perf stat: Add per-core aggregation
[linux-3.10.git] / 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/util.h"
47 #include "util/parse-options.h"
48 #include "util/parse-events.h"
49 #include "util/event.h"
50 #include "util/evlist.h"
51 #include "util/evsel.h"
52 #include "util/debug.h"
53 #include "util/color.h"
54 #include "util/stat.h"
55 #include "util/header.h"
56 #include "util/cpumap.h"
57 #include "util/thread.h"
58 #include "util/thread_map.h"
59
60 #include <stdlib.h>
61 #include <sys/prctl.h>
62 #include <locale.h>
63
64 #define DEFAULT_SEPARATOR       " "
65 #define CNTR_NOT_SUPPORTED      "<not supported>"
66 #define CNTR_NOT_COUNTED        "<not counted>"
67
68 static void print_stat(int argc, const char **argv);
69 static void print_counter_aggr(struct perf_evsel *counter, char *prefix);
70 static void print_counter(struct perf_evsel *counter, char *prefix);
71 static void print_aggr(char *prefix);
72
73 static struct perf_evlist       *evsel_list;
74
75 static struct perf_target       target = {
76         .uid    = UINT_MAX,
77 };
78
79 enum aggr_mode {
80         AGGR_NONE,
81         AGGR_GLOBAL,
82         AGGR_SOCKET,
83         AGGR_CORE,
84 };
85
86 static int                      run_count                       =  1;
87 static bool                     no_inherit                      = false;
88 static bool                     scale                           =  true;
89 static enum aggr_mode           aggr_mode                       = AGGR_GLOBAL;
90 static pid_t                    child_pid                       = -1;
91 static bool                     null_run                        =  false;
92 static int                      detailed_run                    =  0;
93 static bool                     big_num                         =  true;
94 static int                      big_num_opt                     =  -1;
95 static const char               *csv_sep                        = NULL;
96 static bool                     csv_output                      = false;
97 static bool                     group                           = false;
98 static FILE                     *output                         = NULL;
99 static const char               *pre_cmd                        = NULL;
100 static const char               *post_cmd                       = NULL;
101 static bool                     sync_run                        = false;
102 static unsigned int             interval                        = 0;
103 static bool                     forever                         = false;
104 static struct timespec          ref_time;
105 static struct cpu_map           *aggr_map;
106 static int                      (*aggr_get_id)(struct cpu_map *m, int cpu);
107
108 static volatile int done = 0;
109
110 struct perf_stat {
111         struct stats      res_stats[3];
112 };
113
114 static inline void diff_timespec(struct timespec *r, struct timespec *a,
115                                  struct timespec *b)
116 {
117         r->tv_sec = a->tv_sec - b->tv_sec;
118         if (a->tv_nsec < b->tv_nsec) {
119                 r->tv_nsec = a->tv_nsec + 1000000000L - b->tv_nsec;
120                 r->tv_sec--;
121         } else {
122                 r->tv_nsec = a->tv_nsec - b->tv_nsec ;
123         }
124 }
125
126 static inline struct cpu_map *perf_evsel__cpus(struct perf_evsel *evsel)
127 {
128         return (evsel->cpus && !target.cpu_list) ? evsel->cpus : evsel_list->cpus;
129 }
130
131 static inline int perf_evsel__nr_cpus(struct perf_evsel *evsel)
132 {
133         return perf_evsel__cpus(evsel)->nr;
134 }
135
136 static void perf_evsel__reset_stat_priv(struct perf_evsel *evsel)
137 {
138         memset(evsel->priv, 0, sizeof(struct perf_stat));
139 }
140
141 static int perf_evsel__alloc_stat_priv(struct perf_evsel *evsel)
142 {
143         evsel->priv = zalloc(sizeof(struct perf_stat));
144         return evsel->priv == NULL ? -ENOMEM : 0;
145 }
146
147 static void perf_evsel__free_stat_priv(struct perf_evsel *evsel)
148 {
149         free(evsel->priv);
150         evsel->priv = NULL;
151 }
152
153 static int perf_evsel__alloc_prev_raw_counts(struct perf_evsel *evsel)
154 {
155         void *addr;
156         size_t sz;
157
158         sz = sizeof(*evsel->counts) +
159              (perf_evsel__nr_cpus(evsel) * sizeof(struct perf_counts_values));
160
161         addr = zalloc(sz);
162         if (!addr)
163                 return -ENOMEM;
164
165         evsel->prev_raw_counts =  addr;
166
167         return 0;
168 }
169
170 static void perf_evsel__free_prev_raw_counts(struct perf_evsel *evsel)
171 {
172         free(evsel->prev_raw_counts);
173         evsel->prev_raw_counts = NULL;
174 }
175
176 static void perf_evlist__free_stats(struct perf_evlist *evlist)
177 {
178         struct perf_evsel *evsel;
179
180         list_for_each_entry(evsel, &evlist->entries, node) {
181                 perf_evsel__free_stat_priv(evsel);
182                 perf_evsel__free_counts(evsel);
183                 perf_evsel__free_prev_raw_counts(evsel);
184         }
185 }
186
187 static int perf_evlist__alloc_stats(struct perf_evlist *evlist, bool alloc_raw)
188 {
189         struct perf_evsel *evsel;
190
191         list_for_each_entry(evsel, &evlist->entries, node) {
192                 if (perf_evsel__alloc_stat_priv(evsel) < 0 ||
193                     perf_evsel__alloc_counts(evsel, perf_evsel__nr_cpus(evsel)) < 0 ||
194                     (alloc_raw && perf_evsel__alloc_prev_raw_counts(evsel) < 0))
195                         goto out_free;
196         }
197
198         return 0;
199
200 out_free:
201         perf_evlist__free_stats(evlist);
202         return -1;
203 }
204
205 static struct stats runtime_nsecs_stats[MAX_NR_CPUS];
206 static struct stats runtime_cycles_stats[MAX_NR_CPUS];
207 static struct stats runtime_stalled_cycles_front_stats[MAX_NR_CPUS];
208 static struct stats runtime_stalled_cycles_back_stats[MAX_NR_CPUS];
209 static struct stats runtime_branches_stats[MAX_NR_CPUS];
210 static struct stats runtime_cacherefs_stats[MAX_NR_CPUS];
211 static struct stats runtime_l1_dcache_stats[MAX_NR_CPUS];
212 static struct stats runtime_l1_icache_stats[MAX_NR_CPUS];
213 static struct stats runtime_ll_cache_stats[MAX_NR_CPUS];
214 static struct stats runtime_itlb_cache_stats[MAX_NR_CPUS];
215 static struct stats runtime_dtlb_cache_stats[MAX_NR_CPUS];
216 static struct stats walltime_nsecs_stats;
217
218 static void perf_stat__reset_stats(struct perf_evlist *evlist)
219 {
220         struct perf_evsel *evsel;
221
222         list_for_each_entry(evsel, &evlist->entries, node) {
223                 perf_evsel__reset_stat_priv(evsel);
224                 perf_evsel__reset_counts(evsel, perf_evsel__nr_cpus(evsel));
225         }
226
227         memset(runtime_nsecs_stats, 0, sizeof(runtime_nsecs_stats));
228         memset(runtime_cycles_stats, 0, sizeof(runtime_cycles_stats));
229         memset(runtime_stalled_cycles_front_stats, 0, sizeof(runtime_stalled_cycles_front_stats));
230         memset(runtime_stalled_cycles_back_stats, 0, sizeof(runtime_stalled_cycles_back_stats));
231         memset(runtime_branches_stats, 0, sizeof(runtime_branches_stats));
232         memset(runtime_cacherefs_stats, 0, sizeof(runtime_cacherefs_stats));
233         memset(runtime_l1_dcache_stats, 0, sizeof(runtime_l1_dcache_stats));
234         memset(runtime_l1_icache_stats, 0, sizeof(runtime_l1_icache_stats));
