tools/perf/builtin-record.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * builtin-record.c
   4  *
   5  * Builtin record command: Record the profile of a workload
   6  * (or a CPU, or a PID) into the perf.data output file - for
   7  * later analysis via perf report.
   8  */
   9 #include "builtin.h"
  10
  11 #include "perf.h"
  12
  13 #include "util/build-id.h"
  14 #include "util/util.h"
  15 #include <subcmd/parse-options.h>
  16 #include "util/parse-events.h"
  17 #include "util/config.h"
  18
  19 #include "util/callchain.h"
  20 #include "util/cgroup.h"
  21 #include "util/header.h"
  22 #include "util/event.h"
  23 #include "util/evlist.h"
  24 #include "util/evsel.h"
  25 #include "util/debug.h"
  26 #include "util/drv_configs.h"
  27 #include "util/session.h"
  28 #include "util/tool.h"
  29 #include "util/symbol.h"
  30 #include "util/cpumap.h"
  31 #include "util/thread_map.h"
  32 #include "util/data.h"
  33 #include "util/perf_regs.h"
  34 #include "util/auxtrace.h"
  35 #include "util/tsc.h"
  36 #include "util/parse-branch-options.h"
  37 #include "util/parse-regs-options.h"
  38 #include "util/llvm-utils.h"
  39 #include "util/bpf-loader.h"
  40 #include "util/trigger.h"
  41 #include "util/perf-hooks.h"
  42 #include "util/time-utils.h"
  43 #include "util/units.h"
  44 #include "asm/bug.h"
  45
  46 #include <errno.h>
  47 #include <inttypes.h>
  48 #include <poll.h>
  49 #include <unistd.h>
  50 #include <sched.h>
  51 #include <signal.h>
  52 #include <sys/mman.h>
  53 #include <sys/wait.h>
  54 #include <linux/time64.h>
  55
  56 struct switch_output {
  57         bool             enabled;
  58         bool             signal;
  59         unsigned long    size;
  60         unsigned long    time;
  61         const char      *str;
  62         bool             set;
  63 };
  64
  65 struct record {
  66         struct perf_tool        tool;
  67         struct record_opts      opts;
  68         u64                     bytes_written;
  69         struct perf_data        data;
  70         struct auxtrace_record  *itr;
  71         struct perf_evlist      *evlist;
  72         struct perf_session     *session;
  73         const char              *progname;
  74         int                     realtime_prio;
  75         bool                    no_buildid;
  76         bool                    no_buildid_set;
  77         bool                    no_buildid_cache;
  78         bool                    no_buildid_cache_set;
  79         bool                    buildid_all;
  80         bool                    timestamp_filename;
  81         bool                    timestamp_boundary;
  82         struct switch_output    switch_output;
  83         unsigned long long      samples;
  84 };
  85
  86 static volatile int auxtrace_record__snapshot_started;
  87 static DEFINE_TRIGGER(auxtrace_snapshot_trigger);
  88 static DEFINE_TRIGGER(switch_output_trigger);
  89
  90 static bool switch_output_signal(struct record *rec)
  91 {
  92         return rec->switch_output.signal &&
  93                trigger_is_ready(&switch_output_trigger);
  94 }
  95
  96 static bool switch_output_size(struct record *rec)
  97 {
  98         return rec->switch_output.size &&
  99                trigger_is_ready(&switch_output_trigger) &&
 100                (rec->bytes_written >= rec->switch_output.size);
 101 }
 102
 103 static bool switch_output_time(struct record *rec)
 104 {
 105         return rec->switch_output.time &&
 106                trigger_is_ready(&switch_output_trigger);
 107 }
 108
 109 static int record__write(struct record *rec, void *bf, size_t size)
 110 {
 111         if (perf_data__write(rec->session->data, bf, size) < 0) {
 112                 pr_err("failed to write perf data, error: %m\n");
 113                 return -1;
 114         }
 115
 116         rec->bytes_written += size;
 117
 118         if (switch_output_size(rec))
 119                 trigger_hit(&switch_output_trigger);
 120
 121         return 0;
 122 }
 123
 124 static int process_synthesized_event(struct perf_tool *tool,
 125                                      union perf_event *event,
 126                                      struct perf_sample *sample __maybe_unused,
 127                                      struct machine *machine __maybe_unused)
 128 {
 129         struct record *rec = container_of(tool, struct record, tool);
 130         return record__write(rec, event, event->header.size);
 131 }
 132
 133 static int record__pushfn(void *to, void *bf, size_t size)
 134 {
 135         struct record *rec = to;
 136
 137         rec->samples++;
 138         return record__write(rec, bf, size);
 139 }
 140
 141 static volatile int done;
 142 static volatile int signr = -1;
 143 static volatile int child_finished;
 144
 145 static void sig_handler(int sig)
 146 {
 147         if (sig == SIGCHLD)
 148                 child_finished = 1;
 149         else
 150                 signr = sig;
 151
 152         done = 1;
 153 }
 154
 155 static void sigsegv_handler(int sig)
 156 {
 157         perf_hooks__recover();
 158         sighandler_dump_stack(sig);
 159 }
 160
 161 static void record__sig_exit(void)
 162 {
 163         if (signr == -1)
 164                 return;
 165
 166         signal(signr, SIG_DFL);
 167         raise(signr);
 168 }
 169
 170 #ifdef HAVE_AUXTRACE_SUPPORT
 171
 172 static int record__process_auxtrace(struct perf_tool *tool,
 173                                     union perf_event *event, void *data1,
 174                                     size_t len1, void *data2, size_t len2)
 175 {
 176         struct record *rec = container_of(tool, struct record, tool);
 177         struct perf_data *data = &rec->data;
 178         size_t padding;
 179         u8 pad[8] = {0};
 180
 181         if (!perf_data__is_pipe(data)) {
 182                 off_t file_offset;
 183                 int fd = perf_data__fd(data);
 184                 int err;
 185
 186                 file_offset = lseek(fd, 0, SEEK_CUR);
 187                 if (file_offset == -1)
 188                         return -1;
 189                 err = auxtrace_index__auxtrace_event(&rec->session->auxtrace_index,
 190                                                      event, file_offset);
 191                 if (err)
 192                         return err;
 193         }
 194
 195         /* event.auxtrace.size includes padding, see __auxtrace_mmap__read() */
 196         padding = (len1 + len2) & 7;
 197         if (padding)
 198                 padding = 8 - padding;
 199
 200         record__write(rec, event, event->header.size);
 201         record__write(rec, data1, len1);
 202         if (len2)
 203                 record__write(rec, data2, len2);
 204         record__write(rec, &pad, padding);
 205
 206         return 0;
 207 }
 208
 209 static int record__auxtrace_mmap_read(struct record *rec,
 210                                       struct auxtrace_mmap *mm)
 211 {
 212         int ret;
 213
 214         ret = auxtrace_mmap__read(mm, rec->itr, &rec->tool,
 215                                   record__process_auxtrace);
 216         if (ret < 0)
 217                 return ret;
 218
 219         if (ret)
 220                 rec->samples++;
 221
 222         return 0;
 223 }
 224
 225 static int record__auxtrace_mmap_read_snapshot(struct record *rec,
 226                                                struct auxtrace_mmap *mm)
 227 {
 228         int ret;
 229
 230         ret = auxtrace_mmap__read_snapshot(mm, rec->itr, &rec->tool,
 231                                            record__process_auxtrace,
 232                                            rec->opts.auxtrace_snapshot_size);
 233         if (ret < 0)
 234                 return ret;
 235
 236         if (ret)
 237                 rec->samples++;
 238
 239         return 0;
 240 }
 241
 242 static int record__auxtrace_read_snapshot_all(struct record *rec)
 243 {
 244         int i;
 245         int rc = 0;
 246
 247         for (i = 0; i < rec->evlist->nr_mmaps; i++) {
 248                 struct auxtrace_mmap *mm =
 249                                 &rec->evlist->mmap[i].auxtrace_mmap;
 250
 251                 if (!mm->base)
 252                         continue;
 253
 254                 if (record__auxtrace_mmap_read_snapshot(rec, mm) != 0) {
 255                         rc = -1;
 256                         goto out;
 257                 }
 258         }
 259 out:
 260         return rc;
 261 }
 262
 263 static void record__read_auxtrace_snapshot(struct record *rec)
 264 {
 265         pr_debug("Recording AUX area tracing snapshot\n");
 266         if (record__auxtrace_read_snapshot_all(rec) < 0) {
 267                 trigger_error(&auxtrace_snapshot_trigger);
 268         } else {
 269                 if (auxtrace_record__snapshot_finish(rec->itr))
 270                         trigger_error(&auxtrace_snapshot_trigger);
 271                 else
 272                         trigger_ready(&auxtrace_snapshot_trigger);
 273         }
 274 }
 275
 276 #else
 277
 278 static inline
 279 int record__auxtrace_mmap_read(struct record *rec __maybe_unused,
 280                                struct auxtrace_mmap *mm __maybe_unused)
 281 {
 282         return 0;
 283 }
 284
 285 static inline
 286 void record__read_auxtrace_snapshot(struct record *rec __maybe_unused)
 287 {
 288 }
 289
 290 static inline
 291 int auxtrace_record__snapshot_start(struct auxtrace_record *itr __maybe_unused)
 292 {
 293         return 0;
 294 }
 295
 296 #endif
 297
 298 static int record__mmap_evlist(struct record *rec,
 299                                struct perf_evlist *evlist)
 300 {
 301         struct record_opts *opts = &rec->opts;
 302         char msg[512];
 303
 304         if (perf_evlist__mmap_ex(evlist, opts->mmap_pages,
 305                                  opts->auxtrace_mmap_pages,
 306                                  opts->auxtrace_snapshot_mode) < 0) {
 307                 if (errno == EPERM) {
 308                         pr_err("Permission error mapping pages.\n"
 309                                "Consider increasing "
 310                                "/proc/sys/kernel/perf_event_mlock_kb,\n"
 311                                "or try again with a smaller value of -m/--mmap_pages.\n"
 312                                "(current value: %u,%u)\n",
 313                                opts->mmap_pages, opts->auxtrace_mmap_pages);
 314                         return -errno;
 315                 } else {
 316                         pr_err("failed to mmap with %d (%s)\n", errno,
 317                                 str_error_r(errno, msg, sizeof(msg)));
 318                         if (errno)
 319                                 return -errno;
 320                         else
 321                                 return -EINVAL;
 322                 }
 323         }
 324         return 0;
 325 }
 326
 327 static int record__mmap(struct record *rec)
 328 {
 329         return record__mmap_evlist(rec, rec->evlist);
 330 }
 331
 332 static int record__open(struct record *rec)
 333 {
 334         char msg[BUFSIZ];
 335         struct perf_evsel *pos;
 336         struct perf_evlist *evlist = rec->evlist;
 337         struct perf_session *session = rec->session;
 338         struct record_opts *opts = &rec->opts;
 339         struct perf_evsel_config_term *err_term;
 340         int rc = 0;
 341
 342         /*
 343          * For initial_delay we need to add a dummy event so that we can track
 344          * PERF_RECORD_MMAP while we wait for the initial delay to enable the
 345          * real events, the ones asked by the user.
