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/symbol.h"
  26 #include "util/cpumap.h"
  27 #include "util/thread_map.h"
  28
  29 #include <unistd.h>
  30 #include <sched.h>
  31 #include <sys/mman.h>
  32
  33 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
  34
  35 enum write_mode_t {
  36         WRITE_FORCE,
  37         WRITE_APPEND
  38 };
  39
  40 static u64                      user_interval                   = ULLONG_MAX;
  41 static u64                      default_interval                =      0;
  42
  43 static unsigned int             page_size;
  44 static unsigned int             mmap_pages                      = UINT_MAX;
  45 static unsigned int             user_freq                       = UINT_MAX;
  46 static int                      freq                            =   1000;
  47 static int                      output;
  48 static int                      pipe_output                     =      0;
  49 static const char               *output_name                    = NULL;
  50 static int                      group                           =      0;
  51 static int                      realtime_prio                   =      0;
  52 static bool                     nodelay                         =  false;
  53 static bool                     raw_samples                     =  false;
  54 static bool                     sample_id_all_avail             =   true;
  55 static bool                     system_wide                     =  false;
  56 static pid_t                    target_pid                      =     -1;
  57 static pid_t                    target_tid                      =     -1;
  58 static pid_t                    child_pid                       =     -1;
  59 static bool                     no_inherit                      =  false;
  60 static enum write_mode_t        write_mode                      = WRITE_FORCE;
  61 static bool                     call_graph                      =  false;
  62 static bool                     inherit_stat                    =  false;
  63 static bool                     no_samples                      =  false;
  64 static bool                     sample_address                  =  false;
  65 static bool                     sample_time                     =  false;
  66 static bool                     no_buildid                      =  false;
  67 static bool                     no_buildid_cache                =  false;
  68 static struct perf_evlist       *evsel_list;
  69
  70 static long                     samples                         =      0;
  71 static u64                      bytes_written                   =      0;
  72
  73 static int                      file_new                        =      1;
  74 static off_t                    post_processing_offset;
  75
  76 static struct perf_session      *session;
  77 static const char               *cpu_list;
  78
  79 static void advance_output(size_t size)
  80 {
  81         bytes_written += size;
  82 }
  83
  84 static void write_output(void *buf, size_t size)
  85 {
  86         while (size) {
  87                 int ret = write(output, buf, size);
  88
  89                 if (ret < 0)
  90                         die("failed to write");
  91
  92                 size -= ret;
  93                 buf += ret;
  94
  95                 bytes_written += ret;
  96         }
  97 }
  98
  99 static int process_synthesized_event(union perf_event *event,
 100                                      struct perf_sample *sample __used,
 101                                      struct perf_session *self __used)
 102 {
 103         write_output(event, event->header.size);
 104         return 0;
 105 }
 106
 107 static void mmap_read(struct perf_mmap *md)
 108 {
 109         unsigned int head = perf_mmap__read_head(md);
 110         unsigned int old = md->prev;
 111         unsigned char *data = md->base + page_size;
 112         unsigned long size;
 113         void *buf;
 114
 115         if (old == head)
 116                 return;
 117
 118         samples++;
 119
 120         size = head - old;
 121
 122         if ((old & md->mask) + size != (head & md->mask)) {
 123                 buf = &data[old & md->mask];
 124                 size = md->mask + 1 - (old & md->mask);
 125                 old += size;
 126
 127                 write_output(buf, size);
 128         }
 129
 130         buf = &data[old & md->mask];
 131         size = head - old;
 132         old += size;
 133
 134         write_output(buf, size);
 135
 136         md->prev = old;
 137         perf_mmap__write_tail(md, old);
 138 }
 139
 140 static volatile int done = 0;
 141 static volatile int signr = -1;
 142
 143 static void sig_handler(int sig)
 144 {
 145         done = 1;
 146         signr = sig;
 147 }
 148
 149 static void sig_atexit(void)
 150 {
 151         if (child_pid > 0)
 152                 kill(child_pid, SIGTERM);
 153
 154         if (signr == -1 || signr == SIGUSR1)
 155                 return;
 156
 157         signal(signr, SIG_DFL);
 158         kill(getpid(), signr);
 159 }
 160
 161 static void config_attr(struct perf_evsel *evsel, struct perf_evlist *evlist)
 162 {
 163         struct perf_event_attr *attr = &evsel->attr;
 164         int track = !evsel->idx; /* only the first counter needs these */
 165
 166         attr->read_format       = PERF_FORMAT_TOTAL_TIME_ENABLED |
 167                                   PERF_FORMAT_TOTAL_TIME_RUNNING |
 168                                   PERF_FORMAT_ID;
 169
 170         attr->sample_type       |= PERF_SAMPLE_IP | PERF_SAMPLE_TID;
 171
 172         if (evlist->nr_entries > 1)
 173                 attr->sample_type |= PERF_SAMPLE_ID;
 174
 175         /*
 176          * We default some events to a 1 default interval. But keep
 177          * it a weak assumption overridable by the user.
