tools/perf/builtin-stat.c

   1 /*
   2  * builtin-stat.c
   3  *
   4  * Builtin stat command: Give a precise performance counters summary
   5  * overview about any workload, CPU or specific PID.
   6  *
   7  * Sample output:
   8
   9    $ perf stat ./hackbench 10
  10
  11   Time: 0.118
  12
  13   Performance counter stats for './hackbench 10':
  14
  15        1708.761321 task-clock                #   11.037 CPUs utilized
  16             41,190 context-switches          #    0.024 M/sec
  17              6,735 CPU-migrations            #    0.004 M/sec
  18             17,318 page-faults               #    0.010 M/sec
  19      5,205,202,243 cycles                    #    3.046 GHz
  20      3,856,436,920 stalled-cycles-frontend   #   74.09% frontend cycles idle
  21      1,600,790,871 stalled-cycles-backend    #   30.75% backend  cycles idle
  22      2,603,501,247 instructions              #    0.50  insns per cycle
  23                                              #    1.48  stalled cycles per insn
  24        484,357,498 branches                  #  283.455 M/sec
  25          6,388,934 branch-misses             #    1.32% of all branches
  26
  27         0.154822978  seconds time elapsed
  28
  29  *
  30  * Copyright (C) 2008-2011, Red Hat Inc, Ingo Molnar <mingo@redhat.com>
  31  *
  32  * Improvements and fixes by:
  33  *
  34  *   Arjan van de Ven <arjan@linux.intel.com>
  35  *   Yanmin Zhang <yanmin.zhang@intel.com>
  36  *   Wu Fengguang <fengguang.wu@intel.com>
  37  *   Mike Galbraith <efault@gmx.de>
  38  *   Paul Mackerras <paulus@samba.org>
  39  *   Jaswinder Singh Rajput <jaswinder@kernel.org>
  40  *
  41  * Released under the GPL v2. (and only v2, not any later version)
  42  */
  43
  44 #include "perf.h"
  45 #include "builtin.h"
  46 #include "util/cgroup.h"
  47 #include "util/util.h"
  48 #include <subcmd/parse-options.h>
  49 #include "util/parse-events.h"
  50 #include "util/pmu.h"
  51 #include "util/event.h"
  52 #include "util/evlist.h"
  53 #include "util/evsel.h"
  54 #include "util/debug.h"
  55 #include "util/drv_configs.h"
  56 #include "util/color.h"
  57 #include "util/stat.h"
  58 #include "util/header.h"
  59 #include "util/cpumap.h"
  60 #include "util/thread.h"
  61 #include "util/thread_map.h"
  62 #include "util/counts.h"
  63 #include "util/group.h"
  64 #include "util/session.h"
  65 #include "util/tool.h"
  66 #include "util/group.h"
  67 #include "util/string2.h"
  68 #include "util/metricgroup.h"
  69 #include "asm/bug.h"
  70
  71 #include <linux/time64.h>
  72 #include <api/fs/fs.h>
  73 #include <errno.h>
  74 #include <signal.h>
  75 #include <stdlib.h>
  76 #include <sys/prctl.h>
  77 #include <inttypes.h>
  78 #include <locale.h>
  79 #include <math.h>
  80 #include <sys/types.h>
  81 #include <sys/stat.h>
  82 #include <sys/wait.h>
  83 #include <unistd.h>
  84
  85 #include "sane_ctype.h"
  86
  87 #define DEFAULT_SEPARATOR       " "
  88 #define CNTR_NOT_SUPPORTED      "<not supported>"
  89 #define CNTR_NOT_COUNTED        "<not counted>"
  90 #define FREEZE_ON_SMI_PATH      "devices/cpu/freeze_on_smi"
  91
  92 static void print_counters(struct timespec *ts, int argc, const char **argv);
  93
  94 /* Default events used for perf stat -T */
  95 static const char *transaction_attrs = {
  96         "task-clock,"
  97         "{"
  98         "instructions,"
  99         "cycles,"
 100         "cpu/cycles-t/,"
 101         "cpu/tx-start/,"
 102         "cpu/el-start/,"
 103         "cpu/cycles-ct/"
 104         "}"
 105 };
 106
 107 /* More limited version when the CPU does not have all events. */
 108 static const char * transaction_limited_attrs = {
 109         "task-clock,"
 110         "{"
 111         "instructions,"
 112         "cycles,"
 113         "cpu/cycles-t/,"
 114         "cpu/tx-start/"
 115         "}"
 116 };
 117
 118 static const char * topdown_attrs[] = {
 119         "topdown-total-slots",
 120         "topdown-slots-retired",
 121         "topdown-recovery-bubbles",
 122         "topdown-fetch-bubbles",
 123         "topdown-slots-issued",
 124         NULL,
 125 };
 126
 127 static const char *smi_cost_attrs = {
 128         "{"
 129         "msr/aperf/,"
 130         "msr/smi/,"
 131         "cycles"
 132         "}"
 133 };
 134
 135 static struct perf_evlist       *evsel_list;
 136
 137 static struct rblist             metric_events;
 138
 139 static struct target target = {
 140         .uid    = UINT_MAX,
 141 };
 142
 143 typedef int (*aggr_get_id_t)(struct cpu_map *m, int cpu);
 144
 145 static int                      run_count                       =  1;
 146 static bool                     no_inherit                      = false;
 147 static volatile pid_t           child_pid                       = -1;
 148 static bool                     null_run                        =  false;
 149 static int                      detailed_run                    =  0;
 150 static bool                     transaction_run;
 151 static bool                     topdown_run                     = false;
 152 static bool                     smi_cost                        = false;
 153 static bool                     smi_reset                       = false;
 154 static bool                     big_num                         =  true;
 155 static int                      big_num_opt                     =  -1;
 156 static const char               *csv_sep                        = NULL;
 157 static bool                     csv_output                      = false;
 158 static bool                     group                           = false;
 159 static const char               *pre_cmd                        = NULL;
 160 static const char               *post_cmd                       = NULL;
 161 static bool                     sync_run                        = false;
 162 static unsigned int             initial_delay                   = 0;
 163 static unsigned int             unit_width                      = 4; /* strlen("unit") */
 164 static bool                     forever                         = false;
 165 static bool                     metric_only                     = false;
 166 static bool                     force_metric_only               = false;
 167 static bool                     no_merge                        = false;
 168 static struct timespec          ref_time;
 169 static struct cpu_map           *aggr_map;
 170 static aggr_get_id_t            aggr_get_id;
 171 static bool                     append_file;
 172 static const char               *output_name;
 173 static int                      output_fd;
 174 static int                      print_free_counters_hint;
 175
 176 struct perf_stat {
 177         bool                     record;
 178         struct perf_data         data;
 179         struct perf_session     *session;
 180         u64                      bytes_written;
 181         struct perf_tool         tool;
 182         bool                     maps_allocated;
 183         struct cpu_map          *cpus;
 184         struct thread_map       *threads;
 185         enum aggr_mode           aggr_mode;
 186 };
 187
 188 static struct perf_stat         perf_stat;
 189 #define STAT_RECORD             perf_stat.record
 190
 191 static volatile int done = 0;
 192
 193 static struct perf_stat_config stat_config = {
 194         .aggr_mode      = AGGR_GLOBAL,
 195         .scale          = true,
 196 };
 197
 198 static bool is_duration_time(struct perf_evsel *evsel)
 199 {
 200         return !strcmp(evsel->name, "duration_time");
 201 }
 202
 203 static inline void diff_timespec(struct timespec *r, struct timespec *a,
 204                                  struct timespec *b)
 205 {
 206         r->tv_sec = a->tv_sec - b->tv_sec;
 207         if (a->tv_nsec < b->tv_nsec) {
 208                 r->tv_nsec = a->tv_nsec + NSEC_PER_SEC - b->tv_nsec;
 209                 r->tv_sec--;
 210         } else {
 211                 r->tv_nsec = a->tv_nsec - b->tv_nsec ;
 212         }
 213 }
 214
 215 static void perf_stat__reset_stats(void)
 216 {
 217         perf_evlist__reset_stats(evsel_list);
 218         perf_stat__reset_shadow_stats();
 219 }
 220
 221 static int create_perf_stat_counter(struct perf_evsel *evsel)
 222 {
 223         struct perf_event_attr *attr = &evsel->attr;
 224         struct perf_evsel *leader = evsel->leader;
 225
 226         if (stat_config.scale) {
 227                 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
 228                                     PERF_FORMAT_TOTAL_TIME_RUNNING;
 229         }
 230
 231         /*
 232          * The event is part of non trivial group, let's enable
 233          * the group read (for leader) and ID retrieval for all
 234          * members.
 235          */
 236         if (leader->nr_members > 1)
 237                 attr->read_format |= PERF_FORMAT_ID|PERF_FORMAT_GROUP;
 238
 239         attr->inherit = !no_inherit;
 240
 241         /*
 242          * Some events get initialized with sample_(period/type) set,
 243          * like tracepoints. Clear it up for counting.
 244          */
 245         attr->sample_period = 0;
 246
 247         /*
 248          * But set sample_type to PERF_SAMPLE_IDENTIFIER, which should be harmless
 249          * while avoiding that older tools show confusing messages.
 250          *
 251          * However for pipe sessions we need to keep it zero,
 252          * because script's perf_evsel__check_attr is triggered
 253          * by attr->sample_type != 0, and we can't run it on
 254          * stat sessions.
 255          */
 256         if (!(STAT_RECORD && perf_stat.data.is_pipe))
 257                 attr->sample_type = PERF_SAMPLE_IDENTIFIER;
 258
 259         /*
 260          * Disabling all counters initially, they will be enabled
 261          * either manually by us or by kernel via enable_on_exec
 262          * set later.
 263          */
 264         if (perf_evsel__is_group_leader(evsel)) {
 265                 attr->disabled = 1;
 266
 267                 /*
 268                  * In case of initial_delay we enable tracee
 269                  * events manually.
 270                  */
 271                 if (target__none(&target) && !initial_delay)
 272                         attr->enable_on_exec = 1;
 273         }
 274
 275         if (target__has_cpu(&target))
 276                 return perf_evsel__open_per_cpu(evsel, perf_evsel__cpus(evsel));
 277
 278         return perf_evsel__open_per_thread(evsel, evsel_list->threads);
 279 }
 280
 281 /*
 282  * Does the counter have nsecs as a unit?
 283  */
 284 static inline int nsec_counter(struct perf_evsel *evsel)
 285 {
 286         if (perf_evsel__match(evsel, SOFTWARE, SW_CPU_CLOCK) ||
 287             perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
 288                 return 1;
 289
 290         return 0;
 291 }
 292
 293 static int process_synthesized_event(struct perf_tool *tool __maybe_unused,
 294                                      union perf_event *event,
 295                                      struct perf_sample *sample __maybe_unused,
 296                                      struct machine *machine __maybe_unused)
 297 {
 298         if (perf_data__write(&perf_stat.data, event, event->header.size) < 0) {
 299                 pr_err("failed to write perf data, error: %m\n");
 300                 return -1;
 301         }
 302
 303         perf_stat.bytes_written += event->header.size;
 304         return 0;
 305 }
 306
 307 static int write_stat_round_event(u64 tm, u64 type)
 308 {
 309         return perf_event__synthesize_stat_round(NULL, tm, type,
 310                                                  process_synthesized_event,
 311                                                  NULL);
 312 }
 313
 314 #define WRITE_STAT_ROUND_EVENT(time, interval) \
 315         write_stat_round_event(time, PERF_STAT_ROUND_TYPE__ ## interval)
 316
 317 #define SID(e, x, y) xyarray__entry(e->sample_id, x, y)
 318
 319 static int
 320 perf_evsel__write_stat_event(struct perf_evsel *counter, u32 cpu, u32 thread,
 321                              struct perf_counts_values *count)
 322 {
 323         struct perf_sample_id *sid = SID(counter, cpu, thread);
 324
 325         return perf_event__synthesize_stat(NULL, cpu, thread, sid->id, count,
 326                                            process_synthesized_event, NULL);
 327 }
 328
 329 /*
 330  * Read out the results of a single counter:
 331  * do not aggregate counts across CPUs in system-wide mode
 332  */
 333 static int read_counter(struct perf_evsel *counter)
 334 {
 335         int nthreads = thread_map__nr(evsel_list->threads);
 336         int ncpus, cpu, thread;
 337
 338         if (target__has_cpu(&target))
 339                 ncpus = perf_evsel__nr_cpus(counter);
 340         else
 341                 ncpus = 1;
 342
 343         if (!counter->supported)
 344                 return -ENOENT;
 345
 346         if (counter->system_wide)
 347                 nthreads = 1;
 348
 349         for (thread = 0; thread < nthreads; thread++) {
 350                 for (cpu = 0; cpu < ncpus; cpu++) {
 351                         struct perf_counts_values *count;
 352
 353                         count = perf_counts(counter->counts, cpu, thread);
 354
 355                         /*
 356                          * The leader's group read loads data into its group members
 357                          * (via perf_evsel__read_counter) and sets threir count->loaded.
