?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
perf stat: Print cache misses as percentage
[linux-2.6.git] / tools / perf / builtin-stat.c
blob0de3a2002f49af3f479f48c6c216f686e13962f9
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 Time: 0.104
11
12 Performance counter stats for '/home/mingo/hackbench':
13
14 1255.538611 task clock ticks # 10.143 CPU utilization factor
15 54011 context switches # 0.043 M/sec
16 385 CPU migrations # 0.000 M/sec
17 17755 pagefaults # 0.014 M/sec
18 3808323185 CPU cycles # 3033.219 M/sec
19 1575111190 instructions # 1254.530 M/sec
20 17367895 cache references # 13.833 M/sec
21 7674421 cache misses # 6.112 M/sec
22
23 Wall-clock time elapsed: 123.786620 msecs
24
25 *
26 * Copyright (C) 2008, Red Hat Inc, Ingo Molnar <mingo@redhat.com>
27 *
28 * Improvements and fixes by:
29 *
30 * Arjan van de Ven <arjan@linux.intel.com>
31 * Yanmin Zhang <yanmin.zhang@intel.com>
32 * Wu Fengguang <fengguang.wu@intel.com>
33 * Mike Galbraith <efault@gmx.de>
34 * Paul Mackerras <paulus@samba.org>
35 * Jaswinder Singh Rajput <jaswinder@kernel.org>
36 *
37 * Released under the GPL v2. (and only v2, not any later version)
38 */
39
40 #include "perf.h"
41 #include "builtin.h"
42 #include "util/util.h"
43 #include "util/parse-options.h"
44 #include "util/parse-events.h"
45 #include "util/event.h"
46 #include "util/evlist.h"
47 #include "util/evsel.h"
48 #include "util/debug.h"
49 #include "util/header.h"
50 #include "util/cpumap.h"
51 #include "util/thread.h"
52 #include "util/thread_map.h"
53
54 #include <sys/prctl.h>
55 #include <math.h>
56 #include <locale.h>
57
58 #define DEFAULT_SEPARATOR " "
59
60 static struct perf_event_attr default_attrs[] = {
61
62 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK },
63 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES },
64 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS },
65 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS },
66
67 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES },
68 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS },
69 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS },
70 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES },
71 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CACHE_REFERENCES },
72 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CACHE_MISSES },
73
74 };
75
76 struct perf_evlist *evsel_list;
77
78 static bool system_wide = false;
79 static int run_idx = 0;
80
81 static int run_count = 1;
82 static bool no_inherit = false;
83 static bool scale = true;
84 static bool no_aggr = false;
85 static pid_t target_pid = -1;
86 static pid_t target_tid = -1;
87 static pid_t child_pid = -1;
88 static bool null_run = false;
89 static bool big_num = true;
90 static int big_num_opt = -1;
91 static const char *cpu_list;
92 static const char *csv_sep = NULL;
93 static bool csv_output = false;
94
95 static volatile int done = 0;
96
97 struct stats
98 {
99 double n, mean, M2;
100 };
101
102 struct perf_stat {
103 struct stats res_stats[3];
104 };
105
106 static int perf_evsel__alloc_stat_priv(struct perf_evsel *evsel)
107 {
108 evsel->priv = zalloc(sizeof(struct perf_stat));
