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/callchain.h"
  18 #include "util/cgroup.h"
  19 #include "util/header.h"
  20 #include "util/event.h"
  21 #include "util/evlist.h"
  22 #include "util/evsel.h"
  23 #include "util/debug.h"
  24 #include "util/session.h"
  25 #include "util/tool.h"
  26 #include "util/symbol.h"
  27 #include "util/cpumap.h"
  28 #include "util/thread_map.h"
  29 #include "util/data.h"
  30
  31 #include <unistd.h>
  32 #include <sched.h>
  33 #include <sys/mman.h>
  34
  35
  36 struct record {
  37         struct perf_tool        tool;
  38         struct record_opts      opts;
  39         u64                     bytes_written;
  40         struct perf_data_file   file;
  41         struct perf_evlist      *evlist;
  42         struct perf_session     *session;
  43         const char              *progname;
  44         int                     realtime_prio;
  45         bool                    no_buildid;
  46         bool                    no_buildid_cache;
  47         long                    samples;
  48 };
  49
  50 static int record__write(struct record *rec, void *bf, size_t size)
  51 {
  52         if (perf_data_file__write(rec->session->file, bf, size) < 0) {
  53                 pr_err("failed to write perf data, error: %m\n");
  54                 return -1;
  55         }
  56
  57         rec->bytes_written += size;
  58         return 0;
  59 }
  60
  61 static int process_synthesized_event(struct perf_tool *tool,
  62                                      union perf_event *event,
  63                                      struct perf_sample *sample __maybe_unused,
  64                                      struct machine *machine __maybe_unused)
  65 {
  66         struct record *rec = container_of(tool, struct record, tool);
  67         return record__write(rec, event, event->header.size);
  68 }
  69
  70 static int record__mmap_read(struct record *rec, int idx)
  71 {
  72         struct perf_mmap *md = &rec->evlist->mmap[idx];
  73         unsigned int head = perf_mmap__read_head(md);
  74         unsigned int old = md->prev;
  75         unsigned char *data = md->base + page_size;
  76         unsigned long size;
  77         void *buf;
  78         int rc = 0;
  79
  80         if (old == head)
  81                 return 0;
  82
  83         rec->samples++;
  84
  85         size = head - old;
  86
  87         if ((old & md->mask) + size != (head & md->mask)) {
  88                 buf = &data[old & md->mask];
  89                 size = md->mask + 1 - (old & md->mask);
  90                 old += size;
  91
  92                 if (record__write(rec, buf, size) < 0) {
  93                         rc = -1;
  94                         goto out;
  95                 }
  96         }
  97
  98         buf = &data[old & md->mask];
  99         size = head - old;
 100         old += size;
 101
 102         if (record__write(rec, buf, size) < 0) {
 103                 rc = -1;
 104                 goto out;
 105         }
 106
 107         md->prev = old;
 108         perf_evlist__mmap_consume(rec->evlist, idx);
 109 out:
 110         return rc;
 111 }
 112
 113 static volatile int done = 0;
 114 static volatile int signr = -1;
 115 static volatile int child_finished = 0;
 116
 117 static void sig_handler(int sig)
 118 {
 119         if (sig == SIGCHLD)
 120                 child_finished = 1;
 121         else
 122                 signr = sig;
 123
 124         done = 1;
 125 }
 126
 127 static void record__sig_exit(void)
 128 {
 129         if (signr == -1)
 130                 return;
 131
 132         signal(signr, SIG_DFL);
 133         raise(signr);
 134 }
 135
 136 static int record__open(struct record *rec)
 137 {
 138         char msg[512];
 139         struct perf_evsel *pos;
 140         struct perf_evlist *evlist = rec->evlist;
 141         struct perf_session *session = rec->session;
 142         struct record_opts *opts = &rec->opts;
 143         int rc = 0;
 144
 145         perf_evlist__config(evlist, opts);
 146
 147         evlist__for_each(evlist, pos) {
 148 try_again:
 149                 if (perf_evsel__open(pos, evlist->cpus, evlist->threads) < 0) {
 150                         if (perf_evsel__fallback(pos, errno, msg, sizeof(msg))) {
 151                                 if (verbose)
 152                                         ui__warning("%s\n", msg);
 153                                 goto try_again;
 154                         }
 155
 156                         rc = -errno;
 157                         perf_evsel__open_strerror(pos, &opts->target,
 158                                                   errno, msg, sizeof(msg));
 159                         ui__error("%s\n", msg);
 160                         goto out;
 161                 }
 162         }
 163
 164         if (perf_evlist__apply_filters(evlist)) {
