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/color.h"
  56 #include "util/stat.h"
  57 #include "util/header.h"
  58 #include "util/cpumap.h"
  59 #include "util/thread.h"
  60 #include "util/thread_map.h"
  61 #include "util/counts.h"
  62 #include "util/session.h"
  63 #include "util/tool.h"
  64 #include "asm/bug.h"
  65
  66 #include <stdlib.h>
  67 #include <sys/prctl.h>
  68 #include <locale.h>
  69 #include <math.h>
  70
  71 #define DEFAULT_SEPARATOR       " "
  72 #define CNTR_NOT_SUPPORTED      "<not supported>"
  73 #define CNTR_NOT_COUNTED        "<not counted>"
  74
  75 static void print_counters(struct timespec *ts, int argc, const char **argv);
  76
  77 /* Default events used for perf stat -T */
  78 static const char *transaction_attrs = {
  79         "task-clock,"
  80         "{"
  81         "instructions,"
  82         "cycles,"
  83         "cpu/cycles-t/,"
  84         "cpu/tx-start/,"
  85         "cpu/el-start/,"
  86         "cpu/cycles-ct/"
  87         "}"
  88 };
  89
  90 /* More limited version when the CPU does not have all events. */
  91 static const char * transaction_limited_attrs = {
  92         "task-clock,"
  93         "{"
  94         "instructions,"
  95         "cycles,"
  96         "cpu/cycles-t/,"
  97         "cpu/tx-start/"
  98         "}"
  99 };
 100
 101 static struct perf_evlist       *evsel_list;
 102
 103 static struct target target = {
 104         .uid    = UINT_MAX,
 105 };
 106
 107 typedef int (*aggr_get_id_t)(struct cpu_map *m, int cpu);
 108
 109 static int                      run_count                       =  1;
 110 static bool                     no_inherit                      = false;
 111 static volatile pid_t           child_pid                       = -1;
 112 static bool                     null_run                        =  false;
 113 static int                      detailed_run                    =  0;
 114 static bool                     transaction_run;
 115 static bool                     big_num                         =  true;
 116 static int                      big_num_opt                     =  -1;
 117 static const char               *csv_sep                        = NULL;
 118 static bool                     csv_output                      = false;
 119 static bool                     group                           = false;
 120 static const char               *pre_cmd                        = NULL;
 121 static const char               *post_cmd                       = NULL;
 122 static bool                     sync_run                        = false;
 123 static unsigned int             initial_delay                   = 0;
 124 static unsigned int             unit_width                      = 4; /* strlen("unit") */
 125 static bool                     forever                         = false;
 126 static bool                     metric_only                     = false;
 127 static struct timespec          ref_time;
 128 static struct cpu_map           *aggr_map;
 129 static aggr_get_id_t            aggr_get_id;
 130 static bool                     append_file;
 131 static const char               *output_name;
 132 static int                      output_fd;
 133
 134 struct perf_stat {
 135         bool                     record;
 136         struct perf_data_file    file;
 137         struct perf_session     *session;
 138         u64                      bytes_written;
 139         struct perf_tool         tool;
 140         bool                     maps_allocated;
 141         struct cpu_map          *cpus;
 142         struct thread_map       *threads;
 143         enum aggr_mode           aggr_mode;
 144 };
 145
 146 static struct perf_stat         perf_stat;
 147 #define STAT_RECORD             perf_stat.record
 148
 149 static volatile int done = 0;
 150
 151 static struct perf_stat_config stat_config = {
 152         .aggr_mode      = AGGR_GLOBAL,
 153         .scale          = true,
 154 };
 155
 156 static inline void diff_timespec(struct timespec *r, struct timespec *a,
 157                                  struct timespec *b)
 158 {
 159         r->tv_sec = a->tv_sec - b->tv_sec;
 160         if (a->tv_nsec < b->tv_nsec) {
 161                 r->tv_nsec = a->tv_nsec + 1000000000L - b->tv_nsec;
 162                 r->tv_sec--;
 163         } else {
 164                 r->tv_nsec = a->tv_nsec - b->tv_nsec ;
 165         }
 166 }
 167
 168 static void perf_stat__reset_stats(void)
 169 {
 170         perf_evlist__reset_stats(evsel_list);
 171         perf_stat__reset_shadow_stats();
 172 }
 173
 174 static int create_perf_stat_counter(struct perf_evsel *evsel)
 175 {
 176         struct perf_event_attr *attr = &evsel->attr;
 177
 178         if (stat_config.scale)
 179                 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
 180                                     PERF_FORMAT_TOTAL_TIME_RUNNING;
 181
 182         attr->inherit = !no_inherit;
 183
 184         /*
 185          * Some events get initialized with sample_(period/type) set,
 186          * like tracepoints. Clear it up for counting.
 187          */
 188         attr->sample_period = 0;
 189
 190         /*
 191          * But set sample_type to PERF_SAMPLE_IDENTIFIER, which should be harmless
 192          * while avoiding that older tools show confusing messages.
 193          *
 194          * However for pipe sessions we need to keep it zero,
 195          * because script's perf_evsel__check_attr is triggered
 196          * by attr->sample_type != 0, and we can't run it on
 197          * stat sessions.
 198          */
 199         if (!(STAT_RECORD && perf_stat.file.is_pipe))
 200                 attr->sample_type = PERF_SAMPLE_IDENTIFIER;
 201
 202         /*
 203          * Disabling all counters initially, they will be enabled
 204          * either manually by us or by kernel via enable_on_exec
 205          * set later.
 206          */
 207         if (perf_evsel__is_group_leader(evsel)) {
 208                 attr->disabled = 1;
 209
 210                 /*
 211                  * In case of initial_delay we enable tracee
 212                  * events manually.
 213                  */
 214                 if (target__none(&target) && !initial_delay)
 215                         attr->enable_on_exec = 1;
 216         }
 217
 218         if (target__has_cpu(&target))
 219                 return perf_evsel__open_per_cpu(evsel, perf_evsel__cpus(evsel));
 220
 221         return perf_evsel__open_per_thread(evsel, evsel_list->threads);
 222 }
 223
 224 /*
 225  * Does the counter have nsecs as a unit?
 226  */
 227 static inline int nsec_counter(struct perf_evsel *evsel)
 228 {
 229         if (perf_evsel__match(evsel, SOFTWARE, SW_CPU_CLOCK) ||
 230             perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
 231                 return 1;
 232
 233         return 0;
 234 }
 235
 236 static int process_synthesized_event(struct perf_tool *tool __maybe_unused,
 237                                      union perf_event *event,
 238                                      struct perf_sample *sample __maybe_unused,
 239                                      struct machine *machine __maybe_unused)
 240 {
 241         if (perf_data_file__write(&perf_stat.file, event, event->header.size) < 0) {
 242                 pr_err("failed to write perf data, error: %m\n");
 243                 return -1;
 244         }
 245
 246         perf_stat.bytes_written += event->header.size;
 247         return 0;
 248 }
 249
 250 static int write_stat_round_event(u64 tm, u64 type)
 251 {
 252         return perf_event__synthesize_stat_round(NULL, tm, type,
 253                                                  process_synthesized_event,
 254                                                  NULL);
 255 }
 256
 257 #define WRITE_STAT_ROUND_EVENT(time, interval) \
 258         write_stat_round_event(time, PERF_STAT_ROUND_TYPE__ ## interval)
 259
 260 #define SID(e, x, y) xyarray__entry(e->sample_id, x, y)
 261
 262 static int
 263 perf_evsel__write_stat_event(struct perf_evsel *counter, u32 cpu, u32 thread,
 264                              struct perf_counts_values *count)
 265 {
 266         struct perf_sample_id *sid = SID(counter, cpu, thread);
 267
 268         return perf_event__synthesize_stat(NULL, cpu, thread, sid->id, count,
 269                                            process_synthesized_event, NULL);
 270 }
 271
 272 /*
 273  * Read out the results of a single counter:
 274  * do not aggregate counts across CPUs in system-wide mode
 275  */
 276 static int read_counter(struct perf_evsel *counter)
 277 {
 278         int nthreads = thread_map__nr(evsel_list->threads);
 279         int ncpus = perf_evsel__nr_cpus(counter);
 280         int cpu, thread;
 281
 282         if (!counter->supported)
 283                 return -ENOENT;
 284
 285         if (counter->system_wide)
 286                 nthreads = 1;
 287
 288         for (thread = 0; thread < nthreads; thread++) {
 289                 for (cpu = 0; cpu < ncpus; cpu++) {
 290                         struct perf_counts_values *count;
 291
 292                         count = perf_counts(counter->counts, cpu, thread);
 293                         if (perf_evsel__read(counter, cpu, thread, count))
 294                                 return -1;
 295
 296                         if (STAT_RECORD) {
 297                                 if (perf_evsel__write_stat_event(counter, cpu, thread, count)) {
 298                                         pr_err("failed to write stat event\n");
 299                                         return -1;
 300                                 }
 301                         }
 302
 303                         if (verbose > 1) {
 304                                 fprintf(stat_config.output,
 305                                         "%s: %d: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
 306                                                 perf_evsel__name(counter),
 307                                                 cpu,
 308                                                 count->val, count->ena, count->run);
 309                         }
 310                 }
 311         }
 312
 313         return 0;
 314 }
 315
 316 static void read_counters(bool close_counters)
 317 {
 318         struct perf_evsel *counter;
 319
 320         evlist__for_each(evsel_list, counter) {
 321                 if (read_counter(counter))
 322                         pr_debug("failed to read counter %s\n", counter->name);
 323
 324                 if (perf_stat_process_counter(&stat_config, counter))
 325                         pr_warning("failed to process counter %s\n", counter->name);
 326
 327                 if (close_counters) {
 328                         perf_evsel__close_fd(counter, perf_evsel__nr_cpus(counter),
 329                                              thread_map__nr(evsel_list->threads));
 330                 }
 331         }
 332 }
 333
 334 static void process_interval(void)
 335 {
 336         struct timespec ts, rs;
 337
 338         read_counters(false);
 339
 340         clock_gettime(CLOCK_MONOTONIC, &ts);
 341         diff_timespec(&rs, &ts, &ref_time);
 342
 343         if (STAT_RECORD) {
 344                 if (WRITE_STAT_ROUND_EVENT(rs.tv_sec * NSECS_PER_SEC + rs.tv_nsec, INTERVAL))
 345                         pr_err("failed to write stat round event\n");
 346         }
 347
 348         print_counters(&rs, 0, NULL);
 349 }
 350
 351 static void enable_counters(void)
 352 {
 353         if (initial_delay)
 354                 usleep(initial_delay * 1000);
 355
 356         /*
 357          * We need to enable counters only if:
 358          * - we don't have tracee (attaching to task or cpu)
 359          * - we have initial delay configured
 360          */
 361         if (!target__none(&target) || initial_delay)
 362                 perf_evlist__enable(evsel_list);
 363 }
 364
 365 static volatile int workload_exec_errno;
 366
 367 /*
 368  * perf_evlist__prepare_workload will send a SIGUSR1
 369  * if the fork fails, since we asked by setting its
 370  * want_signal to true.
