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 #define _FILE_OFFSET_BITS 64
   9
  10 #include "builtin.h"
  11
  12 #include "perf.h"
  13
  14 #include "util/build-id.h"
  15 #include "util/util.h"
  16 #include "util/parse-options.h"
  17 #include "util/parse-events.h"
  18
  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
  30 #include <unistd.h>
  31 #include <sched.h>
  32 #include <sys/mman.h>
  33
  34 #ifndef HAVE_ON_EXIT
  35 #ifndef ATEXIT_MAX
  36 #define ATEXIT_MAX 32
  37 #endif
  38 static int __on_exit_count = 0;
  39 typedef void (*on_exit_func_t) (int, void *);
  40 static on_exit_func_t __on_exit_funcs[ATEXIT_MAX];
  41 static void *__on_exit_args[ATEXIT_MAX];
  42 static int __exitcode = 0;
  43 static void __handle_on_exit_funcs(void);
  44 static int on_exit(on_exit_func_t function, void *arg);
  45 #define exit(x) (exit)(__exitcode = (x))
  46
  47 static int on_exit(on_exit_func_t function, void *arg)
  48 {
  49         if (__on_exit_count == ATEXIT_MAX)
  50                 return -ENOMEM;
  51         else if (__on_exit_count == 0)
  52                 atexit(__handle_on_exit_funcs);
  53         __on_exit_funcs[__on_exit_count] = function;
  54         __on_exit_args[__on_exit_count++] = arg;
  55         return 0;
  56 }
  57
  58 static void __handle_on_exit_funcs(void)
  59 {
  60         int i;
  61         for (i = 0; i < __on_exit_count; i++)
  62                 __on_exit_funcs[i] (__exitcode, __on_exit_args[i]);
  63 }
  64 #endif
  65
  66 enum write_mode_t {
  67         WRITE_FORCE,
  68         WRITE_APPEND
  69 };
  70
  71 struct perf_record {
  72         struct perf_tool        tool;
  73         struct perf_record_opts opts;
  74         u64                     bytes_written;
  75         const char              *output_name;
  76         struct perf_evlist      *evlist;
  77         struct perf_session     *session;
  78         const char              *progname;
  79         int                     output;
  80         unsigned int            page_size;
  81         int                     realtime_prio;
  82         enum write_mode_t       write_mode;
  83         bool                    no_buildid;
  84         bool                    no_buildid_cache;
  85         bool                    force;
  86         bool                    file_new;
  87         bool                    append_file;
  88         long                    samples;
  89         off_t                   post_processing_offset;
  90 };
  91
  92 static void advance_output(struct perf_record *rec, size_t size)
  93 {
  94         rec->bytes_written += size;
  95 }
  96
  97 static int write_output(struct perf_record *rec, void *buf, size_t size)
  98 {
  99         while (size) {
 100                 int ret = write(rec->output, buf, size);
 101
 102                 if (ret < 0) {
 103                         pr_err("failed to write\n");
 104                         return -1;
 105                 }
 106
 107                 size -= ret;
 108                 buf += ret;
 109
 110                 rec->bytes_written += ret;
 111         }
 112
 113         return 0;
 114 }
 115
 116 static int process_synthesized_event(struct perf_tool *tool,
 117                                      union perf_event *event,
 118                                      struct perf_sample *sample __maybe_unused,
 119                                      struct machine *machine __maybe_unused)
 120 {
 121         struct perf_record *rec = container_of(tool, struct perf_record, tool);
 122         if (write_output(rec, event, event->header.size) < 0)
 123                 return -1;
 124
 125         return 0;
 126 }
 127
 128 static int perf_record__mmap_read(struct perf_record *rec,
 129                                    struct perf_mmap *md)
 130 {
 131         unsigned int head = perf_mmap__read_head(md);
 132         unsigned int old = md->prev;
 133         unsigned char *data = md->base + rec->page_size;
 134         unsigned long size;
 135         void *buf;
 136         int rc = 0;
 137
 138         if (old == head)
 139                 return 0;
 140
 141         rec->samples++;
 142
 143         size = head - old;
 144
 145         if ((old & md->mask) + size != (head & md->mask)) {
 146                 buf = &data[old & md->mask];
 147                 size = md->mask + 1 - (old & md->mask);
 148                 old += size;
 149
 150                 if (write_output(rec, buf, size) < 0) {
 151                         rc = -1;
 152                         goto out;
 153                 }
 154         }
 155
 156         buf = &data[old & md->mask];
 157         size = head - old;
 158         old += size;
 159
 160         if (write_output(rec, buf, size) < 0) {
 161                 rc = -1;
 162                 goto out;
 163         }
 164
 165         md->prev = old;
 166         perf_mmap__write_tail(md, old);
 167
 168 out:
 169         return rc;
 170 }
 171
 172 static volatile int done = 0;
 173 static volatile int signr = -1;
 174 static volatile int child_finished = 0;
 175
 176 static void sig_handler(int sig)
 177 {
 178         if (sig == SIGCHLD)
 179                 child_finished = 1;
 180
 181         done = 1;
 182         signr = sig;
 183 }
 184
 185 static void perf_record__sig_exit(int exit_status __maybe_unused, void *arg)
 186 {
 187         struct perf_record *rec = arg;
 188         int status;
 189
 190         if (rec->evlist->workload.pid > 0) {
 191                 if (!child_finished)
 192                         kill(rec->evlist->workload.pid, SIGTERM);
 193
 194                 wait(&status);
 195                 if (WIFSIGNALED(status))
