tools/perf/builtin-stat.c

   1 /*
   2  * builtin-stat.c
   3  *
   4  * Builtin stat command: Give a precise performance counters summary
   5  * overview about any workload, CPU or specific PID.
   6  *
   7  * Sample output:
   8
   9    $ perf stat ./hackbench 10
  10
  11   Time: 0.118
  12
  13   Performance counter stats for './hackbench 10':
  14
  15        1708.761321 task-clock                #   11.037 CPUs utilized
  16             41,190 context-switches          #    0.024 M/sec
  17              6,735 CPU-migrations            #    0.004 M/sec
  18             17,318 page-faults               #    0.010 M/sec
  19      5,205,202,243 cycles                    #    3.046 GHz
  20      3,856,436,920 stalled-cycles-frontend   #   74.09% frontend cycles idle
  21      1,600,790,871 stalled-cycles-backend    #   30.75% backend  cycles idle
  22      2,603,501,247 instructions              #    0.50  insns per cycle
  23                                              #    1.48  stalled cycles per insn
  24        484,357,498 branches                  #  283.455 M/sec
  25          6,388,934 branch-misses             #    1.32% of all branches
  26
  27         0.154822978  seconds time elapsed
  28
  29  *
  30  * Copyright (C) 2008-2011, Red Hat Inc, Ingo Molnar <mingo@redhat.com>
  31  *
  32  * Improvements and fixes by:
  33  *
  34  *   Arjan van de Ven <arjan@linux.intel.com>
  35  *   Yanmin Zhang <yanmin.zhang@intel.com>
  36  *   Wu Fengguang <fengguang.wu@intel.com>
  37  *   Mike Galbraith <efault@gmx.de>
  38  *   Paul Mackerras <paulus@samba.org>
  39  *   Jaswinder Singh Rajput <jaswinder@kernel.org>
  40  *
  41  * Released under the GPL v2. (and only v2, not any later version)
  42  */
  43
  44 #include "perf.h"
  45 #include "builtin.h"
  46 #include "util/cgroup.h"
  47 #include "util/util.h"
  48 #include <subcmd/parse-options.h>
  49 #include "util/parse-events.h"
  50 #include "util/pmu.h"
  51 #include "util/event.h"
  52 #include "util/evlist.h"
  53 #include "util/evsel.h"
  54 #include "util/debug.h"
  55 #include "util/color.h"
  56 #include "util/stat.h"
  57 #include "util/header.h"
  58 #include "util/cpumap.h"
  59 #include "util/thread.h"
  60 #include "util/thread_map.h"
  61 #include "util/counts.h"
  62 #include "util/session.h"
  63 #include "util/tool.h"
  64 #include "asm/bug.h"
  65
  66 #include <stdlib.h>
  67 #include <sys/prctl.h>
  68 #include <locale.h>
  69
  70 #define DEFAULT_SEPARATOR       " "
  71 #define CNTR_NOT_SUPPORTED      "<not supported>"
  72 #define CNTR_NOT_COUNTED        "<not counted>"
  73
  74 static void print_counters(struct timespec *ts, int argc, const char **argv);
  75
  76 /* Default events used for perf stat -T */
  77 static const char *transaction_attrs = {
  78         "task-clock,"
  79         "{"
  80         "instructions,"
  81         "cycles,"
  82         "cpu/cycles-t/,"
  83         "cpu/tx-start/,"
  84         "cpu/el-start/,"
  85         "cpu/cycles-ct/"
  86         "}"
  87 };
  88
  89 /* More limited version when the CPU does not have all events. */
  90 static const char * transaction_limited_attrs = {
  91         "task-clock,"
  92         "{"
  93         "instructions,"
  94         "cycles,"
  95         "cpu/cycles-t/,"
  96         "cpu/tx-start/"
  97         "}"
  98 };
  99
 100 static struct perf_evlist       *evsel_list;
 101
 102 static struct target target = {
 103         .uid    = UINT_MAX,
 104 };
 105
 106 typedef int (*aggr_get_id_t)(struct cpu_map *m, int cpu);
 107
 108 static int                      run_count                       =  1;
 109 static bool                     no_inherit                      = false;
 110 static volatile pid_t           child_pid                       = -1;
 111 static bool                     null_run                        =  false;
 112 static int                      detailed_run                    =  0;
 113 static bool                     transaction_run;
 114 static bool                     big_num                         =  true;
 115 static int                      big_num_opt                     =  -1;
 116 static const char               *csv_sep                        = NULL;
 117 static bool                     csv_output                      = false;
 118 static bool                     group                           = false;
 119 static const char               *pre_cmd                        = NULL;
 120 static const char               *post_cmd                       = NULL;
 121 static bool                     sync_run                        = false;
 122 static unsigned int             initial_delay                   = 0;
 123 static unsigned int             unit_width                      = 4; /* strlen("unit") */
 124 static bool                     forever                         = false;
 125 static bool                     metric_only                     = false;
 126 static struct timespec          ref_time;
 127 static struct cpu_map           *aggr_map;
 128 static aggr_get_id_t            aggr_get_id;
 129 static bool                     append_file;
 130 static const char               *output_name;
 131 static int                      output_fd;
 132
 133 struct perf_stat {
 134         bool                     record;
 135         struct perf_data_file    file;
 136         struct perf_session     *session;
 137         u64                      bytes_written;
 138         struct perf_tool         tool;
 139         bool                     maps_allocated;
 140         struct cpu_map          *cpus;
 141         struct thread_map       *threads;
 142         enum aggr_mode           aggr_mode;
 143 };
 144
 145 static struct perf_stat         perf_stat;
 146 #define STAT_RECORD             perf_stat.record
 147
 148 static volatile int done = 0;
 149
 150 static struct perf_stat_config stat_config = {
 151         .aggr_mode      = AGGR_GLOBAL,
 152         .scale          = true,
 153 };
 154
 155 static inline void diff_timespec(struct timespec *r, struct timespec *a,
 156                                  struct timespec *b)
 157 {
 158         r->tv_sec = a->tv_sec - b->tv_sec;
 159         if (a->tv_nsec < b->tv_nsec) {
 160                 r->tv_nsec = a->tv_nsec + 1000000000L - b->tv_nsec;
 161                 r->tv_sec--;
 162         } else {
 163                 r->tv_nsec = a->tv_nsec - b->tv_nsec ;
 164         }
 165 }
 166
 167 static void perf_stat__reset_stats(void)
 168 {
 169         perf_evlist__reset_stats(evsel_list);
 170         perf_stat__reset_shadow_stats();
 171 }
 172
 173 static int create_perf_stat_counter(struct perf_evsel *evsel)
 174 {
 175         struct perf_event_attr *attr = &evsel->attr;
 176
 177         if (stat_config.scale)
 178                 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
 179                                     PERF_FORMAT_TOTAL_TIME_RUNNING;
 180
 181         attr->inherit = !no_inherit;
 182
 183         /*
 184          * Some events get initialized with sample_(period/type) set,
 185          * like tracepoints. Clear it up for counting.
 186          */
 187         attr->sample_period = 0;
 188
 189         /*
 190          * But set sample_type to PERF_SAMPLE_IDENTIFIER, which should be harmless
 191          * while avoiding that older tools show confusing messages.
 192          *
 193          * However for pipe sessions we need to keep it zero,
 194          * because script's perf_evsel__check_attr is triggered
 195          * by attr->sample_type != 0, and we can't run it on
 196          * stat sessions.
 197          */
 198         if (!(STAT_RECORD && perf_stat.file.is_pipe))
 199                 attr->sample_type = PERF_SAMPLE_IDENTIFIER;
 200
 201         /*
 202          * Disabling all counters initially, they will be enabled
 203          * either manually by us or by kernel via enable_on_exec
 204          * set later.
 205          */
 206         if (perf_evsel__is_group_leader(evsel)) {
 207                 attr->disabled = 1;
 208
 209                 /*
 210                  * In case of initial_delay we enable tracee
 211                  * events manually.
 212                  */
 213                 if (target__none(&target) && !initial_delay)
 214                         attr->enable_on_exec = 1;
 215         }
 216
 217         if (target__has_cpu(&target))
 218                 return perf_evsel__open_per_cpu(evsel, perf_evsel__cpus(evsel));
 219
 220         return perf_evsel__open_per_thread(evsel, evsel_list->threads);
 221 }
 222
 223 /*
 224  * Does the counter have nsecs as a unit?
