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 /*
  28  * magic2 = "PERFILE2"
  29  * must be a numerical value to let the endianness
  30  * determine the memory layout. That way we are able
  31  * to detect endianness when reading the perf.data file
  32  * back.
  33  *
  34  * we check for legacy (PERFFILE) format.
  35  */
  36 static const char *__perf_magic1 = "PERFFILE";
  37 static const u64 __perf_magic2    = 0x32454c4946524550ULL;
  38 static const u64 __perf_magic2_sw = 0x50455246494c4532ULL;
  39
  40 #define PERF_MAGIC      __perf_magic2
  41
  42 struct perf_file_attr {
  43         struct perf_event_attr  attr;
  44         struct perf_file_section        ids;
  45 };
  46
  47 void perf_header__set_feat(struct perf_header *header, int feat)
  48 {
  49         set_bit(feat, header->adds_features);
  50 }
  51
  52 void perf_header__clear_feat(struct perf_header *header, int feat)
  53 {
  54         clear_bit(feat, header->adds_features);
  55 }
  56
  57 bool perf_header__has_feat(const struct perf_header *header, int feat)
  58 {
  59         return test_bit(feat, header->adds_features);
  60 }
  61
  62 static int do_write(int fd, const void *buf, size_t size)
  63 {
  64         while (size) {
  65                 int ret = write(fd, buf, size);
  66
  67                 if (ret < 0)
  68                         return -errno;
  69
  70                 size -= ret;
  71                 buf += ret;
  72         }
  73
  74         return 0;
  75 }
  76
  77 int write_padded(int fd, const void *bf, size_t count, size_t count_aligned)
  78 {
  79         static const char zero_buf[NAME_ALIGN];
  80         int err = do_write(fd, bf, count);
  81
  82         if (!err)
  83                 err = do_write(fd, zero_buf, count_aligned - count);
  84
  85         return err;
  86 }
  87
  88 #define string_size(str)                                                \
  89         (PERF_ALIGN((strlen(str) + 1), NAME_ALIGN) + sizeof(u32))
  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 static int write_tracing_data(int fd, struct perf_header *h __maybe_unused,
 139                             struct perf_evlist *evlist)
 140 {
 141         return read_tracing_data(fd, &evlist->entries);
 142 }
 143
 144
 145 static int write_build_id(int fd, struct perf_header *h,
 146                           struct perf_evlist *evlist __maybe_unused)
 147 {
 148         struct perf_session *session;
 149         int err;
 150
 151         session = container_of(h, struct perf_session, header);
 152
 153         if (!perf_session__read_build_ids(session, true))
 154                 return -1;
 155
 156         err = perf_session__write_buildid_table(session, fd);
 157         if (err < 0) {
 158                 pr_debug("failed to write buildid table\n");
 159                 return err;
 160         }
 161         perf_session__cache_build_ids(session);
 162
 163         return 0;
 164 }
 165
 166 static int write_hostname(int fd, struct perf_header *h __maybe_unused,
 167                           struct perf_evlist *evlist __maybe_unused)
 168 {
 169         struct utsname uts;
 170         int ret;
 171
 172         ret = uname(&uts);
 173         if (ret < 0)
 174                 return -1;
 175
 176         return do_write_string(fd, uts.nodename);
 177 }
 178
 179 static int write_osrelease(int fd, struct perf_header *h __maybe_unused,
 180                            struct perf_evlist *evlist __maybe_unused)
 181 {
 182         struct utsname uts;
 183         int ret;
 184
 185         ret = uname(&uts);
 186         if (ret < 0)
 187                 return -1;
 188
 189         return do_write_string(fd, uts.release);
 190 }
 191
 192 static int write_arch(int fd, struct perf_header *h __maybe_unused,
 193                       struct perf_evlist *evlist __maybe_unused)
 194 {
 195         struct utsname uts;
 196         int ret;
 197
 198         ret = uname(&uts);
 199         if (ret < 0)
 200                 return -1;
 201
 202         return do_write_string(fd, uts.machine);
 203 }
 204
 205 static int write_version(int fd, struct perf_header *h __maybe_unused,
 206                          struct perf_evlist *evlist __maybe_unused)
 207 {
 208         return do_write_string(fd, perf_version_string);
 209 }
 210
 211 static int __write_cpudesc(int fd, const char *cpuinfo_proc)
 212 {
 213         FILE *file;
 214         char *buf = NULL;
 215         char *s, *p;
 216         const char *search = cpuinfo_proc;
 217         size_t len = 0;
 218         int ret = -1;
 219
 220         if (!search)
 221                 return -1;
 222
 223         file = fopen("/proc/cpuinfo", "r");
 224         if (!file)
 225                 return -1;
 226
 227         while (getline(&buf, &len, file) > 0) {
 228                 ret = strncmp(buf, search, strlen(search));
 229                 if (!ret)
 230                         break;
 231         }
 232
 233         if (ret) {
 234                 ret = -1;
 235                 goto done;
 236         }
 237
 238         s = buf;
 239
 240         p = strchr(buf, ':');
 241         if (p && *(p+1) == ' ' && *(p+2))
 242                 s = p + 2;
 243         p = strchr(s, '\n');
 244         if (p)
 245                 *p = '\0';
 246
 247         /* squash extra space characters (branding string) */
 248         p = s;
 249         while (*p) {
 250                 if (isspace(*p)) {
 251                         char *r = p + 1;
 252                         char *q = r;
 253                         *p = ' ';
 254                         while (*q && isspace(*q))
 255                                 q++;
 256                         if (q != (p+1))
 257                                 while ((*r++ = *q++));
 258                 }
 259                 p++;
 260         }
 261         ret = do_write_string(fd, s);
 262 done:
 263         free(buf);
 264         fclose(file);
 265         return ret;
 266 }
 267
 268 static int write_cpudesc(int fd, struct perf_header *h __maybe_unused,
 269                        struct perf_evlist *evlist __maybe_unused)
 270 {
 271 #ifndef CPUINFO_PROC
 272 #define CPUINFO_PROC {"model name", }
 273 #endif
 274         const char *cpuinfo_procs[] = CPUINFO_PROC;
 275         unsigned int i;
 276
 277         for (i = 0; i < ARRAY_SIZE(cpuinfo_procs); i++) {
 278                 int ret;
 279                 ret = __write_cpudesc(fd, cpuinfo_procs[i]);
 280                 if (ret >= 0)
 281                         return ret;
 282         }
 283         return -1;
 284 }
 285
 286
 287 static int write_nrcpus(int fd, struct perf_header *h __maybe_unused,
 288                         struct perf_evlist *evlist __maybe_unused)
 289 {
 290         long nr;
 291         u32 nrc, nra;
 292         int ret;
 293
 294         nr = sysconf(_SC_NPROCESSORS_CONF);
 295         if (nr < 0)
 296                 return -1;
 297
 298         nrc = (u32)(nr & UINT_MAX);
 299
 300         nr = sysconf(_SC_NPROCESSORS_ONLN);
 301         if (nr < 0)
 302                 return -1;
 303
 304         nra = (u32)(nr & UINT_MAX);
 305
 306         ret = do_write(fd, &nrc, sizeof(nrc));
 307         if (ret < 0)
 308                 return ret;
 309
 310         return do_write(fd, &nra, sizeof(nra));
 311 }
 312
 313 static int write_event_desc(int fd, struct perf_header *h __maybe_unused,
 314                             struct perf_evlist *evlist)
 315 {
 316         struct perf_evsel *evsel;
 317         u32 nre, nri, sz;
 318         int ret;
 319
 320         nre = evlist->nr_entries;
 321
 322         /*
 323          * write number of events
 324          */
 325         ret = do_write(fd, &nre, sizeof(nre));
 326         if (ret < 0)
 327                 return ret;
 328
 329         /*
 330          * size of perf_event_attr struct
 331          */
 332         sz = (u32)sizeof(evsel->attr);
 333         ret = do_write(fd, &sz, sizeof(sz));
 334         if (ret < 0)
 335                 return ret;
 336
 337         evlist__for_each(evlist, evsel) {
 338                 ret = do_write(fd, &evsel->attr, sz);
 339                 if (ret < 0)
 340                         return ret;
 341                 /*
 342                  * write number of unique id per event
 343                  * there is one id per instance of an event
 344                  *
 345                  * copy into an nri to be independent of the
 346                  * type of ids,
 347                  */
 348                 nri = evsel->ids;
 349                 ret = do_write(fd, &nri, sizeof(nri));
 350                 if (ret < 0)
 351                         return ret;
 352
 353                 /*
 354                  * write event string as passed on cmdline
 355                  */
 356                 ret = do_write_string(fd, perf_evsel__name(evsel));
 357                 if (ret < 0)
 358                         return ret;
 359                 /*
 360                  * write unique ids for this event
 361                  */
 362                 ret = do_write(fd, evsel->id, evsel->ids * sizeof(u64));
 363                 if (ret < 0)
 364                         return ret;
 365         }
 366         return 0;
 367 }
 368
 369 static int write_cmdline(int fd, struct perf_header *h __maybe_unused,
 370                          struct perf_evlist *evlist __maybe_unused)
 371 {
 372         char buf[MAXPATHLEN];
 373         char proc[32];
 374         u32 n;
 375         int i, ret;
 376
 377         /*
 378          * actual atual path to perf binary
 379          */
 380         sprintf(proc, "/proc/%d/exe", getpid());
 381         ret = readlink(proc, buf, sizeof(buf));
 382         if (ret <= 0)
 383                 return -1;
 384
 385         /* readlink() does not add null termination */
 386         buf[ret] = '\0';
 387
 388         /* account for binary path */
 389         n = perf_env.nr_cmdline + 1;
 390
 391         ret = do_write(fd, &n, sizeof(n));
 392         if (ret < 0)
 393                 return ret;
 394
 395         ret = do_write_string(fd, buf);
 396         if (ret < 0)
 397                 return ret;
 398
 399         for (i = 0 ; i < perf_env.nr_cmdline; i++) {
 400                 ret = do_write_string(fd, perf_env.cmdline_argv[i]);
 401                 if (ret < 0)
 402                         return ret;
 403         }
 404         return 0;
 405 }
 406
 407 #define CORE_SIB_FMT \
 408         "/sys/devices/system/cpu/cpu%d/topology/core_siblings_list"
 409 #define THRD_SIB_FMT \
 410         "/sys/devices/system/cpu/cpu%d/topology/thread_siblings_list"
 411
 412 struct cpu_topo {
 413         u32 cpu_nr;
 414         u32 core_sib;
 415         u32 thread_sib;
 416         char **core_siblings;
 417         char **thread_siblings;
 418 };
 419
 420 static int build_cpu_topo(struct cpu_topo *tp, int cpu)
 421 {
 422         FILE *fp;
 423         char filename[MAXPATHLEN];
 424         char *buf = NULL, *p;
 425         size_t len = 0;
 426         ssize_t sret;
 427         u32 i = 0;
 428         int ret = -1;
 429
 430         sprintf(filename, CORE_SIB_FMT, cpu);
 431         fp = fopen(filename, "r");
 432         if (!fp)
 433                 goto try_threads;
 434
 435         sret = getline(&buf, &len, fp);
 436         fclose(fp);
 437         if (sret <= 0)
 438                 goto try_threads;
 439
 440         p = strchr(buf, '\n');
 441         if (p)
 442                 *p = '\0';
 443
 444         for (i = 0; i < tp->core_sib; i++) {
 445                 if (!strcmp(buf, tp->core_siblings[i]))
 446                         break;
 447         }
 448         if (i == tp->core_sib) {
 449                 tp->core_siblings[i] = buf;
 450                 tp->core_sib++;
 451                 buf = NULL;
 452                 len = 0;
 453         }
 454         ret = 0;
 455
 456 try_threads:
 457         sprintf(filename, THRD_SIB_FMT, cpu);
 458         fp = fopen(filename, "r");
 459         if (!fp)
 460                 goto done;
 461
 462         if (getline(&buf, &len, fp) <= 0)
 463                 goto done;
 464
 465         p = strchr(buf, '\n');
 466         if (p)
 467                 *p = '\0';
 468
 469         for (i = 0; i < tp->thread_sib; i++) {
 470                 if (!strcmp(buf, tp->thread_siblings[i]))
 471                         break;
 472         }
 473         if (i == tp->thread_sib) {
 474                 tp->thread_siblings[i] = buf;
 475                 tp->thread_sib++;
 476                 buf = NULL;
 477         }
 478         ret = 0;
 479 done:
 480         if(fp)
 481                 fclose(fp);
 482         free(buf);
 483         return ret;
 484 }
 485
 486 static void free_cpu_topo(struct cpu_topo *tp)
 487 {
 488         u32 i;
 489
 490         if (!tp)
 491                 return;
 492
