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 int write_auxtrace(int fd, struct perf_header *h,
 873                           struct perf_evlist *evlist __maybe_unused)
 874 {
 875         struct perf_session *session;
 876         int err;
 877
 878         session = container_of(h, struct perf_session, header);
 879
 880         err = auxtrace_index__write(fd, &session->auxtrace_index);
 881         if (err < 0)
 882                 pr_err("Failed to write auxtrace index\n");
 883         return err;
 884 }
 885
 886 static void print_hostname(struct perf_header *ph, int fd __maybe_unused,
 887                            FILE *fp)
 888 {
 889         fprintf(fp, "# hostname : %s\n", ph->env.hostname);
 890 }
 891
 892 static void print_osrelease(struct perf_header *ph, int fd __maybe_unused,
 893                             FILE *fp)
 894 {
 895         fprintf(fp, "# os release : %s\n", ph->env.os_release);
 896 }
 897
 898 static void print_arch(struct perf_header *ph, int fd __maybe_unused, FILE *fp)
 899 {
 900         fprintf(fp, "# arch : %s\n", ph->env.arch);
 901 }
 902
 903 static void print_cpudesc(struct perf_header *ph, int fd __maybe_unused,
 904                           FILE *fp)
 905 {
 906         fprintf(fp, "# cpudesc : %s\n", ph->env.cpu_desc);
 907 }
 908
 909 static void print_nrcpus(struct perf_header *ph, int fd __maybe_unused,
 910                          FILE *fp)
 911 {
 912         fprintf(fp, "# nrcpus online : %u\n", ph->env.nr_cpus_online);
 913         fprintf(fp, "# nrcpus avail : %u\n", ph->env.nr_cpus_avail);
 914 }
 915
 916 static void print_version(struct perf_header *ph, int fd __maybe_unused,
 917                           FILE *fp)
 918 {
 919         fprintf(fp, "# perf version : %s\n", ph->env.version);
 920 }
 921
 922 static void print_cmdline(struct perf_header *ph, int fd __maybe_unused,
 923                           FILE *fp)
 924 {
 925         int nr, i;
 926
 927         nr = ph->env.nr_cmdline;
 928
 929         fprintf(fp, "# cmdline : ");
 930
 931         for (i = 0; i < nr; i++)
 932                 fprintf(fp, "%s ", ph->env.cmdline_argv[i]);
 933         fputc('\n', fp);
 934 }
 935
 936 static void print_cpu_topology(struct perf_header *ph, int fd __maybe_unused,
 937                                FILE *fp)
 938 {
 939         int nr, i;
 940         char *str;
 941
 942         nr = ph->env.nr_sibling_cores;
 943         str = ph->env.sibling_cores;
 944
 945         for (i = 0; i < nr; i++) {
 946                 fprintf(fp, "# sibling cores   : %s\n", str);
 947                 str += strlen(str) + 1;
 948         }
 949
 950         nr = ph->env.nr_sibling_threads;
 951         str = ph->env.sibling_threads;
 952
 953         for (i = 0; i < nr; i++) {
 954                 fprintf(fp, "# sibling threads : %s\n", str);
 955                 str += strlen(str) + 1;
 956         }
 957 }
 958
 959 static void free_event_desc(struct perf_evsel *events)
 960 {
 961         struct perf_evsel *evsel;
 962
 963         if (!events)
 964                 return;
 965
 966         for (evsel = events; evsel->attr.size; evsel++) {
 967                 zfree(&evsel->name);
 968                 zfree(&evsel->id);
 969         }
 970
 971         free(events);
 972 }
 973
 974 static struct perf_evsel *
 975 read_event_desc(struct perf_header *ph, int fd)
 976 {
 977         struct perf_evsel *evsel, *events = NULL;
 978         u64 *id;
 979         void *buf = NULL;
 980         u32 nre, sz, nr, i, j;
 981         ssize_t ret;
 982         size_t msz;
 983
 984         /* number of events */
 985         ret = readn(fd, &nre, sizeof(nre));
 986         if (ret != (ssize_t)sizeof(nre))
 987                 goto error;
 988
 989         if (ph->needs_swap)
 990                 nre = bswap_32(nre);
 991
 992         ret = readn(fd, &sz, sizeof(sz));
 993         if (ret != (ssize_t)sizeof(sz))
 994                 goto error;
 995
 996         if (ph->needs_swap)
 997                 sz = bswap_32(sz);
 998
 999         /* buffer to hold on file attr struct */
1000         buf = malloc(sz);
1001         if (!buf)
1002                 goto error;
1003
1004         /* the last event terminates with evsel->attr.size == 0: */
1005         events = calloc(nre + 1, sizeof(*events));
1006         if (!events)
1007                 goto error;
1008
1009         msz = sizeof(evsel->attr);
1010         if (sz < msz)
1011                 msz = sz;
1012
1013         for (i = 0, evsel = events; i < nre; evsel++, i++) {
1014                 evsel->idx = i;
1015
1016                 /*
1017                  * must read entire on-file attr struct to
1018                  * sync up with layout.
1019                  */
1020                 ret = readn(fd, buf, sz);
1021                 if (ret != (ssize_t)sz)
1022                         goto error;
1023
1024                 if (ph->needs_swap)
1025                         perf_event__attr_swap(buf);
1026
1027                 memcpy(&evsel->attr, buf, msz);
1028
1029                 ret = readn(fd, &nr, sizeof(nr));
1030                 if (ret != (ssize_t)sizeof(nr))
1031                         goto error;
1032
1033                 if (ph->needs_swap) {
1034                         nr = bswap_32(nr);
1035                         evsel->needs_swap = true;
1036                 }
1037
1038                 evsel->name = do_read_string(fd, ph);
1039
1040                 if (!nr)
1041                         continue;
1042
1043                 id = calloc(nr, sizeof(*id));
1044                 if (!id)
1045                         goto error;
1046                 evsel->ids = nr;
1047                 evsel->id = id;
1048
1049                 for (j = 0 ; j < nr; j++) {
1050                         ret = readn(fd, id, sizeof(*id));
1051                         if (ret != (ssize_t)sizeof(*id))
1052                                 goto error;
1053                         if (ph->needs_swap)
1054                                 *id = bswap_64(*id);
1055                         id++;
1056                 }
1057         }
1058 out:
1059         free(buf);
1060         return events;
1061 error:
1062         free_event_desc(events);
1063         events = NULL;
1064         goto out;
1065 }
1066
1067 static int __desc_attr__fprintf(FILE *fp, const char *name, const char *val,
1068                                 void *priv __attribute__((unused)))
1069 {
1070         return fprintf(fp, ", %s = %s", name, val);
1071 }
1072
1073 static void print_event_desc(struct perf_header *ph, int fd, FILE *fp)
1074 {
1075         struct perf_evsel *evsel, *events = read_event_desc(ph, fd);
1076         u32 j;
1077         u64 *id;
1078
1079         if (!events) {
1080                 fprintf(fp, "# event desc: not available or unable to read\n");
1081                 return;
1082         }
1083
1084         for (evsel = events; evsel->attr.size; evsel++) {
1085                 fprintf(fp, "# event : name = %s, ", evsel->name);
1086
1087                 if (evsel->ids) {
1088                         fprintf(fp, ", id = {");
1089                         for (j = 0, id = evsel->id; j < evsel->ids; j++, id++) {
1090                                 if (j)
1091                                         fputc(',', fp);
1092                                 fprintf(fp, " %"PRIu64, *id);
1093                         }
1094                         fprintf(fp, " }");
1095                 }
1096
1097                 perf_event_attr__fprintf(fp, &evsel->attr, __desc_attr__fprintf, NULL);
1098
1099                 fputc('\n', fp);
1100         }
1101
1102         free_event_desc(events);
1103 }
1104