 225  */
 226 static inline int nsec_counter(struct perf_evsel *evsel)
 227 {
 228         if (perf_evsel__match(evsel, SOFTWARE, SW_CPU_CLOCK) ||
 229             perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
 230                 return 1;
 231
 232         return 0;
 233 }
 234
 235 static int process_synthesized_event(struct perf_tool *tool __maybe_unused,
 236                                      union perf_event *event,
 237                                      struct perf_sample *sample __maybe_unused,
 238                                      struct machine *machine __maybe_unused)
 239 {
 240         if (perf_data_file__write(&perf_stat.file, event, event->header.size) < 0) {
 241                 pr_err("failed to write perf data, error: %m\n");
 242                 return -1;
 243         }
 244
 245         perf_stat.bytes_written += event->header.size;
 246         return 0;
 247 }
 248
 249 static int write_stat_round_event(u64 tm, u64 type)
 250 {
 251         return perf_event__synthesize_stat_round(NULL, tm, type,
 252                                                  process_synthesized_event,
 253                                                  NULL);
 254 }
 255
 256 #define WRITE_STAT_ROUND_EVENT(time, interval) \
 257         write_stat_round_event(time, PERF_STAT_ROUND_TYPE__ ## interval)
 258
 259 #define SID(e, x, y) xyarray__entry(e->sample_id, x, y)
 260
 261 static int
 262 perf_evsel__write_stat_event(struct perf_evsel *counter, u32 cpu, u32 thread,
 263                              struct perf_counts_values *count)
 264 {
 265         struct perf_sample_id *sid = SID(counter, cpu, thread);
 266
 267         return perf_event__synthesize_stat(NULL, cpu, thread, sid->id, count,
 268                                            process_synthesized_event, NULL);
 269 }
 270
 271 /*
 272  * Read out the results of a single counter:
 273  * do not aggregate counts across CPUs in system-wide mode
 274  */
 275 static int read_counter(struct perf_evsel *counter)
 276 {
 277         int nthreads = thread_map__nr(evsel_list->threads);
 278         int ncpus = perf_evsel__nr_cpus(counter);
 279         int cpu, thread;
 280
 281         if (!counter->supported)
 282                 return -ENOENT;
 283
 284         if (counter->system_wide)
 285                 nthreads = 1;
 286
 287         for (thread = 0; thread < nthreads; thread++) {
 288                 for (cpu = 0; cpu < ncpus; cpu++) {
 289                         struct perf_counts_values *count;
 290
 291                         count = perf_counts(counter->counts, cpu, thread);
 292                         if (perf_evsel__read(counter, cpu, thread, count))
 293                                 return -1;
 294
 295                         if (STAT_RECORD) {
 296                                 if (perf_evsel__write_stat_event(counter, cpu, thread, count)) {
 297                                         pr_err("failed to write stat event\n");
 298                                         return -1;
 299                                 }
 300                         }
 301
 302                         if (verbose > 1) {
 303                                 fprintf(stat_config.output,
 304                                         "%s: %d: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
 305                                                 perf_evsel__name(counter),
 306                                                 cpu,
 307                                                 count->val, count->ena, count->run);
 308                         }
 309                 }
 310         }
 311
 312         return 0;
 313 }
 314
 315 static void read_counters(bool close_counters)
 316 {
 317         struct perf_evsel *counter;
 318
 319         evlist__for_each(evsel_list, counter) {
 320                 if (read_counter(counter))
 321                         pr_debug("failed to read counter %s\n", counter->name);
 322
 323                 if (perf_stat_process_counter(&stat_config, counter))
 324                         pr_warning("failed to process counter %s\n", counter->name);
 325
 326                 if (close_counters) {
 327                         perf_evsel__close_fd(counter, perf_evsel__nr_cpus(counter),
 328                                              thread_map__nr(evsel_list->threads));
 329                 }
 330         }
 331 }
 332
 333 static void process_interval(void)
 334 {
 335         struct timespec ts, rs;
 336
 337         read_counters(false);
 338
 339         clock_gettime(CLOCK_MONOTONIC, &ts);
 340         diff_timespec(&rs, &ts, &ref_time);
 341
 342         if (STAT_RECORD) {
 343                 if (WRITE_STAT_ROUND_EVENT(rs.tv_sec * NSECS_PER_SEC + rs.tv_nsec, INTERVAL))
 344                         pr_err("failed to write stat round event\n");
 345         }
 346
 347         print_counters(&rs, 0, NULL);
 348 }
 349
 350 static void enable_counters(void)
 351 {
 352         if (initial_delay)
 353                 usleep(initial_delay * 1000);
 354
 355         /*
 356          * We need to enable counters only if:
 357          * - we don't have tracee (attaching to task or cpu)
 358          * - we have initial delay configured
 359          */
 360         if (!target__none(&target) || initial_delay)
 361                 perf_evlist__enable(evsel_list);
 362 }
 363
 364 static volatile int workload_exec_errno;
 365
 366 /*
 367  * perf_evlist__prepare_workload will send a SIGUSR1
 368  * if the fork fails, since we asked by setting its
 369  * want_signal to true.
 370  */
 371 static void workload_exec_failed_signal(int signo __maybe_unused, siginfo_t *info,
 372                                         void *ucontext __maybe_unused)
 373 {
 374         workload_exec_errno = info->si_value.sival_int;
 375 }
 376
 377 static bool has_unit(struct perf_evsel *counter)
 378 {
 379         return counter->unit && *counter->unit;
 380 }
 381
 382 static bool has_scale(struct perf_evsel *counter)
 383 {
 384         return counter->scale != 1;
 385 }
 386
 387 static int perf_stat_synthesize_config(bool is_pipe)
 388 {
 389         struct perf_evsel *counter;
 390         int err;
 391
 392         if (is_pipe) {
 393                 err = perf_event__synthesize_attrs(NULL, perf_stat.session,
 394                                                    process_synthesized_event);
 395                 if (err < 0) {
 396                         pr_err("Couldn't synthesize attrs.\n");
 397                         return err;
 398                 }
 399         }
 400
 401         /*
 402          * Synthesize other events stuff not carried within
 403          * attr event - unit, scale, name
 404          */
 405         evlist__for_each(evsel_list, counter) {
 406                 if (!counter->supported)
 407                         continue;
 408
 409                 /*
 410                  * Synthesize unit and scale only if it's defined.
 411                  */
 412                 if (has_unit(counter)) {
 413                         err = perf_event__synthesize_event_update_unit(NULL, counter, process_synthesized_event);
 414                         if (err < 0) {
 415                                 pr_err("Couldn't synthesize evsel unit.\n");
 416                                 return err;
 417                         }
 418                 }
 419
 420                 if (has_scale(counter)) {
 421                         err = perf_event__synthesize_event_update_scale(NULL, counter, process_synthesized_event);
 422                         if (err < 0) {
 423                                 pr_err("Couldn't synthesize evsel scale.\n");
 424                                 return err;
 425                         }
 426                 }
 427
 428                 if (counter->own_cpus) {
 429                         err = perf_event__synthesize_event_update_cpus(NULL, counter, process_synthesized_event);
 430                         if (err < 0) {
 431                                 pr_err("Couldn't synthesize evsel scale.\n");
 432                                 return err;
 433                         }
 434                 }
 435
 436                 /*
 437                  * Name is needed only for pipe output,
 438                  * perf.data carries event names.
 439                  */
 440                 if (is_pipe) {
 441                         err = perf_event__synthesize_event_update_name(NULL, counter, process_synthesized_event);
 442                         if (err < 0) {
 443                                 pr_err("Couldn't synthesize evsel name.\n");
 444                                 return err;
 445                         }
 446                 }
 447         }
 448
 449         err = perf_event__synthesize_thread_map2(NULL, evsel_list->threads,
 450                                                 process_synthesized_event,
 451                                                 NULL);
 452         if (err < 0) {
 453                 pr_err("Couldn't synthesize thread map.\n");
 454                 return err;
 455         }
 456
 457         err = perf_event__synthesize_cpu_map(NULL, evsel_list->cpus,
 458                                              process_synthesized_event, NULL);
 459         if (err < 0) {
 460                 pr_err("Couldn't synthesize thread map.\n");
 461                 return err;
 462         }
 463
 464         err = perf_event__synthesize_stat_config(NULL, &stat_config,
 465                                                  process_synthesized_event, NULL);
 466         if (err < 0) {
 467                 pr_err("Couldn't synthesize config.\n");
 468                 return err;
 469         }
 470
 471         return 0;
 472 }
 473
 474 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
 475
 476 static int __store_counter_ids(struct perf_evsel *counter,
 477                                struct cpu_map *cpus,
 478                                struct thread_map *threads)
 479 {
 480         int cpu, thread;
 481
 482         for (cpu = 0; cpu < cpus->nr; cpu++) {
 483                 for (thread = 0; thread < threads->nr; thread++) {
 484                         int fd = FD(counter, cpu, thread);
 485
 486                         if (perf_evlist__id_add_fd(evsel_list, counter,
 487                                                    cpu, thread, fd) < 0)
 488                                 return -1;
 489                 }
 490         }
 491
 492         return 0;
 493 }
 494
 495 static int store_counter_ids(struct perf_evsel *counter)
 496 {
 497         struct cpu_map *cpus = counter->cpus;
 498         struct thread_map *threads = counter->threads;
 499
 500         if (perf_evsel__alloc_id(counter, cpus->nr, threads->nr))
 501                 return -ENOMEM;
 502
 503         return __store_counter_ids(counter, cpus, threads);
 504 }
 505
 506 static int __run_perf_stat(int argc, const char **argv)
 507 {
 508         int interval = stat_config.interval;
 509         char msg[512];
 510         unsigned long long t0, t1;
 511         struct perf_evsel *counter;
 512         struct timespec ts;
 513         size_t l;
 514         int status = 0;
 515         const bool forks = (argc > 0);
 516         bool is_pipe = STAT_RECORD ? perf_stat.file.is_pipe : false;
 517
 518         if (interval) {
 519                 ts.tv_sec  = interval / 1000;
 520                 ts.tv_nsec = (interval % 1000) * 1000000;
 521         } else {
 522                 ts.tv_sec  = 1;
 523                 ts.tv_nsec = 0;
 524         }
 525
 526         if (forks) {
 527                 if (perf_evlist__prepare_workload(evsel_list, &target, argv, is_pipe,
 528                                                   workload_exec_failed_signal) < 0) {
 529                         perror("failed to prepare workload");
 530                         return -1;
 531                 }
 532                 child_pid = evsel_list->workload.pid;
 533         }
 534
 535         if (group)
 536                 perf_evlist__set_leader(evsel_list);
 537
 538         evlist__for_each(evsel_list, counter) {
 539 try_again:
 540                 if (create_perf_stat_counter(counter) < 0) {
 541                         /*
 542                          * PPC returns ENXIO for HW counters until 2.6.37
 543                          * (behavior changed with commit b0a873e).
