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
  65 static long                     samples                         =      0;
  66 static u64                      bytes_written                   =      0;
  67
  68 static struct pollfd            *event_array;
  69
  70 static int                      nr_poll                         =      0;
  71 static int                      nr_cpu                          =      0;
  72
  73 static int                      file_new                        =      1;
  74 static off_t                    post_processing_offset;
  75
  76 static struct perf_session      *session;
  77
  78 struct mmap_data {
  79         int                     counter;
  80         void                    *base;
  81         unsigned int            mask;
  82         unsigned int            prev;
  83 };
  84
  85 static struct mmap_data         *mmap_array[MAX_NR_CPUS][MAX_COUNTERS];
  86
  87 static unsigned long mmap_read_head(struct mmap_data *md)
  88 {
  89         struct perf_event_mmap_page *pc = md->base;
  90         long head;
  91
  92         head = pc->data_head;
  93         rmb();
  94
  95         return head;
  96 }
  97
  98 static void mmap_write_tail(struct mmap_data *md, unsigned long tail)
  99 {
 100         struct perf_event_mmap_page *pc = md->base;
 101
 102         /*
 103          * ensure all reads are done before we write the tail out.
 104          */
 105         /* mb(); */
 106         pc->data_tail = tail;
 107 }
 108
 109 static void advance_output(size_t size)
 110 {
 111         bytes_written += size;
 112 }
 113
 114 static void write_output(void *buf, size_t size)
 115 {
 116         while (size) {
 117                 int ret = write(output, buf, size);
 118
 119                 if (ret < 0)
 120                         die("failed to write");
 121
 122                 size -= ret;
 123                 buf += ret;
 124
 125                 bytes_written += ret;
 126         }
 127 }
 128
 129 static int process_synthesized_event(event_t *event,
 130                                      struct perf_session *self __used)
 131 {
 132         write_output(event, event->header.size);
 133         return 0;
 134 }
 135
 136 static void mmap_read(struct mmap_data *md)
 137 {
 138         unsigned int head = mmap_read_head(md);
 139         unsigned int old = md->prev;
 140         unsigned char *data = md->base + page_size;
 141         unsigned long size;
 142         void *buf;
 143         int diff;
 144
 145         /*
 146          * If we're further behind than half the buffer, there's a chance
 147          * the writer will bite our tail and mess up the samples under us.
 148          *
 149          * If we somehow ended up ahead of the head, we got messed up.
 150          *
 151          * In either case, truncate and restart at head.
 152          */
 153         diff = head - old;
 154         if (diff < 0) {
 155                 fprintf(stderr, "WARNING: failed to keep up with mmap data\n");
 156                 /*
 157                  * head points to a known good entry, start there.
 158                  */
 159                 old = head;
 160         }
 161
 162         if (old != head)
 163                 samples++;
 164
 165         size = head - old;
 166
 167         if ((old & md->mask) + size != (head & md->mask)) {
 168                 buf = &data[old & md->mask];
 169                 size = md->mask + 1 - (old & md->mask);
 170                 old += size;
 171
 172                 write_output(buf, size);
 173         }
 174
 175         buf = &data[old & md->mask];
 176         size = head - old;
 177         old += size;
 178
 179         write_output(buf, size);
 180
 181         md->prev = old;
 182         mmap_write_tail(md, old);
 183 }
 184
 185 static volatile int done = 0;
 186 static volatile int signr = -1;
 187
 188 static void sig_handler(int sig)
 189 {
 190         done = 1;
 191         signr = sig;
 192 }
 193
 194 static void sig_atexit(void)
 195 {
 196         if (child_pid != -1)
 197                 kill(child_pid, SIGTERM);
 198
 199         if (signr == -1)
 200                 return;
 201
 202         signal(signr, SIG_DFL);
 203         kill(getpid(), signr);
 204 }
 205
 206 static int group_fd;
 207
 208 static struct perf_header_attr *get_header_attr(struct perf_event_attr *a, int nr)
 209 {
 210         struct perf_header_attr *h_attr;
 211
 212         if (nr < session->header.attrs) {
 213                 h_attr = session->header.attr[nr];
 214         } else {
 215                 h_attr = perf_header_attr__new(a);
 216                 if (h_attr != NULL)
 217                         if (perf_header__add_attr(&session->header, h_attr) < 0) {
 218                                 perf_header_attr__delete(h_attr);
 219                                 h_attr = NULL;
