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         u64 bytes_written = rec->bytes_written;
 242         int i;
 243         int rc = 0;
 244
 245         for (i = 0; i < rec->evlist->nr_mmaps; i++) {
 246                 if (rec->evlist->mmap[i].base) {
 247                         if (record__mmap_read(rec, &rec->evlist->mmap[i]) != 0) {
 248                                 rc = -1;
 249                                 goto out;
 250                         }
 251                 }
 252         }
 253
 254         /*
 255          * Mark the round finished in case we wrote
 256          * at least one event.
 257          */
 258         if (bytes_written != rec->bytes_written)
 259                 rc = record__write(rec, &finished_round_event, sizeof(finished_round_event));
 260
 261 out:
 262         return rc;
 263 }
 264
 265 static void record__init_features(struct record *rec)
 266 {
 267         struct perf_session *session = rec->session;
 268         int feat;
 269
 270         for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++)
 271                 perf_header__set_feat(&session->header, feat);
 272
 273         if (rec->no_buildid)
 274                 perf_header__clear_feat(&session->header, HEADER_BUILD_ID);
 275
 276         if (!have_tracepoints(&rec->evlist->entries))
 277                 perf_header__clear_feat(&session->header, HEADER_TRACING_DATA);
 278
 279         if (!rec->opts.branch_stack)
 280                 perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK);
 281 }
 282
 283 static volatile int workload_exec_errno;
 284
 285 /*
 286  * perf_evlist__prepare_workload will send a SIGUSR1
 287  * if the fork fails, since we asked by setting its
 288  * want_signal to true.
 289  */
 290 static void workload_exec_failed_signal(int signo __maybe_unused,
 291                                         siginfo_t *info,
 292                                         void *ucontext __maybe_unused)
 293 {
 294         workload_exec_errno = info->si_value.sival_int;
 295         done = 1;
 296         child_finished = 1;
 297 }
 298
 299 static int __cmd_record(struct record *rec, int argc, const char **argv)
 300 {
 301         int err;
 302         int status = 0;
 303         unsigned long waking = 0;
 304         const bool forks = argc > 0;
 305         struct machine *machine;
 306         struct perf_tool *tool = &rec->tool;
 307         struct record_opts *opts = &rec->opts;
 308         struct perf_data_file *file = &rec->file;
 309         struct perf_session *session;
 310         bool disabled = false;
 311
 312         rec->progname = argv[0];
 313
 314         atexit(record__sig_exit);
 315         signal(SIGCHLD, sig_handler);
 316         signal(SIGINT, sig_handler);
 317         signal(SIGTERM, sig_handler);
 318
 319         session = perf_session__new(file, false, NULL);
 320         if (session == NULL) {
 321                 pr_err("Perf session creation failed.\n");
 322                 return -1;
 323         }
 324
 325         rec->session = session;
 326
 327         record__init_features(rec);
 328
 329         if (forks) {
 330                 err = perf_evlist__prepare_workload(rec->evlist, &opts->target,
 331                                                     argv, file->is_pipe,
 332                                                     workload_exec_failed_signal);
 333                 if (err < 0) {
 334                         pr_err("Couldn't run the workload!\n");
 335                         status = err;
 336                         goto out_delete_session;
 337                 }
 338         }
 339
 340         if (record__open(rec) != 0) {
 341                 err = -1;
 342                 goto out_child;
 343         }
 344
 345         if (!rec->evlist->nr_groups)
 346                 perf_header__clear_feat(&session->header, HEADER_GROUP_DESC);
 347
 348         if (file->is_pipe) {
 349                 err = perf_header__write_pipe(file->fd);
 350                 if (err < 0)
 351                         goto out_child;
 352         } else {
 353                 err = perf_session__write_header(session, rec->evlist,
 354                                                  file->fd, false);
 355                 if (err < 0)
 356                         goto out_child;
 357         }
 358
 359         if (!rec->no_buildid
 360             && !perf_header__has_feat(&session->header, HEADER_BUILD_ID)) {
 361                 pr_err("Couldn't generate buildids. "
 362                        "Use --no-buildid to profile anyway.\n");
 363                 err = -1;
 364                 goto out_child;
 365         }
 366
 367         machine = &session->machines.host;
 368
 369         if (file->is_pipe) {
 370                 err = perf_event__synthesize_attrs(tool, session,
 371                                                    process_synthesized_event);
