SAFE public projects git trees. - safe/jmp/linux-2.6/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/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
  29 enum write_mode_t {
  30         WRITE_FORCE,
  31         WRITE_APPEND
  32 };
  33
  34 static int                      *fd[MAX_NR_CPUS][MAX_COUNTERS];
  35
  36 static unsigned int             user_interval                   = UINT_MAX;
  37 static long                     default_interval                =      0;
  38
  39 static int                      nr_cpus                         =      0;
  40 static unsigned int             page_size;
  41 static unsigned int             mmap_pages                      =    128;
  42 static unsigned int             user_freq                       = UINT_MAX;
  43 static int                      freq                            =   1000;
  44 static int                      output;
  45 static int                      pipe_output                     =      0;
  46 static const char               *output_name                    = "perf.data";
  47 static int                      group                           =      0;
  48 static unsigned int             realtime_prio                   =      0;
  49 static bool                     raw_samples                     =  false;
  50 static bool                     system_wide                     =  false;
  51 static int                      profile_cpu                     =     -1;
  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                     inherit                         =   true;
  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                     multiplex                       =  false;
  64 static int                      multiplex_fd                    =     -1;
  65
  66 static long                     samples                         =      0;
  67 static struct timeval           last_read;
  68 static struct timeval           this_read;
  69
  70 static u64                      bytes_written                   =      0;
  71
  72 static struct pollfd            *event_array;
  73
  74 static int                      nr_poll                         =      0;
  75 static int                      nr_cpu                          =      0;
  76
  77 static int                      file_new                        =      1;
  78 static off_t                    post_processing_offset;
  79
  80 static struct perf_session      *session;
  81
  82 struct mmap_data {
  83         int                     counter;
  84         void                    *base;
  85         unsigned int            mask;
  86         unsigned int            prev;
  87 };
  88
  89 static struct mmap_data         *mmap_array[MAX_NR_CPUS][MAX_COUNTERS];
  90
  91 static unsigned long mmap_read_head(struct mmap_data *md)
  92 {
  93         struct perf_event_mmap_page *pc = md->base;
  94         long head;
  95
  96         head = pc->data_head;
  97         rmb();
  98
  99         return head;
 100 }
 101
 102 static void mmap_write_tail(struct mmap_data *md, unsigned long tail)
 103 {
 104         struct perf_event_mmap_page *pc = md->base;
 105
 106         /*
 107          * ensure all reads are done before we write the tail out.
 108          */
 109         /* mb(); */
 110         pc->data_tail = tail;
 111 }
 112
 113 static void advance_output(size_t size)
 114 {
 115         bytes_written += size;
 116 }
 117
 118 static void write_output(void *buf, size_t size)
 119 {
 120         while (size) {
 121                 int ret = write(output, buf, size);
 122
 123                 if (ret < 0)
 124                         die("failed to write");
 125
 126                 size -= ret;
 127                 buf += ret;
 128
 129                 bytes_written += ret;
 130         }
 131 }
 132
 133 static int process_synthesized_event(event_t *event,
 134                                      struct perf_session *self __used)
 135 {
 136         write_output(event, event->header.size);
 137         return 0;
 138 }
 139
 140 static void mmap_read(struct mmap_data *md)
 141 {
 142         unsigned int head = mmap_read_head(md);
 143         unsigned int old = md->prev;
 144         unsigned char *data = md->base + page_size;
 145         unsigned long size;
 146         void *buf;
 147         int diff;
 148
 149         gettimeofday(&this_read, NULL);
 150
 151         /*
 152          * If we're further behind than half the buffer, there's a chance
 153          * the writer will bite our tail and mess up the samples under us.
 154          *
 155          * If we somehow ended up ahead of the head, we got messed up.
 156          *
 157          * In either case, truncate and restart at head.
 158          */
 159         diff = head - old;
 160         if (diff < 0) {
 161                 struct timeval iv;
 162                 unsigned long msecs;
 163
 164                 timersub(&this_read, &last_read, &iv);
 165                 msecs = iv.tv_sec*1000 + iv.tv_usec/1000;
 166
 167                 fprintf(stderr, "WARNING: failed to keep up with mmap data."
