perf stat: Fix IPC and other formulas with -A
[firefly-linux-kernel-4.4.55.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/cgroup.h"
47 #include "util/util.h"
48 #include "util/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
62 #include <stdlib.h>
63 #include <sys/prctl.h>
64 #include <locale.h>
65
66 #define DEFAULT_SEPARATOR       " "
67 #define CNTR_NOT_SUPPORTED      "<not supported>"
68 #define CNTR_NOT_COUNTED        "<not counted>"
69
70 static void print_stat(int argc, const char **argv);
71 static void print_counter_aggr(struct perf_evsel *counter, char *prefix);
72 static void print_counter(struct perf_evsel *counter, char *prefix);
73 static void print_aggr(char *prefix);
74
75 /* Default events used for perf stat -T */
76 static const char * const transaction_attrs[] = {
77         "task-clock",
78         "{"
79         "instructions,"
80         "cycles,"
81         "cpu/cycles-t/,"
82         "cpu/tx-start/,"
83         "cpu/el-start/,"
84         "cpu/cycles-ct/"
85         "}"
86 };
87
88 /* More limited version when the CPU does not have all events. */
89 static const char * const transaction_limited_attrs[] = {
90         "task-clock",
91         "{"
92         "instructions,"
93         "cycles,"
94         "cpu/cycles-t/,"
95         "cpu/tx-start/"
96         "}"
97 };
98
99 /* must match transaction_attrs and the beginning limited_attrs */
100 enum {
101         T_TASK_CLOCK,
102         T_INSTRUCTIONS,
103         T_CYCLES,
104         T_CYCLES_IN_TX,
105         T_TRANSACTION_START,
106         T_ELISION_START,
107         T_CYCLES_IN_TX_CP,
108 };
109
110 static struct perf_evlist       *evsel_list;
111
112 static struct target target = {
113         .uid    = UINT_MAX,
114 };
115
116 enum aggr_mode {
117         AGGR_NONE,
118         AGGR_GLOBAL,
119         AGGR_SOCKET,
120         AGGR_CORE,
121 };
122
123 static int                      run_count                       =  1;
124 static bool                     no_inherit                      = false;
125 static bool                     scale                           =  true;
126 static enum aggr_mode           aggr_mode                       = AGGR_GLOBAL;
127 static volatile pid_t           child_pid                       = -1;
128 static bool                     null_run                        =  false;
129 static int                      detailed_run                    =  0;
130 static bool                     transaction_run;
131 static bool                     big_num                         =  true;
132 static int                      big_num_opt                     =  -1;
133 static const char               *csv_sep                        = NULL;
134 static bool                     csv_output                      = false;
135 static bool                     group                           = false;
136 static FILE                     *output                         = NULL;
137 static const char               *pre_cmd                        = NULL;
138 static const char               *post_cmd                       = NULL;
139 static bool                     sync_run                        = false;
140 static unsigned int             interval                        = 0;
141 static unsigned int             initial_delay                   = 0;
142 static unsigned int             unit_width                      = 4; /* strlen("unit") */
143 static bool                     forever                         = false;
144 static struct timespec          ref_time;
145 static struct cpu_map           *aggr_map;
146 static int                      (*aggr_get_id)(struct cpu_map *m, int cpu);
147
148 static volatile int done = 0;
149
150 struct perf_stat {
151         struct stats      res_stats[3];
152 };
153
154 static inline void diff_timespec(struct timespec *r, struct timespec *a,
155                                  struct timespec *b)
156 {
157         r->tv_sec = a->tv_sec - b->tv_sec;
158         if (a->tv_nsec < b->tv_nsec) {
159                 r->tv_nsec = a->tv_nsec + 1000000000L - b->tv_nsec;
160                 r->tv_sec--;
161         } else {
162                 r->tv_nsec = a->tv_nsec - b->tv_nsec ;
163         }
164 }
165
166 static inline struct cpu_map *perf_evsel__cpus(struct perf_evsel *evsel)
167 {
168         return (evsel->cpus && !target.cpu_list) ? evsel->cpus : evsel_list->cpus;
169 }
170
171 static inline int perf_evsel__nr_cpus(struct perf_evsel *evsel)
172 {
173         return perf_evsel__cpus(evsel)->nr;
174 }
175
176 static void perf_evsel__reset_stat_priv(struct perf_evsel *evsel)
177 {
178         int i;
179         struct perf_stat *ps = evsel->priv;
180
181         for (i = 0; i < 3; i++)
182                 init_stats(&ps->res_stats[i]);
183 }
184
185 static int perf_evsel__alloc_stat_priv(struct perf_evsel *evsel)
186 {
187         evsel->priv = zalloc(sizeof(struct perf_stat));
188         if (evsel->priv == NULL)
189                 return -ENOMEM;
190         perf_evsel__reset_stat_priv(evsel);
191         return 0;
192 }
193
194 static void perf_evsel__free_stat_priv(struct perf_evsel *evsel)
195 {
196         zfree(&evsel->priv);
197 }
198
199 static int perf_evsel__alloc_prev_raw_counts(struct perf_evsel *evsel)
200 {
201         void *addr;
202         size_t sz;
203
204         sz = sizeof(*evsel->counts) +
205              (perf_evsel__nr_cpus(evsel) * sizeof(struct perf_counts_values));
206
207         addr = zalloc(sz);
208         if (!addr)
209                 return -ENOMEM;
210
211         evsel->prev_raw_counts =  addr;
212
213         return 0;
214 }
215
216 static void perf_evsel__free_prev_raw_counts(struct perf_evsel *evsel)
217 {
218         zfree(&evsel->prev_raw_counts);
219 }
220
221 static void perf_evlist__free_stats(struct perf_evlist *evlist)
222 {
223         struct perf_evsel *evsel;
224
225         evlist__for_each(evlist, evsel) {
226                 perf_evsel__free_stat_priv(evsel);
227                 perf_evsel__free_counts(evsel);
228                 perf_evsel__free_prev_raw_counts(evsel);
229         }
230 }
231
232 static int perf_evlist__alloc_stats(struct perf_evlist *evlist, bool alloc_raw)
233 {
234         struct perf_evsel *evsel;
235
236         evlist__for_each(evlist, evsel) {
237                 if (perf_evsel__alloc_stat_priv(evsel) < 0 ||
238                     perf_evsel__alloc_counts(evsel, perf_evsel__nr_cpus(evsel)) < 0 ||
239                     (alloc_raw && perf_evsel__alloc_prev_raw_counts(evsel) < 0))
240                         goto out_free;
241         }
242
243         return 0;
244
245 out_free:
246         perf_evlist__free_stats(evlist);
247         return -1;
248 }
249
250 static struct stats runtime_nsecs_stats[MAX_NR_CPUS];
251 static struct stats runtime_cycles_stats[MAX_NR_CPUS];
252 static struct stats runtime_stalled_cycles_front_stats[MAX_NR_CPUS];
253 static struct stats runtime_stalled_cycles_back_stats[MAX_NR_CPUS];
254 static struct stats runtime_branches_stats[MAX_NR_CPUS];
255 static struct stats runtime_cacherefs_stats[MAX_NR_CPUS];
256 static struct stats runtime_l1_dcache_stats[MAX_NR_CPUS];
257 static struct stats runtime_l1_icache_stats[MAX_NR_CPUS];
258 static struct stats runtime_ll_cache_stats[MAX_NR_CPUS];
259 static struct stats runtime_itlb_cache_stats[MAX_NR_CPUS];
260 static struct stats runtime_dtlb_cache_stats[MAX_NR_CPUS];
261 static struct stats runtime_cycles_in_tx_stats[MAX_NR_CPUS];
262 static struct stats walltime_nsecs_stats;
263 static struct stats runtime_transaction_stats[MAX_NR_CPUS];
264 static struct stats runtime_elision_stats[MAX_NR_CPUS];
265
266 static void perf_stat__reset_stats(struct perf_evlist *evlist)
267 {
268         struct perf_evsel *evsel;
269
270         evlist__for_each(evlist, evsel) {
271                 perf_evsel__reset_stat_priv(evsel);
272                 perf_evsel__reset_counts(evsel, perf_evsel__nr_cpus(evsel));
273         }
274
275         memset(runtime_nsecs_stats, 0, sizeof(runtime_nsecs_stats));
276         memset(runtime_cycles_stats, 0, sizeof(runtime_cycles_stats));
277         memset(runtime_stalled_cycles_front_stats, 0, sizeof(runtime_stalled_cycles_front_stats));
278         memset(runtime_stalled_cycles_back_stats, 0, sizeof(runtime_stalled_cycles_back_stats));
279         memset(runtime_branches_stats, 0, sizeof(runtime_branches_stats));
280         memset(runtime_cacherefs_stats, 0, sizeof(runtime_cacherefs_stats));
281         memset(runtime_l1_dcache_stats, 0, sizeof(runtime_l1_dcache_stats));
282         memset(runtime_l1_icache_stats, 0, sizeof(runtime_l1_icache_stats));
283         memset(runtime_ll_cache_stats, 0, sizeof(runtime_ll_cache_stats));
284         memset(runtime_itlb_cache_stats, 0, sizeof(runtime_itlb_cache_stats));
285         memset(runtime_dtlb_cache_stats, 0, sizeof(runtime_dtlb_cache_stats));
286         memset(runtime_cycles_in_tx_stats, 0,
287                         sizeof(runtime_cycles_in_tx_stats));
288         memset(runtime_transaction_stats, 0,
289                 sizeof(runtime_transaction_stats));
290         memset(runtime_elision_stats, 0, sizeof(runtime_elision_stats));
291         memset(&walltime_nsecs_stats, 0, sizeof(walltime_nsecs_stats));
292 }
293
294 static int create_perf_stat_counter(struct perf_evsel *evsel)
295 {
296         struct perf_event_attr *attr = &evsel->attr;
297
298         if (scale)
299                 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
300                                     PERF_FORMAT_TOTAL_TIME_RUNNING;
301
302         attr->inherit = !no_inherit;
303
304         if (target__has_cpu(&target))
305                 return perf_evsel__open_per_cpu(evsel, perf_evsel__cpus(evsel));
306
307         if (!target__has_task(&target) && perf_evsel__is_group_leader(evsel)) {
308                 attr->disabled = 1;
309                 if (!initial_delay)
310                         attr->enable_on_exec = 1;
311         }
312
313         return perf_evsel__open_per_thread(evsel, evsel_list->threads);
314 }
315
316 /*
317  * Does the counter have nsecs as a unit?