 346          */
 347         if (opts->initial_delay) {
 348                 if (perf_evlist__add_dummy(evlist))
 349                         return -ENOMEM;
 350
 351                 pos = perf_evlist__first(evlist);
 352                 pos->tracking = 0;
 353                 pos = perf_evlist__last(evlist);
 354                 pos->tracking = 1;
 355                 pos->attr.enable_on_exec = 1;
 356         }
 357
 358         perf_evlist__config(evlist, opts, &callchain_param);
 359
 360         evlist__for_each_entry(evlist, pos) {
 361 try_again:
 362                 if (perf_evsel__open(pos, pos->cpus, pos->threads) < 0) {
 363                         if (perf_evsel__fallback(pos, errno, msg, sizeof(msg))) {
 364                                 if (verbose > 0)
 365                                         ui__warning("%s\n", msg);
 366                                 goto try_again;
 367                         }
 368
 369                         rc = -errno;
 370                         perf_evsel__open_strerror(pos, &opts->target,
 371                                                   errno, msg, sizeof(msg));
 372                         ui__error("%s\n", msg);
 373                         goto out;
 374                 }
 375
 376                 pos->supported = true;
 377         }
 378
 379         if (perf_evlist__apply_filters(evlist, &pos)) {
 380                 pr_err("failed to set filter \"%s\" on event %s with %d (%s)\n",
 381                         pos->filter, perf_evsel__name(pos), errno,
 382                         str_error_r(errno, msg, sizeof(msg)));
 383                 rc = -1;
 384                 goto out;
 385         }
 386
 387         if (perf_evlist__apply_drv_configs(evlist, &pos, &err_term)) {
 388                 pr_err("failed to set config \"%s\" on event %s with %d (%s)\n",
 389                       err_term->val.drv_cfg, perf_evsel__name(pos), errno,
 390                       str_error_r(errno, msg, sizeof(msg)));
 391                 rc = -1;
 392                 goto out;
 393         }
 394
 395         rc = record__mmap(rec);
 396         if (rc)
 397                 goto out;
 398
 399         session->evlist = evlist;
 400         perf_session__set_id_hdr_size(session);
 401 out:
 402         return rc;
 403 }
 404
 405 static int process_sample_event(struct perf_tool *tool,
 406                                 union perf_event *event,
 407                                 struct perf_sample *sample,
 408                                 struct perf_evsel *evsel,
 409                                 struct machine *machine)
 410 {
 411         struct record *rec = container_of(tool, struct record, tool);
 412
 413         if (rec->evlist->first_sample_time == 0)
 414                 rec->evlist->first_sample_time = sample->time;
 415
 416         rec->evlist->last_sample_time = sample->time;
 417
 418         if (rec->buildid_all)
 419                 return 0;
 420
 421         rec->samples++;
 422         return build_id__mark_dso_hit(tool, event, sample, evsel, machine);
 423 }
 424
 425 static int process_buildids(struct record *rec)
 426 {
 427         struct perf_data *data = &rec->data;
 428         struct perf_session *session = rec->session;
 429
 430         if (data->size == 0)
 431                 return 0;
 432
 433         /*
 434          * During this process, it'll load kernel map and replace the
 435          * dso->long_name to a real pathname it found.  In this case
 436          * we prefer the vmlinux path like
 437          *   /lib/modules/3.16.4/build/vmlinux
 438          *
 439          * rather than build-id path (in debug directory).
 440          *   $HOME/.debug/.build-id/f0/6e17aa50adf4d00b88925e03775de107611551
 441          */
 442         symbol_conf.ignore_vmlinux_buildid = true;
 443
 444         /*
 445          * If --buildid-all is given, it marks all DSO regardless of hits,
 446          * so no need to process samples. But if timestamp_boundary is enabled,
 447          * it still needs to walk on all samples to get the timestamps of
 448          * first/last samples.
 449          */
 450         if (rec->buildid_all && !rec->timestamp_boundary)
 451                 rec->tool.sample = NULL;
 452
 453         return perf_session__process_events(session);
 454 }
 455
 456 static void perf_event__synthesize_guest_os(struct machine *machine, void *data)
 457 {
 458         int err;
 459         struct perf_tool *tool = data;
 460         /*
 461          *As for guest kernel when processing subcommand record&report,
 462          *we arrange module mmap prior to guest kernel mmap and trigger
 463          *a preload dso because default guest module symbols are loaded
 464          *from guest kallsyms instead of /lib/modules/XXX/XXX. This
 465          *method is used to avoid symbol missing when the first addr is
 466          *in module instead of in guest kernel.
 467          */
 468         err = perf_event__synthesize_modules(tool, process_synthesized_event,
 469                                              machine);
 470         if (err < 0)
 471                 pr_err("Couldn't record guest kernel [%d]'s reference"
 472                        " relocation symbol.\n", machine->pid);
 473
 474         /*
 475          * We use _stext for guest kernel because guest kernel's /proc/kallsyms
 476          * have no _text sometimes.
 477          */
 478         err = perf_event__synthesize_kernel_mmap(tool, process_synthesized_event,
 479                                                  machine);
 480         if (err < 0)
 481                 pr_err("Couldn't record guest kernel [%d]'s reference"
 482                        " relocation symbol.\n", machine->pid);
 483 }
 484
 485 static struct perf_event_header finished_round_event = {
 486         .size = sizeof(struct perf_event_header),
 487         .type = PERF_RECORD_FINISHED_ROUND,
 488 };
 489
 490 static int record__mmap_read_evlist(struct record *rec, struct perf_evlist *evlist,
 491                                     bool overwrite)
 492 {
 493         u64 bytes_written = rec->bytes_written;
 494         int i;
 495         int rc = 0;
 496         struct perf_mmap *maps;
 497
 498         if (!evlist)
 499                 return 0;
 500
 501         maps = overwrite ? evlist->overwrite_mmap : evlist->mmap;
 502         if (!maps)
 503                 return 0;
 504
 505         if (overwrite && evlist->bkw_mmap_state != BKW_MMAP_DATA_PENDING)
 506                 return 0;
 507
 508         for (i = 0; i < evlist->nr_mmaps; i++) {
 509                 struct auxtrace_mmap *mm = &maps[i].auxtrace_mmap;
 510
 511                 if (maps[i].base) {
 512                         if (perf_mmap__push(&maps[i], overwrite, rec, record__pushfn) != 0) {
 513                                 rc = -1;
 514                                 goto out;
 515                         }
 516                 }
 517
 518                 if (mm->base && !rec->opts.auxtrace_snapshot_mode &&
 519                     record__auxtrace_mmap_read(rec, mm) != 0) {
 520                         rc = -1;
 521                         goto out;
 522                 }
 523         }
 524
 525         /*
 526          * Mark the round finished in case we wrote
 527          * at least one event.
