Pileus Git - ~andy/linux/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 static u64                      sample_type;
  40
  41 static int                      nr_cpus                         =      0;
  42 static unsigned int             page_size;
  43 static unsigned int             mmap_pages                      =    128;
  44 static unsigned int             user_freq                       = UINT_MAX;
  45 static int                      freq                            =   1000;
  46 static int                      output;
  47 static int                      pipe_output                     =      0;
  48 static const char               *output_name                    = "perf.data";
  49 static int                      group                           =      0;
  50 static int                      realtime_prio                   =      0;
  51 static bool                     raw_samples                     =  false;
  52 static bool                     system_wide                     =  false;
  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                     no_buildid                      =  false;
  65 static bool                     no_buildid_cache                =  false;
  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 static const char               *cpu_list;
  80
  81 struct mmap_data {
  82         int                     counter;
  83         void                    *base;
  84         unsigned int            mask;
  85         unsigned int            prev;
  86 };
  87
  88 static struct mmap_data         mmap_array[MAX_NR_CPUS];
  89
  90 static unsigned long mmap_read_head(struct mmap_data *md)
  91 {
  92         struct perf_event_mmap_page *pc = md->base;
  93         long head;
  94
  95         head = pc->data_head;
  96         rmb();
  97
  98         return head;
  99 }
 100
 101 static void mmap_write_tail(struct mmap_data *md, unsigned long tail)
 102 {
 103         struct perf_event_mmap_page *pc = md->base;
 104
 105         /*
 106          * ensure all reads are done before we write the tail out.
 107          */
 108         /* mb(); */
 109         pc->data_tail = tail;
 110 }
 111
 112 static void advance_output(size_t size)
 113 {
 114         bytes_written += size;
 115 }
 116
 117 static void write_output(void *buf, size_t size)
 118 {
 119         while (size) {
 120                 int ret = write(output, buf, size);
 121
 122                 if (ret < 0)
 123                         die("failed to write");
 124
 125                 size -= ret;
 126                 buf += ret;
 127
 128                 bytes_written += ret;
 129         }
 130 }
 131
 132 static int process_synthesized_event(event_t *event,
 133                                      struct sample_data *sample __used,
 134                                      struct perf_session *self __used)
 135 {
 136         write_output(event, event->header.size);
 137         return 0;
 138 }
 139
 140 static void mmap_read(struct mmap_data *md)
 141 {
 142         unsigned int head = mmap_read_head(md);
 143         unsigned int old = md->prev;
 144         unsigned char *data = md->base + page_size;
 145         unsigned long size;
 146         void *buf;
 147         int diff;
 148
 149         /*
 150          * If we're further behind than half the buffer, there's a chance
 151          * the writer will bite our tail and mess up the samples under us.
 152          *
 153          * If we somehow ended up ahead of the head, we got messed up.
 154          *
 155          * In either case, truncate and restart at head.
 156          */
 157         diff = head - old;
 158         if (diff < 0) {
 159                 fprintf(stderr, "WARNING: failed to keep up with mmap data\n");
 160                 /*
 161                  * head points to a known good entry, start there.
