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