 528          */
 529         if (bytes_written != rec->bytes_written)
 530                 rc = record__write(rec, &finished_round_event, sizeof(finished_round_event));
 531
 532         if (overwrite)
 533                 perf_evlist__toggle_bkw_mmap(evlist, BKW_MMAP_EMPTY);
 534 out:
 535         return rc;
 536 }
 537
 538 static int record__mmap_read_all(struct record *rec)
 539 {
 540         int err;
 541
 542         err = record__mmap_read_evlist(rec, rec->evlist, false);
 543         if (err)
 544                 return err;
 545
 546         return record__mmap_read_evlist(rec, rec->evlist, true);
 547 }
 548
 549 static void record__init_features(struct record *rec)
 550 {
 551         struct perf_session *session = rec->session;
 552         int feat;
 553
 554         for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++)
 555                 perf_header__set_feat(&session->header, feat);
 556
 557         if (rec->no_buildid)
 558                 perf_header__clear_feat(&session->header, HEADER_BUILD_ID);
 559
 560         if (!have_tracepoints(&rec->evlist->entries))
 561                 perf_header__clear_feat(&session->header, HEADER_TRACING_DATA);
 562
 563         if (!rec->opts.branch_stack)
 564                 perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK);
 565
 566         if (!rec->opts.full_auxtrace)
 567                 perf_header__clear_feat(&session->header, HEADER_AUXTRACE);
 568
 569         perf_header__clear_feat(&session->header, HEADER_STAT);
 570 }
 571
 572 static void
 573 record__finish_output(struct record *rec)
 574 {
 575         struct perf_data *data = &rec->data;
 576         int fd = perf_data__fd(data);
 577
 578         if (data->is_pipe)
 579                 return;
 580
 581         rec->session->header.data_size += rec->bytes_written;
 582         data->size = lseek(perf_data__fd(data), 0, SEEK_CUR);
 583
 584         if (!rec->no_buildid) {
 585                 process_buildids(rec);
 586
 587                 if (rec->buildid_all)
 588                         dsos__hit_all(rec->session);
 589         }
 590         perf_session__write_header(rec->session, rec->evlist, fd, true);
 591
 592         return;
 593 }
 594
 595 static int record__synthesize_workload(struct record *rec, bool tail)
 596 {
 597         int err;
 598         struct thread_map *thread_map;
 599
 600         if (rec->opts.tail_synthesize != tail)
 601                 return 0;
 602
 603         thread_map = thread_map__new_by_tid(rec->evlist->workload.pid);
 604         if (thread_map == NULL)
 605                 return -1;
 606
 607         err = perf_event__synthesize_thread_map(&rec->tool, thread_map,
 608                                                  process_synthesized_event,
 609                                                  &rec->session->machines.host,
 610                                                  rec->opts.sample_address,
 611                                                  rec->opts.proc_map_timeout);
 612         thread_map__put(thread_map);
 613         return err;
 614 }
 615
 616 static int record__synthesize(struct record *rec, bool tail);
 617
 618 static int
 619 record__switch_output(struct record *rec, bool at_exit)
 620 {
 621         struct perf_data *data = &rec->data;
 622         int fd, err;
 623
 624         /* Same Size:      "2015122520103046"*/
 625         char timestamp[] = "InvalidTimestamp";
 626
 627         record__synthesize(rec, true);
 628         if (target__none(&rec->opts.target))
 629                 record__synthesize_workload(rec, true);
 630
 631         rec->samples = 0;
 632         record__finish_output(rec);
 633         err = fetch_current_timestamp(timestamp, sizeof(timestamp));
 634         if (err) {
 635                 pr_err("Failed to get current timestamp\n");
 636                 return -EINVAL;
 637         }
 638
 639         fd = perf_data__switch(data, timestamp,
 640                                     rec->session->header.data_offset,
 641                                     at_exit);
 642         if (fd >= 0 && !at_exit) {
 643                 rec->bytes_written = 0;
 644                 rec->session->header.data_size = 0;
 645         }
 646
 647         if (!quiet)
 648                 fprintf(stderr, "[ perf record: Dump %s.%s ]\n",
 649                         data->file.path, timestamp);
 650
 651         /* Output tracking events */
 652         if (!at_exit) {
 653                 record__synthesize(rec, false);
 654
 655                 /*
 656                  * In 'perf record --switch-output' without -a,
 657                  * record__synthesize() in record__switch_output() won't
 658                  * generate tracking events because there's no thread_map
 659                  * in evlist. Which causes newly created perf.data doesn't
 660                  * contain map and comm information.
 661                  * Create a fake thread_map and directly call
 662                  * perf_event__synthesize_thread_map() for those events.
 663                  */
 664                 if (target__none(&rec->opts.target))
 665                         record__synthesize_workload(rec, false);
 666         }
 667         return fd;
 668 }
 669
 670 static volatile int workload_exec_errno;
 671
 672 /*
 673  * perf_evlist__prepare_workload will send a SIGUSR1
 674  * if the fork fails, since we asked by setting its
 675  * want_signal to true.
 676  */
 677 static void workload_exec_failed_signal(int signo __maybe_unused,
 678                                         siginfo_t *info,
 679                                         void *ucontext __maybe_unused)
 680 {
 681         workload_exec_errno = info->si_value.sival_int;
 682         done = 1;
 683         child_finished = 1;
 684 }
 685
 686 static void snapshot_sig_handler(int sig);
 687 static void alarm_sig_handler(int sig);
 688
 689 int __weak
 690 perf_event__synth_time_conv(const struct perf_event_mmap_page *pc __maybe_unused,
 691                             struct perf_tool *tool __maybe_unused,
 692                             perf_event__handler_t process __maybe_unused,
 693                             struct machine *machine __maybe_unused)
 694 {
 695         return 0;
 696 }
 697
 698 static const struct perf_event_mmap_page *
 699 perf_evlist__pick_pc(struct perf_evlist *evlist)
 700 {
 701         if (evlist) {
 702                 if (evlist->mmap && evlist->mmap[0].base)
 703                         return evlist->mmap[0].base;
 704                 if (evlist->overwrite_mmap && evlist->overwrite_mmap[0].base)
 705                         return evlist->overwrite_mmap[0].base;
 706         }
 707         return NULL;
 708 }
 709
 710 static const struct perf_event_mmap_page *record__pick_pc(struct record *rec)
 711 {
 712         const struct perf_event_mmap_page *pc;
 713
 714         pc = perf_evlist__pick_pc(rec->evlist);
 715         if (pc)
 716                 return pc;
 717         return NULL;
 718 }
 719
 720 static int record__synthesize(struct record *rec, bool tail)
 721 {
 722         struct perf_session *session = rec->session;
 723         struct machine *machine = &session->machines.host;
 724         struct perf_data *data = &rec->data;
 725         struct record_opts *opts = &rec->opts;
 726         struct perf_tool *tool = &rec->tool;
 727         int fd = perf_data__fd(data);
 728         int err = 0;
 729
 730         if (rec->opts.tail_synthesize != tail)
 731                 return 0;
 732
 733         if (data->is_pipe) {
 734                 err = perf_event__synthesize_features(
 735                         tool, session, rec->evlist, process_synthesized_event);
 736                 if (err < 0) {
 737                         pr_err("Couldn't synthesize features.\n");
 738                         return err;
 739                 }
 740
 741                 err = perf_event__synthesize_attrs(tool, session,
 742                                                    process_synthesized_event);
 743                 if (err < 0) {
 744                         pr_err("Couldn't synthesize attrs.\n");
 745                         goto out;
 746                 }
 747
 748                 if (have_tracepoints(&rec->evlist->entries)) {
 749                         /*
 750                          * FIXME err <= 0 here actually means that
 751                          * there were no tracepoints so its not really
 752                          * an error, just that we don't need to
 753                          * synthesize anything.  We really have to
 754                          * return this more properly and also
 755                          * propagate errors that now are calling die()
 756                          */
 757                         err = perf_event__synthesize_tracing_data(tool, fd, rec->evlist,
 758                                                                   process_synthesized_event);
 759                         if (err <= 0) {
 760                                 pr_err("Couldn't record tracing data.\n");
 761                                 goto out;
 762                         }
 763                         rec->bytes_written += err;
 764                 }
 765         }
 766
 767         err = perf_event__synth_time_conv(record__pick_pc(rec), tool,
 768                                           process_synthesized_event, machine);
 769         if (err)
 770                 goto out;
 771
 772         if (rec->opts.full_auxtrace) {
 773                 err = perf_event__synthesize_auxtrace_info(rec->itr, tool,
 774                                         session, process_synthesized_event);
 775                 if (err)
 776                         goto out;
 777         }
 778
 779         if (!perf_evlist__exclude_kernel(rec->evlist)) {
 780                 err = perf_event__synthesize_kernel_mmap(tool, process_synthesized_event,
 781                                                          machine);
 782                 WARN_ONCE(err < 0, "Couldn't record kernel reference relocation symbol\n"
 783                                    "Symbol resolution may be skewed if relocation was used (e.g. kexec).\n"
