Pileus Git - ~andy/linux/blob - 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 #define SID(e, x, y) xyarray__entry(e->id, x, y)
  35
  36 enum write_mode_t {
  37         WRITE_FORCE,
  38         WRITE_APPEND
  39 };
  40
  41 static u64                      user_interval                   = ULLONG_MAX;
  42 static u64                      default_interval                =      0;
  43 static u64                      sample_type;
  44
  45 static unsigned int             page_size;
  46 static unsigned int             mmap_pages                      =    128;
  47 static unsigned int             user_freq                       = UINT_MAX;
  48 static int                      freq                            =   1000;
  49 static int                      output;
  50 static int                      pipe_output                     =      0;
  51 static const char               *output_name                    = NULL;
  52 static int                      group                           =      0;
  53 static int                      realtime_prio                   =      0;
  54 static bool                     nodelay                         =  false;
  55 static bool                     raw_samples                     =  false;
  56 static bool                     sample_id_all_avail             =   true;
  57 static bool                     system_wide                     =  false;
  58 static pid_t                    target_pid                      =     -1;
  59 static pid_t                    target_tid                      =     -1;
  60 static pid_t                    child_pid                       =     -1;
  61 static bool                     no_inherit                      =  false;
  62 static enum write_mode_t        write_mode                      = WRITE_FORCE;
  63 static bool                     call_graph                      =  false;
  64 static bool                     inherit_stat                    =  false;
  65 static bool                     no_samples                      =  false;
  66 static bool                     sample_address                  =  false;
  67 static bool                     sample_time                     =  false;
  68 static bool                     no_buildid                      =  false;
  69 static bool                     no_buildid_cache                =  false;
  70 static struct perf_evlist       *evsel_list;
  71
  72 static long                     samples                         =      0;
  73 static u64                      bytes_written                   =      0;
  74
  75 static int                      file_new                        =      1;
  76 static off_t                    post_processing_offset;
  77
  78 static struct perf_session      *session;
  79 static const char               *cpu_list;
  80
  81 static void advance_output(size_t size)
  82 {
  83         bytes_written += size;
  84 }
  85
  86 static void write_output(void *buf, size_t size)
  87 {
  88         while (size) {
  89                 int ret = write(output, buf, size);
  90
  91                 if (ret < 0)
  92                         die("failed to write");
  93
  94                 size -= ret;
  95                 buf += ret;
  96
  97                 bytes_written += ret;
  98         }
  99 }
 100
 101 static int process_synthesized_event(union perf_event *event,
 102                                      struct perf_sample *sample __used,
 103                                      struct perf_session *self __used)
 104 {
 105         write_output(event, event->header.size);
 106         return 0;
 107 }
 108
 109 static void mmap_read(struct perf_mmap *md)
 110 {
 111         unsigned int head = perf_mmap__read_head(md);
 112         unsigned int old = md->prev;
 113         unsigned char *data = md->base + page_size;
 114         unsigned long size;
 115         void *buf;
 116
 117         if (old == head)
 118                 return;
 119
 120         samples++;
 121
 122         size = head - old;
 123
 124         if ((old & md->mask) + size != (head & md->mask)) {
 125                 buf = &data[old & md->mask];
 126                 size = md->mask + 1 - (old & md->mask);
 127                 old += size;
 128
 129                 write_output(buf, size);
 130         }
 131
 132         buf = &data[old & md->mask];
 133         size = head - old;
 134         old += size;
 135
 136         write_output(buf, size);
 137
 138         md->prev = old;
 139         perf_mmap__write_tail(md, old);
 140 }
 141
 142 static volatile int done = 0;
 143 static volatile int signr = -1;
 144
 145 static void sig_handler(int sig)
 146 {
 147         done = 1;
 148         signr = sig;
 149 }
 150
 151 static void sig_atexit(void)
 152 {
 153         if (child_pid > 0)
 154                 kill(child_pid, SIGTERM);
 155
 156         if (signr == -1 || signr == SIGUSR1)
 157                 return;
 158
 159         signal(signr, SIG_DFL);
