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