tools/perf/builtin-record.c

   1 /*
   2  * builtin-record.c
   3  *
   4  * Builtin record command: Record the profile of a workload
   5  * (or a CPU, or a PID) into the perf.data output file - for
   6  * later analysis via perf report.
   7  */
   8 #include "builtin.h"
   9
  10 #include "perf.h"
  11
  12 #include "util/build-id.h"
  13 #include "util/util.h"
  14 #include <subcmd/parse-options.h>
  15 #include "util/parse-events.h"
  16
  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 #include "util/bpf-loader.h"
  36
  37 #include <unistd.h>
  38 #include <sched.h>
  39 #include <sys/mman.h>
  40
  41
  42 struct record {
  43         struct perf_tool        tool;
  44         struct record_opts      opts;
  45         u64                     bytes_written;
  46         struct perf_data_file   file;
  47         struct auxtrace_record  *itr;
  48         struct perf_evlist      *evlist;
  49         struct perf_session     *session;
  50         const char              *progname;
  51         int                     realtime_prio;
  52         bool                    no_buildid;
  53         bool                    no_buildid_set;
  54         bool                    no_buildid_cache;
  55         bool                    no_buildid_cache_set;
  56         bool                    buildid_all;
  57         unsigned long long      samples;
  58 };
  59
  60 static int record__write(struct record *rec, void *bf, size_t size)
  61 {
  62         if (perf_data_file__write(rec->session->file, bf, size) < 0) {
  63                 pr_err("failed to write perf data, error: %m\n");
  64                 return -1;
  65         }
  66
  67         rec->bytes_written += size;
  68         return 0;
  69 }
  70
  71 static int process_synthesized_event(struct perf_tool *tool,
  72                                      union perf_event *event,
  73                                      struct perf_sample *sample __maybe_unused,
  74                                      struct machine *machine __maybe_unused)
  75 {
  76         struct record *rec = container_of(tool, struct record, tool);
  77         return record__write(rec, event, event->header.size);
  78 }
  79
  80 static int record__mmap_read(struct record *rec, int idx)
  81 {
  82         struct perf_mmap *md = &rec->evlist->mmap[idx];
  83         u64 head = perf_mmap__read_head(md);
  84         u64 old = md->prev;
  85         unsigned char *data = md->base + page_size;
  86         unsigned long size;
  87         void *buf;
  88         int rc = 0;
  89
  90         if (old == head)
  91                 return 0;
  92
  93         rec->samples++;
  94
  95         size = head - old;
  96
  97         if ((old & md->mask) + size != (head & md->mask)) {
  98                 buf = &data[old & md->mask];
  99                 size = md->mask + 1 - (old & md->mask);
 100                 old += size;
 101
 102                 if (record__write(rec, buf, size) < 0) {
 103                         rc = -1;
 104                         goto out;
 105                 }
 106         }
 107
 108         buf = &data[old & md->mask];
 109         size = head - old;
 110         old += size;
 111
 112         if (record__write(rec, buf, size) < 0) {
 113                 rc = -1;
 114                 goto out;
 115         }
 116
 117         md->prev = old;
 118         perf_evlist__mmap_consume(rec->evlist, idx);
 119 out:
 120         return rc;
 121 }
 122
 123 static volatile int done;
 124 static volatile int signr = -1;
 125 static volatile int child_finished;
 126 static volatile int auxtrace_snapshot_enabled;
 127 static volatile int auxtrace_snapshot_err;
 128 static volatile int auxtrace_record__snapshot_started;
 129
 130 static void sig_handler(int sig)
 131 {
 132         if (sig == SIGCHLD)
 133                 child_finished = 1;
 134         else
 135                 signr = sig;
 136
 137         done = 1;
 138 }
 139
 140 static void record__sig_exit(void)
 141 {
 142         if (signr == -1)
 143                 return;
 144
 145         signal(signr, SIG_DFL);
 146         raise(signr);
 147 }
 148
 149 #ifdef HAVE_AUXTRACE_SUPPORT
 150
 151 static int record__process_auxtrace(struct perf_tool *tool,
 152                                     union perf_event *event, void *data1,
 153                                     size_t len1, void *data2, size_t len2)
 154 {
 155         struct record *rec = container_of(tool, struct record, tool);
 156         struct perf_data_file *file = &rec->file;
 157         size_t padding;
 158         u8 pad[8] = {0};
 159
 160         if (!perf_data_file__is_pipe(file)) {
 161                 off_t file_offset;
 162                 int fd = perf_data_file__fd(file);
 163                 int err;
 164
 165                 file_offset = lseek(fd, 0, SEEK_CUR);
 166                 if (file_offset == -1)
 167                         return -1;
 168                 err = auxtrace_index__auxtrace_event(&rec->session->auxtrace_index,
 169                                                      event, file_offset);
 170                 if (err)
 171                         return err;
 172         }
 173
 174         /* event.auxtrace.size includes padding, see __auxtrace_mmap__read() */
 175         padding = (len1 + len2) & 7;
 176         if (padding)
 177                 padding = 8 - padding;
 178
 179         record__write(rec, event, event->header.size);
 180         record__write(rec, data1, len1);
 181         if (len2)
 182                 record__write(rec, data2, len2);
 183         record__write(rec, &pad, padding);
 184
 185         return 0;
 186 }
 187
 188 static int record__auxtrace_mmap_read(struct record *rec,
 189                                       struct auxtrace_mmap *mm)
 190 {
 191         int ret;
 192
 193         ret = auxtrace_mmap__read(mm, rec->itr, &rec->tool,
 194                                   record__process_auxtrace);
 195         if (ret < 0)
 196                 return ret;
 197
 198         if (ret)
 199                 rec->samples++;
 200
 201         return 0;
 202 }
 203
 204 static int record__auxtrace_mmap_read_snapshot(struct record *rec,
 205                                                struct auxtrace_mmap *mm)
 206 {
 207         int ret;
 208
 209         ret = auxtrace_mmap__read_snapshot(mm, rec->itr, &rec->tool,
 210                                            record__process_auxtrace,
 211                                            rec->opts.auxtrace_snapshot_size);
 212         if (ret < 0)
 213                 return ret;
 214
 215         if (ret)
