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, struct perf_evlist *evlist, int idx)
 136 {
 137         struct perf_mmap *md = &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(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(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(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_evlist(struct record *rec, struct perf_evlist *evlist)
 502 {
 503         u64 bytes_written = rec->bytes_written;
 504         int i;
 505         int rc = 0;
 506
 507         if (!evlist)
 508                 return 0;
 509
 510         for (i = 0; i < evlist->nr_mmaps; i++) {
 511                 struct auxtrace_mmap *mm = &evlist->mmap[i].auxtrace_mmap;
 512
 513                 if (evlist->mmap[i].base) {
 514                         if (record__mmap_read(rec, evlist, i) != 0) {
 515                                 rc = -1;
 516                                 goto out;
 517                         }
 518                 }
 519
 520                 if (mm->base && !rec->opts.auxtrace_snapshot_mode &&
 521                     record__auxtrace_mmap_read(rec, mm) != 0) {
 522                         rc = -1;
 523                         goto out;
 524                 }
 525         }
 526
 527         /*
 528          * Mark the round finished in case we wrote
 529          * at least one event.
 530          */
 531         if (bytes_written != rec->bytes_written)
 532                 rc = record__write(rec, &finished_round_event, sizeof(finished_round_event));
 533
 534 out:
 535         return rc;
 536 }
 537
 538 static int record__mmap_read_all(struct record *rec)
 539 {
 540         int err;
 541
 542         err = record__mmap_read_evlist(rec, rec->evlist);
 543         if (err)
 544                 return err;
 545
 546         return err;
 547 }
 548
 549 static void record__init_features(struct record *rec)
 550 {
 551         struct perf_session *session = rec->session;
 552         int feat;
 553
 554         for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++)
 555                 perf_header__set_feat(&session->header, feat);
 556
 557         if (rec->no_buildid)
 558                 perf_header__clear_feat(&session->header, HEADER_BUILD_ID);
 559
 560         if (!have_tracepoints(&rec->evlist->entries))
 561                 perf_header__clear_feat(&session->header, HEADER_TRACING_DATA);
 562
 563         if (!rec->opts.branch_stack)
 564                 perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK);
 565
 566         if (!rec->opts.full_auxtrace)
 567                 perf_header__clear_feat(&session->header, HEADER_AUXTRACE);
 568
 569         perf_header__clear_feat(&session->header, HEADER_STAT);
 570 }
 571
 572 static void
 573 record__finish_output(struct record *rec)
 574 {
 575         struct perf_data_file *file = &rec->file;
 576         int fd = perf_data_file__fd(file);
 577
 578         if (file->is_pipe)
 579                 return;
 580
 581         rec->session->header.data_size += rec->bytes_written;
 582         file->size = lseek(perf_data_file__fd(file), 0, SEEK_CUR);
 583
 584         if (!rec->no_buildid) {
 585                 process_buildids(rec);
 586
 587                 if (rec->buildid_all)
 588                         dsos__hit_all(rec->session);
 589         }
 590         perf_session__write_header(rec->session, rec->evlist, fd, true);
 591
 592         return;
 593 }
 594
 595 static int record__synthesize_workload(struct record *rec)
 596 {
 597         struct {
 598                 struct thread_map map;
 599                 struct thread_map_data map_data;
 600         } thread_map;
 601
 602         thread_map.map.nr = 1;
 603         thread_map.map.map[0].pid = rec->evlist->workload.pid;
 604         thread_map.map.map[0].comm = NULL;
 605         return perf_event__synthesize_thread_map(&rec->tool, &thread_map.map,
 606                                                  process_synthesized_event,
 607                                                  &rec->session->machines.host,
 608                                                  rec->opts.sample_address,
 609                                                  rec->opts.proc_map_timeout);
 610 }
 611
 612 static int record__synthesize(struct record *rec);
 613
 614 static int
 615 record__switch_output(struct record *rec, bool at_exit)
 616 {
 617         struct perf_data_file *file = &rec->file;
 618         int fd, err;
 619
 620         /* Same Size:      "2015122520103046"*/
 621         char timestamp[] = "InvalidTimestamp";
 622
 623         rec->samples = 0;
 624         record__finish_output(rec);
 625         err = fetch_current_timestamp(timestamp, sizeof(timestamp));
 626         if (err) {
 627                 pr_err("Failed to get current timestamp\n");
 628                 return -EINVAL;
 629         }
 630
 631         fd = perf_data_file__switch(file, timestamp,
 632                                     rec->session->header.data_offset,
 633                                     at_exit);
 634         if (fd >= 0 && !at_exit) {
 635                 rec->bytes_written = 0;
 636                 rec->session->header.data_size = 0;
 637         }
 638
 639         if (!quiet)
 640                 fprintf(stderr, "[ perf record: Dump %s.%s ]\n",
 641                         file->path, timestamp);
 642
 643         /* Output tracking events */
 644         if (!at_exit) {
 645                 record__synthesize(rec);
 646
 647                 /*
 648                  * In 'perf record --switch-output' without -a,
 649                  * record__synthesize() in record__switch_output() won't
 650                  * generate tracking events because there's no thread_map
 651                  * in evlist. Which causes newly created perf.data doesn't
 652                  * contain map and comm information.
 653                  * Create a fake thread_map and directly call
 654                  * perf_event__synthesize_thread_map() for those events.