 358                          */
 359                         if (!count->loaded &&
 360                             perf_evsel__read_counter(counter, cpu, thread)) {
 361                                 counter->counts->scaled = -1;
 362                                 perf_counts(counter->counts, cpu, thread)->ena = 0;
 363                                 perf_counts(counter->counts, cpu, thread)->run = 0;
 364                                 return -1;
 365                         }
 366
 367                         count->loaded = false;
 368
 369                         if (STAT_RECORD) {
 370                                 if (perf_evsel__write_stat_event(counter, cpu, thread, count)) {
 371                                         pr_err("failed to write stat event\n");
 372                                         return -1;
 373                                 }
 374                         }
 375
 376                         if (verbose > 1) {
 377                                 fprintf(stat_config.output,
 378                                         "%s: %d: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
 379                                                 perf_evsel__name(counter),
 380                                                 cpu,
 381                                                 count->val, count->ena, count->run);
 382                         }
 383                 }
 384         }
 385
 386         return 0;
 387 }
 388
 389 static void read_counters(void)
 390 {
 391         struct perf_evsel *counter;
 392         int ret;
 393
 394         evlist__for_each_entry(evsel_list, counter) {
 395                 ret = read_counter(counter);
 396                 if (ret)
 397                         pr_debug("failed to read counter %s\n", counter->name);
 398
 399                 if (ret == 0 && perf_stat_process_counter(&stat_config, counter))
 400                         pr_warning("failed to process counter %s\n", counter->name);
 401         }
 402 }
 403
 404 static void process_interval(void)
 405 {
 406         struct timespec ts, rs;
 407
 408         read_counters();
 409
 410         clock_gettime(CLOCK_MONOTONIC, &ts);
 411         diff_timespec(&rs, &ts, &ref_time);
 412
 413         if (STAT_RECORD) {
 414                 if (WRITE_STAT_ROUND_EVENT(rs.tv_sec * NSEC_PER_SEC + rs.tv_nsec, INTERVAL))
 415                         pr_err("failed to write stat round event\n");
 416         }
 417
 418         init_stats(&walltime_nsecs_stats);
 419         update_stats(&walltime_nsecs_stats, stat_config.interval * 1000000);
 420         print_counters(&rs, 0, NULL);
 421 }
 422
 423 static void enable_counters(void)
 424 {
 425         if (initial_delay)
 426                 usleep(initial_delay * USEC_PER_MSEC);
 427
 428         /*
 429          * We need to enable counters only if:
 430          * - we don't have tracee (attaching to task or cpu)
 431          * - we have initial delay configured
 432          */
 433         if (!target__none(&target) || initial_delay)
 434                 perf_evlist__enable(evsel_list);
 435 }
 436
 437 static void disable_counters(void)
 438 {
 439         /*
 440          * If we don't have tracee (attaching to task or cpu), counters may
 441          * still be running. To get accurate group ratios, we must stop groups
 442          * from counting before reading their constituent counters.
 443          */
 444         if (!target__none(&target))
 445                 perf_evlist__disable(evsel_list);
 446 }
 447
 448 static volatile int workload_exec_errno;
 449
 450 /*
 451  * perf_evlist__prepare_workload will send a SIGUSR1
 452  * if the fork fails, since we asked by setting its
 453  * want_signal to true.
 454  */
 455 static void workload_exec_failed_signal(int signo __maybe_unused, siginfo_t *info,
 456                                         void *ucontext __maybe_unused)
 457 {
 458         workload_exec_errno = info->si_value.sival_int;
 459 }
 460
 461 static bool has_unit(struct perf_evsel *counter)
 462 {
 463         return counter->unit && *counter->unit;
 464 }
 465
 466 static bool has_scale(struct perf_evsel *counter)
 467 {
 468         return counter->scale != 1;
 469 }
 470
 471 static int perf_stat_synthesize_config(bool is_pipe)
 472 {
 473         struct perf_evsel *counter;
 474         int err;
 475
 476         if (is_pipe) {
 477                 err = perf_event__synthesize_attrs(NULL, perf_stat.session,
 478                                                    process_synthesized_event);
 479                 if (err < 0) {
 480                         pr_err("Couldn't synthesize attrs.\n");
 481                         return err;
 482                 }
 483         }
 484
 485         /*
 486          * Synthesize other events stuff not carried within
 487          * attr event - unit, scale, name
 488          */
 489         evlist__for_each_entry(evsel_list, counter) {
 490                 if (!counter->supported)
 491                         continue;
 492
 493                 /*
 494                  * Synthesize unit and scale only if it's defined.
 495                  */
 496                 if (has_unit(counter)) {
 497                         err = perf_event__synthesize_event_update_unit(NULL, counter, process_synthesized_event);
 498                         if (err < 0) {
 499                                 pr_err("Couldn't synthesize evsel unit.\n");
 500                                 return err;
 501                         }
 502                 }
 503
 504                 if (has_scale(counter)) {
 505                         err = perf_event__synthesize_event_update_scale(NULL, counter, process_synthesized_event);
 506                         if (err < 0) {
 507                                 pr_err("Couldn't synthesize evsel scale.\n");
 508                                 return err;
 509                         }
 510                 }
 511
 512                 if (counter->own_cpus) {
 513                         err = perf_event__synthesize_event_update_cpus(NULL, counter, process_synthesized_event);
 514                         if (err < 0) {
 515                                 pr_err("Couldn't synthesize evsel scale.\n");
 516                                 return err;
 517                         }
 518                 }
 519
 520                 /*
 521                  * Name is needed only for pipe output,
 522                  * perf.data carries event names.
 523                  */
 524                 if (is_pipe) {
 525                         err = perf_event__synthesize_event_update_name(NULL, counter, process_synthesized_event);
 526                         if (err < 0) {
 527                                 pr_err("Couldn't synthesize evsel name.\n");
 528                                 return err;
 529                         }
 530                 }
 531         }
 532
 533         err = perf_event__synthesize_thread_map2(NULL, evsel_list->threads,
 534                                                 process_synthesized_event,
 535                                                 NULL);
 536         if (err < 0) {
 537                 pr_err("Couldn't synthesize thread map.\n");
 538                 return err;
 539         }
 540
 541         err = perf_event__synthesize_cpu_map(NULL, evsel_list->cpus,
 542                                              process_synthesized_event, NULL);
 543         if (err < 0) {
 544                 pr_err("Couldn't synthesize thread map.\n");
 545                 return err;
 546         }
 547
 548         err = perf_event__synthesize_stat_config(NULL, &stat_config,
 549                                                  process_synthesized_event, NULL);
 550         if (err < 0) {
 551                 pr_err("Couldn't synthesize config.\n");
 552                 return err;
 553         }
 554
 555         return 0;
 556 }
 557
 558 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
 559
 560 static int __store_counter_ids(struct perf_evsel *counter,
 561                                struct cpu_map *cpus,
 562                                struct thread_map *threads)
 563 {
 564         int cpu, thread;
 565
 566         for (cpu = 0; cpu < cpus->nr; cpu++) {
 567                 for (thread = 0; thread < threads->nr; thread++) {
 568                         int fd = FD(counter, cpu, thread);
 569
 570                         if (perf_evlist__id_add_fd(evsel_list, counter,
 571                                                    cpu, thread, fd) < 0)
 572                                 return -1;
 573                 }
 574         }
 575
 576         return 0;
 577 }
 578
 579 static int store_counter_ids(struct perf_evsel *counter)
 580 {
 581         struct cpu_map *cpus = counter->cpus;
 582         struct thread_map *threads = counter->threads;
 583
 584         if (perf_evsel__alloc_id(counter, cpus->nr, threads->nr))
 585                 return -ENOMEM;
 586
 587         return __store_counter_ids(counter, cpus, threads);
 588 }
 589
 590 static bool perf_evsel__should_store_id(struct perf_evsel *counter)
 591 {
 592         return STAT_RECORD || counter->attr.read_format & PERF_FORMAT_ID;
 593 }
 594
 595 static struct perf_evsel *perf_evsel__reset_weak_group(struct perf_evsel *evsel)
 596 {
 597         struct perf_evsel *c2, *leader;
 598         bool is_open = true;
 599
 600         leader = evsel->leader;
 601         pr_debug("Weak group for %s/%d failed\n",
 602                         leader->name, leader->nr_members);
 603
 604         /*
 605          * for_each_group_member doesn't work here because it doesn't
 606          * include the first entry.
 607          */
 608         evlist__for_each_entry(evsel_list, c2) {
 609                 if (c2 == evsel)
 610                         is_open = false;
 611                 if (c2->leader == leader) {
 612                         if (is_open)
 613                                 perf_evsel__close(c2);
 614                         c2->leader = c2;
 615                         c2->nr_members = 0;
 616                 }
 617         }
 618         return leader;
 619 }
 620
 621 static int __run_perf_stat(int argc, const char **argv)
 622 {
 623         int interval = stat_config.interval;
 624         char msg[BUFSIZ];
 625         unsigned long long t0, t1;
 626         struct perf_evsel *counter;
 627         struct timespec ts;
 628         size_t l;
 629         int status = 0;
 630         const bool forks = (argc > 0);
 631         bool is_pipe = STAT_RECORD ? perf_stat.data.is_pipe : false;
 632         struct perf_evsel_config_term *err_term;
 633
 634         if (interval) {
 635                 ts.tv_sec  = interval / USEC_PER_MSEC;
 636                 ts.tv_nsec = (interval % USEC_PER_MSEC) * NSEC_PER_MSEC;
 637         } else {
 638                 ts.tv_sec  = 1;
 639                 ts.tv_nsec = 0;
 640         }
 641
 642         if (forks) {
 643                 if (perf_evlist__prepare_workload(evsel_list, &target, argv, is_pipe,
 644                                                   workload_exec_failed_signal) < 0) {
 645                         perror("failed to prepare workload");
 646                         return -1;
 647                 }
 648                 child_pid = evsel_list->workload.pid;
 649         }
 650
 651         if (group)
 652                 perf_evlist__set_leader(evsel_list);
 653
 654         evlist__for_each_entry(evsel_list, counter) {
 655 try_again:
 656                 if (create_perf_stat_counter(counter) < 0) {
 657
 658                         /* Weak group failed. Reset the group. */
 659                         if ((errno == EINVAL || errno == EBADF) &&
 660                             counter->leader != counter &&
 661                             counter->weak_group) {
 662                                 counter = perf_evsel__reset_weak_group(counter);
 663                                 goto try_again;
 664                         }
 665
 666                         /*
 667                          * PPC returns ENXIO for HW counters until 2.6.37
 668                          * (behavior changed with commit b0a873e).
