tools/perf/builtin-record.c

   1 /*
   2  * builtin-record.c
   3  *
   4  * Builtin record command: Record the profile of a workload
   5  * (or a CPU, or a PID) into the perf.data output file - for
   6  * later analysis via perf report.
   7  */
   8 #include "builtin.h"
   9
  10 #include "perf.h"
  11
  12 #include "util/build-id.h"
  13 #include "util/util.h"
  14 #include "util/parse-options.h"
  15 #include "util/parse-events.h"
  16
  17 #include "util/header.h"
  18 #include "util/event.h"
  19 #include "util/evlist.h"
  20 #include "util/evsel.h"
  21 #include "util/debug.h"
  22 #include "util/session.h"
  23 #include "util/tool.h"
  24 #include "util/symbol.h"
  25 #include "util/cpumap.h"
  26 #include "util/thread_map.h"
  27 #include "util/data.h"
  28
  29 #include <unistd.h>
  30 #include <sched.h>
  31 #include <sys/mman.h>
  32
  33
  34 struct record {
  35         struct perf_tool        tool;
  36         struct record_opts      opts;
  37         u64                     bytes_written;
  38         struct perf_data_file   file;
  39         struct perf_evlist      *evlist;
  40         struct perf_session     *session;
  41         const char              *progname;
  42         int                     realtime_prio;
  43         bool                    no_buildid;
  44         bool                    no_buildid_cache;
  45         long                    samples;
  46 };
  47
  48 static int record__write(struct record *rec, void *bf, size_t size)
  49 {
  50         if (perf_data_file__write(rec->session->file, bf, size) < 0) {
  51                 pr_err("failed to write perf data, error: %m\n");
  52                 return -1;
  53         }
  54
  55         rec->bytes_written += size;
  56         return 0;
  57 }
  58
  59 static int process_synthesized_event(struct perf_tool *tool,
  60                                      union perf_event *event,
  61                                      struct perf_sample *sample __maybe_unused,
  62                                      struct machine *machine __maybe_unused)
  63 {
  64         struct record *rec = container_of(tool, struct record, tool);
  65         return record__write(rec, event, event->header.size);
  66 }
  67
  68 static int record__mmap_read(struct record *rec, struct perf_mmap *md)
  69 {
  70         unsigned int head = perf_mmap__read_head(md);
  71         unsigned int old = md->prev;
  72         unsigned char *data = md->base + page_size;
  73         unsigned long size;
  74         void *buf;
  75         int rc = 0;
  76
  77         if (old == head)
  78                 return 0;
  79
  80         rec->samples++;
  81
  82         size = head - old;
  83
  84         if ((old & md->mask) + size != (head & md->mask)) {
  85                 buf = &data[old & md->mask];
  86                 size = md->mask + 1 - (old & md->mask);
  87                 old += size;
  88
  89                 if (record__write(rec, buf, size) < 0) {
  90                         rc = -1;
  91                         goto out;
  92                 }
  93         }
  94
  95         buf = &data[old & md->mask];
  96         size = head - old;
  97         old += size;
  98
  99         if (record__write(rec, buf, size) < 0) {
 100                 rc = -1;
 101                 goto out;
 102         }
 103
 104         md->prev = old;
 105         perf_mmap__write_tail(md, old);
 106
 107 out:
 108         return rc;
 109 }
 110
 111 static volatile int done = 0;
 112 static volatile int signr = -1;
 113 static volatile int child_finished = 0;
 114
 115 static void sig_handler(int sig)
 116 {
 117         if (sig == SIGCHLD)
 118                 child_finished = 1;
 119         else
 120                 signr = sig;
 121
 122         done = 1;
 123 }
 124
 125 static void record__sig_exit(void)
 126 {
 127         if (signr == -1)
 128                 return;
 129
 130         signal(signr, SIG_DFL);
 131         raise(signr);
 132 }
 133
 134 static int record__open(struct record *rec)
 135 {
 136         char msg[512];
 137         struct perf_evsel *pos;
 138         struct perf_evlist *evlist = rec->evlist;
 139         struct perf_session *session = rec->session;
 140         struct record_opts *opts = &rec->opts;
 141         int rc = 0;
 142
 143         perf_evlist__config(evlist, opts);
 144
 145         evlist__for_each(evlist, pos) {
 146 try_again:
 147                 if (perf_evsel__open(pos, evlist->cpus, evlist->threads) < 0) {
 148                         if (perf_evsel__fallback(pos, errno, msg, sizeof(msg))) {
