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, &pos)) {
 165                 error("failed to set filter \"%s\" on event %s with %d (%s)\n",
 166                         pos->filter, perf_evsel__name(pos), errno,
 167                         strerror_r(errno, msg, sizeof(msg)));
 168                 rc = -1;
 169                 goto out;
 170         }
 171
 172         if (perf_evlist__mmap(evlist, opts->mmap_pages, false) < 0) {
 173                 if (errno == EPERM) {
 174                         pr_err("Permission error mapping pages.\n"
 175                                "Consider increasing "
 176                                "/proc/sys/kernel/perf_event_mlock_kb,\n"
 177                                "or try again with a smaller value of -m/--mmap_pages.\n"
 178                                "(current value: %u)\n", opts->mmap_pages);
 179                         rc = -errno;
 180                 } else {
 181                         pr_err("failed to mmap with %d (%s)\n", errno,
 182                                 strerror_r(errno, msg, sizeof(msg)));
 183                         rc = -errno;
 184                 }
 185                 goto out;
 186         }
 187
 188         session->evlist = evlist;
 189         perf_session__set_id_hdr_size(session);
 190 out:
 191         return rc;
 192 }
 193
 194 static int process_sample_event(struct perf_tool *tool,
 195                                 union perf_event *event,
 196                                 struct perf_sample *sample,
 197                                 struct perf_evsel *evsel,
 198                                 struct machine *machine)
 199 {
 200         struct record *rec = container_of(tool, struct record, tool);
 201
 202         rec->samples++;
 203
 204         return build_id__mark_dso_hit(tool, event, sample, evsel, machine);
 205 }
 206
 207 static int process_buildids(struct record *rec)
 208 {
 209         struct perf_data_file *file  = &rec->file;
 210         struct perf_session *session = rec->session;
 211
 212         u64 size = lseek(perf_data_file__fd(file), 0, SEEK_CUR);
 213         if (size == 0)
 214                 return 0;
 215
 216         file->size = size;
 217
 218         /*
 219          * During this process, it'll load kernel map and replace the
 220          * dso->long_name to a real pathname it found.  In this case
 221          * we prefer the vmlinux path like
 222          *   /lib/modules/3.16.4/build/vmlinux
 223          *
 224          * rather than build-id path (in debug directory).
 225          *   $HOME/.debug/.build-id/f0/6e17aa50adf4d00b88925e03775de107611551
 226          */
 227         symbol_conf.ignore_vmlinux_buildid = true;
 228
 229         return perf_session__process_events(session);
 230 }
 231
 232 static void perf_event__synthesize_guest_os(struct machine *machine, void *data)
 233 {
 234         int err;
 235         struct perf_tool *tool = data;
 236         /*
 237          *As for guest kernel when processing subcommand record&report,
 238          *we arrange module mmap prior to guest kernel mmap and trigger
 239          *a preload dso because default guest module symbols are loaded
 240          *from guest kallsyms instead of /lib/modules/XXX/XXX. This
 241          *method is used to avoid symbol missing when the first addr is
 242          *in module instead of in guest kernel.
 243          */
 244         err = perf_event__synthesize_modules(tool, process_synthesized_event,
 245                                              machine);
 246         if (err < 0)
 247                 pr_err("Couldn't record guest kernel [%d]'s reference"
 248                        " relocation symbol.\n", machine->pid);
 249
 250         /*
 251          * We use _stext for guest kernel because guest kernel's /proc/kallsyms
 252          * have no _text sometimes.
 253          */
 254         err = perf_event__synthesize_kernel_mmap(tool, process_synthesized_event,
 255                                                  machine);
 256         if (err < 0)
 257                 pr_err("Couldn't record guest kernel [%d]'s reference"
 258                        " relocation symbol.\n", machine->pid);
 259 }
 260
 261 static struct perf_event_header finished_round_event = {
 262         .size = sizeof(struct perf_event_header),
 263         .type = PERF_RECORD_FINISHED_ROUND,
 264 };
 265
 266 static int record__mmap_read_all(struct record *rec)
 267 {
 268         u64 bytes_written = rec->bytes_written;
 269         int i;
 270         int rc = 0;
 271
 272         for (i = 0; i < rec->evlist->nr_mmaps; i++) {
 273                 if (rec->evlist->mmap[i].base) {
 274                         if (record__mmap_read(rec, i) != 0) {
 275                                 rc = -1;
 276                                 goto out;
 277                         }
 278                 }
 279         }
 280
 281         /*
 282          * Mark the round finished in case we wrote
 283          * at least one event.