109 return evsel->priv == NULL ? -ENOMEM : 0;
110 }
111
112 static void perf_evsel__free_stat_priv(struct perf_evsel *evsel)
113 {
114 free(evsel->priv);
115 evsel->priv = NULL;
116 }
117
118 static void update_stats(struct stats *stats, u64 val)
119 {
120 double delta;
121
122 stats->n++;
123 delta = val - stats->mean;
124 stats->mean += delta / stats->n;
125 stats->M2 += delta*(val - stats->mean);
126 }
127
128 static double avg_stats(struct stats *stats)
129 {
130 return stats->mean;
131 }
132
133 /*
134 * http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
135 *
136 * (\Sum n_i^2) - ((\Sum n_i)^2)/n
137 * s^2 = -------------------------------
138 * n - 1
139 *
140 * http://en.wikipedia.org/wiki/Stddev
141 *
142 * The std dev of the mean is related to the std dev by:
143 *
144 * s
145 * s_mean = -------
146 * sqrt(n)
147 *
148 */
149 static double stddev_stats(struct stats *stats)
150 {
151 double variance = stats->M2 / (stats->n - 1);
152 double variance_mean = variance / stats->n;
153
154 return sqrt(variance_mean);
155 }
156
157 struct stats runtime_nsecs_stats[MAX_NR_CPUS];
158 struct stats runtime_cycles_stats[MAX_NR_CPUS];
159 struct stats runtime_branches_stats[MAX_NR_CPUS];
160 struct stats runtime_cacherefs_stats[MAX_NR_CPUS];
161 struct stats walltime_nsecs_stats;
162
163 static int create_perf_stat_counter(struct perf_evsel *evsel)
164 {
165 struct perf_event_attr *attr = &evsel->attr;
166
167 if (scale)
168 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
169 PERF_FORMAT_TOTAL_TIME_RUNNING;
170
171 attr->inherit = !no_inherit;
172
173 if (system_wide)
174 return perf_evsel__open_per_cpu(evsel, evsel_list->cpus, false);
175
176 if (target_pid == -1 && target_tid == -1) {
177 attr->disabled = 1;
178 attr->enable_on_exec = 1;
179 }
180
181 return perf_evsel__open_per_thread(evsel, evsel_list->threads, false);
182 }
183
184 /*
185 * Does the counter have nsecs as a unit?
186 */
187 static inline int nsec_counter(struct perf_evsel *evsel)
188 {
189 if (perf_evsel__match(evsel, SOFTWARE, SW_CPU_CLOCK) ||
190 perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
191 return 1;
192
193 return 0;
194 }
195
196 /*
197 * Read out the results of a single counter:
198 * aggregate counts across CPUs in system-wide mode
199 */
200 static int read_counter_aggr(struct perf_evsel *counter)
201 {
202 struct perf_stat *ps = counter->priv;
203 u64 *count = counter->counts->aggr.values;
204 int i;
205
206 if (__perf_evsel__read(counter, evsel_list->cpus->nr,
207 evsel_list->threads->nr, scale) < 0)
208 return -1;
209
210 for (i = 0; i < 3; i++)
211 update_stats(&ps->res_stats[i], count[i]);
212
213 if (verbose) {
214 fprintf(stderr, "%s: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
215 event_name(counter), count[0], count[1], count[2]);
216 }
217
218 /*
219 * Save the full runtime - to allow normalization during printout:
220 */
221 if (perf_evsel__match(counter, SOFTWARE, SW_TASK_CLOCK))
222 update_stats(&runtime_nsecs_stats[0], count[0]);
223 else if (perf_evsel__match(counter, HARDWARE, HW_CPU_CYCLES))
224 update_stats(&runtime_cycles_stats[0], count[0]);
225 else if (perf_evsel__match(counter, HARDWARE, HW_BRANCH_INSTRUCTIONS))