 216                 rec->samples++;
 217
 218         return 0;
 219 }
 220
 221 static int record__auxtrace_read_snapshot_all(struct record *rec)
 222 {
 223         int i;
 224         int rc = 0;
 225
 226         for (i = 0; i < rec->evlist->nr_mmaps; i++) {
 227                 struct auxtrace_mmap *mm =
 228                                 &rec->evlist->mmap[i].auxtrace_mmap;
 229
 230                 if (!mm->base)
 231                         continue;
 232
 233                 if (record__auxtrace_mmap_read_snapshot(rec, mm) != 0) {
 234                         rc = -1;
 235                         goto out;
 236                 }
 237         }
 238 out:
 239         return rc;
 240 }
 241
 242 static void record__read_auxtrace_snapshot(struct record *rec)
 243 {
 244         pr_debug("Recording AUX area tracing snapshot\n");
 245         if (record__auxtrace_read_snapshot_all(rec) < 0) {
 246                 auxtrace_snapshot_err = -1;
 247         } else {
 248                 auxtrace_snapshot_err = auxtrace_record__snapshot_finish(rec->itr);
 249                 if (!auxtrace_snapshot_err)
 250                         auxtrace_snapshot_enabled = 1;
 251         }
 252 }
 253
 254 #else
 255
 256 static inline
 257 int record__auxtrace_mmap_read(struct record *rec __maybe_unused,
 258                                struct auxtrace_mmap *mm __maybe_unused)
 259 {
 260         return 0;
 261 }
 262
 263 static inline
 264 void record__read_auxtrace_snapshot(struct record *rec __maybe_unused)
 265 {
 266 }
 267
 268 static inline
 269 int auxtrace_record__snapshot_start(struct auxtrace_record *itr __maybe_unused)
 270 {
 271         return 0;
 272 }
 273
 274 #endif
 275
 276 static int record__open(struct record *rec)
 277 {
 278         char msg[512];
 279         struct perf_evsel *pos;
 280         struct perf_evlist *evlist = rec->evlist;
 281         struct perf_session *session = rec->session;
 282         struct record_opts *opts = &rec->opts;
 283         int rc = 0;
 284
 285         perf_evlist__config(evlist, opts);
 286
 287         evlist__for_each(evlist, pos) {
 288 try_again:
 289                 if (perf_evsel__open(pos, pos->cpus, pos->threads) < 0) {
 290                         if (perf_evsel__fallback(pos, errno, msg, sizeof(msg))) {
 291                                 if (verbose)
 292                                         ui__warning("%s\n", msg);
 293                                 goto try_again;
 294                         }
 295
 296                         rc = -errno;
 297                         perf_evsel__open_strerror(pos, &opts->target,
 298                                                   errno, msg, sizeof(msg));
 299                         ui__error("%s\n", msg);
 300                         goto out;
 301                 }
 302         }
 303
 304         if (perf_evlist__apply_filters(evlist, &pos)) {
 305                 error("failed to set filter \"%s\" on event %s with %d (%s)\n",
 306                         pos->filter, perf_evsel__name(pos), errno,
 307                         strerror_r(errno, msg, sizeof(msg)));
 308                 rc = -1;
 309                 goto out;
 310         }
 311
 312         if (perf_evlist__mmap_ex(evlist, opts->mmap_pages, false,
 313                                  opts->auxtrace_mmap_pages,
 314                                  opts->auxtrace_snapshot_mode) < 0) {
 315                 if (errno == EPERM) {
 316                         pr_err("Permission error mapping pages.\n"
 317                                "Consider increasing "
 318                                "/proc/sys/kernel/perf_event_mlock_kb,\n"
 319                                "or try again with a smaller value of -m/--mmap_pages.\n"
 320                                "(current value: %u,%u)\n",
 321                                opts->mmap_pages, opts->auxtrace_mmap_pages);
 322                         rc = -errno;
 323                 } else {
 324                         pr_err("failed to mmap with %d (%s)\n", errno,
 325                                 strerror_r(errno, msg, sizeof(msg)));
 326                         rc = -errno;
 327                 }
 328                 goto out;
 329         }
 330
 331         session->evlist = evlist;
 332         perf_session__set_id_hdr_size(session);
 333 out:
 334         return rc;
 335 }
 336
 337 static int process_sample_event(struct perf_tool *tool,
 338                                 union perf_event *event,
 339                                 struct perf_sample *sample,
 340                                 struct perf_evsel *evsel,
 341                                 struct machine *machine)
 342 {
 343         struct record *rec = container_of(tool, struct record, tool);
 344
 345         rec->samples++;
 346
 347         return build_id__mark_dso_hit(tool, event, sample, evsel, machine);
 348 }
 349
 350 static int process_buildids(struct record *rec)
 351 {
 352         struct perf_data_file *file  = &rec->file;
 353         struct perf_session *session = rec->session;
 354
 355         if (file->size == 0)
 356                 return 0;
 357
 358         /*
 359          * During this process, it'll load kernel map and replace the
 360          * dso->long_name to a real pathname it found.  In this case
 361          * we prefer the vmlinux path like
 362          *   /lib/modules/3.16.4/build/vmlinux
 363          *
 364          * rather than build-id path (in debug directory).
 365          *   $HOME/.debug/.build-id/f0/6e17aa50adf4d00b88925e03775de107611551
 366          */
 367         symbol_conf.ignore_vmlinux_buildid = true;
 368
 369         /*
 370          * If --buildid-all is given, it marks all DSO regardless of hits,
 371          * so no need to process samples.
 372          */
 373         if (rec->buildid_all)
 374                 rec->tool.sample = NULL;
 375
 376         return perf_session__process_events(session);
 377 }
 378
 379 static void perf_event__synthesize_guest_os(struct machine *machine, void *data)
 380 {
 381         int err;
 382         struct perf_tool *tool = data;
 383         /*
 384          *As for guest kernel when processing subcommand record&report,
 385          *we arrange module mmap prior to guest kernel mmap and trigger
 386          *a preload dso because default guest module symbols are loaded
 387          *from guest kallsyms instead of /lib/modules/XXX/XXX. This
 388          *method is used to avoid symbol missing when the first addr is
 389          *in module instead of in guest kernel.
