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 #include "util/config.h"
  17
  18 #include "util/callchain.h"
  19 #include "util/cgroup.h"
  20 #include "util/header.h"
  21 #include "util/event.h"
  22 #include "util/evlist.h"
  23 #include "util/evsel.h"
  24 #include "util/debug.h"
  25 #include "util/session.h"
  26 #include "util/tool.h"
  27 #include "util/symbol.h"
  28 #include "util/cpumap.h"
  29 #include "util/thread_map.h"
  30 #include "util/data.h"
  31 #include "util/perf_regs.h"
  32 #include "util/auxtrace.h"
  33 #include "util/tsc.h"
  34 #include "util/parse-branch-options.h"
  35 #include "util/parse-regs-options.h"
  36 #include "util/llvm-utils.h"
  37 #include "util/bpf-loader.h"
  38 #include "util/trigger.h"
  39 #include "asm/bug.h"
  40
  41 #include <unistd.h>
  42 #include <sched.h>
  43 #include <sys/mman.h>
  44 #include <asm/bug.h>
  45
  46
  47 struct record {
  48         struct perf_tool        tool;
  49         struct record_opts      opts;
  50         u64                     bytes_written;
  51         struct perf_data_file   file;
  52         struct auxtrace_record  *itr;
  53         struct perf_evlist      *evlist;
  54         struct perf_session     *session;
  55         const char              *progname;
  56         int                     realtime_prio;
  57         bool                    no_buildid;
  58         bool                    no_buildid_set;
  59         bool                    no_buildid_cache;
  60         bool                    no_buildid_cache_set;
  61         bool                    buildid_all;
  62         bool                    timestamp_filename;
  63         bool                    switch_output;
  64         unsigned long long      samples;
  65 };
  66
  67 static int record__write(struct record *rec, void *bf, size_t size)
  68 {
  69         if (perf_data_file__write(rec->session->file, bf, size) < 0) {
  70                 pr_err("failed to write perf data, error: %m\n");
  71                 return -1;
  72         }
  73
  74         rec->bytes_written += size;
  75         return 0;
  76 }
  77
  78 static int process_synthesized_event(struct perf_tool *tool,
  79                                      union perf_event *event,
  80                                      struct perf_sample *sample __maybe_unused,
  81                                      struct machine *machine __maybe_unused)
  82 {
  83         struct record *rec = container_of(tool, struct record, tool);
  84         return record__write(rec, event, event->header.size);
  85 }
  86
  87 static int
  88 backward_rb_find_range(void *buf, int mask, u64 head, u64 *start, u64 *end)
  89 {
  90         struct perf_event_header *pheader;
  91         u64 evt_head = head;
  92         int size = mask + 1;
  93
  94         pr_debug2("backward_rb_find_range: buf=%p, head=%"PRIx64"\n", buf, head);
  95         pheader = (struct perf_event_header *)(buf + (head & mask));
  96         *start = head;
  97         while (true) {
  98                 if (evt_head - head >= (unsigned int)size) {
  99                         pr_debug("Finshed reading backward ring buffer: rewind\n");
 100                         if (evt_head - head > (unsigned int)size)
 101                                 evt_head -= pheader->size;
 102                         *end = evt_head;
 103                         return 0;
 104                 }
 105
 106                 pheader = (struct perf_event_header *)(buf + (evt_head & mask));
 107
 108                 if (pheader->size == 0) {
 109                         pr_debug("Finshed reading backward ring buffer: get start\n");
 110                         *end = evt_head;
 111                         return 0;
 112                 }
 113
 114                 evt_head += pheader->size;
 115                 pr_debug3("move evt_head: %"PRIx64"\n", evt_head);
 116         }
 117         WARN_ONCE(1, "Shouldn't get here\n");
 118         return -1;
 119 }
 120
 121 static int
 122 rb_find_range(struct perf_evlist *evlist,
 123               void *data, int mask, u64 head, u64 old,
 124               u64 *start, u64 *end)
 125 {
 126         if (!evlist->backward) {
 127                 *start = old;
 128                 *end = head;
 129                 return 0;
 130         }
 131
 132         return backward_rb_find_range(data, mask, head, start, end);
 133 }
 134
 135 static int record__mmap_read(struct record *rec, int idx)
 136 {
 137         struct perf_mmap *md = &rec->evlist->mmap[idx];
 138         u64 head = perf_mmap__read_head(md);
 139         u64 old = md->prev;
 140         u64 end = head, start = old;
 141         unsigned char *data = md->base + page_size;
 142         unsigned long size;
 143         void *buf;
 144         int rc = 0;
 145
 146         if (rb_find_range(rec->evlist, data, md->mask, head,
 147                           old, &start, &end))
 148                 return -1;
 149
 150         if (start == end)
 151                 return 0;
 152
 153         rec->samples++;
 154
 155         size = end - start;
 156         if (size > (unsigned long)(md->mask) + 1) {
 157                 WARN_ONCE(1, "failed to keep up with mmap data. (warn only once)\n");
 158
 159                 md->prev = head;
 160                 perf_evlist__mmap_consume(rec->evlist, idx);
 161                 return 0;
 162         }
 163
 164         if ((start & md->mask) + size != (end & md->mask)) {
 165                 buf = &data[start & md->mask];
 166                 size = md->mask + 1 - (start & md->mask);
 167                 start += size;
 168
 169                 if (record__write(rec, buf, size) < 0) {
 170                         rc = -1;
 171                         goto out;
 172                 }
 173         }
 174
 175         buf = &data[start & md->mask];
 176         size = end - start;
 177         start += size;
 178
 179         if (record__write(rec, buf, size) < 0) {
 180                 rc = -1;
 181                 goto out;
 182         }
 183
 184         md->prev = head;
 185         perf_evlist__mmap_consume(rec->evlist, idx);
 186 out:
 187         return rc;
 188 }
 189
 190 static volatile int done;
 191 static volatile int signr = -1;
 192 static volatile int child_finished;
 193
 194 static volatile int auxtrace_record__snapshot_started;
 195 static DEFINE_TRIGGER(auxtrace_snapshot_trigger);
 196 static DEFINE_TRIGGER(switch_output_trigger);
 197
 198 static void sig_handler(int sig)
 199 {
 200         if (sig == SIGCHLD)
 201                 child_finished = 1;
 202         else
 203                 signr = sig;
 204
 205         done = 1;
 206 }
 207
 208 static void record__sig_exit(void)
 209 {
 210         if (signr == -1)
 211                 return;
 212
 213         signal(signr, SIG_DFL);
 214         raise(signr);
 215 }
 216
 217 #ifdef HAVE_AUXTRACE_SUPPORT
 218
 219 static int record__process_auxtrace(struct perf_tool *tool,
 220                                     union perf_event *event, void *data1,
 221                                     size_t len1, void *data2, size_t len2)
 222 {
 223         struct record *rec = container_of(tool, struct record, tool);
 224         struct perf_data_file *file = &rec->file;
 225         size_t padding;
 226         u8 pad[8] = {0};
 227
 228         if (!perf_data_file__is_pipe(file)) {
 229                 off_t file_offset;
 230                 int fd = perf_data_file__fd(file);
 231                 int err;
 232
 233                 file_offset = lseek(fd, 0, SEEK_CUR);
 234                 if (file_offset == -1)
 235                         return -1;
 236                 err = auxtrace_index__auxtrace_event(&rec->session->auxtrace_index,
 237                                                      event, file_offset);
 238                 if (err)
 239                         return err;
 240         }
