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                 err = perf_event__synthesize_features(
 804                         tool, session, rec->evlist, process_synthesized_event);
 805                 if (err < 0) {
 806                         pr_err("Couldn't synthesize features.\n");
 807                         return err;
 808                 }
 809
 810                 err = perf_event__synthesize_attrs(tool, session,
 811                                                    process_synthesized_event);
 812                 if (err < 0) {
 813                         pr_err("Couldn't synthesize attrs.\n");
 814                         goto out;
 815                 }
 816
 817                 if (have_tracepoints(&rec->evlist->entries)) {
 818                         /*
 819                          * FIXME err <= 0 here actually means that
 820                          * there were no tracepoints so its not really
 821                          * an error, just that we don't need to
 822                          * synthesize anything.  We really have to
 823                          * return this more properly and also
 824                          * propagate errors that now are calling die()
 825                          */
 826                         err = perf_event__synthesize_tracing_data(tool, fd, rec->evlist,
 827                                                                   process_synthesized_event);
 828                         if (err <= 0) {
 829                                 pr_err("Couldn't record tracing data.\n");
 830                                 goto out;
 831                         }
 832                         rec->bytes_written += err;
 833                 }
 834         }
 835
 836         err = perf_event__synth_time_conv(record__pick_pc(rec), tool,
 837                                           process_synthesized_event, machine);
 838         if (err)
 839                 goto out;
 840
 841         if (rec->opts.full_auxtrace) {
 842                 err = perf_event__synthesize_auxtrace_info(rec->itr, tool,
 843                                         session, process_synthesized_event);
 844                 if (err)
 845                         goto out;
 846         }
 847
 848         err = perf_event__synthesize_kernel_mmap(tool, process_synthesized_event,
 849                                                  machine);
 850         WARN_ONCE(err < 0, "Couldn't record kernel reference relocation symbol\n"
 851                            "Symbol resolution may be skewed if relocation was used (e.g. kexec).\n"
 852                            "Check /proc/kallsyms permission or run as root.\n");
 853
 854         err = perf_event__synthesize_modules(tool, process_synthesized_event,
 855                                              machine);
 856         WARN_ONCE(err < 0, "Couldn't record kernel module information.\n"
 857                            "Symbol resolution may be skewed if relocation was used (e.g. kexec).\n"
 858                            "Check /proc/modules permission or run as root.\n");
 859
 860         if (perf_guest) {
 861                 machines__process_guests(&session->machines,
 862                                          perf_event__synthesize_guest_os, tool);
 863         }
 864
 865         err = __machine__synthesize_threads(machine, tool, &opts->target, rec->evlist->threads,
 866                                             process_synthesized_event, opts->sample_address,
 867                                             opts->proc_map_timeout);
 868 out:
 869         return err;
 870 }
 871
 872 static int __cmd_record(struct record *rec, int argc, const char **argv)
 873 {
 874         int err;
 875         int status = 0;
 876         unsigned long waking = 0;
 877         const bool forks = argc > 0;
 878         struct machine *machine;
 879         struct perf_tool *tool = &rec->tool;
 880         struct record_opts *opts = &rec->opts;
 881         struct perf_data_file *file = &rec->file;
 882         struct perf_session *session;
 883         bool disabled = false, draining = false;
 884         int fd;
 885
 886         rec->progname = argv[0];
 887
 888         atexit(record__sig_exit);
 889         signal(SIGCHLD, sig_handler);
 890         signal(SIGINT, sig_handler);
 891         signal(SIGTERM, sig_handler);
 892         signal(SIGSEGV, sigsegv_handler);
 893
 894         if (rec->opts.record_namespaces)
 895                 tool->namespace_events = true;
 896
 897         if (rec->opts.auxtrace_snapshot_mode || rec->switch_output.enabled) {
 898                 signal(SIGUSR2, snapshot_sig_handler);
 899                 if (rec->opts.auxtrace_snapshot_mode)
 900                         trigger_on(&auxtrace_snapshot_trigger);
 901                 if (rec->switch_output.enabled)
 902                         trigger_on(&switch_output_trigger);
 903         } else {
 904                 signal(SIGUSR2, SIG_IGN);
 905         }
 906
 907         session = perf_session__new(file, false, tool);
 908         if (session == NULL) {
 909                 pr_err("Perf session creation failed.\n");
 910                 return -1;
 911         }
 912
 913         fd = perf_data_file__fd(file);
 914         rec->session = session;
 915
 916         record__init_features(rec);
 917
 918         if (forks) {
 919                 err = perf_evlist__prepare_workload(rec->evlist, &opts->target,
 920                                                     argv, file->is_pipe,
 921                                                     workload_exec_failed_signal);
 922                 if (err < 0) {
 923                         pr_err("Couldn't run the workload!\n");
 924                         status = err;
 925                         goto out_delete_session;
 926                 }
 927         }
 928
 929         if (record__open(rec) != 0) {
 930                 err = -1;
 931                 goto out_child;
 932         }
 933
 934         err = bpf__apply_obj_config();
 935         if (err) {
 936                 char errbuf[BUFSIZ];
 937
 938                 bpf__strerror_apply_obj_config(err, errbuf, sizeof(errbuf));
 939                 pr_err("ERROR: Apply config to BPF failed: %s\n",
 940                          errbuf);
 941                 goto out_child;
 942         }
 943
 944         /*
 945          * Normally perf_session__new would do this, but it doesn't have the
 946          * evlist.