 372                 if (err < 0) {
 373                         pr_err("Couldn't synthesize attrs.\n");
 374                         goto out_child;
 375                 }
 376
 377                 if (have_tracepoints(&rec->evlist->entries)) {
 378                         /*
 379                          * FIXME err <= 0 here actually means that
 380                          * there were no tracepoints so its not really
 381                          * an error, just that we don't need to
 382                          * synthesize anything.  We really have to
 383                          * return this more properly and also
 384                          * propagate errors that now are calling die()
 385                          */
 386                         err = perf_event__synthesize_tracing_data(tool, file->fd, rec->evlist,
 387                                                                   process_synthesized_event);
 388                         if (err <= 0) {
 389                                 pr_err("Couldn't record tracing data.\n");
 390                                 goto out_child;
 391                         }
 392                         rec->bytes_written += err;
 393                 }
 394         }
 395
 396         err = perf_event__synthesize_kernel_mmap(tool, process_synthesized_event,
 397                                                  machine);
 398         if (err < 0)
 399                 pr_err("Couldn't record kernel reference relocation symbol\n"
 400                        "Symbol resolution may be skewed if relocation was used (e.g. kexec).\n"
 401                        "Check /proc/kallsyms permission or run as root.\n");
 402
 403         err = perf_event__synthesize_modules(tool, process_synthesized_event,
 404                                              machine);
 405         if (err < 0)
 406                 pr_err("Couldn't record kernel module information.\n"
 407                        "Symbol resolution may be skewed if relocation was used (e.g. kexec).\n"
 408                        "Check /proc/modules permission or run as root.\n");
 409
 410         if (perf_guest) {
 411                 machines__process_guests(&session->machines,
 412                                          perf_event__synthesize_guest_os, tool);
 413         }
 414
 415         err = __machine__synthesize_threads(machine, tool, &opts->target, rec->evlist->threads,
 416                                             process_synthesized_event, opts->sample_address);
 417         if (err != 0)
 418                 goto out_child;
 419
 420         if (rec->realtime_prio) {
 421                 struct sched_param param;
 422
 423                 param.sched_priority = rec->realtime_prio;
 424                 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
 425                         pr_err("Could not set realtime priority.\n");
 426                         err = -1;
 427                         goto out_child;
 428                 }
 429         }
 430
 431         /*
 432          * When perf is starting the traced process, all the events
 433          * (apart from group members) have enable_on_exec=1 set,
 434          * so don't spoil it by prematurely enabling them.
 435          */
 436         if (!target__none(&opts->target) && !opts->initial_delay)
 437                 perf_evlist__enable(rec->evlist);
 438
 439         /*
 440          * Let the child rip
 441          */
 442         if (forks)
 443                 perf_evlist__start_workload(rec->evlist);
 444
 445         if (opts->initial_delay) {
 446                 usleep(opts->initial_delay * 1000);
 447                 perf_evlist__enable(rec->evlist);
 448         }
 449
 450         for (;;) {
 451                 int hits = rec->samples;
 452
 453                 if (record__mmap_read_all(rec) < 0) {
 454                         err = -1;
 455                         goto out_child;
 456                 }
 457
 458                 if (hits == rec->samples) {
 459                         if (done)
 460                                 break;
 461                         err = poll(rec->evlist->pollfd, rec->evlist->nr_fds, -1);
 462                         /*
 463                          * Propagate error, only if there's any. Ignore positive
 464                          * number of returned events and interrupt error.
 465                          */
 466                         if (err > 0 || (err < 0 && errno == EINTR))
 467                                 err = 0;
 468                         waking++;
 469                 }
 470
 471                 /*
 472                  * When perf is starting the traced process, at the end events
 473                  * die with the process and we wait for that. Thus no need to
 474                  * disable events in this case.
