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 #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 static bool                     multiplex                       =  false;
  65 static int                      multiplex_fd                    =     -1;
  66
  67 static long                     samples                         =      0;
  68 static u64                      bytes_written                   =      0;
  69
  70 static struct pollfd            *event_array;
  71
  72 static int                      nr_poll                         =      0;
  73 static int                      nr_cpu                          =      0;
  74
  75 static int                      file_new                        =      1;
  76 static off_t                    post_processing_offset;
  77
  78 static struct perf_session      *session;
  79
  80 struct mmap_data {
  81         int                     counter;
  82         void                    *base;
  83         unsigned int            mask;
  84         unsigned int            prev;
  85 };
  86
  87 static struct mmap_data         *mmap_array[MAX_NR_CPUS][MAX_COUNTERS];
  88
  89 static unsigned long mmap_read_head(struct mmap_data *md)
  90 {
  91         struct perf_event_mmap_page *pc = md->base;
  92         long head;
  93
  94         head = pc->data_head;
  95         rmb();
  96
  97         return head;
  98 }
  99
 100 static void mmap_write_tail(struct mmap_data *md, unsigned long tail)
 101 {
 102         struct perf_event_mmap_page *pc = md->base;
 103
 104         /*
 105          * ensure all reads are done before we write the tail out.
 106          */
 107         /* mb(); */
 108         pc->data_tail = tail;
 109 }
 110
 111 static void advance_output(size_t size)
 112 {
 113         bytes_written += size;
 114 }
 115
 116 static void write_output(void *buf, size_t size)
 117 {
 118         while (size) {
 119                 int ret = write(output, buf, size);
 120
 121                 if (ret < 0)
 122                         die("failed to write");
 123
 124                 size -= ret;
 125                 buf += ret;
 126
 127                 bytes_written += ret;
 128         }
 129 }
 130
 131 static int process_synthesized_event(event_t *event,
 132                                      struct perf_session *self __used)
 133 {
 134         write_output(event, event->header.size);
 135         return 0;
 136 }
 137
 138 static void mmap_read(struct mmap_data *md)
 139 {
 140         unsigned int head = mmap_read_head(md);
 141         unsigned int old = md->prev;
 142         unsigned char *data = md->base + page_size;
 143         unsigned long size;
 144         void *buf;
 145         int diff;
 146
 147         /*
 148          * If we're further behind than half the buffer, there's a chance
 149          * the writer will bite our tail and mess up the samples under us.
 150          *
 151          * If we somehow ended up ahead of the head, we got messed up.
 152          *
 153          * In either case, truncate and restart at head.
 154          */
 155         diff = head - old;
 156         if (diff < 0) {
 157                 fprintf(stderr, "WARNING: failed to keep up with mmap data\n");
 158                 /*
 159                  * head points to a known good entry, start there.
 160                  */
 161                 old = head;
 162         }
 163
 164         if (old != head)
 165                 samples++;
 166
 167         size = head - old;
 168
 169         if ((old & md->mask) + size != (head & md->mask)) {
 170                 buf = &data[old & md->mask];
 171                 size = md->mask + 1 - (old & md->mask);
 172                 old += size;
 173
 174                 write_output(buf, size);
 175         }
 176
 177         buf = &data[old & md->mask];
 178         size = head - old;
 179         old += size;
 180
 181         write_output(buf, size);
 182
 183         md->prev = old;
 184         mmap_write_tail(md, old);
 185 }
 186
 187 static volatile int done = 0;
 188 static volatile int signr = -1;
 189
 190 static void sig_handler(int sig)
 191 {
 192         done = 1;
 193         signr = sig;
 194 }
 195
 196 static void sig_atexit(void)
 197 {
 198         if (child_pid != -1)
 199                 kill(child_pid, SIGTERM);
 200
 201         if (signr == -1)
 202                 return;
 203
 204         signal(signr, SIG_DFL);
 205         kill(getpid(), signr);
 206 }
 207
 208 static int group_fd;
 209
 210 static struct perf_header_attr *get_header_attr(struct perf_event_attr *a, int nr)
 211 {
 212         struct perf_header_attr *h_attr;
 213
 214         if (nr < session->header.attrs) {
 215                 h_attr = session->header.attr[nr];
 216         } else {
 217                 h_attr = perf_header_attr__new(a);
 218                 if (h_attr != NULL)
 219                         if (perf_header__add_attr(&session->header, h_attr) < 0) {
 220                                 perf_header_attr__delete(h_attr);
 221                                 h_attr = NULL;
 222                         }
 223         }
 224
 225         return h_attr;
 226 }
 227
 228 static void create_counter(int counter, int cpu)
 229 {
 230         char *filter = filters[counter];
 231         struct perf_event_attr *attr = attrs + counter;
 232         struct perf_header_attr *h_attr;
 233         int track = !counter; /* only the first counter needs these */
 234         int thread_index;
 235         int ret;
 236         struct {
 237                 u64 count;
 238                 u64 time_enabled;
 239                 u64 time_running;
 240                 u64 id;
 241         } read_data;
 242
 243         attr->read_format       = PERF_FORMAT_TOTAL_TIME_ENABLED |
 244                                   PERF_FORMAT_TOTAL_TIME_RUNNING |
 245                                   PERF_FORMAT_ID;
 246
 247         attr->sample_type       |= PERF_SAMPLE_IP | PERF_SAMPLE_TID;
 248
 249         if (nr_counters > 1)
 250                 attr->sample_type |= PERF_SAMPLE_ID;
 251
 252         /*
 253          * We default some events to a 1 default interval. But keep
 254          * it a weak assumption overridable by the user.
