1 #define _FILE_OFFSET_BITS 64 2 3 #include "util.h" 4 #include <sys/types.h> 5 #include <byteswap.h> 6 #include <unistd.h> 7 #include <stdio.h> 8 #include <stdlib.h> 9 #include <linux/list.h> 10 #include <linux/kernel.h> 11 #include <linux/bitops.h> 12 #include <sys/utsname.h> 13 14 #include "evlist.h" 15 #include "evsel.h" 16 #include "header.h" 17 #include "../perf.h" 18 #include "trace-event.h" 19 #include "session.h" 20 #include "symbol.h" 21 #include "debug.h" 22 #include "cpumap.h" 23 #include "pmu.h" 24 #include "vdso.h" 25 #include "strbuf.h" 26 27 static bool no_buildid_cache = false; 28 29 static int trace_event_count; 30 static struct perf_trace_event_type *trace_events; 31 32 static u32 header_argc; 33 static const char **header_argv; 34 35 int perf_header__push_event(u64 id, const char *name) 36 { 37 struct perf_trace_event_type *nevents; 38 39 if (strlen(name) > MAX_EVENT_NAME) 40 pr_warning("Event %s will be truncated\n", name); 41 42 nevents = realloc(trace_events, (trace_event_count + 1) * sizeof(*trace_events)); 43 if (nevents == NULL) 44 return -ENOMEM; 45 trace_events = nevents; 46 47 memset(&trace_events[trace_event_count], 0, sizeof(struct perf_trace_event_type)); 48 trace_events[trace_event_count].event_id = id; 49 strncpy(trace_events[trace_event_count].name, name, MAX_EVENT_NAME - 1); 50 trace_event_count++; 51 return 0; 52 } 53 54 char *perf_header__find_event(u64 id) 55 { 56 int i; 57 for (i = 0 ; i < trace_event_count; i++) { 58 if (trace_events[i].event_id == id) 59 return trace_events[i].name; 60 } 61 return NULL; 62 } 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 struct perf_file_attr { 80 struct perf_event_attr attr; 81 struct perf_file_section ids; 82 }; 83 84 void perf_header__set_feat(struct perf_header *header, int feat) 85 { 86 set_bit(feat, header->adds_features); 87 } 88 89 void perf_header__clear_feat(struct perf_header *header, int feat) 90 { 91 clear_bit(feat, header->adds_features); 92 } 93 94 bool perf_header__has_feat(const struct perf_header *header, int feat) 95 { 96 return test_bit(feat, header->adds_features); 97 } 98 99 static int do_write(int fd, const void *buf, size_t size) 100 { 101 while (size) { 102 int ret = write(fd, buf, size); 103 104 if (ret < 0) 105 return -errno; 106 107 size -= ret; 108 buf += ret; 109 } 110 111 return 0; 112 } 113 114 #define NAME_ALIGN 64 115 116 static int write_padded(int fd, const void *bf, size_t count, 117 size_t count_aligned) 118 { 119 static const char zero_buf[NAME_ALIGN]; 120 int err = do_write(fd, bf, count); 121 122 if (!err) 123 err = do_write(fd, zero_buf, count_aligned - count); 124 125 return err; 126 } 127 128 static int do_write_string(int fd, const char *str) 129 { 130 u32 len, olen; 131 int ret; 132 133 olen = strlen(str) + 1; 134 len = PERF_ALIGN(olen, NAME_ALIGN); 135 136 /* write len, incl. \0 */ 137 ret = do_write(fd, &len, sizeof(len)); 138 if (ret < 0) 139 return ret; 140 141 return write_padded(fd, str, olen, len); 142 } 143 144 static char *do_read_string(int fd, struct perf_header *ph) 145 { 146 ssize_t sz, ret; 147 u32 len; 148 char *buf; 149 150 sz = read(fd, &len, sizeof(len)); 151 if (sz < (ssize_t)sizeof(len)) 152 return NULL; 153 154 if (ph->needs_swap) 155 len = bswap_32(len); 156 157 buf = malloc(len); 158 if (!buf) 159 return NULL; 160 161 ret = read(fd, buf, len); 162 if (ret == (ssize_t)len) { 163 /* 164 * strings are padded by zeroes 165 * thus the actual strlen of buf 166 * may be less than len 167 */ 168 return buf; 169 } 170 171 free(buf); 172 return NULL; 173 } 174 175 int 176 perf_header__set_cmdline(int argc, const char **argv) 177 { 178 int i; 179 180 /* 181 * If header_argv has already been set, do not override it. 182 * This allows a command to set the cmdline, parse args and 183 * then call another builtin function that implements a 184 * command -- e.g, cmd_kvm calling cmd_record. 185 */ 186 if (header_argv) 187 return 0; 188 189 header_argc = (u32)argc; 190 191 /* do not include NULL termination */ 192 header_argv = calloc(argc, sizeof(char *)); 193 if (!header_argv) 194 return -ENOMEM; 195 196 /* 197 * must copy argv contents because it gets moved 198 * around during option parsing 199 */ 200 for (i = 0; i < argc ; i++) 201 header_argv[i] = argv[i]; 202 203 return 0; 204 } 205 206 #define dsos__for_each_with_build_id(pos, head) \ 207 list_for_each_entry(pos, head, node) \ 208 if (!pos->has_build_id) \ 209 continue; \ 210 else 211 212 static int write_buildid(char *name, size_t name_len, u8 *build_id, 213 pid_t pid, u16 misc, int fd) 214 { 215 int err; 216 struct build_id_event b; 217 size_t len; 218 219 len = name_len + 1; 220 len = PERF_ALIGN(len, NAME_ALIGN); 221 222 memset(&b, 0, sizeof(b)); 223 memcpy(&b.build_id, build_id, BUILD_ID_SIZE); 224 b.pid = pid; 225 b.header.misc = misc; 226 b.header.size = sizeof(b) + len; 227 228 err = do_write(fd, &b, sizeof(b)); 229 if (err < 0) 230 return err; 231 232 return write_padded(fd, name, name_len + 1, len); 233 } 234 235 static int __dsos__write_buildid_table(struct list_head *head, pid_t pid, 236 u16 misc, int fd) 237 { 238 struct dso *pos; 239 240 dsos__for_each_with_build_id(pos, head) { 241 int err; 242 char *name; 243 size_t name_len; 244 245 if (!pos->hit) 246 continue; 247 248 if (is_vdso_map(pos->short_name)) { 249 name = (char *) VDSO__MAP_NAME; 250 name_len = sizeof(VDSO__MAP_NAME) + 1; 251 } else { 252 name = pos->long_name; 253 name_len = pos->long_name_len + 1; 254 } 255 256 err = write_buildid(name, name_len, pos->build_id, 257 pid, misc, fd); 258 if (err) 259 return err; 260 } 261 262 return 0; 263 } 264 265 static int machine__write_buildid_table(struct machine *machine, int fd) 266 { 267 int err; 268 u16 kmisc = PERF_RECORD_MISC_KERNEL, 269 umisc = PERF_RECORD_MISC_USER; 270 271 if (!machine__is_host(machine)) { 272 kmisc = PERF_RECORD_MISC_GUEST_KERNEL; 273 umisc = PERF_RECORD_MISC_GUEST_USER; 274 } 275 276 err = __dsos__write_buildid_table(&machine->kernel_dsos, machine->pid, 277 kmisc, fd); 278 if (err == 0) 279 err = __dsos__write_buildid_table(&machine->user_dsos, 280 machine->pid, umisc, fd); 281 return err; 282 } 283 284 static int dsos__write_buildid_table(struct perf_header *header, int fd) 285 { 286 struct perf_session *session = container_of(header, 287 struct perf_session, header); 288 struct rb_node *nd; 289 int err = machine__write_buildid_table(&session->host_machine, fd); 290 291 if (err) 292 return err; 293 294 for (nd = rb_first(&session->machines); nd; nd = rb_next(nd)) { 295 struct machine *pos = rb_entry(nd, struct machine, rb_node); 296 err = machine__write_buildid_table(pos, fd); 297 if (err) 298 break; 299 } 300 return err; 301 } 302 303 int build_id_cache__add_s(const char *sbuild_id, const char *debugdir, 304 const char *name, bool is_kallsyms, bool is_vdso) 305 { 306 const size_t size = PATH_MAX; 307 char *realname, *filename = zalloc(size), 308 *linkname = zalloc(size), *targetname; 309 int len, err = -1; 310 bool slash = is_kallsyms || is_vdso; 311 312 if (is_kallsyms) { 313 if (symbol_conf.kptr_restrict) { 314 pr_debug("Not caching a kptr_restrict'ed /proc/kallsyms\n"); 315 return 0; 316 } 317 realname = (char *) name; 318 } else 319 realname = realpath(name, NULL); 320 321 if (realname == NULL || filename == NULL || linkname == NULL) 322 goto out_free; 323 324 len = scnprintf(filename, size, "%s%s%s", 325 debugdir, slash ? "/" : "", 326 is_vdso ? VDSO__MAP_NAME : realname); 327 if (mkdir_p(filename, 0755)) 328 goto out_free; 329 330 snprintf(filename + len, size - len, "/%s", sbuild_id); 331 332 if (access(filename, F_OK)) { 333 if (is_kallsyms) { 334 if (copyfile("/proc/kallsyms", filename)) 335 goto out_free; 336 } else if (link(realname, filename) && copyfile(name, filename)) 337 goto out_free; 338 } 339 340 len = scnprintf(linkname, size, "%s/.build-id/%.2s", 341 debugdir, sbuild_id); 342 343 if (access(linkname, X_OK) && mkdir_p(linkname, 0755)) 344 goto out_free; 345 346 snprintf(linkname + len, size - len, "/%s", sbuild_id + 2); 347 targetname = filename + strlen(debugdir) - 5; 348 memcpy(targetname, "../..", 5); 349 350 if (symlink(targetname, linkname) == 0) 351 err = 0; 352 out_free: 353 if (!is_kallsyms) 354 free(realname); 355 free(filename); 356 free(linkname); 357 return err; 358 } 359 360 static int build_id_cache__add_b(const u8 *build_id, size_t build_id_size, 361 const char *name, const char *debugdir, 362 bool is_kallsyms, bool is_vdso) 363 { 364 char sbuild_id[BUILD_ID_SIZE * 2 + 1]; 365 366 build_id__sprintf(build_id, build_id_size, sbuild_id); 367 368 return build_id_cache__add_s(sbuild_id, debugdir, name, 369 is_kallsyms, is_vdso); 370 } 371 372 int build_id_cache__remove_s(const char *sbuild_id, const char *debugdir) 373 { 374 const size_t size = PATH_MAX; 375 char *filename = zalloc(size), 376 *linkname = zalloc(size); 377 int err = -1; 378 379 if (filename == NULL || linkname == NULL) 380 goto out_free; 381 382 snprintf(linkname, size, "%s/.build-id/%.2s/%s", 383 debugdir, sbuild_id, sbuild_id + 2); 384 385 if (access(linkname, F_OK)) 386 goto out_free; 387 388 if (readlink(linkname, filename, size - 1) < 0) 389 goto out_free; 390 391 if (unlink(linkname)) 392 goto out_free; 393 394 /* 395 * Since the link is relative, we must make it absolute: 396 */ 397 snprintf(linkname, size, "%s/.build-id/%.2s/%s", 