1105 static void print_total_mem(struct perf_header *ph, int fd __maybe_unused,
1106                             FILE *fp)
1107 {
1108         fprintf(fp, "# total memory : %Lu kB\n", ph->env.total_mem);
1109 }
1110
1111 static void print_numa_topology(struct perf_header *ph, int fd __maybe_unused,
1112                                 FILE *fp)
1113 {
1114         u32 nr, c, i;
1115         char *str, *tmp;
1116         uint64_t mem_total, mem_free;
1117
1118         /* nr nodes */
1119         nr = ph->env.nr_numa_nodes;
1120         str = ph->env.numa_nodes;
1121
1122         for (i = 0; i < nr; i++) {
1123                 /* node number */
1124                 c = strtoul(str, &tmp, 0);
1125                 if (*tmp != ':')
1126                         goto error;
1127
1128                 str = tmp + 1;
1129                 mem_total = strtoull(str, &tmp, 0);
1130                 if (*tmp != ':')
1131                         goto error;
1132
1133                 str = tmp + 1;
1134                 mem_free = strtoull(str, &tmp, 0);
1135                 if (*tmp != ':')
1136                         goto error;
1137
1138                 fprintf(fp, "# node%u meminfo  : total = %"PRIu64" kB,"
1139                             " free = %"PRIu64" kB\n",
1140                         c, mem_total, mem_free);
1141
1142                 str = tmp + 1;
1143                 fprintf(fp, "# node%u cpu list : %s\n", c, str);
1144
1145                 str += strlen(str) + 1;
1146         }
1147         return;
1148 error:
1149         fprintf(fp, "# numa topology : not available\n");
1150 }
1151
1152 static void print_cpuid(struct perf_header *ph, int fd __maybe_unused, FILE *fp)
1153 {
1154         fprintf(fp, "# cpuid : %s\n", ph->env.cpuid);
1155 }
1156
1157 static void print_branch_stack(struct perf_header *ph __maybe_unused,
1158                                int fd __maybe_unused, FILE *fp)
1159 {
1160         fprintf(fp, "# contains samples with branch stack\n");
1161 }
1162
1163 static void print_auxtrace(struct perf_header *ph __maybe_unused,
1164                            int fd __maybe_unused, FILE *fp)
1165 {
1166         fprintf(fp, "# contains AUX area data (e.g. instruction trace)\n");
1167 }
1168
1169 static void print_pmu_mappings(struct perf_header *ph, int fd __maybe_unused,
1170                                FILE *fp)
1171 {
1172         const char *delimiter = "# pmu mappings: ";
1173         char *str, *tmp;
1174         u32 pmu_num;
1175         u32 type;
1176
1177         pmu_num = ph->env.nr_pmu_mappings;
1178         if (!pmu_num) {
1179                 fprintf(fp, "# pmu mappings: not available\n");
1180                 return;
1181         }
1182
1183         str = ph->env.pmu_mappings;
1184
1185         while (pmu_num) {
1186                 type = strtoul(str, &tmp, 0);
1187                 if (*tmp != ':')
1188                         goto error;
1189
1190                 str = tmp + 1;
1191                 fprintf(fp, "%s%s = %" PRIu32, delimiter, str, type);
1192
1193                 delimiter = ", ";
1194                 str += strlen(str) + 1;
1195                 pmu_num--;
1196         }
1197
1198         fprintf(fp, "\n");
1199
1200         if (!pmu_num)
1201                 return;
1202 error:
1203         fprintf(fp, "# pmu mappings: unable to read\n");
1204 }
1205
1206 static void print_group_desc(struct perf_header *ph, int fd __maybe_unused,
1207                              FILE *fp)
1208 {
1209         struct perf_session *session;
1210         struct perf_evsel *evsel;
1211         u32 nr = 0;
1212
1213         session = container_of(ph, struct perf_session, header);
1214
1215         evlist__for_each(session->evlist, evsel) {
1216                 if (perf_evsel__is_group_leader(evsel) &&
1217                     evsel->nr_members > 1) {
1218                         fprintf(fp, "# group: %s{%s", evsel->group_name ?: "",
1219                                 perf_evsel__name(evsel));
1220
1221                         nr = evsel->nr_members - 1;
1222                 } else if (nr) {
1223                         fprintf(fp, ",%s", perf_evsel__name(evsel));
1224
1225                         if (--nr == 0)
1226                                 fprintf(fp, "}\n");
1227                 }
1228         }
1229 }
1230
1231 static int __event_process_build_id(struct build_id_event *bev,
1232                                     char *filename,
1233                                     struct perf_session *session)
1234 {
1235         int err = -1;
1236         struct machine *machine;
1237         u16 cpumode;
1238         struct dso *dso;
1239         enum dso_kernel_type dso_type;
1240
1241         machine = perf_session__findnew_machine(session, bev->pid);
1242         if (!machine)
1243                 goto out;
1244
1245         cpumode = bev->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
1246
1247         switch (cpumode) {
1248         case PERF_RECORD_MISC_KERNEL:
1249                 dso_type = DSO_TYPE_KERNEL;
1250                 break;
1251         case PERF_RECORD_MISC_GUEST_KERNEL:
1252                 dso_type = DSO_TYPE_GUEST_KERNEL;
1253                 break;
1254         case PERF_RECORD_MISC_USER:
1255         case PERF_RECORD_MISC_GUEST_USER:
1256                 dso_type = DSO_TYPE_USER;
1257                 break;
1258         default:
1259                 goto out;
1260         }
1261
1262         dso = machine__findnew_dso(machine, filename);
1263         if (dso != NULL) {
1264                 char sbuild_id[BUILD_ID_SIZE * 2 + 1];
1265
1266                 dso__set_build_id(dso, &bev->build_id);
1267
1268                 if (!is_kernel_module(filename, cpumode))
1269                         dso->kernel = dso_type;
1270
1271                 build_id__sprintf(dso->build_id, sizeof(dso->build_id),
1272                                   sbuild_id);
1273                 pr_debug("build id event received for %s: %s\n",
1274                          dso->long_name, sbuild_id);
1275                 dso__put(dso);
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_avail = 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_online = 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,
1541                            struct perf_header *ph, int fd,
1542                            void *data __maybe_unused)
1543 {
1544         ssize_t ret;
1545         char *str, *cmdline = NULL, **argv = NULL;
1546         u32 nr, i, len = 0;
1547
1548         ret = readn(fd, &nr, sizeof(nr));
1549         if (ret != sizeof(nr))
1550                 return -1;
1551
1552         if (ph->needs_swap)
1553                 nr = bswap_32(nr);
1554
1555         ph->env.nr_cmdline = nr;
1556
1557         cmdline = zalloc(section->size + nr + 1);
1558         if (!cmdline)
1559                 return -1;
1560
1561         argv = zalloc(sizeof(char *) * (nr + 1));
1562         if (!argv)
1563                 goto error;
1564
1565         for (i = 0; i < nr; i++) {
1566                 str = do_read_string(fd, ph);
1567                 if (!str)
1568                         goto error;
1569
1570                 argv[i] = cmdline + len;
1571                 memcpy(argv[i], str, strlen(str) + 1);
1572                 len += strlen(str) + 1;
1573                 free(str);
1574         }
1575         ph->env.cmdline = cmdline;
1576         ph->env.cmdline_argv = (const char **) argv;
1577         return 0;
1578
1579 error:
1580         free(argv);
1581         free(cmdline);
1582         return -1;
1583 }
1584
1585 static int process_cpu_topology(struct perf_file_section *section __maybe_unused,
1586                                 struct perf_header *ph, int fd,
