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/debug.h"
  22 #include "util/session.h"
  23 #include "util/symbol.h"
  24 #include "util/cpumap.h"
  25
  26 #include <unistd.h>
  27 #include <sched.h>
  28 #include <sys/mman.h>
  29
  30 enum write_mode_t {
  31         WRITE_FORCE,
  32         WRITE_APPEND
  33 };
  34
  35 static int                      *fd[MAX_NR_CPUS][MAX_COUNTERS];
  36
  37 static u64                      user_interval                   = ULLONG_MAX;
  38 static u64                      default_interval                =      0;
  39
  40 static int                      nr_cpus                         =      0;
  41 static unsigned int             page_size;
  42 static unsigned int             mmap_pages                      =    128;
  43 static unsigned int             user_freq                       = UINT_MAX;
  44 static int                      freq                            =   1000;
  45 static int                      output;
  46 static int                      pipe_output                     =      0;
  47 static const char               *output_name                    = "perf.data";
  48 static int                      group                           =      0;
  49 static int                      realtime_prio                   =      0;
  50 static bool                     raw_samples                     =  false;
  51 static bool                     system_wide                     =  false;
  52 static pid_t                    target_pid                      =     -1;
  53 static pid_t                    target_tid                      =     -1;
  54 static pid_t                    *all_tids                       =      NULL;
  55 static int                      thread_num                      =      0;
  56 static pid_t                    child_pid                       =     -1;
  57 static bool                     no_inherit                      =  false;
  58 static enum write_mode_t        write_mode                      = WRITE_FORCE;
  59 static bool                     call_graph                      =  false;
  60 static bool                     inherit_stat                    =  false;
  61 static bool                     no_samples                      =  false;
  62 static bool                     sample_address                  =  false;
  63 static bool                     no_buildid                      =  false;
  64
  65 static long                     samples                         =      0;
  66 static u64                      bytes_written                   =      0;
  67
  68 static struct pollfd            *event_array;
  69
  70 static int                      nr_poll                         =      0;
  71 static int                      nr_cpu                          =      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 struct mmap_data {
  80         int                     counter;
  81         void                    *base;
  82         unsigned int            mask;
  83         unsigned int            prev;
  84 };
  85
  86 static struct mmap_data         mmap_array[MAX_NR_CPUS];
  87
  88 static unsigned long mmap_read_head(struct mmap_data *md)
  89 {
  90         struct perf_event_mmap_page *pc = md->base;
  91         long head;
  92
  93         head = pc->data_head;
  94         rmb();
  95
  96         return head;
  97 }
  98
  99 static void mmap_write_tail(struct mmap_data *md, unsigned long tail)
 100 {
 101         struct perf_event_mmap_page *pc = md->base;
 102
 103         /*
 104          * ensure all reads are done before we write the tail out.
 105          */
 106         /* mb(); */
 107         pc->data_tail = tail;
 108 }
 109
 110 static void advance_output(size_t size)
 111 {
 112         bytes_written += size;
 113 }
 114
 115 static void write_output(void *buf, size_t size)
 116 {
 117         while (size) {
 118                 int ret = write(output, buf, size);
 119
 120                 if (ret < 0)
 121                         die("failed to write");
 122
 123                 size -= ret;
 124                 buf += ret;
 125
 126                 bytes_written += ret;
 127         }
 128 }
 129
 130 static int process_synthesized_event(event_t *event,
 131                                      struct perf_session *self __used)
 132 {
 133         write_output(event, event->header.size);
 134         return 0;
 135 }
 136
 137 static void mmap_read(struct mmap_data *md)
 138 {
 139         unsigned int head = mmap_read_head(md);
 140         unsigned int old = md->prev;
 141         unsigned char *data = md->base + page_size;
 142         unsigned long size;
 143         void *buf;
 144         int diff;
 145
 146         /*
 147          * If we're further behind than half the buffer, there's a chance
 148          * the writer will bite our tail and mess up the samples under us.
