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