tools/perf/builtin-stat.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * builtin-stat.c
   4  *
   5  * Builtin stat command: Give a precise performance counters summary
   6  * overview about any workload, CPU or specific PID.
   7  *
   8  * Sample output:
   9
  10    $ perf stat ./hackbench 10
  11
  12   Time: 0.118
  13
  14   Performance counter stats for './hackbench 10':
  15
  16        1708.761321 task-clock                #   11.037 CPUs utilized
  17             41,190 context-switches          #    0.024 M/sec
  18              6,735 CPU-migrations            #    0.004 M/sec
  19             17,318 page-faults               #    0.010 M/sec
  20      5,205,202,243 cycles                    #    3.046 GHz
  21      3,856,436,920 stalled-cycles-frontend   #   74.09% frontend cycles idle
  22      1,600,790,871 stalled-cycles-backend    #   30.75% backend  cycles idle
  23      2,603,501,247 instructions              #    0.50  insns per cycle
  24                                              #    1.48  stalled cycles per insn
  25        484,357,498 branches                  #  283.455 M/sec
  26          6,388,934 branch-misses             #    1.32% of all branches
  27
  28         0.154822978  seconds time elapsed
  29
  30  *
  31  * Copyright (C) 2008-2011, Red Hat Inc, Ingo Molnar <mingo@redhat.com>
  32  *
  33  * Improvements and fixes by:
  34  *
  35  *   Arjan van de Ven <arjan@linux.intel.com>
  36  *   Yanmin Zhang <yanmin.zhang@intel.com>
  37  *   Wu Fengguang <fengguang.wu@intel.com>
  38  *   Mike Galbraith <efault@gmx.de>
  39  *   Paul Mackerras <paulus@samba.org>
  40  *   Jaswinder Singh Rajput <jaswinder@kernel.org>
  41  */
  42
  43 #include "builtin.h"
  44 #include "perf.h"
  45 #include "util/cgroup.h"
  46 #include <subcmd/parse-options.h>
  47 #include "util/parse-events.h"
  48 #include "util/pmu.h"
  49 #include "util/event.h"
  50 #include "util/evlist.h"
  51 #include "util/evlist-hybrid.h"
  52 #include "util/evsel.h"
  53 #include "util/debug.h"
  54 #include "util/color.h"
  55 #include "util/stat.h"
  56 #include "util/header.h"
  57 #include "util/cpumap.h"
  58 #include "util/thread_map.h"
  59 #include "util/counts.h"
  60 #include "util/topdown.h"
  61 #include "util/session.h"
  62 #include "util/tool.h"
  63 #include "util/string2.h"
  64 #include "util/metricgroup.h"
  65 #include "util/synthetic-events.h"
  66 #include "util/target.h"
  67 #include "util/time-utils.h"
  68 #include "util/top.h"
  69 #include "util/affinity.h"
  70 #include "util/pfm.h"
  71 #include "util/bpf_counter.h"
  72 #include "util/iostat.h"
  73 #include "util/pmu-hybrid.h"
  74 #include "asm/bug.h"
  75
  76 #include <linux/time64.h>
  77 #include <linux/zalloc.h>
  78 #include <api/fs/fs.h>
  79 #include <errno.h>
  80 #include <signal.h>
  81 #include <stdlib.h>
  82 #include <sys/prctl.h>
  83 #include <inttypes.h>
  84 #include <locale.h>
  85 #include <math.h>
  86 #include <sys/types.h>
  87 #include <sys/stat.h>
  88 #include <sys/wait.h>
  89 #include <unistd.h>
  90 #include <sys/time.h>
  91 #include <sys/resource.h>
  92 #include <linux/err.h>
  93
  94 #include <linux/ctype.h>
  95 #include <perf/evlist.h>
  96
  97 #define DEFAULT_SEPARATOR       " "
  98 #define FREEZE_ON_SMI_PATH      "devices/cpu/freeze_on_smi"
  99
 100 static void print_counters(struct timespec *ts, int argc, const char **argv);
 101
 102 /* Default events used for perf stat -T */
 103 static const char *transaction_attrs = {
 104         "task-clock,"
 105         "{"
 106         "instructions,"
 107         "cycles,"
 108         "cpu/cycles-t/,"
 109         "cpu/tx-start/,"
 110         "cpu/el-start/,"
 111         "cpu/cycles-ct/"
 112         "}"
 113 };
 114
 115 /* More limited version when the CPU does not have all events. */
 116 static const char * transaction_limited_attrs = {
 117         "task-clock,"
 118         "{"
 119         "instructions,"
 120         "cycles,"
 121         "cpu/cycles-t/,"
 122         "cpu/tx-start/"
 123         "}"
 124 };
 125
 126 static const char * topdown_attrs[] = {
 127         "topdown-total-slots",
 128         "topdown-slots-retired",
 129         "topdown-recovery-bubbles",
 130         "topdown-fetch-bubbles",
 131         "topdown-slots-issued",
 132         NULL,
 133 };
 134
 135 static const char *topdown_metric_attrs[] = {
 136         "slots",
 137         "topdown-retiring",
 138         "topdown-bad-spec",
 139         "topdown-fe-bound",
 140         "topdown-be-bound",
 141         NULL,
 142 };
 143
 144 static const char *topdown_metric_L2_attrs[] = {
 145         "slots",
 146         "topdown-retiring",
 147         "topdown-bad-spec",
 148         "topdown-fe-bound",
 149         "topdown-be-bound",
 150         "topdown-heavy-ops",
 151         "topdown-br-mispredict",
 152         "topdown-fetch-lat",
 153         "topdown-mem-bound",
 154         NULL,
 155 };
 156
 157 #define TOPDOWN_MAX_LEVEL                       2
 158
 159 static const char *smi_cost_attrs = {
 160         "{"
 161         "msr/aperf/,"
 162         "msr/smi/,"
 163         "cycles"
 164         "}"
 165 };
 166
 167 static struct evlist    *evsel_list;
 168 static bool all_counters_use_bpf = true;
 169
 170 static struct target target = {
 171         .uid    = UINT_MAX,
 172 };
 173
 174 #define METRIC_ONLY_LEN 20
 175
 176 static volatile pid_t           child_pid                       = -1;
 177 static int                      detailed_run                    =  0;
 178 static bool                     transaction_run;
 179 static bool                     topdown_run                     = false;
 180 static bool                     smi_cost                        = false;
 181 static bool                     smi_reset                       = false;
 182 static int                      big_num_opt                     =  -1;
 183 static bool                     group                           = false;
 184 static const char               *pre_cmd                        = NULL;
 185 static const char               *post_cmd                       = NULL;
 186 static bool                     sync_run                        = false;
 187 static bool                     forever                         = false;
 188 static bool                     force_metric_only               = false;
 189 static struct timespec          ref_time;
 190 static bool                     append_file;
 191 static bool                     interval_count;
 192 static const char               *output_name;
 193 static int                      output_fd;
 194
 195 struct perf_stat {
 196         bool                     record;
 197         struct perf_data         data;
 198         struct perf_session     *session;
 199         u64                      bytes_written;
 200         struct perf_tool         tool;
 201         bool                     maps_allocated;
 202         struct perf_cpu_map     *cpus;
 203         struct perf_thread_map *threads;
 204         enum aggr_mode           aggr_mode;
 205 };
 206
 207 static struct perf_stat         perf_stat;
 208 #define STAT_RECORD             perf_stat.record
 209
 210 static volatile int done = 0;
 211
 212 static struct perf_stat_config stat_config = {
 213         .aggr_mode              = AGGR_GLOBAL,
 214         .scale                  = true,
 215         .unit_width             = 4, /* strlen("unit") */
 216         .run_count              = 1,
 217         .metric_only_len        = METRIC_ONLY_LEN,
 218         .walltime_nsecs_stats   = &walltime_nsecs_stats,
 219         .big_num                = true,
 220         .ctl_fd                 = -1,
 221         .ctl_fd_ack             = -1,
 222         .iostat_run             = false,
 223 };
 224
 225 static bool cpus_map_matched(struct evsel *a, struct evsel *b)
 226 {
 227         if (!a->core.cpus && !b->core.cpus)
 228                 return true;
 229
 230         if (!a->core.cpus || !b->core.cpus)
 231                 return false;
 232
 233         if (a->core.cpus->nr != b->core.cpus->nr)
 234                 return false;
 235
 236         for (int i = 0; i < a->core.cpus->nr; i++) {
 237                 if (a->core.cpus->map[i] != b->core.cpus->map[i])
 238                         return false;
 239         }
 240
 241         return true;
 242 }
 243
 244 static void evlist__check_cpu_maps(struct evlist *evlist)
 245 {
 246         struct evsel *evsel, *pos, *leader;
 247         char buf[1024];
 248
 249         if (evlist__has_hybrid(evlist))
 250                 evlist__warn_hybrid_group(evlist);
 251
 252         evlist__for_each_entry(evlist, evsel) {
 253                 leader = evsel__leader(evsel);
 254
 255                 /* Check that leader matches cpus with each member. */
 256                 if (leader == evsel)
 257                         continue;
 258                 if (cpus_map_matched(leader, evsel))
 259                         continue;
 260
 261                 /* If there's mismatch disable the group and warn user. */
 262                 WARN_ONCE(1, "WARNING: grouped events cpus do not match, disabling group:\n");
 263                 evsel__group_desc(leader, buf, sizeof(buf));
 264                 pr_warning("  %s\n", buf);
 265
 266                 if (verbose) {
 267                         cpu_map__snprint(leader->core.cpus, buf, sizeof(buf));
 268                         pr_warning("     %s: %s\n", leader->name, buf);
 269                         cpu_map__snprint(evsel->core.cpus, buf, sizeof(buf));
 270                         pr_warning("     %s: %s\n", evsel->name, buf);
 271                 }
 272
 273                 for_each_group_evsel(pos, leader) {
 274                         evsel__set_leader(pos, pos);
 275                         pos->core.nr_members = 0;
 276                 }
 277                 evsel->core.leader->nr_members = 0;
 278         }
 279 }
 280
 281 static inline void diff_timespec(struct timespec *r, struct timespec *a,
 282                                  struct timespec *b)
 283 {
 284         r->tv_sec = a->tv_sec - b->tv_sec;
 285         if (a->tv_nsec < b->tv_nsec) {
 286                 r->tv_nsec = a->tv_nsec + NSEC_PER_SEC - b->tv_nsec;
 287                 r->tv_sec--;
 288         } else {
 289                 r->tv_nsec = a->tv_nsec - b->tv_nsec ;
 290         }
 291 }
 292
 293 static void perf_stat__reset_stats(void)
 294 {
 295         int i;
 296
 297         evlist__reset_stats(evsel_list);
 298         perf_stat__reset_shadow_stats();
 299
 300         for (i = 0; i < stat_config.stats_num; i++)
 301                 perf_stat__reset_shadow_per_stat(&stat_config.stats[i]);
 302 }
 303
 304 static int process_synthesized_event(struct perf_tool *tool __maybe_unused,
 305                                      union perf_event *event,
 306                                      struct perf_sample *sample __maybe_unused,
 307                                      struct machine *machine __maybe_unused)
 308 {
 309         if (perf_data__write(&perf_stat.data, event, event->header.size) < 0) {
 310                 pr_err("failed to write perf data, error: %m\n");
 311                 return -1;
 312         }
 313
 314         perf_stat.bytes_written += event->header.size;
 315         return 0;
 316 }
 317
 318 static int write_stat_round_event(u64 tm, u64 type)
 319 {
 320         return perf_event__synthesize_stat_round(NULL, tm, type,
 321                                                  process_synthesized_event,
 322                                                  NULL);
 323 }
 324
 325 #define WRITE_STAT_ROUND_EVENT(time, interval) \
 326         write_stat_round_event(time, PERF_STAT_ROUND_TYPE__ ## interval)
 327
 328 #define SID(e, x, y) xyarray__entry(e->core.sample_id, x, y)
 329
 330 static int evsel__write_stat_event(struct evsel *counter, u32 cpu, u32 thread,
 331                                    struct perf_counts_values *count)
 332 {
 333         struct perf_sample_id *sid = SID(counter, cpu, thread);
 334
 335         return perf_event__synthesize_stat(NULL, cpu, thread, sid->id, count,
 336                                            process_synthesized_event, NULL);
 337 }
 338
 339 static int read_single_counter(struct evsel *counter, int cpu,
 340                                int thread, struct timespec *rs)
 341 {
 342         if (counter->tool_event == PERF_TOOL_DURATION_TIME) {
 343                 u64 val = rs->tv_nsec + rs->tv_sec*1000000000ULL;
 344                 struct perf_counts_values *count =
 345                         perf_counts(counter->counts, cpu, thread);
 346                 count->ena = count->run = val;
 347                 count->val = val;
 348                 return 0;
 349         }
 350         return evsel__read_counter(counter, cpu, thread);
 351 }
 352
 353 /*
 354  * Read out the results of a single counter:
 355  * do not aggregate counts across CPUs in system-wide mode
 356  */
 357 static int read_counter_cpu(struct evsel *counter, struct timespec *rs, int cpu)
 358 {
 359         int nthreads = perf_thread_map__nr(evsel_list->core.threads);
 360         int thread;
 361
 362         if (!counter->supported)
 363                 return -ENOENT;
 364
 365         if (counter->core.system_wide)
 366                 nthreads = 1;
 367
 368         for (thread = 0; thread < nthreads; thread++) {
 369                 struct perf_counts_values *count;
 370
 371                 count = perf_counts(counter->counts, cpu, thread);
 372
 373                 /*
 374                  * The leader's group read loads data into its group members
 375                  * (via evsel__read_counter()) and sets their count->loaded.