 241
 242         /* event.auxtrace.size includes padding, see __auxtrace_mmap__read() */
 243         padding = (len1 + len2) & 7;
 244         if (padding)
 245                 padding = 8 - padding;
 246
 247         record__write(rec, event, event->header.size);
 248         record__write(rec, data1, len1);
 249         if (len2)
 250                 record__write(rec, data2, len2);
 251         record__write(rec, &pad, padding);
 252
 253         return 0;
 254 }
 255
 256 static int record__auxtrace_mmap_read(struct record *rec,
 257                                       struct auxtrace_mmap *mm)
 258 {
 259         int ret;
 260
 261         ret = auxtrace_mmap__read(mm, rec->itr, &rec->tool,
 262                                   record__process_auxtrace);
 263         if (ret < 0)
 264                 return ret;
 265
 266         if (ret)
 267                 rec->samples++;
 268
 269         return 0;
 270 }
 271
 272 static int record__auxtrace_mmap_read_snapshot(struct record *rec,
 273                                                struct auxtrace_mmap *mm)
 274 {
 275         int ret;
 276
 277         ret = auxtrace_mmap__read_snapshot(mm, rec->itr, &rec->tool,
 278                                            record__process_auxtrace,
 279                                            rec->opts.auxtrace_snapshot_size);
 280         if (ret < 0)
 281                 return ret;
 282
 283         if (ret)
 284                 rec->samples++;
 285
 286         return 0;
 287 }
 288
 289 static int record__auxtrace_read_snapshot_all(struct record *rec)
 290 {
 291         int i;
 292         int rc = 0;
 293
 294         for (i = 0; i < rec->evlist->nr_mmaps; i++) {
 295                 struct auxtrace_mmap *mm =
 296                                 &rec->evlist->mmap[i].auxtrace_mmap;
 297
 298                 if (!mm->base)
 299                         continue;
 300
 301                 if (record__auxtrace_mmap_read_snapshot(rec, mm) != 0) {
 302                         rc = -1;
 303                         goto out;
 304                 }
 305         }
 306 out:
 307         return rc;
 308 }
 309
 310 static void record__read_auxtrace_snapshot(struct record *rec)
 311 {
 312         pr_debug("Recording AUX area tracing snapshot\n");
 313         if (record__auxtrace_read_snapshot_all(rec) < 0) {
 314                 trigger_error(&auxtrace_snapshot_trigger);
 315         } else {
 316                 if (auxtrace_record__snapshot_finish(rec->itr))
 317                         trigger_error(&auxtrace_snapshot_trigger);
 318                 else
 319                         trigger_ready(&auxtrace_snapshot_trigger);
 320         }
 321 }
 322
 323 #else
 324
 325 static inline
 326 int record__auxtrace_mmap_read(struct record *rec __maybe_unused,
 327                                struct auxtrace_mmap *mm __maybe_unused)
 328 {
 329         return 0;
 330 }
 331
 332 static inline
 333 void record__read_auxtrace_snapshot(struct record *rec __maybe_unused)
 334 {
 335 }
 336
 337 static inline
 338 int auxtrace_record__snapshot_start(struct auxtrace_record *itr __maybe_unused)
 339 {
 340         return 0;
 341 }
 342
 343 #endif
 344
 345 static int record__mmap_evlist(struct record *rec,
 346                                struct perf_evlist *evlist)
 347 {
 348         struct record_opts *opts = &rec->opts;
 349         char msg[512];
 350
 351         if (perf_evlist__mmap_ex(evlist, opts->mmap_pages, false,
 352                                  opts->auxtrace_mmap_pages,
 353                                  opts->auxtrace_snapshot_mode) < 0) {
 354                 if (errno == EPERM) {
 355                         pr_err("Permission error mapping pages.\n"
 356                                "Consider increasing "
 357                                "/proc/sys/kernel/perf_event_mlock_kb,\n"
 358                                "or try again with a smaller value of -m/--mmap_pages.\n"
 359                                "(current value: %u,%u)\n",
 360                                opts->mmap_pages, opts->auxtrace_mmap_pages);
 361                         return -errno;
 362                 } else {
 363                         pr_err("failed to mmap with %d (%s)\n", errno,
 364                                 strerror_r(errno, msg, sizeof(msg)));
 365                         if (errno)
 366                                 return -errno;
 367                         else
 368                                 return -EINVAL;
 369                 }
 370         }
 371         return 0;
 372 }
 373
 374 static int record__mmap(struct record *rec)
 375 {
 376         return record__mmap_evlist(rec, rec->evlist);
 377 }
 378
 379 static int record__open(struct record *rec)
 380 {
 381         char msg[512];
 382         struct perf_evsel *pos;
 383         struct perf_evlist *evlist = rec->evlist;
 384         struct perf_session *session = rec->session;
 385         struct record_opts *opts = &rec->opts;
 386         int rc = 0;
 387
 388         perf_evlist__config(evlist, opts, &callchain_param);
 389
 390         evlist__for_each_entry(evlist, pos) {
 391 try_again:
 392                 if (perf_evsel__open(pos, pos->cpus, pos->threads) < 0) {
 393                         if (perf_evsel__fallback(pos, errno, msg, sizeof(msg))) {
 394                                 if (verbose)
 395                                         ui__warning("%s\n", msg);
 396                                 goto try_again;
 397                         }
 398
 399                         rc = -errno;
 400                         perf_evsel__open_strerror(pos, &opts->target,
 401                                                   errno, msg, sizeof(msg));
 402                         ui__error("%s\n", msg);
 403                         goto out;
 404                 }
 405         }
 406
 407         if (perf_evlist__apply_filters(evlist, &pos)) {
 408                 error("failed to set filter \"%s\" on event %s with %d (%s)\n",
 409                         pos->filter, perf_evsel__name(pos), errno,
 410                         strerror_r(errno, msg, sizeof(msg)));
 411                 rc = -1;
 412                 goto out;
 413         }
 414
 415         rc = record__mmap(rec);
 416         if (rc)
 417                 goto out;
 418
 419         session->evlist = evlist;
 420         perf_session__set_id_hdr_size(session);
 421 out:
 422         return rc;
 423 }
 424
 425 static int process_sample_event(struct perf_tool *tool,
 426                                 union perf_event *event,
 427                                 struct perf_sample *sample,
 428                                 struct perf_evsel *evsel,
 429                                 struct machine *machine)
 430 {
 431         struct record *rec = container_of(tool, struct record, tool);
 432
 433         rec->samples++;
 434
 435         return build_id__mark_dso_hit(tool, event, sample, evsel, machine);
 436 }
 437
 438 static int process_buildids(struct record *rec)
 439 {
 440         struct perf_data_file *file  = &rec->file;
 441         struct perf_session *session = rec->session;
 442
 443         if (file->size == 0)
 444                 return 0;
 445
 446         /*
 447          * During this process, it'll load kernel map and replace the
 448          * dso->long_name to a real pathname it found.  In this case
 449          * we prefer the vmlinux path like
 450          *   /lib/modules/3.16.4/build/vmlinux
 451          *
 452          * rather than build-id path (in debug directory).
 453          *   $HOME/.debug/.build-id/f0/6e17aa50adf4d00b88925e03775de107611551
 454          */
 455         symbol_conf.ignore_vmlinux_buildid = true;
 456
 457         /*
 458          * If --buildid-all is given, it marks all DSO regardless of hits,
 459          * so no need to process samples.
