1 // SPDX-License-Identifier: GPL-2.0 2 #include <errno.h> 3 #include <linux/kernel.h> 4 #include <linux/types.h> 5 #include <inttypes.h> 6 #include <stdlib.h> 7 #include <unistd.h> 8 #include <stdio.h> 9 #include <string.h> 10 #include <sys/param.h> 11 #include <perf/cpumap.h> 12 #include <perf/evlist.h> 13 #include <perf/mmap.h> 14 15 #include "debug.h" 16 #include "dso.h" 17 #include "env.h" 18 #include "parse-events.h" 19 #include "evlist.h" 20 #include "evsel.h" 21 #include "thread_map.h" 22 #include "machine.h" 23 #include "map.h" 24 #include "symbol.h" 25 #include "event.h" 26 #include "record.h" 27 #include "util/mmap.h" 28 #include "util/string2.h" 29 #include "util/synthetic-events.h" 30 #include "util/util.h" 31 #include "thread.h" 32 33 #include "tests.h" 34 35 #include <linux/ctype.h> 36 37 #define BUFSZ 1024 38 #define READLEN 128 39 40 struct state { 41 u64 done[1024]; 42 size_t done_cnt; 43 }; 44 45 static size_t read_objdump_chunk(const char **line, unsigned char **buf, 46 size_t *buf_len) 47 { 48 size_t bytes_read = 0; 49 unsigned char *chunk_start = *buf; 50 51 /* Read bytes */ 52 while (*buf_len > 0) { 53 char c1, c2; 54 55 /* Get 2 hex digits */ 56 c1 = *(*line)++; 57 if (!isxdigit(c1)) 58 break; 59 c2 = *(*line)++; 60 if (!isxdigit(c2)) 61 break; 62 63 /* Store byte and advance buf */ 64 **buf = (hex(c1) << 4) | hex(c2); 65 (*buf)++; 66 (*buf_len)--; 67 bytes_read++; 68 69 /* End of chunk? */ 70 if (isspace(**line)) 71 break; 72 } 73 74 /* 75 * objdump will display raw insn as LE if code endian 76 * is LE and bytes_per_chunk > 1. In that case reverse 77 * the chunk we just read. 78 * 79 * see disassemble_bytes() at binutils/objdump.c for details 80 * how objdump chooses display endian) 81 */ 82 if (bytes_read > 1 && !host_is_bigendian()) { 83 unsigned char *chunk_end = chunk_start + bytes_read - 1; 84 unsigned char tmp; 85 86 while (chunk_start < chunk_end) { 87 tmp = *chunk_start; 88 *chunk_start = *chunk_end; 89 *chunk_end = tmp; 90 chunk_start++; 91 chunk_end--; 92 } 93 } 94 95 return bytes_read; 96 } 97 98 static size_t read_objdump_line(const char *line, unsigned char *buf, 99 size_t buf_len) 100 { 101 const char *p; 102 size_t ret, bytes_read = 0; 103 104 /* Skip to a colon */ 105 p = strchr(line, ':'); 106 if (!p) 107 return 0; 108 p++; 109 110 /* Skip initial spaces */ 111 while (*p) { 112 if (!isspace(*p)) 113 break; 114 p++; 115 } 116 117 do { 118 ret = read_objdump_chunk(&p, &buf, &buf_len); 119 bytes_read += ret; 120 p++; 121 } while (ret > 0); 122 123 /* return number of successfully read bytes */ 124 return bytes_read; 125 } 126 127 static int read_objdump_output(FILE *f, void *buf, size_t *len, u64 start_addr) 128 { 129 char *line = NULL; 130 size_t line_len, off_last = 0; 131 ssize_t ret; 132 int err = 0; 133 u64 addr, last_addr = start_addr; 134 135 while (off_last < *len) { 136 size_t off, read_bytes, written_bytes; 137 unsigned char tmp[BUFSZ]; 138 139 ret = getline(&line, &line_len, f); 140 if (feof(f)) 141 break; 142 if (ret < 0) { 143 pr_debug("getline failed\n"); 144 err = -1; 145 break; 146 } 147 148 /* read objdump data into temporary buffer */ 149 read_bytes = read_objdump_line(line, tmp, sizeof(tmp)); 150 if (!read_bytes) 151 continue; 152 153 if (sscanf(line, "%"PRIx64, &addr) != 1) 154 continue; 155 if (addr < last_addr) { 156 pr_debug("addr going backwards, read beyond section?