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         char *str;
 927
 928         nr = ph->env.nr_cmdline;
 929         str = ph->env.cmdline;
 930
 931         fprintf(fp, "# cmdline : ");
 932
 933         for (i = 0; i < nr; i++) {
 934                 fprintf(fp, "%s ", str);
 935                 str += strlen(str) + 1;
 936         }
 937         fputc('\n', fp);
 938 }
 939
 940 static void print_cpu_topology(struct perf_header *ph, int fd __maybe_unused,
 941                                FILE *fp)
 942 {
 943         int nr, i;
 944         char *str;
 945
 946         nr = ph->env.nr_sibling_cores;
 947         str = ph->env.sibling_cores;
 948
 949         for (i = 0; i < nr; i++) {
 950                 fprintf(fp, "# sibling cores   : %s\n", str);
 951                 str += strlen(str) + 1;
 952         }
 953
 954         nr = ph->env.nr_sibling_threads;
 955         str = ph->env.sibling_threads;
 956
 957         for (i = 0; i < nr; i++) {
 958                 fprintf(fp, "# sibling threads : %s\n", str);
 959                 str += strlen(str) + 1;
 960         }
 961 }
 962
 963 static void free_event_desc(struct perf_evsel *events)
 964 {
 965         struct perf_evsel *evsel;
 966
 967         if (!events)
 968                 return;
 969
 970         for (evsel = events; evsel->attr.size; evsel++) {
 971                 zfree(&evsel->name);
 972                 zfree(&evsel->id);
 973         }
 974
 975         free(events);
 976 }
 977
 978 static struct perf_evsel *
 979 read_event_desc(struct perf_header *ph, int fd)
 980 {
 981         struct perf_evsel *evsel, *events = NULL;
 982         u64 *id;
 983         void *buf = NULL;
 984         u32 nre, sz, nr, i, j;
 985         ssize_t ret;
 986         size_t msz;
 987
 988         /* number of events */
 989         ret = readn(fd, &nre, sizeof(nre));
 990         if (ret != (ssize_t)sizeof(nre))
 991                 goto error;
 992
 993         if (ph->needs_swap)
 994                 nre = bswap_32(nre);
 995
 996         ret = readn(fd, &sz, sizeof(sz));
 997         if (ret != (ssize_t)sizeof(sz))
 998                 goto error;
 999
1000         if (ph->needs_swap)
1001                 sz = bswap_32(sz);
1002
1003         /* buffer to hold on file attr struct */
1004         buf = malloc(sz);
1005         if (!buf)
1006                 goto error;
1007
1008         /* the last event terminates with evsel->attr.size == 0: */
1009         events = calloc(nre + 1, sizeof(*events));
1010         if (!events)
1011                 goto error;
1012
1013         msz = sizeof(evsel->attr);
1014         if (sz < msz)
1015                 msz = sz;
1016
1017         for (i = 0, evsel = events; i < nre; evsel++, i++) {
1018                 evsel->idx = i;
1019
1020                 /*
1021                  * must read entire on-file attr struct to
1022                  * sync up with layout.
1023                  */
1024                 ret = readn(fd, buf, sz);
1025                 if (ret != (ssize_t)sz)
1026                         goto error;
1027
1028                 if (ph->needs_swap)
1029                         perf_event__attr_swap(buf);
1030
1031                 memcpy(&evsel->attr, buf, msz);
1032
1033                 ret = readn(fd, &nr, sizeof(nr));
1034                 if (ret != (ssize_t)sizeof(nr))
1035                         goto error;
1036
1037                 if (ph->needs_swap) {
1038                         nr = bswap_32(nr);
1039                         evsel->needs_swap = true;
1040                 }
1041
1042                 evsel->name = do_read_string(fd, ph);
1043
1044                 if (!nr)
1045                         continue;
1046
1047                 id = calloc(nr, sizeof(*id));
1048                 if (!id)
1049                         goto error;
1050                 evsel->ids = nr;
1051                 evsel->id = id;
1052
1053                 for (j = 0 ; j < nr; j++) {
1054                         ret = readn(fd, id, sizeof(*id));
1055                         if (ret != (ssize_t)sizeof(*id))
1056                                 goto error;
1057                         if (ph->needs_swap)
1058                                 *id = bswap_64(*id);
1059                         id++;
1060                 }
1061         }
1062 out:
1063         free(buf);
1064         return events;
1065 error:
1066         if (events)
1067                 free_event_desc(events);
1068         events = NULL;
1069         goto out;
1070 }
1071
1072 static int __desc_attr__fprintf(FILE *fp, const char *name, const char *val,
1073                                 void *priv __attribute__((unused)))
1074 {
1075         return fprintf(fp, ", %s = %s", name, val);
1076 }
1077
1078 static void print_event_desc(struct perf_header *ph, int fd, FILE *fp)
1079 {
1080         struct perf_evsel *evsel, *events = read_event_desc(ph, fd);
1081         u32 j;
1082         u64 *id;
1083
1084         if (!events) {
1085                 fprintf(fp, "# event desc: not available or unable to read\n");
1086                 return;
1087         }
1088
1089         for (evsel = events; evsel->attr.size; evsel++) {
1090                 fprintf(fp, "# event : name = %s, ", evsel->name);
1091
1092                 if (evsel->ids) {
1093                         fprintf(fp, ", id = {");
1094                         for (j = 0, id = evsel->id; j < evsel->ids; j++, id++) {
1095                                 if (j)
1096                                         fputc(',', fp);
1097                                 fprintf(fp, " %"PRIu64, *id);
1098                         }
1099                         fprintf(fp, " }");
1100                 }
1101
1102                 perf_event_attr__fprintf(fp, &evsel->attr, __desc_attr__fprintf, NULL);
1103
1104                 fputc('\n', fp);
1105         }
1106
1107         free_event_desc(events);
1108 }
1109
1110 static void print_total_mem(struct perf_header *ph, int fd __maybe_unused,
1111                             FILE *fp)
1112 {
1113         fprintf(fp, "# total memory : %Lu kB\n", ph->env.total_mem);
1114 }
1115
1116 static void print_numa_topology(struct perf_header *ph, int fd __maybe_unused,
1117                                 FILE *fp)
1118 {
1119         u32 nr, c, i;
1120         char *str, *tmp;
1121         uint64_t mem_total, mem_free;
1122
1123         /* nr nodes */
1124         nr = ph->env.nr_numa_nodes;
1125         str = ph->env.numa_nodes;
1126
1127         for (i = 0; i < nr; i++) {
1128                 /* node number */
1129                 c = strtoul(str, &tmp, 0);
1130                 if (*tmp != ':')
1131                         goto error;
1132
1133                 str = tmp + 1;
1134                 mem_total = strtoull(str, &tmp, 0);
1135                 if (*tmp != ':')
1136                         goto error;
1137
1138                 str = tmp + 1;
1139                 mem_free = strtoull(str, &tmp, 0);
1140                 if (*tmp != ':')
1141                         goto error;
1142
1143                 fprintf(fp, "# node%u meminfo  : total = %"PRIu64" kB,"
1144                             " free = %"PRIu64" kB\n",
1145                         c, mem_total, mem_free);
1146
1147                 str = tmp + 1;
1148                 fprintf(fp, "# node%u cpu list : %s\n", c, str);
1149
1150                 str += strlen(str) + 1;
1151         }
1152         return;
1153 error:
1154         fprintf(fp, "# numa topology : not available\n");
1155 }
1156
1157 static void print_cpuid(struct perf_header *ph, int fd __maybe_unused, FILE *fp)
1158 {
1159         fprintf(fp, "# cpuid : %s\n", ph->env.cpuid);
1160 }
1161
1162 static void print_branch_stack(struct perf_header *ph __maybe_unused,