226 update_stats(&runtime_branches_stats[0], count[0]);
227 else if (perf_evsel__match(counter, HARDWARE, HW_CACHE_REFERENCES))
228 update_stats(&runtime_cacherefs_stats[0], count[0]);
229
230 return 0;
231 }
232
233 /*
234 * Read out the results of a single counter:
235 * do not aggregate counts across CPUs in system-wide mode
236 */
237 static int read_counter(struct perf_evsel *counter)
238 {
239 u64 *count;
240 int cpu;
241
242 for (cpu = 0; cpu < evsel_list->cpus->nr; cpu++) {
243 if (__perf_evsel__read_on_cpu(counter, cpu, 0, scale) < 0)
244 return -1;
245
246 count = counter->counts->cpu[cpu].values;
247
248 if (perf_evsel__match(counter, SOFTWARE, SW_TASK_CLOCK))
249 update_stats(&runtime_nsecs_stats[cpu], count[0]);
250 if (perf_evsel__match(counter, HARDWARE, HW_CPU_CYCLES))
251 update_stats(&runtime_cycles_stats[cpu], count[0]);
252 if (perf_evsel__match(counter, HARDWARE, HW_BRANCH_INSTRUCTIONS))
253 update_stats(&runtime_branches_stats[cpu], count[0]);
254 }
255
256 return 0;
257 }
258
259 static int run_perf_stat(int argc __used, const char **argv)
260 {
261 unsigned long long t0, t1;
262 struct perf_evsel *counter;
263 int status = 0;
264 int child_ready_pipe[2], go_pipe[2];
265 const bool forks = (argc > 0);
266 char buf;
267
268 if (forks && (pipe(child_ready_pipe) < 0 || pipe(go_pipe) < 0)) {
269 perror("failed to create pipes");
270 exit(1);
271 }
272
273 if (forks) {
274 if ((child_pid = fork()) < 0)
275 perror("failed to fork");
276
277 if (!child_pid) {
278 close(child_ready_pipe[0]);
279 close(go_pipe[1]);
280 fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
281
282 /*
283 * Do a dummy execvp to get the PLT entry resolved,
284 * so we avoid the resolver overhead on the real
285 * execvp call.
286 */
287 execvp("", (char **)argv);
288
289 /*
290 * Tell the parent we're ready to go
291 */
292 close(child_ready_pipe[1]);
293
294 /*
295 * Wait until the parent tells us to go.
296 */
297 if (read(go_pipe[0], &buf, 1) == -1)
298 perror("unable to read pipe");
299
300 execvp(argv[0], (char **)argv);
301
302 perror(argv[0]);
303 exit(-1);
304 }
305
306 if (target_tid == -1 && target_pid == -1 && !system_wide)
307 evsel_list->threads->map[0] = child_pid;
308
309 /*
310 * Wait for the child to be ready to exec.
311 */
312 close(child_ready_pipe[1]);
313 close(go_pipe[0]);
314 if (read(child_ready_pipe[0], &buf, 1) == -1)
315 perror("unable to read pipe");
316 close(child_ready_pipe[0]);
317 }
318
319 list_for_each_entry(counter, &evsel_list->entries, node) {
320 if (create_perf_stat_counter(counter) < 0) {
321 if (errno == -EPERM || errno == -EACCES) {
322 error("You may not have permission to collect %sstats.\n"
323 "\t Consider tweaking"
324 " /proc/sys/kernel/perf_event_paranoid or running as root.",
325 system_wide ? "system-wide " : "");
326 } else if (errno == ENOENT) {
327 error("%s event is not supported. ", event_name(counter));
328 } else {
329 error("open_counter returned with %d (%s). "
330 "/bin/dmesg may provide additional information.\n",
331 errno, strerror(errno));
332 }
333 if (child_pid != -1)
334 kill(child_pid, SIGTERM);
335 die("Not all events could be opened.\n");