 178          */
 179         if (!attr->sample_period || (user_freq != UINT_MAX &&
 180                                      user_interval != ULLONG_MAX)) {
 181                 if (freq) {
 182                         attr->sample_type       |= PERF_SAMPLE_PERIOD;
 183                         attr->freq              = 1;
 184                         attr->sample_freq       = freq;
 185                 } else {
 186                         attr->sample_period = default_interval;
 187                 }
 188         }
 189
 190         if (no_samples)
 191                 attr->sample_freq = 0;
 192
 193         if (inherit_stat)
 194                 attr->inherit_stat = 1;
 195
 196         if (sample_address) {
 197                 attr->sample_type       |= PERF_SAMPLE_ADDR;
 198                 attr->mmap_data = track;
 199         }
 200
 201         if (call_graph)
 202                 attr->sample_type       |= PERF_SAMPLE_CALLCHAIN;
 203
 204         if (system_wide)
 205                 attr->sample_type       |= PERF_SAMPLE_CPU;
 206
 207         if (sample_id_all_avail &&
 208             (sample_time || system_wide || !no_inherit || cpu_list))
 209                 attr->sample_type       |= PERF_SAMPLE_TIME;
 210
 211         if (raw_samples) {
 212                 attr->sample_type       |= PERF_SAMPLE_TIME;
 213                 attr->sample_type       |= PERF_SAMPLE_RAW;
 214                 attr->sample_type       |= PERF_SAMPLE_CPU;
 215         }
 216
 217         if (nodelay) {
 218                 attr->watermark = 0;
 219                 attr->wakeup_events = 1;
 220         }
 221
 222         attr->mmap              = track;
 223         attr->comm              = track;
 224
 225         if (target_pid == -1 && target_tid == -1 && !system_wide) {
 226                 attr->disabled = 1;
 227                 attr->enable_on_exec = 1;
 228         }
 229 }
 230
 231 static bool perf_evlist__equal(struct perf_evlist *evlist,
 232                                struct perf_evlist *other)
 233 {
 234         struct perf_evsel *pos, *pair;
 235
 236         if (evlist->nr_entries != other->nr_entries)
 237                 return false;
 238
 239         pair = list_entry(other->entries.next, struct perf_evsel, node);
 240
 241         list_for_each_entry(pos, &evlist->entries, node) {
 242                 if (memcmp(&pos->attr, &pair->attr, sizeof(pos->attr) != 0))
 243                         return false;
 244                 pair = list_entry(pair->node.next, struct perf_evsel, node);
 245         }
 246
 247         return true;
 248 }
 249
 250 static void open_counters(struct perf_evlist *evlist)
 251 {
 252         struct perf_evsel *pos;
 253
 254         list_for_each_entry(pos, &evlist->entries, node) {
 255                 struct perf_event_attr *attr = &pos->attr;
 256                 /*
 257                  * Check if parse_single_tracepoint_event has already asked for
 258                  * PERF_SAMPLE_TIME.