 784                                    "Check /proc/kallsyms permission or run as root.\n");
 785
 786                 err = perf_event__synthesize_modules(tool, process_synthesized_event,
 787                                                      machine);
 788                 WARN_ONCE(err < 0, "Couldn't record kernel module information.\n"
 789                                    "Symbol resolution may be skewed if relocation was used (e.g. kexec).\n"
 790                                    "Check /proc/modules permission or run as root.\n");
 791         }
 792
 793         if (perf_guest) {
 794                 machines__process_guests(&session->machines,
 795                                          perf_event__synthesize_guest_os, tool);
 796         }
 797
 798         err = perf_event__synthesize_extra_attr(&rec->tool,
 799                                                 rec->evlist,
 800                                                 process_synthesized_event,
 801                                                 data->is_pipe);
 802         if (err)
 803                 goto out;
 804
 805         err = perf_event__synthesize_thread_map2(&rec->tool, rec->evlist->threads,
 806                                                  process_synthesized_event,
 807                                                 NULL);
 808         if (err < 0) {
 809                 pr_err("Couldn't synthesize thread map.\n");
 810                 return err;
 811         }
 812
 813         err = perf_event__synthesize_cpu_map(&rec->tool, rec->evlist->cpus,
 814                                              process_synthesized_event, NULL);
 815         if (err < 0) {
 816                 pr_err("Couldn't synthesize cpu map.\n");
 817                 return err;
 818         }
 819
 820         err = __machine__synthesize_threads(machine, tool, &opts->target, rec->evlist->threads,
 821                                             process_synthesized_event, opts->sample_address,
 822                                             opts->proc_map_timeout, 1);
 823 out:
 824         return err;
 825 }
 826
 827 static int __cmd_record(struct record *rec, int argc, const char **argv)
 828 {
 829         int err;
 830         int status = 0;
 831         unsigned long waking = 0;
 832         const bool forks = argc > 0;
 833         struct machine *machine;
 834         struct perf_tool *tool = &rec->tool;
 835         struct record_opts *opts = &rec->opts;
 836         struct perf_data *data = &rec->data;
 837         struct perf_session *session;
 838         bool disabled = false, draining = false;
 839         int fd;
 840
 841         rec->progname = argv[0];
 842
 843         atexit(record__sig_exit);
 844         signal(SIGCHLD, sig_handler);
 845         signal(SIGINT, sig_handler);
 846         signal(SIGTERM, sig_handler);
 847         signal(SIGSEGV, sigsegv_handler);
 848
 849         if (rec->opts.record_namespaces)
 850                 tool->namespace_events = true;
 851
 852         if (rec->opts.auxtrace_snapshot_mode || rec->switch_output.enabled) {
 853                 signal(SIGUSR2, snapshot_sig_handler);
 854                 if (rec->opts.auxtrace_snapshot_mode)
 855                         trigger_on(&auxtrace_snapshot_trigger);
 856                 if (rec->switch_output.enabled)
 857                         trigger_on(&switch_output_trigger);
 858         } else {
 859                 signal(SIGUSR2, SIG_IGN);
 860         }
 861
 862         session = perf_session__new(data, false, tool);
 863         if (session == NULL) {
 864                 pr_err("Perf session creation failed.\n");
 865                 return -1;
 866         }
 867
 868         fd = perf_data__fd(data);
 869         rec->session = session;
 870
 871         record__init_features(rec);
 872
 873         if (forks) {
 874                 err = perf_evlist__prepare_workload(rec->evlist, &opts->target,
 875                                                     argv, data->is_pipe,
 876                                                     workload_exec_failed_signal);
 877                 if (err < 0) {
 878                         pr_err("Couldn't run the workload!\n");
 879                         status = err;
 880                         goto out_delete_session;
 881                 }
 882         }
 883
 884         /*
 885          * If we have just single event and are sending data
 886          * through pipe, we need to force the ids allocation,
 887          * because we synthesize event name through the pipe
 888          * and need the id for that.
 889          */
 890         if (data->is_pipe && rec->evlist->nr_entries == 1)
 891                 rec->opts.sample_id = true;
 892
 893         if (record__open(rec) != 0) {
 894                 err = -1;
 895                 goto out_child;
 896         }
 897
 898         err = bpf__apply_obj_config();
 899         if (err) {
 900                 char errbuf[BUFSIZ];
 901
 902                 bpf__strerror_apply_obj_config(err, errbuf, sizeof(errbuf));
 903                 pr_err("ERROR: Apply config to BPF failed: %s\n",
 904                          errbuf);
 905                 goto out_child;
 906         }
 907
 908         /*
 909          * Normally perf_session__new would do this, but it doesn't have the
 910          * evlist.
 911          */
 912         if (rec->tool.ordered_events && !perf_evlist__sample_id_all(rec->evlist)) {
 913                 pr_warning("WARNING: No sample_id_all support, falling back to unordered processing\n");
 914                 rec->tool.ordered_events = false;
 915         }
 916
 917         if (!rec->evlist->nr_groups)
 918                 perf_header__clear_feat(&session->header, HEADER_GROUP_DESC);
 919
 920         if (data->is_pipe) {
 921                 err = perf_header__write_pipe(fd);
 922                 if (err < 0)
 923                         goto out_child;
 924         } else {
 925                 err = perf_session__write_header(session, rec->evlist, fd, false);
 926                 if (err < 0)
 927                         goto out_child;
 928         }
 929
 930         if (!rec->no_buildid
 931             && !perf_header__has_feat(&session->header, HEADER_BUILD_ID)) {
 932                 pr_err("Couldn't generate buildids. "
 933                        "Use --no-buildid to profile anyway.\n");
 934                 err = -1;
 935                 goto out_child;
 936         }
 937
 938         machine = &session->machines.host;
 939
 940         err = record__synthesize(rec, false);
 941         if (err < 0)
 942                 goto out_child;
 943
 944         if (rec->realtime_prio) {
 945                 struct sched_param param;
 946
 947                 param.sched_priority = rec->realtime_prio;
 948                 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
 949                         pr_err("Could not set realtime priority.\n");
 950                         err = -1;
 951                         goto out_child;
 952                 }
 953         }
 954
 955         /*
 956          * When perf is starting the traced process, all the events
 957          * (apart from group members) have enable_on_exec=1 set,
 958          * so don't spoil it by prematurely enabling them.
 959          */
 960         if (!target__none(&opts->target) && !opts->initial_delay)
 961                 perf_evlist__enable(rec->evlist);
 962
 963         /*
 964          * Let the child rip
 965          */
 966         if (forks) {
 967                 union perf_event *event;
 968                 pid_t tgid;
 969
 970                 event = malloc(sizeof(event->comm) + machine->id_hdr_size);
 971                 if (event == NULL) {
 972                         err = -ENOMEM;
 973                         goto out_child;
 974                 }
 975
 976                 /*
 977                  * Some H/W events are generated before COMM event
 978                  * which is emitted during exec(), so perf script
 979                  * cannot see a correct process name for those events.
 980                  * Synthesize COMM event to prevent it.
 981                  */
 982                 tgid = perf_event__synthesize_comm(tool, event,
 983                                                    rec->evlist->workload.pid,
 984                                                    process_synthesized_event,
 985                                                    machine);
 986                 free(event);
 987
 988                 if (tgid == -1)
 989                         goto out_child;
 990
 991                 event = malloc(sizeof(event->namespaces) +
 992                                (NR_NAMESPACES * sizeof(struct perf_ns_link_info)) +
 993                                machine->id_hdr_size);
 994                 if (event == NULL) {
 995                         err = -ENOMEM;
 996                         goto out_child;
 997                 }
 998
 999                 /*
1000                  * Synthesize NAMESPACES event for the command specified.
1001                  */
1002                 perf_event__synthesize_namespaces(tool, event,
1003                                                   rec->evlist->workload.pid,
1004                                                   tgid, process_synthesized_event,
1005                                                   machine);
1006                 free(event);
1007
1008                 perf_evlist__start_workload(rec->evlist);
1009         }
1010
1011         if (opts->initial_delay) {
1012                 usleep(opts->initial_delay * USEC_PER_MSEC);
1013                 perf_evlist__enable(rec->evlist);
1014         }
1015
1016         trigger_ready(&auxtrace_snapshot_trigger);
1017         trigger_ready(&switch_output_trigger);
1018         perf_hooks__invoke_record_start();
1019         for (;;) {
1020                 unsigned long long hits = rec->samples;
1021
1022                 /*
1023                  * rec->evlist->bkw_mmap_state is possible to be
1024                  * BKW_MMAP_EMPTY here: when done == true and
1025                  * hits != rec->samples in previous round.