 493         for (i = 0 ; i < tp->core_sib; i++)
 494                 zfree(&tp->core_siblings[i]);
 495
 496         for (i = 0 ; i < tp->thread_sib; i++)
 497                 zfree(&tp->thread_siblings[i]);
 498
 499         free(tp);
 500 }
 501
 502 static struct cpu_topo *build_cpu_topology(void)
 503 {
 504         struct cpu_topo *tp;
 505         void *addr;
 506         u32 nr, i;
 507         size_t sz;
 508         long ncpus;
 509         int ret = -1;
 510
 511         ncpus = sysconf(_SC_NPROCESSORS_CONF);
 512         if (ncpus < 0)
 513                 return NULL;
 514
 515         nr = (u32)(ncpus & UINT_MAX);
 516
 517         sz = nr * sizeof(char *);
 518
 519         addr = calloc(1, sizeof(*tp) + 2 * sz);
 520         if (!addr)
 521                 return NULL;
 522
 523         tp = addr;
 524         tp->cpu_nr = nr;
 525         addr += sizeof(*tp);
 526         tp->core_siblings = addr;
 527         addr += sz;
 528         tp->thread_siblings = addr;
 529
 530         for (i = 0; i < nr; i++) {
 531                 ret = build_cpu_topo(tp, i);
 532                 if (ret < 0)
 533                         break;
 534         }
 535         if (ret) {
 536                 free_cpu_topo(tp);
 537                 tp = NULL;
 538         }
 539         return tp;
 540 }
 541
 542 static int write_cpu_topology(int fd, struct perf_header *h __maybe_unused,
 543                           struct perf_evlist *evlist __maybe_unused)
 544 {
 545         struct cpu_topo *tp;
 546         u32 i;
 547         int ret, j;
 548
 549         tp = build_cpu_topology();
 550         if (!tp)
 551                 return -1;
 552
 553         ret = do_write(fd, &tp->core_sib, sizeof(tp->core_sib));
 554         if (ret < 0)
 555                 goto done;
 556
 557         for (i = 0; i < tp->core_sib; i++) {
 558                 ret = do_write_string(fd, tp->core_siblings[i]);
 559                 if (ret < 0)
 560                         goto done;
 561         }
 562         ret = do_write(fd, &tp->thread_sib, sizeof(tp->thread_sib));
 563         if (ret < 0)
 564                 goto done;
 565
 566         for (i = 0; i < tp->thread_sib; i++) {
 567                 ret = do_write_string(fd, tp->thread_siblings[i]);
 568                 if (ret < 0)
 569                         break;
 570         }
 571
 572         ret = perf_env__read_cpu_topology_map(&perf_env);
 573         if (ret < 0)
 574                 goto done;
 575
 576         for (j = 0; j < perf_env.nr_cpus_avail; j++) {
 577                 ret = do_write(fd, &perf_env.cpu[j].core_id,
 578                                sizeof(perf_env.cpu[j].core_id));
 579                 if (ret < 0)
 580                         return ret;
 581                 ret = do_write(fd, &perf_env.cpu[j].socket_id,
 582                                sizeof(perf_env.cpu[j].socket_id));
 583                 if (ret < 0)
 584                         return ret;
 585         }
 586 done:
 587         free_cpu_topo(tp);
 588         return ret;
 589 }
 590
 591
 592
 593 static int write_total_mem(int fd, struct perf_header *h __maybe_unused,
 594                           struct perf_evlist *evlist __maybe_unused)
 595 {
 596         char *buf = NULL;
 597         FILE *fp;
 598         size_t len = 0;
 599         int ret = -1, n;
 600         uint64_t mem;
 601
 602         fp = fopen("/proc/meminfo", "r");
 603         if (!fp)
 604                 return -1;
 605
 606         while (getline(&buf, &len, fp) > 0) {
 607                 ret = strncmp(buf, "MemTotal:", 9);
 608                 if (!ret)
 609                         break;
 610         }
 611         if (!ret) {
 612                 n = sscanf(buf, "%*s %"PRIu64, &mem);
 613                 if (n == 1)
 614                         ret = do_write(fd, &mem, sizeof(mem));
 615         } else
 616                 ret = -1;
 617         free(buf);
 618         fclose(fp);
 619         return ret;
 620 }
 621
 622 static int write_topo_node(int fd, int node)
 623 {
 624         char str[MAXPATHLEN];
 625         char field[32];
 626         char *buf = NULL, *p;
 627         size_t len = 0;
 628         FILE *fp;
 629         u64 mem_total, mem_free, mem;
 630         int ret = -1;
 631
 632         sprintf(str, "/sys/devices/system/node/node%d/meminfo", node);
 633         fp = fopen(str, "r");
 634         if (!fp)
 635                 return -1;
 636
 637         while (getline(&buf, &len, fp) > 0) {
 638                 /* skip over invalid lines */
 639                 if (!strchr(buf, ':'))
 640                         continue;
 641                 if (sscanf(buf, "%*s %*d %31s %"PRIu64, field, &mem) != 2)
 642                         goto done;
 643                 if (!strcmp(field, "MemTotal:"))
 644                         mem_total = mem;
 645                 if (!strcmp(field, "MemFree:"))
 646                         mem_free = mem;
 647         }
 648
 649         fclose(fp);
 650         fp = NULL;
 651
 652         ret = do_write(fd, &mem_total, sizeof(u64));
 653         if (ret)
 654                 goto done;
 655
 656         ret = do_write(fd, &mem_free, sizeof(u64));
 657         if (ret)
 658                 goto done;
 659
 660         ret = -1;
 661         sprintf(str, "/sys/devices/system/node/node%d/cpulist", node);
 662
 663         fp = fopen(str, "r");
 664         if (!fp)
 665                 goto done;
 666
 667         if (getline(&buf, &len, fp) <= 0)
 668                 goto done;
 669
 670         p = strchr(buf, '\n');
 671         if (p)
 672                 *p = '\0';
 673
 674         ret = do_write_string(fd, buf);
 675 done:
 676         free(buf);
 677         if (fp)
 678                 fclose(fp);
 679         return ret;
 680 }
 681
 682 static int write_numa_topology(int fd, struct perf_header *h __maybe_unused,
 683                           struct perf_evlist *evlist __maybe_unused)
 684 {
 685         char *buf = NULL;
 686         size_t len = 0;
 687         FILE *fp;
 688         struct cpu_map *node_map = NULL;
 689         char *c;
 690         u32 nr, i, j;
 691         int ret = -1;
 692
 693         fp = fopen("/sys/devices/system/node/online", "r");
 694         if (!fp)
 695                 return -1;
 696
 697         if (getline(&buf, &len, fp) <= 0)
 698                 goto done;
 699
 700         c = strchr(buf, '\n');
 701         if (c)
 702                 *c = '\0';
 703
 704         node_map = cpu_map__new(buf);
 705         if (!node_map)
 706                 goto done;
 707
 708         nr = (u32)node_map->nr;
 709
 710         ret = do_write(fd, &nr, sizeof(nr));
 711         if (ret < 0)
 712                 goto done;
 713
 714         for (i = 0; i < nr; i++) {
 715                 j = (u32)node_map->map[i];
 716                 ret = do_write(fd, &j, sizeof(j));
 717                 if (ret < 0)
 718                         break;
 719
 720                 ret = write_topo_node(fd, i);
 721                 if (ret < 0)
 722                         break;
 723         }
 724 done:
 725         free(buf);
 726         fclose(fp);
 727         free(node_map);
 728         return ret;
 729 }
 730
 731 /*
 732  * File format:
 733  *
 734  * struct pmu_mappings {
 735  *      u32     pmu_num;
 736  *      struct pmu_map {
 737  *              u32     type;
 738  *              char    name[];
 739  *      }[pmu_num];
 740  * };
 741  */
 742
 743 static int write_pmu_mappings(int fd, struct perf_header *h __maybe_unused,
 744                               struct perf_evlist *evlist __maybe_unused)
 745 {
 746         struct perf_pmu *pmu = NULL;
 747         off_t offset = lseek(fd, 0, SEEK_CUR);
 748         __u32 pmu_num = 0;
 749         int ret;
 750
 751         /* write real pmu_num later */
 752         ret = do_write(fd, &pmu_num, sizeof(pmu_num));
 753         if (ret < 0)
 754                 return ret;
 755
 756         while ((pmu = perf_pmu__scan(pmu))) {
 757                 if (!pmu->name)
 758                         continue;
 759                 pmu_num++;
 760
 761                 ret = do_write(fd, &pmu->type, sizeof(pmu->type));
 762                 if (ret < 0)
 763                         return ret;
 764
 765                 ret = do_write_string(fd, pmu->name);
 766                 if (ret < 0)
 767                         return ret;
 768         }
 769
 770         if (pwrite(fd, &pmu_num, sizeof(pmu_num), offset) != sizeof(pmu_num)) {
 771                 /* discard all */
 772                 lseek(fd, offset, SEEK_SET);
 773                 return -1;
 774         }
 775
 776         return 0;
 777 }
 778
 779 /*
 780  * File format:
 781  *
 782  * struct group_descs {
 783  *      u32     nr_groups;
 784  *      struct group_desc {
 785  *              char    name[];
 786  *              u32     leader_idx;
 787  *              u32     nr_members;
 788  *      }[nr_groups];
 789  * };
 790  */
 791 static int write_group_desc(int fd, struct perf_header *h __maybe_unused,
 792                             struct perf_evlist *evlist)
 793 {
 794         u32 nr_groups = evlist->nr_groups;
 795         struct perf_evsel *evsel;
 796         int ret;
 797
 798         ret = do_write(fd, &nr_groups, sizeof(nr_groups));
 799         if (ret < 0)
 800                 return ret;
 801
 802         evlist__for_each(evlist, evsel) {
 803                 if (perf_evsel__is_group_leader(evsel) &&
 804                     evsel->nr_members > 1) {
 805                         const char *name = evsel->group_name ?: "{anon_group}";
 806                         u32 leader_idx = evsel->idx;
 807                         u32 nr_members = evsel->nr_members;
 808
 809                         ret = do_write_string(fd, name);
 810                         if (ret < 0)
 811                                 return ret;
 812
 813                         ret = do_write(fd, &leader_idx, sizeof(leader_idx));
 814                         if (ret < 0)
 815                                 return ret;
 816
 817                         ret = do_write(fd, &nr_members, sizeof(nr_members));
 818                         if (ret < 0)
 819                                 return ret;
 820                 }
 821         }
 822         return 0;
 823 }
 824
 825 /*
 826  * default get_cpuid(): nothing gets recorded
 827  * actual implementation must be in arch/$(ARCH)/util/header.c
 828  */
 829 int __attribute__ ((weak)) get_cpuid(char *buffer __maybe_unused,
 830                                      size_t sz __maybe_unused)
 831 {
 832         return -1;
 833 }
 834
 835 static int write_cpuid(int fd, struct perf_header *h __maybe_unused,
 836                        struct perf_evlist *evlist __maybe_unused)
 837 {
 838         char buffer[64];
 839         int ret;
 840
 841         ret = get_cpuid(buffer, sizeof(buffer));
 842         if (!ret)
 843                 goto write_it;
 844
 845         return -1;
 846 write_it:
 847         return do_write_string(fd, buffer);
 848 }
 849
 850 static int write_branch_stack(int fd __maybe_unused,
 851                               struct perf_header *h __maybe_unused,
 852                        struct perf_evlist *evlist __maybe_unused)
 853 {
 854         return 0;
 855 }
 856
 857 static int write_auxtrace(int fd, struct perf_header *h,
 858                           struct perf_evlist *evlist __maybe_unused)
 859 {
 860         struct perf_session *session;
 861         int err;
 862
 863         session = container_of(h, struct perf_session, header);
 864
 865         err = auxtrace_index__write(fd, &session->auxtrace_index);
 866         if (err < 0)
 867                 pr_err("Failed to write auxtrace index\n");
 868         return err;
 869 }
 870
 871 static void print_hostname(struct perf_header *ph, int fd __maybe_unused,
 872                            FILE *fp)
 873 {
 874         fprintf(fp, "# hostname : %s\n", ph->env.hostname);
 875 }
 876
 877 static void print_osrelease(struct perf_header *ph, int fd __maybe_unused,
 878                             FILE *fp)
 879 {
 880         fprintf(fp, "# os release : %s\n", ph->env.os_release);
 881 }
 882
 883 static void print_arch(struct perf_header *ph, int fd __maybe_unused, FILE *fp)
 884 {
 885         fprintf(fp, "# arch : %s\n", ph->env.arch);
 886 }
 887
 888 static void print_cpudesc(struct perf_header *ph, int fd __maybe_unused,
 889                           FILE *fp)
 890 {
 891         fprintf(fp, "# cpudesc : %s\n", ph->env.cpu_desc);
 892 }
 893
 894 static void print_nrcpus(struct perf_header *ph, int fd __maybe_unused,
 895                          FILE *fp)
 896 {
 897         fprintf(fp, "# nrcpus online : %u\n", ph->env.nr_cpus_online);
 898         fprintf(fp, "# nrcpus avail : %u\n", ph->env.nr_cpus_avail);
 899 }
 900
 901 static void print_version(struct perf_header *ph, int fd __maybe_unused,