 168                                 "  Last read %lu msecs ago.\n", msecs);
 169
 170                 /*
 171                  * head points to a known good entry, start there.
 172                  */
 173                 old = head;
 174         }
 175
 176         last_read = this_read;
 177
 178         if (old != head)
 179                 samples++;
 180
 181         size = head - old;
 182
 183         if ((old & md->mask) + size != (head & md->mask)) {
 184                 buf = &data[old & md->mask];
 185                 size = md->mask + 1 - (old & md->mask);
 186                 old += size;
 187
 188                 write_output(buf, size);
 189         }
 190
 191         buf = &data[old & md->mask];
 192         size = head - old;
 193         old += size;
 194
 195         write_output(buf, size);
 196
 197         md->prev = old;
 198         mmap_write_tail(md, old);
 199 }
 200
 201 static volatile int done = 0;
 202 static volatile int signr = -1;
 203
 204 static void sig_handler(int sig)
 205 {
 206         done = 1;
 207         signr = sig;
 208 }
 209
 210 static void sig_atexit(void)
 211 {
 212         if (child_pid != -1)
 213                 kill(child_pid, SIGTERM);
 214
 215         if (signr == -1)
 216                 return;
 217
 218         signal(signr, SIG_DFL);
 219         kill(getpid(), signr);
 220 }
 221
 222 static int group_fd;
 223
 224 static struct perf_header_attr *get_header_attr(struct perf_event_attr *a, int nr)
 225 {
 226         struct perf_header_attr *h_attr;
 227
 228         if (nr < session->header.attrs) {
 229                 h_attr = session->header.attr[nr];
 230         } else {
 231                 h_attr = perf_header_attr__new(a);
 232                 if (h_attr != NULL)
 233                         if (perf_header__add_attr(&session->header, h_attr) < 0) {
 234                                 perf_header_attr__delete(h_attr);
 235                                 h_attr = NULL;
 236                         }
 237         }
 238
 239         return h_attr;
 240 }
 241
 242 static void create_counter(int counter, int cpu)
 243 {
 244         char *filter = filters[counter];
 245         struct perf_event_attr *attr = attrs + counter;
 246         struct perf_header_attr *h_attr;
 247         int track = !counter; /* only the first counter needs these */
 248         int thread_index;
 249         int ret;
 250         struct {
 251                 u64 count;
 252                 u64 time_enabled;
 253                 u64 time_running;
 254                 u64 id;
 255         } read_data;
 256
 257         attr->read_format       = PERF_FORMAT_TOTAL_TIME_ENABLED |
 258                                   PERF_FORMAT_TOTAL_TIME_RUNNING |
 259                                   PERF_FORMAT_ID;
 260
 261         attr->sample_type       |= PERF_SAMPLE_IP | PERF_SAMPLE_TID;
 262
 263         if (nr_counters > 1)
 264                 attr->sample_type |= PERF_SAMPLE_ID;
 265
 266         /*
 267          * We default some events to a 1 default interval. But keep
 268          * it a weak assumption overridable by the user.