 475                  */
 476                 if (done && !disabled && !target__none(&opts->target)) {
 477                         perf_evlist__disable(rec->evlist);
 478                         disabled = true;
 479                 }
 480         }
 481
 482         if (forks && workload_exec_errno) {
 483                 char msg[512];
 484                 const char *emsg = strerror_r(workload_exec_errno, msg, sizeof(msg));
 485                 pr_err("Workload failed: %s\n", emsg);
 486                 err = -1;
 487                 goto out_child;
 488         }
 489
 490         if (!quiet) {
 491                 fprintf(stderr, "[ perf record: Woken up %ld times to write data ]\n", waking);
 492
 493                 /*
 494                  * Approximate RIP event size: 24 bytes.
 495                  */
 496                 fprintf(stderr,
 497                         "[ perf record: Captured and wrote %.3f MB %s (~%" PRIu64 " samples) ]\n",
 498                         (double)rec->bytes_written / 1024.0 / 1024.0,
 499                         file->path,
 500                         rec->bytes_written / 24);
 501         }
 502
 503 out_child:
 504         if (forks) {
 505                 int exit_status;
 506
 507                 if (!child_finished)
 508                         kill(rec->evlist->workload.pid, SIGTERM);
 509
 510                 wait(&exit_status);
 511
 512                 if (err < 0)
 513                         status = err;
 514                 else if (WIFEXITED(exit_status))
 515                         status = WEXITSTATUS(exit_status);
 516                 else if (WIFSIGNALED(exit_status))
 517                         signr = WTERMSIG(exit_status);
 518         } else
 519                 status = err;
 520
 521         if (!err && !file->is_pipe) {
 522                 rec->session->header.data_size += rec->bytes_written;
 523
 524                 if (!rec->no_buildid)
 525                         process_buildids(rec);
 526                 perf_session__write_header(rec->session, rec->evlist,
 527                                            file->fd, true);
 528         }
 529
 530 out_delete_session:
 531         perf_session__delete(session);
 532         return status;
 533 }
 534
 535 #define BRANCH_OPT(n, m) \
 536         { .name = n, .mode = (m) }
 537
 538 #define BRANCH_END { .name = NULL }
 539
 540 struct branch_mode {
 541         const char *name;
 542         int mode;
 543 };
 544
 545 static const struct branch_mode branch_modes[] = {
 546         BRANCH_OPT("u", PERF_SAMPLE_BRANCH_USER),
 547         BRANCH_OPT("k", PERF_SAMPLE_BRANCH_KERNEL),
 548         BRANCH_OPT("hv", PERF_SAMPLE_BRANCH_HV),
 549         BRANCH_OPT("any", PERF_SAMPLE_BRANCH_ANY),
 550         BRANCH_OPT("any_call", PERF_SAMPLE_BRANCH_ANY_CALL),
 551         BRANCH_OPT("any_ret", PERF_SAMPLE_BRANCH_ANY_RETURN),
 552         BRANCH_OPT("ind_call", PERF_SAMPLE_BRANCH_IND_CALL),
 553         BRANCH_OPT("abort_tx", PERF_SAMPLE_BRANCH_ABORT_TX),
 554         BRANCH_OPT("in_tx", PERF_SAMPLE_BRANCH_IN_TX),
 555         BRANCH_OPT("no_tx", PERF_SAMPLE_BRANCH_NO_TX),
 556         BRANCH_OPT("cond", PERF_SAMPLE_BRANCH_COND),
 557         BRANCH_END
 558 };
 559
 560 static int
 561 parse_branch_stack(const struct option *opt, const char *str, int unset)
 562 {
 563 #define ONLY_PLM \
 564         (PERF_SAMPLE_BRANCH_USER        |\
 565          PERF_SAMPLE_BRANCH_KERNEL      |\
 566          PERF_SAMPLE_BRANCH_HV)
 567
 568         uint64_t *mode = (uint64_t *)opt->value;
 569         const struct branch_mode *br;
 570         char *s, *os = NULL, *p;
 571         int ret = -1;
 572
 573         if (unset)
 574                 return 0;
 575
 576         /*
 577          * cannot set it twice, -b + --branch-filter for instance
 578          */
 579         if (*mode)
 580                 return -1;
 581
 582         /* str may be NULL in case no arg is passed to -b */
 583         if (str) {
