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(void *data, int mask, u64 head, u64 old,
 123               u64 *start, u64 *end, bool backward)
 124 {
 125         if (!backward) {
 126                 *start = old;
 127                 *end = head;
 128                 return 0;
 129         }
 130
 131         return backward_rb_find_range(data, mask, head, start, end);
 132 }
 133
 134 static int
 135 record__mmap_read(struct record *rec, struct perf_mmap *md,
 136                   bool overwrite, bool backward)
 137 {
 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(data, md->mask, head,
 147                           old, &start, &end, backward))
 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_mmap__consume(md, overwrite || backward);
 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_mmap__consume(md, overwrite || backward);
 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                                 str_error_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                         str_error_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                                     bool backward)
 503 {
 504         u64 bytes_written = rec->bytes_written;
 505         int i;
 506         int rc = 0;
 507         struct perf_mmap *maps;
 508
 509         if (!evlist)
 510                 return 0;
 511
 512         maps = backward ? evlist->backward_mmap : evlist->mmap;
 513         if (!maps)
 514                 return 0;
 515
 516         if (backward && evlist->bkw_mmap_state != BKW_MMAP_DATA_PENDING)
 517                 return 0;
 518
 519         for (i = 0; i < evlist->nr_mmaps; i++) {
 520                 struct auxtrace_mmap *mm = &maps[i].auxtrace_mmap;
 521
 522                 if (maps[i].base) {
 523                         if (record__mmap_read(rec, &maps[i],
 524                                               evlist->overwrite, backward) != 0) {
 525                                 rc = -1;
 526                                 goto out;
 527                         }
 528                 }
 529
 530                 if (mm->base && !rec->opts.auxtrace_snapshot_mode &&
 531                     record__auxtrace_mmap_read(rec, mm) != 0) {
 532                         rc = -1;
 533                         goto out;
 534                 }
 535         }
 536
 537         /*
 538          * Mark the round finished in case we wrote
 539          * at least one event.
 540          */
 541         if (bytes_written != rec->bytes_written)
 542                 rc = record__write(rec, &finished_round_event, sizeof(finished_round_event));
 543
 544         if (backward)
 545                 perf_evlist__toggle_bkw_mmap(evlist, BKW_MMAP_EMPTY);
 546 out:
 547         return rc;
 548 }
 549
 550 static int record__mmap_read_all(struct record *rec)
 551 {
 552         int err;
 553
 554         err = record__mmap_read_evlist(rec, rec->evlist, false);
 555         if (err)
 556                 return err;
 557
 558         return record__mmap_read_evlist(rec, rec->evlist, true);
 559 }
 560
 561 static void record__init_features(struct record *rec)
 562 {
 563         struct perf_session *session = rec->session;
 564         int feat;
 565
 566         for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++)
 567                 perf_header__set_feat(&session->header, feat);
 568
 569         if (rec->no_buildid)
 570                 perf_header__clear_feat(&session->header, HEADER_BUILD_ID);
 571
 572         if (!have_tracepoints(&rec->evlist->entries))
 573                 perf_header__clear_feat(&session->header, HEADER_TRACING_DATA);
 574
 575         if (!rec->opts.branch_stack)
 576                 perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK);
 577
 578         if (!rec->opts.full_auxtrace)
 579                 perf_header__clear_feat(&session->header, HEADER_AUXTRACE);
 580
 581         perf_header__clear_feat(&session->header, HEADER_STAT);
 582 }
 583
 584 static void
 585 record__finish_output(struct record *rec)
 586 {
 587         struct perf_data_file *file = &rec->file;
 588         int fd = perf_data_file__fd(file);
 589
 590         if (file->is_pipe)
 591                 return;
 592
 593         rec->session->header.data_size += rec->bytes_written;
 594         file->size = lseek(perf_data_file__fd(file), 0, SEEK_CUR);
 595
 596         if (!rec->no_buildid) {
 597                 process_buildids(rec);
 598
 599                 if (rec->buildid_all)
 600                         dsos__hit_all(rec->session);
 601         }
 602         perf_session__write_header(rec->session, rec->evlist, fd, true);
 603
 604         return;
 605 }
 606
 607 static int record__synthesize_workload(struct record *rec, bool tail)
 608 {
 609         struct {
 610                 struct thread_map map;
 611                 struct thread_map_data map_data;
 612         } thread_map;
 613
 614         if (rec->opts.tail_synthesize != tail)
 615                 return 0;
 616
 617         thread_map.map.nr = 1;
 618         thread_map.map.map[0].pid = rec->evlist->workload.pid;
 619         thread_map.map.map[0].comm = NULL;
 620         return perf_event__synthesize_thread_map(&rec->tool, &thread_map.map,
 621                                                  process_synthesized_event,
 622                                                  &rec->session->machines.host,
 623                                                  rec->opts.sample_address,
 624                                                  rec->opts.proc_map_timeout);
 625 }
 626
 627 static int record__synthesize(struct record *rec, bool tail);
 628
 629 static int
 630 record__switch_output(struct record *rec, bool at_exit)
 631 {
 632         struct perf_data_file *file = &rec->file;
 633         int fd, err;
 634
 635         /* Same Size:      "2015122520103046"*/
 636         char timestamp[] = "InvalidTimestamp";
 637
 638         record__synthesize(rec, true);
 639         if (target__none(&rec->opts.target))
 640                 record__synthesize_workload(rec, true);
 641
 642         rec->samples = 0;
 643         record__finish_output(rec);
 644         err = fetch_current_timestamp(timestamp, sizeof(timestamp));
 645         if (err) {
 646                 pr_err("Failed to get current timestamp\n");
 647                 return -EINVAL;
 648         }
 649
 650         fd = perf_data_file__switch(file, timestamp,
 651                                     rec->session->header.data_offset,
 652                                     at_exit);
 653         if (fd >= 0 && !at_exit) {
 654                 rec->bytes_written = 0;
 655                 rec->session->header.data_size = 0;
 656         }
 657
 658         if (!quiet)
 659                 fprintf(stderr, "[ perf record: Dump %s.%s ]\n",
 660                         file->path, timestamp);
 661
 662         /* Output tracking events */
 663         if (!at_exit) {
 664                 record__synthesize(rec, false);
 665
 666                 /*
 667                  * In 'perf record --switch-output' without -a,
 668                  * record__synthesize() in record__switch_output() won't
 669                  * generate tracking events because there's no thread_map
 670                  * in evlist. Which causes newly created perf.data doesn't
 671                  * contain map and comm information.