 255          */
 256         if (!attr->sample_period || (user_freq != UINT_MAX &&
 257                                      user_interval != ULLONG_MAX)) {
 258                 if (freq) {
 259                         attr->sample_type       |= PERF_SAMPLE_PERIOD;
 260                         attr->freq              = 1;
 261                         attr->sample_freq       = freq;
 262                 } else {
 263                         attr->sample_period = default_interval;
 264                 }
 265         }
 266
 267         if (no_samples)
 268                 attr->sample_freq = 0;
 269
 270         if (inherit_stat)
 271                 attr->inherit_stat = 1;
 272
 273         if (sample_address)
 274                 attr->sample_type       |= PERF_SAMPLE_ADDR;
 275
 276         if (call_graph)
 277                 attr->sample_type       |= PERF_SAMPLE_CALLCHAIN;
 278
 279         if (raw_samples) {
 280                 attr->sample_type       |= PERF_SAMPLE_TIME;
 281                 attr->sample_type       |= PERF_SAMPLE_RAW;
 282                 attr->sample_type       |= PERF_SAMPLE_CPU;
 283         }
 284
 285         attr->mmap              = track;
 286         attr->comm              = track;
 287         attr->inherit           = !no_inherit;
 288         if (target_pid == -1 && target_tid == -1 && !system_wide) {
 289                 attr->disabled = 1;
 290                 attr->enable_on_exec = 1;
 291         }
 292
 293         for (thread_index = 0; thread_index < thread_num; thread_index++) {
 294 try_again:
 295                 fd[nr_cpu][counter][thread_index] = sys_perf_event_open(attr,
 296                                 all_tids[thread_index], cpu, group_fd, 0);
 297
 298                 if (fd[nr_cpu][counter][thread_index] < 0) {
 299                         int err = errno;
 300
 301                         if (err == EPERM || err == EACCES)
 302                                 die("Permission error - are you root?\n"
 303                                         "\t Consider tweaking"
 304                                         " /proc/sys/kernel/perf_event_paranoid.\n");
 305                         else if (err ==  ENODEV && profile_cpu != -1) {
 306                                 die("No such device - did you specify"
 307                                         " an out-of-range profile CPU?\n");
 308                         }
 309
 310                         /*
 311                          * If it's cycles then fall back to hrtimer
 312                          * based cpu-clock-tick sw counter, which
 313                          * is always available even if no PMU support:
 314                          */
 315                         if (attr->type == PERF_TYPE_HARDWARE
 316                                         && attr->config == PERF_COUNT_HW_CPU_CYCLES) {
 317
 318                                 if (verbose)
 319                                         warning(" ... trying to fall back to cpu-clock-ticks\n");
 320                                 attr->type = PERF_TYPE_SOFTWARE;
 321                                 attr->config = PERF_COUNT_SW_CPU_CLOCK;
 322                                 goto try_again;
 323                         }
 324                         printf("\n");
 325                         error("perfcounter syscall returned with %d (%s)\n",
 326                                         fd[nr_cpu][counter][thread_index], strerror(err));
 327
 328 #if defined(__i386__) || defined(__x86_64__)
 329                         if (attr->type == PERF_TYPE_HARDWARE && err == EOPNOTSUPP)
 330                                 die("No hardware sampling interrupt available."