 196                         psignal(WTERMSIG(status), rec->progname);
 197         }
 198
 199         if (signr == -1 || signr == SIGUSR1)
 200                 return;
 201
 202         signal(signr, SIG_DFL);
 203         kill(getpid(), signr);
 204 }
 205
 206 static bool perf_evlist__equal(struct perf_evlist *evlist,
 207                                struct perf_evlist *other)
 208 {
 209         struct perf_evsel *pos, *pair;
 210
 211         if (evlist->nr_entries != other->nr_entries)
 212                 return false;
 213
 214         pair = perf_evlist__first(other);
 215
 216         list_for_each_entry(pos, &evlist->entries, node) {
 217                 if (memcmp(&pos->attr, &pair->attr, sizeof(pos->attr) != 0))
 218                         return false;
 219                 pair = perf_evsel__next(pair);
 220         }
 221
 222         return true;
 223 }
 224
 225 static int perf_record__open(struct perf_record *rec)
 226 {
 227         char msg[512];
 228         struct perf_evsel *pos;
 229         struct perf_evlist *evlist = rec->evlist;
 230         struct perf_session *session = rec->session;
 231         struct perf_record_opts *opts = &rec->opts;
 232         int rc = 0;
 233
 234         perf_evlist__config(evlist, opts);
 235
 236         list_for_each_entry(pos, &evlist->entries, node) {
 237 try_again:
 238                 if (perf_evsel__open(pos, evlist->cpus, evlist->threads) < 0) {
 239                         if (perf_evsel__fallback(pos, errno, msg, sizeof(msg))) {
 240                                 if (verbose)
 241                                         ui__warning("%s\n", msg);
 242                                 goto try_again;
 243                         }
 244
 245                         rc = -errno;
 246                         perf_evsel__open_strerror(pos, &opts->target,
 247                                                   errno, msg, sizeof(msg));
 248                         ui__error("%s\n", msg);
 249                         goto out;
 250                 }
 251         }
 252
 253         if (perf_evlist__apply_filters(evlist)) {
 254                 error("failed to set filter with %d (%s)\n", errno,
 255                         strerror(errno));
 256                 rc = -1;
 257                 goto out;
 258         }
 259
 260         if (perf_evlist__mmap(evlist, opts->mmap_pages, false) < 0) {
 261                 if (errno == EPERM) {
 262                         pr_err("Permission error mapping pages.\n"
 263                                "Consider increasing "
 264                                "/proc/sys/kernel/perf_event_mlock_kb,\n"
 265                                "or try again with a smaller value of -m/--mmap_pages.\n"
 266                                "(current value: %d)\n", opts->mmap_pages);
 267                         rc = -errno;
 268                 } else if (!is_power_of_2(opts->mmap_pages) &&
 269                            (opts->mmap_pages != UINT_MAX)) {
 270                         pr_err("--mmap_pages/-m value must be a power of two.");
 271                         rc = -EINVAL;
 272                 } else {
 273                         pr_err("failed to mmap with %d (%s)\n", errno, strerror(errno));
 274                         rc = -errno;
 275                 }
 276                 goto out;
 277         }
 278
 279         if (rec->file_new)
 280                 session->evlist = evlist;
 281         else {
 282                 if (!perf_evlist__equal(session->evlist, evlist)) {
 283                         fprintf(stderr, "incompatible append\n");
 284                         rc = -1;
 285                         goto out;
 286                 }
 287         }
 288
 289         perf_session__set_id_hdr_size(session);
 290 out:
 291         return rc;
 292 }
 293
 294 static int process_buildids(struct perf_record *rec)
 295 {
 296         u64 size = lseek(rec->output, 0, SEEK_CUR);
 297
 298         if (size == 0)
 299                 return 0;
 300
 301         rec->session->fd = rec->output;
 302         return __perf_session__process_events(rec->session, rec->post_processing_offset,
 303                                               size - rec->post_processing_offset,
 304                                               size, &build_id__mark_dso_hit_ops);
 305 }
 306
 307 static void perf_record__exit(int status, void *arg)
 308 {
 309         struct perf_record *rec = arg;
 310
 311         if (status != 0)
 312                 return;
 313
 314         if (!rec->opts.pipe_output) {
 315                 rec->session->header.data_size += rec->bytes_written;
 316
 317                 if (!rec->no_buildid)
 318                         process_buildids(rec);
 319                 perf_session__write_header(rec->session, rec->evlist,
 320                                            rec->output, true);
 321                 perf_session__delete(rec->session);
 322                 perf_evlist__delete(rec->evlist);
 323                 symbol__exit();
 324         }
 325 }
 326
 327 static void perf_event__synthesize_guest_os(struct machine *machine, void *data)
 328 {
 329         int err;
 330         struct perf_tool *tool = data;
 331         /*
 332          *As for guest kernel when processing subcommand record&report,
 333          *we arrange module mmap prior to guest kernel mmap and trigger
 334          *a preload dso because default guest module symbols are loaded
 335          *from guest kallsyms instead of /lib/modules/XXX/XXX. This
 336          *method is used to avoid symbol missing when the first addr is
 337          *in module instead of in guest kernel.