 165                 error("failed to set filter with %d (%s)\n", errno,
 166                         strerror_r(errno, msg, sizeof(msg)));
 167                 rc = -1;
 168                 goto out;
 169         }
 170
 171         if (perf_evlist__mmap(evlist, opts->mmap_pages, false) < 0) {
 172                 if (errno == EPERM) {
 173                         pr_err("Permission error mapping pages.\n"
 174                                "Consider increasing "
 175                                "/proc/sys/kernel/perf_event_mlock_kb,\n"
 176                                "or try again with a smaller value of -m/--mmap_pages.\n"
 177                                "(current value: %u)\n", opts->mmap_pages);
 178                         rc = -errno;
 179                 } else {
 180                         pr_err("failed to mmap with %d (%s)\n", errno,
 181                                 strerror_r(errno, msg, sizeof(msg)));
 182                         rc = -errno;
 183                 }
 184                 goto out;
 185         }
 186
 187         session->evlist = evlist;
 188         perf_session__set_id_hdr_size(session);
 189 out:
 190         return rc;
 191 }
 192
 193 static int process_sample_event(struct perf_tool *tool,
 194                                 union perf_event *event,
 195                                 struct perf_sample *sample,
 196                                 struct perf_evsel *evsel,
 197                                 struct machine *machine)
 198 {
 199         struct record *rec = container_of(tool, struct record, tool);
 200
 201         rec->samples++;
 202
 203         return build_id__mark_dso_hit(tool, event, sample, evsel, machine);
 204 }
 205
 206 static int process_buildids(struct record *rec)
 207 {
 208         struct perf_data_file *file  = &rec->file;
 209         struct perf_session *session = rec->session;
 210
 211         u64 size = lseek(perf_data_file__fd(file), 0, SEEK_CUR);
 212         if (size == 0)
 213                 return 0;
 214
 215         file->size = size;
 216
 217         /*
 218          * During this process, it'll load kernel map and replace the
 219          * dso->long_name to a real pathname it found.  In this case
 220          * we prefer the vmlinux path like
 221          *   /lib/modules/3.16.4/build/vmlinux
 222          *
 223          * rather than build-id path (in debug directory).
 224          *   $HOME/.debug/.build-id/f0/6e17aa50adf4d00b88925e03775de107611551
 225          */
 226         symbol_conf.ignore_vmlinux_buildid = true;
 227
 228         return perf_session__process_events(session, &rec->tool);
 229 }
 230
 231 static void perf_event__synthesize_guest_os(struct machine *machine, void *data)
 232 {
 233         int err;
 234         struct perf_tool *tool = data;
 235         /*
 236          *As for guest kernel when processing subcommand record&report,
 237          *we arrange module mmap prior to guest kernel mmap and trigger
 238          *a preload dso because default guest module symbols are loaded
 239          *from guest kallsyms instead of /lib/modules/XXX/XXX. This
 240          *method is used to avoid symbol missing when the first addr is
 241          *in module instead of in guest kernel.
 242          */
 243         err = perf_event__synthesize_modules(tool, process_synthesized_event,
 244                                              machine);
 245         if (err < 0)
 246                 pr_err("Couldn't record guest kernel [%d]'s reference"
 247                        " relocation symbol.\n", machine->pid);
 248
 249         /*
 250          * We use _stext for guest kernel because guest kernel's /proc/kallsyms
 251          * have no _text sometimes.
 252          */
 253         err = perf_event__synthesize_kernel_mmap(tool, process_synthesized_event,
 254                                                  machine);
 255         if (err < 0)
 256                 pr_err("Couldn't record guest kernel [%d]'s reference"
 257                        " relocation symbol.\n", machine->pid);
 258 }
 259
 260 static struct perf_event_header finished_round_event = {
 261         .size = sizeof(struct perf_event_header),
 262         .type = PERF_RECORD_FINISHED_ROUND,
 263 };
 264
 265 static int record__mmap_read_all(struct record *rec)
 266 {
 267         u64 bytes_written = rec->bytes_written;
 268         int i;
 269         int rc = 0;
 270
 271         for (i = 0; i < rec->evlist->nr_mmaps; i++) {
 272                 if (rec->evlist->mmap[i].base) {
 273                         if (record__mmap_read(rec, i) != 0) {
 274                                 rc = -1;
 275                                 goto out;
 276                         }
 277                 }
 278         }
 279
 280         /*
 281          * Mark the round finished in case we wrote
 282          * at least one event.