 371  */
 372 static void workload_exec_failed_signal(int signo __maybe_unused, siginfo_t *info,
 373                                         void *ucontext __maybe_unused)
 374 {
 375         workload_exec_errno = info->si_value.sival_int;
 376 }
 377
 378 static bool has_unit(struct perf_evsel *counter)
 379 {
 380         return counter->unit && *counter->unit;
 381 }
 382
 383 static bool has_scale(struct perf_evsel *counter)
 384 {
 385         return counter->scale != 1;
 386 }
 387
 388 static int perf_stat_synthesize_config(bool is_pipe)
 389 {
 390         struct perf_evsel *counter;
 391         int err;
 392
 393         if (is_pipe) {
 394                 err = perf_event__synthesize_attrs(NULL, perf_stat.session,
 395                                                    process_synthesized_event);
 396                 if (err < 0) {
 397                         pr_err("Couldn't synthesize attrs.\n");
 398                         return err;
 399                 }
 400         }
 401
 402         /*
 403          * Synthesize other events stuff not carried within
 404          * attr event - unit, scale, name
 405          */
 406         evlist__for_each(evsel_list, counter) {
 407                 if (!counter->supported)
 408                         continue;
 409
 410                 /*
 411                  * Synthesize unit and scale only if it's defined.
 412                  */
 413                 if (has_unit(counter)) {
 414                         err = perf_event__synthesize_event_update_unit(NULL, counter, process_synthesized_event);
 415                         if (err < 0) {
 416                                 pr_err("Couldn't synthesize evsel unit.\n");
 417                                 return err;
 418                         }
 419                 }
 420
 421                 if (has_scale(counter)) {
 422                         err = perf_event__synthesize_event_update_scale(NULL, counter, process_synthesized_event);
 423                         if (err < 0) {
 424                                 pr_err("Couldn't synthesize evsel scale.\n");
 425                                 return err;
 426                         }
 427                 }
 428
 429                 if (counter->own_cpus) {
 430                         err = perf_event__synthesize_event_update_cpus(NULL, counter, process_synthesized_event);
 431                         if (err < 0) {
 432                                 pr_err("Couldn't synthesize evsel scale.\n");
 433                                 return err;
 434                         }
 435                 }
 436
 437                 /*
 438                  * Name is needed only for pipe output,
 439                  * perf.data carries event names.
 440                  */
 441                 if (is_pipe) {
 442                         err = perf_event__synthesize_event_update_name(NULL, counter, process_synthesized_event);
 443                         if (err < 0) {
 444                                 pr_err("Couldn't synthesize evsel name.\n");
 445                                 return err;
 446                         }
 447                 }
 448         }
 449
 450         err = perf_event__synthesize_thread_map2(NULL, evsel_list->threads,
 451                                                 process_synthesized_event,
 452                                                 NULL);
 453         if (err < 0) {
 454                 pr_err("Couldn't synthesize thread map.\n");
 455                 return err;
 456         }
 457
 458         err = perf_event__synthesize_cpu_map(NULL, evsel_list->cpus,
 459                                              process_synthesized_event, NULL);
 460         if (err < 0) {
 461                 pr_err("Couldn't synthesize thread map.\n");
 462                 return err;
 463         }
 464
 465         err = perf_event__synthesize_stat_config(NULL, &stat_config,
 466                                                  process_synthesized_event, NULL);
 467         if (err < 0) {
 468                 pr_err("Couldn't synthesize config.\n");
 469                 return err;
 470         }
 471
 472         return 0;
 473 }
 474
 475 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
 476
 477 static int __store_counter_ids(struct perf_evsel *counter,
 478                                struct cpu_map *cpus,
 479                                struct thread_map *threads)
 480 {
 481         int cpu, thread;
 482
 483         for (cpu = 0; cpu < cpus->nr; cpu++) {
 484                 for (thread = 0; thread < threads->nr; thread++) {
 485                         int fd = FD(counter, cpu, thread);
 486
 487                         if (perf_evlist__id_add_fd(evsel_list, counter,
 488                                                    cpu, thread, fd) < 0)
 489                                 return -1;
 490                 }
 491         }
 492
 493         return 0;
 494 }
 495
 496 static int store_counter_ids(struct perf_evsel *counter)
 497 {
 498         struct cpu_map *cpus = counter->cpus;
 499         struct thread_map *threads = counter->threads;
 500
 501         if (perf_evsel__alloc_id(counter, cpus->nr, threads->nr))
 502                 return -ENOMEM;
 503
 504         return __store_counter_ids(counter, cpus, threads);
 505 }
 506
 507 static int __run_perf_stat(int argc, const char **argv)
 508 {
 509         int interval = stat_config.interval;
 510         char msg[512];
 511         unsigned long long t0, t1;
 512         struct perf_evsel *counter;
 513         struct timespec ts;
 514         size_t l;
 515         int status = 0;
 516         const bool forks = (argc > 0);
 517         bool is_pipe = STAT_RECORD ? perf_stat.file.is_pipe : false;
 518
 519         if (interval) {
 520                 ts.tv_sec  = interval / 1000;
 521                 ts.tv_nsec = (interval % 1000) * 1000000;
 522         } else {
 523                 ts.tv_sec  = 1;
 524                 ts.tv_nsec = 0;
 525         }
 526
 527         if (forks) {
 528                 if (perf_evlist__prepare_workload(evsel_list, &target, argv, is_pipe,
 529                                                   workload_exec_failed_signal) < 0) {
 530                         perror("failed to prepare workload");
 531                         return -1;
 532                 }
 533                 child_pid = evsel_list->workload.pid;
 534         }
 535
 536         if (group)
 537                 perf_evlist__set_leader(evsel_list);
 538
 539         evlist__for_each(evsel_list, counter) {
 540                 if (create_perf_stat_counter(counter) < 0) {
 541                         /*
 542                          * PPC returns ENXIO for HW counters until 2.6.37
 543                          * (behavior changed with commit b0a873e).
 544                          */
 545                         if (errno == EINVAL || errno == ENOSYS ||
 546                             errno == ENOENT || errno == EOPNOTSUPP ||
 547                             errno == ENXIO) {
 548                                 if (verbose)
 549                                         ui__warning("%s event is not supported by the kernel.\n",
 550                                                     perf_evsel__name(counter));
 551                                 counter->supported = false;
 552
 553                                 if ((counter->leader != counter) ||
 554                                     !(counter->leader->nr_members > 1))
 555                                         continue;
 556                         }
 557
 558                         perf_evsel__open_strerror(counter, &target,
 559                                                   errno, msg, sizeof(msg));
 560                         ui__error("%s\n", msg);
 561
 562                         if (child_pid != -1)
 563                                 kill(child_pid, SIGTERM);
 564
 565                         return -1;
 566                 }
 567                 counter->supported = true;
 568
 569                 l = strlen(counter->unit);
 570                 if (l > unit_width)
 571                         unit_width = l;
 572
 573                 if (STAT_RECORD && store_counter_ids(counter))
 574                         return -1;
 575         }
 576
 577         if (perf_evlist__apply_filters(evsel_list, &counter)) {
 578                 error("failed to set filter \"%s\" on event %s with %d (%s)\n",
 579                         counter->filter, perf_evsel__name(counter), errno,
 580                         strerror_r(errno, msg, sizeof(msg)));
 581                 return -1;
 582         }
 583
 584         if (STAT_RECORD) {
 585                 int err, fd = perf_data_file__fd(&perf_stat.file);
 586
 587                 if (is_pipe) {
 588                         err = perf_header__write_pipe(perf_data_file__fd(&perf_stat.file));
 589                 } else {
 590                         err = perf_session__write_header(perf_stat.session, evsel_list,
 591                                                          fd, false);
 592                 }
 593
 594                 if (err < 0)
 595                         return err;
 596
 597                 err = perf_stat_synthesize_config(is_pipe);
 598                 if (err < 0)
 599                         return err;
 600         }
 601
 602         /*
 603          * Enable counters and exec the command:
 604          */
 605         t0 = rdclock();
 606         clock_gettime(CLOCK_MONOTONIC, &ref_time);
 607
 608         if (forks) {
 609                 perf_evlist__start_workload(evsel_list);
 610                 enable_counters();
 611
 612                 if (interval) {
 613                         while (!waitpid(child_pid, &status, WNOHANG)) {
 614                                 nanosleep(&ts, NULL);