 460          */
 461         if (rec->buildid_all)
 462                 rec->tool.sample = NULL;
 463
 464         return perf_session__process_events(session);
 465 }
 466
 467 static void perf_event__synthesize_guest_os(struct machine *machine, void *data)
 468 {
 469         int err;
 470         struct perf_tool *tool = data;
 471         /*
 472          *As for guest kernel when processing subcommand record&report,
 473          *we arrange module mmap prior to guest kernel mmap and trigger
 474          *a preload dso because default guest module symbols are loaded
 475          *from guest kallsyms instead of /lib/modules/XXX/XXX. This
 476          *method is used to avoid symbol missing when the first addr is
 477          *in module instead of in guest kernel.
 478          */
 479         err = perf_event__synthesize_modules(tool, process_synthesized_event,
 480                                              machine);
 481         if (err < 0)
 482                 pr_err("Couldn't record guest kernel [%d]'s reference"
 483                        " relocation symbol.\n", machine->pid);
 484
 485         /*
 486          * We use _stext for guest kernel because guest kernel's /proc/kallsyms
 487          * have no _text sometimes.
 488          */
 489         err = perf_event__synthesize_kernel_mmap(tool, process_synthesized_event,
 490                                                  machine);
 491         if (err < 0)
 492                 pr_err("Couldn't record guest kernel [%d]'s reference"
 493                        " relocation symbol.\n", machine->pid);
 494 }
 495
 496 static struct perf_event_header finished_round_event = {
 497         .size = sizeof(struct perf_event_header),
 498         .type = PERF_RECORD_FINISHED_ROUND,
 499 };
 500
 501 static int record__mmap_read_all(struct record *rec)
 502 {
 503         u64 bytes_written = rec->bytes_written;
 504         int i;
 505         int rc = 0;
 506
 507         for (i = 0; i < rec->evlist->nr_mmaps; i++) {
 508                 struct auxtrace_mmap *mm = &rec->evlist->mmap[i].auxtrace_mmap;
 509
 510                 if (rec->evlist->mmap[i].base) {
 511                         if (record__mmap_read(rec, i) != 0) {
 512                                 rc = -1;
 513                                 goto out;
 514                         }
 515                 }
 516
 517                 if (mm->base && !rec->opts.auxtrace_snapshot_mode &&
 518                     record__auxtrace_mmap_read(rec, mm) != 0) {
 519                         rc = -1;
 520                         goto out;
 521                 }
 522         }
 523
 524         /*
 525          * Mark the round finished in case we wrote
 526          * at least one event.
 527          */
 528         if (bytes_written != rec->bytes_written)
 529                 rc = record__write(rec, &finished_round_event, sizeof(finished_round_event));
 530
 531 out:
 532         return rc;
 533 }
 534
 535 static void record__init_features(struct record *rec)
 536 {
 537         struct perf_session *session = rec->session;
 538         int feat;
 539
 540         for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++)
 541                 perf_header__set_feat(&session->header, feat);
 542
 543         if (rec->no_buildid)
 544                 perf_header__clear_feat(&session->header, HEADER_BUILD_ID);
 545
 546         if (!have_tracepoints(&rec->evlist->entries))
 547                 perf_header__clear_feat(&session->header, HEADER_TRACING_DATA);
 548
 549         if (!rec->opts.branch_stack)
 550                 perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK);
 551
 552         if (!rec->opts.full_auxtrace)
 553                 perf_header__clear_feat(&session->header, HEADER_AUXTRACE);
 554
 555         perf_header__clear_feat(&session->header, HEADER_STAT);
 556 }
 557
 558 static void
 559 record__finish_output(struct record *rec)
 560 {
 561         struct perf_data_file *file = &rec->file;
 562         int fd = perf_data_file__fd(file);
 563
 564         if (file->is_pipe)
 565                 return;
 566
 567         rec->session->header.data_size += rec->bytes_written;
 568         file->size = lseek(perf_data_file__fd(file), 0, SEEK_CUR);
 569
 570         if (!rec->no_buildid) {
 571                 process_buildids(rec);
 572
 573                 if (rec->buildid_all)
 574                         dsos__hit_all(rec->session);
 575         }
 576         perf_session__write_header(rec->session, rec->evlist, fd, true);
 577
 578         return;
 579 }
 580
 581 static int record__synthesize_workload(struct record *rec)
 582 {
 583         struct {
 584                 struct thread_map map;
 585                 struct thread_map_data map_data;
 586         } thread_map;
 587
 588         thread_map.map.nr = 1;
 589         thread_map.map.map[0].pid = rec->evlist->workload.pid;
 590         thread_map.map.map[0].comm = NULL;
 591         return perf_event__synthesize_thread_map(&rec->tool, &thread_map.map,
 592                                                  process_synthesized_event,
 593                                                  &rec->session->machines.host,
 594                                                  rec->opts.sample_address,
 595                                                  rec->opts.proc_map_timeout);
 596 }
 597
 598 static int record__synthesize(struct record *rec);
 599
 600 static int
 601 record__switch_output(struct record *rec, bool at_exit)
 602 {
 603         struct perf_data_file *file = &rec->file;
 604         int fd, err;
 605
 606         /* Same Size:      "2015122520103046"*/
 607         char timestamp[] = "InvalidTimestamp";
 608
 609         rec->samples = 0;
 610         record__finish_output(rec);
 611         err = fetch_current_timestamp(timestamp, sizeof(timestamp));
 612         if (err) {
 613                 pr_err("Failed to get current timestamp\n");
 614                 return -EINVAL;
 615         }
 616
 617         fd = perf_data_file__switch(file, timestamp,
 618                                     rec->session->header.data_offset,
 619                                     at_exit);
 620         if (fd >= 0 && !at_exit) {
 621                 rec->bytes_written = 0;
 622                 rec->session->header.data_size = 0;
 623         }
 624
 625         if (!quiet)
 626                 fprintf(stderr, "[ perf record: Dump %s.%s ]\n",
 627                         file->path, timestamp);
 628
 629         /* Output tracking events */
 630         if (!at_exit) {
 631                 record__synthesize(rec);
 632
 633                 /*
 634                  * In 'perf record --switch-output' without -a,
 635                  * record__synthesize() in record__switch_output() won't
 636                  * generate tracking events because there's no thread_map
 637                  * in evlist. Which causes newly created perf.data doesn't
 638                  * contain map and comm information.
 639                  * Create a fake thread_map and directly call
 640                  * perf_event__synthesize_thread_map() for those events.
 641                  */
 642                 if (target__none(&rec->opts.target))
 643                         record__synthesize_workload(rec);
 644         }
 645         return fd;
 646 }
 647
 648 static volatile int workload_exec_errno;
 649
 650 /*
 651  * perf_evlist__prepare_workload will send a SIGUSR1
 652  * if the fork fails, since we asked by setting its
 653  * want_signal to true.
