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/evsel.h"
  22 #include "util/debug.h"
  23 #include "util/session.h"
  24 #include "util/symbol.h"
  25 #include "util/cpumap.h"
  26
  27 #include <unistd.h>
  28 #include <sched.h>
  29 #include <sys/mman.h>
  30
  31 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
  32
  33 enum write_mode_t {
  34         WRITE_FORCE,
  35         WRITE_APPEND
  36 };
  37
  38 static u64                      user_interval                   = ULLONG_MAX;
  39 static u64                      default_interval                =      0;
  40 static u64                      sample_type;
  41
  42 static struct cpu_map           *cpus;
  43 static unsigned int             page_size;
  44 static unsigned int             mmap_pages                      =    128;
  45 static unsigned int             user_freq                       = UINT_MAX;
  46 static int                      freq                            =   1000;
  47 static int                      output;
  48 static int                      pipe_output                     =      0;
  49 static const char               *output_name                    = "perf.data";
  50 static int                      group                           =      0;
  51 static int                      realtime_prio                   =      0;
  52 static bool                     raw_samples                     =  false;
  53 static bool                     sample_id_all_avail             =   true;
  54 static bool                     system_wide                     =  false;
  55 static pid_t                    target_pid                      =     -1;
  56 static pid_t                    target_tid                      =     -1;
  57 static struct thread_map        *threads;
  58 static pid_t                    child_pid                       =     -1;
  59 static bool                     no_inherit                      =  false;
  60 static enum write_mode_t        write_mode                      = WRITE_FORCE;
  61 static bool                     call_graph                      =  false;
  62 static bool                     inherit_stat                    =  false;
  63 static bool                     no_samples                      =  false;
  64 static bool                     sample_address                  =  false;
  65 static bool                     sample_time                     =  false;
  66 static bool                     no_buildid                      =  false;
  67 static bool                     no_buildid_cache                =  false;
  68
  69 static long                     samples                         =      0;
  70 static u64                      bytes_written                   =      0;
  71
  72 static struct pollfd            *event_array;
  73
  74 static int                      nr_poll                         =      0;
  75 static int                      nr_cpu                          =      0;
  76
  77 static int                      file_new                        =      1;
  78 static off_t                    post_processing_offset;
  79
  80 static struct perf_session      *session;
  81 static const char               *cpu_list;
  82
  83 struct mmap_data {
  84         void                    *base;
  85         unsigned int            mask;
  86         unsigned int            prev;
  87 };
  88
  89 static struct mmap_data         mmap_array[MAX_NR_CPUS];
  90
  91 static unsigned long mmap_read_head(struct mmap_data *md)
  92 {
  93         struct perf_event_mmap_page *pc = md->base;
  94         long head;
  95
  96         head = pc->data_head;
  97         rmb();
  98
  99         return head;
 100 }
 101
 102 static void mmap_write_tail(struct mmap_data *md, unsigned long tail)
 103 {
 104         struct perf_event_mmap_page *pc = md->base;
 105
 106         /*
 107          * ensure all reads are done before we write the tail out.
 108          */
 109         /* mb(); */
 110         pc->data_tail = tail;
 111 }
 112
 113 static void advance_output(size_t size)
 114 {
 115         bytes_written += size;
 116 }
 117
 118 static void write_output(void *buf, size_t size)
 119 {
 120         while (size) {
 121                 int ret = write(output, buf, size);
 122
 123                 if (ret < 0)
 124                         die("failed to write");
 125
 126                 size -= ret;
 127                 buf += ret;
 128
 129                 bytes_written += ret;
 130         }
 131 }
 132
 133 static int process_synthesized_event(event_t *event,
 134                                      struct sample_data *sample __used,
 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         /*
 151          * If we're further behind than half the buffer, there's a chance
 152          * the writer will bite our tail and mess up the samples under us.
 153          *
 154          * If we somehow ended up ahead of the head, we got messed up.
 155          *
 156          * In either case, truncate and restart at head.
 157          */
 158         diff = head - old;
 159         if (diff < 0) {
 160                 fprintf(stderr, "WARNING: failed to keep up with mmap data\n");
 161                 /*
 162                  * head points to a known good entry, start there.