 669                          */
 670                         if (errno == EINVAL || errno == ENOSYS ||
 671                             errno == ENOENT || errno == EOPNOTSUPP ||
 672                             errno == ENXIO) {
 673                                 if (verbose > 0)
 674                                         ui__warning("%s event is not supported by the kernel.\n",
 675                                                     perf_evsel__name(counter));
 676                                 counter->supported = false;
 677
 678                                 if ((counter->leader != counter) ||
 679                                     !(counter->leader->nr_members > 1))
 680                                         continue;
 681                         } else if (perf_evsel__fallback(counter, errno, msg, sizeof(msg))) {
 682                                 if (verbose > 0)
 683                                         ui__warning("%s\n", msg);
 684                                 goto try_again;
 685                         }
 686
 687                         perf_evsel__open_strerror(counter, &target,
 688                                                   errno, msg, sizeof(msg));
 689                         ui__error("%s\n", msg);
 690
 691                         if (child_pid != -1)
 692                                 kill(child_pid, SIGTERM);
 693
 694                         return -1;
 695                 }
 696                 counter->supported = true;
 697
 698                 l = strlen(counter->unit);
 699                 if (l > unit_width)
 700                         unit_width = l;
 701
 702                 if (perf_evsel__should_store_id(counter) &&
 703                     store_counter_ids(counter))
 704                         return -1;
 705         }
 706
 707         if (perf_evlist__apply_filters(evsel_list, &counter)) {
 708                 pr_err("failed to set filter \"%s\" on event %s with %d (%s)\n",
 709                         counter->filter, perf_evsel__name(counter), errno,
 710                         str_error_r(errno, msg, sizeof(msg)));
 711                 return -1;
 712         }
 713
 714         if (perf_evlist__apply_drv_configs(evsel_list, &counter, &err_term)) {
 715                 pr_err("failed to set config \"%s\" on event %s with %d (%s)\n",
 716                       err_term->val.drv_cfg, perf_evsel__name(counter), errno,
 717                       str_error_r(errno, msg, sizeof(msg)));
 718                 return -1;
 719         }
 720
 721         if (STAT_RECORD) {
 722                 int err, fd = perf_data__fd(&perf_stat.data);
 723
 724                 if (is_pipe) {
 725                         err = perf_header__write_pipe(perf_data__fd(&perf_stat.data));
 726                 } else {
 727                         err = perf_session__write_header(perf_stat.session, evsel_list,
 728                                                          fd, false);
 729                 }
 730
 731                 if (err < 0)
 732                         return err;
 733
 734                 err = perf_stat_synthesize_config(is_pipe);
 735                 if (err < 0)
 736                         return err;
 737         }
 738
 739         /*
 740          * Enable counters and exec the command:
 741          */
 742         t0 = rdclock();
 743         clock_gettime(CLOCK_MONOTONIC, &ref_time);
 744
 745         if (forks) {
 746                 perf_evlist__start_workload(evsel_list);
 747                 enable_counters();
 748
 749                 if (interval) {
 750                         while (!waitpid(child_pid, &status, WNOHANG)) {
 751                                 nanosleep(&ts, NULL);
 752                                 process_interval();
 753                         }
 754                 }
 755                 waitpid(child_pid, &status, 0);
 756
 757                 if (workload_exec_errno) {
 758                         const char *emsg = str_error_r(workload_exec_errno, msg, sizeof(msg));
 759                         pr_err("Workload failed: %s\n", emsg);
 760                         return -1;
 761                 }
 762
 763                 if (WIFSIGNALED(status))
 764                         psignal(WTERMSIG(status), argv[0]);
 765         } else {
 766                 enable_counters();
 767                 while (!done) {
 768                         nanosleep(&ts, NULL);
 769                         if (interval)
 770                                 process_interval();
 771                 }
 772         }
 773
 774         disable_counters();
 775
 776         t1 = rdclock();
 777
 778         update_stats(&walltime_nsecs_stats, t1 - t0);
 779
 780         /*
 781          * Closing a group leader splits the group, and as we only disable
 782          * group leaders, results in remaining events becoming enabled. To
 783          * avoid arbitrary skew, we must read all counters before closing any
 784          * group leaders.
 785          */
 786         read_counters();
 787         perf_evlist__close(evsel_list);
 788
 789         return WEXITSTATUS(status);
 790 }
 791
 792 static int run_perf_stat(int argc, const char **argv)
 793 {
 794         int ret;
 795
 796         if (pre_cmd) {
 797                 ret = system(pre_cmd);
 798                 if (ret)
 799                         return ret;
 800         }
 801
 802         if (sync_run)
 803                 sync();
 804
 805         ret = __run_perf_stat(argc, argv);
 806         if (ret)
 807                 return ret;
 808
 809         if (post_cmd) {
 810                 ret = system(post_cmd);
 811                 if (ret)
 812                         return ret;
 813         }
 814
 815         return ret;
 816 }
 817
 818 static void print_running(u64 run, u64 ena)
 819 {
 820         if (csv_output) {
 821                 fprintf(stat_config.output, "%s%" PRIu64 "%s%.2f",
 822                                         csv_sep,
 823                                         run,
 824                                         csv_sep,
 825                                         ena ? 100.0 * run / ena : 100.0);
 826         } else if (run != ena) {
 827                 fprintf(stat_config.output, "  (%.2f%%)", 100.0 * run / ena);
 828         }
 829 }
 830
 831 static void print_noise_pct(double total, double avg)
 832 {
 833         double pct = rel_stddev_stats(total, avg);
 834
 835         if (csv_output)
 836                 fprintf(stat_config.output, "%s%.2f%%", csv_sep, pct);
 837         else if (pct)
 838                 fprintf(stat_config.output, "  ( +-%6.2f%% )", pct);
 839 }
 840
 841 static void print_noise(struct perf_evsel *evsel, double avg)
 842 {
 843         struct perf_stat_evsel *ps;
 844
 845         if (run_count == 1)
 846                 return;
 847
 848         ps = evsel->stats;
 849         print_noise_pct(stddev_stats(&ps->res_stats[0]), avg);
 850 }
 851
 852 static void aggr_printout(struct perf_evsel *evsel, int id, int nr)
 853 {
 854         switch (stat_config.aggr_mode) {
 855         case AGGR_CORE:
 856                 fprintf(stat_config.output, "S%d-C%*d%s%*d%s",
 857                         cpu_map__id_to_socket(id),
 858                         csv_output ? 0 : -8,
 859                         cpu_map__id_to_cpu(id),
 860                         csv_sep,
 861                         csv_output ? 0 : 4,
 862                         nr,
 863                         csv_sep);
 864                 break;
 865         case AGGR_SOCKET:
 866                 fprintf(stat_config.output, "S%*d%s%*d%s",
 867                         csv_output ? 0 : -5,
 868                         id,
 869                         csv_sep,
 870                         csv_output ? 0 : 4,
 871                         nr,
 872                         csv_sep);
 873                         break;
 874         case AGGR_NONE:
 875                 fprintf(stat_config.output, "CPU%*d%s",
 876                         csv_output ? 0 : -4,
 877                         perf_evsel__cpus(evsel)->map[id], csv_sep);
 878                 break;
 879         case AGGR_THREAD:
 880                 fprintf(stat_config.output, "%*s-%*d%s",
 881                         csv_output ? 0 : 16,
 882                         thread_map__comm(evsel->threads, id),
 883                         csv_output ? 0 : -8,
 884                         thread_map__pid(evsel->threads, id),
 885                         csv_sep);
 886                 break;
 887         case AGGR_GLOBAL:
 888         case AGGR_UNSET:
 889         default:
 890                 break;
 891         }
 892 }
 893
 894 struct outstate {
 895         FILE *fh;
 896         bool newline;
 897         const char *prefix;
 898         int  nfields;
 899         int  id, nr;
 900         struct perf_evsel *evsel;
 901 };
 902
 903 #define METRIC_LEN  35
 904
 905 static void new_line_std(void *ctx)
 906 {
 907         struct outstate *os = ctx;
 908
 909         os->newline = true;
 910 }
 911
 912 static void do_new_line_std(struct outstate *os)
 913 {
 914         fputc('\n', os->fh);
 915         fputs(os->prefix, os->fh);
 916         aggr_printout(os->evsel, os->id, os->nr);
 917         if (stat_config.aggr_mode == AGGR_NONE)
 918                 fprintf(os->fh, "        ");
 919         fprintf(os->fh, "                                                 ");
 920 }
 921
 922 static void print_metric_std(void *ctx, const char *color, const char *fmt,
 923                              const char *unit, double val)
 924 {
 925         struct outstate *os = ctx;
 926         FILE *out = os->fh;
 927         int n;
 928         bool newline = os->newline;
 929
 930         os->newline = false;
 931
 932         if (unit == NULL || fmt == NULL) {
 933                 fprintf(out, "%-*s", METRIC_LEN, "");
 934                 return;
 935         }
 936
 937         if (newline)
 938                 do_new_line_std(os);
 939
 940         n = fprintf(out, " # ");
 941         if (color)
 942                 n += color_fprintf(out, color, fmt, val);
 943         else
 944                 n += fprintf(out, fmt, val);
 945         fprintf(out, " %-*s", METRIC_LEN - n - 1, unit);
 946 }
 947
 948 static void new_line_csv(void *ctx)
 949 {
 950         struct outstate *os = ctx;
 951         int i;
 952
 953         fputc('\n', os->fh);
 954         if (os->prefix)
 955                 fprintf(os->fh, "%s%s", os->prefix, csv_sep);
 956         aggr_printout(os->evsel, os->id, os->nr);
 957         for (i = 0; i < os->nfields; i++)
 958                 fputs(csv_sep, os->fh);
 959 }
 960
 961 static void print_metric_csv(void *ctx,
 962                              const char *color __maybe_unused,
 963                              const char *fmt, const char *unit, double val)
 964 {
 965         struct outstate *os = ctx;
 966         FILE *out = os->fh;
 967         char buf[64], *vals, *ends;
 968
 969         if (unit == NULL || fmt == NULL) {
 970                 fprintf(out, "%s%s%s%s", csv_sep, csv_sep, csv_sep, csv_sep);
 971                 return;
 972         }
 973         snprintf(buf, sizeof(buf), fmt, val);
 974         ends = vals = ltrim(buf);
 975         while (isdigit(*ends) || *ends == '.')
 976                 ends++;
 977         *ends = 0;
 978         while (isspace(*unit))
 979                 unit++;
 980         fprintf(out, "%s%s%s%s", csv_sep, vals, csv_sep, unit);
 981 }
 982
 983 #define METRIC_ONLY_LEN 20
 984
 985 /* Filter out some columns that don't work well in metrics only mode */
 986
 987 static bool valid_only_metric(const char *unit)
 988 {
 989         if (!unit)
 990                 return false;
 991         if (strstr(unit, "/sec") ||
 992             strstr(unit, "hz") ||
 993             strstr(unit, "Hz") ||
 994             strstr(unit, "CPUs utilized"))
 995                 return false;
 996         return true;
 997 }
 998
 999 static const char *fixunit(char *buf, struct perf_evsel *evsel,
1000                            const char *unit)
1001 {
1002         if (!strncmp(unit, "of all", 6)) {
1003                 snprintf(buf, 1024, "%s %s", perf_evsel__name(evsel),
1004                          unit);
1005                 return buf;
1006         }
1007         return unit;
1008 }
1009
1010 static void print_metric_only(void *ctx, const char *color, const char *fmt,
1011                               const char *unit, double val)
1012 {
1013         struct outstate *os = ctx;
1014         FILE *out = os->fh;
1015         int n;
1016         char buf[1024];
1017         unsigned mlen = METRIC_ONLY_LEN;
1018
1019         if (!valid_only_metric(unit))
1020                 return;
1021         unit = fixunit(buf, os->evsel, unit);
1022         if (color)
1023                 n = color_fprintf(out, color, fmt, val);
1024         else
1025                 n = fprintf(out, fmt, val);
1026         if (n > METRIC_ONLY_LEN)
1027                 n = METRIC_ONLY_LEN;
1028         if (mlen < strlen(unit))
1029                 mlen = strlen(unit) + 1;
1030         fprintf(out, "%*s", mlen - n, "");
1031 }
1032
1033 static void print_metric_only_csv(void *ctx, const char *color __maybe_unused,
1034                                   const char *fmt,
1035                                   const char *unit, double val)
1036 {
1037         struct outstate *os = ctx;
1038         FILE *out = os->fh;
1039         char buf[64], *vals, *ends;
1040         char tbuf[1024];
1041
1042         if (!valid_only_metric(unit))
1043                 return;
1044         unit = fixunit(tbuf, os->evsel, unit);
1045         snprintf(buf, sizeof buf, fmt, val);
1046         ends = vals = ltrim(buf);
1047         while (isdigit(*ends) || *ends == '.')
