tools/perf/builtin-record.c

   1 /*
   2  * builtin-record.c
   3  *
   4  * Builtin record command: Record the profile of a workload
   5  * (or a CPU, or a PID) into the perf.data output file - for
   6  * later analysis via perf report.
   7  */
   8 #include "builtin.h"
   9
  10 #include "perf.h"
  11
  12 #include "util/build-id.h"
  13 #include "util/util.h"
  14 #include "util/parse-options.h"
  15 #include "util/parse-events.h"
  16
  17 #include "util/callchain.h"
  18 #include "util/cgroup.h"
  19 #include "util/header.h"
  20 #include "util/event.h"
  21 #include "util/evlist.h"
  22 #include "util/evsel.h"
  23 #include "util/debug.h"
  24 #include "util/session.h"
  25 #include "util/tool.h"
  26 #include "util/symbol.h"
  27 #include "util/cpumap.h"
  28 #include "util/thread_map.h"
  29 #include "util/data.h"
  30 #include "util/auxtrace.h"
  31 #include "util/parse-branch-options.h"
  32
  33 #include <unistd.h>
  34 #include <sched.h>
  35 #include <sys/mman.h>
  36
  37
  38 struct record {
  39         struct perf_tool        tool;
  40         struct record_opts      opts;
  41         u64                     bytes_written;
  42         struct perf_data_file   file;
  43         struct auxtrace_record  *itr;
  44         struct perf_evlist      *evlist;
  45         struct perf_session     *session;
  46         const char              *progname;
  47         int                     realtime_prio;
  48         bool                    no_buildid;
  49         bool                    no_buildid_cache;
  50         long                    samples;
  51 };
  52
  53 static int record__write(struct record *rec, void *bf, size_t size)
  54 {
  55         if (perf_data_file__write(rec->session->file, bf, size) < 0) {
  56                 pr_err("failed to write perf data, error: %m\n");
  57                 return -1;
  58         }
  59
  60         rec->bytes_written += size;
  61         return 0;
  62 }
  63
  64 static int process_synthesized_event(struct perf_tool *tool,
  65                                      union perf_event *event,
  66                                      struct perf_sample *sample __maybe_unused,
  67                                      struct machine *machine __maybe_unused)
  68 {
  69         struct record *rec = container_of(tool, struct record, tool);
  70         return record__write(rec, event, event->header.size);
  71 }
  72
  73 static int record__mmap_read(struct record *rec, int idx)
  74 {
  75         struct perf_mmap *md = &rec->evlist->mmap[idx];
  76         u64 head = perf_mmap__read_head(md);
  77         u64 old = md->prev;
  78         unsigned char *data = md->base + page_size;
  79         unsigned long size;
  80         void *buf;
  81         int rc = 0;
  82
  83         if (old == head)
  84                 return 0;
  85
  86         rec->samples++;
  87
  88         size = head - old;
  89
  90         if ((old & md->mask) + size != (head & md->mask)) {
  91                 buf = &data[old & md->mask];
  92                 size = md->mask + 1 - (old & md->mask);
  93                 old += size;
  94
  95                 if (record__write(rec, buf, size) < 0) {
  96                         rc = -1;
  97                         goto out;
  98                 }
  99         }
 100
 101         buf = &data[old & md->mask];
 102         size = head - old;
 103         old += size;
 104
 105         if (record__write(rec, buf, size) < 0) {
 106                 rc = -1;
 107                 goto out;
 108         }
 109
 110         md->prev = old;
 111         perf_evlist__mmap_consume(rec->evlist, idx);
 112 out:
 113         return rc;
 114 }
 115
 116 static volatile int done;
 117 static volatile int signr = -1;
 118 static volatile int child_finished;
 119 static volatile int auxtrace_snapshot_enabled;
 120 static volatile int auxtrace_snapshot_err;
 121 static volatile int auxtrace_record__snapshot_started;
 122
 123 static void sig_handler(int sig)
 124 {
 125         if (sig == SIGCHLD)
 126                 child_finished = 1;
 127         else
 128                 signr = sig;
 129
 130         done = 1;
 131 }
 132
 133 static void record__sig_exit(void)
 134 {
 135         if (signr == -1)
 136                 return;
 137
 138         signal(signr, SIG_DFL);
 139         raise(signr);
 140 }
 141
 142 #ifdef HAVE_AUXTRACE_SUPPORT
 143
 144 static int record__process_auxtrace(struct perf_tool *tool,
 145                                     union perf_event *event, void *data1,
 146                                     size_t len1, void *data2, size_t len2)
 147 {
 148         struct record *rec = container_of(tool, struct record, tool);
 149         struct perf_data_file *file = &rec->file;
 150         size_t padding;
 151         u8 pad[8] = {0};
 152
 153         if (!perf_data_file__is_pipe(file)) {
 154                 off_t file_offset;
 155                 int fd = perf_data_file__fd(file);
 156                 int err;
 157
 158                 file_offset = lseek(fd, 0, SEEK_CUR);
 159                 if (file_offset == -1)
 160                         return -1;
 161                 err = auxtrace_index__auxtrace_event(&rec->session->auxtrace_index,
 162                                                      event, file_offset);
 163                 if (err)
 164                         return err;
 165         }
 166
 167         /* event.auxtrace.size includes padding, see __auxtrace_mmap__read() */
 168         padding = (len1 + len2) & 7;
 169         if (padding)
 170                 padding = 8 - padding;
 171
 172         record__write(rec, event, event->header.size);
 173         record__write(rec, data1, len1);
 174         if (len2)
 175                 record__write(rec, data2, len2);
 176         record__write(rec, &pad, padding);
 177
 178         return 0;
 179 }
 180
 181 static int record__auxtrace_mmap_read(struct record *rec,
 182                                       struct auxtrace_mmap *mm)
 183 {
 184         int ret;
 185
 186         ret = auxtrace_mmap__read(mm, rec->itr, &rec->tool,
 187                                   record__process_auxtrace);