 615                                 process_interval();
 616                         }
 617                 }
 618                 wait(&status);
 619
 620                 if (workload_exec_errno) {
 621                         const char *emsg = strerror_r(workload_exec_errno, msg, sizeof(msg));
 622                         pr_err("Workload failed: %s\n", emsg);
 623                         return -1;
 624                 }
 625
 626                 if (WIFSIGNALED(status))
 627                         psignal(WTERMSIG(status), argv[0]);
 628         } else {
 629                 enable_counters();
 630                 while (!done) {
 631                         nanosleep(&ts, NULL);
 632                         if (interval)
 633                                 process_interval();
 634                 }
 635         }
 636
 637         t1 = rdclock();
 638
 639         update_stats(&walltime_nsecs_stats, t1 - t0);
 640
 641         read_counters(true);
 642
 643         return WEXITSTATUS(status);
 644 }
 645
 646 static int run_perf_stat(int argc, const char **argv)
 647 {
 648         int ret;
 649
 650         if (pre_cmd) {
 651                 ret = system(pre_cmd);
 652                 if (ret)
 653                         return ret;
 654         }
 655
 656         if (sync_run)
 657                 sync();
 658
 659         ret = __run_perf_stat(argc, argv);
 660         if (ret)
 661                 return ret;
 662
 663         if (post_cmd) {
 664                 ret = system(post_cmd);
 665                 if (ret)
 666                         return ret;
 667         }
 668
 669         return ret;
 670 }
 671
 672 static void print_running(u64 run, u64 ena)
 673 {
 674         if (csv_output) {
 675                 fprintf(stat_config.output, "%s%" PRIu64 "%s%.2f",
 676                                         csv_sep,
 677                                         run,
 678                                         csv_sep,
 679                                         ena ? 100.0 * run / ena : 100.0);
 680         } else if (run != ena) {
 681                 fprintf(stat_config.output, "  (%.2f%%)", 100.0 * run / ena);
 682         }
 683 }
 684
 685 static void print_noise_pct(double total, double avg)
 686 {
 687         double pct = rel_stddev_stats(total, avg);
 688
 689         if (csv_output)
 690                 fprintf(stat_config.output, "%s%.2f%%", csv_sep, pct);
 691         else if (pct)
 692                 fprintf(stat_config.output, "  ( +-%6.2f%% )", pct);
 693 }
 694
 695 static void print_noise(struct perf_evsel *evsel, double avg)
 696 {
 697         struct perf_stat_evsel *ps;
 698
 699         if (run_count == 1)
 700                 return;
 701
 702         ps = evsel->priv;
 703         print_noise_pct(stddev_stats(&ps->res_stats[0]), avg);
 704 }
 705
 706 static void aggr_printout(struct perf_evsel *evsel, int id, int nr)
 707 {
 708         switch (stat_config.aggr_mode) {
 709         case AGGR_CORE:
 710                 fprintf(stat_config.output, "S%d-C%*d%s%*d%s",
 711                         cpu_map__id_to_socket(id),
 712                         csv_output ? 0 : -8,
 713                         cpu_map__id_to_cpu(id),
 714                         csv_sep,
 715                         csv_output ? 0 : 4,
 716                         nr,
 717                         csv_sep);
 718                 break;
 719         case AGGR_SOCKET:
 720                 fprintf(stat_config.output, "S%*d%s%*d%s",
 721                         csv_output ? 0 : -5,
 722                         id,
 723                         csv_sep,
 724                         csv_output ? 0 : 4,
 725                         nr,
 726                         csv_sep);
 727                         break;
 728         case AGGR_NONE:
 729                 fprintf(stat_config.output, "CPU%*d%s",
 730                         csv_output ? 0 : -4,
 731                         perf_evsel__cpus(evsel)->map[id], csv_sep);
 732                 break;
 733         case AGGR_THREAD:
 734                 fprintf(stat_config.output, "%*s-%*d%s",
 735                         csv_output ? 0 : 16,
 736                         thread_map__comm(evsel->threads, id),
 737                         csv_output ? 0 : -8,
 738                         thread_map__pid(evsel->threads, id),
 739                         csv_sep);
 740                 break;
 741         case AGGR_GLOBAL:
 742         case AGGR_UNSET:
 743         default:
 744                 break;
 745         }
 746 }
 747
 748 struct outstate {
 749         FILE *fh;
 750         bool newline;
 751         const char *prefix;
 752         int  nfields;
 753         int  id, nr;
 754         struct perf_evsel *evsel;
 755 };
 756
 757 #define METRIC_LEN  35
 758
 759 static void new_line_std(void *ctx)
 760 {
 761         struct outstate *os = ctx;
 762
 763         os->newline = true;
 764 }
 765
 766 static void do_new_line_std(struct outstate *os)
 767 {
 768         fputc('\n', os->fh);
 769         fputs(os->prefix, os->fh);
 770         aggr_printout(os->evsel, os->id, os->nr);
 771         if (stat_config.aggr_mode == AGGR_NONE)
 772                 fprintf(os->fh, "        ");
 773         fprintf(os->fh, "                                                 ");
 774 }
 775
 776 static void print_metric_std(void *ctx, const char *color, const char *fmt,
 777                              const char *unit, double val)
 778 {
 779         struct outstate *os = ctx;
 780         FILE *out = os->fh;
 781         int n;
 782         bool newline = os->newline;
 783
 784         os->newline = false;
 785
 786         if (unit == NULL || fmt == NULL) {
 787                 fprintf(out, "%-*s", METRIC_LEN, "");
 788                 return;
 789         }
 790
 791         if (newline)
 792                 do_new_line_std(os);
 793
 794         n = fprintf(out, " # ");
 795         if (color)
 796                 n += color_fprintf(out, color, fmt, val);
 797         else
 798                 n += fprintf(out, fmt, val);
 799         fprintf(out, " %-*s", METRIC_LEN - n - 1, unit);
 800 }
 801
 802 static void new_line_csv(void *ctx)
 803 {
 804         struct outstate *os = ctx;
 805         int i;
 806
 807         fputc('\n', os->fh);
 808         if (os->prefix)
 809                 fprintf(os->fh, "%s%s", os->prefix, csv_sep);
 810         aggr_printout(os->evsel, os->id, os->nr);
 811         for (i = 0; i < os->nfields; i++)
 812                 fputs(csv_sep, os->fh);
 813 }
 814
 815 static void print_metric_csv(void *ctx,
 816                              const char *color __maybe_unused,
 817                              const char *fmt, const char *unit, double val)
 818 {
 819         struct outstate *os = ctx;
 820         FILE *out = os->fh;
 821         char buf[64], *vals, *ends;
 822
 823         if (unit == NULL || fmt == NULL) {
 824                 fprintf(out, "%s%s%s%s", csv_sep, csv_sep, csv_sep, csv_sep);
 825                 return;
 826         }
 827         snprintf(buf, sizeof(buf), fmt, val);
 828         vals = buf;
 829         while (isspace(*vals))
 830                 vals++;
 831         ends = vals;
 832         while (isdigit(*ends) || *ends == '.')
 833                 ends++;
 834         *ends = 0;
 835         while (isspace(*unit))
 836                 unit++;
 837         fprintf(out, "%s%s%s%s", csv_sep, vals, csv_sep, unit);
 838 }
 839
 840 #define METRIC_ONLY_LEN 20
 841
 842 /* Filter out some columns that don't work well in metrics only mode */
 843
 844 static bool valid_only_metric(const char *unit)
 845 {
 846         if (!unit)
 847                 return false;
 848         if (strstr(unit, "/sec") ||
 849             strstr(unit, "hz") ||
 850             strstr(unit, "Hz") ||
 851             strstr(unit, "CPUs utilized"))
 852                 return false;
 853         return true;
 854 }
 855
 856 static const char *fixunit(char *buf, struct perf_evsel *evsel,
 857                            const char *unit)
 858 {
 859         if (!strncmp(unit, "of all", 6)) {
 860                 snprintf(buf, 1024, "%s %s", perf_evsel__name(evsel),
 861                          unit);
 862                 return buf;
 863         }
 864         return unit;
 865 }
 866
 867 static void print_metric_only(void *ctx, const char *color, const char *fmt,
 868                               const char *unit, double val)
 869 {
 870         struct outstate *os = ctx;
 871         FILE *out = os->fh;
 872         int n;
 873         char buf[1024];
 874         unsigned mlen = METRIC_ONLY_LEN;
 875
 876         if (!valid_only_metric(unit))
 877                 return;
 878         unit = fixunit(buf, os->evsel, unit);
 879         if (color)
 880                 n = color_fprintf(out, color, fmt, val);
 881         else
 882                 n = fprintf(out, fmt, val);
 883         if (n > METRIC_ONLY_LEN)
 884                 n = METRIC_ONLY_LEN;
 885         if (mlen < strlen(unit))
 886                 mlen = strlen(unit) + 1;
 887         fprintf(out, "%*s", mlen - n, "");
 888 }
 889
 890 static void print_metric_only_csv(void *ctx, const char *color __maybe_unused,
 891                                   const char *fmt,
 892                                   const char *unit, double val)
 893 {
 894         struct outstate *os = ctx;
 895         FILE *out = os->fh;
 896         char buf[64], *vals, *ends;
 897         char tbuf[1024];
 898
 899         if (!valid_only_metric(unit))
 900                 return;
 901         unit = fixunit(tbuf, os->evsel, unit);
 902         snprintf(buf, sizeof buf, fmt, val);
 903         vals = buf;
 904         while (isspace(*vals))
 905                 vals++;
 906         ends = vals;
 907         while (isdigit(*ends) || *ends == '.')