1048                 ends++;
1049         *ends = 0;
1050         fprintf(out, "%s%s", vals, csv_sep);
1051 }
1052
1053 static void new_line_metric(void *ctx __maybe_unused)
1054 {
1055 }
1056
1057 static void print_metric_header(void *ctx, const char *color __maybe_unused,
1058                                 const char *fmt __maybe_unused,
1059                                 const char *unit, double val __maybe_unused)
1060 {
1061         struct outstate *os = ctx;
1062         char tbuf[1024];
1063
1064         if (!valid_only_metric(unit))
1065                 return;
1066         unit = fixunit(tbuf, os->evsel, unit);
1067         if (csv_output)
1068                 fprintf(os->fh, "%s%s", unit, csv_sep);
1069         else
1070                 fprintf(os->fh, "%-*s ", METRIC_ONLY_LEN, unit);
1071 }
1072
1073 static void nsec_printout(int id, int nr, struct perf_evsel *evsel, double avg)
1074 {
1075         FILE *output = stat_config.output;
1076         double msecs = avg / NSEC_PER_MSEC;
1077         const char *fmt_v, *fmt_n;
1078         char name[25];
1079
1080         fmt_v = csv_output ? "%.6f%s" : "%18.6f%s";
1081         fmt_n = csv_output ? "%s" : "%-25s";
1082
1083         aggr_printout(evsel, id, nr);
1084
1085         scnprintf(name, sizeof(name), "%s%s",
1086                   perf_evsel__name(evsel), csv_output ? "" : " (msec)");
1087
1088         fprintf(output, fmt_v, msecs, csv_sep);
1089
1090         if (csv_output)
1091                 fprintf(output, "%s%s", evsel->unit, csv_sep);
1092         else
1093                 fprintf(output, "%-*s%s", unit_width, evsel->unit, csv_sep);
1094
1095         fprintf(output, fmt_n, name);
1096
1097         if (evsel->cgrp)
1098                 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
1099 }
1100
1101 static int first_shadow_cpu(struct perf_evsel *evsel, int id)
1102 {
1103         int i;
1104
1105         if (!aggr_get_id)
1106                 return 0;
1107
1108         if (stat_config.aggr_mode == AGGR_NONE)
1109                 return id;
1110
1111         if (stat_config.aggr_mode == AGGR_GLOBAL)
1112                 return 0;
1113
1114         for (i = 0; i < perf_evsel__nr_cpus(evsel); i++) {
1115                 int cpu2 = perf_evsel__cpus(evsel)->map[i];
1116
1117                 if (aggr_get_id(evsel_list->cpus, cpu2) == id)
1118                         return cpu2;
1119         }
1120         return 0;
1121 }
1122
1123 static void abs_printout(int id, int nr, struct perf_evsel *evsel, double avg)
1124 {
1125         FILE *output = stat_config.output;
1126         double sc =  evsel->scale;
1127         const char *fmt;
1128
1129         if (csv_output) {
1130                 fmt = floor(sc) != sc ?  "%.2f%s" : "%.0f%s";
1131         } else {
1132                 if (big_num)
1133                         fmt = floor(sc) != sc ? "%'18.2f%s" : "%'18.0f%s";
1134                 else
1135                         fmt = floor(sc) != sc ? "%18.2f%s" : "%18.0f%s";
1136         }
1137
1138         aggr_printout(evsel, id, nr);
1139
1140         fprintf(output, fmt, avg, csv_sep);
1141
1142         if (evsel->unit)
1143                 fprintf(output, "%-*s%s",
1144                         csv_output ? 0 : unit_width,
1145                         evsel->unit, csv_sep);
1146
1147         fprintf(output, "%-*s", csv_output ? 0 : 25, perf_evsel__name(evsel));
1148
1149         if (evsel->cgrp)
1150                 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
1151 }
1152
1153 static void printout(int id, int nr, struct perf_evsel *counter, double uval,
1154                      char *prefix, u64 run, u64 ena, double noise)
1155 {
1156         struct perf_stat_output_ctx out;
1157         struct outstate os = {
1158                 .fh = stat_config.output,
1159                 .prefix = prefix ? prefix : "",
1160                 .id = id,
1161                 .nr = nr,
1162                 .evsel = counter,
1163         };
1164         print_metric_t pm = print_metric_std;
1165         void (*nl)(void *);
1166
1167         if (metric_only) {
1168                 nl = new_line_metric;
1169                 if (csv_output)
1170                         pm = print_metric_only_csv;
1171                 else
1172                         pm = print_metric_only;
1173         } else
1174                 nl = new_line_std;
1175
1176         if (csv_output && !metric_only) {
1177                 static int aggr_fields[] = {
1178                         [AGGR_GLOBAL] = 0,
1179                         [AGGR_THREAD] = 1,
1180                         [AGGR_NONE] = 1,
1181                         [AGGR_SOCKET] = 2,
1182                         [AGGR_CORE] = 2,
1183                 };
1184
1185                 pm = print_metric_csv;
1186                 nl = new_line_csv;
1187                 os.nfields = 3;
1188                 os.nfields += aggr_fields[stat_config.aggr_mode];
1189                 if (counter->cgrp)
1190                         os.nfields++;
1191         }
1192         if (run == 0 || ena == 0 || counter->counts->scaled == -1) {
1193                 if (metric_only) {
1194                         pm(&os, NULL, "", "", 0);
1195                         return;
1196                 }
1197                 aggr_printout(counter, id, nr);
1198
1199                 fprintf(stat_config.output, "%*s%s",
1200                         csv_output ? 0 : 18,
1201                         counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
1202                         csv_sep);
1203
1204                 if (counter->supported)
1205                         print_free_counters_hint = 1;
1206
1207                 fprintf(stat_config.output, "%-*s%s",
1208                         csv_output ? 0 : unit_width,
1209                         counter->unit, csv_sep);
1210
1211                 fprintf(stat_config.output, "%*s",
1212                         csv_output ? 0 : -25,
1213                         perf_evsel__name(counter));
1214
1215                 if (counter->cgrp)
1216                         fprintf(stat_config.output, "%s%s",
1217                                 csv_sep, counter->cgrp->name);
1218
1219                 if (!csv_output)
1220                         pm(&os, NULL, NULL, "", 0);
1221                 print_noise(counter, noise);
1222                 print_running(run, ena);
1223                 if (csv_output)
1224                         pm(&os, NULL, NULL, "", 0);
1225                 return;
1226         }
1227
1228         if (metric_only)
1229                 /* nothing */;
1230         else if (nsec_counter(counter))
1231                 nsec_printout(id, nr, counter, uval);
1232         else
1233                 abs_printout(id, nr, counter, uval);
1234
1235         out.print_metric = pm;
1236         out.new_line = nl;
1237         out.ctx = &os;
1238         out.force_header = false;
1239
1240         if (csv_output && !metric_only) {
1241                 print_noise(counter, noise);
1242                 print_running(run, ena);
1243         }
1244
1245         perf_stat__print_shadow_stats(counter, uval,
1246                                 first_shadow_cpu(counter, id),
1247                                 &out, &metric_events);
1248         if (!csv_output && !metric_only) {
1249                 print_noise(counter, noise);
1250                 print_running(run, ena);
1251         }
1252 }
1253
1254 static void aggr_update_shadow(void)
1255 {
1256         int cpu, s2, id, s;
1257         u64 val;
1258         struct perf_evsel *counter;
1259
1260         for (s = 0; s < aggr_map->nr; s++) {
1261                 id = aggr_map->map[s];
1262                 evlist__for_each_entry(evsel_list, counter) {
1263                         val = 0;
1264                         for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
1265                                 s2 = aggr_get_id(evsel_list->cpus, cpu);
1266                                 if (s2 != id)
1267                                         continue;
1268                                 val += perf_counts(counter->counts, cpu, 0)->val;
1269                         }
1270                         perf_stat__update_shadow_stats(counter, val,
1271                                                        first_shadow_cpu(counter, id));
1272                 }
1273         }
1274 }
1275
1276 static void collect_all_aliases(struct perf_evsel *counter,
1277                             void (*cb)(struct perf_evsel *counter, void *data,
1278                                        bool first),
1279                             void *data)
1280 {
1281         struct perf_evsel *alias;
1282
1283         alias = list_prepare_entry(counter, &(evsel_list->entries), node);
1284         list_for_each_entry_continue (alias, &evsel_list->entries, node) {
1285                 if (strcmp(perf_evsel__name(alias), perf_evsel__name(counter)) ||
1286                     alias->scale != counter->scale ||
1287                     alias->cgrp != counter->cgrp ||
1288                     strcmp(alias->unit, counter->unit) ||
1289                     nsec_counter(alias) != nsec_counter(counter))
1290                         break;
1291                 alias->merged_stat = true;
1292                 cb(alias, data, false);
1293         }
1294 }
1295
1296 static bool collect_data(struct perf_evsel *counter,
1297                             void (*cb)(struct perf_evsel *counter, void *data,
1298                                        bool first),
1299                             void *data)
1300 {
1301         if (counter->merged_stat)
1302                 return false;
1303         cb(counter, data, true);
1304         if (!no_merge && counter->auto_merge_stats)
1305                 collect_all_aliases(counter, cb, data);
1306         return true;
1307 }
1308
1309 struct aggr_data {
1310         u64 ena, run, val;
1311         int id;
1312         int nr;
1313         int cpu;
1314 };
1315
1316 static void aggr_cb(struct perf_evsel *counter, void *data, bool first)
1317 {
1318         struct aggr_data *ad = data;
1319         int cpu, s2;
1320
1321         for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
1322                 struct perf_counts_values *counts;
1323
1324                 s2 = aggr_get_id(perf_evsel__cpus(counter), cpu);
1325                 if (s2 != ad->id)
1326                         continue;
1327                 if (first)
1328                         ad->nr++;
1329                 counts = perf_counts(counter->counts, cpu, 0);
1330                 /*
1331                  * When any result is bad, make them all to give
1332                  * consistent output in interval mode.
1333                  */
1334                 if (counts->ena == 0 || counts->run == 0 ||
1335                     counter->counts->scaled == -1) {
1336                         ad->ena = 0;
1337                         ad->run = 0;
1338                         break;
1339                 }
1340                 ad->val += counts->val;
1341                 ad->ena += counts->ena;
1342                 ad->run += counts->run;
1343         }
1344 }
1345
1346 static void print_aggr(char *prefix)
1347 {
1348         FILE *output = stat_config.output;
1349         struct perf_evsel *counter;
1350         int s, id, nr;
1351         double uval;
1352         u64 ena, run, val;
1353         bool first;
1354
1355         if (!(aggr_map || aggr_get_id))
1356                 return;
1357
1358         aggr_update_shadow();
1359
1360         /*
1361          * With metric_only everything is on a single line.
1362          * Without each counter has its own line.
