perf report: Add infrastructure for a cycles histogram
[deliverable/linux.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_counters(struct timespec *ts, int argc, const char **argv);
71
72 /* Default events used for perf stat -T */
73 static const char *transaction_attrs = {
74 "task-clock,"
75 "{"
76 "instructions,"
77 "cycles,"
78 "cpu/cycles-t/,"
79 "cpu/tx-start/,"
80 "cpu/el-start/,"
81 "cpu/cycles-ct/"
82 "}"
83 };
84
85 /* More limited version when the CPU does not have all events. */
86 static const char * transaction_limited_attrs = {
87 "task-clock,"
88 "{"
89 "instructions,"
90 "cycles,"
91 "cpu/cycles-t/,"
92 "cpu/tx-start/"
93 "}"
94 };
95
96 static struct perf_evlist *evsel_list;
97
98 static struct target target = {
99 .uid = UINT_MAX,
100 };
101
102 static int run_count = 1;
103 static bool no_inherit = false;
104 static volatile pid_t child_pid = -1;
105 static bool null_run = false;
106 static int detailed_run = 0;
107 static bool transaction_run;
108 static bool big_num = true;
109 static int big_num_opt = -1;
110 static const char *csv_sep = NULL;
111 static bool csv_output = false;
112 static bool group = false;
113 static const char *pre_cmd = NULL;
114 static const char *post_cmd = NULL;
115 static bool sync_run = false;
116 static unsigned int initial_delay = 0;
117 static unsigned int unit_width = 4; /* strlen("unit") */
118 static bool forever = false;
119 static struct timespec ref_time;
120 static struct cpu_map *aggr_map;
121 static int (*aggr_get_id)(struct cpu_map *m, int cpu);
122
123 static volatile int done = 0;
124
125 static struct perf_stat_config stat_config = {
126 .aggr_mode = AGGR_GLOBAL,
127 .scale = true,
128 };
129
130 static inline void diff_timespec(struct timespec *r, struct timespec *a,
131 struct timespec *b)
132 {
133 r->tv_sec = a->tv_sec - b->tv_sec;
134 if (a->tv_nsec < b->tv_nsec) {
135 r->tv_nsec = a->tv_nsec + 1000000000L - b->tv_nsec;
136 r->tv_sec--;
137 } else {
138 r->tv_nsec = a->tv_nsec - b->tv_nsec ;
139 }
140 }
141
142 static void perf_stat__reset_stats(void)
143 {
144 perf_evlist__reset_stats(evsel_list);
145 perf_stat__reset_shadow_stats();
146 }
147
148 static int create_perf_stat_counter(struct perf_evsel *evsel)
149 {
150 struct perf_event_attr *attr = &evsel->attr;
151
152 if (stat_config.scale)
153 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
154 PERF_FORMAT_TOTAL_TIME_RUNNING;
155
156 attr->inherit = !no_inherit;
157
158 if (target__has_cpu(&target))
159 return perf_evsel__open_per_cpu(evsel, perf_evsel__cpus(evsel));
160
161 if (!target__has_task(&target) && perf_evsel__is_group_leader(evsel)) {
162 attr->disabled = 1;
163 if (!initial_delay)
164 attr->enable_on_exec = 1;
165 }
166
167 return perf_evsel__open_per_thread(evsel, evsel_list->threads);
168 }
169
170 /*
171 * Does the counter have nsecs as a unit?
172 */
173 static inline int nsec_counter(struct perf_evsel *evsel)
174 {
175 if (perf_evsel__match(evsel, SOFTWARE, SW_CPU_CLOCK) ||
176 perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
177 return 1;
178
179 return 0;
180 }
181
182 /*
183 * Read out the results of a single counter:
184 * do not aggregate counts across CPUs in system-wide mode
185 */
186 static int read_counter(struct perf_evsel *counter)
187 {
188 int nthreads = thread_map__nr(evsel_list->threads);
189 int ncpus = perf_evsel__nr_cpus(counter);
190 int cpu, thread;
191
192 if (!counter->supported)
193 return -ENOENT;
194
195 if (counter->system_wide)
196 nthreads = 1;
197
198 for (thread = 0; thread < nthreads; thread++) {
199 for (cpu = 0; cpu < ncpus; cpu++) {
200 struct perf_counts_values *count;
201
202 count = perf_counts(counter->counts, cpu, thread);
203 if (perf_evsel__read(counter, cpu, thread, count))
204 return -1;
205 }
206 }
207
208 return 0;
209 }
210
211 static void read_counters(bool close_counters)
212 {
213 struct perf_evsel *counter;
214
215 evlist__for_each(evsel_list, counter) {
216 if (read_counter(counter))
217 pr_warning("failed to read counter %s\n", counter->name);
218
219 if (perf_stat_process_counter(&stat_config, counter))
220 pr_warning("failed to process counter %s\n", counter->name);
221
222 if (close_counters) {
223 perf_evsel__close_fd(counter, perf_evsel__nr_cpus(counter),
224 thread_map__nr(evsel_list->threads));
225 }
226 }
227 }
228
229 static void process_interval(void)
230 {
231 struct timespec ts, rs;
232
233 read_counters(false);
234
235 clock_gettime(CLOCK_MONOTONIC, &ts);
236 diff_timespec(&rs, &ts, &ref_time);
237
238 print_counters(&rs, 0, NULL);
239 }
240
241 static void handle_initial_delay(void)
242 {
243 struct perf_evsel *counter;
244
245 if (initial_delay) {
246 const int ncpus = cpu_map__nr(evsel_list->cpus),
247 nthreads = thread_map__nr(evsel_list->threads);
248
249 usleep(initial_delay * 1000);
250 evlist__for_each(evsel_list, counter)
251 perf_evsel__enable(counter, ncpus, nthreads);
252 }
253 }
254
255 static volatile int workload_exec_errno;
256
257 /*
258 * perf_evlist__prepare_workload will send a SIGUSR1
259 * if the fork fails, since we asked by setting its
260 * want_signal to true.