 672                  * Create a fake thread_map and directly call
 673                  * perf_event__synthesize_thread_map() for those events.
 674                  */
 675                 if (target__none(&rec->opts.target))
 676                         record__synthesize_workload(rec, false);
 677         }
 678         return fd;
 679 }
 680
 681 static volatile int workload_exec_errno;
 682
 683 /*
 684  * perf_evlist__prepare_workload will send a SIGUSR1
 685  * if the fork fails, since we asked by setting its
 686  * want_signal to true.
 687  */
 688 static void workload_exec_failed_signal(int signo __maybe_unused,
 689                                         siginfo_t *info,
 690                                         void *ucontext __maybe_unused)
 691 {
 692         workload_exec_errno = info->si_value.sival_int;
 693         done = 1;
 694         child_finished = 1;
 695 }
 696
 697 static void snapshot_sig_handler(int sig);
 698
 699 int __weak
 700 perf_event__synth_time_conv(const struct perf_event_mmap_page *pc __maybe_unused,
 701                             struct perf_tool *tool __maybe_unused,
 702                             perf_event__handler_t process __maybe_unused,
 703                             struct machine *machine __maybe_unused)
 704 {
 705         return 0;
 706 }
 707
 708 static const struct perf_event_mmap_page *
 709 perf_evlist__pick_pc(struct perf_evlist *evlist)
 710 {
 711         if (evlist) {
 712                 if (evlist->mmap && evlist->mmap[0].base)
 713                         return evlist->mmap[0].base;
 714                 if (evlist->backward_mmap && evlist->backward_mmap[0].base)
 715                         return evlist->backward_mmap[0].base;
 716         }
 717         return NULL;
 718 }
 719
 720 static const struct perf_event_mmap_page *record__pick_pc(struct record *rec)
 721 {
 722         const struct perf_event_mmap_page *pc;
 723
 724         pc = perf_evlist__pick_pc(rec->evlist);
 725         if (pc)
 726                 return pc;
 727         return NULL;
 728 }
 729
 730 static int record__synthesize(struct record *rec, bool tail)
 731 {
 732         struct perf_session *session = rec->session;
 733         struct machine *machine = &session->machines.host;
 734         struct perf_data_file *file = &rec->file;
 735         struct record_opts *opts = &rec->opts;
 736         struct perf_tool *tool = &rec->tool;
 737         int fd = perf_data_file__fd(file);
 738         int err = 0;
 739
 740         if (rec->opts.tail_synthesize != tail)
 741                 return 0;
 742
 743         if (file->is_pipe) {
 744                 err = perf_event__synthesize_attrs(tool, session,
 745                                                    process_synthesized_event);
 746                 if (err < 0) {
 747                         pr_err("Couldn't synthesize attrs.\n");
 748                         goto out;
 749                 }
 750
 751                 if (have_tracepoints(&rec->evlist->entries)) {
 752                         /*
 753                          * FIXME err <= 0 here actually means that
 754                          * there were no tracepoints so its not really
 755                          * an error, just that we don't need to
 756                          * synthesize anything.  We really have to
 757                          * return this more properly and also
 758                          * propagate errors that now are calling die()
 759                          */
 760                         err = perf_event__synthesize_tracing_data(tool, fd, rec->evlist,
 761                                                                   process_synthesized_event);
 762                         if (err <= 0) {
 763                                 pr_err("Couldn't record tracing data.\n");
 764                                 goto out;
 765                         }
 766                         rec->bytes_written += err;
 767                 }
 768         }
 769
 770         err = perf_event__synth_time_conv(record__pick_pc(rec), tool,
 771                                           process_synthesized_event, machine);
 772         if (err)
 773                 goto out;
 774
 775         if (rec->opts.full_auxtrace) {
 776                 err = perf_event__synthesize_auxtrace_info(rec->itr, tool,
 777                                         session, process_synthesized_event);
 778                 if (err)
 779                         goto out;
 780         }
 781
 782         err = perf_event__synthesize_kernel_mmap(tool, process_synthesized_event,
 783                                                  machine);
 784         WARN_ONCE(err < 0, "Couldn't record kernel reference relocation symbol\n"
 785                            "Symbol resolution may be skewed if relocation was used (e.g. kexec).\n"
 786                            "Check /proc/kallsyms permission or run as root.\n");
 787
 788         err = perf_event__synthesize_modules(tool, process_synthesized_event,
 789                                              machine);
 790         WARN_ONCE(err < 0, "Couldn't record kernel module information.\n"
 791                            "Symbol resolution may be skewed if relocation was used (e.g. kexec).\n"
 792                            "Check /proc/modules permission or run as root.\n");
 793
 794         if (perf_guest) {
 795                 machines__process_guests(&session->machines,
 796                                          perf_event__synthesize_guest_os, tool);
 797         }
 798
 799         err = __machine__synthesize_threads(machine, tool, &opts->target, rec->evlist->threads,
 800                                             process_synthesized_event, opts->sample_address,
 801                                             opts->proc_map_timeout);
 802 out:
 803         return err;
 804 }
 805
 806 static int __cmd_record(struct record *rec, int argc, const char **argv)
 807 {
 808         int err;
 809         int status = 0;
 810         unsigned long waking = 0;
 811         const bool forks = argc > 0;
 812         struct machine *machine;
 813         struct perf_tool *tool = &rec->tool;
 814         struct record_opts *opts = &rec->opts;
