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 void print_event_desc(struct perf_header *ph, int fd, FILE *fp)
1059 {
1060         struct perf_evsel *evsel, *events = read_event_desc(ph, fd);
1061         u32 j;
1062         u64 *id;
1063
1064         if (!events) {
1065                 fprintf(fp, "# event desc: not available or unable to read\n");
1066                 return;
1067         }
1068
1069         for (evsel = events; evsel->attr.size; evsel++) {
1070                 fprintf(fp, "# event : name = %s, ", evsel->name);
1071
1072                 fprintf(fp, "type = %d, config = 0x%"PRIx64
1073                             ", config1 = 0x%"PRIx64", config2 = 0x%"PRIx64,
1074                                 evsel->attr.type,
1075                                 (u64)evsel->attr.config,
1076                                 (u64)evsel->attr.config1,
1077                                 (u64)evsel->attr.config2);
1078
1079                 fprintf(fp, ", excl_usr = %d, excl_kern = %d",
1080                                 evsel->attr.exclude_user,
1081                                 evsel->attr.exclude_kernel);
1082
1083                 fprintf(fp, ", excl_host = %d, excl_guest = %d",
1084                                 evsel->attr.exclude_host,
1085                                 evsel->attr.exclude_guest);
1086
1087                 fprintf(fp, ", precise_ip = %d", evsel->attr.precise_ip);
1088
1089                 fprintf(fp, ", attr_mmap2 = %d", evsel->attr.mmap2);
1090                 fprintf(fp, ", attr_mmap  = %d", evsel->attr.mmap);
1091                 fprintf(fp, ", attr_mmap_data = %d", evsel->attr.mmap_data);
1092                 if (evsel->ids) {
1093                         fprintf(fp, ", id = {");
1094                         for (j = 0, id = evsel->id; j < evsel->ids; j++, id++) {
1095                                 if (j)
1096                                         fputc(',', fp);
1097                                 fprintf(fp, " %"PRIu64, *id);
1098                         }
1099                         fprintf(fp, " }");
1100                 }
1101
1102                 fputc('\n', fp);
1103         }
1104
1105         free_event_desc(events);
1106 }
1107
1108 static void print_total_mem(struct perf_header *ph, int fd __maybe_unused,
1109                             FILE *fp)
1110 {
1111         fprintf(fp, "# total memory : %Lu kB\n", ph->env.total_mem);
1112 }
1113
1114 static void print_numa_topology(struct perf_header *ph, int fd __maybe_unused,
1115                                 FILE *fp)
1116 {
1117         u32 nr, c, i;
1118         char *str, *tmp;
1119         uint64_t mem_total, mem_free;
1120
1121         /* nr nodes */
1122         nr = ph->env.nr_numa_nodes;
1123         str = ph->env.numa_nodes;
1124
1125         for (i = 0; i < nr; i++) {
1126                 /* node number */
1127                 c = strtoul(str, &tmp, 0);
1128                 if (*tmp != ':')
1129                         goto error;
1130
1131                 str = tmp + 1;
1132                 mem_total = strtoull(str, &tmp, 0);
1133                 if (*tmp != ':')
1134                         goto error;
1135
1136                 str = tmp + 1;
1137                 mem_free = strtoull(str, &tmp, 0);
1138                 if (*tmp != ':')
1139                         goto error;
1140
1141                 fprintf(fp, "# node%u meminfo  : total = %"PRIu64" kB,"
1142                             " free = %"PRIu64" kB\n",
1143                         c, mem_total, mem_free);
1144
1145                 str = tmp + 1;
1146                 fprintf(fp, "# node%u cpu list : %s\n", c, str);
1147
1148                 str += strlen(str) + 1;
1149         }
1150         return;
1151 error:
1152         fprintf(fp, "# numa topology : not available\n");
1153 }
1154
1155 static void print_cpuid(struct perf_header *ph, int fd __maybe_unused, FILE *fp)
1156 {
1157         fprintf(fp, "# cpuid : %s\n", ph->env.cpuid);
1158 }
1159
1160 static void print_branch_stack(struct perf_header *ph __maybe_unused,
1161                                int fd __maybe_unused, FILE *fp)
1162 {
1163         fprintf(fp, "# contains samples with branch stack\n");
1164 }
1165
1166 static void print_pmu_mappings(struct perf_header *ph, int fd __maybe_unused,
1167                                FILE *fp)
1168 {
1169         const char *delimiter = "# pmu mappings: ";
1170         char *str, *tmp;
1171         u32 pmu_num;
1172         u32 type;
1173
1174         pmu_num = ph->env.nr_pmu_mappings;
1175         if (!pmu_num) {
1176                 fprintf(fp, "# pmu mappings: not available\n");
1177                 return;
1178         }
1179
1180         str = ph->env.pmu_mappings;
1181
1182         while (pmu_num) {
1183                 type = strtoul(str, &tmp, 0);
1184                 if (*tmp != ':')
1185                         goto error;
1186
1187                 str = tmp + 1;
1188                 fprintf(fp, "%s%s = %" PRIu32, delimiter, str, type);
1189
1190                 delimiter = ", ";
1191                 str += strlen(str) + 1;
1192                 pmu_num--;
1193         }
1194
1195         fprintf(fp, "\n");
1196
1197         if (!pmu_num)
1198                 return;
1199 error:
1200         fprintf(fp, "# pmu mappings: unable to read\n");
1201 }
1202
1203 static void print_group_desc(struct perf_header *ph, int fd __maybe_unused,
1204                              FILE *fp)
1205 {
1206         struct perf_session *session;
1207         struct perf_evsel *evsel;
1208         u32 nr = 0;
1209
1210         session = container_of(ph, struct perf_session, header);
1211
1212         evlist__for_each(session->evlist, evsel) {
1213                 if (perf_evsel__is_group_leader(evsel) &&
1214                     evsel->nr_members > 1) {
1215                         fprintf(fp, "# group: %s{%s", evsel->group_name ?: "",
1216                                 perf_evsel__name(evsel));
1217
1218                         nr = evsel->nr_members - 1;
1219                 } else if (nr) {
1220                         fprintf(fp, ",%s", perf_evsel__name(evsel));
1221
1222                         if (--nr == 0)
1223                                 fprintf(fp, "}\n");
1224                 }
1225         }
1226 }
1227
1228 static int __event_process_build_id(struct build_id_event *bev,
1229                                     char *filename,
1230                                     struct perf_session *session)
1231 {
1232         int err = -1;
1233         struct dsos *dsos;
1234         struct machine *machine;
1235         u16 misc;
1236         struct dso *dso;
1237         enum dso_kernel_type dso_type;
1238
1239         machine = perf_session__findnew_machine(session, bev->pid);
1240         if (!machine)
1241                 goto out;
1242
1243         misc = bev->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
1244
1245         switch (misc) {
1246         case PERF_RECORD_MISC_KERNEL:
1247                 dso_type = DSO_TYPE_KERNEL;
1248                 dsos = &machine->kernel_dsos;
1249                 break;
1250         case PERF_RECORD_MISC_GUEST_KERNEL:
1251                 dso_type = DSO_TYPE_GUEST_KERNEL;
1252                 dsos = &machine->kernel_dsos;
1253                 break;
1254         case PERF_RECORD_MISC_USER:
1255         case PERF_RECORD_MISC_GUEST_USER:
1256                 dso_type = DSO_TYPE_USER;
1257                 dsos = &machine->user_dsos;
1258                 break;
1259         default:
1260                 goto out;
1261         }
1262
1263         dso = __dsos__findnew(dsos, filename);
1264         if (dso != NULL) {
1265                 char sbuild_id[BUILD_ID_SIZE * 2 + 1];
1266
1267                 dso__set_build_id(dso, &bev->build_id);
1268
1269                 if (!is_kernel_module(filename, NULL))
1270                         dso->kernel = dso_type;
1271
1272                 build_id__sprintf(dso->build_id, sizeof(dso->build_id),
1273                                   sbuild_id);
1274                 pr_debug("build id event received for %s: %s\n",
1275                          dso->long_name, sbuild_id);
1276         }
1277
1278         err = 0;
1279 out:
1280         return err;
1281 }
1282
1283 static int perf_header__read_build_ids_abi_quirk(struct perf_header *header,
1284                                                  int input, u64 offset, u64 size)
1285 {
1286         struct perf_session *session = container_of(header, struct perf_session, header);
1287         struct {
1288                 struct perf_event_header   header;
1289                 u8                         build_id[PERF_ALIGN(BUILD_ID_SIZE, sizeof(u64))];
1290                 char                       filename[0];
1291         } old_bev;
1292         struct build_id_event bev;
1293         char filename[PATH_MAX];
1294         u64 limit = offset + size;
1295
1296         while (offset < limit) {
1297                 ssize_t len;
1298
1299                 if (readn(input, &old_bev, sizeof(old_bev)) != sizeof(old_bev))
1300                         return -1;
1301
1302                 if (header->needs_swap)
1303                         perf_event_header__bswap(&old_bev.header);
1304
1305                 len = old_bev.header.size - sizeof(old_bev);
1306                 if (readn(input, filename, len) != len)
1307                         return -1;
1308
1309                 bev.header = old_bev.header;
1310
1311                 /*
1312                  * As the pid is the missing value, we need to fill
1313                  * it properly. The header.misc value give us nice hint.
