?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
Linux 4.1.18
[linux/fpc-iii.git] / tools / perf / builtin-stat.c
bloba1f3ffc2786def5a01ea4bf95371d9b573aea71d
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, &counter)) {
688 error("failed to set filter \"%s\" on event %s with %d (%s)\n",
689 counter->filter, perf_evsel__name(counter), errno,
690 strerror_r(errno, msg, sizeof(msg)));
691 return -1;
692 }
693
694 /*
695 * Enable counters and exec the command:
696 */
697 t0 = rdclock();
698 clock_gettime(CLOCK_MONOTONIC, &ref_time);
699
700 if (forks) {
701 perf_evlist__start_workload(evsel_list);
702 handle_initial_delay();
703
704 if (interval) {
705 while (!waitpid(child_pid, &status, WNOHANG)) {
706 nanosleep(&ts, NULL);
707 print_interval();
708 }
709 }
710 wait(&status);
711
712 if (workload_exec_errno) {
713 const char *emsg = strerror_r(workload_exec_errno, msg, sizeof(msg));
714 pr_err("Workload failed: %s\n", emsg);
715 return -1;
716 }
717
718 if (WIFSIGNALED(status))
719 psignal(WTERMSIG(status), argv[0]);
720 } else {
721 handle_initial_delay();
722 while (!done) {
723 nanosleep(&ts, NULL);
724 if (interval)
725 print_interval();
726 }
727 }
728
729 t1 = rdclock();
730
731 update_stats(&walltime_nsecs_stats, t1 - t0);
732
733 if (aggr_mode == AGGR_GLOBAL) {
734 evlist__for_each(evsel_list, counter) {
735 read_counter_aggr(counter);
736 perf_evsel__close_fd(counter, perf_evsel__nr_cpus(counter),
737 thread_map__nr(evsel_list->threads));
738 }
739 } else {
740 evlist__for_each(evsel_list, counter) {
741 read_counter(counter);
742 perf_evsel__close_fd(counter, perf_evsel__nr_cpus(counter), 1);
743 }
744 }
745
746 return WEXITSTATUS(status);
747 }
748
749 static int run_perf_stat(int argc, const char **argv)
750 {
751 int ret;
752
753 if (pre_cmd) {
754 ret = system(pre_cmd);
755 if (ret)
756 return ret;
757 }
758
759 if (sync_run)
760 sync();
761
762 ret = __run_perf_stat(argc, argv);
763 if (ret)
764 return ret;
765
766 if (post_cmd) {
767 ret = system(post_cmd);
768 if (ret)
769 return ret;
770 }
771
772 return ret;
773 }
774
775 static void print_running(u64 run, u64 ena)
776 {
777 if (csv_output) {
778 fprintf(output, "%s%" PRIu64 "%s%.2f",
779 csv_sep,
780 run,
781 csv_sep,
782 ena ? 100.0 * run / ena : 100.0);
783 } else if (run != ena) {
784 fprintf(output, " (%.2f%%)", 100.0 * run / ena);
785 }
786 }
787
788 static void print_noise_pct(double total, double avg)
789 {
790 double pct = rel_stddev_stats(total, avg);
791
792 if (csv_output)
793 fprintf(output, "%s%.2f%%", csv_sep, pct);
794 else if (pct)
795 fprintf(output, " ( +-%6.2f%% )", pct);
796 }
797
798 static void print_noise(struct perf_evsel *evsel, double avg)
799 {
800 struct perf_stat *ps;
801
802 if (run_count == 1)
803 return;
804
805 ps = evsel->priv;
806 print_noise_pct(stddev_stats(&ps->res_stats[0]), avg);
807 }
808
809 static void aggr_printout(struct perf_evsel *evsel, int id, int nr)
810 {
811 switch (aggr_mode) {
812 case AGGR_CORE:
813 fprintf(output, "S%d-C%*d%s%*d%s",
814 cpu_map__id_to_socket(id),
815 csv_output ? 0 : -8,
816 cpu_map__id_to_cpu(id),
817 csv_sep,
818 csv_output ? 0 : 4,
819 nr,
820 csv_sep);
821 break;
822 case AGGR_SOCKET:
823 fprintf(output, "S%*d%s%*d%s",
824 csv_output ? 0 : -5,
825 id,
826 csv_sep,
827 csv_output ? 0 : 4,
828 nr,
829 csv_sep);
830 break;
831 case AGGR_NONE:
832 fprintf(output, "CPU%*d%s",
833 csv_output ? 0 : -4,
834 perf_evsel__cpus(evsel)->map[id], csv_sep);
835 break;
836 case AGGR_GLOBAL:
837 default:
838 break;
839 }
840 }
841
842 static void nsec_printout(int id, int nr, struct perf_evsel *evsel, double avg)
843 {
844 double msecs = avg / 1e6;
845 const char *fmt_v, *fmt_n;
846 char name[25];
847
848 fmt_v = csv_output ? "%.6f%s" : "%18.6f%s";
849 fmt_n = csv_output ? "%s" : "%-25s";
850
851 aggr_printout(evsel, id, nr);
852
853 scnprintf(name, sizeof(name), "%s%s",
854 perf_evsel__name(evsel), csv_output ? "" : " (msec)");
855
856 fprintf(output, fmt_v, msecs, csv_sep);
857
858 if (csv_output)
859 fprintf(output, "%s%s", evsel->unit, csv_sep);
860 else
861 fprintf(output, "%-*s%s", unit_width, evsel->unit, csv_sep);
862
863 fprintf(output, fmt_n, name);
864
865 if (evsel->cgrp)
866 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
867
868 if (csv_output || interval)
869 return;
870
871 if (perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
