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 struct timeval           last_read;
  69 static struct timeval           this_read;
  70
  71 static u64                      bytes_written                   =      0;
  72
  73 static struct pollfd            *event_array;
  74
  75 static int                      nr_poll                         =      0;
  76 static int                      nr_cpu                          =      0;
  77
  78 static int                      file_new                        =      1;
  79 static off_t                    post_processing_offset;
  80
  81 static struct perf_session      *session;
  82
  83 struct mmap_data {
  84         int                     counter;
  85         void                    *base;
  86         unsigned int            mask;
  87         unsigned int            prev;
  88 };
  89
  90 static struct mmap_data         *mmap_array[MAX_NR_CPUS][MAX_COUNTERS];
  91
  92 static unsigned long mmap_read_head(struct mmap_data *md)
  93 {
  94         struct perf_event_mmap_page *pc = md->base;
  95         long head;
  96
  97         head = pc->data_head;
  98         rmb();
  99
 100         return head;
 101 }
 102
 103 static void mmap_write_tail(struct mmap_data *md, unsigned long tail)
 104 {
 105         struct perf_event_mmap_page *pc = md->base;
 106
 107         /*
 108          * ensure all reads are done before we write the tail out.
 109          */
 110         /* mb(); */
 111         pc->data_tail = tail;
 112 }
 113
 114 static void advance_output(size_t size)
 115 {
 116         bytes_written += size;
 117 }
 118
 119 static void write_output(void *buf, size_t size)
 120 {
 121         while (size) {
 122                 int ret = write(output, buf, size);
 123
 124                 if (ret < 0)
 125                         die("failed to write");
 126
 127                 size -= ret;
 128                 buf += ret;
 129
 130                 bytes_written += ret;
 131         }
 132 }
 133
 134 static int process_synthesized_event(event_t *event,
 135                                      struct perf_session *self __used)
 136 {
 137         write_output(event, event->header.size);
 138         return 0;
 139 }
 140
 141 static void mmap_read(struct mmap_data *md)
 142 {
 143         unsigned int head = mmap_read_head(md);
 144         unsigned int old = md->prev;
 145         unsigned char *data = md->base + page_size;
 146         unsigned long size;
 147         void *buf;
 148         int diff;
 149
 150         gettimeofday(&this_read, NULL);
 151
 152         /*
 153          * If we're further behind than half the buffer, there's a chance
 154          * the writer will bite our tail and mess up the samples under us.
 155          *
 156          * If we somehow ended up ahead of the head, we got messed up.
 157          *
 158          * In either case, truncate and restart at head.
 159          */
 160         diff = head - old;
 161         if (diff < 0) {
 162                 struct timeval iv;
 163                 unsigned long msecs;
 164
 165                 timersub(&this_read, &last_read, &iv);
 166                 msecs = iv.tv_sec*1000 + iv.tv_usec/1000;
 167
 168                 fprintf(stderr, "WARNING: failed to keep up with mmap data."
 169                                 "  Last read %lu msecs ago.\n", msecs);
 170
 171                 /*
 172                  * head points to a known good entry, start there.