 902                           FILE *fp)
 903 {
 904         fprintf(fp, "# perf version : %s\n", ph->env.version);
 905 }
 906
 907 static void print_cmdline(struct perf_header *ph, int fd __maybe_unused,
 908                           FILE *fp)
 909 {
 910         int nr, i;
 911
 912         nr = ph->env.nr_cmdline;
 913
 914         fprintf(fp, "# cmdline : ");
 915
 916         for (i = 0; i < nr; i++)
 917                 fprintf(fp, "%s ", ph->env.cmdline_argv[i]);
 918         fputc('\n', fp);
 919 }
 920
 921 static void print_cpu_topology(struct perf_header *ph, int fd __maybe_unused,
 922                                FILE *fp)
 923 {
 924         int nr, i;
 925         char *str;
 926         int cpu_nr = ph->env.nr_cpus_online;
 927
 928         nr = ph->env.nr_sibling_cores;
 929         str = ph->env.sibling_cores;
 930
 931         for (i = 0; i < nr; i++) {
 932                 fprintf(fp, "# sibling cores   : %s\n", str);
 933                 str += strlen(str) + 1;
 934         }
 935
 936         nr = ph->env.nr_sibling_threads;
 937         str = ph->env.sibling_threads;
 938
 939         for (i = 0; i < nr; i++) {
 940                 fprintf(fp, "# sibling threads : %s\n", str);
 941                 str += strlen(str) + 1;
 942         }
 943
 944         if (ph->env.cpu != NULL) {
 945                 for (i = 0; i < cpu_nr; i++)
 946                         fprintf(fp, "# CPU %d: Core ID %d, Socket ID %d\n", i,
 947                                 ph->env.cpu[i].core_id, ph->env.cpu[i].socket_id);
 948         } else
 949                 fprintf(fp, "# Core ID and Socket ID information is not available\n");
 950 }
 951
 952 static void free_event_desc(struct perf_evsel *events)
 953 {
 954         struct perf_evsel *evsel;
 955
 956         if (!events)
 957                 return;
 958
 959         for (evsel = events; evsel->attr.size; evsel++) {
 960                 zfree(&evsel->name);
 961                 zfree(&evsel->id);
 962         }
 963
 964         free(events);
 965 }
 966
 967 static struct perf_evsel *
 968 read_event_desc(struct perf_header *ph, int fd)
 969 {
 970         struct perf_evsel *evsel, *events = NULL;
 971         u64 *id;
 972         void *buf = NULL;
 973         u32 nre, sz, nr, i, j;
 974         ssize_t ret;
 975         size_t msz;
 976
 977         /* number of events */
 978         ret = readn(fd, &nre, sizeof(nre));
 979         if (ret != (ssize_t)sizeof(nre))
 980                 goto error;
 981
 982         if (ph->needs_swap)
 983                 nre = bswap_32(nre);
 984
 985         ret = readn(fd, &sz, sizeof(sz));
 986         if (ret != (ssize_t)sizeof(sz))
 987                 goto error;
 988
 989         if (ph->needs_swap)
 990                 sz = bswap_32(sz);
 991
 992         /* buffer to hold on file attr struct */
 993         buf = malloc(sz);
 994         if (!buf)
 995                 goto error;
 996
 997         /* the last event terminates with evsel->attr.size == 0: */
 998         events = calloc(nre + 1, sizeof(*events));
 999         if (!events)
1000                 goto error;
1001
1002         msz = sizeof(evsel->attr);
1003         if (sz < msz)
1004                 msz = sz;
1005
1006         for (i = 0, evsel = events; i < nre; evsel++, i++) {
1007                 evsel->idx = i;
1008
1009                 /*
1010                  * must read entire on-file attr struct to
1011                  * sync up with layout.
1012                  */
1013                 ret = readn(fd, buf, sz);
1014                 if (ret != (ssize_t)sz)
1015                         goto error;
1016
1017                 if (ph->needs_swap)
1018                         perf_event__attr_swap(buf);
1019
1020                 memcpy(&evsel->attr, buf, msz);
1021
1022                 ret = readn(fd, &nr, sizeof(nr));
1023                 if (ret != (ssize_t)sizeof(nr))
1024                         goto error;
1025
1026                 if (ph->needs_swap) {
1027                         nr = bswap_32(nr);
1028                         evsel->needs_swap = true;
1029                 }
1030
1031                 evsel->name = do_read_string(fd, ph);
1032
1033                 if (!nr)
1034                         continue;
1035
1036                 id = calloc(nr, sizeof(*id));
1037                 if (!id)
1038                         goto error;
1039                 evsel->ids = nr;
1040                 evsel->id = id;
1041
1042                 for (j = 0 ; j < nr; j++) {
1043                         ret = readn(fd, id, sizeof(*id));
1044                         if (ret != (ssize_t)sizeof(*id))
1045                                 goto error;
1046                         if (ph->needs_swap)
1047                                 *id = bswap_64(*id);
1048                         id++;
1049                 }
1050         }
1051 out:
1052         free(buf);
1053         return events;
1054 error:
1055         free_event_desc(events);
1056         events = NULL;
1057         goto out;
1058 }
1059
1060 static int __desc_attr__fprintf(FILE *fp, const char *name, const char *val,
1061                                 void *priv __attribute__((unused)))
1062 {
1063         return fprintf(fp, ", %s = %s", name, val);
1064 }
1065
1066 static void print_event_desc(struct perf_header *ph, int fd, FILE *fp)
1067 {
1068         struct perf_evsel *evsel, *events = read_event_desc(ph, fd);
1069         u32 j;
1070         u64 *id;
1071
1072         if (!events) {
1073                 fprintf(fp, "# event desc: not available or unable to read\n");
1074                 return;
1075         }
1076
1077         for (evsel = events; evsel->attr.size; evsel++) {
1078                 fprintf(fp, "# event : name = %s, ", evsel->name);
1079
1080                 if (evsel->ids) {
1081                         fprintf(fp, ", id = {");
1082                         for (j = 0, id = evsel->id; j < evsel->ids; j++, id++) {
1083                                 if (j)
1084                                         fputc(',', fp);
1085                                 fprintf(fp, " %"PRIu64, *id);
1086                         }
1087                         fprintf(fp, " }");
1088                 }
1089
1090                 perf_event_attr__fprintf(fp, &evsel->attr, __desc_attr__fprintf, NULL);
1091
1092                 fputc('\n', fp);
1093         }
1094
1095         free_event_desc(events);
1096 }
1097
1098 static void print_total_mem(struct perf_header *ph, int fd __maybe_unused,
1099                             FILE *fp)
1100 {
1101         fprintf(fp, "# total memory : %Lu kB\n", ph->env.total_mem);
1102 }
1103
1104 static void print_numa_topology(struct perf_header *ph, int fd __maybe_unused,
1105                                 FILE *fp)
1106 {
1107         u32 nr, c, i;
1108         char *str, *tmp;
1109         uint64_t mem_total, mem_free;
1110
1111         /* nr nodes */
1112         nr = ph->env.nr_numa_nodes;
1113         str = ph->env.numa_nodes;
1114
1115         for (i = 0; i < nr; i++) {
1116                 /* node number */
1117                 c = strtoul(str, &tmp, 0);
1118                 if (*tmp != ':')
1119                         goto error;
1120
1121                 str = tmp + 1;
1122                 mem_total = strtoull(str, &tmp, 0);
1123                 if (*tmp != ':')
1124                         goto error;
1125
1126                 str = tmp + 1;
1127                 mem_free = strtoull(str, &tmp, 0);
1128                 if (*tmp != ':')
1129                         goto error;
1130
1131                 fprintf(fp, "# node%u meminfo  : total = %"PRIu64" kB,"
1132                             " free = %"PRIu64" kB\n",
1133                         c, mem_total, mem_free);
1134
1135                 str = tmp + 1;
1136                 fprintf(fp, "# node%u cpu list : %s\n", c, str);
1137
1138                 str += strlen(str) + 1;
1139         }
1140         return;
1141 error:
1142         fprintf(fp, "# numa topology : not available\n");
1143 }
1144
1145 static void print_cpuid(struct perf_header *ph, int fd __maybe_unused, FILE *fp)
1146 {
1147         fprintf(fp, "# cpuid : %s\n", ph->env.cpuid);
1148 }
1149
1150 static void print_branch_stack(struct perf_header *ph __maybe_unused,
1151                                int fd __maybe_unused, FILE *fp)
1152 {
1153         fprintf(fp, "# contains samples with branch stack\n");
1154 }
1155
1156 static void print_auxtrace(struct perf_header *ph __maybe_unused,
1157                            int fd __maybe_unused, FILE *fp)
1158 {
1159         fprintf(fp, "# contains AUX area data (e.g. instruction trace)\n");
1160 }
1161
1162 static void print_pmu_mappings(struct perf_header *ph, int fd __maybe_unused,
1163                                FILE *fp)
1164 {
1165         const char *delimiter = "# pmu mappings: ";
1166         char *str, *tmp;
1167         u32 pmu_num;
1168         u32 type;
1169
1170         pmu_num = ph->env.nr_pmu_mappings;
1171         if (!pmu_num) {
1172                 fprintf(fp, "# pmu mappings: not available\n");
1173                 return;
1174         }
1175
1176         str = ph->env.pmu_mappings;
1177
1178         while (pmu_num) {
1179                 type = strtoul(str, &tmp, 0);
1180                 if (*tmp != ':')
1181                         goto error;
1182
1183                 str = tmp + 1;
1184                 fprintf(fp, "%s%s = %" PRIu32, delimiter, str, type);
1185
1186                 delimiter = ", ";
1187                 str += strlen(str) + 1;
1188                 pmu_num--;
1189         }
1190
1191         fprintf(fp, "\n");
1192
1193         if (!pmu_num)
1194                 return;
1195 error:
1196         fprintf(fp, "# pmu mappings: unable to read\n");
1197 }
1198
1199 static void print_group_desc(struct perf_header *ph, int fd __maybe_unused,
1200                              FILE *fp)
1201 {
1202         struct perf_session *session;
1203         struct perf_evsel *evsel;
1204         u32 nr = 0;
1205
1206         session = container_of(ph, struct perf_session, header);
1207
1208         evlist__for_each(session->evlist, evsel) {
1209                 if (perf_evsel__is_group_leader(evsel) &&
1210                     evsel->nr_members > 1) {
1211                         fprintf(fp, "# group: %s{%s", evsel->group_name ?: "",
1212                                 perf_evsel__name(evsel));
1213
1214                         nr = evsel->nr_members - 1;
1215                 } else if (nr) {
1216                         fprintf(fp, ",%s", perf_evsel__name(evsel));
1217
1218                         if (--nr == 0)
1219                                 fprintf(fp, "}\n");
1220                 }
1221         }
1222 }
1223
1224 static int __event_process_build_id(struct build_id_event *bev,
1225                                     char *filename,
1226                                     struct perf_session *session)
1227 {
1228         int err = -1;
1229         struct machine *machine;
1230         u16 cpumode;
1231         struct dso *dso;
1232         enum dso_kernel_type dso_type;
1233
1234         machine = perf_session__findnew_machine(session, bev->pid);
1235         if (!machine)
1236                 goto out;
1237
1238         cpumode = bev->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
1239
1240         switch (cpumode) {
1241         case PERF_RECORD_MISC_KERNEL:
1242                 dso_type = DSO_TYPE_KERNEL;
1243                 break;
1244         case PERF_RECORD_MISC_GUEST_KERNEL:
1245                 dso_type = DSO_TYPE_GUEST_KERNEL;
1246                 break;
1247         case PERF_RECORD_MISC_USER:
1248         case PERF_RECORD_MISC_GUEST_USER:
1249                 dso_type = DSO_TYPE_USER;
1250                 break;
1251         default:
1252                 goto out;
1253         }
1254
1255         dso = machine__findnew_dso(machine, filename);
1256         if (dso != NULL) {
1257                 char sbuild_id[BUILD_ID_SIZE * 2 + 1];
1258
1259                 dso__set_build_id(dso, &bev->build_id);
1260
1261                 if (!is_kernel_module(filename, cpumode))
1262                         dso->kernel = dso_type;
1263
1264                 build_id__sprintf(dso->build_id, sizeof(dso->build_id),
1265                                   sbuild_id);
1266                 pr_debug("build id event received for %s: %s\n",
1267                          dso->long_name, sbuild_id);
1268                 dso__put(dso);
1269         }
1270
1271         err = 0;
1272 out:
1273         return err;
1274 }
1275
1276 static int perf_header__read_build_ids_abi_quirk(struct perf_header *header,
1277                                                  int input, u64 offset, u64 size)
1278 {
1279         struct perf_session *session = container_of(header, struct perf_session, header);
1280         struct {