 331                                     " No APIC? If so then you can boot the kernel"
 332                                     " with the \"lapic\" boot parameter to"
 333                                     " force-enable it.\n");
 334 #endif
 335
 336                         die("No CONFIG_PERF_EVENTS=y kernel support configured?\n");
 337                         exit(-1);
 338                 }
 339
 340                 h_attr = get_header_attr(attr, counter);
 341                 if (h_attr == NULL)
 342                         die("nomem\n");
 343
 344                 if (!file_new) {
 345                         if (memcmp(&h_attr->attr, attr, sizeof(*attr))) {
 346                                 fprintf(stderr, "incompatible append\n");
 347                                 exit(-1);
 348                         }
 349                 }
 350
 351                 if (read(fd[nr_cpu][counter][thread_index], &read_data, sizeof(read_data)) == -1) {
 352                         perror("Unable to read perf file descriptor\n");
 353                         exit(-1);
 354                 }
 355
 356                 if (perf_header_attr__add_id(h_attr, read_data.id) < 0) {
 357                         pr_warning("Not enough memory to add id\n");
 358                         exit(-1);
 359                 }
 360
 361                 assert(fd[nr_cpu][counter][thread_index] >= 0);
 362                 fcntl(fd[nr_cpu][counter][thread_index], F_SETFL, O_NONBLOCK);
 363
 364                 /*
 365                  * First counter acts as the group leader:
 366                  */
 367                 if (group && group_fd == -1)
 368                         group_fd = fd[nr_cpu][counter][thread_index];
 369                 if (multiplex && multiplex_fd == -1)
 370                         multiplex_fd = fd[nr_cpu][counter][thread_index];
 371
 372                 if (multiplex && fd[nr_cpu][counter][thread_index] != multiplex_fd) {
 373
 374                         ret = ioctl(fd[nr_cpu][counter][thread_index], PERF_EVENT_IOC_SET_OUTPUT, multiplex_fd);
 375                         assert(ret != -1);
 376                 } else {
 377                         event_array[nr_poll].fd = fd[nr_cpu][counter][thread_index];
 378                         event_array[nr_poll].events = POLLIN;
 379                         nr_poll++;
 380
 381                         mmap_array[nr_cpu][counter][thread_index].counter = counter;
 382                         mmap_array[nr_cpu][counter][thread_index].prev = 0;
 383                         mmap_array[nr_cpu][counter][thread_index].mask = mmap_pages*page_size - 1;
 384                         mmap_array[nr_cpu][counter][thread_index].base = mmap(NULL, (mmap_pages+1)*page_size,
 385                                 PROT_READ|PROT_WRITE, MAP_SHARED, fd[nr_cpu][counter][thread_index], 0);
 386                         if (mmap_array[nr_cpu][counter][thread_index].base == MAP_FAILED) {
 387                                 error("failed to mmap with %d (%s)\n", errno, strerror(errno));
 388                                 exit(-1);
 389                         }
 390                 }
 391
 392                 if (filter != NULL) {
 393                         ret = ioctl(fd[nr_cpu][counter][thread_index],
 394                                         PERF_EVENT_IOC_SET_FILTER, filter);
 395                         if (ret) {
 396                                 error("failed to set filter with %d (%s)\n", errno,
 397                                                 strerror(errno));
 398                                 exit(-1);
 399                         }
 400                 }
 401         }
 402 }
 403
 404 static void open_counters(int cpu)
 405 {
 406         int counter;
 407
 408         group_fd = -1;
 409         for (counter = 0; counter < nr_counters; counter++)
 410                 create_counter(counter, cpu);
 411
 412         nr_cpu++;
 413 }
 414
 415 static int process_buildids(void)
 416 {
 417         u64 size = lseek(output, 0, SEEK_CUR);
 418
 419         if (size == 0)
 420                 return 0;
 421
 422         session->fd = output;
 423         return __perf_session__process_events(session, post_processing_offset,
 424                                               size - post_processing_offset,
 425                                               size, &build_id__mark_dso_hit_ops);
 426 }
 427
 428 static void atexit_header(void)
 429 {
 430         if (!pipe_output) {
 431                 session->header.data_size += bytes_written;
 432
 433                 process_buildids();
 434                 perf_header__write(&session->header, output, true);
 435         }
 436 }
 437
 438 static void event__synthesize_guest_os(struct machine *machine, void *data)
 439 {
 440         int err;
 441         char *guest_kallsyms;
 442         char path[PATH_MAX];
 443         struct perf_session *psession = data;
 444
 445         if (machine__is_host(machine))
 446                 return;
 447
 448         /*
 449          *As for guest kernel when processing subcommand record&report,
 450          *we arrange module mmap prior to guest kernel mmap and trigger
 451          *a preload dso because default guest module symbols are loaded
 452          *from guest kallsyms instead of /lib/modules/XXX/XXX. This
 453          *method is used to avoid symbol missing when the first addr is
 454          *in module instead of in guest kernel.