872 fprintf(output, " # %8.3f CPUs utilized ",
873 avg / avg_stats(&walltime_nsecs_stats));
874 else
875 fprintf(output, " ");
876 }
877
878 /* used for get_ratio_color() */
879 enum grc_type {
880 GRC_STALLED_CYCLES_FE,
881 GRC_STALLED_CYCLES_BE,
882 GRC_CACHE_MISSES,
883 GRC_MAX_NR
884 };
885
886 static const char *get_ratio_color(enum grc_type type, double ratio)
887 {
888 static const double grc_table[GRC_MAX_NR][3] = {
889 [GRC_STALLED_CYCLES_FE] = { 50.0, 30.0, 10.0 },
890 [GRC_STALLED_CYCLES_BE] = { 75.0, 50.0, 20.0 },
891 [GRC_CACHE_MISSES] = { 20.0, 10.0, 5.0 },
892 };
893 const char *color = PERF_COLOR_NORMAL;
894
895 if (ratio > grc_table[type][0])
896 color = PERF_COLOR_RED;
897 else if (ratio > grc_table[type][1])
898 color = PERF_COLOR_MAGENTA;
899 else if (ratio > grc_table[type][2])
900 color = PERF_COLOR_YELLOW;
901
902 return color;
903 }
904
905 static void print_stalled_cycles_frontend(int cpu,
906 struct perf_evsel *evsel
907 __maybe_unused, double avg)
908 {
909 double total, ratio = 0.0;
910 const char *color;
911
912 total = avg_stats(&runtime_cycles_stats[cpu]);
913
914 if (total)
915 ratio = avg / total * 100.0;
916
917 color = get_ratio_color(GRC_STALLED_CYCLES_FE, ratio);
918
919 fprintf(output, " # ");
920 color_fprintf(output, color, "%6.2f%%", ratio);
921 fprintf(output, " frontend cycles idle ");
922 }
923
924 static void print_stalled_cycles_backend(int cpu,
925 struct perf_evsel *evsel
926 __maybe_unused, double avg)
927 {
928 double total, ratio = 0.0;
929 const char *color;
930
931 total = avg_stats(&runtime_cycles_stats[cpu]);
932
933 if (total)
934 ratio = avg / total * 100.0;
935
936 color = get_ratio_color(GRC_STALLED_CYCLES_BE, ratio);
937
938 fprintf(output, " # ");
939 color_fprintf(output, color, "%6.2f%%", ratio);
940 fprintf(output, " backend cycles idle ");
941 }
942
943 static void print_branch_misses(int cpu,
944 struct perf_evsel *evsel __maybe_unused,
945 double avg)
946 {
947 double total, ratio = 0.0;
948 const char *color;
949
950 total = avg_stats(&runtime_branches_stats[cpu]);
951
952 if (total)
953 ratio = avg / total * 100.0;
954
955 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
956
957 fprintf(output, " # ");
958 color_fprintf(output, color, "%6.2f%%", ratio);
959 fprintf(output, " of all branches ");
960 }
961
962 static void print_l1_dcache_misses(int cpu,
963 struct perf_evsel *evsel __maybe_unused,
964 double avg)
965 {
966 double total, ratio = 0.0;
967 const char *color;
968
969 total = avg_stats(&runtime_l1_dcache_stats[cpu]);
970
971 if (total)
972 ratio = avg / total * 100.0;
973
974 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
975
976 fprintf(output, " # ");
977 color_fprintf(output, color, "%6.2f%%", ratio);
978 fprintf(output, " of all L1-dcache hits ");
979 }
980
981 static void print_l1_icache_misses(int cpu,
982 struct perf_evsel *evsel __maybe_unused,
983 double avg)
984 {
985 double total, ratio = 0.0;
986 const char *color;
987
988 total = avg_stats(&runtime_l1_icache_stats[cpu]);
989
990 if (total)
991 ratio = avg / total * 100.0;
992
993 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
994
995 fprintf(output, " # ");
996 color_fprintf(output, color, "%6.2f%%", ratio);
997 fprintf(output, " of all L1-icache hits ");
998 }
999
1000 static void print_dtlb_cache_misses(int cpu,
1001 struct perf_evsel *evsel __maybe_unused,
1002 double avg)
1003 {
1004 double total, ratio = 0.0;
1005 const char *color;
1006
1007 total = avg_stats(&runtime_dtlb_cache_stats[cpu]);
1008
1009 if (total)
1010 ratio = avg / total * 100.0;
1011
1012 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
1013
1014 fprintf(output, " # ");
1015 color_fprintf(output, color, "%6.2f%%", ratio);
1016 fprintf(output, " of all dTLB cache hits ");
1017 }
1018
1019 static void print_itlb_cache_misses(int cpu,
1020 struct perf_evsel *evsel __maybe_unused,
1021 double avg)
1022 {
1023 double total, ratio = 0.0;
1024 const char *color;
1025
1026 total = avg_stats(&runtime_itlb_cache_stats[cpu]);
1027
1028 if (total)
1029 ratio = avg / total * 100.0;
1030
1031 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
1032
1033 fprintf(output, " # ");
1034 color_fprintf(output, color, "%6.2f%%", ratio);
1035 fprintf(output, " of all iTLB cache hits ");
1036 }
1037
1038 static void print_ll_cache_misses(int cpu,
1039 struct perf_evsel *evsel __maybe_unused,