1314                  */
1315                 bev.pid = HOST_KERNEL_ID;
1316                 if (bev.header.misc == PERF_RECORD_MISC_GUEST_USER ||
1317                     bev.header.misc == PERF_RECORD_MISC_GUEST_KERNEL)
1318                         bev.pid = DEFAULT_GUEST_KERNEL_ID;
1319
1320                 memcpy(bev.build_id, old_bev.build_id, sizeof(bev.build_id));
1321                 __event_process_build_id(&bev, filename, session);
1322
1323                 offset += bev.header.size;
1324         }
1325
1326         return 0;
1327 }
1328
1329 static int perf_header__read_build_ids(struct perf_header *header,
1330                                        int input, u64 offset, u64 size)
1331 {
1332         struct perf_session *session = container_of(header, struct perf_session, header);
1333         struct build_id_event bev;
1334         char filename[PATH_MAX];
1335         u64 limit = offset + size, orig_offset = offset;
1336         int err = -1;
1337
1338         while (offset < limit) {
1339                 ssize_t len;
1340
1341                 if (readn(input, &bev, sizeof(bev)) != sizeof(bev))
1342                         goto out;
1343
1344                 if (header->needs_swap)
1345                         perf_event_header__bswap(&bev.header);
1346
1347                 len = bev.header.size - sizeof(bev);
1348                 if (readn(input, filename, len) != len)
1349                         goto out;
1350                 /*
1351                  * The a1645ce1 changeset:
1352                  *
1353                  * "perf: 'perf kvm' tool for monitoring guest performance from host"
1354                  *
1355                  * Added a field to struct build_id_event that broke the file
1356                  * format.
1357                  *
1358                  * Since the kernel build-id is the first entry, process the
1359                  * table using the old format if the well known
1360                  * '[kernel.kallsyms]' string for the kernel build-id has the
1361                  * first 4 characters chopped off (where the pid_t sits).
1362                  */
1363                 if (memcmp(filename, "nel.kallsyms]", 13) == 0) {
1364                         if (lseek(input, orig_offset, SEEK_SET) == (off_t)-1)
1365                                 return -1;
1366                         return perf_header__read_build_ids_abi_quirk(header, input, offset, size);
1367                 }
1368
1369                 __event_process_build_id(&bev, filename, session);
1370
1371                 offset += bev.header.size;
1372         }
1373         err = 0;
1374 out:
1375         return err;
1376 }
1377
1378 static int process_tracing_data(struct perf_file_section *section __maybe_unused,
1379                                 struct perf_header *ph __maybe_unused,
1380                                 int fd, void *data)
1381 {
1382         ssize_t ret = trace_report(fd, data, false);
1383         return ret < 0 ? -1 : 0;
1384 }
1385
1386 static int process_build_id(struct perf_file_section *section,
1387                             struct perf_header *ph, int fd,
1388                             void *data __maybe_unused)
1389 {
1390         if (perf_header__read_build_ids(ph, fd, section->offset, section->size))
1391                 pr_debug("Failed to read buildids, continuing...\n");
1392         return 0;
1393 }
1394
1395 static int process_hostname(struct perf_file_section *section __maybe_unused,
1396                             struct perf_header *ph, int fd,
1397                             void *data __maybe_unused)
1398 {
1399         ph->env.hostname = do_read_string(fd, ph);
1400         return ph->env.hostname ? 0 : -ENOMEM;
1401 }
1402
1403 static int process_osrelease(struct perf_file_section *section __maybe_unused,
1404                              struct perf_header *ph, int fd,
1405                              void *data __maybe_unused)
1406 {
1407         ph->env.os_release = do_read_string(fd, ph);
1408         return ph->env.os_release ? 0 : -ENOMEM;
1409 }
1410
1411 static int process_version(struct perf_file_section *section __maybe_unused,
1412                            struct perf_header *ph, int fd,
1413                            void *data __maybe_unused)
1414 {
1415         ph->env.version = do_read_string(fd, ph);
1416         return ph->env.version ? 0 : -ENOMEM;
1417 }
1418
1419 static int process_arch(struct perf_file_section *section __maybe_unused,
1420                         struct perf_header *ph, int fd,
1421                         void *data __maybe_unused)
1422 {
1423         ph->env.arch = do_read_string(fd, ph);
1424         return ph->env.arch ? 0 : -ENOMEM;
1425 }
1426
1427 static int process_nrcpus(struct perf_file_section *section __maybe_unused,
1428                           struct perf_header *ph, int fd,
1429                           void *data __maybe_unused)
1430 {
1431         ssize_t ret;
1432         u32 nr;
1433
1434         ret = readn(fd, &nr, sizeof(nr));
1435         if (ret != sizeof(nr))
1436                 return -1;
1437
1438         if (ph->needs_swap)
1439                 nr = bswap_32(nr);
1440
1441         ph->env.nr_cpus_online = nr;
1442
1443         ret = readn(fd, &nr, sizeof(nr));
1444         if (ret != sizeof(nr))
1445                 return -1;
1446
1447         if (ph->needs_swap)
1448                 nr = bswap_32(nr);
1449
1450         ph->env.nr_cpus_avail = nr;
1451         return 0;
1452 }
1453
1454 static int process_cpudesc(struct perf_file_section *section __maybe_unused,
1455                            struct perf_header *ph, int fd,
1456                            void *data __maybe_unused)
1457 {
1458         ph->env.cpu_desc = do_read_string(fd, ph);
1459         return ph->env.cpu_desc ? 0 : -ENOMEM;
1460 }
1461
1462 static int process_cpuid(struct perf_file_section *section __maybe_unused,
1463                          struct perf_header *ph,  int fd,
1464                          void *data __maybe_unused)
1465 {
1466         ph->env.cpuid = do_read_string(fd, ph);
1467         return ph->env.cpuid ? 0 : -ENOMEM;
1468 }
1469
1470 static int process_total_mem(struct perf_file_section *section __maybe_unused,
1471                              struct perf_header *ph, int fd,
1472                              void *data __maybe_unused)
1473 {
1474         uint64_t mem;
1475         ssize_t ret;
1476
1477         ret = readn(fd, &mem, sizeof(mem));
1478         if (ret != sizeof(mem))
1479                 return -1;
1480
1481         if (ph->needs_swap)
1482                 mem = bswap_64(mem);
1483
1484         ph->env.total_mem = mem;
1485         return 0;
1486 }
1487
1488 static struct perf_evsel *
1489 perf_evlist__find_by_index(struct perf_evlist *evlist, int idx)
1490 {
1491         struct perf_evsel *evsel;
1492
1493         evlist__for_each(evlist, evsel) {
1494                 if (evsel->idx == idx)
1495                         return evsel;
1496         }
1497
1498         return NULL;
1499 }
1500
1501 static void
1502 perf_evlist__set_event_name(struct perf_evlist *evlist,
1503                             struct perf_evsel *event)
1504 {
1505         struct perf_evsel *evsel;
1506
1507         if (!event->name)
1508                 return;
1509
1510         evsel = perf_evlist__find_by_index(evlist, event->idx);
1511         if (!evsel)
1512                 return;
1513
1514         if (evsel->name)
1515                 return;
1516
1517         evsel->name = strdup(event->name);
1518 }
1519
1520 static int
