xref: /linux/tools/perf/builtin-stat.c (revision f37130533f68711fd6bae2c79950b8e72002bad6)
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/util.h"
47 #include "util/parse-options.h"
48 #include "util/parse-events.h"
49 #include "util/event.h"
50 #include "util/evlist.h"
51 #include "util/evsel.h"
52 #include "util/debug.h"
53 #include "util/color.h"
54 #include "util/stat.h"
55 #include "util/header.h"
56 #include "util/cpumap.h"
57 #include "util/thread.h"
58 #include "util/thread_map.h"
59 
60 #include <stdlib.h>
61 #include <sys/prctl.h>
62 #include <locale.h>
63 
64 #define DEFAULT_SEPARATOR	" "
65 #define CNTR_NOT_SUPPORTED	"<not supported>"
66 #define CNTR_NOT_COUNTED	"<not counted>"
67 
68 static void print_stat(int argc, const char **argv);
69 static void print_counter_aggr(struct perf_evsel *counter, char *prefix);
70 static void print_counter(struct perf_evsel *counter, char *prefix);
71 static void print_aggr_socket(char *prefix);
72 
73 static struct perf_evlist	*evsel_list;
74 
75 static struct perf_target	target = {
76 	.uid	= UINT_MAX,
77 };
78 
79 static int			run_count			=  1;
80 static bool			no_inherit			= false;
81 static bool			scale				=  true;
82 static bool			no_aggr				= false;
83 static bool			aggr_socket			= false;
84 static pid_t			child_pid			= -1;
85 static bool			null_run			=  false;
86 static int			detailed_run			=  0;
87 static bool			big_num				=  true;
88 static int			big_num_opt			=  -1;
89 static const char		*csv_sep			= NULL;
90 static bool			csv_output			= false;
91 static bool			group				= false;
92 static FILE			*output				= NULL;
93 static const char		*pre_cmd			= NULL;
94 static const char		*post_cmd			= NULL;
95 static bool			sync_run			= false;
96 static unsigned int		interval			= 0;
97 static struct timespec		ref_time;
98 static struct cpu_map		*sock_map;
99 
100 static volatile int done = 0;
101 
102 struct perf_stat {
103 	struct stats	  res_stats[3];
104 };
105 
106 static inline void diff_timespec(struct timespec *r, struct timespec *a,
107 				 struct timespec *b)
108 {
109 	r->tv_sec = a->tv_sec - b->tv_sec;
110 	if (a->tv_nsec < b->tv_nsec) {
111 		r->tv_nsec = a->tv_nsec + 1000000000L - b->tv_nsec;
112 		r->tv_sec--;
113 	} else {
114 		r->tv_nsec = a->tv_nsec - b->tv_nsec ;
115 	}
116 }
117 
118 static inline struct cpu_map *perf_evsel__cpus(struct perf_evsel *evsel)
119 {
120 	return (evsel->cpus && !target.cpu_list) ? evsel->cpus : evsel_list->cpus;
121 }
122 
123 static inline int perf_evsel__nr_cpus(struct perf_evsel *evsel)
124 {
125 	return perf_evsel__cpus(evsel)->nr;
126 }
127 
128 static int perf_evsel__alloc_stat_priv(struct perf_evsel *evsel)
129 {
130 	evsel->priv = zalloc(sizeof(struct perf_stat));
131 	return evsel->priv == NULL ? -ENOMEM : 0;
132 }
133 
134 static void perf_evsel__free_stat_priv(struct perf_evsel *evsel)
135 {
136 	free(evsel->priv);
137 	evsel->priv = NULL;
138 }
139 
140 static int perf_evsel__alloc_prev_raw_counts(struct perf_evsel *evsel)
141 {
142 	void *addr;
143 	size_t sz;
144 
145 	sz = sizeof(*evsel->counts) +
146 	     (perf_evsel__nr_cpus(evsel) * sizeof(struct perf_counts_values));
147 
148 	addr = zalloc(sz);
149 	if (!addr)
150 		return -ENOMEM;
151 
152 	evsel->prev_raw_counts =  addr;
153 
154 	return 0;
155 }
156 
157 static void perf_evsel__free_prev_raw_counts(struct perf_evsel *evsel)
158 {
159 	free(evsel->prev_raw_counts);
160 	evsel->prev_raw_counts = NULL;
161 }
162 
163 static struct stats runtime_nsecs_stats[MAX_NR_CPUS];
164 static struct stats runtime_cycles_stats[MAX_NR_CPUS];
165 static struct stats runtime_stalled_cycles_front_stats[MAX_NR_CPUS];
166 static struct stats runtime_stalled_cycles_back_stats[MAX_NR_CPUS];
167 static struct stats runtime_branches_stats[MAX_NR_CPUS];
168 static struct stats runtime_cacherefs_stats[MAX_NR_CPUS];
169 static struct stats runtime_l1_dcache_stats[MAX_NR_CPUS];
170 static struct stats runtime_l1_icache_stats[MAX_NR_CPUS];
171 static struct stats runtime_ll_cache_stats[MAX_NR_CPUS];
172 static struct stats runtime_itlb_cache_stats[MAX_NR_CPUS];
173 static struct stats runtime_dtlb_cache_stats[MAX_NR_CPUS];
174 static struct stats walltime_nsecs_stats;
175 
176 static int create_perf_stat_counter(struct perf_evsel *evsel)
177 {
178 	struct perf_event_attr *attr = &evsel->attr;
179 
180 	if (scale)
181 		attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
182 				    PERF_FORMAT_TOTAL_TIME_RUNNING;
183 
184 	attr->inherit = !no_inherit;
185 
186 	if (perf_target__has_cpu(&target))
187 		return perf_evsel__open_per_cpu(evsel, perf_evsel__cpus(evsel));
188 
189 	if (!perf_target__has_task(&target) &&
190 	    perf_evsel__is_group_leader(evsel)) {
191 		attr->disabled = 1;
192 		attr->enable_on_exec = 1;
193 	}
194 
195 	return perf_evsel__open_per_thread(evsel, evsel_list->threads);
196 }
197 
198 /*
199  * Does the counter have nsecs as a unit?