 654  */
 655 static void workload_exec_failed_signal(int signo __maybe_unused,
 656                                         siginfo_t *info,
 657                                         void *ucontext __maybe_unused)
 658 {
 659         workload_exec_errno = info->si_value.sival_int;
 660         done = 1;
 661         child_finished = 1;
 662 }
 663
 664 static void snapshot_sig_handler(int sig);
 665
 666 int __weak
 667 perf_event__synth_time_conv(const struct perf_event_mmap_page *pc __maybe_unused,
 668                             struct perf_tool *tool __maybe_unused,
 669                             perf_event__handler_t process __maybe_unused,
 670                             struct machine *machine __maybe_unused)
 671 {
 672         return 0;
 673 }
 674
 675 static const struct perf_event_mmap_page *record__pick_pc(struct record *rec)
 676 {
 677         if (rec->evlist && rec->evlist->mmap && rec->evlist->mmap[0].base)
 678                 return rec->evlist->mmap[0].base;
 679         return NULL;
 680 }
 681
 682 static int record__synthesize(struct record *rec)
 683 {
 684         struct perf_session *session = rec->session;
 685         struct machine *machine = &session->machines.host;
 686         struct perf_data_file *file = &rec->file;
 687         struct record_opts *opts = &rec->opts;
 688         struct perf_tool *tool = &rec->tool;
 689         int fd = perf_data_file__fd(file);
 690         int err = 0;
 691
 692         if (file->is_pipe) {
 693                 err = perf_event__synthesize_attrs(tool, session,
 694                                                    process_synthesized_event);
 695                 if (err < 0) {
 696                         pr_err("Couldn't synthesize attrs.\n");
 697                         goto out;
 698                 }
 699
 700                 if (have_tracepoints(&rec->evlist->entries)) {
 701                         /*
 702                          * FIXME err <= 0 here actually means that
 703                          * there were no tracepoints so its not really
 704                          * an error, just that we don't need to
 705                          * synthesize anything.  We really have to
 706                          * return this more properly and also
 707                          * propagate errors that now are calling die()
 708                          */
 709                         err = perf_event__synthesize_tracing_data(tool, fd, rec->evlist,
 710                                                                   process_synthesized_event);
 711                         if (err <= 0) {
 712                                 pr_err("Couldn't record tracing data.\n");
 713                                 goto out;
 714                         }
 715                         rec->bytes_written += err;
 716                 }
 717         }
 718
 719         err = perf_event__synth_time_conv(record__pick_pc(rec), tool,
 720                                           process_synthesized_event, machine);
 721         if (err)
 722                 goto out;
 723
 724         if (rec->opts.full_auxtrace) {
 725                 err = perf_event__synthesize_auxtrace_info(rec->itr, tool,
 726                                         session, process_synthesized_event);
 727                 if (err)
 728                         goto out;
 729         }
 730
 731         err = perf_event__synthesize_kernel_mmap(tool, process_synthesized_event,
 732                                                  machine);
 733         WARN_ONCE(err < 0, "Couldn't record kernel reference relocation symbol\n"
 734                            "Symbol resolution may be skewed if relocation was used (e.g. kexec).\n"
 735                            "Check /proc/kallsyms permission or run as root.\n");
 736
 737         err = perf_event__synthesize_modules(tool, process_synthesized_event,
 738                                              machine);
 739         WARN_ONCE(err < 0, "Couldn't record kernel module information.\n"
 740                            "Symbol resolution may be skewed if relocation was used (e.g. kexec).\n"
 741                            "Check /proc/modules permission or run as root.\n");
 742
 743         if (perf_guest) {
 744                 machines__process_guests(&session->machines,
 745                                          perf_event__synthesize_guest_os, tool);
 746         }
 747
 748         err = __machine__synthesize_threads(machine, tool, &opts->target, rec->evlist->threads,
 749                                             process_synthesized_event, opts->sample_address,
 750                                             opts->proc_map_timeout);
 751 out:
 752         return err;
 753 }
 754
 755 static int __cmd_record(struct record *rec, int argc, const char **argv)
 756 {
 757         int err;
 758         int status = 0;
 759         unsigned long waking = 0;
 760         const bool forks = argc > 0;
 761         struct machine *machine;
 762         struct perf_tool *tool = &rec->tool;
 763         struct record_opts *opts = &rec->opts;
 764         struct perf_data_file *file = &rec->file;
 765         struct perf_session *session;
 766         bool disabled = false, draining = false;
 767         int fd;
 768
 769         rec->progname = argv[0];
 770
 771         atexit(record__sig_exit);
 772         signal(SIGCHLD, sig_handler);
 773         signal(SIGINT, sig_handler);
 774         signal(SIGTERM, sig_handler);
 775
 776         if (rec->opts.auxtrace_snapshot_mode || rec->switch_output) {
 777                 signal(SIGUSR2, snapshot_sig_handler);
 778                 if (rec->opts.auxtrace_snapshot_mode)
 779                         trigger_on(&auxtrace_snapshot_trigger);
 780                 if (rec->switch_output)
 781                         trigger_on(&switch_output_trigger);
 782         } else {
 783                 signal(SIGUSR2, SIG_IGN);
 784         }
 785
 786         session = perf_session__new(file, false, tool);
 787         if (session == NULL) {
 788                 pr_err("Perf session creation failed.\n");
 789                 return -1;
 790         }
 791
 792         fd = perf_data_file__fd(file);
 793         rec->session = session;
 794
 795         record__init_features(rec);
 796
 797         if (forks) {
 798                 err = perf_evlist__prepare_workload(rec->evlist, &opts->target,
 799                                                     argv, file->is_pipe,
 800                                                     workload_exec_failed_signal);
 801                 if (err < 0) {
 802                         pr_err("Couldn't run the workload!\n");
 803                         status = err;
 804                         goto out_delete_session;
 805                 }
 806         }
 807
 808         if (record__open(rec) != 0) {
 809                 err = -1;
 810                 goto out_child;
 811         }
 812
 813         err = bpf__apply_obj_config();
 814         if (err) {
 815                 char errbuf[BUFSIZ];
 816
 817                 bpf__strerror_apply_obj_config(err, errbuf, sizeof(errbuf));
 818                 pr_err("ERROR: Apply config to BPF failed: %s\n",
 819                          errbuf);
 820                 goto out_child;
 821         }
 822
 823         /*
 824          * Normally perf_session__new would do this, but it doesn't have the
 825          * evlist.
 826          */
 827         if (rec->tool.ordered_events && !perf_evlist__sample_id_all(rec->evlist)) {
 828                 pr_warning("WARNING: No sample_id_all support, falling back to unordered processing\n");
 829                 rec->tool.ordered_events = false;
 830         }
 831
 832         if (!rec->evlist->nr_groups)
 833                 perf_header__clear_feat(&session->header, HEADER_GROUP_DESC);
 834
 835         if (file->is_pipe) {
 836                 err = perf_header__write_pipe(fd);
 837                 if (err < 0)
 838                         goto out_child;
 839         } else {
 840                 err = perf_session__write_header(session, rec->evlist, fd, false);
 841                 if (err < 0)
 842                         goto out_child;
 843         }
 844
 845         if (!rec->no_buildid
 846             && !perf_header__has_feat(&session->header, HEADER_BUILD_ID)) {
 847                 pr_err("Couldn't generate buildids. "
 848                        "Use --no-buildid to profile anyway.\n");
 849                 err = -1;
 850                 goto out_child;
 851         }
 852
 853         machine = &session->machines.host;
 854
 855         err = record__synthesize(rec);
 856         if (err < 0)
 857                 goto out_child;
 858
 859         if (rec->realtime_prio) {
 860                 struct sched_param param;
 861
 862                 param.sched_priority = rec->realtime_prio;
 863                 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
 864                         pr_err("Could not set realtime priority.\n");
 865                         err = -1;
 866                         goto out_child;
 867                 }
 868         }
 869
 870         /*
 871          * When perf is starting the traced process, all the events
 872          * (apart from group members) have enable_on_exec=1 set,
 873          * so don't spoil it by prematurely enabling them.
