tools/perf/builtin-record.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * builtin-record.c
   4  *
   5  * Builtin record command: Record the profile of a workload
   6  * (or a CPU, or a PID) into the perf.data output file - for
   7  * later analysis via perf report.
   8  */
   9 #include "builtin.h"
  10
  11 #include "perf.h"
  12
  13 #include "util/build-id.h"
  14 #include "util/util.h"
  15 #include <subcmd/parse-options.h>
  16 #include "util/parse-events.h"
  17 #include "util/config.h"
  18
  19 #include "util/callchain.h"
  20 #include "util/cgroup.h"
  21 #include "util/header.h"
  22 #include "util/event.h"
  23 #include "util/evlist.h"
  24 #include "util/evsel.h"
  25 #include "util/debug.h"
  26 #include "util/drv_configs.h"
  27 #include "util/session.h"
  28 #include "util/tool.h"
  29 #include "util/symbol.h"
  30 #include "util/cpumap.h"
  31 #include "util/thread_map.h"
  32 #include "util/data.h"
  33 #include "util/perf_regs.h"
  34 #include "util/auxtrace.h"
  35 #include "util/tsc.h"
  36 #include "util/parse-branch-options.h"
  37 #include "util/parse-regs-options.h"
  38 #include "util/llvm-utils.h"
  39 #include "util/bpf-loader.h"
  40 #include "util/trigger.h"
  41 #include "util/perf-hooks.h"
  42 #include "util/time-utils.h"
  43 #include "util/units.h"
  44 #include "asm/bug.h"
  45
  46 #include <errno.h>
  47 #include <inttypes.h>
  48 #include <poll.h>
  49 #include <unistd.h>
  50 #include <sched.h>
  51 #include <signal.h>
  52 #include <sys/mman.h>
  53 #include <sys/wait.h>
  54 #include <asm/bug.h>
  55 #include <linux/time64.h>
  56
  57 struct switch_output {
  58         bool             enabled;
  59         bool             signal;
  60         unsigned long    size;
  61         unsigned long    time;
  62         const char      *str;
  63         bool             set;
  64 };
  65
  66 struct record {
  67         struct perf_tool        tool;
  68         struct record_opts      opts;
  69         u64                     bytes_written;
  70         struct perf_data_file   file;
  71         struct auxtrace_record  *itr;
  72         struct perf_evlist      *evlist;
  73         struct perf_session     *session;
  74         const char              *progname;
  75         int                     realtime_prio;
  76         bool                    no_buildid;
  77         bool                    no_buildid_set;
  78         bool                    no_buildid_cache;
  79         bool                    no_buildid_cache_set;
  80         bool                    buildid_all;
  81         bool                    timestamp_filename;
  82         struct switch_output    switch_output;
  83         unsigned long long      samples;
  84 };
  85
  86 static volatile int auxtrace_record__snapshot_started;
  87 static DEFINE_TRIGGER(auxtrace_snapshot_trigger);
  88 static DEFINE_TRIGGER(switch_output_trigger);
  89
  90 static bool switch_output_signal(struct record *rec)
  91 {
  92         return rec->switch_output.signal &&
  93                trigger_is_ready(&switch_output_trigger);
  94 }
  95
  96 static bool switch_output_size(struct record *rec)
  97 {
  98         return rec->switch_output.size &&
  99                trigger_is_ready(&switch_output_trigger) &&
 100                (rec->bytes_written >= rec->switch_output.size);
 101 }
 102
 103 static bool switch_output_time(struct record *rec)
 104 {
 105         return rec->switch_output.time &&
 106                trigger_is_ready(&switch_output_trigger);
 107 }
 108
 109 static int record__write(struct record *rec, void *bf, size_t size)
 110 {
 111         if (perf_data_file__write(rec->session->file, bf, size) < 0) {
 112                 pr_err("failed to write perf data, error: %m\n");
 113                 return -1;
 114         }
 115
 116         rec->bytes_written += size;
 117
 118         if (switch_output_size(rec))
 119                 trigger_hit(&switch_output_trigger);
 120
 121         return 0;
 122 }
 123
 124 static int process_synthesized_event(struct perf_tool *tool,
 125                                      union perf_event *event,
 126                                      struct perf_sample *sample __maybe_unused,
 127                                      struct machine *machine __maybe_unused)
 128 {
 129         struct record *rec = container_of(tool, struct record, tool);
 130         return record__write(rec, event, event->header.size);
 131 }
 132
 133 static int
 134 backward_rb_find_range(void *buf, int mask, u64 head, u64 *start, u64 *end)
 135 {
 136         struct perf_event_header *pheader;
 137         u64 evt_head = head;
 138         int size = mask + 1;
 139
 140         pr_debug2("backward_rb_find_range: buf=%p, head=%"PRIx64"\n", buf, head);
 141         pheader = (struct perf_event_header *)(buf + (head & mask));
 142         *start = head;
 143         while (true) {
 144                 if (evt_head - head >= (unsigned int)size) {
 145                         pr_debug("Finished reading backward ring buffer: rewind\n");
 146                         if (evt_head - head > (unsigned int)size)
 147                                 evt_head -= pheader->size;
 148                         *end = evt_head;
 149                         return 0;
 150                 }
 151
 152                 pheader = (struct perf_event_header *)(buf + (evt_head & mask));
 153
 154                 if (pheader->size == 0) {
 155                         pr_debug("Finished reading backward ring buffer: get start\n");
 156                         *end = evt_head;
 157                         return 0;
 158                 }
 159
 160                 evt_head += pheader->size;
 161                 pr_debug3("move evt_head: %"PRIx64"\n", evt_head);
 162         }
 163         WARN_ONCE(1, "Shouldn't get here\n");
 164         return -1;
 165 }
 166
 167 static int
 168 rb_find_range(void *data, int mask, u64 head, u64 old,
 169               u64 *start, u64 *end, bool backward)
 170 {
 171         if (!backward) {
 172                 *start = old;
 173                 *end = head;
 174                 return 0;
 175         }
 176
 177         return backward_rb_find_range(data, mask, head, start, end);
 178 }
 179
 180 static int
 181 record__mmap_read(struct record *rec, struct perf_mmap *md,
 182                   bool overwrite, bool backward)
 183 {
 184         u64 head = perf_mmap__read_head(md);
 185         u64 old = md->prev;
 186         u64 end = head, start = old;
 187         unsigned char *data = md->base + page_size;
 188         unsigned long size;
 189         void *buf;
 190         int rc = 0;
 191
 192         if (rb_find_range(data, md->mask, head,
 193                           old, &start, &end, backward))
 194                 return -1;
 195
 196         if (start == end)
 197                 return 0;
 198
 199         rec->samples++;
 200
 201         size = end - start;
 202         if (size > (unsigned long)(md->mask) + 1) {
 203                 WARN_ONCE(1, "failed to keep up with mmap data. (warn only once)\n");
 204
 205                 md->prev = head;
 206                 perf_mmap__consume(md, overwrite || backward);
 207                 return 0;
 208         }
 209
 210         if ((start & md->mask) + size != (end & md->mask)) {
 211                 buf = &data[start & md->mask];
 212                 size = md->mask + 1 - (start & md->mask);
 213                 start += size;
 214
 215                 if (record__write(rec, buf, size) < 0) {
 216                         rc = -1;
 217                         goto out;
 218                 }
 219         }
 220
 221         buf = &data[start & md->mask];
 222         size = end - start;
 223         start += size;
 224
 225         if (record__write(rec, buf, size) < 0) {
 226                 rc = -1;
 227                 goto out;
 228         }
 229
 230         md->prev = head;
 231         perf_mmap__consume(md, overwrite || backward);
 232 out:
 233         return rc;
 234 }
 235
 236 static volatile int done;
 237 static volatile int signr = -1;
 238 static volatile int child_finished;
 239
 240 static void sig_handler(int sig)
 241 {
 242         if (sig == SIGCHLD)
 243                 child_finished = 1;
 244         else
 245                 signr = sig;
 246
 247         done = 1;
 248 }
 249
 250 static void sigsegv_handler(int sig)
 251 {
 252         perf_hooks__recover();
 253         sighandler_dump_stack(sig);
 254 }
 255
 256 static void record__sig_exit(void)
 257 {
 258         if (signr == -1)
 259                 return;
 260
 261         signal(signr, SIG_DFL);
 262         raise(signr);
 263 }
 264
 265 #ifdef HAVE_AUXTRACE_SUPPORT
 266
 267 static int record__process_auxtrace(struct perf_tool *tool,
 268                                     union perf_event *event, void *data1,
 269                                     size_t len1, void *data2, size_t len2)
 270 {
 271         struct record *rec = container_of(tool, struct record, tool);
 272         struct perf_data_file *file = &rec->file;
 273         size_t padding;
 274         u8 pad[8] = {0};
 275
 276         if (!perf_data_file__is_pipe(file)) {
 277                 off_t file_offset;
 278                 int fd = perf_data_file__fd(file);
 279                 int err;
 280
 281                 file_offset = lseek(fd, 0, SEEK_CUR);
 282                 if (file_offset == -1)