 163                  */
 164                 old = head;
 165         }
 166
 167         if (old != head)
 168                 samples++;
 169
 170         size = head - old;
 171
 172         if ((old & md->mask) + size != (head & md->mask)) {
 173                 buf = &data[old & md->mask];
 174                 size = md->mask + 1 - (old & md->mask);
 175                 old += size;
 176
 177                 write_output(buf, size);
 178         }
 179
 180         buf = &data[old & md->mask];
 181         size = head - old;
 182         old += size;
 183
 184         write_output(buf, size);
 185
 186         md->prev = old;
 187         mmap_write_tail(md, old);
 188 }
 189
 190 static volatile int done = 0;
 191 static volatile int signr = -1;
 192
 193 static void sig_handler(int sig)
 194 {
 195         done = 1;
 196         signr = sig;
 197 }
 198
 199 static void sig_atexit(void)
 200 {
 201         if (child_pid > 0)
 202                 kill(child_pid, SIGTERM);
 203
 204         if (signr == -1 || signr == SIGUSR1)
 205                 return;
 206
 207         signal(signr, SIG_DFL);
 208         kill(getpid(), signr);
 209 }
 210
 211 static int group_fd;
 212
 213 static struct perf_header_attr *get_header_attr(struct perf_event_attr *a, int nr)
 214 {
 215         struct perf_header_attr *h_attr;
 216
 217         if (nr < session->header.attrs) {
 218                 h_attr = session->header.attr[nr];
 219         } else {
 220                 h_attr = perf_header_attr__new(a);
 221                 if (h_attr != NULL)
 222                         if (perf_header__add_attr(&session->header, h_attr) < 0) {
 223                                 perf_header_attr__delete(h_attr);
 224                                 h_attr = NULL;
 225                         }
 226         }
 227
 228         return h_attr;
 229 }
 230
 231 static void create_counter(struct perf_evsel *evsel, int cpu)
 232 {
 233         char *filter = evsel->filter;
 234         struct perf_event_attr *attr = &evsel->attr;
 235         struct perf_header_attr *h_attr;
 236         int track = !evsel->idx; /* only the first counter needs these */
 237         int thread_index;
 238         int ret;
 239         struct {
 240                 u64 count;
 241                 u64 time_enabled;
 242                 u64 time_running;
 243                 u64 id;
 244         } read_data;
 245         /*
 246          * Check if parse_single_tracepoint_event has already asked for
 247          * PERF_SAMPLE_TIME.
 248          *
 249          * XXX this is kludgy but short term fix for problems introduced by
 250          * eac23d1c that broke 'perf script' by having different sample_types
 251          * when using multiple tracepoint events when we use a perf binary
 252          * that tries to use sample_id_all on an older kernel.
 253          *
 254          * We need to move counter creation to perf_session, support
 255          * different sample_types, etc.
 256          */
 257         bool time_needed = attr->sample_type & PERF_SAMPLE_TIME;
 258
 259         attr->read_format       = PERF_FORMAT_TOTAL_TIME_ENABLED |
 260                                   PERF_FORMAT_TOTAL_TIME_RUNNING |
 261                                   PERF_FORMAT_ID;
 262
 263         attr->sample_type       |= PERF_SAMPLE_IP | PERF_SAMPLE_TID;
 264
 265         if (nr_counters > 1)
 266                 attr->sample_type |= PERF_SAMPLE_ID;
 267
 268         /*
 269          * We default some events to a 1 default interval. But keep
 270          * it a weak assumption overridable by the user.