1363          */
1364         for (s = 0; s < aggr_map->nr; s++) {
1365                 struct aggr_data ad;
1366                 if (prefix && metric_only)
1367                         fprintf(output, "%s", prefix);
1368
1369                 ad.id = id = aggr_map->map[s];
1370                 first = true;
1371                 evlist__for_each_entry(evsel_list, counter) {
1372                         if (is_duration_time(counter))
1373                                 continue;
1374
1375                         ad.val = ad.ena = ad.run = 0;
1376                         ad.nr = 0;
1377                         if (!collect_data(counter, aggr_cb, &ad))
1378                                 continue;
1379                         nr = ad.nr;
1380                         ena = ad.ena;
1381                         run = ad.run;
1382                         val = ad.val;
1383                         if (first && metric_only) {
1384                                 first = false;
1385                                 aggr_printout(counter, id, nr);
1386                         }
1387                         if (prefix && !metric_only)
1388                                 fprintf(output, "%s", prefix);
1389
1390                         uval = val * counter->scale;
1391                         printout(id, nr, counter, uval, prefix, run, ena, 1.0);
1392                         if (!metric_only)
1393                                 fputc('\n', output);
1394                 }
1395                 if (metric_only)
1396                         fputc('\n', output);
1397         }
1398 }
1399
1400 static void print_aggr_thread(struct perf_evsel *counter, char *prefix)
1401 {
1402         FILE *output = stat_config.output;
1403         int nthreads = thread_map__nr(counter->threads);
1404         int ncpus = cpu_map__nr(counter->cpus);
1405         int cpu, thread;
1406         double uval;
1407
1408         for (thread = 0; thread < nthreads; thread++) {
1409                 u64 ena = 0, run = 0, val = 0;
1410
1411                 for (cpu = 0; cpu < ncpus; cpu++) {
1412                         val += perf_counts(counter->counts, cpu, thread)->val;
1413                         ena += perf_counts(counter->counts, cpu, thread)->ena;
1414                         run += perf_counts(counter->counts, cpu, thread)->run;
1415                 }
1416
1417                 if (prefix)
1418                         fprintf(output, "%s", prefix);
1419
1420                 uval = val * counter->scale;
1421                 printout(thread, 0, counter, uval, prefix, run, ena, 1.0);
1422                 fputc('\n', output);
1423         }
1424 }
1425
1426 struct caggr_data {
1427         double avg, avg_enabled, avg_running;
1428 };
1429
1430 static void counter_aggr_cb(struct perf_evsel *counter, void *data,
1431                             bool first __maybe_unused)
1432 {
1433         struct caggr_data *cd = data;
1434         struct perf_stat_evsel *ps = counter->stats;
1435
1436         cd->avg += avg_stats(&ps->res_stats[0]);
1437         cd->avg_enabled += avg_stats(&ps->res_stats[1]);
1438         cd->avg_running += avg_stats(&ps->res_stats[2]);
1439 }
1440
1441 /*
1442  * Print out the results of a single counter:
1443  * aggregated counts in system-wide mode
1444  */
1445 static void print_counter_aggr(struct perf_evsel *counter, char *prefix)
1446 {
1447         FILE *output = stat_config.output;
1448         double uval;
1449         struct caggr_data cd = { .avg = 0.0 };
1450
1451         if (!collect_data(counter, counter_aggr_cb, &cd))
1452                 return;
1453
1454         if (prefix && !metric_only)
1455                 fprintf(output, "%s", prefix);
1456
1457         uval = cd.avg * counter->scale;
1458         printout(-1, 0, counter, uval, prefix, cd.avg_running, cd.avg_enabled, cd.avg);
1459         if (!metric_only)
1460                 fprintf(output, "\n");
1461 }
1462
1463 static void counter_cb(struct perf_evsel *counter, void *data,
1464                        bool first __maybe_unused)
1465 {
1466         struct aggr_data *ad = data;
1467
1468         ad->val += perf_counts(counter->counts, ad->cpu, 0)->val;
1469         ad->ena += perf_counts(counter->counts, ad->cpu, 0)->ena;
1470         ad->run += perf_counts(counter->counts, ad->cpu, 0)->run;
1471 }
1472
1473 /*
1474  * Print out the results of a single counter:
1475  * does not use aggregated count in system-wide
1476  */
1477 static void print_counter(struct perf_evsel *counter, char *prefix)
1478 {
1479         FILE *output = stat_config.output;
1480         u64 ena, run, val;
1481         double uval;
1482         int cpu;
1483
1484         for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
1485                 struct aggr_data ad = { .cpu = cpu };
1486
1487                 if (!collect_data(counter, counter_cb, &ad))
1488                         return;
1489                 val = ad.val;
1490                 ena = ad.ena;
1491                 run = ad.run;
1492
1493                 if (prefix)
1494                         fprintf(output, "%s", prefix);
1495
1496                 uval = val * counter->scale;
1497                 printout(cpu, 0, counter, uval, prefix, run, ena, 1.0);
1498
1499                 fputc('\n', output);
1500         }
1501 }
1502
1503 static void print_no_aggr_metric(char *prefix)
1504 {
1505         int cpu;
1506         int nrcpus = 0;
1507         struct perf_evsel *counter;
1508         u64 ena, run, val;
1509         double uval;
1510
1511         nrcpus = evsel_list->cpus->nr;
1512         for (cpu = 0; cpu < nrcpus; cpu++) {
1513                 bool first = true;
1514
1515                 if (prefix)
1516                         fputs(prefix, stat_config.output);
1517                 evlist__for_each_entry(evsel_list, counter) {
1518                         if (is_duration_time(counter))
1519                                 continue;
1520                         if (first) {
1521                                 aggr_printout(counter, cpu, 0);
1522                                 first = false;
1523                         }
1524                         val = perf_counts(counter->counts, cpu, 0)->val;
1525                         ena = perf_counts(counter->counts, cpu, 0)->ena;
1526                         run = perf_counts(counter->counts, cpu, 0)->run;
1527
1528                         uval = val * counter->scale;
1529                         printout(cpu, 0, counter, uval, prefix, run, ena, 1.0);
1530                 }
1531                 fputc('\n', stat_config.output);
1532         }
1533 }
1534
1535 static int aggr_header_lens[] = {
1536         [AGGR_CORE] = 18,
1537         [AGGR_SOCKET] = 12,
1538         [AGGR_NONE] = 6,
1539         [AGGR_THREAD] = 24,
1540         [AGGR_GLOBAL] = 0,
1541 };
1542
1543 static const char *aggr_header_csv[] = {
1544         [AGGR_CORE]     =       "core,cpus,",
1545         [AGGR_SOCKET]   =       "socket,cpus",
1546         [AGGR_NONE]     =       "cpu,",
1547         [AGGR_THREAD]   =       "comm-pid,",
1548         [AGGR_GLOBAL]   =       ""
1549 };
1550
1551 static void print_metric_headers(const char *prefix, bool no_indent)
1552 {
1553         struct perf_stat_output_ctx out;
1554         struct perf_evsel *counter;
1555         struct outstate os = {
1556                 .fh = stat_config.output
1557         };
1558
1559         if (prefix)
1560                 fprintf(stat_config.output, "%s", prefix);
1561
1562         if (!csv_output && !no_indent)
1563                 fprintf(stat_config.output, "%*s",
1564                         aggr_header_lens[stat_config.aggr_mode], "");
1565         if (csv_output) {
1566                 if (stat_config.interval)
1567                         fputs("time,", stat_config.output);
1568                 fputs(aggr_header_csv[stat_config.aggr_mode],
1569                         stat_config.output);
1570         }
1571
1572         /* Print metrics headers only */
1573         evlist__for_each_entry(evsel_list, counter) {
1574                 if (is_duration_time(counter))
1575                         continue;
1576                 os.evsel = counter;
1577                 out.ctx = &os;
1578                 out.print_metric = print_metric_header;
1579                 out.new_line = new_line_metric;
1580                 out.force_header = true;
1581                 os.evsel = counter;
1582                 perf_stat__print_shadow_stats(counter, 0,
1583                                               0,
1584                                               &out,
1585                                               &metric_events);
1586         }
1587         fputc('\n', stat_config.output);
1588 }
1589
1590 static void print_interval(char *prefix, struct timespec *ts)
1591 {
1592         FILE *output = stat_config.output;
1593         static int num_print_interval;
1594
1595         sprintf(prefix, "%6lu.%09lu%s", ts->tv_sec, ts->tv_nsec, csv_sep);
1596
1597         if (num_print_interval == 0 && !csv_output) {
1598                 switch (stat_config.aggr_mode) {
1599                 case AGGR_SOCKET:
1600                         fprintf(output, "#           time socket cpus");
1601                         if (!metric_only)
1602                                 fprintf(output, "             counts %*s events\n", unit_width, "unit");
1603                         break;
1604                 case AGGR_CORE:
1605                         fprintf(output, "#           time core         cpus");
1606                         if (!metric_only)
1607                                 fprintf(output, "             counts %*s events\n", unit_width, "unit");
1608                         break;
1609                 case AGGR_NONE:
1610                         fprintf(output, "#           time CPU");
1611                         if (!metric_only)
1612                                 fprintf(output, "                counts %*s events\n", unit_width, "unit");
1613                         break;
1614                 case AGGR_THREAD:
1615                         fprintf(output, "#           time             comm-pid");
1616                         if (!metric_only)
1617                                 fprintf(output, "                  counts %*s events\n", unit_width, "unit");
1618                         break;
1619                 case AGGR_GLOBAL:
1620                 default:
1621                         fprintf(output, "#           time");
1622                         if (!metric_only)
1623                                 fprintf(output, "             counts %*s events\n", unit_width, "unit");
1624                 case AGGR_UNSET:
1625                         break;
1626                 }
1627         }
1628
1629         if (num_print_interval == 0 && metric_only)
1630                 print_metric_headers(" ", true);
1631         if (++num_print_interval == 25)
1632                 num_print_interval = 0;
1633 }
1634
1635 static void print_header(int argc, const char **argv)
1636 {
1637         FILE *output = stat_config.output;
1638         int i;
1639
1640         fflush(stdout);
1641
1642         if (!csv_output) {
1643                 fprintf(output, "\n");
1644                 fprintf(output, " Performance counter stats for ");
1645                 if (target.system_wide)
1646                         fprintf(output, "\'system wide");
1647                 else if (target.cpu_list)
1648                         fprintf(output, "\'CPU(s) %s", target.cpu_list);
1649                 else if (!target__has_task(&target)) {
1650                         fprintf(output, "\'%s", argv ? argv[0] : "pipe");
1651                         for (i = 1; argv && (i < argc); i++)
1652                                 fprintf(output, " %s", argv[i]);
1653                 } else if (target.pid)
1654                         fprintf(output, "process id \'%s", target.pid);
1655                 else
1656                         fprintf(output, "thread id \'%s", target.tid);
1657
1658                 fprintf(output, "\'");
1659                 if (run_count > 1)
1660                         fprintf(output, " (%d runs)", run_count);
1661                 fprintf(output, ":\n\n");
1662         }
1663 }
1664
1665 static void print_footer(void)
1666 {
1667         FILE *output = stat_config.output;
1668         int n;
1669
1670         if (!null_run)
1671                 fprintf(output, "\n");
1672         fprintf(output, " %17.9f seconds time elapsed",
1673                         avg_stats(&walltime_nsecs_stats) / NSEC_PER_SEC);
1674         if (run_count > 1) {
1675                 fprintf(output, "                                        ");
1676                 print_noise_pct(stddev_stats(&walltime_nsecs_stats),
1677                                 avg_stats(&walltime_nsecs_stats));
1678         }
1679         fprintf(output, "\n\n");
1680
1681         if (print_free_counters_hint &&
1682             sysctl__read_int("kernel/nmi_watchdog", &n) >= 0 &&
1683             n > 0)
1684                 fprintf(output,
1685 "Some events weren't counted. Try disabling the NMI watchdog:\n"
1686 "       echo 0 > /proc/sys/kernel/nmi_watchdog\n"
1687 "       perf stat ...\n"
1688 "       echo 1 > /proc/sys/kernel/nmi_watchdog\n");
1689 }
1690
1691 static void print_counters(struct timespec *ts, int argc, const char **argv)
1692 {
1693         int interval = stat_config.interval;
1694         struct perf_evsel *counter;
1695         char buf[64], *prefix = NULL;
1696
1697         /* Do not print anything if we record to the pipe. */
1698         if (STAT_RECORD && perf_stat.data.is_pipe)
1699                 return;
1700
1701         if (interval)
1702                 print_interval(prefix = buf, ts);
1703         else
1704                 print_header(argc, argv);
1705
1706         if (metric_only) {
1707                 static int num_print_iv;
1708
1709                 if (num_print_iv == 0 && !interval)
1710                         print_metric_headers(prefix, false);
1711                 if (num_print_iv++ == 25)
1712                         num_print_iv = 0;
1713                 if (stat_config.aggr_mode == AGGR_GLOBAL && prefix)
1714                         fprintf(stat_config.output, "%s", prefix);
1715         }
1716
1717         switch (stat_config.aggr_mode) {
1718         case AGGR_CORE:
1719         case AGGR_SOCKET:
1720                 print_aggr(prefix);
1721                 break;
1722         case AGGR_THREAD:
1723                 evlist__for_each_entry(evsel_list, counter) {
1724                         if (is_duration_time(counter))
1725                                 continue;
1726                         print_aggr_thread(counter, prefix);
1727                 }
1728                 break;
1729         case AGGR_GLOBAL:
1730                 evlist__for_each_entry(evsel_list, counter) {
1731                         if (is_duration_time(counter))
1732                                 continue;
1733                         print_counter_aggr(counter, prefix);
1734                 }
1735                 if (metric_only)
1736                         fputc('\n', stat_config.output);
1737                 break;
1738         case AGGR_NONE:
1739                 if (metric_only)
1740                         print_no_aggr_metric(prefix);
1741                 else {
1742                         evlist__for_each_entry(evsel_list, counter) {
1743                                 if (is_duration_time(counter))
1744                                         continue;
1745                                 print_counter(counter, prefix);
1746                         }
1747                 }
1748                 break;
1749         case AGGR_UNSET:
1750         default:
1751                 break;
1752         }
1753
1754         if (!interval && !csv_output)
1755                 print_footer();
1756
1757         fflush(stat_config.output);
1758 }
1759
1760 static volatile int signr = -1;
1761
1762 static void skip_signal(int signo)
1763 {
1764         if ((child_pid == -1) || stat_config.interval)
1765                 done = 1;
1766
1767         signr = signo;
1768         /*
1769          * render child_pid harmless
1770          * won't send SIGTERM to a random
1771          * process in case of race condition
1772          * and fast PID recycling
1773          */
1774         child_pid = -1;
1775 }
1776
1777 static void sig_atexit(void)
1778 {
1779         sigset_t set, oset;
1780
1781         /*
1782          * avoid race condition with SIGCHLD handler
1783          * in skip_signal() which is modifying child_pid
1784          * goal is to avoid send SIGTERM to a random
1785          * process
1786          */
1787         sigemptyset(&set);
1788         sigaddset(&set, SIGCHLD);
1789         sigprocmask(SIG_BLOCK, &set, &oset);
1790
1791         if (child_pid != -1)
1792                 kill(child_pid, SIGTERM);
1793
1794         sigprocmask(SIG_SETMASK, &oset, NULL);
1795
1796         if (signr == -1)
1797                 return;
1798
1799         signal(signr, SIG_DFL);
1800         kill(getpid(), signr);
1801 }
1802
1803 static int stat__set_big_num(const struct option *opt __maybe_unused,
1804                              const char *s __maybe_unused, int unset)
1805 {
1806         big_num_opt = unset ? 0 : 1;
1807         return 0;
1808 }
1809
1810 static int enable_metric_only(const struct option *opt __maybe_unused,
1811                               const char *s __maybe_unused, int unset)
1812 {
1813         force_metric_only = true;
1814         metric_only = !unset;
1815         return 0;
1816 }
1817
1818 static int parse_metric_groups(const struct option *opt,
1819                                const char *str,
1820                                int unset __maybe_unused)
1821 {
1822         return metricgroup__parse_groups(opt, str, &metric_events);
1823 }
1824
1825 static const struct option stat_options[] = {
1826         OPT_BOOLEAN('T', "transaction", &transaction_run,
1827                     "hardware transaction statistics"),
1828         OPT_CALLBACK('e', "event", &evsel_list, "event",
1829                      "event selector. use 'perf list' to list available events",
1830                      parse_events_option),
1831         OPT_CALLBACK(0, "filter", &evsel_list, "filter",
1832                      "event filter", parse_filter),
1833         OPT_BOOLEAN('i', "no-inherit", &no_inherit,
1834                     "child tasks do not inherit counters"),
1835         OPT_STRING('p', "pid", &target.pid, "pid",
1836                    "stat events on existing process id"),
1837         OPT_STRING('t', "tid", &target.tid, "tid",
1838                    "stat events on existing thread id"),
1839         OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
1840                     "system-wide collection from all CPUs"),
1841         OPT_BOOLEAN('g', "group", &group,
1842                     "put the counters into a counter group"),
1843         OPT_BOOLEAN('c', "scale", &stat_config.scale, "scale/normalize counters"),
1844         OPT_INCR('v', "verbose", &verbose,
1845                     "be more verbose (show counter open errors, etc)"),
1846         OPT_INTEGER('r', "repeat", &run_count,
1847                     "repeat command and print average + stddev (max: 100, forever: 0)"),
1848         OPT_BOOLEAN('n', "null", &null_run,
1849                     "null run - dont start any counters"),
1850         OPT_INCR('d', "detailed", &detailed_run,
1851                     "detailed run - start a lot of events"),
1852         OPT_BOOLEAN('S', "sync", &sync_run,
1853                     "call sync() before starting a run"),
1854         OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
1855                            "print large numbers with thousands\' separators",
1856                            stat__set_big_num),
1857         OPT_STRING('C', "cpu", &target.cpu_list, "cpu",
1858                     "list of cpus to monitor in system-wide"),
1859         OPT_SET_UINT('A', "no-aggr", &stat_config.aggr_mode,
1860                     "disable CPU count aggregation", AGGR_NONE),
1861         OPT_BOOLEAN(0, "no-merge", &no_merge, "Do not merge identical named events"),
1862         OPT_STRING('x', "field-separator", &csv_sep, "separator",
1863                    "print counts with custom separator"),
1864         OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
1865                      "monitor event in cgroup name only", parse_cgroups),
1866         OPT_STRING('o', "output", &output_name, "file", "output file name"),
1867         OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
1868         OPT_INTEGER(0, "log-fd", &output_fd,
1869                     "log output to fd, instead of stderr"),
1870         OPT_STRING(0, "pre", &pre_cmd, "command",
1871                         "command to run prior to the measured command"),
1872         OPT_STRING(0, "post", &post_cmd, "command",
1873                         "command to run after to the measured command"),
1874         OPT_UINTEGER('I', "interval-print", &stat_config.interval,
1875                     "print counts at regular interval in ms (>= 10)"),
1876         OPT_SET_UINT(0, "per-socket", &stat_config.aggr_mode,
1877                      "aggregate counts per processor socket", AGGR_SOCKET),
1878         OPT_SET_UINT(0, "per-core", &stat_config.aggr_mode,
1879                      "aggregate counts per physical processor core", AGGR_CORE),
1880         OPT_SET_UINT(0, "per-thread", &stat_config.aggr_mode,
1881                      "aggregate counts per thread", AGGR_THREAD),
1882         OPT_UINTEGER('D', "delay", &initial_delay,
1883                      "ms to wait before starting measurement after program start"),
1884         OPT_CALLBACK_NOOPT(0, "metric-only", &metric_only, NULL,
1885                         "Only print computed metrics. No raw values", enable_metric_only),
1886         OPT_BOOLEAN(0, "topdown", &topdown_run,
1887                         "measure topdown level 1 statistics"),
1888         OPT_BOOLEAN(0, "smi-cost", &smi_cost,
1889                         "measure SMI cost"),
1890         OPT_CALLBACK('M', "metrics", &evsel_list, "metric/metric group list",
1891                      "monitor specified metrics or metric groups (separated by ,)",
1892                      parse_metric_groups),
1893         OPT_END()
1894 };
1895
1896 static int perf_stat__get_socket(struct cpu_map *map, int cpu)
1897 {
1898         return cpu_map__get_socket(map, cpu, NULL);
1899 }
1900
1901 static int perf_stat__get_core(struct cpu_map *map, int cpu)
1902 {
1903         return cpu_map__get_core(map, cpu, NULL);
1904 }
1905
1906 static int cpu_map__get_max(struct cpu_map *map)
1907 {
1908         int i, max = -1;
1909
1910         for (i = 0; i < map->nr; i++) {
1911                 if (map->map[i] > max)
1912                         max = map->map[i];
1913         }
1914
1915         return max;
1916 }
1917
1918 static struct cpu_map *cpus_aggr_map;
1919
1920 static int perf_stat__get_aggr(aggr_get_id_t get_id, struct cpu_map *map, int idx)
1921 {
1922         int cpu;
1923
1924         if (idx >= map->nr)
1925                 return -1;
1926
1927         cpu = map->map[idx];
1928
1929         if (cpus_aggr_map->map[cpu] == -1)
1930                 cpus_aggr_map->map[cpu] = get_id(map, idx);
1931
1932         return cpus_aggr_map->map[cpu];
1933 }
1934
1935 static int perf_stat__get_socket_cached(struct cpu_map *map, int idx)
1936 {
1937         return perf_stat__get_aggr(perf_stat__get_socket, map, idx);
1938 }
1939
1940 static int perf_stat__get_core_cached(struct cpu_map *map, int idx)
1941 {
1942         return perf_stat__get_aggr(perf_stat__get_core, map, idx);
1943 }
1944
1945 static int perf_stat_init_aggr_mode(void)
1946 {
1947         int nr;
1948
1949         switch (stat_config.aggr_mode) {
1950         case AGGR_SOCKET:
1951                 if (cpu_map__build_socket_map(evsel_list->cpus, &aggr_map)) {
1952                         perror("cannot build socket map");
1953                         return -1;
1954                 }
1955                 aggr_get_id = perf_stat__get_socket_cached;
1956                 break;
1957         case AGGR_CORE:
1958                 if (cpu_map__build_core_map(evsel_list->cpus, &aggr_map)) {
1959                         perror("cannot build core map");
1960                         return -1;
1961                 }
1962                 aggr_get_id = perf_stat__get_core_cached;
1963                 break;
1964         case AGGR_NONE:
1965         case AGGR_GLOBAL:
1966         case AGGR_THREAD:
1967         case AGGR_UNSET:
1968         default:
1969                 break;
1970         }
1971
1972         /*
1973          * The evsel_list->cpus is the base we operate on,
1974          * taking the highest cpu number to be the size of
1975          * the aggregation translate cpumap.
1976          */
1977         nr = cpu_map__get_max(evsel_list->cpus);
1978         cpus_aggr_map = cpu_map__empty_new(nr + 1);
1979         return cpus_aggr_map ? 0 : -ENOMEM;
1980 }
1981
1982 static void perf_stat__exit_aggr_mode(void)
1983 {
1984         cpu_map__put(aggr_map);
1985         cpu_map__put(cpus_aggr_map);
1986         aggr_map = NULL;
1987         cpus_aggr_map = NULL;
1988 }
1989
1990 static inline int perf_env__get_cpu(struct perf_env *env, struct cpu_map *map, int idx)
1991 {
1992         int cpu;
1993
1994         if (idx > map->nr)
1995                 return -1;
1996
1997         cpu = map->map[idx];
1998
1999         if (cpu >= env->nr_cpus_avail)
2000                 return -1;
2001
2002         return cpu;
2003 }
2004
2005 static int perf_env__get_socket(struct cpu_map *map, int idx, void *data)
2006 {
2007         struct perf_env *env = data;
2008         int cpu = perf_env__get_cpu(env, map, idx);
2009
2010         return cpu == -1 ? -1 : env->cpu[cpu].socket_id;
2011 }
2012
2013 static int perf_env__get_core(struct cpu_map *map, int idx, void *data)
2014 {
2015         struct perf_env *env = data;
2016         int core = -1, cpu = perf_env__get_cpu(env, map, idx);
2017
2018         if (cpu != -1) {
2019                 int socket_id = env->cpu[cpu].socket_id;
2020
2021                 /*
2022                  * Encode socket in upper 16 bits
2023                  * core_id is relative to socket, and
2024                  * we need a global id. So we combine
2025                  * socket + core id.
2026                  */
2027                 core = (socket_id << 16) | (env->cpu[cpu].core_id & 0xffff);
2028         }
2029
2030         return core;
2031 }
2032
2033 static int perf_env__build_socket_map(struct perf_env *env, struct cpu_map *cpus,
2034                                       struct cpu_map **sockp)
2035 {
2036         return cpu_map__build_map(cpus, sockp, perf_env__get_socket, env);
2037 }
2038
2039 static int perf_env__build_core_map(struct perf_env *env, struct cpu_map *cpus,
2040                                     struct cpu_map **corep)
2041 {
2042         return cpu_map__build_map(cpus, corep, perf_env__get_core, env);
2043 }
2044
2045 static int perf_stat__get_socket_file(struct cpu_map *map, int idx)
2046 {
2047         return perf_env__get_socket(map, idx, &perf_stat.session->header.env);
2048 }
2049
2050 static int perf_stat__get_core_file(struct cpu_map *map, int idx)
2051 {
2052         return perf_env__get_core(map, idx, &perf_stat.session->header.env);
2053 }
2054
2055 static int perf_stat_init_aggr_mode_file(struct perf_stat *st)
2056 {
2057         struct perf_env *env = &st->session->header.env;
2058
2059         switch (stat_config.aggr_mode) {
2060         case AGGR_SOCKET:
2061                 if (perf_env__build_socket_map(env, evsel_list->cpus, &aggr_map)) {
2062                         perror("cannot build socket map");
2063                         return -1;
2064                 }
2065                 aggr_get_id = perf_stat__get_socket_file;
2066                 break;
2067         case AGGR_CORE:
2068                 if (perf_env__build_core_map(env, evsel_list->cpus, &aggr_map)) {
2069                         perror("cannot build core map");
2070                         return -1;
2071                 }
2072                 aggr_get_id = perf_stat__get_core_file;
2073                 break;
2074         case AGGR_NONE:
2075         case AGGR_GLOBAL:
2076         case AGGR_THREAD:
2077         case AGGR_UNSET:
2078         default:
2079                 break;
2080         }
2081
2082         return 0;
2083 }
2084
2085 static int topdown_filter_events(const char **attr, char **str, bool use_group)
2086 {
2087         int off = 0;
2088         int i;
2089         int len = 0;
2090         char *s;
2091
2092         for (i = 0; attr[i]; i++) {
2093                 if (pmu_have_event("cpu", attr[i])) {
2094                         len += strlen(attr[i]) + 1;
2095                         attr[i - off] = attr[i];
2096                 } else
2097                         off++;
2098         }
2099         attr[i - off] = NULL;
2100
2101         *str = malloc(len + 1 + 2);
2102         if (!*str)
2103                 return -1;
2104         s = *str;
2105         if (i - off == 0) {
2106                 *s = 0;
2107                 return 0;
2108         }
2109         if (use_group)
2110                 *s++ = '{';
2111         for (i = 0; attr[i]; i++) {
2112                 strcpy(s, attr[i]);
2113                 s += strlen(s);
2114                 *s++ = ',';
2115         }
2116         if (use_group) {
2117                 s[-1] = '}';
2118                 *s = 0;
2119         } else
2120                 s[-1] = 0;
2121         return 0;
2122 }
2123
2124 __weak bool arch_topdown_check_group(bool *warn)
2125 {
2126         *warn = false;
2127         return false;
2128 }
2129
2130 __weak void arch_topdown_group_warn(void)
2131 {
2132 }
2133
2134 /*
2135  * Add default attributes, if there were no attributes specified or
2136  * if -d/--detailed, -d -d or -d -d -d is used:
2137  */
2138 static int add_default_attributes(void)
2139 {
2140         int err;
2141         struct perf_event_attr default_attrs0[] = {
2142
2143   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK              },
2144   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES        },
2145   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS          },
2146   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS             },