 283          */
 284         if (bytes_written != rec->bytes_written)
 285                 rc = record__write(rec, &finished_round_event, sizeof(finished_round_event));
 286
 287 out:
 288         return rc;
 289 }
 290
 291 static void record__init_features(struct record *rec)
 292 {
 293         struct perf_session *session = rec->session;
 294         int feat;
 295
 296         for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++)
 297                 perf_header__set_feat(&session->header, feat);
 298
 299         if (rec->no_buildid)
 300                 perf_header__clear_feat(&session->header, HEADER_BUILD_ID);
 301
 302         if (!have_tracepoints(&rec->evlist->entries))
 303                 perf_header__clear_feat(&session->header, HEADER_TRACING_DATA);
 304
 305         if (!rec->opts.branch_stack)
 306                 perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK);
 307 }
 308
 309 static volatile int workload_exec_errno;
 310
 311 /*
 312  * perf_evlist__prepare_workload will send a SIGUSR1
 313  * if the fork fails, since we asked by setting its
 314  * want_signal to true.
 315  */
 316 static void workload_exec_failed_signal(int signo __maybe_unused,
 317                                         siginfo_t *info,
 318                                         void *ucontext __maybe_unused)
 319 {
 320         workload_exec_errno = info->si_value.sival_int;
 321         done = 1;
 322         child_finished = 1;
 323 }
 324
 325 static int __cmd_record(struct record *rec, int argc, const char **argv)
 326 {
 327         int err;
 328         int status = 0;
 329         unsigned long waking = 0;
 330         const bool forks = argc > 0;
 331         struct machine *machine;
 332         struct perf_tool *tool = &rec->tool;
 333         struct record_opts *opts = &rec->opts;
 334         struct perf_data_file *file = &rec->file;
 335         struct perf_session *session;
 336         bool disabled = false, draining = false;
 337         int fd;
 338
 339         rec->progname = argv[0];
 340
 341         atexit(record__sig_exit);
 342         signal(SIGCHLD, sig_handler);
 343         signal(SIGINT, sig_handler);
 344         signal(SIGTERM, sig_handler);
 345
 346         session = perf_session__new(file, false, NULL);
 347         if (session == NULL) {
 348                 pr_err("Perf session creation failed.\n");
 349                 return -1;
 350         }
 351
 352         fd = perf_data_file__fd(file);
 353         rec->session = session;
 354
 355         record__init_features(rec);
 356
 357         if (forks) {
 358                 err = perf_evlist__prepare_workload(rec->evlist, &opts->target,
 359                                                     argv, file->is_pipe,
 360                                                     workload_exec_failed_signal);
 361                 if (err < 0) {
 362                         pr_err("Couldn't run the workload!\n");
 363                         status = err;
 364                         goto out_delete_session;
 365                 }
 366         }
 367
 368         if (record__open(rec) != 0) {
 369                 err = -1;
 370                 goto out_child;
 371         }
 372
 373         if (!rec->evlist->nr_groups)
 374                 perf_header__clear_feat(&session->header, HEADER_GROUP_DESC);
 375
 376         if (file->is_pipe) {
 377                 err = perf_header__write_pipe(fd);
 378                 if (err < 0)
 379                         goto out_child;
 380         } else {
 381                 err = perf_session__write_header(session, rec->evlist, fd, false);
 382                 if (err < 0)
 383                         goto out_child;
 384         }
 385
 386         if (!rec->no_buildid
 387             && !perf_header__has_feat(&session->header, HEADER_BUILD_ID)) {
 388                 pr_err("Couldn't generate buildids. "
 389                        "Use --no-buildid to profile anyway.\n");
 390                 err = -1;
 391                 goto out_child;
 392         }
 393
 394         machine = &session->machines.host;
 395
 396         if (file->is_pipe) {
 397                 err = perf_event__synthesize_attrs(tool, session,
 398                                                    process_synthesized_event);
 399                 if (err < 0) {
 400                         pr_err("Couldn't synthesize attrs.\n");
 401                         goto out_child;
 402                 }
 403
 404                 if (have_tracepoints(&rec->evlist->entries)) {
 405                         /*
 406                          * FIXME err <= 0 here actually means that
