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
  64 static long                     samples                         =      0;
  65 static u64                      bytes_written                   =      0;
  66
  67 static struct pollfd            *event_array;
  68
  69 static int                      nr_poll                         =      0;
  70 static int                      nr_cpu                          =      0;
  71
  72 static int                      file_new                        =      1;
  73 static off_t                    post_processing_offset;
  74
  75 static struct perf_session      *session;
  76 static const char               *cpu_list;
  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                 attr->mmap_data = track;
 274         }
 275
 276         if (call_graph)
 277                 attr->sample_type       |= PERF_SAMPLE_CALLCHAIN;
 278
 279         if (system_wide)
 280                 attr->sample_type       |= PERF_SAMPLE_CPU;
 281
 282         if (raw_samples) {
 283                 attr->sample_type       |= PERF_SAMPLE_TIME;
 284                 attr->sample_type       |= PERF_SAMPLE_RAW;
 285                 attr->sample_type       |= PERF_SAMPLE_CPU;
 286         }
 287
 288         attr->mmap              = track;
 289         attr->comm              = track;
 290         attr->inherit           = !no_inherit;
 291         if (target_pid == -1 && target_tid == -1 && !system_wide) {
 292                 attr->disabled = 1;
 293                 attr->enable_on_exec = 1;
 294         }
 295
 296         for (thread_index = 0; thread_index < thread_num; thread_index++) {
 297 try_again:
 298                 fd[nr_cpu][counter][thread_index] = sys_perf_event_open(attr,
 299                                 all_tids[thread_index], cpu, group_fd, 0);
 300
 301                 if (fd[nr_cpu][counter][thread_index] < 0) {
 302                         int err = errno;
 303
 304                         if (err == EPERM || err == EACCES)
 305                                 die("Permission error - are you root?\n"
 306                                         "\t Consider tweaking"
 307                                         " /proc/sys/kernel/perf_event_paranoid.\n");
 308                         else if (err ==  ENODEV && cpu_list) {
 309                                 die("No such device - did you specify"
 310                                         " an out-of-range profile CPU?\n");
 311                         }
 312
 313                         /*
 314                          * If it's cycles then fall back to hrtimer
 315                          * based cpu-clock-tick sw counter, which
 316                          * is always available even if no PMU support:
 317                          */
 318                         if (attr->type == PERF_TYPE_HARDWARE
 319                                         && attr->config == PERF_COUNT_HW_CPU_CYCLES) {
 320
 321                                 if (verbose)
 322                                         warning(" ... trying to fall back to cpu-clock-ticks\n");
 323                                 attr->type = PERF_TYPE_SOFTWARE;
 324                                 attr->config = PERF_COUNT_SW_CPU_CLOCK;
 325                                 goto try_again;
 326                         }
 327                         printf("\n");
 328                         error("perfcounter syscall returned with %d (%s)\n",
 329                                         fd[nr_cpu][counter][thread_index], strerror(err));
 330
 331 #if defined(__i386__) || defined(__x86_64__)
 332                         if (attr->type == PERF_TYPE_HARDWARE && err == EOPNOTSUPP)
 333                                 die("No hardware sampling interrupt available."