 162                  */
 163                 old = head;
 164         }
 165
 166         if (old != head)
 167                 samples++;
 168
 169         size = head - old;
 170
 171         if ((old & md->mask) + size != (head & md->mask)) {
 172                 buf = &data[old & md->mask];
 173                 size = md->mask + 1 - (old & md->mask);
 174                 old += size;
 175
 176                 write_output(buf, size);
 177         }
 178
 179         buf = &data[old & md->mask];
 180         size = head - old;
 181         old += size;
 182
 183         write_output(buf, size);
 184
 185         md->prev = old;
 186         mmap_write_tail(md, old);
 187 }
 188
 189 static volatile int done = 0;
 190 static volatile int signr = -1;
 191
 192 static void sig_handler(int sig)
 193 {
 194         done = 1;
 195         signr = sig;
 196 }
 197
 198 static void sig_atexit(void)
 199 {
 200         if (child_pid > 0)
 201                 kill(child_pid, SIGTERM);
 202
 203         if (signr == -1)
 204                 return;
 205
 206         signal(signr, SIG_DFL);
 207         kill(getpid(), signr);
 208 }
 209
 210 static int group_fd;
 211
 212 static struct perf_header_attr *get_header_attr(struct perf_event_attr *a, int nr)
 213 {
 214         struct perf_header_attr *h_attr;
 215
 216         if (nr < session->header.attrs) {
 217                 h_attr = session->header.attr[nr];
 218         } else {
 219                 h_attr = perf_header_attr__new(a);
 220                 if (h_attr != NULL)
 221                         if (perf_header__add_attr(&session->header, h_attr) < 0) {
 222                                 perf_header_attr__delete(h_attr);
 223                                 h_attr = NULL;
 224                         }
 225         }
 226
 227         return h_attr;
 228 }
 229
 230 static void create_counter(int counter, int cpu)
 231 {
 232         char *filter = filters[counter];
 233         struct perf_event_attr *attr = attrs + counter;
 234         struct perf_header_attr *h_attr;
 235         int track = !counter; /* only the first counter needs these */
 236         int thread_index;
 237         int ret;
 238         struct {
 239                 u64 count;
 240                 u64 time_enabled;
 241                 u64 time_running;
 242                 u64 id;
 243         } read_data;
 244
 245         attr->read_format       = PERF_FORMAT_TOTAL_TIME_ENABLED |
 246                                   PERF_FORMAT_TOTAL_TIME_RUNNING |
 247                                   PERF_FORMAT_ID;
 248
 249         attr->sample_type       |= PERF_SAMPLE_IP | PERF_SAMPLE_TID;
 250
 251         if (nr_counters > 1)
 252                 attr->sample_type |= PERF_SAMPLE_ID;
 253
 254         /*
 255          * We default some events to a 1 default interval. But keep
 256          * it a weak assumption overridable by the user.
 257          */
 258         if (!attr->sample_period || (user_freq != UINT_MAX &&
 259                                      user_interval != ULLONG_MAX)) {
 260                 if (freq) {
 261                         attr->sample_type       |= PERF_SAMPLE_PERIOD;
 262                         attr->freq              = 1;
 263                         attr->sample_freq       = freq;
 264                 } else {
 265                         attr->sample_period = default_interval;
 266                 }
 267         }
 268
 269         if (no_samples)
 270                 attr->sample_freq = 0;
 271
 272         if (inherit_stat)
 273                 attr->inherit_stat = 1;
 274
 275         if (sample_address) {
 276                 attr->sample_type       |= PERF_SAMPLE_ADDR;
 277                 attr->mmap_data = track;
 278         }
 279
 280         if (call_graph)
 281                 attr->sample_type       |= PERF_SAMPLE_CALLCHAIN;
 282
 283         if (system_wide)
 284                 attr->sample_type       |= PERF_SAMPLE_CPU;
 285
 286         if (raw_samples) {
 287                 attr->sample_type       |= PERF_SAMPLE_TIME;
 288                 attr->sample_type       |= PERF_SAMPLE_RAW;
 289                 attr->sample_type       |= PERF_SAMPLE_CPU;
 290         }
 291
 292         if (!sample_type)
 293                 sample_type = attr->sample_type;
 294
