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