 947          */
 948         if (rec->tool.ordered_events && !perf_evlist__sample_id_all(rec->evlist)) {
 949                 pr_warning("WARNING: No sample_id_all support, falling back to unordered processing\n");
 950                 rec->tool.ordered_events = false;
 951         }
 952
 953         if (!rec->evlist->nr_groups)
 954                 perf_header__clear_feat(&session->header, HEADER_GROUP_DESC);
 955
 956         if (file->is_pipe) {
 957                 err = perf_header__write_pipe(fd);
 958                 if (err < 0)
 959                         goto out_child;
 960         } else {
 961                 err = perf_session__write_header(session, rec->evlist, fd, false);
 962                 if (err < 0)
 963                         goto out_child;
 964         }
 965
 966         if (!rec->no_buildid
 967             && !perf_header__has_feat(&session->header, HEADER_BUILD_ID)) {
 968                 pr_err("Couldn't generate buildids. "
 969                        "Use --no-buildid to profile anyway.\n");
 970                 err = -1;
 971                 goto out_child;
 972         }
 973
 974         machine = &session->machines.host;
 975
 976         err = record__synthesize(rec, false);
 977         if (err < 0)
 978                 goto out_child;
 979
 980         if (rec->realtime_prio) {
 981                 struct sched_param param;
 982
 983                 param.sched_priority = rec->realtime_prio;
 984                 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
 985                         pr_err("Could not set realtime priority.\n");
 986                         err = -1;
 987                         goto out_child;
 988                 }
 989         }
 990
 991         /*
 992          * When perf is starting the traced process, all the events
 993          * (apart from group members) have enable_on_exec=1 set,
 994          * so don't spoil it by prematurely enabling them.
 995          */
 996         if (!target__none(&opts->target) && !opts->initial_delay)
 997                 perf_evlist__enable(rec->evlist);
 998
 999         /*
1000          * Let the child rip
1001          */
1002         if (forks) {
1003                 union perf_event *event;
1004                 pid_t tgid;
1005
1006                 event = malloc(sizeof(event->comm) + machine->id_hdr_size);
1007                 if (event == NULL) {
1008                         err = -ENOMEM;
1009                         goto out_child;
1010                 }
1011
1012                 /*
1013                  * Some H/W events are generated before COMM event
1014                  * which is emitted during exec(), so perf script
1015                  * cannot see a correct process name for those events.
1016                  * Synthesize COMM event to prevent it.
1017                  */
1018                 tgid = perf_event__synthesize_comm(tool, event,
1019                                                    rec->evlist->workload.pid,
1020                                                    process_synthesized_event,
1021                                                    machine);
1022                 free(event);
1023
1024                 if (tgid == -1)
1025                         goto out_child;
1026
1027                 event = malloc(sizeof(event->namespaces) +
1028                                (NR_NAMESPACES * sizeof(struct perf_ns_link_info)) +
1029                                machine->id_hdr_size);
1030                 if (event == NULL) {
1031                         err = -ENOMEM;
1032                         goto out_child;
1033                 }
1034
1035                 /*
1036                  * Synthesize NAMESPACES event for the command specified.
1037                  */
1038                 perf_event__synthesize_namespaces(tool, event,
1039                                                   rec->evlist->workload.pid,
1040                                                   tgid, process_synthesized_event,
1041                                                   machine);
1042                 free(event);
1043
1044                 perf_evlist__start_workload(rec->evlist);
1045         }
1046
1047         if (opts->initial_delay) {
1048                 usleep(opts->initial_delay * USEC_PER_MSEC);
1049                 perf_evlist__enable(rec->evlist);
1050         }
1051
1052         trigger_ready(&auxtrace_snapshot_trigger);
1053         trigger_ready(&switch_output_trigger);
1054         perf_hooks__invoke_record_start();
1055         for (;;) {
1056                 unsigned long long hits = rec->samples;
1057
1058                 /*
1059                  * rec->evlist->bkw_mmap_state is possible to be
1060                  * BKW_MMAP_EMPTY here: when done == true and
1061                  * hits != rec->samples in previous round.
1062                  *
1063                  * perf_evlist__toggle_bkw_mmap ensure we never
1064                  * convert BKW_MMAP_EMPTY to BKW_MMAP_DATA_PENDING.
1065                  */
1066                 if (trigger_is_hit(&switch_output_trigger) || done || draining)
1067                         perf_evlist__toggle_bkw_mmap(rec->evlist, BKW_MMAP_DATA_PENDING);
1068
1069                 if (record__mmap_read_all(rec) < 0) {
1070                         trigger_error(&auxtrace_snapshot_trigger);
1071                         trigger_error(&switch_output_trigger);
1072                         err = -1;
1073                         goto out_child;
1074                 }
1075
1076                 if (auxtrace_record__snapshot_started) {
1077                         auxtrace_record__snapshot_started = 0;
1078                         if (!trigger_is_error(&auxtrace_snapshot_trigger))
1079                                 record__read_auxtrace_snapshot(rec);
1080                         if (trigger_is_error(&auxtrace_snapshot_trigger)) {
1081                                 pr_err("AUX area tracing snapshot failed\n");
1082                                 err = -1;
1083                                 goto out_child;
1084                         }
1085                 }
1086
1087                 if (trigger_is_hit(&switch_output_trigger)) {
1088                         /*
1089                          * If switch_output_trigger is hit, the data in
1090                          * overwritable ring buffer should have been collected,
1091                          * so bkw_mmap_state should be set to BKW_MMAP_EMPTY.
1092                          *
1093                          * If SIGUSR2 raise after or during record__mmap_read_all(),
1094                          * record__mmap_read_all() didn't collect data from
1095                          * overwritable ring buffer. Read again.