 271          */
 272         if (!attr->sample_period || (user_freq != UINT_MAX &&
 273                                      user_interval != ULLONG_MAX)) {
 274                 if (freq) {
 275                         attr->sample_type       |= PERF_SAMPLE_PERIOD;
 276                         attr->freq              = 1;
 277                         attr->sample_freq       = freq;
 278                 } else {
 279                         attr->sample_period = default_interval;
 280                 }
 281         }
 282
 283         if (no_samples)
 284                 attr->sample_freq = 0;
 285
 286         if (inherit_stat)
 287                 attr->inherit_stat = 1;
 288
 289         if (sample_address) {
 290                 attr->sample_type       |= PERF_SAMPLE_ADDR;
 291                 attr->mmap_data = track;
 292         }
 293
 294         if (call_graph)
 295                 attr->sample_type       |= PERF_SAMPLE_CALLCHAIN;
 296
 297         if (system_wide)
 298                 attr->sample_type       |= PERF_SAMPLE_CPU;
 299
 300         if (sample_id_all_avail &&
 301             (sample_time || system_wide || !no_inherit || cpu_list))
 302                 attr->sample_type       |= PERF_SAMPLE_TIME;
 303
 304         if (raw_samples) {
 305                 attr->sample_type       |= PERF_SAMPLE_TIME;
 306                 attr->sample_type       |= PERF_SAMPLE_RAW;
 307                 attr->sample_type       |= PERF_SAMPLE_CPU;
 308         }
 309
 310         attr->mmap              = track;
 311         attr->comm              = track;
 312         attr->inherit           = !no_inherit;
 313         if (target_pid == -1 && target_tid == -1 && !system_wide) {
 314                 attr->disabled = 1;
 315                 attr->enable_on_exec = 1;
 316         }
 317 retry_sample_id:
 318         attr->sample_id_all = sample_id_all_avail ? 1 : 0;
 319
 320         for (thread_index = 0; thread_index < threads->nr; thread_index++) {
 321 try_again:
 322                 FD(evsel, nr_cpu, thread_index) = sys_perf_event_open(attr, threads->map[thread_index], cpu, group_fd, 0);
 323
 324                 if (FD(evsel, nr_cpu, thread_index) < 0) {
 325                         int err = errno;
 326
 327                         if (err == EPERM || err == EACCES)
 328                                 die("Permission error - are you root?\n"
 329                                         "\t Consider tweaking"
 330                                         " /proc/sys/kernel/perf_event_paranoid.\n");
 331                         else if (err ==  ENODEV && cpu_list) {
 332                                 die("No such device - did you specify"
 333                                         " an out-of-range profile CPU?\n");
 334                         } else if (err == EINVAL && sample_id_all_avail) {
 335                                 /*
 336                                  * Old kernel, no attr->sample_id_type_all field
 337                                  */
 338                                 sample_id_all_avail = false;
 339                                 if (!sample_time && !raw_samples && !time_needed)
 340                                         attr->sample_type &= ~PERF_SAMPLE_TIME;
 341
 342                                 goto retry_sample_id;
 343                         }
 344
 345                         /*
 346                          * If it's cycles then fall back to hrtimer
 347                          * based cpu-clock-tick sw counter, which
 348                          * is always available even if no PMU support:
 349                          */
 350                         if (attr->type == PERF_TYPE_HARDWARE
 351                                         && attr->config == PERF_COUNT_HW_CPU_CYCLES) {
 352
 353                                 if (verbose)
 354                                         warning(" ... trying to fall back to cpu-clock-ticks\n");
 355                                 attr->type = PERF_TYPE_SOFTWARE;
 356                                 attr->config = PERF_COUNT_SW_CPU_CLOCK;
 357                                 goto try_again;
 358                         }
 359                         printf("\n");
 360                         error("sys_perf_event_open() syscall returned with %d (%s).  /bin/dmesg may provide additional information.\n",
 361                               FD(evsel, nr_cpu, thread_index), strerror(err));
 362
 363 #if defined(__i386__) || defined(__x86_64__)
 364                         if (attr->type == PERF_TYPE_HARDWARE && err == EOPNOTSUPP)
 365                                 die("No hardware sampling interrupt available."