\n"); 157 break; 158 } 159 last_addr = addr; 160 161 /* copy it from temporary buffer to 'buf' according 162 * to address on current objdump line */ 163 off = addr - start_addr; 164 if (off >= *len) 165 break; 166 written_bytes = MIN(read_bytes, *len - off); 167 memcpy(buf + off, tmp, written_bytes); 168 off_last = off + written_bytes; 169 } 170 171 /* len returns number of bytes that could not be read */ 172 *len -= off_last; 173 174 free(line); 175 176 return err; 177 } 178 179 static int read_via_objdump(const char *filename, u64 addr, void *buf, 180 size_t len) 181 { 182 char cmd[PATH_MAX * 2]; 183 const char *fmt; 184 FILE *f; 185 int ret; 186 187 fmt = "%s -z -d --start-address=0x%"PRIx64" --stop-address=0x%"PRIx64" %s"; 188 ret = snprintf(cmd, sizeof(cmd), fmt, "objdump", addr, addr + len, 189 filename); 190 if (ret <= 0 || (size_t)ret >= sizeof(cmd)) 191 return -1; 192 193 pr_debug("Objdump command is: %s\n", cmd); 194 195 /* Ignore objdump errors */ 196 strcat(cmd, " 2>/dev/null"); 197 198 f = popen(cmd, "r"); 199 if (!f) { 200 pr_debug("popen failed\n"); 201 return -1; 202 } 203 204 ret = read_objdump_output(f, buf, &len, addr); 205 if (len) { 206 pr_debug("objdump read too few bytes: %zd\n", len); 207 if (!ret) 208 ret = len; 209 } 210 211 pclose(f); 212 213 return ret; 214 } 215 216 static void dump_buf(unsigned char *buf, size_t len) 217 { 218 size_t i; 219 220 for (i = 0; i < len; i++) { 221 pr_debug("0x%02x ", buf[i]); 222 if (i % 16 == 15) 223 pr_debug("\n"); 224 } 225 pr_debug("\n"); 226 } 227 228 static int read_object_code(u64 addr, size_t len, u8 cpumode, 229 struct thread *thread, struct state *state) 230 { 231 struct addr_location al; 232 unsigned char buf1[BUFSZ] = {0}; 233 unsigned char buf2[BUFSZ] = {0}; 234 size_t ret_len; 235 u64 objdump_addr; 236 const char *objdump_name; 237 char decomp_name[KMOD_DECOMP_LEN]; 238 bool decomp = false; 239 int ret, err = 0; 240 241 pr_debug("Reading object code for memory address: %#"PRIx64"\n", addr); 242 243 if (!thread__find_map(thread, cpumode, addr, &al) || !al.map->dso) { 244 if (cpumode == PERF_RECORD_MISC_HYPERVISOR) { 245 pr_debug("Hypervisor address can not be resolved - skipping\n"); 246 goto out; 247 } 248 249 pr_debug("thread__find_map failed\n"); 250 err = -1; 251 goto out; 252 } 253 254 pr_debug("File is: %s\n", al.map->dso->long_name); 255 256 if (al.map->dso->symtab_type == DSO_BINARY_TYPE__KALLSYMS && 257 !dso__is_kcore(al.map->dso)) { 258 pr_debug("Unexpected kernel address - skipping\n"); 259 goto out; 260 } 261 262 pr_debug("On file address is: %#"PRIx64"\n", al.addr); 263 264 if (len > BUFSZ) 265 len = BUFSZ; 266 267 /* Do not go off the map */ 268 if (addr + len > al.map->end) 269 len = al.map->end - addr; 270 271 /* Read the object code using perf */ 272 ret_len = dso__data_read_offset(al.map->dso, thread->maps->machine, 273 al.addr, buf1, len); 274 if (ret_len != len) { 275 pr_debug("dso__data_read_offset failed\n"); 276 err = -1; 277 goto out; 278 } 279 280 /* 281 * Converting addresses for use by objdump requires more information. 282 * map__load() does that. See map__rip_2objdump() for details. 283 */ 284 if (map__load(al.map)) { 285 err = -1; 286 goto out; 287 } 288 289 /* objdump struggles with kcore - try each map only once */ 290 if (dso__is_kcore(al.map->dso)) { 291 size_t d; 292 293 for (d = 0; d < state->done_cnt; d++) { 294 if (state->done[d] == al.map->start) { 295 pr_debug("kcore map tested already"); 296 pr_debug(" - skipping\n"); 297 goto out; 298 } 299 } 300 if (state->done_cnt >= ARRAY_SIZE(state->done)) { 301 pr_debug("Too many kcore maps - skipping\n"); 302 goto out; 303 } 304 state->done[state->done_cnt++] = al.map->start; 305 } 306 307 objdump_name = al.map->dso->long_name; 308 if (dso__needs_decompress(al.map->dso)) { 309 if (dso__decompress_kmodule_path(al.map->dso, objdump_name, 310 decomp_name, 311 sizeof(decomp_name)) < 0) { 312 pr_debug("decompression failed\n"); 313 err = -1; 314 goto out; 315 } 316 317 decomp = true; 318 objdump_name = decomp_name; 319 } 320 321 /* Read the object code using objdump */ 322 objdump_addr = map__rip_2objdump(al.map, al.addr); 323 ret = read_via_objdump(objdump_name, objdump_addr, buf2, len); 324 325 if (decomp) 326 unlink(objdump_name); 327 328 if (ret > 0) { 329 /* 330 * The kernel maps are inaccurate - assume objdump is right in 331 * that case. 332 */ 333 if (cpumode == PERF_RECORD_MISC_KERNEL || 334 cpumode == PERF_RECORD_MISC_GUEST_KERNEL) { 335 len -= ret; 336 if (len) { 337 pr_debug("Reducing len to %zu\n", len); 338 } else if (dso__is_kcore(al.map->dso)) { 339 /* 340 * objdump cannot handle very large segments 341 * that may be found in kcore. 342 */ 343 pr_debug("objdump failed for kcore"); 344 pr_debug(" - skipping\n"); 345 } else { 346 err = -1; 347 } 348 goto out; 349 } 350 } 351 if (ret < 0) { 352 pr_debug("read_via_objdump failed\n"); 353 err = -1; 354 goto out; 355 } 356 357 /* The results should be identical */ 358 if (memcmp(buf1, buf2, len)) { 359 pr_debug("Bytes read differ from those read by objdump\n"); 360 pr_debug("buf1 (dso):\n"); 361 dump_buf(buf1, len); 362 pr_debug("buf2 (objdump):\n"); 363 dump_buf(buf2, len); 364 err = -1; 365 goto out; 366 } 367 pr_debug("Bytes read match those read by objdump\n"); 368 out: 369 return err; 370 } 371 372 static int process_sample_event(struct machine *machine, 373 struct evlist *evlist, 374 union perf_event *event, struct state *state) 375 { 376 struct perf_sample sample; 377 struct thread *thread; 378 int ret; 379 380 if (evlist__parse_sample(evlist, event, &sample)) { 381 pr_debug("evlist__parse_sample failed\n"); 382 return -1; 383 } 384 385 thread = machine__findnew_thread(machine, sample.pid, sample.tid); 386 if (!thread) { 387 pr_debug("machine__findnew_thread failed\n"); 388 return -1; 389 } 390 391 ret = read_object_code(sample.ip, READLEN, sample.cpumode, thread, state); 392 thread__put(thread); 393 return ret; 394 } 395 396 static int process_event(struct machine *machine, struct evlist *evlist, 397 union perf_event *event, struct state *state) 398 { 399 if (event->header.type == PERF_RECORD_SAMPLE) 400 return process_sample_event(machine, evlist, event, state); 401 402 if (event->header.type == PERF_RECORD_THROTTLE || 403 event->header.type == PERF_RECORD_UNTHROTTLE) 404 return 0; 405 406 if (event->header.type < PERF_RECORD_MAX) { 407 int ret; 408 409 ret = machine__process_event(machine, event, NULL); 410 if (ret < 0) 411 pr_debug("machine__process_event failed, event type %u\n", 412 event->header.type); 413 return ret; 414 } 415 416 return 0; 417 } 418 419 static int process_events(struct machine *machine, struct evlist *evlist, 420 struct state *state) 421 { 422 union perf_event *event; 423 struct mmap *md; 424 int i, ret; 425 426 for (i = 0; i < evlist->core.nr_mmaps; i++) { 427 md = &evlist->mmap[i]; 428 if (perf_mmap__read_init(&md->core) < 0) 429 continue; 430 431 while ((event = perf_mmap__read_event(&md->core)) != NULL) { 432 ret = process_event(machine, evlist, event, state); 433 perf_mmap__consume(&md->core); 434 if (ret < 0) 435 return ret; 436 } 437 perf_mmap__read_done(&md->core); 438 } 439 return 0; 440 } 441 442 static int comp(const void *a, const void *b) 443 { 444 return *(int *)a - *(int *)b; 445 } 446 447 static void do_sort_something(void) 448 { 449 int