 295         attr->mmap              = track;
 296         attr->comm              = track;
 297         attr->inherit           = !no_inherit;
 298         if (target_pid == -1 && target_tid == -1 && !system_wide) {
 299                 attr->disabled = 1;
 300                 attr->enable_on_exec = 1;
 301         }
 302
 303         for (thread_index = 0; thread_index < thread_num; thread_index++) {
 304 try_again:
 305                 fd[nr_cpu][counter][thread_index] = sys_perf_event_open(attr,
 306                                 all_tids[thread_index], cpu, group_fd, 0);
 307
 308                 if (fd[nr_cpu][counter][thread_index] < 0) {
 309                         int err = errno;
 310
 311                         if (err == EPERM || err == EACCES)
 312                                 die("Permission error - are you root?\n"
 313                                         "\t Consider tweaking"
 314                                         " /proc/sys/kernel/perf_event_paranoid.\n");
 315                         else if (err ==  ENODEV && cpu_list) {
 316                                 die("No such device - did you specify"
 317                                         " an out-of-range profile CPU?\n");
 318                         }
 319
 320                         /*
 321                          * If it's cycles then fall back to hrtimer
 322                          * based cpu-clock-tick sw counter, which
 323                          * is always available even if no PMU support:
 324                          */
 325                         if (attr->type == PERF_TYPE_HARDWARE
 326                                         && attr->config == PERF_COUNT_HW_CPU_CYCLES) {
 327
 328                                 if (verbose)
 329                                         warning(" ... trying to fall back to cpu-clock-ticks\n");
 330                                 attr->type = PERF_TYPE_SOFTWARE;
 331                                 attr->config = PERF_COUNT_SW_CPU_CLOCK;
 332                                 goto try_again;
 333                         }
 334                         printf("\n");
 335                         error("sys_perf_event_open() syscall returned with %d (%s).  /bin/dmesg may provide additional information.\n",
 336                                         fd[nr_cpu][counter][thread_index], strerror(err));
 337
 338 #if defined(__i386__) || defined(__x86_64__)
 339                         if (attr->type == PERF_TYPE_HARDWARE && err == EOPNOTSUPP)
 340                                 die("No hardware sampling interrupt available."
 341                                     " No APIC? If so then you can boot the kernel"
 342                                     " with the \"lapic\" boot parameter to"
 343                                     " force-enable it.\n");
 344 #endif
 345
 346                         die("No CONFIG_PERF_EVENTS=y kernel support configured?\n");
 347                         exit(-1);
 348                 }
 349
 350                 h_attr = get_header_attr(attr, counter);
 351                 if (h_attr == NULL)
 352                         die("nomem\n");
 353
 354                 if (!file_new) {
 355                         if (memcmp(&h_attr->attr, attr, sizeof(*attr))) {
 356                                 fprintf(stderr, "incompatible append\n");
 357                                 exit(-1);
 358                         }
 359                 }
 360
 361                 if (read(fd[nr_cpu][counter][thread_index], &read_data, sizeof(read_data)) == -1) {
 362                         perror("Unable to read perf file descriptor");
 363                         exit(-1);
 364                 }
 365
 366                 if (perf_header_attr__add_id(h_attr, read_data.id) < 0) {
 367                         pr_warning("Not enough memory to add id\n");
 368                         exit(-1);
 369                 }
 370
 371                 assert(fd[nr_cpu][counter][thread_index] >= 0);
 372                 fcntl(fd[nr_cpu][counter][thread_index], F_SETFL, O_NONBLOCK);
 373
 374                 /*
 375                  * First counter acts as the group leader:
 376                  */
 377                 if (group && group_fd == -1)
 378                         group_fd = fd[nr_cpu][counter][thread_index];
