kernel/tools/perf/util/header.c

   1 #include "util.h"
   2 #include <sys/types.h>
   3 #include <byteswap.h>
   4 #include <unistd.h>
   5 #include <stdio.h>
   6 #include <stdlib.h>
   7 #include <linux/list.h>
   8 #include <linux/kernel.h>
   9 #include <linux/bitops.h>
  10 #include <sys/utsname.h>
  11
  12 #include "evlist.h"
  13 #include "evsel.h"
  14 #include "header.h"
  15 #include "../perf.h"
  16 #include "trace-event.h"
  17 #include "session.h"
  18 #include "symbol.h"
  19 #include "debug.h"
  20 #include "cpumap.h"
  21 #include "pmu.h"
  22 #include "vdso.h"
  23 #include "strbuf.h"
  24 #include "build-id.h"
  25 #include "data.h"
  26
  27 static u32 header_argc;
  28 static const char **header_argv;
  29
  30 /*
  31  * magic2 = "PERFILE2"
  32  * must be a numerical value to let the endianness
  33  * determine the memory layout. That way we are able
  34  * to detect endianness when reading the perf.data file
  35  * back.
  36  *
  37  * we check for legacy (PERFFILE) format.
  38  */
  39 static const char *__perf_magic1 = "PERFFILE";
  40 static const u64 __perf_magic2    = 0x32454c4946524550ULL;
  41 static const u64 __perf_magic2_sw = 0x50455246494c4532ULL;
  42
  43 #define PERF_MAGIC      __perf_magic2
  44
  45 struct perf_file_attr {
  46         struct perf_event_attr  attr;
  47         struct perf_file_section        ids;
  48 };
  49
  50 void perf_header__set_feat(struct perf_header *header, int feat)
  51 {
  52         set_bit(feat, header->adds_features);
  53 }
  54
  55 void perf_header__clear_feat(struct perf_header *header, int feat)
  56 {
  57         clear_bit(feat, header->adds_features);
  58 }
  59
  60 bool perf_header__has_feat(const struct perf_header *header, int feat)
  61 {
  62         return test_bit(feat, header->adds_features);
  63 }
  64
  65 static int do_write(int fd, const void *buf, size_t size)
  66 {
  67         while (size) {
  68                 int ret = write(fd, buf, size);
  69
  70                 if (ret < 0)
  71                         return -errno;
  72
  73                 size -= ret;
  74                 buf += ret;
  75         }
  76
  77         return 0;
  78 }
  79
  80 int write_padded(int fd, const void *bf, size_t count, size_t count_aligned)
  81 {
  82         static const char zero_buf[NAME_ALIGN];
  83         int err = do_write(fd, bf, count);
  84
  85         if (!err)
  86                 err = do_write(fd, zero_buf, count_aligned - count);
  87
  88         return err;
  89 }
  90
  91 static int do_write_string(int fd, const char *str)
  92 {
  93         u32 len, olen;
  94         int ret;
  95
  96         olen = strlen(str) + 1;
  97         len = PERF_ALIGN(olen, NAME_ALIGN);
  98
  99         /* write len, incl. \0 */
 100         ret = do_write(fd, &len, sizeof(len));
 101         if (ret < 0)
 102                 return ret;
 103
 104         return write_padded(fd, str, olen, len);
 105 }
 106
 107 static char *do_read_string(int fd, struct perf_header *ph)
 108 {
 109         ssize_t sz, ret;
 110         u32 len;
 111         char *buf;
 112
 113         sz = readn(fd, &len, sizeof(len));
 114         if (sz < (ssize_t)sizeof(len))
 115                 return NULL;
 116
 117         if (ph->needs_swap)
 118                 len = bswap_32(len);
 119
 120         buf = malloc(len);
 121         if (!buf)
 122                 return NULL;
 123
 124         ret = readn(fd, buf, len);
 125         if (ret == (ssize_t)len) {
 126                 /*
 127                  * strings are padded by zeroes
 128                  * thus the actual strlen of buf
 129                  * may be less than len
 130                  */
 131                 return buf;
 132         }
 133
 134         free(buf);
 135         return NULL;
 136 }
 137
 138 int
 139 perf_header__set_cmdline(int argc, const char **argv)
 140 {
 141         int i;
 142
 143         /*
 144          * If header_argv has already been set, do not override it.
 145          * This allows a command to set the cmdline, parse args and
 146          * then call another builtin function that implements a
 147          * command -- e.g, cmd_kvm calling cmd_record.
 148          */
 149         if (header_argv)
 150                 return 0;
 151
 152         header_argc = (u32)argc;
 153
 154         /* do not include NULL termination */
 155         header_argv = calloc(argc, sizeof(char *));
 156         if (!header_argv)
 157                 return -ENOMEM;
 158
 159         /*
 160          * must copy argv contents because it gets moved
 161          * around during option parsing
 162          */
 163         for (i = 0; i < argc ; i++)
 164                 header_argv[i] = argv[i];
 165
 166         return 0;
 167 }
 168
 169 static int write_tracing_data(int fd, struct perf_header *h __maybe_unused,
 170                             struct perf_evlist *evlist)
 171 {
 172         return read_tracing_data(fd, &evlist->entries);
 173 }
 174
 175
 176 static int write_build_id(int fd, struct perf_header *h,
 177                           struct perf_evlist *evlist __maybe_unused)
 178 {
 179         struct perf_session *session;
 180         int err;
 181
 182         session = container_of(h, struct perf_session, header);
 183
 184         if (!perf_session__read_build_ids(session, true))
 185                 return -1;
 186
 187         err = perf_session__write_buildid_table(session, fd);
 188         if (err < 0) {
 189                 pr_debug("failed to write buildid table\n");
 190                 return err;
 191         }
 192         perf_session__cache_build_ids(session);
 193
 194         return 0;
 195 }
 196
 197 static int write_hostname(int fd, struct perf_header *h __maybe_unused,
 198                           struct perf_evlist *evlist __maybe_unused)
 199 {
 200         struct utsname uts;
 201         int ret;
 202
 203         ret = uname(&uts);
 204         if (ret < 0)
 205                 return -1;
 206
 207         return do_write_string(fd, uts.nodename);
 208 }
 209
 210 static int write_osrelease(int fd, struct perf_header *h __maybe_unused,
 211                            struct perf_evlist *evlist __maybe_unused)
 212 {
 213         struct utsname uts;
 214         int ret;
 215
 216         ret = uname(&uts);
 217         if (ret < 0)
 218                 return -1;
 219
 220         return do_write_string(fd, uts.release);
 221 }
 222
 223 static int write_arch(int fd, struct perf_header *h __maybe_unused,
 224                       struct perf_evlist *evlist __maybe_unused)
 225 {
 226         struct utsname uts;
 227         int ret;
 228
 229         ret = uname(&uts);
 230         if (ret < 0)
 231                 return -1;
 232
 233         return do_write_string(fd, uts.machine);
 234 }
 235
 236 static int write_version(int fd, struct perf_header *h __maybe_unused,
 237                          struct perf_evlist *evlist __maybe_unused)
 238 {
 239         return do_write_string(fd, perf_version_string);
 240 }
 241
 242 static int __write_cpudesc(int fd, const char *cpuinfo_proc)
 243 {
 244         FILE *file;
 245         char *buf = NULL;
 246         char *s, *p;
 247         const char *search = cpuinfo_proc;
 248         size_t len = 0;
 249         int ret = -1;
 250
 251         if (!search)
 252                 return -1;
 253
 254         file = fopen("/proc/cpuinfo", "r");
 255         if (!file)
 256                 return -1;
 257
 258         while (getline(&buf, &len, file) > 0) {
 259                 ret = strncmp(buf, search, strlen(search));
 260                 if (!ret)
 261                         break;
 262         }
 263
 264         if (ret) {
 265                 ret = -1;
 266                 goto done;
 267         }
 268
 269         s = buf;
 270
 271         p = strchr(buf, ':');
 272         if (p && *(p+1) == ' ' && *(p+2))
 273                 s = p + 2;
 274         p = strchr(s, '\n');
 275         if (p)
 276                 *p = '\0';
 277
 278         /* squash extra space characters (branding string) */
 279         p = s;
 280         while (*p) {
 281                 if (isspace(*p)) {
 282                         char *r = p + 1;
 283                         char *q = r;
 284                         *p = ' ';
 285                         while (*q && isspace(*q))
 286                                 q++;
 287                         if (q != (p+1))
 288                                 while ((*r++ = *q++));
 289                 }
 290                 p++;
 291         }
 292         ret = do_write_string(fd, s);
 293 done:
 294         free(buf);
 295         fclose(file);
 296         return ret;
 297 }
 298
 299 static int write_cpudesc(int fd, struct perf_header *h __maybe_unused,
 300                        struct perf_evlist *evlist __maybe_unused)
 301 {
 302 #ifndef CPUINFO_PROC
 303 #define CPUINFO_PROC {"model name", }
 304 #endif
 305         const char *cpuinfo_procs[] = CPUINFO_PROC;
 306         unsigned int i;
 307
 308         for (i = 0; i < ARRAY_SIZE(cpuinfo_procs); i++) {
 309                 int ret;
 310                 ret = __write_cpudesc(fd, cpuinfo_procs[i]);
 311                 if (ret >= 0)
 312                         return ret;
 313         }
 314         return -1;
 315 }
 316
 317
 318 static int write_nrcpus(int fd, struct perf_header *h __maybe_unused,
 319                         struct perf_evlist *evlist __maybe_unused)
 320 {
 321         long nr;
 322         u32 nrc, nra;
 323         int ret;
 324
 325         nr = sysconf(_SC_NPROCESSORS_CONF);
 326         if (nr < 0)
 327                 return -1;
 328
 329         nrc = (u32)(nr & UINT_MAX);
 330
 331         nr = sysconf(_SC_NPROCESSORS_ONLN);
 332         if (nr < 0)
 333                 return -1;
 334
 335         nra = (u32)(nr & UINT_MAX);
 336
 337         ret = do_write(fd, &nrc, sizeof(nrc));
 338         if (ret < 0)
 339                 return ret;
 340
 341         return do_write(fd, &nra, sizeof(nra));
 342 }
 343
 344 static int write_event_desc(int fd, struct perf_header *h __maybe_unused,
 345                             struct perf_evlist *evlist)
 346 {
 347         struct perf_evsel *evsel;
 348         u32 nre, nri, sz;
 349         int ret;
 350
 351         nre = evlist->nr_entries;
 352
 353         /*
 354          * write number of events
 355          */
 356         ret = do_write(fd, &nre, sizeof(nre));
 357         if (ret < 0)
 358                 return ret;
 359
 360         /*
 361          * size of perf_event_attr struct
 362          */
 363         sz = (u32)sizeof(evsel->attr);
 364         ret = do_write(fd, &sz, sizeof(sz));
 365         if (ret < 0)
 366                 return ret;
 367
 368         evlist__for_each(evlist, evsel) {
 369                 ret = do_write(fd, &evsel->attr, sz);
 370                 if (ret < 0)
 371                         return ret;
 372                 /*
 373                  * write number of unique id per event
 374                  * there is one id per instance of an event
 375                  *
 376                  * copy into an nri to be independent of the
 377                  * type of ids,
 378                  */
 379                 nri = evsel->ids;
 380                 ret = do_write(fd, &nri, sizeof(nri));
 381                 if (ret < 0)
 382                         return ret;
 383
 384                 /*
 385                  * write event string as passed on cmdline
 386                  */
 387                 ret = do_write_string(fd, perf_evsel__name(evsel));
 388                 if (ret < 0)
 389                         return ret;
 390                 /*
 391                  * write unique ids for this event
 392                  */
 393                 ret = do_write(fd, evsel->id, evsel->ids * sizeof(u64));