318  */
319 static inline int nsec_counter(struct perf_evsel *evsel)
320 {
321         if (perf_evsel__match(evsel, SOFTWARE, SW_CPU_CLOCK) ||
322             perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
323                 return 1;
324
325         return 0;
326 }
327
328 static struct perf_evsel *nth_evsel(int n)
329 {
330         static struct perf_evsel **array;
331         static int array_len;
332         struct perf_evsel *ev;
333         int j;
334
335         /* Assumes this only called when evsel_list does not change anymore. */
336         if (!array) {
337                 evlist__for_each(evsel_list, ev)
338                         array_len++;
339                 array = malloc(array_len * sizeof(void *));
340                 if (!array)
341                         exit(ENOMEM);
342                 j = 0;
343                 evlist__for_each(evsel_list, ev)
344                         array[j++] = ev;
345         }
346         if (n < array_len)
347                 return array[n];
348         return NULL;
349 }
350
351 /*
352  * Update various tracking values we maintain to print
353  * more semantic information such as miss/hit ratios,
354  * instruction rates, etc:
355  */
356 static void update_shadow_stats(struct perf_evsel *counter, u64 *count,
357                                 int cpu)
358 {
359         if (perf_evsel__match(counter, SOFTWARE, SW_TASK_CLOCK))
360                 update_stats(&runtime_nsecs_stats[cpu], count[0]);
361         else if (perf_evsel__match(counter, HARDWARE, HW_CPU_CYCLES))
362                 update_stats(&runtime_cycles_stats[cpu], count[0]);
363         else if (transaction_run &&
364                  perf_evsel__cmp(counter, nth_evsel(T_CYCLES_IN_TX)))
365                 update_stats(&runtime_cycles_in_tx_stats[cpu], count[0]);
366         else if (transaction_run &&
367                  perf_evsel__cmp(counter, nth_evsel(T_TRANSACTION_START)))
368                 update_stats(&runtime_transaction_stats[cpu], count[0]);
369         else if (transaction_run &&
370                  perf_evsel__cmp(counter, nth_evsel(T_ELISION_START)))
371                 update_stats(&runtime_elision_stats[cpu], count[0]);
372         else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_FRONTEND))
373                 update_stats(&runtime_stalled_cycles_front_stats[cpu], count[0]);
374         else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_BACKEND))
375                 update_stats(&runtime_stalled_cycles_back_stats[cpu], count[0]);
376         else if (perf_evsel__match(counter, HARDWARE, HW_BRANCH_INSTRUCTIONS))
377                 update_stats(&runtime_branches_stats[cpu], count[0]);
378         else if (perf_evsel__match(counter, HARDWARE, HW_CACHE_REFERENCES))
379                 update_stats(&runtime_cacherefs_stats[cpu], count[0]);
380         else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1D))
381                 update_stats(&runtime_l1_dcache_stats[cpu], count[0]);
382         else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1I))
383                 update_stats(&runtime_l1_icache_stats[cpu], count[0]);
384         else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_LL))
385                 update_stats(&runtime_ll_cache_stats[cpu], count[0]);
386         else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_DTLB))
387                 update_stats(&runtime_dtlb_cache_stats[cpu], count[0]);
388         else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_ITLB))
389                 update_stats(&runtime_itlb_cache_stats[cpu], count[0]);
390 }
391
392 static void zero_per_pkg(struct perf_evsel *counter)
393 {
394         if (counter->per_pkg_mask)
395                 memset(counter->per_pkg_mask, 0, MAX_NR_CPUS);
396 }
397
398 static int check_per_pkg(struct perf_evsel *counter, int cpu, bool *skip)
399 {
400         unsigned long *mask = counter->per_pkg_mask;
401         struct cpu_map *cpus = perf_evsel__cpus(counter);
402         int s;
403
404         *skip = false;
405
406         if (!counter->per_pkg)
407                 return 0;
408
409         if (cpu_map__empty(cpus))
410                 return 0;
411
412         if (!mask) {
413                 mask = zalloc(MAX_NR_CPUS);
414                 if (!mask)
415                         return -ENOMEM;
416
417                 counter->per_pkg_mask = mask;
418         }
419
420         s = cpu_map__get_socket(cpus, cpu);
421         if (s < 0)
422                 return -1;
423
424         *skip = test_and_set_bit(s, mask) == 1;
425         return 0;
426 }
427
428 static int read_cb(struct perf_evsel *evsel, int cpu, int thread __maybe_unused,
429                    struct perf_counts_values *count)
430 {
431         struct perf_counts_values *aggr = &evsel->counts->aggr;
432         static struct perf_counts_values zero;
433         bool skip = false;
434
435         if (check_per_pkg(evsel, cpu, &skip)) {
436                 pr_err("failed to read per-pkg counter\n");
437                 return -1;
438         }
439
440         if (skip)
441                 count = &zero;
442
443         switch (aggr_mode) {
444         case AGGR_CORE:
445         case AGGR_SOCKET:
446         case AGGR_NONE:
447                 if (!evsel->snapshot)
448                         perf_evsel__compute_deltas(evsel, cpu, count);
449                 perf_counts_values__scale(count, scale, NULL);
450                 evsel->counts->cpu[cpu] = *count;
451                 if (aggr_mode == AGGR_NONE)
452                         update_shadow_stats(evsel, count->values, cpu);
453                 break;
454         case AGGR_GLOBAL:
455                 aggr->val += count->val;
456                 if (scale) {
457                         aggr->ena += count->ena;
458                         aggr->run += count->run;
459                 }
460         default:
461                 break;
462         }
463
464         return 0;
465 }
466
467 static int read_counter(struct perf_evsel *counter);
468
469 /*
470  * Read out the results of a single counter:
471  * aggregate counts across CPUs in system-wide mode
472  */
473 static int read_counter_aggr(struct perf_evsel *counter)
474 {
475         struct perf_counts_values *aggr = &counter->counts->aggr;
476         struct perf_stat *ps = counter->priv;
477         u64 *count = counter->counts->aggr.values;
478         int i;
479
480         aggr->val = aggr->ena = aggr->run = 0;
481
482         if (read_counter(counter))
483                 return -1;
484
485         if (!counter->snapshot)
486                 perf_evsel__compute_deltas(counter, -1, aggr);
487         perf_counts_values__scale(aggr, scale, &counter->counts->scaled);
488
489         for (i = 0; i < 3; i++)
490                 update_stats(&ps->res_stats[i], count[i]);
491
492         if (verbose) {
493                 fprintf(output, "%s: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
494                         perf_evsel__name(counter), count[0], count[1], count[2]);
495         }
496
497         /*
498          * Save the full runtime - to allow normalization during printout:
499          */
500         update_shadow_stats(counter, count, 0);
501
502         return 0;
503 }
504
505 /*
506  * Read out the results of a single counter:
507  * do not aggregate counts across CPUs in system-wide mode
508  */
509 static int read_counter(struct perf_evsel *counter)
510 {
511         int nthreads = thread_map__nr(evsel_list->threads);
512         int ncpus = perf_evsel__nr_cpus(counter);
513         int cpu, thread;
514
515         if (!counter->supported)
516                 return -ENOENT;
517
518         if (counter->system_wide)
519                 nthreads = 1;
520
521         if (counter->per_pkg)
522                 zero_per_pkg(counter);
523
524         for (thread = 0; thread < nthreads; thread++) {
525                 for (cpu = 0; cpu < ncpus; cpu++) {
526                         if (perf_evsel__read_cb(counter, cpu, thread, read_cb))
527                                 return -1;
528                 }
529         }
530
531         return 0;
532 }
533
534 static void print_interval(void)
535 {
536         static int num_print_interval;
537         struct perf_evsel *counter;
538         struct perf_stat *ps;
539         struct timespec ts, rs;
540         char prefix[64];
541
542         if (aggr_mode == AGGR_GLOBAL) {
543                 evlist__for_each(evsel_list, counter) {
544                         ps = counter->priv;
545                         memset(ps->res_stats, 0, sizeof(ps->res_stats));
546                         read_counter_aggr(counter);
547                 }
548         } else  {
549                 evlist__for_each(evsel_list, counter) {
550                         ps = counter->priv;
551                         memset(ps->res_stats, 0, sizeof(ps->res_stats));
552                         read_counter(counter);
553                 }
554         }
555
556         clock_gettime(CLOCK_MONOTONIC, &ts);
557         diff_timespec(&rs, &ts, &ref_time);
558         sprintf(prefix, "%6lu.%09lu%s", rs.tv_sec, rs.tv_nsec, csv_sep);
559
560         if (num_print_interval == 0 && !csv_output) {