 379
 380                 if (counter || thread_index) {
 381                         ret = ioctl(fd[nr_cpu][counter][thread_index],
 382                                         PERF_EVENT_IOC_SET_OUTPUT,
 383                                         fd[nr_cpu][0][0]);
 384                         if (ret) {
 385                                 error("failed to set output: %d (%s)\n", errno,
 386                                                 strerror(errno));
 387                                 exit(-1);
 388                         }
 389                 } else {
 390                         mmap_array[nr_cpu].counter = counter;
 391                         mmap_array[nr_cpu].prev = 0;
 392                         mmap_array[nr_cpu].mask = mmap_pages*page_size - 1;
 393                         mmap_array[nr_cpu].base = mmap(NULL, (mmap_pages+1)*page_size,
 394                                 PROT_READ|PROT_WRITE, MAP_SHARED, fd[nr_cpu][counter][thread_index], 0);
 395                         if (mmap_array[nr_cpu].base == MAP_FAILED) {
 396                                 error("failed to mmap with %d (%s)\n", errno, strerror(errno));
 397                                 exit(-1);
 398                         }
 399
 400                         event_array[nr_poll].fd = fd[nr_cpu][counter][thread_index];
 401                         event_array[nr_poll].events = POLLIN;
 402                         nr_poll++;
 403                 }
 404
 405                 if (filter != NULL) {
 406                         ret = ioctl(fd[nr_cpu][counter][thread_index],
 407                                         PERF_EVENT_IOC_SET_FILTER, filter);
 408                         if (ret) {
 409                                 error("failed to set filter with %d (%s)\n", errno,
 410                                                 strerror(errno));
 411                                 exit(-1);
 412                         }
 413                 }
 414         }
 415 }
 416
 417 static void open_counters(int cpu)
 418 {
 419         int counter;
 420
 421         group_fd = -1;
 422         for (counter = 0; counter < nr_counters; counter++)
 423                 create_counter(counter, cpu);
 424
 425         nr_cpu++;
 426 }
 427
 428 static int process_buildids(void)
 429 {
 430         u64 size = lseek(output, 0, SEEK_CUR);
 431
 432         if (size == 0)
 433                 return 0;
 434
 435         session->fd = output;
 436         return __perf_session__process_events(session, post_processing_offset,
 437                                               size - post_processing_offset,
 438                                               size, &build_id__mark_dso_hit_ops);
 439 }
 440
 441 static void atexit_header(void)
 442 {
 443         if (!pipe_output) {
 444                 session->header.data_size += bytes_written;
 445
 446                 if (!no_buildid)
 447                         process_buildids();
 448                 perf_header__write(&session->header, output, true);
 449                 perf_session__delete(session);
 450                 symbol__exit();
 451         }
 452 }
 453
 454 static void event__synthesize_guest_os(struct machine *machine, void *data)
 455 {
 456         int err;
 457         struct perf_session *psession = data;
 458
 459         if (machine__is_host(machine))
 460                 return;
 461
 462         /*
 463          *As for guest kernel when processing subcommand record&report,
 464          *we arrange module mmap prior to guest kernel mmap and trigger
 465          *a preload dso because default guest module symbols are loaded
 466          *from guest kallsyms instead of /lib/modules/XXX/XXX. This
 467          *method is used to avoid symbol missing when the first addr is
 468          *in module instead of in guest kernel.
 469          */
 470         err = event__synthesize_modules(process_synthesized_event,
 471                                         psession, machine);
 472         if (err < 0)
 473                 pr_err("Couldn't record guest kernel [%d]'s reference"
 474                        " relocation symbol.\n", machine->pid);
 475
 476         /*
 477          * We use _stext for guest kernel because guest kernel's /proc/kallsyms
 478          * have no _text sometimes.