 544                          */
 545                         if (errno == EINVAL || errno == ENOSYS ||
 546                             errno == ENOENT || errno == EOPNOTSUPP ||
 547                             errno == ENXIO) {
 548                                 if (verbose)
 549                                         ui__warning("%s event is not supported by the kernel.\n",
 550                                                     perf_evsel__name(counter));
 551                                 counter->supported = false;
 552
 553                                 if ((counter->leader != counter) ||
 554                                     !(counter->leader->nr_members > 1))
 555                                         continue;
 556                         } else if (perf_evsel__fallback(counter, errno, msg, sizeof(msg))) {
 557                                 if (verbose)
 558                                         ui__warning("%s\n", msg);
 559                                 goto try_again;
 560                         }
 561
 562                         perf_evsel__open_strerror(counter, &target,
 563                                                   errno, msg, sizeof(msg));
 564                         ui__error("%s\n", msg);
 565
 566                         if (child_pid != -1)
 567                                 kill(child_pid, SIGTERM);
 568
 569                         return -1;
 570                 }
 571                 counter->supported = true;
 572
 573                 l = strlen(counter->unit);
 574                 if (l > unit_width)
 575                         unit_width = l;
 576
 577                 if (STAT_RECORD && store_counter_ids(counter))
 578                         return -1;
 579         }
 580
 581         if (perf_evlist__apply_filters(evsel_list, &counter)) {
 582                 error("failed to set filter \"%s\" on event %s with %d (%s)\n",
 583                         counter->filter, perf_evsel__name(counter), errno,
 584                         strerror_r(errno, msg, sizeof(msg)));
 585                 return -1;
 586         }
 587
 588         if (STAT_RECORD) {
 589                 int err, fd = perf_data_file__fd(&perf_stat.file);
 590
 591                 if (is_pipe) {
 592                         err = perf_header__write_pipe(perf_data_file__fd(&perf_stat.file));
 593                 } else {
 594                         err = perf_session__write_header(perf_stat.session, evsel_list,
 595                                                          fd, false);
 596                 }
 597
 598                 if (err < 0)
 599                         return err;
 600
 601                 err = perf_stat_synthesize_config(is_pipe);
 602                 if (err < 0)
 603                         return err;
 604         }
 605
 606         /*
 607          * Enable counters and exec the command:
 608          */
 609         t0 = rdclock();
 610         clock_gettime(CLOCK_MONOTONIC, &ref_time);
 611
 612         if (forks) {
 613                 perf_evlist__start_workload(evsel_list);
 614                 enable_counters();
 615
 616                 if (interval) {
 617                         while (!waitpid(child_pid, &status, WNOHANG)) {
 618                                 nanosleep(&ts, NULL);
 619                                 process_interval();
 620                         }
 621                 }
 622                 wait(&status);
 623
 624                 if (workload_exec_errno) {
 625                         const char *emsg = strerror_r(workload_exec_errno, msg, sizeof(msg));
 626                         pr_err("Workload failed: %s\n", emsg);
 627                         return -1;
 628                 }
 629
 630                 if (WIFSIGNALED(status))
 631                         psignal(WTERMSIG(status), argv[0]);
 632         } else {
 633                 enable_counters();
 634                 while (!done) {
 635                         nanosleep(&ts, NULL);
 636                         if (interval)
 637                                 process_interval();
 638                 }
 639         }
 640
 641         t1 = rdclock();
 642
 643         update_stats(&walltime_nsecs_stats, t1 - t0);
 644
 645         read_counters(true);
 646
 647         return WEXITSTATUS(status);
 648 }
 649
 650 static int run_perf_stat(int argc, const char **argv)
 651 {
 652         int ret;
 653
 654         if (pre_cmd) {
 655                 ret = system(pre_cmd);
 656                 if (ret)
 657                         return ret;
 658         }
 659
 660         if (sync_run)
 661                 sync();
 662
 663         ret = __run_perf_stat(argc, argv);
 664         if (ret)
 665                 return ret;
 666
 667         if (post_cmd) {
 668                 ret = system(post_cmd);
 669                 if (ret)
 670                         return ret;
 671         }
 672
 673         return ret;
 674 }
 675
 676 static void print_running(u64 run, u64 ena)
 677 {
 678         if (csv_output) {
 679                 fprintf(stat_config.output, "%s%" PRIu64 "%s%.2f",
 680                                         csv_sep,
 681                                         run,
 682                                         csv_sep,
 683                                         ena ? 100.0 * run / ena : 100.0);
 684         } else if (run != ena) {
 685                 fprintf(stat_config.output, "  (%.2f%%)", 100.0 * run / ena);
 686         }
 687 }
 688
 689 static void print_noise_pct(double total, double avg)
 690 {
 691         double pct = rel_stddev_stats(total, avg);
 692
 693         if (csv_output)
 694                 fprintf(stat_config.output, "%s%.2f%%", csv_sep, pct);
 695         else if (pct)
 696                 fprintf(stat_config.output, "  ( +-%6.2f%% )", pct);
 697 }
 698
 699 static void print_noise(struct perf_evsel *evsel, double avg)
 700 {
 701         struct perf_stat_evsel *ps;
 702
 703         if (run_count == 1)
 704                 return;
 705
 706         ps = evsel->priv;
 707         print_noise_pct(stddev_stats(&ps->res_stats[0]), avg);
 708 }
 709
 710 static void aggr_printout(struct perf_evsel *evsel, int id, int nr)
 711 {
 712         switch (stat_config.aggr_mode) {
 713         case AGGR_CORE:
 714                 fprintf(stat_config.output, "S%d-C%*d%s%*d%s",
 715                         cpu_map__id_to_socket(id),
 716                         csv_output ? 0 : -8,
 717                         cpu_map__id_to_cpu(id),
 718                         csv_sep,
 719                         csv_output ? 0 : 4,
 720                         nr,
 721                         csv_sep);
 722                 break;
 723         case AGGR_SOCKET:
 724                 fprintf(stat_config.output, "S%*d%s%*d%s",
 725                         csv_output ? 0 : -5,
 726                         id,
 727                         csv_sep,
 728                         csv_output ? 0 : 4,
 729                         nr,
 730                         csv_sep);
 731                         break;
 732         case AGGR_NONE:
 733                 fprintf(stat_config.output, "CPU%*d%s",
 734                         csv_output ? 0 : -4,
 735                         perf_evsel__cpus(evsel)->map[id], csv_sep);
 736                 break;
 737         case AGGR_THREAD:
 738                 fprintf(stat_config.output, "%*s-%*d%s",
 739                         csv_output ? 0 : 16,
 740                         thread_map__comm(evsel->threads, id),
 741                         csv_output ? 0 : -8,
 742                         thread_map__pid(evsel->threads, id),
 743                         csv_sep);
 744                 break;
 745         case AGGR_GLOBAL:
 746         case AGGR_UNSET:
 747         default:
 748                 break;
 749         }
 750 }
 751
 752 struct outstate {
 753         FILE *fh;
 754         bool newline;
 755         const char *prefix;
 756         int  nfields;
 757         int  id, nr;
 758         struct perf_evsel *evsel;
 759 };
 760
 761 #define METRIC_LEN  35
 762
 763 static void new_line_std(void *ctx)
 764 {
 765         struct outstate *os = ctx;
 766
 767         os->newline = true;
 768 }
 769
 770 static void do_new_line_std(struct outstate *os)
 771 {
 772         fputc('\n', os->fh);
 773         fputs(os->prefix, os->fh);
 774         aggr_printout(os->evsel, os->id, os->nr);
 775         if (stat_config.aggr_mode == AGGR_NONE)
 776                 fprintf(os->fh, "        ");
 777         fprintf(os->fh, "                                                 ");
 778 }
 779
 780 static void print_metric_std(void *ctx, const char *color, const char *fmt,
 781                              const char *unit, double val)
 782 {
 783         struct outstate *os = ctx;
 784         FILE *out = os->fh;
 785         int n;
 786         bool newline = os->newline;
 787
 788         os->newline = false;
 789
 790         if (unit == NULL || fmt == NULL) {
 791                 fprintf(out, "%-*s", METRIC_LEN, "");
 792                 return;
 793         }
 794
 795         if (newline)
 796                 do_new_line_std(os);
 797
 798         n = fprintf(out, " # ");
 799         if (color)
 800                 n += color_fprintf(out, color, fmt, val);
 801         else
 802                 n += fprintf(out, fmt, val);
 803         fprintf(out, " %-*s", METRIC_LEN - n - 1, unit);
 804 }
 805
 806 static void new_line_csv(void *ctx)
 807 {
 808         struct outstate *os = ctx;
 809         int i;
 810
 811         fputc('\n', os->fh);
 812         if (os->prefix)
 813                 fprintf(os->fh, "%s%s", os->prefix, csv_sep);
 814         aggr_printout(os->evsel, os->id, os->nr);
 815         for (i = 0; i < os->nfields; i++)
 816                 fputs(csv_sep, os->fh);
 817 }
 818
 819 static void print_metric_csv(void *ctx,
 820                              const char *color __maybe_unused,
 821                              const char *fmt, const char *unit, double val)
 822 {
 823         struct outstate *os = ctx;
 824         FILE *out = os->fh;
 825         char buf[64], *vals, *ends;
 826
 827         if (unit == NULL || fmt == NULL) {
 828                 fprintf(out, "%s%s%s%s", csv_sep, csv_sep, csv_sep, csv_sep);
 829                 return;
 830         }
 831         snprintf(buf, sizeof(buf), fmt, val);
 832         vals = buf;
 833         while (isspace(*vals))
 834                 vals++;
 835         ends = vals;
 836         while (isdigit(*ends) || *ends == '.')
 837                 ends++;
 838         *ends = 0;
 839         while (isspace(*unit))
 840                 unit++;
 841         fprintf(out, "%s%s%s%s", csv_sep, vals, csv_sep, unit);
 842 }
 843
 844 #define METRIC_ONLY_LEN 20
 845
 846 /* Filter out some columns that don't work well in metrics only mode */
 847
 848 static bool valid_only_metric(const char *unit)
 849 {
 850         if (!unit)
 851                 return false;
 852         if (strstr(unit, "/sec") ||
 853             strstr(unit, "hz") ||
 854             strstr(unit, "Hz") ||
 855             strstr(unit, "CPUs utilized"))
 856                 return false;
 857         return true;
 858 }
 859
 860 static const char *fixunit(char *buf, struct perf_evsel *evsel,
 861                            const char *unit)
 862 {
 863         if (!strncmp(unit, "of all", 6)) {
 864                 snprintf(buf, 1024, "%s %s", perf_evsel__name(evsel),
 865                          unit);
 866                 return buf;
 867         }
 868         return unit;
 869 }
 870
 871 static void print_metric_only(void *ctx, const char *color, const char *fmt,
 872                               const char *unit, double val)
 873 {
 874         struct outstate *os = ctx;
 875         FILE *out = os->fh;
 876         int n;
 877         char buf[1024];
 878         unsigned mlen = METRIC_ONLY_LEN;
 879
 880         if (!valid_only_metric(unit))
 881                 return;
 882         unit = fixunit(buf, os->evsel, unit);
 883         if (color)
 884                 n = color_fprintf(out, color, fmt, val);
 885         else
 886                 n = fprintf(out, fmt, val);
 887         if (n > METRIC_ONLY_LEN)
 888                 n = METRIC_ONLY_LEN;
 889         if (mlen < strlen(unit))
 890                 mlen = strlen(unit) + 1;
 891         fprintf(out, "%*s", mlen - n, "");
 892 }
 893
 894 static void print_metric_only_csv(void *ctx, const char *color __maybe_unused,
 895                                   const char *fmt,
 896                                   const char *unit, double val)
 897 {
 898         struct outstate *os = ctx;
 899         FILE *out = os->fh;
 900         char buf[64], *vals, *ends;
 901         char tbuf[1024];
 902
 903         if (!valid_only_metric(unit))
 904                 return;
 905         unit = fixunit(tbuf, os->evsel, unit);
 906         snprintf(buf, sizeof buf, fmt, val);
 907         vals = buf;
 908         while (isspace(*vals))
 909                 vals++;
 910         ends = vals;
 911         while (isdigit(*ends) || *ends == '.')