1163                                int fd __maybe_unused, FILE *fp)
1164 {
1165         fprintf(fp, "# contains samples with branch stack\n");
1166 }
1167
1168 static void print_auxtrace(struct perf_header *ph __maybe_unused,
1169                            int fd __maybe_unused, FILE *fp)
1170 {
1171         fprintf(fp, "# contains AUX area data (e.g. instruction trace)\n");
1172 }
1173
1174 static void print_pmu_mappings(struct perf_header *ph, int fd __maybe_unused,
1175                                FILE *fp)
1176 {
1177         const char *delimiter = "# pmu mappings: ";
1178         char *str, *tmp;
1179         u32 pmu_num;
1180         u32 type;
1181
1182         pmu_num = ph->env.nr_pmu_mappings;
1183         if (!pmu_num) {
1184                 fprintf(fp, "# pmu mappings: not available\n");
1185                 return;
1186         }
1187
1188         str = ph->env.pmu_mappings;
1189
1190         while (pmu_num) {
1191                 type = strtoul(str, &tmp, 0);
1192                 if (*tmp != ':')
1193                         goto error;
1194
1195                 str = tmp + 1;
1196                 fprintf(fp, "%s%s = %" PRIu32, delimiter, str, type);
1197
1198                 delimiter = ", ";
1199                 str += strlen(str) + 1;
1200                 pmu_num--;
1201         }
1202
1203         fprintf(fp, "\n");
1204
1205         if (!pmu_num)
1206                 return;
1207 error:
1208         fprintf(fp, "# pmu mappings: unable to read\n");
1209 }
1210
1211 static void print_group_desc(struct perf_header *ph, int fd __maybe_unused,
1212                              FILE *fp)
1213 {
1214         struct perf_session *session;
1215         struct perf_evsel *evsel;
1216         u32 nr = 0;
1217
1218         session = container_of(ph, struct perf_session, header);
1219
1220         evlist__for_each(session->evlist, evsel) {
1221                 if (perf_evsel__is_group_leader(evsel) &&
1222                     evsel->nr_members > 1) {
1223                         fprintf(fp, "# group: %s{%s", evsel->group_name ?: "",
1224                                 perf_evsel__name(evsel));
1225
1226                         nr = evsel->nr_members - 1;
1227                 } else if (nr) {
1228                         fprintf(fp, ",%s", perf_evsel__name(evsel));
1229
1230                         if (--nr == 0)
1231                                 fprintf(fp, "}\n");
1232                 }
1233         }
1234 }
1235
1236 static int __event_process_build_id(struct build_id_event *bev,
1237                                     char *filename,
1238                                     struct perf_session *session)
1239 {
1240         int err = -1;
1241         struct machine *machine;
1242         u16 cpumode;
1243         struct dso *dso;
1244         enum dso_kernel_type dso_type;
1245
1246         machine = perf_session__findnew_machine(session, bev->pid);
1247         if (!machine)
1248                 goto out;
1249
1250         cpumode = bev->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
1251
1252         switch (cpumode) {
1253         case PERF_RECORD_MISC_KERNEL:
1254                 dso_type = DSO_TYPE_KERNEL;
1255                 break;
1256         case PERF_RECORD_MISC_GUEST_KERNEL:
1257                 dso_type = DSO_TYPE_GUEST_KERNEL;
1258                 break;
1259         case PERF_RECORD_MISC_USER:
1260         case PERF_RECORD_MISC_GUEST_USER:
1261                 dso_type = DSO_TYPE_USER;
1262                 break;
1263         default:
1264                 goto out;
1265         }
1266
1267         dso = machine__findnew_dso(machine, filename);
1268         if (dso != NULL) {
1269                 char sbuild_id[BUILD_ID_SIZE * 2 + 1];
1270
1271                 dso__set_build_id(dso, &bev->build_id);
1272
1273                 if (!is_kernel_module(filename, cpumode))
1274                         dso->kernel = dso_type;
1275
1276                 build_id__sprintf(dso->build_id, sizeof(dso->build_id),
1277                                   sbuild_id);
1278                 pr_debug("build id event received for %s: %s\n",
1279                          dso->long_name, sbuild_id);
1280                 dso__put(dso);
1281         }
1282
1283         err = 0;
1284 out:
1285         return err;
1286 }
1287
1288 static int perf_header__read_build_ids_abi_quirk(struct perf_header *header,
1289                                                  int input, u64 offset, u64 size)
1290 {
1291         struct perf_session *session = container_of(header, struct perf_session, header);
1292         struct {
1293                 struct perf_event_header   header;
1294                 u8                         build_id[PERF_ALIGN(BUILD_ID_SIZE, sizeof(u64))];
1295                 char                       filename[0];
1296         } old_bev;
1297         struct build_id_event bev;
1298         char filename[PATH_MAX];
1299         u64 limit = offset + size;
1300
1301         while (offset < limit) {
1302                 ssize_t len;
1303
1304                 if (readn(input, &old_bev, sizeof(old_bev)) != sizeof(old_bev))
1305                         return -1;
1306
1307                 if (header->needs_swap)
1308                         perf_event_header__bswap(&old_bev.header);
1309
1310                 len = old_bev.header.size - sizeof(old_bev);
1311                 if (readn(input, filename, len) != len)
1312                         return -1;
1313
1314                 bev.header = old_bev.header;
1315
1316                 /*
1317                  * As the pid is the missing value, we need to fill
1318                  * it properly. The header.misc value give us nice hint.
1319                  */
1320                 bev.pid = HOST_KERNEL_ID;
1321                 if (bev.header.misc == PERF_RECORD_MISC_GUEST_USER ||
1322                     bev.header.misc == PERF_RECORD_MISC_GUEST_KERNEL)
1323                         bev.pid = DEFAULT_GUEST_KERNEL_ID;
1324
1325                 memcpy(bev.build_id, old_bev.build_id, sizeof(bev.build_id));
1326                 __event_process_build_id(&bev, filename, session);
1327
1328                 offset += bev.header.size;
1329         }
1330
1331         return 0;
1332 }
1333
1334 static int perf_header__read_build_ids(struct perf_header *header,
1335                                        int input, u64 offset, u64 size)
1336 {
1337         struct perf_session *session = container_of(header, struct perf_session, header);
1338         struct build_id_event bev;
1339         char filename[PATH_MAX];
1340         u64 limit = offset + size, orig_offset = offset;
1341         int err = -1;
1342
1343         while (offset < limit) {
1344                 ssize_t len;
1345
1346                 if (readn(input, &bev, sizeof(bev)) != sizeof(bev))
1347                         goto out;
1348
1349                 if (header->needs_swap)
1350                         perf_event_header__bswap(&bev.header);
1351
1352                 len = bev.header.size - sizeof(bev);
1353                 if (readn(input, filename, len) != len)
1354                         goto out;
1355                 /*
1356                  * The a1645ce1 changeset:
1357                  *
1358                  * "perf: 'perf kvm' tool for monitoring guest performance from host"
1359                  *
1360                  * Added a field to struct build_id_event that broke the file
1361                  * format.
1362                  *
1363                  * Since the kernel build-id is the first entry, process the
1364                  * table using the old format if the well known
1365                  * '[kernel.kallsyms]' string for the kernel build-id has the
1366                  * first 4 characters chopped off (where the pid_t sits).