1521 process_event_desc(struct perf_file_section *section __maybe_unused,
1522                    struct perf_header *header, int fd,
1523                    void *data __maybe_unused)
1524 {
1525         struct perf_session *session;
1526         struct perf_evsel *evsel, *events = read_event_desc(header, fd);
1527
1528         if (!events)
1529                 return 0;
1530
1531         session = container_of(header, struct perf_session, header);
1532         for (evsel = events; evsel->attr.size; evsel++)
1533                 perf_evlist__set_event_name(session->evlist, evsel);
1534
1535         free_event_desc(events);
1536
1537         return 0;
1538 }
1539
1540 static int process_cmdline(struct perf_file_section *section __maybe_unused,
1541                            struct perf_header *ph, int fd,
1542                            void *data __maybe_unused)
1543 {
1544         ssize_t ret;
1545         char *str;
1546         u32 nr, i;
1547         struct strbuf sb;
1548
1549         ret = readn(fd, &nr, sizeof(nr));
1550         if (ret != sizeof(nr))
1551                 return -1;
1552
1553         if (ph->needs_swap)
1554                 nr = bswap_32(nr);
1555
1556         ph->env.nr_cmdline = nr;
1557         strbuf_init(&sb, 128);
1558
1559         for (i = 0; i < nr; i++) {
1560                 str = do_read_string(fd, ph);
1561                 if (!str)
1562                         goto error;
1563
1564                 /* include a NULL character at the end */
1565                 strbuf_add(&sb, str, strlen(str) + 1);
1566                 free(str);
1567         }
1568         ph->env.cmdline = strbuf_detach(&sb, NULL);
1569         return 0;
1570
1571 error:
1572         strbuf_release(&sb);
1573         return -1;
1574 }
1575
1576 static int process_cpu_topology(struct perf_file_section *section __maybe_unused,
1577                                 struct perf_header *ph, int fd,
1578                                 void *data __maybe_unused)
1579 {
1580         ssize_t ret;
1581         u32 nr, i;
1582         char *str;
1583         struct strbuf sb;
1584
1585         ret = readn(fd, &nr, sizeof(nr));
1586         if (ret != sizeof(nr))
1587                 return -1;
1588
1589         if (ph->needs_swap)
1590                 nr = bswap_32(nr);
1591
1592         ph->env.nr_sibling_cores = nr;
1593         strbuf_init(&sb, 128);
1594
1595         for (i = 0; i < nr; i++) {
1596                 str = do_read_string(fd, ph);
1597                 if (!str)
1598                         goto error;
1599
1600                 /* include a NULL character at the end */
1601                 strbuf_add(&sb, str, strlen(str) + 1);
1602                 free(str);
1603         }
1604         ph->env.sibling_cores = strbuf_detach(&sb, NULL);
1605
1606         ret = readn(fd, &nr, sizeof(nr));
1607         if (ret != sizeof(nr))
1608                 return -1;
1609
1610         if (ph->needs_swap)
1611                 nr = bswap_32(nr);
1612
1613         ph->env.nr_sibling_threads = nr;
1614
1615         for (i = 0; i < nr; i++) {
1616                 str = do_read_string(fd, ph);
1617                 if (!str)
1618                         goto error;
1619
1620                 /* include a NULL character at the end */
1621                 strbuf_add(&sb, str, strlen(str) + 1);
1622                 free(str);
1623         }
1624         ph->env.sibling_threads = strbuf_detach(&sb, NULL);
1625         return 0;
1626
1627 error:
1628         strbuf_release(&sb);
1629         return -1;
1630 }
1631
1632 static int process_numa_topology(struct perf_file_section *section __maybe_unused,
1633                                  struct perf_header *ph, int fd,
1634                                  void *data __maybe_unused)
1635 {
1636         ssize_t ret;
1637         u32 nr, node, i;
1638         char *str;
1639         uint64_t mem_total, mem_free;
1640         struct strbuf sb;
1641
1642         /* nr nodes */
1643         ret = readn(fd, &nr, sizeof(nr));
1644         if (ret != sizeof(nr))
1645                 goto error;
1646
1647         if (ph->needs_swap)
1648                 nr = bswap_32(nr);
1649
1650         ph->env.nr_numa_nodes = nr;
1651         strbuf_init(&sb, 256);
1652
1653         for (i = 0; i < nr; i++) {
1654                 /* node number */
1655                 ret = readn(fd, &node, sizeof(node));
1656                 if (ret != sizeof(node))
1657                         goto error;
1658
1659                 ret = readn(fd, &mem_total, sizeof(u64));
1660                 if (ret != sizeof(u64))
1661                         goto error;
1662
1663                 ret = readn(fd, &mem_free, sizeof(u64));
1664                 if (ret != sizeof(u64))
1665                         goto error;
1666
1667                 if (ph->needs_swap) {
1668                         node = bswap_32(node);
1669                         mem_total = bswap_64(mem_total);
1670                         mem_free = bswap_64(mem_free);
1671                 }
1672
1673                 strbuf_addf(&sb, "%u:%"PRIu64":%"PRIu64":",
1674                             node, mem_total, mem_free);
1675
1676                 str = do_read_string(fd, ph);
1677                 if (!str)
1678                         goto error;
1679
1680                 /* include a NULL character at the end */
1681                 strbuf_add(&sb, str, strlen(str) + 1);
1682                 free(str);
1683         }
1684         ph->env.numa_nodes = strbuf_detach(&sb, NULL);
1685         return 0;
1686
1687 error:
1688         strbuf_release(&sb);
1689         return -1;
1690 }
1691
1692 static int process_pmu_mappings(struct perf_file_section *section __maybe_unused,
1693                                 struct perf_header *ph, int fd,
1694                                 void *data __maybe_unused)
1695 {
1696         ssize_t ret;
1697         char *name;
1698         u32 pmu_num;
1699         u32 type;
1700         struct strbuf sb;
1701
1702         ret = readn(fd, &pmu_num, sizeof(pmu_num));
1703         if (ret != sizeof(pmu_num))
1704                 return -1;
1705
1706         if (ph->needs_swap)
1707                 pmu_num = bswap_32(pmu_num);
1708
1709         if (!pmu_num) {
1710                 pr_debug("pmu mappings not available\n");
1711                 return 0;
1712         }
1713
1714         ph->env.nr_pmu_mappings = pmu_num;
1715         strbuf_init(&sb, 128);
1716
1717         while (pmu_num) {
1718                 if (readn(fd, &type, sizeof(type)) != sizeof(type))
1719                         goto error;
1720                 if (ph->needs_swap)
1721                         type = bswap_32(type);
1722
1723                 name = do_read_string(fd, ph);
1724                 if (!name)
1725                         goto error;
1726
1727                 strbuf_addf(&sb, "%u:%s", type, name);
1728                 /* include a NULL character at the end */
1729                 strbuf_add(&sb, "", 1);
1730
1731                 free(name);
1732                 pmu_num--;
1733         }
1734         ph->env.pmu_mappings = strbuf_detach(&sb, NULL);
1735         return 0;
1736
1737 error:
1738         strbuf_release(&sb);
1739         return -1;
1740 }
1741
1742 static int process_group_desc(struct perf_file_section *section __maybe_unused,
1743                               struct perf_header *ph, int fd,
1744                               void *data __maybe_unused)
1745 {
1746         size_t ret = -1;
1747         u32 i, nr, nr_groups;
1748         struct perf_session *session;