261 */
262 static void workload_exec_failed_signal(int signo __maybe_unused, siginfo_t *info,
263 void *ucontext __maybe_unused)
264 {
265 workload_exec_errno = info->si_value.sival_int;
266 }
267
268 static int __run_perf_stat(int argc, const char **argv)
269 {
270 int interval = stat_config.interval;
271 char msg[512];
272 unsigned long long t0, t1;
273 struct perf_evsel *counter;
274 struct timespec ts;
275 size_t l;
276 int status = 0;
277 const bool forks = (argc > 0);
278
279 if (interval) {
280 ts.tv_sec = interval / 1000;
281 ts.tv_nsec = (interval % 1000) * 1000000;
282 } else {
283 ts.tv_sec = 1;
284 ts.tv_nsec = 0;
285 }
286
287 if (forks) {
288 if (perf_evlist__prepare_workload(evsel_list, &target, argv, false,
289 workload_exec_failed_signal) < 0) {
290 perror("failed to prepare workload");
291 return -1;
292 }
293 child_pid = evsel_list->workload.pid;
294 }
295
296 if (group)
297 perf_evlist__set_leader(evsel_list);
298
299 evlist__for_each(evsel_list, counter) {
300 if (create_perf_stat_counter(counter) < 0) {
301 /*
302 * PPC returns ENXIO for HW counters until 2.6.37
303 * (behavior changed with commit b0a873e).
304 */
305 if (errno == EINVAL || errno == ENOSYS ||
306 errno == ENOENT || errno == EOPNOTSUPP ||
307 errno == ENXIO) {
308 if (verbose)
309 ui__warning("%s event is not supported by the kernel.\n",
310 perf_evsel__name(counter));
311 counter->supported = false;
312
313 if ((counter->leader != counter) ||
314 !(counter->leader->nr_members > 1))
315 continue;
316 }
317
318 perf_evsel__open_strerror(counter, &target,
319 errno, msg, sizeof(msg));
320 ui__error("%s\n", msg);
321
322 if (child_pid != -1)
323 kill(child_pid, SIGTERM);
324
325 return -1;
326 }
327 counter->supported = true;
328
329 l = strlen(counter->unit);
330 if (l > unit_width)
331 unit_width = l;
332 }
333
334 if (perf_evlist__apply_filters(evsel_list, &counter)) {
335 error("failed to set filter \"%s\" on event %s with %d (%s)\n",
336 counter->filter, perf_evsel__name(counter), errno,
337 strerror_r(errno, msg, sizeof(msg)));
338 return -1;
339 }
340
341 /*
342 * Enable counters and exec the command:
343 */
344 t0 = rdclock();
345 clock_gettime(CLOCK_MONOTONIC, &ref_time);
346
347 if (forks) {
348 perf_evlist__start_workload(evsel_list);
349 handle_initial_delay();
350
351 if (interval) {
352 while (!waitpid(child_pid, &status, WNOHANG)) {
353 nanosleep(&ts, NULL);
354 process_interval();
355 }
356 }
357 wait(&status);
358
359 if (workload_exec_errno) {
360 const char *emsg = strerror_r(workload_exec_errno, msg, sizeof(msg));
361 pr_err("Workload failed: %s\n", emsg);
362 return -1;
363 }
364
365 if (WIFSIGNALED(status))
366 psignal(WTERMSIG(status), argv[0]);
367 } else {
368 handle_initial_delay();
369 while (!done) {
370 nanosleep(&ts, NULL);
371 if (interval)
372 process_interval();
373 }
374 }
375
376 t1 = rdclock();
377
378 update_stats(&walltime_nsecs_stats, t1 - t0);
379
380 read_counters(true);
381
382 return WEXITSTATUS(status);
383 }
384
385 static int run_perf_stat(int argc, const char **argv)
386 {
387 int ret;
388
389 if (pre_cmd) {
390 ret = system(pre_cmd);
391 if (ret)
392 return ret;
393 }
394
395 if (sync_run)
396 sync();
397
398 ret = __run_perf_stat(argc, argv);
399 if (ret)
400 return ret;
401
402 if (post_cmd) {
403 ret = system(post_cmd);
404 if (ret)
405 return ret;
406 }
407
408 return ret;
409 }
410
411 static void print_running(u64 run, u64 ena)
412 {
413 if (csv_output) {
414 fprintf(stat_config.output, "%s%" PRIu64 "%s%.2f",
415 csv_sep,
416 run,
417 csv_sep,
418 ena ? 100.0 * run / ena : 100.0);
419 } else if (run != ena) {
420 fprintf(stat_config.output, " (%.2f%%)", 100.0 * run / ena);
421 }
422 }
423
424 static void print_noise_pct(double total, double avg)
425 {
426 double pct = rel_stddev_stats(total, avg);
427
428 if (csv_output)
429 fprintf(stat_config.output, "%s%.2f%%", csv_sep, pct);
430 else if (pct)
431 fprintf(stat_config.output, " ( +-%6.2f%% )", pct);
432 }
433
434 static void print_noise(struct perf_evsel *evsel, double avg)
435 {
436 struct perf_stat *ps;
437
438 if (run_count == 1)
439 return;
440
441 ps = evsel->priv;
442 print_noise_pct(stddev_stats(&ps->res_stats[0]), avg);
443 }
444
445 static void aggr_printout(struct perf_evsel *evsel, int id, int nr)
446 {
447 switch (stat_config.aggr_mode) {
448 case AGGR_CORE:
449 fprintf(stat_config.output, "S%d-C%*d%s%*d%s",
450 cpu_map__id_to_socket(id),
451 csv_output ? 0 : -8,
452 cpu_map__id_to_cpu(id),
