1 // SPDX-License-Identifier: GPL-2.0 2 #include <errno.h> 3 #include <inttypes.h> 4 #include "string2.h" 5 #include <sys/param.h> 6 #include <sys/types.h> 7 #include <byteswap.h> 8 #include <unistd.h> 9 #include <regex.h> 10 #include <stdio.h> 11 #include <stdlib.h> 12 #include <linux/compiler.h> 13 #include <linux/list.h> 14 #include <linux/kernel.h> 15 #include <linux/bitops.h> 16 #include <linux/string.h> 17 #include <linux/stringify.h> 18 #include <linux/zalloc.h> 19 #include <sys/stat.h> 20 #include <sys/utsname.h> 21 #include <linux/time64.h> 22 #include <dirent.h> 23 #ifdef HAVE_LIBBPF_SUPPORT 24 #include <bpf/libbpf.h> 25 #endif 26 #include <perf/cpumap.h> 27 #include <tools/libc_compat.h> // reallocarray 28 29 #include "dso.h" 30 #include "evlist.h" 31 #include "evsel.h" 32 #include "util/evsel_fprintf.h" 33 #include "header.h" 34 #include "memswap.h" 35 #include "trace-event.h" 36 #include "session.h" 37 #include "symbol.h" 38 #include "debug.h" 39 #include "cpumap.h" 40 #include "pmu.h" 41 #include "pmus.h" 42 #include "vdso.h" 43 #include "strbuf.h" 44 #include "build-id.h" 45 #include "data.h" 46 #include <api/fs/fs.h> 47 #include "asm/bug.h" 48 #include "tool.h" 49 #include "time-utils.h" 50 #include "units.h" 51 #include "util/util.h" // perf_exe() 52 #include "cputopo.h" 53 #include "bpf-event.h" 54 #include "bpf-utils.h" 55 #include "clockid.h" 56 57 #include <linux/ctype.h> 58 #include <internal/lib.h> 59 60 #ifdef HAVE_LIBTRACEEVENT 61 #include <traceevent/event-parse.h> 62 #endif 63 64 /* 65 * magic2 = "PERFILE2" 66 * must be a numerical value to let the endianness 67 * determine the memory layout. That way we are able 68 * to detect endianness when reading the perf.data file 69 * back. 70 * 71 * we check for legacy (PERFFILE) format. 72 */ 73 static const char *__perf_magic1 = "PERFFILE"; 74 static const u64 __perf_magic2 = 0x32454c4946524550ULL; 75 static const u64 __perf_magic2_sw = 0x50455246494c4532ULL; 76 77 #define PERF_MAGIC __perf_magic2 78 79 const char perf_version_string[] = PERF_VERSION; 80 81 struct perf_file_attr { 82 struct perf_event_attr attr; 83 struct perf_file_section ids; 84 }; 85 86 void perf_header__set_feat(struct perf_header *header, int feat) 87 { 88 __set_bit(feat, header->adds_features); 89 } 90 91 void perf_header__clear_feat(struct perf_header *header, int feat) 92 { 93 __clear_bit(feat, header->adds_features); 94 } 95 96 bool perf_header__has_feat(const struct perf_header *header, int feat) 97 { 98 return test_bit(feat, header->adds_features); 99 } 100 101 static int __do_write_fd(struct feat_fd *ff, const void *buf, size_t size) 102 { 103 ssize_t ret = writen(ff->fd, buf, size); 104 105 if (ret != (ssize_t)size) 106 return ret < 0 ? (int)ret : -1; 107 return 0; 108 } 109 110 static int __do_write_buf(struct feat_fd *ff, const void *buf, size_t size) 111 { 112 /* struct perf_event_header::size is u16 */ 113 const size_t max_size = 0xffff - sizeof(struct perf_event_header); 114 size_t new_size = ff->size; 115 void *addr; 116 117 if (size + ff->offset > max_size) 118 return -E2BIG; 119 120 while (size > (new_size - ff->offset)) 121 new_size <<= 1; 122 new_size = min(max_size, new_size); 123 124 if (ff->size < new_size) { 125 addr = realloc(ff->buf, new_size); 126 if (!addr) 127 return -ENOMEM; 128 ff->buf = addr; 129 ff->size = new_size; 130 } 131 132 memcpy(ff->buf + ff->offset, buf, size); 133 ff->offset += size; 134 135 return 0; 136 } 137 138 /* Return: 0 if succeeded, -ERR if failed. */ 139 int do_write(struct feat_fd *ff, const void *buf, size_t size) 140 { 141 if (!ff->buf) 142 return __do_write_fd(ff, buf, size); 143 return __do_write_buf(ff, buf, size); 144 } 145 146 /* Return: 0 if succeeded, -ERR if failed. */ 147 static int do_write_bitmap(struct feat_fd *ff, unsigned long *set, u64 size) 148 { 149 u64 *p = (u64 *) set; 150 int i, ret; 151 152 ret = do_write(ff, &size, sizeof(size)); 153 if (ret < 0) 154 return ret; 155 156 for (i = 0; (u64) i < BITS_TO_U64(size); i++) { 157 ret = do_write(ff, p + i, sizeof(*p)); 158 if (ret < 0) 159 return ret; 160 } 161 162 return 0; 163 } 164 165 /* Return: 0 if succeeded, -ERR if failed. */ 166 int write_padded(struct feat_fd *ff, const void *bf, 167 size_t count, size_t count_aligned) 168 { 169 static const char zero_buf[NAME_ALIGN]; 170 int err = do_write(ff, bf, count); 171 172 if (!err) 173 err = do_write(ff, zero_buf, count_aligned - count); 174 175 return err; 176 } 177 178 #define string_size(str) \ 179 (PERF_ALIGN((strlen(str) + 1), NAME_ALIGN) + sizeof(u32)) 180 181 /* Return: 0 if succeeded, -ERR if failed. */ 182 static int do_write_string(struct feat_fd *ff, const char *str) 183 { 184 u32 len, olen; 185 int ret; 186 187 olen = strlen(str) + 1; 188 len = PERF_ALIGN(olen, NAME_ALIGN); 189 190 /* write len, incl. \0 */ 191 ret = do_write(ff, &len, sizeof(len)); 192 if (ret < 0) 193 return ret; 194 195 return write_padded(ff, str, olen, len); 196 } 197 198 static int __do_read_fd(struct feat_fd *ff, void *addr, ssize_t size) 199 { 200 ssize_t ret = readn(ff->fd, addr, size); 201 202 if (ret != size) 203 return ret < 0 ? (int)ret : -1; 204 return 0; 205 } 206 207 static int __do_read_buf(struct feat_fd *ff, void *addr, ssize_t size) 208 { 209 if (size > (ssize_t)ff->size - ff->offset) 210 return -1; 211 212 memcpy(addr, ff->buf + ff->offset, size); 213 ff->offset += size; 214 215 return 0; 216 217 } 218 219 static int __do_read(struct feat_fd *ff, void *addr, ssize_t size) 220 { 221 if (!ff->buf) 222 return __do_read_fd(ff, addr, size); 223 return __do_read_buf(ff, addr, size); 224 } 225 226 static int do_read_u32(struct feat_fd *ff, u32 *addr) 227 { 228 int ret; 229 230 ret = __do_read(ff, addr, sizeof(*addr)); 231 if (ret) 232 return ret; 233 234 if (ff->ph->needs_swap) 235 *addr = bswap_32(*addr); 236 return 0; 237 } 238 239 static int do_read_u64(struct feat_fd *ff, u64 *addr) 240 { 241 int ret; 242 243 ret = __do_read(ff, addr, sizeof(*addr)); 244 if (ret) 245 return ret; 246 247 if (ff->ph->needs_swap) 248 *addr = bswap_64(*addr); 249 return 0; 250 } 251 252 static char *do_read_string(struct feat_fd *ff) 253 { 254 u32 len; 255 char *buf; 256 257 if (do_read_u32(ff, &len)) 258 return NULL; 259 260 buf = malloc(len); 261 if (!buf) 262 return NULL; 263 264 if (!__do_read(ff, buf, len)) { 265 /* 266 * strings are padded by zeroes 267 * thus the actual strlen of buf 268 * may be less than len 269 */ 270 return buf; 271 } 272 273 free(buf); 274 return NULL; 275 } 276 277 /* Return: 0 if succeeded, -ERR if failed. */ 278 static int do_read_bitmap(struct feat_fd *ff, unsigned long **pset, u64 *psize) 279 { 280 unsigned long *set; 281 u64 size, *p; 282 int i, ret; 283 284 ret = do_read_u64(ff, &size); 285 if (ret) 286 return ret; 287 288 set = bitmap_zalloc(size); 289 if (!set) 290 return -ENOMEM; 291 292 p = (u64 *) set; 293 294 for (i = 0; (u64) i < BITS_TO_U64(size); i++) { 295 ret = do_read_u64(ff, p + i); 296 if (ret < 0) { 297 free(set); 298 return ret; 299 } 300 } 301 302 *pset = set; 303 *psize = size; 304 return 0; 305 } 306 307 #ifdef HAVE_LIBTRACEEVENT 308 static int write_tracing_data(struct feat_fd *ff, 309 struct evlist *evlist) 310 { 311 if (WARN(ff->buf, "Error: calling %s in pipe-mode.\n", __func__)) 312 return -1; 313 314 return read_tracing_data(ff->fd, &evlist->core.entries); 315 } 316 #endif 317 318 static int write_build_id(struct feat_fd *ff, 319 struct evlist *evlist __maybe_unused) 320 { 321 struct perf_session *session; 322 int err; 323 324 session = container_of(ff->ph, struct perf_session, header); 325 326 if (!perf_session__read_build_ids(session, true)) 327 return -1; 328 329 if (WARN(ff->buf, "Error: calling %s in pipe-mode.\n", __func__)) 330 return -1; 331 332 err = perf_session__write_buildid_table(session, ff); 333 if (err < 0) { 334 pr_debug("failed to write buildid table\n"); 335 return err; 336 } 337 perf_session__cache_build_ids(session); 338 339 return 0; 340 } 341 342 static int write_hostname(struct feat_fd *ff, 343 struct evlist *evlist __maybe_unused) 344 { 345 struct utsname uts; 346 int ret; 347 348 ret = uname(&uts); 349 if (ret < 0) 350 return -1; 351 352 return do_write_string(ff, uts.nodename); 353 } 354 355 static int write_osrelease(struct feat_fd *ff, 356 struct evlist *evlist __maybe_unused) 357 { 358 struct utsname uts; 359 int ret; 360 361 ret = uname(&uts); 362 if (ret < 0) 363 return -1; 364 365 return do_write_string(ff, uts.release); 366 } 367 368 static int write_arch(struct feat_fd *ff, 369 struct evlist *evlist __maybe_unused) 370 { 371 struct utsname uts; 372 int ret; 373 374 ret = uname(&uts); 375 if (ret < 0) 376 return -1; 377 378 return do_write_string(ff, uts.machine); 379 } 380 381 static int write_version(struct feat_fd *ff, 382 struct evlist *evlist __maybe_unused) 383 { 384 return do_write_string(ff, perf_version_string); 385 } 386 387 static int __write_cpudesc(struct feat_fd *ff, const char *cpuinfo_proc) 388 { 389 FILE *file; 390 char *buf = NULL; 391 char *s, *p; 392 const char *search = cpuinfo_proc; 393 size_t len = 0; 394 int ret = -1; 395 396 if (!search) 397 return -1; 398 399 file = fopen("/proc/cpuinfo", "r"); 400 if (!file) 401 return -1; 402 403 while (getline(&buf, &len, file) > 0) { 404 ret = strncmp(buf, search, strlen(search)); 405 if (!ret) 406 break; 407 } 408 409 if (ret) { 410 ret = -1; 411 goto done; 412 } 413 414 s = buf; 415 416 p = strchr(buf, ':'); 417 if (p && *(p+1) == ' ' && *(p+2)) 418 s = p + 2; 419 p = strchr(s, '\n'); 420 if (p) 421 *p = '\0'; 422 423 /* squash extra space characters (branding string) */ 424 p = s; 425 while (*p) { 426 if (isspace(*p)) { 427 char *r = p + 1; 428 char *q = skip_spaces(r); 429 *p = ' '; 430 if (q != (p+1)) 431 while ((*r++ = *q++)); 432 } 433 p++; 434 } 435 ret = do_write_string(ff, s); 436 done: 437 free(buf); 438 fclose(file); 439 return ret; 440 } 441 442 static int write_cpudesc(struct feat_fd *ff, 443 struct evlist *evlist __maybe_unused) 444 { 445 #if defined(__powerpc__) || defined(__hppa__) || defined(__sparc__) 446 #define CPUINFO_PROC { "cpu", } 447 #elif defined(__s390__) 448 #define CPUINFO_PROC { "vendor_id", } 449 #elif defined(__sh__) 450 #define CPUINFO_PROC { "cpu type", } 451 #elif defined(__alpha__) || defined(__mips__) 452 #define CPUINFO_PROC { "cpu model", } 453 #elif defined(__arm__) 454 #define CPUINFO_PROC { "model name", "Processor", } 455 #elif defined(__arc__) 456 #define CPUINFO_PROC { "Processor", } 457 #elif defined(__xtensa__) 458 #define CPUINFO_PROC { "core ID", } 459 #elif defined(__loongarch__) 460 #define CPUINFO_PROC { "Model Name", } 461 #else 462 #define CPUINFO_PROC { "model name", } 463 #endif 464 const char *cpuinfo_procs[] = CPUINFO_PROC; 465 #undef CPUINFO_PROC 466 unsigned int i; 467 468 for (i = 0; i < ARRAY_SIZE(cpuinfo_procs); i++) { 469 int ret; 470 ret = __write_cpudesc(ff, cpuinfo_procs[i]); 471 if (ret >= 0) 472 return ret; 473 } 474 return -1; 475 } 476 477 478 static int write_nrcpus(struct feat_fd *ff, 479 struct evlist *evlist __maybe_unused) 480 { 481 long nr; 482 u32 nrc, nra; 483 int ret; 484 485 nrc = cpu__max_present_cpu().cpu; 486 487 nr = sysconf(_SC_NPROCESSORS_ONLN); 488 if (nr < 0) 489 return -1; 490 491 nra = (u32)(nr & UINT_MAX); 492 493 ret = do_write(ff, &nrc, sizeof(nrc)); 494 if (ret < 0) 495 return ret; 496 497 return do_write(ff, &nra, sizeof(nra)); 498 } 499 500 static int write_event_desc(struct feat_fd *ff, 501 struct evlist *evlist) 502 { 503 struct evsel *evsel; 504 u32 nre, nri, sz; 505 int ret; 506 507 nre = evlist->core.nr_entries; 508 509 /* 510 * write number of events 511 */ 512 ret = do_write(ff, &nre, sizeof(nre)); 513 if (ret < 0) 514 return ret; 515 516 /* 517 * size of perf_event_attr struct 518 */ 519 sz = (u32)sizeof(evsel->core.attr); 520 ret = do_write(ff, &sz, sizeof(sz)); 521 if (ret < 0) 522 return ret; 523 524 evlist__for_each_entry(evlist, evsel) { 525 ret = do_write(ff, &evsel->core.attr, sz); 526 if (ret < 0) 527 return ret; 528 /* 529 * write number of unique id per event 530 * there is one id per instance of an event 531 * 532 * copy into an nri to be independent of the 533 * type of ids, 534 */ 535 nri = evsel->core.ids; 536 ret = do_write(ff, &nri, sizeof(nri)); 537 if (ret < 0) 538 return ret; 539 540 /* 541 * write event string as passed on cmdline 542 */ 543 ret = do_write_string(ff, evsel__name(evsel)); 544 if (ret < 0) 545 return ret; 546 /* 547 * write unique ids for this event 548 */ 549 ret = do_write(ff, evsel->core.id, evsel->core.ids * sizeof(u64)); 550 if (ret < 0) 551 return ret; 552 } 553 return 0; 554 } 555 556 static int write_cmdline(struct feat_fd *ff, 557 struct evlist *evlist __maybe_unused) 558 { 559 char pbuf[MAXPATHLEN], *buf; 560 int i, ret, n; 561 562 /* actual path to perf binary */ 563 buf = perf_exe(pbuf, MAXPATHLEN); 564 565 /* account for binary path */ 566 n = perf_env.nr_cmdline + 1; 567 568 ret = do_write(ff, &n, sizeof(n)); 569 if (ret < 0) 570 return ret; 571 572 ret = do_write_string(ff, buf); 573 if (ret < 0) 574 return ret; 575 576 for (i = 0 ; i < perf_env.nr_cmdline; i++) { 577 ret = do_write_string(ff, perf_env.cmdline_argv[i]); 578 if (ret < 0) 579 return ret; 580 } 581 return 0; 582 } 583 584 585 static int write_cpu_topology(struct feat_fd *ff, 586 struct evlist *evlist __maybe_unused) 587 { 588 struct cpu_topology *tp; 589 u32 i; 590 int ret, j; 591 592 tp = cpu_topology__new(); 593 if (!tp) 594 return -1; 595 596 ret = do_write(ff, &tp->package_cpus_lists, sizeof(tp->package_cpus_lists)); 597 if (ret < 0) 598 goto done; 599 600 for (i = 0; i < tp->package_cpus_lists; i++) { 601 ret = do_write_string(ff, tp->package_cpus_list[i]); 602 if (ret < 0) 603 goto done; 604 } 605 ret = do_write(ff, &tp->core_cpus_lists, sizeof(tp->core_cpus_lists)); 606 if (ret < 0) 607 goto done; 608 609 for (i = 0; i < tp->core_cpus_lists; i++) { 610 ret = do_write_string(ff, tp->core_cpus_list[i]); 611 if (ret < 0) 612 break; 613 } 614 615 ret = perf_env__read_cpu_topology_map(&perf_env); 616 if (ret < 0) 617 goto done; 618 619 for (j = 0; j < perf_env.nr_cpus_avail; j++) { 620 ret = do_write(ff, &perf_env.cpu[j].core_id, 621 sizeof(perf_env.cpu[j].core_id)); 622 if (ret < 0) 623 return ret; 624 ret = do_write(ff, &perf_env.cpu[j].socket_id, 625 sizeof(perf_env.cpu[j].socket_id)); 626 if (ret < 0) 627 return ret; 628 } 