200  */
201 static inline int nsec_counter(struct perf_evsel *evsel)
202 {
203 	if (perf_evsel__match(evsel, SOFTWARE, SW_CPU_CLOCK) ||
204 	    perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
205 		return 1;
206 
207 	return 0;
208 }
209 
210 /*
211  * Update various tracking values we maintain to print
212  * more semantic information such as miss/hit ratios,
213  * instruction rates, etc:
214  */
215 static void update_shadow_stats(struct perf_evsel *counter, u64 *count)
216 {
217 	if (perf_evsel__match(counter, SOFTWARE, SW_TASK_CLOCK))
218 		update_stats(&runtime_nsecs_stats[0], count[0]);
219 	else if (perf_evsel__match(counter, HARDWARE, HW_CPU_CYCLES))
220 		update_stats(&runtime_cycles_stats[0], count[0]);
221 	else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_FRONTEND))
222 		update_stats(&runtime_stalled_cycles_front_stats[0], count[0]);
223 	else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_BACKEND))
224 		update_stats(&runtime_stalled_cycles_back_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 	else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1D))
230 		update_stats(&runtime_l1_dcache_stats[0], count[0]);
231 	else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1I))
232 		update_stats(&runtime_l1_icache_stats[0], count[0]);
233 	else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_LL))
234 		update_stats(&runtime_ll_cache_stats[0], count[0]);
235 	else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_DTLB))
236 		update_stats(&runtime_dtlb_cache_stats[0], count[0]);
237 	else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_ITLB))
238 		update_stats(&runtime_itlb_cache_stats[0], count[0]);
239 }
240 
241 /*
242  * Read out the results of a single counter:
243  * aggregate counts across CPUs in system-wide mode
244  */
245 static int read_counter_aggr(struct perf_evsel *counter)
246 {
247 	struct perf_stat *ps = counter->priv;
248 	u64 *count = counter->counts->aggr.values;
249 	int i;
250 
251 	if (__perf_evsel__read(counter, perf_evsel__nr_cpus(counter),
252 			       evsel_list->threads->nr, scale) < 0)
253 		return -1;
254 
255 	for (i = 0; i < 3; i++)
256 		update_stats(&ps->res_stats[i], count[i]);
257 
258 	if (verbose) {
259 		fprintf(output, "%s: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
260 			perf_evsel__name(counter), count[0], count[1], count[2]);
261 	}
262 
263 	/*
264 	 * Save the full runtime - to allow normalization during printout:
265 	 */
266 	update_shadow_stats(counter, count);
267 
268 	return 0;
269 }
270 
271 /*
272  * Read out the results of a single counter:
273  * do not aggregate counts across CPUs in system-wide mode
274  */
275 static int read_counter(struct perf_evsel *counter)
276 {
277 	u64 *count;
278 	int cpu;
279 
280 	for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
281 		if (__perf_evsel__read_on_cpu(counter, cpu, 0, scale) < 0)
282 			return -1;
283 
284 		count = counter->counts->cpu[cpu].values;
285 
286 		update_shadow_stats(counter, count);
287 	}
288 
289 	return 0;
290 }
291 
292 static void print_interval(void)
293 {
294 	static int num_print_interval;
295 	struct perf_evsel *counter;
296 	struct perf_stat *ps;
297 	struct timespec ts, rs;
298 	char prefix[64];
299 
300 	if (no_aggr) {
301 		list_for_each_entry(counter, &evsel_list->entries, node) {
302 			ps = counter->priv;
303 			memset(ps->res_stats, 0, sizeof(ps->res_stats));
304 			read_counter(counter);
305 		}
306 	} else {
307 		list_for_each_entry(counter, &evsel_list->entries, node) {
308 			ps = counter->priv;
309 			memset(ps->res_stats, 0, sizeof(ps->res_stats));
310 			read_counter_aggr(counter);
311 		}
312 	}
313 	clock_gettime(CLOCK_MONOTONIC, &ts);
314 	diff_timespec(&rs, &ts, &ref_time);
315 	sprintf(prefix, "%6lu.%09lu%s", rs.tv_sec, rs.tv_nsec, csv_sep);
316 
317 	if (num_print_interval == 0 && !csv_output) {
318 		if (aggr_socket)
319 			fprintf(output, "#           time socket cpus             counts events\n");
320 		else if (no_aggr)
321 			fprintf(output, "#           time CPU                 counts events\n");
322 		else
323 			fprintf(output, "#           time             counts events\n");
324 	}
325 
326 	if (++num_print_interval == 25)
327 		num_print_interval = 0;
328 
329 	if (aggr_socket)
330 		print_aggr_socket(prefix);
331 	else if (no_aggr) {
332 		list_for_each_entry(counter, &evsel_list->entries, node)
333 			print_counter(counter, prefix);
334 	} else {
335 		list_for_each_entry(counter, &evsel_list->entries, node)
336 			print_counter_aggr(counter, prefix);
337 	}
338 }
339 
340 static int __run_perf_stat(int argc __maybe_unused, const char **argv)
341 {
342 	char msg[512];
343 	unsigned long long t0, t1;
344 	struct perf_evsel *counter;
345 	struct timespec ts;
346 	int status = 0;
347 	int child_ready_pipe[2], go_pipe[2];
348 	const bool forks = (argc > 0);
349 	char buf;
350 
351 	if (interval) {
352 		ts.tv_sec  = interval / 1000;
353 		ts.tv_nsec = (interval % 1000) * 1000000;
354 	} else {
355 		ts.tv_sec  = 1;
356 		ts.tv_nsec = 0;
357 	}
358 
359 	if (aggr_socket
360 	    && cpu_map__build_socket_map(evsel_list->cpus, &sock_map)) {
361 		perror("cannot build socket map");
362 		return -1;
363 	}
364 
365 	if (forks && (pipe(child_ready_pipe) < 0 || pipe(go_pipe) < 0)) {
366 		perror("failed to create pipes");
367 		return -1;
368 	}
369 
370 	if (forks) {
371 		if ((child_pid = fork()) < 0)
372 			perror("failed to fork");
373 
374 		if (!child_pid) {
375 			close(child_ready_pipe[0]);
376 			close(go_pipe[1]);
377 			fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
378 
379 			/*
380 			 * Do a dummy execvp to get the PLT entry resolved,
381 			 * so we avoid the resolver overhead on the real
382 			 * execvp call.
383 			 */
384 			execvp("", (char **)argv);
385 
386 			/*
387 			 * Tell the parent we're ready to go
388 			 */
389 			close(child_ready_pipe[1]);
390 
391 			/*
392 			 * Wait until the parent tells us to go.
393 			 */
394 			if (read(go_pipe[0], &buf, 1) == -1)
395 				perror("unable to read pipe");
396 
397 			execvp(argv[0], (char **)argv);
398 
399 			perror(argv[0]);
400 			exit(-1);
401 		}
402 
403 		if (perf_target__none(&target))
404 			evsel_list->threads->map[0] = child_pid;
405 
406 		/*
407 		 * Wait for the child to be ready to exec.