 655                  */
 656                 if (target__none(&rec->opts.target))
 657                         record__synthesize_workload(rec);
 658         }
 659         return fd;
 660 }
 661
 662 static volatile int workload_exec_errno;
 663
 664 /*
 665  * perf_evlist__prepare_workload will send a SIGUSR1
 666  * if the fork fails, since we asked by setting its
 667  * want_signal to true.
 668  */
 669 static void workload_exec_failed_signal(int signo __maybe_unused,
 670                                         siginfo_t *info,
 671                                         void *ucontext __maybe_unused)
 672 {
 673         workload_exec_errno = info->si_value.sival_int;
 674         done = 1;
 675         child_finished = 1;
 676 }
 677
 678 static void snapshot_sig_handler(int sig);
 679
 680 int __weak
 681 perf_event__synth_time_conv(const struct perf_event_mmap_page *pc __maybe_unused,
 682                             struct perf_tool *tool __maybe_unused,
 683                             perf_event__handler_t process __maybe_unused,
 684                             struct machine *machine __maybe_unused)
 685 {
 686         return 0;
 687 }
 688
 689 static const struct perf_event_mmap_page *record__pick_pc(struct record *rec)
 690 {
 691         if (rec->evlist && rec->evlist->mmap && rec->evlist->mmap[0].base)
 692                 return rec->evlist->mmap[0].base;
 693         return NULL;
 694 }
 695
 696 static int record__synthesize(struct record *rec)
 697 {
 698         struct perf_session *session = rec->session;
 699         struct machine *machine = &session->machines.host;
 700         struct perf_data_file *file = &rec->file;
 701         struct record_opts *opts = &rec->opts;
 702         struct perf_tool *tool = &rec->tool;
 703         int fd = perf_data_file__fd(file);
 704         int err = 0;
 705
 706         if (file->is_pipe) {
 707                 err = perf_event__synthesize_attrs(tool, session,
 708                                                    process_synthesized_event);
 709                 if (err < 0) {
 710                         pr_err("Couldn't synthesize attrs.\n");
 711                         goto out;
 712                 }
 713
 714                 if (have_tracepoints(&rec->evlist->entries)) {
 715                         /*
 716                          * FIXME err <= 0 here actually means that
 717                          * there were no tracepoints so its not really
 718                          * an error, just that we don't need to
 719                          * synthesize anything.  We really have to
 720                          * return this more properly and also
 721                          * propagate errors that now are calling die()
 722                          */
 723                         err = perf_event__synthesize_tracing_data(tool, fd, rec->evlist,
 724                                                                   process_synthesized_event);
 725                         if (err <= 0) {
 726                                 pr_err("Couldn't record tracing data.\n");
 727                                 goto out;
 728                         }
 729                         rec->bytes_written += err;
 730                 }
 731         }
 732
 733         err = perf_event__synth_time_conv(record__pick_pc(rec), tool,
 734                                           process_synthesized_event, machine);
 735         if (err)
 736                 goto out;
 737
 738         if (rec->opts.full_auxtrace) {
 739                 err = perf_event__synthesize_auxtrace_info(rec->itr, tool,
 740                                         session, process_synthesized_event);
 741                 if (err)
 742                         goto out;
 743         }
 744
 745         err = perf_event__synthesize_kernel_mmap(tool, process_synthesized_event,
 746                                                  machine);
 747         WARN_ONCE(err < 0, "Couldn't record kernel reference relocation symbol\n"
 748                            "Symbol resolution may be skewed if relocation was used (e.g. kexec).\n"
 749                            "Check /proc/kallsyms permission or run as root.\n");
 750
 751         err = perf_event__synthesize_modules(tool, process_synthesized_event,
 752                                              machine);
 753         WARN_ONCE(err < 0, "Couldn't record kernel module information.\n"
 754                            "Symbol resolution may be skewed if relocation was used (e.g. kexec).\n"
 755                            "Check /proc/modules permission or run as root.\n");
 756
 757         if (perf_guest) {
 758                 machines__process_guests(&session->machines,
 759                                          perf_event__synthesize_guest_os, tool);
 760         }
 761
 762         err = __machine__synthesize_threads(machine, tool, &opts->target, rec->evlist->threads,
 763                                             process_synthesized_event, opts->sample_address,
 764                                             opts->proc_map_timeout);
 765 out:
 766         return err;
 767 }
 768
 769 static int __cmd_record(struct record *rec, int argc, const char **argv)
 770 {
 771         int err;
 772         int status = 0;
 773         unsigned long waking = 0;
 774         const bool forks = argc > 0;
 775         struct machine *machine;
 776         struct perf_tool *tool = &rec->tool;
 777         struct record_opts *opts = &rec->opts;
 778         struct perf_data_file *file = &rec->file;
 779         struct perf_session *session;
 780         bool disabled = false, draining = false;
 781         int fd;
 782
 783         rec->progname = argv[0];
 784
 785         atexit(record__sig_exit);
 786         signal(SIGCHLD, sig_handler);
 787         signal(SIGINT, sig_handler);
 788         signal(SIGTERM, sig_handler);
 789
 790         if (rec->opts.auxtrace_snapshot_mode || rec->switch_output) {
 791                 signal(SIGUSR2, snapshot_sig_handler);
 792                 if (rec->opts.auxtrace_snapshot_mode)
 793                         trigger_on(&auxtrace_snapshot_trigger);
 794                 if (rec->switch_output)
 795                         trigger_on(&switch_output_trigger);
 796         } else {
 797                 signal(SIGUSR2, SIG_IGN);
 798         }
 799
 800         session = perf_session__new(file, false, tool);
 801         if (session == NULL) {
 802                 pr_err("Perf session creation failed.\n");
 803                 return -1;
 804         }
 805
 806         fd = perf_data_file__fd(file);
 807         rec->session = session;
 808
 809         record__init_features(rec);
 810
 811         if (forks) {
 812                 err = perf_evlist__prepare_workload(rec->evlist, &opts->target,
 813                                                     argv, file->is_pipe,
 814                                                     workload_exec_failed_signal);
 815                 if (err < 0) {
 816                         pr_err("Couldn't run the workload!\n");
 817                         status = err;
 818                         goto out_delete_session;
 819                 }
 820         }
 821
 822         if (record__open(rec) != 0) {
 823                 err = -1;
 824                 goto out_child;
 825         }
 826
 827         err = bpf__apply_obj_config();
 828         if (err) {
 829                 char errbuf[BUFSIZ];
 830
 831                 bpf__strerror_apply_obj_config(err, errbuf, sizeof(errbuf));
 832                 pr_err("ERROR: Apply config to BPF failed: %s\n",
 833                          errbuf);
 834                 goto out_child;
 835         }
 836
 837         /*
 838          * Normally perf_session__new would do this, but it doesn't have the
 839          * evlist.
