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 (dso_type != DSO_TYPE_USER) {
1262                         struct kmod_path m = { .name = NULL, };
1263
1264                         if (!kmod_path__parse_name(&m, filename) && m.kmod)
1265                                 dso__set_short_name(dso, strdup(m.name), true);
1266                         else
1267                                 dso->kernel = dso_type;
1268
1269                         free(m.name);
1270                 }
1271
1272                 build_id__sprintf(dso->build_id, sizeof(dso->build_id),
1273                                   sbuild_id);
1274                 pr_debug("build id event received for %s: %s\n",
1275                          dso->long_name, sbuild_id);
1276                 dso__put(dso);
1277         }
1278
1279         err = 0;
1280 out:
1281         return err;
1282 }
1283
1284 static int perf_header__read_build_ids_abi_quirk(struct perf_header *header,
1285                                                  int input, u64 offset, u64 size)
1286 {
1287         struct perf_session *session = container_of(header, struct perf_session, header);
1288         struct {
1289                 struct perf_event_header   header;
1290                 u8                         build_id[PERF_ALIGN(BUILD_ID_SIZE, sizeof(u64))];
1291                 char                       filename[0];
1292         } old_bev;
1293         struct build_id_event bev;
1294         char filename[PATH_MAX];
1295         u64 limit = offset + size;
1296
1297         while (offset < limit) {
1298                 ssize_t len;
1299
1300                 if (readn(input, &old_bev, sizeof(old_bev)) != sizeof(old_bev))
1301                         return -1;
1302
1303                 if (header->needs_swap)
1304                         perf_event_header__bswap(&old_bev.header);
1305
1306                 len = old_bev.header.size - sizeof(old_bev);
1307                 if (readn(input, filename, len) != len)
1308                         return -1;
1309
1310                 bev.header = old_bev.header;
1311
1312                 /*
1313                  * As the pid is the missing value, we need to fill
1314                  * it properly. The header.misc value give us nice hint.
1315                  */
1316                 bev.pid = HOST_KERNEL_ID;
1317                 if (bev.header.misc == PERF_RECORD_MISC_GUEST_USER ||
1318                     bev.header.misc == PERF_RECORD_MISC_GUEST_KERNEL)
1319                         bev.pid = DEFAULT_GUEST_KERNEL_ID;
1320
1321                 memcpy(bev.build_id, old_bev.build_id, sizeof(bev.build_id));
1322                 __event_process_build_id(&bev, filename, session);
1323
1324                 offset += bev.header.size;
1325         }
1326
1327         return 0;
1328 }
1329
1330 static int perf_header__read_build_ids(struct perf_header *header,
1331                                        int input, u64 offset, u64 size)
1332 {
1333         struct perf_session *session = container_of(header, struct perf_session, header);
1334         struct build_id_event bev;
1335         char filename[PATH_MAX];
1336         u64 limit = offset + size, orig_offset = offset;
1337         int err = -1;
1338
1339         while (offset < limit) {
1340                 ssize_t len;
1341
1342                 if (readn(input, &bev, sizeof(bev)) != sizeof(bev))
1343                         goto out;
1344
1345                 if (header->needs_swap)
1346                         perf_event_header__bswap(&bev.header);
1347
1348                 len = bev.header.size - sizeof(bev);
1349                 if (readn(input, filename, len) != len)
1350                         goto out;
1351                 /*
1352                  * The a1645ce1 changeset:
1353                  *
1354                  * "perf: 'perf kvm' tool for monitoring guest performance from host"
1355                  *
1356                  * Added a field to struct build_id_event that broke the file
1357                  * format.
1358                  *
1359                  * Since the kernel build-id is the first entry, process the
1360                  * table using the old format if the well known
1361                  * '[kernel.kallsyms]' string for the kernel build-id has the
1362                  * first 4 characters chopped off (where the pid_t sits).
1363                  */
1364                 if (memcmp(filename, "nel.kallsyms]", 13) == 0) {
1365                         if (lseek(input, orig_offset, SEEK_SET) == (off_t)-1)
1366                                 return -1;
1367                         return perf_header__read_build_ids_abi_quirk(header, input, offset, size);
1368                 }
1369
1370                 __event_process_build_id(&bev, filename, session);
1371
1372                 offset += bev.header.size;
1373         }
1374         err = 0;
1375 out:
1376         return err;
1377 }
1378
1379 static int process_tracing_data(struct perf_file_section *section __maybe_unused,
1380                                 struct perf_header *ph __maybe_unused,
1381                                 int fd, void *data)
1382 {
1383         ssize_t ret = trace_report(fd, data, false);
1384         return ret < 0 ? -1 : 0;
1385 }
1386
1387 static int process_build_id(struct perf_file_section *section,
1388                             struct perf_header *ph, int fd,
1389                             void *data __maybe_unused)
1390 {
1391         if (perf_header__read_build_ids(ph, fd, section->offset, section->size))
1392                 pr_debug("Failed to read buildids, continuing...\n");
1393         return 0;
1394 }
1395
1396 static int process_hostname(struct perf_file_section *section __maybe_unused,
1397                             struct perf_header *ph, int fd,
1398                             void *data __maybe_unused)
1399 {
1400         ph->env.hostname = do_read_string(fd, ph);
1401         return ph->env.hostname ? 0 : -ENOMEM;
1402 }
1403
1404 static int process_osrelease(struct perf_file_section *section __maybe_unused,
1405                              struct perf_header *ph, int fd,
1406                              void *data __maybe_unused)
1407 {
1408         ph->env.os_release = do_read_string(fd, ph);
1409         return ph->env.os_release ? 0 : -ENOMEM;
1410 }
1411
1412 static int process_version(struct perf_file_section *section __maybe_unused,
1413                            struct perf_header *ph, int fd,
1414                            void *data __maybe_unused)
1415 {
1416         ph->env.version = do_read_string(fd, ph);
1417         return ph->env.version ? 0 : -ENOMEM;
1418 }
1419
1420 static int process_arch(struct perf_file_section *section __maybe_unused,
1421                         struct perf_header *ph, int fd,
1422                         void *data __maybe_unused)
1423 {
1424         ph->env.arch = do_read_string(fd, ph);
1425         return ph->env.arch ? 0 : -ENOMEM;
1426 }
1427
1428 static int process_nrcpus(struct perf_file_section *section __maybe_unused,
1429                           struct perf_header *ph, int fd,
1430                           void *data __maybe_unused)
1431 {
1432         ssize_t ret;
1433         u32 nr;
1434
1435         ret = readn(fd, &nr, sizeof(nr));
1436         if (ret != sizeof(nr))
1437                 return -1;
1438
1439         if (ph->needs_swap)
1440                 nr = bswap_32(nr);
1441
1442         ph->env.nr_cpus_avail = nr;
1443
1444         ret = readn(fd, &nr, sizeof(nr));
1445         if (ret != sizeof(nr))
1446                 return -1;