 338          */
 339         err = perf_event__synthesize_modules(tool, process_synthesized_event,
 340                                              machine);
 341         if (err < 0)
 342                 pr_err("Couldn't record guest kernel [%d]'s reference"
 343                        " relocation symbol.\n", machine->pid);
 344
 345         /*
 346          * We use _stext for guest kernel because guest kernel's /proc/kallsyms
 347          * have no _text sometimes.
 348          */
 349         err = perf_event__synthesize_kernel_mmap(tool, process_synthesized_event,
 350                                                  machine, "_text");
 351         if (err < 0)
 352                 err = perf_event__synthesize_kernel_mmap(tool, process_synthesized_event,
 353                                                          machine, "_stext");
 354         if (err < 0)
 355                 pr_err("Couldn't record guest kernel [%d]'s reference"
 356                        " relocation symbol.\n", machine->pid);
 357 }
 358
 359 static struct perf_event_header finished_round_event = {
 360         .size = sizeof(struct perf_event_header),
 361         .type = PERF_RECORD_FINISHED_ROUND,
 362 };
 363
 364 static int perf_record__mmap_read_all(struct perf_record *rec)
 365 {
 366         int i;
 367         int rc = 0;
 368
 369         for (i = 0; i < rec->evlist->nr_mmaps; i++) {
 370                 if (rec->evlist->mmap[i].base) {
 371                         if (perf_record__mmap_read(rec, &rec->evlist->mmap[i]) != 0) {
 372                                 rc = -1;
 373                                 goto out;
 374                         }
 375                 }
 376         }
 377
 378         if (perf_header__has_feat(&rec->session->header, HEADER_TRACING_DATA))
 379                 rc = write_output(rec, &finished_round_event,
 380                                   sizeof(finished_round_event));
 381
 382 out:
 383         return rc;
 384 }
 385
 386 static int __cmd_record(struct perf_record *rec, int argc, const char **argv)
 387 {
 388         struct stat st;
 389         int flags;
 390         int err, output, feat;
 391         unsigned long waking = 0;
 392         const bool forks = argc > 0;
 393         struct machine *machine;
 394         struct perf_tool *tool = &rec->tool;
 395         struct perf_record_opts *opts = &rec->opts;
 396         struct perf_evlist *evsel_list = rec->evlist;
 397         const char *output_name = rec->output_name;
 398         struct perf_session *session;
 399         bool disabled = false;
 400
 401         rec->progname = argv[0];
 402
 403         rec->page_size = sysconf(_SC_PAGE_SIZE);
 404
 405         on_exit(perf_record__sig_exit, rec);
 406         signal(SIGCHLD, sig_handler);
 407         signal(SIGINT, sig_handler);
 408         signal(SIGUSR1, sig_handler);
 409
 410         if (!output_name) {
 411                 if (!fstat(STDOUT_FILENO, &st) && S_ISFIFO(st.st_mode))
 412                         opts->pipe_output = true;
 413                 else
 414                         rec->output_name = output_name = "perf.data";
 415         }
 416         if (output_name) {
 417                 if (!strcmp(output_name, "-"))
 418                         opts->pipe_output = true;
 419                 else if (!stat(output_name, &st) && st.st_size) {
 420                         if (rec->write_mode == WRITE_FORCE) {
 421                                 char oldname[PATH_MAX];
 422                                 snprintf(oldname, sizeof(oldname), "%s.old",
 423                                          output_name);
 424                                 unlink(oldname);
 425                                 rename(output_name, oldname);
 426                         }
 427                 } else if (rec->write_mode == WRITE_APPEND) {
 428                         rec->write_mode = WRITE_FORCE;
 429                 }
 430         }
 431
 432         flags = O_CREAT|O_RDWR;
 433         if (rec->write_mode == WRITE_APPEND)
 434                 rec->file_new = 0;
 435         else
 436                 flags |= O_TRUNC;
 437
 438         if (opts->pipe_output)
 439                 output = STDOUT_FILENO;
 440         else
 441                 output = open(output_name, flags, S_IRUSR | S_IWUSR);
 442         if (output < 0) {
 443                 perror("failed to create output file");
 444                 return -1;
 445         }
 446
 447         rec->output = output;
 448
 449         session = perf_session__new(output_name, O_WRONLY,
 450                                     rec->write_mode == WRITE_FORCE, false, NULL);
 451         if (session == NULL) {
 452                 pr_err("Not enough memory for reading perf file header\n");
 453                 return -1;
 454         }
 455
 456         rec->session = session;
 457
 458         for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++)
 459                 perf_header__set_feat(&session->header, feat);
 460
 461         if (rec->no_buildid)
 462                 perf_header__clear_feat(&session->header, HEADER_BUILD_ID);
 463
 464         if (!have_tracepoints(&evsel_list->entries))
 465                 perf_header__clear_feat(&session->header, HEADER_TRACING_DATA);
 466
 467         if (!rec->opts.branch_stack)
 468                 perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK);
 469
 470         if (!rec->file_new) {
 471                 err = perf_session__read_header(session, output);
 472                 if (err < 0)
 473                         goto out_delete_session;
 474         }
 475
 476         if (forks) {
 477                 err = perf_evlist__prepare_workload(evsel_list, &opts->target,
 478                                                     argv, opts->pipe_output,
 479                                                     true);
 480                 if (err < 0) {
 481                         pr_err("Couldn't run the workload!\n");
