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/drv_configs.h"
  56 #include "util/color.h"
  57 #include "util/stat.h"
  58 #include "util/header.h"
  59 #include "util/cpumap.h"
  60 #include "util/thread.h"
  61 #include "util/thread_map.h"
  62 #include "util/counts.h"
  63 #include "util/group.h"
  64 #include "util/session.h"
  65 #include "util/tool.h"
  66 #include "util/group.h"
  67 #include "util/string2.h"
  68 #include "asm/bug.h"
  69
  70 #include <linux/time64.h>
  71 #include <api/fs/fs.h>
  72 #include <errno.h>
  73 #include <signal.h>
  74 #include <stdlib.h>
  75 #include <sys/prctl.h>
  76 #include <inttypes.h>
  77 #include <locale.h>
  78 #include <math.h>
  79 #include <sys/types.h>
  80 #include <sys/stat.h>
  81 #include <sys/wait.h>
  82 #include <unistd.h>
  83
  84 #include "sane_ctype.h"
  85
  86 #define DEFAULT_SEPARATOR       " "
  87 #define CNTR_NOT_SUPPORTED      "<not supported>"
  88 #define CNTR_NOT_COUNTED        "<not counted>"
  89
  90 static void print_counters(struct timespec *ts, int argc, const char **argv);
  91
  92 /* Default events used for perf stat -T */
  93 static const char *transaction_attrs = {
  94         "task-clock,"
  95         "{"
  96         "instructions,"
  97         "cycles,"
  98         "cpu/cycles-t/,"
  99         "cpu/tx-start/,"
 100         "cpu/el-start/,"
 101         "cpu/cycles-ct/"
 102         "}"
 103 };
 104
 105 /* More limited version when the CPU does not have all events. */
 106 static const char * transaction_limited_attrs = {
 107         "task-clock,"
 108         "{"
 109         "instructions,"
 110         "cycles,"
 111         "cpu/cycles-t/,"
 112         "cpu/tx-start/"
 113         "}"
 114 };
 115
 116 static const char * topdown_attrs[] = {
 117         "topdown-total-slots",
 118         "topdown-slots-retired",
 119         "topdown-recovery-bubbles",
 120         "topdown-fetch-bubbles",
 121         "topdown-slots-issued",
 122         NULL,
 123 };
 124
 125 static struct perf_evlist       *evsel_list;
 126
 127 static struct target target = {
 128         .uid    = UINT_MAX,
 129 };
 130
 131 typedef int (*aggr_get_id_t)(struct cpu_map *m, int cpu);
 132
 133 static int                      run_count                       =  1;
 134 static bool                     no_inherit                      = false;
 135 static volatile pid_t           child_pid                       = -1;
 136 static bool                     null_run                        =  false;
 137 static int                      detailed_run                    =  0;
 138 static bool                     transaction_run;
 139 static bool                     topdown_run                     = false;
 140 static bool                     big_num                         =  true;
 141 static int                      big_num_opt                     =  -1;
 142 static const char               *csv_sep                        = NULL;
 143 static bool                     csv_output                      = false;
 144 static bool                     group                           = false;
 145 static const char               *pre_cmd                        = NULL;
 146 static const char               *post_cmd                       = NULL;
 147 static bool                     sync_run                        = false;
 148 static unsigned int             initial_delay                   = 0;
 149 static unsigned int             unit_width                      = 4; /* strlen("unit") */
 150 static bool                     forever                         = false;
 151 static bool                     metric_only                     = false;
 152 static bool                     force_metric_only               = false;
 153 static bool                     no_merge                        = false;
 154 static struct timespec          ref_time;
 155 static struct cpu_map           *aggr_map;
 156 static aggr_get_id_t            aggr_get_id;
 157 static bool                     append_file;
 158 static const char               *output_name;
 159 static int                      output_fd;
 160 static int                      print_free_counters_hint;
 161
 162 struct perf_stat {
 163         bool                     record;
 164         struct perf_data_file    file;
 165         struct perf_session     *session;
 166         u64                      bytes_written;
 167         struct perf_tool         tool;
 168         bool                     maps_allocated;
 169         struct cpu_map          *cpus;
 170         struct thread_map       *threads;
 171         enum aggr_mode           aggr_mode;
 172 };
 173
 174 static struct perf_stat         perf_stat;
 175 #define STAT_RECORD             perf_stat.record
 176
 177 static volatile int done = 0;
 178
 179 static struct perf_stat_config stat_config = {
 180         .aggr_mode      = AGGR_GLOBAL,
 181         .scale          = true,
 182 };
 183
 184 static inline void diff_timespec(struct timespec *r, struct timespec *a,
 185                                  struct timespec *b)
 186 {
 187         r->tv_sec = a->tv_sec - b->tv_sec;
 188         if (a->tv_nsec < b->tv_nsec) {
 189                 r->tv_nsec = a->tv_nsec + NSEC_PER_SEC - b->tv_nsec;
 190                 r->tv_sec--;
 191         } else {
 192                 r->tv_nsec = a->tv_nsec - b->tv_nsec ;
 193         }
 194 }
 195
 196 static void perf_stat__reset_stats(void)
 197 {
 198         perf_evlist__reset_stats(evsel_list);
 199         perf_stat__reset_shadow_stats();
 200 }
 201
 202 static int create_perf_stat_counter(struct perf_evsel *evsel)
 203 {
 204         struct perf_event_attr *attr = &evsel->attr;
 205
 206         if (stat_config.scale)
 207                 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
 208                                     PERF_FORMAT_TOTAL_TIME_RUNNING;
 209
 210         attr->inherit = !no_inherit;
 211
 212         /*
 213          * Some events get initialized with sample_(period/type) set,
 214          * like tracepoints. Clear it up for counting.
 215          */
 216         attr->sample_period = 0;
 217
 218         /*
 219          * But set sample_type to PERF_SAMPLE_IDENTIFIER, which should be harmless
 220          * while avoiding that older tools show confusing messages.
 221          *
 222          * However for pipe sessions we need to keep it zero,
 223          * because script's perf_evsel__check_attr is triggered
 224          * by attr->sample_type != 0, and we can't run it on
 225          * stat sessions.
 226          */
 227         if (!(STAT_RECORD && perf_stat.file.is_pipe))
 228                 attr->sample_type = PERF_SAMPLE_IDENTIFIER;
 229
 230         /*
 231          * Disabling all counters initially, they will be enabled
 232          * either manually by us or by kernel via enable_on_exec
 233          * set later.
 234          */
 235         if (perf_evsel__is_group_leader(evsel)) {
 236                 attr->disabled = 1;
 237
 238                 /*
 239                  * In case of initial_delay we enable tracee
 240                  * events manually.
 241                  */
 242                 if (target__none(&target) && !initial_delay)
 243                         attr->enable_on_exec = 1;
 244         }
 245
 246         if (target__has_cpu(&target))
 247                 return perf_evsel__open_per_cpu(evsel, perf_evsel__cpus(evsel));
 248
 249         return perf_evsel__open_per_thread(evsel, evsel_list->threads);
 250 }
 251
 252 /*
 253  * Does the counter have nsecs as a unit?
 254  */
 255 static inline int nsec_counter(struct perf_evsel *evsel)
 256 {
 257         if (perf_evsel__match(evsel, SOFTWARE, SW_CPU_CLOCK) ||
 258             perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
 259                 return 1;
 260
 261         return 0;
 262 }
 263
 264 static int process_synthesized_event(struct perf_tool *tool __maybe_unused,
 265                                      union perf_event *event,
 266                                      struct perf_sample *sample __maybe_unused,
 267                                      struct machine *machine __maybe_unused)
 268 {
 269         if (perf_data_file__write(&perf_stat.file, event, event->header.size) < 0) {
 270                 pr_err("failed to write perf data, error: %m\n");
 271                 return -1;
 272         }
 273
 274         perf_stat.bytes_written += event->header.size;
 275         return 0;
 276 }
 277
 278 static int write_stat_round_event(u64 tm, u64 type)
 279 {
 280         return perf_event__synthesize_stat_round(NULL, tm, type,
 281                                                  process_synthesized_event,
 282                                                  NULL);
 283 }
 284
 285 #define WRITE_STAT_ROUND_EVENT(time, interval) \
 286         write_stat_round_event(time, PERF_STAT_ROUND_TYPE__ ## interval)
 287
 288 #define SID(e, x, y) xyarray__entry(e->sample_id, x, y)
 289
 290 static int
 291 perf_evsel__write_stat_event(struct perf_evsel *counter, u32 cpu, u32 thread,
 292                              struct perf_counts_values *count)
 293 {
 294         struct perf_sample_id *sid = SID(counter, cpu, thread);
 295
 296         return perf_event__synthesize_stat(NULL, cpu, thread, sid->id, count,
 297                                            process_synthesized_event, NULL);
 298 }
 299
 300 /*
 301  * Read out the results of a single counter:
 302  * do not aggregate counts across CPUs in system-wide mode
 303  */
 304 static int read_counter(struct perf_evsel *counter)
 305 {
 306         int nthreads = thread_map__nr(evsel_list->threads);
 307         int ncpus, cpu, thread;
 308
 309         if (target__has_cpu(&target))
 310                 ncpus = perf_evsel__nr_cpus(counter);
 311         else
 312                 ncpus = 1;
 313
 314         if (!counter->supported)
 315                 return -ENOENT;
 316
 317         if (counter->system_wide)
 318                 nthreads = 1;
 319
 320         for (thread = 0; thread < nthreads; thread++) {
 321                 for (cpu = 0; cpu < ncpus; cpu++) {
 322                         struct perf_counts_values *count;
 323
 324                         count = perf_counts(counter->counts, cpu, thread);
 325                         if (perf_evsel__read(counter, cpu, thread, count)) {
 326                                 counter->counts->scaled = -1;
 327                                 perf_counts(counter->counts, cpu, thread)->ena = 0;
 328                                 perf_counts(counter->counts, cpu, thread)->run = 0;
 329                                 return -1;
 330                         }
 331
 332                         if (STAT_RECORD) {
 333                                 if (perf_evsel__write_stat_event(counter, cpu, thread, count)) {
 334                                         pr_err("failed to write stat event\n");
 335                                         return -1;
 336                                 }
 337                         }
 338
 339                         if (verbose > 1) {
 340                                 fprintf(stat_config.output,
 341                                         "%s: %d: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
 342                                                 perf_evsel__name(counter),
 343                                                 cpu,
 344                                                 count->val, count->ena, count->run);
 345                         }
 346                 }
 347         }
 348
 349         return 0;
 350 }
 351
 352 static void read_counters(void)
 353 {
 354         struct perf_evsel *counter;
 355         int ret;
 356
 357         evlist__for_each_entry(evsel_list, counter) {
 358                 ret = read_counter(counter);
 359                 if (ret)
 360                         pr_debug("failed to read counter %s\n", counter->name);
 361
 362                 if (ret == 0 && perf_stat_process_counter(&stat_config, counter))
 363                         pr_warning("failed to process counter %s\n", counter->name);
 364         }
 365 }
 366
 367 static void process_interval(void)
 368 {
 369         struct timespec ts, rs;
 370
 371         read_counters();
 372
 373         clock_gettime(CLOCK_MONOTONIC, &ts);
 374         diff_timespec(&rs, &ts, &ref_time);
 375
 376         if (STAT_RECORD) {
 377                 if (WRITE_STAT_ROUND_EVENT(rs.tv_sec * NSEC_PER_SEC + rs.tv_nsec, INTERVAL))
 378                         pr_err("failed to write stat round event\n");
 379         }
 380
 381         print_counters(&rs, 0, NULL);
 382 }
 383
 384 static void enable_counters(void)
 385 {
 386         if (initial_delay)
 387                 usleep(initial_delay * USEC_PER_MSEC);
 388
 389         /*
 390          * We need to enable counters only if:
 391          * - we don't have tracee (attaching to task or cpu)
 392          * - we have initial delay configured
 393          */
 394         if (!target__none(&target) || initial_delay)
 395                 perf_evlist__enable(evsel_list);
 396 }
 397
 398 static void disable_counters(void)
 399 {
 400         /*
 401          * If we don't have tracee (attaching to task or cpu), counters may
 402          * still be running. To get accurate group ratios, we must stop groups
 403          * from counting before reading their constituent counters.