336 return -1;
337 }
338 }
339
340 if (perf_evlist__set_filters(evsel_list)) {
341 error("failed to set filter with %d (%s)\n", errno,
342 strerror(errno));
343 return -1;
344 }
345
346 /*
347 * Enable counters and exec the command:
348 */
349 t0 = rdclock();
350
351 if (forks) {
352 close(go_pipe[1]);
353 wait(&status);
354 } else {
355 while(!done) sleep(1);
356 }
357
358 t1 = rdclock();
359
360 update_stats(&walltime_nsecs_stats, t1 - t0);
361
362 if (no_aggr) {
363 list_for_each_entry(counter, &evsel_list->entries, node) {
364 read_counter(counter);
365 perf_evsel__close_fd(counter, evsel_list->cpus->nr, 1);
366 }
367 } else {
368 list_for_each_entry(counter, &evsel_list->entries, node) {
369 read_counter_aggr(counter);
370 perf_evsel__close_fd(counter, evsel_list->cpus->nr,
371 evsel_list->threads->nr);
372 }
373 }
374
375 return WEXITSTATUS(status);
376 }
377
378 static void print_noise(struct perf_evsel *evsel, double avg)
379 {
380 struct perf_stat *ps;
381
382 if (run_count == 1)
383 return;
384
385 ps = evsel->priv;
386 fprintf(stderr, " ( +- %7.3f%% )",
387 100 * stddev_stats(&ps->res_stats[0]) / avg);
388 }
389
390 static void nsec_printout(int cpu, struct perf_evsel *evsel, double avg)
391 {
392 double msecs = avg / 1e6;
393 char cpustr[16] = { '\0', };
394 const char *fmt = csv_output ? "%s%.6f%s%s" : "%s%18.6f%s%-24s";
395
396 if (no_aggr)
397 sprintf(cpustr, "CPU%*d%s",
398 csv_output ? 0 : -4,
399 evsel_list->cpus->map[cpu], csv_sep);
400
401 fprintf(stderr, fmt, cpustr, msecs, csv_sep, event_name(evsel));
402
403 if (evsel->cgrp)
404 fprintf(stderr, "%s%s", csv_sep, evsel->cgrp->name);
405
406 if (csv_output)
407 return;
408
409 if (perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
410 fprintf(stderr, " # %10.3f CPUs",
411 avg / avg_stats(&walltime_nsecs_stats));
412 }
413
414 static void abs_printout(int cpu, struct perf_evsel *evsel, double avg)
415 {
416 double total, ratio = 0.0;
417 char cpustr[16] = { '\0', };
418 const char *fmt;
419
420 if (csv_output)
421 fmt = "%s%.0f%s%s";
422 else if (big_num)
423 fmt = "%s%'18.0f%s%-24s";
424 else
425 fmt = "%s%18.0f%s%-24s";
426
427 if (no_aggr)
428 sprintf(cpustr, "CPU%*d%s",
429 csv_output ? 0 : -4,
430 evsel_list->cpus->map[cpu], csv_sep);
431 else
432 cpu = 0;
433
434 fprintf(stderr, fmt, cpustr, avg, csv_sep, event_name(evsel));
435
436 if (evsel->cgrp)
437 fprintf(stderr, "%s%s", csv_sep, evsel->cgrp->name);
438
439 if (csv_output)
440 return;
441
442 if (perf_evsel__match(evsel, HARDWARE, HW_INSTRUCTIONS)) {
443 total = avg_stats(&runtime_cycles_stats[cpu]);
444
445 if (total)
446 ratio = avg / total;
447
448 fprintf(stderr, " # ( %4.2f instructions per cycle )", ratio);
449 } else if (perf_evsel__match(evsel, HARDWARE, HW_BRANCH_MISSES) &&
450 runtime_branches_stats[cpu].n != 0) {
451 total = avg_stats(&runtime_branches_stats[cpu]);
452
453 if (total)
454 ratio = avg * 100 / total;
455
456 fprintf(stderr, " # %10.3f %%", ratio);
457
458 } else if (perf_evsel__match(evsel, HARDWARE, HW_CACHE_MISSES) &&
459 runtime_cacherefs_stats[cpu].n != 0) {
460 total = avg_stats(&runtime_cacherefs_stats[cpu]);
461
462 if (total)
