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 int ret; 233 234 pr_debug("Reading object code for memory address: %#"PRIx64"\n", addr); 235 236 thread__find_addr_map(thread, cpumode, MAP__FUNCTION, addr, &al); 237 if (!al.map || !al.map->dso) { 238 pr_debug("thread__find_addr_map failed\n"); 239 return -1; 240 } 241 242 pr_debug("File is: %s\n", al.map->dso->long_name); 243 244 if (al.map->dso->symtab_type == DSO_BINARY_TYPE__KALLSYMS && 245 !dso__is_kcore(al.map->dso)) { 246 pr_debug("Unexpected kernel address - skipping\n"); 247 return 0; 248 } 249 250 pr_debug("On file address is: %#"PRIx64"\n", al.addr); 251 252 if (len > BUFSZ) 253 len = BUFSZ; 254 255 /* Do not go off the map */ 256 if (addr + len > al.map->end) 257 len = al.map->end - addr; 258 259 /* Read the object code using perf */ 260 ret_len = dso__data_read_offset(al.map->dso, thread->mg->machine, 261 al.addr, buf1, len); 262 if (ret_len != len) { 263 pr_debug("dso__data_read_offset failed\n"); 264 return -1; 265 } 266 267 /* 268 * Converting addresses for use by objdump requires more information. 269 * map__load() does that. See map__rip_2objdump() for details. 270 */ 271 if (map__load(al.map)) 272 return -1; 273 274 /* objdump struggles with kcore - try each map only once */ 275 if (dso__is_kcore(al.map->dso)) { 276 size_t d; 277 278 for (d = 0; d < state->done_cnt; d++) { 279 if (state->done[d] == al.map->start) { 280 pr_debug("kcore map tested already"); 281 pr_debug(" - skipping\n"); 282 return 0; 283 } 284 } 285 if (state->done_cnt >= ARRAY_SIZE(state->done)) { 286 pr_debug("Too many kcore maps - skipping\n"); 287 return 0; 288 } 289 state->done[state->done_cnt++] = al.map->start; 290 } 291 292 /* Read the object code using objdump */ 293 objdump_addr = map__rip_2objdump(al.map, al.addr); 294 ret = read_via_objdump(al.map->dso->long_name, objdump_addr, buf2, len); 295 if (ret > 0) { 296 /* 297 * The kernel maps are inaccurate - assume objdump is right in 298 * that case. 299 */ 300 if (cpumode == PERF_RECORD_MISC_KERNEL || 301 cpumode == PERF_RECORD_MISC_GUEST_KERNEL) { 302 len -= ret; 303 if (len) { 304 pr_debug("Reducing len to %zu\n", len); 305 } else if (dso__is_kcore(al.map->dso)) { 306 /* 307 * objdump cannot handle very large segments 308 * that may be found in kcore. 309 */ 310 pr_debug("objdump failed for kcore"); 311 pr_debug(" - skipping\n"); 312 return 0; 313 } else { 314 return -1; 315 } 316 } 317 } 318 if (ret < 0) { 319 pr_debug("read_via_objdump failed\n"); 320 return -1; 321 } 322 323 /* The results should be identical */ 324 if (memcmp(buf1, buf2, len)) { 325 pr_debug("Bytes read differ from those read by objdump\n"); 326 pr_debug("buf1 (dso):\n"); 327 dump_buf(buf1, len); 328 pr_debug("buf2 (objdump):\n"); 329 dump_buf(buf2, len); 330 return -1; 331 } 332 pr_debug("Bytes read match those read by objdump\n"); 333 334 return 0; 335 } 336 337 static int process_sample_event(struct machine *machine, 338 struct perf_evlist *evlist, 339 union perf_event *event, struct state *state) 340 { 341 struct perf_sample sample; 342 struct thread *thread; 343 int ret; 344 345 if (perf_evlist__parse_sample(evlist, event, &sample)) { 346 pr_debug("perf_evlist__parse_sample failed\n"); 347 return -1; 348 } 349 350 thread = machine__findnew_thread(machine, sample.pid, sample.tid); 351 if (!thread) { 352 pr_debug("machine__findnew_thread failed\n"); 353 return -1; 354 } 355 356 ret = read_object_code(sample.ip, READLEN, sample.cpumode, thread, state); 357 thread__put(thread); 358 return ret; 359 } 360 361 static int process_event(struct machine *machine, struct perf_evlist *evlist, 362 union perf_event *event, struct state *state) 363 { 364 if (event->header.type == PERF_RECORD_SAMPLE) 365 return process_sample_event(machine, evlist, event, state); 366 367 if (event->header.type == PERF_RECORD_THROTTLE || 368 event->header.type == PERF_RECORD_UNTHROTTLE) 369 return 0; 370 371 if (event->header.type < PERF_RECORD_MAX) { 372 int ret; 373 374 ret = machine__process_event(machine, event, NULL); 375 if (ret < 0) 376 pr_debug("machine__process_event failed, event type %u\n", 377 event->header.type); 378 return ret; 379 } 380 381 return 0; 382 } 383 384 static int process_events(struct machine *machine, struct perf_evlist *evlist, 385 struct state *state) 386 { 387 union perf_event *event; 388 int i, ret; 389 390 for (i = 0; i < evlist->nr_mmaps; i++) { 391 while ((event = perf_evlist__mmap_read(evlist, i)) != NULL) { 392 ret = process_event(machine, evlist, event, state); 393 perf_evlist__mmap_consume(evlist, i); 394 if (ret < 0) 395 return ret; 396 } 397 } 398 return 0; 399 } 400 401 static int comp(const void *a, const void *b) 402 { 403 return *(int *)a - *(int *)b; 404 } 405 406 static void do_sort_something(void) 407 { 408 int buf[40960], i; 409 410 for (i = 0; i < (int)ARRAY_SIZE(buf); i++) 411 buf[i] = ARRAY_SIZE(buf) - i - 1; 412 413 qsort(buf, ARRAY_SIZE(buf), sizeof(int), comp); 414 415 for (i = 0; i < (int)ARRAY_SIZE(buf); i++) { 416 if (buf[i] != i) { 417 pr_debug("qsort failed\n"); 418 break; 419 } 420 } 421 } 422 423 static void sort_something(void) 424 { 425 int i; 426 427 for (i = 0; i < 10; i++) 428 do_sort_something(); 429 } 430 431 static void syscall_something(void) 432 { 433 int pipefd[2]; 434 int i; 435 436 for (i = 0; i < 1000; i++) { 437 if (pipe(pipefd) < 0) { 438 pr_debug("pipe failed\n"); 439 break; 440 } 441 close(pipefd[1]); 442 close(pipefd[0]); 443 } 444 } 445 446 static void fs_something(void) 447 { 448 const char *test_file_name = "temp-perf-code-reading-test-file--"; 449 FILE *f; 450 int i; 451 452 for (i = 0; i < 1000; i++) { 453 f = fopen(test_file_name, "w+"); 454 if (f) { 455 fclose(f); 456 unlink(test_file_name); 457 } 458 } 459 } 460 461 static void do_something(void) 462 { 463 fs_something(); 464 465 sort_something(); 466 467 syscall_something(); 468 } 469 470 enum { 471 TEST_CODE_READING_OK, 472 TEST_CODE_READING_NO_VMLINUX, 473 TEST_CODE_READING_NO_KCORE, 474 TEST_CODE_READING_NO_ACCESS, 475 TEST_CODE_READING_NO_KERNEL_OBJ, 476 }; 477 478 static int do_test_code_reading(bool try_kcore) 479 { 480 struct machine *machine; 481 struct thread *thread; 482 struct record_opts opts = { 483 .mmap_pages = UINT_MAX, 484 .user_freq = UINT_MAX, 485 .user_interval = ULLONG_MAX, 486 .freq = 500, 487 .target = { 488 .uses_mmap = true, 489 }, 490 }; 491 struct state state = { 492 .done_cnt = 0, 493 }; 494 struct thread_map *threads = NULL; 495 struct cpu_map *cpus = NULL; 496 struct perf_evlist *evlist = NULL; 497 struct perf_evsel *evsel = NULL; 498 int err = -1, ret; 499 pid_t pid; 500 struct map *map; 501 bool have_vmlinux, have_kcore, excl_kernel = false; 502 503 pid = getpid(); 504 505 machine = machine__new_host(); 506 507 ret = machine__create_kernel_maps(machine); 508 if (ret < 0) { 509 pr_debug("machine__create_kernel_maps failed\n"); 510 goto out_err; 511 } 512 513 /* Force the use of kallsyms instead of vmlinux to try kcore */ 514 if (try_kcore) 515 symbol_conf.kallsyms_name = "/proc/kallsyms"; 516 517 /* Load kernel map */ 518 map = machine__kernel_map(machine); 519 ret = map__load(map); 520 if (ret < 0) { 521 pr_debug("map__load failed\n"); 522 goto out_err; 523 } 524 have_vmlinux = dso__is_vmlinux(map->dso); 525 have_kcore = dso__is_kcore(map->dso); 526 527 /* 2nd time through we just try kcore */ 528 if (try_kcore && !have_kcore) 529 return TEST_CODE_READING_NO_KCORE; 530 531 /* No point getting kernel events if there is no kernel object */ 532 if (!have_vmlinux && !have_kcore) 533 excl_kernel = true; 534 535 threads = thread_map__new_by_tid(pid); 536 if (!threads) { 537 pr_debug("thread_map__new_by_tid failed\n"); 538 goto out_err; 539 } 540 541 ret = perf_event__synthesize_thread_map(NULL, threads, 542 perf_event__process, machine, false, 500); 543 if (ret < 0) { 544 pr_debug("perf_event__synthesize_thread_map failed\n"); 545 goto out_err; 546 } 547 548 thread = machine__findnew_thread(machine, pid, pid); 549 if (!thread) { 550 pr_debug("machine__findnew_thread failed\n"); 551 goto out_put; 552 } 553 554 cpus = cpu_map__new(NULL); 555 if (!cpus) { 556 pr_debug("cpu_map__new failed\n"); 557 goto out_put; 558 } 559 560 while (1) { 561 const char *str; 562 563 evlist = perf_evlist__new(); 564 if (!evlist) { 565 pr_debug("perf_evlist__new failed\n"); 566 goto out_put; 567 } 568 569 perf_evlist__set_maps(evlist, cpus, threads); 570 571 if (excl_kernel) 572 str = "cycles:u"; 573 else 574 str = "cycles"; 575 pr_debug("Parsing event '%s'\n", str); 576 ret = parse_events(evlist, str, NULL); 577 if (ret < 0) { 578 pr_debug("parse_events failed\n"); 579 goto out_put; 580 } 581 582 perf_evlist__config(evlist, &opts, NULL); 583 584 evsel = perf_evlist__first(evlist); 585 586 evsel->attr.comm = 1; 587 evsel->attr.disabled = 1; 588 evsel->attr.enable_on_exec = 0; 589 590 ret = perf_evlist__open(evlist); 591 if (ret < 0) { 592 if (!excl_kernel) { 593 excl_kernel = true; 594 /* 595 * Both cpus and threads are now owned by evlist 596 * and will be freed by following perf_evlist__set_maps 597 * call. Getting refference to keep them alive. 598 */ 599 cpu_map__get(cpus); 600 thread_map__get(threads); 601 perf_evlist__set_maps(evlist, NULL, NULL); 602 perf_evlist__delete(evlist); 603 evlist = NULL; 604 continue; 605 } 606 607 if (verbose > 0) { 608 char errbuf[512]; 609 perf_evlist__strerror_open(evlist, errno, errbuf, sizeof(errbuf)); 610 pr_debug("perf_evlist__open() failed!\n%s\n", errbuf); 611 } 612 613 goto out_put; 614 } 615 break; 616 } 617 618 ret = perf_evlist__mmap(evlist, UINT_MAX, false); 619 if (ret < 0) { 620 pr_debug("perf_evlist__mmap failed\n"); 621 goto out_put; 622 } 623 624 perf_evlist__enable(evlist); 625 626 do_something(); 627 628 perf_evlist__disable(evlist); 629 630 ret = process_events(machine, evlist, &state); 631 if (ret < 0) 632 goto out_put; 633 634 if (!have_vmlinux && !have_kcore && !try_kcore) 635 err = TEST_CODE_READING_NO_KERNEL_OBJ; 636 else if (!have_vmlinux && !try_kcore) 637 err = TEST_CODE_READING_NO_VMLINUX; 638 else if (excl_kernel) 639 err = TEST_CODE_READING_NO_ACCESS; 640 else 641 err = TEST_CODE_READING_OK; 642 out_put: 643 thread__put(thread); 644 out_err: 645 646 if (evlist) { 647 perf_evlist__delete(evlist); 648 } else { 649 cpu_map__put(cpus); 650 thread_map__put(threads); 651 } 652 machine__delete_threads(machine); 653 machine__delete(machine); 654 655 return err; 656 } 657 658 int test__code_reading(int subtest __maybe_unused) 659 { 660 int ret; 661 662 ret = do_test_code_reading(false); 663 if (!ret) 664 ret = do_test_code_reading(true); 665 666 switch (ret) { 667 case TEST_CODE_READING_OK: 668 return 0; 669 case TEST_CODE_READING_NO_VMLINUX: 670 pr_debug("no vmlinux\n"); 671 return 0; 672 case TEST_CODE_READING_NO_KCORE: 673 pr_debug("no kcore\n"); 674 return 0; 675 case TEST_CODE_READING_NO_ACCESS: 676 pr_debug("no access\n"); 677 return 0; 678 case TEST_CODE_READING_NO_KERNEL_OBJ: 679 pr_debug("no kernel obj\n"); 680 return 0; 681 default: 682 return -1; 683 }; 684 } 685