398 debugdir, sbuild_id, filename); 399 400 if (unlink(linkname)) 401 goto out_free; 402 403 err = 0; 404 out_free: 405 free(filename); 406 free(linkname); 407 return err; 408 } 409 410 static int dso__cache_build_id(struct dso *dso, const char *debugdir) 411 { 412 bool is_kallsyms = dso->kernel && dso->long_name[0] != '/'; 413 bool is_vdso = is_vdso_map(dso->short_name); 414 415 return build_id_cache__add_b(dso->build_id, sizeof(dso->build_id), 416 dso->long_name, debugdir, 417 is_kallsyms, is_vdso); 418 } 419 420 static int __dsos__cache_build_ids(struct list_head *head, const char *debugdir) 421 { 422 struct dso *pos; 423 int err = 0; 424 425 dsos__for_each_with_build_id(pos, head) 426 if (dso__cache_build_id(pos, debugdir)) 427 err = -1; 428 429 return err; 430 } 431 432 static int machine__cache_build_ids(struct machine *machine, const char *debugdir) 433 { 434 int ret = __dsos__cache_build_ids(&machine->kernel_dsos, debugdir); 435 ret |= __dsos__cache_build_ids(&machine->user_dsos, debugdir); 436 return ret; 437 } 438 439 static int perf_session__cache_build_ids(struct perf_session *session) 440 { 441 struct rb_node *nd; 442 int ret; 443 char debugdir[PATH_MAX]; 444 445 snprintf(debugdir, sizeof(debugdir), "%s", buildid_dir); 446 447 if (mkdir(debugdir, 0755) != 0 && errno != EEXIST) 448 return -1; 449 450 ret = machine__cache_build_ids(&session->host_machine, debugdir); 451 452 for (nd = rb_first(&session->machines); nd; nd = rb_next(nd)) { 453 struct machine *pos = rb_entry(nd, struct machine, rb_node); 454 ret |= machine__cache_build_ids(pos, debugdir); 455 } 456 return ret ? -1 : 0; 457 } 458 459 static bool machine__read_build_ids(struct machine *machine, bool with_hits) 460 { 461 bool ret = __dsos__read_build_ids(&machine->kernel_dsos, with_hits); 462 ret |= __dsos__read_build_ids(&machine->user_dsos, with_hits); 463 return ret; 464 } 465 466 static bool perf_session__read_build_ids(struct perf_session *session, bool with_hits) 467 { 468 struct rb_node *nd; 469 bool ret = machine__read_build_ids(&session->host_machine, with_hits); 470 471 for (nd = rb_first(&session->machines); nd; nd = rb_next(nd)) { 472 struct machine *pos = rb_entry(nd, struct machine, rb_node); 473 ret |= machine__read_build_ids(pos, with_hits); 474 } 475 476 return ret; 477 } 478 479 static int write_tracing_data(int fd, struct perf_header *h __maybe_unused, 480 struct perf_evlist *evlist) 481 { 482 return read_tracing_data(fd, &evlist->entries); 483 } 484 485 486 static int write_build_id(int fd, struct perf_header *h, 487 struct perf_evlist *evlist __maybe_unused) 488 { 489 struct perf_session *session; 490 int err; 491 492 session = container_of(h, struct perf_session, header); 493 494 if (!perf_session__read_build_ids(session, true)) 495 return -1; 496 497 err = dsos__write_buildid_table(h, fd); 498 if (err < 0) { 499 pr_debug("failed to write buildid table\n"); 500 return err; 501 } 502 if (!no_buildid_cache) 503 perf_session__cache_build_ids(session); 504 505 return 0; 506 } 507 508 static int write_hostname(int fd, struct perf_header *h __maybe_unused, 509 struct perf_evlist *evlist __maybe_unused) 510 { 511 struct utsname uts; 512 int ret; 513 514 ret = uname(&uts); 515 if (ret < 0) 516 return -1; 517 518 return do_write_string(fd, uts.nodename); 519 } 520 521 static int write_osrelease(int fd, struct perf_header *h __maybe_unused, 522 struct perf_evlist *evlist __maybe_unused) 523 { 524 struct utsname uts; 525 int ret; 526 527 ret = uname(&uts); 528 if (ret < 0) 529 return -1; 530 531 return do_write_string(fd, uts.release); 532 } 533 534 static int write_arch(int fd, struct perf_header *h __maybe_unused, 535 struct perf_evlist *evlist __maybe_unused) 536 { 537 struct utsname uts; 538 int ret; 539 540 ret = uname(&uts); 541 if (ret < 0) 542 return -1; 543 544 return do_write_string(fd, uts.machine); 545 } 546 547 static int write_version(int fd, struct perf_header *h __maybe_unused, 548 struct perf_evlist *evlist __maybe_unused) 549 { 550 return do_write_string(fd, perf_version_string); 551 } 552 553 static int write_cpudesc(int fd, struct perf_header *h __maybe_unused, 554 struct perf_evlist *evlist __maybe_unused) 555 { 556 #ifndef CPUINFO_PROC 557 #define CPUINFO_PROC NULL 558 #endif 559 FILE *file; 560 char *buf = NULL; 561 char *s, *p; 562 const char *search = CPUINFO_PROC; 563 size_t len = 0; 564 int ret = -1; 565 566 if (!search) 567 return -1; 568 569 file = fopen("/proc/cpuinfo", "r"); 570 if (!file) 571 return -1; 572 573 while (getline(&buf, &len, file) > 0) { 574 ret = strncmp(buf, search, strlen(search)); 575 if (!ret) 576 break; 577 } 578 579 if (ret) 580 goto done; 581 582 s = buf; 583 584 p = strchr(buf, ':'); 585 if (p && *(p+1) == ' ' && *(p+2)) 586 s = p + 2; 587 p = strchr(s, '\n'); 588 if (p) 589 *p = '\0'; 590 591 /* squash extra space characters (branding string) */ 592 p = s; 593 while (*p) { 594 if (isspace(*p)) { 595 char *r = p + 1; 596 char *q = r; 597 *p = ' '; 598 while (*q && isspace(*q)) 599 q++; 600 if (q != (p+1)) 601 while ((*r++ = *q++)); 602 } 603 p++; 604 } 605 ret = do_write_string(fd, s); 606 done: 607 free(buf); 608 fclose(file); 609 return ret; 610 } 611 612 static int write_nrcpus(int fd, struct perf_header *h __maybe_unused, 613 struct perf_evlist *evlist __maybe_unused) 614 { 615 long nr; 616 u32 nrc, nra; 617 int ret; 618 619 nr = sysconf(_SC_NPROCESSORS_CONF); 620 if (nr < 0) 621 return -1; 622 623 nrc = (u32)(nr & UINT_MAX); 624 625 nr = sysconf(_SC_NPROCESSORS_ONLN); 626 if (nr < 0) 627 return -1; 628 629 nra = (u32)(nr & UINT_MAX); 630 631 ret = do_write(fd, &nrc, sizeof(nrc)); 632 if (ret < 0) 633 return ret; 634 635 return do_write(fd, &nra, sizeof(nra)); 636 } 637 638 static int write_event_desc(int fd, struct perf_header *h __maybe_unused, 639 struct perf_evlist *evlist) 640 { 641 struct perf_evsel *evsel; 642 u32 nre, nri, sz; 643 int ret; 644 645 nre = evlist->nr_entries; 646 647 /* 648 * write number of events 649 */ 650 ret = do_write(fd, &nre, sizeof(nre)); 651 if (ret < 0) 652 return ret; 653 654 /* 655 * size of perf_event_attr struct 656 */ 657 sz = (u32)sizeof(evsel->attr); 658 ret = do_write(fd, &sz, sizeof(sz)); 659 if (ret < 0) 660 return ret; 661 662 list_for_each_entry(evsel, &evlist->entries, node) { 663 664 ret = do_write(fd, &evsel->attr, sz); 665 if (ret < 0) 666 return ret; 667 /* 668 * write number of unique id per event 669 * there is one id per instance of an event 670 * 671 * copy into an nri to be independent of the 672 * type of ids, 673 */ 674 nri = evsel->ids; 675 ret = do_write(fd, &nri, sizeof(nri)); 676 if (ret < 0) 677 return ret; 678 679 /* 680 * write event string as passed on cmdline 681 */ 682 ret = do_write_string(fd, perf_evsel__name(evsel)); 683 if (ret < 0) 684 return ret; 685 /* 686 * write unique ids for this event 687 */ 688 ret = do_write(fd, evsel->id, evsel->ids * sizeof(u64)); 689 if (ret < 0) 690 return ret; 691 } 692 return 0; 693 } 694 695 static int write_cmdline(int fd, struct perf_header *h __maybe_unused, 696 struct perf_evlist *evlist __maybe_unused) 697 { 698 char buf[MAXPATHLEN]; 699 char proc[32]; 700 u32 i, n; 701 int ret; 702 703 /* 704 * actual atual path to perf binary 705 */ 706 sprintf(proc, "/proc/%d/exe", getpid()); 707 ret = readlink(proc, buf, sizeof(buf)); 708 if (ret <= 0) 709 return -1; 710 711 /* readlink() does not add null termination */ 712 buf[ret] = '\0'; 713 714 /* account for binary path */ 715 n = header_argc + 1; 716 717 ret = do_write(fd, &n, sizeof(n)); 718 if (ret < 0) 719 return ret; 720 721 ret = do_write_string(fd, buf); 722 if (ret < 0) 723 return ret; 724 725 for (i = 0 ; i < header_argc; i++) { 726 ret = do_write_string(fd, header_argv[i]); 727 if (ret < 0) 728 return ret; 729 } 730 return 0; 731 } 732 733 #define CORE_SIB_FMT \ 734 "/sys/devices/system/cpu/cpu%d/topology/core_siblings_list" 735 #define THRD_SIB_FMT \ 736 "/sys/devices/system/cpu/cpu%d/topology/thread_siblings_list" 737 738 struct cpu_topo { 739 u32 core_sib; 740 u32 thread_sib; 741 char **core_siblings; 742 char **thread_siblings; 743 }; 744 745 static int build_cpu_topo(struct cpu_topo *tp, int cpu) 746 { 747 FILE *fp; 748 char filename[MAXPATHLEN]; 749 char *buf = NULL, *p; 750 size_t len = 0; 751 u32 i = 0; 752 int ret = -1; 753 754 sprintf(filename, CORE_SIB_FMT, cpu); 755 fp = fopen(filename, "r"); 756 if (!fp) 757 return -1; 758 759 if (getline(&buf, &len, fp) <= 0) 760 goto done; 761 762 fclose(fp); 763 764 p = strchr(buf, '\n'); 765 if (p) 766 *p = '\0'; 767 768 for (i = 0; i < tp->core_sib; i++) { 769 if (!strcmp(buf, tp->core_siblings[i])) 770 break; 771 } 772 if (i == tp->core_sib) { 773 tp->core_siblings[i] = buf; 774 tp->core_sib++; 775 buf = NULL; 776 len = 0; 777 } 778 779 sprintf(filename, THRD_SIB_FMT, cpu); 780 fp = fopen(filename, "r"); 781 if (!fp) 782 goto done; 783 784 if (getline(&buf, &len, fp) <= 0) 785 goto done; 786 787 p = strchr(buf, '\n'); 788 if (p) 789 *p = '\0'; 790 791 for (i = 0; i < tp->thread_sib; i++) { 792 if (!strcmp(buf, tp->thread_siblings[i])) 793 break; 794 } 795 if (i == tp->thread_sib) { 796 tp->thread_siblings[i] = buf; 797 tp->thread_sib++; 798 buf = NULL; 799 } 800 ret = 0; 801 done: 802 if(fp) 803 fclose(fp); 804 free(buf); 805 return ret; 806 } 807 808 static void free_cpu_topo(struct