408 		 */
409 		close(child_ready_pipe[1]);
410 		close(go_pipe[0]);
411 		if (read(child_ready_pipe[0], &buf, 1) == -1)
412 			perror("unable to read pipe");
413 		close(child_ready_pipe[0]);
414 	}
415 
416 	if (group)
417 		perf_evlist__set_leader(evsel_list);
418 
419 	list_for_each_entry(counter, &evsel_list->entries, node) {
420 		if (create_perf_stat_counter(counter) < 0) {
421 			/*
422 			 * PPC returns ENXIO for HW counters until 2.6.37
423 			 * (behavior changed with commit b0a873e).
424 			 */
425 			if (errno == EINVAL || errno == ENOSYS ||
426 			    errno == ENOENT || errno == EOPNOTSUPP ||
427 			    errno == ENXIO) {
428 				if (verbose)
429 					ui__warning("%s event is not supported by the kernel.\n",
430 						    perf_evsel__name(counter));
431 				counter->supported = false;
432 				continue;
433 			}
434 
435 			perf_evsel__open_strerror(counter, &target,
436 						  errno, msg, sizeof(msg));
437 			ui__error("%s\n", msg);
438 
439 			if (child_pid != -1)
440 				kill(child_pid, SIGTERM);
441 
442 			return -1;
443 		}
444 		counter->supported = true;
445 	}
446 
447 	if (perf_evlist__apply_filters(evsel_list)) {
448 		error("failed to set filter with %d (%s)\n", errno,
449 			strerror(errno));
450 		return -1;
451 	}
452 
453 	/*
454 	 * Enable counters and exec the command:
455 	 */
456 	t0 = rdclock();
457 	clock_gettime(CLOCK_MONOTONIC, &ref_time);
458 
459 	if (forks) {
460 		close(go_pipe[1]);
461 		if (interval) {
462 			while (!waitpid(child_pid, &status, WNOHANG)) {
463 				nanosleep(&ts, NULL);
464 				print_interval();
465 			}
466 		}
467 		wait(&status);
468 		if (WIFSIGNALED(status))
469 			psignal(WTERMSIG(status), argv[0]);
470 	} else {
471 		while (!done) {
472 			nanosleep(&ts, NULL);
473 			if (interval)
474 				print_interval();
475 		}
476 	}
477 
478 	t1 = rdclock();
479 
480 	update_stats(&walltime_nsecs_stats, t1 - t0);
481 
482 	if (no_aggr) {
483 		list_for_each_entry(counter, &evsel_list->entries, node) {
484 			read_counter(counter);
485 			perf_evsel__close_fd(counter, perf_evsel__nr_cpus(counter), 1);
486 		}
487 	} else {
488 		list_for_each_entry(counter, &evsel_list->entries, node) {
489 			read_counter_aggr(counter);
490 			perf_evsel__close_fd(counter, perf_evsel__nr_cpus(counter),
491 					     evsel_list->threads->nr);
492 		}
493 	}
494 
495 	return WEXITSTATUS(status);
496 }
497 
498 static int run_perf_stat(int argc __maybe_unused, const char **argv)
499 {
500 	int ret;
501 
502 	if (pre_cmd) {
503 		ret = system(pre_cmd);
504 		if (ret)
505 			return ret;
506 	}
507 
508 	if (sync_run)
509 		sync();
510 
511 	ret = __run_perf_stat(argc, argv);
512 	if (ret)
513 		return ret;
514 
515 	if (post_cmd) {
516 		ret = system(post_cmd);
517 		if (ret)
518 			return ret;
519 	}
520 
521 	return ret;
522 }
523 
524 static void print_noise_pct(double total, double avg)
525 {
526 	double pct = rel_stddev_stats(total, avg);
527 
528 	if (csv_output)
529 		fprintf(output, "%s%.2f%%", csv_sep, pct);
530 	else if (pct)
531 		fprintf(output, "  ( +-%6.2f%% )", pct);
532 }
533 
534 static void print_noise(struct perf_evsel *evsel, double avg)
535 {
536 	struct perf_stat *ps;
537 
538 	if (run_count == 1)
539 		return;
540 
541 	ps = evsel->priv;
542 	print_noise_pct(stddev_stats(&ps->res_stats[0]), avg);
543 }
544 
545 static void nsec_printout(int cpu, int nr, struct perf_evsel *evsel, double avg)
546 {
547 	double msecs = avg / 1e6;
548 	char cpustr[16] = { '\0', };
549 	const char *fmt = csv_output ? "%s%.6f%s%s" : "%s%18.6f%s%-25s";
550 
551 	if (aggr_socket)
552 		sprintf(cpustr, "S%*d%s%*d%s",
553 			csv_output ? 0 : -5,
554 			cpu,
555 			csv_sep,
556 			csv_output ? 0 : 4,
557 			nr,
558 			csv_sep);
559 	else if (no_aggr)
560 		sprintf(cpustr, "CPU%*d%s",
561 			csv_output ? 0 : -4,
562 			perf_evsel__cpus(evsel)->map[cpu], csv_sep);
563 
564 	fprintf(output, fmt, cpustr, msecs, csv_sep, perf_evsel__name(evsel));
565 
566 	if (evsel->cgrp)
567 		fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
568 
569 	if (csv_output || interval)
570 		return;
571 
572 	if (perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
573 		fprintf(output, " # %8.3f CPUs utilized          ",
574 			avg / avg_stats(&walltime_nsecs_stats));
575 	else
576 		fprintf(output, "                                   ");
577 }
578 
579 /* used for get_ratio_color() */
580 enum grc_type {
581 	GRC_STALLED_CYCLES_FE,
582 	GRC_STALLED_CYCLES_BE,
583 	GRC_CACHE_MISSES,
584 	GRC_MAX_NR
585 };
586 
587 static const char *get_ratio_color(enum grc_type type, double ratio)
588 {
589 	static const double grc_table[GRC_MAX_NR][3] = {
590 		[GRC_STALLED_CYCLES_FE] = { 50.0, 30.0, 10.0 },
591 		[GRC_STALLED_CYCLES_BE] = { 75.0, 50.0, 20.0 },
592 		[GRC_CACHE_MISSES] 	= { 20.0, 10.0, 5.0 },
593 	};
594 	const char *color = PERF_COLOR_NORMAL;
595 
596 	if (ratio > grc_table[type][0])
597 		color = PERF_COLOR_RED;
598 	else if (ratio > grc_table[type][1])
599 		color = PERF_COLOR_MAGENTA;
600 	else if (ratio > grc_table[type][2])
601 		color = PERF_COLOR_YELLOW;
602 
603 	return color;
604 }
605 
606 static void print_stalled_cycles_frontend(int cpu,
607 					  struct perf_evsel *evsel
608 					  __maybe_unused, double avg)
609 {
610 	double total, ratio = 0.0;
611 	const char *color;
612 
613 	total = avg_stats(&runtime_cycles_stats[cpu]);
614 
615 	if (total)
616 		ratio = avg / total * 100.0;
617 
618 	color = get_ratio_color(GRC_STALLED_CYCLES_FE, ratio);
619 