 390          */
 391         err = perf_event__synthesize_modules(tool, process_synthesized_event,
 392                                              machine);
 393         if (err < 0)
 394                 pr_err("Couldn't record guest kernel [%d]'s reference"
 395                        " relocation symbol.\n", machine->pid);
 396
 397         /*
 398          * We use _stext for guest kernel because guest kernel's /proc/kallsyms
 399          * have no _text sometimes.
 400          */
 401         err = perf_event__synthesize_kernel_mmap(tool, process_synthesized_event,
 402                                                  machine);
 403         if (err < 0)
 404                 pr_err("Couldn't record guest kernel [%d]'s reference"
 405                        " relocation symbol.\n", machine->pid);
 406 }
 407
 408 static struct perf_event_header finished_round_event = {
 409         .size = sizeof(struct perf_event_header),
 410         .type = PERF_RECORD_FINISHED_ROUND,
 411 };
 412
 413 static int record__mmap_read_all(struct record *rec)
 414 {
 415         u64 bytes_written = rec->bytes_written;
 416         int i;
 417         int rc = 0;
 418
 419         for (i = 0; i < rec->evlist->nr_mmaps; i++) {
 420                 struct auxtrace_mmap *mm = &rec->evlist->mmap[i].auxtrace_mmap;
 421
 422                 if (rec->evlist->mmap[i].base) {
 423                         if (record__mmap_read(rec, i) != 0) {
 424                                 rc = -1;
 425                                 goto out;
 426                         }
 427                 }
 428
 429                 if (mm->base && !rec->opts.auxtrace_snapshot_mode &&
 430                     record__auxtrace_mmap_read(rec, mm) != 0) {
 431                         rc = -1;
 432                         goto out;
 433                 }
 434         }
 435
 436         /*
 437          * Mark the round finished in case we wrote
 438          * at least one event.
 439          */
 440         if (bytes_written != rec->bytes_written)
 441                 rc = record__write(rec, &finished_round_event, sizeof(finished_round_event));
 442
 443 out:
 444         return rc;
 445 }
 446
 447 static void record__init_features(struct record *rec)
 448 {
 449         struct perf_session *session = rec->session;
 450         int feat;
 451
 452         for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++)
 453                 perf_header__set_feat(&session->header, feat);
 454
 455         if (rec->no_buildid)
 456                 perf_header__clear_feat(&session->header, HEADER_BUILD_ID);
 457
 458         if (!have_tracepoints(&rec->evlist->entries))
 459                 perf_header__clear_feat(&session->header, HEADER_TRACING_DATA);
 460
 461         if (!rec->opts.branch_stack)
 462                 perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK);
 463
 464         if (!rec->opts.full_auxtrace)
 465                 perf_header__clear_feat(&session->header, HEADER_AUXTRACE);
 466
 467         perf_header__clear_feat(&session->header, HEADER_STAT);
 468 }
 469
 470 static volatile int workload_exec_errno;
 471
 472 /*
 473  * perf_evlist__prepare_workload will send a SIGUSR1
 474  * if the fork fails, since we asked by setting its
 475  * want_signal to true.
 476  */
 477 static void workload_exec_failed_signal(int signo __maybe_unused,
 478                                         siginfo_t *info,
 479                                         void *ucontext __maybe_unused)
 480 {
 481         workload_exec_errno = info->si_value.sival_int;
 482         done = 1;
 483         child_finished = 1;
 484 }
 485
 486 static void snapshot_sig_handler(int sig);
 487
 488 static int __cmd_record(struct record *rec, int argc, const char **argv)
 489 {
 490         int err;
 491         int status = 0;
 492         unsigned long waking = 0;
 493         const bool forks = argc > 0;
 494         struct machine *machine;
 495         struct perf_tool *tool = &rec->tool;
 496         struct record_opts *opts = &rec->opts;
 497         struct perf_data_file *file = &rec->file;
 498         struct perf_session *session;
 499         bool disabled = false, draining = false;
 500         int fd;
 501
 502         rec->progname = argv[0];
 503
 504         atexit(record__sig_exit);
 505         signal(SIGCHLD, sig_handler);
 506         signal(SIGINT, sig_handler);
 507         signal(SIGTERM, sig_handler);
 508         if (rec->opts.auxtrace_snapshot_mode)
 509                 signal(SIGUSR2, snapshot_sig_handler);
 510         else
 511                 signal(SIGUSR2, SIG_IGN);
 512
 513         session = perf_session__new(file, false, tool);
 514         if (session == NULL) {
 515                 pr_err("Perf session creation failed.\n");
 516                 return -1;
 517         }
 518
 519         fd = perf_data_file__fd(file);
 520         rec->session = session;
 521
 522         record__init_features(rec);
 523
 524         if (forks) {
 525                 err = perf_evlist__prepare_workload(rec->evlist, &opts->target,
 526                                                     argv, file->is_pipe,
 527                                                     workload_exec_failed_signal);
 528                 if (err < 0) {
 529                         pr_err("Couldn't run the workload!\n");
 530                         status = err;
 531                         goto out_delete_session;
 532                 }
 533         }
 534
 535         if (record__open(rec) != 0) {
 536                 err = -1;
 537                 goto out_child;
 538         }
 539
 540         err = bpf__apply_obj_config();
 541         if (err) {
 542                 char errbuf[BUFSIZ];
 543
 544                 bpf__strerror_apply_obj_config(err, errbuf, sizeof(errbuf));
 545                 pr_err("ERROR: Apply config to BPF failed: %s\n",
 546                          errbuf);
 547                 goto out_child;
 548         }
 549
 550         /*
 551          * Normally perf_session__new would do this, but it doesn't have the
 552          * evlist.
