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 #include <internal/threadmap.h>
  97
  98 #define DEFAULT_SEPARATOR       " "
  99 #define FREEZE_ON_SMI_PATH      "devices/cpu/freeze_on_smi"
 100
 101 static void print_counters(struct timespec *ts, int argc, const char **argv);
 102
 103 /* Default events used for perf stat -T */
 104 static const char *transaction_attrs = {
 105         "task-clock,"
 106         "{"
 107         "instructions,"
 108         "cycles,"
 109         "cpu/cycles-t/,"
 110         "cpu/tx-start/,"
 111         "cpu/el-start/,"
 112         "cpu/cycles-ct/"
 113         "}"
 114 };
 115
 116 /* More limited version when the CPU does not have all events. */
 117 static const char * transaction_limited_attrs = {
 118         "task-clock,"
 119         "{"
 120         "instructions,"
 121         "cycles,"
 122         "cpu/cycles-t/,"
 123         "cpu/tx-start/"
 124         "}"
 125 };
 126
 127 static const char * topdown_attrs[] = {
 128         "topdown-total-slots",
 129         "topdown-slots-retired",
 130         "topdown-recovery-bubbles",
 131         "topdown-fetch-bubbles",
 132         "topdown-slots-issued",
 133         NULL,
 134 };
 135
 136 static const char *topdown_metric_attrs[] = {
 137         "slots",
 138         "topdown-retiring",
 139         "topdown-bad-spec",
 140         "topdown-fe-bound",
 141         "topdown-be-bound",
 142         NULL,
 143 };
 144
 145 static const char *topdown_metric_L2_attrs[] = {
 146         "slots",
 147         "topdown-retiring",
 148         "topdown-bad-spec",
 149         "topdown-fe-bound",
 150         "topdown-be-bound",
 151         "topdown-heavy-ops",
 152         "topdown-br-mispredict",
 153         "topdown-fetch-lat",
 154         "topdown-mem-bound",
 155         NULL,
 156 };
 157
 158 #define TOPDOWN_MAX_LEVEL                       2
 159
 160 static const char *smi_cost_attrs = {
 161         "{"
 162         "msr/aperf/,"
 163         "msr/smi/,"
 164         "cycles"
 165         "}"
 166 };
 167
 168 static struct evlist    *evsel_list;
 169 static bool all_counters_use_bpf = true;
 170
 171 static struct target target = {
 172         .uid    = UINT_MAX,
 173 };
 174
 175 #define METRIC_ONLY_LEN 20
 176
 177 static volatile sig_atomic_t    child_pid                       = -1;
 178 static int                      detailed_run                    =  0;
 179 static bool                     transaction_run;
 180 static bool                     topdown_run                     = false;
 181 static bool                     smi_cost                        = false;
 182 static bool                     smi_reset                       = false;
 183 static int                      big_num_opt                     =  -1;
 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 static char                     *metrics;
 195
 196 struct perf_stat {
 197         bool                     record;
 198         struct perf_data         data;
 199         struct perf_session     *session;
 200         u64                      bytes_written;
 201         struct perf_tool         tool;
 202         bool                     maps_allocated;
 203         struct perf_cpu_map     *cpus;
 204         struct perf_thread_map *threads;
 205         enum aggr_mode           aggr_mode;
 206 };
 207
 208 static struct perf_stat         perf_stat;
 209 #define STAT_RECORD             perf_stat.record
 210
 211 static volatile sig_atomic_t done = 0;
 212
 213 static struct perf_stat_config stat_config = {
 214         .aggr_mode              = AGGR_GLOBAL,
 215         .scale                  = true,
 216         .unit_width             = 4, /* strlen("unit") */
 217         .run_count              = 1,
 218         .metric_only_len        = METRIC_ONLY_LEN,
 219         .walltime_nsecs_stats   = &walltime_nsecs_stats,
 220         .ru_stats               = &ru_stats,
 221         .big_num                = true,
 222         .ctl_fd                 = -1,
 223         .ctl_fd_ack             = -1,
 224         .iostat_run             = false,
 225 };
 226
 227 static bool cpus_map_matched(struct evsel *a, struct evsel *b)
 228 {
 229         if (!a->core.cpus && !b->core.cpus)
 230                 return true;
 231
 232         if (!a->core.cpus || !b->core.cpus)
 233                 return false;
 234
 235         if (perf_cpu_map__nr(a->core.cpus) != perf_cpu_map__nr(b->core.cpus))
 236                 return false;
 237
 238         for (int i = 0; i < perf_cpu_map__nr(a->core.cpus); i++) {
 239                 if (perf_cpu_map__cpu(a->core.cpus, i).cpu !=
 240                     perf_cpu_map__cpu(b->core.cpus, i).cpu)
 241                         return false;
 242         }
 243
 244         return true;
 245 }
 246
 247 static void evlist__check_cpu_maps(struct evlist *evlist)
 248 {
 249         struct evsel *evsel, *pos, *leader;
 250         char buf[1024];
 251
 252         if (evlist__has_hybrid(evlist))
 253                 evlist__warn_hybrid_group(evlist);
 254
 255         evlist__for_each_entry(evlist, evsel) {
 256                 leader = evsel__leader(evsel);
 257
 258                 /* Check that leader matches cpus with each member. */
 259                 if (leader == evsel)
 260                         continue;
 261                 if (cpus_map_matched(leader, evsel))
 262                         continue;
 263
 264                 /* If there's mismatch disable the group and warn user. */
 265                 WARN_ONCE(1, "WARNING: grouped events cpus do not match, disabling group:\n");
 266                 evsel__group_desc(leader, buf, sizeof(buf));
 267                 pr_warning("  %s\n", buf);
 268
 269                 if (verbose > 0) {
 270                         cpu_map__snprint(leader->core.cpus, buf, sizeof(buf));
 271                         pr_warning("     %s: %s\n", leader->name, buf);
 272                         cpu_map__snprint(evsel->core.cpus, buf, sizeof(buf));
 273                         pr_warning("     %s: %s\n", evsel->name, buf);
 274                 }
 275
 276                 for_each_group_evsel(pos, leader)
 277                         evsel__remove_from_group(pos, leader);
 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         evlist__reset_stats(evsel_list);
 296         perf_stat__reset_shadow_stats();
 297 }
 298
 299 static int process_synthesized_event(struct perf_tool *tool __maybe_unused,
 300                                      union perf_event *event,
 301                                      struct perf_sample *sample __maybe_unused,
 302                                      struct machine *machine __maybe_unused)
 303 {
 304         if (perf_data__write(&perf_stat.data, event, event->header.size) < 0) {
 305                 pr_err("failed to write perf data, error: %m\n");
 306                 return -1;
 307         }
 308
 309         perf_stat.bytes_written += event->header.size;
 310         return 0;
 311 }
 312
 313 static int write_stat_round_event(u64 tm, u64 type)
 314 {
 315         return perf_event__synthesize_stat_round(NULL, tm, type,
 316                                                  process_synthesized_event,
 317                                                  NULL);
 318 }
 319
 320 #define WRITE_STAT_ROUND_EVENT(time, interval) \
 321         write_stat_round_event(time, PERF_STAT_ROUND_TYPE__ ## interval)
 322
 323 #define SID(e, x, y) xyarray__entry(e->core.sample_id, x, y)
 324
 325 static int evsel__write_stat_event(struct evsel *counter, int cpu_map_idx, u32 thread,
 326                                    struct perf_counts_values *count)
 327 {
 328         struct perf_sample_id *sid = SID(counter, cpu_map_idx, thread);
 329         struct perf_cpu cpu = perf_cpu_map__cpu(evsel__cpus(counter), cpu_map_idx);
 330
 331         return perf_event__synthesize_stat(NULL, cpu, thread, sid->id, count,
 332                                            process_synthesized_event, NULL);
 333 }
 334
 335 static int read_single_counter(struct evsel *counter, int cpu_map_idx,
 336                                int thread, struct timespec *rs)
 337 {
 338         switch(counter->tool_event) {
 339                 case PERF_TOOL_DURATION_TIME: {
 340                         u64 val = rs->tv_nsec + rs->tv_sec*1000000000ULL;
 341                         struct perf_counts_values *count =
 342                                 perf_counts(counter->counts, cpu_map_idx, thread);
 343                         count->ena = count->run = val;
 344                         count->val = val;
 345                         return 0;
 346                 }
 347                 case PERF_TOOL_USER_TIME:
 348                 case PERF_TOOL_SYSTEM_TIME: {
 349                         u64 val;
 350                         struct perf_counts_values *count =
 351                                 perf_counts(counter->counts, cpu_map_idx, thread);
 352                         if (counter->tool_event == PERF_TOOL_USER_TIME)
 353                                 val = ru_stats.ru_utime_usec_stat.mean;
 354                         else
 355                                 val = ru_stats.ru_stime_usec_stat.mean;
 356                         count->ena = count->run = val;
 357                         count->val = val;
 358                         return 0;
 359                 }
 360                 default:
 361                 case PERF_TOOL_NONE:
 362                         return evsel__read_counter(counter, cpu_map_idx, thread);
 363                 case PERF_TOOL_MAX:
 364                         /* This should never be reached */
 365                         return 0;
 366         }
 367 }
 368
 369 /*
 370  * Read out the results of a single counter:
 371  * do not aggregate counts across CPUs in system-wide mode
 372  */
 373 static int read_counter_cpu(struct evsel *counter, struct timespec *rs, int cpu_map_idx)
 374 {
 375         int nthreads = perf_thread_map__nr(evsel_list->core.threads);
 376         int thread;
 377
 378         if (!counter->supported)
 379                 return -ENOENT;
 380
 381         for (thread = 0; thread < nthreads; thread++) {
 382                 struct perf_counts_values *count;
 383
 384                 count = perf_counts(counter->counts, cpu_map_idx, thread);
 385
 386                 /*
 387                  * The leader's group read loads data into its group members
 388                  * (via evsel__read_counter()) and sets their count->loaded.