 283                         return -1;
 284                 err = auxtrace_index__auxtrace_event(&rec->session->auxtrace_index,
 285                                                      event, file_offset);
 286                 if (err)
 287                         return err;
 288         }
 289
 290         /* event.auxtrace.size includes padding, see __auxtrace_mmap__read() */
 291         padding = (len1 + len2) & 7;
 292         if (padding)
 293                 padding = 8 - padding;
 294
 295         record__write(rec, event, event->header.size);
 296         record__write(rec, data1, len1);
 297         if (len2)
 298                 record__write(rec, data2, len2);
 299         record__write(rec, &pad, padding);
 300
 301         return 0;
 302 }
 303
 304 static int record__auxtrace_mmap_read(struct record *rec,
 305                                       struct auxtrace_mmap *mm)
 306 {
 307         int ret;
 308
 309         ret = auxtrace_mmap__read(mm, rec->itr, &rec->tool,
 310                                   record__process_auxtrace);
 311         if (ret < 0)
 312                 return ret;
 313
 314         if (ret)
 315                 rec->samples++;
 316
 317         return 0;
 318 }
 319
 320 static int record__auxtrace_mmap_read_snapshot(struct record *rec,
 321                                                struct auxtrace_mmap *mm)
 322 {
 323         int ret;
 324
 325         ret = auxtrace_mmap__read_snapshot(mm, rec->itr, &rec->tool,
 326                                            record__process_auxtrace,
 327                                            rec->opts.auxtrace_snapshot_size);
 328         if (ret < 0)
 329                 return ret;
 330
 331         if (ret)
 332                 rec->samples++;
 333
 334         return 0;
 335 }
 336
 337 static int record__auxtrace_read_snapshot_all(struct record *rec)
 338 {
 339         int i;
 340         int rc = 0;
 341
 342         for (i = 0; i < rec->evlist->nr_mmaps; i++) {
 343                 struct auxtrace_mmap *mm =
 344                                 &rec->evlist->mmap[i].auxtrace_mmap;
 345
 346                 if (!mm->base)
 347                         continue;
 348
 349                 if (record__auxtrace_mmap_read_snapshot(rec, mm) != 0) {
 350                         rc = -1;
 351                         goto out;
 352                 }
 353         }
 354 out:
 355         return rc;
 356 }
 357
 358 static void record__read_auxtrace_snapshot(struct record *rec)
 359 {
 360         pr_debug("Recording AUX area tracing snapshot\n");
 361         if (record__auxtrace_read_snapshot_all(rec) < 0) {
 362                 trigger_error(&auxtrace_snapshot_trigger);
 363         } else {
 364                 if (auxtrace_record__snapshot_finish(rec->itr))
 365                         trigger_error(&auxtrace_snapshot_trigger);
 366                 else
 367                         trigger_ready(&auxtrace_snapshot_trigger);
 368         }
 369 }
 370
 371 #else
 372
 373 static inline
 374 int record__auxtrace_mmap_read(struct record *rec __maybe_unused,
 375                                struct auxtrace_mmap *mm __maybe_unused)
 376 {
 377         return 0;
 378 }
 379
 380 static inline
 381 void record__read_auxtrace_snapshot(struct record *rec __maybe_unused)
 382 {
 383 }
 384
 385 static inline
 386 int auxtrace_record__snapshot_start(struct auxtrace_record *itr __maybe_unused)
 387 {
 388         return 0;
 389 }
 390
 391 #endif
 392
 393 static int record__mmap_evlist(struct record *rec,
 394                                struct perf_evlist *evlist)
 395 {
 396         struct record_opts *opts = &rec->opts;
 397         char msg[512];
 398
 399         if (perf_evlist__mmap_ex(evlist, opts->mmap_pages, false,
 400                                  opts->auxtrace_mmap_pages,
 401                                  opts->auxtrace_snapshot_mode) < 0) {
 402                 if (errno == EPERM) {
 403                         pr_err("Permission error mapping pages.\n"
 404                                "Consider increasing "
 405                                "/proc/sys/kernel/perf_event_mlock_kb,\n"
 406                                "or try again with a smaller value of -m/--mmap_pages.\n"
 407                                "(current value: %u,%u)\n",
 408                                opts->mmap_pages, opts->auxtrace_mmap_pages);
 409                         return -errno;
 410                 } else {
 411                         pr_err("failed to mmap with %d (%s)\n", errno,
 412                                 str_error_r(errno, msg, sizeof(msg)));
 413                         if (errno)
 414                                 return -errno;
 415                         else
 416                                 return -EINVAL;
 417                 }
 418         }
 419         return 0;
 420 }
 421
 422 static int record__mmap(struct record *rec)
 423 {
 424         return record__mmap_evlist(rec, rec->evlist);
 425 }
 426
 427 static int record__open(struct record *rec)
 428 {
 429         char msg[BUFSIZ];
 430         struct perf_evsel *pos;
 431         struct perf_evlist *evlist = rec->evlist;
 432         struct perf_session *session = rec->session;
 433         struct record_opts *opts = &rec->opts;
 434         struct perf_evsel_config_term *err_term;
 435         int rc = 0;
 436
 437         perf_evlist__config(evlist, opts, &callchain_param);
 438
 439         evlist__for_each_entry(evlist, pos) {
 440 try_again:
 441                 if (perf_evsel__open(pos, pos->cpus, pos->threads) < 0) {
 442                         if (perf_evsel__fallback(pos, errno, msg, sizeof(msg))) {
 443                                 if (verbose > 0)
 444                                         ui__warning("%s\n", msg);
 445                                 goto try_again;
 446                         }
 447
 448                         rc = -errno;
 449                         perf_evsel__open_strerror(pos, &opts->target,
 450                                                   errno, msg, sizeof(msg));
 451                         ui__error("%s\n", msg);
 452                         goto out;
 453                 }
 454         }
 455
 456         if (perf_evlist__apply_filters(evlist, &pos)) {
 457                 pr_err("failed to set filter \"%s\" on event %s with %d (%s)\n",
 458                         pos->filter, perf_evsel__name(pos), errno,
 459                         str_error_r(errno, msg, sizeof(msg)));
 460                 rc = -1;
 461                 goto out;
 462         }
 463
 464         if (perf_evlist__apply_drv_configs(evlist, &pos, &err_term)) {
 465                 pr_err("failed to set config \"%s\" on event %s with %d (%s)\n",
 466                       err_term->val.drv_cfg, perf_evsel__name(pos), errno,
 467                       str_error_r(errno, msg, sizeof(msg)));
 468                 rc = -1;
 469                 goto out;
 470         }
 471
 472         rc = record__mmap(rec);
 473         if (rc)
 474                 goto out;
 475
 476         session->evlist = evlist;
 477         perf_session__set_id_hdr_size(session);
 478 out:
 479         return rc;
 480 }
 481
 482 static int process_sample_event(struct perf_tool *tool,
 483                                 union perf_event *event,
 484                                 struct perf_sample *sample,
 485                                 struct perf_evsel *evsel,
 486                                 struct machine *machine)
 487 {
 488         struct record *rec = container_of(tool, struct record, tool);
 489
 490         rec->samples++;
 491
 492         return build_id__mark_dso_hit(tool, event, sample, evsel, machine);
 493 }
 494
 495 static int process_buildids(struct record *rec)
 496 {
 497         struct perf_data_file *file  = &rec->file;
 498         struct perf_session *session = rec->session;
 499
 500         if (file->size == 0)
 501                 return 0;
 502
 503         /*
 504          * During this process, it'll load kernel map and replace the
 505          * dso->long_name to a real pathname it found.  In this case
 506          * we prefer the vmlinux path like
 507          *   /lib/modules/3.16.4/build/vmlinux
 508          *
 509          * rather than build-id path (in debug directory).
 510          *   $HOME/.debug/.build-id/f0/6e17aa50adf4d00b88925e03775de107611551
 511          */
 512         symbol_conf.ignore_vmlinux_buildid = true;
 513
 514         /*
 515          * If --buildid-all is given, it marks all DSO regardless of hits,
 516          * so no need to process samples.
 517          */
 518         if (rec->buildid_all)
 519                 rec->tool.sample = NULL;
 520
 521         return perf_session__process_events(session);
 522 }
 523
 524 static void perf_event__synthesize_guest_os(struct machine *machine, void *data)
 525 {
 526         int err;
 527         struct perf_tool *tool = data;
 528         /*
 529          *As for guest kernel when processing subcommand record&report,
 530          *we arrange module mmap prior to guest kernel mmap and trigger
 531          *a preload dso because default guest module symbols are loaded
 532          *from guest kallsyms instead of /lib/modules/XXX/XXX. This
 533          *method is used to avoid symbol missing when the first addr is
 534          *in module instead of in guest kernel.