1447
1448         if (ph->needs_swap)
1449                 nr = bswap_32(nr);
1450
1451         ph->env.nr_cpus_online = nr;
1452         return 0;
1453 }
1454
1455 static int process_cpudesc(struct perf_file_section *section __maybe_unused,
1456                            struct perf_header *ph, int fd,
1457                            void *data __maybe_unused)
1458 {
1459         ph->env.cpu_desc = do_read_string(fd, ph);
1460         return ph->env.cpu_desc ? 0 : -ENOMEM;
1461 }
1462
1463 static int process_cpuid(struct perf_file_section *section __maybe_unused,
1464                          struct perf_header *ph,  int fd,
1465                          void *data __maybe_unused)
1466 {
1467         ph->env.cpuid = do_read_string(fd, ph);
1468         return ph->env.cpuid ? 0 : -ENOMEM;
1469 }
1470
1471 static int process_total_mem(struct perf_file_section *section __maybe_unused,
1472                              struct perf_header *ph, int fd,
1473                              void *data __maybe_unused)
1474 {
1475         uint64_t mem;
1476         ssize_t ret;
1477
1478         ret = readn(fd, &mem, sizeof(mem));
1479         if (ret != sizeof(mem))
1480                 return -1;
1481
1482         if (ph->needs_swap)
1483                 mem = bswap_64(mem);
1484
1485         ph->env.total_mem = mem;
1486         return 0;
1487 }
1488
1489 static struct perf_evsel *
1490 perf_evlist__find_by_index(struct perf_evlist *evlist, int idx)
1491 {
1492         struct perf_evsel *evsel;
1493
1494         evlist__for_each(evlist, evsel) {
1495                 if (evsel->idx == idx)
1496                         return evsel;
1497         }
1498
1499         return NULL;
1500 }
1501
1502 static void
1503 perf_evlist__set_event_name(struct perf_evlist *evlist,
1504                             struct perf_evsel *event)
1505 {
1506         struct perf_evsel *evsel;
1507
1508         if (!event->name)
1509                 return;
1510
1511         evsel = perf_evlist__find_by_index(evlist, event->idx);
1512         if (!evsel)
1513                 return;
1514
1515         if (evsel->name)
1516                 return;
1517
1518         evsel->name = strdup(event->name);
1519 }
1520
1521 static int
1522 process_event_desc(struct perf_file_section *section __maybe_unused,
1523                    struct perf_header *header, int fd,
1524                    void *data __maybe_unused)
1525 {
1526         struct perf_session *session;
1527         struct perf_evsel *evsel, *events = read_event_desc(header, fd);
1528
1529         if (!events)
1530                 return 0;
1531
1532         session = container_of(header, struct perf_session, header);
1533         for (evsel = events; evsel->attr.size; evsel++)
1534                 perf_evlist__set_event_name(session->evlist, evsel);
1535
1536         free_event_desc(events);
1537
1538         return 0;
1539 }
1540
1541 static int process_cmdline(struct perf_file_section *section,
1542                            struct perf_header *ph, int fd,
1543                            void *data __maybe_unused)
1544 {
1545         ssize_t ret;
1546         char *str, *cmdline = NULL, **argv = NULL;
1547         u32 nr, i, len = 0;
1548
1549         ret = readn(fd, &nr, sizeof(nr));
1550         if (ret != sizeof(nr))
1551                 return -1;
1552
1553         if (ph->needs_swap)
1554                 nr = bswap_32(nr);
1555
1556         ph->env.nr_cmdline = nr;
1557
1558         cmdline = zalloc(section->size + nr + 1);
1559         if (!cmdline)
1560                 return -1;
1561
1562         argv = zalloc(sizeof(char *) * (nr + 1));
1563         if (!argv)
1564                 goto error;
1565
1566         for (i = 0; i < nr; i++) {
1567                 str = do_read_string(fd, ph);
1568                 if (!str)
1569                         goto error;
1570
1571                 argv[i] = cmdline + len;
1572                 memcpy(argv[i], str, strlen(str) + 1);
1573                 len += strlen(str) + 1;
1574                 free(str);
1575         }
1576         ph->env.cmdline = cmdline;
1577         ph->env.cmdline_argv = (const char **) argv;
1578         return 0;
1579
1580 error:
1581         free(argv);
1582         free(cmdline);
1583         return -1;
1584 }
1585
1586 static int process_cpu_topology(struct perf_file_section *section,
1587                                 struct perf_header *ph, int fd,
1588                                 void *data __maybe_unused)
1589 {
1590         ssize_t ret;
1591         u32 nr, i;
1592         char *str;
1593         struct strbuf sb;
1594         int cpu_nr = ph->env.nr_cpus_online;
1595         u64 size = 0;
1596
1597         ph->env.cpu = calloc(cpu_nr, sizeof(*ph->env.cpu));
1598         if (!ph->env.cpu)
1599                 return -1;
1600
1601         ret = readn(fd, &nr, sizeof(nr));
1602         if (ret != sizeof(nr))
1603                 goto free_cpu;
1604
1605         if (ph->needs_swap)
1606                 nr = bswap_32(nr);
1607
1608         ph->env.nr_sibling_cores = nr;
1609         size += sizeof(u32);
1610         strbuf_init(&sb, 128);
1611
1612         for (i = 0; i < nr; i++) {
1613                 str = do_read_string(fd, ph);
1614                 if (!str)
1615                         goto error;
1616
1617                 /* include a NULL character at the end */
1618                 strbuf_add(&sb, str, strlen(str) + 1);
1619                 size += string_size(str);
1620                 free(str);
1621         }
1622         ph->env.sibling_cores = strbuf_detach(&sb, NULL);
1623
1624         ret = readn(fd, &nr, sizeof(nr));
1625         if (ret != sizeof(nr))
1626                 return -1;
1627
1628         if (ph->needs_swap)
1629                 nr = bswap_32(nr);
1630
1631         ph->env.nr_sibling_threads = nr;
1632         size += sizeof(u32);
1633
1634         for (i = 0; i < nr; i++) {
1635                 str = do_read_string(fd, ph);
1636                 if (!str)
1637                         goto error;
1638
1639                 /* include a NULL character at the end */
1640                 strbuf_add(&sb, str, strlen(str) + 1);
1641                 size += string_size(str);
1642                 free(str);
1643         }
1644         ph->env.sibling_threads = strbuf_detach(&sb, NULL);
1645
1646         /*
1647          * The header may be from old perf,
1648          * which doesn't include core id and socket id information.
1649          */
1650         if (section->size <= size) {
1651                 zfree(&ph->env.cpu);
1652                 return 0;
1653         }
1654
1655         for (i = 0; i < (u32)cpu_nr; i++) {
1656                 ret = readn(fd, &nr, sizeof(nr));
1657                 if (ret != sizeof(nr))
1658                         goto free_cpu;
1659
1660                 if (ph->needs_swap)
1661                         nr = bswap_32(nr);
1662
1663                 if (nr > (u32)cpu_nr) {
1664                         pr_debug("core_id number is too big."
1665                                  "You may need to upgrade the perf tool.\n");
1666                         goto free_cpu;
1667                 }
1668                 ph->env.cpu[i].core_id = nr;
1669
1670                 ret = readn(fd, &nr, sizeof(nr));
1671                 if (ret != sizeof(nr))
1672                         goto free_cpu;
1673
1674                 if (ph->needs_swap)
1675                         nr = bswap_32(nr);
1676
1677                 if (nr > (u32)cpu_nr) {
1678                         pr_debug("socket_id number is too big."