 376                  */
 377                 if (!perf_counts__is_loaded(counter->counts, cpu, thread) &&
 378                     read_single_counter(counter, cpu, thread, rs)) {
 379                         counter->counts->scaled = -1;
 380                         perf_counts(counter->counts, cpu, thread)->ena = 0;
 381                         perf_counts(counter->counts, cpu, thread)->run = 0;
 382                         return -1;
 383                 }
 384
 385                 perf_counts__set_loaded(counter->counts, cpu, thread, false);
 386
 387                 if (STAT_RECORD) {
 388                         if (evsel__write_stat_event(counter, cpu, thread, count)) {
 389                                 pr_err("failed to write stat event\n");
 390                                 return -1;
 391                         }
 392                 }
 393
 394                 if (verbose > 1) {
 395                         fprintf(stat_config.output,
 396                                 "%s: %d: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
 397                                         evsel__name(counter),
 398                                         cpu,
 399                                         count->val, count->ena, count->run);
 400                 }
 401         }
 402
 403         return 0;
 404 }
 405
 406 static int read_affinity_counters(struct timespec *rs)
 407 {
 408         struct evsel *counter;
 409         struct affinity affinity;
 410         int i, ncpus, cpu;
 411
 412         if (all_counters_use_bpf)
 413                 return 0;
 414
 415         if (affinity__setup(&affinity) < 0)
 416                 return -1;
 417
 418         ncpus = perf_cpu_map__nr(evsel_list->core.all_cpus);
 419         if (!target__has_cpu(&target) || target__has_per_thread(&target))
 420                 ncpus = 1;
 421         evlist__for_each_cpu(evsel_list, i, cpu) {
 422                 if (i >= ncpus)
 423                         break;
 424                 affinity__set(&affinity, cpu);
 425
 426                 evlist__for_each_entry(evsel_list, counter) {
 427                         if (evsel__cpu_iter_skip(counter, cpu))
 428                                 continue;
 429                         if (evsel__is_bpf(counter))
 430                                 continue;
 431                         if (!counter->err) {
 432                                 counter->err = read_counter_cpu(counter, rs,
 433                                                                 counter->cpu_iter - 1);
 434                         }
 435                 }
 436         }
 437         affinity__cleanup(&affinity);
 438         return 0;
 439 }
 440
 441 static int read_bpf_map_counters(void)
 442 {
 443         struct evsel *counter;
 444         int err;
 445
 446         evlist__for_each_entry(evsel_list, counter) {
 447                 if (!evsel__is_bpf(counter))
 448                         continue;
 449
 450                 err = bpf_counter__read(counter);
 451                 if (err)
 452                         return err;
 453         }
 454         return 0;
 455 }
 456
 457 static void read_counters(struct timespec *rs)
 458 {
 459         struct evsel *counter;
 460
 461         if (!stat_config.stop_read_counter) {
 462                 if (read_bpf_map_counters() ||
 463                     read_affinity_counters(rs))
 464                         return;
 465         }
 466
 467         evlist__for_each_entry(evsel_list, counter) {
 468                 if (counter->err)
 469                         pr_debug("failed to read counter %s\n", counter->name);
 470                 if (counter->err == 0 && perf_stat_process_counter(&stat_config, counter))
 471                         pr_warning("failed to process counter %s\n", counter->name);
 472                 counter->err = 0;
 473         }
 474 }
 475
 476 static int runtime_stat_new(struct perf_stat_config *config, int nthreads)
 477 {
 478         int i;
 479
 480         config->stats = calloc(nthreads, sizeof(struct runtime_stat));
 481         if (!config->stats)
 482                 return -1;
 483
 484         config->stats_num = nthreads;
 485
 486         for (i = 0; i < nthreads; i++)
 487                 runtime_stat__init(&config->stats[i]);
 488
 489         return 0;
 490 }
 491
 492 static void runtime_stat_delete(struct perf_stat_config *config)
 493 {
 494         int i;
 495
 496         if (!config->stats)
 497                 return;
 498
 499         for (i = 0; i < config->stats_num; i++)
 500                 runtime_stat__exit(&config->stats[i]);
 501
 502         zfree(&config->stats);
 503 }
 504
 505 static void runtime_stat_reset(struct perf_stat_config *config)
 506 {
 507         int i;
 508
 509         if (!config->stats)
 510                 return;
 511
 512         for (i = 0; i < config->stats_num; i++)
 513                 perf_stat__reset_shadow_per_stat(&config->stats[i]);
 514 }
 515
 516 static void process_interval(void)
 517 {
 518         struct timespec ts, rs;
 519
 520         clock_gettime(CLOCK_MONOTONIC, &ts);
 521         diff_timespec(&rs, &ts, &ref_time);
 522
 523         perf_stat__reset_shadow_per_stat(&rt_stat);
 524         runtime_stat_reset(&stat_config);
 525         read_counters(&rs);
 526
 527         if (STAT_RECORD) {
 528                 if (WRITE_STAT_ROUND_EVENT(rs.tv_sec * NSEC_PER_SEC + rs.tv_nsec, INTERVAL))
 529                         pr_err("failed to write stat round event\n");
 530         }
 531
 532         init_stats(&walltime_nsecs_stats);
 533         update_stats(&walltime_nsecs_stats, stat_config.interval * 1000000ULL);
 534         print_counters(&rs, 0, NULL);
 535 }
 536
 537 static bool handle_interval(unsigned int interval, int *times)
 538 {
 539         if (interval) {
 540                 process_interval();
 541                 if (interval_count && !(--(*times)))
 542                         return true;
 543         }
 544         return false;
 545 }
 546
 547 static int enable_counters(void)
 548 {
 549         struct evsel *evsel;
 550         int err;
 551
 552         evlist__for_each_entry(evsel_list, evsel) {
 553                 if (!evsel__is_bpf(evsel))
 554                         continue;
 555
 556                 err = bpf_counter__enable(evsel);
 557                 if (err)
 558                         return err;
 559         }
 560
 561         if (stat_config.initial_delay < 0) {
 562                 pr_info(EVLIST_DISABLED_MSG);
 563                 return 0;
 564         }
 565
 566         if (stat_config.initial_delay > 0) {
 567                 pr_info(EVLIST_DISABLED_MSG);
 568                 usleep(stat_config.initial_delay * USEC_PER_MSEC);
 569         }
 570
 571         /*
 572          * We need to enable counters only if:
 573          * - we don't have tracee (attaching to task or cpu)
 574          * - we have initial delay configured
 575          */
 576         if (!target__none(&target) || stat_config.initial_delay) {
 577                 if (!all_counters_use_bpf)
 578                         evlist__enable(evsel_list);
 579                 if (stat_config.initial_delay > 0)
 580                         pr_info(EVLIST_ENABLED_MSG);
 581         }
 582         return 0;
 583 }
 584
 585 static void disable_counters(void)
 586 {
 587         struct evsel *counter;
 588
 589         /*
 590          * If we don't have tracee (attaching to task or cpu), counters may
 591          * still be running. To get accurate group ratios, we must stop groups
 592          * from counting before reading their constituent counters.
 593          */
 594         if (!target__none(&target)) {
 595                 evlist__for_each_entry(evsel_list, counter)
 596                         bpf_counter__disable(counter);
 597                 if (!all_counters_use_bpf)
 598                         evlist__disable(evsel_list);
 599         }
 600 }
 601
 602 static volatile int workload_exec_errno;
 603
 604 /*
 605  * evlist__prepare_workload will send a SIGUSR1
 606  * if the fork fails, since we asked by setting its
 607  * want_signal to true.
 608  */
 609 static void workload_exec_failed_signal(int signo __maybe_unused, siginfo_t *info,
 610                                         void *ucontext __maybe_unused)
 611 {
 612         workload_exec_errno = info->si_value.sival_int;
 613 }
 614
 615 static bool evsel__should_store_id(struct evsel *counter)
 616 {
 617         return STAT_RECORD || counter->core.attr.read_format & PERF_FORMAT_ID;
 618 }
 619
 620 static bool is_target_alive(struct target *_target,
 621                             struct perf_thread_map *threads)
 622 {
 623         struct stat st;
 624         int i;
 625
 626         if (!target__has_task(_target))
 627                 return true;
 628
 629         for (i = 0; i < threads->nr; i++) {
 630                 char path[PATH_MAX];
 631
 632                 scnprintf(path, PATH_MAX, "%s/%d", procfs__mountpoint(),
 633                           threads->map[i].pid);
 634
 635                 if (!stat(path, &st))
 636                         return true;
 637         }
 638
 639         return false;
 640 }
 641
 642 static void process_evlist(struct evlist *evlist, unsigned int interval)
 643 {
 644         enum evlist_ctl_cmd cmd = EVLIST_CTL_CMD_UNSUPPORTED;
 645
 646         if (evlist__ctlfd_process(evlist, &cmd) > 0) {
 647                 switch (cmd) {
 648                 case EVLIST_CTL_CMD_ENABLE:
 649                         if (interval)
 650                                 process_interval();
 651                         break;
 652                 case EVLIST_CTL_CMD_DISABLE:
 653                         if (interval)
 654                                 process_interval();
 655                         break;
 656                 case EVLIST_CTL_CMD_SNAPSHOT:
 657                 case EVLIST_CTL_CMD_ACK:
 658                 case EVLIST_CTL_CMD_UNSUPPORTED:
 659                 case EVLIST_CTL_CMD_EVLIST:
 660                 case EVLIST_CTL_CMD_STOP:
 661                 case EVLIST_CTL_CMD_PING:
 662                 default:
 663                         break;
 664                 }
 665         }
 666 }
 667
 668 static void compute_tts(struct timespec *time_start, struct timespec *time_stop,
 669                         int *time_to_sleep)
 670 {
 671         int tts = *time_to_sleep;
 672         struct timespec time_diff;
 673
 674         diff_timespec(&time_diff, time_stop, time_start);
 675
 676         tts -= time_diff.tv_sec * MSEC_PER_SEC +
 677                time_diff.tv_nsec / NSEC_PER_MSEC;
 678
 679         if (tts < 0)
 680                 tts = 0;
 681
 682         *time_to_sleep = tts;
 683 }
 684
 685 static int dispatch_events(bool forks, int timeout, int interval, int *times)
 686 {
 687         int child_exited = 0, status = 0;
 688         int time_to_sleep, sleep_time;
 689         struct timespec time_start, time_stop;
 690
 691         if (interval)
 692                 sleep_time = interval;
 693         else if (timeout)
 694                 sleep_time = timeout;
 695         else
 696                 sleep_time = 1000;
 697
 698         time_to_sleep = sleep_time;
 699
 700         while (!done) {
 701                 if (forks)
 702                         child_exited = waitpid(child_pid, &status, WNOHANG);
 703                 else
 704                         child_exited = !is_target_alive(&target, evsel_list->core.threads) ? 1 : 0;
 705
 706                 if (child_exited)
 707                         break;
 708
 709                 clock_gettime(CLOCK_MONOTONIC, &time_start);
 710                 if (!(evlist__poll(evsel_list, time_to_sleep) > 0)) { /* poll timeout or EINTR */
 711                         if (timeout || handle_interval(interval, times))
 712                                 break;
 713                         time_to_sleep = sleep_time;
 714                 } else { /* fd revent */
 715                         process_evlist(evsel_list, interval);
 716                         clock_gettime(CLOCK_MONOTONIC, &time_stop);
 717                         compute_tts(&time_start, &time_stop, &time_to_sleep);
 718                 }
 719         }
 720
 721         return status;
 722 }
 723
 724 enum counter_recovery {
 725         COUNTER_SKIP,
 726         COUNTER_RETRY,
 727         COUNTER_FATAL,
 728 };
 729
 730 static enum counter_recovery stat_handle_error(struct evsel *counter)
 731 {
 732         char msg[BUFSIZ];
 733         /*
 734          * PPC returns ENXIO for HW counters until 2.6.37
 735          * (behavior changed with commit b0a873e).