 535          */
 536         err = perf_event__synthesize_modules(tool, process_synthesized_event,
 537                                              machine);
 538         if (err < 0)
 539                 pr_err("Couldn't record guest kernel [%d]'s reference"
 540                        " relocation symbol.\n", machine->pid);
 541
 542         /*
 543          * We use _stext for guest kernel because guest kernel's /proc/kallsyms
 544          * have no _text sometimes.
 545          */
 546         err = perf_event__synthesize_kernel_mmap(tool, process_synthesized_event,
 547                                                  machine);
 548         if (err < 0)
 549                 pr_err("Couldn't record guest kernel [%d]'s reference"
 550                        " relocation symbol.\n", machine->pid);
 551 }
 552
 553 static struct perf_event_header finished_round_event = {
 554         .size = sizeof(struct perf_event_header),
 555         .type = PERF_RECORD_FINISHED_ROUND,
 556 };
 557
 558 static int record__mmap_read_evlist(struct record *rec, struct perf_evlist *evlist,
 559                                     bool backward)
 560 {
 561         u64 bytes_written = rec->bytes_written;
 562         int i;
 563         int rc = 0;
 564         struct perf_mmap *maps;
 565
 566         if (!evlist)
 567                 return 0;
 568
 569         maps = backward ? evlist->backward_mmap : evlist->mmap;
 570         if (!maps)
 571                 return 0;
 572
 573         if (backward && evlist->bkw_mmap_state != BKW_MMAP_DATA_PENDING)
 574                 return 0;
 575
 576         for (i = 0; i < evlist->nr_mmaps; i++) {
 577                 struct auxtrace_mmap *mm = &maps[i].auxtrace_mmap;
 578
 579                 if (maps[i].base) {
 580                         if (record__mmap_read(rec, &maps[i],
 581                                               evlist->overwrite, backward) != 0) {
 582                                 rc = -1;
 583                                 goto out;
 584                         }
 585                 }
 586
 587                 if (mm->base && !rec->opts.auxtrace_snapshot_mode &&
 588                     record__auxtrace_mmap_read(rec, mm) != 0) {
 589                         rc = -1;
 590                         goto out;
 591                 }
 592         }
 593
 594         /*
 595          * Mark the round finished in case we wrote
 596          * at least one event.
 597          */
 598         if (bytes_written != rec->bytes_written)
 599                 rc = record__write(rec, &finished_round_event, sizeof(finished_round_event));
 600
 601         if (backward)
 602                 perf_evlist__toggle_bkw_mmap(evlist, BKW_MMAP_EMPTY);
 603 out:
 604         return rc;
 605 }
 606
 607 static int record__mmap_read_all(struct record *rec)
 608 {
 609         int err;
 610
 611         err = record__mmap_read_evlist(rec, rec->evlist, false);
 612         if (err)
 613                 return err;
 614
 615         return record__mmap_read_evlist(rec, rec->evlist, true);
 616 }
 617
 618 static void record__init_features(struct record *rec)
 619 {
 620         struct perf_session *session = rec->session;
 621         int feat;
 622
 623         for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++)
 624                 perf_header__set_feat(&session->header, feat);
 625
 626         if (rec->no_buildid)
 627                 perf_header__clear_feat(&session->header, HEADER_BUILD_ID);
 628
 629         if (!have_tracepoints(&rec->evlist->entries))
 630                 perf_header__clear_feat(&session->header, HEADER_TRACING_DATA);
 631
 632         if (!rec->opts.branch_stack)
 633                 perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK);
 634
 635         if (!rec->opts.full_auxtrace)
 636                 perf_header__clear_feat(&session->header, HEADER_AUXTRACE);
 637
 638         perf_header__clear_feat(&session->header, HEADER_STAT);
 639 }
 640
 641 static void
 642 record__finish_output(struct record *rec)
 643 {
 644         struct perf_data_file *file = &rec->file;
 645         int fd = perf_data_file__fd(file);
 646
 647         if (file->is_pipe)
 648                 return;
 649
 650         rec->session->header.data_size += rec->bytes_written;
 651         file->size = lseek(perf_data_file__fd(file), 0, SEEK_CUR);
 652
 653         if (!rec->no_buildid) {
 654                 process_buildids(rec);
 655
 656                 if (rec->buildid_all)
 657                         dsos__hit_all(rec->session);
 658         }
 659         perf_session__write_header(rec->session, rec->evlist, fd, true);
 660
 661         return;
 662 }
 663
 664 static int record__synthesize_workload(struct record *rec, bool tail)
 665 {
 666         int err;
 667         struct thread_map *thread_map;
 668
 669         if (rec->opts.tail_synthesize != tail)
 670                 return 0;
 671
 672         thread_map = thread_map__new_by_tid(rec->evlist->workload.pid);
 673         if (thread_map == NULL)
 674                 return -1;
 675
 676         err = perf_event__synthesize_thread_map(&rec->tool, thread_map,
 677                                                  process_synthesized_event,
 678                                                  &rec->session->machines.host,
 679                                                  rec->opts.sample_address,
 680                                                  rec->opts.proc_map_timeout);
 681         thread_map__put(thread_map);
 682         return err;
 683 }
 684
 685 static int record__synthesize(struct record *rec, bool tail);
 686
 687 static int
 688 record__switch_output(struct record *rec, bool at_exit)
 689 {
 690         struct perf_data_file *file = &rec->file;
 691         int fd, err;
 692
 693         /* Same Size:      "2015122520103046"*/
 694         char timestamp[] = "InvalidTimestamp";
 695
 696         record__synthesize(rec, true);
 697         if (target__none(&rec->opts.target))
 698                 record__synthesize_workload(rec, true);
 699
 700         rec->samples = 0;
 701         record__finish_output(rec);
 702         err = fetch_current_timestamp(timestamp, sizeof(timestamp));
 703         if (err) {
 704                 pr_err("Failed to get current timestamp\n");
 705                 return -EINVAL;
 706         }
 707
 708         fd = perf_data_file__switch(file, timestamp,
 709                                     rec->session->header.data_offset,
 710                                     at_exit);
 711         if (fd >= 0 && !at_exit) {
 712                 rec->bytes_written = 0;
 713                 rec->session->header.data_size = 0;
 714         }
 715
 716         if (!quiet)
 717                 fprintf(stderr, "[ perf record: Dump %s.%s ]\n",
 718                         file->path, timestamp);
 719
 720         /* Output tracking events */
 721         if (!at_exit) {
 722                 record__synthesize(rec, false);
 723
 724                 /*
 725                  * In 'perf record --switch-output' without -a,
 726                  * record__synthesize() in record__switch_output() won't
 727                  * generate tracking events because there's no thread_map
 728                  * in evlist. Which causes newly created perf.data doesn't
 729                  * contain map and comm information.
 730                  * Create a fake thread_map and directly call
 731                  * perf_event__synthesize_thread_map() for those events.
 732                  */
 733                 if (target__none(&rec->opts.target))
 734                         record__synthesize_workload(rec, false);
 735         }
 736         return fd;
 737 }
 738
 739 static volatile int workload_exec_errno;
 740
 741 /*
 742  * perf_evlist__prepare_workload will send a SIGUSR1
 743  * if the fork fails, since we asked by setting its
 744  * want_signal to true.
 745  */
 746 static void workload_exec_failed_signal(int signo __maybe_unused,
 747                                         siginfo_t *info,
 748                                         void *ucontext __maybe_unused)
 749 {
 750         workload_exec_errno = info->si_value.sival_int;
 751         done = 1;
 752         child_finished = 1;
 753 }
 754
 755 static void snapshot_sig_handler(int sig);
 756 static void alarm_sig_handler(int sig);
 757
 758 int __weak
 759 perf_event__synth_time_conv(const struct perf_event_mmap_page *pc __maybe_unused,
 760                             struct perf_tool *tool __maybe_unused,
 761                             perf_event__handler_t process __maybe_unused,
 762                             struct machine *machine __maybe_unused)
 763 {
 764         return 0;
 765 }
 766
 767 static const struct perf_event_mmap_page *
 768 perf_evlist__pick_pc(struct perf_evlist *evlist)
 769 {
 770         if (evlist) {
 771                 if (evlist->mmap && evlist->mmap[0].base)
 772                         return evlist->mmap[0].base;
 773                 if (evlist->backward_mmap && evlist->backward_mmap[0].base)
 774                         return evlist->backward_mmap[0].base;
 775         }
 776         return NULL;
 777 }
 778
 779 static const struct perf_event_mmap_page *record__pick_pc(struct record *rec)
 780 {
 781         const struct perf_event_mmap_page *pc;
 782
 783         pc = perf_evlist__pick_pc(rec->evlist);
 784         if (pc)
 785                 return pc;
 786         return NULL;
 787 }
 788
 789 static int record__synthesize(struct record *rec, bool tail)
 790 {
 791         struct perf_session *session = rec->session;
 792         struct machine *machine = &session->machines.host;
 793         struct perf_data_file *file = &rec->file;
 794         struct record_opts *opts = &rec->opts;
 795         struct perf_tool *tool = &rec->tool;
 796         int fd = perf_data_file__fd(file);
 797         int err = 0;
 798
 799         if (rec->opts.tail_synthesize != tail)
 800                 return 0;
 801
 802         if (file->is_pipe) {
 803                 /*
 804                  * We need to synthesize events first, because some
 805                  * features works on top of them (on report side).
