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/header.h"
  18 #include "util/event.h"
  19 #include "util/evlist.h"
  20 #include "util/evsel.h"
  21 #include "util/debug.h"
  22 #include "util/session.h"
  23 #include "util/tool.h"
  24 #include "util/symbol.h"
  25 #include "util/cpumap.h"
  26 #include "util/thread_map.h"
  27 #include "util/data.h"
  28
  29 #include <unistd.h>
  30 #include <sched.h>
  31 #include <sys/mman.h>
  32
  33
  34 struct record {
  35         struct perf_tool        tool;
  36         struct record_opts      opts;
  37         u64                     bytes_written;
  38         struct perf_data_file   file;
  39         struct perf_evlist      *evlist;
  40         struct perf_session     *session;
  41         const char              *progname;
  42         int                     realtime_prio;
  43         bool                    no_buildid;
  44         bool                    no_buildid_cache;
  45         long                    samples;
  46 };
  47
  48 static int record__write(struct record *rec, void *bf, size_t size)
  49 {
  50         if (perf_data_file__write(rec->session->file, bf, size) < 0) {
  51                 pr_err("failed to write perf data, error: %m\n");
  52                 return -1;
  53         }
  54
  55         rec->bytes_written += size;
  56         return 0;
  57 }
  58
  59 static int process_synthesized_event(struct perf_tool *tool,
  60                                      union perf_event *event,
  61                                      struct perf_sample *sample __maybe_unused,
  62                                      struct machine *machine __maybe_unused)
  63 {
  64         struct record *rec = container_of(tool, struct record, tool);
  65         return record__write(rec, event, event->header.size);
  66 }
  67
  68 static int record__mmap_read(struct record *rec, struct perf_mmap *md)
  69 {
  70         unsigned int head = perf_mmap__read_head(md);
  71         unsigned int old = md->prev;
  72         unsigned char *data = md->base + page_size;
  73         unsigned long size;
  74         void *buf;
  75         int rc = 0;
  76
  77         if (old == head)
  78                 return 0;
  79
  80         rec->samples++;
  81
  82         size = head - old;
  83
  84         if ((old & md->mask) + size != (head & md->mask)) {
  85                 buf = &data[old & md->mask];
  86                 size = md->mask + 1 - (old & md->mask);
  87                 old += size;
  88
  89                 if (record__write(rec, buf, size) < 0) {
  90                         rc = -1;
  91                         goto out;
  92                 }
  93         }
  94
  95         buf = &data[old & md->mask];
  96         size = head - old;
  97         old += size;
  98
  99         if (record__write(rec, buf, size) < 0) {
 100                 rc = -1;
 101                 goto out;
 102         }
 103
 104         md->prev = old;
 105         perf_mmap__write_tail(md, old);
 106
 107 out:
 108         return rc;
 109 }
 110
 111 static volatile int done = 0;
 112 static volatile int signr = -1;
 113 static volatile int child_finished = 0;
 114
 115 static void sig_handler(int sig)
 116 {
 117         if (sig == SIGCHLD)
 118                 child_finished = 1;
 119         else
 120                 signr = sig;
 121
 122         done = 1;
 123 }
 124
 125 static void record__sig_exit(void)
 126 {
 127         if (signr == -1)
 128                 return;
 129
 130         signal(signr, SIG_DFL);
 131         raise(signr);
 132 }
 133
 134 static int record__open(struct record *rec)
 135 {
 136         char msg[512];
 137         struct perf_evsel *pos;
 138         struct perf_evlist *evlist = rec->evlist;
 139         struct perf_session *session = rec->session;
 140         struct record_opts *opts = &rec->opts;
 141         int rc = 0;
 142
 143         perf_evlist__config(evlist, opts);
 144
 145         evlist__for_each(evlist, pos) {
 146 try_again:
 147                 if (perf_evsel__open(pos, evlist->cpus, evlist->threads) < 0) {
 148                         if (perf_evsel__fallback(pos, errno, msg, sizeof(msg))) {
 149                                 if (verbose)
 150                                         ui__warning("%s\n", msg);
 151                                 goto try_again;
 152                         }
 153
 154                         rc = -errno;
 155                         perf_evsel__open_strerror(pos, &opts->target,
 156                                                   errno, msg, sizeof(msg));
 157                         ui__error("%s\n", msg);
 158                         goto out;
 159                 }
 160         }
 161
 162         if (perf_evlist__apply_filters(evlist)) {
 163                 error("failed to set filter with %d (%s)\n", errno,