1749         struct perf_evsel *evsel, *leader = NULL;
1750         struct group_desc {
1751                 char *name;
1752                 u32 leader_idx;
1753                 u32 nr_members;
1754         } *desc;
1755
1756         if (readn(fd, &nr_groups, sizeof(nr_groups)) != sizeof(nr_groups))
1757                 return -1;
1758
1759         if (ph->needs_swap)
1760                 nr_groups = bswap_32(nr_groups);
1761
1762         ph->env.nr_groups = nr_groups;
1763         if (!nr_groups) {
1764                 pr_debug("group desc not available\n");
1765                 return 0;
1766         }
1767
1768         desc = calloc(nr_groups, sizeof(*desc));
1769         if (!desc)
1770                 return -1;
1771
1772         for (i = 0; i < nr_groups; i++) {
1773                 desc[i].name = do_read_string(fd, ph);
1774                 if (!desc[i].name)
1775                         goto out_free;
1776
1777                 if (readn(fd, &desc[i].leader_idx, sizeof(u32)) != sizeof(u32))
1778                         goto out_free;
1779
1780                 if (readn(fd, &desc[i].nr_members, sizeof(u32)) != sizeof(u32))
1781                         goto out_free;
1782
1783                 if (ph->needs_swap) {
1784                         desc[i].leader_idx = bswap_32(desc[i].leader_idx);
1785                         desc[i].nr_members = bswap_32(desc[i].nr_members);
1786                 }
1787         }
1788
1789         /*
1790          * Rebuild group relationship based on the group_desc
1791          */
1792         session = container_of(ph, struct perf_session, header);
1793         session->evlist->nr_groups = nr_groups;
1794
1795         i = nr = 0;
1796         evlist__for_each(session->evlist, evsel) {
1797                 if (evsel->idx == (int) desc[i].leader_idx) {
1798                         evsel->leader = evsel;
1799                         /* {anon_group} is a dummy name */
1800                         if (strcmp(desc[i].name, "{anon_group}")) {
1801                                 evsel->group_name = desc[i].name;
1802                                 desc[i].name = NULL;
1803                         }
1804                         evsel->nr_members = desc[i].nr_members;
1805
1806                         if (i >= nr_groups || nr > 0) {
1807                                 pr_debug("invalid group desc\n");
1808                                 goto out_free;
1809                         }
1810
1811                         leader = evsel;
1812                         nr = evsel->nr_members - 1;
1813                         i++;
1814                 } else if (nr) {
1815                         /* This is a group member */
1816                         evsel->leader = leader;
1817
1818                         nr--;
1819                 }
1820         }
1821
1822         if (i != nr_groups || nr != 0) {
1823                 pr_debug("invalid group desc\n");
1824                 goto out_free;
1825         }
1826
1827         ret = 0;
1828 out_free:
1829         for (i = 0; i < nr_groups; i++)
1830                 zfree(&desc[i].name);
1831         free(desc);
1832
1833         return ret;
1834 }
1835
1836 struct feature_ops {
1837         int (*write)(int fd, struct perf_header *h, struct perf_evlist *evlist);
1838         void (*print)(struct perf_header *h, int fd, FILE *fp);
1839         int (*process)(struct perf_file_section *section,
1840                        struct perf_header *h, int fd, void *data);
1841         const char *name;
1842         bool full_only;
1843 };
1844
1845 #define FEAT_OPA(n, func) \
1846         [n] = { .name = #n, .write = write_##func, .print = print_##func }
1847 #define FEAT_OPP(n, func) \
1848         [n] = { .name = #n, .write = write_##func, .print = print_##func, \
1849                 .process = process_##func }
1850 #define FEAT_OPF(n, func) \
1851         [n] = { .name = #n, .write = write_##func, .print = print_##func, \
1852                 .process = process_##func, .full_only = true }
1853
1854 /* feature_ops not implemented: */
1855 #define print_tracing_data      NULL
1856 #define print_build_id          NULL
1857
1858 static const struct feature_ops feat_ops[HEADER_LAST_FEATURE] = {
1859         FEAT_OPP(HEADER_TRACING_DATA,   tracing_data),
1860         FEAT_OPP(HEADER_BUILD_ID,       build_id),
1861         FEAT_OPP(HEADER_HOSTNAME,       hostname),
1862         FEAT_OPP(HEADER_OSRELEASE,      osrelease),
1863         FEAT_OPP(HEADER_VERSION,        version),
1864         FEAT_OPP(HEADER_ARCH,           arch),
1865         FEAT_OPP(HEADER_NRCPUS,         nrcpus),
1866         FEAT_OPP(HEADER_CPUDESC,        cpudesc),
1867         FEAT_OPP(HEADER_CPUID,          cpuid),
1868         FEAT_OPP(HEADER_TOTAL_MEM,      total_mem),
1869         FEAT_OPP(HEADER_EVENT_DESC,     event_desc),
1870         FEAT_OPP(HEADER_CMDLINE,        cmdline),
1871         FEAT_OPF(HEADER_CPU_TOPOLOGY,   cpu_topology),
1872         FEAT_OPF(HEADER_NUMA_TOPOLOGY,  numa_topology),
1873         FEAT_OPA(HEADER_BRANCH_STACK,   branch_stack),
1874         FEAT_OPP(HEADER_PMU_MAPPINGS,   pmu_mappings),
1875         FEAT_OPP(HEADER_GROUP_DESC,     group_desc),
1876 };
1877
1878 struct header_print_data {
1879         FILE *fp;
1880         bool full; /* extended list of headers */
1881 };
1882
1883 static int perf_file_section__fprintf_info(struct perf_file_section *section,
1884                                            struct perf_header *ph,
1885                                            int feat, int fd, void *data)
1886 {
1887         struct header_print_data *hd = data;
1888
1889         if (lseek(fd, section->offset, SEEK_SET) == (off_t)-1) {
1890                 pr_debug("Failed to lseek to %" PRIu64 " offset for feature "
1891                                 "%d, continuing...\n", section->offset, feat);
1892                 return 0;
1893         }
1894         if (feat >= HEADER_LAST_FEATURE) {
1895                 pr_warning("unknown feature %d\n", feat);
1896                 return 0;
1897         }
1898         if (!feat_ops[feat].print)
1899                 return 0;
1900
1901         if (!feat_ops[feat].full_only || hd->full)
1902                 feat_ops[feat].print(ph, fd, hd->fp);
1903         else
1904                 fprintf(hd->fp, "# %s info available, use -I to display\n",
1905                         feat_ops[feat].name);
1906
1907         return 0;
1908 }
1909
1910 int perf_header__fprintf_info(struct perf_session *session, FILE *fp, bool full)
1911 {
1912         struct header_print_data hd;
1913         struct perf_header *header = &session->header;
1914         int fd = perf_data_file__fd(session->file);
1915         hd.fp = fp;
1916         hd.full = full;
1917
1918         perf_header__process_sections(header, fd, &hd,
1919                                       perf_file_section__fprintf_info);
1920         return 0;
1921 }
1922
1923 static int do_write_feat(int fd, struct perf_header *h, int type,
1924                          struct perf_file_section **p,
1925                          struct perf_evlist *evlist)
1926 {
1927         int err;
1928         int ret = 0;
1929
1930         if (perf_header__has_feat(h, type)) {
1931                 if (!feat_ops[type].write)
1932                         return -1;
1933
1934                 (*p)->offset = lseek(fd, 0, SEEK_CUR);
1935
1936                 err = feat_ops[type].write(fd, h, evlist);
1937                 if (err < 0) {
1938                         pr_debug("failed to write feature %d\n", type);
1939