 269          */
 270         if (!attr->sample_period || (user_freq != UINT_MAX &&
 271                                      user_interval != UINT_MAX)) {
 272                 if (freq) {
 273                         attr->sample_type       |= PERF_SAMPLE_PERIOD;
 274                         attr->freq              = 1;
 275                         attr->sample_freq       = freq;
 276                 } else {
 277                         attr->sample_period = default_interval;
 278                 }
 279         }
 280
 281         if (no_samples)
 282                 attr->sample_freq = 0;
 283
 284         if (inherit_stat)
 285                 attr->inherit_stat = 1;
 286
 287         if (sample_address)
 288                 attr->sample_type       |= PERF_SAMPLE_ADDR;
 289
 290         if (call_graph)
 291                 attr->sample_type       |= PERF_SAMPLE_CALLCHAIN;
 292
 293         if (raw_samples) {
 294                 attr->sample_type       |= PERF_SAMPLE_TIME;
 295                 attr->sample_type       |= PERF_SAMPLE_RAW;
 296                 attr->sample_type       |= PERF_SAMPLE_CPU;
 297         }
 298
 299         attr->mmap              = track;
 300         attr->comm              = track;
 301         attr->inherit           = inherit;
 302         if (target_pid == -1 && !system_wide) {
 303                 attr->disabled = 1;
 304                 attr->enable_on_exec = 1;
 305         }
 306
 307         for (thread_index = 0; thread_index < thread_num; thread_index++) {
 308 try_again:
 309                 fd[nr_cpu][counter][thread_index] = sys_perf_event_open(attr,
 310                                 all_tids[thread_index], cpu, group_fd, 0);
 311
 312                 if (fd[nr_cpu][counter][thread_index] < 0) {
 313                         int err = errno;
 314
 315                         if (err == EPERM || err == EACCES)
 316                                 die("Permission error - are you root?\n"
 317                                         "\t Consider tweaking"
 318                                         " /proc/sys/kernel/perf_event_paranoid.\n");
 319                         else if (err ==  ENODEV && profile_cpu != -1) {
 320                                 die("No such device - did you specify"
 321                                         " an out-of-range profile CPU?\n");
 322                         }
 323
 324                         /*
 325                          * If it's cycles then fall back to hrtimer
 326                          * based cpu-clock-tick sw counter, which
 327                          * is always available even if no PMU support:
 328                          */
 329                         if (attr->type == PERF_TYPE_HARDWARE
 330                                         && attr->config == PERF_COUNT_HW_CPU_CYCLES) {
 331
 332                                 if (verbose)
 333                                         warning(" ... trying to fall back to cpu-clock-ticks\n");
 334                                 attr->type = PERF_TYPE_SOFTWARE;
 335                                 attr->config = PERF_COUNT_SW_CPU_CLOCK;
 336                                 goto try_again;
 337                         }
 338                         printf("\n");
 339                         error("perfcounter syscall returned with %d (%s)\n",
 340                                         fd[nr_cpu][counter][thread_index], strerror(err));
 341
 342 #if defined(__i386__) || defined(__x86_64__)
 343                         if (attr->type == PERF_TYPE_HARDWARE && err == EOPNOTSUPP)
 344                                 die("No hardware sampling interrupt available."
 345                                     " No APIC? If so then you can boot the kernel"
 346                                     " with the \"lapic\" boot parameter to"
 347                                     " force-enable it.\n");
 348 #endif
 349
 350                         die("No CONFIG_PERF_EVENTS=y kernel support configured?\n");
 351                         exit(-1);
 352                 }
 353
 354                 h_attr = get_header_attr(attr, counter);
 355                 if (h_attr == NULL)
 356                         die("nomem\n");
 357
 358                 if (!file_new) {
 359                         if (memcmp(&h_attr->attr, attr, sizeof(*attr))) {
 360                                 fprintf(stderr, "incompatible append\n");
 361                                 exit(-1);
 362                         }
 363                 }
 364
 365                 if (read(fd[nr_cpu][counter][thread_index], &read_data, sizeof(read_data)) == -1) {
 366                         perror("Unable to read perf file descriptor\n");
 367                         exit(-1);
 368                 }
 369
 370                 if (perf_header_attr__add_id(h_attr, read_data.id) < 0) {
 371                         pr_warning("Not enough memory to add id\n");
 372                         exit(-1);
 373                 }
 374
