tools/perf/builtin-record.c

   1 /*
   2  * builtin-record.c
   3  *
   4  * Builtin record command: Record the profile of a workload
   5  * (or a CPU, or a PID) into the perf.data output file - for
   6  * later analysis via perf report.
   7  */
   8 #include "builtin.h"
   9
  10 #include "perf.h"
  11
  12 #include "util/build-id.h"
  13 #include "util/util.h"
  14 #include "util/parse-options.h"
  15 #include "util/parse-events.h"
  16
  17 #include "util/callchain.h"
  18 #include "util/cgroup.h"
  19 #include "util/header.h"
  20 #include "util/event.h"
  21 #include "util/evlist.h"
  22 #include "util/evsel.h"
  23 #include "util/debug.h"
  24 #include "util/session.h"
  25 #include "util/tool.h"
  26 #include "util/symbol.h"
  27 #include "util/cpumap.h"
  28 #include "util/thread_map.h"
  29 #include "util/data.h"
  30 #include "util/perf_regs.h"
  31 #include "util/auxtrace.h"
  32 #include "util/parse-branch-options.h"
  33 #include "util/parse-regs-options.h"
  34 #include "util/llvm-utils.h"
  35
  36 #include <unistd.h>
  37 #include <sched.h>
  38 #include <sys/mman.h>
  39
  40
  41 struct record {
  42         struct perf_tool        tool;
  43         struct record_opts      opts;
  44         u64                     bytes_written;
  45         struct perf_data_file   file;
  46         struct auxtrace_record  *itr;
  47         struct perf_evlist      *evlist;
  48         struct perf_session     *session;
  49         const char              *progname;
  50         int                     realtime_prio;
  51         bool                    no_buildid;
  52         bool                    no_buildid_cache;
  53         unsigned long long      samples;
  54 };
  55
  56 static int record__write(struct record *rec, void *bf, size_t size)
  57 {
  58         if (perf_data_file__write(rec->session->file, bf, size) < 0) {
  59                 pr_err("failed to write perf data, error: %m\n");
  60                 return -1;
  61         }
  62
  63         rec->bytes_written += size;
  64         return 0;
  65 }
  66
  67 static int process_synthesized_event(struct perf_tool *tool,
  68                                      union perf_event *event,
  69                                      struct perf_sample *sample __maybe_unused,
  70                                      struct machine *machine __maybe_unused)
  71 {
  72         struct record *rec = container_of(tool, struct record, tool);
  73         return record__write(rec, event, event->header.size);
  74 }
  75
  76 static int record__mmap_read(struct record *rec, int idx)
  77 {
  78         struct perf_mmap *md = &rec->evlist->mmap[idx];
  79         u64 head = perf_mmap__read_head(md);
  80         u64 old = md->prev;
  81         unsigned char *data = md->base + page_size;
  82         unsigned long size;
  83         void *buf;
  84         int rc = 0;
  85
  86         if (old == head)
  87                 return 0;
  88
  89         rec->samples++;
  90
  91         size = head - old;
  92
  93         if ((old & md->mask) + size != (head & md->mask)) {
  94                 buf = &data[old & md->mask];
  95                 size = md->mask + 1 - (old & md->mask);
  96                 old += size;
  97
  98                 if (record__write(rec, buf, size) < 0) {
  99                         rc = -1;
 100                         goto out;
 101                 }
 102         }
 103
 104         buf = &data[old & md->mask];
 105         size = head - old;
 106         old += size;
 107
 108         if (record__write(rec, buf, size) < 0) {
 109                 rc = -1;
 110                 goto out;
 111         }
 112
 113         md->prev = old;
 114         perf_evlist__mmap_consume(rec->evlist, idx);
 115 out:
 116         return rc;
 117 }
 118
 119 static volatile int done;
 120 static volatile int signr = -1;
 121 static volatile int child_finished;
 122 static volatile int auxtrace_snapshot_enabled;
 123 static volatile int auxtrace_snapshot_err;
 124 static volatile int auxtrace_record__snapshot_started;
 125
 126 static void sig_handler(int sig)
 127 {
 128         if (sig == SIGCHLD)
 129                 child_finished = 1;
 130         else
 131                 signr = sig;
 132
 133         done = 1;
 134 }
 135
 136 static void record__sig_exit(void)
 137 {
 138         if (signr == -1)
 139                 return;
 140
 141         signal(signr, SIG_DFL);
 142         raise(signr);
 143 }
 144
 145 #ifdef HAVE_AUXTRACE_SUPPORT
 146
 147 static int record__process_auxtrace(struct perf_tool *tool,
 148                                     union perf_event *event, void *data1,
 149                                     size_t len1, void *data2, size_t len2)
 150 {
 151         struct record *rec = container_of(tool, struct record, tool);
 152         struct perf_data_file *file = &rec->file;
 153         size_t padding;
 154         u8 pad[8] = {0};
 155
 156         if (!perf_data_file__is_pipe(file)) {
 157                 off_t file_offset;
 158                 int fd = perf_data_file__fd(file);
 159                 int err;
 160
 161                 file_offset = lseek(fd, 0, SEEK_CUR);
 162                 if (file_offset == -1)
 163                         return -1;
 164                 err = auxtrace_index__auxtrace_event(&rec->session->auxtrace_index,
 165                                                      event, file_offset);
 166                 if (err)
 167                         return err;
 168         }
 169
 170         /* event.auxtrace.size includes padding, see __auxtrace_mmap__read() */
 171         padding = (len1 + len2) & 7;
 172         if (padding)
 173                 padding = 8 - padding;
 174
 175         record__write(rec, event, event->header.size);
 176         record__write(rec, data1, len1);
 177         if (len2)
 178                 record__write(rec, data2, len2);
 179         record__write(rec, &pad, padding);
 180
 181         return 0;
 182 }
 183
 184 static int record__auxtrace_mmap_read(struct record *rec,
 185                                       struct auxtrace_mmap *mm)
 186 {
 187         int ret;
 188
 189         ret = auxtrace_mmap__read(mm, rec->itr, &rec->tool,
 190                                   record__process_auxtrace);
 191         if (ret < 0)
 192                 return ret;
 193
 194         if (ret)
 195                 rec->samples++;
 196
 197         return 0;
 198 }
 199