235         memset(runtime_ll_cache_stats, 0, sizeof(runtime_ll_cache_stats));
236         memset(runtime_itlb_cache_stats, 0, sizeof(runtime_itlb_cache_stats));
237         memset(runtime_dtlb_cache_stats, 0, sizeof(runtime_dtlb_cache_stats));
238         memset(&walltime_nsecs_stats, 0, sizeof(walltime_nsecs_stats));
239 }
240
241 static int create_perf_stat_counter(struct perf_evsel *evsel)
242 {
243         struct perf_event_attr *attr = &evsel->attr;
244
245         if (scale)
246                 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
247                                     PERF_FORMAT_TOTAL_TIME_RUNNING;
248
249         attr->inherit = !no_inherit;
250
251         if (perf_target__has_cpu(&target))
252                 return perf_evsel__open_per_cpu(evsel, perf_evsel__cpus(evsel));
253
254         if (!perf_target__has_task(&target) &&
255             perf_evsel__is_group_leader(evsel)) {
256                 attr->disabled = 1;
257                 attr->enable_on_exec = 1;
258         }
259
260         return perf_evsel__open_per_thread(evsel, evsel_list->threads);
261 }
262
263 /*
264  * Does the counter have nsecs as a unit?
265  */
266 static inline int nsec_counter(struct perf_evsel *evsel)
267 {
268         if (perf_evsel__match(evsel, SOFTWARE, SW_CPU_CLOCK) ||
269             perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
270                 return 1;
271
272         return 0;
273 }
274
275 /*
276  * Update various tracking values we maintain to print
277  * more semantic information such as miss/hit ratios,
278  * instruction rates, etc:
279  */
280 static void update_shadow_stats(struct perf_evsel *counter, u64 *count)
281 {
282         if (perf_evsel__match(counter, SOFTWARE, SW_TASK_CLOCK))
283                 update_stats(&runtime_nsecs_stats[0], count[0]);
284         else if (perf_evsel__match(counter, HARDWARE, HW_CPU_CYCLES))
285                 update_stats(&runtime_cycles_stats[0], count[0]);
286         else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_FRONTEND))
287                 update_stats(&runtime_stalled_cycles_front_stats[0], count[0]);
288         else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_BACKEND))
289                 update_stats(&runtime_stalled_cycles_back_stats[0], count[0]);
290         else if (perf_evsel__match(counter, HARDWARE, HW_BRANCH_INSTRUCTIONS))
291                 update_stats(&runtime_branches_stats[0], count[0]);
292         else if (perf_evsel__match(counter, HARDWARE, HW_CACHE_REFERENCES))
293                 update_stats(&runtime_cacherefs_stats[0], count[0]);
294         else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1D))
295                 update_stats(&runtime_l1_dcache_stats[0], count[0]);
296         else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1I))
297                 update_stats(&runtime_l1_icache_stats[0], count[0]);
298         else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_LL))
299                 update_stats(&runtime_ll_cache_stats[0], count[0]);
300         else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_DTLB))
301                 update_stats(&runtime_dtlb_cache_stats[0], count[0]);
302         else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_ITLB))
303                 update_stats(&runtime_itlb_cache_stats[0], count[0]);
304 }
305
306 /*
307  * Read out the results of a single counter:
308  * aggregate counts across CPUs in system-wide mode
309  */
310 static int read_counter_aggr(struct perf_evsel *counter)
311 {
312         struct perf_stat *ps = counter->priv;
313         u64 *count = counter->counts->aggr.values;
314         int i;
315
316         if (__perf_evsel__read(counter, perf_evsel__nr_cpus(counter),
317                                thread_map__nr(evsel_list->threads), scale) < 0)
318                 return -1;
319
320         for (i = 0; i < 3; i++)
321                 update_stats(&ps->res_stats[i], count[i]);
322
323         if (verbose) {
324                 fprintf(output, "%s: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
325                         perf_evsel__name(counter), count[0], count[1], count[2]);
326         }
327
328         /*
329          * Save the full runtime - to allow normalization during printout:
330          */
331         update_shadow_stats(counter, count);
332
333         return 0;
334 }
335
336 /*
337  * Read out the results of a single counter:
338  * do not aggregate counts across CPUs in system-wide mode
339  */
340 static int read_counter(struct perf_evsel *counter)
341 {
342         u64 *count;
343         int cpu;
344
345         for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
346                 if (__perf_evsel__read_on_cpu(counter, cpu, 0, scale) < 0)
347                         return -1;
348
349                 count = counter->counts->cpu[cpu].values;
350
351                 update_shadow_stats(counter, count);
352         }
353
354         return 0;
355 }
356
357 static void print_interval(void)
358 {
359         static int num_print_interval;
360         struct perf_evsel *counter;
361         struct perf_stat *ps;
362         struct timespec ts, rs;
363         char prefix[64];
364
365         if (aggr_mode == AGGR_GLOBAL) {
366                 list_for_each_entry(counter, &evsel_list->entries, node) {
367                         ps = counter->priv;
368                         memset(ps->res_stats, 0, sizeof(ps->res_stats));
369                         read_counter_aggr(counter);
370                 }
371         } else  {
372                 list_for_each_entry(counter, &evsel_list->entries, node) {
373                         ps = counter->priv;
374                         memset(ps->res_stats, 0, sizeof(ps->res_stats));
375                         read_counter(counter);
376                 }
377         }
378
379         clock_gettime(CLOCK_MONOTONIC, &ts);
380         diff_timespec(&rs, &ts, &ref_time);
381         sprintf(prefix, "%6lu.%09lu%s", rs.tv_sec, rs.tv_nsec, csv_sep);
382
383         if (num_print_interval == 0 && !csv_output) {
384                 switch (aggr_mode) {
385                 case AGGR_SOCKET:
386                         fprintf(output, "#           time socket cpus             counts events\n");
387                         break;
388                 case AGGR_CORE:
389                         fprintf(output, "#           time core         cpus             counts events\n");
390                         break;
391                 case AGGR_NONE:
392                         fprintf(output, "#           time CPU                 counts events\n");
393                         break;
394                 case AGGR_GLOBAL:
395                 default:
396                         fprintf(output, "#           time             counts events\n");
397                 }
398         }
399
400         if (++num_print_interval == 25)
401                 num_print_interval = 0;
402
403         switch (aggr_mode) {
404         case AGGR_CORE:
405         case AGGR_SOCKET:
406                 print_aggr(prefix);
407                 break;
408         case AGGR_NONE:
409                 list_for_each_entry(counter, &evsel_list->entries, node)
410                         print_counter(counter, prefix);
411                 break;
412         case AGGR_GLOBAL:
413         default:
414                 list_for_each_entry(counter, &evsel_list->entries, node)
415                         print_counter_aggr(counter, prefix);
416         }
