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                         if (err < 0 && errno == EINTR)
 458                                 err = 0;
 459                         waking++;
 460                 }
 461
 462                 /*
 463                  * When perf is starting the traced process, at the end events
 464                  * die with the process and we wait for that. Thus no need to
 465                  * disable events in this case.
 466                  */
 467                 if (done && !disabled && !target__none(&opts->target)) {
 468                         perf_evlist__disable(rec->evlist);
 469                         disabled = true;
 470                 }
 471         }
 472
 473         if (forks && workload_exec_errno) {
 474                 char msg[512];
 475                 const char *emsg = strerror_r(workload_exec_errno, msg, sizeof(msg));
 476                 pr_err("Workload failed: %s\n", emsg);
 477                 err = -1;
 478                 goto out_child;
 479         }
 480
 481         if (!quiet) {
 482                 fprintf(stderr, "[ perf record: Woken up %ld times to write data ]\n", waking);
 483
 484                 /*
 485                  * Approximate RIP event size: 24 bytes.
 486                  */
 487                 fprintf(stderr,
 488                         "[ perf record: Captured and wrote %.3f MB %s (~%" PRIu64 " samples) ]\n",
 489                         (double)rec->bytes_written / 1024.0 / 1024.0,
 490                         file->path,
 491                         rec->bytes_written / 24);
 492         }
 493
 494 out_child:
 495         if (forks) {
 496                 int exit_status;
 497
 498                 if (!child_finished)
 499                         kill(rec->evlist->workload.pid, SIGTERM);
 500
 501                 wait(&exit_status);
 502
 503                 if (err < 0)
 504                         status = err;
 505                 else if (WIFEXITED(exit_status))
 506                         status = WEXITSTATUS(exit_status);
 507                 else if (WIFSIGNALED(exit_status))
 508                         signr = WTERMSIG(exit_status);
 509         } else
 510                 status = err;
 511
 512         if (!err && !file->is_pipe) {
 513                 rec->session->header.data_size += rec->bytes_written;
 514
 515                 if (!rec->no_buildid)
 516                         process_buildids(rec);
 517                 perf_session__write_header(rec->session, rec->evlist,
 518                                            file->fd, true);
 519         }
 520
 521 out_delete_session:
 522         perf_session__delete(session);
 523         return status;
 524 }
 525
 526 #define BRANCH_OPT(n, m) \
 527         { .name = n, .mode = (m) }
 528
 529 #define BRANCH_END { .name = NULL }
 530
 531 struct branch_mode {
 532         const char *name;
 533         int mode;
 534 };
 535
 536 static const struct branch_mode branch_modes[] = {
 537         BRANCH_OPT("u", PERF_SAMPLE_BRANCH_USER),
 538         BRANCH_OPT("k", PERF_SAMPLE_BRANCH_KERNEL),
 539         BRANCH_OPT("hv", PERF_SAMPLE_BRANCH_HV),
 540         BRANCH_OPT("any", PERF_SAMPLE_BRANCH_ANY),
 541         BRANCH_OPT("any_call", PERF_SAMPLE_BRANCH_ANY_CALL),
 542         BRANCH_OPT("any_ret", PERF_SAMPLE_BRANCH_ANY_RETURN),
 543         BRANCH_OPT("ind_call", PERF_SAMPLE_BRANCH_IND_CALL),
 544         BRANCH_OPT("abort_tx", PERF_SAMPLE_BRANCH_ABORT_TX),
 545         BRANCH_OPT("in_tx", PERF_SAMPLE_BRANCH_IN_TX),
 546         BRANCH_OPT("no_tx", PERF_SAMPLE_BRANCH_NO_TX),
 547         BRANCH_END
 548 };
 549
 550 static int
 551 parse_branch_stack(const struct option *opt, const char *str, int unset)
 552 {
 553 #define ONLY_PLM \
 554         (PERF_SAMPLE_BRANCH_USER        |\
 555          PERF_SAMPLE_BRANCH_KERNEL      |\
 556          PERF_SAMPLE_BRANCH_HV)
 557
 558         uint64_t *mode = (uint64_t *)opt->value;
 559         const struct branch_mode *br;
 560         char *s, *os = NULL, *p;
 561         int ret = -1;
 562
 563         if (unset)
 564                 return 0;
