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