 160         kill(getpid(), signr);
 161 }
 162
 163 static struct perf_header_attr *get_header_attr(struct perf_event_attr *a, int nr)
 164 {
 165         struct perf_header_attr *h_attr;
 166
 167         if (nr < session->header.attrs) {
 168                 h_attr = session->header.attr[nr];
 169         } else {
 170                 h_attr = perf_header_attr__new(a);
 171                 if (h_attr != NULL)
 172                         if (perf_header__add_attr(&session->header, h_attr) < 0) {
 173                                 perf_header_attr__delete(h_attr);
 174                                 h_attr = NULL;
 175                         }
 176         }
 177
 178         return h_attr;
 179 }
 180
 181 static void create_counter(struct perf_evsel *evsel, int cpu)
 182 {
 183         struct perf_event_attr *attr = &evsel->attr;
 184         struct perf_header_attr *h_attr;
 185         struct perf_sample_id *sid;
 186         int thread_index;
 187
 188         for (thread_index = 0; thread_index < evsel_list->threads->nr; thread_index++) {
 189                 h_attr = get_header_attr(attr, evsel->idx);
 190                 if (h_attr == NULL)
 191                         die("nomem\n");
 192
 193                 if (!file_new) {
 194                         if (memcmp(&h_attr->attr, attr, sizeof(*attr))) {
 195                                 fprintf(stderr, "incompatible append\n");
 196                                 exit(-1);
 197                         }
 198                 }
 199
 200                 sid = SID(evsel, cpu, thread_index);
 201                 if (perf_header_attr__add_id(h_attr, sid->id) < 0) {
 202                         pr_warning("Not enough memory to add id\n");
 203                         exit(-1);
 204                 }
 205         }
 206
 207         if (!sample_type)
 208                 sample_type = attr->sample_type;
 209 }
 210
 211 static void config_attr(struct perf_evsel *evsel, struct perf_evlist *evlist)
 212 {
 213         struct perf_event_attr *attr = &evsel->attr;
 214         int track = !evsel->idx; /* only the first counter needs these */
 215
 216         attr->read_format       = PERF_FORMAT_TOTAL_TIME_ENABLED |
 217                                   PERF_FORMAT_TOTAL_TIME_RUNNING |
 218                                   PERF_FORMAT_ID;
 219
 220         attr->sample_type       |= PERF_SAMPLE_IP | PERF_SAMPLE_TID;
 221
 222         if (evlist->nr_entries > 1)
 223                 attr->sample_type |= PERF_SAMPLE_ID;
 224
 225         /*
 226          * We default some events to a 1 default interval. But keep
 227          * it a weak assumption overridable by the user.
 228          */
 229         if (!attr->sample_period || (user_freq != UINT_MAX &&
 230                                      user_interval != ULLONG_MAX)) {
 231                 if (freq) {
 232                         attr->sample_type       |= PERF_SAMPLE_PERIOD;
 233                         attr->freq              = 1;
 234                         attr->sample_freq       = freq;
 235                 } else {
 236                         attr->sample_period = default_interval;
 237                 }
 238         }
 239
 240         if (no_samples)
 241                 attr->sample_freq = 0;
 242
 243         if (inherit_stat)
 244                 attr->inherit_stat = 1;
 245
 246         if (sample_address) {
 247                 attr->sample_type       |= PERF_SAMPLE_ADDR;
 248                 attr->mmap_data = track;
 249         }
 250
 251         if (call_graph)
 252                 attr->sample_type       |= PERF_SAMPLE_CALLCHAIN;
 253
 254         if (system_wide)
 255                 attr->sample_type       |= PERF_SAMPLE_CPU;
 256
 257         if (sample_id_all_avail &&
 258             (sample_time || system_wide || !no_inherit || cpu_list))
 259                 attr->sample_type       |= PERF_SAMPLE_TIME;
 260
 261         if (raw_samples) {
 262                 attr->sample_type       |= PERF_SAMPLE_TIME;
 263                 attr->sample_type       |= PERF_SAMPLE_RAW;
 264                 attr->sample_type       |= PERF_SAMPLE_CPU;
 265         }
 266
 267         if (nodelay) {
 268                 attr->watermark = 0;
 269                 attr->wakeup_events = 1;
 270         }
 271
 272         attr->mmap              = track;
 273         attr->comm              = track;
 274
 275         if (target_pid == -1 && target_tid == -1 && !system_wide) {
 276                 attr->disabled = 1;
 277                 attr->enable_on_exec = 1;
 278         }
 279 }
 280
 281 static void open_counters(struct perf_evlist *evlist)
 282 {
 283         struct perf_evsel *pos;
 284         int cpu;
 285
 286         list_for_each_entry(pos, &evlist->entries, node) {
 287                 struct perf_event_attr *attr = &pos->attr;
 288                 /*
 289                  * Check if parse_single_tracepoint_event has already asked for
 290                  * PERF_SAMPLE_TIME.