 164                         strerror(errno));
 165                 rc = -1;
 166                 goto out;
 167         }
 168
 169         if (perf_evlist__mmap(evlist, opts->mmap_pages, false) < 0) {
 170                 if (errno == EPERM) {
 171                         pr_err("Permission error mapping pages.\n"
 172                                "Consider increasing "
 173                                "/proc/sys/kernel/perf_event_mlock_kb,\n"
 174                                "or try again with a smaller value of -m/--mmap_pages.\n"
 175                                "(current value: %u)\n", opts->mmap_pages);
 176                         rc = -errno;
 177                 } else {
 178                         pr_err("failed to mmap with %d (%s)\n", errno, strerror(errno));
 179                         rc = -errno;
 180                 }
 181                 goto out;
 182         }
 183
 184         session->evlist = evlist;
 185         perf_session__set_id_hdr_size(session);
 186 out:
 187         return rc;
 188 }
 189
 190 static int process_buildids(struct record *rec)
 191 {
 192         struct perf_data_file *file  = &rec->file;
 193         struct perf_session *session = rec->session;
 194         u64 start = session->header.data_offset;
 195
 196         u64 size = lseek(file->fd, 0, SEEK_CUR);
 197         if (size == 0)
 198                 return 0;
 199
 200         return __perf_session__process_events(session, start,
 201                                               size - start,
 202                                               size, &build_id__mark_dso_hit_ops);
 203 }
 204
 205 static void perf_event__synthesize_guest_os(struct machine *machine, void *data)
 206 {
 207         int err;
 208         struct perf_tool *tool = data;
 209         /*
 210          *As for guest kernel when processing subcommand record&report,
 211          *we arrange module mmap prior to guest kernel mmap and trigger
 212          *a preload dso because default guest module symbols are loaded
 213          *from guest kallsyms instead of /lib/modules/XXX/XXX. This
 214          *method is used to avoid symbol missing when the first addr is
 215          *in module instead of in guest kernel.
 216          */
 217         err = perf_event__synthesize_modules(tool, process_synthesized_event,
 218                                              machine);
 219         if (err < 0)
 220                 pr_err("Couldn't record guest kernel [%d]'s reference"
 221                        " relocation symbol.\n", machine->pid);
 222
 223         /*
 224          * We use _stext for guest kernel because guest kernel's /proc/kallsyms
 225          * have no _text sometimes.
 226          */
 227         err = perf_event__synthesize_kernel_mmap(tool, process_synthesized_event,
 228                                                  machine);
 229         if (err < 0)
 230                 pr_err("Couldn't record guest kernel [%d]'s reference"
 231                        " relocation symbol.\n", machine->pid);
 232 }
 233
 234 static struct perf_event_header finished_round_event = {
 235         .size = sizeof(struct perf_event_header),
 236         .type = PERF_RECORD_FINISHED_ROUND,
 237 };
 238
 239 static int record__mmap_read_all(struct record *rec)
 240 {
 241         int i;
 242         int rc = 0;
 243
 244         for (i = 0; i < rec->evlist->nr_mmaps; i++) {
 245                 if (rec->evlist->mmap[i].base) {
 246                         if (record__mmap_read(rec, &rec->evlist->mmap[i]) != 0) {
 247                                 rc = -1;
 248                                 goto out;
 249                         }
 250                 }
 251         }
 252
 253         if (perf_header__has_feat(&rec->session->header, HEADER_TRACING_DATA))
 254                 rc = record__write(rec, &finished_round_event, sizeof(finished_round_event));
 255
 256 out:
 257         return rc;
 258 }
 259
 260 static void record__init_features(struct record *rec)
 261 {
 262         struct perf_session *session = rec->session;
 263         int feat;
 264
 265         for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++)
 266                 perf_header__set_feat(&session->header, feat);
 267
 268         if (rec->no_buildid)
 269                 perf_header__clear_feat(&session->header, HEADER_BUILD_ID);
 270
 271         if (!have_tracepoints(&rec->evlist->entries))
 272                 perf_header__clear_feat(&session->header, HEADER_TRACING_DATA);
 273
 274         if (!rec->opts.branch_stack)
 275                 perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK);
 276 }
 277
 278 static volatile int workload_exec_errno;
 279
 280 /*
 281  * perf_evlist__prepare_workload will send a SIGUSR1
 282  * if the fork fails, since we asked by setting its
 283  * want_signal to true.