453 csv_sep,
454 csv_output ? 0 : 4,
455 nr,
456 csv_sep);
457 break;
458 case AGGR_SOCKET:
459 fprintf(stat_config.output, "S%*d%s%*d%s",
460 csv_output ? 0 : -5,
461 id,
462 csv_sep,
463 csv_output ? 0 : 4,
464 nr,
465 csv_sep);
466 break;
467 case AGGR_NONE:
468 fprintf(stat_config.output, "CPU%*d%s",
469 csv_output ? 0 : -4,
470 perf_evsel__cpus(evsel)->map[id], csv_sep);
471 break;
472 case AGGR_THREAD:
473 fprintf(stat_config.output, "%*s-%*d%s",
474 csv_output ? 0 : 16,
475 thread_map__comm(evsel->threads, id),
476 csv_output ? 0 : -8,
477 thread_map__pid(evsel->threads, id),
478 csv_sep);
479 break;
480 case AGGR_GLOBAL:
481 default:
482 break;
483 }
484 }
485
486 static void nsec_printout(int id, int nr, struct perf_evsel *evsel, double avg)
487 {
488 FILE *output = stat_config.output;
489 double msecs = avg / 1e6;
490 const char *fmt_v, *fmt_n;
491 char name[25];
492
493 fmt_v = csv_output ? "%.6f%s" : "%18.6f%s";
494 fmt_n = csv_output ? "%s" : "%-25s";
495
496 aggr_printout(evsel, id, nr);
497
498 scnprintf(name, sizeof(name), "%s%s",
499 perf_evsel__name(evsel), csv_output ? "" : " (msec)");
500
501 fprintf(output, fmt_v, msecs, csv_sep);
502
503 if (csv_output)
504 fprintf(output, "%s%s", evsel->unit, csv_sep);
505 else
506 fprintf(output, "%-*s%s", unit_width, evsel->unit, csv_sep);
507
508 fprintf(output, fmt_n, name);
509
510 if (evsel->cgrp)
511 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
512
513 if (csv_output || stat_config.interval)
514 return;
515
516 if (perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
517 fprintf(output, " # %8.3f CPUs utilized ",
518 avg / avg_stats(&walltime_nsecs_stats));
519 else
520 fprintf(output, " ");
521 }
522
523 static void abs_printout(int id, int nr, struct perf_evsel *evsel, double avg)
524 {
525 FILE *output = stat_config.output;
526 double sc = evsel->scale;
527 const char *fmt;
528 int cpu = cpu_map__id_to_cpu(id);
529
530 if (csv_output) {
531 fmt = sc != 1.0 ? "%.2f%s" : "%.0f%s";
532 } else {
533 if (big_num)
534 fmt = sc != 1.0 ? "%'18.2f%s" : "%'18.0f%s";
535 else
536 fmt = sc != 1.0 ? "%18.2f%s" : "%18.0f%s";
537 }
538
539 aggr_printout(evsel, id, nr);
540
541 if (stat_config.aggr_mode == AGGR_GLOBAL)
542 cpu = 0;
543
544 fprintf(output, fmt, avg, csv_sep);
545
546 if (evsel->unit)
547 fprintf(output, "%-*s%s",
548 csv_output ? 0 : unit_width,
549 evsel->unit, csv_sep);
550
551 fprintf(output, "%-*s", csv_output ? 0 : 25, perf_evsel__name(evsel));
552
553 if (evsel->cgrp)
554 fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
555
556 if (csv_output || stat_config.interval)
557 return;
558
559 perf_stat__print_shadow_stats(output, evsel, avg, cpu,
560 stat_config.aggr_mode);
561 }
562
563 static void print_aggr(char *prefix)
564 {
565 FILE *output = stat_config.output;
566 struct perf_evsel *counter;
567 int cpu, cpu2, s, s2, id, nr;
568 double uval;
569 u64 ena, run, val;
570
571 if (!(aggr_map || aggr_get_id))
572 return;
573
574 for (s = 0; s < aggr_map->nr; s++) {
575 id = aggr_map->map[s];
576 evlist__for_each(evsel_list, counter) {
577 val = ena = run = 0;
578 nr = 0;
579 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
580 cpu2 = perf_evsel__cpus(counter)->map[cpu];
581 s2 = aggr_get_id(evsel_list->cpus, cpu2);
582 if (s2 != id)
583 continue;
584 val += perf_counts(counter->counts, cpu, 0)->val;
585 ena += perf_counts(counter->counts, cpu, 0)->ena;
586 run += perf_counts(counter->counts, cpu, 0)->run;
587 nr++;
588 }
589 if (prefix)
590 fprintf(output, "%s", prefix);
591
592 if (run == 0 || ena == 0) {
593 aggr_printout(counter, id, nr);
594
595 fprintf(output, "%*s%s",
596 csv_output ? 0 : 18,
597 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
598 csv_sep);
599
600 fprintf(output, "%-*s%s",
601 csv_output ? 0 : unit_width,
602 counter->unit, csv_sep);
603
604 fprintf(output, "%*s",
605 csv_output ? 0 : -25,
606 perf_evsel__name(counter));
607
608 if (counter->cgrp)
609 fprintf(output, "%s%s",
610 csv_sep, counter->cgrp->name);
611
612 print_running(run, ena);
613 fputc('\n', output);
614 continue;
615 }
616 uval = val * counter->scale;
617
618 if (nsec_counter(counter))
619 nsec_printout(id, nr, counter, uval);
620 else
621 abs_printout(id, nr, counter, uval);
622
623 if (!csv_output)
624 print_noise(counter, 1.0);
625
626 print_running(run, ena);
627 fputc('\n', output);
628 }
629 }
630 }
631
632 static void print_aggr_thread(struct perf_evsel *counter, char *prefix)
633 {