cpu_topo *tp) 809 { 810 u32 i; 811 812 if (!tp) 813 return; 814 815 for (i = 0 ; i < tp->core_sib; i++) 816 free(tp->core_siblings[i]); 817 818 for (i = 0 ; i < tp->thread_sib; i++) 819 free(tp->thread_siblings[i]); 820 821 free(tp); 822 } 823 824 static struct cpu_topo *build_cpu_topology(void) 825 { 826 struct cpu_topo *tp; 827 void *addr; 828 u32 nr, i; 829 size_t sz; 830 long ncpus; 831 int ret = -1; 832 833 ncpus = sysconf(_SC_NPROCESSORS_CONF); 834 if (ncpus < 0) 835 return NULL; 836 837 nr = (u32)(ncpus & UINT_MAX); 838 839 sz = nr * sizeof(char *); 840 841 addr = calloc(1, sizeof(*tp) + 2 * sz); 842 if (!addr) 843 return NULL; 844 845 tp = addr; 846 847 addr += sizeof(*tp); 848 tp->core_siblings = addr; 849 addr += sz; 850 tp->thread_siblings = addr; 851 852 for (i = 0; i < nr; i++) { 853 ret = build_cpu_topo(tp, i); 854 if (ret < 0) 855 break; 856 } 857 if (ret) { 858 free_cpu_topo(tp); 859 tp = NULL; 860 } 861 return tp; 862 } 863 864 static int write_cpu_topology(int fd, struct perf_header *h __maybe_unused, 865 struct perf_evlist *evlist __maybe_unused) 866 { 867 struct cpu_topo *tp; 868 u32 i; 869 int ret; 870 871 tp = build_cpu_topology(); 872 if (!tp) 873 return -1; 874 875 ret = do_write(fd, &tp->core_sib, sizeof(tp->core_sib)); 876 if (ret < 0) 877 goto done; 878 879 for (i = 0; i < tp->core_sib; i++) { 880 ret = do_write_string(fd, tp->core_siblings[i]); 881 if (ret < 0) 882 goto done; 883 } 884 ret = do_write(fd, &tp->thread_sib, sizeof(tp->thread_sib)); 885 if (ret < 0) 886 goto done; 887 888 for (i = 0; i < tp->thread_sib; i++) { 889 ret = do_write_string(fd, tp->thread_siblings[i]); 890 if (ret < 0) 891 break; 892 } 893 done: 894 free_cpu_topo(tp); 895 return ret; 896 } 897 898 899 900 static int write_total_mem(int fd, struct perf_header *h __maybe_unused, 901 struct perf_evlist *evlist __maybe_unused) 902 { 903 char *buf = NULL; 904 FILE *fp; 905 size_t len = 0; 906 int ret = -1, n; 907 uint64_t mem; 908 909 fp = fopen("/proc/meminfo", "r"); 910 if (!fp) 911 return -1; 912 913 while (getline(&buf, &len, fp) > 0) { 914 ret = strncmp(buf, "MemTotal:", 9); 915 if (!ret) 916 break; 917 } 918 if (!ret) { 919 n = sscanf(buf, "%*s %"PRIu64, &mem); 920 if (n == 1) 921 ret = do_write(fd, &mem, sizeof(mem)); 922 } 923 free(buf); 924 fclose(fp); 925 return ret; 926 } 927 928 static int write_topo_node(int fd, int node) 929 { 930 char str[MAXPATHLEN]; 931 char field[32]; 932 char *buf = NULL, *p; 933 size_t len = 0; 934 FILE *fp; 935 u64 mem_total, mem_free, mem; 936 int ret = -1; 937 938 sprintf(str, "/sys/devices/system/node/node%d/meminfo", node); 939 fp = fopen(str, "r"); 940 if (!fp) 941 return -1; 942 943 while (getline(&buf, &len, fp) > 0) { 944 /* skip over invalid lines */ 945 if (!strchr(buf, ':')) 946 continue; 947 if (sscanf(buf, "%*s %*d %s %"PRIu64, field, &mem) != 2) 948 goto done; 949 if (!strcmp(field, "MemTotal:")) 950 mem_total = mem; 951 if (!strcmp(field, "MemFree:")) 952 mem_free = mem; 953 } 954 955 fclose(fp); 956 957 ret = do_write(fd, &mem_total, sizeof(u64)); 958 if (ret) 959 goto done; 960 961 ret = do_write(fd, &mem_free, sizeof(u64)); 962 if (ret) 963 goto done; 964 965 ret = -1; 966 sprintf(str, "/sys/devices/system/node/node%d/cpulist", node); 967 968 fp = fopen(str, "r"); 969 if (!fp) 970 goto done; 971 972 if (getline(&buf, &len, fp) <= 0) 973 goto done; 974 975 p = strchr(buf, '\n'); 976 if (p) 977 *p = '\0'; 978 979 ret = do_write_string(fd, buf); 980 done: 981 free(buf); 982 fclose(fp); 983 return ret; 984 } 985 986 static int write_numa_topology(int fd, struct perf_header *h __maybe_unused, 987 struct perf_evlist *evlist __maybe_unused) 988 { 989 char *buf = NULL; 990 size_t len = 0; 991 FILE *fp; 992 struct cpu_map *node_map = NULL; 993 char *c; 994 u32 nr, i, j; 995 int ret = -1; 996 997 fp = fopen("/sys/devices/system/node/online", "r"); 998 if (!fp) 999 return -1; 1000 1001 if (getline(&buf, &len, fp) <= 0) 1002 goto done; 1003 1004 c = strchr(buf, '\n'); 1005 if (c) 1006 *c = '\0'; 1007 1008 node_map = cpu_map__new(buf); 1009 if (!node_map) 1010 goto done; 1011 1012 nr = (u32)node_map->nr; 1013 1014 ret = do_write(fd, &nr, sizeof(nr)); 1015 if (ret < 0) 1016 goto done; 1017 1018 for (i = 0; i < nr; i++) { 1019 j = (u32)node_map->map[i]; 1020 ret = do_write(fd, &j, sizeof(j)); 1021 if (ret < 0) 1022 break; 1023 1024 ret = write_topo_node(fd, i); 1025 if (ret < 0) 1026 break; 1027 } 1028 done: 1029 free(buf); 1030 fclose(fp); 1031 free(node_map); 1032 return ret; 1033 } 1034 1035 /* 1036 * File format: 1037 * 1038 * struct pmu_mappings { 1039 * u32 pmu_num; 1040 * struct pmu_map { 1041 * u32 type; 1042 * char name[]; 1043 * }[pmu_num]; 1044 * }; 1045 */ 1046 1047 static int write_pmu_mappings(int fd, struct perf_header *h __maybe_unused, 1048 struct perf_evlist *evlist __maybe_unused) 1049 { 1050 struct perf_pmu *pmu = NULL; 1051 off_t offset = lseek(fd, 0, SEEK_CUR); 1052 __u32 pmu_num = 0; 1053 1054 /* write real pmu_num later */ 1055 do_write(fd, &pmu_num, sizeof(pmu_num)); 1056 1057 while ((pmu = perf_pmu__scan(pmu))) { 1058 if (!pmu->name) 1059 continue; 1060 pmu_num++; 1061 do_write(fd, &pmu->type, sizeof(pmu->type)); 1062 do_write_string(fd, pmu->name); 1063 } 1064 1065 if (pwrite(fd, &pmu_num, sizeof(pmu_num), offset) != sizeof(pmu_num)) { 1066 /* discard all */ 1067 lseek(fd, offset, SEEK_SET); 1068 return -1; 1069 } 1070 1071 return 0; 1072 } 1073 1074 /* 1075 * default get_cpuid(): nothing gets recorded 1076 * actual implementation must be in arch/$(ARCH)/util/header.c 1077 */ 1078 int __attribute__ ((weak)) get_cpuid(char *buffer __maybe_unused, 1079 size_t sz __maybe_unused) 1080 { 1081 return -1; 1082 } 1083 1084 static int write_cpuid(int fd, struct perf_header *h __maybe_unused, 1085 struct perf_evlist *evlist __maybe_unused) 1086 { 1087 char buffer[64]; 1088 int ret; 1089 1090 ret = get_cpuid(buffer, sizeof(buffer)); 1091 if (!ret) 1092 goto write_it; 1093 1094 return -1; 1095 write_it: 1096 return do_write_string(fd, buffer); 1097 } 1098 1099 static int write_branch_stack(int fd __maybe_unused, 1100 struct perf_header *h __maybe_unused, 1101 struct perf_evlist *evlist __maybe_unused) 1102 { 1103 return 0; 1104 } 1105 1106 static void print_hostname(struct perf_header *ph, int fd __maybe_unused, 1107 FILE *fp) 1108 { 1109 fprintf(fp, "# hostname : %s\n", ph->env.hostname); 1110 } 1111 1112 static void print_osrelease(struct perf_header *ph, int fd __maybe_unused, 1113 FILE *fp) 1114 { 1115 fprintf(fp, "# os release : %s\n", ph->env.os_release); 1116 } 1117 1118 static void print_arch(struct perf_header *ph, int fd __maybe_unused, FILE *fp) 1119 { 1120 fprintf(fp, "# arch : %s\n", ph->env.arch); 1121 } 1122 1123 static void print_cpudesc(struct perf_header *ph, int fd __maybe_unused, 1124 FILE *fp) 1125 { 1126 fprintf(fp, "# cpudesc : %s\n", ph->env.cpu_desc); 1127 } 1128 1129 static void print_nrcpus(struct perf_header *ph, int fd __maybe_unused, 1130 FILE *fp) 1131 { 1132 fprintf(fp, "# nrcpus online : %u\n", ph->env.nr_cpus_online); 1133 fprintf(fp, "# nrcpus avail : %u\n", ph->env.nr_cpus_avail); 1134 } 1135 1136 static void print_version(struct perf_header *ph, int fd __maybe_unused, 1137 FILE *fp) 1138 { 1139 fprintf(fp, "# perf version : %s\n", ph->env.version); 1140 } 1141 1142 static void print_cmdline(struct perf_header *ph, int fd __maybe_unused, 1143 FILE *fp) 1144 { 1145 int nr, i; 1146 char *str; 1147 1148 nr = ph->env.nr_cmdline; 1149 str = ph->env.cmdline; 1150 1151 fprintf(fp, "# cmdline : "); 1152 1153 for (i = 0; i < nr; i++) { 1154 fprintf(fp, "%s ", str); 1155 str += strlen(str) + 1; 1156 } 1157 fputc('\n', fp); 1158 } 1159 1160 static void print_cpu_topology(struct perf_header *ph, int fd __maybe_unused, 1161 FILE *fp) 1162 { 1163 int nr, i; 1164 char *str; 1165 1166 nr = ph->env.nr_sibling_cores; 1167 str = ph->env.sibling_cores; 1168 1169 for (i = 0; i < nr; i++) { 1170 fprintf(fp, "# sibling cores : %s\n", str); 1171 str += strlen(str) + 1; 1172 } 1173 1174 nr = ph->env.nr_sibling_threads; 1175 str = ph->env.sibling_threads; 1176 1177 for (i = 0; i < nr; i++) { 1178 fprintf(fp, "# sibling threads : %s\n", str); 1179 str += strlen(str) + 1; 1180 } 1181 } 1182 1183 static void free_event_desc(struct perf_evsel *events) 1184 { 1185 struct perf_evsel *evsel; 1186 1187 if (!events) 1188 return; 1189 1190 for (evsel = events; evsel->attr.size; evsel++) { 1191 if (evsel->name) 1192 free(evsel->name); 1193 if (evsel->id) 1194 free(evsel->id); 1195 } 1196 1197 free(events); 1198 } 1199 1200 static struct perf_evsel * 1201 read_event_desc(struct perf_header *ph, int fd) 1202 { 1203 struct perf_evsel *evsel, *events = NULL; 1204 u64 *id; 1205 void *buf = NULL; 1206 u32 nre, sz, nr, i, j; 1207 ssize_t ret; 1208 size_t msz; 1209 1210 /* number of events */ 1211 ret = read(fd, &nre, sizeof(nre)); 1212 if (ret != (ssize_t)sizeof(nre)) 1213 goto error; 1214 1215 if (ph->needs_swap) 1216 nre = bswap_32(nre); 1217 1218 ret = read(fd, &sz, sizeof(sz)); 1219 if (ret != (ssize_t)sizeof(sz)) 1220 goto error; 1221 1222 if (ph->needs_swap) 1223 sz = bswap_32(sz); 1224 1225 /* buffer to hold on file attr struct */ 1226 buf = malloc(sz); 1227 if (!buf) 1228 goto error; 1229 1230 /* the last event terminates with evsel->attr.size == 0: */ 1231 events = calloc(nre + 1, sizeof(*events)); 1232 if (!events) 1233 goto error; 1234 1235 msz = sizeof(evsel->attr); 1236 if (sz < msz) 1237 msz = sz; 1238 1239 for (i = 0, evsel = events; i < nre; evsel++, i++) { 1240 evsel->idx = i; 1241 1242 /* 1243 * must read entire on-file attr struct to 1244 * sync up with layout. 