561                 switch (aggr_mode) {
562                 case AGGR_SOCKET:
563                         fprintf(output, "#           time socket cpus             counts %*s events\n", unit_width, "unit");
564                         break;
565                 case AGGR_CORE:
566                         fprintf(output, "#           time core         cpus             counts %*s events\n", unit_width, "unit");
567                         break;
568                 case AGGR_NONE:
569                         fprintf(output, "#           time CPU                counts %*s events\n", unit_width, "unit");
570                         break;
571                 case AGGR_GLOBAL:
572                 default:
573                         fprintf(output, "#           time             counts %*s events\n", unit_width, "unit");
574                 }
575         }
576
577         if (++num_print_interval == 25)
578                 num_print_interval = 0;
579
580         switch (aggr_mode) {
581         case AGGR_CORE:
582         case AGGR_SOCKET:
583                 print_aggr(prefix);
584                 break;
585         case AGGR_NONE:
586                 evlist__for_each(evsel_list, counter)
587                         print_counter(counter, prefix);
588                 break;
589         case AGGR_GLOBAL:
590         default:
591                 evlist__for_each(evsel_list, counter)
592                         print_counter_aggr(counter, prefix);
593         }
594
595         fflush(output);
596 }
597
598 static void handle_initial_delay(void)
599 {
600         struct perf_evsel *counter;
601
602         if (initial_delay) {
603                 const int ncpus = cpu_map__nr(evsel_list->cpus),
604                         nthreads = thread_map__nr(evsel_list->threads);
605
606                 usleep(initial_delay * 1000);
607                 evlist__for_each(evsel_list, counter)
608                         perf_evsel__enable(counter, ncpus, nthreads);
609         }
610 }
611
612 static volatile int workload_exec_errno;
613
614 /*
615  * perf_evlist__prepare_workload will send a SIGUSR1
616  * if the fork fails, since we asked by setting its
617  * want_signal to true.
618  */
619 static void workload_exec_failed_signal(int signo __maybe_unused, siginfo_t *info,
620                                         void *ucontext __maybe_unused)
621 {
622         workload_exec_errno = info->si_value.sival_int;
623 }
624
625 static int __run_perf_stat(int argc, const char **argv)
626 {
627         char msg[512];
628         unsigned long long t0, t1;
629         struct perf_evsel *counter;
630         struct timespec ts;
631         size_t l;
632         int status = 0;
633         const bool forks = (argc > 0);
634
635         if (interval) {
636                 ts.tv_sec  = interval / 1000;
637                 ts.tv_nsec = (interval % 1000) * 1000000;
638         } else {
639                 ts.tv_sec  = 1;
640                 ts.tv_nsec = 0;
641         }
642
643         if (forks) {
644                 if (perf_evlist__prepare_workload(evsel_list, &target, argv, false,
645                                                   workload_exec_failed_signal) < 0) {
646                         perror("failed to prepare workload");
647                         return -1;
648                 }
649                 child_pid = evsel_list->workload.pid;
650         }
651
652         if (group)
653                 perf_evlist__set_leader(evsel_list);
654
655         evlist__for_each(evsel_list, counter) {
656                 if (create_perf_stat_counter(counter) < 0) {
657                         /*
658                          * PPC returns ENXIO for HW counters until 2.6.37
659                          * (behavior changed with commit b0a873e).
660                          */
661                         if (errno == EINVAL || errno == ENOSYS ||
662                             errno == ENOENT || errno == EOPNOTSUPP ||
663                             errno == ENXIO) {
664                                 if (verbose)
665                                         ui__warning("%s event is not supported by the kernel.\n",
666                                                     perf_evsel__name(counter));
667                                 counter->supported = false;
668                                 continue;
669                         }
670
671                         perf_evsel__open_strerror(counter, &target,
672                                                   errno, msg, sizeof(msg));
673                         ui__error("%s\n", msg);
674
675                         if (child_pid != -1)
676                                 kill(child_pid, SIGTERM);
677
678                         return -1;
679                 }
680                 counter->supported = true;
681
682                 l = strlen(counter->unit);
683                 if (l > unit_width)
684                         unit_width = l;
685         }
686
687         if (perf_evlist__apply_filters(evsel_list)) {
688                 error("failed to set filter with %d (%s)\n", errno,
689                         strerror_r(errno, msg, sizeof(msg)));
690                 return -1;
691         }
692
693         /*
694          * Enable counters and exec the command:
695          */
696         t0 = rdclock();
697         clock_gettime(CLOCK_MONOTONIC, &ref_time);
698
699         if (forks) {
700                 perf_evlist__start_workload(evsel_list);
701                 handle_initial_delay();
702
703                 if (interval) {
704                         while (!waitpid(child_pid, &status, WNOHANG)) {
705                                 nanosleep(&ts, NULL);
706                                 print_interval();
707                         }
708                 }
709                 wait(&status);
710
711                 if (workload_exec_errno) {
712                         const char *emsg = strerror_r(workload_exec_errno, msg, sizeof(msg));
713                         pr_err("Workload failed: %s\n", emsg);
714                         return -1;
715                 }
716
717                 if (WIFSIGNALED(status))
718                         psignal(WTERMSIG(status), argv[0]);
719         } else {
720                 handle_initial_delay();
721                 while (!done) {
722                         nanosleep(&ts, NULL);
723                         if (interval)
724                                 print_interval();
725                 }
726         }
727
728         t1 = rdclock();
729
730         update_stats(&walltime_nsecs_stats, t1 - t0);
731
732         if (aggr_mode == AGGR_GLOBAL) {
733                 evlist__for_each(evsel_list, counter) {
734                         read_counter_aggr(counter);
735                         perf_evsel__close_fd(counter, perf_evsel__nr_cpus(counter),
736                                              thread_map__nr(evsel_list->threads));
737                 }
738         } else {
739                 evlist__for_each(evsel_list, counter) {
740                         read_counter(counter);
741                         perf_evsel__close_fd(counter, perf_evsel__nr_cpus(counter), 1);
742                 }
743         }
744
745         return WEXITSTATUS(status);
746 }
747
748 static int run_perf_stat(int argc, const char **argv)
749 {
750         int ret;
751
752         if (pre_cmd) {
753                 ret = system(pre_cmd);
754                 if (ret)
755                         return ret;
756         }
757
758         if (sync_run)
759                 sync();
760
761         ret = __run_perf_stat(argc, argv);
762         if (ret)
763                 return ret;
764
765         if (post_cmd) {
766                 ret = system(post_cmd);
767                 if (ret)
768                         return ret;
769         }
770
771         return ret;
772 }
773
774 static void print_running(u64 run, u64 ena)
775 {
776         if (csv_output) {
777                 fprintf(output, "%s%" PRIu64 "%s%.2f",
778                                         csv_sep,
779                                         run,
780                                         csv_sep,
781                                         ena ? 100.0 * run / ena : 100.0);
782         } else if (run != ena) {
783                 fprintf(output, "  (%.2f%%)", 100.0 * run / ena);
784         }
785 }
786
787 static void print_noise_pct(double total, double avg)
788 {
789         double pct = rel_stddev_stats(total, avg);
790
791         if (csv_output)
792                 fprintf(output, "%s%.2f%%", csv_sep, pct);
793         else if (pct)
794                 fprintf(output, "  ( +-%6.2f%% )", pct);
795 }
796
797 static void print_noise(struct perf_evsel *evsel, double avg)
798 {
799         struct perf_stat *ps;
800
801         if (run_count == 1)
802                 return;
803
804         ps = evsel->priv;
805         print_noise_pct(stddev_stats(&ps->res_stats[0]), avg);
806 }
807
808 static void aggr_printout(struct perf_evsel *evsel, int id, int nr)
809 {
810         switch (aggr_mode) {
811         case AGGR_CORE:
812                 fprintf(output, "S%d-C%*d%s%*d%s",
813                         cpu_map__id_to_socket(id),