 334                                     " No APIC? If so then you can boot the kernel"
 335                                     " with the \"lapic\" boot parameter to"
 336                                     " force-enable it.\n");
 337 #endif
 338
 339                         die("No CONFIG_PERF_EVENTS=y kernel support configured?\n");
 340                         exit(-1);
 341                 }
 342
 343                 h_attr = get_header_attr(attr, counter);
 344                 if (h_attr == NULL)
 345                         die("nomem\n");
 346
 347                 if (!file_new) {
 348                         if (memcmp(&h_attr->attr, attr, sizeof(*attr))) {
 349                                 fprintf(stderr, "incompatible append\n");
 350                                 exit(-1);
 351                         }
 352                 }
 353
 354                 if (read(fd[nr_cpu][counter][thread_index], &read_data, sizeof(read_data)) == -1) {
 355                         perror("Unable to read perf file descriptor\n");
 356                         exit(-1);
 357                 }
 358
 359                 if (perf_header_attr__add_id(h_attr, read_data.id) < 0) {
 360                         pr_warning("Not enough memory to add id\n");
 361                         exit(-1);
 362                 }
 363
 364                 assert(fd[nr_cpu][counter][thread_index] >= 0);
 365                 fcntl(fd[nr_cpu][counter][thread_index], F_SETFL, O_NONBLOCK);
 366
 367                 /*
 368                  * First counter acts as the group leader:
 369                  */
 370                 if (group && group_fd == -1)
 371                         group_fd = fd[nr_cpu][counter][thread_index];
 372
 373                 if (counter || thread_index) {
 374                         ret = ioctl(fd[nr_cpu][counter][thread_index],
 375                                         PERF_EVENT_IOC_SET_OUTPUT,
 376                                         fd[nr_cpu][0][0]);
 377                         if (ret) {
 378                                 error("failed to set output: %d (%s)\n", errno,
 379                                                 strerror(errno));
 380                                 exit(-1);
 381                         }
 382                 } else {
 383                         mmap_array[nr_cpu].counter = counter;
 384                         mmap_array[nr_cpu].prev = 0;
 385                         mmap_array[nr_cpu].mask = mmap_pages*page_size - 1;
 386                         mmap_array[nr_cpu].base = mmap(NULL, (mmap_pages+1)*page_size,
 387                                 PROT_READ|PROT_WRITE, MAP_SHARED, fd[nr_cpu][counter][thread_index], 0);
 388                         if (mmap_array[nr_cpu].base == MAP_FAILED) {
 389                                 error("failed to mmap with %d (%s)\n", errno, strerror(errno));
 390                                 exit(-1);
 391                         }
 392
 393                         event_array[nr_poll].fd = fd[nr_cpu][counter][thread_index];
 394                         event_array[nr_poll].events = POLLIN;
 395                         nr_poll++;
 396                 }
 397
 398                 if (filter != NULL) {
 399                         ret = ioctl(fd[nr_cpu][counter][thread_index],
 400                                         PERF_EVENT_IOC_SET_FILTER, filter);
 401                         if (ret) {
 402                                 error("failed to set filter with %d (%s)\n", errno,
 403                                                 strerror(errno));
 404                                 exit(-1);
 405                         }
 406                 }
 407         }
 408 }
 409
 410 static void open_counters(int cpu)
 411 {
 412         int counter;
 413
 414         group_fd = -1;
 415         for (counter = 0; counter < nr_counters; counter++)
 416                 create_counter(counter, cpu);
 417
 418         nr_cpu++;
 419 }
 420
 421 static int process_buildids(void)
 422 {
 423         u64 size = lseek(output, 0, SEEK_CUR);
 424
 425         if (size == 0)
 426                 return 0;
 427
 428         session->fd = output;
 429         return __perf_session__process_events(session, post_processing_offset,
 430                                               size - post_processing_offset,
 431                                               size, &build_id__mark_dso_hit_ops);
 432 }
 433
 434 static void atexit_header(void)
 435 {
 436         if (!pipe_output) {
 437                 session->header.data_size += bytes_written;
 438
 439                 process_buildids();
 440                 perf_header__write(&session->header, output, true);
 441         }
 442 }
 443
 444 static void event__synthesize_guest_os(struct machine *machine, void *data)
 445 {
 446         int err;
 447         char *guest_kallsyms;
 448         char path[PATH_MAX];
 449         struct perf_session *psession = data;
 450
 451         if (machine__is_host(machine))
 452                 return;
 453
 454         /*
 455          *As for guest kernel when processing subcommand record&report,
 456          *we arrange module mmap prior to guest kernel mmap and trigger
 457          *a preload dso because default guest module symbols are loaded
 458          *from guest kallsyms instead of /lib/modules/XXX/XXX. This
 459          *method is used to avoid symbol missing when the first addr is
 460          *in module instead of in guest kernel.