 259                  *
 260                  * XXX this is kludgy but short term fix for problems introduced by
 261                  * eac23d1c that broke 'perf script' by having different sample_types
 262                  * when using multiple tracepoint events when we use a perf binary
 263                  * that tries to use sample_id_all on an older kernel.
 264                  *
 265                  * We need to move counter creation to perf_session, support
 266                  * different sample_types, etc.
 267                  */
 268                 bool time_needed = attr->sample_type & PERF_SAMPLE_TIME;
 269
 270                 config_attr(pos, evlist);
 271 retry_sample_id:
 272                 attr->sample_id_all = sample_id_all_avail ? 1 : 0;
 273 try_again:
 274                 if (perf_evsel__open(pos, evlist->cpus, evlist->threads, group,
 275                                      !no_inherit) < 0) {
 276                         int err = errno;
 277
 278                         if (err == EPERM || err == EACCES) {
 279                                 ui__warning_paranoid();
 280                                 exit(EXIT_FAILURE);
 281                         } else if (err ==  ENODEV && cpu_list) {
 282                                 die("No such device - did you specify"
 283                                         " an out-of-range profile CPU?\n");
 284                         } else if (err == EINVAL && sample_id_all_avail) {
 285                                 /*
 286                                  * Old kernel, no attr->sample_id_type_all field
 287                                  */
 288                                 sample_id_all_avail = false;
 289                                 if (!sample_time && !raw_samples && !time_needed)
 290                                         attr->sample_type &= ~PERF_SAMPLE_TIME;
 291
 292                                 goto retry_sample_id;
 293                         }
 294
 295                         /*
 296                          * If it's cycles then fall back to hrtimer
 297                          * based cpu-clock-tick sw counter, which
 298                          * is always available even if no PMU support:
 299                          */
 300                         if (attr->type == PERF_TYPE_HARDWARE
 301                                         && attr->config == PERF_COUNT_HW_CPU_CYCLES) {
 302
 303                                 if (verbose)
 304                                         ui__warning("The cycles event is not supported, "
 305                                                     "trying to fall back to cpu-clock-ticks\n");
 306                                 attr->type = PERF_TYPE_SOFTWARE;
 307                                 attr->config = PERF_COUNT_SW_CPU_CLOCK;
 308                                 goto try_again;
 309                         }
 310
 311                         if (err == ENOENT) {
 312                                 ui__warning("The %s event is not supported.\n",
 313                                             event_name(pos));
 314                                 exit(EXIT_FAILURE);
 315                         }
 316
 317                         printf("\n");
 318                         error("sys_perf_event_open() syscall returned with %d (%s).  /bin/dmesg may provide additional information.\n",
 319                               err, strerror(err));
 320
 321 #if defined(__i386__) || defined(__x86_64__)
 322                         if (attr->type == PERF_TYPE_HARDWARE && err == EOPNOTSUPP)
 323                                 die("No hardware sampling interrupt available."