 806                  */
 807                 err = perf_event__synthesize_attrs(tool, session,
 808                                                    process_synthesized_event);
 809                 if (err < 0) {
 810                         pr_err("Couldn't synthesize attrs.\n");
 811                         goto out;
 812                 }
 813
 814                 err = perf_event__synthesize_features(tool, session, rec->evlist,
 815                                                       process_synthesized_event);
 816                 if (err < 0) {
 817                         pr_err("Couldn't synthesize features.\n");
 818                         return err;
 819                 }
 820
 821                 if (have_tracepoints(&rec->evlist->entries)) {
 822                         /*
 823                          * FIXME err <= 0 here actually means that
 824                          * there were no tracepoints so its not really
 825                          * an error, just that we don't need to
 826                          * synthesize anything.  We really have to
 827                          * return this more properly and also
 828                          * propagate errors that now are calling die()
 829                          */
 830                         err = perf_event__synthesize_tracing_data(tool, fd, rec->evlist,
 831                                                                   process_synthesized_event);
 832                         if (err <= 0) {
 833                                 pr_err("Couldn't record tracing data.\n");
 834                                 goto out;
 835                         }
 836                         rec->bytes_written += err;
 837                 }
 838         }
 839
 840         err = perf_event__synth_time_conv(record__pick_pc(rec), tool,
 841                                           process_synthesized_event, machine);
 842         if (err)
 843                 goto out;
 844
 845         if (rec->opts.full_auxtrace) {
 846                 err = perf_event__synthesize_auxtrace_info(rec->itr, tool,
 847                                         session, process_synthesized_event);
 848                 if (err)
 849                         goto out;
 850         }
 851
 852         err = perf_event__synthesize_kernel_mmap(tool, process_synthesized_event,
 853                                                  machine);
 854         WARN_ONCE(err < 0, "Couldn't record kernel reference relocation symbol\n"
 855                            "Symbol resolution may be skewed if relocation was used (e.g. kexec).\n"
 856                            "Check /proc/kallsyms permission or run as root.\n");
 857
 858         err = perf_event__synthesize_modules(tool, process_synthesized_event,
 859                                              machine);
 860         WARN_ONCE(err < 0, "Couldn't record kernel module information.\n"
 861                            "Symbol resolution may be skewed if relocation was used (e.g. kexec).\n"
 862                            "Check /proc/modules permission or run as root.\n");
 863
 864         if (perf_guest) {
 865                 machines__process_guests(&session->machines,
 866                                          perf_event__synthesize_guest_os, tool);
 867         }
 868
 869         err = __machine__synthesize_threads(machine, tool, &opts->target, rec->evlist->threads,
 870                                             process_synthesized_event, opts->sample_address,
 871                                             opts->proc_map_timeout);
 872 out:
 873         return err;
 874 }
 875
 876 static int __cmd_record(struct record *rec, int argc, const char **argv)
 877 {
 878         int err;
 879         int status = 0;
 880         unsigned long waking = 0;
 881         const bool forks = argc > 0;
 882         struct machine *machine;
 883         struct perf_tool *tool = &rec->tool;
 884         struct record_opts *opts = &rec->opts;
 885         struct perf_data_file *file = &rec->file;
 886         struct perf_session *session;
 887         bool disabled = false, draining = false;
 888         int fd;
 889
 890         rec->progname = argv[0];
 891
 892         atexit(record__sig_exit);
 893         signal(SIGCHLD, sig_handler);
 894         signal(SIGINT, sig_handler);
 895         signal(SIGTERM, sig_handler);
 896         signal(SIGSEGV, sigsegv_handler);
 897
 898         if (rec->opts.record_namespaces)
 899                 tool->namespace_events = true;
 900
 901         if (rec->opts.auxtrace_snapshot_mode || rec->switch_output.enabled) {
 902                 signal(SIGUSR2, snapshot_sig_handler);
 903                 if (rec->opts.auxtrace_snapshot_mode)
 904                         trigger_on(&auxtrace_snapshot_trigger);
 905                 if (rec->switch_output.enabled)
 906                         trigger_on(&switch_output_trigger);
 907         } else {
 908                 signal(SIGUSR2, SIG_IGN);
 909         }
 910
 911         session = perf_session__new(file, false, tool);
 912         if (session == NULL) {
 913                 pr_err("Perf session creation failed.\n");
 914                 return -1;
 915         }
 916
 917         fd = perf_data_file__fd(file);
 918         rec->session = session;
 919
 920         record__init_features(rec);
 921
 922         if (forks) {
 923                 err = perf_evlist__prepare_workload(rec->evlist, &opts->target,
 924                                                     argv, file->is_pipe,
 925                                                     workload_exec_failed_signal);
 926                 if (err < 0) {
 927                         pr_err("Couldn't run the workload!\n");
 928                         status = err;
 929                         goto out_delete_session;
 930                 }
 931         }
 932
 933         if (record__open(rec) != 0) {
 934                 err = -1;
 935                 goto out_child;
 936         }
 937
 938         err = bpf__apply_obj_config();
 939         if (err) {
 940                 char errbuf[BUFSIZ];
 941
 942                 bpf__strerror_apply_obj_config(err, errbuf, sizeof(errbuf));
 943                 pr_err("ERROR: Apply config to BPF failed: %s\n",
 944                          errbuf);
 945                 goto out_child;
 946         }
 947
 948         /*
 949          * Normally perf_session__new would do this, but it doesn't have the
 950          * evlist.
 951          */
 952         if (rec->tool.ordered_events && !perf_evlist__sample_id_all(rec->evlist)) {
 953                 pr_warning("WARNING: No sample_id_all support, falling back to unordered processing\n");
 954                 rec->tool.ordered_events = false;
 955         }
 956
 957         if (!rec->evlist->nr_groups)
 958                 perf_header__clear_feat(&session->header, HEADER_GROUP_DESC);
 959
 960         if (file->is_pipe) {
 961                 err = perf_header__write_pipe(fd);
 962                 if (err < 0)
 963                         goto out_child;
 964         } else {
 965                 err = perf_session__write_header(session, rec->evlist, fd, false);
 966                 if (err < 0)
 967                         goto out_child;
 968         }
 969
 970         if (!rec->no_buildid
 971             && !perf_header__has_feat(&session->header, HEADER_BUILD_ID)) {
 972                 pr_err("Couldn't generate buildids. "
 973                        "Use --no-buildid to profile anyway.\n");
 974                 err = -1;
 975                 goto out_child;
 976         }
 977
 978         machine = &session->machines.host;
 979
 980         err = record__synthesize(rec, false);
 981         if (err < 0)
 982                 goto out_child;
 983
 984         if (rec->realtime_prio) {
 985                 struct sched_param param;
 986
 987                 param.sched_priority = rec->realtime_prio;
 988                 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
 989                         pr_err("Could not set realtime priority.\n");
 990                         err = -1;
 991                         goto out_child;
 992                 }
 993         }
 994
 995         /*
 996          * When perf is starting the traced process, all the events
 997          * (apart from group members) have enable_on_exec=1 set,
 998          * so don't spoil it by prematurely enabling them.
 999          */
1000         if (!target__none(&opts->target) && !opts->initial_delay)
1001                 perf_evlist__enable(rec->evlist);
1002
1003         /*
1004          * Let the child rip
1005          */
1006         if (forks) {
1007                 union perf_event *event;
1008                 pid_t tgid;
1009
1010                 event = malloc(sizeof(event->comm) + machine->id_hdr_size);
1011                 if (event == NULL) {
1012                         err = -ENOMEM;
1013                         goto out_child;
1014                 }
1015
1016                 /*
1017                  * Some H/W events are generated before COMM event
1018                  * which is emitted during exec(), so perf script
1019                  * cannot see a correct process name for those events.
1020                  * Synthesize COMM event to prevent it.
1021                  */
1022                 tgid = perf_event__synthesize_comm(tool, event,
1023                                                    rec->evlist->workload.pid,
1024                                                    process_synthesized_event,
1025                                                    machine);
1026                 free(event);
1027
1028                 if (tgid == -1)
1029                         goto out_child;
1030
1031                 event = malloc(sizeof(event->namespaces) +
1032                                (NR_NAMESPACES * sizeof(struct perf_ns_link_info)) +
1033                                machine->id_hdr_size);
1034                 if (event == NULL) {
1035                         err = -ENOMEM;
1036                         goto out_child;
1037                 }
1038
1039                 /*
1040                  * Synthesize NAMESPACES event for the command specified.