 455          */
 456         err = event__synthesize_modules(process_synthesized_event,
 457                                         psession, machine);
 458         if (err < 0)
 459                 pr_err("Couldn't record guest kernel [%d]'s reference"
 460                        " relocation symbol.\n", machine->pid);
 461
 462         if (machine__is_default_guest(machine))
 463                 guest_kallsyms = (char *) symbol_conf.default_guest_kallsyms;
 464         else {
 465                 sprintf(path, "%s/proc/kallsyms", machine->root_dir);
 466                 guest_kallsyms = path;
 467         }
 468
 469         /*
 470          * We use _stext for guest kernel because guest kernel's /proc/kallsyms
 471          * have no _text sometimes.
 472          */
 473         err = event__synthesize_kernel_mmap(process_synthesized_event,
 474                                             psession, machine, "_text");
 475         if (err < 0)
 476                 err = event__synthesize_kernel_mmap(process_synthesized_event,
 477                                                     psession, machine, "_stext");
 478         if (err < 0)
 479                 pr_err("Couldn't record guest kernel [%d]'s reference"
 480                        " relocation symbol.\n", machine->pid);
 481 }
 482
 483 static struct perf_event_header finished_round_event = {
 484         .size = sizeof(struct perf_event_header),
 485         .type = PERF_RECORD_FINISHED_ROUND,
 486 };
 487
 488 static void mmap_read_all(void)
 489 {
 490         int i, counter, thread;
 491
 492         for (i = 0; i < nr_cpu; i++) {
 493                 for (counter = 0; counter < nr_counters; counter++) {
 494                         for (thread = 0; thread < thread_num; thread++) {
 495                                 if (mmap_array[i][counter][thread].base)
 496                                         mmap_read(&mmap_array[i][counter][thread]);
 497                         }
 498
 499                 }
 500         }
 501
 502         if (perf_header__has_feat(&session->header, HEADER_TRACE_INFO))
 503                 write_output(&finished_round_event, sizeof(finished_round_event));
 504 }
 505
 506 static int __cmd_record(int argc, const char **argv)
 507 {
 508         int i, counter;
 509         struct stat st;
 510         pid_t pid = 0;
 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 (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         if ((!system_wide && no_inherit) || profile_cpu != -1) {
 631                 open_counters(profile_cpu);
 632         } else {
 633                 nr_cpus = read_cpu_map();
 634                 for (i = 0; i < nr_cpus; i++)
 635                         open_counters(cpumap[i]);
 636         }
 637
 638         if (pipe_output) {
 639                 err = perf_header__write_pipe(output);
 640                 if (err < 0)
 641                         return err;
 642         } else if (file_new) {
 643                 err = perf_header__write(&session->header, output, false);
 644                 if (err < 0)
 645                         return err;
 646         }
 647
 648         post_processing_offset = lseek(output, 0, SEEK_CUR);
 649
 650         if (pipe_output) {
 651                 err = event__synthesize_attrs(&session->header,
 652                                               process_synthesized_event,
 653                                               session);
 654                 if (err < 0) {
 655                         pr_err("Couldn't synthesize attrs.\n");
 656                         return err;
 657                 }
 658
 659                 err = event__synthesize_event_types(process_synthesized_event,
 660                                                     session);
 661                 if (err < 0) {
 662                         pr_err("Couldn't synthesize event_types.\n");
 663                         return err;
 664                 }
 665
 666                 if (have_tracepoints(attrs, nr_counters)) {
 667                         /*
 668                          * FIXME err <= 0 here actually means that
 669                          * there were no tracepoints so its not really
 670                          * an error, just that we don't need to
 671                          * synthesize anything.  We really have to
 672                          * return this more properly and also
 673                          * propagate errors that now are calling die()
 674                          */
 675                         err = event__synthesize_tracing_data(output, attrs,
 676                                                              nr_counters,
 677                                                              process_synthesized_event,
 678                                                              session);
 679                         if (err <= 0) {