 366                                     " No APIC? If so then you can boot the kernel"
 367                                     " with the \"lapic\" boot parameter to"
 368                                     " force-enable it.\n");
 369 #endif
 370
 371                         die("No CONFIG_PERF_EVENTS=y kernel support configured?\n");
 372                         exit(-1);
 373                 }
 374
 375                 h_attr = get_header_attr(attr, evsel->idx);
 376                 if (h_attr == NULL)
 377                         die("nomem\n");
 378
 379                 if (!file_new) {
 380                         if (memcmp(&h_attr->attr, attr, sizeof(*attr))) {
 381                                 fprintf(stderr, "incompatible append\n");
 382                                 exit(-1);
 383                         }
 384                 }
 385
 386                 if (read(FD(evsel, nr_cpu, thread_index), &read_data, sizeof(read_data)) == -1) {
 387                         perror("Unable to read perf file descriptor");
 388                         exit(-1);
 389                 }
 390
 391                 if (perf_header_attr__add_id(h_attr, read_data.id) < 0) {
 392                         pr_warning("Not enough memory to add id\n");
 393                         exit(-1);
 394                 }
 395
 396                 assert(FD(evsel, nr_cpu, thread_index) >= 0);
 397                 fcntl(FD(evsel, nr_cpu, thread_index), F_SETFL, O_NONBLOCK);
 398
 399                 /*
 400                  * First counter acts as the group leader:
 401                  */
 402                 if (group && group_fd == -1)
 403                         group_fd = FD(evsel, nr_cpu, thread_index);
 404
 405                 if (evsel->idx || thread_index) {
 406                         struct perf_evsel *first;
 407                         first = list_entry(evsel_list.next, struct perf_evsel, node);
 408                         ret = ioctl(FD(evsel, nr_cpu, thread_index),
 409                                     PERF_EVENT_IOC_SET_OUTPUT,
 410                                     FD(first, nr_cpu, 0));
 411                         if (ret) {
 412                                 error("failed to set output: %d (%s)\n", errno,
 413                                                 strerror(errno));
 414                                 exit(-1);
 415                         }
 416                 } else {
 417                         mmap_array[nr_cpu].prev = 0;
 418                         mmap_array[nr_cpu].mask = mmap_pages*page_size - 1;
 419                         mmap_array[nr_cpu].base = mmap(NULL, (mmap_pages+1)*page_size,
 420                                 PROT_READ | PROT_WRITE, MAP_SHARED, FD(evsel, nr_cpu, thread_index), 0);
 421                         if (mmap_array[nr_cpu].base == MAP_FAILED) {
 422                                 error("failed to mmap with %d (%s)\n", errno, strerror(errno));
 423                                 exit(-1);
 424                         }
 425
 426                         event_array[nr_poll].fd = FD(evsel, nr_cpu, thread_index);
 427                         event_array[nr_poll].events = POLLIN;
 428                         nr_poll++;
 429                 }
 430
 431                 if (filter != NULL) {
 432                         ret = ioctl(FD(evsel, nr_cpu, thread_index),
 433                                     PERF_EVENT_IOC_SET_FILTER, filter);
 434                         if (ret) {
 435                                 error("failed to set filter with %d (%s)\n", errno,
 436                                                 strerror(errno));
 437                                 exit(-1);
 438                         }
 439                 }
 440         }
 441
 442         if (!sample_type)
 443                 sample_type = attr->sample_type;
 444 }
 445
 446 static void open_counters(int cpu)
 447 {
 448         struct perf_evsel *pos;
 449
 450         group_fd = -1;
 451
 452         list_for_each_entry(pos, &evsel_list, node)
 453                 create_counter(pos, cpu);
 454
 455         nr_cpu++;
 456 }
 457
 458 static int process_buildids(void)
 459 {
 460         u64 size = lseek(output, 0, SEEK_CUR);
 461
 462         if (size == 0)
 463                 return 0;
 464
 465         session->fd = output;
 466         return __perf_session__process_events(session, post_processing_offset,
 467                                               size - post_processing_offset,
 468                                               size, &build_id__mark_dso_hit_ops);
 469 }
 470
 471 static void atexit_header(void)
 472 {
 473         if (!pipe_output) {
 474                 session->header.data_size += bytes_written;
 475
 476                 if (!no_buildid)
 477                         process_buildids();
 478                 perf_header__write(&session->header, output, true);
 479                 perf_session__delete(session);
 480                 perf_evsel_list__delete();
 481                 symbol__exit();
 482         }
 483 }
 484
 485 static void event__synthesize_guest_os(struct machine *machine, void *data)
 486 {
 487         int err;
 488         struct perf_session *psession = data;
 489
 490         if (machine__is_host(machine))
 491                 return;
 492
 493         /*
 494          *As for guest kernel when processing subcommand record&report,
 495          *we arrange module mmap prior to guest kernel mmap and trigger
 496          *a preload dso because default guest module symbols are loaded
 497          *from guest kallsyms instead of /lib/modules/XXX/XXX. This
 498          *method is used to avoid symbol missing when the first addr is
 499          *in module instead of in guest kernel.