 815         struct perf_data_file *file = &rec->file;
 816         struct perf_session *session;
 817         bool disabled = false, draining = false;
 818         int fd;
 819
 820         rec->progname = argv[0];
 821
 822         atexit(record__sig_exit);
 823         signal(SIGCHLD, sig_handler);
 824         signal(SIGINT, sig_handler);
 825         signal(SIGTERM, sig_handler);
 826
 827         if (rec->opts.auxtrace_snapshot_mode || rec->switch_output) {
 828                 signal(SIGUSR2, snapshot_sig_handler);
 829                 if (rec->opts.auxtrace_snapshot_mode)
 830                         trigger_on(&auxtrace_snapshot_trigger);
 831                 if (rec->switch_output)
 832                         trigger_on(&switch_output_trigger);
 833         } else {
 834                 signal(SIGUSR2, SIG_IGN);
 835         }
 836
 837         session = perf_session__new(file, false, tool);
 838         if (session == NULL) {
 839                 pr_err("Perf session creation failed.\n");
 840                 return -1;
 841         }
 842
 843         fd = perf_data_file__fd(file);
 844         rec->session = session;
 845
 846         record__init_features(rec);
 847
 848         if (forks) {
 849                 err = perf_evlist__prepare_workload(rec->evlist, &opts->target,
 850                                                     argv, file->is_pipe,
 851                                                     workload_exec_failed_signal);
 852                 if (err < 0) {
 853                         pr_err("Couldn't run the workload!\n");
 854                         status = err;
 855                         goto out_delete_session;
 856                 }
 857         }
 858
 859         if (record__open(rec) != 0) {
 860                 err = -1;
 861                 goto out_child;
 862         }
 863
 864         err = bpf__apply_obj_config();
 865         if (err) {
 866                 char errbuf[BUFSIZ];
 867
 868                 bpf__strerror_apply_obj_config(err, errbuf, sizeof(errbuf));
 869                 pr_err("ERROR: Apply config to BPF failed: %s\n",
 870                          errbuf);
 871                 goto out_child;
 872         }
 873
 874         /*
 875          * Normally perf_session__new would do this, but it doesn't have the
 876          * evlist.
 877          */
 878         if (rec->tool.ordered_events && !perf_evlist__sample_id_all(rec->evlist)) {
 879                 pr_warning("WARNING: No sample_id_all support, falling back to unordered processing\n");
 880                 rec->tool.ordered_events = false;
 881         }
 882
 883         if (!rec->evlist->nr_groups)
 884                 perf_header__clear_feat(&session->header, HEADER_GROUP_DESC);
 885
 886         if (file->is_pipe) {
 887                 err = perf_header__write_pipe(fd);
 888                 if (err < 0)
 889                         goto out_child;
 890         } else {
 891                 err = perf_session__write_header(session, rec->evlist, fd, false);
 892                 if (err < 0)
 893                         goto out_child;
 894         }
 895
 896         if (!rec->no_buildid
 897             && !perf_header__has_feat(&session->header, HEADER_BUILD_ID)) {
 898                 pr_err("Couldn't generate buildids. "
 899                        "Use --no-buildid to profile anyway.\n");
 900                 err = -1;
 901                 goto out_child;
 902         }
 903
 904         machine = &session->machines.host;
 905
 906         err = record__synthesize(rec, false);
 907         if (err < 0)
 908                 goto out_child;
 909
 910         if (rec->realtime_prio) {
 911                 struct sched_param param;
 912
 913                 param.sched_priority = rec->realtime_prio;
 914                 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
 915                         pr_err("Could not set realtime priority.\n");
 916                         err = -1;
 917                         goto out_child;
 918                 }
 919         }
 920
 921         /*
 922          * When perf is starting the traced process, all the events
 923          * (apart from group members) have enable_on_exec=1 set,
 924          * so don't spoil it by prematurely enabling them.
 925          */
 926         if (!target__none(&opts->target) && !opts->initial_delay)
 927                 perf_evlist__enable(rec->evlist);
 928
 929         /*
 930          * Let the child rip
 931          */
 932         if (forks) {
 933                 union perf_event *event;
 934
 935                 event = malloc(sizeof(event->comm) + machine->id_hdr_size);
 936                 if (event == NULL) {
 937                         err = -ENOMEM;
 938                         goto out_child;
 939                 }
 940
 941                 /*
 942                  * Some H/W events are generated before COMM event
 943                  * which is emitted during exec(), so perf script
 944                  * cannot see a correct process name for those events.
 945                  * Synthesize COMM event to prevent it.
 946                  */
 947                 perf_event__synthesize_comm(tool, event,
 948                                             rec->evlist->workload.pid,
 949                                             process_synthesized_event,
 950                                             machine);
 951                 free(event);
 952
 953                 perf_evlist__start_workload(rec->evlist);
 954         }
 955
 956         if (opts->initial_delay) {
 957                 usleep(opts->initial_delay * 1000);
 958                 perf_evlist__enable(rec->evlist);
 959         }
 960
 961         trigger_ready(&auxtrace_snapshot_trigger);
 962         trigger_ready(&switch_output_trigger);
 963         for (;;) {
 964                 unsigned long long hits = rec->samples;
 965
 966                 /*
 967                  * rec->evlist->bkw_mmap_state is possible to be
 968                  * BKW_MMAP_EMPTY here: when done == true and
 969                  * hits != rec->samples in previous round.
 970                  *
 971                  * perf_evlist__toggle_bkw_mmap ensure we never
 972                  * convert BKW_MMAP_EMPTY to BKW_MMAP_DATA_PENDING.