 584                 /* because str is read-only */
 585                 s = os = strdup(str);
 586                 if (!s)
 587                         return -1;
 588
 589                 for (;;) {
 590                         p = strchr(s, ',');
 591                         if (p)
 592                                 *p = '\0';
 593
 594                         for (br = branch_modes; br->name; br++) {
 595                                 if (!strcasecmp(s, br->name))
 596                                         break;
 597                         }
 598                         if (!br->name) {
 599                                 ui__warning("unknown branch filter %s,"
 600                                             " check man page\n", s);
 601                                 goto error;
 602                         }
 603
 604                         *mode |= br->mode;
 605
 606                         if (!p)
 607                                 break;
 608
 609                         s = p + 1;
 610                 }
 611         }
 612         ret = 0;
 613
 614         /* default to any branch */
 615         if ((*mode & ~ONLY_PLM) == 0) {
 616                 *mode = PERF_SAMPLE_BRANCH_ANY;
 617         }
 618 error:
 619         free(os);
 620         return ret;
 621 }
 622
 623 #ifdef HAVE_DWARF_UNWIND_SUPPORT
 624 static int get_stack_size(char *str, unsigned long *_size)
 625 {
 626         char *endptr;
 627         unsigned long size;
 628         unsigned long max_size = round_down(USHRT_MAX, sizeof(u64));
 629
 630         size = strtoul(str, &endptr, 0);
 631
 632         do {
 633                 if (*endptr)
 634                         break;
 635
 636                 size = round_up(size, sizeof(u64));
 637                 if (!size || size > max_size)
 638                         break;
 639
 640                 *_size = size;
 641                 return 0;
 642
 643         } while (0);
 644
 645         pr_err("callchain: Incorrect stack dump size (max %ld): %s\n",
 646                max_size, str);
 647         return -1;
 648 }
 649 #endif /* HAVE_DWARF_UNWIND_SUPPORT */
 650
 651 int record_parse_callchain(const char *arg, struct record_opts *opts)
 652 {
 653         char *tok, *name, *saveptr = NULL;
 654         char *buf;
 655         int ret = -1;
 656
 657         /* We need buffer that we know we can write to. */
 658         buf = malloc(strlen(arg) + 1);
 659         if (!buf)
 660                 return -ENOMEM;
 661
 662         strcpy(buf, arg);
 663
 664         tok = strtok_r((char *)buf, ",", &saveptr);
 665         name = tok ? : (char *)buf;
 666
 667         do {
 668                 /* Framepointer style */
 669                 if (!strncmp(name, "fp", sizeof("fp"))) {
 670                         if (!strtok_r(NULL, ",", &saveptr)) {
 671                                 opts->call_graph = CALLCHAIN_FP;
 672                                 ret = 0;
 673                         } else
 674                                 pr_err("callchain: No more arguments "
 675                                        "needed for -g fp\n");
 676                         break;
 677
 678 #ifdef HAVE_DWARF_UNWIND_SUPPORT
 679                 /* Dwarf style */
 680                 } else if (!strncmp(name, "dwarf", sizeof("dwarf"))) {
 681                         const unsigned long default_stack_dump_size = 8192;
 682
 683                         ret = 0;
 684                         opts->call_graph = CALLCHAIN_DWARF;
 685                         opts->stack_dump_size = default_stack_dump_size;
 686
 687                         tok = strtok_r(NULL, ",", &saveptr);
 688                         if (tok) {
 689                                 unsigned long size = 0;
 690
 691                                 ret = get_stack_size(tok, &size);
 692                                 opts->stack_dump_size = size;
 693                         }
 694 #endif /* HAVE_DWARF_UNWIND_SUPPORT */