1041                  */
1042                 perf_event__synthesize_namespaces(tool, event,
1043                                                   rec->evlist->workload.pid,
1044                                                   tgid, process_synthesized_event,
1045                                                   machine);
1046                 free(event);
1047
1048                 perf_evlist__start_workload(rec->evlist);
1049         }
1050
1051         if (opts->initial_delay) {
1052                 usleep(opts->initial_delay * USEC_PER_MSEC);
1053                 perf_evlist__enable(rec->evlist);
1054         }
1055
1056         trigger_ready(&auxtrace_snapshot_trigger);
1057         trigger_ready(&switch_output_trigger);
1058         perf_hooks__invoke_record_start();
1059         for (;;) {
1060                 unsigned long long hits = rec->samples;
1061
1062                 /*
1063                  * rec->evlist->bkw_mmap_state is possible to be
1064                  * BKW_MMAP_EMPTY here: when done == true and
1065                  * hits != rec->samples in previous round.
1066                  *
1067                  * perf_evlist__toggle_bkw_mmap ensure we never
1068                  * convert BKW_MMAP_EMPTY to BKW_MMAP_DATA_PENDING.
1069                  */
1070                 if (trigger_is_hit(&switch_output_trigger) || done || draining)
1071                         perf_evlist__toggle_bkw_mmap(rec->evlist, BKW_MMAP_DATA_PENDING);
1072
1073                 if (record__mmap_read_all(rec) < 0) {
1074                         trigger_error(&auxtrace_snapshot_trigger);
1075                         trigger_error(&switch_output_trigger);
1076                         err = -1;
1077                         goto out_child;
1078                 }
1079
1080                 if (auxtrace_record__snapshot_started) {
1081                         auxtrace_record__snapshot_started = 0;
1082                         if (!trigger_is_error(&auxtrace_snapshot_trigger))
1083                                 record__read_auxtrace_snapshot(rec);
1084                         if (trigger_is_error(&auxtrace_snapshot_trigger)) {
1085                                 pr_err("AUX area tracing snapshot failed\n");
1086                                 err = -1;
1087                                 goto out_child;
1088                         }
1089                 }
1090
1091                 if (trigger_is_hit(&switch_output_trigger)) {
1092                         /*
1093                          * If switch_output_trigger is hit, the data in
1094                          * overwritable ring buffer should have been collected,
1095                          * so bkw_mmap_state should be set to BKW_MMAP_EMPTY.
1096                          *
1097                          * If SIGUSR2 raise after or during record__mmap_read_all(),
1098                          * record__mmap_read_all() didn't collect data from
1099                          * overwritable ring buffer. Read again.
1100                          */
1101                         if (rec->evlist->bkw_mmap_state == BKW_MMAP_RUNNING)
1102                                 continue;
1103                         trigger_ready(&switch_output_trigger);
1104
1105                         /*
1106                          * Reenable events in overwrite ring buffer after
1107                          * record__mmap_read_all(): we should have collected
1108                          * data from it.
1109                          */
1110                         perf_evlist__toggle_bkw_mmap(rec->evlist, BKW_MMAP_RUNNING);
1111
1112                         if (!quiet)
1113                                 fprintf(stderr, "[ perf record: dump data: Woken up %ld times ]\n",
1114                                         waking);
1115                         waking = 0;
1116                         fd = record__switch_output(rec, false);
1117                         if (fd < 0) {
1118                                 pr_err("Failed to switch to new file\n");
1119                                 trigger_error(&switch_output_trigger);
1120                                 err = fd;
1121                                 goto out_child;
1122                         }
1123
1124                         /* re-arm the alarm */
1125                         if (rec->switch_output.time)
1126                                 alarm(rec->switch_output.time);
1127                 }
1128
1129                 if (hits == rec->samples) {
1130                         if (done || draining)
1131                                 break;
1132                         err = perf_evlist__poll(rec->evlist, -1);
1133                         /*
1134                          * Propagate error, only if there's any. Ignore positive
1135                          * number of returned events and interrupt error.
1136                          */
1137                         if (err > 0 || (err < 0 && errno == EINTR))
1138                                 err = 0;
1139                         waking++;
1140
1141                         if (perf_evlist__filter_pollfd(rec->evlist, POLLERR | POLLHUP) == 0)
1142                                 draining = true;
1143                 }
1144
1145                 /*
1146                  * When perf is starting the traced process, at the end events
1147                  * die with the process and we wait for that. Thus no need to
1148                  * disable events in this case.
1149                  */
1150                 if (done && !disabled && !target__none(&opts->target)) {
1151                         trigger_off(&auxtrace_snapshot_trigger);
1152                         perf_evlist__disable(rec->evlist);
1153                         disabled = true;
1154                 }
1155         }
1156         trigger_off(&auxtrace_snapshot_trigger);
1157         trigger_off(&switch_output_trigger);
1158
1159         if (forks && workload_exec_errno) {
1160                 char msg[STRERR_BUFSIZE];
1161                 const char *emsg = str_error_r(workload_exec_errno, msg, sizeof(msg));
1162                 pr_err("Workload failed: %s\n", emsg);
1163                 err = -1;
1164                 goto out_child;
1165         }
1166
1167         if (!quiet)
1168                 fprintf(stderr, "[ perf record: Woken up %ld times to write data ]\n", waking);
1169
1170         if (target__none(&rec->opts.target))
1171                 record__synthesize_workload(rec, true);
1172
1173 out_child:
1174         if (forks) {
1175                 int exit_status;
1176
1177                 if (!child_finished)
1178                         kill(rec->evlist->workload.pid, SIGTERM);
1179
1180                 wait(&exit_status);
1181
1182                 if (err < 0)
1183                         status = err;
1184                 else if (WIFEXITED(exit_status))
1185                         status = WEXITSTATUS(exit_status);
1186                 else if (WIFSIGNALED(exit_status))
1187                         signr = WTERMSIG(exit_status);
1188         } else
1189                 status = err;
1190
1191         record__synthesize(rec, true);
1192         /* this will be recalculated during process_buildids() */
1193         rec->samples = 0;
1194
1195         if (!err) {
1196                 if (!rec->timestamp_filename) {
1197                         record__finish_output(rec);
1198                 } else {
1199                         fd = record__switch_output(rec, true);
1200                         if (fd < 0) {
1201                                 status = fd;
1202                                 goto out_delete_session;
1203                         }
1204                 }
1205         }
1206
1207         perf_hooks__invoke_record_end();
1208
1209         if (!err && !quiet) {
1210                 char samples[128];
1211                 const char *postfix = rec->timestamp_filename ?
1212                                         ".<timestamp>" : "";
1213
1214                 if (rec->samples && !rec->opts.full_auxtrace)
1215                         scnprintf(samples, sizeof(samples),
1216                                   " (%" PRIu64 " samples)", rec->samples);
1217                 else
1218                         samples[0] = '\0';
1219
1220                 fprintf(stderr, "[ perf record: Captured and wrote %.3f MB %s%s%s ]\n",
1221                         perf_data_file__size(file) / 1024.0 / 1024.0,
1222                         file->path, postfix, samples);
1223         }
1224
1225 out_delete_session:
1226         perf_session__delete(session);
1227         return status;
1228 }
1229
1230 static void callchain_debug(struct callchain_param *callchain)
1231 {
1232         static const char *str[CALLCHAIN_MAX] = { "NONE", "FP", "DWARF", "LBR" };
1233
1234         pr_debug("callchain: type %s\n", str[callchain->record_mode]);
1235
1236         if (callchain->record_mode == CALLCHAIN_DWARF)
1237                 pr_debug("callchain: stack dump size %d\n",
1238                          callchain->dump_size);
1239 }
1240
1241 int record_opts__parse_callchain(struct record_opts *record,
1242                                  struct callchain_param *callchain,
1243                                  const char *arg, bool unset)
1244 {
1245         int ret;
1246         callchain->enabled = !unset;
1247
1248         /* --no-call-graph */
1249         if (unset) {
1250                 callchain->record_mode = CALLCHAIN_NONE;
1251                 pr_debug("callchain: disabled\n");
1252                 return 0;
1253         }
1254
1255         ret = parse_callchain_record_opt(arg, callchain);
1256         if (!ret) {
1257                 /* Enable data address sampling for DWARF unwind. */
1258                 if (callchain->record_mode == CALLCHAIN_DWARF)
1259                         record->sample_address = true;
1260                 callchain_debug(callchain);
1261         }
1262
1263         return ret;
1264 }
1265
1266 int record_parse_callchain_opt(const struct option *opt,
1267                                const char *arg,
1268                                int unset)
1269 {
1270         return record_opts__parse_callchain(opt->value, &callchain_param, arg, unset);
1271 }
1272
1273 int record_callchain_opt(const struct option *opt,
1274                          const char *arg __maybe_unused,
1275                          int unset __maybe_unused)
1276 {
1277         struct callchain_param *callchain = opt->value;
1278
1279         callchain->enabled = true;
1280
1281         if (callchain->record_mode == CALLCHAIN_NONE)
1282                 callchain->record_mode = CALLCHAIN_FP;
1283
1284         callchain_debug(callchain);
1285         return 0;
1286 }
1287
1288 static int perf_record_config(const char *var, const char *value, void *cb)
1289 {
1290         struct record *rec = cb;
1291
1292         if (!strcmp(var, "record.build-id")) {
1293                 if (!strcmp(value, "cache"))
1294                         rec->no_buildid_cache = false;
1295                 else if (!strcmp(value, "no-cache"))
1296                         rec->no_buildid_cache = true;
1297                 else if (!strcmp(value, "skip"))
1298                         rec->no_buildid = true;
1299                 else
1300                         return -1;
1301                 return 0;
1302         }
1303         if (!strcmp(var, "record.call-graph"))
1304                 var = "call-graph.record-mode"; /* fall-through */
1305
1306         return perf_default_config(var, value, cb);
1307 }
1308
1309 struct clockid_map {
1310         const char *name;
1311         int clockid;
1312 };
1313
1314 #define CLOCKID_MAP(n, c)       \
1315         { .name = n, .clockid = (c), }
1316
1317 #define CLOCKID_END     { .name = NULL, }
1318
1319
1320 /*
1321  * Add the missing ones, we need to build on many distros...
