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/auxtrace.h"
  31
  32 #include <unistd.h>
  33 #include <sched.h>
  34 #include <sys/mman.h>
  35
  36
  37 struct record {
  38         struct perf_tool        tool;
  39         struct record_opts      opts;
  40         u64                     bytes_written;
  41         struct perf_data_file   file;
  42         struct auxtrace_record  *itr;
  43         struct perf_evlist      *evlist;
  44         struct perf_session     *session;
  45         const char              *progname;
  46         int                     realtime_prio;
  47         bool                    no_buildid;
  48         bool                    no_buildid_cache;
  49         long                    samples;
  50 };
  51
  52 static int record__write(struct record *rec, void *bf, size_t size)
  53 {
  54         if (perf_data_file__write(rec->session->file, bf, size) < 0) {
  55                 pr_err("failed to write perf data, error: %m\n");
  56                 return -1;
  57         }
  58
  59         rec->bytes_written += size;
  60         return 0;
  61 }
  62
  63 static int process_synthesized_event(struct perf_tool *tool,
  64                                      union perf_event *event,
  65                                      struct perf_sample *sample __maybe_unused,
  66                                      struct machine *machine __maybe_unused)
  67 {
  68         struct record *rec = container_of(tool, struct record, tool);
  69         return record__write(rec, event, event->header.size);
  70 }
  71
  72 static int record__mmap_read(struct record *rec, int idx)
  73 {
  74         struct perf_mmap *md = &rec->evlist->mmap[idx];
  75         u64 head = perf_mmap__read_head(md);
  76         u64 old = md->prev;
  77         unsigned char *data = md->base + page_size;
  78         unsigned long size;
  79         void *buf;
  80         int rc = 0;
  81
  82         if (old == head)
  83                 return 0;
  84
  85         rec->samples++;
  86
  87         size = head - old;
  88
  89         if ((old & md->mask) + size != (head & md->mask)) {
  90                 buf = &data[old & md->mask];
  91                 size = md->mask + 1 - (old & md->mask);
  92                 old += size;
  93
  94                 if (record__write(rec, buf, size) < 0) {
  95                         rc = -1;
  96                         goto out;
  97                 }
  98         }
  99
 100         buf = &data[old & md->mask];
 101         size = head - old;
 102         old += size;
 103
 104         if (record__write(rec, buf, size) < 0) {
 105                 rc = -1;
 106                 goto out;
 107         }
 108
 109         md->prev = old;
 110         perf_evlist__mmap_consume(rec->evlist, idx);
 111 out:
 112         return rc;
 113 }
 114
 115 static int record__process_auxtrace(struct perf_tool *tool,
 116                                     union perf_event *event, void *data1,
 117                                     size_t len1, void *data2, size_t len2)
 118 {
 119         struct record *rec = container_of(tool, struct record, tool);
 120         struct perf_data_file *file = &rec->file;
 121         size_t padding;
 122         u8 pad[8] = {0};
 123
 124         if (!perf_data_file__is_pipe(file)) {
 125                 off_t file_offset;
 126                 int fd = perf_data_file__fd(file);
 127                 int err;
 128
 129                 file_offset = lseek(fd, 0, SEEK_CUR);
 130                 if (file_offset == -1)
 131                         return -1;
 132                 err = auxtrace_index__auxtrace_event(&rec->session->auxtrace_index,
 133                                                      event, file_offset);
 134                 if (err)
 135                         return err;
 136         }
 137
 138         /* event.auxtrace.size includes padding, see __auxtrace_mmap__read() */
 139         padding = (len1 + len2) & 7;
 140         if (padding)
 141                 padding = 8 - padding;
 142
 143         record__write(rec, event, event->header.size);
 144         record__write(rec, data1, len1);
 145         if (len2)
 146                 record__write(rec, data2, len2);
 147         record__write(rec, &pad, padding);
 148
 149         return 0;
 150 }
 151
 152 static int record__auxtrace_mmap_read(struct record *rec,
 153                                       struct auxtrace_mmap *mm)
 154 {
 155         int ret;
 156
 157         ret = auxtrace_mmap__read(mm, rec->itr, &rec->tool,
 158                                   record__process_auxtrace);
 159         if (ret < 0)
 160                 return ret;
 161
 162         if (ret)
 163                 rec->samples++;
 164
 165         return 0;
 166 }
 167
 168 static volatile int done = 0;
 169 static volatile int signr = -1;
 170 static volatile int child_finished = 0;
 171
 172 static void sig_handler(int sig)
 173 {
 174         if (sig == SIGCHLD)
 175                 child_finished = 1;
 176         else
 177                 signr = sig;
 178
 179         done = 1;
 180 }
 181
 182 static void record__sig_exit(void)
 183 {
 184         if (signr == -1)
 185                 return;
 186
 187         signal(signr, SIG_DFL);
 188         raise(signr);
 189 }
 190
 191 static int record__open(struct record *rec)
 192 {
 193         char msg[512];
 194         struct perf_evsel *pos;
 195         struct perf_evlist *evlist = rec->evlist;
 196         struct perf_session *session = rec->session;
 197         struct record_opts *opts = &rec->opts;
 198         int rc = 0;
 199
 200         perf_evlist__config(evlist, opts);
 201
 202         evlist__for_each(evlist, pos) {
 203 try_again:
 204                 if (perf_evsel__open(pos, evlist->cpus, evlist->threads) < 0) {
 205                         if (perf_evsel__fallback(pos, errno, msg, sizeof(msg))) {
 206                                 if (verbose)
