tools/perf/util/header.c

   1 #include "util.h"
   2 #include <sys/types.h>
   3 #include <byteswap.h>
   4 #include <unistd.h>
   5 #include <stdio.h>
   6 #include <stdlib.h>
   7 #include <linux/list.h>
   8 #include <linux/kernel.h>
   9 #include <linux/bitops.h>
  10 #include <sys/utsname.h>
  11
  12 #include "evlist.h"
  13 #include "evsel.h"
  14 #include "header.h"
  15 #include "../perf.h"
  16 #include "trace-event.h"
  17 #include "session.h"
  18 #include "symbol.h"
  19 #include "debug.h"
  20 #include "cpumap.h"
  21 #include "pmu.h"
  22 #include "vdso.h"
  23 #include "strbuf.h"
  24 #include "build-id.h"
  25 #include "data.h"
  26
  27 static u32 header_argc;
  28 static const char **header_argv;
  29
  30 /*
  31  * magic2 = "PERFILE2"
  32  * must be a numerical value to let the endianness
  33  * determine the memory layout. That way we are able
  34  * to detect endianness when reading the perf.data file
  35  * back.
  36  *
  37  * we check for legacy (PERFFILE) format.
  38  */
  39 static const char *__perf_magic1 = "PERFFILE";
  40 static const u64 __perf_magic2    = 0x32454c4946524550ULL;
  41 static const u64 __perf_magic2_sw = 0x50455246494c4532ULL;
  42
  43 #define PERF_MAGIC      __perf_magic2
  44
  45 struct perf_file_attr {
  46         struct perf_event_attr  attr;
  47         struct perf_file_section        ids;
  48 };
  49
  50 void perf_header__set_feat(struct perf_header *header, int feat)
  51 {
  52         set_bit(feat, header->adds_features);
  53 }
  54
  55 void perf_header__clear_feat(struct perf_header *header, int feat)
  56 {
  57         clear_bit(feat, header->adds_features);
  58 }
  59
  60 bool perf_header__has_feat(const struct perf_header *header, int feat)
  61 {
  62         return test_bit(feat, header->adds_features);
  63 }
  64
  65 static int do_write(int fd, const void *buf, size_t size)
  66 {
  67         while (size) {
  68                 int ret = write(fd, buf, size);
  69
  70                 if (ret < 0)
  71                         return -errno;
  72
  73                 size -= ret;
  74                 buf += ret;
  75         }
  76
  77         return 0;
  78 }
  79
  80 int write_padded(int fd, const void *bf, size_t count, size_t count_aligned)
  81 {
  82         static const char zero_buf[NAME_ALIGN];
  83         int err = do_write(fd, bf, count);
  84
  85         if (!err)
  86                 err = do_write(fd, zero_buf, count_aligned - count);
  87
  88         return err;
  89 }
  90
  91 static int do_write_string(int fd, const char *str)
  92 {
  93         u32 len, olen;
  94         int ret;
  95
  96         olen = strlen(str) + 1;
  97         len = PERF_ALIGN(olen, NAME_ALIGN);
  98
  99         /* write len, incl. \0 */
 100         ret = do_write(fd, &len, sizeof(len));
 101         if (ret < 0)
 102                 return ret;
 103
 104         return write_padded(fd, str, olen, len);
 105 }
 106
 107 static char *do_read_string(int fd, struct perf_header *ph)
 108 {
 109         ssize_t sz, ret;
 110         u32 len;
 111         char *buf;
 112
 113         sz = readn(fd, &len, sizeof(len));
 114         if (sz < (ssize_t)sizeof(len))
 115                 return NULL;
 116
 117         if (ph->needs_swap)
 118                 len = bswap_32(len);
 119
 120         buf = malloc(len);
 121         if (!buf)
 122                 return NULL;
 123
 124         ret = readn(fd, buf, len);
 125         if (ret == (ssize_t)len) {
 126                 /*
 127                  * strings are padded by zeroes
 128                  * thus the actual strlen of buf
 129                  * may be less than len
 130                  */
 131                 return buf;
 132         }
 133
 134         free(buf);
 135         return NULL;
 136 }
 137
 138 int
 139 perf_header__set_cmdline(int argc, const char **argv)
 140 {
 141         int i;
 142
 143         /*
 144          * If header_argv has already been set, do not override it.
 145          * This allows a command to set the cmdline, parse args and
 146          * then call another builtin function that implements a
 147          * command -- e.g, cmd_kvm calling cmd_record.
 148          */
 149         if (header_argv)
 150                 return 0;
 151
 152         header_argc = (u32)argc;
 153
 154         /* do not include NULL termination */
 155         header_argv = calloc(argc, sizeof(char *));
 156         if (!header_argv)
 157                 return -ENOMEM;
 158
 159         /*
 160          * must copy argv contents because it gets moved
 161          * around during option parsing
 162          */
 163         for (i = 0; i < argc ; i++)
 164                 header_argv[i] = argv[i];
 165
 166         return 0;
 167 }
 168
 169 static int write_tracing_data(int fd, struct perf_header *h __maybe_unused,
 170                             struct perf_evlist *evlist)
 171 {
 172         return read_tracing_data(fd, &evlist->entries);
 173 }
 174
 175
 176 static int write_build_id(int fd, struct perf_header *h,
 177                           struct perf_evlist *evlist __maybe_unused)
 178 {
 179         struct perf_session *session;
 180         int err;
 181
 182         session = container_of(h, struct perf_session, header);
 183
 184         if (!perf_session__read_build_ids(session, true))
 185                 return -1;
 186
 187         err = perf_session__write_buildid_table(session, fd);
 188         if (err < 0) {
 189                 pr_debug("failed to write buildid table\n");
 190                 return err;
 191         }
 192         perf_session__cache_build_ids(session);
 193
 194         return 0;
 195 }
 196
 197 static int write_hostname(int fd, struct perf_header *h __maybe_unused,
 198                           struct perf_evlist *evlist __maybe_unused)
 199 {
 200         struct utsname uts;
 201         int ret;
 202
 203         ret = uname(&uts);
 204         if (ret < 0)
 205                 return -1;
 206
 207         return do_write_string(fd, uts.nodename);
 208 }
 209
 210 static int write_osrelease(int fd, struct perf_header *h __maybe_unused,
 211                            struct perf_evlist *evlist __maybe_unused)
 212 {
 213         struct utsname uts;
 214         int ret;
 215
 216         ret = uname(&uts);
 217         if (ret < 0)
 218                 return -1;
 219
 220         return do_write_string(fd, uts.release);
 221 }
 222
 223 static int write_arch(int fd, struct perf_header *h __maybe_unused,
 224                       struct perf_evlist *evlist __maybe_unused)
 225 {
 226         struct utsname uts;
 227         int ret;
 228
 229         ret = uname(&uts);
 230         if (ret < 0)
 231                 return -1;
 232
 233         return do_write_string(fd, uts.machine);
 234 }
 235
 236 static int write_version(int fd, struct perf_header *h __maybe_unused,
 237                          struct perf_evlist *evlist __maybe_unused)
 238 {
 239         return do_write_string(fd, perf_version_string);
 240 }
 241
 242 static int __write_cpudesc(int fd, const char *cpuinfo_proc)
 243 {
 244         FILE *file;
 245         char *buf = NULL;
 246         char *s, *p;
 247         const char *search = cpuinfo_proc;
 248         size_t len = 0;
 249         int ret = -1;
 250
 251         if (!search)
 252                 return -1;
 253
 254         file = fopen("/proc/cpuinfo", "r");
 255         if (!file)
 256                 return -1;
 257
 258         while (getline(&buf, &len, file) > 0) {
 259                 ret = strncmp(buf, search, strlen(search));
 260                 if (!ret)
 261                         break;
 262         }
 263
 264         if (ret) {
 265                 ret = -1;
 266                 goto done;
 267         }
 268
 269         s = buf;
 270
 271         p = strchr(buf, ':');
 272         if (p && *(p+1) == ' ' && *(p+2))
 273                 s = p + 2;
 274         p = strchr(s, '\n');
 275         if (p)
 276                 *p = '\0';
 277
 278         /* squash extra space characters (branding string) */
 279         p = s;
 280         while (*p) {
 281                 if (isspace(*p)) {
 282                         char *r = p + 1;
 283                         char *q = r;
 284                         *p = ' ';
 285                         while (*q && isspace(*q))
 286                                 q++;
 287                         if (q != (p+1))
 288                                 while ((*r++ = *q++));
 289                 }
 290                 p++;
 291         }
 292         ret = do_write_string(fd, s);
 293 done:
 294         free(buf);
 295         fclose(file);
 296         return ret;
 297 }
 298
 299 static int write_cpudesc(int fd, struct perf_header *h __maybe_unused,
 300                        struct perf_evlist *evlist __maybe_unused)
 301 {
 302 #ifndef CPUINFO_PROC
 303 #define CPUINFO_PROC {"model name", }
 304 #endif
 305         const char *cpuinfo_procs[] = CPUINFO_PROC;
 306         unsigned int i;
 307
 308         for (i = 0; i < ARRAY_SIZE(cpuinfo_procs); i++) {
 309                 int ret;
 310                 ret = __write_cpudesc(fd, cpuinfo_procs[i]);
 311                 if (ret >= 0)
 312                         return ret;
 313         }
 314         return -1;
 315 }
 316
 317
 318 static int write_nrcpus(int fd, struct perf_header *h __maybe_unused,
 319                         struct perf_evlist *evlist __maybe_unused)
 320 {
 321         long nr;
 322         u32 nrc, nra;
 323         int ret;
 324
 325         nr = sysconf(_SC_NPROCESSORS_CONF);
 326         if (nr < 0)
 327                 return -1;
 328
 329         nrc = (u32)(nr & UINT_MAX);
 330
 331         nr = sysconf(_SC_NPROCESSORS_ONLN);
 332         if (nr < 0)
 333                 return -1;
 334
 335         nra = (u32)(nr & UINT_MAX);
 336
 337         ret = do_write(fd, &nrc, sizeof(nrc));
 338         if (ret < 0)
 339                 return ret;
 340
 341         return do_write(fd, &nra, sizeof(nra));
 342 }
 343
 344 static int write_event_desc(int fd, struct perf_header *h __maybe_unused,
 345                             struct perf_evlist *evlist)
 346 {
 347         struct perf_evsel *evsel;
 348         u32 nre, nri, sz;
 349         int ret;
 350
 351         nre = evlist->nr_entries;
 352
 353         /*
 354          * write number of events
 355          */
 356         ret = do_write(fd, &nre, sizeof(nre));
 357         if (ret < 0)
 358                 return ret;
 359
 360         /*
 361          * size of perf_event_attr struct
 362          */
 363         sz = (u32)sizeof(evsel->attr);
 364         ret = do_write(fd, &sz, sizeof(sz));
 365         if (ret < 0)
 366                 return ret;
 367
 368         evlist__for_each(evlist, evsel) {
 369                 ret = do_write(fd, &evsel->attr, sz);
 370                 if (ret < 0)
 371                         return ret;
 372                 /*
 373                  * write number of unique id per event
 374                  * there is one id per instance of an event
 375                  *
 376                  * copy into an nri to be independent of the
 377                  * type of ids,
 378                  */
 379                 nri = evsel->ids;
 380                 ret = do_write(fd, &nri, sizeof(nri));
 381                 if (ret < 0)
 382                         return ret;
 383
 384                 /*
 385                  * write event string as passed on cmdline
 386                  */
 387                 ret = do_write_string(fd, perf_evsel__name(evsel));
 388                 if (ret < 0)
 389                         return ret;
 390                 /*
 391                  * write unique ids for this event
 392                  */
 393                 ret = do_write(fd, evsel->id, evsel->ids * sizeof(u64));
 394                 if (ret < 0)
 395                         return ret;
 396         }
 397         return 0;
 398 }
 399
 400 static int write_cmdline(int fd, struct perf_header *h __maybe_unused,