1281                 struct perf_event_header   header;
1282                 u8                         build_id[PERF_ALIGN(BUILD_ID_SIZE, sizeof(u64))];
1283                 char                       filename[0];
1284         } old_bev;
1285         struct build_id_event bev;
1286         char filename[PATH_MAX];
1287         u64 limit = offset + size;
1288
1289         while (offset < limit) {
1290                 ssize_t len;
1291
1292                 if (readn(input, &old_bev, sizeof(old_bev)) != sizeof(old_bev))
1293                         return -1;
1294
1295                 if (header->needs_swap)
1296                         perf_event_header__bswap(&old_bev.header);
1297
1298                 len = old_bev.header.size - sizeof(old_bev);
1299                 if (readn(input, filename, len) != len)
1300                         return -1;
1301
1302                 bev.header = old_bev.header;
1303
1304                 /*
1305                  * As the pid is the missing value, we need to fill
1306                  * it properly. The header.misc value give us nice hint.
1307                  */
1308                 bev.pid = HOST_KERNEL_ID;
1309                 if (bev.header.misc == PERF_RECORD_MISC_GUEST_USER ||
1310                     bev.header.misc == PERF_RECORD_MISC_GUEST_KERNEL)
1311                         bev.pid = DEFAULT_GUEST_KERNEL_ID;
1312
1313                 memcpy(bev.build_id, old_bev.build_id, sizeof(bev.build_id));
1314                 __event_process_build_id(&bev, filename, session);
1315
1316                 offset += bev.header.size;
1317         }
1318
1319         return 0;
1320 }
1321
1322 static int perf_header__read_build_ids(struct perf_header *header,
1323                                        int input, u64 offset, u64 size)
1324 {
1325         struct perf_session *session = container_of(header, struct perf_session, header);
1326         struct build_id_event bev;
1327         char filename[PATH_MAX];
1328         u64 limit = offset + size, orig_offset = offset;
1329         int err = -1;
1330
1331         while (offset < limit) {
1332                 ssize_t len;
1333
1334                 if (readn(input, &bev, sizeof(bev)) != sizeof(bev))
1335                         goto out;
1336
1337                 if (header->needs_swap)
1338                         perf_event_header__bswap(&bev.header);
1339
1340                 len = bev.header.size - sizeof(bev);
1341                 if (readn(input, filename, len) != len)
1342                         goto out;
1343                 /*
1344                  * The a1645ce1 changeset:
1345                  *
1346                  * "perf: 'perf kvm' tool for monitoring guest performance from host"
1347                  *
1348                  * Added a field to struct build_id_event that broke the file
1349                  * format.
1350                  *
1351                  * Since the kernel build-id is the first entry, process the
1352                  * table using the old format if the well known
1353                  * '[kernel.kallsyms]' string for the kernel build-id has the
1354                  * first 4 characters chopped off (where the pid_t sits).
1355                  */
1356                 if (memcmp(filename, "nel.kallsyms]", 13) == 0) {
1357                         if (lseek(input, orig_offset, SEEK_SET) == (off_t)-1)
1358                                 return -1;
1359                         return perf_header__read_build_ids_abi_quirk(header, input, offset, size);
1360                 }
1361
1362                 __event_process_build_id(&bev, filename, session);
1363
1364                 offset += bev.header.size;
1365         }
1366         err = 0;
1367 out:
1368         return err;
1369 }
1370
1371 static int process_tracing_data(struct perf_file_section *section __maybe_unused,
1372                                 struct perf_header *ph __maybe_unused,
1373                                 int fd, void *data)
1374 {
1375         ssize_t ret = trace_report(fd, data, false);
1376         return ret < 0 ? -1 : 0;
1377 }
1378
1379 static int process_build_id(struct perf_file_section *section,
1380                             struct perf_header *ph, int fd,
1381                             void *data __maybe_unused)
1382 {
1383         if (perf_header__read_build_ids(ph, fd, section->offset, section->size))
1384                 pr_debug("Failed to read buildids, continuing...\n");
1385         return 0;
1386 }
1387
1388 static int process_hostname(struct perf_file_section *section __maybe_unused,
1389                             struct perf_header *ph, int fd,
1390                             void *data __maybe_unused)
1391 {
1392         ph->env.hostname = do_read_string(fd, ph);
1393         return ph->env.hostname ? 0 : -ENOMEM;
1394 }
1395
1396 static int process_osrelease(struct perf_file_section *section __maybe_unused,
1397                              struct perf_header *ph, int fd,
1398                              void *data __maybe_unused)
1399 {
1400         ph->env.os_release = do_read_string(fd, ph);
1401         return ph->env.os_release ? 0 : -ENOMEM;
1402 }
1403
1404 static int process_version(struct perf_file_section *section __maybe_unused,
1405                            struct perf_header *ph, int fd,
1406                            void *data __maybe_unused)
1407 {
1408         ph->env.version = do_read_string(fd, ph);
1409         return ph->env.version ? 0 : -ENOMEM;
1410 }
1411
1412 static int process_arch(struct perf_file_section *section __maybe_unused,
1413                         struct perf_header *ph, int fd,
1414                         void *data __maybe_unused)
1415 {
1416         ph->env.arch = do_read_string(fd, ph);
1417         return ph->env.arch ? 0 : -ENOMEM;
1418 }
1419
1420 static int process_nrcpus(struct perf_file_section *section __maybe_unused,
1421                           struct perf_header *ph, int fd,
1422                           void *data __maybe_unused)
1423 {
1424         ssize_t ret;
1425         u32 nr;
1426
1427         ret = readn(fd, &nr, sizeof(nr));
1428         if (ret != sizeof(nr))
1429                 return -1;
1430
1431         if (ph->needs_swap)
1432                 nr = bswap_32(nr);
1433
1434         ph->env.nr_cpus_avail = nr;
1435
1436         ret = readn(fd, &nr, sizeof(nr));
1437         if (ret != sizeof(nr))
1438                 return -1;
1439
1440         if (ph->needs_swap)
1441                 nr = bswap_32(nr);
1442
1443         ph->env.nr_cpus_online = nr;
1444         return 0;
1445 }
1446
1447 static int process_cpudesc(struct perf_file_section *section __maybe_unused,
1448                            struct perf_header *ph, int fd,
1449                            void *data __maybe_unused)
1450 {
1451         ph->env.cpu_desc = do_read_string(fd, ph);
1452         return ph->env.cpu_desc ? 0 : -ENOMEM;
1453 }
1454
1455 static int process_cpuid(struct perf_file_section *section __maybe_unused,
1456                          struct perf_header *ph,  int fd,
1457                          void *data __maybe_unused)
1458 {
1459         ph->env.cpuid = do_read_string(fd, ph);
1460         return ph->env.cpuid ? 0 : -ENOMEM;
1461 }
1462
1463 static int process_total_mem(struct perf_file_section *section __maybe_unused,
1464                              struct perf_header *ph, int fd,
1465                              void *data __maybe_unused)
1466 {
1467         uint64_t mem;
1468         ssize_t ret;
1469
1470         ret = readn(fd, &mem, sizeof(mem));
1471         if (ret != sizeof(mem))
1472                 return -1;
1473
1474         if (ph->needs_swap)
1475                 mem = bswap_64(mem);
1476
1477         ph->env.total_mem = mem;
1478         return 0;
1479 }
1480
1481 static struct perf_evsel *
1482 perf_evlist__find_by_index(struct perf_evlist *evlist, int idx)
1483 {
1484         struct perf_evsel *evsel;
1485
1486         evlist__for_each(evlist, evsel) {
1487                 if (evsel->idx == idx)
1488                         return evsel;
1489         }
1490
1491         return NULL;
1492 }
1493
1494 static void
1495 perf_evlist__set_event_name(struct perf_evlist *evlist,
1496                             struct perf_evsel *event)
1497 {
1498         struct perf_evsel *evsel;
1499
1500         if (!event->name)
1501                 return;
1502
1503         evsel = perf_evlist__find_by_index(evlist, event->idx);
1504         if (!evsel)
1505                 return;
1506
1507         if (evsel->name)
1508                 return;
1509
1510         evsel->name = strdup(event->name);
1511 }
1512
1513 static int
1514 process_event_desc(struct perf_file_section *section __maybe_unused,
1515                    struct perf_header *header, int fd,
1516                    void *data __maybe_unused)
1517 {
1518         struct perf_session *session;
1519         struct perf_evsel *evsel, *events = read_event_desc(header, fd);
1520
1521         if (!events)
1522                 return 0;
1523
1524         session = container_of(header, struct perf_session, header);
1525         for (evsel = events; evsel->attr.size; evsel++)
1526                 perf_evlist__set_event_name(session->evlist, evsel);
1527
1528         free_event_desc(events);
1529
1530         return 0;
1531 }
1532
1533 static int process_cmdline(struct perf_file_section *section,
1534                            struct perf_header *ph, int fd,
1535                            void *data __maybe_unused)
1536 {
1537         ssize_t ret;
1538         char *str, *cmdline = NULL, **argv = NULL;
1539         u32 nr, i, len = 0;
1540
1541         ret = readn(fd, &nr, sizeof(nr));
1542         if (ret != sizeof(nr))
1543                 return -1;
1544
1545         if (ph->needs_swap)
1546                 nr = bswap_32(nr);
1547
1548         ph->env.nr_cmdline = nr;
1549
1550         cmdline = zalloc(section->size + nr + 1);
1551         if (!cmdline)
1552                 return -1;
1553
1554         argv = zalloc(sizeof(char *) * (nr + 1));
1555         if (!argv)
1556                 goto error;
1557
1558         for (i = 0; i < nr; i++) {
1559                 str = do_read_string(fd, ph);
1560                 if (!str)
1561                         goto error;
1562
1563                 argv[i] = cmdline + len;
1564                 memcpy(argv[i], str, strlen(str) + 1);
1565                 len += strlen(str) + 1;
1566                 free(str);
1567         }
1568         ph->env.cmdline = cmdline;
1569         ph->env.cmdline_argv = (const char **) argv;
1570         return 0;
1571
1572 error:
1573         free(argv);
1574         free(cmdline);
1575         return -1;
1576 }
1577
1578 static int process_cpu_topology(struct perf_file_section *section,
1579                                 struct perf_header *ph, int fd,
1580                                 void *data __maybe_unused)
1581 {
1582         ssize_t ret;
1583         u32 nr, i;
1584         char *str;
1585         struct strbuf sb;
1586         int cpu_nr = ph->env.nr_cpus_online;
1587         u64 size = 0;
1588
1589         ph->env.cpu = calloc(cpu_nr, sizeof(*ph->env.cpu));
1590         if (!ph->env.cpu)
1591                 return -1;
1592
1593         ret = readn(fd, &nr, sizeof(nr));
1594         if (ret != sizeof(nr))
1595                 goto free_cpu;
1596
1597         if (ph->needs_swap)
1598                 nr = bswap_32(nr);
1599
1600         ph->env.nr_sibling_cores = nr;
1601         size += sizeof(u32);
1602         strbuf_init(&sb, 128);
1603
1604         for (i = 0; i < nr; i++) {
1605                 str = do_read_string(fd, ph);
1606                 if (!str)
1607                         goto error;
1608
1609                 /* include a NULL character at the end */
1610                 strbuf_add(&sb, str, strlen(str) + 1);
1611                 size += string_size(str);
1612                 free(str);
1613         }
1614         ph->env.sibling_cores = strbuf_detach(&sb, NULL);
1615
1616         ret = readn(fd, &nr, sizeof(nr));
1617         if (ret != sizeof(nr))
1618                 return -1;
1619
1620         if (ph->needs_swap)
1621                 nr = bswap_32(nr);
1622
1623         ph->env.nr_sibling_threads = nr;
1624         size += sizeof(u32);
1625
1626         for (i = 0; i < nr; i++) {
1627                 str = do_read_string(fd, ph);
1628                 if (!str)
1629                         goto error;
1630
1631                 /* include a NULL character at the end */
1632                 strbuf_add(&sb, str, strlen(str) + 1);
1633                 size += string_size(str);
1634                 free(str);
1635         }
1636         ph->env.sibling_threads = strbuf_detach(&sb, NULL);
1637
1638         /*
1639          * The header may be from old perf,
1640          * which doesn't include core id and socket id information.