417 }
418
419 static int __run_perf_stat(int argc, const char **argv)
420 {
421         char msg[512];
422         unsigned long long t0, t1;
423         struct perf_evsel *counter;
424         struct timespec ts;
425         int status = 0;
426         const bool forks = (argc > 0);
427
428         if (interval) {
429                 ts.tv_sec  = interval / 1000;
430                 ts.tv_nsec = (interval % 1000) * 1000000;
431         } else {
432                 ts.tv_sec  = 1;
433                 ts.tv_nsec = 0;
434         }
435
436         if (forks) {
437                 if (perf_evlist__prepare_workload(evsel_list, &target, argv,
438                                                   false, false) < 0) {
439                         perror("failed to prepare workload");
440                         return -1;
441                 }
442         }
443
444         if (group)
445                 perf_evlist__set_leader(evsel_list);
446
447         list_for_each_entry(counter, &evsel_list->entries, node) {
448                 if (create_perf_stat_counter(counter) < 0) {
449                         /*
450                          * PPC returns ENXIO for HW counters until 2.6.37
451                          * (behavior changed with commit b0a873e).
452                          */
453                         if (errno == EINVAL || errno == ENOSYS ||
454                             errno == ENOENT || errno == EOPNOTSUPP ||
455                             errno == ENXIO) {
456                                 if (verbose)
457                                         ui__warning("%s event is not supported by the kernel.\n",
458                                                     perf_evsel__name(counter));
459                                 counter->supported = false;
460                                 continue;
461                         }
462
463                         perf_evsel__open_strerror(counter, &target,
464                                                   errno, msg, sizeof(msg));
465                         ui__error("%s\n", msg);
466
467                         if (child_pid != -1)
468                                 kill(child_pid, SIGTERM);
469
470                         return -1;
471                 }
472                 counter->supported = true;
473         }
474
475         if (perf_evlist__apply_filters(evsel_list)) {
476                 error("failed to set filter with %d (%s)\n", errno,
477                         strerror(errno));
478                 return -1;
479         }
480
481         /*
482          * Enable counters and exec the command:
483          */
484         t0 = rdclock();
485         clock_gettime(CLOCK_MONOTONIC, &ref_time);
486
487         if (forks) {
488                 perf_evlist__start_workload(evsel_list);
489
490                 if (interval) {
491                         while (!waitpid(child_pid, &status, WNOHANG)) {
492                                 nanosleep(&ts, NULL);
493                                 print_interval();
494                         }
495                 }
496                 wait(&status);
497                 if (WIFSIGNALED(status))
498                         psignal(WTERMSIG(status), argv[0]);
499         } else {
500                 while (!done) {
501                         nanosleep(&ts, NULL);
502                         if (interval)
503                                 print_interval();
504                 }
505         }
506
507         t1 = rdclock();
508
509         update_stats(&walltime_nsecs_stats, t1 - t0);
510
511         if (aggr_mode == AGGR_GLOBAL) {
512                 list_for_each_entry(counter, &evsel_list->entries, node) {
513                         read_counter_aggr(counter);
514                         perf_evsel__close_fd(counter, perf_evsel__nr_cpus(counter),
515                                              thread_map__nr(evsel_list->threads));
516                 }
517         } else {
518                 list_for_each_entry(counter, &evsel_list->entries, node) {
519                         read_counter(counter);
520                         perf_evsel__close_fd(counter, perf_evsel__nr_cpus(counter), 1);
521                 }
522         }
523
524         return WEXITSTATUS(status);
525 }
526
527 static int run_perf_stat(int argc __maybe_unused, const char **argv)
528 {
529         int ret;
530
531         if (pre_cmd) {
532                 ret = system(pre_cmd);
533                 if (ret)
534                         return ret;
535         }
536
537         if (sync_run)
538                 sync();
539
540         ret = __run_perf_stat(argc, argv);
541         if (ret)
542                 return ret;
543
544         if (post_cmd) {
545                 ret = system(post_cmd);
546                 if (ret)
547                         return ret;
548         }
549
550         return ret;
551 }
552
553 static void print_noise_pct(double total, double avg)
554 {
555         double pct = rel_stddev_stats(total, avg);
556
557         if (csv_output)
558                 fprintf(output, "%s%.2f%%", csv_sep, pct);
559         else if (pct)
560                 fprintf(output, "  ( +-%6.2f%% )", pct);
561 }
562
563 static void print_noise(struct perf_evsel *evsel, double avg)
564 {
565         struct perf_stat *ps;
566
567         if (run_count == 1)
568                 return;
569
570         ps = evsel->priv;
571         print_noise_pct(stddev_stats(&ps->res_stats[0]), avg);
572 }
573
574 static void aggr_printout(struct perf_evsel *evsel, int id, int nr)
575 {
576         switch (aggr_mode) {
577         case AGGR_CORE:
578                 fprintf(output, "S%d-C%*d%s%*d%s",
579                         cpu_map__id_to_socket(id),
580                         csv_output ? 0 : -8,
581                         cpu_map__id_to_cpu(id),
582                         csv_sep,
583                         csv_output ? 0 : 4,
584                         nr,
585                         csv_sep);
586                 break;
587         case AGGR_SOCKET:
588                 fprintf(output, "S%*d%s%*d%s",
589                         csv_output ? 0 : -5,
590                         id,
591                         csv_sep,
592                         csv_output ? 0 : 4,
593                         nr,
594                         csv_sep);
595                         break;
596         case AGGR_NONE:
597                 fprintf(output, "CPU%*d%s",
598                         csv_output ? 0 : -4,
599                         perf_evsel__cpus(evsel)->map[id], csv_sep);
600                 break;
601         case AGGR_GLOBAL:
602         default:
603                 break;
604         }
605 }
606
607 static void nsec_printout(int cpu, int nr, struct perf_evsel *evsel, double avg)
608 {
609         double msecs = avg / 1e6;
610         const char *fmt = csv_output ? "%.6f%s%s" : "%18.6f%s%-25s";
611
612         aggr_printout(evsel, cpu, nr);
613
614         fprintf(output, fmt, msecs, csv_sep, perf_evsel__name(evsel));
615
616         if (evsel->cgrp)
617                 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
618
619         if (csv_output || interval)
620                 return;
621
622         if (perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
623                 fprintf(output, " # %8.3f CPUs utilized          ",
624                         avg / avg_stats(&walltime_nsecs_stats));
625         else
626                 fprintf(output, "                                   ");