 149          *
 150          * If we somehow ended up ahead of the head, we got messed up.
 151          *
 152          * In either case, truncate and restart at head.
 153          */
 154         diff = head - old;
 155         if (diff < 0) {
 156                 fprintf(stderr, "WARNING: failed to keep up with mmap data\n");
 157                 /*
 158                  * head points to a known good entry, start there.
 159                  */
 160                 old = head;
 161         }
 162
 163         if (old != head)
 164                 samples++;
 165
 166         size = head - old;
 167
 168         if ((old & md->mask) + size != (head & md->mask)) {
 169                 buf = &data[old & md->mask];
 170                 size = md->mask + 1 - (old & md->mask);
 171                 old += size;
 172
 173                 write_output(buf, size);
 174         }
 175
 176         buf = &data[old & md->mask];
 177         size = head - old;
 178         old += size;
 179
 180         write_output(buf, size);
 181
 182         md->prev = old;
 183         mmap_write_tail(md, old);
 184 }
 185
 186 static volatile int done = 0;
 187 static volatile int signr = -1;
 188
 189 static void sig_handler(int sig)
 190 {
 191         done = 1;
 192         signr = sig;
 193 }
 194
 195 static void sig_atexit(void)
 196 {
 197         if (child_pid > 0)
 198                 kill(child_pid, SIGTERM);
 199
 200         if (signr == -1)
 201                 return;
 202
 203         signal(signr, SIG_DFL);
 204         kill(getpid(), signr);
 205 }
 206
 207 static int group_fd;
 208
 209 static struct perf_header_attr *get_header_attr(struct perf_event_attr *a, int nr)
 210 {
 211         struct perf_header_attr *h_attr;
 212
 213         if (nr < session->header.attrs) {
 214                 h_attr = session->header.attr[nr];
 215         } else {
 216                 h_attr = perf_header_attr__new(a);
 217                 if (h_attr != NULL)
 218                         if (perf_header__add_attr(&session->header, h_attr) < 0) {
 219                                 perf_header_attr__delete(h_attr);
 220                                 h_attr = NULL;
 221                         }
 222         }
 223
 224         return h_attr;
 225 }
 226
 227 static void create_counter(int counter, int cpu)
 228 {
 229         char *filter = filters[counter];
 230         struct perf_event_attr *attr = attrs + counter;
 231         struct perf_header_attr *h_attr;
 232         int track = !counter; /* only the first counter needs these */
 233         int thread_index;
 234         int ret;
 235         struct {
 236                 u64 count;
 237                 u64 time_enabled;
 238                 u64 time_running;
 239                 u64 id;
 240         } read_data;
 241
 242         attr->read_format       = PERF_FORMAT_TOTAL_TIME_ENABLED |
 243                                   PERF_FORMAT_TOTAL_TIME_RUNNING |
 244                                   PERF_FORMAT_ID;
 245
 246         attr->sample_type       |= PERF_SAMPLE_IP | PERF_SAMPLE_TID;
 247
 248         if (nr_counters > 1)
 249                 attr->sample_type |= PERF_SAMPLE_ID;
 250
 251         /*
 252          * We default some events to a 1 default interval. But keep
 253          * it a weak assumption overridable by the user.