 912                 ends++;
 913         *ends = 0;
 914         fprintf(out, "%s%s", vals, csv_sep);
 915 }
 916
 917 static void new_line_metric(void *ctx __maybe_unused)
 918 {
 919 }
 920
 921 static void print_metric_header(void *ctx, const char *color __maybe_unused,
 922                                 const char *fmt __maybe_unused,
 923                                 const char *unit, double val __maybe_unused)
 924 {
 925         struct outstate *os = ctx;
 926         char tbuf[1024];
 927
 928         if (!valid_only_metric(unit))
 929                 return;
 930         unit = fixunit(tbuf, os->evsel, unit);
 931         if (csv_output)
 932                 fprintf(os->fh, "%s%s", unit, csv_sep);
 933         else
 934                 fprintf(os->fh, "%-*s ", METRIC_ONLY_LEN, unit);
 935 }
 936
 937 static void nsec_printout(int id, int nr, struct perf_evsel *evsel, double avg)
 938 {
 939         FILE *output = stat_config.output;
 940         double msecs = avg / 1e6;
 941         const char *fmt_v, *fmt_n;
 942         char name[25];
 943
 944         fmt_v = csv_output ? "%.6f%s" : "%18.6f%s";
 945         fmt_n = csv_output ? "%s" : "%-25s";
 946
 947         aggr_printout(evsel, id, nr);
 948
 949         scnprintf(name, sizeof(name), "%s%s",
 950                   perf_evsel__name(evsel), csv_output ? "" : " (msec)");
 951
 952         fprintf(output, fmt_v, msecs, csv_sep);
 953
 954         if (csv_output)
 955                 fprintf(output, "%s%s", evsel->unit, csv_sep);
 956         else
 957                 fprintf(output, "%-*s%s", unit_width, evsel->unit, csv_sep);
 958
 959         fprintf(output, fmt_n, name);
 960
 961         if (evsel->cgrp)
 962                 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
 963 }
 964
 965 static int first_shadow_cpu(struct perf_evsel *evsel, int id)
 966 {
 967         int i;
 968
 969         if (!aggr_get_id)
 970                 return 0;
 971
 972         if (stat_config.aggr_mode == AGGR_NONE)
 973                 return id;
 974
 975         if (stat_config.aggr_mode == AGGR_GLOBAL)
 976                 return 0;
 977
 978         for (i = 0; i < perf_evsel__nr_cpus(evsel); i++) {
 979                 int cpu2 = perf_evsel__cpus(evsel)->map[i];
 980
 981                 if (aggr_get_id(evsel_list->cpus, cpu2) == id)
 982                         return cpu2;
 983         }
 984         return 0;
 985 }
 986
 987 static void abs_printout(int id, int nr, struct perf_evsel *evsel, double avg)
 988 {
 989         FILE *output = stat_config.output;
 990         double sc =  evsel->scale;
 991         const char *fmt;
 992
 993         if (csv_output) {
 994                 fmt = sc != 1.0 ?  "%.2f%s" : "%.0f%s";
 995         } else {
 996                 if (big_num)
 997                         fmt = sc != 1.0 ? "%'18.2f%s" : "%'18.0f%s";
 998                 else
 999                         fmt = sc != 1.0 ? "%18.2f%s" : "%18.0f%s";
1000         }
1001
1002         aggr_printout(evsel, id, nr);
1003
1004         fprintf(output, fmt, avg, csv_sep);
1005
1006         if (evsel->unit)
1007                 fprintf(output, "%-*s%s",
1008                         csv_output ? 0 : unit_width,
1009                         evsel->unit, csv_sep);
1010
1011         fprintf(output, "%-*s", csv_output ? 0 : 25, perf_evsel__name(evsel));
1012
1013         if (evsel->cgrp)
1014                 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
1015 }
1016
1017 static void printout(int id, int nr, struct perf_evsel *counter, double uval,
1018                      char *prefix, u64 run, u64 ena, double noise)
1019 {
1020         struct perf_stat_output_ctx out;
1021         struct outstate os = {
1022                 .fh = stat_config.output,
1023                 .prefix = prefix ? prefix : "",
1024                 .id = id,
1025                 .nr = nr,
1026                 .evsel = counter,
1027         };
1028         print_metric_t pm = print_metric_std;
1029         void (*nl)(void *);
1030
1031         if (metric_only) {
1032                 nl = new_line_metric;
1033                 if (csv_output)
1034                         pm = print_metric_only_csv;
1035                 else
1036                         pm = print_metric_only;
1037         } else
1038                 nl = new_line_std;
1039
1040         if (csv_output && !metric_only) {
1041                 static int aggr_fields[] = {
1042                         [AGGR_GLOBAL] = 0,
1043                         [AGGR_THREAD] = 1,
1044                         [AGGR_NONE] = 1,
1045                         [AGGR_SOCKET] = 2,
1046                         [AGGR_CORE] = 2,
1047                 };
1048
1049                 pm = print_metric_csv;
1050                 nl = new_line_csv;
1051                 os.nfields = 3;
1052                 os.nfields += aggr_fields[stat_config.aggr_mode];
1053                 if (counter->cgrp)
1054                         os.nfields++;
1055         }
1056         if (run == 0 || ena == 0 || counter->counts->scaled == -1) {
1057                 if (metric_only) {
1058                         pm(&os, NULL, "", "", 0);
1059                         return;
1060                 }
1061                 aggr_printout(counter, id, nr);
1062
1063                 fprintf(stat_config.output, "%*s%s",
1064                         csv_output ? 0 : 18,
1065                         counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
1066                         csv_sep);
1067
1068                 fprintf(stat_config.output, "%-*s%s",
1069                         csv_output ? 0 : unit_width,
1070                         counter->unit, csv_sep);
1071
1072                 fprintf(stat_config.output, "%*s",
1073                         csv_output ? 0 : -25,
1074                         perf_evsel__name(counter));
1075
1076                 if (counter->cgrp)
1077                         fprintf(stat_config.output, "%s%s",
1078                                 csv_sep, counter->cgrp->name);
1079
1080                 if (!csv_output)
1081                         pm(&os, NULL, NULL, "", 0);
1082                 print_noise(counter, noise);
1083                 print_running(run, ena);
1084                 if (csv_output)
1085                         pm(&os, NULL, NULL, "", 0);
1086                 return;
1087         }
1088
1089         if (metric_only)
1090                 /* nothing */;
1091         else if (nsec_counter(counter))
1092                 nsec_printout(id, nr, counter, uval);
1093         else
1094                 abs_printout(id, nr, counter, uval);
1095
1096         out.print_metric = pm;
1097         out.new_line = nl;
1098         out.ctx = &os;
1099
1100         if (csv_output && !metric_only) {
1101                 print_noise(counter, noise);
1102                 print_running(run, ena);
1103         }
1104
1105         perf_stat__print_shadow_stats(counter, uval,
1106                                 first_shadow_cpu(counter, id),
1107                                 &out);
1108         if (!csv_output && !metric_only) {
1109                 print_noise(counter, noise);
1110                 print_running(run, ena);
1111         }
1112 }
1113
1114 static void aggr_update_shadow(void)
1115 {
1116         int cpu, s2, id, s;
1117         u64 val;
1118         struct perf_evsel *counter;
1119
1120         for (s = 0; s < aggr_map->nr; s++) {
1121                 id = aggr_map->map[s];
1122                 evlist__for_each(evsel_list, counter) {
1123                         val = 0;
1124                         for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
1125                                 s2 = aggr_get_id(evsel_list->cpus, cpu);
1126                                 if (s2 != id)
1127                                         continue;
1128                                 val += perf_counts(counter->counts, cpu, 0)->val;
1129                         }
1130                         val = val * counter->scale;
1131                         perf_stat__update_shadow_stats(counter, &val,
1132                                                        first_shadow_cpu(counter, id));
1133                 }
1134         }
1135 }
1136
1137 static void print_aggr(char *prefix)
1138 {
1139         FILE *output = stat_config.output;
1140         struct perf_evsel *counter;
1141         int cpu, s, s2, id, nr;
1142         double uval;
1143         u64 ena, run, val;
1144         bool first;
1145
1146         if (!(aggr_map || aggr_get_id))
1147                 return;
1148
1149         aggr_update_shadow();
1150
1151         /*
1152          * With metric_only everything is on a single line.
1153          * Without each counter has its own line.