 173                  */
 174                 old = head;
 175         }
 176
 177         last_read = this_read;
 178
 179         if (old != head)
 180                 samples++;
 181
 182         size = head - old;
 183
 184         if ((old & md->mask) + size != (head & md->mask)) {
 185                 buf = &data[old & md->mask];
 186                 size = md->mask + 1 - (old & md->mask);
 187                 old += size;
 188
 189                 write_output(buf, size);
 190         }
 191
 192         buf = &data[old & md->mask];
 193         size = head - old;
 194         old += size;
 195
 196         write_output(buf, size);
 197
 198         md->prev = old;
 199         mmap_write_tail(md, old);
 200 }
 201
 202 static volatile int done = 0;
 203 static volatile int signr = -1;
 204
 205 static void sig_handler(int sig)
 206 {
 207         done = 1;
 208         signr = sig;
 209 }
 210
 211 static void sig_atexit(void)
 212 {
 213         if (child_pid != -1)
 214                 kill(child_pid, SIGTERM);
 215
 216         if (signr == -1)
 217                 return;
 218
 219         signal(signr, SIG_DFL);
 220         kill(getpid(), signr);
 221 }
 222
 223 static int group_fd;
 224
 225 static struct perf_header_attr *get_header_attr(struct perf_event_attr *a, int nr)
 226 {
 227         struct perf_header_attr *h_attr;
 228
 229         if (nr < session->header.attrs) {
 230                 h_attr = session->header.attr[nr];
 231         } else {
 232                 h_attr = perf_header_attr__new(a);
 233                 if (h_attr != NULL)
 234                         if (perf_header__add_attr(&session->header, h_attr) < 0) {
 235                                 perf_header_attr__delete(h_attr);
 236                                 h_attr = NULL;
 237                         }
 238         }
 239
 240         return h_attr;
 241 }
 242
 243 static void create_counter(int counter, int cpu)
 244 {
 245         char *filter = filters[counter];
 246         struct perf_event_attr *attr = attrs + counter;
 247         struct perf_header_attr *h_attr;
 248         int track = !counter; /* only the first counter needs these */
 249         int thread_index;
 250         int ret;
 251         struct {
 252                 u64 count;
 253                 u64 time_enabled;
 254                 u64 time_running;
 255                 u64 id;
 256         } read_data;
 257
 258         attr->read_format       = PERF_FORMAT_TOTAL_TIME_ENABLED |
 259                                   PERF_FORMAT_TOTAL_TIME_RUNNING |
 260                                   PERF_FORMAT_ID;
 261
 262         attr->sample_type       |= PERF_SAMPLE_IP | PERF_SAMPLE_TID;
 263
 264         if (nr_counters > 1)
 265                 attr->sample_type |= PERF_SAMPLE_ID;
 266
 267         /*
 268          * We default some events to a 1 default interval. But keep
 269          * it a weak assumption overridable by the user.
 270          */
 271         if (!attr->sample_period || (user_freq != UINT_MAX &&
 272                                      user_interval != ULLONG_MAX)) {
 273                 if (freq) {
 274                         attr->sample_type       |= PERF_SAMPLE_PERIOD;
 275                         attr->freq              = 1;
 276                         attr->sample_freq       = freq;
 277                 } else {
 278                         attr->sample_period = default_interval;
 279                 }
 280         }
 281
 282         if (no_samples)
 283                 attr->sample_freq = 0;
 284
 285         if (inherit_stat)
 286                 attr->inherit_stat = 1;
 287
 288         if (sample_address)
 289                 attr->sample_type       |= PERF_SAMPLE_ADDR;
 290
 291         if (call_graph)
 292                 attr->sample_type       |= PERF_SAMPLE_CALLCHAIN;
 293
 294         if (raw_samples) {
 295                 attr->sample_type       |= PERF_SAMPLE_TIME;
 296                 attr->sample_type       |= PERF_SAMPLE_RAW;
 297                 attr->sample_type       |= PERF_SAMPLE_CPU;
 298         }
 299
 300         attr->mmap              = track;
 301         attr->comm              = track;
 302         attr->inherit           = !no_inherit;
 303         if (target_pid == -1 && target_tid == -1 && !system_wide) {
 304                 attr->disabled = 1;
 305                 attr->enable_on_exec = 1;
 306         }
 307
 308         for (thread_index = 0; thread_index < thread_num; thread_index++) {
 309 try_again:
 310                 fd[nr_cpu][counter][thread_index] = sys_perf_event_open(attr,
 311                                 all_tids[thread_index], cpu, group_fd, 0);
 312
 313                 if (fd[nr_cpu][counter][thread_index] < 0) {
 314                         int err = errno;
 315
 316                         if (err == EPERM || err == EACCES)
 317                                 die("Permission error - are you root?\n"
 318                                         "\t Consider tweaking"
 319                                         " /proc/sys/kernel/perf_event_paranoid.\n");
 320                         else if (err ==  ENODEV && profile_cpu != -1) {
 321                                 die("No such device - did you specify"
 322                                         " an out-of-range profile CPU?\n");
 323                         }
 324
 325                         /*
 326                          * If it's cycles then fall back to hrtimer
 327                          * based cpu-clock-tick sw counter, which
 328                          * is always available even if no PMU support:
 329                          */
 330                         if (attr->type == PERF_TYPE_HARDWARE
 331                                         && attr->config == PERF_COUNT_HW_CPU_CYCLES) {
 332
 333                                 if (verbose)
 334                                         warning(" ... trying to fall back to cpu-clock-ticks\n");
 335                                 attr->type = PERF_TYPE_SOFTWARE;
 336                                 attr->config = PERF_COUNT_SW_CPU_CLOCK;
 337                                 goto try_again;
 338                         }
 339                         printf("\n");
 340                         error("perfcounter syscall returned with %d (%s)\n",
 341                                         fd[nr_cpu][counter][thread_index], strerror(err));
 342
 343 #if defined(__i386__) || defined(__x86_64__)
 344                         if (attr->type == PERF_TYPE_HARDWARE && err == EOPNOTSUPP)
 345                                 die("No hardware sampling interrupt available."