1367                  */
1368                 if (memcmp(filename, "nel.kallsyms]", 13) == 0) {
1369                         if (lseek(input, orig_offset, SEEK_SET) == (off_t)-1)
1370                                 return -1;
1371                         return perf_header__read_build_ids_abi_quirk(header, input, offset, size);
1372                 }
1373
1374                 __event_process_build_id(&bev, filename, session);
1375
1376                 offset += bev.header.size;
1377         }
1378         err = 0;
1379 out:
1380         return err;
1381 }
1382
1383 static int process_tracing_data(struct perf_file_section *section __maybe_unused,
1384                                 struct perf_header *ph __maybe_unused,
1385                                 int fd, void *data)
1386 {
1387         ssize_t ret = trace_report(fd, data, false);
1388         return ret < 0 ? -1 : 0;
1389 }
1390
1391 static int process_build_id(struct perf_file_section *section,
1392                             struct perf_header *ph, int fd,
1393                             void *data __maybe_unused)
1394 {
1395         if (perf_header__read_build_ids(ph, fd, section->offset, section->size))
1396                 pr_debug("Failed to read buildids, continuing...\n");
1397         return 0;
1398 }
1399
1400 static int process_hostname(struct perf_file_section *section __maybe_unused,
1401                             struct perf_header *ph, int fd,
1402                             void *data __maybe_unused)
1403 {
1404         ph->env.hostname = do_read_string(fd, ph);
1405         return ph->env.hostname ? 0 : -ENOMEM;
1406 }
1407
1408 static int process_osrelease(struct perf_file_section *section __maybe_unused,
1409                              struct perf_header *ph, int fd,
1410                              void *data __maybe_unused)
1411 {
1412         ph->env.os_release = do_read_string(fd, ph);
1413         return ph->env.os_release ? 0 : -ENOMEM;
1414 }
1415
1416 static int process_version(struct perf_file_section *section __maybe_unused,
1417                            struct perf_header *ph, int fd,
1418                            void *data __maybe_unused)
1419 {
1420         ph->env.version = do_read_string(fd, ph);
1421         return ph->env.version ? 0 : -ENOMEM;
1422 }
1423
1424 static int process_arch(struct perf_file_section *section __maybe_unused,
1425                         struct perf_header *ph, int fd,
1426                         void *data __maybe_unused)
1427 {
1428         ph->env.arch = do_read_string(fd, ph);
1429         return ph->env.arch ? 0 : -ENOMEM;
1430 }
1431
1432 static int process_nrcpus(struct perf_file_section *section __maybe_unused,
1433                           struct perf_header *ph, int fd,
1434                           void *data __maybe_unused)
1435 {
1436         ssize_t ret;
1437         u32 nr;
1438
1439         ret = readn(fd, &nr, sizeof(nr));
1440         if (ret != sizeof(nr))
1441                 return -1;
1442
1443         if (ph->needs_swap)
1444                 nr = bswap_32(nr);
1445
1446         ph->env.nr_cpus_online = nr;
1447
1448         ret = readn(fd, &nr, sizeof(nr));
1449         if (ret != sizeof(nr))
1450                 return -1;
1451
1452         if (ph->needs_swap)
1453                 nr = bswap_32(nr);
1454
1455         ph->env.nr_cpus_avail = nr;
1456         return 0;
1457 }
1458
1459 static int process_cpudesc(struct perf_file_section *section __maybe_unused,
1460                            struct perf_header *ph, int fd,
1461                            void *data __maybe_unused)
1462 {
1463         ph->env.cpu_desc = do_read_string(fd, ph);
1464         return ph->env.cpu_desc ? 0 : -ENOMEM;
1465 }
1466
1467 static int process_cpuid(struct perf_file_section *section __maybe_unused,
1468                          struct perf_header *ph,  int fd,
1469                          void *data __maybe_unused)
1470 {
1471         ph->env.cpuid = do_read_string(fd, ph);
1472         return ph->env.cpuid ? 0 : -ENOMEM;
1473 }
1474
1475 static int process_total_mem(struct perf_file_section *section __maybe_unused,
1476                              struct perf_header *ph, int fd,
1477                              void *data __maybe_unused)
1478 {
1479         uint64_t mem;
1480         ssize_t ret;
1481
1482         ret = readn(fd, &mem, sizeof(mem));
1483         if (ret != sizeof(mem))
1484                 return -1;
1485
1486         if (ph->needs_swap)
1487                 mem = bswap_64(mem);
1488
1489         ph->env.total_mem = mem;
1490         return 0;
1491 }
1492
1493 static struct perf_evsel *
1494 perf_evlist__find_by_index(struct perf_evlist *evlist, int idx)
1495 {
1496         struct perf_evsel *evsel;
1497
1498         evlist__for_each(evlist, evsel) {
1499                 if (evsel->idx == idx)
1500                         return evsel;
1501         }
1502
1503         return NULL;
1504 }
1505
1506 static void
1507 perf_evlist__set_event_name(struct perf_evlist *evlist,
1508                             struct perf_evsel *event)
1509 {
1510         struct perf_evsel *evsel;
1511
1512         if (!event->name)
1513                 return;
1514
1515         evsel = perf_evlist__find_by_index(evlist, event->idx);
1516         if (!evsel)
1517                 return;
1518
1519         if (evsel->name)
1520                 return;
1521
1522         evsel->name = strdup(event->name);
1523 }
1524
1525 static int
1526 process_event_desc(struct perf_file_section *section __maybe_unused,
1527                    struct perf_header *header, int fd,
1528                    void *data __maybe_unused)
1529 {
1530         struct perf_session *session;
1531         struct perf_evsel *evsel, *events = read_event_desc(header, fd);
1532
1533         if (!events)
1534                 return 0;
1535
1536         session = container_of(header, struct perf_session, header);
1537         for (evsel = events; evsel->attr.size; evsel++)
1538                 perf_evlist__set_event_name(session->evlist, evsel);
1539
1540         free_event_desc(events);
1541
1542         return 0;
1543 }
1544
1545 static int process_cmdline(struct perf_file_section *section __maybe_unused,
1546                            struct perf_header *ph, int fd,
1547                            void *data __maybe_unused)
1548 {
1549         ssize_t ret;
1550         char *str;
1551         u32 nr, i;
1552         struct strbuf sb;
1553
1554         ret = readn(fd, &nr, sizeof(nr));
1555         if (ret != sizeof(nr))
1556                 return -1;
1557
1558         if (ph->needs_swap)
1559                 nr = bswap_32(nr);
1560
1561         ph->env.nr_cmdline = nr;
1562         strbuf_init(&sb, 128);
1563
1564         for (i = 0; i < nr; i++) {
1565                 str = do_read_string(fd, ph);
1566                 if (!str)
1567                         goto error;
1568
1569                 /* include a NULL character at the end */
1570                 strbuf_add(&sb, str, strlen(str) + 1);
1571                 free(str);
1572         }
1573         ph->env.cmdline = strbuf_detach(&sb, NULL);
1574         return 0;
1575
1576 error:
1577         strbuf_release(&sb);
1578         return -1;
1579 }
1580
1581 static int process_cpu_topology(struct perf_file_section *section __maybe_unused,
1582                                 struct perf_header *ph, int fd,
1583                                 void *data __maybe_unused)
1584 {
1585         ssize_t ret;
1586         u32 nr, i;
1587         char *str;
1588         struct strbuf sb;
1589
1590         ret = readn(fd, &nr, sizeof(nr));
1591         if (ret != sizeof(nr))
1592                 return -1;
1593
1594         if (ph->needs_swap)
1595                 nr = bswap_32(nr);
1596
1597         ph->env.nr_sibling_cores = nr;
1598         strbuf_init(&sb, 128);
1599
1600         for (i = 0; i < nr; i++) {
1601                 str = do_read_string(fd, ph);
1602                 if (!str)
1603                         goto error;
1604
1605                 /* include a NULL character at the end */
1606                 strbuf_add(&sb, str, strlen(str) + 1);
1607                 free(str);
1608         }
1609         ph->env.sibling_cores = strbuf_detach(&sb, NULL);
1610
1611         ret = readn(fd, &nr, sizeof(nr));
1612         if (ret != sizeof(nr))
1613                 return -1;
1614
1615         if (ph->needs_swap)
1616                 nr = bswap_32(nr);
1617
1618         ph->env.nr_sibling_threads = nr;
1619
1620         for (i = 0; i < nr; i++) {
1621                 str = do_read_string(fd, ph);
1622                 if (!str)
1623                         goto error;
1624
1625                 /* include a NULL character at the end */
1626                 strbuf_add(&sb, str, strlen(str) + 1);
1627                 free(str);
1628         }
1629         ph->env.sibling_threads = strbuf_detach(&sb, NULL);
1630         return 0;
1631
1632 error:
1633         strbuf_release(&sb);
1634         return -1;
1635 }
1636
1637 static int process_numa_topology(struct perf_file_section *section __maybe_unused,
1638                                  struct perf_header *ph, int fd,
1639                                  void *data __maybe_unused)
1640 {
1641         ssize_t ret;
1642         u32 nr, node, i;
1643         char *str;
1644         uint64_t mem_total, mem_free;
1645         struct strbuf sb;
1646
1647         /* nr nodes */
1648         ret = readn(fd, &nr, sizeof(nr));
1649         if (ret != sizeof(nr))
1650                 goto error;
1651
1652         if (ph->needs_swap)
1653                 nr = bswap_32(nr);