1154          */
1155         for (s = 0; s < aggr_map->nr; s++) {
1156                 if (prefix && metric_only)
1157                         fprintf(output, "%s", prefix);
1158
1159                 id = aggr_map->map[s];
1160                 first = true;
1161                 evlist__for_each(evsel_list, counter) {
1162                         val = ena = run = 0;
1163                         nr = 0;
1164                         for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
1165                                 s2 = aggr_get_id(perf_evsel__cpus(counter), cpu);
1166                                 if (s2 != id)
1167                                         continue;
1168                                 val += perf_counts(counter->counts, cpu, 0)->val;
1169                                 ena += perf_counts(counter->counts, cpu, 0)->ena;
1170                                 run += perf_counts(counter->counts, cpu, 0)->run;
1171                                 nr++;
1172                         }
1173                         if (first && metric_only) {
1174                                 first = false;
1175                                 aggr_printout(counter, id, nr);
1176                         }
1177                         if (prefix && !metric_only)
1178                                 fprintf(output, "%s", prefix);
1179
1180                         uval = val * counter->scale;
1181                         printout(id, nr, counter, uval, prefix, run, ena, 1.0);
1182                         if (!metric_only)
1183                                 fputc('\n', output);
1184                 }
1185                 if (metric_only)
1186                         fputc('\n', output);
1187         }
1188 }
1189
1190 static void print_aggr_thread(struct perf_evsel *counter, char *prefix)
1191 {
1192         FILE *output = stat_config.output;
1193         int nthreads = thread_map__nr(counter->threads);
1194         int ncpus = cpu_map__nr(counter->cpus);
1195         int cpu, thread;
1196         double uval;
1197
1198         for (thread = 0; thread < nthreads; thread++) {
1199                 u64 ena = 0, run = 0, val = 0;
1200
1201                 for (cpu = 0; cpu < ncpus; cpu++) {
1202                         val += perf_counts(counter->counts, cpu, thread)->val;
1203                         ena += perf_counts(counter->counts, cpu, thread)->ena;
1204                         run += perf_counts(counter->counts, cpu, thread)->run;
1205                 }
1206
1207                 if (prefix)
1208                         fprintf(output, "%s", prefix);
1209
1210                 uval = val * counter->scale;
1211                 printout(thread, 0, counter, uval, prefix, run, ena, 1.0);
1212                 fputc('\n', output);
1213         }
1214 }
1215
1216 /*
1217  * Print out the results of a single counter:
1218  * aggregated counts in system-wide mode
1219  */
1220 static void print_counter_aggr(struct perf_evsel *counter, char *prefix)
1221 {
1222         FILE *output = stat_config.output;
1223         struct perf_stat_evsel *ps = counter->priv;
1224         double avg = avg_stats(&ps->res_stats[0]);
1225         double uval;
1226         double avg_enabled, avg_running;
1227
1228         avg_enabled = avg_stats(&ps->res_stats[1]);
1229         avg_running = avg_stats(&ps->res_stats[2]);
1230
1231         if (prefix && !metric_only)
1232                 fprintf(output, "%s", prefix);
1233
1234         uval = avg * counter->scale;
1235         printout(-1, 0, counter, uval, prefix, avg_running, avg_enabled, avg);
1236         if (!metric_only)
1237                 fprintf(output, "\n");
1238 }
1239
1240 /*
1241  * Print out the results of a single counter:
1242  * does not use aggregated count in system-wide
1243  */
1244 static void print_counter(struct perf_evsel *counter, char *prefix)
1245 {
1246         FILE *output = stat_config.output;
1247         u64 ena, run, val;
1248         double uval;
1249         int cpu;
1250
1251         for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
1252                 val = perf_counts(counter->counts, cpu, 0)->val;
1253                 ena = perf_counts(counter->counts, cpu, 0)->ena;
1254                 run = perf_counts(counter->counts, cpu, 0)->run;
1255
1256                 if (prefix)
1257                         fprintf(output, "%s", prefix);
1258
1259                 uval = val * counter->scale;
1260                 printout(cpu, 0, counter, uval, prefix, run, ena, 1.0);
1261
1262                 fputc('\n', output);
1263         }
1264 }
1265
1266 static void print_no_aggr_metric(char *prefix)
1267 {
1268         int cpu;
1269         int nrcpus = 0;
1270         struct perf_evsel *counter;
1271         u64 ena, run, val;
1272         double uval;
1273
1274         nrcpus = evsel_list->cpus->nr;
1275         for (cpu = 0; cpu < nrcpus; cpu++) {
1276                 bool first = true;
1277
1278                 if (prefix)
1279                         fputs(prefix, stat_config.output);
1280                 evlist__for_each(evsel_list, counter) {
1281                         if (first) {
1282                                 aggr_printout(counter, cpu, 0);
1283                                 first = false;
1284                         }
1285                         val = perf_counts(counter->counts, cpu, 0)->val;
1286                         ena = perf_counts(counter->counts, cpu, 0)->ena;
1287                         run = perf_counts(counter->counts, cpu, 0)->run;
1288
1289                         uval = val * counter->scale;
1290                         printout(cpu, 0, counter, uval, prefix, run, ena, 1.0);
1291                 }
1292                 fputc('\n', stat_config.output);
1293         }
1294 }
1295
1296 static int aggr_header_lens[] = {
1297         [AGGR_CORE] = 18,
1298         [AGGR_SOCKET] = 12,
1299         [AGGR_NONE] = 6,
1300         [AGGR_THREAD] = 24,
1301         [AGGR_GLOBAL] = 0,
1302 };
1303
1304 static void print_metric_headers(char *prefix)
1305 {
1306         struct perf_stat_output_ctx out;
1307         struct perf_evsel *counter;
1308         struct outstate os = {
1309                 .fh = stat_config.output
1310         };
1311
1312         if (prefix)
1313                 fprintf(stat_config.output, "%s", prefix);
1314
1315         if (!csv_output)
1316                 fprintf(stat_config.output, "%*s",
1317                         aggr_header_lens[stat_config.aggr_mode], "");
1318
1319         /* Print metrics headers only */
1320         evlist__for_each(evsel_list, counter) {
1321                 os.evsel = counter;
1322                 out.ctx = &os;
1323                 out.print_metric = print_metric_header;
1324                 out.new_line = new_line_metric;
1325                 os.evsel = counter;
1326                 perf_stat__print_shadow_stats(counter, 0,
1327                                               0,
1328                                               &out);
1329         }
1330         fputc('\n', stat_config.output);
1331 }
1332
1333 static void print_interval(char *prefix, struct timespec *ts)
1334 {
1335         FILE *output = stat_config.output;
1336         static int num_print_interval;
1337
1338         sprintf(prefix, "%6lu.%09lu%s", ts->tv_sec, ts->tv_nsec, csv_sep);
1339
1340         if (num_print_interval == 0 && !csv_output && !metric_only) {
1341                 switch (stat_config.aggr_mode) {
1342                 case AGGR_SOCKET:
1343                         fprintf(output, "#           time socket cpus             counts %*s events\n", unit_width, "unit");
1344                         break;
1345                 case AGGR_CORE:
1346                         fprintf(output, "#           time core         cpus             counts %*s events\n", unit_width, "unit");
1347                         break;
1348                 case AGGR_NONE:
1349                         fprintf(output, "#           time CPU                counts %*s events\n", unit_width, "unit");
1350                         break;
1351                 case AGGR_THREAD:
1352                         fprintf(output, "#           time             comm-pid                  counts %*s events\n", unit_width, "unit");
1353                         break;
1354                 case AGGR_GLOBAL:
1355                 default:
1356                         fprintf(output, "#           time             counts %*s events\n", unit_width, "unit");
1357                 case AGGR_UNSET:
1358                         break;
1359                 }
1360         }
1361
1362         if (++num_print_interval == 25)
1363                 num_print_interval = 0;
1364 }
1365
1366 static void print_header(int argc, const char **argv)
1367 {
1368         FILE *output = stat_config.output;
1369         int i;
1370
1371         fflush(stdout);
1372
1373         if (!csv_output) {
1374                 fprintf(output, "\n");
1375                 fprintf(output, " Performance counter stats for ");
1376                 if (target.system_wide)
1377                         fprintf(output, "\'system wide");
1378                 else if (target.cpu_list)
1379                         fprintf(output, "\'CPU(s) %s", target.cpu_list);
1380                 else if (!target__has_task(&target)) {
1381                         fprintf(output, "\'%s", argv ? argv[0] : "pipe");
1382                         for (i = 1; argv && (i < argc); i++)
1383                                 fprintf(output, " %s", argv[i]);
1384                 } else if (target.pid)
1385                         fprintf(output, "process id \'%s", target.pid);
1386                 else
1387                         fprintf(output, "thread id \'%s", target.tid);
1388
1389                 fprintf(output, "\'");
1390                 if (run_count > 1)
1391                         fprintf(output, " (%d runs)", run_count);
1392                 fprintf(output, ":\n\n");
1393         }
1394 }
1395
1396 static void print_footer(void)
1397 {
1398         FILE *output = stat_config.output;
1399
1400         if (!null_run)
1401                 fprintf(output, "\n");
1402         fprintf(output, " %17.9f seconds time elapsed",
1403                         avg_stats(&walltime_nsecs_stats)/1e9);
1404         if (run_count > 1) {
1405                 fprintf(output, "                                        ");
1406                 print_noise_pct(stddev_stats(&walltime_nsecs_stats),
1407                                 avg_stats(&walltime_nsecs_stats));
1408         }
1409         fprintf(output, "\n\n");
1410 }
1411
1412 static void print_counters(struct timespec *ts, int argc, const char **argv)
1413 {
1414         int interval = stat_config.interval;
1415         struct perf_evsel *counter;
1416         char buf[64], *prefix = NULL;
1417
1418         /* Do not print anything if we record to the pipe. */
1419         if (STAT_RECORD && perf_stat.file.is_pipe)
1420                 return;
1421
1422         if (interval)
1423                 print_interval(prefix = buf, ts);
1424         else
1425                 print_header(argc, argv);
1426
1427         if (metric_only) {
1428                 static int num_print_iv;
1429
1430                 if (num_print_iv == 0)
1431                         print_metric_headers(prefix);
1432                 if (num_print_iv++ == 25)
1433                         num_print_iv = 0;
1434                 if (stat_config.aggr_mode == AGGR_GLOBAL && prefix)
1435                         fprintf(stat_config.output, "%s", prefix);
1436         }
1437
1438         switch (stat_config.aggr_mode) {
1439         case AGGR_CORE:
1440         case AGGR_SOCKET:
1441                 print_aggr(prefix);
1442                 break;
1443         case AGGR_THREAD:
1444                 evlist__for_each(evsel_list, counter)
1445                         print_aggr_thread(counter, prefix);
1446                 break;
1447         case AGGR_GLOBAL:
1448                 evlist__for_each(evsel_list, counter)
1449                         print_counter_aggr(counter, prefix);
1450                 if (metric_only)
1451                         fputc('\n', stat_config.output);
1452                 break;
1453         case AGGR_NONE:
1454                 if (metric_only)
1455                         print_no_aggr_metric(prefix);
1456                 else {
1457                         evlist__for_each(evsel_list, counter)
1458                                 print_counter(counter, prefix);
1459                 }
1460                 break;
1461         case AGGR_UNSET:
1462         default:
1463                 break;
1464         }
1465
1466         if (!interval && !csv_output)
1467                 print_footer();
1468
1469         fflush(stat_config.output);
1470 }
1471
1472 static volatile int signr = -1;
1473
1474 static void skip_signal(int signo)
1475 {
1476         if ((child_pid == -1) || stat_config.interval)
1477                 done = 1;
1478
1479         signr = signo;
1480         /*
1481          * render child_pid harmless
1482          * won't send SIGTERM to a random
1483          * process in case of race condition
1484          * and fast PID recycling
1485          */
1486         child_pid = -1;
1487 }
1488
1489 static void sig_atexit(void)
1490 {
1491         sigset_t set, oset;
1492
1493         /*
1494          * avoid race condition with SIGCHLD handler
1495          * in skip_signal() which is modifying child_pid
1496          * goal is to avoid send SIGTERM to a random
1497          * process
1498          */
1499         sigemptyset(&set);
1500         sigaddset(&set, SIGCHLD);
1501         sigprocmask(SIG_BLOCK, &set, &oset);
1502
1503         if (child_pid != -1)