 207                                         ui__warning("%s\n", msg);
 208                                 goto try_again;
 209                         }
 210
 211                         rc = -errno;
 212                         perf_evsel__open_strerror(pos, &opts->target,
 213                                                   errno, msg, sizeof(msg));
 214                         ui__error("%s\n", msg);
 215                         goto out;
 216                 }
 217         }
 218
 219         if (perf_evlist__apply_filters(evlist, &pos)) {
 220                 error("failed to set filter \"%s\" on event %s with %d (%s)\n",
 221                         pos->filter, perf_evsel__name(pos), errno,
 222                         strerror_r(errno, msg, sizeof(msg)));
 223                 rc = -1;
 224                 goto out;
 225         }
 226
 227         if (perf_evlist__mmap_ex(evlist, opts->mmap_pages, false,
 228                                  opts->auxtrace_mmap_pages, false) < 0) {
 229                 if (errno == EPERM) {
 230                         pr_err("Permission error mapping pages.\n"
 231                                "Consider increasing "
 232                                "/proc/sys/kernel/perf_event_mlock_kb,\n"
 233                                "or try again with a smaller value of -m/--mmap_pages.\n"
 234                                "(current value: %u,%u)\n",
 235                                opts->mmap_pages, opts->auxtrace_mmap_pages);
 236                         rc = -errno;
 237                 } else {
 238                         pr_err("failed to mmap with %d (%s)\n", errno,
 239                                 strerror_r(errno, msg, sizeof(msg)));
 240                         rc = -errno;
 241                 }
 242                 goto out;
 243         }
 244
 245         session->evlist = evlist;
 246         perf_session__set_id_hdr_size(session);
 247 out:
 248         return rc;
 249 }
 250
 251 static int process_sample_event(struct perf_tool *tool,
 252                                 union perf_event *event,
 253                                 struct perf_sample *sample,
 254                                 struct perf_evsel *evsel,
 255                                 struct machine *machine)
 256 {
 257         struct record *rec = container_of(tool, struct record, tool);
 258
 259         rec->samples++;
 260
 261         return build_id__mark_dso_hit(tool, event, sample, evsel, machine);
 262 }
 263
 264 static int process_buildids(struct record *rec)
 265 {
 266         struct perf_data_file *file  = &rec->file;
 267         struct perf_session *session = rec->session;
 268
 269         u64 size = lseek(perf_data_file__fd(file), 0, SEEK_CUR);
 270         if (size == 0)
 271                 return 0;
 272
 273         file->size = size;
 274
 275         /*
 276          * During this process, it'll load kernel map and replace the
 277          * dso->long_name to a real pathname it found.  In this case
 278          * we prefer the vmlinux path like
 279          *   /lib/modules/3.16.4/build/vmlinux
 280          *
 281          * rather than build-id path (in debug directory).
 282          *   $HOME/.debug/.build-id/f0/6e17aa50adf4d00b88925e03775de107611551
 283          */
 284         symbol_conf.ignore_vmlinux_buildid = true;
 285
 286         return perf_session__process_events(session);
 287 }
 288
 289 static void perf_event__synthesize_guest_os(struct machine *machine, void *data)
 290 {
 291         int err;
 292         struct perf_tool *tool = data;
 293         /*
 294          *As for guest kernel when processing subcommand record&report,
 295          *we arrange module mmap prior to guest kernel mmap and trigger
 296          *a preload dso because default guest module symbols are loaded
 297          *from guest kallsyms instead of /lib/modules/XXX/XXX. This
 298          *method is used to avoid symbol missing when the first addr is
 299          *in module instead of in guest kernel.
 300          */
 301         err = perf_event__synthesize_modules(tool, process_synthesized_event,
 302                                              machine);
 303         if (err < 0)
 304                 pr_err("Couldn't record guest kernel [%d]'s reference"
 305                        " relocation symbol.\n", machine->pid);
 306
 307         /*
 308          * We use _stext for guest kernel because guest kernel's /proc/kallsyms
 309          * have no _text sometimes.
 310          */
 311         err = perf_event__synthesize_kernel_mmap(tool, process_synthesized_event,
 312                                                  machine);
 313         if (err < 0)
 314                 pr_err("Couldn't record guest kernel [%d]'s reference"
 315                        " relocation symbol.\n", machine->pid);
 316 }
 317
 318 static struct perf_event_header finished_round_event = {
 319         .size = sizeof(struct perf_event_header),
 320         .type = PERF_RECORD_FINISHED_ROUND,
 321 };
 322
 323 static int record__mmap_read_all(struct record *rec)
 324 {
 325         u64 bytes_written = rec->bytes_written;
 326         int i;
 327         int rc = 0;
 328
 329         for (i = 0; i < rec->evlist->nr_mmaps; i++) {
 330                 struct auxtrace_mmap *mm = &rec->evlist->mmap[i].auxtrace_mmap;
 331
 332                 if (rec->evlist->mmap[i].base) {
 333                         if (record__mmap_read(rec, i) != 0) {
 334                                 rc = -1;
 335                                 goto out;
 336                         }
 337                 }
 338
 339                 if (mm->base &&
 340                     record__auxtrace_mmap_read(rec, mm) != 0) {
 341                         rc = -1;
 342                         goto out;
 343                 }
 344         }
 345
 346         /*
 347          * Mark the round finished in case we wrote
 348          * at least one event.