 404          */
 405         if (!target__none(&target))
 406                 perf_evlist__disable(evsel_list);
 407 }
 408
 409 static volatile int workload_exec_errno;
 410
 411 /*
 412  * perf_evlist__prepare_workload will send a SIGUSR1
 413  * if the fork fails, since we asked by setting its
 414  * want_signal to true.
 415  */
 416 static void workload_exec_failed_signal(int signo __maybe_unused, siginfo_t *info,
 417                                         void *ucontext __maybe_unused)
 418 {
 419         workload_exec_errno = info->si_value.sival_int;
 420 }
 421
 422 static bool has_unit(struct perf_evsel *counter)
 423 {
 424         return counter->unit && *counter->unit;
 425 }
 426
 427 static bool has_scale(struct perf_evsel *counter)
 428 {
 429         return counter->scale != 1;
 430 }
 431
 432 static int perf_stat_synthesize_config(bool is_pipe)
 433 {
 434         struct perf_evsel *counter;
 435         int err;
 436
 437         if (is_pipe) {
 438                 err = perf_event__synthesize_attrs(NULL, perf_stat.session,
 439                                                    process_synthesized_event);
 440                 if (err < 0) {
 441                         pr_err("Couldn't synthesize attrs.\n");
 442                         return err;
 443                 }
 444         }
 445
 446         /*
 447          * Synthesize other events stuff not carried within
 448          * attr event - unit, scale, name
 449          */
 450         evlist__for_each_entry(evsel_list, counter) {
 451                 if (!counter->supported)
 452                         continue;
 453
 454                 /*
 455                  * Synthesize unit and scale only if it's defined.
 456                  */
 457                 if (has_unit(counter)) {
 458                         err = perf_event__synthesize_event_update_unit(NULL, counter, process_synthesized_event);
 459                         if (err < 0) {
 460                                 pr_err("Couldn't synthesize evsel unit.\n");
 461                                 return err;
 462                         }
 463                 }
 464
 465                 if (has_scale(counter)) {
 466                         err = perf_event__synthesize_event_update_scale(NULL, counter, process_synthesized_event);
 467                         if (err < 0) {
 468                                 pr_err("Couldn't synthesize evsel scale.\n");
 469                                 return err;
 470                         }
 471                 }
 472
 473                 if (counter->own_cpus) {
 474                         err = perf_event__synthesize_event_update_cpus(NULL, counter, process_synthesized_event);
 475                         if (err < 0) {
 476                                 pr_err("Couldn't synthesize evsel scale.\n");
 477                                 return err;
 478                         }
 479                 }
 480
 481                 /*
 482                  * Name is needed only for pipe output,
 483                  * perf.data carries event names.
 484                  */
 485                 if (is_pipe) {
 486                         err = perf_event__synthesize_event_update_name(NULL, counter, process_synthesized_event);
 487                         if (err < 0) {
 488                                 pr_err("Couldn't synthesize evsel name.\n");
 489                                 return err;
 490                         }
 491                 }
 492         }
 493
 494         err = perf_event__synthesize_thread_map2(NULL, evsel_list->threads,
 495                                                 process_synthesized_event,
 496                                                 NULL);
 497         if (err < 0) {
 498                 pr_err("Couldn't synthesize thread map.\n");
 499                 return err;
 500         }
 501
 502         err = perf_event__synthesize_cpu_map(NULL, evsel_list->cpus,
 503                                              process_synthesized_event, NULL);
 504         if (err < 0) {
 505                 pr_err("Couldn't synthesize thread map.\n");
 506                 return err;
 507         }
 508
 509         err = perf_event__synthesize_stat_config(NULL, &stat_config,
 510                                                  process_synthesized_event, NULL);
 511         if (err < 0) {
 512                 pr_err("Couldn't synthesize config.\n");
 513                 return err;
 514         }
 515
 516         return 0;
 517 }
 518
 519 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
 520
 521 static int __store_counter_ids(struct perf_evsel *counter,
 522                                struct cpu_map *cpus,
 523                                struct thread_map *threads)
 524 {
 525         int cpu, thread;
 526
 527         for (cpu = 0; cpu < cpus->nr; cpu++) {
 528                 for (thread = 0; thread < threads->nr; thread++) {
 529                         int fd = FD(counter, cpu, thread);
 530
 531                         if (perf_evlist__id_add_fd(evsel_list, counter,
 532                                                    cpu, thread, fd) < 0)
 533                                 return -1;
 534                 }
 535         }
 536
 537         return 0;
 538 }
 539
 540 static int store_counter_ids(struct perf_evsel *counter)
 541 {
 542         struct cpu_map *cpus = counter->cpus;
 543         struct thread_map *threads = counter->threads;
 544
 545         if (perf_evsel__alloc_id(counter, cpus->nr, threads->nr))
 546                 return -ENOMEM;
 547
 548         return __store_counter_ids(counter, cpus, threads);
 549 }
 550
 551 static int __run_perf_stat(int argc, const char **argv)
 552 {
 553         int interval = stat_config.interval;
 554         char msg[BUFSIZ];
 555         unsigned long long t0, t1;
 556         struct perf_evsel *counter;
 557         struct timespec ts;
 558         size_t l;
 559         int status = 0;
 560         const bool forks = (argc > 0);
 561         bool is_pipe = STAT_RECORD ? perf_stat.file.is_pipe : false;
 562         struct perf_evsel_config_term *err_term;
 563
 564         if (interval) {
 565                 ts.tv_sec  = interval / USEC_PER_MSEC;
 566                 ts.tv_nsec = (interval % USEC_PER_MSEC) * NSEC_PER_MSEC;
 567         } else {
 568                 ts.tv_sec  = 1;
 569                 ts.tv_nsec = 0;
 570         }
 571
 572         if (forks) {
 573                 if (perf_evlist__prepare_workload(evsel_list, &target, argv, is_pipe,
 574                                                   workload_exec_failed_signal) < 0) {
 575                         perror("failed to prepare workload");
 576                         return -1;
 577                 }
 578                 child_pid = evsel_list->workload.pid;
 579         }
 580
 581         if (group)
 582                 perf_evlist__set_leader(evsel_list);
 583
 584         evlist__for_each_entry(evsel_list, counter) {
 585 try_again:
 586                 if (create_perf_stat_counter(counter) < 0) {
 587                         /*
 588                          * PPC returns ENXIO for HW counters until 2.6.37
 589                          * (behavior changed with commit b0a873e).
 590                          */
 591                         if (errno == EINVAL || errno == ENOSYS ||
 592                             errno == ENOENT || errno == EOPNOTSUPP ||
 593                             errno == ENXIO) {
 594                                 if (verbose > 0)
 595                                         ui__warning("%s event is not supported by the kernel.\n",
 596                                                     perf_evsel__name(counter));
 597                                 counter->supported = false;
 598
 599                                 if ((counter->leader != counter) ||
 600                                     !(counter->leader->nr_members > 1))
 601                                         continue;
 602                         } else if (perf_evsel__fallback(counter, errno, msg, sizeof(msg))) {
 603                                 if (verbose > 0)
 604                                         ui__warning("%s\n", msg);
 605                                 goto try_again;
 606                         }
 607
 608                         perf_evsel__open_strerror(counter, &target,
 609                                                   errno, msg, sizeof(msg));
 610                         ui__error("%s\n", msg);
 611
 612                         if (child_pid != -1)
 613                                 kill(child_pid, SIGTERM);
 614
 615                         return -1;
 616                 }
 617                 counter->supported = true;
 618
 619                 l = strlen(counter->unit);
 620                 if (l > unit_width)
 621                         unit_width = l;
 622
 623                 if (STAT_RECORD && store_counter_ids(counter))
 624                         return -1;
 625         }
 626
 627         if (perf_evlist__apply_filters(evsel_list, &counter)) {
 628                 error("failed to set filter \"%s\" on event %s with %d (%s)\n",
 629                         counter->filter, perf_evsel__name(counter), errno,
 630                         str_error_r(errno, msg, sizeof(msg)));
 631                 return -1;
 632         }
 633
 634         if (perf_evlist__apply_drv_configs(evsel_list, &counter, &err_term)) {
 635                 error("failed to set config \"%s\" on event %s with %d (%s)\n",
 636                       err_term->val.drv_cfg, perf_evsel__name(counter), errno,
 637                       str_error_r(errno, msg, sizeof(msg)));
 638                 return -1;
 639         }
 640
 641         if (STAT_RECORD) {
 642                 int err, fd = perf_data_file__fd(&perf_stat.file);
 643
 644                 if (is_pipe) {
 645                         err = perf_header__write_pipe(perf_data_file__fd(&perf_stat.file));
 646                 } else {
 647                         err = perf_session__write_header(perf_stat.session, evsel_list,
 648                                                          fd, false);
 649                 }
 650
 651                 if (err < 0)
 652                         return err;
 653
 654                 err = perf_stat_synthesize_config(is_pipe);
 655                 if (err < 0)
 656                         return err;
 657         }
 658
 659         /*
 660          * Enable counters and exec the command:
 661          */
 662         t0 = rdclock();
 663         clock_gettime(CLOCK_MONOTONIC, &ref_time);
 664
 665         if (forks) {
 666                 perf_evlist__start_workload(evsel_list);
 667                 enable_counters();
 668
 669                 if (interval) {
 670                         while (!waitpid(child_pid, &status, WNOHANG)) {
 671                                 nanosleep(&ts, NULL);
 672                                 process_interval();
 673                         }
 674                 }
 675                 wait(&status);
 676
 677                 if (workload_exec_errno) {
 678                         const char *emsg = str_error_r(workload_exec_errno, msg, sizeof(msg));
 679                         pr_err("Workload failed: %s\n", emsg);
 680                         return -1;
 681                 }
 682
 683                 if (WIFSIGNALED(status))
 684                         psignal(WTERMSIG(status), argv[0]);
 685         } else {
 686                 enable_counters();
 687                 while (!done) {
 688                         nanosleep(&ts, NULL);
 689                         if (interval)
 690                                 process_interval();
 691                 }
 692         }
 693
 694         disable_counters();
 695
 696         t1 = rdclock();
 697
 698         update_stats(&walltime_nsecs_stats, t1 - t0);
 699
 700         /*
 701          * Closing a group leader splits the group, and as we only disable
 702          * group leaders, results in remaining events becoming enabled. To
 703          * avoid arbitrary skew, we must read all counters before closing any
 704          * group leaders.