 840          */
 841         if (rec->tool.ordered_events && !perf_evlist__sample_id_all(rec->evlist)) {
 842                 pr_warning("WARNING: No sample_id_all support, falling back to unordered processing\n");
 843                 rec->tool.ordered_events = false;
 844         }
 845
 846         if (!rec->evlist->nr_groups)
 847                 perf_header__clear_feat(&session->header, HEADER_GROUP_DESC);
 848
 849         if (file->is_pipe) {
 850                 err = perf_header__write_pipe(fd);
 851                 if (err < 0)
 852                         goto out_child;
 853         } else {
 854                 err = perf_session__write_header(session, rec->evlist, fd, false);
 855                 if (err < 0)
 856                         goto out_child;
 857         }
 858
 859         if (!rec->no_buildid
 860             && !perf_header__has_feat(&session->header, HEADER_BUILD_ID)) {
 861                 pr_err("Couldn't generate buildids. "
 862                        "Use --no-buildid to profile anyway.\n");
 863                 err = -1;
 864                 goto out_child;
 865         }
 866
 867         machine = &session->machines.host;
 868
 869         err = record__synthesize(rec);
 870         if (err < 0)
 871                 goto out_child;
 872
 873         if (rec->realtime_prio) {
 874                 struct sched_param param;
 875
 876                 param.sched_priority = rec->realtime_prio;
 877                 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
 878                         pr_err("Could not set realtime priority.\n");
 879                         err = -1;
 880                         goto out_child;
 881                 }
 882         }
 883
 884         /*
 885          * When perf is starting the traced process, all the events
 886          * (apart from group members) have enable_on_exec=1 set,
 887          * so don't spoil it by prematurely enabling them.
 888          */
 889         if (!target__none(&opts->target) && !opts->initial_delay)
 890                 perf_evlist__enable(rec->evlist);
 891
 892         /*
 893          * Let the child rip
 894          */
 895         if (forks) {
 896                 union perf_event *event;
 897
 898                 event = malloc(sizeof(event->comm) + machine->id_hdr_size);
 899                 if (event == NULL) {
 900                         err = -ENOMEM;
 901                         goto out_child;
 902                 }
 903
 904                 /*
 905                  * Some H/W events are generated before COMM event
 906                  * which is emitted during exec(), so perf script
 907                  * cannot see a correct process name for those events.
 908                  * Synthesize COMM event to prevent it.
 909                  */
 910                 perf_event__synthesize_comm(tool, event,
 911                                             rec->evlist->workload.pid,
 912                                             process_synthesized_event,
 913                                             machine);
 914                 free(event);
 915
 916                 perf_evlist__start_workload(rec->evlist);
 917         }
 918
 919         if (opts->initial_delay) {
 920                 usleep(opts->initial_delay * 1000);
 921                 perf_evlist__enable(rec->evlist);
 922         }
 923
 924         trigger_ready(&auxtrace_snapshot_trigger);
 925         trigger_ready(&switch_output_trigger);
 926         for (;;) {
 927                 unsigned long long hits = rec->samples;
 928
 929                 if (record__mmap_read_all(rec) < 0) {
 930                         trigger_error(&auxtrace_snapshot_trigger);
 931                         trigger_error(&switch_output_trigger);
 932                         err = -1;
 933                         goto out_child;
 934                 }
 935
 936                 if (auxtrace_record__snapshot_started) {
 937                         auxtrace_record__snapshot_started = 0;
 938                         if (!trigger_is_error(&auxtrace_snapshot_trigger))
 939                                 record__read_auxtrace_snapshot(rec);
 940                         if (trigger_is_error(&auxtrace_snapshot_trigger)) {
 941                                 pr_err("AUX area tracing snapshot failed\n");
 942                                 err = -1;
 943                                 goto out_child;
 944                         }
 945                 }
 946
 947                 if (trigger_is_hit(&switch_output_trigger)) {
 948                         trigger_ready(&switch_output_trigger);
 949
 950                         if (!quiet)
 951                                 fprintf(stderr, "[ perf record: dump data: Woken up %ld times ]\n",
 952                                         waking);
 953                         waking = 0;
 954                         fd = record__switch_output(rec, false);
 955                         if (fd < 0) {
 956                                 pr_err("Failed to switch to new file\n");
 957                                 trigger_error(&switch_output_trigger);
 958                                 err = fd;
 959                                 goto out_child;
 960                         }
 961                 }
 962
 963                 if (hits == rec->samples) {
 964                         if (done || draining)
 965                                 break;
 966                         err = perf_evlist__poll(rec->evlist, -1);
 967                         /*
 968                          * Propagate error, only if there's any. Ignore positive
 969                          * number of returned events and interrupt error.