 874          */
 875         if (!target__none(&opts->target) && !opts->initial_delay)
 876                 perf_evlist__enable(rec->evlist);
 877
 878         /*
 879          * Let the child rip
 880          */
 881         if (forks) {
 882                 union perf_event *event;
 883
 884                 event = malloc(sizeof(event->comm) + machine->id_hdr_size);
 885                 if (event == NULL) {
 886                         err = -ENOMEM;
 887                         goto out_child;
 888                 }
 889
 890                 /*
 891                  * Some H/W events are generated before COMM event
 892                  * which is emitted during exec(), so perf script
 893                  * cannot see a correct process name for those events.
 894                  * Synthesize COMM event to prevent it.
 895                  */
 896                 perf_event__synthesize_comm(tool, event,
 897                                             rec->evlist->workload.pid,
 898                                             process_synthesized_event,
 899                                             machine);
 900                 free(event);
 901
 902                 perf_evlist__start_workload(rec->evlist);
 903         }
 904
 905         if (opts->initial_delay) {
 906                 usleep(opts->initial_delay * 1000);
 907                 perf_evlist__enable(rec->evlist);
 908         }
 909
 910         trigger_ready(&auxtrace_snapshot_trigger);
 911         trigger_ready(&switch_output_trigger);
 912         for (;;) {
 913                 unsigned long long hits = rec->samples;
 914
 915                 if (record__mmap_read_all(rec) < 0) {
 916                         trigger_error(&auxtrace_snapshot_trigger);
 917                         trigger_error(&switch_output_trigger);
 918                         err = -1;
 919                         goto out_child;
 920                 }
 921
 922                 if (auxtrace_record__snapshot_started) {
 923                         auxtrace_record__snapshot_started = 0;
 924                         if (!trigger_is_error(&auxtrace_snapshot_trigger))
 925                                 record__read_auxtrace_snapshot(rec);
 926                         if (trigger_is_error(&auxtrace_snapshot_trigger)) {
 927                                 pr_err("AUX area tracing snapshot failed\n");
 928                                 err = -1;
 929                                 goto out_child;
 930                         }
 931                 }
 932
 933                 if (trigger_is_hit(&switch_output_trigger)) {
 934                         trigger_ready(&switch_output_trigger);
 935
 936                         if (!quiet)
 937                                 fprintf(stderr, "[ perf record: dump data: Woken up %ld times ]\n",
 938                                         waking);
 939                         waking = 0;
 940                         fd = record__switch_output(rec, false);
 941                         if (fd < 0) {
 942                                 pr_err("Failed to switch to new file\n");
 943                                 trigger_error(&switch_output_trigger);
 944                                 err = fd;
 945                                 goto out_child;
 946                         }
 947                 }
 948
 949                 if (hits == rec->samples) {
 950                         if (done || draining)
 951                                 break;
 952                         err = perf_evlist__poll(rec->evlist, -1);
 953                         /*
 954                          * Propagate error, only if there's any. Ignore positive
 955                          * number of returned events and interrupt error.
 956                          */
 957                         if (err > 0 || (err < 0 && errno == EINTR))
 958                                 err = 0;
 959                         waking++;
 960
 961                         if (perf_evlist__filter_pollfd(rec->evlist, POLLERR | POLLHUP) == 0)
 962                                 draining = true;
 963                 }
 964
 965                 /*
 966                  * When perf is starting the traced process, at the end events
 967                  * die with the process and we wait for that. Thus no need to
 968                  * disable events in this case.
 969                  */
 970                 if (done && !disabled && !target__none(&opts->target)) {
 971                         trigger_off(&auxtrace_snapshot_trigger);
 972                         perf_evlist__disable(rec->evlist);
 973                         disabled = true;
 974                 }
 975         }
 976         trigger_off(&auxtrace_snapshot_trigger);
 977         trigger_off(&switch_output_trigger);
 978
 979         if (forks && workload_exec_errno) {
 980                 char msg[STRERR_BUFSIZE];
 981                 const char *emsg = strerror_r(workload_exec_errno, msg, sizeof(msg));
 982                 pr_err("Workload failed: %s\n", emsg);
 983                 err = -1;
 984                 goto out_child;
 985         }
 986
 987         if (!quiet)
 988                 fprintf(stderr, "[ perf record: Woken up %ld times to write data ]\n", waking);
 989
 990 out_child:
 991         if (forks) {
 992                 int exit_status;
 993
 994                 if (!child_finished)
 995                         kill(rec->evlist->workload.pid, SIGTERM);
 996
 997                 wait(&exit_status);
 998
 999                 if (err < 0)
1000                         status = err;
1001                 else if (WIFEXITED(exit_status))
1002                         status = WEXITSTATUS(exit_status);
1003                 else if (WIFSIGNALED(exit_status))
1004                         signr = WTERMSIG(exit_status);
1005         } else
1006                 status = err;
1007
1008         /* this will be recalculated during process_buildids() */
1009         rec->samples = 0;
1010
1011         if (!err) {
1012                 if (!rec->timestamp_filename) {
1013                         record__finish_output(rec);
1014                 } else {
1015                         fd = record__switch_output(rec, true);
1016                         if (fd < 0) {
1017                                 status = fd;
1018                                 goto out_delete_session;
1019                         }
1020                 }
1021         }
1022
1023         if (!err && !quiet) {
1024                 char samples[128];
1025                 const char *postfix = rec->timestamp_filename ?
1026                                         ".<timestamp>" : "";
1027
1028                 if (rec->samples && !rec->opts.full_auxtrace)
1029                         scnprintf(samples, sizeof(samples),
1030                                   " (%" PRIu64 " samples)", rec->samples);
1031                 else
1032                         samples[0] = '\0';
1033
1034                 fprintf(stderr, "[ perf record: Captured and wrote %.3f MB %s%s%s ]\n",
1035                         perf_data_file__size(file) / 1024.0 / 1024.0,
1036                         file->path, postfix, samples);
1037         }
1038
1039 out_delete_session:
1040         perf_session__delete(session);
1041         return status;
1042 }
1043
1044 static void callchain_debug(struct callchain_param *callchain)
1045 {
1046         static const char *str[CALLCHAIN_MAX] = { "NONE", "FP", "DWARF", "LBR" };
1047
1048         pr_debug("callchain: type %s\n", str[callchain->record_mode]);
1049
1050         if (callchain->record_mode == CALLCHAIN_DWARF)
1051                 pr_debug("callchain: stack dump size %d\n",
1052                          callchain->dump_size);
1053 }
1054
1055 int record_opts__parse_callchain(struct record_opts *record,
1056                                  struct callchain_param *callchain,
1057                                  const char *arg, bool unset)
1058 {
1059         int ret;
1060         callchain->enabled = !unset;
1061
1062         /* --no-call-graph */
1063         if (unset) {
1064                 callchain->record_mode = CALLCHAIN_NONE;
1065                 pr_debug("callchain: disabled\n");
1066                 return 0;
1067         }
1068
1069         ret = parse_callchain_record_opt(arg, callchain);
1070         if (!ret) {
1071                 /* Enable data address sampling for DWARF unwind. */
1072                 if (callchain->record_mode == CALLCHAIN_DWARF)
1073                         record->sample_address = true;
1074                 callchain_debug(callchain);
1075         }
1076
1077         return ret;
1078 }
1079
1080 int record_parse_callchain_opt(const struct option *opt,
1081                                const char *arg,
1082                                int unset)
1083 {
1084         return record_opts__parse_callchain(opt->value, &callchain_param, arg, unset);
1085 }
1086
1087 int record_callchain_opt(const struct option *opt,
1088                          const char *arg __maybe_unused,
1089                          int unset __maybe_unused)
1090 {
1091         struct callchain_param *callchain = opt->value;
1092
1093         callchain->enabled = true;
1094
1095         if (callchain->record_mode == CALLCHAIN_NONE)
1096                 callchain->record_mode = CALLCHAIN_FP;
1097
1098         callchain_debug(callchain);
1099         return 0;
1100 }
1101
1102 static int perf_record_config(const char *var, const char *value, void *cb)
1103 {
1104         struct record *rec = cb;
1105
1106         if (!strcmp(var, "record.build-id")) {
1107                 if (!strcmp(value, "cache"))
1108                         rec->no_buildid_cache = false;
1109                 else if (!strcmp(value, "no-cache"))
1110                         rec->no_buildid_cache = true;
1111                 else if (!strcmp(value, "skip"))
1112                         rec->no_buildid = true;
1113                 else
1114                         return -1;
1115                 return 0;
1116         }
1117         if (!strcmp(var, "record.call-graph"))
1118                 var = "call-graph.record-mode"; /* fall-through */
1119
1120         return perf_default_config(var, value, cb);
1121 }
1122
1123 struct clockid_map {
1124         const char *name;
1125         int clockid;
1126 };
1127
1128 #define CLOCKID_MAP(n, c)       \
1129         { .name = n, .clockid = (c), }
1130
1131 #define CLOCKID_END     { .name = NULL, }
1132
1133
1134 /*
1135  * Add the missing ones, we need to build on many distros...