 284  */
 285 static void workload_exec_failed_signal(int signo __maybe_unused,
 286                                         siginfo_t *info,
 287                                         void *ucontext __maybe_unused)
 288 {
 289         workload_exec_errno = info->si_value.sival_int;
 290         done = 1;
 291         child_finished = 1;
 292 }
 293
 294 static int __cmd_record(struct record *rec, int argc, const char **argv)
 295 {
 296         int err;
 297         int status = 0;
 298         unsigned long waking = 0;
 299         const bool forks = argc > 0;
 300         struct machine *machine;
 301         struct perf_tool *tool = &rec->tool;
 302         struct record_opts *opts = &rec->opts;
 303         struct perf_data_file *file = &rec->file;
 304         struct perf_session *session;
 305         bool disabled = false;
 306
 307         rec->progname = argv[0];
 308
 309         atexit(record__sig_exit);
 310         signal(SIGCHLD, sig_handler);
 311         signal(SIGINT, sig_handler);
 312         signal(SIGTERM, sig_handler);
 313
 314         session = perf_session__new(file, false, NULL);
 315         if (session == NULL) {
 316                 pr_err("Perf session creation failed.\n");
 317                 return -1;
 318         }
 319
 320         rec->session = session;
 321
 322         record__init_features(rec);
 323
 324         if (forks) {
 325                 err = perf_evlist__prepare_workload(rec->evlist, &opts->target,
 326                                                     argv, file->is_pipe,
 327                                                     workload_exec_failed_signal);
 328                 if (err < 0) {
 329                         pr_err("Couldn't run the workload!\n");
 330                         status = err;
 331                         goto out_delete_session;
 332                 }
 333         }
 334
 335         if (record__open(rec) != 0) {
 336                 err = -1;
 337                 goto out_child;
 338         }
 339
 340         if (!rec->evlist->nr_groups)
 341                 perf_header__clear_feat(&session->header, HEADER_GROUP_DESC);
 342
 343         if (file->is_pipe) {
 344                 err = perf_header__write_pipe(file->fd);
 345                 if (err < 0)
 346                         goto out_child;
 347         } else {
 348                 err = perf_session__write_header(session, rec->evlist,
 349                                                  file->fd, false);
 350                 if (err < 0)
 351                         goto out_child;
 352         }
 353
 354         if (!rec->no_buildid
 355             && !perf_header__has_feat(&session->header, HEADER_BUILD_ID)) {
 356                 pr_err("Couldn't generate buildids. "
 357                        "Use --no-buildid to profile anyway.\n");
 358                 err = -1;
 359                 goto out_child;
 360         }
 361
 362         machine = &session->machines.host;
 363
 364         if (file->is_pipe) {
 365                 err = perf_event__synthesize_attrs(tool, session,
 366                                                    process_synthesized_event);
 367                 if (err < 0) {
 368                         pr_err("Couldn't synthesize attrs.\n");
 369                         goto out_child;
 370                 }
 371
 372                 if (have_tracepoints(&rec->evlist->entries)) {
 373                         /*
 374                          * FIXME err <= 0 here actually means that
 375                          * there were no tracepoints so its not really
 376                          * an error, just that we don't need to
 377                          * synthesize anything.  We really have to
 378                          * return this more properly and also
 379                          * propagate errors that now are calling die()
 380                          */
 381                         err = perf_event__synthesize_tracing_data(tool, file->fd, rec->evlist,
 382                                                                   process_synthesized_event);
 383                         if (err <= 0) {
 384                                 pr_err("Couldn't record tracing data.\n");
 385                                 goto out_child;
 386                         }
 387                         rec->bytes_written += err;
 388                 }
 389         }
 390
 391         err = perf_event__synthesize_kernel_mmap(tool, process_synthesized_event,
 392                                                  machine);
 393         if (err < 0)
 394                 pr_err("Couldn't record kernel reference relocation symbol\n"
 395                        "Symbol resolution may be skewed if relocation was used (e.g. kexec).\n"
 396                        "Check /proc/kallsyms permission or run as root.\n");
 397
 398         err = perf_event__synthesize_modules(tool, process_synthesized_event,
 399                                              machine);
 400         if (err < 0)
 401                 pr_err("Couldn't record kernel module information.\n"
 402                        "Symbol resolution may be skewed if relocation was used (e.g. kexec).\n"
 403                        "Check /proc/modules permission or run as root.\n");
 404
 405         if (perf_guest) {
 406                 machines__process_guests(&session->machines,
 407                                          perf_event__synthesize_guest_os, tool);
 408         }
 409
 410         err = __machine__synthesize_threads(machine, tool, &opts->target, rec->evlist->threads,
 411                                             process_synthesized_event, opts->sample_address);
 412         if (err != 0)
 413                 goto out_child;
 414
 415         if (rec->realtime_prio) {
 416                 struct sched_param param;
 417
 418                 param.sched_priority = rec->realtime_prio;
 419                 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
 420                         pr_err("Could not set realtime priority.\n");
 421                         err = -1;
 422                         goto out_child;
 423                 }
 424         }
 425
 426         /*
 427          * When perf is starting the traced process, all the events
 428          * (apart from group members) have enable_on_exec=1 set,
 429          * so don't spoil it by prematurely enabling them.