1040 double avg)
1041 {
1042 double total, ratio = 0.0;
1043 const char *color;
1044
1045 total = avg_stats(&runtime_ll_cache_stats[cpu]);
1046
1047 if (total)
1048 ratio = avg / total * 100.0;
1049
1050 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
1051
1052 fprintf(output, " # ");
1053 color_fprintf(output, color, "%6.2f%%", ratio);
1054 fprintf(output, " of all LL-cache hits ");
1055 }
1056
1057 static void abs_printout(int id, int nr, struct perf_evsel *evsel, double avg)
1058 {
1059 double total, ratio = 0.0, total2;
1060 double sc = evsel->scale;
1061 const char *fmt;
1062 int cpu = cpu_map__id_to_cpu(id);
1063
1064 if (csv_output) {
1065 fmt = sc != 1.0 ? "%.2f%s" : "%.0f%s";
1066 } else {
1067 if (big_num)
1068 fmt = sc != 1.0 ? "%'18.2f%s" : "%'18.0f%s";
1069 else
1070 fmt = sc != 1.0 ? "%18.2f%s" : "%18.0f%s";
1071 }
1072
1073 aggr_printout(evsel, id, nr);
1074
1075 if (aggr_mode == AGGR_GLOBAL)
1076 cpu = 0;
1077
1078 fprintf(output, fmt, avg, csv_sep);
1079
1080 if (evsel->unit)
1081 fprintf(output, "%-*s%s",
1082 csv_output ? 0 : unit_width,
1083 evsel->unit, csv_sep);
1084
1085 fprintf(output, "%-*s", csv_output ? 0 : 25, perf_evsel__name(evsel));
1086
1087 if (evsel->cgrp)
1088 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
1089
1090 if (csv_output || interval)
1091 return;
1092
1093 if (perf_evsel__match(evsel, HARDWARE, HW_INSTRUCTIONS)) {
1094 total = avg_stats(&runtime_cycles_stats[cpu]);
1095 if (total) {
1096 ratio = avg / total;
1097 fprintf(output, " # %5.2f insns per cycle ", ratio);
1098 } else {
1099 fprintf(output, " ");
1100 }
1101 total = avg_stats(&runtime_stalled_cycles_front_stats[cpu]);
1102 total = max(total, avg_stats(&runtime_stalled_cycles_back_stats[cpu]));
1103
1104 if (total && avg) {
1105 ratio = total / avg;
1106 fprintf(output, "\n");
1107 if (aggr_mode == AGGR_NONE)
1108 fprintf(output, " ");
1109 fprintf(output, " # %5.2f stalled cycles per insn", ratio);
1110 }
1111
1112 } else if (perf_evsel__match(evsel, HARDWARE, HW_BRANCH_MISSES) &&
1113 runtime_branches_stats[cpu].n != 0) {
1114 print_branch_misses(cpu, evsel, avg);
1115 } else if (
1116 evsel->attr.type == PERF_TYPE_HW_CACHE &&
1117 evsel->attr.config == ( PERF_COUNT_HW_CACHE_L1D |
1118 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
1119 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
1120 runtime_l1_dcache_stats[cpu].n != 0) {
1121 print_l1_dcache_misses(cpu, evsel, avg);
1122 } else if (
1123 evsel->attr.type == PERF_TYPE_HW_CACHE &&
1124 evsel->attr.config == ( PERF_COUNT_HW_CACHE_L1I |
1125 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
1126 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
1127 runtime_l1_icache_stats[cpu].n != 0) {
1128 print_l1_icache_misses(cpu, evsel, avg);
1129 } else if (
1130 evsel->attr.type == PERF_TYPE_HW_CACHE &&
1131 evsel->attr.config == ( PERF_COUNT_HW_CACHE_DTLB |
1132 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
1133 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
1134 runtime_dtlb_cache_stats[cpu].n != 0) {
1135 print_dtlb_cache_misses(cpu, evsel, avg);
1136 } else if (
1137 evsel->attr.type == PERF_TYPE_HW_CACHE &&
1138 evsel->attr.config == ( PERF_COUNT_HW_CACHE_ITLB |
1139 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
1140 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
1141 runtime_itlb_cache_stats[cpu].n != 0) {
1142 print_itlb_cache_misses(cpu, evsel, avg);
1143 } else if (
1144 evsel->attr.type == PERF_TYPE_HW_CACHE &&
1145 evsel->attr.config == ( PERF_COUNT_HW_CACHE_LL |
1146 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
1147 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
1148 runtime_ll_cache_stats[cpu].n != 0) {
1149 print_ll_cache_misses(cpu, evsel, avg);
1150 } else if (perf_evsel__match(evsel, HARDWARE, HW_CACHE_MISSES) &&
1151 runtime_cacherefs_stats[cpu].n != 0) {
1152 total = avg_stats(&runtime_cacherefs_stats[cpu]);
1153
1154 if (total)
1155 ratio = avg * 100 / total;
1156
1157 fprintf(output, " # %8.3f %% of all cache refs ", ratio);
1158
1159 } else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_FRONTEND)) {
1160 print_stalled_cycles_frontend(cpu, evsel, avg);
1161 } else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_BACKEND)) {
1162 print_stalled_cycles_backend(cpu, evsel, avg);
1163 } else if (perf_evsel__match(evsel, HARDWARE, HW_CPU_CYCLES)) {