 705          */
 706         read_counters();
 707         perf_evlist__close(evsel_list);
 708
 709         return WEXITSTATUS(status);
 710 }
 711
 712 static int run_perf_stat(int argc, const char **argv)
 713 {
 714         int ret;
 715
 716         if (pre_cmd) {
 717                 ret = system(pre_cmd);
 718                 if (ret)
 719                         return ret;
 720         }
 721
 722         if (sync_run)
 723                 sync();
 724
 725         ret = __run_perf_stat(argc, argv);
 726         if (ret)
 727                 return ret;
 728
 729         if (post_cmd) {
 730                 ret = system(post_cmd);
 731                 if (ret)
 732                         return ret;
 733         }
 734
 735         return ret;
 736 }
 737
 738 static void print_running(u64 run, u64 ena)
 739 {
 740         if (csv_output) {
 741                 fprintf(stat_config.output, "%s%" PRIu64 "%s%.2f",
 742                                         csv_sep,
 743                                         run,
 744                                         csv_sep,
 745                                         ena ? 100.0 * run / ena : 100.0);
 746         } else if (run != ena) {
 747                 fprintf(stat_config.output, "  (%.2f%%)", 100.0 * run / ena);
 748         }
 749 }
 750
 751 static void print_noise_pct(double total, double avg)
 752 {
 753         double pct = rel_stddev_stats(total, avg);
 754
 755         if (csv_output)
 756                 fprintf(stat_config.output, "%s%.2f%%", csv_sep, pct);
 757         else if (pct)
 758                 fprintf(stat_config.output, "  ( +-%6.2f%% )", pct);
 759 }
 760
 761 static void print_noise(struct perf_evsel *evsel, double avg)
 762 {
 763         struct perf_stat_evsel *ps;
 764
 765         if (run_count == 1)
 766                 return;
 767
 768         ps = evsel->priv;
 769         print_noise_pct(stddev_stats(&ps->res_stats[0]), avg);
 770 }
 771
 772 static void aggr_printout(struct perf_evsel *evsel, int id, int nr)
 773 {
 774         switch (stat_config.aggr_mode) {
 775         case AGGR_CORE:
 776                 fprintf(stat_config.output, "S%d-C%*d%s%*d%s",
 777                         cpu_map__id_to_socket(id),
 778                         csv_output ? 0 : -8,
 779                         cpu_map__id_to_cpu(id),
 780                         csv_sep,
 781                         csv_output ? 0 : 4,
 782                         nr,
 783                         csv_sep);
 784                 break;
 785         case AGGR_SOCKET:
 786                 fprintf(stat_config.output, "S%*d%s%*d%s",
 787                         csv_output ? 0 : -5,
 788                         id,
 789                         csv_sep,
 790                         csv_output ? 0 : 4,
 791                         nr,
 792                         csv_sep);
 793                         break;
 794         case AGGR_NONE:
 795                 fprintf(stat_config.output, "CPU%*d%s",
 796                         csv_output ? 0 : -4,
 797                         perf_evsel__cpus(evsel)->map[id], csv_sep);
 798                 break;
 799         case AGGR_THREAD:
 800                 fprintf(stat_config.output, "%*s-%*d%s",
 801                         csv_output ? 0 : 16,
 802                         thread_map__comm(evsel->threads, id),
 803                         csv_output ? 0 : -8,
 804                         thread_map__pid(evsel->threads, id),
 805                         csv_sep);
 806                 break;
 807         case AGGR_GLOBAL:
 808         case AGGR_UNSET:
 809         default:
 810                 break;
 811         }
 812 }
 813
 814 struct outstate {
 815         FILE *fh;
 816         bool newline;
 817         const char *prefix;
 818         int  nfields;
 819         int  id, nr;
 820         struct perf_evsel *evsel;
 821 };
 822
 823 #define METRIC_LEN  35
 824
 825 static void new_line_std(void *ctx)
 826 {
 827         struct outstate *os = ctx;
 828
 829         os->newline = true;
 830 }
 831
 832 static void do_new_line_std(struct outstate *os)
 833 {
 834         fputc('\n', os->fh);
 835         fputs(os->prefix, os->fh);
 836         aggr_printout(os->evsel, os->id, os->nr);
 837         if (stat_config.aggr_mode == AGGR_NONE)
 838                 fprintf(os->fh, "        ");
 839         fprintf(os->fh, "                                                 ");
 840 }
 841
 842 static void print_metric_std(void *ctx, const char *color, const char *fmt,
 843                              const char *unit, double val)
 844 {
 845         struct outstate *os = ctx;
 846         FILE *out = os->fh;
 847         int n;
 848         bool newline = os->newline;
 849
 850         os->newline = false;
 851
 852         if (unit == NULL || fmt == NULL) {
 853                 fprintf(out, "%-*s", METRIC_LEN, "");
 854                 return;
 855         }
 856
 857         if (newline)
 858                 do_new_line_std(os);
 859
 860         n = fprintf(out, " # ");
 861         if (color)
 862                 n += color_fprintf(out, color, fmt, val);
 863         else
 864                 n += fprintf(out, fmt, val);
 865         fprintf(out, " %-*s", METRIC_LEN - n - 1, unit);
 866 }
 867
 868 static void new_line_csv(void *ctx)
 869 {
 870         struct outstate *os = ctx;
 871         int i;
 872
 873         fputc('\n', os->fh);
 874         if (os->prefix)
 875                 fprintf(os->fh, "%s%s", os->prefix, csv_sep);
 876         aggr_printout(os->evsel, os->id, os->nr);
 877         for (i = 0; i < os->nfields; i++)
 878                 fputs(csv_sep, os->fh);
 879 }
 880
 881 static void print_metric_csv(void *ctx,
 882                              const char *color __maybe_unused,
 883                              const char *fmt, const char *unit, double val)
 884 {
 885         struct outstate *os = ctx;
 886         FILE *out = os->fh;
 887         char buf[64], *vals, *ends;
 888
 889         if (unit == NULL || fmt == NULL) {
 890                 fprintf(out, "%s%s%s%s", csv_sep, csv_sep, csv_sep, csv_sep);
 891                 return;
 892         }
 893         snprintf(buf, sizeof(buf), fmt, val);
 894         ends = vals = ltrim(buf);
 895         while (isdigit(*ends) || *ends == '.')
 896                 ends++;
 897         *ends = 0;
 898         while (isspace(*unit))
 899                 unit++;
 900         fprintf(out, "%s%s%s%s", csv_sep, vals, csv_sep, unit);
 901 }
 902
 903 #define METRIC_ONLY_LEN 20
 904
 905 /* Filter out some columns that don't work well in metrics only mode */
 906
 907 static bool valid_only_metric(const char *unit)
 908 {
 909         if (!unit)
 910                 return false;
 911         if (strstr(unit, "/sec") ||
 912             strstr(unit, "hz") ||
 913             strstr(unit, "Hz") ||
 914             strstr(unit, "CPUs utilized"))
 915                 return false;
 916         return true;
 917 }
 918
 919 static const char *fixunit(char *buf, struct perf_evsel *evsel,
 920                            const char *unit)
 921 {
 922         if (!strncmp(unit, "of all", 6)) {
 923                 snprintf(buf, 1024, "%s %s", perf_evsel__name(evsel),
 924                          unit);
 925                 return buf;
 926         }
 927         return unit;
 928 }
 929
 930 static void print_metric_only(void *ctx, const char *color, const char *fmt,
 931                               const char *unit, double val)
 932 {
 933         struct outstate *os = ctx;
 934         FILE *out = os->fh;
 935         int n;
 936         char buf[1024];
 937         unsigned mlen = METRIC_ONLY_LEN;
 938
 939         if (!valid_only_metric(unit))
 940                 return;
 941         unit = fixunit(buf, os->evsel, unit);
 942         if (color)
 943                 n = color_fprintf(out, color, fmt, val);
 944         else
 945                 n = fprintf(out, fmt, val);
 946         if (n > METRIC_ONLY_LEN)
 947                 n = METRIC_ONLY_LEN;
 948         if (mlen < strlen(unit))
 949                 mlen = strlen(unit) + 1;
 950         fprintf(out, "%*s", mlen - n, "");
 951 }
 952
 953 static void print_metric_only_csv(void *ctx, const char *color __maybe_unused,
 954                                   const char *fmt,
 955                                   const char *unit, double val)
 956 {
 957         struct outstate *os = ctx;
 958         FILE *out = os->fh;
 959         char buf[64], *vals, *ends;
 960         char tbuf[1024];
 961
 962         if (!valid_only_metric(unit))
 963                 return;
 964         unit = fixunit(tbuf, os->evsel, unit);
 965         snprintf(buf, sizeof buf, fmt, val);
 966         ends = vals = ltrim(buf);
 967         while (isdigit(*ends) || *ends == '.')
 968                 ends++;
 969         *ends = 0;
 970         fprintf(out, "%s%s", vals, csv_sep);
 971 }
 972
 973 static void new_line_metric(void *ctx __maybe_unused)
 974 {
 975 }
 976
 977 static void print_metric_header(void *ctx, const char *color __maybe_unused,
 978                                 const char *fmt __maybe_unused,
 979                                 const char *unit, double val __maybe_unused)
 980 {
 981         struct outstate *os = ctx;
 982         char tbuf[1024];
 983
 984         if (!valid_only_metric(unit))
 985                 return;
 986         unit = fixunit(tbuf, os->evsel, unit);
 987         if (csv_output)
 988                 fprintf(os->fh, "%s%s", unit, csv_sep);
 989         else
 990                 fprintf(os->fh, "%-*s ", METRIC_ONLY_LEN, unit);
 991 }
 992
 993 static void nsec_printout(int id, int nr, struct perf_evsel *evsel, double avg)
 994 {
 995         FILE *output = stat_config.output;
 996         double msecs = avg / NSEC_PER_MSEC;
 997         const char *fmt_v, *fmt_n;
 998         char name[25];
 999
1000         fmt_v = csv_output ? "%.6f%s" : "%18.6f%s";
1001         fmt_n = csv_output ? "%s" : "%-25s";
1002
1003         aggr_printout(evsel, id, nr);
1004
1005         scnprintf(name, sizeof(name), "%s%s",
1006                   perf_evsel__name(evsel), csv_output ? "" : " (msec)");
1007
1008         fprintf(output, fmt_v, msecs, csv_sep);
1009
1010         if (csv_output)
1011                 fprintf(output, "%s%s", evsel->unit, csv_sep);
1012         else
1013                 fprintf(output, "%-*s%s", unit_width, evsel->unit, csv_sep);
1014
1015         fprintf(output, fmt_n, name);
1016
1017         if (evsel->cgrp)
1018                 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
1019 }
1020
1021 static int first_shadow_cpu(struct perf_evsel *evsel, int id)
1022 {
1023         int i;
1024
1025         if (!aggr_get_id)
1026                 return 0;
1027
1028         if (stat_config.aggr_mode == AGGR_NONE)
1029                 return id;
1030
1031         if (stat_config.aggr_mode == AGGR_GLOBAL)
1032                 return 0;
1033
1034         for (i = 0; i < perf_evsel__nr_cpus(evsel); i++) {
1035                 int cpu2 = perf_evsel__cpus(evsel)->map[i];
1036
1037                 if (aggr_get_id(evsel_list->cpus, cpu2) == id)
1038                         return cpu2;
1039         }
1040         return 0;
1041 }
1042
1043 static void abs_printout(int id, int nr, struct perf_evsel *evsel, double avg)
1044 {
1045         FILE *output = stat_config.output;
1046         double sc =  evsel->scale;
1047         const char *fmt;
1048
1049         if (csv_output) {
1050                 fmt = floor(sc) != sc ?  "%.2f%s" : "%.0f%s";
1051         } else {
1052                 if (big_num)
1053                         fmt = floor(sc) != sc ? "%'18.2f%s" : "%'18.0f%s";
1054                 else
1055                         fmt = floor(sc) != sc ? "%18.2f%s" : "%18.0f%s";
1056         }
1057
1058         aggr_printout(evsel, id, nr);
1059
1060         fprintf(output, fmt, avg, csv_sep);
1061
1062         if (evsel->unit)
1063                 fprintf(output, "%-*s%s",
1064                         csv_output ? 0 : unit_width,
1065                         evsel->unit, csv_sep);
1066
1067         fprintf(output, "%-*s", csv_output ? 0 : 25, perf_evsel__name(evsel));
1068
1069         if (evsel->cgrp)
1070                 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
1071 }
1072
1073 static void printout(int id, int nr, struct perf_evsel *counter, double uval,
1074                      char *prefix, u64 run, u64 ena, double noise)
1075 {
1076         struct perf_stat_output_ctx out;
1077         struct outstate os = {
1078                 .fh = stat_config.output,
1079                 .prefix = prefix ? prefix : "",
1080                 .id = id,
1081                 .nr = nr,
1082                 .evsel = counter,
1083         };
1084         print_metric_t pm = print_metric_std;
1085         void (*nl)(void *);
1086
1087         if (metric_only) {
1088                 nl = new_line_metric;
1089                 if (csv_output)
1090                         pm = print_metric_only_csv;
1091                 else
1092                         pm = print_metric_only;
1093         } else
1094                 nl = new_line_std;
1095
1096         if (csv_output && !metric_only) {
1097                 static int aggr_fields[] = {
1098                         [AGGR_GLOBAL] = 0,
1099                         [AGGR_THREAD] = 1,
1100                         [AGGR_NONE] = 1,
1101                         [AGGR_SOCKET] = 2,
1102                         [AGGR_CORE] = 2,
1103                 };
1104
1105                 pm = print_metric_csv;
1106                 nl = new_line_csv;
1107                 os.nfields = 3;
1108                 os.nfields += aggr_fields[stat_config.aggr_mode];
1109                 if (counter->cgrp)