634 FILE *output = stat_config.output;
635 int nthreads = thread_map__nr(counter->threads);
636 int ncpus = cpu_map__nr(counter->cpus);
637 int cpu, thread;
638 double uval;
639
640 for (thread = 0; thread < nthreads; thread++) {
641 u64 ena = 0, run = 0, val = 0;
642
643 for (cpu = 0; cpu < ncpus; cpu++) {
644 val += perf_counts(counter->counts, cpu, thread)->val;
645 ena += perf_counts(counter->counts, cpu, thread)->ena;
646 run += perf_counts(counter->counts, cpu, thread)->run;
647 }
648
649 if (prefix)
650 fprintf(output, "%s", prefix);
651
652 uval = val * counter->scale;
653
654 if (nsec_counter(counter))
655 nsec_printout(thread, 0, counter, uval);
656 else
657 abs_printout(thread, 0, counter, uval);
658
659 if (!csv_output)
660 print_noise(counter, 1.0);
661
662 print_running(run, ena);
663 fputc('\n', output);
664 }
665 }
666
667 /*
668 * Print out the results of a single counter:
669 * aggregated counts in system-wide mode
670 */
671 static void print_counter_aggr(struct perf_evsel *counter, char *prefix)
672 {
673 FILE *output = stat_config.output;
674 struct perf_stat *ps = counter->priv;
675 double avg = avg_stats(&ps->res_stats[0]);
676 int scaled = counter->counts->scaled;
677 double uval;
678 double avg_enabled, avg_running;
679
680 avg_enabled = avg_stats(&ps->res_stats[1]);
681 avg_running = avg_stats(&ps->res_stats[2]);
682
683 if (prefix)
684 fprintf(output, "%s", prefix);
685
686 if (scaled == -1 || !counter->supported) {
687 fprintf(output, "%*s%s",
688 csv_output ? 0 : 18,
689 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
690 csv_sep);
691 fprintf(output, "%-*s%s",
692 csv_output ? 0 : unit_width,
693 counter->unit, csv_sep);
694 fprintf(output, "%*s",
695 csv_output ? 0 : -25,
696 perf_evsel__name(counter));
697
698 if (counter->cgrp)
699 fprintf(output, "%s%s", csv_sep, counter->cgrp->name);
700
701 print_running(avg_running, avg_enabled);
702 fputc('\n', output);
703 return;
704 }
705
706 uval = avg * counter->scale;
707
708 if (nsec_counter(counter))
709 nsec_printout(-1, 0, counter, uval);
710 else
711 abs_printout(-1, 0, counter, uval);
712
713 print_noise(counter, avg);
714
715 print_running(avg_running, avg_enabled);
716 fprintf(output, "\n");
717 }
718
719 /*
720 * Print out the results of a single counter:
721 * does not use aggregated count in system-wide
722 */
723 static void print_counter(struct perf_evsel *counter, char *prefix)
724 {
725 FILE *output = stat_config.output;
726 u64 ena, run, val;
727 double uval;
728 int cpu;
729
730 for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
731 val = perf_counts(counter->counts, cpu, 0)->val;
732 ena = perf_counts(counter->counts, cpu, 0)->ena;
733 run = perf_counts(counter->counts, cpu, 0)->run;
734
735 if (prefix)
736 fprintf(output, "%s", prefix);
737
738 if (run == 0 || ena == 0) {
739 fprintf(output, "CPU%*d%s%*s%s",
740 csv_output ? 0 : -4,
741 perf_evsel__cpus(counter)->map[cpu], csv_sep,
742 csv_output ? 0 : 18,
743 counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
744 csv_sep);
745
746 fprintf(output, "%-*s%s",
747 csv_output ? 0 : unit_width,
748 counter->unit, csv_sep);
749
750 fprintf(output, "%*s",
751 csv_output ? 0 : -25,
752 perf_evsel__name(counter));
753
754 if (counter->cgrp)
755 fprintf(output, "%s%s",
756 csv_sep, counter->cgrp->name);
757
758 print_running(run, ena);
759 fputc('\n', output);
760 continue;
761 }
762
763 uval = val * counter->scale;
764
765 if (nsec_counter(counter))
766 nsec_printout(cpu, 0, counter, uval);
767 else
768 abs_printout(cpu, 0, counter, uval);
769
770 if (!csv_output)
771 print_noise(counter, 1.0);
772 print_running(run, ena);
773
774 fputc('\n', output);
775 }
776 }
777
778 static void print_interval(char *prefix, struct timespec *ts)
779 {
780 FILE *output = stat_config.output;
781 static int num_print_interval;
782
783 sprintf(prefix, "%6lu.%09lu%s", ts->tv_sec, ts->tv_nsec, csv_sep);
784
785 if (num_print_interval == 0 && !csv_output) {
786 switch (stat_config.aggr_mode) {
787 case AGGR_SOCKET:
788 fprintf(output, "# time socket cpus counts %*s events\n", unit_width, "unit");
789 break;
790 case AGGR_CORE:
791 fprintf(output, "# time core cpus counts %*s events\n", unit_width, "unit");
792 break;
793 case AGGR_NONE:
794 fprintf(output, "# time CPU counts %*s events\n", unit_width, "unit");
795 break;
796 case AGGR_THREAD:
797 fprintf(output, "# time comm-pid counts %*s events\n", unit_width, "unit");
798 break;
799 case AGGR_GLOBAL:
800 default:
801 fprintf(output, "# time counts %*s events\n", unit_width, "unit");
802 }
803 }
804
805 if (++num_print_interval == 25)
806 num_print_interval = 0;
807 }
808
809 static void print_header(int argc, const char **argv)
810 {
811 FILE *output = stat_config.output;
812 int i;
813
814 fflush(stdout);
815
816 if (!csv_output) {
817 fprintf(output, "\n");
818 fprintf(output, " Performance counter stats for ");
819 if (target.system_wide)
820 fprintf(output, "\'system wide");
821 else if (target.cpu_list)
822 fprintf(output, "\'CPU(s) %s", target.cpu_list);
823 else if (!target__has_task(&target)) {
824 fprintf(output, "\'%s", argv[0]);
825 for (i = 1; i < argc; i++)
826 fprintf(output, " %s", argv[i]);
827 } else if (target.pid)
828 fprintf(output, "process id \'%s", target.pid);
829 else
830 fprintf(output, "thread id \'%s", target.tid);
831
832 fprintf(output, "\'");
833 if (run_count > 1)
834 fprintf(output, " (%d runs)", run_count);
835 fprintf(output, ":\n\n");
836 }
837 }
838
839 static void print_footer(void)
840 {
841 FILE *output = stat_config.output;
842
843 if (!null_run)
844 fprintf(output, "\n");
845 fprintf(output, " %17.9f seconds time elapsed",
846 avg_stats(&walltime_nsecs_stats)/1e9);
847 if (run_count > 1) {
848 fprintf(output, " ");
849 print_noise_pct(stddev_stats(&walltime_nsecs_stats),
850 avg_stats(&walltime_nsecs_stats));
851 }
852 fprintf(output, "\n\n");
853 }
854
855 static void print_counters(struct timespec *ts, int argc, const char **argv)
856 {
857 int interval = stat_config.interval;
858 struct perf_evsel *counter;
859 char buf[64], *prefix = NULL;
860
861 if (interval)
862 print_interval(prefix = buf, ts);
863 else
864 print_header(argc, argv);
865
866 switch (stat_config.aggr_mode) {
867 case AGGR_CORE:
868 case AGGR_SOCKET:
869 print_aggr(prefix);
870 break;
871 case AGGR_THREAD:
872 evlist__for_each(evsel_list, counter)
873 print_aggr_thread(counter, prefix);
874 break;
875 case AGGR_GLOBAL:
876 evlist__for_each(evsel_list, counter)
877 print_counter_aggr(counter, prefix);
878 break;
879 case AGGR_NONE:
880 evlist__for_each(evsel_list, counter)
881 print_counter(counter, prefix);
882 break;
883 default:
884 break;
885 }
886
887 if (!interval && !csv_output)
888 print_footer();
889
890 fflush(stat_config.output);
891 }
892
893 static volatile int signr = -1;
894
895 static void skip_signal(int signo)
896 {
897 if ((child_pid == -1) || stat_config.interval)
898 done = 1;
899
900 signr = signo;
901 /*
902 * render child_pid harmless
903 * won't send SIGTERM to a random
904 * process in case of race condition
905 * and fast PID recycling
906 */
907 child_pid = -1;
908 }
909
910 static void sig_atexit(void)
911 {
912 sigset_t set, oset;
913
914 /*
915 * avoid race condition with SIGCHLD handler
916 * in skip_signal() which is modifying child_pid
917 * goal is to avoid send SIGTERM to a random
918 * process
919 */
920 sigemptyset(&set);
921 sigaddset(&set, SIGCHLD);
922 sigprocmask(SIG_BLOCK, &set, &oset);
923
924 if (child_pid != -1)
925 kill(child_pid, SIGTERM);
926
927 sigprocmask(SIG_SETMASK, &oset, NULL);
928
929 if (signr == -1)
930 return;
931
932 signal(signr, SIG_DFL);
933 kill(getpid(), signr);
934 }
935
936 static int stat__set_big_num(const struct option *opt __maybe_unused,
937 const char *s __maybe_unused, int unset)
938 {
939 big_num_opt = unset ? 0 : 1;
940 return 0;
941 }
942
943 static int perf_stat_init_aggr_mode(void)
944 {
945 switch (stat_config.aggr_mode) {
946 case AGGR_SOCKET:
947 if (cpu_map__build_socket_map(evsel_list->cpus, &aggr_map)) {
948 perror("cannot build socket map");
949 return -1;
950 }
951 aggr_get_id = cpu_map__get_socket;
952 break;
953 case AGGR_CORE:
954 if (cpu_map__build_core_map(evsel_list->cpus, &aggr_map)) {
955 perror("cannot build core map");
956 return -1;
957 }
958 aggr_get_id = cpu_map__get_core;
959 break;
960 case AGGR_NONE:
961 case AGGR_GLOBAL:
962 case AGGR_THREAD:
963 default:
964 break;
965 }
966 return 0;
967 }
968
969 /*
970 * Add default attributes, if there were no attributes specified or
971 * if -d/--detailed, -d -d or -d -d -d is used:
972 */
973 static int add_default_attributes(void)
974 {
975 struct perf_event_attr default_attrs[] = {
976
977 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK },
978 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES },
979 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS },
980 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS },
981
982 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES },
983 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND },
984 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND },
985 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS },
986 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS },
987 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES },
988
989 };
990
991 /*
992 * Detailed stats (-d), covering the L1 and last level data caches:
993 */
994 struct perf_event_attr detailed_attrs[] = {
995
996 { .type = PERF_TYPE_HW_CACHE,
997 .config =
998 PERF_COUNT_HW_CACHE_L1D << 0 |
999 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1000 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1001
1002 { .type = PERF_TYPE_HW_CACHE,
1003 .config =
1004 PERF_COUNT_HW_CACHE_L1D << 0 |
1005 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1006 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1007
1008 { .type = PERF_TYPE_HW_CACHE,
1009 .config =
1010 PERF_COUNT_HW_CACHE_LL << 0 |
1011 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1012 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1013
1014 { .type = PERF_TYPE_HW_CACHE,
1015 .config =
1016 PERF_COUNT_HW_CACHE_LL << 0 |
1017 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1018 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1019 };
1020
1021 /*
1022 * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
1023 */
1024 struct perf_event_attr very_detailed_attrs[] = {
1025
1026 { .type = PERF_TYPE_HW_CACHE,
1027 .config =
1028 PERF_COUNT_HW_CACHE_L1I << 0 |
1029 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1030 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1031
1032 { .type = PERF_TYPE_HW_CACHE,
1033 .config =
1034 PERF_COUNT_HW_CACHE_L1I << 0 |
1035 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1036 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1037
1038 { .type = PERF_TYPE_HW_CACHE,
1039 .config =
1040 PERF_COUNT_HW_CACHE_DTLB << 0 |
1041 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1042 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1043
1044 { .type = PERF_TYPE_HW_CACHE,
1045 .config =
1046 PERF_COUNT_HW_CACHE_DTLB << 0 |
1047 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1048 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1049
1050 { .type = PERF_TYPE_HW_CACHE,
1051 .config =
1052 PERF_COUNT_HW_CACHE_ITLB << 0 |
1053 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1054 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1055
1056 { .type = PERF_TYPE_HW_CACHE,
1057 .config =
1058 PERF_COUNT_HW_CACHE_ITLB << 0 |
1059 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
1060 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1061
1062 };
1063
1064 /*
1065 * Very, very detailed stats (-d -d -d), adding prefetch events:
1066 */
1067 struct perf_event_attr very_very_detailed_attrs[] = {
1068
1069 { .type = PERF_TYPE_HW_CACHE,
1070 .config =
1071 PERF_COUNT_HW_CACHE_L1D << 0 |
1072 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
1073 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
1074
1075 { .type = PERF_TYPE_HW_CACHE,
1076 .config =
1077 PERF_COUNT_HW_CACHE_L1D << 0 |
1078 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
1079 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
1080 };
1081
1082 /* Set attrs if no event is selected and !null_run: */
1083 if (null_run)
1084 return 0;
1085
1086 if (transaction_run) {
1087 int err;
1088 if (pmu_have_event("cpu", "cycles-ct") &&
1089 pmu_have_event("cpu", "el-start"))
1090 err = parse_events(evsel_list, transaction_attrs, NULL);
1091 else
1092 err = parse_events(evsel_list, transaction_limited_attrs, NULL);
1093 if (err) {
1094 fprintf(stderr, "Cannot set up transaction events\n");
1095 return -1;
1096 }
1097 return 0;
1098 }
1099
1100 if (!evsel_list->nr_entries) {
1101 if (perf_evlist__add_default_attrs(evsel_list, default_attrs) < 0)
1102 return -1;
1103 }
1104
1105 /* Detailed events get appended to the event list: */
1106
1107 if (detailed_run < 1)
1108 return 0;
1109
1110 /* Append detailed run extra attributes: */
1111 if (perf_evlist__add_default_attrs(evsel_list, detailed_attrs) < 0)
1112 return -1;
1113
1114 if (detailed_run < 2)
1115 return 0;
1116
1117 /* Append very detailed run extra attributes: */
1118 if (perf_evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0)
1119 return -1;
1120
1121 if (detailed_run < 3)
1122 return 0;