 430          */
 431         if (!target__none(&opts->target) && !opts->initial_delay)
 432                 perf_evlist__enable(rec->evlist);
 433
 434         /*
 435          * Let the child rip
 436          */
 437         if (forks)
 438                 perf_evlist__start_workload(rec->evlist);
 439
 440         if (opts->initial_delay) {
 441                 usleep(opts->initial_delay * 1000);
 442                 perf_evlist__enable(rec->evlist);
 443         }
 444
 445         for (;;) {
 446                 int hits = rec->samples;
 447
 448                 if (record__mmap_read_all(rec) < 0) {
 449                         err = -1;
 450                         goto out_child;
 451                 }
 452
 453                 if (hits == rec->samples) {
 454                         if (done)
 455                                 break;
 456                         err = poll(rec->evlist->pollfd, rec->evlist->nr_fds, -1);
 457                         /*
 458                          * Propagate error, only if there's any. Ignore positive
 459                          * number of returned events and interrupt error.
 460                          */
 461                         if (err > 0 || (err < 0 && errno == EINTR))
 462                                 err = 0;
 463                         waking++;
 464                 }
 465
 466                 /*
 467                  * When perf is starting the traced process, at the end events
 468                  * die with the process and we wait for that. Thus no need to
 469                  * disable events in this case.
 470                  */
 471                 if (done && !disabled && !target__none(&opts->target)) {
 472                         perf_evlist__disable(rec->evlist);
 473                         disabled = true;
 474                 }
 475         }
 476
 477         if (forks && workload_exec_errno) {
 478                 char msg[512];
 479                 const char *emsg = strerror_r(workload_exec_errno, msg, sizeof(msg));
 480                 pr_err("Workload failed: %s\n", emsg);
 481                 err = -1;
 482                 goto out_child;
 483         }
 484
 485         if (!quiet) {
 486                 fprintf(stderr, "[ perf record: Woken up %ld times to write data ]\n", waking);
 487
 488                 /*
 489                  * Approximate RIP event size: 24 bytes.
 490                  */
 491                 fprintf(stderr,
 492                         "[ perf record: Captured and wrote %.3f MB %s (~%" PRIu64 " samples) ]\n",
 493                         (double)rec->bytes_written / 1024.0 / 1024.0,
 494                         file->path,
 495                         rec->bytes_written / 24);
 496         }
 497
 498 out_child:
 499         if (forks) {
 500                 int exit_status;
 501
 502                 if (!child_finished)
 503                         kill(rec->evlist->workload.pid, SIGTERM);
 504
 505                 wait(&exit_status);
 506
 507                 if (err < 0)
 508                         status = err;
 509                 else if (WIFEXITED(exit_status))
 510                         status = WEXITSTATUS(exit_status);
 511                 else if (WIFSIGNALED(exit_status))
 512                         signr = WTERMSIG(exit_status);
 513         } else
 514                 status = err;
 515
 516         if (!err && !file->is_pipe) {
 517                 rec->session->header.data_size += rec->bytes_written;
 518
 519                 if (!rec->no_buildid)
 520                         process_buildids(rec);
 521                 perf_session__write_header(rec->session, rec->evlist,
 522                                            file->fd, true);
 523         }
 524
 525 out_delete_session:
 526         perf_session__delete(session);
 527         return status;
 528 }
 529
 530 #define BRANCH_OPT(n, m) \
 531         { .name = n, .mode = (m) }
 532
 533 #define BRANCH_END { .name = NULL }
 534
 535 struct branch_mode {
 536         const char *name;
 537         int mode;
 538 };
 539
 540 static const struct branch_mode branch_modes[] = {
