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/string2.h"
  67 #include "util/metricgroup.h"
  68 #include "asm/bug.h"
  69
  70 #include <linux/time64.h>
  71 #include <api/fs/fs.h>
  72 #include <errno.h>
  73 #include <signal.h>
  74 #include <stdlib.h>
  75 #include <sys/prctl.h>
  76 #include <inttypes.h>
  77 #include <locale.h>
  78 #include <math.h>
  79 #include <sys/types.h>
  80 #include <sys/stat.h>
  81 #include <sys/wait.h>
  82 #include <unistd.h>
  83 #include <sys/time.h>
  84 #include <sys/resource.h>
  85 #include <sys/wait.h>
  86
  87 #include "sane_ctype.h"
  88
  89 #define DEFAULT_SEPARATOR       " "
  90 #define CNTR_NOT_SUPPORTED      "<not supported>"
  91 #define CNTR_NOT_COUNTED        "<not counted>"
  92 #define FREEZE_ON_SMI_PATH      "devices/cpu/freeze_on_smi"
  93
  94 static void print_counters(struct timespec *ts, int argc, const char **argv);
  95
  96 /* Default events used for perf stat -T */
  97 static const char *transaction_attrs = {
  98         "task-clock,"
  99         "{"
 100         "instructions,"
 101         "cycles,"
 102         "cpu/cycles-t/,"
 103         "cpu/tx-start/,"
 104         "cpu/el-start/,"
 105         "cpu/cycles-ct/"
 106         "}"
 107 };
 108
 109 /* More limited version when the CPU does not have all events. */
 110 static const char * transaction_limited_attrs = {
 111         "task-clock,"
 112         "{"
 113         "instructions,"
 114         "cycles,"
 115         "cpu/cycles-t/,"
 116         "cpu/tx-start/"
 117         "}"
 118 };
 119
 120 static const char * topdown_attrs[] = {
 121         "topdown-total-slots",
 122         "topdown-slots-retired",
 123         "topdown-recovery-bubbles",
 124         "topdown-fetch-bubbles",
 125         "topdown-slots-issued",
 126         NULL,
 127 };
 128
 129 static const char *smi_cost_attrs = {
 130         "{"
 131         "msr/aperf/,"
 132         "msr/smi/,"
 133         "cycles"
 134         "}"
 135 };
 136
 137 static struct perf_evlist       *evsel_list;
 138
 139 static struct rblist             metric_events;
 140
 141 static struct target target = {
 142         .uid    = UINT_MAX,
 143 };
 144
 145 typedef int (*aggr_get_id_t)(struct cpu_map *m, int cpu);
 146
 147 static int                      run_count                       =  1;
 148 static bool                     no_inherit                      = false;
 149 static volatile pid_t           child_pid                       = -1;
 150 static bool                     null_run                        =  false;
 151 static int                      detailed_run                    =  0;
 152 static bool                     transaction_run;
 153 static bool                     topdown_run                     = false;
 154 static bool                     smi_cost                        = false;
 155 static bool                     smi_reset                       = false;
 156 static bool                     big_num                         =  true;
 157 static int                      big_num_opt                     =  -1;
 158 static const char               *csv_sep                        = NULL;
 159 static bool                     csv_output                      = false;
 160 static bool                     group                           = false;
 161 static const char               *pre_cmd                        = NULL;
 162 static const char               *post_cmd                       = NULL;
 163 static bool                     sync_run                        = false;
 164 static unsigned int             initial_delay                   = 0;
 165 static unsigned int             unit_width                      = 4; /* strlen("unit") */
 166 static bool                     forever                         = false;
 167 static bool                     metric_only                     = false;
 168 static bool                     force_metric_only               = false;
 169 static bool                     no_merge                        = false;
 170 static bool                     walltime_run_table              = false;
 171 static struct timespec          ref_time;
 172 static struct cpu_map           *aggr_map;
 173 static aggr_get_id_t            aggr_get_id;
 174 static bool                     append_file;
 175 static bool                     interval_count;
 176 static const char               *output_name;
 177 static int                      output_fd;
 178 static int                      print_free_counters_hint;
 179 static int                      print_mixed_hw_group_error;
 180 static u64                      *walltime_run;
 181 static bool                     ru_display                      = false;
 182 static struct rusage            ru_data;
 183
 184 struct perf_stat {
 185         bool                     record;
 186         struct perf_data         data;
 187         struct perf_session     *session;
 188         u64                      bytes_written;
 189         struct perf_tool         tool;
 190         bool                     maps_allocated;
 191         struct cpu_map          *cpus;
 192         struct thread_map       *threads;
 193         enum aggr_mode           aggr_mode;
 194 };
 195
 196 static struct perf_stat         perf_stat;
 197 #define STAT_RECORD             perf_stat.record
 198
 199 static volatile int done = 0;
 200
 201 static struct perf_stat_config stat_config = {
 202         .aggr_mode      = AGGR_GLOBAL,
 203         .scale          = true,
 204 };
 205
 206 static bool is_duration_time(struct perf_evsel *evsel)
 207 {
 208         return !strcmp(evsel->name, "duration_time");
 209 }
 210
 211 static inline void diff_timespec(struct timespec *r, struct timespec *a,
 212                                  struct timespec *b)
 213 {
 214         r->tv_sec = a->tv_sec - b->tv_sec;
 215         if (a->tv_nsec < b->tv_nsec) {
 216                 r->tv_nsec = a->tv_nsec + NSEC_PER_SEC - b->tv_nsec;
 217                 r->tv_sec--;
 218         } else {
 219                 r->tv_nsec = a->tv_nsec - b->tv_nsec ;
 220         }
 221 }
 222
 223 static void perf_stat__reset_stats(void)
 224 {
 225         int i;
 226
 227         perf_evlist__reset_stats(evsel_list);
 228         perf_stat__reset_shadow_stats();
 229
 230         for (i = 0; i < stat_config.stats_num; i++)
 231                 perf_stat__reset_shadow_per_stat(&stat_config.stats[i]);
 232 }
 233
 234 static int create_perf_stat_counter(struct perf_evsel *evsel)
 235 {
 236         struct perf_event_attr *attr = &evsel->attr;
 237         struct perf_evsel *leader = evsel->leader;
 238
 239         if (stat_config.scale) {
 240                 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
 241                                     PERF_FORMAT_TOTAL_TIME_RUNNING;
 242         }
 243
 244         /*
 245          * The event is part of non trivial group, let's enable
 246          * the group read (for leader) and ID retrieval for all
 247          * members.
 248          */
 249         if (leader->nr_members > 1)
 250                 attr->read_format |= PERF_FORMAT_ID|PERF_FORMAT_GROUP;
 251
 252         attr->inherit = !no_inherit;
 253
 254         /*
 255          * Some events get initialized with sample_(period/type) set,
 256          * like tracepoints. Clear it up for counting.
 257          */
 258         attr->sample_period = 0;
 259
 260         /*
 261          * But set sample_type to PERF_SAMPLE_IDENTIFIER, which should be harmless
 262          * while avoiding that older tools show confusing messages.
 263          *
 264          * However for pipe sessions we need to keep it zero,
 265          * because script's perf_evsel__check_attr is triggered
 266          * by attr->sample_type != 0, and we can't run it on
 267          * stat sessions.
 268          */
 269         if (!(STAT_RECORD && perf_stat.data.is_pipe))
 270                 attr->sample_type = PERF_SAMPLE_IDENTIFIER;
 271
 272         /*
 273          * Disabling all counters initially, they will be enabled
 274          * either manually by us or by kernel via enable_on_exec
 275          * set later.
 276          */
 277         if (perf_evsel__is_group_leader(evsel)) {
 278                 attr->disabled = 1;
 279
 280                 /*
 281                  * In case of initial_delay we enable tracee
 282                  * events manually.
 283                  */
 284                 if (target__none(&target) && !initial_delay)
 285                         attr->enable_on_exec = 1;
 286         }
 287
 288         if (target__has_cpu(&target) && !target__has_per_thread(&target))
 289                 return perf_evsel__open_per_cpu(evsel, perf_evsel__cpus(evsel));
 290
 291         return perf_evsel__open_per_thread(evsel, evsel_list->threads);
 292 }
 293
 294 /*
 295  * Does the counter have nsecs as a unit?
 296  */
 297 static inline int nsec_counter(struct perf_evsel *evsel)
 298 {
 299         if (perf_evsel__match(evsel, SOFTWARE, SW_CPU_CLOCK) ||
 300             perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
 301                 return 1;
 302
 303         return 0;
 304 }
 305
 306 static int process_synthesized_event(struct perf_tool *tool __maybe_unused,
 307                                      union perf_event *event,
 308                                      struct perf_sample *sample __maybe_unused,
 309                                      struct machine *machine __maybe_unused)
 310 {
 311         if (perf_data__write(&perf_stat.data, event, event->header.size) < 0) {
 312                 pr_err("failed to write perf data, error: %m\n");
 313                 return -1;
 314         }
 315
 316         perf_stat.bytes_written += event->header.size;
 317         return 0;
 318 }
 319
 320 static int write_stat_round_event(u64 tm, u64 type)
 321 {
 322         return perf_event__synthesize_stat_round(NULL, tm, type,
 323                                                  process_synthesized_event,
 324                                                  NULL);
 325 }
 326
 327 #define WRITE_STAT_ROUND_EVENT(time, interval) \
 328         write_stat_round_event(time, PERF_STAT_ROUND_TYPE__ ## interval)
 329
 330 #define SID(e, x, y) xyarray__entry(e->sample_id, x, y)
 331
 332 static int
 333 perf_evsel__write_stat_event(struct perf_evsel *counter, u32 cpu, u32 thread,
 334                              struct perf_counts_values *count)
 335 {
 336         struct perf_sample_id *sid = SID(counter, cpu, thread);
 337
 338         return perf_event__synthesize_stat(NULL, cpu, thread, sid->id, count,
 339                                            process_synthesized_event, NULL);
 340 }
 341
 342 /*
 343  * Read out the results of a single counter:
 344  * do not aggregate counts across CPUs in system-wide mode
 345  */
 346 static int read_counter(struct perf_evsel *counter)
 347 {
 348         int nthreads = thread_map__nr(evsel_list->threads);
 349         int ncpus, cpu, thread;
 350
 351         if (target__has_cpu(&target) && !target__has_per_thread(&target))
 352                 ncpus = perf_evsel__nr_cpus(counter);
 353         else
 354                 ncpus = 1;
 355
 356         if (!counter->supported)
 357                 return -ENOENT;
 358
 359         if (counter->system_wide)
 360                 nthreads = 1;
 361
 362         for (thread = 0; thread < nthreads; thread++) {
 363                 for (cpu = 0; cpu < ncpus; cpu++) {
 364                         struct perf_counts_values *count;
 365
 366                         count = perf_counts(counter->counts, cpu, thread);
 367
 368                         /*
 369                          * The leader's group read loads data into its group members
 370                          * (via perf_evsel__read_counter) and sets threir count->loaded.
 371                          */
 372                         if (!count->loaded &&
 373                             perf_evsel__read_counter(counter, cpu, thread)) {
 374                                 counter->counts->scaled = -1;
 375                                 perf_counts(counter->counts, cpu, thread)->ena = 0;
 376                                 perf_counts(counter->counts, cpu, thread)->run = 0;
 377                                 return -1;
 378                         }
 379
 380                         count->loaded = false;
 381
 382                         if (STAT_RECORD) {
 383                                 if (perf_evsel__write_stat_event(counter, cpu, thread, count)) {
 384                                         pr_err("failed to write stat event\n");
 385                                         return -1;
 386                                 }
 387                         }
 388
 389                         if (verbose > 1) {
 390                                 fprintf(stat_config.output,
 391                                         "%s: %d: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
 392                                                 perf_evsel__name(counter),
 393                                                 cpu,
 394                                                 count->val, count->ena, count->run);
 395                         }
 396                 }
 397         }
 398
 399         return 0;
 400 }
 401
 402 static void read_counters(void)
 403 {
 404         struct perf_evsel *counter;
 405         int ret;
 406
 407         evlist__for_each_entry(evsel_list, counter) {
 408                 ret = read_counter(counter);
 409                 if (ret)
 410                         pr_debug("failed to read counter %s\n", counter->name);
 411
 412                 if (ret == 0 && perf_stat_process_counter(&stat_config, counter))
 413                         pr_warning("failed to process counter %s\n", counter->name);
 414         }
 415 }
 416
 417 static void process_interval(void)
 418 {
 419         struct timespec ts, rs;
 420
 421         read_counters();
 422
 423         clock_gettime(CLOCK_MONOTONIC, &ts);
 424         diff_timespec(&rs, &ts, &ref_time);
 425
 426         if (STAT_RECORD) {
 427                 if (WRITE_STAT_ROUND_EVENT(rs.tv_sec * NSEC_PER_SEC + rs.tv_nsec, INTERVAL))
 428                         pr_err("failed to write stat round event\n");
 429         }
 430
 431         init_stats(&walltime_nsecs_stats);
 432         update_stats(&walltime_nsecs_stats, stat_config.interval * 1000000);
 433         print_counters(&rs, 0, NULL);
 434 }
 435
 436 static void enable_counters(void)
 437 {
 438         if (initial_delay)
 439                 usleep(initial_delay * USEC_PER_MSEC);
 440
 441         /*
 442          * We need to enable counters only if:
 443          * - we don't have tracee (attaching to task or cpu)
 444          * - we have initial delay configured
 445          */
 446         if (!target__none(&target) || initial_delay)
 447                 perf_evlist__enable(evsel_list);
 448 }
 449
 450 static void disable_counters(void)
 451 {
 452         /*
 453          * If we don't have tracee (attaching to task or cpu), counters may
 454          * still be running. To get accurate group ratios, we must stop groups
 455          * from counting before reading their constituent counters.