 324                                     " No APIC? If so then you can boot the kernel"
 325                                     " with the \"lapic\" boot parameter to"
 326                                     " force-enable it.\n");
 327 #endif
 328
 329                         die("No CONFIG_PERF_EVENTS=y kernel support configured?\n");
 330                 }
 331         }
 332
 333         if (perf_evlist__set_filters(evlist)) {
 334                 error("failed to set filter with %d (%s)\n", errno,
 335                         strerror(errno));
 336                 exit(-1);
 337         }
 338
 339         if (perf_evlist__mmap(evlist, mmap_pages, false) < 0)
 340                 die("failed to mmap with %d (%s)\n", errno, strerror(errno));
 341
 342         if (file_new)
 343                 session->evlist = evlist;
 344         else {
 345                 if (!perf_evlist__equal(session->evlist, evlist)) {
 346                         fprintf(stderr, "incompatible append\n");
 347                         exit(-1);
 348                 }
 349         }
 350
 351         perf_session__update_sample_type(session);
 352 }
 353
 354 static int process_buildids(void)
 355 {
 356         u64 size = lseek(output, 0, SEEK_CUR);
 357
 358         if (size == 0)
 359                 return 0;
 360
 361         session->fd = output;
 362         return __perf_session__process_events(session, post_processing_offset,
 363                                               size - post_processing_offset,
 364                                               size, &build_id__mark_dso_hit_ops);
 365 }
 366
 367 static void atexit_header(void)
 368 {
 369         if (!pipe_output) {
 370                 session->header.data_size += bytes_written;
 371
 372                 if (!no_buildid)
 373                         process_buildids();
 374                 perf_session__write_header(session, evsel_list, output, true);
 375                 perf_session__delete(session);
 376                 perf_evlist__delete(evsel_list);
 377                 symbol__exit();
 378         }
 379 }
 380
 381 static void perf_event__synthesize_guest_os(struct machine *machine, void *data)
 382 {
 383         int err;
 384         struct perf_session *psession = data;
 385
 386         if (machine__is_host(machine))
 387                 return;
 388
 389         /*
 390          *As for guest kernel when processing subcommand record&report,
 391          *we arrange module mmap prior to guest kernel mmap and trigger
 392          *a preload dso because default guest module symbols are loaded
 393          *from guest kallsyms instead of /lib/modules/XXX/XXX. This
 394          *method is used to avoid symbol missing when the first addr is
 395          *in module instead of in guest kernel.
 396          */
 397         err = perf_event__synthesize_modules(process_synthesized_event,
 398                                              psession, machine);
 399         if (err < 0)
 400                 pr_err("Couldn't record guest kernel [%d]'s reference"
 401                        " relocation symbol.\n", machine->pid);
 402
 403         /*
 404          * We use _stext for guest kernel because guest kernel's /proc/kallsyms
 405          * have no _text sometimes.
 406          */
 407         err = perf_event__synthesize_kernel_mmap(process_synthesized_event,
 408                                                  psession, machine, "_text");
 409         if (err < 0)
 410                 err = perf_event__synthesize_kernel_mmap(process_synthesized_event,
 411                                                          psession, machine,
 412                                                          "_stext");
 413         if (err < 0)
 414                 pr_err("Couldn't record guest kernel [%d]'s reference"
 415                        " relocation symbol.\n", machine->pid);
 416 }
 417
 418 static struct perf_event_header finished_round_event = {
 419         .size = sizeof(struct perf_event_header),
 420         .type = PERF_RECORD_FINISHED_ROUND,
 421 };
 422
 423 static void mmap_read_all(void)
 424 {
 425         int i;
 426
 427         for (i = 0; i < evsel_list->cpus->nr; i++) {
 428                 if (evsel_list->mmap[i].base)
 429                         mmap_read(&evsel_list->mmap[i]);
 430         }
 431
 432         if (perf_header__has_feat(&session->header, HEADER_TRACE_INFO))
 433                 write_output(&finished_round_event, sizeof(finished_round_event));