 346                                     " No APIC? If so then you can boot the kernel"
 347                                     " with the \"lapic\" boot parameter to"
 348                                     " force-enable it.\n");
 349 #endif
 350
 351                         die("No CONFIG_PERF_EVENTS=y kernel support configured?\n");
 352                         exit(-1);
 353                 }
 354
 355                 h_attr = get_header_attr(attr, counter);
 356                 if (h_attr == NULL)
 357                         die("nomem\n");
 358
 359                 if (!file_new) {
 360                         if (memcmp(&h_attr->attr, attr, sizeof(*attr))) {
 361                                 fprintf(stderr, "incompatible append\n");
 362                                 exit(-1);
 363                         }
 364                 }
 365
 366                 if (read(fd[nr_cpu][counter][thread_index], &read_data, sizeof(read_data)) == -1) {
 367                         perror("Unable to read perf file descriptor\n");
 368                         exit(-1);
 369                 }
 370
 371                 if (perf_header_attr__add_id(h_attr, read_data.id) < 0) {
 372                         pr_warning("Not enough memory to add id\n");
 373                         exit(-1);
 374                 }
 375
 376                 assert(fd[nr_cpu][counter][thread_index] >= 0);
 377                 fcntl(fd[nr_cpu][counter][thread_index], F_SETFL, O_NONBLOCK);
 378
 379                 /*
 380                  * First counter acts as the group leader:
 381                  */
 382                 if (group && group_fd == -1)
 383                         group_fd = fd[nr_cpu][counter][thread_index];
 384                 if (multiplex && multiplex_fd == -1)
 385                         multiplex_fd = fd[nr_cpu][counter][thread_index];
 386
 387                 if (multiplex && fd[nr_cpu][counter][thread_index] != multiplex_fd) {
 388
 389                         ret = ioctl(fd[nr_cpu][counter][thread_index], PERF_EVENT_IOC_SET_OUTPUT, multiplex_fd);
 390                         assert(ret != -1);
 391                 } else {
 392                         event_array[nr_poll].fd = fd[nr_cpu][counter][thread_index];
 393                         event_array[nr_poll].events = POLLIN;
 394                         nr_poll++;
 395
 396                         mmap_array[nr_cpu][counter][thread_index].counter = counter;
 397                         mmap_array[nr_cpu][counter][thread_index].prev = 0;
 398                         mmap_array[nr_cpu][counter][thread_index].mask = mmap_pages*page_size - 1;
 399                         mmap_array[nr_cpu][counter][thread_index].base = mmap(NULL, (mmap_pages+1)*page_size,
 400                                 PROT_READ|PROT_WRITE, MAP_SHARED, fd[nr_cpu][counter][thread_index], 0);
 401                         if (mmap_array[nr_cpu][counter][thread_index].base == MAP_FAILED) {
 402                                 error("failed to mmap with %d (%s)\n", errno, strerror(errno));
 403                                 exit(-1);
 404                         }
 405                 }
 406
 407                 if (filter != NULL) {
 408                         ret = ioctl(fd[nr_cpu][counter][thread_index],
 409                                         PERF_EVENT_IOC_SET_FILTER, filter);
 410                         if (ret) {
 411                                 error("failed to set filter with %d (%s)\n", errno,
 412                                                 strerror(errno));
 413                                 exit(-1);
 414                         }
 415                 }
 416         }
 417 }
 418
 419 static void open_counters(int cpu)
 420 {
 421         int counter;
 422
 423         group_fd = -1;
 424         for (counter = 0; counter < nr_counters; counter++)
 425                 create_counter(counter, cpu);
 426
 427         nr_cpu++;
 428 }
 429
 430 static int process_buildids(void)
 431 {
 432         u64 size = lseek(output, 0, SEEK_CUR);
 433
 434         if (size == 0)
 435                 return 0;
 436
 437         session->fd = output;
 438         return __perf_session__process_events(session, post_processing_offset,
 439                                               size - post_processing_offset,
 440                                               size, &build_id__mark_dso_hit_ops);
 441 }
 442
 443 static void atexit_header(void)
 444 {
 445         if (!pipe_output) {
 446                 session->header.data_size += bytes_written;