 220                         }
 221         }
 222
 223         return h_attr;
 224 }
 225
 226 static void create_counter(int counter, int cpu)
 227 {
 228         char *filter = filters[counter];
 229         struct perf_event_attr *attr = attrs + counter;
 230         struct perf_header_attr *h_attr;
 231         int track = !counter; /* only the first counter needs these */
 232         int thread_index;
 233         int ret;
 234         struct {
 235                 u64 count;
 236                 u64 time_enabled;
 237                 u64 time_running;
 238                 u64 id;
 239         } read_data;
 240
 241         attr->read_format       = PERF_FORMAT_TOTAL_TIME_ENABLED |
 242                                   PERF_FORMAT_TOTAL_TIME_RUNNING |
 243                                   PERF_FORMAT_ID;
 244
 245         attr->sample_type       |= PERF_SAMPLE_IP | PERF_SAMPLE_TID;
 246
 247         if (nr_counters > 1)
 248                 attr->sample_type |= PERF_SAMPLE_ID;
 249
 250         /*
 251          * We default some events to a 1 default interval. But keep
 252          * it a weak assumption overridable by the user.
 253          */
 254         if (!attr->sample_period || (user_freq != UINT_MAX &&
 255                                      user_interval != ULLONG_MAX)) {
 256                 if (freq) {
 257                         attr->sample_type       |= PERF_SAMPLE_PERIOD;
 258                         attr->freq              = 1;
 259                         attr->sample_freq       = freq;
 260                 } else {
 261                         attr->sample_period = default_interval;
 262                 }
 263         }
 264
 265         if (no_samples)
 266                 attr->sample_freq = 0;
 267
 268         if (inherit_stat)
 269                 attr->inherit_stat = 1;
 270
 271         if (sample_address)
 272                 attr->sample_type       |= PERF_SAMPLE_ADDR;
 273
 274         if (call_graph)
 275                 attr->sample_type       |= PERF_SAMPLE_CALLCHAIN;
 276
 277         if (raw_samples) {
 278                 attr->sample_type       |= PERF_SAMPLE_TIME;
 279                 attr->sample_type       |= PERF_SAMPLE_RAW;
 280                 attr->sample_type       |= PERF_SAMPLE_CPU;
 281         }
 282
 283         attr->mmap              = track;
 284         attr->comm              = track;
 285         attr->inherit           = !no_inherit;
 286         if (target_pid == -1 && target_tid == -1 && !system_wide) {
 287                 attr->disabled = 1;
 288                 attr->enable_on_exec = 1;
 289         }
 290
 291         for (thread_index = 0; thread_index < thread_num; thread_index++) {
 292 try_again:
 293                 fd[nr_cpu][counter][thread_index] = sys_perf_event_open(attr,
 294                                 all_tids[thread_index], cpu, group_fd, 0);
 295
 296                 if (fd[nr_cpu][counter][thread_index] < 0) {
 297                         int err = errno;
 298
 299                         if (err == EPERM || err == EACCES)
 300                                 die("Permission error - are you root?\n"
 301                                         "\t Consider tweaking"
 302                                         " /proc/sys/kernel/perf_event_paranoid.\n");
 303                         else if (err ==  ENODEV && profile_cpu != -1) {
 304                                 die("No such device - did you specify"
 305                                         " an out-of-range profile CPU?\n");
 306                         }
 307
 308                         /*
 309                          * If it's cycles then fall back to hrtimer
 310                          * based cpu-clock-tick sw counter, which
 311                          * is always available even if no PMU support:
 312                          */
 313                         if (attr->type == PERF_TYPE_HARDWARE
 314                                         && attr->config == PERF_COUNT_HW_CPU_CYCLES) {
 315
 316                                 if (verbose)
 317                                         warning(" ... trying to fall back to cpu-clock-ticks\n");
 318                                 attr->type = PERF_TYPE_SOFTWARE;
 319                                 attr->config = PERF_COUNT_SW_CPU_CLOCK;
 320                                 goto try_again;
 321                         }
 322                         printf("\n");
 323                         error("perfcounter syscall returned with %d (%s)\n",
 324                                         fd[nr_cpu][counter][thread_index], strerror(err));
 325
 326 #if defined(__i386__) || defined(__x86_64__)
 327                         if (attr->type == PERF_TYPE_HARDWARE && err == EOPNOTSUPP)
 328                                 die("No hardware sampling interrupt available."