1679                                  "You may need to upgrade the perf tool.\n");
1680                         goto free_cpu;
1681                 }
1682
1683                 ph->env.cpu[i].socket_id = nr;
1684         }
1685
1686         return 0;
1687
1688 error:
1689         strbuf_release(&sb);
1690 free_cpu:
1691         zfree(&ph->env.cpu);
1692         return -1;
1693 }
1694
1695 static int process_numa_topology(struct perf_file_section *section __maybe_unused,
1696                                  struct perf_header *ph, int fd,
1697                                  void *data __maybe_unused)
1698 {
1699         ssize_t ret;
1700         u32 nr, node, i;
1701         char *str;
1702         uint64_t mem_total, mem_free;
1703         struct strbuf sb;
1704
1705         /* nr nodes */
1706         ret = readn(fd, &nr, sizeof(nr));
1707         if (ret != sizeof(nr))
1708                 goto error;
1709
1710         if (ph->needs_swap)
1711                 nr = bswap_32(nr);
1712
1713         ph->env.nr_numa_nodes = nr;
1714         strbuf_init(&sb, 256);
1715
1716         for (i = 0; i < nr; i++) {
1717                 /* node number */
1718                 ret = readn(fd, &node, sizeof(node));
1719                 if (ret != sizeof(node))
1720                         goto error;
1721
1722                 ret = readn(fd, &mem_total, sizeof(u64));
1723                 if (ret != sizeof(u64))
1724                         goto error;
1725
1726                 ret = readn(fd, &mem_free, sizeof(u64));
1727                 if (ret != sizeof(u64))
1728                         goto error;
1729
1730                 if (ph->needs_swap) {
1731                         node = bswap_32(node);
1732                         mem_total = bswap_64(mem_total);
1733                         mem_free = bswap_64(mem_free);
1734                 }
1735
1736                 strbuf_addf(&sb, "%u:%"PRIu64":%"PRIu64":",
1737                             node, mem_total, mem_free);
1738
1739                 str = do_read_string(fd, ph);
1740                 if (!str)
1741                         goto error;
1742
1743                 /* include a NULL character at the end */
1744                 strbuf_add(&sb, str, strlen(str) + 1);
1745                 free(str);
1746         }
1747         ph->env.numa_nodes = strbuf_detach(&sb, NULL);
1748         return 0;
1749
1750 error:
1751         strbuf_release(&sb);
1752         return -1;
1753 }
1754
1755 static int process_pmu_mappings(struct perf_file_section *section __maybe_unused,
1756                                 struct perf_header *ph, int fd,
1757                                 void *data __maybe_unused)
1758 {
1759         ssize_t ret;
1760         char *name;
1761         u32 pmu_num;
1762         u32 type;
1763         struct strbuf sb;
1764
1765         ret = readn(fd, &pmu_num, sizeof(pmu_num));
1766         if (ret != sizeof(pmu_num))
1767                 return -1;
1768
1769         if (ph->needs_swap)
1770                 pmu_num = bswap_32(pmu_num);
1771
1772         if (!pmu_num) {
1773                 pr_debug("pmu mappings not available\n");
1774                 return 0;
1775         }
1776
1777         ph->env.nr_pmu_mappings = pmu_num;
1778         strbuf_init(&sb, 128);
1779
1780         while (pmu_num) {
1781                 if (readn(fd, &type, sizeof(type)) != sizeof(type))
1782                         goto error;
1783                 if (ph->needs_swap)
1784                         type = bswap_32(type);
1785
1786                 name = do_read_string(fd, ph);
1787                 if (!name)
1788                         goto error;
1789
1790                 strbuf_addf(&sb, "%u:%s", type, name);
1791                 /* include a NULL character at the end */
1792                 strbuf_add(&sb, "", 1);
1793
1794                 if (!strcmp(name, "msr"))
1795                         ph->env.msr_pmu_type = type;
1796
1797                 free(name);
1798                 pmu_num--;
1799         }
1800         ph->env.pmu_mappings = strbuf_detach(&sb, NULL);
1801         return 0;
1802
1803 error:
1804         strbuf_release(&sb);
1805         return -1;
1806 }
1807
1808 static int process_group_desc(struct perf_file_section *section __maybe_unused,
1809                               struct perf_header *ph, int fd,
1810                               void *data __maybe_unused)
1811 {
1812         size_t ret = -1;
1813         u32 i, nr, nr_groups;
1814         struct perf_session *session;
1815         struct perf_evsel *evsel, *leader = NULL;
1816         struct group_desc {
1817                 char *name;
1818                 u32 leader_idx;
1819                 u32 nr_members;
1820         } *desc;
1821
1822         if (readn(fd, &nr_groups, sizeof(nr_groups)) != sizeof(nr_groups))
1823                 return -1;
1824
1825         if (ph->needs_swap)
1826                 nr_groups = bswap_32(nr_groups);
1827
1828         ph->env.nr_groups = nr_groups;
1829         if (!nr_groups) {
1830                 pr_debug("group desc not available\n");
1831                 return 0;
1832         }
1833
1834         desc = calloc(nr_groups, sizeof(*desc));
1835         if (!desc)
1836                 return -1;
1837
1838         for (i = 0; i < nr_groups; i++) {
1839                 desc[i].name = do_read_string(fd, ph);
1840                 if (!desc[i].name)
1841                         goto out_free;
1842
1843                 if (readn(fd, &desc[i].leader_idx, sizeof(u32)) != sizeof(u32))
1844                         goto out_free;
1845
1846                 if (readn(fd, &desc[i].nr_members, sizeof(u32)) != sizeof(u32))
1847                         goto out_free;
1848
1849                 if (ph->needs_swap) {
1850                         desc[i].leader_idx = bswap_32(desc[i].leader_idx);
1851                         desc[i].nr_members = bswap_32(desc[i].nr_members);
1852                 }
1853         }
1854
1855         /*
1856          * Rebuild group relationship based on the group_desc
1857          */
1858         session = container_of(ph, struct perf_session, header);
1859         session->evlist->nr_groups = nr_groups;
1860
1861         i = nr = 0;
1862         evlist__for_each(session->evlist, evsel) {
1863                 if (evsel->idx == (int) desc[i].leader_idx) {
1864                         evsel->leader = evsel;
1865                         /* {anon_group} is a dummy name */
1866                         if (strcmp(desc[i].name, "{anon_group}")) {
1867                                 evsel->group_name = desc[i].name;
1868                                 desc[i].name = NULL;
1869                         }
1870                         evsel->nr_members = desc[i].nr_members;
1871
1872                         if (i >= nr_groups || nr > 0) {
1873                                 pr_debug("invalid group desc\n");
1874                                 goto out_free;
1875                         }
1876
1877                         leader = evsel;
1878                         nr = evsel->nr_members - 1;
1879                         i++;
1880                 } else if (nr) {
1881                         /* This is a group member */
1882                         evsel->leader = leader;
1883
1884                         nr--;
1885                 }
1886         }
1887