 389                  */
 390                 if (!perf_counts__is_loaded(counter->counts, cpu_map_idx, thread) &&
 391                     read_single_counter(counter, cpu_map_idx, thread, rs)) {
 392                         counter->counts->scaled = -1;
 393                         perf_counts(counter->counts, cpu_map_idx, thread)->ena = 0;
 394                         perf_counts(counter->counts, cpu_map_idx, thread)->run = 0;
 395                         return -1;
 396                 }
 397
 398                 perf_counts__set_loaded(counter->counts, cpu_map_idx, thread, false);
 399
 400                 if (STAT_RECORD) {
 401                         if (evsel__write_stat_event(counter, cpu_map_idx, thread, count)) {
 402                                 pr_err("failed to write stat event\n");
 403                                 return -1;
 404                         }
 405                 }
 406
 407                 if (verbose > 1) {
 408                         fprintf(stat_config.output,
 409                                 "%s: %d: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
 410                                         evsel__name(counter),
 411                                         perf_cpu_map__cpu(evsel__cpus(counter),
 412                                                           cpu_map_idx).cpu,
 413                                         count->val, count->ena, count->run);
 414                 }
 415         }
 416
 417         return 0;
 418 }
 419
 420 static int read_affinity_counters(struct timespec *rs)
 421 {
 422         struct evlist_cpu_iterator evlist_cpu_itr;
 423         struct affinity saved_affinity, *affinity;
 424
 425         if (all_counters_use_bpf)
 426                 return 0;
 427
 428         if (!target__has_cpu(&target) || target__has_per_thread(&target))
 429                 affinity = NULL;
 430         else if (affinity__setup(&saved_affinity) < 0)
 431                 return -1;
 432         else
 433                 affinity = &saved_affinity;
 434
 435         evlist__for_each_cpu(evlist_cpu_itr, evsel_list, affinity) {
 436                 struct evsel *counter = evlist_cpu_itr.evsel;
 437
 438                 if (evsel__is_bpf(counter))
 439                         continue;
 440
 441                 if (!counter->err) {
 442                         counter->err = read_counter_cpu(counter, rs,
 443                                                         evlist_cpu_itr.cpu_map_idx);
 444                 }
 445         }
 446         if (affinity)
 447                 affinity__cleanup(&saved_affinity);
 448
 449         return 0;
 450 }
 451
 452 static int read_bpf_map_counters(void)
 453 {
 454         struct evsel *counter;
 455         int err;
 456
 457         evlist__for_each_entry(evsel_list, counter) {
 458                 if (!evsel__is_bpf(counter))
 459                         continue;
 460
 461                 err = bpf_counter__read(counter);
 462                 if (err)
 463                         return err;
 464         }
 465         return 0;
 466 }
 467
 468 static int read_counters(struct timespec *rs)
 469 {
 470         if (!stat_config.stop_read_counter) {
 471                 if (read_bpf_map_counters() ||
 472                     read_affinity_counters(rs))
 473                         return -1;
 474         }
 475         return 0;
 476 }
 477
 478 static void process_counters(void)
 479 {
 480         struct evsel *counter;
 481
 482         evlist__for_each_entry(evsel_list, counter) {
 483                 if (counter->err)
 484                         pr_debug("failed to read counter %s\n", counter->name);
 485                 if (counter->err == 0 && perf_stat_process_counter(&stat_config, counter))
 486                         pr_warning("failed to process counter %s\n", counter->name);
 487                 counter->err = 0;
 488         }
 489
 490         perf_stat_merge_counters(&stat_config, evsel_list);
 491         perf_stat_process_percore(&stat_config, evsel_list);
 492         perf_stat_process_shadow_stats(&stat_config, evsel_list);
 493 }
 494
 495 static void process_interval(void)
 496 {
 497         struct timespec ts, rs;
 498
 499         clock_gettime(CLOCK_MONOTONIC, &ts);
 500         diff_timespec(&rs, &ts, &ref_time);
 501
 502         perf_stat__reset_shadow_per_stat(&rt_stat);
 503         evlist__reset_aggr_stats(evsel_list);
 504
 505         if (read_counters(&rs) == 0)
 506                 process_counters();
 507
 508         if (STAT_RECORD) {
 509                 if (WRITE_STAT_ROUND_EVENT(rs.tv_sec * NSEC_PER_SEC + rs.tv_nsec, INTERVAL))
 510                         pr_err("failed to write stat round event\n");
 511         }
 512
 513         init_stats(&walltime_nsecs_stats);
 514         update_stats(&walltime_nsecs_stats, stat_config.interval * 1000000ULL);
 515         print_counters(&rs, 0, NULL);
 516 }
 517
 518 static bool handle_interval(unsigned int interval, int *times)
 519 {
 520         if (interval) {
 521                 process_interval();
 522                 if (interval_count && !(--(*times)))
 523                         return true;
 524         }
 525         return false;
 526 }
 527
 528 static int enable_counters(void)
 529 {
 530         struct evsel *evsel;
 531         int err;
 532
 533         evlist__for_each_entry(evsel_list, evsel) {
 534                 if (!evsel__is_bpf(evsel))
 535                         continue;
 536
 537                 err = bpf_counter__enable(evsel);
 538                 if (err)
 539                         return err;
 540         }
 541
 542         if (!target__enable_on_exec(&target)) {
 543                 if (!all_counters_use_bpf)
 544                         evlist__enable(evsel_list);
 545         }
 546         return 0;
 547 }
 548
 549 static void disable_counters(void)
 550 {
 551         struct evsel *counter;
 552
 553         /*
 554          * If we don't have tracee (attaching to task or cpu), counters may
 555          * still be running. To get accurate group ratios, we must stop groups
 556          * from counting before reading their constituent counters.
 557          */
 558         if (!target__none(&target)) {
 559                 evlist__for_each_entry(evsel_list, counter)
 560                         bpf_counter__disable(counter);
 561                 if (!all_counters_use_bpf)
 562                         evlist__disable(evsel_list);
 563         }
 564 }
 565
 566 static volatile sig_atomic_t workload_exec_errno;
 567
 568 /*
 569  * evlist__prepare_workload will send a SIGUSR1
 570  * if the fork fails, since we asked by setting its
 571  * want_signal to true.
 572  */
 573 static void workload_exec_failed_signal(int signo __maybe_unused, siginfo_t *info,
 574                                         void *ucontext __maybe_unused)
 575 {
 576         workload_exec_errno = info->si_value.sival_int;
 577 }
 578
 579 static bool evsel__should_store_id(struct evsel *counter)
 580 {
 581         return STAT_RECORD || counter->core.attr.read_format & PERF_FORMAT_ID;
 582 }
 583
 584 static bool is_target_alive(struct target *_target,
 585                             struct perf_thread_map *threads)
 586 {
 587         struct stat st;
 588         int i;
 589
 590         if (!target__has_task(_target))
 591                 return true;
 592
 593         for (i = 0; i < threads->nr; i++) {
 594                 char path[PATH_MAX];
 595
 596                 scnprintf(path, PATH_MAX, "%s/%d", procfs__mountpoint(),
 597                           threads->map[i].pid);
 598
 599                 if (!stat(path, &st))
 600                         return true;
 601         }
 602
 603         return false;
 604 }
 605
 606 static void process_evlist(struct evlist *evlist, unsigned int interval)
 607 {
 608         enum evlist_ctl_cmd cmd = EVLIST_CTL_CMD_UNSUPPORTED;
 609
 610         if (evlist__ctlfd_process(evlist, &cmd) > 0) {
 611                 switch (cmd) {
 612                 case EVLIST_CTL_CMD_ENABLE:
 613                         __fallthrough;
 614                 case EVLIST_CTL_CMD_DISABLE:
 615                         if (interval)
 616                                 process_interval();
 617                         break;
 618                 case EVLIST_CTL_CMD_SNAPSHOT:
 619                 case EVLIST_CTL_CMD_ACK:
 620                 case EVLIST_CTL_CMD_UNSUPPORTED:
 621                 case EVLIST_CTL_CMD_EVLIST:
 622                 case EVLIST_CTL_CMD_STOP:
 623                 case EVLIST_CTL_CMD_PING:
 624                 default:
 625                         break;
 626                 }
 627         }
 628 }
 629
 630 static void compute_tts(struct timespec *time_start, struct timespec *time_stop,
 631                         int *time_to_sleep)
 632 {
 633         int tts = *time_to_sleep;
 634         struct timespec time_diff;
 635
 636         diff_timespec(&time_diff, time_stop, time_start);
 637
 638         tts -= time_diff.tv_sec * MSEC_PER_SEC +
 639                time_diff.tv_nsec / NSEC_PER_MSEC;
 640
 641         if (tts < 0)
 642                 tts = 0;
 643
 644         *time_to_sleep = tts;
 645 }
 646
 647 static int dispatch_events(bool forks, int timeout, int interval, int *times)
 648 {
 649         int child_exited = 0, status = 0;
 650         int time_to_sleep, sleep_time;
 651         struct timespec time_start, time_stop;
 652
 653         if (interval)
 654                 sleep_time = interval;
 655         else if (timeout)
 656                 sleep_time = timeout;
 657         else
 658                 sleep_time = 1000;
 659
 660         time_to_sleep = sleep_time;
 661
 662         while (!done) {
 663                 if (forks)
 664                         child_exited = waitpid(child_pid, &status, WNOHANG);
 665                 else
 666                         child_exited = !is_target_alive(&target, evsel_list->core.threads) ? 1 : 0;
 667
 668                 if (child_exited)
 669                         break;
 670
 671                 clock_gettime(CLOCK_MONOTONIC, &time_start);
 672                 if (!(evlist__poll(evsel_list, time_to_sleep) > 0)) { /* poll timeout or EINTR */
 673                         if (timeout || handle_interval(interval, times))
 674                                 break;
 675                         time_to_sleep = sleep_time;
 676                 } else { /* fd revent */
 677                         process_evlist(evsel_list, interval);
 678                         clock_gettime(CLOCK_MONOTONIC, &time_stop);
 679                         compute_tts(&time_start, &time_stop, &time_to_sleep);
 680                 }
 681         }
 682
 683         return status;
 684 }
 685
 686 enum counter_recovery {
 687         COUNTER_SKIP,
 688         COUNTER_RETRY,
 689         COUNTER_FATAL,
 690 };
 691
 692 static enum counter_recovery stat_handle_error(struct evsel *counter)
 693 {
 694         char msg[BUFSIZ];
 695         /*
 696          * PPC returns ENXIO for HW counters until 2.6.37
 697          * (behavior changed with commit b0a873e).
 698          */
 699         if (errno == EINVAL || errno == ENOSYS ||
 700             errno == ENOENT || errno == EOPNOTSUPP ||
 701             errno == ENXIO) {
 702                 if (verbose > 0)
 703                         ui__warning("%s event is not supported by the kernel.\n",
 704                                     evsel__name(counter));
 705                 counter->supported = false;
 706                 /*
 707                  * errored is a sticky flag that means one of the counter's
 708                  * cpu event had a problem and needs to be reexamined.
 709                  */
 710                 counter->errored = true;
 711
 712                 if ((evsel__leader(counter) != counter) ||
 713                     !(counter->core.leader->nr_members > 1))
 714                         return COUNTER_SKIP;
 715         } else if (evsel__fallback(counter, errno, msg, sizeof(msg))) {
 716                 if (verbose > 0)
 717                         ui__warning("%s\n", msg);
 718                 return COUNTER_RETRY;
 719         } else if (target__has_per_thread(&target) &&
 720                    evsel_list->core.threads &&
 721                    evsel_list->core.threads->err_thread != -1) {
 722                 /*
 723                  * For global --per-thread case, skip current
 724                  * error thread.
 725                  */
 726                 if (!thread_map__remove(evsel_list->core.threads,
 727                                         evsel_list->core.threads->err_thread)) {
 728                         evsel_list->core.threads->err_thread = -1;
 729                         return COUNTER_RETRY;
 730                 }
 731         }
 732
 733         evsel__open_strerror(counter, &target, errno, msg, sizeof(msg));
 734         ui__error("%s\n", msg);
 735
 736         if (child_pid != -1)
 737                 kill(child_pid, SIGTERM);
 738         return COUNTER_FATAL;
 739 }
 740
 741 static int __run_perf_stat(int argc, const char **argv, int run_idx)
 742 {
 743         int interval = stat_config.interval;
 744         int times = stat_config.times;
 745         int timeout = stat_config.timeout;
 746         char msg[BUFSIZ];
 747         unsigned long long t0, t1;
 748         struct evsel *counter;
 749         size_t l;
 750         int status = 0;
 751         const bool forks = (argc > 0);
 752         bool is_pipe = STAT_RECORD ? perf_stat.data.is_pipe : false;
 753         struct evlist_cpu_iterator evlist_cpu_itr;
 754         struct affinity saved_affinity, *affinity = NULL;
 755         int err;
 756         bool second_pass = false;
 757
 758         if (forks) {
 759                 if (evlist__prepare_workload(evsel_list, &target, argv, is_pipe, workload_exec_failed_signal) < 0) {
 760                         perror("failed to prepare workload");
 761                         return -1;
 762                 }
 763                 child_pid = evsel_list->workload.pid;
 764         }
 765
 766         if (!cpu_map__is_dummy(evsel_list->core.user_requested_cpus)) {
 767                 if (affinity__setup(&saved_affinity) < 0)
 768                         return -1;
 769                 affinity = &saved_affinity;
 770         }
 771
 772         evlist__for_each_entry(evsel_list, counter) {
 773                 counter->reset_group = false;
 774                 if (bpf_counter__load(counter, &target))
 775                         return -1;
 776                 if (!evsel__is_bpf(counter))
 777                         all_counters_use_bpf = false;
 778         }
 779
 780         evlist__for_each_cpu(evlist_cpu_itr, evsel_list, affinity) {
 781                 counter = evlist_cpu_itr.evsel;
 782
 783                 /*
 784                  * bperf calls evsel__open_per_cpu() in bperf__load(), so
 785                  * no need to call it again here.