 970                          */
 971                         if (err > 0 || (err < 0 && errno == EINTR))
 972                                 err = 0;
 973                         waking++;
 974
 975                         if (perf_evlist__filter_pollfd(rec->evlist, POLLERR | POLLHUP) == 0)
 976                                 draining = true;
 977                 }
 978
 979                 /*
 980                  * When perf is starting the traced process, at the end events
 981                  * die with the process and we wait for that. Thus no need to
 982                  * disable events in this case.
 983                  */
 984                 if (done && !disabled && !target__none(&opts->target)) {
 985                         trigger_off(&auxtrace_snapshot_trigger);
 986                         perf_evlist__disable(rec->evlist);
 987                         disabled = true;
 988                 }
 989         }
 990         trigger_off(&auxtrace_snapshot_trigger);
 991         trigger_off(&switch_output_trigger);
 992
 993         if (forks && workload_exec_errno) {
 994                 char msg[STRERR_BUFSIZE];
 995                 const char *emsg = strerror_r(workload_exec_errno, msg, sizeof(msg));
 996                 pr_err("Workload failed: %s\n", emsg);
 997                 err = -1;
 998                 goto out_child;
 999         }
1000
1001         if (!quiet)
1002                 fprintf(stderr, "[ perf record: Woken up %ld times to write data ]\n", waking);
1003
1004 out_child:
1005         if (forks) {
1006                 int exit_status;
1007
1008                 if (!child_finished)
1009                         kill(rec->evlist->workload.pid, SIGTERM);
1010
1011                 wait(&exit_status);
1012
1013                 if (err < 0)
1014                         status = err;
1015                 else if (WIFEXITED(exit_status))
1016                         status = WEXITSTATUS(exit_status);
1017                 else if (WIFSIGNALED(exit_status))
1018                         signr = WTERMSIG(exit_status);
1019         } else
1020                 status = err;
1021
1022         /* this will be recalculated during process_buildids() */
1023         rec->samples = 0;
1024
1025         if (!err) {
1026                 if (!rec->timestamp_filename) {
1027                         record__finish_output(rec);
1028                 } else {
1029                         fd = record__switch_output(rec, true);
1030                         if (fd < 0) {
1031                                 status = fd;
1032                                 goto out_delete_session;
1033                         }
1034                 }
1035         }
1036
1037         if (!err && !quiet) {
1038                 char samples[128];
1039                 const char *postfix = rec->timestamp_filename ?
1040                                         ".<timestamp>" : "";
1041
1042                 if (rec->samples && !rec->opts.full_auxtrace)
1043                         scnprintf(samples, sizeof(samples),
1044                                   " (%" PRIu64 " samples)", rec->samples);
1045                 else
1046                         samples[0] = '\0';
1047
1048                 fprintf(stderr, "[ perf record: Captured and wrote %.3f MB %s%s%s ]\n",
1049                         perf_data_file__size(file) / 1024.0 / 1024.0,
1050                         file->path, postfix, samples);
1051         }
1052
1053 out_delete_session:
1054         perf_session__delete(session);
1055         return status;
1056 }
1057
1058 static void callchain_debug(struct callchain_param *callchain)
1059 {
1060         static const char *str[CALLCHAIN_MAX] = { "NONE", "FP", "DWARF", "LBR" };
1061
1062         pr_debug("callchain: type %s\n", str[callchain->record_mode]);
1063
1064         if (callchain->record_mode == CALLCHAIN_DWARF)
1065                 pr_debug("callchain: stack dump size %d\n",
1066                          callchain->dump_size);
1067 }
1068
1069 int record_opts__parse_callchain(struct record_opts *record,
1070                                  struct callchain_param *callchain,
1071                                  const char *arg, bool unset)
1072 {
1073         int ret;
1074         callchain->enabled = !unset;
1075
1076         /* --no-call-graph */
1077         if (unset) {
1078                 callchain->record_mode = CALLCHAIN_NONE;
1079                 pr_debug("callchain: disabled\n");
1080                 return 0;
1081         }
1082
1083         ret = parse_callchain_record_opt(arg, callchain);
1084         if (!ret) {
1085                 /* Enable data address sampling for DWARF unwind. */
1086                 if (callchain->record_mode == CALLCHAIN_DWARF)
1087                         record->sample_address = true;
1088                 callchain_debug(callchain);
1089         }
1090
1091         return ret;
1092 }
1093
1094 int record_parse_callchain_opt(const struct option *opt,
1095                                const char *arg,
1096                                int unset)
1097 {
1098         return record_opts__parse_callchain(opt->value, &callchain_param, arg, unset);
1099 }
1100
1101 int record_callchain_opt(const struct option *opt,
1102                          const char *arg __maybe_unused,
1103                          int unset __maybe_unused)
1104 {
1105         struct callchain_param *callchain = opt->value;
1106
1107         callchain->enabled = true;
1108
1109         if (callchain->record_mode == CALLCHAIN_NONE)
1110                 callchain->record_mode = CALLCHAIN_FP;
1111
1112         callchain_debug(callchain);
1113         return 0;
1114 }
1115
1116 static int perf_record_config(const char *var, const char *value, void *cb)
1117 {
1118         struct record *rec = cb;
1119
1120         if (!strcmp(var, "record.build-id")) {
1121                 if (!strcmp(value, "cache"))
1122                         rec->no_buildid_cache = false;
1123                 else if (!strcmp(value, "no-cache"))
1124                         rec->no_buildid_cache = true;
1125                 else if (!strcmp(value, "skip"))
1126                         rec->no_buildid = true;
1127                 else
1128                         return -1;
1129                 return 0;
1130         }
1131         if (!strcmp(var, "record.call-graph"))
1132                 var = "call-graph.record-mode"; /* fall-through */
1133
1134         return perf_default_config(var, value, cb);
1135 }
1136
1137 struct clockid_map {
1138         const char *name;
1139         int clockid;
1140 };
1141
1142 #define CLOCKID_MAP(n, c)       \
1143         { .name = n, .clockid = (c), }
1144
1145 #define CLOCKID_END     { .name = NULL, }
1146
1147
1148 /*
1149  * Add the missing ones, we need to build on many distros...