 149                                 if (verbose)
 150                                         ui__warning("%s\n", msg);
 151                                 goto try_again;
 152                         }
 153
 154                         rc = -errno;
 155                         perf_evsel__open_strerror(pos, &opts->target,
 156                                                   errno, msg, sizeof(msg));
 157                         ui__error("%s\n", msg);
 158                         goto out;
 159                 }
 160         }
 161
 162         if (perf_evlist__apply_filters(evlist)) {
 163                 error("failed to set filter with %d (%s)\n", errno,
 164                         strerror_r(errno, msg, sizeof(msg)));
 165                 rc = -1;
 166                 goto out;
 167         }
 168
 169         if (perf_evlist__mmap(evlist, opts->mmap_pages, false) < 0) {
 170                 if (errno == EPERM) {
 171                         pr_err("Permission error mapping pages.\n"
 172                                "Consider increasing "
 173                                "/proc/sys/kernel/perf_event_mlock_kb,\n"
 174                                "or try again with a smaller value of -m/--mmap_pages.\n"
 175                                "(current value: %u)\n", opts->mmap_pages);
 176                         rc = -errno;
 177                 } else {
 178                         pr_err("failed to mmap with %d (%s)\n", errno,
 179                                 strerror_r(errno, msg, sizeof(msg)));
 180                         rc = -errno;
 181                 }
 182                 goto out;
 183         }
 184
 185         session->evlist = evlist;
 186         perf_session__set_id_hdr_size(session);
 187 out:
 188         return rc;
 189 }
 190
 191 static int process_buildids(struct record *rec)
 192 {
 193         struct perf_data_file *file  = &rec->file;
 194         struct perf_session *session = rec->session;
 195         u64 start = session->header.data_offset;
 196
 197         u64 size = lseek(file->fd, 0, SEEK_CUR);
 198         if (size == 0)
 199                 return 0;
 200
 201         return __perf_session__process_events(session, start,
 202                                               size - start,
 203                                               size, &build_id__mark_dso_hit_ops);
 204 }
 205
 206 static void perf_event__synthesize_guest_os(struct machine *machine, void *data)
 207 {
 208         int err;
 209         struct perf_tool *tool = data;
 210         /*
 211          *As for guest kernel when processing subcommand record&report,
 212          *we arrange module mmap prior to guest kernel mmap and trigger
 213          *a preload dso because default guest module symbols are loaded
 214          *from guest kallsyms instead of /lib/modules/XXX/XXX. This
 215          *method is used to avoid symbol missing when the first addr is
 216          *in module instead of in guest kernel.
 217          */
 218         err = perf_event__synthesize_modules(tool, process_synthesized_event,
 219                                              machine);
 220         if (err < 0)
 221                 pr_err("Couldn't record guest kernel [%d]'s reference"
 222                        " relocation symbol.\n", machine->pid);
 223
 224         /*
 225          * We use _stext for guest kernel because guest kernel's /proc/kallsyms
 226          * have no _text sometimes.
 227          */
 228         err = perf_event__synthesize_kernel_mmap(tool, process_synthesized_event,
 229                                                  machine);
 230         if (err < 0)
 231                 pr_err("Couldn't record guest kernel [%d]'s reference"
 232                        " relocation symbol.\n", machine->pid);
 233 }
 234
 235 static struct perf_event_header finished_round_event = {
 236         .size = sizeof(struct perf_event_header),
 237         .type = PERF_RECORD_FINISHED_ROUND,
 238 };
 239
 240 static int record__mmap_read_all(struct record *rec)
 241 {
 242         u64 bytes_written = rec->bytes_written;
 243         int i;
 244         int rc = 0;
 245
 246         for (i = 0; i < rec->evlist->nr_mmaps; i++) {
 247                 if (rec->evlist->mmap[i].base) {
 248                         if (record__mmap_read(rec, &rec->evlist->mmap[i]) != 0) {
 249                                 rc = -1;
 250                                 goto out;
 251                         }
 252                 }
 253         }
 254
 255         /*
 256          * Mark the round finished in case we wrote
 257          * at least one event.