 188         if (ret < 0)
 189                 return ret;
 190
 191         if (ret)
 192                 rec->samples++;
 193
 194         return 0;
 195 }
 196
 197 static int record__auxtrace_mmap_read_snapshot(struct record *rec,
 198                                                struct auxtrace_mmap *mm)
 199 {
 200         int ret;
 201
 202         ret = auxtrace_mmap__read_snapshot(mm, rec->itr, &rec->tool,
 203                                            record__process_auxtrace,
 204                                            rec->opts.auxtrace_snapshot_size);
 205         if (ret < 0)
 206                 return ret;
 207
 208         if (ret)
 209                 rec->samples++;
 210
 211         return 0;
 212 }
 213
 214 static int record__auxtrace_read_snapshot_all(struct record *rec)
 215 {
 216         int i;
 217         int rc = 0;
 218
 219         for (i = 0; i < rec->evlist->nr_mmaps; i++) {
 220                 struct auxtrace_mmap *mm =
 221                                 &rec->evlist->mmap[i].auxtrace_mmap;
 222
 223                 if (!mm->base)
 224                         continue;
 225
 226                 if (record__auxtrace_mmap_read_snapshot(rec, mm) != 0) {
 227                         rc = -1;
 228                         goto out;
 229                 }
 230         }
 231 out:
 232         return rc;
 233 }
 234
 235 static void record__read_auxtrace_snapshot(struct record *rec)
 236 {
 237         pr_debug("Recording AUX area tracing snapshot\n");
 238         if (record__auxtrace_read_snapshot_all(rec) < 0) {
 239                 auxtrace_snapshot_err = -1;
 240         } else {
 241                 auxtrace_snapshot_err = auxtrace_record__snapshot_finish(rec->itr);
 242                 if (!auxtrace_snapshot_err)
 243                         auxtrace_snapshot_enabled = 1;
 244         }
 245 }
 246
 247 #else
 248
 249 static inline
 250 int record__auxtrace_mmap_read(struct record *rec __maybe_unused,
 251                                struct auxtrace_mmap *mm __maybe_unused)
 252 {
 253         return 0;
 254 }
 255
 256 static inline
 257 void record__read_auxtrace_snapshot(struct record *rec __maybe_unused)
 258 {
 259 }
 260
 261 static inline
 262 int auxtrace_record__snapshot_start(struct auxtrace_record *itr __maybe_unused)
 263 {
 264         return 0;
 265 }
 266
 267 #endif
 268
 269 static int record__open(struct record *rec)
 270 {
 271         char msg[512];
 272         struct perf_evsel *pos;
 273         struct perf_evlist *evlist = rec->evlist;
 274         struct perf_session *session = rec->session;
 275         struct record_opts *opts = &rec->opts;
 276         int rc = 0;
 277
 278         perf_evlist__config(evlist, opts);
 279
 280         evlist__for_each(evlist, pos) {
 281 try_again:
 282                 if (perf_evsel__open(pos, evlist->cpus, evlist->threads) < 0) {
 283                         if (perf_evsel__fallback(pos, errno, msg, sizeof(msg))) {
 284                                 if (verbose)
 285                                         ui__warning("%s\n", msg);
 286                                 goto try_again;
 287                         }
 288
 289                         rc = -errno;
 290                         perf_evsel__open_strerror(pos, &opts->target,
 291                                                   errno, msg, sizeof(msg));
 292                         ui__error("%s\n", msg);
 293                         goto out;
 294                 }
 295         }
 296
 297         if (perf_evlist__apply_filters(evlist, &pos)) {
 298                 error("failed to set filter \"%s\" on event %s with %d (%s)\n",
 299                         pos->filter, perf_evsel__name(pos), errno,
 300                         strerror_r(errno, msg, sizeof(msg)));
 301                 rc = -1;
 302                 goto out;
 303         }
 304
 305         if (perf_evlist__mmap_ex(evlist, opts->mmap_pages, false,
 306                                  opts->auxtrace_mmap_pages,
 307                                  opts->auxtrace_snapshot_mode) < 0) {
 308                 if (errno == EPERM) {
 309                         pr_err("Permission error mapping pages.\n"
 310                                "Consider increasing "
 311                                "/proc/sys/kernel/perf_event_mlock_kb,\n"
 312                                "or try again with a smaller value of -m/--mmap_pages.\n"
 313                                "(current value: %u,%u)\n",
 314                                opts->mmap_pages, opts->auxtrace_mmap_pages);
 315                         rc = -errno;
 316                 } else {
 317                         pr_err("failed to mmap with %d (%s)\n", errno,
 318                                 strerror_r(errno, msg, sizeof(msg)));
 319                         rc = -errno;
 320                 }
 321                 goto out;
 322         }
 323
 324         session->evlist = evlist;
 325         perf_session__set_id_hdr_size(session);
 326 out:
 327         return rc;
 328 }
 329
 330 static int process_sample_event(struct perf_tool *tool,
 331                                 union perf_event *event,
 332                                 struct perf_sample *sample,
 333                                 struct perf_evsel *evsel,
 334                                 struct machine *machine)
 335 {
 336         struct record *rec = container_of(tool, struct record, tool);
 337
 338         rec->samples++;
 339
 340         return build_id__mark_dso_hit(tool, event, sample, evsel, machine);
 341 }
 342
 343 static int process_buildids(struct record *rec)
 344 {
 345         struct perf_data_file *file  = &rec->file;
 346         struct perf_session *session = rec->session;
 347
 348         if (file->size == 0)
 349                 return 0;
 350
 351         /*
 352          * During this process, it'll load kernel map and replace the
 353          * dso->long_name to a real pathname it found.  In this case
 354          * we prefer the vmlinux path like
 355          *   /lib/modules/3.16.4/build/vmlinux
 356          *
 357          * rather than build-id path (in debug directory).