629 630 if (!tp->die_cpus_lists) 631 goto done; 632 633 ret = do_write(ff, &tp->die_cpus_lists, sizeof(tp->die_cpus_lists)); 634 if (ret < 0) 635 goto done; 636 637 for (i = 0; i < tp->die_cpus_lists; i++) { 638 ret = do_write_string(ff, tp->die_cpus_list[i]); 639 if (ret < 0) 640 goto done; 641 } 642 643 for (j = 0; j < perf_env.nr_cpus_avail; j++) { 644 ret = do_write(ff, &perf_env.cpu[j].die_id, 645 sizeof(perf_env.cpu[j].die_id)); 646 if (ret < 0) 647 return ret; 648 } 649 650 done: 651 cpu_topology__delete(tp); 652 return ret; 653 } 654 655 656 657 static int write_total_mem(struct feat_fd *ff, 658 struct evlist *evlist __maybe_unused) 659 { 660 char *buf = NULL; 661 FILE *fp; 662 size_t len = 0; 663 int ret = -1, n; 664 uint64_t mem; 665 666 fp = fopen("/proc/meminfo", "r"); 667 if (!fp) 668 return -1; 669 670 while (getline(&buf, &len, fp) > 0) { 671 ret = strncmp(buf, "MemTotal:", 9); 672 if (!ret) 673 break; 674 } 675 if (!ret) { 676 n = sscanf(buf, "%*s %"PRIu64, &mem); 677 if (n == 1) 678 ret = do_write(ff, &mem, sizeof(mem)); 679 } else 680 ret = -1; 681 free(buf); 682 fclose(fp); 683 return ret; 684 } 685 686 static int write_numa_topology(struct feat_fd *ff, 687 struct evlist *evlist __maybe_unused) 688 { 689 struct numa_topology *tp; 690 int ret = -1; 691 u32 i; 692 693 tp = numa_topology__new(); 694 if (!tp) 695 return -ENOMEM; 696 697 ret = do_write(ff, &tp->nr, sizeof(u32)); 698 if (ret < 0) 699 goto err; 700 701 for (i = 0; i < tp->nr; i++) { 702 struct numa_topology_node *n = &tp->nodes[i]; 703 704 ret = do_write(ff, &n->node, sizeof(u32)); 705 if (ret < 0) 706 goto err; 707 708 ret = do_write(ff, &n->mem_total, sizeof(u64)); 709 if (ret) 710 goto err; 711 712 ret = do_write(ff, &n->mem_free, sizeof(u64)); 713 if (ret) 714 goto err; 715 716 ret = do_write_string(ff, n->cpus); 717 if (ret < 0) 718 goto err; 719 } 720 721 ret = 0; 722 723 err: 724 numa_topology__delete(tp); 725 return ret; 726 } 727 728 /* 729 * File format: 730 * 731 * struct pmu_mappings { 732 * u32 pmu_num; 733 * struct pmu_map { 734 * u32 type; 735 * char name[]; 736 * }[pmu_num]; 737 * }; 738 */ 739 740 static int write_pmu_mappings(struct feat_fd *ff, 741 struct evlist *evlist __maybe_unused) 742 { 743 struct perf_pmu *pmu = NULL; 744 u32 pmu_num = 0; 745 int ret; 746 747 /* 748 * Do a first pass to count number of pmu to avoid lseek so this 749 * works in pipe mode as well. 750 */ 751 while ((pmu = perf_pmus__scan(pmu))) 752 pmu_num++; 753 754 ret = do_write(ff, &pmu_num, sizeof(pmu_num)); 755 if (ret < 0) 756 return ret; 757 758 while ((pmu = perf_pmus__scan(pmu))) { 759 ret = do_write(ff, &pmu->type, sizeof(pmu->type)); 760 if (ret < 0) 761 return ret; 762 763 ret = do_write_string(ff, pmu->name); 764 if (ret < 0) 765 return ret; 766 } 767 768 return 0; 769 } 770 771 /* 772 * File format: 773 * 774 * struct group_descs { 775 * u32 nr_groups; 776 * struct group_desc { 777 * char name[]; 778 * u32 leader_idx; 779 * u32 nr_members; 780 * }[nr_groups]; 781 * }; 782 */ 783 static int write_group_desc(struct feat_fd *ff, 784 struct evlist *evlist) 785 { 786 u32 nr_groups = evlist__nr_groups(evlist); 787 struct evsel *evsel; 788 int ret; 789 790 ret = do_write(ff, &nr_groups, sizeof(nr_groups)); 791 if (ret < 0) 792 return ret; 793 794 evlist__for_each_entry(evlist, evsel) { 795 if (evsel__is_group_leader(evsel) && evsel->core.nr_members > 1) { 796 const char *name = evsel->group_name ?: "{anon_group}"; 797 u32 leader_idx = evsel->core.idx; 798 u32 nr_members = evsel->core.nr_members; 799 800 ret = do_write_string(ff, name); 801 if (ret < 0) 802 return ret; 803 804 ret = do_write(ff, &leader_idx, sizeof(leader_idx)); 805 if (ret < 0) 806 return ret; 807 808 ret = do_write(ff, &nr_members, sizeof(nr_members)); 809 if (ret < 0) 810 return ret; 811 } 812 } 813 return 0; 814 } 815 816 /* 817 * Return the CPU id as a raw string. 818 * 819 * Each architecture should provide a more precise id string that 820 * can be use to match the architecture's "mapfile". 821 */ 822 char * __weak get_cpuid_str(struct perf_pmu *pmu __maybe_unused) 823 { 824 return NULL; 825 } 826 827 /* Return zero when the cpuid from the mapfile.csv matches the 828 * cpuid string generated on this platform. 829 * Otherwise return non-zero. 830 */ 831 int __weak strcmp_cpuid_str(const char *mapcpuid, const char *cpuid) 832 { 833 regex_t re; 834 regmatch_t pmatch[1]; 835 int match; 836 837 if (regcomp(&re, mapcpuid, REG_EXTENDED) != 0) { 838 /* Warn unable to generate match particular string. */ 839 pr_info("Invalid regular expression %s\n", mapcpuid); 840 return 1; 841 } 842 843 match = !regexec(&re, cpuid, 1, pmatch, 0); 844 regfree(&re); 845 if (match) { 846 size_t match_len = (pmatch[0].rm_eo - pmatch[0].rm_so); 847 848 /* Verify the entire string matched. */ 849 if (match_len == strlen(cpuid)) 850 return 0; 851 } 852 return 1; 853 } 854 855 /* 856 * default get_cpuid(): nothing gets recorded 857 * actual implementation must be in arch/$(SRCARCH)/util/header.c 858 */ 859 int __weak get_cpuid(char *buffer __maybe_unused, size_t sz __maybe_unused) 860 { 861 return ENOSYS; /* Not implemented */ 862 } 863 864 static int write_cpuid(struct feat_fd *ff, 865 struct evlist *evlist __maybe_unused) 866 { 867 char buffer[64]; 868 int ret; 869 870 ret = get_cpuid(buffer, sizeof(buffer)); 871 if (ret) 872 return -1; 873 874 return do_write_string(ff, buffer); 875 } 876 877 static int write_branch_stack(struct feat_fd *ff __maybe_unused, 878 struct evlist *evlist __maybe_unused) 879 { 880 return 0; 881 } 882 883 static int write_auxtrace(struct feat_fd *ff, 884 struct evlist *evlist __maybe_unused) 885 { 886 struct perf_session *session; 887 int err; 888 889 if (WARN(ff->buf, "Error: calling %s in pipe-mode.\n", __func__)) 890 return -1; 891 892 session = container_of(ff->ph, struct perf_session, header); 893 894 err = auxtrace_index__write(ff->fd, &session->auxtrace_index); 895 if (err < 0) 896 pr_err("Failed to write auxtrace index\n"); 897 return err; 898 } 899 900 static int write_clockid(struct feat_fd *ff, 901 struct evlist *evlist __maybe_unused) 902 { 903 return do_write(ff, &ff->ph->env.clock.clockid_res_ns, 904 sizeof(ff->ph->env.clock.clockid_res_ns)); 905 } 906 907 static int write_clock_data(struct feat_fd *ff, 908 struct evlist *evlist __maybe_unused) 909 { 910 u64 *data64; 911 u32 data32; 912 int ret; 913 914 /* version */ 915 data32 = 1; 916 917 ret = do_write(ff, &data32, sizeof(data32)); 918 if (ret < 0) 919 return ret; 920 921 /* clockid */ 922 data32 = ff->ph->env.clock.clockid; 923 924 ret = do_write(ff, &data32, sizeof(data32)); 925 if (ret < 0) 926 return ret; 927 928 /* TOD ref time */ 929 data64 = &ff->ph->env.clock.tod_ns; 930 931 ret = do_write(ff, data64, sizeof(*data64)); 932 if (ret < 0) 933 return ret; 934 935 /* clockid ref time */ 936 data64 = &ff->ph->env.clock.clockid_ns; 937 938 return do_write(ff, data64, sizeof(*data64)); 939 } 940 941 static int write_hybrid_topology(struct feat_fd *ff, 942 struct evlist *evlist __maybe_unused) 943 { 944 struct hybrid_topology *tp; 945 int ret; 946 u32 i; 947 948 tp = hybrid_topology__new(); 949 if (!tp) 950 return -ENOENT; 951 952 ret = do_write(ff, &tp->nr, sizeof(u32)); 953 if (ret < 0) 954 goto err; 955 956 for (i = 0; i < tp->nr; i++) { 957 struct hybrid_topology_node *n = &tp->nodes[i]; 958 959 ret = do_write_string(ff, n->pmu_name); 960 if (ret < 0) 961 goto err; 962 963 ret = do_write_string(ff, n->cpus); 964 if (ret < 0) 965 goto err; 966 } 967 968 ret = 0; 969 970 err: 971 hybrid_topology__delete(tp); 972 return ret; 973 } 974 975 static int write_dir_format(struct feat_fd *ff, 976 struct evlist *evlist __maybe_unused) 977 { 978 struct perf_session *session; 979 struct perf_data *data; 980 981 session = container_of(ff->ph, struct perf_session, header); 982 data = session->data; 983 984 if (WARN_ON(!perf_data__is_dir(data))) 985 return -1; 986 987 return do_write(ff, &data->dir.version, sizeof(data->dir.version)); 988 } 989 990 /* 991 * Check whether a CPU is online 992 * 993 * Returns: 994 * 1 -> if CPU is online 995 * 0 -> if CPU is offline 996 * -1 -> error case 997 */ 998 int is_cpu_online(unsigned int cpu) 999 { 1000 char *str; 1001 size_t strlen; 1002 char buf[256]; 1003 int status = -1; 1004 struct stat statbuf; 1005 1006 snprintf(buf, sizeof(buf), 1007 "/sys/devices/system/cpu/cpu%d", cpu); 1008 if (stat(buf, &statbuf) != 0) 1009 return 0; 1010 1011 /* 1012 * Check if /sys/devices/system/cpu/cpux/online file 1013 * exists. Some cases cpu0 won't have online file since 1014 * it is not expected to be turned off generally. 1015 * In kernels without CONFIG_HOTPLUG_CPU, this 1016 * file won't exist 1017 */ 1018 snprintf(buf, sizeof(buf), 1019 "/sys/devices/system/cpu/cpu%d/online", cpu); 1020 if (stat(buf, &statbuf) != 0) 1021 return 1; 1022 1023 /* 1024 * Read online file using sysfs__read_str. 1025 * If read or open fails, return -1. 1026 * If read succeeds, return value from file 1027 * which gets stored in "str" 1028 */ 1029 snprintf(buf, sizeof(buf), 1030 "devices/system/cpu/cpu%d/online", cpu); 1031 1032 if (sysfs__read_str(buf, &str, &strlen) < 0) 1033 return status; 1034 1035 status = atoi(str); 1036 1037 free(str); 1038 return status; 1039 } 1040 1041 #ifdef HAVE_LIBBPF_SUPPORT 1042 static int write_bpf_prog_info(struct feat_fd *ff, 1043 struct evlist *evlist __maybe_unused) 1044 { 1045 struct perf_env *env = &ff->ph->env; 1046 struct rb_root *root; 1047 struct rb_node *next; 1048 int ret; 1049 1050 down_read(&env->bpf_progs.lock); 1051 1052 ret = do_write(ff, &env->bpf_progs.infos_cnt, 1053 sizeof(env->bpf_progs.infos_cnt)); 1054 if (ret < 0) 1055 goto out; 1056 1057 root = &env->bpf_progs.infos; 1058 next = rb_first(root); 1059 while (next) { 1060 struct bpf_prog_info_node *node; 1061 size_t len; 1062 1063 node = rb_entry(next, struct bpf_prog_info_node, rb_node); 1064 next = rb_next(&node->rb_node); 1065 len = sizeof(struct perf_bpil) + 1066 node->info_linear->data_len; 1067 1068 /* before writing to file, translate address to offset */ 1069 bpil_addr_to_offs(node->info_linear); 1070 ret = do_write(ff, node->info_linear, len); 1071 /* 1072 * translate back to address even when do_write() fails, 1073 * so that this function never changes the data. 1074 */ 1075 bpil_offs_to_addr(node->info_linear); 1076 if (ret < 0) 1077 goto out; 1078 } 1079 out: 1080 up_read(&env->bpf_progs.lock); 1081 return ret; 1082 } 1083 1084 static int write_bpf_btf(struct feat_fd *ff, 1085 struct evlist *evlist __maybe_unused) 1086 { 1087 struct perf_env *env = &ff->ph->env; 1088 struct rb_root *root; 1089 struct rb_node *next; 1090 int ret; 1091 1092 down_read(&env->bpf_progs.lock); 1093 1094 ret = do_write(ff, &env->bpf_progs.btfs_cnt, 1095 sizeof(env->bpf_progs.btfs_cnt)); 1096 1097 if (ret < 0) 1098 goto out; 1099 1100 root = &env->bpf_progs.btfs; 1101 next = rb_first(root); 1102 while (next) { 1103 struct btf_node *node; 1104 1105 node = rb_entry(next, struct btf_node, rb_node); 1106 next = rb_next(&node->rb_node); 1107 ret = do_write(ff, &node->id, 1108 sizeof(u32) * 2 + node->data_size); 1109 if (ret < 0) 1110 goto out; 1111 } 1112 out: 1113 up_read(&env->bpf_progs.lock); 1114 return ret; 1115 } 1116 #endif // HAVE_LIBBPF_SUPPORT 1117 1118 static int cpu_cache_level__sort(const void *a, const void *b) 1119 { 1120 struct cpu_cache_level *cache_a = (struct cpu_cache_level *)a; 1121 struct cpu_cache_level *cache_b = (struct cpu_cache_level *)b; 1122 1123 return cache_a->level - cache_b->level; 1124 } 1125 1126 static bool cpu_cache_level__cmp(struct cpu_cache_level *a, struct cpu_cache_level *b) 1127 { 1128 if (a->level != b->level) 1129 return false; 1130 1131 if (a->line_size != b->line_size) 1132 return false; 1133 1134 if (a->sets != b->sets) 1135 return false; 1136 1137 if (a->ways != b->ways) 1138 return false; 1139 1140 if (strcmp(a->type, b->type)) 1141 return false; 1142 1143 if (strcmp(a->size, b->size)) 1144 return false; 1145 1146 if (strcmp(a->map, b->map)) 1147 return false; 1148 1149 return true; 1150 } 1151 1152 static int cpu_cache_level__read(struct cpu_cache_level *cache, u32 cpu, u16 level) 1153 { 1154 char path[PATH_MAX], file[PATH_MAX]; 1155 struct stat st; 1156 size_t len; 1157 1158 scnprintf(path, PATH_MAX, "devices/system/cpu/cpu%d/cache/index%d/", cpu, level); 1159 scnprintf(file, PATH_MAX, "%s/%s", sysfs__mountpoint(), path); 1160 1161 if (stat(file, &st)) 1162 return 1; 1163 1164 scnprintf(file, PATH_MAX, "%s/level", path); 1165 if (sysfs__read_int(file, (int *) &cache->level)) 1166 return -1; 1167 1168 scnprintf(file, PATH_MAX, "%s/coherency_line_size", path); 1169 if (sysfs__read_int(file, (int *) &cache->line_size)) 1170 return -1; 1171 1172 scnprintf(file, PATH_MAX, "%s/number_of_sets", path); 1173 if (sysfs__read_int(file, (int *) &cache->sets)) 1174 return -1; 1175 1176 scnprintf(file, PATH_MAX, "%s/ways_of_associativity", path); 1177 if (sysfs__read_int(file, (int *) &cache->ways)) 1178 return -1; 1179 1180 scnprintf(file, PATH_MAX, "%s/type", path); 1181 if (sysfs__read_str(file, &cache->type, &len)) 1182 return -1; 1183 1184 cache->type[len] = 0; 1185 cache->type = strim(cache->type); 1186 1187 scnprintf(file, PATH_MAX, "%s/size", path); 1188 if (sysfs__read_str(file, &cache->size, &len)) { 1189 zfree(&cache->type); 1190 return -1; 1191 } 1192 1193 cache->size[len] = 0; 1194 cache->size = strim(cache->size); 1195 1196 scnprintf(file, PATH_MAX, "%s/shared_cpu_list", path); 1197 if (sysfs__read_str(file, &cache->map, &len)) { 1198 zfree(&cache->size); 1199 zfree(&cache->type); 1200 return -1; 1201 } 1202 1203 cache->map[len] = 0; 1204 cache->map = strim(cache->map); 1205 return 0; 1206 } 1207 1208 static void cpu_cache_level__fprintf(FILE *out, struct cpu_cache_level *c) 1209 { 1210 fprintf(out, "L%d %-15s %8s [%s]\n", c->level, c->type, c->size, c->map); 1211 } 1212 1213 /* 1214 * Build caches levels for a particular CPU from the data in 1215 * /sys/devices/system/cpu/cpu<cpu>/cache/ 1216 * The cache level data is stored in caches[] from index at 1217 * *cntp. 1218 */ 1219 int build_caches_for_cpu(u32 cpu, struct cpu_cache_level caches[], u32 *cntp) 1220 { 1221 u16 level; 1222 1223 for (level = 0; level < MAX_CACHE_LVL; level++) { 1224 struct cpu_cache_level c; 1225 int err; 1226 u32 i; 1227 1228 err = cpu_cache_level__read(&c, cpu, level); 