 401                          struct perf_evlist *evlist __maybe_unused)
 402 {
 403         char buf[MAXPATHLEN];
 404         char proc[32];
 405         u32 i, n;
 406         int ret;
 407
 408         /*
 409          * actual atual path to perf binary
 410          */
 411         sprintf(proc, "/proc/%d/exe", getpid());
 412         ret = readlink(proc, buf, sizeof(buf));
 413         if (ret <= 0)
 414                 return -1;
 415
 416         /* readlink() does not add null termination */
 417         buf[ret] = '\0';
 418
 419         /* account for binary path */
 420         n = header_argc + 1;
 421
 422         ret = do_write(fd, &n, sizeof(n));
 423         if (ret < 0)
 424                 return ret;
 425
 426         ret = do_write_string(fd, buf);
 427         if (ret < 0)
 428                 return ret;
 429
 430         for (i = 0 ; i < header_argc; i++) {
 431                 ret = do_write_string(fd, header_argv[i]);
 432                 if (ret < 0)
 433                         return ret;
 434         }
 435         return 0;
 436 }
 437
 438 #define CORE_SIB_FMT \
 439         "/sys/devices/system/cpu/cpu%d/topology/core_siblings_list"
 440 #define THRD_SIB_FMT \
 441         "/sys/devices/system/cpu/cpu%d/topology/thread_siblings_list"
 442
 443 struct cpu_topo {
 444         u32 core_sib;
 445         u32 thread_sib;
 446         char **core_siblings;
 447         char **thread_siblings;
 448 };
 449
 450 static int build_cpu_topo(struct cpu_topo *tp, int cpu)
 451 {
 452         FILE *fp;
 453         char filename[MAXPATHLEN];
 454         char *buf = NULL, *p;
 455         size_t len = 0;
 456         ssize_t sret;
 457         u32 i = 0;
 458         int ret = -1;
 459
 460         sprintf(filename, CORE_SIB_FMT, cpu);
 461         fp = fopen(filename, "r");
 462         if (!fp)
 463                 goto try_threads;
 464
 465         sret = getline(&buf, &len, fp);
 466         fclose(fp);
 467         if (sret <= 0)
 468                 goto try_threads;
 469
 470         p = strchr(buf, '\n');
 471         if (p)
 472                 *p = '\0';
 473
 474         for (i = 0; i < tp->core_sib; i++) {
 475                 if (!strcmp(buf, tp->core_siblings[i]))
 476                         break;
 477         }
 478         if (i == tp->core_sib) {
 479                 tp->core_siblings[i] = buf;
 480                 tp->core_sib++;
 481                 buf = NULL;
 482                 len = 0;
 483         }
 484         ret = 0;
 485
 486 try_threads:
 487         sprintf(filename, THRD_SIB_FMT, cpu);
 488         fp = fopen(filename, "r");
 489         if (!fp)
 490                 goto done;
 491
 492         if (getline(&buf, &len, fp) <= 0)
 493                 goto done;
 494
 495         p = strchr(buf, '\n');
 496         if (p)
 497                 *p = '\0';
 498
 499         for (i = 0; i < tp->thread_sib; i++) {
 500                 if (!strcmp(buf, tp->thread_siblings[i]))
 501                         break;
 502         }
 503         if (i == tp->thread_sib) {
 504                 tp->thread_siblings[i] = buf;
 505                 tp->thread_sib++;
 506                 buf = NULL;
 507         }
 508         ret = 0;
 509 done:
 510         if(fp)
 511                 fclose(fp);
 512         free(buf);
 513         return ret;
 514 }
 515
 516 static void free_cpu_topo(struct cpu_topo *tp)
 517 {
 518         u32 i;
 519
 520         if (!tp)
 521                 return;
 522
 523         for (i = 0 ; i < tp->core_sib; i++)
 524                 zfree(&tp->core_siblings[i]);
 525
 526         for (i = 0 ; i < tp->thread_sib; i++)
 527                 zfree(&tp->thread_siblings[i]);
 528
 529         free(tp);
 530 }
 531
 532 static struct cpu_topo *build_cpu_topology(void)
 533 {
 534         struct cpu_topo *tp;
 535         void *addr;
 536         u32 nr, i;
 537         size_t sz;
 538         long ncpus;
 539         int ret = -1;
 540
 541         ncpus = sysconf(_SC_NPROCESSORS_CONF);
 542         if (ncpus < 0)
 543                 return NULL;
 544
 545         nr = (u32)(ncpus & UINT_MAX);
 546
 547         sz = nr * sizeof(char *);
 548
 549         addr = calloc(1, sizeof(*tp) + 2 * sz);
 550         if (!addr)
 551                 return NULL;
 552
 553         tp = addr;
 554
 555         addr += sizeof(*tp);
 556         tp->core_siblings = addr;
 557         addr += sz;
 558         tp->thread_siblings = addr;
 559
 560         for (i = 0; i < nr; i++) {
 561                 ret = build_cpu_topo(tp, i);
 562                 if (ret < 0)
 563                         break;
 564         }
 565         if (ret) {
 566                 free_cpu_topo(tp);
 567                 tp = NULL;
 568         }
 569         return tp;
 570 }
 571
 572 static int write_cpu_topology(int fd, struct perf_header *h __maybe_unused,
 573                           struct perf_evlist *evlist __maybe_unused)
 574 {
 575         struct cpu_topo *tp;
 576         u32 i;
 577         int ret;
 578
 579         tp = build_cpu_topology();
 580         if (!tp)
 581                 return -1;
 582
 583         ret = do_write(fd, &tp->core_sib, sizeof(tp->core_sib));
 584         if (ret < 0)
 585                 goto done;
 586
 587         for (i = 0; i < tp->core_sib; i++) {
 588                 ret = do_write_string(fd, tp->core_siblings[i]);
 589                 if (ret < 0)
 590                         goto done;
 591         }
 592         ret = do_write(fd, &tp->thread_sib, sizeof(tp->thread_sib));
 593         if (ret < 0)
 594                 goto done;
 595
 596         for (i = 0; i < tp->thread_sib; i++) {
 597                 ret = do_write_string(fd, tp->thread_siblings[i]);
 598                 if (ret < 0)
 599                         break;
 600         }
 601 done:
 602         free_cpu_topo(tp);
 603         return ret;
 604 }
 605
 606
 607
 608 static int write_total_mem(int fd, struct perf_header *h __maybe_unused,
 609                           struct perf_evlist *evlist __maybe_unused)
 610 {
 611         char *buf = NULL;
 612         FILE *fp;
 613         size_t len = 0;
 614         int ret = -1, n;
 615         uint64_t mem;
 616
 617         fp = fopen("/proc/meminfo", "r");
 618         if (!fp)
 619                 return -1;
 620
 621         while (getline(&buf, &len, fp) > 0) {
 622                 ret = strncmp(buf, "MemTotal:", 9);
 623                 if (!ret)
 624                         break;
 625         }
 626         if (!ret) {
 627                 n = sscanf(buf, "%*s %"PRIu64, &mem);
 628                 if (n == 1)
 629                         ret = do_write(fd, &mem, sizeof(mem));
 630         } else
 631                 ret = -1;
 632         free(buf);
 633         fclose(fp);
 634         return ret;
 635 }
 636
 637 static int write_topo_node(int fd, int node)
 638 {
 639         char str[MAXPATHLEN];
 640         char field[32];
 641         char *buf = NULL, *p;
 642         size_t len = 0;
 643         FILE *fp;
 644         u64 mem_total, mem_free, mem;
 645         int ret = -1;
 646
 647         sprintf(str, "/sys/devices/system/node/node%d/meminfo", node);
 648         fp = fopen(str, "r");
 649         if (!fp)
 650                 return -1;
 651
 652         while (getline(&buf, &len, fp) > 0) {
 653                 /* skip over invalid lines */
 654                 if (!strchr(buf, ':'))
 655                         continue;
 656                 if (sscanf(buf, "%*s %*d %31s %"PRIu64, field, &mem) != 2)
 657                         goto done;
 658                 if (!strcmp(field, "MemTotal:"))
 659                         mem_total = mem;
 660                 if (!strcmp(field, "MemFree:"))
 661                         mem_free = mem;
 662         }
 663
 664         fclose(fp);
 665         fp = NULL;
 666
 667         ret = do_write(fd, &mem_total, sizeof(u64));
 668         if (ret)
 669                 goto done;
 670
 671         ret = do_write(fd, &mem_free, sizeof(u64));
 672         if (ret)
 673                 goto done;
 674
 675         ret = -1;
 676         sprintf(str, "/sys/devices/system/node/node%d/cpulist", node);
 677
 678         fp = fopen(str, "r");
 679         if (!fp)
 680                 goto done;
 681
 682         if (getline(&buf, &len, fp) <= 0)
 683                 goto done;
 684
 685         p = strchr(buf, '\n');
 686         if (p)
 687                 *p = '\0';
 688
 689         ret = do_write_string(fd, buf);
 690 done:
 691         free(buf);
 692         if (fp)
 693                 fclose(fp);
 694         return ret;
 695 }
 696
 697 static int write_numa_topology(int fd, struct perf_header *h __maybe_unused,
 698                           struct perf_evlist *evlist __maybe_unused)
 699 {
 700         char *buf = NULL;
 701         size_t len = 0;
 702         FILE *fp;
 703         struct cpu_map *node_map = NULL;
 704         char *c;
 705         u32 nr, i, j;
 706         int ret = -1;
 707
 708         fp = fopen("/sys/devices/system/node/online", "r");
 709         if (!fp)
 710                 return -1;
 711
 712         if (getline(&buf, &len, fp) <= 0)
 713                 goto done;
 714
 715         c = strchr(buf, '\n');
 716         if (c)
 717                 *c = '\0';
 718
 719         node_map = cpu_map__new(buf);
 720         if (!node_map)
 721                 goto done;
 722
 723         nr = (u32)node_map->nr;
 724
 725         ret = do_write(fd, &nr, sizeof(nr));
 726         if (ret < 0)
 727                 goto done;
 728
 729         for (i = 0; i < nr; i++) {
 730                 j = (u32)node_map->map[i];
 731                 ret = do_write(fd, &j, sizeof(j));
 732                 if (ret < 0)
 733                         break;
 734
 735                 ret = write_topo_node(fd, i);
 736                 if (ret < 0)
 737                         break;
 738         }
 739 done:
 740         free(buf);
 741         fclose(fp);
 742         free(node_map);
 743         return ret;
 744 }
 745
 746 /*
 747  * File format:
 748  *
 749  * struct pmu_mappings {
 750  *      u32     pmu_num;
 751  *      struct pmu_map {
 752  *              u32     type;
 753  *              char    name[];
 754  *      }[pmu_num];
 755  * };
 756  */
 757
 758 static int write_pmu_mappings(int fd, struct perf_header *h __maybe_unused,
 759                               struct perf_evlist *evlist __maybe_unused)
 760 {
 761         struct perf_pmu *pmu = NULL;
 762         off_t offset = lseek(fd, 0, SEEK_CUR);
 763         __u32 pmu_num = 0;
 764         int ret;
 765
 766         /* write real pmu_num later */
 767         ret = do_write(fd, &pmu_num, sizeof(pmu_num));
 768         if (ret < 0)
 769                 return ret;
 770
 771         while ((pmu = perf_pmu__scan(pmu))) {
 772                 if (!pmu->name)
 773                         continue;
 774                 pmu_num++;
 775
 776                 ret = do_write(fd, &pmu->type, sizeof(pmu->type));
 777                 if (ret < 0)
 778                         return ret;
 779
 780                 ret = do_write_string(fd, pmu->name);
 781                 if (ret < 0)
 782                         return ret;
 783         }
 784
 785         if (pwrite(fd, &pmu_num, sizeof(pmu_num), offset) != sizeof(pmu_num)) {
 786                 /* discard all */
 787                 lseek(fd, offset, SEEK_SET);
 788                 return -1;
 789         }
 790
 791         return 0;
 792 }
 793
 794 /*
 795  * File format:
 796  *
 797  * struct group_descs {
 798  *      u32     nr_groups;
 799  *      struct group_desc {
 800  *              char    name[];
 801  *              u32     leader_idx;
 802  *              u32     nr_members;
 803  *      }[nr_groups];
 804  * };
 805  */
 806 static int write_group_desc(int fd, struct perf_header *h __maybe_unused,
 807                             struct perf_evlist *evlist)
 808 {
 809         u32 nr_groups = evlist->nr_groups;
 810         struct perf_evsel *evsel;
 811         int ret;
 812