1096                          */
1097                         if (rec->evlist->bkw_mmap_state == BKW_MMAP_RUNNING)
1098                                 continue;
1099                         trigger_ready(&switch_output_trigger);
1100
1101                         /*
1102                          * Reenable events in overwrite ring buffer after
1103                          * record__mmap_read_all(): we should have collected
1104                          * data from it.
1105                          */
1106                         perf_evlist__toggle_bkw_mmap(rec->evlist, BKW_MMAP_RUNNING);
1107
1108                         if (!quiet)
1109                                 fprintf(stderr, "[ perf record: dump data: Woken up %ld times ]\n",
1110                                         waking);
1111                         waking = 0;
1112                         fd = record__switch_output(rec, false);
1113                         if (fd < 0) {
1114                                 pr_err("Failed to switch to new file\n");
1115                                 trigger_error(&switch_output_trigger);
1116                                 err = fd;
1117                                 goto out_child;
1118                         }
1119
1120                         /* re-arm the alarm */
1121                         if (rec->switch_output.time)
1122                                 alarm(rec->switch_output.time);
1123                 }
1124
1125                 if (hits == rec->samples) {
1126                         if (done || draining)
1127                                 break;
1128                         err = perf_evlist__poll(rec->evlist, -1);
1129                         /*
1130                          * Propagate error, only if there's any. Ignore positive
1131                          * number of returned events and interrupt error.
1132                          */
1133                         if (err > 0 || (err < 0 && errno == EINTR))
1134                                 err = 0;
1135                         waking++;
1136
1137                         if (perf_evlist__filter_pollfd(rec->evlist, POLLERR | POLLHUP) == 0)
1138                                 draining = true;
1139                 }
1140
1141                 /*
1142                  * When perf is starting the traced process, at the end events
1143                  * die with the process and we wait for that. Thus no need to
1144                  * disable events in this case.
1145                  */
1146                 if (done && !disabled && !target__none(&opts->target)) {
1147                         trigger_off(&auxtrace_snapshot_trigger);
1148                         perf_evlist__disable(rec->evlist);
1149                         disabled = true;
1150                 }
1151         }
1152         trigger_off(&auxtrace_snapshot_trigger);
1153         trigger_off(&switch_output_trigger);
1154
1155         if (forks && workload_exec_errno) {
1156                 char msg[STRERR_BUFSIZE];
1157                 const char *emsg = str_error_r(workload_exec_errno, msg, sizeof(msg));
1158                 pr_err("Workload failed: %s\n", emsg);
1159                 err = -1;
1160                 goto out_child;
1161         }
1162
1163         if (!quiet)
1164                 fprintf(stderr, "[ perf record: Woken up %ld times to write data ]\n", waking);
1165
1166         if (target__none(&rec->opts.target))
1167                 record__synthesize_workload(rec, true);
1168
1169 out_child:
1170         if (forks) {
1171                 int exit_status;
1172
1173                 if (!child_finished)
1174                         kill(rec->evlist->workload.pid, SIGTERM);
1175
1176                 wait(&exit_status);
1177
1178                 if (err < 0)
1179                         status = err;
1180                 else if (WIFEXITED(exit_status))
1181                         status = WEXITSTATUS(exit_status);
1182                 else if (WIFSIGNALED(exit_status))
1183                         signr = WTERMSIG(exit_status);
1184         } else
1185                 status = err;
1186
1187         record__synthesize(rec, true);
1188         /* this will be recalculated during process_buildids() */
1189         rec->samples = 0;
1190
1191         if (!err) {
1192                 if (!rec->timestamp_filename) {
1193                         record__finish_output(rec);
1194                 } else {
1195                         fd = record__switch_output(rec, true);
1196                         if (fd < 0) {
1197                                 status = fd;
1198                                 goto out_delete_session;
1199                         }
1200                 }
1201         }
1202
1203         perf_hooks__invoke_record_end();
1204
1205         if (!err && !quiet) {
1206                 char samples[128];
1207                 const char *postfix = rec->timestamp_filename ?