 254          */
 255         if (!attr->sample_period || (user_freq != UINT_MAX &&
 256                                      user_interval != ULLONG_MAX)) {
 257                 if (freq) {
 258                         attr->sample_type       |= PERF_SAMPLE_PERIOD;
 259                         attr->freq              = 1;
 260                         attr->sample_freq       = freq;
 261                 } else {
 262                         attr->sample_period = default_interval;
 263                 }
 264         }
 265
 266         if (no_samples)
 267                 attr->sample_freq = 0;
 268
 269         if (inherit_stat)
 270                 attr->inherit_stat = 1;
 271
 272         if (sample_address) {
 273                 attr->sample_type       |= PERF_SAMPLE_ADDR;
 274                 attr->mmap_data = track;
 275         }
 276
 277         if (call_graph)
 278                 attr->sample_type       |= PERF_SAMPLE_CALLCHAIN;
 279
 280         if (system_wide)
 281                 attr->sample_type       |= PERF_SAMPLE_CPU;
 282
 283         if (raw_samples) {
 284                 attr->sample_type       |= PERF_SAMPLE_TIME;
 285                 attr->sample_type       |= PERF_SAMPLE_RAW;
 286                 attr->sample_type       |= PERF_SAMPLE_CPU;
 287         }
 288
 289         attr->mmap              = track;
 290         attr->comm              = track;
 291         attr->inherit           = !no_inherit;
 292         if (target_pid == -1 && target_tid == -1 && !system_wide) {
 293                 attr->disabled = 1;
 294                 attr->enable_on_exec = 1;
 295         }
 296
 297         for (thread_index = 0; thread_index < thread_num; thread_index++) {
 298 try_again:
 299                 fd[nr_cpu][counter][thread_index] = sys_perf_event_open(attr,
 300                                 all_tids[thread_index], cpu, group_fd, 0);
 301
 302                 if (fd[nr_cpu][counter][thread_index] < 0) {
 303                         int err = errno;
 304
 305                         if (err == EPERM || err == EACCES)
 306                                 die("Permission error - are you root?\n"
 307                                         "\t Consider tweaking"
 308                                         " /proc/sys/kernel/perf_event_paranoid.\n");
 309                         else if (err ==  ENODEV && cpu_list) {
 310                                 die("No such device - did you specify"
 311                                         " an out-of-range profile CPU?\n");
 312                         }
 313
 314                         /*
 315                          * If it's cycles then fall back to hrtimer
 316                          * based cpu-clock-tick sw counter, which
 317                          * is always available even if no PMU support:
 318                          */
 319                         if (attr->type == PERF_TYPE_HARDWARE
 320                                         && attr->config == PERF_COUNT_HW_CPU_CYCLES) {
 321
 322                                 if (verbose)
 323                                         warning(" ... trying to fall back to cpu-clock-ticks\n");
 324                                 attr->type = PERF_TYPE_SOFTWARE;
 325                                 attr->config = PERF_COUNT_SW_CPU_CLOCK;
 326                                 goto try_again;
 327                         }
 328                         printf("\n");
 329                         error("perfcounter syscall returned with %d (%s)\n",
 330                                         fd[nr_cpu][counter][thread_index], strerror(err));
 331
 332 #if defined(__i386__) || defined(__x86_64__)
 333                         if (attr->type == PERF_TYPE_HARDWARE && err == EOPNOTSUPP)
 334                                 die("No hardware sampling interrupt available."
 335                                     " No APIC? If so then you can boot the kernel"
 336                                     " with the \"lapic\" boot parameter to"
 337                                     " force-enable it.\n");
 338 #endif
 339
 340                         die("No CONFIG_PERF_EVENTS=y kernel support configured?\n");
 341                         exit(-1);
 342                 }
 343
 344                 h_attr = get_header_attr(attr, counter);
 345                 if (h_attr == NULL)
 346                         die("nomem\n");
 347
 348                 if (!file_new) {
 349                         if (memcmp(&h_attr->attr, attr, sizeof(*attr))) {
 350                                 fprintf(stderr, "incompatible append\n");
 351                                 exit(-1);
 352                         }
 353                 }
 354
 355                 if (read(fd[nr_cpu][counter][thread_index], &read_data, sizeof(read_data)) == -1) {
 356                         perror("Unable to read perf file descriptor\n");