 456          */
 457         if (!target__none(&target))
 458                 perf_evlist__disable(evsel_list);
 459 }
 460
 461 static volatile int workload_exec_errno;
 462
 463 /*
 464  * perf_evlist__prepare_workload will send a SIGUSR1
 465  * if the fork fails, since we asked by setting its
 466  * want_signal to true.
 467  */
 468 static void workload_exec_failed_signal(int signo __maybe_unused, siginfo_t *info,
 469                                         void *ucontext __maybe_unused)
 470 {
 471         workload_exec_errno = info->si_value.sival_int;
 472 }
 473
 474 static int perf_stat_synthesize_config(bool is_pipe)
 475 {
 476         int err;
 477
 478         if (is_pipe) {
 479                 err = perf_event__synthesize_attrs(NULL, perf_stat.session,
 480                                                    process_synthesized_event);
 481                 if (err < 0) {
 482                         pr_err("Couldn't synthesize attrs.\n");
 483                         return err;
 484                 }
 485         }
 486
 487         err = perf_event__synthesize_extra_attr(NULL,
 488                                                 evsel_list,
 489                                                 process_synthesized_event,
 490                                                 is_pipe);
 491
 492         err = perf_event__synthesize_thread_map2(NULL, evsel_list->threads,
 493                                                 process_synthesized_event,
 494                                                 NULL);
 495         if (err < 0) {
 496                 pr_err("Couldn't synthesize thread map.\n");
 497                 return err;
 498         }
 499
 500         err = perf_event__synthesize_cpu_map(NULL, evsel_list->cpus,
 501                                              process_synthesized_event, NULL);
 502         if (err < 0) {
 503                 pr_err("Couldn't synthesize thread map.\n");
 504                 return err;
 505         }
 506
 507         err = perf_event__synthesize_stat_config(NULL, &stat_config,
 508                                                  process_synthesized_event, NULL);
 509         if (err < 0) {
 510                 pr_err("Couldn't synthesize config.\n");
 511                 return err;
 512         }
 513
 514         return 0;
 515 }
 516
 517 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
 518
 519 static int __store_counter_ids(struct perf_evsel *counter)
 520 {
 521         int cpu, thread;
 522
 523         for (cpu = 0; cpu < xyarray__max_x(counter->fd); cpu++) {
 524                 for (thread = 0; thread < xyarray__max_y(counter->fd);
 525                      thread++) {
 526                         int fd = FD(counter, cpu, thread);
 527
 528                         if (perf_evlist__id_add_fd(evsel_list, counter,
 529                                                    cpu, thread, fd) < 0)
 530                                 return -1;
 531                 }
 532         }
 533
 534         return 0;
 535 }
 536
 537 static int store_counter_ids(struct perf_evsel *counter)
 538 {
 539         struct cpu_map *cpus = counter->cpus;
 540         struct thread_map *threads = counter->threads;
 541
 542         if (perf_evsel__alloc_id(counter, cpus->nr, threads->nr))
 543                 return -ENOMEM;
 544
 545         return __store_counter_ids(counter);
 546 }
 547
 548 static bool perf_evsel__should_store_id(struct perf_evsel *counter)
 549 {
 550         return STAT_RECORD || counter->attr.read_format & PERF_FORMAT_ID;
 551 }
 552
 553 static struct perf_evsel *perf_evsel__reset_weak_group(struct perf_evsel *evsel)
 554 {
 555         struct perf_evsel *c2, *leader;
 556         bool is_open = true;
 557
 558         leader = evsel->leader;
 559         pr_debug("Weak group for %s/%d failed\n",
 560                         leader->name, leader->nr_members);
 561
 562         /*
 563          * for_each_group_member doesn't work here because it doesn't
 564          * include the first entry.
 565          */
 566         evlist__for_each_entry(evsel_list, c2) {
 567                 if (c2 == evsel)
 568                         is_open = false;
 569                 if (c2->leader == leader) {
 570                         if (is_open)
 571                                 perf_evsel__close(c2);
 572                         c2->leader = c2;
 573                         c2->nr_members = 0;
 574                 }
 575         }
 576         return leader;
 577 }
 578
 579 static int __run_perf_stat(int argc, const char **argv, int run_idx)
 580 {
 581         int interval = stat_config.interval;
 582         int times = stat_config.times;
 583         int timeout = stat_config.timeout;
 584         char msg[BUFSIZ];
 585         unsigned long long t0, t1;
 586         struct perf_evsel *counter;
 587         struct timespec ts;
 588         size_t l;
 589         int status = 0;
 590         const bool forks = (argc > 0);
 591         bool is_pipe = STAT_RECORD ? perf_stat.data.is_pipe : false;
 592         struct perf_evsel_config_term *err_term;
 593
 594         if (interval) {
 595                 ts.tv_sec  = interval / USEC_PER_MSEC;
 596                 ts.tv_nsec = (interval % USEC_PER_MSEC) * NSEC_PER_MSEC;
 597         } else if (timeout) {
 598                 ts.tv_sec  = timeout / USEC_PER_MSEC;
 599                 ts.tv_nsec = (timeout % USEC_PER_MSEC) * NSEC_PER_MSEC;
 600         } else {
 601                 ts.tv_sec  = 1;
 602                 ts.tv_nsec = 0;
 603         }
 604
 605         if (forks) {
 606                 if (perf_evlist__prepare_workload(evsel_list, &target, argv, is_pipe,
 607                                                   workload_exec_failed_signal) < 0) {
 608                         perror("failed to prepare workload");
 609                         return -1;
 610                 }
 611                 child_pid = evsel_list->workload.pid;
 612         }
 613
 614         if (group)
 615                 perf_evlist__set_leader(evsel_list);
 616
 617         evlist__for_each_entry(evsel_list, counter) {
 618 try_again:
 619                 if (create_perf_stat_counter(counter) < 0) {
 620
 621                         /* Weak group failed. Reset the group. */
 622                         if ((errno == EINVAL || errno == EBADF) &&
 623                             counter->leader != counter &&
 624                             counter->weak_group) {
 625                                 counter = perf_evsel__reset_weak_group(counter);
 626                                 goto try_again;
 627                         }
 628
 629                         /*
 630                          * PPC returns ENXIO for HW counters until 2.6.37
 631                          * (behavior changed with commit b0a873e).
 632                          */
 633                         if (errno == EINVAL || errno == ENOSYS ||
 634                             errno == ENOENT || errno == EOPNOTSUPP ||
 635                             errno == ENXIO) {
 636                                 if (verbose > 0)
 637                                         ui__warning("%s event is not supported by the kernel.\n",
 638                                                     perf_evsel__name(counter));
 639                                 counter->supported = false;
 640
 641                                 if ((counter->leader != counter) ||
 642                                     !(counter->leader->nr_members > 1))
 643                                         continue;
 644                         } else if (perf_evsel__fallback(counter, errno, msg, sizeof(msg))) {
 645                                 if (verbose > 0)
 646                                         ui__warning("%s\n", msg);
 647                                 goto try_again;
 648                         } else if (target__has_per_thread(&target) &&
 649                                    evsel_list->threads &&
 650                                    evsel_list->threads->err_thread != -1) {
 651                                 /*
 652                                  * For global --per-thread case, skip current
 653                                  * error thread.
 654                                  */
 655                                 if (!thread_map__remove(evsel_list->threads,
 656                                                         evsel_list->threads->err_thread)) {
 657                                         evsel_list->threads->err_thread = -1;
 658                                         goto try_again;
 659                                 }
 660                         }
 661
 662                         perf_evsel__open_strerror(counter, &target,
 663                                                   errno, msg, sizeof(msg));
 664                         ui__error("%s\n", msg);
 665
 666                         if (child_pid != -1)
 667                                 kill(child_pid, SIGTERM);
 668
 669                         return -1;
 670                 }
 671                 counter->supported = true;
 672
 673                 l = strlen(counter->unit);
 674                 if (l > unit_width)
 675                         unit_width = l;
 676
 677                 if (perf_evsel__should_store_id(counter) &&
 678                     store_counter_ids(counter))
 679                         return -1;
 680         }
 681
 682         if (perf_evlist__apply_filters(evsel_list, &counter)) {
 683                 pr_err("failed to set filter \"%s\" on event %s with %d (%s)\n",
 684                         counter->filter, perf_evsel__name(counter), errno,
 685                         str_error_r(errno, msg, sizeof(msg)));
 686                 return -1;
 687         }
 688
 689         if (perf_evlist__apply_drv_configs(evsel_list, &counter, &err_term)) {
 690                 pr_err("failed to set config \"%s\" on event %s with %d (%s)\n",
 691                       err_term->val.drv_cfg, perf_evsel__name(counter), errno,
 692                       str_error_r(errno, msg, sizeof(msg)));
 693                 return -1;
 694         }
 695
 696         if (STAT_RECORD) {
 697                 int err, fd = perf_data__fd(&perf_stat.data);
 698
 699                 if (is_pipe) {
 700                         err = perf_header__write_pipe(perf_data__fd(&perf_stat.data));
 701                 } else {
 702                         err = perf_session__write_header(perf_stat.session, evsel_list,
 703                                                          fd, false);
 704                 }
 705
 706                 if (err < 0)
 707                         return err;
 708
 709                 err = perf_stat_synthesize_config(is_pipe);
 710                 if (err < 0)
 711                         return err;
 712         }
 713
 714         /*
 715          * Enable counters and exec the command:
 716          */
 717         t0 = rdclock();
 718         clock_gettime(CLOCK_MONOTONIC, &ref_time);
 719
 720         if (forks) {
 721                 perf_evlist__start_workload(evsel_list);
 722                 enable_counters();
 723
 724                 if (interval || timeout) {
 725                         while (!waitpid(child_pid, &status, WNOHANG)) {
 726                                 nanosleep(&ts, NULL);
 727                                 if (timeout)
 728                                         break;
 729                                 process_interval();
 730                                 if (interval_count && !(--times))
 731                                         break;
 732                         }
 733                 }
 734                 wait4(child_pid, &status, 0, &ru_data);
 735
 736                 if (workload_exec_errno) {
 737                         const char *emsg = str_error_r(workload_exec_errno, msg, sizeof(msg));
 738                         pr_err("Workload failed: %s\n", emsg);
 739                         return -1;
 740                 }
 741
 742                 if (WIFSIGNALED(status))
 743                         psignal(WTERMSIG(status), argv[0]);
 744         } else {
 745                 enable_counters();
 746                 while (!done) {
 747                         nanosleep(&ts, NULL);
 748                         if (timeout)
 749                                 break;
 750                         if (interval) {
 751                                 process_interval();
 752                                 if (interval_count && !(--times))
 753                                         break;
 754                         }
 755                 }
 756         }
 757
 758         disable_counters();
 759
 760         t1 = rdclock();
 761
 762         if (walltime_run_table)
 763                 walltime_run[run_idx] = t1 - t0;
 764
 765         update_stats(&walltime_nsecs_stats, t1 - t0);
 766
 767         /*
 768          * Closing a group leader splits the group, and as we only disable
 769          * group leaders, results in remaining events becoming enabled. To
 770          * avoid arbitrary skew, we must read all counters before closing any
 771          * group leaders.