 500          */
 501         err = event__synthesize_modules(process_synthesized_event,
 502                                         psession, machine);
 503         if (err < 0)
 504                 pr_err("Couldn't record guest kernel [%d]'s reference"
 505                        " relocation symbol.\n", machine->pid);
 506
 507         /*
 508          * We use _stext for guest kernel because guest kernel's /proc/kallsyms
 509          * have no _text sometimes.
 510          */
 511         err = event__synthesize_kernel_mmap(process_synthesized_event,
 512                                             psession, machine, "_text");
 513         if (err < 0)
 514                 err = event__synthesize_kernel_mmap(process_synthesized_event,
 515                                                     psession, machine, "_stext");
 516         if (err < 0)
 517                 pr_err("Couldn't record guest kernel [%d]'s reference"
 518                        " relocation symbol.\n", machine->pid);
 519 }
 520
 521 static struct perf_event_header finished_round_event = {
 522         .size = sizeof(struct perf_event_header),
 523         .type = PERF_RECORD_FINISHED_ROUND,
 524 };
 525
 526 static void mmap_read_all(void)
 527 {
 528         int i;
 529
 530         for (i = 0; i < nr_cpu; i++) {
 531                 if (mmap_array[i].base)
 532                         mmap_read(&mmap_array[i]);
 533         }
 534
 535         if (perf_header__has_feat(&session->header, HEADER_TRACE_INFO))
 536                 write_output(&finished_round_event, sizeof(finished_round_event));
 537 }
 538
 539 static int __cmd_record(int argc, const char **argv)
 540 {
 541         int i;
 542         struct stat st;
 543         int flags;
 544         int err;
 545         unsigned long waking = 0;
 546         int child_ready_pipe[2], go_pipe[2];
 547         const bool forks = argc > 0;
 548         char buf;
 549         struct machine *machine;
 550
 551         page_size = sysconf(_SC_PAGE_SIZE);
 552
 553         atexit(sig_atexit);
 554         signal(SIGCHLD, sig_handler);
 555         signal(SIGINT, sig_handler);
 556         signal(SIGUSR1, sig_handler);
 557
 558         if (forks && (pipe(child_ready_pipe) < 0 || pipe(go_pipe) < 0)) {
 559                 perror("failed to create pipes");
 560                 exit(-1);
 561         }
 562
 563         if (!strcmp(output_name, "-"))
 564                 pipe_output = 1;
 565         else if (!stat(output_name, &st) && st.st_size) {
 566                 if (write_mode == WRITE_FORCE) {
 567                         char oldname[PATH_MAX];
 568                         snprintf(oldname, sizeof(oldname), "%s.old",
 569                                  output_name);
 570                         unlink(oldname);
 571                         rename(output_name, oldname);
 572                 }
 573         } else if (write_mode == WRITE_APPEND) {
 574                 write_mode = WRITE_FORCE;
 575         }
 576
 577         flags = O_CREAT|O_RDWR;
 578         if (write_mode == WRITE_APPEND)
 579                 file_new = 0;
 580         else
 581                 flags |= O_TRUNC;
 582
 583         if (pipe_output)
 584                 output = STDOUT_FILENO;
 585         else
 586                 output = open(output_name, flags, S_IRUSR | S_IWUSR);
 587         if (output < 0) {
 588                 perror("failed to create output file");
 589                 exit(-1);
 590         }
 591
 592         session = perf_session__new(output_name, O_WRONLY,
 593                                     write_mode == WRITE_FORCE, false, NULL);
 594         if (session == NULL) {
 595                 pr_err("Not enough memory for reading perf file header\n");
 596                 return -1;
 597         }
 598
 599         if (!no_buildid)
 600                 perf_header__set_feat(&session->header, HEADER_BUILD_ID);
 601
 602         if (!file_new) {
 603                 err = perf_header__read(session, output);
 604                 if (err < 0)
 605                         goto out_delete_session;
 606         }
 607
 608         if (have_tracepoints(&evsel_list))
 609                 perf_header__set_feat(&session->header, HEADER_TRACE_INFO);
 610
 611         /*
 612          * perf_session__delete(session) will be called at atexit_header()
 613          */
 614         atexit(atexit_header);
 615
 616         if (forks) {
 617                 child_pid = fork();
 618                 if (child_pid < 0) {
 619                         perror("failed to fork");
 620                         exit(-1);
 621                 }
 622
 623                 if (!child_pid) {
 624                         if (pipe_output)
 625                                 dup2(2, 1);
 626                         close(child_ready_pipe[0]);
 627                         close(go_pipe[1]);
 628                         fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
 629
 630                         /*
 631                          * Do a dummy execvp to get the PLT entry resolved,
 632                          * so we avoid the resolver overhead on the real
 633                          * execvp call.