1026                  *
1027                  * perf_evlist__toggle_bkw_mmap ensure we never
1028                  * convert BKW_MMAP_EMPTY to BKW_MMAP_DATA_PENDING.
1029                  */
1030                 if (trigger_is_hit(&switch_output_trigger) || done || draining)
1031                         perf_evlist__toggle_bkw_mmap(rec->evlist, BKW_MMAP_DATA_PENDING);
1032
1033                 if (record__mmap_read_all(rec) < 0) {
1034                         trigger_error(&auxtrace_snapshot_trigger);
1035                         trigger_error(&switch_output_trigger);
1036                         err = -1;
1037                         goto out_child;
1038                 }
1039
1040                 if (auxtrace_record__snapshot_started) {
1041                         auxtrace_record__snapshot_started = 0;
1042                         if (!trigger_is_error(&auxtrace_snapshot_trigger))
1043                                 record__read_auxtrace_snapshot(rec);
1044                         if (trigger_is_error(&auxtrace_snapshot_trigger)) {
1045                                 pr_err("AUX area tracing snapshot failed\n");
1046                                 err = -1;
1047                                 goto out_child;
1048                         }
1049                 }
1050
1051                 if (trigger_is_hit(&switch_output_trigger)) {
1052                         /*
1053                          * If switch_output_trigger is hit, the data in
1054                          * overwritable ring buffer should have been collected,
1055                          * so bkw_mmap_state should be set to BKW_MMAP_EMPTY.
1056                          *
1057                          * If SIGUSR2 raise after or during record__mmap_read_all(),
1058                          * record__mmap_read_all() didn't collect data from
1059                          * overwritable ring buffer. Read again.
1060                          */
1061                         if (rec->evlist->bkw_mmap_state == BKW_MMAP_RUNNING)
1062                                 continue;
1063                         trigger_ready(&switch_output_trigger);
1064
1065                         /*
1066                          * Reenable events in overwrite ring buffer after
1067                          * record__mmap_read_all(): we should have collected
1068                          * data from it.
1069                          */
1070                         perf_evlist__toggle_bkw_mmap(rec->evlist, BKW_MMAP_RUNNING);
1071
1072                         if (!quiet)
1073                                 fprintf(stderr, "[ perf record: dump data: Woken up %ld times ]\n",
1074                                         waking);
1075                         waking = 0;
1076                         fd = record__switch_output(rec, false);
1077                         if (fd < 0) {
1078                                 pr_err("Failed to switch to new file\n");
1079                                 trigger_error(&switch_output_trigger);
1080                                 err = fd;
1081                                 goto out_child;
1082                         }
1083
1084                         /* re-arm the alarm */
1085                         if (rec->switch_output.time)
1086                                 alarm(rec->switch_output.time);
1087                 }
1088
1089                 if (hits == rec->samples) {
1090                         if (done || draining)
1091                                 break;
1092                         err = perf_evlist__poll(rec->evlist, -1);
1093                         /*
1094                          * Propagate error, only if there's any. Ignore positive
1095                          * number of returned events and interrupt error.
1096                          */
1097                         if (err > 0 || (err < 0 && errno == EINTR))
1098                                 err = 0;
1099                         waking++;
1100
1101                         if (perf_evlist__filter_pollfd(rec->evlist, POLLERR | POLLHUP) == 0)
1102                                 draining = true;
1103                 }
1104
1105                 /*
1106                  * When perf is starting the traced process, at the end events
1107                  * die with the process and we wait for that. Thus no need to
1108                  * disable events in this case.
1109                  */
1110                 if (done && !disabled && !target__none(&opts->target)) {
1111                         trigger_off(&auxtrace_snapshot_trigger);
1112                         perf_evlist__disable(rec->evlist);
1113                         disabled = true;
1114                 }
1115         }
1116         trigger_off(&auxtrace_snapshot_trigger);
1117         trigger_off(&switch_output_trigger);
1118
1119         if (forks && workload_exec_errno) {
1120                 char msg[STRERR_BUFSIZE];
1121                 const char *emsg = str_error_r(workload_exec_errno, msg, sizeof(msg));
1122                 pr_err("Workload failed: %s\n", emsg);
1123                 err = -1;
1124                 goto out_child;
1125         }
1126
1127         if (!quiet)
1128                 fprintf(stderr, "[ perf record: Woken up %ld times to write data ]\n", waking);
1129
1130         if (target__none(&rec->opts.target))
1131                 record__synthesize_workload(rec, true);
1132
1133 out_child:
1134         if (forks) {
1135                 int exit_status;
1136
1137                 if (!child_finished)
1138                         kill(rec->evlist->workload.pid, SIGTERM);
1139
1140                 wait(&exit_status);
1141
1142                 if (err < 0)
1143                         status = err;
1144                 else if (WIFEXITED(exit_status))
1145                         status = WEXITSTATUS(exit_status);
1146                 else if (WIFSIGNALED(exit_status))
1147                         signr = WTERMSIG(exit_status);
1148         } else
1149                 status = err;
1150
1151         record__synthesize(rec, true);
1152         /* this will be recalculated during process_buildids() */
1153         rec->samples = 0;
1154
1155         if (!err) {
1156                 if (!rec->timestamp_filename) {
1157                         record__finish_output(rec);
1158                 } else {
1159                         fd = record__switch_output(rec, true);
1160                         if (fd < 0) {
1161                                 status = fd;
1162                                 goto out_delete_session;
1163                         }
1164                 }
1165         }
1166
1167         perf_hooks__invoke_record_end();
1168
1169         if (!err && !quiet) {
1170                 char samples[128];
1171                 const char *postfix = rec->timestamp_filename ?
1172                                         ".<timestamp>" : "";
1173
1174                 if (rec->samples && !rec->opts.full_auxtrace)
1175                         scnprintf(samples, sizeof(samples),
1176                                   " (%" PRIu64 " samples)", rec->samples);
1177                 else
1178                         samples[0] = '\0';
1179
1180                 fprintf(stderr, "[ perf record: Captured and wrote %.3f MB %s%s%s ]\n",
1181                         perf_data__size(data) / 1024.0 / 1024.0,
1182                         data->file.path, postfix, samples);
1183         }
1184
1185 out_delete_session:
1186         perf_session__delete(session);
1187         return status;
1188 }
1189
1190 static void callchain_debug(struct callchain_param *callchain)
1191 {
1192         static const char *str[CALLCHAIN_MAX] = { "NONE", "FP", "DWARF", "LBR" };
1193
1194         pr_debug("callchain: type %s\n", str[callchain->record_mode]);
1195
1196         if (callchain->record_mode == CALLCHAIN_DWARF)
1197                 pr_debug("callchain: stack dump size %d\n",
1198                          callchain->dump_size);
1199 }
1200
1201 int record_opts__parse_callchain(struct record_opts *record,
1202                                  struct callchain_param *callchain,
1203                                  const char *arg, bool unset)
1204 {
1205         int ret;
1206         callchain->enabled = !unset;
1207
1208         /* --no-call-graph */
1209         if (unset) {
1210                 callchain->record_mode = CALLCHAIN_NONE;
1211                 pr_debug("callchain: disabled\n");
1212                 return 0;
1213         }
1214
1215         ret = parse_callchain_record_opt(arg, callchain);
1216         if (!ret) {
1217                 /* Enable data address sampling for DWARF unwind. */
1218                 if (callchain->record_mode == CALLCHAIN_DWARF)
1219                         record->sample_address = true;
1220                 callchain_debug(callchain);
1221         }
1222
1223         return ret;
1224 }
1225
1226 int record_parse_callchain_opt(const struct option *opt,
1227                                const char *arg,
1228                                int unset)
1229 {
1230         return record_opts__parse_callchain(opt->value, &callchain_param, arg, unset);
1231 }
1232
1233 int record_callchain_opt(const struct option *opt,
1234                          const char *arg __maybe_unused,
1235                          int unset __maybe_unused)
1236 {
1237         struct callchain_param *callchain = opt->value;
1238
1239         callchain->enabled = true;
1240
1241         if (callchain->record_mode == CALLCHAIN_NONE)
1242                 callchain->record_mode = CALLCHAIN_FP;
1243
1244         callchain_debug(callchain);
1245         return 0;
1246 }
1247
1248 static int perf_record_config(const char *var, const char *value, void *cb)
1249 {
1250         struct record *rec = cb;
1251
1252         if (!strcmp(var, "record.build-id")) {
1253                 if (!strcmp(value, "cache"))
1254                         rec->no_buildid_cache = false;
1255                 else if (!strcmp(value, "no-cache"))
1256                         rec->no_buildid_cache = true;
1257                 else if (!strcmp(value, "skip"))
1258                         rec->no_buildid = true;
1259                 else
1260                         return -1;
1261                 return 0;
1262         }
1263         if (!strcmp(var, "record.call-graph"))
1264                 var = "call-graph.record-mode"; /* fall-through */
1265
1266         return perf_default_config(var, value, cb);
1267 }
1268
1269 struct clockid_map {
1270         const char *name;
1271         int clockid;
1272 };
1273
1274 #define CLOCKID_MAP(n, c)       \
1275         { .name = n, .clockid = (c), }
1276
1277 #define CLOCKID_END     { .name = NULL, }
1278
1279
1280 /*
1281  * Add the missing ones, we need to build on many distros...