 461          */
 462         err = event__synthesize_modules(process_synthesized_event,
 463                                         psession, machine);
 464         if (err < 0)
 465                 pr_err("Couldn't record guest kernel [%d]'s reference"
 466                        " relocation symbol.\n", machine->pid);
 467
 468         if (machine__is_default_guest(machine))
 469                 guest_kallsyms = (char *) symbol_conf.default_guest_kallsyms;
 470         else {
 471                 sprintf(path, "%s/proc/kallsyms", machine->root_dir);
 472                 guest_kallsyms = path;
 473         }
 474
 475         /*
 476          * We use _stext for guest kernel because guest kernel's /proc/kallsyms
 477          * have no _text sometimes.
 478          */
 479         err = event__synthesize_kernel_mmap(process_synthesized_event,
 480                                             psession, machine, "_text");
 481         if (err < 0)
 482                 err = event__synthesize_kernel_mmap(process_synthesized_event,
 483                                                     psession, machine, "_stext");
 484         if (err < 0)
 485                 pr_err("Couldn't record guest kernel [%d]'s reference"
 486                        " relocation symbol.\n", machine->pid);
 487 }
 488
 489 static struct perf_event_header finished_round_event = {
 490         .size = sizeof(struct perf_event_header),
 491         .type = PERF_RECORD_FINISHED_ROUND,
 492 };
 493
 494 static void mmap_read_all(void)
 495 {
 496         int i;
 497
 498         for (i = 0; i < nr_cpu; i++) {
 499                 if (mmap_array[i].base)
 500                         mmap_read(&mmap_array[i]);
 501         }
 502
 503         if (perf_header__has_feat(&session->header, HEADER_TRACE_INFO))
 504                 write_output(&finished_round_event, sizeof(finished_round_event));
 505 }
 506
 507 static int __cmd_record(int argc, const char **argv)
 508 {
 509         int i, counter;
 510         struct stat st;
 511         int flags;
 512         int err;
 513         unsigned long waking = 0;
 514         int child_ready_pipe[2], go_pipe[2];
 515         const bool forks = argc > 0;
 516         char buf;
 517         struct machine *machine;
 518
 519         page_size = sysconf(_SC_PAGE_SIZE);
 520
 521         atexit(sig_atexit);
 522         signal(SIGCHLD, sig_handler);
 523         signal(SIGINT, sig_handler);
 524
 525         if (forks && (pipe(child_ready_pipe) < 0 || pipe(go_pipe) < 0)) {
 526                 perror("failed to create pipes");
 527                 exit(-1);
 528         }
 529
 530         if (!strcmp(output_name, "-"))
 531                 pipe_output = 1;
 532         else if (!stat(output_name, &st) && st.st_size) {
 533                 if (write_mode == WRITE_FORCE) {
 534                         char oldname[PATH_MAX];
 535                         snprintf(oldname, sizeof(oldname), "%s.old",
 536                                  output_name);
 537                         unlink(oldname);
 538                         rename(output_name, oldname);
 539                 }
 540         } else if (write_mode == WRITE_APPEND) {
 541                 write_mode = WRITE_FORCE;
 542         }
 543
 544         flags = O_CREAT|O_RDWR;
 545         if (write_mode == WRITE_APPEND)
 546                 file_new = 0;
 547         else
 548                 flags |= O_TRUNC;
 549
 550         if (pipe_output)
 551                 output = STDOUT_FILENO;
 552         else
 553                 output = open(output_name, flags, S_IRUSR | S_IWUSR);
 554         if (output < 0) {
 555                 perror("failed to create output file");
 556                 exit(-1);
 557         }
 558
 559         session = perf_session__new(output_name, O_WRONLY,
 560                                     write_mode == WRITE_FORCE, false);
 561         if (session == NULL) {
 562                 pr_err("Not enough memory for reading perf file header\n");
 563                 return -1;
 564         }
 565
 566         if (!file_new) {
 567                 err = perf_header__read(session, output);
 568                 if (err < 0)
 569                         return err;
 570         }
 571
 572         if (have_tracepoints(attrs, nr_counters))
 573                 perf_header__set_feat(&session->header, HEADER_TRACE_INFO);
 574
 575         atexit(atexit_header);
 576
 577         if (forks) {
 578                 child_pid = fork();
 579                 if (child_pid < 0) {
 580                         perror("failed to fork");
 581                         exit(-1);
 582                 }
 583
 584                 if (!child_pid) {
 585                         if (pipe_output)
 586                                 dup2(2, 1);
 587                         close(child_ready_pipe[0]);
 588                         close(go_pipe[1]);
 589                         fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
 590
 591                         /*
 592                          * Do a dummy execvp to get the PLT entry resolved,
 593                          * so we avoid the resolver overhead on the real
 594                          * execvp call.