 447
 448                 process_buildids();
 449                 perf_header__write(&session->header, output, true);
 450         }
 451 }
 452
 453 static void event__synthesize_guest_os(struct machine *machine, void *data)
 454 {
 455         int err;
 456         char *guest_kallsyms;
 457         char path[PATH_MAX];
 458         struct perf_session *psession = data;
 459
 460         if (machine__is_host(machine))
 461                 return;
 462
 463         /*
 464          *As for guest kernel when processing subcommand record&report,
 465          *we arrange module mmap prior to guest kernel mmap and trigger
 466          *a preload dso because default guest module symbols are loaded
 467          *from guest kallsyms instead of /lib/modules/XXX/XXX. This
 468          *method is used to avoid symbol missing when the first addr is
 469          *in module instead of in guest kernel.
 470          */
 471         err = event__synthesize_modules(process_synthesized_event,
 472                                         psession, machine);
 473         if (err < 0)
 474                 pr_err("Couldn't record guest kernel [%d]'s reference"
 475                        " relocation symbol.\n", machine->pid);
 476
 477         if (machine__is_default_guest(machine))
 478                 guest_kallsyms = (char *) symbol_conf.default_guest_kallsyms;
 479         else {
 480                 sprintf(path, "%s/proc/kallsyms", machine->root_dir);
 481                 guest_kallsyms = path;
 482         }
 483
 484         /*
 485          * We use _stext for guest kernel because guest kernel's /proc/kallsyms
 486          * have no _text sometimes.
 487          */
 488         err = event__synthesize_kernel_mmap(process_synthesized_event,
 489                                             psession, machine, "_text");
 490         if (err < 0)
 491                 err = event__synthesize_kernel_mmap(process_synthesized_event,
 492                                                     psession, machine, "_stext");
 493         if (err < 0)
 494                 pr_err("Couldn't record guest kernel [%d]'s reference"
 495                        " relocation symbol.\n", machine->pid);
 496 }
 497
 498 static struct perf_event_header finished_round_event = {
 499         .size = sizeof(struct perf_event_header),
 500         .type = PERF_RECORD_FINISHED_ROUND,
 501 };
 502
 503 static void mmap_read_all(void)
 504 {
 505         int i, counter, thread;
 506
 507         for (i = 0; i < nr_cpu; i++) {
 508                 for (counter = 0; counter < nr_counters; counter++) {
 509                         for (thread = 0; thread < thread_num; thread++) {
 510                                 if (mmap_array[i][counter][thread].base)
 511                                         mmap_read(&mmap_array[i][counter][thread]);
 512                         }
 513
 514                 }
 515         }
 516
 517         if (perf_header__has_feat(&session->header, HEADER_TRACE_INFO))
 518                 write_output(&finished_round_event, sizeof(finished_round_event));
 519 }
 520
 521 static int __cmd_record(int argc, const char **argv)
 522 {
 523         int i, counter;
 524         struct stat st;
 525         pid_t pid = 0;
 526         int flags;
 527         int err;
 528         unsigned long waking = 0;
 529         int child_ready_pipe[2], go_pipe[2];
 530         const bool forks = argc > 0;
 531         char buf;
 532         struct machine *machine;
 533
 534         page_size = sysconf(_SC_PAGE_SIZE);
 535
 536         atexit(sig_atexit);
 537         signal(SIGCHLD, sig_handler);
 538         signal(SIGINT, sig_handler);
 539
 540         if (forks && (pipe(child_ready_pipe) < 0 || pipe(go_pipe) < 0)) {
 541                 perror("failed to create pipes");
 542                 exit(-1);
 543         }
 544
 545         if (!strcmp(output_name, "-"))
 546                 pipe_output = 1;
 547         else if (!stat(output_name, &st) && st.st_size) {
 548                 if (write_mode == WRITE_FORCE) {
 549                         char oldname[PATH_MAX];
 550                         snprintf(oldname, sizeof(oldname), "%s.old",
 551                                  output_name);
 552                         unlink(oldname);
 553                         rename(output_name, oldname);
 554                 }
 555         } else if (write_mode == WRITE_APPEND) {
 556                 write_mode = WRITE_FORCE;
 557         }
 558
 559         flags = O_CREAT|O_RDWR;