 772          */
 773         read_counters();
 774         perf_evlist__close(evsel_list);
 775
 776         return WEXITSTATUS(status);
 777 }
 778
 779 static int run_perf_stat(int argc, const char **argv, int run_idx)
 780 {
 781         int ret;
 782
 783         if (pre_cmd) {
 784                 ret = system(pre_cmd);
 785                 if (ret)
 786                         return ret;
 787         }
 788
 789         if (sync_run)
 790                 sync();
 791
 792         ret = __run_perf_stat(argc, argv, run_idx);
 793         if (ret)
 794                 return ret;
 795
 796         if (post_cmd) {
 797                 ret = system(post_cmd);
 798                 if (ret)
 799                         return ret;
 800         }
 801
 802         return ret;
 803 }
 804
 805 static void print_running(u64 run, u64 ena)
 806 {
 807         if (csv_output) {
 808                 fprintf(stat_config.output, "%s%" PRIu64 "%s%.2f",
 809                                         csv_sep,
 810                                         run,
 811                                         csv_sep,
 812                                         ena ? 100.0 * run / ena : 100.0);
 813         } else if (run != ena) {
 814                 fprintf(stat_config.output, "  (%.2f%%)", 100.0 * run / ena);
 815         }
 816 }
 817
 818 static void print_noise_pct(double total, double avg)
 819 {
 820         double pct = rel_stddev_stats(total, avg);
 821
 822         if (csv_output)
 823                 fprintf(stat_config.output, "%s%.2f%%", csv_sep, pct);
 824         else if (pct)
 825                 fprintf(stat_config.output, "  ( +-%6.2f%% )", pct);
 826 }
 827
 828 static void print_noise(struct perf_evsel *evsel, double avg)
 829 {
 830         struct perf_stat_evsel *ps;
 831
 832         if (run_count == 1)
 833                 return;
 834
 835         ps = evsel->stats;
 836         print_noise_pct(stddev_stats(&ps->res_stats[0]), avg);
 837 }
 838
 839 static void aggr_printout(struct perf_evsel *evsel, int id, int nr)
 840 {
 841         switch (stat_config.aggr_mode) {
 842         case AGGR_CORE:
 843                 fprintf(stat_config.output, "S%d-C%*d%s%*d%s",
 844                         cpu_map__id_to_socket(id),
 845                         csv_output ? 0 : -8,
 846                         cpu_map__id_to_cpu(id),
 847                         csv_sep,
 848                         csv_output ? 0 : 4,
 849                         nr,
 850                         csv_sep);
 851                 break;
 852         case AGGR_SOCKET:
 853                 fprintf(stat_config.output, "S%*d%s%*d%s",
 854                         csv_output ? 0 : -5,
 855                         id,
 856                         csv_sep,
 857                         csv_output ? 0 : 4,
 858                         nr,
 859                         csv_sep);
 860                         break;
 861         case AGGR_NONE:
 862                 fprintf(stat_config.output, "CPU%*d%s",
 863                         csv_output ? 0 : -4,
 864                         perf_evsel__cpus(evsel)->map[id], csv_sep);
 865                 break;
 866         case AGGR_THREAD:
 867                 fprintf(stat_config.output, "%*s-%*d%s",
 868                         csv_output ? 0 : 16,
 869                         thread_map__comm(evsel->threads, id),
 870                         csv_output ? 0 : -8,
 871                         thread_map__pid(evsel->threads, id),
 872                         csv_sep);
 873                 break;
 874         case AGGR_GLOBAL:
 875         case AGGR_UNSET:
 876         default:
 877                 break;
 878         }
 879 }
 880
 881 struct outstate {
 882         FILE *fh;
 883         bool newline;
 884         const char *prefix;
 885         int  nfields;
 886         int  id, nr;
 887         struct perf_evsel *evsel;
 888 };
 889
 890 #define METRIC_LEN  35
 891
 892 static void new_line_std(void *ctx)
 893 {
 894         struct outstate *os = ctx;
 895
 896         os->newline = true;
 897 }
 898
 899 static void do_new_line_std(struct outstate *os)
 900 {
 901         fputc('\n', os->fh);
 902         fputs(os->prefix, os->fh);
 903         aggr_printout(os->evsel, os->id, os->nr);
 904         if (stat_config.aggr_mode == AGGR_NONE)
 905                 fprintf(os->fh, "        ");
 906         fprintf(os->fh, "                                                 ");
 907 }
 908
 909 static void print_metric_std(void *ctx, const char *color, const char *fmt,
 910                              const char *unit, double val)
 911 {
 912         struct outstate *os = ctx;
 913         FILE *out = os->fh;
 914         int n;
 915         bool newline = os->newline;
 916
 917         os->newline = false;
 918
 919         if (unit == NULL || fmt == NULL) {
 920                 fprintf(out, "%-*s", METRIC_LEN, "");
 921                 return;
 922         }
 923
 924         if (newline)
 925                 do_new_line_std(os);
 926
 927         n = fprintf(out, " # ");
 928         if (color)
 929                 n += color_fprintf(out, color, fmt, val);
 930         else
 931                 n += fprintf(out, fmt, val);
 932         fprintf(out, " %-*s", METRIC_LEN - n - 1, unit);
 933 }
 934
 935 static void new_line_csv(void *ctx)
 936 {
 937         struct outstate *os = ctx;
 938         int i;
 939
 940         fputc('\n', os->fh);
 941         if (os->prefix)
 942                 fprintf(os->fh, "%s%s", os->prefix, csv_sep);
 943         aggr_printout(os->evsel, os->id, os->nr);
 944         for (i = 0; i < os->nfields; i++)
 945                 fputs(csv_sep, os->fh);
 946 }
 947
 948 static void print_metric_csv(void *ctx,
 949                              const char *color __maybe_unused,
 950                              const char *fmt, const char *unit, double val)
 951 {
 952         struct outstate *os = ctx;
 953         FILE *out = os->fh;
 954         char buf[64], *vals, *ends;
 955
 956         if (unit == NULL || fmt == NULL) {
 957                 fprintf(out, "%s%s", csv_sep, csv_sep);
 958                 return;
 959         }
 960         snprintf(buf, sizeof(buf), fmt, val);
 961         ends = vals = ltrim(buf);
 962         while (isdigit(*ends) || *ends == '.')
 963                 ends++;
 964         *ends = 0;
 965         while (isspace(*unit))
 966                 unit++;
 967         fprintf(out, "%s%s%s%s", csv_sep, vals, csv_sep, unit);
 968 }
 969
 970 #define METRIC_ONLY_LEN 20
 971
 972 /* Filter out some columns that don't work well in metrics only mode */
 973
 974 static bool valid_only_metric(const char *unit)
 975 {
 976         if (!unit)
 977                 return false;
 978         if (strstr(unit, "/sec") ||
 979             strstr(unit, "hz") ||
 980             strstr(unit, "Hz") ||
 981             strstr(unit, "CPUs utilized"))
 982                 return false;
 983         return true;
 984 }
 985
 986 static const char *fixunit(char *buf, struct perf_evsel *evsel,
 987                            const char *unit)
 988 {
 989         if (!strncmp(unit, "of all", 6)) {
 990                 snprintf(buf, 1024, "%s %s", perf_evsel__name(evsel),
 991                          unit);
 992                 return buf;
 993         }
 994         return unit;
 995 }
 996
 997 static void print_metric_only(void *ctx, const char *color, const char *fmt,
 998                               const char *unit, double val)
 999 {
1000         struct outstate *os = ctx;
1001         FILE *out = os->fh;
1002         int n;
1003         char buf[1024];
1004         unsigned mlen = METRIC_ONLY_LEN;
1005
1006         if (!valid_only_metric(unit))
1007                 return;
1008         unit = fixunit(buf, os->evsel, unit);
1009         if (color)
1010                 n = color_fprintf(out, color, fmt, val);
1011         else
1012                 n = fprintf(out, fmt, val);
1013         if (n > METRIC_ONLY_LEN)
1014                 n = METRIC_ONLY_LEN;
1015         if (mlen < strlen(unit))
1016                 mlen = strlen(unit) + 1;
1017         fprintf(out, "%*s", mlen - n, "");
1018 }
1019
1020 static void print_metric_only_csv(void *ctx, const char *color __maybe_unused,
1021                                   const char *fmt,
1022                                   const char *unit, double val)
1023 {
1024         struct outstate *os = ctx;
1025         FILE *out = os->fh;
1026         char buf[64], *vals, *ends;
1027         char tbuf[1024];
1028
1029         if (!valid_only_metric(unit))
1030                 return;
1031         unit = fixunit(tbuf, os->evsel, unit);
1032         snprintf(buf, sizeof buf, fmt, val);
1033         ends = vals = ltrim(buf);
1034         while (isdigit(*ends) || *ends == '.')
1035                 ends++;
1036         *ends = 0;
1037         fprintf(out, "%s%s", vals, csv_sep);
1038 }
1039
1040 static void new_line_metric(void *ctx __maybe_unused)
1041 {
1042 }
1043
1044 static void print_metric_header(void *ctx, const char *color __maybe_unused,
1045                                 const char *fmt __maybe_unused,
1046                                 const char *unit, double val __maybe_unused)
1047 {
1048         struct outstate *os = ctx;
1049         char tbuf[1024];
1050
1051         if (!valid_only_metric(unit))
1052                 return;
1053         unit = fixunit(tbuf, os->evsel, unit);
1054         if (csv_output)
1055                 fprintf(os->fh, "%s%s", unit, csv_sep);
1056         else
1057                 fprintf(os->fh, "%-*s ", METRIC_ONLY_LEN, unit);
1058 }
1059
1060 static void nsec_printout(int id, int nr, struct perf_evsel *evsel, double avg)
1061 {
1062         FILE *output = stat_config.output;
1063         double msecs = avg / NSEC_PER_MSEC;
1064         const char *fmt_v, *fmt_n;
1065         char name[25];
1066
1067         fmt_v = csv_output ? "%.6f%s" : "%18.6f%s";
1068         fmt_n = csv_output ? "%s" : "%-25s";
1069
1070         aggr_printout(evsel, id, nr);
1071
1072         scnprintf(name, sizeof(name), "%s%s",
1073                   perf_evsel__name(evsel), csv_output ? "" : " (msec)");
1074
1075         fprintf(output, fmt_v, msecs, csv_sep);
1076
1077         if (csv_output)
1078                 fprintf(output, "%s%s", evsel->unit, csv_sep);
1079         else
1080                 fprintf(output, "%-*s%s", unit_width, evsel->unit, csv_sep);
1081
1082         fprintf(output, fmt_n, name);
1083
1084         if (evsel->cgrp)
1085                 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
1086 }
1087
1088 static int first_shadow_cpu(struct perf_evsel *evsel, int id)
1089 {
1090         int i;
1091
1092         if (!aggr_get_id)
1093                 return 0;
1094
1095         if (stat_config.aggr_mode == AGGR_NONE)
1096                 return id;
1097
1098         if (stat_config.aggr_mode == AGGR_GLOBAL)
1099                 return 0;
1100
1101         for (i = 0; i < perf_evsel__nr_cpus(evsel); i++) {
1102                 int cpu2 = perf_evsel__cpus(evsel)->map[i];
1103
1104                 if (aggr_get_id(evsel_list->cpus, cpu2) == id)
1105                         return cpu2;
1106         }
1107         return 0;
1108 }
1109
1110 static void abs_printout(int id, int nr, struct perf_evsel *evsel, double avg)
1111 {
1112         FILE *output = stat_config.output;
1113         double sc =  evsel->scale;
1114         const char *fmt;
1115
1116         if (csv_output) {
1117                 fmt = floor(sc) != sc ?  "%.2f%s" : "%.0f%s";
1118         } else {
1119                 if (big_num)
1120                         fmt = floor(sc) != sc ? "%'18.2f%s" : "%'18.0f%s";
1121                 else
1122                         fmt = floor(sc) != sc ? "%18.2f%s" : "%18.0f%s";
1123         }
1124
1125         aggr_printout(evsel, id, nr);
1126
1127         fprintf(output, fmt, avg, csv_sep);
1128
1129         if (evsel->unit)
1130                 fprintf(output, "%-*s%s",
1131                         csv_output ? 0 : unit_width,
1132                         evsel->unit, csv_sep);
1133
1134         fprintf(output, "%-*s", csv_output ? 0 : 25, perf_evsel__name(evsel));
1135
1136         if (evsel->cgrp)
1137                 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
1138 }
1139
1140 static bool is_mixed_hw_group(struct perf_evsel *counter)
1141 {
1142         struct perf_evlist *evlist = counter->evlist;
1143         u32 pmu_type = counter->attr.type;
1144         struct perf_evsel *pos;
1145
1146         if (counter->nr_members < 2)
1147                 return false;
1148
1149         evlist__for_each_entry(evlist, pos) {
1150                 /* software events can be part of any hardware group */
1151                 if (pos->attr.type == PERF_TYPE_SOFTWARE)
1152                         continue;
1153                 if (pmu_type == PERF_TYPE_SOFTWARE) {
1154                         pmu_type = pos->attr.type;
1155                         continue;
1156                 }
1157                 if (pmu_type != pos->attr.type)
1158                         return true;
1159         }
1160
1161         return false;
1162 }
1163
1164 static void printout(int id, int nr, struct perf_evsel *counter, double uval,
1165                      char *prefix, u64 run, u64 ena, double noise,
1166                      struct runtime_stat *st)
1167 {
1168         struct perf_stat_output_ctx out;
1169         struct outstate os = {
1170                 .fh = stat_config.output,
1171                 .prefix = prefix ? prefix : "",
1172                 .id = id,
1173                 .nr = nr,
1174                 .evsel = counter,
1175         };
1176         print_metric_t pm = print_metric_std;
1177         void (*nl)(void *);
1178
1179         if (metric_only) {
1180                 nl = new_line_metric;
1181                 if (csv_output)
1182                         pm = print_metric_only_csv;
1183                 else
1184                         pm = print_metric_only;
1185         } else
1186                 nl = new_line_std;
1187
1188         if (csv_output && !metric_only) {
1189                 static int aggr_fields[] = {
1190                         [AGGR_GLOBAL] = 0,
1191                         [AGGR_THREAD] = 1,
1192                         [AGGR_NONE] = 1,
1193                         [AGGR_SOCKET] = 2,
1194                         [AGGR_CORE] = 2,
1195                 };
1196
1197                 pm = print_metric_csv;
1198                 nl = new_line_csv;
1199                 os.nfields = 3;
1200                 os.nfields += aggr_fields[stat_config.aggr_mode];
1201                 if (counter->cgrp)
1202                         os.nfields++;
1203         }
1204         if (run == 0 || ena == 0 || counter->counts->scaled == -1) {
1205                 if (metric_only) {
1206                         pm(&os, NULL, "", "", 0);
1207                         return;
1208                 }
1209                 aggr_printout(counter, id, nr);
1210
1211                 fprintf(stat_config.output, "%*s%s",
1212                         csv_output ? 0 : 18,
1213                         counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
1214                         csv_sep);
1215
1216                 if (counter->supported) {
1217                         print_free_counters_hint = 1;
1218                         if (is_mixed_hw_group(counter))
1219                                 print_mixed_hw_group_error = 1;
1220                 }
1221
1222                 fprintf(stat_config.output, "%-*s%s",
1223                         csv_output ? 0 : unit_width,
1224                         counter->unit, csv_sep);
1225
1226                 fprintf(stat_config.output, "%*s",
1227                         csv_output ? 0 : -25,
1228                         perf_evsel__name(counter));
1229
1230                 if (counter->cgrp)
1231                         fprintf(stat_config.output, "%s%s",
1232                                 csv_sep, counter->cgrp->name);
1233
1234                 if (!csv_output)