 786                  */
 787                 if (target.use_bpf)
 788                         break;
 789
 790                 if (counter->reset_group || counter->errored)
 791                         continue;
 792                 if (evsel__is_bpf(counter))
 793                         continue;
 794 try_again:
 795                 if (create_perf_stat_counter(counter, &stat_config, &target,
 796                                              evlist_cpu_itr.cpu_map_idx) < 0) {
 797
 798                         /*
 799                          * Weak group failed. We cannot just undo this here
 800                          * because earlier CPUs might be in group mode, and the kernel
 801                          * doesn't support mixing group and non group reads. Defer
 802                          * it to later.
 803                          * Don't close here because we're in the wrong affinity.
 804                          */
 805                         if ((errno == EINVAL || errno == EBADF) &&
 806                                 evsel__leader(counter) != counter &&
 807                                 counter->weak_group) {
 808                                 evlist__reset_weak_group(evsel_list, counter, false);
 809                                 assert(counter->reset_group);
 810                                 second_pass = true;
 811                                 continue;
 812                         }
 813
 814                         switch (stat_handle_error(counter)) {
 815                         case COUNTER_FATAL:
 816                                 return -1;
 817                         case COUNTER_RETRY:
 818                                 goto try_again;
 819                         case COUNTER_SKIP:
 820                                 continue;
 821                         default:
 822                                 break;
 823                         }
 824
 825                 }
 826                 counter->supported = true;
 827         }
 828
 829         if (second_pass) {
 830                 /*
 831                  * Now redo all the weak group after closing them,
 832                  * and also close errored counters.
 833                  */
 834
 835                 /* First close errored or weak retry */
 836                 evlist__for_each_cpu(evlist_cpu_itr, evsel_list, affinity) {
 837                         counter = evlist_cpu_itr.evsel;
 838
 839                         if (!counter->reset_group && !counter->errored)
 840                                 continue;
 841
 842                         perf_evsel__close_cpu(&counter->core, evlist_cpu_itr.cpu_map_idx);
 843                 }
 844                 /* Now reopen weak */
 845                 evlist__for_each_cpu(evlist_cpu_itr, evsel_list, affinity) {
 846                         counter = evlist_cpu_itr.evsel;
 847
 848                         if (!counter->reset_group)
 849                                 continue;
 850 try_again_reset:
 851                         pr_debug2("reopening weak %s\n", evsel__name(counter));
 852                         if (create_perf_stat_counter(counter, &stat_config, &target,
 853                                                      evlist_cpu_itr.cpu_map_idx) < 0) {
 854
 855                                 switch (stat_handle_error(counter)) {
 856                                 case COUNTER_FATAL:
 857                                         return -1;
 858                                 case COUNTER_RETRY:
 859                                         goto try_again_reset;
 860                                 case COUNTER_SKIP:
 861                                         continue;
 862                                 default:
 863                                         break;
 864                                 }
 865                         }
 866                         counter->supported = true;
 867                 }
 868         }
 869         affinity__cleanup(affinity);
 870
 871         evlist__for_each_entry(evsel_list, counter) {
 872                 if (!counter->supported) {
 873                         perf_evsel__free_fd(&counter->core);
 874                         continue;
 875                 }
 876
 877                 l = strlen(counter->unit);
 878                 if (l > stat_config.unit_width)
 879                         stat_config.unit_width = l;
 880
 881                 if (evsel__should_store_id(counter) &&
 882                     evsel__store_ids(counter, evsel_list))
 883                         return -1;
 884         }
 885
 886         if (evlist__apply_filters(evsel_list, &counter)) {
 887                 pr_err("failed to set filter \"%s\" on event %s with %d (%s)\n",
 888                         counter->filter, evsel__name(counter), errno,
 889                         str_error_r(errno, msg, sizeof(msg)));
 890                 return -1;
 891         }
 892
 893         if (STAT_RECORD) {
 894                 int fd = perf_data__fd(&perf_stat.data);
 895
 896                 if (is_pipe) {
 897                         err = perf_header__write_pipe(perf_data__fd(&perf_stat.data));
 898                 } else {
 899                         err = perf_session__write_header(perf_stat.session, evsel_list,
 900                                                          fd, false);
 901                 }
 902
 903                 if (err < 0)
 904                         return err;
 905
 906                 err = perf_event__synthesize_stat_events(&stat_config, NULL, evsel_list,
 907                                                          process_synthesized_event, is_pipe);
 908                 if (err < 0)
 909                         return err;
 910         }
 911
 912         if (target.initial_delay) {
 913                 pr_info(EVLIST_DISABLED_MSG);
 914         } else {
 915                 err = enable_counters();
 916                 if (err)
 917                         return -1;
 918         }
 919
 920         /* Exec the command, if any */
 921         if (forks)
 922                 evlist__start_workload(evsel_list);
 923
 924         if (target.initial_delay > 0) {
 925                 usleep(target.initial_delay * USEC_PER_MSEC);
 926                 err = enable_counters();
 927                 if (err)
 928                         return -1;
 929
 930                 pr_info(EVLIST_ENABLED_MSG);
 931         }
 932
 933         t0 = rdclock();
 934         clock_gettime(CLOCK_MONOTONIC, &ref_time);
 935
 936         if (forks) {
 937                 if (interval || timeout || evlist__ctlfd_initialized(evsel_list))
 938                         status = dispatch_events(forks, timeout, interval, &times);
 939                 if (child_pid != -1) {
 940                         if (timeout)
 941                                 kill(child_pid, SIGTERM);
 942                         wait4(child_pid, &status, 0, &stat_config.ru_data);
 943                 }
 944
 945                 if (workload_exec_errno) {
 946                         const char *emsg = str_error_r(workload_exec_errno, msg, sizeof(msg));
 947                         pr_err("Workload failed: %s\n", emsg);
 948                         return -1;
 949                 }
 950
 951                 if (WIFSIGNALED(status))
 952                         psignal(WTERMSIG(status), argv[0]);
 953         } else {
 954                 status = dispatch_events(forks, timeout, interval, &times);
 955         }
 956
 957         disable_counters();
 958
 959         t1 = rdclock();
 960
 961         if (stat_config.walltime_run_table)
 962                 stat_config.walltime_run[run_idx] = t1 - t0;
 963
 964         if (interval && stat_config.summary) {
 965                 stat_config.interval = 0;
 966                 stat_config.stop_read_counter = true;
 967                 init_stats(&walltime_nsecs_stats);
 968                 update_stats(&walltime_nsecs_stats, t1 - t0);
 969
 970                 evlist__copy_prev_raw_counts(evsel_list);
 971                 evlist__reset_prev_raw_counts(evsel_list);
 972                 evlist__reset_aggr_stats(evsel_list);
 973                 perf_stat__reset_shadow_per_stat(&rt_stat);
 974         } else {
 975                 update_stats(&walltime_nsecs_stats, t1 - t0);
 976                 update_rusage_stats(&ru_stats, &stat_config.ru_data);
 977         }
 978
 979         /*
 980          * Closing a group leader splits the group, and as we only disable
 981          * group leaders, results in remaining events becoming enabled. To
 982          * avoid arbitrary skew, we must read all counters before closing any
 983          * group leaders.
 984          */
 985         if (read_counters(&(struct timespec) { .tv_nsec = t1-t0 }) == 0)
 986                 process_counters();
 987
 988         /*
 989          * We need to keep evsel_list alive, because it's processed
 990          * later the evsel_list will be closed after.
 991          */
 992         if (!STAT_RECORD)
 993                 evlist__close(evsel_list);
 994
 995         return WEXITSTATUS(status);
 996 }
 997
 998 static int run_perf_stat(int argc, const char **argv, int run_idx)
 999 {
1000         int ret;
1001
1002         if (pre_cmd) {
1003                 ret = system(pre_cmd);
1004                 if (ret)
1005                         return ret;
1006         }
1007
1008         if (sync_run)
1009                 sync();
1010
1011         ret = __run_perf_stat(argc, argv, run_idx);
1012         if (ret)
1013                 return ret;
1014
1015         if (post_cmd) {
1016                 ret = system(post_cmd);
1017                 if (ret)
1018                         return ret;
1019         }
1020
1021         return ret;
1022 }
1023
1024 static void print_counters(struct timespec *ts, int argc, const char **argv)
1025 {
1026         /* Do not print anything if we record to the pipe. */
1027         if (STAT_RECORD && perf_stat.data.is_pipe)
1028                 return;
1029         if (quiet)
1030                 return;
1031
1032         evlist__print_counters(evsel_list, &stat_config, &target, ts, argc, argv);
1033 }
1034
1035 static volatile sig_atomic_t signr = -1;
1036
1037 static void skip_signal(int signo)
1038 {
1039         if ((child_pid == -1) || stat_config.interval)
1040                 done = 1;
1041
1042         signr = signo;
1043         /*
1044          * render child_pid harmless
1045          * won't send SIGTERM to a random
1046          * process in case of race condition
1047          * and fast PID recycling
1048          */
1049         child_pid = -1;
1050 }
1051
1052 static void sig_atexit(void)
1053 {
1054         sigset_t set, oset;
1055
1056         /*
1057          * avoid race condition with SIGCHLD handler
1058          * in skip_signal() which is modifying child_pid
1059          * goal is to avoid send SIGTERM to a random
1060          * process
1061          */
1062         sigemptyset(&set);
1063         sigaddset(&set, SIGCHLD);
1064         sigprocmask(SIG_BLOCK, &set, &oset);
1065
1066         if (child_pid != -1)
1067                 kill(child_pid, SIGTERM);
1068
1069         sigprocmask(SIG_SETMASK, &oset, NULL);
1070
1071         if (signr == -1)
1072                 return;
1073
1074         signal(signr, SIG_DFL);
1075         kill(getpid(), signr);
1076 }
1077
1078 void perf_stat__set_big_num(int set)
1079 {
1080         stat_config.big_num = (set != 0);
1081 }
1082
1083 void perf_stat__set_no_csv_summary(int set)
1084 {
1085         stat_config.no_csv_summary = (set != 0);
1086 }
1087
1088 static int stat__set_big_num(const struct option *opt __maybe_unused,
1089                              const char *s __maybe_unused, int unset)
1090 {
1091         big_num_opt = unset ? 0 : 1;
1092         perf_stat__set_big_num(!unset);
1093         return 0;
1094 }
1095
1096 static int enable_metric_only(const struct option *opt __maybe_unused,
1097                               const char *s __maybe_unused, int unset)
1098 {
1099         force_metric_only = true;
1100         stat_config.metric_only = !unset;
1101         return 0;
1102 }
1103
1104 static int append_metric_groups(const struct option *opt __maybe_unused,
1105                                const char *str,
1106                                int unset __maybe_unused)
1107 {
1108         if (metrics) {
1109                 char *tmp;
1110
1111                 if (asprintf(&tmp, "%s,%s", metrics, str) < 0)
1112                         return -ENOMEM;
1113                 free(metrics);
1114                 metrics = tmp;
1115         } else {
1116                 metrics = strdup(str);
1117                 if (!metrics)
1118                         return -ENOMEM;
1119         }
1120         return 0;
1121 }
1122
1123 static int parse_control_option(const struct option *opt,
1124                                 const char *str,
1125                                 int unset __maybe_unused)
1126 {
1127         struct perf_stat_config *config = opt->value;
1128
1129         return evlist__parse_control(str, &config->ctl_fd, &config->ctl_fd_ack, &config->ctl_fd_close);
1130 }
1131
1132 static int parse_stat_cgroups(const struct option *opt,