 357                         exit(-1);
 358                 }
 359
 360                 if (perf_header_attr__add_id(h_attr, read_data.id) < 0) {
 361                         pr_warning("Not enough memory to add id\n");
 362                         exit(-1);
 363                 }
 364
 365                 assert(fd[nr_cpu][counter][thread_index] >= 0);
 366                 fcntl(fd[nr_cpu][counter][thread_index], F_SETFL, O_NONBLOCK);
 367
 368                 /*
 369                  * First counter acts as the group leader:
 370                  */
 371                 if (group && group_fd == -1)
 372                         group_fd = fd[nr_cpu][counter][thread_index];
 373
 374                 if (counter || thread_index) {
 375                         ret = ioctl(fd[nr_cpu][counter][thread_index],
 376                                         PERF_EVENT_IOC_SET_OUTPUT,
 377                                         fd[nr_cpu][0][0]);
 378                         if (ret) {
 379                                 error("failed to set output: %d (%s)\n", errno,
 380                                                 strerror(errno));
 381                                 exit(-1);
 382                         }
 383                 } else {
 384                         mmap_array[nr_cpu].counter = counter;
 385                         mmap_array[nr_cpu].prev = 0;
 386                         mmap_array[nr_cpu].mask = mmap_pages*page_size - 1;
 387                         mmap_array[nr_cpu].base = mmap(NULL, (mmap_pages+1)*page_size,
 388                                 PROT_READ|PROT_WRITE, MAP_SHARED, fd[nr_cpu][counter][thread_index], 0);
 389                         if (mmap_array[nr_cpu].base == MAP_FAILED) {
 390                                 error("failed to mmap with %d (%s)\n", errno, strerror(errno));
 391                                 exit(-1);
 392                         }
 393
 394                         event_array[nr_poll].fd = fd[nr_cpu][counter][thread_index];
 395                         event_array[nr_poll].events = POLLIN;
 396                         nr_poll++;
 397                 }
 398
 399                 if (filter != NULL) {
 400                         ret = ioctl(fd[nr_cpu][counter][thread_index],
 401                                         PERF_EVENT_IOC_SET_FILTER, filter);
 402                         if (ret) {
 403                                 error("failed to set filter with %d (%s)\n", errno,
 404                                                 strerror(errno));
 405                                 exit(-1);
 406                         }
 407                 }
 408         }
 409 }
 410
 411 static void open_counters(int cpu)
 412 {
 413         int counter;
 414
 415         group_fd = -1;
 416         for (counter = 0; counter < nr_counters; counter++)
 417                 create_counter(counter, cpu);
 418
 419         nr_cpu++;
 420 }
 421
 422 static int process_buildids(void)
 423 {
 424         u64 size = lseek(output, 0, SEEK_CUR);
 425
 426         if (size == 0)
 427                 return 0;
 428
 429         session->fd = output;
 430         return __perf_session__process_events(session, post_processing_offset,
 431                                               size - post_processing_offset,
 432                                               size, &build_id__mark_dso_hit_ops);
 433 }
 434
 435 static void atexit_header(void)
 436 {
 437         if (!pipe_output) {
 438                 session->header.data_size += bytes_written;
 439
 440                 process_buildids();
 441                 perf_header__write(&session->header, output, true);
 442         }
 443 }
 444
 445 static void event__synthesize_guest_os(struct machine *machine, void *data)
 446 {
 447         int err;
 448         struct perf_session *psession = data;
 449
 450         if (machine__is_host(machine))
 451                 return;
 452
 453         /*
 454          *As for guest kernel when processing subcommand record&report,
 455          *we arrange module mmap prior to guest kernel mmap and trigger
 456          *a preload dso because default guest module symbols are loaded
 457          *from guest kallsyms instead of /lib/modules/XXX/XXX. This
 458          *method is used to avoid symbol missing when the first addr is
 459          *in module instead of in guest kernel.
 460          */
 461         err = event__synthesize_modules(process_synthesized_event,
 462                                         psession, machine);
 463         if (err < 0)
 464                 pr_err("Couldn't record guest kernel [%d]'s reference"
 465                        " relocation symbol.\n", machine->pid);
 466
 467         /*
 468          * We use _stext for guest kernel because guest kernel's /proc/kallsyms
 469          * have no _text sometimes.