 695                 } else {
 696                         pr_err("callchain: Unknown --call-graph option "
 697                                "value: %s\n", arg);
 698                         break;
 699                 }
 700
 701         } while (0);
 702
 703         free(buf);
 704         return ret;
 705 }
 706
 707 static void callchain_debug(struct record_opts *opts)
 708 {
 709         static const char *str[CALLCHAIN_MAX] = { "NONE", "FP", "DWARF" };
 710
 711         pr_debug("callchain: type %s\n", str[opts->call_graph]);
 712
 713         if (opts->call_graph == CALLCHAIN_DWARF)
 714                 pr_debug("callchain: stack dump size %d\n",
 715                          opts->stack_dump_size);
 716 }
 717
 718 int record_parse_callchain_opt(const struct option *opt,
 719                                const char *arg,
 720                                int unset)
 721 {
 722         struct record_opts *opts = opt->value;
 723         int ret;
 724
 725         opts->call_graph_enabled = !unset;
 726
 727         /* --no-call-graph */
 728         if (unset) {
 729                 opts->call_graph = CALLCHAIN_NONE;
 730                 pr_debug("callchain: disabled\n");
 731                 return 0;
 732         }
 733
 734         ret = record_parse_callchain(arg, opts);
 735         if (!ret)
 736                 callchain_debug(opts);
 737
 738         return ret;
 739 }
 740
 741 int record_callchain_opt(const struct option *opt,
 742                          const char *arg __maybe_unused,
 743                          int unset __maybe_unused)
 744 {
 745         struct record_opts *opts = opt->value;
 746
 747         opts->call_graph_enabled = !unset;
 748
 749         if (opts->call_graph == CALLCHAIN_NONE)
 750                 opts->call_graph = CALLCHAIN_FP;
 751
 752         callchain_debug(opts);
 753         return 0;
 754 }
 755
 756 static int perf_record_config(const char *var, const char *value, void *cb)
 757 {
 758         struct record *rec = cb;
 759
 760         if (!strcmp(var, "record.call-graph"))
 761                 return record_parse_callchain(value, &rec->opts);
 762
 763         return perf_default_config(var, value, cb);
 764 }
 765
 766 static const char * const record_usage[] = {
 767         "perf record [<options>] [<command>]",
 768         "perf record [<options>] -- <command> [<options>]",
 769         NULL
 770 };
 771
 772 /*
 773  * XXX Ideally would be local to cmd_record() and passed to a record__new
 774  * because we need to have access to it in record__exit, that is called
 775  * after cmd_record() exits, but since record_options need to be accessible to
 776  * builtin-script, leave it here.
 777  *
 778  * At least we don't ouch it in all the other functions here directly.
 779  *
 780  * Just say no to tons of global variables, sigh.
 781  */
 782 static struct record record = {
 783         .opts = {
 784                 .mmap_pages          = UINT_MAX,
 785                 .user_freq           = UINT_MAX,
 786                 .user_interval       = ULLONG_MAX,
 787                 .freq                = 4000,
 788                 .target              = {
 789                         .uses_mmap   = true,
 790                         .default_per_cpu = true,
 791                 },
 792         },
 793 };
 794
 795 #define CALLCHAIN_HELP "setup and enables call-graph (stack chain/backtrace) recording: "
 796
 797 #ifdef HAVE_DWARF_UNWIND_SUPPORT
 798 const char record_callchain_help[] = CALLCHAIN_HELP "fp dwarf";
 799 #else
 800 const char record_callchain_help[] = CALLCHAIN_HELP "fp";
 801 #endif
 802
 803 /*
 804  * XXX Will stay a global variable till we fix builtin-script.c to stop messing
 805  * with it and switch to use the library functions in perf_evlist that came
 806  * from builtin-record.c, i.e. use record_opts,
 807  * perf_evlist__prepare_workload, etc instead of fork+exec'in 'perf record',
 808  * using pipes, etc.