 329                                     " No APIC? If so then you can boot the kernel"
 330                                     " with the \"lapic\" boot parameter to"
 331                                     " force-enable it.\n");
 332 #endif
 333
 334                         die("No CONFIG_PERF_EVENTS=y kernel support configured?\n");
 335                         exit(-1);
 336                 }
 337
 338                 h_attr = get_header_attr(attr, counter);
 339                 if (h_attr == NULL)
 340                         die("nomem\n");
 341
 342                 if (!file_new) {
 343                         if (memcmp(&h_attr->attr, attr, sizeof(*attr))) {
 344                                 fprintf(stderr, "incompatible append\n");
 345                                 exit(-1);
 346                         }
 347                 }
 348
 349                 if (read(fd[nr_cpu][counter][thread_index], &read_data, sizeof(read_data)) == -1) {
 350                         perror("Unable to read perf file descriptor\n");
 351                         exit(-1);
 352                 }
 353
 354                 if (perf_header_attr__add_id(h_attr, read_data.id) < 0) {
 355                         pr_warning("Not enough memory to add id\n");
 356                         exit(-1);
 357                 }
 358
 359                 assert(fd[nr_cpu][counter][thread_index] >= 0);
 360                 fcntl(fd[nr_cpu][counter][thread_index], F_SETFL, O_NONBLOCK);
 361
 362                 /*
 363                  * First counter acts as the group leader:
 364                  */
 365                 if (group && group_fd == -1)
 366                         group_fd = fd[nr_cpu][counter][thread_index];
 367
 368                 event_array[nr_poll].fd = fd[nr_cpu][counter][thread_index];
 369                 event_array[nr_poll].events = POLLIN;
 370                 nr_poll++;
 371
 372                 mmap_array[nr_cpu][counter][thread_index].counter = counter;
 373                 mmap_array[nr_cpu][counter][thread_index].prev = 0;
 374                 mmap_array[nr_cpu][counter][thread_index].mask = mmap_pages*page_size - 1;
 375                 mmap_array[nr_cpu][counter][thread_index].base = mmap(NULL, (mmap_pages+1)*page_size,
 376                         PROT_READ|PROT_WRITE, MAP_SHARED, fd[nr_cpu][counter][thread_index], 0);
 377                 if (mmap_array[nr_cpu][counter][thread_index].base == MAP_FAILED) {
 378                         error("failed to mmap with %d (%s)\n", errno, strerror(errno));
 379                         exit(-1);
 380                 }
 381
 382                 if (filter != NULL) {
 383                         ret = ioctl(fd[nr_cpu][counter][thread_index],
 384                                         PERF_EVENT_IOC_SET_FILTER, filter);
 385                         if (ret) {
 386                                 error("failed to set filter with %d (%s)\n", errno,
 387                                                 strerror(errno));
 388                                 exit(-1);
 389                         }
 390                 }
 391         }
 392 }
 393
 394 static void open_counters(int cpu)
 395 {
 396         int counter;
 397
 398         group_fd = -1;
 399         for (counter = 0; counter < nr_counters; counter++)
 400                 create_counter(counter, cpu);
 401
 402         nr_cpu++;
 403 }
 404
 405 static int process_buildids(void)
 406 {
 407         u64 size = lseek(output, 0, SEEK_CUR);
 408
 409         if (size == 0)
 410                 return 0;
 411
 412         session->fd = output;
 413         return __perf_session__process_events(session, post_processing_offset,
 414                                               size - post_processing_offset,
 415                                               size, &build_id__mark_dso_hit_ops);
 416 }
 417
 418 static void atexit_header(void)
 419 {
 420         if (!pipe_output) {
 421                 session->header.data_size += bytes_written;
 422
 423                 process_buildids();
 424                 perf_header__write(&session->header, output, true);
 425         }
 426 }
 427
 428 static void event__synthesize_guest_os(struct machine *machine, void *data)
 429 {
 430         int err;
 431         char *guest_kallsyms;
 432         char path[PATH_MAX];
 433         struct perf_session *psession = data;
 434
 435         if (machine__is_host(machine))
 436                 return;
 437
 438         /*
 439          *As for guest kernel when processing subcommand record&report,
 440          *we arrange module mmap prior to guest kernel mmap and trigger
 441          *a preload dso because default guest module symbols are loaded
 442          *from guest kallsyms instead of /lib/modules/XXX/XXX. This
 443          *method is used to avoid symbol missing when the first addr is
 444          *in module instead of in guest kernel.