 375                 assert(fd[nr_cpu][counter][thread_index] >= 0);
 376                 fcntl(fd[nr_cpu][counter][thread_index], F_SETFL, O_NONBLOCK);
 377
 378                 /*
 379                  * First counter acts as the group leader:
 380                  */
 381                 if (group && group_fd == -1)
 382                         group_fd = fd[nr_cpu][counter][thread_index];
 383                 if (multiplex && multiplex_fd == -1)
 384                         multiplex_fd = fd[nr_cpu][counter][thread_index];
 385
 386                 if (multiplex && fd[nr_cpu][counter][thread_index] != multiplex_fd) {
 387
 388                         ret = ioctl(fd[nr_cpu][counter][thread_index], PERF_EVENT_IOC_SET_OUTPUT, multiplex_fd);
 389                         assert(ret != -1);
 390                 } else {
 391                         event_array[nr_poll].fd = fd[nr_cpu][counter][thread_index];
 392                         event_array[nr_poll].events = POLLIN;
 393                         nr_poll++;
 394
 395                         mmap_array[nr_cpu][counter][thread_index].counter = counter;
 396                         mmap_array[nr_cpu][counter][thread_index].prev = 0;
 397                         mmap_array[nr_cpu][counter][thread_index].mask = mmap_pages*page_size - 1;
 398                         mmap_array[nr_cpu][counter][thread_index].base = mmap(NULL, (mmap_pages+1)*page_size,
 399                                 PROT_READ|PROT_WRITE, MAP_SHARED, fd[nr_cpu][counter][thread_index], 0);
 400                         if (mmap_array[nr_cpu][counter][thread_index].base == MAP_FAILED) {
 401                                 error("failed to mmap with %d (%s)\n", errno, strerror(errno));
 402                                 exit(-1);
 403                         }
 404                 }
 405
 406                 if (filter != NULL) {
 407                         ret = ioctl(fd[nr_cpu][counter][thread_index],
 408                                         PERF_EVENT_IOC_SET_FILTER, filter);
 409                         if (ret) {
 410                                 error("failed to set filter with %d (%s)\n", errno,
 411                                                 strerror(errno));
 412                                 exit(-1);
 413                         }
 414                 }
 415         }
 416 }
 417
 418 static void open_counters(int cpu)
 419 {
 420         int counter;
 421
 422         group_fd = -1;
 423         for (counter = 0; counter < nr_counters; counter++)
 424                 create_counter(counter, cpu);
 425
 426         nr_cpu++;
 427 }
 428
 429 static int process_buildids(void)
 430 {
 431         u64 size = lseek(output, 0, SEEK_CUR);
 432
 433         if (size == 0)
 434                 return 0;
 435
 436         session->fd = output;
 437         return __perf_session__process_events(session, post_processing_offset,
 438                                               size - post_processing_offset,
 439                                               size, &build_id__mark_dso_hit_ops);
 440 }
 441
 442 static void atexit_header(void)
 443 {
 444         if (!pipe_output) {
 445                 session->header.data_size += bytes_written;
 446
 447                 process_buildids();
 448                 perf_header__write(&session->header, output, true);
 449         } else {
 450                 int err;
 451
 452                 err = event__synthesize_build_ids(process_synthesized_event,
 453                                                   session);
 454                 if (err < 0)
 455                         pr_err("Couldn't synthesize build ids.\n");
 456         }
 457 }
 458
 459 static int __cmd_record(int argc, const char **argv)
 460 {
 461         int i, counter;
 462         struct stat st;
 463         pid_t pid = 0;
 464         int flags;
 465         int err;
 466         unsigned long waking = 0;
 467         int child_ready_pipe[2], go_pipe[2];
 468         const bool forks = argc > 0;
 469         char buf;
 470
 471         page_size = sysconf(_SC_PAGE_SIZE);
 472
 473         atexit(sig_atexit);
 474         signal(SIGCHLD, sig_handler);
 475         signal(SIGINT, sig_handler);
 476
 477         if (forks && (pipe(child_ready_pipe) < 0 || pipe(go_pipe) < 0)) {
 478                 perror("failed to create pipes");
 479                 exit(-1);
 480         }
 481
 482         if (!strcmp(output_name, "-"))
 483                 pipe_output = 1;
 484         else if (!stat(output_name, &st) && st.st_size) {
 485                 if (write_mode == WRITE_FORCE) {
 486                         char oldname[PATH_MAX];
 487                         snprintf(oldname, sizeof(oldname), "%s.old",
 488                                  output_name);
 489                         unlink(oldname);
 490                         rename(output_name, oldname);
 491                 }
 492         } else if (write_mode == WRITE_APPEND) {
 493                 write_mode = WRITE_FORCE;
 494         }
 495
 496         flags = O_CREAT|O_RDWR;