 736          */
 737         if (errno == EINVAL || errno == ENOSYS ||
 738             errno == ENOENT || errno == EOPNOTSUPP ||
 739             errno == ENXIO) {
 740                 if (verbose > 0)
 741                         ui__warning("%s event is not supported by the kernel.\n",
 742                                     evsel__name(counter));
 743                 counter->supported = false;
 744                 /*
 745                  * errored is a sticky flag that means one of the counter's
 746                  * cpu event had a problem and needs to be reexamined.
 747                  */
 748                 counter->errored = true;
 749
 750                 if ((evsel__leader(counter) != counter) ||
 751                     !(counter->core.leader->nr_members > 1))
 752                         return COUNTER_SKIP;
 753         } else if (evsel__fallback(counter, errno, msg, sizeof(msg))) {
 754                 if (verbose > 0)
 755                         ui__warning("%s\n", msg);
 756                 return COUNTER_RETRY;
 757         } else if (target__has_per_thread(&target) &&
 758                    evsel_list->core.threads &&
 759                    evsel_list->core.threads->err_thread != -1) {
 760                 /*
 761                  * For global --per-thread case, skip current
 762                  * error thread.
 763                  */
 764                 if (!thread_map__remove(evsel_list->core.threads,
 765                                         evsel_list->core.threads->err_thread)) {
 766                         evsel_list->core.threads->err_thread = -1;
 767                         return COUNTER_RETRY;
 768                 }
 769         }
 770
 771         evsel__open_strerror(counter, &target, errno, msg, sizeof(msg));
 772         ui__error("%s\n", msg);
 773
 774         if (child_pid != -1)
 775                 kill(child_pid, SIGTERM);
 776         return COUNTER_FATAL;
 777 }
 778
 779 static int __run_perf_stat(int argc, const char **argv, int run_idx)
 780 {
 781         int interval = stat_config.interval;
 782         int times = stat_config.times;
 783         int timeout = stat_config.timeout;
 784         char msg[BUFSIZ];
 785         unsigned long long t0, t1;
 786         struct evsel *counter;
 787         size_t l;
 788         int status = 0;
 789         const bool forks = (argc > 0);
 790         bool is_pipe = STAT_RECORD ? perf_stat.data.is_pipe : false;
 791         struct affinity affinity;
 792         int i, cpu, err;
 793         bool second_pass = false;
 794
 795         if (forks) {
 796                 if (evlist__prepare_workload(evsel_list, &target, argv, is_pipe, workload_exec_failed_signal) < 0) {
 797                         perror("failed to prepare workload");
 798                         return -1;
 799                 }
 800                 child_pid = evsel_list->workload.pid;
 801         }
 802
 803         if (group)
 804                 evlist__set_leader(evsel_list);
 805
 806         if (affinity__setup(&affinity) < 0)
 807                 return -1;
 808
 809         evlist__for_each_entry(evsel_list, counter) {
 810                 counter->reset_group = false;
 811                 if (bpf_counter__load(counter, &target))
 812                         return -1;
 813                 if (!evsel__is_bpf(counter))
 814                         all_counters_use_bpf = false;
 815         }
 816
 817         evlist__for_each_cpu (evsel_list, i, cpu) {
 818                 /*
 819                  * bperf calls evsel__open_per_cpu() in bperf__load(), so
 820                  * no need to call it again here.
 821                  */
 822                 if (target.use_bpf)
 823                         break;
 824                 affinity__set(&affinity, cpu);
 825
 826                 evlist__for_each_entry(evsel_list, counter) {
 827                         if (evsel__cpu_iter_skip(counter, cpu))
 828                                 continue;
 829                         if (counter->reset_group || counter->errored)
 830                                 continue;
 831                         if (evsel__is_bpf(counter))
 832                                 continue;
 833 try_again:
 834                         if (create_perf_stat_counter(counter, &stat_config, &target,
 835                                                      counter->cpu_iter - 1) < 0) {
 836
 837                                 /*
 838                                  * Weak group failed. We cannot just undo this here
 839                                  * because earlier CPUs might be in group mode, and the kernel
 840                                  * doesn't support mixing group and non group reads. Defer
 841                                  * it to later.
 842                                  * Don't close here because we're in the wrong affinity.
 843                                  */
 844                                 if ((errno == EINVAL || errno == EBADF) &&
 845                                     evsel__leader(counter) != counter &&
 846                                     counter->weak_group) {
 847                                         evlist__reset_weak_group(evsel_list, counter, false);
 848                                         assert(counter->reset_group);
 849                                         second_pass = true;
 850                                         continue;
 851                                 }
 852
 853                                 switch (stat_handle_error(counter)) {
 854                                 case COUNTER_FATAL:
 855                                         return -1;
 856                                 case COUNTER_RETRY:
 857                                         goto try_again;
 858                                 case COUNTER_SKIP:
 859                                         continue;
 860                                 default:
 861                                         break;
 862                                 }
 863
 864                         }
 865                         counter->supported = true;
 866                 }
 867         }
 868
 869         if (second_pass) {
 870                 /*
 871                  * Now redo all the weak group after closing them,
 872                  * and also close errored counters.
 873                  */
 874
 875                 evlist__for_each_cpu(evsel_list, i, cpu) {
 876                         affinity__set(&affinity, cpu);
 877                         /* First close errored or weak retry */
 878                         evlist__for_each_entry(evsel_list, counter) {
 879                                 if (!counter->reset_group && !counter->errored)
 880                                         continue;
 881                                 if (evsel__cpu_iter_skip_no_inc(counter, cpu))
 882                                         continue;
 883                                 perf_evsel__close_cpu(&counter->core, counter->cpu_iter);
 884                         }
 885                         /* Now reopen weak */
 886                         evlist__for_each_entry(evsel_list, counter) {
 887                                 if (!counter->reset_group && !counter->errored)
 888                                         continue;
 889                                 if (evsel__cpu_iter_skip(counter, cpu))
 890                                         continue;
 891                                 if (!counter->reset_group)
 892                                         continue;
 893 try_again_reset:
 894                                 pr_debug2("reopening weak %s\n", evsel__name(counter));
 895                                 if (create_perf_stat_counter(counter, &stat_config, &target,
 896                                                              counter->cpu_iter - 1) < 0) {
 897
 898                                         switch (stat_handle_error(counter)) {
 899                                         case COUNTER_FATAL:
 900                                                 return -1;
 901                                         case COUNTER_RETRY:
 902                                                 goto try_again_reset;
 903                                         case COUNTER_SKIP:
 904                                                 continue;
 905                                         default:
 906                                                 break;
 907                                         }
 908                                 }
 909                                 counter->supported = true;
 910                         }
 911                 }
 912         }
 913         affinity__cleanup(&affinity);
 914
 915         evlist__for_each_entry(evsel_list, counter) {
 916                 if (!counter->supported) {
 917                         perf_evsel__free_fd(&counter->core);
 918                         continue;
 919                 }
 920
 921                 l = strlen(counter->unit);
 922                 if (l > stat_config.unit_width)
 923                         stat_config.unit_width = l;
 924
 925                 if (evsel__should_store_id(counter) &&
 926                     evsel__store_ids(counter, evsel_list))
 927                         return -1;
 928         }
 929
 930         if (evlist__apply_filters(evsel_list, &counter)) {
 931                 pr_err("failed to set filter \"%s\" on event %s with %d (%s)\n",
 932                         counter->filter, evsel__name(counter), errno,
 933                         str_error_r(errno, msg, sizeof(msg)));
 934                 return -1;
 935         }
 936
 937         if (STAT_RECORD) {
 938                 int fd = perf_data__fd(&perf_stat.data);
 939
 940                 if (is_pipe) {
 941                         err = perf_header__write_pipe(perf_data__fd(&perf_stat.data));
 942                 } else {
 943                         err = perf_session__write_header(perf_stat.session, evsel_list,
 944                                                          fd, false);
 945                 }
 946
 947                 if (err < 0)
 948                         return err;
 949
 950                 err = perf_event__synthesize_stat_events(&stat_config, NULL, evsel_list,
 951                                                          process_synthesized_event, is_pipe);
 952                 if (err < 0)
 953                         return err;
 954         }
 955
 956         err = enable_counters();
 957         if (err)
 958                 return -1;
 959
 960         /* Exec the command, if any */
 961         if (forks)
 962                 evlist__start_workload(evsel_list);
 963
 964         t0 = rdclock();
 965         clock_gettime(CLOCK_MONOTONIC, &ref_time);
 966
 967         if (forks) {
 968                 if (interval || timeout || evlist__ctlfd_initialized(evsel_list))
 969                         status = dispatch_events(forks, timeout, interval, &times);
 970                 if (child_pid != -1) {
 971                         if (timeout)
 972                                 kill(child_pid, SIGTERM);
 973                         wait4(child_pid, &status, 0, &stat_config.ru_data);
 974                 }
 975
 976                 if (workload_exec_errno) {
 977                         const char *emsg = str_error_r(workload_exec_errno, msg, sizeof(msg));
 978                         pr_err("Workload failed: %s\n", emsg);
 979                         return -1;
 980                 }
 981
 982                 if (WIFSIGNALED(status))
 983                         psignal(WTERMSIG(status), argv[0]);
 984         } else {
 985                 status = dispatch_events(forks, timeout, interval, &times);
 986         }
 987
 988         disable_counters();
 989
 990         t1 = rdclock();
 991
 992         if (stat_config.walltime_run_table)
 993                 stat_config.walltime_run[run_idx] = t1 - t0;
 994
 995         if (interval && stat_config.summary) {
 996                 stat_config.interval = 0;
 997                 stat_config.stop_read_counter = true;
 998                 init_stats(&walltime_nsecs_stats);
 999                 update_stats(&walltime_nsecs_stats, t1 - t0);
1000
1001                 if (stat_config.aggr_mode == AGGR_GLOBAL)
1002                         evlist__save_aggr_prev_raw_counts(evsel_list);
1003
1004                 evlist__copy_prev_raw_counts(evsel_list);
1005                 evlist__reset_prev_raw_counts(evsel_list);
1006                 runtime_stat_reset(&stat_config);
1007                 perf_stat__reset_shadow_per_stat(&rt_stat);
1008         } else
1009                 update_stats(&walltime_nsecs_stats, t1 - t0);
1010
1011         /*
1012          * Closing a group leader splits the group, and as we only disable
1013          * group leaders, results in remaining events becoming enabled. To
1014          * avoid arbitrary skew, we must read all counters before closing any
1015          * group leaders.
1016          */
1017         read_counters(&(struct timespec) { .tv_nsec = t1-t0 });
1018
1019         /*
1020          * We need to keep evsel_list alive, because it's processed
1021          * later the evsel_list will be closed after.