 565
 566         /*
 567          * cannot set it twice, -b + --branch-filter for instance
 568          */
 569         if (*mode)
 570                 return -1;
 571
 572         /* str may be NULL in case no arg is passed to -b */
 573         if (str) {
 574                 /* because str is read-only */
 575                 s = os = strdup(str);
 576                 if (!s)
 577                         return -1;
 578
 579                 for (;;) {
 580                         p = strchr(s, ',');
 581                         if (p)
 582                                 *p = '\0';
 583
 584                         for (br = branch_modes; br->name; br++) {
 585                                 if (!strcasecmp(s, br->name))
 586                                         break;
 587                         }
 588                         if (!br->name) {
 589                                 ui__warning("unknown branch filter %s,"
 590                                             " check man page\n", s);
 591                                 goto error;
 592                         }
 593
 594                         *mode |= br->mode;
 595
 596                         if (!p)
 597                                 break;
 598
 599                         s = p + 1;
 600                 }
 601         }
 602         ret = 0;
 603
 604         /* default to any branch */
 605         if ((*mode & ~ONLY_PLM) == 0) {
 606                 *mode = PERF_SAMPLE_BRANCH_ANY;
 607         }
 608 error:
 609         free(os);
 610         return ret;
 611 }
 612
 613 #ifdef HAVE_DWARF_UNWIND_SUPPORT
 614 static int get_stack_size(char *str, unsigned long *_size)
 615 {
 616         char *endptr;
 617         unsigned long size;
 618         unsigned long max_size = round_down(USHRT_MAX, sizeof(u64));
 619
 620         size = strtoul(str, &endptr, 0);
 621
 622         do {
 623                 if (*endptr)
 624                         break;
 625
 626                 size = round_up(size, sizeof(u64));
 627                 if (!size || size > max_size)
 628                         break;
 629
 630                 *_size = size;
 631                 return 0;
 632
 633         } while (0);
 634
 635         pr_err("callchain: Incorrect stack dump size (max %ld): %s\n",
 636                max_size, str);
 637         return -1;
 638 }
 639 #endif /* HAVE_DWARF_UNWIND_SUPPORT */
 640
 641 int record_parse_callchain(const char *arg, struct record_opts *opts)
 642 {
 643         char *tok, *name, *saveptr = NULL;
 644         char *buf;
 645         int ret = -1;
 646
 647         /* We need buffer that we know we can write to. */
 648         buf = malloc(strlen(arg) + 1);
 649         if (!buf)
 650                 return -ENOMEM;
 651
 652         strcpy(buf, arg);
 653
 654         tok = strtok_r((char *)buf, ",", &saveptr);
 655         name = tok ? : (char *)buf;
 656
 657         do {
 658                 /* Framepointer style */
 659                 if (!strncmp(name, "fp", sizeof("fp"))) {
 660                         if (!strtok_r(NULL, ",", &saveptr)) {
 661                                 opts->call_graph = CALLCHAIN_FP;
 662                                 ret = 0;
 663                         } else
 664                                 pr_err("callchain: No more arguments "
 665                                        "needed for -g fp\n");
 666                         break;
 667
 668 #ifdef HAVE_DWARF_UNWIND_SUPPORT
 669                 /* Dwarf style */
 670                 } else if (!strncmp(name, "dwarf", sizeof("dwarf"))) {
 671                         const unsigned long default_stack_dump_size = 8192;
 672
 673                         ret = 0;
 674                         opts->call_graph = CALLCHAIN_DWARF;
 675                         opts->stack_dump_size = default_stack_dump_size;
 676
 677                         tok = strtok_r(NULL, ",", &saveptr);
 678                         if (tok) {
 679                                 unsigned long size = 0;
 680
 681                                 ret = get_stack_size(tok, &size);