 394                 if (ret < 0)
 395                         return ret;
 396         }
 397         return 0;
 398 }
 399
 400 static int write_cmdline(int fd, struct perf_header *h __maybe_unused,
 401                          struct perf_evlist *evlist __maybe_unused)
 402 {
 403         char buf[MAXPATHLEN];
 404         char proc[32];
 405         u32 i, n;
 406         int ret;
 407
 408         /*
 409          * actual atual path to perf binary
 410          */
 411         sprintf(proc, "/proc/%d/exe", getpid());
 412         ret = readlink(proc, buf, sizeof(buf));
 413         if (ret <= 0)
 414                 return -1;
 415
 416         /* readlink() does not add null termination */
 417         buf[ret] = '\0';
 418
 419         /* account for binary path */
 420         n = header_argc + 1;
 421
 422         ret = do_write(fd, &n, sizeof(n));
 423         if (ret < 0)
 424                 return ret;
 425
 426         ret = do_write_string(fd, buf);
 427         if (ret < 0)
 428                 return ret;
 429
 430         for (i = 0 ; i < header_argc; i++) {
 431                 ret = do_write_string(fd, header_argv[i]);
 432                 if (ret < 0)
 433                         return ret;
 434         }
 435         return 0;
 436 }
 437
 438 #define CORE_SIB_FMT \
 439         "/sys/devices/system/cpu/cpu%d/topology/core_siblings_list"
 440 #define THRD_SIB_FMT \
 441         "/sys/devices/system/cpu/cpu%d/topology/thread_siblings_list"
 442
 443 struct cpu_topo {
 444         u32 core_sib;
 445         u32 thread_sib;
 446         char **core_siblings;
 447         char **thread_siblings;
 448 };
 449
 450 static int build_cpu_topo(struct cpu_topo *tp, int cpu)
 451 {
 452         FILE *fp;
 453         char filename[MAXPATHLEN];
 454         char *buf = NULL, *p;
 455         size_t len = 0;
 456         ssize_t sret;
 457         u32 i = 0;
 458         int ret = -1;
 459
 460         sprintf(filename, CORE_SIB_FMT, cpu);
 461         fp = fopen(filename, "r");
 462         if (!fp)
 463                 goto try_threads;
 464
 465         sret = getline(&buf, &len, fp);
 466         fclose(fp);
 467         if (sret <= 0)
 468                 goto try_threads;
 469
 470         p = strchr(buf, '\n');
 471         if (p)
 472                 *p = '\0';
 473
 474         for (i = 0; i < tp->core_sib; i++) {
 475                 if (!strcmp(buf, tp->core_siblings[i]))
 476                         break;
 477         }
 478         if (i == tp->core_sib) {
 479                 tp->core_siblings[i] = buf;
 480                 tp->core_sib++;
 481                 buf = NULL;
 482                 len = 0;
 483         }
 484         ret = 0;
 485
 486 try_threads:
 487         sprintf(filename, THRD_SIB_FMT, cpu);
 488         fp = fopen(filename, "r");
 489         if (!fp)
 490                 goto done;
 491
 492         if (getline(&buf, &len, fp) <= 0)
 493                 goto done;
 494
 495         p = strchr(buf, '\n');
 496         if (p)
 497                 *p = '\0';
 498
 499         for (i = 0; i < tp->thread_sib; i++) {
 500                 if (!strcmp(buf, tp->thread_siblings[i]))
 501                         break;
 502         }
 503         if (i == tp->thread_sib) {
 504                 tp->thread_siblings[i] = buf;
 505                 tp->thread_sib++;
 506                 buf = NULL;
 507         }
 508         ret = 0;
 509 done:
 510         if(fp)
 511                 fclose(fp);
 512         free(buf);
 513         return ret;
 514 }
 515
 516 static void free_cpu_topo(struct cpu_topo *tp)
 517 {
 518         u32 i;
 519
 520         if (!tp)
 521                 return;
 522
 523         for (i = 0 ; i < tp->core_sib; i++)
 524                 zfree(&tp->core_siblings[i]);
 525
 526         for (i = 0 ; i < tp->thread_sib; i++)
 527                 zfree(&tp->thread_siblings[i]);
 528
 529         free(tp);
 530 }
 531
 532 static struct cpu_topo *build_cpu_topology(void)
 533 {
 534         struct cpu_topo *tp;
 535         void *addr;
 536         u32 nr, i;
 537         size_t sz;
 538         long ncpus;
 539         int ret = -1;
 540
 541         ncpus = sysconf(_SC_NPROCESSORS_CONF);
 542         if (ncpus < 0)
 543                 return NULL;
 544
 545         nr = (u32)(ncpus & UINT_MAX);
 546
 547         sz = nr * sizeof(char *);
 548
 549         addr = calloc(1, sizeof(*tp) + 2 * sz);
 550         if (!addr)
 551                 return NULL;
 552
 553         tp = addr;
 554
 555         addr += sizeof(*tp);
 556         tp->core_siblings = addr;
 557         addr += sz;
 558         tp->thread_siblings = addr;
 559
 560         for (i = 0; i < nr; i++) {
 561                 ret = build_cpu_topo(tp, i);
 562                 if (ret < 0)
 563                         break;
 564         }
 565         if (ret) {
 566                 free_cpu_topo(tp);
 567                 tp = NULL;
 568         }
 569         return tp;
 570 }
 571
 572 static int write_cpu_topology(int fd, struct perf_header *h __maybe_unused,
 573                           struct perf_evlist *evlist __maybe_unused)
 574 {
 575         struct cpu_topo *tp;
 576         u32 i;
 577         int ret;
 578
 579         tp = build_cpu_topology();
 580         if (!tp)
 581                 return -1;
 582
 583         ret = do_write(fd, &tp->core_sib, sizeof(tp->core_sib));
 584         if (ret < 0)
 585                 goto done;
 586
 587         for (i = 0; i < tp->core_sib; i++) {
 588                 ret = do_write_string(fd, tp->core_siblings[i]);
 589                 if (ret < 0)
 590                         goto done;
 591         }
 592         ret = do_write(fd, &tp->thread_sib, sizeof(tp->thread_sib));
 593         if (ret < 0)
 594                 goto done;
 595
 596         for (i = 0; i < tp->thread_sib; i++) {
 597                 ret = do_write_string(fd, tp->thread_siblings[i]);
 598                 if (ret < 0)
 599                         break;
 600         }
 601 done:
 602         free_cpu_topo(tp);
 603         return ret;
 604 }
 605
 606
 607
 608 static int write_total_mem(int fd, struct perf_header *h __maybe_unused,
 609                           struct perf_evlist *evlist __maybe_unused)
 610 {
 611         char *buf = NULL;
 612         FILE *fp;
 613         size_t len = 0;
 614         int ret = -1, n;
 615         uint64_t mem;
 616
 617         fp = fopen("/proc/meminfo", "r");
 618         if (!fp)
 619                 return -1;
 620
 621         while (getline(&buf, &len, fp) > 0) {
 622                 ret = strncmp(buf, "MemTotal:", 9);
 623                 if (!ret)
 624                         break;
 625         }
 626         if (!ret) {
 627                 n = sscanf(buf, "%*s %"PRIu64, &mem);
 628                 if (n == 1)
 629                         ret = do_write(fd, &mem, sizeof(mem));
 630         } else
 631                 ret = -1;
 632         free(buf);
 633         fclose(fp);
 634         return ret;
 635 }
 636
 637 static int write_topo_node(int fd, int node)
 638 {
 639         char str[MAXPATHLEN];
 640         char field[32];
 641         char *buf = NULL, *p;
 642         size_t len = 0;
 643         FILE *fp;
 644         u64 mem_total, mem_free, mem;
 645         int ret = -1;
 646
 647         sprintf(str, "/sys/devices/system/node/node%d/meminfo", node);
 648         fp = fopen(str, "r");
 649         if (!fp)
 650                 return -1;
 651
 652         while (getline(&buf, &len, fp) > 0) {
 653                 /* skip over invalid lines */
 654                 if (!strchr(buf, ':'))
 655                         continue;
 656                 if (sscanf(buf, "%*s %*d %31s %"PRIu64, field, &mem) != 2)
 657                         goto done;
 658                 if (!strcmp(field, "MemTotal:"))
 659                         mem_total = mem;
 660                 if (!strcmp(field, "MemFree:"))
 661                         mem_free = mem;
 662         }
 663
 664         fclose(fp);
 665         fp = NULL;
 666
 667         ret = do_write(fd, &mem_total, sizeof(u64));
 668         if (ret)
 669                 goto done;
 670
 671         ret = do_write(fd, &mem_free, sizeof(u64));
 672         if (ret)
 673                 goto done;
 674
 675         ret = -1;
 676         sprintf(str, "/sys/devices/system/node/node%d/cpulist", node);
 677
 678         fp = fopen(str, "r");
 679         if (!fp)
 680                 goto done;
 681
 682         if (getline(&buf, &len, fp) <= 0)
 683                 goto done;
 684
 685         p = strchr(buf, '\n');
 686         if (p)
 687                 *p = '\0';
 688
 689         ret = do_write_string(fd, buf);
 690 done:
 691         free(buf);
 692         if (fp)
 693                 fclose(fp);
 694         return ret;
 695 }
 696
 697 static int write_numa_topology(int fd, struct perf_header *h __maybe_unused,
 698                           struct perf_evlist *evlist __maybe_unused)
 699 {
 700         char *buf = NULL;
 701         size_t len = 0;
 702         FILE *fp;
 703         struct cpu_map *node_map = NULL;
 704         char *c;
 705         u32 nr, i, j;
 706         int ret = -1;
 707
 708         fp = fopen("/sys/devices/system/node/online", "r");
 709         if (!fp)
 710                 return -1;
 711
 712         if (getline(&buf, &len, fp) <= 0)
 713                 goto done;
 714
 715         c = strchr(buf, '\n');
 716         if (c)
 717                 *c = '\0';
 718
 719         node_map = cpu_map__new(buf);
 720         if (!node_map)
 721                 goto done;
 722
 723         nr = (u32)node_map->nr;
 724
 725         ret = do_write(fd, &nr, sizeof(nr));
 726         if (ret < 0)
 727                 goto done;
 728
 729         for (i = 0; i < nr; i++) {
 730                 j = (u32)node_map->map[i];
 731                 ret = do_write(fd, &j, sizeof(j));
 732                 if (ret < 0)
 733                         break;
 734
 735                 ret = write_topo_node(fd, i);
 736                 if (ret < 0)
 737                         break;
 738         }
 739 done:
 740         free(buf);
 741         fclose(fp);
 742         free(node_map);
 743         return ret;
 744 }
 745
 746 /*
 747  * File format:
 748  *
 749  * struct pmu_mappings {
 750  *      u32     pmu_num;
 751  *      struct pmu_map {
 752  *              u32     type;
 753  *              char    name[];
 754  *      }[pmu_num];
 755  * };
 756  */
 757
 758 static int write_pmu_mappings(int fd, struct perf_header *h __maybe_unused,
 759                               struct perf_evlist *evlist __maybe_unused)
 760 {
 761         struct perf_pmu *pmu = NULL;
 762         off_t offset = lseek(fd, 0, SEEK_CUR);
 763         __u32 pmu_num = 0;
 764         int ret;
 765
 766         /* write real pmu_num later */
 767         ret = do_write(fd, &pmu_num, sizeof(pmu_num));
 768         if (ret < 0)
 769                 return ret;
 770
 771         while ((pmu = perf_pmu__scan(pmu))) {
 772                 if (!pmu->name)
 773                         continue;
 774                 pmu_num++;
 775
 776                 ret = do_write(fd, &pmu->type, sizeof(pmu->type));
 777                 if (ret < 0)
 778                         return ret;
 779
 780                 ret = do_write_string(fd, pmu->name);
 781                 if (ret < 0)
 782                         return ret;
 783         }
 784
 785         if (pwrite(fd, &pmu_num, sizeof(pmu_num), offset) != sizeof(pmu_num)) {
 786                 /* discard all */
 787                 lseek(fd, offset, SEEK_SET);
 788                 return -1;
 789         }
 790
 791         return 0;
 792 }
 793
 794 /*
 795  * File format:
 796  *
 797  * struct group_descs {
 798  *      u32     nr_groups;
 799  *      struct group_desc {
 800  *              char    name[];