627 }
628
629 /* used for get_ratio_color() */
630 enum grc_type {
631         GRC_STALLED_CYCLES_FE,
632         GRC_STALLED_CYCLES_BE,
633         GRC_CACHE_MISSES,
634         GRC_MAX_NR
635 };
636
637 static const char *get_ratio_color(enum grc_type type, double ratio)
638 {
639         static const double grc_table[GRC_MAX_NR][3] = {
640                 [GRC_STALLED_CYCLES_FE] = { 50.0, 30.0, 10.0 },
641                 [GRC_STALLED_CYCLES_BE] = { 75.0, 50.0, 20.0 },
642                 [GRC_CACHE_MISSES]      = { 20.0, 10.0, 5.0 },
643         };
644         const char *color = PERF_COLOR_NORMAL;
645
646         if (ratio > grc_table[type][0])
647                 color = PERF_COLOR_RED;
648         else if (ratio > grc_table[type][1])
649                 color = PERF_COLOR_MAGENTA;
650         else if (ratio > grc_table[type][2])
651                 color = PERF_COLOR_YELLOW;
652
653         return color;
654 }
655
656 static void print_stalled_cycles_frontend(int cpu,
657                                           struct perf_evsel *evsel
658                                           __maybe_unused, double avg)
659 {
660         double total, ratio = 0.0;
661         const char *color;
662
663         total = avg_stats(&runtime_cycles_stats[cpu]);
664
665         if (total)
666                 ratio = avg / total * 100.0;
667
668         color = get_ratio_color(GRC_STALLED_CYCLES_FE, ratio);
669
670         fprintf(output, " #  ");
671         color_fprintf(output, color, "%6.2f%%", ratio);
672         fprintf(output, " frontend cycles idle   ");
673 }
674
675 static void print_stalled_cycles_backend(int cpu,
676                                          struct perf_evsel *evsel
677                                          __maybe_unused, double avg)
678 {
679         double total, ratio = 0.0;
680         const char *color;
681
682         total = avg_stats(&runtime_cycles_stats[cpu]);
683
684         if (total)
685                 ratio = avg / total * 100.0;
686
687         color = get_ratio_color(GRC_STALLED_CYCLES_BE, ratio);
688
689         fprintf(output, " #  ");
690         color_fprintf(output, color, "%6.2f%%", ratio);
691         fprintf(output, " backend  cycles idle   ");
692 }
693
694 static void print_branch_misses(int cpu,
695                                 struct perf_evsel *evsel __maybe_unused,
696                                 double avg)
697 {
698         double total, ratio = 0.0;
699         const char *color;
700
701         total = avg_stats(&runtime_branches_stats[cpu]);
702
703         if (total)
704                 ratio = avg / total * 100.0;
705
706         color = get_ratio_color(GRC_CACHE_MISSES, ratio);
707
708         fprintf(output, " #  ");
709         color_fprintf(output, color, "%6.2f%%", ratio);
710         fprintf(output, " of all branches        ");
711 }
712
713 static void print_l1_dcache_misses(int cpu,
714                                    struct perf_evsel *evsel __maybe_unused,
715                                    double avg)
716 {
717         double total, ratio = 0.0;
718         const char *color;
719
720         total = avg_stats(&runtime_l1_dcache_stats[cpu]);
721
722         if (total)
723                 ratio = avg / total * 100.0;
724
725         color = get_ratio_color(GRC_CACHE_MISSES, ratio);
726
727         fprintf(output, " #  ");
728         color_fprintf(output, color, "%6.2f%%", ratio);
729         fprintf(output, " of all L1-dcache hits  ");
730 }
731
732 static void print_l1_icache_misses(int cpu,
733                                    struct perf_evsel *evsel __maybe_unused,
734                                    double avg)
735 {
736         double total, ratio = 0.0;
737         const char *color;
738
739         total = avg_stats(&runtime_l1_icache_stats[cpu]);
740
741         if (total)
742                 ratio = avg / total * 100.0;
743
744         color = get_ratio_color(GRC_CACHE_MISSES, ratio);
745
746         fprintf(output, " #  ");
747         color_fprintf(output, color, "%6.2f%%", ratio);
748         fprintf(output, " of all L1-icache hits  ");
749 }
750
751 static void print_dtlb_cache_misses(int cpu,
752                                     struct perf_evsel *evsel __maybe_unused,
753                                     double avg)
754 {
755         double total, ratio = 0.0;
756         const char *color;
757
758         total = avg_stats(&runtime_dtlb_cache_stats[cpu]);
759
760         if (total)
761                 ratio = avg / total * 100.0;
762
763         color = get_ratio_color(GRC_CACHE_MISSES, ratio);
764
765         fprintf(output, " #  ");
766         color_fprintf(output, color, "%6.2f%%", ratio);
767         fprintf(output, " of all dTLB cache hits ");
768 }
769
770 static void print_itlb_cache_misses(int cpu,
771                                     struct perf_evsel *evsel __maybe_unused,
772                                     double avg)
773 {
774         double total, ratio = 0.0;
775         const char *color;
776
777         total = avg_stats(&runtime_itlb_cache_stats[cpu]);
778
779         if (total)
780                 ratio = avg / total * 100.0;
781
782         color = get_ratio_color(GRC_CACHE_MISSES, ratio);
783
784         fprintf(output, " #  ");
785         color_fprintf(output, color, "%6.2f%%", ratio);
786         fprintf(output, " of all iTLB cache hits ");
787 }
788
789 static void print_ll_cache_misses(int cpu,
790                                   struct perf_evsel *evsel __maybe_unused,
791                                   double avg)
792 {
793         double total, ratio = 0.0;
794         const char *color;
795
796         total = avg_stats(&runtime_ll_cache_stats[cpu]);
797
798         if (total)
799                 ratio = avg / total * 100.0;
800
801         color = get_ratio_color(GRC_CACHE_MISSES, ratio);
802
803         fprintf(output, " #  ");
804         color_fprintf(output, color, "%6.2f%%", ratio);
805         fprintf(output, " of all LL-cache hits   ");
806 }
807
808 static void abs_printout(int cpu, int nr, struct perf_evsel *evsel, double avg)
809 {
810         double total, ratio = 0.0;
811         const char *fmt;
812
813         if (csv_output)
814                 fmt = "%.0f%s%s";
815         else if (big_num)
816                 fmt = "%'18.0f%s%-25s";
817         else
818                 fmt = "%18.0f%s%-25s";
819
820         aggr_printout(evsel, cpu, nr);
821
822         if (aggr_mode == AGGR_GLOBAL)
823                 cpu = 0;
824
825         fprintf(output, fmt, avg, csv_sep, perf_evsel__name(evsel));
826
827         if (evsel->cgrp)
828                 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
829
830         if (csv_output || interval)
831                 return;
832
833         if (perf_evsel__match(evsel, HARDWARE, HW_INSTRUCTIONS)) {
834                 total = avg_stats(&runtime_cycles_stats[cpu]);
835                 if (total)
836                         ratio = avg / total;
837
838                 fprintf(output, " #   %5.2f  insns per cycle        ", ratio);
839
840                 total = avg_stats(&runtime_stalled_cycles_front_stats[cpu]);
841                 total = max(total, avg_stats(&runtime_stalled_cycles_back_stats[cpu]));
842
843                 if (total && avg) {
844                         ratio = total / avg;
845                         fprintf(output, "\n                                             #   %5.2f  stalled cycles per insn", ratio);