 497         if (write_mode == WRITE_APPEND)
 498                 file_new = 0;
 499         else
 500                 flags |= O_TRUNC;
 501
 502         if (pipe_output)
 503                 output = STDOUT_FILENO;
 504         else
 505                 output = open(output_name, flags, S_IRUSR | S_IWUSR);
 506         if (output < 0) {
 507                 perror("failed to create output file");
 508                 exit(-1);
 509         }
 510
 511         session = perf_session__new(output_name, O_WRONLY,
 512                                     write_mode == WRITE_FORCE);
 513         if (session == NULL) {
 514                 pr_err("Not enough memory for reading perf file header\n");
 515                 return -1;
 516         }
 517
 518         if (!file_new) {
 519                 err = perf_header__read(session, output);
 520                 if (err < 0)
 521                         return err;
 522         }
 523
 524         if (raw_samples) {
 525                 perf_header__set_feat(&session->header, HEADER_TRACE_INFO);
 526         } else {
 527                 for (i = 0; i < nr_counters; i++) {
 528                         if (attrs[i].sample_type & PERF_SAMPLE_RAW) {
 529                                 perf_header__set_feat(&session->header, HEADER_TRACE_INFO);
 530                                 break;
 531                         }
 532                 }
 533         }
 534
 535         atexit(atexit_header);
 536
 537         if (forks) {
 538                 child_pid = fork();
 539                 if (pid < 0) {
 540                         perror("failed to fork");
 541                         exit(-1);
 542                 }
 543
 544                 if (!child_pid) {
 545                         if (pipe_output)
 546                                 dup2(2, 1);
 547                         close(child_ready_pipe[0]);
 548                         close(go_pipe[1]);
 549                         fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
 550
 551                         /*
 552                          * Do a dummy execvp to get the PLT entry resolved,
 553                          * so we avoid the resolver overhead on the real
 554                          * execvp call.
 555                          */
 556                         execvp("", (char **)argv);
 557
 558                         /*
 559                          * Tell the parent we're ready to go
 560                          */
 561                         close(child_ready_pipe[1]);
 562
 563                         /*
 564                          * Wait until the parent tells us to go.
 565                          */
 566                         if (read(go_pipe[0], &buf, 1) == -1)
 567                                 perror("unable to read pipe");
 568
 569                         execvp(argv[0], (char **)argv);
 570
 571                         perror(argv[0]);
 572                         exit(-1);
 573                 }
 574
 575                 if (!system_wide && target_tid == -1 && target_pid == -1)
 576                         all_tids[0] = child_pid;
 577
 578                 close(child_ready_pipe[1]);
 579                 close(go_pipe[0]);
 580                 /*
 581                  * wait for child to settle
 582                  */
 583                 if (read(child_ready_pipe[0], &buf, 1) == -1) {
 584                         perror("unable to read pipe");
 585                         exit(-1);
 586                 }
 587                 close(child_ready_pipe[0]);
 588         }
 589
 590         if ((!system_wide && !inherit) || profile_cpu != -1) {
 591                 open_counters(profile_cpu);
 592         } else {
 593                 nr_cpus = read_cpu_map();
 594                 for (i = 0; i < nr_cpus; i++)
 595                         open_counters(cpumap[i]);
 596         }
 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, output, false);
 604                 if (err < 0)
 605                         return err;
 606         }
 607
 608         post_processing_offset = lseek(output, 0, SEEK_CUR);
 609
 610         if (pipe_output) {
 611                 err = event__synthesize_attrs(&session->header,
 612                                               process_synthesized_event,
 613                                               session);
 614                 if (err < 0) {
 615                         pr_err("Couldn't synthesize attrs.\n");
 616                         return err;
 617                 }
 618
 619                 err = event__synthesize_event_types(process_synthesized_event,
 620                                                     session);
 621                 if (err < 0) {
 622                         pr_err("Couldn't synthesize event_types.\n");
 623                         return err;
 624                 }
 625
 626                 err = event__synthesize_tracing_data(output, attrs,
 627                                                      nr_counters,