1641          */
1642         if (section->size <= size) {
1643                 zfree(&ph->env.cpu);
1644                 return 0;
1645         }
1646
1647         for (i = 0; i < (u32)cpu_nr; i++) {
1648                 ret = readn(fd, &nr, sizeof(nr));
1649                 if (ret != sizeof(nr))
1650                         goto free_cpu;
1651
1652                 if (ph->needs_swap)
1653                         nr = bswap_32(nr);
1654
1655                 if (nr > (u32)cpu_nr) {
1656                         pr_debug("core_id number is too big."
1657                                  "You may need to upgrade the perf tool.\n");
1658                         goto free_cpu;
1659                 }
1660                 ph->env.cpu[i].core_id = nr;
1661
1662                 ret = readn(fd, &nr, sizeof(nr));
1663                 if (ret != sizeof(nr))
1664                         goto free_cpu;
1665
1666                 if (ph->needs_swap)
1667                         nr = bswap_32(nr);
1668
1669                 if (nr > (u32)cpu_nr) {
1670                         pr_debug("socket_id number is too big."
1671                                  "You may need to upgrade the perf tool.\n");
1672                         goto free_cpu;
1673                 }
1674
1675                 ph->env.cpu[i].socket_id = nr;
1676         }
1677
1678         return 0;
1679
1680 error:
1681         strbuf_release(&sb);
1682 free_cpu:
1683         zfree(&ph->env.cpu);
1684         return -1;
1685 }
1686
1687 static int process_numa_topology(struct perf_file_section *section __maybe_unused,
1688                                  struct perf_header *ph, int fd,
1689                                  void *data __maybe_unused)
1690 {
1691         ssize_t ret;
1692         u32 nr, node, i;
1693         char *str;
1694         uint64_t mem_total, mem_free;
1695         struct strbuf sb;
1696
1697         /* nr nodes */
1698         ret = readn(fd, &nr, sizeof(nr));
1699         if (ret != sizeof(nr))
1700                 goto error;
1701
1702         if (ph->needs_swap)
1703                 nr = bswap_32(nr);
1704
1705         ph->env.nr_numa_nodes = nr;
1706         strbuf_init(&sb, 256);
1707
1708         for (i = 0; i < nr; i++) {
1709                 /* node number */
1710                 ret = readn(fd, &node, sizeof(node));
1711                 if (ret != sizeof(node))
1712                         goto error;
1713
1714                 ret = readn(fd, &mem_total, sizeof(u64));
1715                 if (ret != sizeof(u64))
1716                         goto error;
1717
1718                 ret = readn(fd, &mem_free, sizeof(u64));
1719                 if (ret != sizeof(u64))
1720                         goto error;
1721
1722                 if (ph->needs_swap) {
1723                         node = bswap_32(node);
1724                         mem_total = bswap_64(mem_total);
1725                         mem_free = bswap_64(mem_free);
1726                 }
1727
1728                 strbuf_addf(&sb, "%u:%"PRIu64":%"PRIu64":",
1729                             node, mem_total, mem_free);
1730
1731                 str = do_read_string(fd, ph);
1732                 if (!str)
1733                         goto error;
1734
1735                 /* include a NULL character at the end */
1736                 strbuf_add(&sb, str, strlen(str) + 1);
1737                 free(str);
1738         }
1739         ph->env.numa_nodes = strbuf_detach(&sb, NULL);
1740         return 0;
1741
1742 error:
1743         strbuf_release(&sb);
1744         return -1;
1745 }
1746
1747 static int process_pmu_mappings(struct perf_file_section *section __maybe_unused,
1748                                 struct perf_header *ph, int fd,
1749                                 void *data __maybe_unused)
1750 {
1751         ssize_t ret;
1752         char *name;
1753         u32 pmu_num;
1754         u32 type;
1755         struct strbuf sb;
1756
1757         ret = readn(fd, &pmu_num, sizeof(pmu_num));
1758         if (ret != sizeof(pmu_num))
1759                 return -1;
1760
1761         if (ph->needs_swap)
1762                 pmu_num = bswap_32(pmu_num);
1763
1764         if (!pmu_num) {
1765                 pr_debug("pmu mappings not available\n");
1766                 return 0;
1767         }
1768
1769         ph->env.nr_pmu_mappings = pmu_num;
1770         strbuf_init(&sb, 128);
1771
1772         while (pmu_num) {
1773                 if (readn(fd, &type, sizeof(type)) != sizeof(type))
1774                         goto error;
1775                 if (ph->needs_swap)
1776                         type = bswap_32(type);
1777
1778                 name = do_read_string(fd, ph);
1779                 if (!name)
1780                         goto error;
1781
1782                 strbuf_addf(&sb, "%u:%s", type, name);
1783                 /* include a NULL character at the end */
1784                 strbuf_add(&sb, "", 1);
1785
1786                 if (!strcmp(name, "msr"))
1787                         ph->env.msr_pmu_type = type;
1788
1789                 free(name);
1790                 pmu_num--;
1791         }
1792         ph->env.pmu_mappings = strbuf_detach(&sb, NULL);
1793         return 0;
1794
1795 error:
1796         strbuf_release(&sb);
1797         return -1;
1798 }
1799
1800 static int process_group_desc(struct perf_file_section *section __maybe_unused,
1801                               struct perf_header *ph, int fd,
1802                               void *data __maybe_unused)
1803 {
1804         size_t ret = -1;
1805         u32 i, nr, nr_groups;
1806         struct perf_session *session;
1807         struct perf_evsel *evsel, *leader = NULL;
1808         struct group_desc {
1809                 char *name;
1810                 u32 leader_idx;
1811                 u32 nr_members;
1812         } *desc;
1813
1814         if (readn(fd, &nr_groups, sizeof(nr_groups)) != sizeof(nr_groups))
1815                 return -1;
1816
1817         if (ph->needs_swap)
1818                 nr_groups = bswap_32(nr_groups);
1819
1820         ph->env.nr_groups = nr_groups;
1821         if (!nr_groups) {
1822                 pr_debug("group desc not available\n");
1823                 return 0;
1824         }
1825
1826         desc = calloc(nr_groups, sizeof(*desc));
1827         if (!desc)
1828                 return -1;
1829
1830         for (i = 0; i < nr_groups; i++) {
1831                 desc[i].name = do_read_string(fd, ph);
1832                 if (!desc[i].name)
1833                         goto out_free;
1834
1835                 if (readn(fd, &desc[i].leader_idx, sizeof(u32)) != sizeof(u32))
1836                         goto out_free;
1837
1838                 if (readn(fd, &desc[i].nr_members, sizeof(u32)) != sizeof(u32))
1839                         goto out_free;
1840
1841                 if (ph->needs_swap) {
1842                         desc[i].leader_idx = bswap_32(desc[i].leader_idx);
1843                         desc[i].nr_members = bswap_32(desc[i].nr_members);
1844                 }
1845         }
1846
1847         /*
1848          * Rebuild group relationship based on the group_desc
1849          */
1850         session = container_of(ph, struct perf_session, header);
1851         session->evlist->nr_groups = nr_groups;
1852
1853         i = nr = 0;
1854         evlist__for_each(session->evlist, evsel) {
1855                 if (evsel->idx == (int) desc[i].leader_idx) {
1856                         evsel->leader = evsel;
1857                         /* {anon_group} is a dummy name */
1858                         if (strcmp(desc[i].name, "{anon_group}")) {
1859                                 evsel->group_name = desc[i].name;
1860                                 desc[i].name = NULL;
1861                         }
1862                         evsel->nr_members = desc[i].nr_members;
1863
1864                         if (i >= nr_groups || nr > 0) {
1865                                 pr_debug("invalid group desc\n");
1866                                 goto out_free;
1867                         }
1868
1869                         leader = evsel;
1870                         nr = evsel->nr_members - 1;
1871                         i++;
1872                 } else if (nr) {
1873                         /* This is a group member */
1874                         evsel->leader = leader;
1875
1876                         nr--;
1877                 }
1878         }
1879
1880         if (i != nr_groups || nr != 0) {
1881                 pr_debug("invalid group desc\n");
1882                 goto out_free;
1883         }
1884
1885         ret = 0;
1886 out_free:
1887         for (i = 0; i < nr_groups; i++)
1888                 zfree(&desc[i].name);
1889         free(desc);
1890
1891         return ret;
1892 }
1893
1894 static int process_auxtrace(struct perf_file_section *section,
1895                             struct perf_header *ph, int fd,
1896                             void *data __maybe_unused)
1897 {
1898         struct perf_session *session;
1899         int err;
1900
1901         session = container_of(ph, struct perf_session, header);
1902
1903         err = auxtrace_index__process(fd, section->size, session,
1904                                       ph->needs_swap);
1905         if (err < 0)
1906                 pr_err("Failed to process auxtrace index\n");
1907         return err;
1908 }
1909
1910 struct feature_ops {
1911         int (*write)(int fd, struct perf_header *h, struct perf_evlist *evlist);
1912         void (*print)(struct perf_header *h, int fd, FILE *fp);
1913         int (*process)(struct perf_file_section *section,
1914                        struct perf_header *h, int fd, void *data);
1915         const char *name;
1916         bool full_only;
1917 };
1918
1919 #define FEAT_OPA(n, func) \
1920         [n] = { .name = #n, .write = write_##func, .print = print_##func }
1921 #define FEAT_OPP(n, func) \
1922         [n] = { .name = #n, .write = write_##func, .print = print_##func, \
1923                 .process = process_##func }
1924 #define FEAT_OPF(n, func) \
1925         [n] = { .name = #n, .write = write_##func, .print = print_##func, \
1926                 .process = process_##func, .full_only = true }
1927
1928 /* feature_ops not implemented: */
1929 #define print_tracing_data      NULL
1930 #define print_build_id          NULL
1931
1932 static const struct feature_ops feat_ops[HEADER_LAST_FEATURE] = {
1933         FEAT_OPP(HEADER_TRACING_DATA,   tracing_data),
1934         FEAT_OPP(HEADER_BUILD_ID,       build_id),
1935         FEAT_OPP(HEADER_HOSTNAME,       hostname),
1936         FEAT_OPP(HEADER_OSRELEASE,      osrelease),
1937         FEAT_OPP(HEADER_VERSION,        version),