 809  */
 810 const struct option record_options[] = {
 811         OPT_CALLBACK('e', "event", &record.evlist, "event",
 812                      "event selector. use 'perf list' to list available events",
 813                      parse_events_option),
 814         OPT_CALLBACK(0, "filter", &record.evlist, "filter",
 815                      "event filter", parse_filter),
 816         OPT_STRING('p', "pid", &record.opts.target.pid, "pid",
 817                     "record events on existing process id"),
 818         OPT_STRING('t', "tid", &record.opts.target.tid, "tid",
 819                     "record events on existing thread id"),
 820         OPT_INTEGER('r', "realtime", &record.realtime_prio,
 821                     "collect data with this RT SCHED_FIFO priority"),
 822         OPT_BOOLEAN(0, "no-buffering", &record.opts.no_buffering,
 823                     "collect data without buffering"),
 824         OPT_BOOLEAN('R', "raw-samples", &record.opts.raw_samples,
 825                     "collect raw sample records from all opened counters"),
 826         OPT_BOOLEAN('a', "all-cpus", &record.opts.target.system_wide,
 827                             "system-wide collection from all CPUs"),
 828         OPT_STRING('C', "cpu", &record.opts.target.cpu_list, "cpu",
 829                     "list of cpus to monitor"),
 830         OPT_U64('c', "count", &record.opts.user_interval, "event period to sample"),
 831         OPT_STRING('o', "output", &record.file.path, "file",
 832                     "output file name"),
 833         OPT_BOOLEAN_SET('i', "no-inherit", &record.opts.no_inherit,
 834                         &record.opts.no_inherit_set,
 835                         "child tasks do not inherit counters"),
 836         OPT_UINTEGER('F', "freq", &record.opts.user_freq, "profile at this frequency"),
 837         OPT_CALLBACK('m', "mmap-pages", &record.opts.mmap_pages, "pages",
 838                      "number of mmap data pages",
 839                      perf_evlist__parse_mmap_pages),
 840         OPT_BOOLEAN(0, "group", &record.opts.group,
 841                     "put the counters into a counter group"),
 842         OPT_CALLBACK_NOOPT('g', NULL, &record.opts,
 843                            NULL, "enables call-graph recording" ,
 844                            &record_callchain_opt),
 845         OPT_CALLBACK(0, "call-graph", &record.opts,
 846                      "mode[,dump_size]", record_callchain_help,
 847                      &record_parse_callchain_opt),
 848         OPT_INCR('v', "verbose", &verbose,
 849                     "be more verbose (show counter open errors, etc)"),
 850         OPT_BOOLEAN('q', "quiet", &quiet, "don't print any message"),
 851         OPT_BOOLEAN('s', "stat", &record.opts.inherit_stat,
 852                     "per thread counts"),
 853         OPT_BOOLEAN('d', "data", &record.opts.sample_address,
 854                     "Sample addresses"),
 855         OPT_BOOLEAN('T', "timestamp", &record.opts.sample_time, "Sample timestamps"),
 856         OPT_BOOLEAN('P', "period", &record.opts.period, "Sample period"),
 857         OPT_BOOLEAN('n', "no-samples", &record.opts.no_samples,
 858                     "don't sample"),
 859         OPT_BOOLEAN('N', "no-buildid-cache", &record.no_buildid_cache,
 860                     "do not update the buildid cache"),
 861         OPT_BOOLEAN('B', "no-buildid", &record.no_buildid,
 862                     "do not collect buildids in perf.data"),
 863         OPT_CALLBACK('G', "cgroup", &record.evlist, "name",
 864                      "monitor event in cgroup name only",
 865                      parse_cgroups),
 866         OPT_UINTEGER('D', "delay", &record.opts.initial_delay,
 867                   "ms to wait before starting measurement after program start"),
 868         OPT_STRING('u', "uid", &record.opts.target.uid_str, "user",
 869                    "user to profile"),
 870