814                         csv_output ? 0 : -8,
815                         cpu_map__id_to_cpu(id),
816                         csv_sep,
817                         csv_output ? 0 : 4,
818                         nr,
819                         csv_sep);
820                 break;
821         case AGGR_SOCKET:
822                 fprintf(output, "S%*d%s%*d%s",
823                         csv_output ? 0 : -5,
824                         id,
825                         csv_sep,
826                         csv_output ? 0 : 4,
827                         nr,
828                         csv_sep);
829                         break;
830         case AGGR_NONE:
831                 fprintf(output, "CPU%*d%s",
832                         csv_output ? 0 : -4,
833                         perf_evsel__cpus(evsel)->map[id], csv_sep);
834                 break;
835         case AGGR_GLOBAL:
836         default:
837                 break;
838         }
839 }
840
841 static void nsec_printout(int id, int nr, struct perf_evsel *evsel, double avg)
842 {
843         double msecs = avg / 1e6;
844         const char *fmt_v, *fmt_n;
845         char name[25];
846
847         fmt_v = csv_output ? "%.6f%s" : "%18.6f%s";
848         fmt_n = csv_output ? "%s" : "%-25s";
849
850         aggr_printout(evsel, id, nr);
851
852         scnprintf(name, sizeof(name), "%s%s",
853                   perf_evsel__name(evsel), csv_output ? "" : " (msec)");
854
855         fprintf(output, fmt_v, msecs, csv_sep);
856
857         if (csv_output)
858                 fprintf(output, "%s%s", evsel->unit, csv_sep);
859         else
860                 fprintf(output, "%-*s%s", unit_width, evsel->unit, csv_sep);
861
862         fprintf(output, fmt_n, name);
863
864         if (evsel->cgrp)
865                 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
866
867         if (csv_output || interval)
868                 return;
869
870         if (perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
871                 fprintf(output, " # %8.3f CPUs utilized          ",
872                         avg / avg_stats(&walltime_nsecs_stats));
873         else
874                 fprintf(output, "                                   ");
875 }
876
877 /* used for get_ratio_color() */
878 enum grc_type {
879         GRC_STALLED_CYCLES_FE,
880         GRC_STALLED_CYCLES_BE,
881         GRC_CACHE_MISSES,
882         GRC_MAX_NR
883 };
884
885 static const char *get_ratio_color(enum grc_type type, double ratio)
886 {
887         static const double grc_table[GRC_MAX_NR][3] = {
888                 [GRC_STALLED_CYCLES_FE] = { 50.0, 30.0, 10.0 },
889                 [GRC_STALLED_CYCLES_BE] = { 75.0, 50.0, 20.0 },
890                 [GRC_CACHE_MISSES]      = { 20.0, 10.0, 5.0 },
891         };
892         const char *color = PERF_COLOR_NORMAL;
893
894         if (ratio > grc_table[type][0])
895                 color = PERF_COLOR_RED;
896         else if (ratio > grc_table[type][1])
897                 color = PERF_COLOR_MAGENTA;
898         else if (ratio > grc_table[type][2])
899                 color = PERF_COLOR_YELLOW;
900
901         return color;
902 }
903
904 static void print_stalled_cycles_frontend(int cpu,
905                                           struct perf_evsel *evsel
906                                           __maybe_unused, double avg)
907 {
908         double total, ratio = 0.0;
909         const char *color;
910
911         total = avg_stats(&runtime_cycles_stats[cpu]);
912
913         if (total)
914                 ratio = avg / total * 100.0;
915
916         color = get_ratio_color(GRC_STALLED_CYCLES_FE, ratio);
917
918         fprintf(output, " #  ");
919         color_fprintf(output, color, "%6.2f%%", ratio);
920         fprintf(output, " frontend cycles idle   ");
921 }
922
923 static void print_stalled_cycles_backend(int cpu,
924                                          struct perf_evsel *evsel
925                                          __maybe_unused, double avg)
926 {
927         double total, ratio = 0.0;
928         const char *color;
929
930         total = avg_stats(&runtime_cycles_stats[cpu]);
931
932         if (total)
933                 ratio = avg / total * 100.0;
934
935         color = get_ratio_color(GRC_STALLED_CYCLES_BE, ratio);
936
937         fprintf(output, " #  ");
938         color_fprintf(output, color, "%6.2f%%", ratio);
939         fprintf(output, " backend  cycles idle   ");
940 }
941
942 static void print_branch_misses(int cpu,
943                                 struct perf_evsel *evsel __maybe_unused,
944                                 double avg)
945 {
946         double total, ratio = 0.0;
947         const char *color;
948
949         total = avg_stats(&runtime_branches_stats[cpu]);
950
951         if (total)
952                 ratio = avg / total * 100.0;
953
954         color = get_ratio_color(GRC_CACHE_MISSES, ratio);
955
956         fprintf(output, " #  ");
957         color_fprintf(output, color, "%6.2f%%", ratio);
958         fprintf(output, " of all branches        ");
959 }
960
961 static void print_l1_dcache_misses(int cpu,
962                                    struct perf_evsel *evsel __maybe_unused,
963                                    double avg)
964 {
965         double total, ratio = 0.0;
966         const char *color;
967
968         total = avg_stats(&runtime_l1_dcache_stats[cpu]);
969
970         if (total)
971                 ratio = avg / total * 100.0;
972
973         color = get_ratio_color(GRC_CACHE_MISSES, ratio);
974
975         fprintf(output, " #  ");
976         color_fprintf(output, color, "%6.2f%%", ratio);
977         fprintf(output, " of all L1-dcache hits  ");
978 }
979
980 static void print_l1_icache_misses(int cpu,
981                                    struct perf_evsel *evsel __maybe_unused,
982                                    double avg)
983 {
984         double total, ratio = 0.0;
985         const char *color;
986
987         total = avg_stats(&runtime_l1_icache_stats[cpu]);
988
989         if (total)
990                 ratio = avg / total * 100.0;
991
992         color = get_ratio_color(GRC_CACHE_MISSES, ratio);
993
994         fprintf(output, " #  ");
995         color_fprintf(output, color, "%6.2f%%", ratio);
996         fprintf(output, " of all L1-icache hits  ");
997 }
998
999 static void print_dtlb_cache_misses(int cpu,
1000                                     struct perf_evsel *evsel __maybe_unused,
1001                                     double avg)
1002 {
1003         double total, ratio = 0.0;
1004         const char *color;
1005
1006         total = avg_stats(&runtime_dtlb_cache_stats[cpu]);
1007
1008         if (total)
1009                 ratio = avg / total * 100.0;
1010
1011         color = get_ratio_color(GRC_CACHE_MISSES, ratio);
1012
1013         fprintf(output, " #  ");
1014         color_fprintf(output, color, "%6.2f%%", ratio);
1015         fprintf(output, " of all dTLB cache hits ");
1016 }
1017
1018 static void print_itlb_cache_misses(int cpu,
1019                                     struct perf_evsel *evsel __maybe_unused,
1020                                     double avg)
1021 {
1022         double total, ratio = 0.0;
1023         const char *color;
1024
1025         total = avg_stats(&runtime_itlb_cache_stats[cpu]);
1026
1027         if (total)
1028                 ratio = avg / total * 100.0;
1029
1030         color = get_ratio_color(GRC_CACHE_MISSES, ratio);
1031
1032         fprintf(output, " #  ");
1033         color_fprintf(output, color, "%6.2f%%", ratio);
1034         fprintf(output, " of all iTLB cache hits ");
1035 }
1036
1037 static void print_ll_cache_misses(int cpu,
1038                                   struct perf_evsel *evsel __maybe_unused,
1039                                   double avg)
1040 {
1041         double total, ratio = 0.0;
1042         const char *color;
1043
1044         total = avg_stats(&runtime_ll_cache_stats[cpu]);
1045
1046         if (total)
1047                 ratio = avg / total * 100.0;
1048
1049         color = get_ratio_color(GRC_CACHE_MISSES, ratio);
1050
1051         fprintf(output, " #  ");
1052         color_fprintf(output, color, "%6.2f%%", ratio);
1053         fprintf(output, " of all LL-cache hits   ");
1054 }
1055
1056 static void abs_printout(int id, int nr, struct perf_evsel *evsel, double avg)
1057 {
1058         double total, ratio = 0.0, total2;
1059         double sc =  evsel->scale;
1060         const char *fmt;
1061         int cpu = cpu_map__id_to_cpu(id);
1062
1063         if (csv_output) {
1064                 fmt = sc != 1.0 ?  "%.2f%s" : "%.0f%s";
1065         } else {
1066                 if (big_num)
1067                         fmt = sc != 1.0 ? "%'18.2f%s" : "%'18.0f%s";
1068                 else
1069                         fmt = sc != 1.0 ? "%18.2f%s" : "%18.0f%s";
1070         }
1071
1072         aggr_printout(evsel, id, nr);
1073
1074         if (aggr_mode == AGGR_GLOBAL)
1075                 cpu = 0;
1076
1077         fprintf(output, fmt, avg, csv_sep);
1078
1079         if (evsel->unit)
1080                 fprintf(output, "%-*s%s",
1081                         csv_output ? 0 : unit_width,