 284          */
 285         if (bytes_written != rec->bytes_written)
 286                 rc = record__write(rec, &finished_round_event, sizeof(finished_round_event));
 287
 288 out:
 289         return rc;
 290 }
 291
 292 static void record__init_features(struct record *rec)
 293 {
 294         struct perf_session *session = rec->session;
 295         int feat;
 296
 297         for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++)
 298                 perf_header__set_feat(&session->header, feat);
 299
 300         if (rec->no_buildid)
 301                 perf_header__clear_feat(&session->header, HEADER_BUILD_ID);
 302
 303         if (!have_tracepoints(&rec->evlist->entries))
 304                 perf_header__clear_feat(&session->header, HEADER_TRACING_DATA);
 305
 306         if (!rec->opts.branch_stack)
 307                 perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK);
 308 }
 309
 310 static volatile int workload_exec_errno;
 311
 312 /*
 313  * perf_evlist__prepare_workload will send a SIGUSR1
 314  * if the fork fails, since we asked by setting its
 315  * want_signal to true.
 316  */
 317 static void workload_exec_failed_signal(int signo __maybe_unused,
 318                                         siginfo_t *info,
 319                                         void *ucontext __maybe_unused)
 320 {
 321         workload_exec_errno = info->si_value.sival_int;
 322         done = 1;
 323         child_finished = 1;
 324 }
 325
 326 static int __cmd_record(struct record *rec, int argc, const char **argv)
 327 {
 328         int err;
 329         int status = 0;
 330         unsigned long waking = 0;
 331         const bool forks = argc > 0;
 332         struct machine *machine;
 333         struct perf_tool *tool = &rec->tool;
 334         struct record_opts *opts = &rec->opts;
 335         struct perf_data_file *file = &rec->file;
 336         struct perf_session *session;
 337         bool disabled = false, draining = false;
 338         int fd;
 339
 340         rec->progname = argv[0];
 341
 342         atexit(record__sig_exit);
 343         signal(SIGCHLD, sig_handler);
 344         signal(SIGINT, sig_handler);
 345         signal(SIGTERM, sig_handler);
 346
 347         session = perf_session__new(file, false, tool);
 348         if (session == NULL) {
 349                 pr_err("Perf session creation failed.\n");
 350                 return -1;
 351         }
 352
 353         fd = perf_data_file__fd(file);
 354         rec->session = session;
 355
 356         record__init_features(rec);
 357
 358         if (forks) {
 359                 err = perf_evlist__prepare_workload(rec->evlist, &opts->target,
 360                                                     argv, file->is_pipe,
 361                                                     workload_exec_failed_signal);
 362                 if (err < 0) {
 363                         pr_err("Couldn't run the workload!\n");
 364                         status = err;
 365                         goto out_delete_session;
 366                 }
 367         }
 368
 369         if (record__open(rec) != 0) {
 370                 err = -1;
 371                 goto out_child;
 372         }
 373
 374         if (!rec->evlist->nr_groups)
 375                 perf_header__clear_feat(&session->header, HEADER_GROUP_DESC);
 376
 377         if (file->is_pipe) {
 378                 err = perf_header__write_pipe(fd);
 379                 if (err < 0)
 380                         goto out_child;
 381         } else {
 382                 err = perf_session__write_header(session, rec->evlist, fd, false);
 383                 if (err < 0)
 384                         goto out_child;
 385         }
 386
 387         if (!rec->no_buildid
 388             && !perf_header__has_feat(&session->header, HEADER_BUILD_ID)) {
 389                 pr_err("Couldn't generate buildids. "
 390                        "Use --no-buildid to profile anyway.\n");
 391                 err = -1;
 392                 goto out_child;
 393         }
 394
 395         machine = &session->machines.host;
 396
 397         if (file->is_pipe) {
 398                 err = perf_event__synthesize_attrs(tool, session,
 399                                                    process_synthesized_event);
 400                 if (err < 0) {
 401                         pr_err("Couldn't synthesize attrs.\n");
 402                         goto out_child;
 403                 }
 404
 405                 if (have_tracepoints(&rec->evlist->entries)) {
 406                         /*
 407                          * FIXME err <= 0 here actually means that