 445          */
 446         err = event__synthesize_modules(process_synthesized_event,
 447                                         psession, machine);
 448         if (err < 0)
 449                 pr_err("Couldn't record guest kernel [%d]'s reference"
 450                        " relocation symbol.\n", machine->pid);
 451
 452         if (machine__is_default_guest(machine))
 453                 guest_kallsyms = (char *) symbol_conf.default_guest_kallsyms;
 454         else {
 455                 sprintf(path, "%s/proc/kallsyms", machine->root_dir);
 456                 guest_kallsyms = path;
 457         }
 458
 459         /*
 460          * We use _stext for guest kernel because guest kernel's /proc/kallsyms
 461          * have no _text sometimes.
 462          */
 463         err = event__synthesize_kernel_mmap(process_synthesized_event,
 464                                             psession, machine, "_text");
 465         if (err < 0)
 466                 err = event__synthesize_kernel_mmap(process_synthesized_event,
 467                                                     psession, machine, "_stext");
 468         if (err < 0)
 469                 pr_err("Couldn't record guest kernel [%d]'s reference"
 470                        " relocation symbol.\n", machine->pid);
 471 }
 472
 473 static struct perf_event_header finished_round_event = {
 474         .size = sizeof(struct perf_event_header),
 475         .type = PERF_RECORD_FINISHED_ROUND,
 476 };
 477
 478 static void mmap_read_all(void)
 479 {
 480         int i, counter, thread;
 481
 482         for (i = 0; i < nr_cpu; i++) {
 483                 for (counter = 0; counter < nr_counters; counter++) {
 484                         for (thread = 0; thread < thread_num; thread++) {
 485                                 if (mmap_array[i][counter][thread].base)
 486                                         mmap_read(&mmap_array[i][counter][thread]);
 487                         }
 488
 489                 }
 490         }
 491
 492         if (perf_header__has_feat(&session->header, HEADER_TRACE_INFO))
 493                 write_output(&finished_round_event, sizeof(finished_round_event));
 494 }
 495
 496 static int __cmd_record(int argc, const char **argv)
 497 {
 498         int i, counter;
 499         struct stat st;
 500         pid_t pid = 0;
 501         int flags;
 502         int err;
 503         unsigned long waking = 0;
 504         int child_ready_pipe[2], go_pipe[2];
 505         const bool forks = argc > 0;
 506         char buf;
 507         struct machine *machine;
 508
 509         page_size = sysconf(_SC_PAGE_SIZE);
 510
 511         atexit(sig_atexit);
 512         signal(SIGCHLD, sig_handler);
 513         signal(SIGINT, sig_handler);
 514
 515         if (forks && (pipe(child_ready_pipe) < 0 || pipe(go_pipe) < 0)) {
 516                 perror("failed to create pipes");
 517                 exit(-1);
 518         }
 519
 520         if (!strcmp(output_name, "-"))
 521                 pipe_output = 1;
 522         else if (!stat(output_name, &st) && st.st_size) {
 523                 if (write_mode == WRITE_FORCE) {
 524                         char oldname[PATH_MAX];
 525                         snprintf(oldname, sizeof(oldname), "%s.old",
 526                                  output_name);
 527                         unlink(oldname);
 528                         rename(output_name, oldname);
 529                 }
 530         } else if (write_mode == WRITE_APPEND) {
 531                 write_mode = WRITE_FORCE;
 532         }
 533
 534         flags = O_CREAT|O_RDWR;
 535         if (write_mode == WRITE_APPEND)
 536                 file_new = 0;
 537         else
 538                 flags |= O_TRUNC;
 539
 540         if (pipe_output)
 541                 output = STDOUT_FILENO;
 542         else
 543                 output = open(output_name, flags, S_IRUSR | S_IWUSR);
 544         if (output < 0) {
 545                 perror("failed to create output file");
 546                 exit(-1);
 547         }
 548
 549         session = perf_session__new(output_name, O_WRONLY,
 550                                     write_mode == WRITE_FORCE, false);
 551         if (session == NULL) {
 552                 pr_err("Not enough memory for reading perf file header\n");
 553                 return -1;
 554         }
 555
 556         if (!file_new) {
 557                 err = perf_header__read(session, output);
 558                 if (err < 0)
 559                         return err;
 560         }
 561
 562         if (have_tracepoints(attrs, nr_counters))
 563                 perf_header__set_feat(&session->header, HEADER_TRACE_INFO);
 564
 565         atexit(atexit_header);