1888         if (i != nr_groups || nr != 0) {
1889                 pr_debug("invalid group desc\n");
1890                 goto out_free;
1891         }
1892
1893         ret = 0;
1894 out_free:
1895         for (i = 0; i < nr_groups; i++)
1896                 zfree(&desc[i].name);
1897         free(desc);
1898
1899         return ret;
1900 }
1901
1902 static int process_auxtrace(struct perf_file_section *section,
1903                             struct perf_header *ph, int fd,
1904                             void *data __maybe_unused)
1905 {
1906         struct perf_session *session;
1907         int err;
1908
1909         session = container_of(ph, struct perf_session, header);
1910
1911         err = auxtrace_index__process(fd, section->size, session,
1912                                       ph->needs_swap);
1913         if (err < 0)
1914                 pr_err("Failed to process auxtrace index\n");
1915         return err;
1916 }
1917
1918 struct feature_ops {
1919         int (*write)(int fd, struct perf_header *h, struct perf_evlist *evlist);
1920         void (*print)(struct perf_header *h, int fd, FILE *fp);
1921         int (*process)(struct perf_file_section *section,
1922                        struct perf_header *h, int fd, void *data);
1923         const char *name;
1924         bool full_only;
1925 };
1926
1927 #define FEAT_OPA(n, func) \
1928         [n] = { .name = #n, .write = write_##func, .print = print_##func }
1929 #define FEAT_OPP(n, func) \
1930         [n] = { .name = #n, .write = write_##func, .print = print_##func, \
1931                 .process = process_##func }
1932 #define FEAT_OPF(n, func) \
1933         [n] = { .name = #n, .write = write_##func, .print = print_##func, \
1934                 .process = process_##func, .full_only = true }
1935
1936 /* feature_ops not implemented: */
1937 #define print_tracing_data      NULL
1938 #define print_build_id          NULL
1939
1940 static const struct feature_ops feat_ops[HEADER_LAST_FEATURE] = {
1941         FEAT_OPP(HEADER_TRACING_DATA,   tracing_data),
1942         FEAT_OPP(HEADER_BUILD_ID,       build_id),
1943         FEAT_OPP(HEADER_HOSTNAME,       hostname),
1944         FEAT_OPP(HEADER_OSRELEASE,      osrelease),
1945         FEAT_OPP(HEADER_VERSION,        version),
1946         FEAT_OPP(HEADER_ARCH,           arch),
1947         FEAT_OPP(HEADER_NRCPUS,         nrcpus),
1948         FEAT_OPP(HEADER_CPUDESC,        cpudesc),
1949         FEAT_OPP(HEADER_CPUID,          cpuid),
1950         FEAT_OPP(HEADER_TOTAL_MEM,      total_mem),
1951         FEAT_OPP(HEADER_EVENT_DESC,     event_desc),
1952         FEAT_OPP(HEADER_CMDLINE,        cmdline),
1953         FEAT_OPF(HEADER_CPU_TOPOLOGY,   cpu_topology),
1954         FEAT_OPF(HEADER_NUMA_TOPOLOGY,  numa_topology),
1955         FEAT_OPA(HEADER_BRANCH_STACK,   branch_stack),
1956         FEAT_OPP(HEADER_PMU_MAPPINGS,   pmu_mappings),
1957         FEAT_OPP(HEADER_GROUP_DESC,     group_desc),
1958         FEAT_OPP(HEADER_AUXTRACE,       auxtrace),
1959 };
1960
1961 struct header_print_data {
1962         FILE *fp;
1963         bool full; /* extended list of headers */
1964 };
1965
1966 static int perf_file_section__fprintf_info(struct perf_file_section *section,
1967                                            struct perf_header *ph,
1968                                            int feat, int fd, void *data)
1969 {
1970         struct header_print_data *hd = data;
1971
1972         if (lseek(fd, section->offset, SEEK_SET) == (off_t)-1) {
1973                 pr_debug("Failed to lseek to %" PRIu64 " offset for feature "
1974                                 "%d, continuing...\n", section->offset, feat);
1975                 return 0;
1976         }
1977         if (feat >= HEADER_LAST_FEATURE) {
1978                 pr_warning("unknown feature %d\n", feat);
1979                 return 0;
1980         }
1981         if (!feat_ops[feat].print)
1982                 return 0;
1983
1984         if (!feat_ops[feat].full_only || hd->full)
1985                 feat_ops[feat].print(ph, fd, hd->fp);
1986         else
1987                 fprintf(hd->fp, "# %s info available, use -I to display\n",
1988                         feat_ops[feat].name);
1989
1990         return 0;
1991 }
1992
1993 int perf_header__fprintf_info(struct perf_session *session, FILE *fp, bool full)
1994 {
1995         struct header_print_data hd;
1996         struct perf_header *header = &session->header;
1997         int fd = perf_data_file__fd(session->file);
1998         hd.fp = fp;
1999         hd.full = full;
2000
2001         perf_header__process_sections(header, fd, &hd,
2002                                       perf_file_section__fprintf_info);
2003         return 0;
2004 }
2005
2006 static int do_write_feat(int fd, struct perf_header *h, int type,
2007                          struct perf_file_section **p,
2008                          struct perf_evlist *evlist)
2009 {
2010         int err;
2011         int ret = 0;
2012
2013         if (perf_header__has_feat(h, type)) {
2014                 if (!feat_ops[type].write)
2015                         return -1;
2016
2017                 (*p)->offset = lseek(fd, 0, SEEK_CUR);
2018
2019                 err = feat_ops[type].write(fd, h, evlist);
2020                 if (err < 0) {
2021                         pr_debug("failed to write feature %d\n", type);
2022
2023                         /* undo anything written */
2024                         lseek(fd, (*p)->offset, SEEK_SET);
2025
2026                         return -1;
2027                 }
2028                 (*p)->size = lseek(fd, 0, SEEK_CUR) - (*p)->offset;
2029                 (*p)++;
2030         }
2031         return ret;
2032 }
2033
2034 static int perf_header__adds_write(struct perf_header *header,
2035                                    struct perf_evlist *evlist, int fd)
2036 {
2037         int nr_sections;
2038         struct perf_file_section *feat_sec, *p;
2039         int sec_size;
2040         u64 sec_start;
2041         int feat;
2042         int err;
2043
2044         nr_sections = bitmap_weight(header->adds_features, HEADER_FEAT_BITS);
2045         if (!nr_sections)
2046                 return 0;
2047
2048         feat_sec = p = calloc(nr_sections, sizeof(*feat_sec));
2049         if (feat_sec == NULL)
2050                 return -ENOMEM;
2051
2052         sec_size = sizeof(*feat_sec) * nr_sections;
2053
2054         sec_start = header->feat_offset;
2055         lseek(fd, sec_start + sec_size, SEEK_SET);
2056
2057         for_each_set_bit(feat, header->adds_features, HEADER_FEAT_BITS) {
2058                 if (do_write_feat(fd, header, feat, &p, evlist))
2059                         perf_header__clear_feat(header, feat);
2060         }
2061
2062         lseek(fd, sec_start, SEEK_SET);
2063         /*
2064          * may write more than needed due to dropped feature, but
2065          * this is okay, reader will skip the mising entries
2066          */
2067         err = do_write(fd, feat_sec, sec_size);
2068         if (err < 0)
2069                 pr_debug("failed to write feature section\n");
2070         free(feat_sec);
2071         return err;
2072 }
2073
2074 int perf_header__write_pipe(int fd)
2075 {
2076         struct perf_pipe_file_header f_header;
2077         int err;
2078