 482                         goto out_delete_session;
 483                 }
 484         }
 485
 486         if (perf_record__open(rec) != 0) {
 487                 err = -1;
 488                 goto out_delete_session;
 489         }
 490
 491         if (!evsel_list->nr_groups)
 492                 perf_header__clear_feat(&session->header, HEADER_GROUP_DESC);
 493
 494         /*
 495          * perf_session__delete(session) will be called at perf_record__exit()
 496          */
 497         on_exit(perf_record__exit, rec);
 498
 499         if (opts->pipe_output) {
 500                 err = perf_header__write_pipe(output);
 501                 if (err < 0)
 502                         goto out_delete_session;
 503         } else if (rec->file_new) {
 504                 err = perf_session__write_header(session, evsel_list,
 505                                                  output, false);
 506                 if (err < 0)
 507                         goto out_delete_session;
 508         }
 509
 510         if (!rec->no_buildid
 511             && !perf_header__has_feat(&session->header, HEADER_BUILD_ID)) {
 512                 pr_err("Couldn't generate buildids. "
 513                        "Use --no-buildid to profile anyway.\n");
 514                 err = -1;
 515                 goto out_delete_session;
 516         }
 517
 518         rec->post_processing_offset = lseek(output, 0, SEEK_CUR);
 519
 520         machine = &session->machines.host;
 521
 522         if (opts->pipe_output) {
 523                 err = perf_event__synthesize_attrs(tool, session,
 524                                                    process_synthesized_event);
 525                 if (err < 0) {
 526                         pr_err("Couldn't synthesize attrs.\n");
 527                         goto out_delete_session;
 528                 }
 529
 530                 err = perf_event__synthesize_event_types(tool, process_synthesized_event,
 531                                                          machine);
 532                 if (err < 0) {
 533                         pr_err("Couldn't synthesize event_types.\n");
 534                         goto out_delete_session;
 535                 }
 536
 537                 if (have_tracepoints(&evsel_list->entries)) {
 538                         /*
 539                          * FIXME err <= 0 here actually means that
 540                          * there were no tracepoints so its not really
 541                          * an error, just that we don't need to
 542                          * synthesize anything.  We really have to
 543                          * return this more properly and also
 544                          * propagate errors that now are calling die()
 545                          */
 546                         err = perf_event__synthesize_tracing_data(tool, output, evsel_list,
 547                                                                   process_synthesized_event);
 548                         if (err <= 0) {
 549                                 pr_err("Couldn't record tracing data.\n");
 550                                 goto out_delete_session;
 551                         }
 552                         advance_output(rec, err);
 553                 }
 554         }
 555
 556         err = perf_event__synthesize_kernel_mmap(tool, process_synthesized_event,
 557                                                  machine, "_text");
 558         if (err < 0)
 559                 err = perf_event__synthesize_kernel_mmap(tool, process_synthesized_event,
 560                                                          machine, "_stext");
 561         if (err < 0)
 562                 pr_err("Couldn't record kernel reference relocation symbol\n"
 563                        "Symbol resolution may be skewed if relocation was used (e.g. kexec).\n"
 564                        "Check /proc/kallsyms permission or run as root.\n");
 565
 566         err = perf_event__synthesize_modules(tool, process_synthesized_event,
 567                                              machine);
 568         if (err < 0)
 569                 pr_err("Couldn't record kernel module information.\n"
 570                        "Symbol resolution may be skewed if relocation was used (e.g. kexec).\n"
 571                        "Check /proc/modules permission or run as root.\n");
 572
 573         if (perf_guest) {
 574                 machines__process_guests(&session->machines,
 575                                          perf_event__synthesize_guest_os, tool);
 576         }
 577
 578         if (perf_target__has_task(&opts->target))
 579                 err = perf_event__synthesize_thread_map(tool, evsel_list->threads,
 580                                                   process_synthesized_event,
 581                                                   machine);
 582         else if (perf_target__has_cpu(&opts->target))
 583                 err = perf_event__synthesize_threads(tool, process_synthesized_event,
 584                                                machine);
 585         else /* command specified */
 586                 err = 0;
 587
 588         if (err != 0)
 589                 goto out_delete_session;
 590
 591         if (rec->realtime_prio) {
 592                 struct sched_param param;
 593
 594                 param.sched_priority = rec->realtime_prio;
 595                 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
 596                         pr_err("Could not set realtime priority.\n");
 597                         err = -1;
 598                         goto out_delete_session;
 599                 }
 600         }
 601
 602         /*
 603          * When perf is starting the traced process, all the events
 604          * (apart from group members) have enable_on_exec=1 set,
 605          * so don't spoil it by prematurely enabling them.