 349          */
 350         if (bytes_written != rec->bytes_written)
 351                 rc = record__write(rec, &finished_round_event, sizeof(finished_round_event));
 352
 353 out:
 354         return rc;
 355 }
 356
 357 static void record__init_features(struct record *rec)
 358 {
 359         struct perf_session *session = rec->session;
 360         int feat;
 361
 362         for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++)
 363                 perf_header__set_feat(&session->header, feat);
 364
 365         if (rec->no_buildid)
 366                 perf_header__clear_feat(&session->header, HEADER_BUILD_ID);
 367
 368         if (!have_tracepoints(&rec->evlist->entries))
 369                 perf_header__clear_feat(&session->header, HEADER_TRACING_DATA);
 370
 371         if (!rec->opts.branch_stack)
 372                 perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK);
 373
 374         if (!rec->opts.full_auxtrace)
 375                 perf_header__clear_feat(&session->header, HEADER_AUXTRACE);
 376 }
 377
 378 static volatile int workload_exec_errno;
 379
 380 /*
 381  * perf_evlist__prepare_workload will send a SIGUSR1
 382  * if the fork fails, since we asked by setting its
 383  * want_signal to true.
 384  */
 385 static void workload_exec_failed_signal(int signo __maybe_unused,
 386                                         siginfo_t *info,
 387                                         void *ucontext __maybe_unused)
 388 {
 389         workload_exec_errno = info->si_value.sival_int;
 390         done = 1;
 391         child_finished = 1;
 392 }
 393
 394 static int __cmd_record(struct record *rec, int argc, const char **argv)
 395 {
 396         int err;
 397         int status = 0;
 398         unsigned long waking = 0;
 399         const bool forks = argc > 0;
 400         struct machine *machine;
 401         struct perf_tool *tool = &rec->tool;
 402         struct record_opts *opts = &rec->opts;
 403         struct perf_data_file *file = &rec->file;
 404         struct perf_session *session;
 405         bool disabled = false, draining = false;
 406         int fd;
 407
 408         rec->progname = argv[0];
 409
 410         atexit(record__sig_exit);
 411         signal(SIGCHLD, sig_handler);
 412         signal(SIGINT, sig_handler);
 413         signal(SIGTERM, sig_handler);
 414
 415         session = perf_session__new(file, false, tool);
 416         if (session == NULL) {
 417                 pr_err("Perf session creation failed.\n");
 418                 return -1;
 419         }
 420
 421         fd = perf_data_file__fd(file);
 422         rec->session = session;
 423
 424         record__init_features(rec);
 425
 426         if (forks) {
 427                 err = perf_evlist__prepare_workload(rec->evlist, &opts->target,
 428                                                     argv, file->is_pipe,
 429                                                     workload_exec_failed_signal);
 430                 if (err < 0) {
 431                         pr_err("Couldn't run the workload!\n");
 432                         status = err;
 433                         goto out_delete_session;
 434                 }
 435         }
 436
 437         if (record__open(rec) != 0) {
 438                 err = -1;
 439                 goto out_child;
 440         }
 441
 442         if (!rec->evlist->nr_groups)
 443                 perf_header__clear_feat(&session->header, HEADER_GROUP_DESC);
 444
 445         if (file->is_pipe) {
 446                 err = perf_header__write_pipe(fd);
 447                 if (err < 0)
 448                         goto out_child;
 449         } else {
 450                 err = perf_session__write_header(session, rec->evlist, fd, false);
 451                 if (err < 0)
 452                         goto out_child;
 453         }
 454
 455         if (!rec->no_buildid
 456             && !perf_header__has_feat(&session->header, HEADER_BUILD_ID)) {
 457                 pr_err("Couldn't generate buildids. "
 458                        "Use --no-buildid to profile anyway.\n");
 459                 err = -1;
 460                 goto out_child;
 461         }
 462
 463         machine = &session->machines.host;
 464
 465         if (file->is_pipe) {
 466                 err = perf_event__synthesize_attrs(tool, session,
 467                                                    process_synthesized_event);
 468                 if (err < 0) {
 469                         pr_err("Couldn't synthesize attrs.\n");
 470                         goto out_child;
 471                 }
 472
 473                 if (have_tracepoints(&rec->evlist->entries)) {
 474                         /*
 475                          * FIXME err <= 0 here actually means that
 476                          * there were no tracepoints so its not really
 477                          * an error, just that we don't need to
 478                          * synthesize anything.  We really have to
 479                          * return this more properly and also
 480                          * propagate errors that now are calling die()
 481                          */
 482                         err = perf_event__synthesize_tracing_data(tool, fd, rec->evlist,
 483                                                                   process_synthesized_event);
 484                         if (err <= 0) {
 485                                 pr_err("Couldn't record tracing data.\n");
 486                                 goto out_child;
 487                         }
 488                         rec->bytes_written += err;
 489                 }
 490         }
 491
 492         if (rec->opts.full_auxtrace) {
 493                 err = perf_event__synthesize_auxtrace_info(rec->itr, tool,