1940                         /* undo anything written */
1941                         lseek(fd, (*p)->offset, SEEK_SET);
1942
1943                         return -1;
1944                 }
1945                 (*p)->size = lseek(fd, 0, SEEK_CUR) - (*p)->offset;
1946                 (*p)++;
1947         }
1948         return ret;
1949 }
1950
1951 static int perf_header__adds_write(struct perf_header *header,
1952                                    struct perf_evlist *evlist, int fd)
1953 {
1954         int nr_sections;
1955         struct perf_file_section *feat_sec, *p;
1956         int sec_size;
1957         u64 sec_start;
1958         int feat;
1959         int err;
1960
1961         nr_sections = bitmap_weight(header->adds_features, HEADER_FEAT_BITS);
1962         if (!nr_sections)
1963                 return 0;
1964
1965         feat_sec = p = calloc(nr_sections, sizeof(*feat_sec));
1966         if (feat_sec == NULL)
1967                 return -ENOMEM;
1968
1969         sec_size = sizeof(*feat_sec) * nr_sections;
1970
1971         sec_start = header->feat_offset;
1972         lseek(fd, sec_start + sec_size, SEEK_SET);
1973
1974         for_each_set_bit(feat, header->adds_features, HEADER_FEAT_BITS) {
1975                 if (do_write_feat(fd, header, feat, &p, evlist))
1976                         perf_header__clear_feat(header, feat);
1977         }
1978
1979         lseek(fd, sec_start, SEEK_SET);
1980         /*
1981          * may write more than needed due to dropped feature, but
1982          * this is okay, reader will skip the mising entries
1983          */
1984         err = do_write(fd, feat_sec, sec_size);
1985         if (err < 0)
1986                 pr_debug("failed to write feature section\n");
1987         free(feat_sec);
1988         return err;
1989 }
1990
1991 int perf_header__write_pipe(int fd)
1992 {
1993         struct perf_pipe_file_header f_header;
1994         int err;
1995
1996         f_header = (struct perf_pipe_file_header){
1997                 .magic     = PERF_MAGIC,
1998                 .size      = sizeof(f_header),
1999         };
2000
2001         err = do_write(fd, &f_header, sizeof(f_header));
2002         if (err < 0) {
2003                 pr_debug("failed to write perf pipe header\n");
2004                 return err;
2005         }
2006
2007         return 0;
2008 }
2009
2010 int perf_session__write_header(struct perf_session *session,
2011                                struct perf_evlist *evlist,
2012                                int fd, bool at_exit)
2013 {
2014         struct perf_file_header f_header;
2015         struct perf_file_attr   f_attr;
2016         struct perf_header *header = &session->header;
2017         struct perf_evsel *evsel;
2018         u64 attr_offset;
2019         int err;
2020
2021         lseek(fd, sizeof(f_header), SEEK_SET);
2022
2023         evlist__for_each(session->evlist, evsel) {
2024                 evsel->id_offset = lseek(fd, 0, SEEK_CUR);
2025                 err = do_write(fd, evsel->id, evsel->ids * sizeof(u64));
2026                 if (err < 0) {
2027                         pr_debug("failed to write perf header\n");
2028                         return err;
2029                 }
2030         }
2031
2032         attr_offset = lseek(fd, 0, SEEK_CUR);
2033
2034         evlist__for_each(evlist, evsel) {
2035                 f_attr = (struct perf_file_attr){
2036                         .attr = evsel->attr,
2037                         .ids  = {
2038                                 .offset = evsel->id_offset,
2039                                 .size   = evsel->ids * sizeof(u64),
2040                         }
2041                 };
2042                 err = do_write(fd, &f_attr, sizeof(f_attr));
2043                 if (err < 0) {
2044                         pr_debug("failed to write perf header attribute\n");
2045                         return err;
2046                 }
2047         }
2048
2049         if (!header->data_offset)
2050                 header->data_offset = lseek(fd, 0, SEEK_CUR);
2051         header->feat_offset = header->data_offset + header->data_size;
2052
2053         if (at_exit) {
2054                 err = perf_header__adds_write(header, evlist, fd);
2055                 if (err < 0)
2056                         return err;
2057         }
2058
2059         f_header = (struct perf_file_header){
2060                 .magic     = PERF_MAGIC,
2061                 .size      = sizeof(f_header),
2062                 .attr_size = sizeof(f_attr),
2063                 .attrs = {
2064                         .offset = attr_offset,
2065                         .size   = evlist->nr_entries * sizeof(f_attr),
2066                 },
2067                 .data = {
2068                         .offset = header->data_offset,
2069                         .size   = header->data_size,
2070                 },
2071                 /* event_types is ignored, store zeros */
2072         };
2073
2074         memcpy(&f_header.adds_features, &header->adds_features, sizeof(header->adds_features));
2075
2076         lseek(fd, 0, SEEK_SET);
2077         err = do_write(fd, &f_header, sizeof(f_header));
2078         if (err < 0) {
2079                 pr_debug("failed to write perf header\n");
2080                 return err;
2081         }
2082         lseek(fd, header->data_offset + header->data_size, SEEK_SET);
2083
2084         return 0;
2085 }
2086
2087 static int perf_header__getbuffer64(struct perf_header *header,
2088                                     int fd, void *buf, size_t size)
2089 {
2090         if (readn(fd, buf, size) <= 0)
2091                 return -1;
2092
2093         if (header->needs_swap)
2094                 mem_bswap_64(buf, size);
2095
2096         return 0;
2097 }
2098
2099 int perf_header__process_sections(struct perf_header *header, int fd,
2100                                   void *data,
2101                                   int (*process)(struct perf_file_section *section,
2102                                                  struct perf_header *ph,
2103                                                  int feat, int fd, void *data))
2104 {
2105         struct perf_file_section *feat_sec, *sec;
2106         int nr_sections;
2107         int sec_size;
2108         int feat;
2109         int err;
2110
2111         nr_sections = bitmap_weight(header->adds_features, HEADER_FEAT_BITS);
2112         if (!nr_sections)
2113                 return 0;
2114
2115         feat_sec = sec = calloc(nr_sections, sizeof(*feat_sec));
2116         if (!feat_sec)
2117                 return -1;
2118
2119         sec_size = sizeof(*feat_sec) * nr_sections;
2120
2121         lseek(fd, header->feat_offset, SEEK_SET);
2122
2123         err = perf_header__getbuffer64(header, fd, feat_sec, sec_size);
2124         if (err < 0)
2125                 goto out_free;
2126
2127         for_each_set_bit(feat, header->adds_features, HEADER_LAST_FEATURE) {
2128                 err = process(sec++, header, feat, fd, data);
2129                 if (err < 0)
2130                         goto out_free;
2131         }
2132         err = 0;
2133 out_free:
2134         free(feat_sec);
2135         return err;
2136 }
2137
2138 static const int attr_file_abi_sizes[] = {
2139         [0] = PERF_ATTR_SIZE_VER0,
2140         [1] = PERF_ATTR_SIZE_VER1,
2141         [2] = PERF_ATTR_SIZE_VER2,
2142         [3] = PERF_ATTR_SIZE_VER3,
2143         [4] = PERF_ATTR_SIZE_VER4,
2144         0,
2145 };
2146
2147 /*
2148  * In the legacy file format, the magic number is not used to encode endianness.