 200 static int record__auxtrace_mmap_read_snapshot(struct record *rec,
 201                                                struct auxtrace_mmap *mm)
 202 {
 203         int ret;
 204
 205         ret = auxtrace_mmap__read_snapshot(mm, rec->itr, &rec->tool,
 206                                            record__process_auxtrace,
 207                                            rec->opts.auxtrace_snapshot_size);
 208         if (ret < 0)
 209                 return ret;
 210
 211         if (ret)
 212                 rec->samples++;
 213
 214         return 0;
 215 }
 216
 217 static int record__auxtrace_read_snapshot_all(struct record *rec)
 218 {
 219         int i;
 220         int rc = 0;
 221
 222         for (i = 0; i < rec->evlist->nr_mmaps; i++) {
 223                 struct auxtrace_mmap *mm =
 224                                 &rec->evlist->mmap[i].auxtrace_mmap;
 225
 226                 if (!mm->base)
 227                         continue;
 228
 229                 if (record__auxtrace_mmap_read_snapshot(rec, mm) != 0) {
 230                         rc = -1;
 231                         goto out;
 232                 }
 233         }
 234 out:
 235         return rc;
 236 }
 237
 238 static void record__read_auxtrace_snapshot(struct record *rec)
 239 {
 240         pr_debug("Recording AUX area tracing snapshot\n");
 241         if (record__auxtrace_read_snapshot_all(rec) < 0) {
 242                 auxtrace_snapshot_err = -1;
 243         } else {
 244                 auxtrace_snapshot_err = auxtrace_record__snapshot_finish(rec->itr);
 245                 if (!auxtrace_snapshot_err)
 246                         auxtrace_snapshot_enabled = 1;
 247         }
 248 }
 249
 250 #else
 251
 252 static inline
 253 int record__auxtrace_mmap_read(struct record *rec __maybe_unused,
 254                                struct auxtrace_mmap *mm __maybe_unused)
 255 {
 256         return 0;
 257 }
 258
 259 static inline
 260 void record__read_auxtrace_snapshot(struct record *rec __maybe_unused)
 261 {
 262 }
 263
 264 static inline
 265 int auxtrace_record__snapshot_start(struct auxtrace_record *itr __maybe_unused)
 266 {
 267         return 0;
 268 }
 269
 270 #endif
 271
 272 static int record__open(struct record *rec)
 273 {
 274         char msg[512];
 275         struct perf_evsel *pos;
 276         struct perf_evlist *evlist = rec->evlist;
 277         struct perf_session *session = rec->session;
 278         struct record_opts *opts = &rec->opts;
 279         struct perf_evsel_config_term *err_term;
 280         int rc = 0;
 281
 282         perf_evlist__config(evlist, opts);
 283
 284         evlist__for_each(evlist, pos) {
 285 try_again:
 286                 if (perf_evsel__open(pos, pos->cpus, pos->threads) < 0) {
 287                         if (perf_evsel__fallback(pos, errno, msg, sizeof(msg))) {
 288                                 if (verbose)
 289                                         ui__warning("%s\n", msg);
 290                                 goto try_again;
 291                         }
 292
 293                         rc = -errno;
 294                         perf_evsel__open_strerror(pos, &opts->target,
 295                                                   errno, msg, sizeof(msg));
 296                         ui__error("%s\n", msg);
 297                         goto out;
 298                 }
 299         }
 300
 301         if (perf_evlist__apply_filters(evlist, &pos)) {
 302                 error("failed to set filter \"%s\" on event %s with %d (%s)\n",
 303                         pos->filter, perf_evsel__name(pos), errno,
 304                         strerror_r(errno, msg, sizeof(msg)));
 305                 rc = -1;
 306                 goto out;
 307         }
 308
 309         if (perf_evlist__apply_drv_configs(evlist, &pos, &err_term)) {
 310                 error("failed to set config \"%s\" on event %s with %d (%s)\n",
 311                         err_term->val.drv_cfg, perf_evsel__name(pos), errno,
 312                         strerror_r(errno, msg, sizeof(msg)));
 313                 rc = -1;
 314                 goto out;
 315         }
 316
 317         if (perf_evlist__mmap_ex(evlist, opts->mmap_pages, false,
 318                                  opts->auxtrace_mmap_pages,
 319                                  opts->auxtrace_snapshot_mode) < 0) {
 320                 if (errno == EPERM) {
 321                         pr_err("Permission error mapping pages.\n"
 322                                "Consider increasing "
 323                                "/proc/sys/kernel/perf_event_mlock_kb,\n"
 324                                "or try again with a smaller value of -m/--mmap_pages.\n"
 325                                "(current value: %u,%u)\n",
 326                                opts->mmap_pages, opts->auxtrace_mmap_pages);
 327                         rc = -errno;
 328                 } else {
 329                         pr_err("failed to mmap with %d (%s)\n", errno,
 330                                 strerror_r(errno, msg, sizeof(msg)));
 331                         rc = -errno;
 332                 }
 333                 goto out;
 334         }
 335
 336         session->evlist = evlist;
 337         perf_session__set_id_hdr_size(session);
 338 out:
 339         return rc;
 340 }
 341
 342 static int process_sample_event(struct perf_tool *tool,
 343                                 union perf_event *event,
 344                                 struct perf_sample *sample,
 345                                 struct perf_evsel *evsel,
 346                                 struct machine *machine)
 347 {
 348         struct record *rec = container_of(tool, struct record, tool);
 349
 350         rec->samples++;
 351
 352         return build_id__mark_dso_hit(tool, event, sample, evsel, machine);
 353 }
 354
 355 static int process_buildids(struct record *rec)
 356 {
 357         struct perf_data_file *file  = &rec->file;
 358         struct perf_session *session = rec->session;
 359
 360         if (file->size == 0)
 361                 return 0;
 362
 363         /*
 364          * During this process, it'll load kernel map and replace the
 365          * dso->long_name to a real pathname it found.  In this case
 366          * we prefer the vmlinux path like
 367          *   /lib/modules/3.16.4/build/vmlinux
 368          *
 369          * rather than build-id path (in debug directory).