1123
1124 /* Append very, very detailed run extra attributes: */
1125 return perf_evlist__add_default_attrs(evsel_list, very_very_detailed_attrs);
1126 }
1127
1128 int cmd_stat(int argc, const char **argv, const char *prefix __maybe_unused)
1129 {
1130 bool append_file = false;
1131 int output_fd = 0;
1132 const char *output_name = NULL;
1133 const struct option options[] = {
1134 OPT_BOOLEAN('T', "transaction", &transaction_run,
1135 "hardware transaction statistics"),
1136 OPT_CALLBACK('e', "event", &evsel_list, "event",
1137 "event selector. use 'perf list' to list available events",
1138 parse_events_option),
1139 OPT_CALLBACK(0, "filter", &evsel_list, "filter",
1140 "event filter", parse_filter),
1141 OPT_BOOLEAN('i', "no-inherit", &no_inherit,
1142 "child tasks do not inherit counters"),
1143 OPT_STRING('p', "pid", &target.pid, "pid",
1144 "stat events on existing process id"),
1145 OPT_STRING('t', "tid", &target.tid, "tid",
1146 "stat events on existing thread id"),
1147 OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
1148 "system-wide collection from all CPUs"),
1149 OPT_BOOLEAN('g', "group", &group,
1150 "put the counters into a counter group"),
1151 OPT_BOOLEAN('c', "scale", &stat_config.scale, "scale/normalize counters"),
1152 OPT_INCR('v', "verbose", &verbose,
1153 "be more verbose (show counter open errors, etc)"),
1154 OPT_INTEGER('r', "repeat", &run_count,
1155 "repeat command and print average + stddev (max: 100, forever: 0)"),
1156 OPT_BOOLEAN('n', "null", &null_run,
1157 "null run - dont start any counters"),
1158 OPT_INCR('d', "detailed", &detailed_run,
1159 "detailed run - start a lot of events"),
1160 OPT_BOOLEAN('S', "sync", &sync_run,
1161 "call sync() before starting a run"),
1162 OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
1163 "print large numbers with thousands\' separators",
1164 stat__set_big_num),
1165 OPT_STRING('C', "cpu", &target.cpu_list, "cpu",
1166 "list of cpus to monitor in system-wide"),
1167 OPT_SET_UINT('A', "no-aggr", &stat_config.aggr_mode,
1168 "disable CPU count aggregation", AGGR_NONE),
1169 OPT_STRING('x', "field-separator", &csv_sep, "separator",
1170 "print counts with custom separator"),
1171 OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
1172 "monitor event in cgroup name only", parse_cgroups),
1173 OPT_STRING('o', "output", &output_name, "file", "output file name"),
1174 OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
1175 OPT_INTEGER(0, "log-fd", &output_fd,
1176 "log output to fd, instead of stderr"),
1177 OPT_STRING(0, "pre", &pre_cmd, "command",
1178 "command to run prior to the measured command"),
1179 OPT_STRING(0, "post", &post_cmd, "command",
1180 "command to run after to the measured command"),
1181 OPT_UINTEGER('I', "interval-print", &stat_config.interval,
1182 "print counts at regular interval in ms (>= 100)"),
1183 OPT_SET_UINT(0, "per-socket", &stat_config.aggr_mode,
1184 "aggregate counts per processor socket", AGGR_SOCKET),
1185 OPT_SET_UINT(0, "per-core", &stat_config.aggr_mode,
1186 "aggregate counts per physical processor core", AGGR_CORE),
1187 OPT_SET_UINT(0, "per-thread", &stat_config.aggr_mode,
1188 "aggregate counts per thread", AGGR_THREAD),
1189 OPT_UINTEGER('D', "delay", &initial_delay,
1190 "ms to wait before starting measurement after program start"),
1191 OPT_END()
1192 };
1193 const char * const stat_usage[] = {
1194 "perf stat [<options>] [<command>]",
1195 NULL
1196 };
1197 int status = -EINVAL, run_idx;
1198 const char *mode;
1199 FILE *output = stderr;
1200 unsigned int interval;
1201
1202 setlocale(LC_ALL, "");
1203
1204 evsel_list = perf_evlist__new();
1205 if (evsel_list == NULL)
1206 return -ENOMEM;
1207
1208 argc = parse_options(argc, argv, options, stat_usage,
1209 PARSE_OPT_STOP_AT_NON_OPTION);
1210
1211 interval = stat_config.interval;
1212
1213 if (output_name && strcmp(output_name, "-"))
1214 output = NULL;
1215
1216 if (output_name && output_fd) {
1217 fprintf(stderr, "cannot use both --output and --log-fd\n");
1218 parse_options_usage(stat_usage, options, "o", 1);
1219 parse_options_usage(NULL, options, "log-fd", 0);
1220 goto out;
1221 }
1222
1223 if (output_fd < 0) {
1224 fprintf(stderr, "argument to --log-fd must be a > 0\n");
1225 parse_options_usage(stat_usage, options, "log-fd", 0);
1226 goto out;
1227 }
1228
1229 if (!output) {
1230 struct timespec tm;
1231 mode = append_file ? "a" : "w";
1232
1233 output = fopen(output_name, mode);
1234 if (!output) {
