); 619 perf_evlist__set_maps(evlist, NULL, NULL); 620 perf_evlist__delete(evlist); 621 evlist = NULL; 622 continue; 623 } 624 625 if (verbose > 0) { 626 char errbuf[512]; 627 perf_evlist__strerror_open(evlist, errno, errbuf, sizeof(errbuf)); 628 pr_debug("perf_evlist__open() failed!\n%s\n", errbuf); 629 } 630 631 goto out_put; 632 } 633 break; 634 } 635 636 ret = perf_evlist__mmap(evlist, UINT_MAX, false); 637 if (ret < 0) { 638 pr_debug("perf_evlist__mmap failed\n"); 639 goto out_put; 640 } 641 642 perf_evlist__enable(evlist); 643 644 do_something(); 645 646 perf_evlist__disable(evlist); 647 648 ret = process_events(machine, evlist, &state); 649 if (ret < 0) 650 goto out_put; 651 652 if (!have_vmlinux && !have_kcore && !try_kcore) 653 err = TEST_CODE_READING_NO_KERNEL_OBJ; 654 else if (!have_vmlinux && !try_kcore) 655 err = TEST_CODE_READING_NO_VMLINUX; 656 else if (excl_kernel) 657 err = TEST_CODE_READING_NO_ACCESS; 658 else 659 err = TEST_CODE_READING_OK; 660 out_put: 661 thread__put(thread); 662 out_err: 663 664 if (evlist) { 665 perf_evlist__delete(evlist); 666 } else { 667 cpu_map__put(cpus); 668 thread_map__put(threads); 669 } 670 machine__delete_threads(machine); 671 machine__delete(machine); 672 673 return err; 674 } 675 676 int test__code_reading(int subtest __maybe_unused) 677 { 678 int ret; 679 680 ret = do_test_code_reading(false); 681 if (!ret) 682 ret = do_test_code_reading(true); 683 684 switch (ret) { 685 case TEST_CODE_READING_OK: 686 return 0; 687 case TEST_CODE_READING_NO_VMLINUX: 688 pr_debug("no vmlinux\n"); 689 return 0; 690 case TEST_CODE_READING_NO_KCORE: 691 pr_debug("no kcore\n"); 692 return 0; 693 case TEST_CODE_READING_NO_ACCESS: 694 pr_debug("no access\n"); 695 return 0; 696 case TEST_CODE_READING_NO_KERNEL_OBJ: 697 pr_debug("no kernel obj\n"); 698 return 0; 699 default: 700 return -1; 701 }; 702 } 703