846                 }
847
848         } else if (perf_evsel__match(evsel, HARDWARE, HW_BRANCH_MISSES) &&
849                         runtime_branches_stats[cpu].n != 0) {
850                 print_branch_misses(cpu, evsel, avg);
851         } else if (
852                 evsel->attr.type == PERF_TYPE_HW_CACHE &&
853                 evsel->attr.config ==  ( PERF_COUNT_HW_CACHE_L1D |
854                                         ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
855                                         ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
856                         runtime_l1_dcache_stats[cpu].n != 0) {
857                 print_l1_dcache_misses(cpu, evsel, avg);
858         } else if (
859                 evsel->attr.type == PERF_TYPE_HW_CACHE &&
860                 evsel->attr.config ==  ( PERF_COUNT_HW_CACHE_L1I |
861                                         ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
862                                         ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
863                         runtime_l1_icache_stats[cpu].n != 0) {
864                 print_l1_icache_misses(cpu, evsel, avg);
865         } else if (
866                 evsel->attr.type == PERF_TYPE_HW_CACHE &&
867                 evsel->attr.config ==  ( PERF_COUNT_HW_CACHE_DTLB |
868                                         ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
869                                         ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
870                         runtime_dtlb_cache_stats[cpu].n != 0) {
871                 print_dtlb_cache_misses(cpu, evsel, avg);
872         } else if (
873                 evsel->attr.type == PERF_TYPE_HW_CACHE &&
874                 evsel->attr.config ==  ( PERF_COUNT_HW_CACHE_ITLB |
875                                         ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
876                                         ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
877                         runtime_itlb_cache_stats[cpu].n != 0) {
878                 print_itlb_cache_misses(cpu, evsel, avg);
879         } else if (
880                 evsel->attr.type == PERF_TYPE_HW_CACHE &&
881                 evsel->attr.config ==  ( PERF_COUNT_HW_CACHE_LL |
882                                         ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
883                                         ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
884                         runtime_ll_cache_stats[cpu].n != 0) {
885                 print_ll_cache_misses(cpu, evsel, avg);
886         } else if (perf_evsel__match(evsel, HARDWARE, HW_CACHE_MISSES) &&
887                         runtime_cacherefs_stats[cpu].n != 0) {
888                 total = avg_stats(&runtime_cacherefs_stats[cpu]);
889
890                 if (total)
891                         ratio = avg * 100 / total;
892
893                 fprintf(output, " # %8.3f %% of all cache refs    ", ratio);
894
895         } else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_FRONTEND)) {
896                 print_stalled_cycles_frontend(cpu, evsel, avg);
897         } else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_BACKEND)) {
898                 print_stalled_cycles_backend(cpu, evsel, avg);
899         } else if (perf_evsel__match(evsel, HARDWARE, HW_CPU_CYCLES)) {
900                 total = avg_stats(&runtime_nsecs_stats[cpu]);
901
902                 if (total)
903                         ratio = 1.0 * avg / total;
904
905                 fprintf(output, " # %8.3f GHz                    ", ratio);
906         } else if (runtime_nsecs_stats[cpu].n != 0) {
907                 char unit = 'M';
908
909                 total = avg_stats(&runtime_nsecs_stats[cpu]);
910
911                 if (total)
912                         ratio = 1000.0 * avg / total;
913                 if (ratio < 0.001) {
914                         ratio *= 1000;
915                         unit = 'K';
916                 }
917
918                 fprintf(output, " # %8.3f %c/sec                  ", ratio, unit);
919         } else {
920                 fprintf(output, "                                   ");
921         }
922 }
923
924 static void print_aggr(char *prefix)
925 {
926         struct perf_evsel *counter;
927         int cpu, s, s2, id, nr;
928         u64 ena, run, val;
929
930         if (!(aggr_map || aggr_get_id))
931                 return;
932
933         for (s = 0; s < aggr_map->nr; s++) {
934                 id = aggr_map->map[s];
935                 list_for_each_entry(counter, &evsel_list->entries, node) {
936                         val = ena = run = 0;
937                         nr = 0;
938                         for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
939                                 s2 = aggr_get_id(evsel_list->cpus, cpu);
940                                 if (s2 != id)
941                                         continue;
942                                 val += counter->counts->cpu[cpu].val;
943                                 ena += counter->counts->cpu[cpu].ena;
944                                 run += counter->counts->cpu[cpu].run;
945                                 nr++;
946                         }
947                         if (prefix)
948                                 fprintf(output, "%s", prefix);
949
950                         if (run == 0 || ena == 0) {
951                                 aggr_printout(counter, cpu, nr);
952
953                                 fprintf(output, "%*s%s%*s",
954                                         csv_output ? 0 : 18,
955                                         counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
956                                         csv_sep,
957                                         csv_output ? 0 : -24,
958                                         perf_evsel__name(counter));
959
960                                 if (counter->cgrp)
961                                         fprintf(output, "%s%s",
962                                                 csv_sep, counter->cgrp->name);
963
964                                 fputc('\n', output);
965                                 continue;
966                         }
967
968                         if (nsec_counter(counter))
969                                 nsec_printout(id, nr, counter, val);
970                         else
971                                 abs_printout(id, nr, counter, val);
972
973                         if (!csv_output) {
974                                 print_noise(counter, 1.0);
975
976                                 if (run != ena)
977                                         fprintf(output, "  (%.2f%%)",
978                                                 100.0 * run / ena);
979                         }
980                         fputc('\n', output);
981                 }
982         }
983 }
984
985 /*
986  * Print out the results of a single counter:
987  * aggregated counts in system-wide mode
988  */
989 static void print_counter_aggr(struct perf_evsel *counter, char *prefix)
990 {
991         struct perf_stat *ps = counter->priv;
992         double avg = avg_stats(&ps->res_stats[0]);
993         int scaled = counter->counts->scaled;
994
995         if (prefix)
996                 fprintf(output, "%s", prefix);
997
998         if (scaled == -1) {