 566
 567         if (forks) {
 568                 child_pid = fork();
 569                 if (pid < 0) {
 570                         perror("failed to fork");
 571                         exit(-1);
 572                 }
 573
 574                 if (!child_pid) {
 575                         if (pipe_output)
 576                                 dup2(2, 1);
 577                         close(child_ready_pipe[0]);
 578                         close(go_pipe[1]);
 579                         fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
 580
 581                         /*
 582                          * Do a dummy execvp to get the PLT entry resolved,
 583                          * so we avoid the resolver overhead on the real
 584                          * execvp call.
 585                          */
 586                         execvp("", (char **)argv);
 587
 588                         /*
 589                          * Tell the parent we're ready to go
 590                          */
 591                         close(child_ready_pipe[1]);
 592
 593                         /*
 594                          * Wait until the parent tells us to go.
 595                          */
 596                         if (read(go_pipe[0], &buf, 1) == -1)
 597                                 perror("unable to read pipe");
 598
 599                         execvp(argv[0], (char **)argv);
 600
 601                         perror(argv[0]);
 602                         exit(-1);
 603                 }
 604
 605                 if (!system_wide && target_tid == -1 && target_pid == -1)
 606                         all_tids[0] = child_pid;
 607
 608                 close(child_ready_pipe[1]);
 609                 close(go_pipe[0]);
 610                 /*
 611                  * wait for child to settle
 612                  */
 613                 if (read(child_ready_pipe[0], &buf, 1) == -1) {
 614                         perror("unable to read pipe");
 615                         exit(-1);
 616                 }
 617                 close(child_ready_pipe[0]);
 618         }
 619
 620         if ((!system_wide && no_inherit) || profile_cpu != -1) {
 621                 open_counters(profile_cpu);
 622         } else {
 623                 nr_cpus = read_cpu_map();
 624                 for (i = 0; i < nr_cpus; i++)
 625                         open_counters(cpumap[i]);
 626         }
 627
 628         if (pipe_output) {
 629                 err = perf_header__write_pipe(output);
 630                 if (err < 0)
 631                         return err;
 632         } else if (file_new) {
 633                 err = perf_header__write(&session->header, output, false);
 634                 if (err < 0)
 635                         return err;
 636         }
 637
 638         post_processing_offset = lseek(output, 0, SEEK_CUR);
 639
 640         if (pipe_output) {
 641                 err = event__synthesize_attrs(&session->header,
 642                                               process_synthesized_event,
 643                                               session);
 644                 if (err < 0) {
 645                         pr_err("Couldn't synthesize attrs.\n");
 646                         return err;
 647                 }
 648
 649                 err = event__synthesize_event_types(process_synthesized_event,
 650                                                     session);
 651                 if (err < 0) {
 652                         pr_err("Couldn't synthesize event_types.\n");
 653                         return err;
 654                 }
 655
 656                 if (have_tracepoints(attrs, nr_counters)) {
 657                         /*
 658                          * FIXME err <= 0 here actually means that
 659                          * there were no tracepoints so its not really
 660                          * an error, just that we don't need to
 661                          * synthesize anything.  We really have to
 662                          * return this more properly and also
 663                          * propagate errors that now are calling die()
 664                          */
 665                         err = event__synthesize_tracing_data(output, attrs,
 666                                                              nr_counters,
 667                                                              process_synthesized_event,
 668                                                              session);
 669                         if (err <= 0) {
 670                                 pr_err("Couldn't record tracing data.\n");
 671                                 return err;
 672                         }