 801  *              u32     leader_idx;
 802  *              u32     nr_members;
 803  *      }[nr_groups];
 804  * };
 805  */
 806 static int write_group_desc(int fd, struct perf_header *h __maybe_unused,
 807                             struct perf_evlist *evlist)
 808 {
 809         u32 nr_groups = evlist->nr_groups;
 810         struct perf_evsel *evsel;
 811         int ret;
 812
 813         ret = do_write(fd, &nr_groups, sizeof(nr_groups));
 814         if (ret < 0)
 815                 return ret;
 816
 817         evlist__for_each(evlist, evsel) {
 818                 if (perf_evsel__is_group_leader(evsel) &&
 819                     evsel->nr_members > 1) {
 820                         const char *name = evsel->group_name ?: "{anon_group}";
 821                         u32 leader_idx = evsel->idx;
 822                         u32 nr_members = evsel->nr_members;
 823
 824                         ret = do_write_string(fd, name);
 825                         if (ret < 0)
 826                                 return ret;
 827
 828                         ret = do_write(fd, &leader_idx, sizeof(leader_idx));
 829                         if (ret < 0)
 830                                 return ret;
 831
 832                         ret = do_write(fd, &nr_members, sizeof(nr_members));
 833                         if (ret < 0)
 834                                 return ret;
 835                 }
 836         }
 837         return 0;
 838 }
 839
 840 /*
 841  * default get_cpuid(): nothing gets recorded
 842  * actual implementation must be in arch/$(ARCH)/util/header.c
 843  */
 844 int __attribute__ ((weak)) get_cpuid(char *buffer __maybe_unused,
 845                                      size_t sz __maybe_unused)
 846 {
 847         return -1;
 848 }
 849
 850 static int write_cpuid(int fd, struct perf_header *h __maybe_unused,
 851                        struct perf_evlist *evlist __maybe_unused)
 852 {
 853         char buffer[64];
 854         int ret;
 855
 856         ret = get_cpuid(buffer, sizeof(buffer));
 857         if (!ret)
 858                 goto write_it;
 859
 860         return -1;
 861 write_it:
 862         return do_write_string(fd, buffer);
 863 }
 864
 865 static int write_branch_stack(int fd __maybe_unused,
 866                               struct perf_header *h __maybe_unused,
 867                        struct perf_evlist *evlist __maybe_unused)
 868 {
 869         return 0;
 870 }
 871
 872 static void print_hostname(struct perf_header *ph, int fd __maybe_unused,
 873                            FILE *fp)
 874 {
 875         fprintf(fp, "# hostname : %s\n", ph->env.hostname);
 876 }
 877
 878 static void print_osrelease(struct perf_header *ph, int fd __maybe_unused,
 879                             FILE *fp)
 880 {
 881         fprintf(fp, "# os release : %s\n", ph->env.os_release);
 882 }
 883
 884 static void print_arch(struct perf_header *ph, int fd __maybe_unused, FILE *fp)
 885 {
 886         fprintf(fp, "# arch : %s\n", ph->env.arch);
 887 }
 888
 889 static void print_cpudesc(struct perf_header *ph, int fd __maybe_unused,
 890                           FILE *fp)
 891 {
 892         fprintf(fp, "# cpudesc : %s\n", ph->env.cpu_desc);
 893 }
 894
 895 static void print_nrcpus(struct perf_header *ph, int fd __maybe_unused,
 896                          FILE *fp)
 897 {
 898         fprintf(fp, "# nrcpus online : %u\n", ph->env.nr_cpus_online);
 899         fprintf(fp, "# nrcpus avail : %u\n", ph->env.nr_cpus_avail);
 900 }
 901
 902 static void print_version(struct perf_header *ph, int fd __maybe_unused,
 903                           FILE *fp)
 904 {
 905         fprintf(fp, "# perf version : %s\n", ph->env.version);
 906 }
 907
 908 static void print_cmdline(struct perf_header *ph, int fd __maybe_unused,
 909                           FILE *fp)
 910 {
 911         int nr, i;
 912         char *str;
 913
 914         nr = ph->env.nr_cmdline;
 915         str = ph->env.cmdline;
 916
 917         fprintf(fp, "# cmdline : ");
 918
 919         for (i = 0; i < nr; i++) {
 920                 fprintf(fp, "%s ", str);
 921                 str += strlen(str) + 1;
 922         }
 923         fputc('\n', fp);
 924 }
 925
 926 static void print_cpu_topology(struct perf_header *ph, int fd __maybe_unused,
 927                                FILE *fp)
 928 {
 929         int nr, i;
 930         char *str;
 931
 932         nr = ph->env.nr_sibling_cores;
 933         str = ph->env.sibling_cores;
 934
 935         for (i = 0; i < nr; i++) {
 936                 fprintf(fp, "# sibling cores   : %s\n", str);
 937                 str += strlen(str) + 1;
 938         }
 939
 940         nr = ph->env.nr_sibling_threads;
 941         str = ph->env.sibling_threads;
 942
 943         for (i = 0; i < nr; i++) {
 944                 fprintf(fp, "# sibling threads : %s\n", str);
 945                 str += strlen(str) + 1;
 946         }
 947 }
 948
 949 static void free_event_desc(struct perf_evsel *events)
 950 {
 951         struct perf_evsel *evsel;
 952
 953         if (!events)
 954                 return;
 955
 956         for (evsel = events; evsel->attr.size; evsel++) {
 957                 zfree(&evsel->name);
 958                 zfree(&evsel->id);
 959         }
 960
 961         free(events);
 962 }
 963
 964 static struct perf_evsel *
 965 read_event_desc(struct perf_header *ph, int fd)
 966 {
 967         struct perf_evsel *evsel, *events = NULL;
 968         u64 *id;
 969         void *buf = NULL;
 970         u32 nre, sz, nr, i, j;
 971         ssize_t ret;
 972         size_t msz;
 973
 974         /* number of events */
 975         ret = readn(fd, &nre, sizeof(nre));
 976         if (ret != (ssize_t)sizeof(nre))
 977                 goto error;
 978
 979         if (ph->needs_swap)
 980                 nre = bswap_32(nre);
 981
 982         ret = readn(fd, &sz, sizeof(sz));
 983         if (ret != (ssize_t)sizeof(sz))
 984                 goto error;
 985
 986         if (ph->needs_swap)
 987                 sz = bswap_32(sz);
 988
 989         /* buffer to hold on file attr struct */
 990         buf = malloc(sz);
 991         if (!buf)
 992                 goto error;
 993
 994         /* the last event terminates with evsel->attr.size == 0: */
 995         events = calloc(nre + 1, sizeof(*events));
 996         if (!events)
 997                 goto error;
 998
 999         msz = sizeof(evsel->attr);
1000         if (sz < msz)
1001                 msz = sz;
1002
1003         for (i = 0, evsel = events; i < nre; evsel++, i++) {
1004                 evsel->idx = i;
1005
1006                 /*
1007                  * must read entire on-file attr struct to
1008                  * sync up with layout.
1009                  */
1010                 ret = readn(fd, buf, sz);
1011                 if (ret != (ssize_t)sz)
1012                         goto error;
1013
1014                 if (ph->needs_swap)
1015                         perf_event__attr_swap(buf);
1016
1017                 memcpy(&evsel->attr, buf, msz);
1018
1019                 ret = readn(fd, &nr, sizeof(nr));
1020                 if (ret != (ssize_t)sizeof(nr))
1021                         goto error;
1022
1023                 if (ph->needs_swap) {
1024                         nr = bswap_32(nr);
1025                         evsel->needs_swap = true;
1026                 }
1027
1028                 evsel->name = do_read_string(fd, ph);
1029
1030                 if (!nr)
1031                         continue;
1032
1033                 id = calloc(nr, sizeof(*id));
1034                 if (!id)
1035                         goto error;
1036                 evsel->ids = nr;
1037                 evsel->id = id;
1038
1039                 for (j = 0 ; j < nr; j++) {
1040                         ret = readn(fd, id, sizeof(*id));
1041                         if (ret != (ssize_t)sizeof(*id))
1042                                 goto error;
1043                         if (ph->needs_swap)
1044                                 *id = bswap_64(*id);
1045                         id++;
1046                 }
1047         }
1048 out:
1049         free(buf);
1050         return events;
1051 error:
1052         if (events)
1053                 free_event_desc(events);
1054         events = NULL;
1055         goto out;
1056 }
1057
1058 static int __desc_attr__fprintf(FILE *fp, const char *name, const char *val,
1059                                 void *priv __attribute__((unused)))
1060 {
1061         return fprintf(fp, ", %s = %s", name, val);
1062 }
1063
1064 static void print_event_desc(struct perf_header *ph, int fd, FILE *fp)
1065 {
1066         struct perf_evsel *evsel, *events = read_event_desc(ph, fd);
1067         u32 j;
1068         u64 *id;
1069
1070         if (!events) {
1071                 fprintf(fp, "# event desc: not available or unable to read\n");
1072                 return;
1073         }
1074
1075         for (evsel = events; evsel->attr.size; evsel++) {
1076                 fprintf(fp, "# event : name = %s, ", evsel->name);
1077
1078                 if (evsel->ids) {
1079                         fprintf(fp, ", id = {");
1080                         for (j = 0, id = evsel->id; j < evsel->ids; j++, id++) {
1081                                 if (j)
1082                                         fputc(',', fp);
1083                                 fprintf(fp, " %"PRIu64, *id);
1084                         }
1085                         fprintf(fp, " }");
1086                 }
1087
1088                 perf_event_attr__fprintf(fp, &evsel->attr, __desc_attr__fprintf, NULL);
1089
1090                 fputc('\n', fp);
1091         }
1092
1093         free_event_desc(events);
1094 }
1095
1096 static void print_total_mem(struct perf_header *ph, int fd __maybe_unused,
1097                             FILE *fp)
1098 {
1099         fprintf(fp, "# total memory : %Lu kB\n", ph->env.total_mem);
1100 }
1101
1102 static void print_numa_topology(struct perf_header *ph, int fd __maybe_unused,
1103                                 FILE *fp)
1104 {
1105         u32 nr, c, i;
1106         char *str, *tmp;
1107         uint64_t mem_total, mem_free;
1108
1109         /* nr nodes */
1110         nr = ph->env.nr_numa_nodes;
1111         str = ph->env.numa_nodes;
1112
1113         for (i = 0; i < nr; i++) {
1114                 /* node number */
1115                 c = strtoul(str, &tmp, 0);
1116                 if (*tmp != ':')
1117                         goto error;
1118
1119                 str = tmp + 1;
1120                 mem_total = strtoull(str, &tmp, 0);
1121                 if (*tmp != ':')
1122                         goto error;
1123
1124                 str = tmp + 1;
1125                 mem_free = strtoull(str, &tmp, 0);
1126                 if (*tmp != ':')
1127                         goto error;
1128
1129                 fprintf(fp, "# node%u meminfo  : total = %"PRIu64" kB,"
1130                             " free = %"PRIu64" kB\n",
1131                         c, mem_total, mem_free);
1132
1133                 str = tmp + 1;
1134                 fprintf(fp, "# node%u cpu list : %s\n", c, str);
1135
1136                 str += strlen(str) + 1;
1137         }
1138         return;
1139 error:
1140         fprintf(fp, "# numa topology : not available\n");
1141 }
1142
1143 static void print_cpuid(struct perf_header *ph, int fd __maybe_unused, FILE *fp)
1144 {
1145         fprintf(fp, "# cpuid : %s\n", ph->env.cpuid);
1146 }
1147
1148 static void print_branch_stack(struct perf_header *ph __maybe_unused,
1149                                int fd __maybe_unused, FILE *fp)
1150 {
1151         fprintf(fp, "# contains samples with branch stack\n");
1152 }
1153
1154 static void print_pmu_mappings(struct perf_header *ph, int fd __maybe_unused,
1155                                FILE *fp)