 291                  *
 292                  * XXX this is kludgy but short term fix for problems introduced by
 293                  * eac23d1c that broke 'perf script' by having different sample_types
 294                  * when using multiple tracepoint events when we use a perf binary
 295                  * that tries to use sample_id_all on an older kernel.
 296                  *
 297                  * We need to move counter creation to perf_session, support
 298                  * different sample_types, etc.
 299                  */
 300                 bool time_needed = attr->sample_type & PERF_SAMPLE_TIME;
 301
 302                 config_attr(pos, evlist);
 303 retry_sample_id:
 304                 attr->sample_id_all = sample_id_all_avail ? 1 : 0;
 305 try_again:
 306                 if (perf_evsel__open(pos, evlist->cpus, evlist->threads, group,
 307                                      !no_inherit) < 0) {
 308                         int err = errno;
 309
 310                         if (err == EPERM || err == EACCES)
 311                                 die("Permission error - are you root?\n"
 312                                         "\t Consider tweaking"
 313                                         " /proc/sys/kernel/perf_event_paranoid.\n");
 314                         else if (err ==  ENODEV && cpu_list) {
 315                                 die("No such device - did you specify"
 316                                         " an out-of-range profile CPU?\n");
 317                         } else if (err == EINVAL && sample_id_all_avail) {
 318                                 /*
 319                                  * Old kernel, no attr->sample_id_type_all field
 320                                  */
 321                                 sample_id_all_avail = false;
 322                                 if (!sample_time && !raw_samples && !time_needed)
 323                                         attr->sample_type &= ~PERF_SAMPLE_TIME;
 324
 325                                 goto retry_sample_id;
 326                         }
 327
 328                         /*
 329                          * If it's cycles then fall back to hrtimer
 330                          * based cpu-clock-tick sw counter, which
 331                          * is always available even if no PMU support:
 332                          */
 333                         if (attr->type == PERF_TYPE_HARDWARE
 334                                         && attr->config == PERF_COUNT_HW_CPU_CYCLES) {
 335
 336                                 if (verbose)
 337                                         warning(" ... trying to fall back to cpu-clock-ticks\n");
 338                                 attr->type = PERF_TYPE_SOFTWARE;
 339                                 attr->config = PERF_COUNT_SW_CPU_CLOCK;
 340                                 goto try_again;
 341                         }
 342                         printf("\n");
 343                         error("sys_perf_event_open() syscall returned with %d (%s).  /bin/dmesg may provide additional information.\n",
 344                               err, strerror(err));
 345
 346 #if defined(__i386__) || defined(__x86_64__)
 347                         if (attr->type == PERF_TYPE_HARDWARE && err == EOPNOTSUPP)
 348                                 die("No hardware sampling interrupt available."