1654
1655         ph->env.nr_numa_nodes = nr;
1656         strbuf_init(&sb, 256);
1657
1658         for (i = 0; i < nr; i++) {
1659                 /* node number */
1660                 ret = readn(fd, &node, sizeof(node));
1661                 if (ret != sizeof(node))
1662                         goto error;
1663
1664                 ret = readn(fd, &mem_total, sizeof(u64));
1665                 if (ret != sizeof(u64))
1666                         goto error;
1667
1668                 ret = readn(fd, &mem_free, sizeof(u64));
1669                 if (ret != sizeof(u64))
1670                         goto error;
1671
1672                 if (ph->needs_swap) {
1673                         node = bswap_32(node);
1674                         mem_total = bswap_64(mem_total);
1675                         mem_free = bswap_64(mem_free);
1676                 }
1677
1678                 strbuf_addf(&sb, "%u:%"PRIu64":%"PRIu64":",
1679                             node, mem_total, mem_free);
1680
1681                 str = do_read_string(fd, ph);
1682                 if (!str)
1683                         goto error;
1684
1685                 /* include a NULL character at the end */
1686                 strbuf_add(&sb, str, strlen(str) + 1);
1687                 free(str);
1688         }
1689         ph->env.numa_nodes = strbuf_detach(&sb, NULL);
1690         return 0;
1691
1692 error:
1693         strbuf_release(&sb);
1694         return -1;
1695 }
1696
1697 static int process_pmu_mappings(struct perf_file_section *section __maybe_unused,
1698                                 struct perf_header *ph, int fd,
1699                                 void *data __maybe_unused)
1700 {
1701         ssize_t ret;
1702         char *name;
1703         u32 pmu_num;
1704         u32 type;
1705         struct strbuf sb;
1706
1707         ret = readn(fd, &pmu_num, sizeof(pmu_num));
1708         if (ret != sizeof(pmu_num))
1709                 return -1;
1710
1711         if (ph->needs_swap)
1712                 pmu_num = bswap_32(pmu_num);
1713
1714         if (!pmu_num) {
1715                 pr_debug("pmu mappings not available\n");
1716                 return 0;
1717         }
1718
1719         ph->env.nr_pmu_mappings = pmu_num;
1720         strbuf_init(&sb, 128);
1721
1722         while (pmu_num) {
1723                 if (readn(fd, &type, sizeof(type)) != sizeof(type))
1724                         goto error;
1725                 if (ph->needs_swap)
1726                         type = bswap_32(type);
1727
1728                 name = do_read_string(fd, ph);
1729                 if (!name)
1730                         goto error;
1731
1732                 strbuf_addf(&sb, "%u:%s", type, name);
1733                 /* include a NULL character at the end */
1734                 strbuf_add(&sb, "", 1);
1735
1736                 free(name);
1737                 pmu_num--;
1738         }
1739         ph->env.pmu_mappings = strbuf_detach(&sb, NULL);
1740         return 0;
1741
1742 error:
1743         strbuf_release(&sb);
1744         return -1;
1745 }
1746
1747 static int process_group_desc(struct perf_file_section *section __maybe_unused,
1748                               struct perf_header *ph, int fd,
1749                               void *data __maybe_unused)
1750 {
1751         size_t ret = -1;
1752         u32 i, nr, nr_groups;
1753         struct perf_session *session;
1754         struct perf_evsel *evsel, *leader = NULL;
1755         struct group_desc {
1756                 char *name;
1757                 u32 leader_idx;
1758                 u32 nr_members;
1759         } *desc;
1760
1761         if (readn(fd, &nr_groups, sizeof(nr_groups)) != sizeof(nr_groups))
1762                 return -1;
1763
1764         if (ph->needs_swap)
1765                 nr_groups = bswap_32(nr_groups);
1766
1767         ph->env.nr_groups = nr_groups;
1768         if (!nr_groups) {
1769                 pr_debug("group desc not available\n");
1770                 return 0;
1771         }
1772
1773         desc = calloc(nr_groups, sizeof(*desc));
1774         if (!desc)
1775                 return -1;
1776
1777         for (i = 0; i < nr_groups; i++) {
1778                 desc[i].name = do_read_string(fd, ph);
1779                 if (!desc[i].name)
1780                         goto out_free;
1781
1782                 if (readn(fd, &desc[i].leader_idx, sizeof(u32)) != sizeof(u32))
1783                         goto out_free;
1784
1785                 if (readn(fd, &desc[i].nr_members, sizeof(u32)) != sizeof(u32))
1786                         goto out_free;
1787
1788                 if (ph->needs_swap) {
1789                         desc[i].leader_idx = bswap_32(desc[i].leader_idx);
1790                         desc[i].nr_members = bswap_32(desc[i].nr_members);
1791                 }
1792         }
1793
1794         /*
1795          * Rebuild group relationship based on the group_desc
1796          */
1797         session = container_of(ph, struct perf_session, header);
1798         session->evlist->nr_groups = nr_groups;
1799
1800         i = nr = 0;
1801         evlist__for_each(session->evlist, evsel) {
1802                 if (evsel->idx == (int) desc[i].leader_idx) {
1803                         evsel->leader = evsel;
1804                         /* {anon_group} is a dummy name */
1805                         if (strcmp(desc[i].name, "{anon_group}")) {
1806                                 evsel->group_name = desc[i].name;
1807                                 desc[i].name = NULL;
1808                         }
1809                         evsel->nr_members = desc[i].nr_members;
1810
1811                         if (i >= nr_groups || nr > 0) {
1812                                 pr_debug("invalid group desc\n");
1813                                 goto out_free;
1814                         }
1815
1816                         leader = evsel;
1817                         nr = evsel->nr_members - 1;
1818                         i++;
1819                 } else if (nr) {
1820                         /* This is a group member */
1821                         evsel->leader = leader;
1822
1823                         nr--;
1824                 }
1825         }
1826
1827         if (i != nr_groups || nr != 0) {
1828                 pr_debug("invalid group desc\n");
1829                 goto out_free;
1830         }
1831
1832         ret = 0;
1833 out_free:
1834         for (i = 0; i < nr_groups; i++)
1835                 zfree(&desc[i].name);
1836         free(desc);
1837
1838         return ret;
1839 }
1840
1841 static int process_auxtrace(struct perf_file_section *section,
1842                             struct perf_header *ph, int fd,
1843                             void *data __maybe_unused)
1844 {
1845         struct perf_session *session;
1846         int err;
1847
1848         session = container_of(ph, struct perf_session, header);
1849
1850         err = auxtrace_index__process(fd, section->size, session,
1851                                       ph->needs_swap);
1852         if (err < 0)
1853                 pr_err("Failed to process auxtrace index\n");
1854         return err;
1855 }
1856
1857 struct feature_ops {
1858         int (*write)(int fd, struct perf_header *h, struct perf_evlist *evlist);
1859         void (*print)(struct perf_header *h, int fd, FILE *fp);
1860         int (*process)(struct perf_file_section *section,
1861                        struct perf_header *h, int fd, void *data);
1862         const char *name;
1863         bool full_only;
1864 };
1865
1866 #define FEAT_OPA(n, func) \
1867         [n] = { .name = #n, .write = write_##func, .print = print_##func }
1868 #define FEAT_OPP(n, func) \
1869         [n] = { .name = #n, .write = write_##func, .print = print_##func, \
1870                 .process = process_##func }
1871 #define FEAT_OPF(n, func) \
1872         [n] = { .name = #n, .write = write_##func, .print = print_##func, \
1873                 .process = process_##func, .full_only = true }
1874
1875 /* feature_ops not implemented: */
1876 #define print_tracing_data      NULL
1877 #define print_build_id          NULL
1878
1879 static const struct feature_ops feat_ops[HEADER_LAST_FEATURE] = {
1880         FEAT_OPP(HEADER_TRACING_DATA,   tracing_data),
1881         FEAT_OPP(HEADER_BUILD_ID,       build_id),
1882         FEAT_OPP(HEADER_HOSTNAME,       hostname),
1883         FEAT_OPP(HEADER_OSRELEASE,      osrelease),
1884         FEAT_OPP(HEADER_VERSION,        version),
1885         FEAT_OPP(HEADER_ARCH,           arch),
1886         FEAT_OPP(HEADER_NRCPUS,         nrcpus),
1887         FEAT_OPP(HEADER_CPUDESC,        cpudesc),
1888         FEAT_OPP(HEADER_CPUID,          cpuid),
1889         FEAT_OPP(HEADER_TOTAL_MEM,      total_mem),
1890         FEAT_OPP(HEADER_EVENT_DESC,     event_desc),
1891         FEAT_OPP(HEADER_CMDLINE,        cmdline),
1892         FEAT_OPF(HEADER_CPU_TOPOLOGY,   cpu_topology),
1893         FEAT_OPF(HEADER_NUMA_TOPOLOGY,  numa_topology),
1894         FEAT_OPA(HEADER_BRANCH_STACK,   branch_stack),
1895         FEAT_OPP(HEADER_PMU_MAPPINGS,   pmu_mappings),
1896         FEAT_OPP(HEADER_GROUP_DESC,     group_desc),
1897         FEAT_OPP(HEADER_AUXTRACE,       auxtrace),
1898 };
1899
1900 struct header_print_data {
1901         FILE *fp;
1902         bool full; /* extended list of headers */
1903 };
1904
1905 static int perf_file_section__fprintf_info(struct perf_file_section *section,
1906                                            struct perf_header *ph,
1907                                            int feat, int fd, void *data)