2147
2148   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES              },
2149 };
2150         struct perf_event_attr frontend_attrs[] = {
2151   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND },
2152 };
2153         struct perf_event_attr backend_attrs[] = {
2154   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND  },
2155 };
2156         struct perf_event_attr default_attrs1[] = {
2157   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS            },
2158   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS     },
2159   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES           },
2160
2161 };
2162
2163 /*
2164  * Detailed stats (-d), covering the L1 and last level data caches:
2165  */
2166         struct perf_event_attr detailed_attrs[] = {
2167
2168   { .type = PERF_TYPE_HW_CACHE,
2169     .config =
2170          PERF_COUNT_HW_CACHE_L1D                <<  0  |
2171         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
2172         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
2173
2174   { .type = PERF_TYPE_HW_CACHE,
2175     .config =
2176          PERF_COUNT_HW_CACHE_L1D                <<  0  |
2177         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
2178         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
2179
2180   { .type = PERF_TYPE_HW_CACHE,
2181     .config =
2182          PERF_COUNT_HW_CACHE_LL                 <<  0  |
2183         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
2184         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
2185
2186   { .type = PERF_TYPE_HW_CACHE,
2187     .config =
2188          PERF_COUNT_HW_CACHE_LL                 <<  0  |
2189         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
2190         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
2191 };
2192
2193 /*
2194  * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
2195  */
2196         struct perf_event_attr very_detailed_attrs[] = {
2197
2198   { .type = PERF_TYPE_HW_CACHE,
2199     .config =
2200          PERF_COUNT_HW_CACHE_L1I                <<  0  |
2201         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
2202         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
2203
2204   { .type = PERF_TYPE_HW_CACHE,
2205     .config =
2206          PERF_COUNT_HW_CACHE_L1I                <<  0  |
2207         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
2208         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
2209
2210   { .type = PERF_TYPE_HW_CACHE,
2211     .config =
2212          PERF_COUNT_HW_CACHE_DTLB               <<  0  |
2213         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
2214         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
2215
2216   { .type = PERF_TYPE_HW_CACHE,
2217     .config =
2218          PERF_COUNT_HW_CACHE_DTLB               <<  0  |
2219         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
2220         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
2221
2222   { .type = PERF_TYPE_HW_CACHE,
2223     .config =
2224          PERF_COUNT_HW_CACHE_ITLB               <<  0  |
2225         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
2226         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
2227
2228   { .type = PERF_TYPE_HW_CACHE,
2229     .config =
2230          PERF_COUNT_HW_CACHE_ITLB               <<  0  |
2231         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
2232         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
2233
2234 };
2235
2236 /*
2237  * Very, very detailed stats (-d -d -d), adding prefetch events:
2238  */
2239         struct perf_event_attr very_very_detailed_attrs[] = {
2240
2241   { .type = PERF_TYPE_HW_CACHE,
2242     .config =
2243          PERF_COUNT_HW_CACHE_L1D                <<  0  |
2244         (PERF_COUNT_HW_CACHE_OP_PREFETCH        <<  8) |
2245         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
2246
2247   { .type = PERF_TYPE_HW_CACHE,
2248     .config =
2249          PERF_COUNT_HW_CACHE_L1D                <<  0  |
2250         (PERF_COUNT_HW_CACHE_OP_PREFETCH        <<  8) |
2251         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
2252 };
2253
2254         /* Set attrs if no event is selected and !null_run: */
2255         if (null_run)
2256                 return 0;
2257
2258         if (transaction_run) {
2259                 if (pmu_have_event("cpu", "cycles-ct") &&
2260                     pmu_have_event("cpu", "el-start"))
2261                         err = parse_events(evsel_list, transaction_attrs, NULL);
2262                 else
2263                         err = parse_events(evsel_list, transaction_limited_attrs, NULL);
2264                 if (err) {
2265                         fprintf(stderr, "Cannot set up transaction events\n");
2266                         return -1;
2267                 }
2268                 return 0;
2269         }
2270
2271         if (smi_cost) {
2272                 int smi;
2273
2274                 if (sysfs__read_int(FREEZE_ON_SMI_PATH, &smi) < 0) {
2275                         fprintf(stderr, "freeze_on_smi is not supported.\n");
2276                         return -1;
2277                 }
2278
2279                 if (!smi) {
2280                         if (sysfs__write_int(FREEZE_ON_SMI_PATH, 1) < 0) {
2281                                 fprintf(stderr, "Failed to set freeze_on_smi.\n");
2282                                 return -1;
2283                         }
2284                         smi_reset = true;
2285                 }
2286
2287                 if (pmu_have_event("msr", "aperf") &&
2288                     pmu_have_event("msr", "smi")) {
2289                         if (!force_metric_only)
2290                                 metric_only = true;
2291                         err = parse_events(evsel_list, smi_cost_attrs, NULL);
2292                 } else {
2293                         fprintf(stderr, "To measure SMI cost, it needs "
2294                                 "msr/aperf/, msr/smi/ and cpu/cycles/ support\n");
2295                         return -1;
2296                 }
2297                 if (err) {
2298                         fprintf(stderr, "Cannot set up SMI cost events\n");
2299                         return -1;
2300                 }
2301                 return 0;
2302         }
2303
2304         if (topdown_run) {
2305                 char *str = NULL;
2306                 bool warn = false;
2307
2308                 if (stat_config.aggr_mode != AGGR_GLOBAL &&
2309                     stat_config.aggr_mode != AGGR_CORE) {
2310                         pr_err("top down event configuration requires --per-core mode\n");
2311                         return -1;
2312                 }
2313                 stat_config.aggr_mode = AGGR_CORE;
2314                 if (nr_cgroups || !target__has_cpu(&target)) {
2315                         pr_err("top down event configuration requires system-wide mode (-a)\n");
2316                         return -1;
2317                 }
2318
2319                 if (!force_metric_only)
2320                         metric_only = true;
2321                 if (topdown_filter_events(topdown_attrs, &str,
2322                                 arch_topdown_check_group(&warn)) < 0) {
2323                         pr_err("Out of memory\n");
2324                         return -1;
2325                 }
2326                 if (topdown_attrs[0] && str) {
2327                         if (warn)
2328                                 arch_topdown_group_warn();
2329                         err = parse_events(evsel_list, str, NULL);
2330                         if (err) {
2331                                 fprintf(stderr,
2332                                         "Cannot set up top down events %s: %d\n",
2333                                         str, err);
2334                                 free(str);
2335                                 return -1;
2336                         }
2337                 } else {
2338                         fprintf(stderr, "System does not support topdown\n");
2339                         return -1;
2340                 }
2341                 free(str);
2342         }
2343
2344         if (!evsel_list->nr_entries) {
2345                 if (target__has_cpu(&target))
2346                         default_attrs0[0].config = PERF_COUNT_SW_CPU_CLOCK;
2347
2348                 if (perf_evlist__add_default_attrs(evsel_list, default_attrs0) < 0)
2349                         return -1;
2350                 if (pmu_have_event("cpu", "stalled-cycles-frontend")) {
2351                         if (perf_evlist__add_default_attrs(evsel_list,
2352                                                 frontend_attrs) < 0)
2353                                 return -1;
2354                 }
2355                 if (pmu_have_event("cpu", "stalled-cycles-backend")) {
2356                         if (perf_evlist__add_default_attrs(evsel_list,
2357                                                 backend_attrs) < 0)
2358                                 return -1;
2359                 }
2360                 if (perf_evlist__add_default_attrs(evsel_list, default_attrs1) < 0)
2361                         return -1;
2362         }
2363
2364         /* Detailed events get appended to the event list: */
2365
2366         if (detailed_run <  1)
2367                 return 0;
2368
2369         /* Append detailed run extra attributes: */
2370         if (perf_evlist__add_default_attrs(evsel_list, detailed_attrs) < 0)
2371                 return -1;
2372
2373         if (detailed_run < 2)
2374                 return 0;
2375
2376         /* Append very detailed run extra attributes: */
2377         if (perf_evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0)
2378                 return -1;
2379
2380         if (detailed_run < 3)
2381                 return 0;
2382
2383         /* Append very, very detailed run extra attributes: */
2384         return perf_evlist__add_default_attrs(evsel_list, very_very_detailed_attrs);
2385 }
2386
2387 static const char * const stat_record_usage[] = {
2388         "perf stat record [<options>]",
2389         NULL,
2390 };
2391
2392 static void init_features(struct perf_session *session)
2393 {
2394         int feat;
2395
2396         for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++)
2397                 perf_header__set_feat(&session->header, feat);
2398
2399         perf_header__clear_feat(&session->header, HEADER_BUILD_ID);
2400         perf_header__clear_feat(&session->header, HEADER_TRACING_DATA);
2401         perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK);
2402         perf_header__clear_feat(&session->header, HEADER_AUXTRACE);
2403 }
2404
2405 static int __cmd_record(int argc, const char **argv)
2406 {
2407         struct perf_session *session;
2408         struct perf_data *data = &perf_stat.data;
2409
2410         argc = parse_options(argc, argv, stat_options, stat_record_usage,
2411                              PARSE_OPT_STOP_AT_NON_OPTION);
2412
2413         if (output_name)
2414                 data->file.path = output_name;
2415
2416         if (run_count != 1 || forever) {
2417                 pr_err("Cannot use -r option with perf stat record.\n");
2418                 return -1;
2419         }
2420
2421         session = perf_session__new(data, false, NULL);
2422         if (session == NULL) {
2423                 pr_err("Perf session creation failed.\n");
2424                 return -1;
2425         }
2426
2427         init_features(session);
2428
2429         session->evlist   = evsel_list;
2430         perf_stat.session = session;
2431         perf_stat.record  = true;
2432         return argc;
2433 }
2434
2435 static int process_stat_round_event(struct perf_tool *tool __maybe_unused,
2436                                     union perf_event *event,
2437                                     struct perf_session *session)
2438 {
2439         struct stat_round_event *stat_round = &event->stat_round;
2440         struct perf_evsel *counter;
2441         struct timespec tsh, *ts = NULL;
2442         const char **argv = session->header.env.cmdline_argv;
2443         int argc = session->header.env.nr_cmdline;
2444
2445         evlist__for_each_entry(evsel_list, counter)
2446                 perf_stat_process_counter(&stat_config, counter);
2447
2448         if (stat_round->type == PERF_STAT_ROUND_TYPE__FINAL)
2449                 update_stats(&walltime_nsecs_stats, stat_round->time);
2450
2451         if (stat_config.interval && stat_round->time) {
2452                 tsh.tv_sec  = stat_round->time / NSEC_PER_SEC;
2453                 tsh.tv_nsec = stat_round->time % NSEC_PER_SEC;
2454                 ts = &tsh;
2455         }
2456
2457         print_counters(ts, argc, argv);
2458         return 0;
2459 }
2460
2461 static
2462 int process_stat_config_event(struct perf_tool *tool,
2463                               union perf_event *event,
2464                               struct perf_session *session __maybe_unused)
2465 {
2466         struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2467
2468         perf_event__read_stat_config(&stat_config, &event->stat_config);
2469
2470         if (cpu_map__empty(st->cpus)) {
2471                 if (st->aggr_mode != AGGR_UNSET)
2472                         pr_warning("warning: processing task data, aggregation mode not set\n");
2473                 return 0;
2474         }
2475
2476         if (st->aggr_mode != AGGR_UNSET)
2477                 stat_config.aggr_mode = st->aggr_mode;
2478
2479         if (perf_stat.data.is_pipe)
2480                 perf_stat_init_aggr_mode();
2481         else
2482                 perf_stat_init_aggr_mode_file(st);
2483
2484         return 0;
2485 }
2486
2487 static int set_maps(struct perf_stat *st)
2488 {
2489         if (!st->cpus || !st->threads)
2490                 return 0;
2491
2492         if (WARN_ONCE(st->maps_allocated, "stats double allocation\n"))
2493                 return -EINVAL;
2494
2495         perf_evlist__set_maps(evsel_list, st->cpus, st->threads);
2496
2497         if (perf_evlist__alloc_stats(evsel_list, true))
2498                 return -ENOMEM;
2499
2500         st->maps_allocated = true;
2501         return 0;
2502 }
2503
2504 static
2505 int process_thread_map_event(struct perf_tool *tool,
2506                              union perf_event *event,
2507                              struct perf_session *session __maybe_unused)
2508 {
2509         struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2510
2511         if (st->threads) {
2512                 pr_warning("Extra thread map event, ignoring.\n");
2513                 return 0;
2514         }
2515
2516         st->threads = thread_map__new_event(&event->thread_map);
2517         if (!st->threads)
2518                 return -ENOMEM;
2519
2520         return set_maps(st);
2521 }
2522
2523 static
2524 int process_cpu_map_event(struct perf_tool *tool,
2525                           union perf_event *event,
2526                           struct perf_session *session __maybe_unused)
2527 {