2079         f_header = (struct perf_pipe_file_header){
2080                 .magic     = PERF_MAGIC,
2081                 .size      = sizeof(f_header),
2082         };
2083
2084         err = do_write(fd, &f_header, sizeof(f_header));
2085         if (err < 0) {
2086                 pr_debug("failed to write perf pipe header\n");
2087                 return err;
2088         }
2089
2090         return 0;
2091 }
2092
2093 int perf_session__write_header(struct perf_session *session,
2094                                struct perf_evlist *evlist,
2095                                int fd, bool at_exit)
2096 {
2097         struct perf_file_header f_header;
2098         struct perf_file_attr   f_attr;
2099         struct perf_header *header = &session->header;
2100         struct perf_evsel *evsel;
2101         u64 attr_offset;
2102         int err;
2103
2104         lseek(fd, sizeof(f_header), SEEK_SET);
2105
2106         evlist__for_each(session->evlist, evsel) {
2107                 evsel->id_offset = lseek(fd, 0, SEEK_CUR);
2108                 err = do_write(fd, evsel->id, evsel->ids * sizeof(u64));
2109                 if (err < 0) {
2110                         pr_debug("failed to write perf header\n");
2111                         return err;
2112                 }
2113         }
2114
2115         attr_offset = lseek(fd, 0, SEEK_CUR);
2116
2117         evlist__for_each(evlist, evsel) {
2118                 f_attr = (struct perf_file_attr){
2119                         .attr = evsel->attr,
2120                         .ids  = {
2121                                 .offset = evsel->id_offset,
2122                                 .size   = evsel->ids * sizeof(u64),
2123                         }
2124                 };
2125                 err = do_write(fd, &f_attr, sizeof(f_attr));
2126                 if (err < 0) {
2127                         pr_debug("failed to write perf header attribute\n");
2128                         return err;
2129                 }
2130         }
2131
2132         if (!header->data_offset)
2133                 header->data_offset = lseek(fd, 0, SEEK_CUR);
2134         header->feat_offset = header->data_offset + header->data_size;
2135
2136         if (at_exit) {
2137                 err = perf_header__adds_write(header, evlist, fd);
2138                 if (err < 0)
2139                         return err;
2140         }
2141
2142         f_header = (struct perf_file_header){
2143                 .magic     = PERF_MAGIC,
2144                 .size      = sizeof(f_header),
2145                 .attr_size = sizeof(f_attr),
2146                 .attrs = {
2147                         .offset = attr_offset,
2148                         .size   = evlist->nr_entries * sizeof(f_attr),
2149                 },
2150                 .data = {
2151                         .offset = header->data_offset,
2152                         .size   = header->data_size,
2153                 },
2154                 /* event_types is ignored, store zeros */
2155         };
2156
2157         memcpy(&f_header.adds_features, &header->adds_features, sizeof(header->adds_features));
2158
2159         lseek(fd, 0, SEEK_SET);
2160         err = do_write(fd, &f_header, sizeof(f_header));
2161         if (err < 0) {
2162                 pr_debug("failed to write perf header\n");
2163                 return err;
2164         }
2165         lseek(fd, header->data_offset + header->data_size, SEEK_SET);
2166
2167         return 0;
2168 }
2169
2170 static int perf_header__getbuffer64(struct perf_header *header,
2171                                     int fd, void *buf, size_t size)
2172 {
2173         if (readn(fd, buf, size) <= 0)
2174                 return -1;
2175
2176         if (header->needs_swap)
2177                 mem_bswap_64(buf, size);
2178
2179         return 0;
2180 }
2181
2182 int perf_header__process_sections(struct perf_header *header, int fd,
2183                                   void *data,
2184                                   int (*process)(struct perf_file_section *section,
2185                                                  struct perf_header *ph,
2186                                                  int feat, int fd, void *data))
2187 {
2188         struct perf_file_section *feat_sec, *sec;
2189         int nr_sections;
2190         int sec_size;
2191         int feat;
2192         int err;
2193
2194         nr_sections = bitmap_weight(header->adds_features, HEADER_FEAT_BITS);
2195         if (!nr_sections)
2196                 return 0;
2197
2198         feat_sec = sec = calloc(nr_sections, sizeof(*feat_sec));
2199         if (!feat_sec)
2200                 return -1;
2201
2202         sec_size = sizeof(*feat_sec) * nr_sections;
2203
2204         lseek(fd, header->feat_offset, SEEK_SET);
2205
2206         err = perf_header__getbuffer64(header, fd, feat_sec, sec_size);
2207         if (err < 0)
2208                 goto out_free;
2209
2210         for_each_set_bit(feat, header->adds_features, HEADER_LAST_FEATURE) {
2211                 err = process(sec++, header, feat, fd, data);
2212                 if (err < 0)
2213                         goto out_free;
2214         }
2215         err = 0;
2216 out_free:
2217         free(feat_sec);
2218         return err;
2219 }
2220
2221 static const int attr_file_abi_sizes[] = {
2222         [0] = PERF_ATTR_SIZE_VER0,
2223         [1] = PERF_ATTR_SIZE_VER1,
2224         [2] = PERF_ATTR_SIZE_VER2,
2225         [3] = PERF_ATTR_SIZE_VER3,
2226         [4] = PERF_ATTR_SIZE_VER4,
2227         0,
2228 };
2229
2230 /*
2231  * In the legacy file format, the magic number is not used to encode endianness.
2232  * hdr_sz was used to encode endianness. But given that hdr_sz can vary based
2233  * on ABI revisions, we need to try all combinations for all endianness to
2234  * detect the endianness.
2235  */
2236 static int try_all_file_abis(uint64_t hdr_sz, struct perf_header *ph)
2237 {
2238         uint64_t ref_size, attr_size;
2239         int i;
2240
2241         for (i = 0 ; attr_file_abi_sizes[i]; i++) {
2242                 ref_size = attr_file_abi_sizes[i]
2243                          + sizeof(struct perf_file_section);
2244                 if (hdr_sz != ref_size) {
2245                         attr_size = bswap_64(hdr_sz);
2246                         if (attr_size != ref_size)
2247                                 continue;
2248
2249                         ph->needs_swap = true;
2250                 }
2251                 pr_debug("ABI%d perf.data file detected, need_swap=%d\n",
2252                          i,
2253                          ph->needs_swap);
2254                 return 0;
2255         }
2256         /* could not determine endianness */
2257         return -1;
2258 }
2259
2260 #define PERF_PIPE_HDR_VER0      16
2261
2262 static const size_t attr_pipe_abi_sizes[] = {
2263         [0] = PERF_PIPE_HDR_VER0,
2264         0,
2265 };
2266
2267 /*
2268  * In the legacy pipe format, there is an implicit assumption that endiannesss
2269  * between host recording the samples, and host parsing the samples is the
2270  * same. This is not always the case given that the pipe output may always be
2271  * redirected into a file and analyzed on a different machine with possibly a
2272  * different endianness and perf_event ABI revsions in the perf tool itself.