 494                                         session, process_synthesized_event);
 495                 if (err)
 496                         goto out_delete_session;
 497         }
 498
 499         err = perf_event__synthesize_kernel_mmap(tool, process_synthesized_event,
 500                                                  machine);
 501         if (err < 0)
 502                 pr_err("Couldn't record kernel reference relocation symbol\n"
 503                        "Symbol resolution may be skewed if relocation was used (e.g. kexec).\n"
 504                        "Check /proc/kallsyms permission or run as root.\n");
 505
 506         err = perf_event__synthesize_modules(tool, process_synthesized_event,
 507                                              machine);
 508         if (err < 0)
 509                 pr_err("Couldn't record kernel module information.\n"
 510                        "Symbol resolution may be skewed if relocation was used (e.g. kexec).\n"
 511                        "Check /proc/modules permission or run as root.\n");
 512
 513         if (perf_guest) {
 514                 machines__process_guests(&session->machines,
 515                                          perf_event__synthesize_guest_os, tool);
 516         }
 517
 518         err = __machine__synthesize_threads(machine, tool, &opts->target, rec->evlist->threads,
 519                                             process_synthesized_event, opts->sample_address);
 520         if (err != 0)
 521                 goto out_child;
 522
 523         if (rec->realtime_prio) {
 524                 struct sched_param param;
 525
 526                 param.sched_priority = rec->realtime_prio;
 527                 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
 528                         pr_err("Could not set realtime priority.\n");
 529                         err = -1;
 530                         goto out_child;
 531                 }
 532         }
 533
 534         /*
 535          * When perf is starting the traced process, all the events
 536          * (apart from group members) have enable_on_exec=1 set,
 537          * so don't spoil it by prematurely enabling them.
 538          */
 539         if (!target__none(&opts->target) && !opts->initial_delay)
 540                 perf_evlist__enable(rec->evlist);
 541
 542         /*
 543          * Let the child rip
 544          */
 545         if (forks)
 546                 perf_evlist__start_workload(rec->evlist);
 547
 548         if (opts->initial_delay) {
 549                 usleep(opts->initial_delay * 1000);
 550                 perf_evlist__enable(rec->evlist);
 551         }
 552
 553         for (;;) {
 554                 int hits = rec->samples;
 555
 556                 if (record__mmap_read_all(rec) < 0) {
 557                         err = -1;
 558                         goto out_child;
 559                 }
 560
 561                 if (hits == rec->samples) {
 562                         if (done || draining)
 563                                 break;
 564                         err = perf_evlist__poll(rec->evlist, -1);
 565                         /*
 566                          * Propagate error, only if there's any. Ignore positive
 567                          * number of returned events and interrupt error.
 568                          */
 569                         if (err > 0 || (err < 0 && errno == EINTR))
 570                                 err = 0;
 571                         waking++;
 572
 573                         if (perf_evlist__filter_pollfd(rec->evlist, POLLERR | POLLHUP) == 0)
 574                                 draining = true;
 575                 }
 576
 577                 /*
 578                  * When perf is starting the traced process, at the end events
 579                  * die with the process and we wait for that. Thus no need to
 580                  * disable events in this case.
 581                  */
 582                 if (done && !disabled && !target__none(&opts->target)) {
 583                         perf_evlist__disable(rec->evlist);
 584                         disabled = true;
 585                 }
 586         }
 587
 588         if (forks && workload_exec_errno) {
 589                 char msg[STRERR_BUFSIZE];
 590                 const char *emsg = strerror_r(workload_exec_errno, msg, sizeof(msg));
 591                 pr_err("Workload failed: %s\n", emsg);
 592                 err = -1;
 593                 goto out_child;
 594         }
 595
 596         if (!quiet)
 597                 fprintf(stderr, "[ perf record: Woken up %ld times to write data ]\n", waking);
 598
 599 out_child:
 600         if (forks) {
 601                 int exit_status;
 602
 603                 if (!child_finished)
 604                         kill(rec->evlist->workload.pid, SIGTERM);
 605
 606                 wait(&exit_status);
 607
 608                 if (err < 0)
 609                         status = err;
 610                 else if (WIFEXITED(exit_status))
 611                         status = WEXITSTATUS(exit_status);
 612                 else if (WIFSIGNALED(exit_status))
 613                         signr = WTERMSIG(exit_status);
 614         } else
 615                 status = err;
 616
 617         /* this will be recalculated during process_buildids() */
 618         rec->samples = 0;
 619
 620         if (!err && !file->is_pipe) {
 621                 rec->session->header.data_size += rec->bytes_written;
 622
 623                 if (!rec->no_buildid)
 624                         process_buildids(rec);
 625                 perf_session__write_header(rec->session, rec->evlist, fd, true);
 626         }
 627
 628         if (!err && !quiet) {
 629                 char samples[128];