 908                 ends++;
 909         *ends = 0;
 910         fprintf(out, "%s%s", vals, csv_sep);
 911 }
 912
 913 static void new_line_metric(void *ctx __maybe_unused)
 914 {
 915 }
 916
 917 static void print_metric_header(void *ctx, const char *color __maybe_unused,
 918                                 const char *fmt __maybe_unused,
 919                                 const char *unit, double val __maybe_unused)
 920 {
 921         struct outstate *os = ctx;
 922         char tbuf[1024];
 923
 924         if (!valid_only_metric(unit))
 925                 return;
 926         unit = fixunit(tbuf, os->evsel, unit);
 927         if (csv_output)
 928                 fprintf(os->fh, "%s%s", unit, csv_sep);
 929         else
 930                 fprintf(os->fh, "%-*s ", METRIC_ONLY_LEN, unit);
 931 }
 932
 933 static void nsec_printout(int id, int nr, struct perf_evsel *evsel, double avg)
 934 {
 935         FILE *output = stat_config.output;
 936         double msecs = avg / 1e6;
 937         const char *fmt_v, *fmt_n;
 938         char name[25];
 939
 940         fmt_v = csv_output ? "%.6f%s" : "%18.6f%s";
 941         fmt_n = csv_output ? "%s" : "%-25s";
 942
 943         aggr_printout(evsel, id, nr);
 944
 945         scnprintf(name, sizeof(name), "%s%s",
 946                   perf_evsel__name(evsel), csv_output ? "" : " (msec)");
 947
 948         fprintf(output, fmt_v, msecs, csv_sep);
 949
 950         if (csv_output)
 951                 fprintf(output, "%s%s", evsel->unit, csv_sep);
 952         else
 953                 fprintf(output, "%-*s%s", unit_width, evsel->unit, csv_sep);
 954
 955         fprintf(output, fmt_n, name);
 956
 957         if (evsel->cgrp)
 958                 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
 959 }
 960
 961 static int first_shadow_cpu(struct perf_evsel *evsel, int id)
 962 {
 963         int i;
 964
 965         if (!aggr_get_id)
 966                 return 0;
 967
 968         if (stat_config.aggr_mode == AGGR_NONE)
 969                 return id;
 970
 971         if (stat_config.aggr_mode == AGGR_GLOBAL)
 972                 return 0;
 973
 974         for (i = 0; i < perf_evsel__nr_cpus(evsel); i++) {
 975                 int cpu2 = perf_evsel__cpus(evsel)->map[i];
 976
 977                 if (aggr_get_id(evsel_list->cpus, cpu2) == id)
 978                         return cpu2;
 979         }
 980         return 0;
 981 }
 982
 983 static void abs_printout(int id, int nr, struct perf_evsel *evsel, double avg)
 984 {
 985         FILE *output = stat_config.output;
 986         double sc =  evsel->scale;
 987         const char *fmt;
 988
 989         if (csv_output) {
 990                 fmt = floor(sc) != sc ?  "%.2f%s" : "%.0f%s";
 991         } else {
 992                 if (big_num)
 993                         fmt = floor(sc) != sc ? "%'18.2f%s" : "%'18.0f%s";
 994                 else
 995                         fmt = floor(sc) != sc ? "%18.2f%s" : "%18.0f%s";
 996         }
 997
 998         aggr_printout(evsel, id, nr);
 999
1000         fprintf(output, fmt, avg, csv_sep);
1001
1002         if (evsel->unit)
1003                 fprintf(output, "%-*s%s",
1004                         csv_output ? 0 : unit_width,
1005                         evsel->unit, csv_sep);
1006
1007         fprintf(output, "%-*s", csv_output ? 0 : 25, perf_evsel__name(evsel));
1008
1009         if (evsel->cgrp)
1010                 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
1011 }
1012
1013 static void printout(int id, int nr, struct perf_evsel *counter, double uval,
1014                      char *prefix, u64 run, u64 ena, double noise)
1015 {
1016         struct perf_stat_output_ctx out;
1017         struct outstate os = {
1018                 .fh = stat_config.output,
1019                 .prefix = prefix ? prefix : "",
1020                 .id = id,
1021                 .nr = nr,
1022                 .evsel = counter,
1023         };
1024         print_metric_t pm = print_metric_std;
1025         void (*nl)(void *);
1026
1027         if (metric_only) {
1028                 nl = new_line_metric;
1029                 if (csv_output)
1030                         pm = print_metric_only_csv;
1031                 else
1032                         pm = print_metric_only;
1033         } else
1034                 nl = new_line_std;
1035
1036         if (csv_output && !metric_only) {
1037                 static int aggr_fields[] = {
1038                         [AGGR_GLOBAL] = 0,
1039                         [AGGR_THREAD] = 1,
1040                         [AGGR_NONE] = 1,
1041                         [AGGR_SOCKET] = 2,
1042                         [AGGR_CORE] = 2,
1043                 };
1044
1045                 pm = print_metric_csv;
1046                 nl = new_line_csv;
1047                 os.nfields = 3;
1048                 os.nfields += aggr_fields[stat_config.aggr_mode];
1049                 if (counter->cgrp)
1050                         os.nfields++;
1051         }
1052         if (run == 0 || ena == 0 || counter->counts->scaled == -1) {
1053                 if (metric_only) {
1054                         pm(&os, NULL, "", "", 0);
1055                         return;
1056                 }
1057                 aggr_printout(counter, id, nr);
1058
1059                 fprintf(stat_config.output, "%*s%s",
1060                         csv_output ? 0 : 18,
1061                         counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
1062                         csv_sep);
1063
1064                 fprintf(stat_config.output, "%-*s%s",
1065                         csv_output ? 0 : unit_width,
1066                         counter->unit, csv_sep);
1067
1068                 fprintf(stat_config.output, "%*s",
1069                         csv_output ? 0 : -25,
1070                         perf_evsel__name(counter));
1071
1072                 if (counter->cgrp)
1073                         fprintf(stat_config.output, "%s%s",
1074                                 csv_sep, counter->cgrp->name);
1075
1076                 if (!csv_output)
1077                         pm(&os, NULL, NULL, "", 0);
1078                 print_noise(counter, noise);
1079                 print_running(run, ena);
1080                 if (csv_output)
1081                         pm(&os, NULL, NULL, "", 0);
1082                 return;
1083         }
1084
1085         if (metric_only)
1086                 /* nothing */;
1087         else if (nsec_counter(counter))
1088                 nsec_printout(id, nr, counter, uval);
1089         else
1090                 abs_printout(id, nr, counter, uval);
1091
1092         out.print_metric = pm;
1093         out.new_line = nl;
1094         out.ctx = &os;
1095
1096         if (csv_output && !metric_only) {
1097                 print_noise(counter, noise);
1098                 print_running(run, ena);
1099         }
1100
1101         perf_stat__print_shadow_stats(counter, uval,
1102                                 first_shadow_cpu(counter, id),
1103                                 &out);
1104         if (!csv_output && !metric_only) {
1105                 print_noise(counter, noise);
1106                 print_running(run, ena);
1107         }
1108 }
1109
1110 static void aggr_update_shadow(void)
1111 {
1112         int cpu, s2, id, s;
1113         u64 val;
1114         struct perf_evsel *counter;
1115
1116         for (s = 0; s < aggr_map->nr; s++) {
1117                 id = aggr_map->map[s];
1118                 evlist__for_each(evsel_list, counter) {
1119                         val = 0;
1120                         for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
1121                                 s2 = aggr_get_id(evsel_list->cpus, cpu);
1122                                 if (s2 != id)
1123                                         continue;
1124                                 val += perf_counts(counter->counts, cpu, 0)->val;
1125                         }
1126                         val = val * counter->scale;
1127                         perf_stat__update_shadow_stats(counter, &val,
1128                                                        first_shadow_cpu(counter, id));
1129                 }
1130         }
1131 }
1132
1133 static void print_aggr(char *prefix)
1134 {
1135         FILE *output = stat_config.output;
1136         struct perf_evsel *counter;
1137         int cpu, s, s2, id, nr;
1138         double uval;
1139         u64 ena, run, val;
1140         bool first;
1141
1142         if (!(aggr_map || aggr_get_id))
1143                 return;
1144
1145         aggr_update_shadow();
1146
1147         /*
1148          * With metric_only everything is on a single line.
1149          * Without each counter has its own line.
1150          */
1151         for (s = 0; s < aggr_map->nr; s++) {
1152                 if (prefix && metric_only)
1153                         fprintf(output, "%s", prefix);
1154
1155                 id = aggr_map->map[s];
1156                 first = true;
1157                 evlist__for_each(evsel_list, counter) {
1158                         val = ena = run = 0;
1159                         nr = 0;
1160                         for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
1161                                 s2 = aggr_get_id(perf_evsel__cpus(counter), cpu);
1162                                 if (s2 != id)
1163                                         continue;
1164                                 val += perf_counts(counter->counts, cpu, 0)->val;
1165                                 ena += perf_counts(counter->counts, cpu, 0)->ena;
1166                                 run += perf_counts(counter->counts, cpu, 0)->run;
1167                                 nr++;
1168                         }
1169                         if (first && metric_only) {
1170                                 first = false;
1171                                 aggr_printout(counter, id, nr);
1172                         }
1173                         if (prefix && !metric_only)
1174                                 fprintf(output, "%s", prefix);
1175
1176                         uval = val * counter->scale;
1177                         printout(id, nr, counter, uval, prefix, run, ena, 1.0);
1178                         if (!metric_only)
1179                                 fputc('\n', output);
1180                 }
1181                 if (metric_only)
1182                         fputc('\n', output);
1183         }
1184 }
1185
1186 static void print_aggr_thread(struct perf_evsel *counter, char *prefix)
1187 {
1188         FILE *output = stat_config.output;
1189         int nthreads = thread_map__nr(counter->threads);
1190         int ncpus = cpu_map__nr(counter->cpus);
1191         int cpu, thread;
1192         double uval;
1193
1194         for (thread = 0; thread < nthreads; thread++) {
1195                 u64 ena = 0, run = 0, val = 0;
1196
1197                 for (cpu = 0; cpu < ncpus; cpu++) {
1198                         val += perf_counts(counter->counts, cpu, thread)->val;
1199                         ena += perf_counts(counter->counts, cpu, thread)->ena;
1200                         run += perf_counts(counter->counts, cpu, thread)->run;
1201                 }
1202
1203                 if (prefix)
1204                         fprintf(output, "%s", prefix);
1205
1206                 uval = val * counter->scale;
1207                 printout(thread, 0, counter, uval, prefix, run, ena, 1.0);
1208                 fputc('\n', output);
1209         }
1210 }
1211
1212 /*
1213  * Print out the results of a single counter:
1214  * aggregated counts in system-wide mode
1215  */
1216 static void print_counter_aggr(struct perf_evsel *counter, char *prefix)
1217 {
1218         FILE *output = stat_config.output;
1219         struct perf_stat_evsel *ps = counter->priv;
1220         double avg = avg_stats(&ps->res_stats[0]);
1221         double uval;
1222         double avg_enabled, avg_running;
1223
1224         avg_enabled = avg_stats(&ps->res_stats[1]);
1225         avg_running = avg_stats(&ps->res_stats[2]);
1226
1227         if (prefix && !metric_only)
1228                 fprintf(output, "%s", prefix);
1229
1230         uval = avg * counter->scale;
1231         printout(-1, 0, counter, uval, prefix, avg_running, avg_enabled, avg);
1232         if (!metric_only)
1233                 fprintf(output, "\n");
1234 }
1235
1236 /*
1237  * Print out the results of a single counter:
1238  * does not use aggregated count in system-wide
1239  */
1240 static void print_counter(struct perf_evsel *counter, char *prefix)
1241 {
1242         FILE *output = stat_config.output;
1243         u64 ena, run, val;
1244         double uval;
1245         int cpu;
1246
1247         for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
1248                 val = perf_counts(counter->counts, cpu, 0)->val;
1249                 ena = perf_counts(counter->counts, cpu, 0)->ena;
1250                 run = perf_counts(counter->counts, cpu, 0)->run;
1251
1252                 if (prefix)
1253                         fprintf(output, "%s", prefix);
1254
1255                 uval = val * counter->scale;
1256                 printout(cpu, 0, counter, uval, prefix, run, ena, 1.0);
1257
1258                 fputc('\n', output);
1259         }
1260 }
1261
1262 static void print_no_aggr_metric(char *prefix)
1263 {
1264         int cpu;
1265         int nrcpus = 0;
1266         struct perf_evsel *counter;
1267         u64 ena, run, val;
1268         double uval;
1269
1270         nrcpus = evsel_list->cpus->nr;
1271         for (cpu = 0; cpu < nrcpus; cpu++) {
1272                 bool first = true;
1273
1274                 if (prefix)
1275                         fputs(prefix, stat_config.output);
1276                 evlist__for_each(evsel_list, counter) {
1277                         if (first) {
1278                                 aggr_printout(counter, cpu, 0);
1279                                 first = false;
1280                         }
1281                         val = perf_counts(counter->counts, cpu, 0)->val;
1282                         ena = perf_counts(counter->counts, cpu, 0)->ena;
1283                         run = perf_counts(counter->counts, cpu, 0)->run;
1284
1285                         uval = val * counter->scale;
1286                         printout(cpu, 0, counter, uval, prefix, run, ena, 1.0);
1287                 }
1288                 fputc('\n', stat_config.output);
1289         }
1290 }
1291
1292 static int aggr_header_lens[] = {
1293         [AGGR_CORE] = 18,
1294         [AGGR_SOCKET] = 12,
1295         [AGGR_NONE] = 6,
1296         [AGGR_THREAD] = 24,
1297         [AGGR_GLOBAL] = 0,
1298 };
1299
1300 static void print_metric_headers(char *prefix)
1301 {
1302         struct perf_stat_output_ctx out;
1303         struct perf_evsel *counter;
1304         struct outstate os = {
1305                 .fh = stat_config.output
1306         };
1307
1308         if (prefix)
1309                 fprintf(stat_config.output, "%s", prefix);
1310
1311         if (!csv_output)
1312                 fprintf(stat_config.output, "%*s",
1313                         aggr_header_lens[stat_config.aggr_mode], "");
1314
1315         /* Print metrics headers only */
1316         evlist__for_each(evsel_list, counter) {
1317                 os.evsel = counter;
1318                 out.ctx = &os;
1319                 out.print_metric = print_metric_header;
1320                 out.new_line = new_line_metric;
1321                 os.evsel = counter;
1322                 perf_stat__print_shadow_stats(counter, 0,
1323                                               0,
1324                                               &out);
1325         }
1326         fputc('\n', stat_config.output);
1327 }
1328
1329 static void print_interval(char *prefix, struct timespec *ts)
1330 {
1331         FILE *output = stat_config.output;
1332         static int num_print_interval;
1333
1334         sprintf(prefix, "%6lu.%09lu%s", ts->tv_sec, ts->tv_nsec, csv_sep);
1335
1336         if (num_print_interval == 0 && !csv_output && !metric_only) {
1337                 switch (stat_config.aggr_mode) {
1338                 case AGGR_SOCKET:
1339                         fprintf(output, "#           time socket cpus             counts %*s events\n", unit_width, "unit");
1340                         break;
1341                 case AGGR_CORE:
1342                         fprintf(output, "#           time core         cpus             counts %*s events\n", unit_width, "unit");
1343                         break;
1344                 case AGGR_NONE:
1345                         fprintf(output, "#           time CPU                counts %*s events\n", unit_width, "unit");
1346                         break;
1347                 case AGGR_THREAD:
1348                         fprintf(output, "#           time             comm-pid                  counts %*s events\n", unit_width, "unit");
1349                         break;
1350                 case AGGR_GLOBAL:
1351                 default:
1352                         fprintf(output, "#           time             counts %*s events\n", unit_width, "unit");
1353                 case AGGR_UNSET:
1354                         break;
1355                 }
1356         }
1357
1358         if (++num_print_interval == 25)
1359                 num_print_interval = 0;
1360 }
1361
1362 static void print_header(int argc, const char **argv)
1363 {
1364         FILE *output = stat_config.output;
1365         int i;
1366
1367         fflush(stdout);
1368
1369         if (!csv_output) {
1370                 fprintf(output, "\n");
1371                 fprintf(output, " Performance counter stats for ");
1372                 if (target.system_wide)
1373                         fprintf(output, "\'system wide");
1374                 else if (target.cpu_list)
1375                         fprintf(output, "\'CPU(s) %s", target.cpu_list);
1376                 else if (!target__has_task(&target)) {
1377                         fprintf(output, "\'%s", argv ? argv[0] : "pipe");
1378                         for (i = 1; argv && (i < argc); i++)
1379                                 fprintf(output, " %s", argv[i]);
1380                 } else if (target.pid)
1381                         fprintf(output, "process id \'%s", target.pid);
1382                 else
1383                         fprintf(output, "thread id \'%s", target.tid);
1384
1385                 fprintf(output, "\'");
1386                 if (run_count > 1)
1387                         fprintf(output, " (%d runs)", run_count);
1388                 fprintf(output, ":\n\n");
1389         }
1390 }
1391
1392 static void print_footer(void)
1393 {
1394         FILE *output = stat_config.output;
1395
1396         if (!null_run)
1397                 fprintf(output, "\n");
1398         fprintf(output, " %17.9f seconds time elapsed",
1399                         avg_stats(&walltime_nsecs_stats)/1e9);
1400         if (run_count > 1) {
1401                 fprintf(output, "                                        ");
1402                 print_noise_pct(stddev_stats(&walltime_nsecs_stats),
1403                                 avg_stats(&walltime_nsecs_stats));
1404         }
1405         fprintf(output, "\n\n");
1406 }
1407
1408 static void print_counters(struct timespec *ts, int argc, const char **argv)
1409 {
1410         int interval = stat_config.interval;
1411         struct perf_evsel *counter;
1412         char buf[64], *prefix = NULL;
1413
1414         /* Do not print anything if we record to the pipe. */
1415         if (STAT_RECORD && perf_stat.file.is_pipe)
1416                 return;
1417
1418         if (interval)
1419                 print_interval(prefix = buf, ts);
1420         else
1421                 print_header(argc, argv);
1422
1423         if (metric_only) {
1424                 static int num_print_iv;
1425
1426                 if (num_print_iv == 0)
1427                         print_metric_headers(prefix);
1428                 if (num_print_iv++ == 25)
1429                         num_print_iv = 0;
1430                 if (stat_config.aggr_mode == AGGR_GLOBAL && prefix)
1431                         fprintf(stat_config.output, "%s", prefix);
1432         }
1433
1434         switch (stat_config.aggr_mode) {
1435         case AGGR_CORE:
1436         case AGGR_SOCKET:
1437                 print_aggr(prefix);
1438                 break;
1439         case AGGR_THREAD:
1440                 evlist__for_each(evsel_list, counter)
1441                         print_aggr_thread(counter, prefix);
1442                 break;
1443         case AGGR_GLOBAL:
1444                 evlist__for_each(evsel_list, counter)
1445                         print_counter_aggr(counter, prefix);
1446                 if (metric_only)
1447                         fputc('\n', stat_config.output);
1448                 break;
1449         case AGGR_NONE:
1450                 if (metric_only)
1451                         print_no_aggr_metric(prefix);
1452                 else {
1453                         evlist__for_each(evsel_list, counter)
1454                                 print_counter(counter, prefix);
1455                 }
1456                 break;
1457         case AGGR_UNSET:
1458         default:
1459                 break;
1460         }
1461
1462         if (!interval && !csv_output)
1463                 print_footer();
1464
1465         fflush(stat_config.output);
1466 }
1467
1468 static volatile int signr = -1;
1469
1470 static void skip_signal(int signo)
1471 {
1472         if ((child_pid == -1) || stat_config.interval)
1473                 done = 1;
1474
1475         signr = signo;
1476         /*
1477          * render child_pid harmless
1478          * won't send SIGTERM to a random
1479          * process in case of race condition
1480          * and fast PID recycling
1481          */
1482         child_pid = -1;
1483 }
1484
1485 static void sig_atexit(void)
1486 {
1487         sigset_t set, oset;
1488
1489         /*
1490          * avoid race condition with SIGCHLD handler
1491          * in skip_signal() which is modifying child_pid
1492          * goal is to avoid send SIGTERM to a random
1493          * process
1494          */
1495         sigemptyset(&set);
1496         sigaddset(&set, SIGCHLD);
1497         sigprocmask(SIG_BLOCK, &set, &oset);
1498
1499         if (child_pid != -1)
1500                 kill(child_pid, SIGTERM);
1501
1502         sigprocmask(SIG_SETMASK, &oset, NULL);
1503
1504         if (signr == -1)
1505                 return;
1506
1507         signal(signr, SIG_DFL);
1508         kill(getpid(), signr);
1509 }
1510
1511 static int stat__set_big_num(const struct option *opt __maybe_unused,
1512                              const char *s __maybe_unused, int unset)
1513 {
1514         big_num_opt = unset ? 0 : 1;
1515         return 0;
1516 }
1517
1518 static const struct option stat_options[] = {
1519         OPT_BOOLEAN('T', "transaction", &transaction_run,
1520                     "hardware transaction statistics"),
1521         OPT_CALLBACK('e', "event", &evsel_list, "event",
1522                      "event selector. use 'perf list' to list available events",
1523                      parse_events_option),
1524         OPT_CALLBACK(0, "filter", &evsel_list, "filter",
1525                      "event filter", parse_filter),
1526         OPT_BOOLEAN('i', "no-inherit", &no_inherit,
1527                     "child tasks do not inherit counters"),
1528         OPT_STRING('p', "pid", &target.pid, "pid",
1529                    "stat events on existing process id"),
1530         OPT_STRING('t', "tid", &target.tid, "tid",
1531                    "stat events on existing thread id"),
1532         OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
1533                     "system-wide collection from all CPUs"),
1534         OPT_BOOLEAN('g', "group", &group,
1535                     "put the counters into a counter group"),
1536         OPT_BOOLEAN('c', "scale", &stat_config.scale, "scale/normalize counters"),
1537         OPT_INCR('v', "verbose", &verbose,
1538                     "be more verbose (show counter open errors, etc)"),
1539         OPT_INTEGER('r', "repeat", &run_count,
1540                     "repeat command and print average + stddev (max: 100, forever: 0)"),
1541         OPT_BOOLEAN('n', "null", &null_run,
1542                     "null run - dont start any counters"),
1543         OPT_INCR('d', "detailed", &detailed_run,
1544                     "detailed run - start a lot of events"),
1545         OPT_BOOLEAN('S', "sync", &sync_run,
1546                     "call sync() before starting a run"),
1547         OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
1548                            "print large numbers with thousands\' separators",
1549                            stat__set_big_num),
1550         OPT_STRING('C', "cpu", &target.cpu_list, "cpu",