 813         ret = do_write(fd, &nr_groups, sizeof(nr_groups));
 814         if (ret < 0)
 815                 return ret;
 816
 817         evlist__for_each(evlist, evsel) {
 818                 if (perf_evsel__is_group_leader(evsel) &&
 819                     evsel->nr_members > 1) {
 820                         const char *name = evsel->group_name ?: "{anon_group}";
 821                         u32 leader_idx = evsel->idx;
 822                         u32 nr_members = evsel->nr_members;
 823
 824                         ret = do_write_string(fd, name);
 825                         if (ret < 0)
 826                                 return ret;
 827
 828                         ret = do_write(fd, &leader_idx, sizeof(leader_idx));
 829                         if (ret < 0)
 830                                 return ret;
 831
 832                         ret = do_write(fd, &nr_members, sizeof(nr_members));
 833                         if (ret < 0)
 834                                 return ret;
 835                 }
 836         }
 837         return 0;
 838 }
 839
 840 /*
 841  * default get_cpuid(): nothing gets recorded
 842  * actual implementation must be in arch/$(ARCH)/util/header.c
 843  */
 844 int __attribute__ ((weak)) get_cpuid(char *buffer __maybe_unused,
 845                                      size_t sz __maybe_unused)
 846 {
 847         return -1;
 848 }
 849
 850 static int write_cpuid(int fd, struct perf_header *h __maybe_unused,
 851                        struct perf_evlist *evlist __maybe_unused)
 852 {
 853         char buffer[64];
 854         int ret;
 855
 856         ret = get_cpuid(buffer, sizeof(buffer));
 857         if (!ret)
 858                 goto write_it;
 859
 860         return -1;
 861 write_it:
 862         return do_write_string(fd, buffer);
 863 }
 864
 865 static int write_branch_stack(int fd __maybe_unused,
 866                               struct perf_header *h __maybe_unused,
 867                        struct perf_evlist *evlist __maybe_unused)
 868 {
 869         return 0;
 870 }
 871
 872 static int write_auxtrace(int fd, struct perf_header *h,
 873                           struct perf_evlist *evlist __maybe_unused)
 874 {
 875         struct perf_session *session;
 876         int err;
 877
 878         session = container_of(h, struct perf_session, header);
 879
 880         err = auxtrace_index__write(fd, &session->auxtrace_index);
 881         if (err < 0)
 882                 pr_err("Failed to write auxtrace index\n");
 883         return err;
 884 }
 885
 886 static void print_hostname(struct perf_header *ph, int fd __maybe_unused,
 887                            FILE *fp)
 888 {
 889         fprintf(fp, "# hostname : %s\n", ph->env.hostname);
 890 }
 891
 892 static void print_osrelease(struct perf_header *ph, int fd __maybe_unused,
 893                             FILE *fp)
 894 {
 895         fprintf(fp, "# os release : %s\n", ph->env.os_release);
 896 }
 897
 898 static void print_arch(struct perf_header *ph, int fd __maybe_unused, FILE *fp)
 899 {
 900         fprintf(fp, "# arch : %s\n", ph->env.arch);
 901 }
 902
 903 static void print_cpudesc(struct perf_header *ph, int fd __maybe_unused,
 904                           FILE *fp)
 905 {
 906         fprintf(fp, "# cpudesc : %s\n", ph->env.cpu_desc);
 907 }
 908
 909 static void print_nrcpus(struct perf_header *ph, int fd __maybe_unused,
 910                          FILE *fp)
 911 {
 912         fprintf(fp, "# nrcpus online : %u\n", ph->env.nr_cpus_online);
 913         fprintf(fp, "# nrcpus avail : %u\n", ph->env.nr_cpus_avail);
 914 }
 915
 916 static void print_version(struct perf_header *ph, int fd __maybe_unused,
 917                           FILE *fp)
 918 {
 919         fprintf(fp, "# perf version : %s\n", ph->env.version);
 920 }
 921
 922 static void print_cmdline(struct perf_header *ph, int fd __maybe_unused,
 923                           FILE *fp)
 924 {
 925         int nr, i;
 926         char *str;
 927
 928         nr = ph->env.nr_cmdline;
 929         str = ph->env.cmdline;
 930
 931         fprintf(fp, "# cmdline : ");
 932
 933         for (i = 0; i < nr; i++) {
 934                 fprintf(fp, "%s ", str);
 935                 str += strlen(str) + 1;
 936         }
 937         fputc('\n', fp);
 938 }
 939
 940 static void print_cpu_topology(struct perf_header *ph, int fd __maybe_unused,
 941                                FILE *fp)
 942 {
 943         int nr, i;
 944         char *str;
 945
 946         nr = ph->env.nr_sibling_cores;
 947         str = ph->env.sibling_cores;
 948
 949         for (i = 0; i < nr; i++) {
 950                 fprintf(fp, "# sibling cores   : %s\n", str);
 951                 str += strlen(str) + 1;
 952         }
 953
 954         nr = ph->env.nr_sibling_threads;
 955         str = ph->env.sibling_threads;
 956
 957         for (i = 0; i < nr; i++) {
 958                 fprintf(fp, "# sibling threads : %s\n", str);
 959                 str += strlen(str) + 1;
 960         }
 961 }
 962
 963 static void free_event_desc(struct perf_evsel *events)
 964 {
 965         struct perf_evsel *evsel;
 966
 967         if (!events)
 968                 return;
 969
 970         for (evsel = events; evsel->attr.size; evsel++) {
 971                 zfree(&evsel->name);
 972                 zfree(&evsel->id);
 973         }
 974
 975         free(events);
 976 }
 977
 978 static struct perf_evsel *
 979 read_event_desc(struct perf_header *ph, int fd)
 980 {
 981         struct perf_evsel *evsel, *events = NULL;
 982         u64 *id;
 983         void *buf = NULL;
 984         u32 nre, sz, nr, i, j;
 985         ssize_t ret;
 986         size_t msz;
 987
 988         /* number of events */
 989         ret = readn(fd, &nre, sizeof(nre));
 990         if (ret != (ssize_t)sizeof(nre))
 991                 goto error;
 992
 993         if (ph->needs_swap)
 994                 nre = bswap_32(nre);
 995
 996         ret = readn(fd, &sz, sizeof(sz));
 997         if (ret != (ssize_t)sizeof(sz))
 998                 goto error;
 999
1000         if (ph->needs_swap)
1001                 sz = bswap_32(sz);
1002
1003         /* buffer to hold on file attr struct */
1004         buf = malloc(sz);
1005         if (!buf)
1006                 goto error;
1007
1008         /* the last event terminates with evsel->attr.size == 0: */
1009         events = calloc(nre + 1, sizeof(*events));
1010         if (!events)
1011                 goto error;
1012
1013         msz = sizeof(evsel->attr);
1014         if (sz < msz)
1015                 msz = sz;
1016
1017         for (i = 0, evsel = events; i < nre; evsel++, i++) {
1018                 evsel->idx = i;
1019
1020                 /*
1021                  * must read entire on-file attr struct to
1022                  * sync up with layout.
1023                  */
1024                 ret = readn(fd, buf, sz);
1025                 if (ret != (ssize_t)sz)
1026                         goto error;
1027
1028                 if (ph->needs_swap)
1029                         perf_event__attr_swap(buf);
1030
1031                 memcpy(&evsel->attr, buf, msz);
1032
1033                 ret = readn(fd, &nr, sizeof(nr));
1034                 if (ret != (ssize_t)sizeof(nr))
1035                         goto error;
1036
1037                 if (ph->needs_swap) {
1038                         nr = bswap_32(nr);
1039                         evsel->needs_swap = true;
1040                 }
1041
1042                 evsel->name = do_read_string(fd, ph);
1043
1044                 if (!nr)
1045                         continue;
1046
1047                 id = calloc(nr, sizeof(*id));
1048                 if (!id)
1049                         goto error;
1050                 evsel->ids = nr;
1051                 evsel->id = id;
1052
1053                 for (j = 0 ; j < nr; j++) {
1054                         ret = readn(fd, id, sizeof(*id));
1055                         if (ret != (ssize_t)sizeof(*id))
1056                                 goto error;
1057                         if (ph->needs_swap)
1058                                 *id = bswap_64(*id);
1059                         id++;
1060                 }
1061         }
1062 out:
1063         free(buf);
1064         return events;
1065 error:
1066         free_event_desc(events);
1067         events = NULL;
1068         goto out;
1069 }
1070
1071 static int __desc_attr__fprintf(FILE *fp, const char *name, const char *val,
1072                                 void *priv __attribute__((unused)))
1073 {
1074         return fprintf(fp, ", %s = %s", name, val);
1075 }
1076
1077 static void print_event_desc(struct perf_header *ph, int fd, FILE *fp)
1078 {
1079         struct perf_evsel *evsel, *events = read_event_desc(ph, fd);
1080         u32 j;
1081         u64 *id;
1082
1083         if (!events) {
1084                 fprintf(fp, "# event desc: not available or unable to read\n");
1085                 return;
1086         }
1087
1088         for (evsel = events; evsel->attr.size; evsel++) {
1089                 fprintf(fp, "# event : name = %s, ", evsel->name);
1090
1091                 if (evsel->ids) {
1092                         fprintf(fp, ", id = {");
1093                         for (j = 0, id = evsel->id; j < evsel->ids; j++, id++) {
1094                                 if (j)
1095                                         fputc(',', fp);
1096                                 fprintf(fp, " %"PRIu64, *id);
1097                         }
1098                         fprintf(fp, " }");
1099                 }
1100
1101                 perf_event_attr__fprintf(fp, &evsel->attr, __desc_attr__fprintf, NULL);
1102
1103                 fputc('\n', fp);
1104         }
1105
1106         free_event_desc(events);
1107 }
1108
1109 static void print_total_mem(struct perf_header *ph, int fd __maybe_unused,
1110                             FILE *fp)
1111 {
1112         fprintf(fp, "# total memory : %Lu kB\n", ph->env.total_mem);
1113 }
1114
1115 static void print_numa_topology(struct perf_header *ph, int fd __maybe_unused,
1116                                 FILE *fp)
1117 {
1118         u32 nr, c, i;
1119         char *str, *tmp;
1120         uint64_t mem_total, mem_free;
1121
1122         /* nr nodes */
1123         nr = ph->env.nr_numa_nodes;
1124         str = ph->env.numa_nodes;
1125
1126         for (i = 0; i < nr; i++) {
1127                 /* node number */
1128                 c = strtoul(str, &tmp, 0);
1129                 if (*tmp != ':')
1130                         goto error;
1131
1132                 str = tmp + 1;
1133                 mem_total = strtoull(str, &tmp, 0);
1134                 if (*tmp != ':')
1135                         goto error;
1136
1137                 str = tmp + 1;
1138                 mem_free = strtoull(str, &tmp, 0);
1139                 if (*tmp != ':')
1140                         goto error;
1141
1142                 fprintf(fp, "# node%u meminfo  : total = %"PRIu64" kB,"
1143                             " free = %"PRIu64" kB\n",
1144                         c, mem_total, mem_free);
1145
1146                 str = tmp + 1;
1147                 fprintf(fp, "# node%u cpu list : %s\n", c, str);
1148
1149                 str += strlen(str) + 1;
1150         }
1151         return;
1152 error:
1153         fprintf(fp, "# numa topology : not available\n");
1154 }
1155
1156 static void print_cpuid(struct perf_header *ph, int fd __maybe_unused, FILE *fp)
1157 {
1158         fprintf(fp, "# cpuid : %s\n", ph->env.cpuid);
1159 }
1160
1161 static void print_branch_stack(struct perf_header *ph __maybe_unused,
1162                                int fd __maybe_unused, FILE *fp)
1163 {
1164         fprintf(fp, "# contains samples with branch stack\n");
1165 }
1166
1167 static void print_auxtrace(struct perf_header *ph __maybe_unused,
1168                            int fd __maybe_unused, FILE *fp)
1169 {
1170         fprintf(fp, "# contains AUX area data (e.g. instruction trace)\n");
1171 }
1172
1173 static void print_pmu_mappings(struct perf_header *ph, int fd __maybe_unused,