463 ratio = avg * 100 / total;
464
465 fprintf(stderr, " # %10.3f %%", ratio);
466
467 } else if (runtime_nsecs_stats[cpu].n != 0) {
468 total = avg_stats(&runtime_nsecs_stats[cpu]);
469
470 if (total)
471 ratio = 1000.0 * avg / total;
472
473 fprintf(stderr, " # %10.3f M/sec", ratio);
474 } else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES)) {
475 total = avg_stats(&runtime_cycles_stats[cpu]);
476
477 if (total)
478 ratio = avg / total * 100.0;
479
480 fprintf(stderr, " # (%5.2f%% of all cycles )", ratio);
481 }
482 }
483
484 /*
485 * Print out the results of a single counter:
486 * aggregated counts in system-wide mode
487 */
488 static void print_counter_aggr(struct perf_evsel *counter)
489 {
490 struct perf_stat *ps = counter->priv;
491 double avg = avg_stats(&ps->res_stats[0]);
492 int scaled = counter->counts->scaled;
493
494 if (scaled == -1) {
495 fprintf(stderr, "%*s%s%*s",
496 csv_output ? 0 : 18,
497 "<not counted>",
498 csv_sep,
499 csv_output ? 0 : -24,
500 event_name(counter));
501
502 if (counter->cgrp)
503 fprintf(stderr, "%s%s", csv_sep, counter->cgrp->name);
504
505 fputc('\n', stderr);
506 return;
507 }
508
509 if (nsec_counter(counter))
510 nsec_printout(-1, counter, avg);
511 else
512 abs_printout(-1, counter, avg);
513
514 if (csv_output) {
515 fputc('\n', stderr);
516 return;
517 }
518
519 print_noise(counter, avg);
520
521 if (scaled) {
522 double avg_enabled, avg_running;
523
524 avg_enabled = avg_stats(&ps->res_stats[1]);
525 avg_running = avg_stats(&ps->res_stats[2]);
526
527 fprintf(stderr, " (scaled from %.2f%%)",
528 100 * avg_running / avg_enabled);
529 }
530 fprintf(stderr, "\n");
531 }
532
533 /*
534 * Print out the results of a single counter:
535 * does not use aggregated count in system-wide
536 */
537 static void print_counter(struct perf_evsel *counter)
538 {
539 u64 ena, run, val;
540 int cpu;
541
542 for (cpu = 0; cpu < evsel_list->cpus->nr; cpu++) {
543 val = counter->counts->cpu[cpu].val;
544 ena = counter->counts->cpu[cpu].ena;
545 run = counter->counts->cpu[cpu].run;
546 if (run == 0 || ena == 0) {
547 fprintf(stderr, "CPU%*d%s%*s%s%*s",
548 csv_output ? 0 : -4,
549 evsel_list->cpus->map[cpu], csv_sep,
550 csv_output ? 0 : 18,
551 "<not counted>", csv_sep,
552 csv_output ? 0 : -24,
553 event_name(counter));
554
555 if (counter->cgrp)
556 fprintf(stderr, "%s%s", csv_sep, counter->cgrp->name);
557
558 fputc('\n', stderr);
559 continue;
560 }
561
562 if (nsec_counter(counter))
563 nsec_printout(cpu, counter, val);
564 else
565 abs_printout(cpu, counter, val);
566
567 if (!csv_output) {
568 print_noise(counter, 1.0);
569
570 if (run != ena) {
571 fprintf(stderr, " (scaled from %.2f%%)",
572 100.0 * run / ena);
573 }
574 }
575 fputc('\n', stderr);
576 }
577 }
578
579 static void print_stat(int argc, const char **argv)
580 {
581 struct perf_evsel *counter;
582 int i;
583
584 fflush(stdout);
585
586 if (!csv_output) {
587 fprintf(stderr, "\n");
588 fprintf(stderr, " Performance counter stats for ");
589 if(target_pid == -1 && target_tid == -1) {
590 fprintf(stderr, "\'%s", argv[0]);
591 for (i = 1; i < argc; i++)
592 fprintf(stderr, " %s", argv[i]);
593 } else if (target_pid != -1)