 553          */
 554         if (rec->tool.ordered_events && !perf_evlist__sample_id_all(rec->evlist)) {
 555                 pr_warning("WARNING: No sample_id_all support, falling back to unordered processing\n");
 556                 rec->tool.ordered_events = false;
 557         }
 558
 559         if (!rec->evlist->nr_groups)
 560                 perf_header__clear_feat(&session->header, HEADER_GROUP_DESC);
 561
 562         if (file->is_pipe) {
 563                 err = perf_header__write_pipe(fd);
 564                 if (err < 0)
 565                         goto out_child;
 566         } else {
 567                 err = perf_session__write_header(session, rec->evlist, fd, false);
 568                 if (err < 0)
 569                         goto out_child;
 570         }
 571
 572         if (!rec->no_buildid
 573             && !perf_header__has_feat(&session->header, HEADER_BUILD_ID)) {
 574                 pr_err("Couldn't generate buildids. "
 575                        "Use --no-buildid to profile anyway.\n");
 576                 err = -1;
 577                 goto out_child;
 578         }
 579
 580         machine = &session->machines.host;
 581
 582         if (file->is_pipe) {
 583                 err = perf_event__synthesize_attrs(tool, session,
 584                                                    process_synthesized_event);
 585                 if (err < 0) {
 586                         pr_err("Couldn't synthesize attrs.\n");
 587                         goto out_child;
 588                 }
 589
 590                 if (have_tracepoints(&rec->evlist->entries)) {
 591                         /*
 592                          * FIXME err <= 0 here actually means that
 593                          * there were no tracepoints so its not really
 594                          * an error, just that we don't need to
 595                          * synthesize anything.  We really have to
 596                          * return this more properly and also
 597                          * propagate errors that now are calling die()
 598                          */
 599                         err = perf_event__synthesize_tracing_data(tool, fd, rec->evlist,
 600                                                                   process_synthesized_event);
 601                         if (err <= 0) {
 602                                 pr_err("Couldn't record tracing data.\n");
 603                                 goto out_child;
 604                         }
 605                         rec->bytes_written += err;
 606                 }
 607         }
 608
 609         if (rec->opts.full_auxtrace) {
 610                 err = perf_event__synthesize_auxtrace_info(rec->itr, tool,
 611                                         session, process_synthesized_event);
 612                 if (err)
 613                         goto out_delete_session;
 614         }
 615
 616         err = perf_event__synthesize_kernel_mmap(tool, process_synthesized_event,
 617                                                  machine);
 618         if (err < 0)
 619                 pr_err("Couldn't record kernel reference relocation symbol\n"
 620                        "Symbol resolution may be skewed if relocation was used (e.g. kexec).\n"
 621                        "Check /proc/kallsyms permission or run as root.\n");
 622
 623         err = perf_event__synthesize_modules(tool, process_synthesized_event,
 624                                              machine);
 625         if (err < 0)
 626                 pr_err("Couldn't record kernel module information.\n"
 627                        "Symbol resolution may be skewed if relocation was used (e.g. kexec).\n"
 628                        "Check /proc/modules permission or run as root.\n");
 629
 630         if (perf_guest) {
 631                 machines__process_guests(&session->machines,
 632                                          perf_event__synthesize_guest_os, tool);
 633         }
 634
 635         err = __machine__synthesize_threads(machine, tool, &opts->target, rec->evlist->threads,
 636                                             process_synthesized_event, opts->sample_address,
 637                                             opts->proc_map_timeout);
 638         if (err != 0)
 639                 goto out_child;
 640
 641         if (rec->realtime_prio) {
 642                 struct sched_param param;
 643
 644                 param.sched_priority = rec->realtime_prio;
 645                 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
 646                         pr_err("Could not set realtime priority.\n");
 647                         err = -1;
 648                         goto out_child;
 649                 }
 650         }
 651
 652         /*
 653          * When perf is starting the traced process, all the events
 654          * (apart from group members) have enable_on_exec=1 set,
 655          * so don't spoil it by prematurely enabling them.
 656          */
 657         if (!target__none(&opts->target) && !opts->initial_delay)
 658                 perf_evlist__enable(rec->evlist);
 659
 660         /*
 661          * Let the child rip
 662          */
 663         if (forks) {
 664                 union perf_event *event;
 665
 666                 event = malloc(sizeof(event->comm) + machine->id_hdr_size);
 667                 if (event == NULL) {
 668                         err = -ENOMEM;
 669                         goto out_child;
 670                 }
 671
 672                 /*
 673                  * Some H/W events are generated before COMM event
 674                  * which is emitted during exec(), so perf script
 675                  * cannot see a correct process name for those events.
 676                  * Synthesize COMM event to prevent it.
 677                  */
 678                 perf_event__synthesize_comm(tool, event,
 679                                             rec->evlist->workload.pid,
 680                                             process_synthesized_event,
 681                                             machine);
 682                 free(event);
 683
 684                 perf_evlist__start_workload(rec->evlist);
 685         }
 686
 687         if (opts->initial_delay) {
 688                 usleep(opts->initial_delay * 1000);
 689                 perf_evlist__enable(rec->evlist);
 690         }
 691
 692         auxtrace_snapshot_enabled = 1;
 693         for (;;) {
 694                 unsigned long long hits = rec->samples;
 695
 696                 if (record__mmap_read_all(rec) < 0) {
 697                         auxtrace_snapshot_enabled = 0;
 698                         err = -1;
 699                         goto out_child;
 700                 }
 701
 702                 if (auxtrace_record__snapshot_started) {
 703                         auxtrace_record__snapshot_started = 0;
 704                         if (!auxtrace_snapshot_err)
 705                                 record__read_auxtrace_snapshot(rec);
 706                         if (auxtrace_snapshot_err) {
 707                                 pr_err("AUX area tracing snapshot failed\n");
 708                                 err = -1;
 709                                 goto out_child;
 710                         }
 711                 }
 712
 713                 if (hits == rec->samples) {
 714                         if (done || draining)
 715                                 break;
 716                         err = perf_evlist__poll(rec->evlist, -1);
 717                         /*
 718                          * Propagate error, only if there's any. Ignore positive
 719                          * number of returned events and interrupt error.
 720                          */
 721                         if (err > 0 || (err < 0 && errno == EINTR))
 722                                 err = 0;
 723                         waking++;
 724
 725                         if (perf_evlist__filter_pollfd(rec->evlist, POLLERR | POLLHUP) == 0)
 726                                 draining = true;
 727                 }
 728
 729                 /*
 730                  * When perf is starting the traced process, at the end events
 731                  * die with the process and we wait for that. Thus no need to
 732                  * disable events in this case.