1164 total = avg_stats(&runtime_nsecs_stats[cpu]);
1165
1166 if (total) {
1167 ratio = avg / total;
1168 fprintf(output, " # %8.3f GHz ", ratio);
1169 } else {
1170 fprintf(output, " ");
1171 }
1172 } else if (transaction_run &&
1173 perf_evsel__cmp(evsel, nth_evsel(T_CYCLES_IN_TX))) {
1174 total = avg_stats(&runtime_cycles_stats[cpu]);
1175 if (total)
1176 fprintf(output,
1177 " # %5.2f%% transactional cycles ",
1178 100.0 * (avg / total));
1179 } else if (transaction_run &&
1180 perf_evsel__cmp(evsel, nth_evsel(T_CYCLES_IN_TX_CP))) {
1181 total = avg_stats(&runtime_cycles_stats[cpu]);
1182 total2 = avg_stats(&runtime_cycles_in_tx_stats[cpu]);
1183 if (total2 < avg)
1184 total2 = avg;
1185 if (total)
1186 fprintf(output,
1187 " # %5.2f%% aborted cycles ",
1188 100.0 * ((total2-avg) / total));
1189 } else if (transaction_run &&
1190 perf_evsel__cmp(evsel, nth_evsel(T_TRANSACTION_START)) &&
1191 avg > 0 &&
1192 runtime_cycles_in_tx_stats[cpu].n != 0) {
1193 total = avg_stats(&runtime_cycles_in_tx_stats[cpu]);
1194
1195 if (total)
1196 ratio = total / avg;
1197
1198 fprintf(output, " # %8.0f cycles / transaction ", ratio);
1199 } else if (transaction_run &&
1200 perf_evsel__cmp(evsel, nth_evsel(T_ELISION_START)) &&
1201 avg > 0 &&
1202 runtime_cycles_in_tx_stats[cpu].n != 0) {
1203 total = avg_stats(&runtime_cycles_in_tx_stats[cpu]);
1204
1205 if (total)
1206 ratio = total / avg;
1207
1208 fprintf(output, " # %8.0f cycles / elision ", ratio);
1209 } else if (runtime_nsecs_stats[cpu].n != 0) {
1210 char unit = 'M';
1211
1212 total = avg_stats(&runtime_nsecs_stats[cpu]);
1213
1214 if (total)
1215 ratio = 1000.0 * avg / total;
1216 if (ratio < 0.001) {
1217 ratio *= 1000;
1218 unit = 'K';
1219 }
1220
1221 fprintf(output, " # %8.3f %c/sec ", ratio, unit);
1222 } else {
1223 fprintf(output, " ");
1224 }
1225 }
1226
1227 static void print_aggr(char *prefix)
1228 {
1229 struct perf_evsel *counter;
1230 int cpu, s, s2, id, nr;
1231 double uval;
1232 u64 ena, run, val;
1233
1234 if (!(aggr_map || aggr_get_id))
1235 return;
1236
1237 for (s = 0; s < aggr_map->nr; s++) {
1238 id = aggr_map->map[s];
1239 evlist__for_each(evsel_list, counter) {
1240 val = ena = run = 0;
1241 nr = 0;
1242 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
1243 s2 = aggr_get_id(perf_evsel__cpus(counter), cpu);
1244 if (s2 != id)
1245 continue;
1246 val += counter->counts->cpu[cpu].val;
1247 ena += counter->counts->cpu[cpu].ena;
1248 run += counter->counts->cpu[cpu].run;
1249 nr++;
1250 }
1251 if (prefix)
1252 fprintf(output, "%s", prefix);
1253
1254 if (run == 0 || ena == 0) {
1255 aggr_printout(counter, id, nr);
1256
1257 fprintf(output, "%*s%s",
1258 csv_output ? 0 : 18,
1259 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
1260 csv_sep);
1261
1262 fprintf(output, "%-*s%s",
1263 csv_output ? 0 : unit_width,
1264 counter->unit, csv_sep);
1265
1266 fprintf(output, "%*s",
1267 csv_output ? 0 : -25,
1268 perf_evsel__name(counter));
1269
1270 if (counter->cgrp)
1271 fprintf(output, "%s%s",
1272 csv_sep, counter->cgrp->name);
1273
1274 print_running(run, ena);
1275 fputc('\n', output);
1276 continue;
1277 }
1278 uval = val * counter->scale;
1279
1280 if (nsec_counter(counter))
1281 nsec_printout(id, nr, counter, uval);
1282 else
1283 abs_printout(id, nr, counter, uval);
1284
1285 if (!csv_output)
1286 print_noise(counter, 1.0);
1287
1288 print_running(run, ena);
1289 fputc('\n', output);
1290 }
1291 }
1292 }
1293
1294 /*
1295 * Print out the results of a single counter:
1296 * aggregated counts in system-wide mode
1297 */
1298 static void print_counter_aggr(struct perf_evsel *counter, char *prefix)
1299 {
1300 struct perf_stat *ps = counter->priv;
1301 double avg = avg_stats(&ps->res_stats[0]);
1302 int scaled = counter->counts->scaled;
1303 double uval;
1304 double avg_enabled, avg_running;
1305
1306 avg_enabled = avg_stats(&ps->res_stats[1]);
1307 avg_running = avg_stats(&ps->res_stats[2]);
1308
1309 if (prefix)
1310 fprintf(output, "%s", prefix);
1311
1312 if (scaled == -1 || !counter->supported) {
1313 fprintf(output, "%*s%s",
1314 csv_output ? 0 : 18,
1315 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
1316 csv_sep);
1317 fprintf(output, "%-*s%s",
1318 csv_output ? 0 : unit_width,
1319 counter->unit, csv_sep);
1320 fprintf(output, "%*s",
1321 csv_output ? 0 : -25,
1322 perf_evsel__name(counter));
1323