1174                                FILE *fp)
1175 {
1176         const char *delimiter = "# pmu mappings: ";
1177         char *str, *tmp;
1178         u32 pmu_num;
1179         u32 type;
1180
1181         pmu_num = ph->env.nr_pmu_mappings;
1182         if (!pmu_num) {
1183                 fprintf(fp, "# pmu mappings: not available\n");
1184                 return;
1185         }
1186
1187         str = ph->env.pmu_mappings;
1188
1189         while (pmu_num) {
1190                 type = strtoul(str, &tmp, 0);
1191                 if (*tmp != ':')
1192                         goto error;
1193
1194                 str = tmp + 1;
1195                 fprintf(fp, "%s%s = %" PRIu32, delimiter, str, type);
1196
1197                 delimiter = ", ";
1198                 str += strlen(str) + 1;
1199                 pmu_num--;
1200         }
1201
1202         fprintf(fp, "\n");
1203
1204         if (!pmu_num)
1205                 return;
1206 error:
1207         fprintf(fp, "# pmu mappings: unable to read\n");
1208 }
1209
1210 static void print_group_desc(struct perf_header *ph, int fd __maybe_unused,
1211                              FILE *fp)
1212 {
1213         struct perf_session *session;
1214         struct perf_evsel *evsel;
1215         u32 nr = 0;
1216
1217         session = container_of(ph, struct perf_session, header);
1218
1219         evlist__for_each(session->evlist, evsel) {
1220                 if (perf_evsel__is_group_leader(evsel) &&
1221                     evsel->nr_members > 1) {
1222                         fprintf(fp, "# group: %s{%s", evsel->group_name ?: "",
1223                                 perf_evsel__name(evsel));
1224
1225                         nr = evsel->nr_members - 1;
1226                 } else if (nr) {
1227                         fprintf(fp, ",%s", perf_evsel__name(evsel));
1228
1229                         if (--nr == 0)
1230                                 fprintf(fp, "}\n");
1231                 }
1232         }
1233 }
1234
1235 static int __event_process_build_id(struct build_id_event *bev,
1236                                     char *filename,
1237                                     struct perf_session *session)
1238 {
1239         int err = -1;
1240         struct machine *machine;
1241         u16 cpumode;
1242         struct dso *dso;
1243         enum dso_kernel_type dso_type;
1244
1245         machine = perf_session__findnew_machine(session, bev->pid);
1246         if (!machine)
1247                 goto out;
1248
1249         cpumode = bev->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
1250
1251         switch (cpumode) {
1252         case PERF_RECORD_MISC_KERNEL:
1253                 dso_type = DSO_TYPE_KERNEL;
1254                 break;
1255         case PERF_RECORD_MISC_GUEST_KERNEL:
1256                 dso_type = DSO_TYPE_GUEST_KERNEL;
1257                 break;
1258         case PERF_RECORD_MISC_USER:
1259         case PERF_RECORD_MISC_GUEST_USER:
1260                 dso_type = DSO_TYPE_USER;
1261                 break;
1262         default:
1263                 goto out;
1264         }
1265
1266         dso = machine__findnew_dso(machine, filename);
1267         if (dso != NULL) {
1268                 char sbuild_id[BUILD_ID_SIZE * 2 + 1];
1269
1270                 dso__set_build_id(dso, &bev->build_id);
1271
1272                 if (!is_kernel_module(filename, cpumode))
1273                         dso->kernel = dso_type;
1274
1275                 build_id__sprintf(dso->build_id, sizeof(dso->build_id),
1276                                   sbuild_id);
1277                 pr_debug("build id event received for %s: %s\n",
1278                          dso->long_name, sbuild_id);
1279                 dso__put(dso);
1280         }
1281
1282         err = 0;
1283 out:
1284         return err;
1285 }
1286
1287 static int perf_header__read_build_ids_abi_quirk(struct perf_header *header,
1288                                                  int input, u64 offset, u64 size)
1289 {
1290         struct perf_session *session = container_of(header, struct perf_session, header);
1291         struct {
1292                 struct perf_event_header   header;
1293                 u8                         build_id[PERF_ALIGN(BUILD_ID_SIZE, sizeof(u64))];
1294                 char                       filename[0];
1295         } old_bev;
1296         struct build_id_event bev;
1297         char filename[PATH_MAX];
1298         u64 limit = offset + size;
1299
1300         while (offset < limit) {
1301                 ssize_t len;
1302
1303                 if (readn(input, &old_bev, sizeof(old_bev)) != sizeof(old_bev))
1304                         return -1;
1305
1306                 if (header->needs_swap)
1307                         perf_event_header__bswap(&old_bev.header);
1308
1309                 len = old_bev.header.size - sizeof(old_bev);
1310                 if (readn(input, filename, len) != len)
1311                         return -1;
1312
1313                 bev.header = old_bev.header;
1314
1315                 /*
1316                  * As the pid is the missing value, we need to fill
1317                  * it properly. The header.misc value give us nice hint.
1318                  */
1319                 bev.pid = HOST_KERNEL_ID;
1320                 if (bev.header.misc == PERF_RECORD_MISC_GUEST_USER ||
1321                     bev.header.misc == PERF_RECORD_MISC_GUEST_KERNEL)
1322                         bev.pid = DEFAULT_GUEST_KERNEL_ID;
1323
1324                 memcpy(bev.build_id, old_bev.build_id, sizeof(bev.build_id));
1325                 __event_process_build_id(&bev, filename, session);
1326
1327                 offset += bev.header.size;
1328         }
1329
1330         return 0;
1331 }
1332
1333 static int perf_header__read_build_ids(struct perf_header *header,
1334                                        int input, u64 offset, u64 size)
1335 {
1336         struct perf_session *session = container_of(header, struct perf_session, header);
1337         struct build_id_event bev;
1338         char filename[PATH_MAX];
1339         u64 limit = offset + size, orig_offset = offset;
1340         int err = -1;
1341
1342         while (offset < limit) {
1343                 ssize_t len;
1344
1345                 if (readn(input, &bev, sizeof(bev)) != sizeof(bev))
1346                         goto out;
1347
1348                 if (header->needs_swap)
1349                         perf_event_header__bswap(&bev.header);
1350
1351                 len = bev.header.size - sizeof(bev);
1352                 if (readn(input, filename, len) != len)
1353                         goto out;
1354                 /*
1355                  * The a1645ce1 changeset:
1356                  *
1357                  * "perf: 'perf kvm' tool for monitoring guest performance from host"
1358                  *
1359                  * Added a field to struct build_id_event that broke the file
1360                  * format.
1361                  *
1362                  * Since the kernel build-id is the first entry, process the
1363                  * table using the old format if the well known
1364                  * '[kernel.kallsyms]' string for the kernel build-id has the
1365                  * first 4 characters chopped off (where the pid_t sits).
1366                  */
1367                 if (memcmp(filename, "nel.kallsyms]", 13) == 0) {
1368                         if (lseek(input, orig_offset, SEEK_SET) == (off_t)-1)
1369                                 return -1;
1370                         return perf_header__read_build_ids_abi_quirk(header, input, offset, size);
1371                 }
1372
1373                 __event_process_build_id(&bev, filename, session);
1374
1375                 offset += bev.header.size;
1376         }
1377         err = 0;
1378 out:
1379         return err;
1380 }
1381
1382 static int process_tracing_data(struct perf_file_section *section __maybe_unused,
1383                                 struct perf_header *ph __maybe_unused,
1384                                 int fd, void *data)
1385 {
1386         ssize_t ret = trace_report(fd, data, false);
1387         return ret < 0 ? -1 : 0;
1388 }
1389
1390 static int process_build_id(struct perf_file_section *section,
1391                             struct perf_header *ph, int fd,
1392                             void *data __maybe_unused)
1393 {
1394         if (perf_header__read_build_ids(ph, fd, section->offset, section->size))
1395                 pr_debug("Failed to read buildids, continuing...\n");
1396         return 0;
1397 }
1398
1399 static int process_hostname(struct perf_file_section *section __maybe_unused,
1400                             struct perf_header *ph, int fd,
1401                             void *data __maybe_unused)
1402 {
1403         ph->env.hostname = do_read_string(fd, ph);
1404         return ph->env.hostname ? 0 : -ENOMEM;
1405 }
1406
1407 static int process_osrelease(struct perf_file_section *section __maybe_unused,
1408                              struct perf_header *ph, int fd,
1409                              void *data __maybe_unused)
1410 {
1411         ph->env.os_release = do_read_string(fd, ph);
1412         return ph->env.os_release ? 0 : -ENOMEM;
1413 }
1414
1415 static int process_version(struct perf_file_section *section __maybe_unused,
1416                            struct perf_header *ph, int fd,
1417                            void *data __maybe_unused)
1418 {
1419         ph->env.version = do_read_string(fd, ph);
1420         return ph->env.version ? 0 : -ENOMEM;
1421 }
1422
1423 static int process_arch(struct perf_file_section *section __maybe_unused,
1424                         struct perf_header *ph, int fd,
1425                         void *data __maybe_unused)
1426 {
1427         ph->env.arch = do_read_string(fd, ph);
1428         return ph->env.arch ? 0 : -ENOMEM;
1429 }
1430
1431 static int process_nrcpus(struct perf_file_section *section __maybe_unused,
1432                           struct perf_header *ph, int fd,
1433                           void *data __maybe_unused)
1434 {
1435         ssize_t ret;
1436         u32 nr;
1437
1438         ret = readn(fd, &nr, sizeof(nr));
1439         if (ret != sizeof(nr))
1440                 return -1;
1441
1442         if (ph->needs_swap)
1443                 nr = bswap_32(nr);
1444
1445         ph->env.nr_cpus_online = nr;
1446
1447         ret = readn(fd, &nr, sizeof(nr));
1448         if (ret != sizeof(nr))
1449                 return -1;
1450
1451         if (ph->needs_swap)
1452                 nr = bswap_32(nr);
1453
1454         ph->env.nr_cpus_avail = nr;
1455         return 0;
1456 }
1457
1458 static int process_cpudesc(struct perf_file_section *section __maybe_unused,
1459                            struct perf_header *ph, int fd,
1460                            void *data __maybe_unused)
1461 {
1462         ph->env.cpu_desc = do_read_string(fd, ph);
1463         return ph->env.cpu_desc ? 0 : -ENOMEM;
1464 }
1465
1466 static int process_cpuid(struct perf_file_section *section __maybe_unused,
1467                          struct perf_header *ph,  int fd,
1468                          void *data __maybe_unused)
1469 {
1470         ph->env.cpuid = do_read_string(fd, ph);
1471         return ph->env.cpuid ? 0 : -ENOMEM;
1472 }
1473
1474 static int process_total_mem(struct perf_file_section *section __maybe_unused,
1475                              struct perf_header *ph, int fd,
1476                              void *data __maybe_unused)
1477 {
1478         uint64_t mem;
1479         ssize_t ret;
1480
1481         ret = readn(fd, &mem, sizeof(mem));
1482         if (ret != sizeof(mem))
1483                 return -1;
1484
1485         if (ph->needs_swap)
1486                 mem = bswap_64(mem);
1487
1488         ph->env.total_mem = mem;
1489         return 0;
1490 }
1491
1492 static struct perf_evsel *
1493 perf_evlist__find_by_index(struct perf_evlist *evlist, int idx)
1494 {