620 	fprintf(output, " #  ");
621 	color_fprintf(output, color, "%6.2f%%", ratio);
622 	fprintf(output, " frontend cycles idle   ");
623 }
624 
625 static void print_stalled_cycles_backend(int cpu,
626 					 struct perf_evsel *evsel
627 					 __maybe_unused, double avg)
628 {
629 	double total, ratio = 0.0;
630 	const char *color;
631 
632 	total = avg_stats(&runtime_cycles_stats[cpu]);
633 
634 	if (total)
635 		ratio = avg / total * 100.0;
636 
637 	color = get_ratio_color(GRC_STALLED_CYCLES_BE, ratio);
638 
639 	fprintf(output, " #  ");
640 	color_fprintf(output, color, "%6.2f%%", ratio);
641 	fprintf(output, " backend  cycles idle   ");
642 }
643 
644 static void print_branch_misses(int cpu,
645 				struct perf_evsel *evsel __maybe_unused,
646 				double avg)
647 {
648 	double total, ratio = 0.0;
649 	const char *color;
650 
651 	total = avg_stats(&runtime_branches_stats[cpu]);
652 
653 	if (total)
654 		ratio = avg / total * 100.0;
655 
656 	color = get_ratio_color(GRC_CACHE_MISSES, ratio);
657 
658 	fprintf(output, " #  ");
659 	color_fprintf(output, color, "%6.2f%%", ratio);
660 	fprintf(output, " of all branches        ");
661 }
662 
663 static void print_l1_dcache_misses(int cpu,
664 				   struct perf_evsel *evsel __maybe_unused,
665 				   double avg)
666 {
667 	double total, ratio = 0.0;
668 	const char *color;
669 
670 	total = avg_stats(&runtime_l1_dcache_stats[cpu]);
671 
672 	if (total)
673 		ratio = avg / total * 100.0;
674 
675 	color = get_ratio_color(GRC_CACHE_MISSES, ratio);
676 
677 	fprintf(output, " #  ");
678 	color_fprintf(output, color, "%6.2f%%", ratio);
679 	fprintf(output, " of all L1-dcache hits  ");
680 }
681 
682 static void print_l1_icache_misses(int cpu,
683 				   struct perf_evsel *evsel __maybe_unused,
684 				   double avg)
685 {
686 	double total, ratio = 0.0;
687 	const char *color;
688 
689 	total = avg_stats(&runtime_l1_icache_stats[cpu]);
690 
691 	if (total)
692 		ratio = avg / total * 100.0;
693 
694 	color = get_ratio_color(GRC_CACHE_MISSES, ratio);
695 
696 	fprintf(output, " #  ");
697 	color_fprintf(output, color, "%6.2f%%", ratio);
698 	fprintf(output, " of all L1-icache hits  ");
699 }
700 
701 static void print_dtlb_cache_misses(int cpu,
702 				    struct perf_evsel *evsel __maybe_unused,
703 				    double avg)
704 {
705 	double total, ratio = 0.0;
706 	const char *color;
707 
708 	total = avg_stats(&runtime_dtlb_cache_stats[cpu]);
709 
710 	if (total)
711 		ratio = avg / total * 100.0;
712 
713 	color = get_ratio_color(GRC_CACHE_MISSES, ratio);
714 
715 	fprintf(output, " #  ");
716 	color_fprintf(output, color, "%6.2f%%", ratio);
717 	fprintf(output, " of all dTLB cache hits ");
718 }
719 
720 static void print_itlb_cache_misses(int cpu,
721 				    struct perf_evsel *evsel __maybe_unused,
722 				    double avg)
723 {
724 	double total, ratio = 0.0;
725 	const char *color;
726 
727 	total = avg_stats(&runtime_itlb_cache_stats[cpu]);
728 
729 	if (total)
730 		ratio = avg / total * 100.0;
731 
732 	color = get_ratio_color(GRC_CACHE_MISSES, ratio);
733 
734 	fprintf(output, " #  ");
735 	color_fprintf(output, color, "%6.2f%%", ratio);
736 	fprintf(output, " of all iTLB cache hits ");
737 }
738 
739 static void print_ll_cache_misses(int cpu,
740 				  struct perf_evsel *evsel __maybe_unused,
741 				  double avg)
742 {
743 	double total, ratio = 0.0;
744 	const char *color;
745 
746 	total = avg_stats(&runtime_ll_cache_stats[cpu]);
747 
748 	if (total)
749 		ratio = avg / total * 100.0;
750 
751 	color = get_ratio_color(GRC_CACHE_MISSES, ratio);
752 
753 	fprintf(output, " #  ");
754 	color_fprintf(output, color, "%6.2f%%", ratio);
755 	fprintf(output, " of all LL-cache hits   ");
756 }
757 
758 static void abs_printout(int cpu, int nr, struct perf_evsel *evsel, double avg)
759 {
760 	double total, ratio = 0.0;
761 	char cpustr[16] = { '\0', };
762 	const char *fmt;
763 
764 	if (csv_output)
765 		fmt = "%s%.0f%s%s";
766 	else if (big_num)
767 		fmt = "%s%'18.0f%s%-25s";
768 	else
769 		fmt = "%s%18.0f%s%-25s";
770 
771 	if (aggr_socket)
772 		sprintf(cpustr, "S%*d%s%*d%s",
773 			csv_output ? 0 : -5,
774 			cpu,
775 			csv_sep,
776 			csv_output ? 0 : 4,
777 			nr,
778 			csv_sep);
779 	else if (no_aggr)
780 		sprintf(cpustr, "CPU%*d%s",
781 			csv_output ? 0 : -4,
782 			perf_evsel__cpus(evsel)->map[cpu], csv_sep);
783 	else
784 		cpu = 0;
785 
786 	fprintf(output, fmt, cpustr, avg, csv_sep, perf_evsel__name(evsel));
787 
788 	if (evsel->cgrp)
789 		fprintf(output, "%s%s", csv_sep, evsel->cgrp->name);
790 
791 	if (csv_output || interval)
792 		return;
793 
794 	if (perf_evsel__match(evsel, HARDWARE, HW_INSTRUCTIONS)) {
795 		total = avg_stats(&runtime_cycles_stats[cpu]);
796 		if (total)
797 			ratio = avg / total;
798 
799 		fprintf(output, " #   %5.2f  insns per cycle        ", ratio);
800 
801 		total = avg_stats(&runtime_stalled_cycles_front_stats[cpu]);
802 		total = max(total, avg_stats(&runtime_stalled_cycles_back_stats[cpu]));
803 
804 		if (total && avg) {
805 			ratio = total / avg;
806 			fprintf(output, "\n                                             #   %5.2f  stalled cycles per insn", ratio);
807 		}
808 
809 	} else if (perf_evsel__match(evsel, HARDWARE, HW_BRANCH_MISSES) &&
810 			runtime_branches_stats[cpu].n != 0) {
811 		print_branch_misses(cpu, evsel, avg);
812 	} else if (
813 		evsel->attr.type == PERF_TYPE_HW_CACHE &&
814 		evsel->attr.config ==  ( PERF_COUNT_HW_CACHE_L1D |
815 					((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