2273  */
2274 static int try_all_pipe_abis(uint64_t hdr_sz, struct perf_header *ph)
2275 {
2276         u64 attr_size;
2277         int i;
2278
2279         for (i = 0 ; attr_pipe_abi_sizes[i]; i++) {
2280                 if (hdr_sz != attr_pipe_abi_sizes[i]) {
2281                         attr_size = bswap_64(hdr_sz);
2282                         if (attr_size != hdr_sz)
2283                                 continue;
2284
2285                         ph->needs_swap = true;
2286                 }
2287                 pr_debug("Pipe ABI%d perf.data file detected\n", i);
2288                 return 0;
2289         }
2290         return -1;
2291 }
2292
2293 bool is_perf_magic(u64 magic)
2294 {
2295         if (!memcmp(&magic, __perf_magic1, sizeof(magic))
2296                 || magic == __perf_magic2
2297                 || magic == __perf_magic2_sw)
2298                 return true;
2299
2300         return false;
2301 }
2302
2303 static int check_magic_endian(u64 magic, uint64_t hdr_sz,
2304                               bool is_pipe, struct perf_header *ph)
2305 {
2306         int ret;
2307
2308         /* check for legacy format */
2309         ret = memcmp(&magic, __perf_magic1, sizeof(magic));
2310         if (ret == 0) {
2311                 ph->version = PERF_HEADER_VERSION_1;
2312                 pr_debug("legacy perf.data format\n");
2313                 if (is_pipe)
2314                         return try_all_pipe_abis(hdr_sz, ph);
2315
2316                 return try_all_file_abis(hdr_sz, ph);
2317         }
2318         /*
2319          * the new magic number serves two purposes:
2320          * - unique number to identify actual perf.data files
2321          * - encode endianness of file
2322          */
2323         ph->version = PERF_HEADER_VERSION_2;
2324
2325         /* check magic number with one endianness */
2326         if (magic == __perf_magic2)
2327                 return 0;
2328
2329         /* check magic number with opposite endianness */
2330         if (magic != __perf_magic2_sw)
2331                 return -1;
2332
2333         ph->needs_swap = true;
2334
2335         return 0;
2336 }
2337
2338 int perf_file_header__read(struct perf_file_header *header,
2339                            struct perf_header *ph, int fd)
2340 {
2341         ssize_t ret;
2342
2343         lseek(fd, 0, SEEK_SET);
2344
2345         ret = readn(fd, header, sizeof(*header));
2346         if (ret <= 0)
2347                 return -1;
2348
2349         if (check_magic_endian(header->magic,
2350                                header->attr_size, false, ph) < 0) {
2351                 pr_debug("magic/endian check failed\n");
2352                 return -1;
2353         }
2354
2355         if (ph->needs_swap) {
2356                 mem_bswap_64(header, offsetof(struct perf_file_header,
2357                              adds_features));
2358         }
2359
2360         if (header->size != sizeof(*header)) {
2361                 /* Support the previous format */
2362                 if (header->size == offsetof(typeof(*header), adds_features))
2363                         bitmap_zero(header->adds_features, HEADER_FEAT_BITS);
2364                 else
2365                         return -1;
2366         } else if (ph->needs_swap) {
2367                 /*
2368                  * feature bitmap is declared as an array of unsigned longs --
2369                  * not good since its size can differ between the host that
2370                  * generated the data file and the host analyzing the file.
2371                  *
2372                  * We need to handle endianness, but we don't know the size of
2373                  * the unsigned long where the file was generated. Take a best
2374                  * guess at determining it: try 64-bit swap first (ie., file
2375                  * created on a 64-bit host), and check if the hostname feature
2376                  * bit is set (this feature bit is forced on as of fbe96f2).
2377                  * If the bit is not, undo the 64-bit swap and try a 32-bit
2378                  * swap. If the hostname bit is still not set (e.g., older data
2379                  * file), punt and fallback to the original behavior --
2380                  * clearing all feature bits and setting buildid.
2381                  */
2382                 mem_bswap_64(&header->adds_features,
2383                             BITS_TO_U64(HEADER_FEAT_BITS));
2384
2385                 if (!test_bit(HEADER_HOSTNAME, header->adds_features)) {
2386                         /* unswap as u64 */
2387                         mem_bswap_64(&header->adds_features,
2388                                     BITS_TO_U64(HEADER_FEAT_BITS));
2389
2390                         /* unswap as u32 */
2391                         mem_bswap_32(&header->adds_features,
2392                                     BITS_TO_U32(HEADER_FEAT_BITS));
2393                 }
2394
2395                 if (!test_bit(HEADER_HOSTNAME, header->adds_features)) {
2396                         bitmap_zero(header->adds_features, HEADER_FEAT_BITS);
2397                         set_bit(HEADER_BUILD_ID, header->adds_features);
2398                 }
2399         }
2400
2401         memcpy(&ph->adds_features, &header->adds_features,
2402                sizeof(ph->adds_features));
2403
2404         ph->data_offset  = header->data.offset;
2405         ph->data_size    = header->data.size;
2406         ph->feat_offset  = header->data.offset + header->data.size;
2407         return 0;
2408 }
2409
2410 static int perf_file_section__process(struct perf_file_section *section,
2411                                       struct perf_header *ph,
2412                                       int feat, int fd, void *data)
2413 {
2414         if (lseek(fd, section->offset, SEEK_SET) == (off_t)-1) {
2415                 pr_debug("Failed to lseek to %" PRIu64 " offset for feature "
2416                           "%d, continuing...\n", section->offset, feat);
2417                 return 0;
2418         }
2419
2420         if (feat >= HEADER_LAST_FEATURE) {
2421                 pr_debug("unknown feature %d, continuing...\n", feat);
2422                 return 0;
2423         }
2424
2425         if (!feat_ops[feat].process)
2426                 return 0;
2427
2428         return feat_ops[feat].process(section, ph, fd, data);
2429 }
2430
2431 static int perf_file_header__read_pipe(struct perf_pipe_file_header *header,
2432                                        struct perf_header *ph, int fd,
2433                                        bool repipe)
2434 {
2435         ssize_t ret;
2436
2437         ret = readn(fd, header, sizeof(*header));
2438         if (ret <= 0)
2439                 return -1;
2440
2441         if (check_magic_endian(header->magic, header->size, true, ph) < 0) {
2442                 pr_debug("endian/magic failed\n");
2443                 return -1;
2444         }
2445
2446         if (ph->needs_swap)
2447                 header->size = bswap_64(header->size);
2448
2449         if (repipe && do_write(STDOUT_FILENO, header, sizeof(*header)) < 0)