1156 {
1157         const char *delimiter = "# pmu mappings: ";
1158         char *str, *tmp;
1159         u32 pmu_num;
1160         u32 type;
1161
1162         pmu_num = ph->env.nr_pmu_mappings;
1163         if (!pmu_num) {
1164                 fprintf(fp, "# pmu mappings: not available\n");
1165                 return;
1166         }
1167
1168         str = ph->env.pmu_mappings;
1169
1170         while (pmu_num) {
1171                 type = strtoul(str, &tmp, 0);
1172                 if (*tmp != ':')
1173                         goto error;
1174
1175                 str = tmp + 1;
1176                 fprintf(fp, "%s%s = %" PRIu32, delimiter, str, type);
1177
1178                 delimiter = ", ";
1179                 str += strlen(str) + 1;
1180                 pmu_num--;
1181         }
1182
1183         fprintf(fp, "\n");
1184
1185         if (!pmu_num)
1186                 return;
1187 error:
1188         fprintf(fp, "# pmu mappings: unable to read\n");
1189 }
1190
1191 static void print_group_desc(struct perf_header *ph, int fd __maybe_unused,
1192                              FILE *fp)
1193 {
1194         struct perf_session *session;
1195         struct perf_evsel *evsel;
1196         u32 nr = 0;
1197
1198         session = container_of(ph, struct perf_session, header);
1199
1200         evlist__for_each(session->evlist, evsel) {
1201                 if (perf_evsel__is_group_leader(evsel) &&
1202                     evsel->nr_members > 1) {
1203                         fprintf(fp, "# group: %s{%s", evsel->group_name ?: "",
1204                                 perf_evsel__name(evsel));
1205
1206                         nr = evsel->nr_members - 1;
1207                 } else if (nr) {
1208                         fprintf(fp, ",%s", perf_evsel__name(evsel));
1209
1210                         if (--nr == 0)
1211                                 fprintf(fp, "}\n");
1212                 }
1213         }
1214 }
1215
1216 static int __event_process_build_id(struct build_id_event *bev,
1217                                     char *filename,
1218                                     struct perf_session *session)
1219 {
1220         int err = -1;
1221         struct dsos *dsos;
1222         struct machine *machine;
1223         u16 misc;
1224         struct dso *dso;
1225         enum dso_kernel_type dso_type;
1226
1227         machine = perf_session__findnew_machine(session, bev->pid);
1228         if (!machine)
1229                 goto out;
1230
1231         misc = bev->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
1232
1233         switch (misc) {
1234         case PERF_RECORD_MISC_KERNEL:
1235                 dso_type = DSO_TYPE_KERNEL;
1236                 dsos = &machine->kernel_dsos;
1237                 break;
1238         case PERF_RECORD_MISC_GUEST_KERNEL:
1239                 dso_type = DSO_TYPE_GUEST_KERNEL;
1240                 dsos = &machine->kernel_dsos;
1241                 break;
1242         case PERF_RECORD_MISC_USER:
1243         case PERF_RECORD_MISC_GUEST_USER:
1244                 dso_type = DSO_TYPE_USER;
1245                 dsos = &machine->user_dsos;
1246                 break;
1247         default:
1248                 goto out;
1249         }
1250
1251         dso = __dsos__findnew(dsos, filename);
1252         if (dso != NULL) {
1253                 char sbuild_id[BUILD_ID_SIZE * 2 + 1];
1254
1255                 dso__set_build_id(dso, &bev->build_id);
1256
1257                 if (!is_kernel_module(filename))
1258                         dso->kernel = dso_type;
1259
1260                 build_id__sprintf(dso->build_id, sizeof(dso->build_id),
1261                                   sbuild_id);
1262                 pr_debug("build id event received for %s: %s\n",
1263                          dso->long_name, sbuild_id);
1264         }
1265
1266         err = 0;
1267 out:
1268         return err;
1269 }
1270
1271 static int perf_header__read_build_ids_abi_quirk(struct perf_header *header,
1272                                                  int input, u64 offset, u64 size)
1273 {
1274         struct perf_session *session = container_of(header, struct perf_session, header);
1275         struct {
1276                 struct perf_event_header   header;
1277                 u8                         build_id[PERF_ALIGN(BUILD_ID_SIZE, sizeof(u64))];
1278                 char                       filename[0];
1279         } old_bev;
1280         struct build_id_event bev;
1281         char filename[PATH_MAX];
1282         u64 limit = offset + size;
1283
1284         while (offset < limit) {
1285                 ssize_t len;
1286
1287                 if (readn(input, &old_bev, sizeof(old_bev)) != sizeof(old_bev))
1288                         return -1;
1289
1290                 if (header->needs_swap)
1291                         perf_event_header__bswap(&old_bev.header);
1292
1293                 len = old_bev.header.size - sizeof(old_bev);
1294                 if (readn(input, filename, len) != len)
1295                         return -1;
1296
1297                 bev.header = old_bev.header;
1298
1299                 /*
1300                  * As the pid is the missing value, we need to fill
1301                  * it properly. The header.misc value give us nice hint.
1302                  */
1303                 bev.pid = HOST_KERNEL_ID;
1304                 if (bev.header.misc == PERF_RECORD_MISC_GUEST_USER ||
1305                     bev.header.misc == PERF_RECORD_MISC_GUEST_KERNEL)
1306                         bev.pid = DEFAULT_GUEST_KERNEL_ID;
1307
1308                 memcpy(bev.build_id, old_bev.build_id, sizeof(bev.build_id));
1309                 __event_process_build_id(&bev, filename, session);
1310
1311                 offset += bev.header.size;
1312         }
1313
1314         return 0;
1315 }
1316
1317 static int perf_header__read_build_ids(struct perf_header *header,
1318                                        int input, u64 offset, u64 size)
1319 {
1320         struct perf_session *session = container_of(header, struct perf_session, header);
1321         struct build_id_event bev;
1322         char filename[PATH_MAX];
1323         u64 limit = offset + size, orig_offset = offset;
1324         int err = -1;
1325
1326         while (offset < limit) {
1327                 ssize_t len;
1328
1329                 if (readn(input, &bev, sizeof(bev)) != sizeof(bev))
1330                         goto out;
1331
1332                 if (header->needs_swap)
1333                         perf_event_header__bswap(&bev.header);
1334
1335                 len = bev.header.size - sizeof(bev);
1336                 if (readn(input, filename, len) != len)
1337                         goto out;
1338                 /*
1339                  * The a1645ce1 changeset:
1340                  *
1341                  * "perf: 'perf kvm' tool for monitoring guest performance from host"
1342                  *
1343                  * Added a field to struct build_id_event that broke the file
1344                  * format.
1345                  *
1346                  * Since the kernel build-id is the first entry, process the
1347                  * table using the old format if the well known
1348                  * '[kernel.kallsyms]' string for the kernel build-id has the
1349                  * first 4 characters chopped off (where the pid_t sits).
1350                  */
1351                 if (memcmp(filename, "nel.kallsyms]", 13) == 0) {
1352                         if (lseek(input, orig_offset, SEEK_SET) == (off_t)-1)
1353                                 return -1;
1354                         return perf_header__read_build_ids_abi_quirk(header, input, offset, size);
1355                 }
1356
1357                 __event_process_build_id(&bev, filename, session);
1358
1359                 offset += bev.header.size;
1360         }
1361         err = 0;
1362 out:
1363         return err;
1364 }
1365
1366 static int process_tracing_data(struct perf_file_section *section __maybe_unused,
1367                                 struct perf_header *ph __maybe_unused,
1368                                 int fd, void *data)
1369 {
1370         ssize_t ret = trace_report(fd, data, false);
1371         return ret < 0 ? -1 : 0;
1372 }
1373
1374 static int process_build_id(struct perf_file_section *section,
1375                             struct perf_header *ph, int fd,
1376                             void *data __maybe_unused)
1377 {
1378         if (perf_header__read_build_ids(ph, fd, section->offset, section->size))
1379                 pr_debug("Failed to read buildids, continuing...\n");
1380         return 0;
1381 }
1382
1383 static int process_hostname(struct perf_file_section *section __maybe_unused,
1384                             struct perf_header *ph, int fd,
1385                             void *data __maybe_unused)
1386 {
1387         ph->env.hostname = do_read_string(fd, ph);
1388         return ph->env.hostname ? 0 : -ENOMEM;
1389 }
1390
1391 static int process_osrelease(struct perf_file_section *section __maybe_unused,
1392                              struct perf_header *ph, int fd,
1393                              void *data __maybe_unused)
1394 {
1395         ph->env.os_release = do_read_string(fd, ph);
1396         return ph->env.os_release ? 0 : -ENOMEM;
1397 }
1398
1399 static int process_version(struct perf_file_section *section __maybe_unused,
1400                            struct perf_header *ph, int fd,
1401                            void *data __maybe_unused)
1402 {
1403         ph->env.version = do_read_string(fd, ph);
1404         return ph->env.version ? 0 : -ENOMEM;
1405 }
1406
1407 static int process_arch(struct perf_file_section *section __maybe_unused,
1408                         struct perf_header *ph, int fd,
1409                         void *data __maybe_unused)
1410 {
1411         ph->env.arch = do_read_string(fd, ph);
1412         return ph->env.arch ? 0 : -ENOMEM;
1413 }
1414
1415 static int process_nrcpus(struct perf_file_section *section __maybe_unused,
1416                           struct perf_header *ph, int fd,
1417                           void *data __maybe_unused)
1418 {
1419         ssize_t ret;
1420         u32 nr;
1421
1422         ret = readn(fd, &nr, sizeof(nr));
1423         if (ret != sizeof(nr))
1424                 return -1;
1425
1426         if (ph->needs_swap)
1427                 nr = bswap_32(nr);
1428
1429         ph->env.nr_cpus_online = nr;
1430
1431         ret = readn(fd, &nr, sizeof(nr));
1432         if (ret != sizeof(nr))
1433                 return -1;
1434
1435         if (ph->needs_swap)
1436                 nr = bswap_32(nr);
1437
1438         ph->env.nr_cpus_avail = nr;
1439         return 0;
1440 }
1441
1442 static int process_cpudesc(struct perf_file_section *section __maybe_unused,
1443                            struct perf_header *ph, int fd,
1444                            void *data __maybe_unused)
1445 {
1446         ph->env.cpu_desc = do_read_string(fd, ph);
1447         return ph->env.cpu_desc ? 0 : -ENOMEM;
1448 }
1449
1450 static int process_cpuid(struct perf_file_section *section __maybe_unused,
1451                          struct perf_header *ph,  int fd,
1452                          void *data __maybe_unused)
1453 {
1454         ph->env.cpuid = do_read_string(fd, ph);
1455         return ph->env.cpuid ? 0 : -ENOMEM;
1456 }
1457
1458 static int process_total_mem(struct perf_file_section *section __maybe_unused,
1459                              struct perf_header *ph, int fd,
1460                              void *data __maybe_unused)
1461 {
1462         uint64_t mem;
1463         ssize_t ret;
1464
1465         ret = readn(fd, &mem, sizeof(mem));
1466         if (ret != sizeof(mem))
1467                 return -1;
1468