1495         struct perf_evsel *evsel;
1496
1497         evlist__for_each(evlist, evsel) {
1498                 if (evsel->idx == idx)
1499                         return evsel;
1500         }
1501
1502         return NULL;
1503 }
1504
1505 static void
1506 perf_evlist__set_event_name(struct perf_evlist *evlist,
1507                             struct perf_evsel *event)
1508 {
1509         struct perf_evsel *evsel;
1510
1511         if (!event->name)
1512                 return;
1513
1514         evsel = perf_evlist__find_by_index(evlist, event->idx);
1515         if (!evsel)
1516                 return;
1517
1518         if (evsel->name)
1519                 return;
1520
1521         evsel->name = strdup(event->name);
1522 }
1523
1524 static int
1525 process_event_desc(struct perf_file_section *section __maybe_unused,
1526                    struct perf_header *header, int fd,
1527                    void *data __maybe_unused)
1528 {
1529         struct perf_session *session;
1530         struct perf_evsel *evsel, *events = read_event_desc(header, fd);
1531
1532         if (!events)
1533                 return 0;
1534
1535         session = container_of(header, struct perf_session, header);
1536         for (evsel = events; evsel->attr.size; evsel++)
1537                 perf_evlist__set_event_name(session->evlist, evsel);
1538
1539         free_event_desc(events);
1540
1541         return 0;
1542 }
1543
1544 static int process_cmdline(struct perf_file_section *section __maybe_unused,
1545                            struct perf_header *ph, int fd,
1546                            void *data __maybe_unused)
1547 {
1548         ssize_t ret;
1549         char *str;
1550         u32 nr, i;
1551         struct strbuf sb;
1552
1553         ret = readn(fd, &nr, sizeof(nr));
1554         if (ret != sizeof(nr))
1555                 return -1;
1556
1557         if (ph->needs_swap)
1558                 nr = bswap_32(nr);
1559
1560         ph->env.nr_cmdline = nr;
1561         strbuf_init(&sb, 128);
1562
1563         for (i = 0; i < nr; i++) {
1564                 str = do_read_string(fd, ph);
1565                 if (!str)
1566                         goto error;
1567
1568                 /* include a NULL character at the end */
1569                 strbuf_add(&sb, str, strlen(str) + 1);
1570                 free(str);
1571         }
1572         ph->env.cmdline = strbuf_detach(&sb, NULL);
1573         return 0;
1574
1575 error:
1576         strbuf_release(&sb);
1577         return -1;
1578 }
1579
1580 static int process_cpu_topology(struct perf_file_section *section __maybe_unused,
1581                                 struct perf_header *ph, int fd,
1582                                 void *data __maybe_unused)
1583 {
1584         ssize_t ret;
1585         u32 nr, i;
1586         char *str;
1587         struct strbuf sb;
1588
1589         ret = readn(fd, &nr, sizeof(nr));
1590         if (ret != sizeof(nr))
1591                 return -1;
1592
1593         if (ph->needs_swap)
1594                 nr = bswap_32(nr);
1595
1596         ph->env.nr_sibling_cores = nr;
1597         strbuf_init(&sb, 128);
1598
1599         for (i = 0; i < nr; i++) {
1600                 str = do_read_string(fd, ph);
1601                 if (!str)
1602                         goto error;
1603
1604                 /* include a NULL character at the end */
1605                 strbuf_add(&sb, str, strlen(str) + 1);
1606                 free(str);
1607         }
1608         ph->env.sibling_cores = strbuf_detach(&sb, NULL);
1609
1610         ret = readn(fd, &nr, sizeof(nr));
1611         if (ret != sizeof(nr))
1612                 return -1;
1613
1614         if (ph->needs_swap)
1615                 nr = bswap_32(nr);
1616
1617         ph->env.nr_sibling_threads = nr;
1618
1619         for (i = 0; i < nr; i++) {
1620                 str = do_read_string(fd, ph);
1621                 if (!str)
1622                         goto error;
1623
1624                 /* include a NULL character at the end */
1625                 strbuf_add(&sb, str, strlen(str) + 1);
1626                 free(str);
1627         }
1628         ph->env.sibling_threads = strbuf_detach(&sb, NULL);
1629         return 0;
1630
1631 error:
1632         strbuf_release(&sb);
1633         return -1;
1634 }
1635
1636 static int process_numa_topology(struct perf_file_section *section __maybe_unused,
1637                                  struct perf_header *ph, int fd,
1638                                  void *data __maybe_unused)
1639 {
1640         ssize_t ret;
1641         u32 nr, node, i;
1642         char *str;
1643         uint64_t mem_total, mem_free;
1644         struct strbuf sb;
1645
1646         /* nr nodes */
1647         ret = readn(fd, &nr, sizeof(nr));
1648         if (ret != sizeof(nr))
1649                 goto error;
1650
1651         if (ph->needs_swap)
1652                 nr = bswap_32(nr);
1653
1654         ph->env.nr_numa_nodes = nr;
1655         strbuf_init(&sb, 256);
1656
1657         for (i = 0; i < nr; i++) {
1658                 /* node number */
1659                 ret = readn(fd, &node, sizeof(node));
1660                 if (ret != sizeof(node))
1661                         goto error;
1662
1663                 ret = readn(fd, &mem_total, sizeof(u64));
1664                 if (ret != sizeof(u64))
1665                         goto error;
1666
1667                 ret = readn(fd, &mem_free, sizeof(u64));
1668                 if (ret != sizeof(u64))
1669                         goto error;
1670
1671                 if (ph->needs_swap) {
1672                         node = bswap_32(node);
1673                         mem_total = bswap_64(mem_total);
1674                         mem_free = bswap_64(mem_free);
1675                 }
1676
1677                 strbuf_addf(&sb, "%u:%"PRIu64":%"PRIu64":",
1678                             node, mem_total, mem_free);
1679
1680                 str = do_read_string(fd, ph);
1681                 if (!str)
1682                         goto error;
1683
1684                 /* include a NULL character at the end */
1685                 strbuf_add(&sb, str, strlen(str) + 1);
1686                 free(str);
1687         }
1688         ph->env.numa_nodes = strbuf_detach(&sb, NULL);
1689         return 0;
1690
1691 error:
1692         strbuf_release(&sb);
1693         return -1;
1694 }
1695
1696 static int process_pmu_mappings(struct perf_file_section *section __maybe_unused,
1697                                 struct perf_header *ph, int fd,
1698                                 void *data __maybe_unused)
1699 {
1700         ssize_t ret;
1701         char *name;
1702         u32 pmu_num;
1703         u32 type;
1704         struct strbuf sb;
1705
1706         ret = readn(fd, &pmu_num, sizeof(pmu_num));
1707         if (ret != sizeof(pmu_num))
1708                 return -1;
1709
1710         if (ph->needs_swap)
1711                 pmu_num = bswap_32(pmu_num);
1712
1713         if (!pmu_num) {
1714                 pr_debug("pmu mappings not available\n");
1715                 return 0;
1716         }
1717
1718         ph->env.nr_pmu_mappings = pmu_num;
1719         strbuf_init(&sb, 128);
1720
1721         while (pmu_num) {
1722                 if (readn(fd, &type, sizeof(type)) != sizeof(type))
1723                         goto error;
1724                 if (ph->needs_swap)
1725                         type = bswap_32(type);
1726
1727                 name = do_read_string(fd, ph);
1728                 if (!name)
1729                         goto error;
1730
1731                 strbuf_addf(&sb, "%u:%s", type, name);
1732                 /* include a NULL character at the end */
1733                 strbuf_add(&sb, "", 1);
1734
1735                 free(name);
1736                 pmu_num--;
1737         }
1738         ph->env.pmu_mappings = strbuf_detach(&sb, NULL);
1739         return 0;
1740
1741 error:
1742         strbuf_release(&sb);
1743         return -1;
1744 }
1745
1746 static int process_group_desc(struct perf_file_section *section __maybe_unused,
1747                               struct perf_header *ph, int fd,
1748                               void *data __maybe_unused)
1749 {
1750         size_t ret = -1;
1751         u32 i, nr, nr_groups;
1752         struct perf_session *session;
1753         struct perf_evsel *evsel, *leader = NULL;
1754         struct group_desc {
1755                 char *name;
1756                 u32 leader_idx;
1757                 u32 nr_members;
1758         } *desc;
1759
1760         if (readn(fd, &nr_groups, sizeof(nr_groups)) != sizeof(nr_groups))
1761                 return -1;
1762
1763         if (ph->needs_swap)
1764                 nr_groups = bswap_32(nr_groups);
1765
1766         ph->env.nr_groups = nr_groups;
1767         if (!nr_groups) {
1768                 pr_debug("group desc not available\n");
1769                 return 0;
1770         }
1771
1772         desc = calloc(nr_groups, sizeof(*desc));
1773         if (!desc)
1774                 return -1;
1775
1776         for (i = 0; i < nr_groups; i++) {
1777                 desc[i].name = do_read_string(fd, ph);
1778                 if (!desc[i].name)
1779                         goto out_free;
1780
1781                 if (readn(fd, &desc[i].leader_idx, sizeof(u32)) != sizeof(u32))
1782                         goto out_free;
1783
1784                 if (readn(fd, &desc[i].nr_members, sizeof(u32)) != sizeof(u32))
1785                         goto out_free;
1786
1787                 if (ph->needs_swap) {
1788                         desc[i].leader_idx = bswap_32(desc[i].leader_idx);
1789                         desc[i].nr_members = bswap_32(desc[i].nr_members);
1790                 }
1791         }
1792
1793         /*
1794          * Rebuild group relationship based on the group_desc
1795          */
1796         session = container_of(ph, struct perf_session, header);
1797         session->evlist->nr_groups = nr_groups;
1798
1799         i = nr = 0;
1800         evlist__for_each(session->evlist, evsel) {
1801                 if (evsel->idx == (int) desc[i].leader_idx) {
1802                         evsel->leader = evsel;
1803                         /* {anon_group} is a dummy name */
1804                         if (strcmp(desc[i].name, "{anon_group}")) {
1805                                 evsel->group_name = desc[i].name;
1806                                 desc[i].name = NULL;
1807                         }
1808                         evsel->nr_members = desc[i].nr_members;
1809
1810                         if (i >= nr_groups || nr > 0) {
1811                                 pr_debug("invalid group desc\n");
1812                                 goto out_free;
1813                         }
1814
1815                         leader = evsel;
1816                         nr = evsel->nr_members - 1;
1817                         i++;
1818                 } else if (nr) {
1819                         /* This is a group member */
1820                         evsel->leader = leader;
1821
1822                         nr--;
1823                 }
1824         }
1825
1826         if (i != nr_groups || nr != 0) {
1827                 pr_debug("invalid group desc\n");
1828                 goto out_free;
1829         }
1830
1831         ret = 0;
1832 out_free:
1833         for (i = 0; i < nr_groups; i++)
1834                 zfree(&desc[i].name);
1835         free(desc);
1836
1837         return ret;
1838 }
1839
1840 static int process_auxtrace(struct perf_file_section *section,
1841                             struct perf_header *ph, int fd,
1842                             void *data __maybe_unused)
1843 {
1844         struct perf_session *session;
1845         int err;
1846
1847         session = container_of(ph, struct perf_session, header);
1848
1849         err = auxtrace_index__process(fd, section->size, session,
1850                                       ph->needs_swap);
1851         if (err < 0)
1852                 pr_err("Failed to process auxtrace index\n");