 682                                 opts->stack_dump_size = size;
 683                         }
 684 #endif /* HAVE_DWARF_UNWIND_SUPPORT */
 685                 } else {
 686                         pr_err("callchain: Unknown --call-graph option "
 687                                "value: %s\n", arg);
 688                         break;
 689                 }
 690
 691         } while (0);
 692
 693         free(buf);
 694         return ret;
 695 }
 696
 697 static void callchain_debug(struct record_opts *opts)
 698 {
 699         static const char *str[CALLCHAIN_MAX] = { "NONE", "FP", "DWARF" };
 700
 701         pr_debug("callchain: type %s\n", str[opts->call_graph]);
 702
 703         if (opts->call_graph == CALLCHAIN_DWARF)
 704                 pr_debug("callchain: stack dump size %d\n",
 705                          opts->stack_dump_size);
 706 }
 707
 708 int record_parse_callchain_opt(const struct option *opt,
 709                                const char *arg,
 710                                int unset)
 711 {
 712         struct record_opts *opts = opt->value;
 713         int ret;
 714
 715         opts->call_graph_enabled = !unset;
 716
 717         /* --no-call-graph */
 718         if (unset) {
 719                 opts->call_graph = CALLCHAIN_NONE;
 720                 pr_debug("callchain: disabled\n");
 721                 return 0;
 722         }
 723
 724         ret = record_parse_callchain(arg, opts);
 725         if (!ret)
 726                 callchain_debug(opts);
 727
 728         return ret;
 729 }
 730
 731 int record_callchain_opt(const struct option *opt,
 732                          const char *arg __maybe_unused,
 733                          int unset __maybe_unused)
 734 {
 735         struct record_opts *opts = opt->value;
 736
 737         opts->call_graph_enabled = !unset;
 738
 739         if (opts->call_graph == CALLCHAIN_NONE)
 740                 opts->call_graph = CALLCHAIN_FP;
 741
 742         callchain_debug(opts);
 743         return 0;
 744 }
 745
 746 static int perf_record_config(const char *var, const char *value, void *cb)
 747 {
 748         struct record *rec = cb;
 749
 750         if (!strcmp(var, "record.call-graph"))
 751                 return record_parse_callchain(value, &rec->opts);
 752
 753         return perf_default_config(var, value, cb);
 754 }
 755
 756 static const char * const record_usage[] = {
 757         "perf record [<options>] [<command>]",
 758         "perf record [<options>] -- <command> [<options>]",
 759         NULL
 760 };
 761
 762 /*
 763  * XXX Ideally would be local to cmd_record() and passed to a record__new
 764  * because we need to have access to it in record__exit, that is called
 765  * after cmd_record() exits, but since record_options need to be accessible to
 766  * builtin-script, leave it here.
 767  *
 768  * At least we don't ouch it in all the other functions here directly.
 769  *
 770  * Just say no to tons of global variables, sigh.
 771  */
 772 static struct record record = {
 773         .opts = {
 774                 .mmap_pages          = UINT_MAX,
 775                 .user_freq           = UINT_MAX,
 776                 .user_interval       = ULLONG_MAX,
 777                 .freq                = 4000,
 778                 .target              = {
 779                         .uses_mmap   = true,
 780                         .default_per_cpu = true,
 781                 },
 782         },
 783 };
 784
 785 #define CALLCHAIN_HELP "setup and enables call-graph (stack chain/backtrace) recording: "
 786
 787 #ifdef HAVE_DWARF_UNWIND_SUPPORT
 788 const char record_callchain_help[] = CALLCHAIN_HELP "fp dwarf";
 789 #else
 790 const char record_callchain_help[] = CALLCHAIN_HELP "fp";
 791 #endif
 792
 793 /*
 794  * XXX Will stay a global variable till we fix builtin-script.c to stop messing
 795  * with it and switch to use the library functions in perf_evlist that came
 796  * from builtin-record.c, i.e. use record_opts,
 797  * perf_evlist__prepare_workload, etc instead of fork+exec'in 'perf record',
 798  * using pipes, etc.