 606          */
 607         if (!perf_target__none(&opts->target))
 608                 perf_evlist__enable(evsel_list);
 609
 610         /*
 611          * Let the child rip
 612          */
 613         if (forks)
 614                 perf_evlist__start_workload(evsel_list);
 615
 616         for (;;) {
 617                 int hits = rec->samples;
 618
 619                 if (perf_record__mmap_read_all(rec) < 0) {
 620                         err = -1;
 621                         goto out_delete_session;
 622                 }
 623
 624                 if (hits == rec->samples) {
 625                         if (done)
 626                                 break;
 627                         err = poll(evsel_list->pollfd, evsel_list->nr_fds, -1);
 628                         waking++;
 629                 }
 630
 631                 /*
 632                  * When perf is starting the traced process, at the end events
 633                  * die with the process and we wait for that. Thus no need to
 634                  * disable events in this case.
 635                  */
 636                 if (done && !disabled && !perf_target__none(&opts->target)) {
 637                         perf_evlist__disable(evsel_list);
 638                         disabled = true;
 639                 }
 640         }
 641
 642         if (quiet || signr == SIGUSR1)
 643                 return 0;
 644
 645         fprintf(stderr, "[ perf record: Woken up %ld times to write data ]\n", waking);
 646
 647         /*
 648          * Approximate RIP event size: 24 bytes.
 649          */
 650         fprintf(stderr,
 651                 "[ perf record: Captured and wrote %.3f MB %s (~%" PRIu64 " samples) ]\n",
 652                 (double)rec->bytes_written / 1024.0 / 1024.0,
 653                 output_name,
 654                 rec->bytes_written / 24);
 655
 656         return 0;
 657
 658 out_delete_session:
 659         perf_session__delete(session);
 660         return err;
 661 }
 662
 663 #define BRANCH_OPT(n, m) \
 664         { .name = n, .mode = (m) }
 665
 666 #define BRANCH_END { .name = NULL }
 667
 668 struct branch_mode {
 669         const char *name;
 670         int mode;
 671 };
 672
 673 static const struct branch_mode branch_modes[] = {
 674         BRANCH_OPT("u", PERF_SAMPLE_BRANCH_USER),
 675         BRANCH_OPT("k", PERF_SAMPLE_BRANCH_KERNEL),
 676         BRANCH_OPT("hv", PERF_SAMPLE_BRANCH_HV),
 677         BRANCH_OPT("any", PERF_SAMPLE_BRANCH_ANY),
 678         BRANCH_OPT("any_call", PERF_SAMPLE_BRANCH_ANY_CALL),
 679         BRANCH_OPT("any_ret", PERF_SAMPLE_BRANCH_ANY_RETURN),
 680         BRANCH_OPT("ind_call", PERF_SAMPLE_BRANCH_IND_CALL),
 681         BRANCH_END
 682 };
 683
 684 static int
 685 parse_branch_stack(const struct option *opt, const char *str, int unset)
 686 {
 687 #define ONLY_PLM \
 688         (PERF_SAMPLE_BRANCH_USER        |\
 689          PERF_SAMPLE_BRANCH_KERNEL      |\
 690          PERF_SAMPLE_BRANCH_HV)
 691
 692         uint64_t *mode = (uint64_t *)opt->value;
 693         const struct branch_mode *br;
 694         char *s, *os = NULL, *p;
 695         int ret = -1;
 696
 697         if (unset)
 698                 return 0;
 699
 700         /*
 701          * cannot set it twice, -b + --branch-filter for instance
 702          */
 703         if (*mode)
 704                 return -1;
 705
 706         /* str may be NULL in case no arg is passed to -b */
 707         if (str) {
 708                 /* because str is read-only */
 709                 s = os = strdup(str);
 710                 if (!s)
 711                         return -1;
 712
 713                 for (;;) {
 714                         p = strchr(s, ',');
 715                         if (p)
 716                                 *p = '\0';
 717
 718                         for (br = branch_modes; br->name; br++) {
 719                                 if (!strcasecmp(s, br->name))
 720                                         break;
 721                         }
 722                         if (!br->name) {
 723                                 ui__warning("unknown branch filter %s,"
 724                                             " check man page\n", s);
 725                                 goto error;
 726                         }
 727
 728                         *mode |= br->mode;
 729
 730                         if (!p)
 731                                 break;
 732
 733                         s = p + 1;