1133                               const char *str, int unset)
1134 {
1135         if (stat_config.cgroup_list) {
1136                 pr_err("--cgroup and --for-each-cgroup cannot be used together\n");
1137                 return -1;
1138         }
1139
1140         return parse_cgroups(opt, str, unset);
1141 }
1142
1143 static int parse_hybrid_type(const struct option *opt,
1144                              const char *str,
1145                              int unset __maybe_unused)
1146 {
1147         struct evlist *evlist = *(struct evlist **)opt->value;
1148
1149         if (!list_empty(&evlist->core.entries)) {
1150                 fprintf(stderr, "Must define cputype before events/metrics\n");
1151                 return -1;
1152         }
1153
1154         evlist->hybrid_pmu_name = perf_pmu__hybrid_type_to_pmu(str);
1155         if (!evlist->hybrid_pmu_name) {
1156                 fprintf(stderr, "--cputype %s is not supported!\n", str);
1157                 return -1;
1158         }
1159
1160         return 0;
1161 }
1162
1163 static struct option stat_options[] = {
1164         OPT_BOOLEAN('T', "transaction", &transaction_run,
1165                     "hardware transaction statistics"),
1166         OPT_CALLBACK('e', "event", &evsel_list, "event",
1167                      "event selector. use 'perf list' to list available events",
1168                      parse_events_option),
1169         OPT_CALLBACK(0, "filter", &evsel_list, "filter",
1170                      "event filter", parse_filter),
1171         OPT_BOOLEAN('i', "no-inherit", &stat_config.no_inherit,
1172                     "child tasks do not inherit counters"),
1173         OPT_STRING('p', "pid", &target.pid, "pid",
1174                    "stat events on existing process id"),
1175         OPT_STRING('t', "tid", &target.tid, "tid",
1176                    "stat events on existing thread id"),
1177 #ifdef HAVE_BPF_SKEL
1178         OPT_STRING('b', "bpf-prog", &target.bpf_str, "bpf-prog-id",
1179                    "stat events on existing bpf program id"),
1180         OPT_BOOLEAN(0, "bpf-counters", &target.use_bpf,
1181                     "use bpf program to count events"),
1182         OPT_STRING(0, "bpf-attr-map", &target.attr_map, "attr-map-path",
1183                    "path to perf_event_attr map"),
1184 #endif
1185         OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
1186                     "system-wide collection from all CPUs"),
1187         OPT_BOOLEAN(0, "scale", &stat_config.scale,
1188                     "Use --no-scale to disable counter scaling for multiplexing"),
1189         OPT_INCR('v', "verbose", &verbose,
1190                     "be more verbose (show counter open errors, etc)"),
1191         OPT_INTEGER('r', "repeat", &stat_config.run_count,
1192                     "repeat command and print average + stddev (max: 100, forever: 0)"),
1193         OPT_BOOLEAN(0, "table", &stat_config.walltime_run_table,
1194                     "display details about each run (only with -r option)"),
1195         OPT_BOOLEAN('n', "null", &stat_config.null_run,
1196                     "null run - dont start any counters"),
1197         OPT_INCR('d', "detailed", &detailed_run,
1198                     "detailed run - start a lot of events"),
1199         OPT_BOOLEAN('S', "sync", &sync_run,
1200                     "call sync() before starting a run"),
1201         OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
1202                            "print large numbers with thousands\' separators",
1203                            stat__set_big_num),
1204         OPT_STRING('C', "cpu", &target.cpu_list, "cpu",
1205                     "list of cpus to monitor in system-wide"),
1206         OPT_SET_UINT('A', "no-aggr", &stat_config.aggr_mode,
1207                     "disable CPU count aggregation", AGGR_NONE),
1208         OPT_BOOLEAN(0, "no-merge", &stat_config.no_merge, "Do not merge identical named events"),
1209         OPT_BOOLEAN(0, "hybrid-merge", &stat_config.hybrid_merge,
1210                     "Merge identical named hybrid events"),
1211         OPT_STRING('x', "field-separator", &stat_config.csv_sep, "separator",
1212                    "print counts with custom separator"),
1213         OPT_BOOLEAN('j', "json-output", &stat_config.json_output,
1214                    "print counts in JSON format"),
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", &target.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                      append_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", &quiet,
1278                         "don't print any output, messages or warnings (useful with record)"),
1279         OPT_CALLBACK(0, "cputype", &evsel_list, "hybrid cpu type",
1280                      "Only enable events on applying cpu with this type "
1281                      "for hybrid platform (e.g. core or atom)",
1282                      parse_hybrid_type),
1283 #ifdef HAVE_LIBPFM
1284         OPT_CALLBACK(0, "pfm-events", &evsel_list, "event",
1285                 "libpfm4 event selector. use 'perf list' to list available events",
1286                 parse_libpfm_events_option),
1287 #endif
1288         OPT_CALLBACK(0, "control", &stat_config, "fd:ctl-fd[,ack-fd] or fifo:ctl-fifo[,ack-fifo]",
1289                      "Listen on ctl-fd descriptor for command to control measurement ('enable': enable events, 'disable': disable events).\n"
1290                      "\t\t\t  Optionally send control command completion ('ack\\n') to ack-fd descriptor.\n"
1291                      "\t\t\t  Alternatively, ctl-fifo / ack-fifo will be opened and used as ctl-fd / ack-fd.",
1292                       parse_control_option),
1293         OPT_CALLBACK_OPTARG(0, "iostat", &evsel_list, &stat_config, "default",
1294                             "measure I/O performance metrics provided by arch/platform",
1295                             iostat_parse),
1296         OPT_END()
1297 };
1298
1299 static const char *const aggr_mode__string[] = {
1300         [AGGR_CORE] = "core",
1301         [AGGR_DIE] = "die",
1302         [AGGR_GLOBAL] = "global",
1303         [AGGR_NODE] = "node",
1304         [AGGR_NONE] = "none",
1305         [AGGR_SOCKET] = "socket",
1306         [AGGR_THREAD] = "thread",
1307         [AGGR_UNSET] = "unset",
1308 };
1309
1310 static struct aggr_cpu_id perf_stat__get_socket(struct perf_stat_config *config __maybe_unused,
1311                                                 struct perf_cpu cpu)
1312 {
1313         return aggr_cpu_id__socket(cpu, /*data=*/NULL);
1314 }
1315
1316 static struct aggr_cpu_id perf_stat__get_die(struct perf_stat_config *config __maybe_unused,
1317                                              struct perf_cpu cpu)
1318 {
1319         return aggr_cpu_id__die(cpu, /*data=*/NULL);
1320 }
1321
1322 static struct aggr_cpu_id perf_stat__get_core(struct perf_stat_config *config __maybe_unused,
1323                                               struct perf_cpu cpu)
1324 {
1325         return aggr_cpu_id__core(cpu, /*data=*/NULL);
1326 }
1327
1328 static struct aggr_cpu_id perf_stat__get_node(struct perf_stat_config *config __maybe_unused,
1329                                               struct perf_cpu cpu)
1330 {
1331         return aggr_cpu_id__node(cpu, /*data=*/NULL);
1332 }
1333
1334 static struct aggr_cpu_id perf_stat__get_global(struct perf_stat_config *config __maybe_unused,
1335                                                 struct perf_cpu cpu)
1336 {
1337         return aggr_cpu_id__global(cpu, /*data=*/NULL);
1338 }
1339
1340 static struct aggr_cpu_id perf_stat__get_cpu(struct perf_stat_config *config __maybe_unused,
1341                                              struct perf_cpu cpu)
1342 {
1343         return aggr_cpu_id__cpu(cpu, /*data=*/NULL);
1344 }
1345
1346 static struct aggr_cpu_id perf_stat__get_aggr(struct perf_stat_config *config,
1347                                               aggr_get_id_t get_id, struct perf_cpu cpu)
1348 {
1349         struct aggr_cpu_id id;
1350
1351         /* per-process mode - should use global aggr mode */
1352         if (cpu.cpu == -1)
1353                 return get_id(config, cpu);
1354
1355         if (aggr_cpu_id__is_empty(&config->cpus_aggr_map->map[cpu.cpu]))
1356                 config->cpus_aggr_map->map[cpu.cpu] = get_id(config, cpu);
1357
1358         id = config->cpus_aggr_map->map[cpu.cpu];
1359         return id;
1360 }
1361
1362 static struct aggr_cpu_id perf_stat__get_socket_cached(struct perf_stat_config *config,
1363                                                        struct perf_cpu cpu)
1364 {
1365         return perf_stat__get_aggr(config, perf_stat__get_socket, cpu);
1366 }
1367
1368 static struct aggr_cpu_id perf_stat__get_die_cached(struct perf_stat_config *config,
1369                                                     struct perf_cpu cpu)
1370 {
1371         return perf_stat__get_aggr(config, perf_stat__get_die, cpu);
1372 }
1373
1374 static struct aggr_cpu_id perf_stat__get_core_cached(struct perf_stat_config *config,
1375                                                      struct perf_cpu cpu)
1376 {
1377         return perf_stat__get_aggr(config, perf_stat__get_core, cpu);
1378 }
1379
1380 static struct aggr_cpu_id perf_stat__get_node_cached(struct perf_stat_config *config,
1381                                                      struct perf_cpu cpu)
1382 {
1383         return perf_stat__get_aggr(config, perf_stat__get_node, cpu);
1384 }
1385
1386 static struct aggr_cpu_id perf_stat__get_global_cached(struct perf_stat_config *config,
1387                                                        struct perf_cpu cpu)
1388 {
1389         return perf_stat__get_aggr(config, perf_stat__get_global, cpu);
1390 }
1391
1392 static struct aggr_cpu_id perf_stat__get_cpu_cached(struct perf_stat_config *config,
1393                                                     struct perf_cpu cpu)
1394 {
1395         return perf_stat__get_aggr(config, perf_stat__get_cpu, cpu);
1396 }
1397
1398 static aggr_cpu_id_get_t aggr_mode__get_aggr(enum aggr_mode aggr_mode)
1399 {
1400         switch (aggr_mode) {
1401         case AGGR_SOCKET:
1402                 return aggr_cpu_id__socket;
1403         case AGGR_DIE:
1404                 return aggr_cpu_id__die;
1405         case AGGR_CORE:
1406                 return aggr_cpu_id__core;
1407         case AGGR_NODE:
1408                 return aggr_cpu_id__node;
1409         case AGGR_NONE:
1410                 return aggr_cpu_id__cpu;
1411         case AGGR_GLOBAL:
1412                 return aggr_cpu_id__global;
1413         case AGGR_THREAD:
1414         case AGGR_UNSET:
1415         case AGGR_MAX:
1416         default:
1417                 return NULL;
1418         }
1419 }
1420
1421 static aggr_get_id_t aggr_mode__get_id(enum aggr_mode aggr_mode)
1422 {
1423         switch (aggr_mode) {
1424         case AGGR_SOCKET:
1425                 return perf_stat__get_socket_cached;
1426         case AGGR_DIE:
1427                 return perf_stat__get_die_cached;
1428         case AGGR_CORE:
1429                 return perf_stat__get_core_cached;
1430         case AGGR_NODE:
1431                 return perf_stat__get_node_cached;
1432         case AGGR_NONE:
1433                 return perf_stat__get_cpu_cached;
1434         case AGGR_GLOBAL:
1435                 return perf_stat__get_global_cached;
1436         case AGGR_THREAD:
1437         case AGGR_UNSET:
1438         case AGGR_MAX:
1439         default:
1440                 return NULL;
1441         }
1442 }
1443
1444 static int perf_stat_init_aggr_mode(void)
1445 {
1446         int nr;
1447         aggr_cpu_id_get_t get_id = aggr_mode__get_aggr(stat_config.aggr_mode);
1448
1449         if (get_id) {
1450                 bool needs_sort = stat_config.aggr_mode != AGGR_NONE;
1451                 stat_config.aggr_map = cpu_aggr_map__new(evsel_list->core.user_requested_cpus,
1452                                                          get_id, /*data=*/NULL, needs_sort);
1453                 if (!stat_config.aggr_map) {
1454                         pr_err("cannot build %s map", aggr_mode__string[stat_config.aggr_mode]);
1455                         return -1;
1456                 }
1457                 stat_config.aggr_get_id = aggr_mode__get_id(stat_config.aggr_mode);
1458         }
1459
1460         if (stat_config.aggr_mode == AGGR_THREAD) {
1461                 nr = perf_thread_map__nr(evsel_list->core.threads);
1462                 stat_config.aggr_map = cpu_aggr_map__empty_new(nr);
1463                 if (stat_config.aggr_map == NULL)
1464                         return -ENOMEM;
1465
1466                 for (int s = 0; s < nr; s++) {
1467                         struct aggr_cpu_id id = aggr_cpu_id__empty();
1468
1469                         id.thread_idx = s;
1470                         stat_config.aggr_map->map[s] = id;
1471                 }
1472                 return 0;
1473         }
1474
1475         /*
1476          * The evsel_list->cpus is the base we operate on,
1477          * taking the highest cpu number to be the size of
1478          * the aggregation translate cpumap.