 630
 631                 if (rec->samples && !rec->opts.full_auxtrace)
 632                         scnprintf(samples, sizeof(samples),
 633                                   " (%" PRIu64 " samples)", rec->samples);
 634                 else
 635                         samples[0] = '\0';
 636
 637                 fprintf(stderr, "[ perf record: Captured and wrote %.3f MB %s%s ]\n",
 638                         perf_data_file__size(file) / 1024.0 / 1024.0,
 639                         file->path, samples);
 640         }
 641
 642 out_delete_session:
 643         perf_session__delete(session);
 644         return status;
 645 }
 646
 647 #define BRANCH_OPT(n, m) \
 648         { .name = n, .mode = (m) }
 649
 650 #define BRANCH_END { .name = NULL }
 651
 652 struct branch_mode {
 653         const char *name;
 654         int mode;
 655 };
 656
 657 static const struct branch_mode branch_modes[] = {
 658         BRANCH_OPT("u", PERF_SAMPLE_BRANCH_USER),
 659         BRANCH_OPT("k", PERF_SAMPLE_BRANCH_KERNEL),
 660         BRANCH_OPT("hv", PERF_SAMPLE_BRANCH_HV),
 661         BRANCH_OPT("any", PERF_SAMPLE_BRANCH_ANY),
 662         BRANCH_OPT("any_call", PERF_SAMPLE_BRANCH_ANY_CALL),
 663         BRANCH_OPT("any_ret", PERF_SAMPLE_BRANCH_ANY_RETURN),
 664         BRANCH_OPT("ind_call", PERF_SAMPLE_BRANCH_IND_CALL),
 665         BRANCH_OPT("abort_tx", PERF_SAMPLE_BRANCH_ABORT_TX),
 666         BRANCH_OPT("in_tx", PERF_SAMPLE_BRANCH_IN_TX),
 667         BRANCH_OPT("no_tx", PERF_SAMPLE_BRANCH_NO_TX),
 668         BRANCH_OPT("cond", PERF_SAMPLE_BRANCH_COND),
 669         BRANCH_END
 670 };
 671
 672 static int
 673 parse_branch_stack(const struct option *opt, const char *str, int unset)
 674 {
 675 #define ONLY_PLM \
 676         (PERF_SAMPLE_BRANCH_USER        |\
 677          PERF_SAMPLE_BRANCH_KERNEL      |\
 678          PERF_SAMPLE_BRANCH_HV)
 679
 680         uint64_t *mode = (uint64_t *)opt->value;
 681         const struct branch_mode *br;
 682         char *s, *os = NULL, *p;
 683         int ret = -1;
 684
 685         if (unset)
 686                 return 0;
 687
 688         /*
 689          * cannot set it twice, -b + --branch-filter for instance
 690          */
 691         if (*mode)
 692                 return -1;
 693
 694         /* str may be NULL in case no arg is passed to -b */
 695         if (str) {
 696                 /* because str is read-only */
 697                 s = os = strdup(str);
 698                 if (!s)
 699                         return -1;
 700
 701                 for (;;) {
 702                         p = strchr(s, ',');
 703                         if (p)
 704                                 *p = '\0';
 705
 706                         for (br = branch_modes; br->name; br++) {
 707                                 if (!strcasecmp(s, br->name))
 708                                         break;
 709                         }
 710                         if (!br->name) {
 711                                 ui__warning("unknown branch filter %s,"
 712                                             " check man page\n", s);
 713                                 goto error;
 714                         }
 715
 716                         *mode |= br->mode;
 717
 718                         if (!p)
 719                                 break;
 720
 721                         s = p + 1;
 722                 }
 723         }
 724         ret = 0;
 725
 726         /* default to any branch */
 727         if ((*mode & ~ONLY_PLM) == 0) {
 728                 *mode = PERF_SAMPLE_BRANCH_ANY;
 729         }
 730 error:
 731         free(os);
 732         return ret;
 733 }
 734
 735 static void callchain_debug(void)
 736 {
 737         static const char *str[CALLCHAIN_MAX] = { "NONE", "FP", "DWARF", "LBR" };
 738
 739         pr_debug("callchain: type %s\n", str[callchain_param.record_mode]);
 740
 741         if (callchain_param.record_mode == CALLCHAIN_DWARF)
 742                 pr_debug("callchain: stack dump size %d\n",
 743                          callchain_param.dump_size);
 744 }
 745
 746 int record_parse_callchain_opt(const struct option *opt __maybe_unused,
 747                                const char *arg,
 748                                int unset)
 749 {
 750         int ret;
 751
 752         callchain_param.enabled = !unset;
 753
 754         /* --no-call-graph */
 755         if (unset) {
 756                 callchain_param.record_mode = CALLCHAIN_NONE;
 757                 pr_debug("callchain: disabled\n");
 758                 return 0;
 759         }
 760
 761         ret = parse_callchain_record_opt(arg);
 762         if (!ret)
 763                 callchain_debug();
 764
 765         return ret;
 766 }
 767
 768 int record_callchain_opt(const struct option *opt __maybe_unused,
 769                          const char *arg __maybe_unused,
 770                          int unset __maybe_unused)
 771 {
 772         callchain_param.enabled = true;
 773
 774         if (callchain_param.record_mode == CALLCHAIN_NONE)
 775                 callchain_param.record_mode = CALLCHAIN_FP;
 776
 777         callchain_debug();
 778         return 0;
 779 }
 780
 781 static int perf_record_config(const char *var, const char *value, void *cb)
 782 {
 783         if (!strcmp(var, "record.call-graph"))
 784                 var = "call-graph.record-mode"; /* fall-through */
 785
 786         return perf_default_config(var, value, cb);
 787 }
 788
 789 struct clockid_map {
 790         const char *name;
 791         int clockid;
 792 };
 793
 794 #define CLOCKID_MAP(n, c)       \
 795         { .name = n, .clockid = (c), }
 796
 797 #define CLOCKID_END     { .name = NULL, }
 798
 799
 800 /*
 801  * Add the missing ones, we need to build on many distros...