1322  */
1323 #ifndef CLOCK_MONOTONIC_RAW
1324 #define CLOCK_MONOTONIC_RAW 4
1325 #endif
1326 #ifndef CLOCK_BOOTTIME
1327 #define CLOCK_BOOTTIME 7
1328 #endif
1329 #ifndef CLOCK_TAI
1330 #define CLOCK_TAI 11
1331 #endif
1332
1333 static const struct clockid_map clockids[] = {
1334         /* available for all events, NMI safe */
1335         CLOCKID_MAP("monotonic", CLOCK_MONOTONIC),
1336         CLOCKID_MAP("monotonic_raw", CLOCK_MONOTONIC_RAW),
1337
1338         /* available for some events */
1339         CLOCKID_MAP("realtime", CLOCK_REALTIME),
1340         CLOCKID_MAP("boottime", CLOCK_BOOTTIME),
1341         CLOCKID_MAP("tai", CLOCK_TAI),
1342
1343         /* available for the lazy */
1344         CLOCKID_MAP("mono", CLOCK_MONOTONIC),
1345         CLOCKID_MAP("raw", CLOCK_MONOTONIC_RAW),
1346         CLOCKID_MAP("real", CLOCK_REALTIME),
1347         CLOCKID_MAP("boot", CLOCK_BOOTTIME),
1348
1349         CLOCKID_END,
1350 };
1351
1352 static int parse_clockid(const struct option *opt, const char *str, int unset)
1353 {
1354         struct record_opts *opts = (struct record_opts *)opt->value;
1355         const struct clockid_map *cm;
1356         const char *ostr = str;
1357
1358         if (unset) {
1359                 opts->use_clockid = 0;
1360                 return 0;
1361         }
1362
1363         /* no arg passed */
1364         if (!str)
1365                 return 0;
1366
1367         /* no setting it twice */
1368         if (opts->use_clockid)
1369                 return -1;
1370
1371         opts->use_clockid = true;
1372
1373         /* if its a number, we're done */
1374         if (sscanf(str, "%d", &opts->clockid) == 1)
1375                 return 0;
1376
1377         /* allow a "CLOCK_" prefix to the name */
1378         if (!strncasecmp(str, "CLOCK_", 6))
1379                 str += 6;
1380
1381         for (cm = clockids; cm->name; cm++) {
1382                 if (!strcasecmp(str, cm->name)) {
1383                         opts->clockid = cm->clockid;
1384                         return 0;
1385                 }
1386         }
1387
1388         opts->use_clockid = false;
1389         ui__warning("unknown clockid %s, check man page\n", ostr);
1390         return -1;
1391 }
1392
1393 static int record__parse_mmap_pages(const struct option *opt,
1394                                     const char *str,
1395                                     int unset __maybe_unused)
1396 {
1397         struct record_opts *opts = opt->value;
1398         char *s, *p;
1399         unsigned int mmap_pages;
1400         int ret;
1401
1402         if (!str)
1403                 return -EINVAL;
1404
1405         s = strdup(str);
1406         if (!s)
1407                 return -ENOMEM;
1408
1409         p = strchr(s, ',');
1410         if (p)
1411                 *p = '\0';
1412
1413         if (*s) {
1414                 ret = __perf_evlist__parse_mmap_pages(&mmap_pages, s);
1415                 if (ret)
1416                         goto out_free;
1417                 opts->mmap_pages = mmap_pages;
1418         }
1419
1420         if (!p) {
1421                 ret = 0;
1422                 goto out_free;
1423         }
1424
1425         ret = __perf_evlist__parse_mmap_pages(&mmap_pages, p + 1);
1426         if (ret)
1427                 goto out_free;
1428
1429         opts->auxtrace_mmap_pages = mmap_pages;
1430
1431 out_free:
1432         free(s);
1433         return ret;
1434 }
1435
1436 static void switch_output_size_warn(struct record *rec)
1437 {
1438         u64 wakeup_size = perf_evlist__mmap_size(rec->opts.mmap_pages);
1439         struct switch_output *s = &rec->switch_output;
1440
1441         wakeup_size /= 2;
1442
1443         if (s->size < wakeup_size) {
1444                 char buf[100];
1445
1446                 unit_number__scnprintf(buf, sizeof(buf), wakeup_size);
1447                 pr_warning("WARNING: switch-output data size lower than "
1448                            "wakeup kernel buffer size (%s) "
1449                            "expect bigger perf.data sizes\n", buf);
1450         }
1451 }
1452
1453 static int switch_output_setup(struct record *rec)
1454 {
1455         struct switch_output *s = &rec->switch_output;
1456         static struct parse_tag tags_size[] = {
1457                 { .tag  = 'B', .mult = 1       },
1458                 { .tag  = 'K', .mult = 1 << 10 },
1459                 { .tag  = 'M', .mult = 1 << 20 },
1460                 { .tag  = 'G', .mult = 1 << 30 },
1461                 { .tag  = 0 },
1462         };
1463         static struct parse_tag tags_time[] = {
1464                 { .tag  = 's', .mult = 1        },
1465                 { .tag  = 'm', .mult = 60       },
1466                 { .tag  = 'h', .mult = 60*60    },
1467                 { .tag  = 'd', .mult = 60*60*24 },
1468                 { .tag  = 0 },
1469         };
1470         unsigned long val;
1471
1472         if (!s->set)
1473                 return 0;
1474
1475         if (!strcmp(s->str, "signal")) {
1476                 s->signal = true;
1477                 pr_debug("switch-output with SIGUSR2 signal\n");
1478                 goto enabled;
1479         }
1480
1481         val = parse_tag_value(s->str, tags_size);
1482         if (val != (unsigned long) -1) {
1483                 s->size = val;
1484                 pr_debug("switch-output with %s size threshold\n", s->str);
1485                 goto enabled;
1486         }
1487
1488         val = parse_tag_value(s->str, tags_time);
1489         if (val != (unsigned long) -1) {
1490                 s->time = val;
1491                 pr_debug("switch-output with %s time threshold (%lu seconds)\n",
1492                          s->str, s->time);
1493                 goto enabled;
1494         }
1495
1496         return -1;
1497
1498 enabled:
1499         rec->timestamp_filename = true;
1500         s->enabled              = true;
1501
1502         if (s->size && !rec->opts.no_buffering)
1503                 switch_output_size_warn(rec);
1504
1505         return 0;
1506 }
1507
1508 static const char * const __record_usage[] = {
1509         "perf record [<options>] [<command>]",
1510         "perf record [<options>] -- <command> [<options>]",
1511         NULL
1512 };
1513 const char * const *record_usage = __record_usage;
1514
1515 /*
1516  * XXX Ideally would be local to cmd_record() and passed to a record__new
1517  * because we need to have access to it in record__exit, that is called
1518  * after cmd_record() exits, but since record_options need to be accessible to
1519  * builtin-script, leave it here.
1520  *
1521  * At least we don't ouch it in all the other functions here directly.
1522  *
1523  * Just say no to tons of global variables, sigh.