1150  */
1151 #ifndef CLOCK_MONOTONIC_RAW
1152 #define CLOCK_MONOTONIC_RAW 4
1153 #endif
1154 #ifndef CLOCK_BOOTTIME
1155 #define CLOCK_BOOTTIME 7
1156 #endif
1157 #ifndef CLOCK_TAI
1158 #define CLOCK_TAI 11
1159 #endif
1160
1161 static const struct clockid_map clockids[] = {
1162         /* available for all events, NMI safe */
1163         CLOCKID_MAP("monotonic", CLOCK_MONOTONIC),
1164         CLOCKID_MAP("monotonic_raw", CLOCK_MONOTONIC_RAW),
1165
1166         /* available for some events */
1167         CLOCKID_MAP("realtime", CLOCK_REALTIME),
1168         CLOCKID_MAP("boottime", CLOCK_BOOTTIME),
1169         CLOCKID_MAP("tai", CLOCK_TAI),
1170
1171         /* available for the lazy */
1172         CLOCKID_MAP("mono", CLOCK_MONOTONIC),
1173         CLOCKID_MAP("raw", CLOCK_MONOTONIC_RAW),
1174         CLOCKID_MAP("real", CLOCK_REALTIME),
1175         CLOCKID_MAP("boot", CLOCK_BOOTTIME),
1176
1177         CLOCKID_END,
1178 };
1179
1180 static int parse_clockid(const struct option *opt, const char *str, int unset)
1181 {
1182         struct record_opts *opts = (struct record_opts *)opt->value;
1183         const struct clockid_map *cm;
1184         const char *ostr = str;
1185
1186         if (unset) {
1187                 opts->use_clockid = 0;
1188                 return 0;
1189         }
1190
1191         /* no arg passed */
1192         if (!str)
1193                 return 0;
1194
1195         /* no setting it twice */
1196         if (opts->use_clockid)
1197                 return -1;
1198
1199         opts->use_clockid = true;
1200
1201         /* if its a number, we're done */
1202         if (sscanf(str, "%d", &opts->clockid) == 1)
1203                 return 0;
1204
1205         /* allow a "CLOCK_" prefix to the name */
1206         if (!strncasecmp(str, "CLOCK_", 6))
1207                 str += 6;
1208
1209         for (cm = clockids; cm->name; cm++) {
1210                 if (!strcasecmp(str, cm->name)) {
1211                         opts->clockid = cm->clockid;
1212                         return 0;
1213                 }
1214         }
1215
1216         opts->use_clockid = false;
1217         ui__warning("unknown clockid %s, check man page\n", ostr);
1218         return -1;
1219 }
1220
1221 static int record__parse_mmap_pages(const struct option *opt,
1222                                     const char *str,
1223                                     int unset __maybe_unused)
1224 {
1225         struct record_opts *opts = opt->value;
1226         char *s, *p;
1227         unsigned int mmap_pages;
1228         int ret;
1229
1230         if (!str)
1231                 return -EINVAL;
1232
1233         s = strdup(str);
1234         if (!s)
1235                 return -ENOMEM;
1236
1237         p = strchr(s, ',');
1238         if (p)
1239                 *p = '\0';
1240
1241         if (*s) {
1242                 ret = __perf_evlist__parse_mmap_pages(&mmap_pages, s);
1243                 if (ret)
1244                         goto out_free;
1245                 opts->mmap_pages = mmap_pages;
1246         }
1247
1248         if (!p) {
1249                 ret = 0;
1250                 goto out_free;
1251         }
1252
1253         ret = __perf_evlist__parse_mmap_pages(&mmap_pages, p + 1);
1254         if (ret)
1255                 goto out_free;
1256
1257         opts->auxtrace_mmap_pages = mmap_pages;
1258
1259 out_free:
1260         free(s);
1261         return ret;
1262 }
1263
1264 static const char * const __record_usage[] = {
1265         "perf record [<options>] [<command>]",
1266         "perf record [<options>] -- <command> [<options>]",
1267         NULL
1268 };
1269 const char * const *record_usage = __record_usage;
1270
1271 /*
1272  * XXX Ideally would be local to cmd_record() and passed to a record__new
1273  * because we need to have access to it in record__exit, that is called
1274  * after cmd_record() exits, but since record_options need to be accessible to
1275  * builtin-script, leave it here.
1276  *
1277  * At least we don't ouch it in all the other functions here directly.
1278  *
1279  * Just say no to tons of global variables, sigh.