2149  * hdr_sz was used to encode endianness. But given that hdr_sz can vary based
2150  * on ABI revisions, we need to try all combinations for all endianness to
2151  * detect the endianness.
2152  */
2153 static int try_all_file_abis(uint64_t hdr_sz, struct perf_header *ph)
2154 {
2155         uint64_t ref_size, attr_size;
2156         int i;
2157
2158         for (i = 0 ; attr_file_abi_sizes[i]; i++) {
2159                 ref_size = attr_file_abi_sizes[i]
2160                          + sizeof(struct perf_file_section);
2161                 if (hdr_sz != ref_size) {
2162                         attr_size = bswap_64(hdr_sz);
2163                         if (attr_size != ref_size)
2164                                 continue;
2165
2166                         ph->needs_swap = true;
2167                 }
2168                 pr_debug("ABI%d perf.data file detected, need_swap=%d\n",
2169                          i,
2170                          ph->needs_swap);
2171                 return 0;
2172         }
2173         /* could not determine endianness */
2174         return -1;
2175 }
2176
2177 #define PERF_PIPE_HDR_VER0      16
2178
2179 static const size_t attr_pipe_abi_sizes[] = {
2180         [0] = PERF_PIPE_HDR_VER0,
2181         0,
2182 };
2183
2184 /*
2185  * In the legacy pipe format, there is an implicit assumption that endiannesss
2186  * between host recording the samples, and host parsing the samples is the
2187  * same. This is not always the case given that the pipe output may always be
2188  * redirected into a file and analyzed on a different machine with possibly a
2189  * different endianness and perf_event ABI revsions in the perf tool itself.
2190  */
2191 static int try_all_pipe_abis(uint64_t hdr_sz, struct perf_header *ph)
2192 {
2193         u64 attr_size;
2194         int i;
2195
2196         for (i = 0 ; attr_pipe_abi_sizes[i]; i++) {
2197                 if (hdr_sz != attr_pipe_abi_sizes[i]) {
2198                         attr_size = bswap_64(hdr_sz);
2199                         if (attr_size != hdr_sz)
2200                                 continue;
2201
2202                         ph->needs_swap = true;
2203                 }
2204                 pr_debug("Pipe ABI%d perf.data file detected\n", i);
2205                 return 0;
2206         }
2207         return -1;
2208 }
2209
2210 bool is_perf_magic(u64 magic)
2211 {
2212         if (!memcmp(&magic, __perf_magic1, sizeof(magic))
2213                 || magic == __perf_magic2
2214                 || magic == __perf_magic2_sw)
2215                 return true;
2216
2217         return false;
2218 }
2219
2220 static int check_magic_endian(u64 magic, uint64_t hdr_sz,
2221                               bool is_pipe, struct perf_header *ph)
2222 {
2223         int ret;
2224
2225         /* check for legacy format */
2226         ret = memcmp(&magic, __perf_magic1, sizeof(magic));
2227         if (ret == 0) {
2228                 ph->version = PERF_HEADER_VERSION_1;
2229                 pr_debug("legacy perf.data format\n");
2230                 if (is_pipe)
2231                         return try_all_pipe_abis(hdr_sz, ph);
2232
2233                 return try_all_file_abis(hdr_sz, ph);
2234         }
2235         /*
2236          * the new magic number serves two purposes:
2237          * - unique number to identify actual perf.data files
2238          * - encode endianness of file
2239          */
2240         ph->version = PERF_HEADER_VERSION_2;
2241
2242         /* check magic number with one endianness */
2243         if (magic == __perf_magic2)
2244                 return 0;
2245
2246         /* check magic number with opposite endianness */
2247         if (magic != __perf_magic2_sw)
2248                 return -1;
2249
2250         ph->needs_swap = true;
2251
2252         return 0;
2253 }
2254
2255 int perf_file_header__read(struct perf_file_header *header,
2256                            struct perf_header *ph, int fd)
2257 {
2258         ssize_t ret;
2259
2260         lseek(fd, 0, SEEK_SET);
2261
2262         ret = readn(fd, header, sizeof(*header));
2263         if (ret <= 0)
2264                 return -1;
2265
2266         if (check_magic_endian(header->magic,
2267                                header->attr_size, false, ph) < 0) {
2268                 pr_debug("magic/endian check failed\n");
2269                 return -1;
2270         }
2271
2272         if (ph->needs_swap) {
2273                 mem_bswap_64(header, offsetof(struct perf_file_header,
2274                              adds_features));
2275         }
2276
2277         if (header->size != sizeof(*header)) {
2278                 /* Support the previous format */
2279                 if (header->size == offsetof(typeof(*header), adds_features))
2280                         bitmap_zero(header->adds_features, HEADER_FEAT_BITS);
2281                 else
2282                         return -1;
2283         } else if (ph->needs_swap) {
2284                 /*
2285                  * feature bitmap is declared as an array of unsigned longs --
2286                  * not good since its size can differ between the host that
2287                  * generated the data file and the host analyzing the file.
2288                  *
2289                  * We need to handle endianness, but we don't know the size of
2290                  * the unsigned long where the file was generated. Take a best
2291                  * guess at determining it: try 64-bit swap first (ie., file
2292                  * created on a 64-bit host), and check if the hostname feature
2293                  * bit is set (this feature bit is forced on as of fbe96f2).
2294                  * If the bit is not, undo the 64-bit swap and try a 32-bit
2295                  * swap. If the hostname bit is still not set (e.g., older data
2296                  * file), punt and fallback to the original behavior --
2297                  * clearing all feature bits and setting buildid.