 802  */
 803 #ifndef CLOCK_MONOTONIC_RAW
 804 #define CLOCK_MONOTONIC_RAW 4
 805 #endif
 806 #ifndef CLOCK_BOOTTIME
 807 #define CLOCK_BOOTTIME 7
 808 #endif
 809 #ifndef CLOCK_TAI
 810 #define CLOCK_TAI 11
 811 #endif
 812
 813 static const struct clockid_map clockids[] = {
 814         /* available for all events, NMI safe */
 815         CLOCKID_MAP("monotonic", CLOCK_MONOTONIC),
 816         CLOCKID_MAP("monotonic_raw", CLOCK_MONOTONIC_RAW),
 817
 818         /* available for some events */
 819         CLOCKID_MAP("realtime", CLOCK_REALTIME),
 820         CLOCKID_MAP("boottime", CLOCK_BOOTTIME),
 821         CLOCKID_MAP("tai", CLOCK_TAI),
 822
 823         /* available for the lazy */
 824         CLOCKID_MAP("mono", CLOCK_MONOTONIC),
 825         CLOCKID_MAP("raw", CLOCK_MONOTONIC_RAW),
 826         CLOCKID_MAP("real", CLOCK_REALTIME),
 827         CLOCKID_MAP("boot", CLOCK_BOOTTIME),
 828
 829         CLOCKID_END,
 830 };
 831
 832 static int parse_clockid(const struct option *opt, const char *str, int unset)
 833 {
 834         struct record_opts *opts = (struct record_opts *)opt->value;
 835         const struct clockid_map *cm;
 836         const char *ostr = str;
 837
 838         if (unset) {
 839                 opts->use_clockid = 0;
 840                 return 0;
 841         }
 842
 843         /* no arg passed */
 844         if (!str)
 845                 return 0;
 846
 847         /* no setting it twice */
 848         if (opts->use_clockid)
 849                 return -1;
 850
 851         opts->use_clockid = true;
 852
 853         /* if its a number, we're done */
 854         if (sscanf(str, "%d", &opts->clockid) == 1)
 855                 return 0;
 856
 857         /* allow a "CLOCK_" prefix to the name */
 858         if (!strncasecmp(str, "CLOCK_", 6))
 859                 str += 6;
 860
 861         for (cm = clockids; cm->name; cm++) {
 862                 if (!strcasecmp(str, cm->name)) {
 863                         opts->clockid = cm->clockid;
 864                         return 0;
 865                 }
 866         }
 867
 868         opts->use_clockid = false;
 869         ui__warning("unknown clockid %s, check man page\n", ostr);
 870         return -1;
 871 }
 872
 873 static int record__parse_mmap_pages(const struct option *opt,
 874                                     const char *str,
 875                                     int unset __maybe_unused)
 876 {
 877         struct record_opts *opts = opt->value;
 878         char *s, *p;
 879         unsigned int mmap_pages;
 880         int ret;
 881
 882         if (!str)
 883                 return -EINVAL;
 884
 885         s = strdup(str);
 886         if (!s)
 887                 return -ENOMEM;
 888
 889         p = strchr(s, ',');
 890         if (p)
 891                 *p = '\0';
 892
 893         if (*s) {
 894                 ret = __perf_evlist__parse_mmap_pages(&mmap_pages, s);
 895                 if (ret)
 896                         goto out_free;
 897                 opts->mmap_pages = mmap_pages;
 898         }
 899
 900         if (!p) {
 901                 ret = 0;
 902                 goto out_free;
 903         }
 904
 905         ret = __perf_evlist__parse_mmap_pages(&mmap_pages, p + 1);
 906         if (ret)
 907                 goto out_free;
 908
 909         opts->auxtrace_mmap_pages = mmap_pages;
 910
 911 out_free:
 912         free(s);
 913         return ret;
 914 }
 915
 916 static const char * const __record_usage[] = {
 917         "perf record [<options>] [<command>]",
 918         "perf record [<options>] -- <command> [<options>]",
 919         NULL
 920 };
 921 const char * const *record_usage = __record_usage;
 922
 923 /*
 924  * XXX Ideally would be local to cmd_record() and passed to a record__new
 925  * because we need to have access to it in record__exit, that is called
 926  * after cmd_record() exits, but since record_options need to be accessible to
 927  * builtin-script, leave it here.
 928  *
 929  * At least we don't ouch it in all the other functions here directly.
 930  *
 931  * Just say no to tons of global variables, sigh.