 358          *   $HOME/.debug/.build-id/f0/6e17aa50adf4d00b88925e03775de107611551
 359          */
 360         symbol_conf.ignore_vmlinux_buildid = true;
 361
 362         return perf_session__process_events(session);
 363 }
 364
 365 static void perf_event__synthesize_guest_os(struct machine *machine, void *data)
 366 {
 367         int err;
 368         struct perf_tool *tool = data;
 369         /*
 370          *As for guest kernel when processing subcommand record&report,
 371          *we arrange module mmap prior to guest kernel mmap and trigger
 372          *a preload dso because default guest module symbols are loaded
 373          *from guest kallsyms instead of /lib/modules/XXX/XXX. This
 374          *method is used to avoid symbol missing when the first addr is
 375          *in module instead of in guest kernel.
 376          */
 377         err = perf_event__synthesize_modules(tool, process_synthesized_event,
 378                                              machine);
 379         if (err < 0)
 380                 pr_err("Couldn't record guest kernel [%d]'s reference"
 381                        " relocation symbol.\n", machine->pid);
 382
 383         /*
 384          * We use _stext for guest kernel because guest kernel's /proc/kallsyms
 385          * have no _text sometimes.
 386          */
 387         err = perf_event__synthesize_kernel_mmap(tool, process_synthesized_event,
 388                                                  machine);
 389         if (err < 0)
 390                 pr_err("Couldn't record guest kernel [%d]'s reference"
 391                        " relocation symbol.\n", machine->pid);
 392 }
 393
 394 static struct perf_event_header finished_round_event = {
 395         .size = sizeof(struct perf_event_header),
 396         .type = PERF_RECORD_FINISHED_ROUND,
 397 };
 398
 399 static int record__mmap_read_all(struct record *rec)
 400 {
 401         u64 bytes_written = rec->bytes_written;
 402         int i;
 403         int rc = 0;
 404
 405         for (i = 0; i < rec->evlist->nr_mmaps; i++) {
 406                 struct auxtrace_mmap *mm = &rec->evlist->mmap[i].auxtrace_mmap;
 407
 408                 if (rec->evlist->mmap[i].base) {
 409                         if (record__mmap_read(rec, i) != 0) {
 410                                 rc = -1;
 411                                 goto out;
 412                         }
 413                 }
 414
 415                 if (mm->base && !rec->opts.auxtrace_snapshot_mode &&
 416                     record__auxtrace_mmap_read(rec, mm) != 0) {
 417                         rc = -1;
 418                         goto out;
 419                 }
 420         }
 421
 422         /*
 423          * Mark the round finished in case we wrote
 424          * at least one event.
 425          */
 426         if (bytes_written != rec->bytes_written)
 427                 rc = record__write(rec, &finished_round_event, sizeof(finished_round_event));
 428
 429 out:
 430         return rc;
 431 }
 432
 433 static void record__init_features(struct record *rec)
 434 {
 435         struct perf_session *session = rec->session;
 436         int feat;
 437
 438         for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++)
 439                 perf_header__set_feat(&session->header, feat);
 440
 441         if (rec->no_buildid)
 442                 perf_header__clear_feat(&session->header, HEADER_BUILD_ID);
 443
 444         if (!have_tracepoints(&rec->evlist->entries))
 445                 perf_header__clear_feat(&session->header, HEADER_TRACING_DATA);
 446
 447         if (!rec->opts.branch_stack)
 448                 perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK);
 449
 450         if (!rec->opts.full_auxtrace)
 451                 perf_header__clear_feat(&session->header, HEADER_AUXTRACE);
 452 }
 453
 454 static volatile int workload_exec_errno;
 455
 456 /*
 457  * perf_evlist__prepare_workload will send a SIGUSR1
 458  * if the fork fails, since we asked by setting its
 459  * want_signal to true.