 434 }
 435
 436 static int __cmd_record(int argc, const char **argv)
 437 {
 438         int i;
 439         struct stat st;
 440         int flags;
 441         int err;
 442         unsigned long waking = 0;
 443         int child_ready_pipe[2], go_pipe[2];
 444         const bool forks = argc > 0;
 445         char buf;
 446         struct machine *machine;
 447
 448         page_size = sysconf(_SC_PAGE_SIZE);
 449
 450         atexit(sig_atexit);
 451         signal(SIGCHLD, sig_handler);
 452         signal(SIGINT, sig_handler);
 453         signal(SIGUSR1, sig_handler);
 454
 455         if (forks && (pipe(child_ready_pipe) < 0 || pipe(go_pipe) < 0)) {
 456                 perror("failed to create pipes");
 457                 exit(-1);
 458         }
 459
 460         if (!output_name) {
 461                 if (!fstat(STDOUT_FILENO, &st) && S_ISFIFO(st.st_mode))
 462                         pipe_output = 1;
 463                 else
 464                         output_name = "perf.data";
 465         }
 466         if (output_name) {
 467                 if (!strcmp(output_name, "-"))
 468                         pipe_output = 1;
 469                 else if (!stat(output_name, &st) && st.st_size) {
 470                         if (write_mode == WRITE_FORCE) {
 471                                 char oldname[PATH_MAX];
 472                                 snprintf(oldname, sizeof(oldname), "%s.old",
 473                                          output_name);
 474                                 unlink(oldname);
 475                                 rename(output_name, oldname);
 476                         }
 477                 } else if (write_mode == WRITE_APPEND) {
 478                         write_mode = WRITE_FORCE;
 479                 }
 480         }
 481
 482         flags = O_CREAT|O_RDWR;
 483         if (write_mode == WRITE_APPEND)
 484                 file_new = 0;
 485         else
 486                 flags |= O_TRUNC;
 487
 488         if (pipe_output)
 489                 output = STDOUT_FILENO;
 490         else
 491                 output = open(output_name, flags, S_IRUSR | S_IWUSR);
 492         if (output < 0) {
 493                 perror("failed to create output file");
 494                 exit(-1);
 495         }
 496
 497         session = perf_session__new(output_name, O_WRONLY,
 498                                     write_mode == WRITE_FORCE, false, NULL);
 499         if (session == NULL) {
 500                 pr_err("Not enough memory for reading perf file header\n");
 501                 return -1;
 502         }
 503
 504         if (!no_buildid)
 505                 perf_header__set_feat(&session->header, HEADER_BUILD_ID);
 506
 507         if (!file_new) {
 508                 err = perf_session__read_header(session, output);
 509                 if (err < 0)
 510                         goto out_delete_session;
 511         }
 512
 513         if (have_tracepoints(&evsel_list->entries))
 514                 perf_header__set_feat(&session->header, HEADER_TRACE_INFO);
 515
 516         /* 512 kiB: default amount of unprivileged mlocked memory */
 517         if (mmap_pages == UINT_MAX)
 518                 mmap_pages = (512 * 1024) / page_size;
 519
 520         if (forks) {
 521                 child_pid = fork();
 522                 if (child_pid < 0) {
 523                         perror("failed to fork");
 524                         exit(-1);
 525                 }
 526
 527                 if (!child_pid) {
 528                         if (pipe_output)
 529                                 dup2(2, 1);
 530                         close(child_ready_pipe[0]);
 531                         close(go_pipe[1]);
 532                         fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
 533
 534                         /*
 535                          * Do a dummy execvp to get the PLT entry resolved,
 536                          * so we avoid the resolver overhead on the real
 537                          * execvp call.
 538                          */
 539                         execvp("", (char **)argv);
 540
 541                         /*
 542                          * Tell the parent we're ready to go
 543                          */
 544                         close(child_ready_pipe[1]);
 545
 546                         /*
 547                          * Wait until the parent tells us to go.