1229 if (err < 0) 1230 return err; 1231 1232 if (err == 1) 1233 break; 1234 1235 for (i = 0; i < *cntp; i++) { 1236 if (cpu_cache_level__cmp(&c, &caches[i])) 1237 break; 1238 } 1239 1240 if (i == *cntp) { 1241 caches[*cntp] = c; 1242 *cntp = *cntp + 1; 1243 } else 1244 cpu_cache_level__free(&c); 1245 } 1246 1247 return 0; 1248 } 1249 1250 static int build_caches(struct cpu_cache_level caches[], u32 *cntp) 1251 { 1252 u32 nr, cpu, cnt = 0; 1253 1254 nr = cpu__max_cpu().cpu; 1255 1256 for (cpu = 0; cpu < nr; cpu++) { 1257 int ret = build_caches_for_cpu(cpu, caches, &cnt); 1258 1259 if (ret) 1260 return ret; 1261 } 1262 *cntp = cnt; 1263 return 0; 1264 } 1265 1266 static int write_cache(struct feat_fd *ff, 1267 struct evlist *evlist __maybe_unused) 1268 { 1269 u32 max_caches = cpu__max_cpu().cpu * MAX_CACHE_LVL; 1270 struct cpu_cache_level caches[max_caches]; 1271 u32 cnt = 0, i, version = 1; 1272 int ret; 1273 1274 ret = build_caches(caches, &cnt); 1275 if (ret) 1276 goto out; 1277 1278 qsort(&caches, cnt, sizeof(struct cpu_cache_level), cpu_cache_level__sort); 1279 1280 ret = do_write(ff, &version, sizeof(u32)); 1281 if (ret < 0) 1282 goto out; 1283 1284 ret = do_write(ff, &cnt, sizeof(u32)); 1285 if (ret < 0) 1286 goto out; 1287 1288 for (i = 0; i < cnt; i++) { 1289 struct cpu_cache_level *c = &caches[i]; 1290 1291 #define _W(v) \ 1292 ret = do_write(ff, &c->v, sizeof(u32)); \ 1293 if (ret < 0) \ 1294 goto out; 1295 1296 _W(level) 1297 _W(line_size) 1298 _W(sets) 1299 _W(ways) 1300 #undef _W 1301 1302 #define _W(v) \ 1303 ret = do_write_string(ff, (const char *) c->v); \ 1304 if (ret < 0) \ 1305 goto out; 1306 1307 _W(type) 1308 _W(size) 1309 _W(map) 1310 #undef _W 1311 } 1312 1313 out: 1314 for (i = 0; i < cnt; i++) 1315 cpu_cache_level__free(&caches[i]); 1316 return ret; 1317 } 1318 1319 static int write_stat(struct feat_fd *ff __maybe_unused, 1320 struct evlist *evlist __maybe_unused) 1321 { 1322 return 0; 1323 } 1324 1325 static int write_sample_time(struct feat_fd *ff, 1326 struct evlist *evlist) 1327 { 1328 int ret; 1329 1330 ret = do_write(ff, &evlist->first_sample_time, 1331 sizeof(evlist->first_sample_time)); 1332 if (ret < 0) 1333 return ret; 1334 1335 return do_write(ff, &evlist->last_sample_time, 1336 sizeof(evlist->last_sample_time)); 1337 } 1338 1339 1340 static int memory_node__read(struct memory_node *n, unsigned long idx) 1341 { 1342 unsigned int phys, size = 0; 1343 char path[PATH_MAX]; 1344 struct dirent *ent; 1345 DIR *dir; 1346 1347 #define for_each_memory(mem, dir) \ 1348 while ((ent = readdir(dir))) \ 1349 if (strcmp(ent->d_name, ".") && \ 1350 strcmp(ent->d_name, "..") && \ 1351 sscanf(ent->d_name, "memory%u", &mem) == 1) 1352 1353 scnprintf(path, PATH_MAX, 1354 "%s/devices/system/node/node%lu", 1355 sysfs__mountpoint(), idx); 1356 1357 dir = opendir(path); 1358 if (!dir) { 1359 pr_warning("failed: can't open memory sysfs data\n"); 1360 return -1; 1361 } 1362 1363 for_each_memory(phys, dir) { 1364 size = max(phys, size); 1365 } 1366 1367 size++; 1368 1369 n->set = bitmap_zalloc(size); 1370 if (!n->set) { 1371 closedir(dir); 1372 return -ENOMEM; 1373 } 1374 1375 n->node = idx; 1376 n->size = size; 1377 1378 rewinddir(dir); 1379 1380 for_each_memory(phys, dir) { 1381 __set_bit(phys, n->set); 1382 } 1383 1384 closedir(dir); 1385 return 0; 1386 } 1387 1388 static void memory_node__delete_nodes(struct memory_node *nodesp, u64 cnt) 1389 { 1390 for (u64 i = 0; i < cnt; i++) 1391 bitmap_free(nodesp[i].set); 1392 1393 free(nodesp); 1394 } 1395 1396 static int memory_node__sort(const void *a, const void *b) 1397 { 1398 const struct memory_node *na = a; 1399 const struct memory_node *nb = b; 1400 1401 return na->node - nb->node; 1402 } 1403 1404 static int build_mem_topology(struct memory_node **nodesp, u64 *cntp) 1405 { 1406 char path[PATH_MAX]; 1407 struct dirent *ent; 1408 DIR *dir; 1409 int ret = 0; 1410 size_t cnt = 0, size = 0; 1411 struct memory_node *nodes = NULL; 1412 1413 scnprintf(path, PATH_MAX, "%s/devices/system/node/", 1414 sysfs__mountpoint()); 1415 1416 dir = opendir(path); 1417 if (!dir) { 1418 pr_debug2("%s: couldn't read %s, does this arch have topology information?\n", 1419 __func__, path); 1420 return -1; 1421 } 1422 1423 while (!ret && (ent = readdir(dir))) { 1424 unsigned int idx; 1425 int r; 1426 1427 if (!strcmp(ent->d_name, ".") || 1428 !strcmp(ent->d_name, "..")) 1429 continue; 1430 1431 r = sscanf(ent->d_name, "node%u", &idx); 1432 if (r != 1) 1433 continue; 1434 1435 if (cnt >= size) { 1436 struct memory_node *new_nodes = 1437 reallocarray(nodes, cnt + 4, sizeof(*nodes)); 1438 1439 if (!new_nodes) { 1440 pr_err("Failed to write MEM_TOPOLOGY, size %zd nodes\n", size); 1441 ret = -ENOMEM; 1442 goto out; 1443 } 1444 nodes = new_nodes; 1445 size += 4; 1446 } 1447 ret = memory_node__read(&nodes[cnt++], idx); 1448 } 1449 out: 1450 closedir(dir); 1451 if (!ret) { 1452 *cntp = cnt; 1453 *nodesp = nodes; 1454 qsort(nodes, cnt, sizeof(nodes[0]), memory_node__sort); 1455 } else 1456 memory_node__delete_nodes(nodes, cnt); 1457 1458 return ret; 1459 } 1460 1461 /* 1462 * The MEM_TOPOLOGY holds physical memory map for every 1463 * node in system. The format of data is as follows: 1464 * 1465 * 0 - version | for future changes 1466 * 8 - block_size_bytes | /sys/devices/system/memory/block_size_bytes 1467 * 16 - count | number of nodes 1468 * 1469 * For each node we store map of physical indexes for 1470 * each node: 1471 * 1472 * 32 - node id | node index 1473 * 40 - size | size of bitmap 1474 * 48 - bitmap | bitmap of memory indexes that belongs to node 1475 */ 1476 static int write_mem_topology(struct feat_fd *ff __maybe_unused, 1477 struct evlist *evlist __maybe_unused) 1478 { 1479 struct memory_node *nodes = NULL; 1480 u64 bsize, version = 1, i, nr = 0; 1481 int ret; 1482 1483 ret = sysfs__read_xll("devices/system/memory/block_size_bytes", 1484 (unsigned long long *) &bsize); 1485 if (ret) 1486 return ret; 1487 1488 ret = build_mem_topology(&nodes, &nr); 1489 if (ret) 1490 return ret; 1491 1492 ret = do_write(ff, &version, sizeof(version)); 1493 if (ret < 0) 1494 goto out; 1495 1496 ret = do_write(ff, &bsize, sizeof(bsize)); 1497 if (ret < 0) 1498 goto out; 1499 1500 ret = do_write(ff, &nr, sizeof(nr)); 1501 if (ret < 0) 1502 goto out; 1503 1504 for (i = 0; i < nr; i++) { 1505 struct memory_node *n = &nodes[i]; 1506 1507 #define _W(v) \ 1508 ret = do_write(ff, &n->v, sizeof(n->v)); \ 1509 if (ret < 0) \ 1510 goto out; 1511 1512 _W(node) 1513 _W(size) 1514 1515 #undef _W 1516 1517 ret = do_write_bitmap(ff, n->set, n->size); 1518 if (ret < 0) 1519 goto out; 1520 } 1521 1522 out: 1523 memory_node__delete_nodes(nodes, nr); 1524 return ret; 1525 } 1526 1527 static int write_compressed(struct feat_fd *ff __maybe_unused, 1528 struct evlist *evlist __maybe_unused) 1529 { 1530 int ret; 1531 1532 ret = do_write(ff, &(ff->ph->env.comp_ver), sizeof(ff->ph->env.comp_ver)); 1533 if (ret) 1534 return ret; 1535 1536 ret = do_write(ff, &(ff->ph->env.comp_type), sizeof(ff->ph->env.comp_type)); 1537 if (ret) 1538 return ret; 1539 1540 ret = do_write(ff, &(ff->ph->env.comp_level), sizeof(ff->ph->env.comp_level)); 1541 if (ret) 1542 return ret; 1543 1544 ret = do_write(ff, &(ff->ph->env.comp_ratio), sizeof(ff->ph->env.comp_ratio)); 1545 if (ret) 1546 return ret; 1547 1548 return do_write(ff, &(ff->ph->env.comp_mmap_len), sizeof(ff->ph->env.comp_mmap_len)); 1549 } 1550 1551 static int __write_pmu_caps(struct feat_fd *ff, struct perf_pmu *pmu, 1552 bool write_pmu) 1553 { 1554 struct perf_pmu_caps *caps = NULL; 1555 int ret; 1556 1557 ret = do_write(ff, &pmu->nr_caps, sizeof(pmu->nr_caps)); 1558 if (ret < 0) 1559 return ret; 1560 1561 list_for_each_entry(caps, &pmu->caps, list) { 1562 ret = do_write_string(ff, caps->name); 1563 if (ret < 0) 1564 return ret; 1565 1566 ret = do_write_string(ff, caps->value); 1567 if (ret < 0) 1568 return ret; 1569 } 1570 1571 if (write_pmu) { 1572 ret = do_write_string(ff, pmu->name); 1573 if (ret < 0) 1574 return ret; 1575 } 1576 1577 return ret; 1578 } 1579 1580 static int write_cpu_pmu_caps(struct feat_fd *ff, 1581 struct evlist *evlist __maybe_unused) 1582 { 1583 struct perf_pmu *cpu_pmu = perf_pmus__find("cpu"); 1584 int ret; 1585 1586 if (!cpu_pmu) 1587 return -ENOENT; 1588 1589 ret = perf_pmu__caps_parse(cpu_pmu); 1590 if (ret < 0) 1591 return ret; 1592 1593 return __write_pmu_caps(ff, cpu_pmu, false); 1594 } 1595 1596 static int write_pmu_caps(struct feat_fd *ff, 1597 struct evlist *evlist __maybe_unused) 1598 { 1599 struct perf_pmu *pmu = NULL; 1600 int nr_pmu = 0; 1601 int ret; 1602 1603 while ((pmu = perf_pmus__scan(pmu))) { 1604 if (!strcmp(pmu->name, "cpu")) { 1605 /* 1606 * The "cpu" PMU is special and covered by 1607 * HEADER_CPU_PMU_CAPS. Note, core PMUs are 1608 * counted/written here for ARM, s390 and Intel hybrid. 1609 */ 1610 continue; 1611 } 1612 if (perf_pmu__caps_parse(pmu) <= 0) 1613 continue; 1614 nr_pmu++; 1615 } 1616 1617 ret = do_write(ff, &nr_pmu, sizeof(nr_pmu)); 1618 if (ret < 0) 1619 return ret; 1620 1621 if (!nr_pmu) 1622 return 0; 1623 1624 /* 1625 * Note older perf tools assume core PMUs come first, this is a property 1626 * of perf_pmus__scan. 1627 */ 1628 pmu = NULL; 1629 while ((pmu = perf_pmus__scan(pmu))) { 1630 if (!strcmp(pmu->name, "cpu")) { 1631 /* Skip as above. */ 1632 continue; 1633 } 1634 if (perf_pmu__caps_parse(pmu) <= 0) 1635 continue; 1636 ret = __write_pmu_caps(ff, pmu, true); 1637 if (ret < 0) 1638 return ret; 1639 } 1640 return 0; 1641 } 1642 1643 static void print_hostname(struct feat_fd *ff, FILE *fp) 1644 { 1645 fprintf(fp, "# hostname : %s\n", ff->ph->env.hostname); 1646 } 1647 1648 static void print_osrelease(struct feat_fd *ff, FILE *fp) 1649 { 1650 fprintf(fp, "# os release : %s\n", ff->ph->env.os_release); 1651 } 1652 1653 static void print_arch(struct feat_fd *ff, FILE *fp) 1654 { 1655 fprintf(fp, "# arch : %s\n", ff->ph->env.arch); 1656 } 1657 1658 static void print_cpudesc(struct feat_fd *ff, FILE *fp) 1659 { 1660 fprintf(fp, "# cpudesc : %s\n", ff->ph->env.cpu_desc); 1661 } 1662 1663 static void print_nrcpus(struct feat_fd *ff, FILE *fp) 1664 { 1665 fprintf(fp, "# nrcpus online : %u\n", ff->ph->env.nr_cpus_online); 1666 fprintf(fp, "# nrcpus avail : %u\n", ff->ph->env.nr_cpus_avail); 1667 } 1668 1669 static void print_version(struct feat_fd *ff, FILE *fp) 1670 { 1671 fprintf(fp, "# perf version : %s\n", ff->ph->env.version); 1672 } 1673 1674 static void print_cmdline(struct feat_fd *ff, FILE *fp) 1675 { 1676 int nr, i; 1677 1678 nr = ff->ph->env.nr_cmdline; 1679 1680 fprintf(fp, "# cmdline : "); 1681 1682 for (i = 0; i < nr; i++) { 1683 char *argv_i = strdup(ff->ph->env.cmdline_argv[i]); 1684 if (!argv_i) { 1685 fprintf(fp, "%s ", ff->ph->env.cmdline_argv[i]); 1686 } else { 1687 char *mem = argv_i; 1688 do { 1689 char *quote = strchr(argv_i, '\''); 1690 if (!quote) 1691 break; 1692 *quote++ = '\0'; 1693 fprintf(fp, "%s\\\'", argv_i); 1694 argv_i = quote; 1695 } while (1); 1696 fprintf(fp, "%s ", argv_i); 1697 free(mem); 1698 } 1699 } 1700 fputc('\n', fp); 1701 } 1702 1703 static void print_cpu_topology(struct feat_fd *ff, FILE *fp) 1704 { 1705 struct perf_header *ph = ff->ph; 1706 int cpu_nr = ph->env.nr_cpus_avail; 1707 int nr, i; 1708 char *str; 1709 1710 nr = ph->env.nr_sibling_cores; 1711 str = ph->env.sibling_cores; 1712 1713 for (i = 0; i < nr; i++) { 1714 fprintf(fp, "# sibling sockets : %s\n", str); 1715 str += strlen(str) + 1; 1716 } 1717 1718 if (ph->env.nr_sibling_dies) { 1719 nr = ph->env.nr_sibling_dies; 1720 str = ph->env.sibling_dies; 1721 1722 for (i = 0; i < nr; i++) { 1723 fprintf(fp, "# sibling dies : %s\n", str); 1724 str += strlen(str) + 1; 1725 } 1726 } 1727 1728 nr = ph->env.nr_sibling_threads; 1729 str = ph->env.sibling_threads; 1730 1731 for (i = 0; i < nr; i++) { 1732 fprintf(fp, "# sibling threads : %s\n", str); 1733 str += strlen(str) + 1; 1734 } 1735 1736 if (ph->env.nr_sibling_dies) { 1737 if (ph->env.cpu != NULL) { 1738 for (i = 0; i < cpu_nr; i++) 1739 fprintf(fp, "# CPU %d: Core ID %d, " 1740 "Die ID %d, Socket ID %d\n", 1741 i, ph->env.cpu[i].core_id, 1742 ph->env.cpu[i].die_id, 1743 ph->env.cpu[i].socket_id); 1744 } else 1745 fprintf(fp, "# Core ID, Die ID and Socket ID " 1746 "information is not available\n"); 1747 } else { 1748 if (ph->env.cpu != NULL) { 1749 for (i = 0; i < cpu_nr; i++) 1750 fprintf(fp, "# CPU %d: Core ID %d, " 1751 "Socket ID %d\n", 1752 i, ph->env.cpu[i].core_id, 1753 ph->env.cpu[i].socket_id); 1754 } else 1755 fprintf(fp, "# Core ID and Socket ID " 1756 "information is not available\n"); 1757 } 1758 } 1759 1760 static void print_clockid(struct feat_fd *ff, FILE *fp) 1761 { 1762 fprintf(fp, "# clockid frequency: %"PRIu64" MHz\n", 1763 ff->ph->env.clock.clockid_res_ns * 1000); 1764 } 1765 1766 static void print_clock_data(struct feat_fd *ff, FILE *fp) 1767 { 1768 struct timespec clockid_ns; 1769 char tstr[64], date[64]; 1770 struct timeval tod_ns; 1771 clockid_t clockid; 1772 struct tm ltime; 1773 u64 ref; 1774 1775 if (!ff->ph->env.clock.enabled) { 1776 fprintf(fp, "# reference time disabled\n"); 1777 return; 1778 } 1779 1780 /* Compute TOD time. */ 1781 ref = ff->ph->env.clock.tod_ns; 1782 tod_ns.tv_sec = ref / NSEC_PER_SEC; 1783 ref -= tod_ns.tv_sec * NSEC_PER_SEC; 1784 tod_ns.tv_usec = ref / NSEC_PER_USEC; 1785 1786 /* Compute clockid time. */ 1787 ref = ff->ph->env.clock.clockid_ns; 1788 clockid_ns.tv_sec = ref / NSEC_PER_SEC; 1789 ref -= clockid_ns.tv_sec * NSEC_PER_SEC; 1790 clockid_ns.tv_nsec = ref; 1791 1792 clockid = ff->ph->env.clock.clockid; 1793 1794 if (localtime_r(&tod_ns.tv_sec, <ime) == NULL) 1795 snprintf(tstr, sizeof(tstr), "<error>"); 1796 else { 1797 strftime(date, sizeof(date), "%F %T", <ime); 1798 scnprintf(tstr, sizeof(tstr), "%s.%06d", 1799 date, (int) tod_ns.tv_usec); 1800 } 1801 1802 fprintf(fp, "# clockid: %s (%u)\n", clockid_name(clockid), clockid); 1803 fprintf(fp, "# reference time: %s = %ld.%06d (TOD) = %ld.