1324 if (counter->cgrp)
1325 fprintf(output, "%s%s", csv_sep, counter->cgrp->name);
1326
1327 print_running(avg_running, avg_enabled);
1328 fputc('\n', output);
1329 return;
1330 }
1331
1332 uval = avg * counter->scale;
1333
1334 if (nsec_counter(counter))
1335 nsec_printout(-1, 0, counter, uval);
1336 else
1337 abs_printout(-1, 0, counter, uval);
1338
1339 print_noise(counter, avg);
1340
1341 print_running(avg_running, avg_enabled);
1342 fprintf(output, "\n");
1343 }
1344
1345 /*
1346 * Print out the results of a single counter:
1347 * does not use aggregated count in system-wide
1348 */
1349 static void print_counter(struct perf_evsel *counter, char *prefix)
1350 {
1351 u64 ena, run, val;
1352 double uval;
1353 int cpu;
1354
1355 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
1356 val = counter->counts->cpu[cpu].val;
1357 ena = counter->counts->cpu[cpu].ena;
1358 run = counter->counts->cpu[cpu].run;
1359
1360 if (prefix)
1361 fprintf(output, "%s", prefix);
1362
1363 if (run == 0 || ena == 0) {
1364 fprintf(output, "CPU%*d%s%*s%s",
1365 csv_output ? 0 : -4,
1366 perf_evsel__cpus(counter)->map[cpu], csv_sep,
1367 csv_output ? 0 : 18,
1368 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
1369 csv_sep);
1370
1371 fprintf(output, "%-*s%s",
1372 csv_output ? 0 : unit_width,
1373 counter->unit, csv_sep);
1374
1375 fprintf(output, "%*s",
1376 csv_output ? 0 : -25,
1377 perf_evsel__name(counter));
1378
1379 if (counter->cgrp)
1380 fprintf(output, "%s%s",
1381 csv_sep, counter->cgrp->name);
1382
1383 print_running(run, ena);
1384 fputc('\n', output);
1385 continue;
1386 }
1387
1388 uval = val * counter->scale;
1389
1390 if (nsec_counter(counter))
1391 nsec_printout(cpu, 0, counter, uval);
1392 else
1393 abs_printout(cpu, 0, counter, uval);
1394
1395 if (!csv_output)
1396 print_noise(counter, 1.0);
1397 print_running(run, ena);
1398
1399 fputc('\n', output);
1400 }
1401 }
1402
1403 static void print_stat(int argc, const char **argv)
1404 {
1405 struct perf_evsel *counter;
1406 int i;
1407
1408 fflush(stdout);
1409
1410 if (!csv_output) {
1411 fprintf(output, "\n");
1412 fprintf(output, " Performance counter stats for ");
1413 if (target.system_wide)
1414 fprintf(output, "\'system wide");
1415 else if (target.cpu_list)
1416 fprintf(output, "\'CPU(s) %s", target.cpu_list);
1417 else if (!target__has_task(&target)) {
1418 fprintf(output, "\'%s", argv[0]);
1419 for (i = 1; i < argc; i++)
1420 fprintf(output, " %s", argv[i]);
1421 } else if (target.pid)
1422 fprintf(output, "process id \'%s", target.pid);
1423 else
1424 fprintf(output, "thread id \'%s", target.tid);
1425
1426 fprintf(output, "\'");
1427 if (run_count > 1)
1428 fprintf(output, " (%d runs)", run_count);
1429 fprintf(output, ":\n\n");
1430 }
1431
1432 switch (aggr_mode) {
1433 case AGGR_CORE:
1434 case AGGR_SOCKET:
1435 print_aggr(NULL);
1436 break;
1437 case AGGR_GLOBAL:
1438 evlist__for_each(evsel_list, counter)
1439 print_counter_aggr(counter, NULL);
1440 break;
1441 case AGGR_NONE:
1442 evlist__for_each(evsel_list, counter)
1443 print_counter(counter, NULL);
1444 break;
1445 default:
1446 break;
1447 }
1448
1449 if (!csv_output) {
1450 if (!null_run)
1451 fprintf(output, "\n");
1452 fprintf(output, " %17.9f seconds time elapsed",
1453 avg_stats(&walltime_nsecs_stats)/1e9);
1454 if (run_count > 1) {
1455 fprintf(output, " ");
1456 print_noise_pct(stddev_stats(&walltime_nsecs_stats),
1457 avg_stats(&walltime_nsecs_stats));
1458 }
1459 fprintf(output, "\n\n");
1460 }
1461 }
1462
1463 static volatile int signr = -1;
1464
1465 static void skip_signal(int signo)
1466 {
1467 if ((child_pid == -1) || interval)
1468 done = 1;
1469
1470 signr = signo;
1471 /*
1472 * render child_pid harmless
1473 * won't send SIGTERM to a random
1474 * process in case of race condition
1475 * and fast PID recycling
1476 */
1477 child_pid = -1;
1478 }
1479
1480 static void sig_atexit(void)
1481 {
1482 sigset_t set, oset;
1483
1484 /*
1485 * avoid race condition with SIGCHLD handler
1486 * in skip_signal() which is modifying child_pid
1487 * goal is to avoid send SIGTERM to a random
1488 * process
1489 */
1490 sigemptyset(&set);
1491 sigaddset(&set, SIGCHLD);
1492 sigprocmask(SIG_BLOCK, &set, &oset);
1493
1494 if (child_pid != -1)
1495 kill(child_pid, SIGTERM);
1496
1497 sigprocmask(SIG_SETMASK, &oset, NULL);
1498
1499 if (signr == -1)
1500 return;
1501