1208                                         ".<timestamp>" : "";
1209
1210                 if (rec->samples && !rec->opts.full_auxtrace)
1211                         scnprintf(samples, sizeof(samples),
1212                                   " (%" PRIu64 " samples)", rec->samples);
1213                 else
1214                         samples[0] = '\0';
1215
1216                 fprintf(stderr, "[ perf record: Captured and wrote %.3f MB %s%s%s ]\n",
1217                         perf_data_file__size(file) / 1024.0 / 1024.0,
1218                         file->path, postfix, samples);
1219         }
1220
1221 out_delete_session:
1222         perf_session__delete(session);
1223         return status;
1224 }
1225
1226 static void callchain_debug(struct callchain_param *callchain)
1227 {
1228         static const char *str[CALLCHAIN_MAX] = { "NONE", "FP", "DWARF", "LBR" };
1229
1230         pr_debug("callchain: type %s\n", str[callchain->record_mode]);
1231
1232         if (callchain->record_mode == CALLCHAIN_DWARF)
1233                 pr_debug("callchain: stack dump size %d\n",
1234                          callchain->dump_size);
1235 }
1236
1237 int record_opts__parse_callchain(struct record_opts *record,
1238                                  struct callchain_param *callchain,
1239                                  const char *arg, bool unset)
1240 {
1241         int ret;
1242         callchain->enabled = !unset;
1243
1244         /* --no-call-graph */
1245         if (unset) {
1246                 callchain->record_mode = CALLCHAIN_NONE;
1247                 pr_debug("callchain: disabled\n");
1248                 return 0;
1249         }
1250
1251         ret = parse_callchain_record_opt(arg, callchain);
1252         if (!ret) {
1253                 /* Enable data address sampling for DWARF unwind. */
1254                 if (callchain->record_mode == CALLCHAIN_DWARF)
1255                         record->sample_address = true;
1256                 callchain_debug(callchain);
1257         }
1258
1259         return ret;
1260 }
1261
1262 int record_parse_callchain_opt(const struct option *opt,
1263                                const char *arg,
1264                                int unset)
1265 {
1266         return record_opts__parse_callchain(opt->value, &callchain_param, arg, unset);
1267 }
1268
1269 int record_callchain_opt(const struct option *opt,
1270                          const char *arg __maybe_unused,
1271                          int unset __maybe_unused)
1272 {
1273         struct callchain_param *callchain = opt->value;
1274
1275         callchain->enabled = true;
1276
1277         if (callchain->record_mode == CALLCHAIN_NONE)
1278                 callchain->record_mode = CALLCHAIN_FP;
1279
1280         callchain_debug(callchain);
1281         return 0;
1282 }
1283
1284 static int perf_record_config(const char *var, const char *value, void *cb)
1285 {
1286         struct record *rec = cb;
1287
1288         if (!strcmp(var, "record.build-id")) {
1289                 if (!strcmp(value, "cache"))
1290                         rec->no_buildid_cache = false;
1291                 else if (!strcmp(value, "no-cache"))
1292                         rec->no_buildid_cache = true;
1293                 else if (!strcmp(value, "skip"))
1294                         rec->no_buildid = true;
1295                 else
1296                         return -1;
1297                 return 0;
1298         }
1299         if (!strcmp(var, "record.call-graph"))
1300                 var = "call-graph.record-mode"; /* fall-through */
1301
1302         return perf_default_config(var, value, cb);
1303 }
1304
1305 struct clockid_map {
1306         const char *name;
1307         int clockid;
1308 };
1309
1310 #define CLOCKID_MAP(n, c)       \
1311         { .name = n, .clockid = (c), }
1312
1313 #define CLOCKID_END     { .name = NULL, }
1314
1315
1316 /*
1317  * Add the missing ones, we need to build on many distros...
1318  */
1319 #ifndef CLOCK_MONOTONIC_RAW
1320 #define CLOCK_MONOTONIC_RAW 4
1321 #endif
1322 #ifndef CLOCK_BOOTTIME
1323 #define CLOCK_BOOTTIME 7
1324 #endif
1325 #ifndef CLOCK_TAI
1326 #define CLOCK_TAI 11
1327 #endif
1328
1329 static const struct clockid_map clockids[] = {
1330         /* available for all events, NMI safe */
1331         CLOCKID_MAP("monotonic", CLOCK_MONOTONIC),
1332         CLOCKID_MAP("monotonic_raw", CLOCK_MONOTONIC_RAW),
1333
1334         /* available for some events */
1335         CLOCKID_MAP("realtime", CLOCK_REALTIME),
1336         CLOCKID_MAP("boottime", CLOCK_BOOTTIME),
1337         CLOCKID_MAP("tai", CLOCK_TAI),
1338
1339         /* available for the lazy */
1340         CLOCKID_MAP("mono", CLOCK_MONOTONIC),
1341         CLOCKID_MAP("raw", CLOCK_MONOTONIC_RAW),
1342         CLOCKID_MAP("real", CLOCK_REALTIME),
1343         CLOCKID_MAP("boot", CLOCK_BOOTTIME),
1344
1345         CLOCKID_END,
1346 };
1347
1348 static int parse_clockid(const struct option *opt, const char *str, int unset)
1349 {
1350         struct record_opts *opts = (struct record_opts *)opt->value;
1351         const struct clockid_map *cm;
1352         const char *ostr = str;
1353
1354         if (unset) {
1355                 opts->use_clockid = 0;
1356                 return 0;
1357         }
1358
1359         /* no arg passed */
1360         if (!str)
1361                 return 0;
1362
1363         /* no setting it twice */
1364         if (opts->use_clockid)
1365                 return -1;
1366
1367         opts->use_clockid = true;
1368
1369         /* if its a number, we're done */
1370         if (sscanf(str, "%d", &opts->clockid) == 1)
1371                 return 0;
1372
1373         /* allow a "CLOCK_" prefix to the name */
1374         if (!strncasecmp(str, "CLOCK_", 6))
1375                 str += 6;
1376
1377         for (cm = clockids; cm->name; cm++) {
1378                 if (!strcasecmp(str, cm->name)) {
1379                         opts->clockid = cm->clockid;
1380                         return 0;
1381                 }
1382         }
1383
1384         opts->use_clockid = false;
1385         ui__warning("unknown clockid %s, check man page\n", ostr);