2298                  */
2299                 mem_bswap_64(&header->adds_features,
2300                             BITS_TO_U64(HEADER_FEAT_BITS));
2301
2302                 if (!test_bit(HEADER_HOSTNAME, header->adds_features)) {
2303                         /* unswap as u64 */
2304                         mem_bswap_64(&header->adds_features,
2305                                     BITS_TO_U64(HEADER_FEAT_BITS));
2306
2307                         /* unswap as u32 */
2308                         mem_bswap_32(&header->adds_features,
2309                                     BITS_TO_U32(HEADER_FEAT_BITS));
2310                 }
2311
2312                 if (!test_bit(HEADER_HOSTNAME, header->adds_features)) {
2313                         bitmap_zero(header->adds_features, HEADER_FEAT_BITS);
2314                         set_bit(HEADER_BUILD_ID, header->adds_features);
2315                 }
2316         }
2317
2318         memcpy(&ph->adds_features, &header->adds_features,
2319                sizeof(ph->adds_features));
2320
2321         ph->data_offset  = header->data.offset;
2322         ph->data_size    = header->data.size;
2323         ph->feat_offset  = header->data.offset + header->data.size;
2324         return 0;
2325 }
2326
2327 static int perf_file_section__process(struct perf_file_section *section,
2328                                       struct perf_header *ph,
2329                                       int feat, int fd, void *data)
2330 {
2331         if (lseek(fd, section->offset, SEEK_SET) == (off_t)-1) {
2332                 pr_debug("Failed to lseek to %" PRIu64 " offset for feature "
2333                           "%d, continuing...\n", section->offset, feat);
2334                 return 0;
2335         }
2336
2337         if (feat >= HEADER_LAST_FEATURE) {
2338                 pr_debug("unknown feature %d, continuing...\n", feat);
2339                 return 0;
2340         }
2341
2342         if (!feat_ops[feat].process)
2343                 return 0;
2344
2345         return feat_ops[feat].process(section, ph, fd, data);
2346 }
2347
2348 static int perf_file_header__read_pipe(struct perf_pipe_file_header *header,
2349                                        struct perf_header *ph, int fd,
2350                                        bool repipe)
2351 {
2352         ssize_t ret;
2353
2354         ret = readn(fd, header, sizeof(*header));
2355         if (ret <= 0)
2356                 return -1;
2357
2358         if (check_magic_endian(header->magic, header->size, true, ph) < 0) {
2359                 pr_debug("endian/magic failed\n");
2360                 return -1;
2361         }
2362
2363         if (ph->needs_swap)
2364                 header->size = bswap_64(header->size);
2365
2366         if (repipe && do_write(STDOUT_FILENO, header, sizeof(*header)) < 0)
2367                 return -1;
2368
2369         return 0;
2370 }
2371
2372 static int perf_header__read_pipe(struct perf_session *session)
2373 {
2374         struct perf_header *header = &session->header;
2375         struct perf_pipe_file_header f_header;
2376
2377         if (perf_file_header__read_pipe(&f_header, header,
2378                                         perf_data_file__fd(session->file),
2379                                         session->repipe) < 0) {
2380                 pr_debug("incompatible file format\n");
2381                 return -EINVAL;
2382         }
2383
2384         return 0;
2385 }
2386
2387 static int read_attr(int fd, struct perf_header *ph,
2388                      struct perf_file_attr *f_attr)
2389 {
2390         struct perf_event_attr *attr = &f_attr->attr;
2391         size_t sz, left;
2392         size_t our_sz = sizeof(f_attr->attr);
2393         ssize_t ret;
2394
2395         memset(f_attr, 0, sizeof(*f_attr));
2396
2397         /* read minimal guaranteed structure */
2398         ret = readn(fd, attr, PERF_ATTR_SIZE_VER0);
2399         if (ret <= 0) {
2400                 pr_debug("cannot read %d bytes of header attr\n",
2401                          PERF_ATTR_SIZE_VER0);
2402                 return -1;
2403         }
2404
2405         /* on file perf_event_attr size */
2406         sz = attr->size;
2407
2408         if (ph->needs_swap)
2409                 sz = bswap_32(sz);
2410
2411         if (sz == 0) {
2412                 /* assume ABI0 */
2413                 sz =  PERF_ATTR_SIZE_VER0;
2414         } else if (sz > our_sz) {
2415                 pr_debug("file uses a more recent and unsupported ABI"
2416                          " (%zu bytes extra)\n", sz - our_sz);
2417                 return -1;
2418         }
2419         /* what we have not yet read and that we know about */
2420         left = sz - PERF_ATTR_SIZE_VER0;
2421         if (left) {
2422                 void *ptr = attr;
2423                 ptr += PERF_ATTR_SIZE_VER0;
2424
2425                 ret = readn(fd, ptr, left);
2426         }
2427         /* read perf_file_section, ids are read in caller */
2428         ret = readn(fd, &f_attr->ids, sizeof(f_attr->ids));
2429
2430         return ret <= 0 ? -1 : 0;
2431 }
2432
2433 static int perf_evsel__prepare_tracepoint_event(struct perf_evsel *evsel,
2434                                                 struct pevent *pevent)
2435 {
2436         struct event_format *event;
2437         char bf[128];
2438
2439         /* already prepared */
2440         if (evsel->tp_format)
2441                 return 0;
2442
2443         if (pevent == NULL) {
2444                 pr_debug("broken or missing trace data\n");
2445                 return -1;
2446         }
2447
2448         event = pevent_find_event(pevent, evsel->attr.config);
2449         if (event == NULL)
2450                 return -1;
2451
2452         if (!evsel->name) {
2453                 snprintf(bf, sizeof(bf), "%s:%s", event->system, event->name);
2454                 evsel->name = strdup(bf);
2455                 if (evsel->name == NULL)
2456                         return -1;
2457         }
2458
2459         evsel->tp_format = event;
2460         return 0;
2461 }
2462
2463 static int perf_evlist__prepare_tracepoint_events(struct perf_evlist *evlist,
2464                                                   struct pevent *pevent)
2465 {
2466         struct perf_evsel *pos;
2467
2468         evlist__for_each(evlist, pos) {
2469                 if (pos->attr.type == PERF_TYPE_TRACEPOINT &&
2470                     perf_evsel__prepare_tracepoint_event(pos, pevent))
2471                         return -1;
2472         }
2473
2474         return 0;
2475 }
2476
2477 int perf_session__read_header(struct perf_session *session)
2478 {
2479         struct perf_data_file *file = session->file;
2480         struct perf_header *header = &session->header;
2481         struct perf_file_header f_header;
2482         struct perf_file_attr   f_attr;
2483         u64                     f_id;
2484         int nr_attrs, nr_ids, i, j;
2485         int fd = perf_data_file__fd(file);
2486
2487         session->evlist = perf_evlist__new();
2488         if (session->evlist == NULL)
2489                 return -ENOMEM;
2490
2491         if (perf_data_file__is_pipe(file))
2492                 return perf_header__read_pipe(session);
2493
2494         if (perf_file_header__read(&f_header, header, fd) < 0)
2495                 return -EINVAL;
2496
2497         /*
2498          * Sanity check that perf.data was written cleanly; data size is
2499          * initialized to 0 and updated only if the on_exit function is run.
2500          * If data size is still 0 then the file contains only partial
2501          * information.  Just warn user and process it as much as it can.
2502          */
2503         if (f_header.data.size == 0) {
2504                 pr_warning("WARNING: The %s file's data size field is 0 which is unexpected.\n"
2505                            "Was the 'perf record' command properly terminated?\n",
2506                            file->path);
2507         }
2508
2509         nr_attrs = f_header.attrs.size / f_header.attr_size;
2510         lseek(fd, f_header.attrs.offset, SEEK_SET);
2511
2512         for (i = 0; i < nr_attrs; i++) {
2513                 struct perf_evsel *evsel;
2514                 off_t tmp;
2515
2516                 if (read_attr(fd, header, &f_attr) < 0)
2517                         goto out_errno;
2518
2519                 if (header->needs_swap)
2520                         perf_event__attr_swap(&f_attr.attr);
2521
2522                 tmp = lseek(fd, 0, SEEK_CUR);
2523                 evsel = perf_evsel__new(&f_attr.attr);
2524
2525                 if (evsel == NULL)
2526                         goto out_delete_evlist;
2527
2528                 evsel->needs_swap = header->needs_swap;
2529                 /*
2530                  * Do it before so that if perf_evsel__alloc_id fails, this
2531                  * entry gets purged too at perf_evlist__delete().
2532                  */
2533                 perf_evlist__add(session->evlist, evsel);
2534
2535                 nr_ids = f_attr.ids.size / sizeof(u64);
2536                 /*
2537                  * We don't have the cpu and thread maps on the header, so
2538                  * for allocating the perf_sample_id table we fake 1 cpu and
2539                  * hattr->ids threads.