 258          */
 259         if (bytes_written != rec->bytes_written)
 260                 rc = record__write(rec, &finished_round_event, sizeof(finished_round_event));
 261
 262 out:
 263         return rc;
 264 }
 265
 266 static void record__init_features(struct record *rec)
 267 {
 268         struct perf_session *session = rec->session;
 269         int feat;
 270
 271         for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++)
 272                 perf_header__set_feat(&session->header, feat);
 273
 274         if (rec->no_buildid)
 275                 perf_header__clear_feat(&session->header, HEADER_BUILD_ID);
 276
 277         if (!have_tracepoints(&rec->evlist->entries))
 278                 perf_header__clear_feat(&session->header, HEADER_TRACING_DATA);
 279
 280         if (!rec->opts.branch_stack)
 281                 perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK);
 282 }
 283
 284 static volatile int workload_exec_errno;
 285
 286 /*
 287  * perf_evlist__prepare_workload will send a SIGUSR1
 288  * if the fork fails, since we asked by setting its
 289  * want_signal to true.
 290  */
 291 static void workload_exec_failed_signal(int signo __maybe_unused,
 292                                         siginfo_t *info,
 293                                         void *ucontext __maybe_unused)
 294 {
 295         workload_exec_errno = info->si_value.sival_int;
 296         done = 1;
 297         child_finished = 1;
 298 }
 299
 300 static int __cmd_record(struct record *rec, int argc, const char **argv)
 301 {
 302         int err;
 303         int status = 0;
 304         unsigned long waking = 0;
 305         const bool forks = argc > 0;
 306         struct machine *machine;
 307         struct perf_tool *tool = &rec->tool;
 308         struct record_opts *opts = &rec->opts;
 309         struct perf_data_file *file = &rec->file;
 310         struct perf_session *session;
 311         bool disabled = false;
 312
 313         rec->progname = argv[0];
 314
 315         atexit(record__sig_exit);
 316         signal(SIGCHLD, sig_handler);
 317         signal(SIGINT, sig_handler);
 318         signal(SIGTERM, sig_handler);
 319
 320         session = perf_session__new(file, false, NULL);
 321         if (session == NULL) {
 322                 pr_err("Perf session creation failed.\n");
 323                 return -1;
 324         }
 325
 326         rec->session = session;
 327
 328         record__init_features(rec);
 329
 330         if (forks) {
 331                 err = perf_evlist__prepare_workload(rec->evlist, &opts->target,
 332                                                     argv, file->is_pipe,
 333                                                     workload_exec_failed_signal);
 334                 if (err < 0) {
 335                         pr_err("Couldn't run the workload!\n");
 336                         status = err;
 337                         goto out_delete_session;
 338                 }
 339         }
 340
 341         if (record__open(rec) != 0) {
 342                 err = -1;
 343                 goto out_child;
 344         }
 345
 346         if (!rec->evlist->nr_groups)
 347                 perf_header__clear_feat(&session->header, HEADER_GROUP_DESC);
 348
 349         if (file->is_pipe) {
 350                 err = perf_header__write_pipe(file->fd);
 351                 if (err < 0)
 352                         goto out_child;
 353         } else {
 354                 err = perf_session__write_header(session, rec->evlist,
 355                                                  file->fd, false);
 356                 if (err < 0)
 357                         goto out_child;
 358         }
 359
 360         if (!rec->no_buildid
 361             && !perf_header__has_feat(&session->header, HEADER_BUILD_ID)) {
 362                 pr_err("Couldn't generate buildids. "
 363                        "Use --no-buildid to profile anyway.\n");
 364                 err = -1;
 365                 goto out_child;
 366         }
 367
 368         machine = &session->machines.host;
 369
 370         if (file->is_pipe) {
 371                 err = perf_event__synthesize_attrs(tool, session,
 372                                                    process_synthesized_event);
 373                 if (err < 0) {
 374                         pr_err("Couldn't synthesize attrs.\n");
 375                         goto out_child;
 376                 }
 377
 378                 if (have_tracepoints(&rec->evlist->entries)) {
 379                         /*
 380                          * FIXME err <= 0 here actually means that
 381                          * there were no tracepoints so its not really
 382                          * an error, just that we don't need to
 383                          * synthesize anything.  We really have to
 384                          * return this more properly and also
 385                          * propagate errors that now are calling die()
 386                          */
 387                         err = perf_event__synthesize_tracing_data(tool, file->fd, rec->evlist,
 388                                                                   process_synthesized_event);
 389                         if (err <= 0) {
 390                                 pr_err("Couldn't record tracing data.\n");
 391                                 goto out_child;
 392                         }
 393                         rec->bytes_written += err;
 394                 }
 395         }
 396
 397         err = perf_event__synthesize_kernel_mmap(tool, process_synthesized_event,
 398                                                  machine);
 399         if (err < 0)
 400                 pr_err("Couldn't record kernel reference relocation symbol\n"
 401                        "Symbol resolution may be skewed if relocation was used (e.g. kexec).\n"
 402                        "Check /proc/kallsyms permission or run as root.\n");
 403
 404         err = perf_event__synthesize_modules(tool, process_synthesized_event,
 405                                              machine);
 406         if (err < 0)
 407                 pr_err("Couldn't record kernel module information.\n"
 408                        "Symbol resolution may be skewed if relocation was used (e.g. kexec).\n"
 409                        "Check /proc/modules permission or run as root.\n");
 410
 411         if (perf_guest) {
 412                 machines__process_guests(&session->machines,
 413                                          perf_event__synthesize_guest_os, tool);
 414         }
 415
 416         err = __machine__synthesize_threads(machine, tool, &opts->target, rec->evlist->threads,
 417                                             process_synthesized_event, opts->sample_address);
 418         if (err != 0)
 419                 goto out_child;
 420
 421         if (rec->realtime_prio) {
 422                 struct sched_param param;
 423
 424                 param.sched_priority = rec->realtime_prio;
 425                 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
 426                         pr_err("Could not set realtime priority.\n");
 427                         err = -1;
 428                         goto out_child;
 429                 }
 430         }
 431
 432         /*
 433          * When perf is starting the traced process, all the events
 434          * (apart from group members) have enable_on_exec=1 set,
 435          * so don't spoil it by prematurely enabling them.