buf[40960], i; 450 451 for (i = 0; i < (int)ARRAY_SIZE(buf); i++) 452 buf[i] = ARRAY_SIZE(buf) - i - 1; 453 454 qsort(buf, ARRAY_SIZE(buf), sizeof(int), comp); 455 456 for (i = 0; i < (int)ARRAY_SIZE(buf); i++) { 457 if (buf[i] != i) { 458 pr_debug("qsort failed\n"); 459 break; 460 } 461 } 462 } 463 464 static void sort_something(void) 465 { 466 int i; 467 468 for (i = 0; i < 10; i++) 469 do_sort_something(); 470 } 471 472 static void syscall_something(void) 473 { 474 int pipefd[2]; 475 int i; 476 477 for (i = 0; i < 1000; i++) { 478 if (pipe(pipefd) < 0) { 479 pr_debug("pipe failed\n"); 480 break; 481 } 482 close(pipefd[1]); 483 close(pipefd[0]); 484 } 485 } 486 487 static void fs_something(void) 488 { 489 const char *test_file_name = "temp-perf-code-reading-test-file--"; 490 FILE *f; 491 int i; 492 493 for (i = 0; i < 1000; i++) { 494 f = fopen(test_file_name, "w+"); 495 if (f) { 496 fclose(f); 497 unlink(test_file_name); 498 } 499 } 500 } 501 502 #ifdef __s390x__ 503 #include "header.h" // for get_cpuid() 504 #endif 505 506 static const char *do_determine_event(bool excl_kernel) 507 { 508 const char *event = excl_kernel ? "cycles:u" : "cycles"; 509 510 #ifdef __s390x__ 511 char cpuid[128], model[16], model_c[16], cpum_cf_v[16]; 512 unsigned int family; 513 int ret, cpum_cf_a; 514 515 if (get_cpuid(cpuid, sizeof(cpuid))) 516 goto out_clocks; 517 ret = sscanf(cpuid, "%*[^,],%u,%[^,],%[^,],%[^,],%x", &family, model_c, 518 model, cpum_cf_v, &cpum_cf_a); 519 if (ret != 5) /* Not available */ 520 goto out_clocks; 521 if (excl_kernel && (cpum_cf_a & 4)) 522 return event; 523 if (!excl_kernel && (cpum_cf_a & 2)) 524 return event; 525 526 /* Fall through: missing authorization */ 527 out_clocks: 528 event = excl_kernel ? "cpu-clock:u" : "cpu-clock"; 529 530 #endif 531 return event; 532 } 533 534 static void do_something(void) 535 { 536 fs_something(); 537 538 sort_something(); 539 540 syscall_something(); 541 } 542 543 enum { 544 TEST_CODE_READING_OK, 545 TEST_CODE_READING_NO_VMLINUX, 546 TEST_CODE_READING_NO_KCORE, 547 TEST_CODE_READING_NO_ACCESS, 548 TEST_CODE_READING_NO_KERNEL_OBJ, 549 }; 550 551 static int do_test_code_reading(bool try_kcore) 552 { 553 struct machine *machine; 554 struct thread *thread; 555 struct record_opts opts = { 556 .mmap_pages = UINT_MAX, 557 .user_freq = UINT_MAX, 558 .user_interval = ULLONG_MAX, 559 .freq = 500, 560 .target = { 561 .uses_mmap = true, 562 }, 563 }; 564 struct state state = { 565 .done_cnt = 0, 566 }; 567 struct perf_thread_map *threads = NULL; 568 struct perf_cpu_map *cpus = NULL; 569 struct evlist *evlist = NULL; 570 struct evsel *evsel = NULL; 571 int err = -1, ret; 572 pid_t pid; 573 struct map *map; 574 bool have_vmlinux, have_kcore, excl_kernel = false; 575 576 pid = getpid(); 577 578 machine = machine__new_host(); 579 machine->env = &perf_env; 580 581 ret = machine__create_kernel_maps(machine); 582 if (ret < 0) { 583 pr_debug("machine__create_kernel_maps failed\n"); 584 goto out_err; 585 } 586 587 /* Force the use of kallsyms instead of vmlinux to try kcore */ 588 if (try_kcore) 589 symbol_conf.kallsyms_name = "/proc/kallsyms"; 590 591 /* Load kernel map */ 592 map = machine__kernel_map(machine); 593 ret = map__load(map); 594 if (ret < 0) { 595 pr_debug("map__load failed\n"); 596 goto out_err; 597 } 598 have_vmlinux = dso__is_vmlinux(map->dso); 599 have_kcore = dso__is_kcore(map->dso); 600 601 /* 2nd time through we just try kcore */ 602 if (try_kcore && !have_kcore) 603 return TEST_CODE_READING_NO_KCORE; 604 605 /* No point getting kernel events if there is no