 595                          */
 596                         execvp("", (char **)argv);
 597
 598                         /*
 599                          * Tell the parent we're ready to go
 600                          */
 601                         close(child_ready_pipe[1]);
 602
 603                         /*
 604                          * Wait until the parent tells us to go.
 605                          */
 606                         if (read(go_pipe[0], &buf, 1) == -1)
 607                                 perror("unable to read pipe");
 608
 609                         execvp(argv[0], (char **)argv);
 610
 611                         perror(argv[0]);
 612                         exit(-1);
 613                 }
 614
 615                 if (!system_wide && target_tid == -1 && target_pid == -1)
 616                         all_tids[0] = child_pid;
 617
 618                 close(child_ready_pipe[1]);
 619                 close(go_pipe[0]);
 620                 /*
 621                  * wait for child to settle
 622                  */
 623                 if (read(child_ready_pipe[0], &buf, 1) == -1) {
 624                         perror("unable to read pipe");
 625                         exit(-1);
 626                 }
 627                 close(child_ready_pipe[0]);
 628         }
 629
 630         nr_cpus = read_cpu_map(cpu_list);
 631         if (nr_cpus < 1) {
 632                 perror("failed to collect number of CPUs\n");
 633                 return -1;
 634         }
 635
 636         if (!system_wide && no_inherit && !cpu_list) {
 637                 open_counters(-1);
 638         } else {
 639                 for (i = 0; i < nr_cpus; i++)
 640                         open_counters(cpumap[i]);
 641         }
 642
 643         if (pipe_output) {
 644                 err = perf_header__write_pipe(output);
 645                 if (err < 0)
 646                         return err;
 647         } else if (file_new) {
 648                 err = perf_header__write(&session->header, output, false);
 649                 if (err < 0)
 650                         return err;
 651         }
 652
 653         post_processing_offset = lseek(output, 0, SEEK_CUR);
 654
 655         if (pipe_output) {
 656                 err = event__synthesize_attrs(&session->header,
 657                                               process_synthesized_event,
 658                                               session);
 659                 if (err < 0) {
 660                         pr_err("Couldn't synthesize attrs.\n");
 661                         return err;
 662                 }
 663
 664                 err = event__synthesize_event_types(process_synthesized_event,
 665                                                     session);
 666                 if (err < 0) {
 667                         pr_err("Couldn't synthesize event_types.\n");
 668                         return err;
 669                 }
 670
 671                 if (have_tracepoints(attrs, nr_counters)) {
 672                         /*
 673                          * FIXME err <= 0 here actually means that
 674                          * there were no tracepoints so its not really
 675                          * an error, just that we don't need to
 676                          * synthesize anything.  We really have to
 677                          * return this more properly and also
 678                          * propagate errors that now are calling die()
 679                          */
 680                         err = event__synthesize_tracing_data(output, attrs,
 681                                                              nr_counters,
 682                                                              process_synthesized_event,
 683                                                              session);
 684                         if (err <= 0) {
 685                                 pr_err("Couldn't record tracing data.\n");
 686                                 return err;
 687                         }
 688                         advance_output(err);
 689                 }
 690         }
 691
 692         machine = perf_session__find_host_machine(session);
 693         if (!machine) {
 694                 pr_err("Couldn't find native kernel information.\n");
 695                 return -1;
 696         }
 697
 698         err = event__synthesize_kernel_mmap(process_synthesized_event,