1386         return -1;
1387 }
1388
1389 static int record__parse_mmap_pages(const struct option *opt,
1390                                     const char *str,
1391                                     int unset __maybe_unused)
1392 {
1393         struct record_opts *opts = opt->value;
1394         char *s, *p;
1395         unsigned int mmap_pages;
1396         int ret;
1397
1398         if (!str)
1399                 return -EINVAL;
1400
1401         s = strdup(str);
1402         if (!s)
1403                 return -ENOMEM;
1404
1405         p = strchr(s, ',');
1406         if (p)
1407                 *p = '\0';
1408
1409         if (*s) {
1410                 ret = __perf_evlist__parse_mmap_pages(&mmap_pages, s);
1411                 if (ret)
1412                         goto out_free;
1413                 opts->mmap_pages = mmap_pages;
1414         }
1415
1416         if (!p) {
1417                 ret = 0;
1418                 goto out_free;
1419         }
1420
1421         ret = __perf_evlist__parse_mmap_pages(&mmap_pages, p + 1);
1422         if (ret)
1423                 goto out_free;
1424
1425         opts->auxtrace_mmap_pages = mmap_pages;
1426
1427 out_free:
1428         free(s);
1429         return ret;
1430 }
1431
1432 static void switch_output_size_warn(struct record *rec)
1433 {
1434         u64 wakeup_size = perf_evlist__mmap_size(rec->opts.mmap_pages);
1435         struct switch_output *s = &rec->switch_output;
1436
1437         wakeup_size /= 2;
1438
1439         if (s->size < wakeup_size) {
1440                 char buf[100];
1441
1442                 unit_number__scnprintf(buf, sizeof(buf), wakeup_size);
1443                 pr_warning("WARNING: switch-output data size lower than "
1444                            "wakeup kernel buffer size (%s) "
1445                            "expect bigger perf.data sizes\n", buf);
1446         }
1447 }
1448
1449 static int switch_output_setup(struct record *rec)
1450 {
1451         struct switch_output *s = &rec->switch_output;
1452         static struct parse_tag tags_size[] = {
1453                 { .tag  = 'B', .mult = 1       },
1454                 { .tag  = 'K', .mult = 1 << 10 },
1455                 { .tag  = 'M', .mult = 1 << 20 },
1456                 { .tag  = 'G', .mult = 1 << 30 },
1457                 { .tag  = 0 },
1458         };
1459         static struct parse_tag tags_time[] = {
1460                 { .tag  = 's', .mult = 1        },
1461                 { .tag  = 'm', .mult = 60       },
1462                 { .tag  = 'h', .mult = 60*60    },
1463                 { .tag  = 'd', .mult = 60*60*24 },
1464                 { .tag  = 0 },
1465         };
1466         unsigned long val;
1467
1468         if (!s->set)
1469                 return 0;
1470
1471         if (!strcmp(s->str, "signal")) {
1472                 s->signal = true;
1473                 pr_debug("switch-output with SIGUSR2 signal\n");
1474                 goto enabled;
1475         }
1476
1477         val = parse_tag_value(s->str, tags_size);
1478         if (val != (unsigned long) -1) {
1479                 s->size = val;
1480                 pr_debug("switch-output with %s size threshold\n", s->str);
1481                 goto enabled;
1482         }
1483
1484         val = parse_tag_value(s->str, tags_time);
1485         if (val != (unsigned long) -1) {
1486                 s->time = val;
1487                 pr_debug("switch-output with %s time threshold (%lu seconds)\n",
1488                          s->str, s->time);
1489                 goto enabled;
1490         }
1491
1492         return -1;
1493
1494 enabled:
1495         rec->timestamp_filename = true;
1496         s->enabled              = true;
1497
1498         if (s->size && !rec->opts.no_buffering)
1499                 switch_output_size_warn(rec);
1500
1501         return 0;
1502 }
1503
1504 static const char * const __record_usage[] = {
1505         "perf record [<options>] [<command>]",
1506         "perf record [<options>] -- <command> [<options>]",
1507         NULL
1508 };
1509 const char * const *record_usage = __record_usage;
1510
1511 /*
1512  * XXX Ideally would be local to cmd_record() and passed to a record__new
1513  * because we need to have access to it in record__exit, that is called
1514  * after cmd_record() exits, but since record_options need to be accessible to
1515  * builtin-script, leave it here.
1516  *
1517  * At least we don't ouch it in all the other functions here directly.
1518  *
1519  * Just say no to tons of global variables, sigh.
1520  */
1521 static struct record record = {
1522         .opts = {
1523                 .sample_time         = true,
1524                 .mmap_pages          = UINT_MAX,
1525                 .user_freq           = UINT_MAX,
1526                 .user_interval       = ULLONG_MAX,
1527                 .freq                = 4000,
1528                 .target              = {
1529                         .uses_mmap   = true,
1530                         .default_per_cpu = true,
1531                 },
1532                 .proc_map_timeout     = 500,
1533         },
1534         .tool = {
1535                 .sample         = process_sample_event,
1536                 .fork           = perf_event__process_fork,
1537                 .exit           = perf_event__process_exit,
1538                 .comm           = perf_event__process_comm,
1539                 .namespaces     = perf_event__process_namespaces,
1540                 .mmap           = perf_event__process_mmap,
1541                 .mmap2          = perf_event__process_mmap2,
1542                 .ordered_events = true,
1543         },
1544 };
1545
1546 const char record_callchain_help[] = CALLCHAIN_RECORD_HELP
1547         "\n\t\t\t\tDefault: fp";
1548
1549 static bool dry_run;
1550
1551 /*
1552  * XXX Will stay a global variable till we fix builtin-script.c to stop messing
1553  * with it and switch to use the library functions in perf_evlist that came
1554  * from builtin-record.c, i.e. use record_opts,
1555  * perf_evlist__prepare_workload, etc instead of fork+exec'in 'perf record',
1556  * using pipes, etc.