1235                         pm(&os, NULL, NULL, "", 0);
1236                 print_noise(counter, noise);
1237                 print_running(run, ena);
1238                 if (csv_output)
1239                         pm(&os, NULL, NULL, "", 0);
1240                 return;
1241         }
1242
1243         if (metric_only)
1244                 /* nothing */;
1245         else if (nsec_counter(counter))
1246                 nsec_printout(id, nr, counter, uval);
1247         else
1248                 abs_printout(id, nr, counter, uval);
1249
1250         out.print_metric = pm;
1251         out.new_line = nl;
1252         out.ctx = &os;
1253         out.force_header = false;
1254
1255         if (csv_output && !metric_only) {
1256                 print_noise(counter, noise);
1257                 print_running(run, ena);
1258         }
1259
1260         perf_stat__print_shadow_stats(counter, uval,
1261                                 first_shadow_cpu(counter, id),
1262                                 &out, &metric_events, st);
1263         if (!csv_output && !metric_only) {
1264                 print_noise(counter, noise);
1265                 print_running(run, ena);
1266         }
1267 }
1268
1269 static void aggr_update_shadow(void)
1270 {
1271         int cpu, s2, id, s;
1272         u64 val;
1273         struct perf_evsel *counter;
1274
1275         for (s = 0; s < aggr_map->nr; s++) {
1276                 id = aggr_map->map[s];
1277                 evlist__for_each_entry(evsel_list, counter) {
1278                         val = 0;
1279                         for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
1280                                 s2 = aggr_get_id(evsel_list->cpus, cpu);
1281                                 if (s2 != id)
1282                                         continue;
1283                                 val += perf_counts(counter->counts, cpu, 0)->val;
1284                         }
1285                         perf_stat__update_shadow_stats(counter, val,
1286                                         first_shadow_cpu(counter, id),
1287                                         &rt_stat);
1288                 }
1289         }
1290 }
1291
1292 static void uniquify_event_name(struct perf_evsel *counter)
1293 {
1294         char *new_name;
1295         char *config;
1296
1297         if (counter->uniquified_name ||
1298             !counter->pmu_name || !strncmp(counter->name, counter->pmu_name,
1299                                            strlen(counter->pmu_name)))
1300                 return;
1301
1302         config = strchr(counter->name, '/');
1303         if (config) {
1304                 if (asprintf(&new_name,
1305                              "%s%s", counter->pmu_name, config) > 0) {
1306                         free(counter->name);
1307                         counter->name = new_name;
1308                 }
1309         } else {
1310                 if (asprintf(&new_name,
1311                              "%s [%s]", counter->name, counter->pmu_name) > 0) {
1312                         free(counter->name);
1313                         counter->name = new_name;
1314                 }
1315         }
1316
1317         counter->uniquified_name = true;
1318 }
1319
1320 static void collect_all_aliases(struct perf_evsel *counter,
1321                             void (*cb)(struct perf_evsel *counter, void *data,
1322                                        bool first),
1323                             void *data)
1324 {
1325         struct perf_evsel *alias;
1326
1327         alias = list_prepare_entry(counter, &(evsel_list->entries), node);
1328         list_for_each_entry_continue (alias, &evsel_list->entries, node) {
1329                 if (strcmp(perf_evsel__name(alias), perf_evsel__name(counter)) ||
1330                     alias->scale != counter->scale ||
1331                     alias->cgrp != counter->cgrp ||
1332                     strcmp(alias->unit, counter->unit) ||
1333                     nsec_counter(alias) != nsec_counter(counter))
1334                         break;
1335                 alias->merged_stat = true;
1336                 cb(alias, data, false);
1337         }
1338 }
1339
1340 static bool collect_data(struct perf_evsel *counter,
1341                             void (*cb)(struct perf_evsel *counter, void *data,
1342                                        bool first),
1343                             void *data)
1344 {
1345         if (counter->merged_stat)
1346                 return false;
1347         cb(counter, data, true);
1348         if (no_merge)
1349                 uniquify_event_name(counter);
1350         else if (counter->auto_merge_stats)
1351                 collect_all_aliases(counter, cb, data);
1352         return true;
1353 }
1354
1355 struct aggr_data {
1356         u64 ena, run, val;
1357         int id;
1358         int nr;
1359         int cpu;
1360 };
1361
1362 static void aggr_cb(struct perf_evsel *counter, void *data, bool first)
1363 {
1364         struct aggr_data *ad = data;
1365         int cpu, s2;
1366
1367         for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
1368                 struct perf_counts_values *counts;
1369
1370                 s2 = aggr_get_id(perf_evsel__cpus(counter), cpu);
1371                 if (s2 != ad->id)
1372                         continue;
1373                 if (first)
1374                         ad->nr++;
1375                 counts = perf_counts(counter->counts, cpu, 0);
1376                 /*
1377                  * When any result is bad, make them all to give
1378                  * consistent output in interval mode.
1379                  */
1380                 if (counts->ena == 0 || counts->run == 0 ||
1381                     counter->counts->scaled == -1) {
1382                         ad->ena = 0;
1383                         ad->run = 0;
1384                         break;
1385                 }
1386                 ad->val += counts->val;
1387                 ad->ena += counts->ena;
1388                 ad->run += counts->run;
1389         }
1390 }
1391
1392 static void print_aggr(char *prefix)
1393 {
1394         FILE *output = stat_config.output;
1395         struct perf_evsel *counter;
1396         int s, id, nr;
1397         double uval;
1398         u64 ena, run, val;
1399         bool first;
1400
1401         if (!(aggr_map || aggr_get_id))
1402                 return;
1403
1404         aggr_update_shadow();
1405
1406         /*
1407          * With metric_only everything is on a single line.
1408          * Without each counter has its own line.
1409          */
1410         for (s = 0; s < aggr_map->nr; s++) {
1411                 struct aggr_data ad;
1412                 if (prefix && metric_only)
1413                         fprintf(output, "%s", prefix);
1414
1415                 ad.id = id = aggr_map->map[s];
1416                 first = true;
1417                 evlist__for_each_entry(evsel_list, counter) {
1418                         if (is_duration_time(counter))
1419                                 continue;
1420
1421                         ad.val = ad.ena = ad.run = 0;
1422                         ad.nr = 0;
1423                         if (!collect_data(counter, aggr_cb, &ad))
1424                                 continue;
1425                         nr = ad.nr;
1426                         ena = ad.ena;
1427                         run = ad.run;
1428                         val = ad.val;
1429                         if (first && metric_only) {
1430                                 first = false;
1431                                 aggr_printout(counter, id, nr);
1432                         }
1433                         if (prefix && !metric_only)
1434                                 fprintf(output, "%s", prefix);
1435
1436                         uval = val * counter->scale;
1437                         printout(id, nr, counter, uval, prefix, run, ena, 1.0,
1438                                  &rt_stat);
1439                         if (!metric_only)
1440                                 fputc('\n', output);
1441                 }
1442                 if (metric_only)
1443                         fputc('\n', output);
1444         }
1445 }
1446
1447 static int cmp_val(const void *a, const void *b)
1448 {
1449         return ((struct perf_aggr_thread_value *)b)->val -
1450                 ((struct perf_aggr_thread_value *)a)->val;
1451 }
1452
1453 static struct perf_aggr_thread_value *sort_aggr_thread(
1454                                         struct perf_evsel *counter,
1455                                         int nthreads, int ncpus,
1456                                         int *ret)
1457 {
1458         int cpu, thread, i = 0;
1459         double uval;
1460         struct perf_aggr_thread_value *buf;
1461
1462         buf = calloc(nthreads, sizeof(struct perf_aggr_thread_value));
1463         if (!buf)
1464                 return NULL;
1465
1466         for (thread = 0; thread < nthreads; thread++) {
1467                 u64 ena = 0, run = 0, val = 0;
1468
1469                 for (cpu = 0; cpu < ncpus; cpu++) {
1470                         val += perf_counts(counter->counts, cpu, thread)->val;
1471                         ena += perf_counts(counter->counts, cpu, thread)->ena;
1472                         run += perf_counts(counter->counts, cpu, thread)->run;
1473                 }
1474
1475                 uval = val * counter->scale;
1476
1477                 /*
1478                  * Skip value 0 when enabling --per-thread globally,
1479                  * otherwise too many 0 output.
1480                  */
1481                 if (uval == 0.0 && target__has_per_thread(&target))
1482                         continue;
1483
1484                 buf[i].counter = counter;
1485                 buf[i].id = thread;
1486                 buf[i].uval = uval;
1487                 buf[i].val = val;
1488                 buf[i].run = run;
1489                 buf[i].ena = ena;
1490                 i++;
1491         }
1492
1493         qsort(buf, i, sizeof(struct perf_aggr_thread_value), cmp_val);
1494
1495         if (ret)
1496                 *ret = i;
1497
1498         return buf;
1499 }
1500
1501 static void print_aggr_thread(struct perf_evsel *counter, char *prefix)
1502 {
1503         FILE *output = stat_config.output;
1504         int nthreads = thread_map__nr(counter->threads);
1505         int ncpus = cpu_map__nr(counter->cpus);
1506         int thread, sorted_threads, id;
1507         struct perf_aggr_thread_value *buf;
1508
1509         buf = sort_aggr_thread(counter, nthreads, ncpus, &sorted_threads);
1510         if (!buf) {
1511                 perror("cannot sort aggr thread");
1512                 return;
1513         }
1514
1515         for (thread = 0; thread < sorted_threads; thread++) {
1516                 if (prefix)
1517                         fprintf(output, "%s", prefix);
1518
1519                 id = buf[thread].id;
1520                 if (stat_config.stats)
1521                         printout(id, 0, buf[thread].counter, buf[thread].uval,
1522                                  prefix, buf[thread].run, buf[thread].ena, 1.0,
1523                                  &stat_config.stats[id]);
1524                 else
1525                         printout(id, 0, buf[thread].counter, buf[thread].uval,
1526                                  prefix, buf[thread].run, buf[thread].ena, 1.0,
1527                                  &rt_stat);
1528                 fputc('\n', output);
1529         }
1530
1531         free(buf);
1532 }
1533
1534 struct caggr_data {
1535         double avg, avg_enabled, avg_running;
1536 };
1537
1538 static void counter_aggr_cb(struct perf_evsel *counter, void *data,
1539                             bool first __maybe_unused)
1540 {
1541         struct caggr_data *cd = data;
1542         struct perf_stat_evsel *ps = counter->stats;
1543
1544         cd->avg += avg_stats(&ps->res_stats[0]);
1545         cd->avg_enabled += avg_stats(&ps->res_stats[1]);
1546         cd->avg_running += avg_stats(&ps->res_stats[2]);
1547 }
1548
1549 /*
1550  * Print out the results of a single counter:
1551  * aggregated counts in system-wide mode
1552  */
1553 static void print_counter_aggr(struct perf_evsel *counter, char *prefix)
1554 {
1555         FILE *output = stat_config.output;
1556         double uval;
1557         struct caggr_data cd = { .avg = 0.0 };
1558
1559         if (!collect_data(counter, counter_aggr_cb, &cd))
1560                 return;
1561
1562         if (prefix && !metric_only)
1563                 fprintf(output, "%s", prefix);
1564
1565         uval = cd.avg * counter->scale;
1566         printout(-1, 0, counter, uval, prefix, cd.avg_running, cd.avg_enabled,
1567                  cd.avg, &rt_stat);
1568         if (!metric_only)
1569                 fprintf(output, "\n");
1570 }
1571
1572 static void counter_cb(struct perf_evsel *counter, void *data,
1573                        bool first __maybe_unused)
1574 {
1575         struct aggr_data *ad = data;
1576
1577         ad->val += perf_counts(counter->counts, ad->cpu, 0)->val;
1578         ad->ena += perf_counts(counter->counts, ad->cpu, 0)->ena;
1579         ad->run += perf_counts(counter->counts, ad->cpu, 0)->run;
1580 }
1581
1582 /*
1583  * Print out the results of a single counter:
1584  * does not use aggregated count in system-wide
1585  */
1586 static void print_counter(struct perf_evsel *counter, char *prefix)
1587 {
1588         FILE *output = stat_config.output;
1589         u64 ena, run, val;
1590         double uval;
1591         int cpu;
1592
1593         for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
1594                 struct aggr_data ad = { .cpu = cpu };
1595
1596                 if (!collect_data(counter, counter_cb, &ad))
1597                         return;
1598                 val = ad.val;
1599                 ena = ad.ena;
1600                 run = ad.run;
1601
1602                 if (prefix)
1603                         fprintf(output, "%s", prefix);
1604
1605                 uval = val * counter->scale;
1606                 printout(cpu, 0, counter, uval, prefix, run, ena, 1.0,
1607                          &rt_stat);
1608
1609                 fputc('\n', output);
1610         }
1611 }
1612
1613 static void print_no_aggr_metric(char *prefix)
1614 {
1615         int cpu;
1616         int nrcpus = 0;
1617         struct perf_evsel *counter;
1618         u64 ena, run, val;
1619         double uval;
1620
1621         nrcpus = evsel_list->cpus->nr;
1622         for (cpu = 0; cpu < nrcpus; cpu++) {
1623                 bool first = true;
1624
1625                 if (prefix)
1626                         fputs(prefix, stat_config.output);
1627                 evlist__for_each_entry(evsel_list, counter) {
1628                         if (is_duration_time(counter))
1629                                 continue;
1630                         if (first) {
1631                                 aggr_printout(counter, cpu, 0);
1632                                 first = false;
1633                         }
1634                         val = perf_counts(counter->counts, cpu, 0)->val;
1635                         ena = perf_counts(counter->counts, cpu, 0)->ena;
1636                         run = perf_counts(counter->counts, cpu, 0)->run;
1637
1638                         uval = val * counter->scale;
1639                         printout(cpu, 0, counter, uval, prefix, run, ena, 1.0,
1640                                  &rt_stat);
1641                 }
1642                 fputc('\n', stat_config.output);
1643         }
1644 }
1645
1646 static int aggr_header_lens[] = {
1647         [AGGR_CORE] = 18,
1648         [AGGR_SOCKET] = 12,
1649         [AGGR_NONE] = 6,
1650         [AGGR_THREAD] = 24,
1651         [AGGR_GLOBAL] = 0,
1652 };
1653
1654 static const char *aggr_header_csv[] = {
1655         [AGGR_CORE]     =       "core,cpus,",
1656         [AGGR_SOCKET]   =       "socket,cpus",
1657         [AGGR_NONE]     =       "cpu,",