 733                  */
 734                 if (done && !disabled && !target__none(&opts->target)) {
 735                         auxtrace_snapshot_enabled = 0;
 736                         perf_evlist__disable(rec->evlist);
 737                         disabled = true;
 738                 }
 739         }
 740         auxtrace_snapshot_enabled = 0;
 741
 742         if (forks && workload_exec_errno) {
 743                 char msg[STRERR_BUFSIZE];
 744                 const char *emsg = strerror_r(workload_exec_errno, msg, sizeof(msg));
 745                 pr_err("Workload failed: %s\n", emsg);
 746                 err = -1;
 747                 goto out_child;
 748         }
 749
 750         if (!quiet)
 751                 fprintf(stderr, "[ perf record: Woken up %ld times to write data ]\n", waking);
 752
 753 out_child:
 754         if (forks) {
 755                 int exit_status;
 756
 757                 if (!child_finished)
 758                         kill(rec->evlist->workload.pid, SIGTERM);
 759
 760                 wait(&exit_status);
 761
 762                 if (err < 0)
 763                         status = err;
 764                 else if (WIFEXITED(exit_status))
 765                         status = WEXITSTATUS(exit_status);
 766                 else if (WIFSIGNALED(exit_status))
 767                         signr = WTERMSIG(exit_status);
 768         } else
 769                 status = err;
 770
 771         /* this will be recalculated during process_buildids() */
 772         rec->samples = 0;
 773
 774         if (!err && !file->is_pipe) {
 775                 rec->session->header.data_size += rec->bytes_written;
 776                 file->size = lseek(perf_data_file__fd(file), 0, SEEK_CUR);
 777
 778                 if (!rec->no_buildid) {
 779                         process_buildids(rec);
 780
 781                         if (rec->buildid_all)
 782                                 dsos__hit_all(rec->session);
 783                 }
 784                 perf_session__write_header(rec->session, rec->evlist, fd, true);
 785         }
 786
 787         if (!err && !quiet) {
 788                 char samples[128];
 789
 790                 if (rec->samples && !rec->opts.full_auxtrace)
 791                         scnprintf(samples, sizeof(samples),
 792                                   " (%" PRIu64 " samples)", rec->samples);
 793                 else
 794                         samples[0] = '\0';
 795
 796                 fprintf(stderr, "[ perf record: Captured and wrote %.3f MB %s%s ]\n",
 797                         perf_data_file__size(file) / 1024.0 / 1024.0,
 798                         file->path, samples);
 799         }
 800
 801 out_delete_session:
 802         perf_session__delete(session);
 803         return status;
 804 }
 805
 806 static void callchain_debug(void)
 807 {
 808         static const char *str[CALLCHAIN_MAX] = { "NONE", "FP", "DWARF", "LBR" };
 809
 810         pr_debug("callchain: type %s\n", str[callchain_param.record_mode]);
 811
 812         if (callchain_param.record_mode == CALLCHAIN_DWARF)
 813                 pr_debug("callchain: stack dump size %d\n",
 814                          callchain_param.dump_size);
 815 }
 816
 817 int record_parse_callchain_opt(const struct option *opt,
 818                                const char *arg,
 819                                int unset)
 820 {
 821         int ret;
 822         struct record_opts *record = (struct record_opts *)opt->value;
 823
 824         record->callgraph_set = true;
 825         callchain_param.enabled = !unset;
 826
 827         /* --no-call-graph */
 828         if (unset) {
 829                 callchain_param.record_mode = CALLCHAIN_NONE;
 830                 pr_debug("callchain: disabled\n");
 831                 return 0;
 832         }
 833
 834         ret = parse_callchain_record_opt(arg, &callchain_param);
 835         if (!ret) {
 836                 /* Enable data address sampling for DWARF unwind. */
 837                 if (callchain_param.record_mode == CALLCHAIN_DWARF)
 838                         record->sample_address = true;
 839                 callchain_debug();
 840         }
 841
 842         return ret;
 843 }
 844
 845 int record_callchain_opt(const struct option *opt,
 846                          const char *arg __maybe_unused,
 847                          int unset __maybe_unused)
 848 {
 849         struct record_opts *record = (struct record_opts *)opt->value;
 850
 851         record->callgraph_set = true;
 852         callchain_param.enabled = true;
 853
 854         if (callchain_param.record_mode == CALLCHAIN_NONE)
 855                 callchain_param.record_mode = CALLCHAIN_FP;
 856
 857         callchain_debug();
 858         return 0;
 859 }
 860
 861 static int perf_record_config(const char *var, const char *value, void *cb)
 862 {
 863         struct record *rec = cb;
 864
 865         if (!strcmp(var, "record.build-id")) {
 866                 if (!strcmp(value, "cache"))
 867                         rec->no_buildid_cache = false;
 868                 else if (!strcmp(value, "no-cache"))
 869                         rec->no_buildid_cache = true;
 870                 else if (!strcmp(value, "skip"))
 871                         rec->no_buildid = true;
 872                 else
 873                         return -1;
 874                 return 0;
 875         }
 876         if (!strcmp(var, "record.call-graph"))
 877                 var = "call-graph.record-mode"; /* fall-through */
 878
 879         return perf_default_config(var, value, cb);
 880 }
 881
 882 struct clockid_map {
 883         const char *name;
 884         int clockid;
 885 };
 886
 887 #define CLOCKID_MAP(n, c)       \
 888         { .name = n, .clockid = (c), }
 889
 890 #define CLOCKID_END     { .name = NULL, }
 891
 892
 893 /*
 894  * Add the missing ones, we need to build on many distros...