 973                  */
 974                 if (trigger_is_hit(&switch_output_trigger) || done || draining)
 975                         perf_evlist__toggle_bkw_mmap(rec->evlist, BKW_MMAP_DATA_PENDING);
 976
 977                 if (record__mmap_read_all(rec) < 0) {
 978                         trigger_error(&auxtrace_snapshot_trigger);
 979                         trigger_error(&switch_output_trigger);
 980                         err = -1;
 981                         goto out_child;
 982                 }
 983
 984                 if (auxtrace_record__snapshot_started) {
 985                         auxtrace_record__snapshot_started = 0;
 986                         if (!trigger_is_error(&auxtrace_snapshot_trigger))
 987                                 record__read_auxtrace_snapshot(rec);
 988                         if (trigger_is_error(&auxtrace_snapshot_trigger)) {
 989                                 pr_err("AUX area tracing snapshot failed\n");
 990                                 err = -1;
 991                                 goto out_child;
 992                         }
 993                 }
 994
 995                 if (trigger_is_hit(&switch_output_trigger)) {
 996                         /*
 997                          * If switch_output_trigger is hit, the data in
 998                          * overwritable ring buffer should have been collected,
 999                          * so bkw_mmap_state should be set to BKW_MMAP_EMPTY.
1000                          *
1001                          * If SIGUSR2 raise after or during record__mmap_read_all(),
1002                          * record__mmap_read_all() didn't collect data from
1003                          * overwritable ring buffer. Read again.
1004                          */
1005                         if (rec->evlist->bkw_mmap_state == BKW_MMAP_RUNNING)
1006                                 continue;
1007                         trigger_ready(&switch_output_trigger);
1008
1009                         /*
1010                          * Reenable events in overwrite ring buffer after
1011                          * record__mmap_read_all(): we should have collected
1012                          * data from it.
1013                          */
1014                         perf_evlist__toggle_bkw_mmap(rec->evlist, BKW_MMAP_RUNNING);
1015
1016                         if (!quiet)
1017                                 fprintf(stderr, "[ perf record: dump data: Woken up %ld times ]\n",
1018                                         waking);
1019                         waking = 0;
1020                         fd = record__switch_output(rec, false);
1021                         if (fd < 0) {
1022                                 pr_err("Failed to switch to new file\n");
1023                                 trigger_error(&switch_output_trigger);
1024                                 err = fd;
1025                                 goto out_child;
1026                         }
1027                 }
1028
1029                 if (hits == rec->samples) {
1030                         if (done || draining)
1031                                 break;
1032                         err = perf_evlist__poll(rec->evlist, -1);
1033                         /*
1034                          * Propagate error, only if there's any. Ignore positive
1035                          * number of returned events and interrupt error.
1036                          */
1037                         if (err > 0 || (err < 0 && errno == EINTR))
1038                                 err = 0;
1039                         waking++;
1040
1041                         if (perf_evlist__filter_pollfd(rec->evlist, POLLERR | POLLHUP) == 0)
1042                                 draining = true;
1043                 }
1044
1045                 /*
1046                  * When perf is starting the traced process, at the end events
1047                  * die with the process and we wait for that. Thus no need to
1048                  * disable events in this case.
1049                  */
1050                 if (done && !disabled && !target__none(&opts->target)) {
1051                         trigger_off(&auxtrace_snapshot_trigger);
1052                         perf_evlist__disable(rec->evlist);
1053                         disabled = true;
1054                 }
1055         }
1056         trigger_off(&auxtrace_snapshot_trigger);
1057         trigger_off(&switch_output_trigger);
1058
1059         if (forks && workload_exec_errno) {
1060                 char msg[STRERR_BUFSIZE];
1061                 const char *emsg = str_error_r(workload_exec_errno, msg, sizeof(msg));
1062                 pr_err("Workload failed: %s\n", emsg);
1063                 err = -1;
1064                 goto out_child;
1065         }
1066
1067         if (!quiet)
1068                 fprintf(stderr, "[ perf record: Woken up %ld times to write data ]\n", waking);
1069
1070         if (target__none(&rec->opts.target))
1071                 record__synthesize_workload(rec, true);
1072
1073 out_child:
1074         if (forks) {
1075                 int exit_status;
1076
1077                 if (!child_finished)
1078                         kill(rec->evlist->workload.pid, SIGTERM);
1079
1080                 wait(&exit_status);
1081
1082                 if (err < 0)
1083                         status = err;
1084                 else if (WIFEXITED(exit_status))
1085                         status = WEXITSTATUS(exit_status);
1086                 else if (WIFSIGNALED(exit_status))
1087                         signr = WTERMSIG(exit_status);
1088         } else
1089                 status = err;
1090
1091         record__synthesize(rec, true);
1092         /* this will be recalculated during process_buildids() */
1093         rec->samples = 0;
1094
1095         if (!err) {
1096                 if (!rec->timestamp_filename) {
1097                         record__finish_output(rec);
1098                 } else {
1099                         fd = record__switch_output(rec, true);
1100                         if (fd < 0) {
1101                                 status = fd;
1102                                 goto out_delete_session;
1103                         }
1104                 }
1105         }
1106
1107         if (!err && !quiet) {
1108                 char samples[128];
1109                 const char *postfix = rec->timestamp_filename ?