594 fprintf(stderr, "process id \'%d", target_pid);
595 else
596 fprintf(stderr, "thread id \'%d", target_tid);
597
598 fprintf(stderr, "\'");
599 if (run_count > 1)
600 fprintf(stderr, " (%d runs)", run_count);
601 fprintf(stderr, ":\n\n");
602 }
603
604 if (no_aggr) {
605 list_for_each_entry(counter, &evsel_list->entries, node)
606 print_counter(counter);
607 } else {
608 list_for_each_entry(counter, &evsel_list->entries, node)
609 print_counter_aggr(counter);
610 }
611
612 if (!csv_output) {
613 fprintf(stderr, "\n");
614 fprintf(stderr, " %18.9f seconds time elapsed",
615 avg_stats(&walltime_nsecs_stats)/1e9);
616 if (run_count > 1) {
617 fprintf(stderr, " ( +- %7.3f%% )",
618 100*stddev_stats(&walltime_nsecs_stats) /
619 avg_stats(&walltime_nsecs_stats));
620 }
621 fprintf(stderr, "\n\n");
622 }
623 }
624
625 static volatile int signr = -1;
626
627 static void skip_signal(int signo)
628 {
629 if(child_pid == -1)
630 done = 1;
631
632 signr = signo;
633 }
634
635 static void sig_atexit(void)
636 {
637 if (child_pid != -1)
638 kill(child_pid, SIGTERM);
639
640 if (signr == -1)
641 return;
642
643 signal(signr, SIG_DFL);
644 kill(getpid(), signr);
645 }
646
647 static const char * const stat_usage[] = {
648 "perf stat [<options>] [<command>]",
649 NULL
650 };
651
652 static int stat__set_big_num(const struct option *opt __used,
653 const char *s __used, int unset)
654 {
655 big_num_opt = unset ? 0 : 1;
656 return 0;
657 }
658
659 static const struct option options[] = {
660 OPT_CALLBACK('e', "event", &evsel_list, "event",
661 "event selector. use 'perf list' to list available events",
662 parse_events),
663 OPT_CALLBACK(0, "filter", &evsel_list, "filter",
664 "event filter", parse_filter),
665 OPT_BOOLEAN('i', "no-inherit", &no_inherit,
666 "child tasks do not inherit counters"),
667 OPT_INTEGER('p', "pid", &target_pid,
668 "stat events on existing process id"),
669 OPT_INTEGER('t', "tid", &target_tid,
670 "stat events on existing thread id"),
671 OPT_BOOLEAN('a', "all-cpus", &system_wide,
672 "system-wide collection from all CPUs"),
673 OPT_BOOLEAN('c', "scale", &scale,
674 "scale/normalize counters"),
675 OPT_INCR('v', "verbose", &verbose,
676 "be more verbose (show counter open errors, etc)"),
677 OPT_INTEGER('r', "repeat", &run_count,
678 "repeat command and print average + stddev (max: 100)"),
679 OPT_BOOLEAN('n', "null", &null_run,
680 "null run - dont start any counters"),
681 OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
682 "print large numbers with thousands\' separators",
683 stat__set_big_num),
684 OPT_STRING('C', "cpu", &cpu_list, "cpu",
685 "list of cpus to monitor in system-wide"),
686 OPT_BOOLEAN('A', "no-aggr", &no_aggr,
687 "disable CPU count aggregation"),
688 OPT_STRING('x', "field-separator", &csv_sep, "separator",
689 "print counts with custom separator"),
690 OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
691 "monitor event in cgroup name only",
692 parse_cgroups),
693 OPT_END()
694 };
695
696 int cmd_stat(int argc, const char **argv, const char *prefix __used)
697 {
698 struct perf_evsel *pos;
699 int status = -ENOMEM;
700
701 setlocale(LC_ALL, "");
702
703 evsel_list = perf_evlist__new(NULL, NULL);