1110                         os.nfields++;
1111         }
1112         if (run == 0 || ena == 0 || counter->counts->scaled == -1) {
1113                 if (metric_only) {
1114                         pm(&os, NULL, "", "", 0);
1115                         return;
1116                 }
1117                 aggr_printout(counter, id, nr);
1118
1119                 fprintf(stat_config.output, "%*s%s",
1120                         csv_output ? 0 : 18,
1121                         counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
1122                         csv_sep);
1123
1124                 if (counter->supported)
1125                         print_free_counters_hint = 1;
1126
1127                 fprintf(stat_config.output, "%-*s%s",
1128                         csv_output ? 0 : unit_width,
1129                         counter->unit, csv_sep);
1130
1131                 fprintf(stat_config.output, "%*s",
1132                         csv_output ? 0 : -25,
1133                         perf_evsel__name(counter));
1134
1135                 if (counter->cgrp)
1136                         fprintf(stat_config.output, "%s%s",
1137                                 csv_sep, counter->cgrp->name);
1138
1139                 if (!csv_output)
1140                         pm(&os, NULL, NULL, "", 0);
1141                 print_noise(counter, noise);
1142                 print_running(run, ena);
1143                 if (csv_output)
1144                         pm(&os, NULL, NULL, "", 0);
1145                 return;
1146         }
1147
1148         if (metric_only)
1149                 /* nothing */;
1150         else if (nsec_counter(counter))
1151                 nsec_printout(id, nr, counter, uval);
1152         else
1153                 abs_printout(id, nr, counter, uval);
1154
1155         out.print_metric = pm;
1156         out.new_line = nl;
1157         out.ctx = &os;
1158         out.force_header = false;
1159
1160         if (csv_output && !metric_only) {
1161                 print_noise(counter, noise);
1162                 print_running(run, ena);
1163         }
1164
1165         perf_stat__print_shadow_stats(counter, uval,
1166                                 first_shadow_cpu(counter, id),
1167                                 &out);
1168         if (!csv_output && !metric_only) {
1169                 print_noise(counter, noise);
1170                 print_running(run, ena);
1171         }
1172 }
1173
1174 static void aggr_update_shadow(void)
1175 {
1176         int cpu, s2, id, s;
1177         u64 val;
1178         struct perf_evsel *counter;
1179
1180         for (s = 0; s < aggr_map->nr; s++) {
1181                 id = aggr_map->map[s];
1182                 evlist__for_each_entry(evsel_list, counter) {
1183                         val = 0;
1184                         for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
1185                                 s2 = aggr_get_id(evsel_list->cpus, cpu);
1186                                 if (s2 != id)
1187                                         continue;
1188                                 val += perf_counts(counter->counts, cpu, 0)->val;
1189                         }
1190                         val = val * counter->scale;
1191                         perf_stat__update_shadow_stats(counter, &val,
1192                                                        first_shadow_cpu(counter, id));
1193                 }
1194         }
1195 }
1196
1197 static void collect_all_aliases(struct perf_evsel *counter,
1198                             void (*cb)(struct perf_evsel *counter, void *data,
1199                                        bool first),
1200                             void *data)
1201 {
1202         struct perf_evsel *alias;
1203
1204         alias = list_prepare_entry(counter, &(evsel_list->entries), node);
1205         list_for_each_entry_continue (alias, &evsel_list->entries, node) {
1206                 if (strcmp(perf_evsel__name(alias), perf_evsel__name(counter)) ||
1207                     alias->scale != counter->scale ||
1208                     alias->cgrp != counter->cgrp ||
1209                     strcmp(alias->unit, counter->unit) ||
1210                     nsec_counter(alias) != nsec_counter(counter))
1211                         break;
1212                 alias->merged_stat = true;
1213                 cb(alias, data, false);
1214         }
1215 }
1216
1217 static bool collect_data(struct perf_evsel *counter,
1218                             void (*cb)(struct perf_evsel *counter, void *data,
1219                                        bool first),
1220                             void *data)
1221 {
1222         if (counter->merged_stat)
1223                 return false;
1224         cb(counter, data, true);
1225         if (!no_merge)
1226                 collect_all_aliases(counter, cb, data);
1227         return true;
1228 }
1229
1230 struct aggr_data {
1231         u64 ena, run, val;
1232         int id;
1233         int nr;
1234         int cpu;
1235 };
1236
1237 static void aggr_cb(struct perf_evsel *counter, void *data, bool first)
1238 {
1239         struct aggr_data *ad = data;
1240         int cpu, s2;
1241
1242         for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
1243                 struct perf_counts_values *counts;
1244
1245                 s2 = aggr_get_id(perf_evsel__cpus(counter), cpu);
1246                 if (s2 != ad->id)
1247                         continue;
1248                 if (first)
1249                         ad->nr++;
1250                 counts = perf_counts(counter->counts, cpu, 0);
1251                 /*
1252                  * When any result is bad, make them all to give
1253                  * consistent output in interval mode.
1254                  */
1255                 if (counts->ena == 0 || counts->run == 0 ||
1256                     counter->counts->scaled == -1) {
1257                         ad->ena = 0;
1258                         ad->run = 0;
1259                         break;
1260                 }
1261                 ad->val += counts->val;
1262                 ad->ena += counts->ena;
1263                 ad->run += counts->run;
1264         }
1265 }
1266
1267 static void print_aggr(char *prefix)
1268 {
1269         FILE *output = stat_config.output;
1270         struct perf_evsel *counter;
1271         int s, id, nr;
1272         double uval;
1273         u64 ena, run, val;
1274         bool first;
1275
1276         if (!(aggr_map || aggr_get_id))
1277                 return;
1278
1279         aggr_update_shadow();
1280
1281         /*
1282          * With metric_only everything is on a single line.
1283          * Without each counter has its own line.
1284          */
1285         for (s = 0; s < aggr_map->nr; s++) {
1286                 struct aggr_data ad;
1287                 if (prefix && metric_only)
1288                         fprintf(output, "%s", prefix);
1289
1290                 ad.id = id = aggr_map->map[s];
1291                 first = true;
1292                 evlist__for_each_entry(evsel_list, counter) {
1293                         ad.val = ad.ena = ad.run = 0;
1294                         ad.nr = 0;
1295                         if (!collect_data(counter, aggr_cb, &ad))
1296                                 continue;
1297                         nr = ad.nr;
1298                         ena = ad.ena;
1299                         run = ad.run;
1300                         val = ad.val;
1301                         if (first && metric_only) {
1302                                 first = false;
1303                                 aggr_printout(counter, id, nr);
1304                         }
1305                         if (prefix && !metric_only)
1306                                 fprintf(output, "%s", prefix);
1307
1308                         uval = val * counter->scale;
1309                         printout(id, nr, counter, uval, prefix, run, ena, 1.0);
1310                         if (!metric_only)
1311                                 fputc('\n', output);
1312                 }
1313                 if (metric_only)
1314                         fputc('\n', output);
1315         }
1316 }
1317
1318 static void print_aggr_thread(struct perf_evsel *counter, char *prefix)
1319 {
1320         FILE *output = stat_config.output;
1321         int nthreads = thread_map__nr(counter->threads);
1322         int ncpus = cpu_map__nr(counter->cpus);
1323         int cpu, thread;
1324         double uval;
1325
1326         for (thread = 0; thread < nthreads; thread++) {
1327                 u64 ena = 0, run = 0, val = 0;
1328
1329                 for (cpu = 0; cpu < ncpus; cpu++) {
1330                         val += perf_counts(counter->counts, cpu, thread)->val;
1331                         ena += perf_counts(counter->counts, cpu, thread)->ena;
1332                         run += perf_counts(counter->counts, cpu, thread)->run;
1333                 }
1334
1335                 if (prefix)
1336                         fprintf(output, "%s", prefix);
1337
1338                 uval = val * counter->scale;
1339                 printout(thread, 0, counter, uval, prefix, run, ena, 1.0);
1340                 fputc('\n', output);
1341         }
1342 }
1343
1344 struct caggr_data {
1345         double avg, avg_enabled, avg_running;
1346 };
1347
1348 static void counter_aggr_cb(struct perf_evsel *counter, void *data,
1349                             bool first __maybe_unused)
1350 {
1351         struct caggr_data *cd = data;
1352         struct perf_stat_evsel *ps = counter->priv;
1353
1354         cd->avg += avg_stats(&ps->res_stats[0]);
1355         cd->avg_enabled += avg_stats(&ps->res_stats[1]);
1356         cd->avg_running += avg_stats(&ps->res_stats[2]);
1357 }
1358
1359 /*
1360  * Print out the results of a single counter:
1361  * aggregated counts in system-wide mode
1362  */
1363 static void print_counter_aggr(struct perf_evsel *counter, char *prefix)
1364 {
1365         FILE *output = stat_config.output;
1366         double uval;
1367         struct caggr_data cd = { .avg = 0.0 };
1368
1369         if (!collect_data(counter, counter_aggr_cb, &cd))
1370                 return;
1371
1372         if (prefix && !metric_only)
1373                 fprintf(output, "%s", prefix);
1374
1375         uval = cd.avg * counter->scale;
1376         printout(-1, 0, counter, uval, prefix, cd.avg_running, cd.avg_enabled, cd.avg);
1377         if (!metric_only)
1378                 fprintf(output, "\n");
1379 }
1380
1381 static void counter_cb(struct perf_evsel *counter, void *data,
1382                        bool first __maybe_unused)
1383 {
1384         struct aggr_data *ad = data;
1385
1386         ad->val += perf_counts(counter->counts, ad->cpu, 0)->val;
1387         ad->ena += perf_counts(counter->counts, ad->cpu, 0)->ena;
1388         ad->run += perf_counts(counter->counts, ad->cpu, 0)->run;
1389 }
1390
1391 /*
1392  * Print out the results of a single counter:
1393  * does not use aggregated count in system-wide
1394  */
1395 static void print_counter(struct perf_evsel *counter, char *prefix)
1396 {
1397         FILE *output = stat_config.output;
1398         u64 ena, run, val;
1399         double uval;
1400         int cpu;
1401
1402         for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
1403                 struct aggr_data ad = { .cpu = cpu };
1404
1405                 if (!collect_data(counter, counter_cb, &ad))
1406                         return;
1407                 val = ad.val;
1408                 ena = ad.ena;
1409                 run = ad.run;
1410
1411                 if (prefix)
1412                         fprintf(output, "%s", prefix);
1413
1414                 uval = val * counter->scale;
1415                 printout(cpu, 0, counter, uval, prefix, run, ena, 1.0);
1416
1417                 fputc('\n', output);
1418         }
1419 }
1420
1421 static void print_no_aggr_metric(char *prefix)
1422 {
1423         int cpu;
1424         int nrcpus = 0;
1425         struct perf_evsel *counter;
1426         u64 ena, run, val;
1427         double uval;
1428
1429         nrcpus = evsel_list->cpus->nr;
1430         for (cpu = 0; cpu < nrcpus; cpu++) {
1431                 bool first = true;
1432
1433                 if (prefix)
1434                         fputs(prefix, stat_config.output);
1435                 evlist__for_each_entry(evsel_list, counter) {
1436                         if (first) {
1437                                 aggr_printout(counter, cpu, 0);
1438                                 first = false;
1439                         }
1440                         val = perf_counts(counter->counts, cpu, 0)->val;
1441                         ena = perf_counts(counter->counts, cpu, 0)->ena;
1442                         run = perf_counts(counter->counts, cpu, 0)->run;
1443
1444                         uval = val * counter->scale;
1445                         printout(cpu, 0, counter, uval, prefix, run, ena, 1.0);
1446                 }
1447                 fputc('\n', stat_config.output);
1448         }
1449 }
1450
1451 static int aggr_header_lens[] = {
1452         [AGGR_CORE] = 18,
1453         [AGGR_SOCKET] = 12,
1454         [AGGR_NONE] = 6,
1455         [AGGR_THREAD] = 24,
1456         [AGGR_GLOBAL] = 0,
1457 };
1458
1459 static const char *aggr_header_csv[] = {
1460         [AGGR_CORE]     =       "core,cpus,",
1461         [AGGR_SOCKET]   =       "socket,cpus",
1462         [AGGR_NONE]     =       "cpu,",
1463         [AGGR_THREAD]   =       "comm-pid,",
1464         [AGGR_GLOBAL]   =       ""
1465 };
1466
1467 static void print_metric_headers(const char *prefix, bool no_indent)
1468 {
1469         struct perf_stat_output_ctx out;
1470         struct perf_evsel *counter;