1282  */
1283 #ifndef CLOCK_MONOTONIC_RAW
1284 #define CLOCK_MONOTONIC_RAW 4
1285 #endif
1286 #ifndef CLOCK_BOOTTIME
1287 #define CLOCK_BOOTTIME 7
1288 #endif
1289 #ifndef CLOCK_TAI
1290 #define CLOCK_TAI 11
1291 #endif
1292
1293 static const struct clockid_map clockids[] = {
1294         /* available for all events, NMI safe */
1295         CLOCKID_MAP("monotonic", CLOCK_MONOTONIC),
1296         CLOCKID_MAP("monotonic_raw", CLOCK_MONOTONIC_RAW),
1297
1298         /* available for some events */
1299         CLOCKID_MAP("realtime", CLOCK_REALTIME),
1300         CLOCKID_MAP("boottime", CLOCK_BOOTTIME),
1301         CLOCKID_MAP("tai", CLOCK_TAI),
1302
1303         /* available for the lazy */
1304         CLOCKID_MAP("mono", CLOCK_MONOTONIC),
1305         CLOCKID_MAP("raw", CLOCK_MONOTONIC_RAW),
1306         CLOCKID_MAP("real", CLOCK_REALTIME),
1307         CLOCKID_MAP("boot", CLOCK_BOOTTIME),
1308
1309         CLOCKID_END,
1310 };
1311
1312 static int parse_clockid(const struct option *opt, const char *str, int unset)
1313 {
1314         struct record_opts *opts = (struct record_opts *)opt->value;
1315         const struct clockid_map *cm;
1316         const char *ostr = str;
1317
1318         if (unset) {
1319                 opts->use_clockid = 0;
1320                 return 0;
1321         }
1322
1323         /* no arg passed */
1324         if (!str)
1325                 return 0;
1326
1327         /* no setting it twice */
1328         if (opts->use_clockid)
1329                 return -1;
1330
1331         opts->use_clockid = true;
1332
1333         /* if its a number, we're done */
1334         if (sscanf(str, "%d", &opts->clockid) == 1)
1335                 return 0;
1336
1337         /* allow a "CLOCK_" prefix to the name */
1338         if (!strncasecmp(str, "CLOCK_", 6))
1339                 str += 6;
1340
1341         for (cm = clockids; cm->name; cm++) {
1342                 if (!strcasecmp(str, cm->name)) {
1343                         opts->clockid = cm->clockid;
1344                         return 0;
1345                 }
1346         }
1347
1348         opts->use_clockid = false;
1349         ui__warning("unknown clockid %s, check man page\n", ostr);
1350         return -1;
1351 }
1352
1353 static int record__parse_mmap_pages(const struct option *opt,
1354                                     const char *str,
1355                                     int unset __maybe_unused)
1356 {
1357         struct record_opts *opts = opt->value;
1358         char *s, *p;
1359         unsigned int mmap_pages;
1360         int ret;
1361
1362         if (!str)
1363                 return -EINVAL;
1364
1365         s = strdup(str);
1366         if (!s)
1367                 return -ENOMEM;
1368
1369         p = strchr(s, ',');
1370         if (p)
1371                 *p = '\0';
1372
1373         if (*s) {
1374                 ret = __perf_evlist__parse_mmap_pages(&mmap_pages, s);
1375                 if (ret)
1376                         goto out_free;
1377                 opts->mmap_pages = mmap_pages;
1378         }
1379
1380         if (!p) {
1381                 ret = 0;
1382                 goto out_free;
1383         }
1384
1385         ret = __perf_evlist__parse_mmap_pages(&mmap_pages, p + 1);
1386         if (ret)
1387                 goto out_free;
1388
1389         opts->auxtrace_mmap_pages = mmap_pages;
1390
1391 out_free:
1392         free(s);
1393         return ret;
1394 }
1395
1396 static void switch_output_size_warn(struct record *rec)
1397 {
1398         u64 wakeup_size = perf_evlist__mmap_size(rec->opts.mmap_pages);
1399         struct switch_output *s = &rec->switch_output;
1400
1401         wakeup_size /= 2;
1402
1403         if (s->size < wakeup_size) {
1404                 char buf[100];
1405
1406                 unit_number__scnprintf(buf, sizeof(buf), wakeup_size);
1407                 pr_warning("WARNING: switch-output data size lower than "
1408                            "wakeup kernel buffer size (%s) "
1409                            "expect bigger perf.data sizes\n", buf);
1410         }
1411 }
1412
1413 static int switch_output_setup(struct record *rec)
1414 {
1415         struct switch_output *s = &rec->switch_output;
1416         static struct parse_tag tags_size[] = {
1417                 { .tag  = 'B', .mult = 1       },
1418                 { .tag  = 'K', .mult = 1 << 10 },
1419                 { .tag  = 'M', .mult = 1 << 20 },
1420                 { .tag  = 'G', .mult = 1 << 30 },
1421                 { .tag  = 0 },
1422         };
1423         static struct parse_tag tags_time[] = {
1424                 { .tag  = 's', .mult = 1        },
1425                 { .tag  = 'm', .mult = 60       },
1426                 { .tag  = 'h', .mult = 60*60    },
1427                 { .tag  = 'd', .mult = 60*60*24 },
1428                 { .tag  = 0 },
1429         };
1430         unsigned long val;
1431
1432         if (!s->set)
1433                 return 0;
1434
1435         if (!strcmp(s->str, "signal")) {
1436                 s->signal = true;
1437                 pr_debug("switch-output with SIGUSR2 signal\n");
1438                 goto enabled;
1439         }
1440
1441         val = parse_tag_value(s->str, tags_size);
1442         if (val != (unsigned long) -1) {
1443                 s->size = val;
1444                 pr_debug("switch-output with %s size threshold\n", s->str);
1445                 goto enabled;
1446         }
1447
1448         val = parse_tag_value(s->str, tags_time);
1449         if (val != (unsigned long) -1) {
1450                 s->time = val;
1451                 pr_debug("switch-output with %s time threshold (%lu seconds)\n",
1452                          s->str, s->time);
1453                 goto enabled;
1454         }
1455
1456         return -1;
1457
1458 enabled:
1459         rec->timestamp_filename = true;
1460         s->enabled              = true;
1461
1462         if (s->size && !rec->opts.no_buffering)
1463                 switch_output_size_warn(rec);
1464
1465         return 0;
1466 }
1467
1468 static const char * const __record_usage[] = {
1469         "perf record [<options>] [<command>]",
1470         "perf record [<options>] -- <command> [<options>]",
1471         NULL
1472 };
1473 const char * const *record_usage = __record_usage;
1474
1475 /*
1476  * XXX Ideally would be local to cmd_record() and passed to a record__new
1477  * because we need to have access to it in record__exit, that is called
1478  * after cmd_record() exits, but since record_options need to be accessible to
1479  * builtin-script, leave it here.
1480  *
1481  * At least we don't ouch it in all the other functions here directly.
1482  *
1483  * Just say no to tons of global variables, sigh.
1484  */
1485 static struct record record = {
1486         .opts = {
1487                 .sample_time         = true,
1488                 .mmap_pages          = UINT_MAX,
1489                 .user_freq           = UINT_MAX,
1490                 .user_interval       = ULLONG_MAX,
1491                 .freq                = 4000,
1492                 .target              = {
1493                         .uses_mmap   = true,
1494                         .default_per_cpu = true,
1495                 },
1496                 .proc_map_timeout     = 500,
1497         },
1498         .tool = {
1499                 .sample         = process_sample_event,
1500                 .fork           = perf_event__process_fork,
1501                 .exit           = perf_event__process_exit,
1502                 .comm           = perf_event__process_comm,
1503                 .namespaces     = perf_event__process_namespaces,
1504                 .mmap           = perf_event__process_mmap,
1505                 .mmap2          = perf_event__process_mmap2,
1506                 .ordered_events = true,
1507         },
1508 };
1509
1510 const char record_callchain_help[] = CALLCHAIN_RECORD_HELP
1511         "\n\t\t\t\tDefault: fp";
1512
1513 static bool dry_run;
1514
1515 /*
1516  * XXX Will stay a global variable till we fix builtin-script.c to stop messing
1517  * with it and switch to use the library functions in perf_evlist that came
1518  * from builtin-record.c, i.e. use record_opts,
1519  * perf_evlist__prepare_workload, etc instead of fork+exec'in 'perf record',
1520  * using pipes, etc.