1551                     "list of cpus to monitor in system-wide"),
1552         OPT_SET_UINT('A', "no-aggr", &stat_config.aggr_mode,
1553                     "disable CPU count aggregation", AGGR_NONE),
1554         OPT_STRING('x', "field-separator", &csv_sep, "separator",
1555                    "print counts with custom separator"),
1556         OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
1557                      "monitor event in cgroup name only", parse_cgroups),
1558         OPT_STRING('o', "output", &output_name, "file", "output file name"),
1559         OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
1560         OPT_INTEGER(0, "log-fd", &output_fd,
1561                     "log output to fd, instead of stderr"),
1562         OPT_STRING(0, "pre", &pre_cmd, "command",
1563                         "command to run prior to the measured command"),
1564         OPT_STRING(0, "post", &post_cmd, "command",
1565                         "command to run after to the measured command"),
1566         OPT_UINTEGER('I', "interval-print", &stat_config.interval,
1567                     "print counts at regular interval in ms (>= 10)"),
1568         OPT_SET_UINT(0, "per-socket", &stat_config.aggr_mode,
1569                      "aggregate counts per processor socket", AGGR_SOCKET),
1570         OPT_SET_UINT(0, "per-core", &stat_config.aggr_mode,
1571                      "aggregate counts per physical processor core", AGGR_CORE),
1572         OPT_SET_UINT(0, "per-thread", &stat_config.aggr_mode,
1573                      "aggregate counts per thread", AGGR_THREAD),
1574         OPT_UINTEGER('D', "delay", &initial_delay,
1575                      "ms to wait before starting measurement after program start"),
1576         OPT_BOOLEAN(0, "metric-only", &metric_only,
1577                         "Only print computed metrics. No raw values"),
1578         OPT_END()
1579 };
1580
1581 static int perf_stat__get_socket(struct cpu_map *map, int cpu)
1582 {
1583         return cpu_map__get_socket(map, cpu, NULL);
1584 }
1585
1586 static int perf_stat__get_core(struct cpu_map *map, int cpu)
1587 {
1588         return cpu_map__get_core(map, cpu, NULL);
1589 }
1590
1591 static int cpu_map__get_max(struct cpu_map *map)
1592 {
1593         int i, max = -1;
1594
1595         for (i = 0; i < map->nr; i++) {
1596                 if (map->map[i] > max)
1597                         max = map->map[i];
1598         }
1599
1600         return max;
1601 }
1602
1603 static struct cpu_map *cpus_aggr_map;
1604
1605 static int perf_stat__get_aggr(aggr_get_id_t get_id, struct cpu_map *map, int idx)
1606 {
1607         int cpu;
1608
1609         if (idx >= map->nr)
1610                 return -1;
1611
1612         cpu = map->map[idx];
1613
1614         if (cpus_aggr_map->map[cpu] == -1)
1615                 cpus_aggr_map->map[cpu] = get_id(map, idx);
1616
1617         return cpus_aggr_map->map[cpu];
1618 }
1619
1620 static int perf_stat__get_socket_cached(struct cpu_map *map, int idx)
1621 {
1622         return perf_stat__get_aggr(perf_stat__get_socket, map, idx);
1623 }
1624
1625 static int perf_stat__get_core_cached(struct cpu_map *map, int idx)
1626 {
1627         return perf_stat__get_aggr(perf_stat__get_core, map, idx);
1628 }
1629
1630 static int perf_stat_init_aggr_mode(void)
1631 {
1632         int nr;
1633
1634         switch (stat_config.aggr_mode) {
1635         case AGGR_SOCKET:
1636                 if (cpu_map__build_socket_map(evsel_list->cpus, &aggr_map)) {
1637                         perror("cannot build socket map");
1638                         return -1;
1639                 }
1640                 aggr_get_id = perf_stat__get_socket_cached;
1641                 break;
1642         case AGGR_CORE:
1643                 if (cpu_map__build_core_map(evsel_list->cpus, &aggr_map)) {
1644                         perror("cannot build core map");
1645                         return -1;
1646                 }
1647                 aggr_get_id = perf_stat__get_core_cached;
1648                 break;
1649         case AGGR_NONE:
1650         case AGGR_GLOBAL:
1651         case AGGR_THREAD:
1652         case AGGR_UNSET:
1653         default:
1654                 break;
1655         }
1656
1657         /*
1658          * The evsel_list->cpus is the base we operate on,
1659          * taking the highest cpu number to be the size of
1660          * the aggregation translate cpumap.
1661          */
1662         nr = cpu_map__get_max(evsel_list->cpus);
1663         cpus_aggr_map = cpu_map__empty_new(nr + 1);
1664         return cpus_aggr_map ? 0 : -ENOMEM;
1665 }
1666
1667 static void perf_stat__exit_aggr_mode(void)
1668 {
1669         cpu_map__put(aggr_map);
1670         cpu_map__put(cpus_aggr_map);
1671         aggr_map = NULL;
1672         cpus_aggr_map = NULL;
1673 }
1674
1675 static inline int perf_env__get_cpu(struct perf_env *env, struct cpu_map *map, int idx)
1676 {
1677         int cpu;
1678
1679         if (idx > map->nr)
1680                 return -1;
1681
1682         cpu = map->map[idx];
1683
1684         if (cpu >= env->nr_cpus_online)
1685                 return -1;
1686
1687         return cpu;
1688 }
1689
1690 static int perf_env__get_socket(struct cpu_map *map, int idx, void *data)
1691 {
1692         struct perf_env *env = data;
1693         int cpu = perf_env__get_cpu(env, map, idx);
1694
1695         return cpu == -1 ? -1 : env->cpu[cpu].socket_id;
1696 }
1697
1698 static int perf_env__get_core(struct cpu_map *map, int idx, void *data)
1699 {
1700         struct perf_env *env = data;
1701         int core = -1, cpu = perf_env__get_cpu(env, map, idx);
1702
1703         if (cpu != -1) {
1704                 int socket_id = env->cpu[cpu].socket_id;
1705
1706                 /*
1707                  * Encode socket in upper 16 bits
1708                  * core_id is relative to socket, and
1709                  * we need a global id. So we combine
1710                  * socket + core id.
1711                  */
1712                 core = (socket_id << 16) | (env->cpu[cpu].core_id & 0xffff);
1713         }
1714
1715         return core;
1716 }
1717
1718 static int perf_env__build_socket_map(struct perf_env *env, struct cpu_map *cpus,
1719                                       struct cpu_map **sockp)
1720 {
1721         return cpu_map__build_map(cpus, sockp, perf_env__get_socket, env);
1722 }
1723
1724 static int perf_env__build_core_map(struct perf_env *env, struct cpu_map *cpus,
1725                                     struct cpu_map **corep)
1726 {
1727         return cpu_map__build_map(cpus, corep, perf_env__get_core, env);
1728 }
1729
1730 static int perf_stat__get_socket_file(struct cpu_map *map, int idx)
1731 {
1732         return perf_env__get_socket(map, idx, &perf_stat.session->header.env);
1733 }
1734
1735 static int perf_stat__get_core_file(struct cpu_map *map, int idx)
1736 {
1737         return perf_env__get_core(map, idx, &perf_stat.session->header.env);
1738 }
1739
1740 static int perf_stat_init_aggr_mode_file(struct perf_stat *st)
1741 {
1742         struct perf_env *env = &st->session->header.env;
1743
1744         switch (stat_config.aggr_mode) {
1745         case AGGR_SOCKET:
1746                 if (perf_env__build_socket_map(env, evsel_list->cpus, &aggr_map)) {
1747                         perror("cannot build socket map");
1748                         return -1;
1749                 }
1750                 aggr_get_id = perf_stat__get_socket_file;
1751                 break;
1752         case AGGR_CORE:
1753                 if (perf_env__build_core_map(env, evsel_list->cpus, &aggr_map)) {
1754                         perror("cannot build core map");
1755                         return -1;
1756                 }
1757                 aggr_get_id = perf_stat__get_core_file;
1758                 break;
1759         case AGGR_NONE:
1760         case AGGR_GLOBAL:
1761         case AGGR_THREAD:
1762         case AGGR_UNSET:
1763         default:
1764                 break;
1765         }
1766
1767         return 0;
1768 }
1769
1770 /*
1771  * Add default attributes, if there were no attributes specified or
1772  * if -d/--detailed, -d -d or -d -d -d is used:
1773  */
1774 static int add_default_attributes(void)
1775 {
1776         struct perf_event_attr default_attrs0[] = {
1777
1778   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK              },
1779   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES        },
1780   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS          },
1781   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS             },
1782
1783   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES              },
1784 };
1785         struct perf_event_attr frontend_attrs[] = {
1786   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND },
1787 };
1788         struct perf_event_attr backend_attrs[] = {
1789   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND  },
1790 };
1791         struct perf_event_attr default_attrs1[] = {
1792   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS            },
1793   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS     },
1794   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES           },
1795
1796 };
1797
1798 /*
1799  * Detailed stats (-d), covering the L1 and last level data caches:
1800  */
1801         struct perf_event_attr detailed_attrs[] = {
1802
1803   { .type = PERF_TYPE_HW_CACHE,
1804     .config =
1805          PERF_COUNT_HW_CACHE_L1D                <<  0  |
1806         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1807         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
1808
1809   { .type = PERF_TYPE_HW_CACHE,
1810     .config =
1811          PERF_COUNT_HW_CACHE_L1D                <<  0  |
1812         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1813         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
1814
1815   { .type = PERF_TYPE_HW_CACHE,
1816     .config =
1817          PERF_COUNT_HW_CACHE_LL                 <<  0  |
1818         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1819         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
1820
1821   { .type = PERF_TYPE_HW_CACHE,
1822     .config =
1823          PERF_COUNT_HW_CACHE_LL                 <<  0  |
1824         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1825         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
1826 };
1827
1828 /*
1829  * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
1830  */
1831         struct perf_event_attr very_detailed_attrs[] = {
1832
1833   { .type = PERF_TYPE_HW_CACHE,
1834     .config =
1835          PERF_COUNT_HW_CACHE_L1I                <<  0  |
1836         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1837         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
1838
1839   { .type = PERF_TYPE_HW_CACHE,
1840     .config =
1841          PERF_COUNT_HW_CACHE_L1I                <<  0  |
1842         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1843         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
1844
1845   { .type = PERF_TYPE_HW_CACHE,
1846     .config =
1847          PERF_COUNT_HW_CACHE_DTLB               <<  0  |
1848         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1849         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
1850
1851   { .type = PERF_TYPE_HW_CACHE,
1852     .config =
1853          PERF_COUNT_HW_CACHE_DTLB               <<  0  |
1854         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1855         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
1856
1857   { .type = PERF_TYPE_HW_CACHE,
1858     .config =
1859          PERF_COUNT_HW_CACHE_ITLB               <<  0  |
1860         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1861         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
1862
1863   { .type = PERF_TYPE_HW_CACHE,
1864     .config =
1865          PERF_COUNT_HW_CACHE_ITLB               <<  0  |
1866         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1867         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
1868
1869 };
1870
1871 /*
1872  * Very, very detailed stats (-d -d -d), adding prefetch events:
1873  */
1874         struct perf_event_attr very_very_detailed_attrs[] = {
1875
1876   { .type = PERF_TYPE_HW_CACHE,
1877     .config =
1878          PERF_COUNT_HW_CACHE_L1D                <<  0  |
1879         (PERF_COUNT_HW_CACHE_OP_PREFETCH        <<  8) |