%09ld (%s)\n", 1804 tstr, (long) tod_ns.tv_sec, (int) tod_ns.tv_usec, 1805 (long) clockid_ns.tv_sec, clockid_ns.tv_nsec, 1806 clockid_name(clockid)); 1807 } 1808 1809 static void print_hybrid_topology(struct feat_fd *ff, FILE *fp) 1810 { 1811 int i; 1812 struct hybrid_node *n; 1813 1814 fprintf(fp, "# hybrid cpu system:\n"); 1815 for (i = 0; i < ff->ph->env.nr_hybrid_nodes; i++) { 1816 n = &ff->ph->env.hybrid_nodes[i]; 1817 fprintf(fp, "# %s cpu list : %s\n", n->pmu_name, n->cpus); 1818 } 1819 } 1820 1821 static void print_dir_format(struct feat_fd *ff, FILE *fp) 1822 { 1823 struct perf_session *session; 1824 struct perf_data *data; 1825 1826 session = container_of(ff->ph, struct perf_session, header); 1827 data = session->data; 1828 1829 fprintf(fp, "# directory data version : %"PRIu64"\n", data->dir.version); 1830 } 1831 1832 #ifdef HAVE_LIBBPF_SUPPORT 1833 static void print_bpf_prog_info(struct feat_fd *ff, FILE *fp) 1834 { 1835 struct perf_env *env = &ff->ph->env; 1836 struct rb_root *root; 1837 struct rb_node *next; 1838 1839 down_read(&env->bpf_progs.lock); 1840 1841 root = &env->bpf_progs.infos; 1842 next = rb_first(root); 1843 1844 while (next) { 1845 struct bpf_prog_info_node *node; 1846 1847 node = rb_entry(next, struct bpf_prog_info_node, rb_node); 1848 next = rb_next(&node->rb_node); 1849 1850 bpf_event__print_bpf_prog_info(&node->info_linear->info, 1851 env, fp); 1852 } 1853 1854 up_read(&env->bpf_progs.lock); 1855 } 1856 1857 static void print_bpf_btf(struct feat_fd *ff, FILE *fp) 1858 { 1859 struct perf_env *env = &ff->ph->env; 1860 struct rb_root *root; 1861 struct rb_node *next; 1862 1863 down_read(&env->bpf_progs.lock); 1864 1865 root = &env->bpf_progs.btfs; 1866 next = rb_first(root); 1867 1868 while (next) { 1869 struct btf_node *node; 1870 1871 node = rb_entry(next, struct btf_node, rb_node); 1872 next = rb_next(&node->rb_node); 1873 fprintf(fp, "# btf info of id %u\n", node->id); 1874 } 1875 1876 up_read(&env->bpf_progs.lock); 1877 } 1878 #endif // HAVE_LIBBPF_SUPPORT 1879 1880 static void free_event_desc(struct evsel *events) 1881 { 1882 struct evsel *evsel; 1883 1884 if (!events) 1885 return; 1886 1887 for (evsel = events; evsel->core.attr.size; evsel++) { 1888 zfree(&evsel->name); 1889 zfree(&evsel->core.id); 1890 } 1891 1892 free(events); 1893 } 1894 1895 static bool perf_attr_check(struct perf_event_attr *attr) 1896 { 1897 if (attr->__reserved_1 || attr->__reserved_2 || attr->__reserved_3) { 1898 pr_warning("Reserved bits are set unexpectedly. " 1899 "Please update perf tool.\n"); 1900 return false; 1901 } 1902 1903 if (attr->sample_type & ~(PERF_SAMPLE_MAX-1)) { 1904 pr_warning("Unknown sample type (0x%llx) is detected. " 1905 "Please update perf tool.\n", 1906 attr->sample_type); 1907 return false; 1908 } 1909 1910 if (attr->read_format & ~(PERF_FORMAT_MAX-1)) { 1911 pr_warning("Unknown read format (0x%llx) is detected. " 1912 "Please update perf tool.\n", 1913 attr->read_format); 1914 return false; 1915 } 1916 1917 if ((attr->sample_type & PERF_SAMPLE_BRANCH_STACK) && 1918 (attr->branch_sample_type & ~(PERF_SAMPLE_BRANCH_MAX-1))) { 1919 pr_warning("Unknown branch sample type (0x%llx) is detected. " 1920 "Please update perf tool.\n", 1921 attr->branch_sample_type); 1922 1923 return false; 1924 } 1925 1926 return true; 1927 } 1928 1929 static struct evsel *read_event_desc(struct feat_fd *ff) 1930 { 1931 struct evsel *evsel, *events = NULL; 1932 u64 *id; 1933 void *buf = NULL; 1934 u32 nre, sz, nr, i, j; 1935 size_t msz; 1936 1937 /* number of events */ 1938 if (do_read_u32(ff, &nre)) 1939 goto error; 1940 1941 if (do_read_u32(ff, &sz)) 1942 goto error; 1943 1944 /* buffer to hold on file attr struct */ 1945 buf = malloc(sz); 1946 if (!buf) 1947 goto error; 1948 1949 /* the last event terminates with evsel->core.attr.size == 0: */ 1950 events = calloc(nre + 1, sizeof(*events)); 1951 if (!events) 1952 goto error; 1953 1954 msz = sizeof(evsel->core.attr); 1955 if (sz < msz) 1956 msz = sz; 1957 1958 for (i = 0, evsel = events; i < nre; evsel++, i++) { 1959 evsel->core.idx = i; 1960 1961 /* 1962 * must read entire on-file attr struct to 1963 * sync up with layout. 1964 */ 1965 if (__do_read(ff, buf, sz)) 1966 goto error; 1967 1968 if (ff->ph->needs_swap) 1969 perf_event__attr_swap(buf); 1970 1971 memcpy(&evsel->core.attr, buf, msz); 1972 1973 if (!perf_attr_check(&evsel->core.attr)) 1974 goto error; 1975 1976 if (do_read_u32(ff, &nr)) 1977 goto error; 1978 1979 if (ff->ph->needs_swap) 1980 evsel->needs_swap = true; 1981 1982 evsel->name = do_read_string(ff); 1983 if (!evsel->name) 1984 goto error; 1985 1986 if (!nr) 1987 continue; 1988 1989 id = calloc(nr, sizeof(*id)); 1990 if (!id) 1991 goto error; 1992 evsel->core.ids = nr; 1993 evsel->core.id = id; 1994 1995 for (j = 0 ; j < nr; j++) { 1996 if (do_read_u64(ff, id)) 1997 goto error; 1998 id++; 1999 } 2000 } 2001 out: 2002 free(buf); 2003 return events; 2004 error: 2005 free_event_desc(events); 2006 events = NULL; 2007 goto out; 2008 } 2009 2010 static int __desc_attr__fprintf(FILE *fp, const char *name, const char *val, 2011 void *priv __maybe_unused) 2012 { 2013 return fprintf(fp, ", %s = %s", name, val); 2014 } 2015 2016 static void print_event_desc(struct feat_fd *ff, FILE *fp) 2017 { 2018 struct evsel *evsel, *events; 2019 u32 j; 2020 u64 *id; 2021 2022 if (ff->events) 2023 events = ff->events; 2024 else 2025 events = read_event_desc(ff); 2026 2027 if (!events) { 2028 fprintf(fp, "# event desc: not available or unable to read\n"); 2029 return; 2030 } 2031 2032 for (evsel = events; evsel->core.attr.size; evsel++) { 2033 fprintf(fp, "# event : name = %s, ", evsel->name); 2034 2035 if (evsel->core.ids) { 2036 fprintf(fp, ", id = {"); 2037 for (j = 0, id = evsel->core.id; j < evsel->core.ids; j++, id++) { 2038 if (j) 2039 fputc(',', fp); 2040 fprintf(fp, " %"PRIu64, *id); 2041 } 2042 fprintf(fp, " }"); 2043 } 2044 2045 perf_event_attr__fprintf(fp, &evsel->core.attr, __desc_attr__fprintf, NULL); 2046 2047 fputc('\n', fp); 2048 } 2049 2050 free_event_desc(events); 2051 ff->events = NULL; 2052 } 2053 2054 static void print_total_mem(struct feat_fd *ff, FILE *fp) 2055 { 2056 fprintf(fp, "# total memory : %llu kB\n", ff->ph->env.total_mem); 2057 } 2058 2059 static void print_numa_topology(struct feat_fd *ff, FILE *fp) 2060 { 2061 int i; 2062 struct numa_node *n; 2063 2064 for (i = 0; i < ff->ph->env.nr_numa_nodes; i++) { 2065 n = &ff->ph->env.numa_nodes[i]; 2066 2067 fprintf(fp, "# node%u meminfo : total = %"PRIu64" kB," 2068 " free = %"PRIu64" kB\n", 2069 n->node, n->mem_total, n->mem_free); 2070 2071 fprintf(fp, "# node%u cpu list : ", n->node); 2072 cpu_map__fprintf(n->map, fp); 2073 } 2074 } 2075 2076 static void print_cpuid(struct feat_fd *ff, FILE *fp) 2077 { 2078 fprintf(fp, "# cpuid : %s\n", ff->ph->env.cpuid); 2079 } 2080 2081 static void print_branch_stack(struct feat_fd *ff __maybe_unused, FILE *fp) 2082 { 2083 fprintf(fp, "# contains samples with branch stack\n"); 2084 } 2085 2086 static void print_auxtrace(struct feat_fd *ff __maybe_unused, FILE *fp) 2087 { 2088 fprintf(fp, "# contains AUX area data (e.g. instruction trace)\n"); 2089 } 2090 2091 static void print_stat(struct feat_fd *ff __maybe_unused, FILE *fp) 2092 { 2093 fprintf(fp, "# contains stat data\n"); 2094 } 2095 2096 static void print_cache(struct feat_fd *ff, FILE *fp __maybe_unused) 2097 { 2098 int i; 2099 2100 fprintf(fp, "# CPU cache info:\n"); 2101 for (i = 0; i < ff->ph->env.caches_cnt; i++) { 2102 fprintf(fp, "# "); 2103 cpu_cache_level__fprintf(fp, &ff->ph->env.caches[i]); 2104 } 2105 } 2106 2107 static void print_compressed(struct feat_fd *ff, FILE *fp) 2108 { 2109 fprintf(fp, "# compressed : %s, level = %d, ratio = %d\n", 2110 ff->ph->env.comp_type == PERF_COMP_ZSTD ? "Zstd" : "Unknown", 2111 ff->ph->env.comp_level, ff->ph->env.comp_ratio); 2112 } 2113 2114 static void __print_pmu_caps(FILE *fp, int nr_caps, char **caps, char *pmu_name) 2115 { 2116 const char *delimiter = ""; 2117 int i; 2118 2119 if (!nr_caps) { 2120 fprintf(fp, "# %s pmu capabilities: not available\n", pmu_name); 2121 return; 2122 } 2123 2124 fprintf(fp, "# %s pmu capabilities: ", pmu_name); 2125 for (i = 0; i < nr_caps; i++) { 2126 fprintf(fp, "%s%s", delimiter, caps[i]); 2127 delimiter = ", "; 2128 } 2129 2130 fprintf(fp, "\n"); 2131 } 2132 2133 static void print_cpu_pmu_caps(struct feat_fd *ff, FILE *fp) 2134 { 2135 __print_pmu_caps(fp, ff->ph->env.nr_cpu_pmu_caps, 2136 ff->ph->env.cpu_pmu_caps, (char *)"cpu"); 2137 } 2138 2139 static void print_pmu_caps(struct feat_fd *ff, FILE *fp) 2140 { 2141 struct pmu_caps *pmu_caps; 2142 2143 for (int i = 0; i < ff->ph->env.nr_pmus_with_caps; i++) { 2144 pmu_caps = &ff->ph->env.pmu_caps[i]; 2145 __print_pmu_caps(fp, pmu_caps->nr_caps, pmu_caps->caps, 2146 pmu_caps->pmu_name); 2147 } 2148 2149 if (strcmp(perf_env__arch(&ff->ph->env), "x86") == 0 && 2150 perf_env__has_pmu_mapping(&ff->ph->env, "ibs_op")) { 2151 char *max_precise = perf_env__find_pmu_cap(&ff->ph->env, "cpu", "max_precise"); 2152 2153 if (max_precise != NULL && atoi(max_precise) == 0) 2154 fprintf(fp, "# AMD systems uses ibs_op// PMU for some precise events, e.g.: cycles:p, see the 'perf list' man page for further details.\n"); 2155 } 2156 } 2157 2158 static void print_pmu_mappings(struct feat_fd *ff, FILE *fp) 2159 { 2160 const char *delimiter = "# pmu mappings: "; 2161 char *str, *tmp; 2162 u32 pmu_num; 2163 u32 type; 2164 2165 pmu_num = ff->ph->env.nr_pmu_mappings; 2166 if (!pmu_num) { 2167 fprintf(fp, "# pmu mappings: not available\n"); 2168 return; 2169 } 2170 2171 str = ff->ph->env.pmu_mappings; 2172 2173 while (pmu_num) { 2174 type = strtoul(str, &tmp, 0); 2175 if (*tmp != ':') 2176 goto error; 2177 2178 str = tmp + 1; 2179 fprintf(fp, "%s%s = %" PRIu32, delimiter, str, type); 2180 2181 delimiter = ", "; 2182 str += strlen(str) + 1; 2183 pmu_num--; 2184 } 2185 2186 fprintf(fp, "\n"); 2187 2188 if (!pmu_num) 2189 return; 2190 error: 2191 fprintf(fp, "# pmu mappings: unable to read\n"); 2192 } 2193 2194 static void print_group_desc(struct feat_fd *ff, FILE *fp) 2195 { 2196 struct perf_session *session; 2197 struct evsel *evsel; 2198 u32 nr = 0; 2199 2200 session = container_of(ff->ph, struct perf_session, header); 2201 2202 evlist__for_each_entry(session->evlist, evsel) { 2203 if (evsel__is_group_leader(evsel) && evsel->core.nr_members > 1) { 2204 fprintf(fp, "# group: %s{%s", evsel->group_name ?: "", evsel__name(evsel)); 2205 2206 nr = evsel->core.nr_members - 1; 2207 } else if (nr) { 2208 fprintf(fp, ",%s", evsel__name(evsel)); 2209 2210 if (--nr == 0) 2211 fprintf(fp, "}\n"); 2212 } 2213 } 2214 } 2215 2216 static void print_sample_time(struct feat_fd *ff, FILE *fp) 2217 { 2218 struct perf_session *session; 2219 char time_buf[32]; 2220 double d; 2221 2222 session = container_of(ff->ph, struct perf_session, header); 2223 2224 timestamp__scnprintf_usec(session->evlist->first_sample_time, 2225 time_buf, sizeof(time_buf)); 2226 fprintf(fp, "# time of first sample : %s\n", time_buf); 2227 2228 timestamp__scnprintf_usec(session->evlist->last_sample_time, 2229 time_buf, sizeof(time_buf)); 2230 fprintf(fp, "# time of last sample : %s\n", time_buf); 2231 2232 d = (double)(session->evlist->last_sample_time - 2233 session->evlist->first_sample_time) / NSEC_PER_MSEC; 2234 2235 fprintf(fp, "# sample duration : %10.3f ms\n", d); 2236 } 2237 2238 static void memory_node__fprintf(struct memory_node *n, 2239 unsigned long long bsize, FILE *fp) 2240 { 2241 char buf_map[100], buf_size[50]; 2242 unsigned long long size; 2243 2244 size = bsize * bitmap_weight(n->set, n->size); 2245 unit_number__scnprintf(buf_size, 50, size); 2246 2247 bitmap_scnprintf(n->set, n->size, buf_map, 100); 2248 fprintf(fp, "# %3" PRIu64 " [%s]: %s\n", n->node, buf_size, buf_map); 2249 } 2250 2251 static void print_mem_topology(struct feat_fd *ff, FILE *fp) 2252 { 2253 struct memory_node *nodes; 2254 int i, nr; 2255 2256 nodes = ff->ph->env.memory_nodes; 2257 nr = ff->ph->env.nr_memory_nodes; 2258 2259 fprintf(fp, "# memory nodes (nr %d, block size 0x%llx):\n", 2260 nr, ff->ph->env.memory_bsize); 2261 2262 for (i = 0; i < nr; i++) { 2263 memory_node__fprintf(&nodes[i], ff->ph->env.memory_bsize, fp); 2264 } 2265 } 2266 2267 static int __event_process_build_id(struct perf_record_header_build_id *bev, 2268 char *filename, 2269 struct perf_session *session) 2270 { 2271 int err = -1; 2272 struct machine *machine; 2273 u16 cpumode; 2274 struct dso *dso; 2275 enum dso_space_type dso_space; 2276 2277 machine = perf_session__findnew_machine(session, bev->pid); 2278 if (!machine) 2279 goto out; 2280 2281 cpumode = bev->header.misc & PERF_RECORD_MISC_CPUMODE_MASK; 2282 2283 switch (cpumode) { 2284 case PERF_RECORD_MISC_KERNEL: 2285 dso_space = DSO_SPACE__KERNEL; 2286 break; 2287 case PERF_RECORD_MISC_GUEST_KERNEL: 2288 dso_space = DSO_SPACE__KERNEL_GUEST; 2289 break; 2290 case PERF_RECORD_MISC_USER: 2291 case PERF_RECORD_MISC_GUEST_USER: 2292 dso_space = DSO_SPACE__USER; 2293 break; 2294 default: 2295 goto out; 2296 } 2297 2298 dso = machine__findnew_dso(machine, filename); 2299 if (dso != NULL) { 2300 char sbuild_id[SBUILD_ID_SIZE]; 2301 struct build_id bid; 2302 size_t size = BUILD_ID_SIZE; 2303 2304 if (bev->header.misc & PERF_RECORD_MISC_BUILD_ID_SIZE) 2305 size = bev->size; 2306 2307 build_id__init(&bid, bev->data, size); 2308 dso__set_build_id(dso, &bid); 2309 dso->header_build_id = 1; 2310 2311 if (dso_space != DSO_SPACE__USER) { 2312 struct kmod_path m = { .name = NULL, }; 2313 2314 if (!kmod_path__parse_name(&m, filename) && m.kmod) 2315 dso__set_module_info(dso, &m, machine); 2316 2317 dso->kernel = dso_space; 2318 free(m.name); 2319 } 2320 2321 build_id__sprintf(&dso->bid, sbuild_id); 2322 pr_debug("build id event received for %s: %s [%zu]\n", 2323 dso->long_name, sbuild_id, size); 2324 dso__put(dso); 2325 } 2326 2327 err = 0; 2328 out: 2329 return err; 2330 } 2331 2332 static int perf_header__read_build_ids_abi_quirk(struct perf_header *header, 2333 int input, u64 offset, u64 size) 2334 { 2335 struct perf_session *session = container_of(header, struct perf_session, header); 2336 struct { 2337 struct perf_event_header header; 2338 u8 build_id[PERF_ALIGN(BUILD_ID_SIZE, sizeof(u64))]; 2339 char filename[0]; 2340 } old_bev; 2341 struct perf_record_header_build_id bev; 2342 char filename[PATH_MAX]; 2343 u64 limit = offset + size; 2344 2345 while (offset < limit) { 2346 ssize_t len; 2347 2348 if (readn(input, &old_bev, sizeof(old_bev)) != sizeof(old_bev)) 2349 return -1; 2350 2351 if (header->needs_swap) 2352 perf_event_header__bswap(&old_bev.header); 2353 2354 len = old_bev.header.size - sizeof(old_bev); 2355 if (readn(input, filename, len) != len) 2356 return -1; 2357 2358 bev.header = old_bev.header; 2359 2360 /* 2361 * As the pid is the missing value, we need to fill 2362 * it properly. The header.misc value give us nice hint. 2363 */ 2364 bev.pid = HOST_KERNEL_ID; 2365 if (bev.header.misc == PERF_RECORD_MISC_GUEST_USER || 2366 bev.header.misc == PERF_RECORD_MISC_GUEST_KERNEL) 2367 bev.pid = DEFAULT_GUEST_KERNEL_ID; 2368 2369 memcpy(bev.build_id, old_bev.build_id, sizeof(bev.build_id)); 2370 __event_process_build_id(&bev, filename, session); 2371 2372 offset += bev.header.size; 2373 } 2374 2375 return 0; 2376 } 2377 2378 static int perf_header__read_build_ids(struct perf_header *header, 2379 int input, u64 offset, u64 size) 2380 { 2381 struct perf_session *session = container_of(header, struct perf_session, header); 2382 struct perf_record_header_build_id bev; 2383 char filename[PATH_MAX]; 2384 u64 limit = offset + size, orig_offset = offset; 2385 int err = -1; 2386 2387 while (offset < limit) { 2388 ssize_t len; 2389 2390 if (readn(input, &bev, sizeof(bev)) != sizeof(bev)) 2391 goto out; 2392 2393 if (header->needs_swap) 2394 perf_event_header__bswap(&bev.header); 2395 2396 len = bev.header.size - sizeof(bev); 2397 if (readn(input, filename, len) != len) 2398 goto out; 2399 /* 2400 * The a1645ce1 changeset: 2401 * 2402 * "perf: 'perf kvm' tool for monitoring guest performance from host" 2403 * 2404 * Added a field to struct perf_record_header_build_id that broke the file 2405 * format. 