1908 {
1909         struct header_print_data *hd = data;
1910
1911         if (lseek(fd, section->offset, SEEK_SET) == (off_t)-1) {
1912                 pr_debug("Failed to lseek to %" PRIu64 " offset for feature "
1913                                 "%d, continuing...\n", section->offset, feat);
1914                 return 0;
1915         }
1916         if (feat >= HEADER_LAST_FEATURE) {
1917                 pr_warning("unknown feature %d\n", feat);
1918                 return 0;
1919         }
1920         if (!feat_ops[feat].print)
1921                 return 0;
1922
1923         if (!feat_ops[feat].full_only || hd->full)
1924                 feat_ops[feat].print(ph, fd, hd->fp);
1925         else
1926                 fprintf(hd->fp, "# %s info available, use -I to display\n",
1927                         feat_ops[feat].name);
1928
1929         return 0;
1930 }
1931
1932 int perf_header__fprintf_info(struct perf_session *session, FILE *fp, bool full)
1933 {
1934         struct header_print_data hd;
1935         struct perf_header *header = &session->header;
1936         int fd = perf_data_file__fd(session->file);
1937         hd.fp = fp;
1938         hd.full = full;
1939
1940         perf_header__process_sections(header, fd, &hd,
1941                                       perf_file_section__fprintf_info);
1942         return 0;
1943 }
1944
1945 static int do_write_feat(int fd, struct perf_header *h, int type,
1946                          struct perf_file_section **p,
1947                          struct perf_evlist *evlist)
1948 {
1949         int err;
1950         int ret = 0;
1951
1952         if (perf_header__has_feat(h, type)) {
1953                 if (!feat_ops[type].write)
1954                         return -1;
1955
1956                 (*p)->offset = lseek(fd, 0, SEEK_CUR);
1957
1958                 err = feat_ops[type].write(fd, h, evlist);
1959                 if (err < 0) {
1960                         pr_debug("failed to write feature %d\n", type);
1961
1962                         /* undo anything written */
1963                         lseek(fd, (*p)->offset, SEEK_SET);
1964
1965                         return -1;
1966                 }
1967                 (*p)->size = lseek(fd, 0, SEEK_CUR) - (*p)->offset;
1968                 (*p)++;
1969         }
1970         return ret;
1971 }
1972
1973 static int perf_header__adds_write(struct perf_header *header,
1974                                    struct perf_evlist *evlist, int fd)
1975 {
1976         int nr_sections;
1977         struct perf_file_section *feat_sec, *p;
1978         int sec_size;
1979         u64 sec_start;
1980         int feat;
1981         int err;
1982
1983         nr_sections = bitmap_weight(header->adds_features, HEADER_FEAT_BITS);
1984         if (!nr_sections)
1985                 return 0;
1986
1987         feat_sec = p = calloc(nr_sections, sizeof(*feat_sec));
1988         if (feat_sec == NULL)
1989                 return -ENOMEM;
1990
1991         sec_size = sizeof(*feat_sec) * nr_sections;
1992
1993         sec_start = header->feat_offset;
1994         lseek(fd, sec_start + sec_size, SEEK_SET);
1995
1996         for_each_set_bit(feat, header->adds_features, HEADER_FEAT_BITS) {
1997                 if (do_write_feat(fd, header, feat, &p, evlist))
1998                         perf_header__clear_feat(header, feat);
1999         }
2000
2001         lseek(fd, sec_start, SEEK_SET);
2002         /*
2003          * may write more than needed due to dropped feature, but
2004          * this is okay, reader will skip the mising entries
2005          */
2006         err = do_write(fd, feat_sec, sec_size);
2007         if (err < 0)
2008                 pr_debug("failed to write feature section\n");
2009         free(feat_sec);
2010         return err;
2011 }
2012
2013 int perf_header__write_pipe(int fd)
2014 {
2015         struct perf_pipe_file_header f_header;
2016         int err;
2017
2018         f_header = (struct perf_pipe_file_header){
2019                 .magic     = PERF_MAGIC,
2020                 .size      = sizeof(f_header),
2021         };
2022
2023         err = do_write(fd, &f_header, sizeof(f_header));
2024         if (err < 0) {
2025                 pr_debug("failed to write perf pipe header\n");
2026                 return err;
2027         }
2028
2029         return 0;
2030 }
2031
2032 int perf_session__write_header(struct perf_session *session,
2033                                struct perf_evlist *evlist,
2034                                int fd, bool at_exit)
2035 {
2036         struct perf_file_header f_header;
2037         struct perf_file_attr   f_attr;
2038         struct perf_header *header = &session->header;
2039         struct perf_evsel *evsel;
2040         u64 attr_offset;
2041         int err;
2042
2043         lseek(fd, sizeof(f_header), SEEK_SET);
2044
2045         evlist__for_each(session->evlist, evsel) {
2046                 evsel->id_offset = lseek(fd, 0, SEEK_CUR);
2047                 err = do_write(fd, evsel->id, evsel->ids * sizeof(u64));
2048                 if (err < 0) {
2049                         pr_debug("failed to write perf header\n");
2050                         return err;
2051                 }
2052         }
2053
2054         attr_offset = lseek(fd, 0, SEEK_CUR);
2055
2056         evlist__for_each(evlist, evsel) {
2057                 f_attr = (struct perf_file_attr){
2058                         .attr = evsel->attr,
2059                         .ids  = {
2060                                 .offset = evsel->id_offset,
2061                                 .size   = evsel->ids * sizeof(u64),
2062                         }
2063                 };
2064                 err = do_write(fd, &f_attr, sizeof(f_attr));
2065                 if (err < 0) {
2066                         pr_debug("failed to write perf header attribute\n");
2067                         return err;
2068                 }
2069         }
2070
2071         if (!header->data_offset)
2072                 header->data_offset = lseek(fd, 0, SEEK_CUR);
2073         header->feat_offset = header->data_offset + header->data_size;
2074
2075         if (at_exit) {
2076                 err = perf_header__adds_write(header, evlist, fd);
2077                 if (err < 0)
2078                         return err;
2079         }
2080
2081         f_header = (struct perf_file_header){
2082                 .magic     = PERF_MAGIC,
2083                 .size      = sizeof(f_header),
2084                 .attr_size = sizeof(f_attr),
2085                 .attrs = {
2086                         .offset = attr_offset,
2087                         .size   = evlist->nr_entries * sizeof(f_attr),
2088                 },
2089                 .data = {
2090                         .offset = header->data_offset,
2091                         .size   = header->data_size,
2092                 },
2093                 /* event_types is ignored, store zeros */
2094         };
2095
2096         memcpy(&f_header.adds_features, &header->adds_features, sizeof(header->adds_features));
2097
2098         lseek(fd, 0, SEEK_SET);
2099         err = do_write(fd, &f_header, sizeof(f_header));
2100         if (err < 0) {
2101                 pr_debug("failed to write perf header\n");
2102                 return err;
2103         }
2104         lseek(fd, header->data_offset + header->data_size, SEEK_SET);
2105
2106         return 0;
2107 }
2108
2109 static int perf_header__getbuffer64(struct perf_header *header,
2110                                     int fd, void *buf, size_t size)
2111 {
2112         if (readn(fd, buf, size) <= 0)
2113                 return -1;
2114
2115         if (header->needs_swap)
2116                 mem_bswap_64(buf, size);
2117
2118         return 0;
2119 }
2120
2121 int perf_header__process_sections(struct perf_header *header, int fd,
2122                                   void *data,
2123                                   int (*process)(struct perf_file_section *section,
2124                                                  struct perf_header *ph,
2125                                                  int feat, int fd, void *data))
2126 {
2127         struct perf_file_section *feat_sec, *sec;
2128         int nr_sections;
2129         int sec_size;
2130         int feat;
2131         int err;
2132
2133         nr_sections = bitmap_weight(header->adds_features, HEADER_FEAT_BITS);
2134         if (!nr_sections)
2135                 return 0;
2136
2137         feat_sec = sec = calloc(nr_sections, sizeof(*feat_sec));
2138         if (!feat_sec)
2139                 return -1;
2140
2141         sec_size = sizeof(*feat_sec) * nr_sections;
2142
2143         lseek(fd, header->feat_offset, SEEK_SET);
2144
2145         err = perf_header__getbuffer64(header, fd, feat_sec, sec_size);
2146         if (err < 0)
2147                 goto out_free;
2148
2149         for_each_set_bit(feat, header->adds_features, HEADER_LAST_FEATURE) {
2150                 err = process(sec++, header, feat, fd, data);
2151                 if (err < 0)
2152                         goto out_free;
2153         }
2154         err = 0;
2155 out_free:
2156         free(feat_sec);
2157         return err;
2158 }
2159
2160 static const int attr_file_abi_sizes[] = {
2161         [0] = PERF_ATTR_SIZE_VER0,
2162         [1] = PERF_ATTR_SIZE_VER1,
2163         [2] = PERF_ATTR_SIZE_VER2,
2164         [3] = PERF_ATTR_SIZE_VER3,
2165         [4] = PERF_ATTR_SIZE_VER4,
2166         0,
2167 };
2168
2169 /*
2170  * In the legacy file format, the magic number is not used to encode endianness.