 871         OPT_CALLBACK_NOOPT('b', "branch-any", &record.opts.branch_stack,
 872                      "branch any", "sample any taken branches",
 873                      parse_branch_stack),
 874
 875         OPT_CALLBACK('j', "branch-filter", &record.opts.branch_stack,
 876                      "branch filter mask", "branch stack filter modes",
 877                      parse_branch_stack),
 878         OPT_BOOLEAN('W', "weight", &record.opts.sample_weight,
 879                     "sample by weight (on special events only)"),
 880         OPT_BOOLEAN(0, "transaction", &record.opts.sample_transaction,
 881                     "sample transaction flags (special events only)"),
 882         OPT_BOOLEAN(0, "per-thread", &record.opts.target.per_thread,
 883                     "use per-thread mmaps"),
 884         OPT_END()
 885 };
 886
 887 int cmd_record(int argc, const char **argv, const char *prefix __maybe_unused)
 888 {
 889         int err = -ENOMEM;
 890         struct record *rec = &record;
 891         char errbuf[BUFSIZ];
 892
 893         rec->evlist = perf_evlist__new();
 894         if (rec->evlist == NULL)
 895                 return -ENOMEM;
 896
 897         perf_config(perf_record_config, rec);
 898
 899         argc = parse_options(argc, argv, record_options, record_usage,
 900                             PARSE_OPT_STOP_AT_NON_OPTION);
 901         if (!argc && target__none(&rec->opts.target))
 902                 usage_with_options(record_usage, record_options);
 903
 904         if (nr_cgroups && !rec->opts.target.system_wide) {
 905                 ui__error("cgroup monitoring only available in"
 906                           " system-wide mode\n");
 907                 usage_with_options(record_usage, record_options);
 908         }
 909
 910         symbol__init();
 911
 912         if (symbol_conf.kptr_restrict)
 913                 pr_warning(
 914 "WARNING: Kernel address maps (/proc/{kallsyms,modules}) are restricted,\n"
 915 "check /proc/sys/kernel/kptr_restrict.\n\n"
 916 "Samples in kernel functions may not be resolved if a suitable vmlinux\n"
 917 "file is not found in the buildid cache or in the vmlinux path.\n\n"
 918 "Samples in kernel modules won't be resolved at all.\n\n"
 919 "If some relocation was applied (e.g. kexec) symbols may be misresolved\n"
 920 "even with a suitable vmlinux or kallsyms file.\n\n");
 921
 922         if (rec->no_buildid_cache || rec->no_buildid)
 923                 disable_buildid_cache();
 924
 925         if (rec->evlist->nr_entries == 0 &&
 926             perf_evlist__add_default(rec->evlist) < 0) {
 927                 pr_err("Not enough memory for event selector list\n");
 928                 goto out_symbol_exit;
 929         }
 930
 931         if (rec->opts.target.tid && !rec->opts.no_inherit_set)
 932                 rec->opts.no_inherit = true;
 933
 934         err = target__validate(&rec->opts.target);
 935         if (err) {
 936                 target__strerror(&rec->opts.target, err, errbuf, BUFSIZ);
 937                 ui__warning("%s", errbuf);
 938         }
 939
 940         err = target__parse_uid(&rec->opts.target);
 941         if (err) {
 942                 int saved_errno = errno;
 943
 944                 target__strerror(&rec->opts.target, err, errbuf, BUFSIZ);
 945                 ui__error("%s", errbuf);
 946
 947                 err = -saved_errno;
 948                 goto out_symbol_exit;
 949         }
 950
 951         err = -ENOMEM;
 952         if (perf_evlist__create_maps(rec->evlist, &rec->opts.target) < 0)
 953                 usage_with_options(record_usage, record_options);
 954
 955         if (record_opts__config(&rec->opts)) {
 956                 err = -EINVAL;
 957                 goto out_symbol_exit;
 958         }
 959
 960         err = __cmd_record(&record, argc, argv);
 961 out_symbol_exit:
 962         perf_evlist__delete(rec->evlist);
 963         symbol__exit();
 964         return err;
 965 }