1280  */
1281 static struct record record = {
1282         .opts = {
1283                 .sample_time         = true,
1284                 .mmap_pages          = UINT_MAX,
1285                 .user_freq           = UINT_MAX,
1286                 .user_interval       = ULLONG_MAX,
1287                 .freq                = 4000,
1288                 .target              = {
1289                         .uses_mmap   = true,
1290                         .default_per_cpu = true,
1291                 },
1292                 .proc_map_timeout     = 500,
1293         },
1294         .tool = {
1295                 .sample         = process_sample_event,
1296                 .fork           = perf_event__process_fork,
1297                 .exit           = perf_event__process_exit,
1298                 .comm           = perf_event__process_comm,
1299                 .mmap           = perf_event__process_mmap,
1300                 .mmap2          = perf_event__process_mmap2,
1301                 .ordered_events = true,
1302         },
1303 };
1304
1305 const char record_callchain_help[] = CALLCHAIN_RECORD_HELP
1306         "\n\t\t\t\tDefault: fp";
1307
1308 static bool dry_run;
1309
1310 /*
1311  * XXX Will stay a global variable till we fix builtin-script.c to stop messing
1312  * with it and switch to use the library functions in perf_evlist that came
1313  * from builtin-record.c, i.e. use record_opts,
1314  * perf_evlist__prepare_workload, etc instead of fork+exec'in 'perf record',
1315  * using pipes, etc.
1316  */
1317 struct option __record_options[] = {
1318         OPT_CALLBACK('e', "event", &record.evlist, "event",
1319                      "event selector. use 'perf list' to list available events",
1320                      parse_events_option),
1321         OPT_CALLBACK(0, "filter", &record.evlist, "filter",
1322                      "event filter", parse_filter),
1323         OPT_CALLBACK_NOOPT(0, "exclude-perf", &record.evlist,
1324                            NULL, "don't record events from perf itself",
1325                            exclude_perf),
1326         OPT_STRING('p', "pid", &record.opts.target.pid, "pid",
1327                     "record events on existing process id"),
1328         OPT_STRING('t', "tid", &record.opts.target.tid, "tid",
1329                     "record events on existing thread id"),
1330         OPT_INTEGER('r', "realtime", &record.realtime_prio,
1331                     "collect data with this RT SCHED_FIFO priority"),
1332         OPT_BOOLEAN(0, "no-buffering", &record.opts.no_buffering,
1333                     "collect data without buffering"),
1334         OPT_BOOLEAN('R', "raw-samples", &record.opts.raw_samples,
1335                     "collect raw sample records from all opened counters"),
1336         OPT_BOOLEAN('a', "all-cpus", &record.opts.target.system_wide,
1337                             "system-wide collection from all CPUs"),
1338         OPT_STRING('C', "cpu", &record.opts.target.cpu_list, "cpu",
1339                     "list of cpus to monitor"),
1340         OPT_U64('c', "count", &record.opts.user_interval, "event period to sample"),
1341         OPT_STRING('o', "output", &record.file.path, "file",
1342                     "output file name"),
1343         OPT_BOOLEAN_SET('i', "no-inherit", &record.opts.no_inherit,
1344                         &record.opts.no_inherit_set,
1345                         "child tasks do not inherit counters"),
1346         OPT_UINTEGER('F', "freq", &record.opts.user_freq, "profile at this frequency"),
1347         OPT_CALLBACK('m', "mmap-pages", &record.opts, "pages[,pages]",
1348                      "number of mmap data pages and AUX area tracing mmap pages",
1349                      record__parse_mmap_pages),
1350         OPT_BOOLEAN(0, "group", &record.opts.group,
1351                     "put the counters into a counter group"),
1352         OPT_CALLBACK_NOOPT('g', NULL, &callchain_param,
1353                            NULL, "enables call-graph recording" ,
1354                            &record_callchain_opt),
1355         OPT_CALLBACK(0, "call-graph", &record.opts,
1356                      "record_mode[,record_size]", record_callchain_help,
1357                      &record_parse_callchain_opt),
1358         OPT_INCR('v', "verbose", &verbose,
1359                     "be more verbose (show counter open errors, etc)"),
1360         OPT_BOOLEAN('q', "quiet", &quiet, "don't print any message"),
1361         OPT_BOOLEAN('s', "stat", &record.opts.inherit_stat,
1362                     "per thread counts"),
1363         OPT_BOOLEAN('d', "data", &record.opts.sample_address, "Record the sample addresses"),
1364         OPT_BOOLEAN_SET('T', "timestamp", &record.opts.sample_time,
1365                         &record.opts.sample_time_set,
1366                         "Record the sample timestamps"),
1367         OPT_BOOLEAN('P', "period", &record.opts.period, "Record the sample period"),
1368         OPT_BOOLEAN('n', "no-samples", &record.opts.no_samples,
1369                     "don't sample"),
1370         OPT_BOOLEAN_SET('N', "no-buildid-cache", &record.no_buildid_cache,
1371                         &record.no_buildid_cache_set,
1372                         "do not update the buildid cache"),
1373         OPT_BOOLEAN_SET('B', "no-buildid", &record.no_buildid,
1374                         &record.no_buildid_set,
1375                         "do not collect buildids in perf.data"),
1376         OPT_CALLBACK('G', "cgroup", &record.evlist, "name",
1377                      "monitor event in cgroup name only",