 408                          * there were no tracepoints so its not really
 409                          * an error, just that we don't need to
 410                          * synthesize anything.  We really have to
 411                          * return this more properly and also
 412                          * propagate errors that now are calling die()
 413                          */
 414                         err = perf_event__synthesize_tracing_data(tool, fd, rec->evlist,
 415                                                                   process_synthesized_event);
 416                         if (err <= 0) {
 417                                 pr_err("Couldn't record tracing data.\n");
 418                                 goto out_child;
 419                         }
 420                         rec->bytes_written += err;
 421                 }
 422         }
 423
 424         err = perf_event__synthesize_kernel_mmap(tool, process_synthesized_event,
 425                                                  machine);
 426         if (err < 0)
 427                 pr_err("Couldn't record kernel reference relocation symbol\n"
 428                        "Symbol resolution may be skewed if relocation was used (e.g. kexec).\n"
 429                        "Check /proc/kallsyms permission or run as root.\n");
 430
 431         err = perf_event__synthesize_modules(tool, process_synthesized_event,
 432                                              machine);
 433         if (err < 0)
 434                 pr_err("Couldn't record kernel module information.\n"
 435                        "Symbol resolution may be skewed if relocation was used (e.g. kexec).\n"
 436                        "Check /proc/modules permission or run as root.\n");
 437
 438         if (perf_guest) {
 439                 machines__process_guests(&session->machines,
 440                                          perf_event__synthesize_guest_os, tool);
 441         }
 442
 443         err = __machine__synthesize_threads(machine, tool, &opts->target, rec->evlist->threads,
 444                                             process_synthesized_event, opts->sample_address);
 445         if (err != 0)
 446                 goto out_child;
 447
 448         if (rec->realtime_prio) {
 449                 struct sched_param param;
 450
 451                 param.sched_priority = rec->realtime_prio;
 452                 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
 453                         pr_err("Could not set realtime priority.\n");
 454                         err = -1;
 455                         goto out_child;
 456                 }
 457         }
 458
 459         /*
 460          * When perf is starting the traced process, all the events
 461          * (apart from group members) have enable_on_exec=1 set,
 462          * so don't spoil it by prematurely enabling them.
 463          */
 464         if (!target__none(&opts->target) && !opts->initial_delay)
 465                 perf_evlist__enable(rec->evlist);
 466
 467         /*
 468          * Let the child rip
 469          */
 470         if (forks)
 471                 perf_evlist__start_workload(rec->evlist);
 472
 473         if (opts->initial_delay) {
 474                 usleep(opts->initial_delay * 1000);
 475                 perf_evlist__enable(rec->evlist);
 476         }
 477
 478         for (;;) {
 479                 int hits = rec->samples;
 480
 481                 if (record__mmap_read_all(rec) < 0) {
 482                         err = -1;
 483                         goto out_child;
 484                 }
 485
 486                 if (hits == rec->samples) {
 487                         if (done || draining)
 488                                 break;
 489                         err = perf_evlist__poll(rec->evlist, -1);
 490                         /*
 491                          * Propagate error, only if there's any. Ignore positive
 492                          * number of returned events and interrupt error.
 493                          */
 494                         if (err > 0 || (err < 0 && errno == EINTR))
 495                                 err = 0;
 496                         waking++;
 497
 498                         if (perf_evlist__filter_pollfd(rec->evlist, POLLERR | POLLHUP) == 0)
 499                                 draining = true;
 500                 }
 501
 502                 /*
 503                  * When perf is starting the traced process, at the end events
 504                  * die with the process and we wait for that. Thus no need to
 505                  * disable events in this case.