kernel object */ 606 if (!have_vmlinux && !have_kcore) 607 excl_kernel = true; 608 609 threads = thread_map__new_by_tid(pid); 610 if (!threads) { 611 pr_debug("thread_map__new_by_tid failed\n"); 612 goto out_err; 613 } 614 615 ret = perf_event__synthesize_thread_map(NULL, threads, 616 perf_event__process, machine, 617 true, false); 618 if (ret < 0) { 619 pr_debug("perf_event__synthesize_thread_map failed\n"); 620 goto out_err; 621 } 622 623 thread = machine__findnew_thread(machine, pid, pid); 624 if (!thread) { 625 pr_debug("machine__findnew_thread failed\n"); 626 goto out_put; 627 } 628 629 cpus = perf_cpu_map__new(NULL); 630 if (!cpus) { 631 pr_debug("perf_cpu_map__new failed\n"); 632 goto out_put; 633 } 634 635 while (1) { 636 const char *str; 637 638 evlist = evlist__new(); 639 if (!evlist) { 640 pr_debug("evlist__new failed\n"); 641 goto out_put; 642 } 643 644 perf_evlist__set_maps(&evlist->core, cpus, threads); 645 646 str = do_determine_event(excl_kernel); 647 pr_debug("Parsing event '%s'\n", str); 648 ret = parse_event(evlist, str); 649 if (ret < 0) { 650 pr_debug("parse_events failed\n"); 651 goto out_put; 652 } 653 654 evlist__config(evlist, &opts, NULL); 655 656 evsel = evlist__first(evlist); 657 658 evsel->core.attr.comm = 1; 659 evsel->core.attr.disabled = 1; 660 evsel->core.attr.enable_on_exec = 0; 661 662 ret = evlist__open(evlist); 663 if (ret < 0) { 664 if (!excl_kernel) { 665 excl_kernel = true; 666 /* 667 * Both cpus and threads are now owned by evlist 668 * and will be freed by following perf_evlist__set_maps 669 * call. Getting reference to keep them alive. 670 */ 671 perf_cpu_map__get(cpus); 672 perf_thread_map__get(threads); 673 perf_evlist__set_maps(&evlist->core, NULL, NULL); 674 evlist__delete(evlist); 675 evlist = NULL; 676 continue; 677 } 678 679 if (verbose > 0) { 680 char errbuf[512]; 681 evlist__strerror_open(evlist, errno, errbuf, sizeof(errbuf)); 682 pr_debug("perf_evlist__open() failed!\n%s\n", errbuf); 683 } 684 685 goto out_put; 686 } 687 break; 688 } 689 690 ret = evlist__mmap(evlist, UINT_MAX); 691 if (ret < 0) { 692 pr_debug("evlist__mmap failed\n"); 693 goto out_put; 694 } 695 696 evlist__enable(evlist); 697 698 do_something(); 699 700 evlist__disable(evlist); 701 702 ret = process_events(machine, evlist, &state); 703 if (ret < 0) 704 goto out_put; 705 706 if (!have_vmlinux && !have_kcore && !try_kcore) 707 err = TEST_CODE_READING_NO_KERNEL_OBJ; 708 else if (!have_vmlinux && !try_kcore) 709 err = TEST_CODE_READING_NO_VMLINUX; 710 else if (excl_kernel) 711 err = TEST_CODE_READING_NO_ACCESS; 712 else 713 err = TEST_CODE_READING_OK; 714 out_put: 715 thread__put(thread); 716 out_err: 717 evlist__delete(evlist); 718 perf_cpu_map__put(cpus); 719 perf_thread_map__put(threads); 720 machine__delete_threads(machine); 721 machine__delete(machine); 722 723 return err; 724 } 725 726 static int test__code_reading(struct test_suite *test __maybe_unused, int subtest __maybe_unused) 727 { 728 int ret; 729 730 ret = do_test_code_reading(false); 731 if (!ret) 732 ret = do_test_code_reading(true); 733 734 switch (ret) { 735 case TEST_CODE_READING_OK: 736 return 0; 737 case TEST_CODE_READING_NO_VMLINUX: 738 pr_debug("no vmlinux\n"); 739 return 0; 740 case TEST_CODE_READING_NO_KCORE: 741 pr_debug("no kcore\n"); 742 return 0; 743 case TEST_CODE_READING_NO_ACCESS: 744 pr_debug("no access\n"); 745 return 0; 746 case TEST_CODE_READING_NO_KERNEL_OBJ: 747 pr_debug("no kernel obj\n"); 748 return 0; 749 default: 750 return -1; 751 }; 752 } 753 754 DEFINE_SUITE("Object code reading", code_reading); 755