 460  */
 461 static void workload_exec_failed_signal(int signo __maybe_unused,
 462                                         siginfo_t *info,
 463                                         void *ucontext __maybe_unused)
 464 {
 465         workload_exec_errno = info->si_value.sival_int;
 466         done = 1;
 467         child_finished = 1;
 468 }
 469
 470 static void snapshot_sig_handler(int sig);
 471
 472 static int __cmd_record(struct record *rec, int argc, const char **argv)
 473 {
 474         int err;
 475         int status = 0;
 476         unsigned long waking = 0;
 477         const bool forks = argc > 0;
 478         struct machine *machine;
 479         struct perf_tool *tool = &rec->tool;
 480         struct record_opts *opts = &rec->opts;
 481         struct perf_data_file *file = &rec->file;
 482         struct perf_session *session;
 483         bool disabled = false, draining = false;
 484         int fd;
 485
 486         rec->progname = argv[0];
 487
 488         atexit(record__sig_exit);
 489         signal(SIGCHLD, sig_handler);
 490         signal(SIGINT, sig_handler);
 491         signal(SIGTERM, sig_handler);
 492         if (rec->opts.auxtrace_snapshot_mode)
 493                 signal(SIGUSR2, snapshot_sig_handler);
 494         else
 495                 signal(SIGUSR2, SIG_IGN);
 496
 497         session = perf_session__new(file, false, tool);
 498         if (session == NULL) {
 499                 pr_err("Perf session creation failed.\n");
 500                 return -1;
 501         }
 502
 503         fd = perf_data_file__fd(file);
 504         rec->session = session;
 505
 506         record__init_features(rec);
 507
 508         if (forks) {
 509                 err = perf_evlist__prepare_workload(rec->evlist, &opts->target,
 510                                                     argv, file->is_pipe,
 511                                                     workload_exec_failed_signal);
 512                 if (err < 0) {
 513                         pr_err("Couldn't run the workload!\n");
 514                         status = err;
 515                         goto out_delete_session;
 516                 }
 517         }
 518
 519         if (record__open(rec) != 0) {
 520                 err = -1;
 521                 goto out_child;
 522         }
 523
 524         if (!rec->evlist->nr_groups)
 525                 perf_header__clear_feat(&session->header, HEADER_GROUP_DESC);
 526
 527         if (file->is_pipe) {
 528                 err = perf_header__write_pipe(fd);
 529                 if (err < 0)
 530                         goto out_child;
 531         } else {
 532                 err = perf_session__write_header(session, rec->evlist, fd, false);
 533                 if (err < 0)
 534                         goto out_child;
 535         }
 536
 537         if (!rec->no_buildid
 538             && !perf_header__has_feat(&session->header, HEADER_BUILD_ID)) {
 539                 pr_err("Couldn't generate buildids. "
 540                        "Use --no-buildid to profile anyway.\n");
 541                 err = -1;
 542                 goto out_child;
 543         }
 544
 545         machine = &session->machines.host;
 546
 547         if (file->is_pipe) {
 548                 err = perf_event__synthesize_attrs(tool, session,
 549                                                    process_synthesized_event);
 550                 if (err < 0) {
 551                         pr_err("Couldn't synthesize attrs.\n");
 552                         goto out_child;
 553                 }
 554
 555                 if (have_tracepoints(&rec->evlist->entries)) {
 556                         /*
 557                          * FIXME err <= 0 here actually means that
 558                          * there were no tracepoints so its not really
 559                          * an error, just that we don't need to
 560                          * synthesize anything.  We really have to
 561                          * return this more properly and also
 562                          * propagate errors that now are calling die()
 563                          */
 564                         err = perf_event__synthesize_tracing_data(tool, fd, rec->evlist,
 565                                                                   process_synthesized_event);
 566                         if (err <= 0) {
 567                                 pr_err("Couldn't record tracing data.\n");
 568                                 goto out_child;
 569                         }
 570                         rec->bytes_written += err;
 571                 }
 572         }
 573
 574         if (rec->opts.full_auxtrace) {
 575                 err = perf_event__synthesize_auxtrace_info(rec->itr, tool,
 576                                         session, process_synthesized_event);
 577                 if (err)
 578                         goto out_delete_session;
 579         }
 580
 581         err = perf_event__synthesize_kernel_mmap(tool, process_synthesized_event,
 582                                                  machine);
 583         if (err < 0)
 584                 pr_err("Couldn't record kernel reference relocation symbol\n"
 585                        "Symbol resolution may be skewed if relocation was used (e.g. kexec).\n"
 586                        "Check /proc/kallsyms permission or run as root.\n");
 587
 588         err = perf_event__synthesize_modules(tool, process_synthesized_event,
 589                                              machine);
 590         if (err < 0)
 591                 pr_err("Couldn't record kernel module information.\n"
 592                        "Symbol resolution may be skewed if relocation was used (e.g. kexec).\n"
 593                        "Check /proc/modules permission or run as root.\n");
 594
 595         if (perf_guest) {
 596                 machines__process_guests(&session->machines,
 597                                          perf_event__synthesize_guest_os, tool);
 598         }
 599
 600         err = __machine__synthesize_threads(machine, tool, &opts->target, rec->evlist->threads,
 601                                             process_synthesized_event, opts->sample_address,
 602                                             opts->proc_map_timeout);
 603         if (err != 0)
 604                 goto out_child;
 605
 606         if (rec->realtime_prio) {
 607                 struct sched_param param;
 608
 609                 param.sched_priority = rec->realtime_prio;
 610                 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
 611                         pr_err("Could not set realtime priority.\n");
 612                         err = -1;
 613                         goto out_child;
 614                 }
 615         }
 616
 617         /*
 618          * When perf is starting the traced process, all the events
 619          * (apart from group members) have enable_on_exec=1 set,
 620          * so don't spoil it by prematurely enabling them.
 621          */
 622         if (!target__none(&opts->target) && !opts->initial_delay)
 623                 perf_evlist__enable(rec->evlist);
 624
 625         /*
 626          * Let the child rip
 627          */
 628         if (forks)
 629                 perf_evlist__start_workload(rec->evlist);
 630
 631         if (opts->initial_delay) {
 632                 usleep(opts->initial_delay * 1000);
 633                 perf_evlist__enable(rec->evlist);
 634         }
 635
 636         auxtrace_snapshot_enabled = 1;
 637         for (;;) {
 638                 int hits = rec->samples;
 639
 640                 if (record__mmap_read_all(rec) < 0) {
 641                         auxtrace_snapshot_enabled = 0;
 642                         err = -1;
 643                         goto out_child;
 644                 }
 645
 646                 if (auxtrace_record__snapshot_started) {
 647                         auxtrace_record__snapshot_started = 0;
 648                         if (!auxtrace_snapshot_err)
 649                                 record__read_auxtrace_snapshot(rec);
 650                         if (auxtrace_snapshot_err) {
 651                                 pr_err("AUX area tracing snapshot failed\n");
 652                                 err = -1;
 653                                 goto out_child;
 654                         }
 655                 }
 656
 657                 if (hits == rec->samples) {
 658                         if (done || draining)
 659                                 break;
 660                         err = perf_evlist__poll(rec->evlist, -1);
 661                         /*
 662                          * Propagate error, only if there's any. Ignore positive
 663                          * number of returned events and interrupt error.