 470          */
 471         err = event__synthesize_kernel_mmap(process_synthesized_event,
 472                                             psession, machine, "_text");
 473         if (err < 0)
 474                 err = event__synthesize_kernel_mmap(process_synthesized_event,
 475                                                     psession, machine, "_stext");
 476         if (err < 0)
 477                 pr_err("Couldn't record guest kernel [%d]'s reference"
 478                        " relocation symbol.\n", machine->pid);
 479 }
 480
 481 static struct perf_event_header finished_round_event = {
 482         .size = sizeof(struct perf_event_header),
 483         .type = PERF_RECORD_FINISHED_ROUND,
 484 };
 485
 486 static void mmap_read_all(void)
 487 {
 488         int i;
 489
 490         for (i = 0; i < nr_cpu; i++) {
 491                 if (mmap_array[i].base)
 492                         mmap_read(&mmap_array[i]);
 493         }
 494
 495         if (perf_header__has_feat(&session->header, HEADER_TRACE_INFO))
 496                 write_output(&finished_round_event, sizeof(finished_round_event));
 497 }
 498
 499 static int __cmd_record(int argc, const char **argv)
 500 {
 501         int i, counter;
 502         struct stat st;
 503         int flags;
 504         int err;
 505         unsigned long waking = 0;
 506         int child_ready_pipe[2], go_pipe[2];
 507         const bool forks = argc > 0;
 508         char buf;
 509         struct machine *machine;
 510
 511         page_size = sysconf(_SC_PAGE_SIZE);
 512
 513         atexit(sig_atexit);
 514         signal(SIGCHLD, sig_handler);
 515         signal(SIGINT, sig_handler);
 516
 517         if (forks && (pipe(child_ready_pipe) < 0 || pipe(go_pipe) < 0)) {
 518                 perror("failed to create pipes");
 519                 exit(-1);
 520         }
 521
 522         if (!strcmp(output_name, "-"))
 523                 pipe_output = 1;
 524         else if (!stat(output_name, &st) && st.st_size) {
 525                 if (write_mode == WRITE_FORCE) {
 526                         char oldname[PATH_MAX];
 527                         snprintf(oldname, sizeof(oldname), "%s.old",
 528                                  output_name);
 529                         unlink(oldname);
 530                         rename(output_name, oldname);
 531                 }
 532         } else if (write_mode == WRITE_APPEND) {
 533                 write_mode = WRITE_FORCE;
 534         }
 535
 536         flags = O_CREAT|O_RDWR;
 537         if (write_mode == WRITE_APPEND)
 538                 file_new = 0;
 539         else
 540                 flags |= O_TRUNC;
 541
 542         if (pipe_output)
 543                 output = STDOUT_FILENO;
 544         else
 545                 output = open(output_name, flags, S_IRUSR | S_IWUSR);
 546         if (output < 0) {
 547                 perror("failed to create output file");
 548                 exit(-1);
 549         }
 550
 551         session = perf_session__new(output_name, O_WRONLY,
 552                                     write_mode == WRITE_FORCE, false);
 553         if (session == NULL) {
 554                 pr_err("Not enough memory for reading perf file header\n");
 555                 return -1;
 556         }
 557
 558         if (!file_new) {
 559                 err = perf_header__read(session, output);
 560                 if (err < 0)
 561                         return err;
 562         }
 563
 564         if (have_tracepoints(attrs, nr_counters))
 565                 perf_header__set_feat(&session->header, HEADER_TRACE_INFO);
 566
 567         atexit(atexit_header);
 568
 569         if (forks) {
 570                 child_pid = fork();
 571                 if (child_pid < 0) {
 572                         perror("failed to fork");
 573                         exit(-1);
 574                 }
 575
 576                 if (!child_pid) {
 577                         if (pipe_output)
 578                                 dup2(2, 1);
 579                         close(child_ready_pipe[0]);
 580                         close(go_pipe[1]);
 581                         fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
 582
 583                         /*
 584                          * Do a dummy execvp to get the PLT entry resolved,
 585                          * so we avoid the resolver overhead on the real
 586                          * execvp call.