1136  */
1137 #ifndef CLOCK_MONOTONIC_RAW
1138 #define CLOCK_MONOTONIC_RAW 4
1139 #endif
1140 #ifndef CLOCK_BOOTTIME
1141 #define CLOCK_BOOTTIME 7
1142 #endif
1143 #ifndef CLOCK_TAI
1144 #define CLOCK_TAI 11
1145 #endif
1146
1147 static const struct clockid_map clockids[] = {
1148         /* available for all events, NMI safe */
1149         CLOCKID_MAP("monotonic", CLOCK_MONOTONIC),
1150         CLOCKID_MAP("monotonic_raw", CLOCK_MONOTONIC_RAW),
1151
1152         /* available for some events */
1153         CLOCKID_MAP("realtime", CLOCK_REALTIME),
1154         CLOCKID_MAP("boottime", CLOCK_BOOTTIME),
1155         CLOCKID_MAP("tai", CLOCK_TAI),
1156
1157         /* available for the lazy */
1158         CLOCKID_MAP("mono", CLOCK_MONOTONIC),
1159         CLOCKID_MAP("raw", CLOCK_MONOTONIC_RAW),
1160         CLOCKID_MAP("real", CLOCK_REALTIME),
1161         CLOCKID_MAP("boot", CLOCK_BOOTTIME),
1162
1163         CLOCKID_END,
1164 };
1165
1166 static int parse_clockid(const struct option *opt, const char *str, int unset)
1167 {
1168         struct record_opts *opts = (struct record_opts *)opt->value;
1169         const struct clockid_map *cm;
1170         const char *ostr = str;
1171
1172         if (unset) {
1173                 opts->use_clockid = 0;
1174                 return 0;
1175         }
1176
1177         /* no arg passed */
1178         if (!str)
1179                 return 0;
1180
1181         /* no setting it twice */
1182         if (opts->use_clockid)
1183                 return -1;
1184
1185         opts->use_clockid = true;
1186
1187         /* if its a number, we're done */
1188         if (sscanf(str, "%d", &opts->clockid) == 1)
1189                 return 0;
1190
1191         /* allow a "CLOCK_" prefix to the name */
1192         if (!strncasecmp(str, "CLOCK_", 6))
1193                 str += 6;
1194
1195         for (cm = clockids; cm->name; cm++) {
1196                 if (!strcasecmp(str, cm->name)) {
1197                         opts->clockid = cm->clockid;
1198                         return 0;
1199                 }
1200         }
1201
1202         opts->use_clockid = false;
1203         ui__warning("unknown clockid %s, check man page\n", ostr);
1204         return -1;
1205 }
1206
1207 static int record__parse_mmap_pages(const struct option *opt,
1208                                     const char *str,
1209                                     int unset __maybe_unused)
1210 {
1211         struct record_opts *opts = opt->value;
1212         char *s, *p;
1213         unsigned int mmap_pages;
1214         int ret;
1215
1216         if (!str)
1217                 return -EINVAL;
1218
1219         s = strdup(str);
1220         if (!s)
1221                 return -ENOMEM;
1222
1223         p = strchr(s, ',');
1224         if (p)
1225                 *p = '\0';
1226
1227         if (*s) {
1228                 ret = __perf_evlist__parse_mmap_pages(&mmap_pages, s);
1229                 if (ret)
1230                         goto out_free;
1231                 opts->mmap_pages = mmap_pages;
1232         }
1233
1234         if (!p) {
1235                 ret = 0;
1236                 goto out_free;
1237         }
1238
1239         ret = __perf_evlist__parse_mmap_pages(&mmap_pages, p + 1);
1240         if (ret)
1241                 goto out_free;
1242
1243         opts->auxtrace_mmap_pages = mmap_pages;
1244
1245 out_free:
1246         free(s);
1247         return ret;
1248 }
1249
1250 static const char * const __record_usage[] = {
1251         "perf record [<options>] [<command>]",
1252         "perf record [<options>] -- <command> [<options>]",
1253         NULL
1254 };
1255 const char * const *record_usage = __record_usage;
1256
1257 /*
1258  * XXX Ideally would be local to cmd_record() and passed to a record__new
1259  * because we need to have access to it in record__exit, that is called
1260  * after cmd_record() exits, but since record_options need to be accessible to
1261  * builtin-script, leave it here.
1262  *
1263  * At least we don't ouch it in all the other functions here directly.
1264  *
1265  * Just say no to tons of global variables, sigh.
1266  */
1267 static struct record record = {
1268         .opts = {
1269                 .sample_time         = true,
1270                 .mmap_pages          = UINT_MAX,
1271                 .user_freq           = UINT_MAX,
1272                 .user_interval       = ULLONG_MAX,
1273                 .freq                = 4000,
1274                 .target              = {
1275                         .uses_mmap   = true,
1276                         .default_per_cpu = true,
1277                 },
1278                 .proc_map_timeout     = 500,
1279         },
1280         .tool = {
1281                 .sample         = process_sample_event,
1282                 .fork           = perf_event__process_fork,
1283                 .exit           = perf_event__process_exit,
1284                 .comm           = perf_event__process_comm,
1285                 .mmap           = perf_event__process_mmap,
1286                 .mmap2          = perf_event__process_mmap2,
1287                 .ordered_events = true,
1288         },
1289 };
1290
1291 const char record_callchain_help[] = CALLCHAIN_RECORD_HELP
1292         "\n\t\t\t\tDefault: fp";
1293
1294 static bool dry_run;
1295
1296 /*
1297  * XXX Will stay a global variable till we fix builtin-script.c to stop messing
1298  * with it and switch to use the library functions in perf_evlist that came
1299  * from builtin-record.c, i.e. use record_opts,
1300  * perf_evlist__prepare_workload, etc instead of fork+exec'in 'perf record',
1301  * using pipes, etc.