1022          */
1023         if (!STAT_RECORD)
1024                 evlist__close(evsel_list);
1025
1026         return WEXITSTATUS(status);
1027 }
1028
1029 static int run_perf_stat(int argc, const char **argv, int run_idx)
1030 {
1031         int ret;
1032
1033         if (pre_cmd) {
1034                 ret = system(pre_cmd);
1035                 if (ret)
1036                         return ret;
1037         }
1038
1039         if (sync_run)
1040                 sync();
1041
1042         ret = __run_perf_stat(argc, argv, run_idx);
1043         if (ret)
1044                 return ret;
1045
1046         if (post_cmd) {
1047                 ret = system(post_cmd);
1048                 if (ret)
1049                         return ret;
1050         }
1051
1052         return ret;
1053 }
1054
1055 static void print_counters(struct timespec *ts, int argc, const char **argv)
1056 {
1057         /* Do not print anything if we record to the pipe. */
1058         if (STAT_RECORD && perf_stat.data.is_pipe)
1059                 return;
1060         if (stat_config.quiet)
1061                 return;
1062
1063         evlist__print_counters(evsel_list, &stat_config, &target, ts, argc, argv);
1064 }
1065
1066 static volatile int signr = -1;
1067
1068 static void skip_signal(int signo)
1069 {
1070         if ((child_pid == -1) || stat_config.interval)
1071                 done = 1;
1072
1073         signr = signo;
1074         /*
1075          * render child_pid harmless
1076          * won't send SIGTERM to a random
1077          * process in case of race condition
1078          * and fast PID recycling
1079          */
1080         child_pid = -1;
1081 }
1082
1083 static void sig_atexit(void)
1084 {
1085         sigset_t set, oset;
1086
1087         /*
1088          * avoid race condition with SIGCHLD handler
1089          * in skip_signal() which is modifying child_pid
1090          * goal is to avoid send SIGTERM to a random
1091          * process
1092          */
1093         sigemptyset(&set);
1094         sigaddset(&set, SIGCHLD);
1095         sigprocmask(SIG_BLOCK, &set, &oset);
1096
1097         if (child_pid != -1)
1098                 kill(child_pid, SIGTERM);
1099
1100         sigprocmask(SIG_SETMASK, &oset, NULL);
1101
1102         if (signr == -1)
1103                 return;
1104
1105         signal(signr, SIG_DFL);
1106         kill(getpid(), signr);
1107 }
1108
1109 void perf_stat__set_big_num(int set)
1110 {
1111         stat_config.big_num = (set != 0);
1112 }
1113
1114 void perf_stat__set_no_csv_summary(int set)
1115 {
1116         stat_config.no_csv_summary = (set != 0);
1117 }
1118
1119 static int stat__set_big_num(const struct option *opt __maybe_unused,
1120                              const char *s __maybe_unused, int unset)
1121 {
1122         big_num_opt = unset ? 0 : 1;
1123         perf_stat__set_big_num(!unset);
1124         return 0;
1125 }
1126
1127 static int enable_metric_only(const struct option *opt __maybe_unused,
1128                               const char *s __maybe_unused, int unset)
1129 {
1130         force_metric_only = true;
1131         stat_config.metric_only = !unset;
1132         return 0;
1133 }
1134
1135 static int parse_metric_groups(const struct option *opt,
1136                                const char *str,
1137                                int unset __maybe_unused)
1138 {
1139         return metricgroup__parse_groups(opt, str,
1140                                          stat_config.metric_no_group,
1141                                          stat_config.metric_no_merge,
1142                                          &stat_config.metric_events);
1143 }
1144
1145 static int parse_control_option(const struct option *opt,
1146                                 const char *str,
1147                                 int unset __maybe_unused)
1148 {
1149         struct perf_stat_config *config = opt->value;
1150
1151         return evlist__parse_control(str, &config->ctl_fd, &config->ctl_fd_ack, &config->ctl_fd_close);
1152 }
1153
1154 static int parse_stat_cgroups(const struct option *opt,
1155                               const char *str, int unset)
1156 {
1157         if (stat_config.cgroup_list) {
1158                 pr_err("--cgroup and --for-each-cgroup cannot be used together\n");
1159                 return -1;
1160         }
1161
1162         return parse_cgroups(opt, str, unset);
1163 }
1164
1165 static struct option stat_options[] = {
1166         OPT_BOOLEAN('T', "transaction", &transaction_run,
1167                     "hardware transaction statistics"),
1168         OPT_CALLBACK('e', "event", &evsel_list, "event",
1169                      "event selector. use 'perf list' to list available events",
1170                      parse_events_option),
1171         OPT_CALLBACK(0, "filter", &evsel_list, "filter",
1172                      "event filter", parse_filter),
1173         OPT_BOOLEAN('i', "no-inherit", &stat_config.no_inherit,
1174                     "child tasks do not inherit counters"),
1175         OPT_STRING('p', "pid", &target.pid, "pid",
1176                    "stat events on existing process id"),
1177         OPT_STRING('t', "tid", &target.tid, "tid",
1178                    "stat events on existing thread id"),
1179 #ifdef HAVE_BPF_SKEL
1180         OPT_STRING('b', "bpf-prog", &target.bpf_str, "bpf-prog-id",
1181                    "stat events on existing bpf program id"),
1182         OPT_BOOLEAN(0, "bpf-counters", &target.use_bpf,
1183                     "use bpf program to count events"),
1184         OPT_STRING(0, "bpf-attr-map", &target.attr_map, "attr-map-path",
1185                    "path to perf_event_attr map"),
1186 #endif
1187         OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
1188                     "system-wide collection from all CPUs"),
1189         OPT_BOOLEAN('g', "group", &group,
1190                     "put the counters into a counter group"),
1191         OPT_BOOLEAN(0, "scale", &stat_config.scale,
1192                     "Use --no-scale to disable counter scaling for multiplexing"),
1193         OPT_INCR('v', "verbose", &verbose,
1194                     "be more verbose (show counter open errors, etc)"),
1195         OPT_INTEGER('r', "repeat", &stat_config.run_count,
1196                     "repeat command and print average + stddev (max: 100, forever: 0)"),
1197         OPT_BOOLEAN(0, "table", &stat_config.walltime_run_table,
1198                     "display details about each run (only with -r option)"),
1199         OPT_BOOLEAN('n', "null", &stat_config.null_run,
1200                     "null run - dont start any counters"),
1201         OPT_INCR('d', "detailed", &detailed_run,
1202                     "detailed run - start a lot of events"),
1203         OPT_BOOLEAN('S', "sync", &sync_run,
1204                     "call sync() before starting a run"),
1205         OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
1206                            "print large numbers with thousands\' separators",
1207                            stat__set_big_num),
1208         OPT_STRING('C', "cpu", &target.cpu_list, "cpu",
1209                     "list of cpus to monitor in system-wide"),
1210         OPT_SET_UINT('A', "no-aggr", &stat_config.aggr_mode,
1211                     "disable CPU count aggregation", AGGR_NONE),
1212         OPT_BOOLEAN(0, "no-merge", &stat_config.no_merge, "Do not merge identical named events"),
1213         OPT_STRING('x', "field-separator", &stat_config.csv_sep, "separator",
1214                    "print counts with custom separator"),
1215         OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
1216                      "monitor event in cgroup name only", parse_stat_cgroups),
1217         OPT_STRING(0, "for-each-cgroup", &stat_config.cgroup_list, "name",
1218                     "expand events for each cgroup"),
1219         OPT_STRING('o', "output", &output_name, "file", "output file name"),
1220         OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
1221         OPT_INTEGER(0, "log-fd", &output_fd,
1222                     "log output to fd, instead of stderr"),
1223         OPT_STRING(0, "pre", &pre_cmd, "command",
1224                         "command to run prior to the measured command"),
1225         OPT_STRING(0, "post", &post_cmd, "command",
1226                         "command to run after to the measured command"),
1227         OPT_UINTEGER('I', "interval-print", &stat_config.interval,
1228                     "print counts at regular interval in ms "
1229                     "(overhead is possible for values <= 100ms)"),
1230         OPT_INTEGER(0, "interval-count", &stat_config.times,
1231                     "print counts for fixed number of times"),
1232         OPT_BOOLEAN(0, "interval-clear", &stat_config.interval_clear,
1233                     "clear screen in between new interval"),
1234         OPT_UINTEGER(0, "timeout", &stat_config.timeout,
1235                     "stop workload and print counts after a timeout period in ms (>= 10ms)"),
1236         OPT_SET_UINT(0, "per-socket", &stat_config.aggr_mode,
1237                      "aggregate counts per processor socket", AGGR_SOCKET),
1238         OPT_SET_UINT(0, "per-die", &stat_config.aggr_mode,
1239                      "aggregate counts per processor die", AGGR_DIE),
1240         OPT_SET_UINT(0, "per-core", &stat_config.aggr_mode,
1241                      "aggregate counts per physical processor core", AGGR_CORE),
1242         OPT_SET_UINT(0, "per-thread", &stat_config.aggr_mode,
1243                      "aggregate counts per thread", AGGR_THREAD),
1244         OPT_SET_UINT(0, "per-node", &stat_config.aggr_mode,
1245                      "aggregate counts per numa node", AGGR_NODE),
1246         OPT_INTEGER('D', "delay", &stat_config.initial_delay,
1247                     "ms to wait before starting measurement after program start (-1: start with events disabled)"),
1248         OPT_CALLBACK_NOOPT(0, "metric-only", &stat_config.metric_only, NULL,
1249                         "Only print computed metrics. No raw values", enable_metric_only),
1250         OPT_BOOLEAN(0, "metric-no-group", &stat_config.metric_no_group,
1251                        "don't group metric events, impacts multiplexing"),
1252         OPT_BOOLEAN(0, "metric-no-merge", &stat_config.metric_no_merge,
1253                        "don't try to share events between metrics in a group"),
1254         OPT_BOOLEAN(0, "topdown", &topdown_run,
1255                         "measure top-down statistics"),
1256         OPT_UINTEGER(0, "td-level", &stat_config.topdown_level,
1257                         "Set the metrics level for the top-down statistics (0: max level)"),
1258         OPT_BOOLEAN(0, "smi-cost", &smi_cost,
1259                         "measure SMI cost"),
1260         OPT_CALLBACK('M', "metrics", &evsel_list, "metric/metric group list",
1261                      "monitor specified metrics or metric groups (separated by ,)",
1262                      parse_metric_groups),
1263         OPT_BOOLEAN_FLAG(0, "all-kernel", &stat_config.all_kernel,
1264                          "Configure all used events to run in kernel space.",
1265                          PARSE_OPT_EXCLUSIVE),
1266         OPT_BOOLEAN_FLAG(0, "all-user", &stat_config.all_user,
1267                          "Configure all used events to run in user space.",
1268                          PARSE_OPT_EXCLUSIVE),
1269         OPT_BOOLEAN(0, "percore-show-thread", &stat_config.percore_show_thread,
1270                     "Use with 'percore' event qualifier to show the event "
1271                     "counts of one hardware thread by sum up total hardware "
1272                     "threads of same physical core"),
1273         OPT_BOOLEAN(0, "summary", &stat_config.summary,
1274                        "print summary for interval mode"),
1275         OPT_BOOLEAN(0, "no-csv-summary", &stat_config.no_csv_summary,
1276                        "don't print 'summary' for CSV summary output"),
1277         OPT_BOOLEAN(0, "quiet", &stat_config.quiet,
1278                         "don't print output (useful with record)"),
1279 #ifdef HAVE_LIBPFM
1280         OPT_CALLBACK(0, "pfm-events", &evsel_list, "event",
1281                 "libpfm4 event selector. use 'perf list' to list available events",
1282                 parse_libpfm_events_option),
1283 #endif
1284         OPT_CALLBACK(0, "control", &stat_config, "fd:ctl-fd[,ack-fd] or fifo:ctl-fifo[,ack-fifo]",
1285                      "Listen on ctl-fd descriptor for command to control measurement ('enable': enable events, 'disable': disable events).\n"
1286                      "\t\t\t  Optionally send control command completion ('ack\\n') to ack-fd descriptor.\n"
1287                      "\t\t\t  Alternatively, ctl-fifo / ack-fifo will be opened and used as ctl-fd / ack-fd.",
1288                       parse_control_option),
1289         OPT_CALLBACK_OPTARG(0, "iostat", &evsel_list, &stat_config, "default",
1290                             "measure I/O performance metrics provided by arch/platform",
1291                             iostat_parse),
1292         OPT_END()
1293 };
1294
1295 static struct aggr_cpu_id perf_stat__get_socket(struct perf_stat_config *config __maybe_unused,
1296                                  struct perf_cpu_map *map, int cpu)
1297 {
1298         return cpu_map__get_socket(map, cpu, NULL);
1299 }
1300
1301 static struct aggr_cpu_id perf_stat__get_die(struct perf_stat_config *config __maybe_unused,
1302                               struct perf_cpu_map *map, int cpu)
1303 {
1304         return cpu_map__get_die(map, cpu, NULL);
1305 }
1306
1307 static struct aggr_cpu_id perf_stat__get_core(struct perf_stat_config *config __maybe_unused,
1308                                struct perf_cpu_map *map, int cpu)
1309 {
1310         return cpu_map__get_core(map, cpu, NULL);
1311 }
1312
1313 static struct aggr_cpu_id perf_stat__get_node(struct perf_stat_config *config __maybe_unused,