 932  */
 933 static struct record record = {
 934         .opts = {
 935                 .sample_time         = true,
 936                 .mmap_pages          = UINT_MAX,
 937                 .user_freq           = UINT_MAX,
 938                 .user_interval       = ULLONG_MAX,
 939                 .freq                = 4000,
 940                 .target              = {
 941                         .uses_mmap   = true,
 942                         .default_per_cpu = true,
 943                 },
 944         },
 945         .tool = {
 946                 .sample         = process_sample_event,
 947                 .fork           = perf_event__process_fork,
 948                 .comm           = perf_event__process_comm,
 949                 .mmap           = perf_event__process_mmap,
 950                 .mmap2          = perf_event__process_mmap2,
 951         },
 952 };
 953
 954 #define CALLCHAIN_HELP "setup and enables call-graph (stack chain/backtrace) recording: "
 955
 956 #ifdef HAVE_DWARF_UNWIND_SUPPORT
 957 const char record_callchain_help[] = CALLCHAIN_HELP "fp dwarf lbr";
 958 #else
 959 const char record_callchain_help[] = CALLCHAIN_HELP "fp lbr";
 960 #endif
 961
 962 /*
 963  * XXX Will stay a global variable till we fix builtin-script.c to stop messing
 964  * with it and switch to use the library functions in perf_evlist that came
 965  * from builtin-record.c, i.e. use record_opts,
 966  * perf_evlist__prepare_workload, etc instead of fork+exec'in 'perf record',
 967  * using pipes, etc.
 968  */
 969 struct option __record_options[] = {
 970         OPT_CALLBACK('e', "event", &record.evlist, "event",
 971                      "event selector. use 'perf list' to list available events",
 972                      parse_events_option),
 973         OPT_CALLBACK(0, "filter", &record.evlist, "filter",
 974                      "event filter", parse_filter),
 975         OPT_STRING('p', "pid", &record.opts.target.pid, "pid",
 976                     "record events on existing process id"),
 977         OPT_STRING('t', "tid", &record.opts.target.tid, "tid",
 978                     "record events on existing thread id"),
 979         OPT_INTEGER('r', "realtime", &record.realtime_prio,
 980                     "collect data with this RT SCHED_FIFO priority"),
 981         OPT_BOOLEAN(0, "no-buffering", &record.opts.no_buffering,
 982                     "collect data without buffering"),
 983         OPT_BOOLEAN('R', "raw-samples", &record.opts.raw_samples,
 984                     "collect raw sample records from all opened counters"),
 985         OPT_BOOLEAN('a', "all-cpus", &record.opts.target.system_wide,
 986                             "system-wide collection from all CPUs"),
 987         OPT_STRING('C', "cpu", &record.opts.target.cpu_list, "cpu",
 988                     "list of cpus to monitor"),
 989         OPT_U64('c', "count", &record.opts.user_interval, "event period to sample"),
 990         OPT_STRING('o', "output", &record.file.path, "file",
 991                     "output file name"),
 992         OPT_BOOLEAN_SET('i', "no-inherit", &record.opts.no_inherit,
 993                         &record.opts.no_inherit_set,
 994                         "child tasks do not inherit counters"),
 995         OPT_UINTEGER('F', "freq", &record.opts.user_freq, "profile at this frequency"),
 996         OPT_CALLBACK('m', "mmap-pages", &record.opts, "pages[,pages]",
 997                      "number of mmap data pages and AUX area tracing mmap pages",
 998                      record__parse_mmap_pages),
 999         OPT_BOOLEAN(0, "group", &record.opts.group,
1000                     "put the counters into a counter group"),
1001         OPT_CALLBACK_NOOPT('g', NULL, &record.opts,
1002                            NULL, "enables call-graph recording" ,
1003                            &record_callchain_opt),
1004         OPT_CALLBACK(0, "call-graph", &record.opts,
1005                      "mode[,dump_size]", record_callchain_help,
1006                      &record_parse_callchain_opt),
1007         OPT_INCR('v', "verbose", &verbose,
1008                     "be more verbose (show counter open errors, etc)"),
1009         OPT_BOOLEAN('q', "quiet", &quiet, "don't print any message"),
1010         OPT_BOOLEAN('s', "stat", &record.opts.inherit_stat,
1011                     "per thread counts"),
1012         OPT_BOOLEAN('d', "data", &record.opts.sample_address,
1013                     "Sample addresses"),
1014         OPT_BOOLEAN('T', "timestamp", &record.opts.sample_time, "Sample timestamps"),
1015         OPT_BOOLEAN('P', "period", &record.opts.period, "Sample period"),
1016         OPT_BOOLEAN('n', "no-samples", &record.opts.no_samples,
1017                     "don't sample"),
1018         OPT_BOOLEAN('N', "no-buildid-cache", &record.no_buildid_cache,
1019                     "do not update the buildid cache"),
1020         OPT_BOOLEAN('B', "no-buildid", &record.no_buildid,
1021                     "do not collect buildids in perf.data"),
1022         OPT_CALLBACK('G', "cgroup", &record.evlist, "name",
1023                      "monitor event in cgroup name only",
1024                      parse_cgroups),