 664                          */
 665                         if (err > 0 || (err < 0 && errno == EINTR))
 666                                 err = 0;
 667                         waking++;
 668
 669                         if (perf_evlist__filter_pollfd(rec->evlist, POLLERR | POLLHUP) == 0)
 670                                 draining = true;
 671                 }
 672
 673                 /*
 674                  * When perf is starting the traced process, at the end events
 675                  * die with the process and we wait for that. Thus no need to
 676                  * disable events in this case.
 677                  */
 678                 if (done && !disabled && !target__none(&opts->target)) {
 679                         auxtrace_snapshot_enabled = 0;
 680                         perf_evlist__disable(rec->evlist);
 681                         disabled = true;
 682                 }
 683         }
 684         auxtrace_snapshot_enabled = 0;
 685
 686         if (forks && workload_exec_errno) {
 687                 char msg[STRERR_BUFSIZE];
 688                 const char *emsg = strerror_r(workload_exec_errno, msg, sizeof(msg));
 689                 pr_err("Workload failed: %s\n", emsg);
 690                 err = -1;
 691                 goto out_child;
 692         }
 693
 694         if (!quiet)
 695                 fprintf(stderr, "[ perf record: Woken up %ld times to write data ]\n", waking);
 696
 697 out_child:
 698         if (forks) {
 699                 int exit_status;
 700
 701                 if (!child_finished)
 702                         kill(rec->evlist->workload.pid, SIGTERM);
 703
 704                 wait(&exit_status);
 705
 706                 if (err < 0)
 707                         status = err;
 708                 else if (WIFEXITED(exit_status))
 709                         status = WEXITSTATUS(exit_status);
 710                 else if (WIFSIGNALED(exit_status))
 711                         signr = WTERMSIG(exit_status);
 712         } else
 713                 status = err;
 714
 715         /* this will be recalculated during process_buildids() */
 716         rec->samples = 0;
 717
 718         if (!err && !file->is_pipe) {
 719                 rec->session->header.data_size += rec->bytes_written;
 720                 file->size = lseek(perf_data_file__fd(file), 0, SEEK_CUR);
 721
 722                 if (!rec->no_buildid) {
 723                         process_buildids(rec);
 724                         /*
 725                          * We take all buildids when the file contains
 726                          * AUX area tracing data because we do not decode the
 727                          * trace because it would take too long.
 728                          */
 729                         if (rec->opts.full_auxtrace)
 730                                 dsos__hit_all(rec->session);
 731                 }
 732                 perf_session__write_header(rec->session, rec->evlist, fd, true);
 733         }
 734
 735         if (!err && !quiet) {
 736                 char samples[128];
 737
 738                 if (rec->samples && !rec->opts.full_auxtrace)
 739                         scnprintf(samples, sizeof(samples),
 740                                   " (%" PRIu64 " samples)", rec->samples);
 741                 else
 742                         samples[0] = '\0';
 743
 744                 fprintf(stderr, "[ perf record: Captured and wrote %.3f MB %s%s ]\n",
 745                         perf_data_file__size(file) / 1024.0 / 1024.0,
 746                         file->path, samples);
 747         }
 748
 749 out_delete_session:
 750         perf_session__delete(session);
 751         return status;
 752 }
 753
 754 static void callchain_debug(void)
 755 {
 756         static const char *str[CALLCHAIN_MAX] = { "NONE", "FP", "DWARF", "LBR" };
 757
 758         pr_debug("callchain: type %s\n", str[callchain_param.record_mode]);
 759
 760         if (callchain_param.record_mode == CALLCHAIN_DWARF)
 761                 pr_debug("callchain: stack dump size %d\n",
 762                          callchain_param.dump_size);
 763 }
 764
 765 int record_parse_callchain_opt(const struct option *opt,
 766                                const char *arg,
 767                                int unset)
 768 {
 769         int ret;
 770         struct record_opts *record = (struct record_opts *)opt->value;
 771
 772         record->callgraph_set = true;
 773         callchain_param.enabled = !unset;
 774
 775         /* --no-call-graph */
 776         if (unset) {
 777                 callchain_param.record_mode = CALLCHAIN_NONE;
 778                 pr_debug("callchain: disabled\n");
 779                 return 0;
 780         }
 781
 782         ret = parse_callchain_record_opt(arg, &callchain_param);
 783         if (!ret)
 784                 callchain_debug();
 785
 786         return ret;
 787 }
 788
 789 int record_callchain_opt(const struct option *opt,
 790                          const char *arg __maybe_unused,
 791                          int unset __maybe_unused)
 792 {
 793         struct record_opts *record = (struct record_opts *)opt->value;
 794
 795         record->callgraph_set = true;
 796         callchain_param.enabled = true;
 797
 798         if (callchain_param.record_mode == CALLCHAIN_NONE)
 799                 callchain_param.record_mode = CALLCHAIN_FP;
 800
 801         callchain_debug();
 802         return 0;
 803 }
 804
 805 static int perf_record_config(const char *var, const char *value, void *cb)
 806 {
 807         if (!strcmp(var, "record.call-graph"))
 808                 var = "call-graph.record-mode"; /* fall-through */
 809
 810         return perf_default_config(var, value, cb);
 811 }
 812
 813 struct clockid_map {
 814         const char *name;
 815         int clockid;
 816 };
 817
 818 #define CLOCKID_MAP(n, c)       \
 819         { .name = n, .clockid = (c), }
 820
 821 #define CLOCKID_END     { .name = NULL, }
 822
 823
 824 /*
 825  * Add the missing ones, we need to build on many distros...