2406 * 2407 * Since the kernel build-id is the first entry, process the 2408 * table using the old format if the well known 2409 * '[kernel.kallsyms]' string for the kernel build-id has the 2410 * first 4 characters chopped off (where the pid_t sits). 2411 */ 2412 if (memcmp(filename, "nel.kallsyms]", 13) == 0) { 2413 if (lseek(input, orig_offset, SEEK_SET) == (off_t)-1) 2414 return -1; 2415 return perf_header__read_build_ids_abi_quirk(header, input, offset, size); 2416 } 2417 2418 __event_process_build_id(&bev, filename, session); 2419 2420 offset += bev.header.size; 2421 } 2422 err = 0; 2423 out: 2424 return err; 2425 } 2426 2427 /* Macro for features that simply need to read and store a string. */ 2428 #define FEAT_PROCESS_STR_FUN(__feat, __feat_env) \ 2429 static int process_##__feat(struct feat_fd *ff, void *data __maybe_unused) \ 2430 {\ 2431 free(ff->ph->env.__feat_env); \ 2432 ff->ph->env.__feat_env = do_read_string(ff); \ 2433 return ff->ph->env.__feat_env ? 0 : -ENOMEM; \ 2434 } 2435 2436 FEAT_PROCESS_STR_FUN(hostname, hostname); 2437 FEAT_PROCESS_STR_FUN(osrelease, os_release); 2438 FEAT_PROCESS_STR_FUN(version, version); 2439 FEAT_PROCESS_STR_FUN(arch, arch); 2440 FEAT_PROCESS_STR_FUN(cpudesc, cpu_desc); 2441 FEAT_PROCESS_STR_FUN(cpuid, cpuid); 2442 2443 #ifdef HAVE_LIBTRACEEVENT 2444 static int process_tracing_data(struct feat_fd *ff, void *data) 2445 { 2446 ssize_t ret = trace_report(ff->fd, data, false); 2447 2448 return ret < 0 ? -1 : 0; 2449 } 2450 #endif 2451 2452 static int process_build_id(struct feat_fd *ff, void *data __maybe_unused) 2453 { 2454 if (perf_header__read_build_ids(ff->ph, ff->fd, ff->offset, ff->size)) 2455 pr_debug("Failed to read buildids, continuing...\n"); 2456 return 0; 2457 } 2458 2459 static int process_nrcpus(struct feat_fd *ff, void *data __maybe_unused) 2460 { 2461 int ret; 2462 u32 nr_cpus_avail, nr_cpus_online; 2463 2464 ret = do_read_u32(ff, &nr_cpus_avail); 2465 if (ret) 2466 return ret; 2467 2468 ret = do_read_u32(ff, &nr_cpus_online); 2469 if (ret) 2470 return ret; 2471 ff->ph->env.nr_cpus_avail = (int)nr_cpus_avail; 2472 ff->ph->env.nr_cpus_online = (int)nr_cpus_online; 2473 return 0; 2474 } 2475 2476 static int process_total_mem(struct feat_fd *ff, void *data __maybe_unused) 2477 { 2478 u64 total_mem; 2479 int ret; 2480 2481 ret = do_read_u64(ff, &total_mem); 2482 if (ret) 2483 return -1; 2484 ff->ph->env.total_mem = (unsigned long long)total_mem; 2485 return 0; 2486 } 2487 2488 static struct evsel *evlist__find_by_index(struct evlist *evlist, int idx) 2489 { 2490 struct evsel *evsel; 2491 2492 evlist__for_each_entry(evlist, evsel) { 2493 if (evsel->core.idx == idx) 2494 return evsel; 2495 } 2496 2497 return NULL; 2498 } 2499 2500 static void evlist__set_event_name(struct evlist *evlist, struct evsel *event) 2501 { 2502 struct evsel *evsel; 2503 2504 if (!event->name) 2505 return; 2506 2507 evsel = evlist__find_by_index(evlist, event->core.idx); 2508 if (!evsel) 2509 return; 2510 2511 if (evsel->name) 2512 return; 2513 2514 evsel->name = strdup(event->name); 2515 } 2516 2517 static int 2518 process_event_desc(struct feat_fd *ff, void *data __maybe_unused) 2519 { 2520 struct perf_session *session; 2521 struct evsel *evsel, *events = read_event_desc(ff); 2522 2523 if (!events) 2524 return 0; 2525 2526 session = container_of(ff->ph, struct perf_session, header); 2527 2528 if (session->data->is_pipe) { 2529 /* Save events for reading later by print_event_desc, 2530 * since they can't be read again in pipe mode. */ 2531 ff->events = events; 2532 } 2533 2534 for (evsel = events; evsel->core.attr.size; evsel++) 2535 evlist__set_event_name(session->evlist, evsel); 2536 2537 if (!session->data->is_pipe) 2538 free_event_desc(events); 2539 2540 return 0; 2541 } 2542 2543 static int process_cmdline(struct feat_fd *ff, void *data __maybe_unused) 2544 { 2545 char *str, *cmdline = NULL, **argv = NULL; 2546 u32 nr, i, len = 0; 2547 2548 if (do_read_u32(ff, &nr)) 2549 return -1; 2550 2551 ff->ph->env.nr_cmdline = nr; 2552 2553 cmdline = zalloc(ff->size + nr + 1); 2554 if (!cmdline) 2555 return -1; 2556 2557 argv = zalloc(sizeof(char *) * (nr + 1)); 2558 if (!argv) 2559 goto error; 2560 2561 for (i = 0; i < nr; i++) { 2562 str = do_read_string(ff); 2563 if (!str) 2564 goto error; 2565 2566 argv[i] = cmdline + len; 2567 memcpy(argv[i], str, strlen(str) + 1); 2568 len += strlen(str) + 1; 2569 free(str); 2570 } 2571 ff->ph->env.cmdline = cmdline; 2572 ff->ph->env.cmdline_argv = (const char **) argv; 2573 return 0; 2574 2575 error: 2576 free(argv); 2577 free(cmdline); 2578 return -1; 2579 } 2580 2581 static int process_cpu_topology(struct feat_fd *ff, void *data __maybe_unused) 2582 { 2583 u32 nr, i; 2584 char *str = NULL; 2585 struct strbuf sb; 2586 int cpu_nr = ff->ph->env.nr_cpus_avail; 2587 u64 size = 0; 2588 struct perf_header *ph = ff->ph; 2589 bool do_core_id_test = true; 2590 2591 ph->env.cpu = calloc(cpu_nr, sizeof(*ph->env.cpu)); 2592 if (!ph->env.cpu) 2593 return -1; 2594 2595 if (do_read_u32(ff, &nr)) 2596 goto free_cpu; 2597 2598 ph->env.nr_sibling_cores = nr; 2599 size += sizeof(u32); 2600 if (strbuf_init(&sb, 128) < 0) 2601 goto free_cpu; 2602 2603 for (i = 0; i < nr; i++) { 2604 str = do_read_string(ff); 2605 if (!str) 2606 goto error; 2607 2608 /* include a NULL character at the end */ 2609 if (strbuf_add(&sb, str, strlen(str) + 1) < 0) 2610 goto error; 2611 size += string_size(str); 2612 zfree(&str); 2613 } 2614 ph->env.sibling_cores = strbuf_detach(&sb, NULL); 2615 2616 if (do_read_u32(ff, &nr)) 2617 return -1; 2618 2619 ph->env.nr_sibling_threads = nr; 2620 size += sizeof(u32); 2621 2622 for (i = 0; i < nr; i++) { 2623 str = do_read_string(ff); 2624 if (!str) 2625 goto error; 2626 2627 /* include a NULL character at the end */ 2628 if (strbuf_add(&sb, str, strlen(str) + 1) < 0) 2629 goto error; 2630 size += string_size(str); 2631 zfree(&str); 2632 } 2633 ph->env.sibling_threads = strbuf_detach(&sb, NULL); 2634 2635 /* 2636 * The header may be from old perf, 2637 * which doesn't include core id and socket id information. 2638 */ 2639 if (ff->size <= size) { 2640 zfree(&ph->env.cpu); 2641 return 0; 2642 } 2643 2644 /* On s390 the socket_id number is not related to the numbers of cpus. 2645 * The socket_id number might be higher than the numbers of cpus. 2646 * This depends on the configuration. 2647 * AArch64 is the same. 2648 */ 2649 if (ph->env.arch && (!strncmp(ph->env.arch, "s390", 4) 2650 || !strncmp(ph->env.arch, "aarch64", 7))) 2651 do_core_id_test = false; 2652 2653 for (i = 0; i < (u32)cpu_nr; i++) { 2654 if (do_read_u32(ff, &nr)) 2655 goto free_cpu; 2656 2657 ph->env.cpu[i].core_id = nr; 2658 size += sizeof(u32); 2659 2660 if (do_read_u32(ff, &nr)) 2661 goto free_cpu; 2662 2663 if (do_core_id_test && nr != (u32)-1 && nr > (u32)cpu_nr) { 2664 pr_debug("socket_id number is too big." 2665 "You may need to upgrade the perf tool.\n"); 2666 goto free_cpu; 2667 } 2668 2669 ph->env.cpu[i].socket_id = nr; 2670 size += sizeof(u32); 2671 } 2672 2673 /* 2674 * The header may be from old perf, 2675 * which doesn't include die information. 2676 */ 2677 if (ff->size <= size) 2678 return 0; 2679 2680 if (do_read_u32(ff, &nr)) 2681 return -1; 2682 2683 ph->env.nr_sibling_dies = nr; 2684 size += sizeof(u32); 2685 2686 for (i = 0; i < nr; i++) { 2687 str = do_read_string(ff); 2688 if (!str) 2689 goto error; 2690 2691 /* include a NULL character at the end */ 2692 if (strbuf_add(&sb, str, strlen(str) + 1) < 0) 2693 goto error; 2694 size += string_size(str); 2695 zfree(&str); 2696 } 2697 ph->env.sibling_dies = strbuf_detach(&sb, NULL); 2698 2699 for (i = 0; i < (u32)cpu_nr; i++) { 2700 if (do_read_u32(ff, &nr)) 2701 goto free_cpu; 2702 2703 ph->env.cpu[i].die_id = nr; 2704 } 2705 2706 return 0; 2707 2708 error: 2709 strbuf_release(&sb); 2710 zfree(&str); 2711 free_cpu: 2712 zfree(&ph->env.cpu); 2713 return -1; 2714 } 2715 2716 static int process_numa_topology(struct feat_fd *ff, void *data __maybe_unused) 2717 { 2718 struct numa_node *nodes, *n; 2719 u32 nr, i; 2720 char *str; 2721 2722 /* nr nodes */ 2723 if (do_read_u32(ff, &nr)) 2724 return -1; 2725 2726 nodes = zalloc(sizeof(*nodes) * nr); 2727 if (!nodes) 2728 return -ENOMEM; 2729 2730 for (i = 0; i < nr; i++) { 2731 n = &nodes[i]; 2732 2733 /* node number */ 2734 if (do_read_u32(ff, &n->node)) 2735 goto error; 2736 2737 if (do_read_u64(ff, &n->mem_total)) 2738 goto error; 2739 2740 if (do_read_u64(ff, &n->mem_free)) 2741 goto error; 2742 2743 str = do_read_string(ff); 2744 if (!str) 2745 goto error; 2746 2747 n->map = perf_cpu_map__new(str); 2748 free(str); 2749 if (!n->map) 2750 goto error; 2751 } 2752 ff->ph->env.nr_numa_nodes = nr; 2753 ff->ph->env.numa_nodes = nodes; 2754 return 0; 2755 2756 error: 2757 free(nodes); 2758 return -1; 2759 } 2760 2761 static int process_pmu_mappings(struct feat_fd *ff, void *data __maybe_unused) 2762 { 2763 char *name; 2764 u32 pmu_num; 2765 u32 type; 2766 struct strbuf sb; 2767 2768 if (do_read_u32(ff, &pmu_num)) 2769 return -1; 2770 2771 if (!pmu_num) { 2772 pr_debug("pmu mappings not available\n"); 2773 return 0; 2774 } 2775 2776 ff->ph->env.nr_pmu_mappings = pmu_num; 2777 if (strbuf_init(&sb, 128) < 0) 2778 return -1; 2779 2780 while (pmu_num) { 2781 if (do_read_u32(ff, &type)) 2782 goto error; 2783 2784 name = do_read_string(ff); 2785 if (!name) 2786 goto error; 2787 2788 if (strbuf_addf(&sb, "%u:%s", type, name) < 0) 2789 goto error; 2790 /* include a NULL character at the end */ 2791 if (strbuf_add(&sb, "", 1) < 0) 2792 goto error; 2793 2794 if (!strcmp(name, "msr")) 2795 ff->ph->env.msr_pmu_type = type; 2796 2797 free(name); 2798 pmu_num--; 2799 } 2800 ff->ph->env.pmu_mappings = strbuf_detach(&sb, NULL); 2801 return 0; 2802 2803 error: 2804 strbuf_release(&sb); 2805 return -1; 2806 } 2807 2808 static int process_group_desc(struct feat_fd *ff, void *data __maybe_unused) 2809 { 2810 size_t ret = -1; 2811 u32 i, nr, nr_groups; 2812 struct perf_session *session; 2813 struct evsel *evsel, *leader = NULL; 2814 struct group_desc { 2815 char *name; 2816 u32 leader_idx; 2817 u32 nr_members; 2818 } *desc; 2819 2820 if (do_read_u32(ff, &nr_groups)) 2821 return -1; 2822 2823 ff->ph->env.nr_groups = nr_groups; 2824 if (!nr_groups) { 2825 pr_debug("group desc not available\n"); 2826 return 0; 2827 } 2828 2829 desc = calloc(nr_groups, sizeof(*desc)); 2830 if (!desc) 2831 return -1; 2832 2833 for (i = 0; i < nr_groups; i++) { 2834 desc[i].name = do_read_string(ff); 2835 if (!desc[i].name) 2836 goto out_free; 2837 2838 if (do_read_u32(ff, &desc[i].leader_idx)) 2839 goto out_free; 2840 2841 if (do_read_u32(ff, &desc[i].nr_members)) 2842 goto out_free; 2843 } 2844 2845 /* 2846 * Rebuild group relationship based on the group_desc 2847 */ 2848 session = container_of(ff->ph, struct perf_session, header); 2849 2850 i = nr = 0; 2851 evlist__for_each_entry(session->evlist, evsel) { 2852 if (i < nr_groups && evsel->core.idx == (int) desc[i].leader_idx) { 2853 evsel__set_leader(evsel, evsel); 2854 /* {anon_group} is a dummy name */ 2855 if (strcmp(desc[i].name, "{anon_group}")) { 2856 evsel->group_name = desc[i].name; 2857 desc[i].name = NULL; 2858 } 2859 evsel->core.nr_members = desc[i].nr_members; 2860 2861 if (i >= nr_groups || nr > 0) { 2862 pr_debug("invalid group desc\n"); 2863 goto out_free; 2864 } 2865 2866 leader = evsel; 2867 nr = evsel->core.nr_members - 1; 2868 i++; 2869 } else if (nr) { 2870 /* This is a group member */ 2871 evsel__set_leader(evsel, leader); 2872 2873 nr--; 2874 } 2875 } 2876 2877 if (i != nr_groups || nr != 0) { 2878 pr_debug("invalid group desc\n"); 2879 goto out_free; 2880 } 2881 2882 ret = 0; 2883 out_free: 2884 for (i = 0; i < nr_groups; i++) 2885 zfree(&desc[i].name); 2886 free(desc); 2887 2888 return ret; 2889 } 2890 2891 static int process_auxtrace(struct feat_fd *ff, void *data __maybe_unused) 2892 { 2893 struct perf_session *session; 2894 int err; 2895 2896 session = container_of(ff->ph, struct perf_session, header); 2897 2898 err = auxtrace_index__process(ff->fd, ff->size, session, 2899 ff->ph->needs_swap); 2900 if (err < 0) 2901 pr_err("Failed to process auxtrace index\n"); 2902 return err; 2903 } 2904 2905 static int process_cache(struct feat_fd *ff, void *data __maybe_unused) 2906 { 2907 struct cpu_cache_level *caches; 2908 u32 cnt, i, version; 2909 2910 if (do_read_u32(ff, &version)) 2911 return -1; 2912 2913 if (version != 1) 2914 return -1; 2915 2916 if (do_read_u32(ff, &cnt)) 2917 return -1; 2918 2919 caches = zalloc(sizeof(*caches) * cnt); 2920 if (!caches) 2921 return -1; 2922 2923 for (i = 0; i < cnt; i++) { 2924 struct cpu_cache_level *c = &caches[i]; 2925 2926 #define _R(v) \ 2927 if (do_read_u32(ff, &c->v)) \ 2928 goto out_free_caches; \ 2929 2930 _R(level) 2931 _R(line_size) 2932 _R(sets) 2933 _R(ways) 2934 #undef _R 2935 2936 #define _R(v) \ 2937 c->v = do_read_string(ff); \ 2938 if (!c->v) \ 2939 goto out_free_caches; \ 