 479          */
 480         err = event__synthesize_kernel_mmap(process_synthesized_event,
 481                                             psession, machine, "_text");
 482         if (err < 0)
 483                 err = event__synthesize_kernel_mmap(process_synthesized_event,
 484                                                     psession, machine, "_stext");
 485         if (err < 0)
 486                 pr_err("Couldn't record guest kernel [%d]'s reference"
 487                        " relocation symbol.\n", machine->pid);
 488 }
 489
 490 static struct perf_event_header finished_round_event = {
 491         .size = sizeof(struct perf_event_header),
 492         .type = PERF_RECORD_FINISHED_ROUND,
 493 };
 494
 495 static void mmap_read_all(void)
 496 {
 497         int i;
 498
 499         for (i = 0; i < nr_cpu; i++) {
 500                 if (mmap_array[i].base)
 501                         mmap_read(&mmap_array[i]);
 502         }
 503
 504         if (perf_header__has_feat(&session->header, HEADER_TRACE_INFO))
 505                 write_output(&finished_round_event, sizeof(finished_round_event));
 506 }
 507
 508 static int __cmd_record(int argc, const char **argv)
 509 {
 510         int i, counter;
 511         struct stat st;
 512         int flags;
 513         int err;
 514         unsigned long waking = 0;
 515         int child_ready_pipe[2], go_pipe[2];
 516         const bool forks = argc > 0;
 517         char buf;
 518         struct machine *machine;
 519
 520         page_size = sysconf(_SC_PAGE_SIZE);
 521
 522         atexit(sig_atexit);
 523         signal(SIGCHLD, sig_handler);
 524         signal(SIGINT, sig_handler);
 525
 526         if (forks && (pipe(child_ready_pipe) < 0 || pipe(go_pipe) < 0)) {
 527                 perror("failed to create pipes");
 528                 exit(-1);
 529         }
 530
 531         if (!strcmp(output_name, "-"))
 532                 pipe_output = 1;
 533         else if (!stat(output_name, &st) && st.st_size) {
 534                 if (write_mode == WRITE_FORCE) {
 535                         char oldname[PATH_MAX];
 536                         snprintf(oldname, sizeof(oldname), "%s.old",
 537                                  output_name);
 538                         unlink(oldname);
 539                         rename(output_name, oldname);
 540                 }
 541         } else if (write_mode == WRITE_APPEND) {
 542                 write_mode = WRITE_FORCE;
 543         }
 544
 545         flags = O_CREAT|O_RDWR;
 546         if (write_mode == WRITE_APPEND)
 547                 file_new = 0;
 548         else
 549                 flags |= O_TRUNC;
 550
 551         if (pipe_output)
 552                 output = STDOUT_FILENO;
 553         else
 554                 output = open(output_name, flags, S_IRUSR | S_IWUSR);
 555         if (output < 0) {
 556                 perror("failed to create output file");
 557                 exit(-1);
 558         }
 559
 560         session = perf_session__new(output_name, O_WRONLY,
 561                                     write_mode == WRITE_FORCE, false);
 562         if (session == NULL) {
 563                 pr_err("Not enough memory for reading perf file header\n");
 564                 return -1;
 565         }
 566
 567         if (!no_buildid)
 568                 perf_header__set_feat(&session->header, HEADER_BUILD_ID);
 569
 570         if (!file_new) {
 571                 err = perf_header__read(session, output);
 572                 if (err < 0)
 573                         goto out_delete_session;
 574         }
 575
 576         if (have_tracepoints(attrs, nr_counters))
 577                 perf_header__set_feat(&session->header, HEADER_TRACE_INFO);
 578
 579         /*
 580          * perf_session__delete(session) will be called at atexit_header()
 581          */
 582         atexit(atexit_header);
 583
 584         if (forks) {
 585                 child_pid = fork();
 586                 if (child_pid < 0) {
 587                         perror("failed to fork");
 588                         exit(-1);
 589                 }
 590
 591                 if (!child_pid) {
 592                         if (pipe_output)
 593                                 dup2(2, 1);
 594                         close(child_ready_pipe[0]);
 595                         close(go_pipe[1]);
 596                         fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
 597
 598                         /*
 599                          * Do a dummy execvp to get the PLT entry resolved,
 600                          * so we avoid the resolver overhead on the real
 601                          * execvp call.
 602                          */
 603                         execvp("", (char **)argv);
 604
 605                         /*
 606                          * Tell the parent we're ready to go
 607                          */
 608                         close(child_ready_pipe[1]);
 609
 610                         /*
 611                          * Wait until the parent tells us to go.