 370          *   $HOME/.debug/.build-id/f0/6e17aa50adf4d00b88925e03775de107611551
 371          */
 372         symbol_conf.ignore_vmlinux_buildid = true;
 373
 374         return perf_session__process_events(session);
 375 }
 376
 377 static void perf_event__synthesize_guest_os(struct machine *machine, void *data)
 378 {
 379         int err;
 380         struct perf_tool *tool = data;
 381         /*
 382          *As for guest kernel when processing subcommand record&report,
 383          *we arrange module mmap prior to guest kernel mmap and trigger
 384          *a preload dso because default guest module symbols are loaded
 385          *from guest kallsyms instead of /lib/modules/XXX/XXX. This
 386          *method is used to avoid symbol missing when the first addr is
 387          *in module instead of in guest kernel.
 388          */
 389         err = perf_event__synthesize_modules(tool, process_synthesized_event,
 390                                              machine);
 391         if (err < 0)
 392                 pr_err("Couldn't record guest kernel [%d]'s reference"
 393                        " relocation symbol.\n", machine->pid);
 394
 395         /*
 396          * We use _stext for guest kernel because guest kernel's /proc/kallsyms
 397          * have no _text sometimes.
 398          */
 399         err = perf_event__synthesize_kernel_mmap(tool, process_synthesized_event,
 400                                                  machine);
 401         if (err < 0)
 402                 pr_err("Couldn't record guest kernel [%d]'s reference"
 403                        " relocation symbol.\n", machine->pid);
 404 }
 405
 406 static struct perf_event_header finished_round_event = {
 407         .size = sizeof(struct perf_event_header),
 408         .type = PERF_RECORD_FINISHED_ROUND,
 409 };
 410
 411 static int record__mmap_read_all(struct record *rec)
 412 {
 413         u64 bytes_written = rec->bytes_written;
 414         int i;
 415         int rc = 0;
 416
 417         for (i = 0; i < rec->evlist->nr_mmaps; i++) {
 418                 struct auxtrace_mmap *mm = &rec->evlist->mmap[i].auxtrace_mmap;
 419
 420                 if (rec->evlist->mmap[i].base) {
 421                         if (record__mmap_read(rec, i) != 0) {
 422                                 rc = -1;
 423                                 goto out;
 424                         }
 425                 }
 426
 427                 if (mm->base && !rec->opts.auxtrace_snapshot_mode &&
 428                     record__auxtrace_mmap_read(rec, mm) != 0) {
 429                         rc = -1;
 430                         goto out;
 431                 }
 432         }
 433
 434         /*
 435          * Mark the round finished in case we wrote
 436          * at least one event.
 437          */
 438         if (bytes_written != rec->bytes_written)
 439                 rc = record__write(rec, &finished_round_event, sizeof(finished_round_event));
 440
 441 out:
 442         return rc;
 443 }
 444
 445 static void record__init_features(struct record *rec)
 446 {
 447         struct perf_session *session = rec->session;
 448         int feat;
 449
 450         for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++)
 451                 perf_header__set_feat(&session->header, feat);
 452
 453         if (rec->no_buildid)
 454                 perf_header__clear_feat(&session->header, HEADER_BUILD_ID);
 455
 456         if (!have_tracepoints(&rec->evlist->entries))
 457                 perf_header__clear_feat(&session->header, HEADER_TRACING_DATA);
 458
 459         if (!rec->opts.branch_stack)
 460                 perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK);
 461
 462         if (!rec->opts.full_auxtrace)
 463                 perf_header__clear_feat(&session->header, HEADER_AUXTRACE);
 464 }
 465
 466 static volatile int workload_exec_errno;
 467
 468 /*
 469  * perf_evlist__prepare_workload will send a SIGUSR1
 470  * if the fork fails, since we asked by setting its
 471  * want_signal to true.
 472  */
 473 static void workload_exec_failed_signal(int signo __maybe_unused,
 474                                         siginfo_t *info,
 475                                         void *ucontext __maybe_unused)
 476 {
 477         workload_exec_errno = info->si_value.sival_int;
 478         done = 1;
 479         child_finished = 1;
 480 }
 481
 482 static void snapshot_sig_handler(int sig);
 483
 484 static int __cmd_record(struct record *rec, int argc, const char **argv)
 485 {
 486         int err;
 487         int status = 0;
 488         unsigned long waking = 0;
 489         const bool forks = argc > 0;
 490         struct machine *machine;
 491         struct perf_tool *tool = &rec->tool;
 492         struct record_opts *opts = &rec->opts;
 493         struct perf_data_file *file = &rec->file;
 494         struct perf_session *session;
 495         bool disabled = false, draining = false;
 496         int fd;
 497
 498         rec->progname = argv[0];
 499
 500         atexit(record__sig_exit);
 501         signal(SIGCHLD, sig_handler);
 502         signal(SIGINT, sig_handler);
 503         signal(SIGTERM, sig_handler);
 504         if (rec->opts.auxtrace_snapshot_mode)
 505                 signal(SIGUSR2, snapshot_sig_handler);
 506         else
 507                 signal(SIGUSR2, SIG_IGN);
 508
 509         session = perf_session__new(file, false, tool);
 510         if (session == NULL) {
 511                 pr_err("Perf session creation failed.\n");
 512                 return -1;
 513         }
 514
 515         fd = perf_data_file__fd(file);
 516         rec->session = session;
 517
 518         record__init_features(rec);
 519
 520         if (forks) {
 521                 err = perf_evlist__prepare_workload(rec->evlist, &opts->target,
 522                                                     argv, file->is_pipe,
 523                                                     workload_exec_failed_signal);
 524                 if (err < 0) {
 525                         pr_err("Couldn't run the workload!\n");
 526                         status = err;
 527                         goto out_delete_session;
 528                 }
 529         }
 530
 531         if (record__open(rec) != 0) {
 532                 err = -1;
 533                 goto out_child;
 534         }
 535
 536         /*
 537          * Normally perf_session__new would do this, but it doesn't have the
 538          * evlist.