2171  * hdr_sz was used to encode endianness. But given that hdr_sz can vary based
2172  * on ABI revisions, we need to try all combinations for all endianness to
2173  * detect the endianness.
2174  */
2175 static int try_all_file_abis(uint64_t hdr_sz, struct perf_header *ph)
2176 {
2177         uint64_t ref_size, attr_size;
2178         int i;
2179
2180         for (i = 0 ; attr_file_abi_sizes[i]; i++) {
2181                 ref_size = attr_file_abi_sizes[i]
2182                          + sizeof(struct perf_file_section);
2183                 if (hdr_sz != ref_size) {
2184                         attr_size = bswap_64(hdr_sz);
2185                         if (attr_size != ref_size)
2186                                 continue;
2187
2188                         ph->needs_swap = true;
2189                 }
2190                 pr_debug("ABI%d perf.data file detected, need_swap=%d\n",
2191                          i,
2192                          ph->needs_swap);
2193                 return 0;
2194         }
2195         /* could not determine endianness */
2196         return -1;
2197 }
2198
2199 #define PERF_PIPE_HDR_VER0      16
2200
2201 static const size_t attr_pipe_abi_sizes[] = {
2202         [0] = PERF_PIPE_HDR_VER0,
2203         0,
2204 };
2205
2206 /*
2207  * In the legacy pipe format, there is an implicit assumption that endiannesss
2208  * between host recording the samples, and host parsing the samples is the
2209  * same. This is not always the case given that the pipe output may always be
2210  * redirected into a file and analyzed on a different machine with possibly a
2211  * different endianness and perf_event ABI revsions in the perf tool itself.
2212  */
2213 static int try_all_pipe_abis(uint64_t hdr_sz, struct perf_header *ph)
2214 {
2215         u64 attr_size;
2216         int i;
2217
2218         for (i = 0 ; attr_pipe_abi_sizes[i]; i++) {
2219                 if (hdr_sz != attr_pipe_abi_sizes[i]) {
2220                         attr_size = bswap_64(hdr_sz);
2221                         if (attr_size != hdr_sz)
2222                                 continue;
2223
2224                         ph->needs_swap = true;
2225                 }
2226                 pr_debug("Pipe ABI%d perf.data file detected\n", i);
2227                 return 0;
2228         }
2229         return -1;
2230 }
2231
2232 bool is_perf_magic(u64 magic)
2233 {
2234         if (!memcmp(&magic, __perf_magic1, sizeof(magic))
2235                 || magic == __perf_magic2
2236                 || magic == __perf_magic2_sw)
2237                 return true;
2238
2239         return false;
2240 }
2241
2242 static int check_magic_endian(u64 magic, uint64_t hdr_sz,
2243                               bool is_pipe, struct perf_header *ph)
2244 {
2245         int ret;
2246
2247         /* check for legacy format */
2248         ret = memcmp(&magic, __perf_magic1, sizeof(magic));
2249         if (ret == 0) {
2250                 ph->version = PERF_HEADER_VERSION_1;
2251                 pr_debug("legacy perf.data format\n");
2252                 if (is_pipe)
2253                         return try_all_pipe_abis(hdr_sz, ph);
2254
2255                 return try_all_file_abis(hdr_sz, ph);
2256         }
2257         /*
2258          * the new magic number serves two purposes:
2259          * - unique number to identify actual perf.data files
2260          * - encode endianness of file
2261          */
2262         ph->version = PERF_HEADER_VERSION_2;
2263
2264         /* check magic number with one endianness */
2265         if (magic == __perf_magic2)
2266                 return 0;
2267
2268         /* check magic number with opposite endianness */
2269         if (magic != __perf_magic2_sw)
2270                 return -1;
2271
2272         ph->needs_swap = true;
2273
2274         return 0;
2275 }
2276
2277 int perf_file_header__read(struct perf_file_header *header,
2278                            struct perf_header *ph, int fd)
2279 {
2280         ssize_t ret;
2281
2282         lseek(fd, 0, SEEK_SET);
2283
2284         ret = readn(fd, header, sizeof(*header));
2285         if (ret <= 0)
2286                 return -1;
2287
2288         if (check_magic_endian(header->magic,
2289                                header->attr_size, false, ph) < 0) {
2290                 pr_debug("magic/endian check failed\n");
2291                 return -1;
2292         }
2293
2294         if (ph->needs_swap) {
2295                 mem_bswap_64(header, offsetof(struct perf_file_header,
2296                              adds_features));
2297         }
2298
2299         if (header->size != sizeof(*header)) {
2300                 /* Support the previous format */
2301                 if (header->size == offsetof(typeof(*header), adds_features))
2302                         bitmap_zero(header->adds_features, HEADER_FEAT_BITS);
2303                 else
2304                         return -1;
2305         } else if (ph->needs_swap) {
2306                 /*
2307                  * feature bitmap is declared as an array of unsigned longs --
2308                  * not good since its size can differ between the host that
2309                  * generated the data file and the host analyzing the file.
2310                  *
2311                  * We need to handle endianness, but we don't know the size of
2312                  * the unsigned long where the file was generated. Take a best
2313                  * guess at determining it: try 64-bit swap first (ie., file
2314                  * created on a 64-bit host), and check if the hostname feature
2315                  * bit is set (this feature bit is forced on as of fbe96f2).
2316                  * If the bit is not, undo the 64-bit swap and try a 32-bit
2317                  * swap. If the hostname bit is still not set (e.g., older data
2318                  * file), punt and fallback to the original behavior --
2319                  * clearing all feature bits and setting buildid.