 349                                     " No APIC? If so then you can boot the kernel"
 350                                     " with the \"lapic\" boot parameter to"
 351                                     " force-enable it.\n");
 352 #endif
 353
 354                         die("No CONFIG_PERF_EVENTS=y kernel support configured?\n");
 355                 }
 356         }
 357
 358         if (perf_evlist__set_filters(evlist)) {
 359                 error("failed to set filter with %d (%s)\n", errno,
 360                         strerror(errno));
 361                 exit(-1);
 362         }
 363
 364         if (perf_evlist__mmap(evlist, mmap_pages, false) < 0)
 365                 die("failed to mmap with %d (%s)\n", errno, strerror(errno));
 366
 367         for (cpu = 0; cpu < evsel_list->cpus->nr; ++cpu) {
 368                 list_for_each_entry(pos, &evlist->entries, node)
 369                         create_counter(pos, cpu);
 370         }
 371 }
 372
 373 static int process_buildids(void)
 374 {
 375         u64 size = lseek(output, 0, SEEK_CUR);
 376
 377         if (size == 0)
 378                 return 0;
 379
 380         session->fd = output;
 381         return __perf_session__process_events(session, post_processing_offset,
 382                                               size - post_processing_offset,
 383                                               size, &build_id__mark_dso_hit_ops);
 384 }
 385
 386 static void atexit_header(void)
 387 {
 388         if (!pipe_output) {
 389                 session->header.data_size += bytes_written;
 390
 391                 if (!no_buildid)
 392                         process_buildids();
 393                 perf_header__write(&session->header, evsel_list, output, true);
 394                 perf_session__delete(session);
 395                 perf_evlist__delete(evsel_list);
 396                 symbol__exit();
 397         }
 398 }
 399
 400 static void perf_event__synthesize_guest_os(struct machine *machine, void *data)
 401 {
 402         int err;
 403         struct perf_session *psession = data;
 404
 405         if (machine__is_host(machine))
 406                 return;
 407
 408         /*
 409          *As for guest kernel when processing subcommand record&report,
 410          *we arrange module mmap prior to guest kernel mmap and trigger
 411          *a preload dso because default guest module symbols are loaded
 412          *from guest kallsyms instead of /lib/modules/XXX/XXX. This
 413          *method is used to avoid symbol missing when the first addr is
 414          *in module instead of in guest kernel.
 415          */
 416         err = perf_event__synthesize_modules(process_synthesized_event,
 417                                              psession, machine);
 418         if (err < 0)
 419                 pr_err("Couldn't record guest kernel [%d]'s reference"
 420                        " relocation symbol.\n", machine->pid);
 421
 422         /*
 423          * We use _stext for guest kernel because guest kernel's /proc/kallsyms
 424          * have no _text sometimes.
 425          */
 426         err = perf_event__synthesize_kernel_mmap(process_synthesized_event,
 427                                                  psession, machine, "_text");
 428         if (err < 0)
 429                 err = perf_event__synthesize_kernel_mmap(process_synthesized_event,
 430                                                          psession, machine,
 431                                                          "_stext");
 432         if (err < 0)
 433                 pr_err("Couldn't record guest kernel [%d]'s reference"
 434                        " relocation symbol.\n", machine->pid);
 435 }
 436
 437 static struct perf_event_header finished_round_event = {
 438         .size = sizeof(struct perf_event_header),
 439         .type = PERF_RECORD_FINISHED_ROUND,
 440 };
 441
 442 static void mmap_read_all(void)
 443 {
 444         int i;
 445
 446         for (i = 0; i < evsel_list->cpus->nr; i++) {
 447                 if (evsel_list->mmap[i].base)
 448                         mmap_read(&evsel_list->mmap[i]);
 449         }
 450
 451         if (perf_header__has_feat(&session->header, HEADER_TRACE_INFO))