1469         if (ph->needs_swap)
1470                 mem = bswap_64(mem);
1471
1472         ph->env.total_mem = mem;
1473         return 0;
1474 }
1475
1476 static struct perf_evsel *
1477 perf_evlist__find_by_index(struct perf_evlist *evlist, int idx)
1478 {
1479         struct perf_evsel *evsel;
1480
1481         evlist__for_each(evlist, evsel) {
1482                 if (evsel->idx == idx)
1483                         return evsel;
1484         }
1485
1486         return NULL;
1487 }
1488
1489 static void
1490 perf_evlist__set_event_name(struct perf_evlist *evlist,
1491                             struct perf_evsel *event)
1492 {
1493         struct perf_evsel *evsel;
1494
1495         if (!event->name)
1496                 return;
1497
1498         evsel = perf_evlist__find_by_index(evlist, event->idx);
1499         if (!evsel)
1500                 return;
1501
1502         if (evsel->name)
1503                 return;
1504
1505         evsel->name = strdup(event->name);
1506 }
1507
1508 static int
1509 process_event_desc(struct perf_file_section *section __maybe_unused,
1510                    struct perf_header *header, int fd,
1511                    void *data __maybe_unused)
1512 {
1513         struct perf_session *session;
1514         struct perf_evsel *evsel, *events = read_event_desc(header, fd);
1515
1516         if (!events)
1517                 return 0;
1518
1519         session = container_of(header, struct perf_session, header);
1520         for (evsel = events; evsel->attr.size; evsel++)
1521                 perf_evlist__set_event_name(session->evlist, evsel);
1522
1523         free_event_desc(events);
1524
1525         return 0;
1526 }
1527
1528 static int process_cmdline(struct perf_file_section *section __maybe_unused,
1529                            struct perf_header *ph, int fd,
1530                            void *data __maybe_unused)
1531 {
1532         ssize_t ret;
1533         char *str;
1534         u32 nr, i;
1535         struct strbuf sb;
1536
1537         ret = readn(fd, &nr, sizeof(nr));
1538         if (ret != sizeof(nr))
1539                 return -1;
1540
1541         if (ph->needs_swap)
1542                 nr = bswap_32(nr);
1543
1544         ph->env.nr_cmdline = nr;
1545         strbuf_init(&sb, 128);
1546
1547         for (i = 0; i < nr; i++) {
1548                 str = do_read_string(fd, ph);
1549                 if (!str)
1550                         goto error;
1551
1552                 /* include a NULL character at the end */
1553                 strbuf_add(&sb, str, strlen(str) + 1);
1554                 free(str);
1555         }
1556         ph->env.cmdline = strbuf_detach(&sb, NULL);
1557         return 0;
1558
1559 error:
1560         strbuf_release(&sb);
1561         return -1;
1562 }
1563
1564 static int process_cpu_topology(struct perf_file_section *section __maybe_unused,
1565                                 struct perf_header *ph, int fd,
1566                                 void *data __maybe_unused)
1567 {
1568         ssize_t ret;
1569         u32 nr, i;
1570         char *str;
1571         struct strbuf sb;
1572
1573         ret = readn(fd, &nr, sizeof(nr));
1574         if (ret != sizeof(nr))
1575                 return -1;
1576
1577         if (ph->needs_swap)
1578                 nr = bswap_32(nr);
1579
1580         ph->env.nr_sibling_cores = nr;
1581         strbuf_init(&sb, 128);
1582
1583         for (i = 0; i < nr; i++) {
1584                 str = do_read_string(fd, ph);
1585                 if (!str)
1586                         goto error;
1587
1588                 /* include a NULL character at the end */
1589                 strbuf_add(&sb, str, strlen(str) + 1);
1590                 free(str);
1591         }
1592         ph->env.sibling_cores = strbuf_detach(&sb, NULL);
1593
1594         ret = readn(fd, &nr, sizeof(nr));
1595         if (ret != sizeof(nr))
1596                 return -1;
1597
1598         if (ph->needs_swap)
1599                 nr = bswap_32(nr);
1600
1601         ph->env.nr_sibling_threads = nr;
1602
1603         for (i = 0; i < nr; i++) {
1604                 str = do_read_string(fd, ph);
1605                 if (!str)
1606                         goto error;
1607
1608                 /* include a NULL character at the end */
1609                 strbuf_add(&sb, str, strlen(str) + 1);
1610                 free(str);
1611         }
1612         ph->env.sibling_threads = strbuf_detach(&sb, NULL);
1613         return 0;
1614
1615 error:
1616         strbuf_release(&sb);
1617         return -1;
1618 }
1619
1620 static int process_numa_topology(struct perf_file_section *section __maybe_unused,
1621                                  struct perf_header *ph, int fd,
1622                                  void *data __maybe_unused)
1623 {
1624         ssize_t ret;
1625         u32 nr, node, i;
1626         char *str;
1627         uint64_t mem_total, mem_free;
1628         struct strbuf sb;
1629
1630         /* nr nodes */
1631         ret = readn(fd, &nr, sizeof(nr));
1632         if (ret != sizeof(nr))
1633                 goto error;
1634
1635         if (ph->needs_swap)
1636                 nr = bswap_32(nr);
1637
1638         ph->env.nr_numa_nodes = nr;
1639         strbuf_init(&sb, 256);
1640
1641         for (i = 0; i < nr; i++) {
1642                 /* node number */
1643                 ret = readn(fd, &node, sizeof(node));
1644                 if (ret != sizeof(node))
1645                         goto error;
1646
1647                 ret = readn(fd, &mem_total, sizeof(u64));
1648                 if (ret != sizeof(u64))
1649                         goto error;
1650
1651                 ret = readn(fd, &mem_free, sizeof(u64));
1652                 if (ret != sizeof(u64))
1653                         goto error;
1654
1655                 if (ph->needs_swap) {
1656                         node = bswap_32(node);
1657                         mem_total = bswap_64(mem_total);
1658                         mem_free = bswap_64(mem_free);
1659                 }
1660
1661                 strbuf_addf(&sb, "%u:%"PRIu64":%"PRIu64":",
1662                             node, mem_total, mem_free);
1663
1664                 str = do_read_string(fd, ph);
1665                 if (!str)
1666                         goto error;
1667
1668                 /* include a NULL character at the end */
1669                 strbuf_add(&sb, str, strlen(str) + 1);
1670                 free(str);
1671         }
1672         ph->env.numa_nodes = strbuf_detach(&sb, NULL);
1673         return 0;
1674
1675 error:
1676         strbuf_release(&sb);
1677         return -1;
1678 }
1679
1680 static int process_pmu_mappings(struct perf_file_section *section __maybe_unused,
1681                                 struct perf_header *ph, int fd,
1682                                 void *data __maybe_unused)
1683 {
1684         ssize_t ret;
1685         char *name;
1686         u32 pmu_num;
1687         u32 type;
1688         struct strbuf sb;
1689
1690         ret = readn(fd, &pmu_num, sizeof(pmu_num));
1691         if (ret != sizeof(pmu_num))
1692                 return -1;
1693
1694         if (ph->needs_swap)
1695                 pmu_num = bswap_32(pmu_num);
1696
1697         if (!pmu_num) {
1698                 pr_debug("pmu mappings not available\n");
1699                 return 0;
1700         }
1701
1702         ph->env.nr_pmu_mappings = pmu_num;
1703         strbuf_init(&sb, 128);
1704
1705         while (pmu_num) {
1706                 if (readn(fd, &type, sizeof(type)) != sizeof(type))
1707                         goto error;
1708                 if (ph->needs_swap)
1709                         type = bswap_32(type);
1710
1711                 name = do_read_string(fd, ph);
1712                 if (!name)
1713                         goto error;
1714
1715                 strbuf_addf(&sb, "%u:%s", type, name);
1716                 /* include a NULL character at the end */
1717                 strbuf_add(&sb, "", 1);
1718
1719                 free(name);
1720                 pmu_num--;
1721         }
1722         ph->env.pmu_mappings = strbuf_detach(&sb, NULL);
1723         return 0;
1724
1725 error:
1726         strbuf_release(&sb);
1727         return -1;
1728 }
1729
1730 static int process_group_desc(struct perf_file_section *section __maybe_unused,
1731                               struct perf_header *ph, int fd,
1732                               void *data __maybe_unused)
1733 {
1734         size_t ret = -1;
1735         u32 i, nr, nr_groups;
1736         struct perf_session *session;
1737         struct perf_evsel *evsel, *leader = NULL;
1738         struct group_desc {
1739                 char *name;
1740                 u32 leader_idx;
1741                 u32 nr_members;
1742         } *desc;
1743
1744         if (readn(fd, &nr_groups, sizeof(nr_groups)) != sizeof(nr_groups))
1745                 return -1;
1746
1747         if (ph->needs_swap)
1748                 nr_groups = bswap_32(nr_groups);
1749
1750         ph->env.nr_groups = nr_groups;
1751         if (!nr_groups) {
1752                 pr_debug("group desc not available\n");
1753                 return 0;
1754         }
1755
1756         desc = calloc(nr_groups, sizeof(*desc));
1757         if (!desc)
1758                 return -1;
1759
1760         for (i = 0; i < nr_groups; i++) {
1761                 desc[i].name = do_read_string(fd, ph);
1762                 if (!desc[i].name)
1763                         goto out_free;
1764
1765                 if (readn(fd, &desc[i].leader_idx, sizeof(u32)) != sizeof(u32))
1766                         goto out_free;
1767
1768                 if (readn(fd, &desc[i].nr_members, sizeof(u32)) != sizeof(u32))
1769                         goto out_free;
1770
1771                 if (ph->needs_swap) {
1772                         desc[i].leader_idx = bswap_32(desc[i].leader_idx);
1773                         desc[i].nr_members = bswap_32(desc[i].nr_members);
1774                 }
1775         }
1776
1777         /*
1778          * Rebuild group relationship based on the group_desc
1779          */
1780         session = container_of(ph, struct perf_session, header);
1781         session->evlist->nr_groups = nr_groups;
1782
1783         i = nr = 0;
1784         evlist__for_each(session->evlist, evsel) {
1785                 if (evsel->idx == (int) desc[i].leader_idx) {
1786                         evsel->leader = evsel;
1787                         /* {anon_group} is a dummy name */
1788                         if (strcmp(desc[i].name, "{anon_group}")) {
1789                                 evsel->group_name = desc[i].name;
1790                                 desc[i].name = NULL;
1791                         }
1792                         evsel->nr_members = desc[i].nr_members;
1793
1794                         if (i >= nr_groups || nr > 0) {
1795                                 pr_debug("invalid group desc\n");
1796                                 goto out_free;
1797                         }
1798
1799                         leader = evsel;
1800                         nr = evsel->nr_members - 1;
1801                         i++;
1802                 } else if (nr) {
1803                         /* This is a group member */
1804                         evsel->leader = leader;
1805
1806                         nr--;
1807                 }
1808         }
1809
1810         if (i != nr_groups || nr != 0) {
1811                 pr_debug("invalid group desc\n");
1812                 goto out_free;
1813         }
1814
1815         ret = 0;
1816 out_free:
1817         for (i = 0; i < nr_groups; i++)
1818                 zfree(&desc[i].name);
1819         free(desc);
1820
1821         return ret;
1822 }
1823
1824 struct feature_ops {
1825         int (*write)(int fd, struct perf_header *h, struct perf_evlist *evlist);