 436          */
 437         if (!target__none(&opts->target) && !opts->initial_delay)
 438                 perf_evlist__enable(rec->evlist);
 439
 440         /*
 441          * Let the child rip
 442          */
 443         if (forks)
 444                 perf_evlist__start_workload(rec->evlist);
 445
 446         if (opts->initial_delay) {
 447                 usleep(opts->initial_delay * 1000);
 448                 perf_evlist__enable(rec->evlist);
 449         }
 450
 451         for (;;) {
 452                 int hits = rec->samples;
 453
 454                 if (record__mmap_read_all(rec) < 0) {
 455                         err = -1;
 456                         goto out_child;
 457                 }
 458
 459                 if (hits == rec->samples) {
 460                         if (done)
 461                                 break;
 462                         err = poll(rec->evlist->pollfd, rec->evlist->nr_fds, -1);
 463                         /*
 464                          * Propagate error, only if there's any. Ignore positive
 465                          * number of returned events and interrupt error.
 466                          */
 467                         if (err > 0 || (err < 0 && errno == EINTR))
 468                                 err = 0;
 469                         waking++;
 470                 }
 471
 472                 /*
 473                  * When perf is starting the traced process, at the end events
 474                  * die with the process and we wait for that. Thus no need to
 475                  * disable events in this case.
 476                  */
 477                 if (done && !disabled && !target__none(&opts->target)) {
 478                         perf_evlist__disable(rec->evlist);
 479                         disabled = true;
 480                 }
 481         }
 482
 483         if (forks && workload_exec_errno) {
 484                 char msg[STRERR_BUFSIZE];
 485                 const char *emsg = strerror_r(workload_exec_errno, msg, sizeof(msg));
 486                 pr_err("Workload failed: %s\n", emsg);
 487                 err = -1;
 488                 goto out_child;
 489         }
 490
 491         if (!quiet) {
 492                 fprintf(stderr, "[ perf record: Woken up %ld times to write data ]\n", waking);
 493
 494                 /*
 495                  * Approximate RIP event size: 24 bytes.