1587                                 void *data __maybe_unused)
1588 {
1589         ssize_t ret;
1590         u32 nr, i;
1591         char *str;
1592         struct strbuf sb;
1593
1594         ret = readn(fd, &nr, sizeof(nr));
1595         if (ret != sizeof(nr))
1596                 return -1;
1597
1598         if (ph->needs_swap)
1599                 nr = bswap_32(nr);
1600
1601         ph->env.nr_sibling_cores = nr;
1602         strbuf_init(&sb, 128);
1603
1604         for (i = 0; i < nr; i++) {
1605                 str = do_read_string(fd, ph);
1606                 if (!str)
1607                         goto error;
1608
1609                 /* include a NULL character at the end */
1610                 strbuf_add(&sb, str, strlen(str) + 1);
1611                 free(str);
1612         }
1613         ph->env.sibling_cores = strbuf_detach(&sb, NULL);
1614
1615         ret = readn(fd, &nr, sizeof(nr));
1616         if (ret != sizeof(nr))
1617                 return -1;
1618
1619         if (ph->needs_swap)
1620                 nr = bswap_32(nr);
1621
1622         ph->env.nr_sibling_threads = nr;
1623
1624         for (i = 0; i < nr; i++) {
1625                 str = do_read_string(fd, ph);
1626                 if (!str)
1627                         goto error;
1628
1629                 /* include a NULL character at the end */
1630                 strbuf_add(&sb, str, strlen(str) + 1);
1631                 free(str);
1632         }
1633         ph->env.sibling_threads = strbuf_detach(&sb, NULL);
1634         return 0;
1635
1636 error:
1637         strbuf_release(&sb);
1638         return -1;
1639 }
1640
1641 static int process_numa_topology(struct perf_file_section *section __maybe_unused,
1642                                  struct perf_header *ph, int fd,
1643                                  void *data __maybe_unused)
1644 {
1645         ssize_t ret;
1646         u32 nr, node, i;
1647         char *str;
1648         uint64_t mem_total, mem_free;
1649         struct strbuf sb;
1650
1651         /* nr nodes */
1652         ret = readn(fd, &nr, sizeof(nr));
1653         if (ret != sizeof(nr))
1654                 goto error;
1655
1656         if (ph->needs_swap)
1657                 nr = bswap_32(nr);
1658
1659         ph->env.nr_numa_nodes = nr;
1660         strbuf_init(&sb, 256);
1661
1662         for (i = 0; i < nr; i++) {
1663                 /* node number */
1664                 ret = readn(fd, &node, sizeof(node));
1665                 if (ret != sizeof(node))
1666                         goto error;
1667
1668                 ret = readn(fd, &mem_total, sizeof(u64));
1669                 if (ret != sizeof(u64))
1670                         goto error;
1671
1672                 ret = readn(fd, &mem_free, sizeof(u64));
1673                 if (ret != sizeof(u64))
1674                         goto error;
1675
1676                 if (ph->needs_swap) {
1677                         node = bswap_32(node);
1678                         mem_total = bswap_64(mem_total);
1679                         mem_free = bswap_64(mem_free);
1680                 }
1681
1682                 strbuf_addf(&sb, "%u:%"PRIu64":%"PRIu64":",
1683                             node, mem_total, mem_free);
1684
1685                 str = do_read_string(fd, ph);
1686                 if (!str)
1687                         goto error;
1688
1689                 /* include a NULL character at the end */
1690                 strbuf_add(&sb, str, strlen(str) + 1);
1691                 free(str);
1692         }
1693         ph->env.numa_nodes = strbuf_detach(&sb, NULL);
1694         return 0;
1695
1696 error:
1697         strbuf_release(&sb);
1698         return -1;
1699 }
1700
1701 static int process_pmu_mappings(struct perf_file_section *section __maybe_unused,
1702                                 struct perf_header *ph, int fd,
1703                                 void *data __maybe_unused)
1704 {
1705         ssize_t ret;
1706         char *name;
1707         u32 pmu_num;
1708         u32 type;
1709         struct strbuf sb;
1710
1711         ret = readn(fd, &pmu_num, sizeof(pmu_num));
1712         if (ret != sizeof(pmu_num))
1713                 return -1;
1714
1715         if (ph->needs_swap)
1716                 pmu_num = bswap_32(pmu_num);
1717
1718         if (!pmu_num) {
1719                 pr_debug("pmu mappings not available\n");
1720                 return 0;
1721         }
1722
1723         ph->env.nr_pmu_mappings = pmu_num;
1724         strbuf_init(&sb, 128);
1725
1726         while (pmu_num) {
1727                 if (readn(fd, &type, sizeof(type)) != sizeof(type))
1728                         goto error;
1729                 if (ph->needs_swap)
1730                         type = bswap_32(type);
1731
1732                 name = do_read_string(fd, ph);
1733                 if (!name)
1734                         goto error;
1735
1736                 strbuf_addf(&sb, "%u:%s", type, name);
1737                 /* include a NULL character at the end */
1738                 strbuf_add(&sb, "", 1);
1739
1740                 free(name);
1741                 pmu_num--;
1742         }
1743         ph->env.pmu_mappings = strbuf_detach(&sb, NULL);
1744         return 0;
1745
1746 error:
1747         strbuf_release(&sb);
1748         return -1;
1749 }
1750
1751 static int process_group_desc(struct perf_file_section *section __maybe_unused,
1752                               struct perf_header *ph, int fd,
1753                               void *data __maybe_unused)
1754 {
1755         size_t ret = -1;
1756         u32 i, nr, nr_groups;
1757         struct perf_session *session;
1758         struct perf_evsel *evsel, *leader = NULL;
1759         struct group_desc {
1760                 char *name;
1761                 u32 leader_idx;
1762                 u32 nr_members;
1763         } *desc;
1764
1765         if (readn(fd, &nr_groups, sizeof(nr_groups)) != sizeof(nr_groups))
1766                 return -1;
1767
1768         if (ph->needs_swap)
1769                 nr_groups = bswap_32(nr_groups);
1770
1771         ph->env.nr_groups = nr_groups;
1772         if (!nr_groups) {
1773                 pr_debug("group desc not available\n");
1774                 return 0;
1775         }
1776
1777         desc = calloc(nr_groups, sizeof(*desc));
1778         if (!desc)
1779                 return -1;
1780
1781         for (i = 0; i < nr_groups; i++) {
1782                 desc[i].name = do_read_string(fd, ph);
1783                 if (!desc[i].name)
1784                         goto out_free;
1785
1786                 if (readn(fd, &desc[i].leader_idx, sizeof(u32)) != sizeof(u32))
1787                         goto out_free;
1788
1789                 if (readn(fd, &desc[i].nr_members, sizeof(u32)) != sizeof(u32))
1790                         goto out_free;
1791
1792                 if (ph->needs_swap) {
1793                         desc[i].leader_idx = bswap_32(desc[i].leader_idx);
1794                         desc[i].nr_members = bswap_32(desc[i].nr_members);
1795                 }
1796         }
1797
1798         /*
1799          * Rebuild group relationship based on the group_desc
1800          */
1801         session = container_of(ph, struct perf_session, header);
1802         session->evlist->nr_groups = nr_groups;
1803
1804         i = nr = 0;
1805         evlist__for_each(session->evlist, evsel) {
1806                 if (evsel->idx == (int) desc[i].leader_idx) {
1807                         evsel->leader = evsel;
1808                         /* {anon_group} is a dummy name */
1809                         if (strcmp(desc[i].name, "{anon_group}")) {
1810                                 evsel->group_name = desc[i].name;