 506                  */
 507                 if (done && !disabled && !target__none(&opts->target)) {
 508                         perf_evlist__disable(rec->evlist);
 509                         disabled = true;
 510                 }
 511         }
 512
 513         if (forks && workload_exec_errno) {
 514                 char msg[STRERR_BUFSIZE];
 515                 const char *emsg = strerror_r(workload_exec_errno, msg, sizeof(msg));
 516                 pr_err("Workload failed: %s\n", emsg);
 517                 err = -1;
 518                 goto out_child;
 519         }
 520
 521         if (!quiet)
 522                 fprintf(stderr, "[ perf record: Woken up %ld times to write data ]\n", waking);
 523
 524 out_child:
 525         if (forks) {
 526                 int exit_status;
 527
 528                 if (!child_finished)
 529                         kill(rec->evlist->workload.pid, SIGTERM);
 530
 531                 wait(&exit_status);
 532
 533                 if (err < 0)
 534                         status = err;
 535                 else if (WIFEXITED(exit_status))
 536                         status = WEXITSTATUS(exit_status);
 537                 else if (WIFSIGNALED(exit_status))
 538                         signr = WTERMSIG(exit_status);
 539         } else
 540                 status = err;
 541
 542         /* this will be recalculated during process_buildids() */
 543         rec->samples = 0;
 544
 545         if (!err && !file->is_pipe) {
 546                 rec->session->header.data_size += rec->bytes_written;
 547
 548                 if (!rec->no_buildid)
 549                         process_buildids(rec);
 550                 perf_session__write_header(rec->session, rec->evlist, fd, true);
 551         }
 552
 553         if (!err && !quiet) {
 554                 char samples[128];
 555
 556                 if (rec->samples)
 557                         scnprintf(samples, sizeof(samples),
 558                                   " (%" PRIu64 " samples)", rec->samples);
 559                 else
 560                         samples[0] = '\0';
 561
 562                 fprintf(stderr, "[ perf record: Captured and wrote %.3f MB %s%s ]\n",
 563                         perf_data_file__size(file) / 1024.0 / 1024.0,
 564                         file->path, samples);
 565         }
 566
 567 out_delete_session:
 568         perf_session__delete(session);
 569         return status;
 570 }
 571
 572 #define BRANCH_OPT(n, m) \
 573         { .name = n, .mode = (m) }
 574
 575 #define BRANCH_END { .name = NULL }
 576
 577 struct branch_mode {
 578         const char *name;
 579         int mode;
 580 };
 581
 582 static const struct branch_mode branch_modes[] = {
 583         BRANCH_OPT("u", PERF_SAMPLE_BRANCH_USER),
 584         BRANCH_OPT("k", PERF_SAMPLE_BRANCH_KERNEL),
 585         BRANCH_OPT("hv", PERF_SAMPLE_BRANCH_HV),
 586         BRANCH_OPT("any", PERF_SAMPLE_BRANCH_ANY),
 587         BRANCH_OPT("any_call", PERF_SAMPLE_BRANCH_ANY_CALL),
 588         BRANCH_OPT("any_ret", PERF_SAMPLE_BRANCH_ANY_RETURN),
 589         BRANCH_OPT("ind_call", PERF_SAMPLE_BRANCH_IND_CALL),
 590         BRANCH_OPT("abort_tx", PERF_SAMPLE_BRANCH_ABORT_TX),
 591         BRANCH_OPT("in_tx", PERF_SAMPLE_BRANCH_IN_TX),
 592         BRANCH_OPT("no_tx", PERF_SAMPLE_BRANCH_NO_TX),
 593         BRANCH_OPT("cond", PERF_SAMPLE_BRANCH_COND),
 594         BRANCH_END
 595 };
 596
 597 static int
 598 parse_branch_stack(const struct option *opt, const char *str, int unset)
 599 {
 600 #define ONLY_PLM \
 601         (PERF_SAMPLE_BRANCH_USER        |\
 602          PERF_SAMPLE_BRANCH_KERNEL      |\
 603          PERF_SAMPLE_BRANCH_HV)
 604
 605         uint64_t *mode = (uint64_t *)opt->value;
 606         const struct branch_mode *br;