1245 */ 1246 ret = read(fd, buf, sz); 1247 if (ret != (ssize_t)sz) 1248 goto error; 1249 1250 if (ph->needs_swap) 1251 perf_event__attr_swap(buf); 1252 1253 memcpy(&evsel->attr, buf, msz); 1254 1255 ret = read(fd, &nr, sizeof(nr)); 1256 if (ret != (ssize_t)sizeof(nr)) 1257 goto error; 1258 1259 if (ph->needs_swap) { 1260 nr = bswap_32(nr); 1261 evsel->needs_swap = true; 1262 } 1263 1264 evsel->name = do_read_string(fd, ph); 1265 1266 if (!nr) 1267 continue; 1268 1269 id = calloc(nr, sizeof(*id)); 1270 if (!id) 1271 goto error; 1272 evsel->ids = nr; 1273 evsel->id = id; 1274 1275 for (j = 0 ; j < nr; j++) { 1276 ret = read(fd, id, sizeof(*id)); 1277 if (ret != (ssize_t)sizeof(*id)) 1278 goto error; 1279 if (ph->needs_swap) 1280 *id = bswap_64(*id); 1281 id++; 1282 } 1283 } 1284 out: 1285 if (buf) 1286 free(buf); 1287 return events; 1288 error: 1289 if (events) 1290 free_event_desc(events); 1291 events = NULL; 1292 goto out; 1293 } 1294 1295 static void print_event_desc(struct perf_header *ph, int fd, FILE *fp) 1296 { 1297 struct perf_evsel *evsel, *events = read_event_desc(ph, fd); 1298 u32 j; 1299 u64 *id; 1300 1301 if (!events) { 1302 fprintf(fp, "# event desc: not available or unable to read\n"); 1303 return; 1304 } 1305 1306 for (evsel = events; evsel->attr.size; evsel++) { 1307 fprintf(fp, "# event : name = %s, ", evsel->name); 1308 1309 fprintf(fp, "type = %d, config = 0x%"PRIx64 1310 ", config1 = 0x%"PRIx64", config2 = 0x%"PRIx64, 1311 evsel->attr.type, 1312 (u64)evsel->attr.config, 1313 (u64)evsel->attr.config1, 1314 (u64)evsel->attr.config2); 1315 1316 fprintf(fp, ", excl_usr = %d, excl_kern = %d", 1317 evsel->attr.exclude_user, 1318 evsel->attr.exclude_kernel); 1319 1320 fprintf(fp, ", excl_host = %d, excl_guest = %d", 1321 evsel->attr.exclude_host, 1322 evsel->attr.exclude_guest); 1323 1324 fprintf(fp, ", precise_ip = %d", evsel->attr.precise_ip); 1325 1326 if (evsel->ids) { 1327 fprintf(fp, ", id = {"); 1328 for (j = 0, id = evsel->id; j < evsel->ids; j++, id++) { 1329 if (j) 1330 fputc(',', fp); 1331 fprintf(fp, " %"PRIu64, *id); 1332 } 1333 fprintf(fp, " }"); 1334 } 1335 1336 fputc('\n', fp); 1337 } 1338 1339 free_event_desc(events); 1340 } 1341 1342 static void print_total_mem(struct perf_header *ph, int fd __maybe_unused, 1343 FILE *fp) 1344 { 1345 fprintf(fp, "# total memory : %Lu kB\n", ph->env.total_mem); 1346 } 1347 1348 static void print_numa_topology(struct perf_header *ph, int fd __maybe_unused, 1349 FILE *fp) 1350 { 1351 u32 nr, c, i; 1352 char *str, *tmp; 1353 uint64_t mem_total, mem_free; 1354 1355 /* nr nodes */ 1356 nr = ph->env.nr_numa_nodes; 1357 str = ph->env.numa_nodes; 1358 1359 for (i = 0; i < nr; i++) { 1360 /* node number */ 1361 c = strtoul(str, &tmp, 0); 1362 if (*tmp != ':') 1363 goto error; 1364 1365 str = tmp + 1; 1366 mem_total = strtoull(str, &tmp, 0); 1367 if (*tmp != ':') 1368 goto error; 1369 1370 str = tmp + 1; 1371 mem_free = strtoull(str, &tmp, 0); 1372 if (*tmp != ':') 1373 goto error; 1374 1375 fprintf(fp, "# node%u meminfo : total = %"PRIu64" kB," 1376 " free = %"PRIu64" kB\n", 1377 c, mem_total, mem_free); 1378 1379 str = tmp + 1; 1380 fprintf(fp, "# node%u cpu list : %s\n", c, str); 1381 } 1382 return; 1383 error: 1384 fprintf(fp, "# numa topology : not available\n"); 1385 } 1386 1387 static void print_cpuid(struct perf_header *ph, int fd __maybe_unused, FILE *fp) 1388 { 1389 fprintf(fp, "# cpuid : %s\n", ph->env.cpuid); 1390 } 1391 1392 static void print_branch_stack(struct perf_header *ph __maybe_unused, 1393 int fd __maybe_unused, FILE *fp) 1394 { 1395 fprintf(fp, "# contains samples with branch stack\n"); 1396 } 1397 1398 static void print_pmu_mappings(struct perf_header *ph, int fd __maybe_unused, 1399 FILE *fp) 1400 { 1401 const char *delimiter = "# pmu mappings: "; 1402 char *str, *tmp; 1403 u32 pmu_num; 1404 u32 type; 1405 1406 pmu_num = ph->env.nr_pmu_mappings; 1407 if (!pmu_num) { 1408 fprintf(fp, "# pmu mappings: not available\n"); 1409 return; 1410 } 1411 1412 str = ph->env.pmu_mappings; 1413 1414 while (pmu_num) { 1415 type = strtoul(str, &tmp, 0); 1416 if (*tmp != ':') 1417 goto error; 1418 1419 str = tmp + 1; 1420 fprintf(fp, "%s%s = %" PRIu32, delimiter, str, type); 1421 1422 delimiter = ", "; 1423 str += strlen(str) + 1; 1424 pmu_num--; 1425 } 1426 1427 fprintf(fp, "\n"); 1428 1429 if (!pmu_num) 1430 return; 1431 error: 1432 fprintf(fp, "# pmu mappings: unable to read\n"); 1433 } 1434 1435 static int __event_process_build_id(struct build_id_event *bev, 1436 char *filename, 1437 struct perf_session *session) 1438 { 1439 int err = -1; 1440 struct list_head *head; 1441 struct machine *machine; 1442 u16 misc; 1443 struct dso *dso; 1444 enum dso_kernel_type dso_type; 1445 1446 machine = perf_session__findnew_machine(session, bev->pid); 1447 if (!machine) 1448 goto out; 1449 1450 misc = bev->header.misc & PERF_RECORD_MISC_CPUMODE_MASK; 1451 1452 switch (misc) { 1453 case PERF_RECORD_MISC_KERNEL: 1454 dso_type = DSO_TYPE_KERNEL; 1455 head = &machine->kernel_dsos; 1456 break; 1457 case PERF_RECORD_MISC_GUEST_KERNEL: 1458 dso_type = DSO_TYPE_GUEST_KERNEL; 1459 head = &machine->kernel_dsos; 1460 break; 1461 case PERF_RECORD_MISC_USER: 1462 case PERF_RECORD_MISC_GUEST_USER: 1463 dso_type = DSO_TYPE_USER; 1464 head = &machine->user_dsos; 1465 break; 1466 default: 1467 goto out; 1468 } 1469 1470 dso = __dsos__findnew(head, filename); 1471 if (dso != NULL) { 1472 char sbuild_id[BUILD_ID_SIZE * 2 + 1]; 1473 1474 dso__set_build_id(dso, &bev->build_id); 1475 1476 if (filename[0] == '[') 1477 dso->kernel = dso_type; 1478 1479 build_id__sprintf(dso->build_id, sizeof(dso->build_id), 1480 sbuild_id); 1481 pr_debug("build id event received for %s: %s\n", 1482 dso->long_name, sbuild_id); 1483 } 1484 1485 err = 0; 1486 out: 1487 return err; 1488 } 1489 1490 static int perf_header__read_build_ids_abi_quirk(struct perf_header *header, 1491 int input, u64 offset, u64 size) 1492 { 1493 struct perf_session *session = container_of(header, struct perf_session, header); 1494 struct { 1495 struct perf_event_header header; 1496 u8 build_id[PERF_ALIGN(BUILD_ID_SIZE, sizeof(u64))]; 1497 char filename[0]; 1498 } old_bev; 1499 struct build_id_event bev; 1500 char filename[PATH_MAX]; 1501 u64 limit = offset + size; 1502 1503 while (offset < limit) { 1504 ssize_t len; 1505 1506 if (read(input, &old_bev, sizeof(old_bev)) != sizeof(old_bev)) 1507 return -1; 1508 1509 if (header->needs_swap) 1510 perf_event_header__bswap(&old_bev.header); 1511 1512 len = old_bev.header.size - sizeof(old_bev); 1513 if (read(input, filename, len) != len) 1514 return -1; 1515 1516 bev.header = old_bev.header; 1517 1518 /* 1519 * As the pid is the missing value, we need to fill 1520 * it properly. The header.misc value give us nice hint. 1521 */ 1522 bev.pid = HOST_KERNEL_ID; 1523 if (bev.header.misc == PERF_RECORD_MISC_GUEST_USER || 1524 bev.header.misc == PERF_RECORD_MISC_GUEST_KERNEL) 1525 bev.pid = DEFAULT_GUEST_KERNEL_ID; 1526 1527 memcpy(bev.build_id, old_bev.build_id, sizeof(bev.build_id)); 1528 __event_process_build_id(&bev, filename, session); 1529 1530 offset += bev.header.size; 1531 } 1532 1533 return 0; 1534 } 1535 1536 static int perf_header__read_build_ids(struct perf_header *header, 1537 int input, u64 offset, u64 size) 1538 { 1539 struct perf_session *session = container_of(header, struct perf_session, header); 1540 struct build_id_event bev; 1541 char filename[PATH_MAX]; 1542 u64 limit = offset + size, orig_offset = offset; 1543 int err = -1; 1544 1545 while (offset < limit) { 1546 ssize_t len; 1547 1548 if (read(input, &bev, sizeof(bev)) != sizeof(bev)) 1549 goto out; 1550 1551 if (header->needs_swap) 1552 perf_event_header__bswap(&bev.header); 1553 1554 len = bev.header.size - sizeof(bev); 1555 if (read(input, filename, len) != len) 1556 goto out; 1557 /* 1558 * The a1645ce1 changeset: 1559 * 1560 * "perf: 'perf kvm' tool for monitoring guest performance from host" 1561 * 1562 * Added a field to struct build_id_event that broke the file 1563 * format. 1564 * 1565 * Since the kernel build-id is the first entry, process the 1566 * table using the old format if the well known 1567 * '[kernel.kallsyms]' string for the kernel build-id has the 1568 * first 4 characters chopped off (where the pid_t sits). 1569 */ 1570 if (memcmp(filename, "nel.kallsyms]", 13) == 0) { 1571 if (lseek(input, orig_offset, SEEK_SET) == (off_t)-1) 1572 return -1; 1573 return perf_header__read_build_ids_abi_quirk(header, input, offset, size); 1574 } 1575 1576 __event_process_build_id(&bev, filename, session); 1577 1578 offset += bev.header.size; 1579 } 1580 err = 0; 1581 out: 1582 return err; 1583 } 1584 1585 static int process_tracing_data(struct perf_file_section *section __maybe_unused, 1586 struct perf_header *ph __maybe_unused, 1587 int fd, void *data) 1588 { 1589 trace_report(fd, data, false); 1590 return 0; 1591 } 1592 1593 static int process_build_id(struct perf_file_section *section, 1594 struct perf_header *ph, int fd, 1595 void *data __maybe_unused) 1596 { 1597 if (perf_header__read_build_ids(ph, fd, section->offset, section->size)) 1598 pr_debug("Failed to read buildids, continuing...