1524  */
1525 static struct record record = {
1526         .opts = {
1527                 .sample_time         = true,
1528                 .mmap_pages          = UINT_MAX,
1529                 .user_freq           = UINT_MAX,
1530                 .user_interval       = ULLONG_MAX,
1531                 .freq                = 4000,
1532                 .target              = {
1533                         .uses_mmap   = true,
1534                         .default_per_cpu = true,
1535                 },
1536                 .proc_map_timeout     = 500,
1537         },
1538         .tool = {
1539                 .sample         = process_sample_event,
1540                 .fork           = perf_event__process_fork,
1541                 .exit           = perf_event__process_exit,
1542                 .comm           = perf_event__process_comm,
1543                 .namespaces     = perf_event__process_namespaces,
1544                 .mmap           = perf_event__process_mmap,
1545                 .mmap2          = perf_event__process_mmap2,
1546                 .ordered_events = true,
1547         },
1548 };
1549
1550 const char record_callchain_help[] = CALLCHAIN_RECORD_HELP
1551         "\n\t\t\t\tDefault: fp";
1552
1553 static bool dry_run;
1554
1555 /*
1556  * XXX Will stay a global variable till we fix builtin-script.c to stop messing
1557  * with it and switch to use the library functions in perf_evlist that came
1558  * from builtin-record.c, i.e. use record_opts,
1559  * perf_evlist__prepare_workload, etc instead of fork+exec'in 'perf record',
1560  * using pipes, etc.
1561  */
1562 static struct option __record_options[] = {
1563         OPT_CALLBACK('e', "event", &record.evlist, "event",
1564                      "event selector. use 'perf list' to list available events",
1565                      parse_events_option),
1566         OPT_CALLBACK(0, "filter", &record.evlist, "filter",
1567                      "event filter", parse_filter),
1568         OPT_CALLBACK_NOOPT(0, "exclude-perf", &record.evlist,
1569                            NULL, "don't record events from perf itself",
1570                            exclude_perf),
1571         OPT_STRING('p', "pid", &record.opts.target.pid, "pid",
1572                     "record events on existing process id"),
1573         OPT_STRING('t', "tid", &record.opts.target.tid, "tid",
1574                     "record events on existing thread id"),
1575         OPT_INTEGER('r', "realtime", &record.realtime_prio,
1576                     "collect data with this RT SCHED_FIFO priority"),
1577         OPT_BOOLEAN(0, "no-buffering", &record.opts.no_buffering,
1578                     "collect data without buffering"),
1579         OPT_BOOLEAN('R', "raw-samples", &record.opts.raw_samples,
1580                     "collect raw sample records from all opened counters"),
1581         OPT_BOOLEAN('a', "all-cpus", &record.opts.target.system_wide,
1582                             "system-wide collection from all CPUs"),
1583         OPT_STRING('C', "cpu", &record.opts.target.cpu_list, "cpu",
1584                     "list of cpus to monitor"),
1585         OPT_U64('c', "count", &record.opts.user_interval, "event period to sample"),
1586         OPT_STRING('o', "output", &record.file.path, "file",
1587                     "output file name"),
1588         OPT_BOOLEAN_SET('i', "no-inherit", &record.opts.no_inherit,
1589                         &record.opts.no_inherit_set,
1590                         "child tasks do not inherit counters"),
1591         OPT_BOOLEAN(0, "tail-synthesize", &record.opts.tail_synthesize,
1592                     "synthesize non-sample events at the end of output"),
1593         OPT_BOOLEAN(0, "overwrite", &record.opts.overwrite, "use overwrite mode"),
1594         OPT_UINTEGER('F', "freq", &record.opts.user_freq, "profile at this frequency"),
1595         OPT_CALLBACK('m', "mmap-pages", &record.opts, "pages[,pages]",
1596                      "number of mmap data pages and AUX area tracing mmap pages",
1597                      record__parse_mmap_pages),
1598         OPT_BOOLEAN(0, "group", &record.opts.group,
1599                     "put the counters into a counter group"),
1600         OPT_CALLBACK_NOOPT('g', NULL, &callchain_param,
1601                            NULL, "enables call-graph recording" ,
1602                            &record_callchain_opt),
1603         OPT_CALLBACK(0, "call-graph", &record.opts,
1604                      "record_mode[,record_size]", record_callchain_help,
1605                      &record_parse_callchain_opt),
1606         OPT_INCR('v', "verbose", &verbose,
1607                     "be more verbose (show counter open errors, etc)"),
1608         OPT_BOOLEAN('q', "quiet", &quiet, "don't print any message"),
1609         OPT_BOOLEAN('s', "stat", &record.opts.inherit_stat,
1610                     "per thread counts"),
1611         OPT_BOOLEAN('d', "data", &record.opts.sample_address, "Record the sample addresses"),
1612         OPT_BOOLEAN(0, "phys-data", &record.opts.sample_phys_addr,
1613                     "Record the sample physical addresses"),
1614         OPT_BOOLEAN(0, "sample-cpu", &record.opts.sample_cpu, "Record the sample cpu"),
1615         OPT_BOOLEAN_SET('T', "timestamp", &record.opts.sample_time,
1616                         &record.opts.sample_time_set,
1617                         "Record the sample timestamps"),
1618         OPT_BOOLEAN_SET('P', "period", &record.opts.period, &record.opts.period_set,
1619                         "Record the sample period"),
1620         OPT_BOOLEAN('n', "no-samples", &record.opts.no_samples,
1621                     "don't sample"),
1622         OPT_BOOLEAN_SET('N', "no-buildid-cache", &record.no_buildid_cache,
1623                         &record.no_buildid_cache_set,
1624                         "do not update the buildid cache"),
1625         OPT_BOOLEAN_SET('B', "no-buildid", &record.no_buildid,
1626                         &record.no_buildid_set,
1627                         "do not collect buildids in perf.data"),
1628         OPT_CALLBACK('G', "cgroup", &record.evlist, "name",
1629                      "monitor event in cgroup name only",
1630                      parse_cgroups),
1631         OPT_UINTEGER('D', "delay", &record.opts.initial_delay,
1632                   "ms to wait before starting measurement after program start"),
1633         OPT_STRING('u', "uid", &record.opts.target.uid_str, "user",
1634                    "user to profile"),
1635
1636         OPT_CALLBACK_NOOPT('b', "branch-any", &record.opts.branch_stack,
1637                      "branch any", "sample any taken branches",
1638                      parse_branch_stack),
1639
1640         OPT_CALLBACK('j', "branch-filter", &record.opts.branch_stack,
1641                      "branch filter mask", "branch stack filter modes",
1642                      parse_branch_stack),
1643         OPT_BOOLEAN('W', "weight", &record.opts.sample_weight,
1644                     "sample by weight (on special events only)"),
1645         OPT_BOOLEAN(0, "transaction", &record.opts.sample_transaction,
1646                     "sample transaction flags (special events only)"),
1647         OPT_BOOLEAN(0, "per-thread", &record.opts.target.per_thread,
1648                     "use per-thread mmaps"),
1649         OPT_CALLBACK_OPTARG('I', "intr-regs", &record.opts.sample_intr_regs, NULL, "any register",
1650                     "sample selected machine registers on interrupt,"
1651                     " use -I ? to list register names", parse_regs),
1652         OPT_BOOLEAN(0, "running-time", &record.opts.running_time,
1653                     "Record running/enabled time of read (:S) events"),
1654         OPT_CALLBACK('k', "clockid", &record.opts,
1655         "clockid", "clockid to use for events, see clock_gettime()",
1656         parse_clockid),
1657         OPT_STRING_OPTARG('S', "snapshot", &record.opts.auxtrace_snapshot_opts,
1658                           "opts", "AUX area tracing Snapshot Mode", ""),
1659         OPT_UINTEGER(0, "proc-map-timeout", &record.opts.proc_map_timeout,
1660                         "per thread proc mmap processing timeout in ms"),
1661         OPT_BOOLEAN(0, "namespaces", &record.opts.record_namespaces,
1662                     "Record namespaces events"),
1663         OPT_BOOLEAN(0, "switch-events", &record.opts.record_switch_events,
1664                     "Record context switch events"),
1665         OPT_BOOLEAN_FLAG(0, "all-kernel", &record.opts.all_kernel,
1666                          "Configure all used events to run in kernel space.",
1667                          PARSE_OPT_EXCLUSIVE),
1668         OPT_BOOLEAN_FLAG(0, "all-user", &record.opts.all_user,
1669                          "Configure all used events to run in user space.",
1670                          PARSE_OPT_EXCLUSIVE),
1671         OPT_STRING(0, "clang-path", &llvm_param.clang_path, "clang path",
1672                    "clang binary to use for compiling BPF scriptlets"),
1673         OPT_STRING(0, "clang-opt", &llvm_param.clang_opt, "clang options",
1674                    "options passed to clang when compiling BPF scriptlets"),
1675         OPT_STRING(0, "vmlinux", &symbol_conf.vmlinux_name,
1676                    "file", "vmlinux pathname"),
1677         OPT_BOOLEAN(0, "buildid-all", &record.buildid_all,
1678                     "Record build-id of all DSOs regardless of hits"),