 895  */
 896 #ifndef CLOCK_MONOTONIC_RAW
 897 #define CLOCK_MONOTONIC_RAW 4
 898 #endif
 899 #ifndef CLOCK_BOOTTIME
 900 #define CLOCK_BOOTTIME 7
 901 #endif
 902 #ifndef CLOCK_TAI
 903 #define CLOCK_TAI 11
 904 #endif
 905
 906 static const struct clockid_map clockids[] = {
 907         /* available for all events, NMI safe */
 908         CLOCKID_MAP("monotonic", CLOCK_MONOTONIC),
 909         CLOCKID_MAP("monotonic_raw", CLOCK_MONOTONIC_RAW),
 910
 911         /* available for some events */
 912         CLOCKID_MAP("realtime", CLOCK_REALTIME),
 913         CLOCKID_MAP("boottime", CLOCK_BOOTTIME),
 914         CLOCKID_MAP("tai", CLOCK_TAI),
 915
 916         /* available for the lazy */
 917         CLOCKID_MAP("mono", CLOCK_MONOTONIC),
 918         CLOCKID_MAP("raw", CLOCK_MONOTONIC_RAW),
 919         CLOCKID_MAP("real", CLOCK_REALTIME),
 920         CLOCKID_MAP("boot", CLOCK_BOOTTIME),
 921
 922         CLOCKID_END,
 923 };
 924
 925 static int parse_clockid(const struct option *opt, const char *str, int unset)
 926 {
 927         struct record_opts *opts = (struct record_opts *)opt->value;
 928         const struct clockid_map *cm;
 929         const char *ostr = str;
 930
 931         if (unset) {
 932                 opts->use_clockid = 0;
 933                 return 0;
 934         }
 935
 936         /* no arg passed */
 937         if (!str)
 938                 return 0;
 939
 940         /* no setting it twice */
 941         if (opts->use_clockid)
 942                 return -1;
 943
 944         opts->use_clockid = true;
 945
 946         /* if its a number, we're done */
 947         if (sscanf(str, "%d", &opts->clockid) == 1)
 948                 return 0;
 949
 950         /* allow a "CLOCK_" prefix to the name */
 951         if (!strncasecmp(str, "CLOCK_", 6))
 952                 str += 6;
 953
 954         for (cm = clockids; cm->name; cm++) {
 955                 if (!strcasecmp(str, cm->name)) {
 956                         opts->clockid = cm->clockid;
 957                         return 0;
 958                 }
 959         }
 960
 961         opts->use_clockid = false;
 962         ui__warning("unknown clockid %s, check man page\n", ostr);
 963         return -1;
 964 }
 965
 966 static int record__parse_mmap_pages(const struct option *opt,
 967                                     const char *str,
 968                                     int unset __maybe_unused)
 969 {
 970         struct record_opts *opts = opt->value;
 971         char *s, *p;
 972         unsigned int mmap_pages;
 973         int ret;
 974
 975         if (!str)
 976                 return -EINVAL;
 977
 978         s = strdup(str);
 979         if (!s)
 980                 return -ENOMEM;
 981
 982         p = strchr(s, ',');
 983         if (p)
 984                 *p = '\0';
 985
 986         if (*s) {
 987                 ret = __perf_evlist__parse_mmap_pages(&mmap_pages, s);
 988                 if (ret)
 989                         goto out_free;
 990                 opts->mmap_pages = mmap_pages;
 991         }
 992
 993         if (!p) {
 994                 ret = 0;
 995                 goto out_free;
 996         }
 997
 998         ret = __perf_evlist__parse_mmap_pages(&mmap_pages, p + 1);
 999         if (ret)
1000                 goto out_free;
1001
1002         opts->auxtrace_mmap_pages = mmap_pages;
1003
1004 out_free:
1005         free(s);
1006         return ret;
1007 }
1008
1009 static const char * const __record_usage[] = {
1010         "perf record [<options>] [<command>]",
1011         "perf record [<options>] -- <command> [<options>]",
1012         NULL
1013 };
1014 const char * const *record_usage = __record_usage;
1015
1016 /*
1017  * XXX Ideally would be local to cmd_record() and passed to a record__new
1018  * because we need to have access to it in record__exit, that is called
1019  * after cmd_record() exits, but since record_options need to be accessible to
1020  * builtin-script, leave it here.
1021  *
1022  * At least we don't ouch it in all the other functions here directly.
1023  *
1024  * Just say no to tons of global variables, sigh.
1025  */
1026 static struct record record = {
1027         .opts = {
1028                 .sample_time         = true,
1029                 .mmap_pages          = UINT_MAX,
1030                 .user_freq           = UINT_MAX,
1031                 .user_interval       = ULLONG_MAX,
1032                 .freq                = 4000,
1033                 .target              = {
1034                         .uses_mmap   = true,
1035                         .default_per_cpu = true,
1036                 },
1037                 .proc_map_timeout     = 500,
1038         },
1039         .tool = {
1040                 .sample         = process_sample_event,
1041                 .fork           = perf_event__process_fork,
1042                 .exit           = perf_event__process_exit,
1043                 .comm           = perf_event__process_comm,
1044                 .mmap           = perf_event__process_mmap,
1045                 .mmap2          = perf_event__process_mmap2,
1046                 .ordered_events = true,
1047         },
1048 };
1049
1050 const char record_callchain_help[] = CALLCHAIN_RECORD_HELP
1051         "\n\t\t\t\tDefault: fp";
1052
1053 /*
1054  * XXX Will stay a global variable till we fix builtin-script.c to stop messing
1055  * with it and switch to use the library functions in perf_evlist that came
1056  * from builtin-record.c, i.e. use record_opts,
1057  * perf_evlist__prepare_workload, etc instead of fork+exec'in 'perf record',
1058  * using pipes, etc.