2540                  */
2541                 if (perf_evsel__alloc_id(evsel, 1, nr_ids))
2542                         goto out_delete_evlist;
2543
2544                 lseek(fd, f_attr.ids.offset, SEEK_SET);
2545
2546                 for (j = 0; j < nr_ids; j++) {
2547                         if (perf_header__getbuffer64(header, fd, &f_id, sizeof(f_id)))
2548                                 goto out_errno;
2549
2550                         perf_evlist__id_add(session->evlist, evsel, 0, j, f_id);
2551                 }
2552
2553                 lseek(fd, tmp, SEEK_SET);
2554         }
2555
2556         symbol_conf.nr_events = nr_attrs;
2557
2558         perf_header__process_sections(header, fd, &session->tevent,
2559                                       perf_file_section__process);
2560
2561         if (perf_evlist__prepare_tracepoint_events(session->evlist,
2562                                                    session->tevent.pevent))
2563                 goto out_delete_evlist;
2564
2565         return 0;
2566 out_errno:
2567         return -errno;
2568
2569 out_delete_evlist:
2570         perf_evlist__delete(session->evlist);
2571         session->evlist = NULL;
2572         return -ENOMEM;
2573 }
2574
2575 int perf_event__synthesize_attr(struct perf_tool *tool,
2576                                 struct perf_event_attr *attr, u32 ids, u64 *id,
2577                                 perf_event__handler_t process)
2578 {
2579         union perf_event *ev;
2580         size_t size;
2581         int err;
2582
2583         size = sizeof(struct perf_event_attr);
2584         size = PERF_ALIGN(size, sizeof(u64));
2585         size += sizeof(struct perf_event_header);
2586         size += ids * sizeof(u64);
2587
2588         ev = malloc(size);
2589
2590         if (ev == NULL)
2591                 return -ENOMEM;
2592
2593         ev->attr.attr = *attr;
2594         memcpy(ev->attr.id, id, ids * sizeof(u64));
2595
2596         ev->attr.header.type = PERF_RECORD_HEADER_ATTR;
2597         ev->attr.header.size = (u16)size;
2598
2599         if (ev->attr.header.size == size)
2600                 err = process(tool, ev, NULL, NULL);
2601         else
2602                 err = -E2BIG;
2603
2604         free(ev);
2605
2606         return err;
2607 }
2608
2609 int perf_event__synthesize_attrs(struct perf_tool *tool,
2610                                    struct perf_session *session,
2611                                    perf_event__handler_t process)
2612 {
2613         struct perf_evsel *evsel;
2614         int err = 0;
2615
2616         evlist__for_each(session->evlist, evsel) {
2617                 err = perf_event__synthesize_attr(tool, &evsel->attr, evsel->ids,
2618                                                   evsel->id, process);
2619                 if (err) {
2620                         pr_debug("failed to create perf header attribute\n");
2621                         return err;
2622                 }
2623         }
2624
2625         return err;
2626 }
2627
2628 int perf_event__process_attr(struct perf_tool *tool __maybe_unused,
2629                              union perf_event *event,
2630                              struct perf_evlist **pevlist)
2631 {
2632         u32 i, ids, n_ids;
2633         struct perf_evsel *evsel;
2634         struct perf_evlist *evlist = *pevlist;
2635
2636         if (evlist == NULL) {
2637                 *pevlist = evlist = perf_evlist__new();
2638                 if (evlist == NULL)
2639                         return -ENOMEM;
2640         }
2641
2642         evsel = perf_evsel__new(&event->attr.attr);
2643         if (evsel == NULL)
2644                 return -ENOMEM;
2645
2646         perf_evlist__add(evlist, evsel);
2647
2648         ids = event->header.size;
2649         ids -= (void *)&event->attr.id - (void *)event;
2650         n_ids = ids / sizeof(u64);
2651         /*
2652          * We don't have the cpu and thread maps on the header, so
2653          * for allocating the perf_sample_id table we fake 1 cpu and
2654          * hattr->ids threads.
2655          */
2656         if (perf_evsel__alloc_id(evsel, 1, n_ids))
2657                 return -ENOMEM;
2658
2659         for (i = 0; i < n_ids; i++) {
2660                 perf_evlist__id_add(evlist, evsel, 0, i, event->attr.id[i]);
2661         }
2662
2663         symbol_conf.nr_events = evlist->nr_entries;
2664
2665         return 0;
2666 }
2667
2668 int perf_event__synthesize_tracing_data(struct perf_tool *tool, int fd,
2669                                         struct perf_evlist *evlist,
2670                                         perf_event__handler_t process)
2671 {
2672         union perf_event ev;
2673         struct tracing_data *tdata;
2674         ssize_t size = 0, aligned_size = 0, padding;
2675         int err __maybe_unused = 0;
2676
2677         /*
2678          * We are going to store the size of the data followed
2679          * by the data contents. Since the fd descriptor is a pipe,
2680          * we cannot seek back to store the size of the data once
2681          * we know it. Instead we:
2682          *
2683          * - write the tracing data to the temp file
2684          * - get/write the data size to pipe
2685          * - write the tracing data from the temp file
2686          *   to the pipe
2687          */
2688         tdata = tracing_data_get(&evlist->entries, fd, true);
2689         if (!tdata)
2690                 return -1;
2691
2692         memset(&ev, 0, sizeof(ev));
2693
2694         ev.tracing_data.header.type = PERF_RECORD_HEADER_TRACING_DATA;
2695         size = tdata->size;
2696         aligned_size = PERF_ALIGN(size, sizeof(u64));
2697         padding = aligned_size - size;
2698         ev.tracing_data.header.size = sizeof(ev.tracing_data);
2699         ev.tracing_data.size = aligned_size;
2700
2701         process(tool, &ev, NULL, NULL);
2702
2703         /*
2704          * The put function will copy all the tracing data
2705          * stored in temp file to the pipe.
2706          */
2707         tracing_data_put(tdata);
2708
2709         write_padded(fd, NULL, 0, padding);
2710
2711         return aligned_size;
2712 }
2713
2714 int perf_event__process_tracing_data(struct perf_tool *tool __maybe_unused,
2715                                      union perf_event *event,
2716                                      struct perf_session *session)
2717 {
2718         ssize_t size_read, padding, size = event->tracing_data.size;
2719         int fd = perf_data_file__fd(session->file);
2720         off_t offset = lseek(fd, 0, SEEK_CUR);
2721         char buf[BUFSIZ];
2722
2723         /* setup for reading amidst mmap */
2724         lseek(fd, offset + sizeof(struct tracing_data_event),
2725               SEEK_SET);
2726
2727         size_read = trace_report(fd, &session->tevent,
2728                                  session->repipe);
2729         padding = PERF_ALIGN(size_read, sizeof(u64)) - size_read;
2730
2731         if (readn(fd, buf, padding) < 0) {
2732                 pr_err("%s: reading input file", __func__);
2733                 return -1;
2734         }
2735         if (session->repipe) {
2736                 int retw = write(STDOUT_FILENO, buf, padding);
2737                 if (retw <= 0 || retw != padding) {
2738                         pr_err("%s: repiping tracing data padding", __func__);
2739                         return -1;
2740                 }
2741         }
2742
2743         if (size_read + padding != size) {
2744                 pr_err("%s: tracing data size mismatch", __func__);
2745                 return -1;
2746         }
2747
2748         perf_evlist__prepare_tracepoint_events(session->evlist,
2749                                                session->tevent.pevent);
2750
2751         return size_read + padding;
2752 }
2753
2754 int perf_event__synthesize_build_id(struct perf_tool *tool,
2755                                     struct dso *pos, u16 misc,
2756                                     perf_event__handler_t process,
2757                                     struct machine *machine)
2758 {
2759         union perf_event ev;
2760         size_t len;
2761         int err = 0;
2762
2763         if (!pos->hit)
2764                 return err;
2765
2766         memset(&ev, 0, sizeof(ev));
2767
2768         len = pos->long_name_len + 1;
2769         len = PERF_ALIGN(len, NAME_ALIGN);
2770         memcpy(&ev.build_id.build_id, pos->build_id, sizeof(pos->build_id));
2771         ev.build_id.header.type = PERF_RECORD_HEADER_BUILD_ID;
2772         ev.build_id.header.misc = misc;
2773         ev.build_id.pid = machine->pid;
2774         ev.build_id.header.size = sizeof(ev.build_id) + len;
2775         memcpy(&ev.build_id.filename, pos->long_name, pos->long_name_len);
2776
2777         err = process(tool, &ev, NULL, machine);
2778
2779         return err;
2780 }
2781
2782 int perf_event__process_build_id(struct perf_tool *tool __maybe_unused,
2783                                  union perf_event *event,
2784                                  struct perf_session *session)
2785 {
2786         __event_process_build_id(&event->build_id,
2787                                  event->build_id.filename,
2788                                  session);
2789         return 0;
2790 }