2940 2941 _R(type) 2942 _R(size) 2943 _R(map) 2944 #undef _R 2945 } 2946 2947 ff->ph->env.caches = caches; 2948 ff->ph->env.caches_cnt = cnt; 2949 return 0; 2950 out_free_caches: 2951 for (i = 0; i < cnt; i++) { 2952 free(caches[i].type); 2953 free(caches[i].size); 2954 free(caches[i].map); 2955 } 2956 free(caches); 2957 return -1; 2958 } 2959 2960 static int process_sample_time(struct feat_fd *ff, void *data __maybe_unused) 2961 { 2962 struct perf_session *session; 2963 u64 first_sample_time, last_sample_time; 2964 int ret; 2965 2966 session = container_of(ff->ph, struct perf_session, header); 2967 2968 ret = do_read_u64(ff, &first_sample_time); 2969 if (ret) 2970 return -1; 2971 2972 ret = do_read_u64(ff, &last_sample_time); 2973 if (ret) 2974 return -1; 2975 2976 session->evlist->first_sample_time = first_sample_time; 2977 session->evlist->last_sample_time = last_sample_time; 2978 return 0; 2979 } 2980 2981 static int process_mem_topology(struct feat_fd *ff, 2982 void *data __maybe_unused) 2983 { 2984 struct memory_node *nodes; 2985 u64 version, i, nr, bsize; 2986 int ret = -1; 2987 2988 if (do_read_u64(ff, &version)) 2989 return -1; 2990 2991 if (version != 1) 2992 return -1; 2993 2994 if (do_read_u64(ff, &bsize)) 2995 return -1; 2996 2997 if (do_read_u64(ff, &nr)) 2998 return -1; 2999 3000 nodes = zalloc(sizeof(*nodes) * nr); 3001 if (!nodes) 3002 return -1; 3003 3004 for (i = 0; i < nr; i++) { 3005 struct memory_node n; 3006 3007 #define _R(v) \ 3008 if (do_read_u64(ff, &n.v)) \ 3009 goto out; \ 3010 3011 _R(node) 3012 _R(size) 3013 3014 #undef _R 3015 3016 if (do_read_bitmap(ff, &n.set, &n.size)) 3017 goto out; 3018 3019 nodes[i] = n; 3020 } 3021 3022 ff->ph->env.memory_bsize = bsize; 3023 ff->ph->env.memory_nodes = nodes; 3024 ff->ph->env.nr_memory_nodes = nr; 3025 ret = 0; 3026 3027 out: 3028 if (ret) 3029 free(nodes); 3030 return ret; 3031 } 3032 3033 static int process_clockid(struct feat_fd *ff, 3034 void *data __maybe_unused) 3035 { 3036 if (do_read_u64(ff, &ff->ph->env.clock.clockid_res_ns)) 3037 return -1; 3038 3039 return 0; 3040 } 3041 3042 static int process_clock_data(struct feat_fd *ff, 3043 void *_data __maybe_unused) 3044 { 3045 u32 data32; 3046 u64 data64; 3047 3048 /* version */ 3049 if (do_read_u32(ff, &data32)) 3050 return -1; 3051 3052 if (data32 != 1) 3053 return -1; 3054 3055 /* clockid */ 3056 if (do_read_u32(ff, &data32)) 3057 return -1; 3058 3059 ff->ph->env.clock.clockid = data32; 3060 3061 /* TOD ref time */ 3062 if (do_read_u64(ff, &data64)) 3063 return -1; 3064 3065 ff->ph->env.clock.tod_ns = data64; 3066 3067 /* clockid ref time */ 3068 if (do_read_u64(ff, &data64)) 3069 return -1; 3070 3071 ff->ph->env.clock.clockid_ns = data64; 3072 ff->ph->env.clock.enabled = true; 3073 return 0; 3074 } 3075 3076 static int process_hybrid_topology(struct feat_fd *ff, 3077 void *data __maybe_unused) 3078 { 3079 struct hybrid_node *nodes, *n; 3080 u32 nr, i; 3081 3082 /* nr nodes */ 3083 if (do_read_u32(ff, &nr)) 3084 return -1; 3085 3086 nodes = zalloc(sizeof(*nodes) * nr); 3087 if (!nodes) 3088 return -ENOMEM; 3089 3090 for (i = 0; i < nr; i++) { 3091 n = &nodes[i]; 3092 3093 n->pmu_name = do_read_string(ff); 3094 if (!n->pmu_name) 3095 goto error; 3096 3097 n->cpus = do_read_string(ff); 3098 if (!n->cpus) 3099 goto error; 3100 } 3101 3102 ff->ph->env.nr_hybrid_nodes = nr; 3103 ff->ph->env.hybrid_nodes = nodes; 3104 return 0; 3105 3106 error: 3107 for (i = 0; i < nr; i++) { 3108 free(nodes[i].pmu_name); 3109 free(nodes[i].cpus); 3110 } 3111 3112 free(nodes); 3113 return -1; 3114 } 3115 3116 static int process_dir_format(struct feat_fd *ff, 3117 void *_data __maybe_unused) 3118 { 3119 struct perf_session *session; 3120 struct perf_data *data; 3121 3122 session = container_of(ff->ph, struct perf_session, header); 3123 data = session->data; 3124 3125 if (WARN_ON(!perf_data__is_dir(data))) 3126 return -1; 3127 3128 return do_read_u64(ff, &data->dir.version); 3129 } 3130 3131 #ifdef HAVE_LIBBPF_SUPPORT 3132 static int process_bpf_prog_info(struct feat_fd *ff, void *data __maybe_unused) 3133 { 3134 struct bpf_prog_info_node *info_node; 3135 struct perf_env *env = &ff->ph->env; 3136 struct perf_bpil *info_linear; 3137 u32 count, i; 3138 int err = -1; 3139 3140 if (ff->ph->needs_swap) { 3141 pr_warning("interpreting bpf_prog_info from systems with endianness is not yet supported\n"); 3142 return 0; 3143 } 3144 3145 if (do_read_u32(ff, &count)) 3146 return -1; 3147 3148 down_write(&env->bpf_progs.lock); 3149 3150 for (i = 0; i < count; ++i) { 3151 u32 info_len, data_len; 3152 3153 info_linear = NULL; 3154 info_node = NULL; 3155 if (do_read_u32(ff, &info_len)) 3156 goto out; 3157 if (do_read_u32(ff, &data_len)) 3158 goto out; 3159 3160 if (info_len > sizeof(struct bpf_prog_info)) { 3161 pr_warning("detected invalid bpf_prog_info\n"); 3162 goto out; 3163 } 3164 3165 info_linear = malloc(sizeof(struct perf_bpil) + 3166 data_len); 3167 if (!info_linear) 3168 goto out; 3169 info_linear->info_len = sizeof(struct bpf_prog_info); 3170 info_linear->data_len = data_len; 3171 if (do_read_u64(ff, (u64 *)(&info_linear->arrays))) 3172 goto out; 3173 if (__do_read(ff, &info_linear->info, info_len)) 3174 goto out; 3175 if (info_len < sizeof(struct bpf_prog_info)) 3176 memset(((void *)(&info_linear->info)) + info_len, 0, 3177 sizeof(struct bpf_prog_info) - info_len); 3178 3179 if (__do_read(ff, info_linear->data, data_len)) 3180 goto out; 3181 3182 info_node = malloc(sizeof(struct bpf_prog_info_node)); 3183 if (!info_node) 3184 goto out; 3185 3186 /* after reading from file, translate offset to address */ 3187 bpil_offs_to_addr(info_linear); 3188 info_node->info_linear = info_linear; 3189 perf_env__insert_bpf_prog_info(env, info_node); 3190 } 3191 3192 up_write(&env->bpf_progs.lock); 3193 return 0; 3194 out: 3195 free(info_linear); 3196 free(info_node); 3197 up_write(&env->bpf_progs.lock); 3198 return err; 3199 } 3200 3201 static int process_bpf_btf(struct feat_fd *ff, void *data __maybe_unused) 3202 { 3203 struct perf_env *env = &ff->ph->env; 3204 struct btf_node *node = NULL; 3205 u32 count, i; 3206 int err = -1; 3207 3208 if (ff->ph->needs_swap) { 3209 pr_warning("interpreting btf from systems with endianness is not yet supported\n"); 3210 return 0; 3211 } 3212 3213 if (do_read_u32(ff, &count)) 3214 return -1; 3215 3216 down_write(&env->bpf_progs.lock); 3217 3218 for (i = 0; i < count; ++i) { 3219 u32 id, data_size; 3220 3221 if (do_read_u32(ff, &id)) 3222 goto out; 3223 if (do_read_u32(ff, &data_size)) 3224 goto out; 3225 3226 node = malloc(sizeof(struct btf_node) + data_size); 3227 if (!node) 3228 goto out; 3229 3230 node->id = id; 3231 node->data_size = data_size; 3232 3233 if (__do_read(ff, node->data, data_size)) 3234 goto out; 3235 3236 perf_env__insert_btf(env, node); 3237 node = NULL; 3238 } 3239 3240 err = 0; 3241 out: 3242 up_write(&env->bpf_progs.lock); 3243 free(node); 3244 return err; 3245 } 3246 #endif // HAVE_LIBBPF_SUPPORT 3247 3248 static int process_compressed(struct feat_fd *ff, 3249 void *data __maybe_unused) 3250 { 3251 if (do_read_u32(ff, &(ff->ph->env.comp_ver))) 3252 return -1; 3253 3254 if (do_read_u32(ff, &(ff->ph->env.comp_type))) 3255 return -1; 3256 3257 if (do_read_u32(ff, &(ff->ph->env.comp_level))) 3258 return -1; 3259 3260 if (do_read_u32(ff, &(ff->ph->env.comp_ratio))) 3261 return -1; 3262 3263 if (do_read_u32(ff, &(ff->ph->env.comp_mmap_len))) 3264 return -1; 3265 3266 return 0; 3267 } 3268 3269 static int __process_pmu_caps(struct feat_fd *ff, int *nr_caps, 3270 char ***caps, unsigned int *max_branches, 3271 unsigned int *br_cntr_nr, 3272 unsigned int *br_cntr_width) 3273 { 3274 char *name, *value, *ptr; 3275 u32 nr_pmu_caps, i; 3276 3277 *nr_caps = 0; 3278 *caps = NULL; 3279 3280 if (do_read_u32(ff, &nr_pmu_caps)) 3281 return -1; 3282 3283 if (!nr_pmu_caps) 3284 return 0; 3285 3286 *caps = zalloc(sizeof(char *) * nr_pmu_caps); 3287 if (!*caps) 3288 return -1; 3289 3290 for (i = 0; i < nr_pmu_caps; i++) { 3291 name = do_read_string(ff); 3292 if (!name) 3293 goto error; 3294 3295 value = do_read_string(ff); 3296 if (!value) 3297 goto free_name; 3298 3299 if (asprintf(&ptr, "%s=%s", name, value) < 0) 3300 goto free_value; 3301 3302 (*caps)[i] = ptr; 3303 3304 if (!strcmp(name, "branches")) 3305 *max_branches = atoi(value); 3306 3307 if (!strcmp(name, "branch_counter_nr")) 3308 *br_cntr_nr = atoi(value); 3309 3310 if (!strcmp(name, "branch_counter_width")) 3311 *br_cntr_width = atoi(value); 3312 3313 free(value); 3314 free(name); 3315 } 3316 *nr_caps = nr_pmu_caps; 3317 return 0; 3318 3319 free_value: 3320 free(value); 3321 free_name: 3322 free(name); 3323 error: 3324 for (; i > 0; i--) 3325 free((*caps)[i - 1]); 3326 free(*caps); 3327 *caps = NULL; 3328 *nr_caps = 0; 3329 return -1; 3330 } 3331 3332 static int process_cpu_pmu_caps(struct feat_fd *ff, 3333 void *data __maybe_unused) 3334 { 3335 int ret = __process_pmu_caps(ff, &ff->ph->env.nr_cpu_pmu_caps, 3336 &ff->ph->env.cpu_pmu_caps, 3337 &ff->ph->env.max_branches, 3338 &ff->ph->env.br_cntr_nr, 3339 &ff->ph->env.br_cntr_width); 3340 3341 if (!ret && !ff->ph->env.cpu_pmu_caps) 3342 pr_debug("cpu pmu capabilities not available\n"); 3343 return ret; 3344 } 3345 3346 static int process_pmu_caps(struct feat_fd *ff, void *data __maybe_unused) 3347 { 3348 struct pmu_caps *pmu_caps; 3349 u32 nr_pmu, i; 3350 int ret; 3351 int j; 3352 3353 if (do_read_u32(ff, &nr_pmu)) 3354 return -1; 3355 3356 if (!nr_pmu) { 3357 pr_debug("pmu capabilities not available\n"); 3358 return 0; 3359 } 3360 3361 pmu_caps = zalloc(sizeof(*pmu_caps) * nr_pmu); 3362 if (!pmu_caps) 3363 return -ENOMEM; 3364 3365 for (i = 0; i < nr_pmu; i++) { 3366 ret = __process_pmu_caps(ff, &pmu_caps[i].nr_caps, 3367 &pmu_caps[i].caps, 3368 &pmu_caps[i].max_branches, 3369 &pmu_caps[i].br_cntr_nr, 3370 &pmu_caps[i].br_cntr_width); 3371 if (ret) 3372 goto err; 3373 3374 pmu_caps[i].pmu_name = do_read_string(ff); 3375 if (!pmu_caps[i].pmu_name) { 3376 ret = -1; 3377 goto err; 3378 } 3379 if (!pmu_caps[i].nr_caps) { 3380 pr_debug("%s pmu capabilities not available\n", 3381 pmu_caps[i].pmu_name); 3382 } 3383 } 3384 3385 ff->ph->env.nr_pmus_with_caps = nr_pmu; 3386 ff->ph->env.pmu_caps = pmu_caps; 3387 return 0; 3388 3389 err: 3390 for (i = 0; i < nr_pmu; i++) { 3391 for (j = 0; j < pmu_caps[i].nr_caps; j++) 3392 free(pmu_caps[i].caps[j]); 3393 free(pmu_caps[i].caps); 3394 free(pmu_caps[i].pmu_name); 3395 } 3396 3397 free(pmu_caps); 3398 return ret; 3399 } 3400 3401 #define FEAT_OPR(n, func, __full_only) \ 3402 [HEADER_##n] = { \ 3403 .name = __stringify(n), \ 3404 .write = write_##func, \ 3405 .print = print_##func, \ 3406 .full_only = __full_only, \ 3407 .process = process_##func, \ 3408 .synthesize = true \ 3409 } 3410 3411 #define FEAT_OPN(n, func, __full_only) \ 3412 [HEADER_##n] = { \ 3413 .name = __stringify(n), \ 3414 .write = write_##func, \ 3415 .print = print_##func, \ 3416 .full_only = __full_only, \ 3417 .process = process_##func \ 3418 } 3419 3420 /* feature_ops not implemented: */ 3421 #define print_tracing_data NULL 3422 #define print_build_id NULL 3423 3424 #define process_branch_stack NULL 3425 #define process_stat NULL 3426 3427 // Only used in util/synthetic-events.c 3428 const struct perf_header_feature_ops feat_ops[HEADER_LAST_FEATURE]; 3429 3430 const struct perf_header_feature_ops feat_ops[HEADER_LAST_FEATURE] = { 3431 #ifdef HAVE_LIBTRACEEVENT 3432 FEAT_OPN(TRACING_DATA, tracing_data, false), 3433 #endif 3434 FEAT_OPN(BUILD_ID, build_id, false), 3435 FEAT_OPR(HOSTNAME, hostname, false), 3436 FEAT_OPR(OSRELEASE, osrelease, false), 3437 FEAT_OPR(VERSION, version, false), 3438 FEAT_OPR(ARCH, arch, false), 3439 FEAT_OPR(NRCPUS, nrcpus, false), 3440 FEAT_OPR(CPUDESC, cpudesc, false), 3441 FEAT_OPR(CPUID, cpuid, false), 3442 FEAT_OPR(TOTAL_MEM, total_mem, false), 3443 FEAT_OPR(EVENT_DESC, event_desc, false), 3444 FEAT_OPR(CMDLINE, cmdline, false), 3445 FEAT_OPR(CPU_TOPOLOGY, cpu_topology, true), 3446 FEAT_OPR(NUMA_TOPOLOGY, numa_topology, true), 3447 FEAT_OPN(BRANCH_STACK, branch_stack, false), 3448 FEAT_OPR(PMU_MAPPINGS, pmu_mappings, false), 3449 FEAT_OPR(GROUP_DESC, group_desc, false), 3450 FEAT_OPN(AUXTRACE, auxtrace, false), 3451 FEAT_OPN(STAT, stat, false), 3452 FEAT_OPN(CACHE, cache, true), 3453 FEAT_OPR(SAMPLE_TIME, sample_time, false), 3454 FEAT_OPR(MEM_TOPOLOGY, mem_topology, true), 3455 FEAT_OPR(CLOCKID, clockid, false), 3456 FEAT_OPN(DIR_FORMAT, dir_format, false), 3457 #ifdef HAVE_LIBBPF_SUPPORT 3458 FEAT_OPR(BPF_PROG_INFO, bpf_prog_info, false), 3459 FEAT_OPR(BPF_BTF, bpf_btf, false), 3460 #endif 3461 FEAT_OPR(COMPRESSED, compressed, false), 3462 FEAT_OPR(CPU_PMU_CAPS, cpu_pmu_caps, false), 3463 FEAT_OPR(CLOCK_DATA, clock_data, false), 3464 FEAT_OPN(HYBRID_TOPOLOGY, hybrid_topology, true), 3465 FEAT_OPR(PMU_CAPS, pmu_caps, false), 3466 }; 3467 3468 struct header_print_data { 3469 FILE *fp; 3470 bool full; /* extended list of headers */ 3471 }; 3472 3473 static int perf_file_section__fprintf_info(struct perf_file_section *section, 3474 struct perf_header *ph, 3475 int feat, int fd, void *data) 3476 { 3477 struct header_print_data *hd = data; 3478 struct feat_fd ff; 3479 3480 if (lseek(fd, section->offset, SEEK_SET) == (off_t)-1) { 3481 pr_debug("Failed to lseek to %" PRIu64 " offset for feature " 3482 "%d, continuing...