 607         char *s, *os = NULL, *p;
 608         int ret = -1;
 609
 610         if (unset)
 611                 return 0;
 612
 613         /*
 614          * cannot set it twice, -b + --branch-filter for instance
 615          */
 616         if (*mode)
 617                 return -1;
 618
 619         /* str may be NULL in case no arg is passed to -b */
 620         if (str) {
 621                 /* because str is read-only */
 622                 s = os = strdup(str);
 623                 if (!s)
 624                         return -1;
 625
 626                 for (;;) {
 627                         p = strchr(s, ',');
 628                         if (p)
 629                                 *p = '\0';
 630
 631                         for (br = branch_modes; br->name; br++) {
 632                                 if (!strcasecmp(s, br->name))
 633                                         break;
 634                         }
 635                         if (!br->name) {
 636                                 ui__warning("unknown branch filter %s,"
 637                                             " check man page\n", s);
 638                                 goto error;
 639                         }
 640
 641                         *mode |= br->mode;
 642
 643                         if (!p)
 644                                 break;
 645
 646                         s = p + 1;
 647                 }
 648         }
 649         ret = 0;
 650
 651         /* default to any branch */
 652         if ((*mode & ~ONLY_PLM) == 0) {
 653                 *mode = PERF_SAMPLE_BRANCH_ANY;
 654         }
 655 error:
 656         free(os);
 657         return ret;
 658 }
 659
 660 static void callchain_debug(void)
 661 {
 662         static const char *str[CALLCHAIN_MAX] = { "NONE", "FP", "DWARF", "LBR" };
 663
 664         pr_debug("callchain: type %s\n", str[callchain_param.record_mode]);
 665
 666         if (callchain_param.record_mode == CALLCHAIN_DWARF)
 667                 pr_debug("callchain: stack dump size %d\n",
 668                          callchain_param.dump_size);
 669 }
 670
 671 int record_parse_callchain_opt(const struct option *opt __maybe_unused,
 672                                const char *arg,
 673                                int unset)
 674 {
 675         int ret;
 676
 677         callchain_param.enabled = !unset;
 678
 679         /* --no-call-graph */
 680         if (unset) {
 681                 callchain_param.record_mode = CALLCHAIN_NONE;
 682                 pr_debug("callchain: disabled\n");
 683                 return 0;
 684         }
 685
 686         ret = parse_callchain_record_opt(arg);
 687         if (!ret)
 688                 callchain_debug();
 689
 690         return ret;
 691 }
 692
 693 int record_callchain_opt(const struct option *opt __maybe_unused,
 694                          const char *arg __maybe_unused,
 695                          int unset __maybe_unused)
 696 {
 697         callchain_param.enabled = true;
 698
 699         if (callchain_param.record_mode == CALLCHAIN_NONE)
 700                 callchain_param.record_mode = CALLCHAIN_FP;
 701
 702         callchain_debug();
 703         return 0;
 704 }
 705
 706 static int perf_record_config(const char *var, const char *value, void *cb)
 707 {
 708         if (!strcmp(var, "record.call-graph"))
 709                 var = "call-graph.record-mode"; /* fall-through */
 710
 711         return perf_default_config(var, value, cb);
 712 }
 713
 714 struct clockid_map {
 715         const char *name;
 716         int clockid;
 717 };
 718
 719 #define CLOCKID_MAP(n, c)       \
 720         { .name = n, .clockid = (c), }
 721
 722 #define CLOCKID_END     { .name = NULL, }
 723
 724
 725 /*
 726  * Add the missing ones, we need to build on many distros...