1938         FEAT_OPP(HEADER_ARCH,           arch),
1939         FEAT_OPP(HEADER_NRCPUS,         nrcpus),
1940         FEAT_OPP(HEADER_CPUDESC,        cpudesc),
1941         FEAT_OPP(HEADER_CPUID,          cpuid),
1942         FEAT_OPP(HEADER_TOTAL_MEM,      total_mem),
1943         FEAT_OPP(HEADER_EVENT_DESC,     event_desc),
1944         FEAT_OPP(HEADER_CMDLINE,        cmdline),
1945         FEAT_OPF(HEADER_CPU_TOPOLOGY,   cpu_topology),
1946         FEAT_OPF(HEADER_NUMA_TOPOLOGY,  numa_topology),
1947         FEAT_OPA(HEADER_BRANCH_STACK,   branch_stack),
1948         FEAT_OPP(HEADER_PMU_MAPPINGS,   pmu_mappings),
1949         FEAT_OPP(HEADER_GROUP_DESC,     group_desc),
1950         FEAT_OPP(HEADER_AUXTRACE,       auxtrace),
1951 };
1952
1953 struct header_print_data {
1954         FILE *fp;
1955         bool full; /* extended list of headers */
1956 };
1957
1958 static int perf_file_section__fprintf_info(struct perf_file_section *section,
1959                                            struct perf_header *ph,
1960                                            int feat, int fd, void *data)
1961 {
1962         struct header_print_data *hd = data;
1963
1964         if (lseek(fd, section->offset, SEEK_SET) == (off_t)-1) {
1965                 pr_debug("Failed to lseek to %" PRIu64 " offset for feature "
1966                                 "%d, continuing...\n", section->offset, feat);
1967                 return 0;
1968         }
1969         if (feat >= HEADER_LAST_FEATURE) {
1970                 pr_warning("unknown feature %d\n", feat);
1971                 return 0;
1972         }
1973         if (!feat_ops[feat].print)
1974                 return 0;
1975
1976         if (!feat_ops[feat].full_only || hd->full)
1977                 feat_ops[feat].print(ph, fd, hd->fp);
1978         else
1979                 fprintf(hd->fp, "# %s info available, use -I to display\n",
1980                         feat_ops[feat].name);
1981
1982         return 0;
1983 }
1984
1985 int perf_header__fprintf_info(struct perf_session *session, FILE *fp, bool full)
1986 {
1987         struct header_print_data hd;
1988         struct perf_header *header = &session->header;
1989         int fd = perf_data_file__fd(session->file);
1990         hd.fp = fp;
1991         hd.full = full;
1992
1993         perf_header__process_sections(header, fd, &hd,
1994                                       perf_file_section__fprintf_info);
1995         return 0;
1996 }
1997
1998 static int do_write_feat(int fd, struct perf_header *h, int type,
1999                          struct perf_file_section **p,
2000                          struct perf_evlist *evlist)
2001 {
2002         int err;
2003         int ret = 0;
2004
2005         if (perf_header__has_feat(h, type)) {
2006                 if (!feat_ops[type].write)
2007                         return -1;
2008
2009                 (*p)->offset = lseek(fd, 0, SEEK_CUR);
2010
2011                 err = feat_ops[type].write(fd, h, evlist);
2012                 if (err < 0) {
2013                         pr_debug("failed to write feature %d\n", type);
2014
2015                         /* undo anything written */
2016                         lseek(fd, (*p)->offset, SEEK_SET);
2017
2018                         return -1;
2019                 }
2020                 (*p)->size = lseek(fd, 0, SEEK_CUR) - (*p)->offset;
2021                 (*p)++;
2022         }
2023         return ret;
2024 }
2025
2026 static int perf_header__adds_write(struct perf_header *header,
2027                                    struct perf_evlist *evlist, int fd)
2028 {
2029         int nr_sections;
2030         struct perf_file_section *feat_sec, *p;
2031         int sec_size;
2032         u64 sec_start;
2033         int feat;
2034         int err;
2035
2036         nr_sections = bitmap_weight(header->adds_features, HEADER_FEAT_BITS);
2037         if (!nr_sections)
2038                 return 0;
2039
2040         feat_sec = p = calloc(nr_sections, sizeof(*feat_sec));
2041         if (feat_sec == NULL)
2042                 return -ENOMEM;
2043
2044         sec_size = sizeof(*feat_sec) * nr_sections;
2045
2046         sec_start = header->feat_offset;
2047         lseek(fd, sec_start + sec_size, SEEK_SET);
2048
2049         for_each_set_bit(feat, header->adds_features, HEADER_FEAT_BITS) {
2050                 if (do_write_feat(fd, header, feat, &p, evlist))
2051                         perf_header__clear_feat(header, feat);
2052         }
2053
2054         lseek(fd, sec_start, SEEK_SET);
2055         /*
2056          * may write more than needed due to dropped feature, but
2057          * this is okay, reader will skip the mising entries
2058          */
2059         err = do_write(fd, feat_sec, sec_size);
2060         if (err < 0)
2061                 pr_debug("failed to write feature section\n");
2062         free(feat_sec);
2063         return err;
2064 }
2065
2066 int perf_header__write_pipe(int fd)
2067 {
2068         struct perf_pipe_file_header f_header;
2069         int err;
2070
2071         f_header = (struct perf_pipe_file_header){
2072                 .magic     = PERF_MAGIC,
2073                 .size      = sizeof(f_header),
2074         };
2075
2076         err = do_write(fd, &f_header, sizeof(f_header));
2077         if (err < 0) {
2078                 pr_debug("failed to write perf pipe header\n");
2079                 return err;
2080         }
2081
2082         return 0;
2083 }
2084
2085 int perf_session__write_header(struct perf_session *session,
2086                                struct perf_evlist *evlist,
2087                                int fd, bool at_exit)
2088 {
2089         struct perf_file_header f_header;
2090         struct perf_file_attr   f_attr;
2091         struct perf_header *header = &session->header;
2092         struct perf_evsel *evsel;
2093         u64 attr_offset;
2094         int err;
2095
2096         lseek(fd, sizeof(f_header), SEEK_SET);
2097
2098         evlist__for_each(session->evlist, evsel) {
2099                 evsel->id_offset = lseek(fd, 0, SEEK_CUR);
2100                 err = do_write(fd, evsel->id, evsel->ids * sizeof(u64));
2101                 if (err < 0) {
2102                         pr_debug("failed to write perf header\n");
2103                         return err;
2104                 }
2105         }
2106
2107         attr_offset = lseek(fd, 0, SEEK_CUR);
2108
2109         evlist__for_each(evlist, evsel) {
2110                 f_attr = (struct perf_file_attr){
2111                         .attr = evsel->attr,
2112                         .ids  = {
2113                                 .offset = evsel->id_offset,
2114                                 .size   = evsel->ids * sizeof(u64),
2115                         }
2116                 };
2117                 err = do_write(fd, &f_attr, sizeof(f_attr));
2118                 if (err < 0) {
2119                         pr_debug("failed to write perf header attribute\n");
2120                         return err;
2121                 }
2122         }
2123
2124         if (!header->data_offset)
2125                 header->data_offset = lseek(fd, 0, SEEK_CUR);
2126         header->feat_offset = header->data_offset + header->data_size;
2127
2128         if (at_exit) {
2129                 err = perf_header__adds_write(header, evlist, fd);
2130                 if (err < 0)
2131                         return err;
2132         }
2133
2134         f_header = (struct perf_file_header){
2135                 .magic     = PERF_MAGIC,
2136                 .size      = sizeof(f_header),
2137                 .attr_size = sizeof(f_attr),
2138                 .attrs = {
2139                         .offset = attr_offset,
2140                         .size   = evlist->nr_entries * sizeof(f_attr),
2141                 },
2142                 .data = {
2143                         .offset = header->data_offset,
2144                         .size   = header->data_size,
2145                 },
2146                 /* event_types is ignored, store zeros */
2147         };
2148
2149         memcpy(&f_header.adds_features, &header->adds_features, sizeof(header->adds_features));
2150
2151         lseek(fd, 0, SEEK_SET);
2152         err = do_write(fd, &f_header, sizeof(f_header));
2153         if (err < 0) {
2154                 pr_debug("failed to write perf header\n");
2155                 return err;
2156         }
2157         lseek(fd, header->data_offset + header->data_size, SEEK_SET);
2158
2159         return 0;
2160 }
2161
2162 static int perf_header__getbuffer64(struct perf_header *header,
2163                                     int fd, void *buf, size_t size)
2164 {
2165         if (readn(fd, buf, size) <= 0)
2166                 return -1;
2167
2168         if (header->needs_swap)
2169                 mem_bswap_64(buf, size);
2170
2171         return 0;
2172 }
2173
2174 int perf_header__process_sections(struct perf_header *header, int fd,
2175                                   void *data,
2176                                   int (*process)(struct perf_file_section *section,
2177                                                  struct perf_header *ph,
2178                                                  int feat, int fd, void *data))
2179 {
2180         struct perf_file_section *feat_sec, *sec;
2181         int nr_sections;
2182         int sec_size;
2183         int feat;
2184         int err;
2185
2186         nr_sections = bitmap_weight(header->adds_features, HEADER_FEAT_BITS);
2187         if (!nr_sections)
2188                 return 0;
2189
2190         feat_sec = sec = calloc(nr_sections, sizeof(*feat_sec));
2191         if (!feat_sec)
2192                 return -1;
2193
2194         sec_size = sizeof(*feat_sec) * nr_sections;
2195
2196         lseek(fd, header->feat_offset, SEEK_SET);
2197
2198         err = perf_header__getbuffer64(header, fd, feat_sec, sec_size);
2199         if (err < 0)
2200                 goto out_free;
2201
2202         for_each_set_bit(feat, header->adds_features, HEADER_LAST_FEATURE) {
2203                 err = process(sec++, header, feat, fd, data);
2204                 if (err < 0)
2205                         goto out_free;
2206         }
2207         err = 0;
2208 out_free:
2209         free(feat_sec);
2210         return err;
2211 }
2212
2213 static const int attr_file_abi_sizes[] = {
2214         [0] = PERF_ATTR_SIZE_VER0,
2215         [1] = PERF_ATTR_SIZE_VER1,
2216         [2] = PERF_ATTR_SIZE_VER2,
2217         [3] = PERF_ATTR_SIZE_VER3,
2218         [4] = PERF_ATTR_SIZE_VER4,
2219         0,
2220 };
2221
2222 /*
2223  * In the legacy file format, the magic number is not used to encode endianness.