 541         BRANCH_OPT("u", PERF_SAMPLE_BRANCH_USER),
 542         BRANCH_OPT("k", PERF_SAMPLE_BRANCH_KERNEL),
 543         BRANCH_OPT("hv", PERF_SAMPLE_BRANCH_HV),
 544         BRANCH_OPT("any", PERF_SAMPLE_BRANCH_ANY),
 545         BRANCH_OPT("any_call", PERF_SAMPLE_BRANCH_ANY_CALL),
 546         BRANCH_OPT("any_ret", PERF_SAMPLE_BRANCH_ANY_RETURN),
 547         BRANCH_OPT("ind_call", PERF_SAMPLE_BRANCH_IND_CALL),
 548         BRANCH_OPT("abort_tx", PERF_SAMPLE_BRANCH_ABORT_TX),
 549         BRANCH_OPT("in_tx", PERF_SAMPLE_BRANCH_IN_TX),
 550         BRANCH_OPT("no_tx", PERF_SAMPLE_BRANCH_NO_TX),
 551         BRANCH_OPT("cond", PERF_SAMPLE_BRANCH_COND),
 552         BRANCH_END
 553 };
 554
 555 static int
 556 parse_branch_stack(const struct option *opt, const char *str, int unset)
 557 {
 558 #define ONLY_PLM \
 559         (PERF_SAMPLE_BRANCH_USER        |\
 560          PERF_SAMPLE_BRANCH_KERNEL      |\
 561          PERF_SAMPLE_BRANCH_HV)
 562
 563         uint64_t *mode = (uint64_t *)opt->value;
 564         const struct branch_mode *br;
 565         char *s, *os = NULL, *p;
 566         int ret = -1;
 567
 568         if (unset)
 569                 return 0;
 570
 571         /*
 572          * cannot set it twice, -b + --branch-filter for instance
 573          */
 574         if (*mode)
 575                 return -1;
 576
 577         /* str may be NULL in case no arg is passed to -b */
 578         if (str) {
 579                 /* because str is read-only */
 580                 s = os = strdup(str);
 581                 if (!s)
 582                         return -1;
 583
 584                 for (;;) {
 585                         p = strchr(s, ',');
 586                         if (p)
 587                                 *p = '\0';
 588
 589                         for (br = branch_modes; br->name; br++) {
 590                                 if (!strcasecmp(s, br->name))
 591                                         break;
 592                         }
 593                         if (!br->name) {
 594                                 ui__warning("unknown branch filter %s,"
 595                                             " check man page\n", s);
 596                                 goto error;
 597                         }
 598
 599                         *mode |= br->mode;
 600
 601                         if (!p)
 602                                 break;
 603
 604                         s = p + 1;
 605                 }
 606         }
 607         ret = 0;
 608
 609         /* default to any branch */
 610         if ((*mode & ~ONLY_PLM) == 0) {
 611                 *mode = PERF_SAMPLE_BRANCH_ANY;
 612         }
 613 error:
 614         free(os);
 615         return ret;
 616 }
 617
 618 #ifdef HAVE_DWARF_UNWIND_SUPPORT
 619 static int get_stack_size(char *str, unsigned long *_size)
 620 {
 621         char *endptr;
 622         unsigned long size;
 623         unsigned long max_size = round_down(USHRT_MAX, sizeof(u64));
 624
 625         size = strtoul(str, &endptr, 0);
 626
 627         do {
 628                 if (*endptr)
 629                         break;
 630
 631                 size = round_up(size, sizeof(u64));
 632                 if (!size || size > max_size)
 633                         break;
 634
 635                 *_size = size;
 636                 return 0;
 637
 638         } while (0);
 639
 640         pr_err("callchain: Incorrect stack dump size (max %ld): %s\n",
 641                max_size, str);
 642         return -1;
 643 }
 644 #endif /* HAVE_DWARF_UNWIND_SUPPORT */
 645
 646 int record_parse_callchain(const char *arg, struct record_opts *opts)
 647 {
 648         char *tok, *name, *saveptr = NULL;
 649         char *buf;
 650         int ret = -1;
 651
 652         /* We need buffer that we know we can write to. */
 653         buf = malloc(strlen(arg) + 1);
 654         if (!buf)
 655                 return -ENOMEM;
 656
 657         strcpy(buf, arg);
 658
 659         tok = strtok_r((char *)buf, ",", &saveptr);
 660         name = tok ? : (char *)buf;
 661
 662         do {
 663                 /* Framepointer style */