 826  */
 827 #ifndef CLOCK_MONOTONIC_RAW
 828 #define CLOCK_MONOTONIC_RAW 4
 829 #endif
 830 #ifndef CLOCK_BOOTTIME
 831 #define CLOCK_BOOTTIME 7
 832 #endif
 833 #ifndef CLOCK_TAI
 834 #define CLOCK_TAI 11
 835 #endif
 836
 837 static const struct clockid_map clockids[] = {
 838         /* available for all events, NMI safe */
 839         CLOCKID_MAP("monotonic", CLOCK_MONOTONIC),
 840         CLOCKID_MAP("monotonic_raw", CLOCK_MONOTONIC_RAW),
 841
 842         /* available for some events */
 843         CLOCKID_MAP("realtime", CLOCK_REALTIME),
 844         CLOCKID_MAP("boottime", CLOCK_BOOTTIME),
 845         CLOCKID_MAP("tai", CLOCK_TAI),
 846
 847         /* available for the lazy */
 848         CLOCKID_MAP("mono", CLOCK_MONOTONIC),
 849         CLOCKID_MAP("raw", CLOCK_MONOTONIC_RAW),
 850         CLOCKID_MAP("real", CLOCK_REALTIME),
 851         CLOCKID_MAP("boot", CLOCK_BOOTTIME),
 852
 853         CLOCKID_END,
 854 };
 855
 856 static int parse_clockid(const struct option *opt, const char *str, int unset)
 857 {
 858         struct record_opts *opts = (struct record_opts *)opt->value;
 859         const struct clockid_map *cm;
 860         const char *ostr = str;
 861
 862         if (unset) {
 863                 opts->use_clockid = 0;
 864                 return 0;
 865         }
 866
 867         /* no arg passed */
 868         if (!str)
 869                 return 0;
 870
 871         /* no setting it twice */
 872         if (opts->use_clockid)
 873                 return -1;
 874
 875         opts->use_clockid = true;
 876
 877         /* if its a number, we're done */
 878         if (sscanf(str, "%d", &opts->clockid) == 1)
 879                 return 0;
 880
 881         /* allow a "CLOCK_" prefix to the name */
 882         if (!strncasecmp(str, "CLOCK_", 6))
 883                 str += 6;
 884
 885         for (cm = clockids; cm->name; cm++) {
 886                 if (!strcasecmp(str, cm->name)) {
 887                         opts->clockid = cm->clockid;
 888                         return 0;
 889                 }
 890         }
 891
 892         opts->use_clockid = false;
 893         ui__warning("unknown clockid %s, check man page\n", ostr);
 894         return -1;
 895 }
 896
 897 static int record__parse_mmap_pages(const struct option *opt,
 898                                     const char *str,
 899                                     int unset __maybe_unused)
 900 {
 901         struct record_opts *opts = opt->value;
 902         char *s, *p;
 903         unsigned int mmap_pages;
 904         int ret;
 905
 906         if (!str)
 907                 return -EINVAL;
 908
 909         s = strdup(str);
 910         if (!s)
 911                 return -ENOMEM;
 912
 913         p = strchr(s, ',');
 914         if (p)
 915                 *p = '\0';
 916
 917         if (*s) {
 918                 ret = __perf_evlist__parse_mmap_pages(&mmap_pages, s);
 919                 if (ret)
 920                         goto out_free;
 921                 opts->mmap_pages = mmap_pages;
 922         }
 923
 924         if (!p) {
 925                 ret = 0;
 926                 goto out_free;
 927         }
 928
 929         ret = __perf_evlist__parse_mmap_pages(&mmap_pages, p + 1);
 930         if (ret)
 931                 goto out_free;
 932
 933         opts->auxtrace_mmap_pages = mmap_pages;
 934
 935 out_free:
 936         free(s);
 937         return ret;
 938 }
 939
 940 static const char * const __record_usage[] = {
 941         "perf record [<options>] [<command>]",
 942         "perf record [<options>] -- <command> [<options>]",
 943         NULL
 944 };
 945 const char * const *record_usage = __record_usage;
 946
 947 /*
 948  * XXX Ideally would be local to cmd_record() and passed to a record__new
 949  * because we need to have access to it in record__exit, that is called
 950  * after cmd_record() exits, but since record_options need to be accessible to
 951  * builtin-script, leave it here.
 952  *
 953  * At least we don't ouch it in all the other functions here directly.
 954  *
 955  * Just say no to tons of global variables, sigh.