1880         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
1881
1882   { .type = PERF_TYPE_HW_CACHE,
1883     .config =
1884          PERF_COUNT_HW_CACHE_L1D                <<  0  |
1885         (PERF_COUNT_HW_CACHE_OP_PREFETCH        <<  8) |
1886         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
1887 };
1888
1889         /* Set attrs if no event is selected and !null_run: */
1890         if (null_run)
1891                 return 0;
1892
1893         if (transaction_run) {
1894                 int err;
1895                 if (pmu_have_event("cpu", "cycles-ct") &&
1896                     pmu_have_event("cpu", "el-start"))
1897                         err = parse_events(evsel_list, transaction_attrs, NULL);
1898                 else
1899                         err = parse_events(evsel_list, transaction_limited_attrs, NULL);
1900                 if (err) {
1901                         fprintf(stderr, "Cannot set up transaction events\n");
1902                         return -1;
1903                 }
1904                 return 0;
1905         }
1906
1907         if (!evsel_list->nr_entries) {
1908                 if (perf_evlist__add_default_attrs(evsel_list, default_attrs0) < 0)
1909                         return -1;
1910                 if (pmu_have_event("cpu", "stalled-cycles-frontend")) {
1911                         if (perf_evlist__add_default_attrs(evsel_list,
1912                                                 frontend_attrs) < 0)
1913                                 return -1;
1914                 }
1915                 if (pmu_have_event("cpu", "stalled-cycles-backend")) {
1916                         if (perf_evlist__add_default_attrs(evsel_list,
1917                                                 backend_attrs) < 0)
1918                                 return -1;
1919                 }
1920                 if (perf_evlist__add_default_attrs(evsel_list, default_attrs1) < 0)
1921                         return -1;
1922         }
1923
1924         /* Detailed events get appended to the event list: */
1925
1926         if (detailed_run <  1)
1927                 return 0;
1928
1929         /* Append detailed run extra attributes: */
1930         if (perf_evlist__add_default_attrs(evsel_list, detailed_attrs) < 0)
1931                 return -1;
1932
1933         if (detailed_run < 2)
1934                 return 0;
1935
1936         /* Append very detailed run extra attributes: */
1937         if (perf_evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0)
1938                 return -1;
1939
1940         if (detailed_run < 3)
1941                 return 0;
1942
1943         /* Append very, very detailed run extra attributes: */
1944         return perf_evlist__add_default_attrs(evsel_list, very_very_detailed_attrs);
1945 }
1946
1947 static const char * const stat_record_usage[] = {
1948         "perf stat record [<options>]",
1949         NULL,
1950 };
1951
1952 static void init_features(struct perf_session *session)
1953 {
1954         int feat;
1955
1956         for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++)
1957                 perf_header__set_feat(&session->header, feat);
1958
1959         perf_header__clear_feat(&session->header, HEADER_BUILD_ID);
1960         perf_header__clear_feat(&session->header, HEADER_TRACING_DATA);
1961         perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK);
1962         perf_header__clear_feat(&session->header, HEADER_AUXTRACE);
1963 }
1964
1965 static int __cmd_record(int argc, const char **argv)
1966 {
1967         struct perf_session *session;
1968         struct perf_data_file *file = &perf_stat.file;
1969
1970         argc = parse_options(argc, argv, stat_options, stat_record_usage,
1971                              PARSE_OPT_STOP_AT_NON_OPTION);
1972
1973         if (output_name)
1974                 file->path = output_name;
1975
1976         if (run_count != 1 || forever) {
1977                 pr_err("Cannot use -r option with perf stat record.\n");
1978                 return -1;
1979         }
1980
1981         session = perf_session__new(file, false, NULL);
1982         if (session == NULL) {
1983                 pr_err("Perf session creation failed.\n");
1984                 return -1;
1985         }
1986
1987         init_features(session);
1988
1989         session->evlist   = evsel_list;
1990         perf_stat.session = session;
1991         perf_stat.record  = true;
1992         return argc;
1993 }
1994
1995 static int process_stat_round_event(struct perf_tool *tool __maybe_unused,
1996                                     union perf_event *event,
1997                                     struct perf_session *session)
1998 {
1999         struct stat_round_event *stat_round = &event->stat_round;
2000         struct perf_evsel *counter;
2001         struct timespec tsh, *ts = NULL;
2002         const char **argv = session->header.env.cmdline_argv;
2003         int argc = session->header.env.nr_cmdline;
2004
2005         evlist__for_each(evsel_list, counter)
2006                 perf_stat_process_counter(&stat_config, counter);
2007
2008         if (stat_round->type == PERF_STAT_ROUND_TYPE__FINAL)
2009                 update_stats(&walltime_nsecs_stats, stat_round->time);
2010
2011         if (stat_config.interval && stat_round->time) {
2012                 tsh.tv_sec  = stat_round->time / NSECS_PER_SEC;
2013                 tsh.tv_nsec = stat_round->time % NSECS_PER_SEC;
2014                 ts = &tsh;
2015         }
2016
2017         print_counters(ts, argc, argv);
2018         return 0;
2019 }
2020
2021 static
2022 int process_stat_config_event(struct perf_tool *tool __maybe_unused,
2023                               union perf_event *event,
2024                               struct perf_session *session __maybe_unused)
2025 {
2026         struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2027
2028         perf_event__read_stat_config(&stat_config, &event->stat_config);
2029
2030         if (cpu_map__empty(st->cpus)) {
2031                 if (st->aggr_mode != AGGR_UNSET)
2032                         pr_warning("warning: processing task data, aggregation mode not set\n");
2033                 return 0;
2034         }
2035
2036         if (st->aggr_mode != AGGR_UNSET)
2037                 stat_config.aggr_mode = st->aggr_mode;
2038
2039         if (perf_stat.file.is_pipe)
2040                 perf_stat_init_aggr_mode();
2041         else
2042                 perf_stat_init_aggr_mode_file(st);
2043
2044         return 0;
2045 }
2046
2047 static int set_maps(struct perf_stat *st)
2048 {
2049         if (!st->cpus || !st->threads)
2050                 return 0;
2051
2052         if (WARN_ONCE(st->maps_allocated, "stats double allocation\n"))
2053                 return -EINVAL;
2054
2055         perf_evlist__set_maps(evsel_list, st->cpus, st->threads);
2056
2057         if (perf_evlist__alloc_stats(evsel_list, true))
2058                 return -ENOMEM;
2059
2060         st->maps_allocated = true;
2061         return 0;
2062 }
2063
2064 static
2065 int process_thread_map_event(struct perf_tool *tool __maybe_unused,
2066                              union perf_event *event,
2067                              struct perf_session *session __maybe_unused)
2068 {
2069         struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2070
2071         if (st->threads) {
2072                 pr_warning("Extra thread map event, ignoring.\n");
2073                 return 0;
2074         }
2075
2076         st->threads = thread_map__new_event(&event->thread_map);
2077         if (!st->threads)
2078                 return -ENOMEM;
2079
2080         return set_maps(st);
2081 }
2082
2083 static
2084 int process_cpu_map_event(struct perf_tool *tool __maybe_unused,
2085                           union perf_event *event,
2086                           struct perf_session *session __maybe_unused)
2087 {
2088         struct perf_stat *st = container_of(tool, struct perf_stat, tool);
2089         struct cpu_map *cpus;
2090
2091         if (st->cpus) {
2092                 pr_warning("Extra cpu map event, ignoring.\n");
2093                 return 0;
2094         }
2095
2096         cpus = cpu_map__new_data(&event->cpu_map.data);
2097         if (!cpus)
2098                 return -ENOMEM;
2099
2100         st->cpus = cpus;
2101         return set_maps(st);
2102 }
2103
2104 static const char * const stat_report_usage[] = {
2105         "perf stat report [<options>]",
2106         NULL,
2107 };
2108
2109 static struct perf_stat perf_stat = {
2110         .tool = {
2111                 .attr           = perf_event__process_attr,
2112                 .event_update   = perf_event__process_event_update,
2113                 .thread_map     = process_thread_map_event,
2114                 .cpu_map        = process_cpu_map_event,
2115                 .stat_config    = process_stat_config_event,
2116                 .stat           = perf_event__process_stat_event,
2117                 .stat_round     = process_stat_round_event,
2118         },
2119         .aggr_mode = AGGR_UNSET,
2120 };
2121
2122 static int __cmd_report(int argc, const char **argv)
2123 {
2124         struct perf_session *session;
2125         const struct option options[] = {
2126         OPT_STRING('i', "input", &input_name, "file", "input file name"),
2127         OPT_SET_UINT(0, "per-socket", &perf_stat.aggr_mode,
2128                      "aggregate counts per processor socket", AGGR_SOCKET),
2129         OPT_SET_UINT(0, "per-core", &perf_stat.aggr_mode,
2130                      "aggregate counts per physical processor core", AGGR_CORE),
2131         OPT_SET_UINT('A', "no-aggr", &perf_stat.aggr_mode,
2132                      "disable CPU count aggregation", AGGR_NONE),
2133         OPT_END()
2134         };
2135         struct stat st;
2136         int ret;
2137
2138         argc = parse_options(argc, argv, options, stat_report_usage, 0);
2139
2140         if (!input_name || !strlen(input_name)) {
2141                 if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode))
2142                         input_name = "-";
2143                 else
2144                         input_name = "perf.data";
2145         }
2146
2147         perf_stat.file.path = input_name;
2148         perf_stat.file.mode = PERF_DATA_MODE_READ;
2149
2150         session = perf_session__new(&perf_stat.file, false, &perf_stat.tool);
2151         if (session == NULL)
2152                 return -1;
2153
2154         perf_stat.session  = session;
2155         stat_config.output = stderr;
2156         evsel_list         = session->evlist;
2157
2158         ret = perf_session__process_events(session);
2159         if (ret)
2160                 return ret;
2161
2162         perf_session__delete(session);
2163         return 0;
2164 }
2165
2166 int cmd_stat(int argc, const char **argv, const char *prefix __maybe_unused)
2167 {
2168         const char * const stat_usage[] = {
2169                 "perf stat [<options>] [<command>]",
2170                 NULL
2171         };
2172         int status = -EINVAL, run_idx;
2173         const char *mode;
2174         FILE *output = stderr;
2175         unsigned int interval;
2176         const char * const stat_subcommands[] = { "record", "report" };
2177
2178         setlocale(LC_ALL, "");
2179
2180         evsel_list = perf_evlist__new();
2181         if (evsel_list == NULL)
2182                 return -ENOMEM;
2183
2184         parse_events__shrink_config_terms();
2185         argc = parse_options_subcommand(argc, argv, stat_options, stat_subcommands,
2186                                         (const char **) stat_usage,
2187                                         PARSE_OPT_STOP_AT_NON_OPTION);
2188         perf_stat__init_shadow_stats();
2189
2190         if (csv_sep) {
2191                 csv_output = true;
2192                 if (!strcmp(csv_sep, "\\t"))
2193                         csv_sep = "\t";
2194         } else
2195                 csv_sep = DEFAULT_SEPARATOR;
2196
2197         if (argc && !strncmp(argv[0], "rec", 3)) {
2198                 argc = __cmd_record(argc, argv);
2199                 if (argc < 0)
2200                         return -1;
2201         } else if (argc && !strncmp(argv[0], "rep", 3))
2202                 return __cmd_report(argc, argv);
2203
2204         interval = stat_config.interval;
2205
2206         /*
2207          * For record command the -o is already taken care of.