1110                                         ".<timestamp>" : "";
1111
1112                 if (rec->samples && !rec->opts.full_auxtrace)
1113                         scnprintf(samples, sizeof(samples),
1114                                   " (%" PRIu64 " samples)", rec->samples);
1115                 else
1116                         samples[0] = '\0';
1117
1118                 fprintf(stderr, "[ perf record: Captured and wrote %.3f MB %s%s%s ]\n",
1119                         perf_data_file__size(file) / 1024.0 / 1024.0,
1120                         file->path, postfix, samples);
1121         }
1122
1123 out_delete_session:
1124         perf_session__delete(session);
1125         return status;
1126 }
1127
1128 static void callchain_debug(struct callchain_param *callchain)
1129 {
1130         static const char *str[CALLCHAIN_MAX] = { "NONE", "FP", "DWARF", "LBR" };
1131
1132         pr_debug("callchain: type %s\n", str[callchain->record_mode]);
1133
1134         if (callchain->record_mode == CALLCHAIN_DWARF)
1135                 pr_debug("callchain: stack dump size %d\n",
1136                          callchain->dump_size);
1137 }
1138
1139 int record_opts__parse_callchain(struct record_opts *record,
1140                                  struct callchain_param *callchain,
1141                                  const char *arg, bool unset)
1142 {
1143         int ret;
1144         callchain->enabled = !unset;
1145
1146         /* --no-call-graph */
1147         if (unset) {
1148                 callchain->record_mode = CALLCHAIN_NONE;
1149                 pr_debug("callchain: disabled\n");
1150                 return 0;
1151         }
1152
1153         ret = parse_callchain_record_opt(arg, callchain);
1154         if (!ret) {
1155                 /* Enable data address sampling for DWARF unwind. */
1156                 if (callchain->record_mode == CALLCHAIN_DWARF)
1157                         record->sample_address = true;
1158                 callchain_debug(callchain);
1159         }
1160
1161         return ret;
1162 }
1163
1164 int record_parse_callchain_opt(const struct option *opt,
1165                                const char *arg,
1166                                int unset)
1167 {
1168         return record_opts__parse_callchain(opt->value, &callchain_param, arg, unset);
1169 }
1170
1171 int record_callchain_opt(const struct option *opt,
1172                          const char *arg __maybe_unused,
1173                          int unset __maybe_unused)
1174 {
1175         struct callchain_param *callchain = opt->value;
1176
1177         callchain->enabled = true;
1178
1179         if (callchain->record_mode == CALLCHAIN_NONE)
1180                 callchain->record_mode = CALLCHAIN_FP;
1181
1182         callchain_debug(callchain);
1183         return 0;
1184 }
1185
1186 static int perf_record_config(const char *var, const char *value, void *cb)
1187 {
1188         struct record *rec = cb;
1189
1190         if (!strcmp(var, "record.build-id")) {
1191                 if (!strcmp(value, "cache"))
1192                         rec->no_buildid_cache = false;
1193                 else if (!strcmp(value, "no-cache"))
1194                         rec->no_buildid_cache = true;
1195                 else if (!strcmp(value, "skip"))
1196                         rec->no_buildid = true;
1197                 else
1198                         return -1;
1199                 return 0;
1200         }
1201         if (!strcmp(var, "record.call-graph"))
1202                 var = "call-graph.record-mode"; /* fall-through */
1203
1204         return perf_default_config(var, value, cb);
1205 }
1206
1207 struct clockid_map {
1208         const char *name;
1209         int clockid;
1210 };
1211
1212 #define CLOCKID_MAP(n, c)       \
1213         { .name = n, .clockid = (c), }
1214
1215 #define CLOCKID_END     { .name = NULL, }
1216
1217
1218 /*
1219  * Add the missing ones, we need to build on many distros...
1220  */
1221 #ifndef CLOCK_MONOTONIC_RAW
1222 #define CLOCK_MONOTONIC_RAW 4
1223 #endif
1224 #ifndef CLOCK_BOOTTIME
1225 #define CLOCK_BOOTTIME 7
1226 #endif
1227 #ifndef CLOCK_TAI
1228 #define CLOCK_TAI 11
1229 #endif
1230
1231 static const struct clockid_map clockids[] = {
1232         /* available for all events, NMI safe */
1233         CLOCKID_MAP("monotonic", CLOCK_MONOTONIC),
1234         CLOCKID_MAP("monotonic_raw", CLOCK_MONOTONIC_RAW),
1235
1236         /* available for some events */
1237         CLOCKID_MAP("realtime", CLOCK_REALTIME),
1238         CLOCKID_MAP("boottime", CLOCK_BOOTTIME),
1239         CLOCKID_MAP("tai", CLOCK_TAI),
1240
1241         /* available for the lazy */
1242         CLOCKID_MAP("mono", CLOCK_MONOTONIC),
1243         CLOCKID_MAP("raw", CLOCK_MONOTONIC_RAW),
1244         CLOCKID_MAP("real", CLOCK_REALTIME),
1245         CLOCKID_MAP("boot", CLOCK_BOOTTIME),
1246
1247         CLOCKID_END,
1248 };
1249
1250 static int parse_clockid(const struct option *opt, const char *str, int unset)
1251 {
1252         struct record_opts *opts = (struct record_opts *)opt->value;
1253         const struct clockid_map *cm;
1254         const char *ostr = str;
1255
1256         if (unset) {
1257                 opts->use_clockid = 0;
1258                 return 0;
1259         }
1260
1261         /* no arg passed */
1262         if (!str)
1263                 return 0;
1264
1265         /* no setting it twice */
1266         if (opts->use_clockid)
1267                 return -1;
1268
1269         opts->use_clockid = true;
1270
1271         /* if its a number, we're done */
1272         if (sscanf(str, "%d", &opts->clockid) == 1)
1273                 return 0;
1274
1275         /* allow a "CLOCK_" prefix to the name */
1276         if (!strncasecmp(str, "CLOCK_", 6))
1277                 str += 6;
1278
1279         for (cm = clockids; cm->name; cm++) {
1280                 if (!strcasecmp(str, cm->name)) {
1281                         opts->clockid = cm->clockid;
1282                         return 0;
1283                 }
1284         }
1285
1286         opts->use_clockid = false;
1287         ui__warning("unknown clockid %s, check man page\n", ostr);
1288         return -1;
1289 }
1290
1291 static int record__parse_mmap_pages(const struct option *opt,
1292                                     const char *str,
1293                                     int unset __maybe_unused)
1294 {
1295         struct record_opts *opts = opt->value;
1296         char *s, *p;
1297         unsigned int mmap_pages;
1298         int ret;
1299
1300         if (!str)
1301                 return -EINVAL;
1302
1303         s = strdup(str);
1304         if (!s)
1305                 return -ENOMEM;
1306
1307         p = strchr(s, ',');
1308         if (p)
1309                 *p = '\0';
1310
1311         if (*s) {
1312                 ret = __perf_evlist__parse_mmap_pages(&mmap_pages, s);
1313                 if (ret)
1314                         goto out_free;
1315                 opts->mmap_pages = mmap_pages;
1316         }
1317
1318         if (!p) {
1319                 ret = 0;
1320                 goto out_free;
1321         }
1322
1323         ret = __perf_evlist__parse_mmap_pages(&mmap_pages, p + 1);
1324         if (ret)
1325                 goto out_free;
1326
1327         opts->auxtrace_mmap_pages = mmap_pages;
1328
1329 out_free:
1330         free(s);
1331         return ret;
1332 }
1333
1334 static const char * const __record_usage[] = {
1335         "perf record [<options>] [<command>]",
1336         "perf record [<options>] -- <command> [<options>]",
1337         NULL
1338 };
1339 const char * const *record_usage = __record_usage;
1340
1341 /*
1342  * XXX Ideally would be local to cmd_record() and passed to a record__new
1343  * because we need to have access to it in record__exit, that is called
1344  * after cmd_record() exits, but since record_options need to be accessible to
1345  * builtin-script, leave it here.