2320                  */
2321                 mem_bswap_64(&header->adds_features,
2322                             BITS_TO_U64(HEADER_FEAT_BITS));
2323
2324                 if (!test_bit(HEADER_HOSTNAME, header->adds_features)) {
2325                         /* unswap as u64 */
2326                         mem_bswap_64(&header->adds_features,
2327                                     BITS_TO_U64(HEADER_FEAT_BITS));
2328
2329                         /* unswap as u32 */
2330                         mem_bswap_32(&header->adds_features,
2331                                     BITS_TO_U32(HEADER_FEAT_BITS));
2332                 }
2333
2334                 if (!test_bit(HEADER_HOSTNAME, header->adds_features)) {
2335                         bitmap_zero(header->adds_features, HEADER_FEAT_BITS);
2336                         set_bit(HEADER_BUILD_ID, header->adds_features);
2337                 }
2338         }
2339
2340         memcpy(&ph->adds_features, &header->adds_features,
2341                sizeof(ph->adds_features));
2342
2343         ph->data_offset  = header->data.offset;
2344         ph->data_size    = header->data.size;
2345         ph->feat_offset  = header->data.offset + header->data.size;
2346         return 0;
2347 }
2348
2349 static int perf_file_section__process(struct perf_file_section *section,
2350                                       struct perf_header *ph,
2351                                       int feat, int fd, void *data)
2352 {
2353         if (lseek(fd, section->offset, SEEK_SET) == (off_t)-1) {
2354                 pr_debug("Failed to lseek to %" PRIu64 " offset for feature "
2355                           "%d, continuing...\n", section->offset, feat);
2356                 return 0;
2357         }
2358
2359         if (feat >= HEADER_LAST_FEATURE) {
2360                 pr_debug("unknown feature %d, continuing...\n", feat);
2361                 return 0;
2362         }
2363
2364         if (!feat_ops[feat].process)
2365                 return 0;
2366
2367         return feat_ops[feat].process(section, ph, fd, data);
2368 }
2369
2370 static int perf_file_header__read_pipe(struct perf_pipe_file_header *header,
2371                                        struct perf_header *ph, int fd,
2372                                        bool repipe)
2373 {
2374         ssize_t ret;
2375
2376         ret = readn(fd, header, sizeof(*header));
2377         if (ret <= 0)
2378                 return -1;
2379
2380         if (check_magic_endian(header->magic, header->size, true, ph) < 0) {
2381                 pr_debug("endian/magic failed\n");
2382                 return -1;
2383         }
2384
2385         if (ph->needs_swap)
2386                 header->size = bswap_64(header->size);
2387
2388         if (repipe && do_write(STDOUT_FILENO, header, sizeof(*header)) < 0)
2389                 return -1;
2390
2391         return 0;
2392 }
2393
2394 static int perf_header__read_pipe(struct perf_session *session)
2395 {
2396         struct perf_header *header = &session->header;
2397         struct perf_pipe_file_header f_header;
2398
2399         if (perf_file_header__read_pipe(&f_header, header,
2400                                         perf_data_file__fd(session->file),
2401                                         session->repipe) < 0) {
2402                 pr_debug("incompatible file format\n");
2403                 return -EINVAL;
2404         }
2405
2406         return 0;
2407 }
2408
2409 static int read_attr(int fd, struct perf_header *ph,
2410                      struct perf_file_attr *f_attr)
2411 {
2412         struct perf_event_attr *attr = &f_attr->attr;
2413         size_t sz, left;
2414         size_t our_sz = sizeof(f_attr->attr);
2415         ssize_t ret;
2416
2417         memset(f_attr, 0, sizeof(*f_attr));
2418
2419         /* read minimal guaranteed structure */
2420         ret = readn(fd, attr, PERF_ATTR_SIZE_VER0);
2421         if (ret <= 0) {
2422                 pr_debug("cannot read %d bytes of header attr\n",
2423                          PERF_ATTR_SIZE_VER0);
2424                 return -1;
2425         }
2426
2427         /* on file perf_event_attr size */
2428         sz = attr->size;
2429
2430         if (ph->needs_swap)
2431                 sz = bswap_32(sz);
2432
2433         if (sz == 0) {
2434                 /* assume ABI0 */
2435                 sz =  PERF_ATTR_SIZE_VER0;
2436         } else if (sz > our_sz) {
2437                 pr_debug("file uses a more recent and unsupported ABI"
2438                          " (%zu bytes extra)\n", sz - our_sz);
2439                 return -1;
2440         }
2441         /* what we have not yet read and that we know about */
2442         left = sz - PERF_ATTR_SIZE_VER0;
2443         if (left) {
2444                 void *ptr = attr;
2445                 ptr += PERF_ATTR_SIZE_VER0;
2446
2447                 ret = readn(fd, ptr, left);
2448         }
2449         /* read perf_file_section, ids are read in caller */
2450         ret = readn(fd, &f_attr->ids, sizeof(f_attr->ids));
2451
2452         return ret <= 0 ? -1 : 0;
2453 }
2454
2455 static int perf_evsel__prepare_tracepoint_event(struct perf_evsel *evsel,
2456                                                 struct pevent *pevent)
2457 {
2458         struct event_format *event;
2459         char bf[128];
2460
2461         /* already prepared */
2462         if (evsel->tp_format)
2463                 return 0;
2464
2465         if (pevent == NULL) {
2466                 pr_debug("broken or missing trace data\n");
2467                 return -1;
2468         }
2469
2470         event = pevent_find_event(pevent, evsel->attr.config);
2471         if (event == NULL)
2472                 return -1;
2473
2474         if (!evsel->name) {
2475                 snprintf(bf, sizeof(bf), "%s:%s", event->system, event->name);
2476                 evsel->name = strdup(bf);
2477                 if (evsel->name == NULL)
2478                         return -1;
2479         }
2480
2481         evsel->tp_format = event;
2482         return 0;
2483 }
2484
2485 static int perf_evlist__prepare_tracepoint_events(struct perf_evlist *evlist,
2486                                                   struct pevent *pevent)
2487 {
2488         struct perf_evsel *pos;
2489
2490         evlist__for_each(evlist, pos) {
2491                 if (pos->attr.type == PERF_TYPE_TRACEPOINT &&
2492                     perf_evsel__prepare_tracepoint_event(pos, pevent))
2493                         return -1;
2494         }
2495
2496         return 0;
2497 }
2498
2499 int perf_session__read_header(struct perf_session *session)
2500 {
2501         struct perf_data_file *file = session->file;
2502         struct perf_header *header = &session->header;
2503         struct perf_file_header f_header;
2504         struct perf_file_attr   f_attr;
2505         u64                     f_id;
2506         int nr_attrs, nr_ids, i, j;
2507         int fd = perf_data_file__fd(file);
2508
2509         session->evlist = perf_evlist__new();
2510         if (session->evlist == NULL)
2511                 return -ENOMEM;
2512
2513         if (perf_data_file__is_pipe(file))
2514                 return perf_header__read_pipe(session);
2515
2516         if (perf_file_header__read(&f_header, header, fd) < 0)
2517                 return -EINVAL;
2518
2519         /*
2520          * Sanity check that perf.data was written cleanly; data size is
2521          * initialized to 0 and updated only if the on_exit function is run.
2522          * If data size is still 0 then the file contains only partial
2523          * information.  Just warn user and process it as much as it can.
2524          */
2525         if (f_header.data.size == 0) {
2526                 pr_warning("WARNING: The %s file's data size field is 0 which is unexpected.\n"
2527                            "Was the 'perf record' command properly terminated?\n",
2528                            file->path);
2529         }
2530
2531         nr_attrs = f_header.attrs.size / f_header.attr_size;
2532         lseek(fd, f_header.attrs.offset, SEEK_SET);
2533
2534         for (i = 0; i < nr_attrs; i++) {
2535                 struct perf_evsel *evsel;
2536                 off_t tmp;
2537
2538                 if (read_attr(fd, header, &f_attr) < 0)
2539                         goto out_errno;
2540
2541                 if (header->needs_swap) {
2542                         f_attr.ids.size   = bswap_64(f_attr.ids.size);
2543                         f_attr.ids.offset = bswap_64(f_attr.ids.offset);
2544                         perf_event__attr_swap(&f_attr.attr);
2545                 }
2546
2547                 tmp = lseek(fd, 0, SEEK_CUR);
2548                 evsel = perf_evsel__new(&f_attr.attr);
2549
2550                 if (evsel == NULL)
2551                         goto out_delete_evlist;
2552
2553                 evsel->needs_swap = header->needs_swap;
2554                 /*
2555                  * Do it before so that if perf_evsel__alloc_id fails, this
2556                  * entry gets purged too at perf_evlist__delete().
2557                  */
2558                 perf_evlist__add(session->evlist, evsel);
2559
2560                 nr_ids = f_attr.ids.size / sizeof(u64);
2561                 /*
2562                  * We don't have the cpu and thread maps on the header, so
2563                  * for allocating the perf_sample_id table we fake 1 cpu and
2564                  * hattr->ids threads.