 407                          * there were no tracepoints so its not really
 408                          * an error, just that we don't need to
 409                          * synthesize anything.  We really have to
 410                          * return this more properly and also
 411                          * propagate errors that now are calling die()
 412                          */
 413                         err = perf_event__synthesize_tracing_data(tool, fd, rec->evlist,
 414                                                                   process_synthesized_event);
 415                         if (err <= 0) {
 416                                 pr_err("Couldn't record tracing data.\n");
 417                                 goto out_child;
 418                         }
 419                         rec->bytes_written += err;
 420                 }
 421         }
 422
 423         err = perf_event__synthesize_kernel_mmap(tool, process_synthesized_event,
 424                                                  machine);
 425         if (err < 0)
 426                 pr_err("Couldn't record kernel reference relocation symbol\n"
 427                        "Symbol resolution may be skewed if relocation was used (e.g. kexec).\n"
 428                        "Check /proc/kallsyms permission or run as root.\n");
 429
 430         err = perf_event__synthesize_modules(tool, process_synthesized_event,
 431                                              machine);
 432         if (err < 0)
 433                 pr_err("Couldn't record kernel module information.\n"
 434                        "Symbol resolution may be skewed if relocation was used (e.g. kexec).\n"
 435                        "Check /proc/modules permission or run as root.\n");
 436
 437         if (perf_guest) {
 438                 machines__process_guests(&session->machines,
 439                                          perf_event__synthesize_guest_os, tool);
 440         }
 441
 442         err = __machine__synthesize_threads(machine, tool, &opts->target, rec->evlist->threads,
 443                                             process_synthesized_event, opts->sample_address);
 444         if (err != 0)
 445                 goto out_child;
 446
 447         if (rec->realtime_prio) {
 448                 struct sched_param param;
 449
 450                 param.sched_priority = rec->realtime_prio;
 451                 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
 452                         pr_err("Could not set realtime priority.\n");
 453                         err = -1;
 454                         goto out_child;
 455                 }
 456         }
 457
 458         /*
 459          * When perf is starting the traced process, all the events
 460          * (apart from group members) have enable_on_exec=1 set,
 461          * so don't spoil it by prematurely enabling them.
 462          */
 463         if (!target__none(&opts->target) && !opts->initial_delay)
 464                 perf_evlist__enable(rec->evlist);
 465
 466         /*
 467          * Let the child rip
 468          */
 469         if (forks)
 470                 perf_evlist__start_workload(rec->evlist);
 471
 472         if (opts->initial_delay) {
 473                 usleep(opts->initial_delay * 1000);
 474                 perf_evlist__enable(rec->evlist);
 475         }
 476
 477         for (;;) {
 478                 int hits = rec->samples;
 479
 480                 if (record__mmap_read_all(rec) < 0) {
 481                         err = -1;
 482                         goto out_child;
 483                 }
 484
 485                 if (hits == rec->samples) {
 486                         if (done || draining)
 487                                 break;
 488                         err = perf_evlist__poll(rec->evlist, -1);
 489                         /*
 490                          * Propagate error, only if there's any. Ignore positive
 491                          * number of returned events and interrupt error.
 492                          */
 493                         if (err > 0 || (err < 0 && errno == EINTR))
 494                                 err = 0;
 495                         waking++;
 496
 497                         if (perf_evlist__filter_pollfd(rec->evlist, POLLERR | POLLHUP) == 0)
 498                                 draining = true;
 499                 }
 500
 501                 /*
 502                  * When perf is starting the traced process, at the end events
 503                  * die with the process and we wait for that. Thus no need to
 504                  * disable events in this case.