1826         void (*print)(struct perf_header *h, int fd, FILE *fp);
1827         int (*process)(struct perf_file_section *section,
1828                        struct perf_header *h, int fd, void *data);
1829         const char *name;
1830         bool full_only;
1831 };
1832
1833 #define FEAT_OPA(n, func) \
1834         [n] = { .name = #n, .write = write_##func, .print = print_##func }
1835 #define FEAT_OPP(n, func) \
1836         [n] = { .name = #n, .write = write_##func, .print = print_##func, \
1837                 .process = process_##func }
1838 #define FEAT_OPF(n, func) \
1839         [n] = { .name = #n, .write = write_##func, .print = print_##func, \
1840                 .process = process_##func, .full_only = true }
1841
1842 /* feature_ops not implemented: */
1843 #define print_tracing_data      NULL
1844 #define print_build_id          NULL
1845
1846 static const struct feature_ops feat_ops[HEADER_LAST_FEATURE] = {
1847         FEAT_OPP(HEADER_TRACING_DATA,   tracing_data),
1848         FEAT_OPP(HEADER_BUILD_ID,       build_id),
1849         FEAT_OPP(HEADER_HOSTNAME,       hostname),
1850         FEAT_OPP(HEADER_OSRELEASE,      osrelease),
1851         FEAT_OPP(HEADER_VERSION,        version),
1852         FEAT_OPP(HEADER_ARCH,           arch),
1853         FEAT_OPP(HEADER_NRCPUS,         nrcpus),
1854         FEAT_OPP(HEADER_CPUDESC,        cpudesc),
1855         FEAT_OPP(HEADER_CPUID,          cpuid),
1856         FEAT_OPP(HEADER_TOTAL_MEM,      total_mem),
1857         FEAT_OPP(HEADER_EVENT_DESC,     event_desc),
1858         FEAT_OPP(HEADER_CMDLINE,        cmdline),
1859         FEAT_OPF(HEADER_CPU_TOPOLOGY,   cpu_topology),
1860         FEAT_OPF(HEADER_NUMA_TOPOLOGY,  numa_topology),
1861         FEAT_OPA(HEADER_BRANCH_STACK,   branch_stack),
1862         FEAT_OPP(HEADER_PMU_MAPPINGS,   pmu_mappings),
1863         FEAT_OPP(HEADER_GROUP_DESC,     group_desc),
1864 };
1865
1866 struct header_print_data {
1867         FILE *fp;
1868         bool full; /* extended list of headers */
1869 };
1870
1871 static int perf_file_section__fprintf_info(struct perf_file_section *section,
1872                                            struct perf_header *ph,
1873                                            int feat, int fd, void *data)
1874 {
1875         struct header_print_data *hd = data;
1876
1877         if (lseek(fd, section->offset, SEEK_SET) == (off_t)-1) {
1878                 pr_debug("Failed to lseek to %" PRIu64 " offset for feature "
1879                                 "%d, continuing...\n", section->offset, feat);
1880                 return 0;
1881         }
1882         if (feat >= HEADER_LAST_FEATURE) {
1883                 pr_warning("unknown feature %d\n", feat);
1884                 return 0;
1885         }
1886         if (!feat_ops[feat].print)
1887                 return 0;
1888
1889         if (!feat_ops[feat].full_only || hd->full)
1890                 feat_ops[feat].print(ph, fd, hd->fp);
1891         else
1892                 fprintf(hd->fp, "# %s info available, use -I to display\n",
1893                         feat_ops[feat].name);
1894
1895         return 0;
1896 }
1897
1898 int perf_header__fprintf_info(struct perf_session *session, FILE *fp, bool full)
1899 {
1900         struct header_print_data hd;
1901         struct perf_header *header = &session->header;
1902         int fd = perf_data_file__fd(session->file);
1903         hd.fp = fp;
1904         hd.full = full;
1905
1906         perf_header__process_sections(header, fd, &hd,
1907                                       perf_file_section__fprintf_info);
1908         return 0;
1909 }
1910
1911 static int do_write_feat(int fd, struct perf_header *h, int type,
1912                          struct perf_file_section **p,
1913                          struct perf_evlist *evlist)
1914 {
1915         int err;
1916         int ret = 0;
1917
1918         if (perf_header__has_feat(h, type)) {
1919                 if (!feat_ops[type].write)
1920                         return -1;
1921
1922                 (*p)->offset = lseek(fd, 0, SEEK_CUR);
1923
1924                 err = feat_ops[type].write(fd, h, evlist);
1925                 if (err < 0) {
1926                         pr_debug("failed to write feature %d\n", type);
1927
1928                         /* undo anything written */
1929                         lseek(fd, (*p)->offset, SEEK_SET);
1930
1931                         return -1;
1932                 }
1933                 (*p)->size = lseek(fd, 0, SEEK_CUR) - (*p)->offset;
1934                 (*p)++;
1935         }
1936         return ret;
1937 }
1938
1939 static int perf_header__adds_write(struct perf_header *header,
1940                                    struct perf_evlist *evlist, int fd)
1941 {
1942         int nr_sections;
1943         struct perf_file_section *feat_sec, *p;
1944         int sec_size;
1945         u64 sec_start;
1946         int feat;
1947         int err;
1948
1949         nr_sections = bitmap_weight(header->adds_features, HEADER_FEAT_BITS);
1950         if (!nr_sections)
1951                 return 0;
1952
1953         feat_sec = p = calloc(nr_sections, sizeof(*feat_sec));
1954         if (feat_sec == NULL)
1955                 return -ENOMEM;
1956
1957         sec_size = sizeof(*feat_sec) * nr_sections;
1958
1959         sec_start = header->feat_offset;
1960         lseek(fd, sec_start + sec_size, SEEK_SET);
1961
1962         for_each_set_bit(feat, header->adds_features, HEADER_FEAT_BITS) {
1963                 if (do_write_feat(fd, header, feat, &p, evlist))
1964                         perf_header__clear_feat(header, feat);
1965         }
1966
1967         lseek(fd, sec_start, SEEK_SET);
1968         /*
1969          * may write more than needed due to dropped feature, but
1970          * this is okay, reader will skip the mising entries
1971          */
1972         err = do_write(fd, feat_sec, sec_size);
1973         if (err < 0)
1974                 pr_debug("failed to write feature section\n");
1975         free(feat_sec);
1976         return err;
1977 }
1978
1979 int perf_header__write_pipe(int fd)
1980 {
1981         struct perf_pipe_file_header f_header;
1982         int err;
1983
1984         f_header = (struct perf_pipe_file_header){
1985                 .magic     = PERF_MAGIC,
1986                 .size      = sizeof(f_header),
1987         };
1988
1989         err = do_write(fd, &f_header, sizeof(f_header));
1990         if (err < 0) {
1991                 pr_debug("failed to write perf pipe header\n");
1992                 return err;
1993         }
1994
1995         return 0;
1996 }
1997
1998 int perf_session__write_header(struct perf_session *session,
1999                                struct perf_evlist *evlist,
2000                                int fd, bool at_exit)
2001 {
2002         struct perf_file_header f_header;
2003         struct perf_file_attr   f_attr;
2004         struct perf_header *header = &session->header;
2005         struct perf_evsel *evsel;
2006         u64 attr_offset;
2007         int err;
2008
2009         lseek(fd, sizeof(f_header), SEEK_SET);
2010
2011         evlist__for_each(session->evlist, evsel) {
2012                 evsel->id_offset = lseek(fd, 0, SEEK_CUR);
2013                 err = do_write(fd, evsel->id, evsel->ids * sizeof(u64));
2014                 if (err < 0) {
2015                         pr_debug("failed to write perf header\n");
2016                         return err;
2017                 }
2018         }
2019
2020         attr_offset = lseek(fd, 0, SEEK_CUR);
2021
2022         evlist__for_each(evlist, evsel) {
2023                 f_attr = (struct perf_file_attr){
2024                         .attr = evsel->attr,
2025                         .ids  = {
2026                                 .offset = evsel->id_offset,
2027                                 .size   = evsel->ids * sizeof(u64),
2028                         }
2029                 };
2030                 err = do_write(fd, &f_attr, sizeof(f_attr));
2031                 if (err < 0) {
2032                         pr_debug("failed to write perf header attribute\n");
2033                         return err;
2034                 }
2035         }
2036
2037         if (!header->data_offset)
2038                 header->data_offset = lseek(fd, 0, SEEK_CUR);
2039         header->feat_offset = header->data_offset + header->data_size;
2040
2041         if (at_exit) {
2042                 err = perf_header__adds_write(header, evlist, fd);
2043                 if (err < 0)
2044                         return err;
2045         }
2046
2047         f_header = (struct perf_file_header){
2048                 .magic     = PERF_MAGIC,
2049                 .size      = sizeof(f_header),
2050                 .attr_size = sizeof(f_attr),
2051                 .attrs = {
2052                         .offset = attr_offset,
2053                         .size   = evlist->nr_entries * sizeof(f_attr),
2054                 },
2055                 .data = {
2056                         .offset = header->data_offset,
2057                         .size   = header->data_size,
2058                 },
2059                 /* event_types is ignored, store zeros */
2060         };
2061
2062         memcpy(&f_header.adds_features, &header->adds_features, sizeof(header->adds_features));
2063
2064         lseek(fd, 0, SEEK_SET);
2065         err = do_write(fd, &f_header, sizeof(f_header));
2066         if (err < 0) {
2067                 pr_debug("failed to write perf header\n");
2068                 return err;
2069         }
2070         lseek(fd, header->data_offset + header->data_size, SEEK_SET);
2071
2072         return 0;
2073 }
2074
2075 static int perf_header__getbuffer64(struct perf_header *header,
2076                                     int fd, void *buf, size_t size)
2077 {
2078         if (readn(fd, buf, size) <= 0)
2079                 return -1;
2080
2081         if (header->needs_swap)
2082                 mem_bswap_64(buf, size);
2083
2084         return 0;
2085 }
2086
2087 int perf_header__process_sections(struct perf_header *header, int fd,
2088                                   void *data,
2089                                   int (*process)(struct perf_file_section *section,
2090                                                  struct perf_header *ph,
2091                                                  int feat, int fd, void *data))
2092 {
2093         struct perf_file_section *feat_sec, *sec;
2094         int nr_sections;
2095         int sec_size;
2096         int feat;
2097         int err;
2098
2099         nr_sections = bitmap_weight(header->adds_features, HEADER_FEAT_BITS);
2100         if (!nr_sections)
2101                 return 0;
2102
2103         feat_sec = sec = calloc(nr_sections, sizeof(*feat_sec));
2104         if (!feat_sec)
2105                 return -1;
2106
2107         sec_size = sizeof(*feat_sec) * nr_sections;
2108
2109         lseek(fd, header->feat_offset, SEEK_SET);
2110
2111         err = perf_header__getbuffer64(header, fd, feat_sec, sec_size);
2112         if (err < 0)
2113                 goto out_free;
2114
2115         for_each_set_bit(feat, header->adds_features, HEADER_LAST_FEATURE) {
2116                 err = process(sec++, header, feat, fd, data);
2117                 if (err < 0)
2118                         goto out_free;
2119         }
2120         err = 0;
2121 out_free:
2122         free(feat_sec);
2123         return err;
2124 }
2125
2126 static const int attr_file_abi_sizes[] = {
2127         [0] = PERF_ATTR_SIZE_VER0,
2128         [1] = PERF_ATTR_SIZE_VER1,
2129         [2] = PERF_ATTR_SIZE_VER2,
2130         [3] = PERF_ATTR_SIZE_VER3,
2131         [4] = PERF_ATTR_SIZE_VER4,
2132         0,
2133 };
2134
2135 /*
2136  * In the legacy file format, the magic number is not used to encode endianness.