1082                         evsel->unit, csv_sep);
1083
1084         fprintf(output, "%-*s", csv_output ? 0 : 25, perf_evsel__name(evsel));
1085
1086         if (evsel->cgrp)
1087                 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
1088
1089         if (csv_output || interval)
1090                 return;
1091
1092         if (perf_evsel__match(evsel, HARDWARE, HW_INSTRUCTIONS)) {
1093                 total = avg_stats(&runtime_cycles_stats[cpu]);
1094                 if (total) {
1095                         ratio = avg / total;
1096                         fprintf(output, " #   %5.2f  insns per cycle        ", ratio);
1097                 }
1098                 total = avg_stats(&runtime_stalled_cycles_front_stats[cpu]);
1099                 total = max(total, avg_stats(&runtime_stalled_cycles_back_stats[cpu]));
1100
1101                 if (total && avg) {
1102                         ratio = total / avg;
1103                         fprintf(output, "\n");
1104                         if (aggr_mode == AGGR_NONE)
1105                                 fprintf(output, "        ");
1106                         fprintf(output, "                                                  #   %5.2f  stalled cycles per insn", ratio);
1107                 }
1108
1109         } else if (perf_evsel__match(evsel, HARDWARE, HW_BRANCH_MISSES) &&
1110                         runtime_branches_stats[cpu].n != 0) {
1111                 print_branch_misses(cpu, evsel, avg);
1112         } else if (
1113                 evsel->attr.type == PERF_TYPE_HW_CACHE &&
1114                 evsel->attr.config ==  ( PERF_COUNT_HW_CACHE_L1D |
1115                                         ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
1116                                         ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
1117                         runtime_l1_dcache_stats[cpu].n != 0) {
1118                 print_l1_dcache_misses(cpu, evsel, avg);
1119         } else if (
1120                 evsel->attr.type == PERF_TYPE_HW_CACHE &&
1121                 evsel->attr.config ==  ( PERF_COUNT_HW_CACHE_L1I |
1122                                         ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
1123                                         ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
1124                         runtime_l1_icache_stats[cpu].n != 0) {
1125                 print_l1_icache_misses(cpu, evsel, avg);
1126         } else if (
1127                 evsel->attr.type == PERF_TYPE_HW_CACHE &&
1128                 evsel->attr.config ==  ( PERF_COUNT_HW_CACHE_DTLB |
1129                                         ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
1130                                         ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
1131                         runtime_dtlb_cache_stats[cpu].n != 0) {
1132                 print_dtlb_cache_misses(cpu, evsel, avg);
1133         } else if (
1134                 evsel->attr.type == PERF_TYPE_HW_CACHE &&
1135                 evsel->attr.config ==  ( PERF_COUNT_HW_CACHE_ITLB |
1136                                         ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
1137                                         ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
1138                         runtime_itlb_cache_stats[cpu].n != 0) {
1139                 print_itlb_cache_misses(cpu, evsel, avg);
1140         } else if (
1141                 evsel->attr.type == PERF_TYPE_HW_CACHE &&
1142                 evsel->attr.config ==  ( PERF_COUNT_HW_CACHE_LL |
1143                                         ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
1144                                         ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
1145                         runtime_ll_cache_stats[cpu].n != 0) {
1146                 print_ll_cache_misses(cpu, evsel, avg);
1147         } else if (perf_evsel__match(evsel, HARDWARE, HW_CACHE_MISSES) &&
1148                         runtime_cacherefs_stats[cpu].n != 0) {
1149                 total = avg_stats(&runtime_cacherefs_stats[cpu]);
1150
1151                 if (total)
1152                         ratio = avg * 100 / total;
1153
1154                 fprintf(output, " # %8.3f %% of all cache refs    ", ratio);
1155
1156         } else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_FRONTEND)) {
1157                 print_stalled_cycles_frontend(cpu, evsel, avg);
1158         } else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_BACKEND)) {
1159                 print_stalled_cycles_backend(cpu, evsel, avg);
1160         } else if (perf_evsel__match(evsel, HARDWARE, HW_CPU_CYCLES)) {
1161                 total = avg_stats(&runtime_nsecs_stats[cpu]);
1162
1163                 if (total) {
1164                         ratio = avg / total;
1165                         fprintf(output, " # %8.3f GHz                    ", ratio);
1166                 }
1167         } else if (transaction_run &&
1168                    perf_evsel__cmp(evsel, nth_evsel(T_CYCLES_IN_TX))) {
1169                 total = avg_stats(&runtime_cycles_stats[cpu]);
1170                 if (total)
1171                         fprintf(output,
1172                                 " #   %5.2f%% transactional cycles   ",
1173                                 100.0 * (avg / total));
1174         } else if (transaction_run &&
1175                    perf_evsel__cmp(evsel, nth_evsel(T_CYCLES_IN_TX_CP))) {
1176                 total = avg_stats(&runtime_cycles_stats[cpu]);
1177                 total2 = avg_stats(&runtime_cycles_in_tx_stats[cpu]);
1178                 if (total2 < avg)
1179                         total2 = avg;
1180                 if (total)
1181                         fprintf(output,
1182                                 " #   %5.2f%% aborted cycles         ",
1183                                 100.0 * ((total2-avg) / total));
1184         } else if (transaction_run &&
1185                    perf_evsel__cmp(evsel, nth_evsel(T_TRANSACTION_START)) &&
1186                    avg > 0 &&
1187                    runtime_cycles_in_tx_stats[cpu].n != 0) {
1188                 total = avg_stats(&runtime_cycles_in_tx_stats[cpu]);
1189
1190                 if (total)
1191                         ratio = total / avg;
1192
1193                 fprintf(output, " # %8.0f cycles / transaction   ", ratio);
1194         } else if (transaction_run &&
1195                    perf_evsel__cmp(evsel, nth_evsel(T_ELISION_START)) &&
1196                    avg > 0 &&
1197                    runtime_cycles_in_tx_stats[cpu].n != 0) {
1198                 total = avg_stats(&runtime_cycles_in_tx_stats[cpu]);
1199
1200                 if (total)
1201                         ratio = total / avg;
1202
1203                 fprintf(output, " # %8.0f cycles / elision       ", ratio);
1204         } else if (runtime_nsecs_stats[cpu].n != 0) {
1205                 char unit = 'M';
1206
1207                 total = avg_stats(&runtime_nsecs_stats[cpu]);
1208
1209                 if (total)
1210                         ratio = 1000.0 * avg / total;
1211                 if (ratio < 0.001) {
1212                         ratio *= 1000;
1213                         unit = 'K';
1214                 }
1215
1216                 fprintf(output, " # %8.3f %c/sec                  ", ratio, unit);
1217         } else {
1218                 fprintf(output, "                                   ");
1219         }
1220 }
1221
1222 static void print_aggr(char *prefix)
1223 {
1224         struct perf_evsel *counter;
1225         int cpu, cpu2, s, s2, id, nr;
1226         double uval;
1227         u64 ena, run, val;
1228
1229         if (!(aggr_map || aggr_get_id))
1230                 return;
1231
1232         for (s = 0; s < aggr_map->nr; s++) {
1233                 id = aggr_map->map[s];
1234                 evlist__for_each(evsel_list, counter) {
1235                         val = ena = run = 0;
1236                         nr = 0;
1237                         for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
1238                                 cpu2 = perf_evsel__cpus(counter)->map[cpu];
1239                                 s2 = aggr_get_id(evsel_list->cpus, cpu2);
1240                                 if (s2 != id)
1241                                         continue;
1242                                 val += counter->counts->cpu[cpu].val;
1243                                 ena += counter->counts->cpu[cpu].ena;
1244                                 run += counter->counts->cpu[cpu].run;
1245                                 nr++;
1246                         }
1247                         if (prefix)
1248                                 fprintf(output, "%s", prefix);
1249
1250                         if (run == 0 || ena == 0) {
1251                                 aggr_printout(counter, id, nr);
1252
1253                                 fprintf(output, "%*s%s",
1254                                         csv_output ? 0 : 18,
1255                                         counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
1256                                         csv_sep);