1811                                 desc[i].name = NULL;
1812                         }
1813                         evsel->nr_members = desc[i].nr_members;
1814
1815                         if (i >= nr_groups || nr > 0) {
1816                                 pr_debug("invalid group desc\n");
1817                                 goto out_free;
1818                         }
1819
1820                         leader = evsel;
1821                         nr = evsel->nr_members - 1;
1822                         i++;
1823                 } else if (nr) {
1824                         /* This is a group member */
1825                         evsel->leader = leader;
1826
1827                         nr--;
1828                 }
1829         }
1830
1831         if (i != nr_groups || nr != 0) {
1832                 pr_debug("invalid group desc\n");
1833                 goto out_free;
1834         }
1835
1836         ret = 0;
1837 out_free:
1838         for (i = 0; i < nr_groups; i++)
1839                 zfree(&desc[i].name);
1840         free(desc);
1841
1842         return ret;
1843 }
1844
1845 static int process_auxtrace(struct perf_file_section *section,
1846                             struct perf_header *ph, int fd,
1847                             void *data __maybe_unused)
1848 {
1849         struct perf_session *session;
1850         int err;
1851
1852         session = container_of(ph, struct perf_session, header);
1853
1854         err = auxtrace_index__process(fd, section->size, session,
1855                                       ph->needs_swap);
1856         if (err < 0)
1857                 pr_err("Failed to process auxtrace index\n");
1858         return err;
1859 }
1860
1861 struct feature_ops {
1862         int (*write)(int fd, struct perf_header *h, struct perf_evlist *evlist);
1863         void (*print)(struct perf_header *h, int fd, FILE *fp);
1864         int (*process)(struct perf_file_section *section,
1865                        struct perf_header *h, int fd, void *data);
1866         const char *name;
1867         bool full_only;
1868 };
1869
1870 #define FEAT_OPA(n, func) \
1871         [n] = { .name = #n, .write = write_##func, .print = print_##func }
1872 #define FEAT_OPP(n, func) \
1873         [n] = { .name = #n, .write = write_##func, .print = print_##func, \
1874                 .process = process_##func }
1875 #define FEAT_OPF(n, func) \
1876         [n] = { .name = #n, .write = write_##func, .print = print_##func, \
1877                 .process = process_##func, .full_only = true }
1878
1879 /* feature_ops not implemented: */
1880 #define print_tracing_data      NULL
1881 #define print_build_id          NULL
1882
1883 static const struct feature_ops feat_ops[HEADER_LAST_FEATURE] = {
1884         FEAT_OPP(HEADER_TRACING_DATA,   tracing_data),
1885         FEAT_OPP(HEADER_BUILD_ID,       build_id),
1886         FEAT_OPP(HEADER_HOSTNAME,       hostname),
1887         FEAT_OPP(HEADER_OSRELEASE,      osrelease),
1888         FEAT_OPP(HEADER_VERSION,        version),
1889         FEAT_OPP(HEADER_ARCH,           arch),
1890         FEAT_OPP(HEADER_NRCPUS,         nrcpus),
1891         FEAT_OPP(HEADER_CPUDESC,        cpudesc),
1892         FEAT_OPP(HEADER_CPUID,          cpuid),
1893         FEAT_OPP(HEADER_TOTAL_MEM,      total_mem),
1894         FEAT_OPP(HEADER_EVENT_DESC,     event_desc),
1895         FEAT_OPP(HEADER_CMDLINE,        cmdline),
1896         FEAT_OPF(HEADER_CPU_TOPOLOGY,   cpu_topology),
1897         FEAT_OPF(HEADER_NUMA_TOPOLOGY,  numa_topology),
1898         FEAT_OPA(HEADER_BRANCH_STACK,   branch_stack),
1899         FEAT_OPP(HEADER_PMU_MAPPINGS,   pmu_mappings),
1900         FEAT_OPP(HEADER_GROUP_DESC,     group_desc),
1901         FEAT_OPP(HEADER_AUXTRACE,       auxtrace),
1902 };
1903
1904 struct header_print_data {
1905         FILE *fp;
1906         bool full; /* extended list of headers */
1907 };
1908
1909 static int perf_file_section__fprintf_info(struct perf_file_section *section,
1910                                            struct perf_header *ph,
1911                                            int feat, int fd, void *data)
1912 {
1913         struct header_print_data *hd = data;
1914
1915         if (lseek(fd, section->offset, SEEK_SET) == (off_t)-1) {
1916                 pr_debug("Failed to lseek to %" PRIu64 " offset for feature "
1917                                 "%d, continuing...\n", section->offset, feat);
1918                 return 0;
1919         }
1920         if (feat >= HEADER_LAST_FEATURE) {
1921                 pr_warning("unknown feature %d\n", feat);
1922                 return 0;
1923         }
1924         if (!feat_ops[feat].print)
1925                 return 0;
1926
1927         if (!feat_ops[feat].full_only || hd->full)
1928                 feat_ops[feat].print(ph, fd, hd->fp);
1929         else
1930                 fprintf(hd->fp, "# %s info available, use -I to display\n",
1931                         feat_ops[feat].name);
1932
1933         return 0;
1934 }
1935
1936 int perf_header__fprintf_info(struct perf_session *session, FILE *fp, bool full)
1937 {
1938         struct header_print_data hd;
1939         struct perf_header *header = &session->header;
1940         int fd = perf_data_file__fd(session->file);
1941         hd.fp = fp;
1942         hd.full = full;
1943
1944         perf_header__process_sections(header, fd, &hd,
1945                                       perf_file_section__fprintf_info);
1946         return 0;
1947 }
1948
1949 static int do_write_feat(int fd, struct perf_header *h, int type,
1950                          struct perf_file_section **p,
1951                          struct perf_evlist *evlist)
1952 {
1953         int err;
1954         int ret = 0;
1955
1956         if (perf_header__has_feat(h, type)) {
1957                 if (!feat_ops[type].write)
1958                         return -1;
1959
1960                 (*p)->offset = lseek(fd, 0, SEEK_CUR);
1961
1962                 err = feat_ops[type].write(fd, h, evlist);
1963                 if (err < 0) {
1964                         pr_debug("failed to write feature %d\n", type);
1965
1966                         /* undo anything written */
1967                         lseek(fd, (*p)->offset, SEEK_SET);
1968
1969                         return -1;
1970                 }
1971                 (*p)->size = lseek(fd, 0, SEEK_CUR) - (*p)->offset;
1972                 (*p)++;
1973         }
1974         return ret;
1975 }
1976
1977 static int perf_header__adds_write(struct perf_header *header,
1978                                    struct perf_evlist *evlist, int fd)
1979 {
1980         int nr_sections;
1981         struct perf_file_section *feat_sec, *p;
1982         int sec_size;
1983         u64 sec_start;
1984         int feat;
1985         int err;
1986
1987         nr_sections = bitmap_weight(header->adds_features, HEADER_FEAT_BITS);
1988         if (!nr_sections)
1989                 return 0;
1990
1991         feat_sec = p = calloc(nr_sections, sizeof(*feat_sec));
1992         if (feat_sec == NULL)
1993                 return -ENOMEM;
1994
1995         sec_size = sizeof(*feat_sec) * nr_sections;
1996
1997         sec_start = header->feat_offset;
1998         lseek(fd, sec_start + sec_size, SEEK_SET);
1999
2000         for_each_set_bit(feat, header->adds_features, HEADER_FEAT_BITS) {
2001                 if (do_write_feat(fd, header, feat, &p, evlist))
2002                         perf_header__clear_feat(header, feat);
2003         }
2004
2005         lseek(fd, sec_start, SEEK_SET);
2006         /*
2007          * may write more than needed due to dropped feature, but
2008          * this is okay, reader will skip the mising entries