 539          */
 540         if (rec->tool.ordered_events && !perf_evlist__sample_id_all(rec->evlist)) {
 541                 pr_warning("WARNING: No sample_id_all support, falling back to unordered processing\n");
 542                 rec->tool.ordered_events = false;
 543         }
 544
 545         if (!rec->evlist->nr_groups)
 546                 perf_header__clear_feat(&session->header, HEADER_GROUP_DESC);
 547
 548         if (file->is_pipe) {
 549                 err = perf_header__write_pipe(fd);
 550                 if (err < 0)
 551                         goto out_child;
 552         } else {
 553                 err = perf_session__write_header(session, rec->evlist, fd, false);
 554                 if (err < 0)
 555                         goto out_child;
 556         }
 557
 558         if (!rec->no_buildid
 559             && !perf_header__has_feat(&session->header, HEADER_BUILD_ID)) {
 560                 pr_err("Couldn't generate buildids. "
 561                        "Use --no-buildid to profile anyway.\n");
 562                 err = -1;
 563                 goto out_child;
 564         }
 565
 566         machine = &session->machines.host;
 567
 568         if (file->is_pipe) {
 569                 err = perf_event__synthesize_attrs(tool, session,
 570                                                    process_synthesized_event);
 571                 if (err < 0) {
 572                         pr_err("Couldn't synthesize attrs.\n");
 573                         goto out_child;
 574                 }
 575
 576                 if (have_tracepoints(&rec->evlist->entries)) {
 577                         /*
 578                          * FIXME err <= 0 here actually means that
 579                          * there were no tracepoints so its not really
 580                          * an error, just that we don't need to
 581                          * synthesize anything.  We really have to
 582                          * return this more properly and also
 583                          * propagate errors that now are calling die()
 584                          */
 585                         err = perf_event__synthesize_tracing_data(tool, fd, rec->evlist,
 586                                                                   process_synthesized_event);
 587                         if (err <= 0) {
 588                                 pr_err("Couldn't record tracing data.\n");
 589                                 goto out_child;
 590                         }
 591                         rec->bytes_written += err;
 592                 }
 593         }
 594
 595         if (rec->opts.full_auxtrace) {
 596                 err = perf_event__synthesize_auxtrace_info(rec->itr, tool,
 597                                         session, process_synthesized_event);
 598                 if (err)
 599                         goto out_delete_session;
 600         }
 601
 602         err = perf_event__synthesize_kernel_mmap(tool, process_synthesized_event,
 603                                                  machine);
 604         if (err < 0)
 605                 pr_err("Couldn't record kernel reference relocation symbol\n"
 606                        "Symbol resolution may be skewed if relocation was used (e.g. kexec).\n"
 607                        "Check /proc/kallsyms permission or run as root.\n");
 608
 609         err = perf_event__synthesize_modules(tool, process_synthesized_event,
 610                                              machine);
 611         if (err < 0)
 612                 pr_err("Couldn't record kernel module information.\n"
 613                        "Symbol resolution may be skewed if relocation was used (e.g. kexec).\n"
 614                        "Check /proc/modules permission or run as root.\n");
 615
 616         if (perf_guest) {
 617                 machines__process_guests(&session->machines,
 618                                          perf_event__synthesize_guest_os, tool);
 619         }
 620
 621         err = __machine__synthesize_threads(machine, tool, &opts->target, rec->evlist->threads,
 622                                             process_synthesized_event, opts->sample_address,
 623                                             opts->proc_map_timeout);
 624         if (err != 0)
 625                 goto out_child;
 626
 627         if (rec->realtime_prio) {
 628                 struct sched_param param;
 629
 630                 param.sched_priority = rec->realtime_prio;
 631                 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
 632                         pr_err("Could not set realtime priority.\n");
 633                         err = -1;
 634                         goto out_child;
 635                 }
 636         }
 637
 638         /*
 639          * When perf is starting the traced process, all the events
 640          * (apart from group members) have enable_on_exec=1 set,
 641          * so don't spoil it by prematurely enabling them.
 642          */
 643         if (!target__none(&opts->target) && !opts->initial_delay)
 644                 perf_evlist__enable(rec->evlist);
 645
 646         /*
 647          * Let the child rip
 648          */
 649         if (forks) {
 650                 union perf_event *event;
 651
 652                 event = malloc(sizeof(event->comm) + machine->id_hdr_size);
 653                 if (event == NULL) {
 654                         err = -ENOMEM;
 655                         goto out_child;
 656                 }
 657
 658                 /*
 659                  * Some H/W events are generated before COMM event
 660                  * which is emitted during exec(), so perf script
 661                  * cannot see a correct process name for those events.
 662                  * Synthesize COMM event to prevent it.