\n"); 1599 return 0; 1600 } 1601 1602 static int process_hostname(struct perf_file_section *section __maybe_unused, 1603 struct perf_header *ph, int fd, 1604 void *data __maybe_unused) 1605 { 1606 ph->env.hostname = do_read_string(fd, ph); 1607 return ph->env.hostname ? 0 : -ENOMEM; 1608 } 1609 1610 static int process_osrelease(struct perf_file_section *section __maybe_unused, 1611 struct perf_header *ph, int fd, 1612 void *data __maybe_unused) 1613 { 1614 ph->env.os_release = do_read_string(fd, ph); 1615 return ph->env.os_release ? 0 : -ENOMEM; 1616 } 1617 1618 static int process_version(struct perf_file_section *section __maybe_unused, 1619 struct perf_header *ph, int fd, 1620 void *data __maybe_unused) 1621 { 1622 ph->env.version = do_read_string(fd, ph); 1623 return ph->env.version ? 0 : -ENOMEM; 1624 } 1625 1626 static int process_arch(struct perf_file_section *section __maybe_unused, 1627 struct perf_header *ph, int fd, 1628 void *data __maybe_unused) 1629 { 1630 ph->env.arch = do_read_string(fd, ph); 1631 return ph->env.arch ? 0 : -ENOMEM; 1632 } 1633 1634 static int process_nrcpus(struct perf_file_section *section __maybe_unused, 1635 struct perf_header *ph, int fd, 1636 void *data __maybe_unused) 1637 { 1638 size_t ret; 1639 u32 nr; 1640 1641 ret = read(fd, &nr, sizeof(nr)); 1642 if (ret != sizeof(nr)) 1643 return -1; 1644 1645 if (ph->needs_swap) 1646 nr = bswap_32(nr); 1647 1648 ph->env.nr_cpus_online = nr; 1649 1650 ret = read(fd, &nr, sizeof(nr)); 1651 if (ret != sizeof(nr)) 1652 return -1; 1653 1654 if (ph->needs_swap) 1655 nr = bswap_32(nr); 1656 1657 ph->env.nr_cpus_avail = nr; 1658 return 0; 1659 } 1660 1661 static int process_cpudesc(struct perf_file_section *section __maybe_unused, 1662 struct perf_header *ph, int fd, 1663 void *data __maybe_unused) 1664 { 1665 ph->env.cpu_desc = do_read_string(fd, ph); 1666 return ph->env.cpu_desc ? 0 : -ENOMEM; 1667 } 1668 1669 static int process_cpuid(struct perf_file_section *section __maybe_unused, 1670 struct perf_header *ph, int fd, 1671 void *data __maybe_unused) 1672 { 1673 ph->env.cpuid = do_read_string(fd, ph); 1674 return ph->env.cpuid ? 0 : -ENOMEM; 1675 } 1676 1677 static int process_total_mem(struct perf_file_section *section __maybe_unused, 1678 struct perf_header *ph, int fd, 1679 void *data __maybe_unused) 1680 { 1681 uint64_t mem; 1682 size_t ret; 1683 1684 ret = read(fd, &mem, sizeof(mem)); 1685 if (ret != sizeof(mem)) 1686 return -1; 1687 1688 if (ph->needs_swap) 1689 mem = bswap_64(mem); 1690 1691 ph->env.total_mem = mem; 1692 return 0; 1693 } 1694 1695 static struct perf_evsel * 1696 perf_evlist__find_by_index(struct perf_evlist *evlist, int idx) 1697 { 1698 struct perf_evsel *evsel; 1699 1700 list_for_each_entry(evsel, &evlist->entries, node) { 1701 if (evsel->idx == idx) 1702 return evsel; 1703 } 1704 1705 return NULL; 1706 } 1707 1708 static void 1709 perf_evlist__set_event_name(struct perf_evlist *evlist, 1710 struct perf_evsel *event) 1711 { 1712 struct perf_evsel *evsel; 1713 1714 if (!event->name) 1715 return; 1716 1717 evsel = perf_evlist__find_by_index(evlist, event->idx); 1718 if (!evsel) 1719 return; 1720 1721 if (evsel->name) 1722 return; 1723 1724 evsel->name = strdup(event->name); 1725 } 1726 1727 static int 1728 process_event_desc(struct perf_file_section *section __maybe_unused, 1729 struct perf_header *header, int fd, 1730 void *data __maybe_unused) 1731 { 1732 struct perf_session *session; 1733 struct perf_evsel *evsel, *events = read_event_desc(header, fd); 1734 1735 if (!events) 1736 return 0; 1737 1738 session = container_of(header, struct perf_session, header); 1739 for (evsel = events; evsel->attr.size; evsel++) 1740 perf_evlist__set_event_name(session->evlist, evsel); 1741 1742 free_event_desc(events); 1743 1744 return 0; 1745 } 1746 1747 static int process_cmdline(struct perf_file_section *section __maybe_unused, 1748 struct perf_header *ph, int fd, 1749 void *data __maybe_unused) 1750 { 1751 size_t ret; 1752 char *str; 1753 u32 nr, i; 1754 struct strbuf sb; 1755 1756 ret = read(fd, &nr, sizeof(nr)); 1757 if (ret != sizeof(nr)) 1758 return -1; 1759 1760 if (ph->needs_swap) 1761 nr = bswap_32(nr); 1762 1763 ph->env.nr_cmdline = nr; 1764 strbuf_init(&sb, 128); 1765 1766 for (i = 0; i < nr; i++) { 1767 str = do_read_string(fd, ph); 1768 if (!str) 1769 goto error; 1770 1771 /* include a NULL character at the end */ 1772 strbuf_add(&sb, str, strlen(str) + 1); 1773 free(str); 1774 } 1775 ph->env.cmdline = strbuf_detach(&sb, NULL); 1776 return 0; 1777 1778 error: 1779 strbuf_release(&sb); 1780 return -1; 1781 } 1782 1783 static int process_cpu_topology(struct perf_file_section *section __maybe_unused, 1784 struct perf_header *ph, int fd, 1785 void *data __maybe_unused) 1786 { 1787 size_t ret; 1788 u32 nr, i; 1789 char *str; 1790 struct strbuf sb; 1791 1792 ret = read(fd, &nr, sizeof(nr)); 1793 if (ret != sizeof(nr)) 1794 return -1; 1795 1796 if (ph->needs_swap) 1797 nr = bswap_32(nr); 1798 1799 ph->env.nr_sibling_cores = nr; 1800 strbuf_init(&sb, 128); 1801 1802 for (i = 0; i < nr; i++) { 1803 str = do_read_string(fd, ph); 1804 if (!str) 1805 goto error; 1806 1807 /* include a NULL character at the end */ 1808 strbuf_add(&sb, str, strlen(str) + 1); 1809 free(str); 1810 } 1811 ph->env.sibling_cores = strbuf_detach(&sb, NULL); 1812 1813 ret = read(fd, &nr, sizeof(nr)); 1814 if (ret != sizeof(nr)) 1815 return -1; 1816 1817 if (ph->needs_swap) 1818 nr = bswap_32(nr); 1819 1820 ph->env.nr_sibling_threads = nr; 1821 1822 for (i = 0; i < nr; i++) { 1823 str = do_read_string(fd, ph); 1824 if (!str) 1825 goto error; 1826 1827 /* include a NULL character at the end */ 1828 strbuf_add(&sb, str, strlen(str) + 1); 1829 free(str); 1830 } 1831 ph->env.sibling_threads = strbuf_detach(&sb, NULL); 1832 return 0; 1833 1834 error: 1835 strbuf_release(&sb); 1836 return -1; 1837 } 1838 1839 static int process_numa_topology(struct perf_file_section *section __maybe_unused, 1840 struct perf_header *ph, int fd, 1841 void *data __maybe_unused) 1842 { 1843 size_t ret; 1844 u32 nr, node, i; 1845 char *str; 1846 uint64_t mem_total, mem_free; 1847 struct strbuf sb; 1848 1849 /* nr nodes */ 1850 ret = read(fd, &nr, sizeof(nr)); 1851 if (ret != sizeof(nr)) 1852 goto error; 1853 1854 if (ph->needs_swap) 1855 nr = bswap_32(nr); 1856 1857 ph->env.nr_numa_nodes = nr; 1858 strbuf_init(&sb, 256); 1859 1860 for (i = 0; i < nr; i++) { 1861 /* node number */ 1862 ret = read(fd, &node, sizeof(node)); 1863 if (ret != sizeof(node)) 1864 goto error; 1865 1866 ret = read(fd, &mem_total, sizeof(u64)); 1867 if (ret != sizeof(u64)) 1868 goto error; 1869 1870 ret = read(fd, &mem_free, sizeof(u64)); 1871 if (ret != sizeof(u64)) 1872 goto error; 1873 1874 if (ph->needs_swap) { 1875 node = bswap_32(node); 1876 mem_total = bswap_64(mem_total); 1877 mem_free = bswap_64(mem_free); 1878 } 1879 1880 strbuf_addf(&sb, "%u:%"PRIu64":%"PRIu64":", 1881 node, mem_total, mem_free); 1882 1883 str = do_read_string(fd, ph); 1884 if (!str) 1885 goto error; 1886 1887 /* include a NULL character at the end */ 1888 strbuf_add(&sb, str, strlen(str) + 1); 1889 free(str); 1890 } 1891 ph->env.numa_nodes = strbuf_detach(&sb, NULL); 1892 return 0; 1893 1894 error: 1895 strbuf_release(&sb); 1896 return -1; 1897 } 1898 1899 static int process_pmu_mappings(struct perf_file_section *section __maybe_unused, 1900 struct perf_header *ph, int fd, 1901 void *data __maybe_unused) 1902 { 1903 size_t ret; 1904 char *name; 1905 u32 pmu_num; 1906 u32 type; 1907 struct strbuf sb; 1908 1909 ret = read(fd, &pmu_num, sizeof(pmu_num)); 1910 if (ret != sizeof(pmu_num)) 1911 return -1; 1912 1913 if (ph->needs_swap) 1914 pmu_num = bswap_32(pmu_num); 1915 1916 if (!pmu_num) { 1917 pr_debug("pmu mappings not available\n"); 1918 return 0; 1919 } 1920 1921 ph->env.nr_pmu_mappings = pmu_num; 1922 strbuf_init(&sb, 128); 1923 1924 while (pmu_num) { 1925 if (read(fd, &type, sizeof(type)) != sizeof(type)) 1926 goto error; 1927 if (ph->needs_swap) 1928 type = bswap_32(type); 1929 1930 name = do_read_string(fd, ph); 1931 if (!name) 1932 goto error; 1933 1934 strbuf_addf(&sb, "%u:%s", type, name); 1935 /* include a NULL character at the end */ 1936 strbuf_add(&sb, "", 1); 1937 1938 free(name); 1939 pmu_num--; 1940 } 1941 ph->env.pmu_mappings = strbuf_detach(&sb, NULL); 1942 return 0; 1943 1944 error: 1945 strbuf_release(&sb); 1946 return -1; 1947 } 1948 1949 struct feature_ops { 1950 int (*write)(int fd, struct perf_header *h, struct perf_evlist *evlist); 1951 void (*print)(struct perf_header *h, int fd, FILE *fp); 1952 int (*process)(struct perf_file_section *section, 1953 struct perf_header *h, int fd, void *data); 1954 const char *name; 1955 bool full_only; 1956 }; 1957 1958 #define FEAT_OPA(n, func) \ 1959 [n] = { .name = #n, .write = write_##func, .print = print_##func } 1960 #define FEAT_OPP(n, func) \ 1961 [n] = { .name = #n, .write = write_##func, .print = print_##func, \ 1962 .process = process_##func } 1963 #define FEAT_OPF(n, func) \ 1964 [n] = { .name = #n, .write = write_##func, .print = print_##func, \ 1965 .process = process_##func, .full_only = true } 1966 1967 /* feature_ops not implemented: */ 1968 #define print_tracing_data NULL 1969 #define print_build_id NULL 1970 1971 static const struct feature_ops feat_ops[HEADER_LAST_FEATURE] = { 1972 FEAT_OPP(HEADER_TRACING_DATA, tracing_data), 1973 FEAT_OPP(HEADER_BUILD_ID, build_id), 1974 FEAT_OPP(HEADER_HOSTNAME, hostname), 1975 FEAT_OPP(HEADER_OSRELEASE, osrelease), 1976 FEAT_OPP(HEADER_VERSION, version), 1977 FEAT_OPP(HEADER_ARCH, arch), 1978 FEAT_OPP(HEADER_NRCPUS, nrcpus), 1979 FEAT_OPP(HEADER_CPUDESC, cpudesc), 1980 FEAT_OPP(HEADER_CPUID, cpuid), 1981 FEAT_OPP(HEADER_TOTAL_MEM, total_mem), 1982 FEAT_OPP(HEADER_EVENT_DESC, event_desc), 1983 FEAT_OPP(HEADER_CMDLINE, cmdline), 1984 FEAT_OPF(HEADER_CPU_TOPOLOGY, cpu_topology), 1985 FEAT_OPF(HEADER_NUMA_TOPOLOGY, numa_topology), 1986 FEAT_OPA(HEADER_BRANCH_STACK, branch_stack), 1987 FEAT_OPP(HEADER_PMU_MAPPINGS, pmu_mappings), 1988 }; 1989 1990 struct header_print_data { 1991 FILE *fp; 1992 bool full; /* extended list of headers */ 1993 }; 1994 1995 static int perf_file_section__fprintf_info(struct perf_file_section *section, 1996 struct perf_header *ph, 1997 int feat, int fd, void *data) 1998 { 1999 struct header_print_data *hd = data; 2000 2001 if (lseek(fd, section->offset, SEEK_SET) == (off_t)-1) { 2002 pr_debug("Failed to lseek to %" PRIu64 " offset for feature " 2003 "%d, continuing...