2009          */
2010         err = do_write(fd, feat_sec, sec_size);
2011         if (err < 0)
2012                 pr_debug("failed to write feature section\n");
2013         free(feat_sec);
2014         return err;
2015 }
2016
2017 int perf_header__write_pipe(int fd)
2018 {
2019         struct perf_pipe_file_header f_header;
2020         int err;
2021
2022         f_header = (struct perf_pipe_file_header){
2023                 .magic     = PERF_MAGIC,
2024                 .size      = sizeof(f_header),
2025         };
2026
2027         err = do_write(fd, &f_header, sizeof(f_header));
2028         if (err < 0) {
2029                 pr_debug("failed to write perf pipe header\n");
2030                 return err;
2031         }
2032
2033         return 0;
2034 }
2035
2036 int perf_session__write_header(struct perf_session *session,
2037                                struct perf_evlist *evlist,
2038                                int fd, bool at_exit)
2039 {
2040         struct perf_file_header f_header;
2041         struct perf_file_attr   f_attr;
2042         struct perf_header *header = &session->header;
2043         struct perf_evsel *evsel;
2044         u64 attr_offset;
2045         int err;
2046
2047         lseek(fd, sizeof(f_header), SEEK_SET);
2048
2049         evlist__for_each(session->evlist, evsel) {
2050                 evsel->id_offset = lseek(fd, 0, SEEK_CUR);
2051                 err = do_write(fd, evsel->id, evsel->ids * sizeof(u64));
2052                 if (err < 0) {
2053                         pr_debug("failed to write perf header\n");
2054                         return err;
2055                 }
2056         }
2057
2058         attr_offset = lseek(fd, 0, SEEK_CUR);
2059
2060         evlist__for_each(evlist, evsel) {
2061                 f_attr = (struct perf_file_attr){
2062                         .attr = evsel->attr,
2063                         .ids  = {
2064                                 .offset = evsel->id_offset,
2065                                 .size   = evsel->ids * sizeof(u64),
2066                         }
2067                 };
2068                 err = do_write(fd, &f_attr, sizeof(f_attr));
2069                 if (err < 0) {
2070                         pr_debug("failed to write perf header attribute\n");
2071                         return err;
2072                 }
2073         }
2074
2075         if (!header->data_offset)
2076                 header->data_offset = lseek(fd, 0, SEEK_CUR);
2077         header->feat_offset = header->data_offset + header->data_size;
2078
2079         if (at_exit) {
2080                 err = perf_header__adds_write(header, evlist, fd);
2081                 if (err < 0)
2082                         return err;
2083         }
2084
2085         f_header = (struct perf_file_header){
2086                 .magic     = PERF_MAGIC,
2087                 .size      = sizeof(f_header),
2088                 .attr_size = sizeof(f_attr),
2089                 .attrs = {
2090                         .offset = attr_offset,
2091                         .size   = evlist->nr_entries * sizeof(f_attr),
2092                 },
2093                 .data = {
2094                         .offset = header->data_offset,
2095                         .size   = header->data_size,
2096                 },
2097                 /* event_types is ignored, store zeros */
2098         };
2099
2100         memcpy(&f_header.adds_features, &header->adds_features, sizeof(header->adds_features));
2101
2102         lseek(fd, 0, SEEK_SET);
2103         err = do_write(fd, &f_header, sizeof(f_header));
2104         if (err < 0) {
2105                 pr_debug("failed to write perf header\n");
2106                 return err;
2107         }
2108         lseek(fd, header->data_offset + header->data_size, SEEK_SET);
2109
2110         return 0;
2111 }
2112
2113 static int perf_header__getbuffer64(struct perf_header *header,
2114                                     int fd, void *buf, size_t size)
2115 {
2116         if (readn(fd, buf, size) <= 0)
2117                 return -1;
2118
2119         if (header->needs_swap)
2120                 mem_bswap_64(buf, size);
2121
2122         return 0;
2123 }
2124
2125 int perf_header__process_sections(struct perf_header *header, int fd,
2126                                   void *data,
2127                                   int (*process)(struct perf_file_section *section,
2128                                                  struct perf_header *ph,
2129                                                  int feat, int fd, void *data))
2130 {
2131         struct perf_file_section *feat_sec, *sec;
2132         int nr_sections;
2133         int sec_size;
2134         int feat;
2135         int err;
2136
2137         nr_sections = bitmap_weight(header->adds_features, HEADER_FEAT_BITS);
2138         if (!nr_sections)
2139                 return 0;
2140
2141         feat_sec = sec = calloc(nr_sections, sizeof(*feat_sec));
2142         if (!feat_sec)
2143                 return -1;
2144
2145         sec_size = sizeof(*feat_sec) * nr_sections;
2146
2147         lseek(fd, header->feat_offset, SEEK_SET);
2148
2149         err = perf_header__getbuffer64(header, fd, feat_sec, sec_size);
2150         if (err < 0)
2151                 goto out_free;
2152
2153         for_each_set_bit(feat, header->adds_features, HEADER_LAST_FEATURE) {
2154                 err = process(sec++, header, feat, fd, data);
2155                 if (err < 0)
2156                         goto out_free;
2157         }
2158         err = 0;
2159 out_free:
2160         free(feat_sec);
2161         return err;
2162 }
2163
2164 static const int attr_file_abi_sizes[] = {
2165         [0] = PERF_ATTR_SIZE_VER0,
2166         [1] = PERF_ATTR_SIZE_VER1,
2167         [2] = PERF_ATTR_SIZE_VER2,
2168         [3] = PERF_ATTR_SIZE_VER3,
2169         [4] = PERF_ATTR_SIZE_VER4,
2170         0,
2171 };
2172
2173 /*
2174  * In the legacy file format, the magic number is not used to encode endianness.
2175  * hdr_sz was used to encode endianness. But given that hdr_sz can vary based
2176  * on ABI revisions, we need to try all combinations for all endianness to
2177  * detect the endianness.
2178  */
2179 static int try_all_file_abis(uint64_t hdr_sz, struct perf_header *ph)
2180 {
2181         uint64_t ref_size, attr_size;
2182         int i;
2183
2184         for (i = 0 ; attr_file_abi_sizes[i]; i++) {
2185                 ref_size = attr_file_abi_sizes[i]
2186                          + sizeof(struct perf_file_section);
2187                 if (hdr_sz != ref_size) {
2188                         attr_size = bswap_64(hdr_sz);
2189                         if (attr_size != ref_size)
2190                                 continue;
2191
2192                         ph->needs_swap = true;
2193                 }
2194                 pr_debug("ABI%d perf.data file detected, need_swap=%d\n",
2195                          i,
2196                          ph->needs_swap);
2197                 return 0;
2198         }
2199         /* could not determine endianness */
2200         return -1;
2201 }
2202
2203 #define PERF_PIPE_HDR_VER0      16
2204
2205 static const size_t attr_pipe_abi_sizes[] = {
2206         [0] = PERF_PIPE_HDR_VER0,
2207         0,
2208 };
2209
2210 /*
2211  * In the legacy pipe format, there is an implicit assumption that endiannesss
2212  * between host recording the samples, and host parsing the samples is the
2213  * same. This is not always the case given that the pipe output may always be
2214  * redirected into a file and analyzed on a different machine with possibly a
2215  * different endianness and perf_event ABI revsions in the perf tool itself.