 663                  */
 664                 perf_event__synthesize_comm(tool, event,
 665                                             rec->evlist->workload.pid,
 666                                             process_synthesized_event,
 667                                             machine);
 668                 free(event);
 669
 670                 perf_evlist__start_workload(rec->evlist);
 671         }
 672
 673         if (opts->initial_delay) {
 674                 usleep(opts->initial_delay * 1000);
 675                 perf_evlist__enable(rec->evlist);
 676         }
 677
 678         auxtrace_snapshot_enabled = 1;
 679         for (;;) {
 680                 unsigned long long hits = rec->samples;
 681
 682                 if (record__mmap_read_all(rec) < 0) {
 683                         auxtrace_snapshot_enabled = 0;
 684                         err = -1;
 685                         goto out_child;
 686                 }
 687
 688                 if (auxtrace_record__snapshot_started) {
 689                         auxtrace_record__snapshot_started = 0;
 690                         if (!auxtrace_snapshot_err)
 691                                 record__read_auxtrace_snapshot(rec);
 692                         if (auxtrace_snapshot_err) {
 693                                 pr_err("AUX area tracing snapshot failed\n");
 694                                 err = -1;
 695                                 goto out_child;
 696                         }
 697                 }
 698
 699                 if (hits == rec->samples) {
 700                         if (done || draining)
 701                                 break;
 702                         err = perf_evlist__poll(rec->evlist, -1);
 703                         /*
 704                          * Propagate error, only if there's any. Ignore positive
 705                          * number of returned events and interrupt error.
 706                          */
 707                         if (err > 0 || (err < 0 && errno == EINTR))
 708                                 err = 0;
 709                         waking++;
 710
 711                         if (perf_evlist__filter_pollfd(rec->evlist, POLLERR | POLLHUP) == 0)
 712                                 draining = true;
 713                 }
 714
 715                 /*
 716                  * When perf is starting the traced process, at the end events
 717                  * die with the process and we wait for that. Thus no need to
 718                  * disable events in this case.
 719                  */
 720                 if (done && !disabled && !target__none(&opts->target)) {
 721                         auxtrace_snapshot_enabled = 0;
 722                         perf_evlist__disable(rec->evlist);
 723                         disabled = true;
 724                 }
 725         }
 726         auxtrace_snapshot_enabled = 0;
 727
 728         if (forks && workload_exec_errno) {
 729                 char msg[STRERR_BUFSIZE];
 730                 const char *emsg = strerror_r(workload_exec_errno, msg, sizeof(msg));
 731                 pr_err("Workload failed: %s\n", emsg);
 732                 err = -1;
 733                 goto out_child;
 734         }
 735
 736         if (!quiet)
 737                 fprintf(stderr, "[ perf record: Woken up %ld times to write data ]\n", waking);
 738
 739 out_child:
 740         if (forks) {
 741                 int exit_status;
 742
 743                 if (!child_finished)
 744                         kill(rec->evlist->workload.pid, SIGTERM);
 745
 746                 wait(&exit_status);
 747
 748                 if (err < 0)
 749                         status = err;
 750                 else if (WIFEXITED(exit_status))
 751                         status = WEXITSTATUS(exit_status);
 752                 else if (WIFSIGNALED(exit_status))
 753                         signr = WTERMSIG(exit_status);
 754         } else
 755                 status = err;
 756
 757         /* this will be recalculated during process_buildids() */
 758         rec->samples = 0;
 759
 760         if (!err && !file->is_pipe) {
 761                 rec->session->header.data_size += rec->bytes_written;
 762                 file->size = lseek(perf_data_file__fd(file), 0, SEEK_CUR);
 763
 764                 if (!rec->no_buildid) {
 765                         process_buildids(rec);
 766                         /*
 767                          * We take all buildids when the file contains
 768                          * AUX area tracing data because we do not decode the
 769                          * trace because it would take too long.
 770                          */
 771                         if (rec->opts.full_auxtrace)
 772                                 dsos__hit_all(rec->session);
 773                 }
 774                 perf_session__write_header(rec->session, rec->evlist, fd, true);
 775         }
 776
 777         if (!err && !quiet) {
 778                 char samples[128];
 779
 780                 if (rec->samples && !rec->opts.full_auxtrace)
 781                         scnprintf(samples, sizeof(samples),
 782                                   " (%" PRIu64 " samples)", rec->samples);
 783                 else
 784                         samples[0] = '\0';
 785
 786                 fprintf(stderr, "[ perf record: Captured and wrote %.3f MB %s%s ]\n",
 787                         perf_data_file__size(file) / 1024.0 / 1024.0,
 788                         file->path, samples);
 789         }
 790
 791 out_delete_session:
 792         perf_session__delete(session);
 793         return status;
 794 }
 795
 796 static void callchain_debug(void)
 797 {
 798         static const char *str[CALLCHAIN_MAX] = { "NONE", "FP", "DWARF", "LBR" };
 799
 800         pr_debug("callchain: type %s\n", str[callchain_param.record_mode]);
 801
 802         if (callchain_param.record_mode == CALLCHAIN_DWARF)
 803                 pr_debug("callchain: stack dump size %d\n",
 804                          callchain_param.dump_size);
 805 }
 806
 807 int record_parse_callchain_opt(const struct option *opt,
 808                                const char *arg,
 809                                int unset)
 810 {
 811         int ret;
 812         struct record_opts *record = (struct record_opts *)opt->value;
 813
 814         record->callgraph_set = true;
 815         callchain_param.enabled = !unset;
 816
 817         /* --no-call-graph */
 818         if (unset) {
 819                 callchain_param.record_mode = CALLCHAIN_NONE;
 820                 pr_debug("callchain: disabled\n");
 821                 return 0;
 822         }
 823
 824         ret = parse_callchain_record_opt(arg, &callchain_param);
 825         if (!ret)
 826                 callchain_debug();
 827
 828         return ret;
 829 }
 830
 831 int record_callchain_opt(const struct option *opt,
 832                          const char *arg __maybe_unused,
 833                          int unset __maybe_unused)
 834 {
 835         struct record_opts *record = (struct record_opts *)opt->value;
 836
 837         record->callgraph_set = true;
 838         callchain_param.enabled = true;
 839
 840         if (callchain_param.record_mode == CALLCHAIN_NONE)
 841                 callchain_param.record_mode = CALLCHAIN_FP;
 842
 843         callchain_debug();
 844         return 0;
 845 }
 846
 847 static int perf_record_config(const char *var, const char *value, void *cb)
 848 {
 849         if (!strcmp(var, "record.call-graph"))
 850                 var = "call-graph.record-mode"; /* fall-through */
 851
 852         return perf_default_config(var, value, cb);
 853 }
 854
 855 struct clockid_map {
 856         const char *name;
 857         int clockid;
 858 };
 859
 860 #define CLOCKID_MAP(n, c)       \
 861         { .name = n, .clockid = (c), }
 862
 863 #define CLOCKID_END     { .name = NULL, }
 864
 865
 866 /*
 867  * Add the missing ones, we need to build on many distros...