\n", section->offset, feat); 2004 return 0; 2005 } 2006 if (feat >= HEADER_LAST_FEATURE) { 2007 pr_warning("unknown feature %d\n", feat); 2008 return 0; 2009 } 2010 if (!feat_ops[feat].print) 2011 return 0; 2012 2013 if (!feat_ops[feat].full_only || hd->full) 2014 feat_ops[feat].print(ph, fd, hd->fp); 2015 else 2016 fprintf(hd->fp, "# %s info available, use -I to display\n", 2017 feat_ops[feat].name); 2018 2019 return 0; 2020 } 2021 2022 int perf_header__fprintf_info(struct perf_session *session, FILE *fp, bool full) 2023 { 2024 struct header_print_data hd; 2025 struct perf_header *header = &session->header; 2026 int fd = session->fd; 2027 hd.fp = fp; 2028 hd.full = full; 2029 2030 perf_header__process_sections(header, fd, &hd, 2031 perf_file_section__fprintf_info); 2032 return 0; 2033 } 2034 2035 static int do_write_feat(int fd, struct perf_header *h, int type, 2036 struct perf_file_section **p, 2037 struct perf_evlist *evlist) 2038 { 2039 int err; 2040 int ret = 0; 2041 2042 if (perf_header__has_feat(h, type)) { 2043 if (!feat_ops[type].write) 2044 return -1; 2045 2046 (*p)->offset = lseek(fd, 0, SEEK_CUR); 2047 2048 err = feat_ops[type].write(fd, h, evlist); 2049 if (err < 0) { 2050 pr_debug("failed to write feature %d\n", type); 2051 2052 /* undo anything written */ 2053 lseek(fd, (*p)->offset, SEEK_SET); 2054 2055 return -1; 2056 } 2057 (*p)->size = lseek(fd, 0, SEEK_CUR) - (*p)->offset; 2058 (*p)++; 2059 } 2060 return ret; 2061 } 2062 2063 static int perf_header__adds_write(struct perf_header *header, 2064 struct perf_evlist *evlist, int fd) 2065 { 2066 int nr_sections; 2067 struct perf_file_section *feat_sec, *p; 2068 int sec_size; 2069 u64 sec_start; 2070 int feat; 2071 int err; 2072 2073 nr_sections = bitmap_weight(header->adds_features, HEADER_FEAT_BITS); 2074 if (!nr_sections) 2075 return 0; 2076 2077 feat_sec = p = calloc(sizeof(*feat_sec), nr_sections); 2078 if (feat_sec == NULL) 2079 return -ENOMEM; 2080 2081 sec_size = sizeof(*feat_sec) * nr_sections; 2082 2083 sec_start = header->data_offset + header->data_size; 2084 lseek(fd, sec_start + sec_size, SEEK_SET); 2085 2086 for_each_set_bit(feat, header->adds_features, HEADER_FEAT_BITS) { 2087 if (do_write_feat(fd, header, feat, &p, evlist)) 2088 perf_header__clear_feat(header, feat); 2089 } 2090 2091 lseek(fd, sec_start, SEEK_SET); 2092 /* 2093 * may write more than needed due to dropped feature, but 2094 * this is okay, reader will skip the mising entries 2095 */ 2096 err = do_write(fd, feat_sec, sec_size); 2097 if (err < 0) 2098 pr_debug("failed to write feature section\n"); 2099 free(feat_sec); 2100 return err; 2101 } 2102 2103 int perf_header__write_pipe(int fd) 2104 { 2105 struct perf_pipe_file_header f_header; 2106 int err; 2107 2108 f_header = (struct perf_pipe_file_header){ 2109 .magic = PERF_MAGIC, 2110 .size = sizeof(f_header), 2111 }; 2112 2113 err = do_write(fd, &f_header, sizeof(f_header)); 2114 if (err < 0) { 2115 pr_debug("failed to write perf pipe header\n"); 2116 return err; 2117 } 2118 2119 return 0; 2120 } 2121 2122 int perf_session__write_header(struct perf_session *session, 2123 struct perf_evlist *evlist, 2124 int fd, bool at_exit) 2125 { 2126 struct perf_file_header f_header; 2127 struct perf_file_attr f_attr; 2128 struct perf_header *header = &session->header; 2129 struct perf_evsel *evsel, *pair = NULL; 2130 int err; 2131 2132 lseek(fd, sizeof(f_header), SEEK_SET); 2133 2134 if (session->evlist != evlist) 2135 pair = perf_evlist__first(session->evlist); 2136 2137 list_for_each_entry(evsel, &evlist->entries, node) { 2138 evsel->id_offset = lseek(fd, 0, SEEK_CUR); 2139 err = do_write(fd, evsel->id, evsel->ids * sizeof(u64)); 2140 if (err < 0) { 2141 out_err_write: 2142 pr_debug("failed to write perf header\n"); 2143 return err; 2144 } 2145 if (session->evlist != evlist) { 2146 err = do_write(fd, pair->id, pair->ids * sizeof(u64)); 2147 if (err < 0) 2148 goto out_err_write; 2149 evsel->ids += pair->ids; 2150 pair = perf_evsel__next(pair); 2151 } 2152 } 2153 2154 header->attr_offset = lseek(fd, 0, SEEK_CUR); 2155 2156 list_for_each_entry(evsel, &evlist->entries, node) { 2157 f_attr = (struct perf_file_attr){ 2158 .attr = evsel->attr, 2159 .ids = { 2160 .offset = evsel->id_offset, 2161 .size = evsel->ids * sizeof(u64), 2162 } 2163 }; 2164 err = do_write(fd, &f_attr, sizeof(f_attr)); 2165 if (err < 0) { 2166 pr_debug("failed to write perf header attribute\n"); 2167 return err; 2168 } 2169 } 2170 2171 header->event_offset = lseek(fd, 0, SEEK_CUR); 2172 header->event_size = trace_event_count * sizeof(struct perf_trace_event_type); 2173 if (trace_events) { 2174 err = do_write(fd, trace_events, header->event_size); 2175 if (err < 0) { 2176 pr_debug("failed to write perf header events\n"); 2177 return err; 2178 } 2179 } 2180 2181 header->data_offset = lseek(fd, 0, SEEK_CUR); 2182 2183 if (at_exit) { 2184 err = perf_header__adds_write(header, evlist, fd); 2185 if (err < 0) 2186 return err; 2187 } 2188 2189 f_header = (struct perf_file_header){ 2190 .magic = PERF_MAGIC, 2191 .size = sizeof(f_header), 2192 .attr_size = sizeof(f_attr), 2193 .attrs = { 2194 .offset = header->attr_offset, 2195 .size = evlist->nr_entries * sizeof(f_attr), 2196 }, 2197 .data = { 2198 .offset = header->data_offset, 2199 .size = header->data_size, 2200 }, 2201 .event_types = { 2202 .offset = header->event_offset, 2203 .size = header->event_size, 2204 }, 2205 }; 2206 2207 memcpy(&f_header.adds_features, &header->adds_features, sizeof(header->adds_features)); 2208 2209 lseek(fd, 0, SEEK_SET); 2210 err = do_write(fd, &f_header, sizeof(f_header)); 2211 if (err < 0) { 2212 pr_debug("failed to write perf header\n"); 2213 return err; 2214 } 2215 lseek(fd, header->data_offset + header->data_size, SEEK_SET); 2216 2217 header->frozen = 1; 2218 return 0; 2219 } 2220 2221 static int perf_header__getbuffer64(struct perf_header *header, 2222 int fd, void *buf, size_t size) 2223 { 2224 if (readn(fd, buf, size) <= 0) 2225 return -1; 2226 2227 if (header->needs_swap) 2228 mem_bswap_64(buf, size); 2229 2230 return 0; 2231 } 2232 2233 int perf_header__process_sections(struct perf_header *header, int fd, 2234 void *data, 2235 int (*process)(struct perf_file_section *section, 2236 struct perf_header *ph, 2237 int feat, int fd, void *data)) 2238 { 2239 struct perf_file_section *feat_sec, *sec; 2240 int nr_sections; 2241 int sec_size; 2242 int feat; 2243 int err; 2244 2245 nr_sections = bitmap_weight(header->adds_features, HEADER_FEAT_BITS); 2246 if (!nr_sections) 2247 return 0; 2248 2249 feat_sec = sec = calloc(sizeof(*feat_sec), nr_sections); 2250 if (!feat_sec) 2251 return -1; 2252 2253 sec_size = sizeof(*feat_sec) * nr_sections; 2254 2255 lseek(fd, header->data_offset + header->data_size, SEEK_SET); 2256 2257 err = perf_header__getbuffer64(header, fd, feat_sec, sec_size); 2258 if (err < 0) 2259 goto out_free; 2260 2261 for_each_set_bit(feat, header->adds_features, HEADER_LAST_FEATURE) { 2262 err = process(sec++, header, feat, fd, data); 2263 if (err < 0) 2264 goto out_free; 2265 } 2266 err = 0; 2267 out_free: 2268 free(feat_sec); 2269 return err; 2270 } 2271 2272 static const int attr_file_abi_sizes[] = { 2273 [0] = PERF_ATTR_SIZE_VER0, 2274 [1] = PERF_ATTR_SIZE_VER1, 2275 [2] = PERF_ATTR_SIZE_VER2, 2276 [3] = PERF_ATTR_SIZE_VER3, 2277 0, 2278 }; 2279 2280 /* 2281 * In the legacy file format, the magic number is not used to encode endianness. 2282 * hdr_sz was used to encode endianness. But given that hdr_sz can vary based 2283 * on ABI revisions, we need to try all combinations for all endianness to 2284 * detect the endianness. 