\n", section->offset, feat); 3483 return 0; 3484 } 3485 if (feat >= HEADER_LAST_FEATURE) { 3486 pr_warning("unknown feature %d\n", feat); 3487 return 0; 3488 } 3489 if (!feat_ops[feat].print) 3490 return 0; 3491 3492 ff = (struct feat_fd) { 3493 .fd = fd, 3494 .ph = ph, 3495 }; 3496 3497 if (!feat_ops[feat].full_only || hd->full) 3498 feat_ops[feat].print(&ff, hd->fp); 3499 else 3500 fprintf(hd->fp, "# %s info available, use -I to display\n", 3501 feat_ops[feat].name); 3502 3503 return 0; 3504 } 3505 3506 int perf_header__fprintf_info(struct perf_session *session, FILE *fp, bool full) 3507 { 3508 struct header_print_data hd; 3509 struct perf_header *header = &session->header; 3510 int fd = perf_data__fd(session->data); 3511 struct stat st; 3512 time_t stctime; 3513 int ret, bit; 3514 3515 hd.fp = fp; 3516 hd.full = full; 3517 3518 ret = fstat(fd, &st); 3519 if (ret == -1) 3520 return -1; 3521 3522 stctime = st.st_mtime; 3523 fprintf(fp, "# captured on : %s", ctime(&stctime)); 3524 3525 fprintf(fp, "# header version : %u\n", header->version); 3526 fprintf(fp, "# data offset : %" PRIu64 "\n", header->data_offset); 3527 fprintf(fp, "# data size : %" PRIu64 "\n", header->data_size); 3528 fprintf(fp, "# feat offset : %" PRIu64 "\n", header->feat_offset); 3529 3530 perf_header__process_sections(header, fd, &hd, 3531 perf_file_section__fprintf_info); 3532 3533 if (session->data->is_pipe) 3534 return 0; 3535 3536 fprintf(fp, "# missing features: "); 3537 for_each_clear_bit(bit, header->adds_features, HEADER_LAST_FEATURE) { 3538 if (bit) 3539 fprintf(fp, "%s ", feat_ops[bit].name); 3540 } 3541 3542 fprintf(fp, "\n"); 3543 return 0; 3544 } 3545 3546 struct header_fw { 3547 struct feat_writer fw; 3548 struct feat_fd *ff; 3549 }; 3550 3551 static int feat_writer_cb(struct feat_writer *fw, void *buf, size_t sz) 3552 { 3553 struct header_fw *h = container_of(fw, struct header_fw, fw); 3554 3555 return do_write(h->ff, buf, sz); 3556 } 3557 3558 static int do_write_feat(struct feat_fd *ff, int type, 3559 struct perf_file_section **p, 3560 struct evlist *evlist, 3561 struct feat_copier *fc) 3562 { 3563 int err; 3564 int ret = 0; 3565 3566 if (perf_header__has_feat(ff->ph, type)) { 3567 if (!feat_ops[type].write) 3568 return -1; 3569 3570 if (WARN(ff->buf, "Error: calling %s in pipe-mode.\n", __func__)) 3571 return -1; 3572 3573 (*p)->offset = lseek(ff->fd, 0, SEEK_CUR); 3574 3575 /* 3576 * Hook to let perf inject copy features sections from the input 3577 * file. 3578 */ 3579 if (fc && fc->copy) { 3580 struct header_fw h = { 3581 .fw.write = feat_writer_cb, 3582 .ff = ff, 3583 }; 3584 3585 /* ->copy() returns 0 if the feature was not copied */ 3586 err = fc->copy(fc, type, &h.fw); 3587 } else { 3588 err = 0; 3589 } 3590 if (!err) 3591 err = feat_ops[type].write(ff, evlist); 3592 if (err < 0) { 3593 pr_debug("failed to write feature %s\n", feat_ops[type].name); 3594 3595 /* undo anything written */ 3596 lseek(ff->fd, (*p)->offset, SEEK_SET); 3597 3598 return -1; 3599 } 3600 (*p)->size = lseek(ff->fd, 0, SEEK_CUR) - (*p)->offset; 3601 (*p)++; 3602 } 3603 return ret; 3604 } 3605 3606 static int perf_header__adds_write(struct perf_header *header, 3607 struct evlist *evlist, int fd, 3608 struct feat_copier *fc) 3609 { 3610 int nr_sections; 3611 struct feat_fd ff = { 3612 .fd = fd, 3613 .ph = header, 3614 }; 3615 struct perf_file_section *feat_sec, *p; 3616 int sec_size; 3617 u64 sec_start; 3618 int feat; 3619 int err; 3620 3621 nr_sections = bitmap_weight(header->adds_features, HEADER_FEAT_BITS); 3622 if (!nr_sections) 3623 return 0; 3624 3625 feat_sec = p = calloc(nr_sections, sizeof(*feat_sec)); 3626 if (feat_sec == NULL) 3627 return -ENOMEM; 3628 3629 sec_size = sizeof(*feat_sec) * nr_sections; 3630 3631 sec_start = header->feat_offset; 3632 lseek(fd, sec_start + sec_size, SEEK_SET); 3633 3634 for_each_set_bit(feat, header->adds_features, HEADER_FEAT_BITS) { 3635 if (do_write_feat(&ff, feat, &p, evlist, fc)) 3636 perf_header__clear_feat(header, feat); 3637 } 3638 3639 lseek(fd, sec_start, SEEK_SET); 3640 /* 3641 * may write more than needed due to dropped feature, but 3642 * this is okay, reader will skip the missing entries 3643 */ 3644 err = do_write(&ff, feat_sec, sec_size); 3645 if (err < 0) 3646 pr_debug("failed to write feature section\n"); 3647 free(ff.buf); /* TODO: added to silence clang-tidy. */ 3648 free(feat_sec); 3649 return err; 3650 } 3651 3652 int perf_header__write_pipe(int fd) 3653 { 3654 struct perf_pipe_file_header f_header; 3655 struct feat_fd ff = { 3656 .fd = fd, 3657 }; 3658 int err; 3659 3660 f_header = (struct perf_pipe_file_header){ 3661 .magic = PERF_MAGIC, 3662 .size = sizeof(f_header), 3663 }; 3664 3665 err = do_write(&ff, &f_header, sizeof(f_header)); 3666 if (err < 0) { 3667 pr_debug("failed to write perf pipe header\n"); 3668 return err; 3669 } 3670 free(ff.buf); 3671 return 0; 3672 } 3673 3674 static int perf_session__do_write_header(struct perf_session *session, 3675 struct evlist *evlist, 3676 int fd, bool at_exit, 3677 struct feat_copier *fc) 3678 { 3679 struct perf_file_header f_header; 3680 struct perf_file_attr f_attr; 3681 struct perf_header *header = &session->header; 3682 struct evsel *evsel; 3683 struct feat_fd ff = { 3684 .fd = fd, 3685 }; 3686 u64 attr_offset; 3687 int err; 3688 3689 lseek(fd, sizeof(f_header), SEEK_SET); 3690 3691 evlist__for_each_entry(session->evlist, evsel) { 3692 evsel->id_offset = lseek(fd, 0, SEEK_CUR); 3693 err = do_write(&ff, evsel->core.id, evsel->core.ids * sizeof(u64)); 3694 if (err < 0) { 3695 pr_debug("failed to write perf header\n"); 3696 free(ff.buf); 3697 return err; 3698 } 3699 } 3700 3701 attr_offset = lseek(ff.fd, 0, SEEK_CUR); 3702 3703 evlist__for_each_entry(evlist, evsel) { 3704 if (evsel->core.attr.size < sizeof(evsel->core.attr)) { 3705 /* 3706 * We are likely in "perf inject" and have read 3707 * from an older file. Update attr size so that 3708 * reader gets the right offset to the ids. 3709 */ 3710 evsel->core.attr.size = sizeof(evsel->core.attr); 3711 } 3712 f_attr = (struct perf_file_attr){ 3713 .attr = evsel->core.attr, 3714 .ids = { 3715 .offset = evsel->id_offset, 3716 .size = evsel->core.ids * sizeof(u64), 3717 } 3718 }; 3719 err = do_write(&ff, &f_attr, sizeof(f_attr)); 3720 if (err < 0) { 3721 pr_debug("failed to write perf header attribute\n"); 3722 free(ff.buf); 3723 return err; 3724 } 3725 } 3726 3727 if (!header->data_offset) 3728 header->data_offset = lseek(fd, 0, SEEK_CUR); 3729 header->feat_offset = header->data_offset + header->data_size; 3730 3731 if (at_exit) { 3732 err = perf_header__adds_write(header, evlist, fd, fc); 3733 if (err < 0) { 3734 free(ff.buf); 3735 return err; 3736 } 3737 } 3738 3739 f_header = (struct perf_file_header){ 3740 .magic = PERF_MAGIC, 3741 .size = sizeof(f_header), 3742 .attr_size = sizeof(f_attr), 3743 .attrs = { 3744 .offset = attr_offset, 3745 .size = evlist->core.nr_entries * sizeof(f_attr), 3746 }, 3747 .data = { 3748 .offset = header->data_offset, 3749 .size = header->data_size, 3750 }, 3751 /* event_types is ignored, store zeros */ 3752 }; 3753 3754 memcpy(&f_header.adds_features, &header->adds_features, sizeof(header->adds_features)); 3755 3756 lseek(fd, 0, SEEK_SET); 3757 err = do_write(&ff, &f_header, sizeof(f_header)); 3758 free(ff.buf); 3759 if (err < 0) { 3760 pr_debug("failed to write perf header\n"); 3761 return err; 3762 } 3763 lseek(fd, header->data_offset + header->data_size, SEEK_SET); 3764 3765 return 0; 3766 } 3767 3768 int perf_session__write_header(struct perf_session *session, 3769 struct evlist *evlist, 3770 int fd, bool at_exit) 3771 { 3772 return perf_session__do_write_header(session, evlist, fd, at_exit, NULL); 3773 } 3774 3775 size_t perf_session__data_offset(const struct evlist *evlist) 3776 { 3777 struct evsel *evsel; 3778 size_t data_offset; 3779 3780 data_offset = sizeof(struct perf_file_header); 3781 evlist__for_each_entry(evlist, evsel) { 3782 data_offset += evsel->core.ids * sizeof(u64); 3783 } 3784 data_offset += evlist->core.nr_entries * sizeof(struct perf_file_attr); 3785 3786 return data_offset; 3787 } 3788 3789 int perf_session__inject_header(struct perf_session *session, 3790 struct evlist *evlist, 3791 int fd, 3792 struct feat_copier *fc) 3793 { 3794 return perf_session__do_write_header(session, evlist, fd, true, fc); 3795 } 3796 3797 static int perf_header__getbuffer64(struct perf_header *header, 3798 int fd, void *buf, size_t size) 3799 { 3800 if (readn(fd, buf, size) <= 0) 3801 return -1; 3802 3803 if (header->needs_swap) 3804 mem_bswap_64(buf, size); 3805 3806 return 0; 3807 } 3808 3809 int perf_header__process_sections(struct perf_header *header, int fd, 3810 void *data, 3811 int (*process)(struct perf_file_section *section, 3812 struct perf_header *ph, 3813 int feat, int fd, void *data)) 3814 { 3815 struct perf_file_section *feat_sec, *sec; 3816 int nr_sections; 3817 int sec_size; 3818 int feat; 3819 int err; 3820 3821 nr_sections = bitmap_weight(header->adds_features, HEADER_FEAT_BITS); 3822 if (!nr_sections) 3823 return 0; 3824 3825 feat_sec = sec = calloc(nr_sections, sizeof(*feat_sec)); 3826 if (!feat_sec) 3827 return -1; 3828 3829 sec_size = sizeof(*feat_sec) * nr_sections; 3830 3831 lseek(fd, header->feat_offset, SEEK_SET); 3832 3833 err = perf_header__getbuffer64(header, fd, feat_sec, sec_size); 3834 if (err < 0) 3835 goto out_free; 3836 3837 for_each_set_bit(feat, header->adds_features, HEADER_LAST_FEATURE) { 3838 err = process(sec++, header, feat, fd, data); 3839 if (err < 0) 3840 goto out_free; 3841 } 3842 err = 0; 3843 out_free: 3844 free(feat_sec); 3845 return err; 3846 } 3847 3848 static const int attr_file_abi_sizes[] = { 3849 [0] = PERF_ATTR_SIZE_VER0, 3850 [1] = PERF_ATTR_SIZE_VER1, 3851 [2] = PERF_ATTR_SIZE_VER2, 3852 [3] = PERF_ATTR_SIZE_VER3, 3853 [4] = PERF_ATTR_SIZE_VER4, 3854 0, 3855 }; 3856 3857 /* 3858 * In the legacy file format, the magic number is not used to encode endianness. 3859 * hdr_sz was used to encode endianness. But given that hdr_sz can vary based 3860 * on ABI revisions, we need to try all combinations for all endianness to 3861 * detect the endianness. 3862 */ 3863 static int try_all_file_abis(uint64_t hdr_sz, struct perf_header *ph) 3864 { 3865 uint64_t ref_size, attr_size; 3866 int i; 3867 3868 for (i = 0 ; attr_file_abi_sizes[i]; i++) { 3869 ref_size = attr_file_abi_sizes[i] 3870 + sizeof(struct perf_file_section); 3871 if (hdr_sz != ref_size) { 3872 attr_size = bswap_64(hdr_sz); 3873 if (attr_size != ref_size) 3874 continue; 3875 3876 ph->needs_swap = true; 3877 } 3878 pr_debug("ABI%d perf.data file detected, need_swap=%d\n", 3879 i, 3880 ph->needs_swap); 3881 return 0; 3882 } 3883 /* could not determine endianness */ 3884 return -1; 3885 } 3886 3887 #define PERF_PIPE_HDR_VER0 16 3888 3889 static const size_t attr_pipe_abi_sizes[] = { 3890 [0] = PERF_PIPE_HDR_VER0, 3891 0, 3892 }; 3893 3894 /* 3895 * In the legacy pipe format, there is an implicit assumption that endianness 3896 * between host recording the samples, and host parsing the samples is the 3897 * same. This is not always the case given that the pipe output may always be 3898 * redirected into a file and analyzed on a different machine with possibly a 3899 * different endianness and perf_event ABI revisions in the perf tool itself. 3900 */ 3901 static int try_all_pipe_abis(uint64_t hdr_sz, struct perf_header *ph) 3902 { 3903 u64 attr_size; 3904 int i; 3905 3906 for (i = 0 ; attr_pipe_abi_sizes[i]; i++) { 3907 if (hdr_sz != attr_pipe_abi_sizes[i]) { 3908 attr_size = bswap_64(hdr_sz); 3909 if (attr_size != hdr_sz) 3910 continue; 3911 3912 ph->needs_swap = true; 3913 } 3914 pr_debug("Pipe ABI%d perf.data file detected\n", i); 3915 return 0; 3916 } 3917 return -1; 3918 } 3919 3920 bool is_perf_magic(u64 magic) 3921 { 3922 if (!memcmp(&magic, __perf_magic1, sizeof(magic)) 3923 || magic == __perf_magic2 3924 || magic == __perf_magic2_sw) 3925 return true; 3926 3927 return false; 3928 } 3929 3930 static int check_magic_endian(u64 magic, uint64_t hdr_sz, 3931 bool is_pipe, struct perf_header *ph) 3932 { 3933 int ret; 3934 3935 /* check for legacy format */ 3936 ret = memcmp(&magic, __perf_magic1, sizeof(magic)); 3937 if (ret == 0) { 3938 ph->version = PERF_HEADER_VERSION_1; 3939 pr_debug("legacy perf.data format\n"); 3940 if (is_pipe) 3941 return try_all_pipe_abis(hdr_sz, ph); 3942 3943 return try_all_file_abis(hdr_sz, ph); 3944 } 3945 /* 3946 * the new magic number serves two purposes: 3947 * - unique number to identify actual perf.data files 3948 * - encode endianness of file 3949 */ 3950 ph->version = PERF_HEADER_VERSION_2; 3951 3952 /* check magic number with one endianness */ 3953 if (magic == __perf_magic2) 3954 return 0; 3955 3956 /* check magic number with opposite endianness */ 3957 if (magic != __perf_magic2_sw) 3958 return -1; 3959 3960 ph->needs_swap = true; 3961 3962 return 0; 3963 } 3964 3965 int perf_file_header__read(struct perf_file_header *header, 3966 struct perf_header *ph, int fd) 3967 { 3968 ssize_t ret; 3969 3970 lseek(fd, 0, SEEK_SET); 3971 3972 ret = readn(fd, header, sizeof(*header)); 3973 if (ret <= 0) 3974 return -1; 3975 3976 if (check_magic_endian(header->magic, 3977 header->attr_size, false, ph) < 0) { 3978 pr_debug("magic/endian check failed\n"); 3979 return -1; 3980 } 3981 3982 if (ph->needs_swap) { 3983 mem_bswap_64(header, offsetof(struct perf_file_header, 3984 adds_features)); 3985 } 3986 3987 if (header->size != sizeof(*header)) { 3988 /* Support the previous format */ 3989 if (header->size == offsetof(typeof(*header), adds_features)) 3990 bitmap_zero(header->adds_features, HEADER_FEAT_BITS); 3991 else 3992 return -1; 3993 } else if (ph->needs_swap) { 3994 /* 3995 * feature bitmap is declared as an array of unsigned longs -- 3996 * not good since its size can differ between the host that 3997 * generated the data file and the host analyzing the file. 3998 * 3999 * We need to handle endianness, but we don't know the size of 4000 * the unsigned long where the file was generated. Take a best 4001 * guess at determining it: try 64-bit swap first (ie., file 4002 * created on a 64-bit host), and check if the hostname feature 4003 * bit is set (this feature bit is forced on as of fbe96f2). 4004 * If the bit is not, undo the 64-bit swap and try a 32-bit 4005 * swap. If the hostname bit is still not set (e.g., older data 4006 * file), punt and fallback to the original behavior -- 4007 * clearing all feature bits and setting buildid. 4008 */ 4009 mem_bswap_64(&header->adds_features, 4010 BITS_TO_U64(HEADER_FEAT_BITS)); 4011 4012 if (!test_bit(HEADER_HOSTNAME, header->adds_features)) { 4013 /* unswap as u64 */ 4014 mem_bswap_64(&header->adds_features, 4015 BITS_TO_U64(HEADER_FEAT_BITS)); 4016 4017 /* unswap as u32 */ 4018 mem_bswap_32(&header->adds_features, 4019 BITS_TO_U32(HEADER_FEAT_BITS)); 4020 } 4021 4022 if (!test_bit(HEADER_HOSTNAME, header->adds_features)) { 4023 bitmap_zero(header->adds_features, HEADER_FEAT_BITS); 4024 __set_bit(HEADER_BUILD_ID, header->adds_features); 4025 } 4026 } 4027 4028 memcpy(&ph->adds_features, &header->adds_features, 4029 sizeof(ph->adds_features)); 4030 4031 ph->data_offset = header->data.offset; 4032 ph->data_size = header->data.size; 4033 ph->feat_offset = header->data.offset + header->data.size; 4034 return 0; 4035 } 4036 4037 static int perf_file_section__process(struct perf_file_section *section, 4038 struct perf_header *ph, 4039 int feat, int fd, void *data) 4040 { 4041 struct feat_fd fdd = { 4042 .fd = fd, 4043 .ph = ph, 4044 .size = section->size, 4045 .offset = section->offset, 4046 }; 4047 4048 if (lseek(fd, section->offset, SEEK_SET) == (off_t)-1) { 4049 pr_debug("Failed to lseek to %" PRIu64 " offset for feature " 4050 "%d, continuing...