 868  */
 869 #ifndef CLOCK_MONOTONIC_RAW
 870 #define CLOCK_MONOTONIC_RAW 4
 871 #endif
 872 #ifndef CLOCK_BOOTTIME
 873 #define CLOCK_BOOTTIME 7
 874 #endif
 875 #ifndef CLOCK_TAI
 876 #define CLOCK_TAI 11
 877 #endif
 878
 879 static const struct clockid_map clockids[] = {
 880         /* available for all events, NMI safe */
 881         CLOCKID_MAP("monotonic", CLOCK_MONOTONIC),
 882         CLOCKID_MAP("monotonic_raw", CLOCK_MONOTONIC_RAW),
 883
 884         /* available for some events */
 885         CLOCKID_MAP("realtime", CLOCK_REALTIME),
 886         CLOCKID_MAP("boottime", CLOCK_BOOTTIME),
 887         CLOCKID_MAP("tai", CLOCK_TAI),
 888
 889         /* available for the lazy */
 890         CLOCKID_MAP("mono", CLOCK_MONOTONIC),
 891         CLOCKID_MAP("raw", CLOCK_MONOTONIC_RAW),
 892         CLOCKID_MAP("real", CLOCK_REALTIME),
 893         CLOCKID_MAP("boot", CLOCK_BOOTTIME),
 894
 895         CLOCKID_END,
 896 };
 897
 898 static int parse_clockid(const struct option *opt, const char *str, int unset)
 899 {
 900         struct record_opts *opts = (struct record_opts *)opt->value;
 901         const struct clockid_map *cm;
 902         const char *ostr = str;
 903
 904         if (unset) {
 905                 opts->use_clockid = 0;
 906                 return 0;
 907         }
 908
 909         /* no arg passed */
 910         if (!str)
 911                 return 0;
 912
 913         /* no setting it twice */
 914         if (opts->use_clockid)
 915                 return -1;
 916
 917         opts->use_clockid = true;
 918
 919         /* if its a number, we're done */
 920         if (sscanf(str, "%d", &opts->clockid) == 1)
 921                 return 0;
 922
 923         /* allow a "CLOCK_" prefix to the name */
 924         if (!strncasecmp(str, "CLOCK_", 6))
 925                 str += 6;
 926
 927         for (cm = clockids; cm->name; cm++) {
 928                 if (!strcasecmp(str, cm->name)) {
 929                         opts->clockid = cm->clockid;
 930                         return 0;
 931                 }
 932         }
 933
 934         opts->use_clockid = false;
 935         ui__warning("unknown clockid %s, check man page\n", ostr);
 936         return -1;
 937 }
 938
 939 static int record__parse_mmap_pages(const struct option *opt,
 940                                     const char *str,
 941                                     int unset __maybe_unused)
 942 {
 943         struct record_opts *opts = opt->value;
 944         char *s, *p;
 945         unsigned int mmap_pages;
 946         int ret;
 947
 948         if (!str)
 949                 return -EINVAL;
 950
 951         s = strdup(str);
 952         if (!s)
 953                 return -ENOMEM;
 954
 955         p = strchr(s, ',');
 956         if (p)
 957                 *p = '\0';
 958
 959         if (*s) {
 960                 ret = __perf_evlist__parse_mmap_pages(&mmap_pages, s);
 961                 if (ret)
 962                         goto out_free;
 963                 opts->mmap_pages = mmap_pages;
 964         }
 965
 966         if (!p) {
 967                 ret = 0;
 968                 goto out_free;
 969         }
 970
 971         ret = __perf_evlist__parse_mmap_pages(&mmap_pages, p + 1);
 972         if (ret)
 973                 goto out_free;
 974
 975         opts->auxtrace_mmap_pages = mmap_pages;
 976
 977 out_free:
 978         free(s);
 979         return ret;
 980 }
 981
 982 static const char * const __record_usage[] = {
 983         "perf record [<options>] [<command>]",
 984         "perf record [<options>] -- <command> [<options>]",
 985         NULL
 986 };
 987 const char * const *record_usage = __record_usage;
 988
 989 /*
 990  * XXX Ideally would be local to cmd_record() and passed to a record__new
 991  * because we need to have access to it in record__exit, that is called
 992  * after cmd_record() exits, but since record_options need to be accessible to
 993  * builtin-script, leave it here.
 994  *
 995  * At least we don't ouch it in all the other functions here directly.
 996  *
 997  * Just say no to tons of global variables, sigh.