 612                          */
 613                         if (read(go_pipe[0], &buf, 1) == -1)
 614                                 perror("unable to read pipe");
 615
 616                         execvp(argv[0], (char **)argv);
 617
 618                         perror(argv[0]);
 619                         exit(-1);
 620                 }
 621
 622                 if (!system_wide && target_tid == -1 && target_pid == -1)
 623                         all_tids[0] = child_pid;
 624
 625                 close(child_ready_pipe[1]);
 626                 close(go_pipe[0]);
 627                 /*
 628                  * wait for child to settle
 629                  */
 630                 if (read(child_ready_pipe[0], &buf, 1) == -1) {
 631                         perror("unable to read pipe");
 632                         exit(-1);
 633                 }
 634                 close(child_ready_pipe[0]);
 635         }
 636
 637         nr_cpus = read_cpu_map(cpu_list);
 638         if (nr_cpus < 1) {
 639                 perror("failed to collect number of CPUs");
 640                 return -1;
 641         }
 642
 643         if (!system_wide && no_inherit && !cpu_list) {
 644                 open_counters(-1);
 645         } else {
 646                 for (i = 0; i < nr_cpus; i++)
 647                         open_counters(cpumap[i]);
 648         }
 649
 650         perf_session__set_sample_type(session, sample_type);
 651
 652         if (pipe_output) {
 653                 err = perf_header__write_pipe(output);
 654                 if (err < 0)
 655                         return err;
 656         } else if (file_new) {
 657                 err = perf_header__write(&session->header, output, false);
 658                 if (err < 0)
 659                         return err;
 660         }
 661
 662         post_processing_offset = lseek(output, 0, SEEK_CUR);
 663
 664         if (pipe_output) {
 665                 err = event__synthesize_attrs(&session->header,
 666                                               process_synthesized_event,
 667                                               session);
 668                 if (err < 0) {
 669                         pr_err("Couldn't synthesize attrs.\n");
 670                         return err;
 671                 }
 672
 673                 err = event__synthesize_event_types(process_synthesized_event,
 674                                                     session);
 675                 if (err < 0) {
 676                         pr_err("Couldn't synthesize event_types.\n");
 677                         return err;
 678                 }
 679
 680                 if (have_tracepoints(attrs, nr_counters)) {
 681                         /*
 682                          * FIXME err <= 0 here actually means that
 683                          * there were no tracepoints so its not really
 684                          * an error, just that we don't need to
 685                          * synthesize anything.  We really have to
 686                          * return this more properly and also
 687                          * propagate errors that now are calling die()
 688                          */
 689                         err = event__synthesize_tracing_data(output, attrs,
 690                                                              nr_counters,
 691                                                              process_synthesized_event,
 692                                                              session);
 693                         if (err <= 0) {
 694                                 pr_err("Couldn't record tracing data.\n");
 695                                 return err;
 696                         }
 697                         advance_output(err);
 698                 }
 699         }
 700
 701         machine = perf_session__find_host_machine(session);
 702         if (!machine) {
 703                 pr_err("Couldn't find native kernel information.\n");
 704                 return -1;
 705         }
 706
 707         err = event__synthesize_kernel_mmap(process_synthesized_event,
 708                                             session, machine, "_text");
 709         if (err < 0)
 710                 err = event__synthesize_kernel_mmap(process_synthesized_event,
 711                                                     session, machine, "_stext");
 712         if (err < 0)
 713                 pr_err("Couldn't record kernel reference relocation symbol\n"
 714                        "Symbol resolution may be skewed if relocation was used (e.g. kexec).\n"
 715                        "Check /proc/kallsyms permission or run as root.\n");
 716
 717         err = event__synthesize_modules(process_synthesized_event,
 718                                         session, machine);
 719         if (err < 0)
 720                 pr_err("Couldn't record kernel module information.\n"
 721                        "Symbol resolution may be skewed if relocation was used (e.g. kexec).\n"
 722                        "Check /proc/modules permission or run as root.\n");
 723
 724         if (perf_guest)
 725                 perf_session__process_machines(session, event__synthesize_guest_os);
 726
 727         if (!system_wide)
 728                 event__synthesize_thread(target_tid, process_synthesized_event,
 729                                          session);
 730         else
 731                 event__synthesize_threads(process_synthesized_event, session);
 732
 733         if (realtime_prio) {
 734                 struct sched_param param;
 735
 736                 param.sched_priority = realtime_prio;
 737                 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
 738                         pr_err("Could not set realtime priority.\n");
 739                         exit(-1);
 740                 }
 741         }
 742
 743         /*
 744          * Let the child rip
 745          */
 746         if (forks)
 747                 close(go_pipe[1]);
 748
 749         for (;;) {
 750                 int hits = samples;
 751                 int thread;
 752
 753                 mmap_read_all();
 754
 755                 if (hits == samples) {
 756                         if (done)
 757                                 break;
 758                         err = poll(event_array, nr_poll, -1);
 759                         waking++;
 760                 }
 761
 762                 if (done) {
 763                         for (i = 0; i < nr_cpu; i++) {
 764                                 for (counter = 0;
 765                                         counter < nr_counters;
 766                                         counter++) {
 767                                         for (thread = 0;
 768                                                 thread < thread_num;
 769                                                 thread++)
 770                                                 ioctl(fd[i][counter][thread],
 771                                                         PERF_EVENT_IOC_DISABLE);
 772                                 }
 773                         }
 774                 }
 775         }
 776
 777         if (quiet)
 778                 return 0;
 779
 780         fprintf(stderr, "[ perf record: Woken up %ld times to write data ]\n", waking);
 781
 782         /*
 783          * Approximate RIP event size: 24 bytes.