 664                 if (!strncmp(name, "fp", sizeof("fp"))) {
 665                         if (!strtok_r(NULL, ",", &saveptr)) {
 666                                 opts->call_graph = CALLCHAIN_FP;
 667                                 ret = 0;
 668                         } else
 669                                 pr_err("callchain: No more arguments "
 670                                        "needed for -g fp\n");
 671                         break;
 672
 673 #ifdef HAVE_DWARF_UNWIND_SUPPORT
 674                 /* Dwarf style */
 675                 } else if (!strncmp(name, "dwarf", sizeof("dwarf"))) {
 676                         const unsigned long default_stack_dump_size = 8192;
 677
 678                         ret = 0;
 679                         opts->call_graph = CALLCHAIN_DWARF;
 680                         opts->stack_dump_size = default_stack_dump_size;
 681
 682                         tok = strtok_r(NULL, ",", &saveptr);
 683                         if (tok) {
 684                                 unsigned long size = 0;
 685
 686                                 ret = get_stack_size(tok, &size);
 687                                 opts->stack_dump_size = size;
 688                         }
 689 #endif /* HAVE_DWARF_UNWIND_SUPPORT */
 690                 } else {
 691                         pr_err("callchain: Unknown --call-graph option "
 692                                "value: %s\n", arg);
 693                         break;
 694                 }
 695
 696         } while (0);
 697
 698         free(buf);
 699         return ret;
 700 }
 701
 702 static void callchain_debug(struct record_opts *opts)
 703 {
 704         static const char *str[CALLCHAIN_MAX] = { "NONE", "FP", "DWARF" };
 705
 706         pr_debug("callchain: type %s\n", str[opts->call_graph]);
 707
 708         if (opts->call_graph == CALLCHAIN_DWARF)
 709                 pr_debug("callchain: stack dump size %d\n",
 710                          opts->stack_dump_size);
 711 }
 712
 713 int record_parse_callchain_opt(const struct option *opt,
 714                                const char *arg,
 715                                int unset)
 716 {
 717         struct record_opts *opts = opt->value;
 718         int ret;
 719
 720         opts->call_graph_enabled = !unset;
 721
 722         /* --no-call-graph */
 723         if (unset) {
 724                 opts->call_graph = CALLCHAIN_NONE;
 725                 pr_debug("callchain: disabled\n");
 726                 return 0;
 727         }
 728
 729         ret = record_parse_callchain(arg, opts);
 730         if (!ret)
 731                 callchain_debug(opts);
 732
 733         return ret;
 734 }
 735
 736 int record_callchain_opt(const struct option *opt,
 737                          const char *arg __maybe_unused,
 738                          int unset __maybe_unused)
 739 {
 740         struct record_opts *opts = opt->value;
 741
 742         opts->call_graph_enabled = !unset;
 743
 744         if (opts->call_graph == CALLCHAIN_NONE)
 745                 opts->call_graph = CALLCHAIN_FP;
 746
 747         callchain_debug(opts);
 748         return 0;
 749 }
 750
 751 static int perf_record_config(const char *var, const char *value, void *cb)
 752 {
 753         struct record *rec = cb;
 754
 755         if (!strcmp(var, "record.call-graph"))
 756                 return record_parse_callchain(value, &rec->opts);
 757
 758         return perf_default_config(var, value, cb);
 759 }
 760
 761 static const char * const record_usage[] = {
 762         "perf record [<options>] [<command>]",
 763         "perf record [<options>] -- <command> [<options>]",
 764         NULL
 765 };
 766
 767 /*
 768  * XXX Ideally would be local to cmd_record() and passed to a record__new
 769  * because we need to have access to it in record__exit, that is called
 770  * after cmd_record() exits, but since record_options need to be accessible to
 771  * builtin-script, leave it here.
 772  *
 773  * At least we don't ouch it in all the other functions here directly.
 774  *
 775  * Just say no to tons of global variables, sigh.