1502 signal(signr, SIG_DFL);
1503 kill(getpid(), signr);
1504 }
1505
1506 static int stat__set_big_num(const struct option *opt __maybe_unused,
1507 const char *s __maybe_unused, int unset)
1508 {
1509 big_num_opt = unset ? 0 : 1;
1510 return 0;
1511 }
1512
1513 static int perf_stat_init_aggr_mode(void)
1514 {
1515 switch (aggr_mode) {
1516 case AGGR_SOCKET:
1517 if (cpu_map__build_socket_map(evsel_list->cpus, &aggr_map)) {
1518 perror("cannot build socket map");
1519 return -1;
1520 }
1521 aggr_get_id = cpu_map__get_socket;
1522 break;
1523 case AGGR_CORE:
1524 if (cpu_map__build_core_map(evsel_list->cpus, &aggr_map)) {
1525 perror("cannot build core map");
1526 return -1;
1527 }
1528 aggr_get_id = cpu_map__get_core;
1529 break;
1530 case AGGR_NONE:
1531 case AGGR_GLOBAL:
1532 default:
1533 break;
1534 }
1535 return 0;
1536 }
1537
1538 static int setup_events(const char * const *attrs, unsigned len)
1539 {
1540 unsigned i;
1541
1542 for (i = 0; i < len; i++) {
1543 if (parse_events(evsel_list, attrs[i]))
1544 return -1;
1545 }
1546 return 0;
1547 }
1548
1549 /*
1550 * Add default attributes, if there were no attributes specified or
1551 * if -d/--detailed, -d -d or -d -d -d is used:
1552 */
1553 static int add_default_attributes(void)
1554 {
1555 struct perf_event_attr default_attrs[] = {
1556
1557 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK },
1558 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES },
1559 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS },
1560 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS },
1561
1562 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES },
1563 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND },
1564 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND },
1565 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS },
1566 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS },
1567 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES },
1568
1569 };
1570
1571 /*
1572 * Detailed stats (-d), covering the L1 and last level data caches:
1573 */
1574 struct perf_event_attr detailed_attrs[] = {
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_ACCESS << 16) },
1581
1582 { .type = PERF_TYPE_HW_CACHE,
1583 .config =
1584 PERF_COUNT_HW_CACHE_L1D << 0 |
1585 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1586 (PERF_COUNT_HW_CACHE_RESULT_MISS << 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_ACCESS << 16) },
1593
1594 { .type = PERF_TYPE_HW_CACHE,
1595 .config =
1596 PERF_COUNT_HW_CACHE_LL << 0 |
1597 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1598 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1599 };
1600
1601 /*
1602 * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
1603 */
1604 struct perf_event_attr very_detailed_attrs[] = {
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_ACCESS << 16) },
1611
1612 { .type = PERF_TYPE_HW_CACHE,
1613 .config =
1614 PERF_COUNT_HW_CACHE_L1I << 0 |
1615 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1616 (PERF_COUNT_HW_CACHE_RESULT_MISS << 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_ACCESS << 16) },
1623
1624 { .type = PERF_TYPE_HW_CACHE,
1625 .config =
1626 PERF_COUNT_HW_CACHE_DTLB << 0 |
1627 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1628 (PERF_COUNT_HW_CACHE_RESULT_MISS << 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_ACCESS << 16) },
1635
1636 { .type = PERF_TYPE_HW_CACHE,
1637 .config =
1638 PERF_COUNT_HW_CACHE_ITLB << 0 |
1639 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1640 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1641
1642 };
1643
1644 /*
1645 * Very, very detailed stats (-d -d -d), adding prefetch events:
1646 */
1647 struct perf_event_attr very_very_detailed_attrs[] = {
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_ACCESS << 16) },
1654
1655 { .type = PERF_TYPE_HW_CACHE,
1656 .config =
1657 PERF_COUNT_HW_CACHE_L1D << 0 |
1658 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
1659 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1660 };
1661
1662 /* Set attrs if no event is selected and !null_run: */
1663 if (null_run)
1664 return 0;
1665
1666 if (transaction_run) {
1667 int err;
1668 if (pmu_have_event("cpu", "cycles-ct") &&
1669 pmu_have_event("cpu", "el-start"))
1670 err = setup_events(transaction_attrs,
1671 ARRAY_SIZE(transaction_attrs));
1672 else