 587                          */
 588                         execvp("", (char **)argv);
 589
 590                         /*
 591                          * Tell the parent we're ready to go
 592                          */
 593                         close(child_ready_pipe[1]);
 594
 595                         /*
 596                          * Wait until the parent tells us to go.
 597                          */
 598                         if (read(go_pipe[0], &buf, 1) == -1)
 599                                 perror("unable to read pipe");
 600
 601                         execvp(argv[0], (char **)argv);
 602
 603                         perror(argv[0]);
 604                         exit(-1);
 605                 }
 606
 607                 if (!system_wide && target_tid == -1 && target_pid == -1)
 608                         all_tids[0] = child_pid;
 609
 610                 close(child_ready_pipe[1]);
 611                 close(go_pipe[0]);
 612                 /*
 613                  * wait for child to settle
 614                  */
 615                 if (read(child_ready_pipe[0], &buf, 1) == -1) {
 616                         perror("unable to read pipe");
 617                         exit(-1);
 618                 }
 619                 close(child_ready_pipe[0]);
 620         }
 621
 622         nr_cpus = read_cpu_map(cpu_list);
 623         if (nr_cpus < 1) {
 624                 perror("failed to collect number of CPUs\n");
 625                 return -1;
 626         }
 627
 628         if (!system_wide && no_inherit && !cpu_list) {
 629                 open_counters(-1);
 630         } else {
 631                 for (i = 0; i < nr_cpus; i++)
 632                         open_counters(cpumap[i]);
 633         }
 634
 635         if (pipe_output) {
 636                 err = perf_header__write_pipe(output);
 637                 if (err < 0)
 638                         return err;
 639         } else if (file_new) {
 640                 err = perf_header__write(&session->header, output, false);
 641                 if (err < 0)
 642                         return err;
 643         }
 644
 645         post_processing_offset = lseek(output, 0, SEEK_CUR);
 646
 647         if (pipe_output) {
 648                 err = event__synthesize_attrs(&session->header,
 649                                               process_synthesized_event,
 650                                               session);
 651                 if (err < 0) {
 652                         pr_err("Couldn't synthesize attrs.\n");
 653                         return err;
 654                 }
 655
 656                 err = event__synthesize_event_types(process_synthesized_event,
 657                                                     session);
 658                 if (err < 0) {
 659                         pr_err("Couldn't synthesize event_types.\n");
 660                         return err;
 661                 }
 662
 663                 if (have_tracepoints(attrs, nr_counters)) {
 664                         /*
 665                          * FIXME err <= 0 here actually means that
 666                          * there were no tracepoints so its not really
 667                          * an error, just that we don't need to
 668                          * synthesize anything.  We really have to
 669                          * return this more properly and also
 670                          * propagate errors that now are calling die()
 671                          */
 672                         err = event__synthesize_tracing_data(output, attrs,
 673                                                              nr_counters,
 674                                                              process_synthesized_event,
 675                                                              session);
 676                         if (err <= 0) {
 677                                 pr_err("Couldn't record tracing data.\n");
 678                                 return err;
 679                         }
 680                         advance_output(err);
 681                 }
 682         }
 683
 684         machine = perf_session__find_host_machine(session);
 685         if (!machine) {
 686                 pr_err("Couldn't find native kernel information.\n");
 687                 return -1;
 688         }
 689
 690         err = event__synthesize_kernel_mmap(process_synthesized_event,
 691                                             session, machine, "_text");