2285 */ 2286 static int try_all_file_abis(uint64_t hdr_sz, struct perf_header *ph) 2287 { 2288 uint64_t ref_size, attr_size; 2289 int i; 2290 2291 for (i = 0 ; attr_file_abi_sizes[i]; i++) { 2292 ref_size = attr_file_abi_sizes[i] 2293 + sizeof(struct perf_file_section); 2294 if (hdr_sz != ref_size) { 2295 attr_size = bswap_64(hdr_sz); 2296 if (attr_size != ref_size) 2297 continue; 2298 2299 ph->needs_swap = true; 2300 } 2301 pr_debug("ABI%d perf.data file detected, need_swap=%d\n", 2302 i, 2303 ph->needs_swap); 2304 return 0; 2305 } 2306 /* could not determine endianness */ 2307 return -1; 2308 } 2309 2310 #define PERF_PIPE_HDR_VER0 16 2311 2312 static const size_t attr_pipe_abi_sizes[] = { 2313 [0] = PERF_PIPE_HDR_VER0, 2314 0, 2315 }; 2316 2317 /* 2318 * In the legacy pipe format, there is an implicit assumption that endiannesss 2319 * between host recording the samples, and host parsing the samples is the 2320 * same. This is not always the case given that the pipe output may always be 2321 * redirected into a file and analyzed on a different machine with possibly a 2322 * different endianness and perf_event ABI revsions in the perf tool itself. 2323 */ 2324 static int try_all_pipe_abis(uint64_t hdr_sz, struct perf_header *ph) 2325 { 2326 u64 attr_size; 2327 int i; 2328 2329 for (i = 0 ; attr_pipe_abi_sizes[i]; i++) { 2330 if (hdr_sz != attr_pipe_abi_sizes[i]) { 2331 attr_size = bswap_64(hdr_sz); 2332 if (attr_size != hdr_sz) 2333 continue; 2334 2335 ph->needs_swap = true; 2336 } 2337 pr_debug("Pipe ABI%d perf.data file detected\n", i); 2338 return 0; 2339 } 2340 return -1; 2341 } 2342 2343 static int check_magic_endian(u64 magic, uint64_t hdr_sz, 2344 bool is_pipe, struct perf_header *ph) 2345 { 2346 int ret; 2347 2348 /* check for legacy format */ 2349 ret = memcmp(&magic, __perf_magic1, sizeof(magic)); 2350 if (ret == 0) { 2351 pr_debug("legacy perf.data format\n"); 2352 if (is_pipe) 2353 return try_all_pipe_abis(hdr_sz, ph); 2354 2355 return try_all_file_abis(hdr_sz, ph); 2356 } 2357 /* 2358 * the new magic number serves two purposes: 2359 * - unique number to identify actual perf.data files 2360 * - encode endianness of file 2361 */ 2362 2363 /* check magic number with one endianness */ 2364 if (magic == __perf_magic2) 2365 return 0; 2366 2367 /* check magic number with opposite endianness */ 2368 if (magic != __perf_magic2_sw) 2369 return -1; 2370 2371 ph->needs_swap = true; 2372 2373 return 0; 2374 } 2375 2376 int perf_file_header__read(struct perf_file_header *header, 2377 struct perf_header *ph, int fd) 2378 { 2379 int ret; 2380 2381 lseek(fd, 0, SEEK_SET); 2382 2383 ret = readn(fd, header, sizeof(*header)); 2384 if (ret <= 0) 2385 return -1; 2386 2387 if (check_magic_endian(header->magic, 2388 header->attr_size, false, ph) < 0) { 2389 pr_debug("magic/endian check failed\n"); 2390 return -1; 2391 } 2392 2393 if (ph->needs_swap) { 2394 mem_bswap_64(header, offsetof(struct perf_file_header, 2395 adds_features)); 2396 } 2397 2398 if (header->size != sizeof(*header)) { 2399 /* Support the previous format */ 2400 if (header->size == offsetof(typeof(*header), adds_features)) 2401 bitmap_zero(header->adds_features, HEADER_FEAT_BITS); 2402 else 2403 return -1; 2404 } else if (ph->needs_swap) { 2405 /* 2406 * feature bitmap is declared as an array of unsigned longs -- 2407 * not good since its size can differ between the host that 2408 * generated the data file and the host analyzing the file. 2409 * 2410 * We need to handle endianness, but we don't know the size of 2411 * the unsigned long where the file was generated. Take a best 2412 * guess at determining it: try 64-bit swap first (ie., file 2413 * created on a 64-bit host), and check if the hostname feature 2414 * bit is set (this feature bit is forced on as of fbe96f2). 2415 * If the bit is not, undo the 64-bit swap and try a 32-bit 2416 * swap. If the hostname bit is still not set (e.g., older data 2417 * file), punt and fallback to the original behavior -- 2418 * clearing all feature bits and setting buildid. 2419 */ 2420 mem_bswap_64(&header->adds_features, 2421 BITS_TO_U64(HEADER_FEAT_BITS)); 2422 2423 if (!test_bit(HEADER_HOSTNAME, header->adds_features)) { 2424 /* unswap as u64 */ 2425 mem_bswap_64(&header->adds_features, 2426 BITS_TO_U64(HEADER_FEAT_BITS)); 2427 2428 /* unswap as u32 */ 2429 mem_bswap_32(&header->adds_features, 2430 BITS_TO_U32(HEADER_FEAT_BITS)); 2431 } 2432 2433 if (!test_bit(HEADER_HOSTNAME, header->adds_features)) { 2434 bitmap_zero(header->adds_features, HEADER_FEAT_BITS); 2435 set_bit(HEADER_BUILD_ID, header->adds_features); 2436 } 2437 } 2438 2439 memcpy(&ph->adds_features, &header->adds_features, 2440 sizeof(ph->adds_features)); 2441 2442 ph->event_offset = header->event_types.offset; 2443 ph->event_size = header->event_types.size; 2444 ph->data_offset = header->data.offset; 2445 ph->data_size = header->data.size; 2446 return 0; 2447 } 2448 2449 static int perf_file_section__process(struct perf_file_section *section, 2450 struct perf_header *ph, 2451 int feat, int fd, void *data) 2452 { 2453 if (lseek(fd, section->offset, SEEK_SET) == (off_t)-1) { 2454 pr_debug("Failed to lseek to %" PRIu64 " offset for feature " 2455 "%d, continuing...\n", section->offset, feat); 2456 return 0; 2457 } 2458 2459 if (feat >= HEADER_LAST_FEATURE) { 2460 pr_debug("unknown feature %d, continuing...\n", feat); 2461 return 0; 2462 } 2463 2464 if (!feat_ops[feat].process) 2465 return 0; 2466 2467 return feat_ops[feat].process(section, ph, fd, data); 2468 } 2469 2470 static int perf_file_header__read_pipe(struct perf_pipe_file_header *header, 2471 struct perf_header *ph, int fd, 2472 bool repipe) 2473 { 2474 int ret; 2475 2476 ret = readn(fd, header, sizeof(*header)); 2477 if (ret <= 0) 2478 return -1; 2479 2480 if (check_magic_endian(header->magic, header->size, true, ph) < 0) { 2481 pr_debug("endian/magic failed\n"); 2482 return -1; 2483 } 2484 2485 if (ph->needs_swap) 2486 header->size = bswap_64(header->size); 2487 2488 if (repipe && do_write(STDOUT_FILENO, header, sizeof(*header)) < 0) 2489 return -1; 2490 2491 return 0; 2492 } 2493 2494 static int perf_header__read_pipe(struct perf_session *session, int fd) 2495 { 2496 struct perf_header *header = &session->header; 2497 struct perf_pipe_file_header f_header; 2498 2499 if (perf_file_header__read_pipe(&f_header, header, fd, 2500 session->repipe) < 0) { 2501 pr_debug("incompatible file format\n"); 2502 return -EINVAL; 2503 } 2504 2505 session->fd = fd; 2506 2507 return 0; 2508 } 2509 2510 static int read_attr(int fd, struct perf_header *ph, 2511 struct perf_file_attr *f_attr) 2512 { 2513 struct perf_event_attr *attr = &f_attr->attr; 2514 size_t sz, left; 2515 size_t our_sz = sizeof(f_attr->attr); 2516 int ret; 2517 2518 memset(f_attr, 0, sizeof(*f_attr)); 2519 2520 /* read minimal guaranteed structure */ 2521 ret = readn(fd, attr, PERF_ATTR_SIZE_VER0); 2522 if (ret <= 0) { 2523 pr_debug("cannot read %d bytes of header attr\n", 2524 PERF_ATTR_SIZE_VER0); 2525 return -1; 2526 } 2527 2528 /* on file perf_event_attr size */ 2529 sz = attr->size; 2530 2531 if (ph->needs_swap) 2532 sz = bswap_32(sz); 2533 2534 if (sz == 0) { 2535 /* assume ABI0 */ 2536 sz = PERF_ATTR_SIZE_VER0; 2537 } else if (sz > our_sz) { 2538 pr_debug("file uses a more recent and unsupported ABI" 2539 " (%zu bytes extra)\n", sz - our_sz); 2540 return -1; 2541 } 2542 /* what we have not yet read and that we know about */ 2543 left = sz - PERF_ATTR_SIZE_VER0; 2544 if (left) { 2545 void *ptr = attr; 2546 ptr += PERF_ATTR_SIZE_VER0; 2547 2548 ret = readn(fd, ptr, left); 2549 } 2550 /* read perf_file_section, ids are read in caller */ 2551 ret = readn(fd, &f_attr->ids, sizeof(f_attr->ids)); 2552 2553 return ret <= 0 ? -1 : 0; 2554 } 2555 2556 static int perf_evsel__prepare_tracepoint_event(struct perf_evsel *evsel, 2557 struct pevent *pevent) 2558 { 2559 struct event_format *event; 2560 char bf[128]; 2561 2562 /* already prepared */ 2563 if (evsel->tp_format) 2564 return 0; 2565 2566 event = pevent_find_event(pevent, evsel->attr.config); 2567 if (event == NULL) 2568 return -1; 2569 2570 if (!evsel->name) { 2571 snprintf(bf, sizeof(bf), "%s:%s", event->system, event->name); 2572 evsel->name = strdup(bf); 2573 if (evsel->name == NULL) 2574 return -1; 2575 } 2576 2577 evsel->tp_format = event; 2578 return 0; 2579 } 2580 2581 static int perf_evlist__prepare_tracepoint_events(struct perf_evlist *evlist, 2582 struct pevent *pevent) 2583 { 2584 struct perf_evsel *pos; 2585 2586 list_for_each_entry(pos, &evlist->entries, node) { 2587 if (pos->attr.type == PERF_TYPE_TRACEPOINT && 2588 perf_evsel__prepare_tracepoint_event(pos, pevent)) 2589 return -1; 2590 } 2591 2592 return 0; 2593 } 2594 2595 int perf_session__read_header(struct perf_session *session, int fd) 2596 { 2597 struct perf_header *header = &session->header; 2598 struct perf_file_header f_header; 2599 struct perf_file_attr f_attr; 2600 u64 f_id; 2601 int nr_attrs, nr_ids, i, j; 2602 2603 session->evlist = perf_evlist__new(NULL, NULL); 2604 if (session->evlist == NULL) 2605 return -ENOMEM; 2606 2607 if (session->fd_pipe) 2608 return perf_header__read_pipe(session, fd); 2609 2610 if (perf_file_header__read(&f_header, header, fd) < 0) 2611 return -EINVAL; 2612 2613 nr_attrs = f_header.attrs.size / f_header.attr_size; 2614 lseek(fd, f_header.attrs.offset, SEEK_SET); 2615 2616 for (i = 0; i < nr_attrs; i++) { 2617 struct perf_evsel *evsel; 2618 off_t tmp; 2619 2620 if (read_attr(fd, header, &f_attr) < 0) 2621 goto out_errno; 2622 2623 if (header->needs_swap) 2624 perf_event__attr_swap(&f_attr.attr); 2625 2626 tmp = lseek(fd, 0, SEEK_CUR); 2627 evsel = perf_evsel__new(&f_attr.attr, i); 2628 2629 if (evsel == NULL) 2630 goto out_delete_evlist; 2631 2632 evsel->needs_swap = header->needs_swap; 2633 /* 2634 * Do it before so that if perf_evsel__alloc_id fails, this 2635 * entry gets purged too at perf_evlist__delete(). 