816 					((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
817 			runtime_l1_dcache_stats[cpu].n != 0) {
818 		print_l1_dcache_misses(cpu, evsel, avg);
819 	} else if (
820 		evsel->attr.type == PERF_TYPE_HW_CACHE &&
821 		evsel->attr.config ==  ( PERF_COUNT_HW_CACHE_L1I |
822 					((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
823 					((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
824 			runtime_l1_icache_stats[cpu].n != 0) {
825 		print_l1_icache_misses(cpu, evsel, avg);
826 	} else if (
827 		evsel->attr.type == PERF_TYPE_HW_CACHE &&
828 		evsel->attr.config ==  ( PERF_COUNT_HW_CACHE_DTLB |
829 					((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
830 					((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
831 			runtime_dtlb_cache_stats[cpu].n != 0) {
832 		print_dtlb_cache_misses(cpu, evsel, avg);
833 	} else if (
834 		evsel->attr.type == PERF_TYPE_HW_CACHE &&
835 		evsel->attr.config ==  ( PERF_COUNT_HW_CACHE_ITLB |
836 					((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
837 					((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
838 			runtime_itlb_cache_stats[cpu].n != 0) {
839 		print_itlb_cache_misses(cpu, evsel, avg);
840 	} else if (
841 		evsel->attr.type == PERF_TYPE_HW_CACHE &&
842 		evsel->attr.config ==  ( PERF_COUNT_HW_CACHE_LL |
843 					((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
844 					((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
845 			runtime_ll_cache_stats[cpu].n != 0) {
846 		print_ll_cache_misses(cpu, evsel, avg);
847 	} else if (perf_evsel__match(evsel, HARDWARE, HW_CACHE_MISSES) &&
848 			runtime_cacherefs_stats[cpu].n != 0) {
849 		total = avg_stats(&runtime_cacherefs_stats[cpu]);
850 
851 		if (total)
852 			ratio = avg * 100 / total;
853 
854 		fprintf(output, " # %8.3f %% of all cache refs    ", ratio);
855 
856 	} else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_FRONTEND)) {
857 		print_stalled_cycles_frontend(cpu, evsel, avg);
858 	} else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_BACKEND)) {
859 		print_stalled_cycles_backend(cpu, evsel, avg);
860 	} else if (perf_evsel__match(evsel, HARDWARE, HW_CPU_CYCLES)) {
861 		total = avg_stats(&runtime_nsecs_stats[cpu]);
862 
863 		if (total)
864 			ratio = 1.0 * avg / total;
865 
866 		fprintf(output, " # %8.3f GHz                    ", ratio);
867 	} else if (runtime_nsecs_stats[cpu].n != 0) {
868 		char unit = 'M';
869 
870 		total = avg_stats(&runtime_nsecs_stats[cpu]);
871 
872 		if (total)
873 			ratio = 1000.0 * avg / total;
874 		if (ratio < 0.001) {
875 			ratio *= 1000;
876 			unit = 'K';
877 		}
878 
879 		fprintf(output, " # %8.3f %c/sec                  ", ratio, unit);
880 	} else {
881 		fprintf(output, "                                   ");
882 	}
883 }
884 
885 static void print_aggr_socket(char *prefix)
886 {
887 	struct perf_evsel *counter;
888 	u64 ena, run, val;
889 	int cpu, s, s2, sock, nr;
890 
891 	if (!sock_map)
892 		return;
893 
894 	for (s = 0; s < sock_map->nr; s++) {
895 		sock = cpu_map__socket(sock_map, s);
896 		list_for_each_entry(counter, &evsel_list->entries, node) {
897 			val = ena = run = 0;
898 			nr = 0;
899 			for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
900 				s2 = cpu_map__get_socket(evsel_list->cpus, cpu);
901 				if (s2 != sock)
902 					continue;
903 				val += counter->counts->cpu[cpu].val;
904 				ena += counter->counts->cpu[cpu].ena;
905 				run += counter->counts->cpu[cpu].run;
906 				nr++;
907 			}
908 			if (prefix)
909 				fprintf(output, "%s", prefix);
910 
911 			if (run == 0 || ena == 0) {
912 				fprintf(output, "S%*d%s%*d%s%*s%s%*s",
913 					csv_output ? 0 : -5,
914 					s,
915 					csv_sep,
916 					csv_output ? 0 : 4,
917 					nr,
918 					csv_sep,
919 					csv_output ? 0 : 18,
920 					counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
921 					csv_sep,
922 					csv_output ? 0 : -24,
923 					perf_evsel__name(counter));
924 				if (counter->cgrp)
925 					fprintf(output, "%s%s",
926 						csv_sep, counter->cgrp->name);
927 
928 				fputc('\n', output);
929 				continue;
930 			}
931 
932 			if (nsec_counter(counter))
933 				nsec_printout(sock, nr, counter, val);
934 			else
935 				abs_printout(sock, nr, counter, val);
936 
937 			if (!csv_output) {
938 				print_noise(counter, 1.0);
939 
940 				if (run != ena)
941 					fprintf(output, "  (%.2f%%)",
942 						100.0 * run / ena);
943 			}
944 			fputc('\n', output);
945 		}
946 	}
947 }
948 
949 /*
950  * Print out the results of a single counter:
951  * aggregated counts in system-wide mode
952  */
953 static void print_counter_aggr(struct perf_evsel *counter, char *prefix)
954 {
955 	struct perf_stat *ps = counter->priv;
956 	double avg = avg_stats(&ps->res_stats[0]);
957 	int scaled = counter->counts->scaled;
958 
959 	if (prefix)
960 		fprintf(output, "%s", prefix);
961 
962 	if (scaled == -1) {
963 		fprintf(output, "%*s%s%*s",
964 			csv_output ? 0 : 18,
965 			counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
966 			csv_sep,
967 			csv_output ? 0 : -24,
968 			perf_evsel__name(counter));
969 
970 		if (counter->cgrp)
971 			fprintf(output, "%s%s", csv_sep, counter->cgrp->name);
972 
973 		fputc('\n', output);
974 		return;
975 	}
976 
977 	if (nsec_counter(counter))
978 		nsec_printout(-1, 0, counter, avg);
979 	else
980 		abs_printout(-1, 0, counter, avg);
981 
982 	print_noise(counter, avg);
983 
984 	if (csv_output) {
985 		fputc('\n', output);