 505                  */
 506                 if (done && !disabled && !target__none(&opts->target)) {
 507                         perf_evlist__disable(rec->evlist);
 508                         disabled = true;
 509                 }
 510         }
 511
 512         if (forks && workload_exec_errno) {
 513                 char msg[STRERR_BUFSIZE];
 514                 const char *emsg = strerror_r(workload_exec_errno, msg, sizeof(msg));
 515                 pr_err("Workload failed: %s\n", emsg);
 516                 err = -1;
 517                 goto out_child;
 518         }
 519
 520         if (!quiet)
 521                 fprintf(stderr, "[ perf record: Woken up %ld times to write data ]\n", waking);
 522
 523 out_child:
 524         if (forks) {
 525                 int exit_status;
 526
 527                 if (!child_finished)
 528                         kill(rec->evlist->workload.pid, SIGTERM);
 529
 530                 wait(&exit_status);
 531
 532                 if (err < 0)
 533                         status = err;
 534                 else if (WIFEXITED(exit_status))
 535                         status = WEXITSTATUS(exit_status);
 536                 else if (WIFSIGNALED(exit_status))
 537                         signr = WTERMSIG(exit_status);
 538         } else
 539                 status = err;
 540
 541         /* this will be recalculated during process_buildids() */
 542         rec->samples = 0;
 543
 544         if (!err && !file->is_pipe) {
 545                 rec->session->header.data_size += rec->bytes_written;
 546
 547                 if (!rec->no_buildid)
 548                         process_buildids(rec);
 549                 perf_session__write_header(rec->session, rec->evlist, fd, true);
 550         }
 551
 552         if (!err && !quiet) {
 553                 char samples[128];
 554
 555                 if (rec->samples)
 556                         scnprintf(samples, sizeof(samples),
 557                                   " (%" PRIu64 " samples)", rec->samples);
 558                 else
 559                         samples[0] = '\0';
 560
 561                 fprintf(stderr, "[ perf record: Captured and wrote %.3f MB %s%s ]\n",
 562                         perf_data_file__size(file) / 1024.0 / 1024.0,
 563                         file->path, samples);
 564         }
 565
 566 out_delete_session:
 567         perf_session__delete(session);
 568         return status;
 569 }
 570
 571 #define BRANCH_OPT(n, m) \
 572         { .name = n, .mode = (m) }
 573
 574 #define BRANCH_END { .name = NULL }
 575
 576 struct branch_mode {
 577         const char *name;
 578         int mode;
 579 };
 580
 581 static const struct branch_mode branch_modes[] = {
 582         BRANCH_OPT("u", PERF_SAMPLE_BRANCH_USER),
 583         BRANCH_OPT("k", PERF_SAMPLE_BRANCH_KERNEL),
 584         BRANCH_OPT("hv", PERF_SAMPLE_BRANCH_HV),
 585         BRANCH_OPT("any", PERF_SAMPLE_BRANCH_ANY),
 586         BRANCH_OPT("any_call", PERF_SAMPLE_BRANCH_ANY_CALL),
 587         BRANCH_OPT("any_ret", PERF_SAMPLE_BRANCH_ANY_RETURN),
 588         BRANCH_OPT("ind_call", PERF_SAMPLE_BRANCH_IND_CALL),
 589         BRANCH_OPT("abort_tx", PERF_SAMPLE_BRANCH_ABORT_TX),
 590         BRANCH_OPT("in_tx", PERF_SAMPLE_BRANCH_IN_TX),
 591         BRANCH_OPT("no_tx", PERF_SAMPLE_BRANCH_NO_TX),
 592         BRANCH_OPT("cond", PERF_SAMPLE_BRANCH_COND),
 593         BRANCH_END
 594 };
 595
 596 static int
 597 parse_branch_stack(const struct option *opt, const char *str, int unset)
 598 {
 599 #define ONLY_PLM \
 600         (PERF_SAMPLE_BRANCH_USER        |\
 601          PERF_SAMPLE_BRANCH_KERNEL      |\
 602          PERF_SAMPLE_BRANCH_HV)
 603
 604         uint64_t *mode = (uint64_t *)opt->value;
 605         const struct branch_mode *br;
 606         char *s, *os = NULL, *p;
 607         int ret = -1;
 608
 609         if (unset)
 610                 return 0;
 611
 612         /*
 613          * cannot set it twice, -b + --branch-filter for instance
 614          */
 615         if (*mode)
 616                 return -1;
 617
 618         /* str may be NULL in case no arg is passed to -b */
 619         if (str) {
 620                 /* because str is read-only */
 621                 s = os = strdup(str);
 622                 if (!s)
 623                         return -1;
 624
 625                 for (;;) {
 626                         p = strchr(s, ',');
 627                         if (p)
 628                                 *p = '\0';
 629
 630                         for (br = branch_modes; br->name; br++) {
 631                                 if (!strcasecmp(s, br->name))
 632                                         break;
 633                         }