2224  * hdr_sz was used to encode endianness. But given that hdr_sz can vary based
2225  * on ABI revisions, we need to try all combinations for all endianness to
2226  * detect the endianness.
2227  */
2228 static int try_all_file_abis(uint64_t hdr_sz, struct perf_header *ph)
2229 {
2230         uint64_t ref_size, attr_size;
2231         int i;
2232
2233         for (i = 0 ; attr_file_abi_sizes[i]; i++) {
2234                 ref_size = attr_file_abi_sizes[i]
2235                          + sizeof(struct perf_file_section);
2236                 if (hdr_sz != ref_size) {
2237                         attr_size = bswap_64(hdr_sz);
2238                         if (attr_size != ref_size)
2239                                 continue;
2240
2241                         ph->needs_swap = true;
2242                 }
2243                 pr_debug("ABI%d perf.data file detected, need_swap=%d\n",
2244                          i,
2245                          ph->needs_swap);
2246                 return 0;
2247         }
2248         /* could not determine endianness */
2249         return -1;
2250 }
2251
2252 #define PERF_PIPE_HDR_VER0      16
2253
2254 static const size_t attr_pipe_abi_sizes[] = {
2255         [0] = PERF_PIPE_HDR_VER0,
2256         0,
2257 };
2258
2259 /*
2260  * In the legacy pipe format, there is an implicit assumption that endiannesss
2261  * between host recording the samples, and host parsing the samples is the
2262  * same. This is not always the case given that the pipe output may always be
2263  * redirected into a file and analyzed on a different machine with possibly a
2264  * different endianness and perf_event ABI revsions in the perf tool itself.
2265  */
2266 static int try_all_pipe_abis(uint64_t hdr_sz, struct perf_header *ph)
2267 {
2268         u64 attr_size;
2269         int i;
2270
2271         for (i = 0 ; attr_pipe_abi_sizes[i]; i++) {
2272                 if (hdr_sz != attr_pipe_abi_sizes[i]) {
2273                         attr_size = bswap_64(hdr_sz);
2274                         if (attr_size != hdr_sz)
2275                                 continue;
2276
2277                         ph->needs_swap = true;
2278                 }
2279                 pr_debug("Pipe ABI%d perf.data file detected\n", i);
2280                 return 0;
2281         }
2282         return -1;
2283 }
2284
2285 bool is_perf_magic(u64 magic)
2286 {
2287         if (!memcmp(&magic, __perf_magic1, sizeof(magic))
2288                 || magic == __perf_magic2
2289                 || magic == __perf_magic2_sw)
2290                 return true;
2291
2292         return false;
2293 }
2294
2295 static int check_magic_endian(u64 magic, uint64_t hdr_sz,
2296                               bool is_pipe, struct perf_header *ph)
2297 {
2298         int ret;
2299
2300         /* check for legacy format */
2301         ret = memcmp(&magic, __perf_magic1, sizeof(magic));
2302         if (ret == 0) {
2303                 ph->version = PERF_HEADER_VERSION_1;
2304                 pr_debug("legacy perf.data format\n");
2305                 if (is_pipe)
2306                         return try_all_pipe_abis(hdr_sz, ph);
2307
2308                 return try_all_file_abis(hdr_sz, ph);
2309         }
2310         /*
2311          * the new magic number serves two purposes:
2312          * - unique number to identify actual perf.data files
2313          * - encode endianness of file
2314          */
2315         ph->version = PERF_HEADER_VERSION_2;
2316
2317         /* check magic number with one endianness */
2318         if (magic == __perf_magic2)
2319                 return 0;
2320
2321         /* check magic number with opposite endianness */
2322         if (magic != __perf_magic2_sw)
2323                 return -1;
2324
2325         ph->needs_swap = true;
2326
2327         return 0;
2328 }
2329
2330 int perf_file_header__read(struct perf_file_header *header,
2331                            struct perf_header *ph, int fd)
2332 {
2333         ssize_t ret;
2334
2335         lseek(fd, 0, SEEK_SET);
2336
2337         ret = readn(fd, header, sizeof(*header));
2338         if (ret <= 0)
2339                 return -1;
2340
2341         if (check_magic_endian(header->magic,
2342                                header->attr_size, false, ph) < 0) {
2343                 pr_debug("magic/endian check failed\n");
2344                 return -1;
2345         }
2346
2347         if (ph->needs_swap) {
2348                 mem_bswap_64(header, offsetof(struct perf_file_header,
2349                              adds_features));
2350         }
2351
2352         if (header->size != sizeof(*header)) {
2353                 /* Support the previous format */
2354                 if (header->size == offsetof(typeof(*header), adds_features))
2355                         bitmap_zero(header->adds_features, HEADER_FEAT_BITS);
2356                 else
2357                         return -1;
2358         } else if (ph->needs_swap) {
2359                 /*
2360                  * feature bitmap is declared as an array of unsigned longs --
2361                  * not good since its size can differ between the host that
2362                  * generated the data file and the host analyzing the file.
2363                  *
2364                  * We need to handle endianness, but we don't know the size of
2365                  * the unsigned long where the file was generated. Take a best
2366                  * guess at determining it: try 64-bit swap first (ie., file
2367                  * created on a 64-bit host), and check if the hostname feature
2368                  * bit is set (this feature bit is forced on as of fbe96f2).
2369                  * If the bit is not, undo the 64-bit swap and try a 32-bit
2370                  * swap. If the hostname bit is still not set (e.g., older data
2371                  * file), punt and fallback to the original behavior --
2372                  * clearing all feature bits and setting buildid.
2373                  */
2374                 mem_bswap_64(&header->adds_features,
2375                             BITS_TO_U64(HEADER_FEAT_BITS));
2376
2377                 if (!test_bit(HEADER_HOSTNAME, header->adds_features)) {
2378                         /* unswap as u64 */
2379                         mem_bswap_64(&header->adds_features,
2380                                     BITS_TO_U64(HEADER_FEAT_BITS));
2381
2382                         /* unswap as u32 */
2383                         mem_bswap_32(&header->adds_features,
2384                                     BITS_TO_U32(HEADER_FEAT_BITS));
2385                 }
2386
2387                 if (!test_bit(HEADER_HOSTNAME, header->adds_features)) {
2388                         bitmap_zero(header->adds_features, HEADER_FEAT_BITS);
2389                         set_bit(HEADER_BUILD_ID, header->adds_features);
2390                 }
2391         }
2392
2393         memcpy(&ph->adds_features, &header->adds_features,
2394                sizeof(ph->adds_features));
2395
2396         ph->data_offset  = header->data.offset;
2397         ph->data_size    = header->data.size;
2398         ph->feat_offset  = header->data.offset + header->data.size;
2399         return 0;
2400 }
2401
2402 static int perf_file_section__process(struct perf_file_section *section,
2403                                       struct perf_header *ph,
2404                                       int feat, int fd, void *data)
2405 {
2406         if (lseek(fd, section->offset, SEEK_SET) == (off_t)-1) {
2407                 pr_debug("Failed to lseek to %" PRIu64 " offset for feature "
2408                           "%d, continuing...\n", section->offset, feat);
2409                 return 0;
2410         }
2411
2412         if (feat >= HEADER_LAST_FEATURE) {
2413                 pr_debug("unknown feature %d, continuing...\n", feat);
2414                 return 0;
2415         }
2416
2417         if (!feat_ops[feat].process)
2418                 return 0;
2419
2420         return feat_ops[feat].process(section, ph, fd, data);
2421 }
2422
2423 static int perf_file_header__read_pipe(struct perf_pipe_file_header *header,
2424                                        struct perf_header *ph, int fd,
2425                                        bool repipe)
2426 {
2427         ssize_t ret;
2428
2429         ret = readn(fd, header, sizeof(*header));
2430         if (ret <= 0)
2431                 return -1;
2432
2433         if (check_magic_endian(header->magic, header->size, true, ph) < 0) {
2434                 pr_debug("endian/magic failed\n");
2435                 return -1;
2436         }
2437
2438         if (ph->needs_swap)
2439                 header->size = bswap_64(header->size);
2440
2441         if (repipe && do_write(STDOUT_FILENO, header, sizeof(*header)) < 0)
2442                 return -1;
2443
2444         return 0;
2445 }
2446
2447 static int perf_header__read_pipe(struct perf_session *session)
2448 {
2449         struct perf_header *header = &session->header;
2450         struct perf_pipe_file_header f_header;
2451
2452         if (perf_file_header__read_pipe(&f_header, header,
2453                                         perf_data_file__fd(session->file),
2454                                         session->repipe) < 0) {
2455                 pr_debug("incompatible file format\n");
2456                 return -EINVAL;
2457         }
2458
2459         return 0;
2460 }
2461
2462 static int read_attr(int fd, struct perf_header *ph,
2463                      struct perf_file_attr *f_attr)
2464 {
2465         struct perf_event_attr *attr = &f_attr->attr;
2466         size_t sz, left;
2467         size_t our_sz = sizeof(f_attr->attr);
2468         ssize_t ret;
2469
2470         memset(f_attr, 0, sizeof(*f_attr));
2471
2472         /* read minimal guaranteed structure */
2473         ret = readn(fd, attr, PERF_ATTR_SIZE_VER0);
2474         if (ret <= 0) {
2475                 pr_debug("cannot read %d bytes of header attr\n",
2476                          PERF_ATTR_SIZE_VER0);
2477                 return -1;
2478         }
2479
2480         /* on file perf_event_attr size */
2481         sz = attr->size;
2482
2483         if (ph->needs_swap)
2484                 sz = bswap_32(sz);
2485
2486         if (sz == 0) {
2487                 /* assume ABI0 */
2488                 sz =  PERF_ATTR_SIZE_VER0;
2489         } else if (sz > our_sz) {
2490                 pr_debug("file uses a more recent and unsupported ABI"
2491                          " (%zu bytes extra)\n", sz - our_sz);
2492                 return -1;
2493         }
2494         /* what we have not yet read and that we know about */
2495         left = sz - PERF_ATTR_SIZE_VER0;
2496         if (left) {
2497                 void *ptr = attr;
2498                 ptr += PERF_ATTR_SIZE_VER0;
2499
2500                 ret = readn(fd, ptr, left);
2501         }
2502         /* read perf_file_section, ids are read in caller */
2503         ret = readn(fd, &f_attr->ids, sizeof(f_attr->ids));
2504
2505         return ret <= 0 ? -1 : 0;
2506 }
2507
2508 static int perf_evsel__prepare_tracepoint_event(struct perf_evsel *evsel,
2509                                                 struct pevent *pevent)
2510 {
2511         struct event_format *event;
2512         char bf[128];
2513
2514         /* already prepared */
2515         if (evsel->tp_format)
2516                 return 0;
2517
2518         if (pevent == NULL) {
2519                 pr_debug("broken or missing trace data\n");
2520                 return -1;
2521         }
2522
2523         event = pevent_find_event(pevent, evsel->attr.config);
2524         if (event == NULL)
2525                 return -1;
2526
2527         if (!evsel->name) {
2528                 snprintf(bf, sizeof(bf), "%s:%s", event->system, event->name);
2529                 evsel->name = strdup(bf);
2530                 if (evsel->name == NULL)
2531                         return -1;
2532         }
2533
2534         evsel->tp_format = event;
2535         return 0;
2536 }
2537
2538 static int perf_evlist__prepare_tracepoint_events(struct perf_evlist *evlist,
2539                                                   struct pevent *pevent)
2540 {
2541         struct perf_evsel *pos;
2542
2543         evlist__for_each(evlist, pos) {
2544                 if (pos->attr.type == PERF_TYPE_TRACEPOINT &&
2545                     perf_evsel__prepare_tracepoint_event(pos, pevent))
2546                         return -1;
2547         }
2548
2549         return 0;
2550 }
2551
2552 int perf_session__read_header(struct perf_session *session)
2553 {
2554         struct perf_data_file *file = session->file;
2555         struct perf_header *header = &session->header;
2556         struct perf_file_header f_header;
2557         struct perf_file_attr   f_attr;
2558         u64                     f_id;
2559         int nr_attrs, nr_ids, i, j;
2560         int fd = perf_data_file__fd(file);
2561
2562         session->evlist = perf_evlist__new();
2563         if (session->evlist == NULL)
2564                 return -ENOMEM;
2565
2566         session->evlist->env = &header->env;
2567         session->machines.host.env = &header->env;
2568         if (perf_data_file__is_pipe(file))
2569                 return perf_header__read_pipe(session);
2570
2571         if (perf_file_header__read(&f_header, header, fd) < 0)
2572                 return -EINVAL;
2573
2574         /*
2575          * Sanity check that perf.data was written cleanly; data size is
2576          * initialized to 0 and updated only if the on_exit function is run.