1378                      parse_cgroups),
1379         OPT_UINTEGER('D', "delay", &record.opts.initial_delay,
1380                   "ms to wait before starting measurement after program start"),
1381         OPT_STRING('u', "uid", &record.opts.target.uid_str, "user",
1382                    "user to profile"),
1383
1384         OPT_CALLBACK_NOOPT('b', "branch-any", &record.opts.branch_stack,
1385                      "branch any", "sample any taken branches",
1386                      parse_branch_stack),
1387
1388         OPT_CALLBACK('j', "branch-filter", &record.opts.branch_stack,
1389                      "branch filter mask", "branch stack filter modes",
1390                      parse_branch_stack),
1391         OPT_BOOLEAN('W', "weight", &record.opts.sample_weight,
1392                     "sample by weight (on special events only)"),
1393         OPT_BOOLEAN(0, "transaction", &record.opts.sample_transaction,
1394                     "sample transaction flags (special events only)"),
1395         OPT_BOOLEAN(0, "per-thread", &record.opts.target.per_thread,
1396                     "use per-thread mmaps"),
1397         OPT_CALLBACK_OPTARG('I', "intr-regs", &record.opts.sample_intr_regs, NULL, "any register",
1398                     "sample selected machine registers on interrupt,"
1399                     " use -I ? to list register names", parse_regs),
1400         OPT_BOOLEAN(0, "running-time", &record.opts.running_time,
1401                     "Record running/enabled time of read (:S) events"),
1402         OPT_CALLBACK('k', "clockid", &record.opts,
1403         "clockid", "clockid to use for events, see clock_gettime()",
1404         parse_clockid),
1405         OPT_STRING_OPTARG('S', "snapshot", &record.opts.auxtrace_snapshot_opts,
1406                           "opts", "AUX area tracing Snapshot Mode", ""),
1407         OPT_UINTEGER(0, "proc-map-timeout", &record.opts.proc_map_timeout,
1408                         "per thread proc mmap processing timeout in ms"),
1409         OPT_BOOLEAN(0, "switch-events", &record.opts.record_switch_events,
1410                     "Record context switch events"),
1411         OPT_BOOLEAN_FLAG(0, "all-kernel", &record.opts.all_kernel,
1412                          "Configure all used events to run in kernel space.",
1413                          PARSE_OPT_EXCLUSIVE),
1414         OPT_BOOLEAN_FLAG(0, "all-user", &record.opts.all_user,
1415                          "Configure all used events to run in user space.",
1416                          PARSE_OPT_EXCLUSIVE),
1417         OPT_STRING(0, "clang-path", &llvm_param.clang_path, "clang path",
1418                    "clang binary to use for compiling BPF scriptlets"),
1419         OPT_STRING(0, "clang-opt", &llvm_param.clang_opt, "clang options",
1420                    "options passed to clang when compiling BPF scriptlets"),
1421         OPT_STRING(0, "vmlinux", &symbol_conf.vmlinux_name,
1422                    "file", "vmlinux pathname"),
1423         OPT_BOOLEAN(0, "buildid-all", &record.buildid_all,
1424                     "Record build-id of all DSOs regardless of hits"),
1425         OPT_BOOLEAN(0, "timestamp-filename", &record.timestamp_filename,
1426                     "append timestamp to output filename"),
1427         OPT_BOOLEAN(0, "switch-output", &record.switch_output,
1428                     "Switch output when receive SIGUSR2"),
1429         OPT_BOOLEAN(0, "dry-run", &dry_run,
1430                     "Parse options then exit"),
1431         OPT_END()
1432 };
1433
1434 struct option *record_options = __record_options;
1435
1436 int cmd_record(int argc, const char **argv, const char *prefix __maybe_unused)
1437 {
1438         int err;
1439         struct record *rec = &record;
1440         char errbuf[BUFSIZ];
1441
1442 #ifndef HAVE_LIBBPF_SUPPORT
1443 # define set_nobuild(s, l, c) set_option_nobuild(record_options, s, l, "NO_LIBBPF=1", c)
1444         set_nobuild('\0', "clang-path", true);
1445         set_nobuild('\0', "clang-opt", true);
1446 # undef set_nobuild
1447 #endif
1448
1449 #ifndef HAVE_BPF_PROLOGUE
1450 # if !defined (HAVE_DWARF_SUPPORT)
1451 #  define REASON  "NO_DWARF=1"
1452 # elif !defined (HAVE_LIBBPF_SUPPORT)
1453 #  define REASON  "NO_LIBBPF=1"
1454 # else
1455 #  define REASON  "this architecture doesn't support BPF prologue"
1456 # endif
1457 # define set_nobuild(s, l, c) set_option_nobuild(record_options, s, l, REASON, c)
1458         set_nobuild('\0', "vmlinux", true);
1459 # undef set_nobuild
1460 # undef REASON
1461 #endif
1462
1463         rec->evlist = perf_evlist__new();
1464         if (rec->evlist == NULL)
1465                 return -ENOMEM;
1466
1467         perf_config(perf_record_config, rec);
1468
1469         argc = parse_options(argc, argv, record_options, record_usage,
1470                             PARSE_OPT_STOP_AT_NON_OPTION);
1471         if (!argc && target__none(&rec->opts.target))
1472                 usage_with_options(record_usage, record_options);
1473
1474         if (nr_cgroups && !rec->opts.target.system_wide) {
1475                 usage_with_options_msg(record_usage, record_options,
1476                         "cgroup monitoring only available in system-wide mode");
1477
1478         }
1479         if (rec->opts.record_switch_events &&
1480             !perf_can_record_switch_events()) {
1481                 ui__error("kernel does not support recording context switch events\n");
1482                 parse_options_usage(record_usage, record_options, "switch-events", 0);