2208          */
2209         if (!STAT_RECORD && output_name && strcmp(output_name, "-"))
2210                 output = NULL;
2211
2212         if (output_name && output_fd) {
2213                 fprintf(stderr, "cannot use both --output and --log-fd\n");
2214                 parse_options_usage(stat_usage, stat_options, "o", 1);
2215                 parse_options_usage(NULL, stat_options, "log-fd", 0);
2216                 goto out;
2217         }
2218
2219         if (metric_only && stat_config.aggr_mode == AGGR_THREAD) {
2220                 fprintf(stderr, "--metric-only is not supported with --per-thread\n");
2221                 goto out;
2222         }
2223
2224         if (metric_only && run_count > 1) {
2225                 fprintf(stderr, "--metric-only is not supported with -r\n");
2226                 goto out;
2227         }
2228
2229         if (output_fd < 0) {
2230                 fprintf(stderr, "argument to --log-fd must be a > 0\n");
2231                 parse_options_usage(stat_usage, stat_options, "log-fd", 0);
2232                 goto out;
2233         }
2234
2235         if (!output) {
2236                 struct timespec tm;
2237                 mode = append_file ? "a" : "w";
2238
2239                 output = fopen(output_name, mode);
2240                 if (!output) {
2241                         perror("failed to create output file");
2242                         return -1;
2243                 }
2244                 clock_gettime(CLOCK_REALTIME, &tm);
2245                 fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
2246         } else if (output_fd > 0) {
2247                 mode = append_file ? "a" : "w";
2248                 output = fdopen(output_fd, mode);
2249                 if (!output) {
2250                         perror("Failed opening logfd");
2251                         return -errno;
2252                 }
2253         }
2254
2255         stat_config.output = output;
2256
2257         /*
2258          * let the spreadsheet do the pretty-printing
2259          */
2260         if (csv_output) {
2261                 /* User explicitly passed -B? */
2262                 if (big_num_opt == 1) {
2263                         fprintf(stderr, "-B option not supported with -x\n");
2264                         parse_options_usage(stat_usage, stat_options, "B", 1);
2265                         parse_options_usage(NULL, stat_options, "x", 1);
2266                         goto out;
2267                 } else /* Nope, so disable big number formatting */
2268                         big_num = false;
2269         } else if (big_num_opt == 0) /* User passed --no-big-num */
2270                 big_num = false;
2271
2272         if (!argc && target__none(&target))
2273                 usage_with_options(stat_usage, stat_options);
2274
2275         if (run_count < 0) {
2276                 pr_err("Run count must be a positive number\n");
2277                 parse_options_usage(stat_usage, stat_options, "r", 1);
2278                 goto out;
2279         } else if (run_count == 0) {
2280                 forever = true;
2281                 run_count = 1;
2282         }
2283
2284         if ((stat_config.aggr_mode == AGGR_THREAD) && !target__has_task(&target)) {
2285                 fprintf(stderr, "The --per-thread option is only available "
2286                         "when monitoring via -p -t options.\n");
2287                 parse_options_usage(NULL, stat_options, "p", 1);
2288                 parse_options_usage(NULL, stat_options, "t", 1);
2289                 goto out;
2290         }
2291
2292         /*
2293          * no_aggr, cgroup are for system-wide only
2294          * --per-thread is aggregated per thread, we dont mix it with cpu mode
2295          */
2296         if (((stat_config.aggr_mode != AGGR_GLOBAL &&
2297               stat_config.aggr_mode != AGGR_THREAD) || nr_cgroups) &&
2298             !target__has_cpu(&target)) {
2299                 fprintf(stderr, "both cgroup and no-aggregation "
2300                         "modes only available in system-wide mode\n");
2301
2302                 parse_options_usage(stat_usage, stat_options, "G", 1);
2303                 parse_options_usage(NULL, stat_options, "A", 1);
2304                 parse_options_usage(NULL, stat_options, "a", 1);
2305                 goto out;
2306         }
2307
2308         if (add_default_attributes())
2309                 goto out;
2310
2311         target__validate(&target);
2312
2313         if (perf_evlist__create_maps(evsel_list, &target) < 0) {
2314                 if (target__has_task(&target)) {
2315                         pr_err("Problems finding threads of monitor\n");
2316                         parse_options_usage(stat_usage, stat_options, "p", 1);
2317                         parse_options_usage(NULL, stat_options, "t", 1);
2318                 } else if (target__has_cpu(&target)) {
2319                         perror("failed to parse CPUs map");
2320                         parse_options_usage(stat_usage, stat_options, "C", 1);
2321                         parse_options_usage(NULL, stat_options, "a", 1);
2322                 }
2323                 goto out;
2324         }
2325
2326         /*
2327          * Initialize thread_map with comm names,
2328          * so we could print it out on output.
2329          */
2330         if (stat_config.aggr_mode == AGGR_THREAD)
2331                 thread_map__read_comms(evsel_list->threads);
2332
2333         if (interval && interval < 100) {
2334                 if (interval < 10) {
2335                         pr_err("print interval must be >= 10ms\n");
2336                         parse_options_usage(stat_usage, stat_options, "I", 1);
2337                         goto out;
2338                 } else
2339                         pr_warning("print interval < 100ms. "
2340                                    "The overhead percentage could be high in some cases. "
2341                                    "Please proceed with caution.\n");
2342         }
2343
2344         if (perf_evlist__alloc_stats(evsel_list, interval))
2345                 goto out;
2346
2347         if (perf_stat_init_aggr_mode())
2348                 goto out;
2349
2350         /*
2351          * We dont want to block the signals - that would cause
2352          * child tasks to inherit that and Ctrl-C would not work.
2353          * What we want is for Ctrl-C to work in the exec()-ed
2354          * task, but being ignored by perf stat itself:
2355          */
2356         atexit(sig_atexit);
2357         if (!forever)
2358                 signal(SIGINT,  skip_signal);
2359         signal(SIGCHLD, skip_signal);
2360         signal(SIGALRM, skip_signal);
2361         signal(SIGABRT, skip_signal);
2362
2363         status = 0;
2364         for (run_idx = 0; forever || run_idx < run_count; run_idx++) {
2365                 if (run_count != 1 && verbose)
2366                         fprintf(output, "[ perf stat: executing run #%d ... ]\n",
2367                                 run_idx + 1);
2368
2369                 status = run_perf_stat(argc, argv);
2370                 if (forever && status != -1) {
2371                         print_counters(NULL, argc, argv);
2372                         perf_stat__reset_stats();
2373                 }
2374         }
2375
2376         if (!forever && status != -1 && !interval)
2377                 print_counters(NULL, argc, argv);
2378
2379         if (STAT_RECORD) {
2380                 /*
2381                  * We synthesize the kernel mmap record just so that older tools
2382                  * don't emit warnings about not being able to resolve symbols
2383                  * due to /proc/sys/kernel/kptr_restrict settings and instear provide
2384                  * a saner message about no samples being in the perf.data file.
2385                  *
2386                  * This also serves to suppress a warning about f_header.data.size == 0
2387                  * in header.c at the moment 'perf stat record' gets introduced, which
2388                  * is not really needed once we start adding the stat specific PERF_RECORD_
2389                  * records, but the need to suppress the kptr_restrict messages in older
2390                  * tools remain  -acme
2391                  */
2392                 int fd = perf_data_file__fd(&perf_stat.file);
2393                 int err = perf_event__synthesize_kernel_mmap((void *)&perf_stat,
2394                                                              process_synthesized_event,
2395                                                              &perf_stat.session->machines.host);
2396                 if (err) {
2397                         pr_warning("Couldn't synthesize the kernel mmap record, harmless, "
2398                                    "older tools may produce warnings about this file\n.");
2399                 }
2400
2401                 if (!interval) {
2402                         if (WRITE_STAT_ROUND_EVENT(walltime_nsecs_stats.max, FINAL))
2403                                 pr_err("failed to write stat round event\n");
2404                 }
2405
2406                 if (!perf_stat.file.is_pipe) {
2407                         perf_stat.session->header.data_size += perf_stat.bytes_written;
2408                         perf_session__write_header(perf_stat.session, evsel_list, fd, true);
2409                 }
2410
2411                 perf_session__delete(perf_stat.session);
2412         }
2413
2414         perf_stat__exit_aggr_mode();
2415         perf_evlist__free_stats(evsel_list);
2416 out:
2417         perf_evlist__delete(evsel_list);
2418         return status;
2419 }