1673 err = setup_events(transaction_limited_attrs,
1674 ARRAY_SIZE(transaction_limited_attrs));
1675 if (err < 0) {
1676 fprintf(stderr, "Cannot set up transaction events\n");
1677 return -1;
1678 }
1679 return 0;
1680 }
1681
1682 if (!evsel_list->nr_entries) {
1683 if (perf_evlist__add_default_attrs(evsel_list, default_attrs) < 0)
1684 return -1;
1685 }
1686
1687 /* Detailed events get appended to the event list: */
1688
1689 if (detailed_run < 1)
1690 return 0;
1691
1692 /* Append detailed run extra attributes: */
1693 if (perf_evlist__add_default_attrs(evsel_list, detailed_attrs) < 0)
1694 return -1;
1695
1696 if (detailed_run < 2)
1697 return 0;
1698
1699 /* Append very detailed run extra attributes: */
1700 if (perf_evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0)
1701 return -1;
1702
1703 if (detailed_run < 3)
1704 return 0;
1705
1706 /* Append very, very detailed run extra attributes: */
1707 return perf_evlist__add_default_attrs(evsel_list, very_very_detailed_attrs);
1708 }
1709
1710 int cmd_stat(int argc, const char **argv, const char *prefix __maybe_unused)
1711 {
1712 bool append_file = false;
1713 int output_fd = 0;
1714 const char *output_name = NULL;
1715 const struct option options[] = {
1716 OPT_BOOLEAN('T', "transaction", &transaction_run,
1717 "hardware transaction statistics"),
1718 OPT_CALLBACK('e', "event", &evsel_list, "event",
1719 "event selector. use 'perf list' to list available events",
1720 parse_events_option),
1721 OPT_CALLBACK(0, "filter", &evsel_list, "filter",
1722 "event filter", parse_filter),
1723 OPT_BOOLEAN('i', "no-inherit", &no_inherit,
1724 "child tasks do not inherit counters"),
1725 OPT_STRING('p', "pid", &target.pid, "pid",
1726 "stat events on existing process id"),
1727 OPT_STRING('t', "tid", &target.tid, "tid",
1728 "stat events on existing thread id"),
1729 OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
1730 "system-wide collection from all CPUs"),
1731 OPT_BOOLEAN('g', "group", &group,
1732 "put the counters into a counter group"),
1733 OPT_BOOLEAN('c', "scale", &scale, "scale/normalize counters"),
1734 OPT_INCR('v', "verbose", &verbose,
1735 "be more verbose (show counter open errors, etc)"),
1736 OPT_INTEGER('r', "repeat", &run_count,
1737 "repeat command and print average + stddev (max: 100, forever: 0)"),
1738 OPT_BOOLEAN('n', "null", &null_run,
1739 "null run - dont start any counters"),
1740 OPT_INCR('d', "detailed", &detailed_run,
1741 "detailed run - start a lot of events"),
1742 OPT_BOOLEAN('S', "sync", &sync_run,
1743 "call sync() before starting a run"),
1744 OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
1745 "print large numbers with thousands\' separators",
1746 stat__set_big_num),
1747 OPT_STRING('C', "cpu", &target.cpu_list, "cpu",
1748 "list of cpus to monitor in system-wide"),
1749 OPT_SET_UINT('A', "no-aggr", &aggr_mode,
1750 "disable CPU count aggregation", AGGR_NONE),
1751 OPT_STRING('x', "field-separator", &csv_sep, "separator",
1752 "print counts with custom separator"),
1753 OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
1754 "monitor event in cgroup name only", parse_cgroups),
1755 OPT_STRING('o', "output", &output_name, "file", "output file name"),
1756 OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
1757 OPT_INTEGER(0, "log-fd", &output_fd,
1758 "log output to fd, instead of stderr"),
1759 OPT_STRING(0, "pre", &pre_cmd, "command",
1760 "command to run prior to the measured command"),
1761 OPT_STRING(0, "post", &post_cmd, "command",
1762 "command to run after to the measured command"),
1763 OPT_UINTEGER('I', "interval-print", &interval,
1764 "print counts at regular interval in ms (>= 100)"),
1765 OPT_SET_UINT(0, "per-socket", &aggr_mode,
1766 "aggregate counts per processor socket", AGGR_SOCKET),
1767 OPT_SET_UINT(0, "per-core", &aggr_mode,
1768 "aggregate counts per physical processor core", AGGR_CORE),
1769 OPT_UINTEGER('D', "delay", &initial_delay,
1770 "ms to wait before starting measurement after program start"),
1771 OPT_END()
1772 };
1773 const char * const stat_usage[] = {
1774 "perf stat [<options>] [<command>]",
1775 NULL
1776 };
1777 int status = -EINVAL, run_idx;
1778 const char *mode;
1779
1780 setlocale(LC_ALL, "");
1781
1782 evsel_list = perf_evlist__new();
1783 if (evsel_list == NULL)
1784 return -ENOMEM;
1785
1786 argc = parse_options(argc, argv, options, stat_usage,
1787 PARSE_OPT_STOP_AT_NON_OPTION);