999                 fprintf(output, "%*s%s%*s",
1000                         csv_output ? 0 : 18,
1001                         counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
1002                         csv_sep,
1003                         csv_output ? 0 : -24,
1004                         perf_evsel__name(counter));
1005
1006                 if (counter->cgrp)
1007                         fprintf(output, "%s%s", csv_sep, counter->cgrp->name);
1008
1009                 fputc('\n', output);
1010                 return;
1011         }
1012
1013         if (nsec_counter(counter))
1014                 nsec_printout(-1, 0, counter, avg);
1015         else
1016                 abs_printout(-1, 0, counter, avg);
1017
1018         print_noise(counter, avg);
1019
1020         if (csv_output) {
1021                 fputc('\n', output);
1022                 return;
1023         }
1024
1025         if (scaled) {
1026                 double avg_enabled, avg_running;
1027
1028                 avg_enabled = avg_stats(&ps->res_stats[1]);
1029                 avg_running = avg_stats(&ps->res_stats[2]);
1030
1031                 fprintf(output, " [%5.2f%%]", 100 * avg_running / avg_enabled);
1032         }
1033         fprintf(output, "\n");
1034 }
1035
1036 /*
1037  * Print out the results of a single counter:
1038  * does not use aggregated count in system-wide
1039  */
1040 static void print_counter(struct perf_evsel *counter, char *prefix)
1041 {
1042         u64 ena, run, val;
1043         int cpu;
1044
1045         for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
1046                 val = counter->counts->cpu[cpu].val;
1047                 ena = counter->counts->cpu[cpu].ena;
1048                 run = counter->counts->cpu[cpu].run;
1049
1050                 if (prefix)
1051                         fprintf(output, "%s", prefix);
1052
1053                 if (run == 0 || ena == 0) {
1054                         fprintf(output, "CPU%*d%s%*s%s%*s",
1055                                 csv_output ? 0 : -4,
1056                                 perf_evsel__cpus(counter)->map[cpu], csv_sep,
1057                                 csv_output ? 0 : 18,
1058                                 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
1059                                 csv_sep,
1060                                 csv_output ? 0 : -24,
1061                                 perf_evsel__name(counter));
1062
1063                         if (counter->cgrp)
1064                                 fprintf(output, "%s%s",
1065                                         csv_sep, counter->cgrp->name);
1066
1067                         fputc('\n', output);
1068                         continue;
1069                 }
1070
1071                 if (nsec_counter(counter))
1072                         nsec_printout(cpu, 0, counter, val);
1073                 else
1074                         abs_printout(cpu, 0, counter, val);
1075
1076                 if (!csv_output) {
1077                         print_noise(counter, 1.0);
1078
1079                         if (run != ena)
1080                                 fprintf(output, "  (%.2f%%)",
1081                                         100.0 * run / ena);
1082                 }
1083                 fputc('\n', output);
1084         }
1085 }
1086
1087 static void print_stat(int argc, const char **argv)
1088 {
1089         struct perf_evsel *counter;
1090         int i;
1091
1092         fflush(stdout);
1093
1094         if (!csv_output) {
1095                 fprintf(output, "\n");
1096                 fprintf(output, " Performance counter stats for ");
1097                 if (!perf_target__has_task(&target)) {
1098                         fprintf(output, "\'%s", argv[0]);
1099                         for (i = 1; i < argc; i++)
1100                                 fprintf(output, " %s", argv[i]);
1101                 } else if (target.pid)
1102                         fprintf(output, "process id \'%s", target.pid);
1103                 else
1104                         fprintf(output, "thread id \'%s", target.tid);
1105
1106                 fprintf(output, "\'");
1107                 if (run_count > 1)
1108                         fprintf(output, " (%d runs)", run_count);
1109                 fprintf(output, ":\n\n");
1110         }
1111
1112         switch (aggr_mode) {
1113         case AGGR_CORE:
1114         case AGGR_SOCKET:
1115                 print_aggr(NULL);
1116                 break;
1117         case AGGR_GLOBAL:
1118                 list_for_each_entry(counter, &evsel_list->entries, node)
1119                         print_counter_aggr(counter, NULL);
1120                 break;
1121         case AGGR_NONE:
1122                 list_for_each_entry(counter, &evsel_list->entries, node)
1123                         print_counter(counter, NULL);
1124                 break;
1125         default:
1126                 break;
1127         }
1128
1129         if (!csv_output) {
1130                 if (!null_run)
1131                         fprintf(output, "\n");
1132                 fprintf(output, " %17.9f seconds time elapsed",
1133                                 avg_stats(&walltime_nsecs_stats)/1e9);
1134                 if (run_count > 1) {
1135                         fprintf(output, "                                        ");
1136                         print_noise_pct(stddev_stats(&walltime_nsecs_stats),
1137                                         avg_stats(&walltime_nsecs_stats));
1138                 }
1139                 fprintf(output, "\n\n");
1140         }
1141 }
1142
1143 static volatile int signr = -1;
1144
1145 static void skip_signal(int signo)
1146 {
1147         if ((child_pid == -1) || interval)
1148                 done = 1;
1149
1150         signr = signo;
1151 }
1152
1153 static void sig_atexit(void)
1154 {
1155         if (child_pid != -1)
1156                 kill(child_pid, SIGTERM);
1157
1158         if (signr == -1)
1159                 return;
1160
1161         signal(signr, SIG_DFL);
1162         kill(getpid(), signr);
1163 }
1164
1165 static int stat__set_big_num(const struct option *opt __maybe_unused,
1166                              const char *s __maybe_unused, int unset)
1167 {
1168         big_num_opt = unset ? 0 : 1;
1169         return 0;
1170 }
1171
1172 static int perf_stat_init_aggr_mode(void)
1173 {
1174         switch (aggr_mode) {
1175         case AGGR_SOCKET:
1176                 if (cpu_map__build_socket_map(evsel_list->cpus, &aggr_map)) {
1177                         perror("cannot build socket map");
1178                         return -1;
1179                 }
1180                 aggr_get_id = cpu_map__get_socket;
1181                 break;
1182         case AGGR_CORE:
1183                 if (cpu_map__build_core_map(evsel_list->cpus, &aggr_map)) {
1184                         perror("cannot build core map");
1185                         return -1;
1186                 }
1187                 aggr_get_id = cpu_map__get_core;
1188                 break;
1189         case AGGR_NONE:
1190         case AGGR_GLOBAL:
1191         default:
1192                 break;
1193         }
1194         return 0;
1195 }
1196
1197
1198 /*
1199  * Add default attributes, if there were no attributes specified or
1200  * if -d/--detailed, -d -d or -d -d -d is used:
1201  */
1202 static int add_default_attributes(void)
1203 {
1204         struct perf_event_attr default_attrs[] = {