1479          */
1480         if (evsel_list->core.user_requested_cpus)
1481                 nr = perf_cpu_map__max(evsel_list->core.user_requested_cpus).cpu;
1482         else
1483                 nr = 0;
1484         stat_config.cpus_aggr_map = cpu_aggr_map__empty_new(nr + 1);
1485         return stat_config.cpus_aggr_map ? 0 : -ENOMEM;
1486 }
1487
1488 static void cpu_aggr_map__delete(struct cpu_aggr_map *map)
1489 {
1490         if (map) {
1491                 WARN_ONCE(refcount_read(&map->refcnt) != 0,
1492                           "cpu_aggr_map refcnt unbalanced\n");
1493                 free(map);
1494         }
1495 }
1496
1497 static void cpu_aggr_map__put(struct cpu_aggr_map *map)
1498 {
1499         if (map && refcount_dec_and_test(&map->refcnt))
1500                 cpu_aggr_map__delete(map);
1501 }
1502
1503 static void perf_stat__exit_aggr_mode(void)
1504 {
1505         cpu_aggr_map__put(stat_config.aggr_map);
1506         cpu_aggr_map__put(stat_config.cpus_aggr_map);
1507         stat_config.aggr_map = NULL;
1508         stat_config.cpus_aggr_map = NULL;
1509 }
1510
1511 static struct aggr_cpu_id perf_env__get_socket_aggr_by_cpu(struct perf_cpu cpu, void *data)
1512 {
1513         struct perf_env *env = data;
1514         struct aggr_cpu_id id = aggr_cpu_id__empty();
1515
1516         if (cpu.cpu != -1)
1517                 id.socket = env->cpu[cpu.cpu].socket_id;
1518
1519         return id;
1520 }
1521
1522 static struct aggr_cpu_id perf_env__get_die_aggr_by_cpu(struct perf_cpu cpu, void *data)
1523 {
1524         struct perf_env *env = data;
1525         struct aggr_cpu_id id = aggr_cpu_id__empty();
1526
1527         if (cpu.cpu != -1) {
1528                 /*
1529                  * die_id is relative to socket, so start
1530                  * with the socket ID and then add die to
1531                  * make a unique ID.
1532                  */
1533                 id.socket = env->cpu[cpu.cpu].socket_id;
1534                 id.die = env->cpu[cpu.cpu].die_id;
1535         }
1536
1537         return id;
1538 }
1539
1540 static struct aggr_cpu_id perf_env__get_core_aggr_by_cpu(struct perf_cpu cpu, void *data)
1541 {
1542         struct perf_env *env = data;
1543         struct aggr_cpu_id id = aggr_cpu_id__empty();
1544
1545         if (cpu.cpu != -1) {
1546                 /*
1547                  * core_id is relative to socket and die,
1548                  * we need a global id. So we set
1549                  * socket, die id and core id
1550                  */
1551                 id.socket = env->cpu[cpu.cpu].socket_id;
1552                 id.die = env->cpu[cpu.cpu].die_id;
1553                 id.core = env->cpu[cpu.cpu].core_id;
1554         }
1555
1556         return id;
1557 }
1558
1559 static struct aggr_cpu_id perf_env__get_cpu_aggr_by_cpu(struct perf_cpu cpu, void *data)
1560 {
1561         struct perf_env *env = data;
1562         struct aggr_cpu_id id = aggr_cpu_id__empty();
1563
1564         if (cpu.cpu != -1) {
1565                 /*
1566                  * core_id is relative to socket and die,
1567                  * we need a global id. So we set
1568                  * socket, die id and core id
1569                  */
1570                 id.socket = env->cpu[cpu.cpu].socket_id;
1571                 id.die = env->cpu[cpu.cpu].die_id;
1572                 id.core = env->cpu[cpu.cpu].core_id;
1573                 id.cpu = cpu;
1574         }
1575
1576         return id;
1577 }
1578
1579 static struct aggr_cpu_id perf_env__get_node_aggr_by_cpu(struct perf_cpu cpu, void *data)
1580 {
1581         struct aggr_cpu_id id = aggr_cpu_id__empty();
1582
1583         id.node = perf_env__numa_node(data, cpu);
1584         return id;
1585 }
1586
1587 static struct aggr_cpu_id perf_env__get_global_aggr_by_cpu(struct perf_cpu cpu __maybe_unused,
1588                                                            void *data __maybe_unused)
1589 {
1590         struct aggr_cpu_id id = aggr_cpu_id__empty();
1591
1592         /* it always aggregates to the cpu 0 */
1593         id.cpu = (struct perf_cpu){ .cpu = 0 };
1594         return id;
1595 }
1596
1597 static struct aggr_cpu_id perf_stat__get_socket_file(struct perf_stat_config *config __maybe_unused,
1598                                                      struct perf_cpu cpu)
1599 {
1600         return perf_env__get_socket_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1601 }
1602 static struct aggr_cpu_id perf_stat__get_die_file(struct perf_stat_config *config __maybe_unused,
1603                                                   struct perf_cpu cpu)
1604 {
1605         return perf_env__get_die_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1606 }
1607
1608 static struct aggr_cpu_id perf_stat__get_core_file(struct perf_stat_config *config __maybe_unused,
1609                                                    struct perf_cpu cpu)
1610 {
1611         return perf_env__get_core_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1612 }
1613
1614 static struct aggr_cpu_id perf_stat__get_cpu_file(struct perf_stat_config *config __maybe_unused,
1615                                                   struct perf_cpu cpu)
1616 {
1617         return perf_env__get_cpu_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1618 }
1619
1620 static struct aggr_cpu_id perf_stat__get_node_file(struct perf_stat_config *config __maybe_unused,
1621                                                    struct perf_cpu cpu)
1622 {
1623         return perf_env__get_node_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1624 }
1625
1626 static struct aggr_cpu_id perf_stat__get_global_file(struct perf_stat_config *config __maybe_unused,
1627                                                      struct perf_cpu cpu)
1628 {
1629         return perf_env__get_global_aggr_by_cpu(cpu, &perf_stat.session->header.env);
1630 }
1631
1632 static aggr_cpu_id_get_t aggr_mode__get_aggr_file(enum aggr_mode aggr_mode)
1633 {
1634         switch (aggr_mode) {
1635         case AGGR_SOCKET:
1636                 return perf_env__get_socket_aggr_by_cpu;
1637         case AGGR_DIE:
1638                 return perf_env__get_die_aggr_by_cpu;
1639         case AGGR_CORE:
1640                 return perf_env__get_core_aggr_by_cpu;
1641         case AGGR_NODE:
1642                 return perf_env__get_node_aggr_by_cpu;
1643         case AGGR_GLOBAL:
1644                 return perf_env__get_global_aggr_by_cpu;
1645         case AGGR_NONE:
1646                 return perf_env__get_cpu_aggr_by_cpu;
1647         case AGGR_THREAD:
1648         case AGGR_UNSET:
1649         case AGGR_MAX:
1650         default:
1651                 return NULL;
1652         }
1653 }
1654
1655 static aggr_get_id_t aggr_mode__get_id_file(enum aggr_mode aggr_mode)
1656 {
1657         switch (aggr_mode) {
1658         case AGGR_SOCKET:
1659                 return perf_stat__get_socket_file;
1660         case AGGR_DIE:
1661                 return perf_stat__get_die_file;
1662         case AGGR_CORE:
1663                 return perf_stat__get_core_file;
1664         case AGGR_NODE:
1665                 return perf_stat__get_node_file;
1666         case AGGR_GLOBAL:
1667                 return perf_stat__get_global_file;
1668         case AGGR_NONE:
1669                 return perf_stat__get_cpu_file;
1670         case AGGR_THREAD:
1671         case AGGR_UNSET:
1672         case AGGR_MAX:
1673         default:
1674                 return NULL;
1675         }
1676 }
1677
1678 static int perf_stat_init_aggr_mode_file(struct perf_stat *st)
1679 {
1680         struct perf_env *env = &st->session->header.env;
1681         aggr_cpu_id_get_t get_id = aggr_mode__get_aggr_file(stat_config.aggr_mode);
1682         bool needs_sort = stat_config.aggr_mode != AGGR_NONE;
1683
1684         if (stat_config.aggr_mode == AGGR_THREAD) {
1685                 int nr = perf_thread_map__nr(evsel_list->core.threads);
1686
1687                 stat_config.aggr_map = cpu_aggr_map__empty_new(nr);
1688                 if (stat_config.aggr_map == NULL)
1689                         return -ENOMEM;
1690
1691                 for (int s = 0; s < nr; s++) {
1692                         struct aggr_cpu_id id = aggr_cpu_id__empty();
1693
1694                         id.thread_idx = s;
1695                         stat_config.aggr_map->map[s] = id;
1696                 }
1697                 return 0;
1698         }
1699
1700         if (!get_id)
1701                 return 0;
1702
1703         stat_config.aggr_map = cpu_aggr_map__new(evsel_list->core.user_requested_cpus,
1704                                                  get_id, env, needs_sort);
1705         if (!stat_config.aggr_map) {
1706                 pr_err("cannot build %s map", aggr_mode__string[stat_config.aggr_mode]);
1707                 return -1;
1708         }
1709         stat_config.aggr_get_id = aggr_mode__get_id_file(stat_config.aggr_mode);
1710         return 0;
1711 }
1712
1713 /*
1714  * Add default attributes, if there were no attributes specified or
1715  * if -d/--detailed, -d -d or -d -d -d is used:
1716  */
1717 static int add_default_attributes(void)
1718 {
1719         int err;
1720         struct perf_event_attr default_attrs0[] = {
1721
1722   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK              },
1723   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES        },
1724   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS          },
1725   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS             },
1726
1727   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES              },
1728 };
1729         struct perf_event_attr frontend_attrs[] = {
1730   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND },
1731 };
1732         struct perf_event_attr backend_attrs[] = {
1733   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND  },
1734 };
1735         struct perf_event_attr default_attrs1[] = {
1736   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS            },
1737   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS     },
1738   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES           },
1739
1740 };
1741
1742 /*
1743  * Detailed stats (-d), covering the L1 and last level data caches:
1744  */
1745         struct perf_event_attr detailed_attrs[] = {
1746
1747   { .type = PERF_TYPE_HW_CACHE,
1748     .config =
1749          PERF_COUNT_HW_CACHE_L1D                <<  0  |
1750         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1751         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
1752
1753   { .type = PERF_TYPE_HW_CACHE,
1754     .config =
1755          PERF_COUNT_HW_CACHE_L1D                <<  0  |
1756         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1757         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
1758
1759   { .type = PERF_TYPE_HW_CACHE,
1760     .config =
1761          PERF_COUNT_HW_CACHE_LL                 <<  0  |
1762         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1763         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
1764
1765   { .type = PERF_TYPE_HW_CACHE,
1766     .config =
1767          PERF_COUNT_HW_CACHE_LL                 <<  0  |
1768         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1769         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
1770 };
1771
1772 /*
1773  * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
1774  */
1775         struct perf_event_attr very_detailed_attrs[] = {
1776
1777   { .type = PERF_TYPE_HW_CACHE,
1778     .config =
1779          PERF_COUNT_HW_CACHE_L1I                <<  0  |
1780         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1781         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
1782
1783   { .type = PERF_TYPE_HW_CACHE,
1784     .config =
1785          PERF_COUNT_HW_CACHE_L1I                <<  0  |
1786         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1787         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
1788
1789   { .type = PERF_TYPE_HW_CACHE,
1790     .config =
1791          PERF_COUNT_HW_CACHE_DTLB               <<  0  |
1792         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1793         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
1794
1795   { .type = PERF_TYPE_HW_CACHE,
1796     .config =
1797          PERF_COUNT_HW_CACHE_DTLB               <<  0  |
1798         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1799         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
1800
1801   { .type = PERF_TYPE_HW_CACHE,
1802     .config =
1803          PERF_COUNT_HW_CACHE_ITLB               <<  0  |
1804         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1805         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
1806
1807   { .type = PERF_TYPE_HW_CACHE,
1808     .config =
1809          PERF_COUNT_HW_CACHE_ITLB               <<  0  |
1810         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1811         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
1812
1813 };
1814
1815 /*
1816  * Very, very detailed stats (-d -d -d), adding prefetch events:
1817  */
1818         struct perf_event_attr very_very_detailed_attrs[] = {
1819
1820   { .type = PERF_TYPE_HW_CACHE,
1821     .config =
1822          PERF_COUNT_HW_CACHE_L1D                <<  0  |
1823         (PERF_COUNT_HW_CACHE_OP_PREFETCH        <<  8) |
1824         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
1825
1826   { .type = PERF_TYPE_HW_CACHE,
1827     .config =
1828          PERF_COUNT_HW_CACHE_L1D                <<  0  |
1829         (PERF_COUNT_HW_CACHE_OP_PREFETCH        <<  8) |
1830         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
1831 };
1832
1833         struct perf_event_attr default_null_attrs[] = {};
1834
1835         /* Set attrs if no event is selected and !null_run: */
1836         if (stat_config.null_run)
1837                 return 0;
1838
1839         if (transaction_run) {
1840                 struct parse_events_error errinfo;
1841                 /* Handle -T as -M transaction. Once platform specific metrics
1842                  * support has been added to the json files, all architectures
1843                  * will use this approach. To determine transaction support
1844                  * on an architecture test for such a metric name.