1257
1258                                 fprintf(output, "%-*s%s",
1259                                         csv_output ? 0 : unit_width,
1260                                         counter->unit, csv_sep);
1261
1262                                 fprintf(output, "%*s",
1263                                         csv_output ? 0 : -25,
1264                                         perf_evsel__name(counter));
1265
1266                                 if (counter->cgrp)
1267                                         fprintf(output, "%s%s",
1268                                                 csv_sep, counter->cgrp->name);
1269
1270                                 print_running(run, ena);
1271                                 fputc('\n', output);
1272                                 continue;
1273                         }
1274                         uval = val * counter->scale;
1275
1276                         if (nsec_counter(counter))
1277                                 nsec_printout(id, nr, counter, uval);
1278                         else
1279                                 abs_printout(id, nr, counter, uval);
1280
1281                         if (!csv_output)
1282                                 print_noise(counter, 1.0);
1283
1284                         print_running(run, ena);
1285                         fputc('\n', output);
1286                 }
1287         }
1288 }
1289
1290 /*
1291  * Print out the results of a single counter:
1292  * aggregated counts in system-wide mode
1293  */
1294 static void print_counter_aggr(struct perf_evsel *counter, char *prefix)
1295 {
1296         struct perf_stat *ps = counter->priv;
1297         double avg = avg_stats(&ps->res_stats[0]);
1298         int scaled = counter->counts->scaled;
1299         double uval;
1300         double avg_enabled, avg_running;
1301
1302         avg_enabled = avg_stats(&ps->res_stats[1]);
1303         avg_running = avg_stats(&ps->res_stats[2]);
1304
1305         if (prefix)
1306                 fprintf(output, "%s", prefix);
1307
1308         if (scaled == -1 || !counter->supported) {
1309                 fprintf(output, "%*s%s",
1310                         csv_output ? 0 : 18,
1311                         counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
1312                         csv_sep);
1313                 fprintf(output, "%-*s%s",
1314                         csv_output ? 0 : unit_width,
1315                         counter->unit, csv_sep);
1316                 fprintf(output, "%*s",
1317                         csv_output ? 0 : -25,
1318                         perf_evsel__name(counter));
1319
1320                 if (counter->cgrp)
1321                         fprintf(output, "%s%s", csv_sep, counter->cgrp->name);
1322
1323                 print_running(avg_running, avg_enabled);
1324                 fputc('\n', output);
1325                 return;
1326         }
1327
1328         uval = avg * counter->scale;
1329
1330         if (nsec_counter(counter))
1331                 nsec_printout(-1, 0, counter, uval);
1332         else
1333                 abs_printout(-1, 0, counter, uval);
1334
1335         print_noise(counter, avg);
1336
1337         print_running(avg_running, avg_enabled);
1338         fprintf(output, "\n");
1339 }
1340
1341 /*
1342  * Print out the results of a single counter:
1343  * does not use aggregated count in system-wide
1344  */
1345 static void print_counter(struct perf_evsel *counter, char *prefix)
1346 {
1347         u64 ena, run, val;
1348         double uval;
1349         int cpu;
1350
1351         for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
1352                 val = counter->counts->cpu[cpu].val;
1353                 ena = counter->counts->cpu[cpu].ena;
1354                 run = counter->counts->cpu[cpu].run;
1355
1356                 if (prefix)
1357                         fprintf(output, "%s", prefix);
1358
1359                 if (run == 0 || ena == 0) {
1360                         fprintf(output, "CPU%*d%s%*s%s",
1361                                 csv_output ? 0 : -4,
1362                                 perf_evsel__cpus(counter)->map[cpu], csv_sep,
1363                                 csv_output ? 0 : 18,
1364                                 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
1365                                 csv_sep);
1366
1367                                 fprintf(output, "%-*s%s",
1368                                         csv_output ? 0 : unit_width,
1369                                         counter->unit, csv_sep);
1370
1371                                 fprintf(output, "%*s",
1372                                         csv_output ? 0 : -25,
1373                                         perf_evsel__name(counter));
1374
1375                         if (counter->cgrp)
1376                                 fprintf(output, "%s%s",
1377                                         csv_sep, counter->cgrp->name);
1378
1379                         print_running(run, ena);
1380                         fputc('\n', output);
1381                         continue;
1382                 }
1383
1384                 uval = val * counter->scale;
1385
1386                 if (nsec_counter(counter))
1387                         nsec_printout(cpu, 0, counter, uval);
1388                 else
1389                         abs_printout(cpu, 0, counter, uval);
1390
1391                 if (!csv_output)
1392                         print_noise(counter, 1.0);
1393                 print_running(run, ena);
1394
1395                 fputc('\n', output);
1396         }
1397 }
1398
1399 static void print_stat(int argc, const char **argv)
1400 {
1401         struct perf_evsel *counter;
1402         int i;
1403
1404         fflush(stdout);
1405
1406         if (!csv_output) {
1407                 fprintf(output, "\n");
1408                 fprintf(output, " Performance counter stats for ");
1409                 if (target.system_wide)
1410                         fprintf(output, "\'system wide");
1411                 else if (target.cpu_list)
1412                         fprintf(output, "\'CPU(s) %s", target.cpu_list);
1413                 else if (!target__has_task(&target)) {
1414                         fprintf(output, "\'%s", argv[0]);
1415                         for (i = 1; i < argc; i++)
1416                                 fprintf(output, " %s", argv[i]);
1417                 } else if (target.pid)
1418                         fprintf(output, "process id \'%s", target.pid);
1419                 else
1420                         fprintf(output, "thread id \'%s", target.tid);
1421
1422                 fprintf(output, "\'");
1423                 if (run_count > 1)
1424                         fprintf(output, " (%d runs)", run_count);
1425                 fprintf(output, ":\n\n");
1426         }
1427
1428         switch (aggr_mode) {
1429         case AGGR_CORE:
1430         case AGGR_SOCKET:
1431                 print_aggr(NULL);
1432                 break;
1433         case AGGR_GLOBAL:
1434                 evlist__for_each(evsel_list, counter)
1435                         print_counter_aggr(counter, NULL);
1436                 break;
1437         case AGGR_NONE:
1438                 evlist__for_each(evsel_list, counter)
1439                         print_counter(counter, NULL);
1440                 break;
1441         default:
1442                 break;
1443         }
1444
1445         if (!csv_output) {
1446                 if (!null_run)
1447                         fprintf(output, "\n");
1448                 fprintf(output, " %17.9f seconds time elapsed",
1449                                 avg_stats(&walltime_nsecs_stats)/1e9);
1450                 if (run_count > 1) {
1451                         fprintf(output, "                                        ");
1452                         print_noise_pct(stddev_stats(&walltime_nsecs_stats),
1453                                         avg_stats(&walltime_nsecs_stats));
1454                 }
1455                 fprintf(output, "\n\n");
1456         }
1457 }
1458
1459 static volatile int signr = -1;
1460
1461 static void skip_signal(int signo)
1462 {
1463         if ((child_pid == -1) || interval)
1464                 done = 1;
1465
1466         signr = signo;
1467         /*
1468          * render child_pid harmless
1469          * won't send SIGTERM to a random
1470          * process in case of race condition
1471          * and fast PID recycling
1472          */
1473         child_pid = -1;
1474 }
1475
1476 static void sig_atexit(void)
1477 {
1478         sigset_t set, oset;
1479
1480         /*
1481          * avoid race condition with SIGCHLD handler
1482          * in skip_signal() which is modifying child_pid
1483          * goal is to avoid send SIGTERM to a random
1484          * process
1485          */
1486         sigemptyset(&set);
1487         sigaddset(&set, SIGCHLD);
1488         sigprocmask(SIG_BLOCK, &set, &oset);
1489
1490         if (child_pid != -1)
1491                 kill(child_pid, SIGTERM);
1492
1493         sigprocmask(SIG_SETMASK, &oset, NULL);
1494
1495         if (signr == -1)