\n", section->offset, feat); 4051 return 0; 4052 } 4053 4054 if (feat >= HEADER_LAST_FEATURE) { 4055 pr_debug("unknown feature %d, continuing...\n", feat); 4056 return 0; 4057 } 4058 4059 if (!feat_ops[feat].process) 4060 return 0; 4061 4062 return feat_ops[feat].process(&fdd, data); 4063 } 4064 4065 static int perf_file_header__read_pipe(struct perf_pipe_file_header *header, 4066 struct perf_header *ph, 4067 struct perf_data* data, 4068 bool repipe, int repipe_fd) 4069 { 4070 struct feat_fd ff = { 4071 .fd = repipe_fd, 4072 .ph = ph, 4073 }; 4074 ssize_t ret; 4075 4076 ret = perf_data__read(data, header, sizeof(*header)); 4077 if (ret <= 0) 4078 return -1; 4079 4080 if (check_magic_endian(header->magic, header->size, true, ph) < 0) { 4081 pr_debug("endian/magic failed\n"); 4082 return -1; 4083 } 4084 4085 if (ph->needs_swap) 4086 header->size = bswap_64(header->size); 4087 4088 if (repipe && do_write(&ff, header, sizeof(*header)) < 0) 4089 return -1; 4090 4091 return 0; 4092 } 4093 4094 static int perf_header__read_pipe(struct perf_session *session, int repipe_fd) 4095 { 4096 struct perf_header *header = &session->header; 4097 struct perf_pipe_file_header f_header; 4098 4099 if (perf_file_header__read_pipe(&f_header, header, session->data, 4100 session->repipe, repipe_fd) < 0) { 4101 pr_debug("incompatible file format\n"); 4102 return -EINVAL; 4103 } 4104 4105 return f_header.size == sizeof(f_header) ? 0 : -1; 4106 } 4107 4108 static int read_attr(int fd, struct perf_header *ph, 4109 struct perf_file_attr *f_attr) 4110 { 4111 struct perf_event_attr *attr = &f_attr->attr; 4112 size_t sz, left; 4113 size_t our_sz = sizeof(f_attr->attr); 4114 ssize_t ret; 4115 4116 memset(f_attr, 0, sizeof(*f_attr)); 4117 4118 /* read minimal guaranteed structure */ 4119 ret = readn(fd, attr, PERF_ATTR_SIZE_VER0); 4120 if (ret <= 0) { 4121 pr_debug("cannot read %d bytes of header attr\n", 4122 PERF_ATTR_SIZE_VER0); 4123 return -1; 4124 } 4125 4126 /* on file perf_event_attr size */ 4127 sz = attr->size; 4128 4129 if (ph->needs_swap) 4130 sz = bswap_32(sz); 4131 4132 if (sz == 0) { 4133 /* assume ABI0 */ 4134 sz = PERF_ATTR_SIZE_VER0; 4135 } else if (sz > our_sz) { 4136 pr_debug("file uses a more recent and unsupported ABI" 4137 " (%zu bytes extra)\n", sz - our_sz); 4138 return -1; 4139 } 4140 /* what we have not yet read and that we know about */ 4141 left = sz - PERF_ATTR_SIZE_VER0; 4142 if (left) { 4143 void *ptr = attr; 4144 ptr += PERF_ATTR_SIZE_VER0; 4145 4146 ret = readn(fd, ptr, left); 4147 } 4148 /* read perf_file_section, ids are read in caller */ 4149 ret = readn(fd, &f_attr->ids, sizeof(f_attr->ids)); 4150 4151 return ret <= 0 ? -1 : 0; 4152 } 4153 4154 #ifdef HAVE_LIBTRACEEVENT 4155 static int evsel__prepare_tracepoint_event(struct evsel *evsel, struct tep_handle *pevent) 4156 { 4157 struct tep_event *event; 4158 char bf[128]; 4159 4160 /* already prepared */ 4161 if (evsel->tp_format) 4162 return 0; 4163 4164 if (pevent == NULL) { 4165 pr_debug("broken or missing trace data\n"); 4166 return -1; 4167 } 4168 4169 event = tep_find_event(pevent, evsel->core.attr.config); 4170 if (event == NULL) { 4171 pr_debug("cannot find event format for %d\n", (int)evsel->core.attr.config); 4172 return -1; 4173 } 4174 4175 if (!evsel->name) { 4176 snprintf(bf, sizeof(bf), "%s:%s", event->system, event->name); 4177 evsel->name = strdup(bf); 4178 if (evsel->name == NULL) 4179 return -1; 4180 } 4181 4182 evsel->tp_format = event; 4183 return 0; 4184 } 4185 4186 static int evlist__prepare_tracepoint_events(struct evlist *evlist, struct tep_handle *pevent) 4187 { 4188 struct evsel *pos; 4189 4190 evlist__for_each_entry(evlist, pos) { 4191 if (pos->core.attr.type == PERF_TYPE_TRACEPOINT && 4192 evsel__prepare_tracepoint_event(pos, pevent)) 4193 return -1; 4194 } 4195 4196 return 0; 4197 } 4198 #endif 4199 4200 int perf_session__read_header(struct perf_session *session, int repipe_fd) 4201 { 4202 struct perf_data *data = session->data; 4203 struct perf_header *header = &session->header; 4204 struct perf_file_header f_header; 4205 struct perf_file_attr f_attr; 4206 u64 f_id; 4207 int nr_attrs, nr_ids, i, j, err; 4208 int fd = perf_data__fd(data); 4209 4210 session->evlist = evlist__new(); 4211 if (session->evlist == NULL) 4212 return -ENOMEM; 4213 4214 session->evlist->env = &header->env; 4215 session->machines.host.env = &header->env; 4216 4217 /* 4218 * We can read 'pipe' data event from regular file, 4219 * check for the pipe header regardless of source. 4220 */ 4221 err = perf_header__read_pipe(session, repipe_fd); 4222 if (!err || perf_data__is_pipe(data)) { 4223 data->is_pipe = true; 4224 return err; 4225 } 4226 4227 if (perf_file_header__read(&f_header, header, fd) < 0) 4228 return -EINVAL; 4229 4230 if (header->needs_swap && data->in_place_update) { 4231 pr_err("In-place update not supported when byte-swapping is required\n"); 4232 return -EINVAL; 4233 } 4234 4235 /* 4236 * Sanity check that perf.data was written cleanly; data size is 4237 * initialized to 0 and updated only if the on_exit function is run. 4238 * If data size is still 0 then the file contains only partial 4239 * information. Just warn user and process it as much as it can. 4240 */ 4241 if (f_header.data.size == 0) { 4242 pr_warning("WARNING: The %s file's data size field is 0 which is unexpected.\n" 4243 "Was the 'perf record' command properly terminated?\n", 4244 data->file.path); 4245 } 4246 4247 if (f_header.attr_size == 0) { 4248 pr_err("ERROR: The %s file's attr size field is 0 which is unexpected.\n" 4249 "Was the 'perf record' command properly terminated?\n", 4250 data->file.path); 4251 return -EINVAL; 4252 } 4253 4254 nr_attrs = f_header.attrs.size / f_header.attr_size; 4255 lseek(fd, f_header.attrs.offset, SEEK_SET); 4256 4257 for (i = 0; i < nr_attrs; i++) { 4258 struct evsel *evsel; 4259 off_t tmp; 4260 4261 if (read_attr(fd, header, &f_attr) < 0) 4262 goto out_errno; 4263 4264 if (header->needs_swap) { 4265 f_attr.ids.size = bswap_64(f_attr.ids.size); 4266 f_attr.ids.offset = bswap_64(f_attr.ids.offset); 4267 perf_event__attr_swap(&f_attr.attr); 4268 } 4269 4270 tmp = lseek(fd, 0, SEEK_CUR); 4271 evsel = evsel__new(&f_attr.attr); 4272 4273 if (evsel == NULL) 4274 goto out_delete_evlist; 4275 4276 evsel->needs_swap = header->needs_swap; 4277 /* 4278 * Do it before so that if perf_evsel__alloc_id fails, this 4279 * entry gets purged too at evlist__delete(). 4280 */ 4281 evlist__add(session->evlist, evsel); 4282 4283 nr_ids = f_attr.ids.size / sizeof(u64); 4284 /* 4285 * We don't have the cpu and thread maps on the header, so 4286 * for allocating the perf_sample_id table we fake 1 cpu and 4287 * hattr->ids threads. 4288 */ 4289 if (perf_evsel__alloc_id(&evsel->core, 1, nr_ids)) 4290 goto out_delete_evlist; 4291 4292 lseek(fd, f_attr.ids.offset, SEEK_SET); 4293 4294 for (j = 0; j < nr_ids; j++) { 4295 if (perf_header__getbuffer64(header, fd, &f_id, sizeof(f_id))) 4296 goto out_errno; 4297 4298 perf_evlist__id_add(&session->evlist->core, &evsel->core, 0, j, f_id); 4299 } 4300 4301 lseek(fd, tmp, SEEK_SET); 4302 } 4303 4304 #ifdef HAVE_LIBTRACEEVENT 4305 perf_header__process_sections(header, fd, &session->tevent, 4306 perf_file_section__process); 4307 4308 if (evlist__prepare_tracepoint_events(session->evlist, session->tevent.pevent)) 4309 goto out_delete_evlist; 4310 #else 4311 perf_header__process_sections(header, fd, NULL, perf_file_section__process); 4312 #endif 4313 4314 return 0; 4315 out_errno: 4316 return -errno; 4317 4318 out_delete_evlist: 4319 evlist__delete(session->evlist); 4320 session->evlist = NULL; 4321 return -ENOMEM; 4322 } 4323 4324 int perf_event__process_feature(struct perf_session *session, 4325 union perf_event *event) 4326 { 4327 struct perf_tool *tool = session->tool; 4328 struct feat_fd ff = { .fd = 0 }; 4329 struct perf_record_header_feature *fe = (struct perf_record_header_feature *)event; 4330 int type = fe->header.type; 4331 u64 feat = fe->feat_id; 4332 int ret = 0; 4333 4334 if (type < 0 || type >= PERF_RECORD_HEADER_MAX) { 4335 pr_warning("invalid record type %d in pipe-mode\n", type); 4336 return 0; 4337 } 4338 if (feat == HEADER_RESERVED || feat >= HEADER_LAST_FEATURE) { 4339 pr_warning("invalid record type %d in pipe-mode\n", type); 4340 return -1; 4341 } 4342 4343 if (!feat_ops[feat].process) 4344 return 0; 4345 4346 ff.buf = (void *)fe->data; 4347 ff.size = event->header.size - sizeof(*fe); 4348 ff.ph = &session->header; 4349 4350 if (feat_ops[feat].process(&ff, NULL)) { 4351 ret = -1; 4352 goto out; 4353 } 4354 4355 if (!feat_ops[feat].print || !tool->show_feat_hdr) 4356 goto out; 4357 4358 if (!feat_ops[feat].full_only || 4359 tool->show_feat_hdr >= SHOW_FEAT_HEADER_FULL_INFO) { 4360 feat_ops[feat].print(&ff, stdout); 4361 } else { 4362 fprintf(stdout, "# %s info available, use -I to display\n", 4363 feat_ops[feat].name); 4364 } 4365 out: 4366 free_event_desc(ff.events); 4367 return ret; 4368 } 4369 4370 size_t perf_event__fprintf_event_update(union perf_event *event, FILE *fp) 4371 { 4372 struct perf_record_event_update *ev = &event->event_update; 4373 struct perf_cpu_map *map; 4374 size_t ret; 4375 4376 ret = fprintf(fp, "\n... id: %" PRI_lu64 "\n", ev->id); 4377 4378 switch (ev->type) { 4379 case PERF_EVENT_UPDATE__SCALE: 4380 ret += fprintf(fp, "... scale: %f\n", ev->scale.scale); 4381 break; 4382 case PERF_EVENT_UPDATE__UNIT: 4383 ret += fprintf(fp, "... unit: %s\n", ev->unit); 4384 break; 4385 case PERF_EVENT_UPDATE__NAME: 4386 ret += fprintf(fp, "... name: %s\n", ev->name); 4387 break; 4388 case PERF_EVENT_UPDATE__CPUS: 4389 ret += fprintf(fp, "... "); 4390 4391 map = cpu_map__new_data(&ev->cpus.cpus); 4392 if (map) 4393 ret += cpu_map__fprintf(map, fp); 4394 else 4395 ret += fprintf(fp, "failed to get cpus\n"); 4396 break; 4397 default: 4398 ret += fprintf(fp, "... unknown type\n"); 4399 break; 4400 } 4401 4402 return ret; 4403 } 4404 4405 int perf_event__process_attr(struct perf_tool *tool __maybe_unused, 4406 union perf_event *event, 4407 struct evlist **pevlist) 4408 { 4409 u32 i, n_ids; 4410 u64 *ids; 4411 struct evsel *evsel; 4412 struct evlist *evlist = *pevlist; 4413 4414 if (evlist == NULL) { 4415 *pevlist = evlist = evlist__new(); 4416 if (evlist == NULL) 4417 return -ENOMEM; 4418 } 4419 4420 evsel = evsel__new(&event->attr.attr); 4421 if (evsel == NULL) 4422 return -ENOMEM; 4423 4424 evlist__add(evlist, evsel); 4425 4426 n_ids = event->header.size - sizeof(event->header) - event->attr.attr.size; 4427 n_ids = n_ids / sizeof(u64); 4428 /* 4429 * We don't have the cpu and thread maps on the header, so 4430 * for allocating the perf_sample_id table we fake 1 cpu and 4431 * hattr->ids threads. 4432 */ 4433 if (perf_evsel__alloc_id(&evsel->core, 1, n_ids)) 4434 return -ENOMEM; 4435 4436 ids = perf_record_header_attr_id(event); 4437 for (i = 0; i < n_ids; i++) { 4438 perf_evlist__id_add(&evlist->core, &evsel->core, 0, i, ids[i]); 4439 } 4440 4441 return 0; 4442 } 4443 4444 int perf_event__process_event_update(struct perf_tool *tool __maybe_unused, 4445 union perf_event *event, 4446 struct evlist **pevlist) 4447 { 4448 struct perf_record_event_update *ev = &event->event_update; 4449 struct evlist *evlist; 4450 struct evsel *evsel; 4451 struct perf_cpu_map *map; 4452 4453 if (dump_trace) 4454 perf_event__fprintf_event_update(event, stdout); 4455 4456 if (!pevlist || *pevlist == NULL) 4457 return -EINVAL; 4458 4459 evlist = *pevlist; 4460 4461 evsel = evlist__id2evsel(evlist, ev->id); 4462 if (evsel == NULL) 4463 return -EINVAL; 4464 4465 switch (ev->type) { 4466 case PERF_EVENT_UPDATE__UNIT: 4467 free((char *)evsel->unit); 4468 evsel->unit = strdup(ev->unit); 4469 break; 4470 case PERF_EVENT_UPDATE__NAME: 4471 free(evsel->name); 4472 evsel->name = strdup(ev->name); 4473 break; 4474 case PERF_EVENT_UPDATE__SCALE: 4475 evsel->scale = ev->scale.scale; 4476 break; 4477 case PERF_EVENT_UPDATE__CPUS: 4478 map = cpu_map__new_data(&ev->cpus.cpus); 4479 if (map) { 4480 perf_cpu_map__put(evsel->core.own_cpus); 4481 evsel->core.own_cpus = map; 4482 } else 4483 pr_err("failed to get event_update cpus\n"); 4484 default: 4485 break; 4486 } 4487 4488 return 0; 4489 } 4490 4491 #ifdef HAVE_LIBTRACEEVENT 4492 int perf_event__process_tracing_data(struct perf_session *session, 4493 union perf_event *event) 4494 { 4495 ssize_t size_read, padding, size = event->tracing_data.size; 4496 int fd = perf_data__fd(session->data); 4497 char buf[BUFSIZ]; 4498 4499 /* 4500 * The pipe fd is already in proper place and in any case 4501 * we can't move it, and we'd screw the case where we read 4502 * 'pipe' data from regular file. The trace_report reads 4503 * data from 'fd' so we need to set it directly behind the 4504 * event, where the tracing data starts. 4505 */ 4506 if (!perf_data__is_pipe(session->data)) { 4507 off_t offset = lseek(fd, 0, SEEK_CUR); 4508 4509 /* setup for reading amidst mmap */ 4510 lseek(fd, offset + sizeof(struct perf_record_header_tracing_data), 4511 SEEK_SET); 4512 } 4513 4514 size_read = trace_report(fd, &session->tevent, 4515 session->repipe); 4516 padding = PERF_ALIGN(size_read, sizeof(u64)) - size_read; 4517 4518 if (readn(fd, buf, padding) < 0) { 4519 pr_err("%s: reading input file", __func__); 4520 return -1; 4521 } 4522 if (session->repipe) { 4523 int retw = write(STDOUT_FILENO, buf, padding); 4524 if (retw <= 0 || retw != padding) { 4525 pr_err("%s: repiping tracing data padding", __func__); 4526 return -1; 4527 } 4528 } 4529 4530 if (size_read + padding != size) { 4531 pr_err("%s: tracing data size mismatch", __func__); 4532 return -1; 4533 } 4534 4535 evlist__prepare_tracepoint_events(session->evlist, session->tevent.pevent); 4536 4537 return size_read + padding; 4538 } 4539 #endif 4540 4541 int perf_event__process_build_id(struct perf_session *session, 4542 union perf_event *event) 4543 { 4544 __event_process_build_id(&event->build_id, 4545 event->build_id.filename, 4546 session); 4547 return 0; 4548 } 4549