 692         if (err < 0)
 693                 err = event__synthesize_kernel_mmap(process_synthesized_event,
 694                                                     session, machine, "_stext");
 695         if (err < 0) {
 696                 pr_err("Couldn't record kernel reference relocation symbol.\n");
 697                 return err;
 698         }
 699
 700         err = event__synthesize_modules(process_synthesized_event,
 701                                         session, machine);
 702         if (err < 0) {
 703                 pr_err("Couldn't record kernel reference relocation symbol.\n");
 704                 return err;
 705         }
 706         if (perf_guest)
 707                 perf_session__process_machines(session, event__synthesize_guest_os);
 708
 709         if (!system_wide)
 710                 event__synthesize_thread(target_tid, process_synthesized_event,
 711                                          session);
 712         else
 713                 event__synthesize_threads(process_synthesized_event, session);
 714
 715         if (realtime_prio) {
 716                 struct sched_param param;
 717
 718                 param.sched_priority = realtime_prio;
 719                 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
 720                         pr_err("Could not set realtime priority.\n");
 721                         exit(-1);
 722                 }
 723         }
 724
 725         /*
 726          * Let the child rip
 727          */
 728         if (forks)
 729                 close(go_pipe[1]);
 730
 731         for (;;) {
 732                 int hits = samples;
 733                 int thread;
 734
 735                 mmap_read_all();
 736
 737                 if (hits == samples) {
 738                         if (done)
 739                                 break;
 740                         err = poll(event_array, nr_poll, -1);
 741                         waking++;
 742                 }
 743
 744                 if (done) {
 745                         for (i = 0; i < nr_cpu; i++) {
 746                                 for (counter = 0;
 747                                         counter < nr_counters;
 748                                         counter++) {
 749                                         for (thread = 0;
 750                                                 thread < thread_num;
 751                                                 thread++)
 752                                                 ioctl(fd[i][counter][thread],
 753                                                         PERF_EVENT_IOC_DISABLE);
 754                                 }
 755                         }
 756                 }
 757         }
 758
 759         fprintf(stderr, "[ perf record: Woken up %ld times to write data ]\n", waking);
 760
 761         /*
 762          * Approximate RIP event size: 24 bytes.
 763          */
 764         fprintf(stderr,
 765                 "[ perf record: Captured and wrote %.3f MB %s (~%lld samples) ]\n",
 766                 (double)bytes_written / 1024.0 / 1024.0,
 767                 output_name,
 768                 bytes_written / 24);
 769
 770         return 0;
 771 }
 772
 773 static const char * const record_usage[] = {
 774         "perf record [<options>] [<command>]",
 775         "perf record [<options>] -- <command> [<options>]",
 776         NULL
 777 };
 778
 779 static bool force, append_file;
 780
 781 static const struct option options[] = {
 782         OPT_CALLBACK('e', "event", NULL, "event",
 783                      "event selector. use 'perf list' to list available events",
 784                      parse_events),
 785         OPT_CALLBACK(0, "filter", NULL, "filter",
 786                      "event filter", parse_filter),
 787         OPT_INTEGER('p', "pid", &target_pid,
 788                     "record events on existing process id"),
 789         OPT_INTEGER('t', "tid", &target_tid,
 790                     "record events on existing thread id"),
 791         OPT_INTEGER('r', "realtime", &realtime_prio,
 792                     "collect data with this RT SCHED_FIFO priority"),
 793         OPT_BOOLEAN('R', "raw-samples", &raw_samples,
 794                     "collect raw sample records from all opened counters"),
 795         OPT_BOOLEAN('a', "all-cpus", &system_wide,
 796                             "system-wide collection from all CPUs"),
 797         OPT_BOOLEAN('A', "append", &append_file,