 496                  */
 497                 fprintf(stderr,
 498                         "[ perf record: Captured and wrote %.3f MB %s (~%" PRIu64 " samples) ]\n",
 499                         (double)rec->bytes_written / 1024.0 / 1024.0,
 500                         file->path,
 501                         rec->bytes_written / 24);
 502         }
 503
 504 out_child:
 505         if (forks) {
 506                 int exit_status;
 507
 508                 if (!child_finished)
 509                         kill(rec->evlist->workload.pid, SIGTERM);
 510
 511                 wait(&exit_status);
 512
 513                 if (err < 0)
 514                         status = err;
 515                 else if (WIFEXITED(exit_status))
 516                         status = WEXITSTATUS(exit_status);
 517                 else if (WIFSIGNALED(exit_status))
 518                         signr = WTERMSIG(exit_status);
 519         } else
 520                 status = err;
 521
 522         if (!err && !file->is_pipe) {
 523                 rec->session->header.data_size += rec->bytes_written;
 524
 525                 if (!rec->no_buildid)
 526                         process_buildids(rec);
 527                 perf_session__write_header(rec->session, rec->evlist,
 528                                            file->fd, true);
 529         }
 530
 531 out_delete_session:
 532         perf_session__delete(session);
 533         return status;
 534 }
 535
 536 #define BRANCH_OPT(n, m) \
 537         { .name = n, .mode = (m) }
 538
 539 #define BRANCH_END { .name = NULL }
 540
 541 struct branch_mode {
 542         const char *name;
 543         int mode;
 544 };
 545
 546 static const struct branch_mode branch_modes[] = {
 547         BRANCH_OPT("u", PERF_SAMPLE_BRANCH_USER),
 548         BRANCH_OPT("k", PERF_SAMPLE_BRANCH_KERNEL),
 549         BRANCH_OPT("hv", PERF_SAMPLE_BRANCH_HV),
 550         BRANCH_OPT("any", PERF_SAMPLE_BRANCH_ANY),
 551         BRANCH_OPT("any_call", PERF_SAMPLE_BRANCH_ANY_CALL),
 552         BRANCH_OPT("any_ret", PERF_SAMPLE_BRANCH_ANY_RETURN),
 553         BRANCH_OPT("ind_call", PERF_SAMPLE_BRANCH_IND_CALL),
 554         BRANCH_OPT("abort_tx", PERF_SAMPLE_BRANCH_ABORT_TX),
 555         BRANCH_OPT("in_tx", PERF_SAMPLE_BRANCH_IN_TX),
 556         BRANCH_OPT("no_tx", PERF_SAMPLE_BRANCH_NO_TX),
 557         BRANCH_OPT("cond", PERF_SAMPLE_BRANCH_COND),
 558         BRANCH_END
 559 };
 560
 561 static int
 562 parse_branch_stack(const struct option *opt, const char *str, int unset)
 563 {
 564 #define ONLY_PLM \
 565         (PERF_SAMPLE_BRANCH_USER        |\
 566          PERF_SAMPLE_BRANCH_KERNEL      |\
 567          PERF_SAMPLE_BRANCH_HV)
 568
 569         uint64_t *mode = (uint64_t *)opt->value;
 570         const struct branch_mode *br;
 571         char *s, *os = NULL, *p;
 572         int ret = -1;
 573
 574         if (unset)
 575                 return 0;
 576
 577         /*
 578          * cannot set it twice, -b + --branch-filter for instance
 579          */
 580         if (*mode)
 581                 return -1;
 582
 583         /* str may be NULL in case no arg is passed to -b */
 584         if (str) {
 585                 /* because str is read-only */
 586                 s = os = strdup(str);
 587                 if (!s)
 588                         return -1;
 589
 590                 for (;;) {
 591                         p = strchr(s, ',');
 592                         if (p)
 593                                 *p = '\0';
 594
 595                         for (br = branch_modes; br->name; br++) {
 596                                 if (!strcasecmp(s, br->name))
 597                                         break;
 598                         }
 599                         if (!br->name) {
 600                                 ui__warning("unknown branch filter %s,"
 601                                             " check man page\n", s);
 602                                 goto error;
 603                         }
 604
 605                         *mode |= br->mode;
 606
 607                         if (!p)
 608                                 break;
 609
 610                         s = p + 1;
 611                 }
 612         }
 613         ret = 0;
 614
 615         /* default to any branch */
 616         if ((*mode & ~ONLY_PLM) == 0) {
 617                 *mode = PERF_SAMPLE_BRANCH_ANY;
 618         }
 619 error:
 620         free(os);
 621         return ret;
 622 }
 623
 624 #ifdef HAVE_DWARF_UNWIND_SUPPORT
 625 static int get_stack_size(char *str, unsigned long *_size)
 626 {
 627         char *endptr;
 628         unsigned long size;
 629         unsigned long max_size = round_down(USHRT_MAX, sizeof(u64));
 630
 631         size = strtoul(str, &endptr, 0);
 632
 633         do {
 634                 if (*endptr)
 635                         break;
 636
 637                 size = round_up(size, sizeof(u64));
 638                 if (!size || size > max_size)