 548                          */
 549                         if (read(go_pipe[0], &buf, 1) == -1)
 550                                 perror("unable to read pipe");
 551
 552                         execvp(argv[0], (char **)argv);
 553
 554                         perror(argv[0]);
 555                         kill(getppid(), SIGUSR1);
 556                         exit(-1);
 557                 }
 558
 559                 if (!system_wide && target_tid == -1 && target_pid == -1)
 560                         evsel_list->threads->map[0] = child_pid;
 561
 562                 close(child_ready_pipe[1]);
 563                 close(go_pipe[0]);
 564                 /*
 565                  * wait for child to settle
 566                  */
 567                 if (read(child_ready_pipe[0], &buf, 1) == -1) {
 568                         perror("unable to read pipe");
 569                         exit(-1);
 570                 }
 571                 close(child_ready_pipe[0]);
 572         }
 573
 574         open_counters(evsel_list);
 575
 576         /*
 577          * perf_session__delete(session) will be called at atexit_header()
 578          */
 579         atexit(atexit_header);
 580
 581         if (pipe_output) {
 582                 err = perf_header__write_pipe(output);
 583                 if (err < 0)
 584                         return err;
 585         } else if (file_new) {
 586                 err = perf_session__write_header(session, evsel_list,
 587                                                  output, false);
 588                 if (err < 0)
 589                         return err;
 590         }
 591
 592         post_processing_offset = lseek(output, 0, SEEK_CUR);
 593
 594         if (pipe_output) {
 595                 err = perf_session__synthesize_attrs(session,
 596                                                      process_synthesized_event);
 597                 if (err < 0) {
 598                         pr_err("Couldn't synthesize attrs.\n");
 599                         return err;
 600                 }
 601
 602                 err = perf_event__synthesize_event_types(process_synthesized_event,
 603                                                          session);
 604                 if (err < 0) {
 605                         pr_err("Couldn't synthesize event_types.\n");
 606                         return err;
 607                 }
 608
 609                 if (have_tracepoints(&evsel_list->entries)) {
 610                         /*
 611                          * FIXME err <= 0 here actually means that
 612                          * there were no tracepoints so its not really
 613                          * an error, just that we don't need to
 614                          * synthesize anything.  We really have to
 615                          * return this more properly and also
 616                          * propagate errors that now are calling die()
 617                          */
 618                         err = perf_event__synthesize_tracing_data(output, evsel_list,
 619                                                                   process_synthesized_event,
 620                                                                   session);
 621                         if (err <= 0) {
 622                                 pr_err("Couldn't record tracing data.\n");
 623                                 return err;
 624                         }
 625                         advance_output(err);
 626                 }
 627         }
 628
 629         machine = perf_session__find_host_machine(session);
 630         if (!machine) {
 631                 pr_err("Couldn't find native kernel information.\n");
 632                 return -1;
 633         }
 634
 635         err = perf_event__synthesize_kernel_mmap(process_synthesized_event,
 636                                                  session, machine, "_text");
 637         if (err < 0)
 638                 err = perf_event__synthesize_kernel_mmap(process_synthesized_event,
 639                                                          session, machine, "_stext");
 640         if (err < 0)
 641                 pr_err("Couldn't record kernel reference relocation symbol\n"
 642                        "Symbol resolution may be skewed if relocation was used (e.g. kexec).\n"
 643                        "Check /proc/kallsyms permission or run as root.\n");
 644
 645         err = perf_event__synthesize_modules(process_synthesized_event,
 646                                              session, machine);
 647         if (err < 0)
 648                 pr_err("Couldn't record kernel module information.\n"
 649                        "Symbol resolution may be skewed if relocation was used (e.g. kexec).\n"
 650                        "Check /proc/modules permission or run as root.\n");
 651
 652         if (perf_guest)
 653                 perf_session__process_machines(session,
 654                                                perf_event__synthesize_guest_os);
 655
 656         if (!system_wide)
 657                 perf_event__synthesize_thread_map(evsel_list->threads,
 658                                                   process_synthesized_event,
 659                                                   session);
 660         else
 661                 perf_event__synthesize_threads(process_synthesized_event,
 662                                                session);
 663
 664         if (realtime_prio) {
 665                 struct sched_param param;
 666
 667                 param.sched_priority = realtime_prio;
 668                 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
 669                         pr_err("Could not set realtime priority.\n");
 670                         exit(-1);
 671                 }
 672         }
 673
 674         /*
 675          * Let the child rip
 676          */
 677         if (forks)
 678                 close(go_pipe[1]);
 679
 680         for (;;) {
 681                 int hits = samples;
 682                 int thread;
 683
 684                 mmap_read_all();
 685
 686                 if (hits == samples) {
 687                         if (done)
 688                                 break;
 689                         err = poll(evsel_list->pollfd, evsel_list->nr_fds, -1);
 690                         waking++;
 691                 }
 692
 693                 if (done) {
 694                         for (i = 0; i < evsel_list->cpus->nr; i++) {
 695                                 struct perf_evsel *pos;
 696
 697                                 list_for_each_entry(pos, &evsel_list->entries, node) {
 698                                         for (thread = 0;
 699                                                 thread < evsel_list->threads->nr;
 700                                                 thread++)
 701                                                 ioctl(FD(pos, i, thread),
 702                                                         PERF_EVENT_IOC_DISABLE);
 703                                 }
 704                         }
 705                 }
 706         }
 707
 708         if (quiet || signr == SIGUSR1)
 709                 return 0;
 710
 711         fprintf(stderr, "[ perf record: Woken up %ld times to write data ]\n", waking);
 712
 713         /*
 714          * Approximate RIP event size: 24 bytes.