1346  *
1347  * At least we don't ouch it in all the other functions here directly.
1348  *
1349  * Just say no to tons of global variables, sigh.
1350  */
1351 static struct record record = {
1352         .opts = {
1353                 .sample_time         = true,
1354                 .mmap_pages          = UINT_MAX,
1355                 .user_freq           = UINT_MAX,
1356                 .user_interval       = ULLONG_MAX,
1357                 .freq                = 4000,
1358                 .target              = {
1359                         .uses_mmap   = true,
1360                         .default_per_cpu = true,
1361                 },
1362                 .proc_map_timeout     = 500,
1363         },
1364         .tool = {
1365                 .sample         = process_sample_event,
1366                 .fork           = perf_event__process_fork,
1367                 .exit           = perf_event__process_exit,
1368                 .comm           = perf_event__process_comm,
1369                 .mmap           = perf_event__process_mmap,
1370                 .mmap2          = perf_event__process_mmap2,
1371                 .ordered_events = true,
1372         },
1373 };
1374
1375 const char record_callchain_help[] = CALLCHAIN_RECORD_HELP
1376         "\n\t\t\t\tDefault: fp";
1377
1378 static bool dry_run;
1379
1380 /*
1381  * XXX Will stay a global variable till we fix builtin-script.c to stop messing
1382  * with it and switch to use the library functions in perf_evlist that came
1383  * from builtin-record.c, i.e. use record_opts,
1384  * perf_evlist__prepare_workload, etc instead of fork+exec'in 'perf record',
1385  * using pipes, etc.
1386  */
1387 struct option __record_options[] = {
1388         OPT_CALLBACK('e', "event", &record.evlist, "event",
1389                      "event selector. use 'perf list' to list available events",
1390                      parse_events_option),
1391         OPT_CALLBACK(0, "filter", &record.evlist, "filter",
1392                      "event filter", parse_filter),
1393         OPT_CALLBACK_NOOPT(0, "exclude-perf", &record.evlist,
1394                            NULL, "don't record events from perf itself",
1395                            exclude_perf),
1396         OPT_STRING('p', "pid", &record.opts.target.pid, "pid",
1397                     "record events on existing process id"),
1398         OPT_STRING('t', "tid", &record.opts.target.tid, "tid",
1399                     "record events on existing thread id"),
1400         OPT_INTEGER('r', "realtime", &record.realtime_prio,
1401                     "collect data with this RT SCHED_FIFO priority"),
1402         OPT_BOOLEAN(0, "no-buffering", &record.opts.no_buffering,
1403                     "collect data without buffering"),
1404         OPT_BOOLEAN('R', "raw-samples", &record.opts.raw_samples,
1405                     "collect raw sample records from all opened counters"),
1406         OPT_BOOLEAN('a', "all-cpus", &record.opts.target.system_wide,
1407                             "system-wide collection from all CPUs"),
1408         OPT_STRING('C', "cpu", &record.opts.target.cpu_list, "cpu",
1409                     "list of cpus to monitor"),
1410         OPT_U64('c', "count", &record.opts.user_interval, "event period to sample"),
1411         OPT_STRING('o', "output", &record.file.path, "file",
1412                     "output file name"),
1413         OPT_BOOLEAN_SET('i', "no-inherit", &record.opts.no_inherit,
1414                         &record.opts.no_inherit_set,
1415                         "child tasks do not inherit counters"),
1416         OPT_BOOLEAN(0, "tail-synthesize", &record.opts.tail_synthesize,
1417                     "synthesize non-sample events at the end of output"),
1418         OPT_BOOLEAN(0, "overwrite", &record.opts.overwrite, "use overwrite mode"),
1419         OPT_UINTEGER('F', "freq", &record.opts.user_freq, "profile at this frequency"),
1420         OPT_CALLBACK('m', "mmap-pages", &record.opts, "pages[,pages]",
1421                      "number of mmap data pages and AUX area tracing mmap pages",
1422                      record__parse_mmap_pages),
1423         OPT_BOOLEAN(0, "group", &record.opts.group,
1424                     "put the counters into a counter group"),
1425         OPT_CALLBACK_NOOPT('g', NULL, &callchain_param,
1426                            NULL, "enables call-graph recording" ,
1427                            &record_callchain_opt),
1428         OPT_CALLBACK(0, "call-graph", &record.opts,
1429                      "record_mode[,record_size]", record_callchain_help,
1430                      &record_parse_callchain_opt),
1431         OPT_INCR('v', "verbose", &verbose,
1432                     "be more verbose (show counter open errors, etc)"),
1433         OPT_BOOLEAN('q', "quiet", &quiet, "don't print any message"),
1434         OPT_BOOLEAN('s', "stat", &record.opts.inherit_stat,
1435                     "per thread counts"),
1436         OPT_BOOLEAN('d', "data", &record.opts.sample_address, "Record the sample addresses"),
1437         OPT_BOOLEAN(0, "sample-cpu", &record.opts.sample_cpu, "Record the sample cpu"),
1438         OPT_BOOLEAN_SET('T', "timestamp", &record.opts.sample_time,
1439                         &record.opts.sample_time_set,
1440                         "Record the sample timestamps"),
1441         OPT_BOOLEAN('P', "period", &record.opts.period, "Record the sample period"),
1442         OPT_BOOLEAN('n', "no-samples", &record.opts.no_samples,
1443                     "don't sample"),
1444         OPT_BOOLEAN_SET('N', "no-buildid-cache", &record.no_buildid_cache,
1445                         &record.no_buildid_cache_set,
1446                         "do not update the buildid cache"),