986 		return;
987 	}
988 
989 	if (scaled) {
990 		double avg_enabled, avg_running;
991 
992 		avg_enabled = avg_stats(&ps->res_stats[1]);
993 		avg_running = avg_stats(&ps->res_stats[2]);
994 
995 		fprintf(output, " [%5.2f%%]", 100 * avg_running / avg_enabled);
996 	}
997 	fprintf(output, "\n");
998 }
999 
1000 /*
1001  * Print out the results of a single counter:
1002  * does not use aggregated count in system-wide
1003  */
1004 static void print_counter(struct perf_evsel *counter, char *prefix)
1005 {
1006 	u64 ena, run, val;
1007 	int cpu;
1008 
1009 	for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
1010 		val = counter->counts->cpu[cpu].val;
1011 		ena = counter->counts->cpu[cpu].ena;
1012 		run = counter->counts->cpu[cpu].run;
1013 
1014 		if (prefix)
1015 			fprintf(output, "%s", prefix);
1016 
1017 		if (run == 0 || ena == 0) {
1018 			fprintf(output, "CPU%*d%s%*s%s%*s",
1019 				csv_output ? 0 : -4,
1020 				perf_evsel__cpus(counter)->map[cpu], csv_sep,
1021 				csv_output ? 0 : 18,
1022 				counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
1023 				csv_sep,
1024 				csv_output ? 0 : -24,
1025 				perf_evsel__name(counter));
1026 
1027 			if (counter->cgrp)
1028 				fprintf(output, "%s%s",
1029 					csv_sep, counter->cgrp->name);
1030 
1031 			fputc('\n', output);
1032 			continue;
1033 		}
1034 
1035 		if (nsec_counter(counter))
1036 			nsec_printout(cpu, 0, counter, val);
1037 		else
1038 			abs_printout(cpu, 0, counter, val);
1039 
1040 		if (!csv_output) {
1041 			print_noise(counter, 1.0);
1042 
1043 			if (run != ena)
1044 				fprintf(output, "  (%.2f%%)",
1045 					100.0 * run / ena);
1046 		}
1047 		fputc('\n', output);
1048 	}
1049 }
1050 
1051 static void print_stat(int argc, const char **argv)
1052 {
1053 	struct perf_evsel *counter;
1054 	int i;
1055 
1056 	fflush(stdout);
1057 
1058 	if (!csv_output) {
1059 		fprintf(output, "\n");
1060 		fprintf(output, " Performance counter stats for ");
1061 		if (!perf_target__has_task(&target)) {
1062 			fprintf(output, "\'%s", argv[0]);
1063 			for (i = 1; i < argc; i++)
1064 				fprintf(output, " %s", argv[i]);
1065 		} else if (target.pid)
1066 			fprintf(output, "process id \'%s", target.pid);
1067 		else
1068 			fprintf(output, "thread id \'%s", target.tid);
1069 
1070 		fprintf(output, "\'");
1071 		if (run_count > 1)
1072 			fprintf(output, " (%d runs)", run_count);
1073 		fprintf(output, ":\n\n");
1074 	}
1075 
1076 	if (aggr_socket)
1077 		print_aggr_socket(NULL);
1078 	else if (no_aggr) {
1079 		list_for_each_entry(counter, &evsel_list->entries, node)
1080 			print_counter(counter, NULL);
1081 	} else {
1082 		list_for_each_entry(counter, &evsel_list->entries, node)
1083 			print_counter_aggr(counter, NULL);
1084 	}
1085 
1086 	if (!csv_output) {
1087 		if (!null_run)
1088 			fprintf(output, "\n");
1089 		fprintf(output, " %17.9f seconds time elapsed",
1090 				avg_stats(&walltime_nsecs_stats)/1e9);
1091 		if (run_count > 1) {
1092 			fprintf(output, "                                        ");
1093 			print_noise_pct(stddev_stats(&walltime_nsecs_stats),
1094 					avg_stats(&walltime_nsecs_stats));
1095 		}
1096 		fprintf(output, "\n\n");
1097 	}
1098 }
1099 
1100 static volatile int signr = -1;
1101 
1102 static void skip_signal(int signo)
1103 {
1104 	if ((child_pid == -1) || interval)
1105 		done = 1;
1106 
1107 	signr = signo;
1108 }
1109 
1110 static void sig_atexit(void)
1111 {
1112 	if (child_pid != -1)
1113 		kill(child_pid, SIGTERM);
1114 
1115 	if (signr == -1)
1116 		return;
1117 
1118 	signal(signr, SIG_DFL);
1119 	kill(getpid(), signr);
1120 }
1121 
1122 static int stat__set_big_num(const struct option *opt __maybe_unused,
1123 			     const char *s __maybe_unused, int unset)
1124 {
1125 	big_num_opt = unset ? 0 : 1;
1126 	return 0;
1127 }
1128 
1129 /*
1130  * Add default attributes, if there were no attributes specified or
1131  * if -d/--detailed, -d -d or -d -d -d is used:
1132  */
1133 static int add_default_attributes(void)
1134 {
1135 	struct perf_event_attr default_attrs[] = {
1136 
1137   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK		},
1138   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES	},
1139   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS		},
1140   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS		},
1141 
1142   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES		},
1143   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND	},
1144   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND	},
1145   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS		},
1146   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS	},
1147   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES		},
1148 
1149 };
1150 
1151 /*
1152  * Detailed stats (-d), covering the L1 and last level data caches:
1153  */
1154 	struct perf_event_attr detailed_attrs[] = {
1155 
1156   { .type = PERF_TYPE_HW_CACHE,
1157     .config =
1158 	 PERF_COUNT_HW_CACHE_L1D		<<  0  |
1159 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1160 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1161 
1162   { .type = PERF_TYPE_HW_CACHE,
1163     .config =
1164 	 PERF_COUNT_HW_CACHE_L1D		<<  0  |
1165 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1166 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1167 
1168   { .type = PERF_TYPE_HW_CACHE,
1169     .config =
1170 	 PERF_COUNT_HW_CACHE_LL			<<  0  |