 727  */
 728 #ifndef CLOCK_MONOTONIC_RAW
 729 #define CLOCK_MONOTONIC_RAW 4
 730 #endif
 731 #ifndef CLOCK_BOOTTIME
 732 #define CLOCK_BOOTTIME 7
 733 #endif
 734 #ifndef CLOCK_TAI
 735 #define CLOCK_TAI 11
 736 #endif
 737
 738 static const struct clockid_map clockids[] = {
 739         /* available for all events, NMI safe */
 740         CLOCKID_MAP("monotonic", CLOCK_MONOTONIC),
 741         CLOCKID_MAP("monotonic_raw", CLOCK_MONOTONIC_RAW),
 742
 743         /* available for some events */
 744         CLOCKID_MAP("realtime", CLOCK_REALTIME),
 745         CLOCKID_MAP("boottime", CLOCK_BOOTTIME),
 746         CLOCKID_MAP("tai", CLOCK_TAI),
 747
 748         /* available for the lazy */
 749         CLOCKID_MAP("mono", CLOCK_MONOTONIC),
 750         CLOCKID_MAP("raw", CLOCK_MONOTONIC_RAW),
 751         CLOCKID_MAP("real", CLOCK_REALTIME),
 752         CLOCKID_MAP("boot", CLOCK_BOOTTIME),
 753
 754         CLOCKID_END,
 755 };
 756
 757 static int parse_clockid(const struct option *opt, const char *str, int unset)
 758 {
 759         struct record_opts *opts = (struct record_opts *)opt->value;
 760         const struct clockid_map *cm;
 761         const char *ostr = str;
 762
 763         if (unset) {
 764                 opts->use_clockid = 0;
 765                 return 0;
 766         }
 767
 768         /* no arg passed */
 769         if (!str)
 770                 return 0;
 771
 772         /* no setting it twice */
 773         if (opts->use_clockid)
 774                 return -1;
 775
 776         opts->use_clockid = true;
 777
 778         /* if its a number, we're done */
 779         if (sscanf(str, "%d", &opts->clockid) == 1)
 780                 return 0;
 781
 782         /* allow a "CLOCK_" prefix to the name */
 783         if (!strncasecmp(str, "CLOCK_", 6))
 784                 str += 6;
 785
 786         for (cm = clockids; cm->name; cm++) {
 787                 if (!strcasecmp(str, cm->name)) {
 788                         opts->clockid = cm->clockid;
 789                         return 0;
 790                 }
 791         }
 792
 793         opts->use_clockid = false;
 794         ui__warning("unknown clockid %s, check man page\n", ostr);
 795         return -1;
 796 }
 797
 798 static const char * const __record_usage[] = {
 799         "perf record [<options>] [<command>]",
 800         "perf record [<options>] -- <command> [<options>]",
 801         NULL
 802 };
 803 const char * const *record_usage = __record_usage;
 804
 805 /*
 806  * XXX Ideally would be local to cmd_record() and passed to a record__new
 807  * because we need to have access to it in record__exit, that is called
 808  * after cmd_record() exits, but since record_options need to be accessible to
 809  * builtin-script, leave it here.
 810  *
 811  * At least we don't ouch it in all the other functions here directly.
 812  *
 813  * Just say no to tons of global variables, sigh.
 814  */
 815 static struct record record = {
 816         .opts = {
 817                 .sample_time         = true,
 818                 .mmap_pages          = UINT_MAX,
 819                 .user_freq           = UINT_MAX,
 820                 .user_interval       = ULLONG_MAX,
 821                 .freq                = 4000,
 822                 .target              = {
 823                         .uses_mmap   = true,
 824                         .default_per_cpu = true,
 825                 },
 826         },
 827         .tool = {
 828                 .sample         = process_sample_event,
 829                 .fork           = perf_event__process_fork,
 830                 .comm           = perf_event__process_comm,
 831                 .mmap           = perf_event__process_mmap,
 832                 .mmap2          = perf_event__process_mmap2,
 833         },
 834 };
 835
 836 #define CALLCHAIN_HELP "setup and enables call-graph (stack chain/backtrace) recording: "
 837
 838 #ifdef HAVE_DWARF_UNWIND_SUPPORT
 839 const char record_callchain_help[] = CALLCHAIN_HELP "fp dwarf lbr";
 840 #else
 841 const char record_callchain_help[] = CALLCHAIN_HELP "fp lbr";
 842 #endif
 843
 844 /*
 845  * XXX Will stay a global variable till we fix builtin-script.c to stop messing
 846  * with it and switch to use the library functions in perf_evlist that came
 847  * from builtin-record.c, i.e. use record_opts,
 848  * perf_evlist__prepare_workload, etc instead of fork+exec'in 'perf record',
 849  * using pipes, etc.