1504                 kill(child_pid, SIGTERM);
1505
1506         sigprocmask(SIG_SETMASK, &oset, NULL);
1507
1508         if (signr == -1)
1509                 return;
1510
1511         signal(signr, SIG_DFL);
1512         kill(getpid(), signr);
1513 }
1514
1515 static int stat__set_big_num(const struct option *opt __maybe_unused,
1516                              const char *s __maybe_unused, int unset)
1517 {
1518         big_num_opt = unset ? 0 : 1;
1519         return 0;
1520 }
1521
1522 static const struct option stat_options[] = {
1523         OPT_BOOLEAN('T', "transaction", &transaction_run,
1524                     "hardware transaction statistics"),
1525         OPT_CALLBACK('e', "event", &evsel_list, "event",
1526                      "event selector. use 'perf list' to list available events",
1527                      parse_events_option),
1528         OPT_CALLBACK(0, "filter", &evsel_list, "filter",
1529                      "event filter", parse_filter),
1530         OPT_BOOLEAN('i', "no-inherit", &no_inherit,
1531                     "child tasks do not inherit counters"),
1532         OPT_STRING('p', "pid", &target.pid, "pid",
1533                    "stat events on existing process id"),
1534         OPT_STRING('t', "tid", &target.tid, "tid",
1535                    "stat events on existing thread id"),
1536         OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
1537                     "system-wide collection from all CPUs"),
1538         OPT_BOOLEAN('g', "group", &group,
1539                     "put the counters into a counter group"),
1540         OPT_BOOLEAN('c', "scale", &stat_config.scale, "scale/normalize counters"),
1541         OPT_INCR('v', "verbose", &verbose,
1542                     "be more verbose (show counter open errors, etc)"),
1543         OPT_INTEGER('r', "repeat", &run_count,
1544                     "repeat command and print average + stddev (max: 100, forever: 0)"),
1545         OPT_BOOLEAN('n', "null", &null_run,
1546                     "null run - dont start any counters"),
1547         OPT_INCR('d', "detailed", &detailed_run,
1548                     "detailed run - start a lot of events"),
1549         OPT_BOOLEAN('S', "sync", &sync_run,
1550                     "call sync() before starting a run"),
1551         OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
1552                            "print large numbers with thousands\' separators",
1553                            stat__set_big_num),
1554         OPT_STRING('C', "cpu", &target.cpu_list, "cpu",
1555                     "list of cpus to monitor in system-wide"),
1556         OPT_SET_UINT('A', "no-aggr", &stat_config.aggr_mode,
1557                     "disable CPU count aggregation", AGGR_NONE),
1558         OPT_STRING('x', "field-separator", &csv_sep, "separator",
1559                    "print counts with custom separator"),
1560         OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
1561                      "monitor event in cgroup name only", parse_cgroups),
1562         OPT_STRING('o', "output", &output_name, "file", "output file name"),
1563         OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
1564         OPT_INTEGER(0, "log-fd", &output_fd,
1565                     "log output to fd, instead of stderr"),
1566         OPT_STRING(0, "pre", &pre_cmd, "command",
1567                         "command to run prior to the measured command"),
1568         OPT_STRING(0, "post", &post_cmd, "command",
1569                         "command to run after to the measured command"),
1570         OPT_UINTEGER('I', "interval-print", &stat_config.interval,
1571                     "print counts at regular interval in ms (>= 10)"),
1572         OPT_SET_UINT(0, "per-socket", &stat_config.aggr_mode,
1573                      "aggregate counts per processor socket", AGGR_SOCKET),
1574         OPT_SET_UINT(0, "per-core", &stat_config.aggr_mode,
1575                      "aggregate counts per physical processor core", AGGR_CORE),
1576         OPT_SET_UINT(0, "per-thread", &stat_config.aggr_mode,
1577                      "aggregate counts per thread", AGGR_THREAD),
1578         OPT_UINTEGER('D', "delay", &initial_delay,
1579                      "ms to wait before starting measurement after program start"),
1580         OPT_BOOLEAN(0, "metric-only", &metric_only,
1581                         "Only print computed metrics. No raw values"),
1582         OPT_END()
1583 };
1584
1585 static int perf_stat__get_socket(struct cpu_map *map, int cpu)
1586 {
1587         return cpu_map__get_socket(map, cpu, NULL);
1588 }
1589
1590 static int perf_stat__get_core(struct cpu_map *map, int cpu)
1591 {
1592         return cpu_map__get_core(map, cpu, NULL);
1593 }
1594
1595 static int cpu_map__get_max(struct cpu_map *map)
1596 {
1597         int i, max = -1;
1598
1599         for (i = 0; i < map->nr; i++) {
1600                 if (map->map[i] > max)
1601                         max = map->map[i];
1602         }
1603
1604         return max;
1605 }
1606
1607 static struct cpu_map *cpus_aggr_map;
1608
1609 static int perf_stat__get_aggr(aggr_get_id_t get_id, struct cpu_map *map, int idx)
1610 {
1611         int cpu;
1612
1613         if (idx >= map->nr)
1614                 return -1;
1615
1616         cpu = map->map[idx];
1617
1618         if (cpus_aggr_map->map[cpu] == -1)
1619                 cpus_aggr_map->map[cpu] = get_id(map, idx);
1620
1621         return cpus_aggr_map->map[cpu];
1622 }
1623
1624 static int perf_stat__get_socket_cached(struct cpu_map *map, int idx)
1625 {
1626         return perf_stat__get_aggr(perf_stat__get_socket, map, idx);
1627 }
1628
1629 static int perf_stat__get_core_cached(struct cpu_map *map, int idx)
1630 {
1631         return perf_stat__get_aggr(perf_stat__get_core, map, idx);
1632 }
1633
1634 static int perf_stat_init_aggr_mode(void)
1635 {
1636         int nr;
1637
1638         switch (stat_config.aggr_mode) {
1639         case AGGR_SOCKET:
1640                 if (cpu_map__build_socket_map(evsel_list->cpus, &aggr_map)) {
1641                         perror("cannot build socket map");
1642                         return -1;
1643                 }
1644                 aggr_get_id = perf_stat__get_socket_cached;
1645                 break;
1646         case AGGR_CORE:
1647                 if (cpu_map__build_core_map(evsel_list->cpus, &aggr_map)) {
1648                         perror("cannot build core map");
1649                         return -1;
1650                 }
1651                 aggr_get_id = perf_stat__get_core_cached;
1652                 break;
1653         case AGGR_NONE:
1654         case AGGR_GLOBAL:
1655         case AGGR_THREAD:
1656         case AGGR_UNSET:
1657         default:
1658                 break;
1659         }
1660
1661         /*
1662          * The evsel_list->cpus is the base we operate on,
1663          * taking the highest cpu number to be the size of
1664          * the aggregation translate cpumap.
1665          */
1666         nr = cpu_map__get_max(evsel_list->cpus);
1667         cpus_aggr_map = cpu_map__empty_new(nr + 1);
1668         return cpus_aggr_map ? 0 : -ENOMEM;
1669 }
1670
1671 static void perf_stat__exit_aggr_mode(void)
1672 {
1673         cpu_map__put(aggr_map);
1674         cpu_map__put(cpus_aggr_map);
1675         aggr_map = NULL;
1676         cpus_aggr_map = NULL;
1677 }
1678
1679 static inline int perf_env__get_cpu(struct perf_env *env, struct cpu_map *map, int idx)
1680 {
1681         int cpu;
1682
1683         if (idx > map->nr)
1684                 return -1;
1685
1686         cpu = map->map[idx];
1687
1688         if (cpu >= env->nr_cpus_online)
1689                 return -1;
1690
1691         return cpu;
1692 }
1693
1694 static int perf_env__get_socket(struct cpu_map *map, int idx, void *data)
1695 {
1696         struct perf_env *env = data;
1697         int cpu = perf_env__get_cpu(env, map, idx);
1698
1699         return cpu == -1 ? -1 : env->cpu[cpu].socket_id;
1700 }
1701
1702 static int perf_env__get_core(struct cpu_map *map, int idx, void *data)
1703 {
1704         struct perf_env *env = data;
1705         int core = -1, cpu = perf_env__get_cpu(env, map, idx);
1706
1707         if (cpu != -1) {
1708                 int socket_id = env->cpu[cpu].socket_id;
1709
1710                 /*
1711                  * Encode socket in upper 16 bits
1712                  * core_id is relative to socket, and
1713                  * we need a global id. So we combine
1714                  * socket + core id.
1715                  */
1716                 core = (socket_id << 16) | (env->cpu[cpu].core_id & 0xffff);
1717         }
1718
1719         return core;
1720 }
1721
1722 static int perf_env__build_socket_map(struct perf_env *env, struct cpu_map *cpus,
1723                                       struct cpu_map **sockp)
1724 {
1725         return cpu_map__build_map(cpus, sockp, perf_env__get_socket, env);
1726 }
1727
1728 static int perf_env__build_core_map(struct perf_env *env, struct cpu_map *cpus,
1729                                     struct cpu_map **corep)
1730 {
1731         return cpu_map__build_map(cpus, corep, perf_env__get_core, env);
1732 }
1733
1734 static int perf_stat__get_socket_file(struct cpu_map *map, int idx)
1735 {
1736         return perf_env__get_socket(map, idx, &perf_stat.session->header.env);
1737 }
1738
1739 static int perf_stat__get_core_file(struct cpu_map *map, int idx)
1740 {
1741         return perf_env__get_core(map, idx, &perf_stat.session->header.env);
1742 }
1743
1744 static int perf_stat_init_aggr_mode_file(struct perf_stat *st)
1745 {
1746         struct perf_env *env = &st->session->header.env;
1747
1748         switch (stat_config.aggr_mode) {
1749         case AGGR_SOCKET:
1750                 if (perf_env__build_socket_map(env, evsel_list->cpus, &aggr_map)) {
1751                         perror("cannot build socket map");
1752                         return -1;
1753                 }
1754                 aggr_get_id = perf_stat__get_socket_file;
1755                 break;
1756         case AGGR_CORE:
1757                 if (perf_env__build_core_map(env, evsel_list->cpus, &aggr_map)) {
1758                         perror("cannot build core map");
1759                         return -1;
1760                 }
1761                 aggr_get_id = perf_stat__get_core_file;
1762                 break;
1763         case AGGR_NONE:
1764         case AGGR_GLOBAL:
1765         case AGGR_THREAD:
1766         case AGGR_UNSET:
1767         default:
1768                 break;
1769         }
1770
1771         return 0;
1772 }
1773
1774 /*
1775  * Add default attributes, if there were no attributes specified or
1776  * if -d/--detailed, -d -d or -d -d -d is used:
1777  */
1778 static int add_default_attributes(void)
1779 {
1780         struct perf_event_attr default_attrs0[] = {
1781
1782   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK              },
1783   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES        },
1784   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS          },
1785   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS             },
1786
1787   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES              },
1788 };
1789         struct perf_event_attr frontend_attrs[] = {
1790   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND },
1791 };
1792         struct perf_event_attr backend_attrs[] = {
1793   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND  },
1794 };
1795         struct perf_event_attr default_attrs1[] = {
1796   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS            },
1797   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS     },
1798   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES           },
1799
1800 };
1801
1802 /*
1803  * Detailed stats (-d), covering the L1 and last level data caches:
1804  */
1805         struct perf_event_attr detailed_attrs[] = {
1806
1807   { .type = PERF_TYPE_HW_CACHE,
1808     .config =
1809          PERF_COUNT_HW_CACHE_L1D                <<  0  |
1810         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1811         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
1812
1813   { .type = PERF_TYPE_HW_CACHE,
1814     .config =
1815          PERF_COUNT_HW_CACHE_L1D                <<  0  |
1816         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1817         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
1818
1819   { .type = PERF_TYPE_HW_CACHE,
1820     .config =
1821          PERF_COUNT_HW_CACHE_LL                 <<  0  |
1822         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1823         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
1824
1825   { .type = PERF_TYPE_HW_CACHE,
1826     .config =
1827          PERF_COUNT_HW_CACHE_LL                 <<  0  |
1828         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1829         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
1830 };
1831
1832 /*
1833  * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
1834  */
1835         struct perf_event_attr very_detailed_attrs[] = {
1836
1837   { .type = PERF_TYPE_HW_CACHE,
1838     .config =
1839          PERF_COUNT_HW_CACHE_L1I                <<  0  |
1840         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1841         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
1842
1843   { .type = PERF_TYPE_HW_CACHE,
1844     .config =
1845          PERF_COUNT_HW_CACHE_L1I                <<  0  |
1846         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1847         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
1848
1849   { .type = PERF_TYPE_HW_CACHE,
1850     .config =
1851          PERF_COUNT_HW_CACHE_DTLB               <<  0  |