2216  */
2217 static int try_all_pipe_abis(uint64_t hdr_sz, struct perf_header *ph)
2218 {
2219         u64 attr_size;
2220         int i;
2221
2222         for (i = 0 ; attr_pipe_abi_sizes[i]; i++) {
2223                 if (hdr_sz != attr_pipe_abi_sizes[i]) {
2224                         attr_size = bswap_64(hdr_sz);
2225                         if (attr_size != hdr_sz)
2226                                 continue;
2227
2228                         ph->needs_swap = true;
2229                 }
2230                 pr_debug("Pipe ABI%d perf.data file detected\n", i);
2231                 return 0;
2232         }
2233         return -1;
2234 }
2235
2236 bool is_perf_magic(u64 magic)
2237 {
2238         if (!memcmp(&magic, __perf_magic1, sizeof(magic))
2239                 || magic == __perf_magic2
2240                 || magic == __perf_magic2_sw)
2241                 return true;
2242
2243         return false;
2244 }
2245
2246 static int check_magic_endian(u64 magic, uint64_t hdr_sz,
2247                               bool is_pipe, struct perf_header *ph)
2248 {
2249         int ret;
2250
2251         /* check for legacy format */
2252         ret = memcmp(&magic, __perf_magic1, sizeof(magic));
2253         if (ret == 0) {
2254                 ph->version = PERF_HEADER_VERSION_1;
2255                 pr_debug("legacy perf.data format\n");
2256                 if (is_pipe)
2257                         return try_all_pipe_abis(hdr_sz, ph);
2258
2259                 return try_all_file_abis(hdr_sz, ph);
2260         }
2261         /*
2262          * the new magic number serves two purposes:
2263          * - unique number to identify actual perf.data files
2264          * - encode endianness of file
2265          */
2266         ph->version = PERF_HEADER_VERSION_2;
2267
2268         /* check magic number with one endianness */
2269         if (magic == __perf_magic2)
2270                 return 0;
2271
2272         /* check magic number with opposite endianness */
2273         if (magic != __perf_magic2_sw)
2274                 return -1;
2275
2276         ph->needs_swap = true;
2277
2278         return 0;
2279 }
2280
2281 int perf_file_header__read(struct perf_file_header *header,
2282                            struct perf_header *ph, int fd)
2283 {
2284         ssize_t ret;
2285
2286         lseek(fd, 0, SEEK_SET);
2287
2288         ret = readn(fd, header, sizeof(*header));
2289         if (ret <= 0)
2290                 return -1;
2291
2292         if (check_magic_endian(header->magic,
2293                                header->attr_size, false, ph) < 0) {
2294                 pr_debug("magic/endian check failed\n");
2295                 return -1;
2296         }
2297
2298         if (ph->needs_swap) {
2299                 mem_bswap_64(header, offsetof(struct perf_file_header,
2300                              adds_features));
2301         }
2302
2303         if (header->size != sizeof(*header)) {
2304                 /* Support the previous format */
2305                 if (header->size == offsetof(typeof(*header), adds_features))
2306                         bitmap_zero(header->adds_features, HEADER_FEAT_BITS);
2307                 else
2308                         return -1;
2309         } else if (ph->needs_swap) {
2310                 /*
2311                  * feature bitmap is declared as an array of unsigned longs --
2312                  * not good since its size can differ between the host that
2313                  * generated the data file and the host analyzing the file.
2314                  *
2315                  * We need to handle endianness, but we don't know the size of
2316                  * the unsigned long where the file was generated. Take a best
2317                  * guess at determining it: try 64-bit swap first (ie., file
2318                  * created on a 64-bit host), and check if the hostname feature
2319                  * bit is set (this feature bit is forced on as of fbe96f2).
2320                  * If the bit is not, undo the 64-bit swap and try a 32-bit
2321                  * swap. If the hostname bit is still not set (e.g., older data
2322                  * file), punt and fallback to the original behavior --
2323                  * clearing all feature bits and setting buildid.
2324                  */
2325                 mem_bswap_64(&header->adds_features,
2326                             BITS_TO_U64(HEADER_FEAT_BITS));
2327
2328                 if (!test_bit(HEADER_HOSTNAME, header->adds_features)) {
2329                         /* unswap as u64 */
2330                         mem_bswap_64(&header->adds_features,
2331                                     BITS_TO_U64(HEADER_FEAT_BITS));
2332
2333                         /* unswap as u32 */
2334                         mem_bswap_32(&header->adds_features,
2335                                     BITS_TO_U32(HEADER_FEAT_BITS));
2336                 }
2337
2338                 if (!test_bit(HEADER_HOSTNAME, header->adds_features)) {
2339                         bitmap_zero(header->adds_features, HEADER_FEAT_BITS);
2340                         set_bit(HEADER_BUILD_ID, header->adds_features);
2341                 }
2342         }
2343
2344         memcpy(&ph->adds_features, &header->adds_features,
2345                sizeof(ph->adds_features));
2346
2347         ph->data_offset  = header->data.offset;
2348         ph->data_size    = header->data.size;
2349         ph->feat_offset  = header->data.offset + header->data.size;
2350         return 0;
2351 }
2352
2353 static int perf_file_section__process(struct perf_file_section *section,
2354                                       struct perf_header *ph,
2355                                       int feat, int fd, void *data)
2356 {
2357         if (lseek(fd, section->offset, SEEK_SET) == (off_t)-1) {
2358                 pr_debug("Failed to lseek to %" PRIu64 " offset for feature "
2359                           "%d, continuing...\n", section->offset, feat);
2360                 return 0;
2361         }
2362
2363         if (feat >= HEADER_LAST_FEATURE) {
2364                 pr_debug("unknown feature %d, continuing...\n", feat);
2365                 return 0;
2366         }
2367
2368         if (!feat_ops[feat].process)
2369                 return 0;
2370
2371         return feat_ops[feat].process(section, ph, fd, data);
2372 }
2373
2374 static int perf_file_header__read_pipe(struct perf_pipe_file_header *header,
2375                                        struct perf_header *ph, int fd,
2376                                        bool repipe)
2377 {
2378         ssize_t ret;
2379
2380         ret = readn(fd, header, sizeof(*header));
2381         if (ret <= 0)
2382                 return -1;
2383
2384         if (check_magic_endian(header->magic, header->size, true, ph) < 0) {
2385                 pr_debug("endian/magic failed\n");
2386                 return -1;