 634                          */
 635                         execvp("", (char **)argv);
 636
 637                         /*
 638                          * Tell the parent we're ready to go
 639                          */
 640                         close(child_ready_pipe[1]);
 641
 642                         /*
 643                          * Wait until the parent tells us to go.
 644                          */
 645                         if (read(go_pipe[0], &buf, 1) == -1)
 646                                 perror("unable to read pipe");
 647
 648                         execvp(argv[0], (char **)argv);
 649
 650                         perror(argv[0]);
 651                         kill(getppid(), SIGUSR1);
 652                         exit(-1);
 653                 }
 654
 655                 if (!system_wide && target_tid == -1 && target_pid == -1)
 656                         threads->map[0] = child_pid;
 657
 658                 close(child_ready_pipe[1]);
 659                 close(go_pipe[0]);
 660                 /*
 661                  * wait for child to settle
 662                  */
 663                 if (read(child_ready_pipe[0], &buf, 1) == -1) {
 664                         perror("unable to read pipe");
 665                         exit(-1);
 666                 }
 667                 close(child_ready_pipe[0]);
 668         }
 669
 670         if (!system_wide && no_inherit && !cpu_list) {
 671                 open_counters(-1);
 672         } else {
 673                 for (i = 0; i < cpus->nr; i++)
 674                         open_counters(cpus->map[i]);
 675         }
 676
 677         perf_session__set_sample_type(session, sample_type);
 678
 679         if (pipe_output) {
 680                 err = perf_header__write_pipe(output);
 681                 if (err < 0)
 682                         return err;
 683         } else if (file_new) {
 684                 err = perf_header__write(&session->header, output, false);
 685                 if (err < 0)
 686                         return err;
 687         }
 688
 689         post_processing_offset = lseek(output, 0, SEEK_CUR);
 690
 691         perf_session__set_sample_id_all(session, sample_id_all_avail);
 692
 693         if (pipe_output) {
 694                 err = event__synthesize_attrs(&session->header,
 695                                               process_synthesized_event,
 696                                               session);
 697                 if (err < 0) {
 698                         pr_err("Couldn't synthesize attrs.\n");
 699                         return err;
 700                 }
 701
 702                 err = event__synthesize_event_types(process_synthesized_event,
 703                                                     session);
 704                 if (err < 0) {
 705                         pr_err("Couldn't synthesize event_types.\n");
 706                         return err;
 707                 }
 708
 709                 if (have_tracepoints(&evsel_list)) {
 710                         /*
 711                          * FIXME err <= 0 here actually means that
 712                          * there were no tracepoints so its not really
 713                          * an error, just that we don't need to
 714                          * synthesize anything.  We really have to
 715                          * return this more properly and also
 716                          * propagate errors that now are calling die()
 717                          */
 718                         err = event__synthesize_tracing_data(output, &evsel_list,
 719                                                              process_synthesized_event,
 720                                                              session);
 721                         if (err <= 0) {
 722                                 pr_err("Couldn't record tracing data.\n");
 723                                 return err;
 724                         }
 725                         advance_output(err);
 726                 }
 727         }
 728
 729         machine = perf_session__find_host_machine(session);
 730         if (!machine) {
 731                 pr_err("Couldn't find native kernel information.\n");
 732                 return -1;
 733         }
 734
 735         err = event__synthesize_kernel_mmap(process_synthesized_event,
 736                                             session, machine, "_text");
 737         if (err < 0)
 738                 err = event__synthesize_kernel_mmap(process_synthesized_event,
 739                                                     session, machine, "_stext");
 740         if (err < 0)
 741                 pr_err("Couldn't record kernel reference relocation symbol\n"