 699                                             session, machine, "_text");
 700         if (err < 0)
 701                 err = event__synthesize_kernel_mmap(process_synthesized_event,
 702                                                     session, machine, "_stext");
 703         if (err < 0) {
 704                 pr_err("Couldn't record kernel reference relocation symbol.\n");
 705                 return err;
 706         }
 707
 708         err = event__synthesize_modules(process_synthesized_event,
 709                                         session, machine);
 710         if (err < 0) {
 711                 pr_err("Couldn't record kernel reference relocation symbol.\n");
 712                 return err;
 713         }
 714         if (perf_guest)
 715                 perf_session__process_machines(session, event__synthesize_guest_os);
 716
 717         if (!system_wide && cpu_list)
 718                 event__synthesize_thread(target_tid, process_synthesized_event,
 719                                          session);
 720         else
 721                 event__synthesize_threads(process_synthesized_event, session);
 722
 723         if (realtime_prio) {
 724                 struct sched_param param;
 725
 726                 param.sched_priority = realtime_prio;
 727                 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
 728                         pr_err("Could not set realtime priority.\n");
 729                         exit(-1);
 730                 }
 731         }
 732
 733         /*
 734          * Let the child rip
 735          */
 736         if (forks)
 737                 close(go_pipe[1]);
 738
 739         for (;;) {
 740                 int hits = samples;
 741                 int thread;
 742
 743                 mmap_read_all();
 744
 745                 if (hits == samples) {
 746                         if (done)
 747                                 break;
 748                         err = poll(event_array, nr_poll, -1);
 749                         waking++;
 750                 }
 751
 752                 if (done) {
 753                         for (i = 0; i < nr_cpu; i++) {
 754                                 for (counter = 0;
 755                                         counter < nr_counters;
 756                                         counter++) {
 757                                         for (thread = 0;
 758                                                 thread < thread_num;
 759                                                 thread++)
 760                                                 ioctl(fd[i][counter][thread],
 761                                                         PERF_EVENT_IOC_DISABLE);
 762                                 }
 763                         }
 764                 }
 765         }
 766
 767         fprintf(stderr, "[ perf record: Woken up %ld times to write data ]\n", waking);
 768
 769         /*
 770          * Approximate RIP event size: 24 bytes.
 771          */
 772         fprintf(stderr,
 773                 "[ perf record: Captured and wrote %.3f MB %s (~%lld samples) ]\n",
 774                 (double)bytes_written / 1024.0 / 1024.0,
 775                 output_name,
 776                 bytes_written / 24);
 777
 778         return 0;
 779 }
 780
 781 static const char * const record_usage[] = {
 782         "perf record [<options>] [<command>]",
 783         "perf record [<options>] -- <command> [<options>]",
 784         NULL
 785 };
 786
 787 static bool force, append_file;
 788
 789 static const struct option options[] = {
 790         OPT_CALLBACK('e', "event", NULL, "event",
 791                      "event selector. use 'perf list' to list available events",
 792                      parse_events),
 793         OPT_CALLBACK(0, "filter", NULL, "filter",
 794                      "event filter", parse_filter),
 795         OPT_INTEGER('p', "pid", &target_pid,
 796                     "record events on existing process id"),
 797         OPT_INTEGER('t', "tid", &target_tid,
 798                     "record events on existing thread id"),
 799         OPT_INTEGER('r', "realtime", &realtime_prio,
 800                     "collect data with this RT SCHED_FIFO priority"),
 801         OPT_BOOLEAN('R', "raw-samples", &raw_samples,
 802                     "collect raw sample records from all opened counters"),
 803         OPT_BOOLEAN('a', "all-cpus", &system_wide,