 639                         break;
 640
 641                 *_size = size;
 642                 return 0;
 643
 644         } while (0);
 645
 646         pr_err("callchain: Incorrect stack dump size (max %ld): %s\n",
 647                max_size, str);
 648         return -1;
 649 }
 650 #endif /* HAVE_DWARF_UNWIND_SUPPORT */
 651
 652 int record_parse_callchain(const char *arg, struct record_opts *opts)
 653 {
 654         char *tok, *name, *saveptr = NULL;
 655         char *buf;
 656         int ret = -1;
 657
 658         /* We need buffer that we know we can write to. */
 659         buf = malloc(strlen(arg) + 1);
 660         if (!buf)
 661                 return -ENOMEM;
 662
 663         strcpy(buf, arg);
 664
 665         tok = strtok_r((char *)buf, ",", &saveptr);
 666         name = tok ? : (char *)buf;
 667
 668         do {
 669                 /* Framepointer style */
 670                 if (!strncmp(name, "fp", sizeof("fp"))) {
 671                         if (!strtok_r(NULL, ",", &saveptr)) {
 672                                 opts->call_graph = CALLCHAIN_FP;
 673                                 ret = 0;
 674                         } else
 675                                 pr_err("callchain: No more arguments "
 676                                        "needed for -g fp\n");
 677                         break;
 678
 679 #ifdef HAVE_DWARF_UNWIND_SUPPORT
 680                 /* Dwarf style */
 681                 } else if (!strncmp(name, "dwarf", sizeof("dwarf"))) {
 682                         const unsigned long default_stack_dump_size = 8192;
 683
 684                         ret = 0;
 685                         opts->call_graph = CALLCHAIN_DWARF;
 686                         opts->stack_dump_size = default_stack_dump_size;
 687
 688                         tok = strtok_r(NULL, ",", &saveptr);
 689                         if (tok) {
 690                                 unsigned long size = 0;
 691
 692                                 ret = get_stack_size(tok, &size);
 693                                 opts->stack_dump_size = size;
 694                         }
 695 #endif /* HAVE_DWARF_UNWIND_SUPPORT */
 696                 } else {
 697                         pr_err("callchain: Unknown --call-graph option "
 698                                "value: %s\n", arg);
 699                         break;
 700                 }
 701
 702         } while (0);
 703
 704         free(buf);
 705         return ret;
 706 }
 707
 708 static void callchain_debug(struct record_opts *opts)
 709 {
 710         static const char *str[CALLCHAIN_MAX] = { "NONE", "FP", "DWARF" };
 711
 712         pr_debug("callchain: type %s\n", str[opts->call_graph]);
 713
 714         if (opts->call_graph == CALLCHAIN_DWARF)
 715                 pr_debug("callchain: stack dump size %d\n",
 716                          opts->stack_dump_size);
 717 }
 718
 719 int record_parse_callchain_opt(const struct option *opt,
 720                                const char *arg,
 721                                int unset)
 722 {
 723         struct record_opts *opts = opt->value;
 724         int ret;
 725
 726         opts->call_graph_enabled = !unset;
 727
 728         /* --no-call-graph */
 729         if (unset) {
 730                 opts->call_graph = CALLCHAIN_NONE;
 731                 pr_debug("callchain: disabled\n");
 732                 return 0;
 733         }
 734
 735         ret = record_parse_callchain(arg, opts);
 736         if (!ret)
 737                 callchain_debug(opts);
 738
 739         return ret;
 740 }
 741
 742 int record_callchain_opt(const struct option *opt,
 743                          const char *arg __maybe_unused,
 744                          int unset __maybe_unused)
 745 {
 746         struct record_opts *opts = opt->value;
 747
 748         opts->call_graph_enabled = !unset;
 749
 750         if (opts->call_graph == CALLCHAIN_NONE)
 751                 opts->call_graph = CALLCHAIN_FP;
 752
 753         callchain_debug(opts);
 754         return 0;
 755 }
 756
 757 static int perf_record_config(const char *var, const char *value, void *cb)
 758 {
 759         struct record *rec = cb;
 760
 761         if (!strcmp(var, "record.call-graph"))
 762                 return record_parse_callchain(value, &rec->opts);
 763
 764         return perf_default_config(var, value, cb);
 765 }
 766
 767 static const char * const record_usage[] = {
 768         "perf record [<options>] [<command>]",
 769         "perf record [<options>] -- <command> [<options>]",
 770         NULL
 771 };
 772
 773 /*
 774  * XXX Ideally would be local to cmd_record() and passed to a record__new
 775  * because we need to have access to it in record__exit, that is called
 776  * after cmd_record() exits, but since record_options need to be accessible to
 777  * builtin-script, leave it here.
 778  *
 779  * At least we don't ouch it in all the other functions here directly.
 780  *
 781  * Just say no to tons of global variables, sigh.