 673                         advance_output(err);
 674                 }
 675         }
 676
 677         machine = perf_session__find_host_machine(session);
 678         if (!machine) {
 679                 pr_err("Couldn't find native kernel information.\n");
 680                 return -1;
 681         }
 682
 683         err = event__synthesize_kernel_mmap(process_synthesized_event,
 684                                             session, machine, "_text");
 685         if (err < 0)
 686                 err = event__synthesize_kernel_mmap(process_synthesized_event,
 687                                                     session, machine, "_stext");
 688         if (err < 0) {
 689                 pr_err("Couldn't record kernel reference relocation symbol.\n");
 690                 return err;
 691         }
 692
 693         err = event__synthesize_modules(process_synthesized_event,
 694                                         session, machine);
 695         if (err < 0) {
 696                 pr_err("Couldn't record kernel reference relocation symbol.\n");
 697                 return err;
 698         }
 699         if (perf_guest)
 700                 perf_session__process_machines(session, event__synthesize_guest_os);
 701
 702         if (!system_wide && profile_cpu == -1)
 703                 event__synthesize_thread(target_tid, process_synthesized_event,
 704                                          session);
 705         else
 706                 event__synthesize_threads(process_synthesized_event, session);
 707
 708         if (realtime_prio) {
 709                 struct sched_param param;
 710
 711                 param.sched_priority = realtime_prio;
 712                 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
 713                         pr_err("Could not set realtime priority.\n");
 714                         exit(-1);
 715                 }
 716         }
 717
 718         /*
 719          * Let the child rip
 720          */
 721         if (forks)
 722                 close(go_pipe[1]);
 723
 724         for (;;) {
 725                 int hits = samples;
 726                 int thread;
 727
 728                 mmap_read_all();
 729
 730                 if (hits == samples) {
 731                         if (done)
 732                                 break;
 733                         err = poll(event_array, nr_poll, -1);
 734                         waking++;
 735                 }
 736
 737                 if (done) {
 738                         for (i = 0; i < nr_cpu; i++) {
 739                                 for (counter = 0;
 740                                         counter < nr_counters;
 741                                         counter++) {
 742                                         for (thread = 0;
 743                                                 thread < thread_num;
 744                                                 thread++)
 745                                                 ioctl(fd[i][counter][thread],
 746                                                         PERF_EVENT_IOC_DISABLE);
 747                                 }
 748                         }
 749                 }
 750         }
 751
 752         fprintf(stderr, "[ perf record: Woken up %ld times to write data ]\n", waking);
 753
 754         /*
 755          * Approximate RIP event size: 24 bytes.
 756          */
 757         fprintf(stderr,
 758                 "[ perf record: Captured and wrote %.3f MB %s (~%lld samples) ]\n",
 759                 (double)bytes_written / 1024.0 / 1024.0,
 760                 output_name,
 761                 bytes_written / 24);
 762
 763         return 0;
 764 }
 765
 766 static const char * const record_usage[] = {
 767         "perf record [<options>] [<command>]",
 768         "perf record [<options>] -- <command> [<options>]",
 769         NULL
 770 };
 771
 772 static bool force, append_file;
 773
 774 static const struct option options[] = {
 775         OPT_CALLBACK('e', "event", NULL, "event",
 776                      "event selector. use 'perf list' to list available events",
 777                      parse_events),
 778         OPT_CALLBACK(0, "filter", NULL, "filter",
 779                      "event filter", parse_filter),
 780         OPT_INTEGER('p', "pid", &target_pid,
 781                     "record events on existing process id"),
 782         OPT_INTEGER('t', "tid", &target_tid,
 783                     "record events on existing thread id"),
 784         OPT_INTEGER('r', "realtime", &realtime_prio,
 785                     "collect data with this RT SCHED_FIFO priority"),
 786         OPT_BOOLEAN('R', "raw-samples", &raw_samples,
 787                     "collect raw sample records from all opened counters"),