1496                 return;
1497
1498         signal(signr, SIG_DFL);
1499         kill(getpid(), signr);
1500 }
1501
1502 static int stat__set_big_num(const struct option *opt __maybe_unused,
1503                              const char *s __maybe_unused, int unset)
1504 {
1505         big_num_opt = unset ? 0 : 1;
1506         return 0;
1507 }
1508
1509 static int perf_stat_init_aggr_mode(void)
1510 {
1511         switch (aggr_mode) {
1512         case AGGR_SOCKET:
1513                 if (cpu_map__build_socket_map(evsel_list->cpus, &aggr_map)) {
1514                         perror("cannot build socket map");
1515                         return -1;
1516                 }
1517                 aggr_get_id = cpu_map__get_socket;
1518                 break;
1519         case AGGR_CORE:
1520                 if (cpu_map__build_core_map(evsel_list->cpus, &aggr_map)) {
1521                         perror("cannot build core map");
1522                         return -1;
1523                 }
1524                 aggr_get_id = cpu_map__get_core;
1525                 break;
1526         case AGGR_NONE:
1527         case AGGR_GLOBAL:
1528         default:
1529                 break;
1530         }
1531         return 0;
1532 }
1533
1534 static int setup_events(const char * const *attrs, unsigned len)
1535 {
1536         unsigned i;
1537
1538         for (i = 0; i < len; i++) {
1539                 if (parse_events(evsel_list, attrs[i]))
1540                         return -1;
1541         }
1542         return 0;
1543 }
1544
1545 /*
1546  * Add default attributes, if there were no attributes specified or
1547  * if -d/--detailed, -d -d or -d -d -d is used:
1548  */
1549 static int add_default_attributes(void)
1550 {
1551         struct perf_event_attr default_attrs[] = {
1552
1553   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK              },
1554   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES        },
1555   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS          },
1556   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS             },
1557
1558   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES              },
1559   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND },
1560   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND  },
1561   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS            },
1562   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS     },
1563   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES           },
1564
1565 };
1566
1567 /*
1568  * Detailed stats (-d), covering the L1 and last level data caches:
1569  */
1570         struct perf_event_attr detailed_attrs[] = {
1571
1572   { .type = PERF_TYPE_HW_CACHE,
1573     .config =
1574          PERF_COUNT_HW_CACHE_L1D                <<  0  |
1575         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1576         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
1577
1578   { .type = PERF_TYPE_HW_CACHE,
1579     .config =
1580          PERF_COUNT_HW_CACHE_L1D                <<  0  |
1581         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1582         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
1583
1584   { .type = PERF_TYPE_HW_CACHE,
1585     .config =
1586          PERF_COUNT_HW_CACHE_LL                 <<  0  |
1587         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1588         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
1589
1590   { .type = PERF_TYPE_HW_CACHE,
1591     .config =
1592          PERF_COUNT_HW_CACHE_LL                 <<  0  |
1593         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1594         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
1595 };
1596
1597 /*
1598  * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
1599  */
1600         struct perf_event_attr very_detailed_attrs[] = {
1601
1602   { .type = PERF_TYPE_HW_CACHE,
1603     .config =
1604          PERF_COUNT_HW_CACHE_L1I                <<  0  |
1605         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1606         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
1607
1608   { .type = PERF_TYPE_HW_CACHE,
1609     .config =
1610          PERF_COUNT_HW_CACHE_L1I                <<  0  |
1611         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1612         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
1613
1614   { .type = PERF_TYPE_HW_CACHE,
1615     .config =
1616          PERF_COUNT_HW_CACHE_DTLB               <<  0  |
1617         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1618         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
1619
1620   { .type = PERF_TYPE_HW_CACHE,
1621     .config =
1622          PERF_COUNT_HW_CACHE_DTLB               <<  0  |
1623         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1624         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
1625
1626   { .type = PERF_TYPE_HW_CACHE,
1627     .config =
1628          PERF_COUNT_HW_CACHE_ITLB               <<  0  |
1629         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1630         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
1631
1632   { .type = PERF_TYPE_HW_CACHE,
1633     .config =
1634          PERF_COUNT_HW_CACHE_ITLB               <<  0  |
1635         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
1636         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
1637
1638 };
1639
1640 /*
1641  * Very, very detailed stats (-d -d -d), adding prefetch events:
1642  */
1643         struct perf_event_attr very_very_detailed_attrs[] = {
1644
1645   { .type = PERF_TYPE_HW_CACHE,
1646     .config =
1647          PERF_COUNT_HW_CACHE_L1D                <<  0  |
1648         (PERF_COUNT_HW_CACHE_OP_PREFETCH        <<  8) |
1649         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
1650
1651   { .type = PERF_TYPE_HW_CACHE,
1652     .config =
1653          PERF_COUNT_HW_CACHE_L1D                <<  0  |
1654         (PERF_COUNT_HW_CACHE_OP_PREFETCH        <<  8) |
1655         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
1656 };
1657
1658         /* Set attrs if no event is selected and !null_run: */
1659         if (null_run)
1660                 return 0;
1661
1662         if (transaction_run) {
1663                 int err;
1664                 if (pmu_have_event("cpu", "cycles-ct") &&
1665                     pmu_have_event("cpu", "el-start"))
1666                         err = setup_events(transaction_attrs,
1667                                         ARRAY_SIZE(transaction_attrs));
1668                 else
1669                         err = setup_events(transaction_limited_attrs,
1670                                  ARRAY_SIZE(transaction_limited_attrs));
1671                 if (err < 0) {
1672                         fprintf(stderr, "Cannot set up transaction events\n");
1673                         return -1;
1674                 }
1675                 return 0;
1676         }
1677
1678         if (!evsel_list->nr_entries) {
1679                 if (perf_evlist__add_default_attrs(evsel_list, default_attrs) < 0)
1680                         return -1;
1681         }
1682
1683         /* Detailed events get appended to the event list: */
1684
1685         if (detailed_run <  1)
1686                 return 0;
1687
1688         /* Append detailed run extra attributes: */
1689         if (perf_evlist__add_default_attrs(evsel_list, detailed_attrs) < 0)
1690                 return -1;
1691
1692         if (detailed_run < 2)
1693                 return 0;
1694
1695         /* Append very detailed run extra attributes: */
1696         if (perf_evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0)
1697                 return -1;
1698
1699         if (detailed_run < 3)
1700                 return 0;
1701
1702         /* Append very, very detailed run extra attributes: */
1703         return perf_evlist__add_default_attrs(evsel_list, very_very_detailed_attrs);
1704 }
1705
1706 int cmd_stat(int argc, const char **argv, const char *prefix __maybe_unused)
1707 {
1708         bool append_file = false;
1709         int output_fd = 0;
1710         const char *output_name = NULL;
1711         const struct option options[] = {
1712         OPT_BOOLEAN('T', "transaction", &transaction_run,
1713                     "hardware transaction statistics"),
1714         OPT_CALLBACK('e', "event", &evsel_list, "event",
1715                      "event selector. use 'perf list' to list available events",
1716                      parse_events_option),
1717         OPT_CALLBACK(0, "filter", &evsel_list, "filter",
1718                      "event filter", parse_filter),
1719         OPT_BOOLEAN('i', "no-inherit", &no_inherit,
1720                     "child tasks do not inherit counters"),
1721         OPT_STRING('p', "pid", &target.pid, "pid",
1722                    "stat events on existing process id"),
1723         OPT_STRING('t', "tid", &target.tid, "tid",
1724                    "stat events on existing thread id"),
1725         OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
1726                     "system-wide collection from all CPUs"),
1727         OPT_BOOLEAN('g', "group", &group,