 560         if (write_mode == WRITE_APPEND)
 561                 file_new = 0;
 562         else
 563                 flags |= O_TRUNC;
 564
 565         if (pipe_output)
 566                 output = STDOUT_FILENO;
 567         else
 568                 output = open(output_name, flags, S_IRUSR | S_IWUSR);
 569         if (output < 0) {
 570                 perror("failed to create output file");
 571                 exit(-1);
 572         }
 573
 574         session = perf_session__new(output_name, O_WRONLY,
 575                                     write_mode == WRITE_FORCE, false);
 576         if (session == NULL) {
 577                 pr_err("Not enough memory for reading perf file header\n");
 578                 return -1;
 579         }
 580
 581         if (!file_new) {
 582                 err = perf_header__read(session, output);
 583                 if (err < 0)
 584                         return err;
 585         }
 586
 587         if (have_tracepoints(attrs, nr_counters))
 588                 perf_header__set_feat(&session->header, HEADER_TRACE_INFO);
 589
 590         atexit(atexit_header);
 591
 592         if (forks) {
 593                 child_pid = fork();
 594                 if (pid < 0) {
 595                         perror("failed to fork");
 596                         exit(-1);
 597                 }
 598
 599                 if (!child_pid) {
 600                         if (pipe_output)
 601                                 dup2(2, 1);
 602                         close(child_ready_pipe[0]);
 603                         close(go_pipe[1]);
 604                         fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
 605
 606                         /*
 607                          * Do a dummy execvp to get the PLT entry resolved,
 608                          * so we avoid the resolver overhead on the real
 609                          * execvp call.
 610                          */
 611                         execvp("", (char **)argv);
 612
 613                         /*
 614                          * Tell the parent we're ready to go
 615                          */
 616                         close(child_ready_pipe[1]);
 617
 618                         /*
 619                          * Wait until the parent tells us to go.
 620                          */
 621                         if (read(go_pipe[0], &buf, 1) == -1)
 622                                 perror("unable to read pipe");
 623
 624                         execvp(argv[0], (char **)argv);
 625
 626                         perror(argv[0]);
 627                         exit(-1);
 628                 }
 629
 630                 if (!system_wide && target_tid == -1 && target_pid == -1)
 631                         all_tids[0] = child_pid;
 632
 633                 close(child_ready_pipe[1]);
 634                 close(go_pipe[0]);
 635                 /*
 636                  * wait for child to settle
 637                  */
 638                 if (read(child_ready_pipe[0], &buf, 1) == -1) {
 639                         perror("unable to read pipe");
 640                         exit(-1);
 641                 }
 642                 close(child_ready_pipe[0]);
 643         }
 644
 645         if ((!system_wide && no_inherit) || profile_cpu != -1) {
 646                 open_counters(profile_cpu);
 647         } else {
 648                 nr_cpus = read_cpu_map();
 649                 for (i = 0; i < nr_cpus; i++)
 650                         open_counters(cpumap[i]);
 651         }
 652
 653         if (pipe_output) {
 654                 err = perf_header__write_pipe(output);
 655                 if (err < 0)
 656                         return err;
 657         } else if (file_new) {
 658                 err = perf_header__write(&session->header, output, false);
 659                 if (err < 0)
 660                         return err;
 661         }
 662
 663         post_processing_offset = lseek(output, 0, SEEK_CUR);
 664
 665         if (pipe_output) {
 666                 err = event__synthesize_attrs(&session->header,
 667                                               process_synthesized_event,
 668                                               session);
 669                 if (err < 0) {
 670                         pr_err("Couldn't synthesize attrs.\n");
 671                         return err;
 672                 }
 673
 674                 err = event__synthesize_event_types(process_synthesized_event,