1788
1789 output = stderr;
1790 if (output_name && strcmp(output_name, "-"))
1791 output = NULL;
1792
1793 if (output_name && output_fd) {
1794 fprintf(stderr, "cannot use both --output and --log-fd\n");
1795 parse_options_usage(stat_usage, options, "o", 1);
1796 parse_options_usage(NULL, options, "log-fd", 0);
1797 goto out;
1798 }
1799
1800 if (output_fd < 0) {
1801 fprintf(stderr, "argument to --log-fd must be a > 0\n");
1802 parse_options_usage(stat_usage, options, "log-fd", 0);
1803 goto out;
1804 }
1805
1806 if (!output) {
1807 struct timespec tm;
1808 mode = append_file ? "a" : "w";
1809
1810 output = fopen(output_name, mode);
1811 if (!output) {
1812 perror("failed to create output file");
1813 return -1;
1814 }
1815 clock_gettime(CLOCK_REALTIME, &tm);
1816 fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
1817 } else if (output_fd > 0) {
1818 mode = append_file ? "a" : "w";
1819 output = fdopen(output_fd, mode);
1820 if (!output) {
1821 perror("Failed opening logfd");
1822 return -errno;
1823 }
1824 }
1825
1826 if (csv_sep) {
1827 csv_output = true;
1828 if (!strcmp(csv_sep, "\\t"))
1829 csv_sep = "\t";
1830 } else
1831 csv_sep = DEFAULT_SEPARATOR;
1832
1833 /*
1834 * let the spreadsheet do the pretty-printing
1835 */
1836 if (csv_output) {
1837 /* User explicitly passed -B? */
1838 if (big_num_opt == 1) {
1839 fprintf(stderr, "-B option not supported with -x\n");
1840 parse_options_usage(stat_usage, options, "B", 1);
1841 parse_options_usage(NULL, options, "x", 1);
1842 goto out;
1843 } else /* Nope, so disable big number formatting */
1844 big_num = false;
1845 } else if (big_num_opt == 0) /* User passed --no-big-num */
1846 big_num = false;
1847
1848 if (!argc && target__none(&target))
1849 usage_with_options(stat_usage, options);
1850
1851 if (run_count < 0) {
1852 pr_err("Run count must be a positive number\n");
1853 parse_options_usage(stat_usage, options, "r", 1);
1854 goto out;
1855 } else if (run_count == 0) {
1856 forever = true;
1857 run_count = 1;
1858 }
1859
1860 /* no_aggr, cgroup are for system-wide only */
1861 if ((aggr_mode != AGGR_GLOBAL || nr_cgroups) &&
1862 !target__has_cpu(&target)) {
1863 fprintf(stderr, "both cgroup and no-aggregation "
1864 "modes only available in system-wide mode\n");
1865
1866 parse_options_usage(stat_usage, options, "G", 1);
1867 parse_options_usage(NULL, options, "A", 1);
1868 parse_options_usage(NULL, options, "a", 1);
1869 goto out;
1870 }
1871
1872 if (add_default_attributes())
1873 goto out;
1874
1875 target__validate(&target);
1876
1877 if (perf_evlist__create_maps(evsel_list, &target) < 0) {
1878 if (target__has_task(&target)) {
1879 pr_err("Problems finding threads of monitor\n");
1880 parse_options_usage(stat_usage, options, "p", 1);
1881 parse_options_usage(NULL, options, "t", 1);
1882 } else if (target__has_cpu(&target)) {
1883 perror("failed to parse CPUs map");
1884 parse_options_usage(stat_usage, options, "C", 1);
1885 parse_options_usage(NULL, options, "a", 1);
1886 }
1887 goto out;
1888 }
1889 if (interval && interval < 100) {
1890 pr_err("print interval must be >= 100ms\n");
1891 parse_options_usage(stat_usage, options, "I", 1);
1892 goto out;
1893 }
1894
1895 if (perf_evlist__alloc_stats(evsel_list, interval))
1896 goto out;
1897
1898 if (perf_stat_init_aggr_mode())
1899 goto out;
1900
1901 /*
1902 * We dont want to block the signals - that would cause
1903 * child tasks to inherit that and Ctrl-C would not work.
1904 * What we want is for Ctrl-C to work in the exec()-ed
1905 * task, but being ignored by perf stat itself:
1906 */
1907 atexit(sig_atexit);
1908 if (!forever)
1909 signal(SIGINT, skip_signal);
1910 signal(SIGCHLD, skip_signal);
1911 signal(SIGALRM, skip_signal);
1912 signal(SIGABRT, skip_signal);
1913
1914 status = 0;
1915 for (run_idx = 0; forever || run_idx < run_count; run_idx++) {
1916 if (run_count != 1 && verbose)
1917 fprintf(output, "[ perf stat: executing run #%d ... ]\n",
1918 run_idx + 1);
1919
1920 status = run_perf_stat(argc, argv);
1921 if (forever && status != -1) {
1922 print_stat(argc, argv);
1923 perf_stat__reset_stats(evsel_list);
1924 }
1925 }
1926
1927 if (!forever && status != -1 && !interval)
1928 print_stat(argc, argv);
1929
1930 perf_evlist__free_stats(evsel_list);
1931 out:
1932 perf_evlist__delete(evsel_list);
1933 return status;
1934 }
Linux with added FPC-III support