1557  */
1558 static struct option __record_options[] = {
1559         OPT_CALLBACK('e', "event", &record.evlist, "event",
1560                      "event selector. use 'perf list' to list available events",
1561                      parse_events_option),
1562         OPT_CALLBACK(0, "filter", &record.evlist, "filter",
1563                      "event filter", parse_filter),
1564         OPT_CALLBACK_NOOPT(0, "exclude-perf", &record.evlist,
1565                            NULL, "don't record events from perf itself",
1566                            exclude_perf),
1567         OPT_STRING('p', "pid", &record.opts.target.pid, "pid",
1568                     "record events on existing process id"),
1569         OPT_STRING('t', "tid", &record.opts.target.tid, "tid",
1570                     "record events on existing thread id"),
1571         OPT_INTEGER('r', "realtime", &record.realtime_prio,
1572                     "collect data with this RT SCHED_FIFO priority"),
1573         OPT_BOOLEAN(0, "no-buffering", &record.opts.no_buffering,
1574                     "collect data without buffering"),
1575         OPT_BOOLEAN('R', "raw-samples", &record.opts.raw_samples,
1576                     "collect raw sample records from all opened counters"),
1577         OPT_BOOLEAN('a', "all-cpus", &record.opts.target.system_wide,
1578                             "system-wide collection from all CPUs"),
1579         OPT_STRING('C', "cpu", &record.opts.target.cpu_list, "cpu",
1580                     "list of cpus to monitor"),
1581         OPT_U64('c', "count", &record.opts.user_interval, "event period to sample"),
1582         OPT_STRING('o', "output", &record.file.path, "file",
1583                     "output file name"),
1584         OPT_BOOLEAN_SET('i', "no-inherit", &record.opts.no_inherit,
1585                         &record.opts.no_inherit_set,
1586                         "child tasks do not inherit counters"),
1587         OPT_BOOLEAN(0, "tail-synthesize", &record.opts.tail_synthesize,
1588                     "synthesize non-sample events at the end of output"),
1589         OPT_BOOLEAN(0, "overwrite", &record.opts.overwrite, "use overwrite mode"),
1590         OPT_UINTEGER('F', "freq", &record.opts.user_freq, "profile at this frequency"),
1591         OPT_CALLBACK('m', "mmap-pages", &record.opts, "pages[,pages]",
1592                      "number of mmap data pages and AUX area tracing mmap pages",
1593                      record__parse_mmap_pages),
1594         OPT_BOOLEAN(0, "group", &record.opts.group,
1595                     "put the counters into a counter group"),
1596         OPT_CALLBACK_NOOPT('g', NULL, &callchain_param,
1597                            NULL, "enables call-graph recording" ,
1598                            &record_callchain_opt),
1599         OPT_CALLBACK(0, "call-graph", &record.opts,
1600                      "record_mode[,record_size]", record_callchain_help,
1601                      &record_parse_callchain_opt),
1602         OPT_INCR('v', "verbose", &verbose,
1603                     "be more verbose (show counter open errors, etc)"),
1604         OPT_BOOLEAN('q', "quiet", &quiet, "don't print any message"),
1605         OPT_BOOLEAN('s', "stat", &record.opts.inherit_stat,
1606                     "per thread counts"),
1607         OPT_BOOLEAN('d', "data", &record.opts.sample_address, "Record the sample addresses"),
1608         OPT_BOOLEAN(0, "phys-data", &record.opts.sample_phys_addr,
1609                     "Record the sample physical addresses"),
1610         OPT_BOOLEAN(0, "sample-cpu", &record.opts.sample_cpu, "Record the sample cpu"),
1611         OPT_BOOLEAN_SET('T', "timestamp", &record.opts.sample_time,
1612                         &record.opts.sample_time_set,
1613                         "Record the sample timestamps"),
1614         OPT_BOOLEAN('P', "period", &record.opts.period, "Record the sample period"),
1615         OPT_BOOLEAN('n', "no-samples", &record.opts.no_samples,
1616                     "don't sample"),
1617         OPT_BOOLEAN_SET('N', "no-buildid-cache", &record.no_buildid_cache,
1618                         &record.no_buildid_cache_set,
1619                         "do not update the buildid cache"),
1620         OPT_BOOLEAN_SET('B', "no-buildid", &record.no_buildid,
1621                         &record.no_buildid_set,
1622                         "do not collect buildids in perf.data"),
1623         OPT_CALLBACK('G', "cgroup", &record.evlist, "name",
1624                      "monitor event in cgroup name only",
1625                      parse_cgroups),
1626         OPT_UINTEGER('D', "delay", &record.opts.initial_delay,
1627                   "ms to wait before starting measurement after program start"),
1628         OPT_STRING('u', "uid", &record.opts.target.uid_str, "user",
1629                    "user to profile"),
1630
1631         OPT_CALLBACK_NOOPT('b', "branch-any", &record.opts.branch_stack,
1632                      "branch any", "sample any taken branches",
1633                      parse_branch_stack),
1634
1635         OPT_CALLBACK('j', "branch-filter", &record.opts.branch_stack,
1636                      "branch filter mask", "branch stack filter modes",
1637                      parse_branch_stack),
1638         OPT_BOOLEAN('W', "weight", &record.opts.sample_weight,
1639                     "sample by weight (on special events only)"),
1640         OPT_BOOLEAN(0, "transaction", &record.opts.sample_transaction,
1641                     "sample transaction flags (special events only)"),
1642         OPT_BOOLEAN(0, "per-thread", &record.opts.target.per_thread,
1643                     "use per-thread mmaps"),
1644         OPT_CALLBACK_OPTARG('I', "intr-regs", &record.opts.sample_intr_regs, NULL, "any register",