1845                  */
1846                 if (metricgroup__has_metric("transaction")) {
1847                         return metricgroup__parse_groups(evsel_list, "transaction",
1848                                                          stat_config.metric_no_group,
1849                                                          stat_config.metric_no_merge,
1850                                                          stat_config.user_requested_cpu_list,
1851                                                          stat_config.system_wide,
1852                                                          &stat_config.metric_events);
1853                 }
1854
1855                 parse_events_error__init(&errinfo);
1856                 if (pmu_have_event("cpu", "cycles-ct") &&
1857                     pmu_have_event("cpu", "el-start"))
1858                         err = parse_events(evsel_list, transaction_attrs,
1859                                            &errinfo);
1860                 else
1861                         err = parse_events(evsel_list,
1862                                            transaction_limited_attrs,
1863                                            &errinfo);
1864                 if (err) {
1865                         fprintf(stderr, "Cannot set up transaction events\n");
1866                         parse_events_error__print(&errinfo, transaction_attrs);
1867                 }
1868                 parse_events_error__exit(&errinfo);
1869                 return err ? -1 : 0;
1870         }
1871
1872         if (smi_cost) {
1873                 struct parse_events_error errinfo;
1874                 int smi;
1875
1876                 if (sysfs__read_int(FREEZE_ON_SMI_PATH, &smi) < 0) {
1877                         fprintf(stderr, "freeze_on_smi is not supported.\n");
1878                         return -1;
1879                 }
1880
1881                 if (!smi) {
1882                         if (sysfs__write_int(FREEZE_ON_SMI_PATH, 1) < 0) {
1883                                 fprintf(stderr, "Failed to set freeze_on_smi.\n");
1884                                 return -1;
1885                         }
1886                         smi_reset = true;
1887                 }
1888
1889                 if (!pmu_have_event("msr", "aperf") ||
1890                     !pmu_have_event("msr", "smi")) {
1891                         fprintf(stderr, "To measure SMI cost, it needs "
1892                                 "msr/aperf/, msr/smi/ and cpu/cycles/ support\n");
1893                         return -1;
1894                 }
1895                 if (!force_metric_only)
1896                         stat_config.metric_only = true;
1897
1898                 parse_events_error__init(&errinfo);
1899                 err = parse_events(evsel_list, smi_cost_attrs, &errinfo);
1900                 if (err) {
1901                         parse_events_error__print(&errinfo, smi_cost_attrs);
1902                         fprintf(stderr, "Cannot set up SMI cost events\n");
1903                 }
1904                 parse_events_error__exit(&errinfo);
1905                 return err ? -1 : 0;
1906         }
1907
1908         if (topdown_run) {
1909                 const char **metric_attrs = topdown_metric_attrs;
1910                 unsigned int max_level = 1;
1911                 char *str = NULL;
1912                 bool warn = false;
1913                 const char *pmu_name = arch_get_topdown_pmu_name(evsel_list, true);
1914
1915                 if (!force_metric_only)
1916                         stat_config.metric_only = true;
1917
1918                 if (pmu_have_event(pmu_name, topdown_metric_L2_attrs[5])) {
1919                         metric_attrs = topdown_metric_L2_attrs;
1920                         max_level = 2;
1921                 }
1922
1923                 if (stat_config.topdown_level > max_level) {
1924                         pr_err("Invalid top-down metrics level. The max level is %u.\n", max_level);
1925                         return -1;
1926                 } else if (!stat_config.topdown_level)
1927                         stat_config.topdown_level = max_level;
1928
1929                 if (topdown_filter_events(metric_attrs, &str, 1, pmu_name) < 0) {
1930                         pr_err("Out of memory\n");
1931                         return -1;
1932                 }
1933
1934                 if (metric_attrs[0] && str) {
1935                         if (!stat_config.interval && !stat_config.metric_only) {
1936                                 fprintf(stat_config.output,
1937                                         "Topdown accuracy may decrease when measuring long periods.\n"
1938                                         "Please print the result regularly, e.g. -I1000\n");
1939                         }
1940                         goto setup_metrics;
1941                 }
1942
1943                 zfree(&str);
1944
1945                 if (stat_config.aggr_mode != AGGR_GLOBAL &&
1946                     stat_config.aggr_mode != AGGR_CORE) {
1947                         pr_err("top down event configuration requires --per-core mode\n");
1948                         return -1;
1949                 }
1950                 stat_config.aggr_mode = AGGR_CORE;
1951                 if (nr_cgroups || !target__has_cpu(&target)) {
1952                         pr_err("top down event configuration requires system-wide mode (-a)\n");
1953                         return -1;
1954                 }
1955
1956                 if (topdown_filter_events(topdown_attrs, &str,
1957                                 arch_topdown_check_group(&warn),
1958                                 pmu_name) < 0) {
1959                         pr_err("Out of memory\n");
1960                         return -1;
1961                 }
1962
1963                 if (topdown_attrs[0] && str) {
1964                         struct parse_events_error errinfo;
1965                         if (warn)
1966                                 arch_topdown_group_warn();
1967 setup_metrics:
1968                         parse_events_error__init(&errinfo);
1969                         err = parse_events(evsel_list, str, &errinfo);
1970                         if (err) {
1971                                 fprintf(stderr,
1972                                         "Cannot set up top down events %s: %d\n",
1973                                         str, err);
1974                                 parse_events_error__print(&errinfo, str);
1975                                 parse_events_error__exit(&errinfo);
1976                                 free(str);
1977                                 return -1;
1978                         }
1979                         parse_events_error__exit(&errinfo);
1980                 } else {
1981                         fprintf(stderr, "System does not support topdown\n");
1982                         return -1;
1983                 }
1984                 free(str);
1985         }
1986
1987         if (!stat_config.topdown_level)
1988                 stat_config.topdown_level = TOPDOWN_MAX_LEVEL;
1989
1990         if (!evsel_list->core.nr_entries) {
1991                 if (target__has_cpu(&target))
1992                         default_attrs0[0].config = PERF_COUNT_SW_CPU_CLOCK;
1993
1994                 if (evlist__add_default_attrs(evsel_list, default_attrs0) < 0)
1995                         return -1;
1996                 if (pmu_have_event("cpu", "stalled-cycles-frontend")) {
1997                         if (evlist__add_default_attrs(evsel_list, frontend_attrs) < 0)
1998                                 return -1;
1999                 }
2000                 if (pmu_have_event("cpu", "stalled-cycles-backend")) {
2001                         if (evlist__add_default_attrs(evsel_list, backend_attrs) < 0)
2002                                 return -1;
2003                 }
2004                 if (evlist__add_default_attrs(evsel_list, default_attrs1) < 0)
2005                         return -1;
2006                 /* Platform specific attrs */
2007                 if (evlist__add_default_attrs(evsel_list, default_null_attrs) < 0)
2008                         return -1;
2009         }
2010
2011         /* Detailed events get appended to the event list: */
2012
2013         if (detailed_run <  1)
2014                 return 0;
2015
2016         /* Append detailed run extra attributes: */
2017         if (evlist__add_default_attrs(evsel_list, detailed_attrs) < 0)
2018                 return -1;
2019
2020         if (detailed_run < 2)
2021                 return 0;
2022
2023         /* Append very detailed run extra attributes: */
2024         if (evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0)
2025                 return -1;
2026
2027         if (detailed_run < 3)
2028                 return 0;
2029
2030         /* Append very, very detailed run extra attributes: */
2031         return evlist__add_default_attrs(evsel_list, very_very_detailed_attrs);
2032 }
2033
2034 static const char * const stat_record_usage[] = {
2035         "perf stat record [<options>]",
2036         NULL,
2037 };
2038
2039 static void init_features(struct perf_session *session)
2040 {
2041         int feat;
2042
2043         for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++)
2044                 perf_header__set_feat(&session->header, feat);
2045
2046         perf_header__clear_feat(&session->header, HEADER_DIR_FORMAT);
2047         perf_header__clear_feat(&session->header, HEADER_BUILD_ID);
2048         perf_header__clear_feat(&session->header, HEADER_TRACING_DATA);
2049         perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK);
2050         perf_header__clear_feat(&session->header, HEADER_AUXTRACE);
2051 }
2052
2053 static int __cmd_record(int argc, const char **argv)
2054 {
2055         struct perf_session *session;
2056         struct perf_data *data = &perf_stat.data;
2057
2058         argc = parse_options(argc, argv, stat_options, stat_record_usage,
2059                              PARSE_OPT_STOP_AT_NON_OPTION);
2060
2061         if (output_name)
2062                 data->path = output_name;
2063
2064         if (stat_config.run_count != 1 || forever) {
2065                 pr_err("Cannot use -r option with perf stat record.\n");
2066                 return -1;
2067         }
2068
2069         session = perf_session__new(data, NULL);
2070         if (IS_ERR(session)) {
2071                 pr_err("Perf session creation failed\n");
2072                 return PTR_ERR(session);
2073         }
2074
2075         init_features(session);
2076
2077         session->evlist   = evsel_list;
2078         perf_stat.session = session;
2079         perf_stat.record  = true;
2080         return argc;
2081 }
2082
2083 static int process_stat_round_event(struct perf_session *session,
2084                                     union perf_event *event)
2085 {
2086         struct perf_record_stat_round *stat_round = &event->stat_round;
2087         struct timespec tsh, *ts = NULL;
2088         const char **argv = session->header.env.cmdline_argv;
2089         int argc = session->header.env.nr_cmdline;
2090
2091         process_counters();
2092
2093         if (stat_round->type == PERF_STAT_ROUND_TYPE__FINAL)
2094                 update_stats(&walltime_nsecs_stats, stat_round->time);
2095
2096         if (stat_config.interval && stat_round->time) {
2097                 tsh.tv_sec  = stat_round->time / NSEC_PER_SEC;
2098                 tsh.tv_nsec = stat_round->time % NSEC_PER_SEC;
2099                 ts = &tsh;
2100         }
2101
2102         print_counters(ts, argc, argv);
2103         return 0;
2104 }
2105
2106 static
2107 int process_stat_config_event(struct perf_session *session,
2108                               union perf_event *event)
2109 {
2110         struct perf_tool *tool = session->tool;
2111         struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2112
2113         perf_event__read_stat_config(&stat_config, &event->stat_config);
2114
2115         if (perf_cpu_map__empty(st->cpus)) {
2116                 if (st->aggr_mode != AGGR_UNSET)
2117                         pr_warning("warning: processing task data, aggregation mode not set\n");
2118         } else if (st->aggr_mode != AGGR_UNSET) {