 998  */
 999 static struct record record = {
1000         .opts = {
1001                 .sample_time         = true,
1002                 .mmap_pages          = UINT_MAX,
1003                 .user_freq           = UINT_MAX,
1004                 .user_interval       = ULLONG_MAX,
1005                 .freq                = 4000,
1006                 .target              = {
1007                         .uses_mmap   = true,
1008                         .default_per_cpu = true,
1009                 },
1010                 .proc_map_timeout     = 500,
1011         },
1012         .tool = {
1013                 .sample         = process_sample_event,
1014                 .fork           = perf_event__process_fork,
1015                 .exit           = perf_event__process_exit,
1016                 .comm           = perf_event__process_comm,
1017                 .mmap           = perf_event__process_mmap,
1018                 .mmap2          = perf_event__process_mmap2,
1019                 .ordered_events = true,
1020         },
1021 };
1022
1023 const char record_callchain_help[] = CALLCHAIN_RECORD_HELP
1024         "\n\t\t\t\tDefault: fp";
1025
1026 /*
1027  * XXX Will stay a global variable till we fix builtin-script.c to stop messing
1028  * with it and switch to use the library functions in perf_evlist that came
1029  * from builtin-record.c, i.e. use record_opts,
1030  * perf_evlist__prepare_workload, etc instead of fork+exec'in 'perf record',
1031  * using pipes, etc.
1032  */
1033 struct option __record_options[] = {
1034         OPT_CALLBACK('e', "event", &record.evlist, "event",
1035                      "event selector. use 'perf list' to list available events",
1036                      parse_events_option),
1037         OPT_CALLBACK(0, "filter", &record.evlist, "filter",
1038                      "event filter", parse_filter),
1039         OPT_CALLBACK_NOOPT(0, "exclude-perf", &record.evlist,
1040                            NULL, "don't record events from perf itself",
1041                            exclude_perf),
1042         OPT_STRING('p', "pid", &record.opts.target.pid, "pid",
1043                     "record events on existing process id"),
1044         OPT_STRING('t', "tid", &record.opts.target.tid, "tid",
1045                     "record events on existing thread id"),
1046         OPT_INTEGER('r', "realtime", &record.realtime_prio,
1047                     "collect data with this RT SCHED_FIFO priority"),
1048         OPT_BOOLEAN(0, "no-buffering", &record.opts.no_buffering,
1049                     "collect data without buffering"),
1050         OPT_BOOLEAN('R', "raw-samples", &record.opts.raw_samples,
1051                     "collect raw sample records from all opened counters"),
1052         OPT_BOOLEAN('a', "all-cpus", &record.opts.target.system_wide,
1053                             "system-wide collection from all CPUs"),
1054         OPT_STRING('C', "cpu", &record.opts.target.cpu_list, "cpu",
1055                     "list of cpus to monitor"),
1056         OPT_U64('c', "count", &record.opts.user_interval, "event period to sample"),
1057         OPT_STRING('o', "output", &record.file.path, "file",
1058                     "output file name"),
1059         OPT_BOOLEAN_SET('i', "no-inherit", &record.opts.no_inherit,
1060                         &record.opts.no_inherit_set,
1061                         "child tasks do not inherit counters"),
1062         OPT_UINTEGER('F', "freq", &record.opts.user_freq, "profile at this frequency"),
1063         OPT_CALLBACK('m', "mmap-pages", &record.opts, "pages[,pages]",
1064                      "number of mmap data pages and AUX area tracing mmap pages",
1065                      record__parse_mmap_pages),
1066         OPT_BOOLEAN(0, "group", &record.opts.group,
1067                     "put the counters into a counter group"),
1068         OPT_CALLBACK_NOOPT('g', NULL, &record.opts,
1069                            NULL, "enables call-graph recording" ,
1070                            &record_callchain_opt),
1071         OPT_CALLBACK(0, "call-graph", &record.opts,
1072                      "record_mode[,record_size]", record_callchain_help,
1073                      &record_parse_callchain_opt),
1074         OPT_INCR('v', "verbose", &verbose,
1075                     "be more verbose (show counter open errors, etc)"),
1076         OPT_BOOLEAN('q', "quiet", &quiet, "don't print any message"),
1077         OPT_BOOLEAN('s', "stat", &record.opts.inherit_stat,
1078                     "per thread counts"),
1079         OPT_BOOLEAN('d', "data", &record.opts.sample_address, "Record the sample addresses"),
1080         OPT_BOOLEAN_SET('T', "timestamp", &record.opts.sample_time,
1081                         &record.opts.sample_time_set,
1082                         "Record the sample timestamps"),
1083         OPT_BOOLEAN('P', "period", &record.opts.period, "Record the sample period"),
1084         OPT_BOOLEAN('n', "no-samples", &record.opts.no_samples,
1085                     "don't sample"),
1086         OPT_BOOLEAN('N', "no-buildid-cache", &record.no_buildid_cache,
1087                     "do not update the buildid cache"),
1088         OPT_BOOLEAN('B', "no-buildid", &record.no_buildid,
1089                     "do not collect buildids in perf.data"),
1090         OPT_CALLBACK('G', "cgroup", &record.evlist, "name",
1091                      "monitor event in cgroup name only",
1092                      parse_cgroups),
1093         OPT_UINTEGER('D', "delay", &record.opts.initial_delay,
1094                   "ms to wait before starting measurement after program start"),
1095         OPT_STRING('u', "uid", &record.opts.target.uid_str, "user",
1096                    "user to profile"),
1097
1098         OPT_CALLBACK_NOOPT('b', "branch-any", &record.opts.branch_stack,
1099                      "branch any", "sample any taken branches",
1100                      parse_branch_stack),
1101
1102         OPT_CALLBACK('j', "branch-filter", &record.opts.branch_stack,
1103                      "branch filter mask", "branch stack filter modes",
1104                      parse_branch_stack),
1105         OPT_BOOLEAN('W', "weight", &record.opts.sample_weight,