 784          */
 785         fprintf(stderr,
 786                 "[ perf record: Captured and wrote %.3f MB %s (~%lld samples) ]\n",
 787                 (double)bytes_written / 1024.0 / 1024.0,
 788                 output_name,
 789                 bytes_written / 24);
 790
 791         return 0;
 792
 793 out_delete_session:
 794         perf_session__delete(session);
 795         return err;
 796 }
 797
 798 static const char * const record_usage[] = {
 799         "perf record [<options>] [<command>]",
 800         "perf record [<options>] -- <command> [<options>]",
 801         NULL
 802 };
 803
 804 static bool force, append_file;
 805
 806 const struct option record_options[] = {
 807         OPT_CALLBACK('e', "event", NULL, "event",
 808                      "event selector. use 'perf list' to list available events",
 809                      parse_events),
 810         OPT_CALLBACK(0, "filter", NULL, "filter",
 811                      "event filter", parse_filter),
 812         OPT_INTEGER('p', "pid", &target_pid,
 813                     "record events on existing process id"),
 814         OPT_INTEGER('t', "tid", &target_tid,
 815                     "record events on existing thread id"),
 816         OPT_INTEGER('r', "realtime", &realtime_prio,
 817                     "collect data with this RT SCHED_FIFO priority"),
 818         OPT_BOOLEAN('R', "raw-samples", &raw_samples,
 819                     "collect raw sample records from all opened counters"),
 820         OPT_BOOLEAN('a', "all-cpus", &system_wide,
 821                             "system-wide collection from all CPUs"),
 822         OPT_BOOLEAN('A', "append", &append_file,
 823                             "append to the output file to do incremental profiling"),
 824         OPT_STRING('C', "cpu", &cpu_list, "cpu",
 825                     "list of cpus to monitor"),
 826         OPT_BOOLEAN('f', "force", &force,
 827                         "overwrite existing data file (deprecated)"),
 828         OPT_U64('c', "count", &user_interval, "event period to sample"),
 829         OPT_STRING('o', "output", &output_name, "file",
 830                     "output file name"),
 831         OPT_BOOLEAN('i', "no-inherit", &no_inherit,
 832                     "child tasks do not inherit counters"),
 833         OPT_UINTEGER('F', "freq", &user_freq, "profile at this frequency"),
 834         OPT_UINTEGER('m', "mmap-pages", &mmap_pages, "number of mmap data pages"),
 835         OPT_BOOLEAN('g', "call-graph", &call_graph,
 836                     "do call-graph (stack chain/backtrace) recording"),
 837         OPT_INCR('v', "verbose", &verbose,
 838                     "be more verbose (show counter open errors, etc)"),
 839         OPT_BOOLEAN('q', "quiet", &quiet, "don't print any message"),
 840         OPT_BOOLEAN('s', "stat", &inherit_stat,
 841                     "per thread counts"),
 842         OPT_BOOLEAN('d', "data", &sample_address,
 843                     "Sample addresses"),
 844         OPT_BOOLEAN('n', "no-samples", &no_samples,
 845                     "don't sample"),
 846         OPT_BOOLEAN('N', "no-buildid-cache", &no_buildid_cache,
 847                     "do not update the buildid cache"),
 848         OPT_BOOLEAN('B', "no-buildid", &no_buildid,
 849                     "do not collect buildids in perf.data"),
 850         OPT_END()
 851 };
 852
 853 int cmd_record(int argc, const char **argv, const char *prefix __used)
 854 {
 855         int i, j, err = -ENOMEM;
 856
 857         argc = parse_options(argc, argv, record_options, record_usage,
 858                             PARSE_OPT_STOP_AT_NON_OPTION);
 859         if (!argc && target_pid == -1 && target_tid == -1 &&
 860                 !system_wide && !cpu_list)
 861                 usage_with_options(record_usage, record_options);
 862
 863         if (force && append_file) {
 864                 fprintf(stderr, "Can't overwrite and append at the same time."