 850  */
 851 struct option __record_options[] = {
 852         OPT_CALLBACK('e', "event", &record.evlist, "event",
 853                      "event selector. use 'perf list' to list available events",
 854                      parse_events_option),
 855         OPT_CALLBACK(0, "filter", &record.evlist, "filter",
 856                      "event filter", parse_filter),
 857         OPT_STRING('p', "pid", &record.opts.target.pid, "pid",
 858                     "record events on existing process id"),
 859         OPT_STRING('t', "tid", &record.opts.target.tid, "tid",
 860                     "record events on existing thread id"),
 861         OPT_INTEGER('r', "realtime", &record.realtime_prio,
 862                     "collect data with this RT SCHED_FIFO priority"),
 863         OPT_BOOLEAN(0, "no-buffering", &record.opts.no_buffering,
 864                     "collect data without buffering"),
 865         OPT_BOOLEAN('R', "raw-samples", &record.opts.raw_samples,
 866                     "collect raw sample records from all opened counters"),
 867         OPT_BOOLEAN('a', "all-cpus", &record.opts.target.system_wide,
 868                             "system-wide collection from all CPUs"),
 869         OPT_STRING('C', "cpu", &record.opts.target.cpu_list, "cpu",
 870                     "list of cpus to monitor"),
 871         OPT_U64('c', "count", &record.opts.user_interval, "event period to sample"),
 872         OPT_STRING('o', "output", &record.file.path, "file",
 873                     "output file name"),
 874         OPT_BOOLEAN_SET('i', "no-inherit", &record.opts.no_inherit,
 875                         &record.opts.no_inherit_set,
 876                         "child tasks do not inherit counters"),
 877         OPT_UINTEGER('F', "freq", &record.opts.user_freq, "profile at this frequency"),
 878         OPT_CALLBACK('m', "mmap-pages", &record.opts.mmap_pages, "pages",
 879                      "number of mmap data pages",
 880                      perf_evlist__parse_mmap_pages),
 881         OPT_BOOLEAN(0, "group", &record.opts.group,
 882                     "put the counters into a counter group"),
 883         OPT_CALLBACK_NOOPT('g', NULL, &record.opts,
 884                            NULL, "enables call-graph recording" ,
 885                            &record_callchain_opt),
 886         OPT_CALLBACK(0, "call-graph", &record.opts,
 887                      "mode[,dump_size]", record_callchain_help,
 888                      &record_parse_callchain_opt),
 889         OPT_INCR('v', "verbose", &verbose,
 890                     "be more verbose (show counter open errors, etc)"),
 891         OPT_BOOLEAN('q', "quiet", &quiet, "don't print any message"),
 892         OPT_BOOLEAN('s', "stat", &record.opts.inherit_stat,
 893                     "per thread counts"),
 894         OPT_BOOLEAN('d', "data", &record.opts.sample_address,
 895                     "Sample addresses"),
 896         OPT_BOOLEAN('T', "timestamp", &record.opts.sample_time, "Sample timestamps"),
 897         OPT_BOOLEAN('P', "period", &record.opts.period, "Sample period"),
 898         OPT_BOOLEAN('n', "no-samples", &record.opts.no_samples,
 899                     "don't sample"),
 900         OPT_BOOLEAN('N', "no-buildid-cache", &record.no_buildid_cache,
 901                     "do not update the buildid cache"),
 902         OPT_BOOLEAN('B', "no-buildid", &record.no_buildid,
 903                     "do not collect buildids in perf.data"),
 904         OPT_CALLBACK('G', "cgroup", &record.evlist, "name",
 905                      "monitor event in cgroup name only",
 906                      parse_cgroups),
 907         OPT_UINTEGER('D', "delay", &record.opts.initial_delay,
 908                   "ms to wait before starting measurement after program start"),
 909         OPT_STRING('u', "uid", &record.opts.target.uid_str, "user",
 910                    "user to profile"),
 911
 912         OPT_CALLBACK_NOOPT('b', "branch-any", &record.opts.branch_stack,
 913                      "branch any", "sample any taken branches",
 914                      parse_branch_stack),
 915
 916         OPT_CALLBACK('j', "branch-filter", &record.opts.branch_stack,