1171 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1172 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1173 
1174   { .type = PERF_TYPE_HW_CACHE,
1175     .config =
1176 	 PERF_COUNT_HW_CACHE_LL			<<  0  |
1177 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1178 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1179 };
1180 
1181 /*
1182  * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
1183  */
1184 	struct perf_event_attr very_detailed_attrs[] = {
1185 
1186   { .type = PERF_TYPE_HW_CACHE,
1187     .config =
1188 	 PERF_COUNT_HW_CACHE_L1I		<<  0  |
1189 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1190 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1191 
1192   { .type = PERF_TYPE_HW_CACHE,
1193     .config =
1194 	 PERF_COUNT_HW_CACHE_L1I		<<  0  |
1195 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1196 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1197 
1198   { .type = PERF_TYPE_HW_CACHE,
1199     .config =
1200 	 PERF_COUNT_HW_CACHE_DTLB		<<  0  |
1201 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1202 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1203 
1204   { .type = PERF_TYPE_HW_CACHE,
1205     .config =
1206 	 PERF_COUNT_HW_CACHE_DTLB		<<  0  |
1207 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1208 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1209 
1210   { .type = PERF_TYPE_HW_CACHE,
1211     .config =
1212 	 PERF_COUNT_HW_CACHE_ITLB		<<  0  |
1213 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1214 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1215 
1216   { .type = PERF_TYPE_HW_CACHE,
1217     .config =
1218 	 PERF_COUNT_HW_CACHE_ITLB		<<  0  |
1219 	(PERF_COUNT_HW_CACHE_OP_READ		<<  8) |
1220 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1221 
1222 };
1223 
1224 /*
1225  * Very, very detailed stats (-d -d -d), adding prefetch events:
1226  */
1227 	struct perf_event_attr very_very_detailed_attrs[] = {
1228 
1229   { .type = PERF_TYPE_HW_CACHE,
1230     .config =
1231 	 PERF_COUNT_HW_CACHE_L1D		<<  0  |
1232 	(PERF_COUNT_HW_CACHE_OP_PREFETCH	<<  8) |
1233 	(PERF_COUNT_HW_CACHE_RESULT_ACCESS	<< 16)				},
1234 
1235   { .type = PERF_TYPE_HW_CACHE,
1236     .config =
1237 	 PERF_COUNT_HW_CACHE_L1D		<<  0  |
1238 	(PERF_COUNT_HW_CACHE_OP_PREFETCH	<<  8) |
1239 	(PERF_COUNT_HW_CACHE_RESULT_MISS	<< 16)				},
1240 };
1241 
1242 	/* Set attrs if no event is selected and !null_run: */
1243 	if (null_run)
1244 		return 0;
1245 
1246 	if (!evsel_list->nr_entries) {
1247 		if (perf_evlist__add_default_attrs(evsel_list, default_attrs) < 0)
1248 			return -1;
1249 	}
1250 
1251 	/* Detailed events get appended to the event list: */
1252 
1253 	if (detailed_run <  1)
1254 		return 0;
1255 
1256 	/* Append detailed run extra attributes: */
1257 	if (perf_evlist__add_default_attrs(evsel_list, detailed_attrs) < 0)
1258 		return -1;
1259 
1260 	if (detailed_run < 2)
1261 		return 0;
1262 
1263 	/* Append very detailed run extra attributes: */
1264 	if (perf_evlist__add_default_attrs(evsel_list, very_detailed_attrs) < 0)
1265 		return -1;
1266 
1267 	if (detailed_run < 3)
1268 		return 0;
1269 
1270 	/* Append very, very detailed run extra attributes: */
1271 	return perf_evlist__add_default_attrs(evsel_list, very_very_detailed_attrs);
1272 }
1273 
1274 int cmd_stat(int argc, const char **argv, const char *prefix __maybe_unused)
1275 {
1276 	bool append_file = false;
1277 	int output_fd = 0;
1278 	const char *output_name	= NULL;
1279 	const struct option options[] = {
1280 	OPT_CALLBACK('e', "event", &evsel_list, "event",
1281 		     "event selector. use 'perf list' to list available events",
1282 		     parse_events_option),
1283 	OPT_CALLBACK(0, "filter", &evsel_list, "filter",
1284 		     "event filter", parse_filter),
1285 	OPT_BOOLEAN('i', "no-inherit", &no_inherit,
1286 		    "child tasks do not inherit counters"),
1287 	OPT_STRING('p', "pid", &target.pid, "pid",
1288 		   "stat events on existing process id"),
1289 	OPT_STRING('t', "tid", &target.tid, "tid",
1290 		   "stat events on existing thread id"),
1291 	OPT_BOOLEAN('a', "all-cpus", &target.system_wide,
1292 		    "system-wide collection from all CPUs"),
1293 	OPT_BOOLEAN('g', "group", &group,
1294 		    "put the counters into a counter group"),
1295 	OPT_BOOLEAN('c', "scale", &scale, "scale/normalize counters"),
1296 	OPT_INCR('v', "verbose", &verbose,
1297 		    "be more verbose (show counter open errors, etc)"),
1298 	OPT_INTEGER('r', "repeat", &run_count,
1299 		    "repeat command and print average + stddev (max: 100)"),
1300 	OPT_BOOLEAN('n', "null", &null_run,
1301 		    "null run - dont start any counters"),
1302 	OPT_INCR('d', "detailed", &detailed_run,
1303 		    "detailed run - start a lot of events"),
1304 	OPT_BOOLEAN('S', "sync", &sync_run,
1305 		    "call sync() before starting a run"),
1306 	OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
1307 			   "print large numbers with thousands\' separators",
1308 			   stat__set_big_num),
1309 	OPT_STRING('C', "cpu", &target.cpu_list, "cpu",
1310 		    "list of cpus to monitor in system-wide"),
1311 	OPT_BOOLEAN('A', "no-aggr", &no_aggr, "disable CPU count aggregation"),
1312 	OPT_STRING('x', "field-separator", &csv_sep, "separator",
1313 		   "print counts with custom separator"),
1314 	OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
1315 		     "monitor event in cgroup name only", parse_cgroups),
1316 	OPT_STRING('o', "output", &output_name, "file", "output file name"),