 804                             "system-wide collection from all CPUs"),
 805         OPT_BOOLEAN('A', "append", &append_file,
 806                             "append to the output file to do incremental profiling"),
 807         OPT_STRING('C', "cpu", &cpu_list, "cpu",
 808                     "list of cpus to monitor"),
 809         OPT_BOOLEAN('f', "force", &force,
 810                         "overwrite existing data file (deprecated)"),
 811         OPT_U64('c', "count", &user_interval, "event period to sample"),
 812         OPT_STRING('o', "output", &output_name, "file",
 813                     "output file name"),
 814         OPT_BOOLEAN('i', "no-inherit", &no_inherit,
 815                     "child tasks do not inherit counters"),
 816         OPT_UINTEGER('F', "freq", &user_freq, "profile at this frequency"),
 817         OPT_UINTEGER('m', "mmap-pages", &mmap_pages, "number of mmap data pages"),
 818         OPT_BOOLEAN('g', "call-graph", &call_graph,
 819                     "do call-graph (stack chain/backtrace) recording"),
 820         OPT_INCR('v', "verbose", &verbose,
 821                     "be more verbose (show counter open errors, etc)"),
 822         OPT_BOOLEAN('s', "stat", &inherit_stat,
 823                     "per thread counts"),
 824         OPT_BOOLEAN('d', "data", &sample_address,
 825                     "Sample addresses"),
 826         OPT_BOOLEAN('n', "no-samples", &no_samples,
 827                     "don't sample"),
 828         OPT_END()
 829 };
 830
 831 int cmd_record(int argc, const char **argv, const char *prefix __used)
 832 {
 833         int i,j;
 834
 835         argc = parse_options(argc, argv, options, record_usage,
 836                             PARSE_OPT_STOP_AT_NON_OPTION);
 837         if (!argc && target_pid == -1 && target_tid == -1 &&
 838                 !system_wide && !cpu_list)
 839                 usage_with_options(record_usage, options);
 840
 841         if (force && append_file) {
 842                 fprintf(stderr, "Can't overwrite and append at the same time."
 843                                 " You need to choose between -f and -A");
 844                 usage_with_options(record_usage, options);
 845         } else if (append_file) {
 846                 write_mode = WRITE_APPEND;
 847         } else {
 848                 write_mode = WRITE_FORCE;
 849         }
 850
 851         symbol__init();
 852
 853         if (!nr_counters) {
 854                 nr_counters     = 1;
 855                 attrs[0].type   = PERF_TYPE_HARDWARE;
 856                 attrs[0].config = PERF_COUNT_HW_CPU_CYCLES;
 857         }
 858
 859         if (target_pid != -1) {
 860                 target_tid = target_pid;
 861                 thread_num = find_all_tid(target_pid, &all_tids);
 862                 if (thread_num <= 0) {
 863                         fprintf(stderr, "Can't find all threads of pid %d\n",
 864                                         target_pid);
 865                         usage_with_options(record_usage, options);
 866                 }
 867         } else {
 868                 all_tids=malloc(sizeof(pid_t));
 869                 if (!all_tids)
 870                         return -ENOMEM;
 871
 872                 all_tids[0] = target_tid;
 873                 thread_num = 1;
 874         }
 875
 876         for (i = 0; i < MAX_NR_CPUS; i++) {
 877                 for (j = 0; j < MAX_COUNTERS; j++) {
 878                         fd[i][j] = malloc(sizeof(int)*thread_num);
 879                         if (!fd[i][j])
 880                                 return -ENOMEM;
 881                 }
 882         }
 883         event_array = malloc(
 884                 sizeof(struct pollfd)*MAX_NR_CPUS*MAX_COUNTERS*thread_num);
 885         if (!event_array)
 886                 return -ENOMEM;
 887
 888         if (user_interval != ULLONG_MAX)
 889                 default_interval = user_interval;
 890         if (user_freq != UINT_MAX)
 891                 freq = user_freq;
 892
 893         /*
 894          * User specified count overrides default frequency.
 895          */
 896         if (default_interval)
 897                 freq = 0;
 898         else if (freq) {
 899                 default_interval = freq;
 900         } else {
 901                 fprintf(stderr, "frequency and count are zero, aborting\n");
 902                 exit(EXIT_FAILURE);
 903         }
 904
 905         return __cmd_record(argc, argv);
 906 }