1852         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1853         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
1854
1855   { .type = PERF_TYPE_HW_CACHE,
1856     .config =
1857          PERF_COUNT_HW_CACHE_DTLB               <<  0  |
1858         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1859         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
1860
1861   { .type = PERF_TYPE_HW_CACHE,
1862     .config =
1863          PERF_COUNT_HW_CACHE_ITLB               <<  0  |
1864         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1865         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
1866
1867   { .type = PERF_TYPE_HW_CACHE,
1868     .config =
1869          PERF_COUNT_HW_CACHE_ITLB               <<  0  |
1870         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1871         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
1872
1873 };
1874
1875 /*
1876  * Very, very detailed stats (-d -d -d), adding prefetch events:
1877  */
1878         struct perf_event_attr very_very_detailed_attrs[] = {
1879
1880   { .type = PERF_TYPE_HW_CACHE,
1881     .config =
1882          PERF_COUNT_HW_CACHE_L1D                <<  0  |
1883         (PERF_COUNT_HW_CACHE_OP_PREFETCH        <<  8) |
1884         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
1885
1886   { .type = PERF_TYPE_HW_CACHE,
1887     .config =
1888          PERF_COUNT_HW_CACHE_L1D                <<  0  |
1889         (PERF_COUNT_HW_CACHE_OP_PREFETCH        <<  8) |
1890         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
1891 };
1892
1893         /* Set attrs if no event is selected and !null_run: */
1894         if (null_run)
1895                 return 0;
1896
1897         if (transaction_run) {
1898                 int err;
1899                 if (pmu_have_event("cpu", "cycles-ct") &&
1900                     pmu_have_event("cpu", "el-start"))
1901                         err = parse_events(evsel_list, transaction_attrs, NULL);
1902                 else
1903                         err = parse_events(evsel_list, transaction_limited_attrs, NULL);
1904                 if (err) {
1905                         fprintf(stderr, "Cannot set up transaction events\n");
1906                         return -1;
1907                 }
1908                 return 0;
1909         }
1910
1911         if (!evsel_list->nr_entries) {
1912                 if (perf_evlist__add_default_attrs(evsel_list, default_attrs0) < 0)
1913                         return -1;
1914                 if (pmu_have_event("cpu", "stalled-cycles-frontend")) {
1915                         if (perf_evlist__add_default_attrs(evsel_list,
1916                                                 frontend_attrs) < 0)
1917                                 return -1;
1918                 }
1919                 if (pmu_have_event("cpu", "stalled-cycles-backend")) {
1920                         if (perf_evlist__add_default_attrs(evsel_list,
1921                                                 backend_attrs) < 0)
1922                                 return -1;
1923                 }
1924                 if (perf_evlist__add_default_attrs(evsel_list, default_attrs1) < 0)
1925                         return -1;
1926         }
1927
1928         /* Detailed events get appended to the event list: */
1929
1930         if (detailed_run <  1)
1931                 return 0;
1932
1933         /* Append detailed run extra attributes: */
1934         if (perf_evlist__add_default_attrs(evsel_list, detailed_attrs) < 0)
1935                 return -1;
1936
1937         if (detailed_run < 2)
1938                 return 0;
1939
1940         /* Append very detailed run extra attributes: */
1941         if (perf_evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0)
1942                 return -1;
1943
1944         if (detailed_run < 3)
1945                 return 0;
1946
1947         /* Append very, very detailed run extra attributes: */
1948         return perf_evlist__add_default_attrs(evsel_list, very_very_detailed_attrs);
1949 }
1950
1951 static const char * const stat_record_usage[] = {
1952         "perf stat record [<options>]",
1953         NULL,
1954 };
1955
1956 static void init_features(struct perf_session *session)
1957 {
1958         int feat;
1959
1960         for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++)
1961                 perf_header__set_feat(&session->header, feat);
1962
1963         perf_header__clear_feat(&session->header, HEADER_BUILD_ID);
1964         perf_header__clear_feat(&session->header, HEADER_TRACING_DATA);
1965         perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK);
1966         perf_header__clear_feat(&session->header, HEADER_AUXTRACE);
1967 }
1968
1969 static int __cmd_record(int argc, const char **argv)
1970 {
1971         struct perf_session *session;
1972         struct perf_data_file *file = &perf_stat.file;
1973
1974         argc = parse_options(argc, argv, stat_options, stat_record_usage,
1975                              PARSE_OPT_STOP_AT_NON_OPTION);
1976
1977         if (output_name)
1978                 file->path = output_name;
1979
1980         if (run_count != 1 || forever) {
1981                 pr_err("Cannot use -r option with perf stat record.\n");
1982                 return -1;
1983         }
1984
1985         session = perf_session__new(file, false, NULL);
1986         if (session == NULL) {
1987                 pr_err("Perf session creation failed.\n");
1988                 return -1;
1989         }
1990
1991         init_features(session);
1992
1993         session->evlist   = evsel_list;
1994         perf_stat.session = session;
1995         perf_stat.record  = true;
1996         return argc;
1997 }
1998
1999 static int process_stat_round_event(struct perf_tool *tool __maybe_unused,
2000                                     union perf_event *event,
2001                                     struct perf_session *session)
2002 {
2003         struct stat_round_event *round = &event->stat_round;
2004         struct perf_evsel *counter;
2005         struct timespec tsh, *ts = NULL;
2006         const char **argv = session->header.env.cmdline_argv;
2007         int argc = session->header.env.nr_cmdline;
2008
2009         evlist__for_each(evsel_list, counter)
2010                 perf_stat_process_counter(&stat_config, counter);
2011
2012         if (round->type == PERF_STAT_ROUND_TYPE__FINAL)
2013                 update_stats(&walltime_nsecs_stats, round->time);
2014
2015         if (stat_config.interval && round->time) {
2016                 tsh.tv_sec  = round->time / NSECS_PER_SEC;
2017                 tsh.tv_nsec = round->time % NSECS_PER_SEC;
2018                 ts = &tsh;
2019         }
2020
2021         print_counters(ts, argc, argv);
2022         return 0;
2023 }
2024
2025 static
2026 int process_stat_config_event(struct perf_tool *tool __maybe_unused,
2027                               union perf_event *event,
2028                               struct perf_session *session __maybe_unused)
2029 {
2030         struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2031
2032         perf_event__read_stat_config(&stat_config, &event->stat_config);
2033
2034         if (cpu_map__empty(st->cpus)) {
2035                 if (st->aggr_mode != AGGR_UNSET)
2036                         pr_warning("warning: processing task data, aggregation mode not set\n");
2037                 return 0;
2038         }
2039
2040         if (st->aggr_mode != AGGR_UNSET)
2041                 stat_config.aggr_mode = st->aggr_mode;
2042
2043         if (perf_stat.file.is_pipe)
2044                 perf_stat_init_aggr_mode();
2045         else
2046                 perf_stat_init_aggr_mode_file(st);
2047
2048         return 0;
2049 }
2050
2051 static int set_maps(struct perf_stat *st)
2052 {
2053         if (!st->cpus || !st->threads)
2054                 return 0;
2055
2056         if (WARN_ONCE(st->maps_allocated, "stats double allocation\n"))
2057                 return -EINVAL;
2058
2059         perf_evlist__set_maps(evsel_list, st->cpus, st->threads);
2060
2061         if (perf_evlist__alloc_stats(evsel_list, true))
2062                 return -ENOMEM;
2063
2064         st->maps_allocated = true;
2065         return 0;
2066 }
2067
2068 static
2069 int process_thread_map_event(struct perf_tool *tool __maybe_unused,
2070                              union perf_event *event,
2071                              struct perf_session *session __maybe_unused)
2072 {
2073         struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2074
2075         if (st->threads) {
2076                 pr_warning("Extra thread map event, ignoring.\n");
2077                 return 0;
2078         }
2079
2080         st->threads = thread_map__new_event(&event->thread_map);
2081         if (!st->threads)
2082                 return -ENOMEM;
2083
2084         return set_maps(st);
2085 }
2086
2087 static
2088 int process_cpu_map_event(struct perf_tool *tool __maybe_unused,
2089                           union perf_event *event,
2090                           struct perf_session *session __maybe_unused)
2091 {
2092         struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2093         struct cpu_map *cpus;
2094
2095         if (st->cpus) {
2096                 pr_warning("Extra cpu map event, ignoring.\n");
2097                 return 0;
2098         }
2099
2100         cpus = cpu_map__new_data(&event->cpu_map.data);
2101         if (!cpus)
2102                 return -ENOMEM;
2103
2104         st->cpus = cpus;
2105         return set_maps(st);
2106 }
2107
2108 static const char * const stat_report_usage[] = {
2109         "perf stat report [<options>]",
2110         NULL,
2111 };
2112
2113 static struct perf_stat perf_stat = {
2114         .tool = {
2115                 .attr           = perf_event__process_attr,
2116                 .event_update   = perf_event__process_event_update,
2117                 .thread_map     = process_thread_map_event,
2118                 .cpu_map        = process_cpu_map_event,
2119                 .stat_config    = process_stat_config_event,
2120                 .stat           = perf_event__process_stat_event,
2121                 .stat_round     = process_stat_round_event,
2122         },
2123         .aggr_mode = AGGR_UNSET,
2124 };
2125
2126 static int __cmd_report(int argc, const char **argv)
2127 {
2128         struct perf_session *session;
2129         const struct option options[] = {
2130         OPT_STRING('i', "input", &input_name, "file", "input file name"),
2131         OPT_SET_UINT(0, "per-socket", &perf_stat.aggr_mode,
2132                      "aggregate counts per processor socket", AGGR_SOCKET),
2133         OPT_SET_UINT(0, "per-core", &perf_stat.aggr_mode,
2134                      "aggregate counts per physical processor core", AGGR_CORE),
2135         OPT_SET_UINT('A', "no-aggr", &perf_stat.aggr_mode,
2136                      "disable CPU count aggregation", AGGR_NONE),
2137         OPT_END()
2138         };
2139         struct stat st;
2140         int ret;
2141
2142         argc = parse_options(argc, argv, options, stat_report_usage, 0);
2143
2144         if (!input_name || !strlen(input_name)) {
2145                 if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode))
2146                         input_name = "-";
2147                 else
2148                         input_name = "perf.data";
2149         }
2150
2151         perf_stat.file.path = input_name;
2152         perf_stat.file.mode = PERF_DATA_MODE_READ;
2153
2154         session = perf_session__new(&perf_stat.file, false, &perf_stat.tool);
2155         if (session == NULL)
2156                 return -1;
2157
2158         perf_stat.session  = session;
2159         stat_config.output = stderr;
2160         evsel_list         = session->evlist;
2161
2162         ret = perf_session__process_events(session);
2163         if (ret)
2164                 return ret;
2165
2166         perf_session__delete(session);
2167         return 0;
2168 }
2169
2170 int cmd_stat(int argc, const char **argv, const char *prefix __maybe_unused)
2171 {
2172         const char * const stat_usage[] = {
2173                 "perf stat [<options>] [<command>]",
2174                 NULL
2175         };
2176         int status = -EINVAL, run_idx;
2177         const char *mode;
2178         FILE *output = stderr;
2179         unsigned int interval;
2180         const char * const stat_subcommands[] = { "record", "report" };
2181
2182         setlocale(LC_ALL, "");
2183
2184         evsel_list = perf_evlist__new();
2185         if (evsel_list == NULL)
2186                 return -ENOMEM;
2187
2188         parse_events__shrink_config_terms();
2189         argc = parse_options_subcommand(argc, argv, stat_options, stat_subcommands,
2190                                         (const char **) stat_usage,
2191                                         PARSE_OPT_STOP_AT_NON_OPTION);
2192         perf_stat__init_shadow_stats();
2193
2194         if (csv_sep) {
2195                 csv_output = true;
2196                 if (!strcmp(csv_sep, "\\t"))
2197                         csv_sep = "\t";
2198         } else
2199                 csv_sep = DEFAULT_SEPARATOR;
2200
2201         if (argc && !strncmp(argv[0], "rec", 3)) {
2202                 argc = __cmd_record(argc, argv);
2203                 if (argc < 0)
2204                         return -1;
2205         } else if (argc && !strncmp(argv[0], "rep", 3))
2206                 return __cmd_report(argc, argv);
2207
2208         interval = stat_config.interval;
2209
2210         /*
2211          * For record command the -o is already taken care of.