 452                 write_output(&finished_round_event, sizeof(finished_round_event));
 453 }
 454
 455 static int __cmd_record(int argc, const char **argv)
 456 {
 457         int i;
 458         struct stat st;
 459         int flags;
 460         int err;
 461         unsigned long waking = 0;
 462         int child_ready_pipe[2], go_pipe[2];
 463         const bool forks = argc > 0;
 464         char buf;
 465         struct machine *machine;
 466
 467         page_size = sysconf(_SC_PAGE_SIZE);
 468
 469         atexit(sig_atexit);
 470         signal(SIGCHLD, sig_handler);
 471         signal(SIGINT, sig_handler);
 472         signal(SIGUSR1, sig_handler);
 473
 474         if (forks && (pipe(child_ready_pipe) < 0 || pipe(go_pipe) < 0)) {
 475                 perror("failed to create pipes");
 476                 exit(-1);
 477         }
 478
 479         if (!output_name) {
 480                 if (!fstat(STDOUT_FILENO, &st) && S_ISFIFO(st.st_mode))
 481                         pipe_output = 1;
 482                 else
 483                         output_name = "perf.data";
 484         }
 485         if (output_name) {
 486                 if (!strcmp(output_name, "-"))
 487                         pipe_output = 1;
 488                 else if (!stat(output_name, &st) && st.st_size) {
 489                         if (write_mode == WRITE_FORCE) {
 490                                 char oldname[PATH_MAX];
 491                                 snprintf(oldname, sizeof(oldname), "%s.old",
 492                                          output_name);
 493                                 unlink(oldname);
 494                                 rename(output_name, oldname);
 495                         }
 496                 } else if (write_mode == WRITE_APPEND) {
 497                         write_mode = WRITE_FORCE;
 498                 }
 499         }
 500
 501         flags = O_CREAT|O_RDWR;
 502         if (write_mode == WRITE_APPEND)
 503                 file_new = 0;
 504         else
 505                 flags |= O_TRUNC;
 506
 507         if (pipe_output)
 508                 output = STDOUT_FILENO;
 509         else
 510                 output = open(output_name, flags, S_IRUSR | S_IWUSR);
 511         if (output < 0) {
 512                 perror("failed to create output file");
 513                 exit(-1);
 514         }
 515
 516         session = perf_session__new(output_name, O_WRONLY,
 517                                     write_mode == WRITE_FORCE, false, NULL);
 518         if (session == NULL) {
 519                 pr_err("Not enough memory for reading perf file header\n");
 520                 return -1;
 521         }
 522
 523         if (!no_buildid)
 524                 perf_header__set_feat(&session->header, HEADER_BUILD_ID);
 525
 526         if (!file_new) {
 527                 err = perf_header__read(session, output);
 528                 if (err < 0)
 529                         goto out_delete_session;
 530         }
 531
 532         if (have_tracepoints(&evsel_list->entries))
 533                 perf_header__set_feat(&session->header, HEADER_TRACE_INFO);
 534
 535         if (forks) {
 536                 child_pid = fork();
 537                 if (child_pid < 0) {
 538                         perror("failed to fork");
 539                         exit(-1);
 540                 }
 541
 542                 if (!child_pid) {
 543                         if (pipe_output)
 544                                 dup2(2, 1);
 545                         close(child_ready_pipe[0]);
 546                         close(go_pipe[1]);
 547                         fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
 548
 549                         /*
 550                          * Do a dummy execvp to get the PLT entry resolved,
 551                          * so we avoid the resolver overhead on the real
 552                          * execvp call.
 553                          */
 554                         execvp("", (char **)argv);
 555
 556                         /*
 557                          * Tell the parent we're ready to go
 558                          */
 559                         close(child_ready_pipe[1]);
 560
 561                         /*
 562                          * Wait until the parent tells us to go.