1853         return err;
1854 }
1855
1856 struct feature_ops {
1857         int (*write)(int fd, struct perf_header *h, struct perf_evlist *evlist);
1858         void (*print)(struct perf_header *h, int fd, FILE *fp);
1859         int (*process)(struct perf_file_section *section,
1860                        struct perf_header *h, int fd, void *data);
1861         const char *name;
1862         bool full_only;
1863 };
1864
1865 #define FEAT_OPA(n, func) \
1866         [n] = { .name = #n, .write = write_##func, .print = print_##func }
1867 #define FEAT_OPP(n, func) \
1868         [n] = { .name = #n, .write = write_##func, .print = print_##func, \
1869                 .process = process_##func }
1870 #define FEAT_OPF(n, func) \
1871         [n] = { .name = #n, .write = write_##func, .print = print_##func, \
1872                 .process = process_##func, .full_only = true }
1873
1874 /* feature_ops not implemented: */
1875 #define print_tracing_data      NULL
1876 #define print_build_id          NULL
1877
1878 static const struct feature_ops feat_ops[HEADER_LAST_FEATURE] = {
1879         FEAT_OPP(HEADER_TRACING_DATA,   tracing_data),
1880         FEAT_OPP(HEADER_BUILD_ID,       build_id),
1881         FEAT_OPP(HEADER_HOSTNAME,       hostname),
1882         FEAT_OPP(HEADER_OSRELEASE,      osrelease),
1883         FEAT_OPP(HEADER_VERSION,        version),
1884         FEAT_OPP(HEADER_ARCH,           arch),
1885         FEAT_OPP(HEADER_NRCPUS,         nrcpus),
1886         FEAT_OPP(HEADER_CPUDESC,        cpudesc),
1887         FEAT_OPP(HEADER_CPUID,          cpuid),
1888         FEAT_OPP(HEADER_TOTAL_MEM,      total_mem),
1889         FEAT_OPP(HEADER_EVENT_DESC,     event_desc),
1890         FEAT_OPP(HEADER_CMDLINE,        cmdline),
1891         FEAT_OPF(HEADER_CPU_TOPOLOGY,   cpu_topology),
1892         FEAT_OPF(HEADER_NUMA_TOPOLOGY,  numa_topology),
1893         FEAT_OPA(HEADER_BRANCH_STACK,   branch_stack),
1894         FEAT_OPP(HEADER_PMU_MAPPINGS,   pmu_mappings),
1895         FEAT_OPP(HEADER_GROUP_DESC,     group_desc),
1896         FEAT_OPP(HEADER_AUXTRACE,       auxtrace),
1897 };
1898
1899 struct header_print_data {
1900         FILE *fp;
1901         bool full; /* extended list of headers */
1902 };
1903
1904 static int perf_file_section__fprintf_info(struct perf_file_section *section,
1905                                            struct perf_header *ph,
1906                                            int feat, int fd, void *data)
1907 {
1908         struct header_print_data *hd = data;
1909
1910         if (lseek(fd, section->offset, SEEK_SET) == (off_t)-1) {
1911                 pr_debug("Failed to lseek to %" PRIu64 " offset for feature "
1912                                 "%d, continuing...\n", section->offset, feat);
1913                 return 0;
1914         }
1915         if (feat >= HEADER_LAST_FEATURE) {
1916                 pr_warning("unknown feature %d\n", feat);
1917                 return 0;
1918         }
1919         if (!feat_ops[feat].print)
1920                 return 0;
1921
1922         if (!feat_ops[feat].full_only || hd->full)
1923                 feat_ops[feat].print(ph, fd, hd->fp);
1924         else
1925                 fprintf(hd->fp, "# %s info available, use -I to display\n",
1926                         feat_ops[feat].name);
1927
1928         return 0;
1929 }
1930
1931 int perf_header__fprintf_info(struct perf_session *session, FILE *fp, bool full)
1932 {
1933         struct header_print_data hd;
1934         struct perf_header *header = &session->header;
1935         int fd = perf_data_file__fd(session->file);
1936         hd.fp = fp;
1937         hd.full = full;
1938
1939         perf_header__process_sections(header, fd, &hd,
1940                                       perf_file_section__fprintf_info);
1941         return 0;
1942 }
1943
1944 static int do_write_feat(int fd, struct perf_header *h, int type,
1945                          struct perf_file_section **p,
1946                          struct perf_evlist *evlist)
1947 {
1948         int err;
1949         int ret = 0;
1950
1951         if (perf_header__has_feat(h, type)) {
1952                 if (!feat_ops[type].write)
1953                         return -1;
1954
1955                 (*p)->offset = lseek(fd, 0, SEEK_CUR);
1956
1957                 err = feat_ops[type].write(fd, h, evlist);
1958                 if (err < 0) {
1959                         pr_debug("failed to write feature %d\n", type);
1960
1961                         /* undo anything written */
1962                         lseek(fd, (*p)->offset, SEEK_SET);
1963
1964                         return -1;
1965                 }
1966                 (*p)->size = lseek(fd, 0, SEEK_CUR) - (*p)->offset;
1967                 (*p)++;
1968         }
1969         return ret;
1970 }
1971
1972 static int perf_header__adds_write(struct perf_header *header,
1973                                    struct perf_evlist *evlist, int fd)
1974 {
1975         int nr_sections;
1976         struct perf_file_section *feat_sec, *p;
1977         int sec_size;
1978         u64 sec_start;
1979         int feat;
1980         int err;
1981
1982         nr_sections = bitmap_weight(header->adds_features, HEADER_FEAT_BITS);
1983         if (!nr_sections)
1984                 return 0;
1985
1986         feat_sec = p = calloc(nr_sections, sizeof(*feat_sec));
1987         if (feat_sec == NULL)
1988                 return -ENOMEM;
1989
1990         sec_size = sizeof(*feat_sec) * nr_sections;
1991
1992         sec_start = header->feat_offset;
1993         lseek(fd, sec_start + sec_size, SEEK_SET);
1994
1995         for_each_set_bit(feat, header->adds_features, HEADER_FEAT_BITS) {
1996                 if (do_write_feat(fd, header, feat, &p, evlist))
1997                         perf_header__clear_feat(header, feat);
1998         }
1999
2000         lseek(fd, sec_start, SEEK_SET);
2001         /*
2002          * may write more than needed due to dropped feature, but
2003          * this is okay, reader will skip the mising entries
2004          */
2005         err = do_write(fd, feat_sec, sec_size);
2006         if (err < 0)
2007                 pr_debug("failed to write feature section\n");
2008         free(feat_sec);
2009         return err;
2010 }
2011
2012 int perf_header__write_pipe(int fd)
2013 {
2014         struct perf_pipe_file_header f_header;
2015         int err;
2016
2017         f_header = (struct perf_pipe_file_header){
2018                 .magic     = PERF_MAGIC,
2019                 .size      = sizeof(f_header),
2020         };
2021
2022         err = do_write(fd, &f_header, sizeof(f_header));
2023         if (err < 0) {
2024                 pr_debug("failed to write perf pipe header\n");
2025                 return err;
2026         }
2027
2028         return 0;
2029 }
2030
2031 int perf_session__write_header(struct perf_session *session,
2032                                struct perf_evlist *evlist,
2033                                int fd, bool at_exit)
2034 {
2035         struct perf_file_header f_header;
2036         struct perf_file_attr   f_attr;
2037         struct perf_header *header = &session->header;
2038         struct perf_evsel *evsel;
2039         u64 attr_offset;
2040         int err;
2041
2042         lseek(fd, sizeof(f_header), SEEK_SET);
2043
2044         evlist__for_each(session->evlist, evsel) {
2045                 evsel->id_offset = lseek(fd, 0, SEEK_CUR);
2046                 err = do_write(fd, evsel->id, evsel->ids * sizeof(u64));
2047                 if (err < 0) {
2048                         pr_debug("failed to write perf header\n");
2049                         return err;
2050                 }
2051         }
2052
2053         attr_offset = lseek(fd, 0, SEEK_CUR);
2054
2055         evlist__for_each(evlist, evsel) {
2056                 f_attr = (struct perf_file_attr){
2057                         .attr = evsel->attr,
2058                         .ids  = {
2059                                 .offset = evsel->id_offset,
2060                                 .size   = evsel->ids * sizeof(u64),
2061                         }
2062                 };
2063                 err = do_write(fd, &f_attr, sizeof(f_attr));
2064                 if (err < 0) {
2065                         pr_debug("failed to write perf header attribute\n");
2066                         return err;
2067                 }
2068         }
2069
2070         if (!header->data_offset)
2071                 header->data_offset = lseek(fd, 0, SEEK_CUR);
2072         header->feat_offset = header->data_offset + header->data_size;
2073
2074         if (at_exit) {
2075                 err = perf_header__adds_write(header, evlist, fd);
2076                 if (err < 0)
2077                         return err;
2078         }
2079
2080         f_header = (struct perf_file_header){
2081                 .magic     = PERF_MAGIC,
2082                 .size      = sizeof(f_header),
2083                 .attr_size = sizeof(f_attr),
2084                 .attrs = {
2085                         .offset = attr_offset,
2086                         .size   = evlist->nr_entries * sizeof(f_attr),
2087                 },
2088                 .data = {
2089                         .offset = header->data_offset,
2090                         .size   = header->data_size,
2091                 },
2092                 /* event_types is ignored, store zeros */
2093         };
2094
2095         memcpy(&f_header.adds_features, &header->adds_features, sizeof(header->adds_features));
2096
2097         lseek(fd, 0, SEEK_SET);
2098         err = do_write(fd, &f_header, sizeof(f_header));
2099         if (err < 0) {
2100                 pr_debug("failed to write perf header\n");
2101                 return err;
2102         }
2103         lseek(fd, header->data_offset + header->data_size, SEEK_SET);
2104
2105         return 0;
2106 }
2107
2108 static int perf_header__getbuffer64(struct perf_header *header,
2109                                     int fd, void *buf, size_t size)
2110 {
2111         if (readn(fd, buf, size) <= 0)
2112                 return -1;
2113
2114         if (header->needs_swap)
2115                 mem_bswap_64(buf, size);
2116
2117         return 0;
2118 }
2119
2120 int perf_header__process_sections(struct perf_header *header, int fd,
2121                                   void *data,
2122                                   int (*process)(struct perf_file_section *section,
2123                                                  struct perf_header *ph,
2124                                                  int feat, int fd, void *data))
2125 {
2126         struct perf_file_section *feat_sec, *sec;
2127         int nr_sections;
2128         int sec_size;
2129         int feat;
2130         int err;
2131
2132         nr_sections = bitmap_weight(header->adds_features, HEADER_FEAT_BITS);
2133         if (!nr_sections)
2134                 return 0;
2135
2136         feat_sec = sec = calloc(nr_sections, sizeof(*feat_sec));
2137         if (!feat_sec)
2138                 return -1;
2139
2140         sec_size = sizeof(*feat_sec) * nr_sections;
2141
2142         lseek(fd, header->feat_offset, SEEK_SET);
2143
2144         err = perf_header__getbuffer64(header, fd, feat_sec, sec_size);
2145         if (err < 0)
2146                 goto out_free;
2147
2148         for_each_set_bit(feat, header->adds_features, HEADER_LAST_FEATURE) {
2149                 err = process(sec++, header, feat, fd, data);
2150                 if (err < 0)
2151                         goto out_free;
2152         }
2153         err = 0;
2154 out_free:
2155         free(feat_sec);
2156         return err;
2157 }
2158
2159 static const int attr_file_abi_sizes[] = {
2160         [0] = PERF_ATTR_SIZE_VER0,
2161         [1] = PERF_ATTR_SIZE_VER1,
2162         [2] = PERF_ATTR_SIZE_VER2,
2163         [3] = PERF_ATTR_SIZE_VER3,
2164         [4] = PERF_ATTR_SIZE_VER4,
2165         0,
2166 };
2167
2168 /*
2169  * In the legacy file format, the magic number is not used to encode endianness.