1302  */
1303 struct option __record_options[] = {
1304         OPT_CALLBACK('e', "event", &record.evlist, "event",
1305                      "event selector. use 'perf list' to list available events",
1306                      parse_events_option),
1307         OPT_CALLBACK(0, "filter", &record.evlist, "filter",
1308                      "event filter", parse_filter),
1309         OPT_CALLBACK_NOOPT(0, "exclude-perf", &record.evlist,
1310                            NULL, "don't record events from perf itself",
1311                            exclude_perf),
1312         OPT_STRING('p', "pid", &record.opts.target.pid, "pid",
1313                     "record events on existing process id"),
1314         OPT_STRING('t', "tid", &record.opts.target.tid, "tid",
1315                     "record events on existing thread id"),
1316         OPT_INTEGER('r', "realtime", &record.realtime_prio,
1317                     "collect data with this RT SCHED_FIFO priority"),
1318         OPT_BOOLEAN(0, "no-buffering", &record.opts.no_buffering,
1319                     "collect data without buffering"),
1320         OPT_BOOLEAN('R', "raw-samples", &record.opts.raw_samples,
1321                     "collect raw sample records from all opened counters"),
1322         OPT_BOOLEAN('a', "all-cpus", &record.opts.target.system_wide,
1323                             "system-wide collection from all CPUs"),
1324         OPT_STRING('C', "cpu", &record.opts.target.cpu_list, "cpu",
1325                     "list of cpus to monitor"),
1326         OPT_U64('c', "count", &record.opts.user_interval, "event period to sample"),
1327         OPT_STRING('o', "output", &record.file.path, "file",
1328                     "output file name"),
1329         OPT_BOOLEAN_SET('i', "no-inherit", &record.opts.no_inherit,
1330                         &record.opts.no_inherit_set,
1331                         "child tasks do not inherit counters"),
1332         OPT_UINTEGER('F', "freq", &record.opts.user_freq, "profile at this frequency"),
1333         OPT_CALLBACK('m', "mmap-pages", &record.opts, "pages[,pages]",
1334                      "number of mmap data pages and AUX area tracing mmap pages",
1335                      record__parse_mmap_pages),
1336         OPT_BOOLEAN(0, "group", &record.opts.group,
1337                     "put the counters into a counter group"),
1338         OPT_CALLBACK_NOOPT('g', NULL, &callchain_param,
1339                            NULL, "enables call-graph recording" ,
1340                            &record_callchain_opt),
1341         OPT_CALLBACK(0, "call-graph", &record.opts,
1342                      "record_mode[,record_size]", record_callchain_help,
1343                      &record_parse_callchain_opt),
1344         OPT_INCR('v', "verbose", &verbose,
1345                     "be more verbose (show counter open errors, etc)"),
1346         OPT_BOOLEAN('q', "quiet", &quiet, "don't print any message"),
1347         OPT_BOOLEAN('s', "stat", &record.opts.inherit_stat,
1348                     "per thread counts"),
1349         OPT_BOOLEAN('d', "data", &record.opts.sample_address, "Record the sample addresses"),
1350         OPT_BOOLEAN_SET('T', "timestamp", &record.opts.sample_time,
1351                         &record.opts.sample_time_set,
1352                         "Record the sample timestamps"),
1353         OPT_BOOLEAN('P', "period", &record.opts.period, "Record the sample period"),
1354         OPT_BOOLEAN('n', "no-samples", &record.opts.no_samples,
1355                     "don't sample"),
1356         OPT_BOOLEAN_SET('N', "no-buildid-cache", &record.no_buildid_cache,
1357                         &record.no_buildid_cache_set,
1358                         "do not update the buildid cache"),
1359         OPT_BOOLEAN_SET('B', "no-buildid", &record.no_buildid,
1360                         &record.no_buildid_set,
1361                         "do not collect buildids in perf.data"),
1362         OPT_CALLBACK('G', "cgroup", &record.evlist, "name",
1363                      "monitor event in cgroup name only",
1364                      parse_cgroups),
1365         OPT_UINTEGER('D', "delay", &record.opts.initial_delay,
1366                   "ms to wait before starting measurement after program start"),
1367         OPT_STRING('u', "uid", &record.opts.target.uid_str, "user",
1368                    "user to profile"),
1369
1370         OPT_CALLBACK_NOOPT('b', "branch-any", &record.opts.branch_stack,
1371                      "branch any", "sample any taken branches",
1372                      parse_branch_stack),
1373
1374         OPT_CALLBACK('j', "branch-filter", &record.opts.branch_stack,
1375                      "branch filter mask", "branch stack filter modes",
1376                      parse_branch_stack),
1377         OPT_BOOLEAN('W', "weight", &record.opts.sample_weight,
1378                     "sample by weight (on special events only)"),
1379         OPT_BOOLEAN(0, "transaction", &record.opts.sample_transaction,
1380                     "sample transaction flags (special events only)"),
1381         OPT_BOOLEAN(0, "per-thread", &record.opts.target.per_thread,
1382                     "use per-thread mmaps"),
1383         OPT_CALLBACK_OPTARG('I', "intr-regs", &record.opts.sample_intr_regs, NULL, "any register",
1384                     "sample selected machine registers on interrupt,"
1385                     " use -I ? to list register names", parse_regs),
1386         OPT_BOOLEAN(0, "running-time", &record.opts.running_time,
1387                     "Record running/enabled time of read (:S) events"),
1388         OPT_CALLBACK('k', "clockid", &record.opts,
1389         "clockid", "clockid to use for events, see clock_gettime()",
1390         parse_clockid),
1391         OPT_STRING_OPTARG('S', "snapshot", &record.opts.auxtrace_snapshot_opts,
1392                           "opts", "AUX area tracing Snapshot Mode", ""),
1393         OPT_UINTEGER(0, "proc-map-timeout", &record.opts.proc_map_timeout,
1394                         "per thread proc mmap processing timeout in ms"),
1395         OPT_BOOLEAN(0, "switch-events", &record.opts.record_switch_events,
1396                     "Record context switch events"),
1397         OPT_BOOLEAN_FLAG(0, "all-kernel", &record.opts.all_kernel,
1398                          "Configure all used events to run in kernel space.",
1399                          PARSE_OPT_EXCLUSIVE),
1400         OPT_BOOLEAN_FLAG(0, "all-user", &record.opts.all_user,
1401                          "Configure all used events to run in user space.",
1402                          PARSE_OPT_EXCLUSIVE),
1403         OPT_STRING(0, "clang-path", &llvm_param.clang_path, "clang path",
1404                    "clang binary to use for compiling BPF scriptlets"),
1405         OPT_STRING(0, "clang-opt", &llvm_param.clang_opt, "clang options",
1406                    "options passed to clang when compiling BPF scriptlets"),
1407         OPT_STRING(0, "vmlinux", &symbol_conf.vmlinux_name,
1408                    "file", "vmlinux pathname"),
1409         OPT_BOOLEAN(0, "buildid-all", &record.buildid_all,
1410                     "Record build-id of all DSOs regardless of hits"),
1411         OPT_BOOLEAN(0, "timestamp-filename", &record.timestamp_filename,
1412                     "append timestamp to output filename"),
1413         OPT_BOOLEAN(0, "switch-output", &record.switch_output,
1414                     "Switch output when receive SIGUSR2"),
1415         OPT_BOOLEAN(0, "dry-run", &dry_run,
1416                     "Parse options then exit"),
1417         OPT_END()