1205
1206   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK              },
1207   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES        },
1208   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS          },
1209   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS             },
1210
1211   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES              },
1212   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND },
1213   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND  },
1214   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS            },
1215   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS     },
1216   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES           },
1217
1218 };
1219
1220 /*
1221  * Detailed stats (-d), covering the L1 and last level data caches:
1222  */
1223         struct perf_event_attr detailed_attrs[] = {
1224
1225   { .type = PERF_TYPE_HW_CACHE,
1226     .config =
1227          PERF_COUNT_HW_CACHE_L1D                <<  0  |
1228         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1229         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
1230
1231   { .type = PERF_TYPE_HW_CACHE,
1232     .config =
1233          PERF_COUNT_HW_CACHE_L1D                <<  0  |
1234         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1235         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
1236
1237   { .type = PERF_TYPE_HW_CACHE,
1238     .config =
1239          PERF_COUNT_HW_CACHE_LL                 <<  0  |
1240         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1241         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
1242
1243   { .type = PERF_TYPE_HW_CACHE,
1244     .config =
1245          PERF_COUNT_HW_CACHE_LL                 <<  0  |
1246         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1247         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
1248 };
1249
1250 /*
1251  * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
1252  */
1253         struct perf_event_attr very_detailed_attrs[] = {
1254
1255   { .type = PERF_TYPE_HW_CACHE,
1256     .config =
1257          PERF_COUNT_HW_CACHE_L1I                <<  0  |
1258         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1259         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
1260
1261   { .type = PERF_TYPE_HW_CACHE,
1262     .config =
1263          PERF_COUNT_HW_CACHE_L1I                <<  0  |
1264         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1265         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
1266
1267   { .type = PERF_TYPE_HW_CACHE,
1268     .config =
1269          PERF_COUNT_HW_CACHE_DTLB               <<  0  |
1270         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1271         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
1272
1273   { .type = PERF_TYPE_HW_CACHE,
1274     .config =
1275          PERF_COUNT_HW_CACHE_DTLB               <<  0  |
1276         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1277         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
1278
1279   { .type = PERF_TYPE_HW_CACHE,
1280     .config =
1281          PERF_COUNT_HW_CACHE_ITLB               <<  0  |
1282         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1283         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
1284
1285   { .type = PERF_TYPE_HW_CACHE,
1286     .config =
1287          PERF_COUNT_HW_CACHE_ITLB               <<  0  |
1288         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1289         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
1290
1291 };
1292
1293 /*
1294  * Very, very detailed stats (-d -d -d), adding prefetch events:
1295  */
1296         struct perf_event_attr very_very_detailed_attrs[] = {
1297
1298   { .type = PERF_TYPE_HW_CACHE,
1299     .config =
1300          PERF_COUNT_HW_CACHE_L1D                <<  0  |
1301         (PERF_COUNT_HW_CACHE_OP_PREFETCH        <<  8) |
1302         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
1303
1304   { .type = PERF_TYPE_HW_CACHE,
1305     .config =
1306          PERF_COUNT_HW_CACHE_L1D                <<  0  |
1307         (PERF_COUNT_HW_CACHE_OP_PREFETCH        <<  8) |
1308         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
1309 };
1310
1311         /* Set attrs if no event is selected and !null_run: */
1312         if (null_run)
1313                 return 0;
1314
1315         if (!evsel_list->nr_entries) {
1316                 if (perf_evlist__add_default_attrs(evsel_list, default_attrs) < 0)
1317                         return -1;
1318         }
1319
1320         /* Detailed events get appended to the event list: */
1321
1322         if (detailed_run <  1)
1323                 return 0;
1324
1325         /* Append detailed run extra attributes: */
1326         if (perf_evlist__add_default_attrs(evsel_list, detailed_attrs) < 0)
1327                 return -1;
1328
1329         if (detailed_run < 2)
1330                 return 0;
1331
1332         /* Append very detailed run extra attributes: */
1333         if (perf_evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0)
1334                 return -1;
1335
1336         if (detailed_run < 3)
1337                 return 0;
1338
1339         /* Append very, very detailed run extra attributes: */
1340         return perf_evlist__add_default_attrs(evsel_list, very_very_detailed_attrs);
1341 }
1342
1343 int cmd_stat(int argc, const char **argv, const char *prefix __maybe_unused)
1344 {
1345         bool append_file = false;
1346         int output_fd = 0;
1347         const char *output_name = NULL;
1348         const struct option options[] = {
1349         OPT_CALLBACK('e', "event", &evsel_list, "event",
1350                      "event selector. use 'perf list' to list available events",
1351                      parse_events_option),
1352         OPT_CALLBACK(0, "filter", &evsel_list, "filter",
1353                      "event filter", parse_filter),
1354         OPT_BOOLEAN('i', "no-inherit", &no_inherit,
1355                     "child tasks do not inherit counters"),
1356         OPT_STRING('p', "pid", &target.pid, "pid",
1357                    "stat events on existing process id"),
1358         OPT_STRING('t', "tid", &target.tid, "tid",
1359                    "stat events on existing thread id"),
1360         OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
1361                     "system-wide collection from all CPUs"),
1362         OPT_BOOLEAN('g', "group", &group,
1363                     "put the counters into a counter group"),
1364         OPT_BOOLEAN('c', "scale", &scale, "scale/normalize counters"),
1365         OPT_INCR('v', "verbose", &verbose,
1366                     "be more verbose (show counter open errors, etc)"),
1367         OPT_INTEGER('r', "repeat", &run_count,
1368                     "repeat command and print average + stddev (max: 100, forever: 0)"),
1369         OPT_BOOLEAN('n', "null", &null_run,
1370                     "null run - dont start any counters"),
1371         OPT_INCR('d', "detailed", &detailed_run,
1372                     "detailed run - start a lot of events"),
1373         OPT_BOOLEAN('S', "sync", &sync_run,
1374                     "call sync() before starting a run"),
1375         OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL, 
1376                            "print large numbers with thousands\' separators",