 788         OPT_BOOLEAN('a', "all-cpus", &system_wide,
 789                             "system-wide collection from all CPUs"),
 790         OPT_BOOLEAN('A', "append", &append_file,
 791                             "append to the output file to do incremental profiling"),
 792         OPT_INTEGER('C', "profile_cpu", &profile_cpu,
 793                             "CPU to profile on"),
 794         OPT_BOOLEAN('f', "force", &force,
 795                         "overwrite existing data file (deprecated)"),
 796         OPT_U64('c', "count", &user_interval, "event period to sample"),
 797         OPT_STRING('o', "output", &output_name, "file",
 798                     "output file name"),
 799         OPT_BOOLEAN('i', "no-inherit", &no_inherit,
 800                     "child tasks do not inherit counters"),
 801         OPT_UINTEGER('F', "freq", &user_freq, "profile at this frequency"),
 802         OPT_UINTEGER('m', "mmap-pages", &mmap_pages, "number of mmap data pages"),
 803         OPT_BOOLEAN('g', "call-graph", &call_graph,
 804                     "do call-graph (stack chain/backtrace) recording"),
 805         OPT_INCR('v', "verbose", &verbose,
 806                     "be more verbose (show counter open errors, etc)"),
 807         OPT_BOOLEAN('s', "stat", &inherit_stat,
 808                     "per thread counts"),
 809         OPT_BOOLEAN('d', "data", &sample_address,
 810                     "Sample addresses"),
 811         OPT_BOOLEAN('n', "no-samples", &no_samples,
 812                     "don't sample"),
 813         OPT_END()
 814 };
 815
 816 int cmd_record(int argc, const char **argv, const char *prefix __used)
 817 {
 818         int i,j;
 819
 820         argc = parse_options(argc, argv, options, record_usage,
 821                             PARSE_OPT_STOP_AT_NON_OPTION);
 822         if (!argc && target_pid == -1 && target_tid == -1 &&
 823                 !system_wide && profile_cpu == -1)
 824                 usage_with_options(record_usage, options);
 825
 826         if (force && append_file) {
 827                 fprintf(stderr, "Can't overwrite and append at the same time."
 828                                 " You need to choose between -f and -A");
 829                 usage_with_options(record_usage, options);
 830         } else if (append_file) {
 831                 write_mode = WRITE_APPEND;
 832         } else {
 833                 write_mode = WRITE_FORCE;
 834         }
 835
 836         symbol__init();
 837
 838         if (!nr_counters) {
 839                 nr_counters     = 1;
 840                 attrs[0].type   = PERF_TYPE_HARDWARE;
 841                 attrs[0].config = PERF_COUNT_HW_CPU_CYCLES;
 842         }
 843
 844         if (target_pid != -1) {
 845                 target_tid = target_pid;
 846                 thread_num = find_all_tid(target_pid, &all_tids);
 847                 if (thread_num <= 0) {
 848                         fprintf(stderr, "Can't find all threads of pid %d\n",
 849                                         target_pid);
 850                         usage_with_options(record_usage, options);
 851                 }
 852         } else {
 853                 all_tids=malloc(sizeof(pid_t));
 854                 if (!all_tids)
 855                         return -ENOMEM;
 856
 857                 all_tids[0] = target_tid;
 858                 thread_num = 1;
 859         }
 860
 861         for (i = 0; i < MAX_NR_CPUS; i++) {
 862                 for (j = 0; j < MAX_COUNTERS; j++) {
 863                         fd[i][j] = malloc(sizeof(int)*thread_num);
 864                         mmap_array[i][j] = zalloc(
 865                                 sizeof(struct mmap_data)*thread_num);
 866                         if (!fd[i][j] || !mmap_array[i][j])
 867                                 return -ENOMEM;
 868                 }
 869         }
 870         event_array = malloc(
 871                 sizeof(struct pollfd)*MAX_NR_CPUS*MAX_COUNTERS*thread_num);
 872         if (!event_array)
 873                 return -ENOMEM;
 874
 875         if (user_interval != ULLONG_MAX)
 876                 default_interval = user_interval;
 877         if (user_freq != UINT_MAX)
 878                 freq = user_freq;
 879
 880         /*
 881          * User specified count overrides default frequency.
 882          */
 883         if (default_interval)
 884                 freq = 0;
 885         else if (freq) {
 886                 default_interval = freq;
 887         } else {
 888                 fprintf(stderr, "frequency and count are zero, aborting\n");
 889                 exit(EXIT_FAILURE);
 890         }
 891
 892         return __cmd_record(argc, argv);
 893 }