 628                                                      process_synthesized_event,
 629                                                      session);
 630                 if (err <= 0) {
 631                         pr_err("Couldn't record tracing data.\n");
 632                         return err;
 633                 }
 634
 635                 advance_output(err);
 636         }
 637
 638         err = event__synthesize_kernel_mmap(process_synthesized_event,
 639                                             session, "_text");
 640         if (err < 0)
 641                 err = event__synthesize_kernel_mmap(process_synthesized_event,
 642                                                     session, "_stext");
 643         if (err < 0) {
 644                 pr_err("Couldn't record kernel reference relocation symbol.\n");
 645                 return err;
 646         }
 647
 648         err = event__synthesize_modules(process_synthesized_event, session);
 649         if (err < 0) {
 650                 pr_err("Couldn't record kernel reference relocation symbol.\n");
 651                 return err;
 652         }
 653
 654         if (!system_wide && profile_cpu == -1)
 655                 event__synthesize_thread(target_tid, process_synthesized_event,
 656                                          session);
 657         else
 658                 event__synthesize_threads(process_synthesized_event, session);
 659
 660         if (realtime_prio) {
 661                 struct sched_param param;
 662
 663                 param.sched_priority = realtime_prio;
 664                 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
 665                         pr_err("Could not set realtime priority.\n");
 666                         exit(-1);
 667                 }
 668         }
 669
 670         /*
 671          * Let the child rip
 672          */
 673         if (forks)
 674                 close(go_pipe[1]);
 675
 676         for (;;) {
 677                 int hits = samples;
 678                 int thread;
 679
 680                 for (i = 0; i < nr_cpu; i++) {
 681                         for (counter = 0; counter < nr_counters; counter++) {
 682                                 for (thread = 0;
 683                                         thread < thread_num; thread++) {
 684                                         if (mmap_array[i][counter][thread].base)
 685                                                 mmap_read(&mmap_array[i][counter][thread]);
 686                                 }
 687
 688                         }
 689                 }
 690
 691                 if (hits == samples) {
 692                         if (done)
 693                                 break;
 694                         err = poll(event_array, nr_poll, -1);
 695                         waking++;
 696                 }
 697
 698                 if (done) {
 699                         for (i = 0; i < nr_cpu; i++) {
 700                                 for (counter = 0;
 701                                         counter < nr_counters;
 702                                         counter++) {
 703                                         for (thread = 0;
 704                                                 thread < thread_num;
 705                                                 thread++)
 706                                                 ioctl(fd[i][counter][thread],
 707                                                         PERF_EVENT_IOC_DISABLE);
 708                                 }
 709                         }
 710                 }
 711         }
 712
 713         fprintf(stderr, "[ perf record: Woken up %ld times to write data ]\n", waking);
 714
 715         /*
 716          * Approximate RIP event size: 24 bytes.
 717          */
 718         fprintf(stderr,
 719                 "[ perf record: Captured and wrote %.3f MB %s (~%lld samples) ]\n",
 720                 (double)bytes_written / 1024.0 / 1024.0,
 721                 output_name,
 722                 bytes_written / 24);
 723
 724         return 0;
 725 }
 726
 727 static const char * const record_usage[] = {
 728         "perf record [<options>] [<command>]",
 729         "perf record [<options>] -- <command> [<options>]",
 730         NULL
 731 };
 732
 733 static bool force, append_file;
 734
 735 static const struct option options[] = {
 736         OPT_CALLBACK('e', "event", NULL, "event",
 737                      "event selector. use 'perf list' to list available events",
 738                      parse_events),
 739         OPT_CALLBACK(0, "filter", NULL, "filter",
 740                      "event filter", parse_filter),
 741         OPT_INTEGER('p', "pid", &target_pid,
 742                     "record events on existing process id"),
 743         OPT_INTEGER('t', "tid", &target_tid,
 744                     "record events on existing thread id"),
 745         OPT_INTEGER('r', "realtime", &realtime_prio,
 746                     "collect data with this RT SCHED_FIFO priority"),