1471         struct outstate os = {
1472                 .fh = stat_config.output
1473         };
1474
1475         if (prefix)
1476                 fprintf(stat_config.output, "%s", prefix);
1477
1478         if (!csv_output && !no_indent)
1479                 fprintf(stat_config.output, "%*s",
1480                         aggr_header_lens[stat_config.aggr_mode], "");
1481         if (csv_output) {
1482                 if (stat_config.interval)
1483                         fputs("time,", stat_config.output);
1484                 fputs(aggr_header_csv[stat_config.aggr_mode],
1485                         stat_config.output);
1486         }
1487
1488         /* Print metrics headers only */
1489         evlist__for_each_entry(evsel_list, counter) {
1490                 os.evsel = counter;
1491                 out.ctx = &os;
1492                 out.print_metric = print_metric_header;
1493                 out.new_line = new_line_metric;
1494                 out.force_header = true;
1495                 os.evsel = counter;
1496                 perf_stat__print_shadow_stats(counter, 0,
1497                                               0,
1498                                               &out);
1499         }
1500         fputc('\n', stat_config.output);
1501 }
1502
1503 static void print_interval(char *prefix, struct timespec *ts)
1504 {
1505         FILE *output = stat_config.output;
1506         static int num_print_interval;
1507
1508         sprintf(prefix, "%6lu.%09lu%s", ts->tv_sec, ts->tv_nsec, csv_sep);
1509
1510         if (num_print_interval == 0 && !csv_output) {
1511                 switch (stat_config.aggr_mode) {
1512                 case AGGR_SOCKET:
1513                         fprintf(output, "#           time socket cpus");
1514                         if (!metric_only)
1515                                 fprintf(output, "             counts %*s events\n", unit_width, "unit");
1516                         break;
1517                 case AGGR_CORE:
1518                         fprintf(output, "#           time core         cpus");
1519                         if (!metric_only)
1520                                 fprintf(output, "             counts %*s events\n", unit_width, "unit");
1521                         break;
1522                 case AGGR_NONE:
1523                         fprintf(output, "#           time CPU");
1524                         if (!metric_only)
1525                                 fprintf(output, "                counts %*s events\n", unit_width, "unit");
1526                         break;
1527                 case AGGR_THREAD:
1528                         fprintf(output, "#           time             comm-pid");
1529                         if (!metric_only)
1530                                 fprintf(output, "                  counts %*s events\n", unit_width, "unit");
1531                         break;
1532                 case AGGR_GLOBAL:
1533                 default:
1534                         fprintf(output, "#           time");
1535                         if (!metric_only)
1536                                 fprintf(output, "             counts %*s events\n", unit_width, "unit");
1537                 case AGGR_UNSET:
1538                         break;
1539                 }
1540         }
1541
1542         if (num_print_interval == 0 && metric_only)
1543                 print_metric_headers(" ", true);
1544         if (++num_print_interval == 25)
1545                 num_print_interval = 0;
1546 }
1547
1548 static void print_header(int argc, const char **argv)
1549 {
1550         FILE *output = stat_config.output;
1551         int i;
1552
1553         fflush(stdout);
1554
1555         if (!csv_output) {
1556                 fprintf(output, "\n");
1557                 fprintf(output, " Performance counter stats for ");
1558                 if (target.system_wide)
1559                         fprintf(output, "\'system wide");
1560                 else if (target.cpu_list)
1561                         fprintf(output, "\'CPU(s) %s", target.cpu_list);
1562                 else if (!target__has_task(&target)) {
1563                         fprintf(output, "\'%s", argv ? argv[0] : "pipe");
1564                         for (i = 1; argv && (i < argc); i++)
1565                                 fprintf(output, " %s", argv[i]);
1566                 } else if (target.pid)
1567                         fprintf(output, "process id \'%s", target.pid);
1568                 else
1569                         fprintf(output, "thread id \'%s", target.tid);
1570
1571                 fprintf(output, "\'");
1572                 if (run_count > 1)
1573                         fprintf(output, " (%d runs)", run_count);
1574                 fprintf(output, ":\n\n");
1575         }
1576 }
1577
1578 static void print_footer(void)
1579 {
1580         FILE *output = stat_config.output;
1581
1582         if (!null_run)
1583                 fprintf(output, "\n");
1584         fprintf(output, " %17.9f seconds time elapsed",
1585                         avg_stats(&walltime_nsecs_stats) / NSEC_PER_SEC);
1586         if (run_count > 1) {
1587                 fprintf(output, "                                        ");
1588                 print_noise_pct(stddev_stats(&walltime_nsecs_stats),
1589                                 avg_stats(&walltime_nsecs_stats));
1590         }
1591         fprintf(output, "\n\n");
1592
1593         if (print_free_counters_hint)
1594                 fprintf(output,
1595 "Some events weren't counted. Try disabling the NMI watchdog:\n"
1596 "       echo 0 > /proc/sys/kernel/nmi_watchdog\n"
1597 "       perf stat ...\n"
1598 "       echo 1 > /proc/sys/kernel/nmi_watchdog\n");
1599 }
1600
1601 static void print_counters(struct timespec *ts, int argc, const char **argv)
1602 {
1603         int interval = stat_config.interval;
1604         struct perf_evsel *counter;
1605         char buf[64], *prefix = NULL;
1606
1607         /* Do not print anything if we record to the pipe. */
1608         if (STAT_RECORD && perf_stat.file.is_pipe)
1609                 return;
1610
1611         if (interval)
1612                 print_interval(prefix = buf, ts);
1613         else
1614                 print_header(argc, argv);
1615
1616         if (metric_only) {
1617                 static int num_print_iv;
1618
1619                 if (num_print_iv == 0 && !interval)
1620                         print_metric_headers(prefix, false);
1621                 if (num_print_iv++ == 25)
1622                         num_print_iv = 0;
1623                 if (stat_config.aggr_mode == AGGR_GLOBAL && prefix)
1624                         fprintf(stat_config.output, "%s", prefix);
1625         }
1626
1627         switch (stat_config.aggr_mode) {
1628         case AGGR_CORE:
1629         case AGGR_SOCKET:
1630                 print_aggr(prefix);
1631                 break;
1632         case AGGR_THREAD:
1633                 evlist__for_each_entry(evsel_list, counter)
1634                         print_aggr_thread(counter, prefix);
1635                 break;
1636         case AGGR_GLOBAL:
1637                 evlist__for_each_entry(evsel_list, counter)
1638                         print_counter_aggr(counter, prefix);
1639                 if (metric_only)
1640                         fputc('\n', stat_config.output);
1641                 break;
1642         case AGGR_NONE:
1643                 if (metric_only)
1644                         print_no_aggr_metric(prefix);
1645                 else {
1646                         evlist__for_each_entry(evsel_list, counter)
1647                                 print_counter(counter, prefix);
1648                 }
1649                 break;
1650         case AGGR_UNSET:
1651         default:
1652                 break;
1653         }
1654
1655         if (!interval && !csv_output)
1656                 print_footer();
1657
1658         fflush(stat_config.output);
1659 }
1660
1661 static volatile int signr = -1;
1662
1663 static void skip_signal(int signo)
1664 {
1665         if ((child_pid == -1) || stat_config.interval)
1666                 done = 1;
1667
1668         signr = signo;
1669         /*
1670          * render child_pid harmless
1671          * won't send SIGTERM to a random
1672          * process in case of race condition
1673          * and fast PID recycling
1674          */
1675         child_pid = -1;
1676 }
1677
1678 static void sig_atexit(void)
1679 {
1680         sigset_t set, oset;
1681
1682         /*
1683          * avoid race condition with SIGCHLD handler
1684          * in skip_signal() which is modifying child_pid
1685          * goal is to avoid send SIGTERM to a random
1686          * process
1687          */
1688         sigemptyset(&set);
1689         sigaddset(&set, SIGCHLD);
1690         sigprocmask(SIG_BLOCK, &set, &oset);
1691
1692         if (child_pid != -1)
1693                 kill(child_pid, SIGTERM);
1694
1695         sigprocmask(SIG_SETMASK, &oset, NULL);
1696
1697         if (signr == -1)
1698                 return;
1699
1700         signal(signr, SIG_DFL);
1701         kill(getpid(), signr);
1702 }
1703
1704 static int stat__set_big_num(const struct option *opt __maybe_unused,
1705                              const char *s __maybe_unused, int unset)
1706 {
1707         big_num_opt = unset ? 0 : 1;
1708         return 0;
1709 }
1710
1711 static int enable_metric_only(const struct option *opt __maybe_unused,
1712                               const char *s __maybe_unused, int unset)
1713 {
1714         force_metric_only = true;
1715         metric_only = !unset;
1716         return 0;
1717 }
1718
1719 static const struct option stat_options[] = {
1720         OPT_BOOLEAN('T', "transaction", &transaction_run,
1721                     "hardware transaction statistics"),
1722         OPT_CALLBACK('e', "event", &evsel_list, "event",
1723                      "event selector. use 'perf list' to list available events",
1724                      parse_events_option),
1725         OPT_CALLBACK(0, "filter", &evsel_list, "filter",
1726                      "event filter", parse_filter),
1727         OPT_BOOLEAN('i', "no-inherit", &no_inherit,
1728                     "child tasks do not inherit counters"),
1729         OPT_STRING('p', "pid", &target.pid, "pid",
1730                    "stat events on existing process id"),
1731         OPT_STRING('t', "tid", &target.tid, "tid",
1732                    "stat events on existing thread id"),
1733         OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
1734                     "system-wide collection from all CPUs"),
1735         OPT_BOOLEAN('g', "group", &group,
1736                     "put the counters into a counter group"),
1737         OPT_BOOLEAN('c', "scale", &stat_config.scale, "scale/normalize counters"),
1738         OPT_INCR('v', "verbose", &verbose,
1739                     "be more verbose (show counter open errors, etc)"),
1740         OPT_INTEGER('r', "repeat", &run_count,
1741                     "repeat command and print average + stddev (max: 100, forever: 0)"),
1742         OPT_BOOLEAN('n', "null", &null_run,
1743                     "null run - dont start any counters"),
1744         OPT_INCR('d', "detailed", &detailed_run,
1745                     "detailed run - start a lot of events"),
1746         OPT_BOOLEAN('S', "sync", &sync_run,
1747                     "call sync() before starting a run"),
1748         OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
1749                            "print large numbers with thousands\' separators",
1750                            stat__set_big_num),
1751         OPT_STRING('C', "cpu", &target.cpu_list, "cpu",
1752                     "list of cpus to monitor in system-wide"),
1753         OPT_SET_UINT('A', "no-aggr", &stat_config.aggr_mode,
1754                     "disable CPU count aggregation", AGGR_NONE),
1755         OPT_BOOLEAN(0, "no-merge", &no_merge, "Do not merge identical named events"),
1756         OPT_STRING('x', "field-separator", &csv_sep, "separator",
1757                    "print counts with custom separator"),
1758         OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
1759                      "monitor event in cgroup name only", parse_cgroups),
1760         OPT_STRING('o', "output", &output_name, "file", "output file name"),
1761         OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
1762         OPT_INTEGER(0, "log-fd", &output_fd,
1763                     "log output to fd, instead of stderr"),
1764         OPT_STRING(0, "pre", &pre_cmd, "command",
1765                         "command to run prior to the measured command"),
1766         OPT_STRING(0, "post", &post_cmd, "command",
1767                         "command to run after to the measured command"),
1768         OPT_UINTEGER('I', "interval-print", &stat_config.interval,
1769                     "print counts at regular interval in ms (>= 10)"),
1770         OPT_SET_UINT(0, "per-socket", &stat_config.aggr_mode,
1771                      "aggregate counts per processor socket", AGGR_SOCKET),
1772         OPT_SET_UINT(0, "per-core", &stat_config.aggr_mode,
1773                      "aggregate counts per physical processor core", AGGR_CORE),
1774         OPT_SET_UINT(0, "per-thread", &stat_config.aggr_mode,
1775                      "aggregate counts per thread", AGGR_THREAD),
1776         OPT_UINTEGER('D', "delay", &initial_delay,
1777                      "ms to wait before starting measurement after program start"),
1778         OPT_CALLBACK_NOOPT(0, "metric-only", &metric_only, NULL,
1779                         "Only print computed metrics. No raw values", enable_metric_only),