2137  * hdr_sz was used to encode endianness. But given that hdr_sz can vary based
2138  * on ABI revisions, we need to try all combinations for all endianness to
2139  * detect the endianness.
2140  */
2141 static int try_all_file_abis(uint64_t hdr_sz, struct perf_header *ph)
2142 {
2143         uint64_t ref_size, attr_size;
2144         int i;
2145
2146         for (i = 0 ; attr_file_abi_sizes[i]; i++) {
2147                 ref_size = attr_file_abi_sizes[i]
2148                          + sizeof(struct perf_file_section);
2149                 if (hdr_sz != ref_size) {
2150                         attr_size = bswap_64(hdr_sz);
2151                         if (attr_size != ref_size)
2152                                 continue;
2153
2154                         ph->needs_swap = true;
2155                 }
2156                 pr_debug("ABI%d perf.data file detected, need_swap=%d\n",
2157                          i,
2158                          ph->needs_swap);
2159                 return 0;
2160         }
2161         /* could not determine endianness */
2162         return -1;
2163 }
2164
2165 #define PERF_PIPE_HDR_VER0      16
2166
2167 static const size_t attr_pipe_abi_sizes[] = {
2168         [0] = PERF_PIPE_HDR_VER0,
2169         0,
2170 };
2171
2172 /*
2173  * In the legacy pipe format, there is an implicit assumption that endiannesss
2174  * between host recording the samples, and host parsing the samples is the
2175  * same. This is not always the case given that the pipe output may always be
2176  * redirected into a file and analyzed on a different machine with possibly a
2177  * different endianness and perf_event ABI revsions in the perf tool itself.
2178  */
2179 static int try_all_pipe_abis(uint64_t hdr_sz, struct perf_header *ph)
2180 {
2181         u64 attr_size;
2182         int i;
2183
2184         for (i = 0 ; attr_pipe_abi_sizes[i]; i++) {
2185                 if (hdr_sz != attr_pipe_abi_sizes[i]) {
2186                         attr_size = bswap_64(hdr_sz);
2187                         if (attr_size != hdr_sz)
2188                                 continue;
2189
2190                         ph->needs_swap = true;
2191                 }
2192                 pr_debug("Pipe ABI%d perf.data file detected\n", i);
2193                 return 0;
2194         }
2195         return -1;
2196 }
2197
2198 bool is_perf_magic(u64 magic)
2199 {
2200         if (!memcmp(&magic, __perf_magic1, sizeof(magic))
2201                 || magic == __perf_magic2
2202                 || magic == __perf_magic2_sw)
2203                 return true;
2204
2205         return false;
2206 }
2207
2208 static int check_magic_endian(u64 magic, uint64_t hdr_sz,
2209                               bool is_pipe, struct perf_header *ph)
2210 {
2211         int ret;
2212
2213         /* check for legacy format */
2214         ret = memcmp(&magic, __perf_magic1, sizeof(magic));
2215         if (ret == 0) {
2216                 ph->version = PERF_HEADER_VERSION_1;
2217                 pr_debug("legacy perf.data format\n");
2218                 if (is_pipe)
2219                         return try_all_pipe_abis(hdr_sz, ph);
2220
2221                 return try_all_file_abis(hdr_sz, ph);
2222         }
2223         /*
2224          * the new magic number serves two purposes:
2225          * - unique number to identify actual perf.data files
2226          * - encode endianness of file
2227          */
2228         ph->version = PERF_HEADER_VERSION_2;
2229
2230         /* check magic number with one endianness */
2231         if (magic == __perf_magic2)
2232                 return 0;
2233
2234         /* check magic number with opposite endianness */
2235         if (magic != __perf_magic2_sw)
2236                 return -1;
2237
2238         ph->needs_swap = true;
2239
2240         return 0;
2241 }
2242
2243 int perf_file_header__read(struct perf_file_header *header,
2244                            struct perf_header *ph, int fd)
2245 {
2246         ssize_t ret;
2247
2248         lseek(fd, 0, SEEK_SET);
2249
2250         ret = readn(fd, header, sizeof(*header));
2251         if (ret <= 0)
2252                 return -1;
2253
2254         if (check_magic_endian(header->magic,
2255                                header->attr_size, false, ph) < 0) {
2256                 pr_debug("magic/endian check failed\n");
2257                 return -1;
2258         }
2259
2260         if (ph->needs_swap) {
2261                 mem_bswap_64(header, offsetof(struct perf_file_header,
2262                              adds_features));
2263         }
2264
2265         if (header->size != sizeof(*header)) {
2266                 /* Support the previous format */
2267                 if (header->size == offsetof(typeof(*header), adds_features))
2268                         bitmap_zero(header->adds_features, HEADER_FEAT_BITS);
2269                 else
2270                         return -1;
2271         } else if (ph->needs_swap) {
2272                 /*
2273                  * feature bitmap is declared as an array of unsigned longs --
2274                  * not good since its size can differ between the host that
2275                  * generated the data file and the host analyzing the file.
2276                  *
2277                  * We need to handle endianness, but we don't know the size of
2278                  * the unsigned long where the file was generated. Take a best
2279                  * guess at determining it: try 64-bit swap first (ie., file
2280                  * created on a 64-bit host), and check if the hostname feature
2281                  * bit is set (this feature bit is forced on as of fbe96f2).
2282                  * If the bit is not, undo the 64-bit swap and try a 32-bit
2283                  * swap. If the hostname bit is still not set (e.g., older data
2284                  * file), punt and fallback to the original behavior --
2285                  * clearing all feature bits and setting buildid.
2286                  */
2287                 mem_bswap_64(&header->adds_features,
2288                             BITS_TO_U64(HEADER_FEAT_BITS));
2289
2290                 if (!test_bit(HEADER_HOSTNAME, header->adds_features)) {
2291                         /* unswap as u64 */
2292                         mem_bswap_64(&header->adds_features,
2293                                     BITS_TO_U64(HEADER_FEAT_BITS));
2294
2295                         /* unswap as u32 */
2296                         mem_bswap_32(&header->adds_features,
2297                                     BITS_TO_U32(HEADER_FEAT_BITS));
2298                 }
2299
2300                 if (!test_bit(HEADER_HOSTNAME, header->adds_features)) {
2301                         bitmap_zero(header->adds_features, HEADER_FEAT_BITS);
2302                         set_bit(HEADER_BUILD_ID, header->adds_features);
2303                 }
2304         }
2305
2306         memcpy(&ph->adds_features, &header->adds_features,
2307                sizeof(ph->adds_features));
2308
2309         ph->data_offset  = header->data.offset;
2310         ph->data_size    = header->data.size;
2311         ph->feat_offset  = header->data.offset + header->data.size;
2312         return 0;
2313 }
2314
2315 static int perf_file_section__process(struct perf_file_section *section,
2316                                       struct perf_header *ph,
2317                                       int feat, int fd, void *data)
2318 {
2319         if (lseek(fd, section->offset, SEEK_SET) == (off_t)-1) {
2320                 pr_debug("Failed to lseek to %" PRIu64 " offset for feature "
2321                           "%d, continuing...\n", section->offset, feat);
2322                 return 0;
2323         }
2324
2325         if (feat >= HEADER_LAST_FEATURE) {
2326                 pr_debug("unknown feature %d, continuing...\n", feat);
2327                 return 0;
2328         }
2329
2330         if (!feat_ops[feat].process)
2331                 return 0;
2332
2333         return feat_ops[feat].process(section, ph, fd, data);
2334 }
2335
2336 static int perf_file_header__read_pipe(struct perf_pipe_file_header *header,
2337                                        struct perf_header *ph, int fd,
2338                                        bool repipe)
2339 {
2340         ssize_t ret;
2341
2342         ret = readn(fd, header, sizeof(*header));
2343         if (ret <= 0)
2344                 return -1;
2345
2346         if (check_magic_endian(header->magic, header->size, true, ph) < 0) {
2347                 pr_debug("endian/magic failed\n");
2348                 return -1;
2349         }
2350
2351         if (ph->needs_swap)
2352                 header->size = bswap_64(header->size);
2353
2354         if (repipe && do_write(STDOUT_FILENO, header, sizeof(*header)) < 0)
2355                 return -1;
2356
2357         return 0;
2358 }
2359
2360 static int perf_header__read_pipe(struct perf_session *session)
2361 {
2362         struct perf_header *header = &session->header;
2363         struct perf_pipe_file_header f_header;
2364
2365         if (perf_file_header__read_pipe(&f_header, header,
2366                                         perf_data_file__fd(session->file),
2367                                         session->repipe) < 0) {
2368                 pr_debug("incompatible file format\n");
2369                 return -EINVAL;
2370         }
2371
2372         return 0;
2373 }
2374
2375 static int read_attr(int fd, struct perf_header *ph,
2376                      struct perf_file_attr *f_attr)
2377 {
2378         struct perf_event_attr *attr = &f_attr->attr;
2379         size_t sz, left;
2380         size_t our_sz = sizeof(f_attr->attr);
2381         ssize_t ret;
2382
2383         memset(f_attr, 0, sizeof(*f_attr));
2384
2385         /* read minimal guaranteed structure */
2386         ret = readn(fd, attr, PERF_ATTR_SIZE_VER0);
2387         if (ret <= 0) {
2388                 pr_debug("cannot read %d bytes of header attr\n",
2389                          PERF_ATTR_SIZE_VER0);
2390                 return -1;
2391         }
2392
2393         /* on file perf_event_attr size */
2394         sz = attr->size;
2395
2396         if (ph->needs_swap)
2397                 sz = bswap_32(sz);
2398
2399         if (sz == 0) {
2400                 /* assume ABI0 */
2401                 sz =  PERF_ATTR_SIZE_VER0;
2402         } else if (sz > our_sz) {
2403                 pr_debug("file uses a more recent and unsupported ABI"
2404                          " (%zu bytes extra)\n", sz - our_sz);
2405                 return -1;
2406         }
2407         /* what we have not yet read and that we know about */
2408         left = sz - PERF_ATTR_SIZE_VER0;
2409         if (left) {
2410                 void *ptr = attr;
2411                 ptr += PERF_ATTR_SIZE_VER0;
2412
2413                 ret = readn(fd, ptr, left);
2414         }
2415         /* read perf_file_section, ids are read in caller */
2416         ret = readn(fd, &f_attr->ids, sizeof(f_attr->ids));
2417
2418         return ret <= 0 ? -1 : 0;
2419 }
2420
2421 static int perf_evsel__prepare_tracepoint_event(struct perf_evsel *evsel,
2422                                                 struct pevent *pevent)
2423 {
2424         struct event_format *event;
2425         char bf[128];
2426
2427         /* already prepared */
2428         if (evsel->tp_format)
2429                 return 0;
2430
2431         if (pevent == NULL) {
2432                 pr_debug("broken or missing trace data\n");
2433                 return -1;
2434         }
2435
2436         event = pevent_find_event(pevent, evsel->attr.config);
2437         if (event == NULL)
2438                 return -1;
2439
2440         if (!evsel->name) {
2441                 snprintf(bf, sizeof(bf), "%s:%s", event->system, event->name);
2442                 evsel->name = strdup(bf);
2443                 if (evsel->name == NULL)
2444                         return -1;
2445         }
2446
2447         evsel->tp_format = event;
2448         return 0;
2449 }
2450
2451 static int perf_evlist__prepare_tracepoint_events(struct perf_evlist *evlist,
2452                                                   struct pevent *pevent)
2453 {
2454         struct perf_evsel *pos;
2455
2456         evlist__for_each(evlist, pos) {
2457                 if (pos->attr.type == PERF_TYPE_TRACEPOINT &&
2458                     perf_evsel__prepare_tracepoint_event(pos, pevent))
2459                         return -1;
2460         }
2461
2462         return 0;
2463 }
2464
2465 int perf_session__read_header(struct perf_session *session)
2466 {
2467         struct perf_data_file *file = session->file;
2468         struct perf_header *header = &session->header;
2469         struct perf_file_header f_header;
2470         struct perf_file_attr   f_attr;
2471         u64                     f_id;
2472         int nr_attrs, nr_ids, i, j;
2473         int fd = perf_data_file__fd(file);
2474
2475         session->evlist = perf_evlist__new();
2476         if (session->evlist == NULL)
2477                 return -ENOMEM;
2478
2479         if (perf_data_file__is_pipe(file))
2480                 return perf_header__read_pipe(session);
2481
2482         if (perf_file_header__read(&f_header, header, fd) < 0)
2483                 return -EINVAL;
2484
2485         /*
2486          * Sanity check that perf.data was written cleanly; data size is
2487          * initialized to 0 and updated only if the on_exit function is run.