 563                          */
 564                         if (read(go_pipe[0], &buf, 1) == -1)
 565                                 perror("unable to read pipe");
 566
 567                         execvp(argv[0], (char **)argv);
 568
 569                         perror(argv[0]);
 570                         kill(getppid(), SIGUSR1);
 571                         exit(-1);
 572                 }
 573
 574                 if (!system_wide && target_tid == -1 && target_pid == -1)
 575                         evsel_list->threads->map[0] = child_pid;
 576
 577                 close(child_ready_pipe[1]);
 578                 close(go_pipe[0]);
 579                 /*
 580                  * wait for child to settle
 581                  */
 582                 if (read(child_ready_pipe[0], &buf, 1) == -1) {
 583                         perror("unable to read pipe");
 584                         exit(-1);
 585                 }
 586                 close(child_ready_pipe[0]);
 587         }
 588
 589         open_counters(evsel_list);
 590
 591         perf_session__set_sample_type(session, sample_type);
 592
 593         /*
 594          * perf_session__delete(session) will be called at atexit_header()
 595          */
 596         atexit(atexit_header);
 597
 598         if (pipe_output) {
 599                 err = perf_header__write_pipe(output);
 600                 if (err < 0)
 601                         return err;
 602         } else if (file_new) {
 603                 err = perf_header__write(&session->header, evsel_list,
 604                                          output, false);
 605                 if (err < 0)
 606                         return err;
 607         }
 608
 609         post_processing_offset = lseek(output, 0, SEEK_CUR);
 610
 611         perf_session__set_sample_id_all(session, sample_id_all_avail);
 612
 613         if (pipe_output) {
 614                 err = perf_event__synthesize_attrs(&session->header,
 615                                                    process_synthesized_event,
 616                                                    session);
 617                 if (err < 0) {
 618                         pr_err("Couldn't synthesize attrs.\n");
 619                         return err;
 620                 }
 621
 622                 err = perf_event__synthesize_event_types(process_synthesized_event,
 623                                                          session);
 624                 if (err < 0) {
 625                         pr_err("Couldn't synthesize event_types.\n");
 626                         return err;
 627                 }
 628
 629                 if (have_tracepoints(&evsel_list->entries)) {
 630                         /*
 631                          * FIXME err <= 0 here actually means that
 632                          * there were no tracepoints so its not really
 633                          * an error, just that we don't need to
 634                          * synthesize anything.  We really have to
 635                          * return this more properly and also
 636                          * propagate errors that now are calling die()
 637                          */
 638                         err = perf_event__synthesize_tracing_data(output, evsel_list,
 639                                                                   process_synthesized_event,
 640                                                                   session);
 641                         if (err <= 0) {
 642                                 pr_err("Couldn't record tracing data.\n");
 643                                 return err;
 644                         }
 645                         advance_output(err);
 646                 }
 647         }
 648
 649         machine = perf_session__find_host_machine(session);
 650         if (!machine) {
 651                 pr_err("Couldn't find native kernel information.\n");
 652                 return -1;
 653         }
 654
 655         err = perf_event__synthesize_kernel_mmap(process_synthesized_event,
 656                                                  session, machine, "_text");
 657         if (err < 0)
 658                 err = perf_event__synthesize_kernel_mmap(process_synthesized_event,
 659                                                          session, machine, "_stext");
 660         if (err < 0)
 661                 pr_err("Couldn't record kernel reference relocation symbol\n"
 662                        "Symbol resolution may be skewed if relocation was used (e.g. kexec).\n"
 663                        "Check /proc/kallsyms permission or run as root.\n");
 664
 665         err = perf_event__synthesize_modules(process_synthesized_event,
 666                                              session, machine);
 667         if (err < 0)
 668                 pr_err("Couldn't record kernel module information.\n"
 669                        "Symbol resolution may be skewed if relocation was used (e.g. kexec).\n"
 670                        "Check /proc/modules permission or run as root.\n");
 671
 672         if (perf_guest)
 673                 perf_session__process_machines(session,
 674                                                perf_event__synthesize_guest_os);
 675
 676         if (!system_wide)
 677                 perf_event__synthesize_thread_map(evsel_list->threads,
 678                                                   process_synthesized_event,
 679                                                   session);
 680         else
 681                 perf_event__synthesize_threads(process_synthesized_event,
 682                                                session);
 683
 684         if (realtime_prio) {
 685                 struct sched_param param;
 686
 687                 param.sched_priority = realtime_prio;
 688                 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
 689                         pr_err("Could not set realtime priority.\n");
 690                         exit(-1);
 691                 }
 692         }
 693
 694         /*
 695          * Let the child rip
 696          */
 697         if (forks)
 698                 close(go_pipe[1]);
 699
 700         for (;;) {
 701                 int hits = samples;
 702                 int thread;
 703
 704                 mmap_read_all();
 705
 706                 if (hits == samples) {
 707                         if (done)
 708                                 break;
 709                         err = poll(evsel_list->pollfd, evsel_list->nr_fds, -1);
 710                         waking++;
 711                 }
 712
 713                 if (done) {
 714                         for (i = 0; i < evsel_list->cpus->nr; i++) {
 715                                 struct perf_evsel *pos;
 716
 717                                 list_for_each_entry(pos, &evsel_list->entries, node) {
 718                                         for (thread = 0;
 719                                                 thread < evsel_list->threads->nr;
 720                                                 thread++)
 721                                                 ioctl(FD(pos, i, thread),
 722                                                         PERF_EVENT_IOC_DISABLE);
 723                                 }
 724                         }
 725                 }
 726         }
 727
 728         if (quiet || signr == SIGUSR1)
 729                 return 0;
 730
 731         fprintf(stderr, "[ perf record: Woken up %ld times to write data ]\n", waking);
 732
 733         /*
 734          * Approximate RIP event size: 24 bytes.