2119                 stat_config.aggr_mode = st->aggr_mode;
2120         }
2121
2122         if (perf_stat.data.is_pipe)
2123                 perf_stat_init_aggr_mode();
2124         else
2125                 perf_stat_init_aggr_mode_file(st);
2126
2127         if (stat_config.aggr_map) {
2128                 int nr_aggr = stat_config.aggr_map->nr;
2129
2130                 if (evlist__alloc_aggr_stats(session->evlist, nr_aggr) < 0) {
2131                         pr_err("cannot allocate aggr counts\n");
2132                         return -1;
2133                 }
2134         }
2135         return 0;
2136 }
2137
2138 static int set_maps(struct perf_stat *st)
2139 {
2140         if (!st->cpus || !st->threads)
2141                 return 0;
2142
2143         if (WARN_ONCE(st->maps_allocated, "stats double allocation\n"))
2144                 return -EINVAL;
2145
2146         perf_evlist__set_maps(&evsel_list->core, st->cpus, st->threads);
2147
2148         if (evlist__alloc_stats(&stat_config, evsel_list, /*alloc_raw=*/true))
2149                 return -ENOMEM;
2150
2151         st->maps_allocated = true;
2152         return 0;
2153 }
2154
2155 static
2156 int process_thread_map_event(struct perf_session *session,
2157                              union perf_event *event)
2158 {
2159         struct perf_tool *tool = session->tool;
2160         struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2161
2162         if (st->threads) {
2163                 pr_warning("Extra thread map event, ignoring.\n");
2164                 return 0;
2165         }
2166
2167         st->threads = thread_map__new_event(&event->thread_map);
2168         if (!st->threads)
2169                 return -ENOMEM;
2170
2171         return set_maps(st);
2172 }
2173
2174 static
2175 int process_cpu_map_event(struct perf_session *session,
2176                           union perf_event *event)
2177 {
2178         struct perf_tool *tool = session->tool;
2179         struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2180         struct perf_cpu_map *cpus;
2181
2182         if (st->cpus) {
2183                 pr_warning("Extra cpu map event, ignoring.\n");
2184                 return 0;
2185         }
2186
2187         cpus = cpu_map__new_data(&event->cpu_map.data);
2188         if (!cpus)
2189                 return -ENOMEM;
2190
2191         st->cpus = cpus;
2192         return set_maps(st);
2193 }
2194
2195 static const char * const stat_report_usage[] = {
2196         "perf stat report [<options>]",
2197         NULL,
2198 };
2199
2200 static struct perf_stat perf_stat = {
2201         .tool = {
2202                 .attr           = perf_event__process_attr,
2203                 .event_update   = perf_event__process_event_update,
2204                 .thread_map     = process_thread_map_event,
2205                 .cpu_map        = process_cpu_map_event,
2206                 .stat_config    = process_stat_config_event,
2207                 .stat           = perf_event__process_stat_event,
2208                 .stat_round     = process_stat_round_event,
2209         },
2210         .aggr_mode = AGGR_UNSET,
2211 };
2212
2213 static int __cmd_report(int argc, const char **argv)
2214 {
2215         struct perf_session *session;
2216         const struct option options[] = {
2217         OPT_STRING('i', "input", &input_name, "file", "input file name"),
2218         OPT_SET_UINT(0, "per-socket", &perf_stat.aggr_mode,
2219                      "aggregate counts per processor socket", AGGR_SOCKET),
2220         OPT_SET_UINT(0, "per-die", &perf_stat.aggr_mode,
2221                      "aggregate counts per processor die", AGGR_DIE),
2222         OPT_SET_UINT(0, "per-core", &perf_stat.aggr_mode,
2223                      "aggregate counts per physical processor core", AGGR_CORE),
2224         OPT_SET_UINT(0, "per-node", &perf_stat.aggr_mode,
2225                      "aggregate counts per numa node", AGGR_NODE),
2226         OPT_SET_UINT('A', "no-aggr", &perf_stat.aggr_mode,
2227                      "disable CPU count aggregation", AGGR_NONE),
2228         OPT_END()
2229         };
2230         struct stat st;
2231         int ret;
2232
2233         argc = parse_options(argc, argv, options, stat_report_usage, 0);
2234
2235         if (!input_name || !strlen(input_name)) {
2236                 if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode))
2237                         input_name = "-";
2238                 else
2239                         input_name = "perf.data";
2240         }
2241
2242         perf_stat__init_shadow_stats();
2243
2244         perf_stat.data.path = input_name;
2245         perf_stat.data.mode = PERF_DATA_MODE_READ;
2246
2247         session = perf_session__new(&perf_stat.data, &perf_stat.tool);
2248         if (IS_ERR(session))
2249                 return PTR_ERR(session);
2250
2251         perf_stat.session  = session;
2252         stat_config.output = stderr;
2253         evsel_list         = session->evlist;
2254
2255         ret = perf_session__process_events(session);
2256         if (ret)
2257                 return ret;
2258
2259         perf_session__delete(session);
2260         return 0;
2261 }
2262
2263 static void setup_system_wide(int forks)
2264 {
2265         /*
2266          * Make system wide (-a) the default target if
2267          * no target was specified and one of following
2268          * conditions is met:
2269          *
2270          *   - there's no workload specified
2271          *   - there is workload specified but all requested
2272          *     events are system wide events
2273          */
2274         if (!target__none(&target))
2275                 return;
2276
2277         if (!forks)
2278                 target.system_wide = true;
2279         else {
2280                 struct evsel *counter;
2281
2282                 evlist__for_each_entry(evsel_list, counter) {
2283                         if (!counter->core.requires_cpu &&
2284                             strcmp(counter->name, "duration_time")) {
2285                                 return;
2286                         }
2287                 }
2288
2289                 if (evsel_list->core.nr_entries)
2290                         target.system_wide = true;
2291         }
2292 }
2293
2294 int cmd_stat(int argc, const char **argv)
2295 {
2296         const char * const stat_usage[] = {
2297                 "perf stat [<options>] [<command>]",
2298                 NULL
2299         };
2300         int status = -EINVAL, run_idx, err;
2301         const char *mode;
2302         FILE *output = stderr;
2303         unsigned int interval, timeout;
2304         const char * const stat_subcommands[] = { "record", "report" };
2305         char errbuf[BUFSIZ];
2306
2307         setlocale(LC_ALL, "");
2308
2309         evsel_list = evlist__new();
2310         if (evsel_list == NULL)
2311                 return -ENOMEM;
2312
2313         parse_events__shrink_config_terms();
2314
2315         /* String-parsing callback-based options would segfault when negated */
2316         set_option_flag(stat_options, 'e', "event", PARSE_OPT_NONEG);
2317         set_option_flag(stat_options, 'M', "metrics", PARSE_OPT_NONEG);
2318         set_option_flag(stat_options, 'G', "cgroup", PARSE_OPT_NONEG);
2319
2320         argc = parse_options_subcommand(argc, argv, stat_options, stat_subcommands,
2321                                         (const char **) stat_usage,
2322                                         PARSE_OPT_STOP_AT_NON_OPTION);
2323
2324         if (stat_config.csv_sep) {
2325                 stat_config.csv_output = true;
2326                 if (!strcmp(stat_config.csv_sep, "\\t"))
2327                         stat_config.csv_sep = "\t";
2328         } else
2329                 stat_config.csv_sep = DEFAULT_SEPARATOR;
2330
2331         if (argc && strlen(argv[0]) > 2 && strstarts("record", argv[0])) {
2332                 argc = __cmd_record(argc, argv);
2333                 if (argc < 0)
2334                         return -1;
2335         } else if (argc && strlen(argv[0]) > 2 && strstarts("report", argv[0]))
2336                 return __cmd_report(argc, argv);
2337
2338         interval = stat_config.interval;
2339         timeout = stat_config.timeout;
2340
2341         /*
2342          * For record command the -o is already taken care of.
2343          */
2344         if (!STAT_RECORD && output_name && strcmp(output_name, "-"))
2345                 output = NULL;
2346
2347         if (output_name && output_fd) {
2348                 fprintf(stderr, "cannot use both --output and --log-fd\n");
2349                 parse_options_usage(stat_usage, stat_options, "o", 1);
2350                 parse_options_usage(NULL, stat_options, "log-fd", 0);
2351                 goto out;
2352         }
2353
2354         if (stat_config.metric_only && stat_config.aggr_mode == AGGR_THREAD) {
2355                 fprintf(stderr, "--metric-only is not supported with --per-thread\n");
2356                 goto out;
2357         }
2358
2359         if (stat_config.metric_only && stat_config.run_count > 1) {
2360                 fprintf(stderr, "--metric-only is not supported with -r\n");
2361                 goto out;
2362         }
2363
2364         if (stat_config.walltime_run_table && stat_config.run_count <= 1) {
2365                 fprintf(stderr, "--table is only supported with -r\n");
2366                 parse_options_usage(stat_usage, stat_options, "r", 1);
2367                 parse_options_usage(NULL, stat_options, "table", 0);
2368                 goto out;
2369         }
2370
2371         if (output_fd < 0) {
2372                 fprintf(stderr, "argument to --log-fd must be a > 0\n");
2373                 parse_options_usage(stat_usage, stat_options, "log-fd", 0);
2374                 goto out;
2375         }
2376
2377         if (!output && !quiet) {
2378                 struct timespec tm;
2379                 mode = append_file ? "a" : "w";
2380
2381                 output = fopen(output_name, mode);
2382                 if (!output) {
2383                         perror("failed to create output file");
2384                         return -1;
2385                 }
2386                 clock_gettime(CLOCK_REALTIME, &tm);
2387                 fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
2388         } else if (output_fd > 0) {
2389                 mode = append_file ? "a" : "w";
2390                 output = fdopen(output_fd, mode);
2391                 if (!output) {
2392                         perror("Failed opening logfd");
2393                         return -errno;
2394                 }
2395         }
2396
2397         if (stat_config.interval_clear && !isatty(fileno(output))) {
2398                 fprintf(stderr, "--interval-clear does not work with output\n");
2399                 parse_options_usage(stat_usage, stat_options, "o", 1);
2400                 parse_options_usage(NULL, stat_options, "log-fd", 0);
2401                 parse_options_usage(NULL, stat_options, "interval-clear", 0);
2402                 return -1;
2403         }
2404
2405         stat_config.output = output;
2406
2407         /*
2408          * let the spreadsheet do the pretty-printing
2409          */
2410         if (stat_config.csv_output) {
2411                 /* User explicitly passed -B? */
2412                 if (big_num_opt == 1) {
2413                         fprintf(stderr, "-B option not supported with -x\n");
2414                         parse_options_usage(stat_usage, stat_options, "B", 1);
2415                         parse_options_usage(NULL, stat_options, "x", 1);
2416                         goto out;
2417                 } else /* Nope, so disable big number formatting */
2418                         stat_config.big_num = false;
2419         } else if (big_num_opt == 0) /* User passed --no-big-num */
2420                 stat_config.big_num = false;
2421
2422         err = target__validate(&target);
2423         if (err) {
2424                 target__strerror(&target, err, errbuf, BUFSIZ);
2425                 pr_warning("%s\n", errbuf);
2426         }
2427
2428         setup_system_wide(argc);
2429
2430         /*
2431          * Display user/system times only for single
2432          * run and when there's specified tracee.