1780         OPT_BOOLEAN(0, "topdown", &topdown_run,
1781                         "measure topdown level 1 statistics"),
1782         OPT_END()
1783 };
1784
1785 static int perf_stat__get_socket(struct cpu_map *map, int cpu)
1786 {
1787         return cpu_map__get_socket(map, cpu, NULL);
1788 }
1789
1790 static int perf_stat__get_core(struct cpu_map *map, int cpu)
1791 {
1792         return cpu_map__get_core(map, cpu, NULL);
1793 }
1794
1795 static int cpu_map__get_max(struct cpu_map *map)
1796 {
1797         int i, max = -1;
1798
1799         for (i = 0; i < map->nr; i++) {
1800                 if (map->map[i] > max)
1801                         max = map->map[i];
1802         }
1803
1804         return max;
1805 }
1806
1807 static struct cpu_map *cpus_aggr_map;
1808
1809 static int perf_stat__get_aggr(aggr_get_id_t get_id, struct cpu_map *map, int idx)
1810 {
1811         int cpu;
1812
1813         if (idx >= map->nr)
1814                 return -1;
1815
1816         cpu = map->map[idx];
1817
1818         if (cpus_aggr_map->map[cpu] == -1)
1819                 cpus_aggr_map->map[cpu] = get_id(map, idx);
1820
1821         return cpus_aggr_map->map[cpu];
1822 }
1823
1824 static int perf_stat__get_socket_cached(struct cpu_map *map, int idx)
1825 {
1826         return perf_stat__get_aggr(perf_stat__get_socket, map, idx);
1827 }
1828
1829 static int perf_stat__get_core_cached(struct cpu_map *map, int idx)
1830 {
1831         return perf_stat__get_aggr(perf_stat__get_core, map, idx);
1832 }
1833
1834 static int perf_stat_init_aggr_mode(void)
1835 {
1836         int nr;
1837
1838         switch (stat_config.aggr_mode) {
1839         case AGGR_SOCKET:
1840                 if (cpu_map__build_socket_map(evsel_list->cpus, &aggr_map)) {
1841                         perror("cannot build socket map");
1842                         return -1;
1843                 }
1844                 aggr_get_id = perf_stat__get_socket_cached;
1845                 break;
1846         case AGGR_CORE:
1847                 if (cpu_map__build_core_map(evsel_list->cpus, &aggr_map)) {
1848                         perror("cannot build core map");
1849                         return -1;
1850                 }
1851                 aggr_get_id = perf_stat__get_core_cached;
1852                 break;
1853         case AGGR_NONE:
1854         case AGGR_GLOBAL:
1855         case AGGR_THREAD:
1856         case AGGR_UNSET:
1857         default:
1858                 break;
1859         }
1860
1861         /*
1862          * The evsel_list->cpus is the base we operate on,
1863          * taking the highest cpu number to be the size of
1864          * the aggregation translate cpumap.
1865          */
1866         nr = cpu_map__get_max(evsel_list->cpus);
1867         cpus_aggr_map = cpu_map__empty_new(nr + 1);
1868         return cpus_aggr_map ? 0 : -ENOMEM;
1869 }
1870
1871 static void perf_stat__exit_aggr_mode(void)
1872 {
1873         cpu_map__put(aggr_map);
1874         cpu_map__put(cpus_aggr_map);
1875         aggr_map = NULL;
1876         cpus_aggr_map = NULL;
1877 }
1878
1879 static inline int perf_env__get_cpu(struct perf_env *env, struct cpu_map *map, int idx)
1880 {
1881         int cpu;
1882
1883         if (idx > map->nr)
1884                 return -1;
1885
1886         cpu = map->map[idx];
1887
1888         if (cpu >= env->nr_cpus_avail)
1889                 return -1;
1890
1891         return cpu;
1892 }
1893
1894 static int perf_env__get_socket(struct cpu_map *map, int idx, void *data)
1895 {
1896         struct perf_env *env = data;
1897         int cpu = perf_env__get_cpu(env, map, idx);
1898
1899         return cpu == -1 ? -1 : env->cpu[cpu].socket_id;
1900 }
1901
1902 static int perf_env__get_core(struct cpu_map *map, int idx, void *data)
1903 {
1904         struct perf_env *env = data;
1905         int core = -1, cpu = perf_env__get_cpu(env, map, idx);
1906
1907         if (cpu != -1) {
1908                 int socket_id = env->cpu[cpu].socket_id;
1909
1910                 /*
1911                  * Encode socket in upper 16 bits
1912                  * core_id is relative to socket, and
1913                  * we need a global id. So we combine
1914                  * socket + core id.
1915                  */
1916                 core = (socket_id << 16) | (env->cpu[cpu].core_id & 0xffff);
1917         }
1918
1919         return core;
1920 }
1921
1922 static int perf_env__build_socket_map(struct perf_env *env, struct cpu_map *cpus,
1923                                       struct cpu_map **sockp)
1924 {
1925         return cpu_map__build_map(cpus, sockp, perf_env__get_socket, env);
1926 }
1927
1928 static int perf_env__build_core_map(struct perf_env *env, struct cpu_map *cpus,
1929                                     struct cpu_map **corep)
1930 {
1931         return cpu_map__build_map(cpus, corep, perf_env__get_core, env);
1932 }
1933
1934 static int perf_stat__get_socket_file(struct cpu_map *map, int idx)
1935 {
1936         return perf_env__get_socket(map, idx, &perf_stat.session->header.env);
1937 }
1938
1939 static int perf_stat__get_core_file(struct cpu_map *map, int idx)
1940 {
1941         return perf_env__get_core(map, idx, &perf_stat.session->header.env);
1942 }
1943
1944 static int perf_stat_init_aggr_mode_file(struct perf_stat *st)
1945 {
1946         struct perf_env *env = &st->session->header.env;
1947
1948         switch (stat_config.aggr_mode) {
1949         case AGGR_SOCKET:
1950                 if (perf_env__build_socket_map(env, evsel_list->cpus, &aggr_map)) {
1951                         perror("cannot build socket map");
1952                         return -1;
1953                 }
1954                 aggr_get_id = perf_stat__get_socket_file;
1955                 break;
1956         case AGGR_CORE:
1957                 if (perf_env__build_core_map(env, evsel_list->cpus, &aggr_map)) {
1958                         perror("cannot build core map");
1959                         return -1;
1960                 }
1961                 aggr_get_id = perf_stat__get_core_file;
1962                 break;
1963         case AGGR_NONE:
1964         case AGGR_GLOBAL:
1965         case AGGR_THREAD:
1966         case AGGR_UNSET:
1967         default:
1968                 break;
1969         }
1970
1971         return 0;
1972 }
1973
1974 static int topdown_filter_events(const char **attr, char **str, bool use_group)
1975 {
1976         int off = 0;
1977         int i;
1978         int len = 0;
1979         char *s;
1980
1981         for (i = 0; attr[i]; i++) {
1982                 if (pmu_have_event("cpu", attr[i])) {
1983                         len += strlen(attr[i]) + 1;
1984                         attr[i - off] = attr[i];
1985                 } else
1986                         off++;
1987         }
1988         attr[i - off] = NULL;
1989
1990         *str = malloc(len + 1 + 2);
1991         if (!*str)
1992                 return -1;
1993         s = *str;
1994         if (i - off == 0) {
1995                 *s = 0;
1996                 return 0;
1997         }
1998         if (use_group)
1999                 *s++ = '{';
2000         for (i = 0; attr[i]; i++) {
2001                 strcpy(s, attr[i]);
2002                 s += strlen(s);
2003                 *s++ = ',';
2004         }
2005         if (use_group) {
2006                 s[-1] = '}';
2007                 *s = 0;
2008         } else
2009                 s[-1] = 0;
2010         return 0;
2011 }
2012
2013 __weak bool arch_topdown_check_group(bool *warn)
2014 {
2015         *warn = false;
2016         return false;
2017 }
2018
2019 __weak void arch_topdown_group_warn(void)
2020 {
2021 }
2022
2023 /*
2024  * Add default attributes, if there were no attributes specified or
2025  * if -d/--detailed, -d -d or -d -d -d is used:
2026  */
2027 static int add_default_attributes(void)
2028 {
2029         int err;
2030         struct perf_event_attr default_attrs0[] = {
2031
2032   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK              },
2033   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES        },
2034   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS          },
2035   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS             },
2036
2037   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES              },
2038 };
2039         struct perf_event_attr frontend_attrs[] = {
2040   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND },
2041 };
2042         struct perf_event_attr backend_attrs[] = {
2043   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND  },
2044 };
2045         struct perf_event_attr default_attrs1[] = {
2046   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS            },
2047   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS     },
2048   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES           },
2049
2050 };
2051
2052 /*
2053  * Detailed stats (-d), covering the L1 and last level data caches:
2054  */
2055         struct perf_event_attr detailed_attrs[] = {
2056
2057   { .type = PERF_TYPE_HW_CACHE,
2058     .config =
2059          PERF_COUNT_HW_CACHE_L1D                <<  0  |
2060         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
2061         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
2062
2063   { .type = PERF_TYPE_HW_CACHE,
2064     .config =
2065          PERF_COUNT_HW_CACHE_L1D                <<  0  |
2066         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
2067         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
2068
2069   { .type = PERF_TYPE_HW_CACHE,
2070     .config =
2071          PERF_COUNT_HW_CACHE_LL                 <<  0  |
2072         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
2073         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
2074
2075   { .type = PERF_TYPE_HW_CACHE,
2076     .config =
2077          PERF_COUNT_HW_CACHE_LL                 <<  0  |
2078         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
2079         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
2080 };
2081
2082 /*
2083  * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
2084  */
2085         struct perf_event_attr very_detailed_attrs[] = {
2086
2087   { .type = PERF_TYPE_HW_CACHE,
2088     .config =
2089          PERF_COUNT_HW_CACHE_L1I                <<  0  |
2090         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
2091         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
2092
2093   { .type = PERF_TYPE_HW_CACHE,
2094     .config =
2095          PERF_COUNT_HW_CACHE_L1I                <<  0  |
2096         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
2097         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
2098
2099   { .type = PERF_TYPE_HW_CACHE,
2100     .config =
2101          PERF_COUNT_HW_CACHE_DTLB               <<  0  |
2102         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
2103         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
2104
2105   { .type = PERF_TYPE_HW_CACHE,
2106     .config =
2107          PERF_COUNT_HW_CACHE_DTLB               <<  0  |
2108         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
2109         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
2110
2111   { .type = PERF_TYPE_HW_CACHE,
2112     .config =
2113          PERF_COUNT_HW_CACHE_ITLB               <<  0  |
2114         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
2115         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
2116
2117   { .type = PERF_TYPE_HW_CACHE,
2118     .config =
2119          PERF_COUNT_HW_CACHE_ITLB               <<  0  |
2120         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
2121         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
2122
2123 };
2124
2125 /*
2126  * Very, very detailed stats (-d -d -d), adding prefetch events:
2127  */
2128         struct perf_event_attr very_very_detailed_attrs[] = {
2129
2130   { .type = PERF_TYPE_HW_CACHE,
2131     .config =
2132          PERF_COUNT_HW_CACHE_L1D                <<  0  |
2133         (PERF_COUNT_HW_CACHE_OP_PREFETCH        <<  8) |
2134         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
2135
2136   { .type = PERF_TYPE_HW_CACHE,
2137     .config =
2138          PERF_COUNT_HW_CACHE_L1D                <<  0  |
2139         (PERF_COUNT_HW_CACHE_OP_PREFETCH        <<  8) |
2140         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
2141 };
2142
2143         /* Set attrs if no event is selected and !null_run: */
2144         if (null_run)
2145                 return 0;
2146
2147         if (transaction_run) {
2148                 if (pmu_have_event("cpu", "cycles-ct") &&
2149                     pmu_have_event("cpu", "el-start"))
2150                         err = parse_events(evsel_list, transaction_attrs, NULL);
2151                 else
2152                         err = parse_events(evsel_list, transaction_limited_attrs, NULL);
2153                 if (err) {
2154                         fprintf(stderr, "Cannot set up transaction events\n");
2155                         return -1;
2156                 }
2157                 return 0;
2158         }
2159
2160         if (topdown_run) {
2161                 char *str = NULL;
2162                 bool warn = false;
2163
2164                 if (stat_config.aggr_mode != AGGR_GLOBAL &&
2165                     stat_config.aggr_mode != AGGR_CORE) {
2166                         pr_err("top down event configuration requires --per-core mode\n");
2167                         return -1;
2168                 }
2169                 stat_config.aggr_mode = AGGR_CORE;
2170                 if (nr_cgroups || !target__has_cpu(&target)) {
2171                         pr_err("top down event configuration requires system-wide mode (-a)\n");