1483                 return -EINVAL;
1484         }
1485
1486         if (rec->switch_output)
1487                 rec->timestamp_filename = true;
1488
1489         if (!rec->itr) {
1490                 rec->itr = auxtrace_record__init(rec->evlist, &err);
1491                 if (err)
1492                         return err;
1493         }
1494
1495         err = auxtrace_parse_snapshot_options(rec->itr, &rec->opts,
1496                                               rec->opts.auxtrace_snapshot_opts);
1497         if (err)
1498                 return err;
1499
1500         if (dry_run)
1501                 return 0;
1502
1503         err = bpf__setup_stdout(rec->evlist);
1504         if (err) {
1505                 bpf__strerror_setup_stdout(rec->evlist, err, errbuf, sizeof(errbuf));
1506                 pr_err("ERROR: Setup BPF stdout failed: %s\n",
1507                          errbuf);
1508                 return err;
1509         }
1510
1511         err = -ENOMEM;
1512
1513         symbol__init(NULL);
1514
1515         if (symbol_conf.kptr_restrict)
1516                 pr_warning(
1517 "WARNING: Kernel address maps (/proc/{kallsyms,modules}) are restricted,\n"
1518 "check /proc/sys/kernel/kptr_restrict.\n\n"
1519 "Samples in kernel functions may not be resolved if a suitable vmlinux\n"
1520 "file is not found in the buildid cache or in the vmlinux path.\n\n"
1521 "Samples in kernel modules won't be resolved at all.\n\n"
1522 "If some relocation was applied (e.g. kexec) symbols may be misresolved\n"
1523 "even with a suitable vmlinux or kallsyms file.\n\n");
1524
1525         if (rec->no_buildid_cache || rec->no_buildid) {
1526                 disable_buildid_cache();
1527         } else if (rec->switch_output) {
1528                 /*
1529                  * In 'perf record --switch-output', disable buildid
1530                  * generation by default to reduce data file switching
1531                  * overhead. Still generate buildid if they are required
1532                  * explicitly using
1533                  *
1534                  *  perf record --signal-trigger --no-no-buildid \
1535                  *              --no-no-buildid-cache
1536                  *
1537                  * Following code equals to:
1538                  *
1539                  * if ((rec->no_buildid || !rec->no_buildid_set) &&
1540                  *     (rec->no_buildid_cache || !rec->no_buildid_cache_set))
1541                  *         disable_buildid_cache();
1542                  */
1543                 bool disable = true;
1544
1545                 if (rec->no_buildid_set && !rec->no_buildid)
1546                         disable = false;
1547                 if (rec->no_buildid_cache_set && !rec->no_buildid_cache)
1548                         disable = false;
1549                 if (disable) {
1550                         rec->no_buildid = true;
1551                         rec->no_buildid_cache = true;
1552                         disable_buildid_cache();
1553                 }
1554         }
1555
1556         if (rec->evlist->nr_entries == 0 &&
1557             perf_evlist__add_default(rec->evlist) < 0) {
1558                 pr_err("Not enough memory for event selector list\n");
1559                 goto out_symbol_exit;
1560         }
1561
1562         if (rec->opts.target.tid && !rec->opts.no_inherit_set)
1563                 rec->opts.no_inherit = true;
1564
1565         err = target__validate(&rec->opts.target);
1566         if (err) {
1567                 target__strerror(&rec->opts.target, err, errbuf, BUFSIZ);
1568                 ui__warning("%s", errbuf);
1569         }
1570
1571         err = target__parse_uid(&rec->opts.target);
1572         if (err) {
1573                 int saved_errno = errno;
1574
1575                 target__strerror(&rec->opts.target, err, errbuf, BUFSIZ);
1576                 ui__error("%s", errbuf);
1577
1578                 err = -saved_errno;
1579                 goto out_symbol_exit;
1580         }
1581
1582         err = -ENOMEM;
1583         if (perf_evlist__create_maps(rec->evlist, &rec->opts.target) < 0)
1584                 usage_with_options(record_usage, record_options);
1585
1586         err = auxtrace_record__options(rec->itr, rec->evlist, &rec->opts);
1587         if (err)
1588                 goto out_symbol_exit;
1589
1590         /*
1591          * We take all buildids when the file contains
1592          * AUX area tracing data because we do not decode the
1593          * trace because it would take too long.
1594          */
1595         if (rec->opts.full_auxtrace)
1596                 rec->buildid_all = true;
1597
1598         if (record_opts__config(&rec->opts)) {
1599                 err = -EINVAL;
1600                 goto out_symbol_exit;
1601         }
1602
1603         err = __cmd_record(&record, argc, argv);
1604 out_symbol_exit:
1605         perf_evlist__delete(rec->evlist);
1606         symbol__exit();
1607         auxtrace_record__free(rec->itr);
1608         return err;
1609 }
1610
1611 static void snapshot_sig_handler(int sig __maybe_unused)
1612 {
1613         if (trigger_is_ready(&auxtrace_snapshot_trigger)) {
1614                 trigger_hit(&auxtrace_snapshot_trigger);
1615                 auxtrace_record__snapshot_started = 1;
1616                 if (auxtrace_record__snapshot_start(record.itr))
1617                         trigger_error(&auxtrace_snapshot_trigger);
1618         }
1619
1620         if (trigger_is_ready(&switch_output_trigger))
1621                 trigger_hit(&switch_output_trigger);
1622 }