2433          */
2434         if ((stat_config.run_count == 1) && target__none(&target))
2435                 stat_config.ru_display = true;
2436
2437         if (stat_config.run_count < 0) {
2438                 pr_err("Run count must be a positive number\n");
2439                 parse_options_usage(stat_usage, stat_options, "r", 1);
2440                 goto out;
2441         } else if (stat_config.run_count == 0) {
2442                 forever = true;
2443                 stat_config.run_count = 1;
2444         }
2445
2446         if (stat_config.walltime_run_table) {
2447                 stat_config.walltime_run = zalloc(stat_config.run_count * sizeof(stat_config.walltime_run[0]));
2448                 if (!stat_config.walltime_run) {
2449                         pr_err("failed to setup -r option");
2450                         goto out;
2451                 }
2452         }
2453
2454         if ((stat_config.aggr_mode == AGGR_THREAD) &&
2455                 !target__has_task(&target)) {
2456                 if (!target.system_wide || target.cpu_list) {
2457                         fprintf(stderr, "The --per-thread option is only "
2458                                 "available when monitoring via -p -t -a "
2459                                 "options or only --per-thread.\n");
2460                         parse_options_usage(NULL, stat_options, "p", 1);
2461                         parse_options_usage(NULL, stat_options, "t", 1);
2462                         goto out;
2463                 }
2464         }
2465
2466         /*
2467          * no_aggr, cgroup are for system-wide only
2468          * --per-thread is aggregated per thread, we dont mix it with cpu mode
2469          */
2470         if (((stat_config.aggr_mode != AGGR_GLOBAL &&
2471               stat_config.aggr_mode != AGGR_THREAD) ||
2472              (nr_cgroups || stat_config.cgroup_list)) &&
2473             !target__has_cpu(&target)) {
2474                 fprintf(stderr, "both cgroup and no-aggregation "
2475                         "modes only available in system-wide mode\n");
2476
2477                 parse_options_usage(stat_usage, stat_options, "G", 1);
2478                 parse_options_usage(NULL, stat_options, "A", 1);
2479                 parse_options_usage(NULL, stat_options, "a", 1);
2480                 parse_options_usage(NULL, stat_options, "for-each-cgroup", 0);
2481                 goto out;
2482         }
2483
2484         if (stat_config.iostat_run) {
2485                 status = iostat_prepare(evsel_list, &stat_config);
2486                 if (status)
2487                         goto out;
2488                 if (iostat_mode == IOSTAT_LIST) {
2489                         iostat_list(evsel_list, &stat_config);
2490                         goto out;
2491                 } else if (verbose > 0)
2492                         iostat_list(evsel_list, &stat_config);
2493                 if (iostat_mode == IOSTAT_RUN && !target__has_cpu(&target))
2494                         target.system_wide = true;
2495         }
2496
2497         if ((stat_config.aggr_mode == AGGR_THREAD) && (target.system_wide))
2498                 target.per_thread = true;
2499
2500         stat_config.system_wide = target.system_wide;
2501         if (target.cpu_list) {
2502                 stat_config.user_requested_cpu_list = strdup(target.cpu_list);
2503                 if (!stat_config.user_requested_cpu_list) {
2504                         status = -ENOMEM;
2505                         goto out;
2506                 }
2507         }
2508
2509         /*
2510          * Metric parsing needs to be delayed as metrics may optimize events
2511          * knowing the target is system-wide.
2512          */
2513         if (metrics) {
2514                 metricgroup__parse_groups(evsel_list, metrics,
2515                                         stat_config.metric_no_group,
2516                                         stat_config.metric_no_merge,
2517                                         stat_config.user_requested_cpu_list,
2518                                         stat_config.system_wide,
2519                                         &stat_config.metric_events);
2520                 zfree(&metrics);
2521         }
2522         perf_stat__init_shadow_stats();
2523
2524         if (add_default_attributes())
2525                 goto out;
2526
2527         if (stat_config.cgroup_list) {
2528                 if (nr_cgroups > 0) {
2529                         pr_err("--cgroup and --for-each-cgroup cannot be used together\n");
2530                         parse_options_usage(stat_usage, stat_options, "G", 1);
2531                         parse_options_usage(NULL, stat_options, "for-each-cgroup", 0);
2532                         goto out;
2533                 }
2534
2535                 if (evlist__expand_cgroup(evsel_list, stat_config.cgroup_list,
2536                                           &stat_config.metric_events, true) < 0) {
2537                         parse_options_usage(stat_usage, stat_options,
2538                                             "for-each-cgroup", 0);
2539                         goto out;
2540                 }
2541         }
2542
2543         if (evlist__fix_hybrid_cpus(evsel_list, target.cpu_list)) {
2544                 pr_err("failed to use cpu list %s\n", target.cpu_list);
2545                 goto out;
2546         }
2547
2548         target.hybrid = perf_pmu__has_hybrid();
2549         if (evlist__create_maps(evsel_list, &target) < 0) {
2550                 if (target__has_task(&target)) {
2551                         pr_err("Problems finding threads of monitor\n");
2552                         parse_options_usage(stat_usage, stat_options, "p", 1);
2553                         parse_options_usage(NULL, stat_options, "t", 1);
2554                 } else if (target__has_cpu(&target)) {
2555                         perror("failed to parse CPUs map");
2556                         parse_options_usage(stat_usage, stat_options, "C", 1);
2557                         parse_options_usage(NULL, stat_options, "a", 1);
2558                 }
2559                 goto out;
2560         }
2561
2562         evlist__check_cpu_maps(evsel_list);
2563
2564         /*
2565          * Initialize thread_map with comm names,
2566          * so we could print it out on output.
2567          */
2568         if (stat_config.aggr_mode == AGGR_THREAD) {
2569                 thread_map__read_comms(evsel_list->core.threads);
2570         }
2571
2572         if (stat_config.aggr_mode == AGGR_NODE)
2573                 cpu__setup_cpunode_map();
2574
2575         if (stat_config.times && interval)
2576                 interval_count = true;
2577         else if (stat_config.times && !interval) {
2578                 pr_err("interval-count option should be used together with "
2579                                 "interval-print.\n");
2580                 parse_options_usage(stat_usage, stat_options, "interval-count", 0);
2581                 parse_options_usage(stat_usage, stat_options, "I", 1);
2582                 goto out;
2583         }
2584
2585         if (timeout && timeout < 100) {
2586                 if (timeout < 10) {
2587                         pr_err("timeout must be >= 10ms.\n");
2588                         parse_options_usage(stat_usage, stat_options, "timeout", 0);
2589                         goto out;
2590                 } else
2591                         pr_warning("timeout < 100ms. "
2592                                    "The overhead percentage could be high in some cases. "
2593                                    "Please proceed with caution.\n");
2594         }
2595         if (timeout && interval) {
2596                 pr_err("timeout option is not supported with interval-print.\n");
2597                 parse_options_usage(stat_usage, stat_options, "timeout", 0);
2598                 parse_options_usage(stat_usage, stat_options, "I", 1);
2599                 goto out;
2600         }
2601
2602         if (perf_stat_init_aggr_mode())
2603                 goto out;
2604
2605         if (evlist__alloc_stats(&stat_config, evsel_list, interval))
2606                 goto out;
2607
2608         /*
2609          * Set sample_type to PERF_SAMPLE_IDENTIFIER, which should be harmless
2610          * while avoiding that older tools show confusing messages.
2611          *
2612          * However for pipe sessions we need to keep it zero,
2613          * because script's perf_evsel__check_attr is triggered
2614          * by attr->sample_type != 0, and we can't run it on
2615          * stat sessions.
2616          */
2617         stat_config.identifier = !(STAT_RECORD && perf_stat.data.is_pipe);
2618
2619         /*
2620          * We dont want to block the signals - that would cause
2621          * child tasks to inherit that and Ctrl-C would not work.
2622          * What we want is for Ctrl-C to work in the exec()-ed
2623          * task, but being ignored by perf stat itself:
2624          */
2625         atexit(sig_atexit);
2626         if (!forever)
2627                 signal(SIGINT,  skip_signal);
2628         signal(SIGCHLD, skip_signal);
2629         signal(SIGALRM, skip_signal);
2630         signal(SIGABRT, skip_signal);
2631
2632         if (evlist__initialize_ctlfd(evsel_list, stat_config.ctl_fd, stat_config.ctl_fd_ack))
2633                 goto out;
2634
2635         /* Enable ignoring missing threads when -p option is defined. */
2636         evlist__first(evsel_list)->ignore_missing_thread = target.pid;
2637         status = 0;
2638         for (run_idx = 0; forever || run_idx < stat_config.run_count; run_idx++) {
2639                 if (stat_config.run_count != 1 && verbose > 0)
2640                         fprintf(output, "[ perf stat: executing run #%d ... ]\n",
2641                                 run_idx + 1);
2642
2643                 if (run_idx != 0)
2644                         evlist__reset_prev_raw_counts(evsel_list);
2645
2646                 status = run_perf_stat(argc, argv, run_idx);
2647                 if (forever && status != -1 && !interval) {
2648                         print_counters(NULL, argc, argv);
2649                         perf_stat__reset_stats();
2650                 }
2651         }
2652
2653         if (!forever && status != -1 && (!interval || stat_config.summary))
2654                 print_counters(NULL, argc, argv);
2655
2656         evlist__finalize_ctlfd(evsel_list);
2657
2658         if (STAT_RECORD) {
2659                 /*
2660                  * We synthesize the kernel mmap record just so that older tools
2661                  * don't emit warnings about not being able to resolve symbols
2662                  * due to /proc/sys/kernel/kptr_restrict settings and instead provide
2663                  * a saner message about no samples being in the perf.data file.
2664                  *
2665                  * This also serves to suppress a warning about f_header.data.size == 0
2666                  * in header.c at the moment 'perf stat record' gets introduced, which
2667                  * is not really needed once we start adding the stat specific PERF_RECORD_
2668                  * records, but the need to suppress the kptr_restrict messages in older
2669                  * tools remain  -acme
2670                  */
2671                 int fd = perf_data__fd(&perf_stat.data);
2672
2673                 err = perf_event__synthesize_kernel_mmap((void *)&perf_stat,
2674                                                          process_synthesized_event,
2675                                                          &perf_stat.session->machines.host);
2676                 if (err) {
2677                         pr_warning("Couldn't synthesize the kernel mmap record, harmless, "
2678                                    "older tools may produce warnings about this file\n.");
2679                 }
2680
2681                 if (!interval) {
2682                         if (WRITE_STAT_ROUND_EVENT(walltime_nsecs_stats.max, FINAL))
2683                                 pr_err("failed to write stat round event\n");
2684                 }
2685
2686                 if (!perf_stat.data.is_pipe) {
2687                         perf_stat.session->header.data_size += perf_stat.bytes_written;
2688                         perf_session__write_header(perf_stat.session, evsel_list, fd, true);
2689                 }
2690
2691                 evlist__close(evsel_list);
2692                 perf_session__delete(perf_stat.session);
2693         }
2694
2695         perf_stat__exit_aggr_mode();
2696         evlist__free_stats(evsel_list);
2697 out:
2698         if (stat_config.iostat_run)
2699                 iostat_release(evsel_list);
2700
2701         zfree(&stat_config.walltime_run);
2702         zfree(&stat_config.user_requested_cpu_list);
2703
2704         if (smi_cost && smi_reset)
2705                 sysfs__write_int(FREEZE_ON_SMI_PATH, 0);
2706
2707         evlist__delete(evsel_list);
2708
2709         metricgroup__rblist_exit(&stat_config.metric_events);
2710         evlist__close_control(stat_config.ctl_fd, stat_config.ctl_fd_ack, &stat_config.ctl_fd_close);
2711
2712         return status;
2713 }