 634                         if (!br->name) {
 635                                 ui__warning("unknown branch filter %s,"
 636                                             " check man page\n", s);
 637                                 goto error;
 638                         }
 639
 640                         *mode |= br->mode;
 641
 642                         if (!p)
 643                                 break;
 644
 645                         s = p + 1;
 646                 }
 647         }
 648         ret = 0;
 649
 650         /* default to any branch */
 651         if ((*mode & ~ONLY_PLM) == 0) {
 652                 *mode = PERF_SAMPLE_BRANCH_ANY;
 653         }
 654 error:
 655         free(os);
 656         return ret;
 657 }
 658
 659 static void callchain_debug(void)
 660 {
 661         static const char *str[CALLCHAIN_MAX] = { "NONE", "FP", "DWARF" };
 662
 663         pr_debug("callchain: type %s\n", str[callchain_param.record_mode]);
 664
 665         if (callchain_param.record_mode == CALLCHAIN_DWARF)
 666                 pr_debug("callchain: stack dump size %d\n",
 667                          callchain_param.dump_size);
 668 }
 669
 670 int record_parse_callchain_opt(const struct option *opt __maybe_unused,
 671                                const char *arg,
 672                                int unset)
 673 {
 674         int ret;
 675
 676         callchain_param.enabled = !unset;
 677
 678         /* --no-call-graph */
 679         if (unset) {
 680                 callchain_param.record_mode = CALLCHAIN_NONE;
 681                 pr_debug("callchain: disabled\n");
 682                 return 0;
 683         }
 684
 685         ret = parse_callchain_record_opt(arg);
 686         if (!ret)
 687                 callchain_debug();
 688
 689         return ret;
 690 }
 691
 692 int record_callchain_opt(const struct option *opt __maybe_unused,
 693                          const char *arg __maybe_unused,
 694                          int unset __maybe_unused)
 695 {
 696         callchain_param.enabled = true;
 697
 698         if (callchain_param.record_mode == CALLCHAIN_NONE)
 699                 callchain_param.record_mode = CALLCHAIN_FP;
 700
 701         callchain_debug();
 702         return 0;
 703 }
 704
 705 static int perf_record_config(const char *var, const char *value, void *cb)
 706 {
 707         if (!strcmp(var, "record.call-graph"))
 708                 var = "call-graph.record-mode"; /* fall-through */
 709
 710         return perf_default_config(var, value, cb);
 711 }
 712
 713 static const char * const __record_usage[] = {
 714         "perf record [<options>] [<command>]",
 715         "perf record [<options>] -- <command> [<options>]",
 716         NULL
 717 };
 718 const char * const *record_usage = __record_usage;
 719
 720 /*
 721  * XXX Ideally would be local to cmd_record() and passed to a record__new
 722  * because we need to have access to it in record__exit, that is called
 723  * after cmd_record() exits, but since record_options need to be accessible to
 724  * builtin-script, leave it here.
 725  *
 726  * At least we don't ouch it in all the other functions here directly.
 727  *
 728  * Just say no to tons of global variables, sigh.
 729  */
 730 static struct record record = {
 731         .opts = {
 732                 .sample_time         = true,
 733                 .mmap_pages          = UINT_MAX,
 734                 .user_freq           = UINT_MAX,
 735                 .user_interval       = ULLONG_MAX,
 736                 .freq                = 4000,
 737                 .target              = {
 738                         .uses_mmap   = true,
 739                         .default_per_cpu = true,
 740                 },
 741         },
 742         .tool = {
 743                 .sample         = process_sample_event,
 744                 .fork           = perf_event__process_fork,
 745                 .comm           = perf_event__process_comm,
 746                 .mmap           = perf_event__process_mmap,
 747                 .mmap2          = perf_event__process_mmap2,
 748         },
 749 };
 750
 751 #define CALLCHAIN_HELP "setup and enables call-graph (stack chain/backtrace) recording: "
 752
 753 #ifdef HAVE_DWARF_UNWIND_SUPPORT
 754 const char record_callchain_help[] = CALLCHAIN_HELP "fp dwarf";
 755 #else
 756 const char record_callchain_help[] = CALLCHAIN_HELP "fp";
 757 #endif
 758
 759 /*
 760  * XXX Will stay a global variable till we fix builtin-script.c to stop messing
 761  * with it and switch to use the library functions in perf_evlist that came
 762  * from builtin-record.c, i.e. use record_opts,
 763  * perf_evlist__prepare_workload, etc instead of fork+exec'in 'perf record',
 764  * using pipes, etc.