 680                                 pr_err("Couldn't record tracing data.\n");
 681                                 return err;
 682                         }
 683                         advance_output(err);
 684                 }
 685         }
 686
 687         machine = perf_session__find_host_machine(session);
 688         if (!machine) {
 689                 pr_err("Couldn't find native kernel information.\n");
 690                 return -1;
 691         }
 692
 693         err = event__synthesize_kernel_mmap(process_synthesized_event,
 694                                             session, machine, "_text");
 695         if (err < 0)
 696                 err = event__synthesize_kernel_mmap(process_synthesized_event,
 697                                                     session, machine, "_stext");
 698         if (err < 0) {
 699                 pr_err("Couldn't record kernel reference relocation symbol.\n");
 700                 return err;
 701         }
 702
 703         err = event__synthesize_modules(process_synthesized_event,
 704                                         session, machine);
 705         if (err < 0) {
 706                 pr_err("Couldn't record kernel reference relocation symbol.\n");
 707                 return err;
 708         }
 709         if (perf_guest)
 710                 perf_session__process_machines(session, event__synthesize_guest_os);
 711
 712         if (!system_wide && profile_cpu == -1)
 713                 event__synthesize_thread(target_tid, process_synthesized_event,
 714                                          session);
 715         else
 716                 event__synthesize_threads(process_synthesized_event, session);
 717
 718         if (realtime_prio) {
 719                 struct sched_param param;
 720
 721                 param.sched_priority = realtime_prio;
 722                 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
 723                         pr_err("Could not set realtime priority.\n");
 724                         exit(-1);
 725                 }
 726         }
 727
 728         /*
 729          * Let the child rip
 730          */
 731         if (forks)
 732                 close(go_pipe[1]);
 733
 734         for (;;) {
 735                 int hits = samples;
 736                 int thread;
 737
 738                 mmap_read_all();
 739
 740                 if (hits == samples) {
 741                         if (done)
 742                                 break;
 743                         err = poll(event_array, nr_poll, -1);
 744                         waking++;
 745                 }
 746
 747                 if (done) {
 748                         for (i = 0; i < nr_cpu; i++) {
 749                                 for (counter = 0;
 750                                         counter < nr_counters;
 751                                         counter++) {
 752                                         for (thread = 0;
 753                                                 thread < thread_num;
 754                                                 thread++)
 755                                                 ioctl(fd[i][counter][thread],
 756                                                         PERF_EVENT_IOC_DISABLE);
 757                                 }
 758                         }
 759                 }
 760         }
 761
 762         fprintf(stderr, "[ perf record: Woken up %ld times to write data ]\n", waking);
 763
 764         /*
 765          * Approximate RIP event size: 24 bytes.
 766          */
 767         fprintf(stderr,
 768                 "[ perf record: Captured and wrote %.3f MB %s (~%lld samples) ]\n",
 769                 (double)bytes_written / 1024.0 / 1024.0,
 770                 output_name,
 771                 bytes_written / 24);
 772
 773         return 0;
 774 }
 775
 776 static const char * const record_usage[] = {
 777         "perf record [<options>] [<command>]",
 778         "perf record [<options>] -- <command> [<options>]",
 779         NULL
 780 };
 781
 782 static bool force, append_file;
 783
 784 static const struct option options[] = {
 785         OPT_CALLBACK('e', "event", NULL, "event",
 786                      "event selector. use 'perf list' to list available events",
 787                      parse_events),
 788         OPT_CALLBACK(0, "filter", NULL, "filter",
 789                      "event filter", parse_filter),
 790         OPT_INTEGER('p', "pid", &target_pid,
 791                     "record events on existing process id"),
 792         OPT_INTEGER('t', "tid", &target_tid,
 793                     "record events on existing thread id"),
 794         OPT_INTEGER('r', "realtime", &realtime_prio,
 795                     "collect data with this RT SCHED_FIFO priority"),
 796         OPT_BOOLEAN('R', "raw-samples", &raw_samples,