1679         OPT_BOOLEAN(0, "timestamp-filename", &record.timestamp_filename,
1680                     "append timestamp to output filename"),
1681         OPT_STRING_OPTARG_SET(0, "switch-output", &record.switch_output.str,
1682                           &record.switch_output.set, "signal,size,time",
1683                           "Switch output when receive SIGUSR2 or cross size,time threshold",
1684                           "signal"),
1685         OPT_BOOLEAN(0, "dry-run", &dry_run,
1686                     "Parse options then exit"),
1687         OPT_END()
1688 };
1689
1690 struct option *record_options = __record_options;
1691
1692 int cmd_record(int argc, const char **argv)
1693 {
1694         int err;
1695         struct record *rec = &record;
1696         char errbuf[BUFSIZ];
1697
1698 #ifndef HAVE_LIBBPF_SUPPORT
1699 # define set_nobuild(s, l, c) set_option_nobuild(record_options, s, l, "NO_LIBBPF=1", c)
1700         set_nobuild('\0', "clang-path", true);
1701         set_nobuild('\0', "clang-opt", true);
1702 # undef set_nobuild
1703 #endif
1704
1705 #ifndef HAVE_BPF_PROLOGUE
1706 # if !defined (HAVE_DWARF_SUPPORT)
1707 #  define REASON  "NO_DWARF=1"
1708 # elif !defined (HAVE_LIBBPF_SUPPORT)
1709 #  define REASON  "NO_LIBBPF=1"
1710 # else
1711 #  define REASON  "this architecture doesn't support BPF prologue"
1712 # endif
1713 # define set_nobuild(s, l, c) set_option_nobuild(record_options, s, l, REASON, c)
1714         set_nobuild('\0', "vmlinux", true);
1715 # undef set_nobuild
1716 # undef REASON
1717 #endif
1718
1719         rec->evlist = perf_evlist__new();
1720         if (rec->evlist == NULL)
1721                 return -ENOMEM;
1722
1723         err = perf_config(perf_record_config, rec);
1724         if (err)
1725                 return err;
1726
1727         argc = parse_options(argc, argv, record_options, record_usage,
1728                             PARSE_OPT_STOP_AT_NON_OPTION);
1729         if (quiet)
1730                 perf_quiet_option();
1731
1732         /* Make system wide (-a) the default target. */
1733         if (!argc && target__none(&rec->opts.target))
1734                 rec->opts.target.system_wide = true;
1735
1736         if (nr_cgroups && !rec->opts.target.system_wide) {
1737                 usage_with_options_msg(record_usage, record_options,
1738                         "cgroup monitoring only available in system-wide mode");
1739
1740         }
1741         if (rec->opts.record_switch_events &&
1742             !perf_can_record_switch_events()) {
1743                 ui__error("kernel does not support recording context switch events\n");
1744                 parse_options_usage(record_usage, record_options, "switch-events", 0);
1745                 return -EINVAL;
1746         }
1747
1748         if (switch_output_setup(rec)) {
1749                 parse_options_usage(record_usage, record_options, "switch-output", 0);
1750                 return -EINVAL;
1751         }
1752
1753         if (rec->switch_output.time) {
1754                 signal(SIGALRM, alarm_sig_handler);
1755                 alarm(rec->switch_output.time);
1756         }
1757
1758         if (!rec->itr) {
1759                 rec->itr = auxtrace_record__init(rec->evlist, &err);
1760                 if (err)
1761                         goto out;
1762         }
1763
1764         err = auxtrace_parse_snapshot_options(rec->itr, &rec->opts,
1765                                               rec->opts.auxtrace_snapshot_opts);
1766         if (err)
1767                 goto out;
1768
1769         /*
1770          * Allow aliases to facilitate the lookup of symbols for address
1771          * filters. Refer to auxtrace_parse_filters().
1772          */
1773         symbol_conf.allow_aliases = true;
1774
1775         symbol__init(NULL);
1776
1777         err = auxtrace_parse_filters(rec->evlist);
1778         if (err)
1779                 goto out;
1780
1781         if (dry_run)
1782                 goto out;
1783
1784         err = bpf__setup_stdout(rec->evlist);
1785         if (err) {
1786                 bpf__strerror_setup_stdout(rec->evlist, err, errbuf, sizeof(errbuf));
1787                 pr_err("ERROR: Setup BPF stdout failed: %s\n",
1788                          errbuf);
1789                 goto out;
1790         }
1791
1792         err = -ENOMEM;
1793
1794         if (symbol_conf.kptr_restrict)
1795                 pr_warning(
1796 "WARNING: Kernel address maps (/proc/{kallsyms,modules}) are restricted,\n"
1797 "check /proc/sys/kernel/kptr_restrict.\n\n"
1798 "Samples in kernel functions may not be resolved if a suitable vmlinux\n"
1799 "file is not found in the buildid cache or in the vmlinux path.\n\n"
1800 "Samples in kernel modules won't be resolved at all.\n\n"
1801 "If some relocation was applied (e.g. kexec) symbols may be misresolved\n"
1802 "even with a suitable vmlinux or kallsyms file.\n\n");
1803
1804         if (rec->no_buildid_cache || rec->no_buildid) {
1805                 disable_buildid_cache();
1806         } else if (rec->switch_output.enabled) {
1807                 /*
1808                  * In 'perf record --switch-output', disable buildid
1809                  * generation by default to reduce data file switching
1810                  * overhead. Still generate buildid if they are required
1811                  * explicitly using
1812                  *
1813                  *  perf record --switch-output --no-no-buildid \
1814                  *              --no-no-buildid-cache
1815                  *
1816                  * Following code equals to:
1817                  *
1818                  * if ((rec->no_buildid || !rec->no_buildid_set) &&
1819                  *     (rec->no_buildid_cache || !rec->no_buildid_cache_set))
1820                  *         disable_buildid_cache();
1821                  */
1822                 bool disable = true;
1823
1824                 if (rec->no_buildid_set && !rec->no_buildid)
1825                         disable = false;
1826                 if (rec->no_buildid_cache_set && !rec->no_buildid_cache)
1827                         disable = false;
1828                 if (disable) {
1829                         rec->no_buildid = true;
1830                         rec->no_buildid_cache = true;
1831                         disable_buildid_cache();
1832                 }
1833         }
1834
1835         if (record.opts.overwrite)
1836                 record.opts.tail_synthesize = true;
1837
1838         if (rec->evlist->nr_entries == 0 &&
1839             __perf_evlist__add_default(rec->evlist, !record.opts.no_samples) < 0) {
1840                 pr_err("Not enough memory for event selector list\n");
1841                 goto out;
1842         }
1843
1844         if (rec->opts.target.tid && !rec->opts.no_inherit_set)
1845                 rec->opts.no_inherit = true;
1846
1847         err = target__validate(&rec->opts.target);
1848         if (err) {
1849                 target__strerror(&rec->opts.target, err, errbuf, BUFSIZ);
1850                 ui__warning("%s", errbuf);
1851         }
1852
1853         err = target__parse_uid(&rec->opts.target);
1854         if (err) {
1855                 int saved_errno = errno;
1856
1857                 target__strerror(&rec->opts.target, err, errbuf, BUFSIZ);
1858                 ui__error("%s", errbuf);
1859
1860                 err = -saved_errno;
1861                 goto out;
1862         }
1863
1864         /* Enable ignoring missing threads when -u/-p option is defined. */
1865         rec->opts.ignore_missing_thread = rec->opts.target.uid != UINT_MAX || rec->opts.target.pid;
1866
1867         err = -ENOMEM;
1868         if (perf_evlist__create_maps(rec->evlist, &rec->opts.target) < 0)
1869                 usage_with_options(record_usage, record_options);
1870
1871         err = auxtrace_record__options(rec->itr, rec->evlist, &rec->opts);
1872         if (err)
1873                 goto out;
1874
1875         /*
1876          * We take all buildids when the file contains
1877          * AUX area tracing data because we do not decode the
1878          * trace because it would take too long.
1879          */
1880         if (rec->opts.full_auxtrace)
1881                 rec->buildid_all = true;
1882
1883         if (record_opts__config(&rec->opts)) {
1884                 err = -EINVAL;
1885                 goto out;
1886         }
1887
1888         err = __cmd_record(&record, argc, argv);
1889 out:
1890         perf_evlist__delete(rec->evlist);
1891         symbol__exit();
1892         auxtrace_record__free(rec->itr);
1893         return err;
1894 }
1895
1896 static void snapshot_sig_handler(int sig __maybe_unused)
1897 {
1898         struct record *rec = &record;
1899
1900         if (trigger_is_ready(&auxtrace_snapshot_trigger)) {
1901                 trigger_hit(&auxtrace_snapshot_trigger);
1902                 auxtrace_record__snapshot_started = 1;
1903                 if (auxtrace_record__snapshot_start(record.itr))
1904                         trigger_error(&auxtrace_snapshot_trigger);
1905         }
1906
1907         if (switch_output_signal(rec))
1908                 trigger_hit(&switch_output_trigger);
1909 }
1910
1911 static void alarm_sig_handler(int sig __maybe_unused)
1912 {
1913         struct record *rec = &record;
1914
1915         if (switch_output_time(rec))
1916                 trigger_hit(&switch_output_trigger);
1917 }