1645                     "sample selected machine registers on interrupt,"
1646                     " use -I ? to list register names", parse_regs),
1647         OPT_BOOLEAN(0, "running-time", &record.opts.running_time,
1648                     "Record running/enabled time of read (:S) events"),
1649         OPT_CALLBACK('k', "clockid", &record.opts,
1650         "clockid", "clockid to use for events, see clock_gettime()",
1651         parse_clockid),
1652         OPT_STRING_OPTARG('S', "snapshot", &record.opts.auxtrace_snapshot_opts,
1653                           "opts", "AUX area tracing Snapshot Mode", ""),
1654         OPT_UINTEGER(0, "proc-map-timeout", &record.opts.proc_map_timeout,
1655                         "per thread proc mmap processing timeout in ms"),
1656         OPT_BOOLEAN(0, "namespaces", &record.opts.record_namespaces,
1657                     "Record namespaces events"),
1658         OPT_BOOLEAN(0, "switch-events", &record.opts.record_switch_events,
1659                     "Record context switch events"),
1660         OPT_BOOLEAN_FLAG(0, "all-kernel", &record.opts.all_kernel,
1661                          "Configure all used events to run in kernel space.",
1662                          PARSE_OPT_EXCLUSIVE),
1663         OPT_BOOLEAN_FLAG(0, "all-user", &record.opts.all_user,
1664                          "Configure all used events to run in user space.",
1665                          PARSE_OPT_EXCLUSIVE),
1666         OPT_STRING(0, "clang-path", &llvm_param.clang_path, "clang path",
1667                    "clang binary to use for compiling BPF scriptlets"),
1668         OPT_STRING(0, "clang-opt", &llvm_param.clang_opt, "clang options",
1669                    "options passed to clang when compiling BPF scriptlets"),
1670         OPT_STRING(0, "vmlinux", &symbol_conf.vmlinux_name,
1671                    "file", "vmlinux pathname"),
1672         OPT_BOOLEAN(0, "buildid-all", &record.buildid_all,
1673                     "Record build-id of all DSOs regardless of hits"),
1674         OPT_BOOLEAN(0, "timestamp-filename", &record.timestamp_filename,
1675                     "append timestamp to output filename"),
1676         OPT_STRING_OPTARG_SET(0, "switch-output", &record.switch_output.str,
1677                           &record.switch_output.set, "signal,size,time",
1678                           "Switch output when receive SIGUSR2 or cross size,time threshold",
1679                           "signal"),
1680         OPT_BOOLEAN(0, "dry-run", &dry_run,
1681                     "Parse options then exit"),
1682         OPT_END()
1683 };
1684
1685 struct option *record_options = __record_options;
1686
1687 int cmd_record(int argc, const char **argv)
1688 {
1689         int err;
1690         struct record *rec = &record;
1691         char errbuf[BUFSIZ];
1692
1693 #ifndef HAVE_LIBBPF_SUPPORT
1694 # define set_nobuild(s, l, c) set_option_nobuild(record_options, s, l, "NO_LIBBPF=1", c)
1695         set_nobuild('\0', "clang-path", true);
1696         set_nobuild('\0', "clang-opt", true);
1697 # undef set_nobuild
1698 #endif
1699
1700 #ifndef HAVE_BPF_PROLOGUE
1701 # if !defined (HAVE_DWARF_SUPPORT)
1702 #  define REASON  "NO_DWARF=1"
1703 # elif !defined (HAVE_LIBBPF_SUPPORT)
1704 #  define REASON  "NO_LIBBPF=1"
1705 # else
1706 #  define REASON  "this architecture doesn't support BPF prologue"
1707 # endif
1708 # define set_nobuild(s, l, c) set_option_nobuild(record_options, s, l, REASON, c)
1709         set_nobuild('\0', "vmlinux", true);
1710 # undef set_nobuild
1711 # undef REASON
1712 #endif
1713
1714         rec->evlist = perf_evlist__new();
1715         if (rec->evlist == NULL)
1716                 return -ENOMEM;
1717
1718         err = perf_config(perf_record_config, rec);
1719         if (err)
1720                 return err;
1721
1722         argc = parse_options(argc, argv, record_options, record_usage,
1723                             PARSE_OPT_STOP_AT_NON_OPTION);
1724         if (quiet)
1725                 perf_quiet_option();
1726
1727         /* Make system wide (-a) the default target. */
1728         if (!argc && target__none(&rec->opts.target))
1729                 rec->opts.target.system_wide = true;
1730
1731         if (nr_cgroups && !rec->opts.target.system_wide) {
1732                 usage_with_options_msg(record_usage, record_options,
1733                         "cgroup monitoring only available in system-wide mode");
1734
1735         }
1736         if (rec->opts.record_switch_events &&
1737             !perf_can_record_switch_events()) {
1738                 ui__error("kernel does not support recording context switch events\n");
1739                 parse_options_usage(record_usage, record_options, "switch-events", 0);
1740                 return -EINVAL;
1741         }
1742
1743         if (switch_output_setup(rec)) {
1744                 parse_options_usage(record_usage, record_options, "switch-output", 0);
1745                 return -EINVAL;
1746         }
1747
1748         if (rec->switch_output.time) {
1749                 signal(SIGALRM, alarm_sig_handler);
1750                 alarm(rec->switch_output.time);
1751         }
1752
1753         if (!rec->itr) {
1754                 rec->itr = auxtrace_record__init(rec->evlist, &err);
1755                 if (err)
1756                         goto out;
1757         }
1758
1759         err = auxtrace_parse_snapshot_options(rec->itr, &rec->opts,
1760                                               rec->opts.auxtrace_snapshot_opts);
1761         if (err)
1762                 goto out;
1763
1764         /*
1765          * Allow aliases to facilitate the lookup of symbols for address
1766          * filters. Refer to auxtrace_parse_filters().