2172                         return -1;
2173                 }
2174
2175                 if (!force_metric_only)
2176                         metric_only = true;
2177                 if (topdown_filter_events(topdown_attrs, &str,
2178                                 arch_topdown_check_group(&warn)) < 0) {
2179                         pr_err("Out of memory\n");
2180                         return -1;
2181                 }
2182                 if (topdown_attrs[0] && str) {
2183                         if (warn)
2184                                 arch_topdown_group_warn();
2185                         err = parse_events(evsel_list, str, NULL);
2186                         if (err) {
2187                                 fprintf(stderr,
2188                                         "Cannot set up top down events %s: %d\n",
2189                                         str, err);
2190                                 free(str);
2191                                 return -1;
2192                         }
2193                 } else {
2194                         fprintf(stderr, "System does not support topdown\n");
2195                         return -1;
2196                 }
2197                 free(str);
2198         }
2199
2200         if (!evsel_list->nr_entries) {
2201                 if (target__has_cpu(&target))
2202                         default_attrs0[0].config = PERF_COUNT_SW_CPU_CLOCK;
2203
2204                 if (perf_evlist__add_default_attrs(evsel_list, default_attrs0) < 0)
2205                         return -1;
2206                 if (pmu_have_event("cpu", "stalled-cycles-frontend")) {
2207                         if (perf_evlist__add_default_attrs(evsel_list,
2208                                                 frontend_attrs) < 0)
2209                                 return -1;
2210                 }
2211                 if (pmu_have_event("cpu", "stalled-cycles-backend")) {
2212                         if (perf_evlist__add_default_attrs(evsel_list,
2213                                                 backend_attrs) < 0)
2214                                 return -1;
2215                 }
2216                 if (perf_evlist__add_default_attrs(evsel_list, default_attrs1) < 0)
2217                         return -1;
2218         }
2219
2220         /* Detailed events get appended to the event list: */
2221
2222         if (detailed_run <  1)
2223                 return 0;
2224
2225         /* Append detailed run extra attributes: */
2226         if (perf_evlist__add_default_attrs(evsel_list, detailed_attrs) < 0)
2227                 return -1;
2228
2229         if (detailed_run < 2)
2230                 return 0;
2231
2232         /* Append very detailed run extra attributes: */
2233         if (perf_evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0)
2234                 return -1;
2235
2236         if (detailed_run < 3)
2237                 return 0;
2238
2239         /* Append very, very detailed run extra attributes: */
2240         return perf_evlist__add_default_attrs(evsel_list, very_very_detailed_attrs);
2241 }
2242
2243 static const char * const stat_record_usage[] = {
2244         "perf stat record [<options>]",
2245         NULL,
2246 };
2247
2248 static void init_features(struct perf_session *session)
2249 {
2250         int feat;
2251
2252         for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++)
2253                 perf_header__set_feat(&session->header, feat);
2254
2255         perf_header__clear_feat(&session->header, HEADER_BUILD_ID);
2256         perf_header__clear_feat(&session->header, HEADER_TRACING_DATA);
2257         perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK);
2258         perf_header__clear_feat(&session->header, HEADER_AUXTRACE);
2259 }
2260
2261 static int __cmd_record(int argc, const char **argv)
2262 {
2263         struct perf_session *session;
2264         struct perf_data_file *file = &perf_stat.file;
2265
2266         argc = parse_options(argc, argv, stat_options, stat_record_usage,
2267                              PARSE_OPT_STOP_AT_NON_OPTION);
2268
2269         if (output_name)
2270                 file->path = output_name;
2271
2272         if (run_count != 1 || forever) {
2273                 pr_err("Cannot use -r option with perf stat record.\n");
2274                 return -1;
2275         }
2276
2277         session = perf_session__new(file, false, NULL);
2278         if (session == NULL) {
2279                 pr_err("Perf session creation failed.\n");
2280                 return -1;
2281         }
2282
2283         init_features(session);
2284
2285         session->evlist   = evsel_list;
2286         perf_stat.session = session;
2287         perf_stat.record  = true;
2288         return argc;
2289 }
2290
2291 static int process_stat_round_event(struct perf_tool *tool __maybe_unused,
2292                                     union perf_event *event,
2293                                     struct perf_session *session)
2294 {
2295         struct stat_round_event *stat_round = &event->stat_round;
2296         struct perf_evsel *counter;
2297         struct timespec tsh, *ts = NULL;
2298         const char **argv = session->header.env.cmdline_argv;
2299         int argc = session->header.env.nr_cmdline;
2300
2301         evlist__for_each_entry(evsel_list, counter)
2302                 perf_stat_process_counter(&stat_config, counter);
2303
2304         if (stat_round->type == PERF_STAT_ROUND_TYPE__FINAL)
2305                 update_stats(&walltime_nsecs_stats, stat_round->time);
2306
2307         if (stat_config.interval && stat_round->time) {
2308                 tsh.tv_sec  = stat_round->time / NSEC_PER_SEC;
2309                 tsh.tv_nsec = stat_round->time % NSEC_PER_SEC;
2310                 ts = &tsh;
2311         }
2312
2313         print_counters(ts, argc, argv);
2314         return 0;
2315 }
2316
2317 static
2318 int process_stat_config_event(struct perf_tool *tool,
2319                               union perf_event *event,
2320                               struct perf_session *session __maybe_unused)
2321 {
2322         struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2323
2324         perf_event__read_stat_config(&stat_config, &event->stat_config);
2325
2326         if (cpu_map__empty(st->cpus)) {
2327                 if (st->aggr_mode != AGGR_UNSET)
2328                         pr_warning("warning: processing task data, aggregation mode not set\n");
2329                 return 0;
2330         }
2331
2332         if (st->aggr_mode != AGGR_UNSET)
2333                 stat_config.aggr_mode = st->aggr_mode;
2334
2335         if (perf_stat.file.is_pipe)
2336                 perf_stat_init_aggr_mode();
2337         else
2338                 perf_stat_init_aggr_mode_file(st);
2339
2340         return 0;
2341 }
2342
2343 static int set_maps(struct perf_stat *st)
2344 {
2345         if (!st->cpus || !st->threads)
2346                 return 0;
2347
2348         if (WARN_ONCE(st->maps_allocated, "stats double allocation\n"))
2349                 return -EINVAL;
2350
2351         perf_evlist__set_maps(evsel_list, st->cpus, st->threads);
2352
2353         if (perf_evlist__alloc_stats(evsel_list, true))
2354                 return -ENOMEM;
2355
2356         st->maps_allocated = true;
2357         return 0;
2358 }
2359
2360 static
2361 int process_thread_map_event(struct perf_tool *tool,
2362                              union perf_event *event,
2363                              struct perf_session *session __maybe_unused)
2364 {
2365         struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2366
2367         if (st->threads) {
2368                 pr_warning("Extra thread map event, ignoring.\n");
2369                 return 0;
2370         }
2371
2372         st->threads = thread_map__new_event(&event->thread_map);
2373         if (!st->threads)
2374                 return -ENOMEM;
2375
2376         return set_maps(st);
2377 }
2378
2379 static
2380 int process_cpu_map_event(struct perf_tool *tool,
2381                           union perf_event *event,
2382                           struct perf_session *session __maybe_unused)
2383 {
2384         struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2385         struct cpu_map *cpus;
2386
2387         if (st->cpus) {
2388                 pr_warning("Extra cpu map event, ignoring.\n");
2389                 return 0;
2390         }
2391
2392         cpus = cpu_map__new_data(&event->cpu_map.data);
2393         if (!cpus)
2394                 return -ENOMEM;
2395
2396         st->cpus = cpus;
2397         return set_maps(st);
2398 }
2399
2400 static const char * const stat_report_usage[] = {
2401         "perf stat report [<options>]",
2402         NULL,
2403 };
2404
2405 static struct perf_stat perf_stat = {
2406         .tool = {
2407                 .attr           = perf_event__process_attr,
2408                 .event_update   = perf_event__process_event_update,
2409                 .thread_map     = process_thread_map_event,
2410                 .cpu_map        = process_cpu_map_event,
2411                 .stat_config    = process_stat_config_event,
2412                 .stat           = perf_event__process_stat_event,
2413                 .stat_round     = process_stat_round_event,
2414         },
2415         .aggr_mode = AGGR_UNSET,
2416 };
2417
2418 static int __cmd_report(int argc, const char **argv)
2419 {
2420         struct perf_session *session;
2421         const struct option options[] = {
2422         OPT_STRING('i', "input", &input_name, "file", "input file name"),
2423         OPT_SET_UINT(0, "per-socket", &perf_stat.aggr_mode,
2424                      "aggregate counts per processor socket", AGGR_SOCKET),
2425         OPT_SET_UINT(0, "per-core", &perf_stat.aggr_mode,
2426                      "aggregate counts per physical processor core", AGGR_CORE),
2427         OPT_SET_UINT('A', "no-aggr", &perf_stat.aggr_mode,
2428                      "disable CPU count aggregation", AGGR_NONE),
2429         OPT_END()
2430         };
2431         struct stat st;
2432         int ret;
2433
2434         argc = parse_options(argc, argv, options, stat_report_usage, 0);
2435
2436         if (!input_name || !strlen(input_name)) {
2437                 if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode))
2438                         input_name = "-";
2439                 else
2440                         input_name = "perf.data";
2441         }
2442
2443         perf_stat.file.path = input_name;
2444         perf_stat.file.mode = PERF_DATA_MODE_READ;
2445
2446         session = perf_session__new(&perf_stat.file, false, &perf_stat.tool);
2447         if (session == NULL)
2448                 return -1;
2449
2450         perf_stat.session  = session;
2451         stat_config.output = stderr;
2452         evsel_list         = session->evlist;
2453
2454         ret = perf_session__process_events(session);
2455         if (ret)
2456                 return ret;
2457
2458         perf_session__delete(session);
2459         return 0;
2460 }
2461
2462 static void setup_system_wide(int forks)
2463 {
2464         /*
2465          * Make system wide (-a) the default target if
2466          * no target was specified and one of following
2467          * conditions is met:
2468          *
2469          *   - there's no workload specified
2470          *   - there is workload specified but all requested
2471          *     events are system wide events
2472          */
2473         if (!target__none(&target))
2474                 return;
2475
2476         if (!forks)
2477                 target.system_wide = true;
2478         else {
2479                 struct perf_evsel *counter;
2480
2481                 evlist__for_each_entry(evsel_list, counter) {
2482                         if (!counter->system_wide)
2483                                 return;
2484                 }
2485
2486                 if (evsel_list->nr_entries)
2487                         target.system_wide = true;
2488         }
2489 }
2490
2491 int cmd_stat(int argc, const char **argv)
2492 {
2493         const char * const stat_usage[] = {
2494                 "perf stat [<options>] [<command>]",
2495                 NULL
2496         };
2497         int status = -EINVAL, run_idx;
2498         const char *mode;
2499         FILE *output = stderr;
2500         unsigned int interval;
2501         const char * const stat_subcommands[] = { "record", "report" };
2502
2503         setlocale(LC_ALL, "");
2504
2505         evsel_list = perf_evlist__new();
2506         if (evsel_list == NULL)
2507                 return -ENOMEM;
2508
2509         parse_events__shrink_config_terms();
2510         argc = parse_options_subcommand(argc, argv, stat_options, stat_subcommands,
2511                                         (const char **) stat_usage,
2512                                         PARSE_OPT_STOP_AT_NON_OPTION);
2513         perf_stat__collect_metric_expr(evsel_list);
2514         perf_stat__init_shadow_stats();
2515
2516         if (csv_sep) {
2517                 csv_output = true;
2518                 if (!strcmp(csv_sep, "\\t"))
2519                         csv_sep = "\t";
2520         } else
2521                 csv_sep = DEFAULT_SEPARATOR;
2522
2523         if (argc && !strncmp(argv[0], "rec", 3)) {
2524                 argc = __cmd_record(argc, argv);
2525                 if (argc < 0)
2526                         return -1;
2527         } else if (argc && !strncmp(argv[0], "rep", 3))
2528                 return __cmd_report(argc, argv);
2529
2530         interval = stat_config.interval;
2531
2532         /*
2533          * For record command the -o is already taken care of.