1314                                struct perf_cpu_map *map, int cpu)
1315 {
1316         return cpu_map__get_node(map, cpu, NULL);
1317 }
1318
1319 static struct aggr_cpu_id perf_stat__get_aggr(struct perf_stat_config *config,
1320                                aggr_get_id_t get_id, struct perf_cpu_map *map, int idx)
1321 {
1322         int cpu;
1323         struct aggr_cpu_id id = cpu_map__empty_aggr_cpu_id();
1324
1325         if (idx >= map->nr)
1326                 return id;
1327
1328         cpu = map->map[idx];
1329
1330         if (cpu_map__aggr_cpu_id_is_empty(config->cpus_aggr_map->map[cpu]))
1331                 config->cpus_aggr_map->map[cpu] = get_id(config, map, idx);
1332
1333         id = config->cpus_aggr_map->map[cpu];
1334         return id;
1335 }
1336
1337 static struct aggr_cpu_id perf_stat__get_socket_cached(struct perf_stat_config *config,
1338                                         struct perf_cpu_map *map, int idx)
1339 {
1340         return perf_stat__get_aggr(config, perf_stat__get_socket, map, idx);
1341 }
1342
1343 static struct aggr_cpu_id perf_stat__get_die_cached(struct perf_stat_config *config,
1344                                         struct perf_cpu_map *map, int idx)
1345 {
1346         return perf_stat__get_aggr(config, perf_stat__get_die, map, idx);
1347 }
1348
1349 static struct aggr_cpu_id perf_stat__get_core_cached(struct perf_stat_config *config,
1350                                       struct perf_cpu_map *map, int idx)
1351 {
1352         return perf_stat__get_aggr(config, perf_stat__get_core, map, idx);
1353 }
1354
1355 static struct aggr_cpu_id perf_stat__get_node_cached(struct perf_stat_config *config,
1356                                       struct perf_cpu_map *map, int idx)
1357 {
1358         return perf_stat__get_aggr(config, perf_stat__get_node, map, idx);
1359 }
1360
1361 static bool term_percore_set(void)
1362 {
1363         struct evsel *counter;
1364
1365         evlist__for_each_entry(evsel_list, counter) {
1366                 if (counter->percore)
1367                         return true;
1368         }
1369
1370         return false;
1371 }
1372
1373 static int perf_stat_init_aggr_mode(void)
1374 {
1375         int nr;
1376
1377         switch (stat_config.aggr_mode) {
1378         case AGGR_SOCKET:
1379                 if (cpu_map__build_socket_map(evsel_list->core.cpus, &stat_config.aggr_map)) {
1380                         perror("cannot build socket map");
1381                         return -1;
1382                 }
1383                 stat_config.aggr_get_id = perf_stat__get_socket_cached;
1384                 break;
1385         case AGGR_DIE:
1386                 if (cpu_map__build_die_map(evsel_list->core.cpus, &stat_config.aggr_map)) {
1387                         perror("cannot build die map");
1388                         return -1;
1389                 }
1390                 stat_config.aggr_get_id = perf_stat__get_die_cached;
1391                 break;
1392         case AGGR_CORE:
1393                 if (cpu_map__build_core_map(evsel_list->core.cpus, &stat_config.aggr_map)) {
1394                         perror("cannot build core map");
1395                         return -1;
1396                 }
1397                 stat_config.aggr_get_id = perf_stat__get_core_cached;
1398                 break;
1399         case AGGR_NODE:
1400                 if (cpu_map__build_node_map(evsel_list->core.cpus, &stat_config.aggr_map)) {
1401                         perror("cannot build core map");
1402                         return -1;
1403                 }
1404                 stat_config.aggr_get_id = perf_stat__get_node_cached;
1405                 break;
1406         case AGGR_NONE:
1407                 if (term_percore_set()) {
1408                         if (cpu_map__build_core_map(evsel_list->core.cpus,
1409                                                     &stat_config.aggr_map)) {
1410                                 perror("cannot build core map");
1411                                 return -1;
1412                         }
1413                         stat_config.aggr_get_id = perf_stat__get_core_cached;
1414                 }
1415                 break;
1416         case AGGR_GLOBAL:
1417         case AGGR_THREAD:
1418         case AGGR_UNSET:
1419         default:
1420                 break;
1421         }
1422
1423         /*
1424          * The evsel_list->cpus is the base we operate on,
1425          * taking the highest cpu number to be the size of
1426          * the aggregation translate cpumap.
1427          */
1428         nr = perf_cpu_map__max(evsel_list->core.cpus);
1429         stat_config.cpus_aggr_map = cpu_aggr_map__empty_new(nr + 1);
1430         return stat_config.cpus_aggr_map ? 0 : -ENOMEM;
1431 }
1432
1433 static void cpu_aggr_map__delete(struct cpu_aggr_map *map)
1434 {
1435         if (map) {
1436                 WARN_ONCE(refcount_read(&map->refcnt) != 0,
1437                           "cpu_aggr_map refcnt unbalanced\n");
1438                 free(map);
1439         }
1440 }
1441
1442 static void cpu_aggr_map__put(struct cpu_aggr_map *map)
1443 {
1444         if (map && refcount_dec_and_test(&map->refcnt))
1445                 cpu_aggr_map__delete(map);
1446 }
1447
1448 static void perf_stat__exit_aggr_mode(void)
1449 {
1450         cpu_aggr_map__put(stat_config.aggr_map);
1451         cpu_aggr_map__put(stat_config.cpus_aggr_map);
1452         stat_config.aggr_map = NULL;
1453         stat_config.cpus_aggr_map = NULL;
1454 }
1455
1456 static inline int perf_env__get_cpu(struct perf_env *env, struct perf_cpu_map *map, int idx)
1457 {
1458         int cpu;
1459
1460         if (idx > map->nr)
1461                 return -1;
1462
1463         cpu = map->map[idx];
1464
1465         if (cpu >= env->nr_cpus_avail)
1466                 return -1;
1467
1468         return cpu;
1469 }
1470
1471 static struct aggr_cpu_id perf_env__get_socket(struct perf_cpu_map *map, int idx, void *data)
1472 {
1473         struct perf_env *env = data;
1474         int cpu = perf_env__get_cpu(env, map, idx);
1475         struct aggr_cpu_id id = cpu_map__empty_aggr_cpu_id();
1476
1477         if (cpu != -1)
1478                 id.socket = env->cpu[cpu].socket_id;
1479
1480         return id;
1481 }
1482
1483 static struct aggr_cpu_id perf_env__get_die(struct perf_cpu_map *map, int idx, void *data)
1484 {
1485         struct perf_env *env = data;
1486         struct aggr_cpu_id id = cpu_map__empty_aggr_cpu_id();
1487         int cpu = perf_env__get_cpu(env, map, idx);
1488
1489         if (cpu != -1) {
1490                 /*
1491                  * die_id is relative to socket, so start
1492                  * with the socket ID and then add die to
1493                  * make a unique ID.
1494                  */
1495                 id.socket = env->cpu[cpu].socket_id;
1496                 id.die = env->cpu[cpu].die_id;
1497         }
1498
1499         return id;
1500 }
1501
1502 static struct aggr_cpu_id perf_env__get_core(struct perf_cpu_map *map, int idx, void *data)
1503 {
1504         struct perf_env *env = data;
1505         struct aggr_cpu_id id = cpu_map__empty_aggr_cpu_id();
1506         int cpu = perf_env__get_cpu(env, map, idx);
1507
1508         if (cpu != -1) {
1509                 /*
1510                  * core_id is relative to socket and die,
1511                  * we need a global id. So we set
1512                  * socket, die id and core id
1513                  */
1514                 id.socket = env->cpu[cpu].socket_id;
1515                 id.die = env->cpu[cpu].die_id;
1516                 id.core = env->cpu[cpu].core_id;
1517         }
1518
1519         return id;
1520 }
1521
1522 static struct aggr_cpu_id perf_env__get_node(struct perf_cpu_map *map, int idx, void *data)
1523 {
1524         int cpu = perf_env__get_cpu(data, map, idx);
1525         struct aggr_cpu_id id = cpu_map__empty_aggr_cpu_id();
1526
1527         id.node = perf_env__numa_node(data, cpu);
1528         return id;
1529 }
1530
1531 static int perf_env__build_socket_map(struct perf_env *env, struct perf_cpu_map *cpus,
1532                                       struct cpu_aggr_map **sockp)
1533 {
1534         return cpu_map__build_map(cpus, sockp, perf_env__get_socket, env);
1535 }
1536
1537 static int perf_env__build_die_map(struct perf_env *env, struct perf_cpu_map *cpus,
1538                                    struct cpu_aggr_map **diep)
1539 {
1540         return cpu_map__build_map(cpus, diep, perf_env__get_die, env);
1541 }
1542
1543 static int perf_env__build_core_map(struct perf_env *env, struct perf_cpu_map *cpus,
1544                                     struct cpu_aggr_map **corep)
1545 {
1546         return cpu_map__build_map(cpus, corep, perf_env__get_core, env);
1547 }
1548
1549 static int perf_env__build_node_map(struct perf_env *env, struct perf_cpu_map *cpus,
1550                                     struct cpu_aggr_map **nodep)
1551 {
1552         return cpu_map__build_map(cpus, nodep, perf_env__get_node, env);
1553 }
1554
1555 static struct aggr_cpu_id perf_stat__get_socket_file(struct perf_stat_config *config __maybe_unused,
1556                                       struct perf_cpu_map *map, int idx)
1557 {
1558         return perf_env__get_socket(map, idx, &perf_stat.session->header.env);
1559 }
1560 static struct aggr_cpu_id perf_stat__get_die_file(struct perf_stat_config *config __maybe_unused,
1561                                    struct perf_cpu_map *map, int idx)
1562 {
1563         return perf_env__get_die(map, idx, &perf_stat.session->header.env);
1564 }
1565
1566 static struct aggr_cpu_id perf_stat__get_core_file(struct perf_stat_config *config __maybe_unused,
1567                                     struct perf_cpu_map *map, int idx)
1568 {
1569         return perf_env__get_core(map, idx, &perf_stat.session->header.env);
1570 }
1571
1572 static struct aggr_cpu_id perf_stat__get_node_file(struct perf_stat_config *config __maybe_unused,
1573                                     struct perf_cpu_map *map, int idx)
1574 {
1575         return perf_env__get_node(map, idx, &perf_stat.session->header.env);
1576 }
1577
1578 static int perf_stat_init_aggr_mode_file(struct perf_stat *st)
1579 {
1580         struct perf_env *env = &st->session->header.env;
1581
1582         switch (stat_config.aggr_mode) {
1583         case AGGR_SOCKET:
1584                 if (perf_env__build_socket_map(env, evsel_list->core.cpus, &stat_config.aggr_map)) {
1585                         perror("cannot build socket map");
1586                         return -1;
1587                 }
1588                 stat_config.aggr_get_id = perf_stat__get_socket_file;
1589                 break;
1590         case AGGR_DIE:
1591                 if (perf_env__build_die_map(env, evsel_list->core.cpus, &stat_config.aggr_map)) {
1592                         perror("cannot build die map");
1593                         return -1;
1594                 }
1595                 stat_config.aggr_get_id = perf_stat__get_die_file;
1596                 break;
1597         case AGGR_CORE:
1598                 if (perf_env__build_core_map(env, evsel_list->core.cpus, &stat_config.aggr_map)) {
1599                         perror("cannot build core map");
1600                         return -1;
1601                 }
1602                 stat_config.aggr_get_id = perf_stat__get_core_file;
1603                 break;
1604         case AGGR_NODE:
1605                 if (perf_env__build_node_map(env, evsel_list->core.cpus, &stat_config.aggr_map)) {
1606                         perror("cannot build core map");
1607                         return -1;
1608                 }
1609                 stat_config.aggr_get_id = perf_stat__get_node_file;
1610                 break;
1611         case AGGR_NONE:
1612         case AGGR_GLOBAL:
1613         case AGGR_THREAD:
1614         case AGGR_UNSET:
1615         default:
1616                 break;
1617         }
1618
1619         return 0;
1620 }
1621
1622 /*
1623  * Add default attributes, if there were no attributes specified or
1624  * if -d/--detailed, -d -d or -d -d -d is used:
1625  */
1626 static int add_default_attributes(void)
1627 {
1628         int err;
1629         struct perf_event_attr default_attrs0[] = {
1630
1631   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK              },
1632   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES        },
1633   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS          },
1634   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS             },
1635
1636   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES              },
1637 };
1638         struct perf_event_attr frontend_attrs[] = {
1639   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND },
1640 };
1641         struct perf_event_attr backend_attrs[] = {
1642   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND  },
1643 };
1644         struct perf_event_attr default_attrs1[] = {
1645   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS            },
1646   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS     },
1647   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES           },
1648
1649 };
1650         struct perf_event_attr default_sw_attrs[] = {
1651   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK              },
1652   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES        },
1653   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS          },
1654   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS             },
1655 };
1656
1657 /*
1658  * Detailed stats (-d), covering the L1 and last level data caches:
1659  */
1660         struct perf_event_attr detailed_attrs[] = {
1661
1662   { .type = PERF_TYPE_HW_CACHE,
1663     .config =
1664          PERF_COUNT_HW_CACHE_L1D                <<  0  |
1665         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1666         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
1667