1418 };
1419
1420 struct option *record_options = __record_options;
1421
1422 int cmd_record(int argc, const char **argv, const char *prefix __maybe_unused)
1423 {
1424         int err;
1425         struct record *rec = &record;
1426         char errbuf[BUFSIZ];
1427
1428 #ifndef HAVE_LIBBPF_SUPPORT
1429 # define set_nobuild(s, l, c) set_option_nobuild(record_options, s, l, "NO_LIBBPF=1", c)
1430         set_nobuild('\0', "clang-path", true);
1431         set_nobuild('\0', "clang-opt", true);
1432 # undef set_nobuild
1433 #endif
1434
1435 #ifndef HAVE_BPF_PROLOGUE
1436 # if !defined (HAVE_DWARF_SUPPORT)
1437 #  define REASON  "NO_DWARF=1"
1438 # elif !defined (HAVE_LIBBPF_SUPPORT)
1439 #  define REASON  "NO_LIBBPF=1"
1440 # else
1441 #  define REASON  "this architecture doesn't support BPF prologue"
1442 # endif
1443 # define set_nobuild(s, l, c) set_option_nobuild(record_options, s, l, REASON, c)
1444         set_nobuild('\0', "vmlinux", true);
1445 # undef set_nobuild
1446 # undef REASON
1447 #endif
1448
1449         rec->evlist = perf_evlist__new();
1450         if (rec->evlist == NULL)
1451                 return -ENOMEM;
1452
1453         perf_config(perf_record_config, rec);
1454
1455         argc = parse_options(argc, argv, record_options, record_usage,
1456                             PARSE_OPT_STOP_AT_NON_OPTION);
1457         if (!argc && target__none(&rec->opts.target))
1458                 usage_with_options(record_usage, record_options);
1459
1460         if (nr_cgroups && !rec->opts.target.system_wide) {
1461                 usage_with_options_msg(record_usage, record_options,
1462                         "cgroup monitoring only available in system-wide mode");
1463
1464         }
1465         if (rec->opts.record_switch_events &&
1466             !perf_can_record_switch_events()) {
1467                 ui__error("kernel does not support recording context switch events\n");
1468                 parse_options_usage(record_usage, record_options, "switch-events", 0);
1469                 return -EINVAL;
1470         }
1471
1472         if (rec->switch_output)
1473                 rec->timestamp_filename = true;
1474
1475         if (!rec->itr) {
1476                 rec->itr = auxtrace_record__init(rec->evlist, &err);
1477                 if (err)
1478                         return err;
1479         }
1480
1481         err = auxtrace_parse_snapshot_options(rec->itr, &rec->opts,
1482                                               rec->opts.auxtrace_snapshot_opts);
1483         if (err)
1484                 return err;
1485
1486         if (dry_run)
1487                 return 0;
1488
1489         err = bpf__setup_stdout(rec->evlist);
1490         if (err) {
1491                 bpf__strerror_setup_stdout(rec->evlist, err, errbuf, sizeof(errbuf));
1492                 pr_err("ERROR: Setup BPF stdout failed: %s\n",
1493                          errbuf);
1494                 return err;
1495         }
1496
1497         err = -ENOMEM;
1498
1499         symbol__init(NULL);
1500
1501         if (symbol_conf.kptr_restrict)
1502                 pr_warning(
1503 "WARNING: Kernel address maps (/proc/{kallsyms,modules}) are restricted,\n"
1504 "check /proc/sys/kernel/kptr_restrict.\n\n"
1505 "Samples in kernel functions may not be resolved if a suitable vmlinux\n"
1506 "file is not found in the buildid cache or in the vmlinux path.\n\n"
1507 "Samples in kernel modules won't be resolved at all.\n\n"
1508 "If some relocation was applied (e.g. kexec) symbols may be misresolved\n"
1509 "even with a suitable vmlinux or kallsyms file.\n\n");
1510
1511         if (rec->no_buildid_cache || rec->no_buildid) {
1512                 disable_buildid_cache();
1513         } else if (rec->switch_output) {
1514                 /*
1515                  * In 'perf record --switch-output', disable buildid
1516                  * generation by default to reduce data file switching
1517                  * overhead. Still generate buildid if they are required
1518                  * explicitly using
1519                  *
1520                  *  perf record --signal-trigger --no-no-buildid \
1521                  *              --no-no-buildid-cache
1522                  *
1523                  * Following code equals to:
1524                  *
1525                  * if ((rec->no_buildid || !rec->no_buildid_set) &&
1526                  *     (rec->no_buildid_cache || !rec->no_buildid_cache_set))
1527                  *         disable_buildid_cache();
1528                  */
1529                 bool disable = true;
1530
1531                 if (rec->no_buildid_set && !rec->no_buildid)
1532                         disable = false;
1533                 if (rec->no_buildid_cache_set && !rec->no_buildid_cache)
1534                         disable = false;
1535                 if (disable) {
1536                         rec->no_buildid = true;
1537                         rec->no_buildid_cache = true;
1538                         disable_buildid_cache();
1539                 }
1540         }
1541
1542         if (rec->evlist->nr_entries == 0 &&
1543             perf_evlist__add_default(rec->evlist) < 0) {
1544                 pr_err("Not enough memory for event selector list\n");
1545                 goto out_symbol_exit;
1546         }
1547
1548         if (rec->opts.target.tid && !rec->opts.no_inherit_set)
1549                 rec->opts.no_inherit = true;
1550
1551         err = target__validate(&rec->opts.target);
1552         if (err) {
1553                 target__strerror(&rec->opts.target, err, errbuf, BUFSIZ);
1554                 ui__warning("%s", errbuf);
1555         }
1556
1557         err = target__parse_uid(&rec->opts.target);
1558         if (err) {
1559                 int saved_errno = errno;
1560
1561                 target__strerror(&rec->opts.target, err, errbuf, BUFSIZ);
1562                 ui__error("%s", errbuf);
1563
1564                 err = -saved_errno;
1565                 goto out_symbol_exit;
1566         }
1567
1568         err = -ENOMEM;
1569         if (perf_evlist__create_maps(rec->evlist, &rec->opts.target) < 0)
1570                 usage_with_options(record_usage, record_options);
1571
1572         err = auxtrace_record__options(rec->itr, rec->evlist, &rec->opts);
1573         if (err)
1574                 goto out_symbol_exit;
1575
1576         /*
1577          * We take all buildids when the file contains
1578          * AUX area tracing data because we do not decode the
1579          * trace because it would take too long.
1580          */
1581         if (rec->opts.full_auxtrace)
1582                 rec->buildid_all = true;
1583
1584         if (record_opts__config(&rec->opts)) {
1585                 err = -EINVAL;
1586                 goto out_symbol_exit;
1587         }
1588
1589         err = __cmd_record(&record, argc, argv);
1590 out_symbol_exit:
1591         perf_evlist__delete(rec->evlist);
1592         symbol__exit();
1593         auxtrace_record__free(rec->itr);
1594         return err;
1595 }
1596
1597 static void snapshot_sig_handler(int sig __maybe_unused)
1598 {
1599         if (trigger_is_ready(&auxtrace_snapshot_trigger)) {
1600                 trigger_hit(&auxtrace_snapshot_trigger);
1601                 auxtrace_record__snapshot_started = 1;
1602                 if (auxtrace_record__snapshot_start(record.itr))
1603                         trigger_error(&auxtrace_snapshot_trigger);
1604         }
1605
1606         if (trigger_is_ready(&switch_output_trigger))
1607                 trigger_hit(&switch_output_trigger);
1608 }