2450                 return -1;
2451
2452         return 0;
2453 }
2454
2455 static int perf_header__read_pipe(struct perf_session *session)
2456 {
2457         struct perf_header *header = &session->header;
2458         struct perf_pipe_file_header f_header;
2459
2460         if (perf_file_header__read_pipe(&f_header, header,
2461                                         perf_data_file__fd(session->file),
2462                                         session->repipe) < 0) {
2463                 pr_debug("incompatible file format\n");
2464                 return -EINVAL;
2465         }
2466
2467         return 0;
2468 }
2469
2470 static int read_attr(int fd, struct perf_header *ph,
2471                      struct perf_file_attr *f_attr)
2472 {
2473         struct perf_event_attr *attr = &f_attr->attr;
2474         size_t sz, left;
2475         size_t our_sz = sizeof(f_attr->attr);
2476         ssize_t ret;
2477
2478         memset(f_attr, 0, sizeof(*f_attr));
2479
2480         /* read minimal guaranteed structure */
2481         ret = readn(fd, attr, PERF_ATTR_SIZE_VER0);
2482         if (ret <= 0) {
2483                 pr_debug("cannot read %d bytes of header attr\n",
2484                          PERF_ATTR_SIZE_VER0);
2485                 return -1;
2486         }
2487
2488         /* on file perf_event_attr size */
2489         sz = attr->size;
2490
2491         if (ph->needs_swap)
2492                 sz = bswap_32(sz);
2493
2494         if (sz == 0) {
2495                 /* assume ABI0 */
2496                 sz =  PERF_ATTR_SIZE_VER0;
2497         } else if (sz > our_sz) {
2498                 pr_debug("file uses a more recent and unsupported ABI"
2499                          " (%zu bytes extra)\n", sz - our_sz);
2500                 return -1;
2501         }
2502         /* what we have not yet read and that we know about */
2503         left = sz - PERF_ATTR_SIZE_VER0;
2504         if (left) {
2505                 void *ptr = attr;
2506                 ptr += PERF_ATTR_SIZE_VER0;
2507
2508                 ret = readn(fd, ptr, left);
2509         }
2510         /* read perf_file_section, ids are read in caller */
2511         ret = readn(fd, &f_attr->ids, sizeof(f_attr->ids));
2512
2513         return ret <= 0 ? -1 : 0;
2514 }
2515
2516 static int perf_evsel__prepare_tracepoint_event(struct perf_evsel *evsel,
2517                                                 struct pevent *pevent)
2518 {
2519         struct event_format *event;
2520         char bf[128];
2521
2522         /* already prepared */
2523         if (evsel->tp_format)
2524                 return 0;
2525
2526         if (pevent == NULL) {
2527                 pr_debug("broken or missing trace data\n");
2528                 return -1;
2529         }
2530
2531         event = pevent_find_event(pevent, evsel->attr.config);
2532         if (event == NULL)
2533                 return -1;
2534
2535         if (!evsel->name) {
2536                 snprintf(bf, sizeof(bf), "%s:%s", event->system, event->name);
2537                 evsel->name = strdup(bf);
2538                 if (evsel->name == NULL)
2539                         return -1;
2540         }
2541
2542         evsel->tp_format = event;
2543         return 0;
2544 }
2545
2546 static int perf_evlist__prepare_tracepoint_events(struct perf_evlist *evlist,
2547                                                   struct pevent *pevent)
2548 {
2549         struct perf_evsel *pos;
2550
2551         evlist__for_each(evlist, pos) {
2552                 if (pos->attr.type == PERF_TYPE_TRACEPOINT &&
2553                     perf_evsel__prepare_tracepoint_event(pos, pevent))
2554                         return -1;
2555         }
2556
2557         return 0;
2558 }
2559
2560 int perf_session__read_header(struct perf_session *session)
2561 {
2562         struct perf_data_file *file = session->file;
2563         struct perf_header *header = &session->header;
2564         struct perf_file_header f_header;
2565         struct perf_file_attr   f_attr;
2566         u64                     f_id;
2567         int nr_attrs, nr_ids, i, j;
2568         int fd = perf_data_file__fd(file);
2569
2570         session->evlist = perf_evlist__new();
2571         if (session->evlist == NULL)
2572                 return -ENOMEM;
2573
2574         session->evlist->env = &header->env;
2575         session->machines.host.env = &header->env;
2576         if (perf_data_file__is_pipe(file))
2577                 return perf_header__read_pipe(session);
2578
2579         if (perf_file_header__read(&f_header, header, fd) < 0)
2580                 return -EINVAL;
2581
2582         /*
2583          * Sanity check that perf.data was written cleanly; data size is
2584          * initialized to 0 and updated only if the on_exit function is run.
2585          * If data size is still 0 then the file contains only partial
2586          * information.  Just warn user and process it as much as it can.
2587          */
2588         if (f_header.data.size == 0) {
2589                 pr_warning("WARNING: The %s file's data size field is 0 which is unexpected.\n"
2590                            "Was the 'perf record' command properly terminated?\n",
2591                            file->path);
2592         }
2593
2594         if (f_header.attr_size == 0) {
2595                 pr_err("ERROR: The %s file's attr size field is 0 which is unexpected.\n"
2596                        "Was the 'perf record' command properly terminated?\n",
2597                        file->path);
2598                 return -EINVAL;
2599         }
2600
2601         nr_attrs = f_header.attrs.size / f_header.attr_size;
2602         lseek(fd, f_header.attrs.offset, SEEK_SET);
2603
2604         for (i = 0; i < nr_attrs; i++) {
2605                 struct perf_evsel *evsel;
2606                 off_t tmp;
2607
2608                 if (read_attr(fd, header, &f_attr) < 0)
2609                         goto out_errno;
2610
2611                 if (header->needs_swap) {
2612                         f_attr.ids.size   = bswap_64(f_attr.ids.size);
2613                         f_attr.ids.offset = bswap_64(f_attr.ids.offset);
2614                         perf_event__attr_swap(&f_attr.attr);
2615                 }
2616
2617                 tmp = lseek(fd, 0, SEEK_CUR);
2618                 evsel = perf_evsel__new(&f_attr.attr);
2619
2620                 if (evsel == NULL)
2621                         goto out_delete_evlist;
2622
2623                 evsel->needs_swap = header->needs_swap;
2624                 /*
2625                  * Do it before so that if perf_evsel__alloc_id fails, this
2626                  * entry gets purged too at perf_evlist__delete().
2627                  */
2628                 perf_evlist__add(session->evlist, evsel);
2629
2630                 nr_ids = f_attr.ids.size / sizeof(u64);
2631                 /*
2632                  * We don't have the cpu and thread maps on the header, so
2633                  * for allocating the perf_sample_id table we fake 1 cpu and
2634                  * hattr->ids threads.