1728                     "put the counters into a counter group"),
1729         OPT_BOOLEAN('c', "scale", &scale, "scale/normalize counters"),
1730         OPT_INCR('v', "verbose", &verbose,
1731                     "be more verbose (show counter open errors, etc)"),
1732         OPT_INTEGER('r', "repeat", &run_count,
1733                     "repeat command and print average + stddev (max: 100, forever: 0)"),
1734         OPT_BOOLEAN('n', "null", &null_run,
1735                     "null run - dont start any counters"),
1736         OPT_INCR('d', "detailed", &detailed_run,
1737                     "detailed run - start a lot of events"),
1738         OPT_BOOLEAN('S', "sync", &sync_run,
1739                     "call sync() before starting a run"),
1740         OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
1741                            "print large numbers with thousands\' separators",
1742                            stat__set_big_num),
1743         OPT_STRING('C', "cpu", &target.cpu_list, "cpu",
1744                     "list of cpus to monitor in system-wide"),
1745         OPT_SET_UINT('A', "no-aggr", &aggr_mode,
1746                     "disable CPU count aggregation", AGGR_NONE),
1747         OPT_STRING('x', "field-separator", &csv_sep, "separator",
1748                    "print counts with custom separator"),
1749         OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
1750                      "monitor event in cgroup name only", parse_cgroups),
1751         OPT_STRING('o', "output", &output_name, "file", "output file name"),
1752         OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
1753         OPT_INTEGER(0, "log-fd", &output_fd,
1754                     "log output to fd, instead of stderr"),
1755         OPT_STRING(0, "pre", &pre_cmd, "command",
1756                         "command to run prior to the measured command"),
1757         OPT_STRING(0, "post", &post_cmd, "command",
1758                         "command to run after to the measured command"),
1759         OPT_UINTEGER('I', "interval-print", &interval,
1760                     "print counts at regular interval in ms (>= 100)"),
1761         OPT_SET_UINT(0, "per-socket", &aggr_mode,
1762                      "aggregate counts per processor socket", AGGR_SOCKET),
1763         OPT_SET_UINT(0, "per-core", &aggr_mode,
1764                      "aggregate counts per physical processor core", AGGR_CORE),
1765         OPT_UINTEGER('D', "delay", &initial_delay,
1766                      "ms to wait before starting measurement after program start"),
1767         OPT_END()
1768         };
1769         const char * const stat_usage[] = {
1770                 "perf stat [<options>] [<command>]",
1771                 NULL
1772         };
1773         int status = -EINVAL, run_idx;
1774         const char *mode;
1775
1776         setlocale(LC_ALL, "");
1777
1778         evsel_list = perf_evlist__new();
1779         if (evsel_list == NULL)
1780                 return -ENOMEM;
1781
1782         argc = parse_options(argc, argv, options, stat_usage,
1783                 PARSE_OPT_STOP_AT_NON_OPTION);
1784
1785         output = stderr;
1786         if (output_name && strcmp(output_name, "-"))
1787                 output = NULL;
1788
1789         if (output_name && output_fd) {
1790                 fprintf(stderr, "cannot use both --output and --log-fd\n");
1791                 parse_options_usage(stat_usage, options, "o", 1);
1792                 parse_options_usage(NULL, options, "log-fd", 0);
1793                 goto out;
1794         }
1795
1796         if (output_fd < 0) {
1797                 fprintf(stderr, "argument to --log-fd must be a > 0\n");
1798                 parse_options_usage(stat_usage, options, "log-fd", 0);
1799                 goto out;
1800         }
1801
1802         if (!output) {
1803                 struct timespec tm;
1804                 mode = append_file ? "a" : "w";
1805
1806                 output = fopen(output_name, mode);
1807                 if (!output) {
1808                         perror("failed to create output file");
1809                         return -1;
1810                 }
1811                 clock_gettime(CLOCK_REALTIME, &tm);
1812                 fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
1813         } else if (output_fd > 0) {
1814                 mode = append_file ? "a" : "w";
1815                 output = fdopen(output_fd, mode);
1816                 if (!output) {
1817                         perror("Failed opening logfd");
1818                         return -errno;
1819                 }
1820         }
1821
1822         if (csv_sep) {
1823                 csv_output = true;
1824                 if (!strcmp(csv_sep, "\\t"))
1825                         csv_sep = "\t";
1826         } else
1827                 csv_sep = DEFAULT_SEPARATOR;
1828
1829         /*
1830          * let the spreadsheet do the pretty-printing
1831          */
1832         if (csv_output) {
1833                 /* User explicitly passed -B? */
1834                 if (big_num_opt == 1) {
1835                         fprintf(stderr, "-B option not supported with -x\n");
1836                         parse_options_usage(stat_usage, options, "B", 1);
1837                         parse_options_usage(NULL, options, "x", 1);
1838                         goto out;
1839                 } else /* Nope, so disable big number formatting */
1840                         big_num = false;
1841         } else if (big_num_opt == 0) /* User passed --no-big-num */
1842                 big_num = false;
1843
1844         if (!argc && target__none(&target))
1845                 usage_with_options(stat_usage, options);
1846
1847         if (run_count < 0) {
1848                 pr_err("Run count must be a positive number\n");
1849                 parse_options_usage(stat_usage, options, "r", 1);
1850                 goto out;
1851         } else if (run_count == 0) {
1852                 forever = true;
1853                 run_count = 1;
1854         }
1855
1856         /* no_aggr, cgroup are for system-wide only */
1857         if ((aggr_mode != AGGR_GLOBAL || nr_cgroups) &&
1858             !target__has_cpu(&target)) {
1859                 fprintf(stderr, "both cgroup and no-aggregation "
1860                         "modes only available in system-wide mode\n");
1861
1862                 parse_options_usage(stat_usage, options, "G", 1);
1863                 parse_options_usage(NULL, options, "A", 1);
1864                 parse_options_usage(NULL, options, "a", 1);
1865                 goto out;
1866         }
1867
1868         if (add_default_attributes())
1869                 goto out;
1870
1871         target__validate(&target);
1872
1873         if (perf_evlist__create_maps(evsel_list, &target) < 0) {
1874                 if (target__has_task(&target)) {
1875                         pr_err("Problems finding threads of monitor\n");
1876                         parse_options_usage(stat_usage, options, "p", 1);
1877                         parse_options_usage(NULL, options, "t", 1);
1878                 } else if (target__has_cpu(&target)) {
1879                         perror("failed to parse CPUs map");
1880                         parse_options_usage(stat_usage, options, "C", 1);
1881                         parse_options_usage(NULL, options, "a", 1);
1882                 }
1883                 goto out;
1884         }
1885         if (interval && interval < 100) {
1886                 pr_err("print interval must be >= 100ms\n");
1887                 parse_options_usage(stat_usage, options, "I", 1);
1888                 goto out;
1889         }
1890
1891         if (perf_evlist__alloc_stats(evsel_list, interval))
1892                 goto out;
1893
1894         if (perf_stat_init_aggr_mode())
1895                 goto out;
1896
1897         /*
1898          * We dont want to block the signals - that would cause
1899          * child tasks to inherit that and Ctrl-C would not work.
1900          * What we want is for Ctrl-C to work in the exec()-ed
1901          * task, but being ignored by perf stat itself:
1902          */
1903         atexit(sig_atexit);
1904         if (!forever)
1905                 signal(SIGINT,  skip_signal);
1906         signal(SIGCHLD, skip_signal);
1907         signal(SIGALRM, skip_signal);
1908         signal(SIGABRT, skip_signal);
1909
1910         status = 0;
1911         for (run_idx = 0; forever || run_idx < run_count; run_idx++) {
1912                 if (run_count != 1 && verbose)
1913                         fprintf(output, "[ perf stat: executing run #%d ... ]\n",
1914                                 run_idx + 1);
1915
1916                 status = run_perf_stat(argc, argv);
1917                 if (forever && status != -1) {
1918                         print_stat(argc, argv);
1919                         perf_stat__reset_stats(evsel_list);
1920                 }
1921         }
1922
1923         if (!forever && status != -1 && !interval)
1924                 print_stat(argc, argv);
1925
1926         perf_evlist__free_stats(evsel_list);
1927 out:
1928         perf_evlist__delete(evsel_list);
1929         return status;
1930 }