1447         OPT_BOOLEAN_SET('B', "no-buildid", &record.no_buildid,
1448                         &record.no_buildid_set,
1449                         "do not collect buildids in perf.data"),
1450         OPT_CALLBACK('G', "cgroup", &record.evlist, "name",
1451                      "monitor event in cgroup name only",
1452                      parse_cgroups),
1453         OPT_UINTEGER('D', "delay", &record.opts.initial_delay,
1454                   "ms to wait before starting measurement after program start"),
1455         OPT_STRING('u', "uid", &record.opts.target.uid_str, "user",
1456                    "user to profile"),
1457
1458         OPT_CALLBACK_NOOPT('b', "branch-any", &record.opts.branch_stack,
1459                      "branch any", "sample any taken branches",
1460                      parse_branch_stack),
1461
1462         OPT_CALLBACK('j', "branch-filter", &record.opts.branch_stack,
1463                      "branch filter mask", "branch stack filter modes",
1464                      parse_branch_stack),
1465         OPT_BOOLEAN('W', "weight", &record.opts.sample_weight,
1466                     "sample by weight (on special events only)"),
1467         OPT_BOOLEAN(0, "transaction", &record.opts.sample_transaction,
1468                     "sample transaction flags (special events only)"),
1469         OPT_BOOLEAN(0, "per-thread", &record.opts.target.per_thread,
1470                     "use per-thread mmaps"),
1471         OPT_CALLBACK_OPTARG('I', "intr-regs", &record.opts.sample_intr_regs, NULL, "any register",
1472                     "sample selected machine registers on interrupt,"
1473                     " use -I ? to list register names", parse_regs),
1474         OPT_BOOLEAN(0, "running-time", &record.opts.running_time,
1475                     "Record running/enabled time of read (:S) events"),
1476         OPT_CALLBACK('k', "clockid", &record.opts,
1477         "clockid", "clockid to use for events, see clock_gettime()",
1478         parse_clockid),
1479         OPT_STRING_OPTARG('S', "snapshot", &record.opts.auxtrace_snapshot_opts,
1480                           "opts", "AUX area tracing Snapshot Mode", ""),
1481         OPT_UINTEGER(0, "proc-map-timeout", &record.opts.proc_map_timeout,
1482                         "per thread proc mmap processing timeout in ms"),
1483         OPT_BOOLEAN(0, "switch-events", &record.opts.record_switch_events,
1484                     "Record context switch events"),
1485         OPT_BOOLEAN_FLAG(0, "all-kernel", &record.opts.all_kernel,
1486                          "Configure all used events to run in kernel space.",
1487                          PARSE_OPT_EXCLUSIVE),
1488         OPT_BOOLEAN_FLAG(0, "all-user", &record.opts.all_user,
1489                          "Configure all used events to run in user space.",
1490                          PARSE_OPT_EXCLUSIVE),
1491         OPT_STRING(0, "clang-path", &llvm_param.clang_path, "clang path",
1492                    "clang binary to use for compiling BPF scriptlets"),
1493         OPT_STRING(0, "clang-opt", &llvm_param.clang_opt, "clang options",
1494                    "options passed to clang when compiling BPF scriptlets"),
1495         OPT_STRING(0, "vmlinux", &symbol_conf.vmlinux_name,
1496                    "file", "vmlinux pathname"),
1497         OPT_BOOLEAN(0, "buildid-all", &record.buildid_all,
1498                     "Record build-id of all DSOs regardless of hits"),
1499         OPT_BOOLEAN(0, "timestamp-filename", &record.timestamp_filename,
1500                     "append timestamp to output filename"),
1501         OPT_BOOLEAN(0, "switch-output", &record.switch_output,
1502                     "Switch output when receive SIGUSR2"),
1503         OPT_BOOLEAN(0, "dry-run", &dry_run,
1504                     "Parse options then exit"),
1505         OPT_END()
1506 };
1507
1508 struct option *record_options = __record_options;
1509
1510 int cmd_record(int argc, const char **argv, const char *prefix __maybe_unused)
1511 {
1512         int err;
1513         struct record *rec = &record;
1514         char errbuf[BUFSIZ];
1515
1516 #ifndef HAVE_LIBBPF_SUPPORT
1517 # define set_nobuild(s, l, c) set_option_nobuild(record_options, s, l, "NO_LIBBPF=1", c)
1518         set_nobuild('\0', "clang-path", true);
1519         set_nobuild('\0', "clang-opt", true);
1520 # undef set_nobuild
1521 #endif
1522
1523 #ifndef HAVE_BPF_PROLOGUE
1524 # if !defined (HAVE_DWARF_SUPPORT)
1525 #  define REASON  "NO_DWARF=1"
1526 # elif !defined (HAVE_LIBBPF_SUPPORT)
1527 #  define REASON  "NO_LIBBPF=1"
1528 # else
1529 #  define REASON  "this architecture doesn't support BPF prologue"
1530 # endif
1531 # define set_nobuild(s, l, c) set_option_nobuild(record_options, s, l, REASON, c)
1532         set_nobuild('\0', "vmlinux", true);
1533 # undef set_nobuild
1534 # undef REASON
1535 #endif
1536
1537         rec->evlist = perf_evlist__new();
1538         if (rec->evlist == NULL)
1539                 return -ENOMEM;
1540
1541         perf_config(perf_record_config, rec);
1542
1543         argc = parse_options(argc, argv, record_options, record_usage,
1544                             PARSE_OPT_STOP_AT_NON_OPTION);
1545         if (!argc && target__none(&rec->opts.target))
1546                 usage_with_options(record_usage, record_options);
1547
1548         if (nr_cgroups && !rec->opts.target.system_wide) {
1549                 usage_with_options_msg(record_usage, record_options,
1550                         "cgroup monitoring only available in system-wide mode");
1551
1552         }
1553         if (rec->opts.record_switch_events &&
1554             !perf_can_record_switch_events()) {
1555                 ui__error("kernel does not support recording context switch events\n");