 734                 }
 735         }
 736         ret = 0;
 737
 738         /* default to any branch */
 739         if ((*mode & ~ONLY_PLM) == 0) {
 740                 *mode = PERF_SAMPLE_BRANCH_ANY;
 741         }
 742 error:
 743         free(os);
 744         return ret;
 745 }
 746
 747 #ifdef LIBUNWIND_SUPPORT
 748 static int get_stack_size(char *str, unsigned long *_size)
 749 {
 750         char *endptr;
 751         unsigned long size;
 752         unsigned long max_size = round_down(USHRT_MAX, sizeof(u64));
 753
 754         size = strtoul(str, &endptr, 0);
 755
 756         do {
 757                 if (*endptr)
 758                         break;
 759
 760                 size = round_up(size, sizeof(u64));
 761                 if (!size || size > max_size)
 762                         break;
 763
 764                 *_size = size;
 765                 return 0;
 766
 767         } while (0);
 768
 769         pr_err("callchain: Incorrect stack dump size (max %ld): %s\n",
 770                max_size, str);
 771         return -1;
 772 }
 773 #endif /* LIBUNWIND_SUPPORT */
 774
 775 int record_parse_callchain_opt(const struct option *opt,
 776                                const char *arg, int unset)
 777 {
 778         struct perf_record_opts *opts = opt->value;
 779         char *tok, *name, *saveptr = NULL;
 780         char *buf;
 781         int ret = -1;
 782
 783         /* --no-call-graph */
 784         if (unset)
 785                 return 0;
 786
 787         /* We specified default option if none is provided. */
 788         BUG_ON(!arg);
 789
 790         /* We need buffer that we know we can write to. */
 791         buf = malloc(strlen(arg) + 1);
 792         if (!buf)
 793                 return -ENOMEM;
 794
 795         strcpy(buf, arg);
 796
 797         tok = strtok_r((char *)buf, ",", &saveptr);
 798         name = tok ? : (char *)buf;
 799
 800         do {
 801                 /* Framepointer style */
 802                 if (!strncmp(name, "fp", sizeof("fp"))) {
 803                         if (!strtok_r(NULL, ",", &saveptr)) {
 804                                 opts->call_graph = CALLCHAIN_FP;
 805                                 ret = 0;
 806                         } else
 807                                 pr_err("callchain: No more arguments "
 808                                        "needed for -g fp\n");
 809                         break;
 810
 811 #ifdef LIBUNWIND_SUPPORT
 812                 /* Dwarf style */
 813                 } else if (!strncmp(name, "dwarf", sizeof("dwarf"))) {
 814                         const unsigned long default_stack_dump_size = 8192;
 815
 816                         ret = 0;
 817                         opts->call_graph = CALLCHAIN_DWARF;
 818                         opts->stack_dump_size = default_stack_dump_size;
 819
 820                         tok = strtok_r(NULL, ",", &saveptr);
 821                         if (tok) {
 822                                 unsigned long size = 0;
 823
 824                                 ret = get_stack_size(tok, &size);
 825                                 opts->stack_dump_size = size;
 826                         }
 827
 828                         if (!ret)
 829                                 pr_debug("callchain: stack dump size %d\n",
 830                                          opts->stack_dump_size);
 831 #endif /* LIBUNWIND_SUPPORT */
 832                 } else {
 833                         pr_err("callchain: Unknown -g option "
 834                                "value: %s\n", arg);
 835                         break;
 836                 }
 837
 838         } while (0);
 839
 840         free(buf);
 841
 842         if (!ret)
 843                 pr_debug("callchain: type %d\n", opts->call_graph);
 844
 845         return ret;
 846 }
 847
 848 static const char * const record_usage[] = {
 849         "perf record [<options>] [<command>]",
 850         "perf record [<options>] -- <command> [<options>]",
 851         NULL
 852 };
 853
 854 /*
 855  * XXX Ideally would be local to cmd_record() and passed to a perf_record__new
 856  * because we need to have access to it in perf_record__exit, that is called
 857  * after cmd_record() exits, but since record_options need to be accessible to
 858  * builtin-script, leave it here.
 859  *
 860  * At least we don't ouch it in all the other functions here directly.
 861  *
 862  * Just say no to tons of global variables, sigh.