1317 	OPT_BOOLEAN(0, "append", &append_file, "append to the output file"),
1318 	OPT_INTEGER(0, "log-fd", &output_fd,
1319 		    "log output to fd, instead of stderr"),
1320 	OPT_STRING(0, "pre", &pre_cmd, "command",
1321 			"command to run prior to the measured command"),
1322 	OPT_STRING(0, "post", &post_cmd, "command",
1323 			"command to run after to the measured command"),
1324 	OPT_UINTEGER('I', "interval-print", &interval,
1325 		    "print counts at regular interval in ms (>= 100)"),
1326 	OPT_BOOLEAN(0, "aggr-socket", &aggr_socket, "aggregate counts per processor socket"),
1327 	OPT_END()
1328 	};
1329 	const char * const stat_usage[] = {
1330 		"perf stat [<options>] [<command>]",
1331 		NULL
1332 	};
1333 	struct perf_evsel *pos;
1334 	int status = -ENOMEM, run_idx;
1335 	const char *mode;
1336 
1337 	setlocale(LC_ALL, "");
1338 
1339 	evsel_list = perf_evlist__new(NULL, NULL);
1340 	if (evsel_list == NULL)
1341 		return -ENOMEM;
1342 
1343 	argc = parse_options(argc, argv, options, stat_usage,
1344 		PARSE_OPT_STOP_AT_NON_OPTION);
1345 
1346 	output = stderr;
1347 	if (output_name && strcmp(output_name, "-"))
1348 		output = NULL;
1349 
1350 	if (output_name && output_fd) {
1351 		fprintf(stderr, "cannot use both --output and --log-fd\n");
1352 		usage_with_options(stat_usage, options);
1353 	}
1354 
1355 	if (output_fd < 0) {
1356 		fprintf(stderr, "argument to --log-fd must be a > 0\n");
1357 		usage_with_options(stat_usage, options);
1358 	}
1359 
1360 	if (!output) {
1361 		struct timespec tm;
1362 		mode = append_file ? "a" : "w";
1363 
1364 		output = fopen(output_name, mode);
1365 		if (!output) {
1366 			perror("failed to create output file");
1367 			return -1;
1368 		}
1369 		clock_gettime(CLOCK_REALTIME, &tm);
1370 		fprintf(output, "# started on %s\n", ctime(&tm.tv_sec));
1371 	} else if (output_fd > 0) {
1372 		mode = append_file ? "a" : "w";
1373 		output = fdopen(output_fd, mode);
1374 		if (!output) {
1375 			perror("Failed opening logfd");
1376 			return -errno;
1377 		}
1378 	}
1379 
1380 	if (csv_sep) {
1381 		csv_output = true;
1382 		if (!strcmp(csv_sep, "\\t"))
1383 			csv_sep = "\t";
1384 	} else
1385 		csv_sep = DEFAULT_SEPARATOR;
1386 
1387 	/*
1388 	 * let the spreadsheet do the pretty-printing
1389 	 */
1390 	if (csv_output) {
1391 		/* User explicitly passed -B? */
1392 		if (big_num_opt == 1) {
1393 			fprintf(stderr, "-B option not supported with -x\n");
1394 			usage_with_options(stat_usage, options);
1395 		} else /* Nope, so disable big number formatting */
1396 			big_num = false;
1397 	} else if (big_num_opt == 0) /* User passed --no-big-num */
1398 		big_num = false;
1399 
1400 	if (!argc && !perf_target__has_task(&target))
1401 		usage_with_options(stat_usage, options);
1402 	if (run_count <= 0)
1403 		usage_with_options(stat_usage, options);
1404 
1405 	/* no_aggr, cgroup are for system-wide only */
1406 	if ((no_aggr || nr_cgroups) && !perf_target__has_cpu(&target)) {
1407 		fprintf(stderr, "both cgroup and no-aggregation "
1408 			"modes only available in system-wide mode\n");
1409 
1410 		usage_with_options(stat_usage, options);
1411 	}
1412 
1413 	if (aggr_socket) {
1414 		if (!perf_target__has_cpu(&target)) {
1415 			fprintf(stderr, "--aggr-socket only available in system-wide mode (-a)\n");
1416 			usage_with_options(stat_usage, options);
1417 		}
1418 		no_aggr = true;
1419 	}
1420 
1421 	if (add_default_attributes())
1422 		goto out;
1423 
1424 	perf_target__validate(&target);
1425 
1426 	if (perf_evlist__create_maps(evsel_list, &target) < 0) {
1427 		if (perf_target__has_task(&target))
1428 			pr_err("Problems finding threads of monitor\n");
1429 		if (perf_target__has_cpu(&target))
1430 			perror("failed to parse CPUs map");
1431 
1432 		usage_with_options(stat_usage, options);
1433 		return -1;
1434 	}
1435 	if (interval && interval < 100) {
1436 		pr_err("print interval must be >= 100ms\n");
1437 		usage_with_options(stat_usage, options);
1438 		return -1;
1439 	}
1440 
1441 	list_for_each_entry(pos, &evsel_list->entries, node) {
1442 		if (perf_evsel__alloc_stat_priv(pos) < 0 ||
1443 		    perf_evsel__alloc_counts(pos, perf_evsel__nr_cpus(pos)) < 0)
1444 			goto out_free_fd;
1445 	}
1446 	if (interval) {
1447 		list_for_each_entry(pos, &evsel_list->entries, node) {
1448 			if (perf_evsel__alloc_prev_raw_counts(pos) < 0)
1449 				goto out_free_fd;
1450 		}
1451 	}
1452 
1453 	/*
1454 	 * We dont want to block the signals - that would cause
1455 	 * child tasks to inherit that and Ctrl-C would not work.
1456 	 * What we want is for Ctrl-C to work in the exec()-ed
1457 	 * task, but being ignored by perf stat itself:
1458 	 */
1459 	atexit(sig_atexit);
1460 	signal(SIGINT,  skip_signal);
1461 	signal(SIGCHLD, skip_signal);
1462 	signal(SIGALRM, skip_signal);
1463 	signal(SIGABRT, skip_signal);
1464 
1465 	status = 0;
1466 	for (run_idx = 0; run_idx < run_count; run_idx++) {
1467 		if (run_count != 1 && verbose)
1468 			fprintf(output, "[ perf stat: executing run #%d ... ]\n",
1469 				run_idx + 1);
1470 
1471 		status = run_perf_stat(argc, argv);
1472 	}
1473 
1474 	if (status != -1 && !interval)
1475 		print_stat(argc, argv);
1476 out_free_fd:
1477 	list_for_each_entry(pos, &evsel_list->entries, node) {
1478 		perf_evsel__free_stat_priv(pos);
1479 		perf_evsel__free_counts(pos);
1480 		perf_evsel__free_prev_raw_counts(pos);
1481 	}
1482 	perf_evlist__delete_maps(evsel_list);
1483 out:
1484 	perf_evlist__delete(evsel_list);
1485 	return status;
1486 }
1487