 747         OPT_BOOLEAN('R', "raw-samples", &raw_samples,
 748                     "collect raw sample records from all opened counters"),
 749         OPT_BOOLEAN('a', "all-cpus", &system_wide,
 750                             "system-wide collection from all CPUs"),
 751         OPT_BOOLEAN('A', "append", &append_file,
 752                             "append to the output file to do incremental profiling"),
 753         OPT_INTEGER('C', "profile_cpu", &profile_cpu,
 754                             "CPU to profile on"),
 755         OPT_BOOLEAN('f', "force", &force,
 756                         "overwrite existing data file (deprecated)"),
 757         OPT_LONG('c', "count", &user_interval,
 758                     "event period to sample"),
 759         OPT_STRING('o', "output", &output_name, "file",
 760                     "output file name"),
 761         OPT_BOOLEAN('i', "inherit", &inherit,
 762                     "child tasks inherit counters"),
 763         OPT_INTEGER('F', "freq", &user_freq,
 764                     "profile at this frequency"),
 765         OPT_INTEGER('m', "mmap-pages", &mmap_pages,
 766                     "number of mmap data pages"),
 767         OPT_BOOLEAN('g', "call-graph", &call_graph,
 768                     "do call-graph (stack chain/backtrace) recording"),
 769         OPT_INCR('v', "verbose", &verbose,
 770                     "be more verbose (show counter open errors, etc)"),
 771         OPT_BOOLEAN('s', "stat", &inherit_stat,
 772                     "per thread counts"),
 773         OPT_BOOLEAN('d', "data", &sample_address,
 774                     "Sample addresses"),
 775         OPT_BOOLEAN('n', "no-samples", &no_samples,
 776                     "don't sample"),
 777         OPT_BOOLEAN('M', "multiplex", &multiplex,
 778                     "multiplex counter output in a single channel"),
 779         OPT_END()
 780 };
 781
 782 int cmd_record(int argc, const char **argv, const char *prefix __used)
 783 {
 784         int i,j;
 785
 786         argc = parse_options(argc, argv, options, record_usage,
 787                             PARSE_OPT_STOP_AT_NON_OPTION);
 788         if (!argc && target_pid == -1 && target_tid == -1 &&
 789                 !system_wide && profile_cpu == -1)
 790                 usage_with_options(record_usage, options);
 791
 792         if (force && append_file) {
 793                 fprintf(stderr, "Can't overwrite and append at the same time."
 794                                 " You need to choose between -f and -A");
 795                 usage_with_options(record_usage, options);
 796         } else if (append_file) {
 797                 write_mode = WRITE_APPEND;
 798         } else {
 799                 write_mode = WRITE_FORCE;
 800         }
 801
 802         symbol__init();
 803
 804         if (!nr_counters) {
 805                 nr_counters     = 1;
 806                 attrs[0].type   = PERF_TYPE_HARDWARE;
 807                 attrs[0].config = PERF_COUNT_HW_CPU_CYCLES;
 808         }
 809
 810         if (target_pid != -1) {
 811                 target_tid = target_pid;
 812                 thread_num = find_all_tid(target_pid, &all_tids);
 813                 if (thread_num <= 0) {
 814                         fprintf(stderr, "Can't find all threads of pid %d\n",
 815                                         target_pid);
 816                         usage_with_options(record_usage, options);
 817                 }
 818         } else {
 819                 all_tids=malloc(sizeof(pid_t));
 820                 if (!all_tids)
 821                         return -ENOMEM;
 822
 823                 all_tids[0] = target_tid;
 824                 thread_num = 1;
 825         }
 826
 827         for (i = 0; i < MAX_NR_CPUS; i++) {
 828                 for (j = 0; j < MAX_COUNTERS; j++) {
 829                         fd[i][j] = malloc(sizeof(int)*thread_num);
 830                         mmap_array[i][j] = zalloc(
 831                                 sizeof(struct mmap_data)*thread_num);
 832                         if (!fd[i][j] || !mmap_array[i][j])
 833                                 return -ENOMEM;
 834                 }
 835         }
 836         event_array = malloc(
 837                 sizeof(struct pollfd)*MAX_NR_CPUS*MAX_COUNTERS*thread_num);
 838         if (!event_array)
 839                 return -ENOMEM;
 840
 841         if (user_interval != UINT_MAX)
 842                 default_interval = user_interval;
 843         if (user_freq != UINT_MAX)
 844                 freq = user_freq;
 845
 846         /*
 847          * User specified count overrides default frequency.
 848          */
 849         if (default_interval)
 850                 freq = 0;
 851         else if (freq) {
 852                 default_interval = freq;
 853         } else {
 854                 fprintf(stderr, "frequency and count are zero, aborting\n");
 855                 exit(EXIT_FAILURE);
 856         }
 857
 858         return __cmd_record(argc, argv);
 859 }