 735          */
 736         fprintf(stderr,
 737                 "[ perf record: Captured and wrote %.3f MB %s (~%" PRIu64 " samples) ]\n",
 738                 (double)bytes_written / 1024.0 / 1024.0,
 739                 output_name,
 740                 bytes_written / 24);
 741
 742         return 0;
 743
 744 out_delete_session:
 745         perf_session__delete(session);
 746         return err;
 747 }
 748
 749 static const char * const record_usage[] = {
 750         "perf record [<options>] [<command>]",
 751         "perf record [<options>] -- <command> [<options>]",
 752         NULL
 753 };
 754
 755 static bool force, append_file;
 756
 757 const struct option record_options[] = {
 758         OPT_CALLBACK('e', "event", &evsel_list, "event",
 759                      "event selector. use 'perf list' to list available events",
 760                      parse_events),
 761         OPT_CALLBACK(0, "filter", &evsel_list, "filter",
 762                      "event filter", parse_filter),
 763         OPT_INTEGER('p', "pid", &target_pid,
 764                     "record events on existing process id"),
 765         OPT_INTEGER('t', "tid", &target_tid,
 766                     "record events on existing thread id"),
 767         OPT_INTEGER('r', "realtime", &realtime_prio,
 768                     "collect data with this RT SCHED_FIFO priority"),
 769         OPT_BOOLEAN('D', "no-delay", &nodelay,
 770                     "collect data without buffering"),
 771         OPT_BOOLEAN('R', "raw-samples", &raw_samples,
 772                     "collect raw sample records from all opened counters"),
 773         OPT_BOOLEAN('a', "all-cpus", &system_wide,
 774                             "system-wide collection from all CPUs"),
 775         OPT_BOOLEAN('A', "append", &append_file,
 776                             "append to the output file to do incremental profiling"),
 777         OPT_STRING('C', "cpu", &cpu_list, "cpu",
 778                     "list of cpus to monitor"),
 779         OPT_BOOLEAN('f', "force", &force,
 780                         "overwrite existing data file (deprecated)"),
 781         OPT_U64('c', "count", &user_interval, "event period to sample"),
 782         OPT_STRING('o', "output", &output_name, "file",
 783                     "output file name"),
 784         OPT_BOOLEAN('i', "no-inherit", &no_inherit,
 785                     "child tasks do not inherit counters"),
 786         OPT_UINTEGER('F', "freq", &user_freq, "profile at this frequency"),
 787         OPT_UINTEGER('m', "mmap-pages", &mmap_pages, "number of mmap data pages"),
 788         OPT_BOOLEAN('g', "call-graph", &call_graph,
 789                     "do call-graph (stack chain/backtrace) recording"),
 790         OPT_INCR('v', "verbose", &verbose,
 791                     "be more verbose (show counter open errors, etc)"),
 792         OPT_BOOLEAN('q', "quiet", &quiet, "don't print any message"),
 793         OPT_BOOLEAN('s', "stat", &inherit_stat,
 794                     "per thread counts"),
 795         OPT_BOOLEAN('d', "data", &sample_address,
 796                     "Sample addresses"),
 797         OPT_BOOLEAN('T', "timestamp", &sample_time, "Sample timestamps"),
 798         OPT_BOOLEAN('n', "no-samples", &no_samples,
 799                     "don't sample"),
 800         OPT_BOOLEAN('N', "no-buildid-cache", &no_buildid_cache,
 801                     "do not update the buildid cache"),
 802         OPT_BOOLEAN('B', "no-buildid", &no_buildid,
 803                     "do not collect buildids in perf.data"),