 742                        "Symbol resolution may be skewed if relocation was used (e.g. kexec).\n"
 743                        "Check /proc/kallsyms permission or run as root.\n");
 744
 745         err = event__synthesize_modules(process_synthesized_event,
 746                                         session, machine);
 747         if (err < 0)
 748                 pr_err("Couldn't record kernel module information.\n"
 749                        "Symbol resolution may be skewed if relocation was used (e.g. kexec).\n"
 750                        "Check /proc/modules permission or run as root.\n");
 751
 752         if (perf_guest)
 753                 perf_session__process_machines(session, event__synthesize_guest_os);
 754
 755         if (!system_wide)
 756                 event__synthesize_thread(target_tid, process_synthesized_event,
 757                                          session);
 758         else
 759                 event__synthesize_threads(process_synthesized_event, session);
 760
 761         if (realtime_prio) {
 762                 struct sched_param param;
 763
 764                 param.sched_priority = realtime_prio;
 765                 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
 766                         pr_err("Could not set realtime priority.\n");
 767                         exit(-1);
 768                 }
 769         }
 770
 771         /*
 772          * Let the child rip
 773          */
 774         if (forks)
 775                 close(go_pipe[1]);
 776
 777         for (;;) {
 778                 int hits = samples;
 779                 int thread;
 780
 781                 mmap_read_all();
 782
 783                 if (hits == samples) {
 784                         if (done)
 785                                 break;
 786                         err = poll(event_array, nr_poll, -1);
 787                         waking++;
 788                 }
 789
 790                 if (done) {
 791                         for (i = 0; i < nr_cpu; i++) {
 792                                 struct perf_evsel *pos;
 793
 794                                 list_for_each_entry(pos, &evsel_list, node) {
 795                                         for (thread = 0;
 796                                                 thread < threads->nr;
 797                                                 thread++)
 798                                                 ioctl(FD(pos, i, thread),
 799                                                         PERF_EVENT_IOC_DISABLE);
 800                                 }
 801                         }
 802                 }
 803         }
 804
 805         if (quiet || signr == SIGUSR1)
 806                 return 0;
 807
 808         fprintf(stderr, "[ perf record: Woken up %ld times to write data ]\n", waking);
 809
 810         /*
 811          * Approximate RIP event size: 24 bytes.
 812          */
 813         fprintf(stderr,
 814                 "[ perf record: Captured and wrote %.3f MB %s (~%lld samples) ]\n",
 815                 (double)bytes_written / 1024.0 / 1024.0,
 816                 output_name,
 817                 bytes_written / 24);
 818
 819         return 0;
 820
 821 out_delete_session:
 822         perf_session__delete(session);
 823         return err;
 824 }
 825
 826 static const char * const record_usage[] = {
 827         "perf record [<options>] [<command>]",
 828         "perf record [<options>] -- <command> [<options>]",
 829         NULL
 830 };
 831
 832 static bool force, append_file;
 833
 834 const struct option record_options[] = {
 835         OPT_CALLBACK('e', "event", NULL, "event",
 836                      "event selector. use 'perf list' to list available events",
 837                      parse_events),
 838         OPT_CALLBACK(0, "filter", NULL, "filter",
 839                      "event filter", parse_filter),
 840         OPT_INTEGER('p', "pid", &target_pid,
 841                     "record events on existing process id"),
 842         OPT_INTEGER('t', "tid", &target_tid,
 843                     "record events on existing thread id"),
 844         OPT_INTEGER('r', "realtime", &realtime_prio,
 845                     "collect data with this RT SCHED_FIFO priority"),
 846         OPT_BOOLEAN('R', "raw-samples", &raw_samples,
 847                     "collect raw sample records from all opened counters"),
 848         OPT_BOOLEAN('a', "all-cpus", &system_wide,
 849                             "system-wide collection from all CPUs"),
 850         OPT_BOOLEAN('A', "append", &append_file,
 851                             "append to the output file to do incremental profiling"),
 852         OPT_STRING('C', "cpu", &cpu_list, "cpu",
 853                     "list of cpus to monitor"),