2528         struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2529         struct cpu_map *cpus;
2530
2531         if (st->cpus) {
2532                 pr_warning("Extra cpu map event, ignoring.\n");
2533                 return 0;
2534         }
2535
2536         cpus = cpu_map__new_data(&event->cpu_map.data);
2537         if (!cpus)
2538                 return -ENOMEM;
2539
2540         st->cpus = cpus;
2541         return set_maps(st);
2542 }
2543
2544 static const char * const stat_report_usage[] = {
2545         "perf stat report [<options>]",
2546         NULL,
2547 };
2548
2549 static struct perf_stat perf_stat = {
2550         .tool = {
2551                 .attr           = perf_event__process_attr,
2552                 .event_update   = perf_event__process_event_update,
2553                 .thread_map     = process_thread_map_event,
2554                 .cpu_map        = process_cpu_map_event,
2555                 .stat_config    = process_stat_config_event,
2556                 .stat           = perf_event__process_stat_event,
2557                 .stat_round     = process_stat_round_event,
2558         },
2559         .aggr_mode = AGGR_UNSET,
2560 };
2561
2562 static int __cmd_report(int argc, const char **argv)
2563 {
2564         struct perf_session *session;
2565         const struct option options[] = {
2566         OPT_STRING('i', "input", &input_name, "file", "input file name"),
2567         OPT_SET_UINT(0, "per-socket", &perf_stat.aggr_mode,
2568                      "aggregate counts per processor socket", AGGR_SOCKET),
2569         OPT_SET_UINT(0, "per-core", &perf_stat.aggr_mode,
2570                      "aggregate counts per physical processor core", AGGR_CORE),
2571         OPT_SET_UINT('A', "no-aggr", &perf_stat.aggr_mode,
2572                      "disable CPU count aggregation", AGGR_NONE),
2573         OPT_END()
2574         };
2575         struct stat st;
2576         int ret;
2577
2578         argc = parse_options(argc, argv, options, stat_report_usage, 0);
2579
2580         if (!input_name || !strlen(input_name)) {
2581                 if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode))
2582                         input_name = "-";
2583                 else
2584                         input_name = "perf.data";
2585         }
2586
2587         perf_stat.data.file.path = input_name;
2588         perf_stat.data.mode      = PERF_DATA_MODE_READ;
2589
2590         session = perf_session__new(&perf_stat.data, false, &perf_stat.tool);
2591         if (session == NULL)
2592                 return -1;
2593
2594         perf_stat.session  = session;
2595         stat_config.output = stderr;
2596         evsel_list         = session->evlist;
2597
2598         ret = perf_session__process_events(session);
2599         if (ret)
2600                 return ret;
2601
2602         perf_session__delete(session);
2603         return 0;
2604 }
2605
2606 static void setup_system_wide(int forks)
2607 {
2608         /*
2609          * Make system wide (-a) the default target if
2610          * no target was specified and one of following
2611          * conditions is met:
2612          *
2613          *   - there's no workload specified
2614          *   - there is workload specified but all requested
2615          *     events are system wide events
2616          */
2617         if (!target__none(&target))
2618                 return;
2619
2620         if (!forks)
2621                 target.system_wide = true;
2622         else {
2623                 struct perf_evsel *counter;
2624
2625                 evlist__for_each_entry(evsel_list, counter) {
2626                         if (!counter->system_wide)
2627                                 return;
2628                 }
2629
2630                 if (evsel_list->nr_entries)
2631                         target.system_wide = true;
2632         }
2633 }
2634
2635 int cmd_stat(int argc, const char **argv)
2636 {
2637         const char * const stat_usage[] = {
2638                 "perf stat [<options>] [<command>]",
2639                 NULL
2640         };
2641         int status = -EINVAL, run_idx;
2642         const char *mode;
2643         FILE *output = stderr;
2644         unsigned int interval;
2645         const char * const stat_subcommands[] = { "record", "report" };
2646
2647         setlocale(LC_ALL, "");
2648
2649         evsel_list = perf_evlist__new();
2650         if (evsel_list == NULL)
2651                 return -ENOMEM;
2652
2653         parse_events__shrink_config_terms();
2654         argc = parse_options_subcommand(argc, argv, stat_options, stat_subcommands,
2655                                         (const char **) stat_usage,
2656                                         PARSE_OPT_STOP_AT_NON_OPTION);
2657         perf_stat__collect_metric_expr(evsel_list);
2658         perf_stat__init_shadow_stats();
2659
2660         if (csv_sep) {
2661                 csv_output = true;
2662                 if (!strcmp(csv_sep, "\\t"))
2663                         csv_sep = "\t";
2664         } else
2665                 csv_sep = DEFAULT_SEPARATOR;
2666
2667         if (argc && !strncmp(argv[0], "rec", 3)) {
2668                 argc = __cmd_record(argc, argv);
2669                 if (argc < 0)
2670                         return -1;
2671         } else if (argc && !strncmp(argv[0], "rep", 3))
2672                 return __cmd_report(argc, argv);
2673
2674         interval = stat_config.interval;
2675
2676         /*
2677          * For record command the -o is already taken care of.
2678          */
2679         if (!STAT_RECORD && output_name && strcmp(output_name, "-"))
2680                 output = NULL;
2681
2682         if (output_name && output_fd) {
2683                 fprintf(stderr, "cannot use both --output and --log-fd\n");
2684                 parse_options_usage(stat_usage, stat_options, "o", 1);
2685                 parse_options_usage(NULL, stat_options, "log-fd", 0);
2686                 goto out;
2687         }
2688
2689         if (metric_only && stat_config.aggr_mode == AGGR_THREAD) {
2690                 fprintf(stderr, "--metric-only is not supported with --per-thread\n");
2691                 goto out;
2692         }
2693
2694         if (metric_only && run_count > 1) {
2695                 fprintf(stderr, "--metric-only is not supported with -r\n");
2696                 goto out;
2697         }
2698
2699         if (output_fd < 0) {
2700                 fprintf(stderr, "argument to --log-fd must be a > 0\n");
2701                 parse_options_usage(stat_usage, stat_options, "log-fd", 0);
2702                 goto out;
2703         }
2704
2705         if (!output) {
2706                 struct timespec tm;
2707                 mode = append_file ? "a" : "w";
2708
2709                 output = fopen(output_name, mode);
2710                 if (!output) {
2711                         perror("failed to create output file");
2712                         return -1;
2713                 }
2714                 clock_gettime(CLOCK_REALTIME, &tm);
2715                 fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
2716         } else if (output_fd > 0) {
2717                 mode = append_file ? "a" : "w";
2718                 output = fdopen(output_fd, mode);
2719                 if (!output) {
2720                         perror("Failed opening logfd");
2721                         return -errno;
2722                 }
2723         }
2724
2725         stat_config.output = output;
2726
2727         /*
2728          * let the spreadsheet do the pretty-printing
2729          */
2730         if (csv_output) {
2731                 /* User explicitly passed -B? */
2732                 if (big_num_opt == 1) {
2733                         fprintf(stderr, "-B option not supported with -x\n");
2734                         parse_options_usage(stat_usage, stat_options, "B", 1);
2735                         parse_options_usage(NULL, stat_options, "x", 1);
2736                         goto out;
2737                 } else /* Nope, so disable big number formatting */
2738                         big_num = false;
2739         } else if (big_num_opt == 0) /* User passed --no-big-num */
2740                 big_num = false;
2741
2742         setup_system_wide(argc);
2743
2744         if (run_count < 0) {
2745                 pr_err("Run count must be a positive number\n");
2746                 parse_options_usage(stat_usage, stat_options, "r", 1);
2747                 goto out;
2748         } else if (run_count == 0) {
2749                 forever = true;
2750                 run_count = 1;
2751         }
2752
2753         if ((stat_config.aggr_mode == AGGR_THREAD) && !target__has_task(&target)) {
2754                 fprintf(stderr, "The --per-thread option is only available "
2755                         "when monitoring via -p -t options.\n");
2756                 parse_options_usage(NULL, stat_options, "p", 1);
2757                 parse_options_usage(NULL, stat_options, "t", 1);
2758                 goto out;
2759         }
2760
2761         /*
2762          * no_aggr, cgroup are for system-wide only
2763          * --per-thread is aggregated per thread, we dont mix it with cpu mode
2764          */
2765         if (((stat_config.aggr_mode != AGGR_GLOBAL &&
2766               stat_config.aggr_mode != AGGR_THREAD) || nr_cgroups) &&
2767             !target__has_cpu(&target)) {
2768                 fprintf(stderr, "both cgroup and no-aggregation "
2769                         "modes only available in system-wide mode\n");
2770
2771                 parse_options_usage(stat_usage, stat_options, "G", 1);
2772                 parse_options_usage(NULL, stat_options, "A", 1);
2773                 parse_options_usage(NULL, stat_options, "a", 1);
2774                 goto out;
2775         }
2776
2777         if (add_default_attributes())
2778                 goto out;
2779
2780         target__validate(&target);
2781
2782         if (perf_evlist__create_maps(evsel_list, &target) < 0) {
2783                 if (target__has_task(&target)) {
2784                         pr_err("Problems finding threads of monitor\n");
2785                         parse_options_usage(stat_usage, stat_options, "p", 1);
2786                         parse_options_usage(NULL, stat_options, "t", 1);
2787                 } else if (target__has_cpu(&target)) {
2788                         perror("failed to parse CPUs map");
2789                         parse_options_usage(stat_usage, stat_options, "C", 1);
2790                         parse_options_usage(NULL, stat_options, "a", 1);
2791                 }
2792                 goto out;
2793         }
2794
2795         /*
2796          * Initialize thread_map with comm names,
2797          * so we could print it out on output.
2798          */
2799         if (stat_config.aggr_mode == AGGR_THREAD)
2800                 thread_map__read_comms(evsel_list->threads);
2801
2802         if (interval && interval < 100) {
2803                 if (interval < 10) {
2804                         pr_err("print interval must be >= 10ms\n");
2805                         parse_options_usage(stat_usage, stat_options, "I", 1);
2806                         goto out;
2807                 } else
2808                         pr_warning("print interval < 100ms. "
2809                                    "The overhead percentage could be high in some cases. "
2810                                    "Please proceed with caution.\n");
2811         }
2812
2813         if (perf_evlist__alloc_stats(evsel_list, interval))
2814                 goto out;
2815
2816         if (perf_stat_init_aggr_mode())
2817                 goto out;
2818
2819         /*
2820          * We dont want to block the signals - that would cause
2821          * child tasks to inherit that and Ctrl-C would not work.
2822          * What we want is for Ctrl-C to work in the exec()-ed
2823          * task, but being ignored by perf stat itself:
2824          */
2825         atexit(sig_atexit);
2826         if (!forever)
2827                 signal(SIGINT,  skip_signal);
2828         signal(SIGCHLD, skip_signal);
2829         signal(SIGALRM, skip_signal);
2830         signal(SIGABRT, skip_signal);
2831
2832         status = 0;
2833         for (run_idx = 0; forever || run_idx < run_count; run_idx++) {
2834                 if (run_count != 1 && verbose > 0)
2835                         fprintf(output, "[ perf stat: executing run #%d ... ]\n",
2836                                 run_idx + 1);
2837
2838                 status = run_perf_stat(argc, argv);
2839                 if (forever && status != -1) {
2840                         print_counters(NULL, argc, argv);
2841                         perf_stat__reset_stats();
2842                 }
2843         }
2844
2845         if (!forever && status != -1 && !interval)
2846                 print_counters(NULL, argc, argv);
2847
2848         if (STAT_RECORD) {
2849                 /*
2850                  * We synthesize the kernel mmap record just so that older tools
2851                  * don't emit warnings about not being able to resolve symbols
2852                  * due to /proc/sys/kernel/kptr_restrict settings and instear provide
2853                  * a saner message about no samples being in the perf.data file.
2854                  *
2855                  * This also serves to suppress a warning about f_header.data.size == 0
2856                  * in header.c at the moment 'perf stat record' gets introduced, which
2857                  * is not really needed once we start adding the stat specific PERF_RECORD_
2858                  * records, but the need to suppress the kptr_restrict messages in older
2859                  * tools remain  -acme
2860                  */
2861                 int fd = perf_data__fd(&perf_stat.data);
2862                 int err = perf_event__synthesize_kernel_mmap((void *)&perf_stat,
2863                                                              process_synthesized_event,
2864                                                              &perf_stat.session->machines.host);
2865                 if (err) {
2866                         pr_warning("Couldn't synthesize the kernel mmap record, harmless, "
2867                                    "older tools may produce warnings about this file\n.");
2868                 }
2869
2870                 if (!interval) {
2871                         if (WRITE_STAT_ROUND_EVENT(walltime_nsecs_stats.max, FINAL))
2872                                 pr_err("failed to write stat round event\n");
2873                 }
2874
2875                 if (!perf_stat.data.is_pipe) {
2876                         perf_stat.session->header.data_size += perf_stat.bytes_written;
2877                         perf_session__write_header(perf_stat.session, evsel_list, fd, true);
2878                 }
2879
2880                 perf_session__delete(perf_stat.session);
2881         }
2882
2883         perf_stat__exit_aggr_mode();
2884         perf_evlist__free_stats(evsel_list);
2885 out:
2886         if (smi_cost && smi_reset)
2887                 sysfs__write_int(FREEZE_ON_SMI_PATH, 0);
2888
2889         perf_evlist__delete(evsel_list);
2890         return status;
2891 }