 863  */
 864 static struct perf_record record = {
 865         .opts = {
 866                 .mmap_pages          = UINT_MAX,
 867                 .user_freq           = UINT_MAX,
 868                 .user_interval       = ULLONG_MAX,
 869                 .freq                = 4000,
 870                 .target              = {
 871                         .uses_mmap   = true,
 872                 },
 873         },
 874         .write_mode = WRITE_FORCE,
 875         .file_new   = true,
 876 };
 877
 878 #define CALLCHAIN_HELP "do call-graph (stack chain/backtrace) recording: "
 879
 880 #ifdef LIBUNWIND_SUPPORT
 881 const char record_callchain_help[] = CALLCHAIN_HELP "[fp] dwarf";
 882 #else
 883 const char record_callchain_help[] = CALLCHAIN_HELP "[fp]";
 884 #endif
 885
 886 /*
 887  * XXX Will stay a global variable till we fix builtin-script.c to stop messing
 888  * with it and switch to use the library functions in perf_evlist that came
 889  * from builtin-record.c, i.e. use perf_record_opts,
 890  * perf_evlist__prepare_workload, etc instead of fork+exec'in 'perf record',
 891  * using pipes, etc.
 892  */
 893 const struct option record_options[] = {
 894         OPT_CALLBACK('e', "event", &record.evlist, "event",
 895                      "event selector. use 'perf list' to list available events",
 896                      parse_events_option),
 897         OPT_CALLBACK(0, "filter", &record.evlist, "filter",
 898                      "event filter", parse_filter),
 899         OPT_STRING('p', "pid", &record.opts.target.pid, "pid",
 900                     "record events on existing process id"),
 901         OPT_STRING('t', "tid", &record.opts.target.tid, "tid",
 902                     "record events on existing thread id"),
 903         OPT_INTEGER('r', "realtime", &record.realtime_prio,
 904                     "collect data with this RT SCHED_FIFO priority"),
 905         OPT_BOOLEAN('D', "no-delay", &record.opts.no_delay,
 906                     "collect data without buffering"),
 907         OPT_BOOLEAN('R', "raw-samples", &record.opts.raw_samples,
 908                     "collect raw sample records from all opened counters"),
 909         OPT_BOOLEAN('a', "all-cpus", &record.opts.target.system_wide,
 910                             "system-wide collection from all CPUs"),
 911         OPT_BOOLEAN('A', "append", &record.append_file,
 912                             "append to the output file to do incremental profiling"),
 913         OPT_STRING('C', "cpu", &record.opts.target.cpu_list, "cpu",
 914                     "list of cpus to monitor"),
 915         OPT_BOOLEAN('f', "force", &record.force,
 916                         "overwrite existing data file (deprecated)"),
 917         OPT_U64('c', "count", &record.opts.user_interval, "event period to sample"),
 918         OPT_STRING('o', "output", &record.output_name, "file",
 919                     "output file name"),
 920         OPT_BOOLEAN('i', "no-inherit", &record.opts.no_inherit,
 921                     "child tasks do not inherit counters"),
 922         OPT_UINTEGER('F', "freq", &record.opts.user_freq, "profile at this frequency"),
 923         OPT_UINTEGER('m', "mmap-pages", &record.opts.mmap_pages,
 924                      "number of mmap data pages"),
 925         OPT_BOOLEAN(0, "group", &record.opts.group,
 926                     "put the counters into a counter group"),
 927         OPT_CALLBACK_DEFAULT('g', "call-graph", &record.opts,
 928                              "mode[,dump_size]", record_callchain_help,
 929                              &record_parse_callchain_opt, "fp"),
 930         OPT_INCR('v', "verbose", &verbose,
 931                     "be more verbose (show counter open errors, etc)"),
 932         OPT_BOOLEAN('q', "quiet", &quiet, "don't print any message"),
 933         OPT_BOOLEAN('s', "stat", &record.opts.inherit_stat,
 934                     "per thread counts"),
 935         OPT_BOOLEAN('d', "data", &record.opts.sample_address,
 936                     "Sample addresses"),
 937         OPT_BOOLEAN('T', "timestamp", &record.opts.sample_time, "Sample timestamps"),
 938         OPT_BOOLEAN('P', "period", &record.opts.period, "Sample period"),
 939         OPT_BOOLEAN('n', "no-samples", &record.opts.no_samples,
 940                     "don't sample"),
 941         OPT_BOOLEAN('N', "no-buildid-cache", &record.no_buildid_cache,
 942                     "do not update the buildid cache"),
 943         OPT_BOOLEAN('B', "no-buildid", &record.no_buildid,
 944                     "do not collect buildids in perf.data"),
 945         OPT_CALLBACK('G', "cgroup", &record.evlist, "name",
 946                      "monitor event in cgroup name only",
 947                      parse_cgroups),
 948         OPT_STRING('u', "uid", &record.opts.target.uid_str, "user",
 949                    "user to profile"),
 950
 951         OPT_CALLBACK_NOOPT('b', "branch-any", &record.opts.branch_stack,
 952                      "branch any", "sample any taken branches",
 953                      parse_branch_stack),
 954
 955         OPT_CALLBACK('j', "branch-filter", &record.opts.branch_stack,
 956                      "branch filter mask", "branch stack filter modes",
 957                      parse_branch_stack),
 958         OPT_END()
 959 };
 960
 961 int cmd_record(int argc, const char **argv, const char *prefix __maybe_unused)
 962 {
 963         int err = -ENOMEM;
 964         struct perf_evsel *pos;
 965         struct perf_evlist *evsel_list;
 966         struct perf_record *rec = &record;
 967         char errbuf[BUFSIZ];
 968
 969         evsel_list = perf_evlist__new();
 970         if (evsel_list == NULL)
 971                 return -ENOMEM;
 972
 973         rec->evlist = evsel_list;
 974
 975         argc = parse_options(argc, argv, record_options, record_usage,
 976                             PARSE_OPT_STOP_AT_NON_OPTION);
 977         if (!argc && perf_target__none(&rec->opts.target))
 978                 usage_with_options(record_usage, record_options);
 979
 980         if (rec->force && rec->append_file) {
 981                 ui__error("Can't overwrite and append at the same time."