1767          */
1768         symbol_conf.allow_aliases = true;
1769
1770         symbol__init(NULL);
1771
1772         err = auxtrace_parse_filters(rec->evlist);
1773         if (err)
1774                 goto out;
1775
1776         if (dry_run)
1777                 goto out;
1778
1779         err = bpf__setup_stdout(rec->evlist);
1780         if (err) {
1781                 bpf__strerror_setup_stdout(rec->evlist, err, errbuf, sizeof(errbuf));
1782                 pr_err("ERROR: Setup BPF stdout failed: %s\n",
1783                          errbuf);
1784                 goto out;
1785         }
1786
1787         err = -ENOMEM;
1788
1789         if (symbol_conf.kptr_restrict)
1790                 pr_warning(
1791 "WARNING: Kernel address maps (/proc/{kallsyms,modules}) are restricted,\n"
1792 "check /proc/sys/kernel/kptr_restrict.\n\n"
1793 "Samples in kernel functions may not be resolved if a suitable vmlinux\n"
1794 "file is not found in the buildid cache or in the vmlinux path.\n\n"
1795 "Samples in kernel modules won't be resolved at all.\n\n"
1796 "If some relocation was applied (e.g. kexec) symbols may be misresolved\n"
1797 "even with a suitable vmlinux or kallsyms file.\n\n");
1798
1799         if (rec->no_buildid_cache || rec->no_buildid) {
1800                 disable_buildid_cache();
1801         } else if (rec->switch_output.enabled) {
1802                 /*
1803                  * In 'perf record --switch-output', disable buildid
1804                  * generation by default to reduce data file switching
1805                  * overhead. Still generate buildid if they are required
1806                  * explicitly using
1807                  *
1808                  *  perf record --switch-output --no-no-buildid \
1809                  *              --no-no-buildid-cache
1810                  *
1811                  * Following code equals to:
1812                  *
1813                  * if ((rec->no_buildid || !rec->no_buildid_set) &&
1814                  *     (rec->no_buildid_cache || !rec->no_buildid_cache_set))
1815                  *         disable_buildid_cache();
1816                  */
1817                 bool disable = true;
1818
1819                 if (rec->no_buildid_set && !rec->no_buildid)
1820                         disable = false;
1821                 if (rec->no_buildid_cache_set && !rec->no_buildid_cache)
1822                         disable = false;
1823                 if (disable) {
1824                         rec->no_buildid = true;
1825                         rec->no_buildid_cache = true;
1826                         disable_buildid_cache();
1827                 }
1828         }
1829
1830         if (record.opts.overwrite)
1831                 record.opts.tail_synthesize = true;
1832
1833         if (rec->evlist->nr_entries == 0 &&
1834             __perf_evlist__add_default(rec->evlist, !record.opts.no_samples) < 0) {
1835                 pr_err("Not enough memory for event selector list\n");
1836                 goto out;
1837         }
1838
1839         if (rec->opts.target.tid && !rec->opts.no_inherit_set)
1840                 rec->opts.no_inherit = true;
1841
1842         err = target__validate(&rec->opts.target);
1843         if (err) {
1844                 target__strerror(&rec->opts.target, err, errbuf, BUFSIZ);
1845                 ui__warning("%s", errbuf);
1846         }
1847
1848         err = target__parse_uid(&rec->opts.target);
1849         if (err) {
1850                 int saved_errno = errno;
1851
1852                 target__strerror(&rec->opts.target, err, errbuf, BUFSIZ);
1853                 ui__error("%s", errbuf);
1854
1855                 err = -saved_errno;
1856                 goto out;
1857         }
1858
1859         /* Enable ignoring missing threads when -u option is defined. */
1860         rec->opts.ignore_missing_thread = rec->opts.target.uid != UINT_MAX;
1861
1862         err = -ENOMEM;
1863         if (perf_evlist__create_maps(rec->evlist, &rec->opts.target) < 0)
1864                 usage_with_options(record_usage, record_options);
1865
1866         err = auxtrace_record__options(rec->itr, rec->evlist, &rec->opts);
1867         if (err)
1868                 goto out;
1869
1870         /*
1871          * We take all buildids when the file contains
1872          * AUX area tracing data because we do not decode the
1873          * trace because it would take too long.
1874          */
1875         if (rec->opts.full_auxtrace)
1876                 rec->buildid_all = true;
1877
1878         if (record_opts__config(&rec->opts)) {
1879                 err = -EINVAL;
1880                 goto out;
1881         }
1882
1883         err = __cmd_record(&record, argc, argv);
1884 out:
1885         perf_evlist__delete(rec->evlist);
1886         symbol__exit();
1887         auxtrace_record__free(rec->itr);
1888         return err;
1889 }
1890
1891 static void snapshot_sig_handler(int sig __maybe_unused)
1892 {
1893         struct record *rec = &record;
1894
1895         if (trigger_is_ready(&auxtrace_snapshot_trigger)) {
1896                 trigger_hit(&auxtrace_snapshot_trigger);
1897                 auxtrace_record__snapshot_started = 1;
1898                 if (auxtrace_record__snapshot_start(record.itr))
1899                         trigger_error(&auxtrace_snapshot_trigger);
1900         }
1901
1902         if (switch_output_signal(rec))
1903                 trigger_hit(&switch_output_trigger);
1904 }
1905
1906 static void alarm_sig_handler(int sig __maybe_unused)
1907 {
1908         struct record *rec = &record;
1909
1910         if (switch_output_time(rec))
1911                 trigger_hit(&switch_output_trigger);
1912 }