1059  */
1060 struct option __record_options[] = {
1061         OPT_CALLBACK('e', "event", &record.evlist, "event",
1062                      "event selector. use 'perf list' to list available events",
1063                      parse_events_option),
1064         OPT_CALLBACK(0, "filter", &record.evlist, "filter",
1065                      "event filter", parse_filter),
1066         OPT_CALLBACK_NOOPT(0, "exclude-perf", &record.evlist,
1067                            NULL, "don't record events from perf itself",
1068                            exclude_perf),
1069         OPT_STRING('p', "pid", &record.opts.target.pid, "pid",
1070                     "record events on existing process id"),
1071         OPT_STRING('t', "tid", &record.opts.target.tid, "tid",
1072                     "record events on existing thread id"),
1073         OPT_INTEGER('r', "realtime", &record.realtime_prio,
1074                     "collect data with this RT SCHED_FIFO priority"),
1075         OPT_BOOLEAN(0, "no-buffering", &record.opts.no_buffering,
1076                     "collect data without buffering"),
1077         OPT_BOOLEAN('R', "raw-samples", &record.opts.raw_samples,
1078                     "collect raw sample records from all opened counters"),
1079         OPT_BOOLEAN('a', "all-cpus", &record.opts.target.system_wide,
1080                             "system-wide collection from all CPUs"),
1081         OPT_STRING('C', "cpu", &record.opts.target.cpu_list, "cpu",
1082                     "list of cpus to monitor"),
1083         OPT_U64('c', "count", &record.opts.user_interval, "event period to sample"),
1084         OPT_STRING('o', "output", &record.file.path, "file",
1085                     "output file name"),
1086         OPT_BOOLEAN_SET('i', "no-inherit", &record.opts.no_inherit,
1087                         &record.opts.no_inherit_set,
1088                         "child tasks do not inherit counters"),
1089         OPT_UINTEGER('F', "freq", &record.opts.user_freq, "profile at this frequency"),
1090         OPT_CALLBACK('m', "mmap-pages", &record.opts, "pages[,pages]",
1091                      "number of mmap data pages and AUX area tracing mmap pages",
1092                      record__parse_mmap_pages),
1093         OPT_BOOLEAN(0, "group", &record.opts.group,
1094                     "put the counters into a counter group"),
1095         OPT_CALLBACK_NOOPT('g', NULL, &record.opts,
1096                            NULL, "enables call-graph recording" ,
1097                            &record_callchain_opt),
1098         OPT_CALLBACK(0, "call-graph", &record.opts,
1099                      "record_mode[,record_size]", record_callchain_help,
1100                      &record_parse_callchain_opt),
1101         OPT_INCR('v', "verbose", &verbose,
1102                     "be more verbose (show counter open errors, etc)"),
1103         OPT_BOOLEAN('q', "quiet", &quiet, "don't print any message"),
1104         OPT_BOOLEAN('s', "stat", &record.opts.inherit_stat,
1105                     "per thread counts"),
1106         OPT_BOOLEAN('d', "data", &record.opts.sample_address, "Record the sample addresses"),
1107         OPT_BOOLEAN_SET('T', "timestamp", &record.opts.sample_time,
1108                         &record.opts.sample_time_set,
1109                         "Record the sample timestamps"),
1110         OPT_BOOLEAN('P', "period", &record.opts.period, "Record the sample period"),
1111         OPT_BOOLEAN('n', "no-samples", &record.opts.no_samples,
1112                     "don't sample"),
1113         OPT_BOOLEAN_SET('N', "no-buildid-cache", &record.no_buildid_cache,
1114                         &record.no_buildid_cache_set,
1115                         "do not update the buildid cache"),
1116         OPT_BOOLEAN_SET('B', "no-buildid", &record.no_buildid,
1117                         &record.no_buildid_set,
1118                         "do not collect buildids in perf.data"),
1119         OPT_CALLBACK('G', "cgroup", &record.evlist, "name",
1120                      "monitor event in cgroup name only",
1121                      parse_cgroups),
1122         OPT_UINTEGER('D', "delay", &record.opts.initial_delay,
1123                   "ms to wait before starting measurement after program start"),
1124         OPT_STRING('u', "uid", &record.opts.target.uid_str, "user",
1125                    "user to profile"),
1126
1127         OPT_CALLBACK_NOOPT('b', "branch-any", &record.opts.branch_stack,
1128                      "branch any", "sample any taken branches",
1129                      parse_branch_stack),
1130
1131         OPT_CALLBACK('j', "branch-filter", &record.opts.branch_stack,
1132                      "branch filter mask", "branch stack filter modes",
1133                      parse_branch_stack),
1134         OPT_BOOLEAN('W', "weight", &record.opts.sample_weight,
1135                     "sample by weight (on special events only)"),
1136         OPT_BOOLEAN(0, "transaction", &record.opts.sample_transaction,
1137                     "sample transaction flags (special events only)"),
1138         OPT_BOOLEAN(0, "per-thread", &record.opts.target.per_thread,
1139                     "use per-thread mmaps"),
1140         OPT_CALLBACK_OPTARG('I', "intr-regs", &record.opts.sample_intr_regs, NULL, "any register",
1141                     "sample selected machine registers on interrupt,"
1142                     " use -I ? to list register names", parse_regs),
1143         OPT_BOOLEAN(0, "running-time", &record.opts.running_time,
1144                     "Record running/enabled time of read (:S) events"),
1145         OPT_CALLBACK('k', "clockid", &record.opts,
1146         "clockid", "clockid to use for events, see clock_gettime()",
1147         parse_clockid),
1148         OPT_STRING_OPTARG('S', "snapshot", &record.opts.auxtrace_snapshot_opts,
1149                           "opts", "AUX area tracing Snapshot Mode", ""),
1150         OPT_UINTEGER(0, "proc-map-timeout", &record.opts.proc_map_timeout,
1151                         "per thread proc mmap processing timeout in ms"),
1152         OPT_BOOLEAN(0, "switch-events", &record.opts.record_switch_events,
1153                     "Record context switch events"),
1154         OPT_BOOLEAN_FLAG(0, "all-kernel", &record.opts.all_kernel,
1155                          "Configure all used events to run in kernel space.",