2565                  */
2566                 if (perf_evsel__alloc_id(evsel, 1, nr_ids))
2567                         goto out_delete_evlist;
2568
2569                 lseek(fd, f_attr.ids.offset, SEEK_SET);
2570
2571                 for (j = 0; j < nr_ids; j++) {
2572                         if (perf_header__getbuffer64(header, fd, &f_id, sizeof(f_id)))
2573                                 goto out_errno;
2574
2575                         perf_evlist__id_add(session->evlist, evsel, 0, j, f_id);
2576                 }
2577
2578                 lseek(fd, tmp, SEEK_SET);
2579         }
2580
2581         symbol_conf.nr_events = nr_attrs;
2582
2583         perf_header__process_sections(header, fd, &session->tevent,
2584                                       perf_file_section__process);
2585
2586         if (perf_evlist__prepare_tracepoint_events(session->evlist,
2587                                                    session->tevent.pevent))
2588                 goto out_delete_evlist;
2589
2590         return 0;
2591 out_errno:
2592         return -errno;
2593
2594 out_delete_evlist:
2595         perf_evlist__delete(session->evlist);
2596         session->evlist = NULL;
2597         return -ENOMEM;
2598 }
2599
2600 int perf_event__synthesize_attr(struct perf_tool *tool,
2601                                 struct perf_event_attr *attr, u32 ids, u64 *id,
2602                                 perf_event__handler_t process)
2603 {
2604         union perf_event *ev;
2605         size_t size;
2606         int err;
2607
2608         size = sizeof(struct perf_event_attr);
2609         size = PERF_ALIGN(size, sizeof(u64));
2610         size += sizeof(struct perf_event_header);
2611         size += ids * sizeof(u64);
2612
2613         ev = malloc(size);
2614
2615         if (ev == NULL)
2616                 return -ENOMEM;
2617
2618         ev->attr.attr = *attr;
2619         memcpy(ev->attr.id, id, ids * sizeof(u64));
2620
2621         ev->attr.header.type = PERF_RECORD_HEADER_ATTR;
2622         ev->attr.header.size = (u16)size;
2623
2624         if (ev->attr.header.size == size)
2625                 err = process(tool, ev, NULL, NULL);
2626         else
2627                 err = -E2BIG;
2628
2629         free(ev);
2630
2631         return err;
2632 }
2633
2634 int perf_event__synthesize_attrs(struct perf_tool *tool,
2635                                    struct perf_session *session,
2636                                    perf_event__handler_t process)
2637 {
2638         struct perf_evsel *evsel;
2639         int err = 0;
2640
2641         evlist__for_each(session->evlist, evsel) {
2642                 err = perf_event__synthesize_attr(tool, &evsel->attr, evsel->ids,
2643                                                   evsel->id, process);
2644                 if (err) {
2645                         pr_debug("failed to create perf header attribute\n");
2646                         return err;
2647                 }
2648         }
2649
2650         return err;
2651 }
2652
2653 int perf_event__process_attr(struct perf_tool *tool __maybe_unused,
2654                              union perf_event *event,
2655                              struct perf_evlist **pevlist)
2656 {
2657         u32 i, ids, n_ids;
2658         struct perf_evsel *evsel;
2659         struct perf_evlist *evlist = *pevlist;
2660
2661         if (evlist == NULL) {
2662                 *pevlist = evlist = perf_evlist__new();
2663                 if (evlist == NULL)
2664                         return -ENOMEM;
2665         }
2666
2667         evsel = perf_evsel__new(&event->attr.attr);
2668         if (evsel == NULL)
2669                 return -ENOMEM;
2670
2671         perf_evlist__add(evlist, evsel);
2672
2673         ids = event->header.size;
2674         ids -= (void *)&event->attr.id - (void *)event;
2675         n_ids = ids / sizeof(u64);
2676         /*
2677          * We don't have the cpu and thread maps on the header, so
2678          * for allocating the perf_sample_id table we fake 1 cpu and
2679          * hattr->ids threads.
2680          */
2681         if (perf_evsel__alloc_id(evsel, 1, n_ids))
2682                 return -ENOMEM;
2683
2684         for (i = 0; i < n_ids; i++) {
2685                 perf_evlist__id_add(evlist, evsel, 0, i, event->attr.id[i]);
2686         }
2687
2688         symbol_conf.nr_events = evlist->nr_entries;
2689
2690         return 0;
2691 }
2692
2693 int perf_event__synthesize_tracing_data(struct perf_tool *tool, int fd,
2694                                         struct perf_evlist *evlist,
2695                                         perf_event__handler_t process)
2696 {
2697         union perf_event ev;
2698         struct tracing_data *tdata;
2699         ssize_t size = 0, aligned_size = 0, padding;
2700         int err __maybe_unused = 0;
2701
2702         /*
2703          * We are going to store the size of the data followed
2704          * by the data contents. Since the fd descriptor is a pipe,
2705          * we cannot seek back to store the size of the data once
2706          * we know it. Instead we:
2707          *
2708          * - write the tracing data to the temp file
2709          * - get/write the data size to pipe
2710          * - write the tracing data from the temp file
2711          *   to the pipe
2712          */
2713         tdata = tracing_data_get(&evlist->entries, fd, true);
2714         if (!tdata)
2715                 return -1;
2716
2717         memset(&ev, 0, sizeof(ev));
2718
2719         ev.tracing_data.header.type = PERF_RECORD_HEADER_TRACING_DATA;
2720         size = tdata->size;
2721         aligned_size = PERF_ALIGN(size, sizeof(u64));
2722         padding = aligned_size - size;
2723         ev.tracing_data.header.size = sizeof(ev.tracing_data);
2724         ev.tracing_data.size = aligned_size;
2725
2726         process(tool, &ev, NULL, NULL);
2727
2728         /*
2729          * The put function will copy all the tracing data
2730          * stored in temp file to the pipe.
2731          */
2732         tracing_data_put(tdata);
2733
2734         write_padded(fd, NULL, 0, padding);
2735
2736         return aligned_size;
2737 }
2738
2739 int perf_event__process_tracing_data(struct perf_tool *tool __maybe_unused,
2740                                      union perf_event *event,
2741                                      struct perf_session *session)
2742 {
2743         ssize_t size_read, padding, size = event->tracing_data.size;
2744         int fd = perf_data_file__fd(session->file);
2745         off_t offset = lseek(fd, 0, SEEK_CUR);
2746         char buf[BUFSIZ];
2747
2748         /* setup for reading amidst mmap */
2749         lseek(fd, offset + sizeof(struct tracing_data_event),
2750               SEEK_SET);
2751
2752         size_read = trace_report(fd, &session->tevent,
2753                                  session->repipe);
2754         padding = PERF_ALIGN(size_read, sizeof(u64)) - size_read;
2755
2756         if (readn(fd, buf, padding) < 0) {
2757                 pr_err("%s: reading input file", __func__);
2758                 return -1;
2759         }
2760         if (session->repipe) {
2761                 int retw = write(STDOUT_FILENO, buf, padding);
2762                 if (retw <= 0 || retw != padding) {
2763                         pr_err("%s: repiping tracing data padding", __func__);
2764                         return -1;
2765                 }
2766         }
2767
2768         if (size_read + padding != size) {
2769                 pr_err("%s: tracing data size mismatch", __func__);
2770                 return -1;
2771         }
2772
2773         perf_evlist__prepare_tracepoint_events(session->evlist,
2774                                                session->tevent.pevent);
2775
2776         return size_read + padding;
2777 }
2778
2779 int perf_event__synthesize_build_id(struct perf_tool *tool,
2780                                     struct dso *pos, u16 misc,
2781                                     perf_event__handler_t process,
2782                                     struct machine *machine)
2783 {
2784         union perf_event ev;
2785         size_t len;
2786         int err = 0;
2787
2788         if (!pos->hit)
2789                 return err;
2790
2791         memset(&ev, 0, sizeof(ev));
2792
2793         len = pos->long_name_len + 1;
2794         len = PERF_ALIGN(len, NAME_ALIGN);
2795         memcpy(&ev.build_id.build_id, pos->build_id, sizeof(pos->build_id));
2796         ev.build_id.header.type = PERF_RECORD_HEADER_BUILD_ID;
2797         ev.build_id.header.misc = misc;
2798         ev.build_id.pid = machine->pid;
2799         ev.build_id.header.size = sizeof(ev.build_id) + len;
2800         memcpy(&ev.build_id.filename, pos->long_name, pos->long_name_len);
2801
2802         err = process(tool, &ev, NULL, machine);
2803
2804         return err;
2805 }
2806
2807 int perf_event__process_build_id(struct perf_tool *tool __maybe_unused,
2808                                  union perf_event *event,
2809                                  struct perf_session *session)
2810 {
2811         __event_process_build_id(&event->build_id,
2812                                  event->build_id.filename,
2813                                  session);
2814         return 0;
2815 }