 956  */
 957 static struct record record = {
 958         .opts = {
 959                 .sample_time         = true,
 960                 .mmap_pages          = UINT_MAX,
 961                 .user_freq           = UINT_MAX,
 962                 .user_interval       = ULLONG_MAX,
 963                 .freq                = 4000,
 964                 .target              = {
 965                         .uses_mmap   = true,
 966                         .default_per_cpu = true,
 967                 },
 968                 .proc_map_timeout     = 500,
 969         },
 970         .tool = {
 971                 .sample         = process_sample_event,
 972                 .fork           = perf_event__process_fork,
 973                 .comm           = perf_event__process_comm,
 974                 .mmap           = perf_event__process_mmap,
 975                 .mmap2          = perf_event__process_mmap2,
 976         },
 977 };
 978
 979 #define CALLCHAIN_HELP "setup and enables call-graph (stack chain/backtrace) recording: "
 980
 981 #ifdef HAVE_DWARF_UNWIND_SUPPORT
 982 const char record_callchain_help[] = CALLCHAIN_HELP "fp dwarf lbr";
 983 #else
 984 const char record_callchain_help[] = CALLCHAIN_HELP "fp lbr";
 985 #endif
 986
 987 /*
 988  * XXX Will stay a global variable till we fix builtin-script.c to stop messing
 989  * with it and switch to use the library functions in perf_evlist that came
 990  * from builtin-record.c, i.e. use record_opts,
 991  * perf_evlist__prepare_workload, etc instead of fork+exec'in 'perf record',
 992  * using pipes, etc.
 993  */
 994 struct option __record_options[] = {
 995         OPT_CALLBACK('e', "event", &record.evlist, "event",
 996                      "event selector. use 'perf list' to list available events",
 997                      parse_events_option),
 998         OPT_CALLBACK(0, "filter", &record.evlist, "filter",
 999                      "event filter", parse_filter),
1000         OPT_CALLBACK_NOOPT(0, "exclude-perf", &record.evlist,
1001                            NULL, "don't record events from perf itself",
1002                            exclude_perf),
1003         OPT_STRING('p', "pid", &record.opts.target.pid, "pid",
1004                     "record events on existing process id"),
1005         OPT_STRING('t', "tid", &record.opts.target.tid, "tid",
1006                     "record events on existing thread id"),
1007         OPT_INTEGER('r', "realtime", &record.realtime_prio,
1008                     "collect data with this RT SCHED_FIFO priority"),
1009         OPT_BOOLEAN(0, "no-buffering", &record.opts.no_buffering,
1010                     "collect data without buffering"),
1011         OPT_BOOLEAN('R', "raw-samples", &record.opts.raw_samples,
1012                     "collect raw sample records from all opened counters"),
1013         OPT_BOOLEAN('a', "all-cpus", &record.opts.target.system_wide,
1014                             "system-wide collection from all CPUs"),
1015         OPT_STRING('C', "cpu", &record.opts.target.cpu_list, "cpu",
1016                     "list of cpus to monitor"),
1017         OPT_U64('c', "count", &record.opts.user_interval, "event period to sample"),
1018         OPT_STRING('o', "output", &record.file.path, "file",
1019                     "output file name"),
1020         OPT_BOOLEAN_SET('i', "no-inherit", &record.opts.no_inherit,
1021                         &record.opts.no_inherit_set,
1022                         "child tasks do not inherit counters"),
1023         OPT_UINTEGER('F', "freq", &record.opts.user_freq, "profile at this frequency"),
1024         OPT_CALLBACK('m', "mmap-pages", &record.opts, "pages[,pages]",
1025                      "number of mmap data pages and AUX area tracing mmap pages",
1026                      record__parse_mmap_pages),
1027         OPT_BOOLEAN(0, "group", &record.opts.group,
1028                     "put the counters into a counter group"),
1029         OPT_CALLBACK_NOOPT('g', NULL, &record.opts,
1030                            NULL, "enables call-graph recording" ,
1031                            &record_callchain_opt),
1032         OPT_CALLBACK(0, "call-graph", &record.opts,
1033                      "mode[,dump_size]", record_callchain_help,
1034                      &record_parse_callchain_opt),
1035         OPT_INCR('v', "verbose", &verbose,
1036                     "be more verbose (show counter open errors, etc)"),
1037         OPT_BOOLEAN('q', "quiet", &quiet, "don't print any message"),
1038         OPT_BOOLEAN('s', "stat", &record.opts.inherit_stat,
1039                     "per thread counts"),
1040         OPT_BOOLEAN('d', "data", &record.opts.sample_address, "Record the sample addresses"),
1041         OPT_BOOLEAN_SET('T', "timestamp", &record.opts.sample_time,
1042                         &record.opts.sample_time_set,
1043                         "Record the sample timestamps"),
1044         OPT_BOOLEAN('P', "period", &record.opts.period, "Record the sample period"),
1045         OPT_BOOLEAN('n', "no-samples", &record.opts.no_samples,
1046                     "don't sample"),
1047         OPT_BOOLEAN('N', "no-buildid-cache", &record.no_buildid_cache,
1048                     "do not update the buildid cache"),
1049         OPT_BOOLEAN('B', "no-buildid", &record.no_buildid,
1050                     "do not collect buildids in perf.data"),
1051         OPT_CALLBACK('G', "cgroup", &record.evlist, "name",
1052                      "monitor event in cgroup name only",
1053                      parse_cgroups),
1054         OPT_UINTEGER('D', "delay", &record.opts.initial_delay,
1055                   "ms to wait before starting measurement after program start"),
1056         OPT_STRING('u', "uid", &record.opts.target.uid_str, "user",
1057                    "user to profile"),
1058
1059         OPT_CALLBACK_NOOPT('b', "branch-any", &record.opts.branch_stack,
1060                      "branch any", "sample any taken branches",
1061                      parse_branch_stack),
1062