 797                     "collect raw sample records from all opened counters"),
 798         OPT_BOOLEAN('a', "all-cpus", &system_wide,
 799                             "system-wide collection from all CPUs"),
 800         OPT_BOOLEAN('A', "append", &append_file,
 801                             "append to the output file to do incremental profiling"),
 802         OPT_INTEGER('C', "profile_cpu", &profile_cpu,
 803                             "CPU to profile on"),
 804         OPT_BOOLEAN('f', "force", &force,
 805                         "overwrite existing data file (deprecated)"),
 806         OPT_U64('c', "count", &user_interval, "event period to sample"),
 807         OPT_STRING('o', "output", &output_name, "file",
 808                     "output file name"),
 809         OPT_BOOLEAN('i', "no-inherit", &no_inherit,
 810                     "child tasks do not inherit counters"),
 811         OPT_UINTEGER('F', "freq", &user_freq, "profile at this frequency"),
 812         OPT_UINTEGER('m', "mmap-pages", &mmap_pages, "number of mmap data pages"),
 813         OPT_BOOLEAN('g', "call-graph", &call_graph,
 814                     "do call-graph (stack chain/backtrace) recording"),
 815         OPT_INCR('v', "verbose", &verbose,
 816                     "be more verbose (show counter open errors, etc)"),
 817         OPT_BOOLEAN('s', "stat", &inherit_stat,
 818                     "per thread counts"),
 819         OPT_BOOLEAN('d', "data", &sample_address,
 820                     "Sample addresses"),
 821         OPT_BOOLEAN('n', "no-samples", &no_samples,
 822                     "don't sample"),
 823         OPT_BOOLEAN('M', "multiplex", &multiplex,
 824                     "multiplex counter output in a single channel"),
 825         OPT_END()
 826 };
 827
 828 int cmd_record(int argc, const char **argv, const char *prefix __used)
 829 {
 830         int i,j;
 831
 832         argc = parse_options(argc, argv, options, record_usage,
 833                             PARSE_OPT_STOP_AT_NON_OPTION);
 834         if (!argc && target_pid == -1 && target_tid == -1 &&
 835                 !system_wide && profile_cpu == -1)
 836                 usage_with_options(record_usage, options);
 837
 838         if (force && append_file) {
 839                 fprintf(stderr, "Can't overwrite and append at the same time."
 840                                 " You need to choose between -f and -A");
 841                 usage_with_options(record_usage, options);
 842         } else if (append_file) {
 843                 write_mode = WRITE_APPEND;
 844         } else {
 845                 write_mode = WRITE_FORCE;
 846         }
 847
 848         symbol__init();
 849
 850         if (!nr_counters) {
 851                 nr_counters     = 1;
 852                 attrs[0].type   = PERF_TYPE_HARDWARE;
 853                 attrs[0].config = PERF_COUNT_HW_CPU_CYCLES;
 854         }
 855
 856         if (target_pid != -1) {
 857                 target_tid = target_pid;
 858                 thread_num = find_all_tid(target_pid, &all_tids);
 859                 if (thread_num <= 0) {
 860                         fprintf(stderr, "Can't find all threads of pid %d\n",
 861                                         target_pid);
 862                         usage_with_options(record_usage, options);
 863                 }
 864         } else {
 865                 all_tids=malloc(sizeof(pid_t));
 866                 if (!all_tids)
 867                         return -ENOMEM;
 868
 869                 all_tids[0] = target_tid;
 870                 thread_num = 1;
 871         }
 872
 873         for (i = 0; i < MAX_NR_CPUS; i++) {
 874                 for (j = 0; j < MAX_COUNTERS; j++) {
 875                         fd[i][j] = malloc(sizeof(int)*thread_num);
 876                         mmap_array[i][j] = zalloc(
 877                                 sizeof(struct mmap_data)*thread_num);
 878                         if (!fd[i][j] || !mmap_array[i][j])
 879                                 return -ENOMEM;
 880                 }
 881         }
 882         event_array = malloc(
 883                 sizeof(struct pollfd)*MAX_NR_CPUS*MAX_COUNTERS*thread_num);
 884         if (!event_array)
 885                 return -ENOMEM;
 886
 887         if (user_interval != ULLONG_MAX)
 888                 default_interval = user_interval;
 889         if (user_freq != UINT_MAX)
 890                 freq = user_freq;
 891
 892         /*
 893          * User specified count overrides default frequency.
 894          */
 895         if (default_interval)
 896                 freq = 0;
 897         else if (freq) {
 898                 default_interval = freq;
 899         } else {
 900                 fprintf(stderr, "frequency and count are zero, aborting\n");
 901                 exit(EXIT_FAILURE);
 902         }
 903
 904         return __cmd_record(argc, argv);
 905 }