2170  * hdr_sz was used to encode endianness. But given that hdr_sz can vary based
2171  * on ABI revisions, we need to try all combinations for all endianness to
2172  * detect the endianness.
2173  */
2174 static int try_all_file_abis(uint64_t hdr_sz, struct perf_header *ph)
2175 {
2176         uint64_t ref_size, attr_size;
2177         int i;
2178
2179         for (i = 0 ; attr_file_abi_sizes[i]; i++) {
2180                 ref_size = attr_file_abi_sizes[i]
2181                          + sizeof(struct perf_file_section);
2182                 if (hdr_sz != ref_size) {
2183                         attr_size = bswap_64(hdr_sz);
2184                         if (attr_size != ref_size)
2185                                 continue;
2186
2187                         ph->needs_swap = true;
2188                 }
2189                 pr_debug("ABI%d perf.data file detected, need_swap=%d\n",
2190                          i,
2191                          ph->needs_swap);
2192                 return 0;
2193         }
2194         /* could not determine endianness */
2195         return -1;
2196 }
2197
2198 #define PERF_PIPE_HDR_VER0      16
2199
2200 static const size_t attr_pipe_abi_sizes[] = {
2201         [0] = PERF_PIPE_HDR_VER0,
2202         0,
2203 };
2204
2205 /*
2206  * In the legacy pipe format, there is an implicit assumption that endiannesss
2207  * between host recording the samples, and host parsing the samples is the
2208  * same. This is not always the case given that the pipe output may always be
2209  * redirected into a file and analyzed on a different machine with possibly a
2210  * different endianness and perf_event ABI revsions in the perf tool itself.
2211  */
2212 static int try_all_pipe_abis(uint64_t hdr_sz, struct perf_header *ph)
2213 {
2214         u64 attr_size;
2215         int i;
2216
2217         for (i = 0 ; attr_pipe_abi_sizes[i]; i++) {
2218                 if (hdr_sz != attr_pipe_abi_sizes[i]) {
2219                         attr_size = bswap_64(hdr_sz);
2220                         if (attr_size != hdr_sz)
2221                                 continue;
2222
2223                         ph->needs_swap = true;
2224                 }
2225                 pr_debug("Pipe ABI%d perf.data file detected\n", i);
2226                 return 0;
2227         }
2228         return -1;
2229 }
2230
2231 bool is_perf_magic(u64 magic)
2232 {
2233         if (!memcmp(&magic, __perf_magic1, sizeof(magic))
2234                 || magic == __perf_magic2
2235                 || magic == __perf_magic2_sw)
2236                 return true;
2237
2238         return false;
2239 }
2240
2241 static int check_magic_endian(u64 magic, uint64_t hdr_sz,
2242                               bool is_pipe, struct perf_header *ph)
2243 {
2244         int ret;
2245
2246         /* check for legacy format */
2247         ret = memcmp(&magic, __perf_magic1, sizeof(magic));
2248         if (ret == 0) {
2249                 ph->version = PERF_HEADER_VERSION_1;
2250                 pr_debug("legacy perf.data format\n");
2251                 if (is_pipe)
2252                         return try_all_pipe_abis(hdr_sz, ph);
2253
2254                 return try_all_file_abis(hdr_sz, ph);
2255         }
2256         /*
2257          * the new magic number serves two purposes:
2258          * - unique number to identify actual perf.data files
2259          * - encode endianness of file
2260          */
2261         ph->version = PERF_HEADER_VERSION_2;
2262
2263         /* check magic number with one endianness */
2264         if (magic == __perf_magic2)
2265                 return 0;
2266
2267         /* check magic number with opposite endianness */
2268         if (magic != __perf_magic2_sw)
2269                 return -1;
2270
2271         ph->needs_swap = true;
2272
2273         return 0;
2274 }
2275
2276 int perf_file_header__read(struct perf_file_header *header,
2277                            struct perf_header *ph, int fd)
2278 {
2279         ssize_t ret;
2280
2281         lseek(fd, 0, SEEK_SET);
2282
2283         ret = readn(fd, header, sizeof(*header));
2284         if (ret <= 0)
2285                 return -1;
2286
2287         if (check_magic_endian(header->magic,
2288                                header->attr_size, false, ph) < 0) {
2289                 pr_debug("magic/endian check failed\n");
2290                 return -1;
2291         }
2292
2293         if (ph->needs_swap) {
2294                 mem_bswap_64(header, offsetof(struct perf_file_header,
2295                              adds_features));
2296         }
2297
2298         if (header->size != sizeof(*header)) {
2299                 /* Support the previous format */
2300                 if (header->size == offsetof(typeof(*header), adds_features))
2301                         bitmap_zero(header->adds_features, HEADER_FEAT_BITS);
2302                 else
2303                         return -1;
2304         } else if (ph->needs_swap) {
2305                 /*
2306                  * feature bitmap is declared as an array of unsigned longs --
2307                  * not good since its size can differ between the host that
2308                  * generated the data file and the host analyzing the file.
2309                  *
2310                  * We need to handle endianness, but we don't know the size of
2311                  * the unsigned long where the file was generated. Take a best
2312                  * guess at determining it: try 64-bit swap first (ie., file
2313                  * created on a 64-bit host), and check if the hostname feature
2314                  * bit is set (this feature bit is forced on as of fbe96f2).
2315                  * If the bit is not, undo the 64-bit swap and try a 32-bit
2316                  * swap. If the hostname bit is still not set (e.g., older data
2317                  * file), punt and fallback to the original behavior --
2318                  * clearing all feature bits and setting buildid.
2319                  */
2320                 mem_bswap_64(&header->adds_features,
2321                             BITS_TO_U64(HEADER_FEAT_BITS));
2322
2323                 if (!test_bit(HEADER_HOSTNAME, header->adds_features)) {
2324                         /* unswap as u64 */
2325                         mem_bswap_64(&header->adds_features,
2326                                     BITS_TO_U64(HEADER_FEAT_BITS));
2327
2328                         /* unswap as u32 */
2329                         mem_bswap_32(&header->adds_features,
2330                                     BITS_TO_U32(HEADER_FEAT_BITS));
2331                 }
2332
2333                 if (!test_bit(HEADER_HOSTNAME, header->adds_features)) {
2334                         bitmap_zero(header->adds_features, HEADER_FEAT_BITS);
2335                         set_bit(HEADER_BUILD_ID, header->adds_features);
2336                 }
2337         }
2338
2339         memcpy(&ph->adds_features, &header->adds_features,
2340                sizeof(ph->adds_features));
2341
2342         ph->data_offset  = header->data.offset;
2343         ph->data_size    = header->data.size;
2344         ph->feat_offset  = header->data.offset + header->data.size;
2345         return 0;
2346 }
2347
2348 static int perf_file_section__process(struct perf_file_section *section,
2349                                       struct perf_header *ph,
2350                                       int feat, int fd, void *data)
2351 {
2352         if (lseek(fd, section->offset, SEEK_SET) == (off_t)-1) {
2353                 pr_debug("Failed to lseek to %" PRIu64 " offset for feature "
2354                           "%d, continuing...\n", section->offset, feat);
2355                 return 0;
2356         }
2357
2358         if (feat >= HEADER_LAST_FEATURE) {
2359                 pr_debug("unknown feature %d, continuing...\n", feat);
2360                 return 0;
2361         }
2362
2363         if (!feat_ops[feat].process)
2364                 return 0;
2365
2366         return feat_ops[feat].process(section, ph, fd, data);
2367 }
2368
2369 static int perf_file_header__read_pipe(struct perf_pipe_file_header *header,
2370                                        struct perf_header *ph, int fd,
2371                                        bool repipe)
2372 {
2373         ssize_t ret;
2374
2375         ret = readn(fd, header, sizeof(*header));
2376         if (ret <= 0)
2377                 return -1;
2378
2379         if (check_magic_endian(header->magic, header->size, true, ph) < 0) {
2380                 pr_debug("endian/magic failed\n");
2381                 return -1;
2382         }
2383
2384         if (ph->needs_swap)
2385                 header->size = bswap_64(header->size);
2386
2387         if (repipe && do_write(STDOUT_FILENO, header, sizeof(*header)) < 0)
2388                 return -1;
2389
2390         return 0;
2391 }
2392
2393 static int perf_header__read_pipe(struct perf_session *session)
2394 {
2395         struct perf_header *header = &session->header;
2396         struct perf_pipe_file_header f_header;
2397
2398         if (perf_file_header__read_pipe(&f_header, header,
2399                                         perf_data_file__fd(session->file),
2400                                         session->repipe) < 0) {
2401                 pr_debug("incompatible file format\n");
2402                 return -EINVAL;
2403         }
2404
2405         return 0;
2406 }
2407
2408 static int read_attr(int fd, struct perf_header *ph,
2409                      struct perf_file_attr *f_attr)
2410 {
2411         struct perf_event_attr *attr = &f_attr->attr;
2412         size_t sz, left;
2413         size_t our_sz = sizeof(f_attr->attr);
2414         ssize_t ret;
2415
2416         memset(f_attr, 0, sizeof(*f_attr));
2417
2418         /* read minimal guaranteed structure */
2419         ret = readn(fd, attr, PERF_ATTR_SIZE_VER0);
2420         if (ret <= 0) {
2421                 pr_debug("cannot read %d bytes of header attr\n",
2422                          PERF_ATTR_SIZE_VER0);
2423                 return -1;
2424         }
2425
2426         /* on file perf_event_attr size */
2427         sz = attr->size;
2428
2429         if (ph->needs_swap)
2430                 sz = bswap_32(sz);
2431
2432         if (sz == 0) {
2433                 /* assume ABI0 */
2434                 sz =  PERF_ATTR_SIZE_VER0;
2435         } else if (sz > our_sz) {
2436                 pr_debug("file uses a more recent and unsupported ABI"
2437                          " (%zu bytes extra)\n", sz - our_sz);
2438                 return -1;
2439         }
2440         /* what we have not yet read and that we know about */
2441         left = sz - PERF_ATTR_SIZE_VER0;
2442         if (left) {
2443                 void *ptr = attr;
2444                 ptr += PERF_ATTR_SIZE_VER0;
2445
2446                 ret = readn(fd, ptr, left);
2447         }
2448         /* read perf_file_section, ids are read in caller */
2449         ret = readn(fd, &f_attr->ids, sizeof(f_attr->ids));
2450
2451         return ret <= 0 ? -1 : 0;
2452 }
2453
2454 static int perf_evsel__prepare_tracepoint_event(struct perf_evsel *evsel,
2455                                                 struct pevent *pevent)
2456 {
2457         struct event_format *event;
2458         char bf[128];
2459
2460         /* already prepared */
2461         if (evsel->tp_format)
2462                 return 0;
2463
2464         if (pevent == NULL) {
2465                 pr_debug("broken or missing trace data\n");
2466                 return -1;
2467         }
2468
2469         event = pevent_find_event(pevent, evsel->attr.config);
2470         if (event == NULL)
2471                 return -1;
2472
2473         if (!evsel->name) {
2474                 snprintf(bf, sizeof(bf), "%s:%s", event->system, event->name);
2475                 evsel->name = strdup(bf);
2476                 if (evsel->name == NULL)
2477                         return -1;
2478         }
2479
2480         evsel->tp_format = event;
2481         return 0;
2482 }
2483
2484 static int perf_evlist__prepare_tracepoint_events(struct perf_evlist *evlist,
2485                                                   struct pevent *pevent)
2486 {
2487         struct perf_evsel *pos;
2488
2489         evlist__for_each(evlist, pos) {
2490                 if (pos->attr.type == PERF_TYPE_TRACEPOINT &&
2491                     perf_evsel__prepare_tracepoint_event(pos, pevent))
2492                         return -1;
2493         }
2494
2495         return 0;
2496 }
2497
2498 int perf_session__read_header(struct perf_session *session)
2499 {
2500         struct perf_data_file *file = session->file;
2501         struct perf_header *header = &session->header;
2502         struct perf_file_header f_header;
2503         struct perf_file_attr   f_attr;
2504         u64                     f_id;
2505         int nr_attrs, nr_ids, i, j;
2506         int fd = perf_data_file__fd(file);
2507
2508         session->evlist = perf_evlist__new();
2509         if (session->evlist == NULL)
2510                 return -ENOMEM;
2511
2512         if (perf_data_file__is_pipe(file))
2513                 return perf_header__read_pipe(session);
2514
2515         if (perf_file_header__read(&f_header, header, fd) < 0)
2516                 return -EINVAL;
2517
2518         /*
2519          * Sanity check that perf.data was written cleanly; data size is
2520          * initialized to 0 and updated only if the on_exit function is run.