 715          */
 716         fprintf(stderr,
 717                 "[ perf record: Captured and wrote %.3f MB %s (~%" PRIu64 " samples) ]\n",
 718                 (double)bytes_written / 1024.0 / 1024.0,
 719                 output_name,
 720                 bytes_written / 24);
 721
 722         return 0;
 723
 724 out_delete_session:
 725         perf_session__delete(session);
 726         return err;
 727 }
 728
 729 static const char * const record_usage[] = {
 730         "perf record [<options>] [<command>]",
 731         "perf record [<options>] -- <command> [<options>]",
 732         NULL
 733 };
 734
 735 static bool force, append_file;
 736
 737 const struct option record_options[] = {
 738         OPT_CALLBACK('e', "event", &evsel_list, "event",
 739                      "event selector. use 'perf list' to list available events",
 740                      parse_events),
 741         OPT_CALLBACK(0, "filter", &evsel_list, "filter",
 742                      "event filter", parse_filter),
 743         OPT_INTEGER('p', "pid", &target_pid,
 744                     "record events on existing process id"),
 745         OPT_INTEGER('t', "tid", &target_tid,
 746                     "record events on existing thread id"),
 747         OPT_INTEGER('r', "realtime", &realtime_prio,
 748                     "collect data with this RT SCHED_FIFO priority"),
 749         OPT_BOOLEAN('D', "no-delay", &nodelay,
 750                     "collect data without buffering"),
 751         OPT_BOOLEAN('R', "raw-samples", &raw_samples,
 752                     "collect raw sample records from all opened counters"),
 753         OPT_BOOLEAN('a', "all-cpus", &system_wide,
 754                             "system-wide collection from all CPUs"),
 755         OPT_BOOLEAN('A', "append", &append_file,
 756                             "append to the output file to do incremental profiling"),
 757         OPT_STRING('C', "cpu", &cpu_list, "cpu",
 758                     "list of cpus to monitor"),
 759         OPT_BOOLEAN('f', "force", &force,
 760                         "overwrite existing data file (deprecated)"),
 761         OPT_U64('c', "count", &user_interval, "event period to sample"),
 762         OPT_STRING('o', "output", &output_name, "file",
 763                     "output file name"),
 764         OPT_BOOLEAN('i', "no-inherit", &no_inherit,
 765                     "child tasks do not inherit counters"),
 766         OPT_UINTEGER('F', "freq", &user_freq, "profile at this frequency"),
 767         OPT_UINTEGER('m', "mmap-pages", &mmap_pages, "number of mmap data pages"),
 768         OPT_BOOLEAN('g', "call-graph", &call_graph,
 769                     "do call-graph (stack chain/backtrace) recording"),
 770         OPT_INCR('v', "verbose", &verbose,
 771                     "be more verbose (show counter open errors, etc)"),
 772         OPT_BOOLEAN('q', "quiet", &quiet, "don't print any message"),
 773         OPT_BOOLEAN('s', "stat", &inherit_stat,
 774                     "per thread counts"),
 775         OPT_BOOLEAN('d', "data", &sample_address,
 776                     "Sample addresses"),
 777         OPT_BOOLEAN('T', "timestamp", &sample_time, "Sample timestamps"),
 778         OPT_BOOLEAN('n', "no-samples", &no_samples,
 779                     "don't sample"),
 780         OPT_BOOLEAN('N', "no-buildid-cache", &no_buildid_cache,
 781                     "do not update the buildid cache"),
 782         OPT_BOOLEAN('B', "no-buildid", &no_buildid,
 783                     "do not collect buildids in perf.data"),