 782  */
 783 static struct record record = {
 784         .opts = {
 785                 .sample_time         = true,
 786                 .mmap_pages          = UINT_MAX,
 787                 .user_freq           = UINT_MAX,
 788                 .user_interval       = ULLONG_MAX,
 789                 .freq                = 4000,
 790                 .target              = {
 791                         .uses_mmap   = true,
 792                         .default_per_cpu = true,
 793                 },
 794         },
 795 };
 796
 797 #define CALLCHAIN_HELP "setup and enables call-graph (stack chain/backtrace) recording: "
 798
 799 #ifdef HAVE_DWARF_UNWIND_SUPPORT
 800 const char record_callchain_help[] = CALLCHAIN_HELP "fp dwarf";
 801 #else
 802 const char record_callchain_help[] = CALLCHAIN_HELP "fp";
 803 #endif
 804
 805 /*
 806  * XXX Will stay a global variable till we fix builtin-script.c to stop messing
 807  * with it and switch to use the library functions in perf_evlist that came
 808  * from builtin-record.c, i.e. use record_opts,
 809  * perf_evlist__prepare_workload, etc instead of fork+exec'in 'perf record',
 810  * using pipes, etc.
 811  */
 812 const struct option record_options[] = {
 813         OPT_CALLBACK('e', "event", &record.evlist, "event",
 814                      "event selector. use 'perf list' to list available events",
 815                      parse_events_option),
 816         OPT_CALLBACK(0, "filter", &record.evlist, "filter",
 817                      "event filter", parse_filter),
 818         OPT_STRING('p', "pid", &record.opts.target.pid, "pid",
 819                     "record events on existing process id"),
 820         OPT_STRING('t', "tid", &record.opts.target.tid, "tid",
 821                     "record events on existing thread id"),
 822         OPT_INTEGER('r', "realtime", &record.realtime_prio,
 823                     "collect data with this RT SCHED_FIFO priority"),
 824         OPT_BOOLEAN(0, "no-buffering", &record.opts.no_buffering,
 825                     "collect data without buffering"),
 826         OPT_BOOLEAN('R', "raw-samples", &record.opts.raw_samples,
 827                     "collect raw sample records from all opened counters"),
 828         OPT_BOOLEAN('a', "all-cpus", &record.opts.target.system_wide,
 829                             "system-wide collection from all CPUs"),
 830         OPT_STRING('C', "cpu", &record.opts.target.cpu_list, "cpu",
 831                     "list of cpus to monitor"),
 832         OPT_U64('c', "count", &record.opts.user_interval, "event period to sample"),
 833         OPT_STRING('o', "output", &record.file.path, "file",
 834                     "output file name"),
 835         OPT_BOOLEAN_SET('i', "no-inherit", &record.opts.no_inherit,
 836                         &record.opts.no_inherit_set,
 837                         "child tasks do not inherit counters"),
 838         OPT_UINTEGER('F', "freq", &record.opts.user_freq, "profile at this frequency"),
 839         OPT_CALLBACK('m', "mmap-pages", &record.opts.mmap_pages, "pages",
 840                      "number of mmap data pages",
 841                      perf_evlist__parse_mmap_pages),
 842         OPT_BOOLEAN(0, "group", &record.opts.group,
 843                     "put the counters into a counter group"),
 844         OPT_CALLBACK_NOOPT('g', NULL, &record.opts,
 845                            NULL, "enables call-graph recording" ,
 846                            &record_callchain_opt),
 847         OPT_CALLBACK(0, "call-graph", &record.opts,
 848                      "mode[,dump_size]", record_callchain_help,
 849                      &record_parse_callchain_opt),
 850         OPT_INCR('v', "verbose", &verbose,
 851                     "be more verbose (show counter open errors, etc)"),
 852         OPT_BOOLEAN('q', "quiet", &quiet, "don't print any message"),
 853         OPT_BOOLEAN('s', "stat", &record.opts.inherit_stat,
 854                     "per thread counts"),
 855         OPT_BOOLEAN('d', "data", &record.opts.sample_address,
 856                     "Sample addresses"),
 857         OPT_BOOLEAN('T', "timestamp", &record.opts.sample_time, "Sample timestamps"),
 858         OPT_BOOLEAN('P', "period", &record.opts.period, "Sample period"),
 859         OPT_BOOLEAN('n', "no-samples", &record.opts.no_samples,
 860                     "don't sample"),
 861         OPT_BOOLEAN('N', "no-buildid-cache", &record.no_buildid_cache,
 862                     "do not update the buildid cache"),