2521          * If data size is still 0 then the file contains only partial
2522          * information.  Just warn user and process it as much as it can.
2523          */
2524         if (f_header.data.size == 0) {
2525                 pr_warning("WARNING: The %s file's data size field is 0 which is unexpected.\n"
2526                            "Was the 'perf record' command properly terminated?\n",
2527                            file->path);
2528         }
2529
2530         nr_attrs = f_header.attrs.size / f_header.attr_size;
2531         lseek(fd, f_header.attrs.offset, SEEK_SET);
2532
2533         for (i = 0; i < nr_attrs; i++) {
2534                 struct perf_evsel *evsel;
2535                 off_t tmp;
2536
2537                 if (read_attr(fd, header, &f_attr) < 0)
2538                         goto out_errno;
2539
2540                 if (header->needs_swap) {
2541                         f_attr.ids.size   = bswap_64(f_attr.ids.size);
2542                         f_attr.ids.offset = bswap_64(f_attr.ids.offset);
2543                         perf_event__attr_swap(&f_attr.attr);
2544                 }
2545
2546                 tmp = lseek(fd, 0, SEEK_CUR);
2547                 evsel = perf_evsel__new(&f_attr.attr);
2548
2549                 if (evsel == NULL)
2550                         goto out_delete_evlist;
2551
2552                 evsel->needs_swap = header->needs_swap;
2553                 /*
2554                  * Do it before so that if perf_evsel__alloc_id fails, this
2555                  * entry gets purged too at perf_evlist__delete().
2556                  */
2557                 perf_evlist__add(session->evlist, evsel);
2558
2559                 nr_ids = f_attr.ids.size / sizeof(u64);
2560                 /*
2561                  * We don't have the cpu and thread maps on the header, so
2562                  * for allocating the perf_sample_id table we fake 1 cpu and
2563                  * hattr->ids threads.
2564                  */
2565                 if (perf_evsel__alloc_id(evsel, 1, nr_ids))
2566                         goto out_delete_evlist;
2567
2568                 lseek(fd, f_attr.ids.offset, SEEK_SET);
2569
2570                 for (j = 0; j < nr_ids; j++) {
2571                         if (perf_header__getbuffer64(header, fd, &f_id, sizeof(f_id)))
2572                                 goto out_errno;
2573
2574                         perf_evlist__id_add(session->evlist, evsel, 0, j, f_id);
2575                 }
2576
2577                 lseek(fd, tmp, SEEK_SET);
2578         }
2579
2580         symbol_conf.nr_events = nr_attrs;
2581
2582         perf_header__process_sections(header, fd, &session->tevent,
2583                                       perf_file_section__process);
2584
2585         if (perf_evlist__prepare_tracepoint_events(session->evlist,
2586                                                    session->tevent.pevent))
2587                 goto out_delete_evlist;
2588
2589         return 0;
2590 out_errno:
2591         return -errno;
2592
2593 out_delete_evlist:
2594         perf_evlist__delete(session->evlist);
2595         session->evlist = NULL;
2596         return -ENOMEM;
2597 }
2598
2599 int perf_event__synthesize_attr(struct perf_tool *tool,
2600                                 struct perf_event_attr *attr, u32 ids, u64 *id,
2601                                 perf_event__handler_t process)
2602 {
2603         union perf_event *ev;
2604         size_t size;
2605         int err;
2606
2607         size = sizeof(struct perf_event_attr);
2608         size = PERF_ALIGN(size, sizeof(u64));
2609         size += sizeof(struct perf_event_header);
2610         size += ids * sizeof(u64);
2611
2612         ev = malloc(size);
2613
2614         if (ev == NULL)
2615                 return -ENOMEM;
2616
2617         ev->attr.attr = *attr;
2618         memcpy(ev->attr.id, id, ids * sizeof(u64));
2619
2620         ev->attr.header.type = PERF_RECORD_HEADER_ATTR;
2621         ev->attr.header.size = (u16)size;
2622
2623         if (ev->attr.header.size == size)
2624                 err = process(tool, ev, NULL, NULL);
2625         else
2626                 err = -E2BIG;
2627
2628         free(ev);
2629
2630         return err;
2631 }
2632
2633 int perf_event__synthesize_attrs(struct perf_tool *tool,
2634                                    struct perf_session *session,
2635                                    perf_event__handler_t process)
2636 {
2637         struct perf_evsel *evsel;
2638         int err = 0;
2639
2640         evlist__for_each(session->evlist, evsel) {
2641                 err = perf_event__synthesize_attr(tool, &evsel->attr, evsel->ids,
2642                                                   evsel->id, process);
2643                 if (err) {
2644                         pr_debug("failed to create perf header attribute\n");
2645                         return err;
2646                 }
2647         }
2648
2649         return err;
2650 }
2651
2652 int perf_event__process_attr(struct perf_tool *tool __maybe_unused,
2653                              union perf_event *event,
2654                              struct perf_evlist **pevlist)
2655 {
2656         u32 i, ids, n_ids;
2657         struct perf_evsel *evsel;
2658         struct perf_evlist *evlist = *pevlist;
2659
2660         if (evlist == NULL) {
2661                 *pevlist = evlist = perf_evlist__new();
2662                 if (evlist == NULL)
2663                         return -ENOMEM;
2664         }
2665
2666         evsel = perf_evsel__new(&event->attr.attr);
2667         if (evsel == NULL)
2668                 return -ENOMEM;
2669
2670         perf_evlist__add(evlist, evsel);
2671
2672         ids = event->header.size;
2673         ids -= (void *)&event->attr.id - (void *)event;
2674         n_ids = ids / sizeof(u64);
2675         /*
2676          * We don't have the cpu and thread maps on the header, so
2677          * for allocating the perf_sample_id table we fake 1 cpu and
2678          * hattr->ids threads.
2679          */
2680         if (perf_evsel__alloc_id(evsel, 1, n_ids))
2681                 return -ENOMEM;
2682
2683         for (i = 0; i < n_ids; i++) {
2684                 perf_evlist__id_add(evlist, evsel, 0, i, event->attr.id[i]);
2685         }
2686
2687         symbol_conf.nr_events = evlist->nr_entries;
2688
2689         return 0;
2690 }
2691
2692 int perf_event__synthesize_tracing_data(struct perf_tool *tool, int fd,
2693                                         struct perf_evlist *evlist,
2694                                         perf_event__handler_t process)
2695 {
2696         union perf_event ev;
2697         struct tracing_data *tdata;
2698         ssize_t size = 0, aligned_size = 0, padding;
2699         int err __maybe_unused = 0;
2700
2701         /*
2702          * We are going to store the size of the data followed
2703          * by the data contents. Since the fd descriptor is a pipe,
2704          * we cannot seek back to store the size of the data once
2705          * we know it. Instead we:
2706          *
2707          * - write the tracing data to the temp file
2708          * - get/write the data size to pipe
2709          * - write the tracing data from the temp file
2710          *   to the pipe
2711          */
2712         tdata = tracing_data_get(&evlist->entries, fd, true);
2713         if (!tdata)
2714                 return -1;
2715
2716         memset(&ev, 0, sizeof(ev));
2717
2718         ev.tracing_data.header.type = PERF_RECORD_HEADER_TRACING_DATA;
2719         size = tdata->size;
2720         aligned_size = PERF_ALIGN(size, sizeof(u64));
2721         padding = aligned_size - size;
2722         ev.tracing_data.header.size = sizeof(ev.tracing_data);
2723         ev.tracing_data.size = aligned_size;
2724
2725         process(tool, &ev, NULL, NULL);
2726
2727         /*
2728          * The put function will copy all the tracing data
2729          * stored in temp file to the pipe.
2730          */
2731         tracing_data_put(tdata);
2732
2733         write_padded(fd, NULL, 0, padding);
2734
2735         return aligned_size;
2736 }
2737
2738 int perf_event__process_tracing_data(struct perf_tool *tool __maybe_unused,
2739                                      union perf_event *event,
2740                                      struct perf_session *session)
2741 {
2742         ssize_t size_read, padding, size = event->tracing_data.size;
2743         int fd = perf_data_file__fd(session->file);
2744         off_t offset = lseek(fd, 0, SEEK_CUR);
2745         char buf[BUFSIZ];
2746
2747         /* setup for reading amidst mmap */
2748         lseek(fd, offset + sizeof(struct tracing_data_event),
2749               SEEK_SET);
2750
2751         size_read = trace_report(fd, &session->tevent,
2752                                  session->repipe);
2753         padding = PERF_ALIGN(size_read, sizeof(u64)) - size_read;
2754
2755         if (readn(fd, buf, padding) < 0) {
2756                 pr_err("%s: reading input file", __func__);
2757                 return -1;
2758         }
2759         if (session->repipe) {
2760                 int retw = write(STDOUT_FILENO, buf, padding);
2761                 if (retw <= 0 || retw != padding) {
2762                         pr_err("%s: repiping tracing data padding", __func__);
2763                         return -1;
2764                 }
2765         }
2766
2767         if (size_read + padding != size) {
2768                 pr_err("%s: tracing data size mismatch", __func__);
2769                 return -1;
2770         }
2771
2772         perf_evlist__prepare_tracepoint_events(session->evlist,
2773                                                session->tevent.pevent);
2774
2775         return size_read + padding;
2776 }
2777
2778 int perf_event__synthesize_build_id(struct perf_tool *tool,
2779                                     struct dso *pos, u16 misc,
2780                                     perf_event__handler_t process,
2781                                     struct machine *machine)
2782 {
2783         union perf_event ev;
2784         size_t len;
2785         int err = 0;
2786
2787         if (!pos->hit)
2788                 return err;
2789
2790         memset(&ev, 0, sizeof(ev));
2791
2792         len = pos->long_name_len + 1;
2793         len = PERF_ALIGN(len, NAME_ALIGN);
2794         memcpy(&ev.build_id.build_id, pos->build_id, sizeof(pos->build_id));
2795         ev.build_id.header.type = PERF_RECORD_HEADER_BUILD_ID;
2796         ev.build_id.header.misc = misc;
2797         ev.build_id.pid = machine->pid;
2798         ev.build_id.header.size = sizeof(ev.build_id) + len;
2799         memcpy(&ev.build_id.filename, pos->long_name, pos->long_name_len);
2800
2801         err = process(tool, &ev, NULL, machine);
2802
2803         return err;
2804 }
2805
2806 int perf_event__process_build_id(struct perf_tool *tool __maybe_unused,
2807                                  union perf_event *event,
2808                                  struct perf_session *session)
2809 {
2810         __event_process_build_id(&event->build_id,
2811                                  event->build_id.filename,
2812                                  session);
2813         return 0;
2814 }