1668   { .type = PERF_TYPE_HW_CACHE,
1669     .config =
1670          PERF_COUNT_HW_CACHE_L1D                <<  0  |
1671         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1672         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
1673
1674   { .type = PERF_TYPE_HW_CACHE,
1675     .config =
1676          PERF_COUNT_HW_CACHE_LL                 <<  0  |
1677         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1678         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
1679
1680   { .type = PERF_TYPE_HW_CACHE,
1681     .config =
1682          PERF_COUNT_HW_CACHE_LL                 <<  0  |
1683         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1684         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
1685 };
1686
1687 /*
1688  * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
1689  */
1690         struct perf_event_attr very_detailed_attrs[] = {
1691
1692   { .type = PERF_TYPE_HW_CACHE,
1693     .config =
1694          PERF_COUNT_HW_CACHE_L1I                <<  0  |
1695         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1696         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
1697
1698   { .type = PERF_TYPE_HW_CACHE,
1699     .config =
1700          PERF_COUNT_HW_CACHE_L1I                <<  0  |
1701         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1702         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
1703
1704   { .type = PERF_TYPE_HW_CACHE,
1705     .config =
1706          PERF_COUNT_HW_CACHE_DTLB               <<  0  |
1707         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1708         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
1709
1710   { .type = PERF_TYPE_HW_CACHE,
1711     .config =
1712          PERF_COUNT_HW_CACHE_DTLB               <<  0  |
1713         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1714         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
1715
1716   { .type = PERF_TYPE_HW_CACHE,
1717     .config =
1718          PERF_COUNT_HW_CACHE_ITLB               <<  0  |
1719         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1720         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
1721
1722   { .type = PERF_TYPE_HW_CACHE,
1723     .config =
1724          PERF_COUNT_HW_CACHE_ITLB               <<  0  |
1725         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1726         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
1727
1728 };
1729
1730 /*
1731  * Very, very detailed stats (-d -d -d), adding prefetch events:
1732  */
1733         struct perf_event_attr very_very_detailed_attrs[] = {
1734
1735   { .type = PERF_TYPE_HW_CACHE,
1736     .config =
1737          PERF_COUNT_HW_CACHE_L1D                <<  0  |
1738         (PERF_COUNT_HW_CACHE_OP_PREFETCH        <<  8) |
1739         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
1740
1741   { .type = PERF_TYPE_HW_CACHE,
1742     .config =
1743          PERF_COUNT_HW_CACHE_L1D                <<  0  |
1744         (PERF_COUNT_HW_CACHE_OP_PREFETCH        <<  8) |
1745         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
1746 };
1747         struct parse_events_error errinfo;
1748
1749         /* Set attrs if no event is selected and !null_run: */
1750         if (stat_config.null_run)
1751                 return 0;
1752
1753         bzero(&errinfo, sizeof(errinfo));
1754         if (transaction_run) {
1755                 /* Handle -T as -M transaction. Once platform specific metrics
1756                  * support has been added to the json files, all architectures
1757                  * will use this approach. To determine transaction support
1758                  * on an architecture test for such a metric name.
1759                  */
1760                 if (metricgroup__has_metric("transaction")) {
1761                         struct option opt = { .value = &evsel_list };
1762
1763                         return metricgroup__parse_groups(&opt, "transaction",
1764                                                          stat_config.metric_no_group,
1765                                                         stat_config.metric_no_merge,
1766                                                          &stat_config.metric_events);
1767                 }
1768
1769                 if (pmu_have_event("cpu", "cycles-ct") &&
1770                     pmu_have_event("cpu", "el-start"))
1771                         err = parse_events(evsel_list, transaction_attrs,
1772                                            &errinfo);
1773                 else
1774                         err = parse_events(evsel_list,
1775                                            transaction_limited_attrs,
1776                                            &errinfo);
1777                 if (err) {
1778                         fprintf(stderr, "Cannot set up transaction events\n");
1779                         parse_events_print_error(&errinfo, transaction_attrs);
1780                         return -1;
1781                 }
1782                 return 0;
1783         }
1784
1785         if (smi_cost) {
1786                 int smi;
1787
1788                 if (sysfs__read_int(FREEZE_ON_SMI_PATH, &smi) < 0) {
1789                         fprintf(stderr, "freeze_on_smi is not supported.\n");
1790                         return -1;
1791                 }
1792
1793                 if (!smi) {
1794                         if (sysfs__write_int(FREEZE_ON_SMI_PATH, 1) < 0) {
1795                                 fprintf(stderr, "Failed to set freeze_on_smi.\n");
1796                                 return -1;
1797                         }
1798                         smi_reset = true;
1799                 }
1800
1801                 if (pmu_have_event("msr", "aperf") &&
1802                     pmu_have_event("msr", "smi")) {
1803                         if (!force_metric_only)
1804                                 stat_config.metric_only = true;
1805                         err = parse_events(evsel_list, smi_cost_attrs, &errinfo);
1806                 } else {
1807                         fprintf(stderr, "To measure SMI cost, it needs "
1808                                 "msr/aperf/, msr/smi/ and cpu/cycles/ support\n");
1809                         parse_events_print_error(&errinfo, smi_cost_attrs);
1810                         return -1;
1811                 }
1812                 if (err) {
1813                         parse_events_print_error(&errinfo, smi_cost_attrs);
1814                         fprintf(stderr, "Cannot set up SMI cost events\n");
1815                         return -1;
1816                 }
1817                 return 0;
1818         }
1819
1820         if (topdown_run) {
1821                 const char **metric_attrs = topdown_metric_attrs;
1822                 unsigned int max_level = 1;
1823                 char *str = NULL;
1824                 bool warn = false;
1825
1826                 if (!force_metric_only)
1827                         stat_config.metric_only = true;
1828
1829                 if (pmu_have_event("cpu", topdown_metric_L2_attrs[5])) {
1830                         metric_attrs = topdown_metric_L2_attrs;
1831                         max_level = 2;
1832                 }
1833
1834                 if (stat_config.topdown_level > max_level) {
1835                         pr_err("Invalid top-down metrics level. The max level is %u.\n", max_level);
1836                         return -1;
1837                 } else if (!stat_config.topdown_level)
1838                         stat_config.topdown_level = max_level;
1839
1840                 if (topdown_filter_events(metric_attrs, &str, 1) < 0) {
1841                         pr_err("Out of memory\n");
1842                         return -1;
1843                 }
1844                 if (metric_attrs[0] && str) {
1845                         if (!stat_config.interval && !stat_config.metric_only) {
1846                                 fprintf(stat_config.output,
1847                                         "Topdown accuracy may decrease when measuring long periods.\n"
1848                                         "Please print the result regularly, e.g. -I1000\n");
1849                         }
1850                         goto setup_metrics;
1851                 }
1852
1853                 zfree(&str);
1854
1855                 if (stat_config.aggr_mode != AGGR_GLOBAL &&
1856                     stat_config.aggr_mode != AGGR_CORE) {
1857                         pr_err("top down event configuration requires --per-core mode\n");
1858                         return -1;
1859                 }
1860                 stat_config.aggr_mode = AGGR_CORE;
1861                 if (nr_cgroups || !target__has_cpu(&target)) {
1862                         pr_err("top down event configuration requires system-wide mode (-a)\n");
1863                         return -1;
1864                 }
1865
1866                 if (topdown_filter_events(topdown_attrs, &str,
1867                                 arch_topdown_check_group(&warn)) < 0) {
1868                         pr_err("Out of memory\n");
1869                         return -1;
1870                 }
1871                 if (topdown_attrs[0] && str) {
1872                         if (warn)
1873                                 arch_topdown_group_warn();
1874 setup_metrics:
1875                         err = parse_events(evsel_list, str, &errinfo);
1876                         if (err) {
1877                                 fprintf(stderr,
1878                                         "Cannot set up top down events %s: %d\n",
1879                                         str, err);
1880                                 parse_events_print_error(&errinfo, str);
1881                                 free(str);
1882                                 return -1;
1883                         }
1884                 } else {
1885                         fprintf(stderr, "System does not support topdown\n");
1886                         return -1;
1887                 }
1888                 free(str);
1889         }
1890
1891         if (!evsel_list->core.nr_entries) {
1892                 if (perf_pmu__has_hybrid()) {
1893                         const char *hybrid_str = "cycles,instructions,branches,branch-misses";
1894
1895                         if (target__has_cpu(&target))
1896                                 default_sw_attrs[0].config = PERF_COUNT_SW_CPU_CLOCK;
1897
1898                         if (evlist__add_default_attrs(evsel_list,
1899                                                       default_sw_attrs) < 0) {
1900                                 return -1;
1901                         }
1902
1903                         err = parse_events(evsel_list, hybrid_str, &errinfo);
1904                         if (err) {
1905                                 fprintf(stderr,
1906                                         "Cannot set up hybrid events %s: %d\n",
1907                                         hybrid_str, err);
1908                                 parse_events_print_error(&errinfo, hybrid_str);
1909                                 return -1;
1910                         }
1911                         return err;
1912                 }
1913
1914                 if (target__has_cpu(&target))
1915                         default_attrs0[0].config = PERF_COUNT_SW_CPU_CLOCK;
1916
1917                 if (evlist__add_default_attrs(evsel_list, default_attrs0) < 0)
1918                         return -1;
1919                 if (pmu_have_event("cpu", "stalled-cycles-frontend")) {
1920                         if (evlist__add_default_attrs(evsel_list, frontend_attrs) < 0)
1921                                 return -1;
1922                 }
1923                 if (pmu_have_event("cpu", "stalled-cycles-backend")) {
1924                         if (evlist__add_default_attrs(evsel_list, backend_attrs) < 0)
1925                                 return -1;
1926                 }
1927                 if (evlist__add_default_attrs(evsel_list, default_attrs1) < 0)
1928                         return -1;
1929
1930                 stat_config.topdown_level = TOPDOWN_MAX_LEVEL;
1931                 if (arch_evlist__add_default_attrs(evsel_list) < 0)
1932                         return -1;
1933         }
1934
1935         /* Detailed events get appended to the event list: */
1936
1937         if (detailed_run <  1)
1938                 return 0;
1939
1940         /* Append detailed run extra attributes: */
1941         if (evlist__add_default_attrs(evsel_list, detailed_attrs) < 0)
1942                 return -1;
1943
1944         if (detailed_run < 2)
1945                 return 0;
1946
1947         /* Append very detailed run extra attributes: */
1948         if (evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0)
1949                 return -1;
1950
1951         if (detailed_run < 3)
1952                 return 0;
1953
1954         /* Append very, very detailed run extra attributes: */
1955         return evlist__add_default_attrs(evsel_list, very_very_detailed_attrs);
1956 }
1957
1958 static const char * const stat_record_usage[] = {
1959         "perf stat record [<options>]",
1960         NULL,
1961 };
1962
1963 static void init_features(struct perf_session *session)
1964 {
1965         int feat;
1966
1967         for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++)
1968                 perf_header__set_feat(&session->header, feat);
1969
1970         perf_header__clear_feat(&session->header, HEADER_DIR_FORMAT);
1971         perf_header__clear_feat(&session->header, HEADER_BUILD_ID);
1972         perf_header__clear_feat(&session->header, HEADER_TRACING_DATA);
1973         perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK);
1974         perf_header__clear_feat(&session->header, HEADER_AUXTRACE);
1975 }
1976
1977 static int __cmd_record(int argc, const char **argv)
1978 {
1979         struct perf_session *session;
1980         struct perf_data *data = &perf_stat.data;
1981