2534          */
2535         if (!STAT_RECORD && output_name && strcmp(output_name, "-"))
2536                 output = NULL;
2537
2538         if (output_name && output_fd) {
2539                 fprintf(stderr, "cannot use both --output and --log-fd\n");
2540                 parse_options_usage(stat_usage, stat_options, "o", 1);
2541                 parse_options_usage(NULL, stat_options, "log-fd", 0);
2542                 goto out;
2543         }
2544
2545         if (metric_only && stat_config.aggr_mode == AGGR_THREAD) {
2546                 fprintf(stderr, "--metric-only is not supported with --per-thread\n");
2547                 goto out;
2548         }
2549
2550         if (metric_only && run_count > 1) {
2551                 fprintf(stderr, "--metric-only is not supported with -r\n");
2552                 goto out;
2553         }
2554
2555         if (output_fd < 0) {
2556                 fprintf(stderr, "argument to --log-fd must be a > 0\n");
2557                 parse_options_usage(stat_usage, stat_options, "log-fd", 0);
2558                 goto out;
2559         }
2560
2561         if (!output) {
2562                 struct timespec tm;
2563                 mode = append_file ? "a" : "w";
2564
2565                 output = fopen(output_name, mode);
2566                 if (!output) {
2567                         perror("failed to create output file");
2568                         return -1;
2569                 }
2570                 clock_gettime(CLOCK_REALTIME, &tm);
2571                 fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
2572         } else if (output_fd > 0) {
2573                 mode = append_file ? "a" : "w";
2574                 output = fdopen(output_fd, mode);
2575                 if (!output) {
2576                         perror("Failed opening logfd");
2577                         return -errno;
2578                 }
2579         }
2580
2581         stat_config.output = output;
2582
2583         /*
2584          * let the spreadsheet do the pretty-printing
2585          */
2586         if (csv_output) {
2587                 /* User explicitly passed -B? */
2588                 if (big_num_opt == 1) {
2589                         fprintf(stderr, "-B option not supported with -x\n");
2590                         parse_options_usage(stat_usage, stat_options, "B", 1);
2591                         parse_options_usage(NULL, stat_options, "x", 1);
2592                         goto out;
2593                 } else /* Nope, so disable big number formatting */
2594                         big_num = false;
2595         } else if (big_num_opt == 0) /* User passed --no-big-num */
2596                 big_num = false;
2597
2598         setup_system_wide(argc);
2599
2600         if (run_count < 0) {
2601                 pr_err("Run count must be a positive number\n");
2602                 parse_options_usage(stat_usage, stat_options, "r", 1);
2603                 goto out;
2604         } else if (run_count == 0) {
2605                 forever = true;
2606                 run_count = 1;
2607         }
2608
2609         if ((stat_config.aggr_mode == AGGR_THREAD) && !target__has_task(&target)) {
2610                 fprintf(stderr, "The --per-thread option is only available "
2611                         "when monitoring via -p -t options.\n");
2612                 parse_options_usage(NULL, stat_options, "p", 1);
2613                 parse_options_usage(NULL, stat_options, "t", 1);
2614                 goto out;
2615         }
2616
2617         /*
2618          * no_aggr, cgroup are for system-wide only
2619          * --per-thread is aggregated per thread, we dont mix it with cpu mode
2620          */
2621         if (((stat_config.aggr_mode != AGGR_GLOBAL &&
2622               stat_config.aggr_mode != AGGR_THREAD) || nr_cgroups) &&
2623             !target__has_cpu(&target)) {
2624                 fprintf(stderr, "both cgroup and no-aggregation "
2625                         "modes only available in system-wide mode\n");
2626
2627                 parse_options_usage(stat_usage, stat_options, "G", 1);
2628                 parse_options_usage(NULL, stat_options, "A", 1);
2629                 parse_options_usage(NULL, stat_options, "a", 1);
2630                 goto out;
2631         }
2632
2633         if (add_default_attributes())
2634                 goto out;
2635
2636         target__validate(&target);
2637
2638         if (perf_evlist__create_maps(evsel_list, &target) < 0) {
2639                 if (target__has_task(&target)) {
2640                         pr_err("Problems finding threads of monitor\n");
2641                         parse_options_usage(stat_usage, stat_options, "p", 1);
2642                         parse_options_usage(NULL, stat_options, "t", 1);
2643                 } else if (target__has_cpu(&target)) {
2644                         perror("failed to parse CPUs map");
2645                         parse_options_usage(stat_usage, stat_options, "C", 1);
2646                         parse_options_usage(NULL, stat_options, "a", 1);
2647                 }
2648                 goto out;
2649         }
2650
2651         /*
2652          * Initialize thread_map with comm names,
2653          * so we could print it out on output.
2654          */
2655         if (stat_config.aggr_mode == AGGR_THREAD)
2656                 thread_map__read_comms(evsel_list->threads);
2657
2658         if (interval && interval < 100) {
2659                 if (interval < 10) {
2660                         pr_err("print interval must be >= 10ms\n");
2661                         parse_options_usage(stat_usage, stat_options, "I", 1);
2662                         goto out;
2663                 } else
2664                         pr_warning("print interval < 100ms. "
2665                                    "The overhead percentage could be high in some cases. "
2666                                    "Please proceed with caution.\n");
2667         }
2668
2669         if (perf_evlist__alloc_stats(evsel_list, interval))
2670                 goto out;
2671
2672         if (perf_stat_init_aggr_mode())
2673                 goto out;
2674
2675         /*
2676          * We dont want to block the signals - that would cause
2677          * child tasks to inherit that and Ctrl-C would not work.
2678          * What we want is for Ctrl-C to work in the exec()-ed
2679          * task, but being ignored by perf stat itself:
2680          */
2681         atexit(sig_atexit);
2682         if (!forever)
2683                 signal(SIGINT,  skip_signal);
2684         signal(SIGCHLD, skip_signal);
2685         signal(SIGALRM, skip_signal);
2686         signal(SIGABRT, skip_signal);
2687
2688         status = 0;
2689         for (run_idx = 0; forever || run_idx < run_count; run_idx++) {
2690                 if (run_count != 1 && verbose > 0)
2691                         fprintf(output, "[ perf stat: executing run #%d ... ]\n",
2692                                 run_idx + 1);
2693
2694                 status = run_perf_stat(argc, argv);
2695                 if (forever && status != -1) {
2696                         print_counters(NULL, argc, argv);
2697                         perf_stat__reset_stats();
2698                 }
2699         }
2700
2701         if (!forever && status != -1 && !interval)
2702                 print_counters(NULL, argc, argv);
2703
2704         if (STAT_RECORD) {
2705                 /*
2706                  * We synthesize the kernel mmap record just so that older tools
2707                  * don't emit warnings about not being able to resolve symbols
2708                  * due to /proc/sys/kernel/kptr_restrict settings and instear provide
2709                  * a saner message about no samples being in the perf.data file.
2710                  *
2711                  * This also serves to suppress a warning about f_header.data.size == 0
2712                  * in header.c at the moment 'perf stat record' gets introduced, which
2713                  * is not really needed once we start adding the stat specific PERF_RECORD_
2714                  * records, but the need to suppress the kptr_restrict messages in older
2715                  * tools remain  -acme
2716                  */
2717                 int fd = perf_data_file__fd(&perf_stat.file);
2718                 int err = perf_event__synthesize_kernel_mmap((void *)&perf_stat,
2719                                                              process_synthesized_event,
2720                                                              &perf_stat.session->machines.host);
2721                 if (err) {
2722                         pr_warning("Couldn't synthesize the kernel mmap record, harmless, "
2723                                    "older tools may produce warnings about this file\n.");
2724                 }
2725
2726                 if (!interval) {
2727                         if (WRITE_STAT_ROUND_EVENT(walltime_nsecs_stats.max, FINAL))
2728                                 pr_err("failed to write stat round event\n");
2729                 }
2730
2731                 if (!perf_stat.file.is_pipe) {
2732                         perf_stat.session->header.data_size += perf_stat.bytes_written;
2733                         perf_session__write_header(perf_stat.session, evsel_list, fd, true);
2734                 }
2735
2736                 perf_session__delete(perf_stat.session);
2737         }
2738
2739         perf_stat__exit_aggr_mode();
2740         perf_evlist__free_stats(evsel_list);
2741 out:
2742         perf_evlist__delete(evsel_list);
2743         return status;
2744 }