1658         [AGGR_THREAD]   =       "comm-pid,",
1659         [AGGR_GLOBAL]   =       ""
1660 };
1661
1662 static void print_metric_headers(const char *prefix, bool no_indent)
1663 {
1664         struct perf_stat_output_ctx out;
1665         struct perf_evsel *counter;
1666         struct outstate os = {
1667                 .fh = stat_config.output
1668         };
1669
1670         if (prefix)
1671                 fprintf(stat_config.output, "%s", prefix);
1672
1673         if (!csv_output && !no_indent)
1674                 fprintf(stat_config.output, "%*s",
1675                         aggr_header_lens[stat_config.aggr_mode], "");
1676         if (csv_output) {
1677                 if (stat_config.interval)
1678                         fputs("time,", stat_config.output);
1679                 fputs(aggr_header_csv[stat_config.aggr_mode],
1680                         stat_config.output);
1681         }
1682
1683         /* Print metrics headers only */
1684         evlist__for_each_entry(evsel_list, counter) {
1685                 if (is_duration_time(counter))
1686                         continue;
1687                 os.evsel = counter;
1688                 out.ctx = &os;
1689                 out.print_metric = print_metric_header;
1690                 out.new_line = new_line_metric;
1691                 out.force_header = true;
1692                 os.evsel = counter;
1693                 perf_stat__print_shadow_stats(counter, 0,
1694                                               0,
1695                                               &out,
1696                                               &metric_events,
1697                                               &rt_stat);
1698         }
1699         fputc('\n', stat_config.output);
1700 }
1701
1702 static void print_interval(char *prefix, struct timespec *ts)
1703 {
1704         FILE *output = stat_config.output;
1705         static int num_print_interval;
1706
1707         sprintf(prefix, "%6lu.%09lu%s", ts->tv_sec, ts->tv_nsec, csv_sep);
1708
1709         if (num_print_interval == 0 && !csv_output) {
1710                 switch (stat_config.aggr_mode) {
1711                 case AGGR_SOCKET:
1712                         fprintf(output, "#           time socket cpus");
1713                         if (!metric_only)
1714                                 fprintf(output, "             counts %*s events\n", unit_width, "unit");
1715                         break;
1716                 case AGGR_CORE:
1717                         fprintf(output, "#           time core         cpus");
1718                         if (!metric_only)
1719                                 fprintf(output, "             counts %*s events\n", unit_width, "unit");
1720                         break;
1721                 case AGGR_NONE:
1722                         fprintf(output, "#           time CPU");
1723                         if (!metric_only)
1724                                 fprintf(output, "                counts %*s events\n", unit_width, "unit");
1725                         break;
1726                 case AGGR_THREAD:
1727                         fprintf(output, "#           time             comm-pid");
1728                         if (!metric_only)
1729                                 fprintf(output, "                  counts %*s events\n", unit_width, "unit");
1730                         break;
1731                 case AGGR_GLOBAL:
1732                 default:
1733                         fprintf(output, "#           time");
1734                         if (!metric_only)
1735                                 fprintf(output, "             counts %*s events\n", unit_width, "unit");
1736                 case AGGR_UNSET:
1737                         break;
1738                 }
1739         }
1740
1741         if (num_print_interval == 0 && metric_only)
1742                 print_metric_headers(" ", true);
1743         if (++num_print_interval == 25)
1744                 num_print_interval = 0;
1745 }
1746
1747 static void print_header(int argc, const char **argv)
1748 {
1749         FILE *output = stat_config.output;
1750         int i;
1751
1752         fflush(stdout);
1753
1754         if (!csv_output) {
1755                 fprintf(output, "\n");
1756                 fprintf(output, " Performance counter stats for ");
1757                 if (target.system_wide)
1758                         fprintf(output, "\'system wide");
1759                 else if (target.cpu_list)
1760                         fprintf(output, "\'CPU(s) %s", target.cpu_list);
1761                 else if (!target__has_task(&target)) {
1762                         fprintf(output, "\'%s", argv ? argv[0] : "pipe");
1763                         for (i = 1; argv && (i < argc); i++)
1764                                 fprintf(output, " %s", argv[i]);
1765                 } else if (target.pid)
1766                         fprintf(output, "process id \'%s", target.pid);
1767                 else
1768                         fprintf(output, "thread id \'%s", target.tid);
1769
1770                 fprintf(output, "\'");
1771                 if (run_count > 1)
1772                         fprintf(output, " (%d runs)", run_count);
1773                 fprintf(output, ":\n\n");
1774         }
1775 }
1776
1777 static int get_precision(double num)
1778 {
1779         if (num > 1)
1780                 return 0;
1781
1782         return lround(ceil(-log10(num)));
1783 }
1784
1785 static void print_table(FILE *output, int precision, double avg)
1786 {
1787         char tmp[64];
1788         int idx, indent = 0;
1789
1790         scnprintf(tmp, 64, " %17.*f", precision, avg);
1791         while (tmp[indent] == ' ')
1792                 indent++;
1793
1794         fprintf(output, "%*s# Table of individual measurements:\n", indent, "");
1795
1796         for (idx = 0; idx < run_count; idx++) {
1797                 double run = (double) walltime_run[idx] / NSEC_PER_SEC;
1798                 int h, n = 1 + abs((int) (100.0 * (run - avg)/run) / 5);
1799
1800                 fprintf(output, " %17.*f (%+.*f) ",
1801                         precision, run, precision, run - avg);
1802
1803                 for (h = 0; h < n; h++)
1804                         fprintf(output, "#");
1805
1806                 fprintf(output, "\n");
1807         }
1808
1809         fprintf(output, "\n%*s# Final result:\n", indent, "");
1810 }
1811
1812 static double timeval2double(struct timeval *t)
1813 {
1814         return t->tv_sec + (double) t->tv_usec/USEC_PER_SEC;
1815 }
1816
1817 static void print_footer(void)
1818 {
1819         double avg = avg_stats(&walltime_nsecs_stats) / NSEC_PER_SEC;
1820         FILE *output = stat_config.output;
1821         int n;
1822
1823         if (!null_run)
1824                 fprintf(output, "\n");
1825
1826         if (run_count == 1) {
1827                 fprintf(output, " %17.9f seconds time elapsed", avg);
1828
1829                 if (ru_display) {
1830                         double ru_utime = timeval2double(&ru_data.ru_utime);
1831                         double ru_stime = timeval2double(&ru_data.ru_stime);
1832
1833                         fprintf(output, "\n\n");
1834                         fprintf(output, " %17.9f seconds user\n", ru_utime);
1835                         fprintf(output, " %17.9f seconds sys\n", ru_stime);
1836                 }
1837         } else {
1838                 double sd = stddev_stats(&walltime_nsecs_stats) / NSEC_PER_SEC;
1839                 /*
1840                  * Display at most 2 more significant
1841                  * digits than the stddev inaccuracy.
1842                  */
1843                 int precision = get_precision(sd) + 2;
1844
1845                 if (walltime_run_table)
1846                         print_table(output, precision, avg);
1847
1848                 fprintf(output, " %17.*f +- %.*f seconds time elapsed",
1849                         precision, avg, precision, sd);
1850
1851                 print_noise_pct(sd, avg);
1852         }
1853         fprintf(output, "\n\n");
1854
1855         if (print_free_counters_hint &&
1856             sysctl__read_int("kernel/nmi_watchdog", &n) >= 0 &&
1857             n > 0)
1858                 fprintf(output,
1859 "Some events weren't counted. Try disabling the NMI watchdog:\n"
1860 "       echo 0 > /proc/sys/kernel/nmi_watchdog\n"
1861 "       perf stat ...\n"
1862 "       echo 1 > /proc/sys/kernel/nmi_watchdog\n");
1863
1864         if (print_mixed_hw_group_error)
1865                 fprintf(output,
1866                         "The events in group usually have to be from "
1867                         "the same PMU. Try reorganizing the group.\n");
1868 }
1869
1870 static void print_counters(struct timespec *ts, int argc, const char **argv)
1871 {
1872         int interval = stat_config.interval;
1873         struct perf_evsel *counter;
1874         char buf[64], *prefix = NULL;
1875
1876         /* Do not print anything if we record to the pipe. */
1877         if (STAT_RECORD && perf_stat.data.is_pipe)
1878                 return;
1879
1880         if (interval)
1881                 print_interval(prefix = buf, ts);
1882         else
1883                 print_header(argc, argv);
1884
1885         if (metric_only) {
1886                 static int num_print_iv;
1887
1888                 if (num_print_iv == 0 && !interval)
1889                         print_metric_headers(prefix, false);
1890                 if (num_print_iv++ == 25)
1891                         num_print_iv = 0;
1892                 if (stat_config.aggr_mode == AGGR_GLOBAL && prefix)
1893                         fprintf(stat_config.output, "%s", prefix);
1894         }
1895
1896         switch (stat_config.aggr_mode) {
1897         case AGGR_CORE:
1898         case AGGR_SOCKET:
1899                 print_aggr(prefix);
1900                 break;
1901         case AGGR_THREAD:
1902                 evlist__for_each_entry(evsel_list, counter) {
1903                         if (is_duration_time(counter))
1904                                 continue;
1905                         print_aggr_thread(counter, prefix);
1906                 }
1907                 break;
1908         case AGGR_GLOBAL:
1909                 evlist__for_each_entry(evsel_list, counter) {
1910                         if (is_duration_time(counter))
1911                                 continue;
1912                         print_counter_aggr(counter, prefix);
1913                 }
1914                 if (metric_only)
1915                         fputc('\n', stat_config.output);
1916                 break;
1917         case AGGR_NONE:
1918                 if (metric_only)
1919                         print_no_aggr_metric(prefix);
1920                 else {
1921                         evlist__for_each_entry(evsel_list, counter) {
1922                                 if (is_duration_time(counter))
1923                                         continue;
1924                                 print_counter(counter, prefix);
1925                         }
1926                 }
1927                 break;
1928         case AGGR_UNSET:
1929         default:
1930                 break;
1931         }
1932
1933         if (!interval && !csv_output)
1934                 print_footer();
1935
1936         fflush(stat_config.output);
1937 }
1938
1939 static volatile int signr = -1;
1940
1941 static void skip_signal(int signo)
1942 {
1943         if ((child_pid == -1) || stat_config.interval)
1944                 done = 1;
1945
1946         signr = signo;
1947         /*
1948          * render child_pid harmless
1949          * won't send SIGTERM to a random
1950          * process in case of race condition
1951          * and fast PID recycling
1952          */
1953         child_pid = -1;
1954 }
1955
1956 static void sig_atexit(void)
1957 {
1958         sigset_t set, oset;
1959
1960         /*
1961          * avoid race condition with SIGCHLD handler
1962          * in skip_signal() which is modifying child_pid
1963          * goal is to avoid send SIGTERM to a random
1964          * process
1965          */
1966         sigemptyset(&set);
1967         sigaddset(&set, SIGCHLD);
1968         sigprocmask(SIG_BLOCK, &set, &oset);
1969
1970         if (child_pid != -1)
1971                 kill(child_pid, SIGTERM);
1972
1973         sigprocmask(SIG_SETMASK, &oset, NULL);
1974
1975         if (signr == -1)
1976                 return;
1977
1978         signal(signr, SIG_DFL);
1979         kill(getpid(), signr);
1980 }
1981
1982 static int stat__set_big_num(const struct option *opt __maybe_unused,
1983                              const char *s __maybe_unused, int unset)
1984 {
1985         big_num_opt = unset ? 0 : 1;
1986         return 0;
1987 }
1988
1989 static int enable_metric_only(const struct option *opt __maybe_unused,
1990                               const char *s __maybe_unused, int unset)
1991 {
1992         force_metric_only = true;
1993         metric_only = !unset;
1994         return 0;
1995 }
1996
1997 static int parse_metric_groups(const struct option *opt,
1998                                const char *str,
1999                                int unset __maybe_unused)
2000 {
2001         return metricgroup__parse_groups(opt, str, &metric_events);
2002 }
2003
2004 static const struct option stat_options[] = {
2005         OPT_BOOLEAN('T', "transaction", &transaction_run,
2006                     "hardware transaction statistics"),
2007         OPT_CALLBACK('e', "event", &evsel_list, "event",
2008                      "event selector. use 'perf list' to list available events",
2009                      parse_events_option),
2010         OPT_CALLBACK(0, "filter", &evsel_list, "filter",
2011                      "event filter", parse_filter),
2012         OPT_BOOLEAN('i', "no-inherit", &no_inherit,
2013                     "child tasks do not inherit counters"),
2014         OPT_STRING('p', "pid", &target.pid, "pid",
2015                    "stat events on existing process id"),
2016         OPT_STRING('t', "tid", &target.tid, "tid",
2017                    "stat events on existing thread id"),
2018         OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
2019                     "system-wide collection from all CPUs"),
2020         OPT_BOOLEAN('g', "group", &group,
2021                     "put the counters into a counter group"),
2022         OPT_BOOLEAN('c', "scale", &stat_config.scale, "scale/normalize counters"),
2023         OPT_INCR('v', "verbose", &verbose,
2024                     "be more verbose (show counter open errors, etc)"),
2025         OPT_INTEGER('r', "repeat", &run_count,
2026                     "repeat command and print average + stddev (max: 100, forever: 0)"),
2027         OPT_BOOLEAN(0, "table", &walltime_run_table,
2028                     "display details about each run (only with -r option)"),
2029         OPT_BOOLEAN('n', "null", &null_run,
2030                     "null run - dont start any counters"),
2031         OPT_INCR('d', "detailed", &detailed_run,
2032                     "detailed run - start a lot of events"),
2033         OPT_BOOLEAN('S', "sync", &sync_run,
2034                     "call sync() before starting a run"),
2035         OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
2036                            "print large numbers with thousands\' separators",
2037                            stat__set_big_num),
2038         OPT_STRING('C', "cpu", &target.cpu_list, "cpu",
2039                     "list of cpus to monitor in system-wide"),
2040         OPT_SET_UINT('A', "no-aggr", &stat_config.aggr_mode,
2041                     "disable CPU count aggregation", AGGR_NONE),
2042         OPT_BOOLEAN(0, "no-merge", &no_merge, "Do not merge identical named events"),
2043         OPT_STRING('x', "field-separator", &csv_sep, "separator",
2044                    "print counts with custom separator"),
2045         OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