2636 */ 2637 perf_evlist__add(session->evlist, evsel); 2638 2639 nr_ids = f_attr.ids.size / sizeof(u64); 2640 /* 2641 * We don't have the cpu and thread maps on the header, so 2642 * for allocating the perf_sample_id table we fake 1 cpu and 2643 * hattr->ids threads. 2644 */ 2645 if (perf_evsel__alloc_id(evsel, 1, nr_ids)) 2646 goto out_delete_evlist; 2647 2648 lseek(fd, f_attr.ids.offset, SEEK_SET); 2649 2650 for (j = 0; j < nr_ids; j++) { 2651 if (perf_header__getbuffer64(header, fd, &f_id, sizeof(f_id))) 2652 goto out_errno; 2653 2654 perf_evlist__id_add(session->evlist, evsel, 0, j, f_id); 2655 } 2656 2657 lseek(fd, tmp, SEEK_SET); 2658 } 2659 2660 symbol_conf.nr_events = nr_attrs; 2661 2662 if (f_header.event_types.size) { 2663 lseek(fd, f_header.event_types.offset, SEEK_SET); 2664 trace_events = malloc(f_header.event_types.size); 2665 if (trace_events == NULL) 2666 return -ENOMEM; 2667 if (perf_header__getbuffer64(header, fd, trace_events, 2668 f_header.event_types.size)) 2669 goto out_errno; 2670 trace_event_count = f_header.event_types.size / sizeof(struct perf_trace_event_type); 2671 } 2672 2673 perf_header__process_sections(header, fd, &session->pevent, 2674 perf_file_section__process); 2675 2676 lseek(fd, header->data_offset, SEEK_SET); 2677 2678 if (perf_evlist__prepare_tracepoint_events(session->evlist, 2679 session->pevent)) 2680 goto out_delete_evlist; 2681 2682 header->frozen = 1; 2683 return 0; 2684 out_errno: 2685 return -errno; 2686 2687 out_delete_evlist: 2688 perf_evlist__delete(session->evlist); 2689 session->evlist = NULL; 2690 return -ENOMEM; 2691 } 2692 2693 int perf_event__synthesize_attr(struct perf_tool *tool, 2694 struct perf_event_attr *attr, u32 ids, u64 *id, 2695 perf_event__handler_t process) 2696 { 2697 union perf_event *ev; 2698 size_t size; 2699 int err; 2700 2701 size = sizeof(struct perf_event_attr); 2702 size = PERF_ALIGN(size, sizeof(u64)); 2703 size += sizeof(struct perf_event_header); 2704 size += ids * sizeof(u64); 2705 2706 ev = malloc(size); 2707 2708 if (ev == NULL) 2709 return -ENOMEM; 2710 2711 ev->attr.attr = *attr; 2712 memcpy(ev->attr.id, id, ids * sizeof(u64)); 2713 2714 ev->attr.header.type = PERF_RECORD_HEADER_ATTR; 2715 ev->attr.header.size = (u16)size; 2716 2717 if (ev->attr.header.size == size) 2718 err = process(tool, ev, NULL, NULL); 2719 else 2720 err = -E2BIG; 2721 2722 free(ev); 2723 2724 return err; 2725 } 2726 2727 int perf_event__synthesize_attrs(struct perf_tool *tool, 2728 struct perf_session *session, 2729 perf_event__handler_t process) 2730 { 2731 struct perf_evsel *evsel; 2732 int err = 0; 2733 2734 list_for_each_entry(evsel, &session->evlist->entries, node) { 2735 err = perf_event__synthesize_attr(tool, &evsel->attr, evsel->ids, 2736 evsel->id, process); 2737 if (err) { 2738 pr_debug("failed to create perf header attribute\n"); 2739 return err; 2740 } 2741 } 2742 2743 return err; 2744 } 2745 2746 int perf_event__process_attr(union perf_event *event, 2747 struct perf_evlist **pevlist) 2748 { 2749 u32 i, ids, n_ids; 2750 struct perf_evsel *evsel; 2751 struct perf_evlist *evlist = *pevlist; 2752 2753 if (evlist == NULL) { 2754 *pevlist = evlist = perf_evlist__new(NULL, NULL); 2755 if (evlist == NULL) 2756 return -ENOMEM; 2757 } 2758 2759 evsel = perf_evsel__new(&event->attr.attr, evlist->nr_entries); 2760 if (evsel == NULL) 2761 return -ENOMEM; 2762 2763 perf_evlist__add(evlist, evsel); 2764 2765 ids = event->header.size; 2766 ids -= (void *)&event->attr.id - (void *)event; 2767 n_ids = ids / sizeof(u64); 2768 /* 2769 * We don't have the cpu and thread maps on the header, so 2770 * for allocating the perf_sample_id table we fake 1 cpu and 2771 * hattr->ids threads. 2772 */ 2773 if (perf_evsel__alloc_id(evsel, 1, n_ids)) 2774 return -ENOMEM; 2775 2776 for (i = 0; i < n_ids; i++) { 2777 perf_evlist__id_add(evlist, evsel, 0, i, event->attr.id[i]); 2778 } 2779 2780 return 0; 2781 } 2782 2783 int perf_event__synthesize_event_type(struct perf_tool *tool, 2784 u64 event_id, char *name, 2785 perf_event__handler_t process, 2786 struct machine *machine) 2787 { 2788 union perf_event ev; 2789 size_t size = 0; 2790 int err = 0; 2791 2792 memset(&ev, 0, sizeof(ev)); 2793 2794 ev.event_type.event_type.event_id = event_id; 2795 memset(ev.event_type.event_type.name, 0, MAX_EVENT_NAME); 2796 strncpy(ev.event_type.event_type.name, name, MAX_EVENT_NAME - 1); 2797 2798 ev.event_type.header.type = PERF_RECORD_HEADER_EVENT_TYPE; 2799 size = strlen(ev.event_type.event_type.name); 2800 size = PERF_ALIGN(size, sizeof(u64)); 2801 ev.event_type.header.size = sizeof(ev.event_type) - 2802 (sizeof(ev.event_type.event_type.name) - size); 2803 2804 err = process(tool, &ev, NULL, machine); 2805 2806 return err; 2807 } 2808 2809 int perf_event__synthesize_event_types(struct perf_tool *tool, 2810 perf_event__handler_t process, 2811 struct machine *machine) 2812 { 2813 struct perf_trace_event_type *type; 2814 int i, err = 0; 2815 2816 for (i = 0; i < trace_event_count; i++) { 2817 type = &trace_events[i]; 2818 2819 err = perf_event__synthesize_event_type(tool, type->event_id, 2820 type->name, process, 2821 machine); 2822 if (err) { 2823 pr_debug("failed to create perf header event type\n"); 2824 return err; 2825 } 2826 } 2827 2828 return err; 2829 } 2830 2831 int perf_event__process_event_type(struct perf_tool *tool __maybe_unused, 2832 union perf_event *event) 2833 { 2834 if (perf_header__push_event(event->event_type.event_type.event_id, 2835 event->event_type.event_type.name) < 0) 2836 return -ENOMEM; 2837 2838 return 0; 2839 } 2840 2841 int perf_event__synthesize_tracing_data(struct perf_tool *tool, int fd, 2842 struct perf_evlist *evlist, 2843 perf_event__handler_t process) 2844 { 2845 union perf_event ev; 2846 struct tracing_data *tdata; 2847 ssize_t size = 0, aligned_size = 0, padding; 2848 int err __maybe_unused = 0; 2849 2850 /* 2851 * We are going to store the size of the data followed 2852 * by the data contents. Since the fd descriptor is a pipe, 2853 * we cannot seek back to store the size of the data once 2854 * we know it. Instead we: 2855 * 2856 * - write the tracing data to the temp file 2857 * - get/write the data size to pipe 2858 * - write the tracing data from the temp file 2859 * to the pipe 2860 */ 2861 tdata = tracing_data_get(&evlist->entries, fd, true); 2862 if (!tdata) 2863 return -1; 2864 2865 memset(&ev, 0, sizeof(ev)); 2866 2867 ev.tracing_data.header.type = PERF_RECORD_HEADER_TRACING_DATA; 2868 size = tdata->size; 2869 aligned_size = PERF_ALIGN(size, sizeof(u64)); 2870 padding = aligned_size - size; 2871 ev.tracing_data.header.size = sizeof(ev.tracing_data); 2872 ev.tracing_data.size = aligned_size; 2873 2874 process(tool, &ev, NULL, NULL); 2875 2876 /* 2877 * The put function will copy all the tracing data 2878 * stored in temp file to the pipe. 2879 */ 2880 tracing_data_put(tdata); 2881 2882 write_padded(fd, NULL, 0, padding); 2883 2884 return aligned_size; 2885 } 2886 2887 int perf_event__process_tracing_data(union perf_event *event, 2888 struct perf_session *session) 2889 { 2890 ssize_t size_read, padding, size = event->tracing_data.size; 2891 off_t offset = lseek(session->fd, 0, SEEK_CUR); 2892 char buf[BUFSIZ]; 2893 2894 /* setup for reading amidst mmap */ 2895 lseek(session->fd, offset + sizeof(struct tracing_data_event), 2896 SEEK_SET); 2897 2898 size_read = trace_report(session->fd, &session->pevent, 2899 session->repipe); 2900 padding = PERF_ALIGN(size_read, sizeof(u64)) - size_read; 2901 2902 if (read(session->fd, buf, padding) < 0) 2903 die("reading input file"); 2904 if (session->repipe) { 2905 int retw = write(STDOUT_FILENO, buf, padding); 2906 if (retw <= 0 || retw != padding) 2907 die("repiping tracing data padding"); 2908 } 2909 2910 if (size_read + padding != size) 2911 die("tracing data size mismatch"); 2912 2913 perf_evlist__prepare_tracepoint_events(session->evlist, 2914 session->pevent); 2915 2916 return size_read + padding; 2917 } 2918 2919 int perf_event__synthesize_build_id(struct perf_tool *tool, 2920 struct dso *pos, u16 misc, 2921 perf_event__handler_t process, 2922 struct machine *machine) 2923 { 2924 union perf_event ev; 2925 size_t len; 2926 int err = 0; 2927 2928 if (!pos->hit) 2929 return err; 2930 2931 memset(&ev, 0, sizeof(ev)); 2932 2933 len = pos->long_name_len + 1; 2934 len = PERF_ALIGN(len, NAME_ALIGN); 2935 memcpy(&ev.build_id.build_id, pos->build_id, sizeof(pos->build_id)); 2936 ev.build_id.header.type = PERF_RECORD_HEADER_BUILD_ID; 2937 ev.build_id.header.misc = misc; 2938 ev.build_id.pid = machine->pid; 2939 ev.build_id.header.size = sizeof(ev.build_id) + len; 2940 memcpy(&ev.build_id.filename, pos->long_name, pos->long_name_len); 2941 2942 err = process(tool, &ev, NULL, machine); 2943 2944 return err; 2945 } 2946 2947 int perf_event__process_build_id(struct perf_tool *tool __maybe_unused, 2948 union perf_event *event, 2949 struct perf_session *session) 2950 { 2951 __event_process_build_id(&event->build_id, 2952 event->build_id.filename, 2953 session); 2954 return 0; 2955 } 2956 2957 void disable_buildid_cache(void) 2958 { 2959 no_buildid_cache = true; 2960 } 2961