1521  */
1522 static struct option __record_options[] = {
1523         OPT_CALLBACK('e', "event", &record.evlist, "event",
1524                      "event selector. use 'perf list' to list available events",
1525                      parse_events_option),
1526         OPT_CALLBACK(0, "filter", &record.evlist, "filter",
1527                      "event filter", parse_filter),
1528         OPT_CALLBACK_NOOPT(0, "exclude-perf", &record.evlist,
1529                            NULL, "don't record events from perf itself",
1530                            exclude_perf),
1531         OPT_STRING('p', "pid", &record.opts.target.pid, "pid",
1532                     "record events on existing process id"),
1533         OPT_STRING('t', "tid", &record.opts.target.tid, "tid",
1534                     "record events on existing thread id"),
1535         OPT_INTEGER('r', "realtime", &record.realtime_prio,
1536                     "collect data with this RT SCHED_FIFO priority"),
1537         OPT_BOOLEAN(0, "no-buffering", &record.opts.no_buffering,
1538                     "collect data without buffering"),
1539         OPT_BOOLEAN('R', "raw-samples", &record.opts.raw_samples,
1540                     "collect raw sample records from all opened counters"),
1541         OPT_BOOLEAN('a', "all-cpus", &record.opts.target.system_wide,
1542                             "system-wide collection from all CPUs"),
1543         OPT_STRING('C', "cpu", &record.opts.target.cpu_list, "cpu",
1544                     "list of cpus to monitor"),
1545         OPT_U64('c', "count", &record.opts.user_interval, "event period to sample"),
1546         OPT_STRING('o', "output", &record.data.file.path, "file",
1547                     "output file name"),
1548         OPT_BOOLEAN_SET('i', "no-inherit", &record.opts.no_inherit,
1549                         &record.opts.no_inherit_set,
1550                         "child tasks do not inherit counters"),
1551         OPT_BOOLEAN(0, "tail-synthesize", &record.opts.tail_synthesize,
1552                     "synthesize non-sample events at the end of output"),
1553         OPT_BOOLEAN(0, "overwrite", &record.opts.overwrite, "use overwrite mode"),
1554         OPT_UINTEGER('F', "freq", &record.opts.user_freq, "profile at this frequency"),
1555         OPT_CALLBACK('m', "mmap-pages", &record.opts, "pages[,pages]",
1556                      "number of mmap data pages and AUX area tracing mmap pages",
1557                      record__parse_mmap_pages),
1558         OPT_BOOLEAN(0, "group", &record.opts.group,
1559                     "put the counters into a counter group"),
1560         OPT_CALLBACK_NOOPT('g', NULL, &callchain_param,
1561                            NULL, "enables call-graph recording" ,
1562                            &record_callchain_opt),
1563         OPT_CALLBACK(0, "call-graph", &record.opts,
1564                      "record_mode[,record_size]", record_callchain_help,
1565                      &record_parse_callchain_opt),
1566         OPT_INCR('v', "verbose", &verbose,
1567                     "be more verbose (show counter open errors, etc)"),
1568         OPT_BOOLEAN('q', "quiet", &quiet, "don't print any message"),
1569         OPT_BOOLEAN('s', "stat", &record.opts.inherit_stat,
1570                     "per thread counts"),
1571         OPT_BOOLEAN('d', "data", &record.opts.sample_address, "Record the sample addresses"),
1572         OPT_BOOLEAN(0, "phys-data", &record.opts.sample_phys_addr,
1573                     "Record the sample physical addresses"),
1574         OPT_BOOLEAN(0, "sample-cpu", &record.opts.sample_cpu, "Record the sample cpu"),
1575         OPT_BOOLEAN_SET('T', "timestamp", &record.opts.sample_time,
1576                         &record.opts.sample_time_set,
1577                         "Record the sample timestamps"),
1578         OPT_BOOLEAN_SET('P', "period", &record.opts.period, &record.opts.period_set,
1579                         "Record the sample period"),
1580         OPT_BOOLEAN('n', "no-samples", &record.opts.no_samples,
1581                     "don't sample"),
1582         OPT_BOOLEAN_SET('N', "no-buildid-cache", &record.no_buildid_cache,
1583                         &record.no_buildid_cache_set,
1584                         "do not update the buildid cache"),
1585         OPT_BOOLEAN_SET('B', "no-buildid", &record.no_buildid,
1586                         &record.no_buildid_set,
1587                         "do not collect buildids in perf.data"),
1588         OPT_CALLBACK('G', "cgroup", &record.evlist, "name",
1589                      "monitor event in cgroup name only",
1590                      parse_cgroups),
1591         OPT_UINTEGER('D', "delay", &record.opts.initial_delay,
1592                   "ms to wait before starting measurement after program start"),
1593         OPT_STRING('u', "uid", &record.opts.target.uid_str, "user",
1594                    "user to profile"),
1595
1596         OPT_CALLBACK_NOOPT('b', "branch-any", &record.opts.branch_stack,
1597                      "branch any", "sample any taken branches",
1598                      parse_branch_stack),
1599
1600         OPT_CALLBACK('j', "branch-filter", &record.opts.branch_stack,
1601                      "branch filter mask", "branch stack filter modes",
1602                      parse_branch_stack),
1603         OPT_BOOLEAN('W', "weight", &record.opts.sample_weight,
1604                     "sample by weight (on special events only)"),
1605         OPT_BOOLEAN(0, "transaction", &record.opts.sample_transaction,
1606                     "sample transaction flags (special events only)"),
1607         OPT_BOOLEAN(0, "per-thread", &record.opts.target.per_thread,
1608                     "use per-thread mmaps"),
1609         OPT_CALLBACK_OPTARG('I', "intr-regs", &record.opts.sample_intr_regs, NULL, "any register",
1610                     "sample selected machine registers on interrupt,"
1611                     " use -I ? to list register names", parse_regs),
1612         OPT_CALLBACK_OPTARG(0, "user-regs", &record.opts.sample_user_regs, NULL, "any register",
1613                     "sample selected machine registers on interrupt,"
1614                     " use -I ? to list register names", parse_regs),
1615         OPT_BOOLEAN(0, "running-time", &record.opts.running_time,
1616                     "Record running/enabled time of read (:S) events"),
1617         OPT_CALLBACK('k', "clockid", &record.opts,
1618         "clockid", "clockid to use for events, see clock_gettime()",
1619         parse_clockid),
1620         OPT_STRING_OPTARG('S', "snapshot", &record.opts.auxtrace_snapshot_opts,
1621                           "opts", "AUX area tracing Snapshot Mode", ""),
1622         OPT_UINTEGER(0, "proc-map-timeout", &record.opts.proc_map_timeout,
1623                         "per thread proc mmap processing timeout in ms"),
1624         OPT_BOOLEAN(0, "namespaces", &record.opts.record_namespaces,
1625                     "Record namespaces events"),
1626         OPT_BOOLEAN(0, "switch-events", &record.opts.record_switch_events,
1627                     "Record context switch events"),
1628         OPT_BOOLEAN_FLAG(0, "all-kernel", &record.opts.all_kernel,
1629                          "Configure all used events to run in kernel space.",
1630                          PARSE_OPT_EXCLUSIVE),
1631         OPT_BOOLEAN_FLAG(0, "all-user", &record.opts.all_user,
1632                          "Configure all used events to run in user space.",
1633                          PARSE_OPT_EXCLUSIVE),
1634         OPT_STRING(0, "clang-path", &llvm_param.clang_path, "clang path",
1635                    "clang binary to use for compiling BPF scriptlets"),
1636         OPT_STRING(0, "clang-opt", &llvm_param.clang_opt, "clang options",
1637                    "options passed to clang when compiling BPF scriptlets"),
1638         OPT_STRING(0, "vmlinux", &symbol_conf.vmlinux_name,
1639                    "file", "vmlinux pathname"),
1640         OPT_BOOLEAN(0, "buildid-all", &record.buildid_all,
1641                     "Record build-id of all DSOs regardless of hits"),
1642         OPT_BOOLEAN(0, "timestamp-filename", &record.timestamp_filename,
1643                     "append timestamp to output filename"),
1644         OPT_BOOLEAN(0, "timestamp-boundary", &record.timestamp_boundary,
1645                     "Record timestamp boundary (time of first/last samples)"),
1646         OPT_STRING_OPTARG_SET(0, "switch-output", &record.switch_output.str,
1647                           &record.switch_output.set, "signal,size,time",
1648                           "Switch output when receive SIGUSR2 or cross size,time threshold",
1649                           "signal"),
1650         OPT_BOOLEAN(0, "dry-run", &dry_run,
1651                     "Parse options then exit"),
1652         OPT_END()
1653 };
1654