2046                      "monitor event in cgroup name only", parse_cgroups),
2047         OPT_STRING('o', "output", &output_name, "file", "output file name"),
2048         OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
2049         OPT_INTEGER(0, "log-fd", &output_fd,
2050                     "log output to fd, instead of stderr"),
2051         OPT_STRING(0, "pre", &pre_cmd, "command",
2052                         "command to run prior to the measured command"),
2053         OPT_STRING(0, "post", &post_cmd, "command",
2054                         "command to run after to the measured command"),
2055         OPT_UINTEGER('I', "interval-print", &stat_config.interval,
2056                     "print counts at regular interval in ms "
2057                     "(overhead is possible for values <= 100ms)"),
2058         OPT_INTEGER(0, "interval-count", &stat_config.times,
2059                     "print counts for fixed number of times"),
2060         OPT_UINTEGER(0, "timeout", &stat_config.timeout,
2061                     "stop workload and print counts after a timeout period in ms (>= 10ms)"),
2062         OPT_SET_UINT(0, "per-socket", &stat_config.aggr_mode,
2063                      "aggregate counts per processor socket", AGGR_SOCKET),
2064         OPT_SET_UINT(0, "per-core", &stat_config.aggr_mode,
2065                      "aggregate counts per physical processor core", AGGR_CORE),
2066         OPT_SET_UINT(0, "per-thread", &stat_config.aggr_mode,
2067                      "aggregate counts per thread", AGGR_THREAD),
2068         OPT_UINTEGER('D', "delay", &initial_delay,
2069                      "ms to wait before starting measurement after program start"),
2070         OPT_CALLBACK_NOOPT(0, "metric-only", &metric_only, NULL,
2071                         "Only print computed metrics. No raw values", enable_metric_only),
2072         OPT_BOOLEAN(0, "topdown", &topdown_run,
2073                         "measure topdown level 1 statistics"),
2074         OPT_BOOLEAN(0, "smi-cost", &smi_cost,
2075                         "measure SMI cost"),
2076         OPT_CALLBACK('M', "metrics", &evsel_list, "metric/metric group list",
2077                      "monitor specified metrics or metric groups (separated by ,)",
2078                      parse_metric_groups),
2079         OPT_END()
2080 };
2081
2082 static int perf_stat__get_socket(struct cpu_map *map, int cpu)
2083 {
2084         return cpu_map__get_socket(map, cpu, NULL);
2085 }
2086
2087 static int perf_stat__get_core(struct cpu_map *map, int cpu)
2088 {
2089         return cpu_map__get_core(map, cpu, NULL);
2090 }
2091
2092 static int cpu_map__get_max(struct cpu_map *map)
2093 {
2094         int i, max = -1;
2095
2096         for (i = 0; i < map->nr; i++) {
2097                 if (map->map[i] > max)
2098                         max = map->map[i];
2099         }
2100
2101         return max;
2102 }
2103
2104 static struct cpu_map *cpus_aggr_map;
2105
2106 static int perf_stat__get_aggr(aggr_get_id_t get_id, struct cpu_map *map, int idx)
2107 {
2108         int cpu;
2109
2110         if (idx >= map->nr)
2111                 return -1;
2112
2113         cpu = map->map[idx];
2114
2115         if (cpus_aggr_map->map[cpu] == -1)
2116                 cpus_aggr_map->map[cpu] = get_id(map, idx);
2117
2118         return cpus_aggr_map->map[cpu];
2119 }
2120
2121 static int perf_stat__get_socket_cached(struct cpu_map *map, int idx)
2122 {
2123         return perf_stat__get_aggr(perf_stat__get_socket, map, idx);
2124 }
2125
2126 static int perf_stat__get_core_cached(struct cpu_map *map, int idx)
2127 {
2128         return perf_stat__get_aggr(perf_stat__get_core, map, idx);
2129 }
2130
2131 static int perf_stat_init_aggr_mode(void)
2132 {
2133         int nr;
2134
2135         switch (stat_config.aggr_mode) {
2136         case AGGR_SOCKET:
2137                 if (cpu_map__build_socket_map(evsel_list->cpus, &aggr_map)) {
2138                         perror("cannot build socket map");
2139                         return -1;
2140                 }
2141                 aggr_get_id = perf_stat__get_socket_cached;
2142                 break;
2143         case AGGR_CORE:
2144                 if (cpu_map__build_core_map(evsel_list->cpus, &aggr_map)) {
2145                         perror("cannot build core map");
2146                         return -1;
2147                 }
2148                 aggr_get_id = perf_stat__get_core_cached;
2149                 break;
2150         case AGGR_NONE:
2151         case AGGR_GLOBAL:
2152         case AGGR_THREAD:
2153         case AGGR_UNSET:
2154         default:
2155                 break;
2156         }
2157
2158         /*
2159          * The evsel_list->cpus is the base we operate on,
2160          * taking the highest cpu number to be the size of
2161          * the aggregation translate cpumap.
2162          */
2163         nr = cpu_map__get_max(evsel_list->cpus);
2164         cpus_aggr_map = cpu_map__empty_new(nr + 1);
2165         return cpus_aggr_map ? 0 : -ENOMEM;
2166 }
2167
2168 static void perf_stat__exit_aggr_mode(void)
2169 {
2170         cpu_map__put(aggr_map);
2171         cpu_map__put(cpus_aggr_map);
2172         aggr_map = NULL;
2173         cpus_aggr_map = NULL;
2174 }
2175
2176 static inline int perf_env__get_cpu(struct perf_env *env, struct cpu_map *map, int idx)
2177 {
2178         int cpu;
2179
2180         if (idx > map->nr)
2181                 return -1;
2182
2183         cpu = map->map[idx];
2184
2185         if (cpu >= env->nr_cpus_avail)
2186                 return -1;
2187
2188         return cpu;
2189 }
2190
2191 static int perf_env__get_socket(struct cpu_map *map, int idx, void *data)
2192 {
2193         struct perf_env *env = data;
2194         int cpu = perf_env__get_cpu(env, map, idx);
2195
2196         return cpu == -1 ? -1 : env->cpu[cpu].socket_id;
2197 }
2198
2199 static int perf_env__get_core(struct cpu_map *map, int idx, void *data)
2200 {
2201         struct perf_env *env = data;
2202         int core = -1, cpu = perf_env__get_cpu(env, map, idx);
2203
2204         if (cpu != -1) {
2205                 int socket_id = env->cpu[cpu].socket_id;
2206
2207                 /*
2208                  * Encode socket in upper 16 bits
2209                  * core_id is relative to socket, and
2210                  * we need a global id. So we combine
2211                  * socket + core id.
2212                  */
2213                 core = (socket_id << 16) | (env->cpu[cpu].core_id & 0xffff);
2214         }
2215
2216         return core;
2217 }
2218
2219 static int perf_env__build_socket_map(struct perf_env *env, struct cpu_map *cpus,
2220                                       struct cpu_map **sockp)
2221 {
2222         return cpu_map__build_map(cpus, sockp, perf_env__get_socket, env);
2223 }
2224
2225 static int perf_env__build_core_map(struct perf_env *env, struct cpu_map *cpus,
2226                                     struct cpu_map **corep)
2227 {
2228         return cpu_map__build_map(cpus, corep, perf_env__get_core, env);
2229 }
2230
2231 static int perf_stat__get_socket_file(struct cpu_map *map, int idx)
2232 {
2233         return perf_env__get_socket(map, idx, &perf_stat.session->header.env);
2234 }
2235
2236 static int perf_stat__get_core_file(struct cpu_map *map, int idx)
2237 {
2238         return perf_env__get_core(map, idx, &perf_stat.session->header.env);
2239 }
2240
2241 static int perf_stat_init_aggr_mode_file(struct perf_stat *st)
2242 {
2243         struct perf_env *env = &st->session->header.env;
2244
2245         switch (stat_config.aggr_mode) {
2246         case AGGR_SOCKET:
2247                 if (perf_env__build_socket_map(env, evsel_list->cpus, &aggr_map)) {
2248                         perror("cannot build socket map");
2249                         return -1;
2250                 }
2251                 aggr_get_id = perf_stat__get_socket_file;
2252                 break;
2253         case AGGR_CORE:
2254                 if (perf_env__build_core_map(env, evsel_list->cpus, &aggr_map)) {
2255                         perror("cannot build core map");
2256                         return -1;
2257                 }
2258                 aggr_get_id = perf_stat__get_core_file;
2259                 break;
2260         case AGGR_NONE:
2261         case AGGR_GLOBAL:
2262         case AGGR_THREAD:
2263         case AGGR_UNSET:
2264         default:
2265                 break;
2266         }
2267
2268         return 0;
2269 }
2270
2271 static int topdown_filter_events(const char **attr, char **str, bool use_group)
2272 {
2273         int off = 0;
2274         int i;
2275         int len = 0;
2276         char *s;
2277
2278         for (i = 0; attr[i]; i++) {
2279                 if (pmu_have_event("cpu", attr[i])) {
2280                         len += strlen(attr[i]) + 1;
2281                         attr[i - off] = attr[i];
2282                 } else
2283                         off++;
2284         }
2285         attr[i - off] = NULL;
2286
2287         *str = malloc(len + 1 + 2);
2288         if (!*str)
2289                 return -1;
2290         s = *str;
2291         if (i - off == 0) {
2292                 *s = 0;
2293                 return 0;
2294         }
2295         if (use_group)
2296                 *s++ = '{';
2297         for (i = 0; attr[i]; i++) {
2298                 strcpy(s, attr[i]);
2299                 s += strlen(s);
2300                 *s++ = ',';
2301         }
2302         if (use_group) {
2303                 s[-1] = '}';
2304                 *s = 0;
2305         } else
2306                 s[-1] = 0;
2307         return 0;
2308 }
2309
2310 __weak bool arch_topdown_check_group(bool *warn)
2311 {
2312         *warn = false;
2313         return false;
2314 }
2315
2316 __weak void arch_topdown_group_warn(void)
2317 {
2318 }
2319
2320 /*
2321  * Add default attributes, if there were no attributes specified or
2322  * if -d/--detailed, -d -d or -d -d -d is used:
2323  */
2324 static int add_default_attributes(void)
2325 {
2326         int err;
2327         struct perf_event_attr default_attrs0[] = {
2328
2329   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK              },
2330   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES        },
2331   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS          },
2332   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS             },
2333
2334   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES              },
2335 };
2336         struct perf_event_attr frontend_attrs[] = {
2337   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND },
2338 };
2339         struct perf_event_attr backend_attrs[] = {
2340   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND  },
2341 };
2342         struct perf_event_attr default_attrs1[] = {
2343   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS            },
2344   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS     },
2345   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES           },
2346
2347 };
2348
2349 /*
2350  * Detailed stats (-d), covering the L1 and last level data caches:
2351  */
2352         struct perf_event_attr detailed_attrs[] = {
2353
2354   { .type = PERF_TYPE_HW_CACHE,
2355     .config =
2356          PERF_COUNT_HW_CACHE_L1D                <<  0  |
2357         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
2358         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
2359
2360   { .type = PERF_TYPE_HW_CACHE,
2361     .config =
2362          PERF_COUNT_HW_CACHE_L1D                <<  0  |
2363         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
2364         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
2365
2366   { .type = PERF_TYPE_HW_CACHE,
2367     .config =
2368          PERF_COUNT_HW_CACHE_LL                 <<  0  |
2369         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
2370         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
2371
2372   { .type = PERF_TYPE_HW_CACHE,
2373     .config =
2374          PERF_COUNT_HW_CACHE_LL                 <<  0  |
2375         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
2376         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
2377 };
2378
2379 /*
2380  * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
2381  */
2382         struct perf_event_attr very_detailed_attrs[] = {
2383
2384   { .type = PERF_TYPE_HW_CACHE,
2385     .config =
2386          PERF_COUNT_HW_CACHE_L1I                <<  0  |
2387         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
2388         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
2389
2390   { .type = PERF_TYPE_HW_CACHE,
2391     .config =
2392          PERF_COUNT_HW_CACHE_L1I                <<  0  |
2393         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
2394         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
2395
2396   { .type = PERF_TYPE_HW_CACHE,
2397     .config =
2398          PERF_COUNT_HW_CACHE_DTLB               <<  0  |
2399         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
2400         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
2401
2402   { .type = PERF_TYPE_HW_CACHE,
2403     .config =
2404          PERF_COUNT_HW_CACHE_DTLB               <<  0  |
2405         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
2406         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
2407
2408   { .type = PERF_TYPE_HW_CACHE,
2409     .config =
2410          PERF_COUNT_HW_CACHE_ITLB               <<  0  |
2411         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
2412         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
2413
2414   { .type = PERF_TYPE_HW_CACHE,
2415     .config =
2416          PERF_COUNT_HW_CACHE_ITLB               <<  0  |
2417         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
2418         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
2419
2420 };
2421
2422 /*
2423  * Very, very detailed stats (-d -d -d), adding prefetch events:
2424  */
2425         struct perf_event_attr very_very_detailed_attrs[] = {
2426
2427   { .type = PERF_TYPE_HW_CACHE,
2428     .config =
2429          PERF_COUNT_HW_CACHE_L1D                <<  0  |
2430         (PERF_COUNT_HW_CACHE_OP_PREFETCH        <<  8) |
2431         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
2432
2433   { .type = PERF_TYPE_HW_CACHE,
2434     .config =
2435          PERF_COUNT_HW_CACHE_L1D                <<  0  |
2436         (PERF_COUNT_HW_CACHE_OP_PREFETCH        <<  8) |
2437         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
2438 };
2439
2440         /* Set attrs if no event is selected and !null_run: */
2441         if (null_run)
2442                 return 0;
2443
2444         if (transaction_run) {
2445                 struct parse_events_error errinfo;
2446
2447                 if (pmu_have_event("cpu", "cycles-ct") &&
2448                     pmu_have_event("cpu", "el-start"))
2449                         err = parse_events(evsel_list, transaction_attrs,
2450                                            &errinfo);
2451                 else
2452                         err = parse_events(evsel_list,
2453                                            transaction_limited_attrs,
2454                                            &errinfo);
2455                 if (err) {
2456                         fprintf(stderr, "Cannot set up transaction events\n");
2457                         return -1;
2458                 }
2459                 return 0;
2460         }
2461
2462         if (smi_cost) {
2463                 int smi;
2464
2465                 if (sysfs__read_int(FREEZE_ON_SMI_PATH, &smi) < 0) {
2466                         fprintf(stderr, "freeze_on_smi is not supported.\n");
2467                         return -1;
2468                 }
2469
2470                 if (!smi) {
2471                         if (sysfs__write_int(FREEZE_ON_SMI_PATH, 1) < 0) {
2472                                 fprintf(stderr, "Failed to set freeze_on_smi.\n");
2473                                 return -1;
2474                         }