1982         argc = parse_options(argc, argv, stat_options, stat_record_usage,
1983                              PARSE_OPT_STOP_AT_NON_OPTION);
1984
1985         if (output_name)
1986                 data->path = output_name;
1987
1988         if (stat_config.run_count != 1 || forever) {
1989                 pr_err("Cannot use -r option with perf stat record.\n");
1990                 return -1;
1991         }
1992
1993         session = perf_session__new(data, NULL);
1994         if (IS_ERR(session)) {
1995                 pr_err("Perf session creation failed\n");
1996                 return PTR_ERR(session);
1997         }
1998
1999         init_features(session);
2000
2001         session->evlist   = evsel_list;
2002         perf_stat.session = session;
2003         perf_stat.record  = true;
2004         return argc;
2005 }
2006
2007 static int process_stat_round_event(struct perf_session *session,
2008                                     union perf_event *event)
2009 {
2010         struct perf_record_stat_round *stat_round = &event->stat_round;
2011         struct evsel *counter;
2012         struct timespec tsh, *ts = NULL;
2013         const char **argv = session->header.env.cmdline_argv;
2014         int argc = session->header.env.nr_cmdline;
2015
2016         evlist__for_each_entry(evsel_list, counter)
2017                 perf_stat_process_counter(&stat_config, counter);
2018
2019         if (stat_round->type == PERF_STAT_ROUND_TYPE__FINAL)
2020                 update_stats(&walltime_nsecs_stats, stat_round->time);
2021
2022         if (stat_config.interval && stat_round->time) {
2023                 tsh.tv_sec  = stat_round->time / NSEC_PER_SEC;
2024                 tsh.tv_nsec = stat_round->time % NSEC_PER_SEC;
2025                 ts = &tsh;
2026         }
2027
2028         print_counters(ts, argc, argv);
2029         return 0;
2030 }
2031
2032 static
2033 int process_stat_config_event(struct perf_session *session,
2034                               union perf_event *event)
2035 {
2036         struct perf_tool *tool = session->tool;
2037         struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2038
2039         perf_event__read_stat_config(&stat_config, &event->stat_config);
2040
2041         if (perf_cpu_map__empty(st->cpus)) {
2042                 if (st->aggr_mode != AGGR_UNSET)
2043                         pr_warning("warning: processing task data, aggregation mode not set\n");
2044                 return 0;
2045         }
2046
2047         if (st->aggr_mode != AGGR_UNSET)
2048                 stat_config.aggr_mode = st->aggr_mode;
2049
2050         if (perf_stat.data.is_pipe)
2051                 perf_stat_init_aggr_mode();
2052         else
2053                 perf_stat_init_aggr_mode_file(st);
2054
2055         return 0;
2056 }
2057
2058 static int set_maps(struct perf_stat *st)
2059 {
2060         if (!st->cpus || !st->threads)
2061                 return 0;
2062
2063         if (WARN_ONCE(st->maps_allocated, "stats double allocation\n"))
2064                 return -EINVAL;
2065
2066         perf_evlist__set_maps(&evsel_list->core, st->cpus, st->threads);
2067
2068         if (evlist__alloc_stats(evsel_list, true))
2069                 return -ENOMEM;
2070
2071         st->maps_allocated = true;
2072         return 0;
2073 }
2074
2075 static
2076 int process_thread_map_event(struct perf_session *session,
2077                              union perf_event *event)
2078 {
2079         struct perf_tool *tool = session->tool;
2080         struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2081
2082         if (st->threads) {
2083                 pr_warning("Extra thread map event, ignoring.\n");
2084                 return 0;
2085         }
2086
2087         st->threads = thread_map__new_event(&event->thread_map);
2088         if (!st->threads)
2089                 return -ENOMEM;
2090
2091         return set_maps(st);
2092 }
2093
2094 static
2095 int process_cpu_map_event(struct perf_session *session,
2096                           union perf_event *event)
2097 {
2098         struct perf_tool *tool = session->tool;
2099         struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2100         struct perf_cpu_map *cpus;
2101
2102         if (st->cpus) {
2103                 pr_warning("Extra cpu map event, ignoring.\n");
2104                 return 0;
2105         }
2106
2107         cpus = cpu_map__new_data(&event->cpu_map.data);
2108         if (!cpus)
2109                 return -ENOMEM;
2110
2111         st->cpus = cpus;
2112         return set_maps(st);
2113 }
2114
2115 static const char * const stat_report_usage[] = {
2116         "perf stat report [<options>]",
2117         NULL,
2118 };
2119
2120 static struct perf_stat perf_stat = {
2121         .tool = {
2122                 .attr           = perf_event__process_attr,
2123                 .event_update   = perf_event__process_event_update,
2124                 .thread_map     = process_thread_map_event,
2125                 .cpu_map        = process_cpu_map_event,
2126                 .stat_config    = process_stat_config_event,
2127                 .stat           = perf_event__process_stat_event,
2128                 .stat_round     = process_stat_round_event,
2129         },
2130         .aggr_mode = AGGR_UNSET,
2131 };
2132
2133 static int __cmd_report(int argc, const char **argv)
2134 {
2135         struct perf_session *session;
2136         const struct option options[] = {
2137         OPT_STRING('i', "input", &input_name, "file", "input file name"),
2138         OPT_SET_UINT(0, "per-socket", &perf_stat.aggr_mode,
2139                      "aggregate counts per processor socket", AGGR_SOCKET),
2140         OPT_SET_UINT(0, "per-die", &perf_stat.aggr_mode,
2141                      "aggregate counts per processor die", AGGR_DIE),
2142         OPT_SET_UINT(0, "per-core", &perf_stat.aggr_mode,
2143                      "aggregate counts per physical processor core", AGGR_CORE),
2144         OPT_SET_UINT(0, "per-node", &perf_stat.aggr_mode,
2145                      "aggregate counts per numa node", AGGR_NODE),
2146         OPT_SET_UINT('A', "no-aggr", &perf_stat.aggr_mode,
2147                      "disable CPU count aggregation", AGGR_NONE),
2148         OPT_END()
2149         };
2150         struct stat st;
2151         int ret;
2152
2153         argc = parse_options(argc, argv, options, stat_report_usage, 0);
2154
2155         if (!input_name || !strlen(input_name)) {
2156                 if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode))
2157                         input_name = "-";
2158                 else
2159                         input_name = "perf.data";
2160         }
2161
2162         perf_stat.data.path = input_name;
2163         perf_stat.data.mode = PERF_DATA_MODE_READ;
2164
2165         session = perf_session__new(&perf_stat.data, &perf_stat.tool);
2166         if (IS_ERR(session))
2167                 return PTR_ERR(session);
2168
2169         perf_stat.session  = session;
2170         stat_config.output = stderr;
2171         evsel_list         = session->evlist;
2172
2173         ret = perf_session__process_events(session);
2174         if (ret)
2175                 return ret;
2176
2177         perf_session__delete(session);
2178         return 0;
2179 }
2180
2181 static void setup_system_wide(int forks)
2182 {
2183         /*
2184          * Make system wide (-a) the default target if
2185          * no target was specified and one of following
2186          * conditions is met:
2187          *
2188          *   - there's no workload specified
2189          *   - there is workload specified but all requested
2190          *     events are system wide events
2191          */
2192         if (!target__none(&target))
2193                 return;
2194
2195         if (!forks)
2196                 target.system_wide = true;
2197         else {
2198                 struct evsel *counter;
2199
2200                 evlist__for_each_entry(evsel_list, counter) {
2201                         if (!counter->core.system_wide &&
2202                             strcmp(counter->name, "duration_time")) {
2203                                 return;
2204                         }
2205                 }
2206
2207                 if (evsel_list->core.nr_entries)
2208                         target.system_wide = true;
2209         }
2210 }
2211
2212 int cmd_stat(int argc, const char **argv)
2213 {
2214         const char * const stat_usage[] = {
2215                 "perf stat [<options>] [<command>]",
2216                 NULL
2217         };
2218         int status = -EINVAL, run_idx, err;
2219         const char *mode;
2220         FILE *output = stderr;
2221         unsigned int interval, timeout;
2222         const char * const stat_subcommands[] = { "record", "report" };
2223         char errbuf[BUFSIZ];
2224
2225         setlocale(LC_ALL, "");
2226
2227         evsel_list = evlist__new();
2228         if (evsel_list == NULL)
2229                 return -ENOMEM;
2230
2231         parse_events__shrink_config_terms();
2232
2233         /* String-parsing callback-based options would segfault when negated */
2234         set_option_flag(stat_options, 'e', "event", PARSE_OPT_NONEG);
2235         set_option_flag(stat_options, 'M', "metrics", PARSE_OPT_NONEG);
2236         set_option_flag(stat_options, 'G', "cgroup", PARSE_OPT_NONEG);
2237
2238         argc = parse_options_subcommand(argc, argv, stat_options, stat_subcommands,
2239                                         (const char **) stat_usage,
2240                                         PARSE_OPT_STOP_AT_NON_OPTION);
2241         perf_stat__collect_metric_expr(evsel_list);
2242         perf_stat__init_shadow_stats();
2243
2244         if (stat_config.csv_sep) {
2245                 stat_config.csv_output = true;
2246                 if (!strcmp(stat_config.csv_sep, "\\t"))
2247                         stat_config.csv_sep = "\t";
2248         } else
2249                 stat_config.csv_sep = DEFAULT_SEPARATOR;
2250
2251         if (argc && !strncmp(argv[0], "rec", 3)) {
2252                 argc = __cmd_record(argc, argv);
2253                 if (argc < 0)
2254                         return -1;
2255         } else if (argc && !strncmp(argv[0], "rep", 3))
2256                 return __cmd_report(argc, argv);
2257
2258         interval = stat_config.interval;
2259         timeout = stat_config.timeout;
2260
2261         /*
2262          * For record command the -o is already taken care of.
2263          */
2264         if (!STAT_RECORD && output_name && strcmp(output_name, "-"))
2265                 output = NULL;
2266
2267         if (output_name && output_fd) {
2268                 fprintf(stderr, "cannot use both --output and --log-fd\n");
2269                 parse_options_usage(stat_usage, stat_options, "o", 1);
2270                 parse_options_usage(NULL, stat_options, "log-fd", 0);
2271                 goto out;
2272         }
2273
2274         if (stat_config.metric_only && stat_config.aggr_mode == AGGR_THREAD) {
2275                 fprintf(stderr, "--metric-only is not supported with --per-thread\n");
2276                 goto out;
2277         }
2278
2279         if (stat_config.metric_only && stat_config.run_count > 1) {
2280                 fprintf(stderr, "--metric-only is not supported with -r\n");
2281                 goto out;
2282         }
2283
2284         if (stat_config.walltime_run_table && stat_config.run_count <= 1) {
2285                 fprintf(stderr, "--table is only supported with -r\n");
2286                 parse_options_usage(stat_usage, stat_options, "r", 1);
2287                 parse_options_usage(NULL, stat_options, "table", 0);
2288                 goto out;
2289         }
2290
2291         if (output_fd < 0) {
2292                 fprintf(stderr, "argument to --log-fd must be a > 0\n");
2293                 parse_options_usage(stat_usage, stat_options, "log-fd", 0);
2294                 goto out;
2295         }
2296
2297         if (!output && !stat_config.quiet) {
2298                 struct timespec tm;
2299                 mode = append_file ? "a" : "w";
2300
2301                 output = fopen(output_name, mode);
2302                 if (!output) {
2303                         perror("failed to create output file");
2304                         return -1;
2305                 }
2306                 clock_gettime(CLOCK_REALTIME, &tm);
2307                 fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
2308         } else if (output_fd > 0) {
2309                 mode = append_file ? "a" : "w";
2310                 output = fdopen(output_fd, mode);
2311                 if (!output) {
2312                         perror("Failed opening logfd");
2313                         return -errno;
2314                 }
2315         }
2316
2317         stat_config.output = output;
2318
2319         /*
2320          * let the spreadsheet do the pretty-printing
2321          */
2322         if (stat_config.csv_output) {
2323                 /* User explicitly passed -B? */
2324                 if (big_num_opt == 1) {
2325                         fprintf(stderr, "-B option not supported with -x\n");
2326                         parse_options_usage(stat_usage, stat_options, "B", 1);
2327                         parse_options_usage(NULL, stat_options, "x", 1);
2328                         goto out;
2329                 } else /* Nope, so disable big number formatting */
2330                         stat_config.big_num = false;
2331         } else if (big_num_opt == 0) /* User passed --no-big-num */
2332                 stat_config.big_num = false;
2333
2334         err = target__validate(&target);
2335         if (err) {
2336                 target__strerror(&target, err, errbuf, BUFSIZ);
2337                 pr_warning("%s\n", errbuf);
2338         }
2339
2340         setup_system_wide(argc);
2341
2342         /*
2343          * Display user/system times only for single
2344          * run and when there's specified tracee.