 784         OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
 785                      "monitor event in cgroup name only",
 786                      parse_cgroups),
 787         OPT_END()
 788 };
 789
 790 int cmd_record(int argc, const char **argv, const char *prefix __used)
 791 {
 792         int err = -ENOMEM;
 793         struct perf_evsel *pos;
 794
 795         evsel_list = perf_evlist__new(NULL, NULL);
 796         if (evsel_list == NULL)
 797                 return -ENOMEM;
 798
 799         argc = parse_options(argc, argv, record_options, record_usage,
 800                             PARSE_OPT_STOP_AT_NON_OPTION);
 801         if (!argc && target_pid == -1 && target_tid == -1 &&
 802                 !system_wide && !cpu_list)
 803                 usage_with_options(record_usage, record_options);
 804
 805         if (force && append_file) {
 806                 fprintf(stderr, "Can't overwrite and append at the same time."
 807                                 " You need to choose between -f and -A");
 808                 usage_with_options(record_usage, record_options);
 809         } else if (append_file) {
 810                 write_mode = WRITE_APPEND;
 811         } else {
 812                 write_mode = WRITE_FORCE;
 813         }
 814
 815         if (nr_cgroups && !system_wide) {
 816                 fprintf(stderr, "cgroup monitoring only available in"
 817                         " system-wide mode\n");
 818                 usage_with_options(record_usage, record_options);
 819         }
 820
 821         symbol__init();
 822
 823         if (no_buildid_cache || no_buildid)
 824                 disable_buildid_cache();
 825
 826         if (evsel_list->nr_entries == 0 &&
 827             perf_evlist__add_default(evsel_list) < 0) {
 828                 pr_err("Not enough memory for event selector list\n");
 829                 goto out_symbol_exit;
 830         }
 831
 832         if (target_pid != -1)
 833                 target_tid = target_pid;
 834
 835         if (perf_evlist__create_maps(evsel_list, target_pid,
 836                                      target_tid, cpu_list) < 0)
 837                 usage_with_options(record_usage, record_options);
 838
 839         list_for_each_entry(pos, &evsel_list->entries, node) {
 840                 if (perf_evsel__alloc_fd(pos, evsel_list->cpus->nr,
 841                                          evsel_list->threads->nr) < 0)
 842                         goto out_free_fd;
 843                 if (perf_header__push_event(pos->attr.config, event_name(pos)))
 844                         goto out_free_fd;
 845         }
 846
 847         if (perf_evlist__alloc_pollfd(evsel_list) < 0)
 848                 goto out_free_fd;
 849
 850         if (user_interval != ULLONG_MAX)
 851                 default_interval = user_interval;
 852         if (user_freq != UINT_MAX)
 853                 freq = user_freq;
 854
 855         /*
 856          * User specified count overrides default frequency.
 857          */
 858         if (default_interval)
 859                 freq = 0;
 860         else if (freq) {
 861                 default_interval = freq;
 862         } else {
 863                 fprintf(stderr, "frequency and count are zero, aborting\n");
 864                 err = -EINVAL;
 865                 goto out_free_fd;
 866         }
 867
 868         err = __cmd_record(argc, argv);
 869 out_free_fd:
 870         perf_evlist__delete_maps(evsel_list);
 871 out_symbol_exit:
 872         symbol__exit();
 873         return err;
 874 }