 798                             "append to the output file to do incremental profiling"),
 799         OPT_STRING('C', "cpu", &cpu_list, "cpu",
 800                     "list of cpus to monitor"),
 801         OPT_BOOLEAN('f', "force", &force,
 802                         "overwrite existing data file (deprecated)"),
 803         OPT_U64('c', "count", &user_interval, "event period to sample"),
 804         OPT_STRING('o', "output", &output_name, "file",
 805                     "output file name"),
 806         OPT_BOOLEAN('i', "no-inherit", &no_inherit,
 807                     "child tasks do not inherit counters"),
 808         OPT_UINTEGER('F', "freq", &user_freq, "profile at this frequency"),
 809         OPT_UINTEGER('m', "mmap-pages", &mmap_pages, "number of mmap data pages"),
 810         OPT_BOOLEAN('g', "call-graph", &call_graph,
 811                     "do call-graph (stack chain/backtrace) recording"),
 812         OPT_INCR('v', "verbose", &verbose,
 813                     "be more verbose (show counter open errors, etc)"),
 814         OPT_BOOLEAN('s', "stat", &inherit_stat,
 815                     "per thread counts"),
 816         OPT_BOOLEAN('d', "data", &sample_address,
 817                     "Sample addresses"),
 818         OPT_BOOLEAN('n', "no-samples", &no_samples,
 819                     "don't sample"),
 820         OPT_BOOLEAN('N', "no-buildid-cache", &no_buildid,
 821                     "do not update the buildid cache"),
 822         OPT_END()
 823 };
 824
 825 int cmd_record(int argc, const char **argv, const char *prefix __used)
 826 {
 827         int i,j;
 828
 829         argc = parse_options(argc, argv, options, record_usage,
 830                             PARSE_OPT_STOP_AT_NON_OPTION);
 831         if (!argc && target_pid == -1 && target_tid == -1 &&
 832                 !system_wide && !cpu_list)
 833                 usage_with_options(record_usage, options);
 834
 835         if (force && append_file) {
 836                 fprintf(stderr, "Can't overwrite and append at the same time."
 837                                 " You need to choose between -f and -A");
 838                 usage_with_options(record_usage, options);
 839         } else if (append_file) {
 840                 write_mode = WRITE_APPEND;
 841         } else {
 842                 write_mode = WRITE_FORCE;
 843         }
 844
 845         symbol__init();
 846         if (no_buildid)
 847                 disable_buildid_cache();
 848
 849         if (!nr_counters) {
 850                 nr_counters     = 1;
 851                 attrs[0].type   = PERF_TYPE_HARDWARE;
 852                 attrs[0].config = PERF_COUNT_HW_CPU_CYCLES;
 853         }
 854
 855         if (target_pid != -1) {
 856                 target_tid = target_pid;
 857                 thread_num = find_all_tid(target_pid, &all_tids);
 858                 if (thread_num <= 0) {
 859                         fprintf(stderr, "Can't find all threads of pid %d\n",
 860                                         target_pid);
 861                         usage_with_options(record_usage, options);
 862                 }
 863         } else {
 864                 all_tids=malloc(sizeof(pid_t));
 865                 if (!all_tids)
 866                         return -ENOMEM;
 867
 868                 all_tids[0] = target_tid;
 869                 thread_num = 1;
 870         }
 871
 872         for (i = 0; i < MAX_NR_CPUS; i++) {
 873                 for (j = 0; j < MAX_COUNTERS; j++) {
 874                         fd[i][j] = malloc(sizeof(int)*thread_num);
 875                         if (!fd[i][j])
 876                                 return -ENOMEM;
 877                 }
 878         }
 879         event_array = malloc(
 880                 sizeof(struct pollfd)*MAX_NR_CPUS*MAX_COUNTERS*thread_num);
 881         if (!event_array)
 882                 return -ENOMEM;
 883
 884         if (user_interval != ULLONG_MAX)
 885                 default_interval = user_interval;
 886         if (user_freq != UINT_MAX)
 887                 freq = user_freq;
 888
 889         /*
 890          * User specified count overrides default frequency.
 891          */
 892         if (default_interval)
 893                 freq = 0;
 894         else if (freq) {
 895                 default_interval = freq;
 896         } else {
 897                 fprintf(stderr, "frequency and count are zero, aborting\n");
 898                 exit(EXIT_FAILURE);
 899         }
 900
 901         return __cmd_record(argc, argv);
 902 }