2387         }
2388
2389         if (ph->needs_swap)
2390                 header->size = bswap_64(header->size);
2391
2392         if (repipe && do_write(STDOUT_FILENO, header, sizeof(*header)) < 0)
2393                 return -1;
2394
2395         return 0;
2396 }
2397
2398 static int perf_header__read_pipe(struct perf_session *session)
2399 {
2400         struct perf_header *header = &session->header;
2401         struct perf_pipe_file_header f_header;
2402
2403         if (perf_file_header__read_pipe(&f_header, header,
2404                                         perf_data_file__fd(session->file),
2405                                         session->repipe) < 0) {
2406                 pr_debug("incompatible file format\n");
2407                 return -EINVAL;
2408         }
2409
2410         return 0;
2411 }
2412
2413 static int read_attr(int fd, struct perf_header *ph,
2414                      struct perf_file_attr *f_attr)
2415 {
2416         struct perf_event_attr *attr = &f_attr->attr;
2417         size_t sz, left;
2418         size_t our_sz = sizeof(f_attr->attr);
2419         ssize_t ret;
2420
2421         memset(f_attr, 0, sizeof(*f_attr));
2422
2423         /* read minimal guaranteed structure */
2424         ret = readn(fd, attr, PERF_ATTR_SIZE_VER0);
2425         if (ret <= 0) {
2426                 pr_debug("cannot read %d bytes of header attr\n",
2427                          PERF_ATTR_SIZE_VER0);
2428                 return -1;
2429         }
2430
2431         /* on file perf_event_attr size */
2432         sz = attr->size;
2433
2434         if (ph->needs_swap)
2435                 sz = bswap_32(sz);
2436
2437         if (sz == 0) {
2438                 /* assume ABI0 */
2439                 sz =  PERF_ATTR_SIZE_VER0;
2440         } else if (sz > our_sz) {
2441                 pr_debug("file uses a more recent and unsupported ABI"
2442                          " (%zu bytes extra)\n", sz - our_sz);
2443                 return -1;
2444         }
2445         /* what we have not yet read and that we know about */
2446         left = sz - PERF_ATTR_SIZE_VER0;
2447         if (left) {
2448                 void *ptr = attr;
2449                 ptr += PERF_ATTR_SIZE_VER0;
2450
2451                 ret = readn(fd, ptr, left);
2452         }
2453         /* read perf_file_section, ids are read in caller */
2454         ret = readn(fd, &f_attr->ids, sizeof(f_attr->ids));
2455
2456         return ret <= 0 ? -1 : 0;
2457 }
2458
2459 static int perf_evsel__prepare_tracepoint_event(struct perf_evsel *evsel,
2460                                                 struct pevent *pevent)
2461 {
2462         struct event_format *event;
2463         char bf[128];
2464
2465         /* already prepared */
2466         if (evsel->tp_format)
2467                 return 0;
2468
2469         if (pevent == NULL) {
2470                 pr_debug("broken or missing trace data\n");
2471                 return -1;
2472         }
2473
2474         event = pevent_find_event(pevent, evsel->attr.config);
2475         if (event == NULL)
2476                 return -1;
2477
2478         if (!evsel->name) {
2479                 snprintf(bf, sizeof(bf), "%s:%s", event->system, event->name);
2480                 evsel->name = strdup(bf);
2481                 if (evsel->name == NULL)
2482                         return -1;
2483         }
2484
2485         evsel->tp_format = event;
2486         return 0;
2487 }
2488
2489 static int perf_evlist__prepare_tracepoint_events(struct perf_evlist *evlist,
2490                                                   struct pevent *pevent)
2491 {
2492         struct perf_evsel *pos;
2493
2494         evlist__for_each(evlist, pos) {
2495                 if (pos->attr.type == PERF_TYPE_TRACEPOINT &&
2496                     perf_evsel__prepare_tracepoint_event(pos, pevent))
2497                         return -1;
2498         }
2499
2500         return 0;
2501 }
2502
2503 int perf_session__read_header(struct perf_session *session)
2504 {
2505         struct perf_data_file *file = session->file;
2506         struct perf_header *header = &session->header;
2507         struct perf_file_header f_header;
2508         struct perf_file_attr   f_attr;
2509         u64                     f_id;
2510         int nr_attrs, nr_ids, i, j;
2511         int fd = perf_data_file__fd(file);
2512
2513         session->evlist = perf_evlist__new();
2514         if (session->evlist == NULL)
2515                 return -ENOMEM;
2516
2517         session->evlist->env = &header->env;
2518         if (perf_data_file__is_pipe(file))
2519                 return perf_header__read_pipe(session);
2520
2521         if (perf_file_header__read(&f_header, header, fd) < 0)
2522                 return -EINVAL;
2523
2524         /*
2525          * Sanity check that perf.data was written cleanly; data size is
2526          * initialized to 0 and updated only if the on_exit function is run.
2527          * If data size is still 0 then the file contains only partial
2528          * information.  Just warn user and process it as much as it can.
2529          */
2530         if (f_header.data.size == 0) {
2531                 pr_warning("WARNING: The %s file's data size field is 0 which is unexpected.\n"
2532                            "Was the 'perf record' command properly terminated?\n",
2533                            file->path);
2534         }
2535
2536         nr_attrs = f_header.attrs.size / f_header.attr_size;
2537         lseek(fd, f_header.attrs.offset, SEEK_SET);
2538
2539         for (i = 0; i < nr_attrs; i++) {
2540                 struct perf_evsel *evsel;
2541                 off_t tmp;
2542
2543                 if (read_attr(fd, header, &f_attr) < 0)
2544                         goto out_errno;
2545
2546                 if (header->needs_swap) {
2547                         f_attr.ids.size   = bswap_64(f_attr.ids.size);
2548                         f_attr.ids.offset = bswap_64(f_attr.ids.offset);
2549                         perf_event__attr_swap(&f_attr.attr);
2550                 }
2551
2552                 tmp = lseek(fd, 0, SEEK_CUR);
2553                 evsel = perf_evsel__new(&f_attr.attr);
2554
2555                 if (evsel == NULL)
2556                         goto out_delete_evlist;
2557
2558                 evsel->needs_swap = header->needs_swap;
2559                 /*
2560                  * Do it before so that if perf_evsel__alloc_id fails, this
2561                  * entry gets purged too at perf_evlist__delete().
2562                  */
2563                 perf_evlist__add(session->evlist, evsel);
2564
2565                 nr_ids = f_attr.ids.size / sizeof(u64);
2566                 /*
2567                  * We don't have the cpu and thread maps on the header, so
2568                  * for allocating the perf_sample_id table we fake 1 cpu and
2569                  * hattr->ids threads.