1025         OPT_UINTEGER('D', "delay", &record.opts.initial_delay,
1026                   "ms to wait before starting measurement after program start"),
1027         OPT_STRING('u', "uid", &record.opts.target.uid_str, "user",
1028                    "user to profile"),
1029
1030         OPT_CALLBACK_NOOPT('b', "branch-any", &record.opts.branch_stack,
1031                      "branch any", "sample any taken branches",
1032                      parse_branch_stack),
1033
1034         OPT_CALLBACK('j', "branch-filter", &record.opts.branch_stack,
1035                      "branch filter mask", "branch stack filter modes",
1036                      parse_branch_stack),
1037         OPT_BOOLEAN('W', "weight", &record.opts.sample_weight,
1038                     "sample by weight (on special events only)"),
1039         OPT_BOOLEAN(0, "transaction", &record.opts.sample_transaction,
1040                     "sample transaction flags (special events only)"),
1041         OPT_BOOLEAN(0, "per-thread", &record.opts.target.per_thread,
1042                     "use per-thread mmaps"),
1043         OPT_BOOLEAN('I', "intr-regs", &record.opts.sample_intr_regs,
1044                     "Sample machine registers on interrupt"),
1045         OPT_BOOLEAN(0, "running-time", &record.opts.running_time,
1046                     "Record running/enabled time of read (:S) events"),
1047         OPT_CALLBACK('k', "clockid", &record.opts,
1048         "clockid", "clockid to use for events, see clock_gettime()",
1049         parse_clockid),
1050         OPT_END()
1051 };
1052
1053 struct option *record_options = __record_options;
1054
1055 int cmd_record(int argc, const char **argv, const char *prefix __maybe_unused)
1056 {
1057         int err;
1058         struct record *rec = &record;
1059         char errbuf[BUFSIZ];
1060
1061         rec->evlist = perf_evlist__new();
1062         if (rec->evlist == NULL)
1063                 return -ENOMEM;
1064
1065         perf_config(perf_record_config, rec);
1066
1067         argc = parse_options(argc, argv, record_options, record_usage,
1068                             PARSE_OPT_STOP_AT_NON_OPTION);
1069         if (!argc && target__none(&rec->opts.target))
1070                 usage_with_options(record_usage, record_options);
1071
1072         if (nr_cgroups && !rec->opts.target.system_wide) {
1073                 ui__error("cgroup monitoring only available in"
1074                           " system-wide mode\n");
1075                 usage_with_options(record_usage, record_options);
1076         }
1077
1078         if (!rec->itr) {
1079                 rec->itr = auxtrace_record__init(rec->evlist, &err);
1080                 if (err)
1081                         return err;
1082         }
1083
1084         err = -ENOMEM;
1085
1086         symbol__init(NULL);
1087
1088         if (symbol_conf.kptr_restrict)
1089                 pr_warning(
1090 "WARNING: Kernel address maps (/proc/{kallsyms,modules}) are restricted,\n"
1091 "check /proc/sys/kernel/kptr_restrict.\n\n"
1092 "Samples in kernel functions may not be resolved if a suitable vmlinux\n"
1093 "file is not found in the buildid cache or in the vmlinux path.\n\n"
1094 "Samples in kernel modules won't be resolved at all.\n\n"
1095 "If some relocation was applied (e.g. kexec) symbols may be misresolved\n"
1096 "even with a suitable vmlinux or kallsyms file.\n\n");
1097
1098         if (rec->no_buildid_cache || rec->no_buildid)
1099                 disable_buildid_cache();
1100
1101         if (rec->evlist->nr_entries == 0 &&
1102             perf_evlist__add_default(rec->evlist) < 0) {
1103                 pr_err("Not enough memory for event selector list\n");
1104                 goto out_symbol_exit;
1105         }
1106
1107         if (rec->opts.target.tid && !rec->opts.no_inherit_set)
1108                 rec->opts.no_inherit = true;
1109
1110         err = target__validate(&rec->opts.target);
1111         if (err) {
1112                 target__strerror(&rec->opts.target, err, errbuf, BUFSIZ);
1113                 ui__warning("%s", errbuf);
1114         }
1115
1116         err = target__parse_uid(&rec->opts.target);
1117         if (err) {
1118                 int saved_errno = errno;
1119
1120                 target__strerror(&rec->opts.target, err, errbuf, BUFSIZ);
1121                 ui__error("%s", errbuf);
1122
1123                 err = -saved_errno;
1124                 goto out_symbol_exit;
1125         }
1126
1127         err = -ENOMEM;
1128         if (perf_evlist__create_maps(rec->evlist, &rec->opts.target) < 0)
1129                 usage_with_options(record_usage, record_options);
1130
1131         err = auxtrace_record__options(rec->itr, rec->evlist, &rec->opts);
1132         if (err)
1133                 goto out_symbol_exit;
1134
1135         if (record_opts__config(&rec->opts)) {
1136                 err = -EINVAL;
1137                 goto out_symbol_exit;
1138         }
1139
1140         err = __cmd_record(&record, argc, argv);
1141 out_symbol_exit:
1142         perf_evlist__delete(rec->evlist);
1143         symbol__exit();
1144         auxtrace_record__free(rec->itr);
1145         return err;
1146 }