2475                         smi_reset = true;
2476                 }
2477
2478                 if (pmu_have_event("msr", "aperf") &&
2479                     pmu_have_event("msr", "smi")) {
2480                         if (!force_metric_only)
2481                                 metric_only = true;
2482                         err = parse_events(evsel_list, smi_cost_attrs, NULL);
2483                 } else {
2484                         fprintf(stderr, "To measure SMI cost, it needs "
2485                                 "msr/aperf/, msr/smi/ and cpu/cycles/ support\n");
2486                         return -1;
2487                 }
2488                 if (err) {
2489                         fprintf(stderr, "Cannot set up SMI cost events\n");
2490                         return -1;
2491                 }
2492                 return 0;
2493         }
2494
2495         if (topdown_run) {
2496                 char *str = NULL;
2497                 bool warn = false;
2498
2499                 if (stat_config.aggr_mode != AGGR_GLOBAL &&
2500                     stat_config.aggr_mode != AGGR_CORE) {
2501                         pr_err("top down event configuration requires --per-core mode\n");
2502                         return -1;
2503                 }
2504                 stat_config.aggr_mode = AGGR_CORE;
2505                 if (nr_cgroups || !target__has_cpu(&target)) {
2506                         pr_err("top down event configuration requires system-wide mode (-a)\n");
2507                         return -1;
2508                 }
2509
2510                 if (!force_metric_only)
2511                         metric_only = true;
2512                 if (topdown_filter_events(topdown_attrs, &str,
2513                                 arch_topdown_check_group(&warn)) < 0) {
2514                         pr_err("Out of memory\n");
2515                         return -1;
2516                 }
2517                 if (topdown_attrs[0] && str) {
2518                         if (warn)
2519                                 arch_topdown_group_warn();
2520                         err = parse_events(evsel_list, str, NULL);
2521                         if (err) {
2522                                 fprintf(stderr,
2523                                         "Cannot set up top down events %s: %d\n",
2524                                         str, err);
2525                                 free(str);
2526                                 return -1;
2527                         }
2528                 } else {
2529                         fprintf(stderr, "System does not support topdown\n");
2530                         return -1;
2531                 }
2532                 free(str);
2533         }
2534
2535         if (!evsel_list->nr_entries) {
2536                 if (target__has_cpu(&target))
2537                         default_attrs0[0].config = PERF_COUNT_SW_CPU_CLOCK;
2538
2539                 if (perf_evlist__add_default_attrs(evsel_list, default_attrs0) < 0)
2540                         return -1;
2541                 if (pmu_have_event("cpu", "stalled-cycles-frontend")) {
2542                         if (perf_evlist__add_default_attrs(evsel_list,
2543                                                 frontend_attrs) < 0)
2544                                 return -1;
2545                 }
2546                 if (pmu_have_event("cpu", "stalled-cycles-backend")) {
2547                         if (perf_evlist__add_default_attrs(evsel_list,
2548                                                 backend_attrs) < 0)
2549                                 return -1;
2550                 }
2551                 if (perf_evlist__add_default_attrs(evsel_list, default_attrs1) < 0)
2552                         return -1;
2553         }
2554
2555         /* Detailed events get appended to the event list: */
2556
2557         if (detailed_run <  1)
2558                 return 0;
2559
2560         /* Append detailed run extra attributes: */
2561         if (perf_evlist__add_default_attrs(evsel_list, detailed_attrs) < 0)
2562                 return -1;
2563
2564         if (detailed_run < 2)
2565                 return 0;
2566
2567         /* Append very detailed run extra attributes: */
2568         if (perf_evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0)
2569                 return -1;
2570
2571         if (detailed_run < 3)
2572                 return 0;
2573
2574         /* Append very, very detailed run extra attributes: */
2575         return perf_evlist__add_default_attrs(evsel_list, very_very_detailed_attrs);
2576 }
2577
2578 static const char * const stat_record_usage[] = {
2579         "perf stat record [<options>]",
2580         NULL,
2581 };
2582
2583 static void init_features(struct perf_session *session)
2584 {
2585         int feat;
2586
2587         for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++)
2588                 perf_header__set_feat(&session->header, feat);
2589
2590         perf_header__clear_feat(&session->header, HEADER_BUILD_ID);
2591         perf_header__clear_feat(&session->header, HEADER_TRACING_DATA);
2592         perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK);
2593         perf_header__clear_feat(&session->header, HEADER_AUXTRACE);
2594 }
2595
2596 static int __cmd_record(int argc, const char **argv)
2597 {
2598         struct perf_session *session;
2599         struct perf_data *data = &perf_stat.data;
2600
2601         argc = parse_options(argc, argv, stat_options, stat_record_usage,
2602                              PARSE_OPT_STOP_AT_NON_OPTION);
2603
2604         if (output_name)
2605                 data->file.path = output_name;
2606
2607         if (run_count != 1 || forever) {
2608                 pr_err("Cannot use -r option with perf stat record.\n");
2609                 return -1;
2610         }
2611
2612         session = perf_session__new(data, false, NULL);
2613         if (session == NULL) {
2614                 pr_err("Perf session creation failed.\n");
2615                 return -1;
2616         }
2617
2618         init_features(session);
2619
2620         session->evlist   = evsel_list;
2621         perf_stat.session = session;
2622         perf_stat.record  = true;
2623         return argc;
2624 }
2625
2626 static int process_stat_round_event(struct perf_tool *tool __maybe_unused,
2627                                     union perf_event *event,
2628                                     struct perf_session *session)
2629 {
2630         struct stat_round_event *stat_round = &event->stat_round;
2631         struct perf_evsel *counter;
2632         struct timespec tsh, *ts = NULL;
2633         const char **argv = session->header.env.cmdline_argv;
2634         int argc = session->header.env.nr_cmdline;
2635
2636         evlist__for_each_entry(evsel_list, counter)
2637                 perf_stat_process_counter(&stat_config, counter);
2638
2639         if (stat_round->type == PERF_STAT_ROUND_TYPE__FINAL)
2640                 update_stats(&walltime_nsecs_stats, stat_round->time);
2641
2642         if (stat_config.interval && stat_round->time) {
2643                 tsh.tv_sec  = stat_round->time / NSEC_PER_SEC;
2644                 tsh.tv_nsec = stat_round->time % NSEC_PER_SEC;
2645                 ts = &tsh;
2646         }
2647
2648         print_counters(ts, argc, argv);
2649         return 0;
2650 }
2651
2652 static
2653 int process_stat_config_event(struct perf_tool *tool,
2654                               union perf_event *event,
2655                               struct perf_session *session __maybe_unused)
2656 {
2657         struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2658
2659         perf_event__read_stat_config(&stat_config, &event->stat_config);
2660
2661         if (cpu_map__empty(st->cpus)) {
2662                 if (st->aggr_mode != AGGR_UNSET)
2663                         pr_warning("warning: processing task data, aggregation mode not set\n");
2664                 return 0;
2665         }
2666
2667         if (st->aggr_mode != AGGR_UNSET)
2668                 stat_config.aggr_mode = st->aggr_mode;
2669
2670         if (perf_stat.data.is_pipe)
2671                 perf_stat_init_aggr_mode();
2672         else
2673                 perf_stat_init_aggr_mode_file(st);
2674
2675         return 0;
2676 }
2677
2678 static int set_maps(struct perf_stat *st)
2679 {
2680         if (!st->cpus || !st->threads)
2681                 return 0;
2682
2683         if (WARN_ONCE(st->maps_allocated, "stats double allocation\n"))
2684                 return -EINVAL;
2685
2686         perf_evlist__set_maps(evsel_list, st->cpus, st->threads);
2687
2688         if (perf_evlist__alloc_stats(evsel_list, true))
2689                 return -ENOMEM;
2690
2691         st->maps_allocated = true;
2692         return 0;
2693 }
2694
2695 static
2696 int process_thread_map_event(struct perf_tool *tool,
2697                              union perf_event *event,
2698                              struct perf_session *session __maybe_unused)
2699 {
2700         struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2701
2702         if (st->threads) {
2703                 pr_warning("Extra thread map event, ignoring.\n");
2704                 return 0;
2705         }
2706
2707         st->threads = thread_map__new_event(&event->thread_map);
2708         if (!st->threads)
2709                 return -ENOMEM;
2710
2711         return set_maps(st);
2712 }
2713
2714 static
2715 int process_cpu_map_event(struct perf_tool *tool,
2716                           union perf_event *event,
2717                           struct perf_session *session __maybe_unused)
2718 {
2719         struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2720         struct cpu_map *cpus;
2721
2722         if (st->cpus) {
2723                 pr_warning("Extra cpu map event, ignoring.\n");
2724                 return 0;
2725         }
2726
2727         cpus = cpu_map__new_data(&event->cpu_map.data);
2728         if (!cpus)
2729                 return -ENOMEM;
2730
2731         st->cpus = cpus;
2732         return set_maps(st);
2733 }
2734
2735 static int runtime_stat_new(struct perf_stat_config *config, int nthreads)
2736 {
2737         int i;
2738
2739         config->stats = calloc(nthreads, sizeof(struct runtime_stat));
2740         if (!config->stats)
2741                 return -1;
2742
2743         config->stats_num = nthreads;
2744
2745         for (i = 0; i < nthreads; i++)
2746                 runtime_stat__init(&config->stats[i]);
2747
2748         return 0;
2749 }
2750
2751 static void runtime_stat_delete(struct perf_stat_config *config)
2752 {
2753         int i;
2754
2755         if (!config->stats)
2756                 return;
2757
2758         for (i = 0; i < config->stats_num; i++)
2759                 runtime_stat__exit(&config->stats[i]);
2760
2761         free(config->stats);
2762 }
2763
2764 static const char * const stat_report_usage[] = {
2765         "perf stat report [<options>]",
2766         NULL,
2767 };
2768
2769 static struct perf_stat perf_stat = {
2770         .tool = {
2771                 .attr           = perf_event__process_attr,
2772                 .event_update   = perf_event__process_event_update,
2773                 .thread_map     = process_thread_map_event,
2774                 .cpu_map        = process_cpu_map_event,
2775                 .stat_config    = process_stat_config_event,
2776                 .stat           = perf_event__process_stat_event,
2777                 .stat_round     = process_stat_round_event,
2778         },
2779         .aggr_mode = AGGR_UNSET,
2780 };
2781
2782 static int __cmd_report(int argc, const char **argv)
2783 {
2784         struct perf_session *session;
2785         const struct option options[] = {
2786         OPT_STRING('i', "input", &input_name, "file", "input file name"),
2787         OPT_SET_UINT(0, "per-socket", &perf_stat.aggr_mode,
2788                      "aggregate counts per processor socket", AGGR_SOCKET),
2789         OPT_SET_UINT(0, "per-core", &perf_stat.aggr_mode,
2790                      "aggregate counts per physical processor core", AGGR_CORE),
2791         OPT_SET_UINT('A', "no-aggr", &perf_stat.aggr_mode,
2792                      "disable CPU count aggregation", AGGR_NONE),
2793         OPT_END()
2794         };
2795         struct stat st;
2796         int ret;
2797
2798         argc = parse_options(argc, argv, options, stat_report_usage, 0);
2799
2800         if (!input_name || !strlen(input_name)) {
2801                 if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode))
2802                         input_name = "-";
2803                 else
2804                         input_name = "perf.data";
2805         }
2806
2807         perf_stat.data.file.path = input_name;
2808         perf_stat.data.mode      = PERF_DATA_MODE_READ;
2809
2810         session = perf_session__new(&perf_stat.data, false, &perf_stat.tool);
2811         if (session == NULL)
2812                 return -1;
2813
2814         perf_stat.session  = session;
2815         stat_config.output = stderr;
2816         evsel_list         = session->evlist;
2817
2818         ret = perf_session__process_events(session);
2819         if (ret)
2820                 return ret;
2821
2822         perf_session__delete(session);
2823         return 0;
2824 }
2825
2826 static void setup_system_wide(int forks)
2827 {
2828         /*
2829          * Make system wide (-a) the default target if
2830          * no target was specified and one of following
2831          * conditions is met:
2832          *
2833          *   - there's no workload specified
2834          *   - there is workload specified but all requested
2835          *     events are system wide events
2836          */
2837         if (!target__none(&target))
2838                 return;
2839
2840         if (!forks)
2841                 target.system_wide = true;
2842         else {
2843                 struct perf_evsel *counter;
2844
2845                 evlist__for_each_entry(evsel_list, counter) {
2846                         if (!counter->system_wide)
2847                                 return;
2848                 }
2849
2850                 if (evsel_list->nr_entries)
2851                         target.system_wide = true;
2852         }
2853 }
2854
2855 int cmd_stat(int argc, const char **argv)
2856 {
2857         const char * const stat_usage[] = {
2858                 "perf stat [<options>] [<command>]",
2859                 NULL
2860         };
2861         int status = -EINVAL, run_idx;
2862         const char *mode;
2863         FILE *output = stderr;
2864         unsigned int interval, timeout;
2865         const char * const stat_subcommands[] = { "record", "report" };
2866
2867         setlocale(LC_ALL, "");
2868
2869         evsel_list = perf_evlist__new();
2870         if (evsel_list == NULL)
2871                 return -ENOMEM;
2872
2873         parse_events__shrink_config_terms();
2874         argc = parse_options_subcommand(argc, argv, stat_options, stat_subcommands,
2875                                         (const char **) stat_usage,
2876                                         PARSE_OPT_STOP_AT_NON_OPTION);
2877         perf_stat__collect_metric_expr(evsel_list);
2878         perf_stat__init_shadow_stats();
2879
2880         if (csv_sep) {
2881                 csv_output = true;
2882                 if (!strcmp(csv_sep, "\\t"))
2883                         csv_sep = "\t";
2884         } else
2885                 csv_sep = DEFAULT_SEPARATOR;
2886
2887         if (argc && !strncmp(argv[0], "rec", 3)) {
2888                 argc = __cmd_record(argc, argv);
2889                 if (argc < 0)
2890                         return -1;
2891         } else if (argc && !strncmp(argv[0], "rep", 3))
2892                 return __cmd_report(argc, argv);
2893
2894         interval = stat_config.interval;
2895         timeout = stat_config.timeout;
2896
2897         /*
2898          * For record command the -o is already taken care of.