2212          */
2213         if (!STAT_RECORD && output_name && strcmp(output_name, "-"))
2214                 output = NULL;
2215
2216         if (output_name && output_fd) {
2217                 fprintf(stderr, "cannot use both --output and --log-fd\n");
2218                 parse_options_usage(stat_usage, stat_options, "o", 1);
2219                 parse_options_usage(NULL, stat_options, "log-fd", 0);
2220                 goto out;
2221         }
2222
2223         if (metric_only && stat_config.aggr_mode == AGGR_THREAD) {
2224                 fprintf(stderr, "--metric-only is not supported with --per-thread\n");
2225                 goto out;
2226         }
2227
2228         if (metric_only && run_count > 1) {
2229                 fprintf(stderr, "--metric-only is not supported with -r\n");
2230                 goto out;
2231         }
2232
2233         if (output_fd < 0) {
2234                 fprintf(stderr, "argument to --log-fd must be a > 0\n");
2235                 parse_options_usage(stat_usage, stat_options, "log-fd", 0);
2236                 goto out;
2237         }
2238
2239         if (!output) {
2240                 struct timespec tm;
2241                 mode = append_file ? "a" : "w";
2242
2243                 output = fopen(output_name, mode);
2244                 if (!output) {
2245                         perror("failed to create output file");
2246                         return -1;
2247                 }
2248                 clock_gettime(CLOCK_REALTIME, &tm);
2249                 fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
2250         } else if (output_fd > 0) {
2251                 mode = append_file ? "a" : "w";
2252                 output = fdopen(output_fd, mode);
2253                 if (!output) {
2254                         perror("Failed opening logfd");
2255                         return -errno;
2256                 }
2257         }
2258
2259         stat_config.output = output;
2260
2261         /*
2262          * let the spreadsheet do the pretty-printing
2263          */
2264         if (csv_output) {
2265                 /* User explicitly passed -B? */
2266                 if (big_num_opt == 1) {
2267                         fprintf(stderr, "-B option not supported with -x\n");
2268                         parse_options_usage(stat_usage, stat_options, "B", 1);
2269                         parse_options_usage(NULL, stat_options, "x", 1);
2270                         goto out;
2271                 } else /* Nope, so disable big number formatting */
2272                         big_num = false;
2273         } else if (big_num_opt == 0) /* User passed --no-big-num */
2274                 big_num = false;
2275
2276         if (!argc && target__none(&target))
2277                 usage_with_options(stat_usage, stat_options);
2278
2279         if (run_count < 0) {
2280                 pr_err("Run count must be a positive number\n");
2281                 parse_options_usage(stat_usage, stat_options, "r", 1);
2282                 goto out;
2283         } else if (run_count == 0) {
2284                 forever = true;
2285                 run_count = 1;
2286         }
2287
2288         if ((stat_config.aggr_mode == AGGR_THREAD) && !target__has_task(&target)) {
2289                 fprintf(stderr, "The --per-thread option is only available "
2290                         "when monitoring via -p -t options.\n");
2291                 parse_options_usage(NULL, stat_options, "p", 1);
2292                 parse_options_usage(NULL, stat_options, "t", 1);
2293                 goto out;
2294         }
2295
2296         /*
2297          * no_aggr, cgroup are for system-wide only
2298          * --per-thread is aggregated per thread, we dont mix it with cpu mode
2299          */
2300         if (((stat_config.aggr_mode != AGGR_GLOBAL &&
2301               stat_config.aggr_mode != AGGR_THREAD) || nr_cgroups) &&
2302             !target__has_cpu(&target)) {
2303                 fprintf(stderr, "both cgroup and no-aggregation "
2304                         "modes only available in system-wide mode\n");
2305
2306                 parse_options_usage(stat_usage, stat_options, "G", 1);
2307                 parse_options_usage(NULL, stat_options, "A", 1);
2308                 parse_options_usage(NULL, stat_options, "a", 1);
2309                 goto out;
2310         }
2311
2312         if (add_default_attributes())
2313                 goto out;
2314
2315         target__validate(&target);
2316
2317         if (perf_evlist__create_maps(evsel_list, &target) < 0) {
2318                 if (target__has_task(&target)) {
2319                         pr_err("Problems finding threads of monitor\n");
2320                         parse_options_usage(stat_usage, stat_options, "p", 1);
2321                         parse_options_usage(NULL, stat_options, "t", 1);
2322                 } else if (target__has_cpu(&target)) {
2323                         perror("failed to parse CPUs map");
2324                         parse_options_usage(stat_usage, stat_options, "C", 1);
2325                         parse_options_usage(NULL, stat_options, "a", 1);
2326                 }
2327                 goto out;
2328         }
2329
2330         /*
2331          * Initialize thread_map with comm names,
2332          * so we could print it out on output.
2333          */
2334         if (stat_config.aggr_mode == AGGR_THREAD)
2335                 thread_map__read_comms(evsel_list->threads);
2336
2337         if (interval && interval < 100) {
2338                 if (interval < 10) {
2339                         pr_err("print interval must be >= 10ms\n");
2340                         parse_options_usage(stat_usage, stat_options, "I", 1);
2341                         goto out;
2342                 } else
2343                         pr_warning("print interval < 100ms. "
2344                                    "The overhead percentage could be high in some cases. "
2345                                    "Please proceed with caution.\n");
2346         }
2347
2348         if (perf_evlist__alloc_stats(evsel_list, interval))
2349                 goto out;
2350
2351         if (perf_stat_init_aggr_mode())
2352                 goto out;
2353
2354         /*
2355          * We dont want to block the signals - that would cause
2356          * child tasks to inherit that and Ctrl-C would not work.
2357          * What we want is for Ctrl-C to work in the exec()-ed
2358          * task, but being ignored by perf stat itself:
2359          */
2360         atexit(sig_atexit);
2361         if (!forever)
2362                 signal(SIGINT,  skip_signal);
2363         signal(SIGCHLD, skip_signal);
2364         signal(SIGALRM, skip_signal);
2365         signal(SIGABRT, skip_signal);
2366
2367         status = 0;
2368         for (run_idx = 0; forever || run_idx < run_count; run_idx++) {
2369                 if (run_count != 1 && verbose)
2370                         fprintf(output, "[ perf stat: executing run #%d ... ]\n",
2371                                 run_idx + 1);
2372
2373                 status = run_perf_stat(argc, argv);
2374                 if (forever && status != -1) {
2375                         print_counters(NULL, argc, argv);
2376                         perf_stat__reset_stats();
2377                 }
2378         }
2379
2380         if (!forever && status != -1 && !interval)
2381                 print_counters(NULL, argc, argv);
2382
2383         if (STAT_RECORD) {
2384                 /*
2385                  * We synthesize the kernel mmap record just so that older tools
2386                  * don't emit warnings about not being able to resolve symbols
2387                  * due to /proc/sys/kernel/kptr_restrict settings and instear provide
2388                  * a saner message about no samples being in the perf.data file.
2389                  *
2390                  * This also serves to suppress a warning about f_header.data.size == 0
2391                  * in header.c at the moment 'perf stat record' gets introduced, which
2392                  * is not really needed once we start adding the stat specific PERF_RECORD_
2393                  * records, but the need to suppress the kptr_restrict messages in older
2394                  * tools remain  -acme
2395                  */
2396                 int fd = perf_data_file__fd(&perf_stat.file);
2397                 int err = perf_event__synthesize_kernel_mmap((void *)&perf_stat,
2398                                                              process_synthesized_event,
2399                                                              &perf_stat.session->machines.host);
2400                 if (err) {
2401                         pr_warning("Couldn't synthesize the kernel mmap record, harmless, "
2402                                    "older tools may produce warnings about this file\n.");
2403                 }
2404
2405                 if (!interval) {
2406                         if (WRITE_STAT_ROUND_EVENT(walltime_nsecs_stats.max, FINAL))
2407                                 pr_err("failed to write stat round event\n");
2408                 }
2409
2410                 if (!perf_stat.file.is_pipe) {
2411                         perf_stat.session->header.data_size += perf_stat.bytes_written;
2412                         perf_session__write_header(perf_stat.session, evsel_list, fd, true);
2413                 }
2414
2415                 perf_session__delete(perf_stat.session);
2416         }
2417
2418         perf_stat__exit_aggr_mode();
2419         perf_evlist__free_stats(evsel_list);
2420 out:
2421         perf_evlist__delete(evsel_list);
2422         return status;
2423 }