 804         OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
 805                      "monitor event in cgroup name only",
 806                      parse_cgroups),
 807         OPT_END()
 808 };
 809
 810 int cmd_record(int argc, const char **argv, const char *prefix __used)
 811 {
 812         int err = -ENOMEM;
 813         struct perf_evsel *pos;
 814
 815         evsel_list = perf_evlist__new(NULL, NULL);
 816         if (evsel_list == NULL)
 817                 return -ENOMEM;
 818
 819         argc = parse_options(argc, argv, record_options, record_usage,
 820                             PARSE_OPT_STOP_AT_NON_OPTION);
 821         if (!argc && target_pid == -1 && target_tid == -1 &&
 822                 !system_wide && !cpu_list)
 823                 usage_with_options(record_usage, record_options);
 824
 825         if (force && append_file) {
 826                 fprintf(stderr, "Can't overwrite and append at the same time."
 827                                 " You need to choose between -f and -A");
 828                 usage_with_options(record_usage, record_options);
 829         } else if (append_file) {
 830                 write_mode = WRITE_APPEND;
 831         } else {
 832                 write_mode = WRITE_FORCE;
 833         }
 834
 835         if (nr_cgroups && !system_wide) {
 836                 fprintf(stderr, "cgroup monitoring only available in"
 837                         " system-wide mode\n");
 838                 usage_with_options(record_usage, record_options);
 839         }
 840
 841         symbol__init();
 842
 843         if (no_buildid_cache || no_buildid)
 844                 disable_buildid_cache();
 845
 846         if (evsel_list->nr_entries == 0 &&
 847             perf_evlist__add_default(evsel_list) < 0) {
 848                 pr_err("Not enough memory for event selector list\n");
 849                 goto out_symbol_exit;
 850         }
 851
 852         if (target_pid != -1)
 853                 target_tid = target_pid;
 854
 855         if (perf_evlist__create_maps(evsel_list, target_pid,
 856                                      target_tid, cpu_list) < 0)
 857                 usage_with_options(record_usage, record_options);
 858
 859         list_for_each_entry(pos, &evsel_list->entries, node) {
 860                 if (perf_evsel__alloc_fd(pos, evsel_list->cpus->nr,
 861                                          evsel_list->threads->nr) < 0)
 862                         goto out_free_fd;
 863                 if (perf_header__push_event(pos->attr.config, event_name(pos)))
 864                         goto out_free_fd;
 865         }
 866
 867         if (perf_evlist__alloc_pollfd(evsel_list) < 0)
 868                 goto out_free_fd;
 869
 870         if (user_interval != ULLONG_MAX)
 871                 default_interval = user_interval;
 872         if (user_freq != UINT_MAX)
 873                 freq = user_freq;
 874
 875         /*
 876          * User specified count overrides default frequency.
 877          */
 878         if (default_interval)
 879                 freq = 0;
 880         else if (freq) {
 881                 default_interval = freq;
 882         } else {
 883                 fprintf(stderr, "frequency and count are zero, aborting\n");
 884                 err = -EINVAL;
 885                 goto out_free_fd;
 886         }
 887
 888         err = __cmd_record(argc, argv);
 889 out_free_fd:
 890         perf_evlist__delete_maps(evsel_list);
 891 out_symbol_exit:
 892         symbol__exit();
 893         return err;
 894 }