2570                  */
2571                 if (perf_evsel__alloc_id(evsel, 1, nr_ids))
2572                         goto out_delete_evlist;
2573
2574                 lseek(fd, f_attr.ids.offset, SEEK_SET);
2575
2576                 for (j = 0; j < nr_ids; j++) {
2577                         if (perf_header__getbuffer64(header, fd, &f_id, sizeof(f_id)))
2578                                 goto out_errno;
2579
2580                         perf_evlist__id_add(session->evlist, evsel, 0, j, f_id);
2581                 }
2582
2583                 lseek(fd, tmp, SEEK_SET);
2584         }
2585
2586         symbol_conf.nr_events = nr_attrs;
2587
2588         perf_header__process_sections(header, fd, &session->tevent,
2589                                       perf_file_section__process);
2590
2591         if (perf_evlist__prepare_tracepoint_events(session->evlist,
2592                                                    session->tevent.pevent))
2593                 goto out_delete_evlist;
2594
2595         return 0;
2596 out_errno:
2597         return -errno;
2598
2599 out_delete_evlist:
2600         perf_evlist__delete(session->evlist);
2601         session->evlist = NULL;
2602         return -ENOMEM;
2603 }
2604
2605 int perf_event__synthesize_attr(struct perf_tool *tool,
2606                                 struct perf_event_attr *attr, u32 ids, u64 *id,
2607                                 perf_event__handler_t process)
2608 {
2609         union perf_event *ev;
2610         size_t size;
2611         int err;
2612
2613         size = sizeof(struct perf_event_attr);
2614         size = PERF_ALIGN(size, sizeof(u64));
2615         size += sizeof(struct perf_event_header);
2616         size += ids * sizeof(u64);
2617
2618         ev = malloc(size);
2619
2620         if (ev == NULL)
2621                 return -ENOMEM;
2622
2623         ev->attr.attr = *attr;
2624         memcpy(ev->attr.id, id, ids * sizeof(u64));
2625
2626         ev->attr.header.type = PERF_RECORD_HEADER_ATTR;
2627         ev->attr.header.size = (u16)size;
2628
2629         if (ev->attr.header.size == size)
2630                 err = process(tool, ev, NULL, NULL);
2631         else
2632                 err = -E2BIG;
2633
2634         free(ev);
2635
2636         return err;
2637 }
2638
2639 int perf_event__synthesize_attrs(struct perf_tool *tool,
2640                                    struct perf_session *session,
2641                                    perf_event__handler_t process)
2642 {
2643         struct perf_evsel *evsel;
2644         int err = 0;
2645
2646         evlist__for_each(session->evlist, evsel) {
2647                 err = perf_event__synthesize_attr(tool, &evsel->attr, evsel->ids,
2648                                                   evsel->id, process);
2649                 if (err) {
2650                         pr_debug("failed to create perf header attribute\n");
2651                         return err;
2652                 }
2653         }
2654
2655         return err;
2656 }
2657
2658 int perf_event__process_attr(struct perf_tool *tool __maybe_unused,
2659                              union perf_event *event,
2660                              struct perf_evlist **pevlist)
2661 {
2662         u32 i, ids, n_ids;
2663         struct perf_evsel *evsel;
2664         struct perf_evlist *evlist = *pevlist;
2665
2666         if (evlist == NULL) {
2667                 *pevlist = evlist = perf_evlist__new();
2668                 if (evlist == NULL)
2669                         return -ENOMEM;
2670         }
2671
2672         evsel = perf_evsel__new(&event->attr.attr);
2673         if (evsel == NULL)
2674                 return -ENOMEM;
2675
2676         perf_evlist__add(evlist, evsel);
2677
2678         ids = event->header.size;
2679         ids -= (void *)&event->attr.id - (void *)event;
2680         n_ids = ids / sizeof(u64);
2681         /*
2682          * We don't have the cpu and thread maps on the header, so
2683          * for allocating the perf_sample_id table we fake 1 cpu and
2684          * hattr->ids threads.
2685          */
2686         if (perf_evsel__alloc_id(evsel, 1, n_ids))
2687                 return -ENOMEM;
2688
2689         for (i = 0; i < n_ids; i++) {
2690                 perf_evlist__id_add(evlist, evsel, 0, i, event->attr.id[i]);
2691         }
2692
2693         symbol_conf.nr_events = evlist->nr_entries;
2694
2695         return 0;
2696 }
2697
2698 int perf_event__synthesize_tracing_data(struct perf_tool *tool, int fd,
2699                                         struct perf_evlist *evlist,
2700                                         perf_event__handler_t process)
2701 {
2702         union perf_event ev;
2703         struct tracing_data *tdata;
2704         ssize_t size = 0, aligned_size = 0, padding;
2705         int err __maybe_unused = 0;
2706
2707         /*
2708          * We are going to store the size of the data followed
2709          * by the data contents. Since the fd descriptor is a pipe,
2710          * we cannot seek back to store the size of the data once
2711          * we know it. Instead we:
2712          *
2713          * - write the tracing data to the temp file
2714          * - get/write the data size to pipe
2715          * - write the tracing data from the temp file
2716          *   to the pipe
2717          */
2718         tdata = tracing_data_get(&evlist->entries, fd, true);
2719         if (!tdata)
2720                 return -1;
2721
2722         memset(&ev, 0, sizeof(ev));
2723
2724         ev.tracing_data.header.type = PERF_RECORD_HEADER_TRACING_DATA;
2725         size = tdata->size;
2726         aligned_size = PERF_ALIGN(size, sizeof(u64));
2727         padding = aligned_size - size;
2728         ev.tracing_data.header.size = sizeof(ev.tracing_data);
2729         ev.tracing_data.size = aligned_size;
2730
2731         process(tool, &ev, NULL, NULL);
2732
2733         /*
2734          * The put function will copy all the tracing data
2735          * stored in temp file to the pipe.
2736          */
2737         tracing_data_put(tdata);
2738
2739         write_padded(fd, NULL, 0, padding);
2740
2741         return aligned_size;
2742 }
2743
2744 int perf_event__process_tracing_data(struct perf_tool *tool __maybe_unused,
2745                                      union perf_event *event,
2746                                      struct perf_session *session)
2747 {
2748         ssize_t size_read, padding, size = event->tracing_data.size;
2749         int fd = perf_data_file__fd(session->file);
2750         off_t offset = lseek(fd, 0, SEEK_CUR);
2751         char buf[BUFSIZ];
2752
2753         /* setup for reading amidst mmap */
2754         lseek(fd, offset + sizeof(struct tracing_data_event),
2755               SEEK_SET);
2756
2757         size_read = trace_report(fd, &session->tevent,
2758                                  session->repipe);
2759         padding = PERF_ALIGN(size_read, sizeof(u64)) - size_read;
2760
2761         if (readn(fd, buf, padding) < 0) {
2762                 pr_err("%s: reading input file", __func__);
2763                 return -1;
2764         }
2765         if (session->repipe) {
2766                 int retw = write(STDOUT_FILENO, buf, padding);
2767                 if (retw <= 0 || retw != padding) {
2768                         pr_err("%s: repiping tracing data padding", __func__);
2769                         return -1;
2770                 }
2771         }
2772
2773         if (size_read + padding != size) {
2774                 pr_err("%s: tracing data size mismatch", __func__);
2775                 return -1;
2776         }
2777
2778         perf_evlist__prepare_tracepoint_events(session->evlist,
2779                                                session->tevent.pevent);
2780
2781         return size_read + padding;
2782 }
2783
2784 int perf_event__synthesize_build_id(struct perf_tool *tool,
2785                                     struct dso *pos, u16 misc,
2786                                     perf_event__handler_t process,
2787                                     struct machine *machine)
2788 {
2789         union perf_event ev;
2790         size_t len;
2791         int err = 0;
2792
2793         if (!pos->hit)
2794                 return err;
2795
2796         memset(&ev, 0, sizeof(ev));
2797
2798         len = pos->long_name_len + 1;
2799         len = PERF_ALIGN(len, NAME_ALIGN);
2800         memcpy(&ev.build_id.build_id, pos->build_id, sizeof(pos->build_id));
2801         ev.build_id.header.type = PERF_RECORD_HEADER_BUILD_ID;
2802         ev.build_id.header.misc = misc;
2803         ev.build_id.pid = machine->pid;
2804         ev.build_id.header.size = sizeof(ev.build_id) + len;
2805         memcpy(&ev.build_id.filename, pos->long_name, pos->long_name_len);
2806
2807         err = process(tool, &ev, NULL, machine);
2808
2809         return err;
2810 }
2811
2812 int perf_event__process_build_id(struct perf_tool *tool __maybe_unused,
2813                                  union perf_event *event,
2814                                  struct perf_session *session)
2815 {
2816         __event_process_build_id(&event->build_id,
2817                                  event->build_id.filename,
2818                                  session);
2819         return 0;
2820 }