2345          */
2346         if ((stat_config.run_count == 1) && target__none(&target))
2347                 stat_config.ru_display = true;
2348
2349         if (stat_config.run_count < 0) {
2350                 pr_err("Run count must be a positive number\n");
2351                 parse_options_usage(stat_usage, stat_options, "r", 1);
2352                 goto out;
2353         } else if (stat_config.run_count == 0) {
2354                 forever = true;
2355                 stat_config.run_count = 1;
2356         }
2357
2358         if (stat_config.walltime_run_table) {
2359                 stat_config.walltime_run = zalloc(stat_config.run_count * sizeof(stat_config.walltime_run[0]));
2360                 if (!stat_config.walltime_run) {
2361                         pr_err("failed to setup -r option");
2362                         goto out;
2363                 }
2364         }
2365
2366         if ((stat_config.aggr_mode == AGGR_THREAD) &&
2367                 !target__has_task(&target)) {
2368                 if (!target.system_wide || target.cpu_list) {
2369                         fprintf(stderr, "The --per-thread option is only "
2370                                 "available when monitoring via -p -t -a "
2371                                 "options or only --per-thread.\n");
2372                         parse_options_usage(NULL, stat_options, "p", 1);
2373                         parse_options_usage(NULL, stat_options, "t", 1);
2374                         goto out;
2375                 }
2376         }
2377
2378         /*
2379          * no_aggr, cgroup are for system-wide only
2380          * --per-thread is aggregated per thread, we dont mix it with cpu mode
2381          */
2382         if (((stat_config.aggr_mode != AGGR_GLOBAL &&
2383               stat_config.aggr_mode != AGGR_THREAD) ||
2384              (nr_cgroups || stat_config.cgroup_list)) &&
2385             !target__has_cpu(&target)) {
2386                 fprintf(stderr, "both cgroup and no-aggregation "
2387                         "modes only available in system-wide mode\n");
2388
2389                 parse_options_usage(stat_usage, stat_options, "G", 1);
2390                 parse_options_usage(NULL, stat_options, "A", 1);
2391                 parse_options_usage(NULL, stat_options, "a", 1);
2392                 parse_options_usage(NULL, stat_options, "for-each-cgroup", 0);
2393                 goto out;
2394         }
2395
2396         if (stat_config.iostat_run) {
2397                 status = iostat_prepare(evsel_list, &stat_config);
2398                 if (status)
2399                         goto out;
2400                 if (iostat_mode == IOSTAT_LIST) {
2401                         iostat_list(evsel_list, &stat_config);
2402                         goto out;
2403                 } else if (verbose)
2404                         iostat_list(evsel_list, &stat_config);
2405                 if (iostat_mode == IOSTAT_RUN && !target__has_cpu(&target))
2406                         target.system_wide = true;
2407         }
2408
2409         if (add_default_attributes())
2410                 goto out;
2411
2412         if (stat_config.cgroup_list) {
2413                 if (nr_cgroups > 0) {
2414                         pr_err("--cgroup and --for-each-cgroup cannot be used together\n");
2415                         parse_options_usage(stat_usage, stat_options, "G", 1);
2416                         parse_options_usage(NULL, stat_options, "for-each-cgroup", 0);
2417                         goto out;
2418                 }
2419
2420                 if (evlist__expand_cgroup(evsel_list, stat_config.cgroup_list,
2421                                           &stat_config.metric_events, true) < 0) {
2422                         parse_options_usage(stat_usage, stat_options,
2423                                             "for-each-cgroup", 0);
2424                         goto out;
2425                 }
2426         }
2427
2428         if ((stat_config.aggr_mode == AGGR_THREAD) && (target.system_wide))
2429                 target.per_thread = true;
2430
2431         if (evlist__fix_hybrid_cpus(evsel_list, target.cpu_list)) {
2432                 pr_err("failed to use cpu list %s\n", target.cpu_list);
2433                 goto out;
2434         }
2435
2436         target.hybrid = perf_pmu__has_hybrid();
2437         if (evlist__create_maps(evsel_list, &target) < 0) {
2438                 if (target__has_task(&target)) {
2439                         pr_err("Problems finding threads of monitor\n");
2440                         parse_options_usage(stat_usage, stat_options, "p", 1);
2441                         parse_options_usage(NULL, stat_options, "t", 1);
2442                 } else if (target__has_cpu(&target)) {
2443                         perror("failed to parse CPUs map");
2444                         parse_options_usage(stat_usage, stat_options, "C", 1);
2445                         parse_options_usage(NULL, stat_options, "a", 1);
2446                 }
2447                 goto out;
2448         }
2449
2450         evlist__check_cpu_maps(evsel_list);
2451
2452         /*
2453          * Initialize thread_map with comm names,
2454          * so we could print it out on output.
2455          */
2456         if (stat_config.aggr_mode == AGGR_THREAD) {
2457                 thread_map__read_comms(evsel_list->core.threads);
2458                 if (target.system_wide) {
2459                         if (runtime_stat_new(&stat_config,
2460                                 perf_thread_map__nr(evsel_list->core.threads))) {
2461                                 goto out;
2462                         }
2463                 }
2464         }
2465
2466         if (stat_config.aggr_mode == AGGR_NODE)
2467                 cpu__setup_cpunode_map();
2468
2469         if (stat_config.times && interval)
2470                 interval_count = true;
2471         else if (stat_config.times && !interval) {
2472                 pr_err("interval-count option should be used together with "
2473                                 "interval-print.\n");
2474                 parse_options_usage(stat_usage, stat_options, "interval-count", 0);
2475                 parse_options_usage(stat_usage, stat_options, "I", 1);
2476                 goto out;
2477         }
2478
2479         if (timeout && timeout < 100) {
2480                 if (timeout < 10) {
2481                         pr_err("timeout must be >= 10ms.\n");
2482                         parse_options_usage(stat_usage, stat_options, "timeout", 0);
2483                         goto out;
2484                 } else
2485                         pr_warning("timeout < 100ms. "
2486                                    "The overhead percentage could be high in some cases. "
2487                                    "Please proceed with caution.\n");
2488         }
2489         if (timeout && interval) {
2490                 pr_err("timeout option is not supported with interval-print.\n");
2491                 parse_options_usage(stat_usage, stat_options, "timeout", 0);
2492                 parse_options_usage(stat_usage, stat_options, "I", 1);
2493                 goto out;
2494         }
2495
2496         if (evlist__alloc_stats(evsel_list, interval))
2497                 goto out;
2498
2499         if (perf_stat_init_aggr_mode())
2500                 goto out;
2501
2502         /*
2503          * Set sample_type to PERF_SAMPLE_IDENTIFIER, which should be harmless
2504          * while avoiding that older tools show confusing messages.
2505          *
2506          * However for pipe sessions we need to keep it zero,
2507          * because script's perf_evsel__check_attr is triggered
2508          * by attr->sample_type != 0, and we can't run it on
2509          * stat sessions.
2510          */
2511         stat_config.identifier = !(STAT_RECORD && perf_stat.data.is_pipe);
2512
2513         /*
2514          * We dont want to block the signals - that would cause
2515          * child tasks to inherit that and Ctrl-C would not work.
2516          * What we want is for Ctrl-C to work in the exec()-ed
2517          * task, but being ignored by perf stat itself:
2518          */
2519         atexit(sig_atexit);
2520         if (!forever)
2521                 signal(SIGINT,  skip_signal);
2522         signal(SIGCHLD, skip_signal);
2523         signal(SIGALRM, skip_signal);
2524         signal(SIGABRT, skip_signal);
2525
2526         if (evlist__initialize_ctlfd(evsel_list, stat_config.ctl_fd, stat_config.ctl_fd_ack))
2527                 goto out;
2528
2529         status = 0;
2530         for (run_idx = 0; forever || run_idx < stat_config.run_count; run_idx++) {
2531                 if (stat_config.run_count != 1 && verbose > 0)
2532                         fprintf(output, "[ perf stat: executing run #%d ... ]\n",
2533                                 run_idx + 1);
2534
2535                 if (run_idx != 0)
2536                         evlist__reset_prev_raw_counts(evsel_list);
2537
2538                 status = run_perf_stat(argc, argv, run_idx);
2539                 if (forever && status != -1 && !interval) {
2540                         print_counters(NULL, argc, argv);
2541                         perf_stat__reset_stats();
2542                 }
2543         }
2544
2545         if (!forever && status != -1 && (!interval || stat_config.summary))
2546                 print_counters(NULL, argc, argv);
2547
2548         evlist__finalize_ctlfd(evsel_list);
2549
2550         if (STAT_RECORD) {
2551                 /*
2552                  * We synthesize the kernel mmap record just so that older tools
2553                  * don't emit warnings about not being able to resolve symbols
2554                  * due to /proc/sys/kernel/kptr_restrict settings and instead provide
2555                  * a saner message about no samples being in the perf.data file.
2556                  *
2557                  * This also serves to suppress a warning about f_header.data.size == 0
2558                  * in header.c at the moment 'perf stat record' gets introduced, which
2559                  * is not really needed once we start adding the stat specific PERF_RECORD_
2560                  * records, but the need to suppress the kptr_restrict messages in older
2561                  * tools remain  -acme
2562                  */
2563                 int fd = perf_data__fd(&perf_stat.data);
2564
2565                 err = perf_event__synthesize_kernel_mmap((void *)&perf_stat,
2566                                                          process_synthesized_event,
2567                                                          &perf_stat.session->machines.host);
2568                 if (err) {
2569                         pr_warning("Couldn't synthesize the kernel mmap record, harmless, "
2570                                    "older tools may produce warnings about this file\n.");
2571                 }
2572
2573                 if (!interval) {
2574                         if (WRITE_STAT_ROUND_EVENT(walltime_nsecs_stats.max, FINAL))
2575                                 pr_err("failed to write stat round event\n");
2576                 }
2577
2578                 if (!perf_stat.data.is_pipe) {
2579                         perf_stat.session->header.data_size += perf_stat.bytes_written;
2580                         perf_session__write_header(perf_stat.session, evsel_list, fd, true);
2581                 }
2582
2583                 evlist__close(evsel_list);
2584                 perf_session__delete(perf_stat.session);
2585         }
2586
2587         perf_stat__exit_aggr_mode();
2588         evlist__free_stats(evsel_list);
2589 out:
2590         if (stat_config.iostat_run)
2591                 iostat_release(evsel_list);
2592
2593         zfree(&stat_config.walltime_run);
2594
2595         if (smi_cost && smi_reset)
2596                 sysfs__write_int(FREEZE_ON_SMI_PATH, 0);
2597
2598         evlist__delete(evsel_list);
2599
2600         metricgroup__rblist_exit(&stat_config.metric_events);
2601         runtime_stat_delete(&stat_config);
2602         evlist__close_control(stat_config.ctl_fd, stat_config.ctl_fd_ack, &stat_config.ctl_fd_close);
2603
2604         return status;
2605 }