 799  */
 800 const struct option record_options[] = {
 801         OPT_CALLBACK('e', "event", &record.evlist, "event",
 802                      "event selector. use 'perf list' to list available events",
 803                      parse_events_option),
 804         OPT_CALLBACK(0, "filter", &record.evlist, "filter",
 805                      "event filter", parse_filter),
 806         OPT_STRING('p', "pid", &record.opts.target.pid, "pid",
 807                     "record events on existing process id"),
 808         OPT_STRING('t', "tid", &record.opts.target.tid, "tid",
 809                     "record events on existing thread id"),
 810         OPT_INTEGER('r', "realtime", &record.realtime_prio,
 811                     "collect data with this RT SCHED_FIFO priority"),
 812         OPT_BOOLEAN(0, "no-buffering", &record.opts.no_buffering,
 813                     "collect data without buffering"),
 814         OPT_BOOLEAN('R', "raw-samples", &record.opts.raw_samples,
 815                     "collect raw sample records from all opened counters"),
 816         OPT_BOOLEAN('a', "all-cpus", &record.opts.target.system_wide,
 817                             "system-wide collection from all CPUs"),
 818         OPT_STRING('C', "cpu", &record.opts.target.cpu_list, "cpu",
 819                     "list of cpus to monitor"),
 820         OPT_U64('c', "count", &record.opts.user_interval, "event period to sample"),
 821         OPT_STRING('o', "output", &record.file.path, "file",
 822                     "output file name"),
 823         OPT_BOOLEAN_SET('i', "no-inherit", &record.opts.no_inherit,
 824                         &record.opts.no_inherit_set,
 825                         "child tasks do not inherit counters"),
 826         OPT_UINTEGER('F', "freq", &record.opts.user_freq, "profile at this frequency"),
 827         OPT_CALLBACK('m', "mmap-pages", &record.opts.mmap_pages, "pages",
 828                      "number of mmap data pages",
 829                      perf_evlist__parse_mmap_pages),
 830         OPT_BOOLEAN(0, "group", &record.opts.group,
 831                     "put the counters into a counter group"),
 832         OPT_CALLBACK_NOOPT('g', NULL, &record.opts,
 833                            NULL, "enables call-graph recording" ,
 834                            &record_callchain_opt),
 835         OPT_CALLBACK(0, "call-graph", &record.opts,
 836                      "mode[,dump_size]", record_callchain_help,
 837                      &record_parse_callchain_opt),
 838         OPT_INCR('v', "verbose", &verbose,
 839                     "be more verbose (show counter open errors, etc)"),
 840         OPT_BOOLEAN('q', "quiet", &quiet, "don't print any message"),
 841         OPT_BOOLEAN('s', "stat", &record.opts.inherit_stat,
 842                     "per thread counts"),
 843         OPT_BOOLEAN('d', "data", &record.opts.sample_address,
 844                     "Sample addresses"),
 845         OPT_BOOLEAN('T', "timestamp", &record.opts.sample_time, "Sample timestamps"),
 846         OPT_BOOLEAN('P', "period", &record.opts.period, "Sample period"),
 847         OPT_BOOLEAN('n', "no-samples", &record.opts.no_samples,
 848                     "don't sample"),
 849         OPT_BOOLEAN('N', "no-buildid-cache", &record.no_buildid_cache,
 850                     "do not update the buildid cache"),
 851         OPT_BOOLEAN('B', "no-buildid", &record.no_buildid,
 852                     "do not collect buildids in perf.data"),
 853         OPT_CALLBACK('G', "cgroup", &record.evlist, "name",
 854                      "monitor event in cgroup name only",
 855                      parse_cgroups),
 856         OPT_UINTEGER('D', "delay", &record.opts.initial_delay,
 857                   "ms to wait before starting measurement after program start"),
 858         OPT_STRING('u', "uid", &record.opts.target.uid_str, "user",
 859                    "user to profile"),
 860