1655 struct option *record_options = __record_options;
1656
1657 int cmd_record(int argc, const char **argv)
1658 {
1659         int err;
1660         struct record *rec = &record;
1661         char errbuf[BUFSIZ];
1662
1663 #ifndef HAVE_LIBBPF_SUPPORT
1664 # define set_nobuild(s, l, c) set_option_nobuild(record_options, s, l, "NO_LIBBPF=1", c)
1665         set_nobuild('\0', "clang-path", true);
1666         set_nobuild('\0', "clang-opt", true);
1667 # undef set_nobuild
1668 #endif
1669
1670 #ifndef HAVE_BPF_PROLOGUE
1671 # if !defined (HAVE_DWARF_SUPPORT)
1672 #  define REASON  "NO_DWARF=1"
1673 # elif !defined (HAVE_LIBBPF_SUPPORT)
1674 #  define REASON  "NO_LIBBPF=1"
1675 # else
1676 #  define REASON  "this architecture doesn't support BPF prologue"
1677 # endif
1678 # define set_nobuild(s, l, c) set_option_nobuild(record_options, s, l, REASON, c)
1679         set_nobuild('\0', "vmlinux", true);
1680 # undef set_nobuild
1681 # undef REASON
1682 #endif
1683
1684         rec->evlist = perf_evlist__new();
1685         if (rec->evlist == NULL)
1686                 return -ENOMEM;
1687
1688         err = perf_config(perf_record_config, rec);
1689         if (err)
1690                 return err;
1691
1692         argc = parse_options(argc, argv, record_options, record_usage,
1693                             PARSE_OPT_STOP_AT_NON_OPTION);
1694         if (quiet)
1695                 perf_quiet_option();
1696
1697         /* Make system wide (-a) the default target. */
1698         if (!argc && target__none(&rec->opts.target))
1699                 rec->opts.target.system_wide = true;
1700
1701         if (nr_cgroups && !rec->opts.target.system_wide) {
1702                 usage_with_options_msg(record_usage, record_options,
1703                         "cgroup monitoring only available in system-wide mode");
1704
1705         }
1706         if (rec->opts.record_switch_events &&
1707             !perf_can_record_switch_events()) {
1708                 ui__error("kernel does not support recording context switch events\n");
1709                 parse_options_usage(record_usage, record_options, "switch-events", 0);
1710                 return -EINVAL;
1711         }
1712
1713         if (switch_output_setup(rec)) {
1714                 parse_options_usage(record_usage, record_options, "switch-output", 0);
1715                 return -EINVAL;
1716         }
1717
1718         if (rec->switch_output.time) {
1719                 signal(SIGALRM, alarm_sig_handler);
1720                 alarm(rec->switch_output.time);
1721         }
1722
1723         if (!rec->itr) {
1724                 rec->itr = auxtrace_record__init(rec->evlist, &err);
1725                 if (err)
1726                         goto out;
1727         }
1728
1729         err = auxtrace_parse_snapshot_options(rec->itr, &rec->opts,
1730                                               rec->opts.auxtrace_snapshot_opts);
1731         if (err)
1732                 goto out;
1733
1734         /*
1735          * Allow aliases to facilitate the lookup of symbols for address
1736          * filters. Refer to auxtrace_parse_filters().
1737          */
1738         symbol_conf.allow_aliases = true;
1739
1740         symbol__init(NULL);
1741
1742         err = auxtrace_parse_filters(rec->evlist);
1743         if (err)
1744                 goto out;
1745
1746         if (dry_run)
1747                 goto out;
1748
1749         err = bpf__setup_stdout(rec->evlist);
1750         if (err) {
1751                 bpf__strerror_setup_stdout(rec->evlist, err, errbuf, sizeof(errbuf));
1752                 pr_err("ERROR: Setup BPF stdout failed: %s\n",
1753                          errbuf);
1754                 goto out;
1755         }
1756
1757         err = -ENOMEM;
1758
1759         if (symbol_conf.kptr_restrict && !perf_evlist__exclude_kernel(rec->evlist))
1760                 pr_warning(
1761 "WARNING: Kernel address maps (/proc/{kallsyms,modules}) are restricted,\n"
1762 "check /proc/sys/kernel/kptr_restrict.\n\n"
1763 "Samples in kernel functions may not be resolved if a suitable vmlinux\n"
1764 "file is not found in the buildid cache or in the vmlinux path.\n\n"
1765 "Samples in kernel modules won't be resolved at all.\n\n"
1766 "If some relocation was applied (e.g. kexec) symbols may be misresolved\n"
1767 "even with a suitable vmlinux or kallsyms file.\n\n");
1768
1769         if (rec->no_buildid_cache || rec->no_buildid) {
1770                 disable_buildid_cache();
1771         } else if (rec->switch_output.enabled) {
1772                 /*
1773                  * In 'perf record --switch-output', disable buildid
1774                  * generation by default to reduce data file switching
1775                  * overhead. Still generate buildid if they are required
1776                  * explicitly using
1777                  *
1778                  *  perf record --switch-output --no-no-buildid \
1779                  *              --no-no-buildid-cache
1780                  *
1781                  * Following code equals to:
1782                  *
1783                  * if ((rec->no_buildid || !rec->no_buildid_set) &&
1784                  *     (rec->no_buildid_cache || !rec->no_buildid_cache_set))
1785                  *         disable_buildid_cache();
1786                  */
1787                 bool disable = true;
1788
1789                 if (rec->no_buildid_set && !rec->no_buildid)
1790                         disable = false;
1791                 if (rec->no_buildid_cache_set && !rec->no_buildid_cache)
1792                         disable = false;
1793                 if (disable) {
1794                         rec->no_buildid = true;
1795                         rec->no_buildid_cache = true;
1796                         disable_buildid_cache();
1797                 }
1798         }
1799
1800         if (record.opts.overwrite)
1801                 record.opts.tail_synthesize = true;
1802
1803         if (rec->evlist->nr_entries == 0 &&
1804             __perf_evlist__add_default(rec->evlist, !record.opts.no_samples) < 0) {
1805                 pr_err("Not enough memory for event selector list\n");
1806                 goto out;
1807         }
1808
1809         if (rec->opts.target.tid && !rec->opts.no_inherit_set)
1810                 rec->opts.no_inherit = true;
1811
1812         err = target__validate(&rec->opts.target);
1813         if (err) {
1814                 target__strerror(&rec->opts.target, err, errbuf, BUFSIZ);
1815                 ui__warning("%s", errbuf);
1816         }
1817
1818         err = target__parse_uid(&rec->opts.target);
1819         if (err) {
1820                 int saved_errno = errno;
1821
1822                 target__strerror(&rec->opts.target, err, errbuf, BUFSIZ);
1823                 ui__error("%s", errbuf);
1824
1825                 err = -saved_errno;
1826                 goto out;
1827         }
1828
1829         /* Enable ignoring missing threads when -u/-p option is defined. */
1830         rec->opts.ignore_missing_thread = rec->opts.target.uid != UINT_MAX || rec->opts.target.pid;
1831
1832         err = -ENOMEM;
1833         if (perf_evlist__create_maps(rec->evlist, &rec->opts.target) < 0)
1834                 usage_with_options(record_usage, record_options);
1835
1836         err = auxtrace_record__options(rec->itr, rec->evlist, &rec->opts);
1837         if (err)
1838                 goto out;
1839
1840         /*
1841          * We take all buildids when the file contains
1842          * AUX area tracing data because we do not decode the
1843          * trace because it would take too long.
1844          */
1845         if (rec->opts.full_auxtrace)
1846                 rec->buildid_all = true;
1847
1848         if (record_opts__config(&rec->opts)) {
1849                 err = -EINVAL;
1850                 goto out;
1851         }
1852
1853         err = __cmd_record(&record, argc, argv);
1854 out:
1855         perf_evlist__delete(rec->evlist);
1856         symbol__exit();
1857         auxtrace_record__free(rec->itr);
1858         return err;
1859 }
1860
1861 static void snapshot_sig_handler(int sig __maybe_unused)
1862 {
1863         struct record *rec = &record;
1864
1865         if (trigger_is_ready(&auxtrace_snapshot_trigger)) {
1866                 trigger_hit(&auxtrace_snapshot_trigger);
1867                 auxtrace_record__snapshot_started = 1;
1868                 if (auxtrace_record__snapshot_start(record.itr))
1869                         trigger_error(&auxtrace_snapshot_trigger);
1870         }
1871
1872         if (switch_output_signal(rec))
1873                 trigger_hit(&switch_output_trigger);
1874 }
1875
1876 static void alarm_sig_handler(int sig __maybe_unused)
1877 {
1878         struct record *rec = &record;
1879
1880         if (switch_output_time(rec))
1881                 trigger_hit(&switch_output_trigger);
1882 }