2488          * If data size is still 0 then the file contains only partial
2489          * information.  Just warn user and process it as much as it can.
2490          */
2491         if (f_header.data.size == 0) {
2492                 pr_warning("WARNING: The %s file's data size field is 0 which is unexpected.\n"
2493                            "Was the 'perf record' command properly terminated?\n",
2494                            file->path);
2495         }
2496
2497         nr_attrs = f_header.attrs.size / f_header.attr_size;
2498         lseek(fd, f_header.attrs.offset, SEEK_SET);
2499
2500         for (i = 0; i < nr_attrs; i++) {
2501                 struct perf_evsel *evsel;
2502                 off_t tmp;
2503
2504                 if (read_attr(fd, header, &f_attr) < 0)
2505                         goto out_errno;
2506
2507                 if (header->needs_swap) {
2508                         f_attr.ids.size   = bswap_64(f_attr.ids.size);
2509                         f_attr.ids.offset = bswap_64(f_attr.ids.offset);
2510                         perf_event__attr_swap(&f_attr.attr);
2511                 }
2512
2513                 tmp = lseek(fd, 0, SEEK_CUR);
2514                 evsel = perf_evsel__new(&f_attr.attr);
2515
2516                 if (evsel == NULL)
2517                         goto out_delete_evlist;
2518
2519                 evsel->needs_swap = header->needs_swap;
2520                 /*
2521                  * Do it before so that if perf_evsel__alloc_id fails, this
2522                  * entry gets purged too at perf_evlist__delete().
2523                  */
2524                 perf_evlist__add(session->evlist, evsel);
2525
2526                 nr_ids = f_attr.ids.size / sizeof(u64);
2527                 /*
2528                  * We don't have the cpu and thread maps on the header, so
2529                  * for allocating the perf_sample_id table we fake 1 cpu and
2530                  * hattr->ids threads.
2531                  */
2532                 if (perf_evsel__alloc_id(evsel, 1, nr_ids))
2533                         goto out_delete_evlist;
2534
2535                 lseek(fd, f_attr.ids.offset, SEEK_SET);
2536
2537                 for (j = 0; j < nr_ids; j++) {
2538                         if (perf_header__getbuffer64(header, fd, &f_id, sizeof(f_id)))
2539                                 goto out_errno;
2540
2541                         perf_evlist__id_add(session->evlist, evsel, 0, j, f_id);
2542                 }
2543
2544                 lseek(fd, tmp, SEEK_SET);
2545         }
2546
2547         symbol_conf.nr_events = nr_attrs;
2548
2549         perf_header__process_sections(header, fd, &session->tevent,
2550                                       perf_file_section__process);
2551
2552         if (perf_evlist__prepare_tracepoint_events(session->evlist,
2553                                                    session->tevent.pevent))
2554                 goto out_delete_evlist;
2555
2556         return 0;
2557 out_errno:
2558         return -errno;
2559
2560 out_delete_evlist:
2561         perf_evlist__delete(session->evlist);
2562         session->evlist = NULL;
2563         return -ENOMEM;
2564 }
2565
2566 int perf_event__synthesize_attr(struct perf_tool *tool,
2567                                 struct perf_event_attr *attr, u32 ids, u64 *id,
2568                                 perf_event__handler_t process)
2569 {
2570         union perf_event *ev;
2571         size_t size;
2572         int err;
2573
2574         size = sizeof(struct perf_event_attr);
2575         size = PERF_ALIGN(size, sizeof(u64));
2576         size += sizeof(struct perf_event_header);
2577         size += ids * sizeof(u64);
2578
2579         ev = malloc(size);
2580
2581         if (ev == NULL)
2582                 return -ENOMEM;
2583
2584         ev->attr.attr = *attr;
2585         memcpy(ev->attr.id, id, ids * sizeof(u64));
2586
2587         ev->attr.header.type = PERF_RECORD_HEADER_ATTR;
2588         ev->attr.header.size = (u16)size;
2589
2590         if (ev->attr.header.size == size)
2591                 err = process(tool, ev, NULL, NULL);
2592         else
2593                 err = -E2BIG;
2594
2595         free(ev);
2596
2597         return err;
2598 }
2599
2600 int perf_event__synthesize_attrs(struct perf_tool *tool,
2601                                    struct perf_session *session,
2602                                    perf_event__handler_t process)
2603 {
2604         struct perf_evsel *evsel;
2605         int err = 0;
2606
2607         evlist__for_each(session->evlist, evsel) {
2608                 err = perf_event__synthesize_attr(tool, &evsel->attr, evsel->ids,
2609                                                   evsel->id, process);
2610                 if (err) {
2611                         pr_debug("failed to create perf header attribute\n");
2612                         return err;
2613                 }
2614         }
2615
2616         return err;
2617 }
2618
2619 int perf_event__process_attr(struct perf_tool *tool __maybe_unused,
2620                              union perf_event *event,
2621                              struct perf_evlist **pevlist)
2622 {
2623         u32 i, ids, n_ids;
2624         struct perf_evsel *evsel;
2625         struct perf_evlist *evlist = *pevlist;
2626
2627         if (evlist == NULL) {
2628                 *pevlist = evlist = perf_evlist__new();
2629                 if (evlist == NULL)
2630                         return -ENOMEM;
2631         }
2632
2633         evsel = perf_evsel__new(&event->attr.attr);
2634         if (evsel == NULL)
2635                 return -ENOMEM;
2636
2637         perf_evlist__add(evlist, evsel);
2638
2639         ids = event->header.size;
2640         ids -= (void *)&event->attr.id - (void *)event;
2641         n_ids = ids / sizeof(u64);
2642         /*
2643          * We don't have the cpu and thread maps on the header, so
2644          * for allocating the perf_sample_id table we fake 1 cpu and
2645          * hattr->ids threads.
2646          */
2647         if (perf_evsel__alloc_id(evsel, 1, n_ids))
2648                 return -ENOMEM;
2649
2650         for (i = 0; i < n_ids; i++) {
2651                 perf_evlist__id_add(evlist, evsel, 0, i, event->attr.id[i]);
2652         }
2653
2654         symbol_conf.nr_events = evlist->nr_entries;
2655
2656         return 0;
2657 }
2658
2659 int perf_event__synthesize_tracing_data(struct perf_tool *tool, int fd,
2660                                         struct perf_evlist *evlist,
2661                                         perf_event__handler_t process)
2662 {
2663         union perf_event ev;
2664         struct tracing_data *tdata;
2665         ssize_t size = 0, aligned_size = 0, padding;
2666         int err __maybe_unused = 0;
2667
2668         /*
2669          * We are going to store the size of the data followed
2670          * by the data contents. Since the fd descriptor is a pipe,
2671          * we cannot seek back to store the size of the data once
2672          * we know it. Instead we:
2673          *
2674          * - write the tracing data to the temp file
2675          * - get/write the data size to pipe
2676          * - write the tracing data from the temp file
2677          *   to the pipe
2678          */
2679         tdata = tracing_data_get(&evlist->entries, fd, true);
2680         if (!tdata)
2681                 return -1;
2682
2683         memset(&ev, 0, sizeof(ev));
2684
2685         ev.tracing_data.header.type = PERF_RECORD_HEADER_TRACING_DATA;
2686         size = tdata->size;
2687         aligned_size = PERF_ALIGN(size, sizeof(u64));
2688         padding = aligned_size - size;
2689         ev.tracing_data.header.size = sizeof(ev.tracing_data);
2690         ev.tracing_data.size = aligned_size;
2691
2692         process(tool, &ev, NULL, NULL);
2693
2694         /*
2695          * The put function will copy all the tracing data
2696          * stored in temp file to the pipe.
2697          */
2698         tracing_data_put(tdata);
2699
2700         write_padded(fd, NULL, 0, padding);
2701
2702         return aligned_size;
2703 }
2704
2705 int perf_event__process_tracing_data(struct perf_tool *tool __maybe_unused,
2706                                      union perf_event *event,
2707                                      struct perf_session *session)
2708 {
2709         ssize_t size_read, padding, size = event->tracing_data.size;
2710         int fd = perf_data_file__fd(session->file);
2711         off_t offset = lseek(fd, 0, SEEK_CUR);
2712         char buf[BUFSIZ];
2713
2714         /* setup for reading amidst mmap */
2715         lseek(fd, offset + sizeof(struct tracing_data_event),
2716               SEEK_SET);
2717
2718         size_read = trace_report(fd, &session->tevent,
2719                                  session->repipe);
2720         padding = PERF_ALIGN(size_read, sizeof(u64)) - size_read;
2721
2722         if (readn(fd, buf, padding) < 0) {
2723                 pr_err("%s: reading input file", __func__);
2724                 return -1;
2725         }
2726         if (session->repipe) {
2727                 int retw = write(STDOUT_FILENO, buf, padding);
2728                 if (retw <= 0 || retw != padding) {
2729                         pr_err("%s: repiping tracing data padding", __func__);
2730                         return -1;
2731                 }
2732         }
2733
2734         if (size_read + padding != size) {
2735                 pr_err("%s: tracing data size mismatch", __func__);
2736                 return -1;
2737         }
2738
2739         perf_evlist__prepare_tracepoint_events(session->evlist,
2740                                                session->tevent.pevent);
2741
2742         return size_read + padding;
2743 }
2744
2745 int perf_event__synthesize_build_id(struct perf_tool *tool,
2746                                     struct dso *pos, u16 misc,
2747                                     perf_event__handler_t process,
2748                                     struct machine *machine)
2749 {
2750         union perf_event ev;
2751         size_t len;
2752         int err = 0;
2753
2754         if (!pos->hit)
2755                 return err;
2756
2757         memset(&ev, 0, sizeof(ev));
2758
2759         len = pos->long_name_len + 1;
2760         len = PERF_ALIGN(len, NAME_ALIGN);
2761         memcpy(&ev.build_id.build_id, pos->build_id, sizeof(pos->build_id));
2762         ev.build_id.header.type = PERF_RECORD_HEADER_BUILD_ID;
2763         ev.build_id.header.misc = misc;
2764         ev.build_id.pid = machine->pid;
2765         ev.build_id.header.size = sizeof(ev.build_id) + len;
2766         memcpy(&ev.build_id.filename, pos->long_name, pos->long_name_len);
2767
2768         err = process(tool, &ev, NULL, machine);
2769
2770         return err;
2771 }
2772
2773 int perf_event__process_build_id(struct perf_tool *tool __maybe_unused,
2774                                  union perf_event *event,
2775                                  struct perf_session *session)
2776 {
2777         __event_process_build_id(&event->build_id,
2778                                  event->build_id.filename,
2779                                  session);
2780         return 0;
2781 }