 861         OPT_CALLBACK_NOOPT('b', "branch-any", &record.opts.branch_stack,
 862                      "branch any", "sample any taken branches",
 863                      parse_branch_stack),
 864
 865         OPT_CALLBACK('j', "branch-filter", &record.opts.branch_stack,
 866                      "branch filter mask", "branch stack filter modes",
 867                      parse_branch_stack),
 868         OPT_BOOLEAN('W', "weight", &record.opts.sample_weight,
 869                     "sample by weight (on special events only)"),
 870         OPT_BOOLEAN(0, "transaction", &record.opts.sample_transaction,
 871                     "sample transaction flags (special events only)"),
 872         OPT_BOOLEAN(0, "per-thread", &record.opts.target.per_thread,
 873                     "use per-thread mmaps"),
 874         OPT_END()
 875 };
 876
 877 int cmd_record(int argc, const char **argv, const char *prefix __maybe_unused)
 878 {
 879         int err = -ENOMEM;
 880         struct record *rec = &record;
 881         char errbuf[BUFSIZ];
 882
 883         rec->evlist = perf_evlist__new();
 884         if (rec->evlist == NULL)
 885                 return -ENOMEM;
 886
 887         perf_config(perf_record_config, rec);
 888
 889         argc = parse_options(argc, argv, record_options, record_usage,
 890                             PARSE_OPT_STOP_AT_NON_OPTION);
 891         if (!argc && target__none(&rec->opts.target))
 892                 usage_with_options(record_usage, record_options);
 893
 894         if (nr_cgroups && !rec->opts.target.system_wide) {
 895                 ui__error("cgroup monitoring only available in"
 896                           " system-wide mode\n");
 897                 usage_with_options(record_usage, record_options);
 898         }
 899
 900         symbol__init();
 901
 902         if (symbol_conf.kptr_restrict)
 903                 pr_warning(
 904 "WARNING: Kernel address maps (/proc/{kallsyms,modules}) are restricted,\n"
 905 "check /proc/sys/kernel/kptr_restrict.\n\n"
 906 "Samples in kernel functions may not be resolved if a suitable vmlinux\n"
 907 "file is not found in the buildid cache or in the vmlinux path.\n\n"
 908 "Samples in kernel modules won't be resolved at all.\n\n"
 909 "If some relocation was applied (e.g. kexec) symbols may be misresolved\n"
 910 "even with a suitable vmlinux or kallsyms file.\n\n");
 911
 912         if (rec->no_buildid_cache || rec->no_buildid)
 913                 disable_buildid_cache();
 914
 915         if (rec->evlist->nr_entries == 0 &&
 916             perf_evlist__add_default(rec->evlist) < 0) {
 917                 pr_err("Not enough memory for event selector list\n");
 918                 goto out_symbol_exit;
 919         }
 920
 921         if (rec->opts.target.tid && !rec->opts.no_inherit_set)
 922                 rec->opts.no_inherit = true;
 923
 924         err = target__validate(&rec->opts.target);
 925         if (err) {
 926                 target__strerror(&rec->opts.target, err, errbuf, BUFSIZ);
 927                 ui__warning("%s", errbuf);
 928         }
 929
 930         err = target__parse_uid(&rec->opts.target);
 931         if (err) {
 932                 int saved_errno = errno;
 933
 934                 target__strerror(&rec->opts.target, err, errbuf, BUFSIZ);
 935                 ui__error("%s", errbuf);
 936
 937                 err = -saved_errno;
 938                 goto out_symbol_exit;
 939         }
 940
 941         err = -ENOMEM;
 942         if (perf_evlist__create_maps(rec->evlist, &rec->opts.target) < 0)
 943                 usage_with_options(record_usage, record_options);
 944
 945         if (record_opts__config(&rec->opts)) {
 946                 err = -EINVAL;
 947                 goto out_symbol_exit;
 948         }
 949
 950         err = __cmd_record(&record, argc, argv);
 951 out_symbol_exit:
 952         perf_evlist__delete(rec->evlist);
 953         symbol__exit();
 954         return err;
 955 }