 776  */
 777 static struct record record = {
 778         .opts = {
 779                 .mmap_pages          = UINT_MAX,
 780                 .user_freq           = UINT_MAX,
 781                 .user_interval       = ULLONG_MAX,
 782                 .freq                = 4000,
 783                 .target              = {
 784                         .uses_mmap   = true,
 785                         .default_per_cpu = true,
 786                 },
 787         },
 788 };
 789
 790 #define CALLCHAIN_HELP "setup and enables call-graph (stack chain/backtrace) recording: "
 791
 792 #ifdef HAVE_DWARF_UNWIND_SUPPORT
 793 const char record_callchain_help[] = CALLCHAIN_HELP "fp dwarf";
 794 #else
 795 const char record_callchain_help[] = CALLCHAIN_HELP "fp";
 796 #endif
 797
 798 /*
 799  * XXX Will stay a global variable till we fix builtin-script.c to stop messing
 800  * with it and switch to use the library functions in perf_evlist that came
 801  * from builtin-record.c, i.e. use record_opts,
 802  * perf_evlist__prepare_workload, etc instead of fork+exec'in 'perf record',
 803  * using pipes, etc.
 804  */
 805 const struct option record_options[] = {
 806         OPT_CALLBACK('e', "event", &record.evlist, "event",
 807                      "event selector. use 'perf list' to list available events",
 808                      parse_events_option),
 809         OPT_CALLBACK(0, "filter", &record.evlist, "filter",
 810                      "event filter", parse_filter),
 811         OPT_STRING('p', "pid", &record.opts.target.pid, "pid",
 812                     "record events on existing process id"),
 813         OPT_STRING('t', "tid", &record.opts.target.tid, "tid",
 814                     "record events on existing thread id"),
 815         OPT_INTEGER('r', "realtime", &record.realtime_prio,
 816                     "collect data with this RT SCHED_FIFO priority"),
 817         OPT_BOOLEAN(0, "no-buffering", &record.opts.no_buffering,
 818                     "collect data without buffering"),
 819         OPT_BOOLEAN('R', "raw-samples", &record.opts.raw_samples,
 820                     "collect raw sample records from all opened counters"),
 821         OPT_BOOLEAN('a', "all-cpus", &record.opts.target.system_wide,
 822                             "system-wide collection from all CPUs"),
 823         OPT_STRING('C', "cpu", &record.opts.target.cpu_list, "cpu",
 824                     "list of cpus to monitor"),
 825         OPT_U64('c', "count", &record.opts.user_interval, "event period to sample"),
 826         OPT_STRING('o', "output", &record.file.path, "file",
 827                     "output file name"),
 828         OPT_BOOLEAN_SET('i', "no-inherit", &record.opts.no_inherit,
 829                         &record.opts.no_inherit_set,
 830                         "child tasks do not inherit counters"),
 831         OPT_UINTEGER('F', "freq", &record.opts.user_freq, "profile at this frequency"),
 832         OPT_CALLBACK('m', "mmap-pages", &record.opts.mmap_pages, "pages",
 833                      "number of mmap data pages",
 834                      perf_evlist__parse_mmap_pages),
 835         OPT_BOOLEAN(0, "group", &record.opts.group,
 836                     "put the counters into a counter group"),
 837         OPT_CALLBACK_NOOPT('g', NULL, &record.opts,
 838                            NULL, "enables call-graph recording" ,
 839                            &record_callchain_opt),
 840         OPT_CALLBACK(0, "call-graph", &record.opts,
 841                      "mode[,dump_size]", record_callchain_help,
 842                      &record_parse_callchain_opt),
 843         OPT_INCR('v', "verbose", &verbose,
 844                     "be more verbose (show counter open errors, etc)"),
 845         OPT_BOOLEAN('q', "quiet", &quiet, "don't print any message"),
 846         OPT_BOOLEAN('s', "stat", &record.opts.inherit_stat,
 847                     "per thread counts"),
 848         OPT_BOOLEAN('d', "data", &record.opts.sample_address,
 849                     "Sample addresses"),
 850         OPT_BOOLEAN('T', "timestamp", &record.opts.sample_time, "Sample timestamps"),
 851         OPT_BOOLEAN('P', "period", &record.opts.period, "Sample period"),
 852         OPT_BOOLEAN('n', "no-samples", &record.opts.no_samples,
 853                     "don't sample"),
 854         OPT_BOOLEAN('N', "no-buildid-cache", &record.no_buildid_cache,
 855                     "do not update the buildid cache"),
 856         OPT_BOOLEAN('B', "no-buildid", &record.no_buildid,