 765  */
 766 struct option __record_options[] = {
 767         OPT_CALLBACK('e', "event", &record.evlist, "event",
 768                      "event selector. use 'perf list' to list available events",
 769                      parse_events_option),
 770         OPT_CALLBACK(0, "filter", &record.evlist, "filter",
 771                      "event filter", parse_filter),
 772         OPT_STRING('p', "pid", &record.opts.target.pid, "pid",
 773                     "record events on existing process id"),
 774         OPT_STRING('t', "tid", &record.opts.target.tid, "tid",
 775                     "record events on existing thread id"),
 776         OPT_INTEGER('r', "realtime", &record.realtime_prio,
 777                     "collect data with this RT SCHED_FIFO priority"),
 778         OPT_BOOLEAN(0, "no-buffering", &record.opts.no_buffering,
 779                     "collect data without buffering"),
 780         OPT_BOOLEAN('R', "raw-samples", &record.opts.raw_samples,
 781                     "collect raw sample records from all opened counters"),
 782         OPT_BOOLEAN('a', "all-cpus", &record.opts.target.system_wide,
 783                             "system-wide collection from all CPUs"),
 784         OPT_STRING('C', "cpu", &record.opts.target.cpu_list, "cpu",
 785                     "list of cpus to monitor"),
 786         OPT_U64('c', "count", &record.opts.user_interval, "event period to sample"),
 787         OPT_STRING('o', "output", &record.file.path, "file",
 788                     "output file name"),
 789         OPT_BOOLEAN_SET('i', "no-inherit", &record.opts.no_inherit,
 790                         &record.opts.no_inherit_set,
 791                         "child tasks do not inherit counters"),
 792         OPT_UINTEGER('F', "freq", &record.opts.user_freq, "profile at this frequency"),
 793         OPT_CALLBACK('m', "mmap-pages", &record.opts.mmap_pages, "pages",
 794                      "number of mmap data pages",
 795                      perf_evlist__parse_mmap_pages),
 796         OPT_BOOLEAN(0, "group", &record.opts.group,
 797                     "put the counters into a counter group"),
 798         OPT_CALLBACK_NOOPT('g', NULL, &record.opts,
 799                            NULL, "enables call-graph recording" ,
 800                            &record_callchain_opt),
 801         OPT_CALLBACK(0, "call-graph", &record.opts,
 802                      "mode[,dump_size]", record_callchain_help,
 803                      &record_parse_callchain_opt),
 804         OPT_INCR('v', "verbose", &verbose,
 805                     "be more verbose (show counter open errors, etc)"),
 806         OPT_BOOLEAN('q', "quiet", &quiet, "don't print any message"),
 807         OPT_BOOLEAN('s', "stat", &record.opts.inherit_stat,
 808                     "per thread counts"),
 809         OPT_BOOLEAN('d', "data", &record.opts.sample_address,
 810                     "Sample addresses"),
 811         OPT_BOOLEAN('T', "timestamp", &record.opts.sample_time, "Sample timestamps"),
 812         OPT_BOOLEAN('P', "period", &record.opts.period, "Sample period"),
 813         OPT_BOOLEAN('n', "no-samples", &record.opts.no_samples,
 814                     "don't sample"),
 815         OPT_BOOLEAN('N', "no-buildid-cache", &record.no_buildid_cache,
 816                     "do not update the buildid cache"),
 817         OPT_BOOLEAN('B', "no-buildid", &record.no_buildid,
 818                     "do not collect buildids in perf.data"),
 819         OPT_CALLBACK('G', "cgroup", &record.evlist, "name",
 820                      "monitor event in cgroup name only",
 821                      parse_cgroups),
 822         OPT_UINTEGER('D', "delay", &record.opts.initial_delay,
 823                   "ms to wait before starting measurement after program start"),
 824         OPT_STRING('u', "uid", &record.opts.target.uid_str, "user",
 825                    "user to profile"),
 826
 827         OPT_CALLBACK_NOOPT('b', "branch-any", &record.opts.branch_stack,
 828                      "branch any", "sample any taken branches",