1106                     "sample by weight (on special events only)"),
1107         OPT_BOOLEAN(0, "transaction", &record.opts.sample_transaction,
1108                     "sample transaction flags (special events only)"),
1109         OPT_BOOLEAN(0, "per-thread", &record.opts.target.per_thread,
1110                     "use per-thread mmaps"),
1111         OPT_CALLBACK_OPTARG('I', "intr-regs", &record.opts.sample_intr_regs, NULL, "any register",
1112                     "sample selected machine registers on interrupt,"
1113                     " use -I ? to list register names", parse_regs),
1114         OPT_BOOLEAN(0, "running-time", &record.opts.running_time,
1115                     "Record running/enabled time of read (:S) events"),
1116         OPT_CALLBACK('k', "clockid", &record.opts,
1117         "clockid", "clockid to use for events, see clock_gettime()",
1118         parse_clockid),
1119         OPT_STRING_OPTARG('S', "snapshot", &record.opts.auxtrace_snapshot_opts,
1120                           "opts", "AUX area tracing Snapshot Mode", ""),
1121         OPT_UINTEGER(0, "proc-map-timeout", &record.opts.proc_map_timeout,
1122                         "per thread proc mmap processing timeout in ms"),
1123         OPT_BOOLEAN(0, "switch-events", &record.opts.record_switch_events,
1124                     "Record context switch events"),
1125 #ifdef HAVE_LIBBPF_SUPPORT
1126         OPT_STRING(0, "clang-path", &llvm_param.clang_path, "clang path",
1127                    "clang binary to use for compiling BPF scriptlets"),
1128         OPT_STRING(0, "clang-opt", &llvm_param.clang_opt, "clang options",
1129                    "options passed to clang when compiling BPF scriptlets"),
1130 #endif
1131         OPT_END()
1132 };
1133
1134 struct option *record_options = __record_options;
1135
1136 int cmd_record(int argc, const char **argv, const char *prefix __maybe_unused)
1137 {
1138         int err;
1139         struct record *rec = &record;
1140         char errbuf[BUFSIZ];
1141
1142         rec->evlist = perf_evlist__new();
1143         if (rec->evlist == NULL)
1144                 return -ENOMEM;
1145
1146         perf_config(perf_record_config, rec);
1147
1148         argc = parse_options(argc, argv, record_options, record_usage,
1149                             PARSE_OPT_STOP_AT_NON_OPTION);
1150         if (!argc && target__none(&rec->opts.target))
1151                 usage_with_options(record_usage, record_options);
1152
1153         if (nr_cgroups && !rec->opts.target.system_wide) {
1154                 usage_with_options_msg(record_usage, record_options,
1155                         "cgroup monitoring only available in system-wide mode");
1156
1157         }
1158         if (rec->opts.record_switch_events &&
1159             !perf_can_record_switch_events()) {
1160                 ui__error("kernel does not support recording context switch events\n");
1161                 parse_options_usage(record_usage, record_options, "switch-events", 0);
1162                 return -EINVAL;
1163         }
1164
1165         if (!rec->itr) {
1166                 rec->itr = auxtrace_record__init(rec->evlist, &err);
1167                 if (err)
1168                         return err;
1169         }
1170
1171         err = auxtrace_parse_snapshot_options(rec->itr, &rec->opts,
1172                                               rec->opts.auxtrace_snapshot_opts);
1173         if (err)
1174                 return err;
1175
1176         err = -ENOMEM;
1177
1178         symbol__init(NULL);
1179
1180         if (symbol_conf.kptr_restrict)
1181                 pr_warning(
1182 "WARNING: Kernel address maps (/proc/{kallsyms,modules}) are restricted,\n"
1183 "check /proc/sys/kernel/kptr_restrict.\n\n"
1184 "Samples in kernel functions may not be resolved if a suitable vmlinux\n"
1185 "file is not found in the buildid cache or in the vmlinux path.\n\n"
1186 "Samples in kernel modules won't be resolved at all.\n\n"
1187 "If some relocation was applied (e.g. kexec) symbols may be misresolved\n"
1188 "even with a suitable vmlinux or kallsyms file.\n\n");
1189
1190         if (rec->no_buildid_cache || rec->no_buildid)
1191                 disable_buildid_cache();
1192
1193         if (rec->evlist->nr_entries == 0 &&
1194             perf_evlist__add_default(rec->evlist) < 0) {
1195                 pr_err("Not enough memory for event selector list\n");
1196                 goto out_symbol_exit;
1197         }
1198
1199         if (rec->opts.target.tid && !rec->opts.no_inherit_set)
1200                 rec->opts.no_inherit = true;
1201
1202         err = target__validate(&rec->opts.target);
1203         if (err) {
1204                 target__strerror(&rec->opts.target, err, errbuf, BUFSIZ);
1205                 ui__warning("%s", errbuf);
1206         }
1207
1208         err = target__parse_uid(&rec->opts.target);
1209         if (err) {
1210                 int saved_errno = errno;
1211
1212                 target__strerror(&rec->opts.target, err, errbuf, BUFSIZ);
1213                 ui__error("%s", errbuf);
1214
1215                 err = -saved_errno;
1216                 goto out_symbol_exit;
1217         }
1218
1219         err = -ENOMEM;
1220         if (perf_evlist__create_maps(rec->evlist, &rec->opts.target) < 0)
1221                 usage_with_options(record_usage, record_options);
1222
1223         err = auxtrace_record__options(rec->itr, rec->evlist, &rec->opts);
1224         if (err)
1225                 goto out_symbol_exit;
1226
1227         if (record_opts__config(&rec->opts)) {
1228                 err = -EINVAL;
1229                 goto out_symbol_exit;
1230         }
1231
1232         err = __cmd_record(&record, argc, argv);
1233 out_symbol_exit:
1234         perf_evlist__delete(rec->evlist);
1235         symbol__exit();
1236         auxtrace_record__free(rec->itr);
1237         return err;
1238 }
1239
1240 static void snapshot_sig_handler(int sig __maybe_unused)
1241 {
1242         if (!auxtrace_snapshot_enabled)
1243                 return;
1244         auxtrace_snapshot_enabled = 0;
1245         auxtrace_snapshot_err = auxtrace_record__snapshot_start(record.itr);
1246         auxtrace_record__snapshot_started = 1;
1247 }