1556                 parse_options_usage(record_usage, record_options, "switch-events", 0);
1557                 return -EINVAL;
1558         }
1559
1560         if (rec->switch_output)
1561                 rec->timestamp_filename = true;
1562
1563         if (!rec->itr) {
1564                 rec->itr = auxtrace_record__init(rec->evlist, &err);
1565                 if (err)
1566                         return err;
1567         }
1568
1569         err = auxtrace_parse_snapshot_options(rec->itr, &rec->opts,
1570                                               rec->opts.auxtrace_snapshot_opts);
1571         if (err)
1572                 return err;
1573
1574         if (dry_run)
1575                 return 0;
1576
1577         err = bpf__setup_stdout(rec->evlist);
1578         if (err) {
1579                 bpf__strerror_setup_stdout(rec->evlist, err, errbuf, sizeof(errbuf));
1580                 pr_err("ERROR: Setup BPF stdout failed: %s\n",
1581                          errbuf);
1582                 return err;
1583         }
1584
1585         err = -ENOMEM;
1586
1587         symbol__init(NULL);
1588
1589         if (symbol_conf.kptr_restrict)
1590                 pr_warning(
1591 "WARNING: Kernel address maps (/proc/{kallsyms,modules}) are restricted,\n"
1592 "check /proc/sys/kernel/kptr_restrict.\n\n"
1593 "Samples in kernel functions may not be resolved if a suitable vmlinux\n"
1594 "file is not found in the buildid cache or in the vmlinux path.\n\n"
1595 "Samples in kernel modules won't be resolved at all.\n\n"
1596 "If some relocation was applied (e.g. kexec) symbols may be misresolved\n"
1597 "even with a suitable vmlinux or kallsyms file.\n\n");
1598
1599         if (rec->no_buildid_cache || rec->no_buildid) {
1600                 disable_buildid_cache();
1601         } else if (rec->switch_output) {
1602                 /*
1603                  * In 'perf record --switch-output', disable buildid
1604                  * generation by default to reduce data file switching
1605                  * overhead. Still generate buildid if they are required
1606                  * explicitly using
1607                  *
1608                  *  perf record --signal-trigger --no-no-buildid \
1609                  *              --no-no-buildid-cache
1610                  *
1611                  * Following code equals to:
1612                  *
1613                  * if ((rec->no_buildid || !rec->no_buildid_set) &&
1614                  *     (rec->no_buildid_cache || !rec->no_buildid_cache_set))
1615                  *         disable_buildid_cache();
1616                  */
1617                 bool disable = true;
1618
1619                 if (rec->no_buildid_set && !rec->no_buildid)
1620                         disable = false;
1621                 if (rec->no_buildid_cache_set && !rec->no_buildid_cache)
1622                         disable = false;
1623                 if (disable) {
1624                         rec->no_buildid = true;
1625                         rec->no_buildid_cache = true;
1626                         disable_buildid_cache();
1627                 }
1628         }
1629
1630         if (record.opts.overwrite)
1631                 record.opts.tail_synthesize = true;
1632
1633         if (rec->evlist->nr_entries == 0 &&
1634             perf_evlist__add_default(rec->evlist) < 0) {
1635                 pr_err("Not enough memory for event selector list\n");
1636                 goto out_symbol_exit;
1637         }
1638
1639         if (rec->opts.target.tid && !rec->opts.no_inherit_set)
1640                 rec->opts.no_inherit = true;
1641
1642         err = target__validate(&rec->opts.target);
1643         if (err) {
1644                 target__strerror(&rec->opts.target, err, errbuf, BUFSIZ);
1645                 ui__warning("%s", errbuf);
1646         }
1647
1648         err = target__parse_uid(&rec->opts.target);
1649         if (err) {
1650                 int saved_errno = errno;
1651
1652                 target__strerror(&rec->opts.target, err, errbuf, BUFSIZ);
1653                 ui__error("%s", errbuf);
1654
1655                 err = -saved_errno;
1656                 goto out_symbol_exit;
1657         }
1658
1659         err = -ENOMEM;
1660         if (perf_evlist__create_maps(rec->evlist, &rec->opts.target) < 0)
1661                 usage_with_options(record_usage, record_options);
1662
1663         err = auxtrace_record__options(rec->itr, rec->evlist, &rec->opts);
1664         if (err)
1665                 goto out_symbol_exit;
1666
1667         /*
1668          * We take all buildids when the file contains
1669          * AUX area tracing data because we do not decode the
1670          * trace because it would take too long.
1671          */
1672         if (rec->opts.full_auxtrace)
1673                 rec->buildid_all = true;
1674
1675         if (record_opts__config(&rec->opts)) {
1676                 err = -EINVAL;
1677                 goto out_symbol_exit;
1678         }
1679
1680         err = __cmd_record(&record, argc, argv);
1681 out_symbol_exit:
1682         perf_evlist__delete(rec->evlist);
1683         symbol__exit();
1684         auxtrace_record__free(rec->itr);
1685         return err;
1686 }
1687
1688 static void snapshot_sig_handler(int sig __maybe_unused)
1689 {
1690         if (trigger_is_ready(&auxtrace_snapshot_trigger)) {
1691                 trigger_hit(&auxtrace_snapshot_trigger);
1692                 auxtrace_record__snapshot_started = 1;
1693                 if (auxtrace_record__snapshot_start(record.itr))
1694                         trigger_error(&auxtrace_snapshot_trigger);
1695         }
1696
1697         if (trigger_is_ready(&switch_output_trigger))
1698                 trigger_hit(&switch_output_trigger);
1699 }