 675                                                     session);
 676                 if (err < 0) {
 677                         pr_err("Couldn't synthesize event_types.\n");
 678                         return err;
 679                 }
 680
 681                 if (have_tracepoints(attrs, nr_counters)) {
 682                         /*
 683                          * FIXME err <= 0 here actually means that
 684                          * there were no tracepoints so its not really
 685                          * an error, just that we don't need to
 686                          * synthesize anything.  We really have to
 687                          * return this more properly and also
 688                          * propagate errors that now are calling die()
 689                          */
 690                         err = event__synthesize_tracing_data(output, attrs,
 691                                                              nr_counters,
 692                                                              process_synthesized_event,
 693                                                              session);
 694                         if (err <= 0) {
 695                                 pr_err("Couldn't record tracing data.\n");
 696                                 return err;
 697                         }
 698                         advance_output(err);
 699                 }
 700         }
 701
 702         machine = perf_session__find_host_machine(session);
 703         if (!machine) {
 704                 pr_err("Couldn't find native kernel information.\n");
 705                 return -1;
 706         }
 707
 708         err = event__synthesize_kernel_mmap(process_synthesized_event,
 709                                             session, machine, "_text");
 710         if (err < 0)
 711                 err = event__synthesize_kernel_mmap(process_synthesized_event,
 712                                                     session, machine, "_stext");
 713         if (err < 0) {
 714                 pr_err("Couldn't record kernel reference relocation symbol.\n");
 715                 return err;
 716         }
 717
 718         err = event__synthesize_modules(process_synthesized_event,
 719                                         session, machine);
 720         if (err < 0) {
 721                 pr_err("Couldn't record kernel reference relocation symbol.\n");
 722                 return err;
 723         }
 724         if (perf_guest)
 725                 perf_session__process_machines(session, event__synthesize_guest_os);
 726
 727         if (!system_wide && profile_cpu == -1)
 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         fprintf(stderr, "[ perf record: Woken up %ld times to write data ]\n", waking);
 778
 779         /*
 780          * Approximate RIP event size: 24 bytes.
 781          */
 782         fprintf(stderr,
 783                 "[ perf record: Captured and wrote %.3f MB %s (~%lld samples) ]\n",
 784                 (double)bytes_written / 1024.0 / 1024.0,
 785                 output_name,
 786                 bytes_written / 24);
 787
 788         return 0;
 789 }
 790
 791 static const char * const record_usage[] = {
 792         "perf record [<options>] [<command>]",
 793         "perf record [<options>] -- <command> [<options>]",
 794         NULL
 795 };
 796
 797 static bool force, append_file;
 798
 799 static const struct option options[] = {
 800         OPT_CALLBACK('e', "event", NULL, "event",
 801                      "event selector. use 'perf list' to list available events",
 802                      parse_events),
 803         OPT_CALLBACK(0, "filter", NULL, "filter",
 804                      "event filter", parse_filter),
 805         OPT_INTEGER('p', "pid", &target_pid,
 806                     "record events on existing process id"),
 807         OPT_INTEGER('t', "tid", &target_tid,
 808                     "record events on existing thread id"),
 809         OPT_INTEGER('r', "realtime", &realtime_prio,
 810                     "collect data with this RT SCHED_FIFO priority"),
 811         OPT_BOOLEAN('R', "raw-samples", &raw_samples,
 812                     "collect raw sample records from all opened counters"),
 813         OPT_BOOLEAN('a', "all-cpus", &system_wide,
 814                             "system-wide collection from all CPUs"),
 815         OPT_BOOLEAN('A', "append", &append_file,
 816                             "append to the output file to do incremental profiling"),