 854         OPT_BOOLEAN('f', "force", &force,
 855                         "overwrite existing data file (deprecated)"),
 856         OPT_U64('c', "count", &user_interval, "event period to sample"),
 857         OPT_STRING('o', "output", &output_name, "file",
 858                     "output file name"),
 859         OPT_BOOLEAN('i', "no-inherit", &no_inherit,
 860                     "child tasks do not inherit counters"),
 861         OPT_UINTEGER('F', "freq", &user_freq, "profile at this frequency"),
 862         OPT_UINTEGER('m', "mmap-pages", &mmap_pages, "number of mmap data pages"),
 863         OPT_BOOLEAN('g', "call-graph", &call_graph,
 864                     "do call-graph (stack chain/backtrace) recording"),
 865         OPT_INCR('v', "verbose", &verbose,
 866                     "be more verbose (show counter open errors, etc)"),
 867         OPT_BOOLEAN('q', "quiet", &quiet, "don't print any message"),
 868         OPT_BOOLEAN('s', "stat", &inherit_stat,
 869                     "per thread counts"),
 870         OPT_BOOLEAN('d', "data", &sample_address,
 871                     "Sample addresses"),
 872         OPT_BOOLEAN('T', "timestamp", &sample_time, "Sample timestamps"),
 873         OPT_BOOLEAN('n', "no-samples", &no_samples,
 874                     "don't sample"),
 875         OPT_BOOLEAN('N', "no-buildid-cache", &no_buildid_cache,
 876                     "do not update the buildid cache"),
 877         OPT_BOOLEAN('B', "no-buildid", &no_buildid,
 878                     "do not collect buildids in perf.data"),
 879         OPT_END()
 880 };
 881
 882 int cmd_record(int argc, const char **argv, const char *prefix __used)
 883 {
 884         int err = -ENOMEM;
 885         struct perf_evsel *pos;
 886
 887         argc = parse_options(argc, argv, record_options, record_usage,
 888                             PARSE_OPT_STOP_AT_NON_OPTION);
 889         if (!argc && target_pid == -1 && target_tid == -1 &&
 890                 !system_wide && !cpu_list)
 891                 usage_with_options(record_usage, record_options);
 892
 893         if (force && append_file) {
 894                 fprintf(stderr, "Can't overwrite and append at the same time."
 895                                 " You need to choose between -f and -A");
 896                 usage_with_options(record_usage, record_options);
 897         } else if (append_file) {
 898                 write_mode = WRITE_APPEND;
 899         } else {
 900                 write_mode = WRITE_FORCE;
 901         }
 902
 903         symbol__init();
 904
 905         if (no_buildid_cache || no_buildid)
 906                 disable_buildid_cache();
 907
 908         if (list_empty(&evsel_list) && perf_evsel_list__create_default() < 0) {
 909                 pr_err("Not enough memory for event selector list\n");
 910                 goto out_symbol_exit;
 911         }
 912
 913         if (target_pid != -1)
 914                 target_tid = target_pid;
 915
 916         threads = thread_map__new(target_pid, target_tid);
 917         if (threads == NULL) {
 918                 pr_err("Problems finding threads of monitor\n");
 919                 usage_with_options(record_usage, record_options);
 920         }
 921
 922         cpus = cpu_map__new(cpu_list);
 923         if (cpus == NULL) {
 924                 perror("failed to parse CPUs map");
 925                 return -1;
 926         }
 927
 928         list_for_each_entry(pos, &evsel_list, node) {
 929                 if (perf_evsel__alloc_fd(pos, cpus->nr, threads->nr) < 0)
 930                         goto out_free_fd;
 931         }
 932         event_array = malloc((sizeof(struct pollfd) * MAX_NR_CPUS *
 933                               MAX_COUNTERS * threads->nr));
 934         if (!event_array)
 935                 goto out_free_fd;
 936
 937         if (user_interval != ULLONG_MAX)
 938                 default_interval = user_interval;
 939         if (user_freq != UINT_MAX)
 940                 freq = user_freq;
 941
 942         /*
 943          * User specified count overrides default frequency.
 944          */
 945         if (default_interval)
 946                 freq = 0;
 947         else if (freq) {
 948                 default_interval = freq;
 949         } else {
 950                 fprintf(stderr, "frequency and count are zero, aborting\n");
 951                 err = -EINVAL;
 952                 goto out_free_event_array;
 953         }
 954
 955         err = __cmd_record(argc, argv);
 956
 957 out_free_event_array:
 958         free(event_array);
 959 out_free_fd:
 960         thread_map__delete(threads);
 961         threads = NULL;
 962 out_symbol_exit:
 963         symbol__exit();
 964         return err;
 965 }