1377                            stat__set_big_num),
1378         OPT_STRING('C', "cpu", &target.cpu_list, "cpu",
1379                     "list of cpus to monitor in system-wide"),
1380         OPT_SET_UINT('A', "no-aggr", &aggr_mode,
1381                     "disable CPU count aggregation", AGGR_NONE),
1382         OPT_STRING('x', "field-separator", &csv_sep, "separator",
1383                    "print counts with custom separator"),
1384         OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
1385                      "monitor event in cgroup name only", parse_cgroups),
1386         OPT_STRING('o', "output", &output_name, "file", "output file name"),
1387         OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
1388         OPT_INTEGER(0, "log-fd", &output_fd,
1389                     "log output to fd, instead of stderr"),
1390         OPT_STRING(0, "pre", &pre_cmd, "command",
1391                         "command to run prior to the measured command"),
1392         OPT_STRING(0, "post", &post_cmd, "command",
1393                         "command to run after to the measured command"),
1394         OPT_UINTEGER('I', "interval-print", &interval,
1395                     "print counts at regular interval in ms (>= 100)"),
1396         OPT_SET_UINT(0, "per-socket", &aggr_mode,
1397                      "aggregate counts per processor socket", AGGR_SOCKET),
1398         OPT_SET_UINT(0, "per-core", &aggr_mode,
1399                      "aggregate counts per physical processor core", AGGR_CORE),
1400         OPT_END()
1401         };
1402         const char * const stat_usage[] = {
1403                 "perf stat [<options>] [<command>]",
1404                 NULL
1405         };
1406         int status = -ENOMEM, run_idx;
1407         const char *mode;
1408
1409         setlocale(LC_ALL, "");
1410
1411         evsel_list = perf_evlist__new();
1412         if (evsel_list == NULL)
1413                 return -ENOMEM;
1414
1415         argc = parse_options(argc, argv, options, stat_usage,
1416                 PARSE_OPT_STOP_AT_NON_OPTION);
1417
1418         output = stderr;
1419         if (output_name && strcmp(output_name, "-"))
1420                 output = NULL;
1421
1422         if (output_name && output_fd) {
1423                 fprintf(stderr, "cannot use both --output and --log-fd\n");
1424                 usage_with_options(stat_usage, options);
1425         }
1426
1427         if (output_fd < 0) {
1428                 fprintf(stderr, "argument to --log-fd must be a > 0\n");
1429                 usage_with_options(stat_usage, options);
1430         }
1431
1432         if (!output) {
1433                 struct timespec tm;
1434                 mode = append_file ? "a" : "w";
1435
1436                 output = fopen(output_name, mode);
1437                 if (!output) {
1438                         perror("failed to create output file");
1439                         return -1;
1440                 }
1441                 clock_gettime(CLOCK_REALTIME, &tm);
1442                 fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
1443         } else if (output_fd > 0) {
1444                 mode = append_file ? "a" : "w";
1445                 output = fdopen(output_fd, mode);
1446                 if (!output) {
1447                         perror("Failed opening logfd");
1448                         return -errno;
1449                 }
1450         }
1451
1452         if (csv_sep) {
1453                 csv_output = true;
1454                 if (!strcmp(csv_sep, "\\t"))
1455                         csv_sep = "\t";
1456         } else
1457                 csv_sep = DEFAULT_SEPARATOR;
1458
1459         /*
1460          * let the spreadsheet do the pretty-printing
1461          */
1462         if (csv_output) {
1463                 /* User explicitly passed -B? */
1464                 if (big_num_opt == 1) {
1465                         fprintf(stderr, "-B option not supported with -x\n");
1466                         usage_with_options(stat_usage, options);
1467                 } else /* Nope, so disable big number formatting */
1468                         big_num = false;
1469         } else if (big_num_opt == 0) /* User passed --no-big-num */
1470                 big_num = false;
1471
1472         if (!argc && !perf_target__has_task(&target))
1473                 usage_with_options(stat_usage, options);
1474         if (run_count < 0) {
1475                 usage_with_options(stat_usage, options);
1476         } else if (run_count == 0) {
1477                 forever = true;
1478                 run_count = 1;
1479         }
1480
1481         /* no_aggr, cgroup are for system-wide only */
1482         if ((aggr_mode != AGGR_GLOBAL || nr_cgroups)
1483              && !perf_target__has_cpu(&target)) {
1484                 fprintf(stderr, "both cgroup and no-aggregation "
1485                         "modes only available in system-wide mode\n");
1486
1487                 usage_with_options(stat_usage, options);
1488                 return -1;
1489         }
1490
1491         if (add_default_attributes())
1492                 goto out;
1493
1494         perf_target__validate(&target);
1495
1496         if (perf_evlist__create_maps(evsel_list, &target) < 0) {
1497                 if (perf_target__has_task(&target))
1498                         pr_err("Problems finding threads of monitor\n");
1499                 if (perf_target__has_cpu(&target))
1500                         perror("failed to parse CPUs map");
1501
1502                 usage_with_options(stat_usage, options);
1503                 return -1;
1504         }
1505         if (interval && interval < 100) {
1506                 pr_err("print interval must be >= 100ms\n");
1507                 usage_with_options(stat_usage, options);
1508                 return -1;
1509         }
1510
1511         if (perf_evlist__alloc_stats(evsel_list, interval))
1512                 goto out_free_maps;
1513
1514         if (perf_stat_init_aggr_mode())
1515                 goto out;
1516
1517         /*
1518          * We dont want to block the signals - that would cause
1519          * child tasks to inherit that and Ctrl-C would not work.
1520          * What we want is for Ctrl-C to work in the exec()-ed
1521          * task, but being ignored by perf stat itself:
1522          */
1523         atexit(sig_atexit);
1524         if (!forever)
1525                 signal(SIGINT,  skip_signal);
1526         signal(SIGCHLD, skip_signal);
1527         signal(SIGALRM, skip_signal);
1528         signal(SIGABRT, skip_signal);
1529
1530         status = 0;
1531         for (run_idx = 0; forever || run_idx < run_count; run_idx++) {
1532                 if (run_count != 1 && verbose)
1533                         fprintf(output, "[ perf stat: executing run #%d ... ]\n",
1534                                 run_idx + 1);
1535
1536                 status = run_perf_stat(argc, argv);
1537                 if (forever && status != -1) {
1538                         print_stat(argc, argv);
1539                         perf_stat__reset_stats(evsel_list);
1540                 }
1541         }
1542
1543         if (!forever && status != -1 && !interval)
1544                 print_stat(argc, argv);
1545
1546         perf_evlist__free_stats(evsel_list);
1547 out_free_maps:
1548         perf_evlist__delete_maps(evsel_list);
1549 out:
1550         perf_evlist__delete(evsel_list);
1551         return status;
1552 }