1063         OPT_CALLBACK('j', "branch-filter", &record.opts.branch_stack,
1064                      "branch filter mask", "branch stack filter modes",
1065                      parse_branch_stack),
1066         OPT_BOOLEAN('W', "weight", &record.opts.sample_weight,
1067                     "sample by weight (on special events only)"),
1068         OPT_BOOLEAN(0, "transaction", &record.opts.sample_transaction,
1069                     "sample transaction flags (special events only)"),
1070         OPT_BOOLEAN(0, "per-thread", &record.opts.target.per_thread,
1071                     "use per-thread mmaps"),
1072         OPT_BOOLEAN('I', "intr-regs", &record.opts.sample_intr_regs,
1073                     "Sample machine registers on interrupt"),
1074         OPT_BOOLEAN(0, "running-time", &record.opts.running_time,
1075                     "Record running/enabled time of read (:S) events"),
1076         OPT_CALLBACK('k', "clockid", &record.opts,
1077         "clockid", "clockid to use for events, see clock_gettime()",
1078         parse_clockid),
1079         OPT_STRING_OPTARG('S', "snapshot", &record.opts.auxtrace_snapshot_opts,
1080                           "opts", "AUX area tracing Snapshot Mode", ""),
1081         OPT_UINTEGER(0, "proc-map-timeout", &record.opts.proc_map_timeout,
1082                         "per thread proc mmap processing timeout in ms"),
1083         OPT_BOOLEAN(0, "switch-events", &record.opts.record_switch_events,
1084                     "Record context switch events"),
1085         OPT_END()
1086 };
1087
1088 struct option *record_options = __record_options;
1089
1090 int cmd_record(int argc, const char **argv, const char *prefix __maybe_unused)
1091 {
1092         int err;
1093         struct record *rec = &record;
1094         char errbuf[BUFSIZ];
1095
1096         rec->evlist = perf_evlist__new();
1097         if (rec->evlist == NULL)
1098                 return -ENOMEM;
1099
1100         perf_config(perf_record_config, rec);
1101
1102         argc = parse_options(argc, argv, record_options, record_usage,
1103                             PARSE_OPT_STOP_AT_NON_OPTION);
1104         if (!argc && target__none(&rec->opts.target))
1105                 usage_with_options(record_usage, record_options);
1106
1107         if (nr_cgroups && !rec->opts.target.system_wide) {
1108                 ui__error("cgroup monitoring only available in"
1109                           " system-wide mode\n");
1110                 usage_with_options(record_usage, record_options);
1111         }
1112         if (rec->opts.record_switch_events &&
1113             !perf_can_record_switch_events()) {
1114                 ui__error("kernel does not support recording context switch events (--switch-events option)\n");
1115                 usage_with_options(record_usage, record_options);
1116         }
1117
1118         if (!rec->itr) {
1119                 rec->itr = auxtrace_record__init(rec->evlist, &err);
1120                 if (err)
1121                         return err;
1122         }
1123
1124         err = auxtrace_parse_snapshot_options(rec->itr, &rec->opts,
1125                                               rec->opts.auxtrace_snapshot_opts);
1126         if (err)
1127                 return err;
1128
1129         err = -ENOMEM;
1130
1131         symbol__init(NULL);
1132
1133         if (symbol_conf.kptr_restrict)
1134                 pr_warning(
1135 "WARNING: Kernel address maps (/proc/{kallsyms,modules}) are restricted,\n"
1136 "check /proc/sys/kernel/kptr_restrict.\n\n"
1137 "Samples in kernel functions may not be resolved if a suitable vmlinux\n"
1138 "file is not found in the buildid cache or in the vmlinux path.\n\n"
1139 "Samples in kernel modules won't be resolved at all.\n\n"
1140 "If some relocation was applied (e.g. kexec) symbols may be misresolved\n"
1141 "even with a suitable vmlinux or kallsyms file.\n\n");
1142
1143         if (rec->no_buildid_cache || rec->no_buildid)
1144                 disable_buildid_cache();
1145
1146         if (rec->evlist->nr_entries == 0 &&
1147             perf_evlist__add_default(rec->evlist) < 0) {
1148                 pr_err("Not enough memory for event selector list\n");
1149                 goto out_symbol_exit;
1150         }
1151
1152         if (rec->opts.target.tid && !rec->opts.no_inherit_set)
1153                 rec->opts.no_inherit = true;
1154
1155         err = target__validate(&rec->opts.target);
1156         if (err) {
1157                 target__strerror(&rec->opts.target, err, errbuf, BUFSIZ);
1158                 ui__warning("%s", errbuf);
1159         }
1160
1161         err = target__parse_uid(&rec->opts.target);
1162         if (err) {
1163                 int saved_errno = errno;
1164
1165                 target__strerror(&rec->opts.target, err, errbuf, BUFSIZ);
1166                 ui__error("%s", errbuf);
1167
1168                 err = -saved_errno;
1169                 goto out_symbol_exit;
1170         }
1171
1172         err = -ENOMEM;
1173         if (perf_evlist__create_maps(rec->evlist, &rec->opts.target) < 0)
1174                 usage_with_options(record_usage, record_options);
1175
1176         err = auxtrace_record__options(rec->itr, rec->evlist, &rec->opts);
1177         if (err)
1178                 goto out_symbol_exit;
1179
1180         if (record_opts__config(&rec->opts)) {
1181                 err = -EINVAL;
1182                 goto out_symbol_exit;
1183         }
1184
1185         err = __cmd_record(&record, argc, argv);
1186 out_symbol_exit:
1187         perf_evlist__delete(rec->evlist);
1188         symbol__exit();
1189         auxtrace_record__free(rec->itr);
1190         return err;
1191 }
1192
1193 static void snapshot_sig_handler(int sig __maybe_unused)
1194 {
1195         if (!auxtrace_snapshot_enabled)
1196                 return;
1197         auxtrace_snapshot_enabled = 0;
1198         auxtrace_snapshot_err = auxtrace_record__snapshot_start(record.itr);
1199         auxtrace_record__snapshot_started = 1;
1200 }