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