 917                      "branch filter mask", "branch stack filter modes",
 918                      parse_branch_stack),
 919         OPT_BOOLEAN('W', "weight", &record.opts.sample_weight,
 920                     "sample by weight (on special events only)"),
 921         OPT_BOOLEAN(0, "transaction", &record.opts.sample_transaction,
 922                     "sample transaction flags (special events only)"),
 923         OPT_BOOLEAN(0, "per-thread", &record.opts.target.per_thread,
 924                     "use per-thread mmaps"),
 925         OPT_BOOLEAN('I', "intr-regs", &record.opts.sample_intr_regs,
 926                     "Sample machine registers on interrupt"),
 927         OPT_BOOLEAN(0, "running-time", &record.opts.running_time,
 928                     "Record running/enabled time of read (:S) events"),
 929         OPT_CALLBACK('k', "clockid", &record.opts,
 930         "clockid", "clockid to use for events, see clock_gettime()",
 931         parse_clockid),
 932         OPT_END()
 933 };
 934
 935 struct option *record_options = __record_options;
 936
 937 int cmd_record(int argc, const char **argv, const char *prefix __maybe_unused)
 938 {
 939         int err = -ENOMEM;
 940         struct record *rec = &record;
 941         char errbuf[BUFSIZ];
 942
 943         rec->evlist = perf_evlist__new();
 944         if (rec->evlist == NULL)
 945                 return -ENOMEM;
 946
 947         perf_config(perf_record_config, rec);
 948
 949         argc = parse_options(argc, argv, record_options, record_usage,
 950                             PARSE_OPT_STOP_AT_NON_OPTION);
 951         if (!argc && target__none(&rec->opts.target))
 952                 usage_with_options(record_usage, record_options);
 953
 954         if (nr_cgroups && !rec->opts.target.system_wide) {
 955                 ui__error("cgroup monitoring only available in"
 956                           " system-wide mode\n");
 957                 usage_with_options(record_usage, record_options);
 958         }
 959
 960         symbol__init(NULL);
 961
 962         if (symbol_conf.kptr_restrict)
 963                 pr_warning(
 964 "WARNING: Kernel address maps (/proc/{kallsyms,modules}) are restricted,\n"
 965 "check /proc/sys/kernel/kptr_restrict.\n\n"
 966 "Samples in kernel functions may not be resolved if a suitable vmlinux\n"
 967 "file is not found in the buildid cache or in the vmlinux path.\n\n"
 968 "Samples in kernel modules won't be resolved at all.\n\n"
 969 "If some relocation was applied (e.g. kexec) symbols may be misresolved\n"
 970 "even with a suitable vmlinux or kallsyms file.\n\n");
 971
 972         if (rec->no_buildid_cache || rec->no_buildid)
 973                 disable_buildid_cache();
 974
 975         if (rec->evlist->nr_entries == 0 &&
 976             perf_evlist__add_default(rec->evlist) < 0) {
 977                 pr_err("Not enough memory for event selector list\n");
 978                 goto out_symbol_exit;
 979         }
 980
 981         if (rec->opts.target.tid && !rec->opts.no_inherit_set)
 982                 rec->opts.no_inherit = true;
 983
 984         err = target__validate(&rec->opts.target);
 985         if (err) {
 986                 target__strerror(&rec->opts.target, err, errbuf, BUFSIZ);
 987                 ui__warning("%s", errbuf);
 988         }
 989
 990         err = target__parse_uid(&rec->opts.target);
 991         if (err) {
 992                 int saved_errno = errno;
 993
 994                 target__strerror(&rec->opts.target, err, errbuf, BUFSIZ);
 995                 ui__error("%s", errbuf);
 996
 997                 err = -saved_errno;
 998                 goto out_symbol_exit;
 999         }
1000
1001         err = -ENOMEM;
1002         if (perf_evlist__create_maps(rec->evlist, &rec->opts.target) < 0)
1003                 usage_with_options(record_usage, record_options);
1004
1005         if (record_opts__config(&rec->opts)) {
1006                 err = -EINVAL;
1007                 goto out_symbol_exit;
1008         }
1009
1010         err = __cmd_record(&record, argc, argv);
1011 out_symbol_exit:
1012         perf_evlist__delete(rec->evlist);
1013         symbol__exit();
1014         return err;
1015 }