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