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