 817         OPT_INTEGER('C', "profile_cpu", &profile_cpu,
 818                             "CPU to profile on"),
 819         OPT_BOOLEAN('f', "force", &force,
 820                         "overwrite existing data file (deprecated)"),
 821         OPT_U64('c', "count", &user_interval, "event period to sample"),
 822         OPT_STRING('o', "output", &output_name, "file",
 823                     "output file name"),
 824         OPT_BOOLEAN('i', "no-inherit", &no_inherit,
 825                     "child tasks do not inherit counters"),
 826         OPT_UINTEGER('F', "freq", &user_freq, "profile at this frequency"),
 827         OPT_UINTEGER('m', "mmap-pages", &mmap_pages, "number of mmap data pages"),
 828         OPT_BOOLEAN('g', "call-graph", &call_graph,
 829                     "do call-graph (stack chain/backtrace) recording"),
 830         OPT_INCR('v', "verbose", &verbose,
 831                     "be more verbose (show counter open errors, etc)"),
 832         OPT_BOOLEAN('s', "stat", &inherit_stat,
 833                     "per thread counts"),
 834         OPT_BOOLEAN('d', "data", &sample_address,
 835                     "Sample addresses"),
 836         OPT_BOOLEAN('n', "no-samples", &no_samples,
 837                     "don't sample"),
 838         OPT_BOOLEAN('M', "multiplex", &multiplex,
 839                     "multiplex counter output in a single channel"),
 840         OPT_END()
 841 };
 842
 843 int cmd_record(int argc, const char **argv, const char *prefix __used)
 844 {
 845         int i,j;
 846
 847         argc = parse_options(argc, argv, options, record_usage,
 848                             PARSE_OPT_STOP_AT_NON_OPTION);
 849         if (!argc && target_pid == -1 && target_tid == -1 &&
 850                 !system_wide && profile_cpu == -1)
 851                 usage_with_options(record_usage, options);
 852
 853         if (force && append_file) {
 854                 fprintf(stderr, "Can't overwrite and append at the same time."
 855                                 " You need to choose between -f and -A");
 856                 usage_with_options(record_usage, options);
 857         } else if (append_file) {
 858                 write_mode = WRITE_APPEND;
 859         } else {
 860                 write_mode = WRITE_FORCE;
 861         }
 862
 863         symbol__init();
 864
 865         if (!nr_counters) {
 866                 nr_counters     = 1;
 867                 attrs[0].type   = PERF_TYPE_HARDWARE;
 868                 attrs[0].config = PERF_COUNT_HW_CPU_CYCLES;
 869         }
 870
 871         if (target_pid != -1) {
 872                 target_tid = target_pid;
 873                 thread_num = find_all_tid(target_pid, &all_tids);
 874                 if (thread_num <= 0) {
 875                         fprintf(stderr, "Can't find all threads of pid %d\n",
 876                                         target_pid);
 877                         usage_with_options(record_usage, options);
 878                 }
 879         } else {
 880                 all_tids=malloc(sizeof(pid_t));
 881                 if (!all_tids)
 882                         return -ENOMEM;
 883
 884                 all_tids[0] = target_tid;
 885                 thread_num = 1;
 886         }
 887
 888         for (i = 0; i < MAX_NR_CPUS; i++) {
 889                 for (j = 0; j < MAX_COUNTERS; j++) {
 890                         fd[i][j] = malloc(sizeof(int)*thread_num);
 891                         mmap_array[i][j] = zalloc(
 892                                 sizeof(struct mmap_data)*thread_num);
 893                         if (!fd[i][j] || !mmap_array[i][j])
 894                                 return -ENOMEM;
 895                 }
 896         }
 897         event_array = malloc(
 898                 sizeof(struct pollfd)*MAX_NR_CPUS*MAX_COUNTERS*thread_num);
 899         if (!event_array)
 900                 return -ENOMEM;
 901
 902         if (user_interval != ULLONG_MAX)
 903                 default_interval = user_interval;
 904         if (user_freq != UINT_MAX)
 905                 freq = user_freq;
 906
 907         /*
 908          * User specified count overrides default frequency.
 909          */
 910         if (default_interval)
 911                 freq = 0;
 912         else if (freq) {
 913                 default_interval = freq;
 914         } else {
 915                 fprintf(stderr, "frequency and count are zero, aborting\n");
 916                 exit(EXIT_FAILURE);
 917         }
 918
 919         return __cmd_record(argc, argv);
 920 }