2635                  */
2636                 if (perf_evsel__alloc_id(evsel, 1, nr_ids))
2637                         goto out_delete_evlist;
2638
2639                 lseek(fd, f_attr.ids.offset, SEEK_SET);
2640
2641                 for (j = 0; j < nr_ids; j++) {
2642                         if (perf_header__getbuffer64(header, fd, &f_id, sizeof(f_id)))
2643                                 goto out_errno;
2644
2645                         perf_evlist__id_add(session->evlist, evsel, 0, j, f_id);
2646                 }
2647
2648                 lseek(fd, tmp, SEEK_SET);
2649         }
2650
2651         symbol_conf.nr_events = nr_attrs;
2652
2653         perf_header__process_sections(header, fd, &session->tevent,
2654                                       perf_file_section__process);
2655
2656         if (perf_evlist__prepare_tracepoint_events(session->evlist,
2657                                                    session->tevent.pevent))
2658                 goto out_delete_evlist;
2659
2660         return 0;
2661 out_errno:
2662         return -errno;
2663
2664 out_delete_evlist:
2665         perf_evlist__delete(session->evlist);
2666         session->evlist = NULL;
2667         return -ENOMEM;
2668 }
2669
2670 int perf_event__synthesize_attr(struct perf_tool *tool,
2671                                 struct perf_event_attr *attr, u32 ids, u64 *id,
2672                                 perf_event__handler_t process)
2673 {
2674         union perf_event *ev;
2675         size_t size;
2676         int err;
2677
2678         size = sizeof(struct perf_event_attr);
2679         size = PERF_ALIGN(size, sizeof(u64));
2680         size += sizeof(struct perf_event_header);
2681         size += ids * sizeof(u64);
2682
2683         ev = zalloc(size);
2684
2685         if (ev == NULL)
2686                 return -ENOMEM;
2687
2688         ev->attr.attr = *attr;
2689         memcpy(ev->attr.id, id, ids * sizeof(u64));
2690
2691         ev->attr.header.type = PERF_RECORD_HEADER_ATTR;
2692         ev->attr.header.size = (u16)size;
2693
2694         if (ev->attr.header.size == size)
2695                 err = process(tool, ev, NULL, NULL);
2696         else
2697                 err = -E2BIG;
2698
2699         free(ev);
2700
2701         return err;
2702 }
2703
2704 int perf_event__synthesize_attrs(struct perf_tool *tool,
2705                                    struct perf_session *session,
2706                                    perf_event__handler_t process)
2707 {
2708         struct perf_evsel *evsel;
2709         int err = 0;
2710
2711         evlist__for_each(session->evlist, evsel) {
2712                 err = perf_event__synthesize_attr(tool, &evsel->attr, evsel->ids,
2713                                                   evsel->id, process);
2714                 if (err) {
2715                         pr_debug("failed to create perf header attribute\n");
2716                         return err;
2717                 }
2718         }
2719
2720         return err;
2721 }
2722
2723 int perf_event__process_attr(struct perf_tool *tool __maybe_unused,
2724                              union perf_event *event,
2725                              struct perf_evlist **pevlist)
2726 {
2727         u32 i, ids, n_ids;
2728         struct perf_evsel *evsel;
2729         struct perf_evlist *evlist = *pevlist;
2730
2731         if (evlist == NULL) {
2732                 *pevlist = evlist = perf_evlist__new();
2733                 if (evlist == NULL)
2734                         return -ENOMEM;
2735         }
2736
2737         evsel = perf_evsel__new(&event->attr.attr);
2738         if (evsel == NULL)
2739                 return -ENOMEM;
2740
2741         perf_evlist__add(evlist, evsel);
2742
2743         ids = event->header.size;
2744         ids -= (void *)&event->attr.id - (void *)event;
2745         n_ids = ids / sizeof(u64);
2746         /*
2747          * We don't have the cpu and thread maps on the header, so
2748          * for allocating the perf_sample_id table we fake 1 cpu and
2749          * hattr->ids threads.
2750          */
2751         if (perf_evsel__alloc_id(evsel, 1, n_ids))
2752                 return -ENOMEM;
2753
2754         for (i = 0; i < n_ids; i++) {
2755                 perf_evlist__id_add(evlist, evsel, 0, i, event->attr.id[i]);
2756         }
2757
2758         symbol_conf.nr_events = evlist->nr_entries;
2759
2760         return 0;
2761 }
2762
2763 int perf_event__synthesize_tracing_data(struct perf_tool *tool, int fd,
2764                                         struct perf_evlist *evlist,
2765                                         perf_event__handler_t process)
2766 {
2767         union perf_event ev;
2768         struct tracing_data *tdata;
2769         ssize_t size = 0, aligned_size = 0, padding;
2770         int err __maybe_unused = 0;
2771
2772         /*
2773          * We are going to store the size of the data followed
2774          * by the data contents. Since the fd descriptor is a pipe,
2775          * we cannot seek back to store the size of the data once
2776          * we know it. Instead we:
2777          *
2778          * - write the tracing data to the temp file
2779          * - get/write the data size to pipe
2780          * - write the tracing data from the temp file
2781          *   to the pipe
2782          */
2783         tdata = tracing_data_get(&evlist->entries, fd, true);
2784         if (!tdata)
2785                 return -1;
2786
2787         memset(&ev, 0, sizeof(ev));
2788
2789         ev.tracing_data.header.type = PERF_RECORD_HEADER_TRACING_DATA;
2790         size = tdata->size;
2791         aligned_size = PERF_ALIGN(size, sizeof(u64));
2792         padding = aligned_size - size;
2793         ev.tracing_data.header.size = sizeof(ev.tracing_data);
2794         ev.tracing_data.size = aligned_size;
2795
2796         process(tool, &ev, NULL, NULL);
2797
2798         /*
2799          * The put function will copy all the tracing data
2800          * stored in temp file to the pipe.
2801          */
2802         tracing_data_put(tdata);
2803
2804         write_padded(fd, NULL, 0, padding);
2805
2806         return aligned_size;
2807 }
2808
2809 int perf_event__process_tracing_data(struct perf_tool *tool __maybe_unused,
2810                                      union perf_event *event,
2811                                      struct perf_session *session)
2812 {
2813         ssize_t size_read, padding, size = event->tracing_data.size;
2814         int fd = perf_data_file__fd(session->file);
2815         off_t offset = lseek(fd, 0, SEEK_CUR);
2816         char buf[BUFSIZ];
2817
2818         /* setup for reading amidst mmap */
2819         lseek(fd, offset + sizeof(struct tracing_data_event),
2820               SEEK_SET);
2821
2822         size_read = trace_report(fd, &session->tevent,
2823                                  session->repipe);
2824         padding = PERF_ALIGN(size_read, sizeof(u64)) - size_read;
2825
2826         if (readn(fd, buf, padding) < 0) {
2827                 pr_err("%s: reading input file", __func__);
2828                 return -1;
2829         }
2830         if (session->repipe) {
2831                 int retw = write(STDOUT_FILENO, buf, padding);
2832                 if (retw <= 0 || retw != padding) {
2833                         pr_err("%s: repiping tracing data padding", __func__);
2834                         return -1;
2835                 }
2836         }
2837
2838         if (size_read + padding != size) {
2839                 pr_err("%s: tracing data size mismatch", __func__);
2840                 return -1;
2841         }
2842
2843         perf_evlist__prepare_tracepoint_events(session->evlist,
2844                                                session->tevent.pevent);
2845
2846         return size_read + padding;
2847 }
2848
2849 int perf_event__synthesize_build_id(struct perf_tool *tool,
2850                                     struct dso *pos, u16 misc,
2851                                     perf_event__handler_t process,
2852                                     struct machine *machine)
2853 {
2854         union perf_event ev;
2855         size_t len;
2856         int err = 0;
2857
2858         if (!pos->hit)
2859                 return err;
2860
2861         memset(&ev, 0, sizeof(ev));
2862
2863         len = pos->long_name_len + 1;
2864         len = PERF_ALIGN(len, NAME_ALIGN);
2865         memcpy(&ev.build_id.build_id, pos->build_id, sizeof(pos->build_id));
2866         ev.build_id.header.type = PERF_RECORD_HEADER_BUILD_ID;
2867         ev.build_id.header.misc = misc;
2868         ev.build_id.pid = machine->pid;
2869         ev.build_id.header.size = sizeof(ev.build_id) + len;
2870         memcpy(&ev.build_id.filename, pos->long_name, pos->long_name_len);
2871
2872         err = process(tool, &ev, NULL, machine);
2873
2874         return err;
2875 }
2876
2877 int perf_event__process_build_id(struct perf_tool *tool __maybe_unused,
2878                                  union perf_event *event,
2879                                  struct perf_session *session)
2880 {
2881         __event_process_build_id(&event->build_id,
2882                                  event->build_id.filename,
2883                                  session);
2884         return 0;
2885 }