704 if (evsel_list == NULL)
705 return -ENOMEM;
706
707 argc = parse_options(argc, argv, options, stat_usage,
708 PARSE_OPT_STOP_AT_NON_OPTION);
709
710 if (csv_sep)
711 csv_output = true;
712 else
713 csv_sep = DEFAULT_SEPARATOR;
714
715 /*
716 * let the spreadsheet do the pretty-printing
717 */
718 if (csv_output) {
719 /* User explicitely passed -B? */
720 if (big_num_opt == 1) {
721 fprintf(stderr, "-B option not supported with -x\n");
722 usage_with_options(stat_usage, options);
723 } else /* Nope, so disable big number formatting */
724 big_num = false;
725 } else if (big_num_opt == 0) /* User passed --no-big-num */
726 big_num = false;
727
728 if (!argc && target_pid == -1 && target_tid == -1)
729 usage_with_options(stat_usage, options);
730 if (run_count <= 0)
731 usage_with_options(stat_usage, options);
732
733 /* no_aggr, cgroup are for system-wide only */
734 if ((no_aggr || nr_cgroups) && !system_wide) {
735 fprintf(stderr, "both cgroup and no-aggregation "
736 "modes only available in system-wide mode\n");
737
738 usage_with_options(stat_usage, options);
739 }
740
741 /* Set attrs and nr_counters if no event is selected and !null_run */
742 if (!null_run && !evsel_list->nr_entries) {
743 size_t c;
744
745 for (c = 0; c < ARRAY_SIZE(default_attrs); ++c) {
746 pos = perf_evsel__new(&default_attrs[c], c);
747 if (pos == NULL)
748 goto out;
749 perf_evlist__add(evsel_list, pos);
750 }
751 }
752
753 if (target_pid != -1)
754 target_tid = target_pid;
755
756 evsel_list->threads = thread_map__new(target_pid, target_tid);
757 if (evsel_list->threads == NULL) {
758 pr_err("Problems finding threads of monitor\n");
759 usage_with_options(stat_usage, options);
760 }
761
762 if (system_wide)
763 evsel_list->cpus = cpu_map__new(cpu_list);
764 else
765 evsel_list->cpus = cpu_map__dummy_new();
766
767 if (evsel_list->cpus == NULL) {
768 perror("failed to parse CPUs map");
769 usage_with_options(stat_usage, options);
770 return -1;
771 }
772
773 list_for_each_entry(pos, &evsel_list->entries, node) {
774 if (perf_evsel__alloc_stat_priv(pos) < 0 ||
775 perf_evsel__alloc_counts(pos, evsel_list->cpus->nr) < 0 ||
776 perf_evsel__alloc_fd(pos, evsel_list->cpus->nr, evsel_list->threads->nr) < 0)
777 goto out_free_fd;
778 }
779
780 /*
781 * We dont want to block the signals - that would cause
782 * child tasks to inherit that and Ctrl-C would not work.
783 * What we want is for Ctrl-C to work in the exec()-ed
784 * task, but being ignored by perf stat itself:
785 */
786 atexit(sig_atexit);
787 signal(SIGINT, skip_signal);
788 signal(SIGALRM, skip_signal);
789 signal(SIGABRT, skip_signal);
790
791 status = 0;
792 for (run_idx = 0; run_idx < run_count; run_idx++) {
793 if (run_count != 1 && verbose)
794 fprintf(stderr, "[ perf stat: executing run #%d ... ]\n", run_idx + 1);
795 status = run_perf_stat(argc, argv);
796 }
797
798 if (status != -1)
799 print_stat(argc, argv);
800 out_free_fd:
801 list_for_each_entry(pos, &evsel_list->entries, node)
802 perf_evsel__free_stat_priv(pos);
803 perf_evlist__delete_maps(evsel_list);
804 out:
805 perf_evlist__delete(evsel_list);
806 return status;
807 }
Linus' kernel tree