2899          */
2900         if (!STAT_RECORD && output_name && strcmp(output_name, "-"))
2901                 output = NULL;
2902
2903         if (output_name && output_fd) {
2904                 fprintf(stderr, "cannot use both --output and --log-fd\n");
2905                 parse_options_usage(stat_usage, stat_options, "o", 1);
2906                 parse_options_usage(NULL, stat_options, "log-fd", 0);
2907                 goto out;
2908         }
2909
2910         if (metric_only && stat_config.aggr_mode == AGGR_THREAD) {
2911                 fprintf(stderr, "--metric-only is not supported with --per-thread\n");
2912                 goto out;
2913         }
2914
2915         if (metric_only && run_count > 1) {
2916                 fprintf(stderr, "--metric-only is not supported with -r\n");
2917                 goto out;
2918         }
2919
2920         if (walltime_run_table && run_count <= 1) {
2921                 fprintf(stderr, "--table is only supported with -r\n");
2922                 parse_options_usage(stat_usage, stat_options, "r", 1);
2923                 parse_options_usage(NULL, stat_options, "table", 0);
2924                 goto out;
2925         }
2926
2927         if (output_fd < 0) {
2928                 fprintf(stderr, "argument to --log-fd must be a > 0\n");
2929                 parse_options_usage(stat_usage, stat_options, "log-fd", 0);
2930                 goto out;
2931         }
2932
2933         if (!output) {
2934                 struct timespec tm;
2935                 mode = append_file ? "a" : "w";
2936
2937                 output = fopen(output_name, mode);
2938                 if (!output) {
2939                         perror("failed to create output file");
2940                         return -1;
2941                 }
2942                 clock_gettime(CLOCK_REALTIME, &tm);
2943                 fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
2944         } else if (output_fd > 0) {
2945                 mode = append_file ? "a" : "w";
2946                 output = fdopen(output_fd, mode);
2947                 if (!output) {
2948                         perror("Failed opening logfd");
2949                         return -errno;
2950                 }
2951         }
2952
2953         stat_config.output = output;
2954
2955         /*
2956          * let the spreadsheet do the pretty-printing
2957          */
2958         if (csv_output) {
2959                 /* User explicitly passed -B? */
2960                 if (big_num_opt == 1) {
2961                         fprintf(stderr, "-B option not supported with -x\n");
2962                         parse_options_usage(stat_usage, stat_options, "B", 1);
2963                         parse_options_usage(NULL, stat_options, "x", 1);
2964                         goto out;
2965                 } else /* Nope, so disable big number formatting */
2966                         big_num = false;
2967         } else if (big_num_opt == 0) /* User passed --no-big-num */
2968                 big_num = false;
2969
2970         setup_system_wide(argc);
2971
2972         /*
2973          * Display user/system times only for single
2974          * run and when there's specified tracee.
2975          */
2976         if ((run_count == 1) && target__none(&target))
2977                 ru_display = true;
2978
2979         if (run_count < 0) {
2980                 pr_err("Run count must be a positive number\n");
2981                 parse_options_usage(stat_usage, stat_options, "r", 1);
2982                 goto out;
2983         } else if (run_count == 0) {
2984                 forever = true;
2985                 run_count = 1;
2986         }
2987
2988         if (walltime_run_table) {
2989                 walltime_run = zalloc(run_count * sizeof(walltime_run[0]));
2990                 if (!walltime_run) {
2991                         pr_err("failed to setup -r option");
2992                         goto out;
2993                 }
2994         }
2995
2996         if ((stat_config.aggr_mode == AGGR_THREAD) &&
2997                 !target__has_task(&target)) {
2998                 if (!target.system_wide || target.cpu_list) {
2999                         fprintf(stderr, "The --per-thread option is only "
3000                                 "available when monitoring via -p -t -a "
3001                                 "options or only --per-thread.\n");
3002                         parse_options_usage(NULL, stat_options, "p", 1);
3003                         parse_options_usage(NULL, stat_options, "t", 1);
3004                         goto out;
3005                 }
3006         }
3007
3008         /*
3009          * no_aggr, cgroup are for system-wide only
3010          * --per-thread is aggregated per thread, we dont mix it with cpu mode
3011          */
3012         if (((stat_config.aggr_mode != AGGR_GLOBAL &&
3013               stat_config.aggr_mode != AGGR_THREAD) || nr_cgroups) &&
3014             !target__has_cpu(&target)) {
3015                 fprintf(stderr, "both cgroup and no-aggregation "
3016                         "modes only available in system-wide mode\n");
3017
3018                 parse_options_usage(stat_usage, stat_options, "G", 1);
3019                 parse_options_usage(NULL, stat_options, "A", 1);
3020                 parse_options_usage(NULL, stat_options, "a", 1);
3021                 goto out;
3022         }
3023
3024         if (add_default_attributes())
3025                 goto out;
3026
3027         target__validate(&target);
3028
3029         if ((stat_config.aggr_mode == AGGR_THREAD) && (target.system_wide))
3030                 target.per_thread = true;
3031
3032         if (perf_evlist__create_maps(evsel_list, &target) < 0) {
3033                 if (target__has_task(&target)) {
3034                         pr_err("Problems finding threads of monitor\n");
3035                         parse_options_usage(stat_usage, stat_options, "p", 1);
3036                         parse_options_usage(NULL, stat_options, "t", 1);
3037                 } else if (target__has_cpu(&target)) {
3038                         perror("failed to parse CPUs map");
3039                         parse_options_usage(stat_usage, stat_options, "C", 1);
3040                         parse_options_usage(NULL, stat_options, "a", 1);
3041                 }
3042                 goto out;
3043         }
3044
3045         /*
3046          * Initialize thread_map with comm names,
3047          * so we could print it out on output.
3048          */
3049         if (stat_config.aggr_mode == AGGR_THREAD) {
3050                 thread_map__read_comms(evsel_list->threads);
3051                 if (target.system_wide) {
3052                         if (runtime_stat_new(&stat_config,
3053                                 thread_map__nr(evsel_list->threads))) {
3054                                 goto out;
3055                         }
3056                 }
3057         }
3058
3059         if (stat_config.times && interval)
3060                 interval_count = true;
3061         else if (stat_config.times && !interval) {
3062                 pr_err("interval-count option should be used together with "
3063                                 "interval-print.\n");
3064                 parse_options_usage(stat_usage, stat_options, "interval-count", 0);
3065                 parse_options_usage(stat_usage, stat_options, "I", 1);
3066                 goto out;
3067         }
3068
3069         if (timeout && timeout < 100) {
3070                 if (timeout < 10) {
3071                         pr_err("timeout must be >= 10ms.\n");
3072                         parse_options_usage(stat_usage, stat_options, "timeout", 0);
3073                         goto out;
3074                 } else
3075                         pr_warning("timeout < 100ms. "
3076                                    "The overhead percentage could be high in some cases. "
3077                                    "Please proceed with caution.\n");
3078         }
3079         if (timeout && interval) {
3080                 pr_err("timeout option is not supported with interval-print.\n");
3081                 parse_options_usage(stat_usage, stat_options, "timeout", 0);
3082                 parse_options_usage(stat_usage, stat_options, "I", 1);
3083                 goto out;
3084         }
3085
3086         if (perf_evlist__alloc_stats(evsel_list, interval))
3087                 goto out;
3088
3089         if (perf_stat_init_aggr_mode())
3090                 goto out;
3091
3092         /*
3093          * We dont want to block the signals - that would cause
3094          * child tasks to inherit that and Ctrl-C would not work.
3095          * What we want is for Ctrl-C to work in the exec()-ed
3096          * task, but being ignored by perf stat itself:
3097          */
3098         atexit(sig_atexit);
3099         if (!forever)
3100                 signal(SIGINT,  skip_signal);
3101         signal(SIGCHLD, skip_signal);
3102         signal(SIGALRM, skip_signal);
3103         signal(SIGABRT, skip_signal);
3104
3105         status = 0;
3106         for (run_idx = 0; forever || run_idx < run_count; run_idx++) {
3107                 if (run_count != 1 && verbose > 0)
3108                         fprintf(output, "[ perf stat: executing run #%d ... ]\n",
3109                                 run_idx + 1);
3110
3111                 status = run_perf_stat(argc, argv, run_idx);
3112                 if (forever && status != -1) {
3113                         print_counters(NULL, argc, argv);
3114                         perf_stat__reset_stats();
3115                 }
3116         }
3117
3118         if (!forever && status != -1 && !interval)
3119                 print_counters(NULL, argc, argv);
3120
3121         if (STAT_RECORD) {
3122                 /*
3123                  * We synthesize the kernel mmap record just so that older tools
3124                  * don't emit warnings about not being able to resolve symbols
3125                  * due to /proc/sys/kernel/kptr_restrict settings and instear provide
3126                  * a saner message about no samples being in the perf.data file.
3127                  *
3128                  * This also serves to suppress a warning about f_header.data.size == 0
3129                  * in header.c at the moment 'perf stat record' gets introduced, which
3130                  * is not really needed once we start adding the stat specific PERF_RECORD_
3131                  * records, but the need to suppress the kptr_restrict messages in older
3132                  * tools remain  -acme
3133                  */
3134                 int fd = perf_data__fd(&perf_stat.data);
3135                 int err = perf_event__synthesize_kernel_mmap((void *)&perf_stat,
3136                                                              process_synthesized_event,
3137                                                              &perf_stat.session->machines.host);
3138                 if (err) {
3139                         pr_warning("Couldn't synthesize the kernel mmap record, harmless, "
3140                                    "older tools may produce warnings about this file\n.");
3141                 }
3142
3143                 if (!interval) {
3144                         if (WRITE_STAT_ROUND_EVENT(walltime_nsecs_stats.max, FINAL))
3145                                 pr_err("failed to write stat round event\n");
3146                 }
3147
3148                 if (!perf_stat.data.is_pipe) {
3149                         perf_stat.session->header.data_size += perf_stat.bytes_written;
3150                         perf_session__write_header(perf_stat.session, evsel_list, fd, true);
3151                 }
3152
3153                 perf_session__delete(perf_stat.session);
3154         }
3155
3156         perf_stat__exit_aggr_mode();
3157         perf_evlist__free_stats(evsel_list);
3158 out:
3159         free(walltime_run);
3160
3161         if (smi_cost && smi_reset)
3162                 sysfs__write_int(FREEZE_ON_SMI_PATH, 0);
3163
3164         perf_evlist__delete(evsel_list);
3165
3166         runtime_stat_delete(&stat_config);
3167
3168         return status;
3169 }