 829                      parse_branch_stack),
 830
 831         OPT_CALLBACK('j', "branch-filter", &record.opts.branch_stack,
 832                      "branch filter mask", "branch stack filter modes",
 833                      parse_branch_stack),
 834         OPT_BOOLEAN('W', "weight", &record.opts.sample_weight,
 835                     "sample by weight (on special events only)"),
 836         OPT_BOOLEAN(0, "transaction", &record.opts.sample_transaction,
 837                     "sample transaction flags (special events only)"),
 838         OPT_BOOLEAN(0, "per-thread", &record.opts.target.per_thread,
 839                     "use per-thread mmaps"),
 840         OPT_BOOLEAN('I', "intr-regs", &record.opts.sample_intr_regs,
 841                     "Sample machine registers on interrupt"),
 842         OPT_END()
 843 };
 844
 845 struct option *record_options = __record_options;
 846
 847 int cmd_record(int argc, const char **argv, const char *prefix __maybe_unused)
 848 {
 849         int err = -ENOMEM;
 850         struct record *rec = &record;
 851         char errbuf[BUFSIZ];
 852
 853         rec->evlist = perf_evlist__new();
 854         if (rec->evlist == NULL)
 855                 return -ENOMEM;
 856
 857         perf_config(perf_record_config, rec);
 858
 859         argc = parse_options(argc, argv, record_options, record_usage,
 860                             PARSE_OPT_STOP_AT_NON_OPTION);
 861         if (!argc && target__none(&rec->opts.target))
 862                 usage_with_options(record_usage, record_options);
 863
 864         if (nr_cgroups && !rec->opts.target.system_wide) {
 865                 ui__error("cgroup monitoring only available in"
 866                           " system-wide mode\n");
 867                 usage_with_options(record_usage, record_options);
 868         }
 869
 870         symbol__init(NULL);
 871
 872         if (symbol_conf.kptr_restrict)
 873                 pr_warning(
 874 "WARNING: Kernel address maps (/proc/{kallsyms,modules}) are restricted,\n"
 875 "check /proc/sys/kernel/kptr_restrict.\n\n"
 876 "Samples in kernel functions may not be resolved if a suitable vmlinux\n"
 877 "file is not found in the buildid cache or in the vmlinux path.\n\n"
 878 "Samples in kernel modules won't be resolved at all.\n\n"
 879 "If some relocation was applied (e.g. kexec) symbols may be misresolved\n"
 880 "even with a suitable vmlinux or kallsyms file.\n\n");
 881
 882         if (rec->no_buildid_cache || rec->no_buildid)
 883                 disable_buildid_cache();
 884
 885         if (rec->evlist->nr_entries == 0 &&
 886             perf_evlist__add_default(rec->evlist) < 0) {
 887                 pr_err("Not enough memory for event selector list\n");
 888                 goto out_symbol_exit;
 889         }
 890
 891         if (rec->opts.target.tid && !rec->opts.no_inherit_set)
 892                 rec->opts.no_inherit = true;
 893
 894         err = target__validate(&rec->opts.target);
 895         if (err) {
 896                 target__strerror(&rec->opts.target, err, errbuf, BUFSIZ);
 897                 ui__warning("%s", errbuf);
 898         }
 899
 900         err = target__parse_uid(&rec->opts.target);
 901         if (err) {
 902                 int saved_errno = errno;
 903
 904                 target__strerror(&rec->opts.target, err, errbuf, BUFSIZ);
 905                 ui__error("%s", errbuf);
 906
 907                 err = -saved_errno;
 908                 goto out_symbol_exit;
 909         }
 910
 911         err = -ENOMEM;
 912         if (perf_evlist__create_maps(rec->evlist, &rec->opts.target) < 0)
 913                 usage_with_options(record_usage, record_options);
 914
 915         if (record_opts__config(&rec->opts)) {
 916                 err = -EINVAL;
 917                 goto out_symbol_exit;
 918         }
 919
 920         err = __cmd_record(&record, argc, argv);
 921 out_symbol_exit:
 922         perf_evlist__delete(rec->evlist);
 923         symbol__exit();
 924         return err;
 925 }