1235 perror("failed to create output file");
1236 return -1;
1237 }
1238 clock_gettime(CLOCK_REALTIME, &tm);
1239 fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
1240 } else if (output_fd > 0) {
1241 mode = append_file ? "a" : "w";
1242 output = fdopen(output_fd, mode);
1243 if (!output) {
1244 perror("Failed opening logfd");
1245 return -errno;
1246 }
1247 }
1248
1249 stat_config.output = output;
1250
1251 if (csv_sep) {
1252 csv_output = true;
1253 if (!strcmp(csv_sep, "\\t"))
1254 csv_sep = "\t";
1255 } else
1256 csv_sep = DEFAULT_SEPARATOR;
1257
1258 /*
1259 * let the spreadsheet do the pretty-printing
1260 */
1261 if (csv_output) {
1262 /* User explicitly passed -B? */
1263 if (big_num_opt == 1) {
1264 fprintf(stderr, "-B option not supported with -x\n");
1265 parse_options_usage(stat_usage, options, "B", 1);
1266 parse_options_usage(NULL, options, "x", 1);
1267 goto out;
1268 } else /* Nope, so disable big number formatting */
1269 big_num = false;
1270 } else if (big_num_opt == 0) /* User passed --no-big-num */
1271 big_num = false;
1272
1273 if (!argc && target__none(&target))
1274 usage_with_options(stat_usage, options);
1275
1276 if (run_count < 0) {
1277 pr_err("Run count must be a positive number\n");
1278 parse_options_usage(stat_usage, options, "r", 1);
1279 goto out;
1280 } else if (run_count == 0) {
1281 forever = true;
1282 run_count = 1;
1283 }
1284
1285 if ((stat_config.aggr_mode == AGGR_THREAD) && !target__has_task(&target)) {
1286 fprintf(stderr, "The --per-thread option is only available "
1287 "when monitoring via -p -t options.\n");
1288 parse_options_usage(NULL, options, "p", 1);
1289 parse_options_usage(NULL, options, "t", 1);
1290 goto out;
1291 }
1292
1293 /*
1294 * no_aggr, cgroup are for system-wide only
1295 * --per-thread is aggregated per thread, we dont mix it with cpu mode
1296 */
1297 if (((stat_config.aggr_mode != AGGR_GLOBAL &&
1298 stat_config.aggr_mode != AGGR_THREAD) || nr_cgroups) &&
1299 !target__has_cpu(&target)) {
1300 fprintf(stderr, "both cgroup and no-aggregation "
1301 "modes only available in system-wide mode\n");
1302
1303 parse_options_usage(stat_usage, options, "G", 1);
1304 parse_options_usage(NULL, options, "A", 1);
1305 parse_options_usage(NULL, options, "a", 1);
1306 goto out;
1307 }
1308
1309 if (add_default_attributes())
1310 goto out;
1311
1312 target__validate(&target);
1313
1314 if (perf_evlist__create_maps(evsel_list, &target) < 0) {
1315 if (target__has_task(&target)) {
1316 pr_err("Problems finding threads of monitor\n");
1317 parse_options_usage(stat_usage, options, "p", 1);
1318 parse_options_usage(NULL, options, "t", 1);
1319 } else if (target__has_cpu(&target)) {
1320 perror("failed to parse CPUs map");
1321 parse_options_usage(stat_usage, options, "C", 1);
1322 parse_options_usage(NULL, options, "a", 1);
1323 }
1324 goto out;
1325 }
1326
1327 /*
1328 * Initialize thread_map with comm names,
1329 * so we could print it out on output.
1330 */
1331 if (stat_config.aggr_mode == AGGR_THREAD)
1332 thread_map__read_comms(evsel_list->threads);
1333
1334 if (interval && interval < 100) {
1335 pr_err("print interval must be >= 100ms\n");
1336 parse_options_usage(stat_usage, options, "I", 1);
1337 goto out;
1338 }
1339
1340 if (perf_evlist__alloc_stats(evsel_list, interval))
1341 goto out;
1342
1343 if (perf_stat_init_aggr_mode())
1344 goto out;
1345
1346 /*
1347 * We dont want to block the signals - that would cause
1348 * child tasks to inherit that and Ctrl-C would not work.
1349 * What we want is for Ctrl-C to work in the exec()-ed
1350 * task, but being ignored by perf stat itself:
1351 */
1352 atexit(sig_atexit);
1353 if (!forever)
1354 signal(SIGINT, skip_signal);
1355 signal(SIGCHLD, skip_signal);
1356 signal(SIGALRM, skip_signal);
1357 signal(SIGABRT, skip_signal);
1358
1359 status = 0;
1360 for (run_idx = 0; forever || run_idx < run_count; run_idx++) {
1361 if (run_count != 1 && verbose)
1362 fprintf(output, "[ perf stat: executing run #%d ... ]\n",
1363 run_idx + 1);
1364
1365 status = run_perf_stat(argc, argv);
1366 if (forever && status != -1) {
1367 print_counters(NULL, argc, argv);
1368 perf_stat__reset_stats();
1369 }
1370 }
1371
1372 if (!forever && status != -1 && !interval)
1373 print_counters(NULL, argc, argv);
1374
1375 perf_evlist__free_stats(evsel_list);
1376 out:
1377 perf_evlist__delete(evsel_list);
1378 return status;
1379 }
This page took 0.078966 seconds and 5 git commands to generate.