 982                           " You need to choose between -f and -A");
 983                 usage_with_options(record_usage, record_options);
 984         } else if (rec->append_file) {
 985                 rec->write_mode = WRITE_APPEND;
 986         } else {
 987                 rec->write_mode = WRITE_FORCE;
 988         }
 989
 990         if (nr_cgroups && !rec->opts.target.system_wide) {
 991                 ui__error("cgroup monitoring only available in"
 992                           " system-wide mode\n");
 993                 usage_with_options(record_usage, record_options);
 994         }
 995
 996         symbol__init();
 997
 998         if (symbol_conf.kptr_restrict)
 999                 pr_warning(
1000 "WARNING: Kernel address maps (/proc/{kallsyms,modules}) are restricted,\n"
1001 "check /proc/sys/kernel/kptr_restrict.\n\n"
1002 "Samples in kernel functions may not be resolved if a suitable vmlinux\n"
1003 "file is not found in the buildid cache or in the vmlinux path.\n\n"
1004 "Samples in kernel modules won't be resolved at all.\n\n"
1005 "If some relocation was applied (e.g. kexec) symbols may be misresolved\n"
1006 "even with a suitable vmlinux or kallsyms file.\n\n");
1007
1008         if (rec->no_buildid_cache || rec->no_buildid)
1009                 disable_buildid_cache();
1010
1011         if (evsel_list->nr_entries == 0 &&
1012             perf_evlist__add_default(evsel_list) < 0) {
1013                 pr_err("Not enough memory for event selector list\n");
1014                 goto out_symbol_exit;
1015         }
1016
1017         err = perf_target__validate(&rec->opts.target);
1018         if (err) {
1019                 perf_target__strerror(&rec->opts.target, err, errbuf, BUFSIZ);
1020                 ui__warning("%s", errbuf);
1021         }
1022
1023         err = perf_target__parse_uid(&rec->opts.target);
1024         if (err) {
1025                 int saved_errno = errno;
1026
1027                 perf_target__strerror(&rec->opts.target, err, errbuf, BUFSIZ);
1028                 ui__error("%s", errbuf);
1029
1030                 err = -saved_errno;
1031                 goto out_free_fd;
1032         }
1033
1034         err = -ENOMEM;
1035         if (perf_evlist__create_maps(evsel_list, &rec->opts.target) < 0)
1036                 usage_with_options(record_usage, record_options);
1037
1038         list_for_each_entry(pos, &evsel_list->entries, node) {
1039                 if (perf_header__push_event(pos->attr.config, perf_evsel__name(pos)))
1040                         goto out_free_fd;
1041         }
1042
1043         if (rec->opts.user_interval != ULLONG_MAX)
1044                 rec->opts.default_interval = rec->opts.user_interval;
1045         if (rec->opts.user_freq != UINT_MAX)
1046                 rec->opts.freq = rec->opts.user_freq;
1047
1048         /*
1049          * User specified count overrides default frequency.
1050          */
1051         if (rec->opts.default_interval)
1052                 rec->opts.freq = 0;
1053         else if (rec->opts.freq) {
1054                 rec->opts.default_interval = rec->opts.freq;
1055         } else {
1056                 ui__error("frequency and count are zero, aborting\n");
1057                 err = -EINVAL;
1058                 goto out_free_fd;
1059         }
1060
1061         err = __cmd_record(&record, argc, argv);
1062 out_free_fd:
1063         perf_evlist__delete_maps(evsel_list);
1064 out_symbol_exit:
1065         symbol__exit();
1066         return err;
1067 }