 863         OPT_BOOLEAN('B', "no-buildid", &record.no_buildid,
 864                     "do not collect buildids in perf.data"),
 865         OPT_CALLBACK('G', "cgroup", &record.evlist, "name",
 866                      "monitor event in cgroup name only",
 867                      parse_cgroups),
 868         OPT_UINTEGER('D', "delay", &record.opts.initial_delay,
 869                   "ms to wait before starting measurement after program start"),
 870         OPT_STRING('u', "uid", &record.opts.target.uid_str, "user",
 871                    "user to profile"),
 872
 873         OPT_CALLBACK_NOOPT('b', "branch-any", &record.opts.branch_stack,
 874                      "branch any", "sample any taken branches",
 875                      parse_branch_stack),
 876
 877         OPT_CALLBACK('j', "branch-filter", &record.opts.branch_stack,
 878                      "branch filter mask", "branch stack filter modes",
 879                      parse_branch_stack),
 880         OPT_BOOLEAN('W', "weight", &record.opts.sample_weight,
 881                     "sample by weight (on special events only)"),
 882         OPT_BOOLEAN(0, "transaction", &record.opts.sample_transaction,
 883                     "sample transaction flags (special events only)"),
 884         OPT_BOOLEAN(0, "per-thread", &record.opts.target.per_thread,
 885                     "use per-thread mmaps"),
 886         OPT_END()
 887 };
 888
 889 int cmd_record(int argc, const char **argv, const char *prefix __maybe_unused)
 890 {
 891         int err = -ENOMEM;
 892         struct record *rec = &record;
 893         char errbuf[BUFSIZ];
 894
 895         rec->evlist = perf_evlist__new();
 896         if (rec->evlist == NULL)
 897                 return -ENOMEM;
 898
 899         perf_config(perf_record_config, rec);
 900
 901         argc = parse_options(argc, argv, record_options, record_usage,
 902                             PARSE_OPT_STOP_AT_NON_OPTION);
 903         if (!argc && target__none(&rec->opts.target))
 904                 usage_with_options(record_usage, record_options);
 905
 906         if (nr_cgroups && !rec->opts.target.system_wide) {
 907                 ui__error("cgroup monitoring only available in"
 908                           " system-wide mode\n");
 909                 usage_with_options(record_usage, record_options);
 910         }
 911
 912         symbol__init(NULL);
 913
 914         if (symbol_conf.kptr_restrict)
 915                 pr_warning(
 916 "WARNING: Kernel address maps (/proc/{kallsyms,modules}) are restricted,\n"
 917 "check /proc/sys/kernel/kptr_restrict.\n\n"
 918 "Samples in kernel functions may not be resolved if a suitable vmlinux\n"
 919 "file is not found in the buildid cache or in the vmlinux path.\n\n"
 920 "Samples in kernel modules won't be resolved at all.\n\n"
 921 "If some relocation was applied (e.g. kexec) symbols may be misresolved\n"
 922 "even with a suitable vmlinux or kallsyms file.\n\n");
 923
 924         if (rec->no_buildid_cache || rec->no_buildid)
 925                 disable_buildid_cache();
 926
 927         if (rec->evlist->nr_entries == 0 &&
 928             perf_evlist__add_default(rec->evlist) < 0) {
 929                 pr_err("Not enough memory for event selector list\n");
 930                 goto out_symbol_exit;
 931         }
 932
 933         if (rec->opts.target.tid && !rec->opts.no_inherit_set)
 934                 rec->opts.no_inherit = true;
 935
 936         err = target__validate(&rec->opts.target);
 937         if (err) {
 938                 target__strerror(&rec->opts.target, err, errbuf, BUFSIZ);
 939                 ui__warning("%s", errbuf);
 940         }
 941
 942         err = target__parse_uid(&rec->opts.target);
 943         if (err) {
 944                 int saved_errno = errno;
 945
 946                 target__strerror(&rec->opts.target, err, errbuf, BUFSIZ);
 947                 ui__error("%s", errbuf);
 948
 949                 err = -saved_errno;
 950                 goto out_symbol_exit;
 951         }
 952
 953         err = -ENOMEM;
 954         if (perf_evlist__create_maps(rec->evlist, &rec->opts.target) < 0)
 955                 usage_with_options(record_usage, record_options);
 956
 957         if (record_opts__config(&rec->opts)) {
 958                 err = -EINVAL;
 959                 goto out_symbol_exit;
 960         }
 961
 962         err = __cmd_record(&record, argc, argv);
 963 out_symbol_exit:
 964         perf_evlist__delete(rec->evlist);
 965         symbol__exit();
 966         return err;
 967 }