 857                     "do not collect buildids in perf.data"),
 858         OPT_CALLBACK('G', "cgroup", &record.evlist, "name",
 859                      "monitor event in cgroup name only",
 860                      parse_cgroups),
 861         OPT_UINTEGER('D', "delay", &record.opts.initial_delay,
 862                   "ms to wait before starting measurement after program start"),
 863         OPT_STRING('u', "uid", &record.opts.target.uid_str, "user",
 864                    "user to profile"),
 865
 866         OPT_CALLBACK_NOOPT('b', "branch-any", &record.opts.branch_stack,
 867                      "branch any", "sample any taken branches",
 868                      parse_branch_stack),
 869
 870         OPT_CALLBACK('j', "branch-filter", &record.opts.branch_stack,
 871                      "branch filter mask", "branch stack filter modes",
 872                      parse_branch_stack),
 873         OPT_BOOLEAN('W', "weight", &record.opts.sample_weight,
 874                     "sample by weight (on special events only)"),
 875         OPT_BOOLEAN(0, "transaction", &record.opts.sample_transaction,
 876                     "sample transaction flags (special events only)"),
 877         OPT_BOOLEAN(0, "per-thread", &record.opts.target.per_thread,
 878                     "use per-thread mmaps"),
 879         OPT_END()
 880 };
 881
 882 int cmd_record(int argc, const char **argv, const char *prefix __maybe_unused)
 883 {
 884         int err = -ENOMEM;
 885         struct record *rec = &record;
 886         char errbuf[BUFSIZ];
 887
 888         rec->evlist = perf_evlist__new();
 889         if (rec->evlist == NULL)
 890                 return -ENOMEM;
 891
 892         perf_config(perf_record_config, rec);
 893
 894         argc = parse_options(argc, argv, record_options, record_usage,
 895                             PARSE_OPT_STOP_AT_NON_OPTION);
 896         if (!argc && target__none(&rec->opts.target))
 897                 usage_with_options(record_usage, record_options);
 898
 899         if (nr_cgroups && !rec->opts.target.system_wide) {
 900                 ui__error("cgroup monitoring only available in"
 901                           " system-wide mode\n");
 902                 usage_with_options(record_usage, record_options);
 903         }
 904
 905         symbol__init();
 906
 907         if (symbol_conf.kptr_restrict)
 908                 pr_warning(
 909 "WARNING: Kernel address maps (/proc/{kallsyms,modules}) are restricted,\n"
 910 "check /proc/sys/kernel/kptr_restrict.\n\n"
 911 "Samples in kernel functions may not be resolved if a suitable vmlinux\n"
 912 "file is not found in the buildid cache or in the vmlinux path.\n\n"
 913 "Samples in kernel modules won't be resolved at all.\n\n"
 914 "If some relocation was applied (e.g. kexec) symbols may be misresolved\n"
 915 "even with a suitable vmlinux or kallsyms file.\n\n");
 916
 917         if (rec->no_buildid_cache || rec->no_buildid)
 918                 disable_buildid_cache();
 919
 920         if (rec->evlist->nr_entries == 0 &&
 921             perf_evlist__add_default(rec->evlist) < 0) {
 922                 pr_err("Not enough memory for event selector list\n");
 923                 goto out_symbol_exit;
 924         }
 925
 926         if (rec->opts.target.tid && !rec->opts.no_inherit_set)
 927                 rec->opts.no_inherit = true;
 928
 929         err = target__validate(&rec->opts.target);
 930         if (err) {
 931                 target__strerror(&rec->opts.target, err, errbuf, BUFSIZ);
 932                 ui__warning("%s", errbuf);
 933         }
 934
 935         err = target__parse_uid(&rec->opts.target);
 936         if (err) {
 937                 int saved_errno = errno;
 938
 939                 target__strerror(&rec->opts.target, err, errbuf, BUFSIZ);
 940                 ui__error("%s", errbuf);
 941
 942                 err = -saved_errno;
 943                 goto out_symbol_exit;
 944         }
 945
 946         err = -ENOMEM;
 947         if (perf_evlist__create_maps(rec->evlist, &rec->opts.target) < 0)
 948                 usage_with_options(record_usage, record_options);
 949
 950         if (record_opts__config(&rec->opts)) {
 951                 err = -EINVAL;
 952                 goto out_symbol_exit;
 953         }
 954
 955         err = __cmd_record(&record, argc, argv);
 956 out_symbol_exit:
 957         perf_evlist__delete(rec->evlist);
 958         symbol__exit();
 959         return err;
 960 }