 865                                 " You need to choose between -f and -A");
 866                 usage_with_options(record_usage, record_options);
 867         } else if (append_file) {
 868                 write_mode = WRITE_APPEND;
 869         } else {
 870                 write_mode = WRITE_FORCE;
 871         }
 872
 873         symbol__init();
 874
 875         if (no_buildid_cache || no_buildid)
 876                 disable_buildid_cache();
 877
 878         if (!nr_counters) {
 879                 nr_counters     = 1;
 880                 attrs[0].type   = PERF_TYPE_HARDWARE;
 881                 attrs[0].config = PERF_COUNT_HW_CPU_CYCLES;
 882         }
 883
 884         if (target_pid != -1) {
 885                 target_tid = target_pid;
 886                 thread_num = find_all_tid(target_pid, &all_tids);
 887                 if (thread_num <= 0) {
 888                         fprintf(stderr, "Can't find all threads of pid %d\n",
 889                                         target_pid);
 890                         usage_with_options(record_usage, record_options);
 891                 }
 892         } else {
 893                 all_tids=malloc(sizeof(pid_t));
 894                 if (!all_tids)
 895                         goto out_symbol_exit;
 896
 897                 all_tids[0] = target_tid;
 898                 thread_num = 1;
 899         }
 900
 901         for (i = 0; i < MAX_NR_CPUS; i++) {
 902                 for (j = 0; j < MAX_COUNTERS; j++) {
 903                         fd[i][j] = malloc(sizeof(int)*thread_num);
 904                         if (!fd[i][j])
 905                                 goto out_free_fd;
 906                 }
 907         }
 908         event_array = malloc(
 909                 sizeof(struct pollfd)*MAX_NR_CPUS*MAX_COUNTERS*thread_num);
 910         if (!event_array)
 911                 goto out_free_fd;
 912
 913         if (user_interval != ULLONG_MAX)
 914                 default_interval = user_interval;
 915         if (user_freq != UINT_MAX)
 916                 freq = user_freq;
 917
 918         /*
 919          * User specified count overrides default frequency.
 920          */
 921         if (default_interval)
 922                 freq = 0;
 923         else if (freq) {
 924                 default_interval = freq;
 925         } else {
 926                 fprintf(stderr, "frequency and count are zero, aborting\n");
 927                 err = -EINVAL;
 928                 goto out_free_event_array;
 929         }
 930
 931         err = __cmd_record(argc, argv);
 932
 933 out_free_event_array:
 934         free(event_array);
 935 out_free_fd:
 936         for (i = 0; i < MAX_NR_CPUS; i++) {
 937                 for (j = 0; j < MAX_COUNTERS; j++)
 938                         free(fd[i][j]);
 939         }
 940         free(all_tids);
 941         all_tids = NULL;
 942 out_symbol_exit:
 943         symbol__exit();
 944         return err;
 945 }