1156                          PARSE_OPT_EXCLUSIVE),
1157         OPT_BOOLEAN_FLAG(0, "all-user", &record.opts.all_user,
1158                          "Configure all used events to run in user space.",
1159                          PARSE_OPT_EXCLUSIVE),
1160         OPT_STRING(0, "clang-path", &llvm_param.clang_path, "clang path",
1161                    "clang binary to use for compiling BPF scriptlets"),
1162         OPT_STRING(0, "clang-opt", &llvm_param.clang_opt, "clang options",
1163                    "options passed to clang when compiling BPF scriptlets"),
1164         OPT_STRING(0, "vmlinux", &symbol_conf.vmlinux_name,
1165                    "file", "vmlinux pathname"),
1166         OPT_BOOLEAN(0, "buildid-all", &record.buildid_all,
1167                     "Record build-id of all DSOs regardless of hits"),
1168         OPT_END()
1169 };
1170
1171 struct option *record_options = __record_options;
1172
1173 int cmd_record(int argc, const char **argv, const char *prefix __maybe_unused)
1174 {
1175         int err;
1176         struct record *rec = &record;
1177         char errbuf[BUFSIZ];
1178
1179 #ifndef HAVE_LIBBPF_SUPPORT
1180 # define set_nobuild(s, l, c) set_option_nobuild(record_options, s, l, "NO_LIBBPF=1", c)
1181         set_nobuild('\0', "clang-path", true);
1182         set_nobuild('\0', "clang-opt", true);
1183 # undef set_nobuild
1184 #endif
1185
1186 #ifndef HAVE_BPF_PROLOGUE
1187 # if !defined (HAVE_DWARF_SUPPORT)
1188 #  define REASON  "NO_DWARF=1"
1189 # elif !defined (HAVE_LIBBPF_SUPPORT)
1190 #  define REASON  "NO_LIBBPF=1"
1191 # else
1192 #  define REASON  "this architecture doesn't support BPF prologue"
1193 # endif
1194 # define set_nobuild(s, l, c) set_option_nobuild(record_options, s, l, REASON, c)
1195         set_nobuild('\0', "vmlinux", true);
1196 # undef set_nobuild
1197 # undef REASON
1198 #endif
1199
1200         rec->evlist = perf_evlist__new();
1201         if (rec->evlist == NULL)
1202                 return -ENOMEM;
1203
1204         perf_config(perf_record_config, rec);
1205
1206         argc = parse_options(argc, argv, record_options, record_usage,
1207                             PARSE_OPT_STOP_AT_NON_OPTION);
1208         if (!argc && target__none(&rec->opts.target))
1209                 usage_with_options(record_usage, record_options);
1210
1211         if (nr_cgroups && !rec->opts.target.system_wide) {
1212                 usage_with_options_msg(record_usage, record_options,
1213                         "cgroup monitoring only available in system-wide mode");
1214
1215         }
1216         if (rec->opts.record_switch_events &&
1217             !perf_can_record_switch_events()) {
1218                 ui__error("kernel does not support recording context switch events\n");
1219                 parse_options_usage(record_usage, record_options, "switch-events", 0);
1220                 return -EINVAL;
1221         }
1222
1223         if (!rec->itr) {
1224                 rec->itr = auxtrace_record__init(rec->evlist, &err);
1225                 if (err)
1226                         return err;
1227         }
1228
1229         err = auxtrace_parse_snapshot_options(rec->itr, &rec->opts,
1230                                               rec->opts.auxtrace_snapshot_opts);
1231         if (err)
1232                 return err;
1233
1234         err = -ENOMEM;
1235
1236         symbol__init(NULL);
1237
1238         if (symbol_conf.kptr_restrict)
1239                 pr_warning(
1240 "WARNING: Kernel address maps (/proc/{kallsyms,modules}) are restricted,\n"
1241 "check /proc/sys/kernel/kptr_restrict.\n\n"
1242 "Samples in kernel functions may not be resolved if a suitable vmlinux\n"
1243 "file is not found in the buildid cache or in the vmlinux path.\n\n"
1244 "Samples in kernel modules won't be resolved at all.\n\n"
1245 "If some relocation was applied (e.g. kexec) symbols may be misresolved\n"
1246 "even with a suitable vmlinux or kallsyms file.\n\n");
1247
1248         if (rec->no_buildid_cache || rec->no_buildid)
1249                 disable_buildid_cache();
1250
1251         if (rec->evlist->nr_entries == 0 &&
1252             perf_evlist__add_default(rec->evlist) < 0) {
1253                 pr_err("Not enough memory for event selector list\n");
1254                 goto out_symbol_exit;
1255         }
1256
1257         if (rec->opts.target.tid && !rec->opts.no_inherit_set)
1258                 rec->opts.no_inherit = true;
1259
1260         err = target__validate(&rec->opts.target);
1261         if (err) {
1262                 target__strerror(&rec->opts.target, err, errbuf, BUFSIZ);
1263                 ui__warning("%s", errbuf);
1264         }
1265
1266         err = target__parse_uid(&rec->opts.target);
1267         if (err) {
1268                 int saved_errno = errno;
1269
1270                 target__strerror(&rec->opts.target, err, errbuf, BUFSIZ);
1271                 ui__error("%s", errbuf);
1272
1273                 err = -saved_errno;
1274                 goto out_symbol_exit;
1275         }
1276
1277         err = -ENOMEM;
1278         if (perf_evlist__create_maps(rec->evlist, &rec->opts.target) < 0)
1279                 usage_with_options(record_usage, record_options);
1280
1281         err = auxtrace_record__options(rec->itr, rec->evlist, &rec->opts);
1282         if (err)
1283                 goto out_symbol_exit;
1284
1285         /*
1286          * We take all buildids when the file contains
1287          * AUX area tracing data because we do not decode the
1288          * trace because it would take too long.
1289          */
1290         if (rec->opts.full_auxtrace)
1291                 rec->buildid_all = true;
1292
1293         if (record_opts__config(&rec->opts)) {
1294                 err = -EINVAL;
1295                 goto out_symbol_exit;
1296         }
1297
1298         err = __cmd_record(&record, argc, argv);
1299 out_symbol_exit:
1300         perf_evlist__delete(rec->evlist);
1301         symbol__exit();
1302         auxtrace_record__free(rec->itr);
1303         return err;
1304 }
1305
1306 static void snapshot_sig_handler(int sig __maybe_unused)
1307 {
1308         if (!auxtrace_snapshot_enabled)
1309                 return;
1310         auxtrace_snapshot_enabled = 0;
1311         auxtrace_snapshot_err = auxtrace_record__snapshot_start(record.itr);
1312         auxtrace_record__snapshot_started = 1;
1313 }