2577          * If data size is still 0 then the file contains only partial
2578          * information.  Just warn user and process it as much as it can.
2579          */
2580         if (f_header.data.size == 0) {
2581                 pr_warning("WARNING: The %s file's data size field is 0 which is unexpected.\n"
2582                            "Was the 'perf record' command properly terminated?\n",
2583                            file->path);
2584         }
2585
2586         nr_attrs = f_header.attrs.size / f_header.attr_size;
2587         lseek(fd, f_header.attrs.offset, SEEK_SET);
2588
2589         for (i = 0; i < nr_attrs; i++) {
2590                 struct perf_evsel *evsel;
2591                 off_t tmp;
2592
2593                 if (read_attr(fd, header, &f_attr) < 0)
2594                         goto out_errno;
2595
2596                 if (header->needs_swap) {
2597                         f_attr.ids.size   = bswap_64(f_attr.ids.size);
2598                         f_attr.ids.offset = bswap_64(f_attr.ids.offset);
2599                         perf_event__attr_swap(&f_attr.attr);
2600                 }
2601
2602                 tmp = lseek(fd, 0, SEEK_CUR);
2603                 evsel = perf_evsel__new(&f_attr.attr);
2604
2605                 if (evsel == NULL)
2606                         goto out_delete_evlist;
2607
2608                 evsel->needs_swap = header->needs_swap;
2609                 /*
2610                  * Do it before so that if perf_evsel__alloc_id fails, this
2611                  * entry gets purged too at perf_evlist__delete().
2612                  */
2613                 perf_evlist__add(session->evlist, evsel);
2614
2615                 nr_ids = f_attr.ids.size / sizeof(u64);
2616                 /*
2617                  * We don't have the cpu and thread maps on the header, so
2618                  * for allocating the perf_sample_id table we fake 1 cpu and
2619                  * hattr->ids threads.
2620                  */
2621                 if (perf_evsel__alloc_id(evsel, 1, nr_ids))
2622                         goto out_delete_evlist;
2623
2624                 lseek(fd, f_attr.ids.offset, SEEK_SET);
2625
2626                 for (j = 0; j < nr_ids; j++) {
2627                         if (perf_header__getbuffer64(header, fd, &f_id, sizeof(f_id)))
2628                                 goto out_errno;
2629
2630                         perf_evlist__id_add(session->evlist, evsel, 0, j, f_id);
2631                 }
2632
2633                 lseek(fd, tmp, SEEK_SET);
2634         }
2635
2636         symbol_conf.nr_events = nr_attrs;
2637
2638         perf_header__process_sections(header, fd, &session->tevent,
2639                                       perf_file_section__process);
2640
2641         if (perf_evlist__prepare_tracepoint_events(session->evlist,
2642                                                    session->tevent.pevent))
2643                 goto out_delete_evlist;
2644
2645         return 0;
2646 out_errno:
2647         return -errno;
2648
2649 out_delete_evlist:
2650         perf_evlist__delete(session->evlist);
2651         session->evlist = NULL;
2652         return -ENOMEM;
2653 }
2654
2655 int perf_event__synthesize_attr(struct perf_tool *tool,
2656                                 struct perf_event_attr *attr, u32 ids, u64 *id,
2657                                 perf_event__handler_t process)
2658 {
2659         union perf_event *ev;
2660         size_t size;
2661         int err;
2662
2663         size = sizeof(struct perf_event_attr);
2664         size = PERF_ALIGN(size, sizeof(u64));
2665         size += sizeof(struct perf_event_header);
2666         size += ids * sizeof(u64);
2667
2668         ev = malloc(size);
2669
2670         if (ev == NULL)
2671                 return -ENOMEM;
2672
2673         ev->attr.attr = *attr;
2674         memcpy(ev->attr.id, id, ids * sizeof(u64));
2675
2676         ev->attr.header.type = PERF_RECORD_HEADER_ATTR;
2677         ev->attr.header.size = (u16)size;
2678
2679         if (ev->attr.header.size == size)
2680                 err = process(tool, ev, NULL, NULL);
2681         else
2682                 err = -E2BIG;
2683
2684         free(ev);
2685
2686         return err;
2687 }
2688
2689 int perf_event__synthesize_attrs(struct perf_tool *tool,
2690                                    struct perf_session *session,
2691                                    perf_event__handler_t process)
2692 {
2693         struct perf_evsel *evsel;
2694         int err = 0;
2695
2696         evlist__for_each(session->evlist, evsel) {
2697                 err = perf_event__synthesize_attr(tool, &evsel->attr, evsel->ids,
2698                                                   evsel->id, process);
2699                 if (err) {
2700                         pr_debug("failed to create perf header attribute\n");
2701                         return err;
2702                 }
2703         }
2704
2705         return err;
2706 }
2707
2708 int perf_event__process_attr(struct perf_tool *tool __maybe_unused,
2709                              union perf_event *event,
2710                              struct perf_evlist **pevlist)
2711 {
2712         u32 i, ids, n_ids;
2713         struct perf_evsel *evsel;
2714         struct perf_evlist *evlist = *pevlist;
2715
2716         if (evlist == NULL) {
2717                 *pevlist = evlist = perf_evlist__new();
2718                 if (evlist == NULL)
2719                         return -ENOMEM;
2720         }
2721
2722         evsel = perf_evsel__new(&event->attr.attr);
2723         if (evsel == NULL)
2724                 return -ENOMEM;
2725
2726         perf_evlist__add(evlist, evsel);
2727
2728         ids = event->header.size;
2729         ids -= (void *)&event->attr.id - (void *)event;
2730         n_ids = ids / sizeof(u64);
2731         /*
2732          * We don't have the cpu and thread maps on the header, so
2733          * for allocating the perf_sample_id table we fake 1 cpu and
2734          * hattr->ids threads.
2735          */
2736         if (perf_evsel__alloc_id(evsel, 1, n_ids))
2737                 return -ENOMEM;
2738
2739         for (i = 0; i < n_ids; i++) {
2740                 perf_evlist__id_add(evlist, evsel, 0, i, event->attr.id[i]);
2741         }
2742
2743         symbol_conf.nr_events = evlist->nr_entries;
2744
2745         return 0;
2746 }
2747
2748 int perf_event__synthesize_tracing_data(struct perf_tool *tool, int fd,
2749                                         struct perf_evlist *evlist,
2750                                         perf_event__handler_t process)
2751 {
2752         union perf_event ev;
2753         struct tracing_data *tdata;
2754         ssize_t size = 0, aligned_size = 0, padding;
2755         int err __maybe_unused = 0;
2756
2757         /*
2758          * We are going to store the size of the data followed
2759          * by the data contents. Since the fd descriptor is a pipe,
2760          * we cannot seek back to store the size of the data once
2761          * we know it. Instead we:
2762          *
2763          * - write the tracing data to the temp file
2764          * - get/write the data size to pipe
2765          * - write the tracing data from the temp file
2766          *   to the pipe
2767          */
2768         tdata = tracing_data_get(&evlist->entries, fd, true);
2769         if (!tdata)
2770                 return -1;
2771
2772         memset(&ev, 0, sizeof(ev));
2773
2774         ev.tracing_data.header.type = PERF_RECORD_HEADER_TRACING_DATA;
2775         size = tdata->size;
2776         aligned_size = PERF_ALIGN(size, sizeof(u64));
2777         padding = aligned_size - size;
2778         ev.tracing_data.header.size = sizeof(ev.tracing_data);
2779         ev.tracing_data.size = aligned_size;
2780
2781         process(tool, &ev, NULL, NULL);
2782
2783         /*
2784          * The put function will copy all the tracing data
2785          * stored in temp file to the pipe.
2786          */
2787         tracing_data_put(tdata);
2788
2789         write_padded(fd, NULL, 0, padding);
2790
2791         return aligned_size;
2792 }
2793
2794 int perf_event__process_tracing_data(struct perf_tool *tool __maybe_unused,
2795                                      union perf_event *event,
2796                                      struct perf_session *session)
2797 {
2798         ssize_t size_read, padding, size = event->tracing_data.size;
2799         int fd = perf_data_file__fd(session->file);
2800         off_t offset = lseek(fd, 0, SEEK_CUR);
2801         char buf[BUFSIZ];
2802
2803         /* setup for reading amidst mmap */
2804         lseek(fd, offset + sizeof(struct tracing_data_event),
2805               SEEK_SET);
2806
2807         size_read = trace_report(fd, &session->tevent,
2808                                  session->repipe);
2809         padding = PERF_ALIGN(size_read, sizeof(u64)) - size_read;
2810
2811         if (readn(fd, buf, padding) < 0) {
2812                 pr_err("%s: reading input file", __func__);
2813                 return -1;
2814         }
2815         if (session->repipe) {
2816                 int retw = write(STDOUT_FILENO, buf, padding);
2817                 if (retw <= 0 || retw != padding) {
2818                         pr_err("%s: repiping tracing data padding", __func__);
2819                         return -1;
2820                 }
2821         }
2822
2823         if (size_read + padding != size) {
2824                 pr_err("%s: tracing data size mismatch", __func__);
2825                 return -1;
2826         }
2827
2828         perf_evlist__prepare_tracepoint_events(session->evlist,
2829                                                session->tevent.pevent);
2830
2831         return size_read + padding;
2832 }
2833
2834 int perf_event__synthesize_build_id(struct perf_tool *tool,
2835                                     struct dso *pos, u16 misc,
2836                                     perf_event__handler_t process,
2837                                     struct machine *machine)
2838 {
2839         union perf_event ev;
2840         size_t len;
2841         int err = 0;
2842
2843         if (!pos->hit)
2844                 return err;
2845
2846         memset(&ev, 0, sizeof(ev));
2847
2848         len = pos->long_name_len + 1;
2849         len = PERF_ALIGN(len, NAME_ALIGN);
2850         memcpy(&ev.build_id.build_id, pos->build_id, sizeof(pos->build_id));
2851         ev.build_id.header.type = PERF_RECORD_HEADER_BUILD_ID;
2852         ev.build_id.header.misc = misc;
2853         ev.build_id.pid = machine->pid;
2854         ev.build_id.header.size = sizeof(ev.build_id) + len;
2855         memcpy(&ev.build_id.filename, pos->long_name, pos->long_name_len);
2856
2857         err = process(tool, &ev, NULL, machine);
2858
2859         return err;
2860 }
2861
2862 int perf_event__process_build_id(struct perf_tool *tool __maybe_unused,
2863                                  union perf_event *event,
2864                                  struct perf_session *session)
2865 {
2866         __event_process_build_id(&event->build_id,
2867                                  event->build_id.filename,
2868                                  session);
2869         return 0;
2870 }